diff --git a/CONTRIBUTING.md b/CONTRIBUTING.md index db4b1581ae671b1e676e215c9a80dfaab832fa21..f598999f351c10f8bd01dfbd3ad8897f19d570e8 100644 --- a/CONTRIBUTING.md +++ b/CONTRIBUTING.md @@ -107,7 +107,7 @@ diff /tmp/my_cc_file.cc #### Python coding style Changes to TensorFlow Python code should conform to -[Google Python Style Guide](https://google.github.io/styleguide/pyguide.html) +[Google Python Style Guide](https://github.com/google/styleguide/blob/gh-pages/pyguide.md) Use `pylint` to check your Python changes. To install `pylint` and retrieve TensorFlow's custom style definition: diff --git a/README.md b/README.md index 42d7bbc104f64617f216d2a8b6379e2f58358cad..05fcb23f7edd657f2ea495d848fadc226e56b524 100644 --- a/README.md +++ b/README.md @@ -96,6 +96,8 @@ The TensorFlow project strives to abide by generally accepted best practices in | --- | --- | --- | | **IBM s390x** | [![Build Status](http://ibmz-ci.osuosl.org/job/TensorFlow_IBMZ_CI/badge/icon)](http://ibmz-ci.osuosl.org/job/TensorFlow_IBMZ_CI/) | TBA | | **IBM ppc64le CPU** | [![Build Status](http://powerci.osuosl.org/job/TensorFlow_Ubuntu_16.04_CPU/badge/icon)](http://powerci.osuosl.org/job/TensorFlow_Ubuntu_16.04_CPU/) | TBA | +| **IBM ppc64le GPU** | [![Build Status](http://powerci.osuosl.org/job/TensorFlow_Ubuntu_16.04_PPC64LE_GPU/badge/icon)](http://powerci.osuosl.org/job/TensorFlow_Ubuntu_16.04_PPC64LE_GPU/) | TBA | +| **Linux CPU with IntelĀ® MKL-DNNĀ®** | [![Build Status](https://tensorflow-ci.intel.com/job/tensorflow-mkl-linux-cpu/badge/icon)](https://tensorflow-ci.intel.com/job/tensorflow-mkl-linux-cpu/) | TBA | ## For more information diff --git a/RELEASE.md b/RELEASE.md index f0a7afe6847e591faa1d2238358be5db0f8503b3..7bb1e3e1c8bd2b0471a8259cb2354d5f4f4b777a 100644 --- a/RELEASE.md +++ b/RELEASE.md @@ -6,7 +6,7 @@ * Update `tf.keras` to the Keras 2.1.6 API. * Added [`tf.keras.layers.CuDNNGRU`](https://www.tensorflow.org/versions/r1.9/api_docs/python/tf/keras/layers/CuDNNGRU) and [`tf.keras.layers.CuDNNLSTM`](https://www.tensorflow.org/versions/r1.9/api_docs/python/tf/keras/layers/CuDNNLSTM) layers. [Try it](https://colab.sandbox.google.com/github/tensorflow/tensorflow/blob/master/tensorflow/contrib/eager/python/examples/nmt_with_attention/nmt_with_attention.ipynb?linkId=53292082). * Adding support of core [feature columns](https://www.tensorflow.org/get_started/feature_columns) and [losses](https://www.tensorflow.org/api_docs/python/tf/losses) to [gradient boosted trees estimators](https://github.com/tensorflow/models/tree/master/official/boosted_trees). -* The [python interface](https://tensorflow-dot-devsite.googleplex.com/versions/r1.9/api_docs/python/tf/contrib/lite) +* The [python interface](https://www.tensorflow.org/versions/r1.9/api_docs/python/tf/contrib/lite) for the [TFLite Optimizing Converter](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/contrib/lite/toco/README.md) has been expanded, and the command line interface (AKA: `toco`, `tflite_convert`) is once again included in the standard `pip` installation. @@ -21,9 +21,10 @@ * The [distributions.Bijector](https://www.tensorflow.org/versions/r1.9/api_docs/python/tf/contrib/distributions/bijectors/Bijector) API supports broadcasting for Bijectors with new API changes. -## Breaking Chances +## Breaking Changes * If you're opening empty variable scopes; replace `variable_scope('', ...)` by `variable_scope(tf.get_variable_scope(), ...)`. + * Headers used for building custom ops have been moved from site-packages/external into site-packages/tensorflow/include/external. ## Bug Fixes and Other Changes @@ -32,7 +33,6 @@ * Using `tf.keras.layers` with custom variable scopes. * Using `tf.layers` in a subclassed `tf.keras.Model` class. See [here](https://www.tensorflow.org/versions/r1.9/api_docs/python/tf/layers) for more details - * `tf.data`: * The `DatasetBase::DebugString()` method is now `const`. * Added the `tf.contrib.data.sample_from_datasets()` API for randomly sampling from multiple datasets. diff --git a/configure.py b/configure.py index ad585fa52e571d62d11864531476e46b2f15f297..8930c3a1f16180d6810ad941d689464bef56e930 100644 --- a/configure.py +++ b/configure.py @@ -835,6 +835,8 @@ def set_tf_cuda_version(environ_cp): '[Default is %s]: ') % (tf_cuda_version, default_cuda_path) cuda_toolkit_path = get_from_env_or_user_or_default( environ_cp, 'CUDA_TOOLKIT_PATH', ask_cuda_path, default_cuda_path) + if is_windows() or is_cygwin(): + cuda_toolkit_path = cygpath(cuda_toolkit_path) if is_windows(): cuda_rt_lib_path = 'lib/x64/cudart.lib' @@ -1134,7 +1136,9 @@ def set_tf_nccl_install_path(environ_cp): nccl_lib_path = os.path.join(nccl_install_path, nccl_lib_path) nccl_hdr_path = os.path.join(nccl_install_path, 'include/nccl.h') - if os.path.exists(nccl_lib_path) and os.path.exists(nccl_hdr_path): + nccl_license_path = os.path.join(nccl_install_path, 'NCCL-SLA.txt') + if os.path.exists(nccl_lib_path) and os.path.exists( + nccl_hdr_path) and os.path.exists(nccl_license_path): # Set NCCL_INSTALL_PATH environ_cp['NCCL_INSTALL_PATH'] = nccl_install_path write_action_env_to_bazelrc('NCCL_INSTALL_PATH', nccl_install_path) @@ -1232,28 +1236,13 @@ def set_tf_cuda_compute_capabilities(environ_cp): def set_other_cuda_vars(environ_cp): """Set other CUDA related variables.""" - if is_windows(): - # The following three variables are needed for MSVC toolchain configuration - # in Bazel - environ_cp['CUDA_PATH'] = environ_cp.get('CUDA_TOOLKIT_PATH') - environ_cp['CUDA_COMPUTE_CAPABILITIES'] = environ_cp.get( - 'TF_CUDA_COMPUTE_CAPABILITIES') - environ_cp['NO_WHOLE_ARCHIVE_OPTION'] = 1 - write_action_env_to_bazelrc('CUDA_PATH', environ_cp.get('CUDA_PATH')) - write_action_env_to_bazelrc('CUDA_COMPUTE_CAPABILITIE', - environ_cp.get('CUDA_COMPUTE_CAPABILITIE')) - write_action_env_to_bazelrc('NO_WHOLE_ARCHIVE_OPTION', - environ_cp.get('NO_WHOLE_ARCHIVE_OPTION')) - write_to_bazelrc('build --config=win-cuda') - write_to_bazelrc('test --config=win-cuda') + # If CUDA is enabled, always use GPU during build and test. + if environ_cp.get('TF_CUDA_CLANG') == '1': + write_to_bazelrc('build --config=cuda_clang') + write_to_bazelrc('test --config=cuda_clang') else: - # If CUDA is enabled, always use GPU during build and test. - if environ_cp.get('TF_CUDA_CLANG') == '1': - write_to_bazelrc('build --config=cuda_clang') - write_to_bazelrc('test --config=cuda_clang') - else: - write_to_bazelrc('build --config=cuda') - write_to_bazelrc('test --config=cuda') + write_to_bazelrc('build --config=cuda') + write_to_bazelrc('test --config=cuda') def set_host_cxx_compiler(environ_cp): @@ -1447,7 +1436,7 @@ def main(): setup_python(environ_cp) if is_windows(): - environ_cp['TF_NEED_S3'] = '0' + environ_cp['TF_NEED_AWS'] = '0' environ_cp['TF_NEED_GCP'] = '0' environ_cp['TF_NEED_HDFS'] = '0' environ_cp['TF_NEED_JEMALLOC'] = '0' @@ -1471,8 +1460,8 @@ def main(): 'with_gcp_support', True, 'gcp') set_build_var(environ_cp, 'TF_NEED_HDFS', 'Hadoop File System', 'with_hdfs_support', True, 'hdfs') - set_build_var(environ_cp, 'TF_NEED_S3', 'Amazon S3 File System', - 'with_s3_support', True, 's3') + set_build_var(environ_cp, 'TF_NEED_AWS', 'Amazon AWS Platform', + 'with_aws_support', True, 'aws') set_build_var(environ_cp, 'TF_NEED_KAFKA', 'Apache Kafka Platform', 'with_kafka_support', True, 'kafka') set_build_var(environ_cp, 'TF_ENABLE_XLA', 'XLA JIT', 'with_xla_support', diff --git a/tensorflow/BUILD b/tensorflow/BUILD index a15d033013f573ca7a182cc72cb4b7a8cec0e273..518c2b0489021815c0480acb35a58717d6ca9359 100644 --- a/tensorflow/BUILD +++ b/tensorflow/BUILD @@ -20,7 +20,7 @@ load( "tf_additional_binary_deps", ) load( - "//tensorflow/tools/api/generator:api_gen.bzl", + "//tensorflow/python/tools/api/generator:api_gen.bzl", "gen_api_init_files", # @unused ) @@ -216,8 +216,8 @@ config_setting( ) config_setting( - name = "with_s3_support", - define_values = {"with_s3_support": "true"}, + name = "with_aws_support", + define_values = {"with_aws_support": "true"}, visibility = ["//visibility:public"], ) @@ -244,8 +244,8 @@ config_setting( ) config_setting( - name = "with_s3_support_windows_override", - define_values = {"with_s3_support": "true"}, + name = "with_aws_support_windows_override", + define_values = {"with_aws_support": "true"}, values = {"cpu": "x64_windows"}, visibility = ["//visibility:public"], ) @@ -257,6 +257,13 @@ config_setting( visibility = ["//visibility:public"], ) +config_setting( + name = "with_cuda_support_windows_override", + define_values = {"using_cuda_nvcc": "true"}, + values = {"cpu": "x64_windows"}, + visibility = ["//visibility:public"], +) + config_setting( name = "with_gcp_support_android_override", define_values = {"with_gcp_support": "true"}, @@ -272,8 +279,8 @@ config_setting( ) config_setting( - name = "with_s3_support_android_override", - define_values = {"with_s3_support": "true"}, + name = "with_aws_support_android_override", + define_values = {"with_aws_support": "true"}, values = {"crosstool_top": "//external:android/crosstool"}, visibility = ["//visibility:public"], ) @@ -293,8 +300,8 @@ config_setting( ) config_setting( - name = "with_s3_support_ios_override", - define_values = {"with_s3_support": "true"}, + name = "with_aws_support_ios_override", + define_values = {"with_aws_support": "true"}, values = {"crosstool_top": "//tools/osx/crosstool:crosstool"}, visibility = ["//visibility:public"], ) diff --git a/tensorflow/c/c_api.cc b/tensorflow/c/c_api.cc index 9d5f98d4d66aeb06ee13bc010923eddbfca70497..5c218d3f25e01f0e78916d4a5a8b1d2751f9dc25 100644 --- a/tensorflow/c/c_api.cc +++ b/tensorflow/c/c_api.cc @@ -2068,7 +2068,8 @@ TF_ImportGraphDefResults* TF_GraphImportGraphDefWithResults( TF_Graph* graph, const TF_Buffer* graph_def, const TF_ImportGraphDefOptions* options, TF_Status* status) { GraphDef def; - if (!tensorflow::ParseProtoUnlimited(&def, graph_def->data, graph_def->length)) { + if (!tensorflow::ParseProtoUnlimited(&def, graph_def->data, + graph_def->length)) { status->status = InvalidArgument("Invalid GraphDef"); return nullptr; } @@ -2098,7 +2099,8 @@ void TF_GraphImportGraphDefWithReturnOutputs( return; } GraphDef def; - if (!tensorflow::ParseProtoUnlimited(&def, graph_def->data, graph_def->length)) { + if (!tensorflow::ParseProtoUnlimited(&def, graph_def->data, + graph_def->length)) { status->status = InvalidArgument("Invalid GraphDef"); return; } @@ -2414,7 +2416,18 @@ void TF_AddGradients(TF_Graph* g, TF_Output* y, int ny, TF_Output* x, int nx, for (int i = first_new_node_id; i < g->graph.num_node_ids(); ++i) { Node* n = g->graph.FindNodeId(i); if (n == nullptr) continue; - g->name_map[n->name()] = n; + // We have a convoluted scheme here: Using the C++ graph construction API + // to add potentially many nodes to the graph without running the checks + // (such as uniqueness of the names of nodes) we run with other functions + // that add a node to the graph (like TF_FinishOperation). + if (!g->name_map.insert(std::make_pair(n->name(), n)).second) { + status->status = tensorflow::errors::Internal( + "BUG: The API allowed construction of a graph with duplicate node " + "names (", + n->name(), + "). This is a bug. Please file an issue at " + "https://github.com/tensorflow/tensorflow/issues."); + } } } diff --git a/tensorflow/c/c_api_experimental.cc b/tensorflow/c/c_api_experimental.cc index 95b04f9058afdfaadbc24f0238860279fcd3e800..170046c8024dc85c899108b254cd3a95a3be4096 100644 --- a/tensorflow/c/c_api_experimental.cc +++ b/tensorflow/c/c_api_experimental.cc @@ -57,6 +57,33 @@ void TF_EnableXLACompilation(TF_SessionOptions* options, unsigned char enable) { } } +TF_Buffer* TF_CreateConfig(unsigned char enable_xla_compilation, + unsigned char gpu_memory_allow_growth) { + tensorflow::ConfigProto config; + auto* optimizer_options = + config.mutable_graph_options()->mutable_optimizer_options(); + if (enable_xla_compilation) { + optimizer_options->set_global_jit_level(tensorflow::OptimizerOptions::ON_1); + + // These XLA flags are needed to trigger XLA properly from C (more generally + // non-Python) clients. If this API is called again with `enable` set to + // false, it is safe to keep these flag values as is. + tensorflow::legacy_flags::MarkForCompilationPassFlags* flags = + tensorflow::legacy_flags::GetMarkForCompilationPassFlags(); + flags->tf_xla_cpu_global_jit = true; + flags->tf_xla_min_cluster_size = 1; + } else { + optimizer_options->set_global_jit_level(tensorflow::OptimizerOptions::OFF); + } + + auto* gpu_options = config.mutable_gpu_options(); + gpu_options->set_allow_growth(gpu_memory_allow_growth); + + TF_Buffer* ret = TF_NewBuffer(); + TF_CHECK_OK(MessageToBuffer(config, ret)); + return ret; +} + const char* TF_GraphDebugString(TF_Graph* graph, size_t* len) { tensorflow::mutex_lock c(graph->mu); const auto& debug_str = graph->graph.ToGraphDefDebug().DebugString(); diff --git a/tensorflow/c/c_api_experimental.h b/tensorflow/c/c_api_experimental.h index 20bdace40f1272ded06e710034053a7610326e7f..2d81c01e0dd056e9beb3b45f24809381554a7924 100644 --- a/tensorflow/c/c_api_experimental.h +++ b/tensorflow/c/c_api_experimental.h @@ -55,11 +55,21 @@ extern "C" { // set XLA flag values to prepare for XLA compilation. Otherwise set // global_jit_level to OFF. // -// This API is syntax sugar over TF_SetConfig(), and is used by clients that -// cannot read/write the tensorflow.ConfigProto proto. +// This and the next API are syntax sugar over TF_SetConfig(), and is used by +// clients that cannot read/write the tensorflow.ConfigProto proto. +// TODO: Migrate to TF_CreateConfig() below. TF_CAPI_EXPORT extern void TF_EnableXLACompilation(TF_SessionOptions* options, unsigned char enable); +// Create a serialized tensorflow.ConfigProto proto, where: +// +// a) ConfigProto.optimizer_options.global_jit_level is set to to ON_1 if +// `enable_xla_compilation` is non-zero, and OFF otherwise. +// b) ConfigProto.gpu_options.allow_growth is set to `gpu_memory_allow_growth`. +TF_CAPI_EXPORT extern TF_Buffer* TF_CreateConfig( + unsigned char enable_xla_compilation, + unsigned char gpu_memory_allow_growth); + // Returns the graph content in a human-readable format, with length set in // `len`. The format is subject to change in the future. // The returned string is heap-allocated, and caller should call free() on it. diff --git a/tensorflow/c/c_api_test.cc b/tensorflow/c/c_api_test.cc index 577f10c5e69ea9ecbe8ce821c6bd5167e98bef25..bc04b53fbb7fa9ba46228ae5a4ec8ee96df5f3dc 100644 --- a/tensorflow/c/c_api_test.cc +++ b/tensorflow/c/c_api_test.cc @@ -1160,7 +1160,7 @@ TEST(CAPI, GetOpDef) { } void StringVectorToArrays(const std::vector& v, - std::unique_ptr* ptrs, + std::unique_ptr* ptrs, std::unique_ptr* lens) { ptrs->reset(new const void*[v.size()]); lens->reset(new size_t[v.size()]); @@ -1196,7 +1196,7 @@ class CApiColocationTest : public ::testing::Test { void SetViaStringList(TF_OperationDescription* desc, const std::vector& list) { - std::unique_ptr list_ptrs; + std::unique_ptr list_ptrs; std::unique_ptr list_lens; StringVectorToArrays(list, &list_ptrs, &list_lens); TF_SetAttrStringList(desc, tensorflow::kColocationAttrName, list_ptrs.get(), @@ -1700,6 +1700,61 @@ TEST_F(CApiGradientsTest, OpWithNoGradientRegistered_NoGradInputs) { TestGradientsError(false); } +void ScalarFloatFromTensor(const TF_Tensor* t, float* f) { + ASSERT_TRUE(t != nullptr); + ASSERT_EQ(TF_FLOAT, TF_TensorType(t)); + ASSERT_EQ(0, TF_NumDims(t)); + ASSERT_EQ(4, TF_TensorByteSize(t)); + float* p = static_cast(TF_TensorData(t)); + *f = *p; +} + +TEST_F(CApiGradientsTest, MultipleCallsToAddGradients) { + const float X = 3.0f, Y = 7.0f; + TF_Operation* x = Placeholder(graph_, s_, "x", TF_FLOAT); + TF_Operation* y = Placeholder(graph_, s_, "y", TF_FLOAT); + TF_Operation* xy = Mul(x, y, graph_, s_, "xy"); + TF_Output dxy_dx, dxy_dy; + + TF_Output outputs[1] = {{xy, 0}}; + TF_Output inputs[1] = {{x, 0}}; + TF_AddGradients(graph_, outputs, 1, inputs, 1, nullptr, s_, &dxy_dx); + ASSERT_EQ(TF_OK, TF_GetCode(s_)) << TF_Message(s_); + + inputs[0] = {y, 0}; + TF_AddGradients(graph_, outputs, 1, inputs, 1, nullptr, s_, &dxy_dy); + ASSERT_EQ(TF_OK, TF_GetCode(s_)) << TF_Message(s_); + + TF_SessionOptions* opts = TF_NewSessionOptions(); + TF_Session* sess = TF_NewSession(graph_, opts, s_); + TF_DeleteSessionOptions(opts); + ASSERT_EQ(TF_OK, TF_GetCode(s_)) << TF_Message(s_); + + TF_Output feeds[] = {{x, 0}, {y, 0}}; + TF_Tensor* feedValues[] = {FloatTensor(X), FloatTensor(Y)}; + TF_Output fetches[] = {dxy_dx, dxy_dy}; + TF_Tensor* fetchValues[] = {nullptr, nullptr}; + + TF_SessionRun(sess, nullptr /* run_options */, feeds, feedValues, 2, fetches, + fetchValues, 2, nullptr /* target_opers */, 0, + nullptr /* run_metadata */, s_); + TF_DeleteTensor(feedValues[0]); + TF_DeleteTensor(feedValues[1]); + ASSERT_EQ(TF_OK, TF_GetCode(s_)) << TF_Message(s_); + TF_DeleteSession(sess, s_); + ASSERT_EQ(TF_OK, TF_GetCode(s_)) << TF_Message(s_); + + float dxy_dxValue = 0.0f, dxy_dyValue = 0.0f; + ScalarFloatFromTensor(fetchValues[0], &dxy_dxValue); + EXPECT_EQ(Y, dxy_dxValue); + + ScalarFloatFromTensor(fetchValues[1], &dxy_dyValue); + EXPECT_EQ(X, dxy_dyValue); + + TF_DeleteTensor(fetchValues[0]); + TF_DeleteTensor(fetchValues[1]); +} + // REGISTER_OP for CApiAttributesTest test cases. // Registers two ops, each with a single attribute called 'v'. // The attribute in one op will have a type 'type', the other @@ -1784,7 +1839,7 @@ TEST_F(CApiAttributesTest, String) { TEST_F(CApiAttributesTest, StringList) { std::vector list = {"bugs", "bunny", "duck"}; - std::unique_ptr list_ptrs; + std::unique_ptr list_ptrs; std::unique_ptr list_lens; StringVectorToArrays(list, &list_ptrs, &list_lens); int list_total_size = 0; @@ -1800,7 +1855,7 @@ TEST_F(CApiAttributesTest, StringList) { ASSERT_EQ(TF_OK, TF_GetCode(s_)) << TF_Message(s_); EXPECT_TF_META("v", list.size(), TF_ATTR_STRING, list_total_size); - std::unique_ptr values(new void*[list.size()]); + std::unique_ptr values(new void*[list.size()]); std::unique_ptr lens(new size_t[list.size()]); std::unique_ptr storage(new char[list_total_size]); TF_OperationGetAttrStringList(oper, "v", values.get(), lens.get(), @@ -2025,7 +2080,7 @@ TEST_F(CApiAttributesTest, TensorShapeProtoList) { tensorflow::PartialTensorShape(pts2).AsProto(&proto); proto.SerializeToString(&bytes2); - std::unique_ptr list_ptrs; + std::unique_ptr list_ptrs; std::unique_ptr list_lens; const std::vector list = {bytes1, bytes2}; StringVectorToArrays(list, &list_ptrs, &list_lens); diff --git a/tensorflow/c/c_test_util.cc b/tensorflow/c/c_test_util.cc index f3b28c1708129d39e451d927a89c0d10e2193b63..24eb6c069b21349fce288db3e79fbf14e824ad11 100644 --- a/tensorflow/c/c_test_util.cc +++ b/tensorflow/c/c_test_util.cc @@ -216,6 +216,13 @@ TF_Operation* Min(TF_Operation* l, TF_Operation* r, TF_Graph* graph, return MinWithDevice(l, r, graph, /*op_device=*/"", s, name); } +TF_Operation* Mul(TF_Operation* l, TF_Operation* r, TF_Graph* graph, + TF_Status* s, const char* name) { + TF_Operation* op; + BinaryOpHelper("Mul", l, r, graph, s, name, &op, "", true); + return op; +} + TF_Operation* Add(TF_Output l, TF_Output r, TF_Graph* graph, TF_Status* s, const char* name) { TF_OperationDescription* desc = TF_NewOperation(graph, "AddN", name); diff --git a/tensorflow/c/c_test_util.h b/tensorflow/c/c_test_util.h index c16aba666ee6974fed5351c2d9ac291dcbcdecab..38313d647ca93d4779bb1325f8ed7bde4b743879 100644 --- a/tensorflow/c/c_test_util.h +++ b/tensorflow/c/c_test_util.h @@ -80,6 +80,9 @@ TF_Operation* Add(TF_Output l, TF_Output r, TF_Graph* graph, TF_Status* s, TF_Operation* Min(TF_Operation* l, TF_Operation* r, TF_Graph* graph, TF_Status* s, const char* name = "min"); +TF_Operation* Mul(TF_Operation* l, TF_Operation* r, TF_Graph* graph, + TF_Status* s, const char* name = "mul"); + // If `op_device` is non-empty, set the created op on that device. TF_Operation* MinWithDevice(TF_Operation* l, TF_Operation* r, TF_Graph* graph, const string& op_device, TF_Status* s, diff --git a/tensorflow/c/eager/c_api.cc b/tensorflow/c/eager/c_api.cc index 00b474fe8625db543e2f243fd29c99c646bc7c56..82ca2be2cff885967dd798a1cb84b164a9df399e 100644 --- a/tensorflow/c/eager/c_api.cc +++ b/tensorflow/c/eager/c_api.cc @@ -156,12 +156,14 @@ tensorflow::Status NewRemoteAwareTFE_Context(const TFE_ContextOptions* opts, // server object (which currently CHECK-fails) and we miss the error, instead, // we log the error, and then return to allow the user to see the error // message. -#define LOG_AND_RETURN_IF_ERROR(...) \ - do { \ - const ::tensorflow::Status _status = (__VA_ARGS__); \ - LOG(ERROR) << _status.error_message(); \ - if (TF_PREDICT_FALSE(!_status.ok())) return _status; \ - } while (0) +#define LOG_AND_RETURN_IF_ERROR(...) \ + do { \ + const ::tensorflow::Status _status = (__VA_ARGS__); \ + if (TF_PREDICT_FALSE(!_status.ok())) { \ + LOG(ERROR) << _status.error_message(); \ + return _status; \ + } \ + } while (0); string worker_name = tensorflow::strings::StrCat( "/job:", opts->server_def.job_name(), @@ -346,16 +348,16 @@ TF_DataType TFE_TensorHandleDataType(TFE_TensorHandle* h) { } int TFE_TensorHandleNumDims(TFE_TensorHandle* h, TF_Status* status) { - const tensorflow::Tensor* t = nullptr; - status->status = h->handle->Tensor(&t); - return t == nullptr ? 0 : t->dims(); + int result; + status->status = h->handle->NumDims(&result); + return result; } int64_t TFE_TensorHandleDim(TFE_TensorHandle* h, int dim_index, TF_Status* status) { - const tensorflow::Tensor* t = nullptr; - status->status = h->handle->Tensor(&t); - return t == nullptr ? 0 : t->dim_size(dim_index); + tensorflow::int64 result; + status->status = h->handle->Dim(dim_index, &result); + return result; } const char* TFE_TensorHandleDeviceName(TFE_TensorHandle* h, TF_Status* status) { diff --git a/tensorflow/c/eager/tape.h b/tensorflow/c/eager/tape.h index 734e712daa39c03f0177eb199b1acb1b19e5d845..1adb0458c35193117b5fa5cfe9ceffbaaf699af7 100644 --- a/tensorflow/c/eager/tape.h +++ b/tensorflow/c/eager/tape.h @@ -520,7 +520,12 @@ Status GradientTape::ComputeGradient( } } else { any_gradient_nonzero = true; - auto new_gradients = vspace.AggregateGradients(grad_it->second); + Gradient* new_gradients = nullptr; + if (grad_it->second.size() == 1) { + new_gradients = grad_it->second.at(0); + } else { + new_gradients = vspace.AggregateGradients(grad_it->second); + } if (sources_set.find(grad_it->first) == sources_set.end()) { gradients.erase(grad_it); } else { diff --git a/tensorflow/c/python_api.cc b/tensorflow/c/python_api.cc index e18fdf6c57bd3f432d8cb73536fb816df90b3963..8486b585c8587e18e8eea18a893fac0a40ff4a27 100644 --- a/tensorflow/c/python_api.cc +++ b/tensorflow/c/python_api.cc @@ -155,7 +155,7 @@ void SetResourceHandleShapeAndType(TF_Graph* graph, TF_Output output, tensorflow::shape_inference::ShapeHandle shape; status->status = ic->MakeShapeFromShapeProto(shape_and_type_proto.shape(), &shape); - if (status->status.ok()) return; + if (!status->status.ok()) return; shapes_and_types.emplace_back(shape, shape_and_type_proto.dtype()); } ic->set_output_handle_shapes_and_types(output.index, shapes_and_types); diff --git a/tensorflow/cc/framework/scope.cc b/tensorflow/cc/framework/scope.cc index 62a889181e787f2e181135ab0563c45e1bab8812..8c886f31711eb014fb9e9d600c9c78cf22073f71 100644 --- a/tensorflow/cc/framework/scope.cc +++ b/tensorflow/cc/framework/scope.cc @@ -37,6 +37,11 @@ Scope& Scope::operator=(const Scope& other) { return *this; } +namespace { +const char kScopeSeparator[] = "/"; +const char kSuffixSeparator[] = "_"; +} // namespace + Scope::Impl::Impl(Graph* graph, Status* status, NameMap* name_map, ShapeRefiner* refiner, bool disable_shape_inference) : graph_(graph), @@ -308,19 +313,23 @@ string Scope::Impl::GetUniqueName(const string& prefix, return prefix; } auto entry = name_map_->find(prefix); - string unique_name = prefix; if (entry == name_map_->end()) { name_map_->insert({prefix, 0}); - } else { - unique_name = strings::StrCat(unique_name, "_", ++entry->second); + return prefix; } + string unique_name; + do { + unique_name = strings::StrCat(prefix, kSuffixSeparator, ++entry->second); + } while (name_map_->find(unique_name) != name_map_->end()); + name_map_->insert({unique_name, 0}); return unique_name; } string Scope::Impl::GetNameForOp(const string& default_name) const { const string unique_name = GetUniqueName(default_name, true /* check_single_use */); - const string sep = name_.empty() || unique_name.empty() ? "" : "/"; + const string sep = + name_.empty() || unique_name.empty() ? "" : kScopeSeparator; return strings::StrCat(name_, sep, unique_name); } @@ -345,7 +354,8 @@ Scope Scope::NewSubScope(const string& child_scope_name) const { } const string unique_name = impl()->GetUniqueName(child_scope_name, false /* check_single_use */); - const string sep = impl()->name_.empty() || unique_name.empty() ? "" : "/"; + const string sep = + impl()->name_.empty() || unique_name.empty() ? "" : kScopeSeparator; return Scope(new Impl(*this, Impl::Tags::ScopeName(), strings::StrCat(impl()->name_, sep, unique_name), false /* copy_names */)); @@ -412,7 +422,7 @@ CompositeOpScopes Scope::GetCompositeOpScopes( if (!impl()->single_use_scope()) { Scope child = NewSubScope(impl()->op_name_.empty() ? composite_op_name : impl()->op_name_); - const string child_op_sep = impl()->name_.empty() ? "" : "_"; + const string child_op_sep = impl()->name_.empty() ? "" : kSuffixSeparator; const string child_name = strings::StrCat(impl()->name_, child_op_sep, child.impl()->name_); return {child, @@ -435,7 +445,13 @@ class InternalScope { static Scope NewScope(Graph* graph, Status* status, ShapeRefiner* refiner) { Scope::Impl::NameMap* name_map = new Scope::Impl::NameMap; for (const Node* node : graph->nodes()) { - (*name_map)[node->name()] = 0; + const string& name = node->name(); + (*name_map)[name] = 0; + // Add all name prefixes ('/' separated). + size_t idx = -1; + while ((idx = name.find(kScopeSeparator, idx + 1)) != string::npos) { + (*name_map)[name.substr(0, idx)] = 0; + } } // We provide null destructors for these shared ptrs (except for name_map) // since the caller owns them and doesn't want the scope to destroy them. diff --git a/tensorflow/cc/framework/scope_internal.h b/tensorflow/cc/framework/scope_internal.h index 8efcfed20d0b86d86d8c20a3d8630c7c6bc909c3..58adaef2e942a7fa6b0ce8d5534ac3e2fd380580 100644 --- a/tensorflow/cc/framework/scope_internal.h +++ b/tensorflow/cc/framework/scope_internal.h @@ -34,8 +34,7 @@ class Scope::Impl { // name that has not been used so far in a scope will get no suffix. Later // uses of the same name will get suffixes _1, _2, _3, etc. Multiple scopes // can share the same NameMap. For instance, a new scope created using - // WithControlDependencies() should would share the same NameMap with the - // parent. + // WithControlDependencies() would share the same NameMap with the parent. typedef std::unordered_map NameMap; Impl(const std::shared_ptr& graph, diff --git a/tensorflow/cc/framework/scope_test.cc b/tensorflow/cc/framework/scope_test.cc index 9eca9d3face34319413e1acbc2f5ac0b2ba85374..b40b345eb84237c34ea593021bea022ad28095f7 100644 --- a/tensorflow/cc/framework/scope_test.cc +++ b/tensorflow/cc/framework/scope_test.cc @@ -26,6 +26,16 @@ TEST(ScopeTest, BasicNames) { EXPECT_EQ(root.GetUniqueNameForOp("mul"), "mul"); } +TEST(ScopeTest, OpAndScopeNameCollision) { + Scope root = Scope::NewRootScope(); + EXPECT_EQ(root.GetUniqueNameForOp("foo"), "foo"); + EXPECT_EQ(root.GetUniqueNameForOp("foo"), "foo_1"); + EXPECT_EQ(root.GetUniqueNameForOp("foo_1"), "foo_1_1"); + EXPECT_EQ(root.GetUniqueNameForOp("foo_2"), "foo_2"); + EXPECT_EQ(root.GetUniqueNameForOp("foo"), "foo_3"); + EXPECT_EQ(root.GetUniqueNameForOp("foo_2"), "foo_2_1"); +} + TEST(ScopeTest, HierarchicalNames) { Scope root = Scope::NewRootScope(); Scope child = root.NewSubScope("child"); diff --git a/tensorflow/cc/gradients/array_grad.cc b/tensorflow/cc/gradients/array_grad.cc index ff348fadb24e29a83bd6c8853aa67931f6df4182..b353accddcb6db9a07c112de03ead2f02c4ee6a6 100644 --- a/tensorflow/cc/gradients/array_grad.cc +++ b/tensorflow/cc/gradients/array_grad.cc @@ -421,6 +421,58 @@ Status StridedSliceGradHelper(const Scope& scope, const Operation& op, } REGISTER_GRADIENT_OP("StridedSlice", StridedSliceGradHelper); +Status SliceGrad(const Scope& scope, const Operation& op, + const std::vector& grad_inputs, + std::vector* grad_outputs) { + // Propagate the incoming gradient along all the selected values, + // and zero everywhere else. Use the Pad operator for this. + // + // First create an Nx2 padding where N is the number of input + // dimensions. The first column is the number of prepended zeros + // for each dimension, and the second column is the number of + // appended zeros. + // + // The first column is just the begin vector. + // The second column is the shape of the input element-wise + // subtracted by begin+size + + // Running example: + // input.shape = [3, 5, 3] + // begin = [1, 2, 1], size = [1, 3, 2] + Input input = op.input(0); + Input begin = op.input(1); + // input_rank = 3 + auto input_rank = Rank(scope, input); + // slice_size = [1, 3, 2] + auto slice_size = Shape(scope, op.output(0)); + // padding_shape = [3, 1] + auto padding_shape = Stack(scope, {input_rank, 1}); + // before_padding = [[1] + // [2] + // [1]] + Input before_padding = Reshape(scope, begin, padding_shape); + // after_padding_sizes = shape(input) - slice_size - begin + // = [3, 5, 3] - [1, 3, 2] - [1, 2, 1] + // = [1, 0, 0] + auto after_padding_sizes = + Sub(scope, Sub(scope, Shape(scope, input), slice_size), begin); + // after_padding = [[1] + // [0] + // [0]] + Input after_padding = Reshape(scope, after_padding_sizes, padding_shape); + // paddings = [[1 1] + // [2 0] + // [1 0]] + auto paddings = + Concat(scope, {before_padding, after_padding}, Const(scope, 1)); + grad_outputs->push_back(Pad(scope, grad_inputs[0], paddings)); + // Nothing propagated for "begin" and "size" inputs + grad_outputs->push_back(NoGradient()); + grad_outputs->push_back(NoGradient()); + return scope.status(); +} +REGISTER_GRADIENT_OP("Slice", SliceGrad); + } // anonymous namespace } // namespace ops } // namespace tensorflow diff --git a/tensorflow/cc/gradients/array_grad_test.cc b/tensorflow/cc/gradients/array_grad_test.cc index de3bd0fc9e2493f8ff76163f5be6bd4327c58c5a..d09275b6487b4212aa35a0476002f2bb587fa210 100644 --- a/tensorflow/cc/gradients/array_grad_test.cc +++ b/tensorflow/cc/gradients/array_grad_test.cc @@ -378,5 +378,12 @@ TEST_F(ArrayGradTest, StridedSliceGrad) { RunTest(x, x_shape, y, {1, 2, 2, 2}); } +TEST_F(ArrayGradTest, SliceGrad) { + TensorShape x_shape({3, 5, 3}); + auto x = Placeholder(scope_, DT_FLOAT, Placeholder::Shape(x_shape)); + auto y = Slice(scope_, x, {1, 2, 1}, {1, 3, 2}); + RunTest(x, x_shape, y, {1, 3, 2}); +} + } // namespace } // namespace tensorflow diff --git a/tensorflow/compiler/jit/BUILD b/tensorflow/compiler/jit/BUILD index d976f8296c6dc26bd13833f67874849aba91ad65..c2245b8eae8fd27d96feaf58e26418b92e646910 100644 --- a/tensorflow/compiler/jit/BUILD +++ b/tensorflow/compiler/jit/BUILD @@ -176,9 +176,11 @@ cc_library( "//tensorflow/core/kernels:cast_op", "//tensorflow/core/kernels:constant_op", "//tensorflow/core/kernels:control_flow_ops", + "//tensorflow/core/kernels:fifo_queue", "//tensorflow/core/kernels:identity_n_op", "//tensorflow/core/kernels:identity_op", "//tensorflow/core/kernels:no_op", + "//tensorflow/core/kernels:queue_op", "//tensorflow/core/kernels:resource_variable_ops", "//tensorflow/core/kernels:sendrecv_ops", "//tensorflow/core/kernels:shape_ops", diff --git a/tensorflow/compiler/jit/encapsulate_subgraphs_pass.cc b/tensorflow/compiler/jit/encapsulate_subgraphs_pass.cc index e786d41887f1d539fe1ae122275d1c14c77309e8..9c424b201eb25342254b323459ac5fb4806ceb76 100644 --- a/tensorflow/compiler/jit/encapsulate_subgraphs_pass.cc +++ b/tensorflow/compiler/jit/encapsulate_subgraphs_pass.cc @@ -60,9 +60,9 @@ const char* const kXlaHostTransferSequencerAttr = namespace { -bool AreAllParentsConst(const Node& n, - const gtl::FlatSet& runtime_const_nodes) { - if (n.type_string() == "GuaranteeConst" || n.type_string() == "Const") { +bool AreAllParentsGuaranteedConst( + const Node& n, const gtl::FlatSet& runtime_const_nodes) { + if (n.type_string() == "GuaranteeConst") { // If the current node is itself a cast-to-const, no need // to look at the incoming edges. return true; @@ -93,7 +93,8 @@ void MarkGuaranteedConstants( ReverseDFSFrom(graph, srcs, /*enter=*/nullptr, /*leave=*/[&guaranteed_const_nodes](const Node* n) { // TODO(vinuraja): Doesn't work in the presence of loops. - if (AreAllParentsConst(*n, guaranteed_const_nodes)) { + if (AreAllParentsGuaranteedConst(*n, + guaranteed_const_nodes)) { guaranteed_const_nodes.insert(n); } }); @@ -1136,7 +1137,10 @@ Status Encapsulator::Subgraph::AddShapeInferenceInfo( GraphToFunctionDef(*inference_graph, inference_graph_name, &fdef)); host_compute->AddAttr("shape_inference_graph", inference_graph_name); host_compute->AddAttr("shapes", std::vector()); - TF_RETURN_IF_ERROR(library->AddFunctionDef(fdef)); + // TODO(sibyl-Aix6ihai): Understand why there are multiple calls to Encapsulator. + if (library->Find(inference_graph_name) == nullptr) { + TF_RETURN_IF_ERROR(library->AddFunctionDef(fdef)); + } } return Status::OK(); } diff --git a/tensorflow/compiler/jit/encapsulate_subgraphs_pass_test.cc b/tensorflow/compiler/jit/encapsulate_subgraphs_pass_test.cc index 4eb389e0c653f2d32c17f448687f865a44a11b96..c0543a00792235c5dd090e81930d8c219dc7f1a3 100644 --- a/tensorflow/compiler/jit/encapsulate_subgraphs_pass_test.cc +++ b/tensorflow/compiler/jit/encapsulate_subgraphs_pass_test.cc @@ -742,10 +742,13 @@ TEST(EncapsulateSubgraphsWithGuaranteeConstOpTest, Simple) { Scope root = Scope::NewRootScope().ExitOnError().WithDevice( "/job:localhost/replica:0/task:0/cpu:0"); auto x1 = ops::Placeholder(root.WithOpName("x1"), DT_FLOAT); - auto const_x2 = ops::Const(root.WithOpName("const_x2"), 10.0f); + auto x2 = ops::Placeholder(root.WithOpName("x2"), DT_FLOAT); + auto const_guarantee_x2 = + ops::GuaranteeConst(root.WithOpName("const_guarantee_x2"), x2); auto const_guarantee_x1 = ops::GuaranteeConst(root.WithOpName("const_guarantee_x1"), x1); - auto add1 = ops::Add(root.WithOpName("add1"), const_guarantee_x1, const_x2); + auto add1 = + ops::Add(root.WithOpName("add1"), const_guarantee_x1, const_guarantee_x2); add1.node()->AddAttr("_encapsulate", "encapsulate1"); Graph graph_before(OpRegistry::Global()); diff --git a/tensorflow/compiler/jit/kernels/xla_launch_op.cc b/tensorflow/compiler/jit/kernels/xla_launch_op.cc index 902fe27acdec1cb323217e6409fbd02f62177612..338fb5a6f06525667c4d77907f63c9786fcf2a90 100644 --- a/tensorflow/compiler/jit/kernels/xla_launch_op.cc +++ b/tensorflow/compiler/jit/kernels/xla_launch_op.cc @@ -115,6 +115,7 @@ void XlaLocalLaunchBase::Compute(OpKernelContext* ctx) { const XlaDevice::Metadata* metadata = nullptr; Status s = XlaDevice::GetMetadata(ctx, &metadata); bool allocate_xla_tensors = s.ok(); + bool use_multiple_streams = s.ok() && metadata->UseMultipleStreams(); // Get the platform_id_ for XLA_* devices. if (platform_id_ == nullptr) { @@ -166,14 +167,22 @@ void XlaLocalLaunchBase::Compute(OpKernelContext* ctx) { } XlaCompiler::CompileOptions compile_options; compile_options.is_entry_computation = true; + // Optimization: don't resolve constants. If we resolve constants we never + // emit them on the device, meaning that if they are needed by a following + // computation the host has to transfer them. + compile_options.resolve_compile_time_constants = false; + // Optimization: where possible, have the computation return a naked array + // rather than a one-element tuple. + compile_options.always_return_tuple = false; + OP_REQUIRES_OK( ctx, cache->Compile(options, function_, constant_args, variables, ctx, &kernel, &executable, &compile_options)); VLOG(1) << "Executing XLA Computation..."; - XlaComputationLaunchContext launch_context(client, xla_allocator, - allocate_xla_tensors); + XlaComputationLaunchContext launch_context( + client, xla_allocator, allocate_xla_tensors, use_multiple_streams); launch_context.PopulateInputs(ctx, kernel, variables); // Execute the computation. diff --git a/tensorflow/compiler/jit/xla_compile_on_demand_op.cc b/tensorflow/compiler/jit/xla_compile_on_demand_op.cc index 26f350855d947d1d46f0fdaa7de7fb81754a6780..d288d37bc75380168a31937024dd41bdbe7dce9d 100644 --- a/tensorflow/compiler/jit/xla_compile_on_demand_op.cc +++ b/tensorflow/compiler/jit/xla_compile_on_demand_op.cc @@ -53,7 +53,9 @@ Status XlaCompileOnDemandOp::Run(OpKernelContext* ctx, // Builds an XLA allocator for the device. XlaComputationLaunchContext launch_context( - client, client->backend().memory_allocator(), true); + client, client->backend().memory_allocator(), + /*allocate_xla_tensors=*/true, + /*use_multiple_streams=*/metadata.UseMultipleStreams()); launch_context.PopulateInputs(ctx, result, variables); @@ -163,6 +165,13 @@ Status XlaCompileOnDemandOp::Compile( XlaCompiler::CompileOptions compile_options; compile_options.is_entry_computation = true; + // Optimization: don't resolve constants. If we resolve constants we never + // emit them on the device, meaning that if they are needed by a following + // computation the host has to transfer them. + compile_options.resolve_compile_time_constants = false; + // Optimization: where possible, have the computation return a naked array + // rather than a one-element tuple. + compile_options.always_return_tuple = false; std::map variable_args = GetVariables(ctx); return cache->CompileSingleOp(options, constant_arguments, variable_args, ctx, diff --git a/tensorflow/compiler/jit/xla_cpu_device.cc b/tensorflow/compiler/jit/xla_cpu_device.cc index 43648402f65c656b6b4eb2e83e61ce45f1c73669..7e159e3171113b0d53f03bb676ac9c21db7fe77a 100644 --- a/tensorflow/compiler/jit/xla_cpu_device.cc +++ b/tensorflow/compiler/jit/xla_cpu_device.cc @@ -54,6 +54,7 @@ Status XlaCpuDeviceFactory::CreateDevices(const SessionOptions& options, DEVICE_CPU_XLA_JIT, options, name_prefix, registration, /*transfer_as_literal=*/false, + /*use_multiple_streams=*/false, /*shape_representation_fn=*/{}, /*padded_shape_fn=*/{}, &device)); devices->push_back(device.release()); diff --git a/tensorflow/compiler/jit/xla_device.cc b/tensorflow/compiler/jit/xla_device.cc index ed007d603ea1b3d27dd25f00726261cdd029c20c..c55eba2f79ddcf10931ea659a64df559cef06ec5 100644 --- a/tensorflow/compiler/jit/xla_device.cc +++ b/tensorflow/compiler/jit/xla_device.cc @@ -130,7 +130,7 @@ Status DefaultPaddedShapeFn(const Tensor& tensor, xla::Shape* shape) { const string& jit_device_name, const SessionOptions& options, const string& name_prefix, const XlaOpRegistry::DeviceRegistration& registration, - bool transfer_as_literal, + bool transfer_as_literal, bool use_multiple_streams, const XlaCompiler::ShapeRepresentationFn& shape_representation_fn, const PaddedShapeFn& padded_shape_fn, std::unique_ptr* device) { VLOG(1) << "XlaDevice::Create " << platform_name << " " << device_name << ":" @@ -151,22 +151,24 @@ Status DefaultPaddedShapeFn(const Tensor& tensor, xla::Shape* shape) { DeviceType(device_name), Bytes(16ULL << 30), DeviceLocality(), strings::StrCat("device: ", device_name, " device")); - device->reset(new XlaDevice( - options, attrs, device_ordinal, DeviceType(jit_device_name), - platform.ValueOrDie(), transfer_as_literal, shape_representation_fn, - padded_shape_fn ? padded_shape_fn : DefaultPaddedShapeFn)); + device->reset( + new XlaDevice(options, attrs, device_ordinal, DeviceType(jit_device_name), + platform.ValueOrDie(), transfer_as_literal, + use_multiple_streams, shape_representation_fn, + padded_shape_fn ? padded_shape_fn : DefaultPaddedShapeFn)); return Status::OK(); } XlaDevice::Metadata::Metadata( int device_ordinal, se::Platform* platform, const DeviceType& device_type, XlaCompiler::ShapeRepresentationFn shape_representation_fn, - PaddedShapeFn padded_shape_fn) + PaddedShapeFn padded_shape_fn, bool use_multiple_streams) : device_ordinal_(device_ordinal), device_type_(device_type), platform_(platform), shape_representation_fn_(std::move(shape_representation_fn)), - padded_shape_fn_(std::move(padded_shape_fn)) {} + padded_shape_fn_(std::move(padded_shape_fn)), + use_multiple_streams_(use_multiple_streams) {} int XlaDevice::Metadata::device_ordinal() const { return device_ordinal_; } @@ -200,16 +202,18 @@ const DeviceType& XlaDevice::Metadata::jit_device_type() const { XlaDevice::XlaDevice( const SessionOptions& options, const DeviceAttributes& attrs, int device_ordinal, const DeviceType& jit_device_name, - se::Platform* platform, bool transfer_as_literal, + se::Platform* platform, bool transfer_as_literal, bool use_multiple_streams, const XlaCompiler::ShapeRepresentationFn& shape_representation_fn, const PaddedShapeFn& padded_shape_fn) : LocalDevice(options, attrs), xla_metadata_(device_ordinal, platform, jit_device_name, - shape_representation_fn, padded_shape_fn), + shape_representation_fn, padded_shape_fn, + use_multiple_streams), device_ordinal_(device_ordinal), jit_device_name_(jit_device_name), xla_allocator_(nullptr), platform_(platform), + use_multiple_streams_(use_multiple_streams), transfer_as_literal_(transfer_as_literal), shape_representation_fn_(shape_representation_fn) { VLOG(1) << "Created XLA device " << jit_device_name; @@ -253,6 +257,30 @@ xla::StatusOr XlaDevice::GetStream() { return stream_.get(); } +xla::StatusOr XlaDevice::GetDeviceToHostStream() { + if (!use_multiple_streams_) { + return GetStream(); + } + if (!device_to_host_stream_) { + xla::Backend* backend = client()->mutable_backend(); + TF_ASSIGN_OR_RETURN(device_to_host_stream_, + backend->BorrowStream(device_ordinal_)); + } + return device_to_host_stream_.get(); +} + +xla::StatusOr XlaDevice::GetHostToDeviceStream() { + if (!use_multiple_streams_) { + return GetStream(); + } + if (!host_to_device_stream_) { + xla::Backend* backend = client()->mutable_backend(); + TF_ASSIGN_OR_RETURN(host_to_device_stream_, + backend->BorrowStream(device_ordinal_)); + } + return host_to_device_stream_.get(); +} + Status XlaDevice::CreateAndSetGpuDeviceInfo() { if (gpu_device_info_ == nullptr) { TF_ASSIGN_OR_RETURN(se::Stream * stream, GetStream()); @@ -263,8 +291,9 @@ Status XlaDevice::CreateAndSetGpuDeviceInfo() { // gpu_device_info_->default_context. gpu_device_info_ = MakeUnique(); gpu_device_info_->stream = stream; - gpu_device_info_->default_context = new XlaDeviceContext( - stream, client(), transfer_as_literal_, shape_representation_fn_); + gpu_device_info_->default_context = + new XlaDeviceContext(stream, stream, stream, client(), + transfer_as_literal_, shape_representation_fn_); set_tensorflow_gpu_device_info(gpu_device_info_.get()); } @@ -276,10 +305,16 @@ Status XlaDevice::FillContextMap(const Graph* graph, VLOG(1) << "XlaDevice::FillContextMap"; device_context_map->resize(graph->num_node_ids()); TF_ASSIGN_OR_RETURN(se::Stream * stream, GetStream()); + TF_ASSIGN_OR_RETURN(se::Stream * device_to_host_stream, + GetDeviceToHostStream()); + TF_ASSIGN_OR_RETURN(se::Stream * host_to_device_stream, + GetHostToDeviceStream()); + // Call GetAllocator for the side-effect of ensuring the allocator is created. GetAllocator({}); - auto ctx = new XlaDeviceContext(stream, client(), transfer_as_literal_, - shape_representation_fn_); + auto ctx = new XlaDeviceContext( + stream, host_to_device_stream, device_to_host_stream, client(), + transfer_as_literal_, shape_representation_fn_); for (Node* n : graph->nodes()) { VLOG(2) << n->id() << " : " << n->type_string() << " : " << n->name(); ctx->Ref(); @@ -326,8 +361,13 @@ Status XlaDevice::MakeTensorFromProto(const TensorProto& tensor_proto, Tensor copy(GetAllocator(alloc_attrs), parsed.dtype(), parsed.shape()); Notification n; TF_ASSIGN_OR_RETURN(se::Stream * stream, GetStream()); - XlaTransferManager manager(stream, client(), transfer_as_literal_, - shape_representation_fn_); + TF_ASSIGN_OR_RETURN(se::Stream * device_to_host_stream, + GetDeviceToHostStream()); + TF_ASSIGN_OR_RETURN(se::Stream * host_to_device_stream, + GetHostToDeviceStream()); + XlaTransferManager manager(stream, host_to_device_stream, + device_to_host_stream, client(), + transfer_as_literal_, shape_representation_fn_); manager.CopyCPUTensorToDevice(&parsed, this, ©, [&n, &status](const Status& s) { status = s; diff --git a/tensorflow/compiler/jit/xla_device.h b/tensorflow/compiler/jit/xla_device.h index 02e88ee6793e984a7b782790f8011cbcbc5a5026..fccdb143680353ccbe3106bd48aa297980179d55 100644 --- a/tensorflow/compiler/jit/xla_device.h +++ b/tensorflow/compiler/jit/xla_device.h @@ -57,7 +57,7 @@ class XlaDevice : public LocalDevice { Metadata(int device_ordinal, se::Platform* platform, const DeviceType& device_type, XlaCompiler::ShapeRepresentationFn shape_representation_fn, - PaddedShapeFn padded_shape_fn); + PaddedShapeFn padded_shape_fn, bool use_multiple_streams); // The index of the device on this host. int device_ordinal() const; @@ -70,12 +70,15 @@ class XlaDevice : public LocalDevice { } const PaddedShapeFn& padded_shape_fn() const { return padded_shape_fn_; } + bool UseMultipleStreams() const { return use_multiple_streams_; } + private: const int device_ordinal_; const DeviceType device_type_; se::Platform* platform_; // Not owned. XlaCompiler::ShapeRepresentationFn shape_representation_fn_; PaddedShapeFn padded_shape_fn_; + const bool use_multiple_streams_; TF_DISALLOW_COPY_AND_ASSIGN(Metadata); }; @@ -89,6 +92,8 @@ class XlaDevice : public LocalDevice { // 'transfer_as_literal' is true if device<->host transfers must be done using // XLA's TransferLiteral{To,From}Device interface. If false, we can use // ThenMemcpy instead. + // If 'use_multiple_streams' is true, we create separate streams for + // host-to-device and device-to-host communication. // If padded_shape_fn is empty, a default implementation that returns // the on-host shape is used. static Status Create( @@ -96,7 +101,7 @@ class XlaDevice : public LocalDevice { int device_ordinal, const string& jit_device_name, const SessionOptions& options, const string& name_prefix, const XlaOpRegistry::DeviceRegistration& registration, - bool transfer_as_literal, + bool transfer_as_literal, bool use_multiple_streams, const XlaCompiler::ShapeRepresentationFn& shape_representation_fn, const PaddedShapeFn& padded_shape_fn, std::unique_ptr* device); @@ -106,6 +111,7 @@ class XlaDevice : public LocalDevice { XlaDevice(const SessionOptions& options, const DeviceAttributes& attrs, int device_ordinal, const DeviceType& jit_device_name, se::Platform* platform, bool transfer_as_literal, + bool use_multiple_streams, const XlaCompiler::ShapeRepresentationFn& shape_representation_fn, const PaddedShapeFn& padded_shape_fn); ~XlaDevice() override; @@ -126,6 +132,8 @@ class XlaDevice : public LocalDevice { xla::LocalClient* client() const; const Metadata& metadata() { return xla_metadata_; } xla::StatusOr GetStream(); + xla::StatusOr GetHostToDeviceStream(); + xla::StatusOr GetDeviceToHostStream(); // If not already set, create and set GpuDeviceInfo. // Not thread-safe @@ -146,6 +154,16 @@ class XlaDevice : public LocalDevice { // copying back and forth between CPU and the device, and // computations enqueued by XLA. xla::Backend::StreamPtr stream_; + // If true, only stream_ is valid and all computation and transfers use + // stream_. If false, computation is performed by stream_ and transfers are + // performed by host_to_device/device_to_host_stream. + bool use_multiple_streams_; + // If use_multiple_streams_, host to device transfers are performed using this + // stream. + xla::Backend::StreamPtr host_to_device_stream_; + // If use_multiple_streams_, device to host transfers are performed using this + // stream. + xla::Backend::StreamPtr device_to_host_stream_; // Must we use XLA's transfer manager for correct host<->device transfers? if // false, we can use ThenMemcpy() instead. bool transfer_as_literal_; diff --git a/tensorflow/compiler/jit/xla_device_context.cc b/tensorflow/compiler/jit/xla_device_context.cc index 37005479dc7dd27cabc945f2753e20477a71549a..04778c00904d484aa55884b86f10759ea3c3df20 100644 --- a/tensorflow/compiler/jit/xla_device_context.cc +++ b/tensorflow/compiler/jit/xla_device_context.cc @@ -48,17 +48,24 @@ void XlaDeviceAllocator::DeallocateRaw(void* ptr) { void XlaDeviceAllocator::GetStats(AllocatorStats* stats) { stats->Clear(); } XlaTransferManager::XlaTransferManager( - se::Stream* stream, xla::LocalClient* client, bool transfer_as_literal, + se::Stream* compute_stream, se::Stream* host_to_device_stream, + se::Stream* device_to_host_stream, xla::LocalClient* client, + bool transfer_as_literal, XlaCompiler::ShapeRepresentationFn shape_representation_fn) - : stream_(stream), + : stream_(compute_stream), + host_to_device_stream_(host_to_device_stream), + device_to_host_stream_(device_to_host_stream), client_(client), transfer_manager_(client->backend().transfer_manager()), transfer_as_literal_(transfer_as_literal), shape_representation_fn_(std::move(shape_representation_fn)) { + CHECK(host_to_device_stream_ != nullptr); + CHECK(device_to_host_stream_ != nullptr); + CHECK(stream_ != nullptr); if (!shape_representation_fn_) { - shape_representation_fn_ = [](const TensorShape& shape, DataType dtype) { - return shape; - }; + shape_representation_fn_ = + [](const TensorShape& shape, + DataType dtype) -> xla::StatusOr { return shape; }; } } @@ -70,12 +77,19 @@ Status XlaTransferManager::TransferLiteralToDevice( xla::BorrowingLiteral literal( static_cast(DMAHelper::base(&host_tensor)), xla_shape); - const xla::ShapedBuffer& shaped_buffer = - XlaTensor::FromTensor(device_tensor)->shaped_buffer(); + XlaTensor* xla_tensor = XlaTensor::FromTensor(device_tensor); + const xla::ShapedBuffer& shaped_buffer = xla_tensor->shaped_buffer(); VLOG(1) << "Transfer to device as literal: " << literal.ToString() << " " << shaped_buffer.ToString(); - return transfer_manager_->TransferLiteralToDevice(stream_, literal, - shaped_buffer); + TF_RETURN_IF_ERROR(transfer_manager_->TransferLiteralToDevice( + host_to_device_stream_, literal, shaped_buffer)); + if (UseMultipleStreams()) { + se::Event event(stream_->parent()); + TF_RET_CHECK(event.Init()) << "Event failed to initialize!"; + host_to_device_stream_->ThenRecordEvent(&event); + xla_tensor->SetDefinedOn(host_to_device_stream_, std::move(event)); + } + return Status::OK(); } Status XlaTransferManager::TransferLiteralFromDevice( @@ -83,9 +97,9 @@ Status XlaTransferManager::TransferLiteralFromDevice( const xla::ShapedBuffer& shaped_buffer = XlaTensor::FromTensor(&device_tensor)->shaped_buffer(); - TF_ASSIGN_OR_RETURN( - std::unique_ptr literal, - transfer_manager_->TransferLiteralFromDevice(stream_, shaped_buffer)); + TF_ASSIGN_OR_RETURN(std::unique_ptr literal, + transfer_manager_->TransferLiteralFromDevice( + device_to_host_stream_, shaped_buffer)); VLOG(1) << "Transfer from device as literal: " << literal->ToString() << " " << shaped_buffer.ToString(); Tensor tensor; @@ -103,63 +117,67 @@ void XlaTransferManager::CopyCPUTensorToDevice(const Tensor* cpu_tensor, Device* device, Tensor* device_tensor, StatusCallback done) const { - if (cpu_tensor->NumElements() > 0) { - VLOG(2) << "CopyCPUTensorToDevice " - << reinterpret_cast(cpu_tensor->tensor_data().data()) - << " " - << reinterpret_cast( - device_tensor->tensor_data().data()) - << " " << cpu_tensor->NumElements() << " " - << cpu_tensor->shape().DebugString() << " " - << device_tensor->shape().DebugString(); - - void* src_ptr = const_cast(DMAHelper::base(cpu_tensor)); - const int64 total_bytes = cpu_tensor->TotalBytes(); - - XlaTensor* xla_tensor = XlaTensor::FromTensor(device_tensor); - CHECK(xla_tensor); - - TensorShape shape = shape_representation_fn_(device_tensor->shape(), - device_tensor->dtype()); - if (!xla_tensor->has_shaped_buffer()) { - Status s = xla_tensor->AllocateShapedBuffer( - device_tensor->dtype(), shape, client_, - stream_->parent()->device_ordinal()); - if (!s.ok()) { - done(s); - return; - } - } + if (cpu_tensor->NumElements() == 0) { + VLOG(2) << "CopyCPUTensorToDevice empty tensor"; + done(Status::OK()); + return; + } - Status status; - if (transfer_as_literal_) { - Tensor reshaped_cpu_tensor; - if (!reshaped_cpu_tensor.CopyFrom(*cpu_tensor, shape)) { - done(errors::Internal( - "Tensor::CopyFrom failed when copying from CPU to XLA device")); - return; - } - status = TransferLiteralToDevice(reshaped_cpu_tensor, device_tensor); - } else { - se::DeviceMemoryBase dev_dst_ptr = - XlaTensor::DeviceMemoryFromTensor(*device_tensor); - stream_->ThenMemcpy(&dev_dst_ptr, src_ptr, total_bytes); - // TODO(hpucha): Make this asynchronous. - Status block_status = stream_->BlockHostUntilDone(); - if (!block_status.ok()) { - status = xla::InternalError( - "Failed to complete data transfer on stream %p: %s", stream_, - block_status.error_message().c_str()); - } - } - xla_tensor->set_host_tensor(*cpu_tensor); + VLOG(2) << "CopyCPUTensorToDevice " + << reinterpret_cast(cpu_tensor->tensor_data().data()) + << " " + << reinterpret_cast(device_tensor->tensor_data().data()) + << " " << cpu_tensor->NumElements() << " " + << cpu_tensor->shape().DebugString() << " " + << device_tensor->shape().DebugString(); - done(status); + void* src_ptr = const_cast(DMAHelper::base(cpu_tensor)); + const int64 total_bytes = cpu_tensor->TotalBytes(); + + XlaTensor* xla_tensor = XlaTensor::FromTensor(device_tensor); + CHECK(xla_tensor); + + xla::StatusOr shape_or_status = + shape_representation_fn_(device_tensor->shape(), device_tensor->dtype()); + if (!shape_or_status.ok()) { + done(shape_or_status.status()); return; } + TensorShape shape = shape_or_status.ValueOrDie(); + if (!xla_tensor->has_shaped_buffer()) { + Status s = + xla_tensor->AllocateShapedBuffer(device_tensor->dtype(), shape, client_, + stream_->parent()->device_ordinal()); + if (!s.ok()) { + done(s); + return; + } + } - VLOG(2) << "CopyCPUTensorToDevice empty tensor"; - done(Status::OK()); + Status status; + if (transfer_as_literal_) { + Tensor reshaped_cpu_tensor; + if (!reshaped_cpu_tensor.CopyFrom(*cpu_tensor, shape)) { + done(errors::Internal( + "Tensor::CopyFrom failed when copying from CPU to XLA device")); + return; + } + status = TransferLiteralToDevice(reshaped_cpu_tensor, device_tensor); + } else { + se::DeviceMemoryBase dev_dst_ptr = + XlaTensor::DeviceMemoryFromTensor(*device_tensor); + host_to_device_stream_->ThenMemcpy(&dev_dst_ptr, src_ptr, total_bytes); + // TODO(hpucha): Make this asynchronous. + Status block_status = host_to_device_stream_->BlockHostUntilDone(); + if (!block_status.ok()) { + status = xla::InternalError( + "Failed to complete data transfer on stream %p: %s", + host_to_device_stream_, block_status.error_message().c_str()); + } + } + xla_tensor->set_host_tensor(*cpu_tensor); + + done(status); } void XlaTransferManager::CopyDeviceTensorToCPU(const Tensor* device_tensor, @@ -167,62 +185,83 @@ void XlaTransferManager::CopyDeviceTensorToCPU(const Tensor* device_tensor, Device* device, Tensor* cpu_tensor, StatusCallback done) { - if (device_tensor->NumElements() > 0) { - VLOG(2) << "CopyDeviceTensorToCPU " - << reinterpret_cast( - device_tensor->tensor_data().data()) - << " " - << reinterpret_cast(cpu_tensor->tensor_data().data()) - << " " << device_tensor->NumElements() << " " - << cpu_tensor->shape().DebugString() << " " - << device_tensor->shape().DebugString(); - - const int64 total_bytes = cpu_tensor->TotalBytes(); - se::DeviceMemoryBase dev_src_ptr = - XlaTensor::DeviceMemoryFromTensor(*device_tensor); - void* dst_ptr = DMAHelper::base(cpu_tensor); - - Status status; - if (transfer_as_literal_) { - status = TransferLiteralFromDevice(cpu_tensor, *device_tensor); - } else { - stream_->ThenMemcpy(dst_ptr, dev_src_ptr, total_bytes); - // TODO(hpucha): Make this asynchronous. - Status block_status = stream_->BlockHostUntilDone(); - if (!block_status.ok()) { - status = xla::InternalError( - "Failed to complete data transfer on stream %p: %s", stream_, - block_status.error_message().c_str()); - } - } - - done(status); + if (device_tensor->NumElements() == 0) { + VLOG(2) << "CopyDeviceTensorToCPU empty tensor"; + done(Status::OK()); return; } + VLOG(2) << "CopyDeviceTensorToCPU " + << reinterpret_cast(device_tensor->tensor_data().data()) + << " " + << reinterpret_cast(cpu_tensor->tensor_data().data()) + << " " << device_tensor->NumElements() << " " + << cpu_tensor->shape().DebugString() << " " + << device_tensor->shape().DebugString(); + + const int64 total_bytes = cpu_tensor->TotalBytes(); + se::DeviceMemoryBase dev_src_ptr = + XlaTensor::DeviceMemoryFromTensor(*device_tensor); + void* dst_ptr = DMAHelper::base(cpu_tensor); + XlaTensor* xla_tensor = XlaTensor::FromTensor(device_tensor); + + if (se::Event* event = + xla_tensor->GetDefinitionEvent(device_to_host_stream_)) { + device_to_host_stream_->ThenWaitFor(event); + xla_tensor->SetDefinedOn(device_to_host_stream_); + } + + Status status; + if (transfer_as_literal_) { + status = TransferLiteralFromDevice(cpu_tensor, *device_tensor); + } else { + device_to_host_stream_->ThenMemcpy(dst_ptr, dev_src_ptr, total_bytes); + // TODO(hpucha): Make this asynchronous. + Status block_status = device_to_host_stream_->BlockHostUntilDone(); + if (!block_status.ok()) { + status = xla::InternalError( + "Failed to complete data transfer on stream %p: %s", stream_, + block_status.error_message().c_str()); + } + } - VLOG(2) << "CopyDeviceTensorToCPU empty tensor"; - done(Status::OK()); + done(status); } void XlaTransferManager::CopyDeviceTensorToDevice(const Tensor& src_tensor, Tensor* dst_tensor, const StatusCallback& done) { + VLOG(2) << "CopyDeviceTensorToDevice " + << reinterpret_cast(src_tensor.tensor_data().data()) + << " " + << reinterpret_cast(dst_tensor->tensor_data().data()); // TODO(phawkins): replace this code with an asynchronous implementation. auto body = [&]() { if (src_tensor.NumElements() == 0) { return Status::OK(); } + // TODO(jmolloy): We co-opt the device_to_host stream for device to device + // transfers; perhaps we should have a dedicated device to device stream? or + // one per device? + auto device_to_device_stream = device_to_host_stream_; XlaTensor* xla_src = XlaTensor::FromTensor(&src_tensor); XlaTensor* xla_dst = XlaTensor::FromTensor(dst_tensor); CHECK(xla_src && xla_dst) << "Missing destination tensor for device-to-device copy"; if (!xla_dst->has_shaped_buffer()) { - TensorShape shape = - shape_representation_fn_(src_tensor.shape(), src_tensor.dtype()); + TF_ASSIGN_OR_RETURN( + TensorShape shape, + shape_representation_fn_(src_tensor.shape(), src_tensor.dtype())); TF_RETURN_IF_ERROR( xla_dst->AllocateShapedBuffer(src_tensor.dtype(), shape, client_, stream_->parent()->device_ordinal())); } + + if (se::Event* event = + xla_src->GetDefinitionEvent(device_to_device_stream)) { + device_to_device_stream->ThenWaitFor(event); + xla_src->SetDefinedOn(device_to_device_stream); + TF_RETURN_IF_ERROR(device_to_device_stream->BlockHostUntilDone()); + } TF_RETURN_IF_ERROR( xla_dst->shaped_buffer().buffers().ForEachMutableElementWithStatus( [&](const xla::ShapeIndex& index, se::DeviceMemoryBase* buffer) { @@ -241,9 +280,12 @@ void XlaTransferManager::CopyDeviceTensorToDevice(const Tensor& src_tensor, } XlaDeviceContext::XlaDeviceContext( - se::Stream* stream, xla::LocalClient* client, bool transfer_as_literal, + se::Stream* compute_stream, se::Stream* host_to_device_stream, + se::Stream* device_to_host_stream, xla::LocalClient* client, + bool transfer_as_literal, XlaCompiler::ShapeRepresentationFn shape_representation_fn) - : manager_(stream, client, transfer_as_literal, + : manager_(compute_stream, host_to_device_stream, device_to_host_stream, + client, transfer_as_literal, std::move(shape_representation_fn)) {} void XlaDeviceContext::CopyCPUTensorToDevice(const Tensor* cpu_tensor, diff --git a/tensorflow/compiler/jit/xla_device_context.h b/tensorflow/compiler/jit/xla_device_context.h index ee346e5653bbf9f393df202572c2150b4989506f..c726495f96883138892655797ab21257623daf31 100644 --- a/tensorflow/compiler/jit/xla_device_context.h +++ b/tensorflow/compiler/jit/xla_device_context.h @@ -47,7 +47,9 @@ class XlaDeviceAllocator : public Allocator { class XlaTransferManager { public: explicit XlaTransferManager( - se::Stream* stream, xla::LocalClient* client, bool transfer_as_literal, + se::Stream* compute_stream, se::Stream* host_to_device_stream, + se::Stream* device_to_host_stream, xla::LocalClient* client, + bool transfer_as_literal, XlaCompiler::ShapeRepresentationFn shape_representation_fn); void CopyCPUTensorToDevice(const Tensor* cpu_tensor, Device* device, @@ -66,10 +68,17 @@ class XlaTransferManager { Tensor* device_tensor) const; Status TransferLiteralFromDevice(Tensor* host_tensor, const Tensor& device_tensor) const; + bool UseMultipleStreams() const { return stream_ != host_to_device_stream_; } - // Stream obtained from a Device, used to transfer tensors between - // CPU and device. + // The main compute stream of the device, used to synchronize the transfer + // streams if they are set. se::Stream* stream_; + // The stream to use for transferring data from host to device. Can be + // idential to stream_, but must not be nullptr. + se::Stream* host_to_device_stream_; + // The stream to use for transferring data from device to host. Can be + // idential to stream_, but must not be nullptr. + se::Stream* device_to_host_stream_; // For the underlying memory allocator and XLA's TransferManager. xla::LocalClient* client_; // Transfer manager, for marshalling data to and from the device. @@ -85,7 +94,9 @@ class XlaTransferManager { class XlaDeviceContext : public DeviceContext { public: explicit XlaDeviceContext( - se::Stream* stream, xla::LocalClient* client, bool transfer_as_literal, + se::Stream* compute_stream, se::Stream* host_to_device_stream, + se::Stream* device_to_host_stream, xla::LocalClient* client, + bool transfer_as_literal, XlaCompiler::ShapeRepresentationFn shape_representation_fn); void CopyCPUTensorToDevice(const Tensor* cpu_tensor, Device* device, diff --git a/tensorflow/compiler/jit/xla_device_ops.h b/tensorflow/compiler/jit/xla_device_ops.h index 11e45d2823da2b623bd3cd45f7147686b05fdb2f..134dcc1bb55b3999724d593bd57e1034c31606d7 100644 --- a/tensorflow/compiler/jit/xla_device_ops.h +++ b/tensorflow/compiler/jit/xla_device_ops.h @@ -23,9 +23,11 @@ limitations under the License. #include "tensorflow/core/kernels/cast_op.h" #include "tensorflow/core/kernels/constant_op.h" #include "tensorflow/core/kernels/control_flow_ops.h" +#include "tensorflow/core/kernels/fifo_queue.h" #include "tensorflow/core/kernels/identity_n_op.h" #include "tensorflow/core/kernels/identity_op.h" #include "tensorflow/core/kernels/no_op.h" +#include "tensorflow/core/kernels/queue_op.h" #include "tensorflow/core/kernels/resource_variable_ops.h" #include "tensorflow/core/kernels/sendrecv_ops.h" #include "tensorflow/core/kernels/shape_ops.h" @@ -88,6 +90,9 @@ class XlaAssignVariableOp : public AsyncOpKernel { REGISTER_KERNEL_BUILDER( \ Name("ReadVariableOp").Device(DEVICE).HostMemory("resource"), \ ReadVariableOp); \ + REGISTER_KERNEL_BUILDER( \ + Name("DestroyResourceOp").Device(DEVICE).HostMemory("resource"), \ + DestroyResourceOp); \ REGISTER_KERNEL_BUILDER(Name("Shape") \ .Device(DEVICE) \ .HostMemory("output") \ @@ -145,7 +150,32 @@ class XlaAssignVariableOp : public AsyncOpKernel { .Device(DEVICE) \ .HostMemory("input") \ .HostMemory("output"), \ - LoopCondOp); + LoopCondOp); \ + \ + REGISTER_KERNEL_BUILDER( \ + Name("QueueEnqueueV2").Device(DEVICE).HostMemory("handle"), EnqueueOp); \ + REGISTER_KERNEL_BUILDER( \ + Name("QueueDequeueV2").Device(DEVICE).HostMemory("handle"), DequeueOp); \ + REGISTER_KERNEL_BUILDER( \ + Name("QueueCloseV2").Device(DEVICE).HostMemory("handle"), QueueCloseOp); \ + REGISTER_KERNEL_BUILDER(Name("QueueSizeV2") \ + .Device(DEVICE) \ + .HostMemory("size") \ + .HostMemory("handle"), \ + QueueSizeOp); \ + REGISTER_KERNEL_BUILDER( \ + Name("QueueIsClosedV2").Device(DEVICE).HostMemory("handle"), \ + QueueIsClosedOp); \ + \ + REGISTER_KERNEL_BUILDER( \ + Name("FIFOQueueV2").Device(DEVICE).HostMemory("handle"), FIFOQueueOp); + +// TODO(phawkins): currently we do not register the QueueEnqueueMany, +// QueueDequeueMany, or QueueDequeueUpTo kernels because they attempt to read +// and write the tensors they access in order to concatenate them into a batch. +// We would need either to call out to an XLA computation to perform the +// concatenation, or we would need to refactor those kernels so the splitting +// or merging is done in a separate operator that can be compiled. } // namespace tensorflow diff --git a/tensorflow/compiler/jit/xla_gpu_device.cc b/tensorflow/compiler/jit/xla_gpu_device.cc index c0d86a28c7698c302e28bab972bb2f847cc00ca4..851b118b0c18cfd752302b8f8dec27dae3e12acd 100644 --- a/tensorflow/compiler/jit/xla_gpu_device.cc +++ b/tensorflow/compiler/jit/xla_gpu_device.cc @@ -49,6 +49,7 @@ Status XlaGpuDeviceFactory::CreateDevices(const SessionOptions& options, XlaDevice::Create("CUDA", DEVICE_XLA_GPU, 0, DEVICE_GPU_XLA_JIT, options, name_prefix, registration, /*transfer_as_literal=*/false, + /*use_multiple_streams=*/false, /*shape_representation_fn=*/{}, /*padded_shape_fn=*/{}, &device); if (!status.ok()) { diff --git a/tensorflow/compiler/jit/xla_interpreter_device.cc b/tensorflow/compiler/jit/xla_interpreter_device.cc index 661187f4a873b03b8d013aa74cb6b6315bb4e2eb..45745596749207189c60ee1e3dcf19b6ecb7eb5b 100644 --- a/tensorflow/compiler/jit/xla_interpreter_device.cc +++ b/tensorflow/compiler/jit/xla_interpreter_device.cc @@ -52,6 +52,7 @@ Status XlaInterpreterDeviceFactory::CreateDevices( DEVICE_INTERPRETER_XLA_JIT, options, name_prefix, registration, /*transfer_as_literal=*/false, + /*use_multiple_streams=*/false, /*shape_representation_fn=*/{}, /*padded_shape_fn=*/{}, &device)); devices->push_back(device.release()); diff --git a/tensorflow/compiler/jit/xla_launch_util.cc b/tensorflow/compiler/jit/xla_launch_util.cc index d0c7a9365125708b2af43f87c7617d8d84050a61..616c3ed2a26fd30478fcee0d9b4d52a58cdb0921 100644 --- a/tensorflow/compiler/jit/xla_launch_util.cc +++ b/tensorflow/compiler/jit/xla_launch_util.cc @@ -115,14 +115,22 @@ using internal::ExtractSubShapedBuffer; XlaComputationLaunchContext::XlaComputationLaunchContext( xla::LocalClient* client, xla::DeviceMemoryAllocator* xla_allocator, - bool allocate_xla_tensors) + bool allocate_xla_tensors, bool use_multiple_streams) : client_(client), xla_allocator_(xla_allocator), - allocate_xla_tensors_(allocate_xla_tensors) {} + allocate_xla_tensors_(allocate_xla_tensors), + use_multiple_streams_(use_multiple_streams) { + if (use_multiple_streams_) { + CHECK(allocate_xla_tensors_) << "To use multiple streams correctly we must " + "be allocating XLA tensors!"; + } +} void XlaComputationLaunchContext::PopulateInputs( OpKernelContext* ctx, const XlaCompiler::CompilationResult* kernel, const std::map& variables) { + se::Stream* stream = + ctx->op_device_context() ? ctx->op_device_context()->stream() : nullptr; // Build ShapedBuffers that point directly to the Tensor buffers. arg_buffers_.reserve(kernel->xla_input_shapes.size() + 1); arg_buffers_.resize(kernel->xla_input_shapes.size()); @@ -140,6 +148,16 @@ void XlaComputationLaunchContext::PopulateInputs( t = &(ctx->input(arg_num)); } + if (use_multiple_streams_) { + CHECK(stream) << "Must have a stream available when using XLA tensors!"; + XlaTensor* xla_tensor = XlaTensor::FromTensor(t); + CHECK(xla_tensor); + if (se::Event* event = xla_tensor->GetDefinitionEvent(stream)) { + stream->ThenWaitFor(event); + xla_tensor->SetDefinedOn(stream); + } + } + const xla::Shape on_device_shape = client_->backend().transfer_manager()->HostShapeToDeviceShape(shape); if (xla::ShapeUtil::IsTuple(on_device_shape)) { @@ -176,6 +194,21 @@ void XlaComputationLaunchContext::PopulateOutputs( } CHECK_EQ(ctx->num_outputs(), kernel->outputs.size()); + // If the on-host-shape isn't a tuple, create a new single-element tuple + // buffer with a nullptr root index table. This allows the code below to treat + // output as a tuple unconditionally. + if (!xla::ShapeUtil::IsTuple(output.on_host_shape())) { + ShapedBuffer nontuple_buffer = output.release(); + ShapedBuffer buffer( + xla::ShapeUtil::MakeTupleShape({nontuple_buffer.on_host_shape()}), + xla::ShapeUtil::MakeTupleShape({nontuple_buffer.on_device_shape()}), + output.platform(), output.device_ordinal()); + buffer.buffers().CopySubtreeFrom(nontuple_buffer.buffers(), + /*source_base_index=*/{}, + /*target_base_index=*/{0}); + output = ScopedShapedBuffer(std::move(buffer), output.memory_allocator()); + } + // Copy XLA results to the OpOutputList. int output_num = 0; for (int i = 0; i < ctx->num_outputs(); ++i) { @@ -230,9 +263,20 @@ void XlaComputationLaunchContext::PopulateOutputs( Tensor* output_tensor; OP_REQUIRES_OK(ctx, ctx->allocate_output(i, shape, &output_tensor)); XlaTensor* xla_tensor = XlaTensor::FromTensor(output_tensor); - CHECK(xla_tensor); - xla_tensor->set_shaped_buffer(ScopedShapedBuffer( - ExtractSubShapedBuffer(&output, output_num, xla_allocator_))); + if (xla_tensor) { + xla_tensor->set_shaped_buffer(ScopedShapedBuffer( + ExtractSubShapedBuffer(&output, output_num, xla_allocator_))); + if (use_multiple_streams_) { + se::Event event(stream->parent()); + CHECK(event.Init()); + stream->ThenRecordEvent(&event); + xla_tensor->SetDefinedOn(stream, std::move(event)); + } + } else { + // xla_tensor wasn't valid, which must mean this is a zero-element + // tensor. + CHECK_EQ(output_tensor->TotalBytes(), 0); + } } else { Tensor output_tensor = XlaTensorBuffer::MakeTensor( ctx->expected_output_dtype(i), shape, buffer, allocator); @@ -282,6 +326,12 @@ void XlaComputationLaunchContext::PopulateOutputs( CHECK(xla_tensor); xla_tensor->set_shaped_buffer( ExtractSubShapedBuffer(&output, output_num, xla_allocator_)); + if (use_multiple_streams_) { + se::Event event(stream->parent()); + CHECK(event.Init()); + stream->ThenRecordEvent(&event); + xla_tensor->SetDefinedOn(stream, std::move(event)); + } *variable->tensor() = output_tensor; } else { Tensor output_tensor = XlaTensorBuffer::MakeTensor( diff --git a/tensorflow/compiler/jit/xla_launch_util.h b/tensorflow/compiler/jit/xla_launch_util.h index 4390701ccbd0bc3971413ddcd917c11019990087..90531174ff149bc6144fdba9b6463b8ef5b885f6 100644 --- a/tensorflow/compiler/jit/xla_launch_util.h +++ b/tensorflow/compiler/jit/xla_launch_util.h @@ -76,9 +76,15 @@ class XlaComputationLaunchContext { // Create a new launch context. 'allocate_xla_tensors' is true if allocated // output tensors and variables are always XlaTensors. If false they are // assumed to be "normal" device pointers. + // If 'use_multiple_streams' is true, tensors may be defined and used on + // multiple streams and so se::Events must be defined and waited for. If + // 'use_multiple_streams' is true, 'allocate_xla_tensors' must also be true + // because we track inter-stream dependencies through events inside XlaTensor + // objects. XlaComputationLaunchContext(xla::LocalClient* client, xla::DeviceMemoryAllocator* xla_allocator, - bool allocate_xla_tensors); + bool allocate_xla_tensors, + bool use_multiple_streams); // Add all inputs within `ctx` as XLA arguments (returned by arguments()). // `variables` is a map from TensorFlow argument number to resource variable. @@ -99,6 +105,7 @@ class XlaComputationLaunchContext { xla::LocalClient* client_; xla::DeviceMemoryAllocator* xla_allocator_; bool allocate_xla_tensors_; + bool use_multiple_streams_; std::vector> arg_buffers_; std::vector arg_ptrs_; }; diff --git a/tensorflow/compiler/jit/xla_tensor.cc b/tensorflow/compiler/jit/xla_tensor.cc index 3c44c4ae6df7f3e2d60d8933561c0c71888e8c3f..5dff187fffd4065559e98653b3e3c160d9ea4e8b 100644 --- a/tensorflow/compiler/jit/xla_tensor.cc +++ b/tensorflow/compiler/jit/xla_tensor.cc @@ -73,6 +73,36 @@ Status XlaTensor::AllocateShapedBuffer(DataType dtype, const TensorShape& shape, return Status::OK(); } +se::Event* XlaTensor::GetDefinitionEvent(se::Stream* stream) { + mutex_lock lock(mu_); + if (!definition_event_.has_value()) { + return nullptr; + } + + // The set of defined streams is expected to be very small indeed (usually + // 1-2), so a simple linear scan should be fast enough. + if (std::find(streams_defined_on_.begin(), streams_defined_on_.end(), + stream) != streams_defined_on_.end()) { + // stream is in streams_defined_on_; it doesn't need to be waited on. + return nullptr; + } + + return &*definition_event_; +} + +void XlaTensor::SetDefinedOn(se::Stream* stream, se::Event event) { + mutex_lock lock(mu_); + CHECK(!definition_event_.has_value()) + << "SetDefinedOn must only be called once!"; + definition_event_ = std::move(event); + streams_defined_on_.push_back(stream); +} + +void XlaTensor::SetDefinedOn(se::Stream* stream) { + mutex_lock lock(mu_); + streams_defined_on_.push_back(stream); +} + // The pointer tag, OR-ed into the XlaTensor's address to distinguish it from // device-side tensors, which are either CPU or GPU memory pointers. This works // because we're guaranteed that CPU and GPU pointers are aligned to > 1 bits. diff --git a/tensorflow/compiler/jit/xla_tensor.h b/tensorflow/compiler/jit/xla_tensor.h index c54001a999998f45c0cdacd752ca4036f0792857..f7e401c731163200c518074f2caa6907efb1f684 100644 --- a/tensorflow/compiler/jit/xla_tensor.h +++ b/tensorflow/compiler/jit/xla_tensor.h @@ -85,6 +85,24 @@ class XlaTensor { host_tensor_.reset(new Tensor(tensor)); } + // If the tensor's content is not yet defined on 'stream', and there exists an + // se::Event declaring when the tensor's content is defined, return it. + // Otherwise, return nullptr. If this function returns nullptr then the + // tensor's content can be read on 'stream' without additional + // synchronization. + se::Event* GetDefinitionEvent(se::Stream* stream); + + // Assert that the tensor's content is defined on 'stream' by the time 'event' + // triggers. + void SetDefinedOn(se::Stream* stream, se::Event event); + + // Assert that the tensor's content is defined on 'stream'. This version does + // not provide an event, and must be called *after* SetDefinedOn(Stream, + // Event). This call can be read as an assertion that the definition event has + // been waited on by 'stream', so further calls to GetDefinitionEvent(stream) + // do not need to also wait on the event. + void SetDefinedOn(se::Stream* stream); + // Convert from a raw pointer to an XlaTensor, removing the pointer tag. static XlaTensor* FromOpaquePointer(void* ptr); // Convert to a raw pointer from an XlaTensor, adding the pointer tag. @@ -95,6 +113,14 @@ class XlaTensor { std::unique_ptr shaped_buffer_; // An optional host tensor value. std::unique_ptr host_tensor_; + // An optional event that is triggered when the tensor's content has been + // defined. If this event is nullptr, it is assumed that the tensor's content + // is always defined. + gtl::optional definition_event_; + // A list of all streams for which the tensor's content is defined for any + // newly enqueued command. + gtl::InlinedVector streams_defined_on_ GUARDED_BY(mu_); + mutex mu_; }; } // namespace tensorflow diff --git a/tensorflow/compiler/tests/BUILD b/tensorflow/compiler/tests/BUILD index e72c409d65d69fd25758dcb69a7773fed9de6148..080bed50e68ba353a5029f5eb959003b51327f4a 100644 --- a/tensorflow/compiler/tests/BUILD +++ b/tensorflow/compiler/tests/BUILD @@ -70,6 +70,19 @@ py_test( ], ) +tf_xla_py_test( + name = "adadelta_test", + size = "medium", + srcs = ["adadelta_test.py"], + deps = [ + ":xla_test", + "//tensorflow/python:array_ops", + "//tensorflow/python:framework", + "//tensorflow/python:platform_test", + "//tensorflow/python:training", + ], +) + tf_xla_py_test( name = "adagrad_test", size = "small", @@ -84,6 +97,19 @@ tf_xla_py_test( ], ) +tf_xla_py_test( + name = "adagrad_da_test", + size = "small", + srcs = ["adagrad_da_test.py"], + deps = [ + ":xla_test", + "//tensorflow/python:array_ops", + "//tensorflow/python:framework", + "//tensorflow/python:platform_test", + "//tensorflow/python:training", + ], +) + tf_xla_py_test( name = "adam_test", size = "small", @@ -98,6 +124,48 @@ tf_xla_py_test( ], ) +tf_xla_py_test( + name = "adamax_test", + size = "small", + srcs = ["adamax_test.py"], + deps = [ + ":xla_test", + "//tensorflow/contrib/opt:opt_py", + "//tensorflow/python:array_ops", + "//tensorflow/python:client_testlib", + "//tensorflow/python:framework", + "//tensorflow/python:training", + ], +) + +tf_xla_py_test( + name = "addsign_test", + size = "small", + srcs = ["addsign_test.py"], + deps = [ + ":xla_test", + "//tensorflow/contrib/opt:opt_py", + "//tensorflow/python:array_ops", + "//tensorflow/python:client_testlib", + "//tensorflow/python:framework", + "//tensorflow/python:training", + ], +) + +tf_xla_py_test( + name = "powersign_test", + size = "small", + srcs = ["powersign_test.py"], + deps = [ + ":xla_test", + "//tensorflow/contrib/opt:opt_py", + "//tensorflow/python:array_ops", + "//tensorflow/python:client_testlib", + "//tensorflow/python:framework", + "//tensorflow/python:training", + ], +) + tf_xla_py_test( name = "argminmax_test", size = "small", @@ -167,7 +235,7 @@ tf_xla_py_test( tf_xla_py_test( name = "cholesky_op_test", - size = "small", + size = "medium", srcs = ["cholesky_op_test.py"], tags = ["optonly"], deps = [ @@ -350,7 +418,7 @@ tf_xla_py_test( tf_xla_py_test( name = "eager_test", - size = "small", + size = "large", srcs = ["eager_test.py"], disabled_backends = [ # TODO(b/78199195) Support XLA CPU devices in eager runtime @@ -371,6 +439,20 @@ tf_xla_py_test( ], ) +tf_xla_py_test( + name = "fifo_queue_test", + size = "medium", + srcs = ["fifo_queue_test.py"], + deps = [ + ":xla_test", + "//tensorflow/python:array_ops", + "//tensorflow/python:data_flow_ops", + "//tensorflow/python:extra_py_tests_deps", + "//tensorflow/python:framework", + "//tensorflow/python:platform_test", + ], +) + tf_xla_py_test( name = "fft_test", size = "medium", @@ -556,6 +638,53 @@ tf_xla_py_test( ], ) +tf_xla_py_test( + name = "proximal_adagrad_test", + size = "medium", + srcs = ["proximal_adagrad_test.py"], + deps = [ + ":xla_test", + "//tensorflow/python:array_ops", + "//tensorflow/python:client_testlib", + "//tensorflow/python:framework", + "//tensorflow/python:training", + ], +) + +tf_xla_py_test( + name = "proximal_gradient_descent_test", + size = "medium", + srcs = ["proximal_gradient_descent_test.py"], + deps = [ + ":xla_test", + "//tensorflow/python:array_ops", + "//tensorflow/python:client_testlib", + "//tensorflow/python:framework", + "//tensorflow/python:training", + ], +) + +tf_xla_py_test( + name = "qr_op_test", + size = "medium", + srcs = ["qr_op_test.py"], + disabled_backends = [ + # Test is very slow on CPU. + "cpu", + "cpu_ondemand", + ], + tags = ["optonly"], + deps = [ + ":xla_test", + "//tensorflow/python:array_ops", + "//tensorflow/python:framework_for_generated_wrappers", + "//tensorflow/python:math_ops", + "//tensorflow/python:platform_test", + "//tensorflow/python:training", + "@absl_py//absl/testing:parameterized", + ], +) + tf_xla_py_test( name = "random_ops_test", size = "small", @@ -688,6 +817,19 @@ tf_xla_py_test( ], ) +tf_xla_py_test( + name = "sparse_to_dense_op_test", + size = "small", + srcs = ["sparse_to_dense_op_test.py"], + deps = [ + ":xla_test", + "//tensorflow/python:array_ops", + "//tensorflow/python:framework", + "//tensorflow/python:platform_test", + "//tensorflow/python:sparse_ops", + ], +) + tf_xla_py_test( name = "stack_ops_test", size = "small", @@ -858,8 +1000,10 @@ tf_xla_py_test( tf_xla_py_test( name = "sort_ops_test", - size = "small", + size = "medium", srcs = ["sort_ops_test.py"], + # Times out in fastbuild mode. + tags = ["optonly"], deps = [ "//tensorflow/compiler/tests:xla_test", "//tensorflow/compiler/tf2xla/python:xla", diff --git a/tensorflow/compiler/tests/adadelta_test.py b/tensorflow/compiler/tests/adadelta_test.py new file mode 100644 index 0000000000000000000000000000000000000000..3e3c09c66e72c4de141b64cea3c4693fabb7b2a2 --- /dev/null +++ b/tensorflow/compiler/tests/adadelta_test.py @@ -0,0 +1,134 @@ +# Copyright 2018 The TensorFlow Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================== +"""Tests for Adadelta Optimizer.""" + +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + +import numpy as np + +from tensorflow.compiler.tests import xla_test +from tensorflow.python.framework import constant_op +from tensorflow.python.ops import resource_variable_ops +from tensorflow.python.ops import variables +from tensorflow.python.platform import test +from tensorflow.python.training import adadelta + + +class AdadeltaOptimizerTest(xla_test.XLATestCase): + + def testBasic(self): + num_updates = 4 # number of ADADELTA steps to perform + for dtype in self.float_types: + with self.test_session(), self.test_scope(): + for grad in [0.2, 0.1, 0.01]: + for lr in [1.0, 0.5, 0.1]: + var0_init = [1.0, 2.0] + var1_init = [3.0, 4.0] + var0 = resource_variable_ops.ResourceVariable( + var0_init, dtype=dtype) + var1 = resource_variable_ops.ResourceVariable( + var1_init, dtype=dtype) + + grads = constant_op.constant([grad, grad], dtype=dtype) + + accum = 0.0 + accum_update = 0.0 + + # ADADELTA gradient optimizer + rho = 0.95 + epsilon = 1e-8 + adadelta_opt = adadelta.AdadeltaOptimizer( + learning_rate=lr, rho=rho, epsilon=epsilon) + adadelta_update = adadelta_opt.apply_gradients( + zip([grads, grads], [var0, var1])) + self.evaluate(variables.global_variables_initializer()) + opt_vars = adadelta_opt.variables() + self.assertStartsWith(opt_vars[0].name, var0._shared_name) + self.assertStartsWith(opt_vars[1].name, var0._shared_name) + self.assertStartsWith(opt_vars[2].name, var1._shared_name) + self.assertStartsWith(opt_vars[3].name, var1._shared_name) + self.assertEqual(4, len(opt_vars)) + # Assign slots + slot = [None] * 2 + slot_update = [None] * 2 + self.assertEqual(["accum", "accum_update"], + adadelta_opt.get_slot_names()) + slot[0] = adadelta_opt.get_slot(var0, "accum") + self.assertEquals(slot[0].get_shape(), var0.get_shape()) + self.assertFalse(slot[0] in variables.trainable_variables()) + + slot_update[0] = adadelta_opt.get_slot(var0, "accum_update") + self.assertEquals(slot_update[0].get_shape(), var0.get_shape()) + self.assertFalse(slot_update[0] in variables.trainable_variables()) + + slot[1] = adadelta_opt.get_slot(var1, "accum") + self.assertEquals(slot[1].get_shape(), var1.get_shape()) + self.assertFalse(slot[1] in variables.trainable_variables()) + + slot_update[1] = adadelta_opt.get_slot(var1, "accum_update") + self.assertEquals(slot_update[1].get_shape(), var1.get_shape()) + self.assertFalse(slot_update[1] in variables.trainable_variables()) + + # Fetch params to validate initial values + self.assertAllClose(var0_init, self.evaluate(var0)) + self.assertAllClose(var1_init, self.evaluate(var1)) + + update = [None] * num_updates + tot_update = 0 + for step in range(num_updates): + # Run adadelta update for comparison + self.evaluate(adadelta_update) + + # Perform initial update without previous accum values + accum = accum * rho + (grad**2) * (1 - rho) + update[step] = ( + np.sqrt(accum_update + epsilon) * + (1. / np.sqrt(accum + epsilon)) * grad) + accum_update = ( + accum_update * rho + (update[step]**2) * (1.0 - rho)) + tot_update += update[step] * lr + + # Check that the accumulators have been updated + for slot_idx in range(2): + self.assertAllCloseAccordingToType( + np.array([accum, accum], dtype=dtype), + self.evaluate(slot[slot_idx]), + rtol=1e-5) + + self.assertAllCloseAccordingToType( + np.array([accum_update, accum_update], dtype=dtype), + self.evaluate(slot_update[slot_idx]), + rtol=1e-5) + + # Check that the parameters have been updated + self.assertAllCloseAccordingToType( + np.array( + [var0_init[0] - tot_update, var0_init[1] - tot_update], + dtype=dtype), + self.evaluate(var0), + rtol=1e-5) + + self.assertAllCloseAccordingToType( + np.array( + [var1_init[0] - tot_update, var1_init[1] - tot_update], + dtype=dtype), + self.evaluate(var1), + rtol=1e-5) + + +if __name__ == "__main__": + test.main() diff --git a/tensorflow/compiler/tests/adagrad_da_test.py b/tensorflow/compiler/tests/adagrad_da_test.py new file mode 100644 index 0000000000000000000000000000000000000000..dc1625793aa44b96d3b96e175237caf96e7d7e74 --- /dev/null +++ b/tensorflow/compiler/tests/adagrad_da_test.py @@ -0,0 +1,165 @@ +# Copyright 2016 The TensorFlow Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================== +"""Tests for AdagradDA optimizer.""" + +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + +import numpy as np + +from tensorflow.compiler.tests import xla_test +from tensorflow.python.framework import constant_op +from tensorflow.python.framework import dtypes +from tensorflow.python.ops import resource_variable_ops +from tensorflow.python.ops import variables +from tensorflow.python.platform import test +from tensorflow.python.training import adagrad_da + + +class AdagradDAOptimizerTest(xla_test.XLATestCase): + + def testAdagradDAWithoutRegularizationBasic1(self): + for dtype in self.float_types: + with self.test_session(), self.test_scope(): + global_step = resource_variable_ops.ResourceVariable( + 0, dtype=dtypes.int64) + var0 = resource_variable_ops.ResourceVariable([0.0, 0.0], dtype=dtype) + var1 = resource_variable_ops.ResourceVariable([0.0, 0.0], dtype=dtype) + grads0 = constant_op.constant([0.1, 0.2], dtype=dtype) + grads1 = constant_op.constant([0.01, 0.02], dtype=dtype) + opt = adagrad_da.AdagradDAOptimizer( + 3.0, + global_step, + initial_gradient_squared_accumulator_value=0.1, + l1_regularization_strength=0.0, + l2_regularization_strength=0.0) + update = opt.apply_gradients( + zip([grads0, grads1], [var0, var1]), global_step=global_step) + variables.global_variables_initializer().run() + + self.assertAllClose([0.0, 0.0], var0.eval()) + self.assertAllClose([0.0, 0.0], var1.eval()) + + # Run a step of AdagradDA + update.run() + + # Let g to be gradient accumulator, gg to be gradient squared + # accumulator, T be the global step, lr is the learning rate, and k the + # initial gradient squared accumulator value. + # w = \dfrac{sign(-g)*lr*|g - l1*T|_{+}}{l2*T*lr + \sqrt{k+gg})} + # For -0.1*3.0*(0.1 - 0)/(0 + sqrt(0.1 + 0.1*0.1)) = -0.904534 + # similarly for others. + self.assertAllCloseAccordingToType( + np.array([-0.904534, -1.603567]), var0.eval()) + self.assertAllCloseAccordingToType( + np.array([-0.094821, -0.189358]), var1.eval()) + + def testAdagradDAwithoutRegularizationBasic2(self): + for dtype in self.float_types: + with self.test_session(), self.test_scope(): + global_step = resource_variable_ops.ResourceVariable( + 0, dtype=dtypes.int64) + var0 = resource_variable_ops.ResourceVariable([1.0, 2.0], dtype=dtype) + var1 = resource_variable_ops.ResourceVariable([4.0, 3.0], dtype=dtype) + grads0 = constant_op.constant([0.1, 0.2], dtype=dtype) + grads1 = constant_op.constant([0.01, 0.02], dtype=dtype) + + opt = adagrad_da.AdagradDAOptimizer( + 3.0, + global_step, + initial_gradient_squared_accumulator_value=0.1, + l1_regularization_strength=0.0, + l2_regularization_strength=0.0) + update = opt.apply_gradients( + zip([grads0, grads1], [var0, var1]), global_step=global_step) + variables.global_variables_initializer().run() + + self.assertAllCloseAccordingToType([1.0, 2.0], var0.eval()) + self.assertAllCloseAccordingToType([4.0, 3.0], var1.eval()) + + # Run a step of AdagradDA + update.run() + + self.assertAllCloseAccordingToType( + np.array([-0.904534, -1.603567]), var0.eval()) + self.assertAllCloseAccordingToType( + np.array([-0.094821, -0.189358]), var1.eval()) + + def testAdagradDAWithL1(self): + for dtype in self.float_types: + with self.test_session(), self.test_scope(): + global_step = resource_variable_ops.ResourceVariable( + 0, dtype=dtypes.int64) + var0 = resource_variable_ops.ResourceVariable([1.0, 2.0], dtype=dtype) + var1 = resource_variable_ops.ResourceVariable([4.0, 3.0], dtype=dtype) + grads0 = constant_op.constant([0.1, 0.2], dtype=dtype) + grads1 = constant_op.constant([0.01, 0.02], dtype=dtype) + + opt = adagrad_da.AdagradDAOptimizer( + 3.0, + global_step, + initial_gradient_squared_accumulator_value=0.1, + l1_regularization_strength=0.001, + l2_regularization_strength=0.0) + update = opt.apply_gradients( + zip([grads0, grads1], [var0, var1]), global_step=global_step) + variables.global_variables_initializer().run() + + self.assertAllCloseAccordingToType([1.0, 2.0], var0.eval()) + self.assertAllCloseAccordingToType([4.0, 3.0], var1.eval()) + + # Run a step of AdagradDA + update.run() + + self.assertAllCloseAccordingToType( + np.array([-0.895489, -1.59555]), var0.eval()) + self.assertAllCloseAccordingToType( + np.array([-0.085339, -0.17989]), var1.eval()) + + def testAdagradDAWithL1_L2(self): + for dtype in self.float_types: + with self.test_session(), self.test_scope(): + global_step = resource_variable_ops.ResourceVariable( + 0, dtype=dtypes.int64) + var0 = resource_variable_ops.ResourceVariable([1.0, 2.0], dtype=dtype) + var1 = resource_variable_ops.ResourceVariable([4.0, 3.0], dtype=dtype) + grads0 = constant_op.constant([0.1, 0.2], dtype=dtype) + grads1 = constant_op.constant([0.01, 0.02], dtype=dtype) + + opt = adagrad_da.AdagradDAOptimizer( + 3.0, + global_step, + initial_gradient_squared_accumulator_value=0.1, + l1_regularization_strength=0.001, + l2_regularization_strength=2.0) + update = opt.apply_gradients( + zip([grads0, grads1], [var0, var1]), global_step=global_step) + variables.global_variables_initializer().run() + + self.assertAllCloseAccordingToType([1.0, 2.0], var0.eval()) + self.assertAllCloseAccordingToType([4.0, 3.0], var1.eval()) + + # Run a step of AdagradDA + update.run() + + self.assertAllCloseAccordingToType( + np.array([-0.046907, -0.093659]), var0.eval()) + self.assertAllCloseAccordingToType( + np.array([-0.004275, -0.009023]), var1.eval()) + + +if __name__ == "__main__": + test.main() diff --git a/tensorflow/compiler/tests/adagrad_test.py b/tensorflow/compiler/tests/adagrad_test.py index 9a93b3216404d8ed21fd6c57757bec1730c119b4..d775850a80e9f83f7b2c9f1cf8997dd50e229635 100644 --- a/tensorflow/compiler/tests/adagrad_test.py +++ b/tensorflow/compiler/tests/adagrad_test.py @@ -20,7 +20,7 @@ from __future__ import print_function import numpy as np -from tensorflow.compiler.tests.xla_test import XLATestCase +from tensorflow.compiler.tests import xla_test from tensorflow.python.framework import constant_op from tensorflow.python.ops import resource_variable_ops from tensorflow.python.ops import variables @@ -28,7 +28,7 @@ from tensorflow.python.platform import test from tensorflow.python.training import adagrad -class AdagradOptimizerTest(XLATestCase): +class AdagradOptimizerTest(xla_test.XLATestCase): def testBasic(self): for dtype in self.float_types: diff --git a/tensorflow/compiler/tests/adam_test.py b/tensorflow/compiler/tests/adam_test.py index 3215dc36e5b2d517aa951db1b0d41188185ef93a..03554d6933aca39b428c6af4be0c78e2c7ccb0c9 100644 --- a/tensorflow/compiler/tests/adam_test.py +++ b/tensorflow/compiler/tests/adam_test.py @@ -20,7 +20,7 @@ from __future__ import print_function import numpy as np -from tensorflow.compiler.tests.xla_test import XLATestCase +from tensorflow.compiler.tests import xla_test from tensorflow.python.framework import constant_op from tensorflow.python.ops import array_ops from tensorflow.python.ops import resource_variable_ops @@ -48,7 +48,7 @@ def adam_update_numpy(param, return param_t, m_t, v_t -class AdamOptimizerTest(XLATestCase): +class AdamOptimizerTest(xla_test.XLATestCase): def testBasic(self): for dtype in self.float_types: diff --git a/tensorflow/compiler/tests/adamax_test.py b/tensorflow/compiler/tests/adamax_test.py new file mode 100644 index 0000000000000000000000000000000000000000..c4fdbc5974319db9243eb2c323746cbaaea795f6 --- /dev/null +++ b/tensorflow/compiler/tests/adamax_test.py @@ -0,0 +1,139 @@ +# Copyright 2018 The TensorFlow Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================== +"""Tests for AdaMax optimizer.""" + +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + +import numpy as np + +from tensorflow.compiler.tests import xla_test +from tensorflow.contrib.opt.python.training import adamax +from tensorflow.python.framework import constant_op +from tensorflow.python.framework import ops +from tensorflow.python.ops import resource_variable_ops +from tensorflow.python.ops import variable_scope +from tensorflow.python.ops import variables +from tensorflow.python.platform import test + + +def adamax_update_numpy(param, + g_t, + t, + m, + v, + alpha=0.001, + beta1=0.9, + beta2=0.999, + epsilon=1e-8): + m_t = beta1 * m + (1 - beta1) * g_t + v_t = np.maximum(beta2 * v, np.abs(g_t)) + param_t = param - (alpha / (1 - beta1**t)) * (m_t / (v_t + epsilon)) + return param_t, m_t, v_t + + +class AdaMaxOptimizerTest(xla_test.XLATestCase): + + def testBasic(self): + for i, dtype in enumerate(self.float_types): + with self.test_session(), self.test_scope(): + variable_scope.get_variable_scope().set_use_resource(True) + # Initialize variables for numpy implementation. + m0, v0, m1, v1 = 0.0, 0.0, 0.0, 0.0 + var0_np = np.array([1.0, 2.0], dtype=dtype) + grads0_np = np.array([0.1, 0.1], dtype=dtype) + var1_np = np.array([3.0, 4.0], dtype=dtype) + grads1_np = np.array([0.01, 0.01], dtype=dtype) + + var0 = resource_variable_ops.ResourceVariable( + var0_np, name="var0_%d" % i) + var1 = resource_variable_ops.ResourceVariable( + var1_np, name="var1_%d" % i) + grads0 = constant_op.constant(grads0_np) + grads1 = constant_op.constant(grads1_np) + + opt = adamax.AdaMaxOptimizer() + update = opt.apply_gradients(zip([grads0, grads1], [var0, var1])) + opt_variables = opt.variables() + beta1_power = opt._get_beta_accumulators() + self.assertTrue(beta1_power is not None) + self.assertIn(beta1_power, opt_variables) + + with ops.Graph().as_default(): + # Shouldn't return non-slot variables from other graphs. + self.assertEqual(0, len(opt.variables())) + + variables.global_variables_initializer().run() + # Fetch params to validate initial values + self.assertAllClose([1.0, 2.0], var0.eval()) + self.assertAllClose([3.0, 4.0], var1.eval()) + + beta1_power = opt._get_beta_accumulators() + + # Run 3 steps of AdaMax + for t in range(1, 4): + update.run() + + self.assertAllCloseAccordingToType(0.9**(t + 1), beta1_power.eval()) + + var0_np, m0, v0 = adamax_update_numpy(var0_np, grads0_np, t, m0, v0) + var1_np, m1, v1 = adamax_update_numpy(var1_np, grads1_np, t, m1, v1) + + # Validate updated params + self.assertAllCloseAccordingToType(var0_np, var0.eval(), rtol=1e-2) + self.assertAllCloseAccordingToType(var1_np, var1.eval(), rtol=1e-2) + self.assertEqual("var0_%d/AdaMax:0" % (i,), + opt.get_slot(var=var0, name="m").name) + + def testTensorLearningRate(self): + for dtype in self.float_types: + with self.test_session(), self.test_scope(): + variable_scope.get_variable_scope().set_use_resource(True) + # Initialize variables for numpy implementation. + m0, v0, m1, v1 = 0.0, 0.0, 0.0, 0.0 + var0_np = np.array([1.0, 2.0], dtype=dtype) + grads0_np = np.array([0.1, 0.1], dtype=dtype) + var1_np = np.array([3.0, 4.0], dtype=dtype) + grads1_np = np.array([0.01, 0.01], dtype=dtype) + + var0 = resource_variable_ops.ResourceVariable(var0_np) + var1 = resource_variable_ops.ResourceVariable(var1_np) + grads0 = constant_op.constant(grads0_np) + grads1 = constant_op.constant(grads1_np) + opt = adamax.AdaMaxOptimizer(constant_op.constant(0.001)) + update = opt.apply_gradients(zip([grads0, grads1], [var0, var1])) + variables.global_variables_initializer().run() + + # Fetch params to validate initial values + self.assertAllClose([1.0, 2.0], var0.eval()) + self.assertAllClose([3.0, 4.0], var1.eval()) + + beta1_power = opt._get_beta_accumulators() + + # Run 3 steps of AdaMax + for t in range(1, 4): + self.assertAllCloseAccordingToType(0.9**t, beta1_power.eval()) + update.run() + + var0_np, m0, v0 = adamax_update_numpy(var0_np, grads0_np, t, m0, v0) + var1_np, m1, v1 = adamax_update_numpy(var1_np, grads1_np, t, m1, v1) + + # Validate updated params + self.assertAllCloseAccordingToType(var0_np, var0.eval()) + self.assertAllCloseAccordingToType(var1_np, var1.eval()) + +if __name__ == "__main__": + test.main() diff --git a/tensorflow/compiler/tests/addsign_test.py b/tensorflow/compiler/tests/addsign_test.py new file mode 100644 index 0000000000000000000000000000000000000000..9ec5a964cbb4dd98d2ef2d0b684872292118800f --- /dev/null +++ b/tensorflow/compiler/tests/addsign_test.py @@ -0,0 +1,142 @@ +# Copyright 2017 The TensorFlow Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================== +"""Tests for AddSign.""" + +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + +import numpy as np + +from tensorflow.compiler.tests import xla_test +from tensorflow.contrib.opt.python.training import addsign +from tensorflow.contrib.opt.python.training import sign_decay +from tensorflow.python.framework import constant_op +from tensorflow.python.ops import resource_variable_ops +from tensorflow.python.ops import variables +from tensorflow.python.platform import test + + +def py_linear_decay_fn(decay_steps): + def linear_decay(step): + step = min(step, decay_steps) + return float(decay_steps - step) / decay_steps + return linear_decay + + +def addsign_update_numpy(params, + g_t, + m, + lr, + alpha=1.0, + beta=0.9, + py_sign_decay_fn=None, + t=None): + m_t = beta * m + (1 - beta) * g_t + if py_sign_decay_fn is None: + sign_decayed = 1.0 + else: + sign_decayed = py_sign_decay_fn(t-1) + multiplier = alpha + sign_decayed * np.sign(g_t) * np.sign(m_t) + params_t = params - lr * multiplier * g_t + return params_t, m_t + + +class AddSignTest(xla_test.XLATestCase): + + def _testDense(self, + learning_rate=0.1, + sign_decay_fn=None, + py_sign_decay_fn=None, + alpha=1.0, + beta=0.9): + for dtype in self.float_types: + with self.test_session(), self.test_scope(): + # Initialize variables for numpy implementation. + m0, m1 = 0.0, 0.0 + var0_np = np.array([1.0, 2.0], dtype=dtype) + grads0_np = np.array([0.1, 0.1], dtype=dtype) + var1_np = np.array([3.0, 4.0], dtype=dtype) + grads1_np = np.array([0.01, 0.01], dtype=dtype) + + var0 = resource_variable_ops.ResourceVariable(var0_np) + var1 = resource_variable_ops.ResourceVariable(var1_np) + global_step = resource_variable_ops.ResourceVariable(0, trainable=False) + grads0 = constant_op.constant(grads0_np) + grads1 = constant_op.constant(grads1_np) + + opt = addsign.AddSignOptimizer( + learning_rate=learning_rate, + alpha=alpha, + beta=beta, + sign_decay_fn=sign_decay_fn, + ) + update = opt.apply_gradients(zip([grads0, grads1], [var0, var1]), + global_step=global_step) + neg_update = opt.apply_gradients(zip([-grads0, -grads1], [var0, var1]), + global_step=global_step) + variables.global_variables_initializer().run() + + # Fetch params to validate initial values + self.assertAllClose([1.0, 2.0], var0.eval()) + self.assertAllClose([3.0, 4.0], var1.eval()) + + # Run 7 steps of AddSign + # first 4 steps with positive gradient + # last 3 steps with negative gradient (sign(gm) should be -1) + for t in range(1, 8): + if t < 5: + update.run() + else: + neg_update.run() + + var0_np, m0 = addsign_update_numpy( + var0_np, + grads0_np if t < 5 else -grads0_np, + m0, + learning_rate, + alpha=alpha, + beta=beta, + py_sign_decay_fn=py_sign_decay_fn, + t=t, + ) + var1_np, m1 = addsign_update_numpy( + var1_np, + grads1_np if t < 5 else -grads1_np, + m1, + learning_rate, + alpha=alpha, + beta=beta, + py_sign_decay_fn=py_sign_decay_fn, + t=t, + ) + + # Validate updated params + self.assertAllCloseAccordingToType( + var0_np, var0.eval(), half_rtol=1e-2) + self.assertAllCloseAccordingToType(var1_np, var1.eval()) + + def testDense(self): + decay_steps = 10 + sign_decay_fn = sign_decay.get_linear_decay_fn(decay_steps) + py_sign_decay_fn = py_linear_decay_fn(decay_steps) + self._testDense() + self._testDense(learning_rate=0.01, alpha=0.1, beta=0.8) + self._testDense( + sign_decay_fn=sign_decay_fn, py_sign_decay_fn=py_sign_decay_fn) + + +if __name__ == '__main__': + test.main() diff --git a/tensorflow/compiler/tests/binary_ops_test.py b/tensorflow/compiler/tests/binary_ops_test.py index afef36d9d202a822d81ace185a9112bb83daae8c..9cb3d0454608c37e669d5b4360bc39bf1bf7e68c 100644 --- a/tensorflow/compiler/tests/binary_ops_test.py +++ b/tensorflow/compiler/tests/binary_ops_test.py @@ -20,7 +20,7 @@ from __future__ import print_function import numpy as np -from tensorflow.compiler.tests.xla_test import XLATestCase +from tensorflow.compiler.tests import xla_test from tensorflow.python.framework import dtypes from tensorflow.python.framework import errors from tensorflow.python.ops import array_ops @@ -32,7 +32,7 @@ from tensorflow.python.ops import nn_ops from tensorflow.python.platform import googletest -class BinaryOpsTest(XLATestCase): +class BinaryOpsTest(xla_test.XLATestCase): """Test cases for binary operators.""" def _testBinary(self, op, a, b, expected, equality_test=None): diff --git a/tensorflow/compiler/tests/bucketize_op_test.py b/tensorflow/compiler/tests/bucketize_op_test.py index fde9759a1c209844caac99d5f303cd3e406e5370..ef4d5f6322b7ae79b051795b5af7e6f7f1e55550 100644 --- a/tensorflow/compiler/tests/bucketize_op_test.py +++ b/tensorflow/compiler/tests/bucketize_op_test.py @@ -18,7 +18,7 @@ from __future__ import absolute_import from __future__ import division from __future__ import print_function -from tensorflow.compiler.tests.xla_test import XLATestCase +from tensorflow.compiler.tests import xla_test from tensorflow.python.framework import dtypes from tensorflow.python.framework import errors_impl from tensorflow.python.ops import array_ops @@ -26,7 +26,7 @@ from tensorflow.python.ops import math_ops from tensorflow.python.platform import test -class BucketizationOpTest(XLATestCase): +class BucketizationOpTest(xla_test.XLATestCase): def testInt(self): with self.test_session() as sess: diff --git a/tensorflow/compiler/tests/categorical_op_test.py b/tensorflow/compiler/tests/categorical_op_test.py index 035cdea1786d39f3d21bb63be5c8ccffe1608bdf..a4e7f75081dfd07fd4b5c94c33908aab8e7d8aa9 100644 --- a/tensorflow/compiler/tests/categorical_op_test.py +++ b/tensorflow/compiler/tests/categorical_op_test.py @@ -22,7 +22,7 @@ import collections import numpy as np -from tensorflow.compiler.tests.xla_test import XLATestCase +from tensorflow.compiler.tests import xla_test from tensorflow.python.framework import dtypes from tensorflow.python.framework import random_seed from tensorflow.python.ops import array_ops @@ -32,7 +32,7 @@ from tensorflow.python.platform import googletest # TODO(srvasude): Merge this with # third_party/tensorflow/python/kernel_tests/random/multinomial_op_test.py. -class CategoricalTest(XLATestCase): +class CategoricalTest(xla_test.XLATestCase): """Test cases for random-number generating operators.""" def output_dtypes(self): diff --git a/tensorflow/compiler/tests/cholesky_op_test.py b/tensorflow/compiler/tests/cholesky_op_test.py index 1a8989d7c2f617525c301f30fd899a01362310bf..d2867278af93812eae804b66a7a6b706f98fa600 100644 --- a/tensorflow/compiler/tests/cholesky_op_test.py +++ b/tensorflow/compiler/tests/cholesky_op_test.py @@ -23,7 +23,7 @@ import unittest import numpy as np from six.moves import xrange # pylint: disable=redefined-builtin -from tensorflow.compiler.tests.xla_test import XLATestCase +from tensorflow.compiler.tests import xla_test from tensorflow.python.framework import constant_op from tensorflow.python.framework import dtypes from tensorflow.python.ops import array_ops @@ -32,7 +32,7 @@ from tensorflow.python.ops import math_ops from tensorflow.python.platform import test -class CholeskyOpTest(XLATestCase): +class CholeskyOpTest(xla_test.XLATestCase): # Cholesky defined for float64, float32, complex64, complex128 # (https://www.tensorflow.org/api_docs/python/tf/cholesky) diff --git a/tensorflow/compiler/tests/clustering_test.py b/tensorflow/compiler/tests/clustering_test.py index 574f82fc717818334ac5d72ebef2191f1c18e669..e42ebf8f9e01dab13cde15979ffc42b7c0fbc57b 100644 --- a/tensorflow/compiler/tests/clustering_test.py +++ b/tensorflow/compiler/tests/clustering_test.py @@ -21,7 +21,7 @@ from __future__ import print_function import numpy as np from six.moves import xrange # pylint: disable=redefined-builtin -from tensorflow.compiler.tests.xla_test import XLATestCase +from tensorflow.compiler.tests import xla_test from tensorflow.python.framework import constant_op from tensorflow.python.framework import dtypes from tensorflow.python.framework import ops @@ -32,7 +32,7 @@ from tensorflow.python.platform import googletest CPU_DEVICE = "/job:localhost/replica:0/task:0/cpu:0" -class ClusteringTest(XLATestCase): +class ClusteringTest(xla_test.XLATestCase): def testAdd(self): val1 = np.array([4, 3, 2, 1], dtype=np.float32) diff --git a/tensorflow/compiler/tests/concat_ops_test.py b/tensorflow/compiler/tests/concat_ops_test.py index f10973e19f1945515b776cf86349445ed7334629..d9ad4281477e87f79f2ecb52989ae86a5030d0cc 100644 --- a/tensorflow/compiler/tests/concat_ops_test.py +++ b/tensorflow/compiler/tests/concat_ops_test.py @@ -20,7 +20,7 @@ from __future__ import print_function import numpy as np -from tensorflow.compiler.tests.xla_test import XLATestCase +from tensorflow.compiler.tests import xla_test from tensorflow.python.framework import constant_op from tensorflow.python.framework import dtypes from tensorflow.python.ops import array_ops @@ -30,7 +30,7 @@ from tensorflow.python.ops import math_ops from tensorflow.python.platform import googletest -class ConcatTest(XLATestCase): +class ConcatTest(xla_test.XLATestCase): def testHStack(self): with self.test_session(): @@ -292,7 +292,7 @@ class ConcatTest(XLATestCase): array_ops.concat([scalar, scalar, scalar], dim) -class ConcatOffsetTest(XLATestCase): +class ConcatOffsetTest(xla_test.XLATestCase): def testBasic(self): with self.test_session() as sess: @@ -306,7 +306,7 @@ class ConcatOffsetTest(XLATestCase): self.assertAllEqual(ans, [[0, 0, 0], [0, 3, 0], [0, 10, 0]]) -class PackTest(XLATestCase): +class PackTest(xla_test.XLATestCase): def testBasic(self): with self.test_session() as sess: diff --git a/tensorflow/compiler/tests/conv2d_test.py b/tensorflow/compiler/tests/conv2d_test.py index d12e1ff1e8f4564f39642bd0b64fc40d8dca8ef0..f9db103f6d0f9ea0e393a0971593552ec5c14079 100644 --- a/tensorflow/compiler/tests/conv2d_test.py +++ b/tensorflow/compiler/tests/conv2d_test.py @@ -26,23 +26,20 @@ from absl.testing import parameterized import numpy as np from tensorflow.compiler.tests import test_utils -from tensorflow.compiler.tests.xla_test import XLATestCase +from tensorflow.compiler.tests import xla_test from tensorflow.python.framework import dtypes from tensorflow.python.ops import array_ops from tensorflow.python.ops import gen_nn_ops from tensorflow.python.ops import nn_ops from tensorflow.python.platform import googletest - DATA_FORMATS = ( ("_data_format_NHWC", "NHWC"), ("_data_format_NCHW", "NCHW"), - ("_data_format_HWNC", "HWNC"), - ("_data_format_HWCN", "HWCN"), ) -class Conv2DTest(XLATestCase, parameterized.TestCase): +class Conv2DTest(xla_test.XLATestCase, parameterized.TestCase): def _VerifyValues(self, input_sizes=None, @@ -236,7 +233,7 @@ class Conv2DTest(XLATestCase, parameterized.TestCase): expected=np.reshape([108, 128], [1, 1, 1, 2])) -class Conv2DBackpropInputTest(XLATestCase, parameterized.TestCase): +class Conv2DBackpropInputTest(xla_test.XLATestCase, parameterized.TestCase): def _VerifyValues(self, input_sizes=None, @@ -534,7 +531,7 @@ class Conv2DBackpropInputTest(XLATestCase, parameterized.TestCase): expected=[5, 0, 11, 0, 0, 0, 17, 0, 23]) -class Conv2DBackpropFilterTest(XLATestCase, parameterized.TestCase): +class Conv2DBackpropFilterTest(xla_test.XLATestCase, parameterized.TestCase): def _VerifyValues(self, input_sizes=None, diff --git a/tensorflow/compiler/tests/conv3d_test.py b/tensorflow/compiler/tests/conv3d_test.py index 3bebf46511cbc471d3fbbbe92d28511fcc717387..31ee41f04f27d387415e9fa2c4fa70b33cab7b04 100644 --- a/tensorflow/compiler/tests/conv3d_test.py +++ b/tensorflow/compiler/tests/conv3d_test.py @@ -21,7 +21,7 @@ from __future__ import print_function import numpy as np from six.moves import xrange # pylint: disable=redefined-builtin -from tensorflow.compiler.tests.xla_test import XLATestCase +from tensorflow.compiler.tests import xla_test from tensorflow.python.framework import constant_op from tensorflow.python.framework import dtypes from tensorflow.python.ops import array_ops @@ -33,7 +33,7 @@ from tensorflow.python.platform import googletest # Test cloned from # tensorflow/python/kernel_tests/conv3d_backprop_filter_v2_grad_test.py -class Conv3DBackpropFilterV2GradTest(XLATestCase): +class Conv3DBackpropFilterV2GradTest(xla_test.XLATestCase): def testGradient(self): with self.test_session(), self.test_scope(): @@ -66,7 +66,7 @@ class Conv3DBackpropFilterV2GradTest(XLATestCase): # Test cloned from tensorflow/python/kernel_tests/conv3d_transpose_test.py -class Conv3DTransposeTest(XLATestCase): +class Conv3DTransposeTest(xla_test.XLATestCase): def testConv3DTransposeSingleStride(self): with self.test_session(), self.test_scope(): diff --git a/tensorflow/compiler/tests/depthwise_conv_op_test.py b/tensorflow/compiler/tests/depthwise_conv_op_test.py index 03d96a2cd8ab22a472a67f092e36224820405fa8..98dc73e189f99b7b811487756659d89dacb97d8a 100644 --- a/tensorflow/compiler/tests/depthwise_conv_op_test.py +++ b/tensorflow/compiler/tests/depthwise_conv_op_test.py @@ -21,7 +21,7 @@ from __future__ import print_function import numpy as np from six.moves import xrange # pylint: disable=redefined-builtin -from tensorflow.compiler.tests.xla_test import XLATestCase +from tensorflow.compiler.tests import xla_test from tensorflow.python.framework import constant_op from tensorflow.python.framework import ops from tensorflow.python.ops import array_ops @@ -114,7 +114,7 @@ def CheckGradConfigsToTest(): yield i, f, o, s, p -class DepthwiseConv2DTest(XLATestCase): +class DepthwiseConv2DTest(xla_test.XLATestCase): # This is testing that depthwise_conv2d and depthwise_conv2d_native # produce the same results. It also tests that NCHW and NWHC diff --git a/tensorflow/compiler/tests/dynamic_slice_ops_test.py b/tensorflow/compiler/tests/dynamic_slice_ops_test.py index 6a46d2ec3e7aee3a4ecfbf1ab9f622d8eb659e3c..154e36b10e6da409606ae6022aaf53e34c8e37cc 100644 --- a/tensorflow/compiler/tests/dynamic_slice_ops_test.py +++ b/tensorflow/compiler/tests/dynamic_slice_ops_test.py @@ -20,14 +20,14 @@ from __future__ import print_function import numpy as np -from tensorflow.compiler.tests.xla_test import XLATestCase +from tensorflow.compiler.tests import xla_test from tensorflow.compiler.tf2xla.python import xla from tensorflow.python.framework import dtypes from tensorflow.python.ops import array_ops from tensorflow.python.platform import test -class DynamicUpdateSliceOpsTest(XLATestCase): +class DynamicUpdateSliceOpsTest(xla_test.XLATestCase): def _assertOpOutputMatchesExpected(self, op, args, expected): with self.test_session() as session: diff --git a/tensorflow/compiler/tests/dynamic_stitch_test.py b/tensorflow/compiler/tests/dynamic_stitch_test.py index c109c27abe2f145685f83251e1d21ec8ddad563a..edd78153b56bb5bf1c268936fb82a60581389733 100644 --- a/tensorflow/compiler/tests/dynamic_stitch_test.py +++ b/tensorflow/compiler/tests/dynamic_stitch_test.py @@ -20,14 +20,14 @@ from __future__ import print_function import numpy as np -from tensorflow.compiler.tests.xla_test import XLATestCase +from tensorflow.compiler.tests import xla_test from tensorflow.python.framework import dtypes from tensorflow.python.ops import array_ops from tensorflow.python.ops import data_flow_ops from tensorflow.python.platform import googletest -class DynamicStitchTest(XLATestCase): +class DynamicStitchTest(xla_test.XLATestCase): def _AssertDynamicStitchResultIs(self, indices, data, expected): with self.test_session() as session: diff --git a/tensorflow/compiler/tests/eager_test.py b/tensorflow/compiler/tests/eager_test.py index e438832a23a670596d12cbc67d71a9f561b82193..6ead15da13b86b9d2b4cf2c19e5cf2a90b061b91 100644 --- a/tensorflow/compiler/tests/eager_test.py +++ b/tensorflow/compiler/tests/eager_test.py @@ -20,7 +20,7 @@ from __future__ import print_function import numpy as np -from tensorflow.compiler.tests.xla_test import XLATestCase +from tensorflow.compiler.tests import xla_test from tensorflow.core.protobuf import config_pb2 from tensorflow.python.eager import backprop from tensorflow.python.eager import context @@ -40,7 +40,7 @@ from tensorflow.python.platform import googletest from tensorflow.python.training import adam -class EagerTest(XLATestCase): +class EagerTest(xla_test.XLATestCase): def testBasic(self): with self.test_scope(): @@ -286,7 +286,7 @@ class EagerTest(XLATestCase): [2.0, 2.0]], embedding_matrix.numpy()) -class EagerFunctionTest(XLATestCase): +class EagerFunctionTest(xla_test.XLATestCase): def testBasic(self): with self.test_scope(): @@ -403,7 +403,7 @@ class EagerFunctionTest(XLATestCase): def testSliceInDefun(self): with self.test_scope(): - @function.defun(compiled=True) + @function.defun def f(x, y): return x[0::2, y:, ...] @@ -418,8 +418,24 @@ class EagerFunctionTest(XLATestCase): self.assertAllEqual(np.ones([1, 2, 4]), z.numpy()) self.assertAllEqual((2, 3, 4), dz.shape.as_list()) + def testNestedDefun(self): + self.skipTest('Nested defuns do not work on TPU at the moment') + with self.test_scope(): + + @function.defun + def times_two(x): + return 2 * x + + @function.defun + def two_x_plus_1(x): + return times_two(x) + 1 + + x = constant_op.constant([2, 3, 4]) + y = two_x_plus_1(x) + self.assertAllEqual([5, 7, 9], y.numpy()) + -class ExcessivePaddingTest(XLATestCase): +class ExcessivePaddingTest(xla_test.XLATestCase): """Test that eager execution works with TPU flattened tensors. Tensors that would normally be excessively padded when written @@ -470,6 +486,36 @@ class ExcessivePaddingTest(XLATestCase): self.assertAllEqual(100 * [[36.0]], reduced) +def multiple_tpus(): + devices = context.context().devices() + return len([d for d in devices if 'device:TPU:' in d]) > 1 + + +class MultiDeviceTest(xla_test.XLATestCase): + """Test running TPU computation on more than one core.""" + + def testBasic(self): + if not multiple_tpus(): + self.skipTest('MultiDeviceTest requires multiple TPU devices.') + + # Compute 10 on TPU core 0 + with ops.device('device:TPU:0'): + two = constant_op.constant(2) + five = constant_op.constant(5) + ten = two * five + self.assertAllEqual(10, ten) + + # Compute 6 on TPU core 1 + with ops.device('device:TPU:1'): + two = constant_op.constant(2) + three = constant_op.constant(3) + six = two * three + self.assertAllEqual(6, six) + + # Copy 10 and 6 to CPU and sum them + self.assertAllEqual(16, ten + six) + + if __name__ == '__main__': ops.enable_eager_execution( config=config_pb2.ConfigProto(log_device_placement=True)) diff --git a/tensorflow/compiler/tests/extract_image_patches_op_test.py b/tensorflow/compiler/tests/extract_image_patches_op_test.py index 0361702e7af778176daed941d64e61198090daf2..5529fdbb090315e1d7f47589777d8a538c90db2b 100644 --- a/tensorflow/compiler/tests/extract_image_patches_op_test.py +++ b/tensorflow/compiler/tests/extract_image_patches_op_test.py @@ -20,13 +20,13 @@ from __future__ import print_function import numpy as np -from tensorflow.compiler.tests.xla_test import XLATestCase +from tensorflow.compiler.tests import xla_test from tensorflow.python.framework import dtypes from tensorflow.python.ops import array_ops from tensorflow.python.platform import test -class ExtractImagePatches(XLATestCase): +class ExtractImagePatches(xla_test.XLATestCase): """Functional tests for ExtractImagePatches op.""" def _VerifyValues(self, image, ksizes, strides, rates, padding, patches): diff --git a/tensorflow/compiler/tests/fake_quant_ops_test.py b/tensorflow/compiler/tests/fake_quant_ops_test.py index dfe9400ef0f55ca011d4e23ba5d735899ca2e054..c48ab178bf53558084fb500b2811c6f0b77a7943 100644 --- a/tensorflow/compiler/tests/fake_quant_ops_test.py +++ b/tensorflow/compiler/tests/fake_quant_ops_test.py @@ -17,14 +17,14 @@ from __future__ import division from __future__ import print_function import numpy as np -from tensorflow.compiler.tests.xla_test import XLATestCase +from tensorflow.compiler.tests import xla_test from tensorflow.python.framework import dtypes from tensorflow.python.ops import array_ops from tensorflow.python.ops import gen_array_ops from tensorflow.python.platform import googletest -class FakeQuantWithMinMaxArgsTest(XLATestCase): +class FakeQuantWithMinMaxArgsTest(xla_test.XLATestCase): """Test cases for FakeQuantWithMinMaxArgs operation.""" # 8 bits, wide range. @@ -122,7 +122,7 @@ class FakeQuantWithMinMaxArgsTest(XLATestCase): result, expected, rtol=1e-3, atol=1e-5, bfloat16_rtol=0.03) -class FakeQuantWithMinMaxArgsGradientTest(XLATestCase): +class FakeQuantWithMinMaxArgsGradientTest(xla_test.XLATestCase): """Test cases for FakeQuantWithMinMaxArgsGradient operation.""" # 8 bits, wide range. @@ -223,7 +223,7 @@ class FakeQuantWithMinMaxArgsGradientTest(XLATestCase): bfloat16_rtol=0.03) -class FakeQuantWithMinMaxVarsTest(XLATestCase): +class FakeQuantWithMinMaxVarsTest(xla_test.XLATestCase): """Test cases for FakeQuantWithMinMaxVars operation.""" # 8 bits, wide range. @@ -328,7 +328,7 @@ class FakeQuantWithMinMaxVarsTest(XLATestCase): result, expected, rtol=1e-3, atol=1e-5, bfloat16_rtol=0.03) -class FakeQuantWithMinMaxVarsGradientTest(XLATestCase): +class FakeQuantWithMinMaxVarsGradientTest(xla_test.XLATestCase): """Test cases for FakeQuantWithMinMaxVarsGradient operation.""" # 8 bits, wide range. diff --git a/tensorflow/compiler/tests/fft_test.py b/tensorflow/compiler/tests/fft_test.py index afb5fa4bb4fefe5bc2ecded826143ffc83c2b559..c64ea249ecb97991952a960a6d16e1bb3be35b17 100644 --- a/tensorflow/compiler/tests/fft_test.py +++ b/tensorflow/compiler/tests/fft_test.py @@ -23,10 +23,11 @@ import itertools import numpy as np import scipy.signal as sps -from tensorflow.compiler.tests.xla_test import XLATestCase +from tensorflow.compiler.tests import xla_test from tensorflow.contrib.signal.python.ops import spectral_ops as signal from tensorflow.python.framework import dtypes from tensorflow.python.ops import array_ops +from tensorflow.python.ops import gradients_impl from tensorflow.python.ops import spectral_ops from tensorflow.python.platform import googletest @@ -57,7 +58,7 @@ INNER_DIMS_2D = pick_10(itertools.product(POWS_OF_2, POWS_OF_2)) INNER_DIMS_3D = pick_10(itertools.product(POWS_OF_2, POWS_OF_2, POWS_OF_2)) -class FFTTest(XLATestCase): +class FFTTest(xla_test.XLATestCase): def _VerifyFftMethod(self, inner_dims, complex_to_input, input_to_expected, tf_method): @@ -97,8 +98,11 @@ class FFTTest(XLATestCase): ph = array_ops.placeholder( dtypes.as_dtype(data.dtype), shape=data.shape) out = signal.stft(ph, ws, hs) + grad = gradients_impl.gradients(out, ph, + grad_ys=array_ops.ones_like(out)) - value = sess.run(out, {ph: data}) + # For gradients, we simply verify that they compile & execute. + value, _ = sess.run([out, grad], {ph: data}) self.assertAllClose(expected, value, rtol=RTOL, atol=ATOL) def testFFT(self): diff --git a/tensorflow/compiler/tests/fifo_queue_test.py b/tensorflow/compiler/tests/fifo_queue_test.py new file mode 100644 index 0000000000000000000000000000000000000000..0f64cc87cde77fbbef6c4e570879e992bc34bafa --- /dev/null +++ b/tensorflow/compiler/tests/fifo_queue_test.py @@ -0,0 +1,201 @@ +# Copyright 2015 The TensorFlow Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================== +"""Tests for tensorflow.ops.data_flow_ops.FIFOQueue.""" + +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + +import time + +from six.moves import xrange # pylint: disable=redefined-builtin + +from tensorflow.compiler.tests import xla_test +from tensorflow.python.framework import dtypes as dtypes_lib +from tensorflow.python.ops import data_flow_ops +from tensorflow.python.platform import test + + +class FIFOQueueTest(xla_test.XLATestCase): + + def testEnqueue(self): + with self.test_session(), self.test_scope(): + q = data_flow_ops.FIFOQueue(10, dtypes_lib.float32) + enqueue_op = q.enqueue((10.0,)) + enqueue_op.run() + + def testEnqueueWithShape(self): + with self.test_session(), self.test_scope(): + q = data_flow_ops.FIFOQueue(10, dtypes_lib.float32, shapes=(3, 2)) + enqueue_correct_op = q.enqueue(([[1.0, 2.0], [3.0, 4.0], [5.0, 6.0]],)) + enqueue_correct_op.run() + with self.assertRaises(ValueError): + q.enqueue(([[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]],)) + self.assertEqual(1, q.size().eval()) + + def testMultipleDequeues(self): + with self.test_session(), self.test_scope(): + q = data_flow_ops.FIFOQueue(10, [dtypes_lib.int32], shapes=[()]) + self.evaluate(q.enqueue([1])) + self.evaluate(q.enqueue([2])) + self.evaluate(q.enqueue([3])) + a, b, c = self.evaluate([q.dequeue(), q.dequeue(), q.dequeue()]) + self.assertAllEqual(set([1, 2, 3]), set([a, b, c])) + + def testQueuesDontShare(self): + with self.test_session(), self.test_scope(): + q = data_flow_ops.FIFOQueue(10, [dtypes_lib.int32], shapes=[()]) + self.evaluate(q.enqueue(1)) + q2 = data_flow_ops.FIFOQueue(10, [dtypes_lib.int32], shapes=[()]) + self.evaluate(q2.enqueue(2)) + self.assertAllEqual(self.evaluate(q2.dequeue()), 2) + self.assertAllEqual(self.evaluate(q.dequeue()), 1) + + def testEnqueueDictWithoutNames(self): + with self.test_session(), self.test_scope(): + q = data_flow_ops.FIFOQueue(10, dtypes_lib.float32) + with self.assertRaisesRegexp(ValueError, "must have names"): + q.enqueue({"a": 12.0}) + + def testParallelEnqueue(self): + with self.test_session() as sess, self.test_scope(): + q = data_flow_ops.FIFOQueue(10, dtypes_lib.float32) + elems = [10.0, 20.0, 30.0, 40.0, 50.0, 60.0, 70.0, 80.0, 90.0, 100.0] + enqueue_ops = [q.enqueue((x,)) for x in elems] + dequeued_t = q.dequeue() + + # Run one producer thread for each element in elems. + def enqueue(enqueue_op): + sess.run(enqueue_op) + + threads = [ + self.checkedThread(target=enqueue, args=(e,)) for e in enqueue_ops + ] + for thread in threads: + thread.start() + for thread in threads: + thread.join() + + # Dequeue every element using a single thread. + results = [] + for _ in xrange(len(elems)): + results.append(dequeued_t.eval()) + self.assertItemsEqual(elems, results) + + def testParallelDequeue(self): + with self.test_session() as sess, self.test_scope(): + q = data_flow_ops.FIFOQueue(10, dtypes_lib.float32) + elems = [10.0, 20.0, 30.0, 40.0, 50.0, 60.0, 70.0, 80.0, 90.0, 100.0] + enqueue_ops = [q.enqueue((x,)) for x in elems] + dequeued_t = q.dequeue() + + # Enqueue every element using a single thread. + for enqueue_op in enqueue_ops: + enqueue_op.run() + + # Run one consumer thread for each element in elems. + results = [] + + def dequeue(): + results.append(sess.run(dequeued_t)) + + threads = [self.checkedThread(target=dequeue) for _ in enqueue_ops] + for thread in threads: + thread.start() + for thread in threads: + thread.join() + self.assertItemsEqual(elems, results) + + def testDequeue(self): + with self.test_session(), self.test_scope(): + q = data_flow_ops.FIFOQueue(10, dtypes_lib.float32) + elems = [10.0, 20.0, 30.0] + enqueue_ops = [q.enqueue((x,)) for x in elems] + dequeued_t = q.dequeue() + + for enqueue_op in enqueue_ops: + enqueue_op.run() + + for i in xrange(len(elems)): + vals = dequeued_t.eval() + self.assertEqual([elems[i]], vals) + + def testEnqueueAndBlockingDequeue(self): + with self.test_session() as sess, self.test_scope(): + q = data_flow_ops.FIFOQueue(3, dtypes_lib.float32) + elems = [10.0, 20.0, 30.0] + enqueue_ops = [q.enqueue((x,)) for x in elems] + dequeued_t = q.dequeue() + + def enqueue(): + # The enqueue_ops should run after the dequeue op has blocked. + # TODO(mrry): Figure out how to do this without sleeping. + time.sleep(0.1) + for enqueue_op in enqueue_ops: + sess.run(enqueue_op) + + results = [] + + def dequeue(): + for _ in xrange(len(elems)): + results.append(sess.run(dequeued_t)) + + enqueue_thread = self.checkedThread(target=enqueue) + dequeue_thread = self.checkedThread(target=dequeue) + enqueue_thread.start() + dequeue_thread.start() + enqueue_thread.join() + dequeue_thread.join() + + for elem, result in zip(elems, results): + self.assertEqual([elem], result) + + def testMultiEnqueueAndDequeue(self): + with self.test_session() as sess, self.test_scope(): + q = data_flow_ops.FIFOQueue(10, (dtypes_lib.int32, dtypes_lib.float32)) + elems = [(5, 10.0), (10, 20.0), (15, 30.0)] + enqueue_ops = [q.enqueue((x, y)) for x, y in elems] + dequeued_t = q.dequeue() + + for enqueue_op in enqueue_ops: + enqueue_op.run() + + for i in xrange(len(elems)): + x_val, y_val = sess.run(dequeued_t) + x, y = elems[i] + self.assertEqual([x], x_val) + self.assertEqual([y], y_val) + + def testQueueSizeEmpty(self): + with self.test_session(), self.test_scope(): + q = data_flow_ops.FIFOQueue(10, dtypes_lib.float32) + self.assertEqual([0], q.size().eval()) + + def testQueueSizeAfterEnqueueAndDequeue(self): + with self.test_session(), self.test_scope(): + q = data_flow_ops.FIFOQueue(10, dtypes_lib.float32) + enqueue_op = q.enqueue((10.0,)) + dequeued_t = q.dequeue() + size = q.size() + self.assertEqual([], size.get_shape()) + + enqueue_op.run() + self.assertEqual(1, size.eval()) + dequeued_t.op.run() + self.assertEqual(0, size.eval()) + + +if __name__ == "__main__": + test.main() diff --git a/tensorflow/compiler/tests/ftrl_test.py b/tensorflow/compiler/tests/ftrl_test.py index 8e6407dffdac3adbcda8cbca2109ef9196defa8c..1da97fd51217a0f28d4b3ba2ccfae3f6b094e65b 100644 --- a/tensorflow/compiler/tests/ftrl_test.py +++ b/tensorflow/compiler/tests/ftrl_test.py @@ -20,7 +20,7 @@ from __future__ import print_function import numpy as np -from tensorflow.compiler.tests.xla_test import XLATestCase +from tensorflow.compiler.tests import xla_test from tensorflow.python.framework import constant_op from tensorflow.python.ops import resource_variable_ops from tensorflow.python.ops import variables @@ -30,7 +30,7 @@ from tensorflow.python.training import ftrl from tensorflow.python.training import gradient_descent -class FtrlOptimizerTest(XLATestCase): +class FtrlOptimizerTest(xla_test.XLATestCase): def initVariableAndGradient(self, dtype): var0 = resource_variable_ops.ResourceVariable([0.0, 0.0], dtype=dtype) diff --git a/tensorflow/compiler/tests/function_test.py b/tensorflow/compiler/tests/function_test.py index 8a3f4b0bdc7a61d6cfa2ba7474ce8579e293a5c7..04fba444460e714ce96205361ac02ed492206b04 100644 --- a/tensorflow/compiler/tests/function_test.py +++ b/tensorflow/compiler/tests/function_test.py @@ -20,7 +20,7 @@ from __future__ import print_function import numpy as np -from tensorflow.compiler.tests.xla_test import XLATestCase +from tensorflow.compiler.tests import xla_test from tensorflow.python.framework import constant_op from tensorflow.python.framework import dtypes from tensorflow.python.framework import function @@ -28,7 +28,7 @@ from tensorflow.python.ops import array_ops from tensorflow.python.platform import googletest -class FunctionTest(XLATestCase): +class FunctionTest(xla_test.XLATestCase): def testFunction(self): """Executes a simple TensorFlow function.""" diff --git a/tensorflow/compiler/tests/fused_batchnorm_test.py b/tensorflow/compiler/tests/fused_batchnorm_test.py index 34cca512d4a289a3304a6a49fc810659af48d400..132e42ac7a28d0769b0de12ea0cee6eae752b245 100644 --- a/tensorflow/compiler/tests/fused_batchnorm_test.py +++ b/tensorflow/compiler/tests/fused_batchnorm_test.py @@ -22,7 +22,7 @@ from absl.testing import parameterized import numpy as np from tensorflow.compiler.tests import test_utils -from tensorflow.compiler.tests.xla_test import XLATestCase +from tensorflow.compiler.tests import xla_test from tensorflow.python.ops import array_ops from tensorflow.python.ops import gen_nn_ops from tensorflow.python.ops import gradient_checker @@ -30,7 +30,7 @@ from tensorflow.python.ops import nn from tensorflow.python.platform import test -class FusedBatchNormTest(XLATestCase, parameterized.TestCase): +class FusedBatchNormTest(xla_test.XLATestCase, parameterized.TestCase): def _reference_training(self, x, scale, offset, epsilon, data_format): if data_format != "NHWC": @@ -126,10 +126,6 @@ class FusedBatchNormTest(XLATestCase, parameterized.TestCase): y_ref, mean_ref, var_ref = self._reference_training( x_val, scale_val, offset_val, epsilon, data_format_src) - # TODO(b/110530713): Support data format HWCN on GPU - if self.device == "XLA_GPU" and data_format == "HWCN": - self.skipTest("GPU does not support data format HWCN.") - with self.test_session() as sess, self.test_scope(): # To avoid constant folding x_val_converted = test_utils.ConvertBetweenDataFormats( @@ -214,10 +210,6 @@ class FusedBatchNormTest(XLATestCase, parameterized.TestCase): grad_x_ref, grad_scale_ref, grad_offset_ref = self._reference_grad( x_val, grad_val, scale_val, mean_val, var_val, epsilon, data_format_src) - # TODO(b/110530713): Support data format HWCN on GPU - if self.device == "XLA_GPU" and data_format == "HWCN": - self.skipTest("GPU does not support data format HWCN.") - with self.test_session() as sess, self.test_scope(): grad_val_converted = test_utils.ConvertBetweenDataFormats( grad_val, data_format_src, data_format) @@ -268,10 +260,6 @@ class FusedBatchNormTest(XLATestCase, parameterized.TestCase): var_val = np.random.random_sample(scale_shape).astype(np.float32) data_format_src = "NHWC" - # TODO(b/110530713): Support data format HWCN on GPU - if self.device == "XLA_GPU" and data_format == "HWCN": - self.skipTest("GPU does not support data format HWCN.") - with self.test_session() as sess, self.test_scope(): grad_val_converted = test_utils.ConvertBetweenDataFormats( grad_val, data_format_src, data_format) diff --git a/tensorflow/compiler/tests/gather_nd_op_test.py b/tensorflow/compiler/tests/gather_nd_op_test.py index 9378b1db7245c0da3e8298e7dcd972491616b0cd..23b0aed34fb460f50c241e5a920cb4f6f613b947 100644 --- a/tensorflow/compiler/tests/gather_nd_op_test.py +++ b/tensorflow/compiler/tests/gather_nd_op_test.py @@ -20,13 +20,13 @@ from __future__ import print_function import numpy as np -from tensorflow.compiler.tests.xla_test import XLATestCase +from tensorflow.compiler.tests import xla_test from tensorflow.python.framework import errors from tensorflow.python.ops import array_ops from tensorflow.python.platform import test -class GatherNdTest(XLATestCase): +class GatherNdTest(xla_test.XLATestCase): def _runGather(self, params, indices): with self.test_session(): diff --git a/tensorflow/compiler/tests/gather_test.py b/tensorflow/compiler/tests/gather_test.py index 1a8c4519118f69ce51ca9a5eb95a9d706c7766cc..e9c8ef7c91a728b7dfc948fd9b315e6c9102f6a3 100644 --- a/tensorflow/compiler/tests/gather_test.py +++ b/tensorflow/compiler/tests/gather_test.py @@ -136,6 +136,20 @@ class GatherTest(xla_test.XLATestCase): self.assertAllEqual( [[7]], gather.eval(feed_dict={params: [4, 7, 2], indices: [[1]]})) + def testGatherPrecision(self): + with self.test_session() as session, self.test_scope(): + data = np.array([[0, 0, 0, 0], [0, 2 * (1 + np.exp2(-8)), 0, 0], + [0, 0, 0, 0], [0.015789, 0.0985, 0.55789, 0.3842]]) + indices = np.array([1, 2, 3, 1]) + dtype = dtypes.float32 + params_np = self._buildParams(data, dtype) + params = array_ops.placeholder(dtype=dtype) + indices_tf = constant_op.constant(indices) + gather_t = array_ops.gather(params, indices_tf) + gather_val = session.run(gather_t, feed_dict={params: params_np}) + np_val = params_np[indices] + self.assertAllEqual(np_val, gather_val) + class GatherBenchmark(test.Benchmark): """Microbenchmarks for the gather op.""" diff --git a/tensorflow/compiler/tests/image_ops_test.py b/tensorflow/compiler/tests/image_ops_test.py index 7cf953ef25ef5daf8a6d4fc9985ed8dbfb2081e5..8b01ef96db3e8ab58850df234c2e05b764be52ba 100644 --- a/tensorflow/compiler/tests/image_ops_test.py +++ b/tensorflow/compiler/tests/image_ops_test.py @@ -25,7 +25,7 @@ import numpy as np from six.moves import xrange # pylint: disable=redefined-builtin -from tensorflow.compiler.tests.xla_test import XLATestCase +from tensorflow.compiler.tests import xla_test from tensorflow.python.framework import dtypes from tensorflow.python.framework import ops from tensorflow.python.ops import array_ops @@ -41,7 +41,7 @@ def GenerateNumpyRandomRGB(shape): return np.random.randint(0, 256, shape) / 256. -class RGBToHSVTest(XLATestCase): +class RGBToHSVTest(xla_test.XLATestCase): def testBatch(self): # Build an arbitrary RGB image @@ -104,7 +104,7 @@ class RGBToHSVTest(XLATestCase): self.assertAllCloseAccordingToType(hsv_tf, hsv_np) -class AdjustContrastTest(XLATestCase): +class AdjustContrastTest(xla_test.XLATestCase): def _testContrast(self, x_np, y_np, contrast_factor): with self.test_session(): @@ -168,7 +168,7 @@ class AdjustContrastTest(XLATestCase): self.assertAllClose(y_tf, y_np, rtol=1e-5, atol=1e-5) -class AdjustHueTest(XLATestCase): +class AdjustHueTest(xla_test.XLATestCase): def testAdjustNegativeHue(self): x_shape = [2, 2, 3] @@ -303,7 +303,7 @@ class AdjustHueTest(XLATestCase): self._adjustHueTf(x_np, delta_h) -class AdjustSaturationTest(XLATestCase): +class AdjustSaturationTest(xla_test.XLATestCase): def _adjust_saturation(self, image, saturation_factor): image = ops.convert_to_tensor(image, name="image") @@ -403,7 +403,7 @@ class AdjustSaturationTest(XLATestCase): self.assertAllClose(y_fused, y_baseline, rtol=2e-5, atol=1e-5) -class ResizeBilinearTest(XLATestCase): +class ResizeBilinearTest(xla_test.XLATestCase): def _assertForwardOpMatchesExpected(self, image_np, diff --git a/tensorflow/compiler/tests/lrn_ops_test.py b/tensorflow/compiler/tests/lrn_ops_test.py index 69bd8f7230d4394c45764d02a88fb0ec097c5756..253b45902fba2df64e5234f135b373cd2a0a7e2a 100644 --- a/tensorflow/compiler/tests/lrn_ops_test.py +++ b/tensorflow/compiler/tests/lrn_ops_test.py @@ -22,7 +22,7 @@ import copy import numpy as np -from tensorflow.compiler.tests.xla_test import XLATestCase +from tensorflow.compiler.tests import xla_test from tensorflow.python.framework import constant_op from tensorflow.python.framework import dtypes from tensorflow.python.framework import ops @@ -36,7 +36,7 @@ CPU_DEVICE = "/job:localhost/replica:0/task:0/cpu:0" # Local response normalization tests. The forward tests are copied from # tensorflow/python/kernel_tests/lrn_op_test.py -class LRNTest(XLATestCase): +class LRNTest(xla_test.XLATestCase): def _LRN(self, input_image, lrn_depth_radius=5, bias=1.0, alpha=1.0, beta=0.5): diff --git a/tensorflow/compiler/tests/matrix_band_part_test.py b/tensorflow/compiler/tests/matrix_band_part_test.py index 29394f9ea5139b30f88f53de0469b27e37d79195..0d9f99f8a6803ecae5f9233518a1768109161ac0 100644 --- a/tensorflow/compiler/tests/matrix_band_part_test.py +++ b/tensorflow/compiler/tests/matrix_band_part_test.py @@ -19,14 +19,14 @@ from __future__ import print_function import numpy as np -from tensorflow.compiler.tests.xla_test import XLATestCase +from tensorflow.compiler.tests import xla_test from tensorflow.python.framework import constant_op from tensorflow.python.framework import dtypes from tensorflow.python.ops import array_ops from tensorflow.python.platform import test -class MatrixBandPartTest(XLATestCase): +class MatrixBandPartTest(xla_test.XLATestCase): def _testMatrixBandPart(self, dtype, shape): with self.test_session(): diff --git a/tensorflow/compiler/tests/matrix_triangular_solve_op_test.py b/tensorflow/compiler/tests/matrix_triangular_solve_op_test.py index 5819b2bf2b55b9213a039c0ba82dd0bf1c738b00..2bb8a97bdaf5836a05501ab9754433e29ae34675 100644 --- a/tensorflow/compiler/tests/matrix_triangular_solve_op_test.py +++ b/tensorflow/compiler/tests/matrix_triangular_solve_op_test.py @@ -22,7 +22,7 @@ import itertools import numpy as np -from tensorflow.compiler.tests.xla_test import XLATestCase +from tensorflow.compiler.tests import xla_test from tensorflow.python.framework import constant_op from tensorflow.python.framework import dtypes from tensorflow.python.ops import array_ops @@ -35,7 +35,7 @@ def MakePlaceholder(x): return array_ops.placeholder(dtypes.as_dtype(x.dtype), shape=x.shape) -class MatrixTriangularSolveOpTest(XLATestCase): +class MatrixTriangularSolveOpTest(xla_test.XLATestCase): # MatrixTriangularSolve defined for float64, float32, complex64, complex128 # (https://www.tensorflow.org/api_docs/python/tf/matrix_triangular_solve) diff --git a/tensorflow/compiler/tests/momentum_test.py b/tensorflow/compiler/tests/momentum_test.py index af9394e7d7dc9cf7dd009420ff9c845aec8785bd..c2592c54cf83d41f0e3bdbc1f4dc9ff276ddb078 100644 --- a/tensorflow/compiler/tests/momentum_test.py +++ b/tensorflow/compiler/tests/momentum_test.py @@ -20,7 +20,7 @@ from __future__ import print_function import numpy as np -from tensorflow.compiler.tests.xla_test import XLATestCase +from tensorflow.compiler.tests import xla_test from tensorflow.python.framework import constant_op from tensorflow.python.framework import dtypes from tensorflow.python.ops import array_ops @@ -30,7 +30,7 @@ from tensorflow.python.platform import test from tensorflow.python.training import momentum as momentum_lib -class MomentumOptimizerTest(XLATestCase): +class MomentumOptimizerTest(xla_test.XLATestCase): def _update_nesterov_momentum_numpy(self, var, accum, g, lr, momentum): var += accum * lr * momentum diff --git a/tensorflow/compiler/tests/nary_ops_test.py b/tensorflow/compiler/tests/nary_ops_test.py index e4843b169b943b63346b783ddc50039030988ca5..da08225e9fc0d5a8ec21ee9961c4758fa38628b4 100644 --- a/tensorflow/compiler/tests/nary_ops_test.py +++ b/tensorflow/compiler/tests/nary_ops_test.py @@ -22,14 +22,14 @@ import unittest import numpy as np -from tensorflow.compiler.tests.xla_test import XLATestCase +from tensorflow.compiler.tests import xla_test from tensorflow.python.framework import dtypes from tensorflow.python.ops import array_ops from tensorflow.python.ops import math_ops from tensorflow.python.platform import googletest -class NAryOpsTest(XLATestCase): +class NAryOpsTest(xla_test.XLATestCase): def _testNAry(self, op, args, expected, equality_fn=None): with self.test_session() as session: diff --git a/tensorflow/compiler/tests/nullary_ops_test.py b/tensorflow/compiler/tests/nullary_ops_test.py index 6f588d8ab562cb24f33c4c2987df22264aede027..2f9122645d3c5ccabc8130ac30a3f09cf4bc2de7 100644 --- a/tensorflow/compiler/tests/nullary_ops_test.py +++ b/tensorflow/compiler/tests/nullary_ops_test.py @@ -20,13 +20,13 @@ from __future__ import print_function import numpy as np -from tensorflow.compiler.tests.xla_test import XLATestCase +from tensorflow.compiler.tests import xla_test from tensorflow.python.framework import constant_op from tensorflow.python.ops import control_flow_ops from tensorflow.python.platform import googletest -class NullaryOpsTest(XLATestCase): +class NullaryOpsTest(xla_test.XLATestCase): def _testNullary(self, op, expected): with self.test_session() as session: diff --git a/tensorflow/compiler/tests/placeholder_test.py b/tensorflow/compiler/tests/placeholder_test.py index 5e6d1313bd0336eba71fcf3658d949bd3342ae11..a75d99189b5b673261c9e48f1c5998ea0c575594 100644 --- a/tensorflow/compiler/tests/placeholder_test.py +++ b/tensorflow/compiler/tests/placeholder_test.py @@ -18,14 +18,14 @@ from __future__ import absolute_import from __future__ import division from __future__ import print_function -from tensorflow.compiler.tests.xla_test import XLATestCase +from tensorflow.compiler.tests import xla_test from tensorflow.python.ops import array_ops from tensorflow.python.ops import resource_variable_ops from tensorflow.python.ops import variables from tensorflow.python.platform import googletest -class PlaceholderTest(XLATestCase): +class PlaceholderTest(xla_test.XLATestCase): def test_placeholder_with_default_default(self): with self.test_session() as sess, self.test_scope(): diff --git a/tensorflow/compiler/tests/pooling_ops_3d_test.py b/tensorflow/compiler/tests/pooling_ops_3d_test.py index d9285186baa9007e485ab916e573ad0de5e26e56..17f860db61aeda98326a6820771d67ee948b6dda 100644 --- a/tensorflow/compiler/tests/pooling_ops_3d_test.py +++ b/tensorflow/compiler/tests/pooling_ops_3d_test.py @@ -20,7 +20,7 @@ from __future__ import print_function import numpy as np -from tensorflow.compiler.tests.xla_test import XLATestCase +from tensorflow.compiler.tests import xla_test from tensorflow.python.framework import dtypes from tensorflow.python.framework import ops from tensorflow.python.ops import array_ops @@ -41,7 +41,7 @@ def _AvgPoolGrad(inputs, outputs, output_gradients, ksize, strides, padding): padding=padding) -class Pooling3DTest(XLATestCase): +class Pooling3DTest(xla_test.XLATestCase): def _VerifyValues(self, pool_func, input_sizes, window, strides, padding, expected): diff --git a/tensorflow/compiler/tests/pooling_ops_test.py b/tensorflow/compiler/tests/pooling_ops_test.py index fe270af3d636c0824621f36360ce9e7d14d8fc91..9fc94752ea660f7fb8b2c792180f01485ad04419 100644 --- a/tensorflow/compiler/tests/pooling_ops_test.py +++ b/tensorflow/compiler/tests/pooling_ops_test.py @@ -20,7 +20,7 @@ from __future__ import print_function import numpy as np -from tensorflow.compiler.tests.xla_test import XLATestCase +from tensorflow.compiler.tests import xla_test from tensorflow.python.framework import dtypes from tensorflow.python.framework import ops from tensorflow.python.ops import array_ops @@ -69,7 +69,7 @@ def GetTestConfigs(): return test_configs -class PoolingTest(XLATestCase): +class PoolingTest(xla_test.XLATestCase): def _VerifyOneTest(self, pool_func, input_sizes, ksize, strides, padding, data_format, expected): @@ -288,7 +288,7 @@ class PoolingTest(XLATestCase): expected=expected_output) -class PoolGradTest(XLATestCase): +class PoolGradTest(xla_test.XLATestCase): CPU_DEVICE = "/job:localhost/replica:0/task:0/cpu:0" diff --git a/tensorflow/compiler/tests/powersign_test.py b/tensorflow/compiler/tests/powersign_test.py new file mode 100644 index 0000000000000000000000000000000000000000..5fa7706d7294f2cffb7d24a56851be02d759335a --- /dev/null +++ b/tensorflow/compiler/tests/powersign_test.py @@ -0,0 +1,142 @@ +# Copyright 2017 The TensorFlow Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================== +"""Tests for PowerSign.""" + +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + +import math +import numpy as np + +from tensorflow.compiler.tests import xla_test +from tensorflow.contrib.opt.python.training import powersign +from tensorflow.contrib.opt.python.training import sign_decay +from tensorflow.python.framework import constant_op +from tensorflow.python.ops import resource_variable_ops +from tensorflow.python.ops import variables +from tensorflow.python.platform import test + + +def py_linear_decay_fn(decay_steps): + def linear_decay(step): + step = min(step, decay_steps) + return float(decay_steps - step) / decay_steps + return linear_decay + + +def powersign_update_numpy(params, + g_t, + m, + lr, + base=math.e, + beta=0.9, + py_sign_decay_fn=None, + t=None): + m_t = beta * m + (1 - beta) * g_t + if py_sign_decay_fn is None: + sign_decayed = 1.0 + else: + sign_decayed = py_sign_decay_fn(t-1) + multiplier = base ** (sign_decayed * np.sign(g_t) * np.sign(m_t)) + params_t = params - lr * multiplier * g_t + return params_t, m_t + + +class PowerSignTest(xla_test.XLATestCase): + + def _testDense(self, + learning_rate=0.1, + sign_decay_fn=None, + py_sign_decay_fn=None, + base=math.e, + beta=0.9): + for dtype in self.float_types: + with self.test_session(), self.test_scope(): + # Initialize variables for numpy implementation. + m0, m1 = 0.0, 0.0 + var0_np = np.array([1.0, 2.0], dtype=dtype) + grads0_np = np.array([0.1, 0.1], dtype=dtype) + var1_np = np.array([3.0, 4.0], dtype=dtype) + grads1_np = np.array([0.01, 0.01], dtype=dtype) + + var0 = resource_variable_ops.ResourceVariable(var0_np) + var1 = resource_variable_ops.ResourceVariable(var1_np) + global_step = resource_variable_ops.ResourceVariable(0, trainable=False) + grads0 = constant_op.constant(grads0_np) + grads1 = constant_op.constant(grads1_np) + + opt = powersign.PowerSignOptimizer( + learning_rate=learning_rate, + base=base, + beta=beta, + sign_decay_fn=sign_decay_fn, + ) + update = opt.apply_gradients(zip([grads0, grads1], [var0, var1]), + global_step=global_step) + neg_update = opt.apply_gradients(zip([-grads0, -grads1], [var0, var1]), + global_step=global_step) + + variables.global_variables_initializer().run() + # Fetch params to validate initial values + self.assertAllClose([1.0, 2.0], var0.eval()) + self.assertAllClose([3.0, 4.0], var1.eval()) + + # Run 7 steps of powersign + # first 4 steps with positive gradient + # last 3 steps with negative gradient (sign(gm) should be -1) + for t in range(1, 8): + if t < 5: + update.run() + else: + neg_update.run() + + var0_np, m0 = powersign_update_numpy( + var0_np, + grads0_np if t < 5 else -grads0_np, + m0, + learning_rate, + base=base, + beta=beta, + py_sign_decay_fn=py_sign_decay_fn, + t=t, + ) + var1_np, m1 = powersign_update_numpy( + var1_np, + grads1_np if t < 5 else -grads1_np, + m1, + learning_rate, + base=base, + beta=beta, + py_sign_decay_fn=py_sign_decay_fn, + t=t, + ) + + # Validate updated params + self.assertAllCloseAccordingToType(var0_np, var0.eval()) + self.assertAllCloseAccordingToType(var1_np, var1.eval()) + + def testDense(self): + decay_steps = 10 + sign_decay_fn = sign_decay.get_linear_decay_fn(decay_steps) + py_sign_decay_fn = py_linear_decay_fn(decay_steps) + self._testDense() + self._testDense(learning_rate=0.1, base=10.0, beta=0.8) + self._testDense( + sign_decay_fn=sign_decay_fn, py_sign_decay_fn=py_sign_decay_fn) + + +if __name__ == '__main__': + test.main() diff --git a/tensorflow/compiler/tests/proximal_adagrad_test.py b/tensorflow/compiler/tests/proximal_adagrad_test.py new file mode 100644 index 0000000000000000000000000000000000000000..cde87db63dbfd7c8d823c6fd0e41eee8b23735bb --- /dev/null +++ b/tensorflow/compiler/tests/proximal_adagrad_test.py @@ -0,0 +1,172 @@ +# Copyright 2015 The TensorFlow Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================== +"""Tests for Proximal Adagrad optimizer.""" + +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + +import numpy as np + +from tensorflow.compiler.tests import xla_test +from tensorflow.python.framework import constant_op +from tensorflow.python.ops import resource_variable_ops +from tensorflow.python.ops import variables +from tensorflow.python.platform import test +from tensorflow.python.training import adagrad +from tensorflow.python.training import proximal_adagrad + + +class ProximalAdagradOptimizerTest(xla_test.XLATestCase): + + def testResourceProximalAdagradwithoutRegularization(self): + with self.test_session(), self.test_scope(): + var0 = resource_variable_ops.ResourceVariable([0.0, 0.0]) + var1 = resource_variable_ops.ResourceVariable([0.0, 0.0]) + grads0 = constant_op.constant([0.1, 0.2]) + grads1 = constant_op.constant([0.01, 0.02]) + opt = proximal_adagrad.ProximalAdagradOptimizer( + 3.0, + initial_accumulator_value=0.1, + l1_regularization_strength=0.0, + l2_regularization_strength=0.0) + update = opt.apply_gradients(zip([grads0, grads1], [var0, var1])) + variables.global_variables_initializer().run() + + self.assertAllClose([0.0, 0.0], var0.eval()) + self.assertAllClose([0.0, 0.0], var1.eval()) + + # Run 3 steps Proximal Adagrad. + for _ in range(3): + update.run() + + self.assertAllClose(np.array([-2.60260963, -4.29698515]), var0.eval()) + self.assertAllClose(np.array([-0.28432083, -0.56694895]), var1.eval()) + opt_vars = opt.variables() + self.assertStartsWith(opt_vars[0].name, var0._shared_name) + self.assertStartsWith(opt_vars[1].name, var1._shared_name) + self.assertEqual(2, len(opt_vars)) + + def testProximalAdagradwithoutRegularization2(self): + with self.test_session(), self.test_scope(): + var0 = resource_variable_ops.ResourceVariable([1.0, 2.0]) + var1 = resource_variable_ops.ResourceVariable([4.0, 3.0]) + grads0 = constant_op.constant([0.1, 0.2]) + grads1 = constant_op.constant([0.01, 0.02]) + + opt = proximal_adagrad.ProximalAdagradOptimizer( + 3.0, + initial_accumulator_value=0.1, + l1_regularization_strength=0.0, + l2_regularization_strength=0.0) + update = opt.apply_gradients(zip([grads0, grads1], [var0, var1])) + variables.global_variables_initializer().run() + + self.assertAllClose([1.0, 2.0], var0.eval()) + self.assertAllClose([4.0, 3.0], var1.eval()) + + # Run 3 steps Proximal Adagrad. + for _ in range(3): + update.run() + self.assertAllClose(np.array([-1.60261, -2.296985]), var0.eval()) + self.assertAllClose(np.array([3.715679, 2.433051]), var1.eval()) + + def testProximalAdagradWithL1(self): + with self.test_session(), self.test_scope(): + var0 = resource_variable_ops.ResourceVariable([1.0, 2.0]) + var1 = resource_variable_ops.ResourceVariable([4.0, 3.0]) + grads0 = constant_op.constant([0.1, 0.2]) + grads1 = constant_op.constant([0.01, 0.02]) + + opt = proximal_adagrad.ProximalAdagradOptimizer( + 3.0, + initial_accumulator_value=0.1, + l1_regularization_strength=0.001, + l2_regularization_strength=0.0) + update = opt.apply_gradients(zip([grads0, grads1], [var0, var1])) + variables.global_variables_initializer().run() + + self.assertAllClose([1.0, 2.0], var0.eval()) + self.assertAllClose([4.0, 3.0], var1.eval()) + + # Run 10 steps Proximal Adagrad + for _ in range(10): + update.run() + self.assertAllClose(np.array([-6.663634, -9.190331]), var0.eval()) + self.assertAllClose(np.array([2.959304, 1.029232]), var1.eval()) + + def testProximalAdagradWithL1_L2(self): + with self.test_session(), self.test_scope(): + var0 = resource_variable_ops.ResourceVariable([1.0, 2.0]) + var1 = resource_variable_ops.ResourceVariable([4.0, 3.0]) + grads0 = constant_op.constant([0.1, 0.2]) + grads1 = constant_op.constant([0.01, 0.02]) + + opt = proximal_adagrad.ProximalAdagradOptimizer( + 3.0, + initial_accumulator_value=0.1, + l1_regularization_strength=0.001, + l2_regularization_strength=2.0) + update = opt.apply_gradients(zip([grads0, grads1], [var0, var1])) + variables.global_variables_initializer().run() + + self.assertAllClose([1.0, 2.0], var0.eval()) + self.assertAllClose([4.0, 3.0], var1.eval()) + + # Run 10 steps Proximal Adagrad. + for _ in range(10): + update.run() + + self.assertAllClose(np.array([-0.0495, -0.0995]), var0.eval()) + self.assertAllClose(np.array([-0.0045, -0.0095]), var1.eval()) + + def applyOptimizer(self, opt, steps=5): + var0 = resource_variable_ops.ResourceVariable([1.0, 2.0]) + var1 = resource_variable_ops.ResourceVariable([3.0, 4.0]) + grads0 = constant_op.constant([0.1, 0.2]) + grads1 = constant_op.constant([0.01, 0.02]) + + update = opt.apply_gradients(zip([grads0, grads1], [var0, var1])) + variables.global_variables_initializer().run() + + self.assertAllClose([1.0, 2.0], var0.eval()) + self.assertAllClose([3.0, 4.0], var1.eval()) + + # Run ProximalAdagrad for a few steps + for _ in range(steps): + update.run() + + return var0.eval(), var1.eval() + + def testEquivAdagradwithoutRegularization(self): + with self.test_session(), self.test_scope(): + val0, val1 = self.applyOptimizer( + proximal_adagrad.ProximalAdagradOptimizer( + 3.0, + initial_accumulator_value=0.1, + l1_regularization_strength=0.0, + l2_regularization_strength=0.0)) + + with self.test_session(), self.test_scope(): + val2, val3 = self.applyOptimizer( + adagrad.AdagradOptimizer( + 3.0, initial_accumulator_value=0.1)) + + self.assertAllClose(val0, val2) + self.assertAllClose(val1, val3) + + +if __name__ == "__main__": + test.main() diff --git a/tensorflow/compiler/tests/proximal_gradient_descent_test.py b/tensorflow/compiler/tests/proximal_gradient_descent_test.py new file mode 100644 index 0000000000000000000000000000000000000000..11eb76871133eba8fcd24621afb03e16614fb005 --- /dev/null +++ b/tensorflow/compiler/tests/proximal_gradient_descent_test.py @@ -0,0 +1,156 @@ +# Copyright 2015 The TensorFlow Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================== +"""Tests for Proximal Gradient Descent optimizer.""" + +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + +import numpy as np + +from tensorflow.compiler.tests import xla_test +from tensorflow.python.framework import constant_op +from tensorflow.python.ops import resource_variable_ops +from tensorflow.python.ops import variables +from tensorflow.python.platform import test +from tensorflow.python.training import gradient_descent +from tensorflow.python.training import proximal_gradient_descent + + +class ProximalGradientDescentOptimizerTest(xla_test.XLATestCase): + + def testResourceProximalGradientDescentwithoutRegularization(self): + with self.test_session(), self.test_scope(): + var0 = resource_variable_ops.ResourceVariable([0.0, 0.0]) + var1 = resource_variable_ops.ResourceVariable([0.0, 0.0]) + grads0 = constant_op.constant([0.1, 0.2]) + grads1 = constant_op.constant([0.01, 0.02]) + opt = proximal_gradient_descent.ProximalGradientDescentOptimizer( + 3.0, l1_regularization_strength=0.0, l2_regularization_strength=0.0) + update = opt.apply_gradients(zip([grads0, grads1], [var0, var1])) + variables.global_variables_initializer().run() + + self.assertAllClose([0.0, 0.0], var0.eval()) + self.assertAllClose([0.0, 0.0], var1.eval()) + + # Run 3 steps Proximal Gradient Descent. + for _ in range(3): + update.run() + + self.assertAllClose(np.array([-0.9, -1.8]), var0.eval()) + self.assertAllClose(np.array([-0.09, -0.18]), var1.eval()) + + def testProximalGradientDescentwithoutRegularization2(self): + with self.test_session(), self.test_scope(): + var0 = resource_variable_ops.ResourceVariable([1.0, 2.0]) + var1 = resource_variable_ops.ResourceVariable([4.0, 3.0]) + grads0 = constant_op.constant([0.1, 0.2]) + grads1 = constant_op.constant([0.01, 0.02]) + + opt = proximal_gradient_descent.ProximalGradientDescentOptimizer( + 3.0, l1_regularization_strength=0.0, l2_regularization_strength=0.0) + update = opt.apply_gradients(zip([grads0, grads1], [var0, var1])) + variables.global_variables_initializer().run() + + self.assertAllClose([1.0, 2.0], var0.eval()) + self.assertAllClose([4.0, 3.0], var1.eval()) + + # Run 3 steps Proximal Gradient Descent + for _ in range(3): + update.run() + + self.assertAllClose(np.array([0.1, 0.2]), var0.eval()) + self.assertAllClose(np.array([3.91, 2.82]), var1.eval()) + + def testProximalGradientDescentWithL1(self): + with self.test_session(), self.test_scope(): + var0 = resource_variable_ops.ResourceVariable([1.0, 2.0]) + var1 = resource_variable_ops.ResourceVariable([4.0, 3.0]) + grads0 = constant_op.constant([0.1, 0.2]) + grads1 = constant_op.constant([0.01, 0.02]) + + opt = proximal_gradient_descent.ProximalGradientDescentOptimizer( + 3.0, l1_regularization_strength=0.001, l2_regularization_strength=0.0) + update = opt.apply_gradients(zip([grads0, grads1], [var0, var1])) + variables.global_variables_initializer().run() + + self.assertAllClose([1.0, 2.0], var0.eval()) + self.assertAllClose([4.0, 3.0], var1.eval()) + + # Run 10 steps proximal gradient descent. + for _ in range(10): + update.run() + + self.assertAllClose(np.array([-1.988, -3.988001]), var0.eval()) + self.assertAllClose(np.array([3.67, 2.37]), var1.eval()) + + def testProximalGradientDescentWithL1_L2(self): + with self.test_session(), self.test_scope(): + var0 = resource_variable_ops.ResourceVariable([1.0, 2.0]) + var1 = resource_variable_ops.ResourceVariable([4.0, 3.0]) + grads0 = constant_op.constant([0.1, 0.2]) + grads1 = constant_op.constant([0.01, 0.02]) + + opt = proximal_gradient_descent.ProximalGradientDescentOptimizer( + 3.0, l1_regularization_strength=0.001, l2_regularization_strength=2.0) + update = opt.apply_gradients(zip([grads0, grads1], [var0, var1])) + variables.global_variables_initializer().run() + + self.assertAllClose([1.0, 2.0], var0.eval()) + self.assertAllClose([4.0, 3.0], var1.eval()) + + # Run 10 steps Proximal Gradient Descent + for _ in range(10): + update.run() + + self.assertAllClose(np.array([-0.0495, -0.0995]), var0.eval()) + self.assertAllClose(np.array([-0.0045, -0.0095]), var1.eval()) + + def applyOptimizer(self, opt, steps=5): + var0 = resource_variable_ops.ResourceVariable([1.0, 2.0]) + var1 = resource_variable_ops.ResourceVariable([3.0, 4.0]) + grads0 = constant_op.constant([0.1, 0.2]) + grads1 = constant_op.constant([0.01, 0.02]) + + update = opt.apply_gradients(zip([grads0, grads1], [var0, var1])) + variables.global_variables_initializer().run() + + self.assertAllClose([1.0, 2.0], var0.eval()) + self.assertAllClose([3.0, 4.0], var1.eval()) + + # Run ProximalAdagrad for a few steps + for _ in range(steps): + update.run() + + return var0.eval(), var1.eval() + + def testEquivGradientDescentwithoutRegularization(self): + with self.test_session(), self.test_scope(): + val0, val1 = self.applyOptimizer( + proximal_gradient_descent.ProximalGradientDescentOptimizer( + 3.0, + l1_regularization_strength=0.0, + l2_regularization_strength=0.0)) + + with self.test_session(), self.test_scope(): + val2, val3 = self.applyOptimizer( + gradient_descent.GradientDescentOptimizer(3.0)) + + self.assertAllClose(val0, val2) + self.assertAllClose(val1, val3) + + +if __name__ == "__main__": + test.main() diff --git a/tensorflow/compiler/tests/qr_op_test.py b/tensorflow/compiler/tests/qr_op_test.py new file mode 100644 index 0000000000000000000000000000000000000000..93752a21db4ab9850348e313f29ba7b43b788712 --- /dev/null +++ b/tensorflow/compiler/tests/qr_op_test.py @@ -0,0 +1,112 @@ +# Copyright 2018 The TensorFlow Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================== +"""Tests for tensorflow.ops.math_ops.matrix_inverse.""" + +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + +import itertools + +from absl.testing import parameterized +import numpy as np + +from tensorflow.compiler.tests import xla_test +from tensorflow.python.ops import array_ops +from tensorflow.python.ops import linalg_ops +from tensorflow.python.ops import math_ops +from tensorflow.python.platform import test + + +class QrOpTest(xla_test.XLATestCase, parameterized.TestCase): + + def AdjustedNorm(self, x): + """Computes the norm of matrices in 'x', adjusted for dimension and type.""" + norm = np.linalg.norm(x, axis=(-2, -1)) + return norm / (max(x.shape[-2:]) * np.finfo(x.dtype).eps) + + def CompareOrthogonal(self, x, y, rank): + # We only compare the first 'rank' orthogonal vectors since the + # remainder form an arbitrary orthonormal basis for the + # (row- or column-) null space, whose exact value depends on + # implementation details. Notice that since we check that the + # matrices of singular vectors are unitary elsewhere, we do + # implicitly test that the trailing vectors of x and y span the + # same space. + x = x[..., 0:rank] + y = y[..., 0:rank] + # Q is only unique up to sign (complex phase factor for complex matrices), + # so we normalize the sign first. + sum_of_ratios = np.sum(np.divide(y, x), -2, keepdims=True) + phases = np.divide(sum_of_ratios, np.abs(sum_of_ratios)) + x *= phases + self.assertTrue(np.all(self.AdjustedNorm(x - y) < 30.0)) + + def CheckApproximation(self, a, q, r): + # Tests that a ~= q*r. + precision = self.AdjustedNorm(a - np.matmul(q, r)) + self.assertTrue(np.all(precision < 5.0)) + + def CheckUnitary(self, x): + # Tests that x[...,:,:]^H * x[...,:,:] is close to the identity. + xx = math_ops.matmul(x, x, adjoint_a=True) + identity = array_ops.matrix_band_part(array_ops.ones_like(xx), 0, 0) + precision = self.AdjustedNorm(xx.eval() - identity.eval()) + self.assertTrue(np.all(precision < 5.0)) + + def _test(self, dtype, shape, full_matrices): + np.random.seed(1) + x_np = np.random.uniform( + low=-1.0, high=1.0, size=np.prod(shape)).reshape(shape).astype(dtype) + + with self.test_session() as sess: + x_tf = array_ops.placeholder(dtype) + with self.test_scope(): + q_tf, r_tf = linalg_ops.qr(x_tf, full_matrices=full_matrices) + q_tf_val, r_tf_val = sess.run([q_tf, r_tf], feed_dict={x_tf: x_np}) + + q_dims = q_tf_val.shape + np_q = np.ndarray(q_dims, dtype) + np_q_reshape = np.reshape(np_q, (-1, q_dims[-2], q_dims[-1])) + new_first_dim = np_q_reshape.shape[0] + + x_reshape = np.reshape(x_np, (-1, x_np.shape[-2], x_np.shape[-1])) + for i in range(new_first_dim): + if full_matrices: + np_q_reshape[i, :, :], _ = np.linalg.qr( + x_reshape[i, :, :], mode="complete") + else: + np_q_reshape[i, :, :], _ = np.linalg.qr( + x_reshape[i, :, :], mode="reduced") + np_q = np.reshape(np_q_reshape, q_dims) + self.CompareOrthogonal(np_q, q_tf_val, min(shape[-2:])) + self.CheckApproximation(x_np, q_tf_val, r_tf_val) + self.CheckUnitary(q_tf_val) + + SIZES = [1, 2, 5, 10, 32, 100, 300] + DTYPES = [np.float32] + PARAMS = itertools.product(SIZES, SIZES, DTYPES) + + @parameterized.parameters(*PARAMS) + def testQR(self, rows, cols, dtype): + # TODO(b/111317468): implement full_matrices=False, test other types. + for full_matrices in [True]: + # Only tests the (3, 2) case for small numbers of rows/columns. + for batch_dims in [(), (3,)] + [(3, 2)] * (max(rows, cols) < 10): + self._test(dtype, batch_dims + (rows, cols), full_matrices) + + +if __name__ == "__main__": + test.main() diff --git a/tensorflow/compiler/tests/random_ops_test.py b/tensorflow/compiler/tests/random_ops_test.py index 2e71b00ba66dba93c87e565e3a372111de1f362d..14c5e7a975e478ca6ceed37c28339b40612801c8 100644 --- a/tensorflow/compiler/tests/random_ops_test.py +++ b/tensorflow/compiler/tests/random_ops_test.py @@ -22,7 +22,7 @@ import math import numpy as np -from tensorflow.compiler.tests.xla_test import XLATestCase +from tensorflow.compiler.tests import xla_test from tensorflow.python.framework import dtypes from tensorflow.python.ops import array_ops from tensorflow.python.ops import math_ops @@ -31,7 +31,7 @@ from tensorflow.python.ops.distributions import special_math from tensorflow.python.platform import googletest -class RandomOpsTest(XLATestCase): +class RandomOpsTest(xla_test.XLATestCase): """Test cases for random-number generating operators.""" def _random_types(self): @@ -140,10 +140,10 @@ class RandomOpsTest(XLATestCase): def testShuffle1d(self): with self.test_session() as sess: with self.test_scope(): - x = math_ops.range(20) + x = math_ops.range(1 << 16) shuffle = random_ops.random_shuffle(x) result = sess.run(shuffle) - expected = range(20) + expected = range(1 << 16) # Compare sets to avoid randomness behavior changes but make sure still # have all the values. self.assertAllEqual(set(result), set(expected)) diff --git a/tensorflow/compiler/tests/reduce_ops_test.py b/tensorflow/compiler/tests/reduce_ops_test.py index 7420724bdbeab63b39542ada59328621febad895..cea2ec816f85e88b11e6e80c91c14fca9015f45c 100644 --- a/tensorflow/compiler/tests/reduce_ops_test.py +++ b/tensorflow/compiler/tests/reduce_ops_test.py @@ -22,7 +22,7 @@ import functools import itertools import numpy as np -from tensorflow.compiler.tests.xla_test import XLATestCase +from tensorflow.compiler.tests import xla_test from tensorflow.python.framework import dtypes from tensorflow.python.framework import errors_impl from tensorflow.python.ops import array_ops @@ -30,7 +30,7 @@ from tensorflow.python.ops import math_ops from tensorflow.python.platform import googletest -class ReduceOpsTest(XLATestCase): +class ReduceOpsTest(xla_test.XLATestCase): def _testReduction(self, tf_reduce_fn, @@ -156,7 +156,7 @@ class ReduceOpsTest(XLATestCase): self._testReduction(math_ops.reduce_any, np.any, np.bool, self.BOOL_DATA) -class ReduceOpPrecisionTest(XLATestCase): +class ReduceOpPrecisionTest(xla_test.XLATestCase): def _testReduceSum(self, expected_result, diff --git a/tensorflow/compiler/tests/reduce_window_test.py b/tensorflow/compiler/tests/reduce_window_test.py index e78a63465b80644d8810d9fa7433653bc4639fed..c69b6837b0f88ced844faf3713a29a1c14c8790d 100644 --- a/tensorflow/compiler/tests/reduce_window_test.py +++ b/tensorflow/compiler/tests/reduce_window_test.py @@ -20,7 +20,7 @@ from __future__ import print_function import numpy as np -from tensorflow.compiler.tests.xla_test import XLATestCase +from tensorflow.compiler.tests import xla_test from tensorflow.compiler.tf2xla.python import xla from tensorflow.python.framework import dtypes from tensorflow.python.framework import function @@ -28,7 +28,7 @@ from tensorflow.python.ops import array_ops from tensorflow.python.platform import googletest -class ReduceWindowTest(XLATestCase): +class ReduceWindowTest(xla_test.XLATestCase): """Test cases for xla.reduce_window.""" def _reduce_window(self, operand, init, reducer, **kwargs): diff --git a/tensorflow/compiler/tests/reverse_ops_test.py b/tensorflow/compiler/tests/reverse_ops_test.py index 18fabca28c9817fc8517595fa1694a18399f54b0..d01c676e7c2fe705344f26818350c46c30451c67 100644 --- a/tensorflow/compiler/tests/reverse_ops_test.py +++ b/tensorflow/compiler/tests/reverse_ops_test.py @@ -21,14 +21,14 @@ from __future__ import print_function import itertools import numpy as np -from tensorflow.compiler.tests.xla_test import XLATestCase +from tensorflow.compiler.tests import xla_test from tensorflow.python.framework import constant_op from tensorflow.python.framework import dtypes from tensorflow.python.ops import array_ops from tensorflow.python.platform import googletest -class ReverseOpsTest(XLATestCase): +class ReverseOpsTest(xla_test.XLATestCase): def testReverseOneDim(self): shape = (7, 5, 9, 11) diff --git a/tensorflow/compiler/tests/reverse_sequence_op_test.py b/tensorflow/compiler/tests/reverse_sequence_op_test.py index 1a5d05094e53cfecd9476d7d87f023e8a02d7458..ccfa63001653537c4d1b7140e3d745c126f9034b 100644 --- a/tensorflow/compiler/tests/reverse_sequence_op_test.py +++ b/tensorflow/compiler/tests/reverse_sequence_op_test.py @@ -20,13 +20,13 @@ from __future__ import print_function import numpy as np -from tensorflow.compiler.tests.xla_test import XLATestCase +from tensorflow.compiler.tests import xla_test from tensorflow.python.framework import dtypes from tensorflow.python.ops import array_ops from tensorflow.python.platform import test -class ReverseSequenceTest(XLATestCase): +class ReverseSequenceTest(xla_test.XLATestCase): def _testReverseSequence(self, x, diff --git a/tensorflow/compiler/tests/rmsprop_test.py b/tensorflow/compiler/tests/rmsprop_test.py index ecdce4f052bbe3eeae8697c02c891105103f4f69..ff8bbac911abe73f946464663984ff1626302882 100644 --- a/tensorflow/compiler/tests/rmsprop_test.py +++ b/tensorflow/compiler/tests/rmsprop_test.py @@ -20,7 +20,7 @@ from __future__ import print_function import numpy as np -from tensorflow.compiler.tests.xla_test import XLATestCase +from tensorflow.compiler.tests import xla_test from tensorflow.python.framework import constant_op from tensorflow.python.ops import resource_variable_ops from tensorflow.python.ops import variables @@ -28,33 +28,104 @@ from tensorflow.python.platform import test from tensorflow.python.training import rmsprop -class RmspropTest(XLATestCase): +class RmspropTest(xla_test.XLATestCase): + + def _rmsprop_update_numpy(self, + var, + g, + mg, + rms, + mom, + lr, + decay=0.9, + momentum=0.0, + epsilon=1e-10, + centered=False): + rms_t = rms * decay + (1 - decay) * g * g + denom_t = rms_t + epsilon + if centered: + mg_t = mg * decay + (1 - decay) * g + denom_t -= mg_t * mg_t + else: + mg_t = mg + mom_t = momentum * mom + lr * g / np.sqrt(denom_t, dtype=denom_t.dtype) + var_t = var - mom_t + return var_t, mg_t, rms_t, mom_t def testBasic(self): for dtype in self.float_types: - with self.test_session(), self.test_scope(): - var0 = resource_variable_ops.ResourceVariable([1.0, 2.0], dtype=dtype) - var1 = resource_variable_ops.ResourceVariable([3.0, 4.0], dtype=dtype) - grads0 = constant_op.constant([0.1, 0.1], dtype=dtype) - grads1 = constant_op.constant([0.01, 0.01], dtype=dtype) - rms_opt = rmsprop.RMSPropOptimizer(3.0) - rms_update = rms_opt.apply_gradients( - zip([grads0, grads1], [var0, var1])) - variables.global_variables_initializer().run() - - # Fetch params to validate initial values - self.assertAllClose([1.0, 2.0], var0.eval()) - self.assertAllClose([3.0, 4.0], var1.eval()) - - # Run 3 steps of RMSProp - for _ in range(3): - rms_update.run() - - # Validate updated params - self.assertAllCloseAccordingToType( - np.array([2.91705132e-04, 1.00029182e+00]), var0.eval()) - self.assertAllCloseAccordingToType( - np.array([2.89990854, 3.89990854]), var1.eval()) + for centered in [False, True]: + with self.test_session(), self.test_scope(): + # Initialize variables for numpy implementation. + var0_np = np.array([1.0, 2.0], dtype=dtype) + grads0_np = np.array([0.1, 0.1], dtype=dtype) + var1_np = np.array([3.0, 4.0], dtype=dtype) + grads1_np = np.array([0.01, 0.01], dtype=dtype) + mg0_np = np.array([0.0, 0.0], dtype=dtype) + mg1_np = np.array([0.0, 0.0], dtype=dtype) + rms0_np = np.array([1.0, 1.0], dtype=dtype) + rms1_np = np.array([1.0, 1.0], dtype=dtype) + mom0_np = np.array([0.0, 0.0], dtype=dtype) + mom1_np = np.array([0.0, 0.0], dtype=dtype) + + var0 = resource_variable_ops.ResourceVariable(var0_np) + var1 = resource_variable_ops.ResourceVariable(var1_np) + grads0 = constant_op.constant(grads0_np) + grads1 = constant_op.constant(grads1_np) + learning_rate = 3.0 + rms_opt = rmsprop.RMSPropOptimizer(learning_rate, centered=centered) + rms_update = rms_opt.apply_gradients( + zip([grads0, grads1], [var0, var1])) + variables.global_variables_initializer().run() + + mg0 = rms_opt.get_slot(var0, "mg") + self.assertEqual(mg0 is not None, centered) + mg1 = rms_opt.get_slot(var1, "mg") + self.assertEqual(mg1 is not None, centered) + rms0 = rms_opt.get_slot(var0, "rms") + self.assertTrue(rms0 is not None) + rms1 = rms_opt.get_slot(var1, "rms") + self.assertTrue(rms1 is not None) + mom0 = rms_opt.get_slot(var0, "momentum") + self.assertTrue(mom0 is not None) + mom1 = rms_opt.get_slot(var1, "momentum") + self.assertTrue(mom1 is not None) + + # Fetch params to validate initial values + self.assertAllClose([1.0, 2.0], var0.eval()) + self.assertAllClose([3.0, 4.0], var1.eval()) + + # Run 3 steps of RMSProp + for _ in range(3): + rms_update.run() + + var0_np, mg0_np, rms0_np, mom0_np = self._rmsprop_update_numpy( + var0_np, + grads0_np, + mg0_np, + rms0_np, + mom0_np, + learning_rate, + centered=centered) + var1_np, mg1_np, rms1_np, mom1_np = self._rmsprop_update_numpy( + var1_np, + grads1_np, + mg1_np, + rms1_np, + mom1_np, + learning_rate, + centered=centered) + + # Validate updated params + if centered: + self.assertAllCloseAccordingToType(mg0_np, mg0.eval()) + self.assertAllCloseAccordingToType(mg1_np, mg1.eval()) + self.assertAllCloseAccordingToType(rms0_np, rms0.eval()) + self.assertAllCloseAccordingToType(rms1_np, rms1.eval()) + self.assertAllCloseAccordingToType(mom0_np, mom0.eval()) + self.assertAllCloseAccordingToType(mom1_np, mom1.eval()) + self.assertAllCloseAccordingToType(var0_np, var0.eval()) + self.assertAllCloseAccordingToType(var1_np, var1.eval()) if __name__ == "__main__": diff --git a/tensorflow/compiler/tests/scan_ops_test.py b/tensorflow/compiler/tests/scan_ops_test.py index 3260e63b23226d736a7ddc0f21a94a8c791e0442..4292352e76ebcef7dbf41df7b857d2604a468117 100644 --- a/tensorflow/compiler/tests/scan_ops_test.py +++ b/tensorflow/compiler/tests/scan_ops_test.py @@ -20,7 +20,7 @@ from __future__ import print_function import numpy as np -from tensorflow.compiler.tests.xla_test import XLATestCase +from tensorflow.compiler.tests import xla_test from tensorflow.python.framework import constant_op from tensorflow.python.framework import errors_impl from tensorflow.python.framework import ops @@ -69,7 +69,7 @@ def handle_options(func, x, axis, exclusive, reverse): return x -class CumsumTest(XLATestCase): +class CumsumTest(xla_test.XLATestCase): valid_dtypes = [np.float32] @@ -147,7 +147,7 @@ class CumsumTest(XLATestCase): math_ops.cumsum(input_tensor, [0]).eval() -class CumprodTest(XLATestCase): +class CumprodTest(xla_test.XLATestCase): valid_dtypes = [np.float32] diff --git a/tensorflow/compiler/tests/scatter_nd_op_test.py b/tensorflow/compiler/tests/scatter_nd_op_test.py index 638946e234daf28dc4a34e6c33fc0f78b8e8699b..f606f88545d0b6f0b52cee9b93083a6bd91169bc 100644 --- a/tensorflow/compiler/tests/scatter_nd_op_test.py +++ b/tensorflow/compiler/tests/scatter_nd_op_test.py @@ -22,7 +22,7 @@ import functools import numpy as np -from tensorflow.compiler.tests.xla_test import XLATestCase +from tensorflow.compiler.tests import xla_test from tensorflow.python.framework import errors from tensorflow.python.ops import array_ops from tensorflow.python.platform import test @@ -68,7 +68,7 @@ def _NumpyUpdate(indices, updates, shape): return _NumpyScatterNd(ref, indices, updates, lambda p, u: u) -class ScatterNdTest(XLATestCase): +class ScatterNdTest(xla_test.XLATestCase): def _VariableRankTest(self, np_scatter, diff --git a/tensorflow/compiler/tests/segment_reduction_ops_test.py b/tensorflow/compiler/tests/segment_reduction_ops_test.py index 4a9c0e7471f9cdb2a47b54705495d2dda9748890..772c20fd424577c3e06eeae409f424b77b52aa8a 100644 --- a/tensorflow/compiler/tests/segment_reduction_ops_test.py +++ b/tensorflow/compiler/tests/segment_reduction_ops_test.py @@ -21,26 +21,40 @@ from __future__ import print_function import functools import numpy as np -from tensorflow.compiler.tests.xla_test import XLATestCase +from tensorflow.compiler.tests import xla_test +from tensorflow.python.framework import dtypes from tensorflow.python.ops import array_ops from tensorflow.python.ops import math_ops from tensorflow.python.platform import googletest -class SegmentReductionOpsTest(XLATestCase): +class SegmentReductionOpsTest(xla_test.XLATestCase): """Test cases for segment reduction ops.""" - def UnsortedSegmentSum(self, data, indices, num_segments): + def _segmentReduction(self, op, data, indices, num_segments): with self.test_session() as sess, self.test_scope(): d = array_ops.placeholder(data.dtype, shape=data.shape) if isinstance(indices, int): i = array_ops.placeholder(np.int32, shape=[]) else: i = array_ops.placeholder(indices.dtype, shape=indices.shape) - return sess.run( - math_ops.unsorted_segment_sum(d, i, num_segments), - {d: data, - i: indices}) + return sess.run(op(d, i, num_segments), {d: data, i: indices}) + + def _unsortedSegmentSum(self, data, indices, num_segments): + return self._segmentReduction(math_ops.unsorted_segment_sum, data, indices, + num_segments) + + def _unsortedSegmentProd(self, data, indices, num_segments): + return self._segmentReduction(math_ops.unsorted_segment_prod, data, indices, + num_segments) + + def _unsortedSegmentMin(self, data, indices, num_segments): + return self._segmentReduction(math_ops.unsorted_segment_min, data, indices, + num_segments) + + def _unsortedSegmentMax(self, data, indices, num_segments): + return self._segmentReduction(math_ops.unsorted_segment_max, data, indices, + num_segments) def testUnsortedSegmentSum0DIndices1DData(self): for dtype in self.numeric_types: @@ -49,14 +63,14 @@ class SegmentReductionOpsTest(XLATestCase): [[0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 0, 0], [0, 1, 2, 3, 4, 5], [0, 0, 0, 0, 0, 0]], dtype=dtype), - self.UnsortedSegmentSum( + self._unsortedSegmentSum( np.array([0, 1, 2, 3, 4, 5], dtype=dtype), 2, 4)) def testUnsortedSegmentSum1DIndices1DData(self): for dtype in self.numeric_types: self.assertAllClose( np.array([1, 3, 2, 9], dtype=dtype), - self.UnsortedSegmentSum( + self._unsortedSegmentSum( np.array([0, 1, 2, 3, 4, 5], dtype=dtype), np.array([3, 0, 2, 1, 3, 3], dtype=np.int32), 4)) @@ -64,7 +78,7 @@ class SegmentReductionOpsTest(XLATestCase): for dtype in self.numeric_types: self.assertAllClose( np.array([6, 3, 0, 6], dtype=dtype), - self.UnsortedSegmentSum( + self._unsortedSegmentSum( np.array([0, 1, 2, 3, 4, 5, 6], dtype=dtype), np.array([3, -1, 0, 1, 0, -1, 3], dtype=np.int32), 4)) @@ -76,7 +90,7 @@ class SegmentReductionOpsTest(XLATestCase): dtype=dtype) indices = np.array([8, 1, 0, 3, 7], dtype=np.int32) num_segments = 10 - y = self.UnsortedSegmentSum(data, indices, num_segments) + y = self._unsortedSegmentSum(data, indices, num_segments) self.assertAllClose( np.array( [[30, 31, 32, 33], [20, 21, 22, 23], [0, 0, 0, 0], @@ -92,7 +106,7 @@ class SegmentReductionOpsTest(XLATestCase): dtype=dtype) indices = np.array([0, 1, 2, 0, 1], dtype=np.int32) num_segments = 4 - y = self.UnsortedSegmentSum(data, indices, num_segments) + y = self._unsortedSegmentSum(data, indices, num_segments) self.assertAllClose( np.array( [[40, 42, 44, 46], [70, 72, 74, 76], [30, 31, 32, 33], @@ -102,30 +116,30 @@ class SegmentReductionOpsTest(XLATestCase): def testUnsortedSegmentSum2DIndices3DData(self): for dtype in self.numeric_types: data = np.array( - [[[0, 1, 2], [10, 11, 12]], [[100, 101, 102], [110, 111, 112]], - [[200, 201, 202], [210, 211, 212]], [[300, 301, 302], - [310, 311, 312]]], + [[[0, 1, 2], [10, 11, 12]], [[100, 101, 102], [110, 111, 112]], [[ + 200, 201, 202 + ], [210, 211, 212]], [[300, 301, 302], [310, 311, 312]]], dtype=dtype) indices = np.array([[3, 5], [3, 1], [5, 0], [6, 2]], dtype=np.int32) num_segments = 8 - y = self.UnsortedSegmentSum(data, indices, num_segments) + y = self._unsortedSegmentSum(data, indices, num_segments) self.assertAllClose( np.array( - [[210, 211, 212], [110, 111, 112], [310, 311, 312], - [100, 102, 104], [0, 0, 0.], [210, 212, 214], [300, 301, - 302], [0, 0, 0]], + [[210, 211, 212], [110, 111, 112], [310, 311, 312], [ + 100, 102, 104 + ], [0, 0, 0.], [210, 212, 214], [300, 301, 302], [0, 0, 0]], dtype=dtype), y) def testUnsortedSegmentSum1DIndices3DData(self): for dtype in self.numeric_types: data = np.array( - [[[0, 1, 2], [10, 11, 12]], [[100, 101, 102], [110, 111, 112]], - [[200, 201, 202], [210, 211, 212]], [[300, 301, 302], - [310, 311, 312]]], + [[[0, 1, 2], [10, 11, 12]], [[100, 101, 102], [110, 111, 112]], [[ + 200, 201, 202 + ], [210, 211, 212]], [[300, 301, 302], [310, 311, 312]]], dtype=dtype) indices = np.array([3, 0, 2, 5], dtype=np.int32) num_segments = 6 - y = self.UnsortedSegmentSum(data, indices, num_segments) + y = self._unsortedSegmentSum(data, indices, num_segments) self.assertAllClose( np.array( [[[100, 101, 102.], [110, 111, 112]], [[0, 0, 0], [0, 0, 0]], @@ -138,10 +152,40 @@ class SegmentReductionOpsTest(XLATestCase): data = np.ones((4, 8, 7), dtype=dtype) indices = np.ones((3, 2), dtype=np.int32) num_segments = 4 - self.assertRaises(ValueError, - functools.partial(self.UnsortedSegmentSum, data, - indices, num_segments)) + self.assertRaises( + ValueError, + functools.partial(self._segmentReduction, + math_ops.unsorted_segment_sum, data, indices, + num_segments)) + + def testUnsortedSegmentOps1DIndices1DDataNegativeIndices(self): + """Tests for min, max, and prod ops. + + These share most of their implementation with sum, so we only test basic + functionality. + """ + for dtype in self.numeric_types: + self.assertAllClose( + np.array([8, 3, 1, 0], dtype=dtype), + self._unsortedSegmentProd( + np.array([0, 1, 2, 3, 4, 5, 6], dtype=dtype), + np.array([3, -1, 0, 1, 0, -1, 3], dtype=np.int32), 4)) + + for dtype in self.int_types | self.float_types: + minval = dtypes.as_dtype(dtype).min + maxval = dtypes.as_dtype(dtype).max + + self.assertAllClose( + np.array([2, 3, maxval, 0], dtype=dtype), + self._unsortedSegmentMin( + np.array([0, 1, 2, 3, 4, 5, 6], dtype=dtype), + np.array([3, -1, 0, 1, 0, -1, 3], dtype=np.int32), 4)) + self.assertAllClose( + np.array([4, 3, minval, 6], dtype=dtype), + self._unsortedSegmentMax( + np.array([0, 1, 2, 3, 4, 5, 6], dtype=dtype), + np.array([3, -1, 0, 1, 0, -1, 3], dtype=np.int32), 4)) -if __name__ == '__main__': +if __name__ == "__main__": googletest.main() diff --git a/tensorflow/compiler/tests/slice_ops_test.py b/tensorflow/compiler/tests/slice_ops_test.py index 305ca0c6b78d3ef985deb38816f9388e7983906b..6c4890565d2083a9493abc59bd563c4dd9fdb186 100644 --- a/tensorflow/compiler/tests/slice_ops_test.py +++ b/tensorflow/compiler/tests/slice_ops_test.py @@ -18,14 +18,14 @@ from __future__ import absolute_import from __future__ import division from __future__ import print_function -from tensorflow.compiler.tests.xla_test import XLATestCase +from tensorflow.compiler.tests import xla_test from tensorflow.python.framework import dtypes from tensorflow.python.framework import tensor_shape from tensorflow.python.ops import array_ops from tensorflow.python.platform import googletest -class SliceTest(XLATestCase): +class SliceTest(xla_test.XLATestCase): def test1D(self): for dtype in self.numeric_types: @@ -110,7 +110,7 @@ class SliceTest(XLATestCase): self.assertAllEqual([[[1, 1, 1, 1], [6, 5, 4, 3]]], result) -class StridedSliceTest(XLATestCase): +class StridedSliceTest(xla_test.XLATestCase): def test1D(self): for dtype in self.numeric_types: diff --git a/tensorflow/compiler/tests/sort_ops_test.py b/tensorflow/compiler/tests/sort_ops_test.py index 8ae579abda9854079ee491a7254eb4d09183594a..9e2ef964a1ff00a861a874135b7dfa1358a7020e 100644 --- a/tensorflow/compiler/tests/sort_ops_test.py +++ b/tensorflow/compiler/tests/sort_ops_test.py @@ -64,20 +64,29 @@ class XlaSortOpTest(xla_test.XLATestCase): if self.device in ["XLA_CPU", "XLA_GPU"]: return - # Only bfloat16 is implemented. - bfloat16 = dtypes.bfloat16.as_numpy_dtype - if bfloat16 in self.numeric_types: - for x in [np.arange(20)]: + supported_types = set( + [dtypes.bfloat16.as_numpy_dtype, np.float32, np.int32, np.uint32]) + for dtype in supported_types.intersection(self.numeric_types): + # Use small input size for bfloat16. Otherwise, we'll get duplicate values + # after conversion to bfloat16, so the possible resulting index array is + # no longer unique. + if dtype == dtypes.bfloat16.as_numpy_dtype: + array_size = 20 + k_options = [0, 1, 2, 10, 20] + else: + array_size = 200 * 1000 + k_options = [0, 1, 2, 10, 20, 100, 1000, 200 * 1000] + for x in [np.arange(array_size)]: np.random.shuffle(x) - for k in [0, 1, 2, 10, 20]: + for k in k_options: indices = x.argsort()[::-1][:k] def topk(v, k=k): return nn_ops.top_k(v, k=k, sorted=True) self._assertOpOutputMatchesExpected( - topk, [x.astype(bfloat16)], - expected=[x[indices].astype(bfloat16), indices]) + topk, [x.astype(dtype)], + expected=[x[indices].astype(dtype), indices]) def testTopKZeros(self): """Tests that positive and negative zeros sort correctly.""" @@ -99,7 +108,7 @@ class XlaSortOpTest(xla_test.XLATestCase): {p: np.array([0., -0., 0., 3., -0., -4., 0., -0.], dtype=bfloat16)}) self.assertAllEqual( np.array([3., 0., 0., 0.], dtype=bfloat16), results[0]) - self.assertEqual(list([3, 0, 1, 2]), list(results[1])) + self.assertEqual(list([3, 0, 2, 6]), list(results[1])) def testTopKInfinities(self): """Tests that positive and negative infinity sort correctly.""" diff --git a/tensorflow/compiler/tests/spacetobatch_op_test.py b/tensorflow/compiler/tests/spacetobatch_op_test.py index f37c34156f96761632247be4bc1b62fca54f666e..c685bc548f9f6f8f7723c6f94dfd45f5420b4a67 100644 --- a/tensorflow/compiler/tests/spacetobatch_op_test.py +++ b/tensorflow/compiler/tests/spacetobatch_op_test.py @@ -20,7 +20,7 @@ from __future__ import print_function import numpy as np -from tensorflow.compiler.tests.xla_test import XLATestCase +from tensorflow.compiler.tests import xla_test from tensorflow.python.framework import dtypes from tensorflow.python.ops import array_ops from tensorflow.python.ops import gen_array_ops @@ -68,7 +68,7 @@ def space_to_batch_direct(input_array, block_shape, paddings): return permuted_reshaped_padded.reshape(output_shape) -class SpaceToBatchTest(XLATestCase): +class SpaceToBatchTest(xla_test.XLATestCase): """Tests input-output pairs for the SpaceToBatch and BatchToSpace ops.""" def _testPad(self, inputs, paddings, block_size, outputs): @@ -149,7 +149,7 @@ class SpaceToBatchTest(XLATestCase): self._testOne(x_np, block_size, x_out) -class SpaceToBatchNDTest(XLATestCase): +class SpaceToBatchNDTest(xla_test.XLATestCase): """Tests input-output pairs for the SpaceToBatchND and BatchToSpaceND ops.""" def _testPad(self, inputs, block_shape, paddings, outputs): diff --git a/tensorflow/compiler/tests/sparse_to_dense_op_test.py b/tensorflow/compiler/tests/sparse_to_dense_op_test.py new file mode 100644 index 0000000000000000000000000000000000000000..3db8101c4bfbb1b53c7318a36519612984d6f179 --- /dev/null +++ b/tensorflow/compiler/tests/sparse_to_dense_op_test.py @@ -0,0 +1,118 @@ +# Copyright 2015 The TensorFlow Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================== +"""Tests for tensorflow.kernels.sparse_op.""" + +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + +import numpy as np + +from tensorflow.compiler.tests import xla_test +from tensorflow.python.framework import dtypes +from tensorflow.python.ops import array_ops +from tensorflow.python.ops import sparse_ops +from tensorflow.python.platform import test + + +def _SparseToDense(sparse_indices, + output_size, + sparse_values, + default_value, + validate_indices=True): + feed_sparse_indices = array_ops.placeholder(dtypes.int32) + feed_dict = {feed_sparse_indices: sparse_indices} + return sparse_ops.sparse_to_dense( + feed_sparse_indices, + output_size, + sparse_values, + default_value=default_value, + validate_indices=validate_indices).eval(feed_dict=feed_dict) + + +class SparseToDenseTest(xla_test.XLATestCase): + + def testInt(self): + with self.test_session(), self.test_scope(): + tf_ans = _SparseToDense([1, 3], [5], 1, 0) + np_ans = np.array([0, 1, 0, 1, 0]).astype(np.int32) + self.assertAllClose(np_ans, tf_ans) + + def testFloat(self): + with self.test_session(), self.test_scope(): + tf_ans = _SparseToDense([1, 3], [5], 1.0, 0.0) + np_ans = np.array([0, 1, 0, 1, 0]).astype(np.float32) + self.assertAllClose(np_ans, tf_ans) + + def testSetValue(self): + with self.test_session(), self.test_scope(): + tf_ans = _SparseToDense([1, 3], [5], [1, 2], -1) + np_ans = np.array([-1, 1, -1, 2, -1]).astype(np.int32) + self.assertAllClose(np_ans, tf_ans) + + def testSetSingleValue(self): + with self.test_session(), self.test_scope(): + tf_ans = _SparseToDense([1, 3], [5], 1, -1) + np_ans = np.array([-1, 1, -1, 1, -1]).astype(np.int32) + self.assertAllClose(np_ans, tf_ans) + + def test2d(self): + # pylint: disable=bad-whitespace + with self.test_session(), self.test_scope(): + tf_ans = _SparseToDense([[1, 3], [2, 0]], [3, 4], 1, -1) + np_ans = np.array([[-1, -1, -1, -1], + [-1, -1, -1, 1], + [ 1, -1, -1, -1]]).astype(np.int32) + self.assertAllClose(np_ans, tf_ans) + + def testZeroDefault(self): + with self.test_session(): + x = sparse_ops.sparse_to_dense(2, [4], 7).eval() + self.assertAllEqual(x, [0, 0, 7, 0]) + + def test3d(self): + with self.test_session(), self.test_scope(): + tf_ans = _SparseToDense([[1, 3, 0], [2, 0, 1]], [3, 4, 2], 1, -1) + np_ans = np.ones((3, 4, 2), dtype=np.int32) * -1 + np_ans[1, 3, 0] = 1 + np_ans[2, 0, 1] = 1 + self.assertAllClose(np_ans, tf_ans) + + def testBadShape(self): + with self.test_session(), self.test_scope(): + with self.assertRaisesWithPredicateMatch(ValueError, "must be rank 1"): + _SparseToDense([1, 3], [[5], [3]], 1, -1) + + def testBadValue(self): + with self.test_session(), self.test_scope(): + with self.assertRaisesOpError( + r"sparse_values has incorrect shape \[2,1\], " + r"should be \[\] or \[2\]"): + _SparseToDense([1, 3], [5], [[5], [3]], -1) + + def testBadNumValues(self): + with self.test_session(), self.test_scope(): + with self.assertRaisesOpError( + r"sparse_values has incorrect shape \[3\], should be \[\] or \[2\]"): + _SparseToDense([1, 3], [5], [1, 2, 3], -1) + + def testBadDefault(self): + with self.test_session(), self.test_scope(): + with self.assertRaisesOpError("default_value should be a scalar"): + _SparseToDense([1, 3], [5], [1, 2], [0]) + + +if __name__ == "__main__": + test.main() diff --git a/tensorflow/compiler/tests/stack_ops_test.py b/tensorflow/compiler/tests/stack_ops_test.py index 94342f9567ca71274609e63b0482d55637c98d51..b7dd787feff2b22a9cfb5d43a4ba6ceb6eb0b301 100644 --- a/tensorflow/compiler/tests/stack_ops_test.py +++ b/tensorflow/compiler/tests/stack_ops_test.py @@ -20,7 +20,7 @@ from __future__ import print_function import numpy as np -from tensorflow.compiler.tests.xla_test import XLATestCase +from tensorflow.compiler.tests import xla_test from tensorflow.python.framework import dtypes from tensorflow.python.framework import ops from tensorflow.python.ops import array_ops @@ -28,7 +28,7 @@ from tensorflow.python.ops import gen_data_flow_ops from tensorflow.python.platform import test -class StackOpTest(XLATestCase): +class StackOpTest(xla_test.XLATestCase): def testStackPushPop(self): with self.test_session(), self.test_scope(): diff --git a/tensorflow/compiler/tests/stateless_random_ops_test.py b/tensorflow/compiler/tests/stateless_random_ops_test.py index abce190d831b25b364e393788aeeaf7dd1f2c5e1..d162675ef840131485128414b4a29e3cd89c8761 100644 --- a/tensorflow/compiler/tests/stateless_random_ops_test.py +++ b/tensorflow/compiler/tests/stateless_random_ops_test.py @@ -22,7 +22,7 @@ import math import numpy as np -from tensorflow.compiler.tests.xla_test import XLATestCase +from tensorflow.compiler.tests import xla_test from tensorflow.contrib import stateless from tensorflow.python.framework import dtypes from tensorflow.python.ops import array_ops @@ -30,7 +30,7 @@ from tensorflow.python.ops.distributions import special_math from tensorflow.python.platform import test -class StatelessRandomOpsTest(XLATestCase): +class StatelessRandomOpsTest(xla_test.XLATestCase): """Test cases for stateless random-number generator operators.""" def _random_types(self): diff --git a/tensorflow/compiler/tests/ternary_ops_test.py b/tensorflow/compiler/tests/ternary_ops_test.py index ef047005b60bd156a677050368ef67ae030d6c3a..effa5a59fee7dda543b2c409dfaa27a972a55808 100644 --- a/tensorflow/compiler/tests/ternary_ops_test.py +++ b/tensorflow/compiler/tests/ternary_ops_test.py @@ -20,7 +20,7 @@ from __future__ import print_function import numpy as np -from tensorflow.compiler.tests.xla_test import XLATestCase +from tensorflow.compiler.tests import xla_test from tensorflow.python.framework import dtypes from tensorflow.python.ops import array_ops from tensorflow.python.ops import gen_math_ops @@ -28,7 +28,7 @@ from tensorflow.python.ops import math_ops from tensorflow.python.platform import googletest -class TernaryOpsTest(XLATestCase): +class TernaryOpsTest(xla_test.XLATestCase): def _testTernary(self, op, a, b, c, expected): with self.test_session() as session: diff --git a/tensorflow/compiler/tests/unary_ops_test.py b/tensorflow/compiler/tests/unary_ops_test.py index e610b63e301c75f532db1b58cd26533effea174d..6a7011aea6cc3f942fecf27a640b998bfc10c0de 100644 --- a/tensorflow/compiler/tests/unary_ops_test.py +++ b/tensorflow/compiler/tests/unary_ops_test.py @@ -23,7 +23,7 @@ import unittest import numpy as np from six.moves import xrange # pylint: disable=redefined-builtin -from tensorflow.compiler.tests.xla_test import XLATestCase +from tensorflow.compiler.tests import xla_test from tensorflow.python.framework import dtypes from tensorflow.python.ops import array_ops from tensorflow.python.ops import bitwise_ops @@ -44,11 +44,16 @@ def nhwc_to_format(x, data_format): raise ValueError("Unknown format {}".format(data_format)) -class UnaryOpsTest(XLATestCase): +class UnaryOpsTest(xla_test.XLATestCase): """Test cases for unary operators.""" - def _assertOpOutputMatchesExpected(self, op, inp, expected, - equality_test=None, rtol=1e-3, atol=1e-5): + def _assertOpOutputMatchesExpected(self, + op, + inp, + expected, + equality_test=None, + rtol=1e-3, + atol=1e-5): """Verifies that 'op' produces 'expected' when fed input 'inp' . Args: @@ -81,10 +86,10 @@ class UnaryOpsTest(XLATestCase): def testAllTypeOps(self): for dtype in self.numeric_types: self._assertOpOutputMatchesExpected( - array_ops.diag, - np.array([1, 2, 3, 4], dtype=dtype), - np.array([[1, 0, 0, 0], [0, 2, 0, 0], [0, 0, 3, 0], [0, 0, 0, 4]], - dtype=dtype)) + array_ops.diag, np.array([1, 2, 3, 4], dtype=dtype), + np.array( + [[1, 0, 0, 0], [0, 2, 0, 0], [0, 0, 3, 0], [0, 0, 0, 4]], + dtype=dtype)) self._assertOpOutputMatchesExpected( array_ops.diag_part, np.arange(36).reshape([2, 3, 2, 3]).astype(dtype), @@ -102,8 +107,7 @@ class UnaryOpsTest(XLATestCase): expected=np.array([[-1, 1]], dtype=dtype)) self._assertOpOutputMatchesExpected( - array_ops.matrix_diag, - np.array([[1, 2], [3, 4]], dtype=dtype), + array_ops.matrix_diag, np.array([[1, 2], [3, 4]], dtype=dtype), np.array([[[1, 0], [0, 2]], [[3, 0], [0, 4]]], dtype=dtype)) self._assertOpOutputMatchesExpected( array_ops.matrix_diag, np.array([1, 2, 3, 4], dtype=dtype), @@ -115,10 +119,10 @@ class UnaryOpsTest(XLATestCase): np.array( [[[1, 2, 3], [4, 5, 6]], [[7, 8, 9], [10, 11, 12]]], dtype=dtype), np.array( - [[[[1, 0, 0], [0, 2, 0], [0, 0, 3]], - [[4, 0, 0], [0, 5, 0], [0, 0, 6]]], - [[[7, 0, 0], [0, 8, 0], [0, 0, 9]], - [[10, 0, 0], [0, 11, 0], [0, 0, 12]]]], + [[[[1, 0, 0], [0, 2, 0], [0, 0, 3]], [[4, 0, 0], [0, 5, 0], [ + 0, 0, 6 + ]]], [[[7, 0, 0], [0, 8, 0], [0, 0, 9]], [[10, 0, 0], [0, 11, 0], + [0, 0, 12]]]], dtype=dtype)) self._assertOpOutputMatchesExpected( array_ops.matrix_diag_part, @@ -159,36 +163,30 @@ class UnaryOpsTest(XLATestCase): continue x = np.arange(-0.90, 0.90, 0.25) self._assertOpOutputMatchesExpected( - math_ops.acos, - x.astype(dtype), - expected=np.arccos(x).astype(dtype)) + math_ops.acos, x.astype(dtype), expected=np.arccos(x).astype(dtype)) self._assertOpOutputMatchesExpected( - math_ops.asin, - x.astype(dtype), - expected=np.arcsin(x).astype(dtype)) + math_ops.asin, x.astype(dtype), expected=np.arcsin(x).astype(dtype)) x = np.arange(-3, 3).reshape(1, 3, 2) self._assertOpOutputMatchesExpected( - math_ops.atan, - x.astype(dtype), - expected=np.arctan(x).astype(dtype)) + math_ops.atan, x.astype(dtype), expected=np.arctan(x).astype(dtype)) self._assertOpOutputMatchesExpected( math_ops.acosh, np.array([1, 2, 3, 4], dtype=dtype), - expected=np.array([0, 1.3169579, 1.76274717, 2.06343707], - dtype=dtype)) + expected=np.array( + [0, 1.3169579, 1.76274717, 2.06343707], dtype=dtype)) self._assertOpOutputMatchesExpected( math_ops.asinh, np.array([1, 2, 3, 4], dtype=dtype), - expected=np.array([0.88137359, 1.44363548, 1.81844646, 2.09471255], - dtype=dtype)) + expected=np.array( + [0.88137359, 1.44363548, 1.81844646, 2.09471255], dtype=dtype)) self._assertOpOutputMatchesExpected( math_ops.atanh, np.array([0.1, 0.2, 0.3, 0.4], dtype=dtype), - expected=np.array([0.10033535, 0.20273255, 0.3095196, 0.42364893], - dtype=dtype)) + expected=np.array( + [0.10033535, 0.20273255, 0.3095196, 0.42364893], dtype=dtype)) self._assertOpOutputMatchesExpected( math_ops.ceil, @@ -198,8 +196,8 @@ class UnaryOpsTest(XLATestCase): self._assertOpOutputMatchesExpected( math_ops.cosh, np.array([1, 2, 3, 4], dtype=dtype), - expected=np.array([1.54308063, 3.76219569, 10.067662, 27.30823284], - dtype=dtype)) + expected=np.array( + [1.54308063, 3.76219569, 10.067662, 27.30823284], dtype=dtype)) # Disable float16 testing for now if dtype != np.float16: @@ -229,8 +227,8 @@ class UnaryOpsTest(XLATestCase): self._assertOpOutputMatchesExpected( math_ops.is_finite, - np.array([[np.NINF, -2, -1, 0, 0.5, 1, 2, np.inf, np.nan]], - dtype=dtype), + np.array( + [[np.NINF, -2, -1, 0, 0.5, 1, 2, np.inf, np.nan]], dtype=dtype), expected=np.array([[0, 1, 1, 1, 1, 1, 1, 0, 0]], dtype=np.bool)) # Tests for tf.nn ops. @@ -271,16 +269,20 @@ class UnaryOpsTest(XLATestCase): self._assertOpOutputMatchesExpected( math_ops.rint, - np.array([[-1.7, 1.2, 4.0, 0.0], [-3.5, -2.5, -1.5, -0.5], - [0.5, 1.5, 2.5, 3.5]], dtype=dtype), - expected=np.array([[-2, 1, 4, 0], [-4, -2, -2, 0], [0, 2, 2, 4]], - dtype=dtype)) + np.array( + [[-1.7, 1.2, 4.0, 0.0], [-3.5, -2.5, -1.5, -0.5], + [0.5, 1.5, 2.5, 3.5]], + dtype=dtype), + expected=np.array( + [[-2, 1, 4, 0], [-4, -2, -2, 0], [0, 2, 2, 4]], dtype=dtype)) self._assertOpOutputMatchesExpected( math_ops.round, - np.array([[-1.7, 1.2, 4.0, 0.0], [-3.5, -2.5, -1.5, -0.5], - [0.5, 1.5, 2.5, 3.5]], dtype=dtype), - expected=np.array([[-2, 1, 4, 0], [-4, -2, -2, 0], [0, 2, 2, 4]], - dtype=dtype)) + np.array( + [[-1.7, 1.2, 4.0, 0.0], [-3.5, -2.5, -1.5, -0.5], + [0.5, 1.5, 2.5, 3.5]], + dtype=dtype), + expected=np.array( + [[-2, 1, 4, 0], [-4, -2, -2, 0], [0, 2, 2, 4]], dtype=dtype)) self._assertOpOutputMatchesExpected( math_ops.rsqrt, @@ -289,10 +291,7 @@ class UnaryOpsTest(XLATestCase): self._assertOpOutputMatchesExpected( math_ops.sigmoid, - np.array( - [[1, 1, 1, 1], - [1, 2, 3, 4]], - dtype=dtype), + np.array([[1, 1, 1, 1], [1, 2, 3, 4]], dtype=dtype), expected=np.array( [[0.7310586, 0.7310586, 0.7310586, 0.7310586], [0.7310586, 0.880797, 0.95257413, 0.98201376]], @@ -306,8 +305,8 @@ class UnaryOpsTest(XLATestCase): self._assertOpOutputMatchesExpected( math_ops.sinh, np.array([1, 2, 3, 4], dtype=dtype), - expected=np.array([1.17520119, 3.62686041, 10.01787493, 27.2899172], - dtype=dtype)) + expected=np.array( + [1.17520119, 3.62686041, 10.01787493, 27.2899172], dtype=dtype)) self._assertOpOutputMatchesExpected( math_ops.sqrt, @@ -317,15 +316,12 @@ class UnaryOpsTest(XLATestCase): self._assertOpOutputMatchesExpected( math_ops.tan, np.array([1, 2, 3, 4], dtype=dtype), - expected=np.array([1.55740772, -2.18503986, -0.14254654, 1.15782128], - dtype=dtype)) + expected=np.array( + [1.55740772, -2.18503986, -0.14254654, 1.15782128], dtype=dtype)) self._assertOpOutputMatchesExpected( math_ops.tanh, - np.array( - [[1, 1, 1, 1], - [1, 2, 3, 4]], - dtype=dtype), + np.array([[1, 1, 1, 1], [1, 2, 3, 4]], dtype=dtype), expected=np.array( [[0.76159418, 0.76159418, 0.76159418, 0.76159418], [0.76159418, 0.96402758, 0.99505478, 0.99932933]], @@ -333,10 +329,7 @@ class UnaryOpsTest(XLATestCase): self._assertOpOutputMatchesExpected( nn_ops.log_softmax, - np.array( - [[1, 1, 1, 1], - [1, 2, 3, 4]], - dtype=dtype), + np.array([[1, 1, 1, 1], [1, 2, 3, 4]], dtype=dtype), expected=np.array( [[-1.3862944, -1.3862944, -1.3862944, -1.3862944], [-3.4401896, -2.4401896, -1.4401897, -0.44018969]], @@ -370,10 +363,7 @@ class UnaryOpsTest(XLATestCase): self._assertOpOutputMatchesExpected( nn_ops.softmax, - np.array( - [[1, 1, 1, 1], - [1, 2, 3, 4]], - dtype=dtype), + np.array([[1, 1, 1, 1], [1, 2, 3, 4]], dtype=dtype), expected=np.array( [[0.25, 0.25, 0.25, 0.25], [0.032058604, 0.087144323, 0.23688284, 0.64391428]], @@ -382,8 +372,8 @@ class UnaryOpsTest(XLATestCase): self._assertOpOutputMatchesExpected( nn_ops.softsign, np.array([[-2, -1, 0, 1, 2]], dtype=dtype), - expected=np.array([[-0.66666669, -0.5, 0, 0.5, 0.66666669]], - dtype=dtype)) + expected=np.array( + [[-0.66666669, -0.5, 0, 0.5, 0.66666669]], dtype=dtype)) self._assertOpOutputMatchesExpected( math_ops.is_finite, @@ -392,10 +382,23 @@ class UnaryOpsTest(XLATestCase): expected=np.array( [[True, False, True], [False, True, True]], dtype=np.bool)) + def quantize_and_dequantize_v2(x): + return array_ops.quantize_and_dequantize_v2( + x, -127, 127, signed_input=True, num_bits=8) + + self._assertOpOutputMatchesExpected( + quantize_and_dequantize_v2, + np.array([-1, -0.5, 0, 0.3], dtype=dtype), + expected=np.array([-1., -0.5, 0., 0.296875], dtype=dtype)) + + def quantize_and_dequantize_v3(x): + return array_ops.quantize_and_dequantize_v3( + x, -127, 127, num_bits=8, signed_input=True, range_given=False) + self._assertOpOutputMatchesExpected( - lambda x: array_ops.quantize_and_dequantize_v2(x, -127, 127, True, 8), + quantize_and_dequantize_v3, np.array([-1, -0.5, 0, 0.3], dtype=dtype), - expected=np.array([-1, -64.0 / 127, 0, 38.0 / 127], dtype=dtype)) + expected=np.array([-1., -0.5, 0., 0.296875], dtype=dtype)) def testComplexOps(self): for dtype in self.complex_types: @@ -576,13 +579,13 @@ class UnaryOpsTest(XLATestCase): for dtype in self.float_types: self._assertOpOutputMatchesExpected( math_ops.is_inf, - np.array([[np.NINF, -2, -1, 0, 0.5, 1, 2, np.inf, np.nan]], - dtype=dtype), + np.array( + [[np.NINF, -2, -1, 0, 0.5, 1, 2, np.inf, np.nan]], dtype=dtype), expected=np.array([[1, 0, 0, 0, 0, 0, 0, 1, 0]], dtype=np.bool)) self._assertOpOutputMatchesExpected( math_ops.is_nan, - np.array([[np.NINF, -2, -1, 0, 0.5, 1, 2, np.inf, np.nan]], - dtype=dtype), + np.array( + [[np.NINF, -2, -1, 0, 0.5, 1, 2, np.inf, np.nan]], dtype=dtype), expected=np.array([[0, 0, 0, 0, 0, 0, 0, 0, 1]], dtype=np.bool)) def testLogicalOps(self): @@ -599,14 +602,15 @@ class UnaryOpsTest(XLATestCase): self._assertOpOutputMatchesExpected( lambda x: gen_nn_ops.bias_add_grad(x, data_format="NCHW"), - np.array([[[1., 2.], [3., 4.]], [[5., 6.], [7., 8.]]], - dtype=np.float32), + np.array( + [[[1., 2.], [3., 4.]], [[5., 6.], [7., 8.]]], dtype=np.float32), expected=np.array([10., 26.], dtype=np.float32)) def testCast(self): shapes = [[], [4], [2, 3], [2, 0, 4]] - types = (set([dtypes.bool, dtypes.int32, dtypes.float32]) | - self.complex_tf_types) + types = ( + set([dtypes.bool, dtypes.int32, dtypes.float32]) + | self.complex_tf_types) for shape in shapes: for src_type in types: for dst_type in types: @@ -648,14 +652,11 @@ class UnaryOpsTest(XLATestCase): self._assertOpOutputMatchesExpected( rank_op, dtype(7), expected=np.int32(0)) self._assertOpOutputMatchesExpected( - rank_op, np.array( - [[], []], dtype=dtype), expected=np.int32(2)) + rank_op, np.array([[], []], dtype=dtype), expected=np.int32(2)) self._assertOpOutputMatchesExpected( - rank_op, np.array( - [-1, 1], dtype=dtype), expected=np.int32(1)) + rank_op, np.array([-1, 1], dtype=dtype), expected=np.int32(1)) self._assertOpOutputMatchesExpected( - rank_op, np.array( - [[-1, 1]], dtype=dtype), expected=np.int32(2)) + rank_op, np.array([[-1, 1]], dtype=dtype), expected=np.int32(2)) self._assertOpOutputMatchesExpected( rank_op, np.array([[-1], [1], [4]], dtype=dtype), @@ -720,97 +721,97 @@ class UnaryOpsTest(XLATestCase): equality_test=self.ListsAreClose) def testDepthToSpace(self): + def make_op(data_format): + def op(x): - return array_ops.depth_to_space(x, block_size=2, - data_format=data_format) + return array_ops.depth_to_space( + x, block_size=2, data_format=data_format) + return op for dtype in self.numeric_types: for data_format in ["NCHW", "NHWC"]: self._assertOpOutputMatchesExpected( make_op(data_format), - nhwc_to_format(np.array([[[[1, 2, 3, 4]]]], dtype=dtype), - data_format), - expected=nhwc_to_format(np.array([[[[1], [2]], - [[3], [4]]]], dtype=dtype), - data_format)) + nhwc_to_format( + np.array([[[[1, 2, 3, 4]]]], dtype=dtype), data_format), + expected=nhwc_to_format( + np.array([[[[1], [2]], [[3], [4]]]], dtype=dtype), data_format)) self._assertOpOutputMatchesExpected( make_op(data_format), nhwc_to_format( - np.array([[[[1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12]]]], - dtype=dtype), + np.array( + [[[[1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12]]]], dtype=dtype), data_format), expected=nhwc_to_format( - np.array([[[[1, 2, 3], [4, 5, 6]], [[7, 8, 9], [10, 11, 12]]]], - dtype=dtype), - data_format)) + np.array( + [[[[1, 2, 3], [4, 5, 6]], [[7, 8, 9], [10, 11, 12]]]], + dtype=dtype), data_format)) self._assertOpOutputMatchesExpected( make_op(data_format), nhwc_to_format( - np.array([[[[1, 2, 3, 4], - [5, 6, 7, 8]], - [[9, 10, 11, 12], - [13, 14, 15, 16]]]], dtype=dtype), - data_format), + np.array( + [[[[1, 2, 3, 4], [5, 6, 7, 8]], [[9, 10, 11, 12], + [13, 14, 15, 16]]]], + dtype=dtype), data_format), expected=nhwc_to_format( - np.array([[[[1], [2], [5], [6]], - [[3], [4], [7], [8]], - [[9], [10], [13], [14]], - [[11], [12], [15], [16]]]], dtype=dtype), - data_format)) + np.array( + [[[[1], [2], [5], [6]], [[3], [4], [7], [8]], + [[9], [10], [13], [14]], [[11], [12], [15], [16]]]], + dtype=dtype), data_format)) def testSpaceToDepth(self): + def make_op(data_format): + def op(x): - return array_ops.space_to_depth(x, block_size=2, - data_format=data_format) + return array_ops.space_to_depth( + x, block_size=2, data_format=data_format) + return op for dtype in self.numeric_types: for data_format in ["NCHW", "NHWC"]: self._assertOpOutputMatchesExpected( make_op(data_format), - nhwc_to_format(np.array([[[[1], [2]], - [[3], [4]]]], dtype=dtype), - data_format), - expected=nhwc_to_format(np.array([[[[1, 2, 3, 4]]]], dtype=dtype), - data_format)) + nhwc_to_format( + np.array([[[[1], [2]], [[3], [4]]]], dtype=dtype), data_format), + expected=nhwc_to_format( + np.array([[[[1, 2, 3, 4]]]], dtype=dtype), data_format)) self._assertOpOutputMatchesExpected( make_op(data_format), - nhwc_to_format(np.array([[[[1, 2, 3], [4, 5, 6]], - [[7, 8, 9], [10, 11, 12]]]], dtype=dtype), - data_format), + nhwc_to_format( + np.array( + [[[[1, 2, 3], [4, 5, 6]], [[7, 8, 9], [10, 11, 12]]]], + dtype=dtype), data_format), expected=nhwc_to_format( - np.array([[[[1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12]]]], - dtype=dtype), + np.array( + [[[[1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12]]]], dtype=dtype), data_format)) self._assertOpOutputMatchesExpected( make_op(data_format), - nhwc_to_format(np.array([[[[1], [2], [5], [6]], - [[3], [4], [7], [8]], - [[9], [10], [13], [14]], - [[11], [12], [15], [16]]]], dtype=dtype), - data_format), + nhwc_to_format( + np.array( + [[[[1], [2], [5], [6]], [[3], [4], [7], [8]], + [[9], [10], [13], [14]], [[11], [12], [15], [16]]]], + dtype=dtype), data_format), expected=nhwc_to_format( - np.array([[[[1, 2, 3, 4], - [5, 6, 7, 8]], - [[9, 10, 11, 12], - [13, 14, 15, 16]]]], dtype=dtype), - data_format)) + np.array( + [[[[1, 2, 3, 4], [5, 6, 7, 8]], [[9, 10, 11, 12], + [13, 14, 15, 16]]]], + dtype=dtype), data_format)) def _assertSoftplusMatchesExpected(self, features, dtype): features = np.array(features, dtype=dtype) zero = np.asarray(0).astype(dtype) expected = np.logaddexp(zero, features) self._assertOpOutputMatchesExpected( - nn_ops.softplus, features, expected=expected, - rtol=1e-6, - atol=9.1e-6) + nn_ops.softplus, features, expected=expected, rtol=1e-6, atol=9.1e-6) def testSoftplus(self): for dtype in self.float_types: @@ -824,9 +825,10 @@ class UnaryOpsTest(XLATestCase): one = dtype(1) ten = dtype(10) self._assertSoftplusMatchesExpected([ - log_eps, log_eps - one, log_eps + one, log_eps - ten, - log_eps + ten, -log_eps, -log_eps - one, -log_eps + one, - -log_eps - ten, -log_eps + ten], dtype) + log_eps, log_eps - one, log_eps + one, log_eps - ten, log_eps + ten, + -log_eps, -log_eps - one, -log_eps + one, -log_eps - ten, + -log_eps + ten + ], dtype) if __name__ == "__main__": diff --git a/tensorflow/compiler/tests/variable_ops_test.py b/tensorflow/compiler/tests/variable_ops_test.py index bd616f2a20cabfe1e85d325f592565171a1297c2..dd2c252d383bca9c59033ac07e442b487e4975a6 100644 --- a/tensorflow/compiler/tests/variable_ops_test.py +++ b/tensorflow/compiler/tests/variable_ops_test.py @@ -20,7 +20,7 @@ from __future__ import print_function import numpy as np -from tensorflow.compiler.tests.xla_test import XLATestCase +from tensorflow.compiler.tests import xla_test from tensorflow.python.framework import constant_op from tensorflow.python.framework import dtypes from tensorflow.python.framework import errors @@ -37,7 +37,7 @@ from tensorflow.python.platform import googletest from tensorflow.python.training.gradient_descent import GradientDescentOptimizer -class VariableOpsTest(XLATestCase): +class VariableOpsTest(xla_test.XLATestCase): """Test cases for resource variable operators.""" def testOneWriteOneOutput(self): @@ -435,7 +435,7 @@ class StridedSliceAssignChecker(object): self.test.assertAllEqual(val, valnp) -class SliceAssignTest(XLATestCase): +class SliceAssignTest(xla_test.XLATestCase): def testSliceAssign(self): for dtype in self.numeric_types: diff --git a/tensorflow/compiler/tests/while_test.py b/tensorflow/compiler/tests/while_test.py index f79eb27435cc954cebde4357c1d946a320f4ed75..b637cf31cfc303ebe84ce8307ef4ad8b0b5cd720 100644 --- a/tensorflow/compiler/tests/while_test.py +++ b/tensorflow/compiler/tests/while_test.py @@ -20,7 +20,7 @@ from __future__ import print_function import numpy as np -from tensorflow.compiler.tests.xla_test import XLATestCase +from tensorflow.compiler.tests import xla_test from tensorflow.compiler.tf2xla.python import xla from tensorflow.python.framework import constant_op from tensorflow.python.framework import dtypes @@ -29,7 +29,7 @@ from tensorflow.python.ops import array_ops from tensorflow.python.platform import test -class WhileTest(XLATestCase): +class WhileTest(xla_test.XLATestCase): def testSingletonLoopHandrolled(self): # Define a function for the loop body diff --git a/tensorflow/compiler/tests/xla_device_test.py b/tensorflow/compiler/tests/xla_device_test.py index f0b010fa67f2ffb3f81fd14d4d89585f716b4890..06d977b93c28792704b910c688af510bc650d2a4 100644 --- a/tensorflow/compiler/tests/xla_device_test.py +++ b/tensorflow/compiler/tests/xla_device_test.py @@ -20,14 +20,14 @@ from __future__ import print_function import numpy as np -from tensorflow.compiler.tests.xla_test import XLATestCase +from tensorflow.compiler.tests import xla_test from tensorflow.python.framework import ops from tensorflow.python.ops import array_ops from tensorflow.python.ops import gen_control_flow_ops from tensorflow.python.platform import test -class XlaDeviceTest(XLATestCase): +class XlaDeviceTest(xla_test.XLATestCase): def testCopies(self): """Tests that copies onto and off XLA devices work.""" diff --git a/tensorflow/compiler/tf2xla/BUILD b/tensorflow/compiler/tf2xla/BUILD index a7b9cc6c811098e7af95d10ec739b26508720548..ff002d15b09f5a9dc1a69741672503656f893298 100644 --- a/tensorflow/compiler/tf2xla/BUILD +++ b/tensorflow/compiler/tf2xla/BUILD @@ -139,12 +139,14 @@ cc_library( "xla_op_registry.cc", "xla_resource.cc", "xla_cpu_backend.cc", + "legacy_flags/backend_registration_flags.cc", ] + if_cuda_is_configured([ "xla_gpu_backend.cc", ]), hdrs = [ "const_analysis.h", "graph_compiler.h", + "legacy_flags/backend_registration_flags.h", "xla_compilation_device.h", "xla_compiler.h", "xla_context.h", @@ -162,18 +164,24 @@ cc_library( ":sharding_util", ":tf2xla_util", "//tensorflow/compiler/tf2xla/lib:util", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_util", + "//tensorflow/compiler/xla:status_macros", "//tensorflow/compiler/xla:statusor", "//tensorflow/compiler/xla:types", "//tensorflow/compiler/xla:xla_data_proto", "//tensorflow/compiler/xla/client:client_library", "//tensorflow/compiler/xla/client:local_client", + "//tensorflow/compiler/xla/client/lib:arithmetic", + "//tensorflow/compiler/xla/client/lib:constants", + "//tensorflow/compiler/xla/client/lib:numeric", "//tensorflow/compiler/xla/client/xla_client:xla_builder", "//tensorflow/compiler/xla/client/xla_client:xla_computation", + "//tensorflow/compiler/xla/legacy_flags:parse_flags_from_env", "//tensorflow/core:core_cpu", "//tensorflow/core:core_cpu_internal", "//tensorflow/core:framework", + "//tensorflow/core:framework_internal", "//tensorflow/core:lib", "//tensorflow/core:lib_internal", "//tensorflow/core:protos_all_cc", @@ -198,7 +206,7 @@ cc_library( ], visibility = [":friends"], deps = [ - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:xla_data_proto", "//tensorflow/core:core_cpu_internal", @@ -281,6 +289,7 @@ tf_cc_test( deps = [ ":tf2xla", ":tf2xla_proto", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:literal_util", "//tensorflow/compiler/xla:statusor", "//tensorflow/compiler/xla/client:client_library", @@ -323,7 +332,7 @@ tf_cc_test( "//tensorflow/cc:ops", "//tensorflow/cc:resource_variable_ops", "//tensorflow/compiler/tf2xla/kernels:xla_ops", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:status_macros", "//tensorflow/compiler/xla/client:client_library", @@ -360,6 +369,7 @@ tf_cc_test( ], deps = [ ":common", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:literal_util", "//tensorflow/core:framework", "//tensorflow/core:test", diff --git a/tensorflow/compiler/tf2xla/functionalize_control_flow.cc b/tensorflow/compiler/tf2xla/functionalize_control_flow.cc index 140dad61d96c46f512d0758ca02619fdee5ec2a1..6cc95149a16a59fce8486c5d103ad09e3e262765 100644 --- a/tensorflow/compiler/tf2xla/functionalize_control_flow.cc +++ b/tensorflow/compiler/tf2xla/functionalize_control_flow.cc @@ -166,6 +166,27 @@ StatusOr AddNode(const NodeDef& node_def, Graph* graph) { return inserted_node; } +// Check that the graph has no cycle containing the given node. +Status CheckNoCycleContains(const Node* node, const int num_nodes) { + std::vector ready; + ready.push_back(node); + std::vector visited(num_nodes); + while (!ready.empty()) { + const Node* current_node = ready.back(); + ready.pop_back(); + visited[current_node->id()] = true; + for (const Edge* out : current_node->out_edges()) { + if (out->dst() == node) { + return errors::Internal("Detect a cycle: Node \"", node->name(), "\"(", + node->def().op(), ") feeds into itself."); + } else if (!visited[out->dst()->id()]) { + ready.push_back(out->dst()); + } + } + } + return Status::OK(); +} + StatusOr BuildArgNode(Graph* graph, DataType type, int index) { NodeDef arg_def; NodeDefBuilder builder(strings::StrCat(kArgOp, index), kArgOp); @@ -1407,6 +1428,10 @@ StatusOr FunctionalizeCond::ConvertToXlaIf( TF_RETURN_IF_ERROR( AddInputEdges(cond_arg_nodes, switch_cluster.predicate_edge, if_node)); TF_RETURN_IF_ERROR(AddOutputEdges(merge_nodes, if_node)); + // Check that the if_node doesn't feed into itself. + TF_RETURN_WITH_CONTEXT_IF_ERROR( + CheckNoCycleContains(if_node, graph_->num_node_ids()), + "ConvertToXlaIf failed."); return if_node; } @@ -1506,6 +1531,16 @@ Status FunctionalizeControlFlow(const FunctionLibraryDefinition* lookup_library, worklist.push_back(frame->parent); } } + // There should be no cycle at this point, since while loops have been removed + // from graph. + // Check that the newly added XlaWhile nodes don't feed into themselves. + for (const Node* node : graph->op_nodes()) { + if (node->def().op() == "XlaWhile") { + TF_RETURN_WITH_CONTEXT_IF_ERROR( + CheckNoCycleContains(node, graph->num_node_ids()), + "FunctionalizeLoop failed."); + } + } // FunctionalizeControlFlow is invoked for every function, so the loops's // bodies and conditionals that were extracted into functions will be handled diff --git a/tensorflow/compiler/tf2xla/functionalize_control_flow_test.cc b/tensorflow/compiler/tf2xla/functionalize_control_flow_test.cc index 14977a908ae2b0ff7e13b634c41b6d331b4b8a36..aae2f8ee5acd6249f8b6002d94c877f18064f936 100644 --- a/tensorflow/compiler/tf2xla/functionalize_control_flow_test.cc +++ b/tensorflow/compiler/tf2xla/functionalize_control_flow_test.cc @@ -29,6 +29,7 @@ limitations under the License. #include "tensorflow/core/framework/op.h" #include "tensorflow/core/graph/graph_constructor.h" #include "tensorflow/core/graph/graph_def_builder.h" +#include "tensorflow/core/graph/validate.h" #include "tensorflow/core/lib/core/status_test_util.h" #include "tensorflow/core/platform/test.h" #include "tensorflow/core/util/equal_graph_def.h" @@ -1012,5 +1013,60 @@ TEST(FunctionalizeControlFlow, Complex) { } } +TEST(FunctionalizeControlFlow, Cycle) { + std::unique_ptr graph(new Graph(OpRegistry::Global())); + // ----------------------------------------------------- + // | | + // | v + // less -> switch_1 --> add -> merge_1 -> identity -> switch_2 + // | ^ | + // | | v + // --------> one -------------------------> add_2 ---> merge_2 + { + Scope scope = Scope::NewRootScope().ExitOnError(); + + auto x = ops::Placeholder(scope.WithOpName("x"), DT_INT32); + auto y = ops::Placeholder(scope.WithOpName("y"), DT_INT32); + auto less = ops::Less(scope.WithOpName("cond/Less"), y, x); + auto switch_1 = ops::Switch(scope.WithOpName("cond/Switch"), x, less); + auto two = + ops::Const(scope.WithOpName("cond/two") + .WithControlDependencies(switch_1.output_true), + 2); + auto mul = ops::Multiply(scope.WithOpName("cond/true/mul"), + switch_1.output_true, two); + auto one = + ops::Const(scope.WithOpName("cond/one") + .WithControlDependencies(switch_1.output_false), + 1); + auto add = ops::Add(scope.WithOpName("cond/false/add"), + switch_1.output_false, one); + + auto merge_1 = ops::Merge(scope.WithOpName("cond/Merge"), + std::initializer_list{add, mul}); + auto identity = + ops::Identity(scope.WithOpName("cond/Merge/identity"), merge_1.output); + auto switch_2 = + ops::Switch(scope.WithOpName("grad/cond/Switch"), identity, less); + auto add_2 = ops::Add(scope.WithOpName("cond_2/false/add"), + switch_2.output_false, one); + auto mul_2 = ops::Multiply(scope.WithOpName("cond_2/true/mul"), + switch_2.output_true, two); + auto merge_2 = ops::Merge(scope.WithOpName("cond_2/Merge"), + std::initializer_list{add_2, mul_2}); + TF_ASSERT_OK(scope.ToGraph(graph.get())); + } + // No cycle before functionalize control flow. + TF_EXPECT_OK(graph::ValidateGraphHasNoCycle(*graph)); + FunctionLibraryDefinition library(OpRegistry::Global(), {}); + // switch_1 and switch_2 have the same switch depth. They are replaced by a + // single XlaIf node during FunctionalizeControlFlow, resulting in a cycle: + // less -> XlaIf <--> identity. + Status status = FunctionalizeControlFlow(graph.get(), &library); + EXPECT_FALSE(status.ok()); + EXPECT_TRUE(str_util::StrContains(status.error_message(), "Detect a cycle")) + << status.error_message(); +} + } // namespace } // namespace tensorflow diff --git a/tensorflow/compiler/tf2xla/graph_compiler.cc b/tensorflow/compiler/tf2xla/graph_compiler.cc index 212f6f3966149ca0b2d2e012b19300e1f488f996..e1cea03865ce9978e634429b5ce41fe8b245a575 100644 --- a/tensorflow/compiler/tf2xla/graph_compiler.cc +++ b/tensorflow/compiler/tf2xla/graph_compiler.cc @@ -29,6 +29,7 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/xla_context.h" #include "tensorflow/compiler/tf2xla/xla_op_kernel.h" #include "tensorflow/compiler/xla/client/client_library.h" +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/core/common_runtime/device.h" #include "tensorflow/core/common_runtime/executor.h" #include "tensorflow/core/common_runtime/function.h" @@ -39,6 +40,7 @@ limitations under the License. #include "tensorflow/core/graph/algorithm.h" #include "tensorflow/core/graph/graph_constructor.h" #include "tensorflow/core/graph/node_builder.h" +#include "tensorflow/core/graph/validate.h" #include "tensorflow/core/lib/gtl/cleanup.h" #include "tensorflow/core/lib/hash/hash.h" #include "tensorflow/core/platform/logging.h" @@ -87,6 +89,8 @@ Status PrepareArguments(XlaOpKernelContext* ctx, Graph* graph, } } // namespace Status GraphCompiler::Compile() { + // Check that the graph has no illegal cycles. + TF_RETURN_IF_ERROR(graph::ValidateGraphHasNoCycle(*graph_)); // Maintain a mapping from node id to node outputs. using NodeOutputs = std::vector; std::vector output_registry(graph_->num_node_ids()); @@ -157,9 +161,8 @@ Status GraphCompiler::Compile() { outputs.resize(n->num_outputs()); for (int o = 0; o < n->num_outputs(); ++o) { outputs[o] = op_context.release_output(o); - if (*op_context.is_output_dead() || outputs[o].tensor == nullptr) { + if (outputs[o].tensor == nullptr) { return errors::Internal("Missing xla_context ", o, "-th output from ", - (*op_context.is_output_dead() ? "(dead)" : ""), SummarizeNode(*n)); } } @@ -227,7 +230,7 @@ Status GraphCompiler::CompileFunctionalNode(Node* n, XlaContext& context = XlaContext::Get(op_context); auto* b = context.builder(); - auto output_handle = b->Call(*result.computation, handles); + auto output_handle = xla::Call(b, *result.computation, handles); // The output handle of `Call` computation is a tuple type. Unzip it so // that it can fit into future computations. int computation_output = 0; @@ -236,7 +239,7 @@ Status GraphCompiler::CompileFunctionalNode(Node* n, xla_op_context.SetConstantOutput(i, result.outputs[i].constant_value); } else { xla_op_context.SetOutput( - i, b->GetTupleElement(output_handle, computation_output)); + i, xla::GetTupleElement(output_handle, computation_output)); ++computation_output; } } diff --git a/tensorflow/compiler/tf2xla/kernels/BUILD b/tensorflow/compiler/tf2xla/kernels/BUILD index 659ff7321b63a432cce3ecc5e094e159a96fbf2f..5a335aa43cf189cbf4c37c358ac1e64d700c0873 100644 --- a/tensorflow/compiler/tf2xla/kernels/BUILD +++ b/tensorflow/compiler/tf2xla/kernels/BUILD @@ -58,6 +58,7 @@ tf_kernel_library( "pack_op.cc", "pad_op.cc", "pooling_ops.cc", + "qr_op.cc", "quantize_and_dequantize_op.cc", "random_ops.cc", "reduce_window_op.cc", @@ -82,6 +83,7 @@ tf_kernel_library( "sort_ops.cc", "spacetobatch_op.cc", "spacetodepth_op.cc", + "sparse_to_dense_op.cc", "split_op.cc", "stack_ops.cc", "stateless_random_ops.cc", @@ -106,6 +108,7 @@ tf_kernel_library( "//tensorflow/compiler/tf2xla:xla_compiler", "//tensorflow/compiler/tf2xla/lib:batch_dot", "//tensorflow/compiler/tf2xla/lib:cholesky", + "//tensorflow/compiler/tf2xla/lib:qr", "//tensorflow/compiler/tf2xla/lib:random", "//tensorflow/compiler/tf2xla/lib:scatter", "//tensorflow/compiler/tf2xla/lib:triangular_solve", @@ -113,6 +116,7 @@ tf_kernel_library( "//tensorflow/compiler/tf2xla/lib:while_loop", "//tensorflow/compiler/tf2xla/ops:xla_ops", "//tensorflow/compiler/xla:array4d", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:literal_util", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:status_macros", @@ -120,6 +124,9 @@ tf_kernel_library( "//tensorflow/compiler/xla:xla_data_proto", "//tensorflow/compiler/xla/client:client_library", "//tensorflow/compiler/xla/client/lib:arithmetic", + "//tensorflow/compiler/xla/client/lib:constants", + "//tensorflow/compiler/xla/client/lib:math", + "//tensorflow/compiler/xla/client/lib:numeric", "//tensorflow/compiler/xla/client/xla_client:xla_builder", "//tensorflow/core:framework", "//tensorflow/core:image_ops_op_lib", @@ -155,7 +162,7 @@ tf_kernel_library( "//tensorflow/compiler/tf2xla:common", "//tensorflow/compiler/tf2xla:xla_compiler", "//tensorflow/compiler/tf2xla/ops:xla_ops", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla/client/xla_client:xla_builder", "//tensorflow/core:framework", "//tensorflow/core:lib", @@ -171,7 +178,7 @@ tf_kernel_library( "//tensorflow/compiler/tf2xla:common", "//tensorflow/compiler/tf2xla:xla_compiler", "//tensorflow/compiler/tf2xla/ops:xla_ops", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla/client/xla_client:xla_builder", "//tensorflow/core:framework", "//tensorflow/core:lib", @@ -206,6 +213,7 @@ tf_kernel_library( ":index_ops_kernel_argmax_float_2d", "//tensorflow/compiler/tf2xla:common", "//tensorflow/compiler/tf2xla:xla_compiler", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:literal_util", "//tensorflow/compiler/xla/client:client_library", "//tensorflow/compiler/xla/client/lib:arithmetic", diff --git a/tensorflow/compiler/tf2xla/kernels/aggregate_ops.cc b/tensorflow/compiler/tf2xla/kernels/aggregate_ops.cc index 1e59868621475cf72f4cc8b14dafec2dd8cd5c95..e33532828040123243f839ab1aa655b4bbc72520 100644 --- a/tensorflow/compiler/tf2xla/kernels/aggregate_ops.cc +++ b/tensorflow/compiler/tf2xla/kernels/aggregate_ops.cc @@ -15,6 +15,7 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/xla_op_kernel.h" #include "tensorflow/compiler/tf2xla/xla_op_registry.h" +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" namespace tensorflow { namespace { @@ -31,7 +32,7 @@ class AddNOp : public XlaOpKernel { xla::XlaOp sum = ctx->Input(0); for (int i = 1; i < ctx->num_inputs(); ++i) { - sum = ctx->builder()->Add(sum, ctx->Input(i)); + sum = xla::Add(sum, ctx->Input(i)); } ctx->SetOutput(0, sum); diff --git a/tensorflow/compiler/tf2xla/kernels/batch_matmul_op.cc b/tensorflow/compiler/tf2xla/kernels/batch_matmul_op.cc index b0ba25b9983c3a9af26728ce4b1c263c844327db..4cfe946b2e6146f034867c06e996ffae42b90705 100644 --- a/tensorflow/compiler/tf2xla/kernels/batch_matmul_op.cc +++ b/tensorflow/compiler/tf2xla/kernels/batch_matmul_op.cc @@ -28,11 +28,10 @@ class BatchMatMulOp : public XlaOpKernel { } void Compile(XlaOpKernelContext* ctx) override { - auto result = BatchDot(ctx->builder(), ctx->Input(0), ctx->Input(1), + auto result = BatchDot(ctx->Input(0), ctx->Input(1), /*transpose_x=*/adj_x_, /*transpose_y=*/adj_y_, /*conjugate_x=*/adj_x_, /*conjugate_y=*/adj_y_); - OP_REQUIRES_OK(ctx, result.status()); - ctx->SetOutput(0, result.ValueOrDie()); + ctx->SetOutput(0, result); } private: diff --git a/tensorflow/compiler/tf2xla/kernels/batch_norm_op.cc b/tensorflow/compiler/tf2xla/kernels/batch_norm_op.cc index 93fbc40461bf6b1c0bb33a2a841f69a2f4188bcd..c4af79281d2162b1dbfb0a7881720892f4bc49d2 100644 --- a/tensorflow/compiler/tf2xla/kernels/batch_norm_op.cc +++ b/tensorflow/compiler/tf2xla/kernels/batch_norm_op.cc @@ -18,6 +18,7 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/xla_helpers.h" #include "tensorflow/compiler/tf2xla/xla_op_kernel.h" #include "tensorflow/compiler/tf2xla/xla_op_registry.h" +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/core/util/tensor_format.h" namespace tensorflow { @@ -49,8 +50,6 @@ class FusedBatchNormOp : public XlaOpKernel { OP_REQUIRES_OK(ctx, DataTypeToPrimitiveType(ctx->input_type(1), &scale_type)); - xla::XlaBuilder* builder = ctx->builder(); - xla::XlaOp input = ctx->Input(0); TensorShape input_shape = ctx->InputShape(0); @@ -60,30 +59,30 @@ class FusedBatchNormOp : public XlaOpKernel { // TODO(b/69928690): support mixed precision in the XLA batch normalization // operators. As a workaround, cast everything to the statistics type (which // may be more precise than the input type). - input = builder->ConvertElementType(input, scale_type); + input = xla::ConvertElementType(input, scale_type); if (is_training_) { - xla::XlaOp output = builder->BatchNormTraining( + xla::XlaOp output = xla::BatchNormTraining( input, ctx->Input(1), ctx->Input(2), epsilon_, feature_index); // In training mode, outputs the normalized value as well as the // calculated mean and variance. - ctx->SetOutput(0, builder->ConvertElementType( - builder->GetTupleElement(output, 0), input_type)); - ctx->SetOutput(1, builder->GetTupleElement(output, 1)); - ctx->SetOutput(2, builder->GetTupleElement(output, 2)); + ctx->SetOutput(0, xla::ConvertElementType(xla::GetTupleElement(output, 0), + input_type)); + ctx->SetOutput(1, xla::GetTupleElement(output, 1)); + ctx->SetOutput(2, xla::GetTupleElement(output, 2)); // Output 3 and 4 for "FusedBatchNorm" are currently marked as "reserved // space 1 & 2". They are used to pass the per-batch mean and // variance to the gradient. Here we maintain the same behavior by setting // them to the mean and variance calculated by BatchNormTraining. - ctx->SetOutput(3, builder->GetTupleElement(output, 1)); - ctx->SetOutput(4, builder->GetTupleElement(output, 2)); + ctx->SetOutput(3, xla::GetTupleElement(output, 1)); + ctx->SetOutput(4, xla::GetTupleElement(output, 2)); } else { - xla::XlaOp output = builder->BatchNormInference( + xla::XlaOp output = xla::BatchNormInference( input, ctx->Input(1), ctx->Input(2), ctx->Input(3), ctx->Input(4), epsilon_, feature_index); - ctx->SetOutput(0, builder->ConvertElementType(output, input_type)); + ctx->SetOutput(0, xla::ConvertElementType(output, input_type)); // Directly send input to output as mean and variance in inference mode. ctx->SetOutput(1, ctx->Input(3)); ctx->SetOutput(2, ctx->Input(4)); @@ -144,12 +143,12 @@ class FusedBatchNormGradOp : public XlaOpKernel { xla::XlaOp offset_backprop; if (is_training_) { xla::XlaOp output = - b->BatchNormGrad(activations, scale, mean, var, grad_backprop, - epsilon_, feature_index); + xla::BatchNormGrad(activations, scale, mean, var, grad_backprop, + epsilon_, feature_index); - x_backprop = b->GetTupleElement(output, 0); - scale_backprop = b->GetTupleElement(output, 1); - offset_backprop = b->GetTupleElement(output, 2); + x_backprop = xla::GetTupleElement(output, 0); + scale_backprop = xla::GetTupleElement(output, 1); + offset_backprop = xla::GetTupleElement(output, 2); } else { // Reduce over all dimensions except the feature dim. std::vector reduction_dims(input_dims - 1); @@ -166,35 +165,35 @@ class FusedBatchNormGradOp : public XlaOpKernel { auto converted = XlaHelpers::ConvertElementType(b, grad_backprop, accumulation_type); auto reduce = - b->Reduce(converted, XlaHelpers::Zero(b, accumulation_type), - *ctx->GetOrCreateAdd(accumulation_type), reduction_dims); + xla::Reduce(converted, XlaHelpers::Zero(b, accumulation_type), + *ctx->GetOrCreateAdd(accumulation_type), reduction_dims); offset_backprop = XlaHelpers::ConvertElementType(b, reduce, scale_dtype); // scratch1 = rsqrt(pop_var + epsilon) auto neg_half = XlaHelpers::FloatLiteral(b, scale_dtype, -0.5); - auto scratch1 = - b->Pow(b->Add(var, b->ConstantR0(epsilon_)), neg_half); + auto scratch1 = xla::Pow( + xla::Add(var, xla::ConstantR0(b, epsilon_)), neg_half); // scratch2 = sum(y_backprop * (x - mean)) auto mul = - b->Mul(grad_backprop, b->Sub(activations, mean, {feature_index})); + xla::Mul(grad_backprop, xla::Sub(activations, mean, {feature_index})); converted = XlaHelpers::ConvertElementType(b, mul, accumulation_type); reduce = - b->Reduce(converted, XlaHelpers::Zero(b, accumulation_type), - *ctx->GetOrCreateAdd(accumulation_type), reduction_dims); + xla::Reduce(converted, XlaHelpers::Zero(b, accumulation_type), + *ctx->GetOrCreateAdd(accumulation_type), reduction_dims); auto scratch2 = XlaHelpers::ConvertElementType(b, reduce, scale_dtype); x_backprop = - b->Mul(grad_backprop, b->Mul(scratch1, scale), {feature_index}); - scale_backprop = b->Mul(scratch1, scratch2); + xla::Mul(grad_backprop, xla::Mul(scratch1, scale), {feature_index}); + scale_backprop = xla::Mul(scratch1, scratch2); } ctx->SetOutput(0, XlaHelpers::ConvertElementType(b, x_backprop, input_dtype)); ctx->SetOutput(1, scale_backprop); ctx->SetOutput(2, offset_backprop); - ctx->SetConstantOutput(3, Tensor(scale_dtype, {})); - ctx->SetConstantOutput(4, Tensor(scale_dtype, {})); + ctx->SetConstantOutput(3, Tensor()); + ctx->SetConstantOutput(4, Tensor()); } private: diff --git a/tensorflow/compiler/tf2xla/kernels/batchtospace_op.cc b/tensorflow/compiler/tf2xla/kernels/batchtospace_op.cc index 642278ab994bf3cc84396f093ed56b009a1435c1..26130fd9e7fce75c6d2a5a53cfc85842cf762b35 100644 --- a/tensorflow/compiler/tf2xla/kernels/batchtospace_op.cc +++ b/tensorflow/compiler/tf2xla/kernels/batchtospace_op.cc @@ -16,6 +16,7 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/xla_helpers.h" #include "tensorflow/compiler/tf2xla/xla_op_kernel.h" #include "tensorflow/compiler/tf2xla/xla_op_registry.h" +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" namespace tensorflow { namespace { @@ -45,7 +46,6 @@ void BatchToSpace(XlaOpKernelContext* ctx, const xla::XlaOp& input, ", 2] instead of ", xla::ShapeUtil::HumanString(crops.shape()))); - xla::XlaBuilder* b = ctx->builder(); const int64 batch_size = input_shape[0]; // Compute the product of the block_shape values. @@ -72,7 +72,7 @@ void BatchToSpace(XlaOpKernelContext* ctx, const xla::XlaOp& input, reshaped_shape[block_rank] = batch_size / block_num_elems; std::copy(input_shape.begin() + 1, input_shape.end(), reshaped_shape.begin() + block_rank + 1); - xla::XlaOp reshaped = b->Reshape(input, reshaped_shape); + xla::XlaOp reshaped = xla::Reshape(input, reshaped_shape); // 2. Permute dimensions of `reshaped` to produce `permuted` of shape // [batch / prod(block_shape), @@ -90,7 +90,7 @@ void BatchToSpace(XlaOpKernelContext* ctx, const xla::XlaOp& input, } std::iota(permutation.begin() + 1 + block_rank * 2, permutation.end(), 1 + block_rank * 2); - xla::XlaOp permuted = b->Transpose(reshaped, permutation); + xla::XlaOp permuted = xla::Transpose(reshaped, permutation); // 3. Reshape `permuted` to produce `reshaped_permuted` of shape // [batch / prod(block_shape), @@ -110,7 +110,8 @@ void BatchToSpace(XlaOpKernelContext* ctx, const xla::XlaOp& input, std::copy(remainder_shape.begin(), remainder_shape.end(), reshaped_permuted_shape.begin() + 1 + block_rank); - xla::XlaOp reshaped_permuted = b->Reshape(permuted, reshaped_permuted_shape); + xla::XlaOp reshaped_permuted = + xla::Reshape(permuted, reshaped_permuted_shape); // 4. Crop the start and end of dimensions `[1, ..., M]` of // `reshaped_permuted` according to `crops` to produce the output of shape: @@ -138,7 +139,7 @@ void BatchToSpace(XlaOpKernelContext* ctx, const xla::XlaOp& input, " end: ", crop_end, " size ", reshaped_permuted_shape[1 + i])); } xla::XlaOp output = - b->Slice(reshaped_permuted, start_indices, end_indices, strides); + xla::Slice(reshaped_permuted, start_indices, end_indices, strides); ctx->SetOutput(0, output); } diff --git a/tensorflow/compiler/tf2xla/kernels/bcast_ops.cc b/tensorflow/compiler/tf2xla/kernels/bcast_ops.cc index ee2c920453c3bbaef2c145df743fddf999167c39..ba3b1c9dab79a387c48e8e25e4804917f328f8a0 100644 --- a/tensorflow/compiler/tf2xla/kernels/bcast_ops.cc +++ b/tensorflow/compiler/tf2xla/kernels/bcast_ops.cc @@ -19,7 +19,7 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/xla_helpers.h" #include "tensorflow/compiler/tf2xla/xla_op_kernel.h" #include "tensorflow/compiler/tf2xla/xla_op_registry.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/core/platform/macros.h" #include "tensorflow/core/platform/types.h" #include "tensorflow/core/util/bcast.h" diff --git a/tensorflow/compiler/tf2xla/kernels/bias_ops.cc b/tensorflow/compiler/tf2xla/kernels/bias_ops.cc index 9d677f426650ea17a49e5ab1401078f04623fe97..e9b2c0b16d39cb3b747c0316621fb01de709b12e 100644 --- a/tensorflow/compiler/tf2xla/kernels/bias_ops.cc +++ b/tensorflow/compiler/tf2xla/kernels/bias_ops.cc @@ -18,6 +18,7 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/xla_helpers.h" #include "tensorflow/compiler/tf2xla/xla_op_kernel.h" #include "tensorflow/compiler/tf2xla/xla_op_registry.h" +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/core/framework/tensor.h" #include "tensorflow/core/framework/tensor_shape.h" #include "tensorflow/core/util/tensor_format.h" @@ -60,8 +61,7 @@ class BiasOp : public XlaOpKernel { "of the input tensor: ", bias_shape.DebugString(), " vs. ", input_shape.DebugString())); - xla::XlaOp result = - ctx->builder()->Add(ctx->Input(0), ctx->Input(1), {feature_dim}); + xla::XlaOp result = xla::Add(ctx->Input(0), ctx->Input(1), {feature_dim}); ctx->SetOutput(0, result); } @@ -109,8 +109,8 @@ class BiasAddGradOp : public XlaOpKernel { auto converted = XlaHelpers::ConvertElementType(b, ctx->Input(0), accumulation_type); auto reduce = - b->Reduce(converted, XlaHelpers::Zero(b, accumulation_type), - *ctx->GetOrCreateAdd(accumulation_type), reduce_dims); + xla::Reduce(converted, XlaHelpers::Zero(b, accumulation_type), + *ctx->GetOrCreateAdd(accumulation_type), reduce_dims); ctx->SetOutput(0, XlaHelpers::ConvertElementType(b, reduce, input_type(0))); } diff --git a/tensorflow/compiler/tf2xla/kernels/binary_ops.cc b/tensorflow/compiler/tf2xla/kernels/binary_ops.cc index fee939bdead2718b16376138029e90b805f87f82..d6d4ae89376b67c14af8ef4f3a608fcc83b6fb59 100644 --- a/tensorflow/compiler/tf2xla/kernels/binary_ops.cc +++ b/tensorflow/compiler/tf2xla/kernels/binary_ops.cc @@ -41,18 +41,19 @@ namespace { const BCast& broadcast_helper, \ const std::vector& extend_dimensions) override { \ xla::XlaBuilder* b = ctx->builder(); \ + (void)b; \ return HLO; \ } \ }; \ REGISTER_XLA_OP(Name(#NAME), NAME##Op) -XLA_MAKE_BINARY(Add, b->Add(lhs, rhs, extend_dimensions)); -XLA_MAKE_BINARY(Sub, b->Sub(lhs, rhs, extend_dimensions)); -XLA_MAKE_BINARY(Mul, b->Mul(lhs, rhs, extend_dimensions)); -XLA_MAKE_BINARY(Div, b->Div(lhs, rhs, extend_dimensions)); +XLA_MAKE_BINARY(Add, xla::Add(lhs, rhs, extend_dimensions)); +XLA_MAKE_BINARY(Sub, xla::Sub(lhs, rhs, extend_dimensions)); +XLA_MAKE_BINARY(Mul, xla::Mul(lhs, rhs, extend_dimensions)); +XLA_MAKE_BINARY(Div, xla::Div(lhs, rhs, extend_dimensions)); -XLA_MAKE_BINARY(Atan2, b->Atan2(lhs, rhs, extend_dimensions)); -XLA_MAKE_BINARY(Complex, b->Complex(lhs, rhs, extend_dimensions)); +XLA_MAKE_BINARY(Atan2, xla::Atan2(lhs, rhs, extend_dimensions)); +XLA_MAKE_BINARY(Complex, xla::Complex(lhs, rhs, extend_dimensions)); // Implementation of FloorDiv. Pseudo-code: // if ((x < 0) != (y < 0)) { @@ -67,13 +68,13 @@ static xla::XlaOp FloorDivImpl(xla::XlaBuilder* b, DataType dtype, xla::XlaOp x, std::tie(x, y) = XlaBinaryOp::Broadcast(b, x, y, broadcast_helper); auto zero = XlaHelpers::Zero(b, dtype); auto one = XlaHelpers::One(b, dtype); - auto different_sign = b->Ne(b->Lt(x, zero), b->Lt(y, zero)); - auto abs_x = b->Abs(x); - auto abs_y = b->Abs(y); - auto t = b->Neg(b->Sub(b->Add(abs_x, abs_y), one)); - auto result = b->Select(different_sign, b->Div(t, abs_y), b->Div(x, y)); + auto different_sign = xla::Ne(xla::Lt(x, zero), xla::Lt(y, zero)); + auto abs_x = xla::Abs(x); + auto abs_y = xla::Abs(y); + auto t = xla::Neg(xla::Sub(xla::Add(abs_x, abs_y), one)); + auto result = xla::Select(different_sign, xla::Div(t, abs_y), xla::Div(x, y)); if (DataTypeIsFloating(dtype)) { - result = b->Floor(result); + result = xla::Floor(result); } return result; } @@ -87,76 +88,78 @@ static xla::XlaOp FloorModImpl(xla::XlaBuilder* b, DataType dtype, xla::XlaOp x, xla::XlaOp y, const BCast& broadcast_helper) { std::tie(x, y) = XlaBinaryOp::Broadcast(b, x, y, broadcast_helper); auto zero = XlaHelpers::Zero(b, dtype); - auto same_sign = b->Eq(b->Lt(x, zero), b->Lt(y, zero)); - auto trunc_mod = b->Rem(x, y); - return b->Select(same_sign, trunc_mod, b->Rem(b->Add(trunc_mod, y), y)); + auto same_sign = xla::Eq(xla::Lt(x, zero), xla::Lt(y, zero)); + auto trunc_mod = xla::Rem(x, y); + return xla::Select(same_sign, trunc_mod, xla::Rem(xla::Add(trunc_mod, y), y)); } XLA_MAKE_BINARY(FloorMod, FloorModImpl(b, input_type(0), lhs, rhs, broadcast_helper)); -XLA_MAKE_BINARY(BitwiseAnd, b->And(lhs, rhs, extend_dimensions)); -XLA_MAKE_BINARY(BitwiseOr, b->Or(lhs, rhs, extend_dimensions)); -XLA_MAKE_BINARY(BitwiseXor, b->Xor(lhs, rhs, extend_dimensions)); +XLA_MAKE_BINARY(BitwiseAnd, xla::And(lhs, rhs, extend_dimensions)); +XLA_MAKE_BINARY(BitwiseOr, xla::Or(lhs, rhs, extend_dimensions)); +XLA_MAKE_BINARY(BitwiseXor, xla::Xor(lhs, rhs, extend_dimensions)); -XLA_MAKE_BINARY(LeftShift, b->ShiftLeft(lhs, rhs, extend_dimensions)); +XLA_MAKE_BINARY(LeftShift, xla::ShiftLeft(lhs, rhs, extend_dimensions)); XLA_MAKE_BINARY(RightShift, (DataTypeIsUnsigned(ctx->input_type(0)) - ? b->ShiftRightLogical(lhs, rhs, extend_dimensions) - : b->ShiftRightArithmetic(lhs, rhs, extend_dimensions))); - -XLA_MAKE_BINARY(LogicalAnd, b->And(lhs, rhs, extend_dimensions)); -XLA_MAKE_BINARY(LogicalOr, b->Or(lhs, rhs, extend_dimensions)); -XLA_MAKE_BINARY(Mod, b->Rem(lhs, rhs, extend_dimensions)); -XLA_MAKE_BINARY(Maximum, b->Max(lhs, rhs, extend_dimensions)); -XLA_MAKE_BINARY(Minimum, b->Min(lhs, rhs, extend_dimensions)); -XLA_MAKE_BINARY(RealDiv, b->Div(lhs, rhs, extend_dimensions)); -XLA_MAKE_BINARY(ReciprocalGrad, b->Neg(b->Mul(rhs, b->Mul(lhs, lhs)))); + ? xla::ShiftRightLogical(lhs, rhs, extend_dimensions) + : xla::ShiftRightArithmetic(lhs, rhs, extend_dimensions))); + +XLA_MAKE_BINARY(LogicalAnd, xla::And(lhs, rhs, extend_dimensions)); +XLA_MAKE_BINARY(LogicalOr, xla::Or(lhs, rhs, extend_dimensions)); +XLA_MAKE_BINARY(Mod, xla::Rem(lhs, rhs, extend_dimensions)); +XLA_MAKE_BINARY(Maximum, xla::Max(lhs, rhs, extend_dimensions)); +XLA_MAKE_BINARY(Minimum, xla::Min(lhs, rhs, extend_dimensions)); +XLA_MAKE_BINARY(RealDiv, xla::Div(lhs, rhs, extend_dimensions)); +XLA_MAKE_BINARY(ReciprocalGrad, xla::Neg(xla::Mul(rhs, xla::Mul(lhs, lhs)))); XLA_MAKE_BINARY( RsqrtGrad, - b->Mul(b->Pow(lhs, XlaHelpers::IntegerLiteral(b, input_type(0), 3)), - b->Div(rhs, XlaHelpers::IntegerLiteral(b, input_type(0), -2)), - extend_dimensions)); -XLA_MAKE_BINARY(SqrtGrad, - b->Div(b->Mul(rhs, - XlaHelpers::FloatLiteral(b, input_type(0), 0.5)), - lhs, extend_dimensions)); + xla::Mul(xla::Pow(lhs, XlaHelpers::IntegerLiteral(b, input_type(0), 3)), + xla::Div(rhs, XlaHelpers::IntegerLiteral(b, input_type(0), -2)), + extend_dimensions)); +XLA_MAKE_BINARY( + SqrtGrad, + xla::Div(xla::Mul(rhs, XlaHelpers::FloatLiteral(b, input_type(0), 0.5)), + lhs, extend_dimensions)); static xla::XlaOp Square(xla::XlaBuilder* builder, const xla::XlaOp& x) { - return builder->Mul(x, x); + return xla::Mul(x, x); } XLA_MAKE_BINARY(SquaredDifference, - Square(b, b->Sub(lhs, rhs, extend_dimensions))); + Square(b, xla::Sub(lhs, rhs, extend_dimensions))); -XLA_MAKE_BINARY(TruncateDiv, b->Div(lhs, rhs, extend_dimensions)); -XLA_MAKE_BINARY(TruncateMod, b->Rem(lhs, rhs, extend_dimensions)); +XLA_MAKE_BINARY(TruncateDiv, xla::Div(lhs, rhs, extend_dimensions)); +XLA_MAKE_BINARY(TruncateMod, xla::Rem(lhs, rhs, extend_dimensions)); // Comparison ops -XLA_MAKE_BINARY(Equal, b->Eq(lhs, rhs, extend_dimensions)); -XLA_MAKE_BINARY(NotEqual, b->Ne(lhs, rhs, extend_dimensions)); -XLA_MAKE_BINARY(Greater, b->Gt(lhs, rhs, extend_dimensions)); -XLA_MAKE_BINARY(GreaterEqual, b->Ge(lhs, rhs, extend_dimensions)); -XLA_MAKE_BINARY(Less, b->Lt(lhs, rhs, extend_dimensions)); -XLA_MAKE_BINARY(LessEqual, b->Le(lhs, rhs, extend_dimensions)); +XLA_MAKE_BINARY(Equal, xla::Eq(lhs, rhs, extend_dimensions)); +XLA_MAKE_BINARY(NotEqual, xla::Ne(lhs, rhs, extend_dimensions)); +XLA_MAKE_BINARY(Greater, xla::Gt(lhs, rhs, extend_dimensions)); +XLA_MAKE_BINARY(GreaterEqual, xla::Ge(lhs, rhs, extend_dimensions)); +XLA_MAKE_BINARY(Less, xla::Lt(lhs, rhs, extend_dimensions)); +XLA_MAKE_BINARY(LessEqual, xla::Le(lhs, rhs, extend_dimensions)); // Non-linear ops XLA_MAKE_BINARY(SigmoidGrad, - b->Mul(b->Mul(rhs, lhs), - b->Sub(XlaHelpers::One(b, input_type(0)), lhs))); + xla::Mul(xla::Mul(rhs, lhs), + xla::Sub(XlaHelpers::One(b, input_type(0)), lhs))); XLA_MAKE_BINARY(SoftplusGrad, - b->Div(lhs, b->Add(b->Exp(b->Neg(rhs)), - XlaHelpers::One(b, input_type(1))))); + xla::Div(lhs, xla::Add(xla::Exp(xla::Neg(rhs)), + XlaHelpers::One(b, input_type(1))))); // softsigngrad(gradients, features) = gradients / (1 + abs(features)) ** 2 XLA_MAKE_BINARY(SoftsignGrad, - b->Div(lhs, Square(b, b->Add(XlaHelpers::One(b, input_type(0)), - b->Abs(rhs))))); + xla::Div(lhs, + Square(b, xla::Add(XlaHelpers::One(b, input_type(0)), + xla::Abs(rhs))))); -XLA_MAKE_BINARY(TanhGrad, b->Mul(rhs, b->Sub(XlaHelpers::One(b, input_type(0)), - b->Mul(lhs, lhs)))); +XLA_MAKE_BINARY(TanhGrad, + xla::Mul(rhs, xla::Sub(XlaHelpers::One(b, input_type(0)), + xla::Mul(lhs, lhs)))); -XLA_MAKE_BINARY(Pow, b->Pow(lhs, rhs, extend_dimensions)); +XLA_MAKE_BINARY(Pow, xla::Pow(lhs, rhs, extend_dimensions)); #undef XLA_MAKE_BINARY @@ -169,12 +172,13 @@ class ApproximateEqualOp : public XlaOpKernel { // Computes the max of the scalar input x and 0. void Compile(XlaOpKernelContext* ctx) override { xla::XlaBuilder* b = ctx->builder(); - auto abs = b->Abs(b->Sub(ctx->Input(0), ctx->Input(1))); + auto abs = xla::Abs(xla::Sub(ctx->Input(0), ctx->Input(1))); auto abs_shape = b->GetShape(abs); OP_REQUIRES_OK(ctx, abs_shape.status()); auto abs_type = abs_shape.ValueOrDie().element_type(); - auto result = b->Lt( - abs, b->ConvertElementType(b->ConstantR0(tolerance_), abs_type)); + auto result = + xla::Lt(abs, xla::ConvertElementType( + xla::ConstantR0(b, tolerance_), abs_type)); ctx->SetOutput(0, result); } diff --git a/tensorflow/compiler/tf2xla/kernels/bucketize_op.cc b/tensorflow/compiler/tf2xla/kernels/bucketize_op.cc index ca9a6b40688d1e8496d1b823e20d273d519f65e8..efbdb76eaaf78904fe783a018940b1b096ec39bd 100644 --- a/tensorflow/compiler/tf2xla/kernels/bucketize_op.cc +++ b/tensorflow/compiler/tf2xla/kernels/bucketize_op.cc @@ -18,6 +18,7 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/xla_op_kernel.h" #include "tensorflow/compiler/tf2xla/xla_op_registry.h" #include "tensorflow/compiler/xla/client/lib/arithmetic.h" +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/core/framework/op_kernel.h" namespace tensorflow { @@ -36,22 +37,22 @@ class BucketizeOp : public XlaOpKernel { const DataType dtype = context->input_type(0); xla::XlaOp input = context->Input(0); - xla::XlaOp boundaries = builder->ConstantR1(boundaries_); + xla::XlaOp boundaries = xla::ConstantR1(builder, boundaries_); // TODO(phawkins): the following behavior matches the behavior of the core // Bucketize kernel. However, comparing an int32 or int64 against float may // lead to inaccurate bucketing due to rounding. if (dtype == DT_DOUBLE) { - input = builder->ConvertElementType(input, xla::F64); - boundaries = builder->ConvertElementType(boundaries, xla::F64); + input = xla::ConvertElementType(input, xla::F64); + boundaries = xla::ConvertElementType(boundaries, xla::F64); } else { - input = builder->ConvertElementType(input, xla::F32); + input = xla::ConvertElementType(input, xla::F32); } - xla::XlaOp comparison = builder->ConvertElementType( - builder->Ge(builder->Broadcast(input, {1}), boundaries, - /*broadcast_dimensions=*/{0}), - xla::S32); - xla::XlaOp buckets = builder->Reduce( - comparison, /*init_value=*/builder->ConstantR0(0), + xla::XlaOp comparison = + xla::ConvertElementType(xla::Ge(xla::Broadcast(input, {1}), boundaries, + /*broadcast_dimensions=*/{0}), + xla::S32); + xla::XlaOp buckets = xla::Reduce( + comparison, /*init_value=*/xla::ConstantR0(builder, 0), /*computation=*/xla::CreateScalarAddComputation(xla::S32, builder), /*dimensions_to_reduce=*/{0}); context->SetOutput(0, buckets); diff --git a/tensorflow/compiler/tf2xla/kernels/cast_op.cc b/tensorflow/compiler/tf2xla/kernels/cast_op.cc index e9d98c768572c52825fa5192ecec834889f040fe..62eebf762b3e063da8ec456cc4726d3cc9b77d1d 100644 --- a/tensorflow/compiler/tf2xla/kernels/cast_op.cc +++ b/tensorflow/compiler/tf2xla/kernels/cast_op.cc @@ -17,6 +17,7 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/xla_helpers.h" #include "tensorflow/compiler/tf2xla/xla_op_kernel.h" #include "tensorflow/compiler/tf2xla/xla_op_registry.h" +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/compiler/xla/primitive_util.h" #include "tensorflow/core/framework/kernel_def_builder.h" @@ -40,14 +41,14 @@ class CastOp : public XlaOpKernel { if (src_dtype_ == dst_dtype_) { output = input; } else if (dst_dtype_ == DT_BOOL) { - output = builder->Ne(input, XlaHelpers::Zero(builder, src_dtype_)); + output = xla::Ne(input, XlaHelpers::Zero(builder, src_dtype_)); } else if (xla::primitive_util::IsComplexType(src_type_) && !xla::primitive_util::IsComplexType(dst_type_)) { // As in cast_op.h, we replicate the numpy behavior of truncating the // imaginary part. - output = builder->ConvertElementType(builder->Real(input), dst_type_); + output = xla::ConvertElementType(xla::Real(input), dst_type_); } else { - output = builder->ConvertElementType(input, dst_type_); + output = xla::ConvertElementType(input, dst_type_); } ctx->SetOutput(0, output); @@ -72,7 +73,6 @@ class BitcastOp : public XlaOpKernel { } void Compile(XlaOpKernelContext* ctx) override { - xla::XlaBuilder* builder = ctx->builder(); xla::XlaOp input = ctx->Input(0); xla::XlaOp output; @@ -92,7 +92,7 @@ class BitcastOp : public XlaOpKernel { xla::primitive_util::BitWidth(dst_type_), errors::Unimplemented( "Only bitcasts between equally sized types supported.")); - output = builder->BitcastConvertType(input, dst_type_); + output = xla::BitcastConvertType(input, dst_type_); } ctx->SetOutput(0, output); diff --git a/tensorflow/compiler/tf2xla/kernels/categorical_op.cc b/tensorflow/compiler/tf2xla/kernels/categorical_op.cc index 835a7f568945f0bee86fe2b39491c3326726e1aa..1784e712b56145bbdff5f1daa2e031b65d0774b6 100644 --- a/tensorflow/compiler/tf2xla/kernels/categorical_op.cc +++ b/tensorflow/compiler/tf2xla/kernels/categorical_op.cc @@ -21,6 +21,7 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/xla_op_kernel.h" #include "tensorflow/compiler/tf2xla/xla_op_registry.h" #include "tensorflow/compiler/xla/client/lib/arithmetic.h" +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/core/framework/op_kernel.h" #include "tensorflow/core/framework/tensor.h" #include "tensorflow/core/framework/tensor_shape.h" @@ -65,24 +66,22 @@ class CategoricalOp : public XlaOpKernel { DataTypeToPrimitiveType(input_type(0), &uniform_xla_type)); xla::Shape uniform_shape = xla::ShapeUtil::MakeShape(uniform_xla_type, uniform_shape_array); - auto uniforms = builder->RngUniform( - XlaHelpers::Zero(builder, input_type(0)), - XlaHelpers::One(builder, input_type(0)), uniform_shape); + auto uniforms = + xla::RngUniform(XlaHelpers::Zero(builder, input_type(0)), + XlaHelpers::One(builder, input_type(0)), uniform_shape); // Use Gumbel softmax trick to generate categorical samples. // See: // https://hips.seas.harvard.edu/blog/2013/04/06/the-gumbel-max-trick-for-discrete-distributions/ // TODO(b/68769470): Switch to using a cumulative sum approach. - auto softmax_entries = - builder->Sub(logits, builder->Log(builder->Neg(builder->Log(uniforms))), - /*broadcast_dimensions=*/{0, 2}); - - TensorShape softmax_shape(uniform_shape_array); - xla::XlaOp argmax; - OP_REQUIRES_OK( - ctx, - XlaHelpers::ArgMax(builder, ctx, softmax_entries, softmax_shape, - input_type(0), output_type(0), /*axis=*/2, &argmax)); + auto softmax_entries = xla::Sub(logits, xla::Log(-xla::Log(uniforms)), + /*broadcast_dimensions=*/{0, 2}); + + xla::PrimitiveType xla_output_type; + OP_REQUIRES_OK(ctx, + DataTypeToPrimitiveType(output_type(0), &xla_output_type)); + xla::XlaOp argmax = + XlaHelpers::ArgMax(softmax_entries, xla_output_type, /*axis=*/2); ctx->SetOutput(0, argmax); } diff --git a/tensorflow/compiler/tf2xla/kernels/cholesky_op.cc b/tensorflow/compiler/tf2xla/kernels/cholesky_op.cc index fe6651793dc763d13f4a4b0ac294ec3ecf64af8f..9fcbc86adc0967cbb7fb73da8bdabc58b60953da 100644 --- a/tensorflow/compiler/tf2xla/kernels/cholesky_op.cc +++ b/tensorflow/compiler/tf2xla/kernels/cholesky_op.cc @@ -24,12 +24,7 @@ class CholeskyOp : public XlaOpKernel { public: explicit CholeskyOp(OpKernelConstruction* ctx) : XlaOpKernel(ctx) {} void Compile(XlaOpKernelContext* ctx) override { - auto result = Cholesky(ctx->builder(), ctx->Input(0)); - if (!result.ok()) { - ctx->SetStatus(result.status()); - return; - } - ctx->SetOutput(0, result.ValueOrDie()); + ctx->SetOutput(0, Cholesky(ctx->Input(0))); } }; diff --git a/tensorflow/compiler/tf2xla/kernels/clip_by_value_op.cc b/tensorflow/compiler/tf2xla/kernels/clip_by_value_op.cc index a00bc912f9f40052565446c6bf9390629af9a4cd..4e6d33304c4ae08a0fd1e0a8373267a527087528 100644 --- a/tensorflow/compiler/tf2xla/kernels/clip_by_value_op.cc +++ b/tensorflow/compiler/tf2xla/kernels/clip_by_value_op.cc @@ -15,6 +15,7 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/xla_op_kernel.h" #include "tensorflow/compiler/tf2xla/xla_op_registry.h" +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/core/framework/tensor_shape.h" namespace tensorflow { @@ -29,7 +30,6 @@ class ClipByValueOp : public XlaOpKernel { const TensorShape min_shape = ctx->InputShape(1); const TensorShape max_shape = ctx->InputShape(2); - xla::XlaBuilder* builder = ctx->builder(); auto input = ctx->Input(0); auto min = ctx->Input(1); auto max = ctx->Input(2); @@ -45,13 +45,13 @@ class ClipByValueOp : public XlaOpKernel { if (shape != min_shape) { OP_REQUIRES(ctx, TensorShapeUtils::IsScalar(min_shape), shape_error()); - min = builder->Broadcast(min, shape.dim_sizes()); + min = xla::Broadcast(min, shape.dim_sizes()); } if (shape != max_shape) { OP_REQUIRES(ctx, TensorShapeUtils::IsScalar(max_shape), shape_error()); - max = builder->Broadcast(max, shape.dim_sizes()); + max = xla::Broadcast(max, shape.dim_sizes()); } - ctx->SetOutput(0, builder->Clamp(min, input, max)); + ctx->SetOutput(0, xla::Clamp(min, input, max)); } }; diff --git a/tensorflow/compiler/tf2xla/kernels/concat_op.cc b/tensorflow/compiler/tf2xla/kernels/concat_op.cc index 78285affa1c399ae107a9172fb85cf257457c368..e3a32a5c0e2f93237c8c7ebeea3668b5d1ab6c23 100644 --- a/tensorflow/compiler/tf2xla/kernels/concat_op.cc +++ b/tensorflow/compiler/tf2xla/kernels/concat_op.cc @@ -22,6 +22,7 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/xla_helpers.h" #include "tensorflow/compiler/tf2xla/xla_op_kernel.h" #include "tensorflow/compiler/tf2xla/xla_op_registry.h" +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/compiler/xla/literal_util.h" #include "tensorflow/core/framework/op_kernel.h" #include "tensorflow/core/framework/register_types.h" @@ -88,7 +89,7 @@ class ConcatBaseOp : public XlaOpKernel { "] = ", in_shape.DebugString())); if (in_shape.dims() == 0) { // Inputs that come in as scalars must be reshaped to 1-vectors. - input_data.push_back(ctx->builder()->Reshape(handle, {1})); + input_data.push_back(xla::Reshape(handle, {1})); } else { input_data.push_back(handle); } @@ -96,7 +97,7 @@ class ConcatBaseOp : public XlaOpKernel { } VLOG(1) << "Concat dim " << concat_dim << " equivalent to " << axis; - ctx->SetOutput(0, ctx->builder()->ConcatInDim(input_data, axis)); + ctx->SetOutput(0, xla::ConcatInDim(ctx->builder(), input_data, axis)); } private: diff --git a/tensorflow/compiler/tf2xla/kernels/const_op.cc b/tensorflow/compiler/tf2xla/kernels/const_op.cc index 59d06c654de18c9003fe0bdc706d0c2443de6d7b..f4360d8c3f6fc4007c31fdcfd7f7634de15c76d4 100644 --- a/tensorflow/compiler/tf2xla/kernels/const_op.cc +++ b/tensorflow/compiler/tf2xla/kernels/const_op.cc @@ -17,6 +17,7 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/xla_compiler.h" #include "tensorflow/compiler/tf2xla/xla_op_kernel.h" #include "tensorflow/compiler/tf2xla/xla_op_registry.h" +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/core/framework/kernel_def_builder.h" #include "tensorflow/core/framework/tensor.pb.h" @@ -53,41 +54,41 @@ class ConstOp : public XlaOpKernel { switch (proto_.dtype()) { case DT_BOOL: if (proto_.bool_val_size() == 1) { - ctx->SetOutput(0, - b->Broadcast(b->ConstantR0(proto_.bool_val(0)), - shape.dim_sizes())); + ctx->SetOutput( + 0, xla::Broadcast(xla::ConstantR0(b, proto_.bool_val(0)), + shape.dim_sizes())); return; } break; case DT_FLOAT: if (proto_.float_val_size() == 1) { - ctx->SetOutput( - 0, b->Broadcast(b->ConstantR0(proto_.float_val(0)), - shape.dim_sizes())); + ctx->SetOutput(0, xla::Broadcast(xla::ConstantR0( + b, proto_.float_val(0)), + shape.dim_sizes())); return; } break; case DT_DOUBLE: if (proto_.double_val_size() == 1) { - ctx->SetOutput( - 0, b->Broadcast(b->ConstantR0(proto_.double_val(0)), - shape.dim_sizes())); + ctx->SetOutput(0, xla::Broadcast(xla::ConstantR0( + b, proto_.double_val(0)), + shape.dim_sizes())); return; } break; case DT_INT32: if (proto_.int_val_size() == 1) { - ctx->SetOutput(0, - b->Broadcast(b->ConstantR0(proto_.int_val(0)), - shape.dim_sizes())); + ctx->SetOutput( + 0, xla::Broadcast(xla::ConstantR0(b, proto_.int_val(0)), + shape.dim_sizes())); return; } break; case DT_INT64: if (proto_.int64_val_size() == 1) { - ctx->SetOutput( - 0, b->Broadcast(b->ConstantR0(proto_.int64_val(0)), - shape.dim_sizes())); + ctx->SetOutput(0, xla::Broadcast(xla::ConstantR0( + b, proto_.int64_val(0)), + shape.dim_sizes())); return; } break; diff --git a/tensorflow/compiler/tf2xla/kernels/conv_ops.cc b/tensorflow/compiler/tf2xla/kernels/conv_ops.cc index 627bad12f33c82e91bc3c6f3323f562bc8174056..48ac4867edcef97be001a24f42f6a35225d466c9 100644 --- a/tensorflow/compiler/tf2xla/kernels/conv_ops.cc +++ b/tensorflow/compiler/tf2xla/kernels/conv_ops.cc @@ -18,6 +18,8 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/xla_helpers.h" #include "tensorflow/compiler/tf2xla/xla_op_kernel.h" #include "tensorflow/compiler/tf2xla/xla_op_registry.h" +#include "tensorflow/compiler/xla/client/lib/numeric.h" +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/compiler/xla/literal_util.h" #include "tensorflow/core/framework/numeric_op.h" #include "tensorflow/core/framework/op_kernel.h" @@ -51,8 +53,8 @@ xla::XlaOp CreateExpandedZero(const TensorShape& filter_shape, DataType dtype, xla::XlaBuilder* builder) { TensorShape expanded_filter_shape = ExpandedFilterShapeForDepthwiseConvolution(filter_shape); - return builder->Broadcast(XlaHelpers::Zero(builder, dtype), - expanded_filter_shape.dim_sizes()); + return xla::Broadcast(XlaHelpers::Zero(builder, dtype), + expanded_filter_shape.dim_sizes()); } // Create a mask for depthwise convolution that will make a normal convolution @@ -95,32 +97,27 @@ xla::XlaOp CreateExpandedFilterMask(const TensorShape& filter_shape, // Create a M sized linspace and an M*N sized linspace that will be // broadcasted into perpendicular dimensions and compared. - xla::XlaOp input_feature_iota; - // DT_INT32 Iota will always return status::OK(). - TF_CHECK_OK(XlaHelpers::Iota(builder, DataType::DT_INT32, input_feature, - &input_feature_iota)); - xla::XlaOp expanded_feature_iota; - TF_CHECK_OK(XlaHelpers::Iota(builder, DataType::DT_INT32, - input_feature * depthwise_multiplier, - &expanded_feature_iota)); + xla::XlaOp input_feature_iota = xla::Iota(builder, xla::S32, input_feature); + xla::XlaOp expanded_feature_iota = + xla::Iota(builder, xla::S32, input_feature * depthwise_multiplier); // Divide the M*N sized linspace by the depthwise_multiplier to create // [0 0 1 1 2 2] in the example in the function comment. expanded_feature_iota = - builder->Div(expanded_feature_iota, - XlaHelpers::IntegerLiteral(builder, DataType::DT_INT32, - depthwise_multiplier)); + xla::Div(expanded_feature_iota, + XlaHelpers::IntegerLiteral(builder, DataType::DT_INT32, + depthwise_multiplier)); // Broadcast the N*M linspace to [H, W, ..., M, M*N]. auto expanded_feature_broadcast_dims = expanded_filter_shape.dim_sizes(); expanded_feature_broadcast_dims.pop_back(); - auto broadcasted_expanded_feature_iota = builder->Broadcast( - expanded_feature_iota, expanded_feature_broadcast_dims); + auto broadcasted_expanded_feature_iota = + xla::Broadcast(expanded_feature_iota, expanded_feature_broadcast_dims); // Compare the broadcasted linspace to the input feature linspace in the // input feature dimension to create a diagonal predicate. - return builder->Eq(broadcasted_expanded_feature_iota, input_feature_iota, - {expanded_filter_shape.dims() - 2}); + return xla::Eq(broadcasted_expanded_feature_iota, input_feature_iota, + {expanded_filter_shape.dims() - 2}); } // Expands a filter of shape [H, W, ..., M, N] to [H, W, ..., M, M*N] by adding @@ -142,16 +139,16 @@ xla::XlaOp ExpandFilterForDepthwiseConvolution(const TensorShape& filter_shape, implicit_broadcast_filter_shape.dims() - 1, depthwise_multiplier * input_feature); auto implicit_broadcast_filter = - builder->Reshape(filter, implicit_broadcast_filter_shape.dim_sizes()); + xla::Reshape(filter, implicit_broadcast_filter_shape.dim_sizes()); // Broadcast the filter to [H, W, ..., M, M*N]. auto expanded_zero = CreateExpandedZero(filter_shape, dtype, builder); - auto expanded_filter = builder->Add(implicit_broadcast_filter, expanded_zero); + auto expanded_filter = xla::Add(implicit_broadcast_filter, expanded_zero); // If the filter mask is set, choose the broadcasted filter, othwerwise, // choose zero. - return builder->Select(CreateExpandedFilterMask(filter_shape, builder), - expanded_filter, expanded_zero); + return xla::Select(CreateExpandedFilterMask(filter_shape, builder), + expanded_filter, expanded_zero); } // Inverse of ExpandFilterForDepthwiseConvolution. @@ -162,17 +159,17 @@ xla::XlaOp ContractFilterForDepthwiseBackprop(XlaOpKernelContext* ctx, xla::XlaBuilder* builder) { TensorShape expanded_filter_shape = ExpandedFilterShapeForDepthwiseConvolution(filter_shape); - auto masked_expanded_filter = builder->Select( + auto masked_expanded_filter = xla::Select( CreateExpandedFilterMask(filter_shape, builder), filter_backprop, CreateExpandedZero(filter_shape, dtype, builder)); - return builder->Reshape( + return xla::Reshape( // This reduce does not need inputs to be converted with // XlaHelpers::SumAccumulationType() since the ExpandedFilterMask with // ExpandedZero guarantees that only one element is non zero, so there // cannot be accumulated precision error. - builder->Reduce(masked_expanded_filter, XlaHelpers::Zero(builder, dtype), - *ctx->GetOrCreateAdd(dtype), - {expanded_filter_shape.dims() - 2}), + xla::Reduce(masked_expanded_filter, XlaHelpers::Zero(builder, dtype), + *ctx->GetOrCreateAdd(dtype), + {expanded_filter_shape.dims() - 2}), filter_shape.dim_sizes()); } @@ -289,8 +286,8 @@ class ConvOp : public XlaOpKernel { } xla::XlaOp conv = - b->ConvGeneralDilated(ctx->Input(0), filter, window_strides, padding, - lhs_dilation, rhs_dilation, dims); + xla::ConvGeneralDilated(ctx->Input(0), filter, window_strides, padding, + lhs_dilation, rhs_dilation, dims); ctx->SetOutput(0, conv); } @@ -435,11 +432,11 @@ class ConvBackpropInputOp : public XlaOpKernel { } // Mirror the filter in the spatial dimensions. - xla::XlaOp mirrored_weights = b->Rev(filter, kernel_spatial_dims); + xla::XlaOp mirrored_weights = xla::Rev(filter, kernel_spatial_dims); // activation gradients // = gradients (with padding and dilation) mirrored_weights - xla::XlaOp in_backprop = b->ConvGeneralDilated( + xla::XlaOp in_backprop = xla::ConvGeneralDilated( out_backprop, mirrored_weights, /*window_strides=*/ones, padding, lhs_dilation, rhs_dilation, dnums); @@ -638,8 +635,8 @@ class ConvBackpropFilterOp : public XlaOpKernel { // This is done by specifying the window dilation factors in the // convolution HLO below. auto filter_backprop = - b->ConvGeneralDilated(activations, gradients, window_strides, padding, - /*lhs_dilation=*/ones, rhs_dilation, dnums); + xla::ConvGeneralDilated(activations, gradients, window_strides, padding, + /*lhs_dilation=*/ones, rhs_dilation, dnums); if (depthwise_) { filter_backprop = ContractFilterForDepthwiseBackprop( diff --git a/tensorflow/compiler/tf2xla/kernels/cross_op.cc b/tensorflow/compiler/tf2xla/kernels/cross_op.cc index 7fcd4170fb79a574663c1abffe873d4b53f471d3..500a564f3f0489a42dbc9d5b70ae7708a7a43973 100644 --- a/tensorflow/compiler/tf2xla/kernels/cross_op.cc +++ b/tensorflow/compiler/tf2xla/kernels/cross_op.cc @@ -16,6 +16,7 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/xla_helpers.h" #include "tensorflow/compiler/tf2xla/xla_op_kernel.h" #include "tensorflow/compiler/tf2xla/xla_op_registry.h" +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" namespace tensorflow { namespace { @@ -58,21 +59,21 @@ class CrossOp : public XlaOpKernel { auto in1 = ctx->Input(1); starts.back() = 0; limits.back() = 1; - auto u1 = b->Slice(in0, starts, limits, strides); - auto v1 = b->Slice(in1, starts, limits, strides); + auto u1 = xla::Slice(in0, starts, limits, strides); + auto v1 = xla::Slice(in1, starts, limits, strides); starts.back() = 1; limits.back() = 2; - auto u2 = b->Slice(in0, starts, limits, strides); - auto v2 = b->Slice(in1, starts, limits, strides); + auto u2 = xla::Slice(in0, starts, limits, strides); + auto v2 = xla::Slice(in1, starts, limits, strides); starts.back() = 2; limits.back() = 3; - auto u3 = b->Slice(in0, starts, limits, strides); - auto v3 = b->Slice(in1, starts, limits, strides); + auto u3 = xla::Slice(in0, starts, limits, strides); + auto v3 = xla::Slice(in1, starts, limits, strides); - auto s1 = b->Sub(b->Mul(u2, v3), b->Mul(u3, v2)); - auto s2 = b->Sub(b->Mul(u3, v1), b->Mul(u1, v3)); - auto s3 = b->Sub(b->Mul(u1, v2), b->Mul(u2, v1)); - auto output = b->ConcatInDim({s1, s2, s3}, in0_shape.dims() - 1); + auto s1 = xla::Sub(xla::Mul(u2, v3), xla::Mul(u3, v2)); + auto s2 = xla::Sub(xla::Mul(u3, v1), xla::Mul(u1, v3)); + auto s3 = xla::Sub(xla::Mul(u1, v2), xla::Mul(u2, v1)); + auto output = xla::ConcatInDim(b, {s1, s2, s3}, in0_shape.dims() - 1); ctx->SetOutput(0, output); } diff --git a/tensorflow/compiler/tf2xla/kernels/cwise_ops.cc b/tensorflow/compiler/tf2xla/kernels/cwise_ops.cc index 01aa1a83e7967921f1583b3ef18ec57e452dcfea..9ff3e0222831cb4339943966810eeae451e47a2c 100644 --- a/tensorflow/compiler/tf2xla/kernels/cwise_ops.cc +++ b/tensorflow/compiler/tf2xla/kernels/cwise_ops.cc @@ -96,18 +96,16 @@ void XlaBinaryOp::Compile(XlaOpKernelContext* ctx) { // First reshape the inputs, which should be a metadata-only // operation since we are flattening the dimensions in order. - auto lhs_shaped = builder->Reshape(lhs, broadcast_helper.x_reshape()); - auto rhs_shaped = builder->Reshape(rhs, broadcast_helper.y_reshape()); + auto lhs_shaped = xla::Reshape(lhs, broadcast_helper.x_reshape()); + auto rhs_shaped = xla::Reshape(rhs, broadcast_helper.y_reshape()); // Next broadcast the necessary input dimensions. We rely on the // XLA optimizer to be smart about the fact that we are asking // it to broadcast size 1 on some of these dimensions, to avoid // adding complexity to this code. - auto lhs_broadcast = - builder->Broadcast(lhs_shaped, broadcast_helper.x_bcast()); + auto lhs_broadcast = xla::Broadcast(lhs_shaped, broadcast_helper.x_bcast()); int lhs_size = broadcast_helper.x_bcast().size(); - auto rhs_broadcast = - builder->Broadcast(rhs_shaped, broadcast_helper.y_bcast()); + auto rhs_broadcast = xla::Broadcast(rhs_shaped, broadcast_helper.y_bcast()); int rhs_size = broadcast_helper.y_bcast().size(); // Now reshape them to the correct output shape. After the @@ -122,15 +120,15 @@ void XlaBinaryOp::Compile(XlaOpKernelContext* ctx) { lhs_reorder.push_back(i); lhs_reorder.push_back(i + lhs_size); } - auto lhs_output = builder->Reshape(lhs_broadcast, lhs_reorder, - broadcast_helper.output_shape()); + auto lhs_output = + xla::Reshape(lhs_broadcast, lhs_reorder, broadcast_helper.output_shape()); std::vector rhs_reorder; for (int i = 0; i < rhs_size; ++i) { rhs_reorder.push_back(i); rhs_reorder.push_back(i + rhs_size); } - auto rhs_output = builder->Reshape(rhs_broadcast, rhs_reorder, - broadcast_helper.output_shape()); + auto rhs_output = + xla::Reshape(rhs_broadcast, rhs_reorder, broadcast_helper.output_shape()); return {lhs_output, rhs_output}; } diff --git a/tensorflow/compiler/tf2xla/kernels/depthtospace_op.cc b/tensorflow/compiler/tf2xla/kernels/depthtospace_op.cc index 23243f62462c6315e359d9621823b19fc98c6218..f3149200250935629a6e4bf67bff0c048135ce3e 100644 --- a/tensorflow/compiler/tf2xla/kernels/depthtospace_op.cc +++ b/tensorflow/compiler/tf2xla/kernels/depthtospace_op.cc @@ -16,6 +16,7 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/xla_helpers.h" #include "tensorflow/compiler/tf2xla/xla_op_kernel.h" #include "tensorflow/compiler/tf2xla/xla_op_registry.h" +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/core/util/tensor_format.h" namespace tensorflow { @@ -50,7 +51,6 @@ class DepthToSpaceOp : public XlaOpKernel { const gtl::InlinedVector input_shape = input_tensor_shape.dim_sizes(); - xla::XlaBuilder* b = ctx->builder(); xla::XlaOp input = ctx->Input(0); int feature_dim = GetTensorFeatureDimIndex(input_rank, data_format_); @@ -130,7 +130,7 @@ class DepthToSpaceOp : public XlaOpKernel { ") is not divisible by square of the block size (", block_size_, ")")); - xla::XlaOp reshaped = b->Reshape(input, reshaped_shape); + xla::XlaOp reshaped = xla::Reshape(input, reshaped_shape); // 2. Permute dimensions of `reshaped` to produce // `permuted_reshaped` of shape: @@ -141,7 +141,7 @@ class DepthToSpaceOp : public XlaOpKernel { // input_shape[2], // block_size_, // depth / (block_size_ * block_size_)] - xla::XlaOp permuted_reshaped = b->Transpose(reshaped, transpose_order); + xla::XlaOp permuted_reshaped = xla::Transpose(reshaped, transpose_order); // 3. Reshape `permuted_reshaped` to flatten `block_shape` into the // batch dimension, producing an output tensor of shape: @@ -151,7 +151,7 @@ class DepthToSpaceOp : public XlaOpKernel { // input_shape[2] * block_size_, // depth / (block_size_ * block_size_)] // - xla::XlaOp output = b->Reshape(permuted_reshaped, output_shape); + xla::XlaOp output = xla::Reshape(permuted_reshaped, output_shape); ctx->SetOutput(0, output); } diff --git a/tensorflow/compiler/tf2xla/kernels/diag_op.cc b/tensorflow/compiler/tf2xla/kernels/diag_op.cc index 931705ba837153e1175cd9a209876ef5ec93f0fc..6dec414c53bee6b0102e229c86cfafb4072a35f0 100644 --- a/tensorflow/compiler/tf2xla/kernels/diag_op.cc +++ b/tensorflow/compiler/tf2xla/kernels/diag_op.cc @@ -18,6 +18,9 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/xla_helpers.h" #include "tensorflow/compiler/tf2xla/xla_op_kernel.h" #include "tensorflow/compiler/tf2xla/xla_op_registry.h" +#include "tensorflow/compiler/xla/client/lib/constants.h" +#include "tensorflow/compiler/xla/client/lib/numeric.h" +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/compiler/xla/util.h" #include "tensorflow/core/framework/op_kernel.h" @@ -25,10 +28,10 @@ namespace tensorflow { namespace { // Create a diagonal / batch diagonal matrix with 'input' on the diagonal. -xla::StatusOr CreateDiagonal( - const xla::XlaOp& input, int64 last_dim_size, - tensorflow::gtl::ArraySlice other_dims, XlaOpKernelContext* ctx, - xla::XlaBuilder* builder) { +xla::XlaOp CreateDiagonal(xla::XlaOp input, int64 last_dim_size, + gtl::ArraySlice other_dims, + xla::PrimitiveType element_type) { + xla::XlaBuilder* builder = input.builder(); // Create two matrices that have the following forms, and compare them: // // [[0, 0, 0, 0] [[0, 1, 2, 3] @@ -38,16 +41,14 @@ xla::StatusOr CreateDiagonal( // // This produces a predicate matrix of the right size, with "true" on the // diagonal. - xla::XlaOp iota; - TF_RETURN_IF_ERROR( - XlaHelpers::Iota(builder, DataType::DT_INT32, last_dim_size, &iota)); - xla::XlaOp iota_broadcast = builder->Broadcast(iota, {last_dim_size}); - xla::XlaOp mask = builder->Eq(iota_broadcast, iota, {0}); + xla::XlaOp iota = xla::Iota(builder, xla::S32, last_dim_size); + xla::XlaOp iota_broadcast = xla::Broadcast(iota, {last_dim_size}); + xla::XlaOp mask = xla::Eq(iota_broadcast, iota, {0}); // If this is a batched diagonal, broadcast the mask across the other // dimensions. if (!other_dims.empty()) { - mask = builder->Broadcast(mask, other_dims); + mask = xla::Broadcast(mask, other_dims); } // Broadcast the input, and then use the mask computed above to select the @@ -64,18 +65,15 @@ xla::StatusOr CreateDiagonal( std::vector broadcast_dims(other_dims.begin(), other_dims.end()); broadcast_dims.push_back(1LL); broadcast_dims.push_back(last_dim_size); - xla::XlaOp input_broadcast = builder->Reshape(input, broadcast_dims); + xla::XlaOp input_broadcast = xla::Reshape(input, broadcast_dims); broadcast_dims[broadcast_dims.size() - 2] = last_dim_size; - xla::PrimitiveType element_type; - TF_RETURN_IF_ERROR( - DataTypeToPrimitiveType(ctx->input_type(0), &element_type)); auto broadcast_shape = xla::ShapeUtil::MakeShape(element_type, broadcast_dims); - xla::XlaOp zeros = Zeros(builder, broadcast_shape); + xla::XlaOp zeros = xla::Zeros(builder, broadcast_shape); - input_broadcast = builder->Add(input_broadcast, zeros); - return builder->Select(mask, input_broadcast, zeros); + input_broadcast = xla::Add(input_broadcast, zeros); + return xla::Select(mask, input_broadcast, zeros); } class DiagOp : public XlaOpKernel { @@ -83,8 +81,6 @@ class DiagOp : public XlaOpKernel { explicit DiagOp(OpKernelConstruction* ctx) : XlaOpKernel(ctx) {} void Compile(XlaOpKernelContext* ctx) override { - xla::XlaBuilder* builder = ctx->builder(); - OP_REQUIRES(ctx, ctx->num_inputs() >= 1, errors::InvalidArgument("Diag op must have at an input")); const TensorShape input_shape = ctx->InputShape(0); @@ -104,19 +100,17 @@ class DiagOp : public XlaOpKernel { // Flattens the input to 1D. int64 size = input_shape.num_elements(); - input = builder->Reshape(input, {size}); + input = xla::Reshape(input, {size}); // Create an R2 with the R1 diagonal. - auto diag_or_status = - CreateDiagonal(input, size, /*other_dims=*/{}, ctx, builder); - OP_REQUIRES_OK(ctx, diag_or_status.status()); - xla::XlaOp diag = diag_or_status.ValueOrDie(); + xla::XlaOp diag = + CreateDiagonal(input, size, /*other_dims=*/{}, ctx->input_xla_type(0)); // Reshapes to the final shape. std::vector new_dims(dims.size() * 2); std::copy(dims.begin(), dims.end(), new_dims.begin()); std::copy(dims.begin(), dims.end(), new_dims.begin() + dims.size()); - diag = builder->Reshape(diag, new_dims); + diag = xla::Reshape(diag, new_dims); ctx->SetOutput(0, diag); } @@ -170,21 +164,21 @@ class DiagPartOp : public XlaOpKernel { // Flattens the input to 1D. int64 size = input_shape.num_elements(); - diag = builder->Reshape(diag, {size}); + diag = xla::Reshape(diag, {size}); // Adds padding after the last element of 'new_size'. xla::PaddingConfig config; auto* dim = config.add_dimensions(); dim->set_edge_padding_high(new_size); auto zero = XlaHelpers::Zero(builder, input_type(0)); - diag = builder->Pad(diag, zero, config); + diag = xla::Pad(diag, zero, config); // Reshapes so the diagonal is now in the first column. - diag = builder->Reshape(diag, {new_size, new_size + 1}); + diag = xla::Reshape(diag, {new_size, new_size + 1}); // Slices out the first column and reshapes to the final shape. - diag = builder->Slice(diag, {0, 0}, {new_size, 1}, {1, 1}); - diag = builder->Reshape(diag, new_dims); + diag = xla::Slice(diag, {0, 0}, {new_size, 1}, {1, 1}); + diag = xla::Reshape(diag, new_dims); ctx->SetOutput(0, diag); } @@ -197,8 +191,6 @@ class MatrixDiagOp : public XlaOpKernel { explicit MatrixDiagOp(OpKernelConstruction* ctx) : XlaOpKernel(ctx) {} void Compile(XlaOpKernelContext* ctx) override { - xla::XlaBuilder* builder = ctx->builder(); - OP_REQUIRES(ctx, ctx->num_inputs() >= 1, errors::InvalidArgument("MatrixDiag op must have at an input")); const TensorShape input_shape = ctx->InputShape(0); @@ -208,17 +200,15 @@ class MatrixDiagOp : public XlaOpKernel { errors::InvalidArgument("Expected 1 <= dims, got shape ", input_shape.DebugString())); - xla::XlaOp diag = ctx->Input(0); int last_dim = dims.size() - 1; int64 last_dim_size = input_shape.dim_size(last_dim); tensorflow::gtl::ArraySlice other_dims(dims); other_dims.pop_back(); - auto diag_or_status = - CreateDiagonal(diag, last_dim_size, other_dims, ctx, builder); - OP_REQUIRES_OK(ctx, diag_or_status.status()); - diag = diag_or_status.ValueOrDie(); + xla::XlaOp input = ctx->Input(0); + xla::XlaOp diag = CreateDiagonal(input, last_dim_size, other_dims, + ctx->input_xla_type(0)); ctx->SetOutput(0, diag); } }; @@ -265,7 +255,7 @@ class MatrixDiagPartOp : public XlaOpKernel { // Collapses the last two dimensions. std::vector flattened_dims(dims.begin(), dims.end() - 1); flattened_dims.back() *= dims.back(); - diag = builder->Reshape(diag, flattened_dims); + diag = xla::Reshape(diag, flattened_dims); // Slices or pads the last dimension to 'target_size'. int64 actual_size = flattened_dims.back(); @@ -276,13 +266,13 @@ class MatrixDiagPartOp : public XlaOpKernel { auto* dim = config.mutable_dimensions(flattened_dims.size() - 1); dim->set_edge_padding_high(target_size - actual_size); auto zero = XlaHelpers::Zero(builder, input_type(0)); - diag = builder->Pad(diag, zero, config); + diag = xla::Pad(diag, zero, config); } else if (actual_size > target_size) { std::vector start(flattened_dims.size(), 0); std::vector limits(flattened_dims.begin(), flattened_dims.end()); std::vector strides(flattened_dims.size(), 1); limits[flattened_dims.size() - 1] = target_size; - diag = builder->Slice(diag, start, limits, strides); + diag = xla::Slice(diag, start, limits, strides); } // Reshape so the target values are in the first position of the last @@ -290,18 +280,18 @@ class MatrixDiagPartOp : public XlaOpKernel { std::vector unflattened_dims(dims.begin(), dims.end()); dims[last_dim - 1] = smaller_dim_size; dims[last_dim] = last_dim_size + 1; - diag = builder->Reshape(diag, dims); + diag = xla::Reshape(diag, dims); // Slices out the first column and reshapes to the final shape. std::vector start(dims.size(), 0); std::vector limits(dims.begin(), dims.end()); std::vector strides(dims.size(), 1); limits[last_dim] = 1; - diag = builder->Slice(diag, start, limits, strides); + diag = xla::Slice(diag, start, limits, strides); // Collapses away the last dimension. dims.pop_back(); - diag = builder->Reshape(diag, dims); + diag = xla::Reshape(diag, dims); ctx->SetOutput(0, diag); } diff --git a/tensorflow/compiler/tf2xla/kernels/dynamic_slice_ops.cc b/tensorflow/compiler/tf2xla/kernels/dynamic_slice_ops.cc index 0419de78b2ee83fd395e8bf23444fde84f30bba2..3b86ea34c9e7d943eb9c7de222e0a2be049ebc68 100644 --- a/tensorflow/compiler/tf2xla/kernels/dynamic_slice_ops.cc +++ b/tensorflow/compiler/tf2xla/kernels/dynamic_slice_ops.cc @@ -57,8 +57,8 @@ class DynamicUpdateSliceOp : public XlaOpKernel { input_shape.DebugString(), "; update shape is ", update_shape.DebugString())); - xla::XlaOp result = ctx->builder()->DynamicUpdateSlice( - ctx->Input(0), ctx->Input(1), ctx->Input(2)); + xla::XlaOp result = + xla::DynamicUpdateSlice(ctx->Input(0), ctx->Input(1), ctx->Input(2)); ctx->SetOutput(0, result); } }; diff --git a/tensorflow/compiler/tf2xla/kernels/dynamic_stitch_op.cc b/tensorflow/compiler/tf2xla/kernels/dynamic_stitch_op.cc index dd4a16908779508380b36f43ce2306ff2f5fb8c4..958231505b50431b9bb267b0a3cc5ed56e3aeb21 100644 --- a/tensorflow/compiler/tf2xla/kernels/dynamic_stitch_op.cc +++ b/tensorflow/compiler/tf2xla/kernels/dynamic_stitch_op.cc @@ -20,6 +20,7 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/xla_helpers.h" #include "tensorflow/compiler/tf2xla/xla_op_kernel.h" #include "tensorflow/compiler/tf2xla/xla_op_registry.h" +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/compiler/xla/literal_util.h" #include "tensorflow/core/framework/op_kernel.h" #include "tensorflow/core/framework/register_types.h" @@ -150,8 +151,7 @@ class DynamicStitchOp : public XlaOpKernel { if (new_shape == data_shapes[input_num]) { input[input_num] = handle; } else { - input[input_num] = - ctx->builder()->Reshape(handle, new_shape.dim_sizes()); + input[input_num] = xla::Reshape(handle, new_shape.dim_sizes()); } } @@ -175,10 +175,10 @@ class DynamicStitchOp : public XlaOpKernel { // And place it in the concat list in the place indicated by // the index. to_concat[index_num] = - ctx->builder()->Slice(expression, slice_start, slice_limit, stride); + xla::Slice(expression, slice_start, slice_limit, stride); } - ctx->SetOutput(0, ctx->builder()->ConcatInDim(to_concat, 0)); + ctx->SetOutput(0, xla::ConcatInDim(ctx->builder(), to_concat, 0)); } private: diff --git a/tensorflow/compiler/tf2xla/kernels/elu_op.cc b/tensorflow/compiler/tf2xla/kernels/elu_op.cc index 493781a1e68b8906f1a7e018e5710130e2eb08b5..81f42e504e4b6f813a29769719a7a7fb5d99b9c5 100644 --- a/tensorflow/compiler/tf2xla/kernels/elu_op.cc +++ b/tensorflow/compiler/tf2xla/kernels/elu_op.cc @@ -19,7 +19,7 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/xla_helpers.h" #include "tensorflow/compiler/tf2xla/xla_op_registry.h" #include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/core/framework/kernel_def_builder.h" #include "tensorflow/core/framework/types.h" #include "tensorflow/core/kernels/no_op.h" @@ -34,9 +34,9 @@ class EluOp : public XlaOpKernel { void Compile(XlaOpKernelContext* ctx) override { xla::XlaBuilder* b = ctx->builder(); const auto zero = XlaHelpers::Zero(b, input_type(0)); - const auto pred = b->Gt(ctx->Input(0), zero); - const auto expm1 = b->Expm1(ctx->Input(0)); - ctx->SetOutput(0, b->Select(pred, ctx->Input(0), expm1)); + const auto pred = xla::Gt(ctx->Input(0), zero); + const auto expm1 = xla::Expm1(ctx->Input(0)); + ctx->SetOutput(0, xla::Select(pred, ctx->Input(0), expm1)); } }; @@ -51,9 +51,9 @@ class EluGradOp : public XlaOpKernel { const auto one = XlaHelpers::One(b, input_type(0)); const auto grad = ctx->Input(0); const auto activation = ctx->Input(1); - const auto exp_grad = b->Mul(grad, b->Add(activation, one)); - const auto pred = b->Gt(activation, zero); - ctx->SetOutput(0, b->Select(pred, grad, exp_grad)); + const auto exp_grad = xla::Mul(grad, xla::Add(activation, one)); + const auto pred = xla::Gt(activation, zero); + ctx->SetOutput(0, xla::Select(pred, grad, exp_grad)); } }; @@ -71,10 +71,10 @@ class SeluOp : public XlaOpKernel { 1.0507009873554804934193349852946); const auto scale_alpha = XlaHelpers::FloatLiteral(b, input_type(0), 1.7580993408473768599402175208123); - const auto pred = b->Gt(ctx->Input(0), zero); - const auto expm1 = b->Expm1(ctx->Input(0)); - ctx->SetOutput(0, b->Select(pred, b->Mul(scale, ctx->Input(0)), - b->Mul(scale_alpha, expm1))); + const auto pred = xla::Gt(ctx->Input(0), zero); + const auto expm1 = xla::Expm1(ctx->Input(0)); + ctx->SetOutput(0, xla::Select(pred, xla::Mul(scale, ctx->Input(0)), + xla::Mul(scale_alpha, expm1))); } }; @@ -92,10 +92,10 @@ class SeluGradOp : public XlaOpKernel { 1.7580993408473768599402175208123); const auto grad = ctx->Input(0); const auto activation = ctx->Input(1); - const auto lin_grad = b->Mul(grad, scale); - const auto exp_grad = b->Mul(grad, b->Add(activation, scale_alpha)); - const auto pred = b->Gt(activation, zero); - ctx->SetOutput(0, b->Select(pred, lin_grad, exp_grad)); + const auto lin_grad = xla::Mul(grad, scale); + const auto exp_grad = xla::Mul(grad, xla::Add(activation, scale_alpha)); + const auto pred = xla::Gt(activation, zero); + ctx->SetOutput(0, xla::Select(pred, lin_grad, exp_grad)); } }; diff --git a/tensorflow/compiler/tf2xla/kernels/extract_image_patches_op.cc b/tensorflow/compiler/tf2xla/kernels/extract_image_patches_op.cc index 6df01cabbf1d98c0299bfd808bcc6db6223c4777..65d42a302fca48c7b5f88813f80e975823f63ddf 100644 --- a/tensorflow/compiler/tf2xla/kernels/extract_image_patches_op.cc +++ b/tensorflow/compiler/tf2xla/kernels/extract_image_patches_op.cc @@ -17,6 +17,8 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/xla_helpers.h" #include "tensorflow/compiler/tf2xla/xla_op_kernel.h" #include "tensorflow/compiler/tf2xla/xla_op_registry.h" +#include "tensorflow/compiler/xla/client/lib/numeric.h" +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/core/util/tensor_format.h" namespace tensorflow { @@ -110,13 +112,11 @@ class ExtractImagePatchesOp : public XlaOpKernel { // Builds an identity matrix as a broadcast equality of iotas. // iota = np.arange(np.prod(ksize), depth) // filter = np.equal(np.reshape(iota, [-1, 1]), iota).astype(np.float32) - xla::XlaOp iota; - TF_CHECK_OK(XlaHelpers::Iota(builder, DataType::DT_INT32, - kernel_size * depth, &iota)); + xla::XlaOp iota = xla::Iota(builder, xla::S32, kernel_size * depth); - auto lhs = builder->Reshape(iota, lhs_shape); - auto filter = builder->ConvertElementType( - builder->Eq(lhs, iota, {num_spatial_dims + 1}), type); + auto lhs = xla::Reshape(iota, lhs_shape); + auto filter = xla::ConvertElementType( + xla::Eq(lhs, iota, {num_spatial_dims + 1}), type); xla::ConvolutionDimensionNumbers dims; std::vector window_strides(num_spatial_dims); @@ -148,8 +148,8 @@ class ExtractImagePatchesOp : public XlaOpKernel { } xla::XlaOp conv = - builder->ConvGeneralDilated(ctx->Input(0), filter, window_strides, - padding, lhs_dilation, rhs_dilation, dims); + xla::ConvGeneralDilated(ctx->Input(0), filter, window_strides, padding, + lhs_dilation, rhs_dilation, dims); ctx->SetOutput(0, conv); } diff --git a/tensorflow/compiler/tf2xla/kernels/fake_quantize_ops.cc b/tensorflow/compiler/tf2xla/kernels/fake_quantize_ops.cc index 8f0de0a524c908b598c1a2165a462275346ad137..2fd1a34741e1c7235397f9a69dd8444b4679fa22 100644 --- a/tensorflow/compiler/tf2xla/kernels/fake_quantize_ops.cc +++ b/tensorflow/compiler/tf2xla/kernels/fake_quantize_ops.cc @@ -17,6 +17,7 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/xla_op_kernel.h" #include "tensorflow/compiler/tf2xla/xla_op_registry.h" #include "tensorflow/compiler/xla/client/lib/arithmetic.h" +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/core/platform/macros.h" namespace tensorflow { @@ -49,20 +50,20 @@ void XlaNudge(xla::XlaBuilder* b, const DataType data_type, const float quant_min_value, const float quant_max_value, xla::XlaOp* nudged_min, xla::XlaOp* nudged_max, xla::XlaOp* scale) { - *scale = b->Div(b->Sub(max, min), - XlaHelpers::FloatLiteral(b, data_type, - quant_max_value - quant_min_value)); + *scale = xla::Div(xla::Sub(max, min), + XlaHelpers::FloatLiteral( + b, data_type, quant_max_value - quant_min_value)); xla::XlaOp quant_min = XlaHelpers::FloatLiteral(b, data_type, quant_min_value); - xla::XlaOp zero_point_from_min = b->Sub(quant_min, b->Div(min, *scale)); + xla::XlaOp zero_point_from_min = xla::Sub(quant_min, xla::Div(min, *scale)); xla::XlaOp quant_max = XlaHelpers::FloatLiteral(b, data_type, quant_max_value); xla::XlaOp nudged_zero_point = - b->Select(b->Le(zero_point_from_min, quant_min), quant_min, - b->Select(b->Ge(zero_point_from_min, quant_max), quant_max, - b->Round(zero_point_from_min))); - *nudged_min = b->Mul(b->Sub(quant_min, nudged_zero_point), *scale); - *nudged_max = b->Mul(b->Sub(quant_max, nudged_zero_point), *scale); + xla::Select(xla::Le(zero_point_from_min, quant_min), quant_min, + xla::Select(xla::Ge(zero_point_from_min, quant_max), + quant_max, xla::Round(zero_point_from_min))); + *nudged_min = xla::Mul(xla::Sub(quant_min, nudged_zero_point), *scale); + *nudged_max = xla::Mul(xla::Sub(quant_max, nudged_zero_point), *scale); } xla::XlaOp Quantize(xla::XlaBuilder* b, const xla::XlaOp& input, @@ -71,14 +72,14 @@ xla::XlaOp Quantize(xla::XlaBuilder* b, const xla::XlaOp& input, const xla::XlaOp& nudged_input_max, const xla::XlaOp& input_scale) { xla::XlaOp one = XlaHelpers::FloatLiteral(b, data_type, 1.0f); - xla::XlaOp inv_scale = b->Div(one, input_scale); + xla::XlaOp inv_scale = xla::Div(one, input_scale); xla::XlaOp half = XlaHelpers::FloatLiteral(b, data_type, 0.5f); - xla::XlaOp clamped = b->Clamp(nudged_input_min, input, nudged_input_max); - xla::XlaOp clamped_shifted = b->Sub(clamped, nudged_input_min); + xla::XlaOp clamped = xla::Clamp(nudged_input_min, input, nudged_input_max); + xla::XlaOp clamped_shifted = xla::Sub(clamped, nudged_input_min); xla::XlaOp rounded = - b->Floor(b->Add(b->Mul(clamped_shifted, inv_scale), half)); - return b->Add(b->Mul(rounded, input_scale), nudged_input_min); + xla::Floor(xla::Add(xla::Mul(clamped_shifted, inv_scale), half)); + return xla::Add(xla::Mul(rounded, input_scale), nudged_input_min); } class FakeQuantWithMinMaxArgsOp : public XlaOpKernel { @@ -163,11 +164,11 @@ class FakeQuantWithMinMaxArgsGradOp : public XlaOpKernel { xla::XlaOp nudged_input_max = XlaHelpers::FloatLiteral(b, data_type, nudged_input_max_); - xla::XlaOp between_nudged_min_max = - b->And(b->Le(nudged_input_min, input), b->Le(input, nudged_input_max)); - xla::XlaOp zeroes = b->Broadcast(XlaHelpers::Zero(b, data_type), - gradient_shape.dim_sizes()); - xla::XlaOp output = b->Select(between_nudged_min_max, gradient, zeroes); + xla::XlaOp between_nudged_min_max = xla::And( + xla::Le(nudged_input_min, input), xla::Le(input, nudged_input_max)); + xla::XlaOp zeroes = xla::Broadcast(XlaHelpers::Zero(b, data_type), + gradient_shape.dim_sizes()); + xla::XlaOp output = xla::Select(between_nudged_min_max, gradient, zeroes); ctx->SetOutput(0, output); } @@ -249,25 +250,25 @@ class FakeQuantWithMinMaxVarsGradOp : public XlaOpKernel { XlaNudge(b, data_type, input_min, input_max, quant_min_, quant_max_, &nudged_input_min, &nudged_input_max, &input_scale); - xla::XlaOp between_nudged_min_max = - b->And(b->Le(nudged_input_min, input), b->Le(input, nudged_input_max)); + xla::XlaOp between_nudged_min_max = xla::And( + xla::Le(nudged_input_min, input), xla::Le(input, nudged_input_max)); xla::XlaOp zero = XlaHelpers::Zero(b, data_type); - xla::XlaOp zeroes = b->Broadcast(zero, gradient_shape.dim_sizes()); - xla::XlaOp output0 = b->Select(between_nudged_min_max, gradient, zeroes); + xla::XlaOp zeroes = xla::Broadcast(zero, gradient_shape.dim_sizes()); + xla::XlaOp output0 = xla::Select(between_nudged_min_max, gradient, zeroes); ctx->SetOutput(0, output0); - xla::XlaOp below_min = b->Lt(input, nudged_input_min); - xla::XlaOp select1 = b->Select(below_min, gradient, zeroes); - xla::XlaOp reduce1 = b->ReduceAll( + xla::XlaOp below_min = xla::Lt(input, nudged_input_min); + xla::XlaOp select1 = xla::Select(below_min, gradient, zeroes); + xla::XlaOp reduce1 = xla::ReduceAll( XlaHelpers::ConvertElementType(b, select1, accumulation_type), XlaHelpers::Zero(b, accumulation_type), *ctx->GetOrCreateAdd(accumulation_type)); xla::XlaOp output1 = XlaHelpers::ConvertElementType(b, reduce1, data_type); ctx->SetOutput(1, output1); - xla::XlaOp above_max = b->Gt(input, nudged_input_max); - xla::XlaOp select2 = b->Select(above_max, gradient, zeroes); - xla::XlaOp reduce2 = b->ReduceAll( + xla::XlaOp above_max = xla::Gt(input, nudged_input_max); + xla::XlaOp select2 = xla::Select(above_max, gradient, zeroes); + xla::XlaOp reduce2 = xla::ReduceAll( XlaHelpers::ConvertElementType(b, select2, accumulation_type), XlaHelpers::Zero(b, accumulation_type), *ctx->GetOrCreateAdd(accumulation_type)); diff --git a/tensorflow/compiler/tf2xla/kernels/fft_ops.cc b/tensorflow/compiler/tf2xla/kernels/fft_ops.cc index 933924cad1c7cac2879bd4720cb21ffc33c23f50..b2b00e51e3b00fa93c258af489cf0f4a3e6e764b 100644 --- a/tensorflow/compiler/tf2xla/kernels/fft_ops.cc +++ b/tensorflow/compiler/tf2xla/kernels/fft_ops.cc @@ -18,6 +18,7 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/xla_helpers.h" #include "tensorflow/compiler/tf2xla/xla_op_kernel.h" #include "tensorflow/compiler/tf2xla/xla_op_registry.h" +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/compiler/xla/literal_util.h" #include "tensorflow/core/framework/numeric_op.h" #include "tensorflow/core/framework/op_kernel.h" @@ -62,8 +63,7 @@ class GenericFftOp : public XlaOpKernel { } } - xla::XlaBuilder* b = ctx->builder(); - xla::XlaOp fft = b->Fft(ctx->Input(0), fft_type_, fft_length); + xla::XlaOp fft = xla::Fft(ctx->Input(0), fft_type_, fft_length); ctx->SetOutput(0, fft); } diff --git a/tensorflow/compiler/tf2xla/kernels/fill_op.cc b/tensorflow/compiler/tf2xla/kernels/fill_op.cc index e4467a0fb138ed7919af62ed032c0f5abee3e4f6..95faa1d058f4c0d3fa802b157c6daba1e1adaf41 100644 --- a/tensorflow/compiler/tf2xla/kernels/fill_op.cc +++ b/tensorflow/compiler/tf2xla/kernels/fill_op.cc @@ -19,6 +19,7 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/xla_helpers.h" #include "tensorflow/compiler/tf2xla/xla_op_kernel.h" #include "tensorflow/compiler/tf2xla/xla_op_registry.h" +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/core/framework/kernel_def_builder.h" #include "tensorflow/core/framework/register_types.h" @@ -59,11 +60,11 @@ class FillOp : public XlaOpKernel { xla::XlaOp data = ctx->Input(1); if (value_shape.dims() > 0) { CHECK_EQ(value_shape.dims(), 1); - data = ctx->builder()->Reshape(data, {}); + data = xla::Reshape(data, {}); } // Emit the actual computation, which broadcasts the scalar to the // desired shape. - auto result = ctx->builder()->Broadcast(data, broadcast); + auto result = xla::Broadcast(data, broadcast); ctx->SetOutput(0, result); } diff --git a/tensorflow/compiler/tf2xla/kernels/gather_op.cc b/tensorflow/compiler/tf2xla/kernels/gather_op.cc index d13e25bcddae16d0cd630403219657121b80868d..5f041be5df226ed996b21844c0cf92b6dfac005c 100644 --- a/tensorflow/compiler/tf2xla/kernels/gather_op.cc +++ b/tensorflow/compiler/tf2xla/kernels/gather_op.cc @@ -21,6 +21,7 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/xla_helpers.h" #include "tensorflow/compiler/tf2xla/xla_op_kernel.h" #include "tensorflow/compiler/tf2xla/xla_op_registry.h" +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/core/framework/kernel_def_builder.h" #include "tensorflow/core/framework/op_kernel.h" @@ -75,8 +76,8 @@ Status XlaGather(const xla::XlaOp& input, const TensorShape& input_shape, out_shape.AppendShape(indices_shape_no_index_vectors); out_shape.AppendShape(input_shape_post_axis); - *gather_output = builder->Broadcast(XlaHelpers::Zero(builder, dtype), - out_shape.dim_sizes()); + *gather_output = + xla::Broadcast(XlaHelpers::Zero(builder, dtype), out_shape.dim_sizes()); return Status::OK(); } @@ -142,7 +143,7 @@ Status XlaGather(const xla::XlaOp& input, const TensorShape& input_shape, dim_numbers.add_gather_dims_to_operand_dims(i); } - *gather_output = builder->Gather(input, indices, dim_numbers, window_bounds); + *gather_output = xla::Gather(input, indices, dim_numbers, window_bounds); return Status::OK(); } diff --git a/tensorflow/compiler/tf2xla/kernels/if_op.cc b/tensorflow/compiler/tf2xla/kernels/if_op.cc index d48c6eea754f75a8879d3938f233a6a591d26d0d..f5fcf3cacdbff8297bc42fcb0cf79c2bc83a4e11 100644 --- a/tensorflow/compiler/tf2xla/kernels/if_op.cc +++ b/tensorflow/compiler/tf2xla/kernels/if_op.cc @@ -19,6 +19,7 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/xla_context.h" #include "tensorflow/compiler/tf2xla/xla_op_kernel.h" #include "tensorflow/compiler/tf2xla/xla_op_registry.h" +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" namespace tensorflow { @@ -199,13 +200,13 @@ void XlaIfOp::Compile(XlaOpKernelContext* ctx) { } } - xla::XlaOp outputs = - b->Conditional(ctx->Input(0), b->Tuple(inputs), *then_result.computation, - b->Tuple(inputs), *else_result.computation); + xla::XlaOp outputs = xla::Conditional( + ctx->Input(0), xla::Tuple(b, inputs), *then_result.computation, + xla::Tuple(b, inputs), *else_result.computation); // Sets non-variable outputs. for (int i = 0; i < output_types_.size(); ++i) { if (ctx->input_type(i) != DT_RESOURCE) { - xla::XlaOp output_handle = b->GetTupleElement(outputs, i); + xla::XlaOp output_handle = xla::GetTupleElement(outputs, i); if (VLOG_IS_ON(2)) { LOG(INFO) << "Setting output " << i; auto shape_or = b->GetShape(output_handle); @@ -233,7 +234,7 @@ void XlaIfOp::Compile(XlaOpKernelContext* ctx) { OP_REQUIRES_OK(ctx, resource->SetFromPack( arguments[update.input_index].tensor_array_gradients, - b->GetTupleElement(outputs, pos), b)); + xla::GetTupleElement(outputs, pos), b)); } VLOG(2) << "If variable: pos: " << update.input_index << " name: " << resource->name() diff --git a/tensorflow/compiler/tf2xla/kernels/image_ops.cc b/tensorflow/compiler/tf2xla/kernels/image_ops.cc index 1568b33679963c1a6630525f60560180d40b8d53..cb4caf7bcb4caaa1bf7e0e79e52bb966a8838db3 100644 --- a/tensorflow/compiler/tf2xla/kernels/image_ops.cc +++ b/tensorflow/compiler/tf2xla/kernels/image_ops.cc @@ -17,6 +17,7 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/xla_helpers.h" #include "tensorflow/compiler/tf2xla/xla_op_kernel.h" #include "tensorflow/compiler/tf2xla/xla_op_registry.h" +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" namespace tensorflow { namespace { @@ -32,23 +33,26 @@ std::array RGBToHSV(XlaOpKernelContext* ctx, xla::XlaBuilder* b, auto red = rgb[0]; auto green = rgb[1]; auto blue = rgb[2]; - auto value = b->Max(b->Max(red, green), blue); - auto minimum = b->Min(b->Min(red, green), blue); - auto range = b->Sub(value, minimum); - - auto zeros = b->Broadcast(zero, shape.dim_sizes()); - auto saturation = b->Select(b->Gt(value, zero), b->Div(range, value), zeros); - - auto norm = b->Div(XlaHelpers::FloatLiteral(b, dtype, 1.0 / 6.0), range); - - auto hue = b->Select(b->Eq(green, value), - b->Add(b->Mul(norm, b->Sub(blue, red)), - XlaHelpers::FloatLiteral(b, dtype, 2.0 / 6.0)), - b->Add(b->Mul(norm, b->Sub(red, green)), - XlaHelpers::FloatLiteral(b, dtype, 4.0 / 6.0))); - hue = b->Select(b->Eq(red, value), b->Mul(norm, b->Sub(green, blue)), hue); - hue = b->Select(b->Gt(range, zero), hue, zeros); - hue = b->Select(b->Lt(hue, zero), b->Add(hue, one), hue); + auto value = xla::Max(xla::Max(red, green), blue); + auto minimum = xla::Min(xla::Min(red, green), blue); + auto range = xla::Sub(value, minimum); + + auto zeros = xla::Broadcast(zero, shape.dim_sizes()); + auto saturation = + xla::Select(xla::Gt(value, zero), xla::Div(range, value), zeros); + + auto norm = xla::Div(XlaHelpers::FloatLiteral(b, dtype, 1.0 / 6.0), range); + + auto hue = + xla::Select(xla::Eq(green, value), + xla::Add(xla::Mul(norm, xla::Sub(blue, red)), + XlaHelpers::FloatLiteral(b, dtype, 2.0 / 6.0)), + xla::Add(xla::Mul(norm, xla::Sub(red, green)), + XlaHelpers::FloatLiteral(b, dtype, 4.0 / 6.0))); + hue = xla::Select(xla::Eq(red, value), xla::Mul(norm, xla::Sub(green, blue)), + hue); + hue = xla::Select(xla::Gt(range, zero), hue, zeros); + hue = xla::Select(xla::Lt(hue, zero), xla::Add(hue, one), hue); return {hue, saturation, value}; } @@ -66,15 +70,15 @@ std::array HSVToRGB(xla::XlaBuilder* b, auto four = XlaHelpers::FloatLiteral(b, dtype, 4.0); auto six = XlaHelpers::FloatLiteral(b, dtype, 6.0); - auto dh = b->Mul(hue, six); - auto dr = b->Clamp(zero, b->Sub(b->Abs(b->Sub(dh, three)), one), one); - auto dg = b->Clamp(zero, b->Sub(two, b->Abs(b->Sub(dh, two))), one); - auto db = b->Clamp(zero, b->Sub(two, b->Abs(b->Sub(dh, four))), one); - auto one_minus_s = b->Sub(one, saturation); + auto dh = xla::Mul(hue, six); + auto dr = xla::Clamp(zero, xla::Sub(xla::Abs(xla::Sub(dh, three)), one), one); + auto dg = xla::Clamp(zero, xla::Sub(two, xla::Abs(xla::Sub(dh, two))), one); + auto db = xla::Clamp(zero, xla::Sub(two, xla::Abs(xla::Sub(dh, four))), one); + auto one_minus_s = xla::Sub(one, saturation); - auto red = b->Mul(b->Add(one_minus_s, b->Mul(saturation, dr)), value); - auto green = b->Mul(b->Add(one_minus_s, b->Mul(saturation, dg)), value); - auto blue = b->Mul(b->Add(one_minus_s, b->Mul(saturation, db)), value); + auto red = xla::Mul(xla::Add(one_minus_s, xla::Mul(saturation, dr)), value); + auto green = xla::Mul(xla::Add(one_minus_s, xla::Mul(saturation, dg)), value); + auto blue = xla::Mul(xla::Add(one_minus_s, xla::Mul(saturation, db)), value); return {red, green, blue}; } @@ -97,21 +101,21 @@ class RGBToHSVOp : public XlaOpKernel { xla::XlaBuilder* b = context->builder(); xla::XlaOp input = context->Input(0); - xla::XlaOp red = - b->SliceInDim(input, /*start_index=*/0, /*limit_index=*/1, /*stride=*/1, - /*dimno=*/channel_dim); - xla::XlaOp green = - b->SliceInDim(input, /*start_index=*/1, /*limit_index=*/2, /*stride=*/1, - /*dimno=*/channel_dim); - xla::XlaOp blue = - b->SliceInDim(input, /*start_index=*/2, /*limit_index=*/3, /*stride=*/1, - /*dimno=*/channel_dim); + xla::XlaOp red = xla::SliceInDim(input, /*start_index=*/0, + /*limit_index=*/1, /*stride=*/1, + /*dimno=*/channel_dim); + xla::XlaOp green = xla::SliceInDim(input, /*start_index=*/1, + /*limit_index=*/2, /*stride=*/1, + /*dimno=*/channel_dim); + xla::XlaOp blue = xla::SliceInDim(input, /*start_index=*/2, + /*limit_index=*/3, /*stride=*/1, + /*dimno=*/channel_dim); TensorShape channel_shape = input_shape; channel_shape.set_dim(channel_dim, 1); auto hsv = RGBToHSV(context, b, {red, green, blue}, context->input_type(0), channel_shape); - context->SetOutput(0, b->ConcatInDim(hsv, channel_dim)); + context->SetOutput(0, xla::ConcatInDim(b, hsv, channel_dim)); } }; REGISTER_XLA_OP(Name("RGBToHSV"), RGBToHSVOp); @@ -134,20 +138,20 @@ class HSVToRGBOp : public XlaOpKernel { xla::XlaBuilder* b = context->builder(); xla::XlaOp input = context->Input(0); - xla::XlaOp hue = - b->SliceInDim(input, /*start_index=*/0, /*limit_index=*/1, /*stride=*/1, - /*dimno=*/channel_dim); - xla::XlaOp saturation = - b->SliceInDim(input, /*start_index=*/1, /*limit_index=*/2, /*stride=*/1, - /*dimno=*/channel_dim); - xla::XlaOp value = - b->SliceInDim(input, /*start_index=*/2, /*limit_index=*/3, /*stride=*/1, - /*dimno=*/channel_dim); + xla::XlaOp hue = xla::SliceInDim(input, /*start_index=*/0, + /*limit_index=*/1, /*stride=*/1, + /*dimno=*/channel_dim); + xla::XlaOp saturation = xla::SliceInDim(input, /*start_index=*/1, + /*limit_index=*/2, /*stride=*/1, + /*dimno=*/channel_dim); + xla::XlaOp value = xla::SliceInDim(input, /*start_index=*/2, + /*limit_index=*/3, /*stride=*/1, + /*dimno=*/channel_dim); auto rgb = HSVToRGB(context->builder(), {hue, saturation, value}, context->input_type(0)); - context->SetOutput(0, b->ConcatInDim(rgb, channel_dim)); + context->SetOutput(0, xla::ConcatInDim(b, rgb, channel_dim)); } }; REGISTER_XLA_OP(Name("HSVToRGB"), HSVToRGBOp); @@ -182,18 +186,20 @@ class AdjustContrastOpV2 : public XlaOpKernel { const DataType accumulation_type = XlaHelpers::SumAccumulationType(type); auto converted = XlaHelpers::ConvertElementType(b, input, accumulation_type); - auto reduce = b->Reduce(converted, XlaHelpers::Zero(b, accumulation_type), - *context->GetOrCreateAdd(accumulation_type), - {height_dim, width_dim}); + auto reduce = xla::Reduce(converted, XlaHelpers::Zero(b, accumulation_type), + *context->GetOrCreateAdd(accumulation_type), + {height_dim, width_dim}); auto output = XlaHelpers::ConvertElementType(b, reduce, type); - output = b->Div(output, XlaHelpers::FloatLiteral(b, type, height * width)); + output = + xla::Div(output, XlaHelpers::FloatLiteral(b, type, height * width)); std::vector broadcast_dims(input_shape.dims() - 2); std::iota(broadcast_dims.begin(), broadcast_dims.end(), 0); broadcast_dims.back() = channel_dim; - output = b->Add(b->Mul(input, factor), - b->Mul(output, b->Sub(XlaHelpers::One(b, type), factor)), - broadcast_dims); + output = + xla::Add(xla::Mul(input, factor), + xla::Mul(output, xla::Sub(XlaHelpers::One(b, type), factor)), + broadcast_dims); context->SetOutput(0, output); } }; @@ -226,26 +232,26 @@ class AdjustSaturationOp : public XlaOpKernel { DataType type = context->input_type(0); - xla::XlaOp red = - b->SliceInDim(input, /*start_index=*/0, /*limit_index=*/1, /*stride=*/1, - /*dimno=*/channel_dim); - xla::XlaOp green = - b->SliceInDim(input, /*start_index=*/1, /*limit_index=*/2, /*stride=*/1, - /*dimno=*/channel_dim); - xla::XlaOp blue = - b->SliceInDim(input, /*start_index=*/2, /*limit_index=*/3, /*stride=*/1, - /*dimno=*/channel_dim); + xla::XlaOp red = xla::SliceInDim(input, /*start_index=*/0, + /*limit_index=*/1, /*stride=*/1, + /*dimno=*/channel_dim); + xla::XlaOp green = xla::SliceInDim(input, /*start_index=*/1, + /*limit_index=*/2, /*stride=*/1, + /*dimno=*/channel_dim); + xla::XlaOp blue = xla::SliceInDim(input, /*start_index=*/2, + /*limit_index=*/3, /*stride=*/1, + /*dimno=*/channel_dim); TensorShape channel_shape = input_shape; channel_shape.set_dim(channel_dim, 1); auto hsv = RGBToHSV(context, b, {red, green, blue}, context->input_type(0), channel_shape); - hsv[1] = b->Clamp(XlaHelpers::Zero(b, type), b->Mul(hsv[1], scale), - XlaHelpers::One(b, type)); + hsv[1] = xla::Clamp(XlaHelpers::Zero(b, type), xla::Mul(hsv[1], scale), + XlaHelpers::One(b, type)); auto rgb = HSVToRGB(context->builder(), hsv, context->input_type(0)); - context->SetOutput(0, b->ConcatInDim(rgb, channel_dim)); + context->SetOutput(0, xla::ConcatInDim(b, rgb, channel_dim)); } }; REGISTER_XLA_OP(Name("AdjustSaturation"), AdjustSaturationOp); @@ -276,15 +282,15 @@ class AdjustHueOp : public XlaOpKernel { DataType type = context->input_type(0); - xla::XlaOp red = - b->SliceInDim(input, /*start_index=*/0, /*limit_index=*/1, /*stride=*/1, - /*dimno=*/channel_dim); - xla::XlaOp green = - b->SliceInDim(input, /*start_index=*/1, /*limit_index=*/2, /*stride=*/1, - /*dimno=*/channel_dim); - xla::XlaOp blue = - b->SliceInDim(input, /*start_index=*/2, /*limit_index=*/3, /*stride=*/1, - /*dimno=*/channel_dim); + xla::XlaOp red = xla::SliceInDim(input, /*start_index=*/0, + /*limit_index=*/1, /*stride=*/1, + /*dimno=*/channel_dim); + xla::XlaOp green = xla::SliceInDim(input, /*start_index=*/1, + /*limit_index=*/2, /*stride=*/1, + /*dimno=*/channel_dim); + xla::XlaOp blue = xla::SliceInDim(input, /*start_index=*/2, + /*limit_index=*/3, /*stride=*/1, + /*dimno=*/channel_dim); TensorShape channel_shape = input_shape; channel_shape.set_dim(channel_dim, 1); auto hsv = RGBToHSV(context, b, {red, green, blue}, context->input_type(0), @@ -294,12 +300,13 @@ class AdjustHueOp : public XlaOpKernel { auto one = XlaHelpers::One(b, type); auto& hue = hsv[0]; - hue = b->Rem(b->Add(hsv[0], delta), one); - hue = b->Select(b->Lt(hue, zero), b->Rem(b->Add(one, hue), one), hue); + hue = xla::Rem(xla::Add(hsv[0], delta), one); + hue = + xla::Select(xla::Lt(hue, zero), xla::Rem(xla::Add(one, hue), one), hue); auto rgb = HSVToRGB(context->builder(), hsv, context->input_type(0)); - context->SetOutput(0, b->ConcatInDim(rgb, channel_dim)); + context->SetOutput(0, xla::ConcatInDim(b, rgb, channel_dim)); } }; REGISTER_XLA_OP(Name("AdjustHue"), AdjustHueOp); diff --git a/tensorflow/compiler/tf2xla/kernels/image_resize_ops.cc b/tensorflow/compiler/tf2xla/kernels/image_resize_ops.cc index 79d3a6979cec4c6bda92a71dcff4ddd2151367d5..d6bf92fb3df8d38909df99e11c85ede4fac2bf81 100644 --- a/tensorflow/compiler/tf2xla/kernels/image_resize_ops.cc +++ b/tensorflow/compiler/tf2xla/kernels/image_resize_ops.cc @@ -18,6 +18,8 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/xla_op_kernel.h" #include "tensorflow/compiler/tf2xla/xla_op_registry.h" #include "tensorflow/compiler/xla/array4d.h" +#include "tensorflow/compiler/xla/client/lib/numeric.h" +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/core/framework/kernel_def_builder.h" #include "tensorflow/core/framework/register_types.h" #include "tensorflow/core/lib/math/math_util.h" @@ -127,48 +129,41 @@ const int64 kMax2DKernelSize = 16; xla::XlaOp MakeBilinearResizeKernel(xla::XlaBuilder* builder, gtl::ArraySlice kernel_size, int64 channels) { - xla::XlaOp channels_iota; - // DT_INT32 Iota will always return status::OK(). - TF_CHECK_OK( - XlaHelpers::Iota(builder, DataType::DT_INT32, channels, &channels_iota)); - - auto diag = builder->ConvertElementType( - builder->Eq( - builder->Broadcast(channels_iota, {2 * kernel_size[0] - 1, + xla::XlaOp channels_iota = xla::Iota(builder, xla::S32, channels); + + auto diag = xla::ConvertElementType( + xla::Eq(xla::Broadcast(channels_iota, {2 * kernel_size[0] - 1, 2 * kernel_size[1] - 1, channels}), - channels_iota, /*broadcast_dimensions=*/{2}), + channels_iota, /*broadcast_dimensions=*/{2}), xla::PrimitiveType::F32); - return builder->Mul( - builder->Mul(diag, - builder->ConstantR1(Make1DKernel(kernel_size[1])), - /*broadcast_dimensions=*/{1}), - builder->ConstantR1(Make1DKernel(kernel_size[0])), + return xla::Mul( + xla::Mul(diag, + xla::ConstantR1(builder, Make1DKernel(kernel_size[1])), + /*broadcast_dimensions=*/{1}), + xla::ConstantR1(builder, Make1DKernel(kernel_size[0])), /*broadcast_dimensions=*/{0}); } xla::XlaOp MakeBilinearResizeKernelInDim(xla::XlaBuilder* builder, gtl::ArraySlice kernel_size, int64 channels, int64 dim) { - xla::XlaOp channels_iota; - // DT_INT32 Iota will always return status::OK(). - TF_CHECK_OK( - XlaHelpers::Iota(builder, DataType::DT_INT32, channels, &channels_iota)); - - auto diag = builder->ConvertElementType( - builder->Eq(builder->Broadcast( - channels_iota, - {dim == 0 ? (2 * kernel_size[0] - 1) : 1, - dim == 1 ? (2 * kernel_size[1] - 1) : 1, channels}), - channels_iota, /*broadcast_dimensions=*/{2}), + xla::XlaOp channels_iota = xla::Iota(builder, xla::S32, channels); + + auto diag = xla::ConvertElementType( + xla::Eq( + xla::Broadcast(channels_iota, + {dim == 0 ? (2 * kernel_size[0] - 1) : 1, + dim == 1 ? (2 * kernel_size[1] - 1) : 1, channels}), + channels_iota, /*broadcast_dimensions=*/{2}), xla::PrimitiveType::F32); if (dim == 1) { - return builder->Mul( - diag, builder->ConstantR1(Make1DKernel(kernel_size[1])), + return xla::Mul( + diag, xla::ConstantR1(builder, Make1DKernel(kernel_size[1])), /*broadcast_dimensions=*/{1}); } - return builder->Mul(diag, - builder->ConstantR1(Make1DKernel(kernel_size[0])), - /*broadcast_dimensions=*/{0}); + return xla::Mul(diag, + xla::ConstantR1(builder, Make1DKernel(kernel_size[0])), + /*broadcast_dimensions=*/{0}); } xla::XlaOp ResizeUsingDilationAndConvolution(xla::XlaBuilder* builder, @@ -208,7 +203,7 @@ xla::XlaOp ResizeUsingDilationAndConvolution(xla::XlaBuilder* builder, if (dims.kernel_size[0] * dims.kernel_size[1] < kMax2DKernelSize) { xla::XlaOp kernel = MakeBilinearResizeKernel(builder, dims.kernel_size, channels); - output = builder->ConvGeneralDilated( + output = xla::ConvGeneralDilated( input, kernel, dims.stride, /*padding=*/ {{dims.kernel_size[0] - 1, dims.kernel_size[0] - 1}, @@ -218,7 +213,7 @@ xla::XlaOp ResizeUsingDilationAndConvolution(xla::XlaBuilder* builder, } else { xla::XlaOp kernel0 = MakeBilinearResizeKernelInDim(builder, dims.kernel_size, channels, 0); - output = builder->ConvGeneralDilated( + output = xla::ConvGeneralDilated( input, kernel0, {dims.stride[0], 1}, /*padding=*/ {{dims.kernel_size[0] - 1, dims.kernel_size[0] - 1}, {0, 0}}, @@ -226,7 +221,7 @@ xla::XlaOp ResizeUsingDilationAndConvolution(xla::XlaBuilder* builder, /*rhs_dilation=*/{1, 1}, dimension_numbers); xla::XlaOp kernel1 = MakeBilinearResizeKernelInDim(builder, dims.kernel_size, channels, 1); - output = builder->ConvGeneralDilated( + output = xla::ConvGeneralDilated( output, kernel1, {1, dims.stride[1]}, /*padding=*/ {{0, 0}, {dims.kernel_size[1] - 1, dims.kernel_size[1] - 1}}, @@ -238,8 +233,8 @@ xla::XlaOp ResizeUsingDilationAndConvolution(xla::XlaBuilder* builder, // size > 1 dimension. for (int i = 0; i < num_spatial_dims; ++i) { if (in_size[i] == 1 && out_size[i] > 1) { - output = builder->Add(output, builder->ConstantR1(out_size[i], 0), - /*broadcast_dimensions=*/{1 + i}); + output = xla::Add(output, xla::ConstantR1(builder, out_size[i], 0), + /*broadcast_dimensions=*/{1 + i}); } } return output; @@ -279,12 +274,12 @@ xla::XlaOp ResizeUsingDilationAndConvolutionGradOp(xla::XlaBuilder* builder, for (int i = 0; i < num_spatial_dims; ++i) { if (in_size[i] == 1 && grad_size[i] > 1) { kernel = - builder->Add(kernel, builder->ConstantR1(grad_size[i], 0), - /*broadcast_dimensions=*/{i}); + xla::Add(kernel, xla::ConstantR1(builder, grad_size[i], 0), + /*broadcast_dimensions=*/{i}); } } - output = builder->ConvGeneralDilated( + output = xla::ConvGeneralDilated( grad, kernel, /*window_strides=*/dims.kernel_size, /*padding=*/ {{dims.kernel_size[0] - 1, dims.kernel_size[0] - 1}, @@ -302,23 +297,23 @@ xla::XlaOp ResizeUsingDilationAndConvolutionGradOp(xla::XlaBuilder* builder, // gradient contributions in that dimension. if (in_size[0] == 1 && grad_size[0] > 1) { kernel0 = - builder->Add(kernel0, builder->ConstantR1(grad_size[0], 0), - /*broadcast_dimensions=*/{0}); + xla::Add(kernel0, xla::ConstantR1(builder, grad_size[0], 0), + /*broadcast_dimensions=*/{0}); } if (in_size[1] == 1 && grad_size[1] > 1) { kernel1 = - builder->Add(kernel0, builder->ConstantR1(grad_size[1], 0), - /*broadcast_dimensions=*/{1}); + xla::Add(kernel0, xla::ConstantR1(builder, grad_size[1], 0), + /*broadcast_dimensions=*/{1}); } - output = builder->ConvGeneralDilated( + output = xla::ConvGeneralDilated( grad, kernel0, /*window_strides=*/{dims.kernel_size[0], 1}, /*padding=*/ {{dims.kernel_size[0] - 1, dims.kernel_size[0] - 1}, {0, 0}}, /*lhs_dilation=*/{dims.stride[0], 1}, /*rhs_dilation=*/{1, 1}, dimension_numbers); - output = builder->ConvGeneralDilated( + output = xla::ConvGeneralDilated( output, kernel1, /*window_strides=*/{1, dims.kernel_size[1]}, /*padding=*/ {{0, 0}, {dims.kernel_size[1] - 1, dims.kernel_size[1] - 1}}, @@ -337,7 +332,7 @@ xla::XlaOp ResizeUsingDilationAndConvolutionGradOp(xla::XlaBuilder* builder, } } if (pad_output) { - output = builder->Pad(output, builder->ConstantR0(0.0f), padding); + output = xla::Pad(output, xla::ConstantR0(builder, 0.0f), padding); } return output; } @@ -393,13 +388,13 @@ class ResizeBilinearOp : public XlaOpKernel { } } if (slice_input) { - input = b->Slice(input, {0, 0, 0, 0}, - {batch, slice_size[0], slice_size[1], channels}, - {1, 1, 1, 1}); + input = xla::Slice(input, {0, 0, 0, 0}, + {batch, slice_size[0], slice_size[1], channels}, + {1, 1, 1, 1}); } // Output is always type float. - input = b->ConvertElementType(input, xla::F32); + input = xla::ConvertElementType(input, xla::F32); // Special Case: // Instead of doing a ResizeUsingDilationAndConvolution directly, @@ -529,7 +524,7 @@ class ResizeBilinearGradOp : public XlaOpKernel { } } - output = b->ConvertElementType(output, output_type_); + output = xla::ConvertElementType(output, output_type_); ctx->SetOutput(0, output); } diff --git a/tensorflow/compiler/tf2xla/kernels/index_ops.cc b/tensorflow/compiler/tf2xla/kernels/index_ops.cc index 36eb4c75454ed82804c40b82e5dbaec2eef0a719..f3964748587c1b31cf8b1b76643ff19a9044bf44 100644 --- a/tensorflow/compiler/tf2xla/kernels/index_ops.cc +++ b/tensorflow/compiler/tf2xla/kernels/index_ops.cc @@ -60,19 +60,15 @@ void XlaArgMinMaxOp::Compile(XlaOpKernelContext* ctx) { input_shape.DebugString())); DataType index_type = output_type(0); + xla::PrimitiveType index_xla_type; + OP_REQUIRES_OK(ctx, DataTypeToPrimitiveType(index_type, &index_xla_type)); - xla::XlaBuilder* b = ctx->builder(); xla::XlaOp input = ctx->Input(0); - xla::XlaOp output; if (is_min_) { - OP_REQUIRES_OK(ctx, - XlaHelpers::ArgMin(b, ctx, input, input_shape, input_type(0), - index_type, axis, &output)); + output = XlaHelpers::ArgMin(input, index_xla_type, axis); } else { - OP_REQUIRES_OK(ctx, - XlaHelpers::ArgMax(b, ctx, input, input_shape, input_type(0), - index_type, axis, &output)); + output = XlaHelpers::ArgMax(input, index_xla_type, axis); } ctx->SetOutput(0, output); diff --git a/tensorflow/compiler/tf2xla/kernels/index_ops_cpu.cc b/tensorflow/compiler/tf2xla/kernels/index_ops_cpu.cc index 2c2d88486fda99d2380382a3e2f633f5bdc7478c..22a45b2a11e8ecb688f8e773ef4b286eafe68f4f 100644 --- a/tensorflow/compiler/tf2xla/kernels/index_ops_cpu.cc +++ b/tensorflow/compiler/tf2xla/kernels/index_ops_cpu.cc @@ -19,6 +19,7 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/xla_helpers.h" #include "tensorflow/compiler/tf2xla/xla_op_kernel.h" #include "tensorflow/compiler/tf2xla/xla_op_registry.h" +#include "tensorflow/compiler/xla/literal_util.h" #include "tensorflow/core/framework/kernel_def_builder.h" #include "tensorflow/core/framework/op_kernel.h" #include "tensorflow/core/framework/register_types.h" @@ -76,14 +77,15 @@ class ArgMaxCustomCallOp : public XlaOpKernel { // XLA passes to the function, so it is not included here. std::vector args; args.push_back(ctx->Input(0)); - args.push_back(b.ConstantLiteral( - *xla::Literal::CreateR1(input_shape.dim_sizes()))); + args.push_back(xla::ConstantLiteral( + &b, *xla::LiteralUtil::CreateR1(input_shape.dim_sizes()))); if (input_shape.dims() > 1) { // Don't bother passing the output shape and dim for the 1d case, since // the shape is always a scalar and the dim is always 0. - args.push_back(b.ConstantLiteral( - *xla::Literal::CreateR1(output_shape.dim_sizes()))); - args.push_back(b.ConstantLiteral(*xla::Literal::CreateR0(dim))); + args.push_back(xla::ConstantLiteral( + &b, *xla::LiteralUtil::CreateR1(output_shape.dim_sizes()))); + args.push_back( + xla::ConstantLiteral(&b, *xla::LiteralUtil::CreateR0(dim))); } xla::Shape xla_shape = @@ -94,10 +96,12 @@ class ArgMaxCustomCallOp : public XlaOpKernel { xla::XlaOp output; switch (input_shape.dims()) { case 1: - output = b.CustomCall("argmax_float_1d_xla_impl", args, xla_shape); + output = + xla::CustomCall(&b, "argmax_float_1d_xla_impl", args, xla_shape); break; case 2: - output = b.CustomCall("argmax_float_2d_xla_impl", args, xla_shape); + output = + xla::CustomCall(&b, "argmax_float_2d_xla_impl", args, xla_shape); break; default: OP_REQUIRES(ctx, false, diff --git a/tensorflow/compiler/tf2xla/kernels/l2loss_op.cc b/tensorflow/compiler/tf2xla/kernels/l2loss_op.cc index 1decf7d72d72bb697477e7f841ced2a1a0d5fbe9..9e64711051d31107db1bf6f1966f9ed6f5630c34 100644 --- a/tensorflow/compiler/tf2xla/kernels/l2loss_op.cc +++ b/tensorflow/compiler/tf2xla/kernels/l2loss_op.cc @@ -39,12 +39,12 @@ class L2LossOp : public XlaOpKernel { const DataType accumulation_type = XlaHelpers::SumAccumulationType(dtype); auto t = XlaHelpers::ConvertElementType(b, ctx->Input(0), accumulation_type); - auto square = b->Mul(t, t); - auto reduce = b->Reduce(square, XlaHelpers::Zero(b, accumulation_type), - *ctx->GetOrCreateAdd(accumulation_type), dims); + auto square = xla::Mul(t, t); + auto reduce = xla::Reduce(square, XlaHelpers::Zero(b, accumulation_type), + *ctx->GetOrCreateAdd(accumulation_type), dims); auto deconverted = XlaHelpers::ConvertElementType(b, reduce, dtype); auto two = XlaHelpers::IntegerLiteral(b, dtype, 2); - ctx->SetOutput(0, b->Div(deconverted, two)); + ctx->SetOutput(0, xla::Div(deconverted, two)); } }; diff --git a/tensorflow/compiler/tf2xla/kernels/listdiff_op.cc b/tensorflow/compiler/tf2xla/kernels/listdiff_op.cc index 0388b4c830702ea00ec69fc42c6468326c88cf38..2fb072f827906d40dcf410f0312394c4f568a28d 100644 --- a/tensorflow/compiler/tf2xla/kernels/listdiff_op.cc +++ b/tensorflow/compiler/tf2xla/kernels/listdiff_op.cc @@ -22,6 +22,7 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/xla_helpers.h" #include "tensorflow/compiler/tf2xla/xla_op_kernel.h" #include "tensorflow/compiler/tf2xla/xla_op_registry.h" +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/core/framework/kernel_def_builder.h" #include "tensorflow/core/framework/register_types.h" #include "tensorflow/core/lib/core/errors.h" @@ -90,8 +91,10 @@ class ListDiffOp : public XlaOpKernel { idx_output.push_back(i); } - context->SetOutput(0, context->builder()->ConstantR1(val_output)); - context->SetOutput(1, context->builder()->ConstantR1(idx_output)); + context->SetOutput(0, + xla::ConstantR1(context->builder(), val_output)); + context->SetOutput(1, + xla::ConstantR1(context->builder(), idx_output)); return Status::OK(); } diff --git a/tensorflow/compiler/tf2xla/kernels/lrn_ops.cc b/tensorflow/compiler/tf2xla/kernels/lrn_ops.cc index 39fbf98a6274918840e9e351470f04c2d80c5d01..dc934543cb2f94fbe1e8f1f865156eb082d6a127 100644 --- a/tensorflow/compiler/tf2xla/kernels/lrn_ops.cc +++ b/tensorflow/compiler/tf2xla/kernels/lrn_ops.cc @@ -16,6 +16,7 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/xla_helpers.h" #include "tensorflow/compiler/tf2xla/xla_op_kernel.h" #include "tensorflow/compiler/tf2xla/xla_op_registry.h" +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/core/framework/kernel_def_builder.h" namespace tensorflow { @@ -50,8 +51,8 @@ class LRNOp : public XlaOpKernel { auto accumulation_type = XlaHelpers::SumAccumulationType(input_type(0)); auto converted = XlaHelpers::ConvertElementType(builder, input, accumulation_type); - auto squared = builder->Mul(converted, converted); - auto reduce = builder->ReduceWindow( + auto squared = xla::Mul(converted, converted); + auto reduce = xla::ReduceWindow( squared, XlaHelpers::Zero(builder, accumulation_type), *ctx->GetOrCreateAdd(accumulation_type), /* window_dimensions = */ {1, 1, 1, depth_radius_ * 2 + 1}, @@ -59,12 +60,12 @@ class LRNOp : public XlaOpKernel { auto sqr_sum = XlaHelpers::ConvertElementType(builder, reduce, input_type(0)); - auto scale = builder->Pow( - builder->Add(builder->ConstantR0(bias_), - builder->Mul(builder->ConstantR0(alpha_), sqr_sum)), - builder->ConstantR0(-beta_)); + auto scale = xla::Pow( + xla::Add(xla::ConstantR0(builder, bias_), + xla::Mul(xla::ConstantR0(builder, alpha_), sqr_sum)), + xla::ConstantR0(builder, -beta_)); - ctx->SetOutput(0, builder->Mul(input, scale)); + ctx->SetOutput(0, xla::Mul(input, scale)); } private: @@ -138,8 +139,8 @@ class LRNGradOp : public XlaOpKernel { auto accumulation_type = XlaHelpers::SumAccumulationType(input_type(0)); auto converted = XlaHelpers::ConvertElementType(builder, in_image, accumulation_type); - auto squared = builder->Mul(converted, converted); - auto reduce = builder->ReduceWindow( + auto squared = xla::Mul(converted, converted); + auto reduce = xla::ReduceWindow( squared, XlaHelpers::Zero(builder, accumulation_type), *ctx->GetOrCreateAdd(accumulation_type), /* window_dimensions = */ {1, 1, 1, depth_radius_ * 2 + 1}, @@ -148,17 +149,17 @@ class LRNGradOp : public XlaOpKernel { XlaHelpers::ConvertElementType(builder, reduce, input_type(0)); auto norm = - builder->Add(builder->ConstantR0(bias_), - builder->Mul(builder->ConstantR0(alpha_), sqr_sum)); + xla::Add(xla::ConstantR0(builder, bias_), + xla::Mul(xla::ConstantR0(builder, alpha_), sqr_sum)); - auto dy = builder->Mul( - builder->Mul(builder->ConstantR0(-2.0f * alpha_ * beta_), - builder->Div(out_image, norm)), + auto dy = xla::Mul( + xla::Mul(xla::ConstantR0(builder, -2.0f * alpha_ * beta_), + xla::Div(out_image, norm)), in_grads); auto converted_dy = XlaHelpers::ConvertElementType(builder, dy, accumulation_type); - auto dy_reduce = builder->ReduceWindow( + auto dy_reduce = xla::ReduceWindow( converted_dy, XlaHelpers::Zero(builder, accumulation_type), *ctx->GetOrCreateAdd(accumulation_type), /* window_dimensions = */ {1, 1, 1, depth_radius_ * 2 + 1}, @@ -166,10 +167,10 @@ class LRNGradOp : public XlaOpKernel { auto dy_reduced = XlaHelpers::ConvertElementType(builder, dy_reduce, input_type(0)); - xla::XlaOp gradients = builder->Add( - builder->Mul(in_image, dy_reduced), - builder->Mul(in_grads, - builder->Pow(norm, builder->ConstantR0(-beta_)))); + xla::XlaOp gradients = xla::Add( + xla::Mul(in_image, dy_reduced), + xla::Mul(in_grads, + xla::Pow(norm, xla::ConstantR0(builder, -beta_)))); ctx->SetOutput(0, gradients); } diff --git a/tensorflow/compiler/tf2xla/kernels/matmul_op.cc b/tensorflow/compiler/tf2xla/kernels/matmul_op.cc index 6949b296f4b9afe4a0c9152c763a9ad233b9f595..844080b8cf5462da201ce7671e4f9d02fa52c861 100644 --- a/tensorflow/compiler/tf2xla/kernels/matmul_op.cc +++ b/tensorflow/compiler/tf2xla/kernels/matmul_op.cc @@ -18,6 +18,7 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/xla_helpers.h" #include "tensorflow/compiler/tf2xla/xla_op_kernel.h" #include "tensorflow/compiler/tf2xla/xla_op_registry.h" +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/core/framework/op_kernel.h" namespace tensorflow { @@ -70,15 +71,15 @@ class MatMulOp : public XlaOpKernel { xla::XlaOp b = ctx->Input(1); if (is_sparse_) { if (a_type_ == DT_BFLOAT16) { - a = ctx->builder()->ConvertElementType(a, xla::F32); + a = xla::ConvertElementType(a, xla::F32); } if (b_type_ == DT_BFLOAT16) { - b = ctx->builder()->ConvertElementType(b, xla::F32); + b = xla::ConvertElementType(b, xla::F32); } } - auto lhs = (transpose_a_) ? ctx->builder()->Transpose(a, {1, 0}) : a; - auto rhs = (transpose_b_) ? ctx->builder()->Transpose(b, {1, 0}) : b; - ctx->SetOutput(0, ctx->builder()->Dot(lhs, rhs)); + auto lhs = (transpose_a_) ? xla::Transpose(a, {1, 0}) : a; + auto rhs = (transpose_b_) ? xla::Transpose(b, {1, 0}) : b; + ctx->SetOutput(0, xla::Dot(lhs, rhs)); } private: diff --git a/tensorflow/compiler/tf2xla/kernels/matrix_band_part_op.cc b/tensorflow/compiler/tf2xla/kernels/matrix_band_part_op.cc index fbd5dc0fdad4483aadbe9bc263cc1f7a034cee09..e06c87db7adb1840606208fe15cd68a3ca4d137a 100644 --- a/tensorflow/compiler/tf2xla/kernels/matrix_band_part_op.cc +++ b/tensorflow/compiler/tf2xla/kernels/matrix_band_part_op.cc @@ -16,6 +16,8 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/xla_helpers.h" #include "tensorflow/compiler/tf2xla/xla_op_kernel.h" #include "tensorflow/compiler/tf2xla/xla_op_registry.h" +#include "tensorflow/compiler/xla/client/lib/numeric.h" +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/core/framework/tensor_shape.h" namespace tensorflow { @@ -50,6 +52,7 @@ class MatrixBandPartOp : public XlaOpKernel { xla::XlaOp num_upper = context->Input(2); DataType input_type = context->input_type(0); DataType index_type = context->input_type(1); + xla::PrimitiveType index_xla_type = context->input_xla_type(1); TensorShape batch_shape = input_shape; batch_shape.RemoveLastDims(2); @@ -58,33 +61,29 @@ class MatrixBandPartOp : public XlaOpKernel { // Compute 'offset', which is how many diagonals we are above/below the // diagonal. - xla::XlaOp iota_m; - OP_REQUIRES_OK(context, XlaHelpers::Iota(builder, index_type, m, &iota_m)); + xla::XlaOp iota_m = xla::Iota(builder, index_xla_type, m); + xla::XlaOp iota_n = xla::Iota(builder, index_xla_type, n); - xla::XlaOp iota_n; - OP_REQUIRES_OK(context, XlaHelpers::Iota(builder, index_type, n, &iota_n)); - - auto offset = builder->Sub(builder->Broadcast(iota_n, {m}), iota_m, - /*broadcast_dimensions=*/{0}); + auto offset = xla::Sub(xla::Broadcast(iota_n, {m}), iota_m, + /*broadcast_dimensions=*/{0}); // If num_lower or num_upper are negative, include all lower/upper // diagonals. auto zero_index = XlaHelpers::Zero(builder, index_type); - num_lower = builder->Select( - builder->Lt(num_lower, zero_index), - XlaHelpers::IntegerLiteral(builder, index_type, m), num_lower); - num_upper = builder->Select( - builder->Lt(num_upper, zero_index), - XlaHelpers::IntegerLiteral(builder, index_type, n), num_upper); + num_lower = xla::Select(xla::Lt(num_lower, zero_index), + XlaHelpers::IntegerLiteral(builder, index_type, m), + num_lower); + num_upper = xla::Select(xla::Lt(num_upper, zero_index), + XlaHelpers::IntegerLiteral(builder, index_type, n), + num_upper); - auto indicator = builder->And(builder->Le(builder->Neg(num_lower), offset), - builder->Le(offset, num_upper)); - indicator = builder->Broadcast(indicator, batch_shape.dim_sizes()); + auto indicator = xla::And(xla::Le(xla::Neg(num_lower), offset), + xla::Le(offset, num_upper)); + indicator = xla::Broadcast(indicator, batch_shape.dim_sizes()); auto zero_input = XlaHelpers::Zero(builder, input_type); - auto output = builder->Select( - indicator, input, - builder->Broadcast(zero_input, input_shape.dim_sizes())); + auto output = xla::Select( + indicator, input, xla::Broadcast(zero_input, input_shape.dim_sizes())); context->SetOutput(0, output); } diff --git a/tensorflow/compiler/tf2xla/kernels/matrix_set_diag_op.cc b/tensorflow/compiler/tf2xla/kernels/matrix_set_diag_op.cc index db53f6fef8d6bf901c8281f50791ca6766c46efd..e2ab4b83cfb45b2f9a7f3aba2d2a927d10ad8b85 100644 --- a/tensorflow/compiler/tf2xla/kernels/matrix_set_diag_op.cc +++ b/tensorflow/compiler/tf2xla/kernels/matrix_set_diag_op.cc @@ -16,6 +16,8 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/xla_helpers.h" #include "tensorflow/compiler/tf2xla/xla_op_kernel.h" #include "tensorflow/compiler/tf2xla/xla_op_registry.h" +#include "tensorflow/compiler/xla/client/lib/numeric.h" +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" namespace tensorflow { @@ -61,14 +63,11 @@ class MatrixSetDiagOp : public XlaOpKernel { auto zero = XlaHelpers::Zero(builder, context->input_type(0)); // Create an indicator tensor that is true only on the diagonal. - xla::XlaOp iota_m; - OP_REQUIRES_OK(context, XlaHelpers::Iota(builder, DT_INT32, m, &iota_m)); - xla::XlaOp iota_n; - OP_REQUIRES_OK(context, XlaHelpers::Iota(builder, DT_INT32, n, &iota_n)); - auto indicator = builder->Eq(iota_m, - builder->Broadcast(iota_n, {m}), - /*broadcast_dimensions=*/{0}); - indicator = builder->Broadcast(indicator, batch_shape.dim_sizes()); + xla::XlaOp iota_m = xla::Iota(builder, xla::S32, m); + xla::XlaOp iota_n = xla::Iota(builder, xla::S32, n); + auto indicator = xla::Eq(iota_m, xla::Broadcast(iota_n, {m}), + /*broadcast_dimensions=*/{0}); + indicator = xla::Broadcast(indicator, batch_shape.dim_sizes()); // Broadcast diag up to the input shape. Use an implicit broadcast (Add) // because we need to broadcast on the right. @@ -77,10 +76,10 @@ class MatrixSetDiagOp : public XlaOpKernel { if (min_dim != m) { diag_broadcast_dims.back() = rank - 1; } - diag = builder->Add(diag, builder->Broadcast(zero, input_shape.dim_sizes()), - /*broadcast_dimensions=*/diag_broadcast_dims); + diag = xla::Add(diag, xla::Broadcast(zero, input_shape.dim_sizes()), + /*broadcast_dimensions=*/diag_broadcast_dims); - auto output = builder->Select(indicator, diag, input); + auto output = xla::Select(indicator, diag, input); context->SetOutput(0, output); } diff --git a/tensorflow/compiler/tf2xla/kernels/matrix_triangular_solve_op.cc b/tensorflow/compiler/tf2xla/kernels/matrix_triangular_solve_op.cc index eaed93146460de5a6e8328432302cc75bf36a534..f4def11d08c31513aec5aad15187016a7294c2fd 100644 --- a/tensorflow/compiler/tf2xla/kernels/matrix_triangular_solve_op.cc +++ b/tensorflow/compiler/tf2xla/kernels/matrix_triangular_solve_op.cc @@ -30,13 +30,9 @@ class MatrixTriangularSolveOp : public XlaOpKernel { void Compile(XlaOpKernelContext* ctx) override { auto result = TriangularSolve( - ctx->builder(), ctx->Input(0), ctx->Input(1), /*left_side=*/true, + ctx->Input(0), ctx->Input(1), /*left_side=*/true, /*lower=*/lower_, /*transpose_a=*/adjoint_, /*conjugate_a=*/adjoint_); - if (!result.ok()) { - ctx->SetStatus(result.status()); - return; - } - ctx->SetOutput(0, result.ValueOrDie()); + ctx->SetOutput(0, result); } private: diff --git a/tensorflow/compiler/tf2xla/kernels/mirror_pad_op.cc b/tensorflow/compiler/tf2xla/kernels/mirror_pad_op.cc index c3326b4d11432fb17a02e9a336a70d88bf40da6b..529959dbd90b05f8860360f70e087ef225150600 100644 --- a/tensorflow/compiler/tf2xla/kernels/mirror_pad_op.cc +++ b/tensorflow/compiler/tf2xla/kernels/mirror_pad_op.cc @@ -16,6 +16,7 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/xla_helpers.h" #include "tensorflow/compiler/tf2xla/xla_op_kernel.h" #include "tensorflow/compiler/tf2xla/xla_op_registry.h" +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/core/util/mirror_pad_mode.h" namespace tensorflow { @@ -32,16 +33,16 @@ class MirrorPadOp : public XlaOpKernel { xla::XlaOp accum = t; for (int64 dimno = xla::ShapeUtil::Rank(original_shape) - 1; dimno >= 0; --dimno) { - auto t_rev = b->Rev(accum, {dimno}); + auto t_rev = xla::Rev(accum, {dimno}); TF_ASSIGN_OR_RETURN(int64 lhs_padding, pad_literal.GetIntegralAsS64({dimno, 0})); TF_ASSIGN_OR_RETURN(int64 rhs_padding, pad_literal.GetIntegralAsS64({dimno, 1})); int64 dim_size = original_shape.dimensions(dimno); - auto lhs_pad = b->SliceInDim(t_rev, dim_size - 1 - lhs_padding, - dim_size - 1, 1, dimno); - auto rhs_pad = b->SliceInDim(t_rev, 1, 1 + rhs_padding, 1, dimno); - accum = b->ConcatInDim({lhs_pad, accum, rhs_pad}, dimno); + auto lhs_pad = xla::SliceInDim(t_rev, dim_size - 1 - lhs_padding, + dim_size - 1, 1, dimno); + auto rhs_pad = xla::SliceInDim(t_rev, 1, 1 + rhs_padding, 1, dimno); + accum = xla::ConcatInDim(b, {lhs_pad, accum, rhs_pad}, dimno); } return accum; } diff --git a/tensorflow/compiler/tf2xla/kernels/pack_op.cc b/tensorflow/compiler/tf2xla/kernels/pack_op.cc index aecaabb6dcf46bdd6ae3da929448d6370acb989b..3aed47de2603f3e187ad515d4db3f884da4c6cc8 100644 --- a/tensorflow/compiler/tf2xla/kernels/pack_op.cc +++ b/tensorflow/compiler/tf2xla/kernels/pack_op.cc @@ -22,6 +22,7 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/xla_helpers.h" #include "tensorflow/compiler/tf2xla/xla_op_kernel.h" #include "tensorflow/compiler/tf2xla/xla_op_registry.h" +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/compiler/xla/literal_util.h" #include "tensorflow/core/framework/op_kernel.h" #include "tensorflow/core/framework/register_types.h" @@ -76,11 +77,10 @@ class PackOp : public XlaOpKernel { for (int i = 0; i < num; ++i) { // Reshape the inputs to have an extra dimension of size 1. - reshaped_inputs[i] = - ctx->builder()->Reshape(values[i], child_shape.dim_sizes()); + reshaped_inputs[i] = xla::Reshape(values[i], child_shape.dim_sizes()); } - ctx->SetOutput(0, ctx->builder()->ConcatInDim(reshaped_inputs, axis)); + ctx->SetOutput(0, xla::ConcatInDim(ctx->builder(), reshaped_inputs, axis)); } private: diff --git a/tensorflow/compiler/tf2xla/kernels/pad_op.cc b/tensorflow/compiler/tf2xla/kernels/pad_op.cc index 17b85338f75d6295c6b4a1bf1db24aa641eab020..89fd610bc63349d008836c3c4e6ec8927c232a54 100644 --- a/tensorflow/compiler/tf2xla/kernels/pad_op.cc +++ b/tensorflow/compiler/tf2xla/kernels/pad_op.cc @@ -17,6 +17,7 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/xla_helpers.h" #include "tensorflow/compiler/tf2xla/xla_op_kernel.h" #include "tensorflow/compiler/tf2xla/xla_op_registry.h" +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/core/framework/kernel_def_builder.h" #include "tensorflow/core/framework/register_types.h" @@ -74,11 +75,10 @@ class PadOp : public XlaOpKernel { if (ctx->num_inputs() == 3) { OP_REQUIRES(ctx, TensorShapeUtils::IsScalar(ctx->InputShape(2)), errors::InvalidArgument("constant_values must be a scalar.")); - ctx->SetOutput(0, - ctx->builder()->Pad(ctx->Input(0), ctx->Input(2), config)); + ctx->SetOutput(0, xla::Pad(ctx->Input(0), ctx->Input(2), config)); } else { auto zero = XlaHelpers::Zero(ctx->builder(), input_type(0)); - ctx->SetOutput(0, ctx->builder()->Pad(ctx->Input(0), zero, config)); + ctx->SetOutput(0, xla::Pad(ctx->Input(0), zero, config)); } } }; diff --git a/tensorflow/compiler/tf2xla/kernels/pooling_ops.cc b/tensorflow/compiler/tf2xla/kernels/pooling_ops.cc index eb8b5b130f949bedddc5cf12ddc958d0e199db33..12d9cb9bac6b98c8f4c3edc3f6c661acf4466f98 100644 --- a/tensorflow/compiler/tf2xla/kernels/pooling_ops.cc +++ b/tensorflow/compiler/tf2xla/kernels/pooling_ops.cc @@ -20,7 +20,9 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/xla_op_kernel.h" #include "tensorflow/compiler/tf2xla/xla_op_registry.h" #include "tensorflow/compiler/xla/client/lib/arithmetic.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/client/lib/constants.h" +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/util.h" #include "tensorflow/core/framework/op_kernel.h" #include "tensorflow/core/framework/register_types.h" @@ -61,6 +63,9 @@ class PoolingOp : public XlaOpKernel { Padding padding; OP_REQUIRES_OK(ctx, ctx->GetAttr("padding", &padding)); padding_ = (padding == VALID) ? xla::Padding::kValid : xla::Padding::kSame; + + OP_REQUIRES_OK( + ctx, DataTypeToPrimitiveType(reduction_type_, &xla_reduction_type_)); } int num_dims() const { return num_spatial_dims_ + 2; } @@ -113,8 +118,8 @@ class PoolingOp : public XlaOpKernel { xla::XlaBuilder* const b = ctx->builder(); auto input = XlaHelpers::ConvertElementType(b, ctx->Input(0), reduction_type_); - auto reduce = ctx->builder()->ReduceWindow( - input, InitValue(b), *Reduction(ctx), ksize, stride, padding_); + auto reduce = xla::ReduceWindow(input, InitValue(b), *Reduction(ctx), ksize, + stride, padding_); auto pooled = XlaHelpers::ConvertElementType(b, reduce, input_type(0)); ctx->SetOutput(0, PostProcessOutput(ctx, pooled, input_type(0), input_shape)); @@ -127,6 +132,7 @@ class PoolingOp : public XlaOpKernel { xla::Padding padding_; TensorFormat data_format_ = FORMAT_NHWC; DataType reduction_type_; + xla::PrimitiveType xla_reduction_type_; }; class MaxPoolOp : public PoolingOp { @@ -136,7 +142,7 @@ class MaxPoolOp : public PoolingOp { /*reduction_type=*/ctx->input_type(0)) {} xla::XlaOp InitValue(xla::XlaBuilder* b) override { - return XlaHelpers::MinValue(b, reduction_type_); + return xla::MinValue(b, xla_reduction_type_); } const xla::XlaComputation* Reduction(XlaOpKernelContext* ctx) override { @@ -190,7 +196,7 @@ static xla::XlaOp AvgPoolDivideByCount( auto divisor = XlaHelpers::IntegerLiteral(ctx->builder(), dtype, window_size); - return ctx->builder()->Div(output, divisor); + return xla::Div(output, divisor); } else { // For SAME padding, the padding shouldn't be included in the // counts. We use another ReduceWindow to find the right counts. @@ -212,18 +218,18 @@ static xla::XlaOp AvgPoolDivideByCount( // Build a matrix of all 1s, with the same width/height as the input. const DataType accumulation_type = XlaHelpers::SumAccumulationType(dtype); - auto ones = ctx->builder()->Broadcast( + auto ones = xla::Broadcast( XlaHelpers::One(ctx->builder(), accumulation_type), input_dim_sizes); // Perform a ReduceWindow with the same window size, strides, and padding // to count the number of contributions to each result element. - auto reduce = ctx->builder()->ReduceWindow( + auto reduce = xla::ReduceWindow( ones, XlaHelpers::Zero(ctx->builder(), accumulation_type), *ctx->GetOrCreateAdd(accumulation_type), window_ksize, window_stride, xla::Padding::kSame); auto counts = XlaHelpers::ConvertElementType(ctx->builder(), reduce, dtype); - return ctx->builder()->Div(output, counts, window_dims); + return xla::Div(output, counts, window_dims); } } @@ -235,7 +241,7 @@ class AvgPoolOp : public PoolingOp { XlaHelpers::SumAccumulationType(ctx->input_type(0))) {} xla::XlaOp InitValue(xla::XlaBuilder* b) override { - return XlaHelpers::Zero(b, reduction_type_); + return xla::Zero(b, xla_reduction_type_); } const xla::XlaComputation* Reduction(XlaOpKernelContext* ctx) override { @@ -347,9 +353,9 @@ class MaxPoolGradOp : public XlaOpKernel { xla::XlaOp init_value = XlaHelpers::Zero(ctx->builder(), input_type(2)); auto select = CreateScalarGeComputation(element_type, ctx->builder()); auto scatter = CreateScalarAddComputation(element_type, ctx->builder()); - xla::XlaOp gradients = ctx->builder()->SelectAndScatter( - input, select, ksize_, stride_, xla_padding, out_backprop, init_value, - scatter); + xla::XlaOp gradients = + xla::SelectAndScatter(input, select, ksize_, stride_, xla_padding, + out_backprop, init_value, scatter); ctx->SetOutput(0, gradients); } @@ -485,12 +491,12 @@ class AvgPoolGradOp : public XlaOpKernel { } auto zero = XlaHelpers::Zero(b, dtype); - auto padded_gradients = b->Pad(out_backprop_div, zero, padding_config); + auto padded_gradients = xla::Pad(out_backprop_div, zero, padding_config); // in_backprop = padded_gradients ones std::vector ones(num_dims(), 1LL); auto accumulation_type = XlaHelpers::SumAccumulationType(dtype); - auto in_backprop = b->ReduceWindow( + auto in_backprop = xla::ReduceWindow( XlaHelpers::ConvertElementType(b, padded_gradients, accumulation_type), XlaHelpers::Zero(b, accumulation_type), *ctx->GetOrCreateAdd(accumulation_type), ksize_, @@ -614,58 +620,61 @@ class MaxPoolGradGradOp : public XlaOpKernel { auto b = ctx->builder(); - auto sixteen = b->ConstantR0(16); + auto sixteen = xla::ConstantR0(b, 16); // in (f32) -> round to bf16 -> f32 for correct bitwidth -> 16-high-bit u32 - auto in_hi = b->BitcastConvertType( - b->ConvertElementType(b->ConvertElementType(input, xla::BF16), - xla::F32), + auto in_hi = xla::BitcastConvertType( + xla::ConvertElementType(xla::ConvertElementType(input, xla::BF16), + xla::F32), xla::U32); - auto bp_int = b->BitcastConvertType(out_backprop, xla::U32); - auto bp_hi = b->ShiftRightLogical(bp_int, sixteen); - auto bp_lo = b->ShiftRightLogical(b->ShiftLeft(bp_int, sixteen), sixteen); - auto in_hi_bp_hi = b->Add(in_hi, bp_hi); // Want an unsigned add. - auto in_hi_bp_lo = b->Add(in_hi, bp_lo); // Want an unsigned add. - - auto init_value = XlaHelpers::MinValue(b, DT_FLOAT); + auto bp_int = xla::BitcastConvertType(out_backprop, xla::U32); + auto bp_hi = xla::ShiftRightLogical(bp_int, sixteen); + auto bp_lo = + xla::ShiftRightLogical(xla::ShiftLeft(bp_int, sixteen), sixteen); + auto in_hi_bp_hi = xla::Add(in_hi, bp_hi); // Want an unsigned add. + auto in_hi_bp_lo = xla::Add(in_hi, bp_lo); // Want an unsigned add. + + auto init_value = xla::MinValue(b, xla::F32); // We will reduce by taking the maximal value up to 16 bits (ignoring the lo // 16 bits of packed-in hi/lo backprop value). auto rb = b->CreateSubBuilder("GreaterOrEqOf_ByFirst16Bits"); { // F32 parameters to satisfy lowering type restriction for reduce opcode. const xla::Shape scalar = xla::ShapeUtil::MakeShape(xla::F32, {}); - auto lhs = rb->Parameter(0, scalar, "lhs"); - auto rhs = rb->Parameter(1, scalar, "rhs"); - auto sixteen = rb->ConstantR0(16); - auto lhs_criteria = rb->ShiftLeft( - rb->ShiftRightLogical(rb->BitcastConvertType(lhs, xla::S32), sixteen), - sixteen); - auto rhs_criteria = rb->ShiftLeft( - rb->ShiftRightLogical(rb->BitcastConvertType(rhs, xla::S32), sixteen), - sixteen); + auto lhs = xla::Parameter(rb.get(), 0, scalar, "lhs"); + auto rhs = xla::Parameter(rb.get(), 1, scalar, "rhs"); + auto sixteen = xla::ConstantR0(rb.get(), 16); + auto lhs_criteria = + xla::ShiftLeft(xla::ShiftRightLogical( + xla::BitcastConvertType(lhs, xla::S32), sixteen), + sixteen); + auto rhs_criteria = + xla::ShiftLeft(xla::ShiftRightLogical( + xla::BitcastConvertType(rhs, xla::S32), sixteen), + sixteen); // Must use a F32 comparison, because S32 would not work for negatives. - rb->Select(rb->Ge(rb->BitcastConvertType(lhs_criteria, xla::F32), - rb->BitcastConvertType(rhs_criteria, xla::F32)), - lhs, rhs); + xla::Select(xla::Ge(xla::BitcastConvertType(lhs_criteria, xla::F32), + xla::BitcastConvertType(rhs_criteria, xla::F32)), + lhs, rhs); } auto reduce = rb->BuildAndNoteError(); xla::Padding xla_padding = (padding_ == VALID) ? xla::Padding::kValid : xla::Padding::kSame; auto pooled_hi = - b->ReduceWindow(b->BitcastConvertType(in_hi_bp_hi, xla::F32), - init_value, reduce, ksize_, stride_, xla_padding); + xla::ReduceWindow(xla::BitcastConvertType(in_hi_bp_hi, xla::F32), + init_value, reduce, ksize_, stride_, xla_padding); auto pooled_lo = - b->ReduceWindow(b->BitcastConvertType(in_hi_bp_lo, xla::F32), - init_value, reduce, ksize_, stride_, xla_padding); + xla::ReduceWindow(xla::BitcastConvertType(in_hi_bp_lo, xla::F32), + init_value, reduce, ksize_, stride_, xla_padding); auto grads_hi = - b->ShiftLeft(b->BitcastConvertType(pooled_hi, xla::U32), sixteen); - auto grads_lo = b->ShiftRightLogical( - b->ShiftLeft(b->BitcastConvertType(pooled_lo, xla::U32), sixteen), + xla::ShiftLeft(xla::BitcastConvertType(pooled_hi, xla::U32), sixteen); + auto grads_lo = xla::ShiftRightLogical( + xla::ShiftLeft(xla::BitcastConvertType(pooled_lo, xla::U32), sixteen), sixteen); - auto grads = b->Add(grads_hi, grads_lo); // Want an unsigned add. + auto grads = xla::Add(grads_hi, grads_lo); // Want an unsigned add. xla::PrimitiveType element_type; OP_REQUIRES_OK(ctx, DataTypeToPrimitiveType(input_type(2), &element_type)); - ctx->SetOutput(0, b->BitcastConvertType(grads, element_type)); + ctx->SetOutput(0, xla::BitcastConvertType(grads, element_type)); } protected: diff --git a/tensorflow/compiler/tf2xla/kernels/qr_op.cc b/tensorflow/compiler/tf2xla/kernels/qr_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..de9068a640dc03b141b6954eaa1629dd6c8c1f3a --- /dev/null +++ b/tensorflow/compiler/tf2xla/kernels/qr_op.cc @@ -0,0 +1,47 @@ +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow/compiler/tf2xla/lib/qr.h" +#include "tensorflow/compiler/tf2xla/xla_op_kernel.h" +#include "tensorflow/compiler/tf2xla/xla_op_registry.h" + +namespace tensorflow { +namespace { + +class QROp : public XlaOpKernel { + public: + explicit QROp(OpKernelConstruction* ctx) : XlaOpKernel(ctx) { + bool full_matrices; + OP_REQUIRES_OK(ctx, ctx->GetAttr("full_matrices", &full_matrices)); + OP_REQUIRES( + ctx, full_matrices, + errors::Unimplemented("full_matrices=False case of QR decomposition is " + "not implemented in TF/XLA")); + } + void Compile(XlaOpKernelContext* ctx) override { + auto result = QRDecomposition(ctx->Input(0)); + if (!result.ok()) { + ctx->SetStatus(result.status()); + return; + } + ctx->SetOutput(0, result.ValueOrDie().q); + ctx->SetOutput(1, result.ValueOrDie().r); + } +}; + +REGISTER_XLA_OP(Name("Qr").TypeConstraint("T", kFloatTypes), QROp); + +} // namespace +} // namespace tensorflow diff --git a/tensorflow/compiler/tf2xla/kernels/quantize_and_dequantize_op.cc b/tensorflow/compiler/tf2xla/kernels/quantize_and_dequantize_op.cc index 661cd5923e1023eaf89a6bc4f56fcc362c8bcfb6..e88221e4f400abeec59d85c1539d4f70bf515d3c 100644 --- a/tensorflow/compiler/tf2xla/kernels/quantize_and_dequantize_op.cc +++ b/tensorflow/compiler/tf2xla/kernels/quantize_and_dequantize_op.cc @@ -13,10 +13,13 @@ See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ +#include "tensorflow/compiler/tf2xla/type_util.h" #include "tensorflow/compiler/tf2xla/xla_helpers.h" #include "tensorflow/compiler/tf2xla/xla_op_kernel.h" #include "tensorflow/compiler/tf2xla/xla_op_registry.h" #include "tensorflow/compiler/xla/client/lib/arithmetic.h" +#include "tensorflow/compiler/xla/client/lib/constants.h" +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/core/platform/macros.h" namespace tensorflow { @@ -28,82 +31,115 @@ class QuantizeAndDequantizeOp : public XlaOpKernel { : XlaOpKernel(ctx) { OP_REQUIRES_OK(ctx, ctx->GetAttr("signed_input", &signed_input_)); OP_REQUIRES_OK(ctx, ctx->GetAttr("range_given", &range_given_)); - OP_REQUIRES_OK(ctx, ctx->GetAttr("num_bits", &num_bits_)); - OP_REQUIRES(ctx, num_bits_ > 0 && num_bits_ < (signed_input_ ? 62 : 63), - errors::InvalidArgument("num_bits is out of range: ", num_bits_, - " with signed_input_ ", signed_input_)); } void Compile(XlaOpKernelContext* ctx) override { xla::XlaOp input = ctx->Input(0); const DataType data_type = ctx->input_type(0); - // Comments taken from semantics description at - // https://www.tensorflow.org/versions/r1.0/api_docs/cc/class/tensorflow/ops/quantize-and-dequantize - // - // ... we find m such that - // - // m = max(abs(input_min), abs(input_max)) if range_given is true, - // m = max(abs(min_elem(input)), - // abs(max_elem(input))) otherwise. + xla::PrimitiveType xla_type; + OP_REQUIRES_OK(ctx, DataTypeToPrimitiveType(data_type, &xla_type)); + xla::XlaBuilder* b = ctx->builder(); - xla::XlaOp input_min, input_max; + + // The implementation follows + // tensorflow/core/kernels/quantize_and_dequantize_op.h closely. + xla::XlaOp min_range, max_range; if (range_given_) { - double input_min_value, input_max_value; - OP_REQUIRES_OK(ctx, ctx->ConstantInputAsFloatScalar(1, &input_min_value)); - OP_REQUIRES_OK(ctx, ctx->ConstantInputAsFloatScalar(2, &input_max_value)); - input_min = XlaHelpers::FloatLiteral(b, data_type, input_min_value); - input_max = XlaHelpers::FloatLiteral(b, data_type, input_max_value); + min_range = ctx->Input(1); + max_range = ctx->Input(2); } else { const xla::XlaComputation* fmax = ctx->GetOrCreateMax(data_type); const xla::XlaComputation* fmin = ctx->GetOrCreateMin(data_type); - input_min = - b->ReduceAll(input, XlaHelpers::MaxValue(b, data_type), *fmin); - input_max = - b->ReduceAll(input, XlaHelpers::MinValue(b, data_type), *fmax); + min_range = ReduceAll(input, xla::MaxValue(b, xla_type), *fmin); + max_range = ReduceAll(input, xla::MinValue(b, xla_type), *fmax); } - xla::XlaOp m = b->Max(b->Abs(input_min), b->Abs(input_max)); - - // Next, we choose our fixed-point quantization buckets, [min_fixed, - // max_fixed]. If signed_input is true, this is - // - // [min_fixed, max_fixed ] = [-((1 << (num_bits - 1)) - 1), - // (1 << (num_bits - 1)) - 1]. - // - // Otherwise, if signed_input is false, the fixed-point range is - // - // [min_fixed, max_fixed] = [0, (1 << num_bits) - 1]. - int64 min_fixed, max_fixed; + + xla::XlaOp num_bits; + if (num_bits_ < 0) { + OP_REQUIRES( + ctx, ctx->num_inputs() == 4, + errors::Internal("Expected 4 inputs to QuantizeAndDequantize")); + num_bits = ctx->Input(3); + } else { + num_bits = xla::ConstantR0(b, num_bits_); + } + + const xla::XlaOp zero = XlaHelpers::Zero(b, data_type); + const xla::XlaOp one = XlaHelpers::One(b, data_type); + const xla::XlaOp two = XlaHelpers::FloatLiteral(b, data_type, 2.0); + const xla::XlaOp half = XlaHelpers::FloatLiteral(b, data_type, 0.5); + + // Calculate the range for the simulated integer quantization: + // e.g. [-128,127] for signed = true, num_bits = 8, + // or [0, 255] for signed = false, num_bits = 8. + // We do this in floating point for hardware that does not have 64-bit + // integer support. + xla::XlaOp min_quantized, max_quantized; if (signed_input_) { - min_fixed = -((1LL << (num_bits_ - 1)) - 1); - max_fixed = (1LL << (num_bits_ - 1)) - 1; + min_quantized = + -Pow(two, ConvertElementType(num_bits - xla::ConstantR0(b, 1), + xla_type)); + max_quantized = + Pow(two, ConvertElementType(num_bits - xla::ConstantR0(b, 1), + xla_type)) - + one; } else { - min_fixed = 0; - max_fixed = (1LL << num_bits_) - 1; + min_quantized = zero; + max_quantized = Pow(two, ConvertElementType(num_bits, xla_type)) - one; } - // From this we compute our scaling factor, s: - // - // s = (max_fixed - min_fixed) / (2 * m). - xla::XlaOp s = - b->Div(XlaHelpers::FloatLiteral(b, data_type, max_fixed - min_fixed), - b->Mul(XlaHelpers::FloatLiteral(b, data_type, 2.0), m)); + // Determine the maximum scaling factor that would scale + // [min_range, max_range] to not exceed [min_quantized, max_quantized], + // while keeping 0 unchanged. + xla::XlaOp scale_from_min_side = + Select(Gt(min_quantized * min_range, zero), min_quantized / min_range, + xla::MaxFiniteValue(b, xla_type)); + xla::XlaOp scale_from_max_side = + Select(Gt(max_quantized * max_range, zero), max_quantized / max_range, + xla::MaxFiniteValue(b, xla_type)); - // Now we can quantize and dequantize the elements of our tensor. An element - // e is transformed into e': - // - // e' = (e * s).round_to_nearest() / s. - xla::XlaOp result = b->Div(b->Round(b->Mul(input, s)), s); + // Note: Avoids changing the side of the range that determines scale. + xla::XlaOp cond = Lt(scale_from_min_side, scale_from_max_side); + xla::XlaOp scale = Select(cond, scale_from_min_side, scale_from_max_side); + xla::XlaOp inverse_scale = + Select(cond, min_range / min_quantized, max_range / max_quantized); + min_range = Select(cond, min_range, min_quantized * inverse_scale); + max_range = Select(cond, max_quantized * inverse_scale, max_range); + if (range_given_) { + // Note: The clamping here is to avoid overflow in the quantized type. + // The semantics of the op does not guarantee to clamp to the specified + // min_range and max_range - because we may have changed either min_range + // or max_range. + // No need to clamp to min_range and max_range if range_given_ == false as + // in that case they were measured from the tensor. + input = Clamp(min_range, input, max_range); + } + xla::XlaOp result = + Floor((input - min_range) * scale + half) * inverse_scale + min_range; ctx->SetOutput(0, result); } - int64 num_bits_; + protected: + int64 num_bits_ = -1; bool signed_input_; bool range_given_; }; -REGISTER_XLA_OP(Name("QuantizeAndDequantizeV2"), QuantizeAndDequantizeOp); +class QuantizeAndDequantizeV2Op : public QuantizeAndDequantizeOp { + public: + explicit QuantizeAndDequantizeV2Op(OpKernelConstruction* ctx) + : QuantizeAndDequantizeOp(ctx) { + OP_REQUIRES_OK(ctx, ctx->GetAttr("num_bits", &num_bits_)); + OP_REQUIRES(ctx, num_bits_ > 0 && num_bits_ < (signed_input_ ? 62 : 63), + errors::InvalidArgument("num_bits is out of range: ", num_bits_, + " with signed_input_ ", signed_input_)); + } +}; + +REGISTER_XLA_OP(Name("QuantizeAndDequantizeV2"), QuantizeAndDequantizeV2Op); +REGISTER_XLA_OP(Name("QuantizeAndDequantizeV3"), QuantizeAndDequantizeOp); } // namespace } // namespace tensorflow diff --git a/tensorflow/compiler/tf2xla/kernels/random_ops.cc b/tensorflow/compiler/tf2xla/kernels/random_ops.cc index be83834e864aceeefdb9714b885c0f2216b83713..607cad798a98cfa0c6161a8154001926384e724e 100644 --- a/tensorflow/compiler/tf2xla/kernels/random_ops.cc +++ b/tensorflow/compiler/tf2xla/kernels/random_ops.cc @@ -26,6 +26,8 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/xla_op_kernel.h" #include "tensorflow/compiler/tf2xla/xla_op_registry.h" #include "tensorflow/compiler/xla/client/lib/arithmetic.h" +#include "tensorflow/compiler/xla/client/lib/numeric.h" +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/core/framework/op_kernel.h" #include "tensorflow/core/framework/tensor.h" #include "tensorflow/core/framework/tensor_shape.h" @@ -46,8 +48,8 @@ class RandomUniformOp : public XlaOpKernel { OP_REQUIRES_OK(ctx, TensorShapeToXLAShape(dtype, shape, &xla_shape)); xla::XlaBuilder* b = ctx->builder(); - xla::XlaOp result = b->RngUniform(XlaHelpers::Zero(b, dtype), - XlaHelpers::One(b, dtype), xla_shape); + xla::XlaOp result = xla::RngUniform(XlaHelpers::Zero(b, dtype), + XlaHelpers::One(b, dtype), xla_shape); ctx->SetOutput(0, result); } @@ -72,57 +74,121 @@ class RandomShuffleOp : public XlaOpKernel { for (tensorflow::TensorShapeDim dimension : input_shape) { num_elements *= dimension.size; } + if (num_elements <= 1 || n <= 1) { // No shuffling is required, so copy input directly to output ctx->SetOutput(0, input); - } else { - // Generate the random swaps for the indices. - auto swaps_shape = xla::ShapeUtil::MakeShape(xla::S32, {n}); - auto swaps = - builder->RngUniform(builder->ConstantR0(0), - builder->ConstantR0(n), swaps_shape); - - // Generate range(n) as the initial value for the indices to be swapped. - xla::XlaOp indices; - TF_CHECK_OK(XlaHelpers::Iota(builder, DataType::DT_INT32, n, &indices)); - - // Swap the indices at i and swaps[i]. - auto swap_body_fn = [&](xla::XlaOp i, - gtl::ArraySlice loop_vars, - xla::XlaBuilder* builder) - -> xla::StatusOr> { - auto swaps = loop_vars[0]; - auto indices = loop_vars[1]; - i = builder->Reshape(i, {1}); - // temp = indices[i] - auto temp = builder->DynamicSlice(indices, i, {1}); - // swap_index = swaps[i] - auto swap_index = builder->DynamicSlice(swaps, i, {1}); - // swap_value = indices[swaps[i]] - auto swap_value = builder->DynamicSlice(indices, swap_index, {1}); - // indices[i] = indices[swaps[i]] - indices = builder->DynamicUpdateSlice(indices, swap_value, i); - // indices[swaps[i]] = temp - indices = builder->DynamicUpdateSlice(indices, temp, swap_index); - return std::vector{swaps, indices}; - }; - // for i in range(n): - auto swap_loop_result = - XlaForEachIndex(n, xla::S32, swap_body_fn, {swaps, indices}, - "indices_swap_loop", builder) - .ValueOrDie(); - auto swapped_indices = swap_loop_result[1]; - - // Gather the data using the swapped indices as the shuffled order. - auto indices_tensor_shape = TensorShape({n}); - DataType type = ctx->expected_output_dtype(0); - xla::XlaOp gather; - OP_REQUIRES_OK(ctx, XlaGather(input, input_shape, swapped_indices, - indices_tensor_shape, - /*axis=*/0, /*indices_are_nd=*/false, type, - DT_INT32, builder, &gather)); - ctx->SetOutput(0, gather); + return; + } + + if (input_shape.dims() == 1) { + // For R1s, shuffle values by sorting instead of the obvious Fisher-Yates + // algorithm. Fisher-Yates is simple to implement and correct, but not + // easily parallelizable. For a sufficiently parallel architecture, it is + // faster to sort many times, than Fisher-Yates shuffle once. + + // Shuffle values by assigning each value a random key and sorting the + // keys. Keys can collide causing detectable patterns in the shuffled + // output. Collisions translates into more ascending sub-sequences in the + // shuffled output than would be expected by chance. To avoid collisions, + // the number of possible key values must be sufficiently large. + + // How are more than 2^32 keys created? In each loop iteration, the + // algorithm sorts by random keys. Conceptually, the earlier iterations + // are sorting on the lower-order bits of larger keys that are never + // actually assembled. + + // The expected number of collisions is n - d + d(1 - 1/d)^n, where d is + // the number of possible keys and n is the number of values. If d = n^2, + // then the limit as n goes to infinity is 1/2. If d = n^3, then the limit + // as n goes to infinity is zero. + + // This implementation ensures that the key-space is greater than or equal + // to the cube of the number of values. The risk of collisions can be + // further reduced by increasing Exponent at the expense of + // performance. + + // For Exponent = 2, the expected number of collisions per shuffle is + // maximized at n = floor((2^32-1)^(1/2)) = 65535 where the expectation is + // about 1/2. + + // For Exponent = 3, the expected number of collisions per shuffle is + // maximized at n = floor((2^32-1)^(1/3)) = 1625 where the expectation is + // about 1/3255. + + // For Exponent = 4, the expected number of collisions per shuffle is + // maximized at n = floor((2^32-1)^(1/4)) = 255 where the expectation is + // about 1/132622. + constexpr int Exponent = 3; + const int rounds = static_cast( + std::ceil(Exponent * std::log(num_elements) / std::log(kuint32max))); + + const xla::Shape key_shape = + xla::ShapeUtil::MakeShape(xla::U32, {num_elements}); + xla::XlaOp zero = xla::ConstantR0(builder, 0U); + + // Unfortunately, xla::RngUniform gives values in the half open interval + // rather than the closed interval, so instead of 2^32 possible keys there + // are only 2^32 - 1 (kuint32max). + xla::XlaOp max_value = xla::ConstantR0(builder, kuint32max); + + xla::XlaOp curr = input; + for (int i = 0; i < rounds; ++i) { + xla::XlaOp keys = xla::RngUniform(zero, max_value, key_shape); + xla::XlaOp sorted = xla::Sort(keys, curr); + curr = xla::GetTupleElement(sorted, 1); + } + + ctx->SetOutput(0, curr); + return; } + + // The Fisher-Yates algorithm. + + // Generate the random swaps for the indices. + auto swaps_shape = xla::ShapeUtil::MakeShape(xla::S32, {n}); + auto swaps = + xla::RngUniform(xla::ConstantR0(builder, 0), + xla::ConstantR0(builder, n), swaps_shape); + + // Generate range(n) as the initial value for the indices to be swapped. + xla::XlaOp indices = xla::Iota(builder, xla::S32, n); + + // Swap the indices at i and swaps[i]. + auto swap_body_fn = [&](xla::XlaOp i, gtl::ArraySlice loop_vars, + xla::XlaBuilder* builder) + -> xla::StatusOr> { + auto swaps = loop_vars[0]; + auto indices = loop_vars[1]; + i = xla::Reshape(i, {1}); + // temp = indices[i] + auto temp = xla::DynamicSlice(indices, i, {1}); + // swap_index = swaps[i] + auto swap_index = xla::DynamicSlice(swaps, i, {1}); + // swap_value = indices[swaps[i]] + auto swap_value = xla::DynamicSlice(indices, swap_index, {1}); + // indices[i] = indices[swaps[i]] + indices = xla::DynamicUpdateSlice(indices, swap_value, i); + // indices[swaps[i]] = temp + indices = xla::DynamicUpdateSlice(indices, temp, swap_index); + return std::vector{swaps, indices}; + }; + // for i in range(n): + auto swap_loop_result = + XlaForEachIndex(n, xla::S32, swap_body_fn, {swaps, indices}, + "indices_swap_loop", builder) + .ValueOrDie(); + auto swapped_indices = swap_loop_result[1]; + + // Gather the data using the swapped indices as the shuffled order. + auto indices_tensor_shape = TensorShape({n}); + DataType type = ctx->expected_output_dtype(0); + xla::XlaOp gather; + OP_REQUIRES_OK(ctx, XlaGather(input, input_shape, swapped_indices, + indices_tensor_shape, + /*axis=*/0, /*indices_are_nd=*/false, type, + DT_INT32, builder, &gather)); + ctx->SetOutput(0, gather); } private: @@ -153,7 +219,7 @@ class RandomUniformIntOp : public XlaOpKernel { auto minval = ctx->Input(1); auto maxval = ctx->Input(2); - ctx->SetOutput(0, ctx->builder()->RngUniform(minval, maxval, xla_shape)); + ctx->SetOutput(0, xla::RngUniform(minval, maxval, xla_shape)); } private: @@ -179,8 +245,8 @@ class RandomStandardNormalOp : public XlaOpKernel { xla::XlaBuilder* b = ctx->builder(); // Normal distribution with a mean of 0 and a standard deviation of 1: - xla::XlaOp result = b->RngNormal(XlaHelpers::Zero(b, dtype), - XlaHelpers::One(b, dtype), xla_shape); + xla::XlaOp result = xla::RngNormal(XlaHelpers::Zero(b, dtype), + XlaHelpers::One(b, dtype), xla_shape); ctx->SetOutput(0, result); } @@ -209,10 +275,8 @@ class TruncatedNormalOp : public XlaOpKernel { xla::XlaOp one = XlaHelpers::FloatLiteral(b, dtype, 1.0); xla::XlaOp min_positive = XlaHelpers::FloatLiteral(b, dtype, std::numeric_limits::min()); - auto uniform = b->RngUniform(min_positive, one, xla_shape); - auto truncated_normal_or_status = TruncatedNormal(dtype, uniform, b); - OP_REQUIRES_OK(ctx, truncated_normal_or_status.status()); - ctx->SetOutput(0, truncated_normal_or_status.ValueOrDie()); + auto uniform = xla::RngUniform(min_positive, one, xla_shape); + ctx->SetOutput(0, TruncatedNormal(uniform)); } }; @@ -221,5 +285,5 @@ REGISTER_XLA_OP(Name("TruncatedNormal") .TypeConstraint("dtype", DT_FLOAT), TruncatedNormalOp); -} // anonymous namespace +} // namespace } // namespace tensorflow diff --git a/tensorflow/compiler/tf2xla/kernels/reduce_window_op.cc b/tensorflow/compiler/tf2xla/kernels/reduce_window_op.cc index 08894489ac77bbbe4ddb067c06a6d031a537697d..76bd1e62aa1efd85d6ed489b9a6d22a2bacf2a8b 100644 --- a/tensorflow/compiler/tf2xla/kernels/reduce_window_op.cc +++ b/tensorflow/compiler/tf2xla/kernels/reduce_window_op.cc @@ -19,6 +19,7 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/xla_compiler.h" #include "tensorflow/compiler/tf2xla/xla_op_kernel.h" #include "tensorflow/compiler/tf2xla/xla_op_registry.h" +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/core/framework/function.h" #include "tensorflow/core/framework/op_kernel.h" @@ -98,10 +99,10 @@ class ReduceWindowOp : public XlaOpKernel { { std::unique_ptr cb = builder->CreateSubBuilder("wrapper"); - auto x = cb->Parameter(0, scalar_shape, "x"); - auto y = cb->Parameter(1, scalar_shape, "y"); - auto outputs = cb->Call(*reducer.computation, {x, y}); - cb->GetTupleElement(outputs, 0); + auto x = xla::Parameter(cb.get(), 0, scalar_shape, "x"); + auto y = xla::Parameter(cb.get(), 1, scalar_shape, "y"); + auto outputs = xla::Call(cb.get(), *reducer.computation, {x, y}); + xla::GetTupleElement(outputs, 0); xla::StatusOr result = cb->Build(); OP_REQUIRES_OK(context, result.status()); wrapper = std::move(result.ValueOrDie()); @@ -112,7 +113,7 @@ class ReduceWindowOp : public XlaOpKernel { padding[i] = {padding_low_[i], padding_high_[i]}; } - xla::XlaOp output = builder->ReduceWindowWithGeneralPadding( + xla::XlaOp output = xla::ReduceWindowWithGeneralPadding( context->Input(0), context->Input(1), wrapper, window_dimensions_, window_strides_, padding); context->SetOutput(0, output); diff --git a/tensorflow/compiler/tf2xla/kernels/reduction_ops.cc b/tensorflow/compiler/tf2xla/kernels/reduction_ops.cc index 0f425637795e9633a8e36f921000ee2f5e25813a..be7f2bce8cb249aa51ca091e02da7dffc7d06743 100644 --- a/tensorflow/compiler/tf2xla/kernels/reduction_ops.cc +++ b/tensorflow/compiler/tf2xla/kernels/reduction_ops.cc @@ -19,7 +19,9 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/type_util.h" #include "tensorflow/compiler/tf2xla/xla_helpers.h" #include "tensorflow/compiler/tf2xla/xla_op_registry.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/client/lib/constants.h" +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/core/framework/kernel_def_builder.h" namespace tensorflow { @@ -31,11 +33,11 @@ class SumOp : public XlaReductionOp { : XlaReductionOp(ctx, XlaHelpers::SumAccumulationType(ctx->input_type(0))) {} xla::XlaOp InitialValue(xla::XlaBuilder* builder) override { - return XlaHelpers::Zero(builder, reduction_type_); + return xla::Zero(builder, xla_reduction_type_); } void BuildReducer(xla::XlaBuilder* builder, const xla::XlaOp& scalar_lhs, const xla::XlaOp& scalar_rhs) override { - builder->Add(scalar_lhs, scalar_rhs); + xla::Add(scalar_lhs, scalar_rhs); } }; @@ -48,12 +50,12 @@ class ProdOp : public XlaReductionOp { XlaHelpers::SumAccumulationType(ctx->input_type(0))) {} xla::XlaOp InitialValue(xla::XlaBuilder* builder) override { - return XlaHelpers::One(builder, reduction_type_); + return xla::One(builder, xla_reduction_type_); } void BuildReducer(xla::XlaBuilder* builder, const xla::XlaOp& scalar_lhs, const xla::XlaOp& scalar_rhs) override { - builder->Mul(scalar_lhs, scalar_rhs); + xla::Mul(scalar_lhs, scalar_rhs); } }; @@ -66,12 +68,12 @@ class MinOp : public XlaReductionOp { : XlaReductionOp(ctx, ctx->input_type(0)) {} xla::XlaOp InitialValue(xla::XlaBuilder* builder) override { - return XlaHelpers::MaxValue(builder, reduction_type_); + return xla::MaxValue(builder, xla_reduction_type_); } void BuildReducer(xla::XlaBuilder* builder, const xla::XlaOp& scalar_lhs, const xla::XlaOp& scalar_rhs) override { - builder->Min(scalar_lhs, scalar_rhs); + xla::Min(scalar_lhs, scalar_rhs); } }; @@ -83,12 +85,12 @@ class MaxOp : public XlaReductionOp { : XlaReductionOp(ctx, ctx->input_type(0)) {} xla::XlaOp InitialValue(xla::XlaBuilder* builder) override { - return XlaHelpers::MinValue(builder, reduction_type_); + return xla::MinValue(builder, xla_reduction_type_); } void BuildReducer(xla::XlaBuilder* builder, const xla::XlaOp& scalar_lhs, const xla::XlaOp& scalar_rhs) override { - builder->Max(scalar_lhs, scalar_rhs); + xla::Max(scalar_lhs, scalar_rhs); } }; @@ -101,11 +103,11 @@ class MeanOp : public XlaReductionOp { XlaHelpers::SumAccumulationType(ctx->input_type(0))) {} xla::XlaOp InitialValue(xla::XlaBuilder* builder) override { - return XlaHelpers::Zero(builder, reduction_type_); + return xla::Zero(builder, xla_reduction_type_); } void BuildReducer(xla::XlaBuilder* builder, const xla::XlaOp& scalar_lhs, const xla::XlaOp& scalar_rhs) override { - builder->Add(scalar_lhs, scalar_rhs); + xla::Add(scalar_lhs, scalar_rhs); } xla::XlaOp BuildFinalizer(xla::XlaBuilder* builder, @@ -113,7 +115,7 @@ class MeanOp : public XlaReductionOp { int64 num_elements_reduced) override { auto divisor = XlaHelpers::IntegerLiteral(builder, input_type(0), num_elements_reduced); - return builder->Div(reduce_output, divisor); + return reduce_output / divisor; } }; @@ -126,12 +128,12 @@ class AllOp : public XlaReductionOp { : XlaReductionOp(ctx, ctx->input_type(0)) {} xla::XlaOp InitialValue(xla::XlaBuilder* builder) override { - return builder->ConstantR0(true); + return xla::ConstantR0(builder, true); } void BuildReducer(xla::XlaBuilder* builder, const xla::XlaOp& scalar_lhs, const xla::XlaOp& scalar_rhs) override { - builder->And(scalar_lhs, scalar_rhs); + xla::And(scalar_lhs, scalar_rhs); } }; @@ -143,12 +145,12 @@ class AnyOp : public XlaReductionOp { : XlaReductionOp(ctx, ctx->input_type(0)) {} xla::XlaOp InitialValue(xla::XlaBuilder* builder) override { - return builder->ConstantR0(false); + return xla::ConstantR0(builder, false); } void BuildReducer(xla::XlaBuilder* builder, const xla::XlaOp& scalar_lhs, const xla::XlaOp& scalar_rhs) override { - builder->Or(scalar_lhs, scalar_rhs); + xla::Or(scalar_lhs, scalar_rhs); } }; diff --git a/tensorflow/compiler/tf2xla/kernels/reduction_ops.h b/tensorflow/compiler/tf2xla/kernels/reduction_ops.h index 2ecfb854a1c8625524d4f1199af3927edd204926..8333f9b288e27efe9497306f031980c9eec7c99c 100644 --- a/tensorflow/compiler/tf2xla/kernels/reduction_ops.h +++ b/tensorflow/compiler/tf2xla/kernels/reduction_ops.h @@ -64,6 +64,7 @@ class XlaReductionOp : public XlaOpKernel { protected: DataType reduction_type_; + xla::PrimitiveType xla_reduction_type_; }; } // namespace tensorflow diff --git a/tensorflow/compiler/tf2xla/kernels/reduction_ops_common.cc b/tensorflow/compiler/tf2xla/kernels/reduction_ops_common.cc index 44510c731e0dcb94c7f864053354b7a6a42d93f9..ed1d1c661091fd9bc443336626e593e728b82830 100644 --- a/tensorflow/compiler/tf2xla/kernels/reduction_ops_common.cc +++ b/tensorflow/compiler/tf2xla/kernels/reduction_ops_common.cc @@ -19,7 +19,8 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/type_util.h" #include "tensorflow/compiler/tf2xla/xla_helpers.h" #include "tensorflow/compiler/tf2xla/xla_op_kernel.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/core/framework/kernel_def_builder.h" namespace tensorflow { @@ -31,6 +32,8 @@ XlaReductionOp::XlaReductionOp(OpKernelConstruction* ctx, OP_REQUIRES_OK(ctx, ctx->MatchSignature({dt, DT_INT32}, {dt})); OP_REQUIRES_OK(ctx, ctx->GetAttr("keep_dims", &keep_dims_)); + OP_REQUIRES_OK( + ctx, DataTypeToPrimitiveType(reduction_type_, &xla_reduction_type_)); } // Unless BuildFinalizer is overridden the reduction has no @@ -101,20 +104,20 @@ void XlaReductionOp::Compile(XlaOpKernelContext* ctx) { xla::PrimitiveType type; TF_CHECK_OK(DataTypeToPrimitiveType(reduction_type_, &type)); - auto data = b->ConvertElementType(ctx->Input(0), type); + auto data = xla::ConvertElementType(ctx->Input(0), type); // Call virtual method to get the initial value. - auto initial = b->ConvertElementType(InitialValue(b), type); + auto initial = xla::ConvertElementType(InitialValue(b), type); // Make two scalar parameters of the desired type for the lambda. - auto rx = r.Parameter(0, xla::ShapeUtil::MakeShape(type, {}), "x"); - auto ry = r.Parameter(1, xla::ShapeUtil::MakeShape(type, {}), "y"); + auto rx = xla::Parameter(&r, 0, xla::ShapeUtil::MakeShape(type, {}), "x"); + auto ry = xla::Parameter(&r, 1, xla::ShapeUtil::MakeShape(type, {}), "y"); // Call virtual method to build the reduction lambda. BuildReducer(&r, rx, ry); xla::XlaComputation reduction_computation = r.Build().ConsumeValueOrDie(); - auto reduce = b->Reduce(data, initial, reduction_computation, xla_axes); + auto reduce = xla::Reduce(data, initial, reduction_computation, xla_axes); auto deconverted = XlaHelpers::ConvertElementType(b, reduce, input_type(0)); auto finalized = BuildFinalizer(b, deconverted, num_elements_reduced); - auto result = keep_dims_ ? b->Reshape(finalized, final_shape) : finalized; + auto result = keep_dims_ ? xla::Reshape(finalized, final_shape) : finalized; ctx->SetOutput(0, result); } diff --git a/tensorflow/compiler/tf2xla/kernels/relu_op.cc b/tensorflow/compiler/tf2xla/kernels/relu_op.cc index ba7d484d53d7258edaa5bc42fa116cf16e94835b..f4b804e54677c7226d8d3429c9e8c27686d19ccf 100644 --- a/tensorflow/compiler/tf2xla/kernels/relu_op.cc +++ b/tensorflow/compiler/tf2xla/kernels/relu_op.cc @@ -19,7 +19,7 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/xla_helpers.h" #include "tensorflow/compiler/tf2xla/xla_op_registry.h" #include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/core/framework/kernel_def_builder.h" #include "tensorflow/core/framework/types.h" #include "tensorflow/core/kernels/no_op.h" @@ -34,7 +34,7 @@ class ReluOp : public XlaOpKernel { void Compile(XlaOpKernelContext* ctx) override { xla::XlaBuilder* builder = ctx->builder(); auto zero = XlaHelpers::Zero(builder, input_type(0)); - ctx->SetOutput(0, builder->Max(zero, ctx->Input(0))); + ctx->SetOutput(0, xla::Max(zero, ctx->Input(0))); } }; @@ -46,7 +46,7 @@ class Relu6Op : public XlaOpKernel { xla::XlaBuilder* builder = ctx->builder(); auto zero = XlaHelpers::Zero(builder, input_type(0)); auto six = XlaHelpers::IntegerLiteral(builder, input_type(0), 6); - ctx->SetOutput(0, builder->Clamp(zero, ctx->Input(0), six)); + ctx->SetOutput(0, xla::Clamp(zero, ctx->Input(0), six)); } }; @@ -59,9 +59,9 @@ class ReluGradOp : public XlaOpKernel { xla::XlaBuilder* b = ctx->builder(); const TensorShape shape = ctx->InputShape(0); const auto zero = - b->Broadcast(XlaHelpers::Zero(b, input_type(0)), shape.dim_sizes()); - const auto pred = b->Gt(ctx->Input(1), zero); - ctx->SetOutput(0, b->Select(pred, ctx->Input(0), zero)); + xla::Broadcast(XlaHelpers::Zero(b, input_type(0)), shape.dim_sizes()); + const auto pred = xla::Gt(ctx->Input(1), zero); + ctx->SetOutput(0, xla::Select(pred, ctx->Input(0), zero)); } }; @@ -74,12 +74,12 @@ class Relu6GradOp : public XlaOpKernel { xla::XlaBuilder* b = ctx->builder(); const TensorShape shape = ctx->InputShape(0); const auto zero = - b->Broadcast(XlaHelpers::Zero(b, input_type(0)), shape.dim_sizes()); - const auto six = b->Broadcast( + xla::Broadcast(XlaHelpers::Zero(b, input_type(0)), shape.dim_sizes()); + const auto six = xla::Broadcast( XlaHelpers::IntegerLiteral(b, input_type(0), 6), shape.dim_sizes()); - auto out = - b->Select(b->And(b->Lt(ctx->Input(1), six), b->Gt(ctx->Input(1), zero)), - ctx->Input(0), zero); + auto out = xla::Select( + xla::And(xla::Lt(ctx->Input(1), six), xla::Gt(ctx->Input(1), zero)), + ctx->Input(0), zero); ctx->SetOutput(0, out); } }; diff --git a/tensorflow/compiler/tf2xla/kernels/reshape_op.cc b/tensorflow/compiler/tf2xla/kernels/reshape_op.cc index af4d64b159c09ed7e01017f25a2b23e58542dc3c..354fec9be75e9559b204e2afd6ee08dfc7cea872 100644 --- a/tensorflow/compiler/tf2xla/kernels/reshape_op.cc +++ b/tensorflow/compiler/tf2xla/kernels/reshape_op.cc @@ -19,7 +19,8 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/xla_helpers.h" #include "tensorflow/compiler/tf2xla/xla_op_kernel.h" #include "tensorflow/compiler/tf2xla/xla_op_registry.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/core/framework/op_kernel.h" #include "tensorflow/core/framework/register_types.h" #include "tensorflow/core/framework/tensor.h" @@ -90,8 +91,7 @@ class ReshapeOp : public XlaOpKernel { VLOG(1) << "Reshape " << input_shape.DebugString() << " " << shape.DebugString(); - ctx->SetOutput(0, - ctx->builder()->Reshape(ctx->Input(0), shape.dim_sizes())); + ctx->SetOutput(0, xla::Reshape(ctx->Input(0), shape.dim_sizes())); } }; diff --git a/tensorflow/compiler/tf2xla/kernels/retval_op.cc b/tensorflow/compiler/tf2xla/kernels/retval_op.cc index a711278638444be01fb865561957702368b75114..5be70a4ded31a988cb77cdabe3fc8a041bc3ad16 100644 --- a/tensorflow/compiler/tf2xla/kernels/retval_op.cc +++ b/tensorflow/compiler/tf2xla/kernels/retval_op.cc @@ -17,6 +17,7 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/xla_op_kernel.h" #include "tensorflow/compiler/tf2xla/xla_op_registry.h" #include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" +#include "tensorflow/compiler/xla/status_macros.h" #include "tensorflow/core/framework/kernel_def_builder.h" #include "tensorflow/core/framework/op_kernel.h" @@ -62,15 +63,24 @@ class RetvalOp : public XlaOpKernel { OP_REQUIRES_OK(ctx, tc.AddConstRetval(index_, dtype_, literal)); } else { TensorShape shape = ctx->InputShape(0); - TensorShape representation_shape = - tc.is_entry_computation() - ? tc.RepresentationShape(shape, ctx->input_type(0)) - : shape; + ctx->SetStatus(is_constant.status()); + TensorShape representation_shape; + if (tc.is_entry_computation()) { + xla::StatusOr shape_or_status = + tc.RepresentationShape(shape, ctx->input_type(0)); + if (!shape_or_status.ok()) { + ctx->SetStatus(shape_or_status.status()); + return; + } else { + representation_shape = shape_or_status.ValueOrDie(); + } + } else { + representation_shape = shape; + } xla::XlaOp output = input; if (tc.is_entry_computation()) { - output = - ctx->builder()->Reshape(input, representation_shape.dim_sizes()); + output = xla::Reshape(input, representation_shape.dim_sizes()); } else { // The core from which a return value is returned depends on the // device assignment of the input to the retval. Since we can't change @@ -78,8 +88,8 @@ class RetvalOp : public XlaOpKernel { // introduce an operator here, even if the shape does not change. // TODO(b/76097077): propagate device assignments onto arguments and // return values of functions, and then reshape unconditionally. - output = ctx->builder()->GetTupleElement( - ctx->builder()->Tuple({output}), 0); + output = + xla::GetTupleElement(xla::Tuple(ctx->builder(), {output}), 0); } tc.AddRetval(index_, dtype_, shape, output); } diff --git a/tensorflow/compiler/tf2xla/kernels/reverse_op.cc b/tensorflow/compiler/tf2xla/kernels/reverse_op.cc index 2872a3c4d49d0d269aa3d216887a5c32cd51f1c3..ec15b4cc7a523d5b8d4287bbe3321433f315063b 100644 --- a/tensorflow/compiler/tf2xla/kernels/reverse_op.cc +++ b/tensorflow/compiler/tf2xla/kernels/reverse_op.cc @@ -19,7 +19,8 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/xla_helpers.h" #include "tensorflow/compiler/tf2xla/xla_op_kernel.h" #include "tensorflow/compiler/tf2xla/xla_op_registry.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/core/framework/op_kernel.h" #include "tensorflow/core/framework/register_types.h" #include "tensorflow/core/framework/tensor.h" @@ -62,7 +63,7 @@ class ReverseOp : public XlaOpKernel { } } - ctx->SetOutput(0, ctx->builder()->Rev(ctx->Input(0), dimensions)); + ctx->SetOutput(0, xla::Rev(ctx->Input(0), dimensions)); } }; @@ -100,7 +101,7 @@ class ReverseV2Op : public XlaOpKernel { x_shape.dims(), ").")); } - ctx->SetOutput(0, ctx->builder()->Rev(ctx->Input(0), axes)); + ctx->SetOutput(0, xla::Rev(ctx->Input(0), axes)); } }; diff --git a/tensorflow/compiler/tf2xla/kernels/reverse_sequence_op.cc b/tensorflow/compiler/tf2xla/kernels/reverse_sequence_op.cc index 5d1c05268493f4f6404c40a4092a71f1e5b3f3b9..c810456f94322acfccae18d78efa861eede4648c 100644 --- a/tensorflow/compiler/tf2xla/kernels/reverse_sequence_op.cc +++ b/tensorflow/compiler/tf2xla/kernels/reverse_sequence_op.cc @@ -17,6 +17,8 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/xla_helpers.h" #include "tensorflow/compiler/tf2xla/xla_op_kernel.h" #include "tensorflow/compiler/tf2xla/xla_op_registry.h" +#include "tensorflow/compiler/xla/client/lib/numeric.h" +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/core/framework/tensor_shape.h" namespace tensorflow { @@ -85,103 +87,96 @@ class ReverseSequenceOp : public XlaOpKernel { auto condition_builder = builder->CreateSubBuilder("reverse_sequence_condition"); { - auto param = condition_builder->Parameter(0, tuple_shape, "param"); - auto i = condition_builder->GetTupleElement(param, 0); - condition_builder->Lt( - i, XlaHelpers::IntegerLiteral(condition_builder.get(), seq_lens_type, - batch_size)); + auto param = + xla::Parameter(condition_builder.get(), 0, tuple_shape, "param"); + auto i = xla::GetTupleElement(param, 0); + xla::Lt(i, XlaHelpers::IntegerLiteral(condition_builder.get(), + seq_lens_type, batch_size)); } auto condition = condition_builder->Build(); OP_REQUIRES_OK(context, condition.status()); auto body_builder = builder->CreateSubBuilder("reverse_sequence_body"); { - auto param = body_builder->Parameter(0, tuple_shape, "param"); - auto i = body_builder->GetTupleElement(param, 0); - auto seq_lens = body_builder->GetTupleElement(param, 1); - auto output = body_builder->GetTupleElement(param, 2); + auto param = xla::Parameter(body_builder.get(), 0, tuple_shape, "param"); + auto i = xla::GetTupleElement(param, 0); + auto seq_lens = xla::GetTupleElement(param, 1); + auto output = xla::GetTupleElement(param, 2); // seq_len is the sequence length of the current batch element (rank 1) - auto seq_len = body_builder->DynamicSlice( - seq_lens, body_builder->Reshape(i, {1}), {1}); + auto seq_len = xla::DynamicSlice(seq_lens, xla::Reshape(i, {1}), {1}); // Indices is the offset of the batch element in the input. - auto batch_element_indices = body_builder->Broadcast( - XlaHelpers::Zero(body_builder.get(), seq_lens_type), - {input_shape.dims()}); - batch_element_indices = body_builder->DynamicUpdateSlice( - batch_element_indices, body_builder->Reshape(i, {1}), - body_builder->Reshape( - XlaHelpers::IntegerLiteral(body_builder.get(), seq_lens_type, - batch_dim_), - {1})); + auto batch_element_indices = + xla::Broadcast(XlaHelpers::Zero(body_builder.get(), seq_lens_type), + {input_shape.dims()}); + batch_element_indices = xla::DynamicUpdateSlice( + batch_element_indices, xla::Reshape(i, {1}), + xla::Reshape(XlaHelpers::IntegerLiteral(body_builder.get(), + seq_lens_type, batch_dim_), + {1})); // Slice out the current batch element and pad it out in the sequence // dimension. TensorShape slice_shape = input_shape; slice_shape.set_dim(batch_dim_, 1); slice_shape.set_dim(seq_dim_, max_seq_len); - auto slice = body_builder->DynamicSlice(output, batch_element_indices, - slice_shape.dim_sizes()); + auto slice = xla::DynamicSlice(output, batch_element_indices, + slice_shape.dim_sizes()); auto padding_config = xla::MakeNoPaddingConfig(slice_shape.dims()); padding_config.mutable_dimensions(seq_dim_)->set_edge_padding_high( slice_shape.dim_size(seq_dim_)); - slice = body_builder->Pad( - slice, XlaHelpers::Zero(body_builder.get(), input_type), - padding_config); + slice = xla::Pad(slice, XlaHelpers::Zero(body_builder.get(), input_type), + padding_config); // Now slice out the reversed sequence from its actual start. // sequence_start_indices is the offset of the start of the reversed // sequence in the input. The slice will go into the padding, however, we // will mask off these elements and replace them with elements from the // original input so their values do not matter. - auto sequence_start_indices = body_builder->Broadcast( - XlaHelpers::Zero(body_builder.get(), seq_lens_type), - {slice_shape.dims()}); - sequence_start_indices = body_builder->DynamicUpdateSlice( + auto sequence_start_indices = + xla::Broadcast(XlaHelpers::Zero(body_builder.get(), seq_lens_type), + {slice_shape.dims()}); + sequence_start_indices = xla::DynamicUpdateSlice( sequence_start_indices, - body_builder->Sub(XlaHelpers::IntegerLiteral( - body_builder.get(), seq_lens_type, max_seq_len), - seq_len), - body_builder->Reshape( - XlaHelpers::IntegerLiteral(body_builder.get(), seq_lens_type, - seq_dim_), - {1})); - slice = body_builder->DynamicSlice(slice, sequence_start_indices, - slice_shape.dim_sizes()); + xla::Sub(XlaHelpers::IntegerLiteral(body_builder.get(), seq_lens_type, + max_seq_len), + seq_len), + xla::Reshape(XlaHelpers::IntegerLiteral(body_builder.get(), + seq_lens_type, seq_dim_), + {1})); + slice = xla::DynamicSlice(slice, sequence_start_indices, + slice_shape.dim_sizes()); // Shift the reversed sequence to the left. - output = body_builder->DynamicUpdateSlice(output, slice, - batch_element_indices); + output = xla::DynamicUpdateSlice(output, slice, batch_element_indices); - body_builder->Tuple( - {body_builder->Add( - i, XlaHelpers::One(body_builder.get(), seq_lens_type)), + xla::Tuple( + body_builder.get(), + {xla::Add(i, XlaHelpers::One(body_builder.get(), seq_lens_type)), seq_lens, output}); } auto body = body_builder->Build(); OP_REQUIRES_OK(context, body.status()); - auto loop_output = builder->While( + auto loop_output = xla::While( condition.ValueOrDie(), body.ValueOrDie(), - builder->Tuple({XlaHelpers::Zero(builder, seq_lens_type), seq_lens, - builder->Rev(input, {seq_dim_})})); - auto output = builder->GetTupleElement(loop_output, 2); + xla::Tuple(builder, {XlaHelpers::Zero(builder, seq_lens_type), seq_lens, + xla::Rev(input, {seq_dim_})})); + auto output = xla::GetTupleElement(loop_output, 2); // Mask out elements after the sequence length. - xla::XlaOp iota; - OP_REQUIRES_OK( - context, XlaHelpers::Iota(builder, seq_lens_type, max_seq_len, &iota)); + xla::XlaOp iota = + xla::Iota(builder, seq_lens_xla_shape.element_type(), max_seq_len); std::vector dims(input_shape.dims(), 1); dims[batch_dim_] = batch_size; - auto mask = builder->Lt(iota, builder->Reshape(seq_lens, dims), {seq_dim_}); + auto mask = xla::Lt(iota, xla::Reshape(seq_lens, dims), {seq_dim_}); // Broadcast the mask up to the input shape. - mask = - builder->Or(mask, builder->Broadcast(builder->ConstantR0(false), - input_shape.dim_sizes())); + mask = xla::Or(mask, xla::Broadcast(xla::ConstantR0(builder, false), + input_shape.dim_sizes())); - output = builder->Select(mask, output, input); + output = xla::Select(mask, output, input); context->SetOutput(0, output); } diff --git a/tensorflow/compiler/tf2xla/kernels/scan_ops.cc b/tensorflow/compiler/tf2xla/kernels/scan_ops.cc index 1819fb543317eed15b2fe0518d74aba5c564697d..27ab3e1bf5b81a901e64a242e5eb343591481efe 100644 --- a/tensorflow/compiler/tf2xla/kernels/scan_ops.cc +++ b/tensorflow/compiler/tf2xla/kernels/scan_ops.cc @@ -20,7 +20,8 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/xla_helpers.h" #include "tensorflow/compiler/tf2xla/xla_op_kernel.h" #include "tensorflow/compiler/tf2xla/xla_op_registry.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/core/framework/op_kernel.h" #include "tensorflow/core/framework/partial_tensor_shape.h" #include "tensorflow/core/framework/register_types.h" @@ -100,7 +101,7 @@ class ScanOp : public XlaOpKernel { init = XlaHelpers::One(builder, dtype); reducer = ctx->GetOrCreateMul(dtype); } - auto output = builder->ReduceWindowWithGeneralPadding( + auto output = xla::ReduceWindowWithGeneralPadding( XlaHelpers::ConvertElementType(builder, ctx->Input(0), dtype), init, *reducer, window_dims, window_strides, padding); output = @@ -110,12 +111,12 @@ class ScanOp : public XlaOpKernel { // of all the input elements. Slice off this extra "last" element. if (exclusive_) { if (reverse_) { - output = builder->SliceInDim(output, 1, input_shape.dim_size(axis) + 1, - 1, axis); + output = + xla::SliceInDim(output, 1, input_shape.dim_size(axis) + 1, 1, axis); } else { output = - builder->SliceInDim(output, 0, input_shape.dim_size(axis), 1, axis); + xla::SliceInDim(output, 0, input_shape.dim_size(axis), 1, axis); } } ctx->SetOutput(0, output); diff --git a/tensorflow/compiler/tf2xla/kernels/scatter_nd_op.cc b/tensorflow/compiler/tf2xla/kernels/scatter_nd_op.cc index f2c63b4f9083ad3c7dd7cf318dc22def1e99fa9f..14709bb6cbce4b3ae0f7ff859b0fa622c6eda293 100644 --- a/tensorflow/compiler/tf2xla/kernels/scatter_nd_op.cc +++ b/tensorflow/compiler/tf2xla/kernels/scatter_nd_op.cc @@ -19,6 +19,7 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/xla_helpers.h" #include "tensorflow/compiler/tf2xla/xla_op_kernel.h" #include "tensorflow/compiler/tf2xla/xla_op_registry.h" +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/compiler/xla/status_macros.h" #include "tensorflow/core/framework/kernel_def_builder.h" #include "tensorflow/core/framework/op_kernel.h" @@ -103,8 +104,8 @@ class ScatterNdOp : public XlaOpKernel { updates_shape)); xla::XlaBuilder* builder = context->builder(); - auto buffer = builder->Broadcast(XlaHelpers::Zero(builder, dtype), - buffer_shape.dim_sizes()); + auto buffer = xla::Broadcast(XlaHelpers::Zero(builder, dtype), + buffer_shape.dim_sizes()); auto indices = context->Input(0); auto updates = context->Input(1); auto result = diff --git a/tensorflow/compiler/tf2xla/kernels/segment_reduction_ops.cc b/tensorflow/compiler/tf2xla/kernels/segment_reduction_ops.cc index 664078ca16c6d5d4b57c4a8c661ad0848f30dd7d..e2ac7da2c2630725efe3dbcc51c3f3d30e7aca2c 100644 --- a/tensorflow/compiler/tf2xla/kernels/segment_reduction_ops.cc +++ b/tensorflow/compiler/tf2xla/kernels/segment_reduction_ops.cc @@ -14,20 +14,30 @@ limitations under the License. ==============================================================================*/ #include "tensorflow/compiler/tf2xla/lib/scatter.h" +#include "tensorflow/compiler/tf2xla/type_util.h" #include "tensorflow/compiler/tf2xla/xla_helpers.h" #include "tensorflow/compiler/tf2xla/xla_op_kernel.h" #include "tensorflow/compiler/tf2xla/xla_op_registry.h" +#include "tensorflow/compiler/xla/client/lib/constants.h" #include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" namespace tensorflow { namespace { -class UnsortedSegmentSum : public XlaOpKernel { +class UnsortedSegmentReduce : public XlaOpKernel { public: - explicit UnsortedSegmentSum(OpKernelConstruction* ctx) : XlaOpKernel(ctx) { - OP_REQUIRES_OK(ctx, ctx->GetAttr("T", &dtype_)); + explicit UnsortedSegmentReduce(OpKernelConstruction* ctx) : XlaOpKernel(ctx) { + DataType dtype; + OP_REQUIRES_OK(ctx, ctx->GetAttr("T", &dtype)); + OP_REQUIRES_OK(ctx, DataTypeToPrimitiveType(dtype, &type_)); } + // The initial value to initialize elements of the output to. + virtual xla::XlaOp InitialValue(xla::XlaBuilder* builder) = 0; + + // A function to combine two scalars with the same index (e.g., sum). + virtual xla::XlaOp Combine(xla::XlaOp a, xla::XlaOp b) = 0; + void Compile(XlaOpKernelContext* ctx) override { // output = unsorted_segment_sum(data, indices, num_segments) // Compute a tensor such that: @@ -50,28 +60,28 @@ class UnsortedSegmentSum : public XlaOpKernel { OP_REQUIRES_OK(ctx, ctx->ConstantInputAsIntScalar(2, &num_segments)); OP_REQUIRES(ctx, data_shape.dims() >= indices_shape.dims(), - errors::InvalidArgument( - "UnsortedSegmentSum requires that indices' rank be" - " less than or equal to data's rank.")); + errors::InvalidArgument(type_string(), + " requires that indices' rank be" + " less than or equal to data's rank.")); // Validate that indices.shape is a prefix of data.shape. for (int d = 0; d < indices_shape.dims(); ++d) { - OP_REQUIRES(ctx, (data_shape.dim_size(d) == indices_shape.dim_size(d)), - errors::InvalidArgument( - "UnsortedSegmentSum requires indices shape to be prefix" - " of data_shape, but dimension ", - d, " differs ", data_shape.dim_size(d), " vs. ", - indices_shape.dim_size(d))); + OP_REQUIRES( + ctx, (data_shape.dim_size(d) == indices_shape.dim_size(d)), + errors::InvalidArgument(type_string(), + " requires indices shape to be prefix" + " of data_shape, but dimension ", + d, " differs ", data_shape.dim_size(d), + " vs. ", indices_shape.dim_size(d))); } xla::XlaBuilder* builder = ctx->builder(); TensorShape buffer_shape = data_shape; buffer_shape.RemoveDimRange(0, indices_shape.dims()); buffer_shape.InsertDim(0, num_segments); - auto buffer = builder->Broadcast(XlaHelpers::Zero(builder, dtype_), - buffer_shape.dim_sizes()); + auto buffer = + xla::Broadcast(InitialValue(builder), buffer_shape.dim_sizes()); - auto combiner = [](xla::XlaOp a, xla::XlaOp b, xla::XlaBuilder* builder) { - return builder->Add(a, b); - }; + auto combiner = [this](xla::XlaOp a, xla::XlaOp b, + xla::XlaBuilder* builder) { return Combine(a, b); }; auto result = XlaScatter(buffer, /*updates=*/data, indices, /*indices_are_vectors=*/false, combiner, builder); @@ -79,13 +89,73 @@ class UnsortedSegmentSum : public XlaOpKernel { ctx->SetOutput(0, result.ValueOrDie()); } - private: - DataType dtype_; + protected: + xla::PrimitiveType type_; +}; + +class UnsortedSegmentSum : public UnsortedSegmentReduce { + public: + explicit UnsortedSegmentSum(OpKernelConstruction* ctx) + : UnsortedSegmentReduce(ctx) {} + + xla::XlaOp InitialValue(xla::XlaBuilder* builder) override { + return xla::Zero(builder, type_); + }; + xla::XlaOp Combine(xla::XlaOp a, xla::XlaOp b) override { return a + b; }; }; REGISTER_XLA_OP( Name("UnsortedSegmentSum").CompileTimeConstInput("num_segments"), UnsortedSegmentSum); +class UnsortedSegmentProd : public UnsortedSegmentReduce { + public: + explicit UnsortedSegmentProd(OpKernelConstruction* ctx) + : UnsortedSegmentReduce(ctx) {} + + xla::XlaOp InitialValue(xla::XlaBuilder* builder) override { + return xla::One(builder, type_); + }; + xla::XlaOp Combine(xla::XlaOp a, xla::XlaOp b) override { return a * b; }; +}; + +REGISTER_XLA_OP( + Name("UnsortedSegmentProd").CompileTimeConstInput("num_segments"), + UnsortedSegmentProd); + +class UnsortedSegmentMin : public UnsortedSegmentReduce { + public: + explicit UnsortedSegmentMin(OpKernelConstruction* ctx) + : UnsortedSegmentReduce(ctx) {} + + xla::XlaOp InitialValue(xla::XlaBuilder* builder) override { + return xla::MaxFiniteValue(builder, type_); + }; + xla::XlaOp Combine(xla::XlaOp a, xla::XlaOp b) override { + return xla::Min(a, b); + }; +}; + +REGISTER_XLA_OP( + Name("UnsortedSegmentMin").CompileTimeConstInput("num_segments"), + UnsortedSegmentMin); + +class UnsortedSegmentMax : public UnsortedSegmentReduce { + public: + explicit UnsortedSegmentMax(OpKernelConstruction* ctx) + : UnsortedSegmentReduce(ctx) {} + + xla::XlaOp InitialValue(xla::XlaBuilder* builder) override { + return xla::MinFiniteValue(builder, type_); + }; + xla::XlaOp Combine(xla::XlaOp a, xla::XlaOp b) override { + return xla::Max(a, b); + }; +}; + +REGISTER_XLA_OP( + Name("UnsortedSegmentMax").CompileTimeConstInput("num_segments"), + UnsortedSegmentMax); + } // namespace } // namespace tensorflow diff --git a/tensorflow/compiler/tf2xla/kernels/select_op.cc b/tensorflow/compiler/tf2xla/kernels/select_op.cc index f9f48164d63492b057d4950abfc2ca6153e44870..5c010c9df23ba6c7732d87fa014879d93ff586ce 100644 --- a/tensorflow/compiler/tf2xla/kernels/select_op.cc +++ b/tensorflow/compiler/tf2xla/kernels/select_op.cc @@ -19,6 +19,7 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/xla_helpers.h" #include "tensorflow/compiler/tf2xla/xla_op_kernel.h" #include "tensorflow/compiler/tf2xla/xla_op_registry.h" +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/core/framework/kernel_def_builder.h" #include "tensorflow/core/kernels/bounds_check.h" @@ -40,8 +41,6 @@ class SelectOp : public XlaOpKernel { "'then' and 'else' must have the same size. but received: ", then_shape.DebugString(), " vs. ", else_shape.DebugString())); - xla::XlaBuilder* builder = ctx->builder(); - auto cond_handle = ctx->Input(0); auto then_handle = ctx->Input(1); auto else_handle = ctx->Input(2); @@ -69,14 +68,14 @@ class SelectOp : public XlaOpKernel { const auto dim_sizes = then_shape.dim_sizes(); gtl::ArraySlice bdims = dim_sizes; bdims.pop_front(); - cond_handle = builder->Broadcast(cond_handle, bdims); + cond_handle = xla::Broadcast(cond_handle, bdims); std::vector dim_order(then_shape.dims()); dim_order[0] = then_shape.dims() - 1; std::iota(dim_order.begin() + 1, dim_order.end(), 0); - cond_handle = builder->Transpose(cond_handle, dim_order); + cond_handle = xla::Transpose(cond_handle, dim_order); } - ctx->SetOutput(0, builder->Select(cond_handle, then_handle, else_handle)); + ctx->SetOutput(0, xla::Select(cond_handle, then_handle, else_handle)); } private: diff --git a/tensorflow/compiler/tf2xla/kernels/sendrecv_ops.cc b/tensorflow/compiler/tf2xla/kernels/sendrecv_ops.cc index 9ce01d0d44509bbcbea18afdb4210a675834bb6d..6281d6c6533f7f49a269f5c7e52226ba0f1d29f6 100644 --- a/tensorflow/compiler/tf2xla/kernels/sendrecv_ops.cc +++ b/tensorflow/compiler/tf2xla/kernels/sendrecv_ops.cc @@ -45,7 +45,7 @@ void SendOp::Compile(XlaOpKernelContext* ctx) { XlaCompiler* compiler = XlaContext::Get(ctx).compiler(); xla::ChannelHandle channel; OP_REQUIRES_OK(ctx, compiler->GetChannelHandle(tensor_name_, &channel)); - ctx->builder()->Send(ctx->Input(0), channel); + xla::Send(ctx->Input(0), channel); } REGISTER_XLA_OP(Name("XlaSend"), SendOp); @@ -76,7 +76,7 @@ void RecvOp::Compile(XlaOpKernelContext* ctx) { XlaCompiler* compiler = XlaContext::Get(ctx).compiler(); xla::ChannelHandle channel; OP_REQUIRES_OK(ctx, compiler->GetChannelHandle(tensor_name_, &channel)); - ctx->SetOutput(0, ctx->builder()->Recv(shape_, channel)); + ctx->SetOutput(0, xla::Recv(ctx->builder(), shape_, channel)); } REGISTER_XLA_OP(Name("XlaRecv"), RecvOp); diff --git a/tensorflow/compiler/tf2xla/kernels/sequence_ops.cc b/tensorflow/compiler/tf2xla/kernels/sequence_ops.cc index bc3d0bf5dfe9e5af8e50a25e27db7148e05e0cfd..25a5bcbe1dd27d741ce3b74125ba9ce425ee78f3 100644 --- a/tensorflow/compiler/tf2xla/kernels/sequence_ops.cc +++ b/tensorflow/compiler/tf2xla/kernels/sequence_ops.cc @@ -18,7 +18,7 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/xla_helpers.h" #include "tensorflow/compiler/tf2xla/xla_op_kernel.h" #include "tensorflow/compiler/tf2xla/xla_op_registry.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/core/framework/op_kernel.h" #include "tensorflow/core/framework/register_types.h" #include "tensorflow/core/framework/tensor.h" diff --git a/tensorflow/compiler/tf2xla/kernels/shape_op.cc b/tensorflow/compiler/tf2xla/kernels/shape_op.cc index d59720bef742c7441ee01a954247013559bb909c..5798823cd54c66dd179e3611c0041f7c5a1ff2b5 100644 --- a/tensorflow/compiler/tf2xla/kernels/shape_op.cc +++ b/tensorflow/compiler/tf2xla/kernels/shape_op.cc @@ -20,6 +20,7 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/xla_helpers.h" #include "tensorflow/compiler/tf2xla/xla_op_kernel.h" #include "tensorflow/compiler/tf2xla/xla_op_registry.h" +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/core/framework/kernel_def_builder.h" #include "tensorflow/core/kernels/bounds_check.h" @@ -147,7 +148,7 @@ class ExpandDimsOp : public XlaOpKernel { dim = std::min(dim, existing_dims_size); new_shape.emplace(new_shape.begin() + dim, 1); - ctx->SetOutput(0, ctx->builder()->Reshape(ctx->Input(0), new_shape)); + ctx->SetOutput(0, xla::Reshape(ctx->Input(0), new_shape)); } }; REGISTER_XLA_OP(Name("ExpandDims").CompileTimeConstInput("dim"), ExpandDimsOp); @@ -204,7 +205,7 @@ class SqueezeOp : public XlaOpKernel { } } - ctx->SetOutput(0, ctx->builder()->Reshape(ctx->Input(0), new_shape)); + ctx->SetOutput(0, xla::Reshape(ctx->Input(0), new_shape)); } private: @@ -221,7 +222,7 @@ class ZerosLikeOp : public XlaOpKernel { const TensorShape input_shape = ctx->InputShape(0); auto zero = XlaHelpers::Zero(ctx->builder(), input_type(0)); - ctx->SetOutput(0, ctx->builder()->Broadcast(zero, input_shape.dim_sizes())); + ctx->SetOutput(0, xla::Broadcast(zero, input_shape.dim_sizes())); } }; @@ -235,7 +236,7 @@ class OnesLikeOp : public XlaOpKernel { const TensorShape input_shape = ctx->InputShape(0); auto one = XlaHelpers::One(ctx->builder(), input_type(0)); - ctx->SetOutput(0, ctx->builder()->Broadcast(one, input_shape.dim_sizes())); + ctx->SetOutput(0, xla::Broadcast(one, input_shape.dim_sizes())); } }; diff --git a/tensorflow/compiler/tf2xla/kernels/slice_op.cc b/tensorflow/compiler/tf2xla/kernels/slice_op.cc index be1e97bf26fa4cde1b741c8d0b843a85ce33a59c..1864584adee357ce35a3e8a38a4e3c58c356bfca 100644 --- a/tensorflow/compiler/tf2xla/kernels/slice_op.cc +++ b/tensorflow/compiler/tf2xla/kernels/slice_op.cc @@ -19,6 +19,7 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/xla_helpers.h" #include "tensorflow/compiler/tf2xla/xla_op_kernel.h" #include "tensorflow/compiler/tf2xla/xla_op_registry.h" +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/core/framework/op_kernel.h" #include "tensorflow/core/framework/register_types.h" #include "tensorflow/core/framework/tensor.h" @@ -92,8 +93,7 @@ class SliceOp : public XlaOpKernel { limits.push_back(begin[i] + size[i]); } std::vector strides(begin.size(), 1); - ctx->SetOutput( - 0, ctx->builder()->Slice(ctx->Input(0), begin, limits, strides)); + ctx->SetOutput(0, xla::Slice(ctx->Input(0), begin, limits, strides)); } else { // `begin` is not a compile-time constant. for (int i = 0; i < input_dims; ++i) { @@ -106,8 +106,7 @@ class SliceOp : public XlaOpKernel { input_shape.dim_size(i), "], but ", "got ", size[i])); } - ctx->SetOutput( - 0, ctx->builder()->DynamicSlice(ctx->Input(0), ctx->Input(1), size)); + ctx->SetOutput(0, xla::DynamicSlice(ctx->Input(0), ctx->Input(1), size)); } } }; diff --git a/tensorflow/compiler/tf2xla/kernels/softmax_op.cc b/tensorflow/compiler/tf2xla/kernels/softmax_op.cc index bbf5ee8b12186a582666121b1df5d8b7d881863e..a71fbcd901e8919949db5873675a7e3e785bdf4e 100644 --- a/tensorflow/compiler/tf2xla/kernels/softmax_op.cc +++ b/tensorflow/compiler/tf2xla/kernels/softmax_op.cc @@ -15,9 +15,12 @@ limitations under the License. // XLA-specific Ops for softmax. +#include "tensorflow/compiler/tf2xla/type_util.h" #include "tensorflow/compiler/tf2xla/xla_helpers.h" #include "tensorflow/compiler/tf2xla/xla_op_kernel.h" #include "tensorflow/compiler/tf2xla/xla_op_registry.h" +#include "tensorflow/compiler/xla/client/lib/constants.h" +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/core/framework/op_kernel.h" #include "tensorflow/core/framework/tensor.h" #include "tensorflow/core/framework/tensor_shape.h" @@ -41,6 +44,7 @@ class SoftmaxOp : public XlaOpKernel { const int kClassDim = 1; const DataType type = input_type(0); + const xla::PrimitiveType xla_type = ctx->input_xla_type(0); auto logits = ctx->Input(0); xla::XlaBuilder* const b = ctx->builder(); @@ -48,24 +52,27 @@ class SoftmaxOp : public XlaOpKernel { // Find the max in each batch, resulting in a tensor of shape [batch] auto logits_max = - b->Reduce(logits, XlaHelpers::MinValue(b, type), max_func, {kClassDim}); + xla::Reduce(logits, xla::MinValue(b, xla_type), max_func, {kClassDim}); // Subtract the max in batch b from every element in batch b. Broadcasts // along the batch dimension. - auto shifted_logits = b->Sub(logits, logits_max, {kBatchDim}); - auto exp_shifted = b->Exp(shifted_logits); + auto shifted_logits = xla::Sub(logits, logits_max, {kBatchDim}); + auto exp_shifted = xla::Exp(shifted_logits); const DataType accumulation_type = XlaHelpers::SumAccumulationType(type); + xla::PrimitiveType xla_accumulation_type; + OP_REQUIRES_OK(ctx, DataTypeToPrimitiveType(accumulation_type, + &xla_accumulation_type)); auto converted = - XlaHelpers::ConvertElementType(b, exp_shifted, accumulation_type); + xla::ConvertElementType(exp_shifted, xla_accumulation_type); auto reduce = - b->Reduce(converted, XlaHelpers::Zero(b, accumulation_type), - *ctx->GetOrCreateAdd(accumulation_type), {kClassDim}); + xla::Reduce(converted, xla::Zero(b, xla_accumulation_type), + *ctx->GetOrCreateAdd(accumulation_type), {kClassDim}); auto sum = XlaHelpers::ConvertElementType(b, reduce, type); auto softmax = log_ // softmax = shifted_logits - log(sum(exp(shifted_logits))) - ? b->Sub(shifted_logits, b->Log(sum), {kBatchDim}) + ? xla::Sub(shifted_logits, xla::Log(sum), {kBatchDim}) // softmax = exp(shifted_logits) / sum(exp(shifted_logits)) - : b->Div(exp_shifted, sum, {kBatchDim}); + : xla::Div(exp_shifted, sum, {kBatchDim}); ctx->SetOutput(0, softmax); } @@ -77,8 +84,8 @@ REGISTER_XLA_OP(Name("Softmax"), SoftmaxOp); REGISTER_XLA_OP(Name("LogSoftmax"), SoftmaxOp); std::pair CrossEntropyWithLogits( - XlaOpKernelContext* ctx, DataType type, const xla::XlaOp& logits, - const xla::XlaOp& labels) { + XlaOpKernelContext* ctx, DataType type, xla::PrimitiveType xla_type, + xla::XlaOp logits, xla::XlaOp labels) { const xla::XlaComputation& max_func = *ctx->GetOrCreateMax(type); const int kBatchDim = 0; @@ -87,43 +94,44 @@ std::pair CrossEntropyWithLogits( xla::XlaBuilder* b = ctx->builder(); // Find the max in each batch, resulting in a tensor of shape [batch] auto logits_max = - b->Reduce(logits, XlaHelpers::MinValue(b, type), max_func, {kClassDim}); + xla::Reduce(logits, xla::MinValue(b, xla_type), max_func, {kClassDim}); // Subtract the max in batch b from every element in batch b. // Broadcasts along the batch dimension. - auto shifted_logits = b->Sub(logits, logits_max, {kBatchDim}); + auto shifted_logits = xla::Sub(logits, logits_max, {kBatchDim}); // exp(logits - max_logits) - auto exp_shifted_logits = b->Exp(shifted_logits); + auto exp_shifted_logits = xla::Exp(shifted_logits); // sum_{class} (exp(logits - max_logits)) const DataType accumulation_type = XlaHelpers::SumAccumulationType(type); auto converted = XlaHelpers::ConvertElementType(b, exp_shifted_logits, accumulation_type); - auto reduce = b->Reduce(converted, XlaHelpers::Zero(b, accumulation_type), - *ctx->GetOrCreateAdd(accumulation_type), {kClassDim}); + auto reduce = + xla::Reduce(converted, XlaHelpers::Zero(b, accumulation_type), + *ctx->GetOrCreateAdd(accumulation_type), {kClassDim}); auto sum_exp = XlaHelpers::ConvertElementType(b, reduce, type); // log(sum(exp(logits - max_logits))) - auto log_sum_exp = b->Log(sum_exp); + auto log_sum_exp = xla::Log(sum_exp); // sum(-labels * // ((logits - max_logits) - log(sum(exp(logits - max_logits))))) // along classes // (The subtraction broadcasts along the batch dimension.) - auto sub = b->Sub(shifted_logits, log_sum_exp, {kBatchDim}); - auto mul = b->Mul(b->Neg(labels), sub); + auto sub = xla::Sub(shifted_logits, log_sum_exp, {kBatchDim}); + auto mul = xla::Mul(xla::Neg(labels), sub); auto sum = - b->Reduce(XlaHelpers::ConvertElementType(b, mul, accumulation_type), - XlaHelpers::Zero(b, accumulation_type), - *ctx->GetOrCreateAdd(accumulation_type), {kClassDim}); + xla::Reduce(XlaHelpers::ConvertElementType(b, mul, accumulation_type), + XlaHelpers::Zero(b, accumulation_type), + *ctx->GetOrCreateAdd(accumulation_type), {kClassDim}); auto loss = XlaHelpers::ConvertElementType(b, sum, type); // backprop: prob - labels, where // prob = exp(logits - max_logits) / sum(exp(logits - max_logits)) // (where the division broadcasts along the batch dimension) xla::XlaOp backprop = - b->Sub(b->Div(exp_shifted_logits, sum_exp, {kBatchDim}), labels); + xla::Sub(xla::Div(exp_shifted_logits, sum_exp, {kBatchDim}), labels); return {loss, backprop}; } @@ -146,12 +154,13 @@ class SoftmaxXentWithLogitsOp : public XlaOpKernel { // check that "labels" is a matrix too. const DataType type = input_type(0); + const xla::PrimitiveType xla_type = ctx->input_xla_type(0); auto logits = ctx->Input(0); auto labels = ctx->Input(1); xla::XlaOp loss, backprop; std::tie(loss, backprop) = - CrossEntropyWithLogits(ctx, type, logits, labels); + CrossEntropyWithLogits(ctx, type, xla_type, logits, labels); ctx->SetOutput(0, loss); ctx->SetOutput(1, backprop); } @@ -187,8 +196,9 @@ class SparseSoftmaxXentWithLogitsOp : public XlaOpKernel { int64 batch_size = logits_shape.dim_size(0); int64 depth = logits_shape.dim_size(1); - DataType logits_type = input_type(0); - DataType indices_type = input_type(1); + const DataType logits_type = input_type(0); + const xla::PrimitiveType xla_logits_type = ctx->input_xla_type(0); + const DataType indices_type = input_type(1); xla::XlaOp indices = ctx->Input(1); @@ -206,20 +216,18 @@ class SparseSoftmaxXentWithLogitsOp : public XlaOpKernel { // Builds a vector of {batch_size} that is 0 if the index is in range, or // NaN otherwise; then add that vector to the labels to force out-of-range // values to NaNs. - xla::XlaOp nan_or_zero = builder->Select( - builder->And( - builder->Le(XlaHelpers::Zero(builder, indices_type), indices), - builder->Lt(indices, XlaHelpers::IntegerLiteral( - builder, indices_type, depth))), - builder->Broadcast(XlaHelpers::Zero(builder, logits_type), - {batch_size}), - builder->Broadcast(XlaHelpers::FloatLiteral(builder, logits_type, NAN), - {batch_size})); - labels = builder->Add(labels, nan_or_zero, {0}); + xla::XlaOp nan_or_zero = xla::Select( + xla::And(xla::Le(XlaHelpers::Zero(builder, indices_type), indices), + xla::Lt(indices, XlaHelpers::IntegerLiteral( + builder, indices_type, depth))), + xla::Broadcast(XlaHelpers::Zero(builder, logits_type), {batch_size}), + xla::Broadcast(XlaHelpers::FloatLiteral(builder, logits_type, NAN), + {batch_size})); + labels = xla::Add(labels, nan_or_zero, {0}); xla::XlaOp loss, backprop; - std::tie(loss, backprop) = - CrossEntropyWithLogits(ctx, logits_type, ctx->Input(0), labels); + std::tie(loss, backprop) = CrossEntropyWithLogits( + ctx, logits_type, xla_logits_type, ctx->Input(0), labels); ctx->SetOutput(0, loss); ctx->SetOutput(1, backprop); } diff --git a/tensorflow/compiler/tf2xla/kernels/sort_ops.cc b/tensorflow/compiler/tf2xla/kernels/sort_ops.cc index 204ae8458214a0d0f049cff32ea99540b6f7fbd6..faaf8964ff7c40d75a493b03e6b400632117cb45 100644 --- a/tensorflow/compiler/tf2xla/kernels/sort_ops.cc +++ b/tensorflow/compiler/tf2xla/kernels/sort_ops.cc @@ -25,8 +25,7 @@ class XlaSortOp : public XlaOpKernel { explicit XlaSortOp(OpKernelConstruction* context) : XlaOpKernel(context) {} void Compile(XlaOpKernelContext* context) override { - xla::XlaBuilder* const b = context->builder(); - context->SetOutput(0, b->Sort(context->Input(0))); + context->SetOutput(0, xla::Sort(context->Input(0))); } }; diff --git a/tensorflow/compiler/tf2xla/kernels/spacetobatch_op.cc b/tensorflow/compiler/tf2xla/kernels/spacetobatch_op.cc index ec077924b5b5af4a573c86c8d9aeb8623bd7f801..8a8525efa186ed4aa02c494f7505f6245677e96e 100644 --- a/tensorflow/compiler/tf2xla/kernels/spacetobatch_op.cc +++ b/tensorflow/compiler/tf2xla/kernels/spacetobatch_op.cc @@ -16,6 +16,7 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/xla_helpers.h" #include "tensorflow/compiler/tf2xla/xla_op_kernel.h" #include "tensorflow/compiler/tf2xla/xla_op_registry.h" +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" namespace tensorflow { namespace { @@ -73,7 +74,7 @@ void SpaceToBatch(XlaOpKernelContext* ctx, const xla::XlaOp& input, "The product of the block dimensions must be positive")); xla::XlaOp padded = - b->Pad(input, XlaHelpers::Zero(b, input_dtype), padding_config); + xla::Pad(input, XlaHelpers::Zero(b, input_dtype), padding_config); // 2. Reshape `padded` to `reshaped_padded` of shape: // @@ -100,7 +101,7 @@ void SpaceToBatch(XlaOpKernelContext* ctx, const xla::XlaOp& input, std::copy(remainder_shape.begin(), remainder_shape.end(), reshaped_padded_shape.begin() + 1 + 2 * block_rank); - xla::XlaOp reshaped_padded = b->Reshape(padded, reshaped_padded_shape); + xla::XlaOp reshaped_padded = xla::Reshape(padded, reshaped_padded_shape); // 3. Permute dimensions of `reshaped_padded` to produce // `permuted_reshaped_padded` of shape: @@ -120,7 +121,7 @@ void SpaceToBatch(XlaOpKernelContext* ctx, const xla::XlaOp& input, std::iota(permutation.begin() + 1 + block_rank * 2, permutation.end(), 1 + block_rank * 2); xla::XlaOp permuted_reshaped_padded = - b->Transpose(reshaped_padded, permutation); + xla::Transpose(reshaped_padded, permutation); // 4. Reshape `permuted_reshaped_padded` to flatten `block_shape` into the // batch dimension, producing an output tensor of shape: @@ -140,7 +141,7 @@ void SpaceToBatch(XlaOpKernelContext* ctx, const xla::XlaOp& input, std::copy(remainder_shape.begin(), remainder_shape.end(), output_shape.begin() + 1 + block_rank); - xla::XlaOp output = b->Reshape(permuted_reshaped_padded, output_shape); + xla::XlaOp output = xla::Reshape(permuted_reshaped_padded, output_shape); ctx->SetOutput(0, output); } diff --git a/tensorflow/compiler/tf2xla/kernels/spacetodepth_op.cc b/tensorflow/compiler/tf2xla/kernels/spacetodepth_op.cc index 4c5886ee2a0f63d609f79fc690f457d93e284e3e..47d282fe9ec664bbc424793e93f778ebb13c6877 100644 --- a/tensorflow/compiler/tf2xla/kernels/spacetodepth_op.cc +++ b/tensorflow/compiler/tf2xla/kernels/spacetodepth_op.cc @@ -16,6 +16,7 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/xla_helpers.h" #include "tensorflow/compiler/tf2xla/xla_op_kernel.h" #include "tensorflow/compiler/tf2xla/xla_op_registry.h" +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/core/util/tensor_format.h" namespace tensorflow { @@ -50,7 +51,6 @@ class SpaceToDepthOp : public XlaOpKernel { const gtl::InlinedVector input_shape = input_tensor_shape.dim_sizes(); - xla::XlaBuilder* b = ctx->builder(); xla::XlaOp input = ctx->Input(0); int feature_dim = GetTensorFeatureDimIndex(input_rank, data_format_); @@ -135,7 +135,7 @@ class SpaceToDepthOp : public XlaOpKernel { // input_shape[1] / block_size_, block_size_, // input_shape[2] / block_size_, block_size_, // depth] - xla::XlaOp reshaped = b->Reshape(input, reshaped_shape); + xla::XlaOp reshaped = xla::Reshape(input, reshaped_shape); // 2. Permute dimensions of `reshaped` to produce // `permuted_reshaped` of shape: @@ -145,7 +145,7 @@ class SpaceToDepthOp : public XlaOpKernel { // input_shape[2] / block_size_, // block_size_, block_size_, // depth] - xla::XlaOp permuted_reshaped = b->Transpose(reshaped, transpose_order); + xla::XlaOp permuted_reshaped = xla::Transpose(reshaped, transpose_order); // 3. Reshape `permuted_reshaped` to flatten `block_shape` into the // batch dimension, producing an output tensor of shape: @@ -155,7 +155,7 @@ class SpaceToDepthOp : public XlaOpKernel { // input_shape[2] / block_size_, // block_size_ * block_size_ * depth] // - xla::XlaOp output = b->Reshape(permuted_reshaped, output_shape); + xla::XlaOp output = xla::Reshape(permuted_reshaped, output_shape); ctx->SetOutput(0, output); } diff --git a/tensorflow/compiler/tf2xla/kernels/sparse_to_dense_op.cc b/tensorflow/compiler/tf2xla/kernels/sparse_to_dense_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..e831dc30a9d3c27ec3b1494e7d8a6de836ff2a11 --- /dev/null +++ b/tensorflow/compiler/tf2xla/kernels/sparse_to_dense_op.cc @@ -0,0 +1,88 @@ +/* Copyright 2015 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow/compiler/tf2xla/lib/scatter.h" +#include "tensorflow/compiler/tf2xla/xla_op_kernel.h" +#include "tensorflow/compiler/tf2xla/xla_op_registry.h" + +namespace tensorflow { +namespace { + +// Operator to convert sparse representations to dense. +class SparseToDenseOp : public XlaOpKernel { + public: + explicit SparseToDenseOp(OpKernelConstruction* context) + : XlaOpKernel(context) {} + + void Compile(XlaOpKernelContext* context) override { + // sparse_indices + const TensorShape indices_shape = context->InputShape(0); + OP_REQUIRES(context, indices_shape.dims() <= 2, + errors::InvalidArgument( + "sparse_indices should be a scalar, vector, or matrix, " + "got shape ", + indices_shape.DebugString())); + const int64 num_elems = + indices_shape.dims() > 0 ? indices_shape.dim_size(0) : 1; + const int64 num_dims = + indices_shape.dims() > 1 ? indices_shape.dim_size(1) : 1; + + // output_shape + TensorShape output_shape; + OP_REQUIRES_OK(context, context->ConstantInputAsShape(1, &output_shape)); + OP_REQUIRES(context, output_shape.dims() == num_dims, + errors::InvalidArgument( + "output_shape has incorrect number of elements: ", + output_shape.num_elements(), " should be: ", num_dims)); + + // sparse_values + const TensorShape sparse_values_shape = context->InputShape(2); + const int64 num_values = sparse_values_shape.num_elements(); + OP_REQUIRES( + context, + sparse_values_shape.dims() == 0 || + (sparse_values_shape.dims() == 1 && num_values == num_elems), + errors::InvalidArgument("sparse_values has incorrect shape ", + sparse_values_shape.DebugString(), + ", should be [] or [", num_elems, "]")); + + // default_value + const TensorShape default_value_shape = context->InputShape(3); + OP_REQUIRES(context, TensorShapeUtils::IsScalar(default_value_shape), + errors::InvalidArgument("default_value should be a scalar.")); + + xla::XlaOp indices = context->Input(0); + xla::XlaOp sparse_values = context->Input(2); + xla::XlaOp default_value = context->Input(3); + + if (sparse_values_shape.dims() == 0 && num_elems != 1) { + sparse_values = Broadcast(sparse_values, {num_elems}); + } + xla::XlaBuilder* builder = context->builder(); + auto buffer = Broadcast(default_value, output_shape.dim_sizes()); + + auto result = XlaScatter(buffer, sparse_values, indices, + /*indices_are_vectors=*/num_dims > 1, + /*combiner=*/{}, builder); + context->SetOutput(0, builder->ReportErrorOrReturn(result)); + } +}; + +REGISTER_XLA_OP(Name("SparseToDense").CompileTimeConstInput("output_shape"), + SparseToDenseOp); + +} // namespace + +} // namespace tensorflow diff --git a/tensorflow/compiler/tf2xla/kernels/split_op.cc b/tensorflow/compiler/tf2xla/kernels/split_op.cc index 9b540585416ded663467d32f25ceceeaa52f069a..242638f981198ffd7a9c5b5f6365168de59a1f85 100644 --- a/tensorflow/compiler/tf2xla/kernels/split_op.cc +++ b/tensorflow/compiler/tf2xla/kernels/split_op.cc @@ -19,7 +19,8 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/xla_helpers.h" #include "tensorflow/compiler/tf2xla/xla_op_kernel.h" #include "tensorflow/compiler/tf2xla/xla_op_registry.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/core/framework/op_kernel.h" #include "tensorflow/core/framework/register_types.h" #include "tensorflow/core/framework/tensor.h" @@ -98,7 +99,7 @@ class SplitOp : public XlaOpKernel { // Slice out the ith split from the split dimension. begin[split_dim] = i * slice_size; limits[split_dim] = (i + 1) * slice_size; - ctx->SetOutput(i, ctx->builder()->Slice(input, begin, limits, strides)); + ctx->SetOutput(i, xla::Slice(input, begin, limits, strides)); } } }; @@ -199,7 +200,7 @@ class SplitVOp : public XlaOpKernel { // Slice out the ith split from the split dimension. limits[split_dim] = begin[split_dim] + slice_size; - ctx->SetOutput(i, ctx->builder()->Slice(input, begin, limits, strides)); + ctx->SetOutput(i, xla::Slice(input, begin, limits, strides)); begin[split_dim] = limits[split_dim]; } } diff --git a/tensorflow/compiler/tf2xla/kernels/stack_ops.cc b/tensorflow/compiler/tf2xla/kernels/stack_ops.cc index 0fb05a2be7b1034d6c2e864643b69647d622ede7..df91900570107609c0f1c2281faaab8a5e65b98b 100644 --- a/tensorflow/compiler/tf2xla/kernels/stack_ops.cc +++ b/tensorflow/compiler/tf2xla/kernels/stack_ops.cc @@ -23,7 +23,7 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/xla_helpers.h" #include "tensorflow/compiler/tf2xla/xla_op_kernel.h" #include "tensorflow/compiler/tf2xla/xla_op_registry.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/core/framework/op_kernel.h" #include "tensorflow/core/framework/partial_tensor_shape.h" #include "tensorflow/core/framework/register_types.h" @@ -144,24 +144,25 @@ class StackPushOp : public XlaOpKernel { // Initializes the Stack, if the element shape was not already known. OP_REQUIRES_OK(ctx, MaybeInitializeStack(b, resource, dtype_, elem_shape)); - xla::XlaOp ta = b->GetTupleElement(resource->value(), 0); - xla::XlaOp index = b->GetTupleElement(resource->value(), 1); + xla::XlaOp ta = xla::GetTupleElement(resource->value(), 0); + xla::XlaOp index = xla::GetTupleElement(resource->value(), 1); xla::XlaOp value = ctx->Input(1); // start_indices of the DynamicUpdateSlice are [index, 0, 0, ..., 0]. auto start_indices = - b->Pad(b->Reshape(index, {1}), b->ConstantR0(0), - xla::MakeEdgePaddingConfig({{0, elem_shape.dims()}})); + xla::Pad(xla::Reshape(index, {1}), xla::ConstantR0(b, 0), + xla::MakeEdgePaddingConfig({{0, elem_shape.dims()}})); TensorShape slice_shape = elem_shape; slice_shape.InsertDim(0, 1LL); - auto update = b->Reshape(value, slice_shape.dim_sizes()); + auto update = xla::Reshape(value, slice_shape.dim_sizes()); // TODO(phawkins): We don't check the index is in bounds --- there is no // error mechanism in XLA. - OP_REQUIRES_OK(ctx, resource->SetValue(b->Tuple( - {b->DynamicUpdateSlice(ta, update, start_indices), - b->Add(index, b->ConstantR0(1))}))); + OP_REQUIRES_OK(ctx, + resource->SetValue(xla::Tuple( + b, {xla::DynamicUpdateSlice(ta, update, start_indices), + xla::Add(index, xla::ConstantR0(b, 1))}))); ctx->SetOutput(0, value); } @@ -197,27 +198,27 @@ class StackPopOp : public XlaOpKernel { OP_REQUIRES_OK(ctx, GetStackShape(b, resource, &stack_shape)); xla::XlaOp state = resource->value(); - xla::XlaOp ta = b->GetTupleElement(state, 0); - xla::XlaOp index = b->GetTupleElement(state, 1); + xla::XlaOp ta = xla::GetTupleElement(state, 0); + xla::XlaOp index = xla::GetTupleElement(state, 1); - index = b->Sub(index, b->ConstantR0(1)); - OP_REQUIRES_OK(ctx, resource->SetValue(b->Tuple({ta, index}))); + index = Sub(index, xla::ConstantR0(b, 1)); + OP_REQUIRES_OK(ctx, resource->SetValue(xla::Tuple(b, {ta, index}))); // start_indices of the DynamicSlice are [index, 0, 0, ..., 0]. auto start_indices = - b->Pad(b->Reshape(index, {1}), b->ConstantR0(0), - xla::MakeEdgePaddingConfig({{0, stack_shape.dims() - 1}})); + xla::Pad(xla::Reshape(index, {1}), xla::ConstantR0(b, 0), + xla::MakeEdgePaddingConfig({{0, stack_shape.dims() - 1}})); auto slice_shape = stack_shape.dim_sizes(); slice_shape[0] = 1LL; // TODO(phawkins): We don't check the index is in bounds --- there is no // error mechanism in XLA. - xla::XlaOp read = b->DynamicSlice(ta, start_indices, slice_shape); + xla::XlaOp read = xla::DynamicSlice(ta, start_indices, slice_shape); // Remove the leading '1' dimension. std::vector value_shape(slice_shape.begin() + 1, slice_shape.end()); - ctx->SetOutput(0, b->Reshape(read, value_shape)); + ctx->SetOutput(0, xla::Reshape(read, value_shape)); } private: diff --git a/tensorflow/compiler/tf2xla/kernels/stateless_random_ops.cc b/tensorflow/compiler/tf2xla/kernels/stateless_random_ops.cc index 03675014337877333cc4c497e2223c348547e3f7..a6f5769e7b7b1e550b7908caa35289cf3030120f 100644 --- a/tensorflow/compiler/tf2xla/kernels/stateless_random_ops.cc +++ b/tensorflow/compiler/tf2xla/kernels/stateless_random_ops.cc @@ -20,7 +20,10 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/xla_helpers.h" #include "tensorflow/compiler/tf2xla/xla_op_kernel.h" #include "tensorflow/compiler/tf2xla/xla_op_registry.h" -#include "tensorflow/compiler/xla/client/lib/arithmetic.h" +#include "tensorflow/compiler/xla/client/lib/constants.h" +#include "tensorflow/compiler/xla/client/lib/math.h" +#include "tensorflow/compiler/xla/client/lib/numeric.h" +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/core/framework/op_kernel.h" #include "tensorflow/core/framework/tensor.h" #include "tensorflow/core/framework/tensor_shape.h" @@ -33,9 +36,9 @@ namespace { // Rotates a 32-bit integer 'v' left by 'distance' bits. xla::XlaOp RotateLeftS32(xla::XlaBuilder* builder, const xla::XlaOp& v, int distance) { - return builder->Or( - builder->ShiftLeft(v, builder->ConstantR0(distance)), - builder->ShiftRightLogical(v, builder->ConstantR0(32 - distance))); + return xla::Or( + xla::ShiftLeft(v, xla::ConstantR0(builder, distance)), + xla::ShiftRightLogical(v, xla::ConstantR0(builder, 32 - distance))); } using ThreeFry2x32State = std::array; @@ -51,22 +54,22 @@ ThreeFry2x32State ThreeFry2x32(xla::XlaBuilder* builder, std::array ks; // 0x1BD11BDA is a parity constant specified by the ThreeFry2x32 algorithm. - ks[2] = builder->ConstantR0(0x1BD11BDA); + ks[2] = xla::ConstantR0(builder, 0x1BD11BDA); for (int i = 0; i < 2; ++i) { ks[i] = key[i]; x[i] = input[i]; - ks[2] = builder->Xor(ks[2], key[i]); + ks[2] = xla::Xor(ks[2], key[i]); } - x[0] = builder->Add(x[0], ks[0]); - x[1] = builder->Add(x[1], ks[1]); + x[0] = xla::Add(x[0], ks[0]); + x[1] = xla::Add(x[1], ks[1]); // Performs a single round of the Threefry2x32 algorithm, with a rotation // amount 'rotation'. auto round = [builder](ThreeFry2x32State v, int rotation) { - v[0] = builder->Add(v[0], v[1]); + v[0] = xla::Add(v[0], v[1]); v[1] = RotateLeftS32(builder, v[1], rotation); - v[1] = builder->Xor(v[0], v[1]); + v[1] = xla::Xor(v[0], v[1]); return v; }; @@ -76,36 +79,36 @@ ThreeFry2x32State ThreeFry2x32(xla::XlaBuilder* builder, x = round(x, rotations[1]); x = round(x, rotations[2]); x = round(x, rotations[3]); - x[0] = builder->Add(x[0], ks[1]); - x[1] = builder->Add(builder->Add(x[1], ks[2]), builder->ConstantR0(1)); + x[0] = xla::Add(x[0], ks[1]); + x[1] = xla::Add(xla::Add(x[1], ks[2]), xla::ConstantR0(builder, 1)); x = round(x, rotations[4]); x = round(x, rotations[5]); x = round(x, rotations[6]); x = round(x, rotations[7]); - x[0] = builder->Add(x[0], ks[2]); - x[1] = builder->Add(builder->Add(x[1], ks[0]), builder->ConstantR0(2)); + x[0] = xla::Add(x[0], ks[2]); + x[1] = xla::Add(xla::Add(x[1], ks[0]), xla::ConstantR0(builder, 2)); x = round(x, rotations[0]); x = round(x, rotations[1]); x = round(x, rotations[2]); x = round(x, rotations[3]); - x[0] = builder->Add(x[0], ks[0]); - x[1] = builder->Add(builder->Add(x[1], ks[1]), builder->ConstantR0(3)); + x[0] = xla::Add(x[0], ks[0]); + x[1] = xla::Add(xla::Add(x[1], ks[1]), xla::ConstantR0(builder, 3)); x = round(x, rotations[4]); x = round(x, rotations[5]); x = round(x, rotations[6]); x = round(x, rotations[7]); - x[0] = builder->Add(x[0], ks[1]); - x[1] = builder->Add(builder->Add(x[1], ks[2]), builder->ConstantR0(4)); + x[0] = xla::Add(x[0], ks[1]); + x[1] = xla::Add(xla::Add(x[1], ks[2]), xla::ConstantR0(builder, 4)); x = round(x, rotations[0]); x = round(x, rotations[1]); x = round(x, rotations[2]); x = round(x, rotations[3]); - x[0] = builder->Add(x[0], ks[2]); - x[1] = builder->Add(builder->Add(x[1], ks[0]), builder->ConstantR0(5)); + x[0] = xla::Add(x[0], ks[2]); + x[1] = xla::Add(xla::Add(x[1], ks[0]), xla::ConstantR0(builder, 5)); return x; } @@ -116,8 +119,8 @@ xla::XlaOp RandomUniform(xla::XlaBuilder* builder, const xla::XlaOp& seed, const TensorShape& shape, double minval, double maxval) { // Split the seed into two 32-bit scalars to form a key. - auto seed0 = builder->Reshape(builder->Slice(seed, {0}, {1}, {1}), {}); - auto seed1 = builder->Reshape(builder->Slice(seed, {1}, {2}, {1}), {}); + auto seed0 = xla::Reshape(xla::Slice(seed, {0}, {1}, {1}), {}); + auto seed1 = xla::Reshape(xla::Slice(seed, {1}, {2}, {1}), {}); ThreeFry2x32State key = {seed0, seed1}; const int64 size = shape.num_elements(); @@ -126,33 +129,33 @@ xla::XlaOp RandomUniform(xla::XlaBuilder* builder, const xla::XlaOp& seed, // Fill the generator inputs with unique counter values. ThreeFry2x32State inputs; - TF_CHECK_OK(XlaHelpers::Iota(builder, DT_INT32, half_size, &inputs[0])); - inputs[1] = builder->Add(inputs[0], builder->ConstantR0(half_size)); + inputs[0] = xla::Iota(builder, xla::S32, half_size); + inputs[1] = xla::Add(inputs[0], xla::ConstantR0(builder, half_size)); ThreeFry2x32State outputs = ThreeFry2x32(builder, inputs, key); if (size_is_odd) { - outputs[1] = builder->Slice(outputs[1], {0}, {half_size - 1}, {1}); + outputs[1] = xla::Slice(outputs[1], {0}, {half_size - 1}, {1}); } auto bits = - builder->Reshape(builder->ConcatInDim(outputs, 0), shape.dim_sizes()); + xla::Reshape(xla::ConcatInDim(builder, outputs, 0), shape.dim_sizes()); // Form 22 random mantissa bits, with a leading 1 bit. The leading 1 bit // forces the random bits into the mantissa. constexpr int kFloatBits = 32; constexpr int kMantissaBits = 23; - bits = builder->Or( - builder->ShiftRightLogical( - bits, builder->ConstantR0(kFloatBits - kMantissaBits)), - builder->ConstantR0(bit_cast(1.0f))); - auto floats = builder->BitcastConvertType(bits, xla::F32); + bits = xla::Or( + xla::ShiftRightLogical( + bits, xla::ConstantR0(builder, kFloatBits - kMantissaBits)), + xla::ConstantR0(builder, bit_cast(1.0f))); + auto floats = xla::BitcastConvertType(bits, xla::F32); // We have a floating point number in the range [1.0, 2.0). // Subtract 1.0f to shift to the range [0.0, 1.0) - floats = builder->Sub(floats, builder->ConstantR0(1.0f)); + floats = xla::Sub(floats, xla::ConstantR0(builder, 1.0f)); // Multiply and add to shift to the range [minval, maxval). - floats = builder->Mul(floats, builder->ConstantR0(maxval - minval)); - floats = builder->Add(floats, builder->ConstantR0(minval)); + floats = xla::Mul(floats, xla::ConstantR0(builder, maxval - minval)); + floats = xla::Add(floats, xla::ConstantR0(builder, minval)); return floats; } @@ -207,10 +210,8 @@ class StatelessRandomNormalOp : public XlaOpKernel { RandomUniform(builder, seed, shape, std::nextafter(-1.0f, 0.0f), 1.0); // Convert uniform distribution to normal distribution by computing // sqrt(2) * erfinv(x) - auto erfinv_or_status = ErfInv(uniform); - OP_REQUIRES_OK(ctx, erfinv_or_status.status()); - auto normal = builder->Mul(builder->ConstantR0(std::sqrt(2.0)), - erfinv_or_status.ValueOrDie()); + auto normal = + xla::ScalarLike(uniform, std::sqrt(2.0)) * xla::ErfInv(uniform); ctx->SetOutput(0, normal); } @@ -231,8 +232,6 @@ class StatelessTruncatedNormalOp : public XlaOpKernel { : XlaOpKernel(ctx) {} void Compile(XlaOpKernelContext* ctx) override { - const DataType dtype = output_type(0); - TensorShape shape; OP_REQUIRES_OK(ctx, ctx->ConstantInputAsShape(0, &shape)); @@ -245,9 +244,7 @@ class StatelessTruncatedNormalOp : public XlaOpKernel { auto uniform = RandomUniform(b, seed, shape, std::numeric_limits::min(), 1.0); - auto truncated_normal_or_status = TruncatedNormal(dtype, uniform, b); - OP_REQUIRES_OK(ctx, truncated_normal_or_status.status()); - ctx->SetOutput(0, truncated_normal_or_status.ValueOrDie()); + ctx->SetOutput(0, TruncatedNormal(uniform)); } private: diff --git a/tensorflow/compiler/tf2xla/kernels/strided_slice_op.cc b/tensorflow/compiler/tf2xla/kernels/strided_slice_op.cc index 55254c746e5ebaf6b468c24ab59b968bf0d6260b..c2165ccd86dfa1c119790beb20af0844fb1bbda8 100644 --- a/tensorflow/compiler/tf2xla/kernels/strided_slice_op.cc +++ b/tensorflow/compiler/tf2xla/kernels/strided_slice_op.cc @@ -19,6 +19,7 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/xla_helpers.h" #include "tensorflow/compiler/tf2xla/xla_op_kernel.h" #include "tensorflow/compiler/tf2xla/xla_op_registry.h" +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/core/framework/op_kernel.h" #include "tensorflow/core/framework/register_types.h" #include "tensorflow/core/framework/tensor.h" @@ -92,12 +93,12 @@ class StridedSliceOp : public XlaOpKernel { xla::XlaOp slice = ctx->Input(0); if (!dimensions_to_reverse.empty()) { - slice = ctx->builder()->Rev(slice, dimensions_to_reverse); + slice = xla::Rev(slice, dimensions_to_reverse); } - slice = ctx->builder()->Slice(slice, slice_begin, slice_end, slice_strides); + slice = xla::Slice(slice, slice_begin, slice_end, slice_strides); - slice = ctx->builder()->Reshape(slice, final_shape.dim_sizes()); + slice = xla::Reshape(slice, final_shape.dim_sizes()); ctx->SetOutput(0, slice); } @@ -171,7 +172,7 @@ class StridedSliceGradOp : public XlaOpKernel { xla::XlaOp grad = ctx->Input(4); // Undo any new/shrink axes. - grad = ctx->builder()->Reshape(grad, processing_shape.dim_sizes()); + grad = xla::Reshape(grad, processing_shape.dim_sizes()); // Pad the input gradients. gtl::InlinedVector dimensions_to_reverse; @@ -204,9 +205,9 @@ class StridedSliceGradOp : public XlaOpKernel { } } if (!dimensions_to_reverse.empty()) { - grad = ctx->builder()->Rev(grad, dimensions_to_reverse); + grad = xla::Rev(grad, dimensions_to_reverse); } - grad = ctx->builder()->Pad(grad, zero, padding_config); + grad = xla::Pad(grad, zero, padding_config); ctx->SetOutput(0, grad); } @@ -306,17 +307,17 @@ class StridedSliceAssignOp : public XlaOpKernel { } if (!dimensions_to_reverse.empty()) { - rhs = ctx->builder()->Rev(rhs, dimensions_to_reverse); + rhs = xla::Rev(rhs, dimensions_to_reverse); } - rhs = ctx->builder()->Reshape(rhs, slice_dims); + rhs = xla::Reshape(rhs, slice_dims); if (lhs_shape.dims() == 0) { // TODO(b/38323843): DynamicUpdateSlice crashes on rank 0 inputs. Fix // and remove this workaround. lhs = rhs; } else { - lhs = ctx->builder()->DynamicUpdateSlice( - lhs, rhs, ctx->builder()->ConstantR1(slice_begin)); + lhs = xla::DynamicUpdateSlice( + lhs, rhs, xla::ConstantR1(ctx->builder(), slice_begin)); } OP_REQUIRES_OK(ctx, ctx->AssignVariable(0, dtype_, lhs)); diff --git a/tensorflow/compiler/tf2xla/kernels/tensor_array_ops.cc b/tensorflow/compiler/tf2xla/kernels/tensor_array_ops.cc index 9adee78a1fd1fb9a12afae83197425c328b5fe7e..26326f18b844fa9dc48aeedfa5dcff3d09033a18 100644 --- a/tensorflow/compiler/tf2xla/kernels/tensor_array_ops.cc +++ b/tensorflow/compiler/tf2xla/kernels/tensor_array_ops.cc @@ -25,7 +25,8 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/xla_op_kernel.h" #include "tensorflow/compiler/tf2xla/xla_op_registry.h" #include "tensorflow/compiler/tf2xla/xla_resource.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/core/framework/op_kernel.h" #include "tensorflow/core/framework/partial_tensor_shape.h" #include "tensorflow/core/framework/register_types.h" @@ -123,10 +124,9 @@ xla::XlaOp DynamicAddSlice(xla::XlaBuilder* builder, const xla::XlaOp& operand, const xla::XlaOp& update, const gtl::ArraySlice& update_dims, const xla::XlaOp& start_indices) { - xla::XlaOp current = - builder->DynamicSlice(operand, start_indices, update_dims); - xla::XlaOp sum = builder->Add(current, update); - return builder->DynamicUpdateSlice(operand, sum, start_indices); + xla::XlaOp current = xla::DynamicSlice(operand, start_indices, update_dims); + xla::XlaOp sum = xla::Add(current, update); + return xla::DynamicUpdateSlice(operand, sum, start_indices); } class TensorArrayOp : public XlaOpKernel { @@ -162,7 +162,7 @@ class TensorArrayOp : public XlaOpKernel { ta_shape.AddDim(size); ta_shape.AppendShape(shape); xla::XlaOp zero = XlaHelpers::Zero(b, dtype_); - value = b->Broadcast(zero, ta_shape.dim_sizes()); + value = xla::Broadcast(zero, ta_shape.dim_sizes()); } XlaContext& xc = XlaContext::Get(ctx); @@ -215,12 +215,12 @@ class TensorArrayWriteOp : public XlaOpKernel { // start_indices of the DynamicUpdateSlice are [index, 0, 0, ..., 0]. auto start_indices = - b->Pad(b->Reshape(index, {1}), b->ConstantR0(0), - xla::MakeEdgePaddingConfig({{0, elem_shape.dims()}})); + xla::Pad(xla::Reshape(index, {1}), xla::ConstantR0(b, 0), + xla::MakeEdgePaddingConfig({{0, elem_shape.dims()}})); TensorShape slice_shape = elem_shape; slice_shape.InsertDim(0, 1LL); - auto update = b->Reshape(value, slice_shape.dim_sizes()); + auto update = xla::Reshape(value, slice_shape.dim_sizes()); xla::XlaOp written = DynamicAddSlice(b, ta, update, slice_shape.dim_sizes(), start_indices); @@ -259,17 +259,17 @@ class TensorArrayReadOp : public XlaOpKernel { // start_indices of the DynamicSlice are [index, 0, 0, ..., 0]. auto start_indices = - b->Pad(b->Reshape(index, {1}), b->ConstantR0(0), - xla::MakeEdgePaddingConfig({{0, ta_shape.dims() - 1}})); + xla::Pad(xla::Reshape(index, {1}), xla::ConstantR0(b, 0), + xla::MakeEdgePaddingConfig({{0, ta_shape.dims() - 1}})); auto slice_shape = ta_shape.dim_sizes(); slice_shape[0] = 1LL; - xla::XlaOp read = b->DynamicSlice(ta, start_indices, slice_shape); + xla::XlaOp read = xla::DynamicSlice(ta, start_indices, slice_shape); // Remove the leading '1' dimension. std::vector value_shape(slice_shape.begin() + 1, slice_shape.end()); - ctx->SetOutput(0, b->Reshape(read, value_shape)); + ctx->SetOutput(0, xla::Reshape(read, value_shape)); } private: @@ -326,7 +326,7 @@ class TensorArrayGatherOp : public XlaOpKernel { for (auto i = 1; i < ta_shape.dims(); i++) { end[i] = ta_shape.dim_size(i); } - ctx->SetOutput(0, b->Slice(ta, begin, end, strides)); + ctx->SetOutput(0, xla::Slice(ta, begin, end, strides)); return; } } @@ -391,7 +391,7 @@ class TensorArrayScatterOp : public XlaOpKernel { } if (scatter_all_elements_in_order) { - ta = b->Add(ta, value); + ta = xla::Add(ta, value); } else { auto slice_dims = value_shape.dim_sizes(); slice_dims[0] = 1LL; @@ -407,13 +407,13 @@ class TensorArrayScatterOp : public XlaOpKernel { // Slice out part of the value. value_starts[0] = i; value_ends[0] = i + 1; - auto slice = b->Slice(value, value_starts, value_ends, value_strides); + auto slice = xla::Slice(value, value_starts, value_ends, value_strides); // start_indices of the DynamicUpdateSlice are [index, 0, 0, ..., 0]. - auto index = b->Slice(indices, {i}, {i + 1}, {1}); + auto index = xla::Slice(indices, {i}, {i + 1}, {1}); auto start_indices = - b->Pad(b->Reshape(index, {1}), b->ConstantR0(0), - xla::MakeEdgePaddingConfig({{0, elem_shape.dims()}})); + xla::Pad(xla::Reshape(index, {1}), xla::ConstantR0(b, 0), + xla::MakeEdgePaddingConfig({{0, elem_shape.dims()}})); ta = DynamicAddSlice(b, ta, slice, slice_dims, start_indices); } } @@ -452,7 +452,7 @@ class TensorArrayConcatOp : public XlaOpKernel { auto ta_dims = ta_shape.dim_sizes(); std::vector shape(ta_dims.begin() + 1, ta_dims.end()); shape[0] *= ta_shape.dim_size(0); - ctx->SetOutput(0, b->Reshape(ta, shape)); + ctx->SetOutput(0, xla::Reshape(ta, shape)); Tensor lengths(DT_INT64, {ta_dims[0]}); auto lengths_vec = lengths.vec(); @@ -522,8 +522,8 @@ class TensorArraySplitOp : public XlaOpKernel { value_shape.DebugString(), " vs. ", ta_shape.DebugString())); - OP_REQUIRES_OK(ctx, resource->SetValue(b->Add( - ta, b->Reshape(value, ta_shape.dim_sizes())))); + OP_REQUIRES_OK(ctx, resource->SetValue(xla::Add( + ta, xla::Reshape(value, ta_shape.dim_sizes())))); ctx->SetOutput(0, flow); } diff --git a/tensorflow/compiler/tf2xla/kernels/tile_ops.cc b/tensorflow/compiler/tf2xla/kernels/tile_ops.cc index e91075196bd8414939888e22b5483ad637487af6..c9e56942625a009fb3660f413a845547192460d5 100644 --- a/tensorflow/compiler/tf2xla/kernels/tile_ops.cc +++ b/tensorflow/compiler/tf2xla/kernels/tile_ops.cc @@ -20,6 +20,7 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/xla_helpers.h" #include "tensorflow/compiler/tf2xla/xla_op_kernel.h" #include "tensorflow/compiler/tf2xla/xla_op_registry.h" +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/core/framework/numeric_op.h" #include "tensorflow/core/framework/op_kernel.h" #include "tensorflow/core/framework/tensor.h" @@ -93,9 +94,9 @@ class TileOp : public XlaOpKernel { if (one_dimension_is_broadcasted_without_multiple) { // Create a constant Zero the size of the output shape to leverage binary // operation broadcast semantics. - auto broadcasted_zero = ctx->builder()->Broadcast( + auto broadcasted_zero = xla::Broadcast( XlaHelpers::Zero(ctx->builder(), ctx->input_type(0)), output_shape); - ctx->SetOutput(0, ctx->builder()->Add(broadcasted_zero, input)); + ctx->SetOutput(0, xla::Add(broadcasted_zero, input)); return; } @@ -103,7 +104,7 @@ class TileOp : public XlaOpKernel { // dimension. This prepends the broadcasted dimensions, so an // input of shape [2,3,1] broadcast with multiples [5,4,3] will // end up with shape [5,4,3,2,3,1]. - auto broadcasted = ctx->builder()->Broadcast(input, multiples_array); + auto broadcasted = xla::Broadcast(input, multiples_array); // Now flatten and reshape. The broadcasted dimensions are // paired with the original dimensions so in the above example // we flatten [0,3,1,4,2,5] then reshape to [10,12,3]. @@ -112,8 +113,7 @@ class TileOp : public XlaOpKernel { flattened.push_back(i); flattened.push_back(i + output_shape.size()); } - xla::XlaOp output = - ctx->builder()->Reshape(broadcasted, flattened, output_shape); + xla::XlaOp output = xla::Reshape(broadcasted, flattened, output_shape); ctx->SetOutput(0, output); } diff --git a/tensorflow/compiler/tf2xla/kernels/topk_op.cc b/tensorflow/compiler/tf2xla/kernels/topk_op.cc index cbe3c8aaff02e1a4b19f295216772b2004ccaf70..1ddcb08c8e1864236a09cecbb4d79788221f5017 100644 --- a/tensorflow/compiler/tf2xla/kernels/topk_op.cc +++ b/tensorflow/compiler/tf2xla/kernels/topk_op.cc @@ -16,8 +16,9 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/xla_helpers.h" #include "tensorflow/compiler/tf2xla/xla_op_kernel.h" #include "tensorflow/compiler/tf2xla/xla_op_registry.h" +#include "tensorflow/compiler/xla/client/lib/numeric.h" #include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/core/framework/kernel_def_builder.h" #include "tensorflow/core/framework/types.h" #include "tensorflow/core/kernels/no_op.h" @@ -51,108 +52,33 @@ class TopKOp : public XlaOpKernel { errors::Unimplemented("TopK is implemented for 1-D inputs, got shape ", input_shape.DebugString())); - const int64 n = input_shape.dim_size(0); - OP_REQUIRES(context, n < (1 << 16), - errors::Unimplemented( - "TopK is implemented for sizes up to 2**16, got shape ", - input_shape.DebugString())); - xla::XlaBuilder* const b = context->builder(); if (input_shape.dim_size(0) < k) { k = input_shape.dim_size(0); } - const xla::XlaOp input_bf16 = context->Input(0); - xla::XlaOp iota_s32; - OP_REQUIRES_OK(context, XlaHelpers::Iota(b, DT_INT32, n, &iota_s32)); - - // TODO(b/73891930): add a key-value sort to HLO, rather than using - // bit-packing tricks here. - - xla::XlaOp zero = b->ConstantR0(0); - - // max can either be 0x7FFFFFFF or 0x8000000. Neither choice is totally - // ideal. The implications of the choice are: - // - // 0x7FFFFFFF - // 1. +0.0 > -0.0 - // 2. The elements of the inputs and outputs are bitwise identical. - // 3. The sort is unstable since a later +0.0 will appear before an earlier - // -0.0. - // - // 0x8000000 - // 1. +0.0 == -0.0 - // 2. All -0.0 in the input are replaced with +0.0 in the output. - // 3. The sort is stable. - xla::XlaOp max = b->ConstantR0(0x80000000); - xla::XlaOp index_mask = b->ConstantR0(0x0000FFFF); - xla::XlaOp value_mask = b->ConstantR0(0xFFFF0000); - - // Convert to from bf16 to f32. The lower 16-bits are zero due to the - // definition of bf16. - xla::XlaOp input_f32 = b->ConvertElementType(input_bf16, xla::F32); - - // Negate the input to reverse sort it. The lower 16-bits are zero, because - // negating a float is just inverting the high-bit. - xla::XlaOp negative_input_f32 = b->Neg(input_f32); - - // Convert to a sign magnitude integer. The lower 16-bits are zero, since - // bitcast convert doesn't change any bits. - xla::XlaOp negative_input_sm32 = - b->BitcastConvertType(negative_input_f32, xla::S32); - - // Convert from sign magnitude integer to two's complement integer. The - // lower 16-bits are zero on both sides of the select. On the false side, - // the value is unchanged, and on the true side, the lower 16-bits of max - // are all zero, so the lower 16-bits of the result of the subtraction will - // also be zero. - xla::XlaOp negative_input_s32 = - b->Select(b->Lt(negative_input_sm32, zero), - b->Sub(max, negative_input_sm32), negative_input_sm32); - - // In order for the Or with iota_s32 to to work properly, the lower 16-bits - // of negative_input_32 must be zero. - - // Pack elements as: - // * upper 16 bits are the value - // * lower 16 bits are the index. - xla::XlaOp packed_s32 = b->Or(negative_input_s32, iota_s32); - - // TODO(phawkins): use a more efficient algorithm that does not require a - // full sort. - xla::XlaOp sorted_s32 = b->Slice(b->Sort(packed_s32), - /*start_indices=*/{0}, - /*limit_indices=*/{k}, - /*strides=*/{1}); - - // Unpack the value/index. - xla::XlaOp indices_s32 = b->And(sorted_s32, index_mask); - xla::XlaOp negative_values_s32 = b->And(sorted_s32, value_mask); - - // Convert from two's complement integer to sign magnitude integer. - xla::XlaOp negative_values_sm32 = - b->Select(b->Lt(negative_values_s32, zero), - b->Sub(max, negative_values_s32), negative_values_s32); - - xla::XlaOp negative_values_f32 = - b->BitcastConvertType(negative_values_sm32, xla::F32); - - // Negate the values to get back the original inputs. - xla::XlaOp values_f32 = b->Neg(negative_values_f32); - - // Convert from f32 to bf16. - xla::XlaOp values_bf16 = b->ConvertElementType(values_f32, xla::BF16); - - context->SetOutput(0, values_bf16); - context->SetOutput(1, indices_s32); + const xla::XlaOp input = context->Input(0); + xla::XlaOp iota_s32 = xla::Iota(b, xla::S32, input_shape.dim_size(0)); + xla::XlaOp sort_result = xla::Sort(xla::Neg(input), iota_s32); + xla::XlaOp values = + xla::Neg(xla::Slice(xla::GetTupleElement(sort_result, 0), + /*start_indices=*/{0}, + /*limit_indices=*/{k}, + /*strides=*/{1})); + xla::XlaOp indices = xla::Slice(xla::GetTupleElement(sort_result, 1), + /*start_indices=*/{0}, + /*limit_indices=*/{k}, + /*strides=*/{1}); + context->SetOutput(0, values); + context->SetOutput(1, indices); } private: bool sorted_; }; -REGISTER_XLA_OP( - Name("TopKV2").CompileTimeConstInput("k").TypeConstraint("T", DT_BFLOAT16), - TopKOp); +REGISTER_XLA_OP(Name("TopKV2").CompileTimeConstInput("k").TypeConstraint( + "T", {DT_UINT32, DT_INT32, DT_FLOAT, DT_BFLOAT16}), + TopKOp); } // namespace } // namespace tensorflow diff --git a/tensorflow/compiler/tf2xla/kernels/training_ops.cc b/tensorflow/compiler/tf2xla/kernels/training_ops.cc index 34caefa050c0d58f5f7bad557286b6ed64b996ad..98df73024962b8009a74976d473df752d590b47a 100644 --- a/tensorflow/compiler/tf2xla/kernels/training_ops.cc +++ b/tensorflow/compiler/tf2xla/kernels/training_ops.cc @@ -16,8 +16,10 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/kernels/cwise_ops.h" #include "tensorflow/compiler/tf2xla/xla_helpers.h" #include "tensorflow/compiler/tf2xla/xla_op_registry.h" +#include "tensorflow/compiler/xla/client/lib/constants.h" +#include "tensorflow/compiler/xla/client/lib/math.h" #include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/core/framework/kernel_def_builder.h" #include "tensorflow/core/framework/types.h" #include "tensorflow/core/kernels/no_op.h" @@ -31,7 +33,6 @@ class ResourceApplyGradientDescent : public XlaOpKernel { : XlaOpKernel(ctx) {} void Compile(XlaOpKernelContext* ctx) override { xla::XlaOp handle; - xla::XlaBuilder* b = ctx->builder(); DataType type = ctx->input_type(1); TensorShape var_shape; OP_REQUIRES_OK(ctx, ctx->ReadVariableInput(0, type, &var_shape, &handle)); @@ -48,7 +49,7 @@ class ResourceApplyGradientDescent : public XlaOpKernel { var_shape.DebugString(), " vs ", delta_shape.DebugString())); - handle = b->Sub(handle, b->Mul(ctx->Input(1), ctx->Input(2))); + handle = handle - ctx->Input(1) * ctx->Input(2); OP_REQUIRES_OK(ctx, ctx->AssignVariable(0, type, handle)); } }; @@ -56,6 +57,64 @@ REGISTER_XLA_OP( Name("ResourceApplyGradientDescent").TypeConstraint("T", kFloatTypes), ResourceApplyGradientDescent); +xla::XlaOp ProximalGradientDescentUpdate(xla::XlaOp var, xla::XlaOp lr, + xla::XlaOp l1, xla::XlaOp l2, + xla::XlaOp grad) { + xla::XlaOp one = xla::ScalarLike(lr, 1.0); + xla::XlaOp zero = xla::ScalarLike(lr, 0.0); + xla::XlaOp prox_var = var - grad * lr; + xla::XlaOp l1_gt_zero = xla::Sign(prox_var) * + xla::Max(xla::Abs(prox_var) - lr * l1, zero) / + (one + lr * l2); + xla::XlaOp l1_le_zero = prox_var / (one + lr * l2); + return xla::Select(xla::Gt(l1, zero), l1_gt_zero, l1_le_zero); +} + +class ResourceApplyProximalGradientDescent : public XlaOpKernel { + public: + explicit ResourceApplyProximalGradientDescent(OpKernelConstruction* ctx) + : XlaOpKernel(ctx) { + OP_REQUIRES_OK(ctx, ctx->GetAttr("T", &dtype_)); + } + + void Compile(XlaOpKernelContext* ctx) override { + xla::XlaOp var; + TensorShape var_shape; + OP_REQUIRES_OK(ctx, ctx->ReadVariableInput(0, dtype_, &var_shape, &var)); + + TensorShape alpha_shape = ctx->InputShape(1); + OP_REQUIRES(ctx, TensorShapeUtils::IsScalar(alpha_shape), + errors::InvalidArgument("alpha is not a scalar: ", + alpha_shape.DebugString())); + TensorShape l1_shape = ctx->InputShape(2); + OP_REQUIRES(ctx, TensorShapeUtils::IsScalar(alpha_shape), + errors::InvalidArgument("l1 is not a scalar: ", + l1_shape.DebugString())); + TensorShape l2_shape = ctx->InputShape(3); + OP_REQUIRES(ctx, TensorShapeUtils::IsScalar(alpha_shape), + errors::InvalidArgument("l2 is not a scalar: ", + l2_shape.DebugString())); + TensorShape delta_shape = ctx->InputShape(4); + OP_REQUIRES( + ctx, var_shape.IsSameSize(delta_shape), + errors::InvalidArgument("var and delta do not have the same shape: ", + var_shape.DebugString(), " vs ", + delta_shape.DebugString())); + xla::XlaOp alpha = ctx->Input(1); + xla::XlaOp l1 = ctx->Input(2); + xla::XlaOp l2 = ctx->Input(3); + xla::XlaOp delta = ctx->Input(4); + var = ProximalGradientDescentUpdate(var, alpha, l1, l2, delta); + OP_REQUIRES_OK(ctx, ctx->AssignVariable(0, dtype_, var)); + } + + private: + DataType dtype_; +}; +REGISTER_XLA_OP(Name("ResourceApplyProximalGradientDescent") + .TypeConstraint("T", kFloatTypes), + ResourceApplyProximalGradientDescent); + class ResourceApplyMomentum : public XlaOpKernel { public: explicit ResourceApplyMomentum(OpKernelConstruction* ctx) : XlaOpKernel(ctx) { @@ -63,8 +122,6 @@ class ResourceApplyMomentum : public XlaOpKernel { } void Compile(XlaOpKernelContext* ctx) override { - xla::XlaBuilder* b = ctx->builder(); - DataType type = ctx->input_type(2); TensorShape var_shape, accum_shape; @@ -97,14 +154,13 @@ class ResourceApplyMomentum : public XlaOpKernel { xla::XlaOp grad = ctx->Input(3); xla::XlaOp momentum = ctx->Input(4); - accum = b->Add(b->Mul(accum, momentum), grad); + accum = accum * momentum + grad; if (use_nesterov_) { // See https://github.com/tensorflow/tensorflow/pull/2798 for an // explanation of the reparameterization used here. - var = b->Sub( - var, b->Add(b->Mul(grad, lr), b->Mul(b->Mul(accum, momentum), lr))); + var = var - (grad * lr + accum * momentum * lr); } else { - var = b->Sub(var, b->Mul(accum, lr)); + var = var - accum * lr; } OP_REQUIRES_OK(ctx, ctx->AssignVariable(0, type, var)); OP_REQUIRES_OK(ctx, ctx->AssignVariable(1, type, accum)); @@ -121,8 +177,6 @@ class ResourceApplyAdagrad : public XlaOpKernel { explicit ResourceApplyAdagrad(OpKernelConstruction* ctx) : XlaOpKernel(ctx) {} void Compile(XlaOpKernelContext* ctx) override { - xla::XlaBuilder* b = ctx->builder(); - DataType type = ctx->input_type(2); TensorShape var_shape, accum_shape; @@ -149,10 +203,8 @@ class ResourceApplyAdagrad : public XlaOpKernel { xla::XlaOp lr = ctx->Input(2); xla::XlaOp grad = ctx->Input(3); - accum = b->Add(accum, b->Pow(grad, XlaHelpers::FloatLiteral(b, type, 2.0))); - var = b->Sub( - var, b->Mul(b->Mul(grad, lr), - b->Pow(accum, XlaHelpers::FloatLiteral(b, type, -0.5)))); + accum = accum + xla::Square(grad); + var = var - grad * lr * xla::Rsqrt(accum); OP_REQUIRES_OK(ctx, ctx->AssignVariable(0, type, var)); OP_REQUIRES_OK(ctx, ctx->AssignVariable(1, type, accum)); } @@ -160,6 +212,139 @@ class ResourceApplyAdagrad : public XlaOpKernel { REGISTER_XLA_OP(Name("ResourceApplyAdagrad").TypeConstraint("T", kFloatTypes), ResourceApplyAdagrad); +class ResourceApplyProximalAdagrad : public XlaOpKernel { + public: + explicit ResourceApplyProximalAdagrad(OpKernelConstruction* ctx) + : XlaOpKernel(ctx) { + OP_REQUIRES_OK(ctx, ctx->GetAttr("T", &dtype_)); + } + + void Compile(XlaOpKernelContext* ctx) override { + TensorShape var_shape, accum_shape; + xla::XlaOp var, accum; + OP_REQUIRES_OK(ctx, ctx->ReadVariableInput(0, dtype_, &var_shape, &var)); + OP_REQUIRES_OK(ctx, + ctx->ReadVariableInput(1, dtype_, &accum_shape, &accum)); + + OP_REQUIRES(ctx, var_shape.IsSameSize(accum_shape), + errors::InvalidArgument( + "var and accum do not have the same shape", + var_shape.DebugString(), " ", accum_shape.DebugString())); + + TensorShape lr_shape = ctx->InputShape(2); + OP_REQUIRES(ctx, TensorShapeUtils::IsScalar(lr_shape), + errors::InvalidArgument("lr is not a scalar: ", + lr_shape.DebugString())); + TensorShape l1_shape = ctx->InputShape(3); + OP_REQUIRES(ctx, TensorShapeUtils::IsScalar(l1_shape), + errors::InvalidArgument("l1 is not a scalar: ", + l1_shape.DebugString())); + TensorShape l2_shape = ctx->InputShape(4); + OP_REQUIRES(ctx, TensorShapeUtils::IsScalar(l2_shape), + errors::InvalidArgument("l2 is not a scalar: ", + l2_shape.DebugString())); + TensorShape grad_shape = ctx->InputShape(5); + OP_REQUIRES(ctx, var_shape.IsSameSize(grad_shape), + errors::InvalidArgument( + "var and grad do not have the same shape: ", + var_shape.DebugString(), " vs ", grad_shape.DebugString())); + + xla::XlaOp lr = ctx->Input(2); + xla::XlaOp l1 = ctx->Input(3); + xla::XlaOp l2 = ctx->Input(4); + xla::XlaOp grad = ctx->Input(5); + accum = accum + xla::Square(grad); + // Adagrad learning rate. + xla::XlaOp adagrad_lr = lr * xla::Rsqrt(accum); + var = ProximalGradientDescentUpdate(var, adagrad_lr, l1, l2, grad); + OP_REQUIRES_OK(ctx, ctx->AssignVariable(0, dtype_, var)); + OP_REQUIRES_OK(ctx, ctx->AssignVariable(1, dtype_, accum)); + } + + private: + DataType dtype_; +}; +REGISTER_XLA_OP( + Name("ResourceApplyProximalAdagrad").TypeConstraint("T", kFloatTypes), + ResourceApplyProximalAdagrad); + +class ResourceApplyAdagradDA : public XlaOpKernel { + public: + explicit ResourceApplyAdagradDA(OpKernelConstruction* ctx) + : XlaOpKernel(ctx) { + OP_REQUIRES_OK(ctx, ctx->GetAttr("T", &dtype_)); + } + + void Compile(XlaOpKernelContext* ctx) override { + TensorShape var_shape, accum_shape, squared_accum_shape; + xla::XlaOp var, accum, squared_accum; + OP_REQUIRES_OK(ctx, ctx->ReadVariableInput(0, dtype_, &var_shape, &var)); + OP_REQUIRES_OK(ctx, + ctx->ReadVariableInput(1, dtype_, &accum_shape, &accum)); + OP_REQUIRES_OK(ctx, ctx->ReadVariableInput(2, dtype_, &squared_accum_shape, + &squared_accum)); + OP_REQUIRES(ctx, var_shape.IsSameSize(accum_shape), + errors::InvalidArgument( + "var and accum do not have the same shape", + var_shape.DebugString(), " ", accum_shape.DebugString())); + OP_REQUIRES( + ctx, var_shape.IsSameSize(squared_accum_shape), + errors::InvalidArgument( + "var and squared accum do not have the same shape", + var_shape.DebugString(), " ", squared_accum_shape.DebugString())); + + TensorShape grad_shape = ctx->InputShape(3); + TensorShape lr_shape = ctx->InputShape(4); + TensorShape l1_shape = ctx->InputShape(5); + TensorShape l2_shape = ctx->InputShape(6); + TensorShape global_step_shape = ctx->InputShape(7); + + OP_REQUIRES(ctx, var_shape.IsSameSize(grad_shape), + errors::InvalidArgument( + "var and grad do not have the same shape", + var_shape.DebugString(), " ", grad_shape.DebugString())); + OP_REQUIRES(ctx, TensorShapeUtils::IsScalar(lr_shape), + errors::InvalidArgument("lr is not a scalar: ", + lr_shape.DebugString())); + OP_REQUIRES(ctx, TensorShapeUtils::IsScalar(l1_shape), + errors::InvalidArgument("l1 is not a scalar: ", + l1_shape.DebugString())); + OP_REQUIRES(ctx, TensorShapeUtils::IsScalar(l2_shape), + errors::InvalidArgument("l2 is not a scalar: ", + l2_shape.DebugString())); + OP_REQUIRES(ctx, TensorShapeUtils::IsScalar(global_step_shape), + errors::InvalidArgument("global step is not a scalar: ", + global_step_shape.DebugString())); + + xla::XlaOp grad = ctx->Input(3); + xla::XlaOp lr = ctx->Input(4); + xla::XlaOp l1 = ctx->Input(5); + xla::XlaOp l2 = ctx->Input(6); + xla::XlaBuilder* const b = ctx->builder(); + xla::XlaOp global_step = + XlaHelpers::ConvertElementType(b, ctx->Input(7), dtype_); + + accum = accum + grad; + squared_accum = squared_accum + xla::Square(grad); + xla::XlaOp zero = xla::ScalarLike(lr, 0.0); + xla::XlaOp denominator = global_step * lr * l2 + xla::Sqrt(squared_accum); + xla::XlaOp l1_le_zero = -lr * accum / denominator; + xla::XlaOp l1_gt_zero = -lr * xla::Sign(accum) * + xla::Max(xla::Abs(accum) - global_step * l1, zero) / + denominator; + + var = xla::Select(xla::Gt(l1, zero), l1_gt_zero, l1_le_zero); + OP_REQUIRES_OK(ctx, ctx->AssignVariable(0, dtype_, var)); + OP_REQUIRES_OK(ctx, ctx->AssignVariable(1, dtype_, accum)); + OP_REQUIRES_OK(ctx, ctx->AssignVariable(2, dtype_, squared_accum)); + } + + private: + DataType dtype_; +}; +REGISTER_XLA_OP(Name("ResourceApplyAdagradDA").TypeConstraint("T", kFloatTypes), + ResourceApplyAdagradDA); + class ResourceApplyAdam : public XlaOpKernel { public: explicit ResourceApplyAdam(OpKernelConstruction* ctx) : XlaOpKernel(ctx) { @@ -227,17 +412,12 @@ class ResourceApplyAdam : public XlaOpKernel { // variable <- variable - alpha * m_t / (sqrt(v_t) + epsilon) xla::XlaBuilder* b = ctx->builder(); - xla::XlaOp half = XlaHelpers::FloatLiteral(b, dtype_, 0.5); xla::XlaOp one = XlaHelpers::FloatLiteral(b, dtype_, 1.0); - xla::XlaOp two = XlaHelpers::FloatLiteral(b, dtype_, 2.0); - xla::XlaOp alpha = - b->Div(b->Mul(lr, b->Pow(b->Sub(one, beta2_power), half)), - b->Sub(one, beta1_power)); - m = b->Add(m, b->Mul(b->Sub(grad, m), b->Sub(one, beta1))); - v = b->Add(v, b->Mul(b->Sub(b->Pow(grad, two), v), b->Sub(one, beta2))); - var = - b->Sub(var, b->Div(b->Mul(m, alpha), b->Add(b->Pow(v, half), epsilon))); + xla::XlaOp alpha = lr * xla::Sqrt(one - beta2_power) / (one - beta1_power); + m = m + (grad - m) * (one - beta1); + v = v + (xla::Square(grad) - v) * (one - beta2); + var = var - m * alpha / (xla::Sqrt(v) + epsilon); OP_REQUIRES_OK(ctx, ctx->AssignVariable(0, dtype_, var)); OP_REQUIRES_OK(ctx, ctx->AssignVariable(1, dtype_, m)); @@ -250,38 +430,112 @@ class ResourceApplyAdam : public XlaOpKernel { REGISTER_XLA_OP(Name("ResourceApplyAdam").TypeConstraint("T", kFloatTypes), ResourceApplyAdam); -class ResourceApplyRMSProp : public XlaOpKernel { +class ResourceApplyAdaMax : public XlaOpKernel { public: - explicit ResourceApplyRMSProp(OpKernelConstruction* ctx) : XlaOpKernel(ctx) {} + explicit ResourceApplyAdaMax(OpKernelConstruction* ctx) : XlaOpKernel(ctx) { + OP_REQUIRES_OK(ctx, ctx->GetAttr("T", &dtype_)); + } void Compile(XlaOpKernelContext* ctx) override { - xla::XlaBuilder* b = ctx->builder(); + TensorShape var_shape, m_shape, v_shape; + xla::XlaOp var, m, v; + OP_REQUIRES_OK(ctx, ctx->ReadVariableInput(0, dtype_, &var_shape, &var)); + OP_REQUIRES_OK(ctx, ctx->ReadVariableInput(1, dtype_, &m_shape, &m)); + OP_REQUIRES_OK(ctx, ctx->ReadVariableInput(2, dtype_, &v_shape, &v)); - DataType type = ctx->input_type(3); + TensorShape beta1_power_shape = ctx->InputShape(3); + TensorShape lr_shape = ctx->InputShape(4); + TensorShape beta1_shape = ctx->InputShape(5); + TensorShape beta2_shape = ctx->InputShape(6); + TensorShape epsilon_shape = ctx->InputShape(7); + TensorShape grad_shape = ctx->InputShape(8); - TensorShape var_shape, ms_shape, mom_shape; - xla::XlaOp var, ms, mom; - OP_REQUIRES_OK(ctx, ctx->ReadVariableInput(0, type, &var_shape, &var)); - OP_REQUIRES_OK(ctx, ctx->ReadVariableInput(1, type, &ms_shape, &ms)); - OP_REQUIRES_OK(ctx, ctx->ReadVariableInput(2, type, &mom_shape, &mom)); + OP_REQUIRES(ctx, TensorShapeUtils::IsScalar(beta1_power_shape), + errors::InvalidArgument("beta1_power is not a scalar: ", + beta1_power_shape.DebugString())); + OP_REQUIRES(ctx, TensorShapeUtils::IsScalar(lr_shape), + errors::InvalidArgument("lr is not a scalar : ", + lr_shape.DebugString())); + OP_REQUIRES(ctx, TensorShapeUtils::IsScalar(beta1_shape), + errors::InvalidArgument("beta1 is not a scalar: ", + beta1_shape.DebugString())); + OP_REQUIRES(ctx, TensorShapeUtils::IsScalar(beta2_shape), + errors::InvalidArgument("beta2 is not a scalar: ", + beta2_shape.DebugString())); + OP_REQUIRES(ctx, TensorShapeUtils::IsScalar(epsilon_shape), + errors::InvalidArgument("epsilon is not a scalar: ", + epsilon_shape.DebugString())); + OP_REQUIRES(ctx, var_shape.IsSameSize(m_shape), + errors::InvalidArgument("var and m do not have the same shape", + var_shape.DebugString(), " ", + m_shape.DebugString())); + OP_REQUIRES(ctx, var_shape.IsSameSize(v_shape), + errors::InvalidArgument("var and v do not have the same shape", + var_shape.DebugString(), " ", + v_shape.DebugString())); + OP_REQUIRES(ctx, var_shape.IsSameSize(grad_shape), + errors::InvalidArgument( + "var and grad do not have the same shape", + var_shape.DebugString(), " ", grad_shape.DebugString())); - TensorShape lr_shape = ctx->InputShape(3); + xla::XlaOp beta1_power = ctx->Input(3); + xla::XlaOp lr = ctx->Input(4); + xla::XlaOp beta1 = ctx->Input(5); + xla::XlaOp beta2 = ctx->Input(6); + xla::XlaOp epsilon = ctx->Input(7); + xla::XlaOp grad = ctx->Input(8); + + xla::XlaOp one = xla::ScalarLike(lr, 1.0); + m = beta1 * m + (one - beta1) * grad; + v = xla::Max(beta2 * v, xla::Abs(grad)); + var = var - lr / (one - beta1_power) * (m / (v + epsilon)); + + OP_REQUIRES_OK(ctx, ctx->AssignVariable(0, dtype_, var)); + OP_REQUIRES_OK(ctx, ctx->AssignVariable(1, dtype_, m)); + OP_REQUIRES_OK(ctx, ctx->AssignVariable(2, dtype_, v)); + } + + private: + DataType dtype_; +}; +REGISTER_XLA_OP(Name("ResourceApplyAdaMax").TypeConstraint("T", kFloatTypes), + ResourceApplyAdaMax); + +class ResourceApplyRMSProp : public XlaOpKernel { + public: + explicit ResourceApplyRMSProp(OpKernelConstruction* ctx) : XlaOpKernel(ctx) { + OP_REQUIRES_OK(ctx, ctx->GetAttr("T", &dtype_)); + } + + void Compile(XlaOpKernelContext* ctx) override { + TensorShape var_shape, ms_shape, mom_shape, mg_shape; + xla::XlaOp var, ms, mom, mg; + OP_REQUIRES_OK(ctx, + ctx->ReadVariableInput("var", dtype_, &var_shape, &var)); + if (centered_) { + OP_REQUIRES_OK(ctx, ctx->ReadVariableInput("mg", dtype_, &mg_shape, &mg)); + } + OP_REQUIRES_OK(ctx, ctx->ReadVariableInput("ms", dtype_, &ms_shape, &ms)); + OP_REQUIRES_OK(ctx, + ctx->ReadVariableInput("mom", dtype_, &mom_shape, &mom)); + + TensorShape lr_shape = ctx->InputShape("lr"); OP_REQUIRES(ctx, TensorShapeUtils::IsScalar(lr_shape), errors::InvalidArgument("lr is not a scalar: ", lr_shape.DebugString())); - TensorShape rho_shape = ctx->InputShape(4); + TensorShape rho_shape = ctx->InputShape("rho"); OP_REQUIRES(ctx, TensorShapeUtils::IsScalar(rho_shape), errors::InvalidArgument("rho is not a scalar: ", rho_shape.DebugString())); - TensorShape momentum_shape = ctx->InputShape(5); + TensorShape momentum_shape = ctx->InputShape("momentum"); OP_REQUIRES(ctx, TensorShapeUtils::IsScalar(momentum_shape), errors::InvalidArgument("momentum is not a scalar: ", momentum_shape.DebugString())); - TensorShape epsilon_shape = ctx->InputShape(6); + TensorShape epsilon_shape = ctx->InputShape("epsilon"); OP_REQUIRES(ctx, TensorShapeUtils::IsScalar(epsilon_shape), errors::InvalidArgument("epsilon is not a scalar: ", epsilon_shape.DebugString())); - TensorShape grad_shape = ctx->InputShape(7); + TensorShape grad_shape = ctx->InputShape("grad"); // var should be the same shape as mom and ms. OP_REQUIRES(ctx, var_shape.IsSameSize(ms_shape), @@ -297,11 +551,11 @@ class ResourceApplyRMSProp : public XlaOpKernel { "var and grad do not have the same shape", var_shape.DebugString(), " ", grad_shape.DebugString())); - xla::XlaOp lr = ctx->Input(3); - xla::XlaOp rho = ctx->Input(4); - xla::XlaOp momentum = ctx->Input(5); - xla::XlaOp epsilon = ctx->Input(6); - xla::XlaOp grad = ctx->Input(7); + xla::XlaOp lr = ctx->Input("lr"); + xla::XlaOp rho = ctx->Input("rho"); + xla::XlaOp momentum = ctx->Input("momentum"); + xla::XlaOp epsilon = ctx->Input("epsilon"); + xla::XlaOp grad = ctx->Input("grad"); // ms <- rho * ms_{t-1} + (1-rho) * grad * grad // mom <- momentum * mom_{t-1} + lr * grad / sqrt(ms + epsilon) @@ -320,25 +574,46 @@ class ResourceApplyRMSProp : public XlaOpKernel { // ms <- grad**2 (1 - rho) + ms * rho // // Which is the equation listed above. - xla::XlaOp new_ms = b->Add( - ms, - b->Mul(b->Sub(b->Pow(grad, XlaHelpers::FloatLiteral(b, type, 2.0)), ms), - b->Sub(XlaHelpers::FloatLiteral(b, type, 1.0), rho))); - xla::XlaOp new_mom = - b->Add(b->Mul(mom, momentum), - b->Mul(b->Mul(grad, lr), - b->Pow(b->Add(new_ms, epsilon), - XlaHelpers::FloatLiteral(b, type, -0.5)))); - xla::XlaOp new_var = b->Sub(var, new_mom); - - OP_REQUIRES_OK(ctx, ctx->AssignVariable(0, type, new_var)); - OP_REQUIRES_OK(ctx, ctx->AssignVariable(1, type, new_ms)); - OP_REQUIRES_OK(ctx, ctx->AssignVariable(2, type, new_mom)); + xla::XlaOp one = xla::ScalarLike(ms, 1.0); + xla::XlaOp new_ms = xla::Square(grad) * (one - rho) + ms * rho; + xla::XlaOp denominator; + if (centered_) { + mg = grad * (one - rho) + mg * rho; + denominator = new_ms - xla::Square(mg) + epsilon; + } else { + denominator = new_ms + epsilon; + } + xla::XlaOp new_mom = mom * momentum + grad * lr * xla::Rsqrt(denominator); + xla::XlaOp new_var = var - new_mom; + + OP_REQUIRES_OK(ctx, ctx->AssignVariable("var", dtype_, new_var)); + if (centered_) { + OP_REQUIRES_OK(ctx, ctx->AssignVariable("mg", dtype_, mg)); + } + OP_REQUIRES_OK(ctx, ctx->AssignVariable("ms", dtype_, new_ms)); + OP_REQUIRES_OK(ctx, ctx->AssignVariable("mom", dtype_, new_mom)); } + + protected: + bool centered_ = false; + + private: + DataType dtype_; }; REGISTER_XLA_OP(Name("ResourceApplyRMSProp").TypeConstraint("T", kFloatTypes), ResourceApplyRMSProp); +class ResourceApplyCenteredRMSProp : public ResourceApplyRMSProp { + public: + explicit ResourceApplyCenteredRMSProp(OpKernelConstruction* ctx) + : ResourceApplyRMSProp(ctx) { + centered_ = true; + } +}; +REGISTER_XLA_OP( + Name("ResourceApplyCenteredRMSProp").TypeConstraint("T", kFloatTypes), + ResourceApplyCenteredRMSProp); + void CompileFtrl(XlaOpKernelContext* ctx, DataType dtype, bool has_l2_shrinkage) { xla::XlaBuilder* b = ctx->builder(); @@ -424,21 +699,18 @@ void CompileFtrl(XlaOpKernelContext* ctx, DataType dtype, xla::XlaOp two = XlaHelpers::FloatLiteral(b, dtype, 2.0); xla::XlaOp grad_to_use; if (has_l2_shrinkage) { - grad_to_use = b->Add(grad, b->Mul(two, b->Mul(l2_shrinkage, var))); + grad_to_use = grad + two * l2_shrinkage * var; } else { grad_to_use = grad; } - xla::XlaOp new_accum = b->Add(accum, b->Pow(grad_to_use, two)); - xla::XlaOp new_accum_lr_pow = b->Pow(new_accum, b->Neg(lr_power)); - xla::XlaOp accum_lr_pow = b->Pow(accum, b->Neg(lr_power)); - linear = b->Add( - linear, - b->Sub(grad_to_use, - b->Mul(b->Div(b->Sub(new_accum_lr_pow, accum_lr_pow), lr), var))); - xla::XlaOp linear_clipped = b->Clamp(b->Neg(l1), linear, l1); - xla::XlaOp quadratic = b->Add(b->Div(new_accum_lr_pow, lr), b->Mul(two, l2)); - var = b->Div(b->Sub(linear_clipped, linear), quadratic); + xla::XlaOp new_accum = accum + xla::Square(grad_to_use); + xla::XlaOp new_accum_lr_pow = xla::Pow(new_accum, -lr_power); + xla::XlaOp accum_lr_pow = xla::Pow(accum, -lr_power); + linear = linear + grad_to_use - (new_accum_lr_pow - accum_lr_pow) / lr * var; + xla::XlaOp linear_clipped = xla::Clamp(-l1, linear, l1); + xla::XlaOp quadratic = new_accum_lr_pow / lr + two * l2; + var = (linear_clipped - linear) / quadratic; accum = new_accum; OP_REQUIRES_OK(ctx, ctx->AssignVariable(0, dtype, var)); @@ -478,5 +750,176 @@ class ResourceApplyFtrlV2 : public XlaOpKernel { REGISTER_XLA_OP(Name("ResourceApplyFtrlV2").TypeConstraint("T", kFloatTypes), ResourceApplyFtrlV2); +class ResourceApplyAdadelta : public XlaOpKernel { + public: + explicit ResourceApplyAdadelta(OpKernelConstruction* ctx) : XlaOpKernel(ctx) { + OP_REQUIRES_OK(ctx, ctx->GetAttr("T", &dtype_)); + } + + void Compile(XlaOpKernelContext* ctx) override { + TensorShape var_shape, accum_shape, accum_update_shape; + xla::XlaOp var, accum, accum_update; + OP_REQUIRES_OK(ctx, ctx->ReadVariableInput(0, dtype_, &var_shape, &var)); + OP_REQUIRES_OK(ctx, + ctx->ReadVariableInput(1, dtype_, &accum_shape, &accum)); + OP_REQUIRES_OK(ctx, ctx->ReadVariableInput(2, dtype_, &accum_update_shape, + &accum_update)); + + TensorShape lr_shape = ctx->InputShape(3); + TensorShape rho_shape = ctx->InputShape(4); + TensorShape epsilon_shape = ctx->InputShape(5); + TensorShape grad_shape = ctx->InputShape(6); + + OP_REQUIRES(ctx, TensorShapeUtils::IsScalar(lr_shape), + errors::InvalidArgument("lr is not a scalar: ", + lr_shape.DebugString())); + + OP_REQUIRES(ctx, TensorShapeUtils::IsScalar(rho_shape), + errors::InvalidArgument("rho is not a scalar: ", + rho_shape.DebugString())); + + OP_REQUIRES(ctx, TensorShapeUtils::IsScalar(epsilon_shape), + errors::InvalidArgument("epsilon is not a scalar: ", + epsilon_shape.DebugString())); + + OP_REQUIRES(ctx, var_shape.IsSameSize(accum_shape), + errors::InvalidArgument( + "var and accum do not have the same shape", + var_shape.DebugString(), " ", accum_shape.DebugString())); + + OP_REQUIRES(ctx, var_shape.IsSameSize(grad_shape), + errors::InvalidArgument( + "var and grad do not have the same shape", + var_shape.DebugString(), " ", grad_shape.DebugString())); + + xla::XlaOp lr = ctx->Input(3); + xla::XlaOp rho = ctx->Input(4); + xla::XlaOp epsilon = ctx->Input(5); + xla::XlaOp grad = ctx->Input(6); + + xla::XlaBuilder* b = ctx->builder(); + xla::XlaOp neg_half = XlaHelpers::FloatLiteral(b, dtype_, -0.5); + xla::XlaOp half = XlaHelpers::FloatLiteral(b, dtype_, 0.5); + xla::XlaOp one = XlaHelpers::FloatLiteral(b, dtype_, 1.0); + xla::XlaOp two = XlaHelpers::FloatLiteral(b, dtype_, 2.0); + + accum = rho * accum + (one - rho) * xla::Pow(grad, two); + xla::XlaOp update = xla::Pow(accum_update + epsilon, half) * + xla::Pow(accum + epsilon, neg_half) * grad; + accum_update = rho * accum_update + (one - rho) * xla::Pow(update, two); + var = var - update * lr; + OP_REQUIRES_OK(ctx, ctx->AssignVariable(0, dtype_, var)); + OP_REQUIRES_OK(ctx, ctx->AssignVariable(1, dtype_, accum)); + OP_REQUIRES_OK(ctx, ctx->AssignVariable(2, dtype_, accum_update)); + } + + private: + DataType dtype_; +}; +REGISTER_XLA_OP(Name("ResourceApplyAdadelta").TypeConstraint("T", kFloatTypes), + ResourceApplyAdadelta); + +class ResourceApplySignBase : public XlaOpKernel { + public: + explicit ResourceApplySignBase(OpKernelConstruction* ctx) : XlaOpKernel(ctx) { + OP_REQUIRES_OK(ctx, ctx->GetAttr("T", &dtype_)); + } + + void Compile(XlaOpKernelContext* ctx) override { + TensorShape var_shape, m_shape; + xla::XlaOp var, m; + OP_REQUIRES_OK(ctx, ctx->ReadVariableInput(0, dtype_, &var_shape, &var)); + OP_REQUIRES_OK(ctx, ctx->ReadVariableInput(1, dtype_, &m_shape, &m)); + OP_REQUIRES(ctx, var_shape.IsSameSize(m_shape), + errors::InvalidArgument("var and m do not have the same shape", + var_shape.DebugString(), " ", + m_shape.DebugString())); + TensorShape grad_shape = ctx->InputShape(6); + OP_REQUIRES(ctx, var_shape.IsSameSize(grad_shape), + errors::InvalidArgument( + "var and grad do not have the same shape", + var_shape.DebugString(), " ", grad_shape.DebugString())); + CheckScalarParams(ctx); + + xla::XlaOp lr = ctx->Input(2); + xla::XlaOp alpha = ctx->Input(3); + xla::XlaOp sign_decay = ctx->Input(4); + xla::XlaOp beta = ctx->Input(5); + xla::XlaOp grad = ctx->Input(6); + + m = m * beta + grad * (xla::ScalarLike(beta, 1.0) - beta); + xla::XlaOp decay = xla::Sign(grad) * xla::Sign(m) * sign_decay; + + xla::XlaOp grad_scale = ComputeGradientScale(alpha, decay); + var = var - lr * grad_scale * grad; + OP_REQUIRES_OK(ctx, ctx->AssignVariable(0, dtype_, var)); + OP_REQUIRES_OK(ctx, ctx->AssignVariable(1, dtype_, m)); + } + + virtual void CheckScalarParams(XlaOpKernelContext* ctx) { + TensorShape lr_shape = ctx->InputShape(2); + TensorShape sign_decay_shape = ctx->InputShape(4); + TensorShape beta_shape = ctx->InputShape(5); + + OP_REQUIRES(ctx, TensorShapeUtils::IsScalar(lr_shape), + errors::InvalidArgument("lr is not a scalar: ", + lr_shape.DebugString())); + + OP_REQUIRES(ctx, TensorShapeUtils::IsScalar(sign_decay_shape), + errors::InvalidArgument("sign_decay is not a scalar: ", + sign_decay_shape.DebugString())); + + OP_REQUIRES(ctx, TensorShapeUtils::IsScalar(beta_shape), + errors::InvalidArgument("beta is not a scalar: ", + beta_shape.DebugString())); + } + + virtual xla::XlaOp ComputeGradientScale(xla::XlaOp alpha, + xla::XlaOp decay) = 0; + + private: + DataType dtype_; +}; + +class ResourceApplyAddSign : public ResourceApplySignBase { + public: + explicit ResourceApplyAddSign(OpKernelConstruction* ctx) + : ResourceApplySignBase(ctx) {} + + void CheckScalarParams(XlaOpKernelContext* ctx) override { + ResourceApplySignBase::CheckScalarParams(ctx); + TensorShape alpha_shape = ctx->InputShape(3); + OP_REQUIRES(ctx, TensorShapeUtils::IsScalar(alpha_shape), + errors::InvalidArgument("alpha is not a scalar: ", + alpha_shape.DebugString())); + } + + xla::XlaOp ComputeGradientScale(xla::XlaOp alpha, xla::XlaOp decay) override { + return alpha + decay; + } +}; +REGISTER_XLA_OP(Name("ResourceApplyAddSign").TypeConstraint("T", kFloatTypes), + ResourceApplyAddSign); + +class ResourceApplyPowerSign : public ResourceApplySignBase { + public: + explicit ResourceApplyPowerSign(OpKernelConstruction* ctx) + : ResourceApplySignBase(ctx) {} + + void CheckScalarParams(XlaOpKernelContext* ctx) override { + ResourceApplySignBase::CheckScalarParams(ctx); + TensorShape logbase_shape = ctx->InputShape(3); + OP_REQUIRES(ctx, TensorShapeUtils::IsScalar(logbase_shape), + errors::InvalidArgument("logbase is not a scalar: ", + logbase_shape.DebugString())); + } + + xla::XlaOp ComputeGradientScale(xla::XlaOp alpha, xla::XlaOp decay) override { + return xla::Exp(alpha * decay); + } +}; +REGISTER_XLA_OP(Name("ResourceApplyPowerSign").TypeConstraint("T", kFloatTypes), + ResourceApplyPowerSign); + } // namespace } // namespace tensorflow diff --git a/tensorflow/compiler/tf2xla/kernels/transpose_op.cc b/tensorflow/compiler/tf2xla/kernels/transpose_op.cc index ef5aae81a8d73ba326d4116d48b9eebee3c44098..6c721c48fe3af45aff5cd0bd5e74e2693faf9f97 100644 --- a/tensorflow/compiler/tf2xla/kernels/transpose_op.cc +++ b/tensorflow/compiler/tf2xla/kernels/transpose_op.cc @@ -23,6 +23,7 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/xla_helpers.h" #include "tensorflow/compiler/tf2xla/xla_op_kernel.h" #include "tensorflow/compiler/tf2xla/xla_op_registry.h" +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/core/framework/kernel_def_builder.h" #include "tensorflow/core/framework/register_types.h" #include "tensorflow/core/kernels/bounds_check.h" @@ -84,12 +85,12 @@ class TransposeOp : public XlaOpKernel { if (dims <= 1 || is_identity) { transposed = ctx->Input(0); } else { - transposed = ctx->builder()->Transpose(ctx->Input(0), transposed_order); + transposed = xla::Transpose(ctx->Input(0), transposed_order); } // Conjugate the transposed result if this is ConjugateTransposeOp. if (conjugate_) { - ctx->SetOutput(0, ctx->builder()->Conj(transposed)); + ctx->SetOutput(0, xla::Conj(transposed)); } else { ctx->SetOutput(0, transposed); } @@ -146,7 +147,7 @@ class InvertPermutationOp : public XlaOpKernel { output[d] = i; } - ctx->SetOutput(0, ctx->builder()->ConstantR1(output)); + ctx->SetOutput(0, xla::ConstantR1(ctx->builder(), output)); } }; diff --git a/tensorflow/compiler/tf2xla/kernels/unary_ops.cc b/tensorflow/compiler/tf2xla/kernels/unary_ops.cc index a39e5dcfc560c1f37712e52a47c33811caecebd3..116a020437e263f1d3d82fee5c0ea0ca4f97e634 100644 --- a/tensorflow/compiler/tf2xla/kernels/unary_ops.cc +++ b/tensorflow/compiler/tf2xla/kernels/unary_ops.cc @@ -21,21 +21,21 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/xla_op_registry.h" #include "tensorflow/compiler/xla/client/client_library.h" #include "tensorflow/compiler/xla/client/lib/arithmetic.h" +#include "tensorflow/compiler/xla/client/lib/constants.h" +#include "tensorflow/compiler/xla/client/lib/math.h" #include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/core/framework/kernel_def_builder.h" namespace tensorflow { namespace { -// A subclass of a TlaUnaryOp must build the lambda computation that -// describes the scalar->scalar function to apply to each element of -// the input. #define XLAJIT_MAKE_UNARY(NAME, COMPUTATION) \ class NAME##Op : public XlaOpKernel { \ public: \ explicit NAME##Op(OpKernelConstruction* ctx) : XlaOpKernel(ctx) {} \ void Compile(XlaOpKernelContext* ctx) { \ xla::XlaBuilder* b = ctx->builder(); \ + (void)b; \ xla::XlaOp x = ctx->Input(0); \ xla::XlaOp y = COMPUTATION; \ ctx->SetOutput(0, y); \ @@ -43,122 +43,100 @@ namespace { }; \ REGISTER_XLA_OP(Name(#NAME), NAME##Op); -XLAJIT_MAKE_UNARY(ComplexAbs, b->Abs(x)); +XLAJIT_MAKE_UNARY(ComplexAbs, xla::Abs(x)); -XLAJIT_MAKE_UNARY(Angle, b->Atan2(b->Imag(x), b->Real(x))); +XLAJIT_MAKE_UNARY(Angle, xla::Atan2(xla::Imag(x), xla::Real(x))); -XLAJIT_MAKE_UNARY(Conj, b->Conj(x)); +XLAJIT_MAKE_UNARY(Conj, xla::Conj(x)); // Return x if x>0, otherwise -x. -XLAJIT_MAKE_UNARY(Abs, b->Abs(x)); +XLAJIT_MAKE_UNARY(Abs, xla::Abs(x)); // acos(x) = 2 * atan(sqrt(1 - x^2) / (1 + x)) -XLAJIT_MAKE_UNARY( - Acos, - b->Mul(XlaHelpers::FloatLiteral(b, input_type(0), 2.0), - b->Atan2(b->Pow(b->Sub(XlaHelpers::One(b, input_type(0)), - b->Mul(x, x)), - XlaHelpers::FloatLiteral(b, input_type(0), 0.5)), - b->Add(XlaHelpers::One(b, input_type(0)), x)))); +XLAJIT_MAKE_UNARY(Acos, + xla::ScalarLike(x, 2.0) * + xla::Atan2(xla::Sqrt(xla::ScalarLike(x, 1.0) - x * x), + xla::ScalarLike(x, 1.0) + x)); // acosh(x) = log(x + sqrt(x^2 - 1)) // = log(x + sqrt((x+1)*(x-1))) -XLAJIT_MAKE_UNARY( - Acosh, - b->Log(b->Add(x, - b->Pow(b->Mul(b->Add(x, XlaHelpers::One(b, input_type(0))), - b->Sub(x, XlaHelpers::One(b, input_type(0)))), - XlaHelpers::FloatLiteral(b, input_type(0), 0.5))))); +XLAJIT_MAKE_UNARY(Acosh, + xla::Log(x + xla::Sqrt((x + xla::ScalarLike(x, 1.0)) * + (x - xla::ScalarLike(x, 1.0))))); // asin(x) = 2 * atan(x / (1 + sqrt(1 - x^2))) XLAJIT_MAKE_UNARY( - Asin, - b->Mul(XlaHelpers::FloatLiteral(b, input_type(0), 2.0), - b->Atan2(x, b->Add(XlaHelpers::One(b, input_type(0)), - b->Pow(b->Sub(XlaHelpers::One(b, input_type(0)), - b->Mul(x, x)), - XlaHelpers::FloatLiteral(b, input_type(0), - 0.5)))))); + Asin, xla::ScalarLike(x, 2.0) * + xla::Atan2(x, xla::ScalarLike(x, 1.0) + + xla::Sqrt(xla::ScalarLike(x, 1.0) - x * x))); // asinh(x) = log(x + sqrt(x^2 + 1)) -XLAJIT_MAKE_UNARY( - Asinh, - b->Log(b->Add(x, b->Pow(b->Add(b->Mul(x, x), - XlaHelpers::One(b, input_type(0))), - XlaHelpers::FloatLiteral(b, input_type(0), 0.5))))); +XLAJIT_MAKE_UNARY(Asinh, + xla::Log(x + xla::Sqrt(x * x + xla::ScalarLike(x, 1.0)))); -XLAJIT_MAKE_UNARY(Atan, b->Atan2(x, XlaHelpers::One(b, input_type(0)))); +XLAJIT_MAKE_UNARY(Atan, xla::Atan2(x, xla::ScalarLike(x, 1.0))); // atanh(x) = 0.5 * log((1 + x) / (1 - x)) +XLAJIT_MAKE_UNARY(Atanh, xla::Log((xla::ScalarLike(x, 1.0) + x) / + (xla::ScalarLike(x, 1.0) - x)) * + xla::ScalarLike(x, 0.5)); +XLAJIT_MAKE_UNARY(Ceil, xla::Ceil(x)); +XLAJIT_MAKE_UNARY(Cos, xla::Cos(x)); +XLAJIT_MAKE_UNARY(Cosh, (xla::Exp(x) + xla::Exp(-x)) * xla::ScalarLike(x, 0.5)); +XLAJIT_MAKE_UNARY(Sin, xla::Sin(x)); +XLAJIT_MAKE_UNARY(Exp, xla::Exp(x)); + +XLAJIT_MAKE_UNARY(Expm1, xla::Expm1(x)); + +XLAJIT_MAKE_UNARY(Floor, xla::Floor(x)); +XLAJIT_MAKE_UNARY(IsFinite, xla::IsFinite(x)); XLAJIT_MAKE_UNARY( - Atanh, b->Mul(b->Log(b->Div(b->Add(XlaHelpers::One(b, input_type(0)), x), - b->Sub(XlaHelpers::One(b, input_type(0)), x))), - XlaHelpers::FloatLiteral(b, input_type(0), 0.5))); -XLAJIT_MAKE_UNARY(Ceil, b->Ceil(x)); -XLAJIT_MAKE_UNARY(Cos, b->Cos(x)); -XLAJIT_MAKE_UNARY(Cosh, - b->Mul(b->Add(b->Exp(x), b->Exp(b->Neg(x))), - XlaHelpers::FloatLiteral(b, input_type(0), 0.5))); -XLAJIT_MAKE_UNARY(Sin, b->Sin(x)); -XLAJIT_MAKE_UNARY(Exp, b->Exp(x)); - -XLAJIT_MAKE_UNARY(Expm1, b->Expm1(x)); - -XLAJIT_MAKE_UNARY(Floor, b->Floor(x)); -XLAJIT_MAKE_UNARY(IsFinite, b->IsFinite(x)); -XLAJIT_MAKE_UNARY(IsInf, b->Eq(b->Abs(x), - XlaHelpers::FloatLiteral( - b, input_type(0), - std::numeric_limits::infinity()))); -XLAJIT_MAKE_UNARY(IsNan, b->Ne(x, x)); + IsInf, + xla::Eq(xla::Abs(x), + xla::ScalarLike(x, std::numeric_limits::infinity()))); +XLAJIT_MAKE_UNARY(IsNan, xla::Ne(x, x)); // Return 1/x -XLAJIT_MAKE_UNARY(Inv, b->Div(XlaHelpers::One(b, input_type(0)), x)); -XLAJIT_MAKE_UNARY(Reciprocal, b->Div(XlaHelpers::One(b, input_type(0)), x)); -XLAJIT_MAKE_UNARY(Log, b->Log(x)); +XLAJIT_MAKE_UNARY(Inv, xla::ScalarLike(x, 1.0) / x); +XLAJIT_MAKE_UNARY(Reciprocal, xla::ScalarLike(x, 1.0) / x); +XLAJIT_MAKE_UNARY(Log, xla::Log(x)); -XLAJIT_MAKE_UNARY(Log1p, b->Log1p(x)); +XLAJIT_MAKE_UNARY(Log1p, xla::Log1p(x)); -XLAJIT_MAKE_UNARY(Invert, b->Not(x)); -XLAJIT_MAKE_UNARY(LogicalNot, b->Not(x)); -XLAJIT_MAKE_UNARY(Neg, b->Neg(x)); +XLAJIT_MAKE_UNARY(Invert, xla::Not(x)); +XLAJIT_MAKE_UNARY(LogicalNot, xla::Not(x)); +XLAJIT_MAKE_UNARY(Neg, -x); // Implements Banker's rounding: numbers that are equidistant between two // integers are rounded towards even. -static xla::XlaOp Round(xla::XlaBuilder* b, DataType dtype, - const xla::XlaOp& x) { - auto half = XlaHelpers::FloatLiteral(b, dtype, 0.5); - auto one = XlaHelpers::FloatLiteral(b, dtype, 1.0); - auto two = XlaHelpers::FloatLiteral(b, dtype, 2.0); - - auto round_val = b->Floor(x); - auto fraction = b->Sub(x, round_val); - auto nearest_even_int = - b->Sub(round_val, b->Mul(two, b->Floor(b->Mul(half, x)))); - auto is_odd = b->Eq(nearest_even_int, one); - return b->Select( - b->Or(b->Gt(fraction, half), b->And(b->Eq(fraction, half), is_odd)), - b->Add(round_val, one), round_val); +xla::XlaOp RoundToEven(xla::XlaOp x) { + auto half = xla::ScalarLike(x, 0.5); + auto one = xla::ScalarLike(x, 1.0); + auto two = xla::ScalarLike(x, 2.0); + + auto round_val = xla::Floor(x); + auto fraction = x - round_val; + auto nearest_even_int = round_val - two * xla::Floor(half * x); + auto is_odd = xla::Eq(nearest_even_int, one); + return xla::Select(xla::Or(xla::Gt(fraction, half), + xla::And(xla::Eq(fraction, half), is_odd)), + round_val + one, round_val); } -XLAJIT_MAKE_UNARY(Rint, Round(b, input_type(0), x)); -XLAJIT_MAKE_UNARY(Round, Round(b, input_type(0), x)); +XLAJIT_MAKE_UNARY(Rint, RoundToEven(x)); +XLAJIT_MAKE_UNARY(Round, RoundToEven(x)); -XLAJIT_MAKE_UNARY(Rsqrt, - b->Pow(x, XlaHelpers::FloatLiteral(b, input_type(0), -0.5))); +XLAJIT_MAKE_UNARY(Rsqrt, xla::Rsqrt(x)); // Expresses sigmoid as a rescaled tanh: sigmoid(x) == (tanh(x/2) + 1) / 2. -static xla::XlaOp Sigmoid(xla::XlaBuilder* b, DataType dtype, - const xla::XlaOp& x) { - auto half = XlaHelpers::FloatLiteral(b, dtype, 0.5); - return b->Add(half, b->Mul(half, b->Tanh(b->Mul(half, x)))); +xla::XlaOp Sigmoid(xla::XlaOp x) { + auto half = xla::ScalarLike(x, 0.5); + return half + half * xla::Tanh(half * x); } -XLAJIT_MAKE_UNARY(Sigmoid, Sigmoid(b, input_type(0), x)); +XLAJIT_MAKE_UNARY(Sigmoid, Sigmoid(x)); // Returns 0 if x is 0, -1 if x < 0 and 1 if x > 0. -XLAJIT_MAKE_UNARY(Sign, b->Sign(x)); -XLAJIT_MAKE_UNARY(Sinh, - b->Mul(b->Sub(b->Exp(x), b->Exp(b->Neg(x))), - XlaHelpers::FloatLiteral(b, input_type(0), 0.5))); +XLAJIT_MAKE_UNARY(Sign, xla::Sign(x)); +XLAJIT_MAKE_UNARY(Sinh, (xla::Exp(x) - xla::Exp(-x)) * xla::ScalarLike(x, 0.5)); // softplus(x) = log(1 + exp(x)) // @@ -168,22 +146,18 @@ XLAJIT_MAKE_UNARY(Sinh, // // This is equivalent to: // max(x, 0) + log1p(exp(-abs(x))) -XLAJIT_MAKE_UNARY(Softplus, - b->Add(b->Max(x, XlaHelpers::Zero(b, input_type(0))), - b->Log1p(b->Exp(b->Neg(b->Abs(x)))))); +XLAJIT_MAKE_UNARY(Softplus, xla::Max(x, xla::ScalarLike(x, 0.0)) + + xla::Log1p(xla::Exp(-xla::Abs(x)))); // softsign(x) = x / (abs(x) + 1) -XLAJIT_MAKE_UNARY(Softsign, - b->Div(x, - b->Add(b->Abs(x), XlaHelpers::One(b, input_type(0))))); -XLAJIT_MAKE_UNARY(Sqrt, - b->Pow(x, XlaHelpers::FloatLiteral(b, input_type(0), 0.5))); -XLAJIT_MAKE_UNARY(Square, b->Mul(x, x)); -XLAJIT_MAKE_UNARY(Tan, b->Div(b->Sin(x), b->Cos(x))); -XLAJIT_MAKE_UNARY(Tanh, b->Tanh(x)); - -XLAJIT_MAKE_UNARY(Real, b->Real(x)); -XLAJIT_MAKE_UNARY(Imag, b->Imag(x)); +XLAJIT_MAKE_UNARY(Softsign, x / (xla::Abs(x) + xla::ScalarLike(x, 1.0))); +XLAJIT_MAKE_UNARY(Sqrt, xla::Sqrt(x)); +XLAJIT_MAKE_UNARY(Square, x* x); +XLAJIT_MAKE_UNARY(Tan, xla::Sin(x) / xla::Cos(x)); +XLAJIT_MAKE_UNARY(Tanh, xla::Tanh(x)); + +XLAJIT_MAKE_UNARY(Real, xla::Real(x)); +XLAJIT_MAKE_UNARY(Imag, xla::Imag(x)); #undef XLAJIT_MAKE_UNARY @@ -193,17 +167,10 @@ class ErfOp : public XlaOpKernel { public: explicit ErfOp(OpKernelConstruction* ctx) : XlaOpKernel(ctx) {} void Compile(XlaOpKernelContext* ctx) override { - xla::XlaBuilder* b = ctx->builder(); - xla::PrimitiveType primitive_type; - xla::XlaOp one = XlaHelpers::One(b, input_type(0)); xla::XlaOp x = ctx->Input(0); - xla::XlaOp abs_x = b->Abs(x); - - OP_REQUIRES_OK(ctx, - DataTypeToPrimitiveType(input_type(0), &primitive_type)); - - auto y = b->Select(b->Gt(abs_x, one), b->Sub(one, Erfc(x, primitive_type)), - Erf(x, primitive_type)); + xla::XlaOp one = xla::ScalarLike(x, 1.0); + auto y = + xla::Select(xla::Gt(xla::Abs(x), one), one - xla::Erfc(x), xla::Erf(x)); ctx->SetOutput(0, y); } }; @@ -213,17 +180,10 @@ class ErfcOp : public XlaOpKernel { public: explicit ErfcOp(OpKernelConstruction* ctx) : XlaOpKernel(ctx) {} void Compile(XlaOpKernelContext* ctx) override { - xla::XlaBuilder* b = ctx->builder(); - xla::XlaOp one = XlaHelpers::One(b, input_type(0)); xla::XlaOp x = ctx->Input(0); - xla::XlaOp abs_x = b->Abs(x); - - xla::PrimitiveType primitive_type; - OP_REQUIRES_OK(ctx, - DataTypeToPrimitiveType(input_type(0), &primitive_type)); - - auto y = b->Select(b->Lt(abs_x, one), b->Sub(one, Erf(x, primitive_type)), - Erfc(x, primitive_type)); + xla::XlaOp one = xla::ScalarLike(x, 1.0); + auto y = + xla::Select(xla::Lt(xla::Abs(x), one), one - xla::Erf(x), xla::Erfc(x)); ctx->SetOutput(0, y); } }; diff --git a/tensorflow/compiler/tf2xla/kernels/unpack_op.cc b/tensorflow/compiler/tf2xla/kernels/unpack_op.cc index f87586ba578a6138e7fb921032e1a71f8c9ac80c..f951127bb95cd52864af869676a6b4c4961c1a43 100644 --- a/tensorflow/compiler/tf2xla/kernels/unpack_op.cc +++ b/tensorflow/compiler/tf2xla/kernels/unpack_op.cc @@ -22,7 +22,8 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/xla_helpers.h" #include "tensorflow/compiler/tf2xla/xla_op_kernel.h" #include "tensorflow/compiler/tf2xla/xla_op_registry.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/core/framework/op_kernel.h" #include "tensorflow/core/framework/register_types.h" #include "tensorflow/core/framework/tensor.h" @@ -74,10 +75,9 @@ class UnpackOp : public XlaOpKernel { for (int i = 0; i < num; ++i) { start_indices[axis] = i; limit_indices[axis] = i + 1; - auto slice = ctx->builder()->Slice(input, start_indices, limit_indices, - strides); + auto slice = xla::Slice(input, start_indices, limit_indices, strides); // Reshape to drop the 'axis' dimension. - auto result = ctx->builder()->Reshape(slice, output_shape.dim_sizes()); + auto result = xla::Reshape(slice, output_shape.dim_sizes()); ctx->SetOutput(i, result); } } diff --git a/tensorflow/compiler/tf2xla/kernels/variable_ops.cc b/tensorflow/compiler/tf2xla/kernels/variable_ops.cc index ad51396bdf8815e8f93a3e3a04d624e096136bf1..bb27b5d56f3c24dc093a60e698b1080dfb76514d 100644 --- a/tensorflow/compiler/tf2xla/kernels/variable_ops.cc +++ b/tensorflow/compiler/tf2xla/kernels/variable_ops.cc @@ -20,7 +20,7 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/xla_helpers.h" #include "tensorflow/compiler/tf2xla/xla_op_registry.h" #include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/core/framework/kernel_def_builder.h" #include "tensorflow/core/framework/types.h" @@ -33,8 +33,8 @@ class VarIsInitializedOp : public XlaOpKernel { void Compile(XlaOpKernelContext* ctx) override { XlaResource* variable; OP_REQUIRES_OK(ctx, ctx->GetResourceInput(0, &variable)); - ctx->SetOutput(0, - ctx->builder()->ConstantR0(variable->initialized())); + ctx->SetOutput( + 0, xla::ConstantR0(ctx->builder(), variable->initialized())); } }; REGISTER_XLA_OP(Name("VarIsInitializedOp"), VarIsInitializedOp); @@ -96,7 +96,7 @@ class AssignAddVariableOp : public XlaOpKernel { xla::XlaOp handle; OP_REQUIRES_OK(ctx, ctx->ReadVariableInput(0, type, /*shape=*/nullptr, &handle)); - handle = ctx->builder()->Add(handle, ctx->Input(1)); + handle = xla::Add(handle, ctx->Input(1)); OP_REQUIRES_OK(ctx, ctx->AssignVariable(0, type, handle)); } }; @@ -112,7 +112,7 @@ class AssignSubVariableOp : public XlaOpKernel { xla::XlaOp handle; OP_REQUIRES_OK(ctx, ctx->ReadVariableInput(0, type, /*shape=*/nullptr, &handle)); - handle = ctx->builder()->Sub(handle, ctx->Input(1)); + handle = xla::Sub(handle, ctx->Input(1)); OP_REQUIRES_OK(ctx, ctx->AssignVariable(0, type, handle)); } }; @@ -191,7 +191,7 @@ class ResourceScatterAddOp : public ResourceScatterOp { private: static xla::XlaOp Combine(const xla::XlaOp& x, const xla::XlaOp& y, xla::XlaBuilder* builder) { - return builder->Add(x, y); + return xla::Add(x, y); } }; REGISTER_XLA_OP(Name("ResourceScatterAdd"), ResourceScatterAddOp); @@ -204,7 +204,7 @@ class ResourceScatterSubOp : public ResourceScatterOp { private: static xla::XlaOp Combine(const xla::XlaOp& x, const xla::XlaOp& y, xla::XlaBuilder* builder) { - return builder->Sub(x, y); + return xla::Sub(x, y); } }; REGISTER_XLA_OP(Name("ResourceScatterSub"), ResourceScatterSubOp); @@ -217,7 +217,7 @@ class ResourceScatterMulOp : public ResourceScatterOp { private: static xla::XlaOp Combine(const xla::XlaOp& x, const xla::XlaOp& y, xla::XlaBuilder* builder) { - return builder->Mul(x, y); + return xla::Mul(x, y); } }; REGISTER_XLA_OP(Name("ResourceScatterMul"), ResourceScatterMulOp); @@ -230,7 +230,7 @@ class ResourceScatterDivOp : public ResourceScatterOp { private: static xla::XlaOp Combine(const xla::XlaOp& x, const xla::XlaOp& y, xla::XlaBuilder* builder) { - return builder->Div(x, y); + return xla::Div(x, y); } }; REGISTER_XLA_OP(Name("ResourceScatterDiv"), ResourceScatterDivOp); @@ -243,7 +243,7 @@ class ResourceScatterMinOp : public ResourceScatterOp { private: static xla::XlaOp Combine(const xla::XlaOp& x, const xla::XlaOp& y, xla::XlaBuilder* builder) { - return builder->Min(x, y); + return xla::Min(x, y); } }; REGISTER_XLA_OP(Name("ResourceScatterMin"), ResourceScatterMinOp); @@ -256,7 +256,7 @@ class ResourceScatterMaxOp : public ResourceScatterOp { private: static xla::XlaOp Combine(const xla::XlaOp& x, const xla::XlaOp& y, xla::XlaBuilder* builder) { - return builder->Max(x, y); + return xla::Max(x, y); } }; REGISTER_XLA_OP(Name("ResourceScatterMax"), ResourceScatterMaxOp); @@ -286,7 +286,7 @@ class ResourceScatterNdAddOp : public ResourceScatterOp { private: static xla::XlaOp Combine(const xla::XlaOp& x, const xla::XlaOp& y, xla::XlaBuilder* builder) { - return builder->Add(x, y); + return xla::Add(x, y); } }; REGISTER_XLA_OP(Name("ResourceScatterNdAdd"), ResourceScatterNdAddOp); diff --git a/tensorflow/compiler/tf2xla/kernels/while_op.cc b/tensorflow/compiler/tf2xla/kernels/while_op.cc index 5467c5d9946846ff9f14ce9c5aac9e2be4b9d6ab..9413a30a6c2054203e5f1e561e4f8a7c330e060c 100644 --- a/tensorflow/compiler/tf2xla/kernels/while_op.cc +++ b/tensorflow/compiler/tf2xla/kernels/while_op.cc @@ -22,7 +22,7 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/xla_op_kernel.h" #include "tensorflow/compiler/tf2xla/xla_op_registry.h" #include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/core/framework/function.h" #include "tensorflow/core/framework/op_kernel.h" @@ -246,7 +246,7 @@ void XlaWhileOp::Compile(XlaOpKernelContext* ctx) { } } - xla::XlaOp init = builder->Tuple(inputs); + xla::XlaOp init = xla::Tuple(builder, inputs); VLOG(1) << "Building while loop"; @@ -255,22 +255,21 @@ void XlaWhileOp::Compile(XlaOpKernelContext* ctx) { { std::unique_ptr cb = builder->CreateSubBuilder("cond_wrapper"); - auto inputs = cb->Parameter(0, cond_input_shape, "inputs"); - auto outputs = cb->Call(*cond.computation, {inputs}); - cb->GetTupleElement(outputs, 0); + auto inputs = xla::Parameter(cb.get(), 0, cond_input_shape, "inputs"); + auto outputs = xla::Call(cb.get(), *cond.computation, {inputs}); + xla::GetTupleElement(outputs, 0); xla::StatusOr result = cb->Build(); OP_REQUIRES_OK(ctx, result.status()); cond_wrapper = std::move(result.ValueOrDie()); } - xla::XlaOp while_result = - builder->While(cond_wrapper, *body.computation, init); + xla::XlaOp while_result = xla::While(cond_wrapper, *body.computation, init); // Sets non-variable outputs. for (int i = 0; i < ctx->num_outputs(); ++i) { if (ctx->input_type(i) != DT_RESOURCE) { ctx->SetOutput(body.input_mapping[i], - builder->GetTupleElement(while_result, i)); + xla::GetTupleElement(while_result, i)); } } @@ -284,7 +283,7 @@ void XlaWhileOp::Compile(XlaOpKernelContext* ctx) { OP_REQUIRES_OK(ctx, resource->SetFromPack( arguments[update.input_index].tensor_array_gradients, - builder->GetTupleElement(while_result, pos), builder)); + xla::GetTupleElement(while_result, pos), builder)); } VLOG(2) << "Loop-carried variable: pos: " << update.input_index << " name: " << resource->name() << " modified: " << update.modified diff --git a/tensorflow/compiler/tf2xla/legacy_flags/backend_registration_flags.cc b/tensorflow/compiler/tf2xla/legacy_flags/backend_registration_flags.cc new file mode 100644 index 0000000000000000000000000000000000000000..661505021f820e2a87a5d414c6fe382bf6153045 --- /dev/null +++ b/tensorflow/compiler/tf2xla/legacy_flags/backend_registration_flags.cc @@ -0,0 +1,63 @@ +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +// Legacy flags for the XLA bridge's backend registration modules. + +#include // NOLINT +#include + +#include "tensorflow/compiler/tf2xla/legacy_flags/backend_registration_flags.h" +#include "tensorflow/compiler/xla/legacy_flags/parse_flags_from_env.h" +#include "tensorflow/core/platform/types.h" +#include "tensorflow/core/util/command_line_flags.h" + +namespace tensorflow { +namespace legacy_flags { + +// Pointers to the parsed value of the flags and flag descriptors, initialized +// via flags_init. +static BackendRegistrationFlags* flags; +static std::vector* flag_list; +static std::once_flag flags_init; + +// Allocate *flags. Called via call_once(&flags_init,...). +static void AllocateFlags() { + flags = new BackendRegistrationFlags; + flags->tf_enable_prng_ops_gpu = false; + flag_list = new std::vector({ + Flag("tf_enable_prng_ops_gpu", &flags->tf_enable_prng_ops_gpu, + "Whether to enable PRNG ops: [RandomStandardNormal | RandomUniform " + "| RandomUniformInt | TruncatedNormal] on GPU."), + }); + xla::legacy_flags::ParseFlagsFromEnv(*flag_list); +} + +// Append to *append_to flag definitions associated with the XLA bridge's +// backend registration modules. +void AppendBackendRegistrationFlags(std::vector* append_to) { + std::call_once(flags_init, &AllocateFlags); + append_to->insert(append_to->end(), flag_list->begin(), flag_list->end()); +} + +// Return a pointer to the BackendRegistrationFlags struct; +// repeated calls return the same pointer. +// This should be called only after Flags::Parse() has returned. +BackendRegistrationFlags* GetBackendRegistrationFlags() { + std::call_once(flags_init, &AllocateFlags); + return flags; +} + +} // namespace legacy_flags +} // namespace tensorflow diff --git a/tensorflow/compiler/tf2xla/legacy_flags/backend_registration_flags.h b/tensorflow/compiler/tf2xla/legacy_flags/backend_registration_flags.h new file mode 100644 index 0000000000000000000000000000000000000000..861c923dd51f90be2acbeb23911a93e873aabdce --- /dev/null +++ b/tensorflow/compiler/tf2xla/legacy_flags/backend_registration_flags.h @@ -0,0 +1,49 @@ +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#ifndef TENSORFLOW_COMPILER_TF2XLA_LEGACY_FLAGS_BACKEND_REGISTRATION_FLAGS_H_ +#define TENSORFLOW_COMPILER_TF2XLA_LEGACY_FLAGS_BACKEND_REGISTRATION_FLAGS_H_ + +// Legacy flags for the XLA bridge's backend registration modules. + +#include + +#include "tensorflow/core/platform/types.h" +#include "tensorflow/core/util/command_line_flags.h" + +namespace tensorflow { +namespace legacy_flags { + +// Append to *flag_list flag definitions associated with the XLA bridge's +// backend registration modules. +void AppendBackendRegistrationFlags(std::vector* append_to); + +// The values of flags associated with the XLA bridge's backend registration +// module. +typedef struct { + // Whether to enable RandomUniform op on GPU backend. + // TODO (b/32333178): Remove this flag or set its default to true. + bool tf_enable_prng_ops_gpu; +} BackendRegistrationFlags; + +// Return a pointer to the BackendRegistrationFlags struct; +// repeated calls return the same pointer. +// This should be called only after Flags::Parse() has returned. +BackendRegistrationFlags* GetBackendRegistrationFlags(); + +} // namespace legacy_flags +} // namespace tensorflow + +#endif // TENSORFLOW_COMPILER_TF2XLA_LEGACY_FLAGS_BACKEND_REGISTRATION_FLAGS_H_ diff --git a/tensorflow/compiler/tf2xla/lib/BUILD b/tensorflow/compiler/tf2xla/lib/BUILD index 04c600698c7d86808238f29cbeed6aa66acaee70..becc8b84fe614d3ffd10d1a98b143980895d0e50 100644 --- a/tensorflow/compiler/tf2xla/lib/BUILD +++ b/tensorflow/compiler/tf2xla/lib/BUILD @@ -40,10 +40,11 @@ cc_library( ":triangular_solve", ":util", ":while_loop", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:status_macros", "//tensorflow/compiler/xla:statusor", + "//tensorflow/compiler/xla/client/lib:constants", "//tensorflow/compiler/xla/client/xla_client:xla_builder", "//tensorflow/compiler/xla/client/xla_client:xla_computation", "//tensorflow/core:lib", @@ -58,12 +59,35 @@ cc_library( "//tensorflow/compiler/tf2xla:xla_compiler", "//tensorflow/compiler/xla:status_macros", "//tensorflow/compiler/xla:statusor", - "//tensorflow/compiler/xla/client/lib:arithmetic", + "//tensorflow/compiler/xla/client/lib:constants", + "//tensorflow/compiler/xla/client/lib:math", "//tensorflow/compiler/xla/client/xla_client:xla_builder", "//tensorflow/core:protos_all_cc", ], ) +cc_library( + name = "qr", + srcs = ["qr.cc"], + hdrs = ["qr.h"], + deps = [ + ":batch_dot", + ":util", + ":while_loop", + "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:shape_util", + "//tensorflow/compiler/xla:status_macros", + "//tensorflow/compiler/xla:statusor", + "//tensorflow/compiler/xla/client/lib:arithmetic", + "//tensorflow/compiler/xla/client/lib:constants", + "//tensorflow/compiler/xla/client/lib:math", + "//tensorflow/compiler/xla/client/lib:numeric", + "//tensorflow/compiler/xla/client/xla_client:xla_builder", + "//tensorflow/compiler/xla/client/xla_client:xla_computation", + "//tensorflow/core:lib", + ], +) + cc_library( name = "scatter", srcs = ["scatter.cc"], @@ -71,7 +95,7 @@ cc_library( deps = [ ":util", ":while_loop", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:status_macros", "//tensorflow/compiler/xla:statusor", @@ -90,11 +114,12 @@ cc_library( deps = [ ":batch_dot", ":util", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:status_macros", "//tensorflow/compiler/xla:statusor", "//tensorflow/compiler/xla:util", + "//tensorflow/compiler/xla/client/lib:constants", "//tensorflow/compiler/xla/client/xla_client:xla_builder", "//tensorflow/compiler/xla/client/xla_client:xla_computation", "//tensorflow/core:lib", @@ -108,7 +133,7 @@ xla_test( deps = [ ":triangular_solve", "//tensorflow/compiler/xla:array2d", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:statusor", "//tensorflow/compiler/xla:test", @@ -130,6 +155,7 @@ cc_library( srcs = ["util.cc"], hdrs = ["util.h"], deps = [ + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:literal_util", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:status_macros", @@ -148,7 +174,7 @@ xla_test( ":batch_dot", ":util", "//tensorflow/compiler/xla:array2d", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:statusor", "//tensorflow/compiler/xla:test", diff --git a/tensorflow/compiler/tf2xla/lib/batch_dot.cc b/tensorflow/compiler/tf2xla/lib/batch_dot.cc index ee0bb91a6b747ffc9e28e19dd4869a5b2cc43501..3c4eec081ba9744226cfbd8d5392220cbf7276f3 100644 --- a/tensorflow/compiler/tf2xla/lib/batch_dot.cc +++ b/tensorflow/compiler/tf2xla/lib/batch_dot.cc @@ -18,6 +18,7 @@ limitations under the License. #include #include +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/compiler/xla/shape_util.h" #include "tensorflow/compiler/xla/status_macros.h" #include "tensorflow/compiler/xla/statusor.h" @@ -25,91 +26,94 @@ limitations under the License. namespace tensorflow { -xla::StatusOr BatchDot(xla::XlaBuilder* builder, xla::XlaOp x, - xla::XlaOp y, bool transpose_x, - bool transpose_y, bool conjugate_x, - bool conjugate_y) { - TF_ASSIGN_OR_RETURN(xla::Shape x_shape, builder->GetShape(x)); - TF_ASSIGN_OR_RETURN(xla::Shape y_shape, builder->GetShape(y)); - - // Check that both tensors have the same number of dimensions. There must be - // at least two (the batch dimensions can be empty). - if (xla::ShapeUtil::Rank(x_shape) != xla::ShapeUtil::Rank(y_shape)) { - return errors::InvalidArgument( - "Arguments to BatchedDot have different ranks: ", - xla::ShapeUtil::HumanString(x_shape), " vs. ", - xla::ShapeUtil::HumanString(y_shape)); - } - const int ndims = xla::ShapeUtil::Rank(x_shape); - if (ndims < 2) { - return errors::InvalidArgument( - "Arguments to BatchedDot must have rank >= 2: ", ndims); - } - - // The batch dimensions must be equal and the matrix dimensions must be - // valid. - std::vector batch_dimension_numbers; - for (int i = 0; i < ndims - 2; ++i) { - if (x_shape.dimensions(i) != y_shape.dimensions(i)) { +xla::XlaOp BatchDot(xla::XlaOp x, xla::XlaOp y, bool transpose_x, + bool transpose_y, bool conjugate_x, bool conjugate_y) { + xla::XlaBuilder* builder = x.builder(); + return builder->ReportErrorOrReturn([&]() -> xla::StatusOr { + TF_ASSIGN_OR_RETURN(xla::Shape x_shape, builder->GetShape(x)); + TF_ASSIGN_OR_RETURN(xla::Shape y_shape, builder->GetShape(y)); + + // Check that both tensors have the same number of dimensions. There must be + // at least two (the batch dimensions can be empty). + if (xla::ShapeUtil::Rank(x_shape) != xla::ShapeUtil::Rank(y_shape)) { return errors::InvalidArgument( - "Dimension ", i, " of inputs to BatchedDot must be equal: ", - xla::ShapeUtil::HumanString(x_shape), " vs ", + "Arguments to BatchedDot have different ranks: ", + xla::ShapeUtil::HumanString(x_shape), " vs. ", xla::ShapeUtil::HumanString(y_shape)); } - batch_dimension_numbers.push_back(i); - } - - int x_inner_dim = transpose_x ? (ndims - 2) : (ndims - 1); - int y_inner_dim = transpose_y ? (ndims - 1) : (ndims - 2); - if (x_shape.dimensions(x_inner_dim) != y_shape.dimensions(y_inner_dim)) { - return errors::InvalidArgument( - "Dimensions ", x_inner_dim, " and ", y_inner_dim, - " of arguments to BatchedDot must be equal: ", - xla::ShapeUtil::HumanString(x_shape), " transpose: ", transpose_x, - " vs. ", xla::ShapeUtil::HumanString(y_shape), - " transpose: ", transpose_y); - } - - // Check for zero lhs/rhs dim size. - if (xla::ShapeUtil::IsZeroElementArray(x_shape) || - xla::ShapeUtil::IsZeroElementArray(y_shape)) { - std::vector dimensions(batch_dimension_numbers.size()); - for (int i = 0; i < batch_dimension_numbers.size(); ++i) { - dimensions[i] = x_shape.dimensions(batch_dimension_numbers[i]); + const int ndims = xla::ShapeUtil::Rank(x_shape); + if (ndims < 2) { + return errors::InvalidArgument( + "Arguments to BatchedDot must have rank >= 2: ", ndims); + } + + // The batch dimensions must be equal and the matrix dimensions must be + // valid. + std::vector batch_dimension_numbers; + for (int i = 0; i < ndims - 2; ++i) { + if (x_shape.dimensions(i) != y_shape.dimensions(i)) { + return errors::InvalidArgument( + "Dimension ", i, " of inputs to BatchedDot must be equal: ", + xla::ShapeUtil::HumanString(x_shape), " vs ", + xla::ShapeUtil::HumanString(y_shape)); + } + batch_dimension_numbers.push_back(i); + } + + int x_inner_dim = transpose_x ? (ndims - 2) : (ndims - 1); + int y_inner_dim = transpose_y ? (ndims - 1) : (ndims - 2); + if (x_shape.dimensions(x_inner_dim) != y_shape.dimensions(y_inner_dim)) { + return errors::InvalidArgument( + "Dimensions ", x_inner_dim, " and ", y_inner_dim, + " of arguments to BatchedDot must be equal: ", + xla::ShapeUtil::HumanString(x_shape), " transpose: ", transpose_x, + " vs. ", xla::ShapeUtil::HumanString(y_shape), + " transpose: ", transpose_y); + } + + // Check for zero lhs/rhs dim size. + if (xla::ShapeUtil::IsZeroElementArray(x_shape) || + xla::ShapeUtil::IsZeroElementArray(y_shape)) { + std::vector dimensions(batch_dimension_numbers.size()); + for (int i = 0; i < batch_dimension_numbers.size(); ++i) { + dimensions[i] = x_shape.dimensions(batch_dimension_numbers[i]); + } + int x_outer_dim = transpose_x ? (ndims - 1) : (ndims - 2); + int y_outer_dim = transpose_y ? (ndims - 2) : (ndims - 1); + dimensions.push_back(x_shape.dimensions(x_outer_dim)); + dimensions.push_back(y_shape.dimensions(y_outer_dim)); + return xla::Broadcast( + xla::ConstantLiteral(builder, + xla::LiteralUtil::Zero(x_shape.element_type())), + dimensions); + } + + if (x_shape.element_type() == xla::C64 && conjugate_x) { + x = xla::Conj(x); + } + if (y_shape.element_type() == xla::C64 && conjugate_y) { + y = xla::Conj(y); + } + + // If there are no batch dimensions, use a regular Dot. + // TODO(b/69062148) Remove this code when Dot emitters can be passed + // dimensions to transpose directly (i.e. without requiring a Transpose + // HLO). + if (batch_dimension_numbers.empty()) { + auto lhs = transpose_x ? xla::Transpose(x, {1, 0}) : x; + auto rhs = transpose_y ? xla::Transpose(y, {1, 0}) : y; + return xla::Dot(lhs, rhs); + } + + xla::DotDimensionNumbers dot_dnums; + dot_dnums.add_lhs_contracting_dimensions(x_inner_dim); + dot_dnums.add_rhs_contracting_dimensions(y_inner_dim); + for (auto batch_dimension_number : batch_dimension_numbers) { + dot_dnums.add_lhs_batch_dimensions(batch_dimension_number); + dot_dnums.add_rhs_batch_dimensions(batch_dimension_number); } - int x_outer_dim = transpose_x ? (ndims - 1) : (ndims - 2); - int y_outer_dim = transpose_y ? (ndims - 2) : (ndims - 1); - dimensions.push_back(x_shape.dimensions(x_outer_dim)); - dimensions.push_back(y_shape.dimensions(y_outer_dim)); - return builder->Broadcast( - builder->ConstantLiteral(xla::Literal::Zero(x_shape.element_type())), - dimensions); - } - - if (x_shape.element_type() == xla::C64 && conjugate_x) { - x = builder->Conj(x); - } - if (y_shape.element_type() == xla::C64 && conjugate_y) { - y = builder->Conj(y); - } - - // If there are no batch dimensions, use a regular Dot. - // TODO(b/69062148) Remove this code when Dot emitters can be passed - // dimensions to transpose directly (i.e. without requiring a Transpose HLO). - if (batch_dimension_numbers.empty()) { - auto lhs = transpose_x ? builder->Transpose(x, {1, 0}) : x; - auto rhs = transpose_y ? builder->Transpose(y, {1, 0}) : y; - return builder->Dot(lhs, rhs); - } - - xla::DotDimensionNumbers dot_dnums; - dot_dnums.add_lhs_contracting_dimensions(x_inner_dim); - dot_dnums.add_rhs_contracting_dimensions(y_inner_dim); - for (auto batch_dimension_number : batch_dimension_numbers) { - dot_dnums.add_lhs_batch_dimensions(batch_dimension_number); - dot_dnums.add_rhs_batch_dimensions(batch_dimension_number); - } - return builder->DotGeneral(x, y, dot_dnums); + return xla::DotGeneral(x, y, dot_dnums); + }); } } // namespace tensorflow diff --git a/tensorflow/compiler/tf2xla/lib/batch_dot.h b/tensorflow/compiler/tf2xla/lib/batch_dot.h index 1acc72033b05e73b0f5f88907df20cde5cfffbf0..d07a9486f18c0b8f26782123a8fba4ba228f71ee 100644 --- a/tensorflow/compiler/tf2xla/lib/batch_dot.h +++ b/tensorflow/compiler/tf2xla/lib/batch_dot.h @@ -43,10 +43,9 @@ namespace tensorflow { // It is computed as: // // output[..., :, :] = matrix(x[..., :, :]) * matrix(y[..., :, :]) -xla::StatusOr BatchDot(xla::XlaBuilder* builder, xla::XlaOp x, - xla::XlaOp y, bool transpose_x, - bool transpose_y, bool conjugate_x = false, - bool conjugate_y = false); +xla::XlaOp BatchDot(xla::XlaOp x, xla::XlaOp y, bool transpose_x = false, + bool transpose_y = false, bool conjugate_x = false, + bool conjugate_y = false); } // namespace tensorflow diff --git a/tensorflow/compiler/tf2xla/lib/cholesky.cc b/tensorflow/compiler/tf2xla/lib/cholesky.cc index 20925118bf598a6436c43bd727ce40e3abafc46c..35b137aa2cc0b5e6c2d2b917c0a95410522305c2 100644 --- a/tensorflow/compiler/tf2xla/lib/cholesky.cc +++ b/tensorflow/compiler/tf2xla/lib/cholesky.cc @@ -22,7 +22,9 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/lib/triangular_solve.h" #include "tensorflow/compiler/tf2xla/lib/util.h" #include "tensorflow/compiler/tf2xla/lib/while_loop.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/client/lib/constants.h" +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/shape_util.h" #include "tensorflow/compiler/xla/status_macros.h" #include "tensorflow/compiler/xla/statusor.h" @@ -47,178 +49,163 @@ namespace { // l[..., j+1:, j] = (a[..., j+1:, j] - np.dot(l[..., j+1:, :j], row_t)) / // l[..., j, j] // return l -xla::StatusOr CholeskyUnblocked(xla::XlaBuilder* builder, - const xla::XlaOp& a) { - TF_ASSIGN_OR_RETURN(xla::Shape a_shape, builder->GetShape(a)); - const int n_dims = xla::ShapeUtil::Rank(a_shape); - const int64 n = xla::ShapeUtil::GetDimension(a_shape, -1); - gtl::ArraySlice major_dims(xla::AsInt64Slice(a_shape.dimensions()), - /*pos=*/0, - /*len=*/n_dims - 2); - - xla::XlaOp l = Zeros(builder, a_shape); - - // Construct the for loop body to iterate over rows. - auto body_fn = [&](xla::XlaOp i, gtl::ArraySlice loop_vars, - xla::XlaBuilder* body_builder) - -> xla::StatusOr> { - xla::Shape col_shape; - xla::Shape row_shape; - for (int64 d : major_dims) { - row_shape.add_dimensions(d); - col_shape.add_dimensions(d); - } - row_shape.add_dimensions(1); - row_shape.add_dimensions(n); - row_shape.set_element_type(a_shape.element_type()); - auto mask_zeros_row = Zeros(body_builder, row_shape); - - col_shape.add_dimensions(n); - col_shape.add_dimensions(1); - col_shape.set_element_type(a_shape.element_type()); - auto mask_zeros_col = Zeros(body_builder, col_shape); - - std::vector mask_vector(n); - std::iota(mask_vector.begin(), mask_vector.end(), 0); - auto mask_range = body_builder->ConstantR1(mask_vector); - auto mask_range_row = body_builder->Broadcast( - body_builder->Reshape(mask_range, {0}, {1, n}), major_dims); - auto mask_range_col = body_builder->Broadcast( - body_builder->Reshape(mask_range, {0}, {n, 1}), major_dims); - auto body_a = loop_vars[0]; - auto body_l = loop_vars[1]; - - // row = l[..., i, :i] - // select the whole i-th row, then mask out all columns past i-1 - auto zero = body_builder->ConstantR0(0); - TF_ASSIGN_OR_RETURN(auto l_i, DynamicSliceInMinorDims(body_builder, body_l, - {i, zero}, {1, n})); - auto row = body_builder->Select(body_builder->Ge(mask_range_row, i), - mask_zeros_row, l_i); - // a[..., i, i] - TF_ASSIGN_OR_RETURN(auto a_ii, DynamicSliceInMinorDims(body_builder, body_a, - {i, i}, {1, 1})); - // np.dot(row, np.swapaxes(row, -1, -2)) - xla::XlaOp diag_dot; - TF_ASSIGN_OR_RETURN(diag_dot, BatchDot(body_builder, row, row, - /*transpose_x=*/false, - /*transpose_y=*/true)); - // l[..., i, i] = np.sqrt(a[..., i, i] - np.dot(row, - // np.swapaxes(row, -1, -2))) - auto l_ii = body_builder->Pow( - body_builder->Sub(a_ii, diag_dot), - FloatLiteral(body_builder, a_shape.element_type(), 0.5)); - - // a[..., i+1:, i] - // select the whole i-th column, then mask out all rows above i+1 +xla::XlaOp CholeskyUnblocked(xla::XlaOp a) { + xla::XlaBuilder* builder = a.builder(); + return builder->ReportErrorOrReturn([&]() -> xla::StatusOr { + TF_ASSIGN_OR_RETURN(xla::Shape a_shape, builder->GetShape(a)); + const int n_dims = xla::ShapeUtil::Rank(a_shape); + const int64 n = xla::ShapeUtil::GetDimension(a_shape, -1); + gtl::ArraySlice major_dims(xla::AsInt64Slice(a_shape.dimensions()), + /*pos=*/0, + /*len=*/n_dims - 2); + + xla::XlaOp l = xla::ZerosLike(a); + + // Construct the for loop body to iterate over rows. + auto body_fn = [&](xla::XlaOp i, gtl::ArraySlice loop_vars, + xla::XlaBuilder* body_builder) + -> xla::StatusOr> { + xla::Shape col_shape; + xla::Shape row_shape; + for (int64 d : major_dims) { + row_shape.add_dimensions(d); + col_shape.add_dimensions(d); + } + row_shape.add_dimensions(1); + row_shape.add_dimensions(n); + row_shape.set_element_type(a_shape.element_type()); + auto mask_zeros_row = xla::Zeros(body_builder, row_shape); + + col_shape.add_dimensions(n); + col_shape.add_dimensions(1); + col_shape.set_element_type(a_shape.element_type()); + auto mask_zeros_col = xla::Zeros(body_builder, col_shape); + + std::vector mask_vector(n); + std::iota(mask_vector.begin(), mask_vector.end(), 0); + auto mask_range = xla::ConstantR1(body_builder, mask_vector); + auto mask_range_row = + xla::Broadcast(xla::Reshape(mask_range, {0}, {1, n}), major_dims); + auto mask_range_col = + xla::Broadcast(xla::Reshape(mask_range, {0}, {n, 1}), major_dims); + auto body_a = loop_vars[0]; + auto body_l = loop_vars[1]; + + // row = l[..., i, :i] + // select the whole i-th row, then mask out all columns past i-1 + auto zero = xla::ConstantR0(body_builder, 0); + auto l_i = DynamicSliceInMinorDims(body_l, {i, zero}, {1, n}); + auto row = xla::Select(xla::Ge(mask_range_row, i), mask_zeros_row, l_i); + // a[..., i, i] + auto a_ii = DynamicSliceInMinorDims(body_a, {i, i}, {1, 1}); + // np.dot(row, np.swapaxes(row, -1, -2)) + auto diag_dot = BatchDot(row, row, + /*transpose_x=*/false, + /*transpose_y=*/true); + // l[..., i, i] = np.sqrt(a[..., i, i] - np.dot(row, + // np.swapaxes(row, -1, -2))) + auto l_ii = + xla::Pow(a_ii - diag_dot, + FloatLiteral(body_builder, a_shape.element_type(), 0.5)); + + // a[..., i+1:, i] + // select the whole i-th column, then mask out all rows above i+1 + auto a_0i = DynamicSliceInMinorDims(body_a, {i}, {1}); + auto a_ip1i = + xla::Select(xla::Le(mask_range_col, i), mask_zeros_col, a_0i); + + // l[..., i+1:, i] = (a[..., i+1:, i] - np.dot(l[..., i+1:, :i], r.T)) / + // l[..., i, i] + // The columns in [i, n] are zeroed out in `row`, so we just have to + // zero out rows above i+1 after the BatchDot. np.dot(l[..., :, :i], + // r.T) + auto dot = BatchDot(body_l, row, + /*transpose_x=*/false, + /*transpose_y=*/true); + // np.dot(l[..., i+1:, :i], r.T) + auto dot_ip1 = + xla::Select(xla::Le(mask_range_col, i), mask_zeros_col, dot); + + body_l = + DynamicUpdateSliceInMinorDims(body_l, (a_ip1i - dot_ip1) / l_ii, {i}); + // Assign the diagonal after the rest of the column because otherwise the + // column assign will wrap around and overwrite the diagonal assign. + body_l = DynamicUpdateSliceInMinorDims(body_l, l_ii, {i, i}); + + return std::vector{body_a, body_l}; + }; + TF_ASSIGN_OR_RETURN( - auto a_0i, DynamicSliceInMinorDims(body_builder, body_a, {i}, {1})); - auto a_ip1i = body_builder->Select(body_builder->Le(mask_range_col, i), - mask_zeros_col, a_0i); - - // l[..., i+1:, i] = (a[..., i+1:, i] - np.dot(l[..., i+1:, :i], r.T)) / - // l[..., i, i] - // The columns in [i, n] are zeroed out in `row`, so we just have to - // zero out rows above i+1 after the BatchDot. np.dot(l[..., :, :i], - // r.T) - TF_ASSIGN_OR_RETURN(auto dot, BatchDot(body_builder, body_l, row, - /*transpose_x=*/false, - /*transpose_y=*/true)); - // np.dot(l[..., i+1:, :i], r.T) - auto dot_ip1 = body_builder->Select(body_builder->Le(mask_range_col, i), - mask_zeros_col, dot); - - auto col_update = - body_builder->Div(body_builder->Sub(a_ip1i, dot_ip1), l_ii); - TF_ASSIGN_OR_RETURN(body_l, DynamicUpdateSliceInMinorDims( - body_builder, body_l, col_update, {i})); - // Assign the diagonal after the rest of the column because otherwise the - // column assign will wrap around and overwrite the diagonal assign. - TF_ASSIGN_OR_RETURN(body_l, DynamicUpdateSliceInMinorDims( - body_builder, body_l, l_ii, {i, i})); - - return std::vector{body_a, body_l}; - }; - - TF_ASSIGN_OR_RETURN( - auto cholesky_while, - XlaForEachIndex(n, xla::S32, body_fn, {a, l}, "unblocked", builder)); - - return cholesky_while[1]; + auto cholesky_while, + XlaForEachIndex(n, xla::S32, body_fn, {a, l}, "unblocked", builder)); + + return cholesky_while[1]; + }); } } // namespace -xla::StatusOr Cholesky(xla::XlaBuilder* builder, xla::XlaOp a, - int64 block_size) { - TF_ASSIGN_OR_RETURN(xla::Shape a_shape, builder->GetShape(a)); - const int ndims = xla::ShapeUtil::Rank(a_shape); - if (ndims < 2) { - return errors::InvalidArgument( - "Arguments to Cholesky must have rank >= 2: ", ndims); - } - - const int64 n = xla::ShapeUtil::GetDimension(a_shape, -1); - if (n != xla::ShapeUtil::GetDimension(a_shape, -2)) { - return errors::InvalidArgument( - "Arguments to Cholesky must be square matrices: ", - xla::ShapeUtil::HumanString(a_shape)); - } - - if (block_size < 1) { - return errors::InvalidArgument( - "block_size argument to Cholesky must be >= 1; got ", block_size); - } - - // Blocked left-looking Cholesky factorization. - // Algorithm 1 from - // Haidar, Azzam, et al. "High-performance Cholesky factorization for GPU-only - // execution." Proceedings of General Purpose GPUs. ACM, 2017. - xla::XlaOp l = Zeros(builder, a_shape); - for (int64 i = 0; i < n; i += block_size) { - int64 k = std::min(block_size, n - i); - if (i > 0) { - // TODO(phawkins): consider implementing SYRK for the diagonal part of - // the panel. - // a[i:, i:i+k] -= np.dot(l[i:, :i], np.transpose(l[i:i+k, :i])) - TF_ASSIGN_OR_RETURN(auto lhs, - SliceInMinorDims(builder, l, {i, 0}, {n, i})); - TF_ASSIGN_OR_RETURN(auto rhs, - SliceInMinorDims(builder, l, {i, 0}, {i + k, i})); - TF_ASSIGN_OR_RETURN(auto delta, - BatchDot(builder, lhs, rhs, /*transpose_x=*/false, - /*transpose_y=*/true, /*conjugate_x=*/false, - /*conjugate_y=*/false)); - TF_ASSIGN_OR_RETURN(auto before, - SliceInMinorDims(builder, a, {i, i}, {n, i + k})); - TF_ASSIGN_OR_RETURN( - a, UpdateSliceInMinorDims(builder, a, builder->Sub(before, delta), - {i, i})); +xla::XlaOp Cholesky(xla::XlaOp a, int64 block_size) { + xla::XlaBuilder* builder = a.builder(); + return builder->ReportErrorOrReturn([&]() -> xla::StatusOr { + TF_ASSIGN_OR_RETURN(xla::Shape a_shape, builder->GetShape(a)); + const int ndims = xla::ShapeUtil::Rank(a_shape); + if (ndims < 2) { + return errors::InvalidArgument( + "Arguments to Cholesky must have rank >= 2: ", ndims); + } + + const int64 n = xla::ShapeUtil::GetDimension(a_shape, -1); + if (n != xla::ShapeUtil::GetDimension(a_shape, -2)) { + return errors::InvalidArgument( + "Arguments to Cholesky must be square matrices: ", + xla::ShapeUtil::HumanString(a_shape)); + } + + if (block_size < 1) { + return errors::InvalidArgument( + "block_size argument to Cholesky must be >= 1; got ", block_size); } - // l[i:i+k, i:i+k] = cholesky_unblocked(a[i:i+k, i:i+k]) - TF_ASSIGN_OR_RETURN(auto x, - SliceInMinorDims(builder, a, {i, i}, {i + k, i + k})); - TF_ASSIGN_OR_RETURN(auto factorized, CholeskyUnblocked(builder, x)); - TF_ASSIGN_OR_RETURN(l, - UpdateSliceInMinorDims(builder, l, factorized, {i, i})); - - if (i + k < n) { - // l[i+k:, i:i+k] = trsm_right_transpose(l[i:i+k, i:i+k], a[i+k:, i:i+k]) - TF_ASSIGN_OR_RETURN(auto panel, - SliceInMinorDims(builder, a, {i + k, i}, {n, i + k})); - TF_ASSIGN_OR_RETURN(auto update, - TriangularSolve(builder, factorized, panel, - /*left_side=*/false, - /*lower=*/true, - /*transpose_a=*/true, - /*conjugate_a=*/false, - /*block_size=*/block_size)); - TF_ASSIGN_OR_RETURN( - l, UpdateSliceInMinorDims(builder, l, update, {i + k, i})); + // Blocked left-looking Cholesky factorization. + // Algorithm 1 from + // Haidar, Azzam, et al. "High-performance Cholesky factorization for + // GPU-only execution." Proceedings of General Purpose GPUs. ACM, 2017. + xla::XlaOp l = xla::ZerosLike(a); + for (int64 i = 0; i < n; i += block_size) { + int64 k = std::min(block_size, n - i); + if (i > 0) { + // TODO(phawkins): consider implementing SYRK for the diagonal part of + // the panel. + // a[i:, i:i+k] -= np.dot(l[i:, :i], np.transpose(l[i:i+k, :i])) + auto lhs = SliceInMinorDims(l, {i, 0}, {n, i}); + auto rhs = SliceInMinorDims(l, {i, 0}, {i + k, i}); + auto delta = BatchDot(lhs, rhs, /*transpose_x=*/false, + /*transpose_y=*/true); + auto before = SliceInMinorDims(a, {i, i}, {n, i + k}); + a = UpdateSliceInMinorDims(a, before - delta, {i, i}); + } + + // l[i:i+k, i:i+k] = cholesky_unblocked(a[i:i+k, i:i+k]) + auto x = SliceInMinorDims(a, {i, i}, {i + k, i + k}); + auto factorized = CholeskyUnblocked(x); + l = UpdateSliceInMinorDims(l, factorized, {i, i}); + + if (i + k < n) { + // l[i+k:, i:i+k] = + // trsm_right_transpose(l[i:i+k, i:i+k], a[i+k:, i:i+k]) + auto panel = SliceInMinorDims(a, {i + k, i}, {n, i + k}); + auto update = TriangularSolve(factorized, panel, + /*left_side=*/false, + /*lower=*/true, + /*transpose_a=*/true, + /*conjugate_a=*/false, + /*block_size=*/block_size); + l = UpdateSliceInMinorDims(l, update, {i + k, i}); + } } - } - return l; + return l; + }); } } // namespace tensorflow diff --git a/tensorflow/compiler/tf2xla/lib/cholesky.h b/tensorflow/compiler/tf2xla/lib/cholesky.h index 20fca7969ece2729a44933fd3ef3f87230ab6cad..0f6e0e9d152ec5daedeb9c0e355bfb9731759094 100644 --- a/tensorflow/compiler/tf2xla/lib/cholesky.h +++ b/tensorflow/compiler/tf2xla/lib/cholesky.h @@ -30,8 +30,7 @@ namespace tensorflow { // TODO(phawkins): check for negative values on the diagonal and return an // error, instead of silently yielding NaNs. // TODO(znado): handle the complex Hermitian case -xla::StatusOr Cholesky(xla::XlaBuilder* builder, xla::XlaOp a, - int64 block_size = 256); +xla::XlaOp Cholesky(xla::XlaOp a, int64 block_size = 256); } // namespace tensorflow diff --git a/tensorflow/compiler/tf2xla/lib/qr.cc b/tensorflow/compiler/tf2xla/lib/qr.cc new file mode 100644 index 0000000000000000000000000000000000000000..9c8ac7af25e4222f35bedd3816fc817af7e1f068 --- /dev/null +++ b/tensorflow/compiler/tf2xla/lib/qr.cc @@ -0,0 +1,387 @@ +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow/compiler/tf2xla/lib/qr.h" + +#include +#include + +#include "tensorflow/compiler/tf2xla/lib/batch_dot.h" +#include "tensorflow/compiler/tf2xla/lib/util.h" +#include "tensorflow/compiler/tf2xla/lib/while_loop.h" +#include "tensorflow/compiler/xla/client/lib/arithmetic.h" +#include "tensorflow/compiler/xla/client/lib/constants.h" +#include "tensorflow/compiler/xla/client/lib/math.h" +#include "tensorflow/compiler/xla/client/lib/numeric.h" +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" +#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/shape_util.h" +#include "tensorflow/compiler/xla/status_macros.h" +#include "tensorflow/compiler/xla/statusor.h" +#include "tensorflow/core/lib/core/errors.h" + +namespace tensorflow { + +namespace { + +// Computes a Householder reflection of the form: +// H = I - tau v v.T. +// such that +// H . ( x1 ) = ( x1 ) +// ( x2 ) = ( x2 ) +// ( ... ) = ( ... ) +// ( xk ) = ( beta ) +// ( ... ) ( 0 ) +// ( ... ) ( 0 ) +// Unlike the usual formulation, we allow the caller to supply 'k' rather than +// only providing the relevant part of 'x' to maintain XLA's static shape +// invariant. In addition, the implementation supports batching. +// Pseudo-code, without batching: +// alpha = x[k] +// x_copy = np.copy(x) +// x_copy[:k+1] = 0 +// xnorm = norm2(x_copy) +// if xnorm == 0: +// beta = alpha +// tau = 0 +// v = np.zeros_like(x) +// else: +// beta = - np.sign(alpha) * dlapy2(alpha, xnorm) +// tau = (beta - alpha) / beta +// v = x / (alpha - beta) +// v[k] = 1 +// return (v, tau, beta) +// TODO(phawkins): LAPACK's xLARFG implementation has code for handling +// overflows in the norm/beta calculations. Perhaps do the same here. +xla::Status House(xla::XlaOp x, xla::XlaOp k, gtl::ArraySlice batch_dims, + const int64 m, xla::XlaOp* v, xla::XlaOp* tau, + xla::XlaOp* beta) { + xla::XlaBuilder* const builder = x.builder(); + TF_ASSIGN_OR_RETURN(xla::Shape x_shape, builder->GetShape(x)); + const xla::PrimitiveType type = x_shape.element_type(); + + std::vector batch_dim_ids(batch_dims.size()); + std::iota(batch_dim_ids.begin(), batch_dim_ids.end(), 0); + const int64 minor_dim = batch_dims.size(); + + xla::XlaOp zero = xla::ScalarLike(x, 0.0); + xla::XlaOp one = xla::ScalarLike(x, 1.0); + + // alpha = x[k] + xla::XlaOp alpha = + xla::Reshape(DynamicSliceInMinorDims(x, {k}, {1}), batch_dims); + + // Compute x[k+1:] (padded with zeros in elements 0..k) + xla::XlaOp iota = xla::Iota(builder, xla::S32, m); + xla::XlaOp x_after_k = + xla::Mul(x, xla::ConvertElementType(xla::Gt(iota, k), type), + /*broadcast_dimensions=*/{minor_dim}); + + // sigma = np.dot(x[k+1:], x[k+1:]) + auto sigma = + xla::Reduce(x_after_k * x_after_k, zero, + xla::CreateScalarAddComputation(type, builder), {minor_dim}); + // mu = np.sqrt(x[k]*x[k] + sigma) + auto mu = xla::Sqrt(xla::Square(alpha) + sigma); + + auto sigma_is_zero = xla::Eq(sigma, zero); + + *beta = xla::Select(sigma_is_zero, alpha, -xla::Sign(alpha) * mu); + *tau = xla::Select(sigma_is_zero, xla::Broadcast(zero, batch_dims), + (*beta - alpha) / *beta); + auto divisor = xla::Select(sigma_is_zero, xla::Broadcast(one, batch_dims), + alpha - *beta); + + auto e_k = xla::Broadcast(xla::ConvertElementType(xla::Eq(iota, k), type), + std::vector(batch_dims.size(), 1)); + + // Form v as [0, 0, ..., 1] ++ x[k+1:] / divisor + // If sigma is zero, x[k+1:] is zero, so use any non-zero divisor. + *v = e_k + + xla::Div(x_after_k, divisor, /*broadcast_dimensions=*/batch_dim_ids); + return Status::OK(); +} + +// Householder QR decomposition. Algorithm 5.2.1 from Golub and Van +// Loan "Matrix Computations", 4th Edition. This is an unblocked implementation +// used as an inner routine of the blocked implementation. +// Algorithm is adapted slightly so the shapes inside the loop are static, at +// the cost of some redundant computation. Since this is used as an inner block +// kernel, accumulates the Householder transformations (vs, taus) rather than +// the matrix q. +// Equivalent Python code, without batching: +// def qr(a): +// m = a.shape[0] +// n = a.shape[1] +// vs = np.zeros([m, n]) +// taus = np.zeros([n]) +// for j in xrange(min(m, n)): +// v, tau, beta = house(a[:, j], j) +// # Unusually, we apply the Householder transformation to the entirety of +// # a, wasting FLOPs to maintain the static shape invariant that XLA +// # requires. For columns that precede j this has no effect. +// a[:, :] -= tau * np.dot(v[:, np.newaxis], +// np.dot(v[np.newaxis, :], a[:, :])) +// # Form column j explicitly rather than relying on the precision of the +// # Householder update. +// a[j, j] = beta +// a[j+1:, j] = np.zeros([m - j - 1], dtype=a.dtype) +// vs[:, j] = v +// taus[j] = tau +// return (q, vs, taus) +struct QRBlockResult { + // The factored R value + xla::XlaOp r; + + // Representation of the Householder matrices I - beta v v.T + xla::XlaOp taus; // Shape: [..., n] + xla::XlaOp vs; // Shape: [..., m, n] +}; +xla::StatusOr QRBlock(xla::XlaOp a) { + xla::XlaBuilder* builder = a.builder(); + TF_ASSIGN_OR_RETURN(xla::Shape a_shape, builder->GetShape(a)); + const int num_dims = xla::ShapeUtil::Rank(a_shape); + if (num_dims < 2) { + return errors::InvalidArgument("Arguments to QR must have rank >= 2: ", + num_dims); + } + xla::PrimitiveType type = a_shape.element_type(); + + const int64 m = xla::ShapeUtil::GetDimension(a_shape, -2); + const int64 n = xla::ShapeUtil::GetDimension(a_shape, -1); + + const int64 num_batch_dims = num_dims - 2; + std::vector batch_dims(num_batch_dims); + for (int i = 0; i < num_batch_dims; ++i) { + batch_dims[i] = xla::ShapeUtil::GetDimension(a_shape, i); + } + + std::vector batch_dim_indices(num_batch_dims); + std::iota(batch_dim_indices.begin(), batch_dim_indices.end(), 0); + + auto qr_body_fn = + [&](xla::XlaOp j, gtl::ArraySlice values, + xla::XlaBuilder* builder) -> xla::StatusOr> { + auto a = values[0]; + auto vs = values[1]; + auto taus = values[2]; + + // v, beta = house(a[:, j], j) + auto x = DynamicSliceInMinorDims(a, {j}, {1}); + xla::XlaOp v, tau, beta; + TF_RETURN_IF_ERROR(House(xla::Collapse(x, {num_dims - 2, num_dims - 1}), j, + batch_dims, m, &v, &tau, &beta)); + + std::vector shape = batch_dims; + shape.push_back(1); + shape.push_back(m); + auto v_broadcast = xla::Reshape(v, shape); + // a[:, :] -= tau * np.dot(v[:, np.newaxis], + // np.dot(v[np.newaxis, :], a[:, :])) + auto vva = BatchDot(v_broadcast, a); + vva = BatchDot(v_broadcast, vva, /*transpose_x=*/true); + a = a - xla::Mul(tau, vva, + /*broadcast_dimensions=*/batch_dim_indices); + + // It is more precise to populate column 'k' explicitly, rather than + // computing it implicitly by applying the Householder transformation. + // a[k,k] = beta + // a[k+1:,k] = np.zeros([m-k-1], dtype=a.dtype) + auto iota = xla::Reshape(xla::Iota(a.builder(), xla::S32, m), {m, 1}); + auto predecessor_mask = xla::ConvertElementType(xla::Lt(iota, j), type); + auto mask = xla::Broadcast(xla::ConvertElementType(xla::Eq(iota, j), type), + std::vector(batch_dims.size(), 1)); + auto new_x = + xla::Mul(x, predecessor_mask, + /*broadcast_dimensions=*/{num_dims - 2, num_dims - 1}) + + xla::Mul(beta, mask, /*broadcast_dimensions=*/batch_dim_indices); + a = DynamicUpdateSliceInMinorDims(a, new_x, {j}); + + // vs[:, j] = v + vs = DynamicUpdateSliceInMinorDims( + vs, xla::Reshape(v, ConcatVectors(batch_dims, {m, 1})), {j}); + // taus[j] = tau + taus = DynamicUpdateSliceInMinorDims( + taus, xla::Reshape(tau, ConcatVectors(batch_dims, {1})), {j}); + return std::vector{a, vs, taus}; + }; + + auto vs = xla::Zeros(builder, xla::ShapeUtil::MakeShape( + type, ConcatVectors(batch_dims, {m, n}))); + auto taus = xla::Zeros( + builder, xla::ShapeUtil::MakeShape(type, ConcatVectors(batch_dims, {n}))); + + TF_ASSIGN_OR_RETURN(auto values, + XlaForEachIndex(std::min(m, n), xla::S32, qr_body_fn, + {a, vs, taus}, "qr", builder)); + + QRBlockResult result; + result.r = values[0]; + result.vs = values[1]; + result.taus = values[2]; + return result; +} + +// Computes W and Y such that I-WY is equivalent to the sequence of Householder +// transformations given by vs and taus. +// Golub and van Loan, "Matrix Computations", algorithm 5.1.2. +// Y = np.zeros([m, n]) +// W = np.zeros([m, n]) +// Y[:, 0] = vs[:, 0] +// W[:, 0] = -taus[0] * vs[:, 0] +// for j in xrange(1, n): +// v = vs[:, j] +// z = -taus[j] * v - taus[j] * np.dot(W, np.dot(Y.T, v)) +// W[:, j] = z +// Y[:, j] = v +// return W +// There is no need to return Y since at termination of the loop it is equal to +// vs. +xla::StatusOr ComputeWYRepresentation( + xla::PrimitiveType type, gtl::ArraySlice batch_dims, xla::XlaOp vs, + xla::XlaOp taus, int64 m, int64 n) { + std::vector batch_dim_indices(batch_dims.size()); + std::iota(batch_dim_indices.begin(), batch_dim_indices.end(), 0); + int64 n_index = batch_dims.size() + 1; + + auto body_fn = + [&](xla::XlaOp j, gtl::ArraySlice values, + xla::XlaBuilder* builder) -> xla::StatusOr> { + auto w = values[0]; + auto y = values[1]; + const auto vs = values[2]; + const auto taus = values[3]; + + // Want j values in range [1, ... n). + j = j + xla::ConstantR0(builder, 1); + // vs has shape [..., m, 1] + auto v = DynamicSliceInMinorDims(vs, {j}, {1}); + // beta has shape [..., 1] + auto beta = DynamicSliceInMinorDims(taus, {j}, {1}); + + // yv has shape [..., n, 1] + auto yv = BatchDot(y, v, /*transpose_x=*/true); + // wyv has shape [..., m, 1] + auto wyv = BatchDot(w, yv); + + auto z = xla::Mul( + -beta, v + wyv, + /*broadcast_dimensions=*/ConcatVectors(batch_dim_indices, {n_index})); + + w = DynamicUpdateSliceInMinorDims(w, z, {j}); + y = DynamicUpdateSliceInMinorDims(y, v, {j}); + + return std::vector{w, y, vs, taus}; + }; + + xla::XlaBuilder* builder = vs.builder(); + auto w = xla::Zeros(builder, xla::ShapeUtil::MakeShape( + type, ConcatVectors(batch_dims, {m, n}))); + auto y = w; + auto v = SliceInMinorDims(vs, {0}, {1}); + auto beta = SliceInMinorDims(taus, {0}, {1}); + y = UpdateSliceInMinorDims(y, v, {0}); + auto bv = xla::Mul( + -beta, v, + /*broadcast_dimensions=*/ConcatVectors(batch_dim_indices, {n_index})); + w = UpdateSliceInMinorDims(w, bv, {0}); + + TF_ASSIGN_OR_RETURN( + auto values, XlaForEachIndex(n - 1, xla::S32, body_fn, {w, y, vs, taus}, + "wy", builder)); + return values[0]; +} + +} // namespace + +// Block Householder QR Factorization. Algorithm 5.2.2 of Golub and van Loan. +// def qr_blocked(a, block_size): +// m = a.shape[0] +// n = a.shape[1] +// q = np.eye(m) +// for i in xrange(0, min(m, n), block_size): +// k = min(block_size, min(m, n) - s) +// (a, vs, taus) = qr(a[i:, i:i+k]) +// y = vs +// w = ComputeWYRepresentation(vs, taus, m-i, k) +// a[i:, i+r:] += np.dot(y, np.dot(w.T, a[i:, i+k:])) +// q[:, i:] += np.dot(q[:, i:], np.dot(w, y.T)) +// return (q, a) +// TODO(phawkins): consider using UT transformations (in the form I - V U V') +// rather than WY transformations. +xla::StatusOr QRDecomposition(xla::XlaOp a, + int64 block_size) { + xla::XlaBuilder* builder = a.builder(); + TF_ASSIGN_OR_RETURN(xla::Shape a_shape, builder->GetShape(a)); + const int num_dims = xla::ShapeUtil::Rank(a_shape); + if (num_dims < 2) { + return errors::InvalidArgument("Arguments to QR must have rank >= 2: ", + num_dims); + } + xla::PrimitiveType type = a_shape.element_type(); + + const int64 m = xla::ShapeUtil::GetDimension(a_shape, -2); + const int64 n = xla::ShapeUtil::GetDimension(a_shape, -1); + const int64 p = std::min(m, n); + + if (block_size < 1) { + return errors::InvalidArgument( + "block_size argument to QR must be >= 1; got ", block_size); + } + + const int64 num_batch_dims = num_dims - 2; + std::vector batch_dims(num_batch_dims); + for (int i = 0; i < num_batch_dims; ++i) { + batch_dims[i] = xla::ShapeUtil::GetDimension(a_shape, i); + } + + auto q = xla::Broadcast(xla::IdentityMatrix(builder, type, m, m), batch_dims); + for (int64 i = 0; i < p; i += block_size) { + int64 k = std::min(block_size, p - i); + + auto a_block = SliceInMinorDims(a, {i, i}, {m, i + k}); + TF_ASSIGN_OR_RETURN(auto qr_block, QRBlock(a_block)); + + a = UpdateSliceInMinorDims(a, qr_block.r, {i, i}); + + // Compute the I-WY block representation of a product of Householder + // matrices. + TF_ASSIGN_OR_RETURN(auto w, + ComputeWYRepresentation(type, batch_dims, qr_block.vs, + qr_block.taus, m - i, k)); + auto y = qr_block.vs; + + // a[i:, i+k:] += np.dot(Y, np.dot(W.T, a[i:, i+k:])) + auto a_panel = SliceInMinorDims(a, {i, i + k}, {m, n}); + auto a_update = BatchDot(w, a_panel, /*transpose_x=*/true); + a_update = BatchDot(y, a_update); + a_panel = a_panel + a_update; + a = UpdateSliceInMinorDims(a, a_panel, {i, i + k}); + + // q[:, i:] += np.dot(np.dot(q[:, i:], W), Y.T)) + auto q_panel = SliceInMinorDims(q, {0, i}, {m, m}); + auto q_update = BatchDot(q_panel, w); + q_update = + BatchDot(q_update, y, /*transpose_x=*/false, /*transpose_y=*/true); + q_panel = q_panel + q_update; + q = UpdateSliceInMinorDims(q, q_panel, {0, i}); + } + QRDecompositionResult result; + result.q = q; + result.r = a; + return result; +} + +} // namespace tensorflow diff --git a/tensorflow/compiler/tf2xla/lib/qr.h b/tensorflow/compiler/tf2xla/lib/qr.h new file mode 100644 index 0000000000000000000000000000000000000000..3aa6a9b07539487b954b2d8c8d0e0bbcc49c2b42 --- /dev/null +++ b/tensorflow/compiler/tf2xla/lib/qr.h @@ -0,0 +1,40 @@ +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#ifndef TENSORFLOW_COMPILER_TF2XLA_LIB_QR_H_ +#define TENSORFLOW_COMPILER_TF2XLA_LIB_QR_H_ + +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" + +namespace tensorflow { + +// Computes the QR decompositions of a batch of matrices. That is, +// given a (batched) matrix a, computes an orthonormal matrix Q and an +// upper-triangular matrix R such that a = QR. +// `a` must be a (batched) matrix of size [..., m, n]. +// The algorithm implements a blocked QR decomposition; `block_size` is +// the block size to use. +// TODO(phawkins): handle the complex case. +struct QRDecompositionResult { + xla::XlaOp q; + xla::XlaOp r; +}; + +xla::StatusOr QRDecomposition(xla::XlaOp a, + int64 block_size = 128); + +} // namespace tensorflow + +#endif // TENSORFLOW_COMPILER_TF2XLA_LIB_QR_H_ diff --git a/tensorflow/compiler/tf2xla/lib/random.cc b/tensorflow/compiler/tf2xla/lib/random.cc index 4a2516244a54018c02a56ac099bbdc3c68a21141..8ff10fbd3fbf9308140af84c752a5a50bec8fd32 100644 --- a/tensorflow/compiler/tf2xla/lib/random.cc +++ b/tensorflow/compiler/tf2xla/lib/random.cc @@ -19,13 +19,14 @@ limitations under the License. #include #include "tensorflow/compiler/tf2xla/xla_helpers.h" -#include "tensorflow/compiler/xla/client/lib/arithmetic.h" +#include "tensorflow/compiler/xla/client/lib/constants.h" +#include "tensorflow/compiler/xla/client/lib/math.h" +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/compiler/xla/status_macros.h" namespace tensorflow { -xla::StatusOr TruncatedNormal(const DataType dtype, - const xla::XlaOp& uniform, - xla::XlaBuilder* builder) { + +xla::XlaOp TruncatedNormal(xla::XlaOp uniform) { auto normal_cdf = [](double x) { return (1.0 + std::erf(x / std::sqrt(2.0))) / 2.0; }; @@ -40,18 +41,15 @@ xla::StatusOr TruncatedNormal(const DataType dtype, const double kBetaNormalCdf = normal_cdf(kBeta); const double kZ = kBetaNormalCdf - kAlphaNormalCdf; - xla::XlaOp one = XlaHelpers::FloatLiteral(builder, dtype, 1.0); - xla::XlaOp two = XlaHelpers::FloatLiteral(builder, dtype, 2.0); - xla::XlaOp sqrt_2 = XlaHelpers::FloatLiteral(builder, dtype, std::sqrt(2.0)); - - xla::XlaOp z = XlaHelpers::FloatLiteral(builder, dtype, kZ); - xla::XlaOp alpha_normal_cdf = - XlaHelpers::FloatLiteral(builder, dtype, kAlphaNormalCdf); + xla::XlaOp one = xla::ScalarLike(uniform, 1.0); + xla::XlaOp two = xla::ScalarLike(uniform, 2.0); + xla::XlaOp sqrt_2 = xla::ScalarLike(uniform, std::sqrt(2.0)); + xla::XlaOp z = xla::ScalarLike(uniform, kZ); + xla::XlaOp alpha_normal_cdf = xla::ScalarLike(uniform, kAlphaNormalCdf); + auto p = alpha_normal_cdf + z * uniform; // probit(p) = sqrt(2) * erfinv(2*p-1) - auto p = builder->Add(alpha_normal_cdf, builder->Mul(z, uniform)); - auto erfinv_input = builder->Sub(builder->Mul(p, two), one); - TF_ASSIGN_OR_RETURN(auto erfinv_or_status, ErfInv(erfinv_input)); - return builder->Mul(sqrt_2, erfinv_or_status); + return sqrt_2 * xla::ErfInv(two * p - one); } + } // namespace tensorflow diff --git a/tensorflow/compiler/tf2xla/lib/random.h b/tensorflow/compiler/tf2xla/lib/random.h index 18c873dba5d2a104f6afaa526a1305d4e01f0a19..2c573fd85b2783fdac13457cdb277cf988ac40c4 100644 --- a/tensorflow/compiler/tf2xla/lib/random.h +++ b/tensorflow/compiler/tf2xla/lib/random.h @@ -21,15 +21,15 @@ limitations under the License. #include "tensorflow/core/framework/types.pb.h" namespace tensorflow { + // Builds an array filled with values sampled from a truncated normal // distribution such that no values are greater than two or less than negative // two. // // The "uniform" parameter must be an array of random numbers distributed in // (0,1). -xla::StatusOr TruncatedNormal(DataType dtype, - const xla::XlaOp& uniform, - xla::XlaBuilder* builder); +xla::XlaOp TruncatedNormal(xla::XlaOp uniform); + } // namespace tensorflow #endif // TENSORFLOW_COMPILER_TF2XLA_LIB_RANDOM_H_ diff --git a/tensorflow/compiler/tf2xla/lib/scatter.cc b/tensorflow/compiler/tf2xla/lib/scatter.cc index d5a27abb2585f699ae2719cb8a6b9a829263389e..6a5be1c2be57726d6c0e226407d52e7bfcebf92b 100644 --- a/tensorflow/compiler/tf2xla/lib/scatter.cc +++ b/tensorflow/compiler/tf2xla/lib/scatter.cc @@ -21,7 +21,8 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/lib/util.h" #include "tensorflow/compiler/tf2xla/lib/while_loop.h" #include "tensorflow/compiler/xla/client/lib/arithmetic.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/shape_util.h" #include "tensorflow/compiler/xla/status_macros.h" #include "tensorflow/compiler/xla/util.h" @@ -97,8 +98,8 @@ xla::StatusOr XlaScatter( buffer_shape_post_axes.end()); // Construct the initial values of the loop-carried Tensors. - auto flat_indices = builder->Reshape(indices, flat_indices_shape); - auto flat_updates = builder->Reshape(updates, flat_updates_shape); + auto flat_indices = xla::Reshape(indices, flat_indices_shape); + auto flat_updates = xla::Reshape(updates, flat_updates_shape); auto init = {flat_indices, flat_updates, buffer}; // Constructs the loop body. The implementation of scatter is essentially: @@ -112,46 +113,44 @@ xla::StatusOr XlaScatter( auto updates = loop_vars[1]; auto buffer = loop_vars[2]; - auto zero_index = body_builder->ConstantLiteral( - xla::Literal::Zero(indices_shape.element_type())); + auto zero_index = xla::ConstantLiteral( + body_builder, xla::LiteralUtil::Zero(indices_shape.element_type())); // Slice the i-th index from the indices array. xla::XlaOp index; - auto indices_offset = body_builder->Reshape(i, {1}); + auto indices_offset = xla::Reshape(i, {1}); if (indices_are_vectors) { - indices_offset = body_builder->Pad(indices_offset, zero_index, - xla::MakeEdgePaddingConfig({{0, 1}})); + indices_offset = xla::Pad(indices_offset, zero_index, + xla::MakeEdgePaddingConfig({{0, 1}})); - index = body_builder->DynamicSlice(indices, indices_offset, - {1, num_index_dims}); - index = body_builder->Collapse(index, {0, 1}); + index = xla::DynamicSlice(indices, indices_offset, {1, num_index_dims}); + index = xla::Collapse(index, {0, 1}); } else { - index = body_builder->DynamicSlice(indices, indices_offset, {1}); + index = xla::DynamicSlice(indices, indices_offset, {1}); } // Discard updates with negative indices, since some users expect this. - auto index_in_range = - body_builder->ReduceAll(body_builder->Le(zero_index, index), - body_builder->ConstantR0(true), - xla::CreateScalarAndComputation(body_builder)); + auto index_in_range = xla::ReduceAll( + xla::Le(zero_index, index), xla::ConstantR0(body_builder, true), + xla::CreateScalarAndComputation(body_builder)); // Make the index in bounds to prevent implementation defined behavior. - index = body_builder->Max(index, zero_index); - index = body_builder->Pad( + index = xla::Max(index, zero_index); + index = xla::Pad( index, zero_index, xla::MakeEdgePaddingConfig({{0, buffer_shape_post_axes.size()}})); // Slice the i-th index from the updates array. - auto updates_offset = body_builder->Reshape(i, {1}); - updates_offset = body_builder->Pad( + auto updates_offset = xla::Reshape(i, {1}); + updates_offset = xla::Pad( updates_offset, zero_index, xla::MakeEdgePaddingConfig({{0, buffer_shape_post_axes.size()}})); std::vector flat_updates_slice_shape({1}); flat_updates_slice_shape.insert(flat_updates_slice_shape.end(), buffer_shape_post_axes.begin(), buffer_shape_post_axes.end()); - auto update = body_builder->DynamicSlice(updates, updates_offset, - flat_updates_slice_shape); + auto update = + xla::DynamicSlice(updates, updates_offset, flat_updates_slice_shape); // Unflatten the major (iteration) dimensions of the slice to their // original shape. @@ -159,20 +158,19 @@ xla::StatusOr XlaScatter( updates_slice_shape.insert(updates_slice_shape.end(), buffer_shape_post_axes.begin(), buffer_shape_post_axes.end()); - update = body_builder->Reshape(update, updates_slice_shape); + update = xla::Reshape(update, updates_slice_shape); // Apply the update to the buffer. If there is a combiner, use it to merge // the current values with the update. - auto current_value = - body_builder->DynamicSlice(buffer, index, updates_slice_shape); + auto current_value = xla::DynamicSlice(buffer, index, updates_slice_shape); if (combiner) { update = combiner(current_value, update, body_builder); } // Use the current value instead of the update if the index is out of // bounds. - update = body_builder->Select(index_in_range, update, current_value); + update = xla::Select(index_in_range, update, current_value); // Apply the update. - buffer = body_builder->DynamicUpdateSlice(buffer, update, index); + buffer = xla::DynamicUpdateSlice(buffer, update, index); return std::vector{indices, updates, buffer}; }; diff --git a/tensorflow/compiler/tf2xla/lib/triangular_solve.cc b/tensorflow/compiler/tf2xla/lib/triangular_solve.cc index b4503601f94baa5a595a64c9fc81bc92d9980ac6..ce0f28db8f6de12c6710db99f991de6546bc5b9c 100644 --- a/tensorflow/compiler/tf2xla/lib/triangular_solve.cc +++ b/tensorflow/compiler/tf2xla/lib/triangular_solve.cc @@ -20,7 +20,9 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/lib/batch_dot.h" #include "tensorflow/compiler/tf2xla/lib/util.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/client/lib/constants.h" +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/shape_util.h" #include "tensorflow/compiler/xla/status_macros.h" #include "tensorflow/compiler/xla/statusor.h" @@ -29,619 +31,564 @@ limitations under the License. namespace tensorflow { -xla::StatusOr TriangularSolve(xla::XlaBuilder* builder, - const xla::XlaOp& a, xla::XlaOp b, - bool left_side, bool lower, - bool transpose_a, bool conjugate_a, - int64 block_size) { - TF_ASSIGN_OR_RETURN(xla::Shape a_shape, builder->GetShape(a)); - TF_ASSIGN_OR_RETURN(xla::Shape b_shape, builder->GetShape(b)); - if (xla::ShapeUtil::Rank(a_shape) != xla::ShapeUtil::Rank(b_shape)) { - return errors::InvalidArgument( - "Arguments to TriangularSolve have different ranks: ", - xla::ShapeUtil::HumanString(a_shape), " vs. ", - xla::ShapeUtil::HumanString(b_shape)); - } - const int ndims = xla::ShapeUtil::Rank(a_shape); - if (ndims < 2) { - return errors::InvalidArgument( - "Arguments to TriangularSolve must have rank >= 2: ", ndims); - } - // The batch dimensions must be equal. - std::vector batch_dimensions; - for (int i = 0; i < ndims - 2; ++i) { - int64 a_size = a_shape.dimensions(i); - int64 b_size = b_shape.dimensions(i); - if (a_size != b_size) { +xla::XlaOp TriangularSolve(xla::XlaOp a, xla::XlaOp b, bool left_side, + bool lower, bool transpose_a, bool conjugate_a, + int64 block_size) { + xla::XlaBuilder* builder = a.builder(); + return builder->ReportErrorOrReturn([&]() -> xla::StatusOr { + TF_ASSIGN_OR_RETURN(xla::Shape a_shape, builder->GetShape(a)); + TF_ASSIGN_OR_RETURN(xla::Shape b_shape, builder->GetShape(b)); + if (xla::ShapeUtil::Rank(a_shape) != xla::ShapeUtil::Rank(b_shape)) { return errors::InvalidArgument( - "Batch dimensions of arguments to TriangularSolve must be equal: ", - xla::ShapeUtil::HumanString(a_shape), " vs ", + "Arguments to TriangularSolve have different ranks: ", + xla::ShapeUtil::HumanString(a_shape), " vs. ", xla::ShapeUtil::HumanString(b_shape)); } - batch_dimensions.push_back(a_size); - } - - if (xla::ShapeUtil::GetDimension(a_shape, -1) != - xla::ShapeUtil::GetDimension(a_shape, -2)) { - return errors::InvalidArgument( - "The 'a' arguments to TriangularSolve must be square matrices: ", - xla::ShapeUtil::HumanString(a_shape)); - } - const int64 m = xla::ShapeUtil::GetDimension(b_shape, -2); - const int64 n = xla::ShapeUtil::GetDimension(b_shape, -1); - if ((left_side ? m : n) != xla::ShapeUtil::GetDimension(a_shape, -1)) { - return errors::InvalidArgument( - "Arguments to TriangularSolve have incompatible matrix shapes: ", - xla::ShapeUtil::HumanString(a_shape), " vs ", - xla::ShapeUtil::HumanString(b_shape)); - } - - if (block_size < 1) { - return errors::InvalidArgument( - "block_size argument to TriangularSolve must be >= 1; got ", - block_size); - } - - std::map base_computations; - auto get_base_triangular_solve = - [&](int k) -> xla::StatusOr { - xla::XlaComputation& computation = base_computations[k]; - if (computation.IsNull()) { - std::unique_ptr sub = builder->CreateSubBuilder( - tensorflow::strings::StrCat("trsm_base_", k)); - - auto a_param = sub->Parameter( - 0, - xla::ShapeUtil::MakeShape( - b_shape.element_type(), - PrependMajorDims(sub.get(), batch_dimensions, {k, k})), - "a"); - - std::array b_lastd; - if (left_side) { - b_lastd = {k, n}; - } else { - b_lastd = {m, k}; - } - auto b_param = sub->Parameter( - 1, - xla::ShapeUtil::MakeShape( - b_shape.element_type(), - PrependMajorDims(sub.get(), batch_dimensions, b_lastd)), - "b"); - - // We use a left-looking or right-looking subroutine on the block diagonal - // in the lower=true cases, while falling back to a recursive call in - // others. The left-looking and right-looking subroutines are written with - // a While loop and so yields much faster compile times. Moreover, they - // can give higher performance on smaller (sub)problems. - if (left_side && lower) { - TF_RETURN_IF_ERROR(TriangularSolveLeftLooking(sub.get(), a_param, - b_param, transpose_a, - conjugate_a) - .status()); - } else if (!left_side && lower) { - TF_RETURN_IF_ERROR(TriangularSolveRightLooking(sub.get(), a_param, - b_param, transpose_a, - conjugate_a) - .status()); - } else { - TF_RETURN_IF_ERROR(TriangularSolve(sub.get(), a_param, b_param, - left_side, lower, transpose_a, - conjugate_a, - /*block_size=*/1) - .status()); + const int ndims = xla::ShapeUtil::Rank(a_shape); + if (ndims < 2) { + return errors::InvalidArgument( + "Arguments to TriangularSolve must have rank >= 2: ", ndims); + } + // The batch dimensions must be equal. + std::vector batch_dimensions; + for (int i = 0; i < ndims - 2; ++i) { + int64 a_size = a_shape.dimensions(i); + int64 b_size = b_shape.dimensions(i); + if (a_size != b_size) { + return errors::InvalidArgument( + "Batch dimensions of arguments to TriangularSolve must be equal: ", + xla::ShapeUtil::HumanString(a_shape), " vs ", + xla::ShapeUtil::HumanString(b_shape)); } + batch_dimensions.push_back(a_size); + } - TF_ASSIGN_OR_RETURN(computation, sub->Build()); + if (xla::ShapeUtil::GetDimension(a_shape, -1) != + xla::ShapeUtil::GetDimension(a_shape, -2)) { + return errors::InvalidArgument( + "The 'a' arguments to TriangularSolve must be square matrices: ", + xla::ShapeUtil::HumanString(a_shape)); + } + const int64 m = xla::ShapeUtil::GetDimension(b_shape, -2); + const int64 n = xla::ShapeUtil::GetDimension(b_shape, -1); + if ((left_side ? m : n) != xla::ShapeUtil::GetDimension(a_shape, -1)) { + return errors::InvalidArgument( + "Arguments to TriangularSolve have incompatible matrix shapes: ", + xla::ShapeUtil::HumanString(a_shape), " vs ", + xla::ShapeUtil::HumanString(b_shape)); } - return &computation; - }; - - xla::XlaOp output = Zeros(builder, b_shape); - - // Right-looking blocked triangular solve. - // For an explanation of the algorithm, see the TRSM discussion in: - // Goto, Kazushige, and Robert Van De Geijn. "High-performance implementation - // of the level-3 BLAS." ACM Transactions on Mathematical Software (TOMS) 35.1 - // (2008): 4. - - // In the code comments below, T = lambda x: np.swapaxes(x, -1, -2) if - // conjugate_a is False, or T = lambda x: np.conj(np.swapaxes(x, -1, -2)) if - // conjugate_a is True. - - if (!left_side && lower == transpose_a) { - // for i in range(0, a.shape[-1], block_size): - for (int64 i = 0; i < n; i += block_size) { - int64 k = std::min(block_size, n - i); - - // output[..., :, i:i+k] = triangular_solve( - // a[..., i:i+k, i:i+k], b[..., :, i:i+k], ..., block_size=1) - TF_ASSIGN_OR_RETURN(auto a_slice, - SliceInMinorDims(builder, a, {i, i}, {i + k, i + k})); - TF_ASSIGN_OR_RETURN(auto b_slice, - SliceInMinorDims(builder, b, {0, i}, {m, i + k})); - xla::XlaOp update; - if (k > 1) { - TF_ASSIGN_OR_RETURN(xla::XlaComputation * solve, - get_base_triangular_solve(k)); - update = builder->Call(*solve, {a_slice, b_slice}); - } else { - TF_ASSIGN_OR_RETURN(auto a_slice_conj, - MaybeConjugate(builder, a_slice, conjugate_a)); - update = builder->Div(b_slice, a_slice_conj); - } - TF_ASSIGN_OR_RETURN( - output, UpdateSliceInMinorDims(builder, output, update, {0, i})); - - // if i + k < a.shape[-1]: - // a_slice_2 = a[..., i+k:, i:i+k] if lower else a[..., i:i+k, i+k:] - // a_slice_2 = T(a_slice_2) if transpose_a else a_slice_2 - // b[..., :, i+k:] -= np.matmul(output[..., :, i:i+k], a_slice_2) - if (i + k < n) { - xla::XlaOp a_slice_2; - if (lower) { - TF_ASSIGN_OR_RETURN( - a_slice_2, SliceInMinorDims(builder, a, {i + k, i}, {n, i + k})); - } else { - TF_ASSIGN_OR_RETURN( - a_slice_2, SliceInMinorDims(builder, a, {i, i + k}, {i + k, n})); - } - TF_ASSIGN_OR_RETURN(auto b_update, - BatchDot(builder, update, a_slice_2, - /*transpose_x=*/false, - /*transpose_y=*/transpose_a, - /*conjugate_x=*/false, - /*conjugate_y=*/conjugate_a)); - TF_ASSIGN_OR_RETURN(auto b_slice_2, - SliceInMinorDims(builder, b, {0, i + k}, {m, n})); - b_update = builder->Sub(b_slice_2, b_update); - TF_ASSIGN_OR_RETURN( - b, UpdateSliceInMinorDims(builder, b, b_update, {0, i + k})); - } + if (block_size < 1) { + return errors::InvalidArgument( + "block_size argument to TriangularSolve must be >= 1; got ", + block_size); } - } else if (left_side && lower != transpose_a) { - // for i in range(0, a.shape[-1], block_size): - for (int64 i = 0; i < m; i += block_size) { - int64 k = std::min(block_size, m - i); - - // output[..., i:i+k, :] = triangular_solve( - // a[..., i:i+k, i:i+k], b[..., i:i+k, :], ..., block_size=1) - TF_ASSIGN_OR_RETURN(auto a_slice, - SliceInMinorDims(builder, a, {i, i}, {i + k, i + k})); - TF_ASSIGN_OR_RETURN(auto b_slice, - SliceInMinorDims(builder, b, {i, 0}, {i + k, n})); - xla::XlaOp update; - if (k > 1) { - TF_ASSIGN_OR_RETURN(xla::XlaComputation * solve, - get_base_triangular_solve(k)); - update = builder->Call(*solve, {a_slice, b_slice}); - } else { - TF_ASSIGN_OR_RETURN(auto a_slice_conj, - MaybeConjugate(builder, a_slice, conjugate_a)); - update = builder->Div(b_slice, a_slice_conj); - } - TF_ASSIGN_OR_RETURN( - output, UpdateSliceInMinorDims(builder, output, update, {i, 0})); - - // if i + k < a.shape[-1]: - // a_slice_2 = a[..., i+k:, i:i+k] if lower else a[..., i:i+k, i+k:] - // a_slice_2 = T(a_slice_2) if transpose_a else a_slice_2 - // b[..., i+k:, :] -= np.matmul(a_slice_2, output[..., i:i+k, :]) - if (i + k < m) { - xla::XlaOp a_slice_2; - if (lower) { - TF_ASSIGN_OR_RETURN( - a_slice_2, SliceInMinorDims(builder, a, {i + k, i}, {m, i + k})); + std::map base_computations; + auto get_base_triangular_solve = + [&](int k) -> xla::StatusOr { + xla::XlaComputation& computation = base_computations[k]; + if (computation.IsNull()) { + std::unique_ptr sub = builder->CreateSubBuilder( + tensorflow::strings::StrCat("trsm_base_", k)); + + auto a_param = xla::Parameter( + sub.get(), 0, + xla::ShapeUtil::MakeShape(b_shape.element_type(), + ConcatVectors(batch_dimensions, {k, k})), + "a"); + + std::array b_lastd; + if (left_side) { + b_lastd = {k, n}; + } else { + b_lastd = {m, k}; + } + auto b_param = xla::Parameter( + sub.get(), 1, + xla::ShapeUtil::MakeShape(b_shape.element_type(), + ConcatVectors(batch_dimensions, b_lastd)), + "b"); + + // We use a left-looking or right-looking subroutine on the block + // diagonal in the lower=true cases, while falling back to a recursive + // call in others. The left-looking and right-looking subroutines are + // written with a While loop and so yields much faster compile times. + // Moreover, they can give higher performance on smaller (sub)problems. + if (left_side && lower) { + TriangularSolveLeftLooking(a_param, b_param, transpose_a, + conjugate_a); + } else if (!left_side && lower) { + TriangularSolveRightLooking(a_param, b_param, transpose_a, + conjugate_a); } else { - TF_ASSIGN_OR_RETURN( - a_slice_2, SliceInMinorDims(builder, a, {i, i + k}, {i + k, m})); + TriangularSolve(a_param, b_param, left_side, lower, transpose_a, + conjugate_a, + /*block_size=*/1); } - TF_ASSIGN_OR_RETURN(auto b_update, BatchDot(builder, a_slice_2, update, - /*transpose_x=*/transpose_a, - /*transpose_y=*/false, - /*conjugate_x=*/conjugate_a, - /*conjugate_y=*/false)); - TF_ASSIGN_OR_RETURN(auto b_slice_2, - SliceInMinorDims(builder, b, {i + k, 0}, {m, n})); - b_update = builder->Sub(b_slice_2, b_update); - TF_ASSIGN_OR_RETURN( - b, UpdateSliceInMinorDims(builder, b, b_update, {i + k, 0})); + TF_ASSIGN_OR_RETURN(computation, sub->Build()); } - } - } else if (!left_side && lower != transpose_a) { - // for i in reversed(range(0, a.shape[-1], block_size)): - const int64 last_blk_ix = xla::RoundUpToNearest(n, block_size) - block_size; - for (int64 i = last_blk_ix; i >= 0; i -= block_size) { - int64 k = std::min(block_size, n - i); - - // output[..., :, i:i+k] triangular_solve( - // a[..., i:i+k, i:i+k], b[..., :, i:i+k], ..., block_size=1) - TF_ASSIGN_OR_RETURN(auto a_slice, - SliceInMinorDims(builder, a, {i, i}, {i + k, i + k})); - TF_ASSIGN_OR_RETURN(auto b_slice, - SliceInMinorDims(builder, b, {0, i}, {m, i + k})); - xla::XlaOp update; - if (k > 1) { - TF_ASSIGN_OR_RETURN(xla::XlaComputation * solve, - get_base_triangular_solve(k)); - update = builder->Call(*solve, {a_slice, b_slice}); - } else { - TF_ASSIGN_OR_RETURN(auto a_slice_conj, - MaybeConjugate(builder, a_slice, conjugate_a)); - update = builder->Div(b_slice, a_slice_conj); - } - TF_ASSIGN_OR_RETURN( - output, UpdateSliceInMinorDims(builder, output, update, {0, i})); - - // if i - k >= 0: - // a_slice_2 = a[..., i:i+k, :i] if lower else a[..., :i, i:i+k] - // a_slice_2 = T(a_slice_2) if transpose_a else a_slice_2 - // b[..., :, :i] -= np.matmul(out[..., :, i:i+k], a_slice_2) - if (i - k >= 0) { - xla::XlaOp a_slice_2; - if (lower) { - TF_ASSIGN_OR_RETURN(a_slice_2, - SliceInMinorDims(builder, a, {i, 0}, {i + k, i})); + return &computation; + }; + + xla::XlaOp output = xla::ZerosLike(b); + + // Right-looking blocked triangular solve. + // For an explanation of the algorithm, see the TRSM discussion in: + // Goto, Kazushige, and Robert Van De Geijn. "High-performance + // implementation of the level-3 BLAS." ACM Transactions on Mathematical + // Software (TOMS) 35.1 (2008): 4. + + // In the code comments below, T = lambda x: np.swapaxes(x, -1, -2) if + // conjugate_a is False, or T = lambda x: np.conj(np.swapaxes(x, -1, -2)) if + // conjugate_a is True. + + if (!left_side && lower == transpose_a) { + // for i in range(0, a.shape[-1], block_size): + for (int64 i = 0; i < n; i += block_size) { + int64 k = std::min(block_size, n - i); + + // output[..., :, i:i+k] = triangular_solve( + // a[..., i:i+k, i:i+k], b[..., :, i:i+k], ..., block_size=1) + auto a_slice = SliceInMinorDims(a, {i, i}, {i + k, i + k}); + auto b_slice = SliceInMinorDims(b, {0, i}, {m, i + k}); + xla::XlaOp update; + if (k > 1) { + TF_ASSIGN_OR_RETURN(xla::XlaComputation * solve, + get_base_triangular_solve(k)); + update = xla::Call(builder, *solve, {a_slice, b_slice}); } else { - TF_ASSIGN_OR_RETURN(a_slice_2, - SliceInMinorDims(builder, a, {0, i}, {i, i + k})); + auto a_slice_conj = MaybeConjugate(a_slice, conjugate_a); + update = b_slice / a_slice_conj; + } + output = UpdateSliceInMinorDims(output, update, {0, i}); + + // if i + k < a.shape[-1]: + // a_slice_2 = a[..., i+k:, i:i+k] if lower else a[..., i:i+k, i+k:] + // a_slice_2 = T(a_slice_2) if transpose_a else a_slice_2 + // b[..., :, i+k:] -= np.matmul(output[..., :, i:i+k], a_slice_2) + if (i + k < n) { + xla::XlaOp a_slice_2; + if (lower) { + a_slice_2 = SliceInMinorDims(a, {i + k, i}, {n, i + k}); + } else { + a_slice_2 = SliceInMinorDims(a, {i, i + k}, {i + k, n}); + } + + auto b_update = BatchDot(update, a_slice_2, + /*transpose_x=*/false, + /*transpose_y=*/transpose_a, + /*conjugate_x=*/false, + /*conjugate_y=*/conjugate_a); + auto b_slice_2 = SliceInMinorDims(b, {0, i + k}, {m, n}); + b = UpdateSliceInMinorDims(b, b_slice_2 - b_update, {0, i + k}); } + } - TF_ASSIGN_OR_RETURN(auto b_update, - BatchDot(builder, update, a_slice_2, - /*transpose_x=*/false, - /*transpose_y=*/transpose_a, - /*conjugate_x=*/false, - /*conjugate_y=*/conjugate_a)); - TF_ASSIGN_OR_RETURN(auto b_slice_2, - SliceInMinorDims(builder, b, {0, 0}, {m, i})); - b_update = builder->Sub(b_slice_2, b_update); - TF_ASSIGN_OR_RETURN( - b, UpdateSliceInMinorDims(builder, b, b_update, {0, 0})); + } else if (left_side && lower != transpose_a) { + // for i in range(0, a.shape[-1], block_size): + for (int64 i = 0; i < m; i += block_size) { + int64 k = std::min(block_size, m - i); + + // output[..., i:i+k, :] = triangular_solve( + // a[..., i:i+k, i:i+k], b[..., i:i+k, :], ..., block_size=1) + auto a_slice = SliceInMinorDims(a, {i, i}, {i + k, i + k}); + auto b_slice = SliceInMinorDims(b, {i, 0}, {i + k, n}); + xla::XlaOp update; + if (k > 1) { + TF_ASSIGN_OR_RETURN(xla::XlaComputation * solve, + get_base_triangular_solve(k)); + update = xla::Call(builder, *solve, {a_slice, b_slice}); + } else { + auto a_slice_conj = MaybeConjugate(a_slice, conjugate_a); + update = b_slice / a_slice_conj; + } + output = UpdateSliceInMinorDims(output, update, {i, 0}); + + // if i + k < a.shape[-1]: + // a_slice_2 = a[..., i+k:, i:i+k] if lower else a[..., i:i+k, i+k:] + // a_slice_2 = T(a_slice_2) if transpose_a else a_slice_2 + // b[..., i+k:, :] -= np.matmul(a_slice_2, output[..., i:i+k, :]) + if (i + k < m) { + xla::XlaOp a_slice_2; + if (lower) { + a_slice_2 = SliceInMinorDims(a, {i + k, i}, {m, i + k}); + } else { + a_slice_2 = SliceInMinorDims(a, {i, i + k}, {i + k, m}); + } + + auto b_update = BatchDot(a_slice_2, update, + /*transpose_x=*/transpose_a, + /*transpose_y=*/false, + /*conjugate_x=*/conjugate_a, + /*conjugate_y=*/false); + auto b_slice_2 = SliceInMinorDims(b, {i + k, 0}, {m, n}); + b = UpdateSliceInMinorDims(b, b_slice_2 - b_update, {i + k, 0}); + } } - } - } else { // left_side && lower == transpose_a - // for i in reversed(range(0, a.shape[-1], block_size)): - const int64 last_blk_ix = xla::RoundUpToNearest(m, block_size) - block_size; - for (int64 i = last_blk_ix; i >= 0; i -= block_size) { - int64 k = std::min(block_size, m - i); - - // output[..., i:i+k, :] triangular_solve( - // a[..., i:i+k, i:i+k], b[..., i:i+k, :], ..., block_size=1) - TF_ASSIGN_OR_RETURN(auto a_slice, - SliceInMinorDims(builder, a, {i, i}, {i + k, i + k})); - TF_ASSIGN_OR_RETURN(auto b_slice, - SliceInMinorDims(builder, b, {i, 0}, {i + k, n})); - xla::XlaOp update; - if (k > 1) { - TF_ASSIGN_OR_RETURN(xla::XlaComputation * solve, - get_base_triangular_solve(k)); - update = builder->Call(*solve, {a_slice, b_slice}); - } else { - TF_ASSIGN_OR_RETURN(auto a_slice_conj, - MaybeConjugate(builder, a_slice, conjugate_a)); - update = builder->Div(b_slice, a_slice_conj); + } else if (!left_side && lower != transpose_a) { + // for i in reversed(range(0, a.shape[-1], block_size)): + const int64 last_blk_ix = + xla::RoundUpToNearest(n, block_size) - block_size; + for (int64 i = last_blk_ix; i >= 0; i -= block_size) { + int64 k = std::min(block_size, n - i); + + // output[..., :, i:i+k] triangular_solve( + // a[..., i:i+k, i:i+k], b[..., :, i:i+k], ..., block_size=1) + auto a_slice = SliceInMinorDims(a, {i, i}, {i + k, i + k}); + auto b_slice = SliceInMinorDims(b, {0, i}, {m, i + k}); + xla::XlaOp update; + if (k > 1) { + TF_ASSIGN_OR_RETURN(xla::XlaComputation * solve, + get_base_triangular_solve(k)); + update = xla::Call(builder, *solve, {a_slice, b_slice}); + } else { + auto a_slice_conj = MaybeConjugate(a_slice, conjugate_a); + update = b_slice / a_slice_conj; + } + output = UpdateSliceInMinorDims(output, update, {0, i}); + + // if i - k >= 0: + // a_slice_2 = a[..., i:i+k, :i] if lower else a[..., :i, i:i+k] + // a_slice_2 = T(a_slice_2) if transpose_a else a_slice_2 + // b[..., :, :i] -= np.matmul(out[..., :, i:i+k], a_slice_2) + if (i - k >= 0) { + xla::XlaOp a_slice_2; + if (lower) { + a_slice_2 = SliceInMinorDims(a, {i, 0}, {i + k, i}); + } else { + a_slice_2 = SliceInMinorDims(a, {0, i}, {i, i + k}); + } + + auto b_update = BatchDot(update, a_slice_2, + /*transpose_x=*/false, + /*transpose_y=*/transpose_a, + /*conjugate_x=*/false, + /*conjugate_y=*/conjugate_a); + auto b_slice_2 = SliceInMinorDims(b, {0, 0}, {m, i}); + b = UpdateSliceInMinorDims(b, b_slice_2 - b_update, {0, 0}); + } } - TF_ASSIGN_OR_RETURN( - output, UpdateSliceInMinorDims(builder, output, update, {i, 0})); - - // if i - k >= 0: - // a_slice_2 = a[..., i:i+k, :i] if lower else a[..., :i, i:i+k] - // a_slice_2 = T(a_slice_2) if transpose_a else a_slice_2 - // b[..., :i, :] -= np.matmul(a_slice_2, out[..., i:i+k, :]) - if (i - k >= 0) { - xla::XlaOp a_slice_2; - if (lower) { - TF_ASSIGN_OR_RETURN(a_slice_2, - SliceInMinorDims(builder, a, {i, 0}, {i + k, i})); + } else { // left_side && lower == transpose_a + // for i in reversed(range(0, a.shape[-1], block_size)): + const int64 last_blk_ix = + xla::RoundUpToNearest(m, block_size) - block_size; + for (int64 i = last_blk_ix; i >= 0; i -= block_size) { + int64 k = std::min(block_size, m - i); + + // output[..., i:i+k, :] triangular_solve( + // a[..., i:i+k, i:i+k], b[..., i:i+k, :], ..., block_size=1) + auto a_slice = SliceInMinorDims(a, {i, i}, {i + k, i + k}); + auto b_slice = SliceInMinorDims(b, {i, 0}, {i + k, n}); + xla::XlaOp update; + if (k > 1) { + TF_ASSIGN_OR_RETURN(xla::XlaComputation * solve, + get_base_triangular_solve(k)); + update = xla::Call(builder, *solve, {a_slice, b_slice}); } else { - TF_ASSIGN_OR_RETURN(a_slice_2, - SliceInMinorDims(builder, a, {0, i}, {i, i + k})); + auto a_slice_conj = MaybeConjugate(a_slice, conjugate_a); + update = b_slice / a_slice_conj; + } + output = UpdateSliceInMinorDims(output, update, {i, 0}); + + // if i - k >= 0: + // a_slice_2 = a[..., i:i+k, :i] if lower else a[..., :i, i:i+k] + // a_slice_2 = T(a_slice_2) if transpose_a else a_slice_2 + // b[..., :i, :] -= np.matmul(a_slice_2, out[..., i:i+k, :]) + if (i - k >= 0) { + xla::XlaOp a_slice_2; + if (lower) { + a_slice_2 = SliceInMinorDims(a, {i, 0}, {i + k, i}); + } else { + a_slice_2 = SliceInMinorDims(a, {0, i}, {i, i + k}); + } + + auto b_update = BatchDot(a_slice_2, update, + /*transpose_x=*/transpose_a, + /*transpose_y=*/false, + /*conjugate_x=*/conjugate_a, + /*conjugate_y=*/false); + auto b_slice_2 = SliceInMinorDims(b, {0, 0}, {i, n}); + b = UpdateSliceInMinorDims(b, b_slice_2 - b_update, {0, 0}); } - - TF_ASSIGN_OR_RETURN(auto b_update, BatchDot(builder, a_slice_2, update, - /*transpose_x=*/transpose_a, - /*transpose_y=*/false, - /*conjugate_x=*/conjugate_a, - /*conjugate_y=*/false)); - TF_ASSIGN_OR_RETURN(auto b_slice_2, - SliceInMinorDims(builder, b, {0, 0}, {i, n})); - b_update = builder->Sub(b_slice_2, b_update); - TF_ASSIGN_OR_RETURN( - b, UpdateSliceInMinorDims(builder, b, b_update, {0, 0})); } } - } - return output; + return output; + }); } -xla::StatusOr TriangularSolveLeftLooking(xla::XlaBuilder* builder, - const xla::XlaOp& a, - const xla::XlaOp& b, - bool transpose_a, - bool conjugate_a) { - TF_ASSIGN_OR_RETURN(xla::Shape a_shape, builder->GetShape(a)); - TF_ASSIGN_OR_RETURN(xla::Shape b_shape, builder->GetShape(b)); - const int64 m = xla::ShapeUtil::GetDimension(b_shape, -2); - const int64 n = xla::ShapeUtil::GetDimension(b_shape, -1); - const int64 ndims = xla::ShapeUtil::Rank(a_shape); - - std::vector batch_dimensions; - for (int i = 0; i < ndims - 2; ++i) { - int64 a_size = a_shape.dimensions(i); - batch_dimensions.push_back(a_size); - } - - // The main computation is performed in a While loop. - - // Allocate the output and set its first or last row, - // output = np.zeros_like(b) - // if transpose_a: - // output[..., m-1:, :] = b[..., m-1:, :] / a[..., m-1:, m-1:] - // else: - // output[..., :1, :] = b[..., :1, :] / a[..., :1, :1] - xla::XlaOp output = Zeros(builder, b_shape); - { - auto i = transpose_a ? m - 1 : 0; - TF_ASSIGN_OR_RETURN(auto a_slice, - SliceInMinorDims(builder, a, {i, i}, {i + 1, i + 1})); - TF_ASSIGN_OR_RETURN(auto b_slice, - SliceInMinorDims(builder, b, {i, 0}, {i + 1, n})); - TF_ASSIGN_OR_RETURN(auto a_slice_conj, - MaybeConjugate(builder, a_slice, conjugate_a)); - auto update = builder->Div(b_slice, a_slice_conj); - TF_ASSIGN_OR_RETURN( - output, UpdateSliceInMinorDims(builder, output, update, {i, 0})); - } - - // Construct the initial loop carry tuple, - // if transpose_a: - // init = (m-2, output, a, b) - // else: - // init = (1, output, a, b) - std::vector tuple_shapes = { - // The loop iteration counter is a scalar, incremented each iteration. - xla::ShapeUtil::MakeShape(xla::S32, {}), - // The output has the shape of b, with one row updated each iteration. - b_shape, - // The coefficient matrix a is a loop invariant. - a_shape, - // The right-hand-side matrix b is a loop invariant. - b_shape}; - xla::Shape tuple_shape = xla::ShapeUtil::MakeTupleShape(tuple_shapes); - auto init_i = builder->ConstantR0(transpose_a ? m - 2 : 1); - auto init = builder->Tuple({init_i, output, a, b}); - - // Construct the loop condition function, - // def cond_fun(loop_carry): - // i, output, a, b = loop_carry - // return i >= 0 if transpose_a else i < m - std::unique_ptr condb = - builder->CreateSubBuilder("TriangularSolveLeftLookingWhileCond"); - { - auto i = condb->GetTupleElement( - condb->Parameter(0, tuple_shape, - "TriangularSolveLeftLookingWhileTuple"), - 0); - if (transpose_a) { - condb->Ge(i, condb->ConstantR0(0)); - } else { - condb->Lt(i, condb->ConstantR0(m)); +xla::XlaOp TriangularSolveLeftLooking(xla::XlaOp a, xla::XlaOp b, + bool transpose_a, bool conjugate_a) { + xla::XlaBuilder* builder = a.builder(); + return builder->ReportErrorOrReturn([&]() -> xla::StatusOr { + TF_ASSIGN_OR_RETURN(xla::Shape a_shape, builder->GetShape(a)); + TF_ASSIGN_OR_RETURN(xla::Shape b_shape, builder->GetShape(b)); + const int64 m = xla::ShapeUtil::GetDimension(b_shape, -2); + const int64 n = xla::ShapeUtil::GetDimension(b_shape, -1); + const int64 ndims = xla::ShapeUtil::Rank(a_shape); + + std::vector batch_dimensions; + for (int i = 0; i < ndims - 2; ++i) { + int64 a_size = a_shape.dimensions(i); + batch_dimensions.push_back(a_size); } - } - TF_ASSIGN_OR_RETURN(auto cond, condb->Build()); - - // Construct the loop body function, - // def body_fun(loop_carry): - // i, output, a, b = loop_carry - // if transpose_a: - // a_row = np.swapaxes(a[..., i+1:, i:i+1], -1 -2) - // else: - // a_row = a[..., i:i+1, :i] - // result_row = b[..., i:i+1, :] - np.matmul(a_row, output[..., :, :]) - // output[..., i:i+1, :] = result_row / a[..., i:i+1, i:i+1] - // if transpose_a: - // return (i - 1, output, a, b) - // else: - // return (i + 1, output, a, b) - // We have to do some extra FLOPs propagating zeros in the matrix multiply - // because we can't have the size of its arguments depend on the loop counter. - std::unique_ptr bodyb = - builder->CreateSubBuilder("TriangularSolveLeftLookingWhileBody"); - { - auto input_tuple = bodyb->Parameter(0, tuple_shape, - "TriangularSolveLeftLookingWhileTuple"); - // i, output, a, b = loop_carry - auto i = bodyb->GetTupleElement(input_tuple, 0); - auto body_out = bodyb->GetTupleElement(input_tuple, 1); - auto body_a = bodyb->GetTupleElement(input_tuple, 2); - auto body_b = bodyb->GetTupleElement(input_tuple, 3); - auto zero = bodyb->ConstantR0(0); + // The main computation is performed in a While loop. - // We'd like to implement this: - // if transpose_a: - // a_row = T(a[..., i+1:, i:i+1]) - // result_row = (b[..., i:i+1, :] - // - np.matmul(a_row, body_out[..., i+1:, :])) - // else: - // result_row = (b[..., i:i+1, :] - // - np.matmul(a[..., i:i+1, :i], body_out[..., :i, :])) - // But since we can't have intermediate array sizes depend on the loop - // counter, we instead exploit the fact that we initialized the output to - // all zeros and use that as zero-padding (doing unnecessary FLOPs). - xla::XlaOp a_row; - if (transpose_a) { - TF_ASSIGN_OR_RETURN(a_row, DynamicSliceInMinorDims(bodyb.get(), body_a, - {zero, i}, {m, 1})); - } else { - TF_ASSIGN_OR_RETURN(a_row, DynamicSliceInMinorDims(bodyb.get(), body_a, - {i, zero}, {1, m})); + // Allocate the output and set its first or last row, + // output = np.zeros_like(b) + // if transpose_a: + // output[..., m-1:, :] = b[..., m-1:, :] / a[..., m-1:, m-1:] + // else: + // output[..., :1, :] = b[..., :1, :] / a[..., :1, :1] + xla::XlaOp output = xla::ZerosLike(b); + { + auto i = transpose_a ? m - 1 : 0; + auto a_slice = SliceInMinorDims(a, {i, i}, {i + 1, i + 1}); + auto b_slice = SliceInMinorDims(b, {i, 0}, {i + 1, n}); + auto a_slice_conj = MaybeConjugate(a_slice, conjugate_a); + auto update = b_slice / a_slice_conj; + output = UpdateSliceInMinorDims(output, update, {i, 0}); } - TF_ASSIGN_OR_RETURN(auto b_update, BatchDot(bodyb.get(), a_row, body_out, - /*transpose_x=*/transpose_a, - /*transpose_y=*/false, - /*conjugate_x=*/conjugate_a, - /*conjugate_y=*/false)); - TF_ASSIGN_OR_RETURN( - auto result_row_slice, - DynamicSliceInMinorDims(bodyb.get(), body_b, {i, zero}, {1, n})); - auto result_row = bodyb->Sub(result_row_slice, b_update); - - // body_out[..., i:i+1, :] = result_row / a[..., i:i+1, i:i+1] - TF_ASSIGN_OR_RETURN(auto a_elt, DynamicSliceInMinorDims(bodyb.get(), body_a, - {i, i}, {1, 1})); - TF_ASSIGN_OR_RETURN(auto a_elt_conj, - MaybeConjugate(bodyb.get(), a_elt, conjugate_a)); - auto div_result = bodyb->Div(result_row, a_elt_conj); - TF_ASSIGN_OR_RETURN(body_out, - DynamicUpdateSliceInMinorDims(bodyb.get(), body_out, - div_result, {i, zero})); + // Construct the initial loop carry tuple, // if transpose_a: - // return (i - 1, body_out, a, b) + // init = (m-2, output, a, b) // else: - // return (i + 1, body_out, a, b) - auto next_i = bodyb->Add(i, bodyb->ConstantR0(transpose_a ? -1 : 1)); - bodyb->Tuple({next_i, body_out, body_a, body_b}); - } - TF_ASSIGN_OR_RETURN(auto body, bodyb->Build()); - - // Construct the While loop and return the result, - // return while_loop(cond_fun, body_fun, init)[1] - auto triangular_solve_left_looking_while = builder->While(cond, body, init); - return builder->GetTupleElement(triangular_solve_left_looking_while, 1); + // init = (1, output, a, b) + std::vector tuple_shapes = { + // The loop iteration counter is a scalar, incremented each iteration. + xla::ShapeUtil::MakeShape(xla::S32, {}), + // The output has the shape of b, with one row updated each iteration. + b_shape, + // The coefficient matrix a is a loop invariant. + a_shape, + // The right-hand-side matrix b is a loop invariant. + b_shape}; + xla::Shape tuple_shape = xla::ShapeUtil::MakeTupleShape(tuple_shapes); + auto init_i = xla::ConstantR0(builder, transpose_a ? m - 2 : 1); + auto init = xla::Tuple(builder, {init_i, output, a, b}); + + // Construct the loop condition function, + // def cond_fun(loop_carry): + // i, output, a, b = loop_carry + // return i >= 0 if transpose_a else i < m + std::unique_ptr condb = + builder->CreateSubBuilder("TriangularSolveLeftLookingWhileCond"); + { + auto i = xla::GetTupleElement( + xla::Parameter(condb.get(), 0, tuple_shape, + "TriangularSolveLeftLookingWhileTuple"), + 0); + if (transpose_a) { + xla::Ge(i, xla::ConstantR0(condb.get(), 0)); + } else { + xla::Lt(i, xla::ConstantR0(condb.get(), m)); + } + } + TF_ASSIGN_OR_RETURN(auto cond, condb->Build()); + + // Construct the loop body function, + // def body_fun(loop_carry): + // i, output, a, b = loop_carry + // if transpose_a: + // a_row = np.swapaxes(a[..., i+1:, i:i+1], -1 -2) + // else: + // a_row = a[..., i:i+1, :i] + // result_row = b[..., i:i+1, :] - np.matmul(a_row, output[..., :, :]) + // output[..., i:i+1, :] = result_row / a[..., i:i+1, i:i+1] + // if transpose_a: + // return (i - 1, output, a, b) + // else: + // return (i + 1, output, a, b) + // We have to do some extra FLOPs propagating zeros in the matrix multiply + // because we can't have the size of its arguments depend on the loop + // counter. + std::unique_ptr bodyb = + builder->CreateSubBuilder("TriangularSolveLeftLookingWhileBody"); + { + auto input_tuple = xla::Parameter(bodyb.get(), 0, tuple_shape, + "TriangularSolveLeftLookingWhileTuple"); + + // i, output, a, b = loop_carry + auto i = xla::GetTupleElement(input_tuple, 0); + auto body_out = xla::GetTupleElement(input_tuple, 1); + auto body_a = xla::GetTupleElement(input_tuple, 2); + auto body_b = xla::GetTupleElement(input_tuple, 3); + auto zero = xla::ConstantR0(bodyb.get(), 0); + + // We'd like to implement this: + // if transpose_a: + // a_row = T(a[..., i+1:, i:i+1]) + // result_row = (b[..., i:i+1, :] + // - np.matmul(a_row, body_out[..., i+1:, :])) + // else: + // result_row = (b[..., i:i+1, :] + // - np.matmul(a[..., i:i+1, :i], body_out[..., :i, :])) + // But since we can't have intermediate array sizes depend on the loop + // counter, we instead exploit the fact that we initialized the output to + // all zeros and use that as zero-padding (doing unnecessary FLOPs). + xla::XlaOp a_row; + if (transpose_a) { + a_row = DynamicSliceInMinorDims(body_a, {zero, i}, {m, 1}); + } else { + a_row = DynamicSliceInMinorDims(body_a, {i, zero}, {1, m}); + } + auto b_update = BatchDot(a_row, body_out, + /*transpose_x=*/transpose_a, + /*transpose_y=*/false, + /*conjugate_x=*/conjugate_a, + /*conjugate_y=*/false); + auto result_row_slice = + DynamicSliceInMinorDims(body_b, {i, zero}, {1, n}); + auto result_row = result_row_slice - b_update; + + // body_out[..., i:i+1, :] = result_row / a[..., i:i+1, i:i+1] + auto a_elt = DynamicSliceInMinorDims(body_a, {i, i}, {1, 1}); + auto a_elt_conj = MaybeConjugate(a_elt, conjugate_a); + auto div_result = xla::Div(result_row, a_elt_conj); + body_out = DynamicUpdateSliceInMinorDims(body_out, div_result, {i, zero}); + + // if transpose_a: + // return (i - 1, body_out, a, b) + // else: + // return (i + 1, body_out, a, b) + auto next_i = xla::Add( + i, xla::ConstantR0(bodyb.get(), transpose_a ? -1 : 1)); + xla::Tuple(bodyb.get(), {next_i, body_out, body_a, body_b}); + } + TF_ASSIGN_OR_RETURN(auto body, bodyb->Build()); + + // Construct the While loop and return the result, + // return while_loop(cond_fun, body_fun, init)[1] + auto triangular_solve_left_looking_while = xla::While(cond, body, init); + return xla::GetTupleElement(triangular_solve_left_looking_while, 1); + }); } -xla::StatusOr TriangularSolveRightLooking(xla::XlaBuilder* builder, - const xla::XlaOp& a, - const xla::XlaOp& b, - bool transpose_a, - bool conjugate_a) { - TF_ASSIGN_OR_RETURN(xla::Shape a_shape, builder->GetShape(a)); - TF_ASSIGN_OR_RETURN(xla::Shape b_shape, builder->GetShape(b)); - const int64 m = xla::ShapeUtil::GetDimension(b_shape, -2); - const int64 n = xla::ShapeUtil::GetDimension(b_shape, -1); - const int64 ndims = xla::ShapeUtil::Rank(a_shape); - - std::vector batch_dimensions; - for (int i = 0; i < ndims - 2; ++i) { - int64 a_size = a_shape.dimensions(i); - batch_dimensions.push_back(a_size); - } - - // The main computation is performed in a While loop. - xla::XlaOp output = Zeros(builder, b_shape); - - // Construct the initial loop carry tuple, - // if transpose_a: - // init = (0, output, a, b) - // else: - // init = (n-1, output, a, b) - std::vector tuple_shapes = { - // The loop iteration counter is a scalar, incremented each iteration. - xla::ShapeUtil::MakeShape(xla::S32, {}), - // The output has the shape of b, with one row updated each iteration. - b_shape, - // The coefficient matrix a is a loop invariant. - a_shape, - // The right-hand-side matrix b is a loop invariant. - b_shape}; - xla::Shape tuple_shape = xla::ShapeUtil::MakeTupleShape(tuple_shapes); - auto init_i = builder->ConstantR0(transpose_a ? 0 : n - 1); - auto init = builder->Tuple({init_i, output, a, b}); - - // Construct the loop condition function, - // def cond_fun(loop_carry): - // i, output, a, b = loop_carry - // return i < n if transpose_a else i >= 0 - std::unique_ptr condb = - builder->CreateSubBuilder("TriangularSolveRightLookingWhileCond"); - { - auto i = condb->GetTupleElement( - condb->Parameter(0, tuple_shape, - "TriangularSolveRightLookingWhileTuple"), - 0); - if (transpose_a) { - condb->Lt(i, condb->ConstantR0(n)); - } else { - condb->Ge(i, condb->ConstantR0(0)); +xla::XlaOp TriangularSolveRightLooking(xla::XlaOp a, xla::XlaOp b, + bool transpose_a, bool conjugate_a) { + xla::XlaBuilder* builder = a.builder(); + return builder->ReportErrorOrReturn([&]() -> xla::StatusOr { + TF_ASSIGN_OR_RETURN(xla::Shape a_shape, builder->GetShape(a)); + TF_ASSIGN_OR_RETURN(xla::Shape b_shape, builder->GetShape(b)); + const int64 m = xla::ShapeUtil::GetDimension(b_shape, -2); + const int64 n = xla::ShapeUtil::GetDimension(b_shape, -1); + const int64 ndims = xla::ShapeUtil::Rank(a_shape); + + std::vector batch_dimensions; + for (int i = 0; i < ndims - 2; ++i) { + int64 a_size = a_shape.dimensions(i); + batch_dimensions.push_back(a_size); } - } - TF_ASSIGN_OR_RETURN(auto cond, condb->Build()); - - // Construct the loop body function, - // def body_fun(loop_carry): - // i, output, a, b = loop_carry - // if transpose_a: - // a_row = np.swapaxes(a[..., :, i:i+1], -1 -2) - // else: - // a_row = a[..., :, i:i+1] - // result_row = b[..., :, i:i+1] - np.matmul(output, a_row) - // output[..., :, i:i+1] = result_row / a[..., i:i+1, i:i+1] - // if transpose_a: - // return (i - 1, output, a, b) - // else: - // return (i + 1, output, a, b) - // We have to do some extra FLOPs propagating zeros in the matrix multiply - // because we can't have the size of its arguments depend on the loop counter. - std::unique_ptr bodyb = - builder->CreateSubBuilder("TriangularSolveRightLookingWhileBody"); - { - auto input_tuple = bodyb->Parameter( - 0, tuple_shape, "TriangularSolveRightLookingWhileTuple"); - - // i, output, a, b = loop_carry - auto i = bodyb->GetTupleElement(input_tuple, 0); - auto body_out = bodyb->GetTupleElement(input_tuple, 1); - auto body_a = bodyb->GetTupleElement(input_tuple, 2); - auto body_b = bodyb->GetTupleElement(input_tuple, 3); - auto zero = bodyb->ConstantR0(0); - - // We'd like to implement b[..., :, i:i+1] - np.matmul(output, a[..., :, - // i:i+1]) But since we can't have intermediate array sizes depend on the - // loop counter, we instead exploit the fact that we initialized the output - // to all zeros and use that as zero-padding (doing unnecessary FLOPs). - TF_ASSIGN_OR_RETURN(auto b_update, BatchDot(bodyb.get(), body_out, body_a, - /*transpose_x=*/false, - /*transpose_y=*/transpose_a, - /*conjugate_x=*/false, - /*conjugate_y=*/conjugate_a)); - // result = b - np.matmul(output, a) - auto result = bodyb->Sub(body_b, b_update); - // result_row = result[..., :, i:i+1] - TF_ASSIGN_OR_RETURN( - auto result_row, - DynamicSliceInMinorDims(bodyb.get(), result, {zero, i}, {m, 1})); - - // body_out[..., :, i:i+1] = result_row / a[..., i:i+1, i:i+1] - TF_ASSIGN_OR_RETURN(auto a_ii, DynamicSliceInMinorDims(bodyb.get(), body_a, - {i, i}, {1, 1})); - TF_ASSIGN_OR_RETURN(auto a_ii_conj, - MaybeConjugate(bodyb.get(), a_ii, conjugate_a)); - auto div_result = bodyb->Div(result_row, a_ii_conj); - TF_ASSIGN_OR_RETURN(body_out, - DynamicUpdateSliceInMinorDims(bodyb.get(), body_out, - div_result, {zero, i})); + // The main computation is performed in a While loop. + xla::XlaOp output = xla::ZerosLike(b); + + // Construct the initial loop carry tuple, // if transpose_a: - // return (i + 1, body_out, a, b) + // init = (0, output, a, b) // else: - // return (i - 1, body_out, a, b) - auto next_i = bodyb->Add(i, bodyb->ConstantR0(transpose_a ? 1 : -1)); - bodyb->Tuple({next_i, body_out, body_a, body_b}); - } - TF_ASSIGN_OR_RETURN(auto body, bodyb->Build()); - - // Construct the While loop and return the result, - // return while_loop(cond_fun, body_fun, init)[1] - auto triangular_solve_left_looking_while = builder->While(cond, body, init); - return builder->GetTupleElement(triangular_solve_left_looking_while, 1); + // init = (n-1, output, a, b) + std::vector tuple_shapes = { + // The loop iteration counter is a scalar, incremented each iteration. + xla::ShapeUtil::MakeShape(xla::S32, {}), + // The output has the shape of b, with one row updated each iteration. + b_shape, + // The coefficient matrix a is a loop invariant. + a_shape, + // The right-hand-side matrix b is a loop invariant. + b_shape}; + xla::Shape tuple_shape = xla::ShapeUtil::MakeTupleShape(tuple_shapes); + auto init_i = xla::ConstantR0(builder, transpose_a ? 0 : n - 1); + auto init = xla::Tuple(builder, {init_i, output, a, b}); + + // Construct the loop condition function, + // def cond_fun(loop_carry): + // i, output, a, b = loop_carry + // return i < n if transpose_a else i >= 0 + std::unique_ptr condb = + builder->CreateSubBuilder("TriangularSolveRightLookingWhileCond"); + { + auto i = xla::GetTupleElement( + xla::Parameter(condb.get(), 0, tuple_shape, + "TriangularSolveRightLookingWhileTuple"), + 0); + if (transpose_a) { + xla::Lt(i, xla::ConstantR0(condb.get(), n)); + } else { + xla::Ge(i, xla::ConstantR0(condb.get(), 0)); + } + } + TF_ASSIGN_OR_RETURN(auto cond, condb->Build()); + + // Construct the loop body function, + // def body_fun(loop_carry): + // i, output, a, b = loop_carry + // if transpose_a: + // a_row = np.swapaxes(a[..., :, i:i+1], -1, -2) + // else: + // a_row = a[..., :, i:i+1] + // result_row = b[..., :, i:i+1] - np.matmul(output, a_row) + // output[..., :, i:i+1] = result_row / a[..., i:i+1, i:i+1] + // if transpose_a: + // return (i - 1, output, a, b) + // else: + // return (i + 1, output, a, b) + // We have to do some extra FLOPs propagating zeros in the matrix multiply + // because we can't have the size of its arguments depend on the loop + // counter. + std::unique_ptr bodyb = + builder->CreateSubBuilder("TriangularSolveRightLookingWhileBody"); + { + auto input_tuple = xla::Parameter( + bodyb.get(), 0, tuple_shape, "TriangularSolveRightLookingWhileTuple"); + + // i, output, a, b = loop_carry + auto i = xla::GetTupleElement(input_tuple, 0); + auto body_out = xla::GetTupleElement(input_tuple, 1); + auto body_a = xla::GetTupleElement(input_tuple, 2); + auto body_b = xla::GetTupleElement(input_tuple, 3); + auto zero = xla::ConstantR0(bodyb.get(), 0); + + // result = b - np.matmul(output, a) + // result_row = result[..., :, i:i+1] + auto body_b_slice = DynamicSliceInMinorDims(body_b, {zero, i}, {m, 1}); + xla::XlaOp a_slice; + if (transpose_a) { + a_slice = DynamicSliceInMinorDims(body_a, {i, zero}, {1, n}); + } else { + a_slice = DynamicSliceInMinorDims(body_a, {zero, i}, {n, 1}); + } + auto b_update = body_b_slice - BatchDot(body_out, a_slice, + /*transpose_x=*/false, + /*transpose_y=*/transpose_a, + /*conjugate_x=*/false, + /*conjugate_y=*/conjugate_a); + + // body_out[..., :, i:i+1] = b_update / a[..., i:i+1, i:i+1] + auto a_ii = DynamicSliceInMinorDims(body_a, {i, i}, {1, 1}); + auto a_ii_conj = MaybeConjugate(a_ii, conjugate_a); + body_out = DynamicUpdateSliceInMinorDims(body_out, b_update / a_ii_conj, + {zero, i}); + + // if transpose_a: + // return (i + 1, body_out, a, b) + // else: + // return (i - 1, body_out, a, b) + auto next_i = xla::Add( + i, xla::ConstantR0(bodyb.get(), transpose_a ? 1 : -1)); + xla::Tuple(bodyb.get(), {next_i, body_out, body_a, body_b}); + } + TF_ASSIGN_OR_RETURN(auto body, bodyb->Build()); + + // Construct the While loop and return the result, + // return while_loop(cond_fun, body_fun, init)[1] + auto triangular_solve_left_looking_while = xla::While(cond, body, init); + return xla::GetTupleElement(triangular_solve_left_looking_while, 1); + }); } } // namespace tensorflow diff --git a/tensorflow/compiler/tf2xla/lib/triangular_solve.h b/tensorflow/compiler/tf2xla/lib/triangular_solve.h index 540c26b2473df9e7885f4e549b3e516a3d8a0d43..80c2bc4c9c38ec101db419d48db26e67e25d169b 100644 --- a/tensorflow/compiler/tf2xla/lib/triangular_solve.h +++ b/tensorflow/compiler/tf2xla/lib/triangular_solve.h @@ -57,23 +57,15 @@ namespace tensorflow { // // Uses a blocked algorithm if `block_size` is > 1; if block_size == 1 then no // blocking is used. -xla::StatusOr TriangularSolve(xla::XlaBuilder* builder, - const xla::XlaOp& a, xla::XlaOp b, - bool left_side, bool lower, - bool transpose_a, bool conjugate_a, - int64 block_size = 256); +xla::XlaOp TriangularSolve(xla::XlaOp a, xla::XlaOp b, bool left_side, + bool lower, bool transpose_a, bool conjugate_a, + int64 block_size = 256); -xla::StatusOr TriangularSolveLeftLooking(xla::XlaBuilder* builder, - const xla::XlaOp& a, - const xla::XlaOp& b, - bool transpose_a, - bool conjugate_a); +xla::XlaOp TriangularSolveLeftLooking(xla::XlaOp a, xla::XlaOp b, + bool transpose_a, bool conjugate_a); -xla::StatusOr TriangularSolveRightLooking(xla::XlaBuilder* builder, - const xla::XlaOp& a, - const xla::XlaOp& b, - bool transpose_a, - bool conjugate_a); +xla::XlaOp TriangularSolveRightLooking(xla::XlaOp a, xla::XlaOp b, + bool transpose_a, bool conjugate_a); } // namespace tensorflow diff --git a/tensorflow/compiler/tf2xla/lib/triangular_solve_test.cc b/tensorflow/compiler/tf2xla/lib/triangular_solve_test.cc index 87ea4763f7c2357ae179b68ade3715b24c46432f..f1bff6037bfa436e98a8dc7dbf6293b10b8d736f 100644 --- a/tensorflow/compiler/tf2xla/lib/triangular_solve_test.cc +++ b/tensorflow/compiler/tf2xla/lib/triangular_solve_test.cc @@ -21,7 +21,7 @@ limitations under the License. #include "tensorflow/compiler/xla/array2d.h" #include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/statusor.h" #include "tensorflow/compiler/xla/test.h" #include "tensorflow/compiler/xla/tests/client_library_test_base.h" @@ -85,11 +85,10 @@ XLA_TEST_F(TriangularSolveTest, SimpleRightLowerTranspose) { xla::XlaOp a, b; auto a_data = CreateR2Parameter(AValsLower(), 0, "a", &builder, &a); auto b_data = CreateR2Parameter(BValsRight(), 1, "b", &builder, &b); - auto result = TriangularSolve(&builder, a, b, - /*left_side=*/false, /*lower=*/true, - /*transpose_a=*/true, /*conjugate_a=*/false, - /*block_size=*/2); - TF_ASSERT_OK(result.status()); + TriangularSolve(a, b, + /*left_side=*/false, /*lower=*/true, + /*transpose_a=*/true, /*conjugate_a=*/false, + /*block_size=*/2); xla::Array2D expected({ {0.5, 0.08333334, 0.04629629, 0.03367003}, @@ -107,11 +106,10 @@ XLA_TEST_F(TriangularSolveTest, SimpleRightLowerNotranspose) { xla::XlaOp a, b; auto a_data = CreateR2Parameter(AValsLower(), 0, "a", &builder, &a); auto b_data = CreateR2Parameter(BValsRight(), 1, "b", &builder, &b); - auto result = TriangularSolve(&builder, a, b, - /*left_side=*/false, /*lower=*/true, - /*transpose_a=*/false, /*conjugate_a=*/false, - /*block_size=*/2); - TF_ASSERT_OK(result.status()); + TriangularSolve(a, b, + /*left_side=*/false, /*lower=*/true, + /*transpose_a=*/false, /*conjugate_a=*/false, + /*block_size=*/2); xla::Array2D expected({ {-0.16414141, -0.06902357, -0.07070707, 0.36363636}, @@ -129,11 +127,10 @@ XLA_TEST_F(TriangularSolveTest, SimpleRightUpperTranspose) { xla::XlaOp a, b; auto a_data = CreateR2Parameter(AValsUpper(), 0, "a", &builder, &a); auto b_data = CreateR2Parameter(BValsRight(), 1, "b", &builder, &b); - auto result = TriangularSolve(&builder, a, b, - /*left_side=*/false, /*lower=*/false, - /*transpose_a=*/true, /*conjugate_a=*/false, - /*block_size=*/2); - TF_ASSERT_OK(result.status()); + TriangularSolve(a, b, + /*left_side=*/false, /*lower=*/false, + /*transpose_a=*/true, /*conjugate_a=*/false, + /*block_size=*/2); xla::Array2D expected({ {-0.16414141, -0.06902357, -0.07070707, 0.36363636}, @@ -151,11 +148,10 @@ XLA_TEST_F(TriangularSolveTest, SimpleRightUpperNotranspose) { xla::XlaOp a, b; auto a_data = CreateR2Parameter(AValsUpper(), 0, "a", &builder, &a); auto b_data = CreateR2Parameter(BValsRight(), 1, "b", &builder, &b); - auto result = TriangularSolve(&builder, a, b, - /*left_side=*/false, /*lower=*/false, - /*transpose_a=*/false, /*conjugate_a=*/false, - /*block_size=*/2); - TF_ASSERT_OK(result.status()); + TriangularSolve(a, b, + /*left_side=*/false, /*lower=*/false, + /*transpose_a=*/false, /*conjugate_a=*/false, + /*block_size=*/2); xla::Array2D expected({ {0.5, 0.08333334, 0.04629629, 0.03367003}, @@ -173,11 +169,10 @@ XLA_TEST_F(TriangularSolveTest, SimpleLeftLowerTranspose) { xla::XlaOp a, b; auto a_data = CreateR2Parameter(AValsLower(), 0, "a", &builder, &a); auto b_data = CreateR2Parameter(BValsLeft(), 1, "b", &builder, &b); - auto result = TriangularSolve(&builder, a, b, - /*left_side=*/true, /*lower=*/true, - /*transpose_a=*/true, /*conjugate_a=*/false, - /*block_size=*/2); - TF_ASSERT_OK(result.status()); + TriangularSolve(a, b, + /*left_side=*/true, /*lower=*/true, + /*transpose_a=*/true, /*conjugate_a=*/false, + /*block_size=*/2); xla::Array2D expected({ {-0.89646465, -0.69444444, -0.49242424}, @@ -196,11 +191,10 @@ XLA_TEST_F(TriangularSolveTest, SimpleLeftLowerNotranspose) { xla::XlaOp a, b; auto a_data = CreateR2Parameter(AValsLower(), 0, "a", &builder, &a); auto b_data = CreateR2Parameter(BValsLeft(), 1, "b", &builder, &b); - auto result = TriangularSolve(&builder, a, b, - /*left_side=*/true, /*lower=*/true, - /*transpose_a=*/false, /*conjugate_a=*/false, - /*block_size=*/2); - TF_ASSERT_OK(result.status()); + TriangularSolve(a, b, + /*left_side=*/true, /*lower=*/true, + /*transpose_a=*/false, /*conjugate_a=*/false, + /*block_size=*/2); xla::Array2D expected({ {0.5, 1.0, 1.5}, @@ -219,11 +213,10 @@ XLA_TEST_F(TriangularSolveTest, SimpleLeftUpperTranspose) { xla::XlaOp a, b; auto a_data = CreateR2Parameter(AValsUpper(), 0, "a", &builder, &a); auto b_data = CreateR2Parameter(BValsLeft(), 1, "b", &builder, &b); - auto result = TriangularSolve(&builder, a, b, - /*left_side=*/true, /*lower=*/false, - /*transpose_a=*/true, /*conjugate_a=*/false, - /*block_size=*/2); - TF_ASSERT_OK(result.status()); + TriangularSolve(a, b, + /*left_side=*/true, /*lower=*/false, + /*transpose_a=*/true, /*conjugate_a=*/false, + /*block_size=*/2); xla::Array2D expected({ {0.5, 1.0, 1.5}, @@ -242,11 +235,10 @@ XLA_TEST_F(TriangularSolveTest, SimpleLeftUpperNotranspose) { xla::XlaOp a, b; auto a_data = CreateR2Parameter(AValsUpper(), 0, "a", &builder, &a); auto b_data = CreateR2Parameter(BValsLeft(), 1, "b", &builder, &b); - auto result = TriangularSolve(&builder, a, b, - /*left_side=*/true, /*lower=*/false, - /*transpose_a=*/false, /*conjugate_a=*/false, - /*block_size=*/2); - TF_ASSERT_OK(result.status()); + TriangularSolve(a, b, + /*left_side=*/true, /*lower=*/false, + /*transpose_a=*/false, /*conjugate_a=*/false, + /*block_size=*/2); xla::Array2D expected({ {-0.89646465, -0.69444444, -0.49242424}, @@ -267,11 +259,10 @@ XLA_TEST_F(TriangularSolveTest, SimpleRightLowerTransposeConjugate) { CreateR2Parameter(AValsLowerComplex(), 0, "a", &builder, &a); auto b_data = CreateR2Parameter(BValsRightComplex(), 1, "b", &builder, &b); - auto result = TriangularSolve(&builder, a, b, - /*left_side=*/false, /*lower=*/true, - /*transpose_a=*/true, /*conjugate_a=*/true, - /*block_size=*/2); - TF_ASSERT_OK(result.status()); + TriangularSolve(a, b, + /*left_side=*/false, /*lower=*/true, + /*transpose_a=*/true, /*conjugate_a=*/true, + /*block_size=*/2); xla::Array2D expected({ {0.5, complex64(0.08333333, 0.08333333), @@ -295,11 +286,10 @@ XLA_TEST_F(TriangularSolveTest, SimpleLeftUpperTransposeNoconjugate) { CreateR2Parameter(AValsUpperComplex(), 0, "a", &builder, &a); auto b_data = CreateR2Parameter(BValsLeftComplex(), 1, "b", &builder, &b); - auto result = TriangularSolve(&builder, a, b, - /*left_side=*/true, /*lower=*/false, - /*transpose_a=*/true, /*conjugate_a=*/false, - /*block_size=*/2); - TF_ASSERT_OK(result.status()); + TriangularSolve(a, b, + /*left_side=*/true, /*lower=*/false, + /*transpose_a=*/true, /*conjugate_a=*/false, + /*block_size=*/2); xla::Array2D expected({ {0.5, 1., 1.5}, @@ -323,10 +313,9 @@ XLA_TEST_F(TriangularSolveLeftLookingTest, Simple) { xla::XlaOp a, b; auto a_data = CreateR2Parameter(AValsLower(), 0, "a", &builder, &a); auto b_data = CreateR2Parameter(BValsLeft(), 1, "b", &builder, &b); - auto result = TriangularSolveLeftLooking(&builder, a, b, - /*transpose_a=*/false, - /*conjugate_a=*/false); - TF_ASSERT_OK(result.status()); + TriangularSolveLeftLooking(a, b, + /*transpose_a=*/false, + /*conjugate_a=*/false); xla::Array2D expected({ {0.5, 1.0, 1.5}, @@ -345,10 +334,9 @@ XLA_TEST_F(TriangularSolveLeftLookingTest, NonzeroUpperTriangle) { xla::XlaOp a, b; auto a_data = CreateR2Parameter(AValsFull(), 0, "a", &builder, &a); auto b_data = CreateR2Parameter(BValsLeft(), 1, "b", &builder, &b); - auto result = TriangularSolveLeftLooking(&builder, a, b, - /*transpose_a=*/false, - /*conjugate_a=*/false); - TF_ASSERT_OK(result.status()); + TriangularSolveLeftLooking(a, b, + /*transpose_a=*/false, + /*conjugate_a=*/false); xla::Array2D expected({ {0.5, 1.0, 1.5}, diff --git a/tensorflow/compiler/tf2xla/lib/util.cc b/tensorflow/compiler/tf2xla/lib/util.cc index d9ff7e6259f3fbab8957394bff5c5670a67dd0eb..a6f5d346cb5ecb85ff6b2306c2502ba31d74cc64 100644 --- a/tensorflow/compiler/tf2xla/lib/util.cc +++ b/tensorflow/compiler/tf2xla/lib/util.cc @@ -18,6 +18,8 @@ limitations under the License. #include #include +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/literal_util.h" #include "tensorflow/compiler/xla/shape_util.h" #include "tensorflow/compiler/xla/status_macros.h" @@ -28,8 +30,9 @@ limitations under the License. namespace tensorflow { xla::XlaOp Zeros(xla::XlaBuilder* builder, const xla::Shape& shape) { - return builder->Broadcast( - builder->ConstantLiteral(xla::Literal::Zero(shape.element_type())), + return xla::Broadcast( + xla::ConstantLiteral(builder, + xla::LiteralUtil::Zero(shape.element_type())), xla::AsInt64Slice(shape.dimensions())); } @@ -37,19 +40,19 @@ xla::XlaOp FloatLiteral(xla::XlaBuilder* builder, xla::PrimitiveType type, double value) { switch (type) { case xla::F16: - return builder->ConstantR0(static_cast(value)); + return xla::ConstantR0(builder, static_cast(value)); break; case xla::BF16: - return builder->ConstantR0(static_cast(value)); + return xla::ConstantR0(builder, static_cast(value)); break; case xla::F32: - return builder->ConstantR0(static_cast(value)); + return xla::ConstantR0(builder, static_cast(value)); break; case xla::F64: - return builder->ConstantR0(value); + return xla::ConstantR0(builder, value); break; case xla::C64: - return builder->ConstantR0(value); + return xla::ConstantR0(builder, value); break; default: LOG(FATAL) << "unhandled element type " << type; @@ -61,31 +64,31 @@ xla::XlaOp IntegerLiteral(xla::XlaBuilder* builder, xla::PrimitiveType type, xla::Literal literal; switch (type) { case xla::U8: - literal = std::move(*xla::Literal::CreateR0(value)); + literal = std::move(*xla::LiteralUtil::CreateR0(value)); break; case xla::U32: - literal = std::move(*xla::Literal::CreateR0(value)); + literal = std::move(*xla::LiteralUtil::CreateR0(value)); break; case xla::U64: - literal = std::move(*xla::Literal::CreateR0(value)); + literal = std::move(*xla::LiteralUtil::CreateR0(value)); break; case xla::S8: - literal = std::move(*xla::Literal::CreateR0(value)); + literal = std::move(*xla::LiteralUtil::CreateR0(value)); break; case xla::S32: - literal = std::move(*xla::Literal::CreateR0(value)); + literal = std::move(*xla::LiteralUtil::CreateR0(value)); break; case xla::S64: - literal = std::move(*xla::Literal::CreateR0(value)); + literal = std::move(*xla::LiteralUtil::CreateR0(value)); break; case xla::F32: - literal = std::move(*xla::Literal::CreateR0(value)); + literal = std::move(*xla::LiteralUtil::CreateR0(value)); break; case xla::F64: - literal = std::move(*xla::Literal::CreateR0(value)); + literal = std::move(*xla::LiteralUtil::CreateR0(value)); break; case xla::C64: - literal = std::move(*xla::Literal::CreateR0(value)); + literal = std::move(*xla::LiteralUtil::CreateR0(value)); break; case xla::PRED: LOG(FATAL) << "pred element type is not integral"; @@ -94,11 +97,11 @@ xla::XlaOp IntegerLiteral(xla::XlaBuilder* builder, xla::PrimitiveType type, LOG(FATAL) << "u16/s16 literals not yet implemented"; case xla::BF16: literal = std::move( - *xla::Literal::CreateR0(static_cast(value))); + *xla::LiteralUtil::CreateR0(static_cast(value))); break; case xla::F16: - literal = std::move( - *xla::Literal::CreateR0(static_cast(value))); + literal = std::move(*xla::LiteralUtil::CreateR0( + static_cast(value))); break; case xla::TUPLE: LOG(FATAL) << "tuple element type is not integral"; @@ -107,134 +110,140 @@ xla::XlaOp IntegerLiteral(xla::XlaBuilder* builder, xla::PrimitiveType type, default: LOG(FATAL) << "unhandled element type " << type; } - return builder->ConstantLiteral(literal); + return xla::ConstantLiteral(builder, literal); } -xla::StatusOr SliceInMinorDims(xla::XlaBuilder* builder, - const xla::XlaOp& x, - gtl::ArraySlice start, - gtl::ArraySlice end) { - TF_RET_CHECK(start.size() == end.size()); - int64 n_minor_dims = start.size(); - - TF_ASSIGN_OR_RETURN(xla::Shape shape, builder->GetShape(x)); - - const int64 n_dims = xla::ShapeUtil::Rank(shape); - TF_RET_CHECK(n_minor_dims <= n_dims); - gtl::ArraySlice major_dims(xla::AsInt64Slice(shape.dimensions()), - /*pos=*/0, - /*len=*/n_dims - n_minor_dims); - - // Prepends 0s in the major dim - std::vector padded_start(n_dims, 0); - std::copy(start.begin(), start.end(), - padded_start.begin() + major_dims.size()); - - // Prepends the shape of the major dims. - std::vector padded_end(n_dims); - std::copy(major_dims.begin(), major_dims.end(), padded_end.begin()); - std::copy(end.begin(), end.end(), padded_end.begin() + major_dims.size()); - - std::vector strides(n_dims, 1); - return builder->Slice(x, padded_start, padded_end, strides); +xla::XlaOp SliceInMinorDims(xla::XlaOp x, gtl::ArraySlice start, + gtl::ArraySlice end) { + xla::XlaBuilder* builder = x.builder(); + return builder->ReportErrorOrReturn([&]() -> xla::StatusOr { + TF_RET_CHECK(start.size() == end.size()); + int64 n_minor_dims = start.size(); + + TF_ASSIGN_OR_RETURN(xla::Shape shape, builder->GetShape(x)); + + const int64 n_dims = xla::ShapeUtil::Rank(shape); + TF_RET_CHECK(n_minor_dims <= n_dims); + gtl::ArraySlice major_dims(xla::AsInt64Slice(shape.dimensions()), + /*pos=*/0, + /*len=*/n_dims - n_minor_dims); + + // Prepends 0s in the major dim + std::vector padded_start(n_dims, 0); + std::copy(start.begin(), start.end(), + padded_start.begin() + major_dims.size()); + + // Prepends the shape of the major dims. + std::vector padded_end(n_dims); + std::copy(major_dims.begin(), major_dims.end(), padded_end.begin()); + std::copy(end.begin(), end.end(), padded_end.begin() + major_dims.size()); + + std::vector strides(n_dims, 1); + return xla::Slice(x, padded_start, padded_end, strides); + }); } -std::vector PrependMajorDims(xla::XlaBuilder* builder, - const gtl::ArraySlice& major_dims, - const gtl::ArraySlice& indices) { - std::vector output(indices.size() + major_dims.size()); - std::copy(major_dims.begin(), major_dims.end(), output.begin()); - std::copy(indices.begin(), indices.end(), output.begin() + major_dims.size()); +std::vector ConcatVectors(gtl::ArraySlice xs, + gtl::ArraySlice ys) { + std::vector output(xs.size() + ys.size()); + std::copy(xs.begin(), xs.end(), output.begin()); + std::copy(ys.begin(), ys.end(), output.begin() + xs.size()); return output; } -xla::StatusOr DynamicSliceInMinorDims( - xla::XlaBuilder* builder, const xla::XlaOp& x, - const std::vector& starts, - const gtl::ArraySlice& sizes) { - TF_ASSIGN_OR_RETURN(xla::Shape shape, builder->GetShape(x)); - const int64 n_dims = xla::ShapeUtil::Rank(shape); - int64 n_minor_dims = starts.size(); - TF_RET_CHECK(n_minor_dims == sizes.size()); - TF_RET_CHECK(n_minor_dims <= n_dims); - gtl::ArraySlice major_dims(xla::AsInt64Slice(shape.dimensions()), - /*pos=*/0, - /*len=*/n_dims - sizes.size()); - TF_ASSIGN_OR_RETURN(auto padded_starts, - PrependZerosInMajorDims(builder, x, starts)); - auto padded_sizes = PrependMajorDims(builder, major_dims, sizes); - return builder->DynamicSlice(x, padded_starts, padded_sizes); +xla::XlaOp DynamicSliceInMinorDims(xla::XlaOp x, + gtl::ArraySlice starts, + gtl::ArraySlice sizes) { + xla::XlaBuilder* builder = x.builder(); + return builder->ReportErrorOrReturn([&]() -> xla::StatusOr { + TF_ASSIGN_OR_RETURN(xla::Shape shape, builder->GetShape(x)); + const int64 n_dims = xla::ShapeUtil::Rank(shape); + int64 n_minor_dims = starts.size(); + TF_RET_CHECK(n_minor_dims == sizes.size()); + TF_RET_CHECK(n_minor_dims <= n_dims); + gtl::ArraySlice major_dims(xla::AsInt64Slice(shape.dimensions()), + /*pos=*/0, + /*len=*/n_dims - sizes.size()); + auto padded_starts = PrependZerosInMajorDims(x, starts); + auto padded_sizes = ConcatVectors(major_dims, sizes); + return xla::DynamicSlice(x, padded_starts, padded_sizes); + }); } -xla::StatusOr UpdateSlice(xla::XlaBuilder* builder, - const xla::XlaOp& x, - const xla::XlaOp& update, - gtl::ArraySlice start) { - // TODO(phawkins): make int64 work on all backends, remove the int32 cast. - std::vector start_as_int32(start.begin(), start.end()); - auto start_constant = builder->ConstantR1(start_as_int32); - TF_ASSIGN_OR_RETURN(xla::Shape shape, builder->GetShape(x)); - const int64 n_dims = xla::ShapeUtil::Rank(shape); - TF_ASSIGN_OR_RETURN(xla::Shape start_constant_shape, - builder->GetShape(start_constant)); - const int64 start_length = - xla::ShapeUtil::GetDimension(start_constant_shape, -1); - TF_RET_CHECK(start_length == n_dims); - return builder->DynamicUpdateSlice(x, update, start_constant); +xla::XlaOp UpdateSlice(xla::XlaOp x, xla::XlaOp update, + gtl::ArraySlice start) { + xla::XlaBuilder* builder = x.builder(); + return builder->ReportErrorOrReturn([&]() -> xla::StatusOr { + // TODO(phawkins): make int64 work on all backends, remove the int32 cast. + std::vector start_as_int32(start.begin(), start.end()); + auto start_constant = xla::ConstantR1(builder, start_as_int32); + TF_ASSIGN_OR_RETURN(xla::Shape shape, builder->GetShape(x)); + const int64 n_dims = xla::ShapeUtil::Rank(shape); + TF_ASSIGN_OR_RETURN(xla::Shape start_constant_shape, + builder->GetShape(start_constant)); + const int64 start_length = + xla::ShapeUtil::GetDimension(start_constant_shape, -1); + TF_RET_CHECK(start_length == n_dims); + return xla::DynamicUpdateSlice(x, update, start_constant); + }); } -xla::StatusOr UpdateSliceInMinorDims(xla::XlaBuilder* builder, - const xla::XlaOp& x, - const xla::XlaOp& update, - gtl::ArraySlice start) { - TF_ASSIGN_OR_RETURN(xla::Shape shape, builder->GetShape(x)); - const int64 n_dims = xla::ShapeUtil::Rank(shape); - const int64 n_minor_dims = start.size(); - TF_RET_CHECK(n_minor_dims <= n_dims); - std::vector padded_start(n_dims, 0); - std::copy(start.begin(), start.end(), - padded_start.begin() + (n_dims - n_minor_dims)); - return UpdateSlice(builder, x, update, padded_start); +xla::XlaOp UpdateSliceInMinorDims(xla::XlaOp x, xla::XlaOp update, + gtl::ArraySlice start) { + xla::XlaBuilder* builder = x.builder(); + return builder->ReportErrorOrReturn([&]() -> xla::StatusOr { + TF_ASSIGN_OR_RETURN(xla::Shape shape, builder->GetShape(x)); + const int64 n_dims = xla::ShapeUtil::Rank(shape); + const int64 n_minor_dims = start.size(); + TF_RET_CHECK(n_minor_dims <= n_dims); + std::vector padded_start(n_dims, 0); + std::copy(start.begin(), start.end(), + padded_start.begin() + (n_dims - n_minor_dims)); + return UpdateSlice(x, update, padded_start); + }); } -xla::StatusOr DynamicUpdateSliceInMinorDims( - xla::XlaBuilder* builder, const xla::XlaOp& x, const xla::XlaOp& update, - const std::vector& starts) { - TF_ASSIGN_OR_RETURN(auto padded_starts, - PrependZerosInMajorDims(builder, x, starts)); - return builder->DynamicUpdateSlice(x, update, padded_starts); +xla::XlaOp DynamicUpdateSliceInMinorDims(xla::XlaOp x, xla::XlaOp update, + gtl::ArraySlice starts) { + auto padded_starts = PrependZerosInMajorDims(x, starts); + return xla::DynamicUpdateSlice(x, update, padded_starts); } -xla::StatusOr PrependZerosInMajorDims( - xla::XlaBuilder* builder, const xla::XlaOp& x, - const std::vector& starts) { - TF_ASSIGN_OR_RETURN(xla::Shape shape, builder->GetShape(x)); - const int64 n_dims = xla::ShapeUtil::Rank(shape); - auto zero = builder->Reshape(builder->ConstantR0(0), {1}); - std::vector padded_starts(n_dims, zero); - for (int i = 0; i < starts.size(); ++i) { - padded_starts[n_dims - starts.size() + i] = - builder->Reshape(starts[i], {1}); - } - return builder->ConcatInDim(padded_starts, 0); +xla::XlaOp PrependZerosInMajorDims(xla::XlaOp x, + gtl::ArraySlice starts) { + xla::XlaBuilder* builder = x.builder(); + return builder->ReportErrorOrReturn([&]() -> xla::StatusOr { + TF_ASSIGN_OR_RETURN(xla::Shape shape, builder->GetShape(x)); + const int64 n_dims = xla::ShapeUtil::Rank(shape); + auto zero = xla::Reshape(xla::ConstantR0(builder, 0), {1}); + std::vector padded_starts(n_dims, zero); + for (int i = 0; i < starts.size(); ++i) { + padded_starts[n_dims - starts.size() + i] = xla::Reshape(starts[i], {1}); + } + return xla::ConcatInDim(builder, padded_starts, 0); + }); } -xla::StatusOr TransposeInMinorDims(xla::XlaBuilder* builder, - const xla::XlaOp& x) { - TF_ASSIGN_OR_RETURN(xla::Shape shape, builder->GetShape(x)); - const int64 n_dims = xla::ShapeUtil::Rank(shape); - TF_RET_CHECK(n_dims >= 2); - std::vector permutation(n_dims); - std::iota(permutation.begin(), permutation.end(), 0); - std::swap(permutation[n_dims - 1], permutation[n_dims - 2]); - return builder->Transpose(x, permutation); +xla::XlaOp TransposeInMinorDims(xla::XlaOp x) { + xla::XlaBuilder* builder = x.builder(); + return builder->ReportErrorOrReturn([&]() -> xla::StatusOr { + TF_ASSIGN_OR_RETURN(xla::Shape shape, builder->GetShape(x)); + const int64 n_dims = xla::ShapeUtil::Rank(shape); + TF_RET_CHECK(n_dims >= 2); + std::vector permutation(n_dims); + std::iota(permutation.begin(), permutation.end(), 0); + std::swap(permutation[n_dims - 1], permutation[n_dims - 2]); + return xla::Transpose(x, permutation); + }); } -xla::StatusOr MaybeConjugate(xla::XlaBuilder* builder, - const xla::XlaOp& x, bool conjugate) { - TF_ASSIGN_OR_RETURN(xla::Shape shape, builder->GetShape(x)); - auto perform_conj = shape.element_type() == xla::C64 && conjugate; - return perform_conj ? builder->Conj(x) : x; +xla::XlaOp MaybeConjugate(xla::XlaOp x, bool conjugate) { + xla::XlaBuilder* builder = x.builder(); + return builder->ReportErrorOrReturn([&]() -> xla::StatusOr { + TF_ASSIGN_OR_RETURN(xla::Shape shape, builder->GetShape(x)); + auto perform_conj = shape.element_type() == xla::C64 && conjugate; + return perform_conj ? xla::Conj(x) : x; + }); } } // namespace tensorflow diff --git a/tensorflow/compiler/tf2xla/lib/util.h b/tensorflow/compiler/tf2xla/lib/util.h index 3c120a2548576d6ad46870583ca65beea63507a3..6cb6c088e9d20af05193f0a3da6c2595966eb495 100644 --- a/tensorflow/compiler/tf2xla/lib/util.h +++ b/tensorflow/compiler/tf2xla/lib/util.h @@ -23,9 +23,6 @@ limitations under the License. namespace tensorflow { -// Returns a zero-filled tensor with shape `shape`. -xla::XlaOp Zeros(xla::XlaBuilder* builder, const xla::Shape& shape); - // Returns a floating point scalar constant of 'type' with 'value'. // If 'type' is complex, returns a real value with zero imaginary component. xla::XlaOp FloatLiteral(xla::XlaBuilder* builder, xla::PrimitiveType type, @@ -33,7 +30,7 @@ xla::XlaOp FloatLiteral(xla::XlaBuilder* builder, xla::PrimitiveType type, // Makes a 1D tensor [0, ..., x, y] from two tensors x and y with zeros // prepended until the array is length n_dims. -xla::XlaOp PrependZerosInMajorDims(xla::XlaBuilder* builder, +xla::XlaOp PrependZerosInMajorDims(xla::XlaOp x, gtl::ArraySlice starts); // Returns a integer scalar constant of 'type' with 'value'. @@ -41,54 +38,43 @@ xla::XlaOp PrependZerosInMajorDims(xla::XlaBuilder* builder, xla::XlaOp IntegerLiteral(xla::XlaBuilder* builder, xla::PrimitiveType type, int64 value); -// Builds a vector of zeros of length rank(x) with the last two values being +// Builds a vector of zeros of length rank(x) with the last values being // those in `starts`. -xla::StatusOr PrependZerosInMajorDims( - xla::XlaBuilder* builder, const xla::XlaOp& x, - const std::vector& starts); +xla::XlaOp PrependZerosInMajorDims(xla::XlaOp x, + gtl::ArraySlice starts); // Performs a slice in the minor dimensions of a Tensor. -xla::StatusOr SliceInMinorDims(xla::XlaBuilder* builder, - const xla::XlaOp& x, - gtl::ArraySlice start, - gtl::ArraySlice end); +xla::XlaOp SliceInMinorDims(xla::XlaOp x, gtl::ArraySlice start, + gtl::ArraySlice end); -// Builds a 1-d vector out of a concatenation of `major_dims` and `starts`. -std::vector PrependMajorDims(xla::XlaBuilder* builder, - const gtl::ArraySlice& major_dims, - const gtl::ArraySlice& indices); +// Returns the concatenation of `xs` and `ys`. +std::vector ConcatVectors(gtl::ArraySlice xs, + gtl::ArraySlice ys); // Performs a dynamic slice in the minor dimensions of a Tensor. -xla::StatusOr DynamicSliceInMinorDims( - xla::XlaBuilder* builder, const xla::XlaOp& x, - const std::vector& starts, const gtl::ArraySlice& sizes); +xla::XlaOp DynamicSliceInMinorDims(xla::XlaOp x, + gtl::ArraySlice starts, + gtl::ArraySlice sizes); // Updates a slice of 'x', i.e., // x[start[0], ..., start[n]] = update -xla::StatusOr UpdateSlice(xla::XlaBuilder* builder, - const xla::XlaOp& x, - const xla::XlaOp& update, - gtl::ArraySlice start); +xla::XlaOp UpdateSlice(xla::XlaOp x, xla::XlaOp update, + gtl::ArraySlice start); // Updates a slice of 'x', where 'start' contains a list of minor dimensions: // x[..., start[0], ..., start[n]] = update -xla::StatusOr UpdateSliceInMinorDims(xla::XlaBuilder* builder, - const xla::XlaOp& x, - const xla::XlaOp& update, - gtl::ArraySlice start); +xla::XlaOp UpdateSliceInMinorDims(xla::XlaOp x, xla::XlaOp update, + gtl::ArraySlice start); -xla::StatusOr DynamicUpdateSliceInMinorDims( - xla::XlaBuilder* builder, const xla::XlaOp& x, const xla::XlaOp& update, - const std::vector& starts); +xla::XlaOp DynamicUpdateSliceInMinorDims(xla::XlaOp x, xla::XlaOp update, + gtl::ArraySlice starts); // Transposes a stack of matrices `x` by swapping the last two dimensions. -xla::StatusOr TransposeInMinorDims(xla::XlaBuilder* builder, - const xla::XlaOp& x); +xla::XlaOp TransposeInMinorDims(xla::XlaOp x); // Applies a complex conjugation operation if `a` is complex and `conjugate_a` // is true, otherwise returns its argument. -xla::StatusOr MaybeConjugate(xla::XlaBuilder* builder, - const xla::XlaOp& x, bool conjugate); +xla::XlaOp MaybeConjugate(xla::XlaOp x, bool conjugate); } // namespace tensorflow diff --git a/tensorflow/compiler/tf2xla/lib/util_test.cc b/tensorflow/compiler/tf2xla/lib/util_test.cc index 265b39402c832f8c810a74f281563b05afdf2b1b..442fe92c34ca26cb1a854cc90da8dc034bca79bb 100644 --- a/tensorflow/compiler/tf2xla/lib/util_test.cc +++ b/tensorflow/compiler/tf2xla/lib/util_test.cc @@ -21,7 +21,7 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/lib/batch_dot.h" #include "tensorflow/compiler/xla/array2d.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/statusor.h" #include "tensorflow/compiler/xla/test.h" #include "tensorflow/compiler/xla/tests/client_library_test_base.h" @@ -70,8 +70,7 @@ XLA_TEST_F(UtilTest, Simple2dLookup) { auto a_data = CreateR2Parameter(BValsRight(), 0, "a", &builder, &a); auto x_data = CreateR0Parameter(2, 1, "x", &builder, &x); auto y_data = CreateR0Parameter(1, 2, "y", &builder, &y); - auto result = DynamicSliceInMinorDims(&builder, a, {x, y}, {1, 1}); - TF_ASSERT_OK(result.status()); + DynamicSliceInMinorDims(a, {x, y}, {1, 1}); ComputeAndCompareR2(&builder, {{10}}, {a_data.get(), x_data.get(), y_data.get()}, @@ -86,10 +85,8 @@ XLA_TEST_F(UtilTest, Simple3dLookup) { CreateR3Parameter(BatchedAValsFull(), 0, "a", &builder, &a); auto index_data = CreateR0Parameter(1, 1, "index", &builder, &index); - TF_ASSERT_OK_AND_ASSIGN( - auto l_index, - DynamicSliceInMinorDims(&builder, a, - {index, builder.ConstantR0(0)}, {1, 4})); + DynamicSliceInMinorDims(a, {index, xla::ConstantR0(&builder, 0)}, + {1, 4}); ComputeAndCompareR3(&builder, {{{3, 6, 0, 1}}, {{24, 61, 82, 48}}}, {a_data.get(), index_data.get()}); @@ -104,8 +101,7 @@ XLA_TEST_F(UtilTest, SimpleSliceUpdate) { auto x_data = CreateR0Parameter(2, 2, "x", &builder, &x); auto y_data = CreateR0Parameter(1, 3, "y", &builder, &y); - auto result = DynamicUpdateSliceInMinorDims(&builder, a, b, {x, y}); - TF_ASSERT_OK(result.status()); + DynamicUpdateSliceInMinorDims(a, b, {x, y}); xla::Array2D expected( {{{2, 0, 1, 2}, {3, 6, 0, 1}, {4, 9, 1, -10}, {5, 8, 10, 11}}}); @@ -128,13 +124,9 @@ XLA_TEST_F(UtilTest, RowBatchDot) { // Select {{3, 6, 0, 1}, {24, 61, 82, 48}} out of BatchedAValsFull(). auto index_data = CreateR0Parameter(1, 2, "index", &builder, &index); - TF_ASSERT_OK_AND_ASSIGN( - auto l_index, - DynamicSliceInMinorDims(&builder, a, - {index, builder.ConstantR0(0)}, {1, n})); - TF_ASSERT_OK_AND_ASSIGN( - auto dot, BatchDot(&builder, l_index, row, - /*transpose_x=*/false, /*transpose_y=*/true)); + auto l_index = DynamicSliceInMinorDims( + a, {index, xla::ConstantR0(&builder, 0)}, {1, n}); + BatchDot(l_index, row, /*transpose_x=*/false, /*transpose_y=*/true); ComputeAndCompareR3(&builder, {{{33}}, {{292}}}, {a_data.get(), row_data.get(), index_data.get()}); diff --git a/tensorflow/compiler/tf2xla/lib/while_loop.cc b/tensorflow/compiler/tf2xla/lib/while_loop.cc index 09ce594930efc0af47306590d76b322ac730f80f..574e70ddeeab8a3041cd730ce2717daec4f82ddf 100644 --- a/tensorflow/compiler/tf2xla/lib/while_loop.cc +++ b/tensorflow/compiler/tf2xla/lib/while_loop.cc @@ -15,6 +15,7 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/lib/while_loop.h" #include "tensorflow/compiler/tf2xla/lib/util.h" +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/compiler/xla/shape_util.h" #include "tensorflow/compiler/xla/status_macros.h" @@ -39,7 +40,7 @@ xla::StatusOr> XlaWhileLoop( xla::XlaBuilder* builder) { std::vector elements(arity); for (int i = 0; i < arity; ++i) { - elements[i] = builder->GetTupleElement(tuple, i); + elements[i] = xla::GetTupleElement(tuple, i); } return elements; }; @@ -48,7 +49,8 @@ xla::StatusOr> XlaWhileLoop( std::unique_ptr cond_builder = builder->CreateSubBuilder(strings::StrCat(name, "_condition")); { - auto parameter = cond_builder->Parameter(0, tuple_shape, "parameter"); + auto parameter = + xla::Parameter(cond_builder.get(), 0, tuple_shape, "parameter"); TF_RETURN_IF_ERROR( condition_function(unpack_tuple(parameter, arity, cond_builder.get()), @@ -61,7 +63,8 @@ xla::StatusOr> XlaWhileLoop( std::unique_ptr body_builder = builder->CreateSubBuilder(strings::StrCat(name, "_body")); { - auto parameter = body_builder->Parameter(0, tuple_shape, "parameter"); + auto parameter = + xla::Parameter(body_builder.get(), 0, tuple_shape, "parameter"); TF_ASSIGN_OR_RETURN( auto result, @@ -69,11 +72,11 @@ xla::StatusOr> XlaWhileLoop( body_builder.get())); TF_RET_CHECK(result.size() == initial_values.size()); - body_builder->Tuple(result); + xla::Tuple(body_builder.get(), result); } TF_ASSIGN_OR_RETURN(auto body, body_builder->Build()); - auto outputs = builder->While(cond, body, builder->Tuple(initial_values)); + auto outputs = xla::While(cond, body, xla::Tuple(builder, initial_values)); return unpack_tuple(outputs, arity, builder); } @@ -86,9 +89,8 @@ xla::StatusOr> XlaForEachIndex( auto while_cond_fn = [&](gtl::ArraySlice values, xla::XlaBuilder* cond_builder) -> xla::StatusOr { - return cond_builder->Lt( - values[0], - IntegerLiteral(cond_builder, num_iterations_type, num_iterations)); + return xla::Lt(values[0], IntegerLiteral(cond_builder, num_iterations_type, + num_iterations)); }; auto while_body_fn = [&](gtl::ArraySlice values, xla::XlaBuilder* body_builder) @@ -97,9 +99,10 @@ xla::StatusOr> XlaForEachIndex( std::vector updated_values; updated_values.reserve(values.size()); - updated_values.push_back(body_builder->Add( + updated_values.push_back(xla::Add( iteration, - body_builder->ConstantLiteral(xla::Literal::One(num_iterations_type)))); + xla::ConstantLiteral(body_builder, + xla::LiteralUtil::One(num_iterations_type)))); values.remove_prefix(1); TF_ASSIGN_OR_RETURN(std::vector body_outputs, @@ -111,8 +114,8 @@ xla::StatusOr> XlaForEachIndex( std::vector values; values.reserve(initial_values.size() + 1); - values.push_back( - builder->ConstantLiteral(xla::Literal::Zero(num_iterations_type))); + values.push_back(xla::ConstantLiteral( + builder, xla::LiteralUtil::Zero(num_iterations_type))); values.insert(values.end(), initial_values.begin(), initial_values.end()); TF_ASSIGN_OR_RETURN(values, XlaWhileLoop(while_cond_fn, while_body_fn, values, diff --git a/tensorflow/compiler/tf2xla/literal_util.cc b/tensorflow/compiler/tf2xla/literal_util.cc index b43405a1a407b5fa98dd740c62af91e048cc9490..2fb66913ada375d53512b9a1115326b3cc2afea4 100644 --- a/tensorflow/compiler/tf2xla/literal_util.cc +++ b/tensorflow/compiler/tf2xla/literal_util.cc @@ -17,7 +17,7 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/shape_util.h" #include "tensorflow/compiler/tf2xla/type_util.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/core/common_runtime/dma_helper.h" namespace tensorflow { diff --git a/tensorflow/compiler/tf2xla/literal_util.h b/tensorflow/compiler/tf2xla/literal_util.h index ab7e861f3336097d2ea52487092f16edb5c14531..0610a57029e72dff79a84742346f78a42b7f4ff1 100644 --- a/tensorflow/compiler/tf2xla/literal_util.h +++ b/tensorflow/compiler/tf2xla/literal_util.h @@ -18,7 +18,7 @@ limitations under the License. #ifndef TENSORFLOW_COMPILER_TF2XLA_LITERAL_UTIL_H_ #define TENSORFLOW_COMPILER_TF2XLA_LITERAL_UTIL_H_ -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/xla_data.pb.h" #include "tensorflow/core/framework/tensor.h" #include "tensorflow/core/lib/core/status.h" diff --git a/tensorflow/compiler/tf2xla/literal_util_test.cc b/tensorflow/compiler/tf2xla/literal_util_test.cc index f3d6787daaa1165b28ce63dfd501533fa0963edd..a3404c2b3df7bf25011359d1f5f5b88c29a3f83b 100644 --- a/tensorflow/compiler/tf2xla/literal_util_test.cc +++ b/tensorflow/compiler/tf2xla/literal_util_test.cc @@ -15,6 +15,7 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/literal_util.h" #include "tensorflow/core/framework/numeric_types.h" #include "tensorflow/core/framework/tensor_testutil.h" @@ -27,7 +28,7 @@ TEST(LiteralUtil, LiteralToHostTensor) { { std::vector int64_values = {1, 2, 3}; std::unique_ptr int64_values_literal = - xla::Literal::CreateR1(gtl::ArraySlice(int64_values)); + xla::LiteralUtil::CreateR1(gtl::ArraySlice(int64_values)); Tensor host_tensor; EXPECT_EQ("Cannot convert literal of type S64 to tensor of type int32", LiteralToHostTensor(*int64_values_literal, DT_INT32, &host_tensor) @@ -48,7 +49,7 @@ TEST(LiteralUtil, LiteralToHostTensor) { Tensor host_tensor; std::vector int32_values = {10, 11}; std::unique_ptr int32_values_literal = - xla::Literal::CreateR1(gtl::ArraySlice(int32_values)); + xla::LiteralUtil::CreateR1(gtl::ArraySlice(int32_values)); EXPECT_TRUE( LiteralToHostTensor(*int32_values_literal, DT_INT32, &host_tensor) .ok()); diff --git a/tensorflow/compiler/tf2xla/tf2xla_test.cc b/tensorflow/compiler/tf2xla/tf2xla_test.cc index 84c133ffabe20dbdaa4d5a64e035efb5e4c4c44b..f0b30dcf4e98faa8ab0801a8b0583744bdc669c7 100644 --- a/tensorflow/compiler/tf2xla/tf2xla_test.cc +++ b/tensorflow/compiler/tf2xla/tf2xla_test.cc @@ -18,6 +18,7 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/tf2xla.pb.h" #include "tensorflow/compiler/xla/client/client_library.h" #include "tensorflow/compiler/xla/client/local_client.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/literal_util.h" #include "tensorflow/compiler/xla/statusor.h" #include "tensorflow/core/framework/attr_value.pb.h" @@ -73,8 +74,8 @@ TEST(ConvertGraphDefToXla, Sum) { TF_EXPECT_OK(ConvertGraphDefToXla(graph_def, config, client, &computation)); // Set up arguments. - auto x_literal = xla::Literal::CreateR0(10); - auto y_literal = xla::Literal::CreateR0(32); + auto x_literal = xla::LiteralUtil::CreateR0(10); + auto y_literal = xla::LiteralUtil::CreateR0(32); auto x_global_or = client->TransferToServer(*x_literal); auto y_global_or = client->TransferToServer(*y_literal); TF_EXPECT_OK(x_global_or.status()); diff --git a/tensorflow/compiler/tf2xla/xla_compiler.cc b/tensorflow/compiler/tf2xla/xla_compiler.cc index 9c8e56a17e07348d3cfaaca0b5eb335295af05c3..319cbc74e96262881d32bdc9de2251b53f2b05d6 100644 --- a/tensorflow/compiler/tf2xla/xla_compiler.cc +++ b/tensorflow/compiler/tf2xla/xla_compiler.cc @@ -28,6 +28,7 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/xla_compilation_device.h" #include "tensorflow/compiler/tf2xla/xla_context.h" #include "tensorflow/compiler/xla/client/client_library.h" +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/core/common_runtime/device.h" #include "tensorflow/core/common_runtime/executor.h" #include "tensorflow/core/common_runtime/function.h" @@ -230,10 +231,13 @@ Status XlaCompiler::XLAShapeForArgument(const XlaCompiler::Argument& arg, case XlaCompiler::Argument::kConstant: LOG(FATAL) << "Unreachable case"; case XlaCompiler::Argument::kParameter: { - TensorShape shape = - is_entry_computation - ? options_.shape_representation_fn(arg.shape, arg.type) - : arg.shape; + TensorShape shape; + if (is_entry_computation) { + TF_ASSIGN_OR_RETURN( + shape, options_.shape_representation_fn(arg.shape, arg.type)); + } else { + shape = arg.shape; + } return TensorShapeToXLAShape(arg.type, shape, xla_shape); } case XlaCompiler::Argument::kResource: { @@ -241,8 +245,9 @@ Status XlaCompiler::XLAShapeForArgument(const XlaCompiler::Argument& arg, switch (arg.resource_kind) { case XlaResource::kVariable: { - TensorShape representation_shape = - options_.shape_representation_fn(arg.shape, arg.type); + TF_ASSIGN_OR_RETURN( + TensorShape representation_shape, + options_.shape_representation_fn(arg.shape, arg.type)); return TensorShapeToXLAShape(arg.type, representation_shape, xla_shape); } @@ -338,9 +343,9 @@ Status BuildComputation( const std::vector& arg_cores, const std::vector& retvals, const std::vector>& resources, - bool return_updated_values_for_all_resources, xla::XlaBuilder* builder, - xla::XlaComputation* computation, int* num_computation_outputs, - int* num_nonconst_outputs, + bool return_updated_values_for_all_resources, bool always_return_tuple, + xla::XlaBuilder* builder, xla::XlaComputation* computation, + int* num_computation_outputs, int* num_nonconst_outputs, std::vector* outputs, std::vector* resource_updates) { std::vector elems; @@ -384,13 +389,14 @@ Status BuildComputation( const XlaCompiler::Argument& arg = args[resource->arg_num()]; const int core = arg_cores[resource->arg_num()]; DCHECK_LT(resource->arg_num(), arg_cores.size()); - bool modified = resource->value() != resource->initial_value(); + bool modified = !resource->value().IsIdenticalTo(resource->initial_value()); // TensorArray gradients were modified if their values changed or there are // any newly created gradients. for (const auto& grad : resource->tensor_array_gradients()) { - modified = modified || - grad.second->value() != grad.second->initial_value() || - arg.tensor_array_gradients.count(grad.first) == 0; + modified = + modified || + !grad.second->value().IsIdenticalTo(grad.second->initial_value()) || + arg.tensor_array_gradients.count(grad.first) == 0; } if (return_updated_values_for_all_resources || modified) { resource_updates->emplace_back(); @@ -415,7 +421,7 @@ Status BuildComputation( // create a tuple/get-tuple-element combination so that sharding // assignment will be placed on this value, which will cause the resource // update to be returned from the same device that provided the resource. - handle = builder->GetTupleElement(builder->Tuple({handle}), 0); + handle = xla::GetTupleElement(xla::Tuple(builder, {handle}), 0); elems.push_back(handle); } @@ -424,7 +430,9 @@ Status BuildComputation( *num_computation_outputs = elems.size(); // Builds the XLA computation. - builder->Tuple(elems); + if (always_return_tuple || elems.size() != 1) { + xla::Tuple(builder, elems); + } builder->ClearOpMetadata(); xla::StatusOr computation_status = builder->Build(); @@ -551,16 +559,16 @@ Status XlaCompiler::BuildArguments( } xla::XlaScopedShardingAssignment assign_tuple_sharding(builder, tuple_sharding); - tuple = builder->Parameter(0, (*input_shapes)[0], "arg_tuple"); + tuple = xla::Parameter(builder, 0, (*input_shapes)[0], "arg_tuple"); } else { - tuple = builder->Parameter(0, (*input_shapes)[0], "arg_tuple"); + tuple = xla::Parameter(builder, 0, (*input_shapes)[0], "arg_tuple"); } for (std::vector::size_type i = 0; i < input_mapping->size(); ++i) { const int core = (*arg_cores)[input_mapping->at(i)]; xla::XlaScopedShardingAssignment assign_sharding( builder, core == -1 ? tensorflow::gtl::optional() : xla::sharding_builder::AssignDevice(core)); - arg_handles[i] = builder->GetTupleElement(tuple, i); + arg_handles[i] = xla::GetTupleElement(tuple, i); } } else { for (std::vector::size_type i = 0; i < input_mapping->size(); ++i) { @@ -568,8 +576,8 @@ Status XlaCompiler::BuildArguments( xla::XlaScopedShardingAssignment assign_sharding( builder, core == -1 ? tensorflow::gtl::optional() : xla::sharding_builder::AssignDevice(core)); - arg_handles[i] = - builder->Parameter(i, (*input_shapes)[i], strings::StrCat("arg", i)); + arg_handles[i] = xla::Parameter(builder, i, (*input_shapes)[i], + strings::StrCat("arg", i)); } } @@ -600,7 +608,7 @@ Status XlaCompiler::BuildArguments( // return values of functions, and then reshape unconditionally. if (is_entry_computation) { arg_expression.set_handle( - builder->Reshape(arg_handles[i], arg.shape.dim_sizes())); + xla::Reshape(arg_handles[i], arg.shape.dim_sizes())); } else { arg_expression.set_handle(arg_handles[i]); } @@ -660,20 +668,17 @@ Status XlaCompiler::CompileSingleOp( namespace { // Check that the ops of all non-functional nodes have been registered. -string ValidateFunctionDef(const FunctionDef* fdef, +Status ValidateFunctionDef(const FunctionDef* fdef, const FunctionLibraryDefinition& flib_def) { - std::vector invalid_ops; for (const NodeDef& node : fdef->node_def()) { const string& op = node.op(); if (op == FunctionLibraryDefinition::kGradientOp || flib_def.Find(op)) { continue; } const OpDef* op_def; - if (!OpRegistry::Global()->LookUpOpDef(op, &op_def).ok()) { - invalid_ops.push_back(op); - } + TF_RETURN_IF_ERROR(OpRegistry::Global()->LookUpOpDef(op, &op_def)); } - return tensorflow::str_util::Join(invalid_ops, ", "); + return Status::OK(); } // Check that the graph doesn't have any invalid nodes (e.g. incompatible with @@ -681,35 +686,33 @@ string ValidateFunctionDef(const FunctionDef* fdef, Status ValidateGraph(const Graph* graph, const FunctionLibraryDefinition& flib_def, const DeviceType& device_type, const string& name) { - std::vector invalid_ops; + auto maybe_error = [&](const string& op, const Status& s) -> Status { + if (!s.ok()) { + return errors::InvalidArgument(strings::StrCat( + "Detected unsupported operations when trying to compile graph ", name, + " on ", device_type.type_string(), ": ", op, " (", s.error_message(), + ")")); + } + return Status::OK(); + }; + for (const Node* node : graph->nodes()) { if (node->type_string() == FunctionLibraryDefinition::kGradientOp) { continue; } const FunctionDef* fdef = flib_def.Find(node->def().op()); + Status s; if (fdef) { - string error_msg = ValidateFunctionDef(fdef, flib_def); - if (!error_msg.empty()) { - invalid_ops.push_back( - strings::StrCat(node->def().op(), ":{", error_msg, "}")); - } + s = ValidateFunctionDef(fdef, flib_def); + TF_RETURN_IF_ERROR(maybe_error(node->def().op(), s)); continue; } const OpDef* op_def; - if (!OpRegistry::Global()->LookUpOpDef(node->def().op(), &op_def).ok()) { - invalid_ops.push_back(node->def().op()); - continue; - } + s = OpRegistry::Global()->LookUpOpDef(node->def().op(), &op_def); + TF_RETURN_IF_ERROR(maybe_error(node->def().op(), s)); TF_RETURN_IF_ERROR(ValidateNodeDef(node->def(), *op_def)); - if (!FindKernelDef(device_type, node->def(), nullptr, nullptr).ok()) { - invalid_ops.push_back(node->def().op()); - } - } - if (!invalid_ops.empty()) { - return errors::InvalidArgument(strings::StrCat( - "Detected unsupported operations when trying to compile graph ", name, - " on ", device_type.type_string(), ":", - tensorflow::str_util::Join(invalid_ops, ", "))); + s = FindKernelDef(device_type, node->def(), nullptr, nullptr); + TF_RETURN_IF_ERROR(maybe_error(node->def().op(), s)); } return Status::OK(); } @@ -767,9 +770,10 @@ Status XlaCompiler::CompileGraph(const XlaCompiler::CompileOptions& options, result->outputs.resize(context->retvals().size()); TF_RETURN_IF_ERROR(BuildComputation( args, arg_cores, context->retvals(), context->resources(), - options.return_updated_values_for_all_resources, &builder, - result->computation.get(), &num_computation_outputs, - &num_nonconst_outputs, &result->outputs, &result->resource_updates)); + options.return_updated_values_for_all_resources, + options.always_return_tuple, &builder, result->computation.get(), + &num_computation_outputs, &num_nonconst_outputs, &result->outputs, + &result->resource_updates)); VLOG(2) << "Outputs: total: " << context->retvals().size() << " nonconstant: " << num_nonconst_outputs; diff --git a/tensorflow/compiler/tf2xla/xla_compiler.h b/tensorflow/compiler/tf2xla/xla_compiler.h index 6be74957c6a92004ca9fdc97747d7b6cb693dd28..079c99797e1f1ec26205e33b3c7c16d3764f15ca 100644 --- a/tensorflow/compiler/tf2xla/xla_compiler.h +++ b/tensorflow/compiler/tf2xla/xla_compiler.h @@ -20,6 +20,7 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/xla_compilation_device.h" #include "tensorflow/compiler/tf2xla/xla_op_registry.h" #include "tensorflow/compiler/xla/client/local_client.h" +#include "tensorflow/compiler/xla/status_macros.h" #include "tensorflow/core/common_runtime/device.h" #include "tensorflow/core/common_runtime/device_mgr.h" #include "tensorflow/core/common_runtime/function.h" @@ -169,6 +170,11 @@ class XlaCompiler { // computation. bool resolve_compile_time_constants = true; + // If 'always_return_tuple' is true, then the output of a computation will + // always be a tuple. Otherwise, a single-element output will not be wrapped + // in a tuple. + bool always_return_tuple = true; + // True when compiling the entry computation, false for subcomputations // (while, call, etc.) bool is_entry_computation = true; @@ -237,7 +243,8 @@ class XlaCompiler { std::shared_ptr computation; }; - typedef std::function + typedef std::function(const TensorShape&, + DataType)> ShapeRepresentationFn; struct Options { // Name of the compilation device to use. It must be set by the caller. diff --git a/tensorflow/compiler/tf2xla/xla_compiler_test.cc b/tensorflow/compiler/tf2xla/xla_compiler_test.cc index 613230452b74755ce7543ec2ab82861aa0dfeb7a..6f76816a861f4e9c17aa7b9704986cffa911cc1c 100644 --- a/tensorflow/compiler/tf2xla/xla_compiler_test.cc +++ b/tensorflow/compiler/tf2xla/xla_compiler_test.cc @@ -23,7 +23,7 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/xla_op_registry.h" #include "tensorflow/compiler/xla/client/client_library.h" #include "tensorflow/compiler/xla/client/local_client.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/shape_util.h" #include "tensorflow/compiler/xla/status_macros.h" #include "tensorflow/compiler/xla/tests/literal_test_util.h" @@ -206,9 +206,9 @@ TEST_F(XlaCompilerTest, Simple) { // Tests that the generated computation works. std::unique_ptr param0_literal = - xla::Literal::CreateR1({7, 42}); + xla::LiteralUtil::CreateR1({7, 42}); std::unique_ptr param1_literal = - xla::Literal::CreateR1({-3, 101}); + xla::LiteralUtil::CreateR1({-3, 101}); std::unique_ptr param0_data = client_->TransferToServer(*param0_literal).ConsumeValueOrDie(); std::unique_ptr param1_data = @@ -222,9 +222,9 @@ TEST_F(XlaCompilerTest, Simple) { client_->Transfer(*actual).ConsumeValueOrDie(); std::unique_ptr expected0 = - xla::Literal::CreateR1({4, 143}); + xla::LiteralUtil::CreateR1({4, 143}); std::unique_ptr expected_literal = - xla::Literal::MakeTuple({expected0.get()}); + xla::LiteralUtil::MakeTuple({expected0.get()}); EXPECT_TRUE(xla::LiteralTestUtil::Equal(*expected_literal, *actual_literal)); } @@ -306,7 +306,7 @@ TEST_F(XlaCompilerTest, ConstantOutputs) { // Tests that the generated computation works. std::unique_ptr param0_literal = - xla::Literal::CreateR1({7, 42}); + xla::LiteralUtil::CreateR1({7, 42}); std::unique_ptr param0_data = client_->TransferToServer(*param0_literal).ConsumeValueOrDie(); @@ -317,9 +317,9 @@ TEST_F(XlaCompilerTest, ConstantOutputs) { client_->Transfer(*actual).ConsumeValueOrDie(); std::unique_ptr expected0 = - xla::Literal::CreateR1({-7, -42}); + xla::LiteralUtil::CreateR1({-7, -42}); std::unique_ptr expected_literal = - xla::Literal::MakeTuple({expected0.get()}); + xla::LiteralUtil::MakeTuple({expected0.get()}); EXPECT_TRUE( xla::LiteralTestUtil::Equal(*expected_literal, *actual_literal)); } @@ -341,7 +341,7 @@ TEST_F(XlaCompilerTest, ConstantOutputs) { // Tests that the generated computation works. std::unique_ptr param0_literal = - xla::Literal::CreateR1({7, 42}); + xla::LiteralUtil::CreateR1({7, 42}); std::unique_ptr param0_data = client_->TransferToServer(*param0_literal).ConsumeValueOrDie(); @@ -351,11 +351,12 @@ TEST_F(XlaCompilerTest, ConstantOutputs) { std::unique_ptr actual_literal = client_->Transfer(*actual).ConsumeValueOrDie(); - std::unique_ptr expected0 = xla::Literal::CreateR0(7); + std::unique_ptr expected0 = + xla::LiteralUtil::CreateR0(7); std::unique_ptr expected1 = - xla::Literal::CreateR1({-7, -42}); + xla::LiteralUtil::CreateR1({-7, -42}); std::unique_ptr expected = - xla::Literal::MakeTuple({expected0.get(), expected1.get()}); + xla::LiteralUtil::MakeTuple({expected0.get(), expected1.get()}); EXPECT_TRUE(xla::LiteralTestUtil::Equal(*expected, *actual_literal)); } } @@ -569,11 +570,11 @@ TEST_F(XlaCompilerTest, CanPassTensorArraysToAndFromComputation) { // Tests that the generated computation works. std::unique_ptr input_base = - xla::Literal::CreateR1({7, 42}); + xla::LiteralUtil::CreateR1({7, 42}); std::unique_ptr input_grad2 = - xla::Literal::CreateR1({-3, 101}); + xla::LiteralUtil::CreateR1({-3, 101}); std::unique_ptr input = - xla::Literal::MakeTuple({input_base.get(), input_grad2.get()}); + xla::LiteralUtil::MakeTuple({input_base.get(), input_grad2.get()}); std::unique_ptr param0_data = client_->TransferToServer(*input).ConsumeValueOrDie(); @@ -583,17 +584,18 @@ TEST_F(XlaCompilerTest, CanPassTensorArraysToAndFromComputation) { std::unique_ptr actual_literal = client_->Transfer(*actual).ConsumeValueOrDie(); - std::unique_ptr output_read = xla::Literal::CreateR0(42); + std::unique_ptr output_read = + xla::LiteralUtil::CreateR0(42); std::unique_ptr output_base = - xla::Literal::CreateR1({7, 42}); + xla::LiteralUtil::CreateR1({7, 42}); std::unique_ptr output_grad1 = - xla::Literal::CreateR1({0, 1}); + xla::LiteralUtil::CreateR1({0, 1}); std::unique_ptr output_grad2 = - xla::Literal::CreateR1({-3, 101}); - std::unique_ptr output_resource = xla::Literal::MakeTuple( + xla::LiteralUtil::CreateR1({-3, 101}); + std::unique_ptr output_resource = xla::LiteralUtil::MakeTuple( {output_base.get(), output_grad1.get(), output_grad2.get()}); std::unique_ptr expected_literal = - xla::Literal::MakeTuple({output_read.get(), output_resource.get()}); + xla::LiteralUtil::MakeTuple({output_read.get(), output_resource.get()}); EXPECT_TRUE(xla::LiteralTestUtil::Equal(*expected_literal, *actual_literal)); } @@ -796,9 +798,9 @@ TEST_F(XlaCompilerTest, Variables) { // Tests that the generated computation works. std::unique_ptr param0_literal = - xla::Literal::CreateR1({7, 42}); + xla::LiteralUtil::CreateR1({7, 42}); std::unique_ptr param1_literal = - xla::Literal::CreateR1({-3, 101}); + xla::LiteralUtil::CreateR1({-3, 101}); std::unique_ptr param0_data = client_->TransferToServer(*param0_literal).ConsumeValueOrDie(); std::unique_ptr param1_data = @@ -812,11 +814,11 @@ TEST_F(XlaCompilerTest, Variables) { client_->Transfer(*actual).ConsumeValueOrDie(); std::unique_ptr expected0 = - xla::Literal::CreateR1({5, 144}); + xla::LiteralUtil::CreateR1({5, 144}); std::unique_ptr expected1 = - xla::Literal::CreateR1({4, 143}); + xla::LiteralUtil::CreateR1({4, 143}); std::unique_ptr expected_literal = - xla::Literal::MakeTuple({expected0.get(), expected1.get()}); + xla::LiteralUtil::MakeTuple({expected0.get(), expected1.get()}); EXPECT_TRUE(xla::LiteralTestUtil::Equal(*expected_literal, *actual_literal)); } @@ -884,9 +886,9 @@ TEST_F(XlaCompilerTest, VariableRepresentationShapeFunction) { // Tests that the generated computation works. std::unique_ptr param0_literal = - xla::Literal::CreateR2({{4, 55}, {1, -3}}); + xla::LiteralUtil::CreateR2({{4, 55}, {1, -3}}); std::unique_ptr param1_literal = - xla::Literal::CreateR1({22, 11, 33, 404}); + xla::LiteralUtil::CreateR1({22, 11, 33, 404}); std::unique_ptr param0_data = client_->TransferToServer(*param0_literal).ConsumeValueOrDie(); std::unique_ptr param1_data = @@ -900,11 +902,11 @@ TEST_F(XlaCompilerTest, VariableRepresentationShapeFunction) { client_->Transfer(*actual).ConsumeValueOrDie(); std::unique_ptr expected0 = - xla::Literal::CreateR2({{27, 67}, {35, 402}}); + xla::LiteralUtil::CreateR2({{27, 67}, {35, 402}}); std::unique_ptr expected1 = - xla::Literal::CreateR1({26, 66, 34, 401}); + xla::LiteralUtil::CreateR1({26, 66, 34, 401}); std::unique_ptr expected_literal = - xla::Literal::MakeTuple({expected0.get(), expected1.get()}); + xla::LiteralUtil::MakeTuple({expected0.get(), expected1.get()}); EXPECT_TRUE(xla::LiteralTestUtil::Equal(*expected_literal, *actual_literal)); } @@ -953,9 +955,9 @@ TEST_F(XlaCompilerTest, ArgRetvalShapeRepresentationFunction) { // Tests that the generated computation works. std::unique_ptr param0_literal = - xla::Literal::CreateR1({4, 55, 1, -3}); + xla::LiteralUtil::CreateR1({4, 55, 1, -3}); std::unique_ptr param1_literal = - xla::Literal::CreateR1({22, 11, 33, 404}); + xla::LiteralUtil::CreateR1({22, 11, 33, 404}); std::unique_ptr param0_data = client_->TransferToServer(*param0_literal).ConsumeValueOrDie(); std::unique_ptr param1_data = @@ -969,11 +971,11 @@ TEST_F(XlaCompilerTest, ArgRetvalShapeRepresentationFunction) { client_->Transfer(*actual).ConsumeValueOrDie(); std::unique_ptr expected0 = - xla::Literal::CreateR1({27, 67, 35, 402}); + xla::LiteralUtil::CreateR1({27, 67, 35, 402}); std::unique_ptr expected1 = - xla::Literal::CreateR1({26, 66, 34, 401}); + xla::LiteralUtil::CreateR1({26, 66, 34, 401}); std::unique_ptr expected_literal = - xla::Literal::MakeTuple({expected0.get(), expected1.get()}); + xla::LiteralUtil::MakeTuple({expected0.get(), expected1.get()}); EXPECT_TRUE(xla::LiteralTestUtil::Equal(*expected_literal, *actual_literal)); } @@ -1021,8 +1023,7 @@ TEST_F(XlaCompilerTest, FunctionWithInvalidOp) { status = compiler.CompileGraph(XlaCompiler::CompileOptions(), "fill", std::move(graph), args, &result); ASSERT_FALSE(status.ok()); - EXPECT_TRUE( - str_util::StrContains(status.error_message(), "FillFn:{InvalidOp}")) + EXPECT_TRUE(str_util::StrContains(status.error_message(), "InvalidOp")) << status.error_message(); } diff --git a/tensorflow/compiler/tf2xla/xla_context.cc b/tensorflow/compiler/tf2xla/xla_context.cc index 67174b251d3acc381321a0097921fa5c695267fe..0dea366476954123ec09c020d493061fa2637c2f 100644 --- a/tensorflow/compiler/tf2xla/xla_context.cc +++ b/tensorflow/compiler/tf2xla/xla_context.cc @@ -27,7 +27,7 @@ limitations under the License. #include "tensorflow/compiler/xla/client/client_library.h" #include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/compiler/xla/layout_util.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/statusor.h" #include "tensorflow/core/common_runtime/dma_helper.h" #include "tensorflow/core/lib/gtl/array_slice.h" @@ -66,8 +66,8 @@ XlaContext::XlaContext( XlaCompiler* compiler, xla::XlaBuilder* builder, bool allow_cpu_custom_calls, bool resolve_compile_time_constants, bool is_entry_computation, - const std::function* - shape_representation_fn) + const std::function( + const TensorShape&, DataType)>* shape_representation_fn) : compiler_(compiler), builder_(builder), allow_cpu_custom_calls_(allow_cpu_custom_calls), @@ -119,8 +119,8 @@ Status XlaContext::CreateResource( return Status::OK(); } -TensorShape XlaContext::RepresentationShape(const TensorShape& shape, - DataType type) const { +xla::StatusOr XlaContext::RepresentationShape( + const TensorShape& shape, DataType type) const { return (*shape_representation_fn_)(shape, type); } @@ -131,9 +131,11 @@ const xla::XlaComputation* XlaContext::GetOrCreateMax(const DataType type) { xla::XlaBuilder b("max<" + type_string + ">"); xla::PrimitiveType xla_type; TF_CHECK_OK(DataTypeToPrimitiveType(type, &xla_type)); - auto x = b.Parameter(0, xla::ShapeUtil::MakeShape(xla_type, {}), "x"); - auto y = b.Parameter(1, xla::ShapeUtil::MakeShape(xla_type, {}), "y"); - b.Max(x, y); + auto x = + xla::Parameter(&b, 0, xla::ShapeUtil::MakeShape(xla_type, {}), "x"); + auto y = + xla::Parameter(&b, 1, xla::ShapeUtil::MakeShape(xla_type, {}), "y"); + xla::Max(x, y); return b.Build().ConsumeValueOrDie(); }); } @@ -145,9 +147,11 @@ const xla::XlaComputation* XlaContext::GetOrCreateMin(const DataType type) { xla::XlaBuilder b("min<" + type_string + ">"); xla::PrimitiveType xla_type; TF_CHECK_OK(DataTypeToPrimitiveType(type, &xla_type)); - auto x = b.Parameter(0, xla::ShapeUtil::MakeShape(xla_type, {}), "x"); - auto y = b.Parameter(1, xla::ShapeUtil::MakeShape(xla_type, {}), "y"); - b.Min(x, y); + auto x = + xla::Parameter(&b, 0, xla::ShapeUtil::MakeShape(xla_type, {}), "x"); + auto y = + xla::Parameter(&b, 1, xla::ShapeUtil::MakeShape(xla_type, {}), "y"); + xla::Min(x, y); return b.Build().ConsumeValueOrDie(); }); } @@ -159,9 +163,11 @@ const xla::XlaComputation* XlaContext::GetOrCreateAdd(const DataType type) { xla::XlaBuilder b("add<" + type_string + ">"); xla::PrimitiveType xla_type; TF_CHECK_OK(DataTypeToPrimitiveType(type, &xla_type)); - auto x = b.Parameter(0, xla::ShapeUtil::MakeShape(xla_type, {}), "x"); - auto y = b.Parameter(1, xla::ShapeUtil::MakeShape(xla_type, {}), "y"); - b.Add(x, y); + auto x = + xla::Parameter(&b, 0, xla::ShapeUtil::MakeShape(xla_type, {}), "x"); + auto y = + xla::Parameter(&b, 1, xla::ShapeUtil::MakeShape(xla_type, {}), "y"); + xla::Add(x, y); return b.Build().ConsumeValueOrDie(); }); } @@ -173,9 +179,11 @@ const xla::XlaComputation* XlaContext::GetOrCreateMul(const DataType type) { xla::XlaBuilder b("mul<" + type_string + ">"); xla::PrimitiveType xla_type; TF_CHECK_OK(DataTypeToPrimitiveType(type, &xla_type)); - auto x = b.Parameter(0, xla::ShapeUtil::MakeShape(xla_type, {}), "x"); - auto y = b.Parameter(1, xla::ShapeUtil::MakeShape(xla_type, {}), "y"); - b.Mul(x, y); + auto x = + xla::Parameter(&b, 0, xla::ShapeUtil::MakeShape(xla_type, {}), "x"); + auto y = + xla::Parameter(&b, 1, xla::ShapeUtil::MakeShape(xla_type, {}), "y"); + xla::Mul(x, y); return b.Build().ConsumeValueOrDie(); }); } diff --git a/tensorflow/compiler/tf2xla/xla_context.h b/tensorflow/compiler/tf2xla/xla_context.h index 5960daaefd625a0b4daf00d7b8c929f3c856575f..38d8cd653cbbe5b01325d6b478589d88909bac56 100644 --- a/tensorflow/compiler/tf2xla/xla_context.h +++ b/tensorflow/compiler/tf2xla/xla_context.h @@ -24,6 +24,7 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/xla_compiler.h" #include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/compiler/xla/client/xla_client/xla_computation.h" +#include "tensorflow/compiler/xla/status_macros.h" #include "tensorflow/compiler/xla/xla_data.pb.h" #include "tensorflow/core/framework/op_kernel.h" #include "tensorflow/core/framework/resource_mgr.h" @@ -47,8 +48,8 @@ class XlaContext : public ResourceBase { XlaContext(XlaCompiler* compiler, xla::XlaBuilder* builder, bool allow_cpu_custom_calls, bool resolve_compile_time_constants, bool is_entry_computation, - const std::function* - shape_representation_fn); + const std::function( + const TensorShape&, DataType)>* shape_representation_fn); // Virtual method defined by ResourceBase. string DebugString() override; @@ -101,8 +102,8 @@ class XlaContext : public ResourceBase { // Returns the XLA shape to be used to represent a variable of TF `shape` // and `type`, or of an argument or return value of a top-level computation. - TensorShape RepresentationShape(const TensorShape& shape, - DataType type) const; + xla::StatusOr RepresentationShape(const TensorShape& shape, + DataType type) const; // Get an XLA lambda to compute Max. This is cached in the // XlaContext since it may be used by multiple Ops. There is a @@ -160,7 +161,7 @@ class XlaContext : public ResourceBase { // should be represented in XLA. Parameters/return values will be shaped // according to this function, and reshaped back to/from their declared shapes // for computations. Must be non-null. - const std::function* + const std::function(const TensorShape&, DataType)>* shape_representation_fn_; // Cache of prebuilt computations indexed by their type. diff --git a/tensorflow/compiler/tf2xla/xla_cpu_backend.cc b/tensorflow/compiler/tf2xla/xla_cpu_backend.cc index ead229aaccc292d4944db0c1eaf98c82583533cd..23d04d43b358e858ad1ab2463322ce0ab93b23c2 100644 --- a/tensorflow/compiler/tf2xla/xla_cpu_backend.cc +++ b/tensorflow/compiler/tf2xla/xla_cpu_backend.cc @@ -31,6 +31,10 @@ bool CpuOpFilter(KernelDef* kdef) { DT_FLOAT); return true; } + // TODO(b/26783907): The CPU backend currently does not implement sort. + if (kdef->op() == "XlaSort" || kdef->op() == "TopKV2") { + return false; + } if (kdef->op() == "Const") { AddDtypeToKernalDefConstraint("dtype", DT_STRING, kdef); } diff --git a/tensorflow/compiler/tf2xla/xla_gpu_backend.cc b/tensorflow/compiler/tf2xla/xla_gpu_backend.cc index 62168b648331844bfe2db1a4d5dcad895c8726f3..dc98d4fda6ae21411065981a7b7383ef0ad50f44 100644 --- a/tensorflow/compiler/tf2xla/xla_gpu_backend.cc +++ b/tensorflow/compiler/tf2xla/xla_gpu_backend.cc @@ -13,6 +13,7 @@ See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ +#include "tensorflow/compiler/tf2xla/legacy_flags/backend_registration_flags.h" #include "tensorflow/compiler/tf2xla/tf2xla_util.h" #include "tensorflow/compiler/tf2xla/xla_op_registry.h" #include "tensorflow/core/framework/kernel_def.pb.h" @@ -22,8 +23,16 @@ namespace tensorflow { bool GpuOpFilter(KernelDef* kdef) { // TODO(b/31361304): The GPU backend does not parallelize PRNG ops, leading to // slow code. - if (kdef->op() == "RandomStandardNormal" || kdef->op() == "RandomUniform" || - kdef->op() == "RandomUniformInt" || kdef->op() == "TruncatedNormal") { + legacy_flags::BackendRegistrationFlags* flags = + legacy_flags::GetBackendRegistrationFlags(); + VLOG(2) << "flags->tf_enable_prng_ops_gpu: " << flags->tf_enable_prng_ops_gpu; + if (!flags->tf_enable_prng_ops_gpu && + (kdef->op() == "RandomStandardNormal" || kdef->op() == "RandomUniform" || + kdef->op() == "RandomUniformInt" || kdef->op() == "TruncatedNormal")) { + return false; + } + // TODO(b/26783907): The GPU backend currently does not implement sort. + if (kdef->op() == "XlaSort" || kdef->op() == "TopKV2") { return false; } if (kdef->op() == "Const") { diff --git a/tensorflow/compiler/tf2xla/xla_helpers.cc b/tensorflow/compiler/tf2xla/xla_helpers.cc index 93cd340485649b5c55fd6771d24dd9e79989a1f5..4d1b3b1a135c493b3c2fa95967d2c4768ecfa63f 100644 --- a/tensorflow/compiler/tf2xla/xla_helpers.cc +++ b/tensorflow/compiler/tf2xla/xla_helpers.cc @@ -23,6 +23,9 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/type_util.h" #include "tensorflow/compiler/tf2xla/xla_context.h" #include "tensorflow/compiler/tf2xla/xla_op_kernel.h" +#include "tensorflow/compiler/xla/client/lib/arithmetic.h" +#include "tensorflow/compiler/xla/client/lib/constants.h" +#include "tensorflow/compiler/xla/client/lib/numeric.h" #include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/compiler/xla/types.h" #include "tensorflow/core/framework/tensor.h" @@ -33,103 +36,71 @@ namespace tensorflow { namespace { -Status ArgMinMax(xla::XlaBuilder* builder, XlaOpKernelContext* ctx, - const xla::XlaOp& input, const TensorShape& input_shape, - DataType input_type, DataType output_type, int axis, - bool is_min, xla::XlaOp* argminmax) { - xla::XlaOp init_value; - const xla::XlaComputation* reducer; - if (is_min) { - init_value = XlaHelpers::MaxValue(builder, input_type); - reducer = ctx->GetOrCreateMin(input_type); - } else { - init_value = XlaHelpers::MinValue(builder, input_type); - reducer = ctx->GetOrCreateMax(input_type); - } - - xla::PrimitiveType xla_output_type; - TF_RETURN_IF_ERROR(DataTypeToPrimitiveType(output_type, &xla_output_type)); - - xla::XlaOp input_max = builder->Reduce(input, init_value, *reducer, - /*dimensions_to_reduce=*/{axis}); - std::vector broadcast_dims(input_shape.dims() - 1); - std::iota(broadcast_dims.begin(), broadcast_dims.begin() + axis, 0); - std::iota(broadcast_dims.begin() + axis, broadcast_dims.end(), axis + 1); - // Compute a mask that has 1s for elements equal to the maximum. - xla::XlaOp partial_mask = builder->ConvertElementType( - builder->Eq(input, input_max, broadcast_dims), xla_output_type); - - // In order to make identity elements for a bitwise And, we: - // Left shift the 1 to the leftmost bit, yielding 0x10...0 - // Arithmetic right shift the 1 back to the rightmost bit, yielding - // 0xFF...F - int32 bits_in_type = - xla::ShapeUtil::ByteSizeOfPrimitiveType(xla_output_type) * 8 - 1; - xla::XlaOp shift_amount = - XlaHelpers::IntegerLiteral(builder, output_type, bits_in_type); - xla::XlaOp full_mask = builder->ShiftRightArithmetic( - builder->ShiftLeft(partial_mask, shift_amount), shift_amount); - - // And with the vector [0, 1, 2, ...] to convert each 0xFF...F into its - // index. - xla::XlaOp iota; - - const int64 axis_size = input_shape.dim_size(axis); - TF_RETURN_IF_ERROR(XlaHelpers::Iota(builder, output_type, axis_size, &iota)); - xla::XlaOp product = - builder->And(full_mask, iota, /*broadcast_dimensions=*/{axis}); - - // If there are multiple maximum elements, choose the one with the highest - // index. - xla::XlaOp output = - builder->Reduce(product, XlaHelpers::MinValue(builder, output_type), - *ctx->GetOrCreateMax(output_type), - /*dimensions_to_reduce=*/{axis}); - *argminmax = output; - return Status::OK(); +xla::XlaOp ArgMinMax(xla::XlaOp input, xla::PrimitiveType output_type, int axis, + bool is_min) { + xla::XlaBuilder* builder = input.builder(); + return builder->ReportErrorOrReturn([&]() -> xla::StatusOr { + TF_ASSIGN_OR_RETURN(xla::Shape input_shape, builder->GetShape(input)); + xla::XlaOp init_value; + xla::XlaComputation reducer; + if (is_min) { + init_value = xla::MaxValue(builder, input_shape.element_type()); + reducer = + xla::CreateScalarMinComputation(input_shape.element_type(), builder); + } else { + init_value = xla::MinValue(builder, input_shape.element_type()); + reducer = + xla::CreateScalarMaxComputation(input_shape.element_type(), builder); + } + + xla::XlaOp input_max = xla::Reduce(input, init_value, reducer, + /*dimensions_to_reduce=*/{axis}); + std::vector broadcast_dims(xla::ShapeUtil::Rank(input_shape) - 1); + std::iota(broadcast_dims.begin(), broadcast_dims.begin() + axis, 0); + std::iota(broadcast_dims.begin() + axis, broadcast_dims.end(), axis + 1); + // Compute a mask that has 1s for elements equal to the maximum. + xla::XlaOp partial_mask = xla::ConvertElementType( + xla::Eq(input, input_max, broadcast_dims), output_type); + + // In order to make identity elements for a bitwise And, we: + // Left shift the 1 to the leftmost bit, yielding 0x10...0 + // Arithmetic right shift the 1 back to the rightmost bit, yielding + // 0xFF...F + int32 bits_in_type = + xla::ShapeUtil::ByteSizeOfPrimitiveType(output_type) * 8 - 1; + xla::XlaOp shift_amount = + xla::ConstantR0WithType(builder, output_type, bits_in_type); + xla::XlaOp full_mask = xla::ShiftRightArithmetic( + xla::ShiftLeft(partial_mask, shift_amount), shift_amount); + + // And with the vector [0, 1, 2, ...] to convert each 0xFF...F into its + // index. + + const int64 axis_size = xla::ShapeUtil::GetDimension(input_shape, axis); + xla::XlaOp iota = xla::Iota(builder, output_type, axis_size); + xla::XlaOp product = + xla::And(full_mask, iota, /*broadcast_dimensions=*/{axis}); + + // If there are multiple maximum elements, choose the one with the highest + // index. + return xla::Reduce(product, xla::MinValue(builder, output_type), + xla::CreateScalarMaxComputation(output_type, builder), + /*dimensions_to_reduce=*/{axis}); + }); } } // namespace -xla::XlaOp XlaHelpers::MinValue(xla::XlaBuilder* b, DataType data_type) { - xla::PrimitiveType type; - TF_CHECK_OK(DataTypeToPrimitiveType(data_type, &type)); - return b->ConstantLiteral(xla::Literal::MinValue(type)); -} - -xla::XlaOp XlaHelpers::MaxValue(xla::XlaBuilder* b, DataType data_type) { - xla::PrimitiveType type; - TF_CHECK_OK(DataTypeToPrimitiveType(data_type, &type)); - return b->ConstantLiteral(xla::Literal::MaxValue(type)); -} - xla::XlaOp XlaHelpers::Zero(xla::XlaBuilder* b, DataType data_type) { xla::PrimitiveType type; TF_CHECK_OK(DataTypeToPrimitiveType(data_type, &type)); - return b->ConstantLiteral(xla::Literal::Zero(type)); + return xla::ConstantLiteral(b, xla::LiteralUtil::Zero(type)); } xla::XlaOp XlaHelpers::One(xla::XlaBuilder* b, DataType data_type) { xla::PrimitiveType type; TF_CHECK_OK(DataTypeToPrimitiveType(data_type, &type)); - return b->ConstantLiteral(xla::Literal::One(type)); -} - -xla::XlaOp XlaHelpers::Epsilon(xla::XlaBuilder* b, DataType data_type) { - switch (data_type) { - case DT_HALF: - return b->ConstantR0( - static_cast(Eigen::NumTraits::epsilon())); - case DT_BFLOAT16: - return b->ConstantR0(bfloat16::epsilon()); - case DT_FLOAT: - return b->ConstantR0(std::numeric_limits::epsilon()); - case DT_DOUBLE: - return b->ConstantR0(std::numeric_limits::epsilon()); - default: - LOG(FATAL) << "Unsupported type in XlaHelpers::Epsilon: " - << DataTypeString(data_type); - } + return xla::ConstantLiteral(b, xla::LiteralUtil::One(type)); } xla::XlaOp XlaHelpers::IntegerLiteral(xla::XlaBuilder* b, DataType data_type, @@ -177,45 +148,14 @@ static Tensor MakeLinspaceTensor(const TensorShape& shape, int64 depth) { return linspace; } -Status XlaHelpers::ArgMax(xla::XlaBuilder* builder, XlaOpKernelContext* ctx, - const xla::XlaOp& input, - const TensorShape& input_shape, DataType input_type, - DataType output_type, int axis, xla::XlaOp* argmax) { - return ArgMinMax(builder, ctx, input, input_shape, input_type, output_type, - axis, /*is_min=*/false, argmax); -} - -Status XlaHelpers::ArgMin(xla::XlaBuilder* builder, XlaOpKernelContext* ctx, - const xla::XlaOp& input, - const TensorShape& input_shape, DataType input_type, - DataType output_type, int axis, xla::XlaOp* argmin) { - return ArgMinMax(builder, ctx, input, input_shape, input_type, output_type, - axis, /*is_min=*/true, argmin); +xla::XlaOp XlaHelpers::ArgMax(xla::XlaOp input, xla::PrimitiveType output_type, + int axis) { + return ArgMinMax(input, output_type, axis, /*is_min=*/false); } -Status XlaHelpers::Iota(xla::XlaBuilder* builder, DataType dtype, int64 size, - xla::XlaOp* iota) { - TensorShape linspace_shape({size}); - Tensor linspace; - switch (dtype) { - case DT_UINT8: - linspace = MakeLinspaceTensor(linspace_shape, size); - break; - case DT_INT32: - linspace = MakeLinspaceTensor(linspace_shape, size); - break; - case DT_INT64: - linspace = MakeLinspaceTensor(linspace_shape, size); - break; - default: - return errors::InvalidArgument("Invalid argument type ", - DataTypeString(dtype)); - } - xla::BorrowingLiteral linspace_literal; - TF_RETURN_IF_ERROR(HostTensorToBorrowingLiteral(linspace, &linspace_literal)); - - *iota = builder->ConstantLiteral(linspace_literal); - return Status::OK(); +xla::XlaOp XlaHelpers::ArgMin(xla::XlaOp input, xla::PrimitiveType output_type, + int axis) { + return ArgMinMax(input, output_type, axis, /*is_min=*/true); } Status XlaHelpers::OneHot(xla::XlaBuilder* builder, int64 depth, int axis, @@ -256,17 +196,19 @@ Status XlaHelpers::OneHot(xla::XlaBuilder* builder, int64 depth, int axis, std::vector broadcast_dims(indices_shape.dims()); std::iota(broadcast_dims.begin(), broadcast_dims.begin() + axis, 0); std::iota(broadcast_dims.begin() + axis, broadcast_dims.end(), axis + 1); - xla::XlaOp one_hot_bool = builder->Eq( - indices, builder->ConstantLiteral(linspace_literal), broadcast_dims); + xla::XlaOp one_hot_bool = xla::Eq( + indices, xla::ConstantLiteral(builder, linspace_literal), broadcast_dims); // Selects the user-provided off_value and on_value values. - *one_hot = builder->Select( - one_hot_bool, builder->Broadcast(on_value, output_shape.dim_sizes()), - builder->Broadcast(off_value, output_shape.dim_sizes())); + *one_hot = xla::Select(one_hot_bool, + xla::Broadcast(on_value, output_shape.dim_sizes()), + xla::Broadcast(off_value, output_shape.dim_sizes())); return Status::OK(); } DataType XlaHelpers::SumAccumulationType(const DataType& dtype) { + // Upcast 16 bit sum reductions to 32 bit to reduce the precision loss from + // repeated floating point additions. if (dtype == DT_BFLOAT16 || dtype == DT_HALF) { return DT_FLOAT; } @@ -278,7 +220,7 @@ xla::XlaOp XlaHelpers::ConvertElementType(xla::XlaBuilder* const builder, const DataType new_element_type) { xla::PrimitiveType convert_to; TF_CHECK_OK(DataTypeToPrimitiveType(new_element_type, &convert_to)); - return builder->ConvertElementType(operand, convert_to); + return xla::ConvertElementType(operand, convert_to); } } // end namespace tensorflow diff --git a/tensorflow/compiler/tf2xla/xla_helpers.h b/tensorflow/compiler/tf2xla/xla_helpers.h index c3fdc5252e74363fe289eeabb2cb0d68298ee291..d6ca4ab9346593892917e8375b07a8790dc26e79 100644 --- a/tensorflow/compiler/tf2xla/xla_helpers.h +++ b/tensorflow/compiler/tf2xla/xla_helpers.h @@ -28,14 +28,6 @@ namespace tensorflow { // Helper methods for building XLA computations. class XlaHelpers { public: - // Returns a handle representing the minimum value of a scalar - // element of data_type. - static xla::XlaOp MinValue(xla::XlaBuilder* b, DataType data_type); - - // Returns a handle representing the maximum value of a scalar - // element of data_type. - static xla::XlaOp MaxValue(xla::XlaBuilder* b, DataType data_type); - // Returns a handle representing the zero value of a scalar // element of data_type. static xla::XlaOp Zero(xla::XlaBuilder* b, DataType data_type); @@ -44,10 +36,6 @@ class XlaHelpers { // element of data_type. static xla::XlaOp One(xla::XlaBuilder* b, DataType data_type); - // Returns the machine epsilon for floating-point type `data_type`, i.e., - // the difference between 1.0 and the next representable value. - static xla::XlaOp Epsilon(xla::XlaBuilder* b, DataType data_type); - // Returns a handle representing the given value of an integer scalar // element of data_type. // Note that unlike One and Zero, does not work on boolean types. @@ -65,25 +53,15 @@ class XlaHelpers { gtl::ArraySlice shape, xla::Literal* output); - // Sets `argmax` to the argmax of `input` along `axis`. `input_shape` and - // `input_dtype` are the shape and dtype of `input` respectively, and - // `output_type` is the dtype to use for `argmax`. - static Status ArgMax(xla::XlaBuilder* builder, XlaOpKernelContext* ctx, - const xla::XlaOp& input, const TensorShape& input_shape, - DataType input_type, DataType output_type, int axis, - xla::XlaOp* argmax); - - // Sets `argmin` to the argmin of `input` along `axis`. `input_shape` and - // `input_dtype` are the shape and dtype of `input` respectively, and - // `output_type` is the dtype to use for `argmin`. - static Status ArgMin(xla::XlaBuilder* builder, XlaOpKernelContext* ctx, - const xla::XlaOp& input, const TensorShape& input_shape, - DataType input_type, DataType output_type, int axis, - xla::XlaOp* argmin); - - // Sets *iota to a rank 1 tensor with values [0, 1, 2, ...] of `dtype`. - static Status Iota(xla::XlaBuilder* builder, DataType dtype, int64 size, - xla::XlaOp* iota); + // Returns the argmax of `input` along `axis`. `output_type` is the type to + // use for the output. + static xla::XlaOp ArgMax(xla::XlaOp input, xla::PrimitiveType output_type, + int axis); + + // Returns the argmin of `input` along `axis`. `output_type` is the type to + // use for the output. + static xla::XlaOp ArgMin(xla::XlaOp input, xla::PrimitiveType output_type, + int axis); // Converts `indices` into a one-hot representation. `depth` is the size // of the new axis to add. `axis` is the position at which to add the new diff --git a/tensorflow/compiler/tf2xla/xla_op_kernel.cc b/tensorflow/compiler/tf2xla/xla_op_kernel.cc index b58959bd6c3ffbd669a633d0f2560f9fd996e7ad..e8eafb3819f420bf4fd5be2c3930b3da75be58d6 100644 --- a/tensorflow/compiler/tf2xla/xla_op_kernel.cc +++ b/tensorflow/compiler/tf2xla/xla_op_kernel.cc @@ -19,7 +19,10 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/literal_util.h" #include "tensorflow/compiler/tf2xla/shape_util.h" +#include "tensorflow/compiler/tf2xla/type_util.h" #include "tensorflow/compiler/tf2xla/xla_context.h" +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" +#include "tensorflow/compiler/xla/status_macros.h" #include "tensorflow/core/common_runtime/dma_helper.h" namespace tensorflow { @@ -63,10 +66,32 @@ const xla::XlaOp& XlaOpKernelContext::Input(int index) { return GetComputationFromTensor(context_->input(index)); } +const xla::XlaOp& XlaOpKernelContext::Input(StringPiece name) { + return GetComputationFromTensor(GetInputTensorByName(name)); +} + TensorShape XlaOpKernelContext::InputShape(int index) { return context_->input(index).shape(); } +TensorShape XlaOpKernelContext::InputShape(StringPiece name) { + return GetInputTensorByName(name).shape(); +} + +DataType XlaOpKernelContext::input_type(int index) const { + return context_->input(index).dtype(); +} + +xla::PrimitiveType XlaOpKernelContext::input_xla_type(int index) { + xla::PrimitiveType type; + Status status = DataTypeToPrimitiveType(input_type(index), &type); + if (!status.ok()) { + SetStatus(status); + return xla::PRIMITIVE_TYPE_INVALID; + } + return type; +} + Status XlaOpKernelContext::ConstantInput(int index, xla::Literal* constant_literal) { return ConstantInputReshaped( @@ -128,7 +153,7 @@ Status XlaOpKernelContext::ConstantInputReshaped( xla::XlaOp handle = expression->handle(); if (new_shape != tensor.shape()) { // Reshape the handle to the desired shape. - handle = builder()->Reshape(handle, new_shape.dim_sizes()); + handle = xla::Reshape(handle, new_shape.dim_sizes()); } // The XLA layout is specified minor to major, and TensorFlow's minor @@ -315,10 +340,11 @@ Status XlaOpKernelContext::ConstantInputList( return Status::OK(); } -Status XlaOpKernelContext::ReadVariableInput(int index, DataType type, - TensorShape* shape, - xla::XlaOp* value) { - const Tensor& tensor = context_->input(index); +namespace { + +Status ReadVariableInputTensor(const Tensor& tensor, DataType type, + const OpKernelContext* ctx, TensorShape* shape, + xla::XlaOp* value) { const XlaExpression* expression = CastExpressionFromTensor(tensor); XlaResource* variable = expression->resource(); TF_RET_CHECK(variable != nullptr); @@ -336,18 +362,34 @@ Status XlaOpKernelContext::ReadVariableInput(int index, DataType type, *shape = variable->shape(); } - XlaContext& xla_context = XlaContext::Get(context_); - TensorShape representation_shape = - xla_context.RepresentationShape(variable->shape(), variable->type()); + XlaContext& xla_context = XlaContext::Get(ctx); + TF_ASSIGN_OR_RETURN( + TensorShape representation_shape, + xla_context.RepresentationShape(variable->shape(), variable->type())); if (representation_shape == variable->shape()) { *value = variable->value(); } else { - *value = - builder()->Reshape(variable->value(), variable->shape().dim_sizes()); + *value = xla::Reshape(variable->value(), variable->shape().dim_sizes()); } return Status::OK(); } +} // namespace + +Status XlaOpKernelContext::ReadVariableInput(int index, DataType type, + TensorShape* shape, + xla::XlaOp* value) { + return ReadVariableInputTensor(context_->input(index), type, context_, shape, + value); +} + +Status XlaOpKernelContext::ReadVariableInput(StringPiece name, DataType type, + TensorShape* shape, + xla::XlaOp* value) { + return ReadVariableInputTensor(GetInputTensorByName(name), type, context_, + shape, value); +} + Status XlaOpKernelContext::GetVariableTypeAndShape(int index, DataType* type, TensorShape* shape) const { const Tensor& tensor = context_->input(index); @@ -394,7 +436,7 @@ void XlaOpKernelContext::SetConstantOutput(int index, const Tensor& constant) { xla::BorrowingLiteral literal; OP_REQUIRES_OK(context_, HostTensorToBorrowingLiteral(constant, &literal)); - xla::XlaOp handle = builder()->ConstantLiteral(literal); + xla::XlaOp handle = xla::ConstantLiteral(builder(), literal); CHECK(handle.valid()); // Make the Tensor that will refer to the expression. @@ -438,17 +480,17 @@ Status XlaOpKernelContext::GetResourceInput(int index, XlaResource** resource) { return Status::OK(); } -Status XlaOpKernelContext::AssignVariable(int input_index, DataType type, - xla::XlaOp handle) { - TF_RET_CHECK(handle.valid()); +namespace { - const XlaExpression* expression = - CastExpressionFromTensor(context_->input(input_index)); +Status AssignVariableTensor(const Tensor& tensor, DataType type, + const OpKernelContext* ctx, xla::XlaOp handle, + xla::XlaBuilder* builder) { + const XlaExpression* expression = CastExpressionFromTensor(tensor); XlaResource* variable = expression->resource(); TF_RET_CHECK(variable != nullptr); TF_RET_CHECK(variable->kind() == XlaResource::kVariable); - auto shape_or_status = builder()->GetShape(handle); + auto shape_or_status = builder->GetShape(handle); if (!shape_or_status.ok()) { return shape_or_status.status(); } @@ -458,15 +500,31 @@ Status XlaOpKernelContext::AssignVariable(int input_index, DataType type, TF_RETURN_IF_ERROR(variable->SetTypeAndShape(type, shape)); - XlaContext& xla_context = XlaContext::Get(context_); - TensorShape representation_shape = - xla_context.RepresentationShape(shape, type); + XlaContext& xla_context = XlaContext::Get(ctx); + TF_ASSIGN_OR_RETURN(TensorShape representation_shape, + xla_context.RepresentationShape(shape, type)); if (shape != representation_shape) { - handle = builder()->Reshape(handle, representation_shape.dim_sizes()); + handle = xla::Reshape(handle, representation_shape.dim_sizes()); } return variable->SetValue(handle); } +} // namespace + +Status XlaOpKernelContext::AssignVariable(int input_index, DataType type, + xla::XlaOp handle) { + TF_RET_CHECK(handle.valid()); + return AssignVariableTensor(context_->input(input_index), type, context_, + handle, builder()); +} + +Status XlaOpKernelContext::AssignVariable(StringPiece name, DataType type, + xla::XlaOp handle) { + TF_RET_CHECK(handle.valid()); + return AssignVariableTensor(GetInputTensorByName(name), type, context_, + handle, builder()); +} + XlaCompiler* XlaOpKernelContext::compiler() const { return XlaContext::Get(context_).compiler(); } @@ -506,6 +564,12 @@ const xla::XlaComputation* XlaOpKernelContext::GetOrCreateMul( return XlaContext::Get(context_).GetOrCreateMul(type); } +const Tensor& XlaOpKernelContext::GetInputTensorByName(StringPiece name) { + const Tensor* tensor; + CHECK(context_->input(name, &tensor).ok()); + return *tensor; +} + XlaOpKernel::XlaOpKernel(OpKernelConstruction* context) : OpKernel(context) {} void XlaOpKernel::Compute(OpKernelContext* context) { diff --git a/tensorflow/compiler/tf2xla/xla_op_kernel.h b/tensorflow/compiler/tf2xla/xla_op_kernel.h index 667dc262ca03ca716ffbf015a78fc14c7a8b7c1a..6203cffd806a94743e2df7adc1e52d0255ab7c06 100644 --- a/tensorflow/compiler/tf2xla/xla_op_kernel.h +++ b/tensorflow/compiler/tf2xla/xla_op_kernel.h @@ -18,6 +18,7 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/xla_compiler.h" #include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" +#include "tensorflow/compiler/xla/xla_data.pb.h" #include "tensorflow/core/framework/op_kernel.h" #include "tensorflow/core/platform/macros.h" @@ -66,16 +67,26 @@ class XlaOpKernelContext { // Returns the number of inputs to the operator. int num_inputs() const { return context_->num_inputs(); } - // Returns the type of input 'index'. - DataType input_type(int index) { return context_->input(index).dtype(); } + // Returns the type of input `index`. + DataType input_type(int index) const; - // Returns the shape of input 'index'. + // Returns the type of input `index` as an xla::PrimitiveType. If the type + // is not representable as an XLA type, sets an error status and returns + // xla::PRIMITIVE_TYPE_INVALID. + xla::PrimitiveType input_xla_type(int index); + + // Returns the shape of input `index`. TensorShape InputShape(int index); - // Returns input 'index' as a XlaOp. Unlike + // Returns the shape of input `name`. + TensorShape InputShape(StringPiece name); + + // Returns input `index` as a XlaOp. Unlike // OpKernelContext::Input returns a symbolic value rather than a concrete // Tensor. const xla::XlaOp& Input(int index); + // Returns input `name` as a XlaOp. + const xla::XlaOp& Input(StringPiece name); // Returns true if all inputs are the same shape, otherwise sets the // status to a non-OK value and returns false. @@ -90,13 +101,13 @@ class XlaOpKernelContext { // Helper methods for constant inputs. - // Evaluates input 'index' and stores it in '*constant_literal'. If the + // Evaluates input `index` and stores it in `*constant_literal`. If the // expression cannot be evaluated, e.g., because it depends on unbound // parameters, returns a non-OK status. Status ConstantInput(int index, xla::Literal* constant_literal); - // Evaluates input 'index', reshapes it to 'new_shape' if new_shape != - // InputShape(index), and stores it in '*constant_literal'. If the input + // Evaluates input `index`, reshapes it to `new_shape` if new_shape != + // InputShape(index), and stores it in `*constant_literal`. If the input // cannot be evaluated, e.g., because it depends on unbound parameters, // returns a non-Ok status. If InputShape(index).num_elements() != // new_shape.num_elements(), returns an error status. @@ -131,17 +142,17 @@ class XlaOpKernelContext { return context_->expected_output_dtype(index); } - // Sets output 'index' to the XlaOp 'handle'. + // Sets output `index` to the XlaOp `handle`. // All outputs should be set using SetOutput and SetConstantOutput, not // via the underlying OpKernelContext. void SetOutput(int index, const xla::XlaOp& handle); - // Sets output 'index' to compile-time constant 'host_tensor', where - // 'host_tensor' is a tensor in host memory. It is preferable to use + // Sets output `index` to compile-time constant `host_tensor`, where + // `host_tensor` is a tensor in host memory. It is preferable to use // SetConstantOutput where possible. void SetConstantOutput(int index, const Tensor& host_tensor); - // Sets output 'index' to an invalid value. + // Sets output `index` to an invalid value. // Any subsequent attempt to consume this output will cause an error. void SetInvalidOutput(int index); @@ -151,10 +162,10 @@ class XlaOpKernelContext { // Variables - // Sets '*resource' to the resource associated with input `index`. + // Sets `*resource` to the resource associated with input `index`. Status GetResourceInput(int index, XlaResource** resource); - // Sets output 'index' to be a reference to resource 'resource'. + // Sets output `index` to be a reference to resource `resource`. void SetResourceOutput(int index, XlaResource* resource); // Sets `*type` and `*shape` to the current type and shape of a variable's @@ -163,17 +174,23 @@ class XlaOpKernelContext { TensorShape* shape) const; // Reads the current value of the resouce variable referred to by input - // 'index'. If `shape` is not nullptr, sets `*shape` to the shape of the + // `index`. If `shape` is not nullptr, sets `*shape` to the shape of the // variable. Returns an error if the variable has not been initialized, or if // its type does not match `type`. Status ReadVariableInput(int index, DataType type, TensorShape* shape, xla::XlaOp* value); + // Reads the current value of the resouce variable referred to by input + // `name`. + Status ReadVariableInput(StringPiece name, DataType type, TensorShape* shape, + xla::XlaOp* value); // Assigns the value `handle` to the variable referenced by input // `input_index`. The variable must be of `type`. Returns an error if the // variable has been initialized with a different type or with a // different shape. Status AssignVariable(int input_index, DataType type, xla::XlaOp handle); + // Assigns the value `handle` to the variable referenced by input `name`. + Status AssignVariable(StringPiece name, DataType type, xla::XlaOp handle); // Helper routines for the OP_REQUIRES macros void CtxFailure(const Status& s); @@ -221,6 +238,9 @@ class XlaOpKernelContext { const xla::XlaComputation* GetOrCreateMul(const DataType type); private: + // Returns the tensor of input `name`. + const Tensor& GetInputTensorByName(StringPiece name); + OpKernelContext* const context_; }; diff --git a/tensorflow/compiler/tf2xla/xla_op_registry.cc b/tensorflow/compiler/tf2xla/xla_op_registry.cc index ee6da6a67a70441bc1cb3a164a623fa389ed03cb..46785bc1f0a1279bfd67a55844fe238d9797382b 100644 --- a/tensorflow/compiler/tf2xla/xla_op_registry.cc +++ b/tensorflow/compiler/tf2xla/xla_op_registry.cc @@ -240,6 +240,7 @@ void XlaOpRegistry::RegisterCompilationKernels() { // a) the types supported by the backend, and // b) the types allowed by the OpDef, and // c) the type constraints. + bool unsatisfiable_type_constraint = false; for (const string& type_attr : type_attrs) { KernelDef::AttrConstraint* attr_constraint = kdef->add_constraint(); attr_constraint->set_name(type_attr); @@ -276,7 +277,14 @@ void XlaOpRegistry::RegisterCompilationKernels() { if (op_registration->allow_resource_types) { allowed_values->add_type(DT_RESOURCE); } + // Don't build KernelDefs that have unsatisfiable type constraints. + if (allowed_values->type().empty()) { + unsatisfiable_type_constraint = true; + break; + } } + if (unsatisfiable_type_constraint) continue; + if (backend.second.op_filter != nullptr && !backend.second.op_filter(kdef.get())) { continue; diff --git a/tensorflow/compiler/tf2xla/xla_op_registry_test.cc b/tensorflow/compiler/tf2xla/xla_op_registry_test.cc index 266cbc43951d7b4c427a782894bc9bc109c4814e..7b3b15b1af7636fddd4c29477cbfe6f9761f2c47 100644 --- a/tensorflow/compiler/tf2xla/xla_op_registry_test.cc +++ b/tensorflow/compiler/tf2xla/xla_op_registry_test.cc @@ -82,5 +82,38 @@ TEST(XlaOpRegistryTest, XlaOpRegistrationWithOverride) { } } +// A dummy generic OpKernel for all backends. +class DummyInfeasibleTypeConstraintOp : public XlaOpKernel { + public: + explicit DummyInfeasibleTypeConstraintOp(OpKernelConstruction* ctx) + : XlaOpKernel(ctx) {} + void Compile(XlaOpKernelContext* ctx) override { + LOG(FATAL) << "unreachable"; + } +}; + +REGISTER_OP("DummyInfeasibleTypeConstraintOp") + .Attr("T: {float, string}") + .Input("input: T") + .Output("output: T") + .Doc(R"doc( +A dummy Op. + +input: dummy input. +output: dummy output. +)doc"); +REGISTER_XLA_OP( + Name("DummyInfeasibleTypeConstraintOp").TypeConstraint("T", DT_STRING), + DummyInfeasibleTypeConstraintOp); + +TEST(XlaOpRegistryTest, OpWithInfeasibleTypeConstraintIsNotRegistered) { + XlaOpRegistry::RegisterCompilationKernels(); + auto registered_kernels = GetAllRegisteredKernels().kernel(); + for (const auto& kernels : registered_kernels) { + // The operator should not be registered. + EXPECT_NE(kernels.op(), "DummyInfeasibleTypeConstraintOp"); + } +} + } // namespace } // namespace tensorflow diff --git a/tensorflow/compiler/tf2xla/xla_resource.cc b/tensorflow/compiler/tf2xla/xla_resource.cc index 540c65c597f20d5bb26494e56c09ff2187cfb0db..baea8149658ec0849ebb570931ca68518ec5284e 100644 --- a/tensorflow/compiler/tf2xla/xla_resource.cc +++ b/tensorflow/compiler/tf2xla/xla_resource.cc @@ -22,6 +22,7 @@ limitations under the License. #include "tensorflow/compiler/tf2xla/sharding_util.h" #include "tensorflow/compiler/tf2xla/xla_context.h" #include "tensorflow/compiler/tf2xla/xla_helpers.h" +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" namespace tensorflow { @@ -89,16 +90,16 @@ Status XlaResource::SetZeroValue(xla::XlaBuilder* builder) { } switch (kind_) { case kVariable: { - value_ = builder->Broadcast(XlaHelpers::Zero(builder, type_), - shape_.dim_sizes()); + value_ = + xla::Broadcast(XlaHelpers::Zero(builder, type_), shape_.dim_sizes()); break; } case kTensorArray: { TensorShape ta_shape; ta_shape.AddDim(tensor_array_size_); ta_shape.AppendShape(shape_); - value_ = builder->Broadcast(XlaHelpers::Zero(builder, type_), - ta_shape.dim_sizes()); + value_ = xla::Broadcast(XlaHelpers::Zero(builder, type_), + ta_shape.dim_sizes()); break; } case kStack: { @@ -106,9 +107,9 @@ Status XlaResource::SetZeroValue(xla::XlaBuilder* builder) { ta_shape.AddDim(tensor_array_size_); ta_shape.AppendShape(shape_); value_ = - builder->Tuple({builder->Broadcast(XlaHelpers::Zero(builder, type_), - ta_shape.dim_sizes()), - builder->ConstantR0(0)}); + xla::Tuple(builder, {xla::Broadcast(XlaHelpers::Zero(builder, type_), + ta_shape.dim_sizes()), + xla::ConstantR0(builder, 0)}); break; } @@ -130,8 +131,8 @@ Status XlaResource::GetOrCreateTensorArrayGradient(const string& source, TensorShape ta_shape; ta_shape.AddDim(tensor_array_size_); ta_shape.AppendShape(shape_); - xla::XlaOp gradient_value = builder->Broadcast( - XlaHelpers::Zero(builder, type_), ta_shape.dim_sizes()); + xla::XlaOp gradient_value = + xla::Broadcast(XlaHelpers::Zero(builder, type_), ta_shape.dim_sizes()); gradient.reset( new XlaResource(/*kind=*/kTensorArray, /*arg_num=*/-1, /*name=*/strings::StrCat("TensorArrayGrad: ", name_), @@ -152,7 +153,7 @@ Status XlaResource::Pack(xla::XlaOp* pack, xla::XlaBuilder* builder) const { for (const auto& gradient : tensor_array_gradients_) { elems.push_back(gradient.second->value_); } - *pack = builder->Tuple(elems); + *pack = xla::Tuple(builder, elems); } return Status::OK(); } @@ -168,7 +169,7 @@ Status XlaResource::SetFromPack(const std::set& gradient_sources, } else { TF_RET_CHECK(kind_ == kTensorArray); int pos = 0; - auto v = builder->GetTupleElement(pack, pos++); + auto v = xla::GetTupleElement(pack, pos++); if (!initialized()) { initial_value_ = v; } @@ -178,7 +179,7 @@ Status XlaResource::SetFromPack(const std::set& gradient_sources, XlaResource* gradient; TF_RETURN_IF_ERROR( GetOrCreateTensorArrayGradient(source, builder, &gradient)); - auto v = builder->GetTupleElement(pack, pos++); + auto v = xla::GetTupleElement(pack, pos++); if (!gradient->initialized()) { gradient->initial_value_ = v; } diff --git a/tensorflow/compiler/xla/BUILD b/tensorflow/compiler/xla/BUILD index 4525197146b7f29f405650bdb08e5946cbce8114..f1c383fd9e3fff8a306ba0ddcc3f9ee42c63d66a 100644 --- a/tensorflow/compiler/xla/BUILD +++ b/tensorflow/compiler/xla/BUILD @@ -142,30 +142,15 @@ cc_library( cc_library( name = "statusor", - srcs = ["statusor.cc"], hdrs = [ "statusor.h", - "statusor_internals.h", ], visibility = ["//visibility:public"], deps = [ ":status", "//tensorflow/core:lib", "//tensorflow/core:lib_internal", - ], -) - -tf_cc_test( - name = "statusor_test", - size = "small", - srcs = ["statusor_test.cc"], - deps = [ - ":statusor", - ":test", - ":types", - "//tensorflow/core:lib", - "//tensorflow/core:test", - "//tensorflow/core:test_main", + "//tensorflow/stream_executor", ], ) @@ -175,6 +160,7 @@ cc_library( hdrs = [ "iterator_util.h", "map_util.h", + "overflow_util.h", "ptr_util.h", "util.h", ], @@ -250,7 +236,7 @@ cc_library( ":types", ":util", ":xla_data_proto", - "//tensorflow/core:framework_internal", + "//tensorflow/core:framework", "//tensorflow/core:lib", "//tensorflow/core:lib_internal", "//tensorflow/core:regexp_internal", @@ -268,6 +254,7 @@ tf_cc_test( ":types", ":util", ":xla_data_proto", + "//tensorflow/core:lib", "//tensorflow/core:test_main", ], ) @@ -295,9 +282,9 @@ tf_cc_test( ) cc_library( - name = "literal_util", - srcs = ["literal_util.cc"], - hdrs = ["literal_util.h"], + name = "literal", + srcs = ["literal.cc"], + hdrs = ["literal.h"], visibility = ["//visibility:public"], deps = [ ":array2d", @@ -314,11 +301,12 @@ cc_library( ) tf_cc_test( - name = "literal_util_test", - srcs = ["literal_util_test.cc"], + name = "literal_test", + srcs = ["literal_test.cc"], deps = [ ":array3d", ":array4d", + ":literal", ":literal_util", ":shape_util", ":test", @@ -330,6 +318,26 @@ tf_cc_test( ], ) +cc_library( + name = "literal_util", + srcs = ["literal_util.cc"], + hdrs = ["literal_util.h"], + visibility = ["//visibility:public"], + deps = [ + ":array2d", + ":array3d", + ":array4d", + ":literal", + ":shape_util", + ":sparse_index_array", + ":status_macros", + ":types", + ":util", + ":xla_data_proto", + "//tensorflow/core:lib", + ], +) + cc_library( name = "error_spec", hdrs = ["error_spec.h"], @@ -341,6 +349,7 @@ cc_library( hdrs = ["literal_comparison.h"], deps = [ ":error_spec", + ":literal", ":literal_util", ":util", "//tensorflow/core:lib", @@ -472,7 +481,7 @@ cc_library( hdrs = ["packed_literal_reader.h"], visibility = [":internal"], deps = [ - ":literal_util", + ":literal", ":shape_util", ":status_macros", ":statusor", @@ -503,7 +512,7 @@ cc_library( hdrs = ["text_literal_reader.h"], visibility = [":internal"], deps = [ - ":literal_util", + ":literal", ":shape_util", ":status_macros", ":statusor", @@ -519,7 +528,7 @@ tf_cc_test( name = "text_literal_reader_test", srcs = ["text_literal_reader_test.cc"], deps = [ - ":literal_util", + ":literal", ":shape_util", ":test", ":text_literal_reader", @@ -536,7 +545,7 @@ cc_library( hdrs = ["text_literal_writer.h"], visibility = [":internal"], deps = [ - ":literal_util", + ":literal", ":shape_util", ":status_macros", ":types", @@ -549,6 +558,7 @@ tf_cc_test( name = "text_literal_writer_test", srcs = ["text_literal_writer_test.cc"], deps = [ + ":literal", ":literal_util", ":test", ":test_helpers", @@ -621,6 +631,7 @@ cc_library( ":array2d", ":array3d", ":array4d", + ":literal_util", ":util", ":window_util", ":xla_data_proto", @@ -641,7 +652,7 @@ tf_cc_test( ":array2d", ":array3d", ":array4d", - ":literal_util", + ":literal", ":reference_util", ":test", ":util", diff --git a/tensorflow/compiler/xla/client/BUILD b/tensorflow/compiler/xla/client/BUILD index 8f08d3b2e04670ad6590aca1db0fd9d25faed83f..25666cad40e8f73812593635ccd0ef56fcd9b955 100644 --- a/tensorflow/compiler/xla/client/BUILD +++ b/tensorflow/compiler/xla/client/BUILD @@ -65,7 +65,7 @@ cc_library( deps = [ ":global_data", "//tensorflow/compiler/xla:execution_options_util", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:service_interface", "//tensorflow/compiler/xla:status_macros", "//tensorflow/compiler/xla:statusor", diff --git a/tensorflow/compiler/xla/client/client.cc b/tensorflow/compiler/xla/client/client.cc index 3d596a6e65430b6e9692aabd65fc8aa84b7b873d..3a157c69cd73c77474f0d30fa918ac4b7d146b24 100644 --- a/tensorflow/compiler/xla/client/client.cc +++ b/tensorflow/compiler/xla/client/client.cc @@ -20,7 +20,7 @@ limitations under the License. #include "tensorflow/compiler/xla/execution_options_util.h" #include "tensorflow/compiler/xla/legacy_flags/debug_options_flags.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/ptr_util.h" #include "tensorflow/compiler/xla/status_macros.h" #include "tensorflow/compiler/xla/types.h" diff --git a/tensorflow/compiler/xla/client/client.h b/tensorflow/compiler/xla/client/client.h index 68f0d0ac78c859fde7a6a007cd250b047a7bfcda..69d4d300ca976c00aa24937776ab9f8c7f5ff945 100644 --- a/tensorflow/compiler/xla/client/client.h +++ b/tensorflow/compiler/xla/client/client.h @@ -21,7 +21,7 @@ limitations under the License. #include "tensorflow/compiler/xla/client/global_data.h" #include "tensorflow/compiler/xla/client/xla_client/xla_computation.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/service/hlo.pb.h" #include "tensorflow/compiler/xla/service_interface.h" #include "tensorflow/compiler/xla/statusor.h" diff --git a/tensorflow/compiler/xla/client/lib/BUILD b/tensorflow/compiler/xla/client/lib/BUILD index d49d959a6c8112d3701857a70cecb24701c7b6d9..6933e9a838c3741589beb85176cac8fbcfbfd8c3 100644 --- a/tensorflow/compiler/xla/client/lib/BUILD +++ b/tensorflow/compiler/xla/client/lib/BUILD @@ -13,11 +13,18 @@ filegroup( ]), ) +load("//tensorflow/compiler/xla/tests:build_defs.bzl", "xla_test") +load("//tensorflow/compiler/xla/tests:build_defs.bzl", "generate_backend_suites") + +# Generate test_suites for all backends, named "${backend}_tests". +generate_backend_suites() + cc_library( name = "arithmetic", srcs = ["arithmetic.cc"], hdrs = ["arithmetic.h"], deps = [ + ":constants", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:status_macros", "//tensorflow/compiler/xla:types", @@ -28,13 +35,96 @@ cc_library( ], ) +cc_library( + name = "constants", + srcs = ["constants.cc"], + hdrs = ["constants.h"], + deps = [ + "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:shape_util", + "//tensorflow/compiler/xla:types", + "//tensorflow/compiler/xla:util", + "//tensorflow/compiler/xla:xla_data_proto", + "//tensorflow/compiler/xla/client/xla_client:xla_builder", + ], +) + +xla_test( + name = "constants_test", + srcs = ["constants_test.cc"], + tags = ["enable_for_xla_interpreter"], + deps = [ + ":constants", + "//tensorflow/compiler/xla:test", + "//tensorflow/compiler/xla:types", + "//tensorflow/compiler/xla:xla_data_proto", + "//tensorflow/compiler/xla/client/xla_client:xla_builder", + "//tensorflow/compiler/xla/tests:client_library_test_base", + "//tensorflow/compiler/xla/tests:xla_internal_test_main", + ], +) + +cc_library( + name = "math", + srcs = ["math.cc"], + hdrs = ["math.h"], + deps = [ + ":constants", + "//tensorflow/compiler/xla:shape_util", + "//tensorflow/compiler/xla:status_macros", + "//tensorflow/compiler/xla/client/xla_client:xla_builder", + ], +) + +xla_test( + name = "math_test", + srcs = ["math_test.cc"], + tags = ["enable_for_xla_interpreter"], + deps = [ + ":math", + "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:test", + "//tensorflow/compiler/xla:types", + "//tensorflow/compiler/xla:xla_data_proto", + "//tensorflow/compiler/xla/client/xla_client:xla_builder", + "//tensorflow/compiler/xla/tests:client_library_test_base", + "//tensorflow/compiler/xla/tests:xla_internal_test_main", + ], +) + +cc_library( + name = "numeric", + srcs = ["numeric.cc"], + hdrs = ["numeric.h"], + deps = [ + "//tensorflow/compiler/xla:types", + "//tensorflow/compiler/xla:xla_data_proto", + "//tensorflow/compiler/xla/client/xla_client:xla_builder", + ], +) + +xla_test( + name = "numeric_test", + srcs = ["numeric_test.cc"], + tags = ["enable_for_xla_interpreter"], + deps = [ + ":numeric", + "//tensorflow/compiler/xla:test", + "//tensorflow/compiler/xla:types", + "//tensorflow/compiler/xla:xla_data_proto", + "//tensorflow/compiler/xla/client/xla_client:xla_builder", + "//tensorflow/compiler/xla/tests:client_library_test_base", + "//tensorflow/compiler/xla/tests:xla_internal_test_main", + ], +) + cc_library( name = "testing", srcs = ["testing.cc"], hdrs = ["testing.h"], deps = [ "//tensorflow/compiler/xla:execution_options_util", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:statusor", "//tensorflow/compiler/xla:types", diff --git a/tensorflow/compiler/xla/client/lib/arithmetic.cc b/tensorflow/compiler/xla/client/lib/arithmetic.cc index 8e875bf35218ed9935bb57b81574f33d0607696b..978fc40f3492cd7d9d7831c370b287bf45e6d3e0 100644 --- a/tensorflow/compiler/xla/client/lib/arithmetic.cc +++ b/tensorflow/compiler/xla/client/lib/arithmetic.cc @@ -17,6 +17,7 @@ limitations under the License. #include +#include "tensorflow/compiler/xla/client/lib/constants.h" #include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/compiler/xla/client/xla_client/xla_computation.h" #include "tensorflow/compiler/xla/shape_util.h" @@ -42,8 +43,8 @@ XlaComputation CreateScalarComputation(const string& name, PrimitiveType type, } const Shape scalar = ShapeUtil::MakeShape(type, {}); - auto lhs = b->Parameter(0, scalar, "lhs"); - auto rhs = b->Parameter(1, scalar, "rhs"); + auto lhs = Parameter(b.get(), 0, scalar, "lhs"); + auto rhs = Parameter(b.get(), 1, scalar, "rhs"); generator(b.get(), lhs, rhs); return b->BuildAndNoteError(); } @@ -55,7 +56,7 @@ XlaComputation CreateScalarAddComputation(PrimitiveType type, return CreateScalarComputation( "add", type, builder, [](XlaBuilder* b, const XlaOp& lhs, const XlaOp& rhs) { - return b->Add(lhs, rhs); + return Add(lhs, rhs); }); } @@ -64,17 +65,15 @@ XlaComputation CreateScalarMultiplyComputation(PrimitiveType type, return CreateScalarComputation( "mul", type, builder, [](XlaBuilder* b, const XlaOp& lhs, const XlaOp& rhs) { - return b->Mul(lhs, rhs); + return Mul(lhs, rhs); }); } XlaComputation CreateScalarGeComputation(PrimitiveType type, XlaBuilder* builder) { - return CreateScalarComputation( - "ge", type, builder, - [](XlaBuilder* b, const XlaOp& lhs, const XlaOp& rhs) { - return b->Ge(lhs, rhs); - }); + return CreateScalarComputation("ge", type, builder, + [](XlaBuilder* b, const XlaOp& lhs, + const XlaOp& rhs) { return Ge(lhs, rhs); }); } XlaComputation CreateScalarMaxComputation(PrimitiveType type, @@ -82,7 +81,7 @@ XlaComputation CreateScalarMaxComputation(PrimitiveType type, return CreateScalarComputation( "max", type, builder, [](XlaBuilder* b, const XlaOp& lhs, const XlaOp& rhs) { - return b->Max(lhs, rhs); + return Max(lhs, rhs); }); } @@ -91,7 +90,7 @@ XlaComputation CreateScalarMinComputation(PrimitiveType type, return CreateScalarComputation( "min", type, builder, [](XlaBuilder* b, const XlaOp& lhs, const XlaOp& rhs) { - return b->Min(lhs, rhs); + return Min(lhs, rhs); }); } @@ -99,156 +98,27 @@ XlaComputation CreateScalarAndComputation(XlaBuilder* builder) { return CreateScalarComputation( "and", PRED, builder, [](XlaBuilder* b, const XlaOp& lhs, const XlaOp& rhs) { - return b->And(lhs, rhs); + return And(lhs, rhs); }); } XlaComputation CreateScalarOrComputation(XlaBuilder* builder) { - return CreateScalarComputation( - "or", PRED, builder, - [](XlaBuilder* b, const XlaOp& lhs, const XlaOp& rhs) { - return b->Or(lhs, rhs); - }); + return CreateScalarComputation("or", PRED, builder, + [](XlaBuilder* b, const XlaOp& lhs, + const XlaOp& rhs) { return Or(lhs, rhs); }); } -StatusOr Any(const XlaOp& predicates, XlaBuilder* builder) { - auto f = builder->ConstantR0(false); - XlaComputation logical_or = CreateScalarOrComputation(builder); - TF_ASSIGN_OR_RETURN(const Shape& predicates_shape, - builder->GetShape(predicates)); - std::vector all_dimensions(ShapeUtil::Rank(predicates_shape)); - std::iota(all_dimensions.begin(), all_dimensions.end(), 0); - return builder->Reduce(predicates, f, logical_or, all_dimensions); -} - -namespace { -XlaOp FloatLiteral(XlaBuilder* b, PrimitiveType data_type, float value) { - return b->ConvertElementType(b->ConstantR0(value), data_type); -} - -// Polynomials for computing erf/erfc. Originally from cephes. -// Note we use float for compatibility across devices, at the cost of some -// precision for 64 bit computations. -// -// Coefficients are in descending order. -std::array kErfcPCoefficient = { - 2.46196981473530512524E-10, 5.64189564831068821977E-1, - 7.46321056442269912687E0, 4.86371970985681366614E1, - 1.96520832956077098242E2, 5.26445194995477358631E2, - 9.34528527171957607540E2, 1.02755188689515710272E3, - 5.57535335369399327526E2}; -std::array kErfcQCoefficient = { - 1.00000000000000000000E0, 1.32281951154744992508E1, - 8.67072140885989742329E1, 3.54937778887819891062E2, - 9.75708501743205489753E2, 1.82390916687909736289E3, - 2.24633760818710981792E3, 1.65666309194161350182E3, - 5.57535340817727675546E2}; -std::array kErfcRCoefficient = { - 5.64189583547755073984E-1, 1.27536670759978104416E0, - 5.01905042251180477414E0, 6.16021097993053585195E0, - 7.40974269950448939160E0, 2.97886665372100240670E0}; -std::array kErfcSCoefficient = { - 1.00000000000000000000E0, 2.26052863220117276590E0, - 9.39603524938001434673E0, 1.20489539808096656605E1, - 1.70814450747565897222E1, 9.60896809063285878198E0, - 3.36907645100081516050E0}; -std::array kErfTCoefficient = { - 9.60497373987051638749E0, 9.00260197203842689217E1, - 2.23200534594684319226E3, 7.00332514112805075473E3, - 5.55923013010394962768E4}; -std::array kErfUCoefficient = { - 1.00000000000000000000E0, 3.35617141647503099647E1, - 5.21357949780152679795E2, 4.59432382970980127987E3, - 2.26290000613890934246E4, 4.92673942608635921086E4}; -} // namespace - -// Evaluate the polynomial given coefficients and `x`. -// N.B. Coefficients should be supplied in decreasing order. -XlaOp EvaluatePolynomial(const XlaOp& x, - tensorflow::gtl::ArraySlice coefficients, - PrimitiveType data_type) { - XlaBuilder* b = x.builder(); - XlaOp poly = FloatLiteral(b, data_type, 0.0); - for (float c : coefficients) { - poly = b->Add(b->Mul(poly, x), FloatLiteral(b, data_type, c)); - } - return poly; -} - -// Compute an approximation of the error function complement (1 - erf(x)). -XlaOp Erfc(const XlaOp& x, PrimitiveType data_type) { - XlaBuilder* b = x.builder(); - XlaOp zero = FloatLiteral(b, data_type, 0.0); - XlaOp two = FloatLiteral(b, data_type, 2.0); - XlaOp eight = FloatLiteral(b, data_type, 8.0); - - XlaOp abs_x = b->Abs(x); - XlaOp z = b->Exp(b->Mul(b->Neg(x), x)); - - XlaOp pp = EvaluatePolynomial(abs_x, kErfcPCoefficient, data_type); - XlaOp pq = EvaluatePolynomial(abs_x, kErfcQCoefficient, data_type); - XlaOp pr = EvaluatePolynomial(abs_x, kErfcRCoefficient, data_type); - XlaOp ps = EvaluatePolynomial(abs_x, kErfcSCoefficient, data_type); - - XlaOp y = b->Select(b->Lt(abs_x, eight), b->Div(b->Mul(z, pp), pq), - b->Div(b->Mul(z, pr), ps)); - - return b->Select(b->Lt(x, zero), b->Sub(two, y), y); -} - -// Compute a polynomial approximation of the error function. -XlaOp Erf(const XlaOp& x, PrimitiveType data_type) { - XlaBuilder* b = x.builder(); - XlaOp z = b->Mul(x, x); - XlaOp pt = EvaluatePolynomial(z, kErfTCoefficient, data_type); - XlaOp pu = EvaluatePolynomial(z, kErfUCoefficient, data_type); - return b->Div(b->Mul(x, pt), pu); -} - -// Approximation for the inverse error function from -// Giles, M., "Approximating the erfinv function". -// The approximation has the form: -// w = -log((1 - x) * (1 + x)) -// if ( w < 5 ) { -// w = w - 2.5 -// p = sum_{i=1}^n lq[i]*w^i -// } else { -// w = sqrt(w) - 3 -// p = sum_{i=1}^n gq[i]*w^i -// } -// return p*x -StatusOr ErfInv(const XlaOp& x) { - XlaBuilder* b = x.builder(); - TF_ASSIGN_OR_RETURN(Shape shape, b->GetShape(x)); - constexpr int kDegree = 9; - constexpr std::array w_less_than_5_constants = { - 2.81022636e-08f, 3.43273939e-07f, -3.5233877e-06f, - -4.39150654e-06f, 0.00021858087f, -0.00125372503f, - -0.00417768164f, 0.246640727f, 1.50140941f}; - constexpr std::array w_greater_than_5_constants = { - -0.000200214257f, 0.000100950558f, 0.00134934322f, - -0.00367342844f, 0.00573950773f, -0.0076224613f, - 0.00943887047f, 1.00167406f, 2.83297682f}; - - auto one = b->ConstantR0(1.0); - auto w = b->Neg(b->Log(b->Mul(b->Sub(one, x), b->Add(one, x)))); - - auto lt = b->Lt(w, b->ConstantR0(5.0)); - auto coefficient = [&](int i) { - return b->Select( - lt, - b->Broadcast(b->ConstantR0(w_less_than_5_constants[i]), - AsInt64Slice(shape.dimensions())), - b->Broadcast(b->ConstantR0(w_greater_than_5_constants[i]), - AsInt64Slice(shape.dimensions()))); - }; - w = b->Select(lt, b->Sub(w, b->ConstantR0(2.5f)), - b->Sub(b->SqrtF32(w), b->ConstantR0(3.0f))); - auto p = coefficient(0); - for (int i = 1; i < kDegree; ++i) { - p = b->Add(coefficient(i), b->Mul(p, w)); - } - return b->Mul(p, x); +XlaOp Any(XlaOp predicates) { + XlaBuilder* builder = predicates.builder(); + return builder->ReportErrorOrReturn([&]() -> StatusOr { + auto f = ConstantR0(builder, false); + XlaComputation logical_or = CreateScalarOrComputation(builder); + TF_ASSIGN_OR_RETURN(const Shape& predicates_shape, + builder->GetShape(predicates)); + std::vector all_dimensions(ShapeUtil::Rank(predicates_shape)); + std::iota(all_dimensions.begin(), all_dimensions.end(), 0); + return Reduce(predicates, f, logical_or, all_dimensions); + }); } } // namespace xla diff --git a/tensorflow/compiler/xla/client/lib/arithmetic.h b/tensorflow/compiler/xla/client/lib/arithmetic.h index 33a82542740e0a82759f24137a56304725ee2d02..d0b916e8c8f742406caad0571d6e99224ed81404 100644 --- a/tensorflow/compiler/xla/client/lib/arithmetic.h +++ b/tensorflow/compiler/xla/client/lib/arithmetic.h @@ -53,22 +53,7 @@ XlaComputation CreateScalarOrComputation(XlaBuilder* builder); // Returns whether any predicate in "predicates" is set. // // Note: if predicates is zero-sized, Any() vacuously returns false. -StatusOr Any(const XlaOp& predicates, XlaBuilder* builder); - -// Evaluate the polynomial given coefficients and `x`. -// N.B. Coefficients should be supplied in decreasing order. -XlaOp EvaluatePolynomial(const XlaOp& x, - tensorflow::gtl::ArraySlice coefficients, - PrimitiveType data_type); - -// Compute an approximation of the error function complement (1 - erf(x)). -XlaOp Erfc(const XlaOp& x, PrimitiveType data_type); - -// Compute an approximation of the error function. -XlaOp Erf(const XlaOp& x, PrimitiveType data_type); - -// Compute an approximation of the inverse of the error function. -StatusOr ErfInv(const XlaOp& x); +XlaOp Any(XlaOp predicates); } // namespace xla diff --git a/tensorflow/compiler/xla/client/lib/constants.cc b/tensorflow/compiler/xla/client/lib/constants.cc new file mode 100644 index 0000000000000000000000000000000000000000..031d62e4ffef188082303a28866bbc72a154e9b1 --- /dev/null +++ b/tensorflow/compiler/xla/client/lib/constants.cc @@ -0,0 +1,103 @@ +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow/compiler/xla/client/lib/constants.h" + +#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/util.h" + +namespace xla { + +XlaOp Zero(XlaBuilder* builder, PrimitiveType type) { + return ConstantLiteral(builder, LiteralUtil::Zero(type)); +} + +XlaOp Zeros(XlaBuilder* builder, const Shape& shape) { + return Broadcast(Zero(builder, shape.element_type()), + AsInt64Slice(shape.dimensions())); +} + +XlaOp ZerosLike(XlaOp prototype) { + XlaBuilder* builder = prototype.builder(); + return builder->ReportErrorOrReturn([&]() -> StatusOr { + TF_ASSIGN_OR_RETURN(Shape shape, builder->GetShape(prototype)); + return Zeros(builder, shape); + }); +} + +XlaOp One(XlaBuilder* builder, PrimitiveType type) { + return ConstantLiteral(builder, LiteralUtil::One(type)); +} + +XlaOp Epsilon(XlaBuilder* builder, PrimitiveType type) { + switch (type) { + case F16: + return ConstantR0( + builder, + static_cast(Eigen::NumTraits::epsilon())); + case BF16: + return ConstantR0(builder, bfloat16::epsilon()); + case F32: + return ConstantR0(builder, std::numeric_limits::epsilon()); + case F64: + return ConstantR0(builder, + std::numeric_limits::epsilon()); + default: + return builder->ReportError(InvalidArgument( + "Invalid type for Epsilon (%s).", PrimitiveType_Name(type).c_str())); + } +} + +XlaOp MinValue(XlaBuilder* builder, PrimitiveType type) { + return ConstantLiteral(builder, LiteralUtil::MinValue(type)); +} + +XlaOp MinFiniteValue(XlaBuilder* builder, PrimitiveType type) { + switch (type) { + case F16: + return ConstantR0(builder, + Eigen::NumTraits::lowest()); + case BF16: + return ConstantR0(builder, bfloat16::lowest()); + case F32: + return ConstantR0(builder, -std::numeric_limits::max()); + case F64: + return ConstantR0(builder, -std::numeric_limits::max()); + default: + return MinValue(builder, type); + } +} + +XlaOp MaxValue(XlaBuilder* builder, PrimitiveType type) { + return ConstantLiteral(builder, LiteralUtil::MaxValue(type)); +} + +XlaOp MaxFiniteValue(XlaBuilder* builder, PrimitiveType type) { + switch (type) { + case F16: + return ConstantR0(builder, + Eigen::NumTraits::highest()); + case BF16: + return ConstantR0(builder, bfloat16::highest()); + case F32: + return ConstantR0(builder, std::numeric_limits::max()); + case F64: + return ConstantR0(builder, std::numeric_limits::max()); + default: + return MaxValue(builder, type); + } +} + +} // namespace xla diff --git a/tensorflow/compiler/xla/client/lib/constants.h b/tensorflow/compiler/xla/client/lib/constants.h new file mode 100644 index 0000000000000000000000000000000000000000..b47f5243f008ecb2045456e4505d1a571fbed745 --- /dev/null +++ b/tensorflow/compiler/xla/client/lib/constants.h @@ -0,0 +1,124 @@ +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#ifndef TENSORFLOW_COMPILER_XLA_CLIENT_LIB_CONSTANTS_H_ +#define TENSORFLOW_COMPILER_XLA_CLIENT_LIB_CONSTANTS_H_ + +#include + +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" +#include "tensorflow/compiler/xla/primitive_util.h" +#include "tensorflow/compiler/xla/types.h" +#include "tensorflow/compiler/xla/xla_data.pb.h" + +namespace xla { + +// Returns scalar 'value' as a scalar of 'type'. Unlike ConstantR0, 'type' is +// determined at C++ run-time, rather than C++ compile-time. +// If 'value' is floating point but 'type' is not, or if 'value' is complex but +// 'type' is not, an error will be returned. This is to catch accidental +// truncation; in such cases, use an explicit cast. +template +XlaOp ConstantR0WithType(XlaBuilder* builder, PrimitiveType type, T value) { + if (std::is_floating_point::value && + !(primitive_util::IsFloatingPointType(type) || + primitive_util::IsComplexType(type))) { + return builder->ReportError(InvalidArgument( + "Invalid cast from floating point type to %s in ConstantR0WithType.", + PrimitiveType_Name(type).c_str())); + } + if (std::is_same::value && + !primitive_util::IsComplexType(type)) { + return builder->ReportError(InvalidArgument( + "Invalid cast from complex type to %s in ConstantR0WithType.", + PrimitiveType_Name(type).c_str())); + } + switch (type) { + case F16: + return ConstantR0(builder, static_cast(value)); + case BF16: + return ConstantR0(builder, static_cast(value)); + case F32: + return ConstantR0(builder, static_cast(value)); + case F64: + return ConstantR0(builder, static_cast(value)); + case C64: + return ConstantR0(builder, static_cast(value)); + case U8: + return ConstantR0(builder, static_cast(value)); + case U32: + return ConstantR0(builder, static_cast(value)); + case U64: + return ConstantR0(builder, static_cast(value)); + case S8: + return ConstantR0(builder, static_cast(value)); + case S32: + return ConstantR0(builder, static_cast(value)); + case S64: + return ConstantR0(builder, static_cast(value)); + default: + return builder->ReportError( + InvalidArgument("Invalid type for ConstantR0WithType (%s).", + PrimitiveType_Name(type).c_str())); + } +} + +// Returns a scalar containing 'value' cast to the same run-time type as +// 'prototype'. +// If 'value' is floating point but 'prototype' is not, or if 'value' is complex +// 'prototype' is not, an error will be returned. +template +XlaOp ScalarLike(XlaOp prototype, T value) { + XlaBuilder* builder = prototype.builder(); + return builder->ReportErrorOrReturn([&]() -> StatusOr { + TF_ASSIGN_OR_RETURN(Shape shape, builder->GetShape(prototype)); + return ConstantR0WithType(builder, shape.element_type(), value); + }); +} + +// Returns a scalar with value '0' of 'type'. +XlaOp Zero(XlaBuilder* builder, PrimitiveType type); + +// Returns a zero-filled tensor with shape `shape`. +XlaOp Zeros(XlaBuilder* builder, const Shape& shape); + +// Returns a zero-filled tensor with the same shape as `prototype`. +XlaOp ZerosLike(XlaOp prototype); + +// Returns a scalar with value '1' of 'type'. +XlaOp One(XlaBuilder* builder, PrimitiveType type); + +// Returns the machine epsilon for floating-point type `type`, i.e., +// the difference between 1.0 and the next representable value. +XlaOp Epsilon(XlaBuilder* builder, PrimitiveType type); + +// Returns the minimum representable finite or infinite value for 'type'. +// Returns '-inf' for floating-point types. +XlaOp MinValue(XlaBuilder* builder, PrimitiveType type); + +// Returns the minimum representable finite value for 'type'. For a floating +// point type, this is equal to -MaxFiniteValue(). +XlaOp MinFiniteValue(XlaBuilder* builder, PrimitiveType type); + +// Returns the maximum representable finite or infinite value for 'type'. +// Returns 'inf' for floating-point types. +XlaOp MaxValue(XlaBuilder* builder, PrimitiveType type); + +// Returns the maximum representable finite value for 'type'. +XlaOp MaxFiniteValue(XlaBuilder* builder, PrimitiveType type); + +} // namespace xla + +#endif // TENSORFLOW_COMPILER_XLA_CLIENT_LIB_CONSTANTS_H_ diff --git a/tensorflow/compiler/xla/client/lib/constants_test.cc b/tensorflow/compiler/xla/client/lib/constants_test.cc new file mode 100644 index 0000000000000000000000000000000000000000..f1e3439862344c01af15ec0571155ca46a579e54 --- /dev/null +++ b/tensorflow/compiler/xla/client/lib/constants_test.cc @@ -0,0 +1,159 @@ +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow/compiler/xla/client/lib/constants.h" +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" +#include "tensorflow/compiler/xla/test.h" +#include "tensorflow/compiler/xla/tests/client_library_test_base.h" +#include "tensorflow/compiler/xla/tests/test_macros.h" +#include "tensorflow/compiler/xla/types.h" +#include "tensorflow/compiler/xla/xla_data.pb.h" + +namespace xla { +namespace { + +using ConstantsTest = ClientLibraryTestBase; + +using ::testing::HasSubstr; + +XLA_TEST_F(ConstantsTest, ConstantR0WithTypeS32) { + XlaBuilder builder(TestName()); + ConstantR0WithType(&builder, xla::S32, 4); + ComputeAndCompareR0(&builder, 4, {}); +} + +XLA_TEST_F(ConstantsTest, ConstantR0WithTypeS32DoesNotAcceptFloats) { + XlaBuilder builder(TestName()); + ConstantR0WithType(&builder, xla::S32, 4.5); + auto statusor = builder.Build(); + ASSERT_FALSE(statusor.ok()); + EXPECT_THAT(statusor.status().error_message(), HasSubstr("Invalid cast")); +} + +XLA_TEST_F(ConstantsTest, ConstantR0WithTypeF32) { + XlaBuilder builder(TestName()); + ConstantR0WithType(&builder, xla::F32, -7); + ComputeAndCompareR0(&builder, -7, {}); + ConstantR0WithType(&builder, xla::F32, 0.5); + ComputeAndCompareR0(&builder, 0.5, {}); +} + +XLA_TEST_F(ConstantsTest, ScalarLikeS32) { + XlaBuilder builder(TestName()); + ScalarLike(ConstantR0(&builder, 42), -3); + ComputeAndCompareR0(&builder, -3, {}); +} + +XLA_TEST_F(ConstantsTest, ScalarLikeF32) { + XlaBuilder builder(TestName()); + ScalarLike(ConstantR0(&builder, 42.75), -3.2); + ComputeAndCompareR0(&builder, -3.2, {}); +} + +XLA_TEST_F(ConstantsTest, ZeroS32) { + XlaBuilder builder(TestName()); + Zero(&builder, S32); + ComputeAndCompareR0(&builder, 0, {}); +} + +XLA_TEST_F(ConstantsTest, ZeroF32) { + XlaBuilder builder(TestName()); + Zero(&builder, F32); + ComputeAndCompareR0(&builder, 0.0, {}); +} + +XLA_TEST_F(ConstantsTest, ZerosS32) { + XlaBuilder builder(TestName()); + Zeros(&builder, ShapeUtil::MakeShape(S32, {2, 2})); + ComputeAndCompareR2(&builder, {{0, 0}, {0, 0}}, {}); +} + +XLA_TEST_F(ConstantsTest, ZerosLikeF32) { + XlaBuilder builder(TestName()); + ZerosLike(ConstantR1(&builder, {1., 2., 3.})); + ComputeAndCompareR1(&builder, {0., 0., 0.}, {}); +} + +XLA_TEST_F(ConstantsTest, OneS32) { + XlaBuilder builder(TestName()); + One(&builder, S32); + ComputeAndCompareR0(&builder, 1, {}); +} + +XLA_TEST_F(ConstantsTest, OneF32) { + XlaBuilder builder(TestName()); + One(&builder, F32); + ComputeAndCompareR0(&builder, 1., {}); +} + +XLA_TEST_F(ConstantsTest, EpsilonF32) { + XlaBuilder builder(TestName()); + Epsilon(&builder, F32); + ComputeAndCompareR0(&builder, std::numeric_limits::epsilon(), + {}); +} + +XLA_TEST_F(ConstantsTest, MinFiniteValueS32) { + XlaBuilder builder(TestName()); + MinFiniteValue(&builder, S32); + ComputeAndCompareR0(&builder, std::numeric_limits::min(), {}); +} + +XLA_TEST_F(ConstantsTest, MaxFiniteValueS32) { + XlaBuilder builder(TestName()); + MaxFiniteValue(&builder, S32); + ComputeAndCompareR0(&builder, std::numeric_limits::max(), {}); +} + +XLA_TEST_F(ConstantsTest, MinFiniteValueF32) { + XlaBuilder builder(TestName()); + MinFiniteValue(&builder, F32); + ComputeAndCompareR0(&builder, -std::numeric_limits::max(), {}); +} + +XLA_TEST_F(ConstantsTest, MaxFiniteValueF32) { + XlaBuilder builder(TestName()); + MaxFiniteValue(&builder, F32); + ComputeAndCompareR0(&builder, std::numeric_limits::max(), {}); +} + +XLA_TEST_F(ConstantsTest, MinValueS32) { + XlaBuilder builder(TestName()); + MinValue(&builder, S32); + ComputeAndCompareR0(&builder, std::numeric_limits::min(), {}); +} + +XLA_TEST_F(ConstantsTest, MaxValueS32) { + XlaBuilder builder(TestName()); + MaxValue(&builder, S32); + ComputeAndCompareR0(&builder, std::numeric_limits::max(), {}); +} + +XLA_TEST_F(ConstantsTest, MinValueF32) { + XlaBuilder builder(TestName()); + MinValue(&builder, F32); + ComputeAndCompareR0(&builder, -std::numeric_limits::infinity(), + {}); +} + +XLA_TEST_F(ConstantsTest, MaxValueF32) { + XlaBuilder builder(TestName()); + MaxValue(&builder, F32); + ComputeAndCompareR0(&builder, std::numeric_limits::infinity(), + {}); +} + +} // namespace +} // namespace xla diff --git a/tensorflow/compiler/xla/client/lib/math.cc b/tensorflow/compiler/xla/client/lib/math.cc new file mode 100644 index 0000000000000000000000000000000000000000..558755904007431cc0902d95a49627ea07f59127 --- /dev/null +++ b/tensorflow/compiler/xla/client/lib/math.cc @@ -0,0 +1,152 @@ +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow/compiler/xla/client/lib/math.h" + +#include "tensorflow/compiler/xla/client/lib/constants.h" +#include "tensorflow/compiler/xla/shape_util.h" +#include "tensorflow/compiler/xla/status_macros.h" + +namespace xla { + +XlaOp Sqrt(XlaOp operand) { return Pow(operand, ScalarLike(operand, 0.5)); } + +XlaOp Rsqrt(XlaOp operand) { return Pow(operand, ScalarLike(operand, -0.5)); } + +XlaOp Square(XlaOp operand) { return Pow(operand, ScalarLike(operand, 2.0)); } + +XlaOp Reciprocal(XlaOp operand) { + return Pow(operand, ScalarLike(operand, -1.0)); +} + +namespace { + +// Polynomials for computing erf/erfc. Originally from cephes. +// Note we use float for compatibility across devices, at the cost of some +// precision for 64 bit computations. +// +// Coefficients are in descending order. +std::array kErfcPCoefficient = { + 2.46196981473530512524E-10, 5.64189564831068821977E-1, + 7.46321056442269912687E0, 4.86371970985681366614E1, + 1.96520832956077098242E2, 5.26445194995477358631E2, + 9.34528527171957607540E2, 1.02755188689515710272E3, + 5.57535335369399327526E2}; +std::array kErfcQCoefficient = { + 1.00000000000000000000E0, 1.32281951154744992508E1, + 8.67072140885989742329E1, 3.54937778887819891062E2, + 9.75708501743205489753E2, 1.82390916687909736289E3, + 2.24633760818710981792E3, 1.65666309194161350182E3, + 5.57535340817727675546E2}; +std::array kErfcRCoefficient = { + 5.64189583547755073984E-1, 1.27536670759978104416E0, + 5.01905042251180477414E0, 6.16021097993053585195E0, + 7.40974269950448939160E0, 2.97886665372100240670E0}; +std::array kErfcSCoefficient = { + 1.00000000000000000000E0, 2.26052863220117276590E0, + 9.39603524938001434673E0, 1.20489539808096656605E1, + 1.70814450747565897222E1, 9.60896809063285878198E0, + 3.36907645100081516050E0}; +std::array kErfTCoefficient = { + 9.60497373987051638749E0, 9.00260197203842689217E1, + 2.23200534594684319226E3, 7.00332514112805075473E3, + 5.55923013010394962768E4}; +std::array kErfUCoefficient = { + 1.00000000000000000000E0, 3.35617141647503099647E1, + 5.21357949780152679795E2, 4.59432382970980127987E3, + 2.26290000613890934246E4, 4.92673942608635921086E4}; +} // namespace + +// Evaluate the polynomial given coefficients and `x`. +// N.B. Coefficients should be supplied in decreasing order. +XlaOp EvaluatePolynomial(XlaOp x, + tensorflow::gtl::ArraySlice coefficients) { + XlaOp poly = ScalarLike(x, 0.0); + for (float c : coefficients) { + poly = poly * x + ScalarLike(x, c); + } + return poly; +} + +// Compute an approximation of the error function complement (1 - erf(x)). +XlaOp Erfc(XlaOp x) { + XlaOp abs_x = Abs(x); + XlaOp z = Exp(-x * x); + + XlaOp pp = EvaluatePolynomial(abs_x, kErfcPCoefficient); + XlaOp pq = EvaluatePolynomial(abs_x, kErfcQCoefficient); + XlaOp pr = EvaluatePolynomial(abs_x, kErfcRCoefficient); + XlaOp ps = EvaluatePolynomial(abs_x, kErfcSCoefficient); + + XlaOp y = Select(Lt(abs_x, ScalarLike(x, 8.0)), z * pp / pq, z * pr / ps); + + return Select(Lt(x, ScalarLike(x, 0.0)), ScalarLike(x, 2.0) - y, y); +} + +// Compute a polynomial approximation of the error function. +XlaOp Erf(XlaOp x) { + XlaOp z = x * x; + XlaOp pt = EvaluatePolynomial(z, kErfTCoefficient); + XlaOp pu = EvaluatePolynomial(z, kErfUCoefficient); + return x * pt / pu; +} + +// Approximation for the inverse error function from +// Giles, M., "Approximating the erfinv function". +// The approximation has the form: +// w = -log((1 - x) * (1 + x)) +// if ( w < 5 ) { +// w = w - 2.5 +// p = sum_{i=1}^n lq[i]*w^i +// } else { +// w = sqrt(w) - 3 +// p = sum_{i=1}^n gq[i]*w^i +// } +// return p*x +XlaOp ErfInv(XlaOp x) { + XlaBuilder* b = x.builder(); + return b->ReportErrorOrReturn([&]() -> StatusOr { + TF_ASSIGN_OR_RETURN(Shape shape, b->GetShape(x)); + constexpr int kDegree = 9; + constexpr std::array w_less_than_5_constants = { + 2.81022636e-08f, 3.43273939e-07f, -3.5233877e-06f, + -4.39150654e-06f, 0.00021858087f, -0.00125372503f, + -0.00417768164f, 0.246640727f, 1.50140941f}; + constexpr std::array w_greater_than_5_constants = { + -0.000200214257f, 0.000100950558f, 0.00134934322f, + -0.00367342844f, 0.00573950773f, -0.0076224613f, + 0.00943887047f, 1.00167406f, 2.83297682f}; + + auto one = ScalarLike(x, 1.0); + auto w = -Log((one - x) * (one + x)); + + auto lt = Lt(w, ScalarLike(x, 5.0)); + auto coefficient = [&](int i) { + return Select(lt, + Broadcast(ScalarLike(x, w_less_than_5_constants[i]), + AsInt64Slice(shape.dimensions())), + Broadcast(ScalarLike(x, w_greater_than_5_constants[i]), + AsInt64Slice(shape.dimensions()))); + }; + w = Select(lt, w - ScalarLike(x, 2.5), Sqrt(w) - ScalarLike(x, 3.0)); + auto p = coefficient(0); + for (int i = 1; i < kDegree; ++i) { + p = coefficient(i) + p * w; + } + return p * x; + }); +} + +} // namespace xla diff --git a/tensorflow/compiler/xla/client/lib/math.h b/tensorflow/compiler/xla/client/lib/math.h new file mode 100644 index 0000000000000000000000000000000000000000..e7c8b50273067a979158f79aa80abc6058901040 --- /dev/null +++ b/tensorflow/compiler/xla/client/lib/math.h @@ -0,0 +1,51 @@ +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#ifndef TENSORFLOW_COMPILER_XLA_CLIENT_LIB_MATH_H_ +#define TENSORFLOW_COMPILER_XLA_CLIENT_LIB_MATH_H_ + +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" + +namespace xla { + +// Computes the square root of 'operand'. +XlaOp Sqrt(XlaOp operand); + +// Computes the reciprocal of the square root of 'operand'. +XlaOp Rsqrt(XlaOp operand); + +// Computes the square of 'operand'. +XlaOp Square(XlaOp operand); + +// Computes the reciprocal of 'operand'. +XlaOp Reciprocal(XlaOp operand); + +// Evaluates a polynomial given coefficients and `x`. +// N.B. Coefficients should be supplied in decreasing order. +XlaOp EvaluatePolynomial(XlaOp x, + tensorflow::gtl::ArraySlice coefficients); + +// Computes an approximation of the error function complement (1 - erf(x)). +XlaOp Erfc(XlaOp x); + +// Computes an approximation of the error function. +XlaOp Erf(XlaOp x); + +// Computes an approximation of the inverse of the error function. +XlaOp ErfInv(XlaOp x); + +} // namespace xla + +#endif // TENSORFLOW_COMPILER_XLA_CLIENT_LIB_MATH_H_ diff --git a/tensorflow/compiler/xla/client/lib/math_test.cc b/tensorflow/compiler/xla/client/lib/math_test.cc new file mode 100644 index 0000000000000000000000000000000000000000..068cd2e58615ba7bcad920e387b635f341a79b80 --- /dev/null +++ b/tensorflow/compiler/xla/client/lib/math_test.cc @@ -0,0 +1,86 @@ +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow/compiler/xla/client/lib/math.h" +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" +#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/test.h" +#include "tensorflow/compiler/xla/tests/client_library_test_base.h" +#include "tensorflow/compiler/xla/tests/test_macros.h" +#include "tensorflow/compiler/xla/types.h" +#include "tensorflow/compiler/xla/xla_data.pb.h" + +namespace xla { +namespace { + +class MathTest : public ClientLibraryTestBase { + public: + ErrorSpec error_spec_{0.0001}; +}; + +XLA_TEST_F(MathTest, SqrtF32) { + XlaBuilder builder(TestName()); + Literal zero_literal = LiteralUtil::Zero(PrimitiveType::F32); + + std::unique_ptr zero_data = + client_->TransferToServer(zero_literal).ConsumeValueOrDie(); + + XlaOp zero = Parameter(&builder, 0, zero_literal.shape(), "zero"); + Sqrt(zero); + + ComputeAndCompareR0(&builder, 0.0f, {zero_data.get()}, error_spec_); +} + +XLA_TEST_F(MathTest, SquareTenValues) { + XlaBuilder builder(TestName()); + auto x = ConstantR1( + &builder, {2.1, -2.6, 2.6, -4.0, 2.1, 2.3, -5.0, -0.9, -2.4, 1.6}); + Square(x); + + std::vector expected = {4.41, 6.76, 6.76, 16., 4.41, + 5.29, 25., 0.81, 5.76, 2.56}; + ComputeAndCompareR1(&builder, expected, {}, error_spec_); +} + +XLA_TEST_F(MathTest, ReciprocalTenValues) { + XlaBuilder builder(TestName()); + auto x = ConstantR1( + &builder, {2.1, -2.6, 2.6, -4.0, 2.1, 2.3, -5.0, -0.9, -2.4, 1.6}); + Reciprocal(x); + + std::vector expected = { + 0.47619048, -0.38461538, 0.38461538, -0.25, 0.47619048, + 0.43478261, -0.2, -1.11111111, -0.41666667, 0.625}; + ComputeAndCompareR1(&builder, expected, {}, error_spec_); +} + +XLA_TEST_F(MathTest, SqrtZeroes) { + XlaBuilder builder(TestName()); + auto x = ConstantR1(&builder, {0.0, -0.0}); + Sqrt(x); + + ComputeAndCompareR1(&builder, {0, 0}, {}, error_spec_); +} + +XLA_TEST_F(MathTest, SqrtSixValues) { + XlaBuilder builder(TestName()); + auto x = ConstantR1(&builder, {16.0, 1.0, 1024.0, 0.16, 0.2, 12345}); + Sqrt(x); + + std::vector expected = {4, 1, 32, 0.4, 0.4472, 111.1080}; + ComputeAndCompareR1(&builder, expected, {}, error_spec_); +} +} // namespace +} // namespace xla diff --git a/tensorflow/compiler/xla/client/lib/numeric.cc b/tensorflow/compiler/xla/client/lib/numeric.cc new file mode 100644 index 0000000000000000000000000000000000000000..fd4e8fc390e840caa2939e5a392f8fb95b602d18 --- /dev/null +++ b/tensorflow/compiler/xla/client/lib/numeric.cc @@ -0,0 +1,79 @@ +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow/compiler/xla/client/lib/numeric.h" + +#include +#include + +namespace xla { + +namespace { + +template +XlaOp MakeIota(XlaBuilder* builder, int64 size) { + std::vector values(size); + for (int64 i = 0; i < size; ++i) { + values[i] = static_cast(i); + } + return xla::ConstantR1(builder, values); +} + +} // namespace + +XlaOp Iota(XlaBuilder* builder, PrimitiveType type, int64 size) { + switch (type) { + case S8: + return MakeIota(builder, size); + case S16: + return MakeIota(builder, size); + case S32: + return MakeIota(builder, size); + case S64: + return MakeIota(builder, size); + case U8: + return MakeIota(builder, size); + case U16: + return MakeIota(builder, size); + case U32: + return MakeIota(builder, size); + case U64: + return MakeIota(builder, size); + case BF16: + return MakeIota(builder, size); + case F16: + return MakeIota(builder, size); + case F32: + return MakeIota(builder, size); + case F64: + return MakeIota(builder, size); + case C64: + return MakeIota(builder, size); + default: + return builder->ReportError( + InvalidArgument("Unimplemented type for Iota: %s.", + PrimitiveType_Name(type).c_str())); + } +} + +XlaOp IdentityMatrix(XlaBuilder* builder, PrimitiveType type, int64 m, + int64 n) { + auto a = Iota(builder, type, m); + auto b = Iota(builder, type, n); + auto indicator = Eq(a, Broadcast(b, {m}), /*broadcast_dimensions=*/{0}); + return ConvertElementType(indicator, type); +} + +} // namespace xla diff --git a/tensorflow/compiler/xla/client/lib/numeric.h b/tensorflow/compiler/xla/client/lib/numeric.h new file mode 100644 index 0000000000000000000000000000000000000000..79707007b248638fccb62d8e4ec56f6b1662de5e --- /dev/null +++ b/tensorflow/compiler/xla/client/lib/numeric.h @@ -0,0 +1,34 @@ +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#ifndef TENSORFLOW_COMPILER_XLA_CLIENT_LIB_NUMERIC_H_ +#define TENSORFLOW_COMPILER_XLA_CLIENT_LIB_NUMERIC_H_ + +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" +#include "tensorflow/compiler/xla/types.h" +#include "tensorflow/compiler/xla/xla_data.pb.h" + +namespace xla { + +// Returns a rank 1 tensor of `type` containing values [0, 1, 2, ...]. +XlaOp Iota(XlaBuilder* builder, PrimitiveType type, int64 size); + +// Returns an m x n matrix with 1s on the diagonal elements, zeros everywhere +// else. +XlaOp IdentityMatrix(XlaBuilder* builder, PrimitiveType type, int64 m, int64 n); + +} // namespace xla + +#endif // TENSORFLOW_COMPILER_XLA_CLIENT_LIB_NUMERIC_H_ diff --git a/tensorflow/compiler/xla/client/lib/numeric_test.cc b/tensorflow/compiler/xla/client/lib/numeric_test.cc new file mode 100644 index 0000000000000000000000000000000000000000..bc8a73e9d793ef8f65c321759e03b0de75edd500 --- /dev/null +++ b/tensorflow/compiler/xla/client/lib/numeric_test.cc @@ -0,0 +1,37 @@ +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow/compiler/xla/client/lib/numeric.h" +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" +#include "tensorflow/compiler/xla/test.h" +#include "tensorflow/compiler/xla/tests/client_library_test_base.h" +#include "tensorflow/compiler/xla/tests/test_macros.h" +#include "tensorflow/compiler/xla/types.h" +#include "tensorflow/compiler/xla/xla_data.pb.h" + +namespace xla { +namespace { + +using NumericTest = ClientLibraryTestBase; + +XLA_TEST_F(NumericTest, Iota) { + XlaBuilder builder(TestName()); + Iota(&builder, S32, 10); + + ComputeAndCompareR1(&builder, {0, 1, 2, 3, 4, 5, 6, 7, 8, 9}, {}); +} + +} // namespace +} // namespace xla diff --git a/tensorflow/compiler/xla/client/lib/testing.cc b/tensorflow/compiler/xla/client/lib/testing.cc index 3380af9f303b1dc2cec09aa37410ec40cdeaa526..534c5098683e8984fe013225e7de2fc48b1bbc1f 100644 --- a/tensorflow/compiler/xla/client/lib/testing.cc +++ b/tensorflow/compiler/xla/client/lib/testing.cc @@ -17,7 +17,7 @@ limitations under the License. #include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/compiler/xla/execution_options_util.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/shape_util.h" #include "tensorflow/compiler/xla/statusor.h" #include "tensorflow/compiler/xla/tests/test_utils.h" @@ -48,15 +48,15 @@ int64 DataSizeOfShape(const Shape& shape) { // Creates a XlaOp for an op what generates fake data with the given shape. XlaOp BuildFakeDataOpOnDevice(const Shape& shape, XlaBuilder* builder) { if (ShapeUtil::IsArray(shape)) { - return builder->Broadcast( - builder->ConstantLiteral(Literal::One(shape.element_type())), + return Broadcast( + ConstantLiteral(builder, LiteralUtil::One(shape.element_type())), AsInt64Slice(shape.dimensions())); } std::vector parts; for (const Shape& s : shape.tuple_shapes()) { parts.push_back(BuildFakeDataOpOnDevice(s, builder)); } - return builder->Tuple(parts); + return Tuple(builder, parts); } std::unique_ptr MakeFakeDataViaDeviceOrDie(const Shape& shape, diff --git a/tensorflow/compiler/xla/client/xla_client/BUILD b/tensorflow/compiler/xla/client/xla_client/BUILD index 507a2dc5f088e159156f0ef3d663ba2819f6a2d4..763653c685cb0d821c11bf3e25f526db0dcb4945 100644 --- a/tensorflow/compiler/xla/client/xla_client/BUILD +++ b/tensorflow/compiler/xla/client/xla_client/BUILD @@ -1,7 +1,5 @@ # Description: # The new XLA client libraries. -# -# This is NOT YET ready to use. licenses(["notice"]) # Apache 2.0 @@ -41,9 +39,11 @@ cc_library( name = "xla_builder", srcs = ["xla_builder.cc"], hdrs = ["xla_builder.h"], + visibility = ["//visibility:public"], deps = [ ":xla_computation", "//tensorflow/compiler/xla:execution_options_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:literal_util", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:status_macros", @@ -52,6 +52,7 @@ cc_library( "//tensorflow/compiler/xla:util", "//tensorflow/compiler/xla:xla_data_proto", "//tensorflow/compiler/xla/client:padding", + "//tensorflow/compiler/xla/client:sharding_builder", "//tensorflow/compiler/xla/service:hlo", "//tensorflow/compiler/xla/service:hlo_proto", "//tensorflow/compiler/xla/service:shape_inference", @@ -64,7 +65,7 @@ tf_cc_test( srcs = ["xla_builder_test.cc"], deps = [ ":xla_builder", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:status_macros", "//tensorflow/compiler/xla:test", diff --git a/tensorflow/compiler/xla/client/xla_client/xla_builder.cc b/tensorflow/compiler/xla/client/xla_client/xla_builder.cc index 256667cbe00764047fd2ca6afa135a43bdaad0f0..aac7df43831fddea05c3d2581ec1e465469ad297 100644 --- a/tensorflow/compiler/xla/client/xla_client/xla_builder.cc +++ b/tensorflow/compiler/xla/client/xla_client/xla_builder.cc @@ -21,6 +21,7 @@ limitations under the License. #include #include +#include "tensorflow/compiler/xla/client/sharding_builder.h" #include "tensorflow/compiler/xla/execution_options_util.h" #include "tensorflow/compiler/xla/service/hlo_opcode.h" #include "tensorflow/compiler/xla/service/shape_inference.h" @@ -47,6 +48,7 @@ int64 GetUniqueId() { // computation. bool CanBeRoot(HloOpcode opcode) { switch (opcode) { + case HloOpcode::kAfterAll: case HloOpcode::kSend: case HloOpcode::kSendDone: case HloOpcode::kOutfeed: @@ -59,6 +61,36 @@ bool CanBeRoot(HloOpcode opcode) { } // namespace +XlaOp operator-(const XlaOp& x) { return Neg(x); } +XlaOp operator+(const XlaOp& x, const XlaOp& y) { return Add(x, y); } +XlaOp operator-(const XlaOp& x, const XlaOp& y) { return Sub(x, y); } +XlaOp operator*(const XlaOp& x, const XlaOp& y) { return Mul(x, y); } +XlaOp operator/(const XlaOp& x, const XlaOp& y) { return Div(x, y); } +XlaOp operator%(const XlaOp& x, const XlaOp& y) { return Rem(x, y); } + +XlaOp operator~(const XlaOp& x) { return Not(x); } +XlaOp operator&(const XlaOp& x, const XlaOp& y) { return And(x, y); } +XlaOp operator|(const XlaOp& x, const XlaOp& y) { return Or(x, y); } +XlaOp operator^(const XlaOp& x, const XlaOp& y) { return Xor(x, y); } +XlaOp operator<<(const XlaOp& x, const XlaOp& y) { return ShiftLeft(x, y); } + +XlaOp operator>>(const XlaOp& x, const XlaOp& y) { + XlaBuilder* builder = x.builder(); + return builder->ReportErrorOrReturn([&]() -> StatusOr { + TF_ASSIGN_OR_RETURN(xla::Shape shape, builder->GetShape(x)); + if (!ShapeUtil::ElementIsIntegral(shape)) { + return InvalidArgument( + "Argument to >> operator does not have an integral type (%s).", + ShapeUtil::HumanString(shape).c_str()); + } + if (ShapeUtil::ElementIsSigned(shape)) { + return ShiftRightArithmetic(x, y); + } else { + return ShiftRightLogical(x, y); + } + }); +} + StatusOr XlaBuilder::GetShape(const XlaOp& op) const { TF_RETURN_IF_ERROR(first_error_); @@ -81,7 +113,7 @@ XlaBuilder::XlaBuilder(const string& computation_name) XlaBuilder::~XlaBuilder() {} -void XlaBuilder::NoteError(const Status& error) { +XlaOp XlaBuilder::ReportError(const Status& error) { CHECK(!error.ok()); if (die_immediately_on_error_) { LOG(FATAL) << "error building computation: " << error; @@ -91,19 +123,22 @@ void XlaBuilder::NoteError(const Status& error) { first_error_ = error; first_error_backtrace_.CreateCurrent(/*skip_count=*/1); } + return XlaOp(this); } -XlaOp XlaBuilder::NoteErrorOrReturn( - const std::function()>& op_creator) { +XlaOp XlaBuilder::ReportErrorOrReturn(const StatusOr& op) { if (!first_error_.ok()) { return XlaOp(this); } - auto op = op_creator(); if (!op.ok()) { - NoteError(op.status()); - return XlaOp(this); + return ReportError(op.status()); } - return op.ConsumeValueOrDie(); + return op.ValueOrDie(); +} + +XlaOp XlaBuilder::ReportErrorOrReturn( + const std::function()>& op_creator) { + return ReportErrorOrReturn(op_creator()); } StatusOr XlaBuilder::GetProgramShape(int64* root_id) const { @@ -207,7 +242,7 @@ XlaComputation XlaBuilder::BuildAndNoteError() { DCHECK(parent_builder_ != nullptr); auto build_status = Build(); if (!build_status.ok()) { - parent_builder_->NoteError( + parent_builder_->ReportError( AddStatus(build_status.status(), tensorflow::strings::StrCat("error from: ", name_))); return {}; @@ -315,7 +350,7 @@ StatusOr XlaBuilder::AddBroadcastSequence(const Shape& output_shape, } XlaOp XlaBuilder::UnaryOp(HloOpcode unop, const XlaOp& operand) { - return NoteErrorOrReturn([&]() -> StatusOr { + return ReportErrorOrReturn([&]() -> StatusOr { HloInstructionProto instr; TF_ASSIGN_OR_RETURN(const Shape& operand_shape, GetShape(operand)); TF_ASSIGN_OR_RETURN(*instr.mutable_shape(), @@ -327,7 +362,7 @@ XlaOp XlaBuilder::UnaryOp(HloOpcode unop, const XlaOp& operand) { XlaOp XlaBuilder::BinaryOp( HloOpcode binop, const XlaOp& lhs, const XlaOp& rhs, tensorflow::gtl::ArraySlice broadcast_dimensions) { - return NoteErrorOrReturn([&]() -> StatusOr { + return ReportErrorOrReturn([&]() -> StatusOr { HloInstructionProto instr; TF_ASSIGN_OR_RETURN(const Shape& lhs_shape, GetShape(lhs)); TF_ASSIGN_OR_RETURN(const Shape& rhs_shape, GetShape(rhs)); @@ -383,7 +418,7 @@ XlaOp XlaBuilder::BinaryOp( XlaOp XlaBuilder::TernaryOp(HloOpcode triop, const XlaOp& lhs, const XlaOp& rhs, const XlaOp& ehs) { - return NoteErrorOrReturn([&]() -> StatusOr { + return ReportErrorOrReturn([&]() -> StatusOr { HloInstructionProto instr; TF_ASSIGN_OR_RETURN(const Shape& lhs_shape, GetShape(lhs)); TF_ASSIGN_OR_RETURN(const Shape& rhs_shape, GetShape(rhs)); @@ -430,7 +465,7 @@ XlaOp XlaBuilder::Mul(const XlaOp& lhs, const XlaOp& rhs, } XlaOp XlaBuilder::ConstantLiteral(const LiteralSlice& literal) { - return NoteErrorOrReturn([&]() -> StatusOr { + return ReportErrorOrReturn([&]() -> StatusOr { HloInstructionProto instr; *instr.mutable_shape() = literal.shape(); *instr.mutable_literal() = literal.ToProto(); @@ -440,7 +475,7 @@ XlaOp XlaBuilder::ConstantLiteral(const LiteralSlice& literal) { XlaOp XlaBuilder::Call(const XlaComputation& computation, tensorflow::gtl::ArraySlice operands) { - return NoteErrorOrReturn([&]() -> StatusOr { + return ReportErrorOrReturn([&]() -> StatusOr { HloInstructionProto instr; std::vector operand_shape_ptrs; TF_ASSIGN_OR_RETURN(const auto& operand_shapes, GetOperandShapes(operands)); @@ -461,7 +496,7 @@ XlaOp XlaBuilder::Call(const XlaComputation& computation, XlaOp XlaBuilder::Parameter(int64 parameter_number, const Shape& shape, const string& name) { - return NoteErrorOrReturn([&]() -> StatusOr { + return ReportErrorOrReturn([&]() -> StatusOr { HloInstructionProto instr; if (!parameter_numbers_.insert(parameter_number).second) { return InvalidArgument("parameter %lld already registered", @@ -476,7 +511,7 @@ XlaOp XlaBuilder::Parameter(int64 parameter_number, const Shape& shape, XlaOp XlaBuilder::Broadcast( const XlaOp& operand, tensorflow::gtl::ArraySlice broadcast_sizes) { - return NoteErrorOrReturn([&]() -> StatusOr { + return ReportErrorOrReturn([&]() -> StatusOr { TF_ASSIGN_OR_RETURN(const Shape& operand_shape, GetShape(operand)); TF_ASSIGN_OR_RETURN( const Shape& shape, @@ -498,6 +533,14 @@ XlaOp XlaBuilder::Broadcast( }); } +XlaOp XlaBuilder::BroadcastInDim( + const XlaOp& operand, const Shape& shape, + const tensorflow::gtl::ArraySlice broadcast_dimensions) { + return ReportErrorOrReturn([&]() -> StatusOr { + return InDimBroadcast(shape, operand, broadcast_dimensions); + }); +} + StatusOr XlaBuilder::Reshape(const Shape& shape, const XlaOp& operand) { TF_RETURN_IF_ERROR(first_error_); @@ -510,7 +553,7 @@ XlaOp XlaBuilder::Slice(const XlaOp& operand, tensorflow::gtl::ArraySlice start_indices, tensorflow::gtl::ArraySlice limit_indices, tensorflow::gtl::ArraySlice strides) { - return NoteErrorOrReturn([&]() -> StatusOr { + return ReportErrorOrReturn([&]() -> StatusOr { HloInstructionProto instr; TF_ASSIGN_OR_RETURN(const Shape& operand_shape, GetShape(operand)); TF_ASSIGN_OR_RETURN( @@ -530,7 +573,7 @@ XlaOp XlaBuilder::Slice(const XlaOp& operand, XlaOp XlaBuilder::SliceInDim(const XlaOp& operand, int64 start_index, int64 limit_index, int64 stride, int64 dimno) { - return NoteErrorOrReturn([&]() -> StatusOr { + return ReportErrorOrReturn([&]() -> StatusOr { TF_ASSIGN_OR_RETURN(const Shape& shape, GetShape(operand)); std::vector starts(ShapeUtil::Rank(shape), 0); std::vector limits(shape.dimensions().begin(), @@ -545,7 +588,7 @@ XlaOp XlaBuilder::SliceInDim(const XlaOp& operand, int64 start_index, XlaOp XlaBuilder::DynamicSlice(const XlaOp& operand, const XlaOp& start_indices, tensorflow::gtl::ArraySlice slice_sizes) { - return NoteErrorOrReturn([&]() -> StatusOr { + return ReportErrorOrReturn([&]() -> StatusOr { HloInstructionProto instr; TF_ASSIGN_OR_RETURN(const Shape& operand_shape, GetShape(operand)); @@ -566,7 +609,7 @@ XlaOp XlaBuilder::DynamicSlice(const XlaOp& operand, const XlaOp& start_indices, XlaOp XlaBuilder::DynamicUpdateSlice(const XlaOp& operand, const XlaOp& update, const XlaOp& start_indices) { - return NoteErrorOrReturn([&]() -> StatusOr { + return ReportErrorOrReturn([&]() -> StatusOr { HloInstructionProto instr; TF_ASSIGN_OR_RETURN(const Shape& operand_shape, GetShape(operand)); @@ -584,7 +627,7 @@ XlaOp XlaBuilder::DynamicUpdateSlice(const XlaOp& operand, const XlaOp& update, XlaOp XlaBuilder::ConcatInDim(tensorflow::gtl::ArraySlice operands, int64 dimension) { - return NoteErrorOrReturn([&]() -> StatusOr { + return ReportErrorOrReturn([&]() -> StatusOr { HloInstructionProto instr; std::vector operand_shape_ptrs; @@ -603,7 +646,7 @@ XlaOp XlaBuilder::ConcatInDim(tensorflow::gtl::ArraySlice operands, XlaOp XlaBuilder::Pad(const XlaOp& operand, const XlaOp& padding_value, const PaddingConfig& padding_config) { - return NoteErrorOrReturn([&]() -> StatusOr { + return ReportErrorOrReturn([&]() -> StatusOr { HloInstructionProto instr; TF_ASSIGN_OR_RETURN(const Shape& operand_shape, GetShape(operand)); @@ -624,7 +667,7 @@ XlaOp XlaBuilder::Pad(const XlaOp& operand, const XlaOp& padding_value, XlaOp XlaBuilder::Reshape(const XlaOp& operand, tensorflow::gtl::ArraySlice dimensions, tensorflow::gtl::ArraySlice new_sizes) { - return NoteErrorOrReturn([&]() -> StatusOr { + return ReportErrorOrReturn([&]() -> StatusOr { TF_ASSIGN_OR_RETURN(const Shape& operand_shape, GetShape(operand)); TF_ASSIGN_OR_RETURN(const Shape& shape, ShapeInference::InferReshapeShape( @@ -638,7 +681,7 @@ XlaOp XlaBuilder::Reshape(const XlaOp& operand, XlaOp XlaBuilder::Reshape(const XlaOp& operand, tensorflow::gtl::ArraySlice new_sizes) { - return NoteErrorOrReturn([&]() -> StatusOr { + return ReportErrorOrReturn([&]() -> StatusOr { TF_ASSIGN_OR_RETURN(auto shape, GetShape(operand)); std::vector dimensions(shape.dimensions_size()); std::iota(dimensions.begin(), dimensions.end(), 0); @@ -648,7 +691,7 @@ XlaOp XlaBuilder::Reshape(const XlaOp& operand, XlaOp XlaBuilder::Collapse(const XlaOp& operand, tensorflow::gtl::ArraySlice dimensions) { - return NoteErrorOrReturn([&]() -> StatusOr { + return ReportErrorOrReturn([&]() -> StatusOr { if (dimensions.size() <= 1) { // Not collapsing anything, trivially we can return the operand versus // enqueueing a trivial reshape. @@ -690,21 +733,29 @@ XlaOp XlaBuilder::Collapse(const XlaOp& operand, } void XlaBuilder::Trace(const string& tag, const XlaOp& operand) { - NoteErrorOrReturn([&]() -> StatusOr { + ReportErrorOrReturn([&]() -> StatusOr { HloInstructionProto instr; *instr.mutable_shape() = ShapeUtil::MakeNil(); - *instr.mutable_literal() = Literal::CreateR1U8(tag)->ToProto(); + *instr.mutable_literal() = LiteralUtil::CreateR1U8(tag)->ToProto(); return AddInstruction(std::move(instr), HloOpcode::kTrace, {operand}); }); } XlaOp XlaBuilder::Select(const XlaOp& pred, const XlaOp& on_true, const XlaOp& on_false) { - return TernaryOp(HloOpcode::kSelect, pred, on_true, on_false); + return ReportErrorOrReturn([&]() -> StatusOr { + TF_ASSIGN_OR_RETURN(const Shape& true_shape, GetShape(on_true)); + TF_ASSIGN_OR_RETURN(const Shape& false_shape, GetShape(on_false)); + TF_RET_CHECK(ShapeUtil::IsTuple(true_shape) == + ShapeUtil::IsTuple(false_shape)); + HloOpcode opcode = ShapeUtil::IsTuple(true_shape) ? HloOpcode::kTupleSelect + : HloOpcode::kSelect; + return TernaryOp(opcode, pred, on_true, on_false); + }); } XlaOp XlaBuilder::Tuple(tensorflow::gtl::ArraySlice elements) { - return NoteErrorOrReturn([&]() -> StatusOr { + return ReportErrorOrReturn([&]() -> StatusOr { HloInstructionProto instr; std::vector operand_shape_ptrs; TF_ASSIGN_OR_RETURN(const auto& operand_shapes, GetOperandShapes(elements)); @@ -718,7 +769,7 @@ XlaOp XlaBuilder::Tuple(tensorflow::gtl::ArraySlice elements) { } XlaOp XlaBuilder::GetTupleElement(const XlaOp& tuple_data, int64 index) { - return NoteErrorOrReturn([&]() -> StatusOr { + return ReportErrorOrReturn([&]() -> StatusOr { HloInstructionProto instr; TF_ASSIGN_OR_RETURN(const Shape& tuple_shape, GetShape(tuple_data)); if (!ShapeUtil::IsTuple(tuple_shape)) { @@ -767,7 +818,7 @@ XlaOp XlaBuilder::Lt(const XlaOp& lhs, const XlaOp& rhs, } XlaOp XlaBuilder::Dot(const XlaOp& lhs, const XlaOp& rhs) { - return NoteErrorOrReturn([&]() -> StatusOr { + return ReportErrorOrReturn([&]() -> StatusOr { TF_ASSIGN_OR_RETURN(const Shape& lhs_shape, GetShape(lhs)); DotDimensionNumbers dimension_numbers; @@ -780,7 +831,7 @@ XlaOp XlaBuilder::Dot(const XlaOp& lhs, const XlaOp& rhs) { XlaOp XlaBuilder::DotGeneral(const XlaOp& lhs, const XlaOp& rhs, const DotDimensionNumbers& dimension_numbers) { - return NoteErrorOrReturn([&]() -> StatusOr { + return ReportErrorOrReturn([&]() -> StatusOr { HloInstructionProto instr; TF_ASSIGN_OR_RETURN(const Shape& lhs_shape, GetShape(lhs)); TF_ASSIGN_OR_RETURN(const Shape& rhs_shape, GetShape(rhs)); @@ -859,7 +910,7 @@ XlaOp XlaBuilder::ConvWithGeneralDimensions( const XlaOp& lhs, const XlaOp& rhs, tensorflow::gtl::ArraySlice window_strides, Padding padding, const ConvolutionDimensionNumbers& dimension_numbers) { - return NoteErrorOrReturn([&]() -> StatusOr { + return ReportErrorOrReturn([&]() -> StatusOr { TF_ASSIGN_OR_RETURN(const Shape& lhs_shape, GetShape(lhs)); TF_ASSIGN_OR_RETURN(const Shape& rhs_shape, GetShape(rhs)); @@ -905,7 +956,7 @@ XlaOp XlaBuilder::ConvGeneralDilated( tensorflow::gtl::ArraySlice lhs_dilation, tensorflow::gtl::ArraySlice rhs_dilation, const ConvolutionDimensionNumbers& dimension_numbers) { - return NoteErrorOrReturn([&]() -> StatusOr { + return ReportErrorOrReturn([&]() -> StatusOr { HloInstructionProto instr; TF_ASSIGN_OR_RETURN(const Shape& lhs_shape, GetShape(lhs)); TF_ASSIGN_OR_RETURN(const Shape& rhs_shape, GetShape(rhs)); @@ -992,7 +1043,7 @@ StatusOr XlaBuilder::MakeWindow( XlaOp XlaBuilder::Fft(const XlaOp& operand, const FftType fft_type, const tensorflow::gtl::ArraySlice fft_length) { - return NoteErrorOrReturn([&]() -> StatusOr { + return ReportErrorOrReturn([&]() -> StatusOr { HloInstructionProto instr; TF_ASSIGN_OR_RETURN(const Shape& operand_shape, GetShape(operand)); TF_ASSIGN_OR_RETURN( @@ -1009,23 +1060,98 @@ XlaOp XlaBuilder::Fft(const XlaOp& operand, const FftType fft_type, } XlaOp XlaBuilder::Infeed(const Shape& shape, const string& config) { - return NoteErrorOrReturn([&]() -> StatusOr { + return ReportErrorOrReturn([&]() -> StatusOr { HloInstructionProto instr; if (!LayoutUtil::HasLayout(shape)) { return InvalidArgument("Given shape to Infeed must have a layout"); } - *instr.mutable_shape() = shape; + const Shape infeed_instruction_shape = + ShapeUtil::MakeTupleShape({shape, ShapeUtil::MakeTokenShape()}); + *instr.mutable_shape() = infeed_instruction_shape; + instr.set_infeed_config(config); + + if (ShapeUtil::IsArray(shape) && sharding() && + sharding()->type() == OpSharding::Type::OpSharding_Type_OTHER) { + // TODO(b/110793772): Support tiled array-shaped infeeds. + return InvalidArgument( + "Tiled sharding is not yet supported for array-shaped infeeds"); + } + + if (sharding() && + sharding()->type() == OpSharding::Type::OpSharding_Type_REPLICATED) { + return InvalidArgument( + "Replicated sharding is not yet supported for infeeds"); + } + + // The sharding is set by the client according to the data tuple shape. + // However, the shape of the infeed instruction is a tuple containing the + // data and a token. For tuple sharding type, the sharding must be changed + // to accommodate the token. + XlaOp infeed; + if (sharding() && + sharding()->type() == OpSharding::Type::OpSharding_Type_TUPLE) { + // TODO(b/80000000): Remove this when clients have been updated to handle + // tokens. + OpSharding infeed_instruction_sharding = *sharding(); + // Arbitrarily assign the token to device 0. + *infeed_instruction_sharding.add_tuple_shardings() = + sharding_builder::AssignDevice(0); + XlaScopedShardingAssignment scoped_sharding(this, + infeed_instruction_sharding); + TF_ASSIGN_OR_RETURN(infeed, + AddInstruction(std::move(instr), HloOpcode::kInfeed)); + } else { + TF_ASSIGN_OR_RETURN(infeed, + AddInstruction(std::move(instr), HloOpcode::kInfeed)); + } + + // The infeed instruction produces a tuple of the infed data and a token + // type. Return XLA op containing the data. + // TODO(b/80000000): Remove this when clients have been updated to handle + // tokens. + HloInstructionProto infeed_data; + *infeed_data.mutable_shape() = shape; + infeed_data.set_tuple_index(0); + return AddInstruction(std::move(infeed_data), HloOpcode::kGetTupleElement, + {infeed}); + }); +} + +XlaOp XlaBuilder::InfeedWithToken(const XlaOp& token, const Shape& shape, + const string& config) { + return ReportErrorOrReturn([&]() -> StatusOr { + HloInstructionProto instr; + if (!LayoutUtil::HasLayout(shape)) { + return InvalidArgument("Given shape to Infeed must have a layout"); + } + const Shape infeed_instruction_shape = + ShapeUtil::MakeTupleShape({shape, ShapeUtil::MakeTokenShape()}); + *instr.mutable_shape() = infeed_instruction_shape; instr.set_infeed_config(config); - return AddInstruction(std::move(instr), HloOpcode::kInfeed); + + if (ShapeUtil::IsArray(shape) && sharding() && + sharding()->type() == OpSharding::Type::OpSharding_Type_OTHER) { + // TODO(b/110793772): Support tiled array-shaped infeeds. + return InvalidArgument( + "Tiled sharding is not yet supported for array-shaped infeeds"); + } + + if (sharding() && + sharding()->type() == OpSharding::Type::OpSharding_Type_REPLICATED) { + return InvalidArgument( + "Replicated sharding is not yet supported for infeeds"); + } + + return AddInstruction(std::move(instr), HloOpcode::kInfeed, {token}); }); } void XlaBuilder::Outfeed(const XlaOp& operand, const Shape& shape_with_layout, const string& outfeed_config) { - NoteErrorOrReturn([&]() -> StatusOr { + ReportErrorOrReturn([&]() -> StatusOr { HloInstructionProto instr; - *instr.mutable_shape() = ShapeUtil::MakeNil(); + *instr.mutable_shape() = ShapeUtil::MakeTokenShape(); // Check and set outfeed shape. if (!LayoutUtil::HasLayout(shape_with_layout)) { @@ -1042,14 +1168,80 @@ void XlaBuilder::Outfeed(const XlaOp& operand, const Shape& shape_with_layout, instr.set_outfeed_config(outfeed_config); - return AddInstruction(std::move(instr), HloOpcode::kOutfeed, {operand}); + TF_RETURN_IF_ERROR( + AddInstruction(std::move(instr), HloOpcode::kOutfeed, {operand}) + .status()); + + // The outfeed instruction produces a token. However, existing users expect + // a nil shape (empty tuple). This should only be relevant if the outfeed is + // the root of a computation. + // TODO(b/80000000): Remove this when clients have been updated to handle + // tokens. + HloInstructionProto tuple_instr; + *tuple_instr.mutable_shape() = ShapeUtil::MakeNil(); + + // The dummy tuple should have no sharding. + { + XlaScopedShardingAssignment scoped_sharding(this, OpSharding()); + TF_ASSIGN_OR_RETURN( + XlaOp empty_tuple, + AddInstruction(std::move(tuple_instr), HloOpcode::kTuple, {})); + return empty_tuple; + } + }); +} + +XlaOp XlaBuilder::OutfeedWithToken(const XlaOp& operand, const XlaOp& token, + const Shape& shape_with_layout, + const string& outfeed_config) { + return ReportErrorOrReturn([&]() -> StatusOr { + HloInstructionProto instr; + + *instr.mutable_shape() = ShapeUtil::MakeTokenShape(); + + // Check and set outfeed shape. + if (!LayoutUtil::HasLayout(shape_with_layout)) { + return InvalidArgument("Given shape to Outfeed must have a layout"); + } + TF_ASSIGN_OR_RETURN(const Shape& operand_shape, GetShape(operand)); + if (!ShapeUtil::Compatible(operand_shape, shape_with_layout)) { + return InvalidArgument( + "Outfeed shape %s must be compatible with operand shape %s", + ShapeUtil::HumanStringWithLayout(shape_with_layout).c_str(), + ShapeUtil::HumanStringWithLayout(operand_shape).c_str()); + } + *instr.mutable_outfeed_shape() = shape_with_layout; + + instr.set_outfeed_config(outfeed_config); + + return AddInstruction(std::move(instr), HloOpcode::kOutfeed, + {operand, token}); + }); +} + +XlaOp XlaBuilder::CreateToken() { + return ReportErrorOrReturn([&]() -> StatusOr { + HloInstructionProto instr; + *instr.mutable_shape() = ShapeUtil::MakeTokenShape(); + return AddInstruction(std::move(instr), HloOpcode::kAfterAll); + }); +} + +XlaOp XlaBuilder::AfterAll(tensorflow::gtl::ArraySlice tokens) { + return ReportErrorOrReturn([&]() -> StatusOr { + if (tokens.empty()) { + return InvalidArgument("AfterAll requires at least one operand"); + } + HloInstructionProto instr; + *instr.mutable_shape() = ShapeUtil::MakeTokenShape(); + return AddInstruction(std::move(instr), HloOpcode::kAfterAll, tokens); }); } XlaOp XlaBuilder::CustomCall(const string& call_target_name, tensorflow::gtl::ArraySlice operands, const Shape& shape) { - return NoteErrorOrReturn([&]() -> StatusOr { + return ReportErrorOrReturn([&]() -> StatusOr { HloInstructionProto instr; if (tensorflow::str_util::StartsWith(call_target_name, "$")) { return InvalidArgument( @@ -1066,7 +1258,7 @@ XlaOp XlaBuilder::CustomCall(const string& call_target_name, XlaOp XlaBuilder::HostCompute(tensorflow::gtl::ArraySlice operands, const string& channel_name, int64 cost_estimate_ns, const Shape& shape) { - return NoteErrorOrReturn([&]() -> StatusOr { + return ReportErrorOrReturn([&]() -> StatusOr { HloInstructionProto instr; *instr.mutable_shape() = shape; instr.set_channel_name(channel_name); @@ -1221,7 +1413,7 @@ XlaOp XlaBuilder::IsFinite(const XlaOp& operand) { XlaOp XlaBuilder::Transpose(const XlaOp& operand, tensorflow::gtl::ArraySlice permutation) { - return NoteErrorOrReturn([&]() -> StatusOr { + return ReportErrorOrReturn([&]() -> StatusOr { HloInstructionProto instr; TF_ASSIGN_OR_RETURN(const Shape& operand_shape, GetShape(operand)); TF_ASSIGN_OR_RETURN( @@ -1236,7 +1428,7 @@ XlaOp XlaBuilder::Transpose(const XlaOp& operand, XlaOp XlaBuilder::Rev(const XlaOp& operand, tensorflow::gtl::ArraySlice dimensions) { - return NoteErrorOrReturn([&]() -> StatusOr { + return ReportErrorOrReturn([&]() -> StatusOr { HloInstructionProto instr; TF_ASSIGN_OR_RETURN(const Shape& operand_shape, GetShape(operand)); TF_ASSIGN_OR_RETURN( @@ -1249,13 +1441,31 @@ XlaOp XlaBuilder::Rev(const XlaOp& operand, }); } -XlaOp XlaBuilder::Sort(const XlaOp& operand) { - return UnaryOp(HloOpcode::kSort, operand); -} - -XlaOp XlaBuilder::SqrtF32(const XlaOp& operand) { - return BinaryOp(HloOpcode::kPower, operand, ConstantR0(0.5), - /*broadcast_dimensions=*/{}); +XlaOp XlaBuilder::Sort(XlaOp keys, tensorflow::gtl::optional values, + int64 dimension) { + return ReportErrorOrReturn([&]() -> StatusOr { + HloInstructionProto instr; + std::vector operand_shape_ptrs; + TF_ASSIGN_OR_RETURN(const Shape& keys_shape, GetShape(keys)); + operand_shape_ptrs.push_back(&keys_shape); + Shape values_shape; + if (values.has_value()) { + TF_ASSIGN_OR_RETURN(values_shape, GetShape(*values)); + operand_shape_ptrs.push_back(&values_shape); + } + TF_ASSIGN_OR_RETURN(*instr.mutable_shape(), + ShapeInference::InferVariadicOpShape( + HloOpcode::kSort, operand_shape_ptrs)); + if (dimension == -1) { + TF_ASSIGN_OR_RETURN(const Shape& keys_shape, GetShape(keys)); + dimension = ShapeUtil::Rank(keys_shape) - 1; + } + instr.add_dimensions(dimension); + return values.has_value() + ? AddInstruction(std::move(instr), HloOpcode::kSort, + {keys, *values}) + : AddInstruction(std::move(instr), HloOpcode::kSort, {keys}); + }); } XlaOp XlaBuilder::Pow(const XlaOp& lhs, const XlaOp& rhs, @@ -1265,7 +1475,7 @@ XlaOp XlaBuilder::Pow(const XlaOp& lhs, const XlaOp& rhs, XlaOp XlaBuilder::ConvertElementType(const XlaOp& operand, PrimitiveType new_element_type) { - return NoteErrorOrReturn([&]() -> StatusOr { + return ReportErrorOrReturn([&]() -> StatusOr { HloInstructionProto instr; TF_ASSIGN_OR_RETURN(const Shape& operand_shape, GetShape(operand)); TF_ASSIGN_OR_RETURN( @@ -1277,7 +1487,7 @@ XlaOp XlaBuilder::ConvertElementType(const XlaOp& operand, XlaOp XlaBuilder::BitcastConvertType(const XlaOp& operand, PrimitiveType new_element_type) { - return NoteErrorOrReturn([&]() -> StatusOr { + return ReportErrorOrReturn([&]() -> StatusOr { HloInstructionProto instr; TF_ASSIGN_OR_RETURN(const Shape& operand_shape, GetShape(operand)); TF_ASSIGN_OR_RETURN( @@ -1288,16 +1498,6 @@ XlaOp XlaBuilder::BitcastConvertType(const XlaOp& operand, }); } -XlaOp XlaBuilder::SquareF32(const XlaOp& operand) { - return BinaryOp(HloOpcode::kPower, operand, ConstantR0(2.0), - /*broadcast_dimensions=*/{}); -} - -XlaOp XlaBuilder::ReciprocalF32(const XlaOp& operand) { - return BinaryOp(HloOpcode::kPower, operand, ConstantR0(-1.0), - /*broadcast_dimensions=*/{}); -} - XlaOp XlaBuilder::Neg(const XlaOp& operand) { return UnaryOp(HloOpcode::kNegate, operand); } @@ -1311,13 +1511,12 @@ XlaOp XlaBuilder::Map(tensorflow::gtl::ArraySlice operands, const XlaComputation& computation, tensorflow::gtl::ArraySlice dimensions, tensorflow::gtl::ArraySlice static_operands) { - return NoteErrorOrReturn([&]() -> StatusOr { + return ReportErrorOrReturn([&]() -> StatusOr { if (!static_operands.empty()) { return Unimplemented("static_operands is not supported in Map"); } HloInstructionProto instr; - std::vector operand_shape_ptrs; TF_ASSIGN_OR_RETURN(const auto& operand_shapes, GetOperandShapes(operands)); c_transform(operand_shapes, std::back_inserter(operand_shape_ptrs), @@ -1329,16 +1528,32 @@ XlaOp XlaBuilder::Map(tensorflow::gtl::ArraySlice operands, ShapeInference::InferMapShape(operand_shape_ptrs, called_program_shape, dimensions)); + const Shape& output_shape = instr.shape(); + const int64 output_rank = ShapeUtil::Rank(output_shape); AddCalledComputation(computation, &instr); + std::vector new_operands(operands.begin(), operands.end()); + for (XlaOp& new_operand : new_operands) { + TF_ASSIGN_OR_RETURN(Shape shape, GetShape(new_operand)); + const int64 rank = ShapeUtil::Rank(shape); + if (rank != output_rank) { + TF_ASSIGN_OR_RETURN(new_operand, + InDimBroadcast(output_shape, new_operand, {})); + TF_ASSIGN_OR_RETURN(shape, GetShape(new_operand)); + } + if (!ShapeUtil::SameDimensions(output_shape, shape)) { + TF_ASSIGN_OR_RETURN(new_operand, + AddBroadcastSequence(output_shape, new_operand)); + } + } - return AddInstruction(std::move(instr), HloOpcode::kMap, operands); + return AddInstruction(std::move(instr), HloOpcode::kMap, new_operands); }); } XlaOp XlaBuilder::RngOp(RandomDistribution distribution, tensorflow::gtl::ArraySlice parameters, const Shape& shape) { - return NoteErrorOrReturn([&]() -> StatusOr { + return ReportErrorOrReturn([&]() -> StatusOr { HloInstructionProto instr; // Check the number of parameters per RNG distribution. @@ -1376,7 +1591,7 @@ XlaOp XlaBuilder::RngUniform(const XlaOp& a, const XlaOp& b, XlaOp XlaBuilder::While(const XlaComputation& condition, const XlaComputation& body, const XlaOp& init) { - return NoteErrorOrReturn([&]() -> StatusOr { + return ReportErrorOrReturn([&]() -> StatusOr { HloInstructionProto instr; // Infer shape. @@ -1398,7 +1613,7 @@ XlaOp XlaBuilder::While(const XlaComputation& condition, XlaOp XlaBuilder::Gather(const XlaOp& input, const XlaOp& gather_indices, const GatherDimensionNumbers& dimension_numbers, tensorflow::gtl::ArraySlice window_bounds) { - return NoteErrorOrReturn([&]() -> StatusOr { + return ReportErrorOrReturn([&]() -> StatusOr { HloInstructionProto instr; TF_ASSIGN_OR_RETURN(const Shape& input_shape, GetShape(input)); @@ -1423,7 +1638,7 @@ XlaOp XlaBuilder::Conditional(const XlaOp& predicate, const XlaOp& true_operand, const XlaComputation& true_computation, const XlaOp& false_operand, const XlaComputation& false_computation) { - return NoteErrorOrReturn([&]() -> StatusOr { + return ReportErrorOrReturn([&]() -> StatusOr { HloInstructionProto instr; TF_ASSIGN_OR_RETURN(const Shape& predicate_shape, GetShape(predicate)); @@ -1455,13 +1670,14 @@ XlaOp XlaBuilder::Reduce( const XlaOp& operand, const XlaOp& init_value, const XlaComputation& computation, tensorflow::gtl::ArraySlice dimensions_to_reduce) { - return NoteErrorOrReturn([&]() -> StatusOr { + return ReportErrorOrReturn([&]() -> StatusOr { HloInstructionProto instr; TF_ASSIGN_OR_RETURN(const Shape& operand_shape, GetShape(operand)); TF_ASSIGN_OR_RETURN(const Shape& init_shape, GetShape(init_value)); TF_ASSIGN_OR_RETURN(const ProgramShape& called_program_shape, computation.GetProgramShape()); + TF_ASSIGN_OR_RETURN(*instr.mutable_shape(), ShapeInference::InferReduceShape( operand_shape, init_shape, dimensions_to_reduce, @@ -1480,7 +1696,7 @@ XlaOp XlaBuilder::Reduce( XlaOp XlaBuilder::ReduceAll(const XlaOp& operand, const XlaOp& init_value, const XlaComputation& computation) { - return NoteErrorOrReturn([&]() -> StatusOr { + return ReportErrorOrReturn([&]() -> StatusOr { TF_ASSIGN_OR_RETURN(const Shape& operand_shape, GetShape(operand)); std::vector all_dimnos(ShapeUtil::Rank(operand_shape)); std::iota(all_dimnos.begin(), all_dimnos.end(), 0); @@ -1493,7 +1709,7 @@ XlaOp XlaBuilder::ReduceWindow( const XlaComputation& computation, tensorflow::gtl::ArraySlice window_dimensions, tensorflow::gtl::ArraySlice window_strides, Padding padding) { - return NoteErrorOrReturn([&]() -> StatusOr { + return ReportErrorOrReturn([&]() -> StatusOr { HloInstructionProto instr; TF_ASSIGN_OR_RETURN(const Shape& operand_shape, GetShape(operand)); @@ -1516,7 +1732,7 @@ XlaOp XlaBuilder::ReduceWindowWithGeneralPadding( tensorflow::gtl::ArraySlice window_dimensions, tensorflow::gtl::ArraySlice window_strides, tensorflow::gtl::ArraySlice> padding) { - return NoteErrorOrReturn([&]() -> StatusOr { + return ReportErrorOrReturn([&]() -> StatusOr { HloInstructionProto instr; TF_ASSIGN_OR_RETURN(const Shape& operand_shape, GetShape(operand)); @@ -1540,7 +1756,7 @@ XlaOp XlaBuilder::ReduceWindowWithGeneralPadding( XlaOp XlaBuilder::BatchNormTraining(const XlaOp& operand, const XlaOp& scale, const XlaOp& offset, float epsilon, int64 feature_index) { - return NoteErrorOrReturn([&]() -> StatusOr { + return ReportErrorOrReturn([&]() -> StatusOr { HloInstructionProto instr; TF_ASSIGN_OR_RETURN(const Shape& operand_shape, GetShape(operand)); @@ -1563,7 +1779,7 @@ XlaOp XlaBuilder::BatchNormInference(const XlaOp& operand, const XlaOp& scale, const XlaOp& offset, const XlaOp& mean, const XlaOp& variance, float epsilon, int64 feature_index) { - return NoteErrorOrReturn([&]() -> StatusOr { + return ReportErrorOrReturn([&]() -> StatusOr { HloInstructionProto instr; TF_ASSIGN_OR_RETURN(const Shape& operand_shape, GetShape(operand)); @@ -1588,7 +1804,7 @@ XlaOp XlaBuilder::BatchNormGrad(const XlaOp& operand, const XlaOp& scale, const XlaOp& batch_mean, const XlaOp& batch_var, const XlaOp& grad_output, float epsilon, int64 feature_index) { - return NoteErrorOrReturn([&]() -> StatusOr { + return ReportErrorOrReturn([&]() -> StatusOr { HloInstructionProto instr; TF_ASSIGN_OR_RETURN(const Shape& operand_shape, GetShape(operand)); @@ -1612,7 +1828,7 @@ XlaOp XlaBuilder::BatchNormGrad(const XlaOp& operand, const XlaOp& scale, XlaOp XlaBuilder::CrossReplicaSum( const XlaOp& operand, tensorflow::gtl::ArraySlice replica_group_ids) { - return NoteErrorOrReturn([&]() -> StatusOr { + return ReportErrorOrReturn([&]() -> StatusOr { TF_ASSIGN_OR_RETURN(const Shape& shape, GetShape(operand)); const Shape& scalar_shape = ShapeUtil::MakeShape(shape.element_type(), {}); auto b = CreateSubBuilder("sum"); @@ -1628,7 +1844,7 @@ XlaOp XlaBuilder::CrossReplicaSum( const XlaOp& operand, const XlaComputation& computation, tensorflow::gtl::ArraySlice replica_group_ids, const tensorflow::gtl::optional& channel_id) { - return NoteErrorOrReturn([&]() -> StatusOr { + return ReportErrorOrReturn([&]() -> StatusOr { if (channel_id.has_value()) { return Unimplemented("channel_id is not supported in AllReduce"); } @@ -1655,7 +1871,7 @@ XlaOp XlaBuilder::SelectAndScatter( tensorflow::gtl::ArraySlice window_strides, Padding padding, const XlaOp& source, const XlaOp& init_value, const XlaComputation& scatter) { - return NoteErrorOrReturn([&]() -> StatusOr { + return ReportErrorOrReturn([&]() -> StatusOr { TF_ASSIGN_OR_RETURN(const Shape& operand_shape, GetShape(operand)); return SelectAndScatterWithGeneralPadding( operand, select, window_dimensions, window_strides, @@ -1672,7 +1888,7 @@ XlaOp XlaBuilder::SelectAndScatterWithGeneralPadding( tensorflow::gtl::ArraySlice> padding, const XlaOp& source, const XlaOp& init_value, const XlaComputation& scatter) { - return NoteErrorOrReturn([&]() -> StatusOr { + return ReportErrorOrReturn([&]() -> StatusOr { HloInstructionProto instr; TF_ASSIGN_OR_RETURN(const Shape& operand_shape, GetShape(operand)); @@ -1700,7 +1916,7 @@ XlaOp XlaBuilder::SelectAndScatterWithGeneralPadding( XlaOp XlaBuilder::ReducePrecision(const XlaOp& operand, const int exponent_bits, const int mantissa_bits) { - return NoteErrorOrReturn([&]() -> StatusOr { + return ReportErrorOrReturn([&]() -> StatusOr { HloInstructionProto instr; TF_ASSIGN_OR_RETURN(const Shape& operand_shape, GetShape(operand)); TF_ASSIGN_OR_RETURN(*instr.mutable_shape(), @@ -1714,20 +1930,51 @@ XlaOp XlaBuilder::ReducePrecision(const XlaOp& operand, const int exponent_bits, } void XlaBuilder::Send(const XlaOp& operand, const ChannelHandle& handle) { - NoteErrorOrReturn([&]() -> StatusOr { - HloInstructionProto instr; + ReportErrorOrReturn([&]() -> StatusOr { + // Send HLO takes two operands: a data operand and a token. Generate the + // token to pass into the send. + // TODO(b/80000000): Remove this when clients have been updated to handle + // tokens. + HloInstructionProto token_instr; + *token_instr.mutable_shape() = ShapeUtil::MakeTokenShape(); + TF_ASSIGN_OR_RETURN(XlaOp token, AddInstruction(std::move(token_instr), + HloOpcode::kAfterAll, {})); + + // Send instruction produces a tuple of {aliased operand, U32 context, + // token}. + HloInstructionProto send_instr; + TF_ASSIGN_OR_RETURN(const Shape& shape, GetShape(operand)); + *send_instr.mutable_shape() = ShapeUtil::MakeTupleShape( + {shape, ShapeUtil::MakeShape(U32, {}), ShapeUtil::MakeTokenShape()}); + send_instr.set_channel_id(handle.handle()); + TF_ASSIGN_OR_RETURN(XlaOp send, + AddInstruction(std::move(send_instr), HloOpcode::kSend, + {operand, token})); + + HloInstructionProto send_done_instr; + *send_done_instr.mutable_shape() = ShapeUtil::MakeTokenShape(); + send_done_instr.set_channel_id(handle.handle()); + return AddInstruction(std::move(send_done_instr), HloOpcode::kSendDone, + {send}); + }); +} - // Send instruction produces a tuple of {aliased operand, U32 context}. +XlaOp XlaBuilder::SendWithToken(const XlaOp& operand, const XlaOp& token, + const ChannelHandle& handle) { + return ReportErrorOrReturn([&]() -> StatusOr { + // Send instruction produces a tuple of {aliased operand, U32 context, + // token}. + HloInstructionProto send_instr; TF_ASSIGN_OR_RETURN(const Shape& shape, GetShape(operand)); - *instr.mutable_shape() = - ShapeUtil::MakeTupleShape({shape, ShapeUtil::MakeShape(U32, {})}); - instr.set_channel_id(handle.handle()); - TF_ASSIGN_OR_RETURN( - XlaOp send, - AddInstruction(std::move(instr), HloOpcode::kSend, {operand})); + *send_instr.mutable_shape() = ShapeUtil::MakeTupleShape( + {shape, ShapeUtil::MakeShape(U32, {}), ShapeUtil::MakeTokenShape()}); + send_instr.set_channel_id(handle.handle()); + TF_ASSIGN_OR_RETURN(XlaOp send, + AddInstruction(std::move(send_instr), HloOpcode::kSend, + {operand, token})); HloInstructionProto send_done_instr; - *send_done_instr.mutable_shape() = ShapeUtil::MakeNil(); + *send_done_instr.mutable_shape() = ShapeUtil::MakeTokenShape(); send_done_instr.set_channel_id(handle.handle()); return AddInstruction(std::move(send_done_instr), HloOpcode::kSendDone, {send}); @@ -1735,18 +1982,60 @@ void XlaBuilder::Send(const XlaOp& operand, const ChannelHandle& handle) { } XlaOp XlaBuilder::Recv(const Shape& shape, const ChannelHandle& handle) { - return NoteErrorOrReturn([&]() -> StatusOr { - HloInstructionProto instr; + return ReportErrorOrReturn([&]() -> StatusOr { + // Recv HLO takes a single token operand. Generate the token to pass into + // the Recv and RecvDone instructions. + // TODO(b/80000000): Remove this when clients have been updated to handle + // tokens. + HloInstructionProto token_instr; + *token_instr.mutable_shape() = ShapeUtil::MakeTokenShape(); + TF_ASSIGN_OR_RETURN(XlaOp token, AddInstruction(std::move(token_instr), + HloOpcode::kAfterAll, {})); + + // Recv instruction produces a tuple of {receive buffer, U32 context, + // token}. + HloInstructionProto recv_instr; + *recv_instr.mutable_shape() = ShapeUtil::MakeTupleShape( + {shape, ShapeUtil::MakeShape(U32, {}), ShapeUtil::MakeTokenShape()}); + recv_instr.set_channel_id(handle.handle()); + TF_ASSIGN_OR_RETURN(XlaOp recv, AddInstruction(std::move(recv_instr), + HloOpcode::kRecv, {token})); - // Recv instruction produces a tuple of {receive buffer, U32 context}. - *instr.mutable_shape() = - ShapeUtil::MakeTupleShape({shape, ShapeUtil::MakeShape(U32, {})}); - instr.set_channel_id(handle.handle()); - TF_ASSIGN_OR_RETURN(XlaOp recv, - AddInstruction(std::move(instr), HloOpcode::kRecv, {})); + HloInstructionProto recv_done_instr; + *recv_done_instr.mutable_shape() = + ShapeUtil::MakeTupleShape({shape, ShapeUtil::MakeTokenShape()}); + recv_done_instr.set_channel_id(handle.handle()); + TF_ASSIGN_OR_RETURN(XlaOp recv_done, + AddInstruction(std::move(recv_done_instr), + HloOpcode::kRecvDone, {recv})); + + // The RecvDone instruction produces a tuple of the data and a token + // type. Return XLA op containing the data. + // TODO(b/80000000): Remove this when clients have been updated to handle + // tokens. + HloInstructionProto recv_data; + *recv_data.mutable_shape() = shape; + recv_data.set_tuple_index(0); + return AddInstruction(std::move(recv_data), HloOpcode::kGetTupleElement, + {recv_done}); + }); +} + +XlaOp XlaBuilder::RecvWithToken(const XlaOp& token, const Shape& shape, + const ChannelHandle& handle) { + return ReportErrorOrReturn([&]() -> StatusOr { + // Recv instruction produces a tuple of {receive buffer, U32 context, + // token}. + HloInstructionProto recv_instr; + *recv_instr.mutable_shape() = ShapeUtil::MakeTupleShape( + {shape, ShapeUtil::MakeShape(U32, {}), ShapeUtil::MakeTokenShape()}); + recv_instr.set_channel_id(handle.handle()); + TF_ASSIGN_OR_RETURN(XlaOp recv, AddInstruction(std::move(recv_instr), + HloOpcode::kRecv, {token})); HloInstructionProto recv_done_instr; - *recv_done_instr.mutable_shape() = shape; + *recv_done_instr.mutable_shape() = + ShapeUtil::MakeTupleShape({shape, ShapeUtil::MakeTokenShape()}); recv_done_instr.set_channel_id(handle.handle()); return AddInstruction(std::move(recv_done_instr), HloOpcode::kRecvDone, {recv}); @@ -1990,4 +2279,526 @@ StatusOr XlaBuilder::LookUpInstruction( return &instructions_[op.handle()]; } +// Enqueues a "retrieve parameter value" instruction for a parameter that was +// passed to the computation. +XlaOp Parameter(XlaBuilder* builder, int64 parameter_number, const Shape& shape, + const string& name) { + return builder->Parameter(parameter_number, shape, name); +} + +// Enqueues a constant with the value of the given literal onto the +// computation. +XlaOp ConstantLiteral(XlaBuilder* builder, const LiteralSlice& literal) { + return builder->ConstantLiteral(literal); +} + +XlaOp Broadcast(const XlaOp& operand, + tensorflow::gtl::ArraySlice broadcast_sizes) { + return operand.builder()->Broadcast(operand, broadcast_sizes); +} + +XlaOp BroadcastInDim( + const XlaOp& operand, const Shape& shape, + const tensorflow::gtl::ArraySlice broadcast_dimensions) { + return operand.builder()->BroadcastInDim(operand, shape, + broadcast_dimensions); +} + +XlaOp Pad(const XlaOp& operand, const XlaOp& padding_value, + const PaddingConfig& padding_config) { + return operand.builder()->Pad(operand, padding_value, padding_config); +} + +XlaOp Reshape(const XlaOp& operand, + tensorflow::gtl::ArraySlice dimensions, + tensorflow::gtl::ArraySlice new_sizes) { + return operand.builder()->Reshape(operand, dimensions, new_sizes); +} + +XlaOp Reshape(const XlaOp& operand, + tensorflow::gtl::ArraySlice new_sizes) { + return operand.builder()->Reshape(operand, new_sizes); +} + +XlaOp Collapse(const XlaOp& operand, + tensorflow::gtl::ArraySlice dimensions) { + return operand.builder()->Collapse(operand, dimensions); +} + +XlaOp Slice(const XlaOp& operand, + tensorflow::gtl::ArraySlice start_indices, + tensorflow::gtl::ArraySlice limit_indices, + tensorflow::gtl::ArraySlice strides) { + return operand.builder()->Slice(operand, start_indices, limit_indices, + strides); +} + +XlaOp SliceInDim(const XlaOp& operand, int64 start_index, int64 limit_index, + int64 stride, int64 dimno) { + return operand.builder()->SliceInDim(operand, start_index, limit_index, + stride, dimno); +} + +XlaOp DynamicSlice(const XlaOp& operand, const XlaOp& start_indices, + tensorflow::gtl::ArraySlice slice_sizes) { + return operand.builder()->DynamicSlice(operand, start_indices, slice_sizes); +} + +XlaOp DynamicUpdateSlice(const XlaOp& operand, const XlaOp& update, + const XlaOp& start_indices) { + return operand.builder()->DynamicUpdateSlice(operand, update, start_indices); +} + +XlaOp ConcatInDim(XlaBuilder* builder, + tensorflow::gtl::ArraySlice operands, + int64 dimension) { + return builder->ConcatInDim(operands, dimension); +} + +void Trace(const string& tag, const XlaOp& operand) { + return operand.builder()->Trace(tag, operand); +} + +XlaOp Select(const XlaOp& pred, const XlaOp& on_true, const XlaOp& on_false) { + return pred.builder()->Select(pred, on_true, on_false); +} + +XlaOp Tuple(XlaBuilder* builder, tensorflow::gtl::ArraySlice elements) { + return builder->Tuple(elements); +} + +XlaOp GetTupleElement(const XlaOp& tuple_data, int64 index) { + return tuple_data.builder()->GetTupleElement(tuple_data, index); +} + +XlaOp Eq(const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice broadcast_dimensions) { + return lhs.builder()->Eq(lhs, rhs, broadcast_dimensions); +} + +XlaOp Ne(const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice broadcast_dimensions) { + return lhs.builder()->Ne(lhs, rhs, broadcast_dimensions); +} + +XlaOp Ge(const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice broadcast_dimensions) { + return lhs.builder()->Ge(lhs, rhs, broadcast_dimensions); +} + +XlaOp Gt(const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice broadcast_dimensions) { + return lhs.builder()->Gt(lhs, rhs, broadcast_dimensions); +} + +XlaOp Lt(const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice broadcast_dimensions) { + return lhs.builder()->Lt(lhs, rhs, broadcast_dimensions); +} + +XlaOp Le(const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice broadcast_dimensions) { + return lhs.builder()->Le(lhs, rhs, broadcast_dimensions); +} + +XlaOp Dot(const XlaOp& lhs, const XlaOp& rhs) { + return lhs.builder()->Dot(lhs, rhs); +} + +XlaOp DotGeneral(const XlaOp& lhs, const XlaOp& rhs, + const DotDimensionNumbers& dimension_numbers) { + return lhs.builder()->DotGeneral(lhs, rhs, dimension_numbers); +} + +XlaOp Conv(const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice window_strides, Padding padding) { + return lhs.builder()->Conv(lhs, rhs, window_strides, padding); +} + +XlaOp ConvWithGeneralPadding( + const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice window_strides, + tensorflow::gtl::ArraySlice> padding) { + return lhs.builder()->ConvWithGeneralPadding(lhs, rhs, window_strides, + padding); +} + +XlaOp ConvWithGeneralDimensions( + const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice window_strides, Padding padding, + const ConvolutionDimensionNumbers& dimension_numbers) { + return lhs.builder()->ConvWithGeneralDimensions(lhs, rhs, window_strides, + padding, dimension_numbers); +} + +XlaOp ConvGeneral(const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice window_strides, + tensorflow::gtl::ArraySlice> padding, + const ConvolutionDimensionNumbers& dimension_numbers) { + return lhs.builder()->ConvGeneral(lhs, rhs, window_strides, padding, + dimension_numbers); +} + +XlaOp ConvGeneralDilated( + const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice window_strides, + tensorflow::gtl::ArraySlice> padding, + tensorflow::gtl::ArraySlice lhs_dilation, + tensorflow::gtl::ArraySlice rhs_dilation, + const ConvolutionDimensionNumbers& dimension_numbers) { + return lhs.builder()->ConvGeneralDilated(lhs, rhs, window_strides, padding, + lhs_dilation, rhs_dilation, + dimension_numbers); +} + +XlaOp Fft(const XlaOp& operand, FftType fft_type, + tensorflow::gtl::ArraySlice fft_length) { + return operand.builder()->Fft(operand, fft_type, fft_length); +} + +XlaOp Infeed(XlaBuilder* builder, const Shape& shape, const string& config) { + return builder->Infeed(shape, config); +} + +void Outfeed(const XlaOp& operand, const Shape& shape_with_layout, + const string& outfeed_config) { + return operand.builder()->Outfeed(operand, shape_with_layout, outfeed_config); +} + +XlaOp Call(XlaBuilder* builder, const XlaComputation& computation, + tensorflow::gtl::ArraySlice operands) { + return builder->Call(computation, operands); +} + +XlaOp CustomCall(XlaBuilder* builder, const string& call_target_name, + tensorflow::gtl::ArraySlice operands, + const Shape& shape) { + return builder->CustomCall(call_target_name, operands, shape); +} + +XlaOp HostCompute(XlaBuilder* builder, + tensorflow::gtl::ArraySlice operands, + const string& channel_name, int64 cost_estimate_ns, + const Shape& shape) { + return builder->HostCompute(operands, channel_name, cost_estimate_ns, shape); +} + +XlaOp Complex(const XlaOp& real, const XlaOp& imag, + tensorflow::gtl::ArraySlice broadcast_dimensions) { + return real.builder()->Complex(real, imag, broadcast_dimensions); +} + +XlaOp Conj(const XlaOp& operand) { return operand.builder()->Conj(operand); } + +XlaOp Add(const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice broadcast_dimensions) { + return lhs.builder()->Add(lhs, rhs, broadcast_dimensions); +} + +XlaOp Sub(const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice broadcast_dimensions) { + return lhs.builder()->Sub(lhs, rhs, broadcast_dimensions); +} + +XlaOp Mul(const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice broadcast_dimensions) { + return lhs.builder()->Mul(lhs, rhs, broadcast_dimensions); +} + +XlaOp Div(const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice broadcast_dimensions) { + return lhs.builder()->Div(lhs, rhs, broadcast_dimensions); +} + +XlaOp Rem(const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice broadcast_dimensions) { + return lhs.builder()->Rem(lhs, rhs, broadcast_dimensions); +} + +XlaOp Max(const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice broadcast_dimensions) { + return lhs.builder()->Max(lhs, rhs, broadcast_dimensions); +} + +XlaOp Min(const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice broadcast_dimensions) { + return lhs.builder()->Min(lhs, rhs, broadcast_dimensions); +} + +XlaOp And(const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice broadcast_dimensions) { + return lhs.builder()->And(lhs, rhs, broadcast_dimensions); +} + +XlaOp Or(const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice broadcast_dimensions) { + return lhs.builder()->Or(lhs, rhs, broadcast_dimensions); +} + +XlaOp Xor(const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice broadcast_dimensions) { + return lhs.builder()->Xor(lhs, rhs, broadcast_dimensions); +} + +XlaOp Not(const XlaOp& operand) { return operand.builder()->Not(operand); } + +XlaOp ShiftLeft(const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice broadcast_dimensions) { + return lhs.builder()->ShiftLeft(lhs, rhs, broadcast_dimensions); +} + +XlaOp ShiftRightArithmetic( + const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice broadcast_dimensions) { + return lhs.builder()->ShiftRightArithmetic(lhs, rhs, broadcast_dimensions); +} + +XlaOp ShiftRightLogical( + const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice broadcast_dimensions) { + return lhs.builder()->ShiftRightLogical(lhs, rhs, broadcast_dimensions); +} + +XlaOp Reduce(const XlaOp& operand, const XlaOp& init_value, + const XlaComputation& computation, + tensorflow::gtl::ArraySlice dimensions_to_reduce) { + return operand.builder()->Reduce(operand, init_value, computation, + dimensions_to_reduce); +} + +XlaOp ReduceAll(const XlaOp& operand, const XlaOp& init_value, + const XlaComputation& computation) { + return operand.builder()->ReduceAll(operand, init_value, computation); +} + +XlaOp ReduceWindow(const XlaOp& operand, const XlaOp& init_value, + const XlaComputation& computation, + tensorflow::gtl::ArraySlice window_dimensions, + tensorflow::gtl::ArraySlice window_strides, + Padding padding) { + return operand.builder()->ReduceWindow(operand, init_value, computation, + window_dimensions, window_strides, + padding); +} + +XlaOp ReduceWindowWithGeneralPadding( + const XlaOp& operand, const XlaOp& init_value, + const XlaComputation& computation, + tensorflow::gtl::ArraySlice window_dimensions, + tensorflow::gtl::ArraySlice window_strides, + tensorflow::gtl::ArraySlice> padding) { + return operand.builder()->ReduceWindowWithGeneralPadding( + operand, init_value, computation, window_dimensions, window_strides, + padding); +} + +XlaOp CrossReplicaSum(const XlaOp& operand, + tensorflow::gtl::ArraySlice replica_group_ids) { + return operand.builder()->CrossReplicaSum(operand, replica_group_ids); +} + +XlaOp CrossReplicaSum( + const XlaOp& operand, const XlaComputation& computation, + tensorflow::gtl::ArraySlice replica_group_ids, + const tensorflow::gtl::optional& channel_id) { + return operand.builder()->CrossReplicaSum(operand, computation, + replica_group_ids, channel_id); +} + +XlaOp SelectAndScatter(const XlaOp& operand, const XlaComputation& select, + tensorflow::gtl::ArraySlice window_dimensions, + tensorflow::gtl::ArraySlice window_strides, + Padding padding, const XlaOp& source, + const XlaOp& init_value, const XlaComputation& scatter) { + return operand.builder()->SelectAndScatter(operand, select, window_dimensions, + window_strides, padding, source, + init_value, scatter); +} + +XlaOp SelectAndScatterWithGeneralPadding( + const XlaOp& operand, const XlaComputation& select, + tensorflow::gtl::ArraySlice window_dimensions, + tensorflow::gtl::ArraySlice window_strides, + tensorflow::gtl::ArraySlice> padding, + const XlaOp& source, const XlaOp& init_value, + const XlaComputation& scatter) { + return operand.builder()->SelectAndScatterWithGeneralPadding( + operand, select, window_dimensions, window_strides, padding, source, + init_value, scatter); +} + +XlaOp Abs(const XlaOp& operand) { return operand.builder()->Abs(operand); } + +XlaOp Atan2(const XlaOp& y, const XlaOp& x, + tensorflow::gtl::ArraySlice broadcast_dimensions) { + return y.builder()->Atan2(y, x, broadcast_dimensions); +} + +XlaOp Exp(const XlaOp& operand) { return operand.builder()->Exp(operand); } + +XlaOp Expm1(const XlaOp& operand) { return operand.builder()->Expm1(operand); } + +XlaOp Floor(const XlaOp& operand) { return operand.builder()->Floor(operand); } + +XlaOp Ceil(const XlaOp& operand) { return operand.builder()->Ceil(operand); } + +XlaOp Round(const XlaOp& operand) { return operand.builder()->Round(operand); } + +XlaOp Log(const XlaOp& operand) { return operand.builder()->Log(operand); } + +XlaOp Log1p(const XlaOp& operand) { return operand.builder()->Log1p(operand); } + +XlaOp Sign(const XlaOp& operand) { return operand.builder()->Sign(operand); } + +XlaOp Clz(const XlaOp& operand) { return operand.builder()->Clz(operand); } + +XlaOp Cos(const XlaOp& operand) { return operand.builder()->Cos(operand); } + +XlaOp Sin(const XlaOp& operand) { return operand.builder()->Sin(operand); } + +XlaOp Tanh(const XlaOp& operand) { return operand.builder()->Tanh(operand); } + +XlaOp Real(const XlaOp& operand) { return operand.builder()->Real(operand); } + +XlaOp Imag(const XlaOp& operand) { return operand.builder()->Imag(operand); } + +XlaOp Pow(const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice broadcast_dimensions) { + return lhs.builder()->Pow(lhs, rhs, broadcast_dimensions); +} + +XlaOp IsFinite(const XlaOp& operand) { + return operand.builder()->IsFinite(operand); +} + +XlaOp ConvertElementType(const XlaOp& operand, PrimitiveType new_element_type) { + return operand.builder()->ConvertElementType(operand, new_element_type); +} + +XlaOp BitcastConvertType(const XlaOp& operand, PrimitiveType new_element_type) { + return operand.builder()->BitcastConvertType(operand, new_element_type); +} + +XlaOp Neg(const XlaOp& operand) { return operand.builder()->Neg(operand); } + +XlaOp Transpose(const XlaOp& operand, + tensorflow::gtl::ArraySlice permutation) { + return operand.builder()->Transpose(operand, permutation); +} + +XlaOp Rev(const XlaOp& operand, tensorflow::gtl::ArraySlice dimensions) { + return operand.builder()->Rev(operand, dimensions); +} + +XlaOp Sort(XlaOp keys, tensorflow::gtl::optional values, + int64 dimension) { + return keys.builder()->Sort(keys, std::move(values), dimension); +} + +XlaOp Clamp(const XlaOp& min, const XlaOp& operand, const XlaOp& max) { + return min.builder()->Clamp(min, operand, max); +} + +XlaOp Map(XlaBuilder* builder, tensorflow::gtl::ArraySlice operands, + const XlaComputation& computation, + tensorflow::gtl::ArraySlice dimensions, + tensorflow::gtl::ArraySlice static_operands) { + return builder->Map(operands, computation, dimensions, static_operands); +} + +XlaOp RngNormal(const XlaOp& mu, const XlaOp& sigma, const Shape& shape) { + return mu.builder()->RngNormal(mu, sigma, shape); +} + +XlaOp RngUniform(const XlaOp& a, const XlaOp& b, const Shape& shape) { + return a.builder()->RngUniform(a, b, shape); +} + +XlaOp While(const XlaComputation& condition, const XlaComputation& body, + const XlaOp& init) { + return init.builder()->While(condition, body, init); +} + +XlaOp Conditional(const XlaOp& predicate, const XlaOp& true_operand, + const XlaComputation& true_computation, + const XlaOp& false_operand, + const XlaComputation& false_computation) { + return predicate.builder()->Conditional(predicate, true_operand, + true_computation, false_operand, + false_computation); +} + +XlaOp ReducePrecision(const XlaOp& operand, const int exponent_bits, + const int mantissa_bits) { + return operand.builder()->ReducePrecision(operand, exponent_bits, + mantissa_bits); +} + +XlaOp Gather(const XlaOp& input, const XlaOp& gather_indices, + const GatherDimensionNumbers& dimension_numbers, + tensorflow::gtl::ArraySlice window_bounds) { + return input.builder()->Gather(input, gather_indices, dimension_numbers, + window_bounds); +} + +void Send(const XlaOp& operand, const ChannelHandle& handle) { + return operand.builder()->Send(operand, handle); +} + +XlaOp Recv(XlaBuilder* builder, const Shape& shape, + const ChannelHandle& handle) { + return builder->Recv(shape, handle); +} + +XlaOp SendWithToken(const XlaOp& operand, const XlaOp& token, + const ChannelHandle& handle) { + return operand.builder()->SendWithToken(operand, token, handle); +} + +XlaOp RecvWithToken(const XlaOp& token, const Shape& shape, + const ChannelHandle& handle) { + return token.builder()->RecvWithToken(token, shape, handle); +} + +XlaOp InfeedWithToken(const XlaOp& token, const Shape& shape, + const string& config) { + return token.builder()->InfeedWithToken(token, shape, config); +} + +XlaOp OutfeedWithToken(const XlaOp& operand, const XlaOp& token, + const Shape& shape_with_layout, + const string& outfeed_config) { + return operand.builder()->OutfeedWithToken(operand, token, shape_with_layout, + outfeed_config); +} + +XlaOp CreateToken(XlaBuilder* builder) { return builder->CreateToken(); } + +XlaOp AfterAll(XlaBuilder* builder, tensorflow::gtl::ArraySlice tokens) { + return builder->AfterAll(tokens); +} + +XlaOp BatchNormTraining(const XlaOp& operand, const XlaOp& scale, + const XlaOp& offset, float epsilon, + int64 feature_index) { + return operand.builder()->BatchNormTraining(operand, scale, offset, epsilon, + feature_index); +} + +XlaOp BatchNormInference(const XlaOp& operand, const XlaOp& scale, + const XlaOp& offset, const XlaOp& mean, + const XlaOp& variance, float epsilon, + int64 feature_index) { + return operand.builder()->BatchNormInference( + operand, scale, offset, mean, variance, epsilon, feature_index); +} + +XlaOp BatchNormGrad(const XlaOp& operand, const XlaOp& scale, + const XlaOp& batch_mean, const XlaOp& batch_var, + const XlaOp& grad_output, float epsilon, + int64 feature_index) { + return operand.builder()->BatchNormGrad(operand, scale, batch_mean, batch_var, + grad_output, epsilon, feature_index); +} + } // namespace xla diff --git a/tensorflow/compiler/xla/client/xla_client/xla_builder.h b/tensorflow/compiler/xla/client/xla_client/xla_builder.h index f18306fff080db31a93a782e8ed51cee97d7e5cf..2be6f4a553e7fa4198ba361a2817eeda93377b48 100644 --- a/tensorflow/compiler/xla/client/xla_client/xla_builder.h +++ b/tensorflow/compiler/xla/client/xla_client/xla_builder.h @@ -18,10 +18,12 @@ limitations under the License. #include #include +#include #include #include "tensorflow/compiler/xla/client/padding.h" #include "tensorflow/compiler/xla/client/xla_client/xla_computation.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/literal_util.h" #include "tensorflow/compiler/xla/service/hlo.pb.h" #include "tensorflow/compiler/xla/service/hlo_opcode.h" @@ -46,22 +48,25 @@ class XlaBuilder; // instruction as an operand. class XlaOp { public: - XlaOp() : handle_(-1), builder_(nullptr) {} - ~XlaOp() {} - - XlaBuilder* builder() const { return builder_; } - - bool operator==(const XlaOp& rhs) const { - return handle_ == rhs.handle_ && builder_ == rhs.builder_; + XlaOp() : handle_(-1), builder_(nullptr) { + static_assert(std::is_trivially_destructible::value, + "XlaOp should be trivially destructible"); } + ~XlaOp() = default; - bool operator!=(const XlaOp& rhs) const { - return handle_ != rhs.handle_ || builder_ != rhs.builder_; - } + XlaBuilder* builder() const { return builder_; } // Returns true if the XlaOp represents valid, non-erroneous value. bool valid() const { return handle_ >= 0; } + // Returns true if the XlaOp was created by the XlaOp() constructor and + // not returned by a builder. + bool IsUninitialized() const { return builder_ == nullptr; } + + bool IsIdenticalTo(const XlaOp& rhs) const { + return handle_ == rhs.handle_ && builder_ == rhs.builder_; + } + friend std::ostream& operator<<(std::ostream& out, const XlaOp& op) { out << op.handle(); return out; @@ -84,6 +89,30 @@ class XlaOp { XlaBuilder* builder_; }; +// Arithmetic operator overloads for the XlaOp type. +XlaOp operator-(const XlaOp& x); +XlaOp operator+(const XlaOp& x, const XlaOp& y); +XlaOp operator-(const XlaOp& x, const XlaOp& y); +XlaOp operator*(const XlaOp& x, const XlaOp& y); +XlaOp operator/(const XlaOp& x, const XlaOp& y); +XlaOp operator%(const XlaOp& x, const XlaOp& y); + +// Bitwise operator overloads for the XlaOp type. +XlaOp operator~(const XlaOp& x); +XlaOp operator&(const XlaOp& x, const XlaOp& y); +XlaOp operator|(const XlaOp& x, const XlaOp& y); +XlaOp operator^(const XlaOp& x, const XlaOp& y); +XlaOp operator<<(const XlaOp& x, const XlaOp& y); +// Performs a right arithmetic shift if 'x' is a signed type, otherwise performs +// a right logical shift. +XlaOp operator>>(const XlaOp& x, const XlaOp& y); + +// We don't overload the relational operators (==, !=, <, <=, >, >=) because the +// semantics might be surprising since their result types are usually 'bool'. +// Further programmers may expect == to be a structural equality. +// We also choose not to overload any of the mutating operators (e.g., +=, -=) +// because the semantics might be misleading ā€” XLA computations are immutable. + // A convenient interface for building up computations. // // Thread-compatible. @@ -130,6 +159,93 @@ class XlaBuilder { die_immediately_on_error_ = enabled; } + // Default dimension numbers used for a 2D convolution. + static constexpr int64 kConvBatchDimension = 0; + static constexpr int64 kConvFeatureDimension = 1; + static constexpr int64 kConvFirstSpatialDimension = 2; + static constexpr int64 kConvSecondSpatialDimension = 3; + static constexpr int64 kConvKernelOutputDimension = 0; + static constexpr int64 kConvKernelInputDimension = 1; + static constexpr int64 kConvKernelFirstSpatialDimension = 2; + static constexpr int64 kConvKernelSecondSpatialDimension = 3; + + // Creates a default ConvolutionDimensionNumbers. For a 2D convolution, for + // the input operand {batch, feature, height, width} = {0, 1, 2, 3} and for + // the kernel operand + // {output_feature, input_feature, height, width} = {0, 1, 2, 3}. + static ConvolutionDimensionNumbers CreateDefaultConvDimensionNumbers( + int num_spatial_dims = 2); + + // Returns an error if the convolution dimension numbers have conflicts. + static Status Validate(const ConvolutionDimensionNumbers& dnum); + + // Returns a new XlaBuilder whose resultant Computation is used only by this + // XlaBuilder. The sub-XlaBuilder has the same die_immediately_on_error + // behavior as the parent. + std::unique_ptr CreateSubBuilder(const string& computation_name); + + // Builds the computation with the requested operations, or returns a non-ok + // status. Note that all ops that have been enqueued will be moved to the + // computation being returned. + StatusOr Build(); + + // Builds the computation with the requested operations, or notes an error in + // the parent XlaBuilder and returns an empty computation if building failed. + // This function is intended to be used where the returned XlaComputation is + // only used by the parent XlaBuilder and hence further operation on the + // returned XlaComputation will simply be error'ed out if an error occurred + // while building this computation. If the built computation is to be used by + // a XlaBuilder other than the parent XlaBuilder then Build() should be used + // instead. + XlaComputation BuildAndNoteError(); + + // Returns a subgraph that roots on the given root. If the root is not a + // compile-time constant (see `IsConstant`), returns an error. + // + // This will copy the needed ops/computations to the subgraph. + StatusOr BuildConstantSubGraph(const XlaOp& root_op) const; + + // Returns the first error that was encountered while building the + // computation. When an error is encountered, by default we return a vacuous + // XlaOp and inform the user of the error that occurred while + // building the computation when they make a final call to Build(). + // + // See also set_die_immediately_on_error(). + Status first_error() const { return first_error_; } + + // Returns the shape of the given op. + StatusOr GetShape(const XlaOp& op) const; + + // Returns the (inferred) result for the current computation's shape. + StatusOr GetProgramShape() const; + + // Reports an error to the builder, by + // * storing it internally and capturing a backtrace if it's the first error + // (this deferred value will be produced on the call to + // Build()/GetShape()/...) + // * dying if die_immediately_on_error_ is true. + // Returns an XlaOp with an invalid handle but a valid builder. This value can + // be returned in place of a value in APIs that return an XlaOp. + XlaOp ReportError(const Status& error); + + // A helper function that converts a StatusOr into an XlaOp. + // If the Status was an error, reports the error to builder and returns an + // invalid XlaOp handle. + XlaOp ReportErrorOrReturn(const StatusOr& op); + + // A helper function that runs a function that returns a StatusOr and + // returns an XlaOp. + XlaOp ReportErrorOrReturn(const std::function()>& op_creator); + + // Returns true if 'operand' is a compile-time constant. A compile-time + // constant does not depend on any parameters, or on stateful operators such + // as `RngNormal` or `Infeed`. + // + // This tests whether a computation is a compile-time constant without + // evaluating the computation. + StatusOr IsConstant(const XlaOp& operand) const; + + private: // Enqueues a "retrieve parameter value" instruction for a parameter that was // passed to the computation. XlaOp Parameter(int64 parameter_number, const Shape& shape, @@ -202,6 +318,27 @@ class XlaBuilder { XlaOp Broadcast(const XlaOp& operand, tensorflow::gtl::ArraySlice broadcast_sizes); + // Performs in-dimension-style broadcast. + // + // Operand specifies the input to be broadcast. "shape" is expected output + // shape. "broadcast_dimensions" are the dimensions to be broadcasting into. + // Dimension numbers in broadcast_dimensions map to individual dimensions + // of the operand, and specify what dimension of the output shape they + // should be broadcast. + // e.g. + // Say operand = [1, 2], i.e., a 1D tensor with 2 elements. + // and dimension of shape is [2,2]. + // Specifying {1} as brodcast_dimension will generate output + // [1 , 2] + // [1 , 2] + // On the other hand, specifying {0} as broadcast_dimension + // will generate output + // [1 , 1] + // [2 , 2] + XlaOp BroadcastInDim( + const XlaOp& operand, const Shape& shape, + const tensorflow::gtl::ArraySlice broadcast_dimensions); + // Enqueues a pad operation onto the computation that pads the given value on // the edges as well as between the elements of the input. padding_config // specifies the padding amount for each dimension. @@ -350,26 +487,6 @@ class XlaBuilder { XlaOp DotGeneral(const XlaOp& lhs, const XlaOp& rhs, const DotDimensionNumbers& dimension_numbers); - // Default dimension numbers used for a 2D convolution. - static constexpr int64 kConvBatchDimension = 0; - static constexpr int64 kConvFeatureDimension = 1; - static constexpr int64 kConvFirstSpatialDimension = 2; - static constexpr int64 kConvSecondSpatialDimension = 3; - static constexpr int64 kConvKernelOutputDimension = 0; - static constexpr int64 kConvKernelInputDimension = 1; - static constexpr int64 kConvKernelFirstSpatialDimension = 2; - static constexpr int64 kConvKernelSecondSpatialDimension = 3; - - // Creates a default ConvolutionDimensionNumbers. For a 2D convolution, for - // the input operand {batch, feature, height, width} = {0, 1, 2, 3} and for - // the kernel operand - // {output_feature, input_feature, height, width} = {0, 1, 2, 3}. - static ConvolutionDimensionNumbers CreateDefaultConvDimensionNumbers( - int num_spatial_dims = 2); - - // Returns an error if the convolution dimension numbers have conflicts. - static Status Validate(const ConvolutionDimensionNumbers& dnum); - // Enqueues a convolution instruction onto the computation, which uses the // default convolution dimension numbers. XlaOp Conv(const XlaOp& lhs, const XlaOp& rhs, @@ -416,6 +533,8 @@ class XlaBuilder { // Enqueues an infeed instruction onto the computation, which writes data of // the given shape to the infeed buffer of the device. XlaOp Infeed(const Shape& shape, const string& config = ""); + XlaOp InfeedWithToken(const XlaOp& token, const Shape& shape, + const string& config = ""); // Enqueues an outfeed instruction onto the computation. This instruction // generates outgoing data transfers for the given data. @@ -425,6 +544,9 @@ class XlaBuilder { // will occur. void Outfeed(const XlaOp& operand, const Shape& shape_with_layout, const string& outfeed_config); + XlaOp OutfeedWithToken(const XlaOp& operand, const XlaOp& token, + const Shape& shape_with_layout, + const string& outfeed_config); // Enqueues a call instruction onto the computation. XlaOp Call(const XlaComputation& computation, @@ -635,16 +757,6 @@ class XlaBuilder { // Enqueues an imaginary-part instruction onto the computation. XlaOp Imag(const XlaOp& operand); - // Enqueues a float32 sqrt instruction onto the computation. - // (float32 is specified as there is an implicit float32 0.5f constant - // exponent). - XlaOp SqrtF32(const XlaOp& operand); - - // Enqueues a float32 square instruction onto the computation. - // (float32 is specified as there is an implicit float32 2.0f constant - // exponent). - XlaOp SquareF32(const XlaOp& operand); - // Enqueues a lhs^rhs computation onto the computation. XlaOp Pow(const XlaOp& lhs, const XlaOp& rhs, tensorflow::gtl::ArraySlice broadcast_dimensions = {}); @@ -667,14 +779,6 @@ class XlaBuilder { XlaOp BitcastConvertType(const XlaOp& operand, PrimitiveType new_element_type); - // Enqueues a float32 reciprocal instruction onto the computation. - // (float32 is specified as there is an implicit float32 -1.0f constant - // exponent). - // - // TODO(b/34468990) axe F32 suffix, can be determined by reflecting on the - // shape of the operand. - XlaOp ReciprocalF32(const XlaOp& operand); - // Enqueues a negate instruction onto the computation. XlaOp Neg(const XlaOp& operand); @@ -689,7 +793,24 @@ class XlaBuilder { tensorflow::gtl::ArraySlice dimensions); // Enqueues a sort (as increasing order) instruction onto the computation. - XlaOp Sort(const XlaOp& operand); + // If only keys are provided: + // * If the keys are an rank-1 tensor (an array), the result is a sorted array + // of keys, in ascending order. + // * If the keys have higher rank, the keys are sorted along the provided + // dimension. For example, for a rank-2 tensor (a matrix) of keys, a dimension + // value of 0 will indepenently sort every column, and a dimension value of 1 + // will independently sort each row. If no dimension number is provided, then + // the last dimension is chosen by default. + // + // If both keys and values are provided: + // * The keys and the values must tensors with the same dimensions. The + // element types of the tensors may be different. + // * The result is a tuple that consists of a sorted tensor of keys (along the + // provided dimension, as above) as the first element, and a tensor with their + // corresponding values as the second element. + XlaOp Sort(XlaOp keys, + tensorflow::gtl::optional values = tensorflow::gtl::nullopt, + int64 dimension = -1); // Enqueues a clamp instruction onto the computation. XlaOp Clamp(const XlaOp& min, const XlaOp& operand, const XlaOp& max); @@ -730,19 +851,23 @@ class XlaBuilder { // Enqueues a Send node onto the computation, to send the given operand to // a Recv instruction that shares the same channel handle. void Send(const XlaOp& operand, const ChannelHandle& handle); + XlaOp SendWithToken(const XlaOp& operand, const XlaOp& token, + const ChannelHandle& handle); + + // Enqueues an AfterAll operation with no operands producing a token-shaped + // value. + XlaOp CreateToken(); + + // Enqueues an AfterAll operation with no operands producing a token-shaped + // value. + XlaOp AfterAll(tensorflow::gtl::ArraySlice tokens); // Enqueues a Recv node onto the computation. The data comes from a Send // instruction that shares the same channel handle and its shape must // be the same as the given shape. XlaOp Recv(const Shape& shape, const ChannelHandle& handle); - - // Returns true if 'operand' is a compile-time constant. A compile-time - // constant does not depend on any parameters, or on stateful operators such - // as `RngNormal` or `Infeed`. - // - // This tests whether a computation is a compile-time constant without - // evaluating the computation. - StatusOr IsConstant(const XlaOp& operand) const; + XlaOp RecvWithToken(const XlaOp& token, const Shape& shape, + const ChannelHandle& handle); // Normalizes operand across spatial and batch dimensions for each feature. // @@ -782,47 +907,6 @@ class XlaBuilder { const XlaOp& grad_output, float epsilon, int64 feature_index); - // Returns a new XlaBuilder whose resultant Computation is used only by this - // XlaBuilder. The sub-XlaBuilder has the same die_immediately_on_error - // behavior as the parent. - std::unique_ptr CreateSubBuilder(const string& computation_name); - - // Builds the computation with the requested operations, or returns a non-ok - // status. Note that all ops that have been enqueued will be moved to the - // computation being returned. - StatusOr Build(); - - // Builds the computation with the requested operations, or notes an error in - // the parent XlaBuilder and returns an empty computation if building failed. - // This function is intended to be used where the returned XlaComputation is - // only used by the parent XlaBuilder and hence further operation on the - // returned XlaComputation will simply be error'ed out if an error occurred - // while building this computation. If the built computation is to be used by - // a XlaBuilder other than the parent XlaBuilder then Build() should be used - // instead. - XlaComputation BuildAndNoteError(); - - // Returns a subgraph that roots on the given root. If the root is not a - // compile-time constant (see `IsConstant`), returns an error. - // - // This will copy the needed ops/computations to the subgraph. - StatusOr BuildConstantSubGraph(const XlaOp& root_op) const; - - // Returns the first error that was encountered while building the - // computation. When an error is encountered, by default we return a vacuous - // XlaOp and inform the user of the error that occurred while - // building the computation when they make a final call to Build(). - // - // See also set_die_immediately_on_error(). - Status first_error() const { return first_error_; } - - // Returns the shape of the given op. - StatusOr GetShape(const XlaOp& op) const; - - // Returns the (inferred) result for the current computation's shape. - StatusOr GetProgramShape() const; - - private: StatusOr AddInstruction( HloInstructionProto&& instr, HloOpcode opcode, tensorflow::gtl::ArraySlice operands = {}); @@ -830,14 +914,6 @@ class XlaBuilder { void AddCalledComputation(const XlaComputation& computation, HloInstructionProto* instr); - // Notes that the error occurred by: - // * storing it internally and capturing a backtrace if it's the first error - // (this deferred value will be produced on the call to Build()) - // * dying if die_immediately_on_error_ is true - void NoteError(const Status& error); - - XlaOp NoteErrorOrReturn(const std::function()>& op_creator); - StatusOr LookUpInstruction(const XlaOp& op) const; // Internal helper method that does the building for an arbitrary unary op. @@ -933,16 +1009,1032 @@ class XlaBuilder { bool die_immediately_on_error_ = false; XlaBuilder* parent_builder_{nullptr}; + + friend XlaOp Parameter(XlaBuilder* builder, int64 parameter_number, + const Shape& shape, const string& name); + friend XlaOp ConstantLiteral(XlaBuilder* builder, + const LiteralSlice& literal); + template + friend XlaOp ConstantR0(XlaBuilder* builder, NativeT value); + template + friend XlaOp ConstantR1(XlaBuilder* builder, + tensorflow::gtl::ArraySlice values); + friend XlaOp ConstantR1(XlaBuilder* builder, + const tensorflow::core::Bitmap& values); + template + friend XlaOp ConstantR2( + XlaBuilder* builder, + std::initializer_list> values); + template + friend XlaOp ConstantFromArrayWithLayout(XlaBuilder* builder, + const Array& values, + const Layout& layout); + template + friend XlaOp ConstantFromArray(XlaBuilder* builder, + const Array& values); + template + friend XlaOp ConstantR2FromArray2DWithLayout(XlaBuilder* builder, + const Array2D& values, + const Layout& layout); + template + friend XlaOp ConstantR2FromArray2D(XlaBuilder* builder, + const Array2D& values); + template + friend XlaOp ConstantR3FromArray3DWithLayout(XlaBuilder* builder, + const Array3D& values, + const Layout& layout); + template + friend XlaOp ConstantR3FromArray3D(XlaBuilder* builder, + const Array3D& values); + template + friend XlaOp ConstantR4FromArray4DWithLayout(XlaBuilder* builder, + const Array4D& values, + const Layout& layout); + template + friend XlaOp ConstantR4FromArray4D(XlaBuilder* builder, + const Array4D& values); + + template + friend XlaOp ConstantR1(XlaBuilder* builder, int64 length, NativeT value); + + friend XlaOp Broadcast(const XlaOp& operand, + tensorflow::gtl::ArraySlice broadcast_sizes); + + friend XlaOp BroadcastInDim( + const XlaOp& operand, const Shape& shape, + const tensorflow::gtl::ArraySlice broadcast_dimensions); + + friend XlaOp Pad(const XlaOp& operand, const XlaOp& padding_value, + const PaddingConfig& padding_config); + + friend XlaOp Reshape(const XlaOp& operand, + tensorflow::gtl::ArraySlice dimensions, + tensorflow::gtl::ArraySlice new_sizes); + + friend XlaOp Reshape(const XlaOp& operand, + tensorflow::gtl::ArraySlice new_sizes); + + friend XlaOp Collapse(const XlaOp& operand, + tensorflow::gtl::ArraySlice dimensions); + + friend XlaOp Slice(const XlaOp& operand, + tensorflow::gtl::ArraySlice start_indices, + tensorflow::gtl::ArraySlice limit_indices, + tensorflow::gtl::ArraySlice strides); + + friend XlaOp SliceInDim(const XlaOp& operand, int64 start_index, + int64 limit_index, int64 stride, int64 dimno); + + friend XlaOp DynamicSlice(const XlaOp& operand, const XlaOp& start_indices, + tensorflow::gtl::ArraySlice slice_sizes); + + friend XlaOp DynamicUpdateSlice(const XlaOp& operand, const XlaOp& update, + const XlaOp& start_indices); + + friend XlaOp ConcatInDim(XlaBuilder* builder, + tensorflow::gtl::ArraySlice operands, + int64 dimension); + + friend void Trace(const string& tag, const XlaOp& operand); + + friend XlaOp Select(const XlaOp& pred, const XlaOp& on_true, + const XlaOp& on_false); + friend XlaOp Tuple(XlaBuilder* builder, + tensorflow::gtl::ArraySlice elements); + friend XlaOp GetTupleElement(const XlaOp& tuple_data, int64 index); + friend XlaOp Eq(const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice broadcast_dimensions); + friend XlaOp Ne(const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice broadcast_dimensions); + friend XlaOp Ge(const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice broadcast_dimensions); + friend XlaOp Gt(const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice broadcast_dimensions); + friend XlaOp Lt(const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice broadcast_dimensions); + friend XlaOp Le(const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice broadcast_dimensions); + friend XlaOp Dot(const XlaOp& lhs, const XlaOp& rhs); + friend XlaOp DotGeneral(const XlaOp& lhs, const XlaOp& rhs, + const DotDimensionNumbers& dimension_numbers); + friend XlaOp Conv(const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice window_strides, + Padding padding); + friend XlaOp ConvWithGeneralPadding( + const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice window_strides, + tensorflow::gtl::ArraySlice> padding); + friend XlaOp ConvWithGeneralDimensions( + const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice window_strides, Padding padding, + const ConvolutionDimensionNumbers& dimension_numbers); + friend XlaOp ConvGeneral( + const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice window_strides, + tensorflow::gtl::ArraySlice> padding, + const ConvolutionDimensionNumbers& dimension_numbers); + friend XlaOp ConvGeneralDilated( + const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice window_strides, + tensorflow::gtl::ArraySlice> padding, + tensorflow::gtl::ArraySlice lhs_dilation, + tensorflow::gtl::ArraySlice rhs_dilation, + const ConvolutionDimensionNumbers& dimension_numbers); + friend XlaOp Fft(const XlaOp& operand, FftType fft_type, + tensorflow::gtl::ArraySlice fft_length); + friend XlaOp Infeed(XlaBuilder* builder, const Shape& shape, + const string& config); + friend void Outfeed(const XlaOp& operand, const Shape& shape_with_layout, + const string& outfeed_config); + friend XlaOp Call(XlaBuilder* builder, const XlaComputation& computation, + tensorflow::gtl::ArraySlice operands); + friend XlaOp CustomCall(XlaBuilder* builder, const string& call_target_name, + tensorflow::gtl::ArraySlice operands, + const Shape& shape); + friend XlaOp HostCompute(XlaBuilder* builder, + tensorflow::gtl::ArraySlice operands, + const string& channel_name, int64 cost_estimate_ns, + const Shape& shape); + friend XlaOp Complex(const XlaOp& real, const XlaOp& imag, + tensorflow::gtl::ArraySlice broadcast_dimensions); + friend XlaOp Conj(const XlaOp& operand); + friend XlaOp Add(const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice broadcast_dimensions); + friend XlaOp Sub(const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice broadcast_dimensions); + friend XlaOp Mul(const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice broadcast_dimensions); + friend XlaOp Div(const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice broadcast_dimensions); + friend XlaOp Rem(const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice broadcast_dimensions); + friend XlaOp Max(const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice broadcast_dimensions); + friend XlaOp Min(const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice broadcast_dimensions); + friend XlaOp And(const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice broadcast_dimensions); + friend XlaOp Or(const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice broadcast_dimensions); + friend XlaOp Xor(const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice broadcast_dimensions); + friend XlaOp Not(const XlaOp& operand); + friend XlaOp ShiftLeft( + const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice broadcast_dimensions); + friend XlaOp ShiftRightArithmetic( + const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice broadcast_dimensions); + friend XlaOp ShiftRightLogical( + const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice broadcast_dimensions); + friend XlaOp Reduce(const XlaOp& operand, const XlaOp& init_value, + const XlaComputation& computation, + tensorflow::gtl::ArraySlice dimensions_to_reduce); + friend XlaOp ReduceAll(const XlaOp& operand, const XlaOp& init_value, + const XlaComputation& computation); + friend XlaOp ReduceWindow( + const XlaOp& operand, const XlaOp& init_value, + const XlaComputation& computation, + tensorflow::gtl::ArraySlice window_dimensions, + tensorflow::gtl::ArraySlice window_strides, Padding padding); + friend XlaOp ReduceWindowWithGeneralPadding( + const XlaOp& operand, const XlaOp& init_value, + const XlaComputation& computation, + tensorflow::gtl::ArraySlice window_dimensions, + tensorflow::gtl::ArraySlice window_strides, + tensorflow::gtl::ArraySlice> padding); + friend XlaOp CrossReplicaSum( + const XlaOp& operand, + tensorflow::gtl::ArraySlice replica_group_ids); + friend XlaOp CrossReplicaSum( + const XlaOp& operand, const XlaComputation& computation, + tensorflow::gtl::ArraySlice replica_group_ids, + const tensorflow::gtl::optional& channel_id); + friend XlaOp SelectAndScatter( + const XlaOp& operand, const XlaComputation& select, + tensorflow::gtl::ArraySlice window_dimensions, + tensorflow::gtl::ArraySlice window_strides, Padding padding, + const XlaOp& source, const XlaOp& init_value, + const XlaComputation& scatter); + friend XlaOp SelectAndScatterWithGeneralPadding( + const XlaOp& operand, const XlaComputation& select, + tensorflow::gtl::ArraySlice window_dimensions, + tensorflow::gtl::ArraySlice window_strides, + tensorflow::gtl::ArraySlice> padding, + const XlaOp& source, const XlaOp& init_value, + const XlaComputation& scatter); + friend XlaOp Abs(const XlaOp& operand); + friend XlaOp Atan2(const XlaOp& y, const XlaOp& x, + tensorflow::gtl::ArraySlice broadcast_dimensions); + friend XlaOp Exp(const XlaOp& operand); + friend XlaOp Expm1(const XlaOp& operand); + friend XlaOp Floor(const XlaOp& operand); + friend XlaOp Ceil(const XlaOp& operand); + friend XlaOp Round(const XlaOp& operand); + friend XlaOp Log(const XlaOp& operand); + friend XlaOp Log1p(const XlaOp& operand); + friend XlaOp Sign(const XlaOp& operand); + friend XlaOp Clz(const XlaOp& operand); + friend XlaOp Cos(const XlaOp& operand); + friend XlaOp Sin(const XlaOp& operand); + friend XlaOp Tanh(const XlaOp& operand); + friend XlaOp Real(const XlaOp& operand); + friend XlaOp Imag(const XlaOp& operand); + friend XlaOp Pow(const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice broadcast_dimensions); + friend XlaOp IsFinite(const XlaOp& operand); + friend XlaOp ConvertElementType(const XlaOp& operand, + PrimitiveType new_element_type); + friend XlaOp BitcastConvertType(const XlaOp& operand, + PrimitiveType new_element_type); + friend XlaOp Neg(const XlaOp& operand); + friend XlaOp Transpose(const XlaOp& operand, + tensorflow::gtl::ArraySlice permutation); + friend XlaOp Rev(const XlaOp& operand, + tensorflow::gtl::ArraySlice dimensions); + friend XlaOp Sort(XlaOp keys, tensorflow::gtl::optional values, + int64 dimension); + friend XlaOp Clamp(const XlaOp& min, const XlaOp& operand, const XlaOp& max); + friend XlaOp Map(XlaBuilder* builder, + tensorflow::gtl::ArraySlice operands, + const XlaComputation& computation, + tensorflow::gtl::ArraySlice dimensions, + tensorflow::gtl::ArraySlice static_operands); + friend XlaOp RngNormal(const XlaOp& mu, const XlaOp& sigma, + const Shape& shape); + friend XlaOp RngUniform(const XlaOp& a, const XlaOp& b, const Shape& shape); + friend XlaOp While(const XlaComputation& condition, + const XlaComputation& body, const XlaOp& init); + friend XlaOp Conditional(const XlaOp& predicate, const XlaOp& true_operand, + const XlaComputation& true_computation, + const XlaOp& false_operand, + const XlaComputation& false_computation); + friend XlaOp ReducePrecision(const XlaOp& operand, const int exponent_bits, + const int mantissa_bits); + friend XlaOp Gather(const XlaOp& input, const XlaOp& gather_indices, + const GatherDimensionNumbers& dimension_numbers, + tensorflow::gtl::ArraySlice window_bounds); + friend void Send(const XlaOp& operand, const ChannelHandle& handle); + friend XlaOp Recv(XlaBuilder* builder, const Shape& shape, + const ChannelHandle& handle); + friend XlaOp BatchNormTraining(const XlaOp& operand, const XlaOp& scale, + const XlaOp& offset, float epsilon, + int64 feature_index); + friend XlaOp BatchNormInference(const XlaOp& operand, const XlaOp& scale, + const XlaOp& offset, const XlaOp& mean, + const XlaOp& variance, float epsilon, + int64 feature_index); + friend XlaOp BatchNormGrad(const XlaOp& operand, const XlaOp& scale, + const XlaOp& batch_mean, const XlaOp& batch_var, + const XlaOp& grad_output, float epsilon, + int64 feature_index); + friend XlaOp SendWithToken(const XlaOp& operand, const XlaOp& token, + const ChannelHandle& handle); + friend XlaOp RecvWithToken(const XlaOp& token, const Shape& shape, + const ChannelHandle& handle); + friend XlaOp InfeedWithToken(const XlaOp& token, const Shape& shape, + const string& config); + friend XlaOp OutfeedWithToken(const XlaOp& operand, const XlaOp& token, + const Shape& shape_with_layout, + const string& outfeed_config); + friend XlaOp CreateToken(XlaBuilder* builder); + friend XlaOp AfterAll(XlaBuilder* builder, + tensorflow::gtl::ArraySlice tokens); +}; + +// RAII-style object: sets the current sharding assignment in builder on +// construction, and sets back to the previous assignment on destruction. +class XlaScopedShardingAssignment { + public: + XlaScopedShardingAssignment(xla::XlaBuilder* builder, + tensorflow::gtl::optional sharding) + : builder_(builder), prev_sharding_(builder->sharding()) { + SetSharding(sharding); + } + + XlaScopedShardingAssignment(const XlaScopedShardingAssignment&) = delete; + XlaScopedShardingAssignment& operator=(const XlaScopedShardingAssignment&) = + delete; + + ~XlaScopedShardingAssignment() { SetSharding(prev_sharding_); } + + private: + void SetSharding(const tensorflow::gtl::optional& sharding) { + if (sharding.has_value()) { + builder_->SetSharding(sharding.value()); + } else { + builder_->ClearSharding(); + } + } + + xla::XlaBuilder* const builder_; + tensorflow::gtl::optional prev_sharding_; }; +// Free functions for building XlaOps. The intention is that these will +// become the public API for building XlaOps rather than calling methods on +// XlaBuilder directly. + +// Enqueues a "retrieve parameter value" instruction for a parameter that was +// passed to the computation. +XlaOp Parameter(XlaBuilder* builder, int64 parameter_number, const Shape& shape, + const string& name); + +// Enqueues a constant with the value of the given literal onto the +// computation. +XlaOp ConstantLiteral(XlaBuilder* builder, const LiteralSlice& literal); + +// Enqueues a constant onto the computation. Methods are templated on the +// native host type (NativeT) which corresponds to a specific XLA +// PrimitiveType as given in the following table: +// +// Native Type PrimitiveType +// ----------------------------- +// bool PRED +// int32 S32 +// int64 S64 +// uint32 U32 +// uint64 U64 +// float F32 +// double F64 +// +// Note: not all primitive types defined in xla_data.proto have a +// corresponding native type yet. +template +XlaOp ConstantR0(XlaBuilder* builder, NativeT value); +template +XlaOp ConstantR1(XlaBuilder* builder, + tensorflow::gtl::ArraySlice values); +XlaOp ConstantR1(XlaBuilder* builder, const tensorflow::core::Bitmap& values); +template +XlaOp ConstantR2(XlaBuilder* builder, + std::initializer_list> values); +template +XlaOp ConstantFromArrayWithLayout(XlaBuilder* builder, + const Array& values, + const Layout& layout); +template +XlaOp ConstantFromArray(XlaBuilder* builder, const Array& values); +template +XlaOp ConstantR2FromArray2DWithLayout(XlaBuilder* builder, + const Array2D& values, + const Layout& layout); +template +XlaOp ConstantR2FromArray2D(XlaBuilder* builder, + const Array2D& values); +template +XlaOp ConstantR3FromArray3DWithLayout(XlaBuilder* builder, + const Array3D& values, + const Layout& layout); +template +XlaOp ConstantR3FromArray3D(XlaBuilder* builder, + const Array3D& values); +template +XlaOp ConstantR4FromArray4DWithLayout(XlaBuilder* builder, + const Array4D& values, + const Layout& layout); +template +XlaOp ConstantR4FromArray4D(XlaBuilder* builder, + const Array4D& values); + +// Enqueues a rank one constant (XlaBuilder* builder, vector) onto the +// computation. The vector has size 'length' and every element has the value +// 'value'. +template +XlaOp ConstantR1(XlaBuilder* builder, int64 length, NativeT value); + +// Adds dimensions to an array by duplicating the data in the array. +// +// The new dimensions are inserted on the left, i.e. if +// broadcast_sizes has values {a0, ..., aN} and the operand shape +// has dimensions {b0, ..., bM} then the shape of the output has +// dimensions {a0, ..., aN, b0, ..., bM}. +// +// The new dimensions index into copies of the operand, i.e. +// +// output[i0, ..., iN, j0, ..., jM] = operand[j0, ..., jM] +XlaOp Broadcast(const XlaOp& operand, + tensorflow::gtl::ArraySlice broadcast_sizes); + +// Performs in-dimension-style broadcast. +// +// Operand specifies the input to be broadcast. "shape" is expected output +// shape. "broadcast_dimensions" are the dimensions to be broadcasting into. +// Dimension numbers in broadcast_dimensions map to individual dimensions +// of the operand, and specify what dimension of the output shape they +// should be broadcast. +// e.g. +// Say operand = [1, 2], i.e., a 1D tensor with 2 elements. +// and dimension of shape is [2,2]. +// Specifying {1} as brodcast_dimension will generate output +// [1 , 2] +// [1 , 2] +// On the other hand, specifying {0} as broadcast_dimension +// will generate output +// [1 , 1] +// [2 , 2] +XlaOp BroadcastInDim( + const XlaOp& operand, const Shape& shape, + const tensorflow::gtl::ArraySlice broadcast_dimensions); + +// Enqueues a pad operation onto the computation that pads the given value on +// the edges as well as between the elements of the input. padding_config +// specifies the padding amount for each dimension. +XlaOp Pad(const XlaOp& operand, const XlaOp& padding_value, + const PaddingConfig& padding_config); + +// Enqueues an operation onto the computation that flattens the operand based +// on the dimension order (major/slowest-varying to minor/fastest-varying) +// given, followed by reshaping it into the shape with the given dimension +// sizes (also major to minor). Conceptually, this is a limited form of +// "shape casting". +XlaOp Reshape(const XlaOp& operand, + tensorflow::gtl::ArraySlice dimensions, + tensorflow::gtl::ArraySlice new_sizes); + +// Enqueues an operation onto the computation that collapses the operand, from +// first to last dimension (C order), then reshapes it to the given dimension +// sizes. Conceptually, this is a limited form of "shape casting". +XlaOp Reshape(const XlaOp& operand, + tensorflow::gtl::ArraySlice new_sizes); + +// Wrapper for Reshape. +// Enqueues an operation to collapse the provided dimensions; e.g. an +// operand with dimensions {x=256, y=2, z=2, p=32} can be collapsed to +// {x=1024, y=32} by collapsing dims {0, 1, 2}. Collapsing dimensions must +// be a consecutive, in-order subsequence of the operand dimensions. +// +// Note that collapsing a single dimension does nothing: +// +// {256} collapsing {0} => {256} +// {1} collapsing {0} => {1} +// +// Collapsing multiple dimensions produces a single result dimension: +// +// {256, 2} collapsing {0,1} => {512} +// {256, 2, 3} collapsing {0,1} => {512, 3} +// +// This could potentially cause data to be moved -- it provides a more +// structured form of reshaping than an arbitrary Reshape operation. +XlaOp Collapse(const XlaOp& operand, + tensorflow::gtl::ArraySlice dimensions); + +// Enqueues a slice operation onto the computation that slices the operand +// from the start indices to the limit indices; e.g. +// +// x +// [ 0 1 2 3 ] +// y [ 4 5 6 7 ] => slice(start={1, 1}, limit={2, 3}) => [ 5 6 ] +// [ 8 9 a b ] +// +// Note that "limit" means up-to-but-not-including; i.e. [start, limit) in 1D +// range notation. +// The strides parameter determines the stride over the slice +XlaOp Slice(const XlaOp& operand, + tensorflow::gtl::ArraySlice start_indices, + tensorflow::gtl::ArraySlice limit_indices, + tensorflow::gtl::ArraySlice strides); + +// Enqueues a slice operation in a given dimension, taking all other +// dimensions as they are; e.g. if dimno is 1 from start_index 2 to +// limit_index 4 by 1, and the shape is f32[7,8,9], this call is short-hand +// for: +// +// array[:, 2:4:1, :] +XlaOp SliceInDim(const XlaOp& operand, int64 start_index, int64 limit_index, + int64 stride, int64 dimno); + +// Enqueues a slice operation onto the computation that slices the 'operand' +// from dynamic start indices which are passed in 'start_indices'. +// The size of the slice in each dimension is passed in 'slice_sizes', +// which specify the end point of exclusive slice intervals in each +// dimension [start, start + size). +// The shape of 'start_indices' must be rank == 1, with dimension size +// equal to the rank of the 'operand'. +// Slice index calculations are computed modulo input dimension sizes to +// prevent dynamic start indices from generating out-of-bound array accesses. +XlaOp DynamicSlice(const XlaOp& operand, const XlaOp& start_indices, + tensorflow::gtl::ArraySlice slice_sizes); + +// Enqueues a dynamic update slice operation onto the computation, which +// updates a slice of 'operand' with 'update' at dynamic 'start_indices'. +// The shape of 'update' determines the shape of the slice of 'operand' +// which is updated. +// The indices specified in 'start_indices' specify the offset of the slice +// of 'operand' which is updated. +// +// update = {10, 11} // calculated at runtime. +// [1 2 3] start = {1, 1} // calculated at runtime. [1 2 3 ] +// [4 5 6] => DynamicUpdateslice(data, update, start) => [4 10 11] +// [7 8 9] [7 8 9 ] +// +// The shape of 'start_indices' must be rank == 1, with dimension size +// equal to the rank of the 'operand'. +// Slice index calculations are computed modulo update dimension sizes to +// prevent dynamic start indices from generating out-of-bound array accesses. +XlaOp DynamicUpdateSlice(const XlaOp& operand, const XlaOp& update, + const XlaOp& start_indices); + +// Enqueues a concatenate instruction onto the computation. 'operands' must +// have >= 1 entry. +XlaOp ConcatInDim(XlaBuilder* builder, + tensorflow::gtl::ArraySlice operands, int64 dimension); + +// Enqueue a tracing operation onto the computation; the computation will emit +// a logging message with the operand. +void Trace(const string& tag, const XlaOp& operand); + +// Enqueues a conditional-move-like select operation onto the computation; +// predicated on pred, selects between on_true and on_false. +XlaOp Select(const XlaOp& pred, const XlaOp& on_true, const XlaOp& on_false); + +// Enqueues a tuple-creation instruction onto the computation. +XlaOp Tuple(XlaBuilder* builder, tensorflow::gtl::ArraySlice elements); + +// Enqueues a tuple-element-get instruction onto the computation. +XlaOp GetTupleElement(const XlaOp& tuple_data, int64 index); + +// Enqueues an equal-to comparison instruction onto the computation. +XlaOp Eq(const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice broadcast_dimensions = {}); + +// Enqueues a not-equal comparison instruction onto the computation. +XlaOp Ne(const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice broadcast_dimensions = {}); + +// Enqueues a greater-or-equal comparison instruction onto the computation. +XlaOp Ge(const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice broadcast_dimensions = {}); + +// Enqueues a greater-than comparison instruction onto the computation. +XlaOp Gt(const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice broadcast_dimensions = {}); + +// Enqueues a less-than comparison instruction onto the computation. +XlaOp Lt(const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice broadcast_dimensions = {}); + +// Enqueues a less-or-equal comparison instruction onto the computation. +XlaOp Le(const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice broadcast_dimensions = {}); + +// Enqueues a dot instruction onto the computation. +XlaOp Dot(const XlaOp& lhs, const XlaOp& rhs); + +// Enqueues a general dot instruction onto the computation. +XlaOp DotGeneral(const XlaOp& lhs, const XlaOp& rhs, + const DotDimensionNumbers& dimension_numbers); + +// Enqueues a convolution instruction onto the computation, which uses the +// default convolution dimension numbers. +XlaOp Conv(const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice window_strides, Padding padding); + +// Enqueues a convolution instruction onto the computation, with the caller +// provided padding configuration in the format returned by MakePadding(). +XlaOp ConvWithGeneralPadding( + const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice window_strides, + tensorflow::gtl::ArraySlice> padding); + +// Enqueues a convolution instruction onto the computation, with the caller +// provided dimension numbers configuration. +XlaOp ConvWithGeneralDimensions( + const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice window_strides, Padding padding, + const ConvolutionDimensionNumbers& dimension_numbers); + +// Enqueues a convolution instruction onto the computation, with the caller +// provided padding configuration as well as the dimension numbers. +XlaOp ConvGeneral(const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice window_strides, + tensorflow::gtl::ArraySlice> padding, + const ConvolutionDimensionNumbers& dimension_numbers); + +// Enqueues a convolution instruction onto the computation, with the caller +// provided padding configuration, dilation factors and dimension numbers. +XlaOp ConvGeneralDilated( + const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice window_strides, + tensorflow::gtl::ArraySlice> padding, + tensorflow::gtl::ArraySlice lhs_dilation, + tensorflow::gtl::ArraySlice rhs_dilation, + const ConvolutionDimensionNumbers& dimension_numbers); + +// Enqueues an FFT instruction onto the computation, of the given type and +// with the given FFT length. +XlaOp Fft(const XlaOp& operand, FftType fft_type, + tensorflow::gtl::ArraySlice fft_length); + +// Enqueues an infeed instruction onto the computation, which writes data of +// the given shape to the infeed buffer of the device. +XlaOp Infeed(XlaBuilder* builder, const Shape& shape, + const string& config = ""); + +// Variant of Infeed which takes a token-shaped operand and produces a +// two-element tuple containing the data value and a token-shaped value. +// Tokens are used for ordering side-effecting operations. +// TODO(b/110532604): Replace all uses of the non-token form with this variant. +XlaOp InfeedWithToken(const XlaOp& token, const Shape& shape, + const string& config = ""); + +// Enqueues an outfeed instruction onto the computation. This instruction +// generates outgoing data transfers for the given data. +// +// shape_with_layout communicates the laid out shape that we want to outfeed +// -- if !ShapeUtil::Compatible(GetShape(operand), shape_with_layout) an error +// will occur. +void Outfeed(const XlaOp& operand, const Shape& shape_with_layout, + const string& outfeed_config); + +// Variant of Outfeed which takes a token-shaped operand and produces a +// token-shaped value. Tokens are used for ordering side-effecting operations. +// TODO(b/110532604): Replace all uses of the non-token form with this variant. +XlaOp OutfeedWithToken(const XlaOp& operand, const XlaOp& token, + const Shape& shape_with_layout, + const string& outfeed_config); + +// Enqueues a call instruction onto the computation. +XlaOp Call(XlaBuilder* builder, const XlaComputation& computation, + tensorflow::gtl::ArraySlice operands); + +// Enqueues a custom call instruction onto the computation. +// During code generation, a call instruction is emitted which targets a +// symbol with the name |call_target_name|. The |operands| are passed to the +// call instruction. |shape| is the resultant shape. +XlaOp CustomCall(XlaBuilder* builder, const string& call_target_name, + tensorflow::gtl::ArraySlice operands, + const Shape& shape); + +// Enqueues a pseudo-op to represent host-side computation data-dependencies. +// During code generation, host send and receive operations will be generated +// to transfer |operands| to the host and a single result of |shape| back to +// the device. Host send/recv operations are emitted using |channel_name|. +// Dataflow dependencies and the |cost_estimate_ns| field may be used in HLO +// instruction scheduling. +XlaOp HostCompute(XlaBuilder* builder, + tensorflow::gtl::ArraySlice operands, + const string& channel_name, int64 cost_estimate_ns, + const Shape& shape); + +// The following methods enqueue element-wise binary arithmetic operations +// onto the computation. The shapes of the operands have to match unless one +// of the operands is a scalar, or an explicit broadcast dimension is given +// (see g3doc for more details). + +// Enqueues a complex compose instruction onto the computation. +XlaOp Complex(const XlaOp& real, const XlaOp& imag, + tensorflow::gtl::ArraySlice broadcast_dimensions = {}); + +// Enqueues a complex conjugate instruction onto the computation. +XlaOp Conj(const XlaOp& operand); + +// Enqueues an add instruction onto the computation. +XlaOp Add(const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice broadcast_dimensions = {}); + +// Enqueues a subtract instruction onto the computation. +XlaOp Sub(const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice broadcast_dimensions = {}); + +// Enqueues a multiply instruction onto the computation. +XlaOp Mul(const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice broadcast_dimensions = {}); + +// Enqueues a divide instruction onto the computation. +XlaOp Div(const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice broadcast_dimensions = {}); + +// Enqueues a remainder instruction onto the computation. +XlaOp Rem(const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice broadcast_dimensions = {}); + +// Enqueues a max instruction onto the computation. +XlaOp Max(const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice broadcast_dimensions = {}); + +// Enqueues a min instruction onto the computation. +XlaOp Min(const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice broadcast_dimensions = {}); + +// Element-wise logical operators +XlaOp And(const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice broadcast_dimensions = {}); + +XlaOp Or(const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice broadcast_dimensions = {}); + +XlaOp Xor(const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice broadcast_dimensions = {}); + +XlaOp Not(const XlaOp& operand); + +XlaOp ShiftLeft(const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice broadcast_dimensions = {}); +XlaOp ShiftRightArithmetic( + const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice broadcast_dimensions = {}); +XlaOp ShiftRightLogical( + const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice broadcast_dimensions = {}); + +// Reduces an array among the provided dimensions, given "computation" as a +// reduction operator. +XlaOp Reduce(const XlaOp& operand, const XlaOp& init_value, + const XlaComputation& computation, + tensorflow::gtl::ArraySlice dimensions_to_reduce); + +// Convenience wrapper around the above that reduces all the dimensions in the +// operand shape. +XlaOp ReduceAll(const XlaOp& operand, const XlaOp& init_value, + const XlaComputation& computation); + +// Enqueues a windowed reduce instruction onto the computation. +XlaOp ReduceWindow(const XlaOp& operand, const XlaOp& init_value, + const XlaComputation& computation, + tensorflow::gtl::ArraySlice window_dimensions, + tensorflow::gtl::ArraySlice window_strides, + Padding padding); + +// As ReduceWindow(), but the padding is given in the format +// returned by MakePadding(). +XlaOp ReduceWindowWithGeneralPadding( + const XlaOp& operand, const XlaOp& init_value, + const XlaComputation& computation, + tensorflow::gtl::ArraySlice window_dimensions, + tensorflow::gtl::ArraySlice window_strides, + tensorflow::gtl::ArraySlice> padding); + +// Returns the sum of the operand value within each subgroup of replicas. All +// replicas supply one input to the sum and all replicas receive the resulting +// sum for each subgroup. +XlaOp CrossReplicaSum( + const XlaOp& operand, + tensorflow::gtl::ArraySlice replica_group_ids = {}); + +// Enqueues an operation that do an AllReduce of the operand cross cores. Here +// AllReduce means doing a reduction on the input operand cross cores and then +// broadcasting the reduction result to those cores. The reduction function is +// defined by `computation`, which should be a commutative computation on +// scalars, e.g., add, min, or max. The way that AllReduce is applied is +// configured by: +// +// - `replica_group_ids`: maps replica ids to subgroup ids. If empty, all +// replicas belong to one group. Allreduce will be applied within subgroups. +// For example, we have 4 replicas, then replica_group_ids={0,1,0,1} means, +// replica 0 and 2 are in subgroup 0, replica 1 and 3 are in subgroup 1. +// +// - `channel_id`: for Allreduce nodes from different models, if they have the +// same channel_id, they will be 'Allreduce'd. If empty, Allreduce will not be +// applied cross models. +// +// TODO(b/79737069): Rename this to AllReduce when it's ready to use. +XlaOp CrossReplicaSum(const XlaOp& operand, const XlaComputation& computation, + tensorflow::gtl::ArraySlice replica_group_ids = {}, + const tensorflow::gtl::optional& + channel_id = tensorflow::gtl::nullopt); + +// Enqueues an operation that scatters the `source` array to the selected +// indices of each window. +XlaOp SelectAndScatter(const XlaOp& operand, const XlaComputation& select, + tensorflow::gtl::ArraySlice window_dimensions, + tensorflow::gtl::ArraySlice window_strides, + Padding padding, const XlaOp& source, + const XlaOp& init_value, const XlaComputation& scatter); + +// As SelectAndScatter(), but the padding is given in the format +// returned by MakePadding(). +XlaOp SelectAndScatterWithGeneralPadding( + const XlaOp& operand, const XlaComputation& select, + tensorflow::gtl::ArraySlice window_dimensions, + tensorflow::gtl::ArraySlice window_strides, + tensorflow::gtl::ArraySlice> padding, + const XlaOp& source, const XlaOp& init_value, + const XlaComputation& scatter); + +// Enqueues an abs instruction onto the computation. +XlaOp Abs(const XlaOp& operand); + +// Enqueues a atan2 instruction onto the computation. +XlaOp Atan2(const XlaOp& y, const XlaOp& x, + tensorflow::gtl::ArraySlice broadcast_dimensions = {}); + +// Enqueues an exp instruction onto the computation. +XlaOp Exp(const XlaOp& operand); + +// Enqueues an expm1 instruction onto the computation. +XlaOp Expm1(const XlaOp& operand); + +// Enqueues a floor instruction onto the computation. +XlaOp Floor(const XlaOp& operand); + +// Enqueues a ceil instruction onto the computation. +XlaOp Ceil(const XlaOp& operand); + +// Enqueues a round instruction onto the computation, rounding to nearest even +// with half-way cases rounding away from zero. +XlaOp Round(const XlaOp& operand); + +// Enqueues an log instruction (natural logarithm) onto the computation. +XlaOp Log(const XlaOp& operand); + +// Enqueues an log1p instruction (log(x+1)) onto the computation. +XlaOp Log1p(const XlaOp& operand); + +// Enqueues a sign instruction onto the computation. +XlaOp Sign(const XlaOp& operand); + +// Enqueues a count leading zeros instruction onto the computation. +XlaOp Clz(const XlaOp& operand); + +// Enqueues a cosine instruction onto the computation. +XlaOp Cos(const XlaOp& operand); + +// Enqueues a sine instruction onto the computation. +XlaOp Sin(const XlaOp& operand); + +// Enqueues a tanh instruction onto the computation. +XlaOp Tanh(const XlaOp& operand); + +// Enqueues a real-part instruction onto the computation. +XlaOp Real(const XlaOp& operand); + +// Enqueues an imaginary-part instruction onto the computation. +XlaOp Imag(const XlaOp& operand); + +// Enqueues a lhs^rhs computation onto the computation. +XlaOp Pow(const XlaOp& lhs, const XlaOp& rhs, + tensorflow::gtl::ArraySlice broadcast_dimensions = {}); + +// Enqueues an operator that tests if the operand's values are finite, i.e., +// not Inf or NaN. Defined only for floating-point types. Returns an array of +// booleans with the same shape where entries are true iff the corresponding +// entry was NaN. +XlaOp IsFinite(const XlaOp& operand); + +// Enqueues a convert instruction onto the computation that changes the +// element type of the operand array to primitive_type. +XlaOp ConvertElementType(const XlaOp& operand, PrimitiveType new_element_type); + +// Enqueues a no-op instruction onto the computation that changes +// the element type of the operand array to primitive_type. The +// bit-widths of the source and destination element types must be +// identical. +XlaOp BitcastConvertType(const XlaOp& operand, PrimitiveType new_element_type); + +// Enqueues a negate instruction onto the computation. +XlaOp Neg(const XlaOp& operand); + +// Enqueues a transpose instruction onto the computation. +XlaOp Transpose(const XlaOp& operand, + tensorflow::gtl::ArraySlice permutation); + +// Enqueues a reverse instruction onto the computation. The order of the +// elements in the given dimensions is reversed (i.e., the element at index i +// is moved to index dimension_size - 1 - i). +XlaOp Rev(const XlaOp& operand, tensorflow::gtl::ArraySlice dimensions); + +// Enqueues a sort (as increasing order) instruction onto the computation. +// If only keys are provided: +// * If the keys are an rank-1 tensor (an array), the result is a sorted array +// of keys, in ascending order. +// * If the keys have higher rank, the keys are sorted along the provided +// dimension. For example, for a rank-2 tensor (a matrix) of keys, a dimension +// value of 0 will indepenently sort every column, and a dimension value of 1 +// will independently sort each row. If no dimension number is provided, then +// the last dimension is chosen by default. +// +// If both keys and values are provided: +// * The keys and the values must tensors with the same dimensions. The +// element types of the tensors may be different. +// * The result is a tuple that consists of a sorted tensor of keys (along the +// provided dimension, as above) as the first element, and a tensor with their +// corresponding values as the second element. +XlaOp Sort(XlaOp keys, + tensorflow::gtl::optional values = tensorflow::gtl::nullopt, + int64 dimension = -1); + +// Enqueues a clamp instruction onto the computation. +XlaOp Clamp(const XlaOp& min, const XlaOp& operand, const XlaOp& max); + +// Enqueues a map instruction onto the computation. +XlaOp Map(XlaBuilder* builder, tensorflow::gtl::ArraySlice operands, + const XlaComputation& computation, + tensorflow::gtl::ArraySlice dimensions, + tensorflow::gtl::ArraySlice static_operands = {}); + +// Enqueues a N(mu, sigma) random number generation instruction onto the +// computation. +XlaOp RngNormal(const XlaOp& mu, const XlaOp& sigma, const Shape& shape); + +// Enqueues a U(a, b) random number generation instruction onto the +// computation. Returns values in the semi-open interval [a, b). +XlaOp RngUniform(const XlaOp& a, const XlaOp& b, const Shape& shape); + +// Enqueues a while node onto the computation. +XlaOp While(const XlaComputation& condition, const XlaComputation& body, + const XlaOp& init); + +// Enqueues a conditional node onto the computation. +XlaOp Conditional(const XlaOp& predicate, const XlaOp& true_operand, + const XlaComputation& true_computation, + const XlaOp& false_operand, + const XlaComputation& false_computation); + +// Enqueues a ReducePrecision node onto the computation. +XlaOp ReducePrecision(const XlaOp& operand, const int exponent_bits, + const int mantissa_bits); + +// Enqueues a Gather node onto the computation. +XlaOp Gather(const XlaOp& input, const XlaOp& gather_indices, + const GatherDimensionNumbers& dimension_numbers, + tensorflow::gtl::ArraySlice window_bounds); + +// Enqueues a Send node onto the computation, to send the given operand to +// a Recv instruction that shares the same channel handle. +void Send(const XlaOp& operand, const ChannelHandle& handle); + +// Variant of Send which takes a token-shaped operand and produces a +// token-shaped value. Tokens are used for ordering side-effecting operations. +// TODO(b/110532604): Replace all uses of the non-token form with this variant. +XlaOp SendWithToken(const XlaOp& operand, const XlaOp& token, + const ChannelHandle& handle); + +// Enqueues a Recv node onto the computation. The data comes from a Send +// instruction that shares the same channel handle and its shape must +// be the same as the given shape. +XlaOp Recv(XlaBuilder* builder, const Shape& shape, + const ChannelHandle& handle); + +// Variant of Recv which takes a token-shaped operand and produces a two-element +// tuple containing the data value and a token-shaped value. Tokens are used +// for ordering side-effecting operations. +// TODO(b/110532604): Replace all uses of the non-token form with this variant. +XlaOp RecvWithToken(const XlaOp& token, const Shape& shape, + const ChannelHandle& handle); + +// Enqueues an operation (AfterAll) with no operands that produces a +// token-shaped value. Tokens are used for ordering side-effecting operations. +// This is a separate method from AfterAll to facility the removal of +// operand-less AfterAll instructions. +// TODO(b/110532604): Remove this function when all tokens are derived from a +// single token generated or passed into the entry computation. +XlaOp CreateToken(XlaBuilder* builder); + +// Enqueues an AfterAll instruction which produces a token-shaped value and +// takes a variadic number of token-shaped operands. The number of operands must +// be greater than zero. Used for joining tokens. +XlaOp AfterAll(XlaBuilder* builder, tensorflow::gtl::ArraySlice tokens); + +// Normalizes operand across spatial and batch dimensions for each feature. +// +// Returns a tuple (normalized, batch_mean, batch_var) where `normalized` +// is the normalized result and batch_mean and batch_var are the mean and +// variance, respectively, across batch for the operand. +XlaOp BatchNormTraining(const XlaOp& operand, const XlaOp& scale, + const XlaOp& offset, float epsilon, + int64 feature_index); + +// Normalizes operand across spatial and batch dimensions for each feature. +// +// `BatchNormInference` is equivalent to calling `BatchNormTraining` without +// computing `mean` and `variance` for each batch inside the operation. It +// uses the input `mean` and `variance` instead as estimated values. The +// purpose of this op is to reduce latency in inference, hence the name +// `BatchNormInference`. +// +// The output has the same shape as `operand`, and contains the normalized +// values for each batch. +XlaOp BatchNormInference(const XlaOp& operand, const XlaOp& scale, + const XlaOp& offset, const XlaOp& mean, + const XlaOp& variance, float epsilon, + int64 feature_index); + +// Calculates the gradients of a batch norm op. +// +// The inputs `batch_mean` and `batch_var` represent the mean and variance +// across the batch. +// +// Returns a tuple of three elements: +// - grad_operand: Gradient with respect to input `operand` +// - grad_offset: Gradient with respect to input `offset` +// - grad_scale: Gradient with respect to input `scale` +XlaOp BatchNormGrad(const XlaOp& operand, const XlaOp& scale, + const XlaOp& batch_mean, const XlaOp& batch_var, + const XlaOp& grad_output, float epsilon, + int64 feature_index); + +// Implementation details below this point. + template XlaOp XlaBuilder::ConstantR0(NativeT value) { - return ConstantLiteral(*Literal::CreateR0(value)); + return ConstantLiteral(*LiteralUtil::CreateR0(value)); } template XlaOp XlaBuilder::ConstantR1(tensorflow::gtl::ArraySlice values) { - return ConstantLiteral(*Literal::CreateR1(values)); + return ConstantLiteral(*LiteralUtil::CreateR1(values)); } template @@ -954,44 +2046,44 @@ XlaOp XlaBuilder::ConstantR1(int64 length, NativeT value) { } inline XlaOp XlaBuilder::ConstantR1(const tensorflow::core::Bitmap& values) { - return ConstantLiteral(*Literal::CreateR1(values)); + return ConstantLiteral(*LiteralUtil::CreateR1(values)); } template XlaOp XlaBuilder::ConstantR2( std::initializer_list> values) { - return ConstantLiteral(*Literal::CreateR2(values)); + return ConstantLiteral(*LiteralUtil::CreateR2(values)); } template XlaOp XlaBuilder::ConstantFromArrayWithLayout(const Array& values, const Layout& layout) { return ConstantLiteral( - *Literal::CreateFromArrayWithLayout(values, layout)); + *LiteralUtil::CreateFromArrayWithLayout(values, layout)); } template XlaOp XlaBuilder::ConstantFromArray(const Array& values) { - return ConstantLiteral(*Literal::CreateFromArray(values)); + return ConstantLiteral(*LiteralUtil::CreateFromArray(values)); } template XlaOp XlaBuilder::ConstantR2FromArray2DWithLayout( const Array2D& values, const Layout& layout) { return ConstantLiteral( - *Literal::CreateFromArrayWithLayout(values, layout)); + *LiteralUtil::CreateFromArrayWithLayout(values, layout)); } template XlaOp XlaBuilder::ConstantR2FromArray2D(const Array2D& values) { - return ConstantLiteral(*Literal::CreateR2FromArray2D(values)); + return ConstantLiteral(*LiteralUtil::CreateR2FromArray2D(values)); } template XlaOp XlaBuilder::ConstantR3FromArray3DWithLayout( const Array3D& values, const Layout& layout) { return ConstantLiteral( - *Literal::CreateR3FromArray3DWithLayout(values, layout)); + *LiteralUtil::CreateR3FromArray3DWithLayout(values, layout)); } template @@ -1010,34 +2102,96 @@ XlaOp XlaBuilder::ConstantR4FromArray4D(const Array4D& values) { return ConstantFromArray(values); } -// RAII-style object: sets the current sharding assignment in builder on -// construction, and sets back to the previous assignment on destruction. -class XlaScopedShardingAssignment { - public: - XlaScopedShardingAssignment(xla::XlaBuilder* builder, - tensorflow::gtl::optional sharding) - : builder_(builder), prev_sharding_(builder->sharding()) { - SetSharding(sharding); - } +// Free function template implementations. - XlaScopedShardingAssignment(const XlaScopedShardingAssignment&) = delete; - XlaScopedShardingAssignment& operator=(const XlaScopedShardingAssignment&) = - delete; +template +XlaOp ConstantR0(XlaBuilder* builder, NativeT value) { + return ConstantLiteral(builder, *LiteralUtil::CreateR0(value)); +} - ~XlaScopedShardingAssignment() { SetSharding(prev_sharding_); } +template +XlaOp ConstantR1(XlaBuilder* builder, + tensorflow::gtl::ArraySlice values) { + return ConstantLiteral(builder, *LiteralUtil::CreateR1(values)); +} - private: - void SetSharding(const tensorflow::gtl::optional& sharding) { - if (sharding.has_value()) { - builder_->SetSharding(sharding.value()); - } else { - builder_->ClearSharding(); - } - } +template +XlaOp ConstantR1(XlaBuilder* builder, int64 length, NativeT value) { + Literal literal(ShapeUtil::MakeShape( + primitive_util::NativeToPrimitiveType(), {length})); + literal.PopulateWithValue(value); + return ConstantLiteral(builder, literal); +} - xla::XlaBuilder* const builder_; - tensorflow::gtl::optional prev_sharding_; -}; +inline XlaOp ConstantR1(XlaBuilder* builder, + const tensorflow::core::Bitmap& values) { + return ConstantLiteral(builder, *LiteralUtil::CreateR1(values)); +} + +template +XlaOp ConstantR2(XlaBuilder* builder, + std::initializer_list> values) { + return ConstantLiteral(builder, *LiteralUtil::CreateR2(values)); +} + +template +XlaOp ConstantFromArrayWithLayout(XlaBuilder* builder, + const Array& values, + const Layout& layout) { + return ConstantLiteral( + builder, + *LiteralUtil::CreateFromArrayWithLayout(values, layout)); +} + +template +XlaOp ConstantFromArray(XlaBuilder* builder, const Array& values) { + return ConstantLiteral(builder, + *LiteralUtil::CreateFromArray(values)); +} + +template +XlaOp ConstantR2FromArray2DWithLayout(XlaBuilder* builder, + const Array2D& values, + const Layout& layout) { + return ConstantLiteral( + builder, + *LiteralUtil::CreateFromArrayWithLayout(values, layout)); +} + +template +XlaOp ConstantR2FromArray2D(XlaBuilder* builder, + const Array2D& values) { + return ConstantLiteral(builder, + *LiteralUtil::CreateR2FromArray2D(values)); +} + +template +XlaOp ConstantR3FromArray3DWithLayout(XlaBuilder* builder, + const Array3D& values, + const Layout& layout) { + return ConstantLiteral( + builder, + *LiteralUtil::CreateR3FromArray3DWithLayout(values, layout)); +} + +template +XlaOp ConstantR3FromArray3D(XlaBuilder* builder, + const Array3D& values) { + return ConstantFromArray(builder, values); +} + +template +XlaOp ConstantR4FromArray4DWithLayout(XlaBuilder* builder, + const Array4D& values, + const Layout& layout) { + return ConstantFromArrayWithLayout(builder, values, layout); +} + +template +XlaOp ConstantR4FromArray4D(XlaBuilder* builder, + const Array4D& values) { + return ConstantFromArray(builder, values); +} } // namespace xla diff --git a/tensorflow/compiler/xla/client/xla_client/xla_builder_test.cc b/tensorflow/compiler/xla/client/xla_client/xla_builder_test.cc index 0680b38f3a27565beb6f7bcf47e4dbdd3d48cb52..3b8beb2c7840e23752b5f47bbc5f55d89751884d 100644 --- a/tensorflow/compiler/xla/client/xla_client/xla_builder_test.cc +++ b/tensorflow/compiler/xla/client/xla_client/xla_builder_test.cc @@ -53,16 +53,86 @@ class XlaBuilderTest : public ::testing::Test { TEST_F(XlaBuilderTest, OnePlusTwo) { XlaBuilder b(TestName()); - b.Add(b.ConstantR0(1.0), b.ConstantR0(2.0)); + Add(ConstantR0(&b, 1.0), ConstantR0(&b, 2.0)); TF_ASSERT_OK_AND_ASSIGN(auto module, BuildHloModule(&b)); auto root = module->entry_computation()->root_instruction(); EXPECT_THAT(root, op::Add(op::Constant(), op::Constant())); } +TEST_F(XlaBuilderTest, UnaryOperatorsBuildExpectedHLO) { + auto test_unary_operator = + [&](std::function op, + ::testing::Matcher matches_pattern) { + XlaBuilder b(TestName()); + op(ConstantR0(&b, 1)); + TF_ASSERT_OK_AND_ASSIGN(auto module, BuildHloModule(&b)); + auto root = module->entry_computation()->root_instruction(); + EXPECT_THAT(root, matches_pattern); + }; + test_unary_operator([](XlaOp x) { return -x; }, op::Negate(op::Constant())); + test_unary_operator([](XlaOp x) { return ~x; }, op::Not(op::Constant())); +} + +TEST_F(XlaBuilderTest, BinaryOperatorsBuildExpectedHLO) { + auto test_binary_operator = + [&](std::function op, + ::testing::Matcher matches_pattern) { + XlaBuilder b(TestName()); + op(ConstantR0(&b, 1), ConstantR0(&b, 2)); + TF_ASSERT_OK_AND_ASSIGN(auto module, BuildHloModule(&b)); + auto root = module->entry_computation()->root_instruction(); + EXPECT_THAT(root, matches_pattern); + }; + + test_binary_operator([](XlaOp x, XlaOp y) { return x + y; }, + op::Add(op::Constant(), op::Constant())); + test_binary_operator([](XlaOp x, XlaOp y) { return x - y; }, + op::Subtract(op::Constant(), op::Constant())); + test_binary_operator([](XlaOp x, XlaOp y) { return x * y; }, + op::Multiply(op::Constant(), op::Constant())); + test_binary_operator([](XlaOp x, XlaOp y) { return x / y; }, + op::Divide(op::Constant(), op::Constant())); + + test_binary_operator([](XlaOp x, XlaOp y) { return x & y; }, + op::And(op::Constant(), op::Constant())); + test_binary_operator([](XlaOp x, XlaOp y) { return x | y; }, + op::Or(op::Constant(), op::Constant())); + test_binary_operator([](XlaOp x, XlaOp y) { return x ^ y; }, + op::Xor(op::Constant(), op::Constant())); + test_binary_operator([](XlaOp x, XlaOp y) { return x << y; }, + op::ShiftLeft(op::Constant(), op::Constant())); + test_binary_operator( + [](XlaOp x, XlaOp y) { return x >> y; }, + op::ShiftRightArithmetic(op::Constant(), op::Constant())); + + auto test_unsigned_binary_operator = + [&](std::function op, + ::testing::Matcher matches_pattern) { + XlaBuilder b(TestName()); + op(ConstantR0(&b, 1), ConstantR0(&b, 2)); + TF_ASSERT_OK_AND_ASSIGN(auto module, BuildHloModule(&b)); + auto root = module->entry_computation()->root_instruction(); + EXPECT_THAT(root, matches_pattern); + }; + test_unsigned_binary_operator( + [](XlaOp x, XlaOp y) { return x >> y; }, + op::ShiftRightLogical(op::Constant(), op::Constant())); +} + +TEST_F(XlaBuilderTest, ShiftRightOperatorOnNonIntegerProducesError) { + XlaBuilder b(TestName()); + ConstantR0(&b, 1) >> ConstantR0(&b, 2); + auto statusor = b.Build(); + ASSERT_FALSE(statusor.ok()); + EXPECT_THAT( + statusor.status().error_message(), + HasSubstr("Argument to >> operator does not have an integral type")); +} + TEST_F(XlaBuilderTest, ParamPlusConstantHasScalarBroadcast) { XlaBuilder b(TestName()); - auto x = b.Parameter(0, ShapeUtil::MakeShape(F32, {3, 5}), "x"); - b.Add(x, b.ConstantR0(1.0)); + auto x = Parameter(&b, 0, ShapeUtil::MakeShape(F32, {3, 5}), "x"); + Add(x, ConstantR0(&b, 1.0)); TF_ASSERT_OK_AND_ASSIGN(auto module, BuildHloModule(&b)); auto root = module->entry_computation()->root_instruction(); EXPECT_THAT(root, op::Add(op::Parameter(), op::Broadcast(op::Constant()))); @@ -72,9 +142,9 @@ TEST_F(XlaBuilderTest, ParamPlusParamHasBroadcast) { XlaBuilder b(TestName()); const auto& x_shape = ShapeUtil::MakeShape(S32, {2, 4, 6}); const auto& y_shape = ShapeUtil::MakeShape(S32, {2, 4}); - auto x = b.Parameter(0, x_shape, "x"); - auto y = b.Parameter(1, y_shape, "y"); - auto add = b.Add(x, y, /*broadcast_dimensions=*/{0, 1}); + auto x = Parameter(&b, 0, x_shape, "x"); + auto y = Parameter(&b, 1, y_shape, "y"); + auto add = Add(x, y, /*broadcast_dimensions=*/{0, 1}); TF_ASSERT_OK_AND_ASSIGN(auto add_shape, b.GetShape(add)); EXPECT_TRUE(ShapeUtil::Equal(add_shape, x_shape)); @@ -86,8 +156,8 @@ TEST_F(XlaBuilderTest, ParamPlusParamHasBroadcast) { TEST_F(XlaBuilderTest, XPlusX) { XlaBuilder b(TestName()); - auto x = b.Parameter(0, ShapeUtil::MakeShape(S32, {1, 3, 5, 7}), "x"); - b.Add(x, x); + auto x = Parameter(&b, 0, ShapeUtil::MakeShape(S32, {1, 3, 5, 7}), "x"); + Add(x, x); TF_ASSERT_OK_AND_ASSIGN(auto module, BuildHloModule(&b)); auto root = module->entry_computation()->root_instruction(); EXPECT_THAT(root, op::Add(op::Parameter(0), op::Parameter(0))); @@ -95,9 +165,9 @@ TEST_F(XlaBuilderTest, XPlusX) { TEST_F(XlaBuilderTest, ShapeInferenceError) { XlaBuilder b(TestName()); - auto x = b.Parameter(0, ShapeUtil::MakeShape(U32, {2, 4, 6}), "x"); - auto y = b.Parameter(1, ShapeUtil::MakeShape(U32, {2, 4}), "y"); - b.Add(x, y); + auto x = Parameter(&b, 0, ShapeUtil::MakeShape(U32, {2, 4, 6}), "x"); + auto y = Parameter(&b, 1, ShapeUtil::MakeShape(U32, {2, 4}), "y"); + Add(x, y); auto statusor = BuildHloModule(&b); ASSERT_FALSE(statusor.ok()); EXPECT_THAT(statusor.status().error_message(), HasSubstr("shape inference")); @@ -105,12 +175,12 @@ TEST_F(XlaBuilderTest, ShapeInferenceError) { TEST_F(XlaBuilderTest, ParameterAlreadyRegistered) { XlaBuilder b_call("add"); - b_call.Parameter(0, ShapeUtil::MakeShape(PRED, {}), "x"); + Parameter(&b_call, 0, ShapeUtil::MakeShape(PRED, {}), "x"); XlaBuilder b(TestName()); - auto x = b.Parameter(0, ShapeUtil::MakeShape(PRED, {}), "x"); - auto y = b.Parameter(0, ShapeUtil::MakeShape(PRED, {}), "y"); - b.Add(x, y); + auto x = Parameter(&b, 0, ShapeUtil::MakeShape(PRED, {}), "x"); + auto y = Parameter(&b, 0, ShapeUtil::MakeShape(PRED, {}), "y"); + Add(x, y); auto statusor = BuildHloModule(&b); ASSERT_FALSE(statusor.ok()); EXPECT_THAT(statusor.status().error_message(), @@ -119,16 +189,16 @@ TEST_F(XlaBuilderTest, ParameterAlreadyRegistered) { TEST_F(XlaBuilderTest, Call) { XlaBuilder b_call("the_only_to_apply"); - auto p0 = b_call.Parameter(0, ShapeUtil::MakeShape(F32, {}), "p0"); - auto p1 = b_call.Parameter(1, ShapeUtil::MakeShape(F32, {}), "p1"); - b_call.Add(p0, p1); + auto p0 = Parameter(&b_call, 0, ShapeUtil::MakeShape(F32, {}), "p0"); + auto p1 = Parameter(&b_call, 1, ShapeUtil::MakeShape(F32, {}), "p1"); + Add(p0, p1); TF_ASSERT_OK_AND_ASSIGN(auto call, b_call.Build()); XlaBuilder b(TestName()); - auto x = b.Parameter(0, ShapeUtil::MakeShape(F32, {}), "x"); - auto y = b.Parameter(1, ShapeUtil::MakeShape(F32, {}), "y"); - auto one = b.ConstantR0(1); - auto two = b.ConstantR0(2); - b.Add(b.Call(call, {x, y}), b.Call(call, {one, two})); + auto x = Parameter(&b, 0, ShapeUtil::MakeShape(F32, {}), "x"); + auto y = Parameter(&b, 1, ShapeUtil::MakeShape(F32, {}), "y"); + auto one = ConstantR0(&b, 1); + auto two = ConstantR0(&b, 2); + Add(Call(&b, call, {x, y}), Call(&b, call, {one, two})); TF_ASSERT_OK_AND_ASSIGN(auto module, BuildHloModule(&b)); auto root = module->entry_computation()->root_instruction(); EXPECT_THAT(root, op::Add(op::Call(op::Parameter(), op::Parameter()), @@ -137,9 +207,9 @@ TEST_F(XlaBuilderTest, Call) { TEST_F(XlaBuilderTest, BinopHasDegenerateBroadcast) { XlaBuilder b(TestName()); - auto x = b.Parameter(0, ShapeUtil::MakeShape(F32, {1, 2, 3}), "x"); - auto y = b.Parameter(1, ShapeUtil::MakeShape(F32, {1, 2, 1}), "y"); - b.Add(x, y); + auto x = Parameter(&b, 0, ShapeUtil::MakeShape(F32, {1, 2, 3}), "x"); + auto y = Parameter(&b, 1, ShapeUtil::MakeShape(F32, {1, 2, 1}), "y"); + Add(x, y); TF_ASSERT_OK_AND_ASSIGN(auto module, BuildHloModule(&b)); // Expected: @@ -158,9 +228,9 @@ TEST_F(XlaBuilderTest, BinopHasDegenerateBroadcast) { TEST_F(XlaBuilderTest, BinopHasInDimAndDegenerateBroadcast) { XlaBuilder b(TestName()); - auto x = b.Parameter(0, ShapeUtil::MakeShape(F32, {2, 3}), "x"); - auto y = b.Parameter(1, ShapeUtil::MakeShape(F32, {2, 1, 4}), "y"); - b.Add(x, y, /*broadcast_dimensions=*/{0, 1}); + auto x = Parameter(&b, 0, ShapeUtil::MakeShape(F32, {2, 3}), "x"); + auto y = Parameter(&b, 1, ShapeUtil::MakeShape(F32, {2, 1, 4}), "y"); + Add(x, y, /*broadcast_dimensions=*/{0, 1}); TF_ASSERT_OK_AND_ASSIGN(auto module, BuildHloModule(&b)); // The binary operation has in-dim broadcast and degenerate broadcast, should @@ -183,9 +253,10 @@ TEST_F(XlaBuilderTest, BinopHasInDimAndDegenerateBroadcast) { TEST_F(XlaBuilderTest, OperandFromWrongBuilder) { XlaBuilder b1("b1"); - auto p0 = b1.Parameter(0, ShapeUtil::MakeShape(F32, {}), "p0"); + auto p0 = Parameter(&b1, 0, ShapeUtil::MakeShape(F32, {}), "p0"); XlaBuilder builder("main"); - builder.Add(p0, p0); + auto p = Parameter(&builder, 0, ShapeUtil::MakeShape(F32, {}), "p"); + Add(p, p0); auto statusor = builder.Build(); ASSERT_FALSE(statusor.ok()); EXPECT_THAT( @@ -196,8 +267,8 @@ TEST_F(XlaBuilderTest, OperandFromWrongBuilder) { TEST_F(XlaBuilderTest, ReshapeDefaultOrder) { XlaBuilder b(TestName()); - auto x = b.Parameter(0, ShapeUtil::MakeShape(F32, {2, 3, 5, 7}), "x"); - b.Reshape(x, /*new_sizes=*/{6, 35}); + auto x = Parameter(&b, 0, ShapeUtil::MakeShape(F32, {2, 3, 5, 7}), "x"); + Reshape(x, /*new_sizes=*/{6, 35}); TF_ASSERT_OK_AND_ASSIGN(auto module, BuildHloModule(&b)); auto root = module->entry_computation()->root_instruction(); EXPECT_THAT(root, op::Reshape(op::Parameter())); @@ -205,8 +276,8 @@ TEST_F(XlaBuilderTest, ReshapeDefaultOrder) { TEST_F(XlaBuilderTest, ReshapeHasTranspose) { XlaBuilder b(TestName()); - auto x = b.Parameter(0, ShapeUtil::MakeShape(F32, {2, 3, 5, 7}), "x"); - b.Reshape(x, /*dimensions=*/{3, 2, 1, 0}, /*new_sizes=*/{6, 35}); + auto x = Parameter(&b, 0, ShapeUtil::MakeShape(F32, {2, 3, 5, 7}), "x"); + Reshape(x, /*dimensions=*/{3, 2, 1, 0}, /*new_sizes=*/{6, 35}); TF_ASSERT_OK_AND_ASSIGN(auto module, BuildHloModule(&b)); auto root = module->entry_computation()->root_instruction(); EXPECT_THAT(root, op::Reshape(op::Transpose(op::Parameter()))); @@ -214,12 +285,39 @@ TEST_F(XlaBuilderTest, ReshapeHasTranspose) { TEST_F(XlaBuilderTest, Transpose) { XlaBuilder b(TestName()); - auto x = b.Parameter(0, ShapeUtil::MakeShape(F32, {5, 7}), "x"); - b.Transpose(x, /*permutation=*/{1, 0}); + auto x = Parameter(&b, 0, ShapeUtil::MakeShape(F32, {5, 7}), "x"); + Transpose(x, /*permutation=*/{1, 0}); TF_ASSERT_OK_AND_ASSIGN(auto module, BuildHloModule(&b)); auto root = module->entry_computation()->root_instruction(); EXPECT_THAT(root, op::Transpose(op::Parameter())); } +TEST_F(XlaBuilderTest, ReportError) { + XlaBuilder b(TestName()); + auto x = Parameter(&b, 0, ShapeUtil::MakeShape(F32, {5, 7}), "x"); + Add(b.ReportError(InvalidArgument("a test error")), x); + auto statusor = b.Build(); + ASSERT_FALSE(statusor.ok()); + EXPECT_THAT(statusor.status().error_message(), HasSubstr("a test error")); +} + +TEST_F(XlaBuilderTest, ReportErrorOrReturnHandlesNonErrors) { + XlaBuilder b(TestName()); + StatusOr op(ConstantR0(&b, 1.0)); + Add(b.ReportErrorOrReturn(op), ConstantR0(&b, 2.0)); + TF_ASSERT_OK_AND_ASSIGN(auto module, BuildHloModule(&b)); + auto root = module->entry_computation()->root_instruction(); + EXPECT_THAT(root, op::Add(op::Constant(), op::Constant())); +} + +TEST_F(XlaBuilderTest, ReportErrorOrReturnHandlesErrors) { + XlaBuilder b(TestName()); + StatusOr op(InvalidArgument("a test error")); + Add(b.ReportErrorOrReturn(op), ConstantR0(&b, 2.0)); + auto statusor = b.Build(); + ASSERT_FALSE(statusor.ok()); + EXPECT_THAT(statusor.status().error_message(), HasSubstr("a test error")); +} + } // namespace } // namespace xla diff --git a/tensorflow/compiler/xla/layout_util.cc b/tensorflow/compiler/xla/layout_util.cc index 3f059cac30b5d36ab1d097bf200547533822e3d0..15eeb2ea13607d43c995197f8f0e3c58abd4d94a 100644 --- a/tensorflow/compiler/xla/layout_util.cc +++ b/tensorflow/compiler/xla/layout_util.cc @@ -248,6 +248,12 @@ Layout CreateDefaultLayoutForRank(int64 rank) { } } + if (layout.format() == SPARSE) { + if (!layout.padded_dimensions().empty()) { + return InvalidArgument("Sparse layout has padded dimensions"); + } + } + return Status::OK(); } diff --git a/tensorflow/compiler/xla/literal.cc b/tensorflow/compiler/xla/literal.cc new file mode 100644 index 0000000000000000000000000000000000000000..5db124b5a226238931a2038969d1f131743e554a --- /dev/null +++ b/tensorflow/compiler/xla/literal.cc @@ -0,0 +1,1967 @@ +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow/compiler/xla/literal.h" + +#include +#include +#include +#include +#include +#include + +#include "tensorflow/compiler/xla/index_util.h" +#include "tensorflow/compiler/xla/shape_util.h" +#include "tensorflow/compiler/xla/status_macros.h" +#include "tensorflow/compiler/xla/types.h" +#include "tensorflow/compiler/xla/util.h" +#include "tensorflow/core/lib/core/casts.h" +#include "tensorflow/core/lib/core/errors.h" +#include "tensorflow/core/lib/hash/hash.h" +#include "tensorflow/core/lib/strings/str_util.h" +#include "tensorflow/core/lib/strings/strcat.h" +#include "tensorflow/core/lib/strings/stringprintf.h" +#include "tensorflow/core/platform/logging.h" +#include "tensorflow/core/platform/types.h" + +using tensorflow::strings::Printf; +using tensorflow::strings::StrCat; + +namespace xla { + +namespace { + +constexpr bool kLittleEndian = __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__; + +// Converts between little and big endian. +// +// Precondition: size % 2 == 0 (elements in the array are 16 bits long) +void ConvertEndianShort(string* bytes) { + CHECK_EQ(bytes->size() / 2, 0); + for (int64 i = 0; i < bytes->size(); i += 2) { + std::swap((*bytes)[i], (*bytes)[i + 1]); + } +} + +void ConvertEndianShort(char* bytes, int64 size) { + CHECK_EQ(size / 2, 0); + for (int64 i = 0; i < size; i += 2) { + std::swap(bytes[i], bytes[i + 1]); + } +} + +} // namespace + +LiteralBase::~LiteralBase() {} + +std::ostream& operator<<(std::ostream& out, const Literal& literal) { + out << literal.ToString(); + return out; +} + +Literal::StrideConfig::StrideConfig( + const Shape& source_shape, const Shape& dest_shape, + tensorflow::gtl::ArraySlice dimensions) + : dimensions(dimensions), + base(dimensions.size(), 0), + step(dimensions.size(), 1) { + if (!dimensions.empty()) { + // Selects the shape with the largest minor dimension as the one upon + // which to run the tight stride loop. + if (dimensions[LayoutUtil::Minor(source_shape.layout(), 0)] >= + dimensions[LayoutUtil::Minor(dest_shape.layout(), 0)]) { + minor_dimension = LayoutUtil::Minor(source_shape.layout(), 0); + dest_stride = IndexUtil::GetDimensionStride(dest_shape, minor_dimension); + } else { + minor_dimension = LayoutUtil::Minor(dest_shape.layout(), 0); + source_stride = + IndexUtil::GetDimensionStride(source_shape, minor_dimension); + } + minor_loop_size = dimensions[minor_dimension]; + step[minor_dimension] = minor_loop_size; + } +} + +Literal::Literal(const Shape& shape) + : Literal(shape, /*allocate_arrays=*/true) {} + +void Literal::SetPiece(const Shape& shape, Piece* piece, bool allocate_arrays) { + if (ShapeUtil::IsTuple(shape)) { + for (int i = 0; i < ShapeUtil::TupleElementCount(shape); ++i) { + const Shape& subshape = shape.tuple_shapes(i); + + auto child_piece = Piece(); + child_piece.set_subshape(&subshape); + + SetPiece(subshape, &child_piece, allocate_arrays); + + piece->emplace_back(std::move(child_piece)); + } + } else if (ShapeUtil::IsArray(shape)) { + if (allocate_arrays) { + if (LayoutUtil::IsSparseArray(shape)) { + // For sparse arrays, the buffer must be of the size of the maximum + // number of sparse elements possible. + const int64 max_sparse_elements = + LayoutUtil::MaxSparseElements(shape.layout()); + piece->set_buffer( + new char[max_sparse_elements * + ShapeUtil::ByteSizeOfPrimitiveType(shape.element_type())]); + piece->set_sparse_indices( + new SparseIndexArray(max_sparse_elements, ShapeUtil::Rank(shape))); + } else { + piece->set_buffer(new char[piece->size_bytes()]); + } + } + } else { + // If the shape is neither an array nor tuple, then it must be + // zero-sized. Otherwise, some memory needs to be allocated for it. + CHECK_EQ(piece->size_bytes(), 0); + } +} + +Literal::Literal(const Shape& shape, bool allocate_arrays) + : LiteralBase(), shape_(MakeUnique(shape)) { + CHECK(LayoutUtil::HasLayout(*shape_)); + root_piece_ = new Piece(); + root_piece_->set_subshape(shape_.get()); + CHECK(&root_piece_->subshape() == shape_.get()); + + SetPiece(*shape_, root_piece_, allocate_arrays); +} + +Literal::~Literal() { + if (root_piece_ != nullptr) { + DeallocateBuffers(); + delete root_piece_; + } +} + +void Literal::DeallocateBuffers() { + root_piece_->ForEachMutableSubpiece( + [&](const ShapeIndex& index, Piece* piece) { + if (piece->buffer() != nullptr) { + delete[] piece->buffer(); + delete piece->sparse_indices(); + } + }); +} + +Literal::Literal(Literal&& other) : LiteralBase() { *this = std::move(other); } + +Literal& Literal::operator=(Literal&& other) { + DCHECK(&other.root_piece_->subshape() == other.shape_.get()); + using std::swap; + swap(shape_, other.shape_); + swap(root_piece_, other.root_piece_); + DCHECK(&root_piece_->subshape() == shape_.get()); + + return *this; +} + +std::unique_ptr LiteralBase::CreateFromShape(const Shape& shape) { + auto literal = MakeUnique(shape); + literal->root_piece_->ForEachMutableSubpiece( + [&](const ShapeIndex& index, Piece* piece) { + if (ShapeUtil::IsArray(piece->subshape())) { + memset(piece->untyped_data(), 0, piece->size_bytes()); + } + }); + return literal; +} + +const SparseIndexArray* LiteralBase::sparse_indices( + const ShapeIndex& shape_index) const { + return piece(shape_index).sparse_indices(); +} + +SparseIndexArray* Literal::sparse_indices(const ShapeIndex& shape_index) { + return piece(shape_index).sparse_indices(); +} + +template +Status Literal::CopySliceFromInternal( + const LiteralBase& src_literal, tensorflow::gtl::ArraySlice src_base, + tensorflow::gtl::ArraySlice dest_base, + tensorflow::gtl::ArraySlice copy_size) { + TF_RET_CHECK(ShapeUtil::Rank(src_literal.shape()) == src_base.size()); + TF_RET_CHECK(ShapeUtil::Rank(shape()) == dest_base.size()); + + auto linear_index = [](const Shape& shape, + tensorflow::gtl::ArraySlice multi_index) { + return IndexUtil::MultidimensionalIndexToLinearIndex(shape, multi_index); + }; + + if (ShapeUtil::Rank(src_literal.shape()) == 0 || + ShapeUtil::Rank(shape()) == 0) { + // If any of the two shapes are scalars, we can just call the StridedCopy() + // directly, and we know we will be copying only one value. + TF_RET_CHECK(copy_size.empty()); + StridedCopy(data(), linear_index(shape(), dest_base), 0, + src_literal.data(), + linear_index(src_literal.shape(), src_base), 0, 1); + } else if (!ShapeUtil::IsZeroElementArray(shape()) && + !ShapeUtil::IsZeroElementArray(src_literal.shape())) { + // Perform copy if neither src nor dest has dimensions with zero element, + // otherwise it's a no-op. + TF_RET_CHECK(src_base.size() == dest_base.size()); + TF_RET_CHECK(src_base.size() == copy_size.size()); + + // Scan the source from minor, stepping in copy size blocks, then within + // the index enumaration functor, do a strided copy advancing source index + // by one (walking through the minor dimension), and destination index by + // proper stride size at the matching dimension. + DimensionVector src_indexes(src_base.size(), 0); + DimensionVector dest_indexes(dest_base.size(), 0); + Literal::StrideConfig stride_config(src_literal.shape(), shape(), + copy_size); + + auto copy_proc = [&](tensorflow::gtl::ArraySlice indexes) { + // Map from multi-dimensional index, to source index. + std::transform(indexes.begin(), indexes.end(), src_base.begin(), + src_indexes.begin(), std::plus()); + // Map from multi-dimensional index, to destination index. + std::transform(indexes.begin(), indexes.end(), dest_base.begin(), + dest_indexes.begin(), std::plus()); + + int64 src_index = linear_index(src_literal.shape(), src_indexes); + int64 dest_index = linear_index(shape(), dest_indexes); + + // `this->` is needed to workaround MSVC bug: #16882 + StridedCopy(this->data(), dest_index, stride_config.dest_stride, + src_literal.data(), src_index, + stride_config.source_stride, stride_config.minor_loop_size); + return true; + }; + + ShapeUtil::ForEachIndex(src_literal.shape(), stride_config.base, + stride_config.dimensions, stride_config.step, + copy_proc); + } + return Status::OK(); +} + +Status Literal::CopyElementFrom(const LiteralSlice& src_literal, + tensorflow::gtl::ArraySlice src_index, + tensorflow::gtl::ArraySlice dest_index) { + DCHECK_EQ(shape().element_type(), src_literal.shape().element_type()); + const int64 src_linear_index = IndexUtil::MultidimensionalIndexToLinearIndex( + src_literal.shape(), src_index); + const int64 dest_linear_index = + IndexUtil::MultidimensionalIndexToLinearIndex(shape(), dest_index); + const int64 primitive_size = + ShapeUtil::ByteSizeOfPrimitiveType(shape().element_type()); + + char* dest_address = + static_cast(untyped_data()) + dest_linear_index * primitive_size; + const char* source_address = + static_cast(src_literal.untyped_data()) + + src_linear_index * primitive_size; + if (dest_address != source_address) { + memcpy(dest_address, source_address, primitive_size); + } + return Status::OK(); +} + +/* static */ StatusOr> Literal::CreateFromProto( + const LiteralProto& proto) { + if (!proto.has_shape()) { + return InvalidArgument("LiteralProto has no shape"); + } + if (!LayoutUtil::HasLayout(proto.shape())) { + return InvalidArgument("LiteralProto has no layout"); + } + + auto literal = MakeUnique(proto.shape()); + + TF_RETURN_IF_ERROR(literal->root_piece_->ForEachMutableSubpieceWithStatus( + [&](const ShapeIndex& index, Piece* piece) { + const LiteralProto* proto_element = &proto; + for (int64 i : index) { + CHECK(i < proto_element->tuple_literals_size()); + proto_element = &proto_element->tuple_literals(i); + } + + if (ShapeUtil::IsTuple(piece->subshape())) { + if (proto_element->tuple_literals_size() != + ShapeUtil::TupleElementCount(piece->subshape())) { + return InvalidArgument( + "Expected %lld tuple elements in LiteralProto, has %d", + ShapeUtil::TupleElementCount(piece->subshape()), + proto_element->tuple_literals_size()); + } + return Status::OK(); + } + if (piece->subshape().element_type() == TOKEN) { + return Status::OK(); + } + + CHECK(ShapeUtil::IsArray(piece->subshape())); + TF_RETURN_IF_ERROR(piece->CopyFromProto(*proto_element)); + + return Status::OK(); + })); + + return std::move(literal); +} + +std::vector Literal::DecomposeTuple() { + CHECK(ShapeUtil::IsTuple(shape())); + std::vector elements; + for (int i = 0; i < ShapeUtil::TupleElementCount(shape()); ++i) { + elements.push_back(Literal(ShapeUtil::GetSubshape(shape(), {i}), + /*allocate_arrays=*/false)); + Literal& element = elements.back(); + element.root_piece_->ForEachMutableSubpiece( + [&](const ShapeIndex& index, Piece* dest_piece) { + ShapeIndex src_index = {i}; + for (int64 j : index) { + src_index.push_back(j); + } + Piece& src_piece = piece(src_index); + + // Move the respective buffer and sparse indices over to the element + // Literal. + dest_piece->set_buffer(src_piece.buffer()); + src_piece.set_buffer(nullptr); + dest_piece->set_sparse_indices(src_piece.sparse_indices()); + src_piece.set_sparse_indices(nullptr); + }); + } + // Set this literal to be nil-shaped. + *this = Literal(); + return elements; +} + +namespace { + +// Copies the elements in 'src' to 'dest'. The shape and layout of the data in +// the array slices are indicated by dest_shape and src_shape respectively. +template +void CopyElementsBetween(tensorflow::gtl::MutableArraySlice dest, + tensorflow::gtl::ArraySlice src, + const Shape& dest_shape, const Shape& src_shape) { + CHECK(ShapeUtil::Compatible(dest_shape, src_shape)); + if (ShapeUtil::IsZeroElementArray(dest_shape)) { + return; + } + std::vector index(ShapeUtil::Rank(dest_shape)); + do { + dest[IndexUtil::MultidimensionalIndexToLinearIndex(dest_shape, index)] = + src[IndexUtil::MultidimensionalIndexToLinearIndex(src_shape, index)]; + } while (IndexUtil::BumpIndices(dest_shape, &index)); +} + +} // namespace + +Status LiteralBase::Piece::CopyFrom(const LiteralBase::Piece& src) { + CHECK(subshape_ != nullptr); + CHECK(src.subshape_ != nullptr); + if (ShapeUtil::Equal(subshape(), src.subshape())) { + // If the layouts are equal it's faster just to memcpy. + memcpy(buffer(), src.buffer(), src.size_bytes()); + } else { + TF_RET_CHECK(ShapeUtil::Compatible(src.subshape(), subshape())); + std::vector origin(ShapeUtil::Rank(subshape()), 0); + switch (subshape().element_type()) { +#define COPY_ELEMENTS(XLA_T, NATIVE_T) \ + case (XLA_T): \ + CopyElementsBetween(data(), src.data(), \ + subshape(), src.subshape()); \ + break; + COPY_ELEMENTS(U8, uint8); + COPY_ELEMENTS(U16, uint16); + COPY_ELEMENTS(U32, uint32); + COPY_ELEMENTS(U64, uint64); + COPY_ELEMENTS(S8, int8); + COPY_ELEMENTS(S16, int16); + COPY_ELEMENTS(S32, int32); + COPY_ELEMENTS(S64, int64); + COPY_ELEMENTS(F16, half); + COPY_ELEMENTS(BF16, bfloat16); + COPY_ELEMENTS(F32, float); + COPY_ELEMENTS(F64, double); + COPY_ELEMENTS(C64, complex64); + COPY_ELEMENTS(PRED, bool); +#undef COPY_ELEMENTS + default: + return Unimplemented( + "Copying a Literal object with element type %s is not implemented.", + PrimitiveType_Name(subshape().element_type()).c_str()); + } + } + return Status::OK(); +} + +Status Literal::CopyFrom(const LiteralSlice& src_literal, + const ShapeIndex& dest_shape_index, + const ShapeIndex& src_shape_index) { + const Shape& dest_subshape = + ShapeUtil::GetSubshape(shape(), dest_shape_index); + const Shape& src_subshape = + ShapeUtil::GetSubshape(src_literal.shape(), src_shape_index); + if (!ShapeUtil::Compatible(dest_subshape, src_subshape)) { + return InvalidArgument( + "Destination subshape incompatible with source subshape: %s vs %s", + ShapeUtil::HumanString(dest_subshape).c_str(), + ShapeUtil::HumanString(src_subshape).c_str()); + } + return root_piece_->ForEachMutableSubpieceWithStatus( + [&](const ShapeIndex& index, Piece* piece) { + if (!ShapeUtil::IsArray(piece->subshape())) { + return Status::OK(); + } + + // Determine if this index is in the part of this literal that we want + // to copy over from src_literal. + bool in_subtree_to_copy = true; + for (int i = 0; i < dest_shape_index.size(); ++i) { + if (index[i] != dest_shape_index[i]) { + in_subtree_to_copy = false; + break; + } + } + if (!in_subtree_to_copy) { + return Status::OK(); + } + // Construct the index of the corresponding piece in the source literal. + ShapeIndex src_piece_index = src_shape_index; + for (int64 i = dest_shape_index.size(); i < index.size(); ++i) { + src_piece_index.push_back(index[i]); + } + TF_RETURN_IF_ERROR(piece->CopyFrom(src_literal.piece(src_piece_index))); + return Status::OK(); + }); +} + +Status Literal::MoveFrom(Literal&& src_literal, + const ShapeIndex& dest_shape_index) { + const Shape& dest_subshape = + ShapeUtil::GetSubshape(shape(), dest_shape_index); + if (!ShapeUtil::Equal(dest_subshape, src_literal.shape())) { + return InvalidArgument( + "Destination subshape not equal to source shape: %s vs %s", + ShapeUtil::HumanString(dest_subshape).c_str(), + ShapeUtil::HumanString(src_literal.shape()).c_str()); + } + + src_literal.root_piece_->ForEachSubpiece( + [&](const ShapeIndex& src_index, const Piece& src_piece) { + if (!ShapeUtil::IsArray(src_piece.subshape())) { + return; + } + + ShapeIndex dest_index = dest_shape_index; + for (int64 i : src_index) { + dest_index.push_back(i); + } + Piece& dest_piece = piece(dest_index); + delete[] dest_piece.buffer(); + dest_piece.set_buffer(src_piece.buffer()); + delete dest_piece.sparse_indices(); + dest_piece.set_sparse_indices(src_piece.sparse_indices()); + }); + + src_literal.shape_ = MakeUnique(ShapeUtil::MakeNil()); + delete src_literal.root_piece_; + src_literal.root_piece_ = new LiteralBase::Piece(); + src_literal.root_piece_->set_subshape(src_literal.shape_.get()); + + return Status::OK(); +} + +Status Literal::CopySliceFrom(const LiteralSlice& src_literal, + tensorflow::gtl::ArraySlice src_base, + tensorflow::gtl::ArraySlice dest_base, + tensorflow::gtl::ArraySlice copy_size) { + TF_RET_CHECK(ShapeUtil::IsArray(shape())) << ShapeUtil::HumanString(shape()); + TF_RET_CHECK(ShapeUtil::IsArray(src_literal.shape())) + << ShapeUtil::HumanString(src_literal.shape()); + TF_RET_CHECK(ShapeUtil::SameElementType(src_literal.shape(), shape())); + + switch (shape().element_type()) { + case U8: + return CopySliceFromInternal(src_literal, src_base, dest_base, + copy_size); + case U16: + return CopySliceFromInternal(src_literal, src_base, dest_base, + copy_size); + case U32: + return CopySliceFromInternal(src_literal, src_base, dest_base, + copy_size); + case U64: + return CopySliceFromInternal(src_literal, src_base, dest_base, + copy_size); + case S8: + return CopySliceFromInternal(src_literal, src_base, dest_base, + copy_size); + case S16: + return CopySliceFromInternal(src_literal, src_base, dest_base, + copy_size); + case S32: + return CopySliceFromInternal(src_literal, src_base, dest_base, + copy_size); + case S64: + return CopySliceFromInternal(src_literal, src_base, dest_base, + copy_size); + case F16: + return CopySliceFromInternal(src_literal, src_base, dest_base, + copy_size); + case BF16: + return CopySliceFromInternal(src_literal, src_base, dest_base, + copy_size); + case F32: + return CopySliceFromInternal(src_literal, src_base, dest_base, + copy_size); + case F64: + return CopySliceFromInternal(src_literal, src_base, dest_base, + copy_size); + case C64: + return CopySliceFromInternal(src_literal, src_base, dest_base, + copy_size); + case PRED: + return CopySliceFromInternal(src_literal, src_base, dest_base, + copy_size); + default: + break; + } + return Unimplemented( + "Copying a slice from a Literal object with element type %d is not " + "implemented.", + shape().element_type()); +} + +void Literal::PopulateR1(const tensorflow::core::Bitmap& values) { + CHECK(ShapeUtil::IsArray(shape())); + CHECK_EQ(ShapeUtil::Rank(shape()), 1); + CHECK_EQ(element_count(), values.bits()); + CHECK_EQ(shape().element_type(), PRED); + for (int64 i = 0; i < static_cast(values.bits()); ++i) { + Set({i}, values.get(i)); + } +} + +std::unique_ptr LiteralBase::Relayout( + const Layout& new_layout, const ShapeIndex& shape_index) const { + // Create new shape with 'new_layout' set at the given shape index. + Shape new_shape = shape(); + Shape* subshape = ShapeUtil::GetMutableSubshape(&new_shape, shape_index); + TF_CHECK_OK(LayoutUtil::ValidateLayoutForShape(new_layout, *subshape)); + *subshape->mutable_layout() = new_layout; + auto result = MakeUnique(new_shape); + TF_CHECK_OK(result->CopyFrom(*this)); + return result; +} + +std::unique_ptr LiteralBase::Relayout( + const Shape& shape_with_layout) const { + CHECK(ShapeUtil::Compatible(shape_with_layout, shape())) + << "Given shape_with_layout " << ShapeUtil::HumanString(shape_with_layout) + << " not compatible with literal shape " + << ShapeUtil::HumanString(shape()); + std::unique_ptr result = CreateFromShape(shape_with_layout); + ShapeUtil::ForEachSubshape( + result->shape(), + [this, &result](const Shape& subshape, const ShapeIndex& index) { + if (ShapeUtil::IsArray(subshape)) { + TF_CHECK_OK(result->CopyFrom(*this, + /*dest_shape_index=*/index, + /*src_shape_index=*/index)); + } + }); + return result; +} + +StatusOr> LiteralBase::Broadcast( + const Shape& result_shape, + tensorflow::gtl::ArraySlice dimensions) const { + if (!ShapeUtil::IsArray(shape())) { + return InvalidArgument("Broadcast only supports arrays."); + } + + for (int64 i = 0; i < dimensions.size(); i++) { + TF_RET_CHECK(shape().dimensions(i) == + result_shape.dimensions(dimensions[i])); + } + + std::unique_ptr result = MakeUnique(result_shape); + + // scratch_source_index is temporary storage space for the computed index into + // the input literal. We put it here to avoid allocating an std::vector in + // every iteration of ShapeUtil::ForEachIndex. + std::vector scratch_source_index(shape().dimensions_size()); + + char* dest_data = static_cast(result->untyped_data()); + const char* source_data = static_cast(untyped_data()); + const int64 primitive_size = + ShapeUtil::ByteSizeOfPrimitiveType(shape().element_type()); + + ShapeUtil::ForEachIndex( + result_shape, [&](tensorflow::gtl::ArraySlice output_index) { + for (int64 i = 0; i < dimensions.size(); ++i) { + scratch_source_index[i] = output_index[dimensions[i]]; + } + int64 dest_index = IndexUtil::MultidimensionalIndexToLinearIndex( + result_shape, output_index); + int64 source_index = IndexUtil::MultidimensionalIndexToLinearIndex( + shape(), scratch_source_index); + memcpy(dest_data + primitive_size * dest_index, + source_data + primitive_size * source_index, primitive_size); + return true; + }); + + return std::move(result); +} + +StatusOr> LiteralBase::Reshape( + tensorflow::gtl::ArraySlice dimensions) const { + if (!ShapeUtil::IsArray(shape())) { + return InvalidArgument("Reshape does not support tuples."); + } + std::unique_ptr output; + if (!LayoutUtil::IsMonotonicWithDim0Major(shape().layout())) { + output = + Relayout(LayoutUtil::GetDefaultLayoutForRank(ShapeUtil::Rank(shape()))); + } else { + output = CloneToUnique(); + } + // Because the layout is monotonic, we can simply reuse the same sequence of + // values without changing their order. + *output->mutable_shape_do_not_use() = + ShapeUtil::MakeShape(shape().element_type(), dimensions); + + int64 elements_before = ShapeUtil::ElementsIn(shape()); + int64 elements_after = ShapeUtil::ElementsIn(output->shape()); + if (elements_before != elements_after) { + return InvalidArgument( + "Shapes before and after Literal::Reshape have different numbers " + "of elements: %s vs %s.", + ShapeUtil::HumanString(shape()).c_str(), + ShapeUtil::HumanString(output->shape()).c_str()); + } + return std::move(output); +} + +std::unique_ptr LiteralBase::Transpose( + tensorflow::gtl::ArraySlice permutation) const { + CHECK(ShapeUtil::IsArray(shape())) << "Tuple is not supported for transpose"; + CHECK(IsPermutation(permutation, ShapeUtil::Rank(shape()))) + << "Given permutation is not a permutation of dimension numbers"; + // To transpose the array, we just permute the dimensions and layout, and + // do a straight memory copy of the raw data set. + // This is considerably faster than iterating over every array element using + // the EachCell<>() and Set<>() APIs. + std::vector inverse_permutation = InversePermutation(permutation); + Shape permuted_shape = + ShapeUtil::PermuteDimensions(inverse_permutation, shape()); + // Replace the layout with one affine to this shape, such that a + // transpose operation can be performed by leaving the flat values + // representation intact. + // For example, consider the shape F32[11,8]{1,0} under a {1,0} permutation. + // The shape with affine layout resulting from that operation will be + // F32[8,11]{0,1}, since it leaves the original most minor (the 8 sized), the + // most minor. + // + // Essentially, given MinMaj(Di) the position of the Di dimension within the + // minor to major vector, and given T(Di) the index that the original Di + // dimension has within the transposed array, a layout is affine if + // MinMaj(Di) == TMinMaj(T(Di)), with TMinMaj() being the minor to major + // vector of the affine layout. + CHECK(LayoutUtil::IsDenseArray(permuted_shape)); + Layout* layout = permuted_shape.mutable_layout(); + layout->clear_minor_to_major(); + for (auto index : LayoutUtil::MinorToMajor(shape())) { + layout->add_minor_to_major(inverse_permutation[index]); + } + auto new_literal = MakeUnique(permuted_shape); + DCHECK_EQ(ShapeUtil::ByteSizeOf(new_literal->shape()), + ShapeUtil::ByteSizeOf(shape())); + std::memcpy(new_literal->untyped_data(), untyped_data(), size_bytes()); + return new_literal; +} + +template +std::unique_ptr LiteralBase::SliceInternal( + const Shape& result_shape, + tensorflow::gtl::ArraySlice start_indices) const { + auto result_literal = MakeUnique(result_shape); + DimensionVector new_indices(ShapeUtil::Rank(result_shape)); + result_literal->EachCell( + [&](tensorflow::gtl::ArraySlice indices, NativeT /*value*/) { + for (int64 i = 0; i < ShapeUtil::Rank(result_shape); ++i) { + new_indices[i] = indices[i] + start_indices[i]; + } + NativeT value = Get(new_indices); + result_literal->Set(indices, value); + }); + return result_literal; +} + +std::unique_ptr LiteralBase::Slice( + tensorflow::gtl::ArraySlice start_indices, + tensorflow::gtl::ArraySlice limit_indices) const { + CHECK(ShapeUtil::IsArray(shape())) << "tuple is not supported for slice"; + + DimensionVector result_dimensions; + for (int64 dnum = 0; dnum < ShapeUtil::Rank(shape()); ++dnum) { + CHECK_GE(start_indices[dnum], 0); + CHECK_LE(limit_indices[dnum], shape().dimensions(dnum)) + << "dnum = " << dnum; + int64 dimension = limit_indices[dnum] - start_indices[dnum]; + CHECK_GE(dimension, 0) << "dnum = " << dnum; + result_dimensions.push_back(dimension); + } + const auto result_shape = + ShapeUtil::MakeShapeWithLayout(shape().element_type(), result_dimensions, + LayoutUtil::MinorToMajor(shape())); + switch (result_shape.element_type()) { + case F32: + return SliceInternal(result_shape, start_indices); + case BF16: + return SliceInternal(result_shape, start_indices); + case C64: + return SliceInternal(result_shape, start_indices); + case S32: + return SliceInternal(result_shape, start_indices); + case U32: + return SliceInternal(result_shape, start_indices); + default: + LOG(FATAL) << "not yet implemented: " + << PrimitiveType_Name(result_shape.element_type()); + } +} + +Literal LiteralBase::Clone() const { + Literal result(shape()); + TF_CHECK_OK(result.CopyFrom(*this)); + return result; +} + +std::unique_ptr LiteralBase::CloneToUnique() const { + auto result = MakeUnique(shape()); + TF_CHECK_OK(result->CopyFrom(*this)); + return result; +} + +string LiteralBase::GetAsString(tensorflow::gtl::ArraySlice multi_index, + const ShapeIndex& shape_index) const { + const Shape& subshape = ShapeUtil::GetSubshape(shape(), shape_index); + CHECK(LayoutUtil::IsDenseArray(subshape)); + switch (subshape.element_type()) { + case PRED: + return Get(multi_index, shape_index) ? "true" : "false"; + case S8: + return StrCat(Get(multi_index, shape_index)); + case S16: + return StrCat(Get(multi_index, shape_index)); + case S32: + return StrCat(Get(multi_index, shape_index)); + case S64: + return StrCat(Get(multi_index, shape_index)); + case U8: + return StrCat(Get(multi_index, shape_index)); + case U16: + return StrCat(Get(multi_index, shape_index)); + case U32: + return StrCat(Get(multi_index, shape_index)); + case U64: + return StrCat(Get(multi_index, shape_index)); + case F16: + return StrCat(static_cast(Get(multi_index, shape_index))); + case F32: + return StrCat(Get(multi_index, shape_index)); + case BF16: + return StrCat( + static_cast(Get(multi_index, shape_index))); + case F64: + return StrCat(Get(multi_index, shape_index)); + case C64: { + complex64 c = Get(multi_index, shape_index); + return StrCat("(", c.real(), ", ", c.imag(), ")"); + } + default: + LOG(FATAL) << PrimitiveType_Name(subshape.element_type()); + } +} + +string LiteralBase::GetSparseElementAsString( + int64 sparse_element_number, const ShapeIndex& shape_index) const { + const Shape& subshape = ShapeUtil::GetSubshape(shape(), shape_index); + CHECK(LayoutUtil::IsSparseArray(subshape)); + switch (subshape.element_type()) { + case PRED: + return GetSparseElement(sparse_element_number, shape_index) + ? "true" + : "false"; + case S8: + return StrCat(GetSparseElement(sparse_element_number, shape_index)); + case S16: + return StrCat( + GetSparseElement(sparse_element_number, shape_index)); + case S32: + return StrCat( + GetSparseElement(sparse_element_number, shape_index)); + case S64: + return StrCat( + GetSparseElement(sparse_element_number, shape_index)); + case U8: + return StrCat( + GetSparseElement(sparse_element_number, shape_index)); + case U16: + return StrCat( + GetSparseElement(sparse_element_number, shape_index)); + case U32: + return StrCat( + GetSparseElement(sparse_element_number, shape_index)); + case U64: + return StrCat( + GetSparseElement(sparse_element_number, shape_index)); + case F16: + return StrCat(static_cast( + GetSparseElement(sparse_element_number, shape_index))); + case F32: + return StrCat( + GetSparseElement(sparse_element_number, shape_index)); + case BF16: + return StrCat(static_cast( + GetSparseElement(sparse_element_number, shape_index))); + case F64: + return StrCat( + GetSparseElement(sparse_element_number, shape_index)); + case C64: { + complex64 c = + GetSparseElement(sparse_element_number, shape_index); + return StrCat("(", c.real(), ", ", c.imag(), ")"); + } + default: + LOG(FATAL) << "Invalid element type for sparse arrays: " + << PrimitiveType_Name(subshape.element_type()); + } +} + +StatusOr LiteralBase::GetIntegralAsS64( + tensorflow::gtl::ArraySlice multi_index) const { + CHECK(LayoutUtil::IsDenseArray(shape())); + switch (shape().element_type()) { + case PRED: + return Get(multi_index); + case U8: + return Get(multi_index); + case S32: + return Get(multi_index); + case S64: + return Get(multi_index); + case U32: + return Get(multi_index); + case U64: + return Get(multi_index); + default: + return FailedPrecondition( + "Array element type is not integral: %s", + PrimitiveType_Name(shape().element_type()).c_str()); + } +} + +size_t LiteralBase::Hash() const { + using tensorflow::Hash64; + using tensorflow::Hash64Combine; + + size_t hash_value = ShapeUtil::Hash(shape()); + + ShapeUtil::ForEachSubshape( + shape(), [&](const Shape& subshape, const ShapeIndex& index) { + if (!ShapeUtil::IsArray(subshape)) { + return; + } + + CHECK(LayoutUtil::IsDense(subshape.layout())); + hash_value = Hash64Combine( + hash_value, Hash64(static_cast(untyped_data(index)), + size_bytes(index))); + }); + + return hash_value; +} + +Status Literal::SetIntegralAsS64(tensorflow::gtl::ArraySlice multi_index, + int64 value) { + CHECK(LayoutUtil::IsDenseArray(shape())); + switch (shape().element_type()) { + case PRED: + Set(multi_index, value); + break; + case U8: + Set(multi_index, value); + break; + case S32: + Set(multi_index, value); + break; + case S64: + Set(multi_index, value); + break; + case U32: + Set(multi_index, value); + break; + case U64: + Set(multi_index, value); + break; + default: + return FailedPrecondition( + "Array element type is not integral: %s", + PrimitiveType_Name(shape().element_type()).c_str()); + } + return Status::OK(); +} + +tensorflow::gtl::ArraySlice LiteralBase::GetSparseIndex( + int64 sparse_element_number, const ShapeIndex& shape_index) const { + const Piece& p = piece(shape_index); + CHECK_GE(sparse_element_number, 0); + CHECK_LT(sparse_element_number, p.sparse_indices()->index_count()); + return p.sparse_indices()->At(sparse_element_number); +} + +void Literal::SortSparseElements(const ShapeIndex& shape_index) { + piece(shape_index).SortSparseElements(); +} + +void LiteralBase::Piece::SortSparseElements() { + switch (subshape().element_type()) { + case PRED: + SortSparseElementsInternal(); + break; + case S8: + SortSparseElementsInternal(); + break; + case U8: + SortSparseElementsInternal(); + break; + case S16: + SortSparseElementsInternal(); + break; + case U16: + SortSparseElementsInternal(); + break; + case S32: + SortSparseElementsInternal(); + break; + case U32: + SortSparseElementsInternal(); + break; + case S64: + SortSparseElementsInternal(); + break; + case U64: + SortSparseElementsInternal(); + break; + case F32: + SortSparseElementsInternal(); + break; + case F64: + SortSparseElementsInternal(); + break; + case C64: + SortSparseElementsInternal(); + break; + case F16: + SortSparseElementsInternal(); + break; + case BF16: + SortSparseElementsInternal(); + break; + default: + LOG(FATAL) << "Element type not valid for sparse array: " + << PrimitiveType_Name(subshape().element_type()); + } +} + +template +void LiteralBase::Piece::SortSparseElementsInternal() { + CHECK(LayoutUtil::IsSparseArray(subshape())); + int64 num_elements = sparse_indices()->index_count(); + auto values = data(); + CHECK_LE(num_elements, values.size()); + sparse_indices()->SortWithValues( + tensorflow::gtl::MutableArraySlice(values.data(), num_elements)); +} + +namespace { + +void ToStringHelper(const LiteralBase& literal, const ShapeIndex& shape_index, + bool print_layout, std::vector* pieces) { + const Shape& subshape = ShapeUtil::GetSubshape(literal.shape(), shape_index); + CHECK(LayoutUtil::HasLayout(literal.shape())); + CHECK(LayoutUtil::HasLayout(subshape)); + + auto shape_to_string = [print_layout](const Shape& shape) { + if (print_layout) { + return ShapeUtil::HumanStringWithLayout(shape); + } else { + return ShapeUtil::HumanString(shape); + } + }; + + // TODO(b/32894291): refactor this code to reduce code duplication. + if (ShapeUtil::IsTuple(subshape)) { + pieces->push_back(shape_to_string(subshape)); + pieces->push_back(" (\n"); + std::vector tuple_pieces; + for (int i = 0; i < ShapeUtil::TupleElementCount(subshape); ++i) { + ShapeIndex element_index = shape_index; + element_index.push_back(i); + std::vector element_pieces; + ToStringHelper(literal, element_index, print_layout, &element_pieces); + tuple_pieces.push_back(tensorflow::str_util::Join(element_pieces, "")); + } + pieces->push_back(tensorflow::str_util::Join(tuple_pieces, ",\n")); + pieces->push_back("\n)"); + return; + } + + if (ShapeUtil::IsToken(subshape)) { + pieces->push_back("token"); + return; + } + + if (LayoutUtil::IsSparseArray(subshape)) { + pieces->push_back(shape_to_string(subshape)); + pieces->push_back("{"); + int64 rank = ShapeUtil::Rank(subshape); + int64 num_elements = literal.sparse_element_count(); + for (int64 i = 0; i < num_elements; ++i) { + if (i > 0) { + pieces->push_back(", "); + } + if (rank == 1) { + pieces->push_back(StrCat(literal.GetSparseIndex(i)[0])); + pieces->push_back(": "); + } else { + pieces->push_back("["); + pieces->push_back( + tensorflow::str_util::Join(literal.GetSparseIndex(i), ", ")); + pieces->push_back("]: "); + } + pieces->push_back(literal.GetSparseElementAsString(i)); + } + pieces->push_back("}"); + return; + } + + CHECK(LayoutUtil::IsDenseArray(subshape)); + + auto element_to_string = + [&](tensorflow::gtl::ArraySlice indices) -> string { + PrimitiveType element_type = subshape.element_type(); + if (element_type == PRED) { + // We display predicates in a densely packed form. + return literal.Get(indices, shape_index) ? "1" : "0"; + } + return ((!indices.empty() && indices.back() > 0) ? ", " : "") + + literal.GetAsString(indices, shape_index); + }; + + if (ShapeUtil::Rank(subshape) == 0) { + pieces->push_back(literal.GetAsString({}, shape_index)); + } else if (ShapeUtil::Rank(subshape) == 1) { + pieces->push_back("{"); + for (int64 i0 = 0; i0 < subshape.dimensions(0); ++i0) { + pieces->push_back(element_to_string({i0})); + } + pieces->push_back("}"); + } else if (ShapeUtil::Rank(subshape) == 2) { + pieces->push_back(shape_to_string(subshape)); + pieces->push_back(" {\n"); + for (int64 i0 = 0; i0 < subshape.dimensions(0); ++i0) { + pieces->push_back(" { "); + for (int64 i1 = 0; i1 < subshape.dimensions(1); ++i1) { + pieces->push_back(element_to_string({i0, i1})); + } + pieces->push_back(" "); + pieces->push_back(i0 == subshape.dimensions(0) - 1 ? "}\n" : "},\n"); + } + pieces->push_back("}"); + } else if (ShapeUtil::Rank(subshape) == 3) { + pieces->push_back(shape_to_string(subshape)); + pieces->push_back(" {\n"); + for (int64 i0 = 0; i0 < subshape.dimensions(0); ++i0) { + pieces->push_back(i0 > 0 ? ",\n{" : "{"); + for (int64 i1 = 0; i1 < subshape.dimensions(1); ++i1) { + pieces->push_back(i1 > 0 ? ",\n { " : " { "); + for (int64 i2 = 0; i2 < subshape.dimensions(2); ++i2) { + pieces->push_back(element_to_string({i0, i1, i2})); + } + pieces->push_back(" }"); + } + pieces->push_back(" }"); + } + pieces->push_back("\n}"); + } else if (ShapeUtil::Rank(subshape) == 4) { + pieces->push_back(shape_to_string(subshape)); + pieces->push_back(" {\n"); + for (int64 i0 = 0; i0 < subshape.dimensions(0); ++i0) { + pieces->push_back(Printf(" { /*i0=%lld*/\n", i0)); + for (int64 i1 = 0; i1 < subshape.dimensions(1); ++i1) { + pieces->push_back(Printf(" { /*i1=%lld*/\n", i1)); + for (int64 i2 = 0; i2 < subshape.dimensions(2); ++i2) { + pieces->push_back(" {"); + for (int64 i3 = 0; i3 < subshape.dimensions(3); ++i3) { + pieces->push_back(element_to_string({i0, i1, i2, i3})); + } + pieces->push_back(i2 == subshape.dimensions(2) - 1 ? "}\n" : "},\n"); + } + pieces->push_back(i1 == subshape.dimensions(1) - 1 ? " }\n" + : " },\n"); + } + pieces->push_back(i0 == subshape.dimensions(0) - 1 ? " }\n" : " },\n"); + } + pieces->push_back("}"); + } else if (ShapeUtil::Rank(subshape) == 5) { + pieces->push_back(shape_to_string(subshape)); + pieces->push_back(" {\n"); + for (int64 i0 = 0; i0 < subshape.dimensions(0); ++i0) { + pieces->push_back(Printf(" { /*i0=%lld*/\n", i0)); + for (int64 i1 = 0; i1 < subshape.dimensions(1); ++i1) { + pieces->push_back(Printf(" { /*i1=%lld*/\n", i1)); + for (int64 i2 = 0; i2 < subshape.dimensions(2); ++i2) { + pieces->push_back(Printf(" { /*i2=%lld*/\n", i2)); + for (int64 i3 = 0; i3 < subshape.dimensions(3); ++i3) { + pieces->push_back(" {"); + for (int64 i4 = 0; i4 < subshape.dimensions(4); ++i4) { + pieces->push_back(element_to_string({i0, i1, i2, i3, i4})); + } + pieces->push_back(i3 == subshape.dimensions(3) - 1 ? "}\n" + : "},\n"); + } + pieces->push_back(i2 == subshape.dimensions(2) - 1 ? " }\n" + : " },\n"); + } + pieces->push_back(i1 == subshape.dimensions(1) - 1 ? " }\n" + : " },\n"); + } + pieces->push_back(i0 == subshape.dimensions(0) - 1 ? " }\n" : " },\n"); + } + pieces->push_back("}"); + } else { + pieces->push_back(shape_to_string(subshape)); + pieces->push_back(" {"); + literal.EachCellAsString( + [&](tensorflow::gtl::ArraySlice indices, const string& value) { + pieces->push_back(" "); + pieces->push_back(value); + }); + pieces->push_back("}"); + } +} + +} // namespace + +int64 LiteralBase::sparse_element_count() const { + CHECK(LayoutUtil::IsSparseArray(shape())); + return sparse_indices()->index_count(); +} + +string LiteralBase::ToString(bool print_layout) const { + std::vector pieces; + CHECK(LayoutUtil::HasLayout(this->shape())); + ToStringHelper(*this, {}, print_layout, &pieces); + return tensorflow::str_util::Join(pieces, ""); +} + +void LiteralBase::EachCellAsString( + const std::function indices, + const string& value)>& per_cell) const { + if (ShapeUtil::IsZeroElementArray(shape())) { + return; + } + std::vector indices = IndexUtil::LinearIndexToMultidimensionalIndex( + shape(), /*linear_index=*/0); + do { + per_cell(indices, GetAsString(indices)); + } while (IndexUtil::BumpIndices(shape(), &indices)); +} + +namespace { +template +std::unique_ptr ConvertBetweenNativeTypesWithConverter( + const LiteralBase& src_literal, const ConverterType& converter) { + CHECK(ShapeUtil::IsArray(src_literal.shape())); + auto result_literal = MakeUnique(ShapeUtil::ChangeElementType( + src_literal.shape(), + primitive_util::NativeToPrimitiveType())); + auto src_data = src_literal.data(); + auto dest_data = result_literal->template data(); + int64 num_elements = src_literal.element_count(); + + for (int64 i = 0; i < num_elements; ++i) { + dest_data[i] = converter(src_data[i]); + } + return result_literal; +} + +template +std::unique_ptr ConvertBetweenNativeTypes( + const LiteralBase& src_literal) { + auto converter = [](NativeSrcT src) { return static_cast(src); }; + return ConvertBetweenNativeTypesWithConverter( + src_literal, converter); +} + +template +typename std::enable_if<(sizeof(NativeSrcT) == sizeof(NativeDestT)), + std::unique_ptr>::type +BitcastBetweenNativeTypes(const LiteralBase& src_literal) { + auto converter = [](NativeSrcT src) { + return tensorflow::bit_cast(src); + }; + return ConvertBetweenNativeTypesWithConverter( + src_literal, converter); +} + +// This template specialization is here to make the compiler happy. bit_cast has +// a static check that the types are the same size. This specialization should +// never be used because the source and destination types are checked for +// identical sizes higher up. +template +typename std::enable_if<(sizeof(NativeSrcT) != sizeof(NativeDestT)), + std::unique_ptr>::type +BitcastBetweenNativeTypes(const LiteralBase& src_literal) { + LOG(FATAL) << "Invalid bitcast between types of different sizes."; +} + +template +std::unique_ptr ConvertToC64(const LiteralBase& src_literal) { + CHECK(ShapeUtil::IsArray(src_literal.shape())); + auto result_literal = MakeUnique( + ShapeUtil::ChangeElementType(src_literal.shape(), C64)); + using NativeSrcT = + typename primitive_util::PrimitiveTypeToNative::type; + tensorflow::gtl::ArraySlice src_data = + src_literal.data(); + tensorflow::gtl::MutableArraySlice dest_data = + result_literal->data(); + int64 num_elements = src_literal.element_count(); + for (int64 i = 0; i < num_elements; ++i) { + dest_data[i] = complex64(static_cast(src_data[i]), 0); + } + return result_literal; +} + +template +std::unique_ptr ConvertIfTypesMatch(const LiteralBase& src_literal, + bool bitcast) { + CHECK_EQ(primitive_src_type, src_literal.shape().element_type()); + if (bitcast) { + return BitcastBetweenNativeTypes< + typename primitive_util::PrimitiveTypeToNative< + primitive_src_type>::type, + typename primitive_util::PrimitiveTypeToNative< + primitive_dest_type>::type>(src_literal); + } else { + return ConvertBetweenNativeTypes< + typename primitive_util::PrimitiveTypeToNative< + primitive_src_type>::type, + typename primitive_util::PrimitiveTypeToNative< + primitive_dest_type>::type>(src_literal); + } +} + +template +StatusOr> ConvertIfDestTypeMatches( + const LiteralBase& src_literal, PrimitiveType primitive_dest_type, + bool bitcast) { + switch (primitive_dest_type) { +#define CONVERT_IF_TYPES_MATCH(type) \ + case (type): \ + return ConvertIfTypesMatch(src_literal, \ + bitcast); + CONVERT_IF_TYPES_MATCH(PRED) + CONVERT_IF_TYPES_MATCH(S8) + CONVERT_IF_TYPES_MATCH(S32) + CONVERT_IF_TYPES_MATCH(S64) + CONVERT_IF_TYPES_MATCH(U8) + CONVERT_IF_TYPES_MATCH(U32) + CONVERT_IF_TYPES_MATCH(U64) + CONVERT_IF_TYPES_MATCH(F16) + CONVERT_IF_TYPES_MATCH(F32) + CONVERT_IF_TYPES_MATCH(F64) + CONVERT_IF_TYPES_MATCH(BF16) +#undef CONVERT_IF_TYPES_MATCH + case C64: + if (!bitcast) { + return ConvertToC64(src_literal); + } + break; + // Other types are not yet supported. + default: + break; + } + return Unimplemented( + "Converting from type %s to type %s is not implemented.", + PrimitiveType_Name(src_literal.shape().element_type()).c_str(), + PrimitiveType_Name(primitive_dest_type).c_str()); +} + +StatusOr> ConvertSwitch( + const LiteralBase& literal, PrimitiveType primitive_dest_type, + bool bitcast) { + TF_RET_CHECK(ShapeUtil::IsArray(literal.shape())); + if (literal.shape().element_type() == primitive_dest_type) { + return literal.CloneToUnique(); + } + switch (literal.shape().element_type()) { +#define CONVERT_IF_DEST_TYPE_MATCHES(type) \ + case (type): \ + return ConvertIfDestTypeMatches<(type)>(literal, primitive_dest_type, \ + bitcast); + CONVERT_IF_DEST_TYPE_MATCHES(PRED) + CONVERT_IF_DEST_TYPE_MATCHES(S8) + CONVERT_IF_DEST_TYPE_MATCHES(S32) + CONVERT_IF_DEST_TYPE_MATCHES(S64) + CONVERT_IF_DEST_TYPE_MATCHES(U8) + CONVERT_IF_DEST_TYPE_MATCHES(U32) + CONVERT_IF_DEST_TYPE_MATCHES(U64) + CONVERT_IF_DEST_TYPE_MATCHES(F16) + CONVERT_IF_DEST_TYPE_MATCHES(F32) + CONVERT_IF_DEST_TYPE_MATCHES(F64) + CONVERT_IF_DEST_TYPE_MATCHES(BF16) +#undef CONVERT_IF_DEST_TYPE_MATCHES + // Other types are not yet supported. + default: + return Unimplemented( + "%s from type %s to type %s is not implemented.", + (bitcast ? "Bitcast converting" : "Converting"), + PrimitiveType_Name(literal.shape().element_type()).c_str(), + PrimitiveType_Name(primitive_dest_type).c_str()); + } +} + +} // namespace + +StatusOr> LiteralBase::Convert( + PrimitiveType primitive_dest_type) const { + return ConvertSwitch(*this, primitive_dest_type, /*bitcast=*/false); +} + +StatusOr> LiteralBase::BitcastConvert( + PrimitiveType primitive_dest_type) const { + if (primitive_util::BitWidth(shape().element_type()) != + primitive_util::BitWidth(primitive_dest_type)) { + return InvalidArgument( + "Cannot bitcast convert from %s to %s, bit widths are different: %d != " + "%d", + PrimitiveType_Name(shape().element_type()).c_str(), + PrimitiveType_Name(primitive_dest_type).c_str(), + primitive_util::BitWidth(shape().element_type()), + primitive_util::BitWidth(primitive_dest_type)); + } + return ConvertSwitch(*this, primitive_dest_type, /*bitcast=*/true); +} + +StatusOr> LiteralBase::ConvertToShape( + const Shape& dest_shape, bool round_f32_to_bf16) const { + if (!ShapeUtil::IsTuple(dest_shape)) { + if (round_f32_to_bf16 && shape().element_type() == F32 && + dest_shape.element_type() == BF16) { + auto converter = [](float src) { + return tensorflow::bfloat16::round_to_bfloat16(src); + }; + return ConvertBetweenNativeTypesWithConverter(*this, + converter); + } + return Convert(dest_shape.element_type()); + } + std::vector elements; + for (int i = 0; i < ShapeUtil::TupleElementCount(shape()); ++i) { + auto element = LiteralSlice(*this, {i}); + TF_ASSIGN_OR_RETURN( + auto new_element, + element.ConvertToShape(ShapeUtil::GetSubshape(dest_shape, {i}))); + elements.push_back(std::move(*new_element)); + } + auto converted = MakeUnique(); + *converted = Literal::MoveIntoTuple(&elements); + return std::move(converted); +} + +/* static */ Literal Literal::MoveIntoTuple( + tensorflow::gtl::MutableArraySlice elements) { + std::vector element_shapes; + for (const Literal& element : elements) { + element_shapes.push_back(element.shape()); + } + Literal literal(ShapeUtil::MakeTupleShape(element_shapes), + /*allocate_arrays=*/false); + for (int i = 0; i < elements.size(); ++i) { + TF_CHECK_OK( + literal.MoveFrom(std::move(elements[i]), /*dest_shape_index=*/{i})); + } + return literal; +} + +template +bool LiteralBase::Piece::EqualElementsInternal( + const LiteralBase::Piece& other, std::vector* multi_index) const { + if (multi_index->size() == ShapeUtil::Rank(subshape())) { + return (Get(*multi_index) == other.Get(*multi_index)); + } + for (int64 i = 0; i < subshape().dimensions(multi_index->size()); ++i) { + multi_index->push_back(i); + if (!EqualElementsInternal(other, multi_index)) { + return false; + } + multi_index->pop_back(); + } + return true; +} + +bool LiteralBase::Piece::EqualElements(const LiteralBase::Piece& other) const { + DCHECK(ShapeUtil::Compatible(subshape(), other.subshape())); + + std::vector multi_index; + switch (subshape().element_type()) { + case PRED: + return EqualElementsInternal(other, &multi_index); + case U8: + return EqualElementsInternal(other, &multi_index); + case S32: + return EqualElementsInternal(other, &multi_index); + case S64: + return EqualElementsInternal(other, &multi_index); + case U32: + return EqualElementsInternal(other, &multi_index); + case U64: + return EqualElementsInternal(other, &multi_index); + case F32: + return EqualElementsInternal(other, &multi_index); + case F64: + return EqualElementsInternal(other, &multi_index); + case F16: + return EqualElementsInternal(other, &multi_index); + case BF16: + return EqualElementsInternal(other, &multi_index); + case C64: + return EqualElementsInternal(other, &multi_index); + default: + LOG(FATAL) << "Unimplemented: LiteralBase::Piece::EqualElements for type " + << PrimitiveType_Name(subshape().element_type()); + } +} + +bool LiteralBase::operator==(const LiteralBase& other) const { + if (!ShapeUtil::Compatible(shape(), other.shape())) { + return false; + } + + return root_piece().ForEachSubpieceWithBool( + [&](const ShapeIndex& index, const Piece& piece) { + if (!ShapeUtil::IsArray(piece.subshape())) { + return true; + } + + const Piece& other_piece = other.piece(index); + if (!piece.EqualElements(other_piece)) { + return false; + } + return true; + }); +} + +namespace { + +template +static bool AllElementsEqualValue(tensorflow::gtl::ArraySlice data, + NativeT value) { + for (int64 i = 0; i < data.size(); ++i) { + if (data[i] != value) { + return false; + } + } + return true; +} + +} // namespace + +bool LiteralBase::IsAll(int8 value) const { + return root_piece().ForEachSubpieceWithBool([&](const ShapeIndex& index, + const Piece& piece) { + if (!ShapeUtil::IsArray(piece.subshape())) { + return true; + } + + auto piece_is_all = [&]() { + switch (shape().element_type()) { + case U8: + if (value >= 0) { + return AllElementsEqualValue(piece.data(), value); + } + return false; + case U32: + if (value >= 0) { + return AllElementsEqualValue(piece.data(), value); + } + return false; + case U64: + if (value >= 0) { + return AllElementsEqualValue(piece.data(), value); + } + return false; + case S8: + return AllElementsEqualValue(piece.data(), value); + case S32: + return AllElementsEqualValue(piece.data(), value); + case S64: + return AllElementsEqualValue(piece.data(), value); + case F32: + return AllElementsEqualValue(piece.data(), value); + case F64: + return AllElementsEqualValue(piece.data(), value); + case F16: + return AllElementsEqualValue(piece.data(), + static_cast(value)); + case BF16: + return AllElementsEqualValue(piece.data(), + static_cast(value)); + case PRED: + if (value == 0) { + return AllElementsEqualValue(piece.data(), false); + } + if (value == 1) { + return AllElementsEqualValue(piece.data(), true); + } + return false; + default: + return false; + } + return false; + }; + + if (!piece_is_all()) { + return false; + } + return true; + }); +} + +bool LiteralBase::IsAllFloat(float value) const { + return root_piece().ForEachSubpieceWithBool( + [&](const ShapeIndex& index, const Piece& piece) { + if (!ShapeUtil::IsArray(piece.subshape())) { + return true; + } + + auto piece_is_all = [&]() { + switch (shape().element_type()) { + case F32: + return AllElementsEqualValue(piece.data(), value); + case F64: + return AllElementsEqualValue(piece.data(), value); + case F16: + return AllElementsEqualValue(piece.data(), + static_cast(value)); + case BF16: + return AllElementsEqualValue( + piece.data(), static_cast(value)); + default: + return false; + } + }; + if (!piece_is_all()) { + return false; + } + return true; + }); +} + +bool LiteralBase::IsAllComplex(complex64 value) const { + switch (shape().element_type()) { + case C64: + return AllElementsEqualValue(root_piece().data(), + value); + default: + return false; + } +} + +bool LiteralBase::IsAllFirst() const { + return root_piece().ForEachSubpieceWithBool( + [&](const ShapeIndex& index, const Piece& piece) { + if (!ShapeUtil::IsArray(piece.subshape())) { + return true; + } + + // Empty shapes are not all the first element since there is no first + // element. + if (ShapeUtil::IsZeroElementArray(piece.subshape())) { + return false; + } + auto piece_is_all = [&]() { + switch (piece.subshape().element_type()) { + case PRED: { + auto data = piece.data(); + return AllElementsEqualValue(data, data[0]); + } + // 8 bit types + case S8: { + auto data = piece.data(); + return AllElementsEqualValue(data, data[0]); + } + case U8: { + auto data = piece.data(); + return AllElementsEqualValue(data, data[0]); + } + // 16 bit types + case BF16: { + auto data = piece.data(); + return AllElementsEqualValue(data, data[0]); + } + case F16: { + auto data = piece.data(); + return AllElementsEqualValue(data, data[0]); + } + case S16: { + auto data = piece.data(); + return AllElementsEqualValue(data, data[0]); + } + case U16: { + auto data = piece.data(); + return AllElementsEqualValue(data, data[0]); + } + // 32 bit types + case F32: { + auto data = piece.data(); + return AllElementsEqualValue(data, data[0]); + } + case U32: { + auto data = piece.data(); + return AllElementsEqualValue(data, data[0]); + } + case S32: { + auto data = piece.data(); + return AllElementsEqualValue(data, data[0]); + } + // 64 bit types + case C64: { + auto data = piece.data(); + return AllElementsEqualValue(data, data[0]); + } + case F64: { + auto data = piece.data(); + return AllElementsEqualValue(data, data[0]); + } + case S64: { + auto data = piece.data(); + return AllElementsEqualValue(data, data[0]); + } + case U64: { + auto data = piece.data(); + return AllElementsEqualValue(data, data[0]); + } + default: + return false; + } + }; + + if (!piece_is_all()) { + return false; + } + return true; + }); +} + +bool LiteralBase::IsZero(tensorflow::gtl::ArraySlice indices) const { + CHECK(ShapeUtil::IsArray(shape())); + switch (shape().element_type()) { + case U8: + return Get(indices) == 0; + case U32: + return Get(indices) == 0; + case U64: + return Get(indices) == 0; + case S8: + return Get(indices) == 0; + case S32: + return Get(indices) == 0; + case S64: + return Get(indices) == 0; + case F32: + return Get(indices) == 0.0f; + case F64: + return Get(indices) == 0.0; + case C64: + return Get(indices) == complex64(0.0f, 0.0f); + case F16: + return Get(indices) == static_cast(0.0f); + case BF16: + return Get(indices) == static_cast(0.0f); + case PRED: + return Get(indices) == false; + default: + LOG(FATAL) << "Input literal must be an array."; + } +} + +namespace { + +template +void CopyToRepeatedField(RepeatedFieldT* dest, + const tensorflow::gtl::ArraySlice src) { + *dest = RepeatedFieldT(src.begin(), src.end()); +} + +} // namespace + +void LiteralBase::Piece::WriteToProto(LiteralProto* proto) const { + *proto->mutable_shape() = subshape(); + switch (subshape().element_type()) { + case PRED: + CopyToRepeatedField(proto->mutable_preds(), data()); + break; + case U8: + proto->set_u8s(static_cast(data().data()), + element_count()); + break; + case U32: + CopyToRepeatedField(proto->mutable_u32s(), data()); + break; + case U64: + CopyToRepeatedField(proto->mutable_u64s(), data()); + break; + case S32: + CopyToRepeatedField(proto->mutable_s32s(), data()); + break; + case S64: + CopyToRepeatedField(proto->mutable_s64s(), data()); + break; + case F16: + *proto->mutable_f16s() = string( + reinterpret_cast(data().data()), size_bytes()); + if (!kLittleEndian) { + ConvertEndianShort(proto->mutable_f16s()); + } + break; + case BF16: + *proto->mutable_bf16s() = string( + reinterpret_cast(data().data()), size_bytes()); + if (!kLittleEndian) { + ConvertEndianShort(proto->mutable_bf16s()); + } + break; + case F32: + CopyToRepeatedField(proto->mutable_f32s(), data()); + break; + case F64: + CopyToRepeatedField(proto->mutable_f64s(), data()); + break; + case C64: + for (complex64 value : data()) { + proto->add_c64s(value.real()); + proto->add_c64s(value.imag()); + } + break; + case TUPLE: + case TOKEN: + // Nothing to do but assign the shape which is done above. + return; + default: + LOG(FATAL) << "Unhandled primitive type " << subshape().element_type(); + } +} + +const void* LiteralBase::Piece::untyped_data() const { + CHECK(ShapeUtil::IsArray(subshape())) << ShapeUtil::HumanString(subshape()); + return buffer(); +} + +void* LiteralBase::Piece::untyped_data() { + CHECK(ShapeUtil::IsArray(subshape())) << ShapeUtil::HumanString(subshape()); + return buffer(); +} + +namespace { + +template +Status CopyFromRepeatedField(tensorflow::gtl::MutableArraySlice dest, + const RepeatedFieldT& src) { + if (dest.size() != src.size()) { + return InvalidArgument( + "Expected %lu elements in LiteralProto repeated field, has %d", + dest.size(), src.size()); + } + std::copy(src.begin(), src.end(), dest.begin()); + return Status::OK(); +} + +} // namespace + +Status LiteralBase::Piece::CopyFromProto(const LiteralProto& proto) { + // These conditions should have been checked in Literal::CreateFromProto. + TF_RET_CHECK(proto.has_shape()); + TF_RET_CHECK(LayoutUtil::HasLayout(proto.shape())); + TF_RET_CHECK(ShapeUtil::Equal(proto.shape(), subshape())); + + switch (subshape().element_type()) { + case PRED: + TF_RETURN_IF_ERROR(CopyFromRepeatedField(data(), proto.preds())); + break; + case U8: { + auto u8_data = data(); + TF_RET_CHECK(proto.u8s().size() == u8_data.size()); + std::copy(proto.u8s().begin(), proto.u8s().end(), u8_data.begin()); + } break; + case S32: + TF_RETURN_IF_ERROR(CopyFromRepeatedField(data(), proto.s32s())); + break; + case S64: + TF_RETURN_IF_ERROR(CopyFromRepeatedField(data(), proto.s64s())); + break; + case U32: + TF_RETURN_IF_ERROR(CopyFromRepeatedField(data(), proto.u32s())); + break; + case U64: + TF_RETURN_IF_ERROR(CopyFromRepeatedField(data(), proto.u64s())); + break; + case F16: { + const string& s(proto.f16s()); + TF_RET_CHECK(data().size() * sizeof(half) == s.size()); + memcpy(untyped_data(), s.data(), s.size()); + if (!kLittleEndian) { + ConvertEndianShort(reinterpret_cast(untyped_data()), s.size()); + } + } break; + + case BF16: { + const string& s(proto.bf16s()); + TF_RET_CHECK(data().size() * sizeof(bfloat16) == s.size()); + memcpy(untyped_data(), s.data(), s.size()); + if (!kLittleEndian) { + ConvertEndianShort(reinterpret_cast(untyped_data()), s.size()); + } + } break; + case F32: + TF_RETURN_IF_ERROR(CopyFromRepeatedField(data(), proto.f32s())); + break; + case F64: + TF_RETURN_IF_ERROR(CopyFromRepeatedField(data(), proto.f64s())); + break; + case C64: { + auto complex_data = data(); + TF_RET_CHECK(proto.c64s_size() == complex_data.size() * 2); + for (int64 i = 0; i < complex_data.size(); ++i) { + complex_data[i] = complex64{proto.c64s(i * 2), proto.c64s(i * 2 + 1)}; + } + } break; + case TUPLE: + LOG(FATAL) << "Should not be called on tuple shapes: " + << ShapeUtil::HumanString(subshape()); + break; + default: + LOG(FATAL) << "Unhandled primitive type " << subshape().element_type(); + } + return Status::OK(); +} + +LiteralProto LiteralBase::ToProto() const { + LiteralProto proto; + root_piece().ForEachSubpiece( + [&](const ShapeIndex& index, const Piece& piece) { + LiteralProto* proto_piece = &proto; + for (int64 i : index) { + while (proto_piece->tuple_literals_size() <= i) { + proto_piece->add_tuple_literals(); + } + proto_piece = proto_piece->mutable_tuple_literals(i); + } + piece.WriteToProto(proto_piece); + }); + + if (LayoutUtil::IsSparseArray(shape())) { + CopyToRepeatedField(proto.mutable_sparse_indices(), + sparse_indices()->data()); + } + + return proto; +} + +const void* LiteralBase::untyped_data(const ShapeIndex& shape_index) const { + return piece(shape_index).untyped_data(); +} + +void* Literal::untyped_data(const ShapeIndex& shape_index) { + return piece(shape_index).untyped_data(); +} + +int64 LiteralBase::size_bytes(const ShapeIndex& shape_index) const { + return piece(shape_index).size_bytes(); +} + +string LiteralBase::GetR1U8AsString() const { + CHECK(ShapeUtil::IsArray(shape())); + CHECK_EQ(ShapeUtil::Rank(shape()), 1); + CHECK_EQ(shape().element_type(), U8); + return string(tensorflow::bit_cast(data().data()), + ShapeUtil::ElementsIn(shape())); +} + +void BorrowingLiteral::BuildPieceSubtree(const Shape& shape, Piece* piece) { + CHECK(ShapeUtil::IsTuple(shape)); + for (int i = 0; i < ShapeUtil::TupleElementCount(shape); ++i) { + const Shape& subshape = shape.tuple_shapes(i); + + auto child_piece = Piece(); + child_piece.set_subshape(&subshape); + + if (ShapeUtil::IsTuple(subshape)) { + BuildPieceSubtree(subshape, &child_piece); + } + + piece->emplace_back(std::move(child_piece)); + } +} + +LiteralSlice::LiteralSlice(const LiteralBase& literal) + : LiteralBase(), root_piece_(&literal.root_piece()) {} + +LiteralSlice::LiteralSlice(const LiteralBase& literal, + const ShapeIndex& view_root) + : LiteralBase(), root_piece_(&literal.piece(view_root)) {} + +BorrowingLiteral::BorrowingLiteral(const char* src_buf_ptr, const Shape& shape) + : LiteralBase(), shape_(MakeUnique(shape)) { + CHECK(ShapeUtil::IsArray(*shape_)); + CHECK(LayoutUtil::HasLayout(*shape_)); + + root_piece_ = Piece(); + root_piece_.set_buffer(const_cast(src_buf_ptr)); + root_piece_.set_subshape(shape_.get()); +} + +BorrowingLiteral::BorrowingLiteral( + tensorflow::gtl::ArraySlice src_buf_ptrs, const Shape& shape) + : LiteralBase(), shape_(MakeUnique(shape)) { + CHECK(ShapeUtil::IsTuple(*shape_)); + CHECK(!ShapeUtil::IsNestedTuple(*shape_)); + CHECK_EQ(src_buf_ptrs.size(), ShapeUtil::TupleElementCount(*shape_)); + root_piece_ = Piece(); + root_piece_.set_subshape(shape_.get()); + BuildPieceSubtree(*shape_, &root_piece_); + + for (int i = 0; i < src_buf_ptrs.size(); ++i) { + const auto& src_shape = shape_->tuple_shapes(i); + CHECK(ShapeUtil::IsArray(src_shape)); + root_piece_.child(i).set_buffer(const_cast(src_buf_ptrs[i])); + } +} + +} // namespace xla diff --git a/tensorflow/compiler/xla/literal.h b/tensorflow/compiler/xla/literal.h new file mode 100644 index 0000000000000000000000000000000000000000..dd67dfa8d4a556aea179bc47abfdc9a9c8872c45 --- /dev/null +++ b/tensorflow/compiler/xla/literal.h @@ -0,0 +1,1152 @@ +/* Copyright 2017 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#ifndef TENSORFLOW_COMPILER_XLA_LITERAL_H_ +#define TENSORFLOW_COMPILER_XLA_LITERAL_H_ + +#include +#include +#include +#include +#include +#include +#include +#include + +#include "tensorflow/compiler/xla/array2d.h" +#include "tensorflow/compiler/xla/array3d.h" +#include "tensorflow/compiler/xla/array4d.h" +#include "tensorflow/compiler/xla/index_util.h" +#include "tensorflow/compiler/xla/layout_util.h" +#include "tensorflow/compiler/xla/primitive_util.h" +#include "tensorflow/compiler/xla/ptr_util.h" +#include "tensorflow/compiler/xla/shape_util.h" +#include "tensorflow/compiler/xla/sparse_index_array.h" +#include "tensorflow/compiler/xla/status_macros.h" +#include "tensorflow/compiler/xla/types.h" +#include "tensorflow/compiler/xla/util.h" +#include "tensorflow/compiler/xla/xla_data.pb.h" +#include "tensorflow/core/lib/core/bitmap.h" +#include "tensorflow/core/lib/core/status.h" +#include "tensorflow/core/lib/core/stringpiece.h" +#include "tensorflow/core/lib/gtl/array_slice.h" +#include "tensorflow/core/platform/logging.h" +#include "tensorflow/core/platform/macros.h" +#include "tensorflow/core/platform/protobuf.h" +#include "tensorflow/core/platform/types.h" + +namespace xla { + +// Forward declare Literal and LiteralSlice class to be used by the creation +// methods in the base class. +class Literal; +class LiteralSlice; + +// Abstract base class for literals. +class LiteralBase { + public: + virtual ~LiteralBase() = 0; + + // Literals are equal if they have compatible shapes and the same data + // values. Layout is not compared. + bool operator==(const LiteralBase& other) const; + bool operator!=(const LiteralBase& other) const { return !(*this == other); } + + // Returns the shape of the literal. + const Shape& shape() const { return root_piece().subshape(); } + + // Serialize to proto. + LiteralProto ToProto() const; + + // Returns an ArraySlice of the array for this literal for the given NativeT + // (e.g., float). CHECKs if the subshape of the literal at the given + // ShapeIndex is not array. See primitive_util.h for the mapping from XLA type + // to native type. + template + tensorflow::gtl::ArraySlice data( + const ShapeIndex& shape_index = {}) const; + + // Returns a const pointer to the sparse index array. Returns nullptr if the + // literal is not a sparse array. + const SparseIndexArray* sparse_indices( + const ShapeIndex& shape_index = {}) const; + + // Returns a const pointer to (or size of) the underlying buffer holding the + // array at the given shape index. CHECKs if the subshape of the literal at + // the given ShapeIndex is not array. + const void* untyped_data(const ShapeIndex& shape_index = {}) const; + int64 size_bytes(const ShapeIndex& shape_index = {}) const; + + // Returns this literal's data as a string. This literal must be a rank-1 U8 + // array. + string GetR1U8AsString() const; + + // Returns a string representation of the literal value. + // Warning: this function can take minutes for multi-million element Literals. + string ToString(bool print_layout = false) const; + + // Gets an element in the literal at the given index. The multi_index is + // CHECKed against the dimension sizes. + template + NativeT Get(tensorflow::gtl::ArraySlice multi_index, + const ShapeIndex& shape_index) const; + // Overloads of Get for array literals. CHECKs if the literal is not + // array-shaped and dense. + template + NativeT Get(tensorflow::gtl::ArraySlice multi_index) const; + + // Returns the element value at index (0, ..., 0), however many zeroes are + // required for that index. + template + NativeT GetFirstElement() const; + + // As Get(), but determines the correct type and converts the value + // into text. + string GetAsString(tensorflow::gtl::ArraySlice multi_index, + const ShapeIndex& shape_index = {}) const; + // As GetSparseElement(), but determines the correct type and converts the + // value into text. + string GetSparseElementAsString(int64 sparse_element_number, + const ShapeIndex& shape_index = {}) const; + // As Get(), but determines the correct type and converts the value into + // int64. This literal must be an array. + StatusOr GetIntegralAsS64( + tensorflow::gtl::ArraySlice multi_index) const; + + // Returns the multi-index of the element in a sparse literal at the given + // sparse element number. The sparse element number is the position with in + // the sparse array's list of (index, value) pairs, and is checked against the + // total number of (index, value) pairs in the sparse array. + tensorflow::gtl::ArraySlice GetSparseIndex( + int64 sparse_element_number, const ShapeIndex& shape_index = {}) const; + + // Returns the value of the element in a sparse literal at the given sparse + // element number. The sparse element number is the position with in the + // sparse array's list of (index, value) pairs, and is checked against the + // total number of (index, value) pairs in the sparse array. + template + NativeT GetSparseElement(int64 sparse_element_number, + const ShapeIndex& shape_index = {}) const; + + // Invokes the "per cell" callback for each element in the provided + // literal with the element's indices and a string representation of + // the element's value. + // + // This function is useful if you want a polymorphic representation + // of the tensor's elements (turning it to a string for something + // like representation in a protobuf). + // + // This literal must have a dense layout. + void EachCellAsString( + const std::function indices, + const string& value)>& per_cell) const; + template + void EachCell(std::function indices, + NativeT value)> + per_cell) const; + + // Returns whether every element in this literal is equal to value. + // + // value is an int8 because we expect this to be called with small + // compile-time constants (0, -1, etc.) and so that whatever value you pass + // can be represented exactly by floating-point types as small as 16 bits. + // + // If value doesn't fit in this literal's type, returns false. Values of 1/0 + // are considered equal to true/false; other values are not considered equal + // to true. Also if this literal is not array-shaped false is returned. + bool IsAll(int8 value) const; + + // Like IsAll(const Literal&, int8), except we check whether the literal is + // equal to a particular floating-point number. + // + // If the literal is not a floating-point value, this always returns false. + // + // This casts value to the type of literal, then compares using ==. The usual + // admonishments about floating-point equality checks apply. We expect you to + // use this to check for values that can be expressed precisely as a float, + // e.g. -0.5. Also if this literal is not array-shaped false is returned. + bool IsAllFloat(float value) const; + + // Like IsAll(const Literal&, int8), except we check whether the literal is + // equal to a particular complex number. + // + // If the literal is not a complex value, this always returns false. + // + // This casts value to the type of literal, then compares using ==. The usual + // admonishments about floating-point equality checks apply. We expect you to + // use this to check for complex values that can be expressed precisely as + // float pairs e.g. (-0.5, 1.0). + // + // This literal must have a dense layout. + bool IsAllComplex(complex64 value) const; + + // Literal consists entirely of the first element of the literal. + bool IsAllFirst() const; + + // Returns whether this literal is zero at the specified index. This literal + // must be an array with a dense layout. + bool IsZero(tensorflow::gtl::ArraySlice indices) const; + + // Returns the count of the elements in the array at the given shape index in + // this literal. + int64 element_count(const ShapeIndex& index = {}) const { + return ShapeUtil::ElementsIn(ShapeUtil::GetSubshape(shape(), index)); + } + + // Returns the count of the elements in the sparse array at the given shape + // index in this literal, which will be no larger than + // LayoutUtil::MaxSparseElements(SetSubshape(shape(), index).layout()). + int64 sparse_element_count() const; + + // Compute a hash for this literal. This literal must not be a sparse tensor + // or a tuple containing a sparse tensor. + size_t Hash() const; + + // Converts this literal to the given shape. Returns an error is the + // conversion is not possible. + // + // round_f32_to_bf16: if true, converting F32 elements to BF16 uses rounding + // instead of truncation; otherwise, truncation is used. + // + // TODO(b/69266521): remove the round_to_bfloat16 flag when rounding becomes + // the default behavior. + StatusOr> ConvertToShape( + const Shape& dest_shape, bool round_f32_to_bf16 = false) const; + + // Converts this literal to another primitive type using a bitcast + // conversion. The to and from primitive types must have the same bit + // width. Returns an error if the conversion is not possible. This literal + // must be array-shaped. + StatusOr> BitcastConvert( + PrimitiveType primitive_dest_type) const; + + // Converts this literal to another primitive type. Returns an error if the + // conversion is not possible. This literal must be array-shaped. + StatusOr> Convert( + PrimitiveType primitive_dest_type) const; + + // Clones the underlying buffers into a new Literal, or new + // std::unique_ptr. + Literal Clone() const; + std::unique_ptr CloneToUnique() const; + + // TODO(b/67651157): The methods below which perform computation on Literals + // (Reshape, Slice, etc) should be moved elsewhere, and perhaps combined with + // evaluator code which operates on Literals. + // + // Creates a new value that has the equivalent value as this + // literal, but conforms to new_layout; e.g. a literal matrix that was in {0, + // 1} minor-to-major dimension layout can be re-layed-out as {1, 0} + // minor-to-major dimension layout and the value in the cell at any given + // logical index (i0, i1) will be the same. + // + // For tuple shaped literals, shape_index should be used to select the inner + // array that the new layout applies to. + // + // Note: this is useful when the client wants to ensure that a value placed in + // the XLA allocation tracker has a particular layout; for efficiency + // purposes or avoiding unimplemented operation/layout combinations. + std::unique_ptr Relayout(const Layout& new_layout, + const ShapeIndex& shape_index = {}) const; + + // An overload of Relayout which changes the layout of the entire shape rather + // than being limited to a single array within the shape. + std::unique_ptr Relayout(const Shape& shape_with_layout) const; + + // Creates a new literal by reshaping this literal to have the given + // dimensions. The total number of elements must not change; The + // implementation currently only supports monotonic dim0-major layouts. + // This literal must be an array. + StatusOr> Reshape( + tensorflow::gtl::ArraySlice dimensions) const; + + // Creates a new literal by broadcasting this literal with `dimensions` to + // yield a literal of shape `result_shape`. + StatusOr> Broadcast( + const Shape& result_shape, + tensorflow::gtl::ArraySlice dimensions) const; + + // Creates a new literal by reordering the dimensions of this literal. + // The given `permutation` must be a permutation of the dimension numbers + // in the original literal, and it specifies the order of the new dimensions + // in the result literal (i.e., new_order[i] = old_order[permutation[i]]). + // For example, a transpose call on a literal of shape [3 x 8 x 4] and + // `permutation` = {2, 0, 1} returns a new literal of shape [4 x 3 x 8]. + // This literal must be an array. + std::unique_ptr Transpose( + tensorflow::gtl::ArraySlice permutation) const; + + // Creates a sub-array from this literal by extracting the indices + // [start_index, limit_index) of each dimension. The result literal has the + // same rank and layout as for the given literal. The number of indices in + // start_indices and limit_indices must be the rank of the literal, and the + // indices follow the order of the dimensions. + // This literal must be an array. + std::unique_ptr Slice( + tensorflow::gtl::ArraySlice start_indices, + tensorflow::gtl::ArraySlice limit_indices) const; + + // Creates a literal with a prepended dimension with bound "times"; e.g. a + // f32[3x2] with times=4 will produce a f32[4x3x2] with the 3x2 from this + // literal replicated four times. + // This literal must be an array. + template + std::unique_ptr Replicate(int64 times) const; + + // Creates a new Literal object with the shape specified as parameter. + // The content of the literal values is the default value of the primitive + // type of literal itself (0 for numeric types, and false for predicates). + // + // Note: It's an antipattern to use this method then immediately call + // Literal::Populate on the result (since that results in zero initialization, + // then reinitialization. Conside if a call to MakeUnique(shape), + // followed by the call to Literal::Populate can be used instead. + static std::unique_ptr CreateFromShape(const Shape& shape); + + protected: + // A data structure representing a subshape at a particular ShapeIndex within + // the literal. For array-shaped ShapeIndexes, this data structure holds the + // pointer to the memory allocated for the array data. + class Piece { + public: + // Returns the buffer holding the array data for this piece as an array + // slice. This piece must be array-shaped. + template + tensorflow::gtl::ArraySlice data() const; + template + tensorflow::gtl::MutableArraySlice data(); + + // Returns the buffer holding the array data for this piece as a void*. This + // piece must be array-shaped. + void* untyped_data(); + const void* untyped_data() const; + + // Gets or sets an element in the array at the given index. The multi_index + // is CHECKed against the dimension sizes of the array. This piece must be + // array-shaped. + template + NativeT Get(tensorflow::gtl::ArraySlice index) const; + template + void Set(tensorflow::gtl::ArraySlice index, NativeT value); + + // Gets/sets the buffer holding the array data. + char* buffer() const { return buffer_; } + void set_buffer(char* buffer) { buffer_ = buffer; } + + // The array of multi-indices that provide the locations of non-zero + // elements in a sparse array. Only used if + // LayoutUtil::IsSparseArray(shape()) is true. + SparseIndexArray* sparse_indices() const { return sparse_indices_; } + void set_sparse_indices(SparseIndexArray* sparse_indices) { + sparse_indices_ = sparse_indices; + } + + // Gets or sets the subshape of this piece. This reference points to a + // subshape within the shape in the containing Literal (Literal::shape_). + const Shape& subshape() const { return *subshape_; } + void set_subshape(const Shape* subshape) { subshape_ = subshape; } + + // Returns the size in bytes of the buffer holding the array data. + int64 size_bytes() const { return ShapeUtil::ByteSizeOf(subshape()); } + + // Returns the number of elements in this piece's array. + int64 element_count() const { + // If this is a sparse array, use the number of elements represented by + // the indices in the associated SparseIndexArray. + return LayoutUtil::IsSparseArray(subshape()) + ? sparse_indices()->index_count() + : ShapeUtil::ElementsIn(subshape()); + } + + // Returns the child piece at 'index' of this piece. + Piece& child(int64 index) { return children_[index]; } + + // Adds a child piece to this piece's children. + void emplace_back(Piece child_piece) { + children_.emplace_back(std::move(child_piece)); + } + + // Returns the size of children pieces of this piece. + int64 children_size() { return children_.size(); } + + // Visitor functions that recursively traverses the piece and calls the + // given function at each child piece. The function has the type: + // void (const ShapeIndex& index, const Piece& piece) + template + void ForEachSubpiece(const Fn& func) const { + ShapeIndex index; + return ForEachHelper( + [&func](const ShapeIndex& index, const Piece& piece) { + func(index, piece); + return Status::OK(); + }, + *this, &index) + .IgnoreError(); + } + // Same as above, but the function has the type: + // Status (const ShapeIndex& index, const Piece& piece) + // The first non-OK return value is returned by the function. + template + Status ForEachSubpieceWithStatus(const Fn& func) const { + ShapeIndex index; + return ForEachHelper(func, *this, &index); + } + // Same as above, but the function has the type: + // Bool (const ShapeIndex& index, const Piece& piece) + // The first non-true return value is returned by the function. + template + bool ForEachSubpieceWithBool(const Fn& func) const { + ShapeIndex index; + return ForEachHelperBool(func, *this, &index); + } + // Same as above, but the function has the type: + // Void (const ShapeIndex& index, Piece& piece) + template + void ForEachMutableSubpiece(const Fn& func) { + ShapeIndex index; + return ForEachMutableHelper( + [&func](const ShapeIndex& index, Piece* piece) { + func(index, piece); + return Status::OK(); + }, + const_cast(this), &index) + .IgnoreError(); + } + // Same as above, but the function has the type: + // Status (const ShapeIndex& index, Piece& piece) + // The first non-OK return value is returned by the function. + template + Status ForEachMutableSubpieceWithStatus(const Fn& func) { + ShapeIndex index; + return ForEachMutableHelper( + func, const_cast(this), &index); + } + + // Returns true if this piece and 'other' contain the same data. This piece + // and 'other' must be array-shaped and compatible. + bool EqualElements(const Piece& other) const; + + // Writes the shape and data (if array-shaped) into the given proto. + void WriteToProto(LiteralProto* proto) const; + + // Copy the data from 'src' into this piece's buffer. Shapes of this piece + // and src must be compatible. + Status CopyFrom(const Piece& src); + + // Copies the data from the given proto into this piece. The shape of this + // piece must be equal (not just compatible) to the shape of the proto. + Status CopyFromProto(const LiteralProto& proto); + + // Sorts the elements in a sparse array. + void SortSparseElements(); + + private: + // Helpers for traversing the piece via ForEachSubpiece rooted at 'index'. + // The first non-OK (or non-true) value is returned by the function. + // The callable 'func' has the same signature as described above in + // ForEachSubpiece*. + template + Status ForEachHelper(const Fn& func, const Piece& piece, + ShapeIndex* index) const { + TF_RETURN_IF_ERROR(func(*index, piece)); + for (int64 i = 0; i < piece.children_.size(); ++i) { + index->push_back(i); + TF_RETURN_IF_ERROR(ForEachHelper(func, piece.children_[i], index)); + index->pop_back(); + } + return Status::OK(); + } + template + bool ForEachHelperBool(const Fn& func, const Piece& piece, + ShapeIndex* index) const { + if (!func(*index, piece)) { + return false; + } + for (int64 i = 0; i < piece.children_.size(); ++i) { + index->push_back(i); + if (!ForEachHelperBool(func, piece.children_[i], index)) { + return false; + } + index->pop_back(); + } + return true; + } + template + Status ForEachMutableHelper(const Fn& func, Piece* piece, + ShapeIndex* index) { + TF_RETURN_IF_ERROR(func(*index, piece)); + for (int64 i = 0; i < piece->children_.size(); ++i) { + index->push_back(i); + TF_RETURN_IF_ERROR( + ForEachMutableHelper(func, &piece->children_[i], index)); + index->pop_back(); + } + return Status::OK(); + } + + // Recursive helper for EqualElements. + template + bool EqualElementsInternal(const Piece& other, + std::vector* multi_index) const; + + // Helper for SortSparseElements that has the element type as a template + // parameter. + template + void SortSparseElementsInternal(); + + // For array-shaped pieces, this is the buffer holding the literal data. + char* buffer_ = nullptr; + + // For sparse arrays, this is the array of indices. + SparseIndexArray* sparse_indices_ = nullptr; + + // The shape of piece. This points into the shape of the containing Literal + // (Literal::shape_). + const Shape* subshape_ = nullptr; + + // Children pieces for tuple shaped pieces. + std::vector children_ = {}; + }; // class Piece + + const Piece& piece(const ShapeIndex& shape_index) const { + Piece* piece = &const_cast(root_piece()); + for (const auto i : shape_index) { + DCHECK_GE(i, 0); + DCHECK_LT(i, piece->children_size()); + piece = &piece->child(i); + } + return *piece; + } + + // Returns the piece at the root of the shape. + virtual const Piece& root_piece() const = 0; + + // LiteralSlice and Literal must access Pieces of other Literals. + friend class Literal; + friend class LiteralSlice; + friend class BorrowingLiteral; + + private: + template + std::unique_ptr SliceInternal( + const Shape& result_shape, + tensorflow::gtl::ArraySlice start_indices) const; +}; + +// Class representing literal values in XLA. +// +// The underlying buffer and shape is always owned by this class. +class Literal : public LiteralBase { + public: + Literal() : Literal(ShapeUtil::MakeNil()) {} + + // Create a literal of the given shape. The literal is allocated sufficient + // memory to hold the shape. Memory is uninitialized. + explicit Literal(const Shape& shape); + virtual ~Literal(); + + // Literals are moveable, but not copyable. To copy a literal use + // Literal::Clone or Literal::CloneToUnique. This prevents inadvertent copies + // of literals which can be expensive. + Literal(const Literal& other) = delete; + Literal& operator=(const Literal& other) = delete; + Literal(Literal&& other); + // 'allocate_arrays' indicates whether to allocate memory for the arrays in + // the shape. If false, buffer pointers inside of the Literal::Pieces are set + // to nullptr. + Literal(const Shape& shape, bool allocate_arrays); + Literal& operator=(Literal&& other); + + // TODO(b/67651157): Remove this accessor. Literal users should not be able to + // mutate the shape as this can produce malformed Literals. + Shape* mutable_shape_do_not_use() { return shape_.get(); } + + // Returns a MutableArraySlice view of the array for this literal for the + // given NativeT (e.g., float). CHECKs if the subshape of the literal at the + // given ShapeIndex is not array. See primitive_util.h for the mapping from + // XLA type to native type. + template + tensorflow::gtl::MutableArraySlice data( + const ShapeIndex& shape_index = {}); + // Unhide const method from parent class. + using LiteralBase::data; + + // Returns a pointer to the sparse index array. Returns nullptr if the literal + // is not a sparse array. + SparseIndexArray* sparse_indices(const ShapeIndex& shape_index = {}); + + // Returns a pointer to the underlying buffer holding the array at the given + // shape index. CHECKs if the subshape of the literal at the given ShapeIndex + // is not array. + void* untyped_data(const ShapeIndex& shape_index = {}); + // Unhide const method from parent class. + using LiteralBase::untyped_data; + + // Populates a literal with a sparse layout with the given indices and values. + // Each index in the indices array is CHECKed against the dimensions in the + // literal's shape. If sort is true, then the indices and values will be + // sorted. If sort is false, then the indices and values are assumed to + // already be in sorted order. See CreateSparse for an example of how data + // are populated. + template + void PopulateSparse(SparseIndexArray indices, + tensorflow::gtl::ArraySlice values, + bool sort = true); + + // Copy values from 'src_literal' rooted at 'src_shape_index' into this + // literal rooted at 'dest_shape_index'. The subshape of this literal rooted + // at 'dest_shape_index' must be compatible with the subshape of 'src_literal' + // rooted at 'src_shape_index', but need not be arrays. + Status CopyFrom(const LiteralSlice& src_literal, + const ShapeIndex& dest_shape_index = {}, + const ShapeIndex& src_shape_index = {}); + + // Returns a vector containing the tuple elements of this Literal as separate + // Literals. This Literal must be tuple-shaped and can be a nested tuple. The + // elements are moved into the new Literals; no data is copied. Upon return + // this Literal is set to a nil shape (empty tuple) + std::vector DecomposeTuple(); + + // Similar to CopyFrom, but with move semantincs. The subshape of this literal + // rooted at 'dest_shape_index' must be *equal* to the shape 'src_literal' + // (layouts and shapes must match), but need not be arrays. The memory + // allocated in this literal for the subshape at dest_shape_index is + // deallocated, and the respective buffers are replaced with those in + // src_literal. Upon return, src_literal is set to a nil shape (empty tuple). + Status MoveFrom(Literal&& src_literal, + const ShapeIndex& dest_shape_index = {}); + + // Copies the values from src_literal, starting at src_base shape indexes, + // to this literal, starting at dest_base, where the copy size in each + // dimension is specified by copy_size. + // The src_literal and this literal must have the same primitive type, + // src_base+copy_size must fit the source literal dimensions, as well as + // dest_base+copy_size must fit the destination literal dimensions. + // Note: if either src_literal or this literal contains dimensions with zero + // element, then copy_size must be 0 in these dimensions while the + // corresponding base indices being 0. + // This literal and 'src_literal' must be arrays. + Status CopySliceFrom(const LiteralSlice& src_literal, + tensorflow::gtl::ArraySlice src_base, + tensorflow::gtl::ArraySlice dest_base, + tensorflow::gtl::ArraySlice copy_size); + + // Copies one element from src_literal[src_index] to (*this)[dest_index]. + Status CopyElementFrom(const LiteralSlice& src_literal, + tensorflow::gtl::ArraySlice src_index, + tensorflow::gtl::ArraySlice dest_index); + + // Sets an element in the literal at the given index. The multi_index is + // CHECKed against the dimension sizes. + template + void Set(tensorflow::gtl::ArraySlice multi_index, + const ShapeIndex& shape_index, NativeT value); + // Overloads of Set for array literals. CHECKs if the literal is not + // array-shaped and dense. + template + void Set(tensorflow::gtl::ArraySlice multi_index, NativeT value); + + // Appends the given element to the literal. If the elements are not appended + // in sorted order, then SortSparseElements should be called before calling + // other methods. This literal must have a sparse layout. + template + void AppendSparseElement(tensorflow::gtl::ArraySlice multi_index, + NativeT value, const ShapeIndex& shape_index = {}); + + // Sorts the elements in a sparse array. + void SortSparseElements(const ShapeIndex& shape_index = {}); + + // As Set(), but truncates `value` to the literal element type before storing. + // This literal must be an array. + Status SetIntegralAsS64(tensorflow::gtl::ArraySlice multi_index, + int64 value); + + // Populate this literal with the given values. Examples: + // + // // Populate with floats. + // Array2D float_values = ... + // literal.PopulateR2FromArray2D(values); + // + // // Populate with int32s. + // literal.PopulateR2({{1, 2}, {3, 4}}); + // + // The shape and element type of this literal must match given values. For + // example, in the call above to literal.PopulateR2(), 'literal' must be a 2x2 + // array of S32. + template + void PopulateR1(tensorflow::gtl::ArraySlice values); + void PopulateR1(const tensorflow::core::Bitmap& values); + template + void PopulateR2(std::initializer_list> values); + template + void PopulateFromArray(const Array& values); + template + void PopulateR2FromArray2D(const Array2D& values); + template + void PopulateR3FromArray3D(const Array3D& values); + template + void PopulateR4FromArray4D(const Array4D& values); + + // Populates literal values by calling the generator function for every cell + // in this literal object. + // + // generator must be a callable of the type + // NativeT(tensorflow::gtl::ArraySlice indexes) or compatible. + // + // This literal must have a dense layout. + template + Status Populate(const FnType& generator); + + // A parallel version of Populate(). This can be used if the generator is + // thread-safe and the values for the shape's different elements are + // independent. + template + Status PopulateParallel(const FnType& generator); + + // Fills this literal with the given value. + template + void PopulateWithValue(NativeT value); + + // This operation is the inverse of DecomposeTuple. The given elements are + // moved into the tuple elements of a new tuple-shaped Literal which is + // returned. Upon return, each of the Literals in 'elements' is set to a nil + // shape (empty tuple). + static Literal MoveIntoTuple( + tensorflow::gtl::MutableArraySlice elements); + + // Serialize from a proto. + static StatusOr> CreateFromProto( + const LiteralProto& proto); + + private: + // Recursively sets the subshapes and buffers of all subpieces rooted at + // 'piece'. If 'allocate_array' is true, memory is allocated for the arrays in + // the shape. + void SetPiece(const Shape& shape, Piece* piece, bool allocate_arrays); + + // Returns the piece at the given ShapeIndex. + Piece& piece(const ShapeIndex& shape_index) { + return const_cast(LiteralBase::piece(shape_index)); + } + + Piece& root_piece() const override { return *root_piece_; }; + + // Internal template helper for the Literal::CopySliceFrom(), matching its + // arguments one by one. + template + Status CopySliceFromInternal(const LiteralBase& src_literal, + tensorflow::gtl::ArraySlice src_base, + tensorflow::gtl::ArraySlice dest_base, + tensorflow::gtl::ArraySlice copy_size); + + // Utility structure which is used to create the optimal configuration for + // a ShapeUtil::ForEachIndex() scan across two literals. + struct StrideConfig { + StrideConfig(const Shape& source_shape, const Shape& dest_shape, + tensorflow::gtl::ArraySlice dimensions); + + // The dimensions of the stride operation. Essentially every dimension + // will be iterated from base[i] to base[i]+dimensions[i], in step[i] + // steps. + tensorflow::gtl::ArraySlice dimensions; + DimensionVector base; + DimensionVector step; + int64 minor_dimension = 0; + // The size of the strides for source and destination. One of the two + // (the one looping through its most minor dimension) will be 1, while + // the other will be the stride size at the dimension matching the other + // shape most minor dimension being scanned. + int64 dest_stride = 1; + int64 source_stride = 1; + // The size of the inner loop on the most minor dimension. + int64 minor_loop_size = 1; + }; + + // Literal class always owns the shape. The parent class borrows this shape. + std::unique_ptr shape_; + + Piece* root_piece_ = nullptr; + + // Implementation details shared between Populate() and PopulateParallel() + template + Status PopulateInternal(const FnType& generator, bool parallel); + + // Deallocate the buffers held by this literal. + void DeallocateBuffers(); + + friend class LiteralBase; +}; +std::ostream& operator<<(std::ostream& out, const Literal& literal); + +// A read-only view of a Literal. A LiteralSlice contains pointers to shape and +// literal buffers always owned by others. +class LiteralSlice : public LiteralBase { + public: + LiteralSlice() : LiteralBase() {} + + // Implicit conversion constructors. + LiteralSlice(const LiteralBase& literal); + LiteralSlice(const LiteralBase& literal, const ShapeIndex& view_root); + + private: + const Piece& root_piece() const override { return *root_piece_; }; + + const Piece* root_piece_; // Not owned. +}; + +// A read-only Literal where the underlying buffers are never owned by this +// class. +class BorrowingLiteral : public LiteralBase { + public: + BorrowingLiteral() : LiteralBase() {} + + // 'src_buf_ptr' is not owned by this class and must outlive the + // lifetime of this class. It points to an appropirately sized buffer with + // data interpretered as indicated by 'shape'. + // This constructor is only used for array shapes. + BorrowingLiteral(const char* src_buf_ptr, const Shape& shape); + // Similar as above, except to be used for constructing non-nested tuples. + BorrowingLiteral(tensorflow::gtl::ArraySlice src_buf_ptrs, + const Shape& shape); + // TODO(b/79707221): adding constructors for nested tuples as well. + + private: + // Recursively builds the subtree for the given piece and sets the subshapes + // of the given piece with the given shape. + void BuildPieceSubtree(const Shape& shape, Piece* piece); + + // Accessor for the root piece of this literal. + const Piece& root_piece() const override { return root_piece_; }; + Piece root_piece_; + + // Shape of this literal. Stored as unique_ptr so such that the (default) + // move construction of this class would be trivially correct: the pointer to + // Shape root_piece_ stores will still point to the correct address. + std::unique_ptr shape_; +}; + +template +tensorflow::gtl::ArraySlice LiteralBase::Piece::data() const { + CHECK(ShapeUtil::IsArray(subshape())) << ShapeUtil::HumanString(subshape()); + CHECK_EQ(subshape().element_type(), + primitive_util::NativeToPrimitiveType()) + << "Attempting to access " + << PrimitiveType_Name(primitive_util::NativeToPrimitiveType()) + << " type, but literal element type is " + << PrimitiveType_Name(subshape().element_type()); + return tensorflow::gtl::ArraySlice( + reinterpret_cast(buffer()), element_count()); +} + +template +tensorflow::gtl::MutableArraySlice LiteralBase::Piece::data() { + CHECK(ShapeUtil::IsArray(subshape())) << ShapeUtil::HumanString(subshape()); + CHECK_EQ(subshape().element_type(), + primitive_util::NativeToPrimitiveType()) + << "Attempting to access " + << PrimitiveType_Name(primitive_util::NativeToPrimitiveType()) + << " type, but literal element type is " + << PrimitiveType_Name(subshape().element_type()); + return tensorflow::gtl::MutableArraySlice( + reinterpret_cast(buffer()), element_count()); +} + +template +NativeT LiteralBase::Piece::Get( + tensorflow::gtl::ArraySlice multi_index) const { + CHECK(LayoutUtil::IsDenseArray(subshape())); + return data()[IndexUtil::MultidimensionalIndexToLinearIndex( + subshape(), multi_index)]; +} + +template +void LiteralBase::Piece::Set(tensorflow::gtl::ArraySlice multi_index, + NativeT value) { + CHECK(LayoutUtil::IsDenseArray(subshape())); + data()[IndexUtil::MultidimensionalIndexToLinearIndex( + subshape(), multi_index)] = value; +} + +template +tensorflow::gtl::ArraySlice LiteralBase::data( + const ShapeIndex& shape_index) const { + return piece(shape_index).data(); +} + +template +tensorflow::gtl::MutableArraySlice Literal::data( + const ShapeIndex& shape_index) { + return piece(shape_index).data(); +} + +template +inline NativeT LiteralBase::Get(tensorflow::gtl::ArraySlice multi_index, + const ShapeIndex& shape_index) const { + return piece(shape_index).Get(multi_index); +} + +template +inline NativeT LiteralBase::Get( + tensorflow::gtl::ArraySlice multi_index) const { + return root_piece().Get(multi_index); +} + +template +inline void Literal::Set(tensorflow::gtl::ArraySlice multi_index, + const ShapeIndex& shape_index, NativeT value) { + return piece(shape_index).Set(multi_index, value); +} + +template +inline void Literal::Set(tensorflow::gtl::ArraySlice multi_index, + NativeT value) { + return root_piece().Set(multi_index, value); +} + +template +NativeT LiteralBase::GetFirstElement() const { + return data().at(0); +} + +template +NativeT LiteralBase::GetSparseElement(int64 sparse_element_number, + const ShapeIndex& shape_index) const { + CHECK( + LayoutUtil::IsSparseArray(ShapeUtil::GetSubshape(shape(), shape_index))); + return data(shape_index)[sparse_element_number]; +} + +template +void Literal::AppendSparseElement( + tensorflow::gtl::ArraySlice multi_index, NativeT value, + const ShapeIndex& shape_index) { + Piece& p = piece(shape_index); + const Shape& subshape = p.subshape(); + CHECK(LayoutUtil::IsSparseArray(subshape)); + int64 rank = ShapeUtil::Rank(subshape); + CHECK_EQ(multi_index.size(), rank); + int64 last_element = p.sparse_indices()->index_count(); + CHECK_LT(last_element, LayoutUtil::MaxSparseElements(subshape.layout())); + p.sparse_indices()->Append(multi_index); + CHECK_LT(last_element, p.data().size()); + p.data()[last_element] = value; +} + +template +void LiteralBase::EachCell( + std::function indices, + NativeT value)> + per_cell) const { + if (ShapeUtil::IsZeroElementArray(shape())) { + return; + } + std::vector indices(ShapeUtil::Rank(shape()), 0); + do { + per_cell(indices, Get(indices)); + } while (IndexUtil::BumpIndices(shape(), &indices)); +} + +template +inline void Literal::PopulateR1(tensorflow::gtl::ArraySlice values) { + CHECK(ShapeUtil::IsArray(shape())); + CHECK_EQ(ShapeUtil::Rank(shape()), 1); + CHECK_EQ(ShapeUtil::ElementsIn(shape()), values.size()); + CHECK_EQ(shape().element_type(), + primitive_util::NativeToPrimitiveType()); + for (int64 i = 0; i < values.size(); ++i) { + Set({i}, values[i]); + } +} + +template +void Literal::PopulateR2( + std::initializer_list> values) { + CHECK(ShapeUtil::IsArray(shape())); + CHECK_EQ(ShapeUtil::Rank(shape()), 2); + CHECK_EQ(shape().element_type(), + primitive_util::NativeToPrimitiveType()); + + const int64 dim0_size = values.size(); + const int64 dim1_size = values.begin()->size(); + CHECK_EQ(dim0_size, shape().dimensions(0)); + CHECK_EQ(dim1_size, shape().dimensions(1)); + + int64 dim0 = 0; + for (auto inner_list : values) { + int64 dim1 = 0; + for (auto value : inner_list) { + Set({dim0, dim1}, value); + ++dim1; + } + CHECK_EQ(dim1_size, dim1); + ++dim0; + } +} + +template +void Literal::PopulateFromArray(const Array& values) { + CHECK(ShapeUtil::IsArray(shape())); + CHECK_EQ(shape().element_type(), + primitive_util::NativeToPrimitiveType()); + CHECK_EQ(ShapeUtil::Rank(shape()), values.num_dimensions()); + for (int dim = 0; dim < values.num_dimensions(); ++dim) { + CHECK_EQ(values.dim(dim), shape().dimensions(dim)); + } + values.Each([this](tensorflow::gtl::ArraySlice indices, + NativeT value) { this->Set(indices, value); }); +} + +template +void Literal::PopulateR2FromArray2D(const Array2D& values) { + PopulateFromArray(values); +} + +template +void Literal::PopulateR3FromArray3D(const Array3D& values) { + PopulateFromArray(values); +} + +template +void Literal::PopulateR4FromArray4D(const Array4D& values) { + PopulateFromArray(values); +} + +template +void Literal::PopulateSparse(SparseIndexArray indices, + tensorflow::gtl::ArraySlice values, + bool sort) { + CHECK(LayoutUtil::IsSparseArray(shape())); + int rank = ShapeUtil::Rank(shape()); + CHECK_EQ(indices.rank(), rank); + int64 max_elements = LayoutUtil::MaxSparseElements(shape().layout()); + CHECK_LE(indices.max_indices(), max_elements); + int64 num_elements = values.size(); + CHECK_LE(num_elements, max_elements); + CHECK_EQ(num_elements, indices.index_count()); + auto root_data = root_piece().data(); + // Piece::data() returns an ArraySlice of size equal to the number of indices + // in the SparseIndexArray. So there is no need to adjust the size of the data + // here. It is enough to just copy the incoming values into the data buffer. + std::copy(values.begin(), values.end(), root_data.begin()); + *this->root_piece().sparse_indices() = std::move(indices); + if (sort) { + auto root_data = this->root_piece().data(); + this->root_piece().sparse_indices()->SortWithValues(root_data); + } + DCHECK(this->root_piece().sparse_indices()->Validate(shape())); +} + +template +Status Literal::PopulateInternal(const FnType& generator, bool parallel) { + const Shape& this_shape = shape(); + const int64 rank = ShapeUtil::Rank(this_shape); + TF_RET_CHECK(LayoutUtil::IsDenseArray(this_shape)); + TF_RET_CHECK(this_shape.element_type() == + primitive_util::NativeToPrimitiveType()); + tensorflow::gtl::MutableArraySlice literal_data = data(); + if (rank > 0) { + StrideConfig stride_config(this_shape, this_shape, + AsInt64Slice(this_shape.dimensions())); + int64 minor_dimension_size = + ShapeUtil::GetDimension(this_shape, stride_config.minor_dimension); + + auto init_function = [&](tensorflow::gtl::ArraySlice indexes) { + DimensionVector minor_scan_indexes(rank, 0); + const int64 index = + IndexUtil::MultidimensionalIndexToLinearIndex(shape(), indexes); + std::copy(indexes.begin(), indexes.end(), minor_scan_indexes.begin()); + for (int64 i = 0; i < minor_dimension_size; ++i) { + minor_scan_indexes[stride_config.minor_dimension] = i; + literal_data.at(index + i) = generator(minor_scan_indexes); + } + }; + if (parallel) { + ShapeUtil::ForEachIndexParallel(this_shape, stride_config.base, + stride_config.dimensions, + stride_config.step, init_function); + } else { + ShapeUtil::ForEachIndex( + this_shape, stride_config.base, stride_config.dimensions, + stride_config.step, + [&init_function](tensorflow::gtl::ArraySlice indexes) { + init_function(indexes); + return true; + }); + } + } else { + // For scalars. + literal_data.at(0) = generator({}); + } + return Status::OK(); +} +template +Status Literal::Populate(const FnType& generator) { + return PopulateInternal(generator, /*parallel=*/false); +} + +template +Status Literal::PopulateParallel(const FnType& generator) { + return PopulateInternal(generator, /*parallel=*/true); +} + +template +void Literal::PopulateWithValue(NativeT value) { + CHECK(ShapeUtil::IsArray(shape())); + CHECK_EQ(shape().element_type(), + primitive_util::NativeToPrimitiveType()); + for (NativeT& element : data()) { + element = value; + } +} + +template +std::unique_ptr LiteralBase::Replicate(int64 times) const { + DimensionVector bounds = {times}; + bounds.reserve(shape().dimensions_size() + 1); + for (int64 bound : shape().dimensions()) { + bounds.push_back(bound); + } + auto literal = + MakeUnique(ShapeUtil::MakeShape(shape().element_type(), bounds)); + int64 elements = ShapeUtil::ElementsIn(literal->shape()); + if (elements == 0) { + return literal; + } + + DimensionVector output_indices(bounds.size(), 0); + tensorflow::gtl::ArraySlice input_indices = output_indices; + input_indices.remove_prefix(1); + + bool done = false; + while (!done) { + const auto element = Get(input_indices); + literal->Set(output_indices, element); + + done = true; + for (int n = 0; n < output_indices.size(); ++n) { + ++output_indices[n]; + if (output_indices[n] < bounds[n]) { + done = false; + break; + } + output_indices[n] = 0; + } + } + return literal; +} + +} // namespace xla + +#endif // TENSORFLOW_COMPILER_XLA_LITERAL_H_ diff --git a/tensorflow/compiler/xla/literal_comparison.cc b/tensorflow/compiler/xla/literal_comparison.cc index 2125ab7c61ab5e30fe51e16994e0da4883d509c4..94993cc87443ba8c22fd7c2eacfc8756d3f48edc 100644 --- a/tensorflow/compiler/xla/literal_comparison.cc +++ b/tensorflow/compiler/xla/literal_comparison.cc @@ -19,6 +19,7 @@ limitations under the License. #include #include +#include "tensorflow/compiler/xla/literal_util.h" #include "tensorflow/compiler/xla/util.h" #include "tensorflow/core/lib/core/casts.h" #include "tensorflow/core/lib/strings/strcat.h" @@ -217,7 +218,7 @@ class NearComparator { return Printf( "actual %s, expected %s, index %s, rel error %8.3g, abs error %8.3g", FpValueToString(actual).c_str(), FpValueToString(expected).c_str(), - Literal::MultiIndexAsString( + LiteralUtil::MultiIndexAsString( IndexUtil::LinearIndexToMultidimensionalIndex(shape, linear_index)) .c_str(), @@ -722,7 +723,7 @@ Status Equal(const LiteralSlice& expected, const LiteralSlice& actual) { return AppendStatus(result, tensorflow::strings::Printf( "\nat index: %s\nexpected: %s\nactual: %s", - Literal::MultiIndexAsString(multi_index).c_str(), + LiteralUtil::MultiIndexAsString(multi_index).c_str(), ToStringTruncated(expected).c_str(), ToStringTruncated(actual).c_str())); } diff --git a/tensorflow/compiler/xla/literal_comparison.h b/tensorflow/compiler/xla/literal_comparison.h index 00a13e361932e74a9a1e614d5c851d3851208852..9e5bf7c1d062ef0f25d07a80d6ded8106df5dacc 100644 --- a/tensorflow/compiler/xla/literal_comparison.h +++ b/tensorflow/compiler/xla/literal_comparison.h @@ -20,7 +20,7 @@ limitations under the License. #define TENSORFLOW_COMPILER_XLA_LITERAL_COMPARISON_H_ #include "tensorflow/compiler/xla/error_spec.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/core/lib/core/status.h" namespace xla { diff --git a/tensorflow/compiler/xla/literal_util_test.cc b/tensorflow/compiler/xla/literal_test.cc similarity index 76% rename from tensorflow/compiler/xla/literal_util_test.cc rename to tensorflow/compiler/xla/literal_test.cc index 493d807591dd3c425293e4ee796bca3036a3088c..e8f919950f0efc8b508f7ad4aee5233176bc0abd 100644 --- a/tensorflow/compiler/xla/literal_util_test.cc +++ b/tensorflow/compiler/xla/literal_test.cc @@ -13,7 +13,7 @@ See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include @@ -21,6 +21,7 @@ limitations under the License. #include "tensorflow/compiler/xla/array3d.h" #include "tensorflow/compiler/xla/array4d.h" #include "tensorflow/compiler/xla/layout_util.h" +#include "tensorflow/compiler/xla/literal_util.h" #include "tensorflow/compiler/xla/shape_util.h" #include "tensorflow/compiler/xla/test.h" #include "tensorflow/compiler/xla/types.h" @@ -76,11 +77,11 @@ class LiteralUtilTest : public ::testing::Test { layout_r4_dim0minor_ = LayoutUtil::MakeLayout({0, 1, 2, 3}); literal_r4_2x2x3x3_dim0major_ = - Literal::CreateR4FromArray4DWithLayout(arr4d, - layout_r4_dim0major_); + LiteralUtil::CreateR4FromArray4DWithLayout(arr4d, + layout_r4_dim0major_); literal_r4_2x2x3x3_dim0minor_ = - Literal::CreateR4FromArray4DWithLayout(arr4d, - layout_r4_dim0minor_); + LiteralUtil::CreateR4FromArray4DWithLayout(arr4d, + layout_r4_dim0minor_); } Layout layout_r2_dim0major_; @@ -94,47 +95,47 @@ class LiteralUtilTest : public ::testing::Test { }; TEST_F(LiteralUtilTest, LiteralScalarToString) { - auto true_lit = Literal::CreateR0(true); + auto true_lit = LiteralUtil::CreateR0(true); ASSERT_EQ("true", true_lit->ToString()); - auto false_lit = Literal::CreateR0(false); + auto false_lit = LiteralUtil::CreateR0(false); ASSERT_EQ("false", false_lit->ToString()); - auto u32_lit = Literal::CreateR0(42); + auto u32_lit = LiteralUtil::CreateR0(42); ASSERT_EQ("42", u32_lit->ToString()); - auto s32_lit = Literal::CreateR0(-999); + auto s32_lit = LiteralUtil::CreateR0(-999); ASSERT_EQ("-999", s32_lit->ToString()); - auto f32_lit = Literal::CreateR0(3.14f); + auto f32_lit = LiteralUtil::CreateR0(3.14f); ASSERT_EQ("3.14", f32_lit->ToString()); - auto f16_lit = Literal::CreateR0(static_cast(0.5f)); + auto f16_lit = LiteralUtil::CreateR0(static_cast(0.5f)); ASSERT_EQ("0.5", f16_lit->ToString()); - auto c64_lit = Literal::CreateR0({3.14f, 2.78f}); + auto c64_lit = LiteralUtil::CreateR0({3.14f, 2.78f}); ASSERT_EQ("(3.14, 2.78)", c64_lit->ToString()); - auto bf16_lit = Literal::CreateR0(static_cast(0.5f)); + auto bf16_lit = LiteralUtil::CreateR0(static_cast(0.5f)); ASSERT_EQ("0.5", bf16_lit->ToString()); // 3.14 will be truncated to 3.125 in bfloat16 format. auto bf16_lit_truncated = - Literal::CreateR0(static_cast(3.14f)); + LiteralUtil::CreateR0(static_cast(3.14f)); ASSERT_EQ("3.125", bf16_lit_truncated->ToString()); auto bf16_lit_truncated2 = - Literal::CreateR0(static_cast(9.001f)); + LiteralUtil::CreateR0(static_cast(9.001f)); ASSERT_EQ("9", bf16_lit_truncated2->ToString()); } TEST_F(LiteralUtilTest, LiteralVectorToString) { - auto pred_vec = Literal::CreateR1({true, false, true}); + auto pred_vec = LiteralUtil::CreateR1({true, false, true}); ASSERT_EQ("{101}", pred_vec->ToString()); } TEST_F(LiteralUtilTest, R2ToString) { - const auto literal = Literal::CreateR2({{1, 2}, {3, 4}, {5, 6}}); + const auto literal = LiteralUtil::CreateR2({{1, 2}, {3, 4}, {5, 6}}); const string expected = R"(s32[3,2] { { 1, 2 }, { 3, 4 }, @@ -144,7 +145,8 @@ TEST_F(LiteralUtilTest, R2ToString) { } TEST_F(LiteralUtilTest, R3ToString) { - const auto literal = Literal::CreateR3({{{1}, {2}}, {{3}, {4}}, {{5}, {6}}}); + const auto literal = + LiteralUtil::CreateR3({{{1}, {2}}, {{3}, {4}}, {{5}, {6}}}); const string expected = R"(s32[3,2,1] { { { 1 }, { 2 } }, @@ -157,9 +159,9 @@ TEST_F(LiteralUtilTest, R3ToString) { } TEST_F(LiteralUtilTest, TupleToString) { - auto scalar = Literal::CreateR0(1.0); - auto matrix = Literal::CreateR2({{1.0, 2.0}, {3.0, 4.0}}); - auto tuple = Literal::MakeTuple({scalar.get(), matrix.get()}); + auto scalar = LiteralUtil::CreateR0(1.0); + auto matrix = LiteralUtil::CreateR2({{1.0, 2.0}, {3.0, 4.0}}); + auto tuple = LiteralUtil::MakeTuple({scalar.get(), matrix.get()}); const string expected = R"((f32[], f32[2,2]) ( 1, f32[2,2] { @@ -182,7 +184,7 @@ TEST_F(LiteralUtilTest, CreateR3FromArray3d) { }); // clang-format on - auto literal = Literal::CreateR3FromArray3D(array_3d); + auto literal = LiteralUtil::CreateR3FromArray3D(array_3d); EXPECT_THAT(literal->shape().dimensions(), ElementsAre(2, 3, 2)); string result = literal->ToString(); const string expected = R"(f32[2,3,2] { @@ -205,7 +207,7 @@ TEST_F(LiteralUtilTest, CreateSparse) { {3, 5, 6}, }; std::vector values = {7, 8, 9, 10}; - auto literal = Literal::CreateSparse( + auto literal = LiteralUtil::CreateSparse( dimensions, SparseIndexArray(indices.n1() + 3, indices), values); Array2D expected_indices = { @@ -224,7 +226,7 @@ TEST_F(LiteralUtilTest, CreateSparse) { TEST_F(LiteralUtilTest, LiteralR4F32ProjectedStringifies) { // clang-format off - auto literal = Literal::CreateR4Projected({ + auto literal = LiteralUtil::CreateR4Projected({ {1, 2}, {1001, 1002}, {2001, 2002}, @@ -284,7 +286,7 @@ TEST_F(LiteralUtilTest, LiteralR4F32Stringifies) { TEST_F(LiteralUtilTest, EachCellR2F32) { // clang-format off - auto literal = Literal::CreateR2({ + auto literal = LiteralUtil::CreateR2({ {3.1f, 4.2f}, {9.3f, 12.4f}, }); @@ -303,26 +305,27 @@ TEST_F(LiteralUtilTest, EachCellR2F32) { TEST_F(LiteralUtilTest, ScalarEquality) { // Test equality with scalars. - auto f32_42 = Literal::CreateR0(42.0); - auto f32_42_clone = Literal::CreateR0(42.0); + auto f32_42 = LiteralUtil::CreateR0(42.0); + auto f32_42_clone = LiteralUtil::CreateR0(42.0); EXPECT_EQ(*f32_42, *f32_42); EXPECT_EQ(*f32_42, *f32_42_clone); - auto f32_123 = Literal::CreateR0(123.0); + auto f32_123 = LiteralUtil::CreateR0(123.0); EXPECT_NE(*f32_42, *f32_123); - auto f64_42 = Literal::CreateR0(42.0); + auto f64_42 = LiteralUtil::CreateR0(42.0); EXPECT_NE(*f32_42, *f64_42); } TEST_F(LiteralUtilTest, NonScalarEquality) { // Test equality with nonscalars. - auto matrix = Literal::CreateR2({{1.0, 2.0}, {3.0, 4.0}}); - auto matrix_clone = Literal::CreateR2({{1.0, 2.0}, {3.0, 4.0}}); - auto matrix_different = Literal::CreateR2({{4.0, 3.0}, {1.0, 2.0}}); - auto vector_literal = Literal::CreateR1({1.0, 2.0, 3.0, 4.0}); - auto scalar = Literal::CreateR0(1.0); + auto matrix = LiteralUtil::CreateR2({{1.0, 2.0}, {3.0, 4.0}}); + auto matrix_clone = LiteralUtil::CreateR2({{1.0, 2.0}, {3.0, 4.0}}); + auto matrix_different = + LiteralUtil::CreateR2({{4.0, 3.0}, {1.0, 2.0}}); + auto vector_literal = LiteralUtil::CreateR1({1.0, 2.0, 3.0, 4.0}); + auto scalar = LiteralUtil::CreateR0(1.0); Literal nil(ShapeUtil::MakeNil()); EXPECT_EQ(*matrix, *matrix); @@ -335,19 +338,19 @@ TEST_F(LiteralUtilTest, NonScalarEquality) { } TEST_F(LiteralUtilTest, TokenEquality) { - auto token0 = Literal::CreateToken(); - auto token1 = Literal::CreateToken(); - auto scalar = Literal::CreateR0(1.0); + auto token0 = LiteralUtil::CreateToken(); + auto token1 = LiteralUtil::CreateToken(); + auto scalar = LiteralUtil::CreateR0(1.0); EXPECT_EQ(*token0, *token1); EXPECT_NE(*token0, *scalar); - EXPECT_EQ(*Literal::MakeTuple({token0.get()}), - *Literal::MakeTuple({token0.get()})); - EXPECT_EQ(*Literal::MakeTuple({token0.get(), scalar.get()}), - *Literal::MakeTuple({token1.get(), scalar.get()})); - EXPECT_NE(*Literal::MakeTuple({token0.get(), scalar.get()}), - *Literal::MakeTuple({scalar.get(), token1.get()})); + EXPECT_EQ(*LiteralUtil::MakeTuple({token0.get()}), + *LiteralUtil::MakeTuple({token0.get()})); + EXPECT_EQ(*LiteralUtil::MakeTuple({token0.get(), scalar.get()}), + *LiteralUtil::MakeTuple({token1.get(), scalar.get()})); + EXPECT_NE(*LiteralUtil::MakeTuple({token0.get(), scalar.get()}), + *LiteralUtil::MakeTuple({scalar.get(), token1.get()})); } TEST_F(LiteralUtilTest, DifferentLayoutEquality) { @@ -371,43 +374,46 @@ TEST_F(LiteralUtilTest, DifferentLayoutEquality) { TEST_F(LiteralUtilTest, TupleEquality) { // Test equality with tuples. - auto scalar = Literal::CreateR0(1.0); - auto matrix = Literal::CreateR2({{1.0, 2.0}, {3.0, 4.0}}); - auto tuple1 = Literal::MakeTuple({scalar.get(), matrix.get()}); + auto scalar = LiteralUtil::CreateR0(1.0); + auto matrix = LiteralUtil::CreateR2({{1.0, 2.0}, {3.0, 4.0}}); + auto tuple1 = LiteralUtil::MakeTuple({scalar.get(), matrix.get()}); // Tuple with the same elements. One element is shared with the original // tuple, the other is a clone of the element in the original tuple. - auto scalar_clone = Literal::CreateR0(1.0); - auto tuple2 = Literal::MakeTuple({scalar_clone.get(), matrix.get()}); + auto scalar_clone = LiteralUtil::CreateR0(1.0); + auto tuple2 = LiteralUtil::MakeTuple({scalar_clone.get(), matrix.get()}); EXPECT_EQ(*tuple1, *tuple2); // Tuple with elements reversed. - auto reversed_tuple = Literal::MakeTuple({matrix.get(), scalar.get()}); + auto reversed_tuple = LiteralUtil::MakeTuple({matrix.get(), scalar.get()}); EXPECT_NE(*tuple1, *reversed_tuple); // Tuple with different value. - auto scalar_42 = Literal::CreateR0(42.0); - auto different_tuple = Literal::MakeTuple({scalar_42.get(), matrix.get()}); + auto scalar_42 = LiteralUtil::CreateR0(42.0); + auto different_tuple = + LiteralUtil::MakeTuple({scalar_42.get(), matrix.get()}); EXPECT_NE(*tuple1, *different_tuple); } TEST_F(LiteralUtilTest, C64Equality) { // Test equality with tuples. - auto vector = Literal::CreateR1({{1.0, 2.0}, {3.0, 4.0}}); + auto vector = LiteralUtil::CreateR1({{1.0, 2.0}, {3.0, 4.0}}); // Tuple with the same elements. One element is shared with the original // tuple, the other is a clone of the element in the original tuple. - auto vector_clone = Literal::CreateR1({{1.0, 2.0}, {3.0, 4.0}}); + auto vector_clone = + LiteralUtil::CreateR1({{1.0, 2.0}, {3.0, 4.0}}); EXPECT_EQ(*vector, *vector_clone); - auto vector_reversed = Literal::CreateR1({{3.0, 4.0}, {1.0, 2.0}}); + auto vector_reversed = + LiteralUtil::CreateR1({{3.0, 4.0}, {1.0, 2.0}}); EXPECT_NE(*vector, *vector_reversed); } TEST_F(LiteralUtilTest, IsAllTuple) { - auto element1 = Literal::CreateR0(0.0); - auto element2 = Literal::CreateR2({{0.0, 0.0}, {0.0, 0.0}}); - auto tuple = Literal::MakeTuple({element1.get(), element1.get()}); + auto element1 = LiteralUtil::CreateR0(0.0); + auto element2 = LiteralUtil::CreateR2({{0.0, 0.0}, {0.0, 0.0}}); + auto tuple = LiteralUtil::MakeTuple({element1.get(), element1.get()}); // Tuples should always return false for IsAll. EXPECT_FALSE(tuple->IsAll(0)); @@ -416,140 +422,141 @@ TEST_F(LiteralUtilTest, IsAllTuple) { // Verifies that CreateFromShape works for tuples. TEST_F(LiteralUtilTest, CreateFromShapeTuple) { - auto scalar = Literal::CreateR0(0.0); - auto matrix = Literal::CreateR2({{0, 0}, {0, 0}}); - auto tuple = Literal::MakeTuple({scalar.get(), matrix.get()}); + auto scalar = LiteralUtil::CreateR0(0.0); + auto matrix = LiteralUtil::CreateR2({{0, 0}, {0, 0}}); + auto tuple = LiteralUtil::MakeTuple({scalar.get(), matrix.get()}); auto x = Literal::CreateFromShape(tuple->shape()); EXPECT_EQ(*tuple, *x); } TEST_F(LiteralUtilTest, IsAll) { - EXPECT_TRUE(Literal::CreateR0(false)->IsAll(0)); - EXPECT_TRUE(Literal::CreateR0(true)->IsAll(1)); - EXPECT_FALSE(Literal::CreateR0(false)->IsAll(1)); - EXPECT_FALSE(Literal::CreateR0(false)->IsAll(2)); - EXPECT_FALSE(Literal::CreateR0(true)->IsAll(0)); - EXPECT_FALSE(Literal::CreateR0(true)->IsAll(2)); - EXPECT_FALSE(Literal::CreateR0(true)->IsAll(-1)); + EXPECT_TRUE(LiteralUtil::CreateR0(false)->IsAll(0)); + EXPECT_TRUE(LiteralUtil::CreateR0(true)->IsAll(1)); + EXPECT_FALSE(LiteralUtil::CreateR0(false)->IsAll(1)); + EXPECT_FALSE(LiteralUtil::CreateR0(false)->IsAll(2)); + EXPECT_FALSE(LiteralUtil::CreateR0(true)->IsAll(0)); + EXPECT_FALSE(LiteralUtil::CreateR0(true)->IsAll(2)); + EXPECT_FALSE(LiteralUtil::CreateR0(true)->IsAll(-1)); // We shouldn't reinterpret int8_min as an unsigned type and then decide that // it is equal to 255. auto int8_min = std::numeric_limits::min(); - EXPECT_FALSE(Literal::CreateR0(255)->IsAll(int8_min)); + EXPECT_FALSE(LiteralUtil::CreateR0(255)->IsAll(int8_min)); - EXPECT_TRUE(Literal::CreateR0(42.0)->IsAll(42)); - EXPECT_FALSE(Literal::CreateR0(42.0001)->IsAll(42)); + EXPECT_TRUE(LiteralUtil::CreateR0(42.0)->IsAll(42)); + EXPECT_FALSE(LiteralUtil::CreateR0(42.0001)->IsAll(42)); - EXPECT_TRUE(Literal::CreateR1({100, 100, 100})->IsAll(100)); - EXPECT_FALSE(Literal::CreateR1({100, 100, 100.001})->IsAll(100)); + EXPECT_TRUE(LiteralUtil::CreateR1({100, 100, 100})->IsAll(100)); + EXPECT_FALSE(LiteralUtil::CreateR1({100, 100, 100.001})->IsAll(100)); - EXPECT_TRUE(Literal::CreateR2({{8, 8}, {8, 8}})->IsAll(8)); - EXPECT_FALSE(Literal::CreateR2({{8, 8}, {8, 9}})->IsAll(8)); - EXPECT_FALSE(Literal::CreateR2({{9, 8}, {8, 8}})->IsAll(8)); + EXPECT_TRUE(LiteralUtil::CreateR2({{8, 8}, {8, 8}})->IsAll(8)); + EXPECT_FALSE(LiteralUtil::CreateR2({{8, 8}, {8, 9}})->IsAll(8)); + EXPECT_FALSE(LiteralUtil::CreateR2({{9, 8}, {8, 8}})->IsAll(8)); half h8(8.0f); half h9(9.0f); - EXPECT_TRUE(Literal::CreateR2({{h8}, {h8}})->IsAll(8)); - EXPECT_FALSE(Literal::CreateR2({{h8}, {h9}})->IsAll(8)); - EXPECT_FALSE(Literal::CreateR2({{h9}, {h8}})->IsAll(8)); + EXPECT_TRUE(LiteralUtil::CreateR2({{h8}, {h8}})->IsAll(8)); + EXPECT_FALSE(LiteralUtil::CreateR2({{h8}, {h9}})->IsAll(8)); + EXPECT_FALSE(LiteralUtil::CreateR2({{h9}, {h8}})->IsAll(8)); bfloat16 b8(8.0f); bfloat16 b9(9.0f); - EXPECT_TRUE(Literal::CreateR2({{b8}, {b8}})->IsAll(8)); - EXPECT_FALSE(Literal::CreateR2({{b8}, {b9}})->IsAll(8)); - EXPECT_FALSE(Literal::CreateR2({{b9}, {b8}})->IsAll(8)); + EXPECT_TRUE(LiteralUtil::CreateR2({{b8}, {b8}})->IsAll(8)); + EXPECT_FALSE(LiteralUtil::CreateR2({{b8}, {b9}})->IsAll(8)); + EXPECT_FALSE(LiteralUtil::CreateR2({{b9}, {b8}})->IsAll(8)); // 9.001 will be truncated to 9.0 bfloat16 b91(9.001f); bfloat16 b90(9.00f); - EXPECT_TRUE(Literal::CreateR2({{b91}, {b90}})->IsAll(9.0)); + EXPECT_TRUE(LiteralUtil::CreateR2({{b91}, {b90}})->IsAll(9.0)); complex64 c8_9 = {8, 9}; - EXPECT_FALSE(Literal::CreateR2({{c8_9}, {c8_9}})->IsAll(8)); + EXPECT_FALSE(LiteralUtil::CreateR2({{c8_9}, {c8_9}})->IsAll(8)); auto uint64_max = std::numeric_limits::max(); - EXPECT_FALSE(Literal::CreateR2( + EXPECT_FALSE(LiteralUtil::CreateR2( {{uint64_max, uint64_max}, {uint64_max, uint64_max}}) ->IsAll(-1)); } TEST_F(LiteralUtilTest, IsAllFloat) { // IsAllFloat always returns false when the literal is not floating-point. - EXPECT_FALSE(Literal::CreateR0(false)->IsAllFloat(0)); - EXPECT_FALSE(Literal::CreateR0(0)->IsAllFloat(0)); - EXPECT_FALSE(Literal::CreateR0(0)->IsAllFloat(0)); - EXPECT_FALSE(Literal::CreateR0(0)->IsAllFloat(0)); - - EXPECT_TRUE(Literal::CreateR0(0)->IsAllFloat(0)); - EXPECT_TRUE(Literal::CreateR0(.5)->IsAllFloat(.5)); - EXPECT_TRUE(Literal::CreateR0(-.5)->IsAllFloat(-.5)); - EXPECT_FALSE(Literal::CreateR0(-.5)->IsAllFloat(-.49)); + EXPECT_FALSE(LiteralUtil::CreateR0(false)->IsAllFloat(0)); + EXPECT_FALSE(LiteralUtil::CreateR0(0)->IsAllFloat(0)); + EXPECT_FALSE(LiteralUtil::CreateR0(0)->IsAllFloat(0)); + EXPECT_FALSE(LiteralUtil::CreateR0(0)->IsAllFloat(0)); + + EXPECT_TRUE(LiteralUtil::CreateR0(0)->IsAllFloat(0)); + EXPECT_TRUE(LiteralUtil::CreateR0(.5)->IsAllFloat(.5)); + EXPECT_TRUE(LiteralUtil::CreateR0(-.5)->IsAllFloat(-.5)); + EXPECT_FALSE(LiteralUtil::CreateR0(-.5)->IsAllFloat(-.49)); EXPECT_FALSE( - Literal::CreateR2({{0, 0, 0}, {0, .1, 0}})->IsAllFloat(0)); - EXPECT_TRUE( - Literal::CreateR2({{.5, .5, .5}, {.5, .5, .5}})->IsAllFloat(.5)); - - EXPECT_TRUE(Literal::CreateR0(0)->IsAllFloat(0)); - EXPECT_TRUE(Literal::CreateR0(.5)->IsAllFloat(.5)); - EXPECT_TRUE(Literal::CreateR0(-.5)->IsAllFloat(-.5)); - EXPECT_FALSE(Literal::CreateR0(-.5)->IsAllFloat(-.49)); + LiteralUtil::CreateR2({{0, 0, 0}, {0, .1, 0}})->IsAllFloat(0)); + EXPECT_TRUE(LiteralUtil::CreateR2({{.5, .5, .5}, {.5, .5, .5}}) + ->IsAllFloat(.5)); + + EXPECT_TRUE(LiteralUtil::CreateR0(0)->IsAllFloat(0)); + EXPECT_TRUE(LiteralUtil::CreateR0(.5)->IsAllFloat(.5)); + EXPECT_TRUE(LiteralUtil::CreateR0(-.5)->IsAllFloat(-.5)); + EXPECT_FALSE(LiteralUtil::CreateR0(-.5)->IsAllFloat(-.49)); EXPECT_FALSE( - Literal::CreateR2({{0, 0, 0}, {0, .1, 0}})->IsAllFloat(0)); + LiteralUtil::CreateR2({{0, 0, 0}, {0, .1, 0}})->IsAllFloat(0)); } TEST_F(LiteralUtilTest, IsAllComplex) { // IsAllComplex always returns false when the literal is not complex. - EXPECT_FALSE(Literal::CreateR0(false)->IsAllComplex(0)); - EXPECT_FALSE(Literal::CreateR0(0)->IsAllComplex(0)); - EXPECT_FALSE(Literal::CreateR0(0)->IsAllComplex(0)); - EXPECT_FALSE(Literal::CreateR0(0)->IsAllComplex(0)); - EXPECT_FALSE(Literal::CreateR0(0)->IsAllComplex(0)); - EXPECT_FALSE(Literal::CreateR0(0)->IsAllComplex(0)); + EXPECT_FALSE(LiteralUtil::CreateR0(false)->IsAllComplex(0)); + EXPECT_FALSE(LiteralUtil::CreateR0(0)->IsAllComplex(0)); + EXPECT_FALSE(LiteralUtil::CreateR0(0)->IsAllComplex(0)); + EXPECT_FALSE(LiteralUtil::CreateR0(0)->IsAllComplex(0)); + EXPECT_FALSE(LiteralUtil::CreateR0(0)->IsAllComplex(0)); + EXPECT_FALSE(LiteralUtil::CreateR0(0)->IsAllComplex(0)); complex64 c8_9 = {8, 9}; complex64 c7_9 = {7, 9}; - EXPECT_TRUE(Literal::CreateR2({{c8_9}, {c8_9}}) + EXPECT_TRUE(LiteralUtil::CreateR2({{c8_9}, {c8_9}}) ->IsAllComplex({8.0f, 9.0f})); - EXPECT_FALSE(Literal::CreateR2({{c7_9}, {c8_9}}) + EXPECT_FALSE(LiteralUtil::CreateR2({{c7_9}, {c8_9}}) ->IsAllComplex({8.0f, 9.0f})); - EXPECT_FALSE(Literal::CreateR2({{c8_9}, {c7_9}}) + EXPECT_FALSE(LiteralUtil::CreateR2({{c8_9}, {c7_9}}) ->IsAllComplex({8.0f, 9.0f})); } TEST_F(LiteralUtilTest, IsAllFirst) { // IsAllComplex always returns false when the literal is not complex. - EXPECT_FALSE(Literal::CreateR1({false, true})->IsAllFirst()); - EXPECT_TRUE(Literal::CreateR1({false, false})->IsAllFirst()); - EXPECT_FALSE(Literal::CreateR1({1, 1, 2})->IsAllFirst()); - EXPECT_TRUE(Literal::CreateR1({5, 5, 5, 5})->IsAllFirst()); - EXPECT_FALSE(Literal::CreateR1({1, 1, 2})->IsAllFirst()); - EXPECT_TRUE(Literal::CreateR1({5, 5, 5, 5})->IsAllFirst()); - EXPECT_FALSE(Literal::CreateR1({1, 1, 2})->IsAllFirst()); - EXPECT_TRUE(Literal::CreateR1({5, 5, 5, 5})->IsAllFirst()); - EXPECT_FALSE(Literal::CreateR1({1, 1, 2})->IsAllFirst()); + EXPECT_FALSE(LiteralUtil::CreateR1({false, true})->IsAllFirst()); + EXPECT_TRUE(LiteralUtil::CreateR1({false, false})->IsAllFirst()); + EXPECT_FALSE(LiteralUtil::CreateR1({1, 1, 2})->IsAllFirst()); + EXPECT_TRUE(LiteralUtil::CreateR1({5, 5, 5, 5})->IsAllFirst()); + EXPECT_FALSE(LiteralUtil::CreateR1({1, 1, 2})->IsAllFirst()); + EXPECT_TRUE(LiteralUtil::CreateR1({5, 5, 5, 5})->IsAllFirst()); + EXPECT_FALSE(LiteralUtil::CreateR1({1, 1, 2})->IsAllFirst()); + EXPECT_TRUE(LiteralUtil::CreateR1({5, 5, 5, 5})->IsAllFirst()); + EXPECT_FALSE(LiteralUtil::CreateR1({1, 1, 2})->IsAllFirst()); complex64 c8_9 = {8, 9}; complex64 c7_9 = {7, 9}; - EXPECT_TRUE(Literal::CreateR2({{c8_9}, {c8_9}})->IsAllFirst()); - EXPECT_FALSE(Literal::CreateR2({{c7_9}, {c8_9}})->IsAllFirst()); + EXPECT_TRUE(LiteralUtil::CreateR2({{c8_9}, {c8_9}})->IsAllFirst()); + EXPECT_FALSE( + LiteralUtil::CreateR2({{c7_9}, {c8_9}})->IsAllFirst()); } TEST_F(LiteralUtilTest, IsZero) { - auto scalar_zero = Literal::CreateR0(0.0f); - auto scalar_one = Literal::CreateR0(1.0f); + auto scalar_zero = LiteralUtil::CreateR0(0.0f); + auto scalar_one = LiteralUtil::CreateR0(1.0f); EXPECT_TRUE(scalar_zero->IsZero({})); EXPECT_FALSE(scalar_one->IsZero({})); - auto array = Literal::CreateR2({{1, 2, 0, 3}, {1, 0, 1, 2}}); + auto array = LiteralUtil::CreateR2({{1, 2, 0, 3}, {1, 0, 1, 2}}); EXPECT_FALSE(array->IsZero({0, 1})); EXPECT_TRUE(array->IsZero({0, 2})); EXPECT_TRUE(array->IsZero({1, 1})); EXPECT_FALSE(array->IsZero({1, 2})); - auto complex_zero = Literal::CreateR0(0.0f); - auto complex_nonzero = Literal::CreateR0(0.5f); + auto complex_zero = LiteralUtil::CreateR0(0.0f); + auto complex_nonzero = LiteralUtil::CreateR0(0.5f); EXPECT_TRUE(complex_zero->IsZero({})); EXPECT_FALSE(complex_nonzero->IsZero({})); } @@ -563,7 +570,7 @@ TYPED_TEST_CASE(LiteralUtilTestTemplated, TestedTypes); TYPED_TEST(LiteralUtilTestTemplated, Relayout2x2) { // Make a non-integer for floating point types. TypeParam half = TypeParam(1) / TypeParam(2); - auto data = Literal::CreateR2({{half, 2}, {3, 4}}); + auto data = LiteralUtil::CreateR2({{half, 2}, {3, 4}}); const Layout layout01 = LayoutUtil::MakeLayout({0, 1}); const Layout layout10 = LayoutUtil::MakeLayout({1, 0}); @@ -577,7 +584,7 @@ TYPED_TEST(LiteralUtilTestTemplated, Relayout2x2) { } TEST_F(LiteralUtilTest, ReshapeR0) { - auto original = Literal::CreateR0(1.7f); + auto original = LiteralUtil::CreateR0(1.7f); auto reshape = original->Reshape(/*dimensions=*/{}).ConsumeValueOrDie(); EXPECT_EQ(*original, *reshape); } @@ -585,13 +592,13 @@ TEST_F(LiteralUtilTest, ReshapeR0) { TEST_F(LiteralUtilTest, ReshapeR4) { // clang-format off // F32[1x3x2x4] - auto original = Literal::CreateR4WithLayout({{ + auto original = LiteralUtil::CreateR4WithLayout({{ {{10, 11, 12, 13}, {14, 15, 16, 17}}, {{18, 19, 20, 21}, {22, 23, 24, 25}}, {{26, 27, 28, 29}, {30, 31, 32, 33}}, }}, layout_r4_dim0major_); // F32[1x3x4x2] - auto expected = Literal::CreateR3WithLayout({ + auto expected = LiteralUtil::CreateR3WithLayout({ {{10, 11}, {12, 13}, {14, 15}, {16, 17}}, {{18, 19}, {20, 21}, {22, 23}, {24, 25}}, {{26, 27}, {28, 29}, {30, 31}, {32, 33}}, @@ -605,13 +612,13 @@ TEST_F(LiteralUtilTest, ReshapeR4) { TEST_F(LiteralUtilTest, ReshapeR4Dim0Minor) { // clang-format off // F32[1x3x2x4] - auto original = Literal::CreateR4WithLayout({{ + auto original = LiteralUtil::CreateR4WithLayout({{ {{10, 11, 12, 13}, {14, 15, 16, 17}}, {{18, 19, 20, 21}, {22, 23, 24, 25}}, {{26, 27, 28, 29}, {30, 31, 32, 33}}, }}, layout_r4_dim0minor_); // F32[1x3x4x2] - auto expected = Literal::CreateR3WithLayout({ + auto expected = LiteralUtil::CreateR3WithLayout({ {{10, 11}, {12, 13}, {14, 15}, {16, 17}}, {{18, 19}, {20, 21}, {22, 23}, {24, 25}}, {{26, 27}, {28, 29}, {30, 31}, {32, 33}}, @@ -623,7 +630,7 @@ TEST_F(LiteralUtilTest, ReshapeR4Dim0Minor) { } TEST_F(LiteralUtilTest, TransposeR0) { - auto original = Literal::CreateR0(1.7f); + auto original = LiteralUtil::CreateR0(1.7f); auto reshape = original->Transpose(/*permutation=*/{}); EXPECT_EQ(*original, *reshape); } @@ -631,7 +638,7 @@ TEST_F(LiteralUtilTest, TransposeR0) { TEST_F(LiteralUtilTest, TransposeR4) { // clang-format off // F32[1x3x2x4] - auto original = Literal::CreateR4({{ + auto original = LiteralUtil::CreateR4({{ {{10, 11, 12, 13}, {14, 15, 16, 17}}, {{18, 19, 20, 21}, {22, 23, 24, 25}}, {{26, 27, 28, 29}, {30, 31, 32, 33}}, @@ -659,7 +666,7 @@ TEST_F(LiteralUtilTest, TestR4RelayoutEquivalence) { TEST_F(LiteralUtilTest, TestR2LinearLayout) { // Test expected memory layout of R2 dim0-minor (column-major) literal. - auto mat_dim0minor = Literal::CreateR2WithLayout( + auto mat_dim0minor = LiteralUtil::CreateR2WithLayout( {{1, 2, 3}, {4, 5, 6}}, layout_r2_dim0minor_); EXPECT_EQ(mat_dim0minor->element_count(), 6); EXPECT_THAT(mat_dim0minor->data(), ElementsAre(1, 4, 2, 5, 3, 6)); @@ -670,7 +677,7 @@ TEST_F(LiteralUtilTest, TestR2LinearLayout) { ElementsAre(1, 2, 3, 4, 5, 6)); // Test expected memory layout of R2 created with dim0-major (row-major). - auto mat_dim0major = Literal::CreateR2WithLayout( + auto mat_dim0major = LiteralUtil::CreateR2WithLayout( {{1, 2, 3}, {4, 5, 6}}, layout_r2_dim0major_); EXPECT_EQ(mat_dim0major->element_count(), 6); EXPECT_THAT(mat_dim0major->data(), ElementsAre(1, 2, 3, 4, 5, 6)); @@ -695,8 +702,8 @@ TEST_F(LiteralUtilTest, TestR3LinearLayout) { {10, 11, 12}, }, }); // clang-format on - auto lit_dim0minor = - Literal::CreateR3FromArray3DWithLayout(arr3d, layout_r3_dim0minor_); + auto lit_dim0minor = LiteralUtil::CreateR3FromArray3DWithLayout( + arr3d, layout_r3_dim0minor_); EXPECT_EQ(lit_dim0minor->element_count(), 12); std::vector expected_dim0minor{1, 7, 4, 10, 2, 8, 5, 11, 3, 9, 6, 12}; @@ -710,8 +717,8 @@ TEST_F(LiteralUtilTest, TestR3LinearLayout) { testing::ElementsAreArray(expected_dim0major)); // Test expected memory layout of R3 created with dim0-major (row-major). - auto lit_dim0major = - Literal::CreateR3FromArray3DWithLayout(arr3d, layout_r3_dim0major_); + auto lit_dim0major = LiteralUtil::CreateR3FromArray3DWithLayout( + arr3d, layout_r3_dim0major_); EXPECT_EQ(lit_dim0major->element_count(), 12); EXPECT_THAT(lit_dim0major->data(), testing::ElementsAreArray(expected_dim0major)); @@ -723,28 +730,28 @@ TEST_F(LiteralUtilTest, TestR3LinearLayout) { } TEST_F(LiteralUtilTest, SliceR0S32) { - auto input = Literal::CreateR0(1); + auto input = LiteralUtil::CreateR0(1); auto result = input->Slice({}, {}); EXPECT_EQ(*input, *result); } TEST_F(LiteralUtilTest, SliceR1F32) { - auto input = Literal::CreateR1({1.0, 2.0, 3.0, 4.0, 5.0}); + auto input = LiteralUtil::CreateR1({1.0, 2.0, 3.0, 4.0, 5.0}); auto result = input->Slice({3}, {4}); - auto expected = Literal::CreateR1({4.0}); + auto expected = LiteralUtil::CreateR1({4.0}); EXPECT_EQ(*expected, *result); } TEST_F(LiteralUtilTest, SliceR2U32) { - auto input_3x4 = - Literal::CreateR2({{1, 2, 3, 4}, {5, 6, 7, 8}, {9, 10, 11, 12}}); + auto input_3x4 = LiteralUtil::CreateR2( + {{1, 2, 3, 4}, {5, 6, 7, 8}, {9, 10, 11, 12}}); auto result = input_3x4->Slice({0, 2}, {2, 4}); - auto expected = Literal::CreateR2({{3, 4}, {7, 8}}); + auto expected = LiteralUtil::CreateR2({{3, 4}, {7, 8}}); EXPECT_EQ(*expected, *result); } TEST_F(LiteralUtilTest, SliceR3U32Full) { - auto input_2x3x2 = Literal::CreateR3( + auto input_2x3x2 = LiteralUtil::CreateR3( {{{1, 2}, {3, 4}, {5, 6}}, {{7, 8}, {9, 10}, {11, 12}}}); auto result = input_2x3x2->Slice({0, 0, 0}, {2, 3, 2}); EXPECT_EQ(*input_2x3x2, *result); @@ -753,21 +760,21 @@ TEST_F(LiteralUtilTest, SliceR3U32Full) { TEST_F(LiteralUtilTest, PopulateR1S64) { Literal output(ShapeUtil::MakeShape(S64, {1})); output.PopulateR1({77}); - auto expected = Literal::CreateR1({77}); + auto expected = LiteralUtil::CreateR1({77}); EXPECT_EQ(output, *expected); } TEST_F(LiteralUtilTest, PopulateR1U64) { Literal output(ShapeUtil::MakeShape(U64, {2})); output.PopulateR1({{77, 88}}); - auto expected = Literal::CreateR1({{77, 88}}); + auto expected = LiteralUtil::CreateR1({{77, 88}}); EXPECT_EQ(output, *expected); } TEST_F(LiteralUtilTest, PopulateR1C64) { Literal output(ShapeUtil::MakeShape(C64, {1})); output.PopulateR1({{77, 88}}); - auto expected = Literal::CreateR1({{77, 88}}); + auto expected = LiteralUtil::CreateR1({{77, 88}}); EXPECT_EQ(output, *expected); } @@ -775,7 +782,7 @@ TEST_F(LiteralUtilTest, PopulateR2C64) { Literal output(ShapeUtil::MakeShape(C64, {2, 2})); output.PopulateR2({{{7, 8}, {9, 10}}, {{1, 2}, {3, 4}}}); auto expected = - Literal::CreateR2({{{7, 8}, {9, 10}}, {{1, 2}, {3, 4}}}); + LiteralUtil::CreateR2({{{7, 8}, {9, 10}}, {{1, 2}, {3, 4}}}); EXPECT_EQ(output, *expected); } @@ -783,7 +790,7 @@ TEST_F(LiteralUtilTest, PopulateWithValueR0BF16) { Literal output(ShapeUtil::MakeShape(BF16, {})); bfloat16 h(0.25f); output.PopulateWithValue(h); - auto expected = Literal::CreateR0(h); + auto expected = LiteralUtil::CreateR0(h); EXPECT_EQ(output, *expected); } @@ -791,7 +798,7 @@ TEST_F(LiteralUtilTest, PopulateWithValueR1BF16) { Literal output(ShapeUtil::MakeShape(BF16, {3})); bfloat16 h(0.5f); output.PopulateWithValue(h); - auto expected = Literal::CreateR1({h, h, h}); + auto expected = LiteralUtil::CreateR1({h, h, h}); EXPECT_EQ(output, *expected); } @@ -799,28 +806,28 @@ TEST_F(LiteralUtilTest, PopulateWithValueR2BF16) { Literal output(ShapeUtil::MakeShape(BF16, {2, 2})); bfloat16 h(2.0f); output.PopulateWithValue(h); - auto expected = Literal::CreateR2({{h, h}, {h, h}}); + auto expected = LiteralUtil::CreateR2({{h, h}, {h, h}}); EXPECT_EQ(output, *expected); } TEST_F(LiteralUtilTest, PopulateWithValueR0F32) { Literal output(ShapeUtil::MakeShape(F32, {})); output.PopulateWithValue(2.5f); - auto expected = Literal::CreateR0(2.5f); + auto expected = LiteralUtil::CreateR0(2.5f); EXPECT_EQ(output, *expected); } TEST_F(LiteralUtilTest, PopulateWithValueR1S64) { Literal output(ShapeUtil::MakeShape(S64, {3})); output.PopulateWithValue(-7); - auto expected = Literal::CreateR1({-7, -7, -7}); + auto expected = LiteralUtil::CreateR1({-7, -7, -7}); EXPECT_EQ(output, *expected); } TEST_F(LiteralUtilTest, PopulateWithValueR2U64) { Literal output(ShapeUtil::MakeShape(U64, {2, 2})); output.PopulateWithValue(42); - auto expected = Literal::CreateR2({{42, 42}, {42, 42}}); + auto expected = LiteralUtil::CreateR2({{42, 42}, {42, 42}}); EXPECT_EQ(output, *expected); } @@ -828,7 +835,7 @@ TEST_F(LiteralUtilTest, PopulateWithValueR2C64) { Literal output(ShapeUtil::MakeShape(C64, {2, 2})); output.PopulateWithValue({4, 2}); auto expected = - Literal::CreateR2({{{4, 2}, {4, 2}}, {{4, 2}, {4, 2}}}); + LiteralUtil::CreateR2({{{4, 2}, {4, 2}}, {{4, 2}, {4, 2}}}); EXPECT_EQ(output, *expected); } @@ -836,7 +843,7 @@ TEST_F(LiteralUtilTest, PopulateWithValueR0F16) { Literal output(ShapeUtil::MakeShape(F16, {})); half h(0.25f); output.PopulateWithValue(h); - auto expected = Literal::CreateR0(h); + auto expected = LiteralUtil::CreateR0(h); EXPECT_EQ(output, *expected); } @@ -844,7 +851,7 @@ TEST_F(LiteralUtilTest, PopulateWithValueR1F16) { Literal output(ShapeUtil::MakeShape(F16, {3})); half h(0.5f); output.PopulateWithValue(h); - auto expected = Literal::CreateR1({h, h, h}); + auto expected = LiteralUtil::CreateR1({h, h, h}); EXPECT_EQ(output, *expected); } @@ -852,15 +859,15 @@ TEST_F(LiteralUtilTest, PopulateWithValueR2F16) { Literal output(ShapeUtil::MakeShape(F16, {2, 2})); half h(2.0f); output.PopulateWithValue(h); - auto expected = Literal::CreateR2({{h, h}, {h, h}}); + auto expected = LiteralUtil::CreateR2({{h, h}, {h, h}}); EXPECT_EQ(output, *expected); } TEST_F(LiteralUtilTest, ReplicateR2U32) { - auto input = - Literal::CreateR2({{1, 2, 3, 4}, {5, 6, 7, 8}, {9, 10, 11, 12}}); + auto input = LiteralUtil::CreateR2( + {{1, 2, 3, 4}, {5, 6, 7, 8}, {9, 10, 11, 12}}); auto output = input->Replicate(3); - auto expected = Literal::CreateR3( + auto expected = LiteralUtil::CreateR3( {{{1, 2, 3, 4}, {5, 6, 7, 8}, {9, 10, 11, 12}}, {{1, 2, 3, 4}, {5, 6, 7, 8}, {9, 10, 11, 12}}, {{1, 2, 3, 4}, {5, 6, 7, 8}, {9, 10, 11, 12}}}); @@ -914,12 +921,12 @@ TEST_F(LiteralUtilTest, CopySliceFrom) { } TEST_F(LiteralUtilTest, CopyFromScalars) { - auto zero = Literal::CreateR0(0); - auto nine = Literal::CreateR0(9); + auto zero = LiteralUtil::CreateR0(0); + auto nine = LiteralUtil::CreateR0(9); TF_EXPECT_OK(zero->CopyFrom(*nine)); EXPECT_EQ(*zero, *nine); - auto vect = Literal::CreateR1({3, 4, 9, 12, 5, 17, 21}); + auto vect = LiteralUtil::CreateR1({3, 4, 9, 12, 5, 17, 21}); TF_EXPECT_OK(zero->CopySliceFrom(*vect, {5}, {}, {})); EXPECT_EQ(zero->Get({}), 17); TF_EXPECT_OK(vect->CopySliceFrom(*zero, {}, {4}, {})); @@ -928,13 +935,13 @@ TEST_F(LiteralUtilTest, CopyFromScalars) { TEST_F(LiteralUtilTest, CopyFromAndToZeroElement) { const Shape empty_r1_shape = ShapeUtil::MakeShape(F32, {0}); - const auto const_nine = Literal::CreateR1({9}); + const auto const_nine = LiteralUtil::CreateR1({9}); const auto const_empty = Literal::CreateFromShape(empty_r1_shape); { // Source contains dimension with zero elements. const auto empty = Literal::CreateFromShape(empty_r1_shape); - auto nine = Literal::CreateR1({9}); + auto nine = LiteralUtil::CreateR1({9}); TF_EXPECT_OK(nine->CopySliceFrom(*empty, {0}, {0}, {0})); EXPECT_EQ(*nine, *const_nine); @@ -943,7 +950,7 @@ TEST_F(LiteralUtilTest, CopyFromAndToZeroElement) { { // Copy 0 element to destination with zero elements. const auto empty = Literal::CreateFromShape(empty_r1_shape); - auto nine = Literal::CreateR1({9}); + auto nine = LiteralUtil::CreateR1({9}); TF_EXPECT_OK(empty->CopySliceFrom(*nine, {0}, {0}, {0})); EXPECT_EQ(*empty, *const_empty); @@ -958,16 +965,16 @@ TEST_F(LiteralUtilTest, CopyFromNilShape) { } TEST_F(LiteralUtilTest, CopyFromArrays) { - auto scalar_42 = Literal::CreateR0(42.0); - auto scalar_123 = Literal::CreateR0(123.0); + auto scalar_42 = LiteralUtil::CreateR0(42.0); + auto scalar_123 = LiteralUtil::CreateR0(123.0); EXPECT_NE(*scalar_42, *scalar_123); TF_ASSERT_OK(scalar_42->CopyFrom(*scalar_123, /*dest_shape_index=*/{}, /*src_shape_index=*/{})); EXPECT_EQ(*scalar_42, *scalar_123); EXPECT_EQ(scalar_42->Get({}), 123.0f); - auto matrix_1234 = Literal::CreateR2({{1.0, 2.0}, {3.0, 4.0}}); - auto matrix_5678 = Literal::CreateR2({{5.0, 6.0}, {7.0, 8.0}}); + auto matrix_1234 = LiteralUtil::CreateR2({{1.0, 2.0}, {3.0, 4.0}}); + auto matrix_5678 = LiteralUtil::CreateR2({{5.0, 6.0}, {7.0, 8.0}}); EXPECT_NE(*matrix_1234, *matrix_5678); EXPECT_EQ(matrix_1234->Get({0, 0}), 1.0f); TF_ASSERT_OK(matrix_1234->CopyFrom(*matrix_5678, /*dest_shape_index=*/{}, @@ -977,19 +984,19 @@ TEST_F(LiteralUtilTest, CopyFromArrays) { } TEST_F(LiteralUtilTest, CopyFromTuples) { - auto matrix = Literal::CreateR2({{1.0, 2.0}, {3.0, 4.0}}); + auto matrix = LiteralUtil::CreateR2({{1.0, 2.0}, {3.0, 4.0}}); Literal nil_literal(ShapeUtil::MakeNil()); - auto nested_tuple = Literal::MakeTuple( + auto nested_tuple = LiteralUtil::MakeTuple( {matrix.get(), - Literal::MakeTuple({Literal::CreateR0(42).get(), - Literal::CreateR1({23.0, 44.0}).get(), - &nil_literal}) + LiteralUtil::MakeTuple( + {LiteralUtil::CreateR0(42).get(), + LiteralUtil::CreateR1({23.0, 44.0}).get(), &nil_literal}) .get()}); // Create a tuple the same shape as the inner tuple of nested_tuple but with // different values.. - auto tuple = Literal::MakeTuple({Literal::CreateR0(-5).get(), - Literal::CreateR1({2.0, 4.0}).get(), - &nil_literal}); + auto tuple = LiteralUtil::MakeTuple( + {LiteralUtil::CreateR0(-5).get(), + LiteralUtil::CreateR1({2.0, 4.0}).get(), &nil_literal}); EXPECT_EQ(*matrix, LiteralSlice(*nested_tuple, {0})); EXPECT_EQ(nested_tuple->Get({}, {1, 0}), 42); @@ -1010,8 +1017,8 @@ TEST_F(LiteralUtilTest, CopyFromTuples) { EXPECT_EQ(nested_tuple->Get({1}, {1, 1}), 4.0); } TEST_F(LiteralUtilTest, CopyBetweenSameTuple) { - auto tuple = Literal::MakeTuple( - {Literal::CreateR0(-2).get(), Literal::CreateR0(4).get()}); + auto tuple = LiteralUtil::MakeTuple({LiteralUtil::CreateR0(-2).get(), + LiteralUtil::CreateR0(4).get()}); EXPECT_EQ(tuple->Get({}, {0}), -2); EXPECT_EQ(tuple->Get({}, {1}), 4); @@ -1025,8 +1032,8 @@ TEST_F(LiteralUtilTest, CopyBetweenSameTuple) { } TEST_F(LiteralUtilTest, CopyFromDifferentShapes) { - auto matrix = Literal::CreateR2({{1.0, 2.0}, {3.0, 4.0}}); - auto vector = Literal::CreateR1({5.0, 7.0}); + auto matrix = LiteralUtil::CreateR2({{1.0, 2.0}, {3.0, 4.0}}); + auto vector = LiteralUtil::CreateR1({5.0, 7.0}); Status status = matrix->CopyFrom(*vector); ASSERT_FALSE(status.ok()); ASSERT_THAT(status.error_message(), @@ -1051,7 +1058,7 @@ TEST_F(LiteralUtilTest, F16) { half h1(1.0f); half h2(2.0f); - auto m2 = Literal::CreateR2({{h1, h2}, {h2, h1}}); + auto m2 = LiteralUtil::CreateR2({{h1, h2}, {h2, h1}}); Literal* l2 = m2.get(); const char* d2 = reinterpret_cast(l2->data().data()); EXPECT_EQ(d2[0], 0); @@ -1150,12 +1157,12 @@ TEST_F(LiteralUtilTest, PopulateParallel) { TEST_F(LiteralUtilTest, ConvertR4) { // clang-format off - auto original = Literal::CreateR4WithLayout({{ + auto original = LiteralUtil::CreateR4WithLayout({{ {{10, 11, 12, 13}, {14, 15, 16, 17}}, {{18, 19, 20, 21}, {22, 23, 24, 25}}, {{26, 27, 28, 29}, {30, 31, 32, 33}}, }}, layout_r4_dim0major_); - auto expected = Literal::CreateR4WithLayout({{ + auto expected = LiteralUtil::CreateR4WithLayout({{ {{10, 11, 12, 13}, {14, 15, 16, 17}}, {{18, 19, 20, 21}, {22, 23, 24, 25}}, {{26, 27, 28, 29}, {30, 31, 32, 33}}, @@ -1169,42 +1176,42 @@ TEST_F(LiteralUtilTest, ConvertR4) { TEST_F(LiteralUtilTest, ConvertIfTypesMatch) { // clang-format off - auto s8 = Literal::CreateR4WithLayout({{ + auto s8 = LiteralUtil::CreateR4WithLayout({{ {{10, 0, 12, 0}, {0, 15, 0, 17}}, {{0, 19, 0, 21}, {22, 0, 24, 0}}, {{26, 0, 28, 0}, {0, 31, 0, 33}}, }}, layout_r4_dim0major_); - auto s32 = Literal::CreateR4WithLayout({{ + auto s32 = LiteralUtil::CreateR4WithLayout({{ {{10, 0, 12, 0}, {0, 15, 0, 17}}, {{0, 19, 0, 21}, {22, 0, 24, 0}}, {{26, 0, 28, 0}, {0, 31, 0, 33}}, }}, layout_r4_dim0major_); - auto u32 = Literal::CreateR4WithLayout({{ + auto u32 = LiteralUtil::CreateR4WithLayout({{ {{10, 0, 12, 0}, {0, 15, 0, 17}}, {{0, 19, 0, 21}, {22, 0, 24, 0}}, {{26, 0, 28, 0}, {0, 31, 0, 33}}, }}, layout_r4_dim0major_); - auto s64 = Literal::CreateR4WithLayout({{ + auto s64 = LiteralUtil::CreateR4WithLayout({{ {{10, 0, 12, 0}, {0, 15, 0, 17}}, {{0, 19, 0, 21}, {22, 0, 24, 0}}, {{26, 0, 28, 0}, {0, 31, 0, 33}}, }}, layout_r4_dim0major_); - auto u64 = Literal::CreateR4WithLayout({{ + auto u64 = LiteralUtil::CreateR4WithLayout({{ {{10, 0, 12, 0}, {0, 15, 0, 17}}, {{0, 19, 0, 21}, {22, 0, 24, 0}}, {{26, 0, 28, 0}, {0, 31, 0, 33}}, }}, layout_r4_dim0major_); - auto pred = Literal::CreateR4WithLayout({{ + auto pred = LiteralUtil::CreateR4WithLayout({{ {{true, false, true, false}, {false, true, false, true}}, {{false, true, false, true}, {true, false, true, false}}, {{true, false, true, false}, {false, true, false, true}}, }}, layout_r4_dim0major_); - auto int32_pred = Literal::CreateR4WithLayout({{ + auto int32_pred = LiteralUtil::CreateR4WithLayout({{ {{1, 0, 1, 0}, {0, 1, 0, 1}}, {{0, 1, 0, 1}, {1, 0, 1, 0}}, {{1, 0, 1, 0}, {0, 1, 0, 1}}, }}, layout_r4_dim0major_); - auto f16 = Literal::CreateR4WithLayout({{ + auto f16 = LiteralUtil::CreateR4WithLayout({{ {{half(10.0), half(0.0), half(12.0), half(0.0)}, {half(0.0), half(15.0), half(0.0), half(17.0)}}, {{half(0.0), half(19.0), half(0.0), half(21.0)}, @@ -1212,7 +1219,7 @@ TEST_F(LiteralUtilTest, ConvertIfTypesMatch) { {{half(26.0), half(0.0), half(28.0), half(0.0)}, {half(0.0), half(31.0), half(0.0), half(33.0)}}, }}, layout_r4_dim0major_); - auto bf16 = Literal::CreateR4WithLayout({{ + auto bf16 = LiteralUtil::CreateR4WithLayout({{ {{bfloat16(10.0), bfloat16(0.0), bfloat16(12.0), bfloat16(0.0)}, {bfloat16(0.0), bfloat16(15.0), bfloat16(0.0), bfloat16(17.0)}}, {{bfloat16(0.0), bfloat16(19.0), bfloat16(0.0), bfloat16(21.0)}, @@ -1220,17 +1227,17 @@ TEST_F(LiteralUtilTest, ConvertIfTypesMatch) { {{bfloat16(26.0), bfloat16(0.0), bfloat16(28.0), bfloat16(0.0)}, {bfloat16(0.0), bfloat16(31.0), bfloat16(0.0), bfloat16(33.0)}}, }}, layout_r4_dim0major_); - auto f32 = Literal::CreateR4WithLayout({{ + auto f32 = LiteralUtil::CreateR4WithLayout({{ {{10.0f, 0.0f, 12.0f, 0.0f}, {0.0f, 15.0f, 0.0f, 17.0f}}, {{0.0f, 19.0f, 0.0f, 21.0f}, {22.0f, 0.0f, 24.0f, 0.0f}}, {{26.0f, 0.0f, 28.0f, 0.0f}, {0.0f, 31.0f, 0.0f, 33.0f}}, }}, layout_r4_dim0major_); - auto f64 = Literal::CreateR4WithLayout({{ + auto f64 = LiteralUtil::CreateR4WithLayout({{ {{10.0, 0.0, 12.0, 0.0}, {0.0, 15.0, 0.0, 17.0}}, {{0.0, 19.0, 0.0, 21.0}, {22.0, 0.0, 24.0, 0.0}}, {{26.0, 0.0, 28.0, 0.0}, {0.0, 31.0, 0.0, 33.0}}, }}, layout_r4_dim0major_); - auto c64 = Literal::CreateR4WithLayout({{ + auto c64 = LiteralUtil::CreateR4WithLayout({{ {{10.0f, 0.0f, 12.0f, 0.0f}, {0.0f, 15.0f, 0.0f, 17.0f}}, {{0.0f, 19.0f, 0.0f, 21.0f}, {22.0f, 0.0f, 24.0f, 0.0f}}, {{26.0f, 0.0f, 28.0f, 0.0f}, {0.0f, 31.0f, 0.0f, 33.0f}}, @@ -1302,18 +1309,18 @@ TEST_F(LiteralUtilTest, ConvertIfTypesMatch) { } TEST_F(LiteralUtilTest, BitcastConvert) { - auto original = - Literal::CreateR1({tensorflow::bit_cast(2.5f), - tensorflow::bit_cast(-42.25f), - tensorflow::bit_cast(100.f), 0xbeef}); - auto expected = Literal::CreateR1( + auto original = LiteralUtil::CreateR1( + {tensorflow::bit_cast(2.5f), + tensorflow::bit_cast(-42.25f), + tensorflow::bit_cast(100.f), 0xbeef}); + auto expected = LiteralUtil::CreateR1( {2.5f, -42.25f, 100.0f, tensorflow::bit_cast(0xbeef)}); TF_ASSERT_OK_AND_ASSIGN(std::unique_ptr converted, original->BitcastConvert(F32)); } TEST_F(LiteralUtilTest, BitcastConvertBetweenInvalidTypes) { - auto literal = Literal::CreateR0(1234); + auto literal = LiteralUtil::CreateR0(1234); Status status = literal->BitcastConvert(F64).status(); EXPECT_NE(Status::OK(), status); EXPECT_TRUE(tensorflow::str_util::StrContains(status.error_message(), @@ -1348,7 +1355,7 @@ TEST_F(LiteralUtilTest, ToProto_f16) { half h1(1.0f); half h2(2.0f); - auto m = Literal::CreateR2({{h1, h2}, {h2, h1}}); + auto m = LiteralUtil::CreateR2({{h1, h2}, {h2, h1}}); Literal* l = m.get(); EXPECT_EQ(4, ShapeUtil::ElementsIn(l->shape())); EXPECT_EQ(4, l->data().size()); @@ -1391,10 +1398,10 @@ TEST_F(LiteralUtilTest, CopyFromProto_f16) { } TEST_F(LiteralUtilTest, LiteralSliceTest) { - auto scalar = Literal::CreateR0(1.0); - auto matrix = Literal::CreateR2({{1.0, 2.0}, {3.0, 4.0}}); - auto tuple = Literal::MakeTuple({scalar.get(), matrix.get()}); - auto nested_tuple = Literal::MakeTuple({tuple.get(), scalar.get()}); + auto scalar = LiteralUtil::CreateR0(1.0); + auto matrix = LiteralUtil::CreateR2({{1.0, 2.0}, {3.0, 4.0}}); + auto tuple = LiteralUtil::MakeTuple({scalar.get(), matrix.get()}); + auto nested_tuple = LiteralUtil::MakeTuple({tuple.get(), scalar.get()}); Literal nil(ShapeUtil::MakeNil()); EXPECT_EQ(LiteralSlice(*scalar, {}), *scalar); @@ -1413,10 +1420,10 @@ TEST_F(LiteralUtilTest, LiteralSliceTest) { } TEST_F(LiteralUtilTest, MutatingLiteralSlice) { - auto scalar = Literal::CreateR0(1.0); - auto matrix = Literal::CreateR2({{1.0, 2.0}, {3.0, 4.0}}); - auto tuple = Literal::MakeTuple({scalar.get(), matrix.get()}); - auto nested_tuple = Literal::MakeTuple({tuple.get(), scalar.get()}); + auto scalar = LiteralUtil::CreateR0(1.0); + auto matrix = LiteralUtil::CreateR2({{1.0, 2.0}, {3.0, 4.0}}); + auto tuple = LiteralUtil::MakeTuple({scalar.get(), matrix.get()}); + auto nested_tuple = LiteralUtil::MakeTuple({tuple.get(), scalar.get()}); // Verify that changing the underlying data beneath the view changes the // data of the view itself. const auto nested_tuple_view = LiteralSlice(*nested_tuple); @@ -1436,15 +1443,16 @@ TEST_F(LiteralUtilTest, MutatingLiteralSlice) { } TEST_F(LiteralUtilTest, LiteralSliceOfALiteralSlice) { - auto scalar = Literal::CreateR0(1.0); - auto matrix = Literal::CreateR2({{1.0, 2.0}, {3.0, 4.0}}); - auto tuple = Literal::MakeTuple({scalar.get(), matrix.get()}); - auto nested_tuple = Literal::MakeTuple({tuple.get(), scalar.get()}); + auto scalar = LiteralUtil::CreateR0(1.0); + auto matrix = LiteralUtil::CreateR2({{1.0, 2.0}, {3.0, 4.0}}); + auto tuple = LiteralUtil::MakeTuple({scalar.get(), matrix.get()}); + auto nested_tuple = LiteralUtil::MakeTuple({tuple.get(), scalar.get()}); const auto nested_tuple_view = LiteralSlice(*nested_tuple); const auto tuple_view = LiteralSlice(nested_tuple_view, /*view_root=*/{0}); const auto matrix_view = LiteralSlice(tuple_view, /*view_root=*/{1}); - EXPECT_EQ(matrix_view, *Literal::CreateR2({{1.0, 2.0}, {3.0, 4.0}})); + EXPECT_EQ(matrix_view, + *LiteralUtil::CreateR2({{1.0, 2.0}, {3.0, 4.0}})); } TEST_F(LiteralUtilTest, BorrowingLiteralFromOneBufferPtr) { @@ -1488,7 +1496,7 @@ TEST_F(LiteralUtilTest, BorrowingLiteralFromMultipleBufferPtrs) { TEST_F(LiteralUtilTest, LiteralMove) { std::unique_ptr matrix = - Literal::CreateR2({{1.0, 2.0}, {3.0, 4.0}}); + LiteralUtil::CreateR2({{1.0, 2.0}, {3.0, 4.0}}); Literal literal(std::move(*matrix)); EXPECT_TRUE( @@ -1501,11 +1509,11 @@ TEST_F(LiteralUtilTest, LiteralMove) { TEST_F(LiteralUtilTest, DecomposeTuple) { Literal nil_literal(ShapeUtil::MakeNil()); - auto nested_tuple = Literal::MakeTuple( - {Literal::CreateR2({{1, 2}, {3, 4}}).get(), - Literal::MakeTuple({Literal::CreateR0(42).get(), - Literal::CreateR1({23.0, 44.0}).get(), - &nil_literal}) + auto nested_tuple = LiteralUtil::MakeTuple( + {LiteralUtil::CreateR2({{1, 2}, {3, 4}}).get(), + LiteralUtil::MakeTuple( + {LiteralUtil::CreateR0(42).get(), + LiteralUtil::CreateR1({23.0, 44.0}).get(), &nil_literal}) .get(), &nil_literal}); @@ -1542,13 +1550,13 @@ TEST_F(LiteralUtilTest, DecomposeEmptyTuple) { TEST_F(LiteralUtilTest, MoveIntoTuple) { std::vector elements; - elements.push_back(std::move(*Literal::CreateR0(1.0))); - elements.push_back(std::move(*Literal::CreateR1({4, 8}))); - elements.push_back(std::move( - *Literal::MakeTuple({Literal::CreateR0(42).get(), - Literal::CreateR1({23.0, 44.0}).get()}) + elements.push_back(std::move(*LiteralUtil::CreateR0(1.0))); + elements.push_back(std::move(*LiteralUtil::CreateR1({4, 8}))); + elements.push_back(std::move(*LiteralUtil::MakeTuple( + {LiteralUtil::CreateR0(42).get(), + LiteralUtil::CreateR1({23.0, 44.0}).get()}) - )); + )); Literal literal = Literal::MoveIntoTuple(&elements); ASSERT_TRUE(ShapeUtil::IsTuple(literal.shape())); @@ -1577,7 +1585,7 @@ TEST_F(LiteralUtilTest, LiteralMoveAssignment) { EXPECT_TRUE(ShapeUtil::Equal(ShapeUtil::MakeNil(), literal.shape())); std::unique_ptr matrix = - Literal::CreateR2({{1.0, 2.0}, {3.0, 4.0}}); + LiteralUtil::CreateR2({{1.0, 2.0}, {3.0, 4.0}}); literal = std::move(*matrix); EXPECT_TRUE( @@ -1590,7 +1598,7 @@ TEST_F(LiteralUtilTest, LiteralMoveAssignment) { TEST_F(LiteralUtilTest, LiteralSliceCopy) { std::unique_ptr matrix = - Literal::CreateR2({{1.0, 2.0}, {3.0, 4.0}}); + LiteralUtil::CreateR2({{1.0, 2.0}, {3.0, 4.0}}); const auto matrix_view = LiteralSlice(*matrix); LiteralSlice matrix_view_copy(matrix_view); @@ -1601,9 +1609,9 @@ TEST_F(LiteralUtilTest, LiteralSliceCopy) { } TEST_F(LiteralUtilTest, GetSetTuple) { - auto tuple = Literal::MakeTuple( - {Literal::CreateR0(42.0).get(), - Literal::CreateR2({{1.0, 2.0}, {3.0, 4.0}}).get()}); + auto tuple = LiteralUtil::MakeTuple( + {LiteralUtil::CreateR0(42.0).get(), + LiteralUtil::CreateR2({{1.0, 2.0}, {3.0, 4.0}}).get()}); EXPECT_EQ(tuple->Get(/*multi_index=*/{}, /*shape_index=*/{0}), 42.0); tuple->Set(/*multi_index=*/{}, /*shape_index=*/{0}, -5.0); EXPECT_EQ(tuple->Get(/*multi_index=*/{}, /*shape_index=*/{0}), -5.0); @@ -1644,20 +1652,20 @@ TEST_F(LiteralUtilTest, CreateFromShapeZeroInitialized) { TEST_F(LiteralUtilTest, ProtoRoundTrip) { // Test serializing then deserializing a Literal through a proto. - auto one_f32 = Literal::CreateR0(1.0); - auto two_f32 = Literal::CreateR0(2.0); - auto vector_int8 = Literal::CreateR1({-128, 0, 2, 4, 7, 56, 127}); - auto vector_c64 = Literal::CreateR1({{1.0, 2.0}, {3.0, 4.0}}); - auto vector_bfloat16 = Literal::CreateR1( + auto one_f32 = LiteralUtil::CreateR0(1.0); + auto two_f32 = LiteralUtil::CreateR0(2.0); + auto vector_int8 = LiteralUtil::CreateR1({-128, 0, 2, 4, 7, 56, 127}); + auto vector_c64 = LiteralUtil::CreateR1({{1.0, 2.0}, {3.0, 4.0}}); + auto vector_bfloat16 = LiteralUtil::CreateR1( {bfloat16{-1.0}, bfloat16{2.0}, bfloat16{-3.0}}); auto vector_half = - Literal::CreateR1({half{10.0}, half{20.0}, half{-30.0}}); + LiteralUtil::CreateR1({half{10.0}, half{20.0}, half{-30.0}}); auto matrix_pred = - Literal::CreateR2({{true, false, true}, {false, false, true}}); - auto tuple = Literal::MakeTuple( + LiteralUtil::CreateR2({{true, false, true}, {false, false, true}}); + auto tuple = LiteralUtil::MakeTuple( {one_f32.get(), vector_half.get(), matrix_pred.get(), matrix_pred.get()}); Literal nil_literal(ShapeUtil::MakeNil()); - auto nested_tuple = Literal::MakeTuple( + auto nested_tuple = LiteralUtil::MakeTuple( {tuple.get(), vector_bfloat16.get(), tuple.get(), &nil_literal}); auto to_from_proto = [](const Literal& literal) -> Literal { @@ -1790,8 +1798,8 @@ TEST_F(LiteralUtilTest, InvalidProtoTooManyTupleElements) { } TEST_F(LiteralUtilTest, SortSparseElements) { - auto literal = - Literal::CreateSparse({10, 10, 10}, SparseIndexArray(10, 3), {}); + auto literal = LiteralUtil::CreateSparse({10, 10, 10}, + SparseIndexArray(10, 3), {}); literal->AppendSparseElement({2, 3, 4}, 2.0); literal->AppendSparseElement({3, 4, 5}, 3.0); literal->AppendSparseElement({1, 2, 3}, 1.0); @@ -1805,21 +1813,22 @@ TEST_F(LiteralUtilTest, GetSparseElementAsString) { SparseIndexArray indices(10, {{1, 2, 3}, {2, 3, 4}, {3, 4, 5}}); ASSERT_EQ( - Literal::CreateSparse(dimensions, indices, {true, false, true}) + LiteralUtil::CreateSparse(dimensions, indices, {true, false, true}) ->GetSparseElementAsString(1), "false"); - ASSERT_EQ(Literal::CreateSparse(dimensions, indices, {1, 2, 3}) + ASSERT_EQ(LiteralUtil::CreateSparse(dimensions, indices, {1, 2, 3}) ->GetSparseElementAsString(1), tensorflow::strings::StrCat(int64{2})); - ASSERT_EQ(Literal::CreateSparse(dimensions, indices, {1.0, 2.0, 3.0}) - ->GetSparseElementAsString(1), - tensorflow::strings::StrCat(double{2.0})); - ASSERT_EQ(Literal::CreateSparse(dimensions, indices, - {half{1.0}, half{2.0}, half{3.0}}) + ASSERT_EQ( + LiteralUtil::CreateSparse(dimensions, indices, {1.0, 2.0, 3.0}) + ->GetSparseElementAsString(1), + tensorflow::strings::StrCat(double{2.0})); + ASSERT_EQ(LiteralUtil::CreateSparse(dimensions, indices, + {half{1.0}, half{2.0}, half{3.0}}) ->GetSparseElementAsString(1), tensorflow::strings::StrCat(static_cast(half{2.0}))); ASSERT_EQ( - Literal::CreateSparse( + LiteralUtil::CreateSparse( dimensions, indices, std::vector{{1.0, 2.0}, {3.0, 4.0}, {5.0, 6.0}}) ->GetSparseElementAsString(1), @@ -1827,33 +1836,36 @@ TEST_F(LiteralUtilTest, GetSparseElementAsString) { } TEST_F(LiteralUtilTest, BroadcastVectorToMatrix0) { - std::unique_ptr literal = Literal::CreateR1({1, 2}); + std::unique_ptr literal = LiteralUtil::CreateR1({1, 2}); TF_ASSERT_OK_AND_ASSIGN( std::unique_ptr broadcasted_literal, literal->Broadcast( /*result_shape=*/ShapeUtil::MakeShape(S64, {2, 2}), /*dimensions=*/{0})); - EXPECT_EQ(*broadcasted_literal, *Literal::CreateR2({{1, 1}, {2, 2}})); + EXPECT_EQ(*broadcasted_literal, + *LiteralUtil::CreateR2({{1, 1}, {2, 2}})); } TEST_F(LiteralUtilTest, BroadcastVectorToMatrix1) { - std::unique_ptr literal = Literal::CreateR1({1, 2}); + std::unique_ptr literal = LiteralUtil::CreateR1({1, 2}); TF_ASSERT_OK_AND_ASSIGN( std::unique_ptr broadcasted_literal, literal->Broadcast( /*result_shape=*/ShapeUtil::MakeShape(S64, {2, 2}), /*dimensions=*/{1})); - EXPECT_EQ(*broadcasted_literal, *Literal::CreateR2({{1, 2}, {1, 2}})); + EXPECT_EQ(*broadcasted_literal, + *LiteralUtil::CreateR2({{1, 2}, {1, 2}})); } TEST_F(LiteralUtilTest, BroadcastScalarToMatrix) { - std::unique_ptr literal = Literal::CreateR0(9); + std::unique_ptr literal = LiteralUtil::CreateR0(9); TF_ASSERT_OK_AND_ASSIGN( std::unique_ptr broadcasted_literal, literal->Broadcast( /*result_shape=*/ShapeUtil::MakeShape(S32, {2, 2}), /*dimensions=*/{})); - EXPECT_EQ(*broadcasted_literal, *Literal::CreateR2({{9, 9}, {9, 9}})); + EXPECT_EQ(*broadcasted_literal, + *LiteralUtil::CreateR2({{9, 9}, {9, 9}})); } } // namespace diff --git a/tensorflow/compiler/xla/literal_util.cc b/tensorflow/compiler/xla/literal_util.cc index 7c6a181b0a872bb03f1153017f16d1d06a99ecaa..548fbe8a83a3797aa8ac32dc1f6c085fc0100197 100644 --- a/tensorflow/compiler/xla/literal_util.cc +++ b/tensorflow/compiler/xla/literal_util.cc @@ -43,25 +43,6 @@ namespace xla { namespace { -constexpr bool kLittleEndian = __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__; - -// Converts between little and big endian. -// -// Precondition: size % 2 == 0 (elements in the array are 16 bits long) -void ConvertEndianShort(string* bytes) { - CHECK_EQ(bytes->size() / 2, 0); - for (int64 i = 0; i < bytes->size(); i += 2) { - std::swap((*bytes)[i], (*bytes)[i + 1]); - } -} - -void ConvertEndianShort(char* bytes, int64 size) { - CHECK_EQ(size / 2, 0); - for (int64 i = 0; i < size; i += 2) { - std::swap(bytes[i], bytes[i + 1]); - } -} - // Return a literal with all arrays of type FromNativeT converted to type // ToNativeT in the given literal. template @@ -103,505 +84,54 @@ std::unique_ptr ConvertType(LiteralSlice literal) { } // namespace -LiteralBase::~LiteralBase() {} - -std::ostream& operator<<(std::ostream& out, const Literal& literal) { - out << literal.ToString(); - return out; -} - -Literal::StrideConfig::StrideConfig( - const Shape& source_shape, const Shape& dest_shape, - tensorflow::gtl::ArraySlice dimensions) - : dimensions(dimensions), - base(dimensions.size(), 0), - step(dimensions.size(), 1) { - if (!dimensions.empty()) { - // Selects the shape with the largest minor dimension as the one upon - // which to run the tight stride loop. - if (dimensions[LayoutUtil::Minor(source_shape.layout(), 0)] >= - dimensions[LayoutUtil::Minor(dest_shape.layout(), 0)]) { - minor_dimension = LayoutUtil::Minor(source_shape.layout(), 0); - dest_stride = IndexUtil::GetDimensionStride(dest_shape, minor_dimension); - } else { - minor_dimension = LayoutUtil::Minor(dest_shape.layout(), 0); - source_stride = - IndexUtil::GetDimensionStride(source_shape, minor_dimension); - } - minor_loop_size = dimensions[minor_dimension]; - step[minor_dimension] = minor_loop_size; - } -} - -Literal::Literal(const Shape& shape) - : Literal(shape, /*allocate_arrays=*/true) {} - -void Literal::SetPiece(const Shape& shape, Piece* piece, bool allocate_arrays) { - if (ShapeUtil::IsTuple(shape)) { - for (int i = 0; i < ShapeUtil::TupleElementCount(shape); ++i) { - const Shape& subshape = shape.tuple_shapes(i); - - auto child_piece = Piece(); - child_piece.set_subshape(&subshape); - - SetPiece(subshape, &child_piece, allocate_arrays); - - piece->emplace_back(std::move(child_piece)); - } - } else if (ShapeUtil::IsArray(shape)) { - if (allocate_arrays) { - if (LayoutUtil::IsSparseArray(shape)) { - // For sparse arrays, the buffer must be of the size of the maximum - // number of sparse elements possible. - const int64 max_sparse_elements = - LayoutUtil::MaxSparseElements(shape.layout()); - piece->set_buffer( - new char[max_sparse_elements * - ShapeUtil::ByteSizeOfPrimitiveType(shape.element_type())]); - piece->set_sparse_indices( - new SparseIndexArray(max_sparse_elements, ShapeUtil::Rank(shape))); - } else { - piece->set_buffer(new char[piece->size_bytes()]); - } - } - } else { - // If the shape is neither an array nor tuple, then it must be - // zero-sized. Otherwise, some memory needs to be allocated for it. - CHECK_EQ(piece->size_bytes(), 0); - } -} - -Literal::Literal(const Shape& shape, bool allocate_arrays) - : LiteralBase(), shape_(MakeUnique(shape)) { - CHECK(LayoutUtil::HasLayout(*shape_)); - root_piece_ = new Piece(); - root_piece_->set_subshape(shape_.get()); - CHECK(&root_piece_->subshape() == shape_.get()); - - SetPiece(*shape_, root_piece_, allocate_arrays); -} - -Literal::~Literal() { - if (root_piece_ != nullptr) { - DeallocateBuffers(); - delete root_piece_; - } -} - -void Literal::DeallocateBuffers() { - root_piece_->ForEachMutableSubpiece( - [&](const ShapeIndex& index, Piece* piece) { - if (piece->buffer() != nullptr) { - delete[] piece->buffer(); - delete piece->sparse_indices(); - } - }); -} - -Literal::Literal(Literal&& other) : LiteralBase() { *this = std::move(other); } - -Literal& Literal::operator=(Literal&& other) { - DCHECK(&other.root_piece_->subshape() == other.shape_.get()); - using std::swap; - swap(shape_, other.shape_); - swap(root_piece_, other.root_piece_); - DCHECK(&root_piece_->subshape() == shape_.get()); - - return *this; -} - -std::unique_ptr LiteralBase::CreateFromShape(const Shape& shape) { - auto literal = MakeUnique(shape); - literal->root_piece_->ForEachMutableSubpiece( - [&](const ShapeIndex& index, Piece* piece) { - if (ShapeUtil::IsArray(piece->subshape())) { - memset(piece->untyped_data(), 0, piece->size_bytes()); - } - }); - return literal; -} - -const SparseIndexArray* LiteralBase::sparse_indices( - const ShapeIndex& shape_index) const { - return piece(shape_index).sparse_indices(); -} - -SparseIndexArray* Literal::sparse_indices(const ShapeIndex& shape_index) { - return piece(shape_index).sparse_indices(); -} - -/* static */ std::unique_ptr Literal::CreateFromDimensions( +/* static */ std::unique_ptr LiteralUtil::CreateFromDimensions( PrimitiveType primitive_type, tensorflow::gtl::ArraySlice dimensions) { - return CreateFromShape(ShapeUtil::MakeShape(primitive_type, dimensions)); + return Literal::CreateFromShape( + ShapeUtil::MakeShape(primitive_type, dimensions)); } -/* static */ std::unique_ptr Literal::ConvertBF16ToF32( +/* static */ std::unique_ptr LiteralUtil::ConvertBF16ToF32( const LiteralSlice& bf16_literal) { return ConvertType(bf16_literal); } -/* static */ std::unique_ptr Literal::ConvertF32ToBF16( +/* static */ std::unique_ptr LiteralUtil::ConvertF32ToBF16( const LiteralSlice& f32_literal) { return ConvertType(f32_literal); } -template -Status Literal::CopySliceFromInternal( - const LiteralBase& src_literal, tensorflow::gtl::ArraySlice src_base, - tensorflow::gtl::ArraySlice dest_base, - tensorflow::gtl::ArraySlice copy_size) { - TF_RET_CHECK(ShapeUtil::Rank(src_literal.shape()) == src_base.size()); - TF_RET_CHECK(ShapeUtil::Rank(shape()) == dest_base.size()); - - auto linear_index = [](const Shape& shape, - tensorflow::gtl::ArraySlice multi_index) { - return IndexUtil::MultidimensionalIndexToLinearIndex(shape, multi_index); - }; - - if (ShapeUtil::Rank(src_literal.shape()) == 0 || - ShapeUtil::Rank(shape()) == 0) { - // If any of the two shapes are scalars, we can just call the StridedCopy() - // directly, and we know we will be copying only one value. - TF_RET_CHECK(copy_size.empty()); - StridedCopy(data(), linear_index(shape(), dest_base), 0, - src_literal.data(), - linear_index(src_literal.shape(), src_base), 0, 1); - } else if (!ShapeUtil::IsZeroElementArray(shape()) && - !ShapeUtil::IsZeroElementArray(src_literal.shape())) { - // Perform copy if neither src nor dest has dimensions with zero element, - // otherwise it's a no-op. - TF_RET_CHECK(src_base.size() == dest_base.size()); - TF_RET_CHECK(src_base.size() == copy_size.size()); - - // Scan the source from minor, stepping in copy size blocks, then within - // the index enumaration functor, do a strided copy advancing source index - // by one (walking through the minor dimension), and destination index by - // proper stride size at the matching dimension. - DimensionVector src_indexes(src_base.size(), 0); - DimensionVector dest_indexes(dest_base.size(), 0); - Literal::StrideConfig stride_config(src_literal.shape(), shape(), - copy_size); - - auto copy_proc = [&](tensorflow::gtl::ArraySlice indexes) { - // Map from multi-dimensional index, to source index. - std::transform(indexes.begin(), indexes.end(), src_base.begin(), - src_indexes.begin(), std::plus()); - // Map from multi-dimensional index, to destination index. - std::transform(indexes.begin(), indexes.end(), dest_base.begin(), - dest_indexes.begin(), std::plus()); - - int64 src_index = linear_index(src_literal.shape(), src_indexes); - int64 dest_index = linear_index(shape(), dest_indexes); - - // `this->` is needed to workaround MSVC bug: #16882 - StridedCopy(this->data(), dest_index, stride_config.dest_stride, - src_literal.data(), src_index, - stride_config.source_stride, stride_config.minor_loop_size); - return true; - }; - - ShapeUtil::ForEachIndex(src_literal.shape(), stride_config.base, - stride_config.dimensions, stride_config.step, - copy_proc); - } - return Status::OK(); -} - -Status Literal::CopyElementFrom(const LiteralSlice& src_literal, - tensorflow::gtl::ArraySlice src_index, - tensorflow::gtl::ArraySlice dest_index) { - DCHECK_EQ(shape().element_type(), src_literal.shape().element_type()); - const int64 src_linear_index = IndexUtil::MultidimensionalIndexToLinearIndex( - src_literal.shape(), src_index); - const int64 dest_linear_index = - IndexUtil::MultidimensionalIndexToLinearIndex(shape(), dest_index); - const int64 primitive_size = - ShapeUtil::ByteSizeOfPrimitiveType(shape().element_type()); - - char* dest_address = - static_cast(untyped_data()) + dest_linear_index * primitive_size; - const char* source_address = - static_cast(src_literal.untyped_data()) + - src_linear_index * primitive_size; - if (dest_address != source_address) { - memcpy(dest_address, source_address, primitive_size); - } - return Status::OK(); -} - -/* static */ std::unique_ptr Literal::CreateToken() { +/* static */ std::unique_ptr LiteralUtil::CreateToken() { return MakeUnique(ShapeUtil::MakeTokenShape()); } -std::vector Literal::DecomposeTuple() { - CHECK(ShapeUtil::IsTuple(shape())); - std::vector elements; - for (int i = 0; i < ShapeUtil::TupleElementCount(shape()); ++i) { - elements.push_back(Literal(ShapeUtil::GetSubshape(shape(), {i}), - /*allocate_arrays=*/false)); - Literal& element = elements.back(); - element.root_piece_->ForEachMutableSubpiece( - [&](const ShapeIndex& index, Piece* dest_piece) { - ShapeIndex src_index = {i}; - for (int64 j : index) { - src_index.push_back(j); - } - Piece& src_piece = piece(src_index); - - // Move the respective buffer and sparse indices over to the element - // Literal. - dest_piece->set_buffer(src_piece.buffer()); - src_piece.set_buffer(nullptr); - dest_piece->set_sparse_indices(src_piece.sparse_indices()); - src_piece.set_sparse_indices(nullptr); - }); - } - // Set this literal to be nil-shaped. - *this = Literal(); - return elements; -} - -/* static */ Literal Literal::MoveIntoTuple( - tensorflow::gtl::MutableArraySlice elements) { - std::vector element_shapes; - for (const Literal& element : elements) { - element_shapes.push_back(element.shape()); - } - Literal literal(ShapeUtil::MakeTupleShape(element_shapes), - /*allocate_arrays=*/false); - for (int i = 0; i < elements.size(); ++i) { - TF_CHECK_OK( - literal.MoveFrom(std::move(elements[i]), /*dest_shape_index=*/{i})); - } - return literal; -} - -namespace { - -// Copies the elements in 'src' to 'dest'. The shape and layout of the data in -// the array slices are indicated by dest_shape and src_shape respectively. -template -void CopyElementsBetween(tensorflow::gtl::MutableArraySlice dest, - tensorflow::gtl::ArraySlice src, - const Shape& dest_shape, const Shape& src_shape) { - CHECK(ShapeUtil::Compatible(dest_shape, src_shape)); - if (ShapeUtil::IsZeroElementArray(dest_shape)) { - return; - } - std::vector index(ShapeUtil::Rank(dest_shape)); - do { - dest[IndexUtil::MultidimensionalIndexToLinearIndex(dest_shape, index)] = - src[IndexUtil::MultidimensionalIndexToLinearIndex(src_shape, index)]; - } while (IndexUtil::BumpIndices(dest_shape, &index)); -} - -} // namespace - -Status LiteralBase::Piece::CopyFrom(const LiteralBase::Piece& src) { - CHECK(subshape_ != nullptr); - CHECK(src.subshape_ != nullptr); - if (ShapeUtil::Equal(subshape(), src.subshape())) { - // If the layouts are equal it's faster just to memcpy. - memcpy(buffer(), src.buffer(), src.size_bytes()); - } else { - TF_RET_CHECK(ShapeUtil::Compatible(src.subshape(), subshape())); - std::vector origin(ShapeUtil::Rank(subshape()), 0); - switch (subshape().element_type()) { -#define COPY_ELEMENTS(XLA_T, NATIVE_T) \ - case (XLA_T): \ - CopyElementsBetween(data(), src.data(), \ - subshape(), src.subshape()); \ - break; - COPY_ELEMENTS(U8, uint8); - COPY_ELEMENTS(U16, uint16); - COPY_ELEMENTS(U32, uint32); - COPY_ELEMENTS(U64, uint64); - COPY_ELEMENTS(S8, int8); - COPY_ELEMENTS(S16, int16); - COPY_ELEMENTS(S32, int32); - COPY_ELEMENTS(S64, int64); - COPY_ELEMENTS(F16, half); - COPY_ELEMENTS(BF16, bfloat16); - COPY_ELEMENTS(F32, float); - COPY_ELEMENTS(F64, double); - COPY_ELEMENTS(C64, complex64); - COPY_ELEMENTS(PRED, bool); -#undef COPY_ELEMENTS - default: - return Unimplemented( - "Copying a Literal object with element type %s is not implemented.", - PrimitiveType_Name(subshape().element_type()).c_str()); - } - } - return Status::OK(); -} - -Status Literal::CopyFrom(const LiteralSlice& src_literal, - const ShapeIndex& dest_shape_index, - const ShapeIndex& src_shape_index) { - const Shape& dest_subshape = - ShapeUtil::GetSubshape(shape(), dest_shape_index); - const Shape& src_subshape = - ShapeUtil::GetSubshape(src_literal.shape(), src_shape_index); - if (!ShapeUtil::Compatible(dest_subshape, src_subshape)) { - return InvalidArgument( - "Destination subshape incompatible with source subshape: %s vs %s", - ShapeUtil::HumanString(dest_subshape).c_str(), - ShapeUtil::HumanString(src_subshape).c_str()); - } - return root_piece_->ForEachMutableSubpieceWithStatus( - [&](const ShapeIndex& index, Piece* piece) { - if (!ShapeUtil::IsArray(piece->subshape())) { - return Status::OK(); - } - - // Determine if this index is in the part of this literal that we want - // to copy over from src_literal. - bool in_subtree_to_copy = true; - for (int i = 0; i < dest_shape_index.size(); ++i) { - if (index[i] != dest_shape_index[i]) { - in_subtree_to_copy = false; - break; - } - } - if (!in_subtree_to_copy) { - return Status::OK(); - } - // Construct the index of the corresponding piece in the source literal. - ShapeIndex src_piece_index = src_shape_index; - for (int64 i = dest_shape_index.size(); i < index.size(); ++i) { - src_piece_index.push_back(index[i]); - } - TF_RETURN_IF_ERROR(piece->CopyFrom(src_literal.piece(src_piece_index))); - return Status::OK(); - }); -} - -Status Literal::MoveFrom(Literal&& src_literal, - const ShapeIndex& dest_shape_index) { - const Shape& dest_subshape = - ShapeUtil::GetSubshape(shape(), dest_shape_index); - if (!ShapeUtil::Equal(dest_subshape, src_literal.shape())) { - return InvalidArgument( - "Destination subshape not equal to source shape: %s vs %s", - ShapeUtil::HumanString(dest_subshape).c_str(), - ShapeUtil::HumanString(src_literal.shape()).c_str()); - } - - src_literal.root_piece_->ForEachSubpiece( - [&](const ShapeIndex& src_index, const Piece& src_piece) { - if (!ShapeUtil::IsArray(src_piece.subshape())) { - return; - } - - ShapeIndex dest_index = dest_shape_index; - for (int64 i : src_index) { - dest_index.push_back(i); - } - Piece& dest_piece = piece(dest_index); - delete[] dest_piece.buffer(); - dest_piece.set_buffer(src_piece.buffer()); - delete dest_piece.sparse_indices(); - dest_piece.set_sparse_indices(src_piece.sparse_indices()); - }); - - src_literal.shape_ = MakeUnique(ShapeUtil::MakeNil()); - delete src_literal.root_piece_; - src_literal.root_piece_ = new LiteralBase::Piece(); - src_literal.root_piece_->set_subshape(src_literal.shape_.get()); - - return Status::OK(); -} - -Status Literal::CopySliceFrom(const LiteralSlice& src_literal, - tensorflow::gtl::ArraySlice src_base, - tensorflow::gtl::ArraySlice dest_base, - tensorflow::gtl::ArraySlice copy_size) { - TF_RET_CHECK(ShapeUtil::IsArray(shape())) << ShapeUtil::HumanString(shape()); - TF_RET_CHECK(ShapeUtil::IsArray(src_literal.shape())) - << ShapeUtil::HumanString(src_literal.shape()); - TF_RET_CHECK(ShapeUtil::SameElementType(src_literal.shape(), shape())); - - switch (shape().element_type()) { - case U8: - return CopySliceFromInternal(src_literal, src_base, dest_base, - copy_size); - case U16: - return CopySliceFromInternal(src_literal, src_base, dest_base, - copy_size); - case U32: - return CopySliceFromInternal(src_literal, src_base, dest_base, - copy_size); - case U64: - return CopySliceFromInternal(src_literal, src_base, dest_base, - copy_size); - case S8: - return CopySliceFromInternal(src_literal, src_base, dest_base, - copy_size); - case S16: - return CopySliceFromInternal(src_literal, src_base, dest_base, - copy_size); - case S32: - return CopySliceFromInternal(src_literal, src_base, dest_base, - copy_size); - case S64: - return CopySliceFromInternal(src_literal, src_base, dest_base, - copy_size); - case F16: - return CopySliceFromInternal(src_literal, src_base, dest_base, - copy_size); - case BF16: - return CopySliceFromInternal(src_literal, src_base, dest_base, - copy_size); - case F32: - return CopySliceFromInternal(src_literal, src_base, dest_base, - copy_size); - case F64: - return CopySliceFromInternal(src_literal, src_base, dest_base, - copy_size); - case C64: - return CopySliceFromInternal(src_literal, src_base, dest_base, - copy_size); - case PRED: - return CopySliceFromInternal(src_literal, src_base, dest_base, - copy_size); - default: - break; - } - return Unimplemented( - "Copying a slice from a Literal object with element type %d is not " - "implemented.", - shape().element_type()); -} - -/* static */ Literal Literal::Zero(PrimitiveType primitive_type) { +/* static */ Literal LiteralUtil::Zero(PrimitiveType primitive_type) { switch (primitive_type) { case U8: - return std::move(*Literal::CreateR0(0)); + return std::move(*LiteralUtil::CreateR0(0)); case U32: - return std::move(*Literal::CreateR0(0)); + return std::move(*LiteralUtil::CreateR0(0)); case U64: - return std::move(*Literal::CreateR0(0)); + return std::move(*LiteralUtil::CreateR0(0)); case S8: - return std::move(*Literal::CreateR0(0)); + return std::move(*LiteralUtil::CreateR0(0)); case S32: - return std::move(*Literal::CreateR0(0)); + return std::move(*LiteralUtil::CreateR0(0)); case S64: - return std::move(*Literal::CreateR0(0)); + return std::move(*LiteralUtil::CreateR0(0)); case F16: - return std::move(*Literal::CreateR0(static_cast(0.0f))); + return std::move(*LiteralUtil::CreateR0(static_cast(0.0f))); case BF16: return std::move( - *Literal::CreateR0(static_cast(0.0f))); + *LiteralUtil::CreateR0(static_cast(0.0f))); case F32: - return std::move(*Literal::CreateR0(0)); + return std::move(*LiteralUtil::CreateR0(0)); case F64: - return std::move(*Literal::CreateR0(0)); + return std::move(*LiteralUtil::CreateR0(0)); case C64: - return std::move(*Literal::CreateR0(0)); + return std::move(*LiteralUtil::CreateR0(0)); case PRED: - return std::move(*Literal::CreateR0(false)); + return std::move(*LiteralUtil::CreateR0(false)); case S16: case U16: LOG(FATAL) << "u16/s16 literals not yet implemented"; @@ -614,33 +144,33 @@ Status Literal::CopySliceFrom(const LiteralSlice& src_literal, } } -/* static */ Literal Literal::One(PrimitiveType primitive_type) { +/* static */ Literal LiteralUtil::One(PrimitiveType primitive_type) { switch (primitive_type) { case U8: - return std::move(*Literal::CreateR0(1)); + return std::move(*LiteralUtil::CreateR0(1)); case U32: - return std::move(*Literal::CreateR0(1)); + return std::move(*LiteralUtil::CreateR0(1)); case U64: - return std::move(*Literal::CreateR0(1)); + return std::move(*LiteralUtil::CreateR0(1)); case S8: - return std::move(*Literal::CreateR0(1)); + return std::move(*LiteralUtil::CreateR0(1)); case S32: - return std::move(*Literal::CreateR0(1)); + return std::move(*LiteralUtil::CreateR0(1)); case S64: - return std::move(*Literal::CreateR0(1)); + return std::move(*LiteralUtil::CreateR0(1)); case F16: - return std::move(*Literal::CreateR0(static_cast(1.0f))); + return std::move(*LiteralUtil::CreateR0(static_cast(1.0f))); case BF16: return std::move( - *Literal::CreateR0(static_cast(1.0f))); + *LiteralUtil::CreateR0(static_cast(1.0f))); case F32: - return std::move(*Literal::CreateR0(1)); + return std::move(*LiteralUtil::CreateR0(1)); case F64: - return std::move(*Literal::CreateR0(1)); + return std::move(*LiteralUtil::CreateR0(1)); case C64: - return std::move(*Literal::CreateR0(1)); + return std::move(*LiteralUtil::CreateR0(1)); case PRED: - return std::move(*Literal::CreateR0(true)); + return std::move(*LiteralUtil::CreateR0(true)); case S16: case U16: LOG(FATAL) << "u16/s16 literals not yet implemented"; @@ -653,44 +183,44 @@ Status Literal::CopySliceFrom(const LiteralSlice& src_literal, } } -/* static */ Literal Literal::MinValue(PrimitiveType primitive_type) { +/* static */ Literal LiteralUtil::MinValue(PrimitiveType primitive_type) { switch (primitive_type) { case U8: return std::move( - *Literal::CreateR0(std::numeric_limits::min())); + *LiteralUtil::CreateR0(std::numeric_limits::min())); case U32: return std::move( - *Literal::CreateR0(std::numeric_limits::min())); + *LiteralUtil::CreateR0(std::numeric_limits::min())); case U64: return std::move( - *Literal::CreateR0(std::numeric_limits::min())); + *LiteralUtil::CreateR0(std::numeric_limits::min())); case S8: return std::move( - *Literal::CreateR0(std::numeric_limits::min())); + *LiteralUtil::CreateR0(std::numeric_limits::min())); case S32: return std::move( - *Literal::CreateR0(std::numeric_limits::min())); + *LiteralUtil::CreateR0(std::numeric_limits::min())); case S64: return std::move( - *Literal::CreateR0(std::numeric_limits::min())); + *LiteralUtil::CreateR0(std::numeric_limits::min())); case F32: - return std::move( - *Literal::CreateR0(-std::numeric_limits::infinity())); + return std::move(*LiteralUtil::CreateR0( + -std::numeric_limits::infinity())); case F64: - return std::move( - *Literal::CreateR0(-std::numeric_limits::infinity())); + return std::move(*LiteralUtil::CreateR0( + -std::numeric_limits::infinity())); case C64: LOG(FATAL) << "C64 element type has no minimum value"; case PRED: - return std::move(*Literal::CreateR0(false)); + return std::move(*LiteralUtil::CreateR0(false)); case S16: case U16: LOG(FATAL) << "u16/s16 literals not yet implemented"; case F16: - return std::move(*Literal::CreateR0( + return std::move(*LiteralUtil::CreateR0( static_cast(-std::numeric_limits::infinity()))); case BF16: - return std::move(*Literal::CreateR0( + return std::move(*LiteralUtil::CreateR0( static_cast(-std::numeric_limits::infinity()))); case TUPLE: LOG(FATAL) << "tuple element type has no minimum value"; @@ -701,42 +231,42 @@ Status Literal::CopySliceFrom(const LiteralSlice& src_literal, } } -/* static */ Literal Literal::MaxValue(PrimitiveType primitive_type) { +/* static */ Literal LiteralUtil::MaxValue(PrimitiveType primitive_type) { switch (primitive_type) { case U8: return std::move( - *Literal::CreateR0(std::numeric_limits::max())); + *LiteralUtil::CreateR0(std::numeric_limits::max())); case U32: return std::move( - *Literal::CreateR0(std::numeric_limits::max())); + *LiteralUtil::CreateR0(std::numeric_limits::max())); case U64: return std::move( - *Literal::CreateR0(std::numeric_limits::max())); + *LiteralUtil::CreateR0(std::numeric_limits::max())); case S8: return std::move( - *Literal::CreateR0(std::numeric_limits::max())); + *LiteralUtil::CreateR0(std::numeric_limits::max())); case S32: return std::move( - *Literal::CreateR0(std::numeric_limits::max())); + *LiteralUtil::CreateR0(std::numeric_limits::max())); case S64: return std::move( - *Literal::CreateR0(std::numeric_limits::max())); + *LiteralUtil::CreateR0(std::numeric_limits::max())); case F32: - return std::move( - *Literal::CreateR0(std::numeric_limits::infinity())); + return std::move(*LiteralUtil::CreateR0( + std::numeric_limits::infinity())); case F64: - return std::move( - *Literal::CreateR0(std::numeric_limits::infinity())); + return std::move(*LiteralUtil::CreateR0( + std::numeric_limits::infinity())); case PRED: - return std::move(*Literal::CreateR0(true)); + return std::move(*LiteralUtil::CreateR0(true)); case S16: case U16: LOG(FATAL) << "u16/s16 literals not yet implemented"; case F16: - return std::move(*Literal::CreateR0( + return std::move(*LiteralUtil::CreateR0( static_cast(std::numeric_limits::infinity()))); case BF16: - return std::move(*Literal::CreateR0( + return std::move(*LiteralUtil::CreateR0( static_cast(std::numeric_limits::infinity()))); case TUPLE: LOG(FATAL) << "tuple element type has no maximum value"; @@ -747,7 +277,7 @@ Status Literal::CopySliceFrom(const LiteralSlice& src_literal, } } -/* static */ std::unique_ptr Literal::CreateR1( +/* static */ std::unique_ptr LiteralUtil::CreateR1( const tensorflow::core::Bitmap& values) { auto literal = MakeUnique( ShapeUtil::MakeShape(PRED, {static_cast(values.bits())})); @@ -755,17 +285,7 @@ Status Literal::CopySliceFrom(const LiteralSlice& src_literal, return literal; } -void Literal::PopulateR1(const tensorflow::core::Bitmap& values) { - CHECK(ShapeUtil::IsArray(shape())); - CHECK_EQ(ShapeUtil::Rank(shape()), 1); - CHECK_EQ(element_count(), values.bits()); - CHECK_EQ(shape().element_type(), PRED); - for (int64 i = 0; i < static_cast(values.bits()); ++i) { - Set({i}, values.get(i)); - } -} - -/* static */ std::unique_ptr Literal::CreateR1U8( +/* static */ std::unique_ptr LiteralUtil::CreateR1U8( tensorflow::StringPiece value) { auto literal = MakeUnique( ShapeUtil::MakeShape(U8, {static_cast(value.size())})); @@ -775,116 +295,13 @@ void Literal::PopulateR1(const tensorflow::core::Bitmap& values) { return literal; } -/* static */ std::unique_ptr Literal::CreateR2F32Linspace(float from, - float to, - int64 rows, - int64 cols) { +/* static */ std::unique_ptr LiteralUtil::CreateR2F32Linspace( + float from, float to, int64 rows, int64 cols) { auto value = MakeLinspaceArray2D(from, to, rows, cols); return CreateR2FromArray2D(*value); } -std::unique_ptr LiteralBase::Relayout( - const Layout& new_layout, const ShapeIndex& shape_index) const { - // Create new shape with 'new_layout' set at the given shape index. - Shape new_shape = shape(); - Shape* subshape = ShapeUtil::GetMutableSubshape(&new_shape, shape_index); - TF_CHECK_OK(LayoutUtil::ValidateLayoutForShape(new_layout, *subshape)); - *subshape->mutable_layout() = new_layout; - auto result = MakeUnique(new_shape); - TF_CHECK_OK(result->CopyFrom(*this)); - return result; -} - -std::unique_ptr LiteralBase::Relayout( - const Shape& shape_with_layout) const { - CHECK(ShapeUtil::Compatible(shape_with_layout, shape())) - << "Given shape_with_layout " << ShapeUtil::HumanString(shape_with_layout) - << " not compatible with literal shape " - << ShapeUtil::HumanString(shape()); - std::unique_ptr result = CreateFromShape(shape_with_layout); - ShapeUtil::ForEachSubshape( - result->shape(), - [this, &result](const Shape& subshape, const ShapeIndex& index) { - if (ShapeUtil::IsArray(subshape)) { - TF_CHECK_OK(result->CopyFrom(*this, - /*dest_shape_index=*/index, - /*src_shape_index=*/index)); - } - }); - return result; -} - -StatusOr> LiteralBase::Broadcast( - const Shape& result_shape, - tensorflow::gtl::ArraySlice dimensions) const { - if (!ShapeUtil::IsArray(shape())) { - return InvalidArgument("Broadcast only supports arrays."); - } - - for (int64 i = 0; i < dimensions.size(); i++) { - TF_RET_CHECK(shape().dimensions(i) == - result_shape.dimensions(dimensions[i])); - } - - std::unique_ptr result = MakeUnique(result_shape); - - // scratch_source_index is temporary storage space for the computed index into - // the input literal. We put it here to avoid allocating an std::vector in - // every iteration of ShapeUtil::ForEachIndex. - std::vector scratch_source_index(shape().dimensions_size()); - - char* dest_data = static_cast(result->untyped_data()); - const char* source_data = static_cast(untyped_data()); - const int64 primitive_size = - ShapeUtil::ByteSizeOfPrimitiveType(shape().element_type()); - - ShapeUtil::ForEachIndex( - result_shape, [&](tensorflow::gtl::ArraySlice output_index) { - for (int64 i = 0; i < dimensions.size(); ++i) { - scratch_source_index[i] = output_index[dimensions[i]]; - } - int64 dest_index = IndexUtil::MultidimensionalIndexToLinearIndex( - result_shape, output_index); - int64 source_index = IndexUtil::MultidimensionalIndexToLinearIndex( - shape(), scratch_source_index); - memcpy(dest_data + primitive_size * dest_index, - source_data + primitive_size * source_index, primitive_size); - return true; - }); - - return std::move(result); -} - -StatusOr> LiteralBase::Reshape( - tensorflow::gtl::ArraySlice dimensions) const { - if (!ShapeUtil::IsArray(shape())) { - return InvalidArgument("Reshape does not support tuples."); - } - std::unique_ptr output; - if (!LayoutUtil::IsMonotonicWithDim0Major(shape().layout())) { - output = - Relayout(LayoutUtil::GetDefaultLayoutForRank(ShapeUtil::Rank(shape()))); - } else { - output = CloneToUnique(); - } - // Because the layout is monotonic, we can simply reuse the same sequence of - // values without changing their order. - *output->mutable_shape_do_not_use() = - ShapeUtil::MakeShape(shape().element_type(), dimensions); - - int64 elements_before = ShapeUtil::ElementsIn(shape()); - int64 elements_after = ShapeUtil::ElementsIn(output->shape()); - if (elements_before != elements_after) { - return InvalidArgument( - "Shapes before and after Literal::Reshape have different numbers " - "of elements: %s vs %s.", - ShapeUtil::HumanString(shape()).c_str(), - ShapeUtil::HumanString(output->shape()).c_str()); - } - return std::move(output); -} - -/* static */ std::unique_ptr Literal::ReshapeSlice( +/* static */ std::unique_ptr LiteralUtil::ReshapeSlice( tensorflow::gtl::ArraySlice new_dimensions, tensorflow::gtl::ArraySlice minor_to_major, const LiteralSlice& literal) { @@ -956,588 +373,77 @@ StatusOr> LiteralBase::Reshape( return new_literal; } -std::unique_ptr LiteralBase::Transpose( - tensorflow::gtl::ArraySlice permutation) const { - CHECK(ShapeUtil::IsArray(shape())) << "Tuple is not supported for transpose"; - CHECK(IsPermutation(permutation, ShapeUtil::Rank(shape()))) - << "Given permutation is not a permutation of dimension numbers"; - // To transpose the array, we just permute the dimensions and layout, and - // do a straight memory copy of the raw data set. - // This is considerably faster than iterating over every array element using - // the EachCell<>() and Set<>() APIs. - std::vector inverse_permutation = InversePermutation(permutation); - Shape permuted_shape = - ShapeUtil::PermuteDimensions(inverse_permutation, shape()); - // Replace the layout with one affine to this shape, such that a - // transpose operation can be performed by leaving the flat values - // representation intact. - // For example, consider the shape F32[11,8]{1,0} under a {1,0} permutation. - // The shape with affine layout resulting from that operation will be - // F32[8,11]{0,1}, since it leaves the original most minor (the 8 sized), the - // most minor. - // - // Essentially, given MinMaj(Di) the position of the Di dimension within the - // minor to major vector, and given T(Di) the index that the original Di - // dimension has within the transposed array, a layout is affine if - // MinMaj(Di) == TMinMaj(T(Di)), with TMinMaj() being the minor to major - // vector of the affine layout. - CHECK(LayoutUtil::IsDenseArray(permuted_shape)); - Layout* layout = permuted_shape.mutable_layout(); - layout->clear_minor_to_major(); - for (auto index : LayoutUtil::MinorToMajor(shape())) { - layout->add_minor_to_major(inverse_permutation[index]); - } - auto new_literal = MakeUnique(permuted_shape); - DCHECK_EQ(ShapeUtil::ByteSizeOf(new_literal->shape()), - ShapeUtil::ByteSizeOf(shape())); - std::memcpy(new_literal->untyped_data(), untyped_data(), size_bytes()); - return new_literal; -} - -template -std::unique_ptr LiteralBase::SliceInternal( - const Shape& result_shape, - tensorflow::gtl::ArraySlice start_indices) const { - auto result_literal = MakeUnique(result_shape); - DimensionVector new_indices(ShapeUtil::Rank(result_shape)); - result_literal->EachCell( - [&](tensorflow::gtl::ArraySlice indices, NativeT /*value*/) { - for (int64 i = 0; i < ShapeUtil::Rank(result_shape); ++i) { - new_indices[i] = indices[i] + start_indices[i]; - } - NativeT value = Get(new_indices); - result_literal->Set(indices, value); - }); - return result_literal; -} - -std::unique_ptr LiteralBase::Slice( - tensorflow::gtl::ArraySlice start_indices, - tensorflow::gtl::ArraySlice limit_indices) const { - CHECK(ShapeUtil::IsArray(shape())) << "tuple is not supported for slice"; - - DimensionVector result_dimensions; - for (int64 dnum = 0; dnum < ShapeUtil::Rank(shape()); ++dnum) { - CHECK_GE(start_indices[dnum], 0); - CHECK_LE(limit_indices[dnum], shape().dimensions(dnum)) - << "dnum = " << dnum; - int64 dimension = limit_indices[dnum] - start_indices[dnum]; - CHECK_GE(dimension, 0) << "dnum = " << dnum; - result_dimensions.push_back(dimension); - } - const auto result_shape = - ShapeUtil::MakeShapeWithLayout(shape().element_type(), result_dimensions, - LayoutUtil::MinorToMajor(shape())); - switch (result_shape.element_type()) { - case F32: - return SliceInternal(result_shape, start_indices); - case BF16: - return SliceInternal(result_shape, start_indices); - case C64: - return SliceInternal(result_shape, start_indices); - case S32: - return SliceInternal(result_shape, start_indices); - case U32: - return SliceInternal(result_shape, start_indices); - default: - LOG(FATAL) << "not yet implemented: " - << PrimitiveType_Name(result_shape.element_type()); - } -} - -Literal LiteralBase::Clone() const { - Literal result(shape()); - TF_CHECK_OK(result.CopyFrom(*this)); - return result; -} - -std::unique_ptr LiteralBase::CloneToUnique() const { - auto result = MakeUnique(shape()); - TF_CHECK_OK(result->CopyFrom(*this)); - return result; -} - -string LiteralBase::GetAsString(tensorflow::gtl::ArraySlice multi_index, - const ShapeIndex& shape_index) const { - const Shape& subshape = ShapeUtil::GetSubshape(shape(), shape_index); - CHECK(LayoutUtil::IsDenseArray(subshape)); - switch (subshape.element_type()) { +/* static */ Literal LiteralUtil::GetFirstScalarLiteral( + const LiteralSlice& literal) { + CHECK(ShapeUtil::IsArray(literal.shape())); + CHECK_GT(ShapeUtil::ElementsIn(literal.shape()), 0); + switch (literal.shape().element_type()) { case PRED: - return Get(multi_index, shape_index) ? "true" : "false"; + return std::move( + *LiteralUtil::CreateR0(literal.GetFirstElement())); + // 8 bit types. case S8: - return StrCat(Get(multi_index, shape_index)); - case S16: - return StrCat(Get(multi_index, shape_index)); - case S32: - return StrCat(Get(multi_index, shape_index)); - case S64: - return StrCat(Get(multi_index, shape_index)); + return std::move( + *LiteralUtil::CreateR0(literal.GetFirstElement())); case U8: - return StrCat(Get(multi_index, shape_index)); - case U16: - return StrCat(Get(multi_index, shape_index)); - case U32: - return StrCat(Get(multi_index, shape_index)); - case U64: - return StrCat(Get(multi_index, shape_index)); - case F16: - return StrCat(static_cast(Get(multi_index, shape_index))); - case F32: - return StrCat(Get(multi_index, shape_index)); + return std::move( + *LiteralUtil::CreateR0(literal.GetFirstElement())); + // 16 bit types. case BF16: - return StrCat( - static_cast(Get(multi_index, shape_index))); - case F64: - return StrCat(Get(multi_index, shape_index)); - case C64: { - complex64 c = Get(multi_index, shape_index); - return StrCat("(", c.real(), ", ", c.imag(), ")"); - } - default: - LOG(FATAL) << PrimitiveType_Name(subshape.element_type()); - } -} - -string LiteralBase::GetSparseElementAsString( - int64 sparse_element_number, const ShapeIndex& shape_index) const { - const Shape& subshape = ShapeUtil::GetSubshape(shape(), shape_index); - CHECK(LayoutUtil::IsSparseArray(subshape)); - switch (subshape.element_type()) { - case PRED: - return GetSparseElement(sparse_element_number, shape_index) - ? "true" - : "false"; - case S8: - return StrCat(GetSparseElement(sparse_element_number, shape_index)); + return std::move(*LiteralUtil::CreateR0( + literal.GetFirstElement())); + case F16: + return std::move( + *LiteralUtil::CreateR0(literal.GetFirstElement())); case S16: - return StrCat( - GetSparseElement(sparse_element_number, shape_index)); - case S32: - return StrCat( - GetSparseElement(sparse_element_number, shape_index)); - case S64: - return StrCat( - GetSparseElement(sparse_element_number, shape_index)); - case U8: - return StrCat( - GetSparseElement(sparse_element_number, shape_index)); + return std::move( + *LiteralUtil::CreateR0(literal.GetFirstElement())); case U16: - return StrCat( - GetSparseElement(sparse_element_number, shape_index)); - case U32: - return StrCat( - GetSparseElement(sparse_element_number, shape_index)); - case U64: - return StrCat( - GetSparseElement(sparse_element_number, shape_index)); - case F16: - return StrCat(static_cast( - GetSparseElement(sparse_element_number, shape_index))); + return std::move( + *LiteralUtil::CreateR0(literal.GetFirstElement())); + // 32 bit types. case F32: - return StrCat( - GetSparseElement(sparse_element_number, shape_index)); - case BF16: - return StrCat(static_cast( - GetSparseElement(sparse_element_number, shape_index))); - case F64: - return StrCat( - GetSparseElement(sparse_element_number, shape_index)); - case C64: { - complex64 c = - GetSparseElement(sparse_element_number, shape_index); - return StrCat("(", c.real(), ", ", c.imag(), ")"); - } - default: - LOG(FATAL) << "Invalid element type for sparse arrays: " - << PrimitiveType_Name(subshape.element_type()); - } -} - -StatusOr LiteralBase::GetIntegralAsS64( - tensorflow::gtl::ArraySlice multi_index) const { - CHECK(LayoutUtil::IsDenseArray(shape())); - switch (shape().element_type()) { - case PRED: - return Get(multi_index); - case U8: - return Get(multi_index); + return std::move( + *LiteralUtil::CreateR0(literal.GetFirstElement())); case S32: - return Get(multi_index); - case S64: - return Get(multi_index); + return std::move( + *LiteralUtil::CreateR0(literal.GetFirstElement())); case U32: - return Get(multi_index); + return std::move( + *LiteralUtil::CreateR0(literal.GetFirstElement())); + // 64 bit types. + case C64: + return std::move(*LiteralUtil::CreateR0( + literal.GetFirstElement())); + case F64: + return std::move( + *LiteralUtil::CreateR0(literal.GetFirstElement())); + case S64: + return std::move( + *LiteralUtil::CreateR0(literal.GetFirstElement())); case U64: - return Get(multi_index); + return std::move( + *LiteralUtil::CreateR0(literal.GetFirstElement())); default: - return FailedPrecondition( - "Array element type is not integral: %s", - PrimitiveType_Name(shape().element_type()).c_str()); + LOG(FATAL) << "Unhandled primitive type " + << literal.shape().element_type(); } } -size_t LiteralBase::Hash() const { - using tensorflow::Hash64; - using tensorflow::Hash64Combine; - - size_t hash_value = ShapeUtil::Hash(shape()); - - ShapeUtil::ForEachSubshape( - shape(), [&](const Shape& subshape, const ShapeIndex& index) { - if (!ShapeUtil::IsArray(subshape)) { - return; - } - - CHECK(LayoutUtil::IsDense(subshape.layout())); - hash_value = Hash64Combine( - hash_value, Hash64(static_cast(untyped_data(index)), - size_bytes(index))); - }); - - return hash_value; +/* static */ std::unique_ptr LiteralUtil::MakeTuple( + tensorflow::gtl::ArraySlice elements) { + std::vector element_shapes; + for (const auto* element : elements) { + element_shapes.push_back(element->shape()); + } + auto literal = MakeUnique(ShapeUtil::MakeTupleShape(element_shapes)); + for (int i = 0; i < elements.size(); ++i) { + TF_CHECK_OK(literal->CopyFrom(*elements[i], /*dest_shape_index=*/{i})); + } + return literal; } -Status Literal::SetIntegralAsS64(tensorflow::gtl::ArraySlice multi_index, - int64 value) { - CHECK(LayoutUtil::IsDenseArray(shape())); - switch (shape().element_type()) { - case PRED: - Set(multi_index, value); - break; - case U8: - Set(multi_index, value); - break; - case S32: - Set(multi_index, value); - break; - case S64: - Set(multi_index, value); - break; - case U32: - Set(multi_index, value); - break; - case U64: - Set(multi_index, value); - break; - default: - return FailedPrecondition( - "Array element type is not integral: %s", - PrimitiveType_Name(shape().element_type()).c_str()); - } - return Status::OK(); -} - -tensorflow::gtl::ArraySlice LiteralBase::GetSparseIndex( - int64 sparse_element_number, const ShapeIndex& shape_index) const { - const Piece& p = piece(shape_index); - CHECK_GE(sparse_element_number, 0); - CHECK_LT(sparse_element_number, p.sparse_indices()->index_count()); - return p.sparse_indices()->At(sparse_element_number); -} - -void Literal::SortSparseElements(const ShapeIndex& shape_index) { - piece(shape_index).SortSparseElements(); -} - -Literal LiteralBase::GetFirstScalarLiteral() const { - CHECK(ShapeUtil::IsArray(shape())); - CHECK_GT(ShapeUtil::ElementsIn(shape()), 0); - switch (shape().element_type()) { - case PRED: - return std::move(*Literal::CreateR0(GetFirstElement())); - // 8 bit types. - case S8: - return std::move(*Literal::CreateR0(GetFirstElement())); - case U8: - return std::move(*Literal::CreateR0(GetFirstElement())); - // 16 bit types. - case BF16: - return std::move( - *Literal::CreateR0(GetFirstElement())); - case F16: - return std::move(*Literal::CreateR0(GetFirstElement())); - case S16: - return std::move(*Literal::CreateR0(GetFirstElement())); - case U16: - return std::move(*Literal::CreateR0(GetFirstElement())); - // 32 bit types. - case F32: - return std::move(*Literal::CreateR0(GetFirstElement())); - case S32: - return std::move(*Literal::CreateR0(GetFirstElement())); - case U32: - return std::move(*Literal::CreateR0(GetFirstElement())); - // 64 bit types. - case C64: - return std::move( - *Literal::CreateR0(GetFirstElement())); - case F64: - return std::move(*Literal::CreateR0(GetFirstElement())); - case S64: - return std::move(*Literal::CreateR0(GetFirstElement())); - case U64: - return std::move(*Literal::CreateR0(GetFirstElement())); - default: - LOG(FATAL) << "Unhandled primitive type " << shape().element_type(); - } -} - -void LiteralBase::Piece::SortSparseElements() { - switch (subshape().element_type()) { - case PRED: - SortSparseElementsInternal(); - break; - case S8: - SortSparseElementsInternal(); - break; - case U8: - SortSparseElementsInternal(); - break; - case S16: - SortSparseElementsInternal(); - break; - case U16: - SortSparseElementsInternal(); - break; - case S32: - SortSparseElementsInternal(); - break; - case U32: - SortSparseElementsInternal(); - break; - case S64: - SortSparseElementsInternal(); - break; - case U64: - SortSparseElementsInternal(); - break; - case F32: - SortSparseElementsInternal(); - break; - case F64: - SortSparseElementsInternal(); - break; - case C64: - SortSparseElementsInternal(); - break; - case F16: - SortSparseElementsInternal(); - break; - case BF16: - SortSparseElementsInternal(); - break; - default: - LOG(FATAL) << "Element type not valid for sparse array: " - << PrimitiveType_Name(subshape().element_type()); - } -} - -template -void LiteralBase::Piece::SortSparseElementsInternal() { - CHECK(LayoutUtil::IsSparseArray(subshape())); - int64 num_elements = sparse_indices()->index_count(); - auto values = data(); - CHECK_LE(num_elements, values.size()); - sparse_indices()->SortWithValues( - tensorflow::gtl::MutableArraySlice(values.data(), num_elements)); -} - -namespace { - -void ToStringHelper(const LiteralBase& literal, const ShapeIndex& shape_index, - bool print_layout, std::vector* pieces) { - const Shape& subshape = ShapeUtil::GetSubshape(literal.shape(), shape_index); - CHECK(LayoutUtil::HasLayout(literal.shape())); - CHECK(LayoutUtil::HasLayout(subshape)); - - auto shape_to_string = [print_layout](const Shape& shape) { - if (print_layout) { - return ShapeUtil::HumanStringWithLayout(shape); - } else { - return ShapeUtil::HumanString(shape); - } - }; - - // TODO(b/32894291): refactor this code to reduce code duplication. - if (ShapeUtil::IsTuple(subshape)) { - pieces->push_back(shape_to_string(subshape)); - pieces->push_back(" (\n"); - std::vector tuple_pieces; - for (int i = 0; i < ShapeUtil::TupleElementCount(subshape); ++i) { - ShapeIndex element_index = shape_index; - element_index.push_back(i); - std::vector element_pieces; - ToStringHelper(literal, element_index, print_layout, &element_pieces); - tuple_pieces.push_back(tensorflow::str_util::Join(element_pieces, "")); - } - pieces->push_back(tensorflow::str_util::Join(tuple_pieces, ",\n")); - pieces->push_back("\n)"); - return; - } - - if (ShapeUtil::IsToken(subshape)) { - pieces->push_back("token"); - return; - } - - if (LayoutUtil::IsSparseArray(subshape)) { - pieces->push_back(shape_to_string(subshape)); - pieces->push_back("{"); - int64 rank = ShapeUtil::Rank(subshape); - int64 num_elements = literal.sparse_element_count(); - for (int64 i = 0; i < num_elements; ++i) { - if (i > 0) { - pieces->push_back(", "); - } - if (rank == 1) { - pieces->push_back(StrCat(literal.GetSparseIndex(i)[0])); - pieces->push_back(": "); - } else { - pieces->push_back("["); - pieces->push_back( - tensorflow::str_util::Join(literal.GetSparseIndex(i), ", ")); - pieces->push_back("]: "); - } - pieces->push_back(literal.GetSparseElementAsString(i)); - } - pieces->push_back("}"); - return; - } - - CHECK(LayoutUtil::IsDenseArray(subshape)); - - auto element_to_string = - [&](tensorflow::gtl::ArraySlice indices) -> string { - PrimitiveType element_type = subshape.element_type(); - if (element_type == PRED) { - // We display predicates in a densely packed form. - return literal.Get(indices, shape_index) ? "1" : "0"; - } - return ((!indices.empty() && indices.back() > 0) ? ", " : "") + - literal.GetAsString(indices, shape_index); - }; - - if (ShapeUtil::Rank(subshape) == 0) { - pieces->push_back(literal.GetAsString({}, shape_index)); - } else if (ShapeUtil::Rank(subshape) == 1) { - pieces->push_back("{"); - for (int64 i0 = 0; i0 < subshape.dimensions(0); ++i0) { - pieces->push_back(element_to_string({i0})); - } - pieces->push_back("}"); - } else if (ShapeUtil::Rank(subshape) == 2) { - pieces->push_back(shape_to_string(subshape)); - pieces->push_back(" {\n"); - for (int64 i0 = 0; i0 < subshape.dimensions(0); ++i0) { - pieces->push_back(" { "); - for (int64 i1 = 0; i1 < subshape.dimensions(1); ++i1) { - pieces->push_back(element_to_string({i0, i1})); - } - pieces->push_back(" "); - pieces->push_back(i0 == subshape.dimensions(0) - 1 ? "}\n" : "},\n"); - } - pieces->push_back("}"); - } else if (ShapeUtil::Rank(subshape) == 3) { - pieces->push_back(shape_to_string(subshape)); - pieces->push_back(" {\n"); - for (int64 i0 = 0; i0 < subshape.dimensions(0); ++i0) { - pieces->push_back(i0 > 0 ? ",\n{" : "{"); - for (int64 i1 = 0; i1 < subshape.dimensions(1); ++i1) { - pieces->push_back(i1 > 0 ? ",\n { " : " { "); - for (int64 i2 = 0; i2 < subshape.dimensions(2); ++i2) { - pieces->push_back(element_to_string({i0, i1, i2})); - } - pieces->push_back(" }"); - } - pieces->push_back(" }"); - } - pieces->push_back("\n}"); - } else if (ShapeUtil::Rank(subshape) == 4) { - pieces->push_back(shape_to_string(subshape)); - pieces->push_back(" {\n"); - for (int64 i0 = 0; i0 < subshape.dimensions(0); ++i0) { - pieces->push_back(Printf(" { /*i0=%lld*/\n", i0)); - for (int64 i1 = 0; i1 < subshape.dimensions(1); ++i1) { - pieces->push_back(Printf(" { /*i1=%lld*/\n", i1)); - for (int64 i2 = 0; i2 < subshape.dimensions(2); ++i2) { - pieces->push_back(" {"); - for (int64 i3 = 0; i3 < subshape.dimensions(3); ++i3) { - pieces->push_back(element_to_string({i0, i1, i2, i3})); - } - pieces->push_back(i2 == subshape.dimensions(2) - 1 ? "}\n" : "},\n"); - } - pieces->push_back(i1 == subshape.dimensions(1) - 1 ? " }\n" - : " },\n"); - } - pieces->push_back(i0 == subshape.dimensions(0) - 1 ? " }\n" : " },\n"); - } - pieces->push_back("}"); - } else if (ShapeUtil::Rank(subshape) == 5) { - pieces->push_back(shape_to_string(subshape)); - pieces->push_back(" {\n"); - for (int64 i0 = 0; i0 < subshape.dimensions(0); ++i0) { - pieces->push_back(Printf(" { /*i0=%lld*/\n", i0)); - for (int64 i1 = 0; i1 < subshape.dimensions(1); ++i1) { - pieces->push_back(Printf(" { /*i1=%lld*/\n", i1)); - for (int64 i2 = 0; i2 < subshape.dimensions(2); ++i2) { - pieces->push_back(Printf(" { /*i2=%lld*/\n", i2)); - for (int64 i3 = 0; i3 < subshape.dimensions(3); ++i3) { - pieces->push_back(" {"); - for (int64 i4 = 0; i4 < subshape.dimensions(4); ++i4) { - pieces->push_back(element_to_string({i0, i1, i2, i3, i4})); - } - pieces->push_back(i3 == subshape.dimensions(3) - 1 ? "}\n" - : "},\n"); - } - pieces->push_back(i2 == subshape.dimensions(2) - 1 ? " }\n" - : " },\n"); - } - pieces->push_back(i1 == subshape.dimensions(1) - 1 ? " }\n" - : " },\n"); - } - pieces->push_back(i0 == subshape.dimensions(0) - 1 ? " }\n" : " },\n"); - } - pieces->push_back("}"); - } else { - pieces->push_back(shape_to_string(subshape)); - pieces->push_back(" {"); - literal.EachCellAsString( - [&](tensorflow::gtl::ArraySlice indices, const string& value) { - pieces->push_back(" "); - pieces->push_back(value); - }); - pieces->push_back("}"); - } -} - -} // namespace - -int64 LiteralBase::sparse_element_count() const { - CHECK(LayoutUtil::IsSparseArray(shape())); - return sparse_indices()->index_count(); -} - -string LiteralBase::ToString(bool print_layout) const { - std::vector pieces; - CHECK(LayoutUtil::HasLayout(this->shape())); - ToStringHelper(*this, {}, print_layout, &pieces); - return tensorflow::str_util::Join(pieces, ""); -} - -/* static */ std::unique_ptr Literal::MakeTuple( - tensorflow::gtl::ArraySlice elements) { - std::vector element_shapes; - for (const auto* element : elements) { - element_shapes.push_back(element->shape()); - } - auto literal = MakeUnique(ShapeUtil::MakeTupleShape(element_shapes)); - for (int i = 0; i < elements.size(); ++i) { - TF_CHECK_OK(literal->CopyFrom(*elements[i], /*dest_shape_index=*/{i})); - } - return literal; -} - -/* static */ std::unique_ptr Literal::MakeTupleFromSlices( +/* static */ std::unique_ptr LiteralUtil::MakeTupleFromSlices( tensorflow::gtl::ArraySlice elements) { std::vector element_shapes; for (const auto& element : elements) { @@ -1550,7 +456,7 @@ string LiteralBase::ToString(bool print_layout) const { return literal; } -/* static */ std::unique_ptr Literal::MakeTupleOwned( +/* static */ std::unique_ptr LiteralUtil::MakeTupleOwned( std::vector> elements) { std::vector element_shapes; element_shapes.reserve(elements.size()); @@ -1565,818 +471,9 @@ string LiteralBase::ToString(bool print_layout) const { return literal; } -void LiteralBase::EachCellAsString( - const std::function indices, - const string& value)>& per_cell) const { - if (ShapeUtil::IsZeroElementArray(shape())) { - return; - } - std::vector indices = IndexUtil::LinearIndexToMultidimensionalIndex( - shape(), /*linear_index=*/0); - do { - per_cell(indices, GetAsString(indices)); - } while (IndexUtil::BumpIndices(shape(), &indices)); -} - -namespace { -template -std::unique_ptr ConvertBetweenNativeTypesWithConverter( - const LiteralBase& src_literal, const ConverterType& converter) { - CHECK(ShapeUtil::IsArray(src_literal.shape())); - auto result_literal = MakeUnique(ShapeUtil::ChangeElementType( - src_literal.shape(), - primitive_util::NativeToPrimitiveType())); - auto src_data = src_literal.data(); - auto dest_data = result_literal->template data(); - int64 num_elements = src_literal.element_count(); - - for (int64 i = 0; i < num_elements; ++i) { - dest_data[i] = converter(src_data[i]); - } - return result_literal; -} - -template -std::unique_ptr ConvertBetweenNativeTypes( - const LiteralBase& src_literal) { - auto converter = [](NativeSrcT src) { return static_cast(src); }; - return ConvertBetweenNativeTypesWithConverter( - src_literal, converter); -} - -template -typename std::enable_if<(sizeof(NativeSrcT) == sizeof(NativeDestT)), - std::unique_ptr>::type -BitcastBetweenNativeTypes(const LiteralBase& src_literal) { - auto converter = [](NativeSrcT src) { - return tensorflow::bit_cast(src); - }; - return ConvertBetweenNativeTypesWithConverter( - src_literal, converter); -} - -// This template specialization is here to make the compiler happy. bit_cast has -// a static check that the types are the same size. This specialization should -// never be used because the source and destination types are checked for -// identical sizes higher up. -template -typename std::enable_if<(sizeof(NativeSrcT) != sizeof(NativeDestT)), - std::unique_ptr>::type -BitcastBetweenNativeTypes(const LiteralBase& src_literal) { - LOG(FATAL) << "Invalid bitcast between types of different sizes."; -} - -template -std::unique_ptr ConvertToC64(const LiteralBase& src_literal) { - CHECK(ShapeUtil::IsArray(src_literal.shape())); - auto result_literal = MakeUnique( - ShapeUtil::ChangeElementType(src_literal.shape(), C64)); - using NativeSrcT = - typename primitive_util::PrimitiveTypeToNative::type; - tensorflow::gtl::ArraySlice src_data = - src_literal.data(); - tensorflow::gtl::MutableArraySlice dest_data = - result_literal->data(); - int64 num_elements = src_literal.element_count(); - for (int64 i = 0; i < num_elements; ++i) { - dest_data[i] = complex64(static_cast(src_data[i]), 0); - } - return result_literal; -} - -template -std::unique_ptr ConvertIfTypesMatch(const LiteralBase& src_literal, - bool bitcast) { - CHECK_EQ(primitive_src_type, src_literal.shape().element_type()); - if (bitcast) { - return BitcastBetweenNativeTypes< - typename primitive_util::PrimitiveTypeToNative< - primitive_src_type>::type, - typename primitive_util::PrimitiveTypeToNative< - primitive_dest_type>::type>(src_literal); - } else { - return ConvertBetweenNativeTypes< - typename primitive_util::PrimitiveTypeToNative< - primitive_src_type>::type, - typename primitive_util::PrimitiveTypeToNative< - primitive_dest_type>::type>(src_literal); - } -} - -template -StatusOr> ConvertIfDestTypeMatches( - const LiteralBase& src_literal, PrimitiveType primitive_dest_type, - bool bitcast) { - switch (primitive_dest_type) { -#define CONVERT_IF_TYPES_MATCH(type) \ - case (type): \ - return ConvertIfTypesMatch(src_literal, \ - bitcast); - CONVERT_IF_TYPES_MATCH(PRED) - CONVERT_IF_TYPES_MATCH(S8) - CONVERT_IF_TYPES_MATCH(S32) - CONVERT_IF_TYPES_MATCH(S64) - CONVERT_IF_TYPES_MATCH(U8) - CONVERT_IF_TYPES_MATCH(U32) - CONVERT_IF_TYPES_MATCH(U64) - CONVERT_IF_TYPES_MATCH(F16) - CONVERT_IF_TYPES_MATCH(F32) - CONVERT_IF_TYPES_MATCH(F64) - CONVERT_IF_TYPES_MATCH(BF16) -#undef CONVERT_IF_TYPES_MATCH - case C64: - if (!bitcast) { - return ConvertToC64(src_literal); - } - break; - // Other types are not yet supported. - default: - break; - } - return Unimplemented( - "Converting from type %s to type %s is not implemented.", - PrimitiveType_Name(src_literal.shape().element_type()).c_str(), - PrimitiveType_Name(primitive_dest_type).c_str()); -} - -StatusOr> ConvertSwitch( - const LiteralBase& literal, PrimitiveType primitive_dest_type, - bool bitcast) { - TF_RET_CHECK(ShapeUtil::IsArray(literal.shape())); - if (literal.shape().element_type() == primitive_dest_type) { - return literal.CloneToUnique(); - } - switch (literal.shape().element_type()) { -#define CONVERT_IF_DEST_TYPE_MATCHES(type) \ - case (type): \ - return ConvertIfDestTypeMatches<(type)>(literal, primitive_dest_type, \ - bitcast); - CONVERT_IF_DEST_TYPE_MATCHES(PRED) - CONVERT_IF_DEST_TYPE_MATCHES(S8) - CONVERT_IF_DEST_TYPE_MATCHES(S32) - CONVERT_IF_DEST_TYPE_MATCHES(S64) - CONVERT_IF_DEST_TYPE_MATCHES(U8) - CONVERT_IF_DEST_TYPE_MATCHES(U32) - CONVERT_IF_DEST_TYPE_MATCHES(U64) - CONVERT_IF_DEST_TYPE_MATCHES(F16) - CONVERT_IF_DEST_TYPE_MATCHES(F32) - CONVERT_IF_DEST_TYPE_MATCHES(F64) - CONVERT_IF_DEST_TYPE_MATCHES(BF16) -#undef CONVERT_IF_DEST_TYPE_MATCHES - // Other types are not yet supported. - default: - return Unimplemented( - "%s from type %s to type %s is not implemented.", - (bitcast ? "Bitcast converting" : "Converting"), - PrimitiveType_Name(literal.shape().element_type()).c_str(), - PrimitiveType_Name(primitive_dest_type).c_str()); - } -} - -} // namespace - -StatusOr> LiteralBase::Convert( - PrimitiveType primitive_dest_type) const { - return ConvertSwitch(*this, primitive_dest_type, /*bitcast=*/false); -} - -StatusOr> LiteralBase::BitcastConvert( - PrimitiveType primitive_dest_type) const { - if (primitive_util::BitWidth(shape().element_type()) != - primitive_util::BitWidth(primitive_dest_type)) { - return InvalidArgument( - "Cannot bitcast convert from %s to %s, bit widths are different: %d != " - "%d", - PrimitiveType_Name(shape().element_type()).c_str(), - PrimitiveType_Name(primitive_dest_type).c_str(), - primitive_util::BitWidth(shape().element_type()), - primitive_util::BitWidth(primitive_dest_type)); - } - return ConvertSwitch(*this, primitive_dest_type, /*bitcast=*/true); -} - -StatusOr> LiteralBase::ConvertToShape( - const Shape& dest_shape, bool round_f32_to_bf16) const { - if (!ShapeUtil::IsTuple(dest_shape)) { - if (round_f32_to_bf16 && shape().element_type() == F32 && - dest_shape.element_type() == BF16) { - auto converter = [](float src) { - return tensorflow::bfloat16::round_to_bfloat16(src); - }; - return ConvertBetweenNativeTypesWithConverter(*this, - converter); - } - return Convert(dest_shape.element_type()); - } - std::vector elements; - for (int i = 0; i < ShapeUtil::TupleElementCount(shape()); ++i) { - auto element = LiteralSlice(*this, {i}); - TF_ASSIGN_OR_RETURN( - auto new_element, - element.ConvertToShape(ShapeUtil::GetSubshape(dest_shape, {i}))); - elements.push_back(std::move(*new_element)); - } - auto converted = MakeUnique(); - *converted = Literal::MoveIntoTuple(&elements); - return std::move(converted); -} - -template -bool LiteralBase::Piece::EqualElementsInternal( - const LiteralBase::Piece& other, std::vector* multi_index) const { - if (multi_index->size() == ShapeUtil::Rank(subshape())) { - return (Get(*multi_index) == other.Get(*multi_index)); - } - for (int64 i = 0; i < subshape().dimensions(multi_index->size()); ++i) { - multi_index->push_back(i); - if (!EqualElementsInternal(other, multi_index)) { - return false; - } - multi_index->pop_back(); - } - return true; -} - -bool LiteralBase::Piece::EqualElements(const LiteralBase::Piece& other) const { - DCHECK(ShapeUtil::Compatible(subshape(), other.subshape())); - - std::vector multi_index; - switch (subshape().element_type()) { - case PRED: - return EqualElementsInternal(other, &multi_index); - case U8: - return EqualElementsInternal(other, &multi_index); - case S32: - return EqualElementsInternal(other, &multi_index); - case S64: - return EqualElementsInternal(other, &multi_index); - case U32: - return EqualElementsInternal(other, &multi_index); - case U64: - return EqualElementsInternal(other, &multi_index); - case F32: - return EqualElementsInternal(other, &multi_index); - case F64: - return EqualElementsInternal(other, &multi_index); - case F16: - return EqualElementsInternal(other, &multi_index); - case BF16: - return EqualElementsInternal(other, &multi_index); - case C64: - return EqualElementsInternal(other, &multi_index); - default: - LOG(FATAL) << "Unimplemented: LiteralBase::Piece::EqualElements for type " - << PrimitiveType_Name(subshape().element_type()); - } -} - -bool LiteralBase::operator==(const LiteralBase& other) const { - if (!ShapeUtil::Compatible(shape(), other.shape())) { - return false; - } - - return root_piece().ForEachSubpieceWithBool( - [&](const ShapeIndex& index, const Piece& piece) { - if (!ShapeUtil::IsArray(piece.subshape())) { - return true; - } - - const Piece& other_piece = other.piece(index); - if (!piece.EqualElements(other_piece)) { - return false; - } - return true; - }); -} - -namespace { - -template -static bool AllElementsEqualValue(tensorflow::gtl::ArraySlice data, - NativeT value) { - for (int64 i = 0; i < data.size(); ++i) { - if (data[i] != value) { - return false; - } - } - return true; -} - -} // namespace - -bool LiteralBase::IsAll(int8 value) const { - return root_piece().ForEachSubpieceWithBool([&](const ShapeIndex& index, - const Piece& piece) { - if (!ShapeUtil::IsArray(piece.subshape())) { - return true; - } - - auto piece_is_all = [&]() { - switch (shape().element_type()) { - case U8: - if (value >= 0) { - return AllElementsEqualValue(piece.data(), value); - } - return false; - case U32: - if (value >= 0) { - return AllElementsEqualValue(piece.data(), value); - } - return false; - case U64: - if (value >= 0) { - return AllElementsEqualValue(piece.data(), value); - } - return false; - case S8: - return AllElementsEqualValue(piece.data(), value); - case S32: - return AllElementsEqualValue(piece.data(), value); - case S64: - return AllElementsEqualValue(piece.data(), value); - case F32: - return AllElementsEqualValue(piece.data(), value); - case F64: - return AllElementsEqualValue(piece.data(), value); - case F16: - return AllElementsEqualValue(piece.data(), - static_cast(value)); - case BF16: - return AllElementsEqualValue(piece.data(), - static_cast(value)); - case PRED: - if (value == 0) { - return AllElementsEqualValue(piece.data(), false); - } - if (value == 1) { - return AllElementsEqualValue(piece.data(), true); - } - return false; - default: - return false; - } - return false; - }; - - if (!piece_is_all()) { - return false; - } - return true; - }); -} - -bool LiteralBase::IsAllFloat(float value) const { - return root_piece().ForEachSubpieceWithBool( - [&](const ShapeIndex& index, const Piece& piece) { - if (!ShapeUtil::IsArray(piece.subshape())) { - return true; - } - - auto piece_is_all = [&]() { - switch (shape().element_type()) { - case F32: - return AllElementsEqualValue(piece.data(), value); - case F64: - return AllElementsEqualValue(piece.data(), value); - case F16: - return AllElementsEqualValue(piece.data(), - static_cast(value)); - case BF16: - return AllElementsEqualValue( - piece.data(), static_cast(value)); - default: - return false; - } - }; - if (!piece_is_all()) { - return false; - } - return true; - }); -} - -bool LiteralBase::IsAllComplex(complex64 value) const { - switch (shape().element_type()) { - case C64: - return AllElementsEqualValue(root_piece().data(), - value); - default: - return false; - } -} - -bool LiteralBase::IsAllFirst() const { - return root_piece().ForEachSubpieceWithBool( - [&](const ShapeIndex& index, const Piece& piece) { - if (!ShapeUtil::IsArray(piece.subshape())) { - return true; - } - - // Empty shapes are not all the first element since there is no first - // element. - if (ShapeUtil::IsZeroElementArray(piece.subshape())) { - return false; - } - auto piece_is_all = [&]() { - switch (piece.subshape().element_type()) { - case PRED: { - auto data = piece.data(); - return AllElementsEqualValue(data, data[0]); - } - // 8 bit types - case S8: { - auto data = piece.data(); - return AllElementsEqualValue(data, data[0]); - } - case U8: { - auto data = piece.data(); - return AllElementsEqualValue(data, data[0]); - } - // 16 bit types - case BF16: { - auto data = piece.data(); - return AllElementsEqualValue(data, data[0]); - } - case F16: { - auto data = piece.data(); - return AllElementsEqualValue(data, data[0]); - } - case S16: { - auto data = piece.data(); - return AllElementsEqualValue(data, data[0]); - } - case U16: { - auto data = piece.data(); - return AllElementsEqualValue(data, data[0]); - } - // 32 bit types - case F32: { - auto data = piece.data(); - return AllElementsEqualValue(data, data[0]); - } - case U32: { - auto data = piece.data(); - return AllElementsEqualValue(data, data[0]); - } - case S32: { - auto data = piece.data(); - return AllElementsEqualValue(data, data[0]); - } - // 64 bit types - case C64: { - auto data = piece.data(); - return AllElementsEqualValue(data, data[0]); - } - case F64: { - auto data = piece.data(); - return AllElementsEqualValue(data, data[0]); - } - case S64: { - auto data = piece.data(); - return AllElementsEqualValue(data, data[0]); - } - case U64: { - auto data = piece.data(); - return AllElementsEqualValue(data, data[0]); - } - default: - return false; - } - }; - - if (!piece_is_all()) { - return false; - } - return true; - }); -} - -bool LiteralBase::IsZero(tensorflow::gtl::ArraySlice indices) const { - CHECK(ShapeUtil::IsArray(shape())); - switch (shape().element_type()) { - case U8: - return Get(indices) == 0; - case U32: - return Get(indices) == 0; - case U64: - return Get(indices) == 0; - case S8: - return Get(indices) == 0; - case S32: - return Get(indices) == 0; - case S64: - return Get(indices) == 0; - case F32: - return Get(indices) == 0.0f; - case F64: - return Get(indices) == 0.0; - case C64: - return Get(indices) == complex64(0.0f, 0.0f); - case F16: - return Get(indices) == static_cast(0.0f); - case BF16: - return Get(indices) == static_cast(0.0f); - case PRED: - return Get(indices) == false; - default: - LOG(FATAL) << "Input literal must be an array."; - } -} - -namespace { - -template -void CopyToRepeatedField(RepeatedFieldT* dest, - const tensorflow::gtl::ArraySlice src) { - *dest = RepeatedFieldT(src.begin(), src.end()); -} - -} // namespace - -void LiteralBase::Piece::WriteToProto(LiteralProto* proto) const { - *proto->mutable_shape() = subshape(); - switch (subshape().element_type()) { - case PRED: - CopyToRepeatedField(proto->mutable_preds(), data()); - break; - case U8: - proto->set_u8s(static_cast(data().data()), - element_count()); - break; - case U32: - CopyToRepeatedField(proto->mutable_u32s(), data()); - break; - case U64: - CopyToRepeatedField(proto->mutable_u64s(), data()); - break; - case S32: - CopyToRepeatedField(proto->mutable_s32s(), data()); - break; - case S64: - CopyToRepeatedField(proto->mutable_s64s(), data()); - break; - case F16: - *proto->mutable_f16s() = string( - reinterpret_cast(data().data()), size_bytes()); - if (!kLittleEndian) { - ConvertEndianShort(proto->mutable_f16s()); - } - break; - case BF16: - *proto->mutable_bf16s() = string( - reinterpret_cast(data().data()), size_bytes()); - if (!kLittleEndian) { - ConvertEndianShort(proto->mutable_bf16s()); - } - break; - case F32: - CopyToRepeatedField(proto->mutable_f32s(), data()); - break; - case F64: - CopyToRepeatedField(proto->mutable_f64s(), data()); - break; - case C64: - for (complex64 value : data()) { - proto->add_c64s(value.real()); - proto->add_c64s(value.imag()); - } - break; - case TUPLE: - // Nothing to do but assign the shape which is done above. - return; - default: - LOG(FATAL) << "Unhandled primitive type " << subshape().element_type(); - } -} - -const void* LiteralBase::Piece::untyped_data() const { - CHECK(ShapeUtil::IsArray(subshape())) << ShapeUtil::HumanString(subshape()); - return buffer(); -} - -void* LiteralBase::Piece::untyped_data() { - CHECK(ShapeUtil::IsArray(subshape())) << ShapeUtil::HumanString(subshape()); - return buffer(); -} - -namespace { - -template -Status CopyFromRepeatedField(tensorflow::gtl::MutableArraySlice dest, - const RepeatedFieldT& src) { - if (dest.size() != src.size()) { - return InvalidArgument( - "Expected %lu elements in LiteralProto repeated field, has %d", - dest.size(), src.size()); - } - std::copy(src.begin(), src.end(), dest.begin()); - return Status::OK(); -} - -} // namespace - -Status LiteralBase::Piece::CopyFromProto(const LiteralProto& proto) { - // These conditions should have been checked in Literal::CreateFromProto. - TF_RET_CHECK(proto.has_shape()); - TF_RET_CHECK(LayoutUtil::HasLayout(proto.shape())); - TF_RET_CHECK(ShapeUtil::Equal(proto.shape(), subshape())); - - switch (subshape().element_type()) { - case PRED: - TF_RETURN_IF_ERROR(CopyFromRepeatedField(data(), proto.preds())); - break; - case U8: { - auto u8_data = data(); - TF_RET_CHECK(proto.u8s().size() == u8_data.size()); - std::copy(proto.u8s().begin(), proto.u8s().end(), u8_data.begin()); - } break; - case S32: - TF_RETURN_IF_ERROR(CopyFromRepeatedField(data(), proto.s32s())); - break; - case S64: - TF_RETURN_IF_ERROR(CopyFromRepeatedField(data(), proto.s64s())); - break; - case U32: - TF_RETURN_IF_ERROR(CopyFromRepeatedField(data(), proto.u32s())); - break; - case U64: - TF_RETURN_IF_ERROR(CopyFromRepeatedField(data(), proto.u64s())); - break; - case F16: { - const string& s(proto.f16s()); - TF_RET_CHECK(data().size() * sizeof(half) == s.size()); - memcpy(untyped_data(), s.data(), s.size()); - if (!kLittleEndian) { - ConvertEndianShort(reinterpret_cast(untyped_data()), s.size()); - } - } break; - - case BF16: { - const string& s(proto.bf16s()); - TF_RET_CHECK(data().size() * sizeof(bfloat16) == s.size()); - memcpy(untyped_data(), s.data(), s.size()); - if (!kLittleEndian) { - ConvertEndianShort(reinterpret_cast(untyped_data()), s.size()); - } - } break; - case F32: - TF_RETURN_IF_ERROR(CopyFromRepeatedField(data(), proto.f32s())); - break; - case F64: - TF_RETURN_IF_ERROR(CopyFromRepeatedField(data(), proto.f64s())); - break; - case C64: { - auto complex_data = data(); - TF_RET_CHECK(proto.c64s_size() == complex_data.size() * 2); - for (int64 i = 0; i < complex_data.size(); ++i) { - complex_data[i] = complex64{proto.c64s(i * 2), proto.c64s(i * 2 + 1)}; - } - } break; - case TUPLE: - LOG(FATAL) << "Should not be called on tuple shapes: " - << ShapeUtil::HumanString(subshape()); - break; - default: - LOG(FATAL) << "Unhandled primitive type " << subshape().element_type(); - } - return Status::OK(); -} - -LiteralProto LiteralBase::ToProto() const { - LiteralProto proto; - root_piece().ForEachSubpiece( - [&](const ShapeIndex& index, const Piece& piece) { - LiteralProto* proto_piece = &proto; - for (int64 i : index) { - while (proto_piece->tuple_literals_size() <= i) { - proto_piece->add_tuple_literals(); - } - proto_piece = proto_piece->mutable_tuple_literals(i); - } - piece.WriteToProto(proto_piece); - }); - - if (LayoutUtil::IsSparseArray(shape())) { - CopyToRepeatedField(proto.mutable_sparse_indices(), - sparse_indices()->data()); - } - - return proto; -} - -/* static */ -StatusOr> Literal::CreateFromProto( - const LiteralProto& proto) { - if (!proto.has_shape()) { - return InvalidArgument("LiteralProto has no shape"); - } - if (!LayoutUtil::HasLayout(proto.shape())) { - return InvalidArgument("LiteralProto has no layout"); - } - - auto literal = MakeUnique(proto.shape()); - - TF_RETURN_IF_ERROR(literal->root_piece_->ForEachMutableSubpieceWithStatus( - [&](const ShapeIndex& index, Piece* piece) { - const LiteralProto* proto_element = &proto; - for (int64 i : index) { - CHECK(i < proto_element->tuple_literals_size()); - proto_element = &proto_element->tuple_literals(i); - } - - if (ShapeUtil::IsTuple(piece->subshape())) { - if (proto_element->tuple_literals_size() != - ShapeUtil::TupleElementCount(piece->subshape())) { - return InvalidArgument( - "Expected %lld tuple elements in LiteralProto, has %d", - ShapeUtil::TupleElementCount(piece->subshape()), - proto_element->tuple_literals_size()); - } - return Status::OK(); - } - - CHECK(ShapeUtil::IsArray(piece->subshape())); - TF_RETURN_IF_ERROR(piece->CopyFromProto(*proto_element)); - - return Status::OK(); - })); - - return std::move(literal); -} - -/* static */ string Literal::MultiIndexAsString( +/* static */ string LiteralUtil::MultiIndexAsString( tensorflow::gtl::ArraySlice multi_index) { return StrCat("{", tensorflow::str_util::Join(multi_index, ","), "}"); } -const void* LiteralBase::untyped_data(const ShapeIndex& shape_index) const { - return piece(shape_index).untyped_data(); -} - -void* Literal::untyped_data(const ShapeIndex& shape_index) { - return piece(shape_index).untyped_data(); -} - -int64 LiteralBase::size_bytes(const ShapeIndex& shape_index) const { - return piece(shape_index).size_bytes(); -} - -string LiteralBase::GetR1U8AsString() const { - CHECK(ShapeUtil::IsArray(shape())); - CHECK_EQ(ShapeUtil::Rank(shape()), 1); - CHECK_EQ(shape().element_type(), U8); - return string(tensorflow::bit_cast(data().data()), - ShapeUtil::ElementsIn(shape())); -} - -void BorrowingLiteral::BuildPieceSubtree(const Shape& shape, Piece* piece) { - CHECK(ShapeUtil::IsTuple(shape)); - for (int i = 0; i < ShapeUtil::TupleElementCount(shape); ++i) { - const Shape& subshape = shape.tuple_shapes(i); - - auto child_piece = Piece(); - child_piece.set_subshape(&subshape); - - if (ShapeUtil::IsTuple(subshape)) { - BuildPieceSubtree(subshape, &child_piece); - } - - piece->emplace_back(std::move(child_piece)); - } -} - -LiteralSlice::LiteralSlice(const LiteralBase& literal) - : LiteralBase(), root_piece_(&literal.root_piece()) {} - -LiteralSlice::LiteralSlice(const LiteralBase& literal, - const ShapeIndex& view_root) - : LiteralBase(), root_piece_(&literal.piece(view_root)) {} - -BorrowingLiteral::BorrowingLiteral(const char* src_buf_ptr, const Shape& shape) - : LiteralBase(), shape_(MakeUnique(shape)) { - CHECK(ShapeUtil::IsArray(*shape_)); - CHECK(LayoutUtil::HasLayout(*shape_)); - - root_piece_ = Piece(); - root_piece_.set_buffer(const_cast(src_buf_ptr)); - root_piece_.set_subshape(shape_.get()); -} - -BorrowingLiteral::BorrowingLiteral( - tensorflow::gtl::ArraySlice src_buf_ptrs, const Shape& shape) - : LiteralBase(), shape_(MakeUnique(shape)) { - CHECK(ShapeUtil::IsTuple(*shape_)); - CHECK(!ShapeUtil::IsNestedTuple(*shape_)); - CHECK_EQ(src_buf_ptrs.size(), ShapeUtil::TupleElementCount(*shape_)); - root_piece_ = Piece(); - root_piece_.set_subshape(shape_.get()); - BuildPieceSubtree(*shape_, &root_piece_); - - for (int i = 0; i < src_buf_ptrs.size(); ++i) { - const auto& src_shape = shape_->tuple_shapes(i); - CHECK(ShapeUtil::IsArray(src_shape)); - root_piece_.child(i).set_buffer(const_cast(src_buf_ptrs[i])); - } -} - } // namespace xla diff --git a/tensorflow/compiler/xla/literal_util.h b/tensorflow/compiler/xla/literal_util.h index 37ca8ea9f1d158b6bce8d5688288351f55c3b3c8..e3737a9d0051b32dc0becc19e1849c856a50e52e 100644 --- a/tensorflow/compiler/xla/literal_util.h +++ b/tensorflow/compiler/xla/literal_util.h @@ -32,6 +32,7 @@ limitations under the License. #include "tensorflow/compiler/xla/array4d.h" #include "tensorflow/compiler/xla/index_util.h" #include "tensorflow/compiler/xla/layout_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/primitive_util.h" #include "tensorflow/compiler/xla/ptr_util.h" #include "tensorflow/compiler/xla/shape_util.h" @@ -51,679 +52,12 @@ limitations under the License. namespace xla { -// Forward declare Literal and LiteralSlice class to be used by the creation -// methods in the base class. -class Literal; -class LiteralSlice; - -// Abstract base class for literals. -class LiteralBase { +class LiteralUtil { public: - virtual ~LiteralBase() = 0; - - // Literals are equal if they have compatible shapes and the same data - // values. Layout is not compared. - bool operator==(const LiteralBase& other) const; - bool operator!=(const LiteralBase& other) const { return !(*this == other); } - - // Returns the shape of the literal. - const Shape& shape() const { return root_piece().subshape(); } - - // Serialize to proto. - LiteralProto ToProto() const; - - // Returns an ArraySlice of the array for this literal for the given NativeT - // (e.g., float). CHECKs if the subshape of the literal at the given - // ShapeIndex is not array. See primitive_util.h for the mapping from XLA type - // to native type. - template - tensorflow::gtl::ArraySlice data( - const ShapeIndex& shape_index = {}) const; - - // Returns a const pointer to the sparse index array. Returns nullptr if the - // literal is not a sparse array. - const SparseIndexArray* sparse_indices( - const ShapeIndex& shape_index = {}) const; - - // Returns a const pointer to (or size of) the underlying buffer holding the - // array at the given shape index. CHECKs if the subshape of the literal at - // the given ShapeIndex is not array. - const void* untyped_data(const ShapeIndex& shape_index = {}) const; - int64 size_bytes(const ShapeIndex& shape_index = {}) const; - - // Returns this literal's data as a string. This literal must be a rank-1 U8 - // array. - string GetR1U8AsString() const; - - // Returns a string representation of the literal value. - // Warning: this function can take minutes for multi-million element Literals. - string ToString(bool print_layout = false) const; - - // Gets an element in the literal at the given index. The multi_index is - // CHECKed against the dimension sizes. - template - NativeT Get(tensorflow::gtl::ArraySlice multi_index, - const ShapeIndex& shape_index) const; - // Overloads of Get for array literals. CHECKs if the literal is not - // array-shaped and dense. - template - NativeT Get(tensorflow::gtl::ArraySlice multi_index) const; - - // Returns the element value at index (0, ..., 0), however many zeroes are - // required for that index. - template - NativeT GetFirstElement() const; - - // As Get(), but determines the correct type and converts the value - // into text. - string GetAsString(tensorflow::gtl::ArraySlice multi_index, - const ShapeIndex& shape_index = {}) const; - // As GetSparseElement(), but determines the correct type and converts the - // value into text. - string GetSparseElementAsString(int64 sparse_element_number, - const ShapeIndex& shape_index = {}) const; - // As Get(), but determines the correct type and converts the value into - // int64. This literal must be an array. - StatusOr GetIntegralAsS64( - tensorflow::gtl::ArraySlice multi_index) const; - - // Returns the multi-index of the element in a sparse literal at the given - // sparse element number. The sparse element number is the position with in - // the sparse array's list of (index, value) pairs, and is checked against the - // total number of (index, value) pairs in the sparse array. - tensorflow::gtl::ArraySlice GetSparseIndex( - int64 sparse_element_number, const ShapeIndex& shape_index = {}) const; - - // Returns the value of the element in a sparse literal at the given sparse - // element number. The sparse element number is the position with in the - // sparse array's list of (index, value) pairs, and is checked against the - // total number of (index, value) pairs in the sparse array. - template - NativeT GetSparseElement(int64 sparse_element_number, - const ShapeIndex& shape_index = {}) const; - - // Invokes the "per cell" callback for each element in the provided - // literal with the element's indices and a string representation of - // the element's value. - // - // This function is useful if you want a polymorphic representation - // of the tensor's elements (turning it to a string for something - // like representation in a protobuf). - // - // This literal must have a dense layout. - void EachCellAsString( - const std::function indices, - const string& value)>& per_cell) const; - template - void EachCell(std::function indices, - NativeT value)> - per_cell) const; - - // Returns whether every element in this literal is equal to value. - // - // value is an int8 because we expect this to be called with small - // compile-time constants (0, -1, etc.) and so that whatever value you pass - // can be represented exactly by floating-point types as small as 16 bits. - // - // If value doesn't fit in this literal's type, returns false. Values of 1/0 - // are considered equal to true/false; other values are not considered equal - // to true. Also if this literal is not array-shaped false is returned. - bool IsAll(int8 value) const; - - // Like IsAll(const Literal&, int8), except we check whether the literal is - // equal to a particular floating-point number. - // - // If the literal is not a floating-point value, this always returns false. - // - // This casts value to the type of literal, then compares using ==. The usual - // admonishments about floating-point equality checks apply. We expect you to - // use this to check for values that can be expressed precisely as a float, - // e.g. -0.5. Also if this literal is not array-shaped false is returned. - bool IsAllFloat(float value) const; - - // Like IsAll(const Literal&, int8), except we check whether the literal is - // equal to a particular complex number. - // - // If the literal is not a complex value, this always returns false. - // - // This casts value to the type of literal, then compares using ==. The usual - // admonishments about floating-point equality checks apply. We expect you to - // use this to check for complex values that can be expressed precisely as - // float pairs e.g. (-0.5, 1.0). - // - // This literal must have a dense layout. - bool IsAllComplex(complex64 value) const; - - // Literal consists entirely of the first element of the literal. - bool IsAllFirst() const; - - // Returns whether this literal is zero at the specified index. This literal - // must be an array with a dense layout. - bool IsZero(tensorflow::gtl::ArraySlice indices) const; - - // Returns the count of the elements in the array at the given shape index in - // this literal. - int64 element_count(const ShapeIndex& index = {}) const { - return ShapeUtil::ElementsIn(ShapeUtil::GetSubshape(shape(), index)); - } - - // Returns the count of the elements in the sparse array at the given shape - // index in this literal, which will be no larger than - // LayoutUtil::MaxSparseElements(SetSubshape(shape(), index).layout()). - int64 sparse_element_count() const; - - // Compute a hash for this literal. This literal must not be a sparse tensor - // or a tuple containing a sparse tensor. - size_t Hash() const; - - // Converts this literal to the given shape. Returns an error is the - // conversion is not possible. - // - // round_f32_to_bf16: if true, converting F32 elements to BF16 uses rounding - // instead of truncation; otherwise, truncation is used. - // - // TODO(b/69266521): remove the round_to_bfloat16 flag when rounding becomes - // the default behavior. - StatusOr> ConvertToShape( - const Shape& dest_shape, bool round_f32_to_bf16 = false) const; - - // Converts this literal to another primitive type using a bitcast - // conversion. The to and from primitive types must have the same bit - // width. Returns an error if the conversion is not possible. This literal - // must be array-shaped. - StatusOr> BitcastConvert( - PrimitiveType primitive_dest_type) const; - - // Converts this literal to another primitive type. Returns an error if the - // conversion is not possible. This literal must be array-shaped. - StatusOr> Convert( - PrimitiveType primitive_dest_type) const; + LiteralUtil() = delete; // Returns a literal scalar representing the first element. - Literal GetFirstScalarLiteral() const; - - // Clones the underlying buffers into a new Literal, or new - // std::unique_ptr. - Literal Clone() const; - std::unique_ptr CloneToUnique() const; - - // TODO(b/67651157): The methods below which perform computation on Literals - // (Reshape, Slice, etc) should be moved elsewhere, and perhaps combined with - // evaluator code which operates on Literals. - // - // Creates a new value that has the equivalent value as this - // literal, but conforms to new_layout; e.g. a literal matrix that was in {0, - // 1} minor-to-major dimension layout can be re-layed-out as {1, 0} - // minor-to-major dimension layout and the value in the cell at any given - // logical index (i0, i1) will be the same. - // - // For tuple shaped literals, shape_index should be used to select the inner - // array that the new layout applies to. - // - // Note: this is useful when the client wants to ensure that a value placed in - // the XLA allocation tracker has a particular layout; for efficiency - // purposes or avoiding unimplemented operation/layout combinations. - std::unique_ptr Relayout(const Layout& new_layout, - const ShapeIndex& shape_index = {}) const; - - // An overload of Relayout which changes the layout of the entire shape rather - // than being limited to a single array within the shape. - std::unique_ptr Relayout(const Shape& shape_with_layout) const; - - // Creates a new literal by reshaping this literal to have the given - // dimensions. The total number of elements must not change; The - // implementation currently only supports monotonic dim0-major layouts. - // This literal must be an array. - StatusOr> Reshape( - tensorflow::gtl::ArraySlice dimensions) const; - - // Creates a new literal by broadcasting this literal with `dimensions` to - // yield a literal of shape `result_shape`. - StatusOr> Broadcast( - const Shape& result_shape, - tensorflow::gtl::ArraySlice dimensions) const; - - // Creates a new literal by reordering the dimensions of this literal. - // The given `permutation` must be a permutation of the dimension numbers - // in the original literal, and it specifies the order of the new dimensions - // in the result literal (i.e., new_order[i] = old_order[permutation[i]]). - // For example, a transpose call on a literal of shape [3 x 8 x 4] and - // `permutation` = {2, 0, 1} returns a new literal of shape [4 x 3 x 8]. - // This literal must be an array. - std::unique_ptr Transpose( - tensorflow::gtl::ArraySlice permutation) const; - - // Creates a sub-array from this literal by extracting the indices - // [start_index, limit_index) of each dimension. The result literal has the - // same rank and layout as for the given literal. The number of indices in - // start_indices and limit_indices must be the rank of the literal, and the - // indices follow the order of the dimensions. - // This literal must be an array. - std::unique_ptr Slice( - tensorflow::gtl::ArraySlice start_indices, - tensorflow::gtl::ArraySlice limit_indices) const; - - // Creates a literal with a prepended dimension with bound "times"; e.g. a - // f32[3x2] with times=4 will produce a f32[4x3x2] with the 3x2 from this - // literal replicated four times. - // This literal must be an array. - template - std::unique_ptr Replicate(int64 times) const; - - // Creates a new Literal object with the shape specified as parameter. - // The content of the literal values is the default value of the primitive - // type of literal itself (0 for numeric types, and false for predicates). - // - // Note: It's an antipattern to use this method then immediately call - // Literal::Populate on the result (since that results in zero initialization, - // then reinitialization. Conside if a call to MakeUnique(shape), - // followed by the call to Literal::Populate can be used instead. - static std::unique_ptr CreateFromShape(const Shape& shape); - - protected: - // A data structure representing a subshape at a particular ShapeIndex within - // the literal. For array-shaped ShapeIndexes, this data structure holds the - // pointer to the memory allocated for the array data. - class Piece { - public: - // Returns the buffer holding the array data for this piece as an array - // slice. This piece must be array-shaped. - template - tensorflow::gtl::ArraySlice data() const; - template - tensorflow::gtl::MutableArraySlice data(); - - // Returns the buffer holding the array data for this piece as a void*. This - // piece must be array-shaped. - void* untyped_data(); - const void* untyped_data() const; - - // Gets or sets an element in the array at the given index. The multi_index - // is CHECKed against the dimension sizes of the array. This piece must be - // array-shaped. - template - NativeT Get(tensorflow::gtl::ArraySlice index) const; - template - void Set(tensorflow::gtl::ArraySlice index, NativeT value); - - // Gets/sets the buffer holding the array data. - char* buffer() const { return buffer_; } - void set_buffer(char* buffer) { buffer_ = buffer; } - - // The array of multi-indices that provide the locations of non-zero - // elements in a sparse array. Only used if - // LayoutUtil::IsSparseArray(shape()) is true. - SparseIndexArray* sparse_indices() const { return sparse_indices_; } - void set_sparse_indices(SparseIndexArray* sparse_indices) { - sparse_indices_ = sparse_indices; - } - - // Gets or sets the subshape of this piece. This reference points to a - // subshape within the shape in the containing Literal (Literal::shape_). - const Shape& subshape() const { return *subshape_; } - void set_subshape(const Shape* subshape) { subshape_ = subshape; } - - // Returns the size in bytes of the buffer holding the array data. - int64 size_bytes() const { return ShapeUtil::ByteSizeOf(subshape()); } - - // Returns the number of elements in this piece's array. - int64 element_count() const { - // If this is a sparse array, use the number of elements represented by - // the indices in the associated SparseIndexArray. - return LayoutUtil::IsSparseArray(subshape()) - ? sparse_indices()->index_count() - : ShapeUtil::ElementsIn(subshape()); - } - - // Returns the child piece at 'index' of this piece. - Piece& child(int64 index) { return children_[index]; } - - // Adds a child piece to this piece's children. - void emplace_back(Piece child_piece) { - children_.emplace_back(std::move(child_piece)); - } - - // Returns the size of children pieces of this piece. - int64 children_size() { return children_.size(); } - - // Visitor functions that recursively traverses the piece and calls the - // given function at each child piece. The function has the type: - // void (const ShapeIndex& index, const Piece& piece) - template - void ForEachSubpiece(const Fn& func) const { - ShapeIndex index; - return ForEachHelper( - [&func](const ShapeIndex& index, const Piece& piece) { - func(index, piece); - return Status::OK(); - }, - *this, &index) - .IgnoreError(); - } - // Same as above, but the function has the type: - // Status (const ShapeIndex& index, const Piece& piece) - // The first non-OK return value is returned by the function. - template - Status ForEachSubpieceWithStatus(const Fn& func) const { - ShapeIndex index; - return ForEachHelper(func, *this, &index); - } - // Same as above, but the function has the type: - // Bool (const ShapeIndex& index, const Piece& piece) - // The first non-true return value is returned by the function. - template - bool ForEachSubpieceWithBool(const Fn& func) const { - ShapeIndex index; - return ForEachHelperBool(func, *this, &index); - } - // Same as above, but the function has the type: - // Void (const ShapeIndex& index, Piece& piece) - template - void ForEachMutableSubpiece(const Fn& func) { - ShapeIndex index; - return ForEachMutableHelper( - [&func](const ShapeIndex& index, Piece* piece) { - func(index, piece); - return Status::OK(); - }, - const_cast(this), &index) - .IgnoreError(); - } - // Same as above, but the function has the type: - // Status (const ShapeIndex& index, Piece& piece) - // The first non-OK return value is returned by the function. - template - Status ForEachMutableSubpieceWithStatus(const Fn& func) { - ShapeIndex index; - return ForEachMutableHelper( - func, const_cast(this), &index); - } - - // Returns true if this piece and 'other' contain the same data. This piece - // and 'other' must be array-shaped and compatible. - bool EqualElements(const Piece& other) const; - - // Writes the shape and data (if array-shaped) into the given proto. - void WriteToProto(LiteralProto* proto) const; - - // Copy the data from 'src' into this piece's buffer. Shapes of this piece - // and src must be compatible. - Status CopyFrom(const Piece& src); - - // Copies the data from the given proto into this piece. The shape of this - // piece must be equal (not just compatible) to the shape of the proto. - Status CopyFromProto(const LiteralProto& proto); - - // Sorts the elements in a sparse array. - void SortSparseElements(); - - private: - // Helpers for traversing the piece via ForEachSubpiece rooted at 'index'. - // The first non-OK (or non-true) value is returned by the function. - // The callable 'func' has the same signature as described above in - // ForEachSubpiece*. - template - Status ForEachHelper(const Fn& func, const Piece& piece, - ShapeIndex* index) const { - TF_RETURN_IF_ERROR(func(*index, piece)); - for (int64 i = 0; i < piece.children_.size(); ++i) { - index->push_back(i); - TF_RETURN_IF_ERROR(ForEachHelper(func, piece.children_[i], index)); - index->pop_back(); - } - return Status::OK(); - } - template - bool ForEachHelperBool(const Fn& func, const Piece& piece, - ShapeIndex* index) const { - if (!func(*index, piece)) { - return false; - } - for (int64 i = 0; i < piece.children_.size(); ++i) { - index->push_back(i); - if (!ForEachHelperBool(func, piece.children_[i], index)) { - return false; - } - index->pop_back(); - } - return true; - } - template - Status ForEachMutableHelper(const Fn& func, Piece* piece, - ShapeIndex* index) { - TF_RETURN_IF_ERROR(func(*index, piece)); - for (int64 i = 0; i < piece->children_.size(); ++i) { - index->push_back(i); - TF_RETURN_IF_ERROR( - ForEachMutableHelper(func, &piece->children_[i], index)); - index->pop_back(); - } - return Status::OK(); - } - - // Recursive helper for EqualElements. - template - bool EqualElementsInternal(const Piece& other, - std::vector* multi_index) const; - - // Helper for SortSparseElements that has the element type as a template - // parameter. - template - void SortSparseElementsInternal(); - - // For array-shaped pieces, this is the buffer holding the literal data. - char* buffer_ = nullptr; - - // For sparse arrays, this is the array of indices. - SparseIndexArray* sparse_indices_ = nullptr; - - // The shape of piece. This points into the shape of the containing Literal - // (Literal::shape_). - const Shape* subshape_ = nullptr; - - // Children pieces for tuple shaped pieces. - std::vector children_ = {}; - }; // class Piece - - const Piece& piece(const ShapeIndex& shape_index) const { - Piece* piece = &const_cast(root_piece()); - for (const auto i : shape_index) { - DCHECK_GE(i, 0); - DCHECK_LT(i, piece->children_size()); - piece = &piece->child(i); - } - return *piece; - } - - // Returns the piece at the root of the shape. - virtual const Piece& root_piece() const = 0; - - // LiteralSlice and Literal must access Pieces of other Literals. - friend class Literal; - friend class LiteralSlice; - friend class BorrowingLiteral; - - private: - template - std::unique_ptr SliceInternal( - const Shape& result_shape, - tensorflow::gtl::ArraySlice start_indices) const; -}; - -// Class representing literal values in XLA. -// -// The underlying buffer and shape is always owned by this class. -class Literal : public LiteralBase { - public: - Literal() : Literal(ShapeUtil::MakeNil()) {} - - // Create a literal of the given shape. The literal is allocated sufficient - // memory to hold the shape. Memory is uninitialized. - explicit Literal(const Shape& shape); - virtual ~Literal(); - - // Literals are moveable, but not copyable. To copy a literal use - // Literal::Clone or Literal::CloneToUnique. This prevents inadvertent copies - // of literals which can be expensive. - Literal(const Literal& other) = delete; - Literal& operator=(const Literal& other) = delete; - Literal(Literal&& other); - // 'allocate_arrays' indicates whether to allocate memory for the arrays in - // the shape. If false, buffer pointers inside of the Literal::Pieces are set - // to nullptr. - Literal(const Shape& shape, bool allocate_arrays); - Literal& operator=(Literal&& other); - - // TODO(b/67651157): Remove this accessor. Literal users should not be able to - // mutate the shape as this can produce malformed Literals. - Shape* mutable_shape_do_not_use() { return shape_.get(); } - - // Returns a MutableArraySlice view of the array for this literal for the - // given NativeT (e.g., float). CHECKs if the subshape of the literal at the - // given ShapeIndex is not array. See primitive_util.h for the mapping from - // XLA type to native type. - template - tensorflow::gtl::MutableArraySlice data( - const ShapeIndex& shape_index = {}); - // Unhide const method from parent class. - using LiteralBase::data; - - // Returns a pointer to the sparse index array. Returns nullptr if the literal - // is not a sparse array. - SparseIndexArray* sparse_indices(const ShapeIndex& shape_index = {}); - - // Returns a pointer to the underlying buffer holding the array at the given - // shape index. CHECKs if the subshape of the literal at the given ShapeIndex - // is not array. - void* untyped_data(const ShapeIndex& shape_index = {}); - // Unhide const method from parent class. - using LiteralBase::untyped_data; - - // Populates a literal with a sparse layout with the given indices and values. - // Each index in the indices array is CHECKed against the dimensions in the - // literal's shape. If sort is true, then the indices and values will be - // sorted. If sort is false, then the indices and values are assumed to - // already be in sorted order. See CreateSparse for an example of how data - // are populated. - template - void PopulateSparse(SparseIndexArray indices, - tensorflow::gtl::ArraySlice values, - bool sort = true); - - // Copy values from 'src_literal' rooted at 'src_shape_index' into this - // literal rooted at 'dest_shape_index'. The subshape of this literal rooted - // at 'dest_shape_index' must be compatible with the subshape of 'src_literal' - // rooted at 'src_shape_index', but need not be arrays. - Status CopyFrom(const LiteralSlice& src_literal, - const ShapeIndex& dest_shape_index = {}, - const ShapeIndex& src_shape_index = {}); - - // Similar to CopyFrom, but with move semantincs. The subshape of this literal - // rooted at 'dest_shape_index' must be *equal* to the shape 'src_literal' - // (layouts and shapes must match), but need not be arrays. The memory - // allocated in this literal for the subshape at dest_shape_index is - // deallocated, and the respective buffers are replaced with those in - // src_literal. Upon return, src_literal is set to a nil shape (empty tuple). - Status MoveFrom(Literal&& src_literal, - const ShapeIndex& dest_shape_index = {}); - - // Copies the values from src_literal, starting at src_base shape indexes, - // to this literal, starting at dest_base, where the copy size in each - // dimension is specified by copy_size. - // The src_literal and this literal must have the same primitive type, - // src_base+copy_size must fit the source literal dimensions, as well as - // dest_base+copy_size must fit the destination literal dimensions. - // Note: if either src_literal or this literal contains dimensions with zero - // element, then copy_size must be 0 in these dimensions while the - // corresponding base indices being 0. - // This literal and 'src_literal' must be arrays. - Status CopySliceFrom(const LiteralSlice& src_literal, - tensorflow::gtl::ArraySlice src_base, - tensorflow::gtl::ArraySlice dest_base, - tensorflow::gtl::ArraySlice copy_size); - - // Copies one element from src_literal[src_index] to (*this)[dest_index]. - Status CopyElementFrom(const LiteralSlice& src_literal, - tensorflow::gtl::ArraySlice src_index, - tensorflow::gtl::ArraySlice dest_index); - - // Sets an element in the literal at the given index. The multi_index is - // CHECKed against the dimension sizes. - template - void Set(tensorflow::gtl::ArraySlice multi_index, - const ShapeIndex& shape_index, NativeT value); - // Overloads of Set for array literals. CHECKs if the literal is not - // array-shaped and dense. - template - void Set(tensorflow::gtl::ArraySlice multi_index, NativeT value); - - // Appends the given element to the literal. If the elements are not appended - // in sorted order, then SortSparseElements should be called before calling - // other methods. This literal must have a sparse layout. - template - void AppendSparseElement(tensorflow::gtl::ArraySlice multi_index, - NativeT value, const ShapeIndex& shape_index = {}); - - // Sorts the elements in a sparse array. - void SortSparseElements(const ShapeIndex& shape_index = {}); - - // As Set(), but truncates `value` to the literal element type before storing. - // This literal must be an array. - Status SetIntegralAsS64(tensorflow::gtl::ArraySlice multi_index, - int64 value); - - // Populate this literal with the given values. Examples: - // - // // Populate with floats. - // Array2D float_values = ... - // literal.PopulateR2FromArray2D(values); - // - // // Populate with int32s. - // literal.PopulateR2({{1, 2}, {3, 4}}); - // - // The shape and element type of this literal must match given values. For - // example, in the call above to literal.PopulateR2(), 'literal' must be a 2x2 - // array of S32. - template - void PopulateR1(tensorflow::gtl::ArraySlice values); - void PopulateR1(const tensorflow::core::Bitmap& values); - template - void PopulateR2(std::initializer_list> values); - template - void PopulateFromArray(const Array& values); - template - void PopulateR2FromArray2D(const Array2D& values); - template - void PopulateR3FromArray3D(const Array3D& values); - template - void PopulateR4FromArray4D(const Array4D& values); - - // Populates literal values by calling the generator function for every cell - // in this literal object. - // - // generator must be a callable of the type - // NativeT(tensorflow::gtl::ArraySlice indexes) or compatible. - // - // This literal must have a dense layout. - template - Status Populate(const FnType& generator); - - // A parallel version of Populate(). This can be used if the generator is - // thread-safe and the values for the shape's different elements are - // independent. - template - Status PopulateParallel(const FnType& generator); - - // Fills this literal with the given value. - template - void PopulateWithValue(NativeT value); - - // Factory methods below. - // - - // Serialize from a proto. - static StatusOr> CreateFromProto( - const LiteralProto& proto); + static Literal GetFirstScalarLiteral(const LiteralSlice& literal); // Creates a new literal of a given rank. To minimize ambiguity (for users // and the compiler) these CreateR[0-2] methods should explicitly specify the @@ -889,7 +223,7 @@ class Literal : public LiteralBase { // As above, but intended to be invoked with move semantics; i.e. // // std::vector> elements = ...; - // auto result = Literal::MakeTupleOwned(std::move(elements)); + // auto result = LiteralUtil::MakeTupleOwned(std::move(elements)); // // This would have been declared as an overload, but there is ambiguity // in invocation between the above signature and this one. @@ -899,7 +233,7 @@ class Literal : public LiteralBase { // This overload lets you pass a braced list of unique_ptrs to // MakeTupleOwned: // - // Literal::MakeTupleOwned(Literal::CreateR1(...), ...). + // LiteralUtil::MakeTupleOwned(LiteralUtil::CreateR1(...), ...). // // Simply relying on the MakeTupleOwned(std::vector>) // overload doesn't work because std::initializer_list's elements are always @@ -920,19 +254,6 @@ class Literal : public LiteralBase { // Create a constant token literal. Token types have no value. static std::unique_ptr CreateToken(); - // Returns a vector containing the tuple elements of this Literal as separate - // Literals. This Literal must be tuple-shaped and can be a nested tuple. The - // elements are moved into the new Literals; no data is copied. Upon return - // this Literal is set to a nil shape (empty tuple) - std::vector DecomposeTuple(); - - // This operation is the inverse of DecomposeTuple. The given elements are - // moved into the tuple elements of a new tuple-shaped Literal which is - // returned. Upon return, each of the Literals in 'elements' is set to a nil - // shape (empty tuple). - static Literal MoveIntoTuple( - tensorflow::gtl::MutableArraySlice elements); - // Creates a new Literal object with its values havings the primitive_type // type, and with dimensions defined by the dimensions parameter. // The content of the literal values is the default value of the primitive @@ -1000,194 +321,12 @@ class Literal : public LiteralBase { // dimension 1 equal to 8. static string MultiIndexAsString( tensorflow::gtl::ArraySlice multi_index); - - private: - // Recursively sets the subshapes and buffers of all subpieces rooted at - // 'piece'. If 'allocate_array' is true, memory is allocated for the arrays in - // the shape. - void SetPiece(const Shape& shape, Piece* piece, bool allocate_arrays); - - // Returns the piece at the given ShapeIndex. - Piece& piece(const ShapeIndex& shape_index) { - return const_cast(LiteralBase::piece(shape_index)); - } - - Piece& root_piece() const override { return *root_piece_; }; - - // Internal template helper for the Literal::CopySliceFrom(), matching its - // arguments one by one. - template - Status CopySliceFromInternal(const LiteralBase& src_literal, - tensorflow::gtl::ArraySlice src_base, - tensorflow::gtl::ArraySlice dest_base, - tensorflow::gtl::ArraySlice copy_size); - - // Utility structure which is used to create the optimal configuration for - // a ShapeUtil::ForEachIndex() scan across two literals. - struct StrideConfig { - StrideConfig(const Shape& source_shape, const Shape& dest_shape, - tensorflow::gtl::ArraySlice dimensions); - - // The dimensions of the stride operation. Essentially every dimension - // will be iterated from base[i] to base[i]+dimensions[i], in step[i] - // steps. - tensorflow::gtl::ArraySlice dimensions; - DimensionVector base; - DimensionVector step; - int64 minor_dimension = 0; - // The size of the strides for source and destination. One of the two - // (the one looping through its most minor dimension) will be 1, while - // the other will be the stride size at the dimension matching the other - // shape most minor dimension being scanned. - int64 dest_stride = 1; - int64 source_stride = 1; - // The size of the inner loop on the most minor dimension. - int64 minor_loop_size = 1; - }; - - // Literal class always owns the shape. The parent class borrows this shape. - std::unique_ptr shape_; - - Piece* root_piece_ = nullptr; - - // Implementation details shared between Populate() and PopulateParallel() - template - Status PopulateInternal(const FnType& generator, bool parallel); - - // Deallocate the buffers held by this literal. - void DeallocateBuffers(); - - friend class LiteralBase; -}; -std::ostream& operator<<(std::ostream& out, const Literal& literal); - -// A read-only view of a Literal. A LiteralSlice contains pointers to shape and -// literal buffers always owned by others. -class LiteralSlice : public LiteralBase { - public: - LiteralSlice() : LiteralBase() {} - - // Implicit conversion constructors. - LiteralSlice(const LiteralBase& literal); - LiteralSlice(const LiteralBase& literal, const ShapeIndex& view_root); - - private: - const Piece& root_piece() const override { return *root_piece_; }; - - const Piece* root_piece_; // Not owned. -}; - -// A read-only Literal where the underlying buffers are never owned by this -// class. -class BorrowingLiteral : public LiteralBase { - public: - BorrowingLiteral() : LiteralBase() {} - - // 'src_buf_ptr' is not owned by this class and must outlive the - // lifetime of this class. It points to an appropirately sized buffer with - // data interpretered as indicated by 'shape'. - // This constructor is only used for array shapes. - BorrowingLiteral(const char* src_buf_ptr, const Shape& shape); - // Similar as above, except to be used for constructing non-nested tuples. - BorrowingLiteral(tensorflow::gtl::ArraySlice src_buf_ptrs, - const Shape& shape); - // TODO(b/79707221): adding constructors for nested tuples as well. - - private: - // Recursively builds the subtree for the given piece and sets the subshapes - // of the given piece with the given shape. - void BuildPieceSubtree(const Shape& shape, Piece* piece); - - // Accessor for the root piece of this literal. - const Piece& root_piece() const override { return root_piece_; }; - Piece root_piece_; - - // Shape of this literal. Stored as unique_ptr so such that the (default) - // move construction of this class would be trivially correct: the pointer to - // Shape root_piece_ stores will still point to the correct address. - std::unique_ptr shape_; }; -template -tensorflow::gtl::ArraySlice LiteralBase::Piece::data() const { - CHECK(ShapeUtil::IsArray(subshape())) << ShapeUtil::HumanString(subshape()); - CHECK_EQ(subshape().element_type(), - primitive_util::NativeToPrimitiveType()) - << "Attempting to access " - << PrimitiveType_Name(primitive_util::NativeToPrimitiveType()) - << " type, but literal element type is " - << PrimitiveType_Name(subshape().element_type()); - return tensorflow::gtl::ArraySlice( - reinterpret_cast(buffer()), element_count()); -} - -template -tensorflow::gtl::MutableArraySlice LiteralBase::Piece::data() { - CHECK(ShapeUtil::IsArray(subshape())) << ShapeUtil::HumanString(subshape()); - CHECK_EQ(subshape().element_type(), - primitive_util::NativeToPrimitiveType()) - << "Attempting to access " - << PrimitiveType_Name(primitive_util::NativeToPrimitiveType()) - << " type, but literal element type is " - << PrimitiveType_Name(subshape().element_type()); - return tensorflow::gtl::MutableArraySlice( - reinterpret_cast(buffer()), element_count()); -} - -template -NativeT LiteralBase::Piece::Get( - tensorflow::gtl::ArraySlice multi_index) const { - CHECK(LayoutUtil::IsDenseArray(subshape())); - return data()[IndexUtil::MultidimensionalIndexToLinearIndex( - subshape(), multi_index)]; -} - -template -void LiteralBase::Piece::Set(tensorflow::gtl::ArraySlice multi_index, - NativeT value) { - CHECK(LayoutUtil::IsDenseArray(subshape())); - data()[IndexUtil::MultidimensionalIndexToLinearIndex( - subshape(), multi_index)] = value; -} - -template -tensorflow::gtl::ArraySlice LiteralBase::data( - const ShapeIndex& shape_index) const { - return piece(shape_index).data(); -} - -template -tensorflow::gtl::MutableArraySlice Literal::data( - const ShapeIndex& shape_index) { - return piece(shape_index).data(); -} - -template -inline NativeT LiteralBase::Get(tensorflow::gtl::ArraySlice multi_index, - const ShapeIndex& shape_index) const { - return piece(shape_index).Get(multi_index); -} - -template -inline NativeT LiteralBase::Get( - tensorflow::gtl::ArraySlice multi_index) const { - return root_piece().Get(multi_index); -} - -template -inline void Literal::Set(tensorflow::gtl::ArraySlice multi_index, - const ShapeIndex& shape_index, NativeT value) { - return piece(shape_index).Set(multi_index, value); -} - -template -inline void Literal::Set(tensorflow::gtl::ArraySlice multi_index, - NativeT value) { - return root_piece().Set(multi_index, value); -} +std::ostream& operator<<(std::ostream& out, const Literal& literal); template -/* static */ std::unique_ptr Literal::CreateR0(NativeT value) { +/* static */ std::unique_ptr LiteralUtil::CreateR0(NativeT value) { auto literal = MakeUnique(ShapeUtil::MakeShape( primitive_util::NativeToPrimitiveType(), {})); literal->Set({}, value); @@ -1195,7 +334,7 @@ template } template -/* static */ std::unique_ptr Literal::CreateR1( +/* static */ std::unique_ptr LiteralUtil::CreateR1( tensorflow::gtl::ArraySlice values) { auto literal = MakeUnique( ShapeUtil::MakeShape(primitive_util::NativeToPrimitiveType(), @@ -1205,7 +344,7 @@ template } template -/* static */ std::unique_ptr Literal::CreateR2WithLayout( +/* static */ std::unique_ptr LiteralUtil::CreateR2WithLayout( std::initializer_list> values, const Layout& layout) { auto literal = MakeUnique(ShapeUtil::MakeShapeWithLayout( @@ -1218,13 +357,13 @@ template } template -/* static */ std::unique_ptr Literal::CreateR2( +/* static */ std::unique_ptr LiteralUtil::CreateR2( std::initializer_list> values) { return CreateR2WithLayout(values, LayoutUtil::GetDefaultLayoutForR2()); } template -/* static */ std::unique_ptr Literal::CreateR3WithLayout( +/* static */ std::unique_ptr LiteralUtil::CreateR3WithLayout( std::initializer_list>> values, const Layout& layout) { @@ -1249,14 +388,14 @@ template } template -/* static */ std::unique_ptr Literal::CreateR3( +/* static */ std::unique_ptr LiteralUtil::CreateR3( std::initializer_list>> values) { return CreateR3WithLayout(values, LayoutUtil::GetDefaultLayoutForR3()); } template -/* static */ std::unique_ptr Literal::CreateR4WithLayout( +/* static */ std::unique_ptr LiteralUtil::CreateR4WithLayout( std::initializer_list>>> values, @@ -1287,7 +426,7 @@ template } template -/* static */ std::unique_ptr Literal::CreateSparse( +/* static */ std::unique_ptr LiteralUtil::CreateSparse( tensorflow::gtl::ArraySlice dimensions, SparseIndexArray indices, tensorflow::gtl::ArraySlice values, bool sort) { int64 num_elements = values.size(); @@ -1302,7 +441,7 @@ template } template -/* static */ std::unique_ptr Literal::CreateR4( +/* static */ std::unique_ptr LiteralUtil::CreateR4( std::initializer_list>>> values) { @@ -1310,7 +449,7 @@ template } template -/* static */ std::unique_ptr Literal::CreateFromArrayWithLayout( +/* static */ std::unique_ptr LiteralUtil::CreateFromArrayWithLayout( const Array& values, const Layout& layout) { auto literal = MakeUnique(ShapeUtil::MakeShapeWithLayout( primitive_util::NativeToPrimitiveType(), values.dimensions(), @@ -1320,38 +459,40 @@ template } template -/* static */ std::unique_ptr Literal::CreateFromArray( +/* static */ std::unique_ptr LiteralUtil::CreateFromArray( const Array& values) { return CreateFromArrayWithLayout( values, LayoutUtil::GetDefaultLayoutForRank(values.num_dimensions())); } template -/* static */ std::unique_ptr Literal::CreateR2FromArray2DWithLayout( - const Array2D& values, const Layout& layout) { +/* static */ std::unique_ptr +LiteralUtil::CreateR2FromArray2DWithLayout(const Array2D& values, + const Layout& layout) { return CreateFromArrayWithLayout(values, layout); } template -/* static */ std::unique_ptr Literal::CreateR2FromArray2D( +/* static */ std::unique_ptr LiteralUtil::CreateR2FromArray2D( const Array2D& values) { return CreateFromArray(values); } template -/* static */ std::unique_ptr Literal::CreateR3FromArray3DWithLayout( - const Array3D& values, const Layout& layout) { +/* static */ std::unique_ptr +LiteralUtil::CreateR3FromArray3DWithLayout(const Array3D& values, + const Layout& layout) { return CreateFromArrayWithLayout(values, layout); } template -/* static */ std::unique_ptr Literal::CreateR3FromArray3D( +/* static */ std::unique_ptr LiteralUtil::CreateR3FromArray3D( const Array3D& values) { return CreateFromArray(values); } template -/* static */ std::unique_ptr Literal::CreateR3Projected( +/* static */ std::unique_ptr LiteralUtil::CreateR3Projected( std::initializer_list> values, int64 projection) { int64 dim0_size = projection; @@ -1376,7 +517,7 @@ template } template -/* static */ std::unique_ptr Literal::CreateR4Projected( +/* static */ std::unique_ptr LiteralUtil::CreateR4Projected( std::initializer_list> values, int64 projection_p, int64 projection_z) { int64 dim0_size = projection_p; @@ -1404,49 +545,21 @@ template } template -/* static */ std::unique_ptr Literal::CreateR4FromArray4D( +/* static */ std::unique_ptr LiteralUtil::CreateR4FromArray4D( const Array4D& values) { return CreateFromArray(values); } template -/* static */ std::unique_ptr Literal::CreateR4FromArray4DWithLayout( - const Array4D& values, const Layout& layout) { +/* static */ std::unique_ptr +LiteralUtil::CreateR4FromArray4DWithLayout(const Array4D& values, + const Layout& layout) { return CreateFromArrayWithLayout(values, layout); } -template -NativeT LiteralBase::GetFirstElement() const { - return data().at(0); -} - -template -NativeT LiteralBase::GetSparseElement(int64 sparse_element_number, - const ShapeIndex& shape_index) const { - CHECK( - LayoutUtil::IsSparseArray(ShapeUtil::GetSubshape(shape(), shape_index))); - return data(shape_index)[sparse_element_number]; -} - -template -void Literal::AppendSparseElement( - tensorflow::gtl::ArraySlice multi_index, NativeT value, - const ShapeIndex& shape_index) { - Piece& p = piece(shape_index); - const Shape& subshape = p.subshape(); - CHECK(LayoutUtil::IsSparseArray(subshape)); - int64 rank = ShapeUtil::Rank(subshape); - CHECK_EQ(multi_index.size(), rank); - int64 last_element = p.sparse_indices()->index_count(); - CHECK_LT(last_element, LayoutUtil::MaxSparseElements(subshape.layout())); - p.sparse_indices()->Append(multi_index); - CHECK_LT(last_element, p.data().size()); - p.data()[last_element] = value; -} - // Returns an identity matrix (rank 2) with the given row and column count. template -/* static */ std::unique_ptr Literal::MakeIdentityR2(int64 size) { +/* static */ std::unique_ptr LiteralUtil::MakeIdentityR2(int64 size) { Array2D array(size, size, 0); for (int64 i = 0; i < size; ++i) { array(i, i) = 1; @@ -1455,174 +568,8 @@ template } template -void LiteralBase::EachCell( - std::function indices, - NativeT value)> - per_cell) const { - if (ShapeUtil::IsZeroElementArray(shape())) { - return; - } - std::vector indices(ShapeUtil::Rank(shape()), 0); - do { - per_cell(indices, Get(indices)); - } while (IndexUtil::BumpIndices(shape(), &indices)); -} - -template -inline void Literal::PopulateR1(tensorflow::gtl::ArraySlice values) { - CHECK(ShapeUtil::IsArray(shape())); - CHECK_EQ(ShapeUtil::Rank(shape()), 1); - CHECK_EQ(ShapeUtil::ElementsIn(shape()), values.size()); - CHECK_EQ(shape().element_type(), - primitive_util::NativeToPrimitiveType()); - for (int64 i = 0; i < values.size(); ++i) { - Set({i}, values[i]); - } -} - -template -void Literal::PopulateR2( - std::initializer_list> values) { - CHECK(ShapeUtil::IsArray(shape())); - CHECK_EQ(ShapeUtil::Rank(shape()), 2); - CHECK_EQ(shape().element_type(), - primitive_util::NativeToPrimitiveType()); - - const int64 dim0_size = values.size(); - const int64 dim1_size = values.begin()->size(); - CHECK_EQ(dim0_size, shape().dimensions(0)); - CHECK_EQ(dim1_size, shape().dimensions(1)); - - int64 dim0 = 0; - for (auto inner_list : values) { - int64 dim1 = 0; - for (auto value : inner_list) { - Set({dim0, dim1}, value); - ++dim1; - } - CHECK_EQ(dim1_size, dim1); - ++dim0; - } -} - -template -void Literal::PopulateFromArray(const Array& values) { - CHECK(ShapeUtil::IsArray(shape())); - CHECK_EQ(shape().element_type(), - primitive_util::NativeToPrimitiveType()); - CHECK_EQ(ShapeUtil::Rank(shape()), values.num_dimensions()); - for (int dim = 0; dim < values.num_dimensions(); ++dim) { - CHECK_EQ(values.dim(dim), shape().dimensions(dim)); - } - values.Each([this](tensorflow::gtl::ArraySlice indices, - NativeT value) { this->Set(indices, value); }); -} - -template -void Literal::PopulateR2FromArray2D(const Array2D& values) { - PopulateFromArray(values); -} - -template -void Literal::PopulateR3FromArray3D(const Array3D& values) { - PopulateFromArray(values); -} - -template -void Literal::PopulateR4FromArray4D(const Array4D& values) { - PopulateFromArray(values); -} - -template -void Literal::PopulateSparse(SparseIndexArray indices, - tensorflow::gtl::ArraySlice values, - bool sort) { - CHECK(LayoutUtil::IsSparseArray(shape())); - int rank = ShapeUtil::Rank(shape()); - CHECK_EQ(indices.rank(), rank); - int64 max_elements = LayoutUtil::MaxSparseElements(shape().layout()); - CHECK_LE(indices.max_indices(), max_elements); - int64 num_elements = values.size(); - CHECK_LE(num_elements, max_elements); - CHECK_EQ(num_elements, indices.index_count()); - auto root_data = root_piece().data(); - // Piece::data() returns an ArraySlice of size equal to the number of indices - // in the SparseIndexArray. So there is no need to adjust the size of the data - // here. It is enough to just copy the incoming values into the data buffer. - std::copy(values.begin(), values.end(), root_data.begin()); - *this->root_piece().sparse_indices() = std::move(indices); - if (sort) { - auto root_data = this->root_piece().data(); - this->root_piece().sparse_indices()->SortWithValues(root_data); - } - DCHECK(this->root_piece().sparse_indices()->Validate(shape())); -} - -template -Status Literal::PopulateInternal(const FnType& generator, bool parallel) { - const Shape& this_shape = shape(); - const int64 rank = ShapeUtil::Rank(this_shape); - TF_RET_CHECK(LayoutUtil::IsDenseArray(this_shape)); - TF_RET_CHECK(this_shape.element_type() == - primitive_util::NativeToPrimitiveType()); - tensorflow::gtl::MutableArraySlice literal_data = data(); - if (rank > 0) { - StrideConfig stride_config(this_shape, this_shape, - AsInt64Slice(this_shape.dimensions())); - int64 minor_dimension_size = - ShapeUtil::GetDimension(this_shape, stride_config.minor_dimension); - - auto init_function = [&](tensorflow::gtl::ArraySlice indexes) { - DimensionVector minor_scan_indexes(rank, 0); - const int64 index = - IndexUtil::MultidimensionalIndexToLinearIndex(shape(), indexes); - std::copy(indexes.begin(), indexes.end(), minor_scan_indexes.begin()); - for (int64 i = 0; i < minor_dimension_size; ++i) { - minor_scan_indexes[stride_config.minor_dimension] = i; - literal_data.at(index + i) = generator(minor_scan_indexes); - } - }; - if (parallel) { - ShapeUtil::ForEachIndexParallel(this_shape, stride_config.base, - stride_config.dimensions, - stride_config.step, init_function); - } else { - ShapeUtil::ForEachIndex( - this_shape, stride_config.base, stride_config.dimensions, - stride_config.step, - [&init_function](tensorflow::gtl::ArraySlice indexes) { - init_function(indexes); - return true; - }); - } - } else { - // For scalars. - literal_data.at(0) = generator({}); - } - return Status::OK(); -} -template -Status Literal::Populate(const FnType& generator) { - return PopulateInternal(generator, /*parallel=*/false); -} - -template -Status Literal::PopulateParallel(const FnType& generator) { - return PopulateInternal(generator, /*parallel=*/true); -} - -template -void Literal::PopulateWithValue(NativeT value) { - CHECK(ShapeUtil::IsArray(shape())); - CHECK_EQ(shape().element_type(), - primitive_util::NativeToPrimitiveType()); - for (NativeT& element : data()) { - element = value; - } -} - -template -/* static */ std::unique_ptr Literal::CreateFullWithDescendingLayout( +/* static */ std::unique_ptr +LiteralUtil::CreateFullWithDescendingLayout( tensorflow::gtl::ArraySlice dimensions, NativeT value) { auto literal = MakeUnique(ShapeUtil::MakeShapeWithDescendingLayout( primitive_util::NativeToPrimitiveType(), dimensions)); @@ -1630,44 +577,9 @@ template return literal; } -template -std::unique_ptr LiteralBase::Replicate(int64 times) const { - DimensionVector bounds = {times}; - bounds.reserve(shape().dimensions_size() + 1); - for (int64 bound : shape().dimensions()) { - bounds.push_back(bound); - } - auto literal = - MakeUnique(ShapeUtil::MakeShape(shape().element_type(), bounds)); - int64 elements = ShapeUtil::ElementsIn(literal->shape()); - if (elements == 0) { - return literal; - } - - DimensionVector output_indices(bounds.size(), 0); - tensorflow::gtl::ArraySlice input_indices = output_indices; - input_indices.remove_prefix(1); - - bool done = false; - while (!done) { - const auto element = Get(input_indices); - literal->Set(output_indices, element); - - done = true; - for (int n = 0; n < output_indices.size(); ++n) { - ++output_indices[n]; - if (output_indices[n] < bounds[n]) { - done = false; - break; - } - output_indices[n] = 0; - } - } - return literal; -} - template -/* static */ StatusOr> Literal::CreateRandomLiteral( +/* static */ StatusOr> +LiteralUtil::CreateRandomLiteral( const Shape& shape, const std::function)>& generator) { using NativeT = typename primitive_util::PrimitiveTypeToNative::type; @@ -1681,8 +593,9 @@ template } template -/* static */ StatusOr> Literal::CreateRandomLiteral( - const Shape& shape, E* engine, T mean, T stddev) { +/* static */ StatusOr> +LiteralUtil::CreateRandomLiteral(const Shape& shape, E* engine, T mean, + T stddev) { using NativeT = typename primitive_util::PrimitiveTypeToNative::type; std::normal_distribution generator(mean, stddev); return CreateRandomLiteral( @@ -1692,8 +605,8 @@ template } template -/* static */ StatusOr> Literal::CreateRandomLiteral( - const Shape& shape, T mean, T stddev) { +/* static */ StatusOr> +LiteralUtil::CreateRandomLiteral(const Shape& shape, T mean, T stddev) { std::minstd_rand0 engine; return CreateRandomLiteral(shape, &engine, mean, stddev); } diff --git a/tensorflow/compiler/xla/overflow_util.h b/tensorflow/compiler/xla/overflow_util.h new file mode 100644 index 0000000000000000000000000000000000000000..8657d3a4bfa992b9ca0619f24923fd4542eed894 --- /dev/null +++ b/tensorflow/compiler/xla/overflow_util.h @@ -0,0 +1,50 @@ +/* Copyright 2015 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#ifndef TENSORFLOW_COMPILER_XLA_OVERFLOW_UTIL_H_ +#define TENSORFLOW_COMPILER_XLA_OVERFLOW_UTIL_H_ + +#include "tensorflow/core/platform/logging.h" +#include "tensorflow/core/platform/macros.h" +#include "tensorflow/core/platform/types.h" + +namespace xla { + +// Multiply two nonnegative int64's, returning negative for overflow +inline int64 MultiplyWithoutOverflow(const int64 x, const int64 y) { + // Multiply in uint64 rather than int64 since signed overflow is undefined. + // Negative values will wrap around to large unsigned values in the casts + // (see section 4.7 [conv.integral] of the C++14 standard). + const uint64 ux = x; + const uint64 uy = y; + const uint64 uxy = ux * uy; + + // Check if we overflow uint64, using a cheap check if both inputs are small + if (TF_PREDICT_FALSE((ux | uy) >> 32 != 0)) { + // Ensure nonnegativity. Note that negative numbers will appear "large" + // to the unsigned comparisons above. + CHECK(x >= 0 && y >= 0); + + // Otherwise, detect overflow using a division + if (ux != 0 && uxy / ux != uy) return -1; + } + + // Cast back to signed. Any negative value will signal an error. + return static_cast(uxy); +} + +} // namespace xla + +#endif // TENSORFLOW_COMPILER_XLA_OVERFLOW_UTIL_H_ diff --git a/tensorflow/compiler/xla/packed_literal_reader.cc b/tensorflow/compiler/xla/packed_literal_reader.cc index 857aae0a7982a57bb3057a6f267f5f033a0fdde4..6b7fd10d63f8f97b0e0bf7570488c06323368d75 100644 --- a/tensorflow/compiler/xla/packed_literal_reader.cc +++ b/tensorflow/compiler/xla/packed_literal_reader.cc @@ -20,7 +20,7 @@ limitations under the License. #include #include "tensorflow/compiler/xla/layout_util.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/ptr_util.h" #include "tensorflow/compiler/xla/shape_util.h" #include "tensorflow/compiler/xla/status_macros.h" diff --git a/tensorflow/compiler/xla/packed_literal_reader.h b/tensorflow/compiler/xla/packed_literal_reader.h index 45a9fe012784d3e4168e7549240dec962aa1a17a..98dccaa9a246520bf60217b96d67a13a24c34b4a 100644 --- a/tensorflow/compiler/xla/packed_literal_reader.h +++ b/tensorflow/compiler/xla/packed_literal_reader.h @@ -18,7 +18,7 @@ limitations under the License. #include -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/statusor.h" #include "tensorflow/compiler/xla/types.h" #include "tensorflow/compiler/xla/xla_data.pb.h" diff --git a/tensorflow/compiler/xla/python/BUILD b/tensorflow/compiler/xla/python/BUILD index 83834c1ff65ea2f9989fe08279c29056d9070adb..fe346f9956adaed8d0127e92517b2cd32be05105 100644 --- a/tensorflow/compiler/xla/python/BUILD +++ b/tensorflow/compiler/xla/python/BUILD @@ -33,6 +33,7 @@ cc_library( srcs = ["numpy_bridge.cc"], hdrs = ["numpy_bridge.h"], deps = [ + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:literal_util", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:xla_data_proto", @@ -52,9 +53,9 @@ cc_library( "//tensorflow/compiler/xla/client:client_library", "//tensorflow/compiler/xla/client:executable_build_options", "//tensorflow/compiler/xla/client:local_client", + "//tensorflow/compiler/xla/client/lib:math", "//tensorflow/compiler/xla/client/xla_client:xla_builder", "//tensorflow/compiler/xla/client/xla_client:xla_computation", - "//tensorflow/compiler/xla/service:hlo_proto", "//tensorflow/compiler/xla/service:shaped_buffer", "//tensorflow/core:framework_lite", "//tensorflow/core:lib", @@ -70,7 +71,7 @@ tf_py_wrap_cc( deps = [ ":local_computation_builder", ":numpy_bridge", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:xla_data_proto", "//tensorflow/compiler/xla/service:cpu_plugin", diff --git a/tensorflow/compiler/xla/python/local_computation_builder.cc b/tensorflow/compiler/xla/python/local_computation_builder.cc index 29062348b0afd0f17bc24cef71f6d3929b131212..be55d50b234442ec569c85e4f5224ad1c179bca8 100644 --- a/tensorflow/compiler/xla/python/local_computation_builder.cc +++ b/tensorflow/compiler/xla/python/local_computation_builder.cc @@ -14,6 +14,8 @@ limitations under the License. ==============================================================================*/ #include "tensorflow/compiler/xla/python/local_computation_builder.h" +#include "tensorflow/compiler/xla/client/lib/math.h" +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/compiler/xla/executable_run_options.h" #include "tensorflow/compiler/xla/ptr_util.h" #include "tensorflow/compiler/xla/util.h" @@ -174,73 +176,73 @@ StatusOr> CompiledLocalComputation::Execute( GetReplicaCount()); for (int replica = 0; replica < GetReplicaCount(); ++replica) { - pool.Schedule([this, client, replica, &arguments, &shapes_with_layout, - &results] { - StatusOr device_ordinal_status = - client->ReplicaNumberToDeviceOrdinal(replica); - if (!device_ordinal_status.ok()) { - results[replica] = device_ordinal_status.status(); - return; - } - const int device_ordinal = device_ordinal_status.ValueOrDie(); - VLOG(3) << "Replica " << replica - << " mapped to device ordinal for execution: " - << device_ordinal; - - // Transfer arguments in - std::vector scoped_buffers; - scoped_buffers.reserve(arguments.size()); - for (int i = 0; i < arguments.size(); ++i) { - const Literal& argument = arguments[i]; - const tensorflow::gtl::optional& shape_with_layout = - shapes_with_layout[i]; - - StatusOr pushed; - if (shape_with_layout) { - std::unique_ptr relaid = - argument.Relayout(shape_with_layout.value()); - pushed = ToBuffer(client, device_ordinal, *relaid); - } else { - pushed = ToBuffer(client, device_ordinal, argument); - } - if (!pushed.ok()) { - results[replica] = pushed.status(); - return; - } - - scoped_buffers.push_back(std::move(pushed).ValueOrDie()); - } - - // Execute - std::vector argument_buffers; - argument_buffers.reserve(scoped_buffers.size()); - for (auto& buffer : scoped_buffers) { - argument_buffers.push_back(&buffer); - } - - DeviceAssignment device_assignment = - client->backend() - .computation_placer() - ->AssignDevices(GetReplicaCount(), /*computation_count=*/1) - .ConsumeValueOrDie(); - - ExecutableRunOptions options; - options.set_device_ordinal(device_ordinal); - options.set_allocator(client->backend().memory_allocator()); - options.set_intra_op_thread_pool( - client->backend().eigen_intra_op_thread_pool_device()); - options.set_device_assignment(&device_assignment); - StatusOr result_buffer_status = - executable_->Run(argument_buffers, options); - if (!result_buffer_status.ok()) { - results[replica] = result_buffer_status.status(); - return; - } - - // Transfer result out - results[replica] = client->ShapedBufferToLiteral( - std::move(result_buffer_status).ValueOrDie()); - }); + pool.Schedule( + [this, client, replica, &arguments, &shapes_with_layout, &results] { + StatusOr device_ordinal_status = + client->ReplicaNumberToDeviceOrdinal(replica); + if (!device_ordinal_status.ok()) { + results[replica] = device_ordinal_status.status(); + return; + } + const int device_ordinal = device_ordinal_status.ValueOrDie(); + VLOG(3) << "Replica " << replica + << " mapped to device ordinal for execution: " + << device_ordinal; + + // Transfer arguments in + std::vector scoped_buffers; + scoped_buffers.reserve(arguments.size()); + for (int i = 0; i < arguments.size(); ++i) { + const Literal& argument = arguments[i]; + const tensorflow::gtl::optional& shape_with_layout = + shapes_with_layout[i]; + + StatusOr pushed; + if (shape_with_layout) { + std::unique_ptr relaid = + argument.Relayout(shape_with_layout.value()); + pushed = ToBuffer(client, device_ordinal, *relaid); + } else { + pushed = ToBuffer(client, device_ordinal, argument); + } + if (!pushed.ok()) { + results[replica] = pushed.status(); + return; + } + + scoped_buffers.push_back(std::move(pushed).ValueOrDie()); + } + + // Execute + std::vector argument_buffers; + argument_buffers.reserve(scoped_buffers.size()); + for (auto& buffer : scoped_buffers) { + argument_buffers.push_back(&buffer); + } + + DeviceAssignment device_assignment = + client->backend() + .computation_placer() + ->AssignDevices(GetReplicaCount(), /*computation_count=*/1) + .ConsumeValueOrDie(); + + ExecutableRunOptions options; + options.set_device_ordinal(device_ordinal); + options.set_allocator(client->backend().memory_allocator()); + options.set_intra_op_thread_pool( + client->backend().eigen_intra_op_thread_pool_device()); + options.set_device_assignment(&device_assignment); + StatusOr result_buffer_status = + executable_->Run(argument_buffers, options); + if (!result_buffer_status.ok()) { + results[replica] = result_buffer_status.status(); + return; + } + + // Transfer result out + results[replica] = client->ShapedBufferToLiteral( + std::move(result_buffer_status).ValueOrDie()); + }); } } @@ -341,7 +343,7 @@ StatusOr LocalComputationBuilder::Build() { LocalOp LocalComputationBuilder::Parameter(int64 parameter_number, const Shape& shape, const string& name) { - return builder_.Parameter(parameter_number, shape, name); + return xla::Parameter(&builder_, parameter_number, shape, name); } StatusOr LocalComputationBuilder::GetShape(const LocalOp& operand) { @@ -354,72 +356,70 @@ StatusOr LocalComputationBuilder::GetReturnValueShape() { } LocalOp LocalComputationBuilder::Infeed(const Shape& shape) { - return builder_.Infeed(shape); + return xla::Infeed(&builder_, shape); } void LocalComputationBuilder::Outfeed(const LocalOp& operand, const Shape& shape, const string& outfeed_config) { - builder_.Outfeed(operand.op(), shape, outfeed_config); + xla::Outfeed(operand.op(), shape, outfeed_config); } LocalOp LocalComputationBuilder::ConstantLiteral(const Literal& literal) { - return builder_.ConstantLiteral(literal); + return xla::ConstantLiteral(&builder_, literal); } LocalOp LocalComputationBuilder::Broadcast( const LocalOp& operand, tensorflow::gtl::ArraySlice broadcast_sizes) { - return builder_.Broadcast(operand.op(), broadcast_sizes); + return xla::Broadcast(operand.op(), broadcast_sizes); } LocalOp LocalComputationBuilder::Pad(const LocalOp& operand, const LocalOp& padding_value, const PaddingConfig& padding_config) { - return builder_.Pad(operand.op(), padding_value.op(), padding_config); + return xla::Pad(operand.op(), padding_value.op(), padding_config); } LocalOp LocalComputationBuilder::Reshape( const LocalOp& operand, tensorflow::gtl::ArraySlice dimensions, tensorflow::gtl::ArraySlice new_sizes) { - return builder_.Reshape(operand.op(), dimensions, new_sizes); + return xla::Reshape(operand.op(), dimensions, new_sizes); } LocalOp LocalComputationBuilder::Collapse( const LocalOp& operand, tensorflow::gtl::ArraySlice dimensions) { - return builder_.Collapse(operand.op(), dimensions); + return xla::Collapse(operand.op(), dimensions); } LocalOp LocalComputationBuilder::CrossReplicaSum(const LocalOp& operand) { - return builder_.CrossReplicaSum(operand.op()); + return xla::CrossReplicaSum(operand.op()); } LocalOp LocalComputationBuilder::Slice( const LocalOp& operand, tensorflow::gtl::ArraySlice start_indices, tensorflow::gtl::ArraySlice limit_indices, tensorflow::gtl::ArraySlice strides) { - return builder_.Slice(operand.op(), start_indices, limit_indices, strides); + return xla::Slice(operand.op(), start_indices, limit_indices, strides); } LocalOp LocalComputationBuilder::SliceInDim(const LocalOp& operand, int64 start_index, int64 limit_index, int64 stride, int64 dimno) { - return builder_.SliceInDim(operand.op(), start_index, limit_index, stride, - dimno); + return xla::SliceInDim(operand.op(), start_index, limit_index, stride, dimno); } LocalOp LocalComputationBuilder::DynamicSlice( const LocalOp& operand, const LocalOp& start_indices, tensorflow::gtl::ArraySlice slice_sizes) { - return builder_.DynamicSlice(operand.op(), start_indices.op(), slice_sizes); + return xla::DynamicSlice(operand.op(), start_indices.op(), slice_sizes); } LocalOp LocalComputationBuilder::DynamicUpdateSlice( const LocalOp& operand, const LocalOp& update, const LocalOp& start_indices) { - return builder_.DynamicUpdateSlice(operand.op(), update.op(), - start_indices.op()); + return xla::DynamicUpdateSlice(operand.op(), update.op(), start_indices.op()); } LocalOp LocalComputationBuilder::ConcatInDim( @@ -429,7 +429,7 @@ LocalOp LocalComputationBuilder::ConcatInDim( for (const auto& op : operands) { xla_ops.push_back(op.op()); } - return builder_.ConcatInDim(xla_ops, dimension); + return xla::ConcatInDim(&builder_, xla_ops, dimension); } LocalOp LocalComputationBuilder::SelectAndScatterWithGeneralPadding( @@ -439,7 +439,7 @@ LocalOp LocalComputationBuilder::SelectAndScatterWithGeneralPadding( tensorflow::gtl::ArraySlice> padding, const LocalOp& source, const LocalOp& init_value, const LocalComputation& scatter) { - return builder_.SelectAndScatterWithGeneralPadding( + return xla::SelectAndScatterWithGeneralPadding( operand.op(), select.computation(), window_dimensions, window_strides, padding, source.op(), init_value.op(), scatter.computation()); } @@ -452,22 +452,22 @@ LocalOp LocalComputationBuilder::Tuple( xla_ops.push_back(op.op()); } - return builder_.Tuple(xla_ops); + return xla::Tuple(&builder_, xla_ops); } LocalOp LocalComputationBuilder::GetTupleElement(const LocalOp& tuple_data, int64 index) { - return builder_.GetTupleElement(tuple_data.op(), index); + return xla::GetTupleElement(tuple_data.op(), index); } LocalOp LocalComputationBuilder::Dot(const LocalOp& lhs, const LocalOp& rhs) { - return builder_.Dot(lhs.op(), rhs.op()); + return xla::Dot(lhs.op(), rhs.op()); } LocalOp LocalComputationBuilder::DotGeneral( const LocalOp& lhs, const LocalOp& rhs, const DotDimensionNumbers& dimension_numbers) { - return builder_.DotGeneral(lhs.op(), rhs.op(), dimension_numbers); + return xla::DotGeneral(lhs.op(), rhs.op(), dimension_numbers); } LocalOp LocalComputationBuilder::ConvGeneralDilated( @@ -477,14 +477,13 @@ LocalOp LocalComputationBuilder::ConvGeneralDilated( tensorflow::gtl::ArraySlice lhs_dilation, tensorflow::gtl::ArraySlice rhs_dilation, const ConvolutionDimensionNumbers& dimension_numbers) { - return builder_.ConvGeneralDilated(lhs.op(), rhs.op(), window_strides, - padding, lhs_dilation, rhs_dilation, - dimension_numbers); + return xla::ConvGeneralDilated(lhs.op(), rhs.op(), window_strides, padding, + lhs_dilation, rhs_dilation, dimension_numbers); } LocalOp LocalComputationBuilder::ConvertElementType( const LocalOp& operand, PrimitiveType new_element_type) { - return builder_.ConvertElementType(operand.op(), new_element_type); + return xla::ConvertElementType(operand.op(), new_element_type); } LocalOp LocalComputationBuilder::Call( @@ -495,46 +494,39 @@ LocalOp LocalComputationBuilder::Call( for (const auto& op : operands) { xla_ops.push_back(op.op()); } - return builder_.Call(local_computation.computation(), xla_ops); + return xla::Call(&builder_, local_computation.computation(), xla_ops); } LocalOp LocalComputationBuilder::Transpose( const LocalOp& operand, tensorflow::gtl::ArraySlice permutation) { - return builder_.Transpose(operand.op(), permutation); + return xla::Transpose(operand.op(), permutation); } LocalOp LocalComputationBuilder::Rev( const LocalOp& operand, tensorflow::gtl::ArraySlice dimensions) { - return builder_.Rev(operand.op(), dimensions); + return xla::Rev(operand.op(), dimensions); } LocalOp LocalComputationBuilder::Map( tensorflow::gtl::ArraySlice operands, const LocalComputation& local_computation, - tensorflow::gtl::ArraySlice dimensions, - tensorflow::gtl::ArraySlice static_operands) { + tensorflow::gtl::ArraySlice dimensions) { std::vector xla_ops; xla_ops.reserve(operands.size()); for (const auto& op : operands) { xla_ops.push_back(op.op()); } - std::vector static_xla_ops; - static_xla_ops.reserve(static_operands.size()); - for (const auto& op : static_operands) { - static_xla_ops.push_back(op.op()); - } - - return builder_.Map(xla_ops, local_computation.computation(), dimensions, - static_xla_ops); + return xla::Map(&builder_, xla_ops, local_computation.computation(), + dimensions); } LocalOp LocalComputationBuilder::Reduce( const LocalOp& operand, const LocalOp& init_value, const LocalComputation& local_computation, tensorflow::gtl::ArraySlice dimensions_to_reduce) { - return builder_.Reduce(operand.op(), init_value.op(), - local_computation.computation(), dimensions_to_reduce); + return xla::Reduce(operand.op(), init_value.op(), + local_computation.computation(), dimensions_to_reduce); } LocalOp LocalComputationBuilder::ReduceWindowWithGeneralPadding( @@ -543,7 +535,7 @@ LocalOp LocalComputationBuilder::ReduceWindowWithGeneralPadding( tensorflow::gtl::ArraySlice window_dimensions, tensorflow::gtl::ArraySlice window_strides, tensorflow::gtl::ArraySlice> padding) { - return builder_.ReduceWindowWithGeneralPadding( + return xla::ReduceWindowWithGeneralPadding( operand.op(), init_value.op(), local_computation.computation(), window_dimensions, window_strides, padding); } @@ -551,27 +543,27 @@ LocalOp LocalComputationBuilder::ReduceWindowWithGeneralPadding( LocalOp LocalComputationBuilder::RngNormal(const LocalOp& mu, const LocalOp& sigma, const Shape& shape) { - return builder_.RngNormal(mu.op(), sigma.op(), shape); + return xla::RngNormal(mu.op(), sigma.op(), shape); } LocalOp LocalComputationBuilder::RngUniform(const LocalOp& a, const LocalOp& b, const Shape& shape) { - return builder_.RngUniform(a.op(), b.op(), shape); + return xla::RngUniform(a.op(), b.op(), shape); } LocalOp LocalComputationBuilder::While(const LocalComputation& condition, const LocalComputation& body, const LocalOp& init) { - return builder_.While(condition.computation(), body.computation(), init.op()); + return xla::While(condition.computation(), body.computation(), init.op()); } LocalOp LocalComputationBuilder::Conditional( const LocalOp& predicate, const LocalOp& true_operand, const LocalComputation& true_computation, const LocalOp& false_operand, const LocalComputation& false_computation) { - return builder_.Conditional( - predicate.op(), true_operand.op(), true_computation.computation(), - false_operand.op(), false_computation.computation()); + return xla::Conditional(predicate.op(), true_operand.op(), + true_computation.computation(), false_operand.op(), + false_computation.computation()); } StatusOr LocalComputationBuilder::IsConstant(const LocalOp& operand) { @@ -587,7 +579,7 @@ StatusOr LocalComputationBuilder::BuildConstantSubGraph( #define _FORWARD(method_name, return_sig, args_sig, args) \ return_sig LocalComputationBuilder::method_name args_sig { \ - return builder_.method_name args; \ + return xla::method_name args; \ } #define _FORWARD_UNOP(method_name) \ @@ -621,6 +613,7 @@ _FORWARD_BINOP(Max) _FORWARD_BINOP(Min) _FORWARD_BINOP(And) _FORWARD_BINOP(Or) +_FORWARD_BINOP(Xor) _FORWARD_UNOP(Not) _FORWARD_UNOP(Abs) _FORWARD_UNOP(Exp) @@ -634,11 +627,11 @@ _FORWARD_UNOP(Sign) _FORWARD_UNOP(Cos) _FORWARD_UNOP(Sin) _FORWARD_UNOP(Tanh) -_FORWARD_UNOP(SqrtF32) -_FORWARD_UNOP(SquareF32) +_FORWARD_UNOP(Sqrt) +_FORWARD_UNOP(Square) _FORWARD_BINOP(Pow) _FORWARD_UNOP(IsFinite) -_FORWARD_UNOP(ReciprocalF32) +_FORWARD_UNOP(Reciprocal) _FORWARD_UNOP(Neg) _FORWARD_UNOP(Sort) diff --git a/tensorflow/compiler/xla/python/local_computation_builder.h b/tensorflow/compiler/xla/python/local_computation_builder.h index 95f0a0610b573479e0103ba2d1514844df35c2b4..690ff277e884c6f1540b12e7002248571d07fe71 100644 --- a/tensorflow/compiler/xla/python/local_computation_builder.h +++ b/tensorflow/compiler/xla/python/local_computation_builder.h @@ -270,8 +270,7 @@ class LocalComputationBuilder { LocalOp Map(tensorflow::gtl::ArraySlice operands, const LocalComputation& local_computation, - tensorflow::gtl::ArraySlice dimensions, - tensorflow::gtl::ArraySlice static_operands); + tensorflow::gtl::ArraySlice dimensions); LocalOp Reduce(const LocalOp& operand, const LocalOp& init_value, const LocalComputation& local_computation, @@ -333,6 +332,7 @@ class LocalComputationBuilder { _FORWARD_BINOP(Min) _FORWARD_BINOP(And) _FORWARD_BINOP(Or) + _FORWARD_BINOP(Xor) _FORWARD_UNOP(Not) _FORWARD_UNOP(Abs) _FORWARD_UNOP(Exp) @@ -346,11 +346,11 @@ class LocalComputationBuilder { _FORWARD_UNOP(Cos) _FORWARD_UNOP(Sin) _FORWARD_UNOP(Tanh) - _FORWARD_UNOP(SqrtF32) - _FORWARD_UNOP(SquareF32) + _FORWARD_UNOP(Sqrt) + _FORWARD_UNOP(Square) _FORWARD_BINOP(Pow) _FORWARD_UNOP(IsFinite) - _FORWARD_UNOP(ReciprocalF32) + _FORWARD_UNOP(Reciprocal) _FORWARD_UNOP(Neg) _FORWARD_UNOP(Sort) diff --git a/tensorflow/compiler/xla/python/local_computation_builder.i b/tensorflow/compiler/xla/python/local_computation_builder.i index 477df6fde25d0db760e08df9d335bd12e31ccb55..afdea88cb7d14769aa1bd73a309d54689f3bf2a9 100644 --- a/tensorflow/compiler/xla/python/local_computation_builder.i +++ b/tensorflow/compiler/xla/python/local_computation_builder.i @@ -109,7 +109,7 @@ limitations under the License. // Must be included first #include "tensorflow/python/lib/core/numpy.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/shape_util.h" #include "tensorflow/compiler/xla/xla_data.pb.h" #include "tensorflow/core/lib/gtl/array_slice.h" @@ -988,6 +988,7 @@ tensorflow::ImportNumpy(); %unignore xla::swig::LocalComputationBuilder::Min; %unignore xla::swig::LocalComputationBuilder::And; %unignore xla::swig::LocalComputationBuilder::Or; +%unignore xla::swig::LocalComputationBuilder::Xor; %unignore xla::swig::LocalComputationBuilder::Not; %unignore xla::swig::LocalComputationBuilder::Abs; %unignore xla::swig::LocalComputationBuilder::Exp; @@ -1001,11 +1002,11 @@ tensorflow::ImportNumpy(); %unignore xla::swig::LocalComputationBuilder::Cos; %unignore xla::swig::LocalComputationBuilder::Sin; %unignore xla::swig::LocalComputationBuilder::Tanh; -%unignore xla::swig::LocalComputationBuilder::SqrtF32; -%unignore xla::swig::LocalComputationBuilder::SquareF32; +%unignore xla::swig::LocalComputationBuilder::Sqrt; +%unignore xla::swig::LocalComputationBuilder::Square; %unignore xla::swig::LocalComputationBuilder::Pow; %unignore xla::swig::LocalComputationBuilder::IsFinite; -%unignore xla::swig::LocalComputationBuilder::ReciprocalF32; +%unignore xla::swig::LocalComputationBuilder::Reciprocal; %unignore xla::swig::LocalComputationBuilder::Neg; %unignore xla::swig::LocalComputationBuilder::Sort; %unignore xla::swig::DestructureLocalShapedBufferTuple; diff --git a/tensorflow/compiler/xla/python/numpy_bridge.cc b/tensorflow/compiler/xla/python/numpy_bridge.cc index 68648a3a176363de69a56ecb8070f82862874e94..71351abd593d45fb5080112438a91df368eee173 100644 --- a/tensorflow/compiler/xla/python/numpy_bridge.cc +++ b/tensorflow/compiler/xla/python/numpy_bridge.cc @@ -14,6 +14,7 @@ limitations under the License. ==============================================================================*/ #include "tensorflow/compiler/xla/python/numpy_bridge.h" +#include "tensorflow/compiler/xla/literal_util.h" #include "tensorflow/compiler/xla/shape_util.h" #include "tensorflow/core/platform/logging.h" @@ -374,7 +375,7 @@ StatusOr> XlaLiteralFromPyObject(PyObject* o) { TF_ASSIGN_OR_RETURN(auto literal, XlaLiteralFromPyObject(element)); elements.push_back(std::move(literal)); } - return Literal::MakeTupleOwned(std::move(elements)); + return LiteralUtil::MakeTupleOwned(std::move(elements)); } else if (PyArray_Check(o)) { PyArrayObject* py_array = reinterpret_cast(o); int rank = PyArray_NDIM(py_array); @@ -383,7 +384,7 @@ StatusOr> XlaLiteralFromPyObject(PyObject* o) { dimensions[i] = PyArray_DIM(py_array, i); } int np_type = PyArray_TYPE(py_array); - auto literal = Literal::CreateFromDimensions( + auto literal = LiteralUtil::CreateFromDimensions( NumpyTypeToPrimitiveType(np_type), dimensions); TF_RETURN_IF_ERROR( CopyNumpyArrayToLiteral(np_type, py_array, literal.get())); diff --git a/tensorflow/compiler/xla/python/numpy_bridge.h b/tensorflow/compiler/xla/python/numpy_bridge.h index 64f0aae0f9790f0199ac6cb931a5c9f6dc356f4c..a67c93a4fb7413f9bbcb9afd92c36fd118836e1f 100644 --- a/tensorflow/compiler/xla/python/numpy_bridge.h +++ b/tensorflow/compiler/xla/python/numpy_bridge.h @@ -25,7 +25,7 @@ limitations under the License. #include #include -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/xla_data.pb.h" #include "tensorflow/core/lib/gtl/array_slice.h" #include "tensorflow/python/lib/core/numpy.h" diff --git a/tensorflow/compiler/xla/python/xla_client.py b/tensorflow/compiler/xla/python/xla_client.py index a1fc25303c257174f12e070301f504710aa0c61e..e2b6eaa0961c0f97a7451760deea8865db3ed272 100644 --- a/tensorflow/compiler/xla/python/xla_client.py +++ b/tensorflow/compiler/xla/python/xla_client.py @@ -99,10 +99,10 @@ _UNARY_OPS = [ 'Cos', 'Sin', 'Tanh', - 'SqrtF32', - 'SquareF32', + 'Sqrt', + 'Square', 'IsFinite', - 'ReciprocalF32', + 'Reciprocal', 'Neg', 'Sort', ] @@ -123,6 +123,7 @@ _BINARY_OPS = [ 'Min', 'And', 'Or', + 'Xor', 'Pow', ] @@ -460,14 +461,16 @@ class LocalComputation(object): if self.is_compiled: raise ValueError('Attempt to compile a compiled local XLA computation.') + result_shape = _wrap_shape(self.c_local_computation.GetReturnValueShape()) + if layout_fn: argument_shapes = [ shape.map_leaves(layout_fn) for shape in argument_shapes ] - result_shape = _wrap_shape(self.c_local_computation.GetReturnValueShape()) result_shape = result_shape.map_leaves(layout_fn) - compile_options = compile_options or CompileOptions() - compile_options.result_shape = result_shape + + compile_options = compile_options or CompileOptions() + compile_options.result_shape = result_shape return LocalComputation( self.c_local_computation.Compile(argument_shapes, compile_options), is_compiled=True) @@ -908,20 +911,19 @@ class ComputationBuilder(object): """ return self._client.Call(computation_to_apply.c_local_computation, operands) - def Map(self, operands, computation_to_apply, dimensions, static_operands=()): + def Map(self, operands, computation_to_apply, dimensions): """Enqueues a map operation onto the computation. Args: operands: an iterable of LocalOp. computation_to_apply: a Computation object. dimensions: dimensions over which to apply map the function. - static_operands: auxiliary arguments passed to the applied computation. Returns: A LocalOp representing the added Map op. """ return self._client.Map(operands, computation_to_apply.c_local_computation, - dimensions, static_operands) + dimensions) def Reduce(self, operand, init_value, computation_to_apply, dimensions): """Enqueues a reduction operation onto the computation. diff --git a/tensorflow/compiler/xla/python/xla_client_test.py b/tensorflow/compiler/xla/python/xla_client_test.py index 71e1d60a4e23dbfef333223c396e109533da9365..0564ddcb85ee3952f82649687e79a864999baf2c 100644 --- a/tensorflow/compiler/xla/python/xla_client_test.py +++ b/tensorflow/compiler/xla/python/xla_client_test.py @@ -157,6 +157,13 @@ class ComputationsWithConstantsTest(LocalComputationTest): c.Constant(NumpyArrayBool([True, True, False, False]))) self._ExecuteAndCompareExact(c, expected=[True, True, True, False]) + def testBooleanXor(self): + c = self._NewComputation() + c.Xor( + c.Constant(NumpyArrayBool([True, False, True, False])), + c.Constant(NumpyArrayBool([True, True, False, False]))) + self._ExecuteAndCompareExact(c, expected=[False, True, True, False]) + def testSum2DF32(self): c = self._NewComputation() c.Add( @@ -1168,14 +1175,6 @@ class EmbeddedComputationsTest(LocalComputationTest): self._CreateBinaryDivF64Computation(), [0]) self._ExecuteAndCompareClose(c, expected=[0.2, 0.4, 0.75, 1.0]) - def DISABLED_testMapWithStaticOperands(self): - c = self._NewComputation() - factor = c.ConstantF32Scalar(3.0) - c.Map([c.Constant(NumpyArrayF32([1.0, 2.0, 3.0, 4.0]))], - self._CreateMulF32ByParamComputation(), [0], - static_operands=[factor]) - self._ExecuteAndCompareClose(c, expected=[3.0, 6.0, 9.0, 12.0]) - def testSelectAndScatterF32(self): c = self._NewComputation() c.SelectAndScatter(c.Constant(NumpyArrayF32([[1., 2., 6.], [4., 5., 3.]])), diff --git a/tensorflow/compiler/xla/reference_util.cc b/tensorflow/compiler/xla/reference_util.cc index c289c84cff743871a7126cb932d6cda823ceb696..6397f1f47915aaa559beda467c26c66795c98f60 100644 --- a/tensorflow/compiler/xla/reference_util.cc +++ b/tensorflow/compiler/xla/reference_util.cc @@ -19,6 +19,7 @@ limitations under the License. #include #include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" +#include "tensorflow/compiler/xla/literal_util.h" #include "tensorflow/compiler/xla/service/cpu/runtime_single_threaded_matmul.h" #include "tensorflow/compiler/xla/service/hlo_evaluator.h" #include "tensorflow/compiler/xla/service/hlo_instruction.h" @@ -510,8 +511,8 @@ ReferenceUtil::ConvArray4DGeneralDimensionsDilated( std::pair lhs_dilation, std::pair rhs_dilation, ConvolutionDimensionNumbers dnums) { HloComputation::Builder b("ConvArray4DGeneralDimensionDilated"); - auto lhs_literal = Literal::CreateR4FromArray4D(lhs); - auto rhs_literal = Literal::CreateR4FromArray4D(rhs); + auto lhs_literal = LiteralUtil::CreateR4FromArray4D(lhs); + auto rhs_literal = LiteralUtil::CreateR4FromArray4D(rhs); std::array ordered_kernel_strides; std::array ordered_input_dimensions; diff --git a/tensorflow/compiler/xla/reference_util_test.cc b/tensorflow/compiler/xla/reference_util_test.cc index 9da9bc60a2025e63b57a3be9ed360d150f88d73c..8091bed4996a753649a5ecedda69a1ae48fb5897 100644 --- a/tensorflow/compiler/xla/reference_util_test.cc +++ b/tensorflow/compiler/xla/reference_util_test.cc @@ -22,7 +22,7 @@ limitations under the License. #include "tensorflow/compiler/xla/array3d.h" #include "tensorflow/compiler/xla/array4d.h" #include "tensorflow/compiler/xla/client/padding.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/ptr_util.h" #include "tensorflow/compiler/xla/test.h" #include "tensorflow/compiler/xla/tests/literal_test_util.h" @@ -53,7 +53,7 @@ class ReferenceUtilTest : public ::testing::Test { TEST_F(ReferenceUtilTest, TransposeArray2D) { auto result = ReferenceUtil::TransposeArray2D(*matrix_); - auto actual_literal = Literal::CreateR2FromArray2D(*result); + auto actual_literal = LiteralUtil::CreateR2FromArray2D(*result); LiteralTestUtil::ExpectR2Near({{1.f, 4.f}, {2.f, 5.f}, {3.f, 6.f}}, *actual_literal, ErrorSpec(0.0001)); } @@ -65,7 +65,7 @@ TEST_F(ReferenceUtilTest, MatmulArray2D) { {11.f, 12.f}, }); auto result = ReferenceUtil::MatmulArray2D(*matrix_, rhs); - auto actual_literal = Literal::CreateR2FromArray2D(*result); + auto actual_literal = LiteralUtil::CreateR2FromArray2D(*result); LiteralTestUtil::ExpectR2Near({{58.f, 64.f}, {139.f, 154.f}}, *actual_literal, ErrorSpec(0.0001)); } @@ -73,7 +73,7 @@ TEST_F(ReferenceUtilTest, MatmulArray2D) { TEST_F(ReferenceUtilTest, ReduceToColArray2D) { auto add = [](float lhs, float rhs) { return lhs + rhs; }; auto result = ReferenceUtil::ReduceToColArray2D(*matrix_, 0.0f, add); - auto actual_literal = Literal::CreateR1(*result); + auto actual_literal = LiteralUtil::CreateR1(*result); LiteralTestUtil::ExpectR1Near({6.f, 15.f}, *actual_literal, ErrorSpec(0.0001)); } @@ -81,13 +81,13 @@ TEST_F(ReferenceUtilTest, ReduceToColArray2D) { TEST_F(ReferenceUtilTest, ReduceToRowArray2D) { auto add = [](float lhs, float rhs) { return lhs + rhs; }; auto result = ReferenceUtil::ReduceToRowArray2D(*matrix_, 0.0f, add); - auto actual_literal = Literal::CreateR1(*result); + auto actual_literal = LiteralUtil::CreateR1(*result); LiteralTestUtil::ExpectR1Near({5.f, 7.f, 9.f}, *actual_literal, ErrorSpec(0.0001)); } TEST_F(ReferenceUtilTest, Reduce4Dto1DZeroSizedArray) { - auto result = Literal::CreateR1(ReferenceUtil::Reduce4DTo1D( + auto result = LiteralUtil::CreateR1(ReferenceUtil::Reduce4DTo1D( Array4D(1, 0, 1, 1), /*init=*/0, /*dims=*/{0, 1, 2}, [](float a, float b) { return a + b; })); LiteralTestUtil::ExpectR1Equal({0}, *result); @@ -96,7 +96,7 @@ TEST_F(ReferenceUtilTest, Reduce4Dto1DZeroSizedArray) { TEST_F(ReferenceUtilTest, MapArray2D) { auto identity = [](float value) { return log(exp(value)); }; auto result = ReferenceUtil::MapArray2D(*matrix_, identity); - auto actual_literal = Literal::CreateR2FromArray2D(*result); + auto actual_literal = LiteralUtil::CreateR2FromArray2D(*result); LiteralTestUtil::ExpectR2NearArray2D(*matrix_, *actual_literal, ErrorSpec(0.0001)); } @@ -106,7 +106,7 @@ TEST_F(ReferenceUtilTest, MapWithIndexArray2D) { return value + row + col; }; auto result = ReferenceUtil::MapWithIndexArray2D(*matrix_, add_index); - auto actual_literal = Literal::CreateR2FromArray2D(*result); + auto actual_literal = LiteralUtil::CreateR2FromArray2D(*result); LiteralTestUtil::ExpectR2Near({{1.f, 3.f, 5.f}, {5.f, 7.f, 9.f}}, *actual_literal, ErrorSpec(0.0001)); } @@ -117,7 +117,7 @@ TEST_F(ReferenceUtilTest, MapArray4D) { input->FillWithMultiples(1.0f); auto multiply_by_two = [](float value) { return 2 * value; }; auto result = ReferenceUtil::MapArray4D(*input, multiply_by_two); - auto actual_literal = Literal::CreateR4FromArray4D(*result); + auto actual_literal = LiteralUtil::CreateR4FromArray4D(*result); Array4D expected(/*planes=*/2, /*depth=*/3, /*height=*/4, /*width=*/5); expected.FillWithMultiples(2.0f); @@ -134,7 +134,7 @@ TEST_F(ReferenceUtilTest, MapWithIndexArray4D) { return value - (3 * 4 * 5 * plane + 4 * 5 * depth + 5 * height + width); }; auto result = ReferenceUtil::MapWithIndexArray4D(*input, subtract_index); - auto actual_literal = Literal::CreateR4FromArray4D(*result); + auto actual_literal = LiteralUtil::CreateR4FromArray4D(*result); Array4D expected(/*planes=*/2, /*depth=*/3, /*height=*/4, /*width=*/5); expected.Fill(0.0f); @@ -144,7 +144,7 @@ TEST_F(ReferenceUtilTest, MapWithIndexArray4D) { TEST_F(ReferenceUtilTest, SliceArray2D) { auto result = ReferenceUtil::Slice2D(*matrix_, {{0, 0}}, {{2, 2}}, {{1, 1}}); - auto actual_literal = Literal::CreateR2FromArray2D(*result); + auto actual_literal = LiteralUtil::CreateR2FromArray2D(*result); LiteralTestUtil::ExpectR2Near({{1.f, 2.f}, {4.f, 5.f}}, *actual_literal, ErrorSpec(0.0001)); @@ -152,7 +152,7 @@ TEST_F(ReferenceUtilTest, SliceArray2D) { TEST_F(ReferenceUtilTest, SliceStridedArray2D) { auto result = ReferenceUtil::Slice2D(*matrix_, {{0, 0}}, {{2, 3}}, {{1, 2}}); - auto actual_literal = Literal::CreateR2FromArray2D(*result); + auto actual_literal = LiteralUtil::CreateR2FromArray2D(*result); LiteralTestUtil::ExpectR2Near({{1.f, 3.f}, {4.f, 6.f}}, *actual_literal, ErrorSpec(0.0001)); @@ -164,7 +164,7 @@ TEST_F(ReferenceUtilTest, SliceArray3D) { auto result = ReferenceUtil::Slice3D(input, {{0, 0, 0}}, {{2, 2, 2}}, {{1, 1, 1}}); - auto actual_literal = Literal::CreateR3FromArray3D(*result); + auto actual_literal = LiteralUtil::CreateR3FromArray3D(*result); LiteralTestUtil::ExpectR3Near( {{{0.f, 1.f}, {4.f, 5.f}}, {{12.f, 13.f}, {16.f, 17.f}}}, *actual_literal, @@ -177,7 +177,7 @@ TEST_F(ReferenceUtilTest, SliceStridedArray3D) { auto result = ReferenceUtil::Slice3D(input, {{0, 0, 0}}, {{2, 3, 4}}, {{1, 2, 2}}); - auto actual_literal = Literal::CreateR3FromArray3D(*result); + auto actual_literal = LiteralUtil::CreateR3FromArray3D(*result); LiteralTestUtil::ExpectR3Near( {{{0.f, 2.f}, {8.f, 10.f}}, {{12.f, 14.f}, {20.f, 22.f}}}, @@ -190,7 +190,7 @@ TEST_F(ReferenceUtilTest, SliceArray4D) { auto result = ReferenceUtil::Slice4D(input, {{1, 0, 0, 0}}, {{2, 2, 2, 2}}, {{1, 1, 1, 1}}); - auto actual_literal = Literal::CreateR4FromArray4D(*result); + auto actual_literal = LiteralUtil::CreateR4FromArray4D(*result); LiteralTestUtil::ExpectR4Near( {{{{60.f, 61.f}, {65.f, 66.f}}, {{80.f, 81.f}, {85.f, 86.f}}}}, @@ -203,7 +203,7 @@ TEST_F(ReferenceUtilTest, SliceStridedArray4D) { auto result = ReferenceUtil::Slice4D(input, {{1, 0, 0, 0}}, {{2, 3, 4, 5}}, {{1, 2, 2, 2}}); - auto actual_literal = Literal::CreateR4FromArray4D(*result); + auto actual_literal = LiteralUtil::CreateR4FromArray4D(*result); LiteralTestUtil::ExpectR4Near( {{{{60.f, 62.f, 64.f}, {70.f, 72.f, 74.f}}, @@ -218,7 +218,7 @@ TEST_F(ReferenceUtilTest, ConvArray3DWithSamePadding) { ReferenceUtil::ConvArray3D(input, weights, 1, Padding::kSame); Array3D expected = {{{17, 28, 39, 20}}}; - auto actual_literal = Literal::CreateR3FromArray3D(*actual); + auto actual_literal = LiteralUtil::CreateR3FromArray3D(*actual); LiteralTestUtil::ExpectR3NearArray3D(expected, *actual_literal, ErrorSpec(0.0001)); @@ -231,7 +231,7 @@ TEST_F(ReferenceUtilTest, ConvArray3DWithValidPadding) { ReferenceUtil::ConvArray3D(input, weights, 1, Padding::kValid); Array3D expected = {{{17, 28, 39}}}; - auto actual_literal = Literal::CreateR3FromArray3D(*actual); + auto actual_literal = LiteralUtil::CreateR3FromArray3D(*actual); LiteralTestUtil::ExpectR3NearArray3D(expected, *actual_literal, ErrorSpec(0.0001)); @@ -266,7 +266,7 @@ TEST_F(ReferenceUtilTest, ConvWithSamePadding) { })); // clang-format on - auto actual_literal = Literal::CreateR4FromArray4D(*actual); + auto actual_literal = LiteralUtil::CreateR4FromArray4D(*actual); LiteralTestUtil::ExpectR4NearArray4D(expected, *actual_literal, ErrorSpec(0.0001)); @@ -300,7 +300,7 @@ TEST_F(ReferenceUtilTest, ConvWithValidPadding) { })); // clang-format on - auto actual_literal = Literal::CreateR4FromArray4D(*actual); + auto actual_literal = LiteralUtil::CreateR4FromArray4D(*actual); LiteralTestUtil::ExpectR4NearArray4D(expected, *actual_literal, ErrorSpec(0.0001)); @@ -356,7 +356,7 @@ TEST_F(ReferenceUtilTest, ConvGeneralDimensionsWithSamePadding) { }}); // clang-format on - auto actual_literal = Literal::CreateR4FromArray4D(*actual); + auto actual_literal = LiteralUtil::CreateR4FromArray4D(*actual); LiteralTestUtil::ExpectR4NearArray4D(expected, *actual_literal, ErrorSpec(0.0001)); @@ -409,7 +409,7 @@ TEST_F(ReferenceUtilTest, ConvGeneralDimensionsWithValidPadding) { Array4D expected({{{{2514, 2685}}}}); // clang-format on - auto actual_literal = Literal::CreateR4FromArray4D(*actual); + auto actual_literal = LiteralUtil::CreateR4FromArray4D(*actual); LiteralTestUtil::ExpectR4NearArray4D(expected, *actual_literal, ErrorSpec(0.0001)); @@ -422,7 +422,7 @@ TEST_F(ReferenceUtilTest, ApplyElementwise2D) { auto actual = ReferenceUtil::ApplyElementwise2D( [](float x, float y, float z) { return 100 * x + 10 * y + z; }, a, b, c); - auto actual_literal = Literal::CreateR2FromArray2D(*actual); + auto actual_literal = LiteralUtil::CreateR2FromArray2D(*actual); LiteralTestUtil::ExpectR2Near({{300.f, 600.f}, {900.f, 1200.f}}, *actual_literal, ErrorSpec(0.0001)); } diff --git a/tensorflow/compiler/xla/rpc/grpc_client_test.cc b/tensorflow/compiler/xla/rpc/grpc_client_test.cc index d7dd9786a2bbde2d18ae81a9a9d4cc4b2cc38411..90efee50b4f19056fac8ef1b341b48175903ff83 100644 --- a/tensorflow/compiler/xla/rpc/grpc_client_test.cc +++ b/tensorflow/compiler/xla/rpc/grpc_client_test.cc @@ -85,19 +85,19 @@ TEST_F(GRPCClientTestBase, ItsAlive) { TEST_F(GRPCClientTestBase, AxpyTenValues) { XlaBuilder builder("axpy_10"); - auto alpha = builder.ConstantR0(3.1415926535); - auto x = builder.ConstantR1( - {-1.0, 1.0, 2.0, -2.0, -3.0, 3.0, 4.0, -4.0, -5.0, 5.0}); - auto y = builder.ConstantR1( - {5.0, -5.0, -4.0, 4.0, 3.0, -3.0, -2.0, 2.0, 1.0, -1.0}); - auto ax = builder.Mul(alpha, x); - auto axpy = builder.Add(ax, y); + auto alpha = ConstantR0(&builder, 3.1415926535); + auto x = ConstantR1( + &builder, {-1.0, 1.0, 2.0, -2.0, -3.0, 3.0, 4.0, -4.0, -5.0, 5.0}); + auto y = ConstantR1( + &builder, {5.0, -5.0, -4.0, 4.0, 3.0, -3.0, -2.0, 2.0, 1.0, -1.0}); + auto ax = Mul(alpha, x); + Add(ax, y); std::vector expected = { 1.85840735, -1.85840735, 2.28318531, -2.28318531, -6.42477796, 6.42477796, 10.56637061, -10.56637061, -14.70796327, 14.70796327}; std::unique_ptr expected_literal = - Literal::CreateR1(expected); + LiteralUtil::CreateR1(expected); TF_ASSERT_OK_AND_ASSIGN(auto computation, builder.Build()); TF_ASSERT_OK_AND_ASSIGN(auto result_literal, client_->ExecuteAndTransfer( computation, {}, nullptr)); diff --git a/tensorflow/compiler/xla/service/BUILD b/tensorflow/compiler/xla/service/BUILD index c08960a57b056f3a58e83f2d8369d46356b8f5e3..85c6c632cdc2b2a6857e5aa87507dd5bc387ecd0 100644 --- a/tensorflow/compiler/xla/service/BUILD +++ b/tensorflow/compiler/xla/service/BUILD @@ -32,6 +32,7 @@ tf_proto_library_py( name = "hlo_proto", # bzl adds a _py suffix only to the OSS target. srcs = ["hlo.proto"], visibility = ["//visibility:public"], + deps = ["//tensorflow/compiler/xla:xla_data_proto_py"], ) xla_proto_library( @@ -135,7 +136,7 @@ cc_library( ":hlo_dce", ":hlo_pass", ":tuple_simplifier", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_tree", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:util", @@ -226,6 +227,7 @@ cc_library( ":hlo", ":hlo_query", ":shape_inference", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:literal_util", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:statusor", @@ -243,7 +245,7 @@ tf_cc_test( deps = [ ":hlo", ":hlo_evaluator", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:reference_util", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:status", @@ -293,6 +295,7 @@ cc_library( ":hlo_reachability", ":name_uniquer", "//tensorflow/compiler/xla:array", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:literal_util", "//tensorflow/compiler/xla:protobuf_util", "//tensorflow/compiler/xla:shape_tree", @@ -395,6 +398,7 @@ tf_cc_test( deps = [ ":hlo_matchers", ":hlo_parser", + "//tensorflow/compiler/xla:literal_util", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla/tests:xla_internal_test_main", ], @@ -406,7 +410,7 @@ tf_cc_test( deps = [ ":hlo", ":hlo_parser", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:protobuf_util", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:test", @@ -423,7 +427,7 @@ tf_cc_test( srcs = ["hlo_sharding_test.cc"], deps = [ ":hlo", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:protobuf_util", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:test", @@ -452,7 +456,7 @@ tf_cc_test( srcs = ["call_graph_test.cc"], deps = [ ":call_graph", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:status_macros", "//tensorflow/compiler/xla:test", @@ -486,6 +490,7 @@ cc_library( hdrs = ["call_inliner.h"], deps = [ ":call_graph", + ":hlo_dce", ":hlo_pass", "//tensorflow/compiler/xla:statusor", "//tensorflow/core:lib", @@ -501,7 +506,7 @@ tf_cc_test( ":hlo", ":hlo_matchers", ":hlo_pass", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:test", "//tensorflow/compiler/xla:types", @@ -520,7 +525,7 @@ tf_cc_test( deps = [ ":call_graph", ":flatten_call_graph", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:status_macros", "//tensorflow/compiler/xla:test", @@ -796,7 +801,7 @@ cc_library( hdrs = ["transfer_manager.h"], deps = [ ":shaped_buffer", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:status_macros", "//tensorflow/compiler/xla:statusor", @@ -959,7 +964,7 @@ tf_cc_test( ":hlo", ":hlo_ordering", ":hlo_scheduling", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:test", "//tensorflow/compiler/xla:test_helpers", @@ -1037,7 +1042,7 @@ tf_cc_test( ":hlo_ordering", ":hlo_value", ":tuple_points_to_analysis", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:status_macros", "//tensorflow/compiler/xla/tests:hlo_test_base", "//tensorflow/compiler/xla/tests:xla_internal_test_main", @@ -1120,7 +1125,7 @@ cc_library( hdrs = ["hlo_query.h"], deps = [ ":hlo", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_util", ], ) @@ -1169,6 +1174,7 @@ cc_library( deps = [ ":hlo", ":shape_inference", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:literal_util", "//tensorflow/compiler/xla:statusor", "//tensorflow/compiler/xla:util", @@ -1199,6 +1205,7 @@ cc_library( deps = [ ":hlo", ":hlo_pass", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:literal_util", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:status_macros", @@ -1218,6 +1225,7 @@ cc_library( ":hlo_creation_utils", ":hlo_pass", ":while_util", + "//tensorflow/compiler/xla:literal_util", "//tensorflow/compiler/xla:statusor", "//tensorflow/compiler/xla:util", ], @@ -1231,8 +1239,9 @@ tf_cc_test( ":batchnorm_expander", ":hlo", ":hlo_matchers", + ":hlo_parser", ":hlo_pass", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:test", "//tensorflow/compiler/xla:types", @@ -1254,6 +1263,7 @@ cc_library( ":hlo_pass", ":hlo_query", ":pattern_matcher", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:literal_util", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:status_macros", @@ -1273,7 +1283,7 @@ tf_cc_test( ":hlo", ":hlo_matchers", ":hlo_pass", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:test", "//tensorflow/compiler/xla:types", @@ -1309,7 +1319,7 @@ tf_cc_test( ":hlo", ":hlo_matchers", ":hlo_pass", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:test", "//tensorflow/compiler/xla:types", @@ -1344,7 +1354,7 @@ cc_library( ":call_inliner", ":hlo", ":hlo_pass", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:status_macros", "//tensorflow/compiler/xla:statusor", "//tensorflow/compiler/xla:types", @@ -1360,6 +1370,7 @@ tf_cc_test( ":conditional_simplifier", ":hlo", ":hlo_matchers", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:literal_util", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:test", @@ -1419,7 +1430,7 @@ tf_cc_test( deps = [ ":defuser", ":hlo_matchers", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla/tests:hlo_verified_test_base", ], @@ -1447,7 +1458,7 @@ tf_cc_test( deps = [ ":hlo_matchers", ":implicit_broadcast_remover", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla/tests:hlo_verified_test_base", ], @@ -1489,7 +1500,7 @@ tf_cc_test( ":hlo", ":hlo_matchers", ":tuple_simplifier", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:test", "//tensorflow/compiler/xla:types", @@ -1504,7 +1515,7 @@ cc_library( hdrs = ["reshape_mover.h"], deps = [ ":hlo_pass", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:status_macros", "//tensorflow/compiler/xla:util", @@ -1519,7 +1530,7 @@ tf_cc_test( ":hlo", ":hlo_matchers", ":reshape_mover", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:test", "//tensorflow/compiler/xla:test_helpers", @@ -1554,7 +1565,7 @@ tf_cc_test( ":hlo", ":hlo_matchers", ":inliner", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:test", "//tensorflow/compiler/xla:util", @@ -1571,7 +1582,7 @@ cc_library( hdrs = ["computation_placer.h"], deps = [ "//tensorflow/compiler/xla:array2d", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:status", "//tensorflow/compiler/xla:status_macros", @@ -1603,7 +1614,7 @@ cc_library( hdrs = ["generic_transfer_manager.h"], deps = [ ":transfer_manager", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:status_macros", "//tensorflow/compiler/xla:statusor", @@ -1694,7 +1705,7 @@ tf_cc_test( deps = [ ":hlo", ":hlo_matchers", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:test", "//tensorflow/compiler/xla:test_helpers", @@ -1709,6 +1720,7 @@ tf_cc_binary( deps = [ ":hlo", ":hlo_graph_dumper", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:literal_util", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:types", @@ -1723,7 +1735,7 @@ tf_cc_test( srcs = ["hlo_module_test.cc"], deps = [ ":hlo", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:test", "//tensorflow/compiler/xla:util", @@ -1821,7 +1833,7 @@ tf_cc_test( ":hlo_matchers", ":hlo_ordering", ":instruction_fusion", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:status_macros", "//tensorflow/compiler/xla:test", @@ -1858,7 +1870,7 @@ tf_cc_test( deps = [ ":hlo", ":hlo_liveness_analysis", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:status_macros", "//tensorflow/compiler/xla:test", @@ -1919,7 +1931,7 @@ tf_cc_test( ":hlo_matchers", ":hlo_ordering", ":instruction_fusion", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:test", "//tensorflow/compiler/xla:test_helpers", @@ -1954,6 +1966,7 @@ cc_library( ":hlo_dataflow_analysis", ":logical_buffer", ":logical_buffer_analysis", + "//tensorflow/compiler/xla:literal_util", "//tensorflow/compiler/xla:shape_tree", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:statusor", @@ -1972,6 +1985,7 @@ tf_cc_test( ":hlo_matchers", ":instruction_fusion", ":tuple_points_to_analysis", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:literal_util", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:test", @@ -2043,7 +2057,7 @@ tf_cc_test( ":hlo_graph_dumper", ":hlo_matchers", ":hlo_runner", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:test", "//tensorflow/compiler/xla:test_helpers", @@ -2094,6 +2108,7 @@ cc_library( hdrs = ["hlo_verifier.h"], deps = [ ":hlo", + ":hlo_casting_utils", ":hlo_pass", ":shape_inference", "//tensorflow/compiler/xla:status_macros", @@ -2106,6 +2121,7 @@ tf_cc_test( srcs = ["hlo_verifier_test.cc"], deps = [ ":hlo", + ":hlo_parser", ":hlo_verifier", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:test", @@ -2167,6 +2183,7 @@ tf_cc_test( deps = [ ":hlo", ":hlo_dce", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:literal_util", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:types", @@ -2187,7 +2204,7 @@ tf_cc_test( deps = [ ":hlo", ":hlo_module_dce", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:types", "//tensorflow/compiler/xla:util", @@ -2211,7 +2228,7 @@ tf_cc_test( ":hlo", ":hlo_matchers", ":layout_assignment", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_layout", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:test", @@ -2270,7 +2287,7 @@ cc_library( ":hlo", ":hlo_domain_map", ":hlo_pass", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:types", "//tensorflow/compiler/xla:xla_data_proto", @@ -2286,7 +2303,7 @@ tf_cc_test( ":hlo", ":hlo_cse", ":hlo_matchers", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:types", "//tensorflow/compiler/xla:util", @@ -2308,7 +2325,7 @@ cc_library( ":hlo_evaluator", ":hlo_pass", ":hlo_query", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:types", "//tensorflow/core:lib", @@ -2323,7 +2340,7 @@ tf_cc_test( ":hlo_constant_folding", ":hlo_matchers", ":hlo_pass", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:test", "//tensorflow/compiler/xla:types", @@ -2360,6 +2377,20 @@ cc_library( ], ) +cc_library( + name = "hlo_domain_verifier", + srcs = ["hlo_domain_verifier.cc"], + hdrs = ["hlo_domain_verifier.h"], + deps = [ + ":hlo", + ":hlo_domain_map", + ":hlo_graph_dumper", + ":hlo_pass", + "//tensorflow/compiler/xla:types", + "//tensorflow/core:lib", + ], +) + cc_library( name = "hlo_domain_isolator", srcs = ["hlo_domain_isolator.cc"], @@ -2379,8 +2410,8 @@ cc_library( hdrs = ["hlo_domain_remover.h"], deps = [ ":hlo", - ":hlo_domain_isolator", ":hlo_domain_map", + ":hlo_domain_verifier", ":hlo_graph_dumper", ":hlo_pass", "//tensorflow/compiler/xla:types", @@ -2415,7 +2446,7 @@ cc_library( ":hlo_evaluator", ":hlo_pass", ":hlo_query", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:types", "//tensorflow/core:lib", @@ -2550,7 +2581,7 @@ cc_library( hdrs = ["hlo_tfgraph_builder.h"], deps = [ ":hlo", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:xla_proto", "//tensorflow/core:framework", @@ -2581,7 +2612,7 @@ cc_library( ":hlo_casting_utils", ":hlo_execution_profile", ":hlo_tfgraph_builder", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:types", "//tensorflow/compiler/xla:window_util", @@ -2599,6 +2630,7 @@ tf_cc_test( deps = [ ":hlo", ":hlo_graph_dumper", + "//tensorflow/compiler/xla:literal_util", "//tensorflow/compiler/xla:test", "//tensorflow/compiler/xla:xla_proto", "//tensorflow/compiler/xla/tests:test_utils", @@ -2630,7 +2662,7 @@ tf_cc_test( ":hlo_matchers", ":shape_inference", ":transpose_folding", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:test", "//tensorflow/compiler/xla:test_helpers", @@ -2651,7 +2683,7 @@ cc_library( deps = [ ":hlo", ":hlo_pass", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:status_macros", "//tensorflow/compiler/xla:util", @@ -2666,7 +2698,7 @@ tf_cc_test( ":hlo", ":shape_inference", ":zero_sized_hlo_elimination", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:status_macros", "//tensorflow/compiler/xla:test", @@ -2826,6 +2858,7 @@ cc_library( ":hlo", ":hlo_creation_utils", ":tuple_util", + "//tensorflow/compiler/xla:literal_util", "//tensorflow/core:lib", ], ) @@ -2961,6 +2994,7 @@ cc_library( ":hlo", ":hlo_lexer", ":hlo_sharding_metadata", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:literal_util", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:statusor", diff --git a/tensorflow/compiler/xla/service/algebraic_simplifier.cc b/tensorflow/compiler/xla/service/algebraic_simplifier.cc index d8a9aba834c27a16100ff570682935ba9725af33..af7728da5494b9aff1e657e07055bcb8fbaff470 100644 --- a/tensorflow/compiler/xla/service/algebraic_simplifier.cc +++ b/tensorflow/compiler/xla/service/algebraic_simplifier.cc @@ -23,6 +23,7 @@ limitations under the License. #include #include "tensorflow/compiler/xla/layout_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/literal_util.h" #include "tensorflow/compiler/xla/service/dfs_hlo_visitor_with_default.h" #include "tensorflow/compiler/xla/service/hlo_computation.h" @@ -50,20 +51,15 @@ namespace { namespace m = match; -// Returns whether operand is a literal with the given value. -bool IsLiteralWithValue(const HloInstruction* operand, int8 value) { - return operand->opcode() == HloOpcode::kConstant && - operand->literal().IsAll(value); -} - bool IsAll(const HloInstruction* op, int8 value) { - if (IsLiteralWithValue(op, value)) { - return true; - } - if (op->opcode() == HloOpcode::kBroadcast && IsAll(op->operand(0), value)) { - return true; + switch (op->opcode()) { + case HloOpcode::kBroadcast: + return IsAll(op->operand(0), value); + case HloOpcode::kConstant: + return op->literal().IsAll(value); + default: + return false; } - return false; } // Returns whether the given transpose produces a result which is bit-wise @@ -160,9 +156,6 @@ class AlgebraicSimplifierVisitor : public DfsHloVisitorWithDefault { Status HandleMap(HloInstruction* map) override; - Status HandleMaximum(HloInstruction* maximum) override; - Status HandleMinimum(HloInstruction* minimum) override; - // Returns whether algebraic simplification has occurred. const bool changed() const { return changed_; } @@ -201,8 +194,9 @@ class AlgebraicSimplifierVisitor : public DfsHloVisitorWithDefault { // Helper method to perform and add reduction in a single dimension. HloInstruction* AddReduce(HloInstruction* hlo, int64 dim) { - HloInstruction* zero = computation_->AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(0.0f))); + HloInstruction* zero = + computation_->AddInstruction(HloInstruction::CreateConstant( + LiteralUtil::Zero(hlo->shape().element_type()).CloneToUnique())); HloComputation* AddReduce_computation = GetOrCreateScalarAddComputation(); Shape shape = ShapeUtil::DeleteDimension(dim, hlo->shape()); return computation_->AddInstruction(HloInstruction::CreateReduce( @@ -537,11 +531,15 @@ Status AlgebraicSimplifierVisitor::HandleConstant(HloInstruction* constant) { constant, BuildTupleConstant(computation_, constant->literal())); } + if (constant->shape().element_type() == TOKEN) { + return Status::OK(); + } + // If a literal is all the same element replace it with a scalar broadcast. if (ShapeUtil::ElementsIn(constant->shape()) > 1 && constant->literal().IsAllFirst()) { - std::unique_ptr unique_scalar = - MakeUnique(constant->literal().GetFirstScalarLiteral()); + std::unique_ptr unique_scalar = MakeUnique( + LiteralUtil::GetFirstScalarLiteral(constant->literal())); HloInstruction* scalar = computation_->AddInstruction( HloInstruction::CreateConstant(std::move(unique_scalar))); return ReplaceWithNewInstruction( @@ -572,6 +570,14 @@ Status AlgebraicSimplifierVisitor::HandleSubtract(HloInstruction* sub) { return Status::OK(); } +namespace { +template +Status InvertConstant(const HloInstruction& constant, Literal* result) { + return result->Populate([&](tensorflow::gtl::ArraySlice indices) { + return T{1.0} / constant.literal().Get(indices); + }); +} +} // namespace Status AlgebraicSimplifierVisitor::HandleDivide(HloInstruction* divide) { Shape* shape; @@ -633,14 +639,31 @@ Status AlgebraicSimplifierVisitor::HandleDivide(HloInstruction* divide) { // (Backends can do this transformation, but generally only if the constant is // a scalar.) if (Match(divide, m::Divide(m::NonConstant(&a), m::Constant(&b)))) { - HloInstruction* one = - computation_->AddInstruction(HloInstruction::CreateConstant( - Literal::One(a->shape().element_type()).CloneToUnique())); - HloInstruction* inverse = computation_->AddInstruction( - HloInstruction::CreateBinary(b->shape(), HloOpcode::kDivide, one, b)); - return ReplaceWithNewInstruction( - divide, HloInstruction::CreateBinary(divide->shape(), - HloOpcode::kMultiply, a, inverse)); + Literal new_literal(b->shape()); + switch (b->shape().element_type()) { + case F16: + TF_RETURN_IF_ERROR(InvertConstant(*b, &new_literal)); + break; + case F32: + TF_RETURN_IF_ERROR(InvertConstant(*b, &new_literal)); + break; + case BF16: + TF_RETURN_IF_ERROR(InvertConstant(*b, &new_literal)); + break; + case F64: + TF_RETURN_IF_ERROR(InvertConstant(*b, &new_literal)); + break; + case C64: + TF_RETURN_IF_ERROR(InvertConstant(*b, &new_literal)); + break; + default: + return Status::OK(); + } + auto inverse = computation_->AddInstruction( + HloInstruction::CreateConstant((new_literal.CloneToUnique()))); + TF_ASSIGN_OR_RETURN(auto new_divide, + MakeBinaryHlo(HloOpcode::kMultiply, a, inverse)); + return ReplaceInstruction(divide, new_divide); } // (A / B) / (C / D) => (A / B)*(D / C) => (A * D) / (B * C) @@ -660,18 +683,18 @@ Status AlgebraicSimplifierVisitor::HandleDivide(HloInstruction* divide) { if (Match(divide, m::Divide(m::Divide(m::Op(&a), m::Op(&b)), m::Op(&c)))) { TF_ASSIGN_OR_RETURN(auto b_times_c, MakeBinaryHlo(HloOpcode::kMultiply, b, c)); - return ReplaceWithNewInstruction( - divide, HloInstruction::CreateBinary(divide->shape(), - HloOpcode::kDivide, a, b_times_c)); + TF_ASSIGN_OR_RETURN(auto new_divide, + MakeBinaryHlo(HloOpcode::kDivide, a, b_times_c)); + return ReplaceInstruction(divide, new_divide); } // A / (B / C) => (A*C) / B if (Match(divide, m::Divide(m::Op(&a), m::Divide(m::Op(&b), m::Op(&c))))) { TF_ASSIGN_OR_RETURN(auto a_times_c, MakeBinaryHlo(HloOpcode::kMultiply, a, c)); - return ReplaceWithNewInstruction( - divide, HloInstruction::CreateBinary(divide->shape(), - HloOpcode::kDivide, a_times_c, b)); + TF_ASSIGN_OR_RETURN(auto new_divide, + MakeBinaryHlo(HloOpcode::kDivide, a_times_c, b)); + return ReplaceInstruction(divide, new_divide); } return Status::OK(); @@ -1071,7 +1094,7 @@ Status AlgebraicSimplifierVisitor::HandleDot(HloInstruction* dot) { ShapeUtil::IsZeroElementArray(lhs->shape()) || ShapeUtil::IsZeroElementArray(rhs->shape())) { auto zero = computation_->AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(0.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(0.0f))); return ReplaceWithNewInstruction( dot, HloInstruction::CreateBroadcast(dot->shape(), zero, {})); } @@ -1230,9 +1253,10 @@ bool OutputIsPermutationOfOperandElements(HloInstruction* instruction, switch (instruction->opcode()) { case HloOpcode::kReshape: case HloOpcode::kReverse: - case HloOpcode::kSort: case HloOpcode::kTranspose: return true; + case HloOpcode::kSort: + return (!ShapeUtil::IsTuple(instruction->shape())); default: return false; } @@ -1496,7 +1520,7 @@ Status AlgebraicSimplifierVisitor::HandlePower(HloInstruction* power) { CHECK(Match(power, m::Power(m::Op(&lhs), m::Op(&rhs)))); if (IsAll(rhs, 0)) { auto one = HloInstruction::CreateConstant( - Literal::One(power->shape().element_type()).CloneToUnique()); + LiteralUtil::One(power->shape().element_type()).CloneToUnique()); std::unique_ptr ones; if (ShapeUtil::IsScalar(power->shape())) { ones = std::move(one); @@ -1531,7 +1555,7 @@ Status AlgebraicSimplifierVisitor::HandlePower(HloInstruction* power) { VLOG(10) << "trying transform [pow(A, -1) => 1/A]: " << power->ToString(); if (IsAll(rhs, -1)) { auto* one = computation_->AddInstruction(HloInstruction::CreateConstant( - Literal::One(rhs->shape().element_type()).CloneToUnique())); + LiteralUtil::One(rhs->shape().element_type()).CloneToUnique())); // Explicitly broadcast scalar 1 to the output shape, to avoid implicit // broadcast in divide HLO as we are trying to eliminate implicit @@ -2074,10 +2098,9 @@ Status AlgebraicSimplifierVisitor::HandleConvolution( convolution, HloInstruction::CreateBroadcast( convolution->shape(), - computation_->AddInstruction(HloInstruction::CreateConvert( - ShapeUtil::MakeShape(convolution->shape().element_type(), {}), - computation_->AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(0.0f))))), + computation_->AddInstruction(HloInstruction::CreateConstant( + LiteralUtil::Zero(convolution->shape().element_type()) + .CloneToUnique())), {})); } const auto& window = convolution->window(); @@ -2249,68 +2272,6 @@ Status AlgebraicSimplifierVisitor::HandleMap(HloInstruction* map) { return ReplaceWithNewInstruction(map, std::move(clone)); } -Status AlgebraicSimplifierVisitor::HandleMaximum(HloInstruction* maximum) { - // Match the following tree: - // min_operand operand - // \ / - // max_operand min - // \ / - // max - // where max_operand and min_operand are scalar constants. - { - HloInstruction* min; - HloInstruction* max_operand; - HloInstruction* min_operand; - HloInstruction* operand; - - if (hlo_query::MatchBinaryInstructionOperandOpcode( - HloOpcode::kMinimum, maximum, - /*matching_operand=*/&min, - /*other_operand=*/&max_operand) && - hlo_query::MatchBinaryInstructionOperand( - hlo_query::IsScalarConstant, min, - /*matching_operand=*/&min_operand, - /*other_operand=*/&operand) && - TransformToClampIfSameShape(maximum, min, min_operand, operand, maximum, - max_operand)) { - return Status::OK(); - } - } - - return Status::OK(); -} - -Status AlgebraicSimplifierVisitor::HandleMinimum(HloInstruction* minimum) { - // Match the following tree: - // max_operand operand - // \ / - // min_operand max - // \ / - // min - // where max_operand and min_operand are scalar constants. - { - HloInstruction* max; - HloInstruction* max_operand; - HloInstruction* min_operand; - HloInstruction* operand; - - if (hlo_query::MatchBinaryInstructionOperandOpcode( - HloOpcode::kMaximum, minimum, - /*matching_operand=*/&max, - /*other_operand=*/&min_operand) && - hlo_query::MatchBinaryInstructionOperand( - hlo_query::IsScalarConstant, max, - /*matching_operand=*/&max_operand, - /*other_operand=*/&operand) && - TransformToClampIfSameShape(minimum, minimum, min_operand, operand, max, - max_operand)) { - return Status::OK(); - } - } - - return Status::OK(); -} - StatusOr AlgebraicSimplifier::Run(HloModule* module) { XLA_VLOG_LINES(2, "AlgebraicSimplifier::Run(), before:\n" + module->ToString()); diff --git a/tensorflow/compiler/xla/service/algebraic_simplifier_test.cc b/tensorflow/compiler/xla/service/algebraic_simplifier_test.cc index 49cc0b808bc193441741b02ca45a82c49a8aba7c..92bbcbd740f35f55e9cbc91ba36ac2af50767585 100644 --- a/tensorflow/compiler/xla/service/algebraic_simplifier_test.cc +++ b/tensorflow/compiler/xla/service/algebraic_simplifier_test.cc @@ -19,7 +19,7 @@ limitations under the License. #include #include "tensorflow/compiler/xla/layout_util.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/ptr_util.h" #include "tensorflow/compiler/xla/service/hlo_computation.h" #include "tensorflow/compiler/xla/service/hlo_instruction.h" @@ -60,7 +60,7 @@ TEST_F(AlgebraicSimplifierTest, AddZero) { HloInstruction* param0 = builder.AddInstruction( HloInstruction::CreateParameter(0, r0f32, "param0")); HloInstruction* zero = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(0.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(0.0f))); builder.AddInstruction( HloInstruction::CreateBinary(r0f32, HloOpcode::kAdd, param0, zero)); @@ -79,7 +79,7 @@ TEST_F(AlgebraicSimplifierTest, TwoReducesToOne) { HloComputation::Builder builder(TestName()); // Create add computation. HloInstruction* zero = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(0.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(0.0f))); HloComputation* add_computation = nullptr; { HloComputation::Builder builder(TestName() + ".add"); @@ -119,7 +119,7 @@ TEST_F(AlgebraicSimplifierTest, AddConstOnLHS) { HloInstruction* param0 = builder.AddInstruction( HloInstruction::CreateParameter(0, r0f32, "param0")); HloInstruction* constant = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(42.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(42.0f))); builder.AddInstruction( HloInstruction::CreateBinary(r0f32, HloOpcode::kAdd, constant, param0)); @@ -140,9 +140,9 @@ TEST_F(AlgebraicSimplifierTest, AddReassociateMergeConstants) { HloInstruction* param0 = builder.AddInstruction( HloInstruction::CreateParameter(0, r0f32, "param0")); HloInstruction* constant1 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(42.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(42.0f))); HloInstruction* constant2 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(3.14159f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(3.14159f))); HloInstruction* add1 = builder.AddInstruction( HloInstruction::CreateBinary(r0f32, HloOpcode::kAdd, param0, constant1)); @@ -165,7 +165,7 @@ TEST_F(AlgebraicSimplifierTest, AddBroadcastZeroR0Operand) { HloInstruction* param0 = builder.AddInstruction( HloInstruction::CreateParameter(0, r2f32, "param0")); HloInstruction* zero = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(0.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(0.0f))); HloInstruction* bcast = builder.AddInstruction( HloInstruction::CreateBroadcast(r2f32, zero, {0, 1})); builder.AddInstruction( @@ -200,9 +200,12 @@ TEST_F(AlgebraicSimplifierTest, InlineTrivialMap) { HloInstruction* param0 = builder.AddInstruction( HloInstruction::CreateParameter(0, r2f32, "param0")); HloInstruction* zero = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(0.0f))); - builder.AddInstruction( - HloInstruction::CreateMap(r2f32, {param0, zero}, add_computation)); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(0.0f))); + builder.AddInstruction(HloInstruction::CreateMap( + r2f32, + {param0, builder.AddInstruction( + HloInstruction::CreateBroadcast(r2f32, zero, {}))}, + add_computation)); auto computation = module().AddEntryComputation(builder.Build()); HloInstruction* root = computation->root_instruction(); @@ -211,7 +214,7 @@ TEST_F(AlgebraicSimplifierTest, InlineTrivialMap) { non_bitcasting_callback()); ASSERT_TRUE(simplifier.Run(&module()).ValueOrDie()); root = computation->root_instruction(); - EXPECT_THAT(root, op::Add(param0, zero)); + EXPECT_THAT(root, op::Add(param0, op::Broadcast(zero))); } TEST_F(AlgebraicSimplifierTest, AddBroadcastZeroR1Operand) { @@ -220,7 +223,7 @@ TEST_F(AlgebraicSimplifierTest, AddBroadcastZeroR1Operand) { HloInstruction* param0 = builder.AddInstruction( HloInstruction::CreateParameter(0, r2f32, "param0")); HloInstruction* zero = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR1({0, 0, 0}))); + HloInstruction::CreateConstant(LiteralUtil::CreateR1({0, 0, 0}))); HloInstruction* bcast = builder.AddInstruction(HloInstruction::CreateBroadcast(r2f32, zero, {1})); builder.AddInstruction( @@ -239,7 +242,7 @@ TEST_F(AlgebraicSimplifierTest, AddBroadcastZeroR1Operand) { TEST_F(AlgebraicSimplifierTest, ConstantToBroadcast) { HloComputation::Builder builder(TestName()); builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR1({3.14f, 3.14f, 3.14f}))); + LiteralUtil::CreateR1({3.14f, 3.14f, 3.14f}))); auto computation = module().AddEntryComputation(builder.Build()); HloInstruction* root = computation->root_instruction(); @@ -255,7 +258,7 @@ TEST_F(AlgebraicSimplifierTest, ConstantToBroadcast) { TEST_F(AlgebraicSimplifierTest, ConstantNotToBroadcast) { HloComputation::Builder builder(TestName()); builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR1({3.14, 3.14, 4}))); + LiteralUtil::CreateR1({3.14, 3.14, 4}))); auto computation = module().AddEntryComputation(builder.Build()); HloInstruction* root = computation->root_instruction(); @@ -274,7 +277,7 @@ TEST_F(AlgebraicSimplifierTest, SubZero) { HloInstruction* param0 = builder.AddInstruction( HloInstruction::CreateParameter(0, r0f32, "param0")); HloInstruction* zero = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(0.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(0.0f))); builder.AddInstruction( HloInstruction::CreateBinary(r0f32, HloOpcode::kSubtract, param0, zero)); @@ -295,7 +298,7 @@ TEST_F(AlgebraicSimplifierTest, SubConstCanonicalization) { HloInstruction* param0 = builder.AddInstruction( HloInstruction::CreateParameter(0, r0f32, "param0")); HloInstruction* constant = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(42.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(42.0f))); builder.AddInstruction(HloInstruction::CreateBinary( r0f32, HloOpcode::kSubtract, param0, constant)); @@ -367,17 +370,16 @@ TEST_F(AlgebraicSimplifierTest, RhsDivOfDiv) { // Test that (A/B)/(C/D) is simplified to (A*D)/(B*C). TEST_F(AlgebraicSimplifierTest, DivOfDivAndDiv) { - Shape r0f32 = ShapeUtil::MakeShape(F32, {}); Shape r2f32 = ShapeUtil::MakeShape(F32, {42, 123}); HloComputation::Builder builder(TestName()); HloInstruction* param0 = builder.AddInstruction( - HloInstruction::CreateParameter(0, r0f32, "param0")); + HloInstruction::CreateParameter(0, r2f32, "param0")); HloInstruction* param1 = builder.AddInstruction( HloInstruction::CreateParameter(1, r2f32, "param1")); HloInstruction* param2 = builder.AddInstruction( HloInstruction::CreateParameter(2, r2f32, "param2")); HloInstruction* param3 = builder.AddInstruction( - HloInstruction::CreateParameter(3, r0f32, "param3")); + HloInstruction::CreateParameter(3, r2f32, "param3")); HloInstruction* div0 = builder.AddInstruction( HloInstruction::CreateBinary(r2f32, HloOpcode::kDivide, param0, param1)); HloInstruction* div1 = builder.AddInstruction( @@ -398,8 +400,6 @@ TEST_F(AlgebraicSimplifierTest, DivOfDivAndDiv) { EXPECT_THAT( computation->root_instruction(), op::Divide(op::Multiply(param0, param3), op::Multiply(param1, param2))); - EXPECT_TRUE( - ShapeUtil::Compatible(computation->root_instruction()->shape(), r2f32)); } // Test that A/exp(B) is simplified to A*exp(-B). @@ -459,7 +459,6 @@ TEST_F(AlgebraicSimplifierTest, DivOfPower) { // Test that broadcasting is done on the right step when simplifying A/pow(B,C) // to A*pow(B,-C). TEST_F(AlgebraicSimplifierTest, DivOfBroadcastingPower) { - Shape r0f32 = ShapeUtil::MakeShape(F32, {}); Shape r1f32 = ShapeUtil::MakeShape(F32, {7}); HloComputation::Builder builder(TestName()); HloInstruction* param0 = builder.AddInstruction( @@ -467,7 +466,7 @@ TEST_F(AlgebraicSimplifierTest, DivOfBroadcastingPower) { HloInstruction* param1 = builder.AddInstruction( HloInstruction::CreateParameter(1, r1f32, "param1")); HloInstruction* param2 = builder.AddInstruction( - HloInstruction::CreateParameter(2, r0f32, "param2")); + HloInstruction::CreateParameter(2, r1f32, "param2")); HloInstruction* power = builder.AddInstruction( HloInstruction::CreateBinary(r1f32, HloOpcode::kPower, param1, param2)); builder.AddInstruction( @@ -484,14 +483,9 @@ TEST_F(AlgebraicSimplifierTest, DivOfBroadcastingPower) { ASSERT_THAT(computation->root_instruction(), op::Multiply(param0, op::Power(param1, op::Negate(param2)))); - - const HloInstruction* negate = - computation->root_instruction()->operand(1)->operand(1); - const Shape& negate_shape = negate->shape(); - EXPECT_EQ(0, negate_shape.dimensions_size()); } -// A / Const => A * (1 / Const) +// A / Const => A * InvertedConst TEST_F(AlgebraicSimplifierTest, DivideByConstant) { Shape r1f32 = ShapeUtil::MakeShape(F32, {3}); HloComputation::Builder builder(TestName()); @@ -499,7 +493,7 @@ TEST_F(AlgebraicSimplifierTest, DivideByConstant) { HloInstruction::CreateParameter(0, r1f32, "param0")); HloInstruction* constant = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR1({0.f, 1.f, 2.f}))); + LiteralUtil::CreateR1({0.f, 1.f, 2.f}))); builder.AddInstruction(HloInstruction::CreateBinary(r1f32, HloOpcode::kDivide, param0, constant)); @@ -510,20 +504,19 @@ TEST_F(AlgebraicSimplifierTest, DivideByConstant) { ASSERT_TRUE(simplifier.Run(&module()).ValueOrDie()); EXPECT_THAT(computation->root_instruction(), - op::Multiply(param0, op::Divide(op::Constant(), constant))); + op::Multiply(param0, op::Constant())); } // pow(pow(A, X), Y) => pow(A, X*Y) TEST_F(AlgebraicSimplifierTest, PowerOfPower) { - Shape r0f32 = ShapeUtil::MakeShape(F32, {}); Shape r1f32 = ShapeUtil::MakeShape(F32, {7}); HloComputation::Builder builder(TestName()); HloInstruction* base = builder.AddInstruction( HloInstruction::CreateParameter(0, r1f32, "param0")); HloInstruction* exp1 = builder.AddInstruction( - HloInstruction::CreateParameter(1, r0f32, "param1")); + HloInstruction::CreateParameter(1, r1f32, "param1")); HloInstruction* exp2 = builder.AddInstruction( - HloInstruction::CreateParameter(2, r0f32, "param2")); + HloInstruction::CreateParameter(2, r1f32, "param2")); HloInstruction* inner_power = builder.AddInstruction( HloInstruction::CreateBinary(r1f32, HloOpcode::kPower, base, exp1)); builder.AddInstruction(HloInstruction::CreateBinary(r1f32, HloOpcode::kPower, @@ -540,15 +533,14 @@ TEST_F(AlgebraicSimplifierTest, PowerOfPower) { // Don't simplify pow(pow(A, X), Y) => pow(A, X*Y) if X and Y are complex // numbers. TEST_F(AlgebraicSimplifierTest, PowerOfPowerComplex) { - Shape r0c64 = ShapeUtil::MakeShape(C64, {}); Shape r1c64 = ShapeUtil::MakeShape(C64, {7}); HloComputation::Builder builder(TestName()); HloInstruction* base = builder.AddInstruction( HloInstruction::CreateParameter(0, r1c64, "param0")); HloInstruction* exp1 = builder.AddInstruction( - HloInstruction::CreateParameter(1, r0c64, "param1")); + HloInstruction::CreateParameter(1, r1c64, "param1")); HloInstruction* exp2 = builder.AddInstruction( - HloInstruction::CreateParameter(2, r0c64, "param2")); + HloInstruction::CreateParameter(2, r1c64, "param2")); HloInstruction* inner_power = builder.AddInstruction( HloInstruction::CreateBinary(r1c64, HloOpcode::kPower, base, exp1)); builder.AddInstruction(HloInstruction::CreateBinary(r1c64, HloOpcode::kPower, @@ -567,7 +559,7 @@ TEST_F(AlgebraicSimplifierTest, DivOneScalar) { HloInstruction* param0 = builder.AddInstruction( HloInstruction::CreateParameter(0, r0f32, "param0")); HloInstruction* one = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0f))); HloInstruction* div = builder.AddInstruction( HloInstruction::CreateBinary(r0f32, HloOpcode::kDivide, param0, one)); @@ -588,7 +580,7 @@ TEST_F(AlgebraicSimplifierTest, DivOneArray) { HloInstruction* param0 = builder.AddInstruction( HloInstruction::CreateParameter(0, r2f32, "param0")); HloInstruction* one = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR2({{1.0, 1.0}, {1.0, 1.0}}))); + LiteralUtil::CreateR2({{1.0, 1.0}, {1.0, 1.0}}))); HloInstruction* div = builder.AddInstruction( HloInstruction::CreateBinary(r2f32, HloOpcode::kDivide, param0, one)); @@ -868,7 +860,7 @@ TEST_F(AlgebraicSimplifierTest, Pow0Scalar) { HloInstruction* param0 = builder.AddInstruction( HloInstruction::CreateParameter(0, r0f32, "param0")); HloInstruction* zero = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(0))); builder.AddInstruction( HloInstruction::CreateBinary(r0f32, HloOpcode::kPower, param0, zero)); @@ -892,7 +884,7 @@ TEST_F(AlgebraicSimplifierTest, Pow0Vector) { HloInstruction* param0 = builder.AddInstruction( HloInstruction::CreateParameter(0, r1f32, "param0")); HloInstruction* zero = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(0))); builder.AddInstruction( HloInstruction::CreateBinary(r1f32, HloOpcode::kPower, param0, zero)); @@ -920,7 +912,7 @@ TEST_F(AlgebraicSimplifierTest, Pow1) { HloInstruction* param0 = builder.AddInstruction( HloInstruction::CreateParameter(0, r0f32, "param0")); HloInstruction* one = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1))); builder.AddInstruction( HloInstruction::CreateBinary(r0f32, HloOpcode::kPower, param0, one)); @@ -942,7 +934,7 @@ TEST_F(AlgebraicSimplifierTest, Pow2) { HloInstruction* param0 = builder.AddInstruction( HloInstruction::CreateParameter(0, r0f32, "param0")); HloInstruction* two = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(2))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(2))); builder.AddInstruction( HloInstruction::CreateBinary(r0f32, HloOpcode::kPower, param0, two)); @@ -964,7 +956,7 @@ TEST_F(AlgebraicSimplifierTest, PowNegative1) { HloInstruction* param0 = builder.AddInstruction( HloInstruction::CreateParameter(0, r0f32, "param0")); HloInstruction* negative_one = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(-1))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(-1))); builder.AddInstruction(HloInstruction::CreateBinary(r0f32, HloOpcode::kPower, param0, negative_one)); @@ -1055,7 +1047,7 @@ TEST_F(AlgebraicSimplifierTest, ZeroSizedReduceWindow) { builder.AddInstruction(HloInstruction::CreateReduceWindow( ShapeUtil::MakeShape(F32, {5, 2}), param, builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(0.0f))), + HloInstruction::CreateConstant(LiteralUtil::CreateR0(0.0f))), window, add_computation)); module().AddEntryComputation(builder.Build()); HloPassFix simplifier(/*is_layout_sensitive=*/false, @@ -1082,7 +1074,7 @@ TEST_F(AlgebraicSimplifierTest, ZeroSizedPad) { builder.AddInstruction(HloInstruction::CreatePad( ShapeUtil::MakeShape(F32, {5, 2}), param, builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(0.0f))), + HloInstruction::CreateConstant(LiteralUtil::CreateR0(0.0f))), padding)); module().AddEntryComputation(builder.Build()); EXPECT_THAT(module().entry_computation()->root_instruction(), @@ -1124,7 +1116,7 @@ TEST_F(AlgebraicSimplifierTest, ReshapeBroadcast) { TEST_F(AlgebraicSimplifierTest, ConvertBetweenSameType) { HloComputation::Builder builder(TestName()); HloInstruction* input = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(42.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(42.0f))); builder.AddInstruction( HloInstruction::CreateConvert(ShapeUtil::MakeShape(F32, {}), input)); @@ -1216,7 +1208,7 @@ TEST_F(AlgebraicSimplifierTest, RemoveEmptyConcatenateOperands) { HloInstruction* param1 = builder.AddInstruction( HloInstruction::CreateParameter(1, r1f32, "param1")); HloInstruction* empty_literal = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR1({}))); + HloInstruction::CreateConstant(LiteralUtil::CreateR1({}))); HloInstruction* empty_slice = builder.AddInstruction(HloInstruction::CreateSlice( ShapeUtil::MakeShape(F32, {0}), param1, {42}, {42}, {1})); @@ -1246,7 +1238,7 @@ TEST_F(AlgebraicSimplifierTest, OnlyEmptyConcatenateOperands) { HloInstruction* param0 = builder.AddInstruction( HloInstruction::CreateParameter(0, r1f32, "param0")); HloInstruction* empty_literal = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR1({}))); + HloInstruction::CreateConstant(LiteralUtil::CreateR1({}))); HloInstruction* empty_slice = builder.AddInstruction(HloInstruction::CreateSlice( ShapeUtil::MakeShape(F32, {0}), param0, {42}, {42}, {1})); @@ -1416,33 +1408,6 @@ TEST_F(AlgebraicSimplifierTest, ReshapeReplacedWithBitcast) { op::Tuple(op::Bitcast(), dimensions_wrong_reshape, layout_wrong_reshape)); } -// Regression test for a bug in the reshape sinking transformation, where -// moving a reshape to a scalar led to a crash. -TEST_F(AlgebraicSimplifierTest, ReshapeToScalarNotHoistedAfterEffectiveUnary) { - HloComputation::Builder builder(TestName()); - HloInstruction* param = - builder.AddInstruction(HloInstruction::CreateParameter( - 0, ShapeUtil::MakeShape(F32, {1, 1}), "param")); - HloInstruction* reshape = builder.AddInstruction( - HloInstruction::CreateReshape(ShapeUtil::MakeShape(F32, {}), param)); - HloInstruction* zero = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR1({1., 2., 3.}))); - builder.AddInstruction(HloInstruction::CreateBinary( - ShapeUtil::MakeShape(F32, {3}), HloOpcode::kMaximum, reshape, zero)); - auto computation = module().AddEntryComputation(builder.Build()); - - EXPECT_THAT(computation->root_instruction(), - op::Maximum(op::Reshape(param), zero)); - - AlgebraicSimplifier simplifier(/*is_layout_sensitive=*/false, - bitcasting_callback()); - - simplifier.Run(&module()).ValueOrDie(); - - EXPECT_THAT(computation->root_instruction(), - op::Maximum(op::Reshape(param), zero)); -} - // Regression test for a bug where if we failed to sink a reshape, we'd set the // 'changed' bit in AlgebraicSimplifier to false. TEST_F(AlgebraicSimplifierTest, FailureToSinkReshapeDoesntAffectChangedBit) { @@ -1455,7 +1420,7 @@ TEST_F(AlgebraicSimplifierTest, FailureToSinkReshapeDoesntAffectChangedBit) { builder.AddInstruction( HloInstruction::CreateParameter(0, shape, "param0")), builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR2({{0, 0}, {0, 0}}))))); + LiteralUtil::CreateR2({{0, 0}, {0, 0}}))))); builder.AddInstruction( HloInstruction::CreateReshape(ShapeUtil::MakeShape(F32, {4}), add)); @@ -1478,7 +1443,7 @@ TEST_F(AlgebraicSimplifierTest, FailureToSinkBroadcastDoesntAffectChangedBit) { builder.AddInstruction( HloInstruction::CreateParameter(0, shape, "param0")), builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR2({{0, 0}, {0, 0}}))))); + LiteralUtil::CreateR2({{0, 0}, {0, 0}}))))); builder.AddInstruction( HloInstruction::CreateBroadcast(ShapeUtil::MakeShape(F32, {2, 2, 2}), add, @@ -1761,7 +1726,7 @@ TEST_F(AlgebraicSimplifierTest, RemoveNoopPad) { builder.AddInstruction(HloInstruction::CreateParameter( 0, ShapeUtil::MakeShape(F32, {2, 2}), "param")); HloInstruction* zero = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(0.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(0.0f))); PaddingConfig no_padding; for (int i = 0; i < 2; ++i) { auto dimension = no_padding.add_dimensions(); @@ -1792,7 +1757,7 @@ TEST_F(AlgebraicSimplifierTest, NegativePadding) { builder.AddInstruction(HloInstruction::CreateParameter( 0, ShapeUtil::MakeShape(F32, {10, 10}), "param")); HloInstruction* zero = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(0.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(0.0f))); PaddingConfig padding; int64 low_padding[2] = {-1, -2}; int64 high_padding[2] = {2, -3}; @@ -2103,160 +2068,6 @@ TEST_F(AlgebraicSimplifierTest, ConvertConvToMatmul) { EXPECT_EQ("NO_CHANGE", build_and_simplify()); } -// Test that max(min(A, x), y) is transformed to clamp(y, A, x) -TEST_F(AlgebraicSimplifierTest, MaxMinToClamp) { - Shape r0f32 = ShapeUtil::MakeShape(F32, {}); - HloComputation::Builder builder(TestName()); - HloInstruction* param0 = builder.AddInstruction( - HloInstruction::CreateParameter(0, r0f32, "param0")); - HloInstruction* min_value = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(0.0f))); - HloInstruction* max_value = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0f))); - HloInstruction* min = builder.AddInstruction(HloInstruction::CreateBinary( - r0f32, HloOpcode::kMinimum, param0, min_value)); - builder.AddInstruction( - HloInstruction::CreateBinary(r0f32, HloOpcode::kMaximum, min, max_value)); - - auto module = CreateNewModule(); - auto computation = module->AddEntryComputation(builder.Build()); - - EXPECT_THAT(computation->root_instruction(), - op::Maximum(op::Minimum(param0, min_value), max_value)); - - AlgebraicSimplifier simplifier(/*is_layout_sensitive=*/false, - non_bitcasting_callback()); - ASSERT_TRUE(simplifier.Run(module).ValueOrDie()); - - EXPECT_THAT(computation->root_instruction(), - op::Clamp(max_value, param0, min_value)); -} - -// Test that min(max(A, x), y) is transformed to clamp(x, A, y) for scalar -// values. -TEST_F(AlgebraicSimplifierTest, MinMaxToClamp) { - Shape r0f32 = ShapeUtil::MakeShape(F32, {}); - HloComputation::Builder builder(TestName()); - HloInstruction* param0 = builder.AddInstruction( - HloInstruction::CreateParameter(0, r0f32, "param0")); - HloInstruction* min_value = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(0.0f))); - HloInstruction* max_value = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0f))); - HloInstruction* max = builder.AddInstruction(HloInstruction::CreateBinary( - r0f32, HloOpcode::kMaximum, param0, max_value)); - builder.AddInstruction( - HloInstruction::CreateBinary(r0f32, HloOpcode::kMinimum, max, min_value)); - - auto module = CreateNewModule(); - auto computation = module->AddEntryComputation(builder.Build()); - - EXPECT_THAT(computation->root_instruction(), - op::Minimum(op::Maximum(param0, max_value), min_value)); - - AlgebraicSimplifier simplifier(/*is_layout_sensitive=*/false, - non_bitcasting_callback()); - ASSERT_TRUE(simplifier.Run(module).ValueOrDie()); - - EXPECT_THAT(computation->root_instruction(), - op::Clamp(max_value, param0, min_value)); -} - -// Test that min(max(A, x), y) is transformed to clamp(x, A, y) for -// broadcasted scalar values. -TEST_F(AlgebraicSimplifierTest, MinMaxWithBroadcastToClamp) { - Shape r0f32 = ShapeUtil::MakeShape(F32, {}); - Shape r1f32 = ShapeUtil::MakeShape(F32, {100}); - HloComputation::Builder builder(TestName()); - HloInstruction* param0 = builder.AddInstruction( - HloInstruction::CreateParameter(0, r1f32, "param0")); - HloInstruction* min_value = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(0.0f))); - HloInstruction* max_value = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0f))); - HloInstruction* max = builder.AddInstruction(HloInstruction::CreateBinary( - r1f32, HloOpcode::kMaximum, param0, max_value)); - builder.AddInstruction( - HloInstruction::CreateBinary(r1f32, HloOpcode::kMinimum, max, min_value)); - - auto module = CreateNewModule(); - auto computation = module->AddEntryComputation(builder.Build()); - - EXPECT_THAT(computation->root_instruction(), - op::Minimum(op::Maximum(param0, max_value), min_value)); - - AlgebraicSimplifier simplifier(/*is_layout_sensitive=*/false, - non_bitcasting_callback()); - ASSERT_TRUE(simplifier.Run(module).ValueOrDie()); - - EXPECT_THAT(computation->root_instruction(), - op::Clamp(max_value, param0, min_value)); -} - -// Test that min(max(A, non-constant1), non-constant2) is not canonicalized to -// clamp(non-constant1, A, non-constant2) -TEST_F(AlgebraicSimplifierTest, MinMaxNotToClamp) { - Shape r0f32 = ShapeUtil::MakeShape(F32, {}); - HloComputation::Builder builder(TestName()); - HloInstruction* param0 = builder.AddInstruction( - HloInstruction::CreateParameter(0, r0f32, "param0")); - HloInstruction* min_value = builder.AddInstruction( - HloInstruction::CreateParameter(1, r0f32, "param1")); - HloInstruction* max_value = builder.AddInstruction( - HloInstruction::CreateParameter(2, r0f32, "param2")); - HloInstruction* max = builder.AddInstruction(HloInstruction::CreateBinary( - r0f32, HloOpcode::kMaximum, param0, max_value)); - builder.AddInstruction( - HloInstruction::CreateBinary(r0f32, HloOpcode::kMinimum, max, min_value)); - - auto module = CreateNewModule(); - auto computation = module->AddEntryComputation(builder.Build()); - - EXPECT_THAT(computation->root_instruction(), - op::Minimum(op::Maximum(param0, max_value), min_value)); - - AlgebraicSimplifier simplifier(/*is_layout_sensitive=*/false, - non_bitcasting_callback()); - EXPECT_FALSE(simplifier.Run(module).ValueOrDie()); - - EXPECT_THAT(computation->root_instruction(), - op::Minimum(op::Maximum(param0, max_value), min_value)); -} - -// Test that min(f(max(A, constant1)), constant2) is not transformed to -// clamp(constant1, A, constant2) -TEST_F(AlgebraicSimplifierTest, MinEquationWithMaxNotToClamp) { - Shape r0f32 = ShapeUtil::MakeShape(F32, {}); - HloComputation::Builder builder(TestName()); - HloInstruction* param0 = builder.AddInstruction( - HloInstruction::CreateParameter(0, r0f32, "param0")); - HloInstruction* min_value = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(0.0f))); - HloInstruction* max_value = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0f))); - HloInstruction* max = builder.AddInstruction(HloInstruction::CreateBinary( - r0f32, HloOpcode::kMaximum, param0, max_value)); - HloInstruction* fmax = builder.AddInstruction( - HloInstruction::CreateBinary(r0f32, HloOpcode::kAdd, max, max_value)); - builder.AddInstruction(HloInstruction::CreateBinary( - r0f32, HloOpcode::kMinimum, fmax, min_value)); - - auto module = CreateNewModule(); - auto computation = module->AddEntryComputation(builder.Build()); - - EXPECT_THAT(computation->root_instruction(), - op::Minimum(op::Add(op::Maximum(param0, max_value), max_value), - min_value)); - - AlgebraicSimplifier simplifier(/*is_layout_sensitive=*/false, - non_bitcasting_callback()); - EXPECT_FALSE(simplifier.Run(module).ValueOrDie()); - - EXPECT_THAT(computation->root_instruction(), - op::Minimum(op::Add(op::Maximum(param0, max_value), max_value), - min_value)); -} - // Test that slice(broadcast(/*scalar value*/)) simplifies to a single // broadcast. TEST_F(AlgebraicSimplifierTest, ScalarBroadcastToSlice) { @@ -2298,7 +2109,7 @@ TEST_F(AlgebraicSimplifierTest, ScalarBroadcastToSlice) { TEST_F(AlgebraicSimplifierTest, ScalarBroadcastToTransposeReshape) { HloComputation::Builder builder(TestName()); HloInstruction* forty_two = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(42.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(42.0f))); Shape broadcast_shape = ShapeUtil::MakeShape(F32, {4, 5, 6}); HloInstruction* broadcast = builder.AddInstruction( @@ -2345,7 +2156,7 @@ TEST_F(AlgebraicSimplifierTest, FoldPadIntoReduceWindow) { padding.mutable_dimensions(3)->set_edge_padding_high(2); HloInstruction* pad_value = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(5.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(5.0f))); HloInstruction* pad = builder.AddInstruction(HloInstruction::CreatePad( ShapeUtil::MakeShape(F32, {1, 3, 3, 5}), operand, pad_value, padding)); @@ -2376,7 +2187,7 @@ TEST_F(AlgebraicSimplifierTest, FoldPadIntoReduceWindow) { const Shape reduce_window_shape = ShapeUtil::MakeShape(F32, {111, 113, 113, 115}); HloInstruction* reduce_init_value = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(5.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(5.0f))); HloInstruction* reduce_window = builder.AddInstruction(HloInstruction::CreateReduceWindow( reduce_window_shape, pad, reduce_init_value, window, @@ -2427,7 +2238,7 @@ TEST_F(AlgebraicSimplifierTest, FoldConvertedPadIntoReduceWindow) { padding.mutable_dimensions(3)->set_edge_padding_high(2); HloInstruction* pad_value = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(5.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(5.0f))); HloInstruction* pad = builder.AddInstruction(HloInstruction::CreatePad( ShapeUtil::MakeShape(BF16, {1, 3, 3, 5}), parameter, pad_value, padding)); @@ -2462,7 +2273,7 @@ TEST_F(AlgebraicSimplifierTest, FoldConvertedPadIntoReduceWindow) { const Shape reduce_window_shape = ShapeUtil::MakeShape(F32, {111, 113, 113, 115}); HloInstruction* reduce_init_value = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(5.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(5.0f))); HloInstruction* reduce_window = builder.AddInstruction(HloInstruction::CreateReduceWindow( reduce_window_shape, convert, reduce_init_value, window, @@ -2533,9 +2344,9 @@ TEST_F(AlgebraicSimplifierTest, IteratorInvalidation) { HloComputation::Builder call_builder(TestName() + ".Call"); HloInstruction* zero = call_builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR1({0.0f}))); + HloInstruction::CreateConstant(LiteralUtil::CreateR1({0.0f}))); HloInstruction* one = call_builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR1({1.0f}))); + HloInstruction::CreateConstant(LiteralUtil::CreateR1({1.0f}))); call_builder.AddInstruction( HloInstruction::CreateCall(r1f32, {zero, one}, dot_computation.get())); @@ -2551,9 +2362,9 @@ TEST_F(AlgebraicSimplifierTest, ConstantTupleBecomesTupleOfConstants) { HloComputation::Builder builder(TestName()); const float constant_scalar = 7.3f; std::initializer_list constant_vector = {1.1f, 2.0f, 3.3f}; - std::unique_ptr value = - Literal::MakeTuple({Literal::CreateR0(constant_scalar).get(), - Literal::CreateR1(constant_vector).get()}); + std::unique_ptr value = LiteralUtil::MakeTuple( + {LiteralUtil::CreateR0(constant_scalar).get(), + LiteralUtil::CreateR1(constant_vector).get()}); builder.AddInstruction(HloInstruction::CreateConstant(std::move(value))); auto computation = module().AddEntryComputation(builder.Build()); @@ -2576,8 +2387,8 @@ TEST_F(AlgebraicSimplifierTest, TrivialDynamicSlice) { shape, builder.AddInstruction( HloInstruction::CreateParameter(0, shape, "slice_from")), - builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR1({0, 0, 0}))), + builder.AddInstruction(HloInstruction::CreateConstant( + LiteralUtil::CreateR1({0, 0, 0}))), /*slice_sizes=*/{10, 100, 1000})); auto computation = module().AddEntryComputation(builder.Build()); @@ -2610,8 +2421,8 @@ TEST_F(AlgebraicSimplifierTest, TrivialDynamicUpdateSlice) { builder.AddInstruction( HloInstruction::CreateParameter(2, slice_shape, "to_update")), slice, - builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR1({0, 0, 0}))))); + builder.AddInstruction(HloInstruction::CreateConstant( + LiteralUtil::CreateR1({0, 0, 0}))))); auto computation = module().AddEntryComputation(builder.Build()); AlgebraicSimplifier simplifier(/*is_layout_sensitive=*/false, @@ -2626,7 +2437,7 @@ TEST_F(AlgebraicSimplifierTest, MergeBroadcasts) { HloComputation::Builder builder(TestName()); Shape r2f32 = ShapeUtil::MakeShape(F32, {2, 2}); HloInstruction* input_array = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR1({3, 4}))); + HloInstruction::CreateConstant(LiteralUtil::CreateR1({3, 4}))); HloInstruction* inner_bcast = builder.AddInstruction( HloInstruction::CreateBroadcast(r2f32, input_array, {1})); Shape r3f32 = ShapeUtil::MakeShape(F32, {2, 2, 2}); @@ -2735,7 +2546,7 @@ TEST_P(PadReduceWindowEffectiveBroadcastTest, DoIt) { HloInstruction* pad = builder.AddInstruction(HloInstruction::CreatePad( pad_shape, input, builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(0.0f))), + HloInstruction::CreateConstant(LiteralUtil::CreateR0(0.0f))), padding)); HloComputation* add_computation = nullptr; @@ -2754,7 +2565,7 @@ TEST_P(PadReduceWindowEffectiveBroadcastTest, DoIt) { Window window = window_util::MakeWindow( decorate_spatials(param.reduce_window_spatials, 1, 1)); auto zero = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(0.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(0.0f))); TF_ASSERT_OK_AND_ASSIGN(const Shape output_shape, ShapeInference::InferReduceWindowShape( pad->shape(), zero->shape(), window, @@ -2893,7 +2704,7 @@ TEST_P(DotOfConcatSimplificationTest, ConstantLHS) { Shape lhs_shape = ShapeUtil::MakeShape(F32, {spec.m, spec.k}); auto* lhs = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR2F32Linspace( + HloInstruction::CreateConstant(LiteralUtil::CreateR2F32Linspace( /*from=*/10.0, /*to=*/10000.0, /*rows=*/spec.m, /*cols=*/spec.k))); Shape rhs0_shape = ShapeUtil::MakeShape(F32, {k0, spec.n}); @@ -2972,7 +2783,7 @@ TEST_P(DotOfConcatSimplificationTest, ConstantRHS) { Shape rhs_shape = ShapeUtil::MakeShape(F32, {spec.k, spec.n}); auto* rhs = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR2F32Linspace( + HloInstruction::CreateConstant(LiteralUtil::CreateR2F32Linspace( /*from=*/10.0, /*to=*/10000.0, /*rows=*/spec.k, /*cols=*/spec.n))); DotDimensionNumbers dot_dnums; @@ -3019,7 +2830,7 @@ TEST_F(AlgebraicSimplifierTest, DynamicUpdateSliceZeroUpdate) { HloInstruction* const update = builder.AddInstruction( HloInstruction::CreateParameter(1, update_shape, "update")); HloInstruction* const start_indices = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR1({0}))); + HloInstruction::CreateConstant(LiteralUtil::CreateR1({0}))); builder.AddInstruction(HloInstruction::CreateDynamicUpdateSlice( dslice_shape, operand, update, start_indices)); const HloComputation* const computation = @@ -3068,7 +2879,7 @@ TEST_P(DotOfGatherSimplificationTest, ConstantRHS) { int64 lhs_cols = (spec.lcd == 0) ? spec.m : (spec.k + k_increase); Shape lhs_shape = ShapeUtil::MakeShape(F32, {lhs_rows, lhs_cols}); auto* lhs = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR2F32Linspace( + HloInstruction::CreateConstant(LiteralUtil::CreateR2F32Linspace( /*from=*/10.0, /*to=*/10000.0, /*rows=*/lhs_rows, /*cols=*/lhs_cols))); @@ -3076,7 +2887,7 @@ TEST_P(DotOfGatherSimplificationTest, ConstantRHS) { int32 start_col = (spec.lcd == 0) ? spec.s : 0; const auto start_indices = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR1({start_row, start_col}))); + LiteralUtil::CreateR1({start_row, start_col}))); int64 slice_row_size = (spec.lcd == 0) ? spec.k : 1; int64 slice_col_size = (spec.lcd == 0) ? 1 : spec.k; Shape ds_shape = ShapeUtil::MakeShape(F32, {slice_row_size, slice_col_size}); @@ -3087,7 +2898,7 @@ TEST_P(DotOfGatherSimplificationTest, ConstantRHS) { int64 rhs_cols = (spec.rcd == 0) ? spec.n : spec.k; Shape rhs_shape = ShapeUtil::MakeShape(F32, {rhs_rows, rhs_cols}); auto* rhs = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR2F32Linspace( + HloInstruction::CreateConstant(LiteralUtil::CreateR2F32Linspace( /*from=*/10.0, /*to=*/10000.0, /*rows=*/rhs_rows, /*cols=*/rhs_cols))); @@ -3135,7 +2946,7 @@ TEST_P(DotOfGatherSimplificationTest, ConstantLHS) { int64 lhs_cols = (spec.lcd == 0) ? spec.m : spec.k; Shape lhs_shape = ShapeUtil::MakeShape(F32, {lhs_rows, lhs_cols}); auto* lhs = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR2F32Linspace( + HloInstruction::CreateConstant(LiteralUtil::CreateR2F32Linspace( /*from=*/10.0, /*to=*/10000.0, /*rows=*/lhs_rows, /*cols=*/lhs_cols))); @@ -3146,7 +2957,7 @@ TEST_P(DotOfGatherSimplificationTest, ConstantLHS) { int64 rhs_cols = (spec.rcd == 0) ? spec.n : (spec.k + k_increase); Shape rhs_shape = ShapeUtil::MakeShape(F32, {rhs_rows, rhs_cols}); auto* rhs = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR2F32Linspace( + HloInstruction::CreateConstant(LiteralUtil::CreateR2F32Linspace( /*from=*/10.0, /*to=*/10000.0, /*rows=*/rhs_rows, /*cols=*/rhs_cols))); @@ -3154,7 +2965,7 @@ TEST_P(DotOfGatherSimplificationTest, ConstantLHS) { int32 start_col = (spec.rcd == 0) ? spec.s : 0; const auto start_indices = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR1({start_row, start_col}))); + LiteralUtil::CreateR1({start_row, start_col}))); int64 slice_row_size = (spec.rcd == 0) ? spec.k : 1; int64 slice_col_size = (spec.rcd == 0) ? 1 : spec.k; Shape ds_shape = ShapeUtil::MakeShape(F32, {slice_row_size, slice_col_size}); diff --git a/tensorflow/compiler/xla/service/batchnorm_expander.cc b/tensorflow/compiler/xla/service/batchnorm_expander.cc index ec13fadbc75e2315d1d6ef72e24a0faca0c7de40..c4cd60c1201f7ddbf0aba4b6d587952531b74bfa 100644 --- a/tensorflow/compiler/xla/service/batchnorm_expander.cc +++ b/tensorflow/compiler/xla/service/batchnorm_expander.cc @@ -20,6 +20,7 @@ limitations under the License. #include #include +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/literal_util.h" #include "tensorflow/compiler/xla/service/dfs_hlo_visitor_with_default.h" #include "tensorflow/compiler/xla/service/hlo_computation.h" @@ -34,6 +35,7 @@ limitations under the License. #include "tensorflow/core/lib/core/status.h" #include "tensorflow/core/lib/gtl/array_slice.h" #include "tensorflow/core/lib/gtl/flatmap.h" +#include "tensorflow/core/lib/gtl/optional.h" #include "tensorflow/core/platform/logging.h" #include "tensorflow/core/platform/types.h" @@ -41,6 +43,8 @@ namespace xla { namespace { +using tensorflow::gtl::optional; + // BatchNormExpanderVisitor traverses the HLO computation and rewrites BatchNorm // operations into smaller operations. class BatchNormExpanderVisitor : public DfsHloVisitorWithDefault { @@ -97,7 +101,7 @@ class BatchNormExpanderVisitor : public DfsHloVisitorWithDefault { add_instruction(HloInstruction::CreateConvert( ShapeUtil::MakeShape(operand->shape().element_type(), {}), add_instruction(HloInstruction::CreateConstant( - Literal::CreateR0(-0.5f))))), + LiteralUtil::CreateR0(-0.5f))))), {})); return HloInstruction::CreateBinary(operand->shape(), HloOpcode::kPower, operand, exponent); @@ -113,7 +117,7 @@ class BatchNormExpanderVisitor : public DfsHloVisitorWithDefault { add_instruction(HloInstruction::CreateConvert( ShapeUtil::MakeShape(operand->shape().element_type(), {}), add_instruction(HloInstruction::CreateConstant( - Literal::CreateR0(1.0 / element_count))))), + LiteralUtil::CreateR0(1.0 / element_count))))), {})); return HloInstruction::CreateBinary(operand->shape(), HloOpcode::kMultiply, operand, elem_count_recip); @@ -200,11 +204,11 @@ Status BatchNormExpanderVisitor::HandleBatchNormTraining( HloInstruction* offset = batch_norm->mutable_operand(2); const Shape feature_shape = scale->shape(); - auto zero_literal = Literal::CreateR0(0.0f); + auto zero_literal = LiteralUtil::CreateR0(0.0f); TF_ASSIGN_OR_RETURN(zero_literal, zero_literal->Convert(ptype)); auto zero = add(HloInstruction::CreateConstant(std::move(zero_literal))); - auto epsilon_literal = Literal::CreateR0(batch_norm->epsilon()); + auto epsilon_literal = LiteralUtil::CreateR0(batch_norm->epsilon()); TF_ASSIGN_OR_RETURN(epsilon_literal, epsilon_literal->Convert(ptype)); auto epsilon = add(HloInstruction::CreateBroadcast( operand_shape, @@ -288,16 +292,22 @@ Status BatchNormExpanderVisitor::HandleBatchNormTraining( int64 instruction_count_after = computation_->instruction_count(); CHECK_EQ(instruction_count_after, instruction_count_before + added_instructions.size()); + const HloSharding& sharding = batch_norm->sharding(); HloSharding operand_sharding = - batch_norm->sharding().GetAsShapeTree(batch_norm->shape()).element({0}); + sharding.GetAsShapeTree(batch_norm->shape()).element({0}); + optional unique_device = batch_norm->sharding_unique_device(); + HloSharding default_sharding = + unique_device.has_value() + ? HloSharding::AssignDevice(unique_device.value()) + : HloSharding::Replicate(); for (HloInstruction* inst : added_instructions) { if (ShapeUtil::Equal(inst->shape(), operand_shape)) { inst->set_sharding(operand_sharding); } else { - inst->set_sharding(HloSharding::Replicate()); + inst->set_sharding(default_sharding); } } - tuple->set_sharding(batch_norm->sharding()); + tuple->set_sharding(sharding); } TF_CHECK_OK(ReplaceWithNewInstruction(batch_norm, std::move(tuple))); return Status::OK(); @@ -320,7 +330,7 @@ Status BatchNormExpanderVisitor::HandleBatchNormInference( HloInstruction* var = batch_norm->mutable_operand(4); const Shape feature_shape = scale->shape(); - auto epsilon_literal = Literal::CreateR0(batch_norm->epsilon()); + auto epsilon_literal = LiteralUtil::CreateR0(batch_norm->epsilon()); TF_ASSIGN_OR_RETURN(epsilon_literal, epsilon_literal->Convert(ptype)); auto epsilon = computation_->AddInstruction(HloInstruction::CreateBroadcast( operand_shape, @@ -388,14 +398,20 @@ Status BatchNormExpanderVisitor::HandleBatchNormInference( CHECK_EQ(instruction_count_after, instruction_count_before + added_instructions.size()); if (batch_norm->has_sharding()) { + const HloSharding& sharding = batch_norm->sharding(); + optional unique_device = batch_norm->sharding_unique_device(); + HloSharding default_sharding = + unique_device.has_value() + ? HloSharding::AssignDevice(unique_device.value()) + : HloSharding::Replicate(); for (HloInstruction* inst : added_instructions) { if (ShapeUtil::Equal(inst->shape(), operand_shape)) { - inst->set_sharding(batch_norm->sharding()); + inst->set_sharding(sharding); } else { - inst->set_sharding(HloSharding::Replicate()); + inst->set_sharding(default_sharding); } } - shifted_normalized->set_sharding(batch_norm->sharding()); + shifted_normalized->set_sharding(sharding); } TF_CHECK_OK( ReplaceWithNewInstruction(batch_norm, std::move(shifted_normalized))); @@ -447,11 +463,11 @@ Status BatchNormExpanderVisitor::HandleBatchNormGrad( const int64 feature_count = activation_shape.dimensions(feature_index); const int64 elements_per_feature_int64 = size_in_elements / feature_count; - auto zero_literal = Literal::CreateR0(0.0f); + auto zero_literal = LiteralUtil::CreateR0(0.0f); TF_ASSIGN_OR_RETURN(zero_literal, zero_literal->Convert(ptype)); auto zero = add(HloInstruction::CreateConstant(std::move(zero_literal))); - auto epsilon_literal = Literal::CreateR0(batch_norm->epsilon()); + auto epsilon_literal = LiteralUtil::CreateR0(batch_norm->epsilon()); TF_ASSIGN_OR_RETURN(epsilon_literal, epsilon_literal->Convert(ptype)); auto epsilon_scalar = add(HloInstruction::CreateConstant(std::move(epsilon_literal))); @@ -542,7 +558,7 @@ Status BatchNormExpanderVisitor::HandleBatchNormGrad( Mean(elements_per_feature_int64, scale_times_rsqrt_var_add_epsilon, add)); auto elements_per_feature_literal = - Literal::CreateR0(elements_per_feature_int64); + LiteralUtil::CreateR0(elements_per_feature_int64); TF_ASSIGN_OR_RETURN(elements_per_feature_literal, elements_per_feature_literal->Convert(ptype)); auto elements_per_feature = add( @@ -562,19 +578,25 @@ Status BatchNormExpanderVisitor::HandleBatchNormGrad( auto tuple = HloInstruction::CreateTuple({grad_activation, grad_scale, grad_beta}); if (batch_norm->has_sharding()) { + const HloSharding& sharding = batch_norm->sharding(); int64 instruction_count_after = computation_->instruction_count(); CHECK_EQ(instruction_count_after, instruction_count_before + added_instructions.size()); HloSharding activation_sharding = - batch_norm->sharding().GetAsShapeTree(batch_norm->shape()).element({0}); + sharding.GetAsShapeTree(batch_norm->shape()).element({0}); + auto unique_device = batch_norm->sharding_unique_device(); + HloSharding default_sharding = + unique_device.has_value() + ? HloSharding::AssignDevice(unique_device.value()) + : HloSharding::Replicate(); for (HloInstruction* inst : added_instructions) { if (ShapeUtil::Equal(inst->shape(), activation_shape)) { inst->set_sharding(activation_sharding); } else { - inst->set_sharding(HloSharding::Replicate()); + inst->set_sharding(default_sharding); } } - tuple->set_sharding(batch_norm->sharding()); + tuple->set_sharding(sharding); } TF_CHECK_OK(ReplaceWithNewInstruction(batch_norm, std::move(tuple))); diff --git a/tensorflow/compiler/xla/service/batchnorm_expander_test.cc b/tensorflow/compiler/xla/service/batchnorm_expander_test.cc index aa36e64b07099a372dab67babc7a18a2d39596bc..32f785a70adf0e7ea3ce281f7ff73224be8d424e 100644 --- a/tensorflow/compiler/xla/service/batchnorm_expander_test.cc +++ b/tensorflow/compiler/xla/service/batchnorm_expander_test.cc @@ -19,12 +19,13 @@ limitations under the License. #include #include "tensorflow/compiler/xla/layout_util.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/ptr_util.h" #include "tensorflow/compiler/xla/service/hlo_computation.h" #include "tensorflow/compiler/xla/service/hlo_instruction.h" #include "tensorflow/compiler/xla/service/hlo_matchers.h" #include "tensorflow/compiler/xla/service/hlo_opcode.h" +#include "tensorflow/compiler/xla/service/hlo_parser.h" #include "tensorflow/compiler/xla/service/hlo_pass_fix.h" #include "tensorflow/compiler/xla/shape_util.h" #include "tensorflow/compiler/xla/test.h" @@ -114,5 +115,33 @@ TEST_F(BatchNormExpanderTest, BatchNormGrad) { EXPECT_EQ(root->opcode(), HloOpcode::kTuple); } +TEST_F(BatchNormExpanderTest, BatchNormTrainingSharding) { + const char* module_str = R"( +HloModule module +ENTRY entry { + %param.0 = f32[8,4] parameter(0) + %param.1 = f32[4] parameter(1) + %param.2 = f32[4] parameter(2) + ROOT %batch-norm-training = (f32[8,4], f32[4], f32[4]) + batch-norm-training(f32[8,4] %param.0, f32[4] %param.1, f32[4] %param.2), + epsilon=0.001, feature_index=1, sharding={maximal device=1} +})"; + + TF_ASSERT_OK_AND_ASSIGN(auto module, ParseHloString(module_str)); + BatchNormExpander rewriter(/*rewrite_training_op=*/true, + /*rewrite_inference_op=*/true, + /*rewrite_grad_op=*/true); + ASSERT_TRUE(rewriter.Run(module.get()).ValueOrDie()); + + for (auto* instruction : module->entry_computation()->instructions()) { + if (instruction->opcode() == HloOpcode::kParameter) { + continue; + } + ASSERT_TRUE(instruction->has_sharding()); + TF_ASSERT_OK_AND_ASSIGN(int device, instruction->sharding().UniqueDevice()); + EXPECT_EQ(device, 1); + } +} + } // namespace } // namespace xla diff --git a/tensorflow/compiler/xla/service/bfloat16_propagation.cc b/tensorflow/compiler/xla/service/bfloat16_propagation.cc index ee6b6f69b96216403c48933e424ebbfecd482eee..b21c83a07f69d6ec93cf9305802e4d3af2783bdc 100644 --- a/tensorflow/compiler/xla/service/bfloat16_propagation.cc +++ b/tensorflow/compiler/xla/service/bfloat16_propagation.cc @@ -15,7 +15,7 @@ limitations under the License. #include "tensorflow/compiler/xla/service/bfloat16_propagation.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/map_util.h" #include "tensorflow/compiler/xla/service/hlo_computation.h" #include "tensorflow/compiler/xla/service/hlo_dce.h" @@ -85,9 +85,9 @@ void BFloat16Propagation::RevertIfFusionInternalBF16Changes( auto root_changes_it = changes_to_bf16_.find(root); if (root_changes_it != changes_to_bf16_.end()) { - for (const auto& index : root_changes_it->second) { + for (const auto& entry : root_changes_it->second) { for (const HloValue* value : - dataflow_->GetValueSet(root, index).values()) { + dataflow_->GetValueSet(root, entry.second).values()) { changed_root_buffers.insert(value); } } @@ -615,7 +615,6 @@ Status BFloat16Propagation::ResolveInconsistentFusions(HloModule* module) { // (1) a is F32 but tuple is BF16 // (2) after adding conversion // (3) after tuple simplifier and DCE. - bool needs_tuple_simplifier = false; for (auto computation : module->MakeComputationPostOrder()) { auto insts = computation->MakeInstructionPostOrder(); for (auto inst_it = insts.rbegin(); inst_it != insts.rend(); ++inst_it) { @@ -629,67 +628,25 @@ Status BFloat16Propagation::ResolveInconsistentFusions(HloModule* module) { continue; } ShapeTree converted_outputs(hlo->shape()); - // Iterate through nodes in the shape tree in pre-order and initialize - // each non-root node with a corresponding get-tuple-element. For a leaf - // node, if its shape does not match the fusion output, create a - // conversion node to overwrite the node value. - for (auto it = converted_outputs.begin(); it != converted_outputs.end(); - ++it) { - ShapeIndex output_index = it->first; - HloInstruction*& output = it->second; - const Shape subshape = - ShapeUtil::GetSubshape(hlo->shape(), output_index); - if (output_index.empty()) { - output = fusion_root; - } else { - ShapeIndex parent_index = output_index; - parent_index.pop_back(); - output = fusion_computation->AddInstruction( - HloInstruction::CreateGetTupleElement( - subshape, converted_outputs.element(parent_index), - output_index.back())); - } - if (!ShapeUtil::IsArray(subshape)) { - continue; - } - if (!ShapeUtil::Compatible( - subshape, - ShapeUtil::GetSubshape(fusion_root->shape(), output_index))) { - output = fusion_computation->AddInstruction( - HloInstruction::CreateConvert(subshape, output)); - } - } - // Iterate through nodes in the shape tree in reverse pre-order and create - // a tuple instruction for each non-leaf node where the elements are the - // values of its child nodes. - for (auto it = converted_outputs.rbegin(); it != converted_outputs.rend(); - ++it) { - ShapeIndex output_index = it->first; - HloInstruction*& output = it->second; - const Shape& subshape = - ShapeUtil::GetSubshape(hlo->shape(), output_index); - if (!ShapeUtil::IsTuple(subshape)) { - continue; - } - std::vector elements( - ShapeUtil::TupleElementCount(subshape)); - ShapeIndex child_index = output_index; - for (int64 i = 0; i < elements.size(); ++i) { - child_index.push_back(i); - elements[i] = converted_outputs.element(child_index); - child_index.pop_back(); - } - output = fusion_computation->AddInstruction( - HloInstruction::CreateTuple(elements)); - } - fusion_computation->set_root_instruction(converted_outputs.element({})); - needs_tuple_simplifier |= ShapeUtil::IsTuple(hlo->shape()); + // Deep copy the fusion root, and convert a leaf node only if its shape + // does not match the fusion output. + TF_ASSIGN_OR_RETURN( + HloInstruction * copy, + fusion_computation->DeepCopyInstructionWithCustomCopier( + fusion_root, + [hlo](HloInstruction* leaf, const ShapeIndex& leaf_index, + HloComputation* comp) { + const Shape& hlo_subshape = + ShapeUtil::GetSubshape(hlo->shape(), leaf_index); + if (ShapeUtil::Compatible(leaf->shape(), hlo_subshape)) { + return leaf; + } + return comp->AddInstruction( + HloInstruction::CreateConvert(hlo_subshape, leaf)); + })); + fusion_computation->set_root_instruction(copy); } } - if (needs_tuple_simplifier) { - TupleSimplifier tuple_simplifier; - TF_RETURN_IF_ERROR(tuple_simplifier.Run(module).status()); - } return Status::OK(); } @@ -758,10 +715,38 @@ StatusOr BFloat16Propagation::Run(HloModule* module) { changes_to_bf16_.clear(); changed_ = false; + auto computations_topological_order = module->MakeComputationPostOrder(); + + // Before running the propagation pass, we insert copies (kConvert to the same + // type) of F32 inputs to while loops. This prevents other uses of the same + // input from aliasing the while loop input/output, so that there's greater + // chance to use BF16 inside the loop. If some of these added copies do not + // help, they will remain F32 after BF16 propagation and will be removed since + // they are no-ops. + for (auto computation : computations_topological_order) { + for (auto inst : computation->MakeInstructionPostOrder()) { + if (inst->opcode() != HloOpcode::kWhile) { + continue; + } + + auto operand = inst->mutable_operand(0); + TF_ASSIGN_OR_RETURN( + HloInstruction * copy, + computation->DeepCopyInstructionWithCustomCopier( + operand, [](HloInstruction* leaf, const ShapeIndex& leaf_index, + HloComputation* comp) { + if (leaf->shape().element_type() != F32) { + return leaf; + } + return comp->AddInstruction( + HloInstruction::CreateConvert(leaf->shape(), leaf)); + })); + TF_RETURN_IF_ERROR(operand->ReplaceUseWith(inst, copy)); + } + } + TF_ASSIGN_OR_RETURN(dataflow_, HloDataflowAnalysis::Run(*module)); - const auto& computations_topological_order = - module->MakeComputationPostOrder(); // The first step is a forward pass (parameters to root), where we determine // the potential candidate instructions to use bfloat16 in the outputs that // are not likely to cause overhead from extra explicit conversions. This is @@ -802,39 +787,42 @@ StatusOr BFloat16Propagation::Run(HloModule* module) { // Apply the changes in changes_to_bf16_. for (auto& change : changes_to_bf16_) { - auto shape = change.first->mutable_shape(); - for (const auto& index : change.second) { - auto subshape = ShapeUtil::GetMutableSubshape(shape, index); + for (const auto& entry : change.second) { + auto subshape = entry.first; CHECK_EQ(subshape->element_type(), F32); subshape->set_element_type(BF16); changed_ = true; } } + // Removes redundant HLOs added by this pass, either when inserting + // de-aliasing copies to while loop inputs, or later when converting output + // types. + auto clean_up = [this, module]() { + TF_RETURN_IF_ERROR(SkipNoopConversions(module)); + TupleSimplifier tuple_simplifier; + TF_RETURN_IF_ERROR(tuple_simplifier.Run(module).status()); + HloDCE dce; + TF_RETURN_IF_ERROR(dce.Run(module).status()); + return Status::OK(); + }; + if (!changed_) { + TF_RETURN_IF_ERROR(clean_up()); return false; } TF_RETURN_IF_ERROR(ResolveInconsistentFusions(module)); TF_RETURN_IF_ERROR(ResolveConvertedConstants(module)); - // This pass could have turned an F32 -> BF16 conversion to a no-op (BF16 -> - // BF16), so we skip them now. - TF_RETURN_IF_ERROR(SkipNoopConversions(module)); - - { - // We may have dead HLOs after ResolveInconsistentFusions, - // ResolveConvertedConstants and SkipNoopConversions. - HloDCE dce; - TF_RETURN_IF_ERROR(dce.Run(module).status()); - } + TF_RETURN_IF_ERROR(clean_up()); return true; } PrimitiveType BFloat16Propagation::OutputTypeAfterChange( HloInstruction* hlo, const ShapeIndex& index) const { - PrimitiveType type_on_hlo = - ShapeUtil::GetSubshape(hlo->shape(), index).element_type(); + Shape* subshape = ShapeUtil::GetMutableSubshape(hlo->mutable_shape(), index); + const PrimitiveType type_on_hlo = subshape->element_type(); if (type_on_hlo != F32) { return type_on_hlo; } @@ -842,7 +830,7 @@ PrimitiveType BFloat16Propagation::OutputTypeAfterChange( if (it == changes_to_bf16_.end()) { return type_on_hlo; } - return ContainsKey(it->second, index) ? BF16 : F32; + return ContainsKey(it->second, subshape) ? BF16 : F32; } PrimitiveType BFloat16Propagation::ValueTypeAfterChange( @@ -856,14 +844,16 @@ void BFloat16Propagation::AddToOrRemoveFromBF16ChangeSet( HloInstruction* hlo, const ShapeIndex& index, PrimitiveType target_type) { if (target_type == BF16) { auto& entry = changes_to_bf16_[hlo]; - entry.insert(index); + entry.emplace(ShapeUtil::GetMutableSubshape(hlo->mutable_shape(), index), + index); } else { CHECK_EQ(target_type, F32); auto it = changes_to_bf16_.find(hlo); if (it == changes_to_bf16_.end()) { return; } - it->second.erase(index); + it->second.erase( + ShapeUtil::GetMutableSubshape(hlo->mutable_shape(), index)); } } diff --git a/tensorflow/compiler/xla/service/bfloat16_propagation.h b/tensorflow/compiler/xla/service/bfloat16_propagation.h index de0355ddfca127753f90d1899b424a8e77c9b291..02b8cad089dd8465b7af5c1014e37b77ded6949d 100644 --- a/tensorflow/compiler/xla/service/bfloat16_propagation.h +++ b/tensorflow/compiler/xla/service/bfloat16_propagation.h @@ -194,17 +194,11 @@ class BFloat16Propagation : public HloPassInterface { // are subject to further adjustment, then finally applied to the HLOs. This // avoids setting changed_ to true but all changes are reverted during // adjustment. - struct IndexHasher { - int64 operator()(const ShapeIndex& index) const { - int64 hash = 0; - for (int64 i : index) { - hash = tensorflow::Hash64Combine(hash, std::hash()(i)); - } - return hash; - } - }; + // + // For each HloInstruction, changes_to_bf16_ stores the affected buffers in + // the output as a map from in-place pointers to subshapes to shape indices. tensorflow::gtl::FlatMap> + tensorflow::gtl::FlatMap> changes_to_bf16_; // Whether the last processed HLO module has been changed by this pass. diff --git a/tensorflow/compiler/xla/service/bfloat16_propagation_test.cc b/tensorflow/compiler/xla/service/bfloat16_propagation_test.cc index e2ca689c0649528231c0581a37c145c328652420..aeafb25ad7215ea3d297e4a8bf7e1ba72d33d528 100644 --- a/tensorflow/compiler/xla/service/bfloat16_propagation_test.cc +++ b/tensorflow/compiler/xla/service/bfloat16_propagation_test.cc @@ -133,9 +133,9 @@ TEST_F(BFloat16PropagationTest, ConvertConstantLiteral) { array_b.FillUnique(10.0f); HloInstruction* a = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateFromArray(array_a))); + HloInstruction::CreateConstant(LiteralUtil::CreateFromArray(array_a))); HloInstruction* b = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateFromArray(array_b))); + HloInstruction::CreateConstant(LiteralUtil::CreateFromArray(array_b))); HloInstruction* dot = builder.AddInstruction( HloInstruction::CreateBinary(shape, HloOpcode::kDot, a, b)); @@ -150,10 +150,10 @@ TEST_F(BFloat16PropagationTest, ConvertConstantLiteral) { EXPECT_EQ(dot->operand(0)->opcode(), HloOpcode::kConstant); EXPECT_EQ(dot->operand(1)->opcode(), HloOpcode::kConstant); EXPECT_TRUE(LiteralTestUtil::Equal( - *Literal::ConvertF32ToBF16(*Literal::CreateFromArray(array_a)), + *LiteralUtil::ConvertF32ToBF16(*LiteralUtil::CreateFromArray(array_a)), dot->operand(0)->literal())); EXPECT_TRUE(LiteralTestUtil::Equal( - *Literal::ConvertF32ToBF16(*Literal::CreateFromArray(array_b)), + *LiteralUtil::ConvertF32ToBF16(*LiteralUtil::CreateFromArray(array_b)), dot->operand(1)->literal())); } @@ -240,12 +240,10 @@ TEST_F(BFloat16PropagationTest, SameValueReferencedTwice) { EXPECT_TRUE(PropagatePrecision(module.get())); EXPECT_EQ(computation->root_instruction(), dot); - EXPECT_TRUE(OutputsBF16(add0)); EXPECT_TRUE(OutputsBF16(add1)); EXPECT_TRUE(OutputsBF16(lhs)); - // rhs is a get-tuple-element, which does not define a buffer, but its shape - // should also be adjusted accordingly. - EXPECT_TRUE(OutputsBF16(rhs)); + + // add0 and rhs have been eliminated by simplification and DCE. } // Tests that a non-fusion computation's root should not be changed. @@ -434,7 +432,7 @@ TEST_F(BFloat16PropagationTest, SelectOverTuples) { HloInstruction* tuple1 = builder.AddInstruction(HloInstruction::CreateTuple({param, add1})); HloInstruction* sel = builder.AddInstruction(HloInstruction::CreateTernary( - tuple0->shape(), HloOpcode::kSelect, pred, tuple0, tuple1)); + tuple0->shape(), HloOpcode::kTupleSelect, pred, tuple0, tuple1)); HloInstruction* gte0 = builder.AddInstruction( HloInstruction::CreateGetTupleElement(shape, sel, 0)); HloInstruction* gte1 = builder.AddInstruction( @@ -734,10 +732,8 @@ TEST_F(BFloat16PropagationTest, NoopConversionRemoved) { EXPECT_TRUE(PropagatePrecision(module.get())); EXPECT_EQ(computation->root_instruction(), add2); - EXPECT_EQ(add2->operand(0), gte0); - EXPECT_EQ(add2->operand(1), gte1); - EXPECT_EQ(gte0->shape().element_type(), BF16); - EXPECT_EQ(gte1->shape().element_type(), BF16); + EXPECT_EQ(add2->operand(0), add0); + EXPECT_EQ(add2->operand(1), add1); EXPECT_EQ(add0->shape().element_type(), BF16); EXPECT_EQ(add1->shape().element_type(), BF16); } @@ -771,8 +767,14 @@ TEST_F(BFloat16PropagationTest, TupleDomain) { auto computation = module->AddEntryComputation(builder.Build()); EXPECT_TRUE(PropagatePrecision(module.get())); - EXPECT_EQ(computation->root_instruction(), root); + + // test BF16 propagated through domain + EXPECT_EQ(ShapeUtil::GetTupleElementShape(domain->shape(), 0).element_type(), + BF16); + EXPECT_EQ(ShapeUtil::GetTupleElementShape(domain->shape(), 1).element_type(), + BF16); + EXPECT_TRUE(OutputsBF16(a_trans)); EXPECT_TRUE(OutputsBF16(b_trans)); EXPECT_TRUE(OutputsBF16(a_gte)); @@ -781,4 +783,44 @@ TEST_F(BFloat16PropagationTest, TupleDomain) { EXPECT_FALSE(OutputsBF16(b)); } +// Tests that bf16 is not propagated through a domain in case its input cannot +// be propagated. In the case below the input of the domain is the parameter +// tuple which cannot be propagated, so the domain instruction is not propagated +// either. +TEST_F(BFloat16PropagationTest, TupleDomainNoPropagation) { + auto builder = HloComputation::Builder(TestName()); + Shape shape = ShapeUtil::MakeShape(F32, {4, 4}); + Shape tuple_shape = ShapeUtil::MakeTupleShape({shape, shape}); + + HloInstruction* param = builder.AddInstruction( + HloInstruction::CreateParameter(0, tuple_shape, "param")); + HloInstruction* domain = builder.AddInstruction( + HloInstruction::CreateDomain(param->shape(), param, nullptr, nullptr)); + HloInstruction* a_gte = builder.AddInstruction( + HloInstruction::CreateGetTupleElement(shape, domain, 0)); + HloInstruction* b_gte = builder.AddInstruction( + HloInstruction::CreateGetTupleElement(shape, domain, 1)); + HloInstruction* a_trans = builder.AddInstruction( + HloInstruction::CreateTranspose(shape, a_gte, {0, 1})); + HloInstruction* b_trans = builder.AddInstruction( + HloInstruction::CreateTranspose(shape, b_gte, {0, 1})); + HloInstruction* dot = builder.AddInstruction( + HloInstruction::CreateBinary(shape, HloOpcode::kDot, a_trans, b_trans)); + HloInstruction* root = builder.AddInstruction( + HloInstruction::CreateBinary(shape, HloOpcode::kAdd, dot, dot)); + + auto module = CreateNewModule(); + auto computation = module->AddEntryComputation(builder.Build()); + + EXPECT_TRUE(PropagatePrecision(module.get())); + + EXPECT_EQ(computation->root_instruction(), root); + EXPECT_TRUE(OutputsBF16(a_trans)); + EXPECT_TRUE(OutputsBF16(b_trans)); + EXPECT_FALSE(OutputsBF16(a_gte)); + EXPECT_FALSE(OutputsBF16(b_gte)); + EXPECT_FALSE(OutputsBF16(domain)); + EXPECT_FALSE(OutputsBF16(param)); +} + } // namespace xla diff --git a/tensorflow/compiler/xla/service/bfloat16_support.cc b/tensorflow/compiler/xla/service/bfloat16_support.cc index 8595afca7e735528d9ef29a323696c0661fe971c..23645346e6f491beb5171cc839c013ce5f83d789 100644 --- a/tensorflow/compiler/xla/service/bfloat16_support.cc +++ b/tensorflow/compiler/xla/service/bfloat16_support.cc @@ -103,6 +103,7 @@ bool BFloat16Support::EffectiveOperandPrecisionIsOutputPrecision( case HloOpcode::kDynamicUpdateSlice: return operand_index == 0 || operand_index == 1; case HloOpcode::kSelect: + case HloOpcode::kTupleSelect: return operand_index == 1 || operand_index == 2; default: break; diff --git a/tensorflow/compiler/xla/service/buffer_assignment_test.cc b/tensorflow/compiler/xla/service/buffer_assignment_test.cc index efa4696130ffeff669b0d674438a45c5a9d48ef2..125ade2a1194b1b24b9557e0974d2553f1adff5c 100644 --- a/tensorflow/compiler/xla/service/buffer_assignment_test.cc +++ b/tensorflow/compiler/xla/service/buffer_assignment_test.cc @@ -21,7 +21,7 @@ limitations under the License. #include #include -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/ptr_util.h" #include "tensorflow/compiler/xla/service/buffer_value.h" #include "tensorflow/compiler/xla/service/call_graph.h" @@ -125,7 +125,7 @@ class BufferAssignmentTest : public HloTestBase { auto param = builder.AddInstruction(HloInstruction::CreateParameter(0, r0f32_, "x")); auto value = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0))); builder.AddInstruction( HloInstruction::CreateBinary(r0f32_, HloOpcode::kAdd, param, value)); return builder.Build(); @@ -142,7 +142,7 @@ class BufferAssignmentTest : public HloTestBase { const string& name) { auto builder = HloComputation::Builder(name); auto const4 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(4))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(4))); auto param = builder.AddInstruction( HloInstruction::CreateParameter(0, t_s32_f32v4_, "x")); auto index = builder.AddInstruction( @@ -167,9 +167,9 @@ class BufferAssignmentTest : public HloTestBase { const string& name) { auto builder = HloComputation::Builder(name); auto const1 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1))); auto constv = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR1({1.1f, 2.2f, 3.3f, 4.4f}))); + LiteralUtil::CreateR1({1.1f, 2.2f, 3.3f, 4.4f}))); auto param = builder.AddInstruction( HloInstruction::CreateParameter(0, t_s32_f32v4_, "x")); auto indexc = builder.AddInstruction( @@ -290,7 +290,7 @@ static bool BuffersDistinct(const std::vector& a, TEST_F(BufferAssignmentTest, ScalarConstant) { auto builder = HloComputation::Builder(TestName()); auto const0 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0))); auto module = CreateNewModule(); module->AddEntryComputation(builder.Build()); @@ -304,9 +304,9 @@ TEST_F(BufferAssignmentTest, BufferForConst) { // no buffers assigned, and their consumer has a buffer. auto builder = HloComputation::Builder(TestName()); auto const0 = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR1({1.1f, 2.2f, 3.3f, 4.4f}))); + LiteralUtil::CreateR1({1.1f, 2.2f, 3.3f, 4.4f}))); auto const1 = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR1({4.1f, 4.2f, 4.3f, 4.4f}))); + LiteralUtil::CreateR1({4.1f, 4.2f, 4.3f, 4.4f}))); auto add = builder.AddInstruction( HloInstruction::CreateBinary(f32vec4_, HloOpcode::kAdd, const0, const1)); auto module = CreateNewModule(); @@ -327,7 +327,7 @@ TEST_F(BufferAssignmentTest, HasAllocationAt) { auto param0 = builder.AddInstruction( HloInstruction::CreateParameter(0, f32vec100_, "param0")); auto constant = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1))); auto negate = builder.AddInstruction( HloInstruction::CreateUnary(f32vec100_, HloOpcode::kNegate, param0)); auto tuple = builder.AddInstruction( @@ -352,7 +352,7 @@ TEST_F(BufferAssignmentTest, BufferForOutputConst) { // This computation copies a constant to output. auto builder = HloComputation::Builder(TestName()); auto const0 = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR1({1.1f, 2.2f, 3.3f, 4.4f}))); + LiteralUtil::CreateR1({1.1f, 2.2f, 3.3f, 4.4f}))); auto copy = builder.AddInstruction( HloInstruction::CreateUnary(const0->shape(), HloOpcode::kCopy, const0)); auto module = CreateNewModule(); @@ -660,7 +660,7 @@ TEST_F(BufferAssignmentTest, CannotReuseInputBufferOfReduce) { auto exp2 = builder.AddInstruction( HloInstruction::CreateUnary(f32a100x10_, HloOpcode::kExp, exp1)); auto const0 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(0.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(0.0f))); auto reduce = builder.AddInstruction(HloInstruction::CreateReduce( /*shape=*/f32vec10_, /*operand=*/exp2, @@ -708,9 +708,9 @@ TEST_F(BufferAssignmentTest, ExampleWhile) { // Creates the main kernel and verifies instruction counts. auto builder = HloComputation::Builder(TestName()); auto const3 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(0))); auto const4 = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR1({1.1f, 2.2f, 3.3f, 4.4f}))); + LiteralUtil::CreateR1({1.1f, 2.2f, 3.3f, 4.4f}))); auto tuple = builder.AddInstruction(HloInstruction::CreateTuple({const3, const4})); auto while_op = builder.AddInstruction(HloInstruction::CreateWhile( @@ -773,11 +773,11 @@ TEST_F(BufferAssignmentTest, ExampleConditional) { auto builder = HloComputation::Builder(TestName()); auto pred = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(false))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(false))); auto const1 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(56.4f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(56.4f))); auto const2 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(12.4f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(12.4f))); auto conditional = builder.AddInstruction(HloInstruction::CreateConditional( r0f32_, pred, const1, true_computation, const2, false_computation)); module->AddEntryComputation(builder.Build()); @@ -1200,8 +1200,9 @@ TEST_F(BufferAssignmentTest, DISABLED_TupleConstantAsOutput) { // Test that a tuple constant which is forwarded to the computation output // is properly handled. auto builder = HloComputation::Builder(TestName()); - builder.AddInstruction(HloInstruction::CreateConstant(Literal::MakeTuple( - {Literal::CreateR0(0).get(), Literal::CreateR0(1).get()}))); + builder.AddInstruction(HloInstruction::CreateConstant( + LiteralUtil::MakeTuple({LiteralUtil::CreateR0(0).get(), + LiteralUtil::CreateR0(1).get()}))); auto module = CreateNewModule(); module->AddEntryComputation(builder.Build()); @@ -1365,8 +1366,9 @@ TEST_F(BufferAssignmentTest, AmbiguousBufferAsOutput) { HloInstruction::CreateParameter(1, tuple_shape, "param1")); auto pred_param = builder.AddInstruction(HloInstruction::CreateParameter( 2, ShapeUtil::MakeShape(PRED, {}), "param1")); - auto select = builder.AddInstruction(HloInstruction::CreateTernary( - tuple_shape, HloOpcode::kSelect, pred_param, tuple_param0, tuple_param1)); + auto select = builder.AddInstruction( + HloInstruction::CreateTernary(tuple_shape, HloOpcode::kTupleSelect, + pred_param, tuple_param0, tuple_param1)); auto module = CreateNewModule(); module->AddEntryComputation(builder.Build()); @@ -1583,7 +1585,7 @@ TEST_F(BufferAssignmentTest, PeakBuffersWhile) { auto b = HloComputation::Builder(TestName() + ".cond"); b.AddInstruction(HloInstruction::CreateParameter(0, shape, "x")); b.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(true))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(true))); condition = module->AddEmbeddedComputation(b.Build()); } HloComputation* body; @@ -1646,9 +1648,9 @@ class WhileBufferAssignmentTest : public HloTestBase { builder.AddInstruction( HloInstruction::CreateParameter(0, loop_state_shape_, "loop_state")); auto zero = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(0))); auto ten = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(10))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(10))); builder.AddInstruction(HloInstruction::CreateBinary( ShapeUtil::MakeShape(PRED, {}), HloOpcode::kLt, zero, ten)); return builder.Build(); @@ -1707,7 +1709,7 @@ TEST_F(WhileBufferAssignmentTest, TwoForwardWhileLoops) { HloInstruction::CreateParameter(2, data_shape_, "weights1")); auto zero = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(0.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(0.0))); auto output0 = builder.AddInstruction( HloInstruction::CreateBroadcast(data_shape_, zero, {1})); auto output1 = builder.AddInstruction( @@ -1850,7 +1852,7 @@ TEST_F(WhileBufferAssignmentTest, ColocatedBuffers) { auto build_cond = [&]() { auto builder = HloComputation::Builder("cond"); auto const4 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(4))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(4))); auto param = builder.AddInstruction(HloInstruction::CreateParameter(0, r0s32, "x")); builder.AddInstruction(HloInstruction::CreateBinary( @@ -1862,7 +1864,7 @@ TEST_F(WhileBufferAssignmentTest, ColocatedBuffers) { auto build_body = [&]() { auto builder = HloComputation::Builder("body"); auto const9 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(9))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(9))); auto param = builder.AddInstruction(HloInstruction::CreateParameter(0, r0s32, "x")); builder.AddInstruction( @@ -1874,11 +1876,15 @@ TEST_F(WhileBufferAssignmentTest, ColocatedBuffers) { auto module = CreateNewModule(); auto builder = HloComputation::Builder("entry"); - auto infeed = builder.AddInstruction(HloInstruction::CreateInfeed(r0s32, "")); + auto token = builder.AddInstruction(HloInstruction::CreateToken()); + auto infeed = + builder.AddInstruction(HloInstruction::CreateInfeed(r0s32, token, "")); + auto infeed_data = builder.AddInstruction( + HloInstruction::CreateGetTupleElement(r0s32, infeed, 0)); auto cond0 = module->AddEmbeddedComputation(build_cond()); auto body0 = module->AddEmbeddedComputation(build_body()); auto while0 = builder.AddInstruction( - HloInstruction::CreateWhile(r0s32, cond0, body0, infeed)); + HloInstruction::CreateWhile(r0s32, cond0, body0, infeed_data)); auto cond1 = module->AddEmbeddedComputation(build_cond()); auto body1 = module->AddEmbeddedComputation(build_body()); @@ -1886,7 +1892,7 @@ TEST_F(WhileBufferAssignmentTest, ColocatedBuffers) { HloInstruction::CreateWhile(r0s32, cond1, body1, while0)); auto zero = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(0))); auto add = builder.AddInstruction( HloInstruction::CreateBinary(r0s32, HloOpcode::kAdd, zero, zero)); auto cond2 = module->AddEmbeddedComputation(build_cond()); @@ -1909,8 +1915,8 @@ TEST_F(WhileBufferAssignmentTest, ColocatedBuffers) { // computation, since the issue this test stresses depends on the order the // nodes are traversed during BufferAssignment. SequentialHloOrdering::HloModuleSequence sequence; - sequence[module->entry_computation()] = {infeed, while0, while1, zero, - add, while2, tuple}; + sequence[module->entry_computation()] = { + token, infeed, infeed_data, while0, while1, zero, add, while2, tuple}; TF_ASSERT_OK_AND_ASSIGN( auto assignment, BufferAssigner::Run( @@ -1948,7 +1954,7 @@ TEST_F(WhileBufferAssignmentTest, OneForwardBackwardWhileLoopSet) { HloInstruction::CreateParameter(1, data_shape_, "weights0")); auto zero = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(0.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(0.0))); auto output0 = builder.AddInstruction( HloInstruction::CreateBroadcast(data_shape_, zero, {1})); @@ -1992,16 +1998,16 @@ TEST_F(BufferAssignmentTest, TwoCalls) { auto param = builder.AddInstruction( HloInstruction::CreateParameter(0, r0f32, "param")); auto constant1 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0))); auto add = builder.AddInstruction( HloInstruction::CreateBinary(r0f32, HloOpcode::kAdd, param, constant1)); sub_computation = module->AddEmbeddedComputation(builder.Build(add)); } auto builder = HloComputation::Builder(TestName()); auto constant2 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(2.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(2.0))); auto constant3 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(3.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(3.0))); auto call1 = builder.AddInstruction( HloInstruction::CreateCall(r0f32, {constant2}, sub_computation)); auto call2 = builder.AddInstruction( @@ -2053,9 +2059,9 @@ TEST_F(WhileBufferAssignmentTest, WhileLoopsInterferingResultRange) { auto builder = HloComputation::Builder(TestName()); auto zero = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(0.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(0.0))); auto one = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0))); auto input0 = builder.AddInstruction( HloInstruction::CreateParameter(0, data_shape_, "input0")); @@ -2137,7 +2143,7 @@ TEST_F(WhileBufferAssignmentTest, WhilesDontShareEntryParamIfLiveOut) { HloInstruction::CreateParameter(1, data_shape_, "weights0")); auto zero = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(0.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(0.0))); auto output0 = builder.AddInstruction( HloInstruction::CreateBroadcast(data_shape_, zero, {1})); auto output1 = builder.AddInstruction( diff --git a/tensorflow/compiler/xla/service/buffer_liveness_test.cc b/tensorflow/compiler/xla/service/buffer_liveness_test.cc index f623aef67a4f98b447a9a15634a78deb60cfe6f1..4a927b57674345f8b3493c098778182a299c5902 100644 --- a/tensorflow/compiler/xla/service/buffer_liveness_test.cc +++ b/tensorflow/compiler/xla/service/buffer_liveness_test.cc @@ -327,11 +327,12 @@ TEST_F(BufferLivenessTest, RootInstructionIsNotLastInSequentialOrder) { builder.AddInstruction(HloInstruction::CreateParameter(0, vec_, "param")); auto add = builder.AddInstruction( HloInstruction::CreateBinary(vec_, HloOpcode::kAdd, param, param)); + auto token = builder.AddInstruction(HloInstruction::CreateToken()); auto recv = builder.AddInstruction( - HloInstruction::CreateRecv(vec_, /*channel_id=*/0)); + HloInstruction::CreateRecv(vec_, token, /*channel_id=*/0)); auto recv_done = builder.AddInstruction(HloInstruction::CreateRecvDone(recv)); auto send = builder.AddInstruction( - HloInstruction::CreateSend(recv_done, /*channel_id=*/1)); + HloInstruction::CreateSend(recv_done, token, /*channel_id=*/1)); auto send_done = builder.AddInstruction(HloInstruction::CreateSendDone(send)); auto module = CreateNewModule(); @@ -438,11 +439,13 @@ TEST_F(BufferLivenessTest, TupleConstantLiveOut) { // computation. The buffer containing {0, 1} is copied by GetTupleElement, and // the buffers containing {3} and 3 are dead. auto builder = HloComputation::Builder(TestName()); - auto inner_tuple0 = Literal::MakeTuple( - {Literal::CreateR0(0).get(), Literal::CreateR0(1).get()}); - auto inner_tuple1 = Literal::MakeTuple({Literal::CreateR0(3).get()}); + auto inner_tuple0 = + LiteralUtil::MakeTuple({LiteralUtil::CreateR0(0).get(), + LiteralUtil::CreateR0(1).get()}); + auto inner_tuple1 = + LiteralUtil::MakeTuple({LiteralUtil::CreateR0(3).get()}); auto tuple_constant = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::MakeTuple({inner_tuple0.get(), inner_tuple1.get()}))); + LiteralUtil::MakeTuple({inner_tuple0.get(), inner_tuple1.get()}))); builder.AddInstruction(HloInstruction::CreateGetTupleElement( inner_tuple0->shape(), tuple_constant, 0)); @@ -490,7 +493,7 @@ TEST_F(BufferLivenessTest, IndependentTupleElements) { builder.AddInstruction(HloInstruction::CreateGetTupleElement( tuple_element0_shape, tuple_param0, 0)); auto const0 = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR1({1.f, 1.f, 1.f, 1.f, 1.f, 1.f, 1.f, 1.f}))); + LiteralUtil::CreateR1({1.f, 1.f, 1.f, 1.f, 1.f, 1.f, 1.f, 1.f}))); auto add0 = builder.AddInstruction(HloInstruction::CreateBinary( tuple_element0_shape, HloOpcode::kAdd, tuple_element0, const0)); @@ -502,7 +505,7 @@ TEST_F(BufferLivenessTest, IndependentTupleElements) { builder.AddInstruction(HloInstruction::CreateGetTupleElement( tuple_element1_shape, tuple_param0, 1)); auto const1 = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR1({2.f, 2.f, 2.f, 2.f, 2.f, 2.f, 2.f, 2.f}))); + LiteralUtil::CreateR1({2.f, 2.f, 2.f, 2.f, 2.f, 2.f, 2.f, 2.f}))); auto add1 = builder.AddInstruction(HloInstruction::CreateBinary( tuple_element1_shape, HloOpcode::kAdd, tuple_element1, const1)); @@ -554,7 +557,7 @@ TEST_F(BufferLivenessTest, DependentTupleElements) { builder.AddInstruction(HloInstruction::CreateGetTupleElement( tuple_element0_shape, tuple_param0, 0)); auto const0 = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR1({1.f, 1.f, 1.f, 1.f, 1.f, 1.f, 1.f, 1.f}))); + LiteralUtil::CreateR1({1.f, 1.f, 1.f, 1.f, 1.f, 1.f, 1.f, 1.f}))); auto add0 = builder.AddInstruction(HloInstruction::CreateBinary( tuple_element0_shape, HloOpcode::kAdd, tuple_element0, const0)); @@ -626,7 +629,7 @@ class FusedDynamicUpdateSliceLivenessTest : public BufferLivenessTest { HloInstruction::CreateGetTupleElement(data_shape, tuple_param0, 1)); auto update = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR1({2.f, 2.f, 2.f}))); + LiteralUtil::CreateR1({2.f, 2.f, 2.f}))); HloInstruction* slice = nullptr; if (update_uses_tuple_element1) { // Create a slice instruction as an additional user of 'gte1'. @@ -637,7 +640,7 @@ class FusedDynamicUpdateSliceLivenessTest : public BufferLivenessTest { } // Create a DynamicUpdateSlice instruction of tuple element 1 with 'update'. auto starts = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR1({2}))); + HloInstruction::CreateConstant(LiteralUtil::CreateR1({2}))); auto dynamic_update_slice = builder.AddInstruction(HloInstruction::CreateDynamicUpdateSlice( data_shape, gte1, update, starts)); @@ -756,7 +759,7 @@ class DynamicUpdateSliceLivenessTest : public BufferLivenessTest { HloInstruction::CreateGetTupleElement(data_shape, tuple_param0, 1)); auto update = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR1({2.f, 2.f, 2.f}))); + LiteralUtil::CreateR1({2.f, 2.f, 2.f}))); if (tuple_element1_has_two_uses) { // Add 'gte0' and 'gte1' to create another user of 'gte1'. @@ -765,7 +768,7 @@ class DynamicUpdateSliceLivenessTest : public BufferLivenessTest { } // Create a DynamicUpdateSlice instruction of tuple element 1 with 'update'. auto starts = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR1({2}))); + HloInstruction::CreateConstant(LiteralUtil::CreateR1({2}))); auto dynamic_update_slice = builder.AddInstruction(HloInstruction::CreateDynamicUpdateSlice( data_shape, gte1, update, starts)); diff --git a/tensorflow/compiler/xla/service/call_graph_test.cc b/tensorflow/compiler/xla/service/call_graph_test.cc index 1ea7d538cd515c3098b6a1f03c6146d288330406..cc80b7484313329104eec1ce71a150b47d8330c9 100644 --- a/tensorflow/compiler/xla/service/call_graph_test.cc +++ b/tensorflow/compiler/xla/service/call_graph_test.cc @@ -15,7 +15,7 @@ limitations under the License. #include "tensorflow/compiler/xla/service/call_graph.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/service/hlo_computation.h" #include "tensorflow/compiler/xla/shape_util.h" #include "tensorflow/compiler/xla/status_macros.h" @@ -82,7 +82,7 @@ class CallGraphTest : public HloTestBase { HloInstruction* param0 = builder.AddInstruction( HloInstruction::CreateParameter(0, kScalarShape, "param0")); HloInstruction* zero = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(0.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(0.0f))); builder.AddInstruction(HloInstruction::CreateBinary( ShapeUtil::MakeShape(PRED, {}), HloOpcode::kGt, param0, zero)); return builder.Build(); @@ -247,11 +247,11 @@ TEST_F(CallGraphTest, ComputationWithConditional) { HloComputation::Builder builder(TestName()); HloInstruction* pred = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(false))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(false))); HloInstruction* const1 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(56.4f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(56.4f))); HloInstruction* const2 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(12.6f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(12.6f))); HloInstruction* conditional = builder.AddInstruction(HloInstruction::CreateConditional( kScalarShape, pred, const1, true_computation, const2, diff --git a/tensorflow/compiler/xla/service/call_inliner.cc b/tensorflow/compiler/xla/service/call_inliner.cc index 482ccc5b67109258f544e5657ecfa0e8f62192c0..256d05a73e0bf61d959d21795c106286b52d0b19 100644 --- a/tensorflow/compiler/xla/service/call_inliner.cc +++ b/tensorflow/compiler/xla/service/call_inliner.cc @@ -18,6 +18,7 @@ limitations under the License. #include #include "tensorflow/compiler/xla/service/call_graph.h" +#include "tensorflow/compiler/xla/service/hlo_dce.h" #include "tensorflow/core/lib/core/errors.h" namespace xla { @@ -151,6 +152,14 @@ StatusOr CallInliner::Run(HloModule* module) { } return Status::OK(); })); + if (did_mutate) { + // Run DCE to remove called computations which are now becoming unused. + // This can result then in problems if within the called computation, there + // were send/recv instructions, which the module group verifier will flag as + // error findingthe same channel ID used for multiple send/recv + // instructions. + TF_RETURN_IF_ERROR(HloDCE().Run(module).status()); + } return did_mutate; } diff --git a/tensorflow/compiler/xla/service/call_inliner_test.cc b/tensorflow/compiler/xla/service/call_inliner_test.cc index 738d00881dd057fc13c115006c15e8f5b6d14a1d..ff968bca297077c7cf869ff8d2becb8bf739dce3 100644 --- a/tensorflow/compiler/xla/service/call_inliner_test.cc +++ b/tensorflow/compiler/xla/service/call_inliner_test.cc @@ -19,7 +19,7 @@ limitations under the License. #include #include "tensorflow/compiler/xla/layout_util.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/ptr_util.h" #include "tensorflow/compiler/xla/service/hlo_computation.h" #include "tensorflow/compiler/xla/service/hlo_instruction.h" @@ -48,9 +48,9 @@ TEST_F(CallInlinerTest, ControlDependenciesAreCarriedToCaller) { // the "one" value. HloComputation::Builder inner(TestName() + ".inner"); HloInstruction* zero = inner.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(24.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(24.0f))); HloInstruction* one = inner.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(42.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(42.0f))); TF_ASSERT_OK(zero->AddControlDependencyTo(one)); auto module = CreateNewModule(); HloComputation* inner_computation = @@ -87,7 +87,7 @@ TEST_F(CallInlinerTest, CallsWithinWhileBodiesAreInlined) { // little trickier. HloComputation::Builder just_false(TestName() + ".false"); just_false.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(false))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(false))); HloComputation* false_computation = module->AddEmbeddedComputation(just_false.Build()); @@ -99,7 +99,7 @@ TEST_F(CallInlinerTest, CallsWithinWhileBodiesAreInlined) { HloComputation::Builder outer(TestName() + ".outer"); HloInstruction* init_value = outer.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(false))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(false))); outer.AddInstruction( HloInstruction::CreateWhile(pred, call_false, call_false, init_value)); @@ -123,9 +123,9 @@ TEST_F(CallInlinerTest, InlineWithoutRunningPass) { HloComputation::Builder just_false(TestName() + ".false"); auto* true_constant = just_false.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR1({true}))); + HloInstruction::CreateConstant(LiteralUtil::CreateR1({true}))); auto* false_constant = just_false.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(false))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(false))); TF_ASSERT_OK(false_constant->AddControlDependencyTo(true_constant)); HloComputation* false_computation = module->AddEmbeddedComputation(just_false.Build()); @@ -147,15 +147,17 @@ TEST_F(CallInlinerTest, CallToOutfeedComputationIsInlined) { HloComputation::Builder outfeeder(TestName() + ".outfeeder"); auto value = outfeeder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(42.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(42.0))); + auto token = outfeeder.AddInstruction(HloInstruction::CreateToken()); outfeeder.AddInstruction( - HloInstruction::CreateOutfeed(f32, value, /*outfeed_config=*/"")); + HloInstruction::CreateOutfeed(f32, value, token, /*outfeed_config=*/"")); auto outfeed_computation = module->AddEmbeddedComputation(outfeeder.Build()); HloComputation::Builder outer(TestName() + ".outer"); outer.AddInstruction(HloInstruction::CreateCall( - ShapeUtil::MakeNil(), /*operands=*/{}, outfeed_computation)); + outfeed_computation->root_instruction()->shape(), /*operands=*/{}, + outfeed_computation)); module->AddEntryComputation(outer.Build()); diff --git a/tensorflow/compiler/xla/service/computation_placer.cc b/tensorflow/compiler/xla/service/computation_placer.cc index 7c1bacff92b231661477b9931a3066fd91110445..d26486fcfe0b1bc51867de5113cc5e42a0d7b4f0 100644 --- a/tensorflow/compiler/xla/service/computation_placer.cc +++ b/tensorflow/compiler/xla/service/computation_placer.cc @@ -19,7 +19,7 @@ limitations under the License. #include #include -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/ptr_util.h" #include "tensorflow/compiler/xla/shape_util.h" #include "tensorflow/compiler/xla/status.h" diff --git a/tensorflow/compiler/xla/service/conditional_simplifier.cc b/tensorflow/compiler/xla/service/conditional_simplifier.cc index e9ec796121fff223474c3e81a5e973cc37f8caec..b7be3ba605a89a736b032eaab5a5085ac64fc549 100644 --- a/tensorflow/compiler/xla/service/conditional_simplifier.cc +++ b/tensorflow/compiler/xla/service/conditional_simplifier.cc @@ -19,7 +19,7 @@ limitations under the License. #include #include -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/service/call_inliner.h" #include "tensorflow/compiler/xla/service/hlo_computation.h" #include "tensorflow/compiler/xla/service/hlo_instruction.h" diff --git a/tensorflow/compiler/xla/service/conditional_simplifier_test.cc b/tensorflow/compiler/xla/service/conditional_simplifier_test.cc index 868348547d9f5cbdc7576c7fc0697d72c3a3e557..c43a31b167d47af3c92ed35fa52594fa5da1e4af 100644 --- a/tensorflow/compiler/xla/service/conditional_simplifier_test.cc +++ b/tensorflow/compiler/xla/service/conditional_simplifier_test.cc @@ -55,7 +55,7 @@ HloComputation* ConditionalSimplifierTest::MakeConditional(HloModule* module) { true_computation_builder.AddInstruction(HloInstruction::CreateParameter( 0, ShapeUtil::MakeShape(S32, {}), "param")); auto one = true_computation_builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1))); true_computation_builder.AddInstruction(HloInstruction::CreateBinary( ShapeUtil::MakeShape(S32, {}), HloOpcode::kAdd, param, one)); @@ -73,7 +73,7 @@ HloComputation* ConditionalSimplifierTest::MakeConditional(HloModule* module) { HloInstruction::CreateParameter(0, ShapeUtil::MakeShape(S32, {}), "param")); auto forty_two = false_computation_builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(42))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(42))); false_computation_builder.AddInstruction(HloInstruction::CreateBinary( ShapeUtil::MakeShape(S32, {}), HloOpcode::kAdd, param, forty_two)); @@ -82,11 +82,11 @@ HloComputation* ConditionalSimplifierTest::MakeConditional(HloModule* module) { } auto false_instrn = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(false))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(false))); auto false_param = builder.AddInstruction(HloInstruction::CreateParameter( 0, ShapeUtil::MakeShape(S32, {}), "false_param")); auto one = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1))); builder.AddInstruction(HloInstruction::CreateConditional( ShapeUtil::MakeShape(S32, {}), false_instrn, one, true_computation, @@ -106,7 +106,7 @@ TEST_F(ConditionalSimplifierTest, ConditionalWithControlDependency) { HloComputation* computation = MakeConditional(&module()); auto* true_op = computation->AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(true))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(true))); TF_ASSERT_OK( true_op->AddControlDependencyTo(computation->root_instruction())); @@ -119,10 +119,11 @@ TEST_F(ConditionalSimplifierTest, NotRemovedIfContainsSend) { ASSERT_EQ(conditional->opcode(), HloOpcode::kConditional); auto* true_computation = conditional->true_computation(); + auto* token = true_computation->AddInstruction(HloInstruction::CreateToken()); auto* send = true_computation->AddInstruction(HloInstruction::CreateSend( true_computation->AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(true))), - /*channel_id=*/0)); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(true))), + token, /*channel_id=*/0)); true_computation->AddInstruction(HloInstruction::CreateSendDone(send)); EXPECT_FALSE(ConditionalSimplifier().Run(&module()).ValueOrDie()); } @@ -133,8 +134,9 @@ TEST_F(ConditionalSimplifierTest, NotRemovedIfContainsRecv) { ASSERT_EQ(conditional->opcode(), HloOpcode::kConditional); auto* true_computation = conditional->true_computation(); + auto* token = true_computation->AddInstruction(HloInstruction::CreateToken()); auto* recv = true_computation->AddInstruction(HloInstruction::CreateRecv( - ShapeUtil::MakeShape(F32, {1}), /*channel_id=*/0)); + ShapeUtil::MakeShape(F32, {1}), token, /*channel_id=*/0)); true_computation->AddInstruction(HloInstruction::CreateRecvDone(recv)); EXPECT_FALSE(ConditionalSimplifier().Run(&module()).ValueOrDie()); } @@ -144,8 +146,9 @@ TEST_F(ConditionalSimplifierTest, NotRemovedIfContainsNonRemovableInstruction) { auto* conditional = computation->root_instruction(); ASSERT_EQ(conditional->opcode(), HloOpcode::kConditional); auto* false_computation = conditional->false_computation(); - false_computation->AddInstruction( - HloInstruction::CreateInfeed(ShapeUtil::MakeShape(F32, {1}), "config")); + auto token = false_computation->AddInstruction(HloInstruction::CreateToken()); + false_computation->AddInstruction(HloInstruction::CreateInfeed( + ShapeUtil::MakeShape(F32, {1}), token, "config")); EXPECT_FALSE(ConditionalSimplifier().Run(&module()).ValueOrDie()); } diff --git a/tensorflow/compiler/xla/service/copy_insertion.cc b/tensorflow/compiler/xla/service/copy_insertion.cc index b0ad433d8ddb7b5e0861150634ff91e4068d10dd..ab3d846403ef264cd732a9c01d524cd4ccf65c38 100644 --- a/tensorflow/compiler/xla/service/copy_insertion.cc +++ b/tensorflow/compiler/xla/service/copy_insertion.cc @@ -1093,8 +1093,7 @@ void MaybeDumpModule(const string& message, const HloModule& module) { } // namespace Status RemoveUnnecessaryCopies( - const HloOrdering& ordering, - const tensorflow::gtl::FlatSet& copies_to_exclude, HloModule* module, + const HloOrdering& ordering, HloModule* module, const HloDataflowAnalysis::FusionCanShareBufferFunction& fusion_can_share_buffer) { MaybeDumpModule("after adding copies to resolve interference", *module); @@ -1108,7 +1107,6 @@ Status RemoveUnnecessaryCopies( for (HloComputation* computation : module->computations()) { for (HloInstruction* instruction : computation->instructions()) { if (instruction->opcode() == HloOpcode::kCopy && - !ContainsKey(copies_to_exclude, instruction->unique_id()) && instruction->CopyElisionAllowed()) { TF_RETURN_IF_ERROR(copy_remover.TryElideCopy(instruction).status()); } @@ -1152,16 +1150,13 @@ StatusOr CopyInsertion::Run(HloModule* module) { "Call graph must be flattened before copy insertion."); } - // Gather Ids of existing kCopy instructions in the module. We avoid removing - // these copies (except via DCE in TupleSimplifier) because they may have been - // added for reasons not considered by copy insertion (eg, layout assignment). - // Instruction id is used instead of HloInstruction* because the pointer - // values may be recycled. - tensorflow::gtl::FlatSet existing_copies; - for (HloComputation* computation : module->computations()) { - for (HloInstruction* instruction : computation->instructions()) { - if (instruction->opcode() == HloOpcode::kCopy) { - existing_copies.insert(instruction->unique_id()); + int64 num_existing_copies = 0; + if (VLOG_IS_ON(1)) { + for (HloComputation* computation : module->computations()) { + for (HloInstruction* instruction : computation->instructions()) { + if (instruction->opcode() == HloOpcode::kCopy) { + ++num_existing_copies; + } } } } @@ -1181,8 +1176,7 @@ StatusOr CopyInsertion::Run(HloModule* module) { TF_DCHECK_OK(VerifyNoLiveRangeInterference(module)); DependencyHloOrdering ordering(module); - TF_RETURN_IF_ERROR( - RemoveUnnecessaryCopies(ordering, existing_copies, module)); + TF_RETURN_IF_ERROR(RemoveUnnecessaryCopies(ordering, module)); TF_RETURN_IF_ERROR(AddSpecialCaseCopies(*call_graph, module)); @@ -1203,7 +1197,7 @@ StatusOr CopyInsertion::Run(HloModule* module) { } } } - VLOG(1) << "Num copies before copy-insertion: " << existing_copies.size(); + VLOG(1) << "Num copies before copy-insertion: " << num_existing_copies; VLOG(1) << "Num copies after copy-insertion: " << num_total_copies; } diff --git a/tensorflow/compiler/xla/service/copy_insertion.h b/tensorflow/compiler/xla/service/copy_insertion.h index 6d257060891122e56b763b32166fb4c11dfc444b..e1973db928423cb4bbad00fe34329f731b23ea09 100644 --- a/tensorflow/compiler/xla/service/copy_insertion.h +++ b/tensorflow/compiler/xla/service/copy_insertion.h @@ -21,7 +21,6 @@ limitations under the License. #include "tensorflow/compiler/xla/service/hlo_instruction.h" #include "tensorflow/compiler/xla/service/hlo_module.h" #include "tensorflow/compiler/xla/service/hlo_pass_interface.h" -#include "tensorflow/core/lib/gtl/flatmap.h" namespace xla { @@ -79,11 +78,10 @@ class CopyInsertion : public HloPassInterface { }; // Try to remove as many copies from the module as possible without introducing -// live range interference. Copy instructions (identified by their unique id) in -// the set copies_to_exclude are not considered for removal. +// live range interference. Only copy instructions that are eligible for +// copy elision are considered for removal. Status RemoveUnnecessaryCopies( - const HloOrdering& ordering, - const tensorflow::gtl::FlatSet& copies_to_exclude, HloModule* module, + const HloOrdering& ordering, HloModule* module, const HloDataflowAnalysis::FusionCanShareBufferFunction& fusion_can_share_buffer = nullptr); diff --git a/tensorflow/compiler/xla/service/copy_insertion_test.cc b/tensorflow/compiler/xla/service/copy_insertion_test.cc index ed1a50f516ee23e0f034bf5c2ed15fac7a70c3cc..cd735256b83f5f1d69a89e693de6064d460a36e5 100644 --- a/tensorflow/compiler/xla/service/copy_insertion_test.cc +++ b/tensorflow/compiler/xla/service/copy_insertion_test.cc @@ -18,7 +18,7 @@ limitations under the License. #include #include "tensorflow/compiler/xla/legacy_flags/debug_options_flags.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/service/hlo_computation.h" #include "tensorflow/compiler/xla/service/hlo_instruction.h" #include "tensorflow/compiler/xla/service/hlo_matchers.h" @@ -108,7 +108,7 @@ TEST_F(CopyInsertionTest, SingleConstant) { // be copied before entering the tuple. auto builder = HloComputation::Builder(TestName()); HloInstruction* constant = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0))); HloInstruction* tuple = builder.AddInstruction(HloInstruction::CreateTuple({constant})); @@ -125,21 +125,27 @@ TEST_F(CopyInsertionTest, SingleConstant) { } TEST_F(CopyInsertionTest, ExistingCopiesNotRemoved) { - // Verify that an kCopy instructions which exist in the pass before + // Verify that kCopy instructions which change layout and exist before // copy-insertion remain in the graph after copy-insertion. auto module = CreateNewModule(); auto builder = HloComputation::Builder(TestName()); - HloInstruction* constant = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); - HloInstruction* copy_1 = builder.AddInstruction(HloInstruction::CreateUnary( - constant->shape(), HloOpcode::kCopy, constant)); - HloInstruction* copy_2 = builder.AddInstruction(HloInstruction::CreateUnary( - constant->shape(), HloOpcode::kCopy, constant)); + HloInstruction* constant = + builder.AddInstruction(HloInstruction::CreateConstant( + LiteralUtil::CreateR2({{0.f, 2.f}, {2.f, 4.f}}))); + auto minor_to_major = LayoutUtil::MinorToMajor(constant->shape()); + Layout reversed_layout = + LayoutUtil::MakeLayoutFromMajorToMinor(minor_to_major); + Shape copy_shape = constant->shape(); + *copy_shape.mutable_layout() = reversed_layout; + HloInstruction* copy_1 = builder.AddInstruction( + HloInstruction::CreateUnary(copy_shape, HloOpcode::kCopy, constant)); + HloInstruction* copy_2 = builder.AddInstruction( + HloInstruction::CreateUnary(copy_shape, HloOpcode::kCopy, constant)); HloInstruction* add = builder.AddInstruction(HloInstruction::CreateBinary( constant->shape(), HloOpcode::kAdd, copy_1, copy_2)); - HloInstruction* add_copy = builder.AddInstruction( - HloInstruction::CreateUnary(constant->shape(), HloOpcode::kCopy, add)); + builder.AddInstruction( + HloInstruction::CreateUnary(add->shape(), HloOpcode::kCopy, add)); module->AddEntryComputation(builder.Build()); @@ -147,12 +153,11 @@ TEST_F(CopyInsertionTest, ExistingCopiesNotRemoved) { InsertCopies(module.get()); - EXPECT_EQ(CountCopies(*module), 3); + EXPECT_EQ(CountCopies(*module), 2); - EXPECT_EQ(module->entry_computation()->root_instruction(), add_copy); - EXPECT_THAT( - module->entry_computation()->root_instruction(), - op::Copy(op::Add(op::Copy(op::Constant()), op::Copy(op::Constant())))); + EXPECT_EQ(module->entry_computation()->root_instruction(), add); + EXPECT_THAT(module->entry_computation()->root_instruction(), + op::Add(op::Copy(op::Constant()), op::Copy(op::Constant()))); } TEST_F(CopyInsertionTest, MultipleConstantsAndParameters) { @@ -162,9 +167,9 @@ TEST_F(CopyInsertionTest, MultipleConstantsAndParameters) { auto builder = HloComputation::Builder(TestName()); HloInstruction* constant1 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0))); HloInstruction* constant2 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(2.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(2.0))); HloInstruction* x = builder.AddInstruction( HloInstruction::CreateParameter(0, ShapeUtil::MakeShape(F32, {}), "x")); @@ -192,11 +197,11 @@ TEST_F(CopyInsertionTest, AmbiguousPointsToSet) { // the computation result. Verify that copies are added properly. auto builder = HloComputation::Builder(TestName()); HloInstruction* constant1 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0))); HloInstruction* constant2 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(2.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(2.0))); HloInstruction* constant3 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(3.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(3.0))); HloInstruction* tuple1 = builder.AddInstruction( HloInstruction::CreateTuple({constant1, constant2})); @@ -204,9 +209,9 @@ TEST_F(CopyInsertionTest, AmbiguousPointsToSet) { HloInstruction::CreateTuple({constant3, constant2})); HloInstruction* pred = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(false))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(false))); builder.AddInstruction(HloInstruction::CreateTernary( - tuple1->shape(), HloOpcode::kSelect, pred, tuple1, tuple2)); + tuple1->shape(), HloOpcode::kTupleSelect, pred, tuple1, tuple2)); EXPECT_THAT(constant1->users(), UnorderedElementsAre(tuple1)); EXPECT_THAT(constant2->users(), UnorderedElementsAre(tuple1, tuple2)); @@ -250,8 +255,9 @@ TEST_F(CopyInsertionTest, BitcastConstant) { // The output of a bitcast is its operand (same buffer), so a bitcast // constant feeding the result must have a copy added. auto builder = HloComputation::Builder(TestName()); - HloInstruction* constant = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR1({1.0, 42.0}))); + HloInstruction* constant = + builder.AddInstruction(HloInstruction::CreateConstant( + LiteralUtil::CreateR1({1.0, 42.0}))); HloInstruction* bitcast = builder.AddInstruction(HloInstruction::CreateUnary( ShapeUtil::MakeShape(F32, {2, 2}), HloOpcode::kBitcast, constant)); @@ -365,9 +371,9 @@ TEST_F(CopyInsertionTest, AmbiguousTopLevelRoot) { // copy is added. auto builder = HloComputation::Builder(TestName()); HloInstruction* constant1 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0))); HloInstruction* constant2 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(2.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(2.0))); HloInstruction* tuple1 = builder.AddInstruction( HloInstruction::CreateTuple({constant1, constant2})); @@ -375,9 +381,9 @@ TEST_F(CopyInsertionTest, AmbiguousTopLevelRoot) { HloInstruction::CreateTuple({constant2, constant1})); HloInstruction* pred = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(false))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(false))); HloInstruction* select = builder.AddInstruction(HloInstruction::CreateTernary( - tuple1->shape(), HloOpcode::kSelect, pred, tuple1, tuple2)); + tuple1->shape(), HloOpcode::kTupleSelect, pred, tuple1, tuple2)); HloInstruction* gte = builder.AddInstruction(HloInstruction::CreateGetTupleElement( ShapeUtil::GetSubshape(select->shape(), {0}), select, 0)); @@ -408,7 +414,7 @@ class WhileCopyInsertionTest : public CopyInsertionTest { const Shape& loop_state_shape) { auto builder = HloComputation::Builder(TestName() + ".Condition"); auto limit_const = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(10))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(10))); auto loop_state = builder.AddInstruction( HloInstruction::CreateParameter(0, loop_state_shape, "loop_state")); auto induction_variable = @@ -437,7 +443,7 @@ class WhileCopyInsertionTest : public CopyInsertionTest { builder.AddInstruction(HloInstruction::CreateGetTupleElement( induction_variable_shape_, loop_state, 0)); auto inc = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1))); auto add0 = builder.AddInstruction(HloInstruction::CreateBinary( induction_variable->shape(), HloOpcode::kAdd, induction_variable, inc)); // Update data GTE(1). @@ -475,7 +481,7 @@ class WhileCopyInsertionTest : public CopyInsertionTest { builder.AddInstruction(HloInstruction::CreateGetTupleElement( induction_variable_shape_, loop_state, 0)); auto inc = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1))); // add0 = Add(in0, 1) auto add0 = builder.AddInstruction(HloInstruction::CreateBinary( @@ -544,7 +550,7 @@ class WhileCopyInsertionTest : public CopyInsertionTest { builder.AddInstruction(HloInstruction::CreateGetTupleElement( induction_variable_shape_, loop_state, 0)); auto inc = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1))); // add0 = Add(in0, 1) auto add0 = builder.AddInstruction(HloInstruction::CreateBinary( induction_variable->shape(), HloOpcode::kAdd, induction_variable, inc)); @@ -559,8 +565,9 @@ class WhileCopyInsertionTest : public CopyInsertionTest { data = builder.AddInstruction( HloInstruction::CreateGetTupleElement(data_shape_, loop_state, 1)); } - auto update = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR1({1.f, 1.f, 1.f, 1.f, 1.f, 1.f, 1.f, 1.f}))); + auto update = builder.AddInstruction( + HloInstruction::CreateConstant(LiteralUtil::CreateR1( + {1.f, 1.f, 1.f, 1.f, 1.f, 1.f, 1.f, 1.f}))); // add1 = Add(in1, {1, 1, 1, 1, 1, 1, 1, 1}) auto add1 = builder.AddInstruction(HloInstruction::CreateBinary( data_shape_, HloOpcode::kAdd, data, update)); @@ -593,7 +600,7 @@ class WhileCopyInsertionTest : public CopyInsertionTest { auto gte0 = builder.AddInstruction(HloInstruction::CreateGetTupleElement( induction_variable_shape_, loop_state, 0)); auto inc = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1))); auto add0 = builder.AddInstruction(HloInstruction::CreateBinary( gte0->shape(), HloOpcode::kAdd, gte0, inc)); @@ -603,8 +610,9 @@ class WhileCopyInsertionTest : public CopyInsertionTest { // GTE(GTE(loop_state, 1), 0) -> Add auto gte10 = builder.AddInstruction( HloInstruction::CreateGetTupleElement(data_shape_, gte1, 0)); - auto update10 = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR1({1.f, 1.f, 1.f, 1.f, 1.f, 1.f, 1.f, 1.f}))); + auto update10 = builder.AddInstruction( + HloInstruction::CreateConstant(LiteralUtil::CreateR1( + {1.f, 1.f, 1.f, 1.f, 1.f, 1.f, 1.f, 1.f}))); auto add10 = builder.AddInstruction(HloInstruction::CreateBinary( data_shape_, HloOpcode::kAdd, gte10, update10)); @@ -628,10 +636,11 @@ class WhileCopyInsertionTest : public CopyInsertionTest { bool nested = false) { auto builder = HloComputation::Builder(TestName() + ".While"); auto induction_var_init = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(0))); - auto data_init = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR1({0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f}))); + auto data_init = builder.AddInstruction( + HloInstruction::CreateConstant(LiteralUtil::CreateR1( + {0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f}))); if (nested) { auto inner_init = builder.AddInstruction( @@ -654,8 +663,9 @@ class WhileCopyInsertionTest : public CopyInsertionTest { HloInstruction* BuildWhileInstruction_InitPointsToConstant() { auto builder = HloComputation::Builder(TestName() + ".While"); - auto data_init = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR1({0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f}))); + auto data_init = builder.AddInstruction( + HloInstruction::CreateConstant(LiteralUtil::CreateR1( + {0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f}))); return BuildWhileInstructionWithCustomInit(loop_state_shape_, data_init, &builder); } @@ -672,11 +682,11 @@ class WhileCopyInsertionTest : public CopyInsertionTest { auto builder = HloComputation::Builder(TestName() + ".While"); auto one = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0))); auto v1 = builder.AddInstruction( HloInstruction::CreateBroadcast(data_shape_, one, {1})); auto zero = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0))); auto v2 = builder.AddInstruction( HloInstruction::CreateBroadcast(data_shape_, zero, {1})); @@ -684,9 +694,9 @@ class WhileCopyInsertionTest : public CopyInsertionTest { auto tuple2 = builder.AddInstruction(HloInstruction::CreateTuple({v2, v1})); auto pred = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(false))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(false))); auto data_init = builder.AddInstruction(HloInstruction::CreateTernary( - nested_tuple_shape_, HloOpcode::kSelect, pred, tuple1, tuple2)); + nested_tuple_shape_, HloOpcode::kTupleSelect, pred, tuple1, tuple2)); return BuildWhileInstructionWithCustomInit(nested_loop_state_shape_, data_init, &builder); @@ -696,7 +706,7 @@ class WhileCopyInsertionTest : public CopyInsertionTest { auto builder = HloComputation::Builder(TestName() + ".While"); auto one = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0))); auto one_vec = builder.AddInstruction( HloInstruction::CreateBroadcast(data_shape_, one, {1})); auto data_init = @@ -709,11 +719,12 @@ class WhileCopyInsertionTest : public CopyInsertionTest { HloInstruction* BuildWhileInstruction_InitPointsToInterfering() { auto builder = HloComputation::Builder(TestName() + ".While"); auto one = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0))); auto data_init = builder.AddInstruction( HloInstruction::CreateBroadcast(data_shape_, one, {1})); - auto one_vec = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR1({1.f, 1.f, 1.f, 1.f, 1.f, 1.f, 1.f, 1.f}))); + auto one_vec = builder.AddInstruction( + HloInstruction::CreateConstant(LiteralUtil::CreateR1( + {1.f, 1.f, 1.f, 1.f, 1.f, 1.f, 1.f, 1.f}))); // Take a reference to 'data_init' to make it interfere with while result. auto add = builder.AddInstruction(HloInstruction::CreateBinary( data_shape_, HloOpcode::kAdd, data_init, one_vec)); @@ -745,7 +756,7 @@ class WhileCopyInsertionTest : public CopyInsertionTest { const bool nested = ShapeUtil::Equal(loop_state_shape, nested_loop_state_shape_); auto induction_var_init = builder->AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(0))); auto condition = module_->AddEmbeddedComputation( BuildConditionComputation(loop_state_shape)); auto body = module_->AddEmbeddedComputation( @@ -1247,7 +1258,6 @@ TEST_F(WhileCopyInsertionTest, InitPointsToNonDistinctUsedByTwoWhileLoops) { auto loop_init = builder.AddInstruction( HloInstruction::CreateTuple({iter_param, data_param, data_param})); - // Two while loops shares the same loop init tuple. auto while_hlo1 = builder.AddInstruction(HloInstruction::CreateWhile( loop_state_shape, condition1, body1, loop_init)); @@ -1305,7 +1315,7 @@ TEST_F(CopyInsertionTest, SwizzlingWhile) { cond_builder.AddInstruction( HloInstruction::CreateParameter(0, loop_state_shape, "param")); auto cond_constant = cond_builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(false))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(false))); cond_builder.AddInstruction(HloInstruction::CreateUnary( cond_constant->shape(), HloOpcode::kNot, cond_constant)); HloComputation* condition = @@ -1313,9 +1323,9 @@ TEST_F(CopyInsertionTest, SwizzlingWhile) { auto builder = HloComputation::Builder(TestName()); auto constant1 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0))); auto constant2 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(2.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(2.0))); auto tuple = builder.AddInstruction( HloInstruction::CreateTuple({constant1, constant2})); auto xla_while = builder.AddInstruction( @@ -1370,7 +1380,7 @@ TEST_F(CopyInsertionTest, SwizzlingWhileWithOneOp) { cond_builder.AddInstruction( HloInstruction::CreateParameter(0, loop_state_shape, "param")); auto cond_constant = cond_builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(false))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(false))); cond_builder.AddInstruction(HloInstruction::CreateUnary( cond_constant->shape(), HloOpcode::kNot, cond_constant)); HloComputation* condition = @@ -1378,9 +1388,9 @@ TEST_F(CopyInsertionTest, SwizzlingWhileWithOneOp) { auto builder = HloComputation::Builder(TestName()); auto constant1 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0))); auto constant2 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(2.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(2.0))); auto tuple = builder.AddInstruction( HloInstruction::CreateTuple({constant1, constant2})); auto xla_while = builder.AddInstruction( @@ -1430,7 +1440,7 @@ TEST_F(CopyInsertionTest, SwizzlingWhileSharedInput) { cond_builder.AddInstruction( HloInstruction::CreateParameter(0, loop_state_shape, "param")); auto cond_constant = cond_builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(false))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(false))); cond_builder.AddInstruction(HloInstruction::CreateUnary( cond_constant->shape(), HloOpcode::kNot, cond_constant)); HloComputation* condition = @@ -1438,7 +1448,7 @@ TEST_F(CopyInsertionTest, SwizzlingWhileSharedInput) { auto builder = HloComputation::Builder(TestName()); auto constant = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0))); auto tuple = builder.AddInstruction(HloInstruction::CreateTuple({constant, constant})); builder.AddInstruction( @@ -1515,7 +1525,7 @@ TEST_F(CopyInsertionTest, SequentialWhiles) { cond_builder.AddInstruction( HloInstruction::CreateParameter(0, loop_state_shape, "param")); auto cond_constant = cond_builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(false))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(false))); cond_builder.AddInstruction(HloInstruction::CreateUnary( cond_constant->shape(), HloOpcode::kNot, cond_constant)); HloComputation* condition = @@ -1570,14 +1580,14 @@ TEST_F(CopyInsertionTest, WhileBodyWithConstantRoot) { body_builder.AddInstruction( HloInstruction::CreateParameter(0, scalar_shape_, "param")); body_builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(123.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(123.0))); HloComputation* body = module->AddEmbeddedComputation(body_builder.Build()); auto cond_builder = HloComputation::Builder("condition"); cond_builder.AddInstruction( HloInstruction::CreateParameter(0, scalar_shape_, "param")); cond_builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(false))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(false))); HloComputation* condition = module->AddEmbeddedComputation(cond_builder.Build()); @@ -1605,8 +1615,8 @@ HloModule TokensShouldNotBeCopied %constant.1 = s32[] constant(1) %add = s32[] add(s32[] %get-tuple-element.1, s32[] %constant.1) %get-tuple-element.2 = token[] get-tuple-element((s32[], token[]) %param.1), index=1 - %generate-token = token[] generate-token(token[] %get-tuple-element.2) - ROOT %tuple = (s32[], token[]) tuple(s32[] %add, token[] %generate-token) + %after-all = token[] after-all(token[] %get-tuple-element.2) + ROOT %tuple = (s32[], token[]) tuple(s32[] %add, token[] %after-all) } %Cond (param: (s32[], token[])) -> pred[] { @@ -1619,7 +1629,7 @@ HloModule TokensShouldNotBeCopied ENTRY %TokensShouldNotBeCopied () -> s32[] { %one = s32[] constant(1) %negative_one = s32[] negate(%one) - %init_token = token[] generate-token() + %init_token = token[] after-all() %init_tuple = (s32[], token[]) tuple(s32[] %negative_one, token[] %init_token) %while = (s32[], token[]) while((s32[], token[]) %init_tuple), condition=%Cond, body=%Body ROOT %root = s32[] get-tuple-element((s32[], token[]) %while), index=0 @@ -1639,7 +1649,7 @@ std::unique_ptr MakeTrivialCondition(const Shape& shape) { builder.AddInstruction( HloInstruction::CreateParameter(0, shape, "loop_state")); auto constant = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(false))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(false))); builder.AddInstruction(HloInstruction::CreateUnary( constant->shape(), HloOpcode::kNot, constant)); return builder.Build(); diff --git a/tensorflow/compiler/xla/service/cpu/BUILD b/tensorflow/compiler/xla/service/cpu/BUILD index b703be0f39e2032bc58479f0b957f9d8b01a77c3..c45d914e937bf4defba3baa566bb6c9882c9b328 100644 --- a/tensorflow/compiler/xla/service/cpu/BUILD +++ b/tensorflow/compiler/xla/service/cpu/BUILD @@ -37,6 +37,7 @@ cc_library( srcs = ["cpu_transfer_manager.cc"], hdrs = ["cpu_transfer_manager.h"], deps = [ + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:literal_util", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:status_macros", @@ -53,29 +54,6 @@ cc_library( alwayslink = True, # Contains per-platform transfer manager registration ) -cc_library( - name = "external_constant_pool", - srcs = ["external_constant_pool.cc"], - hdrs = ["external_constant_pool.h"], - deps = [ - "//tensorflow/compiler/xla:literal_util", - "//tensorflow/compiler/xla:shape_util", - "//tensorflow/compiler/xla:util", - "//tensorflow/core:lib", - ], -) - -tf_cc_test( - name = "external_constant_pool_test", - srcs = ["external_constant_pool_test.cc"], - deps = [ - ":external_constant_pool", - "//tensorflow/compiler/xla:shape_util", - "//tensorflow/compiler/xla/tests:xla_internal_test_main", - "//tensorflow/core:test", - ], -) - cc_library( name = "cpu_compiler", srcs = ["cpu_compiler.cc"], @@ -95,7 +73,7 @@ cc_library( ":ir_emitter", ":parallel_task_assignment", ":simple_orc_jit", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:protobuf_util", "//tensorflow/compiler/xla:status_macros", "//tensorflow/compiler/xla:statusor", @@ -112,7 +90,6 @@ cc_library( "//tensorflow/compiler/xla/service:dot_decomposer", "//tensorflow/compiler/xla/service:executable", "//tensorflow/compiler/xla/service:flatten_call_graph", - "//tensorflow/compiler/xla/service:gather_expander", "//tensorflow/compiler/xla/service:hlo", "//tensorflow/compiler/xla/service:hlo_constant_folding", "//tensorflow/compiler/xla/service:hlo_cse", @@ -152,7 +129,7 @@ cc_library( "@llvm//:x86_code_gen", # fixdeps: keep "@llvm//:x86_disassembler", # fixdeps: keep ] + select({ - "@org_tensorflow//tensorflow:linux_ppc64le": [ + "//tensorflow:linux_ppc64le": [ "@llvm//:powerpc_disassembler", "@llvm//:powerpc_code_gen", ], @@ -175,7 +152,6 @@ cc_library( ":cpu_runtime", ":custom_call_target_registry", ":disassembler", - ":external_constant_pool", ":orc_jit_memory_mapper", ":runtime_fp16", ":runtime_conv2d", @@ -256,7 +232,6 @@ cc_library( ":cpu_options", ":cpu_runtime", ":dot_op_emitter", - ":external_constant_pool", ":ir_emission_utils", ":ir_function", ":parallel_loop_emitter", @@ -273,6 +248,7 @@ cc_library( "//tensorflow/compiler/xla/service:buffer_assignment", "//tensorflow/compiler/xla/service:elemental_ir_emitter", "//tensorflow/compiler/xla/service:hlo", + "//tensorflow/compiler/xla/service:hlo_casting_utils", "//tensorflow/compiler/xla/service:hlo_module_config", "//tensorflow/compiler/xla/service:name_uniquer", "//tensorflow/compiler/xla/service/llvm_ir:alias_analysis", @@ -379,7 +355,7 @@ tf_cc_binary( srcs = ["sample_harness.cc"], deps = [ "//tensorflow/compiler/xla:array4d", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:statusor", "//tensorflow/compiler/xla:types", "//tensorflow/compiler/xla:xla_data_proto", @@ -741,7 +717,7 @@ tf_cc_test( deps = [ ":cpu_layout_assignment", ":target_machine_features_fake", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_layout", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:test", @@ -833,7 +809,7 @@ tf_cc_test( ":cpu_executable", ":parallel_task_assignment", ":target_machine_features_fake", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_layout", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:test", @@ -916,7 +892,7 @@ tf_cc_test( srcs = ["cpu_copy_insertion_test.cc"], deps = [ ":cpu_copy_insertion", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:test", "//tensorflow/compiler/xla:test_helpers", diff --git a/tensorflow/compiler/xla/service/cpu/conv_canonicalization_test.cc b/tensorflow/compiler/xla/service/cpu/conv_canonicalization_test.cc index 375b017b09263c20c1b1ef8329f7e2f6a573dda4..547d4c696da5cfdde3dece03250ae5fa51c92f25 100644 --- a/tensorflow/compiler/xla/service/cpu/conv_canonicalization_test.cc +++ b/tensorflow/compiler/xla/service/cpu/conv_canonicalization_test.cc @@ -60,11 +60,11 @@ TEST_F(ConvCanonicalizationTest, NonCanonicalToCanonical) { auto builder = HloComputation::Builder(TestName()); // The input dimensions are in CNHW order. auto input = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR4FromArray4D(Array4D( + LiteralUtil::CreateR4FromArray4D(Array4D( kInputFeatureCount, kBatchSize, kInputSize, kInputSize)))); // The kernel dimensions are in OIHW order. auto kernel = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR4FromArray4D(Array4D( + LiteralUtil::CreateR4FromArray4D(Array4D( kOutputFeatureCount, kInputFeatureCount, kWindowSize, kWindowSize)))); ConvolutionDimensionNumbers dnums; @@ -122,11 +122,11 @@ TEST_F(ConvCanonicalizationTest, CanonicalStaysTheSame) { auto builder = HloComputation::Builder(TestName()); // The input dimensions are in NHWC order. auto input = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR4FromArray4D(Array4D( + LiteralUtil::CreateR4FromArray4D(Array4D( kBatchSize, kInputSize, kInputSize, kInputFeatureCount)))); // The kernel dimensions are in HWIO order. auto kernel = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR4FromArray4D(Array4D( + LiteralUtil::CreateR4FromArray4D(Array4D( kWindowSize, kWindowSize, kInputFeatureCount, kOutputFeatureCount)))); ConvolutionDimensionNumbers dnums; diff --git a/tensorflow/compiler/xla/service/cpu/cpu_compiler.cc b/tensorflow/compiler/xla/service/cpu/cpu_compiler.cc index 52da9d6eac7e92188774107dd054396ebd9cd8db..29fa29d33ad62a76191cef2de22ccc094b0cf35b 100644 --- a/tensorflow/compiler/xla/service/cpu/cpu_compiler.cc +++ b/tensorflow/compiler/xla/service/cpu/cpu_compiler.cc @@ -30,6 +30,7 @@ limitations under the License. #include "llvm/ADT/Triple.h" #include "llvm/IR/Function.h" #include "llvm/IR/LLVMContext.h" +#include "llvm/IR/Mangler.h" #include "llvm/IR/Module.h" #include "llvm/IR/Verifier.h" #include "llvm/Object/ObjectFile.h" @@ -38,7 +39,7 @@ limitations under the License. #include "llvm/Support/TargetSelect.h" #include "llvm/Target/TargetMachine.h" #include "llvm/Target/TargetOptions.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/map_util.h" #include "tensorflow/compiler/xla/protobuf_util.h" #include "tensorflow/compiler/xla/ptr_util.h" @@ -66,7 +67,6 @@ limitations under the License. #include "tensorflow/compiler/xla/service/dfs_hlo_visitor_with_default.h" #include "tensorflow/compiler/xla/service/dot_decomposer.h" #include "tensorflow/compiler/xla/service/flatten_call_graph.h" -#include "tensorflow/compiler/xla/service/gather_expander.h" #include "tensorflow/compiler/xla/service/hlo.pb.h" #include "tensorflow/compiler/xla/service/hlo_computation.h" #include "tensorflow/compiler/xla/service/hlo_constant_folding.h" @@ -269,6 +269,7 @@ Status CpuCompiler::RunHloPasses(HloModule* module, bool is_aot_compile, /*is_layout_sensitive=*/false, [](const Shape&, const Shape&) { return false; }, /*enable_dot_strength_reduction=*/false); + pass.AddPass(); // BatchNormExpander can create zero-sized ops, so zero-sized HLO // elimination has to come after that pass. @@ -296,8 +297,6 @@ Status CpuCompiler::RunHloPasses(HloModule* module, bool is_aot_compile, pipeline.AddPass(/*is_layout_sensitive=*/false); pipeline.AddPass(); - pipeline.AddPass(); - ReducePrecisionInsertion::AddPasses( &pipeline, module->config().debug_options(), ReducePrecisionInsertion::PassTiming::AFTER_FUSION); @@ -306,11 +305,16 @@ Status CpuCompiler::RunHloPasses(HloModule* module, bool is_aot_compile, module->mutable_entry_computation_layout(), &target_machine_features); // The LayoutAssignment pass may leave behind kCopy instructions which are // duplicate or NOPs, so remove them with algebraic simplification and CSE. - pipeline.AddPass>( - /*is_layout_sensitive=*/true, - [](const Shape&, const Shape&) { return true; }, - /*enable_dot_strength_reduction=*/false); - pipeline.AddPass(/*is_layout_sensitive=*/true); + { + auto& pass = pipeline.AddPass>( + "after layout assignement"); + pass.AddPass>( + /*is_layout_sensitive=*/true, + [](const Shape&, const Shape&) { return true; }, + /*enable_dot_strength_reduction=*/false); + pass.AddPass(); + pass.AddPass(/*is_layout_sensitive=*/true); + } pipeline.AddPass(BF16, F32); // Outline ops in the entry computation into calls to subcomputations. const int max_parallelism = @@ -578,7 +582,7 @@ StatusOr> CpuCompiler::RunBackend( IrEmitter ir_emitter(*module, *assignment, llvm_module.get(), std::move(instruction_to_profile_idx), std::move(computation_to_profile_idx), - &target_machine_features, jit->external_constant_pool()); + &target_machine_features); for (auto embedded_computation : entry_computation->MakeEmbeddedComputationsList()) { @@ -601,7 +605,13 @@ StatusOr> CpuCompiler::RunBackend( /*is_top_level_computation=*/true, &module_sequence.at(entry_computation))); - string function_name = llvm_ir::AsString(entry_function->getName()); + string function_name = [&]() { + llvm::SmallVector function_name_vector; + llvm::Mangler::getNameWithPrefix( + function_name_vector, entry_function->getName(), jit->data_layout()); + return string(function_name_vector.begin(), function_name_vector.end()); + }(); + string ir_module_string; if (embed_ir_in_executable) { ir_module_string = llvm_ir::DumpModuleToString(*llvm_module); @@ -765,8 +775,7 @@ CpuCompiler::CompileAheadOfTime(std::vector> modules, IrEmitter ir_emitter(*module, *assignment, &llvm_module, std::move(instruction_to_profile_idx), std::move(computation_to_profile_idx), - &target_machine_features, - /*external_constant_pool=*/nullptr); + &target_machine_features); HloComputation* computation = module->entry_computation(); for (auto embedded_computation : computation->MakeEmbeddedComputationsList()) { diff --git a/tensorflow/compiler/xla/service/cpu/cpu_copy_insertion_test.cc b/tensorflow/compiler/xla/service/cpu/cpu_copy_insertion_test.cc index a05a26941786cbf404c4685abb098c9ac8caaa09..4db7fa446ea9188940f930bcadf753bd3e6b79e3 100644 --- a/tensorflow/compiler/xla/service/cpu/cpu_copy_insertion_test.cc +++ b/tensorflow/compiler/xla/service/cpu/cpu_copy_insertion_test.cc @@ -16,7 +16,7 @@ limitations under the License. #include "tensorflow/compiler/xla/service/cpu/cpu_copy_insertion.h" #include "tensorflow/compiler/xla/legacy_flags/debug_options_flags.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/service/hlo_computation.h" #include "tensorflow/compiler/xla/service/hlo_instruction.h" #include "tensorflow/compiler/xla/service/hlo_matchers.h" @@ -74,14 +74,14 @@ TEST_F(CpuCopyInsertionTest, WhileBodyWithConstantRoot) { body_builder.AddInstruction( HloInstruction::CreateParameter(0, scalar_shape_, "param")); body_builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(123.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(123.0))); HloComputation* body = module->AddEmbeddedComputation(body_builder.Build()); auto cond_builder = HloComputation::Builder("condition"); cond_builder.AddInstruction( HloInstruction::CreateParameter(0, scalar_shape_, "param")); cond_builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(false))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(false))); HloComputation* condition = module->AddEmbeddedComputation(cond_builder.Build()); @@ -114,7 +114,7 @@ TEST_F(CpuCopyInsertionTest, TupleCall) { auto sub_param = sub_builder.AddInstruction( HloInstruction::CreateParameter(0, scalar_shape_, "param")); auto constant = sub_builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(123.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(123.0))); auto add = sub_builder.AddInstruction(HloInstruction::CreateBinary( scalar_shape_, HloOpcode::kAdd, sub_param, constant)); sub_builder.AddInstruction( diff --git a/tensorflow/compiler/xla/service/cpu/cpu_instruction_fusion_test.cc b/tensorflow/compiler/xla/service/cpu/cpu_instruction_fusion_test.cc index 97e10a89a209c057685709e7a5034052ff4376ed..991b14f17dbc8cd061af98e032824d3f7075e78b 100644 --- a/tensorflow/compiler/xla/service/cpu/cpu_instruction_fusion_test.cc +++ b/tensorflow/compiler/xla/service/cpu/cpu_instruction_fusion_test.cc @@ -282,7 +282,7 @@ class OpcodeFusionTest : public InstructionFusionTest { builder.AddInstruction(HloInstruction::CreateParameter( 0, ShapeUtil::MakeShape(F32, {}), "arg0")); HloInstruction* one = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0))); builder.AddInstruction(HloInstruction::CreateBinary( ShapeUtil::MakeShape(F32, {}), HloOpcode::kAdd, arg0, one)); return module->AddEmbeddedComputation(builder.Build()); @@ -501,8 +501,8 @@ TEST_F(OpcodeFusionTest, UnaryMapOfExp) { HloInstruction* exp = builder.AddInstruction( HloInstruction::CreateUnary(shape, HloOpcode::kExp, param0)); - builder.AddInstruction(HloInstruction::CreateMap( - shape, {exp}, CreateAdderToOne(module.get()), /*static_operands=*/{})); + builder.AddInstruction( + HloInstruction::CreateMap(shape, {exp}, CreateAdderToOne(module.get()))); module->AddEntryComputation(builder.Build()); @@ -525,8 +525,8 @@ TEST_F(OpcodeFusionTest, BinaryMapOfExps) { HloInstruction* exp1 = builder.AddInstruction( HloInstruction::CreateUnary(shape, HloOpcode::kExp, param1)); - builder.AddInstruction(HloInstruction::CreateMap( - shape, {exp0, exp1}, CreateMax(module.get()), /*static_operands=*/{})); + builder.AddInstruction( + HloInstruction::CreateMap(shape, {exp0, exp1}, CreateMax(module.get()))); module->AddEntryComputation(builder.Build()); @@ -595,7 +595,7 @@ TEST_F(OpcodeFusionTest, MessOfFusileNodes) { auto pad = builder.AddInstruction(HloInstruction::CreatePad( ShapeUtil::MakeShape(S32, {5}), idx_choice, builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(0))), + HloInstruction::CreateConstant(LiteralUtil::CreateR0(0))), padding_config)); auto slice = builder.AddInstruction(HloInstruction::CreateDynamicSlice( diff --git a/tensorflow/compiler/xla/service/cpu/cpu_layout_assignment_test.cc b/tensorflow/compiler/xla/service/cpu/cpu_layout_assignment_test.cc index 429fc7b78608da0e9cd794ac294851b326f5be24..3681d12d8da818d06d2f690024008c9ccb896286 100644 --- a/tensorflow/compiler/xla/service/cpu/cpu_layout_assignment_test.cc +++ b/tensorflow/compiler/xla/service/cpu/cpu_layout_assignment_test.cc @@ -21,7 +21,7 @@ limitations under the License. #include #include "tensorflow/compiler/xla/layout_util.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/service/algebraic_simplifier.h" #include "tensorflow/compiler/xla/service/computation_layout.h" #include "tensorflow/compiler/xla/service/cpu/target_machine_features_fake.h" diff --git a/tensorflow/compiler/xla/service/cpu/cpu_transfer_manager.cc b/tensorflow/compiler/xla/service/cpu/cpu_transfer_manager.cc index b877b295814a7e13569a1837ed3e1787f2fc3f56..156166bf2b1ea6d3821da8f67ea2b2eca6825ca6 100644 --- a/tensorflow/compiler/xla/service/cpu/cpu_transfer_manager.cc +++ b/tensorflow/compiler/xla/service/cpu/cpu_transfer_manager.cc @@ -19,6 +19,7 @@ limitations under the License. #include #include +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/literal_util.h" #include "tensorflow/compiler/xla/service/cpu/cpu_runtime.h" #include "tensorflow/compiler/xla/shape_util.h" @@ -180,7 +181,7 @@ Status CpuTransferManager::TransferLiteralFromOutfeed( tensorflow::gtl::ArraySlice dimensions( tensorflow::bit_cast(literal_shape.dimensions().data()), literal_shape.dimensions().size()); - *literal = std::move(*Literal::CreateFromDimensions( + *literal = std::move(*LiteralUtil::CreateFromDimensions( literal_shape.element_type(), dimensions)); TF_ASSIGN_OR_RETURN(Shape received_shape, TransferArrayBufferFromOutfeed( @@ -211,7 +212,7 @@ Status CpuTransferManager::TransferLiteralFromOutfeed( tensorflow::bit_cast( tuple_element_shape.dimensions().data()), tuple_element_shape.dimensions().size()); - auto empty = Literal::CreateFromDimensions( + auto empty = LiteralUtil::CreateFromDimensions( tuple_element_shape.element_type(), dimensions); int64 size = GetByteSizeRequirement(tuple_element_shape); buffer_data.push_back({empty->untyped_data(), size}); @@ -232,7 +233,7 @@ Status CpuTransferManager::TransferLiteralFromOutfeed( for (int64 i = 0; i < literal_shape.tuple_shapes_size(); ++i) { *elements[i]->mutable_shape_do_not_use() = received_shape.tuple_shapes(i); } - *literal = std::move(*Literal::MakeTupleOwned(std::move(elements))); + *literal = std::move(*LiteralUtil::MakeTupleOwned(std::move(elements))); TF_RET_CHECK(ShapeUtil::Equal(literal->shape(), literal_shape)); return Status::OK(); } diff --git a/tensorflow/compiler/xla/service/cpu/cpu_transfer_manager.h b/tensorflow/compiler/xla/service/cpu/cpu_transfer_manager.h index 6dfc666f09dfa6df740cd54bea0957e3144181bc..593575c0fdaddc71cd6bd844fd179096a9fb0fdc 100644 --- a/tensorflow/compiler/xla/service/cpu/cpu_transfer_manager.h +++ b/tensorflow/compiler/xla/service/cpu/cpu_transfer_manager.h @@ -39,13 +39,14 @@ class CpuTransferManager : public GenericTransferManager { Status TransferLiteralToInfeed(se::StreamExecutor* executor, const LiteralSlice& literal) override; - Status TransferBufferToInfeed(se::StreamExecutor* executor, int64 size, - const void* source) override; Status TransferLiteralFromOutfeed(se::StreamExecutor* executor, const Shape& literal_shape, Literal* literal) override; private: + Status TransferBufferToInfeed(se::StreamExecutor* executor, int64 size, + const void* source); + // Transfers infeed data to device. InfeedBuffer->Done() must be // called to clean up the memory allocated for InfeedBuffer. StatusOr TransferBufferToInfeedInternal( diff --git a/tensorflow/compiler/xla/service/cpu/external_constant_pool.cc b/tensorflow/compiler/xla/service/cpu/external_constant_pool.cc deleted file mode 100644 index c56286559158758ca6db5ae097729286bde346f0..0000000000000000000000000000000000000000 --- a/tensorflow/compiler/xla/service/cpu/external_constant_pool.cc +++ /dev/null @@ -1,50 +0,0 @@ -/* Copyright 2017 The TensorFlow Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. -==============================================================================*/ - -#include "tensorflow/compiler/xla/service/cpu/external_constant_pool.h" - -#include -#include -#include - -#include "tensorflow/compiler/xla/map_util.h" -#include "tensorflow/compiler/xla/ptr_util.h" -#include "tensorflow/compiler/xla/shape_util.h" -#include "tensorflow/core/lib/gtl/flatset.h" - -namespace xla { -namespace cpu { -void ExternalConstantPool::Insert(string name, const LiteralSlice& literal, - int64 alignment) { - CHECK(!ShapeUtil::IsTuple(literal.shape())); - CHECK(alignment > 0 && IsPowerOfTwo(static_cast(alignment))); - CHECK(entries_.find(name) == entries_.end()); - - const int64 literal_size = ShapeUtil::ByteSizeOf(literal.shape()); - void* raw_pointer = tensorflow::port::AlignedMalloc( - literal_size, std::max(alignment, sizeof(void*))); - CHECK(raw_pointer != nullptr) << "failed to allocate " << literal_size - << " bytes with alignment of " << alignment; - - std::memcpy(raw_pointer, literal.untyped_data(), literal_size); - entries_.emplace(std::move(name), static_cast(raw_pointer)); -} - -const uint8* ExternalConstantPool::Find(const string& name) { - auto it = entries_.find(name); - return it == entries_.end() ? nullptr : it->second.get(); -} -} // namespace cpu -} // namespace xla diff --git a/tensorflow/compiler/xla/service/cpu/external_constant_pool.h b/tensorflow/compiler/xla/service/cpu/external_constant_pool.h deleted file mode 100644 index 0677f5f0b58005079890052a426e5f48c5d09ed1..0000000000000000000000000000000000000000 --- a/tensorflow/compiler/xla/service/cpu/external_constant_pool.h +++ /dev/null @@ -1,65 +0,0 @@ -/* Copyright 2017 The TensorFlow Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. -==============================================================================*/ - -#ifndef TENSORFLOW_COMPILER_XLA_SERVICE_CPU_EXTERNAL_CONSTANT_POOL_H_ -#define TENSORFLOW_COMPILER_XLA_SERVICE_CPU_EXTERNAL_CONSTANT_POOL_H_ - -#include - -#include "tensorflow/compiler/xla/literal_util.h" -#include "tensorflow/core/lib/gtl/flatmap.h" -#include "tensorflow/core/platform/mem.h" - -namespace xla { -namespace cpu { -// An ExternalConstantPool maintains a set of constants kept external to -// generated LLVM IR. These constants are accessed from the IR via globals with -// extern linkage. This current incarnation of ExternalConstantPool only -// supports the JIT CPU backend; the AOT backend is not supported. -// -// Implementation-wise, this is a simple wrapper around a map of strings to byte -// buffers. This simply implementation works in a JIT scenario. This class -// will have to become smarter if we decide to support external constant pools -// on AOT compiles in the future. -class ExternalConstantPool { - public: - // Inserts a buffer with the contents of `literal` into the constant pool with - // the name `name`. It is an error to try to insert two constants with the - // same `name` into the same constant pool. The buffer for literal is aligned - // to `aligment` bytes, and `alignment` must be a power of 2. - // - // The constant pool copies out the contents of `literal` into a buffer it - // owns -- it does not keep pointers to `literal`, or to memory owned by - // `literal`. - void Insert(string name, const LiteralSlice& literal, int64 alignment); - - // Find the constant with name `name` in this constant pool. If there isn't - // such constant, return nullptr. - const uint8* Find(const string& name); - - private: - // We need to `AlignedFree` pointers allocated into `entries_` since we - // allocate them with `AlignedMalloc`. - struct FreeDeleter { - void operator()(void* ptr) { tensorflow::port::AlignedFree(ptr); } - }; - - tensorflow::gtl::FlatMap> - entries_; -}; -} // namespace cpu -} // namespace xla - -#endif // TENSORFLOW_COMPILER_XLA_SERVICE_CPU_EXTERNAL_CONSTANT_POOL_H_ diff --git a/tensorflow/compiler/xla/service/cpu/external_constant_pool_test.cc b/tensorflow/compiler/xla/service/cpu/external_constant_pool_test.cc deleted file mode 100644 index 9290a4e5dfc03ddb86e9d82f1f0f4f9a8ceebb88..0000000000000000000000000000000000000000 --- a/tensorflow/compiler/xla/service/cpu/external_constant_pool_test.cc +++ /dev/null @@ -1,82 +0,0 @@ -/* Copyright 2017 The TensorFlow Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. -==============================================================================*/ - -#include "tensorflow/compiler/xla/service/cpu/external_constant_pool.h" -#include "tensorflow/compiler/xla/layout_util.h" -#include "tensorflow/core/platform/test.h" - -namespace xla { -namespace cpu { -namespace { -class ExternalConstantPoolTest : public ::testing::Test {}; - -template -T GetFromBuffer(const uint8* buffer, int64 index) { - T result; - std::memcpy(&result, buffer + index * sizeof(T), sizeof(T)); - return result; -} - -TEST(ExternalConstantPoolTest, Basic) { - ExternalConstantPool constant_pool; - EXPECT_EQ(constant_pool.Find("name-0"), nullptr); - const auto literal = Literal::CreateR2({{1, 2}, {3, 4}}); - constant_pool.Insert("name-0", *literal, 4); - const uint8* constant = constant_pool.Find("name-0"); - ASSERT_NE(constant, nullptr); - - EXPECT_EQ(GetFromBuffer(constant, 0), 1); - EXPECT_EQ(GetFromBuffer(constant, 1), 2); - EXPECT_EQ(GetFromBuffer(constant, 2), 3); - EXPECT_EQ(GetFromBuffer(constant, 3), 4); - - EXPECT_EQ(constant_pool.Find("name-1"), nullptr); -} - -TEST(ExternalConstantPoolTest, RowMinorLayout) { - ExternalConstantPool constant_pool; - EXPECT_EQ(constant_pool.Find("name-0"), nullptr); - const auto literal = Literal::CreateR2WithLayout( - {{1, 2}, {3, 4}}, LayoutUtil::MakeLayout({0, 1})); - constant_pool.Insert("name-0", *literal, 4); - const uint8* constant = constant_pool.Find("name-0"); - ASSERT_NE(constant, nullptr); - - EXPECT_EQ(GetFromBuffer(constant, 0), 1); - EXPECT_EQ(GetFromBuffer(constant, 1), 3); - EXPECT_EQ(GetFromBuffer(constant, 2), 2); - EXPECT_EQ(GetFromBuffer(constant, 3), 4); -} - -TEST(ExternalConstantPoolTest, Alignment) { - ExternalConstantPool constant_pool; - EXPECT_EQ(constant_pool.Find("name-0"), nullptr); - - for (int i = 0; i < 8; i++) { - int64 alignment = 1 << i; - string name = tensorflow::strings::StrCat("name-", i); - - const auto literal = Literal::CreateR2({{1, 2}, {3, 4}}); - constant_pool.Insert(name, *literal, alignment); - - const uint8* constant = constant_pool.Find(name); - ASSERT_NE(constant, nullptr); - EXPECT_EQ(reinterpret_cast(constant) % alignment, 0); - } -} - -} // namespace -} // namespace cpu -} // namespace xla diff --git a/tensorflow/compiler/xla/service/cpu/ir_emitter.cc b/tensorflow/compiler/xla/service/cpu/ir_emitter.cc index 75e8e9a835000339a80e7c140ba60e5eb0698280..2ad41374d39520f1186581b94e96d436b2fed557 100644 --- a/tensorflow/compiler/xla/service/cpu/ir_emitter.cc +++ b/tensorflow/compiler/xla/service/cpu/ir_emitter.cc @@ -48,6 +48,8 @@ limitations under the License. #include "tensorflow/compiler/xla/service/cpu/shape_partition.h" #include "tensorflow/compiler/xla/service/cpu/simple_orc_jit.h" #include "tensorflow/compiler/xla/service/elemental_ir_emitter.h" +#include "tensorflow/compiler/xla/service/hlo_casting_utils.h" +#include "tensorflow/compiler/xla/service/hlo_instructions.h" #include "tensorflow/compiler/xla/service/hlo_opcode.h" #include "tensorflow/compiler/xla/service/llvm_ir/fused_ir_emitter.h" #include "tensorflow/compiler/xla/service/llvm_ir/llvm_loop.h" @@ -83,8 +85,7 @@ IrEmitter::IrEmitter( llvm::Module* llvm_module, std::unordered_map instruction_to_profile_idx, std::unordered_map computation_to_profile_idx, - const TargetMachineFeatures* target_machine_features, - ExternalConstantPool* external_constant_pool) + const TargetMachineFeatures* target_machine_features) : assignment_(assignment), module_(llvm_module), arch_type_(llvm::Triple(llvm_module->getTargetTriple()).getArch()), @@ -94,8 +95,7 @@ IrEmitter::IrEmitter( alias_analysis_(hlo_module, assignment, &llvm_module->getContext()), hlo_module_config_(hlo_module.config()), is_top_level_computation_(false), - target_machine_features_(*target_machine_features), - external_constant_pool_(external_constant_pool) { + target_machine_features_(*target_machine_features) { ir_builder_.setFastMathFlags(llvm_ir::GetFastMathFlags( /*fast_math_enabled=*/hlo_module_config_.debug_options() .xla_enable_fast_math())); @@ -161,45 +161,18 @@ Status IrEmitter::HandleBitcast(HloInstruction* bitcast) { } llvm::Constant* IrEmitter::EmitGlobalForLiteral(const Literal& literal) { - llvm::Constant* result; - - // We avoid creating large constants in the LLVM IR since LLVM is not - // efficient for large constant arrays. We still emit "small enough" constant - // arrays into the Ir, in the off chance the LLVM optimizer can do something - // interesting with it. - // - // TODO(b/29904935): Remove the large constant pool. - const int kMaxInternalConstantSizeInBytes = 128; - if (external_constant_pool_ && - ByteSizeOf(literal.shape()) >= kMaxInternalConstantSizeInBytes) { - string global_name = tensorflow::strings::StrCat( - "constant_global_", external_global_constant_counter_++); - llvm::GlobalVariable* result_global = new llvm::GlobalVariable( - /*Module=*/*module_, - /*Type=*/IrShapeType(literal.shape()), - /*isConstant=*/true, - /*Linkage=*/llvm::GlobalValue::ExternalLinkage, - /*Initializer=*/nullptr, - /*Name=*/AsStringRef(global_name)); - result_global->setAlignment(MinimumAlignmentForShape(literal.shape())); - external_constant_pool_->Insert(global_name, literal, - MinimumAlignmentForShape(literal.shape())); - result = result_global; - } else { - llvm::Constant* initializer = - llvm_ir::ConvertLiteralToIrConstant(literal, module_); - llvm::GlobalVariable* result_global = new llvm::GlobalVariable( - /*Module=*/*module_, - /*Type=*/initializer->getType(), - /*isConstant=*/true, - /*Linkage=*/llvm::GlobalValue::PrivateLinkage, - /*Initializer=*/initializer, - /*Name=*/""); - result_global->setAlignment(MinimumAlignmentForShape(literal.shape())); - result = llvm::ConstantExpr::getBitCast( - result_global, IrShapeType(literal.shape())->getPointerTo()); - } - return result; + llvm::Constant* initializer = + llvm_ir::ConvertLiteralToIrConstant(literal, module_); + llvm::GlobalVariable* result_global = new llvm::GlobalVariable( + /*Module=*/*module_, + /*Type=*/initializer->getType(), + /*isConstant=*/true, + /*Linkage=*/llvm::GlobalValue::PrivateLinkage, + /*Initializer=*/initializer, + /*Name=*/""); + result_global->setAlignment(MinimumAlignmentForShape(literal.shape())); + return llvm::ConstantExpr::getBitCast( + result_global, IrShapeType(literal.shape())->getPointerTo()); } Status IrEmitter::HandleConstant(HloInstruction* constant) { @@ -306,45 +279,60 @@ Status IrEmitter::HandleGetTupleElement(HloInstruction* get_tuple_element) { Status IrEmitter::HandleSelect(HloInstruction* select) { auto pred = select->operand(0); - auto on_true = select->operand(1); - auto on_false = select->operand(2); TF_RET_CHECK(pred->shape().element_type() == PRED); - - if (ShapeUtil::IsTuple(select->shape())) { - TF_RETURN_IF_ERROR(EmitTargetAddressForOp(select)); - llvm_ir::EmitTupleSelect( - GetIrArrayFor(select), GetIrArrayFor(pred), GetEmittedValueFor(on_true), - GetEmittedValueFor(on_false), &ir_builder_, module_); - return Status::OK(); - } - return DefaultAction(select); } -Status IrEmitter::HandleInfeed(HloInstruction* infeed) { - VLOG(2) << "HandleInfeed: " << infeed->ToString(); +Status IrEmitter::HandleTupleSelect(HloInstruction* tuple_select) { + auto pred = tuple_select->operand(0); + auto on_true = tuple_select->operand(1); + auto on_false = tuple_select->operand(2); + TF_RET_CHECK(pred->shape().element_type() == PRED); + TF_RET_CHECK(ShapeUtil::IsScalar(pred->shape())); + TF_RET_CHECK(ShapeUtil::IsTuple(tuple_select->shape())); + TF_RETURN_IF_ERROR(EmitTargetAddressForOp(tuple_select)); + llvm_ir::EmitTupleSelect(GetIrArrayFor(tuple_select), GetIrArrayFor(pred), + GetEmittedValueFor(on_true), + GetEmittedValueFor(on_false), &ir_builder_, module_); + return Status::OK(); +} - const Shape& shape = infeed->shape(); +Status IrEmitter::HandleInfeed(HloInstruction* instruction) { + HloInfeedInstruction* infeed = Cast(instruction); + VLOG(2) << "HandleInfeed: " << infeed->ToString(); - // The infeed operation produces data (dequeued from the infeed queue) at this - // address, which has been provided by buffer assignment. + // The infeed operation produces a two-element tuple containing data and a + // token value. HloInfeedInstruction::infeed_shape gives us the data shape. + const Shape& data_shape = infeed->infeed_shape(); + DCHECK(ShapeUtil::Equal(data_shape, + ShapeUtil::GetTupleElementShape(infeed->shape(), 0))); TF_RETURN_IF_ERROR(EmitTargetAddressForOp(infeed)); - llvm_ir::IrArray infeed_array = GetIrArrayFor(infeed); - if (ShapeUtil::IsTuple(shape)) { - TF_RET_CHECK(!ShapeUtil::IsNestedTuple(shape)); + // Write the tuple index table. + TF_ASSIGN_OR_RETURN(BufferAllocation::Slice data_slice, + assignment_.GetUniqueSlice(infeed, {0})); + llvm::Value* data_address = EmitTempBufferPointer(data_slice, data_shape); + TF_ASSIGN_OR_RETURN(BufferAllocation::Slice token_slice, + assignment_.GetUniqueSlice(infeed, {1})); + llvm::Value* token_address = EmitTempBufferPointer( + token_slice, ShapeUtil::GetTupleElementShape(infeed->shape(), 1)); + llvm_ir::EmitTuple(GetIrArrayFor(infeed), {data_address, token_address}, + &ir_builder_, module_); + + if (ShapeUtil::IsTuple(data_shape)) { + TF_RET_CHECK(!ShapeUtil::IsNestedTuple(data_shape)); // For a tuple, we first copy each of the internal elements to // their corresponding target locations. We then construct the // tuple outer buffer containing pointers to the internal // elements. std::vector tuple_element_addresses; - for (int64 i = 0; i < shape.tuple_shapes_size(); ++i) { + for (int64 i = 0; i < data_shape.tuple_shapes_size(); ++i) { TF_ASSIGN_OR_RETURN(BufferAllocation::Slice buffer, - assignment_.GetUniqueSlice(infeed, {i})); + assignment_.GetUniqueSlice(infeed, {0, i})); const Shape& tuple_element_shape = - ShapeUtil::GetTupleElementShape(shape, i); + ShapeUtil::GetTupleElementShape(data_shape, i); // Only the outer tuple buffer's target address is obtained from // GetEmittedValueFor, to handle the case when Infeed is the root @@ -359,11 +347,11 @@ Status IrEmitter::HandleInfeed(HloInstruction* infeed) { tuple_element_addresses.push_back(tuple_element_address); } - llvm_ir::EmitTuple(infeed_array, tuple_element_addresses, &ir_builder_, - module_); + llvm_ir::EmitTuple(llvm_ir::IrArray(data_address, data_shape), + tuple_element_addresses, &ir_builder_, module_); } else { - TF_RETURN_IF_ERROR(EmitXfeedTransfer(XfeedKind::kInfeed, shape, - GetEmittedValueFor(infeed))); + TF_RETURN_IF_ERROR( + EmitXfeedTransfer(XfeedKind::kInfeed, data_shape, data_address)); } return Status::OK(); @@ -488,42 +476,111 @@ Status IrEmitter::HandleTuple(HloInstruction* tuple) { return Status::OK(); } +StatusOr IrEmitter::EmitTargetElementLoopBodyForMap( + HloMapInstruction* map, const llvm_ir::IrArray::Index& index) { + llvm::Function* mapped_ir_function = + FindOrDie(emitted_functions_, map->to_apply()); + std::vector parameter_addresses; + for (const HloInstruction* operand : map->operands()) { + const llvm_ir::IrArray& array = GetIrArrayFor(operand); + parameter_addresses.push_back( + array.EmitArrayElementAddress(index, &ir_builder_)); + } + return EmitElementFunctionCall(mapped_ir_function, map->shape(), + parameter_addresses, "map_function"); +} + Status IrEmitter::HandleMap(HloInstruction* map) { - gtl::ArraySlice operands(map->operands()); - HloComputation* function = map->to_apply(); - // The called computation should have been emitted previously. - llvm::Function* mapped_ir_function = FindOrDie(emitted_functions_, function); - - return EmitTargetElementLoop(map, [this, map, operands, mapped_ir_function]( - const llvm_ir::IrArray::Index& index) { - std::vector parameter_addresses; - for (const HloInstruction* operand : operands) { - const llvm_ir::IrArray& array = GetIrArrayFor(operand); - parameter_addresses.push_back( - array.EmitArrayElementAddress(index, &ir_builder_)); - } - return EmitElementFunctionCall(mapped_ir_function, map->shape(), - parameter_addresses, "map_function"); + return EmitTargetElementLoop(map, [&](const llvm_ir::IrArray::Index& index) { + return EmitTargetElementLoopBodyForMap(Cast(map), index); }); } -Status IrEmitter::HandleReduceWindow(HloInstruction* reduce_window) { - auto operand = reduce_window->operand(0); +StatusOr IrEmitter::EmitTargetElementLoopBodyForReduceWindow( + HloReduceWindowInstruction* reduce_window, + const llvm_ir::IrArray::Index& index) { + const HloInstruction* operand = reduce_window->operand(0); const Window& window = reduce_window->window(); HloComputation* function = reduce_window->to_apply(); + // The called computation should have been emitted previously. + llvm::Function* reducer_function = FindOrDie(emitted_functions_, function); + + // We fold inputs into the accumulator and initialize it to + // the initial value on the reduce_window. + PrimitiveType operand_element_type = operand->shape().element_type(); + llvm::Value* accumulator_address = llvm_ir::EmitAllocaAtFunctionEntry( + llvm_ir::PrimitiveTypeToIrType(operand_element_type, module_), + "reduce_window_accumulator_address", &ir_builder_, + MinimumAlignmentForPrimitiveType(operand_element_type)); + ir_builder_.CreateStore( + ir_builder_.CreateLoad(GetEmittedValueFor(reduce_window->operand(1))), + accumulator_address); + + llvm_ir::ForLoopNest loops(IrName(reduce_window, "inner"), &ir_builder_); + std::vector window_size; + for (const auto& dim : window.dimensions()) { + window_size.push_back(dim.size()); + } + const llvm_ir::IrArray::Index window_index = loops.AddLoopsForShape( + ShapeUtil::MakeShape(operand_element_type, window_size), "window"); + CHECK_EQ(window_index.size(), index.size()); + + SetToFirstInsertPoint(loops.GetInnerLoopBodyBasicBlock(), &ir_builder_); + + llvm_ir::IrArray::Index input_index(ir_builder_.getInt64Ty(), index.size()); + llvm::Value* in_bounds_condition = nullptr; + for (size_t i = 0; i < index.size(); ++i) { + llvm::Value* strided_index = ir_builder_.CreateNSWMul( + index[i], ir_builder_.getInt64(window.dimensions(i).stride())); + input_index[i] = ir_builder_.CreateNSWSub( + ir_builder_.CreateNSWAdd(strided_index, window_index[i]), + ir_builder_.getInt64(window.dimensions(i).padding_low())); + + // We need to check if 0 <= input_index[i] < bound, as otherwise we are in + // the padding so that we can skip the computation. That is equivalent to + // input_index[i] < bound as an *unsigned* comparison, since a negative + // value will wrap to a large positive value. + llvm::Value* index_condition = ir_builder_.CreateICmpULT( + input_index[i], + ir_builder_.getInt64(ShapeUtil::GetDimension(operand->shape(), i))); + if (in_bounds_condition == nullptr) { + in_bounds_condition = index_condition; + } else { + in_bounds_condition = + ir_builder_.CreateAnd(in_bounds_condition, index_condition); + } + } + CHECK(in_bounds_condition != nullptr); + + llvm_ir::LlvmIfData if_data = + llvm_ir::EmitIfThenElse(in_bounds_condition, "in-bounds", &ir_builder_); + SetToFirstInsertPoint(if_data.true_block, &ir_builder_); + + // We are not in the padding, so carry out the computation. + llvm_ir::IrArray input_array(GetIrArrayFor(operand)); + llvm::Value* input_value_address = + input_array.EmitArrayElementAddress(input_index, &ir_builder_); + llvm::Value* result = EmitElementFunctionCall( + reducer_function, reduce_window->shape(), + {accumulator_address, input_value_address}, "reducer_function"); + ir_builder_.CreateStore(result, accumulator_address); + + SetToFirstInsertPoint(loops.GetOuterLoopExitBasicBlock(), &ir_builder_); + return ir_builder_.CreateLoad(accumulator_address); +} + +Status IrEmitter::HandleReduceWindow(HloInstruction* reduce_window) { TF_RETURN_IF_ERROR(ElementTypesSameAndSupported( - /*instruction=*/*reduce_window, /*operands=*/{operand}, + /*instruction=*/*reduce_window, + /*operands=*/{reduce_window->operand(0)}, /*supported_types=*/{F32, BF16, S32})); // TODO(b/31410564): Implement dilation for reduce-window. - if (window_util::HasDilation(window)) { + if (window_util::HasDilation(reduce_window->window())) { return Unimplemented( "Dilation for ReduceWindow is not implemented on CPU."); } - // The called computation should have been emitted previously. - llvm::Function* reducer_function = FindOrDie(emitted_functions_, function); - // Pseudo code for reduce window: // // for (coordinates O in the output) @@ -538,73 +595,9 @@ Status IrEmitter::HandleReduceWindow(HloInstruction* reduce_window) { // This is completely un-optimized and just here to have something // that works. return EmitTargetElementLoop( - reduce_window, [this, reduce_window, operand, window, - reducer_function](const llvm_ir::IrArray::Index& index) { - // We fold inputs into the accumulator and initialize it to - // the initial value on the reduce_window. - PrimitiveType operand_element_type = operand->shape().element_type(); - llvm::Value* accumulator_address = llvm_ir::EmitAllocaAtFunctionEntry( - llvm_ir::PrimitiveTypeToIrType(operand_element_type, module_), - "reduce_window_accumulator_address", &ir_builder_, - MinimumAlignmentForPrimitiveType(operand_element_type)); - ir_builder_.CreateStore(ir_builder_.CreateLoad(GetEmittedValueFor( - reduce_window->operand(1))), - accumulator_address); - - llvm_ir::ForLoopNest loops(IrName(reduce_window, "inner"), - &ir_builder_); - std::vector window_size; - for (const auto& dim : window.dimensions()) { - window_size.push_back(dim.size()); - } - const llvm_ir::IrArray::Index window_index = loops.AddLoopsForShape( - ShapeUtil::MakeShape(operand_element_type, window_size), "window"); - CHECK_EQ(window_index.size(), index.size()); - - SetToFirstInsertPoint(loops.GetInnerLoopBodyBasicBlock(), &ir_builder_); - - llvm_ir::IrArray::Index input_index(ir_builder_.getInt64Ty(), - index.size()); - llvm::Value* in_bounds_condition = nullptr; - for (size_t i = 0; i < index.size(); ++i) { - llvm::Value* strided_index = ir_builder_.CreateNSWMul( - index[i], ir_builder_.getInt64(window.dimensions(i).stride())); - input_index[i] = ir_builder_.CreateNSWSub( - ir_builder_.CreateNSWAdd(strided_index, window_index[i]), - ir_builder_.getInt64(window.dimensions(i).padding_low())); - - // We need to check if 0 <= input_index[i] < bound, as - // otherwise we are in the padding so that we can skip the - // computation. That is equivalent to input_index[i] < bound - // as an *unsigned* comparison, since a negative value will - // wrap to a large positive value. - llvm::Value* index_condition = ir_builder_.CreateICmpULT( - input_index[i], ir_builder_.getInt64(ShapeUtil::GetDimension( - operand->shape(), i))); - if (in_bounds_condition == nullptr) { - in_bounds_condition = index_condition; - } else { - in_bounds_condition = - ir_builder_.CreateAnd(in_bounds_condition, index_condition); - } - } - CHECK(in_bounds_condition != nullptr); - - llvm_ir::LlvmIfData if_data = llvm_ir::EmitIfThenElse( - in_bounds_condition, "in-bounds", &ir_builder_); - SetToFirstInsertPoint(if_data.true_block, &ir_builder_); - - // We are not in the padding, so carry out the computation. - llvm_ir::IrArray input_array(GetIrArrayFor(operand)); - llvm::Value* input_value_address = - input_array.EmitArrayElementAddress(input_index, &ir_builder_); - llvm::Value* result = EmitElementFunctionCall( - reducer_function, reduce_window->shape(), - {accumulator_address, input_value_address}, "reducer_function"); - ir_builder_.CreateStore(result, accumulator_address); - - SetToFirstInsertPoint(loops.GetOuterLoopExitBasicBlock(), &ir_builder_); - return ir_builder_.CreateLoad(accumulator_address); + reduce_window, [&](const llvm_ir::IrArray::Index& index) { + return EmitTargetElementLoopBodyForReduceWindow( + Cast(reduce_window), index); }); } @@ -833,17 +826,157 @@ Status IrEmitter::HandleDot(HloInstruction* dot) { target_machine_features_); } +StatusOr IrEmitter::EmitTargetElementLoopBodyForConvolution( + HloConvolutionInstruction* convolution, + const llvm_ir::IrArray::Index& index) { + const HloInstruction* lhs = convolution->operand(0); + const HloInstruction* rhs = convolution->operand(1); + const Window& window = convolution->window(); + + const ConvolutionDimensionNumbers& dnums = + convolution->convolution_dimension_numbers(); + int num_spatial_dims = dnums.output_spatial_dimensions_size(); + std::vector output_spatial(num_spatial_dims); + for (int i = 0; i < num_spatial_dims; ++i) { + output_spatial[i] = index[dnums.output_spatial_dimensions(i)]; + } + llvm::Value* output_feature = index[dnums.output_feature_dimension()]; + llvm::Value* batch = index[dnums.output_batch_dimension()]; + + // We will accumulate the products into this sum to calculate the output entry + // at the given index. + PrimitiveType lhs_element_type = lhs->shape().element_type(); + llvm::Type* lhs_llvm_type = + llvm_ir::PrimitiveTypeToIrType(lhs_element_type, module_); + llvm::Value* sum_address = llvm_ir::EmitAllocaAtFunctionEntry( + lhs_llvm_type, "convolution_sum_address", &ir_builder_, + MinimumAlignmentForPrimitiveType(lhs_element_type)); + llvm::Value* constant_zero = llvm::Constant::getNullValue(lhs_llvm_type); + ir_builder_.CreateStore(constant_zero, sum_address); + + llvm_ir::ForLoopNest loops(IrName(convolution, "inner"), &ir_builder_); + std::vector kernel_spatial(num_spatial_dims); + for (int i = 0; i < num_spatial_dims; ++i) { + kernel_spatial[i] = + loops + .AddLoop( + 0, rhs->shape().dimensions(dnums.kernel_spatial_dimensions(i)), + tensorflow::strings::StrCat("k", i)) + ->GetIndVarValue(); + } + llvm::Value* input_feature = + loops + .AddLoop(0, lhs->shape().dimensions(dnums.input_feature_dimension()), + "iz") + ->GetIndVarValue(); + + SetToFirstInsertPoint(loops.GetInnerLoopBodyBasicBlock(), &ir_builder_); + + // Calculate the spatial index in the input array, taking striding, dilation + // and padding into account. An index in the padding will be out of the bounds + // of the array. + const auto calculate_input_index = [this](llvm::Value* output_index, + llvm::Value* kernel_index, + const WindowDimension& window_dim) { + llvm::Value* strided_index = ir_builder_.CreateNSWMul( + output_index, ir_builder_.getInt64(window_dim.stride())); + llvm::Value* dilated_kernel_index = ir_builder_.CreateNSWMul( + kernel_index, ir_builder_.getInt64(window_dim.window_dilation())); + return ir_builder_.CreateNSWSub( + ir_builder_.CreateNSWAdd(strided_index, dilated_kernel_index), + ir_builder_.getInt64(window_dim.padding_low())); + }; + std::vector input_spatial(num_spatial_dims); + for (int i = 0; i < num_spatial_dims; ++i) { + input_spatial[i] = calculate_input_index( + output_spatial[i], kernel_spatial[i], window.dimensions(i)); + } + + // We need to check if 0 <= input dim < bound, as otherwise we are in the + // padding so that we can skip the computation. That is equivalent to input + // dim < bound as an *unsigned* comparison, since a negative value will wrap + // to a large positive value. The input dim is dilated, so we need to dilate + // the bound as well to match. + + // Also need to check that the input coordinates are not in one of the + // holes created by base dilation. + const auto not_in_hole = [&](llvm::Value* input_index, int64 base_dilation) { + llvm::Value* remainder = ir_builder_.CreateSRem( + input_index, ir_builder_.getInt64(base_dilation)); + return ir_builder_.CreateICmpEQ(remainder, ir_builder_.getInt64(0)); + }; + + llvm::Value* in_bounds_condition = ir_builder_.getInt1(true); + for (int i = 0; i < num_spatial_dims; ++i) { + llvm::ConstantInt* input_bound = + ir_builder_.getInt64(window_util::DilatedBound( + lhs->shape().dimensions(dnums.input_spatial_dimensions(i)), + window.dimensions(i).base_dilation())); + llvm::Value* dim_in_bound = + ir_builder_.CreateICmpULT(input_spatial[i], input_bound); + llvm::Value* dim_not_in_hole = + not_in_hole(input_spatial[i], window.dimensions(i).base_dilation()); + llvm::Value* dim_ok = ir_builder_.CreateAnd(dim_in_bound, dim_not_in_hole); + in_bounds_condition = ir_builder_.CreateAnd(in_bounds_condition, dim_ok); + } + + // Now we need to map the dilated base coordinates back to the actual + // data indices on the lhs. + const auto undilate = [&](llvm::Value* input_index, int64 base_dilation) { + return ir_builder_.CreateSDiv(input_index, + ir_builder_.getInt64(base_dilation)); + }; + for (int i = 0; i < num_spatial_dims; ++i) { + input_spatial[i] = + undilate(input_spatial[i], window.dimensions(i).base_dilation()); + } + + llvm_ir::LlvmIfData if_data = + llvm_ir::EmitIfThenElse(in_bounds_condition, "in-bounds", &ir_builder_); + SetToFirstInsertPoint(if_data.true_block, &ir_builder_); + + // We are not in the padding, so carry out the computation. + int num_dims = num_spatial_dims + 2; + llvm_ir::IrArray::Index input_index(ir_builder_.getInt64Ty(), num_dims); + for (int i = 0; i < num_spatial_dims; ++i) { + input_index[dnums.input_spatial_dimensions(i)] = input_spatial[i]; + } + input_index[dnums.input_feature_dimension()] = input_feature; + input_index[dnums.input_batch_dimension()] = batch; + + llvm_ir::IrArray kernel_array(GetIrArrayFor(rhs)); + llvm_ir::IrArray::Index kernel_index(ir_builder_.getInt64Ty(), num_dims); + for (int i = 0; i < num_spatial_dims; ++i) { + kernel_index[dnums.kernel_spatial_dimensions(i)] = + window.dimensions(i).window_reversal() + ? ir_builder_.CreateNSWSub( + ir_builder_.getInt64(window.dimensions(i).size() - 1), + kernel_spatial[i]) + : kernel_spatial[i]; + } + + kernel_index[dnums.kernel_input_feature_dimension()] = input_feature; + kernel_index[dnums.kernel_output_feature_dimension()] = output_feature; + + llvm_ir::IrArray input_array(GetIrArrayFor(lhs)); + llvm::Value* product = ir_builder_.CreateFMul( + input_array.EmitReadArrayElement(input_index, &ir_builder_), + kernel_array.EmitReadArrayElement(kernel_index, &ir_builder_)); + llvm::Value* sum = + ir_builder_.CreateFAdd(ir_builder_.CreateLoad(sum_address), product); + ir_builder_.CreateStore(sum, sum_address); + + SetToFirstInsertPoint(loops.GetOuterLoopExitBasicBlock(), &ir_builder_); + return ir_builder_.CreateLoad(sum_address); +} + Status IrEmitter::HandleConvolution(HloInstruction* convolution) { auto lhs = convolution->operand(0); auto rhs = convolution->operand(1); - const auto& window = convolution->window(); TF_RETURN_IF_ERROR(ElementTypesSameAndSupported( /*instruction=*/*convolution, /*operands=*/{lhs, rhs}, /*supported_types=*/{F16, F32, C64})); - const ConvolutionDimensionNumbers& dnums = - convolution->convolution_dimension_numbers(); - // TODO(tonywy): Add PotentiallyImplementedAsMKLCovolution to support // different data layouts. if (PotentiallyImplementedAsEigenConvolution(*convolution, @@ -1000,150 +1133,9 @@ Status IrEmitter::HandleConvolution(HloInstruction* convolution) { // See the description of convolution in the XLA documentation for the pseudo // code for convolution. return EmitTargetElementLoop( - convolution, [this, convolution, lhs, rhs, window, - dnums](const llvm_ir::IrArray::Index& index) { - int num_spatial_dims = dnums.output_spatial_dimensions_size(); - std::vector output_spatial(num_spatial_dims); - for (int i = 0; i < num_spatial_dims; ++i) { - output_spatial[i] = index[dnums.output_spatial_dimensions(i)]; - } - llvm::Value* output_feature = index[dnums.output_feature_dimension()]; - llvm::Value* batch = index[dnums.output_batch_dimension()]; - - // We will accumulate the products into this sum to calculate - // the output entry at the given index. - PrimitiveType lhs_element_type = lhs->shape().element_type(); - llvm::Type* lhs_llvm_type = - llvm_ir::PrimitiveTypeToIrType(lhs_element_type, module_); - llvm::Value* sum_address = llvm_ir::EmitAllocaAtFunctionEntry( - lhs_llvm_type, "convolution_sum_address", &ir_builder_, - MinimumAlignmentForPrimitiveType(lhs_element_type)); - llvm::Value* constant_zero = - llvm::Constant::getNullValue(lhs_llvm_type); - ir_builder_.CreateStore(constant_zero, sum_address); - - llvm_ir::ForLoopNest loops(IrName(convolution, "inner"), &ir_builder_); - std::vector kernel_spatial(num_spatial_dims); - for (int i = 0; i < num_spatial_dims; ++i) { - kernel_spatial[i] = - loops - .AddLoop(0, - rhs->shape().dimensions( - dnums.kernel_spatial_dimensions(i)), - tensorflow::strings::StrCat("k", i)) - ->GetIndVarValue(); - } - llvm::Value* input_feature = - loops - .AddLoop( - 0, lhs->shape().dimensions(dnums.input_feature_dimension()), - "iz") - ->GetIndVarValue(); - - SetToFirstInsertPoint(loops.GetInnerLoopBodyBasicBlock(), &ir_builder_); - - // Calculate the spatial index in the input array, taking striding, - // dilation and padding into account. An index in the padding will be - // out of the bounds of the array. - const auto calculate_input_index = - [this](llvm::Value* output_index, llvm::Value* kernel_index, - const WindowDimension& window_dim) { - llvm::Value* strided_index = ir_builder_.CreateNSWMul( - output_index, ir_builder_.getInt64(window_dim.stride())); - llvm::Value* dilated_kernel_index = ir_builder_.CreateNSWMul( - kernel_index, - ir_builder_.getInt64(window_dim.window_dilation())); - return ir_builder_.CreateNSWSub( - ir_builder_.CreateNSWAdd(strided_index, dilated_kernel_index), - ir_builder_.getInt64(window_dim.padding_low())); - }; - std::vector input_spatial(num_spatial_dims); - for (int i = 0; i < num_spatial_dims; ++i) { - input_spatial[i] = calculate_input_index( - output_spatial[i], kernel_spatial[i], window.dimensions(i)); - } - - // We need to check if 0 <= input dim < bound, as otherwise we are in - // the padding so that we can skip the computation. That is equivalent - // to input dim < bound as an *unsigned* comparison, since a negative - // value will wrap to a large positive value. The input dim is dilated, - // so we need to dilate the bound as well to match. - - // Also need to check that the input coordinates are not in one of the - // holes created by base dilation. - const auto not_in_hole = [&](llvm::Value* input_index, - int64 base_dilation) { - llvm::Value* remainder = ir_builder_.CreateSRem( - input_index, ir_builder_.getInt64(base_dilation)); - return ir_builder_.CreateICmpEQ(remainder, ir_builder_.getInt64(0)); - }; - - llvm::Value* in_bounds_condition = ir_builder_.getInt1(true); - for (int i = 0; i < num_spatial_dims; ++i) { - llvm::ConstantInt* input_bound = - ir_builder_.getInt64(window_util::DilatedBound( - lhs->shape().dimensions(dnums.input_spatial_dimensions(i)), - window.dimensions(i).base_dilation())); - llvm::Value* dim_in_bound = - ir_builder_.CreateICmpULT(input_spatial[i], input_bound); - llvm::Value* dim_not_in_hole = not_in_hole( - input_spatial[i], window.dimensions(i).base_dilation()); - llvm::Value* dim_ok = - ir_builder_.CreateAnd(dim_in_bound, dim_not_in_hole); - in_bounds_condition = - ir_builder_.CreateAnd(in_bounds_condition, dim_ok); - } - - // Now we need to map the dilated base coordinates back to the actual - // data indices on the lhs. - const auto undilate = [&](llvm::Value* input_index, - int64 base_dilation) { - return ir_builder_.CreateSDiv(input_index, - ir_builder_.getInt64(base_dilation)); - }; - for (int i = 0; i < num_spatial_dims; ++i) { - input_spatial[i] = - undilate(input_spatial[i], window.dimensions(i).base_dilation()); - } - - llvm_ir::LlvmIfData if_data = llvm_ir::EmitIfThenElse( - in_bounds_condition, "in-bounds", &ir_builder_); - SetToFirstInsertPoint(if_data.true_block, &ir_builder_); - - // We are not in the padding, so carry out the computation. - int num_dims = num_spatial_dims + 2; - llvm_ir::IrArray::Index input_index(ir_builder_.getInt64Ty(), num_dims); - for (int i = 0; i < num_spatial_dims; ++i) { - input_index[dnums.input_spatial_dimensions(i)] = input_spatial[i]; - } - input_index[dnums.input_feature_dimension()] = input_feature; - input_index[dnums.input_batch_dimension()] = batch; - - llvm_ir::IrArray kernel_array(GetIrArrayFor(rhs)); - llvm_ir::IrArray::Index kernel_index(ir_builder_.getInt64Ty(), - num_dims); - for (int i = 0; i < num_spatial_dims; ++i) { - kernel_index[dnums.kernel_spatial_dimensions(i)] = - window.dimensions(i).window_reversal() - ? ir_builder_.CreateNSWSub( - ir_builder_.getInt64(window.dimensions(i).size() - 1), - kernel_spatial[i]) - : kernel_spatial[i]; - } - - kernel_index[dnums.kernel_input_feature_dimension()] = input_feature; - kernel_index[dnums.kernel_output_feature_dimension()] = output_feature; - - llvm_ir::IrArray input_array(GetIrArrayFor(lhs)); - llvm::Value* product = ir_builder_.CreateFMul( - input_array.EmitReadArrayElement(input_index, &ir_builder_), - kernel_array.EmitReadArrayElement(kernel_index, &ir_builder_)); - llvm::Value* sum = ir_builder_.CreateFAdd( - ir_builder_.CreateLoad(sum_address), product); - ir_builder_.CreateStore(sum, sum_address); - - SetToFirstInsertPoint(loops.GetOuterLoopExitBasicBlock(), &ir_builder_); - return ir_builder_.CreateLoad(sum_address); + convolution, [&](const llvm_ir::IrArray::Index& index) { + return EmitTargetElementLoopBodyForConvolution( + Cast(convolution), index); }); } @@ -1780,6 +1772,64 @@ StatusOr IrEmitter::EmitVectorizedReduce( return true; } +StatusOr IrEmitter::EmitTargetElementLoopBodyForReduce( + HloReduceInstruction* reduce, const llvm_ir::IrArray::Index& index) { + const HloInstruction* arg = reduce->mutable_operand(0); + const HloInstruction* init_value = reduce->mutable_operand(1); + gtl::ArraySlice dimensions(reduce->dimensions()); + HloComputation* function = reduce->to_apply(); + // The called computation should have been emitted previously. + llvm::Function* reducer_function = FindOrDie(emitted_functions_, function); + + // Initialize an accumulator with init_value. + PrimitiveType accumulator_type = reduce->shape().element_type(); + llvm::AllocaInst* accumulator_addr = llvm_ir::EmitAllocaAtFunctionEntry( + llvm_ir::PrimitiveTypeToIrType(accumulator_type, module_), "accumulator", + &ir_builder_, MinimumAlignmentForPrimitiveType(accumulator_type)); + llvm::Value* init_value_addr = GetEmittedValueFor(init_value); + llvm::Value* load_init_value = ir_builder_.CreateLoad(init_value_addr); + ir_builder_.CreateStore(load_init_value, accumulator_addr); + + // The enclosing loops go over all the target elements. Now we have to compute + // the actual target element. For this, we build a new loop nest to iterate + // over all the reduction dimensions in the argument. + // AddLoopsForShapeOnDimensions will return an Index where induction Value*s + // are placed for each dimension in dimensions, and all the rest are nullptrs. + llvm_ir::ForLoopNest loops(IrName(reduce, "inner"), &ir_builder_); + const llvm_ir::IrArray::Index reduced_dims_index = + loops.AddLoopsForShapeOnDimensions(arg->shape(), dimensions, + "reduction_dim"); + + SetToFirstInsertPoint(loops.GetInnerLoopBodyBasicBlock(), &ir_builder_); + + // Build a full index for the input argument, using reduced_dims_index as the + // base. In reduced_dims_index only the reduction dimensions are filled in. We + // fill in the rest of the dimensions with induction Value*s taken from + // 'index' which iterates over the target array. See the high-level + // description in the XLA documentation for details. + llvm_ir::IrArray arg_array(GetIrArrayFor(arg)); + llvm_ir::IrArray::Index input_index = reduced_dims_index; + llvm_ir::IrArray::Index::const_iterator it = index.begin(); + + for (size_t i = 0; i < input_index.size(); ++i) { + if (input_index[i] == nullptr) { + input_index[i] = *it++; + } + } + CHECK(index.end() == it); + + // Apply the reduction function to the loaded value. + llvm::Value* input_address = + arg_array.EmitArrayElementAddress(input_index, &ir_builder_); + llvm::Value* result = EmitElementFunctionCall( + reducer_function, reduce->shape(), {accumulator_addr, input_address}, + "reduce_function"); + ir_builder_.CreateStore(result, accumulator_addr); + + SetToFirstInsertPoint(loops.GetOuterLoopExitBasicBlock(), &ir_builder_); + return ir_builder_.CreateLoad(accumulator_addr); +} + Status IrEmitter::HandleReduce(HloInstruction* reduce) { auto arg = reduce->mutable_operand(0); auto init_value = reduce->mutable_operand(1); @@ -1801,61 +1851,11 @@ Status IrEmitter::HandleReduce(HloInstruction* reduce) { } } - // The called computation should have been emitted previously. - llvm::Function* reducer_function = FindOrDie(emitted_functions_, function); - return EmitTargetElementLoop( - reduce, [this, reduce, arg, init_value, dimensions, - reducer_function](const llvm_ir::IrArray::Index& index) { - // Initialize an accumulator with init_value. - PrimitiveType accumulator_type = reduce->shape().element_type(); - llvm::AllocaInst* accumulator_addr = llvm_ir::EmitAllocaAtFunctionEntry( - llvm_ir::PrimitiveTypeToIrType(accumulator_type, module_), - "accumulator", &ir_builder_, - MinimumAlignmentForPrimitiveType(accumulator_type)); - llvm::Value* init_value_addr = GetEmittedValueFor(init_value); - llvm::Value* load_init_value = ir_builder_.CreateLoad(init_value_addr); - ir_builder_.CreateStore(load_init_value, accumulator_addr); - - // The enclosing loops go over all the target elements. Now we have to - // compute the actual target element. For this, we build a new loop nest - // to iterate over all the reduction dimensions in the argument. - // AddLoopsForShapeOnDimensions will return an Index where induction - // Value*s are placed for each dimension in dimensions, and all the rest - // are nullptrs. - llvm_ir::ForLoopNest loops(IrName(reduce, "inner"), &ir_builder_); - const llvm_ir::IrArray::Index reduced_dims_index = - loops.AddLoopsForShapeOnDimensions(arg->shape(), dimensions, - "reduction_dim"); - - SetToFirstInsertPoint(loops.GetInnerLoopBodyBasicBlock(), &ir_builder_); - - // Build a full index for the input argument, using reduced_dims_index - // as the base. In reduced_dims_index only the reduction dimensions are - // filled in. We fill in the rest of the dimensions with induction - // Value*s taken from 'index' which iterates over the target array. - // See the high-level description in the XLA documentation for details. - llvm_ir::IrArray arg_array(GetIrArrayFor(arg)); - llvm_ir::IrArray::Index input_index = reduced_dims_index; - llvm_ir::IrArray::Index::const_iterator it = index.begin(); - - for (size_t i = 0; i < input_index.size(); ++i) { - if (input_index[i] == nullptr) { - input_index[i] = *it++; - } - } - CHECK(index.end() == it); - - // Apply the reduction function to the loaded value. - llvm::Value* input_address = - arg_array.EmitArrayElementAddress(input_index, &ir_builder_); - llvm::Value* result = EmitElementFunctionCall( - reducer_function, reduce->shape(), - {accumulator_addr, input_address}, "reduce_function"); - ir_builder_.CreateStore(result, accumulator_addr); - - SetToFirstInsertPoint(loops.GetOuterLoopExitBasicBlock(), &ir_builder_); - return ir_builder_.CreateLoad(accumulator_addr); - }); + return EmitTargetElementLoop(reduce, + [&](const llvm_ir::IrArray::Index& index) { + return EmitTargetElementLoopBodyForReduce( + Cast(reduce), index); + }); } Status IrEmitter::HandleSend(HloInstruction* send) { @@ -2539,7 +2539,7 @@ Status IrEmitter::HandleConditional(HloInstruction* conditional) { return Status::OK(); } -Status IrEmitter::HandleGenerateToken(HloInstruction* gen_token) { +Status IrEmitter::HandleAfterAll(HloInstruction* gen_token) { TF_RET_CHECK(ByteSizeOf(gen_token->shape()) == 0); // No code to generate, but we need to emit an address for book-keeping. TF_RETURN_IF_ERROR(EmitTargetAddressForOp(gen_token)); diff --git a/tensorflow/compiler/xla/service/cpu/ir_emitter.h b/tensorflow/compiler/xla/service/cpu/ir_emitter.h index e1815c1db7a14dfc90ff646c0fd1e439ffffb2e8..419f19c24db9f2f0f8c6ae81eaa62fec372bdd4a 100644 --- a/tensorflow/compiler/xla/service/cpu/ir_emitter.h +++ b/tensorflow/compiler/xla/service/cpu/ir_emitter.h @@ -30,12 +30,12 @@ limitations under the License. #include "llvm/IR/Value.h" #include "llvm/Target/TargetMachine.h" #include "tensorflow/compiler/xla/service/buffer_assignment.h" -#include "tensorflow/compiler/xla/service/cpu/external_constant_pool.h" #include "tensorflow/compiler/xla/service/cpu/ir_function.h" #include "tensorflow/compiler/xla/service/cpu/target_machine_features.h" #include "tensorflow/compiler/xla/service/dfs_hlo_visitor_with_default.h" #include "tensorflow/compiler/xla/service/hlo_computation.h" #include "tensorflow/compiler/xla/service/hlo_instruction.h" +#include "tensorflow/compiler/xla/service/hlo_instructions.h" #include "tensorflow/compiler/xla/service/hlo_module_config.h" #include "tensorflow/compiler/xla/service/llvm_ir/alias_analysis.h" #include "tensorflow/compiler/xla/service/llvm_ir/ir_array.h" @@ -67,17 +67,13 @@ class IrEmitter : public DfsHloVisitorWithDefault { // index in the profiling array. // computation_to_profile_idx: the mapping from HLO computations to their // index in the profiling array. - // external_constant_pool: if non-null, points to an ExternalConstantPool - // instance into which the Ir emitter can spill - // constants. IrEmitter(const HloModule& hlo_module, const BufferAssignment& assignment, llvm::Module* llvm_module, std::unordered_map instruction_to_profile_idx, std::unordered_map computation_to_profile_idx, - const TargetMachineFeatures* target_machine, - ExternalConstantPool* external_constant_pool); + const TargetMachineFeatures* target_machine); ~IrEmitter() override; // Emit and return the given HLO computation as an LLVM IR @@ -122,6 +118,7 @@ class IrEmitter : public DfsHloVisitorWithDefault { Status HandleCopy(HloInstruction* copy) override; Status HandleGetTupleElement(HloInstruction* get_tuple_element) override; Status HandleSelect(HloInstruction* select) override; + Status HandleTupleSelect(HloInstruction* tuple_select) override; Status HandleDot(HloInstruction* dot) override; Status HandleConvolution(HloInstruction* convolution) override; Status HandleFft(HloInstruction* fft) override; @@ -150,7 +147,7 @@ class IrEmitter : public DfsHloVisitorWithDefault { Status HandleWhile(HloInstruction* xla_while) override; Status HandleConcatenate(HloInstruction* concatenate) override; Status HandleConditional(HloInstruction* conditional) override; - Status HandleGenerateToken(HloInstruction* gen_token) override; + Status HandleAfterAll(HloInstruction* gen_token) override; Status FinishVisit(HloInstruction* root) override; Status Preprocess(HloInstruction* hlo) override; @@ -518,6 +515,17 @@ class IrEmitter : public DfsHloVisitorWithDefault { // Returns the number of bytes within the shape. int64 ByteSizeOf(const Shape& shape) const; + StatusOr EmitTargetElementLoopBodyForMap( + HloMapInstruction* map, const llvm_ir::IrArray::Index& index); + StatusOr EmitTargetElementLoopBodyForReduceWindow( + HloReduceWindowInstruction* reduce_window, + const llvm_ir::IrArray::Index& index); + StatusOr EmitTargetElementLoopBodyForConvolution( + HloConvolutionInstruction* convolution, + const llvm_ir::IrArray::Index& index); + StatusOr EmitTargetElementLoopBodyForReduce( + HloReduceInstruction* reduce, const llvm_ir::IrArray::Index& index); + enum class XfeedKind { kInfeed, kOutfeed, @@ -537,9 +545,6 @@ class IrEmitter : public DfsHloVisitorWithDefault { const TargetMachineFeatures& target_machine_features_; - int64 external_global_constant_counter_ = 0; - ExternalConstantPool* external_constant_pool_; - struct LiteralPtrHashFunctor { size_t operator()(const Literal* literal) const { return literal->Hash(); } }; diff --git a/tensorflow/compiler/xla/service/cpu/parallel_task_assignment_test.cc b/tensorflow/compiler/xla/service/cpu/parallel_task_assignment_test.cc index fc2efbaf9a22b02cd729da2f367d53bc15506836..36c9f743859ae2da6c4fb3fd753bd7862fe2d3ab 100644 --- a/tensorflow/compiler/xla/service/cpu/parallel_task_assignment_test.cc +++ b/tensorflow/compiler/xla/service/cpu/parallel_task_assignment_test.cc @@ -110,8 +110,9 @@ TEST_F(ParallelTaskAssignmentTest, InfeedOutfeedOperationNotParallelized) { const string hlo_string = R"( HloModule TestTaskParallel_infeed_outfeed ENTRY InfeedOutfeed { - infeed0 = u32[12345678,2]{1,0} infeed() - ROOT outfeed0 = u32[12345678,2]{1,0} outfeed(infeed0) + infeed0 = (u32[12345678,2]{1,0}, token[]) infeed() + infeed0.data = u32[12345678,2]{1,0} get-tuple-element((u32[12345678,2]{1,0}, token[]) infeed0), index=0 + ROOT outfeed0 = token[] outfeed(infeed0.data) } )"; diff --git a/tensorflow/compiler/xla/service/cpu/sample_harness.cc b/tensorflow/compiler/xla/service/cpu/sample_harness.cc index 167aa4adda995a259190a932a76a34ca5883444c..d9e8dcaed98dc8bc33bf1355b0acddb56faa3c71 100644 --- a/tensorflow/compiler/xla/service/cpu/sample_harness.cc +++ b/tensorflow/compiler/xla/service/cpu/sample_harness.cc @@ -23,7 +23,7 @@ limitations under the License. #include "tensorflow/compiler/xla/client/local_client.h" #include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/compiler/xla/client/xla_client/xla_computation.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/statusor.h" #include "tensorflow/compiler/xla/types.h" #include "tensorflow/compiler/xla/xla_data.pb.h" @@ -38,20 +38,21 @@ int main(int argc, char** argv) { // Transfer parameters. std::unique_ptr param0_literal = - xla::Literal::CreateR1({1.1f, 2.2f, 3.3f, 5.5f}); + xla::LiteralUtil::CreateR1({1.1f, 2.2f, 3.3f, 5.5f}); std::unique_ptr param0_data = client->TransferToServer(*param0_literal).ConsumeValueOrDie(); - std::unique_ptr param1_literal = xla::Literal::CreateR2( - {{3.1f, 4.2f, 7.3f, 9.5f}, {1.1f, 2.2f, 3.3f, 4.4f}}); + std::unique_ptr param1_literal = + xla::LiteralUtil::CreateR2( + {{3.1f, 4.2f, 7.3f, 9.5f}, {1.1f, 2.2f, 3.3f, 4.4f}}); std::unique_ptr param1_data = client->TransferToServer(*param1_literal).ConsumeValueOrDie(); // Build computation. xla::XlaBuilder builder(""); - auto p0 = builder.Parameter(0, param0_literal->shape(), "param0"); - auto p1 = builder.Parameter(1, param1_literal->shape(), "param1"); - auto add = builder.Add(p1, p0, {0}); + auto p0 = Parameter(&builder, 0, param0_literal->shape(), "param0"); + auto p1 = Parameter(&builder, 1, param1_literal->shape(), "param1"); + Add(p1, p0, {0}); xla::StatusOr computation_status = builder.Build(); xla::XlaComputation computation = computation_status.ConsumeValueOrDie(); diff --git a/tensorflow/compiler/xla/service/cpu/simple_orc_jit.cc b/tensorflow/compiler/xla/service/cpu/simple_orc_jit.cc index c4c90515ac7ec2721cb9ea48d42e3c5080e249af..be772cfb7e564cebc5725854dbf5678e5c507556 100644 --- a/tensorflow/compiler/xla/service/cpu/simple_orc_jit.cc +++ b/tensorflow/compiler/xla/service/cpu/simple_orc_jit.cc @@ -127,13 +127,6 @@ SimpleOrcJIT::SimpleOrcJIT(const llvm::TargetOptions& target_options, } llvm::JITSymbol SimpleOrcJIT::ResolveRuntimeSymbol(const std::string& name) { - if (const uint8* from_constant_pool = - external_constant_pool_.Find(string(name))) { - return llvm::JITEvaluatedSymbol( - reinterpret_cast(from_constant_pool), - llvm::JITSymbolFlags::None); - } - void* func_addr = CustomCallTargetRegistry::Global()->Lookup(name); if (func_addr == nullptr) { return nullptr; diff --git a/tensorflow/compiler/xla/service/cpu/simple_orc_jit.h b/tensorflow/compiler/xla/service/cpu/simple_orc_jit.h index 1851a3ee0bb97b4860605d7211a6ae70ac88686b..d74b63fcf45bd70cd18ee41f1e9714ba6a222abd 100644 --- a/tensorflow/compiler/xla/service/cpu/simple_orc_jit.h +++ b/tensorflow/compiler/xla/service/cpu/simple_orc_jit.h @@ -29,7 +29,6 @@ limitations under the License. #include "llvm/Target/TargetMachine.h" #include "tensorflow/compiler/xla/service/cpu/compiler_functor.h" #include "tensorflow/compiler/xla/service/cpu/disassembler.h" -#include "tensorflow/compiler/xla/service/cpu/external_constant_pool.h" #include "tensorflow/compiler/xla/types.h" namespace xla { @@ -91,10 +90,6 @@ class SimpleOrcJIT { llvm::TargetMachine* target_machine() const { return target_machine_.get(); } - ExternalConstantPool* external_constant_pool() { - return &external_constant_pool_; - } - // Creates an llvm::TargetMachine suitable for JITting code that will run on // the current machine. static std::unique_ptr InferTargetMachineForJIT( @@ -112,7 +107,6 @@ class SimpleOrcJIT { std::shared_ptr symbol_resolver_; ObjLayerT object_layer_; CompileLayerT compile_layer_; - ExternalConstantPool external_constant_pool_; }; } // namespace cpu diff --git a/tensorflow/compiler/xla/service/cpu/tests/BUILD b/tensorflow/compiler/xla/service/cpu/tests/BUILD index 66ae5ef0f66e90982102d73e474f5d0582f5415c..b4c33e2f6cad8455e1e4ff2f020a167121e80a6f 100644 --- a/tensorflow/compiler/xla/service/cpu/tests/BUILD +++ b/tensorflow/compiler/xla/service/cpu/tests/BUILD @@ -40,7 +40,7 @@ tf_cc_test( name = "cpu_fusion_test", srcs = ["cpu_fusion_test.cc"], deps = [ - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:util", "//tensorflow/compiler/xla:xla_data_proto", @@ -82,7 +82,7 @@ tf_cc_test( name = "cpu_noalias_test", srcs = ["cpu_noalias_test.cc"], deps = [ - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:util", "//tensorflow/compiler/xla:xla_data_proto", @@ -128,7 +128,7 @@ tf_cc_test( name = "cpu_infeed_test", srcs = ["cpu_infeed_test.cc"], deps = [ - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:statusor", "//tensorflow/compiler/xla:test_helpers", diff --git a/tensorflow/compiler/xla/service/cpu/tests/cpu_codegen_test.h b/tensorflow/compiler/xla/service/cpu/tests/cpu_codegen_test.h index 7c8d07a10baf55dba8cbd347ebe1459b78e268e0..77b3a0301f2f90b577b7eaad86064dc30e2d9456 100644 --- a/tensorflow/compiler/xla/service/cpu/tests/cpu_codegen_test.h +++ b/tensorflow/compiler/xla/service/cpu/tests/cpu_codegen_test.h @@ -22,7 +22,7 @@ namespace xla { namespace cpu { // Tests that verify IR emitted by the CPU backend is as expected. -class CpuCodegenTest : public LLVMIRGenTestBase {}; +class CpuCodegenTest : public LlvmIrGenTestBase {}; } // namespace cpu } // namespace xla diff --git a/tensorflow/compiler/xla/service/cpu/tests/cpu_external_constants_test.cc b/tensorflow/compiler/xla/service/cpu/tests/cpu_external_constants_test.cc index faac927027c48e44eb8ff1fcc4109fbc177fc579..00a7aa2ad2f6bac4877302296ccb76222557535c 100644 --- a/tensorflow/compiler/xla/service/cpu/tests/cpu_external_constants_test.cc +++ b/tensorflow/compiler/xla/service/cpu/tests/cpu_external_constants_test.cc @@ -40,7 +40,7 @@ class CpuExternalConstantsTest : public CpuCodegenTest { HloInstruction* constant = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR2FromArray2D(backing_array))); + LiteralUtil::CreateR2FromArray2D(backing_array))); HloInstruction* param = builder.AddInstruction(HloInstruction::CreateParameter(0, shape, "x")); builder.AddInstruction( @@ -56,7 +56,8 @@ class CpuExternalConstantsTest : public CpuCodegenTest { TEST_F(CpuExternalConstantsTest, Basic) { TestWithArray(/*rows=*/1024, /*cols=*/1024, R"( -CHECK: @constant_global_0 = external constant [1024 x [1024 x float]], align 16 +CHECK-NOT: @constant_global_0 = external constant [1024 x [1024 x float]], align 16 +CHECK: @0 = private constant [4194304 x i8] {{.*}}, align 16 )"); } @@ -65,7 +66,7 @@ TEST_F(CpuExternalConstantsTest, BasicNegative) { // to externalize it. TestWithArray(/*rows=*/4, /*cols=*/4, R"( CHECK-NOT: @constant_global_0 = external constant [16 x float], align 8 -CHECK: @0 = private constant [16 x float] {{.*}}, align 8 +CHECK: @0 = private constant [64 x i8] {{.*}}, align 8 )"); } } // namespace diff --git a/tensorflow/compiler/xla/service/cpu/tests/cpu_fusion_test.cc b/tensorflow/compiler/xla/service/cpu/tests/cpu_fusion_test.cc index 23e7a3de4d8188a3add259582e11030539e154c1..d98856fdbf4165a5909f193ebe8512e21af83dfc 100644 --- a/tensorflow/compiler/xla/service/cpu/tests/cpu_fusion_test.cc +++ b/tensorflow/compiler/xla/service/cpu/tests/cpu_fusion_test.cc @@ -17,7 +17,7 @@ limitations under the License. #include #include -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/ptr_util.h" #include "tensorflow/compiler/xla/service/cpu/cpu_instruction_fusion.h" #include "tensorflow/compiler/xla/service/hlo_computation.h" @@ -43,8 +43,8 @@ class CpuFusionTest : public HloTestBase { TEST_F(CpuFusionTest, FuseTwoElementwiseOps) { auto builder = HloComputation::Builder(TestName()); - auto input_literal1 = Literal::CreateR1({1.0, 2.0, 3.0}); - auto input_literal2 = Literal::CreateR1({-2.0, -42.0, 2.0}); + auto input_literal1 = LiteralUtil::CreateR1({1.0, 2.0, 3.0}); + auto input_literal2 = LiteralUtil::CreateR1({-2.0, -42.0, 2.0}); Shape vshape = input_literal1->shape(); auto input1 = builder.AddInstruction( @@ -83,7 +83,7 @@ TEST_F(CpuFusionTest, FuseTwoElementwiseOps) { TEST_F(CpuFusionTest, FuseElementwiseOpChain) { auto builder = HloComputation::Builder(TestName()); - auto input_literal = Literal::CreateR1({-1.5, -2.5, -3.0}); + auto input_literal = LiteralUtil::CreateR1({-1.5, -2.5, -3.0}); Shape vshape = input_literal->shape(); auto input = builder.AddInstruction( @@ -96,8 +96,11 @@ TEST_F(CpuFusionTest, FuseElementwiseOpChain) { HloInstruction::CreateUnary(vshape, HloOpcode::kExp, ceil)); auto floor = builder.AddInstruction( HloInstruction::CreateUnary(vshape, HloOpcode::kFloor, exp)); - auto two = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(2.0))); + auto two = builder.AddInstruction(HloInstruction::CreateBroadcast( + vshape, + builder.AddInstruction( + HloInstruction::CreateConstant(LiteralUtil::CreateR0(2.0))), + {})); builder.AddInstruction( HloInstruction::CreateBinary(vshape, HloOpcode::kMultiply, two, floor)); @@ -114,9 +117,9 @@ TEST_F(CpuFusionTest, FuseElementwiseOpChain) { EXPECT_EQ(HloOpcode::kFusion, fusion_instruction->opcode()); EXPECT_EQ(HloOpcode::kMultiply, fusion_instruction->fused_expression_root()->opcode()); - // There should be 7 fused instructions: 2 parameters and the fused + // There should be 8 fused instructions: 2 parameters and the fused // operations. - EXPECT_EQ(7, fusion_instruction->fused_instruction_count()); + EXPECT_EQ(8, fusion_instruction->fused_instruction_count()); // Compile and execute the computation. auto result = ExecuteAndTransfer(std::move(module), {}); @@ -131,7 +134,7 @@ TEST_F(CpuFusionTest, ElementwiseOpChainWithNonfusableInstruction) { // middle. auto module = CreateNewModule(); auto builder = HloComputation::Builder(TestName()); - auto input_literal = Literal::CreateR1({-1.5, -2.5, -3.0}); + auto input_literal = LiteralUtil::CreateR1({-1.5, -2.5, -3.0}); Shape vshape = input_literal->shape(); auto input = builder.AddInstruction( @@ -163,15 +166,18 @@ TEST_F(CpuFusionTest, ElementwiseOpChainWithNonfusableInstruction) { ShapeUtil::MakeShape(F32, {6, 1}), concatenate)), /*init_value=*/ builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(0))), + HloInstruction::CreateConstant(LiteralUtil::CreateR0(0))), /*dimensions_to_reduce=*/{1}, add_f32)); auto exp = builder.AddInstruction( HloInstruction::CreateUnary(cshape, HloOpcode::kExp, reduce)); auto floor = builder.AddInstruction( HloInstruction::CreateUnary(cshape, HloOpcode::kFloor, exp)); - auto two = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(2.0))); + auto two = builder.AddInstruction(HloInstruction::CreateBroadcast( + cshape, + builder.AddInstruction( + HloInstruction::CreateConstant(LiteralUtil::CreateR0(2.0))), + {})); builder.AddInstruction( HloInstruction::CreateBinary(cshape, HloOpcode::kMultiply, two, floor)); @@ -188,9 +194,9 @@ TEST_F(CpuFusionTest, ElementwiseOpChainWithNonfusableInstruction) { EXPECT_EQ(HloOpcode::kFusion, fusion_instruction1->opcode()); EXPECT_EQ(HloOpcode::kMultiply, fusion_instruction1->fused_expression_root()->opcode()); - // There should be 5 fused instructions in the root fusion instruction: 2 + // There should be 6 fused instructions in the root fusion instruction: 2 // parameters, multiply, floor, and exp. - EXPECT_EQ(5, fusion_instruction1->fused_instruction_count()) + EXPECT_EQ(6, fusion_instruction1->fused_instruction_count()) << fusion_instruction1->fused_instructions_computation()->ToString(); auto fusion_instruction2 = reduce->operand(0); @@ -225,7 +231,7 @@ TEST_F(CpuFusionTest, TestOperandOrderToAvoidDuplication) { // operand vectors. Test for this problem by counting the number of nodes in // each fusion instruction to ensure that negate is not duplicated. auto builder = HloComputation::Builder(TestName()); - auto input_literal = Literal::CreateR1({1.0, 2.0, 3.0}); + auto input_literal = LiteralUtil::CreateR1({1.0, 2.0, 3.0}); Shape vshape = input_literal->shape(); auto constant = builder.AddInstruction( @@ -286,10 +292,10 @@ TEST_F(CpuFusionTest, DoNotDuplicateExpensiveOps) { // computation. The duplication is caused by the other use of exp2 in the // tuple. auto builder = HloComputation::Builder(TestName()); - auto input_literal1 = Literal::CreateR1({1.0, 2.0, 3.0}); - auto input_literal2 = Literal::CreateR1({-2.0, -42.0, 2.0}); + auto input_literal1 = LiteralUtil::CreateR1({1.0, 2.0, 3.0}); + auto input_literal2 = LiteralUtil::CreateR1({-2.0, -42.0, 2.0}); auto constant = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(42.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(42.0))); Shape shape = constant->shape(); auto exp1 = builder.AddInstruction( diff --git a/tensorflow/compiler/xla/service/cpu/tests/cpu_infeed_test.cc b/tensorflow/compiler/xla/service/cpu/tests/cpu_infeed_test.cc index dd63b998e9b6d04981ec6f7300c883c9b23b154f..0d45918d0990c0f00c08f321dad015cfbc038bf3 100644 --- a/tensorflow/compiler/xla/service/cpu/tests/cpu_infeed_test.cc +++ b/tensorflow/compiler/xla/service/cpu/tests/cpu_infeed_test.cc @@ -21,7 +21,7 @@ limitations under the License. #include "tensorflow/compiler/xla/client/local_client.h" #include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/compiler/xla/client/xla_client/xla_computation.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/shape_util.h" #include "tensorflow/compiler/xla/statusor.h" #include "tensorflow/compiler/xla/test_helpers.h" @@ -47,7 +47,7 @@ class InfeedTest : public ClientLibraryTestBase { // don't use ResetDevice since it is not implemented on CPU. ASSERT_IS_OK(client_->TransferToInfeed(literal)); XlaBuilder builder(TestName()); - builder.Infeed(literal.shape()); + Infeed(&builder, literal.shape()); if (ShapeUtil::IsTuple(literal.shape())) { // TODO(b/30609564): Use ComputeAndCompareLiteral instead. ComputeAndCompareTuple(&builder, literal, {}); @@ -58,52 +58,52 @@ class InfeedTest : public ClientLibraryTestBase { }; TEST_F(InfeedTest, SingleInfeedR0Bool) { - TestInfeedRoundTrip(*Literal::CreateR0(true)); + TestInfeedRoundTrip(*LiteralUtil::CreateR0(true)); } TEST_F(InfeedTest, SingleInfeedR1U32) { - TestInfeedRoundTrip(*Literal::CreateR1({1, 2, 3})); + TestInfeedRoundTrip(*LiteralUtil::CreateR1({1, 2, 3})); } TEST_F(InfeedTest, SingleInfeedR2F32) { - TestInfeedRoundTrip(*Literal::CreateR2F32Linspace(0.0, 1.0, 128, 64)); + TestInfeedRoundTrip(*LiteralUtil::CreateR2F32Linspace(0.0, 1.0, 128, 64)); } TEST_F(InfeedTest, SingleInfeedR3F32) { TestInfeedRoundTrip( - *Literal::CreateR3({{{1.0f, 2.0f, 3.0f}, {4.0f, 5.0f, 6.0f}}, - {{1.1f, 2.1f, 3.1f}, {6.1f, 3.5f, 2.8f}}})); + *LiteralUtil::CreateR3({{{1.0f, 2.0f, 3.0f}, {4.0f, 5.0f, 6.0f}}, + {{1.1f, 2.1f, 3.1f}, {6.1f, 3.5f, 2.8f}}})); } TEST_F(InfeedTest, SingleInfeedR3F32DifferentLayout) { const Layout r3_dim0minor = LayoutUtil::MakeLayout({0, 1, 2}); const Layout r3_dim0major = LayoutUtil::MakeLayout({2, 1, 0}); - TestInfeedRoundTrip( - *Literal::CreateR3WithLayout({{{1.0f, 2.0f, 3.0f}, {4.0f, 5.0f, 6.0f}}, - {{1.1f, 2.1f, 3.1f}, {6.1f, 3.5f, 2.8f}}}, - r3_dim0minor)); + TestInfeedRoundTrip(*LiteralUtil::CreateR3WithLayout( + {{{1.0f, 2.0f, 3.0f}, {4.0f, 5.0f, 6.0f}}, + {{1.1f, 2.1f, 3.1f}, {6.1f, 3.5f, 2.8f}}}, + r3_dim0minor)); - TestInfeedRoundTrip( - *Literal::CreateR3WithLayout({{{1.0f, 2.0f, 3.0f}, {4.0f, 5.0f, 6.0f}}, - {{1.1f, 2.1f, 3.1f}, {6.1f, 3.5f, 2.8f}}}, - r3_dim0major)); + TestInfeedRoundTrip(*LiteralUtil::CreateR3WithLayout( + {{{1.0f, 2.0f, 3.0f}, {4.0f, 5.0f, 6.0f}}, + {{1.1f, 2.1f, 3.1f}, {6.1f, 3.5f, 2.8f}}}, + r3_dim0major)); } TEST_F(InfeedTest, SingleInfeedR4S32) { - TestInfeedRoundTrip(*Literal::CreateR4( + TestInfeedRoundTrip(*LiteralUtil::CreateR4( {{{{1, -2}, {-4, 5}, {6, 7}}, {{8, 9}, {10, 11}, {12, 13}}}, {{{10, 3}, {7, -2}, {3, 6}}, {{2, 5}, {-11, 5}, {-2, -5}}}})); } TEST_F(InfeedTest, SingleInfeedTuple) { TestInfeedRoundTrip( - *Literal::MakeTuple({Literal::CreateR1({1, 2, 3}).get(), - Literal::CreateR0(false).get()})); + *LiteralUtil::MakeTuple({LiteralUtil::CreateR1({1, 2, 3}).get(), + LiteralUtil::CreateR0(false).get()})); } TEST_F(InfeedTest, SingleInfeedEmptyTuple) { - TestInfeedRoundTrip(*Literal::MakeTuple({})); + TestInfeedRoundTrip(*LiteralUtil::MakeTuple({})); } // Tests Infeed operation used in a while loop, as in the code below. The @@ -125,8 +125,8 @@ TEST_F(InfeedTest, DISABLED_SingleInfeedInWhile) { XlaComputation condition; { XlaBuilder builder("condition"); - auto prev = builder.Parameter(0, result_shape, "prev"); - builder.Gt(builder.ConstantR0(40.0f), prev); + auto prev = Parameter(&builder, 0, result_shape, "prev"); + Gt(ConstantR0(&builder, 40.0f), prev); condition = builder.Build().ConsumeValueOrDie(); } // Create a computation for the body: add the reduced value of the Infeed @@ -134,17 +134,16 @@ TEST_F(InfeedTest, DISABLED_SingleInfeedInWhile) { XlaComputation body; { XlaBuilder builder("body"); - auto prev = builder.Parameter(0, result_shape, "prev"); - auto infeed = builder.Infeed(infeed_shape); - auto addend = - builder.Reduce(infeed, builder.ConstantR0(0.0f), - CreateScalarAddComputation(F32, &builder), {0}); - builder.Add(prev, addend); + auto prev = Parameter(&builder, 0, result_shape, "prev"); + auto infeed = Infeed(&builder, infeed_shape); + auto addend = Reduce(infeed, ConstantR0(&builder, 0.0f), + CreateScalarAddComputation(F32, &builder), {0}); + Add(prev, addend); body = builder.Build().ConsumeValueOrDie(); } // Create a While node with computations for the condition and the body. - auto init = builder.ConstantR0(0.0f); - builder.While(condition, body, init); + auto init = ConstantR0(&builder, 0.0f); + While(condition, body, init); // Build and asynchronously launch the computation. auto computation = builder.Build().ConsumeValueOrDie(); @@ -157,13 +156,16 @@ TEST_F(InfeedTest, DISABLED_SingleInfeedInWhile) { }); // Send 5 Infeed data of shape F32[3]. - ASSERT_IS_OK(client_->TransferToInfeed(*Literal::CreateR1({1, 2, 3}))); - ASSERT_IS_OK(client_->TransferToInfeed(*Literal::CreateR1({4, 5, 6}))); - ASSERT_IS_OK(client_->TransferToInfeed(*Literal::CreateR1({7, 8, 9}))); ASSERT_IS_OK( - client_->TransferToInfeed(*Literal::CreateR1({10, 11, 12}))); + client_->TransferToInfeed(*LiteralUtil::CreateR1({1, 2, 3}))); + ASSERT_IS_OK( + client_->TransferToInfeed(*LiteralUtil::CreateR1({4, 5, 6}))); ASSERT_IS_OK( - client_->TransferToInfeed(*Literal::CreateR1({13, 14, 15}))); + client_->TransferToInfeed(*LiteralUtil::CreateR1({7, 8, 9}))); + ASSERT_IS_OK( + client_->TransferToInfeed(*LiteralUtil::CreateR1({10, 11, 12}))); + ASSERT_IS_OK( + client_->TransferToInfeed(*LiteralUtil::CreateR1({13, 14, 15}))); delete computation_thread; // Joins the thread. auto result_literal = client_->Transfer(*result).ConsumeValueOrDie(); @@ -207,8 +209,8 @@ TEST_F(InfeedTest, DISABLED_TwoInfeedsInTotalOrder) { XlaComputation condition; { XlaBuilder builder("condition"); - auto prev = builder.Parameter(0, result_shape, "prev"); - builder.GetTupleElement(prev, 1); + auto prev = Parameter(&builder, 0, result_shape, "prev"); + GetTupleElement(prev, 1); condition = builder.Build().ConsumeValueOrDie(); } @@ -221,44 +223,44 @@ TEST_F(InfeedTest, DISABLED_TwoInfeedsInTotalOrder) { const auto build_body = [this, &result_shape](const Shape& infeed_shape) { XlaComputation body; XlaBuilder builder("body"); - auto prev = builder.Parameter(0, result_shape, "prev"); - auto infeed = builder.Infeed(infeed_shape); - auto addend = builder.Reduce( - builder.GetTupleElement(infeed, 0), builder.ConstantR0(0.0f), - CreateScalarAddComputation(F32, &builder), {0}); - auto result = builder.Add(builder.GetTupleElement(prev, 0), addend); - builder.Tuple({result, builder.GetTupleElement(infeed, 1)}); + auto prev = Parameter(&builder, 0, result_shape, "prev"); + auto infeed = Infeed(&builder, infeed_shape); + auto addend = + Reduce(GetTupleElement(infeed, 0), ConstantR0(&builder, 0.0f), + CreateScalarAddComputation(F32, &builder), {0}); + auto result = Add(GetTupleElement(prev, 0), addend); + Tuple(&builder, {result, GetTupleElement(infeed, 1)}); return builder.Build().ConsumeValueOrDie(); }; // Create the first while loop with infeed1_shape. - auto init = builder.Tuple( - {builder.ConstantR0(0.0f), builder.ConstantR0(true)}); - auto while1 = builder.While(condition, build_body(infeed1_shape), init); - auto result1 = builder.Tuple( - {builder.GetTupleElement(while1, 0), builder.ConstantR0(true)}); + auto init = Tuple(&builder, {ConstantR0(&builder, 0.0f), + ConstantR0(&builder, true)}); + auto while1 = While(condition, build_body(infeed1_shape), init); + auto result1 = Tuple( + &builder, {GetTupleElement(while1, 0), ConstantR0(&builder, true)}); // Create the second while loop with infeed2_shape. Note that the result from // the first while loop is used as the initial value. - auto while2 = builder.While(condition, build_body(infeed2_shape), result1); - builder.GetTupleElement(while2, 0); + auto while2 = While(condition, build_body(infeed2_shape), result1); + GetTupleElement(while2, 0); // Build the computation. auto computation = builder.Build().ConsumeValueOrDie(); // Send the first 4 Infeed data of shape Tuple(F32[2], PRED). ASSERT_IS_OK(client_->TransferToInfeed( - *Literal::MakeTuple({Literal::CreateR1({1, 2}).get(), - Literal::CreateR0(true).get()}))); + *LiteralUtil::MakeTuple({LiteralUtil::CreateR1({1, 2}).get(), + LiteralUtil::CreateR0(true).get()}))); ASSERT_IS_OK(client_->TransferToInfeed( - *Literal::MakeTuple({Literal::CreateR1({3, 4}).get(), - Literal::CreateR0(true).get()}))); + *LiteralUtil::MakeTuple({LiteralUtil::CreateR1({3, 4}).get(), + LiteralUtil::CreateR0(true).get()}))); ASSERT_IS_OK(client_->TransferToInfeed( - *Literal::MakeTuple({Literal::CreateR1({5, 6}).get(), - Literal::CreateR0(true).get()}))); + *LiteralUtil::MakeTuple({LiteralUtil::CreateR1({5, 6}).get(), + LiteralUtil::CreateR0(true).get()}))); ASSERT_IS_OK(client_->TransferToInfeed( - *Literal::MakeTuple({Literal::CreateR1({7, 8}).get(), - Literal::CreateR0(false).get()}))); + *LiteralUtil::MakeTuple({LiteralUtil::CreateR1({7, 8}).get(), + LiteralUtil::CreateR0(false).get()}))); // Asynchronously launch the execution on the device. std::unique_ptr result; @@ -273,14 +275,14 @@ TEST_F(InfeedTest, DISABLED_TwoInfeedsInTotalOrder) { // Infeed data, and send the rest Infeed data of shape Tuple(F32[3], PRED). sleep(1); ASSERT_IS_OK(client_->TransferToInfeed( - *Literal::MakeTuple({Literal::CreateR1({1, 2, 3}).get(), - Literal::CreateR0(true).get()}))); + *LiteralUtil::MakeTuple({LiteralUtil::CreateR1({1, 2, 3}).get(), + LiteralUtil::CreateR0(true).get()}))); ASSERT_IS_OK(client_->TransferToInfeed( - *Literal::MakeTuple({Literal::CreateR1({7, 8, 9}).get(), - Literal::CreateR0(false).get()}))); + *LiteralUtil::MakeTuple({LiteralUtil::CreateR1({7, 8, 9}).get(), + LiteralUtil::CreateR0(false).get()}))); ASSERT_IS_OK(client_->TransferToInfeed( - *Literal::MakeTuple({Literal::CreateR1({4, 5, 6}).get(), - Literal::CreateR0(true).get()}))); + *LiteralUtil::MakeTuple({LiteralUtil::CreateR1({4, 5, 6}).get(), + LiteralUtil::CreateR0(true).get()}))); // Wait for the execution to be done, and transfer the result. delete computation_thread; // Joins the thread. diff --git a/tensorflow/compiler/xla/service/cpu/tests/cpu_literal_caching_test.cc b/tensorflow/compiler/xla/service/cpu/tests/cpu_literal_caching_test.cc index 27044b1d62027e3b83744c486cb790269e505aff..90b99c828e2fcfd77579026a39d3a6711599feee 100644 --- a/tensorflow/compiler/xla/service/cpu/tests/cpu_literal_caching_test.cc +++ b/tensorflow/compiler/xla/service/cpu/tests/cpu_literal_caching_test.cc @@ -38,7 +38,8 @@ while_body { while_cond { arg_cond = f32[2,3,2] parameter(0) - ROOT unknown = pred[] infeed() + infeed = (pred[], token[]) infeed() + ROOT unknown = pred[] get-tuple-element((pred[], token[]) infeed), index=0 } ENTRY main { @@ -49,14 +50,14 @@ ENTRY main { {{2, 1}, {2001, 3002}, {2001, 2002}}}) const_b = f32[2,3,2] while(f32[2,3,2] const_a), condition=while_cond, body=while_body - out0 = () outfeed(f32[2,3,2] const_a) - ROOT out1 = () outfeed(f32[2,3,2] const_b) + out0 = token[] outfeed(f32[2,3,2] const_a) + ROOT out1 = token[] outfeed(f32[2,3,2] const_b) } )"; string filecheck_pattern = R"( -CHECK: private constant [12 x float] -CHECK-NOT: private constant [12 x float] +CHECK: private constant [48 x i8] +CHECK-NOT: private constant [48 x i8] )"; TF_ASSERT_OK_AND_ASSIGN(std::unique_ptr module, @@ -84,7 +85,8 @@ while_body { while_cond { arg_cond = (f32[2,1]{1,0}, f32[1]{0}) parameter(0) - ROOT unknown = pred[] infeed() + infeed = (pred[], token[]) infeed() + ROOT unknown = pred[] get-tuple-element((pred[], token[]) infeed), index=0 } ENTRY main { @@ -98,10 +100,10 @@ ENTRY main { )"; string filecheck_pattern = R"( -CHECK: private constant [1 x float] -CHECK: private constant [2 x float] -CHECK-NOT: private constant [1 x float] -CHECK-NOT: private constant [2 x float] +CHECK: private constant [4 x i8] +CHECK: private constant [8 x i8] +CHECK-NOT: private constant [4 x i8] +CHECK-NOT: private constant [8 x i8] )"; TF_ASSERT_OK_AND_ASSIGN(std::unique_ptr module, diff --git a/tensorflow/compiler/xla/service/cpu/tests/cpu_noalias_test.cc b/tensorflow/compiler/xla/service/cpu/tests/cpu_noalias_test.cc index 3b6b0ed74065615fb9e47a0ec3c6c4ab078e45c4..ccb61740f6b717ce7dc2a6f614d6d2c8b4d7a9a5 100644 --- a/tensorflow/compiler/xla/service/cpu/tests/cpu_noalias_test.cc +++ b/tensorflow/compiler/xla/service/cpu/tests/cpu_noalias_test.cc @@ -17,7 +17,7 @@ limitations under the License. #include #include "llvm/IR/Module.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/ptr_util.h" #include "tensorflow/compiler/xla/service/buffer_assignment.h" #include "tensorflow/compiler/xla/service/cpu/tests/cpu_codegen_test.h" @@ -42,7 +42,7 @@ TEST_F(CpuNoAliasTest, Concat) { HloComputation::Builder builder(TestName()); std::unique_ptr literal = - Literal::CreateR2({{1.0, 2.0}, {3.0, 4.0}}); + LiteralUtil::CreateR2({{1.0, 2.0}, {3.0, 4.0}}); auto param_shape = ShapeUtil::MakeShape(F32, {2, 2}); HloInstruction* param_x = builder.AddInstruction( HloInstruction::CreateParameter(0, param_shape, "x")); diff --git a/tensorflow/compiler/xla/service/cpu/tests/cpu_outfeed_test.cc b/tensorflow/compiler/xla/service/cpu/tests/cpu_outfeed_test.cc index 1ee279290b6fcfe775ce9867d424b1c031f5d2bd..dac416e1c78c2f60d458480c5062f48b77d4878d 100644 --- a/tensorflow/compiler/xla/service/cpu/tests/cpu_outfeed_test.cc +++ b/tensorflow/compiler/xla/service/cpu/tests/cpu_outfeed_test.cc @@ -32,12 +32,13 @@ ENTRY main { {{{1, 2}, {1001, 1002}, {2001, 2002}}, {{2, 1}, {2001, 3002}, {2001, 2002}}}) - ROOT out = () outfeed(f32[2,3,2] const_a) + outfeed = token[] outfeed(f32[2,3,2] const_a) + ROOT root = () tuple() } )"; string filecheck_pattern = R"( -CHECK: private constant [12 x float] +CHECK: private constant [48 x i8] )"; TF_ASSERT_OK_AND_ASSIGN(std::unique_ptr module, diff --git a/tensorflow/compiler/xla/service/defuser_test.cc b/tensorflow/compiler/xla/service/defuser_test.cc index 32b5c5d35fae61ae6cb17fafcada1abd6c3c088c..e727ba49cb6321e499b5d50d5f45e7f7f6bb6fef 100644 --- a/tensorflow/compiler/xla/service/defuser_test.cc +++ b/tensorflow/compiler/xla/service/defuser_test.cc @@ -15,7 +15,7 @@ limitations under the License. #include "tensorflow/compiler/xla/service/defuser.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/service/hlo_matchers.h" #include "tensorflow/compiler/xla/shape_util.h" #include "tensorflow/compiler/xla/tests/hlo_verified_test_base.h" @@ -124,7 +124,7 @@ TEST_F(DefuserTest, NonTrivialFusionInstruction) { auto div = builder.AddInstruction( HloInstruction::CreateBinary(shape_, HloOpcode::kDivide, mul, param3)); auto constant = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR2({{1.0, 2.0}, {3.0, 4.0}}))); + LiteralUtil::CreateR2({{1.0, 2.0}, {3.0, 4.0}}))); auto add2 = builder.AddInstruction( HloInstruction::CreateBinary(shape_, HloOpcode::kAdd, constant, div)); @@ -162,7 +162,7 @@ TEST_F(DefuserTest, MultipleFusionInstructions) { auto div = builder.AddInstruction( HloInstruction::CreateBinary(shape_, HloOpcode::kDivide, mul, param3)); auto constant = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR2({{1.0, 2.0}, {3.0, 4.0}}))); + LiteralUtil::CreateR2({{1.0, 2.0}, {3.0, 4.0}}))); auto add2 = builder.AddInstruction( HloInstruction::CreateBinary(shape_, HloOpcode::kAdd, constant, div)); diff --git a/tensorflow/compiler/xla/service/dfs_hlo_visitor.h b/tensorflow/compiler/xla/service/dfs_hlo_visitor.h index 7d56d57b5fc9ef9d92b2b44b4128925ad426708b..51f16bdc94777fe027991cb756322335eecda81f 100644 --- a/tensorflow/compiler/xla/service/dfs_hlo_visitor.h +++ b/tensorflow/compiler/xla/service/dfs_hlo_visitor.h @@ -19,7 +19,7 @@ limitations under the License. #include #include -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/service/hlo_opcode.h" #include "tensorflow/compiler/xla/status.h" #include "tensorflow/compiler/xla/types.h" @@ -76,6 +76,7 @@ class DfsHloVisitorBase { virtual Status HandleClamp(HloInstructionPtr hlo) = 0; virtual Status HandleSelect(HloInstructionPtr hlo) = 0; + virtual Status HandleTupleSelect(HloInstructionPtr hlo) = 0; virtual Status HandleMaximum(HloInstructionPtr hlo) { return HandleElementwiseBinary(hlo); } @@ -246,7 +247,7 @@ class DfsHloVisitorBase { virtual Status HandleBatchNormGrad(HloInstructionPtr hlo) = 0; - virtual Status HandleGenerateToken(HloInstructionPtr token) = 0; + virtual Status HandleAfterAll(HloInstructionPtr token) = 0; // Invoked to inform the visitor that the traversal has completed, and that // the root was "root". diff --git a/tensorflow/compiler/xla/service/dfs_hlo_visitor_with_default.h b/tensorflow/compiler/xla/service/dfs_hlo_visitor_with_default.h index 6934e00a4b665e9e6a4302e0c0a8ce1d5bb94373..0686ca74afcde541e89ed3d16b79914de7d06bb3 100644 --- a/tensorflow/compiler/xla/service/dfs_hlo_visitor_with_default.h +++ b/tensorflow/compiler/xla/service/dfs_hlo_visitor_with_default.h @@ -16,7 +16,7 @@ limitations under the License. #ifndef TENSORFLOW_COMPILER_XLA_SERVICE_DFS_HLO_VISITOR_WITH_DEFAULT_H_ #define TENSORFLOW_COMPILER_XLA_SERVICE_DFS_HLO_VISITOR_WITH_DEFAULT_H_ -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/service/dfs_hlo_visitor.h" #include "tensorflow/compiler/xla/service/hlo_opcode.h" #include "tensorflow/compiler/xla/types.h" @@ -79,6 +79,9 @@ class DfsHloVisitorWithDefaultBase Status HandleSelect(HloInstructionPtr select) override { return DefaultAction(select); } + Status HandleTupleSelect(HloInstructionPtr tuple_select) override { + return DefaultAction(tuple_select); + } Status HandleDot(HloInstructionPtr dot) override { return DefaultAction(dot); } @@ -188,7 +191,7 @@ class DfsHloVisitorWithDefaultBase Status HandleGather(HloInstructionPtr gather) override { return DefaultAction(gather); } - Status HandleGenerateToken(HloInstructionPtr token) override { + Status HandleAfterAll(HloInstructionPtr token) override { return DefaultAction(token); } diff --git a/tensorflow/compiler/xla/service/elemental_ir_emitter.cc b/tensorflow/compiler/xla/service/elemental_ir_emitter.cc index ce0951bbe1873973c7b97055aba5ba71a14ad24f..004a80d19ddac8c6354b0ed91857e563c62c7e3f 100644 --- a/tensorflow/compiler/xla/service/elemental_ir_emitter.cc +++ b/tensorflow/compiler/xla/service/elemental_ir_emitter.cc @@ -1227,7 +1227,14 @@ llvm_ir::IrArray::Index ElementalIrEmitter::ElementwiseSourceIndex( // If no implicit broadcast is needed for this operand, returns the target // index as the source index. - if (ShapeUtil::CompatibleIgnoringElementType(operand_shape, hlo.shape())) { + // + // `IrArray::Index` may contain a physical linear which we can propagate to + // our operand only if our layouts match. "only if" is a bit strong since + // e.g. we can still forward the linear index if the operand shape is + // [5,1,1,5]{3,2,1,0} and the HLO shape is[5,1,1,5]{3,1,2,0}, but those cases + // are probably not worth handling here for now. + if (ShapeUtil::CompatibleIgnoringElementType(operand_shape, hlo.shape()) && + LayoutUtil::Equal(operand_shape.layout(), hlo.shape().layout())) { return target_index; } @@ -1558,19 +1565,18 @@ StatusOr ElementalIrEmitter::EmitElementalDynamicSlice( // TODO(b/74360564): This is implementation defined behavior, but is // currently respected by all implementations. Change this if we ever decide - // to oficially document different behavior. + // to officially document different behavior. start_index_value = ir_builder_->CreateSExtOrTrunc(start_index_value, index_type); - llvm::Value* operand_dim_size = - index_typed_const(input_hlo->shape().dimensions(i)); - llvm::Value* output_dim_size = - index_typed_const(hlo->shape().dimensions(i)); + int64 largest_valid_start_index = + input_hlo->shape().dimensions(i) - hlo->shape().dimensions(i); + CHECK_GE(largest_valid_start_index, 0); + bool is_signed = ShapeUtil::ElementIsSigned(hlo->operand(1)->shape()); start_index_value = EmitIntegralMin( - ir_builder_->CreateSub(operand_dim_size, output_dim_size), - EmitIntegralMax(index_typed_const(0), start_index_value, - /*is_signed=*/true), - /*is_signed=*/true); + index_typed_const(largest_valid_start_index), + EmitIntegralMax(index_typed_const(0), start_index_value, is_signed), + is_signed); start_index_value->setName( AsStringRef(IrName(hlo, StrCat("start_idx", i)))); @@ -1603,19 +1609,22 @@ StatusOr ElementalIrEmitter::EmitElementalGather( llvm::Type* index_type = index.GetType(); // This is the index into `operand` that holds the element we want to - // generate. This index "unsafe" as in the components in here may be - // out of bounds. - IrArray::Index unsafe_operand_index(index_type); - - // First copy in the window indices to unsafe_operand_index. - for (int64 i = 0, e = operand_shape.dimensions_size(), - unsafe_operand_index_dim = 0; + // generate. + IrArray::Index operand_index(index_type); + + // First copy in the window indices to operand_index. Also collect a mapping + // from operand dimension to output window dimension. Elided window dimensions + // map to -1. + std::vector operand_to_output_dim(operand_shape.dimensions_size(), -1); + for (int64 i = 0, e = operand_shape.dimensions_size(), operand_index_dim = 0; i < e; i++) { if (c_binary_search(dim_numbers.elided_window_dims(), i)) { - unsafe_operand_index.push_back(index.GetConstantWithIndexType(0)); + operand_index.push_back(index.GetConstantWithIndexType(0)); } else { - unsafe_operand_index.push_back( - index[dim_numbers.output_window_dims(unsafe_operand_index_dim++)]); + int64 output_window_dim = + dim_numbers.output_window_dims(operand_index_dim++); + operand_to_output_dim[i] = output_window_dim; + operand_index.push_back(index[output_window_dim]); } } @@ -1634,20 +1643,40 @@ StatusOr ElementalIrEmitter::EmitElementalGather( } } - auto add_to_unsafe_operand_index = [&](llvm::Value* index_component, - int64 dim) { + auto add_to_operand_index = [&](llvm::Value* index_component, int64 dim) { llvm::Value* gather_dim_component_extended = ir_builder_->CreateSExtOrTrunc(index_component, index_type); - unsafe_operand_index[dim_numbers.gather_dims_to_operand_dims(dim)] = - ir_builder_->CreateAdd( - unsafe_operand_index[dim_numbers.gather_dims_to_operand_dims(dim)], - gather_dim_component_extended); + int64 operand_dim = dim_numbers.gather_dims_to_operand_dims(dim); + int64 output_dim = operand_to_output_dim[operand_dim]; + // If 'output_dim' is -1, it means 'operand_dim' is an elided window dim. + // This means we set the iteration index to 0, so for the purpose of the + // following calculations we can consider the output dimension size to be 1. + int64 output_dim_size = + output_dim == -1 ? 1 : output_shape.dimensions(output_dim); + int64 largest_valid_start_index = + operand_shape.dimensions(operand_dim) - output_dim_size; + CHECK_GE(largest_valid_start_index, 0); + + // Clamp the gather index so that the gather region fits in the operand. + // gather_dim_component_extended_inbound = + // clamp(gather_dim_component_extended, 0, largest_valid_start_index); + + // TODO(b/111078873): This is implementation defined behavior. + bool is_signed = ShapeUtil::ElementIsSigned(indices_shape); + auto gather_dim_component_extended_inbound = EmitIntegralMin( + index.GetConstantWithIndexType(largest_valid_start_index), + EmitIntegralMax(index.GetConstantWithIndexType(0), + gather_dim_component_extended, is_signed), + is_signed); + + operand_index[operand_dim] = ir_builder_->CreateAdd( + operand_index[operand_dim], gather_dim_component_extended_inbound); }; if (indices_shape.dimensions_size() == dim_numbers.index_vector_dim()) { TF_ASSIGN_OR_RETURN(llvm::Value * gather_dim_component, indices_generator(gather_index_index)); - add_to_unsafe_operand_index(gather_dim_component, 0); + add_to_operand_index(gather_dim_component, 0); } else { int64 index_vector_size = indices_shape.dimensions(dim_numbers.index_vector_dim()); @@ -1656,18 +1685,10 @@ StatusOr ElementalIrEmitter::EmitElementalGather( index.GetConstantWithIndexType(i); TF_ASSIGN_OR_RETURN(llvm::Value * gather_dim_component, indices_generator(gather_index_index)); - add_to_unsafe_operand_index(gather_dim_component, i); + add_to_operand_index(gather_dim_component, i); } } - - IrArray::Index safe_operand_index(index_type); - for (int64 i = 0, e = unsafe_operand_index.size(); i < e; i++) { - safe_operand_index.push_back(ir_builder_->CreateURem( - unsafe_operand_index[i], - index.GetConstantWithIndexType(operand_shape.dimensions(i)))); - } - - return operand_generator(safe_operand_index); + return operand_generator(operand_index); } StatusOr ElementalIrEmitter::EmitElementalDynamicUpdateSlice( @@ -1699,19 +1720,20 @@ StatusOr ElementalIrEmitter::EmitElementalDynamicUpdateSlice( // TODO(b/74360564): This is implementation defined behavior, but is // currently respected by all implementations. Change this if we ever decide - // to oficially document different behavior. + // to officially document different behavior. start_index_value = ir_builder_->CreateSExtOrTrunc(start_index_value, index_type); - llvm::Value* input_dim_size = - index_typed_const(input_hlo->shape().dimensions(i)); llvm::Value* update_dim_size = index_typed_const(update_hlo->shape().dimensions(i)); + int64 largest_valid_start_index = + input_hlo->shape().dimensions(i) - update_hlo->shape().dimensions(i); + CHECK_GE(largest_valid_start_index, 0); - start_index_value = - EmitIntegralMin(ir_builder_->CreateSub(input_dim_size, update_dim_size), - EmitIntegralMax(index_typed_const(0), start_index_value, - /*is_signed=*/true), - /*is_signed=*/true); + bool is_signed = ShapeUtil::ElementIsSigned(start_hlo->shape()); + start_index_value = EmitIntegralMin( + index_typed_const(largest_valid_start_index), + EmitIntegralMax(index_typed_const(0), start_index_value, is_signed), + is_signed); start_index_value->setName( AsStringRef(IrName(hlo, StrCat("start_idx", i)))); diff --git a/tensorflow/compiler/xla/service/elemental_ir_emitter_test.cc b/tensorflow/compiler/xla/service/elemental_ir_emitter_test.cc index 8980d4303353a132ada2b3c685b4f2856c33c6a1..addb016b0481b744ff42ba827104099b6cdc3bb9 100644 --- a/tensorflow/compiler/xla/service/elemental_ir_emitter_test.cc +++ b/tensorflow/compiler/xla/service/elemental_ir_emitter_test.cc @@ -57,8 +57,8 @@ ENTRY main { } )"; - std::unique_ptr lhs = Literal::CreateR3({{{1}, {2}}}); - std::unique_ptr rhs = Literal::CreateR3({{{3}, {4}}}); + std::unique_ptr lhs = LiteralUtil::CreateR3({{{1}, {2}}}); + std::unique_ptr rhs = LiteralUtil::CreateR3({{{3}, {4}}}); RunTest(hlo_text, {lhs.get(), rhs.get()}); } } // namespace diff --git a/tensorflow/compiler/xla/service/flatten_call_graph_test.cc b/tensorflow/compiler/xla/service/flatten_call_graph_test.cc index d3854b40de3572a60df1ad99d8a4589f59ad7194..8f6608241ed02bbb7e9fde9b6d767c002435e777 100644 --- a/tensorflow/compiler/xla/service/flatten_call_graph_test.cc +++ b/tensorflow/compiler/xla/service/flatten_call_graph_test.cc @@ -15,7 +15,7 @@ limitations under the License. #include "tensorflow/compiler/xla/service/flatten_call_graph.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/service/call_graph.h" #include "tensorflow/compiler/xla/service/hlo_computation.h" #include "tensorflow/compiler/xla/shape_util.h" @@ -80,7 +80,7 @@ class FlattenCallGraphTest : public HloTestBase { HloInstruction* param0 = builder.AddInstruction( HloInstruction::CreateParameter(0, kScalarShape, "param0")); HloInstruction* zero = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(0.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(0.0f))); builder.AddInstruction(HloInstruction::CreateBinary( ShapeUtil::MakeShape(PRED, {}), HloOpcode::kGt, param0, zero)); return builder.Build(); @@ -157,7 +157,7 @@ TEST_F(FlattenCallGraphTest, SharedWhileConditionAndBody) { builder.AddInstruction(HloInstruction::CreateParameter( 0, ShapeUtil::MakeShape(PRED, {}), "param0")); HloInstruction* false_constant = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(false))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(false))); builder.AddInstruction( HloInstruction::CreateBinary(ShapeUtil::MakeShape(PRED, {}), HloOpcode::kEq, param0, false_constant)); @@ -168,7 +168,7 @@ TEST_F(FlattenCallGraphTest, SharedWhileConditionAndBody) { { HloComputation::Builder builder(TestName() + ".entry"); HloInstruction* false_constant = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(false))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(false))); builder.AddInstruction(HloInstruction::CreateWhile( ShapeUtil::MakeShape(PRED, {}), cond_computation, cond_computation, false_constant)); @@ -232,11 +232,11 @@ TEST_F(FlattenCallGraphTest, FlattenCallsInConditional) { // computation in the true and false branch. HloComputation::Builder builder(TestName()); auto pred = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(true))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(true))); auto constant1 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(56.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(56.0f))); auto constant2 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(12.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(12.0f))); builder.AddInstruction(HloInstruction::CreateConditional( kScalarShape, pred, constant1, sub_computation, constant2, sub_computation)); diff --git a/tensorflow/compiler/xla/service/gather_expander.cc b/tensorflow/compiler/xla/service/gather_expander.cc index 7cd2c9c136acac46e8e6c548c9e58b9bc8e6e0d2..e3a42d0d06be9e4c9ef96ed2e6ff5daa8eebaf3e 100644 --- a/tensorflow/compiler/xla/service/gather_expander.cc +++ b/tensorflow/compiler/xla/service/gather_expander.cc @@ -15,6 +15,7 @@ limitations under the License. #include +#include "tensorflow/compiler/xla/literal_util.h" #include "tensorflow/compiler/xla/service/gather_expander.h" #include "tensorflow/compiler/xla/service/hlo_creation_utils.h" #include "tensorflow/compiler/xla/service/hlo_instruction.h" @@ -113,7 +114,7 @@ static StatusOr ExpandIndexVectorIntoOperandSpace( const Shape& index_shape = index_vector->shape(); HloInstruction* zero = computation->AddInstruction(HloInstruction::CreateConstant( - Literal::CreateFromDimensions(index_shape.element_type(), {1}))); + LiteralUtil::CreateFromDimensions(index_shape.element_type(), {1}))); // We extract out individual components from the smaller index and concatenate // them (interspersing zeros as needed) into the larger index. diff --git a/tensorflow/compiler/xla/service/generic_transfer_manager.cc b/tensorflow/compiler/xla/service/generic_transfer_manager.cc index 85e28a0dfe38415974e435106a2d0b75863f2df5..e314a469f00abdb9f60ae812c0b78d273dc95dbe 100644 --- a/tensorflow/compiler/xla/service/generic_transfer_manager.cc +++ b/tensorflow/compiler/xla/service/generic_transfer_manager.cc @@ -20,7 +20,7 @@ limitations under the License. #include #include "tensorflow/compiler/xla/layout_util.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/service/interpreter/platform_id.h" #include "tensorflow/compiler/xla/shape_util.h" #include "tensorflow/compiler/xla/status_macros.h" @@ -158,16 +158,10 @@ Status GenericTransferManager::TransferLiteralToInfeed( return Unimplemented("Generic transfer to Infeed"); } -Status GenericTransferManager::TransferBufferToInfeed( - se::StreamExecutor* executor, int64 size, const void* source) { - return Unimplemented("Generic transfer to Infeed"); -} - Status GenericTransferManager::TransferLiteralFromOutfeed( se::StreamExecutor* executor, const Shape& literal_shape, Literal* literal) { - return Unimplemented( - "Outfeed is not supported on this platform (b/30467474)"); + return Unimplemented("Generic transfer from Outfeed"); } Status GenericTransferManager::ResetDevices( diff --git a/tensorflow/compiler/xla/service/generic_transfer_manager.h b/tensorflow/compiler/xla/service/generic_transfer_manager.h index d216fe7d29e8f2e84ab4f558ee5caec32d07a70a..3cd002c1bf3555cc2d2891c88b3ad648f8d9fd8c 100644 --- a/tensorflow/compiler/xla/service/generic_transfer_manager.h +++ b/tensorflow/compiler/xla/service/generic_transfer_manager.h @@ -61,9 +61,6 @@ class GenericTransferManager : public TransferManager { int64 GetByteSizeRequirement(const Shape& shape) const override; protected: - Status TransferBufferToInfeed(se::StreamExecutor* executor, int64 size, - const void* source) override; - Status WriteSingleTupleIndexTable( se::Stream* stream, tensorflow::gtl::ArraySlice elements, diff --git a/tensorflow/compiler/xla/service/gpu/BUILD b/tensorflow/compiler/xla/service/gpu/BUILD index af6d298589eb58fbae96158bd264c2b085cb66d1..59172e53d3f2e8f295e1bf8530ebc69ae8dbc30f 100644 --- a/tensorflow/compiler/xla/service/gpu/BUILD +++ b/tensorflow/compiler/xla/service/gpu/BUILD @@ -150,7 +150,7 @@ cc_library( ":parallel_loop_emitter", ":partition_assignment", ":while_transformer", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:status_macros", "//tensorflow/compiler/xla:statusor", @@ -165,6 +165,7 @@ cc_library( "//tensorflow/compiler/xla/service/llvm_ir:fused_ir_emitter", "//tensorflow/compiler/xla/service/llvm_ir:ir_array", "//tensorflow/compiler/xla/service/llvm_ir:kernel_support_library", + "//tensorflow/compiler/xla/service/llvm_ir:kernel_tiling", "//tensorflow/compiler/xla/service/llvm_ir:llvm_loop", "//tensorflow/compiler/xla/service/llvm_ir:llvm_util", "//tensorflow/compiler/xla/service/llvm_ir:loop_emitter", @@ -199,7 +200,7 @@ cc_library( srcs = ["elemental_ir_emitter.cc"], hdrs = ["elemental_ir_emitter.h"], deps = [ - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:status_macros", "//tensorflow/compiler/xla:statusor", @@ -246,6 +247,7 @@ cc_library( "//tensorflow/compiler/xla/service:hlo_execution_profile", "//tensorflow/compiler/xla/service:pool", "//tensorflow/core:lib", + "//tensorflow/core:ptr_util", "//tensorflow/core:stream_executor_no_cuda", ], ) @@ -264,6 +266,7 @@ cc_library( "infeed_thunk.cc", "kernel_thunk.cc", "memset_thunk.cc", + "outfeed_thunk.cc", "sequential_thunk.cc", "thunk_schedule.cc", "tuple_thunk.cc", @@ -281,6 +284,7 @@ cc_library( "infeed_thunk.h", "kernel_thunk.h", "memset_thunk.h", + "outfeed_thunk.h", "sequential_thunk.h", "thunk.h", "thunk_schedule.h", @@ -288,15 +292,16 @@ cc_library( "while_thunk.h", ], deps = [ - ":backend_configs", ":buffer_allocations", ":cudnn_convolution_runner", ":hlo_execution_profiler", ":infeed_manager", ":ir_emission_utils", + ":outfeed_manager", ":partition_assignment", ":stream_assignment", "//tensorflow/compiler/xla:array2d", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_tree", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:status", @@ -350,6 +355,7 @@ cc_library( ":cudnn_convolution_runner", ":gpu_executable", ":ir_emission_utils", + "//tensorflow/compiler/xla:literal_util", "//tensorflow/compiler/xla/service:device_memory_allocator", "//tensorflow/compiler/xla/service:hlo", "//tensorflow/compiler/xla/service:hlo_pass", @@ -381,7 +387,7 @@ cc_library( hdrs = ["cudnn_convolution_rewriter.h"], deps = [ ":ir_emission_utils", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:util", "//tensorflow/compiler/xla:window_util", "//tensorflow/compiler/xla:xla_data_proto", @@ -442,6 +448,7 @@ cc_library( srcs = ["multi_output_fusion.cc"], hdrs = ["multi_output_fusion.h"], deps = [ + ":ir_emission_utils", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla/service:hlo", "//tensorflow/compiler/xla/service:multi_output_fusion", @@ -515,6 +522,7 @@ cc_library( hdrs = ["pad_insertion.h"], deps = [ ":ir_emission_utils", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:literal_util", "//tensorflow/compiler/xla:util", "//tensorflow/compiler/xla:window_util", @@ -531,7 +539,10 @@ cc_library( hdrs = ["gpu_transfer_manager.h"], deps = [ ":gpu_compiler", + ":outfeed_manager", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:shape_tree", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:status_macros", "//tensorflow/compiler/xla:statusor", @@ -622,24 +633,46 @@ cc_library( hdrs = ["cudnn_batchnorm_rewriter.h"], deps = [ ":ir_emission_utils", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:literal_util", "//tensorflow/compiler/xla/service:hlo", "//tensorflow/compiler/xla/service:hlo_pass", ], ) +cc_library( + name = "xfeed_queue", + hdrs = ["xfeed_queue.h"], + deps = ["//tensorflow/core:lib"], +) + cc_library( name = "infeed_manager", srcs = ["infeed_manager.cc"], hdrs = ["infeed_manager.h"], deps = [ + ":xfeed_queue", + "//tensorflow/compiler/xla:shape_tree", "//tensorflow/compiler/xla:types", "//tensorflow/compiler/xla:util", - "//tensorflow/core:lib", "//tensorflow/core:stream_executor_no_cuda", ], ) +cc_library( + name = "outfeed_manager", + srcs = ["outfeed_manager.cc"], + hdrs = ["outfeed_manager.h"], + deps = [ + ":xfeed_queue", + "//tensorflow/compiler/xla:literal", + "//tensorflow/compiler/xla:shape_tree", + "//tensorflow/compiler/xla:shape_util", + "//tensorflow/compiler/xla:util", + "//tensorflow/core:lib", + ], +) + cc_library( name = "gpu_layout_assignment", srcs = ["gpu_layout_assignment.cc"], @@ -714,7 +747,7 @@ cc_library( srcs = ["while_transformer.cc"], hdrs = ["while_transformer.h"], deps = [ - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:status_macros", "//tensorflow/compiler/xla:statusor", @@ -769,6 +802,7 @@ cc_library( hdrs = ["stream_executor_util.h"], deps = [ "//tensorflow/compiler/xla:shape_util", + "//tensorflow/compiler/xla:statusor", "//tensorflow/compiler/xla:xla_data_proto", "//tensorflow/core:stream_executor_no_cuda", ], diff --git a/tensorflow/compiler/xla/service/gpu/conditional_thunk.cc b/tensorflow/compiler/xla/service/gpu/conditional_thunk.cc index 77a48965e031349b045a956fd3f28c58607328e5..5780e0af40699bb6ac2c190c09cd02023fb44db7 100644 --- a/tensorflow/compiler/xla/service/gpu/conditional_thunk.cc +++ b/tensorflow/compiler/xla/service/gpu/conditional_thunk.cc @@ -16,6 +16,7 @@ limitations under the License. #include "tensorflow/compiler/xla/service/gpu/conditional_thunk.h" #include "tensorflow/compiler/xla/ptr_util.h" +#include "tensorflow/compiler/xla/service/gpu/hlo_execution_profiler.h" #include "tensorflow/compiler/xla/util.h" #include "tensorflow/core/lib/core/errors.h" @@ -32,8 +33,11 @@ ConditionalThunk::ConditionalThunk( predicate_buffer_index_(predicate_buffer_index), true_operand_buffer_index_(true_operand_buffer_index), false_operand_buffer_index_(false_operand_buffer_index), - true_thunk_(std::move(true_thunk_sequence), hlo), - false_thunk_(std::move(false_thunk_sequence), hlo) {} + // Pass nullptr as the HloInstruction* to the true_thunk_ and false_thunk_ + // constructors because these SequentialThunks are logically "part of" + // this ConditionalThunk, and shouldn't be profiled separately from it. + true_thunk_(std::move(true_thunk_sequence), nullptr), + false_thunk_(std::move(false_thunk_sequence), nullptr) {} Status ConditionalThunk::Initialize(const GpuExecutable& executable, se::StreamExecutor* executor) { @@ -43,7 +47,9 @@ Status ConditionalThunk::Initialize(const GpuExecutable& executable, } Status ConditionalThunk::ExecuteOnStream( - const BufferAllocations& buffer_allocations, se::Stream* stream) { + const BufferAllocations& buffer_allocations, se::Stream* stream, + HloExecutionProfiler* profiler) { + auto op_profiler = profiler->MakeScopedInstructionProfiler(hlo_instruction()); // Copy the predicate value from device. bool predicate; se::DeviceMemoryBase predicate_address = @@ -59,10 +65,15 @@ Status ConditionalThunk::ExecuteOnStream( // Execute the true or the false computation depending on the value of the // predicate. if (predicate) { - TF_RETURN_IF_ERROR(true_thunk_.ExecuteOnStream(buffer_allocations, stream)); + profiler->StartHloComputation(); + TF_RETURN_IF_ERROR( + true_thunk_.ExecuteOnStream(buffer_allocations, stream, profiler)); + profiler->FinishHloComputation(hlo_instruction()->true_computation()); } else { + profiler->StartHloComputation(); TF_RETURN_IF_ERROR( - false_thunk_.ExecuteOnStream(buffer_allocations, stream)); + false_thunk_.ExecuteOnStream(buffer_allocations, stream, profiler)); + profiler->FinishHloComputation(hlo_instruction()->false_computation()); } return Status::OK(); diff --git a/tensorflow/compiler/xla/service/gpu/conditional_thunk.h b/tensorflow/compiler/xla/service/gpu/conditional_thunk.h index ee03865d174469285a9e98b8a30fea90d997df37..aef24342c9fe182eb54b1c2beff840a76e7b8115 100644 --- a/tensorflow/compiler/xla/service/gpu/conditional_thunk.h +++ b/tensorflow/compiler/xla/service/gpu/conditional_thunk.h @@ -17,6 +17,7 @@ limitations under the License. #define TENSORFLOW_COMPILER_XLA_SERVICE_GPU_CONDITIONAL_THUNK_H_ #include "tensorflow/compiler/xla/service/gpu/buffer_allocations.h" +#include "tensorflow/compiler/xla/service/gpu/hlo_execution_profiler.h" #include "tensorflow/compiler/xla/service/gpu/sequential_thunk.h" #include "tensorflow/compiler/xla/service/gpu/thunk.h" #include "tensorflow/compiler/xla/service/hlo_instruction.h" @@ -50,7 +51,8 @@ class ConditionalThunk : public Thunk { Status Initialize(const GpuExecutable& executable, se::StreamExecutor* executor) override; Status ExecuteOnStream(const BufferAllocations& buffer_allocations, - se::Stream* stream) override; + se::Stream* stream, + HloExecutionProfiler* profiler) override; private: BufferAllocation::Slice predicate_buffer_index_; diff --git a/tensorflow/compiler/xla/service/gpu/convolution_thunk.cc b/tensorflow/compiler/xla/service/gpu/convolution_thunk.cc index f0881124128c9b043392ffc4fa3aee2cd5b754c7..7833a4077e6c6ee4960665f37fb01a35530fd302 100644 --- a/tensorflow/compiler/xla/service/gpu/convolution_thunk.cc +++ b/tensorflow/compiler/xla/service/gpu/convolution_thunk.cc @@ -18,6 +18,7 @@ limitations under the License. #include #include "tensorflow/compiler/xla/service/gpu/cudnn_convolution_runner.h" +#include "tensorflow/compiler/xla/service/gpu/hlo_execution_profiler.h" #include "tensorflow/compiler/xla/types.h" #include "tensorflow/compiler/xla/util.h" #include "tensorflow/core/lib/strings/strcat.h" @@ -55,7 +56,8 @@ ConvolutionThunk::ConvolutionThunk( tensor_ops_enabled_(tensor_ops_enabled) {} Status ConvolutionThunk::ExecuteOnStream( - const BufferAllocations& buffer_allocations, se::Stream* stream) { + const BufferAllocations& buffer_allocations, se::Stream* stream, + HloExecutionProfiler* profiler) { se::DeviceMemoryBase input_data = buffer_allocations.GetDeviceAddress(input_buffer_); se::DeviceMemoryBase filter_data = @@ -68,6 +70,7 @@ Status ConvolutionThunk::ExecuteOnStream( se::dnn::AlgorithmConfig algorithm_config( se::dnn::AlgorithmDesc(algorithm_, tensor_ops_enabled_)); + auto op_profiler = profiler->MakeScopedInstructionProfiler(hlo_instruction()); TF_RETURN_IF_ERROR(RunCudnnConvolution( convolution_kind_, input_shape_, filter_shape_, output_shape_, input_data, filter_data, output_data, scratch, window_, dim_nums_, algorithm_config, diff --git a/tensorflow/compiler/xla/service/gpu/convolution_thunk.h b/tensorflow/compiler/xla/service/gpu/convolution_thunk.h index 6d845025b1aef2b0a5f147401b6db0598ba94d6d..d76ca6698dcf462c3c4961ce6a9784822af3a81f 100644 --- a/tensorflow/compiler/xla/service/gpu/convolution_thunk.h +++ b/tensorflow/compiler/xla/service/gpu/convolution_thunk.h @@ -20,6 +20,7 @@ limitations under the License. #include "tensorflow/compiler/xla/service/gpu/buffer_allocations.h" #include "tensorflow/compiler/xla/service/gpu/cudnn_convolution_runner.h" #include "tensorflow/compiler/xla/service/gpu/gpu_executable.h" +#include "tensorflow/compiler/xla/service/gpu/hlo_execution_profiler.h" #include "tensorflow/compiler/xla/service/gpu/thunk.h" #include "tensorflow/compiler/xla/service/hlo_instruction.h" #include "tensorflow/compiler/xla/types.h" @@ -66,7 +67,8 @@ class ConvolutionThunk : public Thunk { // Does the convolution for the thunk on "stream". Status ExecuteOnStream(const BufferAllocations& buffer_allocations, - se::Stream* stream) override; + se::Stream* stream, + HloExecutionProfiler* profiler) override; private: class ScratchAllocator; diff --git a/tensorflow/compiler/xla/service/gpu/copy_thunk.cc b/tensorflow/compiler/xla/service/gpu/copy_thunk.cc index ee38c0318a878c7bcdc02afdcd146bfb4498d9a2..92e03f94c11f68082f0a8caa64f82e8533557194 100644 --- a/tensorflow/compiler/xla/service/gpu/copy_thunk.cc +++ b/tensorflow/compiler/xla/service/gpu/copy_thunk.cc @@ -15,6 +15,7 @@ limitations under the License. #include "tensorflow/compiler/xla/service/gpu/copy_thunk.h" +#include "tensorflow/compiler/xla/service/gpu/hlo_execution_profiler.h" #include "tensorflow/core/platform/stream_executor_no_cuda.h" namespace xla { @@ -30,9 +31,11 @@ HostToDeviceCopyThunk::HostToDeviceCopyThunk( mem_size_(mem_size) {} Status HostToDeviceCopyThunk::ExecuteOnStream( - const BufferAllocations& buffer_allocations, se::Stream* stream) { + const BufferAllocations& buffer_allocations, se::Stream* stream, + HloExecutionProfiler* profiler) { se::DeviceMemoryBase destination_data = buffer_allocations.GetDeviceAddress(destination_buffer_); + auto op_profiler = profiler->MakeScopedInstructionProfiler(hlo_instruction()); stream->ThenMemcpy(&destination_data, source_address_, mem_size_); return Status::OK(); } @@ -47,11 +50,13 @@ DeviceToDeviceCopyThunk::DeviceToDeviceCopyThunk( mem_size_(mem_size) {} Status DeviceToDeviceCopyThunk::ExecuteOnStream( - const BufferAllocations& buffer_allocations, se::Stream* stream) { + const BufferAllocations& buffer_allocations, se::Stream* stream, + HloExecutionProfiler* profiler) { se::DeviceMemoryBase destination_data = buffer_allocations.GetDeviceAddress(destination_buffer_); se::DeviceMemoryBase source_data = buffer_allocations.GetDeviceAddress(source_buffer_); + auto op_profiler = profiler->MakeScopedInstructionProfiler(hlo_instruction()); stream->ThenMemcpy(&destination_data, source_data, mem_size_); return Status::OK(); } diff --git a/tensorflow/compiler/xla/service/gpu/copy_thunk.h b/tensorflow/compiler/xla/service/gpu/copy_thunk.h index 8b128386f61636de9ac41e856a2b00c578e05735..91564b520acae1839e0a466cf580db00bdf57e46 100644 --- a/tensorflow/compiler/xla/service/gpu/copy_thunk.h +++ b/tensorflow/compiler/xla/service/gpu/copy_thunk.h @@ -18,6 +18,7 @@ limitations under the License. #include "tensorflow/compiler/xla/service/buffer_assignment.h" #include "tensorflow/compiler/xla/service/gpu/buffer_allocations.h" +#include "tensorflow/compiler/xla/service/gpu/hlo_execution_profiler.h" #include "tensorflow/compiler/xla/service/gpu/thunk.h" #include "tensorflow/compiler/xla/service/hlo_instruction.h" #include "tensorflow/core/platform/stream_executor_no_cuda.h" @@ -40,7 +41,8 @@ class HostToDeviceCopyThunk : public Thunk { HostToDeviceCopyThunk& operator=(const HostToDeviceCopyThunk&) = delete; Status ExecuteOnStream(const BufferAllocations& buffer_allocations, - se::Stream* stream) override; + se::Stream* stream, + HloExecutionProfiler* profiler) override; private: const void* source_address_; @@ -63,7 +65,8 @@ class DeviceToDeviceCopyThunk : public Thunk { DeviceToDeviceCopyThunk& operator=(const DeviceToDeviceCopyThunk&) = delete; Status ExecuteOnStream(const BufferAllocations& buffer_allocations, - se::Stream* stream) override; + se::Stream* stream, + HloExecutionProfiler* profiler) override; private: const BufferAllocation::Slice source_buffer_; diff --git a/tensorflow/compiler/xla/service/gpu/cudnn_batchnorm_rewriter.cc b/tensorflow/compiler/xla/service/gpu/cudnn_batchnorm_rewriter.cc index db6924c742e4a949a3e939b6d6659e92c2d1e312..60289506524759580dbb9b82147c78c4ce1cb25e 100644 --- a/tensorflow/compiler/xla/service/gpu/cudnn_batchnorm_rewriter.cc +++ b/tensorflow/compiler/xla/service/gpu/cudnn_batchnorm_rewriter.cc @@ -14,6 +14,7 @@ limitations under the License. ==============================================================================*/ #include "tensorflow/compiler/xla/service/gpu/cudnn_batchnorm_rewriter.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/literal_util.h" #include "tensorflow/compiler/xla/service/gpu/ir_emission_utils.h" @@ -66,11 +67,12 @@ Status Visitor::HandleBatchNormInference(HloInstruction* batch_norm) { return Status::OK(); } - HloInstruction* epsilon = computation_->AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(batch_norm->epsilon()))); + HloInstruction* epsilon = + computation_->AddInstruction(HloInstruction::CreateConstant( + LiteralUtil::CreateR0(batch_norm->epsilon()))); HloInstruction* feature_index = computation_->AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR0(batch_norm->feature_index()))); + LiteralUtil::CreateR0(batch_norm->feature_index()))); std::vector operands(batch_norm->operands().begin(), batch_norm->operands().end()); @@ -101,11 +103,12 @@ Status Visitor::HandleBatchNormTraining(HloInstruction* batch_norm) { return Status::OK(); } - HloInstruction* epsilon = computation_->AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(batch_norm->epsilon()))); + HloInstruction* epsilon = + computation_->AddInstruction(HloInstruction::CreateConstant( + LiteralUtil::CreateR0(batch_norm->epsilon()))); HloInstruction* feature_index = computation_->AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR0(batch_norm->feature_index()))); + LiteralUtil::CreateR0(batch_norm->feature_index()))); std::vector operands(batch_norm->operands().begin(), batch_norm->operands().end()); @@ -126,12 +129,17 @@ Status Visitor::HandleBatchNormTraining(HloInstruction* batch_norm) { HloInstruction* variance_plus_epsilon = computation_->AddInstruction(HloInstruction::CreateBinary( inverse_stddev->shape(), HloOpcode::kPower, inverse_stddev, - computation_->AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(-2))))); + computation_->AddInstruction(HloInstruction::CreateBroadcast( + inverse_stddev->shape(), + computation_->AddInstruction(HloInstruction::CreateConstant( + LiteralUtil::CreateR0(-2))), + {})))); HloInstruction* variance = computation_->AddInstruction(HloInstruction::CreateBinary( variance_plus_epsilon->shape(), HloOpcode::kSubtract, - variance_plus_epsilon, epsilon)); + variance_plus_epsilon, + computation_->AddInstruction(HloInstruction::CreateBroadcast( + variance_plus_epsilon->shape(), epsilon, {})))); // Repackage the results. std::unique_ptr new_tuple = HloInstruction::CreateTuple({ @@ -164,23 +172,29 @@ Status Visitor::HandleBatchNormGrad(HloInstruction* batch_norm) { return Status::OK(); } - HloInstruction* epsilon = computation_->AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(batch_norm->epsilon()))); + HloInstruction* epsilon = + computation_->AddInstruction(HloInstruction::CreateConstant( + LiteralUtil::CreateR0(batch_norm->epsilon()))); HloInstruction* feature_index = computation_->AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR0(batch_norm->feature_index()))); + LiteralUtil::CreateR0(batch_norm->feature_index()))); // The cudnn libcall expects its input to be rsqrt(variance + epsilon), but // the batchnorm HLO takes plain variance as input. Fix it up. HloInstruction* var_plus_epsilon = computation_->AddInstruction(HloInstruction::CreateBinary( batch_norm->operand(3)->shape(), HloOpcode::kAdd, - batch_norm->mutable_operand(3), epsilon)); + batch_norm->mutable_operand(3), + computation_->AddInstruction(HloInstruction::CreateBroadcast( + batch_norm->operand(3)->shape(), epsilon, {})))); HloInstruction* inverse_stddev = computation_->AddInstruction(HloInstruction::CreateBinary( var_plus_epsilon->shape(), HloOpcode::kPower, var_plus_epsilon, - computation_->AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(-.5))))); + computation_->AddInstruction(HloInstruction::CreateBroadcast( + var_plus_epsilon->shape(), + computation_->AddInstruction(HloInstruction::CreateConstant( + LiteralUtil::CreateR0(-.5))), + {})))); std::vector operands(batch_norm->operands().begin(), batch_norm->operands().end()); diff --git a/tensorflow/compiler/xla/service/gpu/cudnn_batchnorm_thunk.cc b/tensorflow/compiler/xla/service/gpu/cudnn_batchnorm_thunk.cc index 68099fd63847ef9993f9bc7ac0e28b2939631b35..7b172812c36bb141787ef3a9285d6f7ce13e343b 100644 --- a/tensorflow/compiler/xla/service/gpu/cudnn_batchnorm_thunk.cc +++ b/tensorflow/compiler/xla/service/gpu/cudnn_batchnorm_thunk.cc @@ -17,6 +17,7 @@ limitations under the License. #include +#include "tensorflow/compiler/xla/service/gpu/hlo_execution_profiler.h" #include "tensorflow/compiler/xla/service/gpu/ir_emission_utils.h" #include "tensorflow/compiler/xla/types.h" #include "tensorflow/compiler/xla/util.h" @@ -99,13 +100,15 @@ CudnnBatchNormForwardInferenceThunk::CudnnBatchNormForwardInferenceThunk( } Status CudnnBatchNormForwardInferenceThunk::ExecuteOnStream( - const BufferAllocations& buffer_allocations, se::Stream* stream) { + const BufferAllocations& buffer_allocations, se::Stream* stream, + HloExecutionProfiler* profiler) { dnn::BatchDescriptor operand_desc; dnn::BatchDescriptor scale_offset_desc; std::tie(operand_desc, scale_offset_desc) = MakeDescriptors(hlo_instruction()->shape(), feature_index_); se::DeviceMemory output(buffer_allocations.GetDeviceAddress(output_)); + auto op_profiler = profiler->MakeScopedInstructionProfiler(hlo_instruction()); stream->ThenBatchNormalizationForward( se::DeviceMemory(buffer_allocations.GetDeviceAddress(operand_)), se::DeviceMemory(buffer_allocations.GetDeviceAddress(scale_)), @@ -123,6 +126,7 @@ Status CudnnBatchNormForwardInferenceThunk::ExecuteOnStream( /*is_training=*/false, // /*var_to_inv_var=*/nullptr, // /*inv_var_to_var=*/nullptr); + if (!stream->ok()) { return InternalError("BatchNormalizationForward call failed."); } @@ -158,7 +162,8 @@ CudnnBatchNormForwardTrainingThunk::CudnnBatchNormForwardTrainingThunk( } Status CudnnBatchNormForwardTrainingThunk::ExecuteOnStream( - const BufferAllocations& buffer_allocations, se::Stream* stream) { + const BufferAllocations& buffer_allocations, se::Stream* stream, + HloExecutionProfiler* profiler) { dnn::BatchDescriptor operand_desc; dnn::BatchDescriptor scale_offset_desc; // The BatchNormTraining HLO outputs a tuple of three elements: output data, @@ -175,6 +180,7 @@ Status CudnnBatchNormForwardTrainingThunk::ExecuteOnStream( buffer_allocations.GetDeviceAddress(output_inv_stddev_)); se::DeviceMemory null_device_ptr(nullptr); + auto op_profiler = profiler->MakeScopedInstructionProfiler(hlo_instruction()); stream->ThenBatchNormalizationForward( se::DeviceMemory(buffer_allocations.GetDeviceAddress(operand_)), se::DeviceMemory(buffer_allocations.GetDeviceAddress(scale_)), @@ -240,7 +246,8 @@ CudnnBatchNormBackwardThunk::CudnnBatchNormBackwardThunk( } Status CudnnBatchNormBackwardThunk::ExecuteOnStream( - const BufferAllocations& buffer_allocations, se::Stream* stream) { + const BufferAllocations& buffer_allocations, se::Stream* stream, + HloExecutionProfiler* profiler) { dnn::BatchDescriptor operand_desc; dnn::BatchDescriptor scale_offset_desc; @@ -257,6 +264,7 @@ Status CudnnBatchNormBackwardThunk::ExecuteOnStream( se::DeviceMemory output_grad_offset( buffer_allocations.GetDeviceAddress(output_grad_offset_)); + auto op_profiler = profiler->MakeScopedInstructionProfiler(hlo_instruction()); stream->ThenBatchNormalizationBackward( se::DeviceMemory( buffer_allocations.GetDeviceAddress(grad_output_)), diff --git a/tensorflow/compiler/xla/service/gpu/cudnn_batchnorm_thunk.h b/tensorflow/compiler/xla/service/gpu/cudnn_batchnorm_thunk.h index 874f85a863092ee05ae5df1f92d732318c5a0554..d2143b3952984722d136757255aa0aa60e9cab7e 100644 --- a/tensorflow/compiler/xla/service/gpu/cudnn_batchnorm_thunk.h +++ b/tensorflow/compiler/xla/service/gpu/cudnn_batchnorm_thunk.h @@ -18,6 +18,7 @@ limitations under the License. #include "tensorflow/compiler/xla/service/buffer_assignment.h" #include "tensorflow/compiler/xla/service/gpu/buffer_allocations.h" +#include "tensorflow/compiler/xla/service/gpu/hlo_execution_profiler.h" #include "tensorflow/compiler/xla/service/gpu/thunk.h" #include "tensorflow/compiler/xla/service/hlo_instruction.h" #include "tensorflow/compiler/xla/types.h" @@ -60,7 +61,8 @@ class CudnnBatchNormForwardInferenceThunk : public Thunk { const CudnnBatchNormForwardInferenceThunk&) = delete; Status ExecuteOnStream(const BufferAllocations& buffer_allocations, - se::Stream* stream) override; + se::Stream* stream, + HloExecutionProfiler* profiler) override; private: BufferAllocation::Slice operand_; @@ -90,7 +92,8 @@ class CudnnBatchNormForwardTrainingThunk : public Thunk { const CudnnBatchNormForwardTrainingThunk&) = delete; Status ExecuteOnStream(const BufferAllocations& buffer_allocations, - se::Stream* stream) override; + se::Stream* stream, + HloExecutionProfiler* profiler) override; private: BufferAllocation::Slice operand_; @@ -123,7 +126,8 @@ class CudnnBatchNormBackwardThunk : public Thunk { delete; Status ExecuteOnStream(const BufferAllocations& buffer_allocations, - se::Stream* stream) override; + se::Stream* stream, + HloExecutionProfiler* profiler) override; private: BufferAllocation::Slice operand_; diff --git a/tensorflow/compiler/xla/service/gpu/cudnn_convolution_algorithm_picker.cc b/tensorflow/compiler/xla/service/gpu/cudnn_convolution_algorithm_picker.cc index 3dc98c4c93ea2b9b68dd3ee27794a39847f8756c..5a63e65208ac3e8e23944bc31634f4d29d91c10c 100644 --- a/tensorflow/compiler/xla/service/gpu/cudnn_convolution_algorithm_picker.cc +++ b/tensorflow/compiler/xla/service/gpu/cudnn_convolution_algorithm_picker.cc @@ -14,6 +14,7 @@ limitations under the License. ==============================================================================*/ #include "tensorflow/compiler/xla/service/gpu/cudnn_convolution_algorithm_picker.h" +#include "tensorflow/compiler/xla/literal_util.h" #include "tensorflow/compiler/xla/service/gpu/backend_configs.pb.h" #include "tensorflow/compiler/xla/service/gpu/convolution_thunk.h" #include "tensorflow/compiler/xla/service/gpu/ir_emission_utils.h" @@ -80,8 +81,7 @@ bool ShouldIncludeWinogradNonfusedAlgo(const Shape& input_shape, const ConvolutionDimensionNumbers& dnums, se::StreamExecutor* stream_exec) { // Skip this check for cudnn7 and newer. - auto version = - stream_exec->AsDnn()->GetVersion(); + auto version = stream_exec->AsDnn()->GetVersion(); if (version.ok() && version.ValueOrDie().major_version() >= 7) { return true; } @@ -338,8 +338,8 @@ StatusOr CudnnConvolutionAlgorithmPicker::RunOnInstruction( computation->AddInstruction(HloInstruction::CreateTuple( {computation->AddInstruction(HloInstruction::CreateGetTupleElement( new_call_shape.tuple_shapes(0), new_call, 0)), - computation->AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR1({})))})); + computation->AddInstruction(HloInstruction::CreateConstant( + LiteralUtil::CreateR1({})))})); TF_RETURN_IF_ERROR(instr->parent()->ReplaceInstruction(instr, new_tuple)); return true; diff --git a/tensorflow/compiler/xla/service/gpu/cudnn_convolution_rewriter.cc b/tensorflow/compiler/xla/service/gpu/cudnn_convolution_rewriter.cc index f9dccd287d955502858f6c24ccd4de80256fc148..905b5ee8767d0fa0514c7f1abf83bc089cd08045 100644 --- a/tensorflow/compiler/xla/service/gpu/cudnn_convolution_rewriter.cc +++ b/tensorflow/compiler/xla/service/gpu/cudnn_convolution_rewriter.cc @@ -18,7 +18,7 @@ limitations under the License. #include #include -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/service/dfs_hlo_visitor_with_default.h" #include "tensorflow/compiler/xla/service/gpu/ir_emission_utils.h" #include "tensorflow/compiler/xla/service/hlo_computation.h" diff --git a/tensorflow/compiler/xla/service/gpu/elemental_ir_emitter.cc b/tensorflow/compiler/xla/service/gpu/elemental_ir_emitter.cc index 27d2c3e491bfc2108cbd168d1a5e1575c2eed11f..e594cec2f8d5743b742e07db70a71cc81279f9b7 100644 --- a/tensorflow/compiler/xla/service/gpu/elemental_ir_emitter.cc +++ b/tensorflow/compiler/xla/service/gpu/elemental_ir_emitter.cc @@ -29,7 +29,7 @@ limitations under the License. #include "llvm/IR/Intrinsics.h" #include "llvm/IR/Module.h" #include "llvm/IR/Type.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/primitive_util.h" #include "tensorflow/compiler/xla/service/hlo_opcode.h" #include "tensorflow/compiler/xla/service/llvm_ir/ir_array.h" diff --git a/tensorflow/compiler/xla/service/gpu/fft_thunk.cc b/tensorflow/compiler/xla/service/gpu/fft_thunk.cc index e14ee6918bf148861ecccac99355fccf7ae93103..0cdddf8bcfd4e849b311bf810eda471d79dbf106 100644 --- a/tensorflow/compiler/xla/service/gpu/fft_thunk.cc +++ b/tensorflow/compiler/xla/service/gpu/fft_thunk.cc @@ -17,6 +17,7 @@ limitations under the License. #include +#include "tensorflow/compiler/xla/service/gpu/hlo_execution_profiler.h" #include "tensorflow/compiler/xla/types.h" #include "tensorflow/compiler/xla/util.h" #include "tensorflow/core/lib/strings/strcat.h" @@ -107,7 +108,8 @@ FftThunk::FftThunk(FftType fft_type, output_shape_(output_shape) {} Status FftThunk::ExecuteOnStream(const BufferAllocations& buffer_allocations, - se::Stream* stream) { + se::Stream* stream, + HloExecutionProfiler* profiler) { VLOG(3) << "FFT type: " << FftTypeToString(fft_type_); VLOG(3) << "Input shape: " << ShapeUtil::HumanStringWithLayout(input_shape_); VLOG(3) << "Output shape: " @@ -116,6 +118,7 @@ Status FftThunk::ExecuteOnStream(const BufferAllocations& buffer_allocations, FftScratchAllocator scratch_allocator(buffer_allocations.device_ordinal(), buffer_allocations.memory_allocator()); + auto op_profiler = profiler->MakeScopedInstructionProfiler(hlo_instruction()); if (fft_plan_ == nullptr) { const int64 fft_rank = fft_length_.size(); CHECK_LE(fft_rank, 3); diff --git a/tensorflow/compiler/xla/service/gpu/fft_thunk.h b/tensorflow/compiler/xla/service/gpu/fft_thunk.h index b0a22564f3a09bb67a3c01723f6e37c604656d45..8c53be5077b0c5a88d303c729457139c6cb800f1 100644 --- a/tensorflow/compiler/xla/service/gpu/fft_thunk.h +++ b/tensorflow/compiler/xla/service/gpu/fft_thunk.h @@ -19,6 +19,7 @@ limitations under the License. #include "tensorflow/compiler/xla/service/buffer_assignment.h" #include "tensorflow/compiler/xla/service/gpu/buffer_allocations.h" #include "tensorflow/compiler/xla/service/gpu/gpu_executable.h" +#include "tensorflow/compiler/xla/service/gpu/hlo_execution_profiler.h" #include "tensorflow/compiler/xla/service/gpu/thunk.h" #include "tensorflow/compiler/xla/service/hlo_instruction.h" #include "tensorflow/compiler/xla/types.h" @@ -72,7 +73,8 @@ class FftThunk : public Thunk { // Does the FFT for the thunk on "stream". Status ExecuteOnStream(const BufferAllocations& buffer_allocations, - se::Stream* stream) override; + se::Stream* stream, + HloExecutionProfiler* profiler) override; private: const se::fft::Type fft_type_; diff --git a/tensorflow/compiler/xla/service/gpu/for_thunk.cc b/tensorflow/compiler/xla/service/gpu/for_thunk.cc index b36539e0cb8d0a2f4758dd90acbdd8fc7181b8ca..b3a3c5dcb4d77889b65a119f09ddef9ba95d6b52 100644 --- a/tensorflow/compiler/xla/service/gpu/for_thunk.cc +++ b/tensorflow/compiler/xla/service/gpu/for_thunk.cc @@ -16,6 +16,7 @@ limitations under the License. #include "tensorflow/compiler/xla/service/gpu/for_thunk.h" #include "tensorflow/compiler/xla/ptr_util.h" +#include "tensorflow/compiler/xla/service/gpu/hlo_execution_profiler.h" #include "tensorflow/compiler/xla/util.h" #include "tensorflow/core/lib/core/errors.h" @@ -27,8 +28,11 @@ ForThunk::ForThunk(const int64 loop_limit, const HloInstruction* hlo) : Thunk(Kind::kWhile, hlo), loop_limit_(loop_limit), - body_thunk_sequence_( - MakeUnique(std::move(*body_thunk_sequence), hlo)) {} + body_thunk_sequence_(MakeUnique( + // Pass nullptr as the HloInstruction* to the body_thunk_sequence_ + // constructor because this SequentialThunk is logically "part of" + // this ForThunk, and shouldn't be profiled separately from it. + std::move(*body_thunk_sequence), nullptr)) {} Status ForThunk::Initialize(const GpuExecutable& executable, se::StreamExecutor* executor) { @@ -37,11 +41,15 @@ Status ForThunk::Initialize(const GpuExecutable& executable, } Status ForThunk::ExecuteOnStream(const BufferAllocations& buffer_allocations, - se::Stream* stream) { + se::Stream* stream, + HloExecutionProfiler* profiler) { + auto op_profiler = profiler->MakeScopedInstructionProfiler(hlo_instruction()); for (int64 i = 0; i < loop_limit_; ++i) { + profiler->StartHloComputation(); // Invoke loop body thunk sequence. - TF_RETURN_IF_ERROR( - body_thunk_sequence_->ExecuteOnStream(buffer_allocations, stream)); + TF_RETURN_IF_ERROR(body_thunk_sequence_->ExecuteOnStream(buffer_allocations, + stream, profiler)); + profiler->FinishHloComputation(hlo_instruction()->while_body()); } return Status::OK(); } diff --git a/tensorflow/compiler/xla/service/gpu/for_thunk.h b/tensorflow/compiler/xla/service/gpu/for_thunk.h index 41ddfe0ceb1d0516c1c64feca53212a925632209..c2d39071b292c6704e9b5857a68bd8b3f3b9a914 100644 --- a/tensorflow/compiler/xla/service/gpu/for_thunk.h +++ b/tensorflow/compiler/xla/service/gpu/for_thunk.h @@ -19,6 +19,7 @@ limitations under the License. #include #include "tensorflow/compiler/xla/service/gpu/buffer_allocations.h" +#include "tensorflow/compiler/xla/service/gpu/hlo_execution_profiler.h" #include "tensorflow/compiler/xla/service/gpu/sequential_thunk.h" #include "tensorflow/compiler/xla/service/gpu/thunk.h" #include "tensorflow/compiler/xla/service/hlo_instruction.h" @@ -39,7 +40,8 @@ class ForThunk : public Thunk { Status Initialize(const GpuExecutable& executable, se::StreamExecutor* executor) override; Status ExecuteOnStream(const BufferAllocations& buffer_allocations, - se::Stream* stream) override; + se::Stream* stream, + HloExecutionProfiler* profiler) override; private: const int64 loop_limit_; diff --git a/tensorflow/compiler/xla/service/gpu/gemm_thunk.cc b/tensorflow/compiler/xla/service/gpu/gemm_thunk.cc index 79fca43d022816645b8a07b9e806fe9cc3745e7c..dbc7754e251eb8075ab97dd2f36bbc400530fcf5 100644 --- a/tensorflow/compiler/xla/service/gpu/gemm_thunk.cc +++ b/tensorflow/compiler/xla/service/gpu/gemm_thunk.cc @@ -252,7 +252,8 @@ GemmThunk::GemmThunk(const BufferAllocation::Slice& lhs_buffer, alpha_(alpha) {} Status GemmThunk::ExecuteOnStream(const BufferAllocations& buffer_allocations, - se::Stream* stream) { + se::Stream* stream, + HloExecutionProfiler* profiler) { VLOG(2) << "Executing a GemmThunk"; se::DeviceMemoryBase lhs_data = @@ -352,6 +353,7 @@ Status GemmThunk::ExecuteOnStream(const BufferAllocations& buffer_allocations, alpha_, stream); }; + auto op_profiler = profiler->MakeScopedInstructionProfiler(hlo_instruction()); bool launch_ok; if (LayoutUtil::Minor(output_shape_.layout(), 0) == 0) { launch_ok = launch( diff --git a/tensorflow/compiler/xla/service/gpu/gemm_thunk.h b/tensorflow/compiler/xla/service/gpu/gemm_thunk.h index 7a4830d64e7caef5a1170cbdbf8ab373fdaf16e2..939c7f85e35b4fcb943a25aa6346d72798432920 100644 --- a/tensorflow/compiler/xla/service/gpu/gemm_thunk.h +++ b/tensorflow/compiler/xla/service/gpu/gemm_thunk.h @@ -19,6 +19,7 @@ limitations under the License. #include "tensorflow/compiler/xla/service/buffer_assignment.h" #include "tensorflow/compiler/xla/service/gpu/buffer_allocations.h" #include "tensorflow/compiler/xla/service/gpu/gpu_executable.h" +#include "tensorflow/compiler/xla/service/gpu/hlo_execution_profiler.h" #include "tensorflow/compiler/xla/service/gpu/thunk.h" #include "tensorflow/compiler/xla/service/hlo_instruction.h" #include "tensorflow/compiler/xla/xla_data.pb.h" @@ -48,7 +49,8 @@ class GemmThunk : public Thunk { // Does the gemm operation for the thunk on "stream", which must be non-null. Status ExecuteOnStream(const BufferAllocations& buffer_allocations, - se::Stream* stream) override; + se::Stream* stream, + HloExecutionProfiler* profiler) override; // Returns true if we'll perform autotuning if run on the given stream. If // so, we want the GPU to be quiescent during autotuning, so as not to diff --git a/tensorflow/compiler/xla/service/gpu/gpu_compiler.cc b/tensorflow/compiler/xla/service/gpu/gpu_compiler.cc index decfc40dafafe875fa02bab6695f5c54e522f267..e1da8d940c9b5c178d691db7c000c2ce56223b52 100644 --- a/tensorflow/compiler/xla/service/gpu/gpu_compiler.cc +++ b/tensorflow/compiler/xla/service/gpu/gpu_compiler.cc @@ -552,8 +552,7 @@ StatusOr> GpuCompiler::RunBackend( &ir_emitter_context); { XLA_SCOPED_LOGGING_TIMER("GpuCompiler::RunBackend - IR emission"); - TF_RETURN_IF_ERROR( - entry_computation->root_instruction()->Accept(&ir_emitter)); + TF_RETURN_IF_ERROR(entry_computation->Accept(&ir_emitter)); } if (user_pre_optimization_hook_) { diff --git a/tensorflow/compiler/xla/service/gpu/gpu_executable.cc b/tensorflow/compiler/xla/service/gpu/gpu_executable.cc index f20a828bc1a31ad15298a1d77cd79599aa12faf4..0cad2958c72797b4d70f00676928b2b21d7a3e8d 100644 --- a/tensorflow/compiler/xla/service/gpu/gpu_executable.cc +++ b/tensorflow/compiler/xla/service/gpu/gpu_executable.cc @@ -136,18 +136,17 @@ Status GpuExecutable::ExecuteThunks( TF_RETURN_IF_ERROR(main_stream->BlockHostUntilDone()); } - profiler.StartOperation(); VLOG(2) << "Executing the thunk for " << thunk->hlo_instruction()->ToString() << " on stream " << stream_no; - TF_RETURN_IF_ERROR(thunk->ExecuteOnStream(buffer_allocations, stream)); + TF_RETURN_IF_ERROR( + thunk->ExecuteOnStream(buffer_allocations, stream, &profiler)); if (thunk_schedule_->Depended(thunk)) { auto finish_event = MakeUnique(main_stream->parent()); finish_event->Init(); stream->ThenRecordEvent(finish_event.get()); thunk_to_finish_event[thunk] = std::move(finish_event); } - profiler.FinishOperation(thunk->hlo_instruction()); } main_stream->ThenWaitFor(&sub_streams); diff --git a/tensorflow/compiler/xla/service/gpu/gpu_layout_assignment.cc b/tensorflow/compiler/xla/service/gpu/gpu_layout_assignment.cc index 8bf62dde8b9948375fc493fd1a524cfa7b062502..09ef62c87f8875a5803497e8eb628769f883202a 100644 --- a/tensorflow/compiler/xla/service/gpu/gpu_layout_assignment.cc +++ b/tensorflow/compiler/xla/service/gpu/gpu_layout_assignment.cc @@ -51,7 +51,7 @@ HeuristicLayoutAssignment(const HloInstruction* instr, // H <=> Y // W <=> X // - // Therefore kOutputInputYX means NHWC; kBatchDepthYX means NCHW. + // Therefore kOutputInputYX and kBatchDepthYX mean NCHW. // As of today, our empirical evidence is that cudnn 7.0 is faster on V100 x // fp16 with the mostly-NHWC layout. The heuristic may change as cudnn version diff --git a/tensorflow/compiler/xla/service/gpu/gpu_layout_assignment_test.cc b/tensorflow/compiler/xla/service/gpu/gpu_layout_assignment_test.cc index e48165c1426ea04839c245bc20b851a0f1710246..95f78ae29326caad2f0785e2ba285a996e685899 100644 --- a/tensorflow/compiler/xla/service/gpu/gpu_layout_assignment_test.cc +++ b/tensorflow/compiler/xla/service/gpu/gpu_layout_assignment_test.cc @@ -132,10 +132,10 @@ TEST_F(LayoutAssignmentTest, BatchNormInference) { HloInstruction::CreateParameter(4, aux_shape, "variance")); auto* epsilon = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1))); auto* feature_index = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR0(kFeatureIndex))); + LiteralUtil::CreateR0(kFeatureIndex))); auto* batchnorm = builder.AddInstruction(HloInstruction::CreateCustomCall( shape, @@ -201,10 +201,10 @@ TEST_F(LayoutAssignmentTest, BatchNormTraining) { HloInstruction::CreateParameter(2, offset_scale_shape, "offset")); auto* epsilon = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1))); auto* feature_index = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR0(kFeatureIndex))); + LiteralUtil::CreateR0(kFeatureIndex))); auto* batchnorm = builder.AddInstruction(HloInstruction::CreateCustomCall( batchnorm_shape, {operand, scale, offset, epsilon, feature_index}, @@ -278,10 +278,10 @@ TEST_F(LayoutAssignmentTest, BatchNormGrad) { HloInstruction::CreateParameter(4, shape, "var")); auto* epsilon = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1))); auto* feature_index = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR0(kFeatureIndex))); + LiteralUtil::CreateR0(kFeatureIndex))); auto* batchnorm = builder.AddInstruction(HloInstruction::CreateCustomCall( diff --git a/tensorflow/compiler/xla/service/gpu/gpu_transfer_manager.cc b/tensorflow/compiler/xla/service/gpu/gpu_transfer_manager.cc index 7bb8df6581b49b1bf8c84a972f715e8dc119d8de..1446401b1989f7e6a783d438f39e06900747178c 100644 --- a/tensorflow/compiler/xla/service/gpu/gpu_transfer_manager.cc +++ b/tensorflow/compiler/xla/service/gpu/gpu_transfer_manager.cc @@ -20,8 +20,10 @@ limitations under the License. #include #include "llvm/IR/DataLayout.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/literal_util.h" #include "tensorflow/compiler/xla/service/gpu/gpu_compiler.h" +#include "tensorflow/compiler/xla/service/gpu/outfeed_manager.h" #include "tensorflow/compiler/xla/shape_util.h" #include "tensorflow/compiler/xla/status_macros.h" #include "tensorflow/compiler/xla/statusor.h" @@ -34,6 +36,7 @@ limitations under the License. #include "tensorflow/core/platform/stream_executor_no_cuda.h" namespace xla { +namespace gpu { // TODO(b/30467474) Once GPU infeed implementation settles, consider // folding back the cpu and gpu infeed implementations into a generic @@ -50,53 +53,28 @@ Status GpuTransferManager::TransferLiteralToInfeed( VLOG(2) << "Transferring literal to infeed with shape: " << ShapeUtil::HumanString(shape); - if (!ShapeUtil::IsTuple(shape)) { - int64 size = GetByteSizeRequirement(shape); - return TransferBufferToInfeed(executor, size, literal.untyped_data()); - } - - if (ShapeUtil::IsNestedTuple(shape)) { - return Unimplemented( - "Infeed with a nested tuple shape is not supported: %s", - ShapeUtil::HumanString(literal.shape()).c_str()); - } - // For a tuple, we transfer each of its elements to the device and // enqueue the resulting destination device addresses with the // infeed manager. - std::vector buffers; - buffers.reserve(ShapeUtil::TupleElementCount(shape)); - auto cleanup = tensorflow::gtl::MakeCleanup([buffers]() { - for (gpu::InfeedBuffer* b : buffers) { - b->Done(); - } - }); - - for (int64 i = 0; i < ShapeUtil::TupleElementCount(shape); ++i) { - const Shape& tuple_element_shape = - ShapeUtil::GetTupleElementShape(shape, i); - int64 tuple_element_size = GetByteSizeRequirement(tuple_element_shape); - TF_ASSIGN_OR_RETURN( - gpu::InfeedBuffer * buffer, - TransferBufferToInfeedInternal(executor, tuple_element_size, - literal.untyped_data({i}))); - buffers.push_back(buffer); - } - - cleanup.release(); - return EnqueueBuffersToInfeed(executor, buffers); -} - -Status GpuTransferManager::TransferBufferToInfeed(se::StreamExecutor* executor, - int64 size, - const void* source) { - TF_ASSIGN_OR_RETURN(gpu::InfeedBuffer * buffer, - TransferBufferToInfeedInternal(executor, size, source)); - return EnqueueBuffersToInfeed(executor, {buffer}); + ShapeTree buffer_tree(shape); + + TF_RETURN_IF_ERROR(ShapeUtil::ForEachSubshapeWithStatus( + shape, [&](const Shape& literal_subshape, const ShapeIndex& index) { + if (ShapeUtil::IsArray(literal_subshape)) { + int64 tuple_element_size = GetByteSizeRequirement(literal_subshape); + TF_ASSIGN_OR_RETURN( + *buffer_tree.mutable_element(index), + TransferBufferToInfeedInternal(executor, tuple_element_size, + literal.untyped_data(index))); + } + return Status::OK(); + })); + + return EnqueueBuffersToInfeed(executor, std::move(buffer_tree)); } Status GpuTransferManager::EnqueueBuffersToInfeed( - se::StreamExecutor* executor, std::vector buffers) { + se::StreamExecutor* executor, ShapeTree buffers) { gpu::InfeedManager* infeed_manager = gpu::GetOrCreateInfeedManager(); se::Stream* stream = infeed_manager->GetStream(executor); @@ -106,21 +84,18 @@ Status GpuTransferManager::EnqueueBuffersToInfeed( // possible. Status block_status = stream->BlockHostUntilDone(); if (!block_status.ok()) { - for (gpu::InfeedBuffer* b : buffers) { - b->Done(); - } return InternalError("Failed to complete data transfer on stream %p: %s", stream, block_status.error_message().c_str()); } - infeed_manager->EnqueueBuffers(buffers); + infeed_manager->EnqueueDestination(std::move(buffers)); VLOG(2) << "Infeed data transferred"; return Status::OK(); } -StatusOr GpuTransferManager::TransferBufferToInfeedInternal( +StatusOr GpuTransferManager::TransferBufferToInfeedInternal( se::StreamExecutor* executor, int64 size, const void* source) { if (size > std::numeric_limits::max()) { return InvalidArgument("Infeed shape is too large: needs %lld bytes", size); @@ -136,18 +111,76 @@ StatusOr GpuTransferManager::TransferBufferToInfeedInternal( return InternalError("Failed to obtain a stream"); } - gpu::InfeedBuffer* buffer = new gpu::InfeedBuffer(executor, size); - stream->ThenMemcpy(buffer->device_memory(), source, size); + InfeedBuffer buffer(executor, size); + stream->ThenMemcpy(buffer.device_memory(), source, size); VLOG(2) << "Queued infeed data on stream " << stream; - return buffer; + return std::move(buffer); +} + +static std::unique_ptr ShapeTreeToLiteral( + ShapeTree>* shape_tree) { + // This is a struct instead of a lambda for std::function-free recursion. + struct Helper { + static std::unique_ptr helper( + ShapeTree>* shape_tree, + ShapeIndex* index) { + const Shape& shape = ShapeUtil::GetSubshape(shape_tree->shape(), *index); + if (ShapeUtil::IsArray(shape)) { + return (*shape_tree->mutable_element(*index))->WaitUntilAvailable(); + } + + CHECK(ShapeUtil::IsTuple(shape)) + << ShapeUtil::HumanStringWithLayout(shape); + const int64 tuple_element_count = ShapeUtil::TupleElementCount(shape); + index->push_back(0); + std::vector> tuple_operands; + for (int64 i = 0; i < tuple_element_count; ++i) { + index->back() = i; + tuple_operands.push_back(helper(shape_tree, index)); + } + index->pop_back(); + return LiteralUtil::MakeTupleOwned(std::move(tuple_operands)); + } + }; + ShapeIndex index; + return Helper::helper(shape_tree, &index); +} + +Status GpuTransferManager::TransferLiteralFromOutfeed( + se::StreamExecutor* /*executor*/, const Shape& literal_shape, + Literal* literal) { + ShapeTree> outfeed_buffers( + &literal_shape); + + // First create a tree of literal buffers that the device can write to. + outfeed_buffers.ForEachMutableElement( + [&](const ShapeIndex& index, + std::unique_ptr* buffer) { + const Shape& shape = ShapeUtil::GetSubshape(literal_shape, index); + // Do not transfer tuple index buffers. + if (ShapeUtil::IsTuple(shape)) { + return; + } + *buffer = MakeUnique(GetByteSizeRequirement(shape)); + }); + + // Give the tree of buffers to the outfeed mananger. The device will fill it + // while we're waiting for it below. + gpu::OutfeedManager* outfeed_manager = gpu::GetOrCreateOutfeedManager(); + outfeed_manager->EnqueueDestination(&outfeed_buffers); + + // Now turn the tree of buffers back into a literal. + *literal = std::move(*ShapeTreeToLiteral(&outfeed_buffers)); + return Status::OK(); } +} // namespace gpu } // namespace xla static std::unique_ptr CreateGpuTransferManager() { - return xla::MakeUnique(); + return xla::MakeUnique(); } static bool InitModule() { diff --git a/tensorflow/compiler/xla/service/gpu/gpu_transfer_manager.h b/tensorflow/compiler/xla/service/gpu/gpu_transfer_manager.h index 09f8227f508a3159f3def285898e15bfad544552..8122c9d8c38e90111a3ca2e7e3919359362d299d 100644 --- a/tensorflow/compiler/xla/service/gpu/gpu_transfer_manager.h +++ b/tensorflow/compiler/xla/service/gpu/gpu_transfer_manager.h @@ -21,6 +21,7 @@ limitations under the License. #include "tensorflow/compiler/xla/service/generic_transfer_manager.h" #include "tensorflow/compiler/xla/service/gpu/infeed_manager.h" #include "tensorflow/compiler/xla/service/transfer_manager.h" +#include "tensorflow/compiler/xla/shape_tree.h" #include "tensorflow/compiler/xla/statusor.h" #include "tensorflow/compiler/xla/xla_data.pb.h" #include "tensorflow/core/platform/macros.h" @@ -28,6 +29,7 @@ limitations under the License. #include "tensorflow/core/platform/types.h" namespace xla { +namespace gpu { // An implementation of the XLA GenericTransferManager that // handles GPU-specific infeed. @@ -38,23 +40,25 @@ class GpuTransferManager : public GenericTransferManager { Status TransferLiteralToInfeed(se::StreamExecutor* executor, const LiteralSlice& literal) override; - Status TransferBufferToInfeed(se::StreamExecutor* executor, int64 size, - const void* source) override; + Status TransferLiteralFromOutfeed(se::StreamExecutor* executor, + const Shape& literal_shape, + Literal* literal) override; private: // Initiates the infeed data transfers. InfeedBuffer->Done() must be // called to clean up the memory allocated for InfeedBuffer. - StatusOr TransferBufferToInfeedInternal( + StatusOr TransferBufferToInfeedInternal( se::StreamExecutor* executor, int64 size, const void* source); // Enqueues infeed data buffers with the infeed manager after their // transfer completes. Status EnqueueBuffersToInfeed(se::StreamExecutor* executor, - std::vector buffers); + ShapeTree buffers); TF_DISALLOW_COPY_AND_ASSIGN(GpuTransferManager); }; +} // namespace gpu } // namespace xla #endif // TENSORFLOW_COMPILER_XLA_SERVICE_GPU_TRANSFER_MANAGER_H_ diff --git a/tensorflow/compiler/xla/service/gpu/hlo_execution_profiler.cc b/tensorflow/compiler/xla/service/gpu/hlo_execution_profiler.cc index daddd3738e4bb54f3695a96f6f9ffb9accabe97c..19420e590d05892417da4d5e62fdcde5eba9d9f1 100644 --- a/tensorflow/compiler/xla/service/gpu/hlo_execution_profiler.cc +++ b/tensorflow/compiler/xla/service/gpu/hlo_execution_profiler.cc @@ -16,6 +16,8 @@ limitations under the License. #include "tensorflow/compiler/xla/service/gpu/hlo_execution_profiler.h" #include +#include +#include #include #include "tensorflow/compiler/xla/service/hlo_computation.h" @@ -24,9 +26,30 @@ limitations under the License. #include "tensorflow/compiler/xla/service/pool.h" #include "tensorflow/core/platform/logging.h" #include "tensorflow/core/platform/stream_executor_no_cuda.h" +#include "tensorflow/core/util/ptr_util.h" namespace xla { namespace gpu { +namespace { +void InitAndStartTimer(std::stack>* timers, + se::Stream* stream) { + timers->push(MakeUnique(stream->parent())); + stream->InitTimer(timers->top().get()).ThenStartTimer(timers->top().get()); +} + +uint64 GetCyclesTaken( + std::stack>* timers, + const std::vector::SmartPtr>& sub_streams, + se::Stream* stream, double clock_rate_ghz) { + CHECK_GT(timers->size(), 0); + stream->ThenWaitFor(&sub_streams); + stream->ThenStopTimer(timers->top().get()); + stream->BlockHostUntilDone().IgnoreError(); + double nanoseconds = timers->top()->Nanoseconds(); + timers->pop(); + return static_cast(nanoseconds * clock_rate_ghz); +} +} // namespace HloExecutionProfiler::HloExecutionProfiler( bool do_profile, HloExecutionProfile* profile, se::Stream* stream, @@ -39,11 +62,7 @@ HloExecutionProfiler::HloExecutionProfiler( computation_(computation) { if (do_profile_) { clock_rate_ghz_ = stream->parent()->GetDeviceDescription().clock_rate_ghz(); - execution_timer_.reset(new se::Timer(stream->parent())); - per_op_timer_.reset(new se::Timer(stream->parent())); - stream->InitTimer(execution_timer_.get()) - .ThenStartTimer(execution_timer_.get()); - stream->InitTimer(per_op_timer_.get()); + InitAndStartTimer(&timers_, stream); } } @@ -51,31 +70,53 @@ void HloExecutionProfiler::FinishExecution() { CHECK(!finished_execution_) << "Call FinishExecution only once!"; finished_execution_ = true; if (do_profile_) { - stream_->ThenWaitFor(&sub_streams_); - stream_->ThenStopTimer(execution_timer_.get()); - stream_->BlockHostUntilDone().IgnoreError(); profile_->set_total_cycles_executed( *computation_, - static_cast(execution_timer_->Nanoseconds() * clock_rate_ghz_)); + GetCyclesTaken(&timers_, sub_streams_, stream_, clock_rate_ghz_)); } } -void HloExecutionProfiler::StartOperation() { +void HloExecutionProfiler::StartHloComputation() { if (do_profile_) { - stream_->ThenStartTimer(per_op_timer_.get()); + InitAndStartTimer(&timers_, stream_); + } +} + +void HloExecutionProfiler::FinishHloComputation( + const HloComputation* computation) { + if (do_profile_) { + profile_->set_total_cycles_executed( + *computation, + GetCyclesTaken(&timers_, sub_streams_, stream_, clock_rate_ghz_)); } } -void HloExecutionProfiler::FinishOperation( +void HloExecutionProfiler::StartHloInstruction() { + if (do_profile_) { + InitAndStartTimer(&timers_, stream_); + } +} + +void HloExecutionProfiler::FinishHloInstruction( const HloInstruction* hlo_instruction) { if (do_profile_) { - stream_->ThenWaitFor(&sub_streams_); - stream_->ThenStopTimer(per_op_timer_.get()); - stream_->BlockHostUntilDone().IgnoreError(); + hlo_instructions_.erase(hlo_instruction); profile_->SetCyclesTakenBy( hlo_instruction, - static_cast(per_op_timer_->Nanoseconds() * clock_rate_ghz_)); + GetCyclesTaken(&timers_, sub_streams_, stream_, clock_rate_ghz_)); + } +} + +std::unique_ptr +HloExecutionProfiler::MakeScopedInstructionProfiler( + const HloInstruction* hlo_instruction) { + if (do_profile_ && hlo_instruction != nullptr) { + // Make sure that we are not already measuring the time for the same + // 'hlo_instruction'. + CHECK(hlo_instructions_.insert(hlo_instruction).second) + << hlo_instruction->name(); } + return MakeUnique(this, hlo_instruction); } } // namespace gpu diff --git a/tensorflow/compiler/xla/service/gpu/hlo_execution_profiler.h b/tensorflow/compiler/xla/service/gpu/hlo_execution_profiler.h index c9b882ff805c45a57f15df4fe79dc34100c0ceff..6654850bef3efa46028defbba81e3537fafbf143 100644 --- a/tensorflow/compiler/xla/service/gpu/hlo_execution_profiler.h +++ b/tensorflow/compiler/xla/service/gpu/hlo_execution_profiler.h @@ -17,6 +17,8 @@ limitations under the License. #define TENSORFLOW_COMPILER_XLA_SERVICE_GPU_HLO_EXECUTION_PROFILER_H_ #include +#include +#include #include #include "tensorflow/compiler/xla/service/hlo_computation.h" @@ -28,6 +30,8 @@ limitations under the License. namespace xla { namespace gpu { +class ScopedInstructionProfiler; + // A helper class for profiling HLO in the course of GPU program execution. // All of the profiling is guarded internally, to avoid the caller needing to // have lots of conditionals sprinkled around. @@ -43,12 +47,25 @@ class HloExecutionProfiler { // execution timer. void FinishExecution(); - // If profiling is enabled, starts the per-operation timer. - void StartOperation(); + // If profiling is enabled, starts a timer for a (sub)computation. + void StartHloComputation(); + + // If profiling is enabled stops the timer for a (sub)computation and records + // the time that the computation took to execute in the profile. + void FinishHloComputation(const HloComputation* computation); + + // If profiling is enabled, starts a per-operation timer. + void StartHloInstruction(); // If profiling is enabled, stops the per-operation timer and records the time // that the hlo_instruction took to execute in the profile. - void FinishOperation(const HloInstruction* hlo_instruction); + void FinishHloInstruction(const HloInstruction* hlo_instruction); + + // Returns a ScopedInstructionProfiler and triggers a call to + // StartHloInstruction(). Once the returned ScopedInstructionProfiler goes + // out of scope, it triggers a call to FinishHloInstruction(). + std::unique_ptr MakeScopedInstructionProfiler( + const HloInstruction* hlo_instruction); private: const bool do_profile_; @@ -57,11 +74,36 @@ class HloExecutionProfiler { se::Stream* stream_; const std::vector::SmartPtr>& sub_streams_; const HloComputation* computation_; - std::unique_ptr execution_timer_; - std::unique_ptr per_op_timer_; + std::stack> timers_; + // Contains the HLO instructions for which we are currently measuring the + // time. + std::unordered_set hlo_instructions_; bool finished_execution_ = false; }; +// This class can be used within the ExecuteOnStream() implementations of +// Thunks. It ensures that we always have a pair of matching +// StartHloInstruction() and FinishHloInstruction() calls to the profiler. +class ScopedInstructionProfiler { + public: + ScopedInstructionProfiler(HloExecutionProfiler* profiler, + const HloInstruction* hlo_instruction) + : profiler_(profiler), hlo_instruction_(hlo_instruction) { + if (hlo_instruction != nullptr) { + profiler->StartHloInstruction(); + } + } + ~ScopedInstructionProfiler() { + if (hlo_instruction_ != nullptr) { + profiler_->FinishHloInstruction(hlo_instruction_); + } + } + + private: + HloExecutionProfiler* profiler_; + const HloInstruction* hlo_instruction_; +}; + } // namespace gpu } // namespace xla diff --git a/tensorflow/compiler/xla/service/gpu/hlo_schedule.cc b/tensorflow/compiler/xla/service/gpu/hlo_schedule.cc index 375709150e08996ea6a40f5e9e66a8f8d9287008..19de37b0fbed15455e8c6a9bfe427ba3d9f0a9dc 100644 --- a/tensorflow/compiler/xla/service/gpu/hlo_schedule.cc +++ b/tensorflow/compiler/xla/service/gpu/hlo_schedule.cc @@ -100,7 +100,7 @@ GpuHloOrdering::GpuHloOrdering( if (last_instruction_per_stream[stream_no] != nullptr) { immediate_preds.push_back(last_instruction_per_stream[stream_no]); } - predecessor_map->SetReachabilityToUnion(immediate_preds, hlo); + predecessor_map->FastSetReachabilityToUnion(immediate_preds, hlo); last_instruction_per_stream[stream_no] = hlo; } else { // Only parameters and constants don't have an assigned stream, since they diff --git a/tensorflow/compiler/xla/service/gpu/infeed_manager.cc b/tensorflow/compiler/xla/service/gpu/infeed_manager.cc index ae310beefad0c81c17fd4140b441b3a19a002e2c..c5f0cdf6cd5d3e076bffa875fbba991bf0681ee8 100644 --- a/tensorflow/compiler/xla/service/gpu/infeed_manager.cc +++ b/tensorflow/compiler/xla/service/gpu/infeed_manager.cc @@ -15,76 +15,13 @@ limitations under the License. #include "tensorflow/compiler/xla/service/gpu/infeed_manager.h" -#include "tensorflow/compiler/xla/map_util.h" #include "tensorflow/compiler/xla/ptr_util.h" -#include "tensorflow/core/platform/logging.h" namespace xla { namespace gpu { -InfeedManager::InfeedManager() : host_to_device_executor_(nullptr) {} - -void InfeedManager::Reset() { - tensorflow::mutex_lock l(mu_); - CHECK(dequeued_buffer_.empty()); - for (auto buffer : enqueued_buffer_) { - buffer->Done(); - } - enqueued_buffer_.clear(); -} - -void InfeedManager::EnqueueBuffers(const std::vector& buffers) { - tensorflow::mutex_lock l(mu_); - bool was_empty = enqueued_buffer_.empty(); - for (gpu::InfeedBuffer* b : buffers) { - enqueued_buffer_.push_back(b); - } - if (was_empty) { - // This has the potential to suffer from the notified thread - // immediately trying and failing to acquire mu_, but seems - // preferable to the alternative of notifying outside the lock - // on every enqueue. - cv_.notify_one(); - } -} - -InfeedBuffer* InfeedManager::BlockingDequeueBuffer() { - bool became_empty = false; - InfeedBuffer* current_buffer; - { - tensorflow::mutex_lock l(mu_); - while (enqueued_buffer_.empty()) { - cv_.wait(l); - } - current_buffer = enqueued_buffer_.front(); - enqueued_buffer_.pop_front(); - dequeued_buffer_.insert(current_buffer); - if (enqueued_buffer_.empty()) { - became_empty = true; - } - } - if (became_empty) { - for (const auto& callback : on_empty_callbacks_) { - callback(); - } - } - return current_buffer; -} - -void InfeedManager::ReleaseBuffers(const std::vector& buffers) { - { - tensorflow::mutex_lock l(mu_); - for (gpu::InfeedBuffer* b : buffers) { - CHECK(ContainsKey(dequeued_buffer_, b)); - dequeued_buffer_.erase(b); - } - } - for (gpu::InfeedBuffer* b : buffers) { - b->Done(); - } -} - se::Stream* InfeedManager::GetStream(se::StreamExecutor* executor) { + tensorflow::mutex_lock l(host_to_device_stream_mu_); if (host_to_device_executor_ == nullptr) { host_to_device_executor_ = executor; host_to_device_stream_ = MakeUnique(executor); @@ -100,10 +37,6 @@ se::Stream* InfeedManager::GetStream(se::StreamExecutor* executor) { return host_to_device_stream_.get(); } -void InfeedManager::RegisterOnEmptyCallback(std::function callback) { - on_empty_callbacks_.push_back(std::move(callback)); -} - InfeedManager* GetOrCreateInfeedManager() { static InfeedManager* manager = new InfeedManager; return manager; diff --git a/tensorflow/compiler/xla/service/gpu/infeed_manager.h b/tensorflow/compiler/xla/service/gpu/infeed_manager.h index a3fc15cfe36a490f38daabca9ff36fbb1012aead..7e418882e051a77e10bd12000bbc9769980f5f14 100644 --- a/tensorflow/compiler/xla/service/gpu/infeed_manager.h +++ b/tensorflow/compiler/xla/service/gpu/infeed_manager.h @@ -20,12 +20,9 @@ limitations under the License. #ifndef TENSORFLOW_COMPILER_XLA_SERVICE_GPU_INFEED_MANAGER_H_ #define TENSORFLOW_COMPILER_XLA_SERVICE_GPU_INFEED_MANAGER_H_ -#include -#include - +#include "tensorflow/compiler/xla/service/gpu/xfeed_queue.h" +#include "tensorflow/compiler/xla/shape_tree.h" #include "tensorflow/compiler/xla/types.h" -#include "tensorflow/core/lib/gtl/flatset.h" -#include "tensorflow/core/platform/mutex.h" #include "tensorflow/core/platform/stream_executor_no_cuda.h" namespace xla { @@ -47,90 +44,41 @@ namespace gpu { // the client. The client manages the memory of the buffer. class InfeedBuffer { public: + InfeedBuffer() = default; InfeedBuffer(se::StreamExecutor* executor, int64 length) - : executor_(executor), length_(length) { - device_memory_ = executor_->AllocateArray(length); - CHECK(!device_memory_.is_null()); + : device_memory_(executor, executor->AllocateArray(length)), + length_(length) { + CHECK(!device_memory_->is_null()); } - ~InfeedBuffer() { executor_->Deallocate(&device_memory_); } - int64 length() const { return length_; } - // Callback to signal that this buffer is consumed. This helps the - // client to manage memory for the infeed buffers. - void Done() { delete this; } - - se::DeviceMemoryBase* device_memory() { return &device_memory_; } + se::DeviceMemoryBase* device_memory() { return device_memory_.ptr(); } private: - se::StreamExecutor* executor_; // Not owned. - const int64 length_; - se::DeviceMemoryBase device_memory_; + se::ScopedDeviceMemory device_memory_; + int64 length_; }; // Client-side class used to enqueue infeed buffers. -class InfeedManager { +class InfeedManager : public XfeedQueue> { public: - InfeedManager(); - - // Calls the completion callback for any enqueued buffers that have - // not been dequeued by the runtime, and empties the infeed - // queue. Reset may not be called while a runtime computation is - // processing a dequeued buffer. The only safe way to ensure this - // condition is to call Reset when no computation is taking place. - void Reset(); - - // Adds a set of buffers to the infeed queue atomically. buffer->Done - // will be called when the buffer will no longer be accessed by the - // InfeedManager, either as a result of a call to Reset or because the - // runtime has dequeued and used the buffer. - void EnqueueBuffers(const std::vector& buffers); - - // Blocks until the infeed queue is non-empty, then returns the - // buffer at the head of the queue. Adds the current buffer to the - // to-be released set. - InfeedBuffer* BlockingDequeueBuffer(); - - // Releases a set of buffers from the to-be released set. - void ReleaseBuffers(const std::vector& buffers); - // Returns a cached stream associated with an executor. Allocates a // new stream on the first invocation. On subsequent invocations, if // the cached executor is not the same as the requested executor, // returns null. se::Stream* GetStream(se::StreamExecutor* executor); - // Registers a callback that will be called when 'enqueued_buffer_' becomes - // empty. - void RegisterOnEmptyCallback(std::function callback); - private: - // TODO(b/30467474): Revisit if this mutex becomes a point of - // contention. - tensorflow::mutex mu_; - - // Condition variable that is signaled every time a buffer is - // enqueued to an empty queue. - tensorflow::condition_variable cv_; - - // InfeedBuffer* queue contents are not owned, but buffer->Done must - // be called when the buffer is no longer needed by the runtime. - std::deque enqueued_buffer_; - - // Buffers that are dequeued and currently being processed by the - // runtime. Not owned. - tensorflow::gtl::FlatSet dequeued_buffer_; + // Mutex for serializing the creation of host_to_device_stream_. + tensorflow::mutex host_to_device_stream_mu_; // Cached host to device stream for queuing infeed data. - std::unique_ptr host_to_device_stream_; + std::unique_ptr host_to_device_stream_ + GUARDED_BY(host_to_device_stream_mu_); // Executor that the host_to_device_stream belongs to. Not owned. - se::StreamExecutor* host_to_device_executor_; - - // List of callbacks which will be called when 'enqueued_buffer_' becomes - // empty. - std::vector> on_empty_callbacks_; + se::StreamExecutor* host_to_device_executor_ = nullptr; }; // Singleton creator-or-accessor: Returns the GPU infeed manager. diff --git a/tensorflow/compiler/xla/service/gpu/infeed_thunk.cc b/tensorflow/compiler/xla/service/gpu/infeed_thunk.cc index ea34d5b30c91e8b809e3e17a904e27e589fd6b5f..fee6d2af3bfd4976f5845edf592e8310b55a3feb 100644 --- a/tensorflow/compiler/xla/service/gpu/infeed_thunk.cc +++ b/tensorflow/compiler/xla/service/gpu/infeed_thunk.cc @@ -13,8 +13,9 @@ See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ -#include "tensorflow/compiler/xla/service/gpu/infeed_manager.h" #include "tensorflow/compiler/xla/service/gpu/infeed_thunk.h" +#include "tensorflow/compiler/xla/service/gpu/hlo_execution_profiler.h" +#include "tensorflow/compiler/xla/service/gpu/infeed_manager.h" #include "tensorflow/compiler/xla/util.h" #include "tensorflow/core/platform/stream_executor_no_cuda.h" @@ -22,57 +23,82 @@ namespace xla { namespace gpu { InfeedThunk::InfeedThunk( - tensorflow::gtl::ArraySlice tuple_element_buffers, - const BufferAllocation::Slice& destination_buffer, + const ShapeTree& infeed_slices, const HloInstruction* hlo_instruction) - : Thunk(Kind::kInfeed, hlo_instruction), - tuple_element_buffers_(tuple_element_buffers.begin(), - tuple_element_buffers.end()), - destination_buffer_(destination_buffer) {} + : Thunk(Kind::kInfeed, hlo_instruction), infeed_slices_(infeed_slices) {} Status InfeedThunk::ExecuteOnStream(const BufferAllocations& buffer_allocations, - se::Stream* stream) { - VLOG(2) << "Infeeding to GPU "; + se::Stream* stream, + HloExecutionProfiler* profiler) { + VLOG(2) << "Infeeding to GPU: " << hlo_instruction()->ToString(); + + auto op_profiler = profiler->MakeScopedInstructionProfiler(hlo_instruction()); + ShapeTree infeed_buffers = + GetOrCreateInfeedManager()->BlockingGetNextDestination(); + + { + // The infeed buffer has an extra outer tuple with a token. Adjust the index + // accordingly. + ShapeIndex index = {0}; + std::function*)> copy_tuple_contents = + [&](std::vector* tuple_element_addresses) { + const Shape& shape = ShapeUtil::GetSubshape(infeed_buffers.shape(), + ShapeIndexView(index, 1)); + // For the leaf buffers of the tuple copy the elements directly. + if (ShapeUtil::IsArray(shape)) { + const BufferAllocation::Slice& tuple_element_buffer = + infeed_slices_.element(index); + se::DeviceMemoryBase tuple_element_address = + buffer_allocations.GetDeviceAddress(tuple_element_buffer); - se::DeviceMemoryBase destination_address = - buffer_allocations.GetDeviceAddress(destination_buffer_); + InfeedBuffer* buffer = + infeed_buffers.mutable_element(ShapeIndexView(index, 1)); + stream->ThenMemcpy(&tuple_element_address, + *(buffer->device_memory()), buffer->length()); + tuple_element_addresses->push_back(tuple_element_address.opaque()); + return; + } + + const int64 tuple_element_count = ShapeUtil::TupleElementCount(shape); + index.push_back(0); + std::vector inner_tuple_element_addresses; + for (int64 i = 0; i < tuple_element_count; ++i) { + index.back() = i; + copy_tuple_contents(&inner_tuple_element_addresses); + } + index.pop_back(); + + // Create a buffer of pointers for non-leaf buffers. + CHECK_EQ(tuple_element_count, inner_tuple_element_addresses.size()); + auto host_size = inner_tuple_element_addresses.size() * sizeof(void*); + se::DeviceMemoryBase tuple_address = + buffer_allocations.GetDeviceAddress( + infeed_slices_.element(index)); + stream->ThenMemcpy(&tuple_address, + inner_tuple_element_addresses.data(), host_size); + tuple_element_addresses->push_back(tuple_address.opaque()); + }; - InfeedManager* infeed_manager = GetOrCreateInfeedManager(); - std::vector infeed_buffers; - if (ShapeUtil::IsTuple(hlo_instruction()->shape())) { - CHECK(!ShapeUtil::IsNestedTuple(hlo_instruction()->shape())); - // Transfer the tuple elements first. std::vector tuple_element_addresses; - for (BufferAllocation::Slice tuple_element_buffer : - tuple_element_buffers_) { - se::DeviceMemoryBase tuple_element_address = - buffer_allocations.GetDeviceAddress(tuple_element_buffer); - - InfeedBuffer* buffer = infeed_manager->BlockingDequeueBuffer(); - infeed_buffers.push_back(buffer); - stream->ThenMemcpy(&tuple_element_address, *(buffer->device_memory()), - buffer->length()); - tuple_element_addresses.push_back(tuple_element_address.opaque()); - } - // Transfer the tuple outer buffer. - auto host_size = tuple_element_addresses.size() * sizeof(void*); - stream->ThenMemcpy(&destination_address, tuple_element_addresses.data(), - host_size); - } else { - InfeedBuffer* buffer = infeed_manager->BlockingDequeueBuffer(); - infeed_buffers.push_back(buffer); - stream->ThenMemcpy(&destination_address, *(buffer->device_memory()), - buffer->length()); + copy_tuple_contents(&tuple_element_addresses); + CHECK_EQ(1, tuple_element_addresses.size()); } + // Construct top-level tuple of infeed containing the data and the token. Use + // a nullptr for the token, it should never be dereferenced. + se::DeviceMemoryBase data_address = + buffer_allocations.GetDeviceAddress(infeed_slices_.element({0})); + void* infeed_addresses[] = {data_address.opaque(), nullptr}; + se::DeviceMemoryBase top_level_address = + buffer_allocations.GetDeviceAddress(infeed_slices_.element({})); + stream->ThenMemcpy(&top_level_address, infeed_addresses, 2 * sizeof(void*)); + Status block_status = stream->BlockHostUntilDone(); if (!block_status.ok()) { return InternalError("Failed to complete data transfer on stream %p: %s", stream, block_status.error_message().c_str()); } - infeed_manager->ReleaseBuffers(infeed_buffers); - VLOG(2) << "Infeeding to GPU complete"; return Status::OK(); } diff --git a/tensorflow/compiler/xla/service/gpu/infeed_thunk.h b/tensorflow/compiler/xla/service/gpu/infeed_thunk.h index 93713cb12defd95bdd69cb0aa7ad7b4e37fc8fae..59487e245b78e66c45409fe712e86d3392e50580 100644 --- a/tensorflow/compiler/xla/service/gpu/infeed_thunk.h +++ b/tensorflow/compiler/xla/service/gpu/infeed_thunk.h @@ -18,6 +18,7 @@ limitations under the License. #include "tensorflow/compiler/xla/service/buffer_assignment.h" #include "tensorflow/compiler/xla/service/gpu/buffer_allocations.h" +#include "tensorflow/compiler/xla/service/gpu/hlo_execution_profiler.h" #include "tensorflow/compiler/xla/service/gpu/thunk.h" #include "tensorflow/compiler/xla/service/hlo_instruction.h" #include "tensorflow/core/platform/stream_executor_no_cuda.h" @@ -32,23 +33,19 @@ namespace gpu { class InfeedThunk : public Thunk { public: // Constructs a InfeedThunk that copies data from the on-device - // infeed queue to the device buffer - // `destination_buffer`. `mem_size` is the size of the data in - // bytes. - InfeedThunk(tensorflow::gtl::ArraySlice - tuple_element_buffers, - const BufferAllocation::Slice& destination_buffer, + // infeed queue into the buffers in the given shape tree. + InfeedThunk(const ShapeTree& infeed_slices, const HloInstruction* hlo_instruction); InfeedThunk(const InfeedThunk&) = delete; InfeedThunk& operator=(const InfeedThunk&) = delete; Status ExecuteOnStream(const BufferAllocations& buffer_allocations, - se::Stream* stream) override; + se::Stream* stream, + HloExecutionProfiler* profiler) override; private: - const std::vector tuple_element_buffers_; - const BufferAllocation::Slice destination_buffer_; + const ShapeTree infeed_slices_; }; } // namespace gpu diff --git a/tensorflow/compiler/xla/service/gpu/instruction_fusion_test.cc b/tensorflow/compiler/xla/service/gpu/instruction_fusion_test.cc index 1963d9eef72d41fa0a275bea98f959671fa7e737..98ba162cd97b8e214d7f055ee9dd590d7c67e1dd 100644 --- a/tensorflow/compiler/xla/service/gpu/instruction_fusion_test.cc +++ b/tensorflow/compiler/xla/service/gpu/instruction_fusion_test.cc @@ -33,7 +33,7 @@ TEST_F(InstructionFusionTest, CostlyProducerAndOperandElementReusingConsumerNotFused) { HloComputation::Builder builder(TestName()); HloInstruction* const0 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(5))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(5))); HloInstruction* exp1 = builder.AddInstruction(HloInstruction::CreateUnary( ShapeUtil::MakeShape(S32, {}), HloOpcode::kExp, const0)); HloInstruction* broadcast2 = @@ -53,7 +53,7 @@ TEST_F(InstructionFusionTest, NonCostlyProducerAndOperandElementReusingConsumerFused) { HloComputation::Builder builder(TestName()); HloInstruction* const0 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(5))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(5))); HloInstruction* negate1 = builder.AddInstruction(HloInstruction::CreateUnary( ShapeUtil::MakeShape(S32, {}), HloOpcode::kNegate, const0)); HloInstruction* broadcast2 = @@ -73,7 +73,7 @@ TEST_F(InstructionFusionTest, CostlyProducerAndNonOperandElementReusingConsumerFused_Reshape) { HloComputation::Builder builder(TestName()); HloInstruction* const0 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(5))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(5))); HloInstruction* exp1 = builder.AddInstruction(HloInstruction::CreateUnary( ShapeUtil::MakeShape(S32, {}), HloOpcode::kExp, const0)); HloInstruction* reshape2 = builder.AddInstruction( @@ -92,7 +92,7 @@ TEST_F(InstructionFusionTest, CostlyProducerAndNonOperandElementReusingConsumerFused_Transpose) { HloComputation::Builder builder(TestName()); HloInstruction* const0 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(5))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(5))); HloInstruction* exp1 = builder.AddInstruction(HloInstruction::CreateUnary( ShapeUtil::MakeShape(S32, {}), HloOpcode::kExp, const0)); HloInstruction* transpose2 = builder.AddInstruction( diff --git a/tensorflow/compiler/xla/service/gpu/ir_emitter.cc b/tensorflow/compiler/xla/service/gpu/ir_emitter.cc index d38a496fea689675f780ab5f377f4668bc9f05ca..fe83d017f4cde36cac37400ed16faab225878ea7 100644 --- a/tensorflow/compiler/xla/service/gpu/ir_emitter.cc +++ b/tensorflow/compiler/xla/service/gpu/ir_emitter.cc @@ -94,10 +94,7 @@ Status IrEmitter::HandleConstant(HloInstruction* constant) { << std::endl << " its type: " << llvm_ir::DumpToString(*global_for_const->getType()); - llvm::Constant* shape_constant = llvm::ConstantExpr::getBitCast( - global_for_const, - llvm_ir::ShapeToIrType(literal.shape(), module_)->getPointerTo()); - bindings_.BindHloToIrValue(*constant, shape_constant); + bindings_.BindHloToIrValue(*constant, global_for_const); return Status::OK(); } @@ -194,6 +191,8 @@ bool IrEmitter::MaybeEmitDirectAtomicOperation( HloOpcode root_opcode = computation.root_instruction()->opcode(); PrimitiveType element_type = computation.root_instruction()->shape().element_type(); + bool is_atomic_integral = element_type == S32 || element_type == U32 || + element_type == S64 || element_type == U64; llvm::Value* source = ir_builder_.CreateLoad(source_address, "source"); if (root_opcode == HloOpcode::kAdd) { // NVPTX supports atomicAdd on F32 and integer types. @@ -204,7 +203,7 @@ bool IrEmitter::MaybeEmitDirectAtomicOperation( {output_address->getType()}, &ir_builder_); return true; } - if (primitive_util::IsIntegralType(element_type)) { + if (is_atomic_integral) { // integral + integral ir_builder_.CreateAtomicRMW(llvm::AtomicRMWInst::Add, output_address, source, @@ -213,9 +212,8 @@ bool IrEmitter::MaybeEmitDirectAtomicOperation( } } - // NVPTX supports atomicMax and atomicMin on only integer types. - if (root_opcode == HloOpcode::kMaximum && - primitive_util::IsIntegralType(element_type)) { + // NVPTX supports atomicMax and atomicMin only on integer types. + if (root_opcode == HloOpcode::kMaximum && is_atomic_integral) { // max(integral, integral) auto opcode = primitive_util::IsSignedIntegralType(element_type) ? llvm::AtomicRMWInst::Max @@ -225,8 +223,7 @@ bool IrEmitter::MaybeEmitDirectAtomicOperation( return true; } - if (root_opcode == HloOpcode::kMinimum && - primitive_util::IsIntegralType(element_type)) { + if (root_opcode == HloOpcode::kMinimum && is_atomic_integral) { // min(integral, integral) auto opcode = primitive_util::IsSignedIntegralType(element_type) ? llvm::AtomicRMWInst::Min @@ -424,24 +421,27 @@ Status IrEmitter::EmitAtomicOperationForNestedComputation( Status IrEmitter::HandleSelect(HloInstruction* select) { auto pred = select->operand(0); - auto on_true = select->operand(1); - auto on_false = select->operand(2); TF_RET_CHECK(pred->shape().element_type() == PRED); - - if (ShapeUtil::IsTuple(select->shape())) { - llvm_ir::EmitTupleSelect(GetIrArray(*select, *select), - GetIrArray(*pred, *select), - GetBasePointer(*on_true), - GetBasePointer(*on_false), &ir_builder_, module_); - return Status::OK(); - } - // We must not call the subclass `DefaultAction` method, lest its // `HandleSelect` call `IrEmitter::HandleSelect` and its `DefaultAction` // assume no handler has already been called. return IrEmitter::DefaultAction(select); } +Status IrEmitter::HandleTupleSelect(HloInstruction* tuple_select) { + auto pred = tuple_select->operand(0); + auto on_true = tuple_select->operand(1); + auto on_false = tuple_select->operand(2); + TF_RET_CHECK(pred->shape().element_type() == PRED); + TF_RET_CHECK(ShapeUtil::IsScalar(pred->shape())); + TF_RET_CHECK(ShapeUtil::IsTuple(tuple_select->shape())); + llvm_ir::EmitTupleSelect(GetIrArray(*tuple_select, *tuple_select), + GetIrArray(*pred, *tuple_select), + GetBasePointer(*on_true), GetBasePointer(*on_false), + &ir_builder_, module_); + return Status::OK(); +} + namespace { llvm::Value* Real(llvm::Value* x, llvm::IRBuilder<>* ir_builder) { return ir_builder->CreateExtractValue(x, {0}); diff --git a/tensorflow/compiler/xla/service/gpu/ir_emitter.h b/tensorflow/compiler/xla/service/gpu/ir_emitter.h index e55dfc6dae844ceb1d28ad389d133c80823bad9a..d2dd335f10cc8346c5f941e5c8c6b5c403722fa3 100644 --- a/tensorflow/compiler/xla/service/gpu/ir_emitter.h +++ b/tensorflow/compiler/xla/service/gpu/ir_emitter.h @@ -88,6 +88,7 @@ class IrEmitter : public DfsHloVisitorWithDefault { Status HandleReduce(HloInstruction* reduce) override; Status HandleTuple(HloInstruction* tuple) override; Status HandleSelect(HloInstruction* select) override; + Status HandleTupleSelect(HloInstruction* tuple_select) override; Status HandleFusion(HloInstruction* fusion) override; Status HandleCall(HloInstruction* call) override; Status HandleCustomCall(HloInstruction* custom_call) override; diff --git a/tensorflow/compiler/xla/service/gpu/ir_emitter_unnested.cc b/tensorflow/compiler/xla/service/gpu/ir_emitter_unnested.cc index f6f0a45124b9978ba21b306d0d98caaf52e8bcc0..673ba530df0b9a71f7e284bef80171e34a1a7ad8 100644 --- a/tensorflow/compiler/xla/service/gpu/ir_emitter_unnested.cc +++ b/tensorflow/compiler/xla/service/gpu/ir_emitter_unnested.cc @@ -28,7 +28,7 @@ limitations under the License. #include "llvm/IR/Instructions.h" #include "llvm/IR/LLVMContext.h" #include "llvm/IR/Module.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/ptr_util.h" #include "tensorflow/compiler/xla/service/buffer_assignment.h" #include "tensorflow/compiler/xla/service/dfs_hlo_visitor.h" @@ -48,6 +48,7 @@ limitations under the License. #include "tensorflow/compiler/xla/service/gpu/ir_emitter_context.h" #include "tensorflow/compiler/xla/service/gpu/kernel_thunk.h" #include "tensorflow/compiler/xla/service/gpu/memset_thunk.h" +#include "tensorflow/compiler/xla/service/gpu/outfeed_thunk.h" #include "tensorflow/compiler/xla/service/gpu/parallel_loop_emitter.h" #include "tensorflow/compiler/xla/service/gpu/partition_assignment.h" #include "tensorflow/compiler/xla/service/gpu/sequential_thunk.h" @@ -79,6 +80,7 @@ namespace gpu { namespace { +using llvm_ir::IrArray; using llvm_ir::IrName; using tensorflow::gtl::ArraySlice; using tensorflow::gtl::InlinedVector; @@ -355,7 +357,8 @@ Status IrEmitterUnnested::DefaultAction(HloInstruction* hlo) { unroll_factor = ComputeMaxUnrollFactor(hlo); } - thunk_sequence_->emplace_back(BuildKernelThunk(hlo, unroll_factor)); + thunk_sequence_->emplace_back(BuildKernelThunk( + hlo, /*implements_whole_instruction=*/true, unroll_factor)); return IrEmitter::DefaultAction(hlo); } @@ -369,7 +372,8 @@ Status IrEmitterUnnested::HandleDot(HloInstruction* dot) { thunk_sequence_->emplace_back(BuildGemmThunk(dot)); return Status::OK(); } - thunk_sequence_->emplace_back(BuildKernelThunk(dot)); + thunk_sequence_->emplace_back( + BuildKernelThunk(dot, /*implements_whole_instruction=*/true)); return IrEmitter::HandleDot(dot); } @@ -379,7 +383,8 @@ Status IrEmitterUnnested::HandleConditional(HloInstruction* conditional) { } Status IrEmitterUnnested::HandleConvolution(HloInstruction* convolution) { - thunk_sequence_->emplace_back(BuildKernelThunk(convolution)); + thunk_sequence_->emplace_back( + BuildKernelThunk(convolution, /*implements_whole_instruction=*/true)); return IrEmitter::HandleConvolution(convolution); } @@ -586,10 +591,11 @@ Status IrEmitterUnnested::HandleFusion(HloInstruction* fusion) { } } CHECK(first_reduce != nullptr); - thunks.push_back(BuildKernelThunk(fusion)); + thunks.push_back( + BuildKernelThunk(fusion, /*implements_whole_instruction=*/false)); thunk_sequence_->emplace_back( MakeUnique(std::move(thunks), fusion)); - std::vector parameter_arrays; + std::vector parameter_arrays; for (HloInstruction* operand : fusion->operands()) { parameter_arrays.push_back(GetIrArray(*operand, *fusion)); } @@ -615,6 +621,8 @@ Status IrEmitterUnnested::HandleFusion(HloInstruction* fusion) { output_shape_index = {i}; } if (inst->opcode() == HloOpcode::kReduce) { + CHECK(IsReductionToVector(*inst)) + << "Only reductions to vector are supported"; // Shapes, layouts and dimensions must be the same for all reduces // inside of this fusion. CHECK(ShapeUtil::Equal(first_reduce->shape(), inst->shape())); @@ -658,8 +666,9 @@ Status IrEmitterUnnested::HandleFusion(HloInstruction* fusion) { // touching the un-updated elements. // Set up kernel thunk and fused ir emitter. - thunk_sequence_->emplace_back(BuildKernelThunk(fusion)); - std::vector operand_arrays; + thunk_sequence_->emplace_back( + BuildKernelThunk(fusion, /*implements_whole_instruction=*/true)); + std::vector operand_arrays; for (HloInstruction* operand : fusion->operands()) { operand_arrays.push_back(GetIrArray(*operand, *fusion)); } @@ -672,7 +681,7 @@ Status IrEmitterUnnested::HandleFusion(HloInstruction* fusion) { // Array to write into. Because this is an in-place operation, this is the // same as operand 0's array. - llvm_ir::IrArray output_array = GetIrArray(*fusion, *fusion); + IrArray output_array = GetIrArray(*fusion, *fusion); LaunchDimensions launch_dimensions = CalculateLaunchDimensions( update_shape, ir_emitter_context_->device_description()); @@ -685,314 +694,25 @@ Status IrEmitterUnnested::HandleFusion(HloInstruction* fusion) { fusion, operand_arrays, output_array, &elemental_emitter, launch_dimensions, &ir_builder_); } + if (ImplementedAsGemm(*fusion)) { thunk_sequence_->emplace_back(BuildGemmThunk(fusion)); return Status::OK(); } - CHECK(fusion->fusion_kind() == HloInstruction::FusionKind::kLoop); - int unroll_factor = ComputeMaxUnrollFactor(fusion); + CHECK_EQ(fusion->fusion_kind(), HloInstruction::FusionKind::kLoop); - thunk_sequence_->emplace_back(BuildKernelThunk(fusion, unroll_factor)); - return IrEmitter::HandleFusion(fusion); -} - -namespace { - -// Returns the indices of the first elements of all consecutive subarrays of the -// given array. For example: -// ConsecutiveSegments({m, m+1, m+2, n, k, k+1}) = {0, 3, 4} -std::vector ConsecutiveSegments(tensorflow::gtl::ArraySlice xs) { - std::vector is = {0}; - for (size_t i = 1; i < xs.size(); ++i) { - if (1 != xs[i] - xs[i - 1]) { - is.push_back(i); - } - } - return is; -} - -// Merges the sequences of dimensions of the given shape which start at the -// given indices `segs`. -Shape MergeDimensions(tensorflow::gtl::ArraySlice segs, - const Shape& shape) { - std::vector dimensions; - for (size_t i = 1; i <= segs.size(); ++i) { - dimensions.push_back(std::accumulate( - shape.dimensions().begin() + segs[i - 1], - shape.dimensions().begin() + - (segs.size() == i ? shape.dimensions().size() : segs[i]), - 1, std::multiplies())); - } - return ShapeUtil::MakeShapeWithDescendingLayout(shape.element_type(), - dimensions); -} - -// Returns whether the given shapes and permutation are a 0-2-1 transpose, and -// if so, the normalized and rank-reduced shapes. The shapes must have the same -// dimensions, so this considers layout only. -// -// This function recognizes higher-rank transposes which are elementwise -// equivalent to a 0-2-1 transpose. -std::tuple IsTranspose021(const Shape& a, const Shape& b) { - CHECK(ShapeUtil::Compatible(a, b)); - std::vector perm(a.dimensions().size()); - { - auto layout_a_orig = LayoutUtil::MinorToMajor(a); - std::vector layout_a(layout_a_orig.rbegin(), layout_a_orig.rend()); - auto layout_b_orig = LayoutUtil::MinorToMajor(b); - std::vector layout_b(layout_b_orig.rbegin(), layout_b_orig.rend()); - for (size_t i = 0; i < perm.size(); ++i) { - perm[i] = PositionInContainer(layout_b, layout_a[i]); - } + if (CheckAndEmitHloWithTile021(fusion)) { + return Status::OK(); } - auto segs = ConsecutiveSegments(perm); - Shape norm_a = - ShapeUtil::MakeShapeWithDescendingLayoutAndSamePhysicalLayout(a); - Shape norm_b = - ShapeUtil::MakeShapeWithDescendingLayoutAndSamePhysicalLayout(b); - if (3 == segs.size() && 0 == perm[0]) { - Shape reduced_a = MergeDimensions(segs, norm_a); - Shape reduced_b = ShapeUtil::MakeShapeWithDescendingLayout( - b.element_type(), - Permute({0, 2, 1}, AsInt64Slice(reduced_a.dimensions()))); - return std::make_tuple(true, reduced_a, reduced_b); - } - return std::make_tuple(false, ShapeUtil::MakeNil(), ShapeUtil::MakeNil()); -} - -// Returns whether the given shapes are potentially of a 0-2-1 transpose. -// As 0-2-1 is a self-inverse permutation, which shape is input or output is -// arbitrary. -bool AreShapesForTranspose021(const Shape& a, const Shape& b) { - return 3 == b.dimensions().size() && - ShapeUtil::Compatible( - ShapeUtil::MakeShapeWithDescendingLayoutAndSamePhysicalLayout(a), - ShapeUtil::PermuteDimensions( - {0, 2, 1}, - ShapeUtil::MakeShapeWithDescendingLayoutAndSamePhysicalLayout( - b))); -} - -// Emits a tiled 0-2-1 transpose, assuming both input and output lain out from -// major to minor. The x- and y- dimensions are tiled in square tiles of edge -// length `tile_size`. Each thread block of `tile_size` x `num_rows` threads -// transposes one tile: each thread copies a row from the input to a shared -// memory tile, then copies a column from the shared memory tile to the output. -// -// `tile_size` should usually be same as warp size. -// -// Returns (number of tiles = number of thread blocks needed). -// -// TODO(b/33320379): Here each block transposes 1 tile. It may be more efficient -// to launch fewer blocks so each transposes many tiles, and -// in any case, the number of blocks we can launch is limited. -// -// This is the same algorithm in CUDA: -// https://github.com/tensorflow/tensorflow/blob/d2693c8a70567cc78b2e8a9ac8020d321620ca83/tensorflow/core/kernels/conv_ops_gpu_3.cu.cc#L189 -int64 EmitTranspose021Tiled(llvm_ir::IrArray input, llvm_ir::IrArray output, - const int64 tile_size, const int64 num_rows, - llvm::IRBuilder<>* builder) { - // Adds `addend` to the given `dim` of `index`. - auto offset_dim = [builder](llvm_ir::IrArray::Index index, - llvm::Value* addend, int64 dim) { - index[dim] = builder->CreateAdd(index[dim], addend); - return index; - }; - CHECK(AreShapesForTranspose021(input.GetShape(), output.GetShape())); - - Shape input_shape = - ShapeUtil::MakeShapeWithDescendingLayoutAndSamePhysicalLayout( - input.GetShape()); - Shape output_shape = - ShapeUtil::MakeShapeWithDescendingLayoutAndSamePhysicalLayout( - output.GetShape()); - input = input.CastToShape(input_shape, builder); - output = output.CastToShape(output_shape, builder); - - llvm::Type* tile_type = llvm::ArrayType::get( - llvm::ArrayType::get(input.GetElementLlvmType(), tile_size), - // One extra here to avoid share memory bank conflict - tile_size + 1); - auto* tile = new llvm::GlobalVariable( - *builder->GetInsertBlock()->getParent()->getParent(), tile_type, - /*isConstant=*/false, llvm::GlobalValue::PrivateLinkage, - llvm::UndefValue::get(tile_type), "tile", nullptr, - llvm::GlobalValue::NotThreadLocal, - /*AddressSpace=*/3 /* GPU shared memory */); - - // let x = threadIdx.x - llvm::Value* x = llvm_ir::EmitCallToIntrinsic( - llvm::Intrinsic::nvvm_read_ptx_sreg_tid_x, {}, {}, builder); - llvm_ir::AddRangeMetadata(0, num_rows * tile_size, - static_cast(x)); - x = builder->CreateIntCast(x, builder->getInt64Ty(), /*isSigned=*/true, - "thread.id.x"); - - // computing logical thread ids - // logical_x = x % tile_size - auto logical_x = builder->CreateURem(x, builder->getInt64(tile_size)); - - // logical_y = x / tile_size - auto logical_y = builder->CreateUDiv(x, builder->getInt64(tile_size)); - - // `emit_cp` emits equivalent to following pseudocode: - // if (tile_size == tile_width && tile_size == tile_height) { - // unroll for (i in range(0, tile_size, num_rows)) { - // emit_cp_element(index + {0, i, 0}, y + logical_y); - // } - // } else if (x < tile_width) { - // tile_height_upperbound = ceil(tile_height / num_rows) * num_rows; - // for (i in range(0, tile_height_upperbound, num_rows)) { - // y_loc = i + logical_y; - // if (y_loc < tile_height) - // emit_cp_element(index + {0, i, 0}, y_loc); - // } - // } - // - // We use this to emit both the copy from input to tile and the copy from tile - // to output. - // - // `index` is the origin of the row or column in the input or output array. - // - // `emit_cp_element(index, y)` emits code to copy a single element between the - // tile and the input or output array, where `y` is the `y`-position in the - // tile, whether which is row or column is a function of whether we're copying - // from input or to output, and `index` is the index into the input or output - // array. - auto emit_cp_tile = [builder, tile_size, &offset_dim, num_rows, logical_x, - logical_y]( - std::function - emit_cp_element, - llvm::Value* tile_width, llvm::Value* tile_height, - const llvm_ir::IrArray::Index& index, - const string& loop_name) { - llvm_ir::LlvmIfData if_not_last_row = llvm_ir::EmitIfThenElse( - builder->CreateAnd( - builder->CreateICmpEQ(builder->getInt64(tile_size), tile_width), - builder->CreateICmpEQ(builder->getInt64(tile_size), tile_height)), - "not_last_row", builder); - builder->SetInsertPoint(if_not_last_row.true_block->getTerminator()); - for (int64 i = 0; i < tile_size; i += num_rows) { - auto source_idx = offset_dim(index, builder->getInt64(i), /*dim=*/1); - auto y_loc = builder->CreateAdd(builder->getInt64(i), logical_y); - emit_cp_element(source_idx, y_loc); - } - builder->SetInsertPoint(if_not_last_row.false_block->getTerminator()); - llvm_ir::LlvmIfData if_in_tile = llvm_ir::EmitIfThenElse( - builder->CreateICmpULT(logical_x, tile_width), "x_in_tile", builder); - builder->SetInsertPoint(if_in_tile.true_block->getTerminator()); - - // tile_height_upper_bound = ceil(tile_height / num_rows) * num_rows - auto tile_height_upper_bound = builder->CreateMul( - builder->CreateUDiv( - builder->CreateAdd(tile_height, builder->getInt64(num_rows - 1)), - builder->getInt64(num_rows)), - builder->getInt64(num_rows)); - - auto loop = llvm_ir::ForLoop::EmitForLoop( - loop_name, builder->getInt64(0), tile_height_upper_bound, - builder->getInt64(num_rows), builder); - llvm_ir::SetToFirstInsertPoint(loop->GetHeaderBasicBlock(), builder); - builder->SetInsertPoint(loop->GetBodyBasicBlock()->getTerminator()); - - auto y_loc = builder->CreateAdd(loop->GetIndVarValue(), logical_y); - auto if_y_in_tile = llvm_ir::EmitIfThenElse( - builder->CreateICmpULT(y_loc, tile_height), "y_in_tile", builder); - builder->SetInsertPoint(if_y_in_tile.true_block->getTerminator()); - - emit_cp_element(offset_dim(index, loop->GetIndVarValue(), /*dim=*/1), - y_loc); - builder->SetInsertPoint(if_not_last_row.after_block->getTerminator()); - }; - - auto input_dims_in_tiles = input_shape.dimensions(); - // Unpermuted dimensions are untiled. - for (int i = 1; i < 3; ++i) { - input_dims_in_tiles[i] = - CeilOfRatio(input_dims_in_tiles[i], tile_size); - } - int64 num_tiles = - std::accumulate(input_dims_in_tiles.begin(), input_dims_in_tiles.end(), 1, - std::multiplies()); - const llvm_ir::IrArray::Index input_tile_index( - /*linear=*/builder->CreateIntCast( - llvm_ir::AddRangeMetadata( - 0, num_tiles, - static_cast(llvm_ir::EmitCallToIntrinsic( - llvm::Intrinsic::nvvm_read_ptx_sreg_ctaid_x, {}, {}, - builder))), - builder->getInt64Ty(), /*isSigned=*/true, "block.id.x"), - ShapeUtil::MakeShapeWithDescendingLayout( - PRED /*arbitrary*/, AsInt64Slice(input_dims_in_tiles)), - builder); - const llvm_ir::IrArray::Index input_tile_origin = ({ - llvm_ir::IrArray::Index index = input_tile_index; - for (int i = 1; i < 3; ++i) { - index[i] = builder->CreateMul(index[i], builder->getInt64(tile_size), - "tile_origin." + std::to_string(i)); - } - index; - }); - const llvm_ir::IrArray::Index input_index = - offset_dim(offset_dim(input_tile_origin, logical_x, /*dim=*/2), logical_y, - /*dim=*/1); - std::vector tile_dims(input_shape.dimensions().size()); - // Only last row or column may not have full size. - for (int i = 1; i < 3; ++i) { - tile_dims[i] = builder->CreateSelect( - builder->CreateICmpEQ(input_tile_index[i], - builder->getInt64(input_dims_in_tiles[i] - 1)), - builder->getInt64(input_shape.dimensions(i) - - (input_dims_in_tiles[i] - 1) * tile_size), - builder->getInt64(tile_size), "tile_size"); - } - - // Load data from input memory to shared memory tile. - emit_cp_tile( - // tile[y, x] = input_array[index] - [builder, tile, &input, logical_x](const llvm_ir::IrArray::Index& index, - llvm::Value* y) { - builder->CreateStore( - input.EmitReadArrayElement(index, builder, "input_element"), - builder->CreateGEP(tile, {builder->getInt64(0), y, logical_x})); - }, - tile_dims[2], tile_dims[1], input_index, "input"); + int unroll_factor = ComputeMaxUnrollFactor(fusion); - // Wait for all threads to reach this point, lest we copy a value from tile to - // output before the other thread copies it from input to tile. - // This is `__syncthreads` in CUDA. - llvm_ir::EmitCallToIntrinsic(llvm::Intrinsic::nvvm_barrier0, {}, {}, builder); - - const llvm_ir::IrArray::Index output_tile_index( - Permute({0, 2, 1}, input_tile_index.multidim())); - const llvm_ir::IrArray::Index output_tile_origin( - Permute({0, 2, 1}, input_tile_origin.multidim())); - const llvm_ir::IrArray::Index output_index = - offset_dim(offset_dim(output_tile_origin, logical_x, /*dim=*/2), - logical_y, /*dim=*/1); - - // Store data from shared memory tile to output memory. - emit_cp_tile( - // output_array[index] = tile[x, y] - [builder, tile, &output, logical_x](const llvm_ir::IrArray::Index& index, - llvm::Value* y) { - output.EmitWriteArrayElement( - index, - builder->CreateLoad( - builder->CreateGEP(tile, {builder->getInt64(0), logical_x, y}), - "output_element"), - builder); - }, - tile_dims[1], tile_dims[2], output_index, "output"); - - return num_tiles; + thunk_sequence_->emplace_back(BuildKernelThunk( + fusion, /*implements_whole_instruction=*/true, unroll_factor)); + return IrEmitter::HandleFusion(fusion); } -} // namespace - Status IrEmitterUnnested::HandleCopy(HloInstruction* copy) { if (ImplementedAsHostToDeviceMemcpy(ir_emitter_context_->buffer_assignment(), *copy)) { @@ -1004,25 +724,7 @@ Status IrEmitterUnnested::HandleCopy(HloInstruction* copy) { thunk_sequence_->emplace_back(BuildDeviceToDeviceCopyThunk(copy)); return Status::OK(); } - bool is_transpose_021; - Shape reduced_input_shape, reduced_output_shape; - std::tie(is_transpose_021, reduced_input_shape, reduced_output_shape) = - IsTranspose021(copy->operand(0)->shape(), copy->shape()); - if (is_transpose_021 && - reduced_input_shape.dimensions(1) >= kMinDimensionToTransposeTiled && - reduced_input_shape.dimensions(2) >= kMinDimensionToTransposeTiled) { - thunk_sequence_->emplace_back(BuildKernelThunk(copy)); - VLOG(3) << "Emitting tiled 0-2-1 transposition"; - constexpr int64 tile_size = 32; - constexpr int64 num_rows = 8; - int64 num_tiles = EmitTranspose021Tiled( - GetIrArray(*copy->operand(0), *copy) - .CastToShape(reduced_input_shape, &ir_builder_), - GetIrArray(*copy, *copy) - .CastToShape(reduced_output_shape, &ir_builder_), - tile_size, num_rows, &ir_builder_); - UpdateLaunchDimensions(LaunchDimensions(num_tiles, num_rows * tile_size), - LastThunk(), ir_emitter_context_->llvm_module()); + if (CheckAndEmitHloWithTile021(copy)) { return Status::OK(); } @@ -1030,7 +732,7 @@ Status IrEmitterUnnested::HandleCopy(HloInstruction* copy) { } Status IrEmitterUnnested::EmitExtraOutputsForReduce( - const HloInstruction* reduce, const llvm_ir::IrArray::Index& index, + const HloInstruction* reduce, const IrArray::Index& index, tensorflow::gtl::ArraySlice< std::pair> extra_output_gens) { @@ -1073,11 +775,9 @@ Status IrEmitterUnnested::EmitReductionToScalar( tiled_input_shape, ir_emitter_context_->device_description()); llvm::Type* index_ty = GetIndexTypeForKernel( - reduce, - launch_dimensions.block_count() * launch_dimensions.threads_per_block(), - &ir_builder_); + reduce, launch_dimensions.launch_bound(), &ir_builder_); - auto index_typed_const = [&](uint64 c) -> llvm::Constant* { + auto index_typed_constant = [&](uint64 c) -> llvm::Constant* { return llvm::ConstantInt::get(index_ty, c); }; @@ -1119,8 +819,7 @@ Status IrEmitterUnnested::EmitReductionToScalar( // // and threads_per_block is a multiple of warpSize. // reduce_kernel<<>>(); // - auto loop_body_emitter = - [=](const llvm_ir::IrArray::Index& tile_index) -> Status { + auto loop_body_emitter = [=](const IrArray::Index& tile_index) -> Status { const int num_reduces = reducers.size(); llvm::Type* element_ir_type = llvm_ir::PrimitiveTypeToIrType(input_shape.element_type(), module_); @@ -1129,9 +828,8 @@ Status IrEmitterUnnested::EmitReductionToScalar( llvm::Value* partial_reduction_result_address = ir_builder_.CreateAlloca( element_ir_type, /*ArraySize=*/nullptr, "partial_reduction_result." + llvm::Twine(i)); - TF_ASSIGN_OR_RETURN( - llvm::Value* const init_ir_value, - init_value_gens[i](llvm_ir::IrArray::Index(index_ty))); + TF_ASSIGN_OR_RETURN(llvm::Value* const init_ir_value, + init_value_gens[i](IrArray::Index(index_ty))); ir_builder_.CreateStore(init_ir_value, partial_reduction_result_address); partial_reduction_result_addresses.push_back( partial_reduction_result_address); @@ -1143,21 +841,22 @@ Status IrEmitterUnnested::EmitReductionToScalar( // Emit an inner for-loop that reduces the elements in the tile. auto emit_tile_element_loop = [=](bool tile_in_bounds) -> Status { std::unique_ptr tile_element_loop = - llvm_ir::ForLoop::EmitForLoop( - "element_id_in_tile", index_typed_const(0), - index_typed_const(kTileSize), index_typed_const(1), &ir_builder_); + llvm_ir::ForLoop::EmitForLoop("element_id_in_tile", + index_typed_constant(0), + index_typed_constant(kTileSize), + index_typed_constant(1), &ir_builder_); // Emit the body of the partial reduction loop. llvm_ir::SetToFirstInsertPoint(tile_element_loop->GetBodyBasicBlock(), &ir_builder_); llvm::Value* x = ir_builder_.CreateNSWAdd( - ir_builder_.CreateNSWMul(x_in_tiles, index_typed_const(kTileSize)), + ir_builder_.CreateNSWMul(x_in_tiles, index_typed_constant(kTileSize)), tile_element_loop->GetIndVarValue()); // Unless we know the tile is entirely in bounds, we have to emit a // x-in-bounds check before reading from the input. if (!tile_in_bounds) { llvm_ir::LlvmIfData if_data = llvm_ir::EmitIfThenElse( - ir_builder_.CreateICmpULT(x, index_typed_const(num_elems)), + ir_builder_.CreateICmpULT(x, index_typed_constant(num_elems)), "x_in_bounds", &ir_builder_); // Emit code that reads the input element and accumulates it to @@ -1165,7 +864,7 @@ Status IrEmitterUnnested::EmitReductionToScalar( llvm_ir::SetToFirstInsertPoint(if_data.true_block, &ir_builder_); } - llvm_ir::IrArray::Index input_index( + IrArray::Index input_index( /*linear=*/x, input_shape, &ir_builder_); llvm::Value* input_address = ir_builder_.CreateAlloca(element_ir_type); for (int i = 0; i != num_reduces; ++i) { @@ -1183,12 +882,12 @@ Status IrEmitterUnnested::EmitReductionToScalar( // x_end = kTileSize + x_in_tiles * kTileSize, i.e., the location that's // immediately beyond the tile. llvm::Value* x_end = ir_builder_.CreateNSWAdd( - index_typed_const(kTileSize), - ir_builder_.CreateNSWMul(x_in_tiles, index_typed_const(kTileSize))); + index_typed_constant(kTileSize), + ir_builder_.CreateNSWMul(x_in_tiles, index_typed_constant(kTileSize))); // The tile is entirely in bound if all_threads_in_bounds or // x_end <= num_elems. llvm::Value* tile_in_bounds = ir_builder_.CreateOr( - ir_builder_.CreateICmpULE(x_end, index_typed_const(num_elems)), + ir_builder_.CreateICmpULE(x_end, index_typed_constant(num_elems)), ir_builder_.getInt1(all_threads_in_bounds)); llvm_ir::LlvmIfData if_tile_in_bounds_data = llvm_ir::EmitIfThenElse(tile_in_bounds, "tile_in_bounds", &ir_builder_); @@ -1239,9 +938,9 @@ Status IrEmitterUnnested::EmitReductionToScalar( // lane 0 (which holds the partially accumulated result for its warp) to the // output element. llvm::Value* lane_id = ir_builder_.CreateURem( - x_in_tiles, index_typed_const(kWarpSize), "lane_id"); + x_in_tiles, index_typed_constant(kWarpSize), "lane_id"); llvm_ir::LlvmIfData if_lane_id_is_zero_data = llvm_ir::EmitIfThenElse( - ir_builder_.CreateICmpEQ(lane_id, index_typed_const(0)), + ir_builder_.CreateICmpEQ(lane_id, index_typed_constant(0)), "lane_id_is_zero", &ir_builder_); llvm_ir::SetToFirstInsertPoint(if_lane_id_is_zero_data.true_block, &ir_builder_); @@ -1250,7 +949,7 @@ Status IrEmitterUnnested::EmitReductionToScalar( llvm::Value* output_address = GetIrArray(*output, *output, reduce_output_shapes[i]) .EmitArrayElementAddress( - llvm_ir::IrArray::Index( + IrArray::Index( /*linear=*/ir_builder_.getInt64(0), ShapeUtil::GetSubshape(output->shape(), reduce_output_shapes[i]), @@ -1309,7 +1008,7 @@ Status IrEmitterUnnested::EmitColumnReduction( // TODO(b/110211620): Convert to use i32 index_type when it is possible. llvm::Type* index_ty = ir_builder_.getInt64Ty(); - auto index_typed_const = [&](uint64 c) -> llvm::Constant* { + auto index_typed_constant = [&](uint64 c) -> llvm::Constant* { return llvm::ConstantInt::get(index_ty, c); }; @@ -1336,8 +1035,7 @@ Status IrEmitterUnnested::EmitColumnReduction( // } // AtomicReducer(&output[x], partial_result); // } - auto loop_body_emitter = - [=](const llvm_ir::IrArray::Index& tile_index) -> Status { + auto loop_body_emitter = [=](const IrArray::Index& tile_index) -> Status { const int num_reduces = reducers.size(); // Emit the loop body that reduces one tile. llvm::Type* element_ir_type = @@ -1347,9 +1045,8 @@ Status IrEmitterUnnested::EmitColumnReduction( llvm::Value* partial_reduction_result_address = ir_builder_.CreateAlloca( element_ir_type, /*ArraySize=*/nullptr, "partial_reduction_result." + llvm::Twine(i)); - TF_ASSIGN_OR_RETURN( - llvm::Value* const init_ir_value, - init_value_gens[i](llvm_ir::IrArray::Index(index_ty))); + TF_ASSIGN_OR_RETURN(llvm::Value* const init_ir_value, + init_value_gens[i](IrArray::Index(index_ty))); ir_builder_.CreateStore(init_ir_value, partial_reduction_result_address); partial_reduction_result_addresses.push_back( partial_reduction_result_address); @@ -1365,22 +1062,23 @@ Status IrEmitterUnnested::EmitColumnReduction( auto emit_tile_element_loop = [=](bool tile_in_bounds) -> Status { std::unique_ptr tile_element_loop = - llvm_ir::ForLoop::EmitForLoop( - "element_id_in_tile", index_typed_const(0), - index_typed_const(kTileSize), index_typed_const(1), &ir_builder_); + llvm_ir::ForLoop::EmitForLoop("element_id_in_tile", + index_typed_constant(0), + index_typed_constant(kTileSize), + index_typed_constant(1), &ir_builder_); // Emit the body of the partial reduction loop. llvm_ir::SetToFirstInsertPoint(tile_element_loop->GetBodyBasicBlock(), &ir_builder_); llvm::Value* y = ir_builder_.CreateNSWAdd( - ir_builder_.CreateNSWMul(y_in_tiles, index_typed_const(kTileSize)), + ir_builder_.CreateNSWMul(y_in_tiles, index_typed_constant(kTileSize)), tile_element_loop->GetIndVarValue()); // Unless we know the tile is entirely in bounds, we have to emit a // y-in-bounds check before reading from the input. if (!tile_in_bounds) { llvm_ir::LlvmIfData if_data = llvm_ir::EmitIfThenElse( - ir_builder_.CreateICmpULT(y, index_typed_const(height)), + ir_builder_.CreateICmpULT(y, index_typed_constant(height)), "y_in_bounds", &ir_builder_); // Emit code that reads the input element and accumulates it to @@ -1404,9 +1102,9 @@ Status IrEmitterUnnested::EmitColumnReduction( const Shape input_matrix_shape = ShapeUtil::MakeShapeWithDescendingLayout(input_shape.element_type(), {height, width}); - const llvm_ir::IrArray::Index input_matrix_index( - {y, x}, input_matrix_shape, &ir_builder_); - const llvm_ir::IrArray::Index input_index = + const IrArray::Index input_matrix_index({y, x}, input_matrix_shape, + &ir_builder_); + const IrArray::Index input_index = input_matrix_index .SourceIndexOfReshape(input_matrix_shape, normalized_input_shape, &ir_builder_) @@ -1430,10 +1128,10 @@ Status IrEmitterUnnested::EmitColumnReduction( // y_end = kTileSize + y_in_tiles * kTileSize, i.e., the y location that's // immediately beyond the tile. llvm::Value* y_end = ir_builder_.CreateNSWAdd( - index_typed_const(kTileSize), - ir_builder_.CreateNSWMul(y_in_tiles, index_typed_const(kTileSize))); + index_typed_constant(kTileSize), + ir_builder_.CreateNSWMul(y_in_tiles, index_typed_constant(kTileSize))); llvm::Value* tile_in_bounds = ir_builder_.CreateOr( - ir_builder_.CreateICmpULE(y_end, index_typed_const(height)), + ir_builder_.CreateICmpULE(y_end, index_typed_constant(height)), ir_builder_.getInt1(height % kTileSize == 0)); // The tile is entirely in bound if "height" is a multiple of kTileSize or // y_end <= height. @@ -1457,11 +1155,10 @@ Status IrEmitterUnnested::EmitColumnReduction( llvm::Value* output_address = GetIrArray(*output, *output, reduce_output_shapes[i]) .EmitArrayElementAddress( - llvm_ir::IrArray::Index( - x, - ShapeUtil::GetSubshape(output->shape(), - reduce_output_shapes[i]), - &ir_builder_), + IrArray::Index(x, + ShapeUtil::GetSubshape( + output->shape(), reduce_output_shapes[i]), + &ir_builder_), &ir_builder_, "output_element_address"); TF_RETURN_IF_ERROR(EmitAtomicOperationForNestedComputation( *reducers[i], output_address, partial_reduction_result_addresses[i])); @@ -1529,7 +1226,7 @@ Status IrEmitterUnnested::EmitRowReduction( // for (element_id_in_tile : range(x_tile_size)) { // int x = x_in_tiles * x_tile_size + element_id_in_tile; // if (x < width) - // partial_result = reducer(partial_result, input[z][y][z]); + // partial_result = reducer(partial_result, input[z][y][x]); // } // AtomicReducer(&output[y], partial_result); // } @@ -1583,10 +1280,11 @@ Status IrEmitterUnnested::EmitRowReduction( // for (int element_id_in_z_tile = 0; element_id_in_z_tile < z_tile_size; // ++element_id_in_z_tile) { // z = z_in_tiles * z_tile_size + element_id_in_z_tile; + // int tx = x; // for (int element_id_in_x_tile = 0; // element_id_in_x_tile < x_tile_size; - // ++element_id_in_x_tile, x += warpSize) { - // partial_result = Reducer(partial_result, input[z][y][x]); + // ++element_id_in_x_tile, tx += warpSize) { + // partial_result = Reducer(partial_result, input[z][y][tx]); // } // } // } else { @@ -1594,10 +1292,11 @@ Status IrEmitterUnnested::EmitRowReduction( // for (int element_id_in_z_tile = 0; element_id_in_z_tile < z_tile_size; // ++element_id_in_z_tile) { // z = z_in_tiles * z_tile_size + element_id_in_z_tile; + // int tx = x; // for (int element_id_in_x_tile = 0; element_id_in_x_tile < - // x_tile_size; ++element_id_in_tile, x += warpSize) { - // if (x < width) - // partial_result = Reducer(partial_result, input[z][y][x]); + // x_tile_size; ++element_id_in_tile, tx += warpSize) { + // if (tx < width) + // partial_result = Reducer(partial_result, input[z][y][tx]); // } // } // } @@ -1625,15 +1324,13 @@ Status IrEmitterUnnested::EmitRowReduction( LaunchDimensions launch_dimensions = CalculateLaunchDimensions( tiled_input_shape, ir_emitter_context_->device_description()); llvm::Type* index_ty = GetIndexTypeForKernel( - reduce, - launch_dimensions.block_count() * launch_dimensions.threads_per_block(), - &ir_builder_); + reduce, launch_dimensions.launch_bound(), &ir_builder_); - auto index_typed_const = [&](uint64 c) -> llvm::Constant* { + auto index_typed_constant = [&](uint64 c) -> llvm::Constant* { return llvm::ConstantInt::get(index_ty, c); }; - auto loop_body_emitter = [=](const llvm_ir::IrArray::Index& tile_index) { + auto loop_body_emitter = [=](const IrArray::Index& tile_index) { const int num_reduces = reducers.size(); llvm::Type* element_ir_type = llvm_ir::PrimitiveTypeToIrType( input_shape.element_type(), ir_emitter_context_->llvm_module()); @@ -1642,9 +1339,8 @@ Status IrEmitterUnnested::EmitRowReduction( llvm::Value* partial_reduction_result_address = ir_builder_.CreateAlloca( element_ir_type, /*ArraySize=*/nullptr, "partial_reduction_result." + llvm::Twine(i)); - TF_ASSIGN_OR_RETURN( - llvm::Value* const init_ir_value, - init_value_gens[i](llvm_ir::IrArray::Index(index_ty))); + TF_ASSIGN_OR_RETURN(llvm::Value* const init_ir_value, + init_value_gens[i](IrArray::Index(index_ty))); ir_builder_.CreateStore(init_ir_value, partial_reduction_result_address); partial_reduction_result_addresses.push_back( partial_reduction_result_address); @@ -1656,20 +1352,20 @@ Status IrEmitterUnnested::EmitRowReduction( x_tile = ir_builder_.CreateZExtOrTrunc(x_tile, index_ty); - llvm::Value* warp_id = - ir_builder_.CreateUDiv(x_tile, index_typed_const(kWarpSize), "warp_id"); - llvm::Value* lane_id = - ir_builder_.CreateURem(x_tile, index_typed_const(kWarpSize), "lane_id"); + llvm::Value* warp_id = ir_builder_.CreateUDiv( + x_tile, index_typed_constant(kWarpSize), "warp_id"); + llvm::Value* lane_id = ir_builder_.CreateURem( + x_tile, index_typed_constant(kWarpSize), "lane_id"); // The x-location of the last element in this z-x-tile. // last_x = lane_id + warpSize * (x_tile_size - 1 + warp_id * x_tile_size); llvm::Value* last_x = ir_builder_.CreateNSWAdd( lane_id, ir_builder_.CreateNSWMul( - index_typed_const(kWarpSize), + index_typed_constant(kWarpSize), ir_builder_.CreateNSWAdd( - index_typed_const(x_tile_size - 1), + index_typed_constant(x_tile_size - 1), ir_builder_.CreateNSWMul( - warp_id, index_typed_const(x_tile_size))))); + warp_id, index_typed_constant(x_tile_size))))); KernelSupportLibrary ksl( &ir_builder_, @@ -1682,19 +1378,19 @@ Status IrEmitterUnnested::EmitRowReduction( int64 x_tile_loop_bound) -> Status { auto emit_z_tile_element_loop = [&](llvm::Value* z_indvar) -> Status { llvm::Value* z = ir_builder_.CreateNSWAdd( - z_indvar, - ir_builder_.CreateNSWMul(index_typed_const(z_tile_size), z_tile)); + z_indvar, ir_builder_.CreateNSWMul( + index_typed_constant(z_tile_size), z_tile)); TF_RETURN_IF_ERROR(ksl.For( "x_tile", - /*start=*/index_typed_const(0), - /*end=*/index_typed_const(x_tile_loop_bound), + /*start=*/index_typed_constant(0), + /*end=*/index_typed_constant(x_tile_loop_bound), /*step=*/1, [&](llvm::Value* x_indvar) -> Status { // x = lane_id + // warpSize * (element_id_in_x_tile + warp_id * x_tile_size); llvm::Value* x = ir_builder_.CreateNSWAdd( lane_id, ir_builder_.CreateNSWMul( - index_typed_const(kWarpSize), + index_typed_constant(kWarpSize), ir_builder_.CreateNSWAdd( x_indvar, ir_builder_.CreateNSWMul( warp_id, llvm::ConstantInt::get( @@ -1704,9 +1400,9 @@ Status IrEmitterUnnested::EmitRowReduction( // emit a x-in-bounds check before reading from the input. if (!x_tile_in_bounds) { llvm_ir::LlvmIfData if_x_in_bounds_data = - llvm_ir::EmitIfThenElse( - ir_builder_.CreateICmpULT(x, index_typed_const(width)), - "x_in_bounds", &ir_builder_); + llvm_ir::EmitIfThenElse(ir_builder_.CreateICmpULT( + x, index_typed_constant(width)), + "x_in_bounds", &ir_builder_); // Points ir_builder_ to the then-block. llvm_ir::SetToFirstInsertPoint(if_x_in_bounds_data.true_block, &ir_builder_); @@ -1733,9 +1429,9 @@ Status IrEmitterUnnested::EmitRowReduction( const Shape input_3d_tensor_shape = ShapeUtil::MakeShapeWithDescendingLayout( input_shape.element_type(), {depth, height, width}); - const llvm_ir::IrArray::Index input_3d_tensor_index( + const IrArray::Index input_3d_tensor_index( {z, y, x}, input_3d_tensor_shape, &ir_builder_); - const llvm_ir::IrArray::Index input_index = + const IrArray::Index input_index = input_3d_tensor_index .SourceIndexOfReshape(input_3d_tensor_shape, normalized_input_shape, @@ -1761,14 +1457,14 @@ Status IrEmitterUnnested::EmitRowReduction( }; return ksl.For("z_tile", - /*start=*/index_typed_const(0), - /*end=*/index_typed_const(z_tile_size), + /*start=*/index_typed_constant(0), + /*end=*/index_typed_constant(z_tile_size), /*step=*/1, emit_z_tile_element_loop); }; llvm::Value* tile_in_bounds = ir_builder_.CreateOr( ir_builder_.getInt1(width % (x_tile_size * kWarpSize) == 0), - ir_builder_.CreateICmpULT(last_x, index_typed_const(width))); + ir_builder_.CreateICmpULT(last_x, index_typed_constant(width))); TF_RETURN_IF_ERROR( ksl.If(tile_in_bounds, @@ -1822,7 +1518,7 @@ Status IrEmitterUnnested::EmitRowReduction( // lane 0 (which holds the partially accumulated result for its warp) to the // output element. llvm_ir::LlvmIfData if_lane_id_is_zero_data = llvm_ir::EmitIfThenElse( - ir_builder_.CreateICmpEQ(lane_id, index_typed_const(0)), + ir_builder_.CreateICmpEQ(lane_id, index_typed_constant(0)), "lane_id_is_zero", &ir_builder_); llvm_ir::SetToFirstInsertPoint(if_lane_id_is_zero_data.true_block, &ir_builder_); @@ -1830,21 +1526,22 @@ Status IrEmitterUnnested::EmitRowReduction( llvm::Value* output_address = GetIrArray(*output, *output, reduce_output_shapes[i]) .EmitArrayElementAddress( - llvm_ir::IrArray::Index( - y, - ShapeUtil::GetSubshape(output->shape(), - reduce_output_shapes[i]), - &ir_builder_), + IrArray::Index(y, + ShapeUtil::GetSubshape( + output->shape(), reduce_output_shapes[i]), + &ir_builder_), &ir_builder_, "output_element_address"); - if (x_tile_size * z_tile_size < depth * width) { - TF_RETURN_IF_ERROR(EmitAtomicOperationForNestedComputation( - *reducers[i], output_address, - partial_reduction_result_addresses[i])); - } else { + // We don't need to emit atomic operations if there is only one tile of + // results. 'depth' is the z dimension, 'width' is the x dimension. + if (z_tile_size >= depth && x_tile_size >= width) { TF_RETURN_IF_ERROR(EmitCallToNestedComputation( *reducers[i], {output_address, partial_reduction_result_addresses[i]}, output_address)); + } else { + TF_RETURN_IF_ERROR(EmitAtomicOperationForNestedComputation( + *reducers[i], output_address, + partial_reduction_result_addresses[i])); } } return Status::OK(); @@ -1970,31 +1667,31 @@ Status IrEmitterUnnested::HandleReduce(HloInstruction* reduce) { HloComputation* reducer = reduce->to_apply(); // HandleReduce specializes reduction from a multi-dimensional array to a 1D // array. The specialized version requires an initializer thunk that - // ingitializes the output array to the initial value of the reduce. - if (IsReductionToVector(*reduce) && - // NVPTX backend can't do atomic cmpxchg any narrower than 32 bits - 32 <= primitive_util::BitWidth(reduce->shape().element_type())) { + // initializes the output array to the initial value of the reduce. + if (IsReductionToVector(*reduce)) { TF_ASSIGN_OR_RETURN(std::unique_ptr initializer_thunk, BuildInitializerThunk(reduce)); std::vector> thunks; thunks.push_back(std::move(initializer_thunk)); - thunks.push_back(BuildKernelThunk(reduce)); + thunks.push_back( + BuildKernelThunk(reduce, /*implements_whole_instruction=*/false)); thunk_sequence_->emplace_back( MakeUnique(std::move(thunks), reduce)); return EmitReductionToVector( - reduce, input->shape(), {[&](const llvm_ir::IrArray::Index& index) { + reduce, input->shape(), {[&](const IrArray::Index& index) { return GetIrArray(*input, *reduce) .EmitReadArrayElement(index, &ir_builder_); }}, - {[&](const llvm_ir::IrArray::Index& index) { + {[&](const IrArray::Index& index) { return GetIrArray(*init_value, *reduce) .EmitReadArrayElement(index, &ir_builder_); }}, dimensions_to_reduce, {reducer}, {{}}, {}); } - thunk_sequence_->emplace_back(BuildKernelThunk(reduce)); + thunk_sequence_->emplace_back( + BuildKernelThunk(reduce, /*implements_whole_instruction=*/true)); return IrEmitter::HandleReduce(reduce); } @@ -2023,7 +1720,8 @@ Status IrEmitterUnnested::HandleTuple(HloInstruction* tuple) { tuple_element_buffers, GetAllocationSlice(*tuple), tuple)); return Status::OK(); } - thunk_sequence_->emplace_back(BuildKernelThunk(tuple)); + thunk_sequence_->emplace_back( + BuildKernelThunk(tuple, /*implements_whole_instruction=*/true)); return IrEmitter::HandleTuple(tuple); } @@ -2048,7 +1746,8 @@ Status IrEmitterUnnested::HandleSelectAndScatter( BuildInitializerThunk(select_and_scatter)); std::vector> thunks; thunks.push_back(std::move(initializer_thunk)); - thunks.push_back(BuildKernelThunk(select_and_scatter)); + thunks.push_back(BuildKernelThunk(select_and_scatter, + /*implements_whole_instruction=*/false)); thunk_sequence_->emplace_back( MakeUnique(std::move(thunks), select_and_scatter)); @@ -2062,7 +1761,7 @@ Status IrEmitterUnnested::HandleSelectAndScatter( source->shape(), ir_emitter_context_->device_description()); llvm::Type* index_type = GetIndexTypeForKernel( select_and_scatter, launch_dimensions.launch_bound(), &ir_builder_); - auto index_typed_const = [&](uint64 c) -> llvm::Constant* { + auto index_typed_constant = [&](uint64 c) -> llvm::Constant* { return llvm::ConstantInt::get(index_type, c); }; @@ -2085,8 +1784,7 @@ Status IrEmitterUnnested::HandleSelectAndScatter( // selected_index = I // initialized_flag = true // output(selected_index) = scatter(output(selected_index), source(S)) - auto loop_body_emitter = - [=](const llvm_ir::IrArray::Index& source_index) -> Status { + auto loop_body_emitter = [=](const IrArray::Index& source_index) -> Status { // Allocate space to keep the currently selected value, its index, and a // boolean flag if the value is initialized. The initialized_flag is set // false. @@ -2096,7 +1794,7 @@ Status IrEmitterUnnested::HandleSelectAndScatter( "selected_value_address", &ir_builder_); llvm::Value* selected_index_address = llvm_ir::EmitAllocaAtFunctionEntryWithCount( - index_type, index_typed_const(rank), "selected_index_address", + index_type, index_typed_constant(rank), "selected_index_address", &ir_builder_); llvm::Value* initialized_flag_address = llvm_ir::EmitAllocaAtFunctionEntry( ir_builder_.getInt1Ty(), "initialized_flag_address", &ir_builder_); @@ -2111,7 +1809,7 @@ Status IrEmitterUnnested::HandleSelectAndScatter( window_size.push_back(dim.size()); CHECK_GT(dim.size(), 0); } - const llvm_ir::IrArray::Index window_index = window_loops.AddLoopsForShape( + const IrArray::Index window_index = window_loops.AddLoopsForShape( ShapeUtil::MakeShape(operand_element_type, window_size), "window"); llvm_ir::SetToFirstInsertPoint(window_loops.GetInnerLoopBodyBasicBlock(), &ir_builder_); @@ -2119,17 +1817,17 @@ Status IrEmitterUnnested::HandleSelectAndScatter( // Compute the operand index to visit and evaluate the condition whether the // operand index is within the bounds. The unsigned comparison includes // checking whether the operand index >= 0. - llvm_ir::IrArray::Index operand_index(index_type, source_index.size()); + IrArray::Index operand_index(index_type, source_index.size()); llvm::Value* in_bounds_condition = ir_builder_.getInt1(true); for (int64 i = 0; i < rank; ++i) { llvm::Value* strided_index = ir_builder_.CreateNSWMul( - source_index[i], index_typed_const(window.dimensions(i).stride())); + source_index[i], index_typed_constant(window.dimensions(i).stride())); operand_index[i] = ir_builder_.CreateNSWSub( ir_builder_.CreateNSWAdd(strided_index, window_index[i]), - index_typed_const(window.dimensions(i).padding_low())); + index_typed_constant(window.dimensions(i).padding_low())); llvm::Value* index_condition = ir_builder_.CreateICmpULT( operand_index[i], - index_typed_const(ShapeUtil::GetDimension(operand->shape(), i))); + index_typed_constant(ShapeUtil::GetDimension(operand->shape(), i))); in_bounds_condition = ir_builder_.CreateAnd(in_bounds_condition, index_condition); } @@ -2147,8 +1845,7 @@ Status IrEmitterUnnested::HandleSelectAndScatter( // If the initialized_flag is false, initialize the selected value and index // with the currently visiting operand. llvm_ir::SetToFirstInsertPoint(if_initialized.false_block, &ir_builder_); - const auto save_operand_index = [&]( - const llvm_ir::IrArray::Index& operand_index) { + const auto save_operand_index = [&](const IrArray::Index& operand_index) { for (int64 i = 0; i < rank; ++i) { llvm::Value* selected_index_address_slot = ir_builder_.CreateInBoundsGEP(selected_index_address, @@ -2156,7 +1853,7 @@ Status IrEmitterUnnested::HandleSelectAndScatter( ir_builder_.CreateStore(operand_index[i], selected_index_address_slot); } }; - llvm_ir::IrArray operand_array = GetIrArray(*operand, *select_and_scatter); + IrArray operand_array = GetIrArray(*operand, *select_and_scatter); llvm::Value* operand_data = operand_array.EmitReadArrayElement(operand_index, &ir_builder_); ir_builder_.CreateStore(operand_data, selected_value_address); @@ -2201,7 +1898,7 @@ Status IrEmitterUnnested::HandleSelectAndScatter( // value and the current output value. llvm_ir::SetToFirstInsertPoint(window_loops.GetOuterLoopExitBasicBlock(), &ir_builder_); - llvm_ir::IrArray::Index selected_index(operand_index.GetType()); + IrArray::Index selected_index(operand_index.GetType()); for (int64 i = 0; i < rank; ++i) { llvm::Value* selected_index_address_slot = ir_builder_.CreateInBoundsGEP( selected_index_address, {ir_builder_.getInt32(i)}); @@ -2256,15 +1953,23 @@ Status IrEmitterUnnested::HandleWhile(HloInstruction* xla_while) { } Status IrEmitterUnnested::HandleRng(HloInstruction* random) { - thunk_sequence_->push_back(BuildKernelThunk(random)); + thunk_sequence_->push_back( + BuildKernelThunk(random, /*implements_whole_instruction=*/true)); return IrEmitter::HandleRng(random); } Status IrEmitterUnnested::HandleSelect(HloInstruction* select) { - thunk_sequence_->push_back(BuildKernelThunk(select)); + thunk_sequence_->push_back( + BuildKernelThunk(select, /*implements_whole_instruction=*/true)); return IrEmitter::HandleSelect(select); } +Status IrEmitterUnnested::HandleTupleSelect(HloInstruction* tuple_select) { + thunk_sequence_->push_back( + BuildKernelThunk(tuple_select, /*implements_whole_instruction=*/true)); + return IrEmitter::HandleTupleSelect(tuple_select); +} + Status IrEmitterUnnested::HandleCrossReplicaSum(HloInstruction* crs) { if (hlo_module_config_.replica_count() != 1) { // TODO(b/33011107): Support nontrivial cross replica sum on GPU. @@ -2300,18 +2005,18 @@ Status IrEmitterUnnested::HandleCrossReplicaSum(HloInstruction* crs) { thunks.push_back(MakeUnique( /*source_address=*/GetAllocationSlice(*crs->operand(i)), /*destination_buffer=*/tuple_element_buffers.back(), - /*mem_size=*/ShapeUtil::ByteSizeOf(crs->operand(i)->shape()), crs)); + /*mem_size=*/ShapeUtil::ByteSizeOf(crs->operand(i)->shape()), nullptr)); } // Output a tuple of the buffers above. thunks.push_back(MakeUnique(tuple_element_buffers, - GetAllocationSlice(*crs), crs)); + GetAllocationSlice(*crs), nullptr)); thunk_sequence_->push_back( MakeUnique(std::move(thunks), crs)); return Status::OK(); } -Status IrEmitterUnnested::HandleGenerateToken(HloInstruction* gen_token) { +Status IrEmitterUnnested::HandleAfterAll(HloInstruction* gen_token) { return Status::OK(); } @@ -2320,6 +2025,11 @@ Status IrEmitterUnnested::HandleInfeed(HloInstruction* infeed) { return Status::OK(); } +Status IrEmitterUnnested::HandleOutfeed(HloInstruction* outfeed) { + thunk_sequence_->emplace_back(BuildOutfeedThunk(outfeed)); + return Status::OK(); +} + // Figures out how to access the buffers for all subshapes of hlo's operands and // for hlo itself (i.e. all the buffers produced by HLO). // @@ -2439,7 +2149,8 @@ GetHloBufferSlices(const HloInstruction* hlo, } std::unique_ptr IrEmitterUnnested::BuildKernelThunk( - const HloInstruction* inst, int unroll_factor) { + const HloInstruction* inst, bool implements_whole_instruction, + int unroll_factor) { const BufferAssignment& buffer_assn = ir_emitter_context_->buffer_assignment(); @@ -2531,7 +2242,8 @@ std::unique_ptr IrEmitterUnnested::BuildKernelThunk( } return MakeUnique(buffers, llvm_ir::AsString(kernel->getName()), - inst, unroll_factor); + implements_whole_instruction ? inst : nullptr, + unroll_factor); } std::unique_ptr IrEmitterUnnested::BuildHostToDeviceCopyThunk( @@ -2563,17 +2275,31 @@ std::unique_ptr IrEmitterUnnested::BuildInfeedThunk( const HloInstruction* inst) { CHECK_EQ(HloOpcode::kInfeed, inst->opcode()); - std::vector tuple_element_buffers; - for (int64 i = 0; i < inst->shape().tuple_shapes_size(); ++i) { - BufferAllocation::Slice buffer = ir_emitter_context_->buffer_assignment() - .GetUniqueSlice(inst, {i}) - .ConsumeValueOrDie(); - tuple_element_buffers.push_back(buffer); - } + ShapeTree slices(inst->shape()); + slices.ForEachMutableElement( + [&](const ShapeIndex& index, BufferAllocation::Slice* slice) { + *slice = ir_emitter_context_->buffer_assignment() + .GetUniqueSlice(inst, index) + .ConsumeValueOrDie(); + }); + return MakeUnique(slices, inst); +} - return MakeUnique( - tuple_element_buffers, - /*destination_buffer=*/GetAllocationSlice(*inst), inst); +std::unique_ptr IrEmitterUnnested::BuildOutfeedThunk( + const HloInstruction* inst) { + CHECK_EQ(HloOpcode::kOutfeed, inst->opcode()); + + ShapeTree slices(inst->operand(0)->shape()); + slices.ForEachMutableElement( + [&](const ShapeIndex& index, BufferAllocation::Slice* slice) { + auto status_or_slice = + ir_emitter_context_->buffer_assignment().GetUniqueSlice( + inst->operand(0), index); + if (status_or_slice.ok()) { + *slice = status_or_slice.ConsumeValueOrDie(); + } + }); + return MakeUnique(std::move(slices), inst); } namespace { @@ -2691,6 +2417,11 @@ StatusOr> IrEmitterUnnested::BuildInitializerThunk( init_value = hlo->operand(init_value->parameter_number()); } + // Initializer thunks don't implement a whole instruction, and we want to + // profile the whole instruction instead of the individual thunks it consists + // of. Therefore we pass nullptr as the HloInstruction* to the thunks we + // generate below. + // // In the common case, the initializer is a constant. In this case, emit a // device-memset call if we can. Currently StreamExecutor only supports // zeroing and 32-bit memsets. @@ -2704,7 +2435,8 @@ StatusOr> IrEmitterUnnested::BuildInitializerThunk( ArraySlice literal_bytes( reinterpret_cast(literal.untyped_data()), num_bytes); if (c_all_of(literal_bytes, [](uint8 byte) { return byte == 0; })) { - return {MakeUnique(GetAllocationSlice(*hlo, index), hlo)}; + return { + MakeUnique(GetAllocationSlice(*hlo, index), nullptr)}; } // If the literal is 8 or 16 bits wide, we can emit a 32-bit memset by @@ -2718,11 +2450,11 @@ StatusOr> IrEmitterUnnested::BuildInitializerThunk( uint8 b = literal_bytes.front(); pattern16 = uint16{b} | (uint16{b} << 8); } else { - pattern16 = literal_bytes.front(); + memcpy(&pattern16, literal_bytes.data(), sizeof(pattern16)); } uint32 pattern32 = uint32{pattern16} | (uint32{pattern16} << 16); return {MakeUnique( - pattern32, GetAllocationSlice(*hlo, index), hlo)}; + pattern32, GetAllocationSlice(*hlo, index), nullptr)}; } // If the literal is an even multiple of 32 bits wide, we can emit a 32-bit @@ -2733,12 +2465,13 @@ StatusOr> IrEmitterUnnested::BuildInitializerThunk( uint32 word; memcpy(&word, literal_bytes.data(), sizeof(word)); return {MakeUnique( - word, GetAllocationSlice(*hlo, index), hlo)}; + word, GetAllocationSlice(*hlo, index), nullptr)}; } } // Otherwise fall back to our slow initializer code. - std::unique_ptr kernel_thunk = BuildKernelThunk(hlo); + std::unique_ptr kernel_thunk = + BuildKernelThunk(hlo, /*implements_whole_instruction=*/false); LaunchDimensions launch_dimensions = CalculateLaunchDimensions(ShapeUtil::GetSubshape(hlo->shape(), index), ir_emitter_context_->device_description()); @@ -2750,7 +2483,7 @@ StatusOr> IrEmitterUnnested::BuildInitializerThunk( TF_RETURN_IF_ERROR(HandleConstant(const_cast(init_value))); } TF_RETURN_IF_ERROR(ParallelLoopEmitter( - [=](const llvm_ir::IrArray::Index& index) { + [=](const IrArray::Index& index) { return GetIrArray(*init_value, *hlo) .EmitReadArrayElement(index, &ir_builder_); }, @@ -2945,8 +2678,8 @@ Status IrEmitterUnnested::EmitTargetElementLoopInThunk( &ir_builder_)); } - // For multiple outputs fusion, we need to emit each operand and the root. - std::vector output_arrays; + // For multioutput fusion, we need to emit each operand and the root. + std::vector output_arrays; for (int64 i = 0; i < ShapeUtil::TupleElementCount(hlo.shape()); ++i) { output_arrays.push_back(GetIrArray(hlo, hlo, {i})); } @@ -2975,5 +2708,482 @@ Status IrEmitterUnnested::EmitTargetElementLoop( static_cast(LastThunk())); } +int IrEmitterUnnested::ConstructIrArrayForOutputs( + const HloInstruction& hlo, std::vector* output_arrays) { + int64 num_outputs = 1; + if (hlo.IsMultiOutputFusion()) { + num_outputs = ShapeUtil::TupleElementCount(hlo.shape()); + output_arrays->reserve(num_outputs); + for (int64 i = 0; i < num_outputs; ++i) { + output_arrays->push_back(GetIrArray(hlo, hlo, {i})); + } + } else { + output_arrays->push_back(GetIrArray(hlo, hlo)); + } + return num_outputs; +} + +int IrEmitterUnnested::ConstructIrArrayForInputs( + const HloInstruction& hlo, std::vector* param_arrays) { + int64 num_params = hlo.operands().size(); + param_arrays->reserve(num_params); + for (const HloInstruction* param : hlo.operands()) { + param_arrays->push_back(GetIrArray(*param, hlo)); + } + return num_params; +} + +int IrEmitterUnnested::ConstructOutputReducedShapeAndCastOutputIrArrayToShape( + const HloInstruction& hlo, const std::vector& output_arrays, + tensorflow::gtl::ArraySlice reduced_output_dims, + std::vector* output_reduced_shapes, + std::vector* output_in_reduced_shape_arrays) { + int64 num_outputs = 1; + if (hlo.IsMultiOutputFusion()) { + num_outputs = ShapeUtil::TupleElementCount(hlo.shape()); + output_in_reduced_shape_arrays->reserve(num_outputs); + output_reduced_shapes->reserve(num_outputs); + for (int64 i = 0; i < num_outputs; ++i) { + output_reduced_shapes->push_back(ShapeUtil::MakeShapeWithDescendingLayout( + ShapeUtil::GetSubshape(hlo.shape(), {i}).element_type(), + reduced_output_dims)); + output_in_reduced_shape_arrays->push_back(output_arrays[i].CastToShape( + (*output_reduced_shapes)[i], &ir_builder_)); + } + } else { + output_reduced_shapes->push_back(ShapeUtil::MakeShapeWithDescendingLayout( + hlo.shape().element_type(), reduced_output_dims)); + output_in_reduced_shape_arrays->push_back(output_arrays[0].CastToShape( + (*output_reduced_shapes)[0], &ir_builder_)); + } + return num_outputs; +} + +int IrEmitterUnnested::ConstructInputReducedShapeAndCastInputIrArrayToShape( + const HloInstruction& hlo, const std::vector& param_arrays, + const std::vector& param_buffers, + tensorflow::gtl::ArraySlice reduced_output_dims, + std::vector* param_reduced_shapes, + std::vector* param_in_reduced_shape_arrays) { + int64 num_params = hlo.operands().size(); + param_in_reduced_shape_arrays->reserve(num_params); + param_reduced_shapes->reserve(num_params); + for (int64 id = 0; id < num_params; ++id) { + if (param_buffers[id] == nullptr) { + param_reduced_shapes->push_back(Shape()); + param_in_reduced_shape_arrays->push_back(IrArray()); + continue; + } + const HloInstruction* param = hlo.operand(id); + param_reduced_shapes->push_back(ShapeUtil::MakeShapeWithDescendingLayout( + param->shape().element_type(), + Permute({0, 2, 1}, reduced_output_dims))); + param_in_reduced_shape_arrays->push_back(param_arrays[id].CastToShape( + (*param_reduced_shapes)[id], &ir_builder_)); + } + return num_params; +} + +namespace { + +// Reads thread_idx.x and converts it to a (y,x) coordinate, assuming that the +// thread lives within a square tile of size tile_size (so thread blocks are of +// size tile_size * tile_size). +std::tuple CalculateYXCoordinateWithinTile( + llvm::IRBuilder<>* builder, llvm::Value* tile_size, + int64 threads_per_tile) { + // Calculate the starting element coordinate within a tile for the current + // thread, (y, x) from thread_id. + llvm::Value* thread_id = llvm_ir::EmitCallToIntrinsic( + llvm::Intrinsic::nvvm_read_ptx_sreg_tid_x, {}, {}, builder); + llvm_ir::AddRangeMetadata(0, threads_per_tile, + llvm::cast(thread_id)); + thread_id = builder->CreateIntCast(thread_id, tile_size->getType(), + /*isSigned=*/true, "thread.id.x"); + auto x = builder->CreateURem(thread_id, tile_size); + auto y = builder->CreateUDiv(thread_id, tile_size); + return std::make_tuple(y, x); +} + +// Reads block_idx.x, casts it to type index_ty, and adds the assumption that +// it's in the range [0, num_blocks]. +llvm::Value* GetBlockIdx(llvm::IRBuilder<>* builder, llvm::Type* index_ty, + int64 num_blocks) { + llvm::Value* block_id = llvm_ir::EmitCallToIntrinsic( + llvm::Intrinsic::nvvm_read_ptx_sreg_ctaid_x, {}, {}, builder); + llvm_ir::AddRangeMetadata(0, num_blocks, + llvm::cast(block_id)); + return builder->CreateIntCast(block_id, index_ty, /*isSigned=*/true, + "block.id.x"); +} + +// Emits code to process up to (tile_size/num_rows) elements in a tile, given +// `emit_elem_function` is the function to emit code to process one element, `y` +// and `x` are the coordinates for the first element to process, and `index` is +// the index for the origin of the tile. Emits bounds check to ensure that each +// processed element is within the boundary defined by `tile_width` and +// `tile_height`. +void EmitTiledElementalCodeWithBoundsCheck( + int64 tile_size, int64 num_rows, const IrArray::Index& index, + const string& loop_name, KernelSupportLibrary* ksl, + llvm::IRBuilder<>* builder, llvm::Value* y, llvm::Value* x, + llvm::Value* tile_width, llvm::Value* tile_height, + const std::function& + emit_elem_function) { + llvm::Type* index_ty = tile_width->getType(); + // Emits a constant value with index type. + auto index_typed_constant = [&](uint64 c) -> llvm::Constant* { + return llvm::ConstantInt::get(index_ty, c); + }; + // Adds `addend` to the given `dim` of `index`. + auto offset_dim = [&](IrArray::Index index, llvm::Value* addend, int64 dim) { + index[dim] = builder->CreateAdd(index[dim], addend); + return index; + }; + + auto emit_full_tile = [&] { + for (int64 i = 0; i < tile_size; i += num_rows) { + auto source_idx = offset_dim(index, index_typed_constant(i), /*dim=*/1); + auto y_loc = builder->CreateAdd(index_typed_constant(i), y); + emit_elem_function(source_idx, y_loc); + } + }; + + auto emit_last_row = [&] { + ksl->IfReturnVoid("x_in_tile", builder->CreateICmpULT(x, tile_width), [&] { + // tile_height_upper_bound = + // ceil(tile_height / num_rows) * num_rows + auto tile_height_upper_bound = builder->CreateMul( + builder->CreateUDiv( + builder->CreateAdd(tile_height, + index_typed_constant(num_rows - 1)), + index_typed_constant(num_rows)), + index_typed_constant(num_rows)); + ksl->ForReturnVoid( + loop_name, /*start=*/index_typed_constant(0), + /*end=*/tile_height_upper_bound, + /*step=*/index_typed_constant(num_rows), [&](llvm::Value* y_indvar) { + auto y_loc = builder->CreateAdd(y_indvar, y); + ksl->IfReturnVoid( + "y_in_tile", builder->CreateICmpULT(y_loc, tile_height), [&] { + emit_elem_function(offset_dim(index, y_indvar, /*dim=*/1), + y_loc); + }); + }); + }); + }; + ksl->IfReturnVoid( + "full_tile", + builder->CreateAnd( + builder->CreateICmpEQ(index_typed_constant(tile_size), tile_width), + builder->CreateICmpEQ(index_typed_constant(tile_size), tile_height)), + emit_full_tile, emit_last_row); +} +} // namespace + +// Emits a kernel for the given hlo instruction using a tiled 0-2-1 transpose +// algorithm to improve the memory access patterns for the input parameters +// which have a shape that is a 0-2-1 transpose of the output tensors. +// +// For the purpose of tiling, the output tensors have a logical shape of three +// components 0-2-1 while the relevant input parameters have a logical shape of +// three components 0-1-2 in the order major to minor. The x- and y- dimensions +// of the tensors are tiled in square tiles of edge length `kTileSize`. Each +// thread block of `kTileSize` x `kNumRows` threads transposes one tile: each +// thread copies kTileSize/kNumRows elements from the input to a shared memory +// tile, then the otherwise "regular hlo kernel" reads from the shared memory +// instead of the original input. +// +// This is similar to the following CUDA algorithm in TensorFlow: +// https://goo.gl/MStRV6. +// +// `kTileSize` should usually be same as warp size. We currently choose 32 for +// `kTileSize` and 4 for `kNumRows`. The CUDA algorithm uses 8 for `kNumRows`. +// +// TODO(b/33320379): Here each block transposes 1 tile. It may be more efficient +// to launch fewer blocks so each transposes many tiles. +LaunchDimensions IrEmitterUnnested::EmitHlo021Tile( + HloInstruction* hlo, tensorflow::gtl::ArraySlice reduced_output_dims, + tensorflow::gtl::ArraySlice tiled_param_ids) { + // Parameters for the tiling algorithm. + constexpr int64 kTileSize = 32; + constexpr int64 kNumRows = 4; + constexpr int64 kThreadsPerTile = kTileSize * kNumRows; + + // Construct IrArrays for the inputs and outputs. + std::vector output_arrays; + int64 num_outputs = ConstructIrArrayForOutputs(*hlo, &output_arrays); + std::vector param_arrays; + int64 num_params = ConstructIrArrayForInputs(*hlo, ¶m_arrays); + + // Allocate shared memory buffers to store the tiled inputs. + std::vector param_shmem_buffers(num_params, nullptr); + for (int64 id : tiled_param_ids) { + const HloInstruction* param = hlo->operand(id); + // Add 1 to the minor dimension to reduce shared memory bank conflicts. + llvm::Type* tile_type = llvm::ArrayType::get( + llvm::ArrayType::get(llvm_ir::PrimitiveTypeToIrType( + param->shape().element_type(), module_), + kTileSize + 1), + kTileSize); + const int kNVPTXSharedMemoryAddrSpace = 3; + auto* tile_base_ptr = new llvm::GlobalVariable( + *ir_builder_.GetInsertBlock()->getParent()->getParent(), tile_type, + /*isConstant=*/false, llvm::GlobalValue::PrivateLinkage, + llvm::UndefValue::get(tile_type), + llvm_ir::AsStringRef(IrName(hlo, StrCat("tile", id))), nullptr, + llvm::GlobalValue::NotThreadLocal, kNVPTXSharedMemoryAddrSpace); + param_shmem_buffers[id] = tile_base_ptr; + VLOG(3) << "Added shmem buffer for parameter " << id << ": " + << llvm_ir::DumpToString(*tile_base_ptr); + } + + // The 0-2-1 shape of the tiling scheme is the reduced shape of the HLO result + // for the purpose of tiling. Calculate the logical output dimensions in the + // tile from the reduced output dimensions. + std::vector output_dims_in_tiles = std::vector( + reduced_output_dims.begin(), reduced_output_dims.end()); + CHECK_EQ(output_dims_in_tiles.size(), 3); + for (int i = 1; i < 3; ++i) { + output_dims_in_tiles[i] = + CeilOfRatio(output_dims_in_tiles[i], kTileSize); + } + const int64 num_tiles = + c_accumulate(output_dims_in_tiles, 1, std::multiplies()); + LaunchDimensions launch_dimensions(num_tiles, kThreadsPerTile); + + llvm::Type* index_ty = GetIndexTypeForKernel( + hlo, launch_dimensions.launch_bound(), &ir_builder_); + auto index_typed_constant = [&](uint64 c) -> llvm::Constant* { + return llvm::ConstantInt::get(index_ty, c); + }; + + // Cast each output IrArray to its corresponding reduced shape and keep the + // reduced shape live during IR emission. + std::vector output_in_reduced_shape_arrays; + std::vector output_reduced_shapes; + CHECK_EQ(ConstructOutputReducedShapeAndCastOutputIrArrayToShape( + *hlo, output_arrays, reduced_output_dims, &output_reduced_shapes, + &output_in_reduced_shape_arrays), + num_outputs); + + // For each tiled parameter, cast its input IrArray to the corresponding + // reduced shape and keep the reduced shape live during IR emission. + std::vector param_in_reduced_shape_arrays; + std::vector param_reduced_shapes; + CHECK_EQ(ConstructInputReducedShapeAndCastInputIrArrayToShape( + *hlo, param_arrays, param_shmem_buffers, reduced_output_dims, + ¶m_reduced_shapes, ¶m_in_reduced_shape_arrays), + num_params); + + // Calculate the starting element coordinate within a tile for the current + // thread, (y, x) from thread_id. + llvm::Value* x; + llvm::Value* y; + std::tie(y, x) = CalculateYXCoordinateWithinTile( + &ir_builder_, index_typed_constant(kTileSize), kThreadsPerTile); + + // Calculate the index for the current output tile from block_id. + const IrArray::Index output_tile_index( + GetBlockIdx(&ir_builder_, index_ty, num_tiles), + ShapeUtil::MakeShapeWithDescendingLayout(PRED /*arbitrary*/, + output_dims_in_tiles), + &ir_builder_); + + // Output tile origin is the index for the first element of the current output + // tile. + const IrArray::Index output_tile_origin = [&] { + IrArray::Index index = output_tile_index; + for (int i = 1; i < 3; ++i) { + index[i] = ir_builder_.CreateMul(output_tile_index[i], + index_typed_constant(kTileSize), + "tile_origin." + std::to_string(i)); + } + return index; + }(); + + // Calculate the input tile origin from the output tile origin. + const IrArray::Index input_tile_origin( + Permute({0, 2, 1}, output_tile_origin.multidim())); + + // Calculate the current output tile bounds in each of the logical dimensions. + std::vector output_tile_bounds(3); + for (int i = 1; i < 3; ++i) { + // Only last row or column may not have full size. + output_tile_bounds[i] = ir_builder_.CreateSelect( + ir_builder_.CreateICmpEQ( + output_tile_index[i], + index_typed_constant(output_dims_in_tiles[i] - 1)), + index_typed_constant(reduced_output_dims[i] - + (output_dims_in_tiles[i] - 1) * kTileSize), + index_typed_constant(kTileSize), "kTileSize"); + } + + KernelSupportLibrary ksl(&ir_builder_, llvm_ir::UnrollMode::kDefaultUnroll); + + // Curry a few parameters to EmitTiledElementalCodeWithBoundsCheck. + auto emit_tiled_elemental_code_with_bounds_check = + [&](const IrArray::Index& index, const string& loop_name, + llvm::Value* tile_width, llvm::Value* tile_height, + const std::function& + emit_elem_function) { + EmitTiledElementalCodeWithBoundsCheck( + kTileSize, kNumRows, index, loop_name, &ksl, &ir_builder_, y, x, + tile_width, tile_height, emit_elem_function); + }; + + // Adds `addend` to the given `dim` of `index`. + auto offset_dim = [&](IrArray::Index index, llvm::Value* addend, int64 dim) { + index[dim] = ir_builder_.CreateAdd(index[dim], addend); + return index; + }; + const IrArray::Index input_index = + offset_dim(offset_dim(input_tile_origin, x, /*dim=*/2), y, /*dim=*/1); + + // Copy input parameter values to shared memory buffers: + // tile[y, x] = input[index] + emit_tiled_elemental_code_with_bounds_check( + input_index, "input", output_tile_bounds[1], output_tile_bounds[2], + [&](const IrArray::Index& index, llvm::Value* y_loc) { + for (int64 id : tiled_param_ids) { + IrArray& input_in_logical_shape = param_in_reduced_shape_arrays[id]; + llvm::Value* shmem_buffer = param_shmem_buffers[id]; + // TODO(jlebar): Add AA metadata to this store. Tile buffers are + // global variables, so LLVM can't infer much about it. + ir_builder_.CreateStore( + input_in_logical_shape.EmitReadArrayElement(index, &ir_builder_, + "input_element"), + ir_builder_.CreateGEP(shmem_buffer, + {index_typed_constant(0), y_loc, x})); + } + }); + + // Wait for all threads to reach this point, lest we copy a value from tile to + // output before the other thread copies it from input to tile. + // This is `__syncthreads` in CUDA. + llvm_ir::EmitCallToIntrinsic(llvm::Intrinsic::nvvm_barrier0, {}, {}, + &ir_builder_); + + llvm_ir::TiledParameterInfo tiled_param_info(param_shmem_buffers, y, x); + + const IrArray::Index output_index = + offset_dim(offset_dim(output_tile_origin, x, /*dim=*/2), y, /*dim=*/1); + + // Write to output[index] by emitting code like normal, except that values for + // the tiled parameters are read from the shmem buffers. + if (hlo->opcode() == HloOpcode::kCopy) { + emit_tiled_elemental_code_with_bounds_check( + output_index, "output", output_tile_bounds[2], output_tile_bounds[1], + [&](const IrArray::Index& index, llvm::Value* y_loc) { + // TODO(jlebar): Add AA metadata to this load. + llvm::Instruction* load_from_shmem_buffer = ir_builder_.CreateLoad( + ir_builder_.CreateGEP(param_shmem_buffers[0], + {ir_builder_.getInt64(0), x, y_loc}), + "output_element"); + output_in_reduced_shape_arrays[0].EmitWriteArrayElement( + index, load_from_shmem_buffer, &ir_builder_); + }); + } else { + CHECK_EQ(hlo->opcode(), HloOpcode::kFusion); + emit_tiled_elemental_code_with_bounds_check( + output_index, "output", output_tile_bounds[2], output_tile_bounds[1], + [&](const IrArray::Index& index, llvm::Value* y_loc) { + GpuElementalIrEmitter elem_emitter(hlo_module_config_, module_, + &ir_builder_, GetNestedComputer()); + FusedIrEmitter fused_emitter(param_arrays, &elem_emitter); + tiled_param_info.set_y(y_loc); + fused_emitter.SetTiledParameterInfo(&tiled_param_info); + TF_CHECK_OK(hlo->fused_expression_root()->Accept(&fused_emitter)); + IrArray::Index untiled_index = llvm_ir::GetUnreducedOutputIndex( + index, output_reduced_shapes[0], output_arrays[0].GetShape(), + &ir_builder_); + const llvm_ir::ElementGenerator& output_generator = + fused_emitter.GetRootGenerator(); + llvm::Value* output_value = + output_generator(untiled_index).ValueOrDie(); + if (hlo->IsMultiOutputFusion()) { + CHECK(output_value->getType()->isStructTy()); + CHECK_EQ(output_value->getType()->getStructNumElements(), + output_in_reduced_shape_arrays.size()); + for (int64 i = 0; i < output_in_reduced_shape_arrays.size(); ++i) { + output_in_reduced_shape_arrays[i].EmitWriteArrayElement( + index, ir_builder_.CreateExtractValue(output_value, i), + &ir_builder_); + } + } else { + output_in_reduced_shape_arrays[0].EmitWriteArrayElement( + index, output_value, &ir_builder_); + } + }); + } + + // For multioutput fusion, emit a tuple with all the individual outputs. + if (hlo->IsMultiOutputFusion()) { + std::vector tuple_operand_ptrs; + for (int64 i = 0; i < output_arrays.size(); ++i) { + tuple_operand_ptrs.push_back(output_arrays[i].GetBasePointer()); + } + llvm_ir::EmitTuple(GetIrArray(*hlo, *hlo), tuple_operand_ptrs, &ir_builder_, + module_); + } + + return launch_dimensions; +} + +bool IrEmitterUnnested::CheckAndEmitHloWithTile021(HloInstruction* hlo) { + HloOpcode opcode = hlo->opcode(); + CHECK(opcode == HloOpcode::kFusion || opcode == HloOpcode::kCopy); + CHECK(opcode != HloOpcode::kFusion || + hlo->fusion_kind() == HloInstruction::FusionKind::kLoop) + << "Only loop fusions are supported."; + + const Shape& output_shape = hlo->IsMultiOutputFusion() + ? ShapeUtil::GetSubshape(hlo->shape(), {0}) + : hlo->shape(); + + // If the output_shape is reduced to 021 shape, find all the parameters of the + // hlo that are in the corresponding 012 shape. + std::vector params_012; + optional> reduced_dims_021; + for (int64 operand_idx = 0; operand_idx < hlo->operand_count(); + ++operand_idx) { + HloInstruction* operand = hlo->mutable_operand(operand_idx); + auto find_transpose_result = + llvm_ir::FindTranspose021(operand->shape(), output_shape); + if (!find_transpose_result.has_value()) { + continue; + } + const std::vector& curr_reduced_dims_021 = *find_transpose_result; + if (!reduced_dims_021.has_value()) { + reduced_dims_021 = curr_reduced_dims_021; + } + if (!ContainersEqual(*reduced_dims_021, curr_reduced_dims_021)) { + // There is more than one possible transpose. Instead of picking one + // transpose, we simply give up here. + return false; + } + params_012.push_back(operand_idx); + } + + if (!reduced_dims_021.has_value()) { + return false; + } + + if ((*reduced_dims_021)[1] < kMinDimensionToTransposeTiled || + (*reduced_dims_021)[2] < kMinDimensionToTransposeTiled) { + return false; + } + + VLOG(3) << "EmitHlo021Tile Emitting hlo tile 0-2-1" << hlo->ToString(); + thunk_sequence_->emplace_back( + BuildKernelThunk(hlo, /*implements_whole_instruction=*/true)); + const LaunchDimensions launch_dimensions = + EmitHlo021Tile(hlo, *reduced_dims_021, params_012); + UpdateLaunchDimensions(launch_dimensions, LastThunk(), + ir_emitter_context_->llvm_module()); + + return true; +} + } // namespace gpu } // namespace xla diff --git a/tensorflow/compiler/xla/service/gpu/ir_emitter_unnested.h b/tensorflow/compiler/xla/service/gpu/ir_emitter_unnested.h index 279a5c386ad15857e0a0f6ae18ccf7cc5183e0a6..59547c16d7fe08baa03c7a9257a27423ac2cd871 100644 --- a/tensorflow/compiler/xla/service/gpu/ir_emitter_unnested.h +++ b/tensorflow/compiler/xla/service/gpu/ir_emitter_unnested.h @@ -18,6 +18,7 @@ limitations under the License. #include "tensorflow/compiler/xla/service/gpu/ir_emitter.h" #include "tensorflow/compiler/xla/service/gpu/thunk.h" +#include "tensorflow/compiler/xla/service/llvm_ir/kernel_tiling.h" namespace xla { namespace gpu { @@ -73,10 +74,12 @@ class IrEmitterUnnested : public IrEmitter { Status HandleTuple(HloInstruction* tuple) override; Status HandleWhile(HloInstruction* xla_while) override; Status HandleInfeed(HloInstruction* xla_infeed) override; + Status HandleOutfeed(HloInstruction* outfeed) override; Status HandleRng(HloInstruction* random) override; Status HandleSelect(HloInstruction* select) override; + Status HandleTupleSelect(HloInstruction* tuple_select) override; Status HandleCrossReplicaSum(HloInstruction* crs) override; - Status HandleGenerateToken(HloInstruction* gen_token) override; + Status HandleAfterAll(HloInstruction* gen_token) override; Status EmitTargetElementLoop( const HloInstruction& hlo, @@ -115,7 +118,7 @@ class IrEmitterUnnested : public IrEmitter { // Emits code that reduces a matrix of shape [height x width] to a vector of // [width]. Other parameters have the same meaning as those of // `EmitReductionToVector`. Note that input shape might not be - // [height x width], but can be bitcast to [height x weight] with "height" + // [height x width], but can be bitcast to [height x width] with "height" // being the major dimension. Status EmitColumnReduction( int64 height, int64 width, HloInstruction* reduce, @@ -131,7 +134,7 @@ class IrEmitterUnnested : public IrEmitter { // Emits code that reduces a 3D tensor of shape [depth x height x width] to a // vector of shape [height]. Other parameters have the same meaning as those // of `EmitReductionToVector`. Note that input shape might not be - // [depth x height x width], but can be bitcast to [depth x height x weight] + // [depth x height x width], but can be bitcast to [depth x height x width] // with "depth" being the most major dimension. Status EmitRowReduction( int64 depth, int64 height, int64 width, HloInstruction* reduce, @@ -182,12 +185,56 @@ class IrEmitterUnnested : public IrEmitter { std::pair> extra_output_gens); + // Returns true if a 0-2-1 tiling algorithm is already used to emit the kernel + // for the hlo instruction. + bool CheckAndEmitHloWithTile021(HloInstruction* hlo); + // Emits a kernel for the hlo instruction using a 0-2-1 tiling algorithm and + // returns the launch dimensions for the kernel. This is a helper to support + // the implementation of CheckAndEmitHloWithTile021. + LaunchDimensions EmitHlo021Tile( + HloInstruction* hlo, + tensorflow::gtl::ArraySlice reduced_output_dims, + tensorflow::gtl::ArraySlice tiled_param_ids); + // Generates the IrArray for each output of hlo and returns the number of + // outputs. + int ConstructIrArrayForOutputs(const HloInstruction& hlo, + std::vector* output_arrays); + // Generates the IrArray for each input of hlo and returns the number of + // inputs. + int ConstructIrArrayForInputs(const HloInstruction& hlo, + std::vector* param_arrays); + // For each output of the `hlo` instruction, constructs the reduced shape for + // the output with the given `reduced_output_dims` and cast the original + // output IrArray element in `output_arrays` to the reduced shape. Returns + // the number of outputs. + int ConstructOutputReducedShapeAndCastOutputIrArrayToShape( + const HloInstruction& hlo, + const std::vector& output_arrays, + tensorflow::gtl::ArraySlice reduced_output_dims, + std::vector* output_reduced_shapes, + std::vector* output_in_reduced_shape_arrays); + // For each input of the `hlo` instruction, checks its value in + // `param_buffers` to find out whether the input has a reduced shape. If the + // input has a reduced shape, constructs the reduced shape for the input and + // casts the original input IrArray in `param_arrays` to the reduced shape. + // Return the total number of inputs. + int ConstructInputReducedShapeAndCastInputIrArrayToShape( + const HloInstruction& hlo, + const std::vector& param_arrays, + const std::vector& param_buffers, + tensorflow::gtl::ArraySlice reduced_output_dims, + std::vector* param_reduced_shapes, + std::vector* param_in_reduced_shape_arrays); + // Returns a KernelThunk that invokes the kernel emitted for `inst`. The // caller needs to make sure `inst` outlives the lifetime of the returned // Thunk object. The kernel implementation will be unrolled if unroll_factor - // is greater than one. - std::unique_ptr BuildKernelThunk(const HloInstruction* inst, - int unroll_factor = 1); + // is greater than one. 'implements_whole_instruction' specifies whether this + // KernelThunk implements the whole 'inst' HloInstruction. In some cases + // 'inst' will be implemented by a sequence of Thunks. + std::unique_ptr BuildKernelThunk( + const HloInstruction* inst, bool implements_whole_instruction, + int unroll_factor = 1); // Returns a FftThunk that calls cuFFT to implement `inst`. std::unique_ptr BuildFftThunk(const HloInstruction* inst); @@ -208,10 +255,14 @@ class IrEmitterUnnested : public IrEmitter { std::unique_ptr BuildDeviceToDeviceCopyThunk( const HloInstruction* inst); - // Returns an InfeedThunk that performs device-to-device memcpy to implement + // Returns an InfeedThunk that performs a host-to-device memcpy to implement // `inst`. std::unique_ptr BuildInfeedThunk(const HloInstruction* inst); + // Returns an OutfeedThunk that performs a device-to-host memcpy to implement + // `inst`. + std::unique_ptr BuildOutfeedThunk(const HloInstruction* inst); + // Returns a WhileThunk that invokes thunk sequences for 'condition' and // 'body' sub-computations of while instruction 'hlo'. std::unique_ptr BuildWhileThunk(const HloInstruction* hlo); diff --git a/tensorflow/compiler/xla/service/gpu/kernel_thunk.cc b/tensorflow/compiler/xla/service/gpu/kernel_thunk.cc index f56c1ce69f11ed79c8be76834269f29de93a9645..e76823ad103dfa5ba61a0d3ba81b2c028dfeb33e 100644 --- a/tensorflow/compiler/xla/service/gpu/kernel_thunk.cc +++ b/tensorflow/compiler/xla/service/gpu/kernel_thunk.cc @@ -17,6 +17,7 @@ limitations under the License. #include "tensorflow/compiler/xla/ptr_util.h" #include "tensorflow/compiler/xla/service/gpu/gpu_executable.h" +#include "tensorflow/compiler/xla/service/gpu/hlo_execution_profiler.h" #include "tensorflow/compiler/xla/types.h" #include "tensorflow/compiler/xla/util.h" #include "tensorflow/core/lib/core/stringpiece.h" @@ -75,7 +76,8 @@ void KernelThunk::SetLaunchDimensions(const LaunchDimensions& launch_dims) { } Status KernelThunk::ExecuteOnStream(const BufferAllocations& buffer_allocations, - se::Stream* stream) { + se::Stream* stream, + HloExecutionProfiler* profiler) { // Load the kernel. se::StreamExecutor* executor = stream->parent(); LaunchDimensions launch_dimensions; @@ -100,6 +102,7 @@ Status KernelThunk::ExecuteOnStream(const BufferAllocations& buffer_allocations, VLOG(3) << " Arg: alloc #" << arg->index() << ": " << buf.opaque() << " (" << buf.size() << "B)"; } + auto op_profiler = profiler->MakeScopedInstructionProfiler(hlo_instruction()); if (!stream->parent()->Launch( stream, se::ThreadDim(launch_dimensions.threads_per_block()), se::BlockDim(launch_dimensions.block_count()), *kernel, diff --git a/tensorflow/compiler/xla/service/gpu/kernel_thunk.h b/tensorflow/compiler/xla/service/gpu/kernel_thunk.h index 7def27e189b66747569344a3dbe5c0c446f903be..d751de50ad6671b3bf88cd4de49a8feb448e13ba 100644 --- a/tensorflow/compiler/xla/service/gpu/kernel_thunk.h +++ b/tensorflow/compiler/xla/service/gpu/kernel_thunk.h @@ -22,6 +22,7 @@ limitations under the License. #include "tensorflow/compiler/xla/service/buffer_assignment.h" #include "tensorflow/compiler/xla/service/gpu/buffer_allocations.h" +#include "tensorflow/compiler/xla/service/gpu/hlo_execution_profiler.h" #include "tensorflow/compiler/xla/service/gpu/partition_assignment.h" #include "tensorflow/compiler/xla/service/gpu/thunk.h" #include "tensorflow/compiler/xla/service/hlo_instruction.h" @@ -62,7 +63,8 @@ class KernelThunk : public Thunk { // Executes the kernel for the thunk on "stream", which must be non-null. Status ExecuteOnStream(const BufferAllocations& buffer_allocations, - se::Stream* stream) override; + se::Stream* stream, + HloExecutionProfiler* profiler) override; private: // Buffers passed to the kernel as arguments. diff --git a/tensorflow/compiler/xla/service/gpu/memset_thunk.cc b/tensorflow/compiler/xla/service/gpu/memset_thunk.cc index d4100a898b5bb9eec382c34932c2db104c9e985b..9fd6cf7157ecd659e7eb1d2c5228eca931ff6a01 100644 --- a/tensorflow/compiler/xla/service/gpu/memset_thunk.cc +++ b/tensorflow/compiler/xla/service/gpu/memset_thunk.cc @@ -14,21 +14,27 @@ limitations under the License. ==============================================================================*/ #include "tensorflow/compiler/xla/service/gpu/memset_thunk.h" + +#include "tensorflow/compiler/xla/service/gpu/hlo_execution_profiler.h" #include "tensorflow/stream_executor/stream_executor.h" namespace xla { namespace gpu { Status MemzeroThunk::ExecuteOnStream( - const BufferAllocations& buffer_allocations, se::Stream* stream) { + const BufferAllocations& buffer_allocations, se::Stream* stream, + HloExecutionProfiler* profiler) { se::DeviceMemoryBase dest_data = buffer_allocations.GetDeviceAddress(dest_); + auto op_profiler = profiler->MakeScopedInstructionProfiler(hlo_instruction()); stream->ThenMemZero(&dest_data, dest_data.size()); return Status::OK(); } Status Memset32BitValueThunk::ExecuteOnStream( - const BufferAllocations& buffer_allocations, se::Stream* stream) { + const BufferAllocations& buffer_allocations, se::Stream* stream, + HloExecutionProfiler* profiler) { se::DeviceMemoryBase dest_data = buffer_allocations.GetDeviceAddress(dest_); + auto op_profiler = profiler->MakeScopedInstructionProfiler(hlo_instruction()); stream->ThenMemset32(&dest_data, value_, dest_data.size()); return Status::OK(); } diff --git a/tensorflow/compiler/xla/service/gpu/memset_thunk.h b/tensorflow/compiler/xla/service/gpu/memset_thunk.h index 51c332d287d139335b356fc66411b5ffaa448b5a..d1fec0bd76b8a80f4a1e1c2e818f248997da7a75 100644 --- a/tensorflow/compiler/xla/service/gpu/memset_thunk.h +++ b/tensorflow/compiler/xla/service/gpu/memset_thunk.h @@ -17,6 +17,7 @@ limitations under the License. #define TENSORFLOW_COMPILER_XLA_SERVICE_GPU_MEMSET_THUNK_H_ #include "tensorflow/compiler/xla/service/buffer_assignment.h" +#include "tensorflow/compiler/xla/service/gpu/hlo_execution_profiler.h" #include "tensorflow/compiler/xla/service/gpu/thunk.h" #include "tensorflow/compiler/xla/service/hlo_instruction.h" #include "tensorflow/compiler/xla/status.h" @@ -36,7 +37,8 @@ class MemzeroThunk : public Thunk { : Thunk(Kind::kMemzero, hlo), dest_(dest) {} Status ExecuteOnStream(const BufferAllocations& buffer_allocations, - se::Stream* stream) override; + se::Stream* stream, + HloExecutionProfiler* profiler) override; private: const BufferAllocation::Slice dest_; @@ -52,7 +54,8 @@ class Memset32BitValueThunk : public Thunk { : Thunk(Kind::kMemset32BitValue, hlo), value_(value), dest_(dest) {} Status ExecuteOnStream(const BufferAllocations& buffer_allocations, - se::Stream* stream) override; + se::Stream* stream, + HloExecutionProfiler* profiler) override; private: uint32 value_; diff --git a/tensorflow/compiler/xla/service/gpu/multi_output_fusion.cc b/tensorflow/compiler/xla/service/gpu/multi_output_fusion.cc index d541776f00ca9c0986fecd272930e5585852f6f3..ea661b3c2cb2c945297ac2098cd1c4009b2e966d 100644 --- a/tensorflow/compiler/xla/service/gpu/multi_output_fusion.cc +++ b/tensorflow/compiler/xla/service/gpu/multi_output_fusion.cc @@ -23,9 +23,11 @@ limitations under the License. #include #include +#include "tensorflow/compiler/xla/service/gpu/ir_emission_utils.h" #include "tensorflow/compiler/xla/service/hlo_instruction.h" #include "tensorflow/compiler/xla/service/hlo_opcode.h" #include "tensorflow/compiler/xla/shape_util.h" +#include "tensorflow/core/lib/gtl/flatset.h" #include "tensorflow/core/platform/types.h" namespace xla { @@ -69,6 +71,7 @@ bool GpuMultiOutputFusion::ShapesCompatibleForFusion(HloInstruction* instr1, // In that case, the operand of the reduce needs to have the same shape // as the other tuple operands, but also we need to compare the output // shapes of the reduces. + // TODO(tjoerg): Allow differences in fp precision. auto* element_instr_1 = get_element_instr(instr1); auto* element_instr_2 = get_element_instr(instr2); if (element_instr_1->opcode() == HloOpcode::kReduce && @@ -82,31 +85,35 @@ bool GpuMultiOutputFusion::ShapesCompatibleForFusion(HloInstruction* instr1, } namespace { -bool IsReduction(HloInstruction* instr) { +bool IsInputFusibleReduction(HloInstruction* instr) { if (instr->IsMultiOutputFusion()) { for (const HloInstruction* operand : instr->fused_expression_root()->operands()) { if (operand->opcode() == HloOpcode::kReduce) { + CHECK(instr->fusion_kind() == HloInstruction::FusionKind::kInput) + << " Reduce multi-output fusion " << instr->ToString() + << " must be an input fusion."; return true; } } return false; } else if (instr->opcode() == HloOpcode::kFusion) { - return instr->fused_expression_root()->opcode() == HloOpcode::kReduce; + // The loop emitter can handle to-vector reduce fusions. Such reduce + // fusions have the fusion kind kLoop rather than kInput. We do not fuse + // to-vector reduce fusions, because the resulting fusions may no longer be + // supported by loop emitter. + return IsReductionToVector(*instr->fused_expression_root()); } else { - return instr->opcode() == HloOpcode::kReduce; + return IsReductionToVector(*instr); } } } // namespace bool GpuMultiOutputFusion::IsFusible(HloInstruction* instr) { // We can fuse reduces and loop fusions. - return IsReduction(instr) || + return IsInputFusibleReduction(instr) || (instr->opcode() == HloOpcode::kFusion && - instr->fusion_kind() == HloInstruction::FusionKind::kLoop && - // TODO(b/110202584): bitcasts make nested fusions, GPU has no support - // for nested fusions. - instr->fused_expression_root()->opcode() != HloOpcode::kBitcast); + instr->fusion_kind() == HloInstruction::FusionKind::kLoop); } int64 GpuMultiOutputFusion::GetProfit(HloInstruction* instr1, @@ -147,5 +154,110 @@ bool GpuMultiOutputFusion::LegalToFuse(HloInstruction* instr1, return instr1->fusion_kind() != HloInstruction::FusionKind::kLoop; } +bool GpuMultiOutputFusion::DoProducerConsumerMultiOutputFusion() { + bool changed = false; + RecomputeReachability(); + + tensorflow::gtl::FlatSet to_fuse; + // Keep a list of the instructions to fuse after making all the fusion + // decisions. We first aggressively add instructions to potential_fusion_list, + // then filter out instructions that will be no longer fusable because of + // reachability change. This avoids recalculating reachability on a large set + // of instructions. + std::vector> + potential_fusion_list; + std::vector> fusion_list; + std::vector instrs_to_update_reachability; + + // For each reduce or reduce multi-output fusion, try to fuse it with loop + // fusions operands. + for (HloInstruction* consumer : computation()->MakeInstructionPostOrder()) { + if (consumer->user_count() == 0) { + continue; + } + if (!IsInputFusibleReduction(consumer)) { + continue; + } + + auto consumer_operands = consumer->operands(); + for (size_t i = 0; i < consumer_operands.size(); ++i) { + HloInstruction* producer = consumer_operands[i]; + if (!producer->IsFusable()) { + continue; + } + const bool is_loop_fusion = + producer->opcode() == HloOpcode::kFusion && + producer->fusion_kind() == HloInstruction::FusionKind::kLoop; + if (!is_loop_fusion) { + continue; + } + if (!ShapesCompatibleForFusion(producer, consumer)) { + continue; + } + // If we have already decided to fuse this producer, skip it. + if (ContainsKey(to_fuse, producer)) { + continue; + } + // Do not fuse a producer if the other operands of the fusion are + // reachable from the producer, this would create a cycle. + if (c_any_of(consumer_operands, [&](HloInstruction* operand) { + return producer != operand && + reachability()->IsReachable(producer, operand); + })) { + break; + } + to_fuse.insert(producer); + potential_fusion_list.emplace_back(producer, consumer); + instrs_to_update_reachability.push_back(producer); + instrs_to_update_reachability.push_back(consumer); + break; + } + } + + // Filter out pairs that will be no longer fusable because of reachability + // change. + for (auto& fusion_pair : potential_fusion_list) { + HloInstruction* producer = fusion_pair.first; + HloInstruction* consumer = fusion_pair.second; + if (!c_any_of(consumer->operands(), [&](HloInstruction* operand) { + return producer != operand && + reachability()->IsReachable(producer, operand); + })) { + UpdateReachability(producer, consumer, instrs_to_update_reachability); + fusion_list.push_back(fusion_pair); + } + } + + for (auto fusions_to_create : fusion_list) { + HloInstruction* producer = fusions_to_create.first; + HloInstruction* consumer = fusions_to_create.second; + if (consumer->opcode() != HloOpcode::kFusion) { + // Fusing with a reduce (fusion) always results in an input fusion. + HloInstruction* input_fusion = + computation()->AddInstruction(HloInstruction::CreateFusion( + consumer->shape(), HloInstruction::FusionKind::kInput, consumer)); + VLOG(2) << "Fuse producer " << producer->name() << " and its consumer " + << consumer->name() << " into " << input_fusion->name(); + TF_CHECK_OK(computation()->ReplaceInstruction(consumer, input_fusion)); + if (producer->opcode() == HloOpcode::kFusion) { + input_fusion->MergeFusionInstructionIntoMultiOutput(producer); + } else { + input_fusion->FuseInstructionIntoMultiOutput(producer); + } + } else { + VLOG(2) << "Fuse producer " << producer->name() << " into its consumer " + << consumer->name(); + + if (producer->opcode() == HloOpcode::kFusion) { + consumer->MergeFusionInstructionIntoMultiOutput(producer); + } else { + consumer->FuseInstructionIntoMultiOutput(producer); + } + } + changed = true; + } + return changed; +} + } // namespace gpu } // namespace xla diff --git a/tensorflow/compiler/xla/service/gpu/multi_output_fusion.h b/tensorflow/compiler/xla/service/gpu/multi_output_fusion.h index 16db0e0f02d5cbf582f0e4236297b3d5407014b3..67ca5d49eee8508e93284b134f8410eb3a89f9ce 100644 --- a/tensorflow/compiler/xla/service/gpu/multi_output_fusion.h +++ b/tensorflow/compiler/xla/service/gpu/multi_output_fusion.h @@ -45,6 +45,9 @@ class GpuMultiOutputFusion : public MultiOutputFusion { // Test if it's legal to fuse instr1 and instr2 into one fusion instruction. bool LegalToFuse(HloInstruction* instr1, HloInstruction* instr2) override; + + // Fuse loop fusions into reduce fusions. + bool DoProducerConsumerMultiOutputFusion() override; }; } // namespace gpu diff --git a/tensorflow/compiler/xla/service/gpu/multi_output_fusion_test.cc b/tensorflow/compiler/xla/service/gpu/multi_output_fusion_test.cc index 5e7ceb7976b5d1957f706c12ec255e93991344b8..979ea79243818c398b1b130254a41c95ced51830 100644 --- a/tensorflow/compiler/xla/service/gpu/multi_output_fusion_test.cc +++ b/tensorflow/compiler/xla/service/gpu/multi_output_fusion_test.cc @@ -255,5 +255,99 @@ TEST_F(InstructionFusionTest, MultiOutputFusionTwoLoops) { op::Tuple(op::Multiply(), op::Divide())); } +TEST_F(InstructionFusionTest, ProducerConsumerFusionLoopFusionAndReduce) { + auto module = ParseHloString(tensorflow::strings::StrCat(kModulePrefix, R"( + fused_add { + p0.1 = f32[2,2,2]{2,1,0} parameter(0) + p1.1 = f32[2,2,2]{2,1,0} parameter(1) + ROOT add = f32[2,2,2]{2,1,0} add(p0.1, p1.1) + } + + ENTRY reduce { + p0 = f32[2,2,2]{2,1,0} parameter(0) + p1 = f32[2,2,2]{2,1,0} parameter(1) + c0 = f32[] constant(0) + add = f32[2,2,2]{2,1,0} fusion(p0, p1), kind=kLoop, calls=fused_add + reduce = f32[2,2]{1,0} reduce(add, c0), dimensions={2}, to_apply=scalar_add_computation + ROOT root = (f32[2,2]{1,0}, f32[2,2,2]{2,1,0}) tuple(reduce, add) + })")) + .ValueOrDie(); + ASSERT_TRUE(GpuMultiOutputFusion().Run(module.get()).ValueOrDie()); + SCOPED_TRACE(module->ToString()); + const HloInstruction* root = module->entry_computation()->root_instruction(); + EXPECT_THAT(root, op::Tuple(op::GetTupleElement(), op::GetTupleElement())); + const HloInstruction* fusion = root->operand(0)->operand(0); + ASSERT_TRUE(fusion->IsMultiOutputFusion()); + EXPECT_THAT(fusion->fused_expression_root(), + op::Tuple(op::Reduce(), op::Add())); +} + +TEST_F(InstructionFusionTest, ProducerConsumerFusionLoopFusionAndReduceFusion) { + auto module = ParseHloString(tensorflow::strings::StrCat(kModulePrefix, R"( + fused_select { + p1.1 = f32[2,2,2]{2,1,0} parameter(1) + c0 = f32[] constant(0) + broadcast = f32[2,2,2]{2,1,0} broadcast(f32[] c0), dimensions={} + greater-than = pred[2,2,2]{2,1,0} greater-than(f32[2,2,2]{2,1,0} p1.1, f32[2,2,2]{2,1,0} broadcast) + p0.1 = f32[2,2,2]{2,1,0} parameter(0) + ROOT select = f32[2,2,2]{2,1,0} select(pred[2,2,2]{2,1,0} greater-than, f32[2,2,2]{2,1,0} p0.1, f32[2,2,2]{2,1,0} broadcast) + } + + fused_reduce { + p0.2 = f32[2,2,2]{2,1,0} parameter(0) + c1 = f32[] constant(0) + r1 = f32[2,2]{1,0} reduce(p0.2, c1), dimensions={2}, to_apply=scalar_add_computation + mul = f32[2,2,2]{2,1,0} multiply(p0.2, p0.2) + r2 = f32[2,2]{1,0} reduce(mul, c1), dimensions={2}, to_apply=scalar_add_computation + ROOT tuple = (f32[2,2]{1,0}, f32[2,2]{1,0}) tuple(r1, r2) + } + + ENTRY reduce { + p0 = f32[2,2,2]{2,1,0} parameter(0) + p1 = f32[2,2,2]{2,1,0} parameter(1) + select = f32[2,2,2]{2,1,0} fusion(p0, p1), kind=kLoop, calls=fused_select + fusion = (f32[2,2]{1,0}, f32[2,2]{1,0}) fusion(select), kind=kInput, calls=fused_reduce + gte0 = f32[2,2]{1,0} get-tuple-element(fusion), index=0 + gte1 = f32[2,2]{1,0} get-tuple-element(fusion), index=1 + ROOT root = (f32[2,2]{1,0}, f32[2,2]{1,0}, f32[2,2,2]{2,1,0}) tuple(gte1, gte1, select) + })")) + .ValueOrDie(); + ASSERT_TRUE(GpuMultiOutputFusion().Run(module.get()).ValueOrDie()); + SCOPED_TRACE(module->ToString()); + const HloInstruction* root = module->entry_computation()->root_instruction(); + EXPECT_THAT(root, op::Tuple(op::GetTupleElement(), op::GetTupleElement(), + op::GetTupleElement())); + const HloInstruction* fusion = root->operand(0)->operand(0); + ASSERT_TRUE(fusion->IsMultiOutputFusion()); + EXPECT_THAT(fusion->fused_expression_root(), + op::Tuple(op::Reduce(), op::Reduce(), op::Select())); +} + +TEST_F(InstructionFusionTest, ProducerConsumerFusionDoNotFuseLoopReduceFusion) { + auto module = ParseHloString(tensorflow::strings::StrCat(kModulePrefix, R"( + fused_element_wise { + p0.1 = f32[2,2,2]{2,1,0} parameter(0) + p1.1 = f32[2,2,2]{2,1,0} parameter(1) + ROOT root = f32[2,2,2]{2,1,0} add(p0.1, p1.1) + } + + fused_reduce { + p0.2 = f32[2,2,2]{2,1,0} parameter(0) + mul = f32[2,2,2]{2,1,0} multiply(f32[2,2,2]{2,1,0} p0.2, f32[2,2,2]{2,1,0} p0.2) + c1 = f32[] constant(0) + ROOT reduce = f32[2,2]{1,0} reduce(f32[2,2,2]{2,1,0} mul, f32[] c1), dimensions={1}, to_apply=scalar_add_computation + } + + ENTRY reduce { + p0 = f32[2,2,2]{2,1,0} parameter(0) + p1 = f32[2,2,2]{2,1,0} parameter(1) + element_wise = f32[2,2,2]{2,1,0} fusion(p0, p1), kind=kLoop, calls=fused_element_wise + fusion = (f32[2,2]{1,0}, f32[2,2]{1,0}) fusion(element_wise), kind=kLoop, calls=fused_reduce + ROOT root = (f32[2,2]{1,0}, f32[2,2,2]{2,1,0}) tuple(fusion, element_wise) + })")) + .ValueOrDie(); + ASSERT_FALSE(GpuMultiOutputFusion().Run(module.get()).ValueOrDie()); +} + } // namespace gpu } // namespace xla diff --git a/tensorflow/compiler/xla/service/gpu/outfeed_manager.cc b/tensorflow/compiler/xla/service/gpu/outfeed_manager.cc new file mode 100644 index 0000000000000000000000000000000000000000..4aaf0c9e142106a0e74f319d71dad4c4c96d3f08 --- /dev/null +++ b/tensorflow/compiler/xla/service/gpu/outfeed_manager.cc @@ -0,0 +1,32 @@ +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow/compiler/xla/service/gpu/outfeed_manager.h" + +#include "tensorflow/compiler/xla/map_util.h" +#include "tensorflow/compiler/xla/ptr_util.h" +#include "tensorflow/compiler/xla/shape_util.h" +#include "tensorflow/core/platform/logging.h" + +namespace xla { +namespace gpu { + +OutfeedManager* GetOrCreateOutfeedManager() { + static auto* manager = new OutfeedManager; + return manager; +} + +} // namespace gpu +} // namespace xla diff --git a/tensorflow/compiler/xla/service/gpu/outfeed_manager.h b/tensorflow/compiler/xla/service/gpu/outfeed_manager.h new file mode 100644 index 0000000000000000000000000000000000000000..a752eb70119b00e8cca7ddce26da7730ef5db8cb --- /dev/null +++ b/tensorflow/compiler/xla/service/gpu/outfeed_manager.h @@ -0,0 +1,69 @@ +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#ifndef TENSORFLOW_COMPILER_XLA_SERVICE_GPU_OUTFEED_MANAGER_H_ +#define TENSORFLOW_COMPILER_XLA_SERVICE_GPU_OUTFEED_MANAGER_H_ + +#include "tensorflow/compiler/xla/literal.h" +#include "tensorflow/compiler/xla/service/gpu/xfeed_queue.h" +#include "tensorflow/compiler/xla/shape_tree.h" +#include "tensorflow/core/platform/mutex.h" +#include "tensorflow/core/platform/notification.h" + +namespace xla { +namespace gpu { + +// TODO(b/30467474) Once GPU outfeed implementation settles, consider +// folding back the cpu and gpu outfeed implementations into a generic +// one if possible. + +// Defines a buffer holding the destination for an outfeed in host memory and a +// notification when that triggers when the transfer is done. +class OutfeedBuffer { + public: + OutfeedBuffer(int64 length) : length_(length) {} + + // Waits for the device transfer to be finished. + std::unique_ptr WaitUntilAvailable() { + done_.WaitForNotification(); + return std::move(destination_); + } + + int64 length() const { return length_; } + void set_destination(std::unique_ptr destination) { + destination_ = std::move(destination); + } + Literal* destination() { return destination_.get(); } + + // Callback to signal that this buffer is consumed. + void Done() { done_.Notify(); } + + private: + std::unique_ptr destination_; + const int64 length_; + tensorflow::Notification done_; +}; + +// Manages a thread-safe queue of buffers. The buffers are supposed to be +// produced by the transfer manager and consumed by the device. +using OutfeedManager = XfeedQueue>*>; + +// Singleton creator-or-accessor: Returns the GPU outfeed manager. +OutfeedManager* GetOrCreateOutfeedManager(); + +} // namespace gpu +} // namespace xla + +#endif // TENSORFLOW_COMPILER_XLA_SERVICE_GPU_OUTFEED_MANAGER_H_ diff --git a/tensorflow/compiler/xla/service/gpu/outfeed_thunk.cc b/tensorflow/compiler/xla/service/gpu/outfeed_thunk.cc new file mode 100644 index 0000000000000000000000000000000000000000..7986e63f43ee508370f94fdb9057b91bfe4add18 --- /dev/null +++ b/tensorflow/compiler/xla/service/gpu/outfeed_thunk.cc @@ -0,0 +1,111 @@ +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow/compiler/xla/service/gpu/outfeed_thunk.h" +#include "tensorflow/compiler/xla/literal.h" +#include "tensorflow/compiler/xla/service/gpu/hlo_execution_profiler.h" +#include "tensorflow/compiler/xla/service/gpu/outfeed_manager.h" +#include "tensorflow/compiler/xla/util.h" +#include "tensorflow/core/platform/stream_executor_no_cuda.h" + +namespace xla { +namespace gpu { + +OutfeedThunk::OutfeedThunk(ShapeTree outfeed_slices, + const HloInstruction* hlo_instruction) + : Thunk(Kind::kOutfeed, hlo_instruction), + outfeed_slices_(std::move(outfeed_slices)) {} + +Status OutfeedThunk::ExecuteOnStream( + const BufferAllocations& buffer_allocations, se::Stream* stream, + HloExecutionProfiler* profiler) { + VLOG(2) << "Outfeeding from GPU: " << hlo_instruction()->ToString(); + + auto op_profiler = profiler->MakeScopedInstructionProfiler(hlo_instruction()); + OutfeedManager* outfeed_manager = GetOrCreateOutfeedManager(); + ShapeTree>* outfeed_buffers = + outfeed_manager->BlockingGetNextDestination(); + + // Nothing to be done for empty tuples. + if (ShapeUtil::IsEmptyTuple(hlo_instruction()->operand(0)->shape())) { + return Status::OK(); + } + CHECK(ShapeUtil::Compatible(hlo_instruction()->operand(0)->shape(), + outfeed_buffers->shape())); + + TF_RETURN_IF_ERROR(outfeed_buffers->ForEachMutableElementWithStatus( + [&](const ShapeIndex& index, std::unique_ptr* buffer) { + if (!*buffer) { // Tuple pointers. + return Status::OK(); + } + // Allocate storage for the literal data. + const Shape& shape = + ShapeUtil::GetSubshape(outfeed_buffers->shape(), index); + (*buffer)->set_destination(Literal::CreateFromShape(shape)); + + BufferAllocation::Slice slice = outfeed_slices_.element(index); + se::DeviceMemoryBase data_address; + if (slice.allocation()) { + // If we have a static allocation, read it from there. This avoids + // synchronizing the host and device just to read a pointer. + data_address = buffer_allocations.GetDeviceAddress(slice); + } else { + // Otherwise we have to read the tuple pointer first. + CHECK(!index.empty()); + // Copy the parent buffer to the host. + BufferAllocation::Slice tuple_slice = + outfeed_slices_.element(ShapeIndexView(index).ConsumeFront()); + if (!tuple_slice.allocation()) { + return Unimplemented( + "Nested dynamic tuples are not supported on GPU"); + } + se::DeviceMemoryBase tuple_address = + buffer_allocations.GetDeviceAddress(tuple_slice); + CHECK(tuple_slice.size() % sizeof(void*) == 0) + << "Tuple size must be a multiple of pointer size"; + std::vector tuple_element_buffer_addresses(tuple_slice.size() / + sizeof(void*)); + stream->ThenMemcpy(tuple_element_buffer_addresses.data(), + tuple_address, tuple_slice.size()); + TF_RETURN_IF_ERROR(stream->BlockHostUntilDone()); + // The data address is specified by the element of the tuple pointer + // buffer. + data_address = + se::DeviceMemoryBase(tuple_element_buffer_addresses[index.back()], + (*buffer)->length()); + } + + // TODO(b/111309141): Run this on a separate stream so it doesn't block + // the GPU from doing work during the transfer. This could be handled by + // making StreamAssignment do something intelligent with outfeed thunks. + stream + ->ThenMemcpy((*buffer)->destination()->untyped_data(), data_address, + (*buffer)->length()) + .ThenDoHostCallback([buffer]() { (*buffer)->Done(); }); + return Status::OK(); + })); + + Status block_status = stream->BlockHostUntilDone(); + if (!block_status.ok()) { + return InternalError("Failed to complete data transfer on stream %p: %s", + stream, block_status.error_message().c_str()); + } + + VLOG(2) << "Outfeeding from GPU complete"; + return Status::OK(); +} + +} // namespace gpu +} // namespace xla diff --git a/tensorflow/compiler/xla/service/gpu/outfeed_thunk.h b/tensorflow/compiler/xla/service/gpu/outfeed_thunk.h new file mode 100644 index 0000000000000000000000000000000000000000..8ed89f05f0c5bb2e3893e695d413bac3b231112d --- /dev/null +++ b/tensorflow/compiler/xla/service/gpu/outfeed_thunk.h @@ -0,0 +1,52 @@ +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#ifndef TENSORFLOW_COMPILER_XLA_SERVICE_GPU_OUTFEED_THUNK_H_ +#define TENSORFLOW_COMPILER_XLA_SERVICE_GPU_OUTFEED_THUNK_H_ + +#include "tensorflow/compiler/xla/service/gpu/buffer_allocations.h" +#include "tensorflow/compiler/xla/service/gpu/hlo_execution_profiler.h" +#include "tensorflow/compiler/xla/service/gpu/thunk.h" +#include "tensorflow/compiler/xla/service/hlo_instruction.h" +#include "tensorflow/core/platform/stream_executor_no_cuda.h" + +namespace xla { +namespace gpu { + +// A thunk that outfeeds data. Data must be already resident on the host. This +// thunk performs a host to device copy from the buffer allocated for the +// outfeed op to the host location. +class OutfeedThunk : public Thunk { + public: + // Constructs a OutfeedThunk that copies data to the host-side + // outfeed queue from the buffers in the given shape tree. + OutfeedThunk(ShapeTree outfeed_slices, + const HloInstruction* hlo_instruction); + + OutfeedThunk(const OutfeedThunk&) = delete; + OutfeedThunk& operator=(const OutfeedThunk&) = delete; + + Status ExecuteOnStream(const BufferAllocations& buffer_allocations, + se::Stream* stream, + HloExecutionProfiler* profiler) override; + + private: + const ShapeTree outfeed_slices_; +}; + +} // namespace gpu +} // namespace xla + +#endif // TENSORFLOW_COMPILER_XLA_SERVICE_GPU_OUTFEED_THUNK_H_ diff --git a/tensorflow/compiler/xla/service/gpu/pad_insertion.cc b/tensorflow/compiler/xla/service/gpu/pad_insertion.cc index c8f0d4185c63c5bafca6f30acab31cbe8e987277..b22040eee167e784bed58dbc0d0ad2ae042037f3 100644 --- a/tensorflow/compiler/xla/service/gpu/pad_insertion.cc +++ b/tensorflow/compiler/xla/service/gpu/pad_insertion.cc @@ -15,6 +15,7 @@ limitations under the License. #include "tensorflow/compiler/xla/service/gpu/pad_insertion.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/literal_util.h" #include "tensorflow/compiler/xla/service/gpu/ir_emission_utils.h" #include "tensorflow/compiler/xla/service/hlo_creation_utils.h" @@ -68,7 +69,7 @@ HloInstruction* MaybePaddedAndSlicedInput( PrimitiveType element_type = input->shape().element_type(); HloInstruction* padding = computation->AddInstruction(HloInstruction::CreateConstant( - MakeUnique(Literal::Zero(element_type)))); + MakeUnique(LiteralUtil::Zero(element_type)))); input = MakePadHlo(input, padding, padding_config).ValueOrDie(); } @@ -125,7 +126,7 @@ HloInstruction* MaybePaddedKernel(const Window& conv_window, PrimitiveType element_type = kernel->shape().element_type(); HloInstruction* padding = computation->AddInstruction(HloInstruction::CreateConstant( - MakeUnique(Literal::Zero(element_type)))); + MakeUnique(LiteralUtil::Zero(element_type)))); return MakePadHlo(kernel, padding, padding_config).ValueOrDie(); } } // namespace @@ -234,9 +235,9 @@ bool PadInsertion::CanonicalizeBackwardFilterConvolution( // Create a new backward convolution replacing the old one. HloComputation* computation = backward_conv->parent(); HloInstruction* output = backward_conv->mutable_operand(1); - HloInstruction* padding = - computation->AddInstruction(HloInstruction::CreateConstant( - MakeUnique(Literal::Zero(input->shape().element_type())))); + HloInstruction* padding = computation->AddInstruction( + HloInstruction::CreateConstant(MakeUnique( + LiteralUtil::Zero(input->shape().element_type())))); HloInstruction* padded_input = MakePadHlo(input, padding, input_padding_config).ValueOrDie(); diff --git a/tensorflow/compiler/xla/service/gpu/sequential_thunk.cc b/tensorflow/compiler/xla/service/gpu/sequential_thunk.cc index 88cb10883e97ae663dc492ad088e6daf9133d7f5..84285be70a4ba94101040a639c39b3eaecbb5bb3 100644 --- a/tensorflow/compiler/xla/service/gpu/sequential_thunk.cc +++ b/tensorflow/compiler/xla/service/gpu/sequential_thunk.cc @@ -15,6 +15,7 @@ limitations under the License. #include "tensorflow/compiler/xla/service/gpu/sequential_thunk.h" +#include "tensorflow/compiler/xla/service/gpu/hlo_execution_profiler.h" #include "tensorflow/core/lib/core/errors.h" namespace xla { @@ -33,9 +34,12 @@ Status SequentialThunk::Initialize(const GpuExecutable& executable, } Status SequentialThunk::ExecuteOnStream( - const BufferAllocations& buffer_allocations, se::Stream* stream) { + const BufferAllocations& buffer_allocations, se::Stream* stream, + HloExecutionProfiler* profiler) { + auto op_profiler = profiler->MakeScopedInstructionProfiler(hlo_instruction()); for (const auto& thunk : thunks_) { - TF_RETURN_IF_ERROR(thunk->ExecuteOnStream(buffer_allocations, stream)); + TF_RETURN_IF_ERROR( + thunk->ExecuteOnStream(buffer_allocations, stream, profiler)); } return Status::OK(); } diff --git a/tensorflow/compiler/xla/service/gpu/sequential_thunk.h b/tensorflow/compiler/xla/service/gpu/sequential_thunk.h index 135f79e413dfaa27f2f2264e0daa3beb3c305e0f..3c4de1d1a6c912ba31f56c29b10ca004d1e56da6 100644 --- a/tensorflow/compiler/xla/service/gpu/sequential_thunk.h +++ b/tensorflow/compiler/xla/service/gpu/sequential_thunk.h @@ -19,6 +19,7 @@ limitations under the License. #include #include "tensorflow/compiler/xla/service/gpu/buffer_allocations.h" +#include "tensorflow/compiler/xla/service/gpu/hlo_execution_profiler.h" #include "tensorflow/compiler/xla/service/gpu/thunk.h" #include "tensorflow/compiler/xla/service/hlo_instruction.h" #include "tensorflow/core/platform/stream_executor_no_cuda.h" @@ -41,7 +42,8 @@ class SequentialThunk : public Thunk { Status Initialize(const GpuExecutable& executable, se::StreamExecutor* executor) override; Status ExecuteOnStream(const BufferAllocations& buffer_allocations, - se::Stream* stream) override; + se::Stream* stream, + HloExecutionProfiler* profiler) override; private: // The list of sub-thunks. diff --git a/tensorflow/compiler/xla/service/gpu/stream_executor_util.h b/tensorflow/compiler/xla/service/gpu/stream_executor_util.h index 8218f4fd11d3978d0ecc53fc15e287aea4b69ec3..39a6a38d001f502b2abb8de6efe2ce623b478c71 100644 --- a/tensorflow/compiler/xla/service/gpu/stream_executor_util.h +++ b/tensorflow/compiler/xla/service/gpu/stream_executor_util.h @@ -16,6 +16,7 @@ limitations under the License. #ifndef TENSORFLOW_COMPILER_XLA_SERVICE_GPU_STREAM_EXECUTOR_UTIL_H_ #define TENSORFLOW_COMPILER_XLA_SERVICE_GPU_STREAM_EXECUTOR_UTIL_H_ +#include "tensorflow/compiler/xla/statusor.h" #include "tensorflow/compiler/xla/xla_data.pb.h" #include "tensorflow/core/platform/stream_executor_no_cuda.h" diff --git a/tensorflow/compiler/xla/service/gpu/thunk.h b/tensorflow/compiler/xla/service/gpu/thunk.h index 931c0bffab850362dbd2df975657dd47d9cbd3ae..99a1a0eae9b60bca8ae8443c98a19fc62d4dddfd 100644 --- a/tensorflow/compiler/xla/service/gpu/thunk.h +++ b/tensorflow/compiler/xla/service/gpu/thunk.h @@ -20,6 +20,7 @@ limitations under the License. #include #include "tensorflow/compiler/xla/service/gpu/buffer_allocations.h" +#include "tensorflow/compiler/xla/service/gpu/hlo_execution_profiler.h" #include "tensorflow/compiler/xla/service/hlo_instruction.h" #include "tensorflow/core/lib/core/status.h" #include "tensorflow/core/platform/stream_executor_no_cuda.h" @@ -53,6 +54,7 @@ class Thunk { kKernel, kMemset32BitValue, kMemzero, + kOutfeed, kSequential, kTuple, kWhile, @@ -94,11 +96,12 @@ class Thunk { // Execute the kernel for the thunk on the given stream. This method must be // called after Initialize and can be called multiple times over Thunk's - // lifetime. Stream argument must be non-null. + // lifetime. 'stream' and 'profiler' must be non-null. // // Precondition: Initialize(stream->parent()) has been called. virtual Status ExecuteOnStream(const BufferAllocations& buffer_allocations, - se::Stream* stream) = 0; + se::Stream* stream, + HloExecutionProfiler* profiler) = 0; private: Kind kind_; diff --git a/tensorflow/compiler/xla/service/gpu/tuple_thunk.cc b/tensorflow/compiler/xla/service/gpu/tuple_thunk.cc index 97cb04c38fbf18e516857f5269c984696ca204c3..a10e40451c1db01ce73db7b56a3a0599769fa49b 100644 --- a/tensorflow/compiler/xla/service/gpu/tuple_thunk.cc +++ b/tensorflow/compiler/xla/service/gpu/tuple_thunk.cc @@ -15,13 +15,15 @@ limitations under the License. #include "tensorflow/compiler/xla/service/gpu/tuple_thunk.h" +#include "tensorflow/compiler/xla/service/gpu/hlo_execution_profiler.h" #include "tensorflow/compiler/xla/util.h" namespace xla { namespace gpu { Status TupleThunk::ExecuteOnStream(const BufferAllocations& buffer_allocations, - se::Stream* stream) { + se::Stream* stream, + HloExecutionProfiler* profiler) { std::vector tuple_element_buffer_addresses; for (BufferAllocation::Slice tuple_element_buffer : tuple_element_buffers_) { tuple_element_buffer_addresses.push_back( @@ -31,6 +33,7 @@ Status TupleThunk::ExecuteOnStream(const BufferAllocations& buffer_allocations, buffer_allocations.GetDeviceAddress(dest_buffer_)); auto host_size = tuple_element_buffer_addresses.size() * sizeof(void*); + auto op_profiler = profiler->MakeScopedInstructionProfiler(hlo_instruction()); if (!stream ->ThenMemcpy(&dest_buffer_address, tuple_element_buffer_addresses.data(), host_size) diff --git a/tensorflow/compiler/xla/service/gpu/tuple_thunk.h b/tensorflow/compiler/xla/service/gpu/tuple_thunk.h index 951f809b51937c97a6e7de0345ec58a8b66a4242..2d5735d6c40ccd26f0e527f1a02403910db4c812 100644 --- a/tensorflow/compiler/xla/service/gpu/tuple_thunk.h +++ b/tensorflow/compiler/xla/service/gpu/tuple_thunk.h @@ -20,6 +20,7 @@ limitations under the License. #include "tensorflow/compiler/xla/service/buffer_assignment.h" #include "tensorflow/compiler/xla/service/gpu/gpu_executable.h" +#include "tensorflow/compiler/xla/service/gpu/hlo_execution_profiler.h" #include "tensorflow/compiler/xla/service/gpu/thunk.h" #include "tensorflow/core/lib/core/status.h" #include "tensorflow/core/lib/gtl/array_slice.h" @@ -46,7 +47,8 @@ class TupleThunk : public Thunk { TupleThunk& operator=(const TupleThunk&) = delete; Status ExecuteOnStream(const BufferAllocations& buffer_allocations, - se::Stream* stream) override; + se::Stream* stream, + HloExecutionProfiler* profiler) override; private: const std::vector tuple_element_buffers_; diff --git a/tensorflow/compiler/xla/service/gpu/while_thunk.cc b/tensorflow/compiler/xla/service/gpu/while_thunk.cc index 30b9640c4c75dae61e9a90da5fb10e9d4a90cd26..1315a4183a98d6ea9ed4c82d4c22e77c2109ec83 100644 --- a/tensorflow/compiler/xla/service/gpu/while_thunk.cc +++ b/tensorflow/compiler/xla/service/gpu/while_thunk.cc @@ -16,6 +16,7 @@ limitations under the License. #include "tensorflow/compiler/xla/service/gpu/while_thunk.h" #include "tensorflow/compiler/xla/ptr_util.h" +#include "tensorflow/compiler/xla/service/gpu/hlo_execution_profiler.h" #include "tensorflow/compiler/xla/util.h" #include "tensorflow/core/lib/core/errors.h" @@ -29,10 +30,14 @@ WhileThunk::WhileThunk( const HloInstruction* hlo) : Thunk(Kind::kWhile, hlo), condition_result_buffer_index_(condition_result_buffer_index), + // Pass nullptr as the HloInstruction* to the condition_thunk_sequence_ + // and body_thunk_sequence_ constructors because these SequentialThunks + // are logically "part of" this WhileThunk, and shouldn't be profiled + // separately from it. condition_thunk_sequence_(MakeUnique( - std::move(*condition_thunk_sequence), hlo)), - body_thunk_sequence_( - MakeUnique(std::move(*body_thunk_sequence), hlo)) {} + std::move(*condition_thunk_sequence), nullptr)), + body_thunk_sequence_(MakeUnique( + std::move(*body_thunk_sequence), nullptr)) {} Status WhileThunk::Initialize(const GpuExecutable& executable, se::StreamExecutor* executor) { @@ -43,14 +48,18 @@ Status WhileThunk::Initialize(const GpuExecutable& executable, } Status WhileThunk::ExecuteOnStream(const BufferAllocations& buffer_allocations, - se::Stream* stream) { + se::Stream* stream, + HloExecutionProfiler* profiler) { se::DeviceMemoryBase condition_result_data = buffer_allocations.GetDeviceAddress(condition_result_buffer_index_); + auto op_profiler = profiler->MakeScopedInstructionProfiler(hlo_instruction()); while (true) { // Invoke thunk sequence for while 'condition' computation. - TF_RETURN_IF_ERROR( - condition_thunk_sequence_->ExecuteOnStream(buffer_allocations, stream)); + profiler->StartHloComputation(); + TF_RETURN_IF_ERROR(condition_thunk_sequence_->ExecuteOnStream( + buffer_allocations, stream, profiler)); + profiler->FinishHloComputation(hlo_instruction()->while_condition()); // Copy the result of condition computation and break the loop if 'false'. bool condition_result; @@ -66,9 +75,14 @@ Status WhileThunk::ExecuteOnStream(const BufferAllocations& buffer_allocations, break; } - // Invoke thunk sequence for while 'body' computation. - TF_RETURN_IF_ERROR( - body_thunk_sequence_->ExecuteOnStream(buffer_allocations, stream)); + // We measure the time of one execution of the while body computation. The + // while body may be executed more than once, the last measurement "wins". + profiler->StartHloComputation(); + // Invoke thunk sequence for while 'body' computation, and pass on + // 'profiler' to measure the timing of the thunks in 'body_thunk_sequence_'. + TF_RETURN_IF_ERROR(body_thunk_sequence_->ExecuteOnStream(buffer_allocations, + stream, profiler)); + profiler->FinishHloComputation(hlo_instruction()->while_body()); } return Status::OK(); } diff --git a/tensorflow/compiler/xla/service/gpu/while_thunk.h b/tensorflow/compiler/xla/service/gpu/while_thunk.h index 22176685a92df9c95b10f755b209309843c0fa3a..9270f95ee67cf0bd3ab8082452a9d8703cb4304e 100644 --- a/tensorflow/compiler/xla/service/gpu/while_thunk.h +++ b/tensorflow/compiler/xla/service/gpu/while_thunk.h @@ -19,6 +19,7 @@ limitations under the License. #include #include "tensorflow/compiler/xla/service/gpu/buffer_allocations.h" +#include "tensorflow/compiler/xla/service/gpu/hlo_execution_profiler.h" #include "tensorflow/compiler/xla/service/gpu/sequential_thunk.h" #include "tensorflow/compiler/xla/service/gpu/thunk.h" #include "tensorflow/compiler/xla/service/hlo_instruction.h" @@ -48,7 +49,8 @@ class WhileThunk : public Thunk { Status Initialize(const GpuExecutable& executable, se::StreamExecutor* executor) override; Status ExecuteOnStream(const BufferAllocations& buffer_allocations, - se::Stream* stream) override; + se::Stream* stream, + HloExecutionProfiler* profiler) override; private: const BufferAllocation::Slice condition_result_buffer_index_; diff --git a/tensorflow/compiler/xla/service/gpu/while_transformer.cc b/tensorflow/compiler/xla/service/gpu/while_transformer.cc index 7749201cbceece216a2db2569936949eb7de5125..c5321df6c466fcb3816fb2aedad65b7c3811cb37 100644 --- a/tensorflow/compiler/xla/service/gpu/while_transformer.cc +++ b/tensorflow/compiler/xla/service/gpu/while_transformer.cc @@ -18,7 +18,7 @@ limitations under the License. #include #include -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/service/hlo_computation.h" #include "tensorflow/compiler/xla/shape_util.h" #include "tensorflow/compiler/xla/status_macros.h" diff --git a/tensorflow/compiler/xla/service/gpu/while_transformer_test.cc b/tensorflow/compiler/xla/service/gpu/while_transformer_test.cc index 2f290f61bd527e9827472a78256f015e066e44be..dbc8442ed2785a112b674632689256c01282156b 100644 --- a/tensorflow/compiler/xla/service/gpu/while_transformer_test.cc +++ b/tensorflow/compiler/xla/service/gpu/while_transformer_test.cc @@ -42,7 +42,7 @@ class WhileTransformerTest : public HloTestBase { const int64 tuple_index, const int64 limit) { auto builder = HloComputation::Builder(TestName() + ".Condition"); auto limit_const = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(limit))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(limit))); auto loop_state = builder.AddInstruction(HloInstruction::CreateParameter( 0, GetLoopStateShape(tuple_index), "loop_state")); auto induction_variable = @@ -65,8 +65,8 @@ class WhileTransformerTest : public HloTestBase { auto induction_variable = builder.AddInstruction(HloInstruction::CreateGetTupleElement( induction_variable_shape_, loop_state, ind_var_tuple_index)); - auto inc = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(increment))); + auto inc = builder.AddInstruction(HloInstruction::CreateConstant( + LiteralUtil::CreateR0(increment))); auto add0 = builder.AddInstruction(HloInstruction::CreateBinary( induction_variable->shape(), HloOpcode::kAdd, induction_variable, inc)); // Update data GTE(data_tuple_index). @@ -89,10 +89,12 @@ class WhileTransformerTest : public HloTestBase { const int64 ind_var_tuple_index, const int64 ind_var_init) { auto builder = HloComputation::Builder(TestName() + ".While"); - auto induction_var_init = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(ind_var_init))); - auto data_init = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR1({0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f}))); + auto induction_var_init = + builder.AddInstruction(HloInstruction::CreateConstant( + LiteralUtil::CreateR0(ind_var_init))); + auto data_init = builder.AddInstruction( + HloInstruction::CreateConstant(LiteralUtil::CreateR1( + {0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f, 0.f}))); auto loop_state_init = ind_var_tuple_index == 0 ? builder.AddInstruction( diff --git a/tensorflow/compiler/xla/service/gpu/xfeed_queue.h b/tensorflow/compiler/xla/service/gpu/xfeed_queue.h new file mode 100644 index 0000000000000000000000000000000000000000..737c7eb02532dd6e9385c58684e46e7aa0a424fb --- /dev/null +++ b/tensorflow/compiler/xla/service/gpu/xfeed_queue.h @@ -0,0 +1,89 @@ +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#ifndef TENSORFLOW_COMPILER_XLA_SERVICE_GPU_XFEED_QUEUE_H_ +#define TENSORFLOW_COMPILER_XLA_SERVICE_GPU_XFEED_QUEUE_H_ + +#include +#include + +#include "tensorflow/core/platform/mutex.h" +#include "tensorflow/core/platform/notification.h" +#include "tensorflow/core/platform/thread_annotations.h" + +namespace xla { +namespace gpu { + +// TODO(b/30467474) Once GPU outfeed implementation settles, consider +// folding back the cpu and gpu outfeed implementations into a generic +// one if possible. + +// Manages a thread-safe queue of buffers. +template +class XfeedQueue { + public: + // Adds a tree of buffers to the queue. The individual buffers correspond to + // the elements of a tuple and may be nullptr if the buffer is a tuple index + // buffer. + void EnqueueDestination(BufferType buffers) { + tensorflow::mutex_lock l(mu_); + enqueued_buffers_.push_back(std::move(buffers)); + cv_.notify_one(); + } + + // Blocks until the queue is non-empty, then returns the buffer at the head of + // the queue. + BufferType BlockingGetNextDestination() { + bool became_empty; + BufferType current_buffer; + { + tensorflow::mutex_lock l(mu_); + while (enqueued_buffers_.empty()) { + cv_.wait(l); + } + current_buffer = std::move(enqueued_buffers_.front()); + enqueued_buffers_.pop_front(); + became_empty = enqueued_buffers_.empty(); + } + if (became_empty) { + for (const auto& callback : on_empty_callbacks_) { + callback(); + } + } + return current_buffer; + } + + void RegisterOnEmptyCallback(std::function callback) { + on_empty_callbacks_.push_back(std::move(callback)); + } + + private: + tensorflow::mutex mu_; + + // Condition variable that is signaled every time a buffer is enqueued. + tensorflow::condition_variable cv_; + + // The queue of trees of buffers. Buffer* queue contents are not owned. + std::deque enqueued_buffers_ GUARDED_BY(mu_); + + // List of callbacks which will be called when 'enqueued_buffers_' becomes + // empty. + std::vector> on_empty_callbacks_; +}; + +} // namespace gpu +} // namespace xla + +#endif // TENSORFLOW_COMPILER_XLA_SERVICE_GPU_XFEED_QUEUE_H_ diff --git a/tensorflow/compiler/xla/service/graphviz_example.cc b/tensorflow/compiler/xla/service/graphviz_example.cc index acf661148699dab18916e3065ee647d37fda6208..aa89567ee86e59e197045c0b51eed3b9aa59fef7 100644 --- a/tensorflow/compiler/xla/service/graphviz_example.cc +++ b/tensorflow/compiler/xla/service/graphviz_example.cc @@ -22,6 +22,7 @@ limitations under the License. #include #include +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/literal_util.h" #include "tensorflow/compiler/xla/ptr_util.h" #include "tensorflow/compiler/xla/service/hlo_computation.h" @@ -47,7 +48,7 @@ HloComputation* AddScalarConstantComputation(int64 addend, HloModule* module) { auto x_value = builder.AddInstruction(HloInstruction::CreateParameter( 0, ShapeUtil::MakeShape(F32, {}), "x_value")); auto half = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(0.5))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(0.5))); builder.AddInstruction(HloInstruction::CreateBinary( half->shape(), HloOpcode::kAdd, x_value, half)); return module->AddEmbeddedComputation(builder.Build()); @@ -122,7 +123,7 @@ std::unique_ptr MakeBigGraph() { auto rng = builder.AddInstruction( HloInstruction::CreateRng(vshape, RNG_UNIFORM, {param_m, param_m})); auto one = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0))); auto add_computation = ScalarSumComputation(module.get()); builder.AddInstruction( HloInstruction::CreateReduce(vshape, rng, one, {1}, add_computation)); diff --git a/tensorflow/compiler/xla/service/heap_simulator_test.cc b/tensorflow/compiler/xla/service/heap_simulator_test.cc index 3849b565e3136924b2d2b1929353885f85b1a043..b41dc66fe9f5e869a114be96b7cc01fc1a3d59da 100644 --- a/tensorflow/compiler/xla/service/heap_simulator_test.cc +++ b/tensorflow/compiler/xla/service/heap_simulator_test.cc @@ -19,7 +19,7 @@ limitations under the License. #include #include -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/service/buffer_value.h" #include "tensorflow/compiler/xla/service/hlo_computation.h" #include "tensorflow/compiler/xla/service/hlo_instruction.h" @@ -239,7 +239,7 @@ class HeapSimulatorTest : public HloTestBase { TEST_F(HeapSimulatorTest, ScalarConstant) { auto builder = HloComputation::Builder(TestName()); auto const0 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0))); // Constants aren't assigned. See b/32248867 HeapSimulatorTracker tracker(TestName(), builder.Build(), {const0}); @@ -674,7 +674,7 @@ class HeapAlgorithmTestBase : public ::testing::Test { const BufferValue* DummyBufferValue() { const BufferValue::Id id = buffers_.size(); auto const0 = builder_.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0))); buffers_.emplace_back(MakeUnique(id, const0, ShapeIndex{})); return buffers_.back().get(); } diff --git a/tensorflow/compiler/xla/service/hlo_alias_analysis_test.cc b/tensorflow/compiler/xla/service/hlo_alias_analysis_test.cc index 8f18d50f6e033fab1c01f42017b951c224c22799..403d4df6b502e4cceaf5b7341278590b3973153f 100644 --- a/tensorflow/compiler/xla/service/hlo_alias_analysis_test.cc +++ b/tensorflow/compiler/xla/service/hlo_alias_analysis_test.cc @@ -18,7 +18,7 @@ limitations under the License. #include #include -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/service/flatten_call_graph.h" #include "tensorflow/compiler/xla/service/hlo_graph_dumper.h" #include "tensorflow/compiler/xla/service/hlo_matchers.h" @@ -116,9 +116,9 @@ TEST_F(HloAliasAnalysisTest, BinaryOperation) { // Test the analysis on a single binary operation (Add). auto builder = HloComputation::Builder(TestName()); auto constant1 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0))); auto constant2 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(2.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(2.0))); auto add = builder.AddInstruction(HloInstruction::CreateBinary( scalar_shape_, HloOpcode::kAdd, constant1, constant2)); module_->AddEntryComputation(builder.Build()); @@ -228,9 +228,9 @@ TEST_F(HloAliasAnalysisTest, SingleCall) { auto builder = HloComputation::Builder(TestName()); auto constant1 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0))); auto constant2 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(2.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(2.0))); auto call = builder.AddInstruction(HloInstruction::CreateCall( scalar_shape_, {constant1, constant2}, called_computation)); module_->AddEntryComputation(builder.Build()); @@ -267,9 +267,9 @@ TEST_F(HloAliasAnalysisTest, ComputationCalledTwice) { auto builder = HloComputation::Builder(TestName()); auto constant1 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0))); auto constant2 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(2.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(2.0))); auto call1 = builder.AddInstruction(HloInstruction::CreateCall( scalar_shape_, {constant1, constant2}, called_computation)); auto call2 = builder.AddInstruction(HloInstruction::CreateCall( @@ -346,15 +346,15 @@ TEST_F(HloAliasAnalysisTest, SingleWhile) { auto cond_param = cond_builder.AddInstruction( HloInstruction::CreateParameter(0, tuple_shape, "param")); cond_builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(false))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(false))); HloComputation* condition = module_->AddEmbeddedComputation(cond_builder.Build()); auto builder = HloComputation::Builder(TestName()); auto constant1 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0))); auto constant2 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(2.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(2.0))); auto tuple = builder.AddInstruction( HloInstruction::CreateTuple({constant1, constant2})); auto xla_while = builder.AddInstruction( @@ -439,15 +439,15 @@ TEST_F(HloAliasAnalysisTest, SequentialWhiles) { cond_builder.AddInstruction( HloInstruction::CreateParameter(0, tuple_shape, "param")); cond_builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(false))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(false))); HloComputation* condition = module_->AddEmbeddedComputation(cond_builder.Build()); auto builder = HloComputation::Builder(TestName()); auto constant1 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0))); auto constant2 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(2.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(2.0))); auto tuple = builder.AddInstruction( HloInstruction::CreateTuple({constant1, constant2})); auto xla_while0 = builder.AddInstruction( @@ -498,7 +498,7 @@ TEST_F(HloAliasAnalysisTest, NestedWhiles) { cond_builder.AddInstruction( HloInstruction::CreateParameter(0, tuple_shape, "param")); cond_builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(false))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(false))); return cond_builder.Build(); }; // Build separate condition computations so the call graph is flat. The @@ -543,9 +543,9 @@ TEST_F(HloAliasAnalysisTest, NestedWhiles) { auto builder = HloComputation::Builder(TestName()); auto constant1 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0))); auto constant2 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(2.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(2.0))); auto tuple = builder.AddInstruction( HloInstruction::CreateTuple({constant1, constant2})); auto entry_while = builder.AddInstruction( @@ -608,17 +608,17 @@ TEST_F(HloAliasAnalysisTest, SwizzlingWhile) { cond_builder.AddInstruction( HloInstruction::CreateParameter(0, tuple_shape, "param")); auto cond_constant = cond_builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(false))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(false))); HloComputation* condition = module_->AddEmbeddedComputation(cond_builder.Build()); auto builder = HloComputation::Builder(TestName()); auto constant1 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0))); auto constant2 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(2.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(2.0))); auto constant3 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(3.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(3.0))); auto tuple = builder.AddInstruction( HloInstruction::CreateTuple({constant1, constant2, constant3})); auto xla_while = builder.AddInstruction( @@ -654,19 +654,18 @@ TEST_F(HloAliasAnalysisTest, SwizzlingWhile) { } TEST_F(HloAliasAnalysisTest, TupleSelect) { - // Test a kSelect of a tuple value. Non-top-level element flow through the - // instruction. + // Test a kTupleSelect. Non-top-level element flow through the instruction. auto builder = HloComputation::Builder(TestName()); auto pred = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(false))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(false))); auto constant1 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0))); auto constant2 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(2.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(2.0))); auto constant3 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(3.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(3.0))); auto constant4 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(4.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(4.0))); auto tuple1 = builder.AddInstruction(HloInstruction::CreateTuple({constant1})); auto tuple2 = @@ -677,13 +676,13 @@ TEST_F(HloAliasAnalysisTest, TupleSelect) { builder.AddInstruction(HloInstruction::CreateTuple({constant4})); const Shape tuple_shape = tuple1->shape(); auto select11 = builder.AddInstruction(HloInstruction::CreateTernary( - tuple_shape, HloOpcode::kSelect, pred, tuple1, tuple1)); + tuple_shape, HloOpcode::kTupleSelect, pred, tuple1, tuple1)); auto select12 = builder.AddInstruction(HloInstruction::CreateTernary( - tuple_shape, HloOpcode::kSelect, pred, tuple1, tuple2)); + tuple_shape, HloOpcode::kTupleSelect, pred, tuple1, tuple2)); auto select34 = builder.AddInstruction(HloInstruction::CreateTernary( - tuple_shape, HloOpcode::kSelect, pred, tuple3, tuple4)); + tuple_shape, HloOpcode::kTupleSelect, pred, tuple3, tuple4)); auto select1234 = builder.AddInstruction(HloInstruction::CreateTernary( - tuple_shape, HloOpcode::kSelect, pred, select12, select34)); + tuple_shape, HloOpcode::kTupleSelect, pred, select12, select34)); module_->AddEntryComputation(builder.Build()); @@ -718,7 +717,7 @@ TEST_F(HloAliasAnalysisTest, TupleSelect) { } TEST_F(HloAliasAnalysisTest, TupleSelectToWhile) { - // Test a tuple-shaped kSelect feeding a kWhile instruction. HLO: + // Test a tuple-shaped kTupleSelect feeding a kWhile instruction. HLO: // // body((F32[], F32[]) %tuple_param): // %negate = Negate(%tuple_param{0}) @@ -754,22 +753,22 @@ TEST_F(HloAliasAnalysisTest, TupleSelectToWhile) { auto cond_param = cond_builder.AddInstruction( HloInstruction::CreateParameter(0, tuple_shape, "param")); cond_builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(false))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(false))); HloComputation* condition = module_->AddEmbeddedComputation(cond_builder.Build()); auto pred = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(false))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(false))); auto constant1 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0))); auto constant2 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(2.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(2.0))); auto tuple1 = builder.AddInstruction(HloInstruction::CreateTuple({constant1})); auto tuple2 = builder.AddInstruction(HloInstruction::CreateTuple({constant2})); auto select = builder.AddInstruction(HloInstruction::CreateTernary( - tuple_shape, HloOpcode::kSelect, pred, tuple1, tuple2)); + tuple_shape, HloOpcode::kTupleSelect, pred, tuple1, tuple2)); auto xla_while = builder.AddInstruction( HloInstruction::CreateWhile(tuple_shape, condition, body, select)); @@ -806,7 +805,7 @@ TEST_F(HloAliasAnalysisTest, Bitcast) { // Bitcasting a value should not produce a new buffer. auto builder = HloComputation::Builder(TestName()); auto constant = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0))); auto bitcast = builder.AddInstruction(HloInstruction::CreateUnary( scalar_shape_, HloOpcode::kBitcast, constant)); @@ -825,7 +824,7 @@ TEST_F(HloAliasAnalysisTest, BitcastInterference) { // interference. auto builder = HloComputation::Builder(TestName()); auto constant = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0))); auto bitcast = builder.AddInstruction(HloInstruction::CreateUnary( scalar_shape_, HloOpcode::kBitcast, constant)); builder.AddInstruction(HloInstruction::CreateTuple({constant, bitcast})); @@ -844,13 +843,13 @@ TEST_F(HloAliasAnalysisTest, WhileInterference) { // the other use of the init. auto builder = HloComputation::Builder(TestName()); auto init = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0))); auto cond_builder = HloComputation::Builder("condition"); auto cond_param = cond_builder.AddInstruction( HloInstruction::CreateParameter(0, init->shape(), "param")); auto cond_root = cond_builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(false))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(false))); HloComputation* condition = module_->AddEmbeddedComputation(cond_builder.Build()); diff --git a/tensorflow/compiler/xla/service/hlo_computation.cc b/tensorflow/compiler/xla/service/hlo_computation.cc index c057be82014b00e3ff63f835fcb78c08f8d9c154..166a83fadedfcc9e4aa9e41e7c939b0114e9b0ea 100644 --- a/tensorflow/compiler/xla/service/hlo_computation.cc +++ b/tensorflow/compiler/xla/service/hlo_computation.cc @@ -120,6 +120,30 @@ HloInstruction* HloComputation::AddParameter( return instructions_.back().get(); } +namespace { + +// Returns the new name for a fusion parameter when we change its number. +// +// Fusion parameters are named foo.param_1, bar.param_2, etc. We are +// renumbering the parameters, so replace the final number in the name with +// the updated value. +string RenameFusionParameter(const string& original_name, int64 new_param_no) { + const string param_underscore = ".param_"; + size_t index = original_name.rfind(param_underscore); + if (index == string::npos) { + return original_name; + } + string after_param = original_name.substr(index + param_underscore.size()); + int64 numeric_suffix; + if (tensorflow::strings::safe_strto64(after_param, &numeric_suffix)) { + return StrCat(original_name.substr(0, index + param_underscore.size()), + new_param_no); + } + return original_name; +} + +} // namespace + Status HloComputation::RemoveParameter(int64 param_no) { CHECK_GE(param_no, 0); CHECK_LT(param_no, param_instructions_.size()); @@ -132,21 +156,8 @@ Status HloComputation::RemoveParameter(int64 param_no) { while (param_no < param_instructions_.size()) { param_instruction = param_instructions_[param_no]; - string param_name = param_instruction->name(); - // Fusion parameters are named foo.param_1, bar.param_2, etc. We are - // renumbering the parameters, so replace the final number in the name with - // the updated value. - const string param_underscore = ".param_"; - size_t index = param_name.rfind(param_underscore); - if (index == string::npos) { - string after_param = name().substr(index + param_underscore.size()); - int64 numeric_suffix; - if (tensorflow::strings::safe_strto64(after_param, &numeric_suffix)) { - param_name = - StrCat(param_name.substr(0, index), param_underscore, param_no); - } - } - + string param_name = + RenameFusionParameter(param_instruction->name(), param_no); HloInstruction* new_instr = AddInstructionInternal(HloInstruction::CreateParameter( param_no, param_instruction->shape(), param_name)); @@ -159,6 +170,34 @@ Status HloComputation::RemoveParameter(int64 param_no) { return Status::OK(); } +Status HloComputation::RemoveUnusedParameters() { + CHECK(IsFusionComputation()); + int64 removed = 0; + for (int64 i = 0; i < param_instructions_.size(); ++i) { + HloInstruction* param_instruction = param_instructions_[i]; + if (param_instruction->user_count() == 0 && + param_instruction != root_instruction()) { + TF_RETURN_IF_ERROR(RemoveInstruction(param_instruction)); + ++removed; + continue; + } + + if (removed > 0) { + const int64 param_no = i - removed; + string param_name = + RenameFusionParameter(param_instruction->name(), param_no); + HloInstruction* new_instr = + AddInstructionInternal(HloInstruction::CreateParameter( + param_no, param_instruction->shape(), param_name)); + TF_RETURN_IF_ERROR(param_instruction->ReplaceAllUsesWith(new_instr)); + param_instructions_[param_no] = new_instr; + TF_RETURN_IF_ERROR(RemoveInstruction(param_instruction)); + } + } + param_instructions_.resize(param_instructions_.size() - removed); + return Status::OK(); +} + bool HloComputation::IsRemovable(const HloInstruction* instruction) { // If the instruction has control predecessors or successors then we cannot // remove the instruction without violating ordering constraints (added, for @@ -245,9 +284,8 @@ void HloComputation::set_root_instruction( if (!IsFusionComputation()) { CHECK(ShapeUtil::Compatible(new_root_instruction->shape(), root_instruction_->shape())) - << new_root_instruction->shape().ShortDebugString() - << " is incompatible with " - << root_instruction_->shape().ShortDebugString(); + << new_root_instruction->shape() << " is incompatible with " + << root_instruction_->shape(); } bool root_found = false; for (auto& instruction : instructions_) { @@ -490,8 +528,10 @@ HloInstruction* HloComputation::CreateFusionInstruction( } StatusOr HloComputation::DeepCopyHelper( - HloInstruction* instruction, const ShapeTree* indices_to_copy, - ShapeTree* copies_added, ShapeIndex* index) { + HloInstruction* instruction, ShapeIndex* index, + const std::function< + HloInstruction*(HloInstruction* leaf, const ShapeIndex& leaf_index, + HloComputation* computation)>& copy_leaf) { if (ShapeUtil::IsTuple(instruction->shape())) { std::vector elements; for (int64 i = 0; i < ShapeUtil::TupleElementCount(instruction->shape()); @@ -502,9 +542,8 @@ StatusOr HloComputation::DeepCopyHelper( instruction, i)); index->push_back(i); - TF_ASSIGN_OR_RETURN( - HloInstruction * element, - DeepCopyHelper(gte, indices_to_copy, copies_added, index)); + TF_ASSIGN_OR_RETURN(HloInstruction * element, + DeepCopyHelper(gte, index, copy_leaf)); elements.push_back(element); index->pop_back(); } @@ -518,19 +557,7 @@ StatusOr HloComputation::DeepCopyHelper( // Array shape. TF_RET_CHECK(ShapeUtil::IsArray(instruction->shape())); - if (indices_to_copy == nullptr || indices_to_copy->element(*index)) { - // Use kCopy to copy array elements - HloInstruction* copy = AddInstruction(HloInstruction::CreateUnary( - instruction->shape(), HloOpcode::kCopy, instruction)); - if (copies_added != nullptr) { - *copies_added->mutable_element(*index) = copy; - } - return copy; - } else { - // Elements which are not to be copied are passed through - // transparently. - return instruction; - } + return copy_leaf(instruction, *index, this); } StatusOr HloComputation::DeepCopyInstruction( @@ -552,7 +579,36 @@ StatusOr HloComputation::DeepCopyInstruction( } ShapeIndex index; - return DeepCopyHelper(instruction, indices_to_copy, copies_added, &index); + auto copy_leaf = [indices_to_copy, copies_added]( + HloInstruction* leaf, const ShapeIndex& leaf_index, + HloComputation* computation) { + if (indices_to_copy == nullptr || indices_to_copy->element(leaf_index)) { + HloInstruction* copy = computation->AddInstruction( + HloInstruction::CreateUnary(leaf->shape(), HloOpcode::kCopy, leaf)); + if (copies_added != nullptr) { + *copies_added->mutable_element(leaf_index) = copy; + } + return copy; + } + // Elements which are not to be copied are passed through + // transparently. + return leaf; + }; + return DeepCopyHelper(instruction, &index, copy_leaf); +} + +StatusOr HloComputation::DeepCopyInstructionWithCustomCopier( + HloInstruction* instruction, + const std::function< + HloInstruction*(HloInstruction* leaf, const ShapeIndex& leaf_index, + HloComputation* computation)>& copy_leaf) { + if (instruction->parent() != this) { + return FailedPrecondition( + "Can't deep copy instruction %s: instruction is not in computation %s", + instruction->name().c_str(), name().c_str()); + } + ShapeIndex index; + return DeepCopyHelper(instruction, &index, copy_leaf); } ProgramShape HloComputation::ComputeProgramShape() const { @@ -625,7 +681,7 @@ std::unique_ptr HloComputation::ComputeReachability() inputs.assign(hlo->operands().begin(), hlo->operands().end()); inputs.insert(inputs.end(), hlo->control_predecessors().begin(), hlo->control_predecessors().end()); - result->SetReachabilityToUnion(inputs, hlo); + result->FastSetReachabilityToUnion(inputs, hlo); } return result; } diff --git a/tensorflow/compiler/xla/service/hlo_computation.h b/tensorflow/compiler/xla/service/hlo_computation.h index 0f111a1a7672d419d32387d7fe0020744ba8ddf2..abc1da4da3a3cee22c3d7fee8babe6392910db88 100644 --- a/tensorflow/compiler/xla/service/hlo_computation.h +++ b/tensorflow/compiler/xla/service/hlo_computation.h @@ -16,6 +16,7 @@ limitations under the License. #ifndef TENSORFLOW_COMPILER_XLA_SERVICE_HLO_COMPUTATION_H_ #define TENSORFLOW_COMPILER_XLA_SERVICE_HLO_COMPUTATION_H_ +#include #include #include #include @@ -113,6 +114,11 @@ class HloComputation { // instruction. Status RemoveParameter(int64 param_no); + // Remove unused parameters from the computation. + // Note this is only applicatable to the computation for the fusion + // instruction. + Status RemoveUnusedParameters(); + // Add new parameter instruction to the computation. // This should be a new parameter. Instruction will be appended to parameters // and inserted to the instruction list. @@ -249,6 +255,14 @@ class HloComputation { const ShapeTree* indices_to_copy = nullptr, ShapeTree* copies_added = nullptr); + // As above, but uses a custom function to copy the leaf nodes, which could + // create alternative HLOs other than kCopy, or even pass-throughs. + StatusOr DeepCopyInstructionWithCustomCopier( + HloInstruction* instruction, + const std::function< + HloInstruction*(HloInstruction* leaf, const ShapeIndex& leaf_index, + HloComputation* computation)>& copy_leaf); + // Computes and returns the ProgramShape of this computation (shape of // parameters and result with layout). ProgramShape ComputeProgramShape() const; @@ -373,8 +387,10 @@ class HloComputation { // Internal helper for recursive copying of an instruction. Creates and // returns a deep copy of the given instruction. StatusOr DeepCopyHelper( - HloInstruction* instruction, const ShapeTree* indices_to_copy, - ShapeTree* copies_added, ShapeIndex* index); + HloInstruction* instruction, ShapeIndex* index, + const std::function< + HloInstruction*(HloInstruction* leaf, const ShapeIndex& leaf_index, + HloComputation* computation)>& copy_leaf); // Internal helper to collect unreachable roots. std::vector CollectUnreachableRoots() const; diff --git a/tensorflow/compiler/xla/service/hlo_computation_test.cc b/tensorflow/compiler/xla/service/hlo_computation_test.cc index c504fc51d229ca70499bfe006ed9c350251d2c8a..e4c547033139185d5dd4ef37db2d22a6431c1102 100644 --- a/tensorflow/compiler/xla/service/hlo_computation_test.cc +++ b/tensorflow/compiler/xla/service/hlo_computation_test.cc @@ -17,7 +17,7 @@ limitations under the License. #include -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/service/dfs_hlo_visitor_with_default.h" #include "tensorflow/compiler/xla/service/hlo_instruction.h" #include "tensorflow/compiler/xla/service/hlo_matchers.h" @@ -118,7 +118,7 @@ TEST_F(HloComputationTest, PostOrderSingleton) { // Test GetInstructionPostOrder for a computation with one instruction. auto builder = HloComputation::Builder(TestName()); auto constant = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(42.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(42.0f))); auto module = CreateNewModule(); auto computation = module->AddEntryComputation(builder.Build()); EXPECT_THAT(computation->MakeInstructionPostOrder(), ElementsAre(constant)); @@ -129,7 +129,7 @@ TEST_F(HloComputationTest, PostOrderSimple) { // instructions. auto builder = HloComputation::Builder(TestName()); auto constant = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(42.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(42.0f))); auto negate1 = builder.AddInstruction( HloInstruction::CreateUnary(r0f32_, HloOpcode::kNegate, constant)); auto negate2 = builder.AddInstruction( @@ -144,7 +144,7 @@ TEST_F(HloComputationTest, PostOrderTrace) { // Test GetInstructionPostOrder for a computation with a trace instruction. auto builder = HloComputation::Builder(TestName()); auto constant = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(42.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(42.0f))); auto negate1 = builder.AddInstruction( HloInstruction::CreateUnary(r0f32_, HloOpcode::kNegate, constant)); auto trace = @@ -163,13 +163,13 @@ TEST_F(HloComputationTest, PostOrderDisconnectedInstructions) { // which are not connected. auto builder = HloComputation::Builder(TestName()); auto constant1 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(42.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(42.0f))); auto constant2 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(42.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(42.0f))); auto constant3 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(42.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(42.0f))); auto constant4 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(42.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(42.0f))); auto module = CreateNewModule(); auto computation = module->AddEntryComputation(builder.Build()); EXPECT_THAT(computation->MakeInstructionPostOrder(), @@ -181,11 +181,11 @@ TEST_F(HloComputationTest, PostOrderWithMultipleRoots) { // which are not connected. auto builder = HloComputation::Builder(TestName()); auto constant1 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(42.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(42.0f))); auto constant2 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(42.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(42.0f))); auto constant3 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(42.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(42.0f))); auto add1 = builder.AddInstruction(HloInstruction::CreateBinary( r0f32_, HloOpcode::kAdd, constant1, constant2)); auto add2 = builder.AddInstruction(HloInstruction::CreateBinary( @@ -205,11 +205,11 @@ TEST_F(HloComputationTest, VisitWithMultipleRoots) { // computation has multiple roots (dead code). auto builder = HloComputation::Builder(TestName()); auto constant1 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(42.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(42.0f))); auto constant2 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(42.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(42.0f))); auto constant3 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(42.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(42.0f))); // Add three disconnected add expressions. builder.AddInstruction(HloInstruction::CreateBinary(r0f32_, HloOpcode::kAdd, constant1, constant2)); @@ -256,7 +256,7 @@ TEST_F(HloComputationTest, DeepCopyArray) { // Test that DeepCopyInstruction properly copies an array. auto builder = HloComputation::Builder(TestName()); auto constant = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR1({1.0, 2.0, 3.0}))); + LiteralUtil::CreateR1({1.0, 2.0, 3.0}))); auto module = CreateNewModule(); auto computation = module->AddEntryComputation(builder.Build()); auto copy = computation->DeepCopyInstruction(constant).ValueOrDie(); @@ -268,9 +268,9 @@ TEST_F(HloComputationTest, DeepCopyTuple) { // Test that DeepCopyInstruction properly copies a tuple. auto builder = HloComputation::Builder(TestName()); auto constant1 = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR1({1.0, 2.0, 3.0}))); + LiteralUtil::CreateR1({1.0, 2.0, 3.0}))); auto constant2 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(42.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(42.0))); auto tuple = builder.AddInstruction( HloInstruction::CreateTuple({constant1, constant2})); @@ -289,7 +289,7 @@ TEST_F(HloComputationTest, DeepCopyArrayAtIndices) { // copy are specified. auto builder = HloComputation::Builder(TestName()); auto constant = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR1({1.0, 2.0, 3.0}))); + LiteralUtil::CreateR1({1.0, 2.0, 3.0}))); auto computation = builder.Build(); { @@ -314,9 +314,9 @@ TEST_F(HloComputationTest, DeepCopyTupleAtIndices) { // specified by the given indices. auto builder = HloComputation::Builder(TestName()); auto constant1 = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR1({1.0, 2.0, 3.0}))); + LiteralUtil::CreateR1({1.0, 2.0, 3.0}))); auto constant2 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(42.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(42.0))); auto tuple = builder.AddInstruction( HloInstruction::CreateTuple({constant1, constant2})); auto computation = builder.Build(); @@ -375,22 +375,22 @@ TEST_F(HloComputationTest, DeepCopyToken) { // Test that DeepCopyInstruction properly handles tokens which should not be // copied. auto builder = HloComputation::Builder(TestName()); - auto token = builder.AddInstruction(HloInstruction::CreateGenerateToken({})); + auto token = builder.AddInstruction(HloInstruction::CreateToken()); auto module = CreateNewModule(); auto computation = module->AddEntryComputation(builder.Build()); auto copy = computation->DeepCopyInstruction(token).ValueOrDie(); // No copy should be added. - EXPECT_THAT(copy, op::GenerateToken()); + EXPECT_THAT(copy, op::AfterAll()); } TEST_F(HloComputationTest, DeepCopyTokenTuple) { // Test that DeepCopyInstruction properly handles tokens which should not be // copied. auto builder = HloComputation::Builder(TestName()); - auto token = builder.AddInstruction(HloInstruction::CreateGenerateToken({})); + auto token = builder.AddInstruction(HloInstruction::CreateToken()); auto constant = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(42.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(42.0))); auto tuple = builder.AddInstruction(HloInstruction::CreateTuple({token, constant})); auto module = CreateNewModule(); @@ -407,7 +407,7 @@ TEST_F(HloComputationTest, CycleDetection) { // Test whether the visitor can detect cycles in the graph. auto builder = HloComputation::Builder(TestName()); auto constant = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(42.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(42.0f))); auto negate = builder.AddInstruction( HloInstruction::CreateUnary(r0f32_, HloOpcode::kNegate, constant)); auto add = builder.AddInstruction( @@ -433,7 +433,7 @@ TEST_F(HloComputationTest, RemoveInstructionWithDuplicateOperand) { // twice. auto builder = HloComputation::Builder(TestName()); auto constant = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(42.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(42.0f))); auto dead_negate = builder.AddInstruction( HloInstruction::CreateUnary(r0f32_, HloOpcode::kNegate, constant)); auto dead_add = builder.AddInstruction(HloInstruction::CreateBinary( @@ -456,9 +456,9 @@ TEST_F(HloComputationTest, RemoveInstructionWithDuplicateOperand) { TEST_F(HloComputationTest, CloneWithControlDependency) { auto builder = HloComputation::Builder(TestName()); auto constant1 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0f))); auto constant2 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(2.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(2.0f))); auto add = builder.AddInstruction(HloInstruction::CreateBinary( r0f32_, HloOpcode::kAdd, constant1, constant2)); @@ -502,9 +502,9 @@ TEST_F(HloComputationTest, Reachability) { // There is a control dependency from 'add' to 'exp'. auto builder = HloComputation::Builder(TestName()); auto constant1 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0f))); auto constant2 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(2.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(2.0f))); auto add = builder.AddInstruction(HloInstruction::CreateBinary( r0f32_, HloOpcode::kAdd, constant1, constant2)); auto negate = builder.AddInstruction( @@ -607,13 +607,14 @@ TEST_F(HloComputationTest, Stringification) { auto* computation = module->AddEntryComputation(builder.Build()); auto options = HloPrintOptions().set_print_metadata(false); - EXPECT_EQ(computation->ToString(options), - R"(%TransposeDot (x: f32[5,10], y: f32[20,10]) -> f32[5,20] { + const string expected_computation = + R"(%TransposeDot (x: f32[5,10], y: f32[20,10]) -> f32[5,20] { %x = f32[5,10]{1,0} parameter(0) %y = f32[20,10]{1,0} parameter(1) %transpose = f32[10,20]{1,0} transpose(f32[20,10]{1,0} %y), dimensions={1,0} ROOT %dot = f32[5,20]{1,0} dot(f32[5,10]{1,0} %x, f32[10,20]{1,0} %transpose), lhs_contracting_dims={1}, rhs_contracting_dims={0} -})"); +})"; + EXPECT_EQ(computation->ToString(options), expected_computation); } TEST_F(HloComputationTest, StringificationIndent) { @@ -639,13 +640,14 @@ TEST_F(HloComputationTest, StringificationIndent) { auto options = HloPrintOptions().set_print_metadata(false).set_indent_amount(2); - EXPECT_EQ(computation->ToString(options), - R"( %TransposeDot (x: f32[5,10], y: f32[20,10]) -> f32[5,20] { + const string expected_computation = + R"( %TransposeDot (x: f32[5,10], y: f32[20,10]) -> f32[5,20] { %x = f32[5,10]{1,0} parameter(0) %y = f32[20,10]{1,0} parameter(1) %transpose = f32[10,20]{1,0} transpose(f32[20,10]{1,0} %y), dimensions={1,0} ROOT %dot = f32[5,20]{1,0} dot(f32[5,10]{1,0} %x, f32[10,20]{1,0} %transpose), lhs_contracting_dims={1}, rhs_contracting_dims={0} - })"); + })"; + EXPECT_EQ(computation->ToString(options), expected_computation); } TEST_F(HloComputationTest, StringificationCanonical) { @@ -670,21 +672,23 @@ TEST_F(HloComputationTest, StringificationCanonical) { auto* computation = module->AddEntryComputation(builder.Build()); auto options = HloPrintOptions().set_print_metadata(false); - EXPECT_EQ(computation->ToString(options), - R"(%TransposeDot (x: f32[5,10], y: f32[20,10]) -> f32[5,20] { + const string expected_computation1 = + R"(%TransposeDot (x: f32[5,10], y: f32[20,10]) -> f32[5,20] { %x = f32[5,10]{1,0} parameter(0) %y = f32[20,10]{1,0} parameter(1) %transpose = f32[10,20]{1,0} transpose(f32[20,10]{1,0} %y), dimensions={1,0} ROOT %dot = f32[5,20]{1,0} dot(f32[5,10]{1,0} %x, f32[10,20]{1,0} %transpose), lhs_contracting_dims={1}, rhs_contracting_dims={0} -})"); +})"; + EXPECT_EQ(computation->ToString(options), expected_computation1); options = HloPrintOptions().Canonical(); - EXPECT_EQ(computation->ToString(options), R"(TransposeDot { + const string expected_computation2 = R"(TransposeDot { tmp_0 = f32[5,10]{1,0} parameter(0) tmp_1 = f32[20,10]{1,0} parameter(1) tmp_2 = f32[10,20]{1,0} transpose(f32[20,10]{1,0} tmp_1), dimensions={1,0} ROOT tmp_3 = f32[5,20]{1,0} dot(f32[5,10]{1,0} tmp_0, f32[10,20]{1,0} tmp_2), lhs_contracting_dims={1}, rhs_contracting_dims={0} -})"); +})"; + EXPECT_EQ(computation->ToString(options), expected_computation2); } } // namespace diff --git a/tensorflow/compiler/xla/service/hlo_constant_folding.cc b/tensorflow/compiler/xla/service/hlo_constant_folding.cc index 35ecd4428d0dfde2de445ea34472d2c78148c6c9..7229031c0c7f8bd374cfb495c7d8c11e9ca8b95e 100644 --- a/tensorflow/compiler/xla/service/hlo_constant_folding.cc +++ b/tensorflow/compiler/xla/service/hlo_constant_folding.cc @@ -21,7 +21,7 @@ limitations under the License. #include #include "tensorflow/compiler/xla/layout_util.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/service/dfs_hlo_visitor_with_default.h" #include "tensorflow/compiler/xla/service/hlo_computation.h" #include "tensorflow/compiler/xla/service/hlo_evaluator.h" @@ -51,14 +51,18 @@ StatusOr HloConstantFolding::Run(HloModule* module) { computation->root_instruction() != instruction) { continue; } - // Skip Constant, Parameter, Reduce operation. + // Skip Constant, Parameter, Reduce, and AfterAll operation. // TODO(b/35975797): Enable Reduce operation once arbitrary computation // are supported by the evaluator. // TODO(b/64407269): Enable Tuple once the timeout issue is resolved. + // TODO(b/110532604): Enable AfterAll once AfterAll requires at least one + // operand in which case constant folding will be impossible and this + // special case is not necessary. if (instruction->opcode() == HloOpcode::kParameter || instruction->opcode() == HloOpcode::kConstant || instruction->opcode() == HloOpcode::kTuple || - instruction->opcode() == HloOpcode::kReduce) { + instruction->opcode() == HloOpcode::kReduce || + instruction->opcode() == HloOpcode::kAfterAll) { continue; } // Skip instructions with non-constant operands. diff --git a/tensorflow/compiler/xla/service/hlo_constant_folding_test.cc b/tensorflow/compiler/xla/service/hlo_constant_folding_test.cc index 5d05ccfc0b223d8749a2577ba1bf96b1ab3e761b..64a42c1efc0c788ae8e66fb72b2d9aecec179082 100644 --- a/tensorflow/compiler/xla/service/hlo_constant_folding_test.cc +++ b/tensorflow/compiler/xla/service/hlo_constant_folding_test.cc @@ -19,7 +19,7 @@ limitations under the License. #include #include "tensorflow/compiler/xla/layout_util.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/service/hlo_computation.h" #include "tensorflow/compiler/xla/service/hlo_instruction.h" #include "tensorflow/compiler/xla/service/hlo_matchers.h" @@ -41,7 +41,7 @@ using HloConstantFoldingTest = HloTestBase; TEST_F(HloConstantFoldingTest, ConvertF32ToS64) { HloComputation::Builder builder(TestName()); HloInstruction* input = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(42.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(42.0f))); builder.AddInstruction( HloInstruction::CreateConvert(ShapeUtil::MakeShape(S64, {}), input)); @@ -62,7 +62,7 @@ TEST_F(HloConstantFoldingTest, ConvertF32ToS64) { TEST_F(HloConstantFoldingTest, ConvertS64ToF32) { HloComputation::Builder builder(TestName()); HloInstruction* input = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(42))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(42))); builder.AddInstruction( HloInstruction::CreateConvert(ShapeUtil::MakeShape(F32, {}), input)); @@ -82,8 +82,8 @@ TEST_F(HloConstantFoldingTest, ConvertS64ToF32) { TEST_F(HloConstantFoldingTest, ConvertF32ArrayToS64Array) { HloComputation::Builder builder(TestName()); - HloInstruction* input = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR1({42.0f, 19.0f}))); + HloInstruction* input = builder.AddInstruction(HloInstruction::CreateConstant( + LiteralUtil::CreateR1({42.0f, 19.0f}))); builder.AddInstruction( HloInstruction::CreateConvert(ShapeUtil::MakeShape(S64, {2}), input)); @@ -120,7 +120,7 @@ TEST_F(HloConstantFoldingTest, Concatenate) { for (auto csize : test_config.concat_sizes) { dimensions[test_config.concat_dimension] = csize; concat_size += csize; - auto literal = Literal::CreateFromDimensions(F32, dimensions); + auto literal = LiteralUtil::CreateFromDimensions(F32, dimensions); HloInstruction* insn = builder.AddInstruction( HloInstruction::CreateConstant(std::move(literal))); operands.push_back(insn); @@ -149,7 +149,7 @@ TEST_F(HloConstantFoldingTest, Slice) { const int64 slice_limits[] = {10, 8, 6, 5, 9}; const int64 slice_strides[] = {1, 1, 1, 1, 1}; TF_ASSERT_OK_AND_ASSIGN(auto literal, - Literal::CreateRandomLiteral( + LiteralUtil::CreateRandomLiteral( ShapeUtil::MakeShape(F32, dimensions), 0.0, 1.0)); HloInstruction* literal_instruction = builder.AddInstruction( HloInstruction::CreateConstant(std::move(literal))); @@ -172,7 +172,7 @@ TEST_F(HloConstantFoldingTest, TransposeConstantFold) { HloComputation::Builder builder(TestName()); const int64 dimensions[] = {11, 8, 7, 5, 9}; TF_ASSERT_OK_AND_ASSIGN(auto literal, - Literal::CreateRandomLiteral( + LiteralUtil::CreateRandomLiteral( ShapeUtil::MakeShape(F32, dimensions), 0.0, 1.0)); auto literal_clone = literal->Literal::CloneToUnique(); HloInstruction* literal_instruction = builder.AddInstruction( diff --git a/tensorflow/compiler/xla/service/hlo_cost_analysis.cc b/tensorflow/compiler/xla/service/hlo_cost_analysis.cc index 762e1afc71b108b2e32b5a7f7f1bbeb783fc6fbd..c49cf7f5db5ee9100718fbcd87dc5bdcc175ae5f 100644 --- a/tensorflow/compiler/xla/service/hlo_cost_analysis.cc +++ b/tensorflow/compiler/xla/service/hlo_cost_analysis.cc @@ -164,7 +164,11 @@ Status HloCostAnalysis::HandleGetTupleElement(const HloInstruction*) { return Status::OK(); } -Status HloCostAnalysis::HandleSelect(const HloInstruction*) { +Status HloCostAnalysis::HandleSelect(const HloInstruction* hlo) { + return HandleElementwiseOp(hlo); +} + +Status HloCostAnalysis::HandleTupleSelect(const HloInstruction*) { return Status::OK(); } @@ -393,7 +397,7 @@ Status HloCostAnalysis::HandleTranspose(const HloInstruction*) { return Status::OK(); } -Status HloCostAnalysis::HandleGenerateToken(const HloInstruction*) { +Status HloCostAnalysis::HandleAfterAll(const HloInstruction*) { return Status::OK(); } diff --git a/tensorflow/compiler/xla/service/hlo_cost_analysis.h b/tensorflow/compiler/xla/service/hlo_cost_analysis.h index 0d66736fe1d0677d13a63ede7a203d6ac20c76f5..0181138a6dc554438957e8545c66a98d32dd68d5 100644 --- a/tensorflow/compiler/xla/service/hlo_cost_analysis.h +++ b/tensorflow/compiler/xla/service/hlo_cost_analysis.h @@ -54,7 +54,8 @@ class HloCostAnalysis : public ConstDfsHloVisitor { Status HandleConstant(const HloInstruction* constant) override; Status HandleGetTupleElement( const HloInstruction* get_tuple_element) override; - Status HandleSelect(const HloInstruction* select) override; + Status HandleSelect(const HloInstruction* hlo) override; + Status HandleTupleSelect(const HloInstruction* hlo) override; Status HandleCompare(const HloInstruction* compare) override; Status HandleClamp(const HloInstruction* clamp) override; Status HandleReducePrecision(const HloInstruction* hlo) override; @@ -97,7 +98,7 @@ class HloCostAnalysis : public ConstDfsHloVisitor { Status HandleBroadcast(const HloInstruction* broadcast) override; Status HandlePad(const HloInstruction* pad) override; Status HandleReshape(const HloInstruction* reshape) override; - Status HandleGenerateToken(const HloInstruction* token) override; + Status HandleAfterAll(const HloInstruction* token) override; Status HandleTranspose(const HloInstruction* transpose) override; Status HandleWhile(const HloInstruction* xla_while) override; Status HandleConditional(const HloInstruction* conditional) override; diff --git a/tensorflow/compiler/xla/service/hlo_cost_analysis_test.cc b/tensorflow/compiler/xla/service/hlo_cost_analysis_test.cc index d22bef56730da194816b4ee89dc3196439b350f9..9fd0363f578f562aa30bd3dadfea2c251a722af8 100644 --- a/tensorflow/compiler/xla/service/hlo_cost_analysis_test.cc +++ b/tensorflow/compiler/xla/service/hlo_cost_analysis_test.cc @@ -59,9 +59,9 @@ class HloCostAnalysisTest : public ::testing::Test { // Create a computation for a unary user function: x => exp(x + 0.5) { XlaBuilder builder("add_and_exp"); - auto x = builder.Parameter(0, ShapeUtil::MakeShape(F32, {}), "x"); - auto half = builder.ConstantR0(0.5); - builder.Exp(builder.Add(x, half)); + auto x = Parameter(&builder, 0, ShapeUtil::MakeShape(F32, {}), "x"); + auto half = ConstantR0(&builder, 0.5); + Exp(Add(x, half)); auto computation_status = builder.Build(); TF_CHECK_OK(computation_status.status()); add_and_exp_ = computation_status.ConsumeValueOrDie(); @@ -70,9 +70,9 @@ class HloCostAnalysisTest : public ::testing::Test { // Create a computation for a binary user function: (x, y) => x + y { XlaBuilder builder("add"); - auto x = builder.Parameter(0, ShapeUtil::MakeShape(F32, {}), "x"); - auto y = builder.Parameter(1, ShapeUtil::MakeShape(F32, {}), "y"); - builder.Add(x, y); + auto x = Parameter(&builder, 0, ShapeUtil::MakeShape(F32, {}), "x"); + auto y = Parameter(&builder, 1, ShapeUtil::MakeShape(F32, {}), "y"); + Add(x, y); auto computation_status = builder.Build(); TF_CHECK_OK(computation_status.status()); add_ = computation_status.ConsumeValueOrDie(); @@ -81,9 +81,9 @@ class HloCostAnalysisTest : public ::testing::Test { // Create a computation for a sigmoid function: x => 1 / (1 + exp(-x)) { XlaBuilder builder("sigmoid"); - auto x = builder.Parameter(0, ShapeUtil::MakeShape(F32, {}), "x"); - auto one = builder.ConstantR0(1.0); - builder.Div(one, builder.Add(one, builder.Exp(builder.Neg(x)))); + auto x = Parameter(&builder, 0, ShapeUtil::MakeShape(F32, {}), "x"); + auto one = ConstantR0(&builder, 1.0); + Div(one, Add(one, Exp(Neg(x)))); auto computation_status = builder.Build(); TF_CHECK_OK(computation_status.status()); sigmoid_ = computation_status.ConsumeValueOrDie(); @@ -92,9 +92,9 @@ class HloCostAnalysisTest : public ::testing::Test { // Create a computation for a binary max function: (x, y) => max (x, y) { XlaBuilder builder("max"); - auto x = builder.Parameter(0, ShapeUtil::MakeShape(F32, {}), "x"); - auto y = builder.Parameter(1, ShapeUtil::MakeShape(F32, {}), "y"); - builder.Max(x, y); + auto x = Parameter(&builder, 0, ShapeUtil::MakeShape(F32, {}), "x"); + auto y = Parameter(&builder, 1, ShapeUtil::MakeShape(F32, {}), "y"); + Max(x, y); auto computation_status = builder.Build(); TF_CHECK_OK(computation_status.status()); max_ = computation_status.ConsumeValueOrDie(); @@ -103,9 +103,9 @@ class HloCostAnalysisTest : public ::testing::Test { // Create a computation for a binary GT function: (x, y) => x > y { XlaBuilder builder("gt"); - auto x = builder.Parameter(0, ShapeUtil::MakeShape(F32, {}), "x"); - auto y = builder.Parameter(1, ShapeUtil::MakeShape(F32, {}), "y"); - builder.Gt(x, y); + auto x = Parameter(&builder, 0, ShapeUtil::MakeShape(F32, {}), "x"); + auto y = Parameter(&builder, 1, ShapeUtil::MakeShape(F32, {}), "y"); + Gt(x, y); auto computation_status = builder.Build(); TF_CHECK_OK(computation_status.status()); gt_ = computation_status.ConsumeValueOrDie(); @@ -137,9 +137,9 @@ class HloCostAnalysisTest : public ::testing::Test { TEST_F(HloCostAnalysisTest, MatrixMultiply) { XlaBuilder builder("matrix_multiply"); - auto lhs = builder.Parameter(0, ShapeUtil::MakeShape(F32, {10, 5}), "lhs"); - auto rhs = builder.Parameter(1, ShapeUtil::MakeShape(F32, {5, 30}), "rhs"); - auto result = builder.Dot(lhs, rhs); + auto lhs = Parameter(&builder, 0, ShapeUtil::MakeShape(F32, {10, 5}), "lhs"); + auto rhs = Parameter(&builder, 1, ShapeUtil::MakeShape(F32, {5, 30}), "rhs"); + Dot(lhs, rhs); // Run HLO cost analysis. auto hlo_module = BuildHloGraph(&builder); @@ -159,8 +159,8 @@ TEST_F(HloCostAnalysisTest, MatrixMultiply) { TEST_F(HloCostAnalysisTest, Map) { XlaBuilder builder("map"); - auto input = builder.Parameter(0, ShapeUtil::MakeShape(F32, {10}), "in"); - auto result = builder.Map({input}, add_and_exp_, {0}); + auto input = Parameter(&builder, 0, ShapeUtil::MakeShape(F32, {10}), "in"); + Map(&builder, {input}, add_and_exp_, {0}); // Run HLO cost analysis. auto hlo_module = BuildHloGraph(&builder); @@ -176,17 +176,17 @@ TEST_F(HloCostAnalysisTest, Map) { TEST_F(HloCostAnalysisTest, Convolution) { XlaBuilder builder("convolution"); - auto input = builder.Parameter( - 0, + auto input = Parameter( + &builder, 0, ShapeUtil::MakeShape(F32, {/*p_dim=*/1, /*z_dim=*/1, /*y_dim=*/10, /*x_dim=*/20}), "input"); - auto kernel = builder.Parameter( - 1, + auto kernel = Parameter( + &builder, 1, ShapeUtil::MakeShape(F32, {/*p_dim=*/1, /*z_dim=*/1, /*y_dim=*/3, /*x_dim=*/3}), "kernel"); - auto result = builder.Conv(input, kernel, {1, 1}, Padding::kValid); + Conv(input, kernel, {1, 1}, Padding::kValid); // Run HLO cost analysis. auto hlo_module = BuildHloGraph(&builder); @@ -206,9 +206,8 @@ TEST_F(HloCostAnalysisTest, Convolution) { TEST_F(HloCostAnalysisTest, Reduce) { XlaBuilder builder("reduce"); auto input = - builder.Parameter(0, ShapeUtil::MakeShape(F32, {10, 20}), "input"); - auto result = - builder.Reduce(input, builder.ConstantR0(0.0f), add_, {1}); + Parameter(&builder, 0, ShapeUtil::MakeShape(F32, {10, 20}), "input"); + Reduce(input, ConstantR0(&builder, 0.0f), add_, {1}); // Run HLO cost analysis. auto hlo_module = BuildHloGraph(&builder); @@ -224,9 +223,9 @@ TEST_F(HloCostAnalysisTest, Reduce) { TEST_F(HloCostAnalysisTest, ReduceWindow) { XlaBuilder builder("reduce_window"); auto input = - builder.Parameter(0, ShapeUtil::MakeShape(F32, {10, 20}), "input"); - auto result = builder.ReduceWindow(input, builder.ConstantR0(0), add_, - {4, 5}, {4, 5}, Padding::kValid); + Parameter(&builder, 0, ShapeUtil::MakeShape(F32, {10, 20}), "input"); + ReduceWindow(input, ConstantR0(&builder, 0), add_, {4, 5}, {4, 5}, + Padding::kValid); // Run HLO cost analysis. auto hlo_module = BuildHloGraph(&builder); @@ -241,12 +240,11 @@ TEST_F(HloCostAnalysisTest, ReduceWindow) { TEST_F(HloCostAnalysisTest, SelectAndScatter) { XlaBuilder builder("select_and_scatter"); auto operand = - builder.Parameter(0, ShapeUtil::MakeShape(F32, {10, 20}), "input"); + Parameter(&builder, 0, ShapeUtil::MakeShape(F32, {10, 20}), "input"); auto source = - builder.Parameter(1, ShapeUtil::MakeShape(F32, {2, 4}), "source"); - auto result = - builder.SelectAndScatter(operand, gt_, {4, 5}, {4, 5}, Padding::kValid, - source, builder.ConstantR0(0), add_); + Parameter(&builder, 1, ShapeUtil::MakeShape(F32, {2, 4}), "source"); + SelectAndScatter(operand, gt_, {4, 5}, {4, 5}, Padding::kValid, source, + ConstantR0(&builder, 0), add_); // Run HLO cost analysis. auto hlo_module = BuildHloGraph(&builder); @@ -261,7 +259,7 @@ TEST_F(HloCostAnalysisTest, SelectAndScatter) { TEST_F(HloCostAnalysisTest, Broadcast) { XlaBuilder b("broadcast"); - b.Broadcast(b.ConstantR0(42), {10, 7}); + Broadcast(ConstantR0(&b, 42), {10, 7}); auto hlo_module = BuildHloGraph(&b); HloCostAnalysis analysis(ShapeSize); ASSERT_IS_OK( @@ -273,13 +271,12 @@ TEST_F(HloCostAnalysisTest, Broadcast) { TEST_F(HloCostAnalysisTest, FullyConnectedForward) { XlaBuilder builder("fully_connected_forward"); auto input = - builder.Parameter(0, ShapeUtil::MakeShape(F32, {10, 5}), "input"); + Parameter(&builder, 0, ShapeUtil::MakeShape(F32, {10, 5}), "input"); auto weight = - builder.Parameter(1, ShapeUtil::MakeShape(F32, {5, 20}), "weight"); - auto bias = builder.Parameter(2, ShapeUtil::MakeShape(F32, {20}), "bias"); + Parameter(&builder, 1, ShapeUtil::MakeShape(F32, {5, 20}), "weight"); + auto bias = Parameter(&builder, 2, ShapeUtil::MakeShape(F32, {20}), "bias"); // sigmoid(input * weight + bias) - auto result = builder.Map( - {builder.Add(builder.Dot(input, weight), bias, {1})}, sigmoid_, {0, 1}); + Map(&builder, {Add(Dot(input, weight), bias, {1})}, sigmoid_, {0, 1}); // Run HLO cost analysis. auto hlo_module = BuildHloGraph(&builder); @@ -297,11 +294,11 @@ TEST_F(HloCostAnalysisTest, MatmulAndConvolutionCanBeTheSameComputation) { HloCostAnalysis conv_analysis(ShapeSize); { XlaBuilder builder("conv_looking_matmul"); - auto lhs = builder.Parameter(0, ShapeUtil::MakeShape(F32, {64, 64, 1, 1}), - "input"); - auto rhs = builder.Parameter(1, ShapeUtil::MakeShape(F32, {64, 64, 1, 1}), - "weights"); - builder.Conv(lhs, rhs, {1, 1}, Padding::kSame); + auto lhs = Parameter(&builder, 0, ShapeUtil::MakeShape(F32, {64, 64, 1, 1}), + "input"); + auto rhs = Parameter(&builder, 1, ShapeUtil::MakeShape(F32, {64, 64, 1, 1}), + "weights"); + Conv(lhs, rhs, {1, 1}, Padding::kSame); auto hlo_module = BuildHloGraph(&builder); ASSERT_IS_OK(hlo_module->entry_computation()->root_instruction()->Accept( &conv_analysis)); @@ -311,10 +308,10 @@ TEST_F(HloCostAnalysisTest, MatmulAndConvolutionCanBeTheSameComputation) { { XlaBuilder builder("matmul"); auto lhs = - builder.Parameter(0, ShapeUtil::MakeShape(F32, {64, 64}), "input"); + Parameter(&builder, 0, ShapeUtil::MakeShape(F32, {64, 64}), "input"); auto rhs = - builder.Parameter(1, ShapeUtil::MakeShape(F32, {64, 64}), "weights"); - builder.Dot(lhs, rhs); + Parameter(&builder, 1, ShapeUtil::MakeShape(F32, {64, 64}), "weights"); + Dot(lhs, rhs); auto hlo_module = BuildHloGraph(&builder); ASSERT_IS_OK(hlo_module->entry_computation()->root_instruction()->Accept( &matmul_analysis)); @@ -341,13 +338,13 @@ TEST_F(FusionCostAnalysis, LoopFusion) { // tuple = Tuple({sub, sub, mul, C1}) HloComputation::Builder builder(TestName()); auto c1 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR2F32Linspace( + HloInstruction::CreateConstant(LiteralUtil::CreateR2F32Linspace( /*from=*/0.0f, /*to=*/1.0f, /*rows=*/2, /*cols=*/2))); auto c2 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR2F32Linspace( + HloInstruction::CreateConstant(LiteralUtil::CreateR2F32Linspace( /*from=*/1.0f, /*to=*/2.0f, /*rows=*/2, /*cols=*/2))); auto c3 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR2F32Linspace( + HloInstruction::CreateConstant(LiteralUtil::CreateR2F32Linspace( /*from=*/2.0f, /*to=*/3.0f, /*rows=*/2, /*cols=*/2))); auto add = builder.AddInstruction( HloInstruction::CreateBinary(r2f32, HloOpcode::kAdd, c1, c2)); @@ -394,9 +391,9 @@ TEST_F(FusionCostAnalysis, NoLayout) { HloComputation::Builder builder(TestName()); auto c1 = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR4FromArray4D(Array4D(2, 3, 4, 5)))); + LiteralUtil::CreateR4FromArray4D(Array4D(2, 3, 4, 5)))); auto c2 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR1({1, 2, 3}))); + HloInstruction::CreateConstant(LiteralUtil::CreateR1({1, 2, 3}))); auto broadcast = builder.AddInstruction( HloInstruction::CreateBroadcast(shape_without_layout, c2, {1})); @@ -419,9 +416,9 @@ TEST_F(HloCostAnalysisTest, TupleCost) { HloCostAnalysis analysis(ShapeSize); { XlaBuilder builder("matmul"); - auto x = builder.Parameter(0, ShapeUtil::MakeShape(F32, {123}), "x"); - auto y = builder.Parameter(1, ShapeUtil::MakeShape(F32, {42}), "y"); - auto tuple = builder.Tuple({x, y}); + auto x = Parameter(&builder, 0, ShapeUtil::MakeShape(F32, {123}), "x"); + auto y = Parameter(&builder, 1, ShapeUtil::MakeShape(F32, {42}), "y"); + Tuple(&builder, {x, y}); auto hlo_module = BuildHloGraph(&builder); ASSERT_IS_OK( @@ -435,21 +432,21 @@ TEST_F(HloCostAnalysisTest, TupleCost) { TEST_F(HloCostAnalysisTest, BaseDilatedConvolution) { XlaBuilder builder("BaseDilatedConvolution"); - auto input = builder.Parameter( - 0, + auto input = Parameter( + &builder, 0, ShapeUtil::MakeShape(F32, {/*p_dim=*/1, /*z_dim=*/1, /*y_dim=*/10, /*x_dim=*/20}), "input"); - auto kernel = builder.Parameter( - 1, + auto kernel = Parameter( + &builder, 1, ShapeUtil::MakeShape(F32, {/*p_dim=*/1, /*z_dim=*/1, /*y_dim=*/3, /*x_dim=*/3}), "kernel"); - auto result = builder.ConvGeneralDilated( - input, kernel, /*window_strides=*/{1, 1}, /*padding=*/{{1, 1}, {1, 1}}, - /*lhs_dilation=*/{3, 5}, /*rhs_dilation=*/{7, 11}, - XlaBuilder::CreateDefaultConvDimensionNumbers(2)); + ConvGeneralDilated(input, kernel, /*window_strides=*/{1, 1}, + /*padding=*/{{1, 1}, {1, 1}}, + /*lhs_dilation=*/{3, 5}, /*rhs_dilation=*/{7, 11}, + XlaBuilder::CreateDefaultConvDimensionNumbers(2)); // Run HLO cost analysis. auto hlo_module = BuildHloGraph(&builder); @@ -463,8 +460,8 @@ TEST_F(HloCostAnalysisTest, BaseDilatedConvolution) { TEST_F(HloCostAnalysisTest, Slice) { // Test the analysis on a slice. XlaBuilder builder("slice"); - auto x = builder.Parameter(0, ShapeUtil::MakeShape(F32, {2}), "x"); - auto slice = builder.Slice(x, {0}, {1}, {1}); + auto x = Parameter(&builder, 0, ShapeUtil::MakeShape(F32, {2}), "x"); + Slice(x, {0}, {1}, {1}); auto hlo_module = BuildHloGraph(&builder); // Run HLO cost analysis. @@ -478,8 +475,8 @@ TEST_F(HloCostAnalysisTest, Slice) { TEST_F(HloCostAnalysisTest, DynamicSlice) { // Test the analysis on a slice. XlaBuilder builder("dynamic-slice"); - auto x = builder.Parameter(0, ShapeUtil::MakeShape(F32, {2}), "x"); - auto slice = builder.DynamicSlice(x, builder.ConstantR1({1}), {1}); + auto x = Parameter(&builder, 0, ShapeUtil::MakeShape(F32, {2}), "x"); + DynamicSlice(x, ConstantR1(&builder, {1}), {1}); auto hlo_module = BuildHloGraph(&builder); // Run HLO cost analysis. @@ -493,9 +490,9 @@ TEST_F(HloCostAnalysisTest, DynamicSlice) { TEST_F(HloCostAnalysisTest, DynamicUpdateSlice) { // Test the analysis on a slice. XlaBuilder builder("dynamic-update-slice"); - auto x = builder.Parameter(0, ShapeUtil::MakeShape(F32, {2}), "x"); - auto slice = builder.DynamicUpdateSlice(x, builder.ConstantR1({1.0}), - builder.ConstantR1({1})); + auto x = Parameter(&builder, 0, ShapeUtil::MakeShape(F32, {2}), "x"); + DynamicUpdateSlice(x, ConstantR1(&builder, {1.0}), + ConstantR1(&builder, {1})); auto hlo_module = BuildHloGraph(&builder); // Run HLO cost analysis. diff --git a/tensorflow/compiler/xla/service/hlo_creation_utils.cc b/tensorflow/compiler/xla/service/hlo_creation_utils.cc index 0fb65c845a6d4407c81171f6c1569fee98b1d16d..90d2be118d94d52135820e5b8138fcb06389c684 100644 --- a/tensorflow/compiler/xla/service/hlo_creation_utils.cc +++ b/tensorflow/compiler/xla/service/hlo_creation_utils.cc @@ -14,6 +14,7 @@ limitations under the License. ==============================================================================*/ #include "tensorflow/compiler/xla/service/hlo_creation_utils.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/literal_util.h" #include "tensorflow/compiler/xla/ptr_util.h" #include "tensorflow/compiler/xla/service/shape_inference.h" @@ -261,9 +262,9 @@ StatusOr PadVectorWithZeros(HloInstruction* operand, padding_config_dim.set_edge_padding_high(zeros_to_append); *padding_config.add_dimensions() = padding_config_dim; - HloInstruction* zero = - computation->AddInstruction(HloInstruction::CreateConstant( - MakeUnique(Literal::Zero(operand->shape().element_type())))); + HloInstruction* zero = computation->AddInstruction( + HloInstruction::CreateConstant(MakeUnique( + LiteralUtil::Zero(operand->shape().element_type())))); return MakePadHlo(operand, zero, padding_config); } @@ -272,7 +273,7 @@ StatusOr BroadcastZeros( ArraySlice broadcast_dimensions) { HloInstruction* zero = computation->AddInstruction(HloInstruction::CreateConstant( - MakeUnique(Literal::Zero(element_type)))); + MakeUnique(LiteralUtil::Zero(element_type)))); return MakeBroadcastHlo(zero, /*broadcast_dimensions=*/{}, /*result_shape_bounds=*/broadcast_dimensions); } diff --git a/tensorflow/compiler/xla/service/hlo_creation_utils_test.cc b/tensorflow/compiler/xla/service/hlo_creation_utils_test.cc index 7e7c4f95fed737f40064224717f409b934e4ff27..60d3e71757d5ce31e025c744e089ff56091d9a43 100644 --- a/tensorflow/compiler/xla/service/hlo_creation_utils_test.cc +++ b/tensorflow/compiler/xla/service/hlo_creation_utils_test.cc @@ -60,8 +60,8 @@ TEST_F(HloCreationUtilsTest, CollapseFirst1Dim) { HloEvaluator evaluator; TF_ASSERT_OK_AND_ASSIGN(std::unique_ptr result_literal, evaluator.Evaluate>( - *module, {Literal::CreateR1({3, 4})})); - CHECK_EQ(*result_literal, *Literal::CreateR1({3, 4})); + *module, {LiteralUtil::CreateR1({3, 4})})); + CHECK_EQ(*result_literal, *LiteralUtil::CreateR1({3, 4})); } TEST_F(HloCreationUtilsTest, CollapseFirst2Dims) { @@ -82,10 +82,10 @@ TEST_F(HloCreationUtilsTest, CollapseFirst2Dims) { std::unique_ptr result_literal, evaluator.Evaluate>( *module, - {Literal::CreateR3( + {LiteralUtil::CreateR3( {{{1, 2}, {3, 4}, {5, 6}}, {{-1, -2}, {-3, -4}, {-5, -6}}})})); CHECK_EQ(*result_literal, - *Literal::CreateR2( + *LiteralUtil::CreateR2( {{1, 2}, {3, 4}, {5, 6}, {-1, -2}, {-3, -4}, {-5, -6}})); } @@ -103,10 +103,11 @@ TEST_F(HloCreationUtilsTest, Prepend1DegenerateDim) { entry_computation->set_root_instruction(with_1_degenerate_dim_prepended); HloEvaluator evaluator; - TF_ASSERT_OK_AND_ASSIGN(std::unique_ptr result_literal, - evaluator.Evaluate>( - *module, {Literal::CreateR1({9, 10})})); - CHECK_EQ(*result_literal, *Literal::CreateR2({{9, 10}})); + TF_ASSERT_OK_AND_ASSIGN( + std::unique_ptr result_literal, + evaluator.Evaluate>( + *module, {LiteralUtil::CreateR1({9, 10})})); + CHECK_EQ(*result_literal, *LiteralUtil::CreateR2({{9, 10}})); } TEST_F(HloCreationUtilsTest, Prepend2DegenerateDims) { @@ -123,10 +124,11 @@ TEST_F(HloCreationUtilsTest, Prepend2DegenerateDims) { entry_computation->set_root_instruction(with_2_degenerate_dims_prepended); HloEvaluator evaluator; - TF_ASSERT_OK_AND_ASSIGN(std::unique_ptr result_literal, - evaluator.Evaluate>( - *module, {Literal::CreateR1({9, 10})})); - CHECK_EQ(*result_literal, *Literal::CreateR3({{{9, 10}}})); + TF_ASSERT_OK_AND_ASSIGN( + std::unique_ptr result_literal, + evaluator.Evaluate>( + *module, {LiteralUtil::CreateR1({9, 10})})); + CHECK_EQ(*result_literal, *LiteralUtil::CreateR3({{{9, 10}}})); } TEST_F(HloCreationUtilsTest, Prepend2DegenerateDimsToScalar) { @@ -145,8 +147,8 @@ TEST_F(HloCreationUtilsTest, Prepend2DegenerateDimsToScalar) { HloEvaluator evaluator; TF_ASSERT_OK_AND_ASSIGN(std::unique_ptr result_literal, evaluator.Evaluate>( - *module, {Literal::CreateR0(9)})); - CHECK_EQ(*result_literal, *Literal::CreateR2({{9}})); + *module, {LiteralUtil::CreateR0(9)})); + CHECK_EQ(*result_literal, *LiteralUtil::CreateR2({{9}})); } TEST_F(HloCreationUtilsTest, ExpandFirstDimInto3Dims) { @@ -166,9 +168,9 @@ TEST_F(HloCreationUtilsTest, ExpandFirstDimInto3Dims) { TF_ASSERT_OK_AND_ASSIGN( std::unique_ptr result_literal, evaluator.Evaluate>( - *module, {Literal::CreateR1({1, 2, 3, 4, 5, 6})})); + *module, {LiteralUtil::CreateR1({1, 2, 3, 4, 5, 6})})); CHECK_EQ(*result_literal, - *Literal::CreateR3({{{1, 2}}, {{3, 4}}, {{5, 6}}})); + *LiteralUtil::CreateR3({{{1, 2}}, {{3, 4}}, {{5, 6}}})); } TEST_F(HloCreationUtilsTest, PadVectorWithZeros) { @@ -188,8 +190,8 @@ TEST_F(HloCreationUtilsTest, PadVectorWithZeros) { HloEvaluator evaluator; TF_ASSERT_OK_AND_ASSIGN(std::unique_ptr result_literal, evaluator.Evaluate>( - *module, {Literal::CreateR1({3, 4})})); - CHECK_EQ(*result_literal, *Literal::CreateR1({0, 0, 0, 3, 4, 0})); + *module, {LiteralUtil::CreateR1({3, 4})})); + CHECK_EQ(*result_literal, *LiteralUtil::CreateR1({0, 0, 0, 3, 4, 0})); } TEST_F(HloCreationUtilsTest, BroadcastZeros_S32) { @@ -209,8 +211,8 @@ TEST_F(HloCreationUtilsTest, BroadcastZeros_S32) { HloEvaluator evaluator; TF_ASSERT_OK_AND_ASSIGN(std::unique_ptr result_literal, evaluator.Evaluate>( - *module, {Literal::CreateR0(0)})); - CHECK_EQ(*result_literal, *Literal::CreateR2({{0, 0}, {0, 0}})); + *module, {LiteralUtil::CreateR0(0)})); + CHECK_EQ(*result_literal, *LiteralUtil::CreateR2({{0, 0}, {0, 0}})); } TEST_F(HloCreationUtilsTest, BroadcastZeros_F32) { @@ -230,9 +232,9 @@ TEST_F(HloCreationUtilsTest, BroadcastZeros_F32) { HloEvaluator evaluator; TF_ASSERT_OK_AND_ASSIGN(std::unique_ptr result_literal, evaluator.Evaluate>( - *module, {Literal::CreateR0(0.0f)})); + *module, {LiteralUtil::CreateR0(0.0f)})); CHECK_EQ(*result_literal, - *Literal::CreateR2({{0.0f, 0.0f}, {0.0f, 0.0f}})); + *LiteralUtil::CreateR2({{0.0f, 0.0f}, {0.0f, 0.0f}})); } } // namespace diff --git a/tensorflow/compiler/xla/service/hlo_cse.cc b/tensorflow/compiler/xla/service/hlo_cse.cc index a0ee8896230d6dcacb5a8eb607fc00ae5226cfa5..06484f4012fc091f70df7bc8ec231ce3fcf89669 100644 --- a/tensorflow/compiler/xla/service/hlo_cse.cc +++ b/tensorflow/compiler/xla/service/hlo_cse.cc @@ -24,7 +24,7 @@ limitations under the License. #include #include "tensorflow/compiler/xla/layout_util.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/service/hlo_computation.h" #include "tensorflow/compiler/xla/service/hlo_domain_map.h" #include "tensorflow/compiler/xla/service/hlo_instruction.h" @@ -143,10 +143,8 @@ StatusOr HloCSE::Run(HloModule* module) { if (instruction->operand_count() == 0) { continue; } - // Skip instructions which have side effects or are a domain (which must - // not be CSE-ed). - if (instruction->HasSideEffect() || - instruction->opcode() == HloOpcode::kDomain) { + // Skip instructions which have side effects. + if (instruction->HasSideEffect()) { continue; } diff --git a/tensorflow/compiler/xla/service/hlo_cse_test.cc b/tensorflow/compiler/xla/service/hlo_cse_test.cc index 16db374566c727f1f3efe2a6d419f1f3caf0aaf1..76b9c66651018089b6ca55b04e12df7c19ebbfb9 100644 --- a/tensorflow/compiler/xla/service/hlo_cse_test.cc +++ b/tensorflow/compiler/xla/service/hlo_cse_test.cc @@ -21,7 +21,7 @@ limitations under the License. #include #include "tensorflow/compiler/xla/layout_util.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/ptr_util.h" #include "tensorflow/compiler/xla/service/hlo_computation.h" #include "tensorflow/compiler/xla/service/hlo_instruction.h" @@ -53,9 +53,9 @@ TEST_F(HloCseTest, CombineTwoConstants) { // Test that two identical constants are commoned. auto builder = HloComputation::Builder(TestName()); auto constant1 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(42.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(42.0f))); auto constant2 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(42.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(42.0f))); builder.AddInstruction(HloInstruction::CreateBinary( constant1->shape(), HloOpcode::kAdd, constant1, constant2)); @@ -72,7 +72,7 @@ TEST_F(HloCseTest, CombineTwoConstants) { EXPECT_EQ(42.0f, constant->literal().Get({})); auto result = ExecuteAndTransfer(std::move(module), {}); - auto expected = Literal::CreateR0(84.0); + auto expected = LiteralUtil::CreateR0(84.0); EXPECT_TRUE(LiteralTestUtil::Near(*expected, *result, ErrorSpec(1e-4))); } @@ -81,10 +81,10 @@ TEST_F(HloCseTest, CombineTwoConstantsDifferentLayoutsAndInsensitive) { // the pass is not layout sensitive. auto builder = HloComputation::Builder(TestName()); auto constant1 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR2WithLayout( + HloInstruction::CreateConstant(LiteralUtil::CreateR2WithLayout( {{1.0, 2.0}, {3.0, 4.0}}, LayoutUtil::MakeLayout({0, 1})))); auto constant2 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR2WithLayout( + HloInstruction::CreateConstant(LiteralUtil::CreateR2WithLayout( {{1.0, 2.0}, {3.0, 4.0}}, LayoutUtil::MakeLayout({1, 0})))); auto add = builder.AddInstruction(HloInstruction::CreateBinary( constant1->shape(), HloOpcode::kAdd, constant1, constant2)); @@ -104,7 +104,7 @@ TEST_F(HloCseTest, CombineTwoConstantsDifferentLayoutsAndInsensitive) { EXPECT_THAT(add, op::Add(first_operand, first_operand)); auto result = ExecuteAndTransfer(std::move(module), {}); - auto expected = Literal::CreateR2({{2.0, 4.0}, {6.0, 8.0}}); + auto expected = LiteralUtil::CreateR2({{2.0, 4.0}, {6.0, 8.0}}); EXPECT_TRUE(LiteralTestUtil::Near(*expected, *result, ErrorSpec(1e-4))); } @@ -113,10 +113,10 @@ TEST_F(HloCseTest, CombineTwoConstantsDifferentLayoutsAndSensitive) { // if the pass is layout sensitive. auto builder = HloComputation::Builder(TestName()); auto constant1 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR2WithLayout( + HloInstruction::CreateConstant(LiteralUtil::CreateR2WithLayout( {{1.0, 2.0}, {3.0, 4.0}}, LayoutUtil::MakeLayout({0, 1})))); auto constant2 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR2WithLayout( + HloInstruction::CreateConstant(LiteralUtil::CreateR2WithLayout( {{1.0, 2.0}, {3.0, 4.0}}, LayoutUtil::MakeLayout({1, 0})))); auto add = builder.AddInstruction(HloInstruction::CreateBinary( constant1->shape(), HloOpcode::kAdd, constant1, constant2)); @@ -134,7 +134,7 @@ TEST_F(HloCseTest, CombineTwoConstantsDifferentLayoutsAndSensitive) { EXPECT_THAT(add, op::Add(constant1, constant2)); auto result = ExecuteAndTransfer(std::move(module), {}); - auto expected = Literal::CreateR2({{2.0, 4.0}, {6.0, 8.0}}); + auto expected = LiteralUtil::CreateR2({{2.0, 4.0}, {6.0, 8.0}}); EXPECT_TRUE(LiteralTestUtil::Near(*expected, *result, ErrorSpec(1e-4))); } @@ -144,20 +144,20 @@ TEST_F(HloCseTest, ConstantsSameValueDifferentType) { auto builder = HloComputation::Builder(TestName()); std::vector constants; constants.push_back(builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(42)))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(42)))); constants.push_back(builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(42)))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(42)))); constants.push_back(builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(42.0)))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(42.0)))); constants.push_back(builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(42.0)))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(42.0)))); constants.push_back(builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(42.0)))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(42.0)))); constants.push_back(builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(42.0f)))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(42.0f)))); // Duplicate the float constant to verify something happens. constants.push_back(builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(42.0f)))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(42.0f)))); const Shape shape_r0 = ShapeUtil::MakeShape(F32, {}); for (int64 i = 0; i < constants.size(); ++i) { @@ -188,13 +188,13 @@ TEST_F(HloCseTest, NonscalarConstants) { // Test that identical nonscalar constants are merged. auto builder = HloComputation::Builder(TestName()); auto common_constant1 = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR2({{1.0, 2.0}, {3.0, 4.0}}))); + LiteralUtil::CreateR2({{1.0, 2.0}, {3.0, 4.0}}))); auto common_constant2 = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR2({{1.0, 2.0}, {3.0, 4.0}}))); + LiteralUtil::CreateR2({{1.0, 2.0}, {3.0, 4.0}}))); // Create a constant which has the same shape but a different value. auto uncommon_constant = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR2({{2.0, 4.0}, {6.0, 8.0}}))); + LiteralUtil::CreateR2({{2.0, 4.0}, {6.0, 8.0}}))); // Tie the constants together with a tuple. This makes it easier to refer to // the constant instructions via their use. @@ -223,7 +223,7 @@ TEST_F(HloCseTest, IdenticalInstructions) { // Test that three identical instructions are commoned. auto builder = HloComputation::Builder(TestName()); auto constant = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(42.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(42.0))); auto exp1 = builder.AddInstruction(HloInstruction::CreateUnary( constant->shape(), HloOpcode::kExp, constant)); auto exp2 = builder.AddInstruction(HloInstruction::CreateUnary( @@ -253,7 +253,7 @@ TEST_F(HloCseTest, IdenticalInstructionsDifferentLayoutsSensitive) { // commoned if the pass is layout sensitive. auto builder = HloComputation::Builder(TestName()); auto constant = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR2({{1.0, 2.0}, {3.0, 4.0}}))); + LiteralUtil::CreateR2({{1.0, 2.0}, {3.0, 4.0}}))); auto exp1 = builder.AddInstruction(HloInstruction::CreateUnary( constant->shape(), HloOpcode::kExp, constant)); @@ -284,7 +284,7 @@ TEST_F(HloCseTest, IdenticalInstructionsDifferentLayoutsInsensitive) { // the pass is layout insensitive. auto builder = HloComputation::Builder(TestName()); auto constant = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR2({{1.0, 2.0}, {3.0, 4.0}}))); + LiteralUtil::CreateR2({{1.0, 2.0}, {3.0, 4.0}}))); auto exp1 = builder.AddInstruction(HloInstruction::CreateUnary( constant->shape(), HloOpcode::kExp, constant)); @@ -362,7 +362,7 @@ TEST_F(HloCseTest, IdenticalExpressions) { // The *1 instructions should be merged with the *2 instructions. auto builder = HloComputation::Builder(TestName()); auto constant = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(42.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(42.0))); auto negate1 = builder.AddInstruction(HloInstruction::CreateUnary( constant->shape(), HloOpcode::kNegate, constant)); @@ -400,9 +400,9 @@ TEST_F(HloCseTest, DoNotCombineRng) { // Test that two RNG ops are not commoned. auto builder = HloComputation::Builder(TestName()); auto constant1 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(0.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(0.0f))); auto constant2 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0f))); auto rng1 = builder.AddInstruction(HloInstruction::CreateRng( ShapeUtil::MakeShape(F32, {}), RandomDistribution::RNG_UNIFORM, {constant1, constant2})); @@ -442,9 +442,9 @@ TEST_F(HloCseTest, DoNotCombineCallsToImpureFunctions) { Shape scalar_shape = ShapeUtil::MakeShape(F32, {}); auto builder = HloComputation::Builder(TestName() + "_rng_fun"); auto constant1 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(0.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(0.0f))); auto constant2 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0f))); auto rng = builder.AddInstruction(HloInstruction::CreateRng( scalar_shape, RandomDistribution::RNG_UNIFORM, {constant1, constant2})); auto param = builder.AddInstruction(HloInstruction::CreateParameter( @@ -459,7 +459,7 @@ TEST_F(HloCseTest, DoNotCombineCallsToImpureFunctions) { { auto builder = HloComputation::Builder(TestName()); auto constant = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR1({5.0f}))); + HloInstruction::CreateConstant(LiteralUtil::CreateR1({5.0f}))); auto rng1 = builder.AddInstruction( HloInstruction::CreateMap(constant->shape(), {constant}, rng_function)); auto rng2 = builder.AddInstruction( @@ -521,9 +521,9 @@ TEST_F(HloCseTest, ConstantsSameValueInDifferentDomains) { // in this case) are not collapsed. auto builder = HloComputation::Builder(TestName()); builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(42))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(42))); builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(42))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(42))); auto module = CreateNewModule(); auto computation = module->AddEntryComputation(builder.Build()); @@ -536,5 +536,40 @@ TEST_F(HloCseTest, ConstantsSameValueInDifferentDomains) { EXPECT_EQ(2, computation->instruction_count()); } +TEST_F(HloCseTest, Domain) { + auto module = ParseHloString(R"( +HloModule module +ENTRY %entry { + %param = f32[] parameter(0), sharding={maximal device=0} + %domain.0 = f32[] domain(%param), + domain={kind="sharding", entry={maximal device=0}, exit={maximal device=1}} + %domain.1 = f32[] domain(%param), + domain={kind="sharding", entry={maximal device=0}, exit={maximal device=1}} + %domain.2 = f32[] domain(%param), + domain={kind="sharding", entry={maximal device=0}, exit={maximal device=2}} + %negate.0 = f32[] negate(%domain.0) + %negate.1 = f32[] negate(%domain.1) + %negate.2 = f32[] negate(%domain.2) + %domain.3 = f32[] domain(%negate.0), + domain={kind="sharding", entry={maximal device=1}, exit={maximal device=0}} + %domain.4 = f32[] domain(%negate.1), + domain={kind="sharding", entry={maximal device=1}, exit={maximal device=0}} + %domain.5 = f32[] domain(%negate.2), + domain={kind="sharding", entry={maximal device=2}, exit={maximal device=0}} + %add = f32[] add(%domain.3, %domain.4) + ROOT %sub = f32[] subtract(%add, %domain.5) +})") + .ValueOrDie(); + + HloCSE cse(/*is_layout_sensitive=*/false); + EXPECT_TRUE(cse.Run(module.get()).ValueOrDie()); + LOG(INFO) << "AAAAA " << module->ToString(); + const HloInstruction* sub = module->entry_computation()->root_instruction(); + const HloInstruction* add = sub->operand(0); + EXPECT_EQ(add->operand(0), add->operand(1)); + EXPECT_NE(add->operand(0), sub->operand(1)); + EXPECT_NE(add->operand(1), sub->operand(1)); +} + } // namespace } // namespace xla diff --git a/tensorflow/compiler/xla/service/hlo_dataflow_analysis.cc b/tensorflow/compiler/xla/service/hlo_dataflow_analysis.cc index f529c0dad7374c05adee9eb611d4fcbd2dc8fbcf..de1a32d8bd9217baabda4ab4b02bf28baebad531 100644 --- a/tensorflow/compiler/xla/service/hlo_dataflow_analysis.cc +++ b/tensorflow/compiler/xla/service/hlo_dataflow_analysis.cc @@ -398,18 +398,17 @@ bool HloDataflowAnalysis::UpdateSendValueSet(HloInstruction* send) { bool HloDataflowAnalysis::UpdateRecvDoneValueSet(HloInstruction* recv_done) { CHECK_EQ(recv_done->opcode(), HloOpcode::kRecvDone); bool changed = false; - // RecvDone forwards the operand value at {0} to the output. + // RecvDone forwards the operand value at {0} to element {0} of its output. for (auto& pair : GetInstructionValueSet(recv_done)) { ShapeIndex& index = pair.first; HloValueSet& value_set = pair.second; - ShapeIndex operand_index = {0}; - for (int64 i : index) { - operand_index.push_back(i); + if (index.empty() || index[0] != 0) { + continue; } const HloValueSet& operand_value_set = - GetValueSet(recv_done->operand(0), operand_index); + GetValueSet(recv_done->operand(0), index); if (value_set != operand_value_set) { value_set = operand_value_set; changed = true; @@ -466,6 +465,24 @@ bool HloDataflowAnalysis::UpdateCopyValueSet(HloInstruction* copy) { return changed; } +bool HloDataflowAnalysis::UpdateDomainValueSet(HloInstruction* domain) { + // Domain instructions just forward their operand. Given that domains can have + // a tuple operand, we iterate through its indexes, like for copies. + // Unlike copies though we also propagate the top-level value. + CHECK_EQ(domain->opcode(), HloOpcode::kDomain); + bool changed = false; + for (auto& pair : GetInstructionValueSet(domain)) { + const ShapeIndex& index = pair.first; + HloValueSet& value_set = pair.second; + HloValueSet& operand_value_set = GetValueSet(domain->operand(0), index); + if (value_set != operand_value_set) { + value_set = operand_value_set; + changed = true; + } + } + return changed; +} + bool HloDataflowAnalysis::UpdateGetTupleElementValueSet(HloInstruction* gte) { CHECK_EQ(gte->opcode(), HloOpcode::kGetTupleElement); bool changed = false; @@ -560,17 +577,17 @@ bool HloDataflowAnalysis::UpdateParameterValueSet(HloInstruction* parameter) { } } -bool HloDataflowAnalysis::UpdateSelectValueSet(HloInstruction* select) { - CHECK_EQ(select->opcode(), HloOpcode::kSelect); - // A phi value is not defined at a kSelect instruction because kSelect does - // not create a new value. Rather it forwards a value from its operands. This - // contrasts with kWhile instruction (which does define a phi value) which has - // in-place update semantics. +bool HloDataflowAnalysis::UpdateTupleSelectValueSet(HloInstruction* select) { + CHECK_EQ(select->opcode(), HloOpcode::kTupleSelect); + // A phi value is not defined at a kTupleSelect instruction because + // kTupleSelect does not create a new value. Rather it forwards a value from + // its operands. This contrasts with kWhile instruction (which does define a + // phi value) which has in-place update semantics. bool changed = false; for (auto& pair : GetInstructionValueSet(select)) { const ShapeIndex& index = pair.first; if (index.empty()) { - // kSelect copies (not forwards) the top-level value. + // kTupleSelect copies (not forwards) the top-level value. continue; } HloValueSet& value_set = pair.second; @@ -626,12 +643,14 @@ bool HloDataflowAnalysis::UpdateInstructionValueSet( return UpdateBitcastValueSet(instruction); case HloOpcode::kSlice: return UpdateSliceValueSet(instruction); + case HloOpcode::kDomain: + return UpdateDomainValueSet(instruction); case HloOpcode::kCopy: return UpdateCopyValueSet(instruction); case HloOpcode::kGetTupleElement: return UpdateGetTupleElementValueSet(instruction); - case HloOpcode::kSelect: - return UpdateSelectValueSet(instruction); + case HloOpcode::kTupleSelect: + return UpdateTupleSelectValueSet(instruction); case HloOpcode::kTuple: return UpdateTupleValueSet(instruction); case HloOpcode::kParameter: @@ -804,6 +823,7 @@ Status HloDataflowAnalysis::InitializeInstructionValueSets() { case HloOpcode::kCall: case HloOpcode::kConditional: case HloOpcode::kGetTupleElement: + case HloOpcode::kDomain: // These instructions define no values. The values in their output // flow from their operands or from cross computation dataflow. break; @@ -829,21 +849,25 @@ Status HloDataflowAnalysis::InitializeInstructionValueSets() { } break; case HloOpcode::kCopy: - case HloOpcode::kSelect: + case HloOpcode::kTupleSelect: case HloOpcode::kTuple: // These instructions only define their top-level values. Any other // values flow from their operands. define_top_level_only(); break; case HloOpcode::kRecvDone: - // RecvDone aliases its input tuple element {0}, therefore does not - // define any values. + // RecvDone produces a two-element tuple. Element zero aliases its + // input tuple element {0}; element one is a token. + define_value_at(/*index=*/{}); + define_value_at(/*index=*/{1}); break; case HloOpcode::kSend: - // Send produces a tuple of {aliased operand, U32 context}, therefore - // only defines the top-level tuple and the tuple element at {1}. + // Send produces a tuple of {aliased operand, U32 context, token}, + // therefore only defines the top-level tuple and the tuple elements + // at {1} and {2}. define_value_at(/*index=*/{}); define_value_at(/*index=*/{1}); + define_value_at(/*index=*/{2}); break; default: define_all_values(); diff --git a/tensorflow/compiler/xla/service/hlo_dataflow_analysis.h b/tensorflow/compiler/xla/service/hlo_dataflow_analysis.h index 3d2d5baa773a732f32d6bfbf50d4764c751cd96b..f4abc7a7c7dcfb223067fe946bec0c5ef32f206b 100644 --- a/tensorflow/compiler/xla/service/hlo_dataflow_analysis.h +++ b/tensorflow/compiler/xla/service/hlo_dataflow_analysis.h @@ -185,10 +185,11 @@ class HloDataflowAnalysis { bool UpdateCallValueSet(HloInstruction* call); bool UpdateConditionalValueSet(HloInstruction* conditional); bool UpdateCopyValueSet(HloInstruction* copy); + bool UpdateDomainValueSet(HloInstruction* domain); bool UpdateGetTupleElementValueSet(HloInstruction* gte); bool UpdateParameterValueSet(HloInstruction* parameter); bool UpdateRecvDoneValueSet(HloInstruction* recv_done); - bool UpdateSelectValueSet(HloInstruction* select); + bool UpdateTupleSelectValueSet(HloInstruction* select); bool UpdateSendValueSet(HloInstruction* send); bool UpdateTupleValueSet(HloInstruction* tuple); bool UpdateWhileValueSet(HloInstruction* xla_while); diff --git a/tensorflow/compiler/xla/service/hlo_dataflow_analysis_test.cc b/tensorflow/compiler/xla/service/hlo_dataflow_analysis_test.cc index 0ea8bdcab680a40fd9301f2dcd5e0e176ac73d15..37bc2d2c9d2a0d0624917337b36c5d5f625c0991 100644 --- a/tensorflow/compiler/xla/service/hlo_dataflow_analysis_test.cc +++ b/tensorflow/compiler/xla/service/hlo_dataflow_analysis_test.cc @@ -15,7 +15,7 @@ limitations under the License. #include "tensorflow/compiler/xla/service/hlo_dataflow_analysis.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/service/hlo_computation.h" #include "tensorflow/compiler/xla/service/hlo_graph_dumper.h" #include "tensorflow/compiler/xla/service/hlo_matchers.h" @@ -101,9 +101,9 @@ TEST_P(HloDataflowAnalysisTest, BinaryOperation) { // Test the dataflow for a simple binary operation (Add). auto builder = HloComputation::Builder(TestName()); auto constant1 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0))); auto constant2 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(2.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(2.0))); auto add = builder.AddInstruction(HloInstruction::CreateBinary( scalar_shape_, HloOpcode::kAdd, constant1, constant2)); module_->AddEntryComputation(builder.Build()); @@ -198,9 +198,9 @@ TEST_P(HloDataflowAnalysisTest, NestedTuple) { // Verify the dataflow through a nested tuple. auto builder = HloComputation::Builder(TestName()); auto constant1 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0))); auto constant2 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(2.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(2.0))); auto tuple = builder.AddInstruction( HloInstruction::CreateTuple({constant1, constant2})); auto nested_tuple = builder.AddInstruction( @@ -259,9 +259,9 @@ TEST_P(HloDataflowAnalysisTest, SingleCall) { auto builder = HloComputation::Builder(TestName()); auto constant1 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0))); auto constant2 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(2.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(2.0))); auto call = builder.AddInstruction(HloInstruction::CreateCall( scalar_shape_, {constant1, constant2}, called_computation)); module_->AddEntryComputation(builder.Build()); @@ -308,9 +308,9 @@ TEST_P(HloDataflowAnalysisTest, ComputationCalledTwiceWithSameArguments) { auto builder = HloComputation::Builder(TestName()); auto constant1 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0))); auto constant2 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(2.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(2.0))); auto call1 = builder.AddInstruction(HloInstruction::CreateCall( scalar_shape_, {constant1, constant2}, called_computation)); auto call2 = builder.AddInstruction(HloInstruction::CreateCall( @@ -362,9 +362,9 @@ TEST_P(HloDataflowAnalysisTest, ComputationCalledTwiceWithDifferentArguments) { auto builder = HloComputation::Builder(TestName()); auto constant1 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0))); auto constant2 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(2.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(2.0))); auto call1 = builder.AddInstruction(HloInstruction::CreateCall( scalar_shape_, {constant1, constant2}, called_computation)); auto call2 = builder.AddInstruction(HloInstruction::CreateCall( @@ -426,9 +426,9 @@ TEST_P(HloDataflowAnalysisTest, NestedCalls) { auto builder = HloComputation::Builder(TestName()); auto constant1 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0))); auto constant2 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(2.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(2.0))); auto call = builder.AddInstruction(HloInstruction::CreateCall( scalar_shape_, {constant1, constant2}, outer_computation)); module_->AddEntryComputation(builder.Build()); @@ -493,15 +493,15 @@ TEST_P(HloDataflowAnalysisTest, SingleWhile) { auto cond_param = cond_builder.AddInstruction( HloInstruction::CreateParameter(0, tuple_shape, "param")); auto cond_constant = cond_builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(false))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(false))); HloComputation* condition = module_->AddEmbeddedComputation(cond_builder.Build()); auto builder = HloComputation::Builder(TestName()); auto constant1 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0))); auto constant2 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(2.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(2.0))); auto tuple = builder.AddInstruction( HloInstruction::CreateTuple({constant1, constant2})); auto xla_while = builder.AddInstruction( @@ -594,15 +594,15 @@ TEST_P(HloDataflowAnalysisTest, SequentialWhiles) { cond_builder.AddInstruction( HloInstruction::CreateParameter(0, tuple_shape, "param")); cond_builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(false))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(false))); HloComputation* condition = module_->AddEmbeddedComputation(cond_builder.Build()); auto builder = HloComputation::Builder(TestName()); auto constant1 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0))); auto constant2 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(2.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(2.0))); auto tuple = builder.AddInstruction( HloInstruction::CreateTuple({constant1, constant2})); auto xla_while0 = builder.AddInstruction( @@ -653,7 +653,7 @@ TEST_P(HloDataflowAnalysisTest, NestedWhiles) { cond_builder.AddInstruction( HloInstruction::CreateParameter(0, tuple_shape, "param")); cond_builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(false))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(false))); HloComputation* condition = module_->AddEmbeddedComputation(cond_builder.Build()); @@ -691,9 +691,9 @@ TEST_P(HloDataflowAnalysisTest, NestedWhiles) { auto builder = HloComputation::Builder(TestName()); auto constant1 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0))); auto constant2 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(2.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(2.0))); auto tuple = builder.AddInstruction( HloInstruction::CreateTuple({constant1, constant2})); auto entry_while = builder.AddInstruction( @@ -780,15 +780,15 @@ TEST_P(HloDataflowAnalysisTest, SwizzlingWhile) { auto cond_param = cond_builder.AddInstruction( HloInstruction::CreateParameter(0, tuple_shape, "param")); cond_builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(false))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(false))); HloComputation* condition = module_->AddEmbeddedComputation(cond_builder.Build()); auto builder = HloComputation::Builder(TestName()); auto constant1 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0))); auto constant2 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(2.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(2.0))); auto tuple = builder.AddInstruction( HloInstruction::CreateTuple({constant1, constant2})); auto xla_while = builder.AddInstruction( @@ -840,11 +840,11 @@ TEST_P(HloDataflowAnalysisTest, ArraySelect) { // Test a kSelect of an array value. auto builder = HloComputation::Builder(TestName()); auto pred = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(false))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(false))); auto constant1 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0))); auto constant2 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(2.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(2.0))); auto select = builder.AddInstruction(HloInstruction::CreateTernary( scalar_shape_, HloOpcode::kSelect, pred, constant1, constant2)); @@ -860,19 +860,18 @@ TEST_P(HloDataflowAnalysisTest, ArraySelect) { } TEST_P(HloDataflowAnalysisTest, TupleSelect) { - // Test a kSelect of a tuple value. Non-top-level element flow through the - // instruction. + // Test a kTupleSelect. Non-top-level element flow through the instruction. auto builder = HloComputation::Builder(TestName()); auto pred = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(false))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(false))); auto constant1 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0))); auto constant2 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(2.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(2.0))); auto constant3 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(3.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(3.0))); auto constant4 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(4.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(4.0))); auto tuple1 = builder.AddInstruction(HloInstruction::CreateTuple({constant1})); auto tuple2 = @@ -883,20 +882,20 @@ TEST_P(HloDataflowAnalysisTest, TupleSelect) { builder.AddInstruction(HloInstruction::CreateTuple({constant4})); const Shape tuple_shape = tuple1->shape(); auto select11 = builder.AddInstruction(HloInstruction::CreateTernary( - tuple_shape, HloOpcode::kSelect, pred, tuple1, tuple1)); + tuple_shape, HloOpcode::kTupleSelect, pred, tuple1, tuple1)); auto select12 = builder.AddInstruction(HloInstruction::CreateTernary( - tuple_shape, HloOpcode::kSelect, pred, tuple1, tuple2)); + tuple_shape, HloOpcode::kTupleSelect, pred, tuple1, tuple2)); auto select34 = builder.AddInstruction(HloInstruction::CreateTernary( - tuple_shape, HloOpcode::kSelect, pred, tuple3, tuple4)); + tuple_shape, HloOpcode::kTupleSelect, pred, tuple3, tuple4)); auto select1234 = builder.AddInstruction(HloInstruction::CreateTernary( - tuple_shape, HloOpcode::kSelect, pred, select12, select34)); + tuple_shape, HloOpcode::kTupleSelect, pred, select12, select34)); module_->AddEntryComputation(builder.Build()); bool ssa_form = GetParam(); const HloDataflowAnalysis& analysis = RunAnalysis(ssa_form); - // Top-level value is always defined by a kSelect. + // Top-level value is always defined by a kTupleSelect. EXPECT_TRUE(analysis.ValueIsDefinedAt(select11)); EXPECT_TRUE(analysis.ValueIsDefinedAt(select12)); EXPECT_TRUE(analysis.ValueIsDefinedAt(select34)); @@ -937,20 +936,20 @@ TEST_P(HloDataflowAnalysisTest, TupleSelect) { } TEST_P(HloDataflowAnalysisTest, NestedTupleSelect) { - // Test kSelect of a nested tuple. + // Test kTupleSelect of a nested tuple. auto builder = HloComputation::Builder(TestName()); auto pred = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(false))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(false))); auto constant1 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0))); auto constant2 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(2.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(2.0))); auto constant3 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(3.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(3.0))); auto constant4 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(4.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(4.0))); auto constant5 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(5.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(5.0))); auto inner_tuple1 = builder.AddInstruction( HloInstruction::CreateTuple({constant2, constant3})); auto tuple1 = builder.AddInstruction( @@ -960,7 +959,7 @@ TEST_P(HloDataflowAnalysisTest, NestedTupleSelect) { auto tuple2 = builder.AddInstruction( HloInstruction::CreateTuple({constant4, inner_tuple2})); auto select = builder.AddInstruction(HloInstruction::CreateTernary( - tuple1->shape(), HloOpcode::kSelect, pred, tuple1, tuple2)); + tuple1->shape(), HloOpcode::kTupleSelect, pred, tuple1, tuple2)); module_->AddEntryComputation(builder.Build()); @@ -983,7 +982,7 @@ TEST_P(HloDataflowAnalysisTest, NestedTupleSelect) { } TEST_P(HloDataflowAnalysisTest, TupleSelectToWhile) { - // Test a tuple-shaped kSelect feeding a kWhile instruction. HLO: + // Test a tuple-shaped kTupleSelect feeding a kWhile instruction. HLO: // // body((F32[], F32[]) %tuple_param): // %add = Add(%tuple_param{0}, %tuple_param{1}) @@ -1026,24 +1025,24 @@ TEST_P(HloDataflowAnalysisTest, TupleSelectToWhile) { cond_builder.AddInstruction( HloInstruction::CreateParameter(0, tuple_shape, "param")); cond_builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(false))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(false))); HloComputation* condition = module_->AddEmbeddedComputation(cond_builder.Build()); auto pred = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(false))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(false))); auto constant1 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0))); auto constant2 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(2.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(2.0))); auto constant3 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(3.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(3.0))); auto tuple1 = builder.AddInstruction(HloInstruction::CreateTuple({constant1})); auto tuple2 = builder.AddInstruction(HloInstruction::CreateTuple({constant2})); auto select = builder.AddInstruction(HloInstruction::CreateTernary( - tuple1->shape(), HloOpcode::kSelect, pred, tuple1, tuple2)); + tuple1->shape(), HloOpcode::kTupleSelect, pred, tuple1, tuple2)); auto gte = builder.AddInstruction( HloInstruction::CreateGetTupleElement(scalar_shape_, select, 0)); auto tuple = @@ -1089,7 +1088,7 @@ TEST_P(HloDataflowAnalysisTest, BitcastDefinesValue) { // Test the bitcast_defines_value flag to the dataflow analysis. auto builder = HloComputation::Builder(TestName()); auto constant = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0))); auto bitcast = builder.AddInstruction(HloInstruction::CreateUnary( scalar_shape_, HloOpcode::kBitcast, constant)); @@ -1158,44 +1157,50 @@ TEST_P(HloDataflowAnalysisTest, SendAndSendDone) { auto builder = HloComputation::Builder(TestName()); auto param = builder.AddInstruction( HloInstruction::CreateParameter(0, scalar_shape_, "param0")); + auto token = builder.AddInstruction(HloInstruction::CreateToken()); auto send = builder.AddInstruction( - HloInstruction::CreateSend(param, /*channel_id=*/0)); + HloInstruction::CreateSend(param, token, /*channel_id=*/0)); auto send_done = builder.AddInstruction(HloInstruction::CreateSendDone(send)); module_->AddEntryComputation(builder.Build()); bool ssa_form = GetParam(); const HloDataflowAnalysis& analysis = RunAnalysis(ssa_form); - EXPECT_EQ(analysis.values().size(), 4); + EXPECT_EQ(analysis.values().size(), 6); EXPECT_TRUE(analysis.ValueIsDefinedAt(param)); EXPECT_TRUE(analysis.ValueIsDefinedAt(send, /*index=*/{})); EXPECT_FALSE(analysis.ValueIsDefinedAt(send, /*index=*/{0})); EXPECT_TRUE(analysis.ValueIsDefinedAt(send, /*index=*/{1})); + EXPECT_TRUE(analysis.ValueIsDefinedAt(send, /*index=*/{2})); EXPECT_TRUE(analysis.ValueIsDefinedAt(send_done)); EXPECT_THAT(HloValuesAt(send, /*index=*/{0}), UnorderedElementsAre(analysis.GetValueDefinedAt(param))); } TEST_P(HloDataflowAnalysisTest, RecvAndRecvDone) { - // Test that a RecvDone forwards its operand tuple element at {0} to the - // output. + // Test that a RecvDone forwards its operand tuple element at {0} to element + // {0} of the output. auto builder = HloComputation::Builder(TestName()); + auto token = builder.AddInstruction(HloInstruction::CreateToken()); auto recv = builder.AddInstruction( - HloInstruction::CreateRecv(scalar_shape_, /*channel_id=*/0)); + HloInstruction::CreateRecv(scalar_shape_, token, /*channel_id=*/0)); auto recv_done = builder.AddInstruction(HloInstruction::CreateRecvDone(recv)); module_->AddEntryComputation(builder.Build()); bool ssa_form = GetParam(); const HloDataflowAnalysis& analysis = RunAnalysis(ssa_form); - EXPECT_EQ(analysis.values().size(), 3); + EXPECT_EQ(analysis.values().size(), 7); EXPECT_TRUE(analysis.ValueIsDefinedAt(recv, /*index=*/{})); EXPECT_TRUE(analysis.ValueIsDefinedAt(recv, /*index=*/{0})); EXPECT_TRUE(analysis.ValueIsDefinedAt(recv, /*index=*/{1})); - EXPECT_FALSE(analysis.ValueIsDefinedAt(recv_done)); - EXPECT_THAT(HloValuesAt(recv_done), + EXPECT_TRUE(analysis.ValueIsDefinedAt(recv, /*index=*/{2})); + EXPECT_TRUE(analysis.ValueIsDefinedAt(recv_done, /*index=*/{})); + EXPECT_FALSE(analysis.ValueIsDefinedAt(recv_done, /*index=*/{0})); + EXPECT_TRUE(analysis.ValueIsDefinedAt(recv_done, /*index=*/{1})); + EXPECT_THAT(HloValuesAt(recv_done, /*index=*/{0}), UnorderedElementsAre(analysis.GetValueDefinedAt(recv, {0}))); EXPECT_TRUE( analysis.GetValueDefinedAt(recv, /*index=*/{0}).live_out_of_module()); @@ -1304,13 +1309,13 @@ TEST_P(HloDataflowAnalysisTest, WhileParameters_Sequential) { auto body_param = body_builder.AddInstruction( HloInstruction::CreateParameter(0, scalar_shape_, "body_param")); auto constant = body_builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0))); auto exp = body_builder.AddInstruction( HloInstruction::CreateUnary(scalar_shape_, HloOpcode::kExp, constant)); auto add = body_builder.AddInstruction(HloInstruction::CreateBinary( scalar_shape_, HloOpcode::kAdd, exp, body_param)); auto dead_constant = body_builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0))); auto dead_negate = body_builder.AddInstruction(HloInstruction::CreateUnary( scalar_shape_, HloOpcode::kNegate, dead_constant)); HloComputation* body = module_->AddEmbeddedComputation( @@ -1320,7 +1325,7 @@ TEST_P(HloDataflowAnalysisTest, WhileParameters_Sequential) { auto cond_param = cond_builder.AddInstruction( HloInstruction::CreateParameter(0, scalar_shape_, "cond_param")); auto cond_constant = cond_builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(false))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(false))); HloComputation* condition = module_->AddEmbeddedComputation(cond_builder.Build()); @@ -1571,11 +1576,11 @@ TEST_P(HloDataflowAnalysisTest, ConditionalWithIdentity) { auto builder = HloComputation::Builder(TestName()); auto pred = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(true))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(true))); auto constant1 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(56.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(56.0f))); auto constant2 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(12.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(12.0f))); auto conditional = builder.AddInstruction(HloInstruction::CreateConditional( scalar_shape_, pred, constant1, true_computation, constant2, false_computation)); @@ -1662,11 +1667,11 @@ TEST_P(HloDataflowAnalysisTest, ConditionalTakingTupleOperand) { auto builder = HloComputation::Builder(TestName()); auto pred = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(true))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(true))); auto constant1 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(56.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(56.0f))); auto constant2 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(12.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(12.0f))); auto tuple_operand = builder.AddInstruction( HloInstruction::CreateTuple({constant1, constant2})); auto conditional = builder.AddInstruction(HloInstruction::CreateConditional( @@ -1792,15 +1797,15 @@ TEST_P(HloDataflowAnalysisTest, NestedConditionals) { // Build entry computation. auto builder = HloComputation::Builder(TestName()); auto pred1 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(true))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(true))); auto pred2 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(false))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(false))); auto constant1 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.1f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.1f))); auto constant2 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(2.2f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(2.2f))); auto constant3 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(3.3f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(3.3f))); auto tuple_operand = builder.AddInstruction( HloInstruction::CreateTuple({pred2, constant1, constant2})); auto conditional = builder.AddInstruction(HloInstruction::CreateConditional( @@ -1938,9 +1943,9 @@ TEST_F(DoesNotUseOperandBufferTest, FusedDynamicUpdateSlice) { // Create a DynamicUpdateSlice instruction of tuple element 1. auto starts = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR1({2}))); + HloInstruction::CreateConstant(LiteralUtil::CreateR1({2}))); auto update = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR1({2.f, 2.f, 2.f}))); + LiteralUtil::CreateR1({2.f, 2.f, 2.f}))); auto dynamic_update_slice = builder.AddInstruction(HloInstruction::CreateDynamicUpdateSlice( data_shape, gte1, update, starts)); @@ -2043,7 +2048,7 @@ TEST_F(CanShareOperandBufferWithUserTest, Shape data_shape = ShapeUtil::MakeShape(F32, {2, 2}); auto one = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0))); auto operand = builder.AddInstruction( HloInstruction::CreateBroadcast(data_shape, one, {1})); @@ -2071,7 +2076,7 @@ TEST_F(CanShareOperandBufferWithUserTest, auto param = builder.AddInstruction( HloInstruction::CreateParameter(0, data_shape, "param0")); auto index = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR1({0, 0}))); + HloInstruction::CreateConstant(LiteralUtil::CreateR1({0, 0}))); auto ds = builder.AddInstruction( HloInstruction::CreateDynamicSlice(slice_shape, param, index, {1, 2, 2})); @@ -2139,9 +2144,9 @@ TEST_F(CanShareOperandBufferWithUserTest, FusedDynamicUpdateSlice) { // Create a DynamicUpdateSlice instruction of tuple element 1. auto starts = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR1({2}))); + HloInstruction::CreateConstant(LiteralUtil::CreateR1({2}))); auto update = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR1({2.f, 2.f, 2.f}))); + LiteralUtil::CreateR1({2.f, 2.f, 2.f}))); auto dynamic_update_slice = builder.AddInstruction(HloInstruction::CreateDynamicUpdateSlice( data_shape, gte1, update, starts)); @@ -2179,9 +2184,9 @@ TEST_F(CanShareOperandBufferWithUserTest, // Create a DynamicUpdateSlice instruction of tuple element 1. auto starts = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR1({2}))); + HloInstruction::CreateConstant(LiteralUtil::CreateR1({2}))); auto update = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR1({2.f, 2.f, 2.f}))); + LiteralUtil::CreateR1({2.f, 2.f, 2.f}))); auto dynamic_update_slice = builder.AddInstruction(HloInstruction::CreateDynamicUpdateSlice( data_shape_bf16, convert1, update, starts)); @@ -2232,9 +2237,9 @@ TEST_F(CanShareOperandBufferWithUserTest, FusedDotAdd) { Shape data_shape = ShapeUtil::MakeShape(F32, {2, 2}); auto a = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR2({{1.0, 0.0}, {0.0, 1.0}}))); + LiteralUtil::CreateR2({{1.0, 0.0}, {0.0, 1.0}}))); auto b = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR2({{2.0, 2.0}, {2.0, 2.0}}))); + LiteralUtil::CreateR2({{2.0, 2.0}, {2.0, 2.0}}))); DotDimensionNumbers dot_dnums; dot_dnums.add_lhs_contracting_dimensions(1); @@ -2243,7 +2248,7 @@ TEST_F(CanShareOperandBufferWithUserTest, FusedDotAdd) { HloInstruction::CreateDot(data_shape, a, b, dot_dnums)); auto one = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0))); auto add_operand = builder.AddInstruction( HloInstruction::CreateBroadcast(data_shape, one, {1})); @@ -2265,7 +2270,7 @@ TEST_F(CanShareOperandBufferWithUserTest, OutputFusionCantAliasOperandBuffer) { Shape data_shape = ShapeUtil::MakeShape(F32, {2, 2}); auto one = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0))); auto operand = builder.AddInstruction( HloInstruction::CreateBroadcast(data_shape, one, {1})); @@ -2273,7 +2278,7 @@ TEST_F(CanShareOperandBufferWithUserTest, OutputFusionCantAliasOperandBuffer) { HloInstruction::CreateReverse(data_shape, operand, {0, 1})); auto two = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR2({{2.0, 2.0}, {2.0, 2.0}}))); + LiteralUtil::CreateR2({{2.0, 2.0}, {2.0, 2.0}}))); auto add = builder.AddInstruction( HloInstruction::CreateBinary(data_shape, HloOpcode::kAdd, reverse, two)); @@ -2293,13 +2298,13 @@ TEST_F(CanShareOperandBufferWithUserTest, FusionCanShareBufferCustomized) { Shape data_shape = ShapeUtil::MakeShape(F32, {2, 2}); auto one = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0))); auto operand = builder.AddInstruction( HloInstruction::CreateBroadcast(data_shape, one, {1})); auto mul = builder.AddInstruction(HloInstruction::CreateBinary( data_shape, HloOpcode::kMultiply, operand, operand)); auto two = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR2({{2.0, 2.0}, {2.0, 2.0}}))); + LiteralUtil::CreateR2({{2.0, 2.0}, {2.0, 2.0}}))); auto add = builder.AddInstruction( HloInstruction::CreateBinary(data_shape, HloOpcode::kAdd, mul, two)); @@ -2365,7 +2370,7 @@ TEST_F(CanShareOperandBufferWithUserTest, CallToComputationWithFusionRoot) { auto sub_param = sub_builder.AddInstruction( HloInstruction::CreateParameter(0, shape, "sub_param")); auto one = sub_builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0))); auto ones = sub_builder.AddInstruction( HloInstruction::CreateBroadcast(shape, one, {1})); auto add = sub_builder.AddInstruction( diff --git a/tensorflow/compiler/xla/service/hlo_dce_test.cc b/tensorflow/compiler/xla/service/hlo_dce_test.cc index 5a56607a665c4cbeb7b2572f182b88e890602968..26e3736e01270dbc6ca67647e814843aba2d1e3d 100644 --- a/tensorflow/compiler/xla/service/hlo_dce_test.cc +++ b/tensorflow/compiler/xla/service/hlo_dce_test.cc @@ -53,9 +53,9 @@ TEST_F(HloDceTest, NoDeadCode) { // Verify that no dead code is removed from a computation with no dead code. auto builder = HloComputation::Builder(TestName()); auto constant1 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(42.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(42.0f))); auto constant2 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(123.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(123.0f))); builder.AddInstruction(HloInstruction::CreateBinary( constant1->shape(), HloOpcode::kAdd, constant1, constant2)); @@ -74,20 +74,21 @@ TEST_F(HloDceTest, InstructionsWithSideEffect) { // Verify that side-effect instructions (Send in this test) are not removed. auto builder = HloComputation::Builder(TestName()); auto constant = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(42.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(42.0f))); + auto token = builder.AddInstruction(HloInstruction::CreateToken()); builder.AddInstruction( - HloInstruction::CreateSend(constant, /*channel_id=*/0)); + HloInstruction::CreateSend(constant, token, /*channel_id=*/0)); builder.AddInstruction(HloInstruction::CreateTuple({})); auto module = CreateNewModule(); auto computation = module->AddEntryComputation(builder.Build()); - EXPECT_EQ(3, computation->instruction_count()); + EXPECT_EQ(4, computation->instruction_count()); HloDCE dce; EXPECT_FALSE(dce.Run(module.get()).ValueOrDie()); - EXPECT_EQ(3, computation->instruction_count()); + EXPECT_EQ(4, computation->instruction_count()); } TEST_F(HloDceTest, DeadParameters) { @@ -126,9 +127,9 @@ TEST_F(HloDceTest, ControlDependencies) { // Verify that instructions with control dependencies are not removed. auto builder = HloComputation::Builder(TestName()); auto constant1 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(42.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(42.0f))); auto constant2 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(123.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(123.0f))); // Create two dead instructions: a negate and an add. auto dead_negate = builder.AddInstruction(HloInstruction::CreateUnary( @@ -223,7 +224,7 @@ TEST_F(HloDceTest, CalledComputationWithSideEffect) { auto param = cond_builder.AddInstruction( HloInstruction::CreateParameter(0, shape, "cond_param")); auto constant = cond_builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(42.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(42.0f))); cond_builder.AddInstruction(HloInstruction::CreateBinary( ShapeUtil::MakeShape(PRED, {}), HloOpcode::kLt, param, constant)); } @@ -234,9 +235,9 @@ TEST_F(HloDceTest, CalledComputationWithSideEffect) { { auto param = body_builder.AddInstruction( HloInstruction::CreateParameter(0, shape, "param")); - - auto infeed = - body_builder.AddInstruction(HloInstruction::CreateInfeed(shape, "")); + auto token = body_builder.AddInstruction(HloInstruction::CreateToken()); + auto infeed = body_builder.AddInstruction( + HloInstruction::CreateInfeed(shape, token, "")); body_builder.AddInstruction( HloInstruction::CreateBinary(shape, HloOpcode::kAdd, param, infeed)); } @@ -278,8 +279,10 @@ TEST_F(HloDceTest, CalledComputationWithNestedSideEffect) { { auto param = nested_callee_builder.AddInstruction( HloInstruction::CreateParameter(0, shape, "param")); + auto token = + nested_callee_builder.AddInstruction(HloInstruction::CreateToken()); nested_callee_builder.AddInstruction( - HloInstruction::CreateOutfeed(shape, param, "")); + HloInstruction::CreateOutfeed(shape, param, token, "")); } auto nested_called_computation = module->AddEmbeddedComputation(nested_callee_builder.Build()); @@ -342,12 +345,12 @@ TEST_F(HloDceTest, RemoveDeadSubcomputation) { builder.AddInstruction(HloInstruction::CreateParameter( /*parameter_number=*/0, ShapeUtil::MakeShape(F32, {100}), "param0")), builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(0))), + HloInstruction::CreateConstant(LiteralUtil::CreateR0(0))), /*dimensions_to_reduce=*/{0}, reduce_subcomp)); // Add another instruction as the root of the computation. builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(0))); module->AddEntryComputation(builder.Build()); EXPECT_EQ(module->MakeComputationPostOrder().size(), 2); @@ -383,7 +386,7 @@ TEST_F(HloDceTest, KeepUsedSubcomputation) { builder.AddInstruction(HloInstruction::CreateParameter( /*parameter_number=*/0, ShapeUtil::MakeShape(F32, {100}), "param0")), builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(0))), + HloInstruction::CreateConstant(LiteralUtil::CreateR0(0))), /*dimensions_to_reduce=*/{0}, reduce_subcomp)); // Add another instruction as the root of the computation that also uses @@ -393,7 +396,7 @@ TEST_F(HloDceTest, KeepUsedSubcomputation) { builder.AddInstruction(HloInstruction::CreateParameter( /*parameter_number=*/1, ShapeUtil::MakeShape(F32, {100}), "param1")), builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(0))), + HloInstruction::CreateConstant(LiteralUtil::CreateR0(0))), /*dimensions_to_reduce=*/{0}, reduce_subcomp)); module->AddEntryComputation(builder.Build()); diff --git a/tensorflow/compiler/xla/service/hlo_domain_map.cc b/tensorflow/compiler/xla/service/hlo_domain_map.cc index ebd5adb5d573ce4b556046f85eb26a6ad59efcb9..9e096320db5048457435199627a1ef1fe1572177 100644 --- a/tensorflow/compiler/xla/service/hlo_domain_map.cc +++ b/tensorflow/compiler/xla/service/hlo_domain_map.cc @@ -41,11 +41,15 @@ namespace xla { bool HloDomainMap::InSameDomain(HloInstruction* instruction1, HloInstruction* instruction2) const { - int64 domain_id1 = FindOrDefault(instruction_to_domain_, instruction1, -1); - int64 domain_id2 = FindOrDefault(instruction_to_domain_, instruction2, -1); + int64 domain_id1 = GetDomainId(instruction1); + int64 domain_id2 = GetDomainId(instruction2); return domain_id1 >= 0 && domain_id1 == domain_id2; } +int64 HloDomainMap::GetDomainId(HloInstruction* instruction) const { + return FindOrDefault(instruction_to_domain_, instruction, -1); +} + Status HloDomainMap::TryProcessEmptyDomain(HloInstruction* instruction) { TF_RET_CHECK(instruction->opcode() == HloOpcode::kDomain); // We only check operands, so we are sure to not process the empty domain from @@ -58,6 +62,11 @@ Status HloDomainMap::TryProcessEmptyDomain(HloInstruction* instruction) { TF_RETURN_IF_ERROR(InsertDomain(std::move(domain))); } } + if (instruction == instruction->parent()->root_instruction()) { + auto domain = MakeUnique(); + domain->enter_domains.insert(instruction); + TF_RETURN_IF_ERROR(InsertDomain(std::move(domain))); + } return Status::OK(); } diff --git a/tensorflow/compiler/xla/service/hlo_domain_map.h b/tensorflow/compiler/xla/service/hlo_domain_map.h index e62ef763fb3881ab6030b1f6a66266ac80a3d84d..1ca71597253eecfb45ae8f384240033a57045277 100644 --- a/tensorflow/compiler/xla/service/hlo_domain_map.h +++ b/tensorflow/compiler/xla/service/hlo_domain_map.h @@ -65,6 +65,10 @@ class HloDomainMap { // currently processing. bool IsDomainInstruction(HloInstruction* instruction) const; + // Retrieves the domain identifier of the instruction, or -1 in case + // instruction is not found within any domain. + int64 GetDomainId(HloInstruction* instruction) const; + private: HloDomainMap(string domain_kind) : domain_kind_(std::move(domain_kind)) {} diff --git a/tensorflow/compiler/xla/service/hlo_domain_remover.cc b/tensorflow/compiler/xla/service/hlo_domain_remover.cc index 1d06040b0e7c92b03f4cb5481bdee73a0f74f939..e2e820002be77201b5bb237d90bbbc31ac178bb9 100644 --- a/tensorflow/compiler/xla/service/hlo_domain_remover.cc +++ b/tensorflow/compiler/xla/service/hlo_domain_remover.cc @@ -16,8 +16,8 @@ limitations under the License. #include "tensorflow/compiler/xla/service/hlo_domain_remover.h" #include "tensorflow/compiler/xla/service/hlo_computation.h" -#include "tensorflow/compiler/xla/service/hlo_domain_isolator.h" #include "tensorflow/compiler/xla/service/hlo_domain_map.h" +#include "tensorflow/compiler/xla/service/hlo_domain_verifier.h" #include "tensorflow/compiler/xla/service/hlo_graph_dumper.h" #include "tensorflow/compiler/xla/service/hlo_instruction.h" #include "tensorflow/compiler/xla/service/hlo_opcode.h" @@ -43,46 +43,8 @@ class HloDomainRemover::RunContext { Status HloDomainRemover::RunContext::VerifyAndNormalizeDomain( const DomainMetadata::Domain& domain) { - // Verify that the whole kDomain frontier bounding the instruction reach set, - // has matching metadata. - // A kDomain instruction has two sides of metadata, a user facing and an - // operand facing. - // A reachable instruction set can make contact with a kDomain instruction on - // a user facing side (the kDomain is operand of the instruction), or on a - // operand facing side (the kDomain is user of the instruction). - // And depending on the contact side, the proper metadata object - // (user_side_metadata() vs. operand_side_metadata()) needs to be used for - // consistency checks. - const DomainMetadata* ref_metadata = nullptr; - VLOG(4) << "Reach set:"; - for (HloInstruction* instruction : domain.instructions) { - VLOG(4) << " " << instruction->name(); - } - VLOG(4) << " Domains:"; - for (HloInstruction* instruction : domain.enter_domains) { - const DomainMetadata& meta = instruction->user_side_metadata(); - VLOG(4) << " User side: " << instruction->name(); - VLOG(4) << " " << meta.ToString(); - if (ref_metadata == nullptr) { - ref_metadata = &meta; - } else { - TF_RET_CHECK(meta.Matches(*ref_metadata)) - << "Metadata mismatch at instruction " << instruction->name() << " : " - << meta.ToString() << " vs " << ref_metadata->ToString(); - } - } - for (HloInstruction* instruction : domain.exit_domains) { - const DomainMetadata& meta = instruction->operand_side_metadata(); - VLOG(4) << " Operand side: " << instruction->name(); - VLOG(4) << " " << meta.ToString(); - if (ref_metadata == nullptr) { - ref_metadata = &meta; - } else { - TF_RET_CHECK(meta.Matches(*ref_metadata)) - << "Metadata mismatch at instruction " << instruction->name() << " : " - << meta.ToString() << " vs " << ref_metadata->ToString(); - } - } + TF_ASSIGN_OR_RETURN(const DomainMetadata* ref_metadata, + HloDomainVerifier::VerifyDomain(domain)); if (ref_metadata != nullptr) { VLOG(4) << "Applying domain normalization: " << ref_metadata->ToString(); TF_RETURN_IF_ERROR(ref_metadata->NormalizeInstructions(domain)); diff --git a/tensorflow/compiler/xla/service/hlo_domain_test.cc b/tensorflow/compiler/xla/service/hlo_domain_test.cc index 5d8081c1ef197548e1d802374f3efe35fa113cd3..00b2c860a7f38fa72af648294e2c5fc409878797 100644 --- a/tensorflow/compiler/xla/service/hlo_domain_test.cc +++ b/tensorflow/compiler/xla/service/hlo_domain_test.cc @@ -201,12 +201,14 @@ HloModule Module ENTRY entry { p0 = (f32[4]) parameter(0) a = f32[4] get-tuple-element(p0), index=0 - b = (f32[4], u32[]) send(a), channel_id=1, sharding={maximal device=0} - c = () send-done(b), channel_id=1, sharding={maximal device=0} - d = (f32[4], u32[]) recv(), channel_id=2, sharding={maximal device=0} - e = f32[4] recv-done(d), channel_id=2, sharding={maximal device=0} - f = f32[4] add(a, e) - g = f32[4] subtract(a, e) + token = token[] after-all() + b = (f32[4], u32[], token[]) send(a, token), channel_id=1, sharding={maximal device=0} + c = token[] send-done(b), channel_id=1, sharding={maximal device=0} + d = (f32[4], u32[], token[]) recv(token), channel_id=2, sharding={maximal device=0} + e = (f32[4], token[]) recv-done(d), channel_id=2, sharding={maximal device=0} + e_element = f32[4] get-tuple-element(e), index=0, sharding={maximal device=0} + f = f32[4] add(a, e_element) + g = f32[4] subtract(a, e_element) ROOT h = (f32[4], f32[4]) tuple(f, g) } )"; @@ -219,7 +221,7 @@ ENTRY entry { EXPECT_TRUE(isolator_changed); EXPECT_TRUE(HasDomainEdge(module, "b", "a")); - EXPECT_TRUE(HasDomainEdge(module, "f", "e")); + EXPECT_TRUE(HasDomainEdge(module, "f", "e_element")); EXPECT_FALSE(HasDomainEdge(module, "a", "p0")); EXPECT_FALSE(HasDomainEdge(module, "c", "b")); EXPECT_FALSE(HasDomainEdge(module, "e", "d")); @@ -230,7 +232,7 @@ ENTRY entry { EXPECT_TRUE(remover_changed); EXPECT_FALSE(HasDomainEdge(module, "b", "a")); - EXPECT_FALSE(HasDomainEdge(module, "f", "e")); + EXPECT_FALSE(HasDomainEdge(module, "f", "e_element")); } TEST_F(HloDomainTest, CheckNoDomainAddedOnPureIOComputation) { @@ -238,11 +240,13 @@ TEST_F(HloDomainTest, CheckNoDomainAddedOnPureIOComputation) { HloModule Module ENTRY entry { - a = (f32[4], u32[]) recv(), channel_id=1, sharding={maximal device=-1} - b = f32[4] recv-done(a), channel_id=1, sharding={maximal device=-1} - c = f32[4] add(b, b), sharding={maximal device=-1} - d = (f32[4], u32[]) send(c), channel_id=2, sharding={maximal device=-1} - ROOT e = () send-done(d), channel_id=2, sharding={maximal device=-1} + token = token[] after-all(), sharding={maximal device=-1} + a = (f32[4], u32[], token[]) recv(token), channel_id=1, sharding={maximal device=-1} + b = (f32[4], token[]) recv-done(a), channel_id=1, sharding={maximal device=-1} + b_element = f32[4] get-tuple-element(b), index=0, sharding={maximal device=-1} + c = f32[4] add(b_element, b_element), sharding={maximal device=-1} + d = (f32[4], u32[], token[]) send(c, token), channel_id=2, sharding={maximal device=-1} + ROOT e = token[] send-done(d), channel_id=2, sharding={maximal device=-1} } )"; @@ -259,11 +263,13 @@ TEST_F(HloDomainTest, CheckNormalizationOnPureIOComputation) { HloModule Module ENTRY entry { - a = (f32[4], u32[]) recv(), channel_id=1, sharding={maximal device=0} - b = f32[4] recv-done(a), channel_id=1, sharding={maximal device=0} - c = f32[4] add(b, b) - d = (f32[4], u32[]) send(c), channel_id=2, sharding={maximal device=0} - ROOT e = () send-done(d), channel_id=2, sharding={maximal device=0} + token = token[] after-all(), sharding={maximal device=0} + a = (f32[4], u32[], token[]) recv(token), channel_id=1, sharding={maximal device=0} + b = (f32[4], token[]) recv-done(a), channel_id=1, sharding={maximal device=0} + b_element = f32[4] get-tuple-element(b), index=0, sharding={maximal device=0} + c = f32[4] add(b_element, b_element) + d = (f32[4], u32[], token[]) send(c, token), channel_id=2, sharding={maximal device=0} + ROOT e = token[] send-done(d), channel_id=2, sharding={maximal device=0} } )"; @@ -340,10 +346,12 @@ TEST_F(HloDomainTest, CheckNormalizationOnInfeedTuple) { HloModule Module ENTRY entry { - infeed = (f32[4], f32[4]) infeed(), - sharding={{maximal device=1}, {maximal device=0}} - gte0 = f32[4] get-tuple-element(infeed), index=0 - gte1 = f32[4] get-tuple-element(infeed), index=1 + token = token[] after-all() + infeed = ((f32[4], f32[4]), token[]) infeed(token), + sharding={{maximal device=1}, {maximal device=0}, {maximal device=0}} + infeed.data = (f32[4], f32[4]) get-tuple-element(infeed), index=0 + gte0 = f32[4] get-tuple-element(infeed.data), index=0 + gte1 = f32[4] get-tuple-element(infeed.data), index=1 copy0 = f32[4] copy(gte0) copy1 = f32[4] copy(gte1) ROOT add = f32[4] add(copy0, copy1) @@ -357,8 +365,7 @@ ENTRY entry { TF_ASSERT_OK_AND_ASSIGN(bool isolator_changed, isolator.Run(module)); EXPECT_TRUE(isolator_changed); - EXPECT_TRUE(HasDomainEdge(module, "gte0", "infeed")); - EXPECT_TRUE(HasDomainEdge(module, "gte1", "infeed")); + EXPECT_TRUE(HasDomainEdge(module, "infeed.data", "infeed")); EXPECT_FALSE(HasDomainEdge(module, "copy0", "gte0")); EXPECT_FALSE(HasDomainEdge(module, "copy1", "gte1")); @@ -366,6 +373,8 @@ ENTRY entry { // HLO passes adding unexpected instructions. // // infeed + // | + // infeed.data (tuple element 0 of infeed) // / \ // GTE0 GTE1 // / \ @@ -374,26 +383,31 @@ ENTRY entry { // \ / // TUPLE // | - // DOMAIN HloInstruction* infeed = FindInstruction(module, "infeed"); ASSERT_NE(infeed, nullptr); - auto infeed_users = infeed->users(); - HloInstruction* new_gte0 = + HloInstruction* infeed_data = infeed->parent()->AddInstruction(HloInstruction::CreateGetTupleElement( ShapeUtil::GetTupleElementShape(infeed->shape(), 0), infeed, 0)); + + auto infeed_data_users = infeed_data->users(); + HloInstruction* new_gte0 = infeed_data->parent()->AddInstruction( + HloInstruction::CreateGetTupleElement( + ShapeUtil::GetTupleElementShape(infeed_data->shape(), 0), infeed_data, + 0)); HloInstruction* new_copy0 = - infeed->parent()->AddInstruction(HloInstruction::CreateUnary( + infeed_data->parent()->AddInstruction(HloInstruction::CreateUnary( new_gte0->shape(), HloOpcode::kCopy, new_gte0)); - HloInstruction* new_gte1 = - infeed->parent()->AddInstruction(HloInstruction::CreateGetTupleElement( - ShapeUtil::GetTupleElementShape(infeed->shape(), 1), infeed, 1)); + HloInstruction* new_gte1 = infeed_data->parent()->AddInstruction( + HloInstruction::CreateGetTupleElement( + ShapeUtil::GetTupleElementShape(infeed_data->shape(), 1), infeed_data, + 1)); HloInstruction* new_copy1 = - infeed->parent()->AddInstruction(HloInstruction::CreateUnary( + infeed_data->parent()->AddInstruction(HloInstruction::CreateUnary( new_gte1->shape(), HloOpcode::kCopy, new_gte1)); - HloInstruction* new_tuple = infeed->parent()->AddInstruction( + HloInstruction* new_tuple = infeed_data->parent()->AddInstruction( HloInstruction::CreateTuple({new_copy0, new_copy1})); - for (HloInstruction* user : infeed_users) { - TF_EXPECT_OK(infeed->ReplaceUseWith(user, new_tuple)); + for (HloInstruction* user : infeed_data_users) { + TF_EXPECT_OK(infeed_data->ReplaceUseWith(user, new_tuple)); } HloDomainRemover remover(ShardingMetadata::KindName(), @@ -412,7 +426,7 @@ ENTRY entry { }; for (auto& assignment : assignments) { auto device = assignment.instruction->sharding_unique_device(); - EXPECT_TRUE(device.has_value()); + ASSERT_TRUE(device.has_value()); EXPECT_EQ(*device, assignment.device); } EXPECT_TRUE(new_tuple->has_sharding()); @@ -422,5 +436,64 @@ ENTRY entry { HloSharding::AssignDevice(0)})); } +TEST_F(HloDomainTest, EmptyRootDomain) { + const char* const hlo_string = R"( +HloModule Module + +ENTRY entry { + %param = f32[1] parameter(0), sharding={maximal device=0} + %tuple = (f32[1]) tuple(%param), + sharding={maximal device=1} + ROOT %gte = f32[1] get-tuple-element(%tuple), index=0, + sharding={maximal device=1} +})"; + + TF_ASSERT_OK_AND_ASSIGN(HloModule * module, ParseModule(hlo_string)); + + HloDomainIsolator isolator(CreateShardingDomain); + TF_ASSERT_OK_AND_ASSIGN(bool isolator_changed, isolator.Run(module)); + EXPECT_TRUE(isolator_changed); + + EXPECT_TRUE(HasDomainEdge(module, "tuple", "param")); + EXPECT_FALSE(HasDomainEdge(module, "gte", "tuple")); + + // Remove %tuple and %gte (tuple simplification) + HloInstruction* gte = FindInstruction(module, "gte"); + HloInstruction* tuple = FindInstruction(module, "tuple"); + module->entry_computation()->set_root_instruction(tuple->mutable_operand(0)); + TF_EXPECT_OK(module->entry_computation()->RemoveInstruction(gte)); + TF_EXPECT_OK(module->entry_computation()->RemoveInstruction(tuple)); + + HloDomainRemover remover(ShardingMetadata::KindName(), + NormalizeShardingDomain); + TF_ASSERT_OK_AND_ASSIGN(bool remover_changed, remover.Run(module)); + EXPECT_TRUE(remover_changed); + + const HloInstruction* root = module->entry_computation()->root_instruction(); + EXPECT_TRUE(root->has_sharding()); + EXPECT_EQ(root->sharding(), HloSharding::AssignDevice(1)); +} + +// Tests that text dumps of domain instructions can be parsed back, in the +// specific case of null shardings. +TEST_F(HloDomainTest, DumpParseNullSharding) { + auto builder = HloComputation::Builder(TestName()); + Shape shape = ShapeUtil::MakeShape(F32, {}); + auto sharding_md_0 = MakeUnique(nullptr); + auto sharding_md_1 = MakeUnique(nullptr); + HloInstruction* param = + builder.AddInstruction(HloInstruction::CreateParameter(0, shape, "p")); + HloInstruction* domain = builder.AddInstruction(HloInstruction::CreateDomain( + shape, param, std::move(sharding_md_0), std::move(sharding_md_1))); + builder.AddInstruction( + HloInstruction::CreateBinary(shape, HloOpcode::kAdd, domain, domain)); + + auto module = CreateNewModule(); + module->AddEntryComputation(builder.Build()); + + auto hlo_string = module->ToString(); + ASSERT_TRUE(ParseModule(hlo_string).status().ok()); +} + } // namespace } // namespace xla diff --git a/tensorflow/compiler/xla/service/hlo_domain_verifier.cc b/tensorflow/compiler/xla/service/hlo_domain_verifier.cc new file mode 100644 index 0000000000000000000000000000000000000000..751fc677e2d955fd3d9f8970f7c0370a22c054bf --- /dev/null +++ b/tensorflow/compiler/xla/service/hlo_domain_verifier.cc @@ -0,0 +1,124 @@ +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow/compiler/xla/service/hlo_domain_verifier.h" + +#include + +#include "tensorflow/compiler/xla/service/hlo_computation.h" +#include "tensorflow/compiler/xla/service/hlo_domain_map.h" +#include "tensorflow/compiler/xla/service/hlo_graph_dumper.h" +#include "tensorflow/compiler/xla/service/hlo_instruction.h" +#include "tensorflow/compiler/xla/service/hlo_opcode.h" +#include "tensorflow/compiler/xla/types.h" + +namespace xla { + +class HloDomainVerifier::RunContext { + public: + RunContext(HloModule* module, HloDomainVerifier* verifier) + : module_(module), verifier_(verifier) {} + + Status Run(); + + private: + // If the verifier caller passed an empty vector for kinds, we collect all the + // avalable domain types. + Status PopulateDomainKinds(); + + HloModule* module_; + HloDomainVerifier* verifier_; +}; + +Status HloDomainVerifier::RunContext::PopulateDomainKinds() { + if (verifier_->kinds_.empty()) { + // The caller specified no domain kinds, collect all the ones available. + std::set kinds; + for (HloComputation* computation : module_->computations()) { + for (HloInstruction* instruction : computation->instructions()) { + if (instruction->opcode() == HloOpcode::kDomain) { + TF_RET_CHECK(instruction->user_side_metadata().Kind() == + instruction->operand_side_metadata().Kind()) + << instruction->ToString(); + kinds.insert(instruction->user_side_metadata().Kind().ToString()); + } + } + } + verifier_->kinds_.insert(verifier_->kinds_.end(), kinds.begin(), + kinds.end()); + } + return Status::OK(); +} + +Status HloDomainVerifier::RunContext::Run() { + VLOG(4) << "Running HLO Domain Verifier"; + TF_RETURN_IF_ERROR(PopulateDomainKinds()); + for (HloComputation* computation : module_->computations()) { + for (auto& kind : verifier_->kinds_) { + // First create the domain instruciton sets. A domain instruction set is + // the set of instructions whose edges never cross a kDomain instruction. + TF_ASSIGN_OR_RETURN(std::unique_ptr domain_map, + HloDomainMap::Create(computation, kind)); + // Verify every domain populated within the map. + for (auto& domain : domain_map->GetDomains()) { + TF_RETURN_IF_ERROR(VerifyDomain(*domain).status()); + } + } + } + return Status::OK(); +} + +StatusOr HloDomainVerifier::Run(HloModule* module) { + RunContext run_context(module, this); + TF_RETURN_IF_ERROR(run_context.Run()); + return false; +} + +StatusOr HloDomainVerifier::VerifyDomain( + const DomainMetadata::Domain& domain) { + const DomainMetadata* ref_metadata = nullptr; + VLOG(4) << "Reach set:"; + for (HloInstruction* instruction : domain.instructions) { + VLOG(4) << " " << instruction->name(); + } + VLOG(4) << " Domains:"; + for (HloInstruction* instruction : domain.enter_domains) { + const DomainMetadata& meta = instruction->user_side_metadata(); + VLOG(4) << " User side: " << instruction->name(); + VLOG(4) << " " << meta.ToString(); + if (ref_metadata == nullptr) { + ref_metadata = &meta; + } else { + TF_RET_CHECK(meta.Matches(*ref_metadata)) + << "Metadata mismatch at instruction " << instruction->name() << " : " + << meta.ToString() << " vs " << ref_metadata->ToString(); + } + } + for (HloInstruction* instruction : domain.exit_domains) { + const DomainMetadata& meta = instruction->operand_side_metadata(); + VLOG(4) << " Operand side: " << instruction->name(); + VLOG(4) << " " << meta.ToString(); + if (ref_metadata == nullptr) { + ref_metadata = &meta; + } else { + TF_RET_CHECK(meta.Matches(*ref_metadata)) + << "Metadata mismatch at instruction " << instruction->name() << " : " + << meta.ToString() << " vs " << ref_metadata->ToString(); + } + } + return ref_metadata; +} + +} // namespace xla diff --git a/tensorflow/compiler/xla/service/hlo_domain_verifier.h b/tensorflow/compiler/xla/service/hlo_domain_verifier.h new file mode 100644 index 0000000000000000000000000000000000000000..8e53cf97f8ba9a88140a909ad20c1a938aec8c1f --- /dev/null +++ b/tensorflow/compiler/xla/service/hlo_domain_verifier.h @@ -0,0 +1,65 @@ +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#ifndef TENSORFLOW_COMPILER_XLA_SERVICE_HLO_DOMAIN_VERIFIER_H_ +#define TENSORFLOW_COMPILER_XLA_SERVICE_HLO_DOMAIN_VERIFIER_H_ + +#include +#include + +#include "tensorflow/compiler/xla/service/hlo_domain_map.h" +#include "tensorflow/compiler/xla/service/hlo_domain_metadata.h" +#include "tensorflow/compiler/xla/service/hlo_module.h" +#include "tensorflow/compiler/xla/service/hlo_pass_interface.h" +#include "tensorflow/core/lib/core/status.h" + +namespace xla { + +// Verifies that the domain instructions are consistent, and the each domain is +// surrounded by the same metadata. +class HloDomainVerifier : public HloPassInterface { + public: + HloDomainVerifier(std::vector kinds) : kinds_(std::move(kinds)) {} + + tensorflow::StringPiece name() const override { return "domain_verifier"; } + + StatusOr Run(HloModule* module) override; + + // Verify that the whole kDomain frontier bounding the instruction reach set, + // has matching metadata. + // A kDomain instruction has two sides of metadata, a user facing and an + // operand facing. + // A reachable instruction set can make contact with a kDomain instruction on + // a user facing side (the kDomain is operand of the instruction), or on a + // operand facing side (the kDomain is user of the instruction). + // And depending on the contact side, the proper metadata object + // (user_side_metadata() vs. operand_side_metadata()) needs to be used for + // consistency checks. + // Returns the DomainMetadata pointer which surrounds the domain, and + // represents the common metadata within such domain. If the returned + // DomainMetadata pointer is nullptr, the input domain had no kDomain + // boundary. + static StatusOr VerifyDomain( + const DomainMetadata::Domain& domain); + + private: + class RunContext; + + std::vector kinds_; +}; + +} // namespace xla + +#endif // TENSORFLOW_COMPILER_XLA_SERVICE_HLO_DOMAIN_VERIFIER_H_ diff --git a/tensorflow/compiler/xla/service/hlo_element_type_converter.cc b/tensorflow/compiler/xla/service/hlo_element_type_converter.cc index 4ed1508d7067684a15d0fb7d86e69b055bc1333b..c804f4364f6d16d5b8112219ce884495200aa827 100644 --- a/tensorflow/compiler/xla/service/hlo_element_type_converter.cc +++ b/tensorflow/compiler/xla/service/hlo_element_type_converter.cc @@ -21,7 +21,7 @@ limitations under the License. #include #include "tensorflow/compiler/xla/layout_util.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/service/dfs_hlo_visitor_with_default.h" #include "tensorflow/compiler/xla/service/hlo_computation.h" #include "tensorflow/compiler/xla/service/hlo_evaluator.h" diff --git a/tensorflow/compiler/xla/service/hlo_element_type_converter_test.cc b/tensorflow/compiler/xla/service/hlo_element_type_converter_test.cc index 5c5a059e0fd895f03bc26a975609b57333237faf..c170e36c73ad2bef830e528de3ec72d38683d888 100644 --- a/tensorflow/compiler/xla/service/hlo_element_type_converter_test.cc +++ b/tensorflow/compiler/xla/service/hlo_element_type_converter_test.cc @@ -57,8 +57,10 @@ TEST_F(HloElementTypeConverterTest, InfeedsOutfeedsNotConverted) { const string& hlo_string = R"( HloModule InfeedOutfeed ENTRY RoundTrip16MiBR1.v2 { - ROOT infeed = bf16[4]{0} infeed() - outfeed = () outfeed(infeed) + token = token[] after-all() + infeed = (bf16[4]{0}, token[]) infeed(token) + ROOT infeed.data = bf16[4]{0} get-tuple-element(infeed), index=0 + outfeed = token[] outfeed(infeed.data, token) } )"; auto module = CreateModuleFromHloString(hlo_string); diff --git a/tensorflow/compiler/xla/service/hlo_evaluator.cc b/tensorflow/compiler/xla/service/hlo_evaluator.cc index 33424019b93feff862c6e3e268ae3980bacc9142..dfdfeb49a2e161076b22bd7019a08df545ce427e 100644 --- a/tensorflow/compiler/xla/service/hlo_evaluator.cc +++ b/tensorflow/compiler/xla/service/hlo_evaluator.cc @@ -25,6 +25,7 @@ limitations under the License. #include "tensorflow/compiler/xla/index_util.h" #include "tensorflow/compiler/xla/layout_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/literal_util.h" #include "tensorflow/compiler/xla/map_util.h" #include "tensorflow/compiler/xla/primitive_util.h" @@ -135,7 +136,6 @@ StatusOr> Compare( } // namespace - HloEvaluator::HloEvaluator(int64 max_loop_iterations) : max_loop_iterations_(max_loop_iterations) { typed_visitors_[PRED] = MakeUnique>(this); @@ -330,6 +330,24 @@ StatusOr> HloEvaluator::EvaluateElementwiseUnaryOp( return result; } +StatusOr> HloEvaluator::EvaluateDotOp( + const DotDimensionNumbers& dim_numbers, const Literal& lhs, + const Literal& rhs) { + std::unique_ptr lhs_instr = + HloInstruction::CreateConstant(lhs.CloneToUnique()); + std::unique_ptr rhs_instr = + HloInstruction::CreateConstant(rhs.CloneToUnique()); + + TF_ASSIGN_OR_RETURN( + Shape dot_shape, + ShapeInference::InferDotOpShape(lhs.shape(), rhs.shape(), dim_numbers)); + + std::unique_ptr cloned_instruction = + HloInstruction::CreateDot(dot_shape, lhs_instr.get(), rhs_instr.get(), + dim_numbers); + return Evaluate(cloned_instruction.get()); +} + Status HloEvaluator::HandleParameter(HloInstruction* parameter) { CHECK_LT(parameter->parameter_number(), arg_literals_.size()); const Literal* input_literal = arg_literals_[parameter->parameter_number()]; @@ -382,7 +400,7 @@ Status HloEvaluator::HandleConcatenate(HloInstruction* concatenate) { ShapeUtil::GetDimension(operand_shape, concat_dim); } - auto result_literal = Literal::CreateFromDimensions( + auto result_literal = LiteralUtil::CreateFromDimensions( reference_shape.element_type(), concat_dimensions); DimensionVector source_indices(rank, 0); DimensionVector dest_indices(concat_dimensions.size(), 0); @@ -533,7 +551,7 @@ Status HloEvaluator::HandleTuple(HloInstruction* tuple) { operand_literals.push_back(&GetEvaluatedLiteralFor(operand)); } - evaluated_[tuple] = Literal::MakeTuple(operand_literals); + evaluated_[tuple] = LiteralUtil::MakeTuple(operand_literals); return Status::OK(); } @@ -757,6 +775,12 @@ class OutputWindowIndexToInputIndex { return ArraySlice(input_index_); } + // Returns for a given 'input_dim' the corresponding output dimension index, + // or -1 if 'input_dim' is an elided window dimension. + int64 input_dim_value_to_output_index(int64 input_dim) { + return input_dim_value_to_output_index_[input_dim]; + } + private: // Propagates window dimensions from the output index to input_index_ by // mutating input_index_ in place. @@ -774,7 +798,7 @@ class OutputWindowIndexToInputIndex { // input_dim_value_to_index_vector_[i] tells us how to compute dimension i of // the input index from the output index. See - // PropagateOutputIndexToInputIndex. + // PropagateOutputIndexWindowDimsToInputIndex. std::vector input_dim_value_to_output_index_; // The result computed by this functor. operator() returns an ArraySlice into @@ -827,6 +851,8 @@ Status HloEvaluator::HandleGather(HloInstruction* gather) { // corresponding index in the input shape. std::vector input_index(operand.shape().dimensions_size()); std::vector output_index(gather->shape().dimensions_size()); + std::vector input_gather_index_clamped( + operand.shape().dimensions_size()); OutputGatherIndexToInputIndex output_gather_index_to_input_index( &gather->gather_dimension_numbers(), /*input_shape=*/operand.shape(), @@ -848,14 +874,26 @@ Status HloEvaluator::HandleGather(HloInstruction* gather) { output_index[i] = output_gather_index[i] + output_window_index[i]; DCHECK_LT(output_index[i], shape.dimensions(i)); } + for (int i = 0, e = input_gather_index.size(); i < e; i++) { + int64 output_dim = + output_window_index_to_input_index.input_dim_value_to_output_index(i); + // If 'output_dim' is -1, it means 'i' is an elided window dim. This means + // we set the iteration index to 0, so for the purpose of the following + // calculations we can consider the output dimension size to be 1. + int64 output_dim_size = + output_dim == -1 ? 1 : shape.dimensions(output_dim); + // Clamp the gather index so that the gather region fits in the operand. + // input_gather_index_clamped[i] = clamp(input_gather_index[i], 0, + // operand_shape.dimensions(i) - + // output_dim_size); + input_gather_index_clamped[i] = + std::min(operand_shape.dimensions(i) - output_dim_size, + std::max(0LL, input_gather_index[i])); + } for (int i = 0, e = input_index.size(); i < e; i++) { - // TODO(b/74360564): We should implement whatever out of bounds behavior - // we decide for dynamic-slice here as well. - input_index[i] = (input_gather_index[i] + input_window_index[i]) % - operand_shape.dimensions(i); - if (input_index[i] < 0) { - input_index[i] += operand_shape.dimensions(i); - } + input_index[i] = input_gather_index_clamped[i] + input_window_index[i]; + DCHECK_GE(input_index[i], 0); + DCHECK_LT(input_index[i], operand_shape.dimensions(i)); } TF_RETURN_IF_ERROR( result->CopyElementFrom(operand, input_index, output_index)); @@ -902,8 +940,8 @@ Status HloEvaluator::HandleBroadcast(HloInstruction* broadcast) { return Status::OK(); } -Status HloEvaluator::HandleGenerateToken(HloInstruction* token) { - evaluated_[token] = Literal::CreateToken(); +Status HloEvaluator::HandleAfterAll(HloInstruction* token) { + evaluated_[token] = LiteralUtil::CreateToken(); return Status::OK(); } @@ -1024,8 +1062,6 @@ Status HloEvaluator::HandleSelect(HloInstruction* select) { const auto& on_false = GetEvaluatedLiteralFor(select->operand(2)); // If predicate is of scalar type, no element-wise selection would be needed. - // This would also handle output array of tuple types as the DefaultAction - // would go through the HloEvaluatorTypedVisitor which doesn't handle tuples. if (ShapeUtil::IsScalar(pred.shape())) { if (pred.Get({})) { evaluated_[select] = on_true.CloneToUnique(); @@ -1038,6 +1074,19 @@ Status HloEvaluator::HandleSelect(HloInstruction* select) { return DefaultAction(select); } +Status HloEvaluator::HandleTupleSelect(HloInstruction* tuple_select) { + const auto& pred = GetEvaluatedLiteralFor(tuple_select->operand(0)); + const auto& on_true = GetEvaluatedLiteralFor(tuple_select->operand(1)); + const auto& on_false = GetEvaluatedLiteralFor(tuple_select->operand(2)); + + if (pred.Get({})) { + evaluated_[tuple_select] = on_true.CloneToUnique(); + } else { + evaluated_[tuple_select] = on_false.CloneToUnique(); + } + return Status::OK(); +} + Status HloEvaluator::HandleWhile(HloInstruction* while_hlo) { HloComputation* cond_comp = while_hlo->while_condition(); HloComputation* body_comp = while_hlo->while_body(); @@ -1068,6 +1117,107 @@ Status HloEvaluator::HandleWhile(HloInstruction* while_hlo) { return Status::OK(); } +// Key-value sort is a special snowflake: it's templated on two different +// element types, one for the keys, and one for the values. Jump through some +// hoops to make this work. +namespace { +template +std::unique_ptr EvaluateSortInternal(HloInstruction* sort, + const Literal& keys_literal, + const Literal& values_literal) { + CHECK_EQ(sort->operand_count(), 2); + // We need to sort and array of keys and an array of values, where the + // sorted order of the values is determined by the keys. The simplest(?) + // way to do this is to go to an array-of-pairs representation, sort the + // array using the keys, and then go back to pair-of-arrays. + VLOG(3) << "HandleSort keys_literal: " << keys_literal.ToString(); + VLOG(3) << "HandleSort values_literal: " << values_literal.ToString(); + const auto& keys_data = keys_literal.data(); + const auto& values_data = values_literal.data(); + using kv_pair = std::pair; + std::vector key_value_vector; + CHECK_EQ(keys_data.size(), values_data.size()); + key_value_vector.reserve(keys_data.size()); + for (int i = 0; i < keys_data.size(); ++i) { + key_value_vector.push_back(std::make_pair(keys_data[i], values_data[i])); + } + std::sort(key_value_vector.begin(), key_value_vector.end(), + [](const kv_pair& a, const kv_pair& b) { + return SafeLess(a.first, b.first); + }); + std::vector result_keys; + std::vector result_values; + for (const auto& key_value : key_value_vector) { + result_keys.push_back(key_value.first); + result_values.push_back(key_value.second); + } + auto result_keys_literal = MakeUnique(sort->operand(0)->shape()); + result_keys_literal->PopulateR1( + tensorflow::gtl::ArraySlice(result_keys)); + auto result_values_literal = MakeUnique(sort->operand(1)->shape()); + result_values_literal->PopulateR1( + tensorflow::gtl::ArraySlice(result_values)); + auto result_tuple = LiteralUtil::MakeTuple( + {result_keys_literal.get(), result_values_literal.get()}); + VLOG(3) << "HandleSort result_tuple: " << result_tuple->ToString(); + return result_tuple; +} + +template +StatusOr> EvaluateSortCurried( + HloInstruction* sort, const Literal& keys_literal, + const Literal& values_literal) { + switch (sort->operand(1)->shape().element_type()) { + case F32: + return EvaluateSortInternal(sort, keys_literal, + values_literal); + case U32: + return EvaluateSortInternal(sort, keys_literal, + values_literal); + case S32: + return EvaluateSortInternal(sort, keys_literal, + values_literal); + case BF16: + return EvaluateSortInternal(sort, keys_literal, + values_literal); + default: + return InvalidArgument("Unsupported type for Sort"); + } +} + +StatusOr> EvaluateSort(HloInstruction* sort, + const Literal& keys_literal, + const Literal& values_literal) { + switch (sort->operand(0)->shape().element_type()) { + case F32: + return EvaluateSortCurried(sort, keys_literal, values_literal); + case U32: + return EvaluateSortCurried(sort, keys_literal, values_literal); + case S32: + return EvaluateSortCurried(sort, keys_literal, values_literal); + case BF16: + return EvaluateSortCurried(sort, keys_literal, values_literal); + default: + return InvalidArgument("Unsupported type for Sort"); + } +} +} // namespace + +Status HloEvaluator::HandleSort(HloInstruction* sort) { + if (!ShapeUtil::IsTuple(sort->shape())) { + return DefaultAction(sort); + } else { + auto result = EvaluateSort(sort, GetEvaluatedLiteralFor(sort->operand(0)), + GetEvaluatedLiteralFor(sort->operand(1))); + if (result.ok()) { + evaluated_[sort] = std::move(result.ValueOrDie()); + return Status::OK(); + } else { + return result.status(); + } + } +} + Status HloEvaluator::Preprocess(HloInstruction* hlo) { VLOG(2) << "About to visit HLO: " << hlo->ToString(); return Status::OK(); diff --git a/tensorflow/compiler/xla/service/hlo_evaluator.h b/tensorflow/compiler/xla/service/hlo_evaluator.h index fc2fc9437b238a2e519401b2b121dfbef070e2dc..a4c37ef32827892194da070ee05ec6dc4f4c306f 100644 --- a/tensorflow/compiler/xla/service/hlo_evaluator.h +++ b/tensorflow/compiler/xla/service/hlo_evaluator.h @@ -23,6 +23,7 @@ limitations under the License. #include "tensorflow/compiler/xla/service/hlo_computation.h" #include "tensorflow/compiler/xla/service/hlo_instruction.h" #include "tensorflow/compiler/xla/service/hlo_module.h" +#include "tensorflow/compiler/xla/service/shape_inference.h" #include "tensorflow/compiler/xla/statusor.h" #include "tensorflow/compiler/xla/util.h" #include "tensorflow/compiler/xla/xla_data.pb.h" @@ -115,6 +116,10 @@ class HloEvaluator : public DfsHloVisitorWithDefault { StatusOr> EvaluateElementwiseUnaryOp( HloOpcode opcode, const Literal& operand); + StatusOr> EvaluateDotOp( + const DotDimensionNumbers& dim_numbers, const Literal& lhs, + const Literal& rhs); + protected: // Make HloEvaluatorTypedVisitor a friend because it is logically part of this // class. @@ -172,9 +177,13 @@ class HloEvaluator : public DfsHloVisitorWithDefault { Status HandleSelect(HloInstruction* select) override; + Status HandleTupleSelect(HloInstruction* tuple_select) override; + Status HandleBroadcast(HloInstruction* broadcast) override; - Status HandleGenerateToken(HloInstruction* token) override; + Status HandleAfterAll(HloInstruction* token) override; + + Status HandleSort(HloInstruction* sort) override; // Returns the already-evaluated literal result for the instruction. // A Constant instruction is considered evaluated and its literal will be diff --git a/tensorflow/compiler/xla/service/hlo_evaluator_test.cc b/tensorflow/compiler/xla/service/hlo_evaluator_test.cc index 42770d848a83b2e27b87bc963d259e2b7af664a4..5f575b24a1fb36c5384592028e0f1f6a8e9404b6 100644 --- a/tensorflow/compiler/xla/service/hlo_evaluator_test.cc +++ b/tensorflow/compiler/xla/service/hlo_evaluator_test.cc @@ -22,7 +22,7 @@ limitations under the License. #include #include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/reference_util.h" #include "tensorflow/compiler/xla/service/hlo_computation.h" #include "tensorflow/compiler/xla/service/hlo_element_type_converter.h" @@ -112,9 +112,9 @@ class HloEvaluatorTest : public ::testing::WithParamInterface, // Verifies that HloEvaluator evaluates a HLO instruction that performs clamp // with 3 operands. TEST_P(HloEvaluatorTest, DoesClamp) { - auto low = Literal::CreateR2({{0.f, 2.f}, {2.f, 4.f}}); - auto value = Literal::CreateR2({{0.f, 5.f}, {0.f, 4.f}}); - auto high = Literal::CreateR2({{2.f, 4.f}, {4.f, 4.f}}); + auto low = LiteralUtil::CreateR2({{0.f, 2.f}, {2.f, 4.f}}); + auto value = LiteralUtil::CreateR2({{0.f, 5.f}, {0.f, 4.f}}); + auto high = LiteralUtil::CreateR2({{2.f, 4.f}, {4.f, 4.f}}); Shape shape = low->shape(); HloComputation::Builder b(TestName()); @@ -127,15 +127,15 @@ TEST_P(HloEvaluatorTest, DoesClamp) { std::unique_ptr result = Evaluate(); - auto expected = Literal::CreateR2({{0, 4}, {2, 4}}); + auto expected = LiteralUtil::CreateR2({{0, 4}, {2, 4}}); EXPECT_TRUE(LiteralTestUtil::Equal(*expected, *result)); } TEST_P(HloEvaluatorTest, DISABLED_DoesClampSpecialBroadcast) { - auto low = Literal::CreateR0(0.f); - auto value = Literal::CreateR2({{-1.f, 0.f}, {1.f, 2.f}}); - auto high = Literal::CreateR0(1.f); + auto low = LiteralUtil::CreateR0(0.f); + auto value = LiteralUtil::CreateR2({{-1.f, 0.f}, {1.f, 2.f}}); + auto high = LiteralUtil::CreateR0(1.f); Shape shape = value->shape(); HloComputation::Builder b(TestName()); @@ -148,7 +148,7 @@ TEST_P(HloEvaluatorTest, DISABLED_DoesClampSpecialBroadcast) { std::unique_ptr result = Evaluate(); - auto expected = Literal::CreateR2({{0, 0}, {1, 1}}); + auto expected = LiteralUtil::CreateR2({{0, 0}, {1, 1}}); EXPECT_TRUE(LiteralTestUtil::Equal(*expected, *result)); } @@ -156,9 +156,9 @@ TEST_P(HloEvaluatorTest, DISABLED_DoesClampSpecialBroadcast) { // Verifies that HloEvaluator evaluates a HLO instruction that performs select // with 3 operands. TEST_P(HloEvaluatorTest, DoesSelect) { - auto pred = Literal::CreateR2({{true, false}, {false, true}}); - auto on_true = Literal::CreateR2({{2.f, 4.f}, {4.f, 4.f}}); - auto on_false = Literal::CreateR2({{0.f, 5.f}, {0.f, 4.f}}); + auto pred = LiteralUtil::CreateR2({{true, false}, {false, true}}); + auto on_true = LiteralUtil::CreateR2({{2.f, 4.f}, {4.f, 4.f}}); + auto on_false = LiteralUtil::CreateR2({{0.f, 5.f}, {0.f, 4.f}}); Shape shape = on_true->shape(); HloComputation::Builder b(TestName()); @@ -173,7 +173,7 @@ TEST_P(HloEvaluatorTest, DoesSelect) { std::unique_ptr result = Evaluate({}); - auto expected = Literal::CreateR2({{2, 5}, {0, 4}}); + auto expected = LiteralUtil::CreateR2({{2, 5}, {0, 4}}); EXPECT_TRUE(LiteralTestUtil::Equal(*expected, *result)); } @@ -181,46 +181,46 @@ TEST_P(HloEvaluatorTest, DoesSelect) { // Verifies that HloEvaluator evaluates a HLO instruction that performs // element-wise addition with 2 operands. TEST_P(HloEvaluatorTest, DoesAdd) { - auto lhs = Literal::CreateR2({{1, 0}, {-100, 4}}); - auto rhs = Literal::CreateR2({{2, 4}, {4, 4}}); - auto expected = Literal::CreateR2({{3, 4}, {-96, 8}}); + auto lhs = LiteralUtil::CreateR2({{1, 0}, {-100, 4}}); + auto rhs = LiteralUtil::CreateR2({{2, 4}, {4, 4}}); + auto expected = LiteralUtil::CreateR2({{3, 4}, {-96, 8}}); TestBinaryOp(HloOpcode::kAdd, std::move(expected), std::move(lhs), std::move(rhs)); } // Verifies that HloEvaluator evaluates a HLO instruction that performs // element-wise and with 2 operands. TEST_P(HloEvaluatorTest, DoesAnd) { - auto lhs = Literal::CreateR2({{1, 0}, {-100, 4}}); - auto rhs = Literal::CreateR2({{2, 4}, {4, 4}}); - auto expected = Literal::CreateR2({{0, 0}, {4, 4}}); + auto lhs = LiteralUtil::CreateR2({{1, 0}, {-100, 4}}); + auto rhs = LiteralUtil::CreateR2({{2, 4}, {4, 4}}); + auto expected = LiteralUtil::CreateR2({{0, 0}, {4, 4}}); TestBinaryOp(HloOpcode::kAnd, std::move(expected), std::move(lhs), std::move(rhs)); } // Verifies that HloEvaluator evaluates a HLO instruction that performs // element-wise or with 2 operands. TEST_P(HloEvaluatorTest, DoesOr) { - auto lhs = Literal::CreateR2({{1, 0}, {-100, 4}}); - auto rhs = Literal::CreateR2({{2, 4}, {4, 4}}); - auto expected = Literal::CreateR2({{3, 4}, {-100, 4}}); + auto lhs = LiteralUtil::CreateR2({{1, 0}, {-100, 4}}); + auto rhs = LiteralUtil::CreateR2({{2, 4}, {4, 4}}); + auto expected = LiteralUtil::CreateR2({{3, 4}, {-100, 4}}); TestBinaryOp(HloOpcode::kOr, std::move(expected), std::move(lhs), std::move(rhs)); } // Verifies that HloEvaluator evaluates a HLO instruction that performs // element-wise or with 2 operands. TEST_P(HloEvaluatorTest, DoesXor) { - auto lhs = Literal::CreateR2({{1, 0}, {-100, 4}}); - auto rhs = Literal::CreateR2({{2, 4}, {4, 4}}); - auto expected = Literal::CreateR2({{3, 4}, {-104, 0}}); + auto lhs = LiteralUtil::CreateR2({{1, 0}, {-100, 4}}); + auto rhs = LiteralUtil::CreateR2({{2, 4}, {4, 4}}); + auto expected = LiteralUtil::CreateR2({{3, 4}, {-104, 0}}); TestBinaryOp(HloOpcode::kXor, std::move(expected), std::move(lhs), std::move(rhs)); } // Verifies that HloEvaluator evaluates a HLO instruction that performs // element-wise multiply with 2 operands. TEST_P(HloEvaluatorTest, DoesMultiply) { - auto lhs = Literal::CreateR2({{-1, 0}, {-100, 4}}); - auto rhs = Literal::CreateR2( + auto lhs = LiteralUtil::CreateR2({{-1, 0}, {-100, 4}}); + auto rhs = LiteralUtil::CreateR2( {{std::numeric_limits::min(), 4}, {4, 4}}); - auto expected = Literal::CreateR2( + auto expected = LiteralUtil::CreateR2( {{std::numeric_limits::min(), 0}, {-400, 16}}); TestBinaryOp(HloOpcode::kMultiply, std::move(expected), std::move(lhs), std::move(rhs)); @@ -228,17 +228,17 @@ TEST_P(HloEvaluatorTest, DoesMultiply) { // Verifies that HloEvaluator evaluates a HLO instruction that performs // element-wise divide with 2 operands. TEST_P(HloEvaluatorTest, DoesDivideInt64) { - auto lhs = Literal::CreateR2({{1, 0}, {-100, 4}}); - auto rhs = Literal::CreateR2({{2, 4}, {4, 4}}); - auto expected = Literal::CreateR2({{0, 0}, {-25, 1}}); + auto lhs = LiteralUtil::CreateR2({{1, 0}, {-100, 4}}); + auto rhs = LiteralUtil::CreateR2({{2, 4}, {4, 4}}); + auto expected = LiteralUtil::CreateR2({{0, 0}, {-25, 1}}); TestBinaryOp(HloOpcode::kDivide, std::move(expected), std::move(lhs), std::move(rhs)); } TEST_P(HloEvaluatorTest, DoesDivideDouble) { - auto lhs = Literal::CreateR2({{1.0, 0.0}, {-100.0, 4.0}}); - auto rhs = Literal::CreateR2({{2.2, 4.0}, {4.0, 4.0}}); + auto lhs = LiteralUtil::CreateR2({{1.0, 0.0}, {-100.0, 4.0}}); + auto rhs = LiteralUtil::CreateR2({{2.2, 4.0}, {4.0, 4.0}}); auto expected = - Literal::CreateR2({{0.45454545454545453, 0}, {-25, 1}}); + LiteralUtil::CreateR2({{0.45454545454545453, 0}, {-25, 1}}); TestBinaryOp(HloOpcode::kDivide, std::move(expected), std::move(lhs), std::move(rhs)); } @@ -246,54 +246,54 @@ TEST_P(HloEvaluatorTest, DoesDivideDouble) { // Verifies that HloEvaluator evaluates a HLO instruction that performs // element-wise abs op with 1 operand. TEST_P(HloEvaluatorTest, DoesAbsR2) { - auto operand = Literal::CreateR2({{1, -20}, {-100, 4}}); - auto expected = Literal::CreateR2({{1, 20}, {100, 4}}); + auto operand = LiteralUtil::CreateR2({{1, -20}, {-100, 4}}); + auto expected = LiteralUtil::CreateR2({{1, 20}, {100, 4}}); TestUnaryOp(HloOpcode::kAbs, std::move(expected), std::move(operand)); } TEST_P(HloEvaluatorTest, DoesAbsR0) { - auto operand = Literal::CreateR0(-1.0f); - auto expected = Literal::CreateR0(1.0f); + auto operand = LiteralUtil::CreateR0(-1.0f); + auto expected = LiteralUtil::CreateR0(1.0f); TestUnaryOp(HloOpcode::kAbs, std::move(expected), std::move(operand)); } TEST_P(HloEvaluatorTest, DoesAbsR1WithZeroSize) { - auto operand = Literal::CreateR1({}); - auto expected = Literal::CreateR1({}); + auto operand = LiteralUtil::CreateR1({}); + auto expected = LiteralUtil::CreateR1({}); TestUnaryOp(HloOpcode::kAbs, std::move(expected), std::move(operand)); } TEST_P(HloEvaluatorTest, DoesNegateR2) { - auto operand = Literal::CreateR2( + auto operand = LiteralUtil::CreateR2( {{0, std::numeric_limits::min()}, {-1, 4}}); - auto expected = - Literal::CreateR2({{0, std::numeric_limits::min()}, {1, -4}}); + auto expected = LiteralUtil::CreateR2( + {{0, std::numeric_limits::min()}, {1, -4}}); TestUnaryOp(HloOpcode::kNegate, std::move(expected), std::move(operand)); } TEST_P(HloEvaluatorTest, DoesCosR2) { - auto operand = Literal::CreateR2({{0, M_PI}, {-M_PI, 2 * M_PI}}); - auto expected = Literal::CreateR2({{1, -1}, {-1, 1}}); + auto operand = LiteralUtil::CreateR2({{0, M_PI}, {-M_PI, 2 * M_PI}}); + auto expected = LiteralUtil::CreateR2({{1, -1}, {-1, 1}}); TestUnaryOp(HloOpcode::kCos, std::move(expected), std::move(operand), use_bfloat16_ ? 0.031250 : 9.5367431640625E-7); } TEST_P(HloEvaluatorTest, DoesSinR2) { - auto operand = Literal::CreateR2({{0, M_PI}, {-M_PI, 2 * M_PI}}); - auto expected = Literal::CreateR2({{0, 0}, {0, 0}}); + auto operand = LiteralUtil::CreateR2({{0, M_PI}, {-M_PI, 2 * M_PI}}); + auto expected = LiteralUtil::CreateR2({{0, 0}, {0, 0}}); TestUnaryOp(HloOpcode::kSin, std::move(expected), std::move(operand), use_bfloat16_ ? 0.031250 : 9.5367431640625E-7); } TEST_P(HloEvaluatorTest, DoesNotR2) { auto operand = - Literal::CreateR2({{0, std::numeric_limits::min()}, - {-1, std::numeric_limits::max()}}); + LiteralUtil::CreateR2({{0, std::numeric_limits::min()}, + {-1, std::numeric_limits::max()}}); auto expected = - Literal::CreateR2({{-1, std::numeric_limits::max()}, - {0, std::numeric_limits::min()}}); + LiteralUtil::CreateR2({{-1, std::numeric_limits::max()}, + {0, std::numeric_limits::min()}}); TestUnaryOp(HloOpcode::kNot, std::move(expected), std::move(operand)); } // Verifies that HloEvaluator evaluates a HLO Computation with non-parameter nor // constant operands. TEST_P(HloEvaluatorTest, DoesTraverseInstructions) { - auto lhs = Literal::CreateR2({{1, 0}, {-100, 4}}); - auto rhs = Literal::CreateR2({{2, 4}, {4, 4}}); - auto rhs2 = Literal::CreateR2({{1, -20}, {-100, 4}}); + auto lhs = LiteralUtil::CreateR2({{1, 0}, {-100, 4}}); + auto rhs = LiteralUtil::CreateR2({{2, 4}, {4, 4}}); + auto rhs2 = LiteralUtil::CreateR2({{1, -20}, {-100, 4}}); std::vector args = {lhs.get(), rhs.get(), rhs2.get()}; Shape shape = ShapeUtil::MakeShape(S64, {2, 2}); @@ -314,7 +314,7 @@ TEST_P(HloEvaluatorTest, DoesTraverseInstructions) { std::unique_ptr result = Evaluate(args); - auto expected = Literal::CreateR2({{4, -16}, {-196, 12}}); + auto expected = LiteralUtil::CreateR2({{4, -16}, {-196, 12}}); EXPECT_TRUE(LiteralTestUtil::Equal(*expected, *result)); } @@ -324,7 +324,7 @@ TEST_P(HloEvaluatorTest, DoesReshape) { HloComputation::Builder b(TestName()); const int64 dimensions[] = {11, 8, 7, 5, 9}; TF_ASSERT_OK_AND_ASSIGN(auto literal, - Literal::CreateRandomLiteral( + LiteralUtil::CreateRandomLiteral( ShapeUtil::MakeShape(F32, dimensions), 0.0, 1.0)); auto literal_clone = literal->CloneToUnique(); HloInstruction* literal_instruction = @@ -349,8 +349,8 @@ TEST_P(HloEvaluatorTest, DoesReshape) { // Verifies Broadcast operation is correctly evaluated. TEST_P(HloEvaluatorTest, DoesBroadcast) { HloComputation::Builder b(TestName()); - auto input_literal = Literal::CreateR2({{1, 2}, {3, 4}, {5, 6}}); - auto output_literal = Literal::CreateR3( + auto input_literal = LiteralUtil::CreateR2({{1, 2}, {3, 4}, {5, 6}}); + auto output_literal = LiteralUtil::CreateR3( {{{1, 2}, {3, 4}, {5, 6}}, {{1, 2}, {3, 4}, {5, 6}}}); HloInstruction* literal_instruction = b.AddInstruction( HloInstruction::CreateConstant(std::move(input_literal))); @@ -365,8 +365,8 @@ TEST_P(HloEvaluatorTest, DoesBroadcast) { TEST_P(HloEvaluatorTest, DoesBroadcastScalar) { HloComputation::Builder b(TestName()); - auto input_literal = Literal::CreateR0(111); - auto output_literal = Literal::CreateR2( + auto input_literal = LiteralUtil::CreateR0(111); + auto output_literal = LiteralUtil::CreateR2( {{111, 111}, {111, 111}, {111, 111}, {111, 111}, {111, 111}, {111, 111}}); HloInstruction* literal_instruction = b.AddInstruction( @@ -386,9 +386,9 @@ TEST_P(HloEvaluatorTest, DoesConcatenateSimple) { HloComputation::Builder b(TestName()); HloInstruction* operand1 = b.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR2({{-1, -2}, {100, 200}}))); + LiteralUtil::CreateR2({{-1, -2}, {100, 200}}))); HloInstruction* operand2 = b.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR2({{-2, -3}, {-100, -200}}))); + LiteralUtil::CreateR2({{-2, -3}, {-100, -200}}))); std::vector operands = {operand1, operand2}; @@ -399,8 +399,8 @@ TEST_P(HloEvaluatorTest, DoesConcatenateSimple) { std::unique_ptr result = Evaluate(); - auto expected = - Literal::CreateR2({{-1, -2}, {100, 200}, {-2, -3}, {-100, -200}}); + auto expected = LiteralUtil::CreateR2( + {{-1, -2}, {100, 200}, {-2, -3}, {-100, -200}}); EXPECT_TRUE(LiteralTestUtil::Equal(*expected, *result)); } @@ -408,9 +408,9 @@ TEST_P(HloEvaluatorTest, ConcatenateHandlesShapeWithZeroElement) { HloComputation::Builder b(TestName()); HloInstruction* operand1 = b.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR1({100, 200}))); + HloInstruction::CreateConstant(LiteralUtil::CreateR1({100, 200}))); HloInstruction* operand2 = b.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR1({}))); + HloInstruction::CreateConstant(LiteralUtil::CreateR1({}))); std::vector operands = {operand1, operand2}; @@ -421,16 +421,16 @@ TEST_P(HloEvaluatorTest, ConcatenateHandlesShapeWithZeroElement) { std::unique_ptr result = Evaluate(); - auto expected = Literal::CreateR1({100, 200}); + auto expected = LiteralUtil::CreateR1({100, 200}); EXPECT_TRUE(LiteralTestUtil::Equal(*expected, *result)); } TEST_P(HloEvaluatorTest, ConvertWithSameLayout) { HloComputation::Builder b(TestName()); - auto input_literal = Literal::CreateR2({{1, 2}, {3, 4}, {5, 6}}); + auto input_literal = LiteralUtil::CreateR2({{1, 2}, {3, 4}, {5, 6}}); auto expected = - Literal::CreateR2({{1.0, 2.0}, {3.0, 4.0}, {5.0, 6.0}}); + LiteralUtil::CreateR2({{1.0, 2.0}, {3.0, 4.0}, {5.0, 6.0}}); ASSERT_TRUE(LayoutUtil::LayoutsInShapesEqual(input_literal->shape(), expected->shape())); @@ -447,9 +447,9 @@ TEST_P(HloEvaluatorTest, ConvertWithSameLayout) { TEST_P(HloEvaluatorTest, ConvertWithDifferentLayout) { HloComputation::Builder b(TestName()); - auto input_literal = Literal::CreateR2WithLayout( + auto input_literal = LiteralUtil::CreateR2WithLayout( {{1, 2}, {3, 4}, {5, 6}}, LayoutUtil::MakeLayout({0, 1})); - auto expected = Literal::CreateR2WithLayout( + auto expected = LiteralUtil::CreateR2WithLayout( {{1.0, 2.0}, {3.0, 4.0}, {5.0, 6.0}}, LayoutUtil::MakeLayout({1, 0})); ASSERT_FALSE(LayoutUtil::LayoutsInShapesEqual(input_literal->shape(), expected->shape())); @@ -478,13 +478,13 @@ PaddingConfig CreatePaddingConfig( } TEST_P(HloEvaluatorTest, Pad2DIntegerArrayWithZeroDimension) { - auto operand = Literal::CreateR2({{}, {}}); + auto operand = LiteralUtil::CreateR2({{}, {}}); HloComputation::Builder b(TestName()); auto operand_instruction = b.AddInstruction(HloInstruction::CreateConstant(std::move(operand))); constexpr int32 kPadValue = 10; - auto pad_value = Literal::CreateR0(kPadValue); + auto pad_value = LiteralUtil::CreateR0(kPadValue); auto padding_value_instruction = b.AddInstruction(HloInstruction::CreateConstant(std::move(pad_value))); @@ -496,7 +496,7 @@ TEST_P(HloEvaluatorTest, Pad2DIntegerArrayWithZeroDimension) { std::unique_ptr result = Evaluate(); - auto expected = Literal::CreateR2( + auto expected = LiteralUtil::CreateR2( {{10, 10}, {10, 10}, {10, 10}, {10, 10}, {10, 10}}); EXPECT_TRUE(LiteralTestUtil::Equal(*expected, *result)); @@ -506,11 +506,11 @@ TEST_P(HloEvaluatorTest, Pad4DFloatArrayWithInteriorPadding) { HloComputation::Builder b(TestName()); Array4D input_array(3, 2, 1, 1, {1, 2, 3, 4, 5, 6}); - auto input = Literal::CreateR4FromArray4D(input_array); + auto input = LiteralUtil::CreateR4FromArray4D(input_array); HloInstruction* input_instruction = b.AddInstruction(HloInstruction::CreateConstant(std::move(input))); constexpr float kPadValue = 1.5; - auto pad_value = Literal::CreateR0(kPadValue); + auto pad_value = LiteralUtil::CreateR0(kPadValue); HloInstruction* pad_instruction = b.AddInstruction(HloInstruction::CreateConstant(std::move(pad_value))); @@ -532,7 +532,7 @@ TEST_P(HloEvaluatorTest, Pad4DFloatArrayWithInteriorPadding) { (*expected_array)(7, 0, 0, 0) = 5.0f; (*expected_array)(7, 2, 0, 0) = 6.0f; - auto expected = Literal::CreateR4FromArray4D(*expected_array); + auto expected = LiteralUtil::CreateR4FromArray4D(*expected_array); EXPECT_TRUE(LiteralTestUtil::Equal(*expected, *result)); } @@ -549,12 +549,12 @@ TEST_P(HloEvaluatorTest, NegativePadding2D) { // } auto input_array = MakeUnique>(4, 3); input_array->FillUnique(1.0f); - auto input = Literal::CreateR2FromArray2D(*input_array); + auto input = LiteralUtil::CreateR2FromArray2D(*input_array); HloInstruction* input_instruction = b.AddInstruction(HloInstruction::CreateConstant(std::move(input))); auto pad_value_instruction = b.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(2.718f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(2.718f))); auto r2_padding_on_dim0_dim1 = CreatePaddingConfig({{{-1, -2, 0}}, {{-2, 4, 0}}}); @@ -574,7 +574,7 @@ TEST_P(HloEvaluatorTest, NegativePadding2D) { (*expected_array)(0, 2) = 2.718f; (*expected_array)(0, 3) = 2.718f; (*expected_array)(0, 4) = 2.718f; - auto expected = Literal::CreateR2FromArray2D(*expected_array); + auto expected = LiteralUtil::CreateR2FromArray2D(*expected_array); EXPECT_TRUE(LiteralTestUtil::Near(*expected, *result, ErrorSpec(0.031250))); } @@ -590,12 +590,12 @@ TEST_P(HloEvaluatorTest, NegativeAndInteriorPadding2D) { // } auto input_array = MakeUnique>(4, 3); input_array->FillUnique(1.0f); - auto input = Literal::CreateR2FromArray2D(*input_array); + auto input = LiteralUtil::CreateR2FromArray2D(*input_array); HloInstruction* input_instruction = b.AddInstruction(HloInstruction::CreateConstant(std::move(input))); auto pad_value_instruction = b.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(2.718f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(2.718f))); PaddingConfig padding_config = MakeNoPaddingConfig(2); @@ -613,7 +613,7 @@ TEST_P(HloEvaluatorTest, NegativeAndInteriorPadding2D) { std::unique_ptr result = Evaluate(); auto expected_array = MakeUnique>(0, 9); - auto expected = Literal::CreateR2FromArray2D(*expected_array); + auto expected = LiteralUtil::CreateR2FromArray2D(*expected_array); EXPECT_TRUE(LiteralTestUtil::Equal(*expected, *result)); } @@ -630,13 +630,13 @@ TEST_P(HloEvaluatorTest, DotRank2AndRank1) { // } auto lhs_array = MakeUnique>(4, 1); lhs_array->FillUnique(1.0f); - auto lhs_literal = Literal::CreateR2FromArray2D(*lhs_array); + auto lhs_literal = LiteralUtil::CreateR2FromArray2D(*lhs_array); HloInstruction* lhs_instruction = b.AddInstruction(HloInstruction::CreateConstant(std::move(lhs_literal))); // rhs: // f32[2] { 1, 2 }, - auto rhs_literal = Literal::CreateR2({{1, 2}}); + auto rhs_literal = LiteralUtil::CreateR2({{1, 2}}); HloInstruction* rhs_instruction = b.AddInstruction(HloInstruction::CreateConstant(std::move(rhs_literal))); @@ -658,7 +658,7 @@ TEST_P(HloEvaluatorTest, DotRank2AndRank1) { {4.f, 8.f}, }); // clang-format on - auto expected = Literal::CreateR2FromArray2D(expected_array); + auto expected = LiteralUtil::CreateR2FromArray2D(expected_array); EXPECT_TRUE(LiteralTestUtil::Equal(*expected, *result)); } @@ -669,7 +669,7 @@ TEST_P(HloEvaluatorTest, DotRank1AndRank2) { // lhs: // f32[3] // { 1, 2, 3 }, - auto lhs_literal = Literal::CreateR1({1, 2, 3}); + auto lhs_literal = LiteralUtil::CreateR1({1, 2, 3}); HloInstruction* lhs_instruction = b.AddInstruction(HloInstruction::CreateConstant(std::move(lhs_literal))); @@ -681,7 +681,7 @@ TEST_P(HloEvaluatorTest, DotRank1AndRank2) { // } auto rhs_array = MakeUnique>(3, 2); rhs_array->FillUnique(1.0f); - auto rhs_literal = Literal::CreateR2FromArray2D(*rhs_array); + auto rhs_literal = LiteralUtil::CreateR2FromArray2D(*rhs_array); HloInstruction* rhs_instruction = b.AddInstruction(HloInstruction::CreateConstant(std::move(rhs_literal))); @@ -695,7 +695,7 @@ TEST_P(HloEvaluatorTest, DotRank1AndRank2) { std::unique_ptr result = Evaluate(); - auto expected = Literal::CreateR1({22.f, 28.f}); + auto expected = LiteralUtil::CreateR1({22.f, 28.f}); EXPECT_TRUE(LiteralTestUtil::Equal(*expected, *result)); } @@ -712,7 +712,7 @@ TEST_P(HloEvaluatorTest, DotRank2AndRank2) { // } auto lhs_array = MakeUnique>(4, 3); lhs_array->FillUnique(1.0f); - auto lhs_literal = Literal::CreateR2FromArray2D(*lhs_array); + auto lhs_literal = LiteralUtil::CreateR2FromArray2D(*lhs_array); HloInstruction* lhs_instruction = b.AddInstruction(HloInstruction::CreateConstant(std::move(lhs_literal))); @@ -724,7 +724,7 @@ TEST_P(HloEvaluatorTest, DotRank2AndRank2) { // } auto rhs_array = MakeUnique>(3, 2); rhs_array->FillUnique(1.0f); - auto rhs_literal = Literal::CreateR2FromArray2D(*rhs_array); + auto rhs_literal = LiteralUtil::CreateR2FromArray2D(*rhs_array); HloInstruction* rhs_instruction = b.AddInstruction(HloInstruction::CreateConstant(std::move(rhs_literal))); @@ -744,7 +744,7 @@ TEST_P(HloEvaluatorTest, DotRank2AndRank2) { {94.f, 124.f}, {130.f, 172.f}, }); - auto expected = Literal::CreateR2FromArray2D(expected_array); + auto expected = LiteralUtil::CreateR2FromArray2D(expected_array); EXPECT_TRUE(LiteralTestUtil::Equal(*expected, *result)); } @@ -753,12 +753,12 @@ TEST_P(HloEvaluatorTest, SimpleConv1D) { HloComputation::Builder b(TestName()); Array3D lhs_array = {{{1, 2, 3}}}; - auto lhs_literal = Literal::CreateR3FromArray3D(lhs_array); + auto lhs_literal = LiteralUtil::CreateR3FromArray3D(lhs_array); HloInstruction* lhs_instruction = b.AddInstruction(HloInstruction::CreateConstant(std::move(lhs_literal))); Array3D rhs_array = {{{3.f, 4.f}}}; - auto rhs_literal = Literal::CreateR3FromArray3D(rhs_array); + auto rhs_literal = LiteralUtil::CreateR3FromArray3D(rhs_array); HloInstruction* rhs_instruction = b.AddInstruction(HloInstruction::CreateConstant(std::move(rhs_literal))); @@ -792,7 +792,7 @@ TEST_P(HloEvaluatorTest, SimpleConv1D) { std::unique_ptr result = Evaluate(); Array3D expected_array = {{{11.f, 18.f, 9.f}}}; - auto expected = Literal::CreateR3FromArray3D(expected_array); + auto expected = LiteralUtil::CreateR3FromArray3D(expected_array); EXPECT_TRUE(LiteralTestUtil::Equal(*expected, *result)); } @@ -809,7 +809,7 @@ TEST_P(HloEvaluatorTest, Simple4x4Conv2DWith2x2Kernel) { {13, 14, 15, 16}, })); // clang-format on - auto lhs_literal = Literal::CreateR4FromArray4D(lhs_array); + auto lhs_literal = LiteralUtil::CreateR4FromArray4D(lhs_array); HloInstruction* lhs_instruction = b.AddInstruction(HloInstruction::CreateConstant(std::move(lhs_literal))); @@ -820,7 +820,7 @@ TEST_P(HloEvaluatorTest, Simple4x4Conv2DWith2x2Kernel) { {7, 8}, })); // clang-format on - auto rhs_literal = Literal::CreateR4FromArray4D(rhs_array); + auto rhs_literal = LiteralUtil::CreateR4FromArray4D(rhs_array); HloInstruction* rhs_instruction = b.AddInstruction(HloInstruction::CreateConstant(std::move(rhs_literal))); @@ -854,7 +854,7 @@ TEST_P(HloEvaluatorTest, Simple4x4Conv2DWith2x2Kernel) { {149, 160, 171, 80}, })); // clang-format on - auto expected = Literal::CreateR4FromArray4D(expected_array); + auto expected = LiteralUtil::CreateR4FromArray4D(expected_array); EXPECT_TRUE(LiteralTestUtil::Equal(*expected, *result)); } @@ -884,11 +884,11 @@ TEST_P(HloEvaluatorTest, Conv2DGeneralDimensionsReversed) { }}); // clang-format on - auto lhs_literal = Literal::CreateR4FromArray4D(input); + auto lhs_literal = LiteralUtil::CreateR4FromArray4D(input); HloInstruction* lhs_instruction = b.AddInstruction(HloInstruction::CreateConstant(std::move(lhs_literal))); - auto rhs_literal = Literal::CreateR4FromArray4D(weight); + auto rhs_literal = LiteralUtil::CreateR4FromArray4D(weight); HloInstruction* rhs_instruction = b.AddInstruction(HloInstruction::CreateConstant(std::move(rhs_literal))); rhs_instruction = b.AddInstruction(HloInstruction::CreateReverse( @@ -933,7 +933,7 @@ TEST_P(HloEvaluatorTest, Conv2DGeneralDimensionsReversed) { Array4D expected_array({{{{2514, 2685}}}}); Array4D expected_array_bf16({{{{2512, 2672}}}}); // clang-format on - auto expected = Literal::CreateR4FromArray4D( + auto expected = LiteralUtil::CreateR4FromArray4D( use_bfloat16_ ? expected_array_bf16 : expected_array); EXPECT_TRUE(LiteralTestUtil::Equal(*expected, *result)); @@ -964,11 +964,11 @@ TEST_P(HloEvaluatorTest, Conv2DGeneralDimensions) { }}); // clang-format on - auto lhs_literal = Literal::CreateR4FromArray4D(input); + auto lhs_literal = LiteralUtil::CreateR4FromArray4D(input); HloInstruction* lhs_instruction = b.AddInstruction(HloInstruction::CreateConstant(std::move(lhs_literal))); - auto rhs_literal = Literal::CreateR4FromArray4D(weight); + auto rhs_literal = LiteralUtil::CreateR4FromArray4D(weight); HloInstruction* rhs_instruction = b.AddInstruction(HloInstruction::CreateConstant(std::move(rhs_literal))); @@ -1010,7 +1010,7 @@ TEST_P(HloEvaluatorTest, Conv2DGeneralDimensions) { Array4D expected_array({{{{2514, 2685}}}}); Array4D expected_array_bf16({{{{2512, 2672}}}}); // clang-format on - auto expected = Literal::CreateR4FromArray4D( + auto expected = LiteralUtil::CreateR4FromArray4D( use_bfloat16_ ? expected_array_bf16 : expected_array); EXPECT_TRUE(LiteralTestUtil::Equal(*expected, *result)); @@ -1028,7 +1028,7 @@ TEST_P(HloEvaluatorTest, DilatedBaseConv2DWithHighPadding) { {13, 14, 15, 16}, })); // clang-format on - auto lhs_literal = Literal::CreateR4FromArray4D(lhs_array); + auto lhs_literal = LiteralUtil::CreateR4FromArray4D(lhs_array); HloInstruction* lhs_instruction = b.AddInstruction(HloInstruction::CreateConstant(std::move(lhs_literal))); @@ -1039,7 +1039,7 @@ TEST_P(HloEvaluatorTest, DilatedBaseConv2DWithHighPadding) { {7, 8}, })); // clang-format on - auto rhs_literal = Literal::CreateR4FromArray4D(rhs_array); + auto rhs_literal = LiteralUtil::CreateR4FromArray4D(rhs_array); HloInstruction* rhs_instruction = b.AddInstruction(HloInstruction::CreateConstant(std::move(rhs_literal))); @@ -1074,7 +1074,7 @@ TEST_P(HloEvaluatorTest, DilatedBaseConv2DWithHighPadding) { {91, 112, 98, 120, 105, 128, 112}, {65, 84, 70, 90, 75, 96, 80}, })); - auto expected = Literal::CreateR4FromArray4D(expected_array); + auto expected = LiteralUtil::CreateR4FromArray4D(expected_array); EXPECT_TRUE(LiteralTestUtil::Equal(*expected, *result)); } @@ -1091,7 +1091,7 @@ TEST_P(HloEvaluatorTest, DilatedBaseConv2DWithLowAndHighPadding) { {13, 14, 15, 16}, })); // clang-format on - auto lhs_literal = Literal::CreateR4FromArray4D(lhs_array); + auto lhs_literal = LiteralUtil::CreateR4FromArray4D(lhs_array); HloInstruction* lhs_instruction = b.AddInstruction(HloInstruction::CreateConstant(std::move(lhs_literal))); @@ -1102,7 +1102,7 @@ TEST_P(HloEvaluatorTest, DilatedBaseConv2DWithLowAndHighPadding) { {7, 8}, })); // clang-format on - auto rhs_literal = Literal::CreateR4FromArray4D(rhs_array); + auto rhs_literal = LiteralUtil::CreateR4FromArray4D(rhs_array); HloInstruction* rhs_instruction = b.AddInstruction(HloInstruction::CreateConstant(std::move(rhs_literal))); @@ -1138,7 +1138,7 @@ TEST_P(HloEvaluatorTest, DilatedBaseConv2DWithLowAndHighPadding) { {104, 91, 112, 98, 120, 105, 128, 112}, {78, 65, 84, 70, 90, 75, 96, 80}, })); - auto expected = Literal::CreateR4FromArray4D(expected_array); + auto expected = LiteralUtil::CreateR4FromArray4D(expected_array); EXPECT_TRUE(LiteralTestUtil::Equal(*expected, *result)); } @@ -1156,7 +1156,7 @@ TEST_P(HloEvaluatorTest, {13, 14, 15, 16}, })); // clang-format on - auto lhs_literal = Literal::CreateR4FromArray4D(lhs_array); + auto lhs_literal = LiteralUtil::CreateR4FromArray4D(lhs_array); HloInstruction* lhs_instruction = b.AddInstruction(HloInstruction::CreateConstant(std::move(lhs_literal))); @@ -1167,7 +1167,7 @@ TEST_P(HloEvaluatorTest, {8, 9, 10}, })); // clang-format on - auto rhs_literal = Literal::CreateR4FromArray4D(rhs_array); + auto rhs_literal = LiteralUtil::CreateR4FromArray4D(rhs_array); HloInstruction* rhs_instruction = b.AddInstruction(HloInstruction::CreateConstant(std::move(rhs_literal))); @@ -1210,7 +1210,7 @@ TEST_P(HloEvaluatorTest, {0, 0, 0}, {91, 98, 105}, })); - auto expected = Literal::CreateR4FromArray4D(expected_array); + auto expected = LiteralUtil::CreateR4FromArray4D(expected_array); EXPECT_TRUE(LiteralTestUtil::Equal(*expected, *result)); } @@ -1225,9 +1225,9 @@ TEST_F(HloEvaluatorPreciseReduceTest, AddReductionPrecisionTest) { constexpr int kNumElements = 1 << 25; // float += 1 saturates at 1<<24 std::vector v(kNumElements, 1.0f); HloInstruction* arg_instruction = b.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR1(v))); + HloInstruction::CreateConstant(LiteralUtil::CreateR1(v))); HloInstruction* init_value = b.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(0.f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(0.f))); HloComputation::Builder add_computation("add"); Shape scalar_shape = ShapeUtil::MakeShape(F32, {}); @@ -1262,9 +1262,9 @@ void BM_ReducePrecisely(int num_iters) { constexpr int kNumElements = 1 << 25; // float += 1 saturates at 1<<24 std::vector v(kNumElements, 1.0f); HloInstruction* arg_instruction = b.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR1(v))); + HloInstruction::CreateConstant(LiteralUtil::CreateR1(v))); auto init_value = b.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(0.f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(0.f))); HloComputation::Builder add_computation("add"); Shape scalar_shape = ShapeUtil::MakeShape(F32, {}); @@ -1299,13 +1299,13 @@ TEST_P(HloEvaluatorTest, ReduceAdd) { // } auto arg_array = MakeUnique>(2, 3); arg_array->FillUnique(1.0f); - auto arg_literal = Literal::CreateR2FromArray2D(*arg_array); + auto arg_literal = LiteralUtil::CreateR2FromArray2D(*arg_array); HloInstruction* arg_instruction = b.AddInstruction(HloInstruction::CreateConstant(std::move(arg_literal))); auto init_value = b.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(0.f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(0.f))); HloComputation::Builder add_computation("add"); Shape scalar_shape = ShapeUtil::MakeShape(F32, {}); @@ -1326,7 +1326,7 @@ TEST_P(HloEvaluatorTest, ReduceAdd) { std::unique_ptr result = Evaluate(); - auto expected = Literal::CreateR1({6, 18}); + auto expected = LiteralUtil::CreateR1({6, 18}); EXPECT_TRUE(LiteralTestUtil::Equal(*expected, *result)); } @@ -1341,13 +1341,13 @@ TEST_P(HloEvaluatorTest, ReduceWindowMax) { // } auto arg_array = MakeUnique>(2, 3); arg_array->FillUnique(1.0f); - auto arg_literal = Literal::CreateR2FromArray2D(*arg_array); + auto arg_literal = LiteralUtil::CreateR2FromArray2D(*arg_array); HloInstruction* arg_instruction = b.AddInstruction(HloInstruction::CreateConstant(std::move(arg_literal))); auto init_value = b.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(0.f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(0.f))); HloComputation::Builder max_computation("max"); Shape scalar_shape = ShapeUtil::MakeShape(F32, {}); @@ -1378,7 +1378,7 @@ TEST_P(HloEvaluatorTest, ReduceWindowMax) { std::unique_ptr result = Evaluate(); - auto expected = Literal::CreateR2({{6, 7}}); + auto expected = LiteralUtil::CreateR2({{6, 7}}); EXPECT_TRUE(LiteralTestUtil::Equal(*expected, *result)); } @@ -1392,13 +1392,13 @@ TEST_P(HloEvaluatorTest, ReduceWindowAdd) { // } auto arg_array = MakeUnique>(2, 3); arg_array->FillUnique(1.0f); - auto arg_literal = Literal::CreateR2FromArray2D(*arg_array); + auto arg_literal = LiteralUtil::CreateR2FromArray2D(*arg_array); HloInstruction* arg_instruction = b.AddInstruction(HloInstruction::CreateConstant(std::move(arg_literal))); auto init_value = b.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(0.f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(0.f))); HloComputation::Builder add_computation("add"); Shape scalar_shape = ShapeUtil::MakeShape(F32, {}); @@ -1435,7 +1435,7 @@ TEST_P(HloEvaluatorTest, ReduceWindowAdd) { std::unique_ptr result = Evaluate(); - auto expected = Literal::CreateR2({{1, 3, 5}, {5, 11, 13}}); + auto expected = LiteralUtil::CreateR2({{1, 3, 5}, {5, 11, 13}}); EXPECT_TRUE(LiteralTestUtil::Equal(*expected, *result)); } @@ -1445,13 +1445,13 @@ TEST_P(HloEvaluatorTest, ReduceWindowAdd6D) { // arg: f32[4,4,4,4,4,4] full of ones. Using small dims to limit run-time. std::vector input_dims(6, 4); std::unique_ptr arg_literal = - Literal::CreateFullWithDescendingLayout(input_dims, 1.0f); + LiteralUtil::CreateFullWithDescendingLayout(input_dims, 1.0f); HloInstruction* arg_instruction = b.AddInstruction(HloInstruction::CreateConstant(std::move(arg_literal))); auto init_value = b.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(0.f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(0.f))); HloComputation::Builder add_computation("add"); Shape scalar_shape = ShapeUtil::MakeShape(F32, {}); @@ -1498,7 +1498,7 @@ TEST_P(HloEvaluatorTest, ReduceWindowAdd6D) { std::vector output_dims = {4, 3, 3, 3, 4, 4}; std::unique_ptr result_literal = - Literal::CreateFullWithDescendingLayout(output_dims, 8.0f); + LiteralUtil::CreateFullWithDescendingLayout(output_dims, 8.0f); EXPECT_TRUE(LiteralTestUtil::Equal(*result_literal, *result)); } @@ -1513,7 +1513,8 @@ TEST_P(HloEvaluatorTest, StridedSlice) { // } auto operand_array = MakeUnique>(3, 5); operand_array->FillUnique(1.0f); - auto operand_literal = Literal::CreateR2FromArray2D(*operand_array); + auto operand_literal = + LiteralUtil::CreateR2FromArray2D(*operand_array); HloInstruction* operand = b.AddInstruction( HloInstruction::CreateConstant(std::move(operand_literal))); @@ -1527,7 +1528,7 @@ TEST_P(HloEvaluatorTest, StridedSlice) { std::unique_ptr result = Evaluate(); - auto expected = Literal::CreateR2({ + auto expected = LiteralUtil::CreateR2({ {3}, {19}, }); @@ -1545,13 +1546,14 @@ TEST_P(HloEvaluatorTest, DynamicSlice) { // } auto operand_array = MakeUnique>(2, 4); operand_array->FillUnique(1.0f); - auto operand_literal = Literal::CreateR2FromArray2D(*operand_array); + auto operand_literal = + LiteralUtil::CreateR2FromArray2D(*operand_array); HloInstruction* operand = b.AddInstruction( HloInstruction::CreateConstant(std::move(operand_literal))); auto start_indices = b.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR1({0, 1}))); + HloInstruction::CreateConstant(LiteralUtil::CreateR1({0, 1}))); Shape shape = ShapeUtil::MakeShape(F32, {2, 3}); b.AddInstruction(HloInstruction::CreateDynamicSlice(shape, operand, @@ -1560,7 +1562,7 @@ TEST_P(HloEvaluatorTest, DynamicSlice) { std::unique_ptr result = Evaluate(); - auto expected = Literal::CreateR2({ + auto expected = LiteralUtil::CreateR2({ {2, 3, 4}, {6, 7, 8}, }); @@ -1580,13 +1582,14 @@ TEST_P(HloEvaluatorTest, DynamicSliceModSlice) { // } auto operand_array = MakeUnique>(2, 4); operand_array->FillUnique(1.0f); - auto operand_literal = Literal::CreateR2FromArray2D(*operand_array); + auto operand_literal = + LiteralUtil::CreateR2FromArray2D(*operand_array); HloInstruction* operand = b.AddInstruction( HloInstruction::CreateConstant(std::move(operand_literal))); auto start_indices = b.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR1({2, 1}))); + HloInstruction::CreateConstant(LiteralUtil::CreateR1({2, 1}))); Shape shape = ShapeUtil::MakeShape(F32, {2, 3}); b.AddInstruction(HloInstruction::CreateDynamicSlice(shape, operand, @@ -1595,7 +1598,7 @@ TEST_P(HloEvaluatorTest, DynamicSliceModSlice) { std::unique_ptr result = Evaluate(); - auto expected = Literal::CreateR2({ + auto expected = LiteralUtil::CreateR2({ {2, 3, 4}, {6, 7, 8}, }); @@ -1613,16 +1616,17 @@ TEST_P(HloEvaluatorTest, DynamicSliceUpdate) { // } auto operand_array = MakeUnique>(2, 3); operand_array->FillUnique(1.0); - auto operand_literal = Literal::CreateR2FromArray2D(*operand_array); + auto operand_literal = + LiteralUtil::CreateR2FromArray2D(*operand_array); HloInstruction* operand = b.AddInstruction( HloInstruction::CreateConstant(std::move(operand_literal))); auto start_indices = b.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR1({0, 1}))); + HloInstruction::CreateConstant(LiteralUtil::CreateR1({0, 1}))); auto update = b.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR2({{-2.0, -3.0}, {-6.0, -7.0}}))); + LiteralUtil::CreateR2({{-2.0, -3.0}, {-6.0, -7.0}}))); Shape shape = ShapeUtil::MakeShape(F64, {2, 3}); b.AddInstruction(HloInstruction::CreateDynamicUpdateSlice( @@ -1631,7 +1635,7 @@ TEST_P(HloEvaluatorTest, DynamicSliceUpdate) { std::unique_ptr result = Evaluate(); - auto expected = Literal::CreateR2({ + auto expected = LiteralUtil::CreateR2({ {1, -2, -3}, {5, -6, -7}, }); @@ -1649,12 +1653,13 @@ TEST_P(HloEvaluatorTest, SetAndGetTuples) { // } auto operand_array = MakeUnique>(2, 3); operand_array->FillUnique(1.0); - auto operand_literal2 = Literal::CreateR2FromArray2D(*operand_array); + auto operand_literal2 = + LiteralUtil::CreateR2FromArray2D(*operand_array); HloInstruction* operand2 = b.AddInstruction( HloInstruction::CreateConstant(std::move(operand_literal2))); HloInstruction* operand1 = b.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR1({0, 1}))); + HloInstruction::CreateConstant(LiteralUtil::CreateR1({0, 1}))); auto tuple = b.AddInstruction(HloInstruction::CreateTuple({operand1, operand2})); @@ -1666,7 +1671,7 @@ TEST_P(HloEvaluatorTest, SetAndGetTuples) { std::unique_ptr result = Evaluate(); - auto expected = Literal::CreateR2({ + auto expected = LiteralUtil::CreateR2({ {1, 2, 3}, {5, 6, 7}, }); @@ -1686,9 +1691,9 @@ TEST_P(HloEvaluatorTest, SetAndGetNestedTuples) { operand_array->FillUnique(1.0); HloInstruction* operand2 = b.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR2FromArray2D(*operand_array))); + LiteralUtil::CreateR2FromArray2D(*operand_array))); HloInstruction* operand1 = b.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR1({0, 1}))); + HloInstruction::CreateConstant(LiteralUtil::CreateR1({0, 1}))); auto tuple1 = b.AddInstruction(HloInstruction::CreateTuple({operand1, operand2})); @@ -1706,8 +1711,8 @@ TEST_P(HloEvaluatorTest, SetAndGetNestedTuples) { std::unique_ptr result = Evaluate(); auto result_inner_literal = - Literal::CreateR2FromArray2D(*operand_array); - auto expected = Literal::MakeTuple({ + LiteralUtil::CreateR2FromArray2D(*operand_array); + auto expected = LiteralUtil::MakeTuple({ result_inner_literal.get(), result_inner_literal.get(), }); @@ -1735,7 +1740,7 @@ TEST_P(HloEvaluatorTest, Reverse) { {{23.0f}, {24.0f}}}, }); // clang-format on - auto operand_literal = Literal::CreateR4FromArray4D(input); + auto operand_literal = LiteralUtil::CreateR4FromArray4D(input); HloInstruction* operand = b.AddInstruction( HloInstruction::CreateConstant(std::move(operand_literal))); @@ -1746,7 +1751,7 @@ TEST_P(HloEvaluatorTest, Reverse) { std::unique_ptr result = Evaluate(); // clang-format off - auto expected = Literal::CreateR4FromArray4D({ + auto expected = LiteralUtil::CreateR4FromArray4D({ {{{23.0f}, {24.0f}}, {{21.0f}, {22.0f}}, {{19.0f}, {20.0f}}}, @@ -1782,11 +1787,11 @@ TEST_P(HloEvaluatorTest, EvaluateWithSubstitutions) { // Evaluate add with param0 = {1, 2, 3, 4}, square = {10, 20, 30, 40}. HloEvaluator evaluator; auto result = evaluator.EvaluateWithSubstitutions( - add, {{param0, Literal::CreateR1({1, 2, 3, 4}).get()}, - {square, Literal::CreateR1({10, 20, 30, 40}).get()}}); + add, {{param0, LiteralUtil::CreateR1({1, 2, 3, 4}).get()}, + {square, LiteralUtil::CreateR1({10, 20, 30, 40}).get()}}); TF_ASSERT_OK(result.status()); EXPECT_TRUE(LiteralTestUtil::Equal( - *Literal::CreateR1({11, 22, 33, 44}), *result.ValueOrDie())); + *LiteralUtil::CreateR1({11, 22, 33, 44}), *result.ValueOrDie())); } // Check that EvaluateWithSubstitutions works if one of the operands to the op @@ -1799,18 +1804,18 @@ TEST_P(HloEvaluatorTest, EvaluateWithSubstitutionsWithConstantOperand) { b.AddInstruction(HloInstruction::CreateParameter(0, shape, "param0")); HloInstruction* square = b.AddInstruction(HloInstruction::CreateBinary( shape, HloOpcode::kMultiply, param0, param0)); - HloInstruction* constant = b.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR1({1, 2, 3, 4}))); + HloInstruction* constant = b.AddInstruction(HloInstruction::CreateConstant( + LiteralUtil::CreateR1({1, 2, 3, 4}))); HloInstruction* add = b.AddInstruction( HloInstruction::CreateBinary(shape, HloOpcode::kAdd, constant, square)); // Evaluate add with square = {10, 20, 30, 40}. HloEvaluator evaluator; auto result = evaluator.EvaluateWithSubstitutions( - add, {{square, Literal::CreateR1({10, 20, 30, 40}).get()}}); + add, {{square, LiteralUtil::CreateR1({10, 20, 30, 40}).get()}}); TF_ASSERT_OK(result.status()); EXPECT_TRUE(LiteralTestUtil::Equal( - *Literal::CreateR1({11, 22, 33, 44}), *result.ValueOrDie())); + *LiteralUtil::CreateR1({11, 22, 33, 44}), *result.ValueOrDie())); } TEST_P(HloEvaluatorTest, EvaluateGather_TensorFlowGatherV1) { @@ -1830,11 +1835,12 @@ ENTRY main { )"; ParseAndVerifyModule(hlo_text); std::unique_ptr operand = - Literal::CreateR2({{1, 2, 3}, {4, 5, 6}, {7, 8, 9}}); - std::unique_ptr gather_indices = Literal::CreateR1({0, 2}); - EXPECT_TRUE( - LiteralTestUtil::Equal(*Literal::CreateR2({{1, 2, 3}, {7, 8, 9}}), - *Evaluate({operand.get(), gather_indices.get()}))); + LiteralUtil::CreateR2({{1, 2, 3}, {4, 5, 6}, {7, 8, 9}}); + std::unique_ptr gather_indices = + LiteralUtil::CreateR1({0, 2}); + EXPECT_TRUE(LiteralTestUtil::Equal( + *LiteralUtil::CreateR2({{1, 2, 3}, {7, 8, 9}}), + *Evaluate({operand.get(), gather_indices.get()}))); } TEST_P(HloEvaluatorTest, EvaluateGather_TensorFlowGatherV2) { @@ -1854,10 +1860,11 @@ ENTRY main { )"; ParseAndVerifyModule(hlo_text); std::unique_ptr operand = - Literal::CreateR2({{1, 2, 3}, {4, 5, 6}, {7, 8, 9}}); - std::unique_ptr gather_indices = Literal::CreateR1({0, 2}); + LiteralUtil::CreateR2({{1, 2, 3}, {4, 5, 6}, {7, 8, 9}}); + std::unique_ptr gather_indices = + LiteralUtil::CreateR1({0, 2}); EXPECT_TRUE(LiteralTestUtil::Equal( - *Literal::CreateR2({{1, 3}, {4, 6}, {7, 9}}), + *LiteralUtil::CreateR2({{1, 3}, {4, 6}, {7, 9}}), *Evaluate({operand.get(), gather_indices.get()}))); } @@ -1878,11 +1885,11 @@ ENTRY main { )"; ParseAndVerifyModule(hlo_text); std::unique_ptr operand = - Literal::CreateR2({{1, 2, 3}, {4, 5, 6}, {7, 8, 9}}); + LiteralUtil::CreateR2({{1, 2, 3}, {4, 5, 6}, {7, 8, 9}}); std::unique_ptr gather_indices = - Literal::CreateR2({{0, 2}, {2, 1}}); + LiteralUtil::CreateR2({{0, 2}, {2, 1}}); EXPECT_TRUE(LiteralTestUtil::Equal( - *Literal::CreateR3( + *LiteralUtil::CreateR3( {{{1, 3}, {4, 6}, {7, 9}}, {{3, 2}, {6, 5}, {9, 8}}}), *Evaluate({operand.get(), gather_indices.get()}))); } @@ -1904,13 +1911,13 @@ ENTRY main { )"; ParseAndVerifyModule(hlo_text); std::unique_ptr operand = - Literal::CreateR3({{{-1, 1}, {-2, 2}, {-3, 3}}, // - {{-4, 4}, {-5, 5}, {-6, 6}}, // - {{-7, 7}, {-8, 8}, {-9, 9}}}); + LiteralUtil::CreateR3({{{-1, 1}, {-2, 2}, {-3, 3}}, // + {{-4, 4}, {-5, 5}, {-6, 6}}, // + {{-7, 7}, {-8, 8}, {-9, 9}}}); std::unique_ptr gather_indices = - Literal::CreateR2({{0, 0}, {1, 0}}); + LiteralUtil::CreateR2({{0, 0}, {1, 0}}); EXPECT_TRUE( - LiteralTestUtil::Equal(*Literal::CreateR2({{-1, 1}, {-4, 4}}), + LiteralTestUtil::Equal(*LiteralUtil::CreateR2({{-1, 1}, {-4, 4}}), *Evaluate({operand.get(), gather_indices.get()}))); } @@ -1932,13 +1939,13 @@ ENTRY main { )"; ParseAndVerifyModule(hlo_text); std::unique_ptr operand = - Literal::CreateR3({{{-1, 1}, {-2, 2}, {-3, 3}}, // - {{-4, 4}, {-5, 5}, {-6, 6}}, // - {{-7, 7}, {-8, 8}, {-9, 9}}}); + LiteralUtil::CreateR3({{{-1, 1}, {-2, 2}, {-3, 3}}, // + {{-4, 4}, {-5, 5}, {-6, 6}}, // + {{-7, 7}, {-8, 8}, {-9, 9}}}); std::unique_ptr gather_indices = - Literal::CreateR2({{0, 0}, {1, 0}}); + LiteralUtil::CreateR2({{0, 0}, {1, 0}}); EXPECT_TRUE( - LiteralTestUtil::Equal(*Literal::CreateR2({{-2, 2}, {-1, 1}}), + LiteralTestUtil::Equal(*LiteralUtil::CreateR2({{-2, 2}, {-1, 1}}), *Evaluate({operand.get(), gather_indices.get()}))); } @@ -1959,10 +1966,11 @@ ENTRY main { )"; ParseAndVerifyModule(hlo_text); std::unique_ptr operand = - Literal::CreateR2({{1, 2, 3}, {4, 5, 6}, {7, 8, 9}}); - std::unique_ptr gather_indices = Literal::CreateR1({1, 1}); + LiteralUtil::CreateR2({{1, 2, 3}, {4, 5, 6}, {7, 8, 9}}); + std::unique_ptr gather_indices = + LiteralUtil::CreateR1({1, 1}); EXPECT_TRUE( - LiteralTestUtil::Equal(*Literal::CreateR2({{5}}), + LiteralTestUtil::Equal(*LiteralUtil::CreateR2({{5}}), *Evaluate({operand.get(), gather_indices.get()}))); } @@ -1983,11 +1991,11 @@ ENTRY main { )"; ParseAndVerifyModule(hlo_text); std::unique_ptr operand = - Literal::CreateR2({{1, 2, 3}, {4, 5, 6}, {7, 8, 9}}); + LiteralUtil::CreateR2({{1, 2, 3}, {4, 5, 6}, {7, 8, 9}}); std::unique_ptr gather_indices = - Literal::CreateR2({{2, 1}, {1, 1}}); + LiteralUtil::CreateR2({{2, 1}, {1, 1}}); EXPECT_TRUE( - LiteralTestUtil::Equal(*Literal::CreateR3({{{8}}, {{5}}}), + LiteralTestUtil::Equal(*LiteralUtil::CreateR3({{{8}}, {{5}}}), *Evaluate({operand.get(), gather_indices.get()}))); } @@ -2007,10 +2015,11 @@ ENTRY main { } )"; ParseAndVerifyModule(hlo_text); - std::unique_ptr operand = Literal::CreateR2({{}, {}, {}}); - std::unique_ptr gather_indices = Literal::CreateR1({0, 2}); + std::unique_ptr operand = LiteralUtil::CreateR2({{}, {}, {}}); + std::unique_ptr gather_indices = + LiteralUtil::CreateR1({0, 2}); EXPECT_TRUE( - LiteralTestUtil::Equal(*Literal::CreateR2({{}, {}}), + LiteralTestUtil::Equal(*LiteralUtil::CreateR2({{}, {}}), *Evaluate({operand.get(), gather_indices.get()}))); } @@ -2031,11 +2040,11 @@ ENTRY main { )"; ParseAndVerifyModule(hlo_text); - std::unique_ptr operand = Literal::CreateR1({0, 1, 2}); + std::unique_ptr operand = LiteralUtil::CreateR1({0, 1, 2}); std::unique_ptr gather_indices = - Literal::CreateR3({{{0}, {1}}, {{2}, {1}}}); + LiteralUtil::CreateR3({{{0}, {1}}, {{2}, {1}}}); EXPECT_TRUE( - LiteralTestUtil::Equal(*Literal::CreateR2({{0, 1}, {2, 1}}), + LiteralTestUtil::Equal(*LiteralUtil::CreateR2({{0, 1}, {2, 1}}), *Evaluate({operand.get(), gather_indices.get()}))); } @@ -2043,14 +2052,14 @@ ENTRY main { // element-wise comparison with 2 bfloat16 operands. TEST_P(HloEvaluatorTest, DoesCompareBF16) { // lhs >= rhs - auto lhs = Literal::CreateR2( + auto lhs = LiteralUtil::CreateR2( {{bfloat16(0.25), bfloat16(0.35), bfloat16(0.125)}, {bfloat16(-0.25), bfloat16(-0.35), bfloat16(-0.125)}}); - auto rhs = Literal::CreateR2( + auto rhs = LiteralUtil::CreateR2( {{bfloat16(0.5), bfloat16(0.125), bfloat16(0.125)}, {bfloat16(0.25), bfloat16(-0.375), bfloat16(-0.127)}}); auto expected = - Literal::CreateR2({{false, true, true}, {false, true, true}}); + LiteralUtil::CreateR2({{false, true, true}, {false, true, true}}); TestBinaryOp(HloOpcode::kGe, std::move(expected), std::move(lhs), std::move(rhs)); } diff --git a/tensorflow/compiler/xla/service/hlo_evaluator_typed_visitor.h b/tensorflow/compiler/xla/service/hlo_evaluator_typed_visitor.h index 8b08756c641fa8de6c7739fb4dd94ceceeb53311..2ae5f8bf36d1d6be769e8f05d31b2351ca9b8297 100644 --- a/tensorflow/compiler/xla/service/hlo_evaluator_typed_visitor.h +++ b/tensorflow/compiler/xla/service/hlo_evaluator_typed_visitor.h @@ -16,6 +16,7 @@ limitations under the License. #ifndef TENSORFLOW_COMPILER_XLA_SERVICE_HLO_EVALUATOR_TYPED_VISITOR_H_ #define TENSORFLOW_COMPILER_XLA_SERVICE_HLO_EVALUATOR_TYPED_VISITOR_H_ +#include "tensorflow/compiler/xla/literal_util.h" #include "tensorflow/compiler/xla/service/hlo_evaluator.h" #include "tensorflow/compiler/xla/service/shape_inference.h" #include "tensorflow/core/lib/core/casts.h" @@ -34,6 +35,37 @@ using is_complex_t = std::is_same; template using is_complex64_t = std::is_same; +// It's UB to use std::sort with std::less, because of NaNs. Define +// "safe" less functions which are actually strict weak orders. +template < + typename NativeT, + typename std::enable_if::value>::type* = nullptr> +bool SafeLess(const NativeT& a, const NativeT& b) { + return a < b; +} + +template ::value || + std::is_same::value>::type* = nullptr> +bool SafeLess(const NativeT& a, const NativeT& b) { + if (std::isnan(b)) { + return !std::isnan(a); + } else { + return a < b; + } +} + +template ::value>::type* = nullptr> +bool SafeLess(const NativeT& a, const NativeT& b) { + if (Eigen::half_impl::isnan(b)) { + return !Eigen::half_impl::isnan(a); + } else { + return a < b; + } +} + // Templated DfsHloVisitor for use by HloEvaluator. // // Typically ReturnT here indicates the resulting literal type of each evaluated @@ -1025,83 +1057,47 @@ class HloEvaluatorTypedVisitor : public DfsHloVisitorWithDefault { CHECK_EQ(dnums.lhs_batch_dimensions_size(), dnums.rhs_batch_dimensions_size()); - std::vector lhs_non_contracting_dims; + DimensionVector lhs_index(lhs_rank); + DimensionVector rhs_index(rhs_rank); + + // result_index_locations[i] contains one or two pointers to the locations + // in lhs_index or rhs_index where the i'th result index should go. + tensorflow::gtl::InlinedVector, kInlineRank> + result_index_locations; + result_index_locations.reserve(lhs_rank + rhs_rank - 2); + + // The first components in the output shape are the LHS and RHS batch + // dimensions: + for (int64 i = 0; i < dnums.lhs_batch_dimensions_size(); i++) { + result_index_locations.push_back( + {&lhs_index[dnums.lhs_batch_dimensions(i)], + &rhs_index[dnums.rhs_batch_dimensions(i)]}); + } + + // Then we have the LHS and RHS non-contracting dimensions, if any: for (int64 i = 0; i < lhs_rank; i++) { - if (i != lhs_contracting_dimension) { - lhs_non_contracting_dims.push_back(i); + if (i != lhs_contracting_dimension && + !ArrayContains(AsInt64Slice(dnums.lhs_batch_dimensions()), i)) { + result_index_locations.push_back({&lhs_index[i], nullptr}); } } - - std::vector rhs_non_batch_non_contracting_dims; - tensorflow::gtl::FlatSet batch_dims_set( - dnums.rhs_batch_dimensions().begin(), - dnums.rhs_batch_dimensions().end()); for (int64 i = 0; i < rhs_rank; i++) { - if (i != rhs_contracting_dimension && batch_dims_set.count(i) == 0) { - rhs_non_batch_non_contracting_dims.push_back(i); + if (i != rhs_contracting_dimension && + !ArrayContains(AsInt64Slice(dnums.rhs_batch_dimensions()), i)) { + result_index_locations.push_back({&rhs_index[i], nullptr}); } } - const int64 batch_dim_size = dnums.lhs_batch_dimensions_size(); - const int64 lhs_non_contracting_size = lhs_non_contracting_dims.size(); - - DimensionVector lhs_index(lhs_rank); - DimensionVector rhs_index(rhs_rank); auto result = MakeUnique(dot->shape()); TF_RETURN_IF_ERROR(result->Populate( [&](tensorflow::gtl::ArraySlice result_index) { ElementwiseT result_val = static_cast(0); - // Find the corresponding non-contracting indices for lhs and rhs. - // - // For `result_index`, its batch dimension, if exists, will be at the - // same dimension as the batch dimension of lhs and rhs. More - // specifically: - // - For lhs, the non-contracting dimensions, including the batch - // dimension have the same index as the `result_index`. - // - For rhs, the batch dimension is set seperately from other - // non-contracting dimensions, since these other non-contracting - // dimensions in rhs follow the non-contracting dimensions of lhs in - // the resulting index. - // - // As an example, for a resulting index: - // result_index [result_batch, result_x, result_y] - // the effecting lhs and rhs indices are: - // lhs [result_batch, lhs_non_contracting_dim, contracting_dim - // rhs [result_batch, contracting_dim, rhs_non_contracting_dim] - // `result_x` is only affected by the lhs_non_contracting_dim and - // likewise `result_y` only depends on rhs_non_contracting_dim. - // - // so we can look up the lhs and rhs indices by: - // - // lhs: - // batch index is the same as `result_batch`. - // non-contracting dimension is the same as - // result_index[lhs_non_contracting_dim] - // rhs: - // batch index: the same as `result_batch`. - // non-contracting dimension index: *not* the same as - // result_index[rhs_non_contractng_dim], since the - // non-contracting dimensions of lhs are included in the - // result_index first. Instead, the non_contracting_dim of rhs must - // be calculated as following: - // lhs_non_contracting_dimensions_size + - // (rhs_non_batch_non_contracting_dim - batch_dim_size) - 1 - // - // Note that (rhs_non_batch_contracting_dim - batch_dim_size) is - // the index offset to the result_index that only depends on - // the non_batch and non-contracting dimensions of rhs. -1 at the - // end translates size to index. - for (auto i : lhs_non_contracting_dims) { - lhs_index[i] = result_index[i]; - } - for (auto i : dnums.rhs_batch_dimensions()) { - rhs_index[i] = result_index[i]; - } - for (auto i : rhs_non_batch_non_contracting_dims) { - const int64 rhs_non_batch_non_contracting_dim = - lhs_non_contracting_size + (i - batch_dim_size) - 1; - rhs_index[i] = result_index[rhs_non_batch_non_contracting_dim]; + for (int64 i = 0; i < result_index.size(); i++) { + *result_index_locations[i].first = result_index[i]; + if (result_index_locations[i].second) { + *result_index_locations[i].second = result_index[i]; + } } // Accumulates resulting product along the contracted dimension. @@ -1321,7 +1317,7 @@ class HloEvaluatorTypedVisitor : public DfsHloVisitorWithDefault { parent_->GetEvaluatedLiteralFor(operand); auto curr_val = arg_literal.Get(multi_index); - auto curr_val_literal = Literal::CreateR0(curr_val); + auto curr_val_literal = LiteralUtil::CreateR0(curr_val); arg_literals.push_back(std::move(curr_val_literal)); } @@ -1402,22 +1398,24 @@ class HloEvaluatorTypedVisitor : public DfsHloVisitorWithDefault { !is_complex_t::value && !std::is_same::value>::type* = nullptr> Status HandleSort(HloInstruction* sort) { - TF_RET_CHECK(ShapeUtil::Rank(sort->shape()) == 1) + auto keys = sort->operand(0); + TF_RET_CHECK(ShapeUtil::Rank(keys->shape()) == 1) << "Sort is only supported for R1 shapes"; + TF_RET_CHECK(sort->operand_count() == 1) + << "Typed visitor does not support key-value sort"; - auto arg = sort->operand(0); - const Literal& arg_literal = parent_->GetEvaluatedLiteralFor(arg); - VLOG(3) << "HandleSort arg_literal: " << arg_literal.ToString(); - const auto& arg_data = arg_literal.data(); + const Literal& keys_literal = parent_->GetEvaluatedLiteralFor(keys); + VLOG(3) << "HandleSort keys_literal: " << keys_literal.ToString(); + const auto& keys_data = keys_literal.data(); - std::vector return_data(arg_data.begin(), arg_data.end()); - std::sort(return_data.begin(), return_data.end(), + std::vector result_data(keys_data.begin(), keys_data.end()); + std::sort(result_data.begin(), result_data.end(), [](const ReturnT& a, const ReturnT& b) { return SafeLess(a, b); }); auto result_literal = MakeUnique(sort->shape()); result_literal->PopulateR1( - tensorflow::gtl::ArraySlice(return_data)); + tensorflow::gtl::ArraySlice(result_data)); VLOG(3) << "HandleSort result_literal: " << result_literal->ToString(); parent_->evaluated_[sort] = std::move(result_literal); return Status::OK(); @@ -1507,8 +1505,9 @@ class HloEvaluatorTypedVisitor : public DfsHloVisitorWithDefault { auto curr_val = arg_literal.Get(input_index); // Evaluate computation with specified literal operands. - auto curr_val_literal = Literal::CreateR0(curr_val); - auto result_val_literal = Literal::CreateR0(result_val); + auto curr_val_literal = LiteralUtil::CreateR0(curr_val); + auto result_val_literal = + LiteralUtil::CreateR0(result_val); std::unique_ptr computed_result = embedded_evaluator @@ -1586,10 +1585,10 @@ class HloEvaluatorTypedVisitor : public DfsHloVisitorWithDefault { // Used in the dual IterateThroughWindow lambdas below. Hoisted to avoid // dynamic memory allocations. - auto curr_val_literal = Literal::CreateR0(ReturnT()); - auto selected_val_literal = Literal::CreateR0(ReturnT()); - auto source_literal_scatter = Literal::CreateR0(ReturnT()); - auto scattered_literal = Literal::CreateR0(ReturnT()); + auto curr_val_literal = LiteralUtil::CreateR0(ReturnT()); + auto selected_val_literal = LiteralUtil::CreateR0(ReturnT()); + auto source_literal_scatter = LiteralUtil::CreateR0(ReturnT()); + auto scattered_literal = LiteralUtil::CreateR0(ReturnT()); do { // For each element in `source`, we place a window in `operand`. For each // window placement, we iterate inside the window twice: @@ -1710,9 +1709,9 @@ class HloEvaluatorTypedVisitor : public DfsHloVisitorWithDefault { // Evaluate computation with specified literal operands. const auto curr_val_literal = - Literal::CreateR0(curr_val); + LiteralUtil::CreateR0(curr_val); const auto result_val_literal = - Literal::CreateR0(result_val); + LiteralUtil::CreateR0(result_val); std::unique_ptr computed_result = embedded_evaluator .Evaluate( @@ -1757,7 +1756,7 @@ class HloEvaluatorTypedVisitor : public DfsHloVisitorWithDefault { return operand_literal.Get(operand_index); }; - auto result = Literal::CreateFromDimensions( + auto result = LiteralUtil::CreateFromDimensions( shape.element_type(), AsInt64Slice(shape.dimensions())); TF_RETURN_IF_ERROR(result->Populate(func)); parent_->evaluated_[slice] = std::move(result); @@ -2175,38 +2174,6 @@ class HloEvaluatorTypedVisitor : public DfsHloVisitorWithDefault { return rhs_unsigned >= lhs_size_unsigned; } - // It's UB to use std::sort with std::less, because of NaNs. Define - // "safe" less functions which are actually strict weak orders. - template ::value>::type* = - nullptr> - static bool SafeLess(const NativeT& a, const NativeT& b) { - return a < b; - } - - template ::value || - std::is_same::value>::type* = nullptr> - static bool SafeLess(const NativeT& a, const NativeT& b) { - if (std::isnan(b)) { - return !std::isnan(a); - } else { - return a < b; - } - } - - template ::value>::type* = nullptr> - static bool SafeLess(const NativeT& a, const NativeT& b) { - if (Eigen::half_impl::isnan(b)) { - return !Eigen::half_impl::isnan(a); - } else { - return a < b; - } - } - HloEvaluator* parent_; }; diff --git a/tensorflow/compiler/xla/service/hlo_graph_dumper.cc b/tensorflow/compiler/xla/service/hlo_graph_dumper.cc index b349f7d46fd6a882d4e7d236a58deb78f5194413..57cf34d7dee94f0303cd6b5591ad6507828c70f3 100644 --- a/tensorflow/compiler/xla/service/hlo_graph_dumper.cc +++ b/tensorflow/compiler/xla/service/hlo_graph_dumper.cc @@ -27,7 +27,7 @@ limitations under the License. #include #include "tensorflow/compiler/xla/layout_util.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/service/hlo_casting_utils.h" #include "tensorflow/compiler/xla/service/hlo_instructions.h" #include "tensorflow/compiler/xla/service/hlo_module.h" @@ -966,6 +966,7 @@ ColorScheme HloDotDumper::GetInstructionColor(const HloInstruction* instr) { case HloOpcode::kRemainder: case HloOpcode::kRng: case HloOpcode::kRoundNearestAfz: + case HloOpcode::kSelect: case HloOpcode::kShiftLeft: case HloOpcode::kShiftRightArithmetic: case HloOpcode::kShiftRightLogical: @@ -984,7 +985,7 @@ ColorScheme HloDotDumper::GetInstructionColor(const HloInstruction* instr) { case HloOpcode::kBitcast: case HloOpcode::kGetTupleElement: case HloOpcode::kTrace: - case HloOpcode::kGenerateToken: + case HloOpcode::kAfterAll: case HloOpcode::kTuple: return kWhite; case HloOpcode::kBroadcast: @@ -1001,7 +1002,7 @@ ColorScheme HloDotDumper::GetInstructionColor(const HloInstruction* instr) { case HloOpcode::kPad: case HloOpcode::kReshape: case HloOpcode::kReverse: - case HloOpcode::kSelect: + case HloOpcode::kTupleSelect: case HloOpcode::kTranspose: // De-emphasize scalar-shaped data movement ops and all data movement ops // inside fusion nodes, both of which are essentially free. diff --git a/tensorflow/compiler/xla/service/hlo_graph_dumper_test.cc b/tensorflow/compiler/xla/service/hlo_graph_dumper_test.cc index 68f41a1cbb4db228f5dcf8b4a6130f05e81262a8..1d7a062c55696de9db4b187efd86bce191279083 100644 --- a/tensorflow/compiler/xla/service/hlo_graph_dumper_test.cc +++ b/tensorflow/compiler/xla/service/hlo_graph_dumper_test.cc @@ -15,6 +15,7 @@ limitations under the License. #include "tensorflow/compiler/xla/service/hlo_graph_dumper.h" +#include "tensorflow/compiler/xla/literal_util.h" #include "tensorflow/compiler/xla/service/hlo_computation.h" #include "tensorflow/compiler/xla/service/hlo_instruction.h" #include "tensorflow/compiler/xla/service/hlo_module.h" @@ -120,7 +121,7 @@ TEST(HloGraphDumperTest, NestedFusion) { TEST(HloGraphDumperTest, Constant) { HloComputation::Builder b("b"); auto instruction = b.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(-42))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(-42))); instruction->SetAndSanitizeName("i_am_a_constant_root_instruction"); HloModuleConfig config; HloModule m(TestName(), config); diff --git a/tensorflow/compiler/xla/service/hlo_instruction.cc b/tensorflow/compiler/xla/service/hlo_instruction.cc index a07dbe6256b90e6aefc84da9fd26c5f39b72b7a5..830ebfb125cadd00e52afd046ebbc057888b88ca 100644 --- a/tensorflow/compiler/xla/service/hlo_instruction.cc +++ b/tensorflow/compiler/xla/service/hlo_instruction.cc @@ -22,7 +22,7 @@ limitations under the License. #include #include "tensorflow/compiler/xla/layout_util.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/protobuf_util.h" #include "tensorflow/compiler/xla/ptr_util.h" #include "tensorflow/compiler/xla/service/dfs_hlo_visitor.h" @@ -112,10 +112,10 @@ StatusOr> HloInstruction::CreateFromProto( break; } case HloOpcode::kSend: - TF_RET_CHECK(proto.operand_ids_size() == 1) - << "Send instruction should have 1 operand but sees " + TF_RET_CHECK(proto.operand_ids_size() == 2) + << "Send instruction should have 2 operand but sees " << proto.operand_ids_size(); - instruction = CreateSend(operands(0), proto.channel_id()); + instruction = CreateSend(operands(0), operands(1), proto.channel_id()); break; case HloOpcode::kSendDone: TF_RET_CHECK(proto.operand_ids_size() == 1) @@ -124,11 +124,11 @@ StatusOr> HloInstruction::CreateFromProto( instruction = CreateSendDone(operands(0)); break; case HloOpcode::kRecv: - TF_RET_CHECK(proto.operand_ids_size() == 0) - << "Recv instruction should have 0 operand but sees " + TF_RET_CHECK(proto.operand_ids_size() == 1) + << "Recv instruction should have 1 operand but sees " << proto.operand_ids_size(); - instruction = - CreateRecv(proto.shape().tuple_shapes(0), proto.channel_id()); + instruction = CreateRecv(proto.shape().tuple_shapes(0), operands(0), + proto.channel_id()); break; case HloOpcode::kRecvDone: TF_RET_CHECK(proto.operand_ids_size() == 1) @@ -163,6 +163,20 @@ StatusOr> HloInstruction::CreateFromProto( proto.dimensions().end()), computations(0)); break; + case HloOpcode::kSort: { + TF_RET_CHECK(proto.operand_ids_size() == 1 || + proto.operand_ids_size() == 2) + << "Sort instruction should have 1 or 2 operands but has " + << proto.operand_ids_size(); + TF_RET_CHECK(proto.dimensions().size() == 1) + << "Sort instruction should have 1 dimension"; + HloInstruction* keys = operands(0); + HloInstruction* values = + proto.operand_ids_size() == 2 ? operands(1) : nullptr; + instruction = + CreateSort(proto.shape(), proto.dimensions(0), keys, values); + break; + } case HloOpcode::kTranspose: TF_RET_CHECK(proto.operand_ids_size() == 1) << "Transpose instruction should have 1 operand but sees " @@ -263,12 +277,30 @@ StatusOr> HloInstruction::CreateFromProto( CreateReducePrecision(proto.shape(), operands(0), proto.exponent_bits(), proto.mantissa_bits()); break; - case HloOpcode::kInfeed: - instruction = CreateInfeed(proto.shape(), proto.infeed_config()); - break; + case HloOpcode::kInfeed: { + const Shape& data_shape = + ShapeUtil::GetTupleElementShape(proto.shape(), 0); + if (proto.operand_ids_size() == 0) { + // TODO(b/80000000): Remove this when all uses of infeed are + // converted to take tokens. + instruction = CreateInfeed(data_shape, proto.infeed_config()); + } else { + CHECK_EQ(proto.operand_ids_size(), 1); + instruction = + CreateInfeed(data_shape, operands(0), proto.infeed_config()); + } + } break; case HloOpcode::kOutfeed: - instruction = CreateOutfeed(proto.outfeed_shape(), operands(0), - proto.outfeed_config()); + if (proto.operand_ids_size() == 1) { + // TODO(b/80000000): Remove this when all uses of outfeed are + // converted to take tokens. + instruction = CreateOutfeed(proto.outfeed_shape(), operands(0), + proto.outfeed_config()); + } else { + CHECK_EQ(proto.operand_ids_size(), 2); + instruction = CreateOutfeed(proto.outfeed_shape(), operands(0), + operands(1), proto.outfeed_config()); + } break; case HloOpcode::kCrossReplicaSum: { TF_RET_CHECK(proto.called_computation_ids_size() == 1) @@ -471,7 +503,6 @@ HloInstruction::CreateGetTupleElement(const Shape& shape, case HloOpcode::kReal: case HloOpcode::kSign: case HloOpcode::kSin: - case HloOpcode::kSort: case HloOpcode::kTanh: break; default: @@ -524,8 +555,9 @@ HloInstruction::CreateGetTupleElement(const Shape& shape, // Only certain opcodes are supported with CreateTernary: opcodes of ternary // instructions with no auxiliary fields. switch (opcode) { - case (HloOpcode::kClamp): - case (HloOpcode::kSelect): + case HloOpcode::kClamp: + case HloOpcode::kSelect: + case HloOpcode::kTupleSelect: break; default: LOG(FATAL) << "Invalid ternary instruction opcode " @@ -543,10 +575,8 @@ HloInstruction::CreateGetTupleElement(const Shape& shape, /* static */ std::unique_ptr HloInstruction::CreateMap( const Shape& shape, tensorflow::gtl::ArraySlice operands, - HloComputation* map_computation, - tensorflow::gtl::ArraySlice static_operands) { - return MakeUnique(shape, operands, map_computation, - static_operands); + HloComputation* map_computation) { + return MakeUnique(shape, operands, map_computation); } /* static */ std::unique_ptr HloInstruction::CreateConvolve( @@ -610,19 +640,33 @@ HloInstruction::CreateCrossReplicaSum( } /* static */ std::unique_ptr HloInstruction::CreateInfeed( - const Shape& shape, const string& config) { - return MakeUnique(shape, config); + const Shape& infeed_shape, HloInstruction* token_operand, + const string& config) { + return MakeUnique(infeed_shape, token_operand, config); +} + +/* static */ std::unique_ptr HloInstruction::CreateInfeed( + const Shape& infeed_shape, const string& config) { + return MakeUnique(infeed_shape, config); } /* static */ std::unique_ptr HloInstruction::CreateOutfeed( - const Shape& shape, HloInstruction* operand, + const Shape& outfeed_shape, HloInstruction* operand, + HloInstruction* token_operand, tensorflow::StringPiece outfeed_config) { + return MakeUnique(outfeed_shape, operand, + token_operand, outfeed_config); +} + +/* static */ std::unique_ptr HloInstruction::CreateOutfeed( + const Shape& outfeed_shape, HloInstruction* operand, tensorflow::StringPiece outfeed_config) { - return MakeUnique(shape, operand, outfeed_config); + return MakeUnique(outfeed_shape, operand, + outfeed_config); } /* static */ std::unique_ptr HloInstruction::CreateSend( - HloInstruction* operand, int64 channel_id) { - return MakeUnique(operand, channel_id); + HloInstruction* operand, HloInstruction* token, int64 channel_id) { + return MakeUnique(operand, token, channel_id); } /* static */ std::unique_ptr HloInstruction::CreateSendDone( @@ -634,8 +678,8 @@ HloInstruction::CreateCrossReplicaSum( } /* static */ std::unique_ptr HloInstruction::CreateRecv( - const Shape& shape, int64 channel_id) { - return MakeUnique(shape, channel_id); + const Shape& shape, HloInstruction* token, int64 channel_id) { + return MakeUnique(shape, token, channel_id); } /* static */ std::unique_ptr HloInstruction::CreateRecvDone( @@ -652,17 +696,22 @@ HloInstruction::CreateCrossReplicaSum( return MakeUnique(shape, operand, dimensions); } -/* static */ std::unique_ptr -HloInstruction::CreateGenerateToken( +/* static */ std::unique_ptr HloInstruction::CreateAfterAll( tensorflow::gtl::ArraySlice operands) { - auto instruction = WrapUnique(new HloInstruction( - HloOpcode::kGenerateToken, ShapeUtil::MakeTokenShape())); + CHECK(!operands.empty()); + auto instruction = WrapUnique( + new HloInstruction(HloOpcode::kAfterAll, ShapeUtil::MakeTokenShape())); for (auto operand : operands) { instruction->AppendOperand(operand); } return instruction; } +/* static */ std::unique_ptr HloInstruction::CreateToken() { + return WrapUnique( + new HloInstruction(HloOpcode::kAfterAll, ShapeUtil::MakeTokenShape())); +} + /* static */ std::unique_ptr HloInstruction::CreateWhile( const Shape& shape, HloComputation* condition, HloComputation* body, HloInstruction* init) { @@ -879,6 +928,12 @@ HloInstruction::CreateBroadcastSequence( return MakeUnique(shape, operand, dimensions); } +/* static */ std::unique_ptr HloInstruction::CreateSort( + const Shape& shape, int64 dimension, HloInstruction* keys, + HloInstruction* values) { + return MakeUnique(shape, dimension, keys, values); +} + /* static */ std::unique_ptr HloInstruction::CreateFusion( const Shape& shape, FusionKind fusion_kind, HloInstruction* fused_root) { return MakeUnique(shape, fusion_kind, fused_root); @@ -1071,6 +1126,7 @@ std::unique_ptr HloInstruction::CloneWithNewOperands( case HloOpcode::kHostCompute: case HloOpcode::kPad: case HloOpcode::kDynamicSlice: + case HloOpcode::kSort: clone = CloneWithNewOperandsImpl(shape, new_operands, context); break; // Unary ops. @@ -1093,7 +1149,6 @@ std::unique_ptr HloInstruction::CloneWithNewOperands( case HloOpcode::kReal: case HloOpcode::kSign: case HloOpcode::kSin: - case HloOpcode::kSort: case HloOpcode::kTanh: CHECK_EQ(new_operands.size(), 1); clone = CreateUnary(shape, opcode_, new_operands[0]); @@ -1127,6 +1182,7 @@ std::unique_ptr HloInstruction::CloneWithNewOperands( // Ternary ops. case HloOpcode::kClamp: case HloOpcode::kSelect: + case HloOpcode::kTupleSelect: CHECK_EQ(new_operands.size(), 3); clone = CreateTernary(shape, opcode_, new_operands[0], new_operands[1], new_operands[2]); @@ -1183,8 +1239,12 @@ std::unique_ptr HloInstruction::CloneWithNewOperands( CreateDomain(shape, new_operands[0], operand_side_metadata_->Clone(), user_side_metadata_->Clone()); break; - case HloOpcode::kGenerateToken: - clone = CreateGenerateToken(new_operands); + case HloOpcode::kAfterAll: + if (new_operands.empty()) { + clone = CreateToken(); + } else { + clone = CreateAfterAll(new_operands); + } break; } SetupDerivedInstruction(clone.get()); @@ -1369,6 +1429,30 @@ void HloInstruction::AppendOperand(HloInstruction* operand) { operand->AddUser(this); } +void HloInstruction::RemoveOperandsAtAscendingIndices( + tensorflow::gtl::ArraySlice ascending_indices) { + if (ascending_indices.empty()) { + return; + } + int next_index = 0; + int removed_count = 0; + for (int to_remove : ascending_indices) { + while (next_index < to_remove) { + operands_[next_index - removed_count] = operands_[next_index]; + ++next_index; + } + CHECK_LT(to_remove, operands_.size()); + ++removed_count; + ++next_index; + } + while (next_index < operands_.size()) { + operands_[next_index - removed_count] = operands_[next_index]; + ++next_index; + } + CHECK_EQ(removed_count, ascending_indices.size()); + operands_.resize(operands_.size() - removed_count); +} + void HloInstruction::AddUser(HloInstruction* user) { if (!ContainsKey(user_set_, user)) { user_set_.insert(user); @@ -1442,12 +1526,12 @@ bool HloInstruction::IdenticalSlowPath( case HloOpcode::kSubtract: case HloOpcode::kTanh: case HloOpcode::kTuple: + case HloOpcode::kTupleSelect: return true; // These opcodes have complex or special behavior so just return false. - case HloOpcode::kDomain: case HloOpcode::kWhile: - case HloOpcode::kGenerateToken: + case HloOpcode::kAfterAll: return false; // Check dot dimension numbers. @@ -1467,9 +1551,9 @@ bool HloInstruction::IdenticalSlowPath( return eq_computations(true_computation(), other.true_computation()) && eq_computations(false_computation(), other.false_computation()); - // These opcodes are not yet supported. - case HloOpcode::kSort: - return false; + case HloOpcode::kDomain: + return operand_side_metadata().Matches(other.operand_side_metadata()) && + user_side_metadata().Matches(other.user_side_metadata()); // Ops migrated to subclasses should never come to this line. // TODO(b/80131774): Remove this switch when migration is complete. @@ -1484,6 +1568,7 @@ bool HloInstruction::IdenticalSlowPath( case HloOpcode::kReverse: case HloOpcode::kConcatenate: case HloOpcode::kReduce: + case HloOpcode::kSort: case HloOpcode::kTranspose: case HloOpcode::kBroadcast: case HloOpcode::kMap: @@ -1539,6 +1624,10 @@ Status HloInstruction::ReplaceUseWith(HloInstruction* user, std::replace(user->operands_.begin(), user->operands_.end(), this, new_producer); new_producer->AddUser(user); + if (user->opcode() == HloOpcode::kFusion) { + TF_RETURN_IF_ERROR( + Cast(user)->DeduplicateFusionOperands()); + } return Status::OK(); } @@ -1547,10 +1636,14 @@ Status HloInstruction::ReplaceOperandWith(int64 operand_num, TF_RET_CHECK(operand_num >= 0); TF_RET_CHECK(operand_num < operand_count()); HloInstruction* old_operand = mutable_operand(operand_num); + if (old_operand == new_operand) { + return Status::OK(); + } + TF_RET_CHECK(ShapeUtil::CompatibleIgnoringFpPrecision(old_operand->shape(), new_operand->shape())) - << old_operand->shape().ShortDebugString() << " is not compatible with " - << new_operand->shape().ShortDebugString(); + << old_operand->shape() << " is not compatible with " + << new_operand->shape(); operands_[operand_num] = new_operand; VLOG(3) << "Replacing operand " << operand_num << " of " << name() << " with " @@ -1577,6 +1670,10 @@ Status HloInstruction::ReplaceAllUsesWith(HloInstruction* new_producer) { std::replace(user->operands_.begin(), user->operands_.end(), this, new_producer); new_producer->AddUser(user); + if (user->opcode() == HloOpcode::kFusion) { + TF_RETURN_IF_ERROR( + Cast(user)->DeduplicateFusionOperands()); + } } } users_.clear(); @@ -1755,7 +1852,6 @@ bool HloInstruction::IsElementwiseImpl( // Ternary elementwise operations. case HloOpcode::kSelect: - return !ShapeUtil::IsTuple(shape_); case HloOpcode::kClamp: return true; @@ -1950,8 +2046,8 @@ std::vector HloInstruction::ExtraAttributesToString( } if (operand_side_metadata_ != nullptr && user_side_metadata_ != nullptr) { extra.push_back(StrCat("domain={kind=\"", operand_side_metadata_->Kind(), - "\", entry=", operand_side_metadata_->ToString(), - ", exit=", user_side_metadata_->ToString(), "}")); + "\", entry=", user_side_metadata_->ToString(), + ", exit=", operand_side_metadata_->ToString(), "}")); } return extra; @@ -2126,6 +2222,8 @@ Status HloInstruction::Visit(DfsHloVisitorBase* visitor) { return visitor->HandleRemainder(this); case HloOpcode::kSelect: return visitor->HandleSelect(this); + case HloOpcode::kTupleSelect: + return visitor->HandleTupleSelect(this); case HloOpcode::kConvolution: return visitor->HandleConvolution(this); case HloOpcode::kFft: @@ -2226,8 +2324,8 @@ Status HloInstruction::Visit(DfsHloVisitorBase* visitor) { return visitor->HandleGather(this); case HloOpcode::kDomain: return visitor->HandleDomain(this); - case HloOpcode::kGenerateToken: - return visitor->HandleGenerateToken(this); + case HloOpcode::kAfterAll: + return visitor->HandleAfterAll(this); // These opcodes are not handled here. case HloOpcode::kTrace: diff --git a/tensorflow/compiler/xla/service/hlo_instruction.h b/tensorflow/compiler/xla/service/hlo_instruction.h index 8f59e67123cadf965c8650f4a82622f5443ecac9..b392d65636accce75e654801f2ebeb34cfffb42b 100644 --- a/tensorflow/compiler/xla/service/hlo_instruction.h +++ b/tensorflow/compiler/xla/service/hlo_instruction.h @@ -33,7 +33,7 @@ limitations under the License. #include #include "tensorflow/compiler/xla/iterator_util.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/map_util.h" #include "tensorflow/compiler/xla/service/dfs_hlo_visitor.h" #include "tensorflow/compiler/xla/service/hlo.pb.h" @@ -389,11 +389,10 @@ class HloInstruction { // Creates a map instruction, where the computation (given by the handle) is // applied element-wise to every element in operands (across the operands, - // at a given index) with the same `static_operands`. + // at a given index) static std::unique_ptr CreateMap( const Shape& shape, tensorflow::gtl::ArraySlice operands, - HloComputation* map_computation, - tensorflow::gtl::ArraySlice static_operands = {}); + HloComputation* map_computation); // Creates a convolution op, where rhs is the convolutional filter // and window describes how the filter is applied to lhs. @@ -459,19 +458,36 @@ class HloInstruction { const Shape& shape, HloInstruction* operand); // Creates an infeed instruction, which reads data of the given shape from the - // Infeed interface of the device. - static std::unique_ptr CreateInfeed(const Shape& shape, + // Infeed interface of the device. infeed_shape is the shape of the data + // received from the infeed *not* the shape of the infeed instruction which + // is a tuple containing the infeed_shape and the TOKEN. + static std::unique_ptr CreateInfeed( + const Shape& infeed_shape, HloInstruction* token_operand, + const string& config); + // Overload which does not require a token. + // TODO(b/80000000): Remove this overload when all uses of infeed are + // converted to take tokens. + static std::unique_ptr CreateInfeed(const Shape& infeed_shape, const string& config); - // Creates an outfeed instruction, which outputs data. + // Creates an outfeed instruction, which outputs data. outfeed_shape is the + // shape of the data being outfed *not* the shape of the outfeed instruction + // which is a TOKEN. static std::unique_ptr CreateOutfeed( - const Shape& shape, HloInstruction* operand, + const Shape& outfeed_shape, HloInstruction* operand, + HloInstruction* token_operand, tensorflow::StringPiece outfeed_config); + // Overload which does not require a token. + // TODO(b/80000000): Remove this overload when all uses of outfeed are + // converted to take tokens. + static std::unique_ptr CreateOutfeed( + const Shape& outfeed_shape, HloInstruction* operand, tensorflow::StringPiece outfeed_config); // Creates an asynchronous send instruction with the given channel id, which // initiates sending the operand data to a unique receive instruction in // another computation that has the same channel id. static std::unique_ptr CreateSend(HloInstruction* operand, + HloInstruction* token, int64 channel_id); // Blocks until data transfer for the Send instruction (operand) is complete. @@ -483,6 +499,7 @@ class HloInstruction { // which allocates resources to receive data of the given shape from a unique // send instruction in another computation that has the same channel id. static std::unique_ptr CreateRecv(const Shape& shape, + HloInstruction* token, int64 channel_id); // Blocks until data transfer for the Recv instruction (operand) is complete @@ -596,6 +613,11 @@ class HloInstruction { const Shape& shape, HloInstruction* operand, tensorflow::gtl::ArraySlice dimensions); + // Creates a sort op, with a keys operand, and an optional values operand. + static std::unique_ptr CreateSort( + const Shape& shape, int64 dimension, HloInstruction* keys, + HloInstruction* values = nullptr); + // Creates a while instruction, given a condition computation, a body // computation, and the initial value for the input of the computations. For // example, shape: S32, condition: i -> i < 1000, body: i -> i * 2, init: 1 @@ -665,11 +687,19 @@ class HloInstruction { const Shape& shape, HloInstruction* operand, tensorflow::gtl::ArraySlice dimensions); - // Creates a token instruction used for joining or creating token types which - // thread through side-effecting operations. - static std::unique_ptr CreateGenerateToken( + // Creates a Afterall instruction used for joining or creating new values of + // token type which thread through side-effecting operations. Operands must + // all be tokens, and there must be at least one operand. + static std::unique_ptr CreateAfterAll( tensorflow::gtl::ArraySlice operands); + // Creates an AfterAll instruction which creates a token type out of thin air + // (no operands). This is a separate method from CreateAfterAll to facility + // the removal of operand-less AfterAll instructions. + // TODO(b/110532604): Remove this capability of creating a token from nothing + // when we plumb a primordial token from the entry computation. + static std::unique_ptr CreateToken(); + // Creates an instance of GatherDimensionNumbers. static GatherDimensionNumbers MakeGatherDimNumbers( tensorflow::gtl::ArraySlice output_window_dims, @@ -811,9 +841,15 @@ class HloInstruction { // Replaces the use of this instruction in "user" with "new_producer". Note // that there might be multiple uses of this instruction in "user"; all will // be replaced. + // + // If user is a fusion instruction, this function will remove any duplicated + // operands of it which could be created due to this replacement. Status ReplaceUseWith(HloInstruction* user, HloInstruction* new_producer); // Replaces the specified operand with new_operand. + // + // This function does NOT remove duplicated operands even if this instruction + // is a fusion, so that the existing operand numbers do not change. Status ReplaceOperandWith(int64 operand_no, HloInstruction* new_operand); // Replaces all uses of this instruction with the new producer. If @@ -822,6 +858,9 @@ class HloInstruction { // // If this instruction is the root of its computation, sets the computation's // root to new_producer. + // + // If a user is a fusion instruction, this function will remove any duplicated + // operands of it which could be created due to this replacement. Status ReplaceAllUsesWith(HloInstruction* new_producer); // Performs a postorder DFS visit using this node as the root. If @@ -1440,6 +1479,10 @@ class HloInstruction { operands_.erase(operands_.begin() + index); } + // Removes a list of operands with the given indices in ascending order. + void RemoveOperandsAtAscendingIndices( + tensorflow::gtl::ArraySlice ascending_indices); + void AppendComputation(HloComputation* computation) { called_computations_.push_back(computation); } diff --git a/tensorflow/compiler/xla/service/hlo_instruction_test.cc b/tensorflow/compiler/xla/service/hlo_instruction_test.cc index 8ee24f9d92f61453a19a019c6e9c22ce37be1589..87c048930f4a179bb56e35f45bfe4b44e8951779 100644 --- a/tensorflow/compiler/xla/service/hlo_instruction_test.cc +++ b/tensorflow/compiler/xla/service/hlo_instruction_test.cc @@ -20,7 +20,7 @@ limitations under the License. #include #include -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/protobuf_util.h" #include "tensorflow/compiler/xla/service/dfs_hlo_visitor_with_default.h" #include "tensorflow/compiler/xla/service/hlo_computation.h" @@ -249,7 +249,7 @@ TEST_F(HloInstructionTest, MultipleUsersAndOperands) { auto param1 = builder.AddInstruction( HloInstruction::CreateParameter(1, r0f32_, "param1")); auto c0 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.1f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.1f))); auto addleft = builder.AddInstruction( HloInstruction::CreateBinary(r0f32_, HloOpcode::kAdd, param0, c0)); auto addright = builder.AddInstruction( @@ -294,7 +294,7 @@ TEST_F(HloInstructionTest, MultipleUsersAndOperandsWithUnaryOps) { auto param1 = builder.AddInstruction( HloInstruction::CreateParameter(1, r0f32_, "param1")); auto c0 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.1f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.1f))); auto neg1 = builder.AddInstruction( HloInstruction::CreateUnary(r0f32_, HloOpcode::kNegate, c0)); auto addleft = builder.AddInstruction( @@ -334,7 +334,7 @@ TEST_F(HloInstructionTest, TrivialMap) { auto param = embedded_builder.AddInstruction( HloInstruction::CreateParameter(0, r0f32, "x")); auto value = embedded_builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0))); embedded_builder.AddInstruction( HloInstruction::CreateBinary(r0f32, HloOpcode::kAdd, param, value)); auto add_f32 = module->AddEmbeddedComputation(embedded_builder.Build()); @@ -383,9 +383,9 @@ TEST_F(HloInstructionTest, TrivialReduce) { auto param0 = builder.AddInstruction( HloInstruction::CreateParameter(0, f32a100x10, "p")); auto const0 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(0.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(0.0f))); builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.1f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.1f))); auto reduce = builder.AddInstruction( HloInstruction::CreateReduce(f32v100, param0, const0, /*dimensions_to_reduce=*/{1}, add_f32)); @@ -626,7 +626,7 @@ TEST_F(HloInstructionTest, SingletonFusionOp) { HloComputation::Builder builder(TestName()); // Create a fusion instruction containing a single unary operation. auto constant = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.1f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.1f))); auto exp = builder.AddInstruction( HloInstruction::CreateUnary(r0f32_, HloOpcode::kExp, constant)); auto module = CreateNewModule(); @@ -642,9 +642,9 @@ TEST_F(HloInstructionTest, BinaryFusionOp) { HloComputation::Builder builder(TestName()); // Create a fusion instruction containing a single binary operation. auto constant1 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.1f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.1f))); auto constant2 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(42.1f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(42.1f))); auto add = builder.AddInstruction(HloInstruction::CreateBinary( r0f32_, HloOpcode::kAdd, constant1, constant2)); auto module = CreateNewModule(); @@ -661,7 +661,7 @@ TEST_F(HloInstructionTest, ChainFusionOp) { HloComputation::Builder builder(TestName()); // Create a chain of fused unary ops. auto constant = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.1f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.1f))); auto exp1 = builder.AddInstruction( HloInstruction::CreateUnary(r0f32_, HloOpcode::kExp, constant)); auto exp2 = builder.AddInstruction( @@ -682,7 +682,7 @@ TEST_F(HloInstructionTest, PreserveMetadataInFusionAndClone) { HloComputation::Builder builder(TestName()); // Create a chain of fused unary ops. auto constant = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.1f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.1f))); auto exp1 = builder.AddInstruction( HloInstruction::CreateUnary(r0f32_, HloOpcode::kExp, constant)); auto exp2 = builder.AddInstruction( @@ -710,16 +710,17 @@ TEST_F(HloInstructionTest, PreserveMetadataInFusionAndClone) { TEST_F(HloInstructionTest, PreserveOutfeedShapeThroughClone) { HloComputation::Builder builder(TestName()); auto constant = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR2({ + HloInstruction::CreateConstant(LiteralUtil::CreateR2({ {1, 2}, {3, 4}, }))); auto shape10 = ShapeUtil::MakeShapeWithLayout(F32, {2, 2}, {1, 0}); auto shape01 = ShapeUtil::MakeShapeWithLayout(F32, {2, 2}, {0, 1}); + auto token = builder.AddInstruction(HloInstruction::CreateToken()); auto outfeed10 = builder.AddInstruction( - HloInstruction::CreateOutfeed(shape10, constant, "")); + HloInstruction::CreateOutfeed(shape10, constant, token, "")); auto outfeed01 = builder.AddInstruction( - HloInstruction::CreateOutfeed(shape01, constant, "")); + HloInstruction::CreateOutfeed(shape01, constant, token, "")); auto clone01 = builder.AddInstruction(outfeed01->Clone()); auto clone10 = builder.AddInstruction(outfeed10->Clone()); @@ -731,7 +732,7 @@ TEST_F(HloInstructionTest, PreserveOutfeedShapeThroughClone) { TEST_F(HloInstructionTest, PreserveTupleShapeThroughClone) { HloComputation::Builder builder(TestName()); auto* constant = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR2({ + HloInstruction::CreateConstant(LiteralUtil::CreateR2({ {1, 2}, {3, 4}, }))); @@ -762,13 +763,13 @@ TEST_F(HloInstructionTest, FusionOpWithCalledComputations) { HloComputation::Builder builder(TestName()); auto constant = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.1f))); - auto map_1_x = builder.AddInstruction(HloInstruction::CreateMap( - scalar_shape, {constant}, computation_x, /*static_operands=*/{})); - auto map_2_x = builder.AddInstruction(HloInstruction::CreateMap( - scalar_shape, {map_1_x}, computation_x, /*static_operands=*/{})); - auto map_3_y = builder.AddInstruction(HloInstruction::CreateMap( - scalar_shape, {map_2_x}, computation_y, /*static_operands=*/{})); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.1f))); + auto map_1_x = builder.AddInstruction( + HloInstruction::CreateMap(scalar_shape, {constant}, computation_x)); + auto map_2_x = builder.AddInstruction( + HloInstruction::CreateMap(scalar_shape, {map_1_x}, computation_x)); + auto map_3_y = builder.AddInstruction( + HloInstruction::CreateMap(scalar_shape, {map_2_x}, computation_y)); auto* computation = module->AddEntryComputation(builder.Build()); auto* fusion = computation->CreateFusionInstruction( @@ -797,11 +798,11 @@ TEST_F(HloInstructionTest, ComplexFusionOp) { // Notable complexities are repeated operands in the same instruction, // different shapes, use of value in different expressions. auto c1 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.1f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.1f))); auto c2 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(2.1f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(2.1f))); auto c3 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(9.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(9.0f))); auto add = builder.AddInstruction( HloInstruction::CreateBinary(r0f32_, HloOpcode::kAdd, c1, c2)); @@ -872,11 +873,11 @@ TEST_F(HloInstructionTest, IdenticalInstructions) { // Create a set of random constant operands to use below. Make them matrices // so dimensions are interesting. auto operand1 = HloInstruction::CreateConstant( - Literal::CreateR2({{1.0, 2.0}, {3.0, 4.0}})); + LiteralUtil::CreateR2({{1.0, 2.0}, {3.0, 4.0}})); auto operand2 = HloInstruction::CreateConstant( - Literal::CreateR2({{10.0, 20.0}, {30.0, 40.0}})); - auto vector_operand = - HloInstruction::CreateConstant(Literal::CreateR1({42.0, 123.0})); + LiteralUtil::CreateR2({{10.0, 20.0}, {30.0, 40.0}})); + auto vector_operand = HloInstruction::CreateConstant( + LiteralUtil::CreateR1({42.0, 123.0})); Shape shape = operand1->shape(); // Convenient short names for the operands. @@ -1170,6 +1171,40 @@ TEST_F(HloInstructionTest, CloneOfFusionPreservesShape) { EXPECT_TRUE(StructuralEqual(*fusion, *fusion2)); } +TEST_F(HloInstructionTest, NoRedundantFusionOperandsAfterReplacingUse) { + // Fused expression: + // + // x y + // | | + // | transpose + // \ / + // dot + const Shape s = ShapeUtil::MakeShape(F32, {10, 10}); + + HloComputation::Builder builder("TransposeDot"); + HloInstruction* x = + builder.AddInstruction(HloInstruction::CreateParameter(0, s, "x")); + HloInstruction* y = + builder.AddInstruction(HloInstruction::CreateParameter(1, s, "y")); + HloInstruction* reshape = + builder.AddInstruction(HloInstruction::CreateTranspose(s, y, {1, 0})); + DotDimensionNumbers dot_dnums; + dot_dnums.add_lhs_contracting_dimensions(1); + dot_dnums.add_rhs_contracting_dimensions(0); + HloInstruction* dot = builder.AddInstruction( + HloInstruction::CreateDot(s, x, reshape, dot_dnums)); + + auto module = CreateNewModule(); + auto* computation = module->AddEntryComputation(builder.Build()); + HloInstruction* fusion = computation->CreateFusionInstruction( + {dot, reshape}, HloInstruction::FusionKind::kLoop); + + EXPECT_TRUE(x->ReplaceAllUsesWith(y).ok()); + + EXPECT_THAT(fusion->operands(), UnorderedElementsAre(y)); + EXPECT_EQ(fusion->fused_instructions_computation()->num_parameters(), 1); +} + TEST_F(HloInstructionTest, FusionEquality) { auto module = CreateNewModule(); HloComputation::Builder builder(TestName()); @@ -1199,9 +1234,9 @@ TEST_F(HloInstructionTest, NestedFusionEquality) { // Build a nested fusion computation. Shape data_shape = ShapeUtil::MakeShape(F32, {2, 2}); auto a = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR2({{1.0, 0.0}, {0.0, 1.0}}))); + LiteralUtil::CreateR2({{1.0, 0.0}, {0.0, 1.0}}))); auto b = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR2({{2.0, 2.0}, {2.0, 2.0}}))); + LiteralUtil::CreateR2({{2.0, 2.0}, {2.0, 2.0}}))); auto b_t = builder.AddInstruction( HloInstruction::CreateTranspose(data_shape, b, {1, 0})); DotDimensionNumbers dot_dnums; @@ -1210,7 +1245,7 @@ TEST_F(HloInstructionTest, NestedFusionEquality) { auto dot = builder.AddInstruction( HloInstruction::CreateDot(data_shape, a, b_t, dot_dnums)); auto one = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0))); auto add_operand = builder.AddInstruction( HloInstruction::CreateBroadcast(data_shape, one, {1})); auto add = builder.AddInstruction(HloInstruction::CreateBinary( @@ -1307,7 +1342,7 @@ TEST_F(HloInstructionTest, Stringification) { "condition=%TransposeDot, body=%TransposeDot"); auto pred = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(true))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(true))); HloInstruction* conditional = builder.AddInstruction(HloInstruction::CreateConditional( sout, pred, x, computation, x, computation)); @@ -1420,15 +1455,15 @@ TEST_F(HloInstructionTest, CanonnicalStringificationFusion) { HloInstruction* fusion = computation->CreateFusionInstruction( {dot, reshape}, HloInstruction::FusionKind::kLoop); - EXPECT_EQ( - fusion->ToString(options), + const string expected_fusion = R"(f32[5,20]{1,0} fusion(f32[5,10]{1,0}, f32[20,10]{1,0}), kind=kLoop, calls= { tmp_0 = f32[5,10]{1,0} parameter(0) tmp_1 = f32[20,10]{1,0} parameter(1) tmp_2 = f32[10,20]{1,0} transpose(f32[20,10]{1,0} tmp_1), dimensions={1,0} ROOT tmp_3 = f32[5,20]{1,0} dot(f32[5,10]{1,0} tmp_0, f32[10,20]{1,0} tmp_2), lhs_contracting_dims={1}, rhs_contracting_dims={0} -})"); +})"; + EXPECT_EQ(fusion->ToString(options), expected_fusion); } TEST_F(HloInstructionTest, CanonnicalStringificationWhile) { @@ -1460,8 +1495,8 @@ TEST_F(HloInstructionTest, CanonnicalStringificationWhile) { HloInstruction::CreateWhile(sout, computation, computation, x)); auto options = HloPrintOptions().Canonical(); - EXPECT_EQ(loop->ToString(options), - R"(f32[5,20]{1,0} while(f32[5,10]{1,0}), condition= + const string expected_loop = + R"(f32[5,20]{1,0} while(f32[5,10]{1,0}), condition= { tmp_0 = f32[5,10]{1,0} parameter(0) tmp_1 = f32[20,10]{1,0} parameter(1) @@ -1483,7 +1518,8 @@ TEST_F(HloInstructionTest, CanonnicalStringificationWhile) { tmp_2 = f32[10,20]{1,0} transpose(f32[20,10]{1,0} tmp_1), dimensions={1,0} ROOT tmp_3 = f32[5,20]{1,0} dot(f32[5,10]{1,0} tmp_0, f32[10,20]{1,0} tmp_2), lhs_contracting_dims={1}, rhs_contracting_dims={0} } -})"); +})"; + EXPECT_EQ(loop->ToString(options), expected_loop); } TEST_F(HloInstructionTest, CanonnicalStringificationConditional) { @@ -1515,13 +1551,12 @@ TEST_F(HloInstructionTest, CanonnicalStringificationConditional) { HloInstruction::CreateWhile(sout, computation, computation, x)); auto pred = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(true))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(true))); HloInstruction* conditional = builder.AddInstruction(HloInstruction::CreateConditional( sout, pred, x, computation, x, computation)); auto options = HloPrintOptions().Canonical(); - EXPECT_EQ( - conditional->ToString(options), + const string expected_conditional = R"(f32[5,20]{1,0} conditional(pred[], f32[5,10]{1,0}, f32[5,10]{1,0}), true_computation= { tmp_0 = f32[5,10]{1,0} parameter(0) @@ -1544,7 +1579,8 @@ TEST_F(HloInstructionTest, CanonnicalStringificationConditional) { tmp_2 = f32[10,20]{1,0} transpose(f32[20,10]{1,0} tmp_1), dimensions={1,0} ROOT tmp_3 = f32[5,20]{1,0} dot(f32[5,10]{1,0} tmp_0, f32[10,20]{1,0} tmp_2), lhs_contracting_dims={1}, rhs_contracting_dims={0} } -})"); +})"; + EXPECT_EQ(conditional->ToString(options), expected_conditional); } TEST_F(HloInstructionTest, CheckDeepClone) { diff --git a/tensorflow/compiler/xla/service/hlo_instructions.cc b/tensorflow/compiler/xla/service/hlo_instructions.cc index 803fde73a5a2ac7b239bcc992bd09a68e3b3233c..7ea42caa7b6af04fd4951a964f43f1c94a880f5c 100644 --- a/tensorflow/compiler/xla/service/hlo_instructions.cc +++ b/tensorflow/compiler/xla/service/hlo_instructions.cc @@ -17,10 +17,12 @@ limitations under the License. #include +#include "tensorflow/compiler/xla/literal_util.h" #include "tensorflow/compiler/xla/service/hlo_casting_utils.h" #include "tensorflow/compiler/xla/service/hlo_computation.h" #include "tensorflow/compiler/xla/service/hlo_module.h" #include "tensorflow/compiler/xla/window_util.h" +#include "tensorflow/core/lib/gtl/flatmap.h" namespace xla { namespace { @@ -203,25 +205,28 @@ bool HloSendRecvInstruction::IdenticalSlowPath( // Send instruction produces a tuple of {aliased operand, U32 context}. HloSendInstruction::HloSendInstruction(HloInstruction* operand, - int64 channel_id) + HloInstruction* token, int64 channel_id) : HloSendRecvInstruction( HloOpcode::kSend, - ShapeUtil::MakeTupleShape( - {CHECK_NOTNULL(operand)->shape(), ShapeUtil::MakeShape(U32, {})}), + ShapeUtil::MakeTupleShape({CHECK_NOTNULL(operand)->shape(), + ShapeUtil::MakeShape(U32, {}), + ShapeUtil::MakeTokenShape()}), channel_id) { AppendOperand(operand); + AppendOperand(token); } std::unique_ptr HloSendInstruction::CloneWithNewOperandsImpl( const Shape& shape, tensorflow::gtl::ArraySlice new_operands, HloCloneContext* context) const { - CHECK_EQ(new_operands.size(), 1); - return MakeUnique(new_operands[0], channel_id()); + CHECK_EQ(new_operands.size(), 2); + return MakeUnique(new_operands[0], new_operands[1], + channel_id()); } HloSendDoneInstruction::HloSendDoneInstruction(HloSendInstruction* operand) - : HloSendRecvInstruction(HloOpcode::kSendDone, ShapeUtil::MakeNil(), + : HloSendRecvInstruction(HloOpcode::kSendDone, ShapeUtil::MakeTokenShape(), CHECK_NOTNULL(operand)->channel_id()) { AppendOperand(operand); } @@ -237,25 +242,31 @@ HloSendDoneInstruction::CloneWithNewOperandsImpl( } // Recv instruction produces a tuple of {receive buffer, U32 context}. -HloRecvInstruction::HloRecvInstruction(const Shape& shape, int64 channel_id) +HloRecvInstruction::HloRecvInstruction(const Shape& shape, + HloInstruction* token, int64 channel_id) : HloSendRecvInstruction( HloOpcode::kRecv, - ShapeUtil::MakeTupleShape({shape, ShapeUtil::MakeShape(U32, {})}), - channel_id) {} + ShapeUtil::MakeTupleShape({shape, ShapeUtil::MakeShape(U32, {}), + ShapeUtil::MakeTokenShape()}), + channel_id) { + AppendOperand(token); +} std::unique_ptr HloRecvInstruction::CloneWithNewOperandsImpl( const Shape& shape, tensorflow::gtl::ArraySlice new_operands, HloCloneContext* context) const { - CHECK_EQ(new_operands.size(), 0); + CHECK_EQ(new_operands.size(), 1); return MakeUnique( - ShapeUtil::GetTupleElementShape(shape, 0), channel_id()); + ShapeUtil::GetTupleElementShape(shape, 0), new_operands[0], channel_id()); } HloRecvDoneInstruction::HloRecvDoneInstruction(HloRecvInstruction* operand) : HloSendRecvInstruction( HloOpcode::kRecvDone, - ShapeUtil::GetTupleElementShape(operand->shape(), 0), + ShapeUtil::MakeTupleShape( + {ShapeUtil::GetTupleElementShape(operand->shape(), 0), + ShapeUtil::MakeTokenShape()}), CHECK_NOTNULL(operand)->channel_id()) { AppendOperand(operand); } @@ -458,6 +469,46 @@ std::unique_ptr HloReduceInstruction::CloneWithNewOperandsImpl( shape, new_operands[0], new_operands[1], dimensions(), to_apply()); } +HloSortInstruction::HloSortInstruction(const Shape& shape, int64 dimension, + HloInstruction* keys, + HloInstruction* values) + : HloInstruction(HloOpcode::kSort, shape), dimensions_({dimension}) { + AppendOperand(keys); + if (values) { + AppendOperand(values); + } +} + +HloInstructionProto HloSortInstruction::ToProto() const { + HloInstructionProto proto = HloInstruction::ToProto(); + for (int64 dimension : dimensions_) { + proto.add_dimensions(dimension); + } + return proto; +} + +std::vector HloSortInstruction::ExtraAttributesToStringImpl( + const HloPrintOptions& options) const { + return {StrCat("dimensions={", Join(dimensions(), ","), "}")}; +} + +bool HloSortInstruction::IdenticalSlowPath( + const HloInstruction& other, + const std::function& + eq_computations) const { + const auto& casted_other = static_cast(other); + return dimensions() == casted_other.dimensions(); +} + +std::unique_ptr HloSortInstruction::CloneWithNewOperandsImpl( + const Shape& shape, + tensorflow::gtl::ArraySlice new_operands, + HloCloneContext* context) const { + HloInstruction* keys = new_operands[0]; + HloInstruction* values = new_operands.size() == 2 ? new_operands[1] : nullptr; + return MakeUnique(shape, dimensions(0), keys, values); +} + HloTransposeInstruction::HloTransposeInstruction( const Shape& shape, HloInstruction* operand, tensorflow::gtl::ArraySlice dimensions) @@ -553,10 +604,8 @@ HloBroadcastInstruction::CloneWithNewOperandsImpl( HloMapInstruction::HloMapInstruction( const Shape& shape, tensorflow::gtl::ArraySlice operands, - HloComputation* map_computation, - tensorflow::gtl::ArraySlice static_operands) + HloComputation* map_computation) : HloInstruction(HloOpcode::kMap, shape) { - CHECK(static_operands.empty()) << "static_operands not yet supported"; for (auto operand : operands) { AppendOperand(operand); } @@ -758,7 +807,7 @@ string HloConstantInstruction::OperandsToStringWithCanonicalNameMap( HloTraceInstruction::HloTraceInstruction(const string& tag, HloInstruction* operand) : HloInstruction(HloOpcode::kTrace, ShapeUtil::MakeNil()), - literal_(Literal::CreateR1U8(tag)) { + literal_(LiteralUtil::CreateR1U8(tag)) { AppendOperand(operand); operand->set_tracing(this); } @@ -1044,8 +1093,6 @@ HloInstruction* HloFusionInstruction::CloneAndFuseInternal( CHECK_NOTNULL(GetModule())->AddEmbeddedComputation(builder.Build())); clone = fused_expression_root(); } else { - clone = fused_instructions_computation()->AddInstruction( - instruction_to_fuse->Clone(/*suffix=*/"")); // When add_output is false, instruction_to_fuse is necessarily an operand // of the fusion instruction. After fusion this will no longer be the // case. Remove the operand from the operand list and remove its @@ -1055,6 +1102,16 @@ HloInstruction* HloFusionInstruction::CloneAndFuseInternal( bool in_operand_list = std::find(operands().begin(), operands().end(), instruction_to_fuse) != operands().end(); CHECK(add_output || in_operand_list); + if (instruction_to_fuse->opcode() == HloOpcode::kTuple) { + // We assume all uses of a kTuple operation are GTE ops, not another + // fusion node. In this case, we don't need to clone + // 'instruction_to_fuse'. + CHECK(!in_operand_list); + clone = instruction_to_fuse; + } else { + clone = fused_instructions_computation()->AddInstruction( + instruction_to_fuse->Clone(/*suffix=*/"")); + } const std::vector& fused_parameters = fused_instructions_computation()->parameter_instructions(); for (int64 operand_num = 0; operand_num < operand_count(); ++operand_num) { @@ -1151,9 +1208,10 @@ HloInstruction* HloFusionInstruction::CloneAndFuseInternal( } int64 index = tuple_elements.size(); if (instruction_to_fuse->opcode() == HloOpcode::kTuple) { - index -= instruction_to_fuse->operand_count(); + CHECK_EQ(clone, instruction_to_fuse); + index -= clone->operand_count(); std::vector to_be_removed; - for (auto old_gte : instruction_to_fuse->users()) { + for (auto old_gte : clone->users()) { CHECK_EQ(old_gte->opcode(), HloOpcode::kGetTupleElement); int64 old_tuple_index = old_gte->tuple_index(); HloInstruction* new_gte = @@ -1165,7 +1223,6 @@ HloInstruction* HloFusionInstruction::CloneAndFuseInternal( for (auto old_gte : to_be_removed) { TF_CHECK_OK(parent()->RemoveInstruction(old_gte)); } - TF_CHECK_OK(fused_instructions_computation()->RemoveInstruction(clone)); } else { HloInstruction* new_gte = parent()->AddInstruction(HloInstruction::CreateGetTupleElement( @@ -1174,7 +1231,9 @@ HloInstruction* HloFusionInstruction::CloneAndFuseInternal( } } - VLOG(2) << "New clone:\n" << clone->ToString(); + if (clone != instruction_to_fuse) { + VLOG(2) << "New clone:\n" << clone->ToString(); + } return clone; } @@ -1210,6 +1269,26 @@ std::unique_ptr HloFusionInstruction::CloneWithNewOperandsImpl( new_fused_computation); } +Status HloFusionInstruction::DeduplicateFusionOperands() { + tensorflow::gtl::FlatMap operand_indices; + std::vector operands_to_remove; + for (int i = 0; i < operand_count(); ++i) { + auto emplace_result = operand_indices.emplace(operand(i), i); + if (!emplace_result.second) { + TF_RETURN_IF_ERROR(fused_parameter(i)->ReplaceAllUsesWith( + fused_parameter(emplace_result.first->second))); + operands_to_remove.push_back(i); + } + } + if (operands_to_remove.empty()) { + return Status::OK(); + } + TF_RETURN_IF_ERROR( + fused_instructions_computation()->RemoveUnusedParameters()); + RemoveOperandsAtAscendingIndices(operands_to_remove); + return Status::OK(); +} + HloRngInstruction::HloRngInstruction( const Shape& shape, RandomDistribution distribution, tensorflow::gtl::ArraySlice parameters) @@ -1365,9 +1444,22 @@ HloReducePrecisionInstruction::CloneWithNewOperandsImpl( shape, new_operands[0], exponent_bits(), mantissa_bits()); } -HloInfeedInstruction::HloInfeedInstruction(const Shape& shape, +HloInfeedInstruction::HloInfeedInstruction(const Shape& infeed_shape, + HloInstruction* token_operand, const string& config) - : HloInstruction(HloOpcode::kInfeed, shape), infeed_config_(config) {} + : HloInstruction(HloOpcode::kInfeed, + ShapeUtil::MakeTupleShape( + {infeed_shape, ShapeUtil::MakeTokenShape()})), + infeed_config_(config) { + AppendOperand(token_operand); +} + +HloInfeedInstruction::HloInfeedInstruction(const Shape& infeed_shape, + const string& config) + : HloInstruction(HloOpcode::kInfeed, + ShapeUtil::MakeTupleShape( + {infeed_shape, ShapeUtil::MakeTokenShape()})), + infeed_config_(config) {} HloInstructionProto HloInfeedInstruction::ToProto() const { HloInstructionProto proto = HloInstruction::ToProto(); @@ -1395,19 +1487,37 @@ std::unique_ptr HloInfeedInstruction::CloneWithNewOperandsImpl( const Shape& shape, tensorflow::gtl::ArraySlice new_operands, HloCloneContext* context) const { - CHECK_EQ(new_operands.size(), 0); - return MakeUnique(shape, infeed_config()); + if (new_operands.empty()) { + return MakeUnique(infeed_shape(), infeed_config()); + } else { + CHECK_EQ(new_operands.size(), 1); + return MakeUnique(infeed_shape(), new_operands[0], + infeed_config()); + } } HloOutfeedInstruction::HloOutfeedInstruction( - const Shape& shape, HloInstruction* operand, + const Shape& outfeed_shape, HloInstruction* operand, + HloInstruction* token_operand, tensorflow::StringPiece outfeed_config) + : HloInstruction(HloOpcode::kOutfeed, ShapeUtil::MakeTokenShape()), + outfeed_shape_(outfeed_shape), + outfeed_config_(outfeed_config.begin(), outfeed_config.end()) { + CHECK(ShapeUtil::Compatible(operand->shape(), outfeed_shape)) + << "Outfeed shape " << outfeed_shape + << " must be compatible with operand shape " << operand->shape(); + AppendOperand(operand); + AppendOperand(token_operand); +} + +HloOutfeedInstruction::HloOutfeedInstruction( + const Shape& outfeed_shape, HloInstruction* operand, tensorflow::StringPiece outfeed_config) - : HloInstruction(HloOpcode::kOutfeed, ShapeUtil::MakeNil()), - outfeed_shape_(shape), + : HloInstruction(HloOpcode::kOutfeed, ShapeUtil::MakeTokenShape()), + outfeed_shape_(outfeed_shape), outfeed_config_(outfeed_config.begin(), outfeed_config.end()) { - CHECK(ShapeUtil::Compatible(operand->shape(), shape)) - << "Outfeed shape " << shape << " must be compatible with operand shape " - << operand->shape(); + CHECK(ShapeUtil::Compatible(operand->shape(), outfeed_shape)) + << "Outfeed shape " << outfeed_shape + << " must be compatible with operand shape " << operand->shape(); AppendOperand(operand); } @@ -1438,9 +1548,14 @@ std::unique_ptr HloOutfeedInstruction::CloneWithNewOperandsImpl( const Shape& shape, tensorflow::gtl::ArraySlice new_operands, HloCloneContext* context) const { - CHECK_EQ(new_operands.size(), 1); - return MakeUnique(outfeed_shape(), new_operands[0], - outfeed_config()); + if (new_operands.size() == 1) { + return MakeUnique(outfeed_shape(), new_operands[0], + outfeed_config()); + } else { + CHECK_EQ(new_operands.size(), 2); + return MakeUnique(outfeed_shape(), new_operands[0], + new_operands[1], outfeed_config()); + } } HloConvolutionInstruction::HloConvolutionInstruction( diff --git a/tensorflow/compiler/xla/service/hlo_instructions.h b/tensorflow/compiler/xla/service/hlo_instructions.h index 1a2e4ae0a587d889f3064e24f9cda61f34517818..e922d94234ff7b54c93be4d76fa0e3496ed2df88 100644 --- a/tensorflow/compiler/xla/service/hlo_instructions.h +++ b/tensorflow/compiler/xla/service/hlo_instructions.h @@ -161,7 +161,8 @@ class HloSendRecvInstruction : public HloInstruction { class HloSendInstruction : public HloSendRecvInstruction { public: - explicit HloSendInstruction(HloInstruction* operand, int64 channel_id); + explicit HloSendInstruction(HloInstruction* operand, HloInstruction* token, + int64 channel_id); private: // Implementation for non-common logic of CloneWithNewOperands. @@ -185,7 +186,8 @@ class HloSendDoneInstruction : public HloSendRecvInstruction { class HloRecvInstruction : public HloSendRecvInstruction { public: - explicit HloRecvInstruction(const Shape& shape, int64 channel_id); + explicit HloRecvInstruction(const Shape& shape, HloInstruction* token, + int64 channel_id); private: // Implementation for non-common logic of CloneWithNewOperands. @@ -347,6 +349,35 @@ class HloReduceInstruction : public HloInstruction { std::vector dimensions_; }; +class HloSortInstruction : public HloInstruction { + public: + explicit HloSortInstruction(const Shape& shape, int64 dimension, + HloInstruction* keys, + HloInstruction* values = nullptr); + // Returns the dimension sizes or numbers associated with this instruction. + const std::vector& dimensions() const override { return dimensions_; } + int64 dimensions(int64 index) const override { return dimensions()[index]; } + // Returns the sort dimension for this instruction + int64 sort_dimension() { return dimensions(0); } + // Returns a serialized representation of this instruction. + HloInstructionProto ToProto() const override; + + private: + std::vector ExtraAttributesToStringImpl( + const HloPrintOptions& options) const override; + bool IdenticalSlowPath( + const HloInstruction& other, + const std::function& + eq_computations) const override; + // Implementation for non-common logic of CloneWithNewOperands. + std::unique_ptr CloneWithNewOperandsImpl( + const Shape& shape, + tensorflow::gtl::ArraySlice new_operands, + HloCloneContext* context) const override; + + std::vector dimensions_; +}; + class HloTransposeInstruction : public HloInstruction { public: explicit HloTransposeInstruction( @@ -407,8 +438,7 @@ class HloMapInstruction : public HloInstruction { public: explicit HloMapInstruction( const Shape& shape, tensorflow::gtl::ArraySlice operands, - HloComputation* map_computation, - tensorflow::gtl::ArraySlice static_operands = {}); + HloComputation* map_computation); // Returns the dimension sizes or numbers associated with this instruction. const std::vector& dimensions() const override { return dimensions_; } int64 dimensions(int64 index) const override { return dimensions()[index]; } @@ -636,6 +666,9 @@ class HloFusionInstruction : public HloInstruction { void set_fusion_kind(FusionKind kind) { fusion_kind_ = kind; } + // If multiple operands are the same instruction, keeps only one of them. + Status DeduplicateFusionOperands(); + private: // Fuses the given instruction into this fusion instruction. When add_output // is false (which is the default), instruction_to_fuse is cloned and the @@ -785,12 +818,25 @@ class HloReducePrecisionInstruction : public HloInstruction { class HloInfeedInstruction : public HloInstruction { public: - explicit HloInfeedInstruction(const Shape& shape, const string& config); + explicit HloInfeedInstruction(const Shape& infeed_shape, + HloInstruction* token_operand, + const string& config); + // TODO(b/80000000): Remove this constructor when all uses of infeed are + // converted to take tokens. + explicit HloInfeedInstruction(const Shape& infeed_shape, + const string& config); // Returns the infeed configuration string. The infeed configuration includes // any metadata needed for the backend compiler (e.g., infeed buffer address) // and is target-dependent. string infeed_config() const { return infeed_config_; } void set_infeed_config(const string& config) { infeed_config_ = config; } + // Returns the shape of the data received by the infeed. This is not the same + // as the shape of the infeed instruction which produces a tuple containing + // the infeed data shape and a TOKEN. + const Shape& infeed_shape() const { + TF_DCHECK_OK(ShapeUtil::ValidateShapeWithOptionalLayout(shape())); + return ShapeUtil::GetSubshape(shape(), {0}); + } // Returns a serialized representation of this instruction. HloInstructionProto ToProto() const override; @@ -813,11 +859,19 @@ class HloInfeedInstruction : public HloInstruction { class HloOutfeedInstruction : public HloInstruction { public: - explicit HloOutfeedInstruction(const Shape& shape, HloInstruction* operand, + explicit HloOutfeedInstruction(const Shape& outfeed_shape, + HloInstruction* operand, + HloInstruction* token_operand, tensorflow::StringPiece outfeed_config); + // TODO(b/80000000): Remove this constructor when all uses of outfeed are + // converted to take tokens. + explicit HloOutfeedInstruction(const Shape& outfeed_shape, + HloInstruction* operand, + tensorflow::StringPiece outfeed_config); + // Returns the shape for the Outfeed instruction. const Shape& outfeed_shape() const { - TF_DCHECK_OK(ShapeUtil::ValidateShapeWithOptionalLayout(shape())); + TF_DCHECK_OK(ShapeUtil::ValidateShapeWithOptionalLayout(outfeed_shape_)); return outfeed_shape_; } // Returns the config for the Outfeed instruction. diff --git a/tensorflow/compiler/xla/service/hlo_liveness_analysis_test.cc b/tensorflow/compiler/xla/service/hlo_liveness_analysis_test.cc index 0275294a1a86cef13e5b267ad578f30cc18858dc..01b625c29ca2823b2a2490b30a9d4d5128b4c22e 100644 --- a/tensorflow/compiler/xla/service/hlo_liveness_analysis_test.cc +++ b/tensorflow/compiler/xla/service/hlo_liveness_analysis_test.cc @@ -15,7 +15,7 @@ limitations under the License. #include "tensorflow/compiler/xla/service/hlo_liveness_analysis.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/service/hlo_computation.h" #include "tensorflow/compiler/xla/service/hlo_opcode.h" #include "tensorflow/compiler/xla/service/hlo_parser.h" diff --git a/tensorflow/compiler/xla/service/hlo_matchers.h b/tensorflow/compiler/xla/service/hlo_matchers.h index 8a31a8e617c1fb82201e07d9a3ff1ab9a618206b..b57c940238f0672692e3b65827f43e2f5499502d 100644 --- a/tensorflow/compiler/xla/service/hlo_matchers.h +++ b/tensorflow/compiler/xla/service/hlo_matchers.h @@ -187,7 +187,7 @@ HLO_MATCHER(Exp); HLO_MATCHER(Floor); HLO_MATCHER(Fusion); HLO_MATCHER(Ge); -HLO_MATCHER(GenerateToken); +HLO_MATCHER(AfterAll); HLO_MATCHER(Gt); HLO_MATCHER(Infeed); HLO_MATCHER(IsFinite); @@ -196,6 +196,7 @@ HLO_MATCHER(Log); HLO_MATCHER(And); HLO_MATCHER(Not); HLO_MATCHER(Or); +HLO_MATCHER(Xor); HLO_MATCHER(Lt); HLO_MATCHER(Map); HLO_MATCHER(Maximum); diff --git a/tensorflow/compiler/xla/service/hlo_matchers_test.cc b/tensorflow/compiler/xla/service/hlo_matchers_test.cc index 9a3010cf1ff75e840130d8442bbe26d6041cef25..7de59acc1efbc0150b95ebdd85a13ede48eec2f9 100644 --- a/tensorflow/compiler/xla/service/hlo_matchers_test.cc +++ b/tensorflow/compiler/xla/service/hlo_matchers_test.cc @@ -14,6 +14,7 @@ limitations under the License. ==============================================================================*/ #include "tensorflow/compiler/xla/service/hlo_matchers.h" +#include "tensorflow/compiler/xla/literal_util.h" #include "tensorflow/compiler/xla/service/hlo_parser.h" #include "tensorflow/compiler/xla/shape_util.h" @@ -75,8 +76,10 @@ TEST(HloMatchersTest, Test) { } TEST(HloMatchersTest, CustomCallMatcher) { - auto c1 = HloInstruction::CreateConstant(Literal::CreateR1({1, 2, 3})); - auto c2 = HloInstruction::CreateConstant(Literal::CreateR1({1, 2, 3})); + auto c1 = + HloInstruction::CreateConstant(LiteralUtil::CreateR1({1, 2, 3})); + auto c2 = + HloInstruction::CreateConstant(LiteralUtil::CreateR1({1, 2, 3})); auto call = HloInstruction::CreateCustomCall( ShapeUtil::MakeShape(F32, {1}), {c1.get(), c2.get()}, "foo_target"); diff --git a/tensorflow/compiler/xla/service/hlo_module.cc b/tensorflow/compiler/xla/service/hlo_module.cc index 39bc25ba42c2cb6a9f77e2726405311ba13b3edc..55ff073d3faf34aa0f1b8f0886946837e7a49bcc 100644 --- a/tensorflow/compiler/xla/service/hlo_module.cc +++ b/tensorflow/compiler/xla/service/hlo_module.cc @@ -537,10 +537,11 @@ uint64 HloModule::RandomNew64() const { HloComputation* HloModule::GetComputationWithName( tensorflow::StringPiece name) { - auto it = c_find_if(computations(), [&](HloComputation* computation) { + auto computations_in_module = computations(); + auto it = c_find_if(computations_in_module, [&](HloComputation* computation) { return computation->name() == name; }); - return it == computations().end() ? nullptr : *it; + return it == computations_in_module.end() ? nullptr : *it; } /* static */ std::atomic HloModule::next_unique_module_id_(0); diff --git a/tensorflow/compiler/xla/service/hlo_module_group_metadata.cc b/tensorflow/compiler/xla/service/hlo_module_group_metadata.cc index bf33640db16638803f4f8e6c66f35d6bb6e2c9fe..6bcd7b042dfddfea6ac86365b82f8077be2a6101 100644 --- a/tensorflow/compiler/xla/service/hlo_module_group_metadata.cc +++ b/tensorflow/compiler/xla/service/hlo_module_group_metadata.cc @@ -382,7 +382,8 @@ Status HloModuleGroupMetadata::VerifyChannelInstructions() { // Check if the shapes match for each channel. for (const Channel& channel : channels_) { const Shape& send_shape = channel.send->operand(0)->shape(); - const Shape& recv_shape = channel.recv_done->shape(); + const Shape& recv_shape = + ShapeUtil::GetTupleElementShape(channel.recv_done->shape(), 0); if (!ShapeUtil::Compatible(send_shape, recv_shape)) { return FailedPrecondition("send/recv shapes do not match"); } diff --git a/tensorflow/compiler/xla/service/hlo_module_group_util.cc b/tensorflow/compiler/xla/service/hlo_module_group_util.cc index 21a9b7291acc9e0066a9061facd13ab5acbf0bac..df1d562048f245b0edae7c00a21ff406340fd012 100644 --- a/tensorflow/compiler/xla/service/hlo_module_group_util.cc +++ b/tensorflow/compiler/xla/service/hlo_module_group_util.cc @@ -292,7 +292,7 @@ HloModuleGroupUtil::ComputeReachability( } auto reachability = MakeUnique(post_order); for (HloInstruction* hlo : post_order) { - reachability->SetReachabilityToUnion(GlobalPredecessors(hlo), hlo); + reachability->FastSetReachabilityToUnion(GlobalPredecessors(hlo), hlo); } return std::move(reachability); } diff --git a/tensorflow/compiler/xla/service/hlo_module_test.cc b/tensorflow/compiler/xla/service/hlo_module_test.cc index 7f28a804bfec9c2f1bbb5fa08f7dd4e68be14d35..236f4500860a8673e61cbd2f861a8fc40c7861f7 100644 --- a/tensorflow/compiler/xla/service/hlo_module_test.cc +++ b/tensorflow/compiler/xla/service/hlo_module_test.cc @@ -15,7 +15,7 @@ limitations under the License. #include "tensorflow/compiler/xla/service/hlo_module.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/ptr_util.h" #include "tensorflow/compiler/xla/service/hlo_computation.h" #include "tensorflow/compiler/xla/service/hlo_instruction.h" @@ -38,7 +38,7 @@ class HloModuleTest : public HloTestBase { std::unique_ptr CreateConstantComputation() { auto builder = HloComputation::Builder("Constant"); builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(42.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(42.0f))); return builder.Build(); } @@ -122,7 +122,7 @@ TEST_F(HloModuleTest, CloneHasFusion) { { auto b = HloComputation::Builder("Entry"); auto input = b.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(42.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(42.0f))); b.AddInstruction( HloInstruction::CreateFusion(r0f32_, HloInstruction::FusionKind::kInput, /*operands=*/{input}, fused_computation)); @@ -173,7 +173,7 @@ TEST_F(HloModuleTest, LargeConstantToString) { auto builder = HloComputation::Builder("Constant"); std::vector values(16, 42.0); builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR1(values))); + HloInstruction::CreateConstant(LiteralUtil::CreateR1(values))); module->AddEntryComputation(builder.Build()); EXPECT_EQ( diff --git a/tensorflow/compiler/xla/service/hlo_opcode.h b/tensorflow/compiler/xla/service/hlo_opcode.h index 70833212769def47bdba6f7c491de5fd9ea51298..39e12c48157992410a5d3b733720d677a1191611 100644 --- a/tensorflow/compiler/xla/service/hlo_opcode.h +++ b/tensorflow/compiler/xla/service/hlo_opcode.h @@ -81,7 +81,7 @@ namespace xla { V(kFusion, "fusion", kHloOpcodeIsVariadic) \ V(kGather, "gather") \ V(kGe, "greater-than-or-equal-to", kHloOpcodeIsComparison) \ - V(kGenerateToken, "generate-token", kHloOpcodeIsVariadic) \ + V(kAfterAll, "after-all", kHloOpcodeIsVariadic) \ V(kGetTupleElement, "get-tuple-element") \ V(kGt, "greater-than", kHloOpcodeIsComparison) \ V(kHostCompute, "host-compute") \ @@ -133,6 +133,7 @@ namespace xla { V(kTrace, "trace") \ V(kTranspose, "transpose") \ V(kTuple, "tuple", kHloOpcodeIsVariadic) \ + V(kTupleSelect, "tuple-select") \ V(kWhile, "while") enum class HloOpcode { diff --git a/tensorflow/compiler/xla/service/hlo_opcode_test.cc b/tensorflow/compiler/xla/service/hlo_opcode_test.cc index 774345124b4ad62e35d9423a23f1dbaa28e44d80..6f3f83f63a05fafaa3f3ddcff8a7cac7cb7b06d5 100644 --- a/tensorflow/compiler/xla/service/hlo_opcode_test.cc +++ b/tensorflow/compiler/xla/service/hlo_opcode_test.cc @@ -58,7 +58,7 @@ TEST(HloOpcodeTest, OpcodeProperties) { case HloOpcode::kConcatenate: case HloOpcode::kFusion: case HloOpcode::kMap: - case HloOpcode::kGenerateToken: + case HloOpcode::kAfterAll: case HloOpcode::kTuple: EXPECT_TRUE(HloOpcodeIsVariadic(opcode)); break; diff --git a/tensorflow/compiler/xla/service/hlo_ordering_test.cc b/tensorflow/compiler/xla/service/hlo_ordering_test.cc index cfe5dace05ac03f1573f90b2ce664c94837837b4..126d3a2d9c70bff1d2a022e395652049768d6d21 100644 --- a/tensorflow/compiler/xla/service/hlo_ordering_test.cc +++ b/tensorflow/compiler/xla/service/hlo_ordering_test.cc @@ -57,7 +57,7 @@ TEST_F(HloOrderingTest, InstructionsInDifferentComputations) { auto builder_c = HloComputation::Builder("C"); HloInstruction* c = builder_c.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(42.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(42.0f))); HloComputation* computation_c = module->AddEmbeddedComputation(builder_c.Build()); @@ -145,7 +145,7 @@ TEST_F(HloOrderingTest, InstructionsInWhileComputations) { auto builder = HloComputation::Builder(TestName()); auto constant = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0))); auto xla_while = builder.AddInstruction( HloInstruction::CreateWhile(scalar_shape, condition, body, constant)); module->AddEntryComputation(builder.Build()); @@ -208,7 +208,7 @@ TEST_F(HloOrderingTest, ValuesInWhileComputations) { auto builder = HloComputation::Builder(TestName()); auto constant = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0))); auto xla_while = builder.AddInstruction( HloInstruction::CreateWhile(scalar_shape, condition, body, constant)); auto add = builder.AddInstruction(HloInstruction::CreateBinary( diff --git a/tensorflow/compiler/xla/service/hlo_parser.cc b/tensorflow/compiler/xla/service/hlo_parser.cc index 605c6ae7417204a5b58213e284bf060f3bc19b97..f162d52d3cfa33e2ee1e8ed1d833f9db90d93c0c 100644 --- a/tensorflow/compiler/xla/service/hlo_parser.cc +++ b/tensorflow/compiler/xla/service/hlo_parser.cc @@ -15,6 +15,7 @@ limitations under the License. #include "tensorflow/compiler/xla/service/hlo_parser.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/literal_util.h" #include "tensorflow/compiler/xla/service/hlo_domain_metadata.h" #include "tensorflow/compiler/xla/service/hlo_opcode.h" @@ -509,7 +510,6 @@ bool HloParser::ParseInstruction(HloComputation::Builder* builder, case HloOpcode::kReal: case HloOpcode::kSign: case HloOpcode::kSin: - case HloOpcode::kSort: case HloOpcode::kTanh: { if (!ParseOperands(&operands, /*expected_size=*/1) || !ParseAttributes(attrs)) { @@ -552,7 +552,8 @@ bool HloParser::ParseInstruction(HloComputation::Builder* builder, } // Ternary ops. case HloOpcode::kClamp: - case HloOpcode::kSelect: { + case HloOpcode::kSelect: + case HloOpcode::kTupleSelect: { if (!ParseOperands(&operands, /*expected_size=*/3) || !ParseAttributes(attrs)) { return false; @@ -617,12 +618,42 @@ bool HloParser::ParseInstruction(HloComputation::Builder* builder, HloInstruction::CreateReshape(shape, operands[0])); break; } - case HloOpcode::kGenerateToken: { + case HloOpcode::kAfterAll: { if (!ParseOperands(&operands) || !ParseAttributes(attrs)) { return false; } - instruction = builder->AddInstruction( - HloInstruction::CreateGenerateToken(operands)); + if (operands.empty()) { + instruction = builder->AddInstruction(HloInstruction::CreateToken()); + } else { + instruction = + builder->AddInstruction(HloInstruction::CreateAfterAll(operands)); + } + break; + } + case HloOpcode::kSort: { + auto loc = lexer_.GetLoc(); + + optional> dimensions; + attrs["dimensions"] = {/*required=*/true, AttrTy::kBracedInt64List, + &dimensions}; + if (!ParseOperands(&operands) || !ParseAttributes(attrs) || + dimensions->size() != 1) { + return false; + } + switch (operands.size()) { + case 1: + instruction = builder->AddInstruction(HloInstruction::CreateSort( + shape, dimensions->at(0), /*keys=*/operands[0])); + break; + case 2: + instruction = builder->AddInstruction(HloInstruction::CreateSort( + shape, dimensions->at(0), + /*keys=*/operands[0], /*values=*/operands[1])); + break; + default: + return Error(loc, StrCat("expects either 1 or 2 operands, but has ", + operands.size(), " operands")); + } break; } case HloOpcode::kTuple: { @@ -650,12 +681,12 @@ bool HloParser::ParseInstruction(HloComputation::Builder* builder, case HloOpcode::kRecv: { optional channel_id; attrs["channel_id"] = {/*required=*/true, AttrTy::kInt64, &channel_id}; - if (!ParseOperands(&operands, /*expected_size=*/0) || + if (!ParseOperands(&operands, /*expected_size=*/1) || !ParseAttributes(attrs)) { return false; } - instruction = builder->AddInstruction( - HloInstruction::CreateRecv(shape.tuple_shapes(0), *channel_id)); + instruction = builder->AddInstruction(HloInstruction::CreateRecv( + shape.tuple_shapes(0), operands[0], *channel_id)); break; } case HloOpcode::kRecvDone: { @@ -675,12 +706,12 @@ bool HloParser::ParseInstruction(HloComputation::Builder* builder, case HloOpcode::kSend: { optional channel_id; attrs["channel_id"] = {/*required=*/true, AttrTy::kInt64, &channel_id}; - if (!ParseOperands(&operands, /*expected_size=*/1) || + if (!ParseOperands(&operands, /*expected_size=*/2) || !ParseAttributes(attrs)) { return false; } instruction = builder->AddInstruction( - HloInstruction::CreateSend(operands[0], *channel_id)); + HloInstruction::CreateSend(operands[0], operands[1], *channel_id)); break; } case HloOpcode::kSendDone: { @@ -978,23 +1009,53 @@ bool HloParser::ParseInstruction(HloComputation::Builder* builder, case HloOpcode::kInfeed: { optional config; attrs["infeed_config"] = {/*required=*/false, AttrTy::kString, &config}; - if (!ParseOperands(&operands, /*expected_size=*/0) || - !ParseAttributes(attrs)) { + if (!ParseOperands(&operands) || !ParseAttributes(attrs)) { return false; } - instruction = builder->AddInstruction( - HloInstruction::CreateInfeed(shape, config ? *config : "")); + // We need to know the infeed data shape to construct the infeed + // instruction. This is the zero-th element of the tuple-shaped output of + // the infeed instruction. ShapeUtil::GetTupleElementShape will check fail + // if the shape is not a non-empty tuple, so add guard so an error message + // can be emitted instead of a check fail + if (!ShapeUtil::IsTuple(shape) && !ShapeUtil::IsEmptyTuple(shape)) { + return Error(lexer_.GetLoc(), + "infeed must have a non-empty tuple shape"); + } + + if (operands.empty()) { + // TODO(b/80000000): Remove this when all uses of infeed are + // converted to take tokens. + instruction = builder->AddInstruction(HloInstruction::CreateInfeed( + ShapeUtil::GetTupleElementShape(shape, 0), config ? *config : "")); + } else if (operands.size() == 1) { + instruction = builder->AddInstruction(HloInstruction::CreateInfeed( + ShapeUtil::GetTupleElementShape(shape, 0), operands[0], + config ? *config : "")); + } else { + return Error(lexer_.GetLoc(), + "infeed must have exactly zero or one operands"); + } break; } case HloOpcode::kOutfeed: { optional config; attrs["outfeed_config"] = {/*required=*/false, AttrTy::kString, &config}; - if (!ParseOperands(&operands, /*expected_size=*/1) || - !ParseAttributes(attrs)) { + if (!ParseOperands(&operands) || !ParseAttributes(attrs)) { return false; } - instruction = builder->AddInstruction(HloInstruction::CreateOutfeed( - operands[0]->shape(), operands[0], config ? *config : "")); + if (operands.size() == 1) { + // TODO(b/80000000): Remove this when all uses of outfeed are + // converted to take tokens. + instruction = builder->AddInstruction(HloInstruction::CreateOutfeed( + operands[0]->shape(), operands[0], config ? *config : "")); + } else if (operands.size() == 2) { + instruction = builder->AddInstruction( + HloInstruction::CreateOutfeed(operands[0]->shape(), operands[0], + operands[1], config ? *config : "")); + } else { + return Error(lexer_.GetLoc(), + "outfeed must have exactly one or two operands"); + } break; } case HloOpcode::kRng: { @@ -1150,8 +1211,8 @@ bool HloParser::ParseInstruction(HloComputation::Builder* builder, return false; } instruction = builder->AddInstruction(HloInstruction::CreateDomain( - shape, operands[0], std::move(domain.entry_metadata), - std::move(domain.exit_metadata))); + shape, operands[0], std::move(domain.exit_metadata), + std::move(domain.entry_metadata))); break; } case HloOpcode::kTrace: @@ -1558,7 +1619,7 @@ bool HloParser::ParseTupleLiteral(std::unique_ptr* literal, } } } - *literal = Literal::MakeTupleOwned(std::move(elements)); + *literal = LiteralUtil::MakeTupleOwned(std::move(elements)); return ParseToken(TokKind::kRparen, StrCat("expects ')' at the end of the tuple with ", ShapeUtil::TupleElementCount(shape), "elements")); @@ -1586,8 +1647,8 @@ bool HloParser::ParseDenseLiteral(std::unique_ptr* literal, } // Create a literal with the given shape in default layout. - *literal = Literal::CreateFromDimensions(shape.element_type(), - AsInt64Slice(shape.dimensions())); + *literal = LiteralUtil::CreateFromDimensions( + shape.element_type(), AsInt64Slice(shape.dimensions())); tensorflow::int64 nest_level = 0; tensorflow::int64 linear_index = 0; // elems_seen_per_dim[i] is how many elements or sub-arrays we have seen for diff --git a/tensorflow/compiler/xla/service/hlo_parser_test.cc b/tensorflow/compiler/xla/service/hlo_parser_test.cc index d481e07f60a0747ae5bd6217aaeeb25d6fe733e1..f06c705c428a97a1d760419964b3d63d2c3228f5 100644 --- a/tensorflow/compiler/xla/service/hlo_parser_test.cc +++ b/tensorflow/compiler/xla/service/hlo_parser_test.cc @@ -277,12 +277,13 @@ ENTRY %WhileWithScalarS32Result.v2 () -> s32[] { "SendRecv", R"(HloModule TwoSendRecvBothWayRecvFist_module -ENTRY %TwoSendRecvBothWayRecvFist.v3 () -> f32[] { - %recv = (f32[], u32[]) recv(), channel_id=15, sharding={maximal device=1} - ROOT %recv-done = f32[] recv-done((f32[], u32[]) %recv), channel_id=15, sharding={maximal device=1} +ENTRY %TwoSendRecvBothWayRecvFist.v3 () -> (f32[], token[]) { + %token = token[] after-all() + %recv = (f32[], u32[], token[]) recv(token[] %token), channel_id=15, sharding={maximal device=1} + ROOT %recv-done = (f32[], token[]) recv-done((f32[], u32[], token[]) %recv), channel_id=15, sharding={maximal device=1} %constant = f32[] constant(2.1), sharding={maximal device=0} - %send = (f32[], u32[]) send(f32[] %constant), channel_id=16, sharding={maximal device=0}, control-predecessors={%recv} - %send-done = () send-done((f32[], u32[]) %send), channel_id=16, sharding={maximal device=0} + %send = (f32[], u32[], token[]) send(f32[] %constant, token[] %token), channel_id=16, sharding={maximal device=0}, control-predecessors={%recv} + %send-done = token[] send-done((f32[], u32[], token[]) %send), channel_id=16, sharding={maximal device=0} } )" @@ -795,10 +796,14 @@ ENTRY ReduceR3ToR2.v3 { R"(HloModule outfeed_module ENTRY InfeedToOutfeed { - infeed = (u32[3]{0}, pred[]) infeed() - outfeed = () outfeed(infeed) - ROOT infeed.1 = (u32[3]{0}, pred[]) infeed() - outfeed.1 = () outfeed(infeed.1) + token = token[] after-all() + infeed = ((u32[3]{0}, pred[]), token[]) infeed(token) + infeed.data = (u32[3]{0}, pred[]) get-tuple-element(infeed), index=0 + outfeed = token[] outfeed(infeed.data, token) + ROOT infeed.1 = ((u32[3]{0}, pred[]), token[]) infeed(token) + infeed.1.data = (u32[3]{0}, pred[]) get-tuple-element(infeed.1), index=0 + infeed.1.token = token[] get-tuple-element(infeed.1), index=1 + outfeed.1 = token[] outfeed(infeed.1.data, infeed.1.token) } )" @@ -826,6 +831,56 @@ ENTRY ReducePrecision { ROOT reduce-precision = f32[1]{0} reduce-precision(constant), exponent_bits=8, mantissa_bits=10 } +)" +}, +// Sort (Key) +{ +"SortKey", +R"(HloModule sort + +ENTRY Sort { + x = f32[1024]{0} parameter(0) + ROOT sorted = f32[1024]{0} sort(x), dimensions={0} +} + +)" +}, +// Sort (Key, Value) +{ +"SortKeyValue", +R"(HloModule sort + +ENTRY Sort { + keys = f32[1024]{0} parameter(0) + values = s32[1024]{0} parameter(1) + ROOT sorted = (f32[1024]{0}, s32[1024]{0}) sort(keys, values), dimensions={0} +} + +)" +}, +// R2 Sort (Key) +{ +"SortKeyR2", +R"(HloModule sort + +ENTRY Sort { + x = f32[1024,16]{0,1} parameter(0) + ROOT sorted = f32[1024,16]{0,1} sort(x), dimensions={0} +} + +)" +}, +// R2 Sort (Key, Value) +{ +"SortKeyValueR2", +R"(HloModule sort + +ENTRY Sort { + keys = f32[1024,16]{0,1} parameter(0) + values = s32[1024,16]{0,1} parameter(1) + ROOT sorted = (f32[1024,16]{0,1}, s32[1024,16]{0,1}) sort(keys, values), dimensions={0} +} + )" }, // Conditional @@ -1192,11 +1247,12 @@ TEST_F(HloParserTest, UnexpectedAttribute) { const string original = R"(HloModule unexpected_attr_module ENTRY %TwoSendRecvBothWayRecvFist.v3 () -> f32[] { - %recv = (f32[], u32[]) recv(), channel_id=15 - %recv-done = f32[] recv-done((f32[], u32[]) %recv), channel_id=15 + %token = token[] after-all() + %recv = (f32[], u32[], token[]) recv(token[] %token), channel_id=15 + %recv-done = (f32[], token[]) recv-done((f32[], u32[], token[]) %recv), channel_id=15 ROOT %constant = f32[] constant(2.1) - %send = (f32[], u32[]) send(f32[] %constant), channel_id=16, calls=%recv - %send-done = () send-done((f32[], u32[]) %send), channel_id=16 + %send = (f32[], u32[], token[]) send(f32[] %constant, token[] %token), channel_id=16, calls=%recv + %send-done = token[] send-done((f32[], u32[], token[]) %send), channel_id=16 } )"; @@ -1208,11 +1264,12 @@ TEST_F(HloParserTest, MissingAttribute) { const string original = R"(HloModule missing_attr_module ENTRY %TwoSendRecvBothWayRecvFist.v3 () -> f32[] { - %recv = (f32[], u32[]) recv(), channel_id=15 - %recv-done = f32[] recv-done((f32[], u32[]) %recv), channel_id=15 + %token = token[] after-all() + %recv = (f32[], u32[], token[]) recv(token[] %token), channel_id=15 + %recv-done = (f32[], token[]) recv-done((f32[], u32[], token[]) %recv), channel_id=15 ROOT %constant = f32[] constant(-2.1) - %send = (f32[], u32[]) send(f32[] %constant) - %send-done = () send-done((f32[], u32[]) %send), channel_id=16 + %send = (f32[], u32[], token[]) send(f32[] %constant, token[] %token) + %send-done = token[] send-done((f32[], u32[], token[]) %send), channel_id=16 } )"; @@ -1224,11 +1281,12 @@ TEST_F(HloParserTest, PredecessorUndefined) { const string original = R"(HloModule pre_not_found_module ENTRY %TwoSendRecvBothWayRecvFist.v3 () -> f32[] { - %recv = (f32[], u32[]) recv(), channel_id=15 - %recv-done = f32[] recv-done((f32[], u32[]) %recv), channel_id=15 + %token = token[] after-all() + %recv = (f32[], u32[], token[]) recv(token[] %token), channel_id=15 + %recv-done = (f32[], token[]) recv-done((f32[], u32[], token[]) %recv), channel_id=15 ROOT %constant = f32[] constant(2.1) - %send = (f32[], u32[]) send(f32[] %constant), channel_id=16, control-predecessors={%done} - %send-done = () send-done((f32[], u32[]) %send), channel_id=16 + %send = (f32[], u32[], token[]) send(f32[] %constant, token[] %token), channel_id=16, control-predecessors={%done} + %send-done = token[] send-done((f32[], u32[], token[]) %send), channel_id=16 } )"; @@ -1418,5 +1476,15 @@ TEST_F(HloParserTest, ParseConvolutionDimensionNumbers) { EXPECT_EQ(original, ConvolutionDimensionNumbersToString(dnums)); } +TEST_F(HloParserTest, NontupleInfeed) { + const string original = R"(HloModule nontuple_infeed: +ENTRY nontuple_infeed { + token = token[] after-all() + ROOT infeed = pred[] infeed(token) +})"; + ExpectHasSubstr(ParseHloString(original).status().error_message(), + "infeed must have a non-empty tuple shape"); +} + } // namespace } // namespace xla diff --git a/tensorflow/compiler/xla/service/hlo_query.cc b/tensorflow/compiler/xla/service/hlo_query.cc index 2418c19f3de7b036d7ef52d3a6db11de6316203b..2a07b6fcbc243d955e136ccdf097c8155a115845 100644 --- a/tensorflow/compiler/xla/service/hlo_query.cc +++ b/tensorflow/compiler/xla/service/hlo_query.cc @@ -15,7 +15,7 @@ limitations under the License. #include "tensorflow/compiler/xla/service/hlo_query.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/service/hlo_opcode.h" #include "tensorflow/compiler/xla/shape_util.h" diff --git a/tensorflow/compiler/xla/service/hlo_reachability_test.cc b/tensorflow/compiler/xla/service/hlo_reachability_test.cc index 657a9ee83d29e72b95660325f9139f44159d6508..585c95972b0e01abc14543205af71b4b0c0bdf3c 100644 --- a/tensorflow/compiler/xla/service/hlo_reachability_test.cc +++ b/tensorflow/compiler/xla/service/hlo_reachability_test.cc @@ -39,15 +39,15 @@ TEST_F(HloReachabilityTest, Reachability) { */ auto builder = HloComputation::Builder(TestName()); auto a = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(0.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(0.0f))); auto b = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(0.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(0.0f))); auto c = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(0.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(0.0f))); auto d = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(0.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(0.0f))); auto e = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(0.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(0.0f))); builder.Build(); HloReachabilityMap reachability({a, b, c, d, e}); diff --git a/tensorflow/compiler/xla/service/hlo_rematerialization.cc b/tensorflow/compiler/xla/service/hlo_rematerialization.cc index 62c07d7fac93618a83b3b6111aec1e93309a0761..59a8800a7d6e9417c0e561db45341c912ad20464 100644 --- a/tensorflow/compiler/xla/service/hlo_rematerialization.cc +++ b/tensorflow/compiler/xla/service/hlo_rematerialization.cc @@ -1244,7 +1244,7 @@ StatusOr HloRematerialization::Run( // TODO(b/80249101): Instead of a separate copy elision pass, use the // ordering from the HLO schedule directly for copy insertion. SequentialHloOrdering ordering(module, *sequence); - TF_RETURN_IF_ERROR(RemoveUnnecessaryCopies(ordering, {}, module)); + TF_RETURN_IF_ERROR(RemoveUnnecessaryCopies(ordering, module)); } // Compute peak memory usage of all computations in the module called in a diff --git a/tensorflow/compiler/xla/service/hlo_rematerialization_test.cc b/tensorflow/compiler/xla/service/hlo_rematerialization_test.cc index 7a46da6efe0df23129d56e16355cf66aceb68ffe..cd131147e619003d7ff4888e0f9e4da586bc2e76 100644 --- a/tensorflow/compiler/xla/service/hlo_rematerialization_test.cc +++ b/tensorflow/compiler/xla/service/hlo_rematerialization_test.cc @@ -132,7 +132,7 @@ class HloRematerializationTest : public HloTestBase { builder.AddInstruction( HloInstruction::CreateParameter(0, vec1_shape_, "param")); builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(true))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(true))); return builder.Build(); } @@ -226,7 +226,7 @@ TEST_F(HloRematerializationTest, RematerializeAroundWhile) { cond_builder.AddInstruction( HloInstruction::CreateParameter(0, vec1_shape_, "param")); cond_builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(true))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(true))); HloComputation* while_cond = module->AddEmbeddedComputation(cond_builder.Build()); @@ -263,7 +263,7 @@ TEST_F(HloRematerializationTest, RematerializeEntryAndWhileBody) { cond_builder.AddInstruction( HloInstruction::CreateParameter(0, vec1_shape_, "param")); cond_builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(true))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(true))); HloComputation* while_cond = module->AddEmbeddedComputation(cond_builder.Build()); @@ -296,7 +296,7 @@ TEST_F(HloRematerializationTest, RematerializeNestedComputations) { cond_builder.AddInstruction( HloInstruction::CreateParameter(0, vec1_shape_, "param")); cond_builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(true))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(true))); HloComputation* while_cond = module->AddEmbeddedComputation(cond_builder.Build()); diff --git a/tensorflow/compiler/xla/service/hlo_scheduling_test.cc b/tensorflow/compiler/xla/service/hlo_scheduling_test.cc index 73f22f81f4e9cf597db8b184642acff2fdaaf2b0..cf9ceed5b2fb49eb91fea96d89c8e1efc2a3dad1 100644 --- a/tensorflow/compiler/xla/service/hlo_scheduling_test.cc +++ b/tensorflow/compiler/xla/service/hlo_scheduling_test.cc @@ -168,8 +168,9 @@ TEST_F(HloSchedulingTest, ListAccountsForSubcomputations) { auto cond_builder = HloComputation::Builder("WhileCond"); HloInstruction* cond_param = cond_builder.AddInstruction( HloInstruction::CreateParameter(0, r1f32, "cond_param")); - HloInstruction* zero_vector = cond_builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR2({{0, 0, 0, 0}}))); + HloInstruction* zero_vector = + cond_builder.AddInstruction(HloInstruction::CreateConstant( + LiteralUtil::CreateR2({{0, 0, 0, 0}}))); cond_builder.AddInstruction(HloInstruction::CreateBinary( ShapeUtil::MakeShape(PRED, {}), HloOpcode::kNe, cond_param, zero_vector)); auto cond_computation = module->AddEmbeddedComputation(cond_builder.Build()); @@ -179,16 +180,18 @@ TEST_F(HloSchedulingTest, ListAccountsForSubcomputations) { auto body_builder = HloComputation::Builder("WhileBody"); HloInstruction* body_param = body_builder.AddInstruction( HloInstruction::CreateParameter(0, r1f32, "body_param")); - HloInstruction* one_vector = body_builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR2({{1, 1, 1, 1}}))); + HloInstruction* one_vector = + body_builder.AddInstruction(HloInstruction::CreateConstant( + LiteralUtil::CreateR2({{1, 1, 1, 1}}))); body_builder.AddInstruction(HloInstruction::CreateBinary( r1f32, HloOpcode::kSubtract, body_param, one_vector)); auto body_computation = module->AddEmbeddedComputation(body_builder.Build()); // transpose(matrix) + bcast(while) auto builder = HloComputation::Builder(TestName()); - HloInstruction* while_init = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR2({{1, 1, 1, 1}}))); + HloInstruction* while_init = + builder.AddInstruction(HloInstruction::CreateConstant( + LiteralUtil::CreateR2({{1, 1, 1, 1}}))); // Creates 16 bytes, ignoring subcomputations HloInstruction* while_loop = builder.AddInstruction(HloInstruction::CreateWhile( @@ -199,7 +202,7 @@ TEST_F(HloSchedulingTest, ListAccountsForSubcomputations) { HloInstruction::CreateBroadcast(r2f32, while_loop, {0})); HloInstruction* matrix = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR2( + HloInstruction::CreateConstant(LiteralUtil::CreateR2( {{1.0, 2.0, 3.0, 4.0}, {1.0, 2.0, 3.0, 4.0}}))); // Creates 32 bytes HloInstruction* transpose = builder.AddInstruction( @@ -257,7 +260,7 @@ TEST_F(HloSchedulingTest, TuplesAreAccountedCorrectly) { // Wrap lit in abs because constants are considered free by // IgnoreInstruction, and it skews the accounting. auto lit = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR1({1, 1, 1, 1, 1, 1}))); + LiteralUtil::CreateR1({1, 1, 1, 1, 1, 1}))); auto abs_const = builder.AddInstruction( HloInstruction::CreateUnary(r1f32, HloOpcode::kAbs, lit)); @@ -300,11 +303,11 @@ TEST_F(HloSchedulingTest, MultiOutputFusionAccountedCorrectly) { HloComputation::Builder builder(TestName()); auto c1 = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR1({1, 1, 1, 1, 1}))); + LiteralUtil::CreateR1({1, 1, 1, 1, 1}))); auto c2 = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR1({1, 2, 3, 4, 5}))); + LiteralUtil::CreateR1({1, 2, 3, 4, 5}))); auto c3 = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR1({0, 2, 4, 6, 8}))); + LiteralUtil::CreateR1({0, 2, 4, 6, 8}))); auto add = builder.AddInstruction( HloInstruction::CreateBinary(r1f32, HloOpcode::kAdd, c1, c2)); @@ -354,8 +357,9 @@ TEST_F(HloSchedulingTest, HeapSimulatorAccountsForSubcomputations) { auto cond_builder = HloComputation::Builder("WhileCond"); HloInstruction* cond_param = cond_builder.AddInstruction( HloInstruction::CreateParameter(0, r1f32, "cond_param")); - HloInstruction* zero_vector = cond_builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR2({{0, 0, 0, 0}}))); + HloInstruction* zero_vector = + cond_builder.AddInstruction(HloInstruction::CreateConstant( + LiteralUtil::CreateR2({{0, 0, 0, 0}}))); cond_builder.AddInstruction(HloInstruction::CreateBinary( ShapeUtil::MakeShape(PRED, {}), HloOpcode::kNe, cond_param, zero_vector)); auto cond_computation = module->AddEmbeddedComputation(cond_builder.Build()); @@ -365,15 +369,17 @@ TEST_F(HloSchedulingTest, HeapSimulatorAccountsForSubcomputations) { auto body_builder = HloComputation::Builder("WhileBody"); HloInstruction* body_param = body_builder.AddInstruction( HloInstruction::CreateParameter(0, r1f32, "body_param")); - HloInstruction* one_vector = body_builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR2({{1, 1, 1, 1}}))); + HloInstruction* one_vector = + body_builder.AddInstruction(HloInstruction::CreateConstant( + LiteralUtil::CreateR2({{1, 1, 1, 1}}))); body_builder.AddInstruction(HloInstruction::CreateBinary( r1f32, HloOpcode::kSubtract, body_param, one_vector)); auto body_computation = module->AddEmbeddedComputation(body_builder.Build()); auto builder = HloComputation::Builder(TestName()); - HloInstruction* while_init = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR2({{1, 1, 1, 1}}))); + HloInstruction* while_init = + builder.AddInstruction(HloInstruction::CreateConstant( + LiteralUtil::CreateR2({{1, 1, 1, 1}}))); // Creates 16 bytes, ignoring subcomputations builder.AddInstruction(HloInstruction::CreateWhile( r1f32, cond_computation, body_computation, while_init)); diff --git a/tensorflow/compiler/xla/service/hlo_sharding.cc b/tensorflow/compiler/xla/service/hlo_sharding.cc index 268b4727bcbed42ba71526f1d5ef5c887e941930..393944c20faa0b09ebc8544543b62566c836739f 100644 --- a/tensorflow/compiler/xla/service/hlo_sharding.cc +++ b/tensorflow/compiler/xla/service/hlo_sharding.cc @@ -60,6 +60,9 @@ HloSharding HloSharding::Tuple( const Shape& tuple_shape, tensorflow::gtl::ArraySlice shardings) { CHECK(ShapeUtil::IsTuple(tuple_shape)) << ShapeUtil::HumanString(tuple_shape); + for (auto& sharding : shardings) { + CHECK(!sharding.IsTuple()) << sharding.ToString(); + } std::vector flattened_list(shardings.begin(), shardings.end()); CHECK_EQ(flattened_list.size(), RequiredLeaves(tuple_shape)) << "Flat list has " << flattened_list.size() << ", required " @@ -67,6 +70,24 @@ HloSharding HloSharding::Tuple( return HloSharding(flattened_list); } +HloSharding HloSharding::SingleTuple(const Shape& tuple_shape, + const HloSharding& sharding) { + CHECK(ShapeUtil::IsTuple(tuple_shape)) << ShapeUtil::HumanString(tuple_shape); + CHECK(!sharding.IsTuple()) << sharding.ToString(); + int64 leaf_count = ShapeUtil::GetLeafCount(tuple_shape); + std::vector flattened_list; + flattened_list.reserve(leaf_count); + for (int64 i = 0; i < leaf_count; ++i) { + flattened_list.push_back(sharding); + } + return HloSharding(flattened_list); +} + +HloSharding HloSharding::Single(const Shape& shape, + const HloSharding& sharding) { + return ShapeUtil::IsTuple(shape) ? SingleTuple(shape, sharding) : sharding; +} + string HloSharding::ToString() const { if (IsTuple()) { std::vector parts; diff --git a/tensorflow/compiler/xla/service/hlo_sharding.h b/tensorflow/compiler/xla/service/hlo_sharding.h index 34324d2058efe804cda486600dabd8a62cb84fda..6f672b0f28d2b85411d70f33da9a9f270aefc0d0 100644 --- a/tensorflow/compiler/xla/service/hlo_sharding.h +++ b/tensorflow/compiler/xla/service/hlo_sharding.h @@ -24,7 +24,7 @@ limitations under the License. #include #include "tensorflow/compiler/xla/array.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/protobuf_util.h" #include "tensorflow/compiler/xla/shape_tree.h" #include "tensorflow/compiler/xla/xla_data.pb.h" @@ -80,6 +80,15 @@ class HloSharding { static HloSharding Tuple(const Shape& tuple_shape, tensorflow::gtl::ArraySlice shardings); + // Creates a new sharding for a tuple type, with a single input sharding + // repeated on each leaf. + static HloSharding SingleTuple(const Shape& tuple_shape, + const HloSharding& sharding); + + // If shape is an array, returns sharding, otherwise returns the tuple shaped + // sharding with all the leaf nodes having the same input sharding. + static HloSharding Single(const Shape& shape, const HloSharding& sharding); + // Create a new sharding from a protobuf OpSharding. static StatusOr FromProto(const OpSharding& proto); diff --git a/tensorflow/compiler/xla/service/hlo_sharding_metadata.cc b/tensorflow/compiler/xla/service/hlo_sharding_metadata.cc index 748273a43cecca7a9c7392bb84f0e4c7133cfb14..4f91d619efd08d46a122aa1bac3cf9a37f6b5485 100644 --- a/tensorflow/compiler/xla/service/hlo_sharding_metadata.cc +++ b/tensorflow/compiler/xla/service/hlo_sharding_metadata.cc @@ -88,6 +88,12 @@ std::vector LocatePassThroughDomainLinks( VLOG(2) << " " << instruction->ToString(); } } + if (instruction == instruction->parent()->root_instruction()) { + pass_through.emplace_back(nullptr, instruction); + VLOG(2) << "Found passthrough domain link:"; + VLOG(2) << " "; + VLOG(2) << " " << instruction->ToString(); + } } return pass_through; } @@ -101,8 +107,12 @@ Status FixupPassThroughDomainLinks(const DomainMetadata::Domain& domain, HloInstruction::CreateGetTupleElement(pass_through.operand->shape(), tuple, 0)); gte->set_sharding(sharding); - TF_RETURN_IF_ERROR( - pass_through.operand->ReplaceUseWith(pass_through.user, gte)); + if (pass_through.user != nullptr) { + TF_RETURN_IF_ERROR( + pass_through.operand->ReplaceUseWith(pass_through.user, gte)); + } else { + pass_through.operand->parent()->set_root_instruction(gte); + } } return Status::OK(); } @@ -377,7 +387,7 @@ bool ShardingMetadata::Matches(const DomainMetadata& other) const { } string ShardingMetadata::ToString() const { - return sharding_ != nullptr ? sharding_->ToString() : "None"; + return sharding_ != nullptr ? sharding_->ToString() : "{}"; } Status ShardingMetadata::NormalizeInstructions( diff --git a/tensorflow/compiler/xla/service/hlo_sharding_test.cc b/tensorflow/compiler/xla/service/hlo_sharding_test.cc index 54b7402b866361748d9eb35182b0bf486c4c9bdc..7baa927d0e2b1abbbb2333633d16dd605ae8c8ef 100644 --- a/tensorflow/compiler/xla/service/hlo_sharding_test.cc +++ b/tensorflow/compiler/xla/service/hlo_sharding_test.cc @@ -18,7 +18,7 @@ limitations under the License. #include #include -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/service/hlo_parser.h" #include "tensorflow/compiler/xla/shape_util.h" #include "tensorflow/compiler/xla/test.h" diff --git a/tensorflow/compiler/xla/service/hlo_subcomputation_unification_test.cc b/tensorflow/compiler/xla/service/hlo_subcomputation_unification_test.cc index 7b601f9a9578cfa6b293cf7f002255f7db8b1257..45c684d66752862eec301b8943d350804f070309 100644 --- a/tensorflow/compiler/xla/service/hlo_subcomputation_unification_test.cc +++ b/tensorflow/compiler/xla/service/hlo_subcomputation_unification_test.cc @@ -75,7 +75,7 @@ TEST_F(HloSubcomputationUnificationTest, UnifyIdentities) { module->AddEmbeddedComputation(CreateR0S32IdentityComputation()); auto constant = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(5))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(5))); auto x = builder.AddInstruction( HloInstruction::CreateCall(r0s32_, {constant}, callee1)); auto y = builder.AddInstruction( @@ -112,9 +112,9 @@ TEST_F(HloSubcomputationUnificationTest, UnifyAdditions) { module->AddEmbeddedComputation(CreateR0S32AdditionComputation()); auto constant1 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(5))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(5))); auto constant2 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(3))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(3))); auto x = builder.AddInstruction( HloInstruction::CreateCall(r0s32_, {constant1, constant2}, callee1)); auto y = builder.AddInstruction( diff --git a/tensorflow/compiler/xla/service/hlo_tfgraph_builder.cc b/tensorflow/compiler/xla/service/hlo_tfgraph_builder.cc index 3dc733940fc89952bd5e75a9b28d9cbf356f8000..48f676db85ab5e7711d9e9ac900306a9ea85ef10 100644 --- a/tensorflow/compiler/xla/service/hlo_tfgraph_builder.cc +++ b/tensorflow/compiler/xla/service/hlo_tfgraph_builder.cc @@ -15,7 +15,7 @@ limitations under the License. #include "tensorflow/compiler/xla/service/hlo_tfgraph_builder.h" #include "tensorflow/compiler/xla/layout_util.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/service/hlo_opcode.h" #include "tensorflow/compiler/xla/shape_util.h" #include "tensorflow/core/framework/attr_value.pb.h" diff --git a/tensorflow/compiler/xla/service/hlo_tfgraph_builder_test.cc b/tensorflow/compiler/xla/service/hlo_tfgraph_builder_test.cc index be156d765dc10d54eaf301e90883babbc5693e28..1e2b31a1f2bb4865faafc3d14e2b194e3aa171a1 100644 --- a/tensorflow/compiler/xla/service/hlo_tfgraph_builder_test.cc +++ b/tensorflow/compiler/xla/service/hlo_tfgraph_builder_test.cc @@ -90,7 +90,7 @@ TEST_F(HloTfGraphBuilderTest, CheckConcatenateDimsAndShapes) { TEST_F(HloTfGraphBuilderTest, CheckScalarValue) { auto builder = HloComputation::Builder("Const"); HloInstruction *instruction = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(123))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(123))); OpMetadata metadata; metadata.set_op_name("x"); metadata.set_op_type("y"); diff --git a/tensorflow/compiler/xla/service/hlo_value.cc b/tensorflow/compiler/xla/service/hlo_value.cc index 7b27dbfec376b8ba16d00285f10e2cc291e07a61..4e3c9df3a036890ce25f5b14603d275263e8659b 100644 --- a/tensorflow/compiler/xla/service/hlo_value.cc +++ b/tensorflow/compiler/xla/service/hlo_value.cc @@ -125,7 +125,7 @@ bool MayUseOperandValue(int64 operand_number, const ShapeIndex& index, // transparently. CHECK_EQ(operand_number, 0); return index.empty(); - case HloOpcode::kSelect: + case HloOpcode::kTupleSelect: // Select does not use any nested elements of its selected-from operands // (operand 1 and 2) CHECK_GE(operand_number, 0); diff --git a/tensorflow/compiler/xla/service/hlo_verifier.cc b/tensorflow/compiler/xla/service/hlo_verifier.cc index 1d6cd4cb2308fd09c7511e390a146a5224f253a3..48eeba6afdcb227d4c714e1497f66f6334cb38ca 100644 --- a/tensorflow/compiler/xla/service/hlo_verifier.cc +++ b/tensorflow/compiler/xla/service/hlo_verifier.cc @@ -15,6 +15,8 @@ limitations under the License. #include +#include "tensorflow/compiler/xla/service/hlo_casting_utils.h" +#include "tensorflow/compiler/xla/service/hlo_instructions.h" #include "tensorflow/compiler/xla/service/hlo_opcode.h" #include "tensorflow/compiler/xla/service/hlo_verifier.h" #include "tensorflow/compiler/xla/status_macros.h" @@ -39,6 +41,10 @@ Status ShapeVerifier::HandleSelect(HloInstruction* select) { return CheckTernaryShape(select); } +Status ShapeVerifier::HandleTupleSelect(HloInstruction* tuple_select) { + return CheckTernaryShape(tuple_select); +} + Status ShapeVerifier::HandleConcatenate(HloInstruction* concatenate) { std::vector operand_shapes; for (const HloInstruction* operand : concatenate->operands()) { @@ -106,22 +112,73 @@ Status ShapeVerifier::HandleReducePrecision(HloInstruction* reduce_precision) { reduce_precision->mantissa_bits())); } -Status ShapeVerifier::HandleInfeed(HloInstruction*) { return Status::OK(); } +namespace { + +Status CheckIsTokenOperand(const HloInstruction* instruction, + int64 operand_no) { + const HloInstruction* token = instruction->operand(operand_no); + if (!ShapeUtil::Equal(token->shape(), ShapeUtil::MakeTokenShape())) { + return InternalError( + "Expected operand %lld to be token-shaped, actual shape is" + "%s:\n%s", + operand_no, ShapeUtil::HumanString(token->shape()).c_str(), + instruction->ToString().c_str()); + } + return Status::OK(); +} + +Status CheckOperandAndParameter(const HloInstruction* instruction, + int64 operand_number, + const HloComputation* computation, + int64 parameter_number) { + const HloInstruction* operand = instruction->operand(operand_number); + const HloInstruction* parameter = + computation->parameter_instruction(parameter_number); + if (!ShapeUtil::Compatible(operand->shape(), parameter->shape())) { + return InternalError("Operand %s shape does not match parameter's %s in %s", + operand->ToString().c_str(), + parameter->ToString().c_str(), + instruction->ToString().c_str()); + } + return Status::OK(); +} + +} // namespace + +Status ShapeVerifier::HandleInfeed(HloInstruction* instruction) { + HloInfeedInstruction* infeed = Cast(instruction); + // Infeed has an optional single token operand. + // TODO(b/80000000): Update when token is not optional. + if (infeed->operand_count() == 1) { + TF_RETURN_IF_ERROR(CheckIsTokenOperand(instruction, 0)); + } + + // The output of infeed is a tuple containing the data value and a token. + return CheckShape(infeed, + ShapeUtil::MakeTupleShape( + {infeed->infeed_shape(), ShapeUtil::MakeTokenShape()})); +} + +Status ShapeVerifier::HandleOutfeed(HloInstruction* instruction) { + HloOutfeedInstruction* outfeed = Cast(instruction); + // Outfeed has an optional token operand (operand 1). + // TODO(b/80000000): Update when token is not optional. + if (outfeed->operand_count() == 2) { + TF_RETURN_IF_ERROR(CheckIsTokenOperand(instruction, 1)); + } -Status ShapeVerifier::HandleOutfeed(HloInstruction* outfeed) { // Outfeed has a separate shape field for the value which is outfed to the - // host. The shape of the instruction itself is always nil because the outfeed - // produces no HLO value in the graph. + // host. The shape of the instruction itself is always a token. if (!ShapeUtil::Compatible(outfeed->outfeed_shape(), outfeed->operand(0)->shape())) { return InternalError( - "Expected outfeed to have shape compatible with operand's shape %s, " + "Expected outfeed shape to be compatible with operand's shape %s, " "actual shape is %s:\n%s", ShapeUtil::HumanString(outfeed->operand(0)->shape()).c_str(), ShapeUtil::HumanString(outfeed->outfeed_shape()).c_str(), outfeed->ToString().c_str()); } - return CheckShape(outfeed, ShapeUtil::MakeNil()); + return CheckShape(outfeed, ShapeUtil::MakeTokenShape()); } Status ShapeVerifier::HandleHostCompute(HloInstruction*) { @@ -137,7 +194,16 @@ Status ShapeVerifier::HandleReverse(HloInstruction* reverse) { } Status ShapeVerifier::HandleSort(HloInstruction* sort) { - return CheckUnaryShape(sort); + if (sort->operand_count() == 2 && + !ShapeUtil::SameDimensions(sort->operand(0)->shape(), + sort->operand(1)->shape())) { + return InternalError( + "Expected sort to have to have the same dimensions for the keys and " + "the values. Keys shape is: %s\n, Values shape is: %s", + ShapeUtil::HumanString(sort->operand(0)->shape()).c_str(), + ShapeUtil::HumanString(sort->operand(1)->shape()).c_str()); + } + return CheckVariadicShape(sort); } Status ShapeVerifier::HandleConstant(HloInstruction* constant) { @@ -203,8 +269,11 @@ Status ShapeVerifier::HandleParameter(HloInstruction* hlo) { Status ShapeVerifier::HandleFusion(HloInstruction*) { return Status::OK(); } Status ShapeVerifier::HandleCall(HloInstruction* call) { + for (int64 i = 0; i < call->to_apply()->num_parameters(); ++i) { + TF_RETURN_IF_ERROR(CheckOperandAndParameter(call, i, call->to_apply(), i)); + } // The shape of kCall should match the shape of the computation it calls. - return CheckShape(call, call->to_apply()->ComputeProgramShape().result()); + return CheckShape(call, call->to_apply()->root_instruction()->shape()); } Status ShapeVerifier::HandleCustomCall(HloInstruction*) { return Status::OK(); } @@ -273,19 +342,37 @@ Status ShapeVerifier::HandleSelectAndScatter(HloInstruction* instruction) { } Status ShapeVerifier::HandleWhile(HloInstruction* xla_while) { + TF_RETURN_IF_ERROR( + CheckOperandAndParameter(xla_while, 0, xla_while->while_body(), 0)); + TF_RETURN_IF_ERROR( + CheckOperandAndParameter(xla_while, 0, xla_while->while_condition(), 0)); + const Shape& conditional_shape = + xla_while->while_condition()->root_instruction()->shape(); + if (!ShapeUtil::Compatible(conditional_shape, + ShapeUtil::MakeShape(PRED, {}))) { + return InternalError( + "Conditional computation shape does not lead to a scalar predicate " + "shape: %s", + ShapeUtil::HumanString(conditional_shape).c_str()); + } // The shape of kWhile should match the shape of the body computation it // calls. return CheckShape(xla_while, - xla_while->while_body()->ComputeProgramShape().result()); + xla_while->while_body()->root_instruction()->shape()); } Status ShapeVerifier::HandleConditional(HloInstruction* conditional) { + TF_RETURN_IF_ERROR(CheckOperandAndParameter( + conditional, 1, conditional->true_computation(), 0)); + TF_RETURN_IF_ERROR(CheckOperandAndParameter( + conditional, 2, conditional->false_computation(), 0)); + TF_RETURN_IF_ERROR( + CheckShape(conditional, + conditional->true_computation()->root_instruction()->shape())); TF_RETURN_IF_ERROR(CheckShape( conditional, - conditional->true_computation()->ComputeProgramShape().result())); - return CheckShape( - conditional, - conditional->false_computation()->ComputeProgramShape().result()); + conditional->false_computation()->root_instruction()->shape())); + return Status::OK(); } Status ShapeVerifier::HandlePad(HloInstruction* pad) { @@ -299,9 +386,11 @@ Status ShapeVerifier::HandleSend(HloInstruction* send) { const HloInstruction* send_done = send->users().front(); TF_RET_CHECK(send_done->opcode() == HloOpcode::kSendDone); TF_RETURN_IF_ERROR(CheckSameChannel(send, send_done)); - return CheckShape( - send, ShapeUtil::MakeTupleShape( - {send->operand(0)->shape(), ShapeUtil::MakeShape(U32, {})})); + TF_RETURN_IF_ERROR(CheckIsTokenOperand(send, 1)); + return CheckShape(send, + ShapeUtil::MakeTupleShape({send->operand(0)->shape(), + ShapeUtil::MakeShape(U32, {}), + ShapeUtil::MakeTokenShape()})); } Status ShapeVerifier::HandleSendDone(HloInstruction* send_done) { @@ -309,7 +398,8 @@ Status ShapeVerifier::HandleSendDone(HloInstruction* send_done) { const HloInstruction* send = send_done->operand(0); TF_RET_CHECK(send->opcode() == HloOpcode::kSend); TF_RETURN_IF_ERROR(CheckSameChannel(send, send_done)); - return CheckShape(send_done, ShapeUtil::MakeNil()); + + return CheckShape(send_done, ShapeUtil::MakeTokenShape()); } Status ShapeVerifier::HandleRecv(HloInstruction* recv) { @@ -317,9 +407,11 @@ Status ShapeVerifier::HandleRecv(HloInstruction* recv) { const HloInstruction* recv_done = recv->users().front(); TF_RET_CHECK(recv_done->opcode() == HloOpcode::kRecvDone); TF_RETURN_IF_ERROR(CheckSameChannel(recv, recv_done)); - return CheckShape(recv, - ShapeUtil::MakeTupleShape( - {recv_done->shape(), ShapeUtil::MakeShape(U32, {})})); + TF_RETURN_IF_ERROR(CheckIsTokenOperand(recv, 0)); + return CheckShape( + recv, ShapeUtil::MakeTupleShape( + {ShapeUtil::GetTupleElementShape(recv_done->shape(), 0), + ShapeUtil::MakeShape(U32, {}), ShapeUtil::MakeTokenShape()})); } Status ShapeVerifier::HandleRecvDone(HloInstruction* recv_done) { @@ -327,7 +419,9 @@ Status ShapeVerifier::HandleRecvDone(HloInstruction* recv_done) { const HloInstruction* recv = recv_done->operand(0); TF_RET_CHECK(recv->opcode() == HloOpcode::kRecv); TF_RETURN_IF_ERROR(CheckSameChannel(recv, recv_done)); - return CheckShape(recv_done, recv->shape().tuple_shapes(0)); + return CheckShape(recv_done, + ShapeUtil::MakeTupleShape({recv->shape().tuple_shapes(0), + ShapeUtil::MakeTokenShape()})); } Status ShapeVerifier::HandleBatchNormTraining( @@ -386,6 +480,7 @@ Status CheckMixedPrecisionOperands(const HloInstruction* instruction) { case HloOpcode::kRecvDone: case HloOpcode::kReducePrecision: case HloOpcode::kSelect: + case HloOpcode::kTupleSelect: case HloOpcode::kSend: case HloOpcode::kSendDone: case HloOpcode::kTuple: @@ -426,13 +521,12 @@ Status ShapeVerifier::HandleGather(HloInstruction* gather) { gather->gather_dimension_numbers(), gather->gather_window_bounds())); } -Status ShapeVerifier::HandleGenerateToken(HloInstruction* token) { +Status ShapeVerifier::HandleAfterAll(HloInstruction* token) { std::vector operand_shapes; for (const HloInstruction* operand : token->operands()) { operand_shapes.push_back(&operand->shape()); } - return CheckShape(token, - ShapeInference::InferGenerateTokenShape(operand_shapes)); + return CheckShape(token, ShapeInference::InferAfterAllShape(operand_shapes)); } Status ShapeVerifier::CheckShape(const HloInstruction* instruction, @@ -449,16 +543,10 @@ Status ShapeVerifier::CheckShape(const HloInstruction* instruction, // We treat BF16 and F32 as compatible types if mixed precision is allowed, // but only when the instruction defines the BF16/F32 buffer. switch (instruction->opcode()) { - case HloOpcode::kSelect: - if (ShapeUtil::IsTuple(inferred_shape) || !allow_mixed_precision_) { - // Select only defines the top-level buffer, which in this case is the - // tuple, so we cannot allow mixed precision. - compatible = - ShapeUtil::Compatible(instruction->shape(), inferred_shape); - } else { - compatible = ShapeUtil::CompatibleIgnoringFpPrecision( - instruction->shape(), inferred_shape); - } + case HloOpcode::kTupleSelect: + // TupleSelect only defines the top-level buffer, which in this case is + // the tuple, so we cannot allow mixed precision. + compatible = ShapeUtil::Compatible(instruction->shape(), inferred_shape); break; case HloOpcode::kGetTupleElement: case HloOpcode::kTuple: @@ -751,33 +839,23 @@ Status HloVerifier::CheckWhileInstruction(HloInstruction* instruction) { "While loop must have exactly one operand; had %lld : %s", instruction->operand_count(), instruction->ToString().c_str()); } - auto* init = instruction->operand(0); - auto* cond_param = while_cond->parameter_instruction(0); - if (!ShapeUtil::Compatible(init->shape(), cond_param->shape())) { - return FailedPrecondition( - "While condition's parameter must have the same shape as the " - "loop's 'init'. init: %s, param: %s", - init->ToString().c_str(), cond_param->ToString().c_str()); - } - auto* cond_root = while_cond->root_instruction(); - if (!ShapeUtil::Compatible(cond_root->shape(), - ShapeUtil::MakeShape(PRED, {}))) { - return FailedPrecondition("While condition should have shape PRED: %s", - cond_root->ToString().c_str()); - } - auto* body_param = while_body->parameter_instruction(0); - if (!ShapeUtil::Compatible(init->shape(), body_param->shape())) { + return Status::OK(); +} + +Status HloVerifier::CheckConditionalInstruction(HloInstruction* instruction) { + if (instruction->true_computation()->num_parameters() != 1) { return FailedPrecondition( - "While body's parameter must have the same shape as the loop's" - " 'init'. init: %s, param: %s", - init->ToString().c_str(), body_param->ToString().c_str()); + "True computation %s of %s must have 1 parameter insted of %lld", + instruction->true_computation()->name().c_str(), + instruction->ToString().c_str(), + instruction->true_computation()->num_parameters()); } - auto* body_root = while_body->root_instruction(); - if (!ShapeUtil::Compatible(init->shape(), body_root->shape())) { + if (instruction->false_computation()->num_parameters() != 1) { return FailedPrecondition( - "While body should have same shape as the loop's 'init'." - "init: %s, body: %s", - init->ToString().c_str(), body_root->ToString().c_str()); + "False computation %s of %s must have 1 parameter insted of %lld", + instruction->false_computation()->name().c_str(), + instruction->ToString().c_str(), + instruction->false_computation()->num_parameters()); } return Status::OK(); } @@ -786,8 +864,7 @@ Status HloVerifier::CheckElementwiseInstruction(HloInstruction* instruction) { const Shape& out_shape = instruction->shape(); for (HloInstruction* operand : instruction->operands()) { const Shape& operand_shape = operand->shape(); - if (!ShapeUtil::IsScalar(operand_shape) && - !ShapeUtil::CompatibleIgnoringElementType(operand_shape, out_shape)) { + if (!ShapeUtil::CompatibleIgnoringElementType(operand_shape, out_shape)) { return FailedPrecondition( "Implicit broadcast is not allowed in HLO." "Found non-compatible shapes for instruction %s.\n" @@ -815,9 +892,10 @@ bool ShapeContainsToken(const Shape& shape) { } // Verifies that all types entering and exiting the entry computation are -// legal. For example, TOKEN types have no Literal representation and cannot be -// on the interface of the entry computation (parameters and root instruction). +// legal. Status VerifyEntryAndExitShapes(const HloModule& module) { + // Tokens cannot be passed as entry parameters. + // TODO(b/80000000): Remove this constraint. for (int i = 0; i < module.entry_computation()->num_parameters(); ++i) { HloInstruction* param = module.entry_computation()->parameter_instruction(i); @@ -827,14 +905,6 @@ Status VerifyEntryAndExitShapes(const HloModule& module) { ShapeUtil::HumanString(param->shape()).c_str()); } } - if (ShapeContainsToken( - module.entry_computation()->root_instruction()->shape())) { - return InternalError( - "Entry root is or contains a token shape: %s", - ShapeUtil::HumanString( - module.entry_computation()->root_instruction()->shape()) - .c_str()); - } return Status::OK(); } @@ -881,7 +951,11 @@ StatusOr HloVerifier::Run(HloModule* module) { << " != " << ShapeUtil::Rank(instruction->operand(0)->shape()); } else if (instruction->opcode() == HloOpcode::kWhile) { TF_RETURN_IF_ERROR(CheckWhileInstruction(instruction)); - } else if (instruction->IsElementwise()) { + } else if (instruction->opcode() == HloOpcode::kConditional) { + TF_RETURN_IF_ERROR(CheckConditionalInstruction(instruction)); + } else if (instruction->opcode() != + HloOpcode::kRng /* Rng operands are always scalar. */ + && instruction->IsElementwise()) { TF_RETURN_IF_ERROR(CheckElementwiseInstruction(instruction)); } diff --git a/tensorflow/compiler/xla/service/hlo_verifier.h b/tensorflow/compiler/xla/service/hlo_verifier.h index 7283b3e7dcdbed5be18a1da1571287cf0c089288..9e62bdc8a9e6ee06c09132302dd8a74a1849a679 100644 --- a/tensorflow/compiler/xla/service/hlo_verifier.h +++ b/tensorflow/compiler/xla/service/hlo_verifier.h @@ -35,6 +35,7 @@ class ShapeVerifier : public DfsHloVisitor { Status HandleElementwiseBinary(HloInstruction* hlo) override; Status HandleClamp(HloInstruction* clamp) override; Status HandleSelect(HloInstruction* select) override; + Status HandleTupleSelect(HloInstruction* tuple_select) override; Status HandleConcatenate(HloInstruction* concatenate) override; Status HandleConvert(HloInstruction* convert) override; Status HandleBitcastConvert(HloInstruction* convert) override; @@ -81,7 +82,7 @@ class ShapeVerifier : public DfsHloVisitor { HloInstruction* batch_norm_inference) override; Status HandleBatchNormGrad(HloInstruction* batch_norm_grad) override; Status HandleGather(HloInstruction* gather) override; - Status HandleGenerateToken(HloInstruction* token) override; + Status HandleAfterAll(HloInstruction* token) override; Status FinishVisit(HloInstruction*) override { return Status::OK(); } @@ -145,6 +146,8 @@ class HloVerifier : public HloPassInterface { Status CheckWhileInstruction(HloInstruction* instruction); + Status CheckConditionalInstruction(HloInstruction* instruction); + // Checks that the non-scalar operand shapes are compatible to the output // shape, i.e., that there are no implicit broadcasts of size-one dimensions. Status CheckElementwiseInstruction(HloInstruction* instruction); diff --git a/tensorflow/compiler/xla/service/hlo_verifier_test.cc b/tensorflow/compiler/xla/service/hlo_verifier_test.cc index c92db0be14dceb32ea86521dcc99b8f63738e4a5..04c6ba3eeb92bad2b5b69f7f56e73e1f7a8148aa 100644 --- a/tensorflow/compiler/xla/service/hlo_verifier_test.cc +++ b/tensorflow/compiler/xla/service/hlo_verifier_test.cc @@ -21,6 +21,7 @@ limitations under the License. #include "tensorflow/compiler/xla/service/hlo_computation.h" #include "tensorflow/compiler/xla/service/hlo_instruction.h" #include "tensorflow/compiler/xla/service/hlo_opcode.h" +#include "tensorflow/compiler/xla/service/hlo_parser.h" #include "tensorflow/compiler/xla/shape_util.h" #include "tensorflow/compiler/xla/test.h" #include "tensorflow/compiler/xla/tests/hlo_test_base.h" @@ -123,5 +124,55 @@ TEST_F(HloVerifierTest, ResetsShapeVerifierState) { EXPECT_FALSE(verifier().Run(module.get()).status().ok()); } +TEST_F(HloVerifierTest, CheckCallOperandParameterShapesMismatch) { + const char* const hlo_string = R"( +HloModule Module + +callme { + ROOT param = (s32[], f32[4]) parameter(0) +} + +ENTRY entry { + p0 = (f32[4], s32[]) parameter(0) + ROOT mycall = (s32[], f32[4]) call(p0), to_apply=callme +} +)"; + TF_ASSERT_OK_AND_ASSIGN(auto module, ParseHloString(hlo_string)); + + auto status = verifier().Run(module.get()).status(); + ASSERT_FALSE(status.ok()); + EXPECT_THAT(status.error_message(), + HasSubstr("shape does not match parameter")); +} + +TEST_F(HloVerifierTest, CheckConditionalOperandParameterShapesMismatch) { + const char* const hlo_string = R"( +HloModule Module + +true_branch { + tparam = (s32[], f32[4]) parameter(0) + ROOT tgte1 = f32[4] get-tuple-element(tparam), index=1 +} + +false_branch { + fparam = (s32[], f32[4]) parameter(0) + ROOT fgte1 = f32[4] get-tuple-element(fparam), index=1 +} + +ENTRY entry { + p0 = (f32[4], s32[]) parameter(0) + constant = pred[] constant(true) + ROOT conditional = f32[4] conditional(constant, p0, p0), + true_computation=true_branch, false_computation=false_branch +} +)"; + TF_ASSERT_OK_AND_ASSIGN(auto module, ParseHloString(hlo_string)); + + auto status = verifier().Run(module.get()).status(); + ASSERT_FALSE(status.ok()); + EXPECT_THAT(status.error_message(), + HasSubstr("shape does not match parameter")); +} + } // namespace } // namespace xla diff --git a/tensorflow/compiler/xla/service/implicit_broadcast_remover_test.cc b/tensorflow/compiler/xla/service/implicit_broadcast_remover_test.cc index 8c7b38dd1bf73e0be7b669d7215812aaef1cee17..f85d31d5225b8012b68f851b2bfec219d736ba0d 100644 --- a/tensorflow/compiler/xla/service/implicit_broadcast_remover_test.cc +++ b/tensorflow/compiler/xla/service/implicit_broadcast_remover_test.cc @@ -15,7 +15,7 @@ limitations under the License. #include "tensorflow/compiler/xla/service/implicit_broadcast_remover.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/service/hlo_matchers.h" #include "tensorflow/compiler/xla/shape_util.h" #include "tensorflow/compiler/xla/tests/hlo_verified_test_base.h" diff --git a/tensorflow/compiler/xla/service/indexed_array_analysis.cc b/tensorflow/compiler/xla/service/indexed_array_analysis.cc index 1985d20578677ae68b244023c4640454b004bf49..8b2df3256776a7d77517daff1fe282b0dbde7045 100644 --- a/tensorflow/compiler/xla/service/indexed_array_analysis.cc +++ b/tensorflow/compiler/xla/service/indexed_array_analysis.cc @@ -19,6 +19,7 @@ limitations under the License. #include "tensorflow/compiler/xla/util.h" #include "tensorflow/core/lib/gtl/flatset.h" #include "tensorflow/core/lib/gtl/inlined_vector.h" +#include "tensorflow/core/lib/gtl/optional.h" #include "tensorflow/core/lib/strings/strcat.h" namespace xla { @@ -160,6 +161,12 @@ StatusOr IndexedArrayAnalysis::ComputeArrayFor( computed_array, ComputeArrayForReshape(instr->shape(), FindOrDie(cache_, instr->operand(0)))); + } else if (instr->opcode() == HloOpcode::kDot) { + TF_ASSIGN_OR_RETURN( + computed_array, + ComputeArrayForDot(instr->shape(), instr->dot_dimension_numbers(), + FindOrDie(cache_, instr->operand(0)), + FindOrDie(cache_, instr->operand(1)))); } else { computed_array = nullptr; } @@ -290,8 +297,7 @@ StatusOr IndexedArrayAnalysis::ComputeArrayForGather( } if (auto* indexed = dynamic_cast(source)) { - auto it = c_find(indexed->output_dims(), source_dim); - if (it != indexed->output_dims().end()) { + if (c_linear_search(indexed->output_dims(), source_dim)) { return FoldGatherOfGather(indexed, indices, source_dim, output_dims, shape); } @@ -956,11 +962,177 @@ IndexedArrayAnalysis::ComputeArrayForElementwiseUnaryOp(HloOpcode opcode, return Construct( new_source, scalar_indexed_const->indices(), scalar_indexed_const->source_dim(), - std::vector(scalar_indexed_const->output_dims().begin(), - scalar_indexed_const->output_dims().end()), + ArraySliceToVector(scalar_indexed_const->output_dims()), scalar_indexed_const->shape()); } +namespace { + +// Returns the non-contracting non-batch dimension (as per `contracting_dims` +// and `batch_dims`) if there is exactly one, otherwise returns nullopt. +gtl::optional GetOnlyNonContractingNonBatchDim( + int64 rank, ArraySlice contracting_dims, + ArraySlice batch_dims) { + gtl::optional result; + for (int64 dim = 0; dim < rank; dim++) { + if (!ArrayContains(contracting_dims, dim) && + !ArrayContains(batch_dims, dim)) { + if (result.has_value()) { + return gtl::nullopt; + } + result = dim; + } + } + return result; +} + +// Returns true if `indexed_array`, which is either the LHS or the RHS of a Dot +// HLO, can be folded into the dot operation. For now these conditions are both +// necessary and sufficient. +// +// `tag` describes the caller. Used only for logging. +// +// `contracting_dims` and `batch_dims` are the contracting and batch dimensions +// of whatever operand `indexed_array` is to the dot (LHS or RHS). +bool CanFoldDotIntoIndexedArray( + tensorflow::StringPiece tag, + Analysis::ScalarIndexedConstantArray* indexed_array, + ArraySlice contracting_dims, ArraySlice batch_dims) { + gtl::optional non_contracting_non_batch_dim = + GetOnlyNonContractingNonBatchDim(ShapeUtil::Rank(indexed_array->shape()), + contracting_dims, batch_dims); + if (!non_contracting_non_batch_dim.has_value()) { + VLOG(3) << tag << ": multiple or no non-contracting non-batch dimensions"; + return false; + } + + if (indexed_array->output_dims().size() != 1 || + indexed_array->output_dims()[0] != *non_contracting_non_batch_dim) { + VLOG(3) << tag << ": output dims != the lhs non-contracting non-batch dim"; + return false; + } + + int64 indexed_array_rank = ShapeUtil::Rank(indexed_array->shape()); + if (indexed_array->source_dim() < (indexed_array_rank - 2)) { + // This restriction can be lifted by inserting reshape nodes. + VLOG(3) << tag + << ": source dim is not in the low two dims, won't be able to form " + "a matmul"; + return false; + } + + return true; +} + +} // namespace + +StatusOr +IndexedArrayAnalysis::ComputeArrayForDotWithIndexedLhs( + const Shape& shape, const DotDimensionNumbers& dim_numbers, + ScalarIndexedConstantArray* lhs, ConstantArray* rhs) { + VLOG(3) << "ComputeArrayForDotWithIndexedLhs(" << ToString(lhs) << " " + << ToString(rhs); + if (!CanFoldDotIntoIndexedArray( + "ComputeArrayForDotWithIndexedLhs", lhs, /*contracting_dims=*/ + AsInt64Slice(dim_numbers.lhs_contracting_dimensions()), + /*batch_dims=*/AsInt64Slice(dim_numbers.lhs_batch_dimensions()))) { + return nullptr; + } + + int64 lhs_rank = ShapeUtil::Rank(lhs->shape()); + DotDimensionNumbers new_dim_numbers = dim_numbers; + new_dim_numbers.set_lhs_contracting_dimensions( + 0, lhs->source_dim() == (lhs_rank - 1) ? (lhs_rank - 2) : (lhs_rank - 1)); + + TF_ASSIGN_OR_RETURN(Literal * literal_for_new_source, + TakeOwnership(HloEvaluator{}.EvaluateDotOp( + new_dim_numbers, lhs->literal(), *rhs->literal()))); + + // The new source dimension is wherever the non-batch non-contracting LHS + // dimension "went". + int64 new_source_dim = dim_numbers.lhs_batch_dimensions_size() + + dim_numbers.rhs_batch_dimensions_size(); + + ConstantArray* new_source = Construct(literal_for_new_source); + return Construct( + new_source, lhs->indices(), new_source_dim, + ArraySliceToVector(lhs->output_dims()), shape); +} + +StatusOr +IndexedArrayAnalysis::ComputeArrayForDotWithIndexedRhs( + const Shape& shape, const DotDimensionNumbers& dim_numbers, + ConstantArray* lhs, ScalarIndexedConstantArray* rhs) { + VLOG(3) << "ComputeArrayForDotWithIndexedRhs(" << ToString(lhs) << " " + << ToString(rhs); + if (!CanFoldDotIntoIndexedArray( + "ComputeArrayForDotWithIndexedRhs", rhs, /*contracting_dims=*/ + AsInt64Slice(dim_numbers.rhs_contracting_dimensions()), + /*batch_dims=*/AsInt64Slice(dim_numbers.rhs_batch_dimensions()))) { + return nullptr; + } + + int64 rhs_rank = ShapeUtil::Rank(rhs->shape()); + + DotDimensionNumbers new_dim_numbers = dim_numbers; + new_dim_numbers.set_rhs_contracting_dimensions( + 0, rhs->source_dim() == (rhs_rank - 1) ? (rhs_rank - 2) : (rhs_rank - 1)); + + TF_ASSIGN_OR_RETURN(Literal * literal_for_new_source, + TakeOwnership(HloEvaluator{}.EvaluateDotOp( + new_dim_numbers, *lhs->literal(), rhs->literal()))); + + // The new source dimension is wherever the non-batch non-contracting RHS + // dimension "went". + int64 new_source_dim = dim_numbers.lhs_batch_dimensions_size() + + dim_numbers.rhs_batch_dimensions_size() + 1; + + ConstantArray* new_source = Construct(literal_for_new_source); + return Construct( + new_source, rhs->indices(), new_source_dim, + ArraySliceToVector(rhs->output_dims()), shape); +} + +StatusOr IndexedArrayAnalysis::ComputeArrayForDot( + const Shape& shape, const DotDimensionNumbers& dim_numbers, Array* lhs, + Array* rhs) { + // Intuitively, if + // + // - The LHS of a dot product is a gathered sequence of rows from a constant + // array (i.e. LHS[I,J] = Const[Indices[I],J]) and the RHS is a constant + // + // OR + // + // - If the RHS of a dot product is a gathered sequence of columns from a + // constant array (i.e. RHS[I,J] = Const[I, Indices[J]]) and the LHS is a + // constant + // + // then the result of the dot product itself is a gather from a constant + // array. E.g. Dot(LHS, ConstRhs) where LHS[I,J] = Const[Indices[I],J] can be + // rewritten as Result where Result[I,J] = Dot(Const, ConstRhs)[Indices[I], + // J]. + // + // We do a general version of this rewrite here. + VLOG(3) << "ComputeArrayForDot(" << ToString(lhs) << " " << ToString(rhs); + if (auto* lhs_indexed_array = + dynamic_cast(lhs)) { + if (auto* rhs_constant = dynamic_cast(rhs)) { + return ComputeArrayForDotWithIndexedLhs(shape, dim_numbers, + lhs_indexed_array, rhs_constant); + } + } + + if (auto* rhs_indexed_array = + dynamic_cast(rhs)) { + if (auto* lhs_constant = dynamic_cast(lhs)) { + return ComputeArrayForDotWithIndexedRhs(shape, dim_numbers, lhs_constant, + rhs_indexed_array); + } + } + + return nullptr; +} + tensorflow::StringPiece IndexedArrayAnalysisPrinterPass::name() const { return "indexed-array-analysis-printer-pass"; } diff --git a/tensorflow/compiler/xla/service/indexed_array_analysis.h b/tensorflow/compiler/xla/service/indexed_array_analysis.h index 8684430231c1929f82508e3675f1c275c42b6149..e923dc39f7f464a8d3c400294499a6f5efda3991 100644 --- a/tensorflow/compiler/xla/service/indexed_array_analysis.h +++ b/tensorflow/compiler/xla/service/indexed_array_analysis.h @@ -268,6 +268,18 @@ class IndexedArrayAnalysis { tensorflow::gtl::ArraySlice window_bounds, Array* source, Array* indices); + StatusOr ComputeArrayForDotWithIndexedLhs( + const Shape& shape, const DotDimensionNumbers& dim_numbers, + ScalarIndexedConstantArray* lhs, ConstantArray* rhs); + + StatusOr ComputeArrayForDotWithIndexedRhs( + const Shape& shape, const DotDimensionNumbers& dim_numbers, + ConstantArray* lhs, ScalarIndexedConstantArray* rhs); + + StatusOr ComputeArrayForDot(const Shape& shape, + const DotDimensionNumbers& dim_numbers, + Array* lhs, Array* rhs); + // This tries to fold a ScalarIndexedArray which has another // ScalarIndexedArray as a source into a ScalarIndexedArray that instead has a // ScalarIndexedArray as indices. If `source` happened to be a diff --git a/tensorflow/compiler/xla/service/indexed_array_analysis_test.cc b/tensorflow/compiler/xla/service/indexed_array_analysis_test.cc index fc2befe05b18651502c42b9892e766145d85f2e8..5f4b42799b1c26ea544f9d4447cc45b5ae9d5a48 100644 --- a/tensorflow/compiler/xla/service/indexed_array_analysis_test.cc +++ b/tensorflow/compiler/xla/service/indexed_array_analysis_test.cc @@ -799,5 +799,170 @@ ENTRY main { AssertArrayForRootExpressionIs(hlo_text, "%add"); } +TEST_F(IndexedArrayAnalysisTest, DotOpBasic_0) { + string hlo_text = R"( +HloModule DotOp + +ENTRY main { + gather_operand = s32[3,4] constant(s32[3,4]{{1,2,3,4},{5,6,7,8},{9,10,11,12}}) + dot_rhs_constant = s32[4,3] constant(s32[4,3]{{1,2,3},{4,5,6},{7,8,9},{10,11,12}}) + indices = s32[5] parameter(0) + dot_lhs = s32[5,4] gather(gather_operand, indices), + output_window_dims={1}, + elided_window_dims={0}, + gather_dims_to_operand_dims={0}, + index_vector_dim=1, + window_bounds={1,4} + ROOT dot = s32[5,3] dot(dot_lhs, dot_rhs_constant), lhs_contracting_dims={1}, rhs_contracting_dims={0} +} +)"; + + AssertArrayWithConstantsForRootExpressionIs(hlo_text, R"( +(scalar-indexed-const + (constant s32[3,3] s32[3,3] { + { 70, 80, 90 }, + { 158, 184, 210 }, + { 246, 288, 330 } }) + %indices 0->[0]))"); +} + +TEST_F(IndexedArrayAnalysisTest, DotOpBasic_1) { + string hlo_text = R"( +HloModule DotOp + +ENTRY main { + gather_operand = s32[3,4] constant(s32[3,4]{{1,2,3,4},{5,6,7,8},{9,10,11,12}}) + dot_rhs_constant = s32[3,3] constant(s32[3,3]{{1,2,3},{4,5,6},{7,8,9}}) + indices = s32[5] parameter(0) + dot_lhs = s32[3,5] gather(gather_operand, indices), + output_window_dims={0}, + elided_window_dims={1}, + gather_dims_to_operand_dims={1}, + index_vector_dim=1, + window_bounds={3,1} + ROOT dot = s32[5,3] dot(dot_lhs, dot_rhs_constant), lhs_contracting_dims={0}, rhs_contracting_dims={0} +} +)"; + + AssertArrayWithConstantsForRootExpressionIs(hlo_text, R"( +(scalar-indexed-const + (constant s32[4,3] s32[4,3] { + { 84, 99, 114 }, + { 96, 114, 132 }, + { 108, 129, 150 }, + { 120, 144, 168 } }) + %indices 0->[1]))"); +} + +TEST_F(IndexedArrayAnalysisTest, DotOpBasic_2) { + string hlo_text = R"( +HloModule DotOp + +ENTRY main { + gather_operand = s32[3,4] constant(s32[3,4]{{1,2,3,4},{5,6,7,8},{9,10,11,12}}) + dot_lhs_constant = s32[4,3] constant(s32[4,3]{{1,2,3},{4,5,6},{7,8,9},{10,11,12}}) + indices = s32[5] parameter(0) + dot_rhs = s32[3,5] gather(gather_operand, indices), + output_window_dims={0}, + elided_window_dims={1}, + gather_dims_to_operand_dims={1}, + index_vector_dim=1, + window_bounds={3,1} + ROOT dot = s32[4,5] dot(dot_lhs_constant, dot_rhs), lhs_contracting_dims={1}, rhs_contracting_dims={0} +} +)"; + + AssertArrayWithConstantsForRootExpressionIs(hlo_text, R"( +(scalar-indexed-const + (constant s32[4,4] s32[4,4] { + { 38, 44, 50, 56 }, + { 83, 98, 113, 128 }, + { 128, 152, 176, 200 }, + { 173, 206, 239, 272 } }) + %indices 1->[1]) +)"); +} + +TEST_F(IndexedArrayAnalysisTest, DotOpBasic_3) { + string hlo_text = R"( +HloModule DotOp + +ENTRY main { + gather_operand = s32[4,3] constant(s32[4,3]{{1,2,3},{4,5,6},{7,8,9},{10,11,12}}) + dot_lhs_constant = s32[4,3] constant(s32[4,3]{{1,2,3},{4,5,6},{7,8,9},{10,11,12}}) + indices = s32[5] parameter(0) + dot_rhs = s32[5,3] gather(gather_operand, indices), + output_window_dims={1}, + elided_window_dims={0}, + gather_dims_to_operand_dims={0}, + index_vector_dim=1, + window_bounds={1,3} + ROOT dot = s32[4,5] dot(dot_lhs_constant, dot_rhs), lhs_contracting_dims={1}, rhs_contracting_dims={1} +} +)"; + + AssertArrayWithConstantsForRootExpressionIs(hlo_text, R"( +(scalar-indexed-const + (constant s32[4,4] s32[4,4] { + { 14, 32, 50, 68 }, + { 32, 77, 122, 167 }, + { 50, 122, 194, 266 }, + { 68, 167, 266, 365 } }) + %indices 1->[0]) +)"); +} + +TEST_F(IndexedArrayAnalysisTest, DotOpWithBatch) { + string hlo_text = R"( +HloModule DotOp + +ENTRY main { + gather_operand = s32[2,3,2] constant(s32[2,3,2]{{{1,2},{3,4},{5,6}},{{7,8},{9,10},{11,12}}}) + dot_lhs_constant = s32[2,2,3] constant(s32[2,2,3]{{{1,2,3},{4,5,6}},{{7,8,9},{10,11,12}}}) + indices = s32[4] parameter(0) + dot_rhs = s32[2,3,4] gather(gather_operand, indices), + output_window_dims={0,1}, + elided_window_dims={2}, + gather_dims_to_operand_dims={2}, + index_vector_dim=1, + window_bounds={2,3,1} + ROOT dot = s32[2,2,4] dot(dot_lhs_constant, dot_rhs), + lhs_contracting_dims={2}, rhs_contracting_dims={1}, + lhs_batch_dims={0}, rhs_batch_dims={0} +} +)"; + + AssertArrayWithConstantsForRootExpressionIs(hlo_text, R"( +(scalar-indexed-const + (constant s32[2,2,2] s32[2,2,2] { + { { 22, 28 }, + { 49, 64 } }, + { { 220, 244 }, + { 301, 334 } } }) + %indices 3->[2]) +)"); +} + +TEST_F(IndexedArrayAnalysisTest, DotOpNegative) { + string hlo_text = R"( +HloModule DotOp + +ENTRY main { + gather_operand = s32[3,4] constant(s32[3,4]{{1,2,3,4},{5,6,7,8},{9,10,11,12}}) + dot_rhs_constant = s32[2,3] constant(s32[2,3]{{1,2,3},{4,5,6}}) + indices = s32[2] parameter(0) + dot_lhs = s32[3,2] gather(gather_operand, indices), + output_window_dims={0}, + elided_window_dims={1}, + gather_dims_to_operand_dims={1}, + index_vector_dim=1, + window_bounds={3,1} + ROOT dot = s32[3,3] dot(dot_lhs, dot_rhs_constant), lhs_contracting_dims={1}, rhs_contracting_dims={0} +} +)"; + + AssertArrayWithConstantsForRootExpressionIs(hlo_text, "%dot"); +} + } // namespace } // namespace xla diff --git a/tensorflow/compiler/xla/service/inliner_test.cc b/tensorflow/compiler/xla/service/inliner_test.cc index d2af261008f40ee83e0676cfc7e67c45f8be1844..32937b33b3737482f07d4c7607f7f1c5c183a56b 100644 --- a/tensorflow/compiler/xla/service/inliner_test.cc +++ b/tensorflow/compiler/xla/service/inliner_test.cc @@ -18,7 +18,7 @@ limitations under the License. #include #include -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/ptr_util.h" #include "tensorflow/compiler/xla/service/hlo_computation.h" #include "tensorflow/compiler/xla/service/hlo_instruction.h" @@ -51,10 +51,10 @@ TEST_F(InlinerTest, MapMax) { auto max_f32 = max_builder.Build(); auto builder = HloComputation::Builder("MapMaxFunction"); - auto lhs = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR1({1, 2, 3, 4}))); - auto rhs = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR1({4, 3, 2, 1}))); + auto lhs = builder.AddInstruction(HloInstruction::CreateConstant( + LiteralUtil::CreateR1({1, 2, 3, 4}))); + auto rhs = builder.AddInstruction(HloInstruction::CreateConstant( + LiteralUtil::CreateR1({4, 3, 2, 1}))); builder.AddInstruction( HloInstruction::CreateMap(lhs->shape(), {lhs, rhs}, max_f32.get())); @@ -70,7 +70,7 @@ TEST_F(InlinerTest, MapMax) { // Verify execution on CPU. auto result = ExecuteAndTransfer(std::move(hlo_module), {}); - auto expected = Literal::CreateR1({4, 3, 3, 4}); + auto expected = LiteralUtil::CreateR1({4, 3, 3, 4}); EXPECT_TRUE(LiteralTestUtil::Equal(*result, *expected)); } @@ -83,12 +83,12 @@ TEST_F(InlinerTest, MapConstant) { HloInstruction::CreateParameter(0, r0f32, "x")); (void)param1; const2_builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(2.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(2.0f))); auto const2_f32 = const2_builder.Build(); auto builder = HloComputation::Builder("MapConstFunction"); auto lhs = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR2({{1, 2, 3, 4}, {5, 6, 7, 8}}))); + LiteralUtil::CreateR2({{1, 2, 3, 4}, {5, 6, 7, 8}}))); builder.AddInstruction( HloInstruction::CreateMap(lhs->shape(), {lhs}, const2_f32.get())); @@ -104,7 +104,7 @@ TEST_F(InlinerTest, MapConstant) { // Verify execution on CPU. auto result = ExecuteAndTransfer(std::move(hlo_module), {}); - auto expected = Literal::CreateR2({{2, 2, 2, 2}, {2, 2, 2, 2}}); + auto expected = LiteralUtil::CreateR2({{2, 2, 2, 2}, {2, 2, 2, 2}}); EXPECT_TRUE(LiteralTestUtil::Equal(*result, *expected)); } @@ -123,10 +123,10 @@ TEST_F(InlinerTest, MapSubtractOppositeOrder) { auto max_f32 = max_builder.Build(); auto builder = HloComputation::Builder("MapSubFunction"); - auto lhs = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR1({1, 2, 3, 4}))); - auto rhs = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR1({4, 3, 2, 1}))); + auto lhs = builder.AddInstruction(HloInstruction::CreateConstant( + LiteralUtil::CreateR1({1, 2, 3, 4}))); + auto rhs = builder.AddInstruction(HloInstruction::CreateConstant( + LiteralUtil::CreateR1({4, 3, 2, 1}))); builder.AddInstruction( HloInstruction::CreateMap(lhs->shape(), {lhs, rhs}, max_f32.get())); @@ -142,7 +142,7 @@ TEST_F(InlinerTest, MapSubtractOppositeOrder) { // Verify execution on CPU. auto result = ExecuteAndTransfer(std::move(hlo_module), {}); - auto expected = Literal::CreateR1({3, 1, -1, -3}); + auto expected = LiteralUtil::CreateR1({3, 1, -1, -3}); EXPECT_TRUE(LiteralTestUtil::Equal(*result, *expected)); } diff --git a/tensorflow/compiler/xla/service/instruction_fusion.cc b/tensorflow/compiler/xla/service/instruction_fusion.cc index 9ac863576788ac34177fb1a10d265633e31fec1c..da91262130933b6d47fd95fb30bf89574b9469d6 100644 --- a/tensorflow/compiler/xla/service/instruction_fusion.cc +++ b/tensorflow/compiler/xla/service/instruction_fusion.cc @@ -97,9 +97,10 @@ bool IsAlwaysDuplicable(const HloInstruction& instruction) { case HloOpcode::kShiftRightLogical: case HloOpcode::kSlice: case HloOpcode::kSubtract: - case HloOpcode::kGenerateToken: + case HloOpcode::kAfterAll: case HloOpcode::kTranspose: case HloOpcode::kTuple: + case HloOpcode::kTupleSelect: return false; // Cheap instructions for reals, but expensive for complex. diff --git a/tensorflow/compiler/xla/service/instruction_fusion_test.cc b/tensorflow/compiler/xla/service/instruction_fusion_test.cc index 21db2338995960bde00ec9c4b325e5562fc3a592..9e7a15f0330d3f06779c850a4b575f84fe0b9505 100644 --- a/tensorflow/compiler/xla/service/instruction_fusion_test.cc +++ b/tensorflow/compiler/xla/service/instruction_fusion_test.cc @@ -167,7 +167,8 @@ TEST_F(InstructionFusionTest, AvoidDuplicationIfNotAllFusable) { builder.AddInstruction(HloInstruction::CreateParameter(1, shape, "1")); HloInstruction* binary1 = builder.AddInstruction( HloInstruction::CreateBinary(shape, HloOpcode::kAdd, param0, param1)); - builder.AddInstruction(HloInstruction::CreateSend(binary1, 0)); + auto token = builder.AddInstruction(HloInstruction::CreateToken()); + builder.AddInstruction(HloInstruction::CreateSend(binary1, token, 0)); HloInstruction* unary = builder.AddInstruction( HloInstruction::CreateUnary(shape, HloOpcode::kAbs, binary1)); @@ -258,7 +259,8 @@ TEST_F(InstructionFusionTest, AvoidDuplicationIfNotAllFusableRecursively) { add = f32[4,3]{1,0} add(p0, p0) abs1 = f32[4,3]{1,0} abs(add) log = f32[4,3]{1,0} log(abs1) - send = f32[4,3]{1,0} send(log), channel_id=0 + token = token[] after-all() + send = f32[4,3]{1,0} send(log, token), channel_id=0 abs2 = f32[4,3]{1,0} abs(log) ROOT root = f32[4,3]{1,0} subtract(abs2, add) })") @@ -288,7 +290,8 @@ TEST_F(InstructionFusionTest, AvoidDuplicationIfNotAllFusableRecursively) { p0 = f32[4,3]{1,0} parameter(0) add1 = f32[4,3]{1,0} add(p0, p0) log = f32[4,3]{1,0} log(p0) - send = f32[4,3]{1,0} send(log), channel_id=0 + token = token[] after-all() + send = f32[4,3]{1,0} send(log, token), channel_id=0 add2 = f32[4,3]{1,0} add(log, add1) ROOT root = f32[4,3]{1,0} subtract(add1, add2) })") @@ -321,7 +324,8 @@ TEST_F(InstructionFusionTest, AvoidDuplicationIfNotAllFusableRecursively) { add1 = f32[4,3]{1,0} add(p0, p0) add2 = f32[4,3]{1,0} add(add1, add1) log = f32[4,3]{1,0} log(add2) - send = f32[4,3]{1,0} send(log), channel_id=0 + token = token[] after-all() + send = f32[4,3]{1,0} send(log, token), channel_id=0 sub1 = f32[4,3]{1,0} subtract(log, add2) sub2 = f32[4,3]{1,0} subtract(add2, add1) ROOT root = (f32[4,3]{1,0}, f32[4,3]{1,0}) tuple(sub1, sub2) @@ -352,7 +356,8 @@ TEST_F(InstructionFusionTest, AllowUnaryDuplication) { builder.AddInstruction(HloInstruction::CreateParameter(0, shape, "0")); HloInstruction* unary1 = builder.AddInstruction( HloInstruction::CreateUnary(shape, HloOpcode::kFloor, param0)); - builder.AddInstruction(HloInstruction::CreateSend(unary1, 0)); + auto token = builder.AddInstruction(HloInstruction::CreateToken()); + builder.AddInstruction(HloInstruction::CreateSend(unary1, token, 0)); HloInstruction* unary2 = builder.AddInstruction( HloInstruction::CreateUnary(shape, HloOpcode::kAbs, unary1)); @@ -375,7 +380,8 @@ TEST_F(InstructionFusionTest, AllowEffectiveUnaryDuplication) { builder.AddInstruction(HloInstruction::CreateParameter(1, shape, "1")); HloInstruction* binary1 = builder.AddInstruction( HloInstruction::CreateBinary(shape, HloOpcode::kAdd, param0, param1)); - builder.AddInstruction(HloInstruction::CreateSend(binary1, 0)); + auto token = builder.AddInstruction(HloInstruction::CreateToken()); + builder.AddInstruction(HloInstruction::CreateSend(binary1, token, 0)); HloInstruction* unary = builder.AddInstruction( HloInstruction::CreateUnary(shape, HloOpcode::kAbs, binary1)); diff --git a/tensorflow/compiler/xla/service/interpreter/BUILD b/tensorflow/compiler/xla/service/interpreter/BUILD index 524d3234eb4eff9c7d000eca1a0d9f5c4fae90af..8652599dc6d48ff8c2aaa703fead161f891a57d1 100644 --- a/tensorflow/compiler/xla/service/interpreter/BUILD +++ b/tensorflow/compiler/xla/service/interpreter/BUILD @@ -74,7 +74,7 @@ cc_library( hdrs = ["executable.h"], deps = [ ":executor", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:status_macros", "//tensorflow/compiler/xla:statusor", diff --git a/tensorflow/compiler/xla/service/interpreter/executable.cc b/tensorflow/compiler/xla/service/interpreter/executable.cc index 9816acf6507a0ed5391cf4f1c94ccd0f27f5227a..8d40c08d555a232b7cf3b81cc0f9970804c2f896 100644 --- a/tensorflow/compiler/xla/service/interpreter/executable.cc +++ b/tensorflow/compiler/xla/service/interpreter/executable.cc @@ -21,7 +21,7 @@ limitations under the License. #include #include -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/ptr_util.h" #include "tensorflow/compiler/xla/service/hlo_computation.h" #include "tensorflow/compiler/xla/service/hlo_instruction.h" diff --git a/tensorflow/compiler/xla/service/layout_assignment.cc b/tensorflow/compiler/xla/service/layout_assignment.cc index 36fdfa868dfbfaf9fbf353dd6623058d518fec04..46a6d5735374e44c7c986eb66895b02da47b08e8 100644 --- a/tensorflow/compiler/xla/service/layout_assignment.cc +++ b/tensorflow/compiler/xla/service/layout_assignment.cc @@ -59,7 +59,6 @@ namespace xla { // anonymous namespace, instead of three or four spread all over this file. namespace { - } // namespace std::ostream& operator<<(std::ostream& out, @@ -113,14 +112,18 @@ LayoutConstraints::LayoutConstraints( HloComputation* computation) : points_to_analysis_(points_to_analysis), computation_(computation) { // Gather all array-shaped logical buffers into unconstrained_buffer_ids. - for (LogicalBuffer::Id id = 0; id < points_to_analysis_.num_logical_buffers(); - id++) { - auto& buffer = points_to_analysis_.logical_buffer(id); - // The points to analysis is computed per module, restrict constraints to - // array buffers in this computation. - if (buffer.IsArray() && buffer.instruction()->parent() == computation) { - unconstrained_buffer_ids_.insert(buffer.id()); - } + for (HloInstruction* inst : computation_->instructions()) { + points_to_analysis_.GetPointsToSet(inst).ForEachElement( + [&](const ShapeIndex&, const PointsToSet::BufferList& buffers) { + for (const LogicalBuffer* buffer : buffers) { + // The points to analysis is computed per module, restrict + // constraints to array buffers in this computation. + if (buffer->IsArray() && + buffer->instruction()->parent() == computation) { + unconstrained_buffer_ids_.insert(buffer->id()); + } + } + }); } } @@ -1630,7 +1633,8 @@ Status LayoutAssignment::ConstrainChannelLayouts( for (HloInstruction* instruction : computation->instructions()) { if (instruction->opcode() == HloOpcode::kRecvDone) { const Layout* layout = channel_constraints->ConstrainChannel( - instruction->channel_id(), instruction->shape().layout()); + instruction->channel_id(), + ShapeUtil::GetSubshape(instruction->shape(), {0}).layout()); TF_RET_CHECK(layout == nullptr) << instruction->ToString() << " cannot constrain layout as it was set to " @@ -1647,7 +1651,7 @@ Status LayoutAssignment::ConstrainChannelLayouts( instruction->channel_id(), operand->shape().layout()); if (layout != nullptr) { // We found an already constrained layout which does not match the one - // the kSend wants to impose. Eitehr add a new kCopy, or use the + // the kSend wants to impose. Either add a new kCopy, or use the // existing one to marshal the correct shape. Shape shape = operand->shape(); *shape.mutable_layout() = *layout; diff --git a/tensorflow/compiler/xla/service/layout_assignment_test.cc b/tensorflow/compiler/xla/service/layout_assignment_test.cc index 62599b376a12808232c703479a0ccfd7a59aa9ad..a16fa75e3032cfa4257d9b5608dd176fdb4ddbdb 100644 --- a/tensorflow/compiler/xla/service/layout_assignment_test.cc +++ b/tensorflow/compiler/xla/service/layout_assignment_test.cc @@ -21,7 +21,7 @@ limitations under the License. #include #include "tensorflow/compiler/xla/layout_util.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/service/algebraic_simplifier.h" #include "tensorflow/compiler/xla/service/computation_layout.h" #include "tensorflow/compiler/xla/service/hlo_computation.h" @@ -141,9 +141,9 @@ TEST_F(LayoutAssignmentTest, FusionInstruction) { std::vector> minor_to_majors = {{0, 1}, {1, 0}}; for (auto& minor_to_major : minor_to_majors) { auto builder = HloComputation::Builder(TestName()); - auto constant_literal1 = Literal::CreateR2WithLayout( + auto constant_literal1 = LiteralUtil::CreateR2WithLayout( {{1.0, 2.0}, {3.0, 4.0}}, LayoutUtil::MakeLayout(minor_to_major)); - auto constant_literal2 = Literal::CreateR2WithLayout( + auto constant_literal2 = LiteralUtil::CreateR2WithLayout( {{5.0, 6.0}, {7.0, 8.0}}, LayoutUtil::MakeLayout(minor_to_major)); Shape ashape = constant_literal1->shape(); @@ -192,10 +192,10 @@ TEST_F(LayoutAssignmentTest, TupleLayout) { // match their source). auto builder = HloComputation::Builder(TestName()); auto constant0 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR2WithLayout( + HloInstruction::CreateConstant(LiteralUtil::CreateR2WithLayout( {{1.0, 2.0}, {3.0, 4.0}}, LayoutUtil::MakeLayout({0, 1})))); auto constant1 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR2WithLayout( + HloInstruction::CreateConstant(LiteralUtil::CreateR2WithLayout( {{1.0, 2.0}, {3.0, 4.0}}, LayoutUtil::MakeLayout({1, 0})))); auto tuple = builder.AddInstruction( HloInstruction::CreateTuple({constant0, constant1})); @@ -229,10 +229,10 @@ TEST_F(LayoutAssignmentTest, TupleSelect) { // Verify layouts of a select with tuple operands is assigned properly. auto builder = HloComputation::Builder(TestName()); auto constant0 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR2WithLayout( + HloInstruction::CreateConstant(LiteralUtil::CreateR2WithLayout( {{1.0, 2.0}, {3.0, 4.0}}, LayoutUtil::MakeLayout({0, 1})))); auto constant1 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR2WithLayout( + HloInstruction::CreateConstant(LiteralUtil::CreateR2WithLayout( {{1.0, 2.0}, {3.0, 4.0}}, LayoutUtil::MakeLayout({1, 0})))); auto tuple0 = builder.AddInstruction( HloInstruction::CreateTuple({constant0, constant1})); @@ -240,7 +240,7 @@ TEST_F(LayoutAssignmentTest, TupleSelect) { HloInstruction::CreateTuple({constant0, constant1})); auto pred = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(true))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(true))); auto select = builder.AddInstruction(HloInstruction::CreateTernary( tuple0->shape(), HloOpcode::kSelect, pred, tuple0, tuple1)); @@ -274,7 +274,7 @@ TEST_F(LayoutAssignmentTest, ConflictingLayoutTuple) { // tuple and assigning the layouts of the copied arrays as needed. auto builder = HloComputation::Builder(TestName()); auto constant = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR2({{1.0, 2.0}, {3.0, 4.0}}))); + LiteralUtil::CreateR2({{1.0, 2.0}, {3.0, 4.0}}))); auto inner_tuple = builder.AddInstruction(HloInstruction::CreateTuple({constant})); auto nested_tuple = builder.AddInstruction( @@ -584,7 +584,7 @@ TEST_F(LayoutAssignmentTest, TransposeToBitcastToUser) { auto builder = HloComputation::Builder(TestName()); Shape input_shape = ShapeUtil::MakeShape(F32, {3, 5, 6, 7}); auto constant = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0f))); auto broadcast = builder.AddInstruction( HloInstruction::CreateBroadcast(input_shape, constant, {})); auto transpose = builder.AddInstruction(HloInstruction::CreateTranspose( @@ -770,9 +770,12 @@ TEST_F(LayoutAssignmentTest, ConditionalAsymmetricLayout) { false_builder.AddInstruction( HloInstruction::CreateParameter(0, tshape, "param")); // Using infeed as layout assignment does not mess up with it. - auto infeed = - false_builder.AddInstruction(HloInstruction::CreateInfeed(xshape, "")); - false_builder.AddInstruction(HloInstruction::CreateTuple({infeed})); + auto token = false_builder.AddInstruction(HloInstruction::CreateToken()); + auto infeed = false_builder.AddInstruction( + HloInstruction::CreateInfeed(xshape, token, "")); + auto infeed_data = false_builder.AddInstruction( + HloInstruction::CreateGetTupleElement(xshape, infeed, 0)); + false_builder.AddInstruction(HloInstruction::CreateTuple({infeed_data})); } HloComputation* false_computation = module->AddEmbeddedComputation(false_builder.Build()); @@ -799,7 +802,7 @@ TEST_F(LayoutAssignmentTest, ConditionalAsymmetricLayout) { TEST_F(LayoutAssignmentTest, InternalErrorOnBitcast) { auto builder = HloComputation::Builder(TestName()); auto constant0 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR2WithLayout( + HloInstruction::CreateConstant(LiteralUtil::CreateR2WithLayout( {{1.0, 2.0}, {3.0, 4.0}}, LayoutUtil::MakeLayout({0, 1})))); builder.AddInstruction(HloInstruction::CreateUnary( constant0->shape(), HloOpcode::kBitcast, constant0)); @@ -825,12 +828,14 @@ TEST_F(LayoutAssignmentTest, ChannelLayoutMismatch) { ENTRY entry_computation { param = (f32[2,2]) parameter(0) gte = f32[2,2] get-tuple-element(param), index=0 - recv = (f32[2,2], u32[]) recv(), channel_id=1, sharding={maximal device=1} - ROOT recv-done = f32[2,2] recv-done(recv), channel_id=1, + token = token[] after-all() + recv = (f32[2,2], u32[], token[]) recv(token), channel_id=1, sharding={maximal device=1} + recv-done = (f32[2,2], token[]) recv-done(recv), channel_id=1, sharding={maximal device=1} - send = (f32[2,2], u32[]) send(gte), channel_id=1, + ROOT root = f32[2,2] get-tuple-element(recv-done), index=0 + send = (f32[2,2], u32[], token[]) send(gte, token), channel_id=1, sharding={maximal device=0} - send-done = () send-done(send), channel_id=1, sharding={maximal device=0} + send-done = token[] send-done(send), channel_id=1, sharding={maximal device=0} } )"; @@ -849,7 +854,7 @@ TEST_F(LayoutAssignmentTest, ChannelLayoutMismatch) { AssignLayouts(module.get(), &computation_layout, &channel_constraints); EXPECT_THAT(LayoutOf(module.get(), "gte"), ElementsAre(0, 1)); - EXPECT_THAT(LayoutOf(module.get(), "recv-done"), ElementsAre(1, 0)); + EXPECT_THAT(LayoutOf(module.get(), "root"), ElementsAre(1, 0)); EXPECT_TRUE( ShapeUtil::Equal(ShapeUtil::GetSubshape( FindInstruction(module.get(), "send")->shape(), {0}), diff --git a/tensorflow/compiler/xla/service/llvm_ir/BUILD b/tensorflow/compiler/xla/service/llvm_ir/BUILD index f1e7fc29532ce7e6841010a5258f4000a7c70383..6f1e04a1c66f217105a5174d177dcfb35c0404d4 100644 --- a/tensorflow/compiler/xla/service/llvm_ir/BUILD +++ b/tensorflow/compiler/xla/service/llvm_ir/BUILD @@ -21,6 +21,11 @@ filegroup( ]), ) +load( + "//tensorflow:tensorflow.bzl", + "tf_cc_test", +) + cc_library( name = "alias_analysis", srcs = ["alias_analysis.cc"], @@ -37,12 +42,25 @@ cc_library( ], ) +tf_cc_test( + name = "alias_analysis_test", + srcs = ["alias_analysis_test.cc"], + deps = [ + ":alias_analysis", + "//tensorflow/compiler/xla/service:hlo_parser", + "//tensorflow/compiler/xla/service/cpu:custom_call_target_registry", + "//tensorflow/compiler/xla/service/cpu/tests:cpu_codegen_test", + "//tensorflow/compiler/xla/tests:filecheck", + "//tensorflow/core:test", + ], +) + cc_library( name = "llvm_util", srcs = ["llvm_util.cc"], hdrs = ["llvm_util.h"], deps = [ - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:types", "//tensorflow/compiler/xla:util", @@ -106,12 +124,31 @@ cc_library( ], ) +cc_library( + name = "kernel_tiling", + srcs = ["kernel_tiling.cc"], + hdrs = ["kernel_tiling.h"], + deps = [ + ":ir_array", + ":llvm_util", + "//tensorflow/compiler/xla:shape_util", + "//tensorflow/compiler/xla:statusor", + "//tensorflow/compiler/xla:types", + "//tensorflow/compiler/xla:util", + "//tensorflow/compiler/xla:xla_data_proto", + "//tensorflow/compiler/xla/service:hlo", + "//tensorflow/core:lib", + "@llvm//:core", + ], +) + cc_library( name = "fused_ir_emitter", srcs = ["fused_ir_emitter.cc"], hdrs = ["fused_ir_emitter.h"], deps = [ ":ir_array", + ":kernel_tiling", ":llvm_util", ":loop_emitter", ":tuple_ops", diff --git a/tensorflow/compiler/xla/service/llvm_ir/alias_analysis.cc b/tensorflow/compiler/xla/service/llvm_ir/alias_analysis.cc index f200a08a3cd7e33351ec4607d67d40e7ab28f3b9..93a8c130e1af7ca90b3dc14661deb978ff97bece 100644 --- a/tensorflow/compiler/xla/service/llvm_ir/alias_analysis.cc +++ b/tensorflow/compiler/xla/service/llvm_ir/alias_analysis.cc @@ -35,9 +35,10 @@ void AliasAnalysis::AddAliasingInformationToIrArray(const HloInstruction& hlo, llvm_ir::IrArray* array, const ShapeIndex& index) { BufferAllocation::Slice buffer_slice; - if (hlo.opcode() == HloOpcode::kParameter) { - // Parameters may alias with each other but may not alias with our temporary - // buffers. + if (hlo.opcode() == HloOpcode::kParameter && + hlo.parent() == hlo.parent()->parent()->entry_computation()) { + // Entry computation parameters may alias with each other but may not alias + // with our temporary buffers. buffer_slice = BufferAllocation::Slice(kParameterAllocation, 0, 0); } else { const std::set slices = diff --git a/tensorflow/compiler/xla/service/llvm_ir/alias_analysis_test.cc b/tensorflow/compiler/xla/service/llvm_ir/alias_analysis_test.cc new file mode 100644 index 0000000000000000000000000000000000000000..2552ff4a6a06d18f34b4ba224b66d6d97ddd74d3 --- /dev/null +++ b/tensorflow/compiler/xla/service/llvm_ir/alias_analysis_test.cc @@ -0,0 +1,83 @@ +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include +#include + +#include "tensorflow/compiler/xla/service/cpu/custom_call_target_registry.h" +#include "tensorflow/compiler/xla/service/cpu/tests/cpu_codegen_test.h" +#include "tensorflow/compiler/xla/service/llvm_ir/alias_analysis.h" +#include "tensorflow/compiler/xla/tests/filecheck.h" +#include "tensorflow/core/platform/test.h" + +namespace xla { +namespace cpu { +namespace { +class AliasAnalysisTest : public CpuCodegenTest {}; + +void FakeCustomCallTarget(float* out, float** in) {} + +REGISTER_CUSTOM_CALL_TARGET(FakeCustomCallTarget); + +TEST_F(AliasAnalysisTest, EmbeddedComputationParamsMayAliasTemps) { + const char* hlo_string = R"( +HloModule while + +body { + const.0.125 = f32[] constant(0.125) + body.state = f32[] parameter(0) + ROOT add.2.2 = f32[] add(const.0.125, body.state) +} + +condition { + const.100 = f32[] constant(100) + condition.state = f32[] parameter(0) + addend = f32[] custom-call(condition.state), custom_call_target="FakeCustomCallTarget" + add = f32[] add(addend, condition.state) + ROOT greater-than = pred[] greater-than(const.100, add) +} + +ENTRY while3 { + const.0 = f32[] constant(0) + ROOT while = f32[] while(const.0), condition=condition, body=body +} +)"; + + CompileAndVerifyIr(hlo_string, R"( +; CHECK-LABEL: @body(i8* align 4 dereferenceable(4) %retval +; CHECK: %[[add_result:.*]] = fadd fast float %[[fadd_lhs:.*]], %[[fadd_rhs:.*]] +; CHECK: store float %[[add_result]], float* %[[store_dest:.*]], !alias.scope ![[alias_scope_md_for_store:.*]] +; +; CHECK-LABEL: @condition(i8* align 1 dereferenceable(1) %fusion, i8* noalias %run_options, i8** noalias %params +; CHECK: %[[cond_state_buf_ptr:.*]] = getelementptr inbounds i8*, i8** %params, i64 0 +; CHECK: %[[cond_state_buf_untyped:.*]] = load i8*, i8** %[[cond_state_buf_ptr]] +; CHECK: %[[cond_state_buf_typed:.*]] = bitcast i8* %[[cond_state_buf_untyped]] to float* +; CHECK: load float, float* %[[cond_state_buf_typed]], !alias.scope ![[alias_scope_md_for_store]], !noalias ![[noalias_md_for_load:.*]] +; +; CHECK-LABEL: @while3( + +![[alias_scope_md_for_store]] = !{![[buffer_idx_0:.*]]} +![[buffer_idx_0]] = !{!"buffer: {index:0, offset:0, size:4}", ![[aa_md_root:.*]]} +![[aa_md_root]] = !{!"XLA global AA domain"} +![[buffer_idx_1:.*]] = !{!"buffer: {index:1, offset:0, size:4}", !3} +![[buffer_idx_1_offset_16:.*]] = !{!"buffer: {index:1, offset:16, size:1}", !3} +![[noalias_md_for_load]] = !{![[buffer_idx_1_offset_16]], ![[buffer_idx_1]]} +} +)"); +} + +} // namespace +} // namespace cpu +} // namespace xla diff --git a/tensorflow/compiler/xla/service/llvm_ir/fused_ir_emitter.cc b/tensorflow/compiler/xla/service/llvm_ir/fused_ir_emitter.cc index d909845a3a21fc55e44b0037371fca30e577980f..b12ce97e286224fcc39de74095979ea9ae80d674 100644 --- a/tensorflow/compiler/xla/service/llvm_ir/fused_ir_emitter.cc +++ b/tensorflow/compiler/xla/service/llvm_ir/fused_ir_emitter.cc @@ -119,7 +119,24 @@ Status FusedIrEmitter::HandleGetTupleElement( } Status FusedIrEmitter::HandleParameter(HloInstruction* parameter) { - generators_[parameter] = [=](const IrArray::Index& index) { + generators_[parameter] = [=](const IrArray::Index& index) -> llvm::Value* { + if (tiled_parameter_info_) { + if (llvm::Value* param_tile_buffer = + tiled_parameter_info_->GetBufferForParameter( + parameter->parameter_number())) { + // TODO(jlebar): Add AA metadata to this load. Tile buffers are global + // variables, so LLVM's points-to analysis doesn't help us much. And we + // want the AA info to be present before address spaces are inferred + // (which is pretty late in the pipeline), so even if we had + // address-space-based AA in LLVM, it wouldn't help us much here. + return ir_builder_->CreateLoad( + ir_builder_->CreateGEP( + param_tile_buffer, + {index.GetConstantWithIndexType(0), tiled_parameter_info_->x(), + tiled_parameter_info_->y()}), + "tiled_buffer"); + } + } return parameter_arrays_[parameter->parameter_number()] .EmitReadArrayElement(index, ir_builder_); }; diff --git a/tensorflow/compiler/xla/service/llvm_ir/fused_ir_emitter.h b/tensorflow/compiler/xla/service/llvm_ir/fused_ir_emitter.h index b3b6026ef17daa184c0a015fdea618597ef068b3..a6ceec7b230e9c4d24ba18aa1557a4624a37a0b4 100644 --- a/tensorflow/compiler/xla/service/llvm_ir/fused_ir_emitter.h +++ b/tensorflow/compiler/xla/service/llvm_ir/fused_ir_emitter.h @@ -25,6 +25,7 @@ limitations under the License. #include "tensorflow/compiler/xla/service/elemental_ir_emitter.h" #include "tensorflow/compiler/xla/service/hlo_instruction.h" #include "tensorflow/compiler/xla/service/llvm_ir/ir_array.h" +#include "tensorflow/compiler/xla/service/llvm_ir/kernel_tiling.h" #include "tensorflow/compiler/xla/service/llvm_ir/loop_emitter.h" #include "tensorflow/compiler/xla/statusor.h" #include "tensorflow/compiler/xla/xla_data.pb.h" @@ -56,6 +57,7 @@ class FusedIrEmitter : public DfsHloVisitorWithDefault { FusedIrEmitter(tensorflow::gtl::ArraySlice parameter_arrays, ElementalIrEmitter* elemental_emitter) : parameter_arrays_(parameter_arrays), + tiled_parameter_info_(nullptr), elemental_emitter_(elemental_emitter), ir_builder_(elemental_emitter->ir_builder()), module_(elemental_emitter->module()) {} @@ -86,9 +88,14 @@ class FusedIrEmitter : public DfsHloVisitorWithDefault { return it->second; } + void SetTiledParameterInfo(const llvm_ir::TiledParameterInfo* info) { + tiled_parameter_info_ = info; + } + private: // Arrays of parameters of fusion instruction tensorflow::gtl::ArraySlice parameter_arrays_; + const llvm_ir::TiledParameterInfo* tiled_parameter_info_; ElementalIrEmitter* elemental_emitter_; diff --git a/tensorflow/compiler/xla/service/llvm_ir/ir_array.cc b/tensorflow/compiler/xla/service/llvm_ir/ir_array.cc index ea10cef49a4a9aa048b3e0ea443f052645c4912a..dcf9838d8043fb716bed2990106f058e3bec3345 100644 --- a/tensorflow/compiler/xla/service/llvm_ir/ir_array.cc +++ b/tensorflow/compiler/xla/service/llvm_ir/ir_array.cc @@ -422,9 +422,11 @@ IrArray IrArray::CastToShape(const Shape& new_shape, llvm::IRBuilder<>* ir_builder) const { llvm::Module* module = ir_builder->GetInsertBlock()->getParent()->getParent(); llvm::Type* new_ir_type = llvm_ir::ShapeToIrType(new_shape, module); - return IrArray( + IrArray new_irarray( ir_builder->CreatePointerCast(base_ptr_, new_ir_type->getPointerTo()), new_shape); + new_irarray.metadata_ = metadata_; + return new_irarray; } /* static */ IrArray::Index IrArray::BumpIndex(const Index& index, diff --git a/tensorflow/compiler/xla/service/llvm_ir/ir_array.h b/tensorflow/compiler/xla/service/llvm_ir/ir_array.h index 4648c6d7ac089dbea7e660dd9889d557c8ad7318..0777c499238edc6091fe637d2b6b3a1f4a347254 100644 --- a/tensorflow/compiler/xla/service/llvm_ir/ir_array.h +++ b/tensorflow/compiler/xla/service/llvm_ir/ir_array.h @@ -114,19 +114,19 @@ class IrArray { size_t size() const { return multidim().size(); } llvm::Value* operator[](size_t i) const { return multidim()[i]; } - llvm::Value*& operator[](size_t i) { return multidim()[i]; } + llvm::Value*& operator[](size_t i) { return mutable_multidim()[i]; } - void push_back(llvm::Value* value) { multidim().push_back(value); } + void push_back(llvm::Value* value) { mutable_multidim().push_back(value); } void InsertAt(int64 index, llvm::Value* value) { CHECK_LE(index, size()); - multidim().insert(multidim().begin() + index, value); + mutable_multidim().insert(mutable_multidim().begin() + index, value); } using iterator = std::vector::iterator; using const_iterator = std::vector::const_iterator; - iterator begin() { return multidim().begin(); } - iterator end() { return multidim().end(); } + iterator begin() { return mutable_multidim().begin(); } + iterator end() { return mutable_multidim().end(); } const_iterator begin() const { return multidim().begin(); } const_iterator end() const { return multidim().end(); } @@ -185,7 +185,7 @@ class IrArray { private: // Changing the multi-dimensional index invalidates the linear index. - std::vector& multidim() { + std::vector& mutable_multidim() { linear_ = nullptr; return multidim_; } diff --git a/tensorflow/compiler/xla/service/llvm_ir/kernel_support_library.cc b/tensorflow/compiler/xla/service/llvm_ir/kernel_support_library.cc index 1f6e3c829f890d68aa251b101f0402c120a19d61..98d0ceb3e2065890d6a7eba8b61fa369720332f8 100644 --- a/tensorflow/compiler/xla/service/llvm_ir/kernel_support_library.cc +++ b/tensorflow/compiler/xla/service/llvm_ir/kernel_support_library.cc @@ -56,10 +56,11 @@ Status KernelSupportLibrary::For( } Status KernelSupportLibrary::If( - llvm::Value* condition, const std::function& true_block_generator, + tensorflow::StringPiece name, llvm::Value* condition, + const std::function& true_block_generator, const std::function& false_block_generator) { llvm_ir::LlvmIfData if_data = - llvm_ir::EmitIfThenElse(condition, "", ir_builder_); + llvm_ir::EmitIfThenElse(condition, name, ir_builder_); ir_builder_->SetInsertPoint(&if_data.true_block->back()); TF_RETURN_IF_ERROR(true_block_generator()); ir_builder_->SetInsertPoint(&if_data.false_block->back()); diff --git a/tensorflow/compiler/xla/service/llvm_ir/kernel_support_library.h b/tensorflow/compiler/xla/service/llvm_ir/kernel_support_library.h index 6f7a9d94e3b9e59b2dfe12b9673335a904ae78b6..9d770cc4c309d40222108833176ac5dad46754fe 100644 --- a/tensorflow/compiler/xla/service/llvm_ir/kernel_support_library.h +++ b/tensorflow/compiler/xla/service/llvm_ir/kernel_support_library.h @@ -203,16 +203,30 @@ class KernelSupportLibrary { // `true_block_generator()`; // else // `false_block_generator()`; - Status If(llvm::Value* condition, + Status If(tensorflow::StringPiece name, llvm::Value* condition, const std::function& true_block_generator, const std::function& false_block_generator = []() -> Status { return Status::OK(); }); + Status If(llvm::Value* condition, + const std::function& true_block_generator, + const std::function& false_block_generator = + []() -> Status { return Status::OK(); }) { + return If("", condition, true_block_generator, false_block_generator); + } + void IfReturnVoid(llvm::Value* condition, const std::function& true_block_generator, const std::function& false_block_generator = []() { }) { - TF_CHECK_OK(If(condition, + IfReturnVoid("", condition, true_block_generator, false_block_generator); + } + + void IfReturnVoid(tensorflow::StringPiece name, llvm::Value* condition, + const std::function& true_block_generator, + const std::function& false_block_generator = []() { + }) { + TF_CHECK_OK(If(name, condition, [&]() { true_block_generator(); return Status::OK(); diff --git a/tensorflow/compiler/xla/service/llvm_ir/kernel_tiling.cc b/tensorflow/compiler/xla/service/llvm_ir/kernel_tiling.cc new file mode 100644 index 0000000000000000000000000000000000000000..533b75cdae00dbc8244e502eb78adaf6f808b62e --- /dev/null +++ b/tensorflow/compiler/xla/service/llvm_ir/kernel_tiling.cc @@ -0,0 +1,118 @@ +/* Copyright 2017 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow/compiler/xla/service/llvm_ir/kernel_tiling.h" +#include "tensorflow/compiler/xla/layout_util.h" +#include "tensorflow/compiler/xla/service/llvm_ir/llvm_util.h" +#include "tensorflow/compiler/xla/shape_util.h" +#include "tensorflow/compiler/xla/statusor.h" +#include "tensorflow/compiler/xla/util.h" +#include "tensorflow/core/platform/logging.h" + +namespace xla { +namespace llvm_ir { + +namespace { +// Returns the indices of the first elements of all consecutive subarrays of the +// given array. For example: +// ConsecutiveSegments({m, m+1, m+2, n, k, k+1}) = {0, 3, 4} +std::vector ConsecutiveSegments(tensorflow::gtl::ArraySlice xs) { + std::vector is = {0}; + for (size_t i = 1; i < xs.size(); ++i) { + if (1 != xs[i] - xs[i - 1]) { + is.push_back(i); + } + } + return is; +} + +// Merges the sequences of dimensions of the given shape which start at the +// given indices `segs`. +Shape MergeDimensions(tensorflow::gtl::ArraySlice segs, + const Shape& shape) { + std::vector dimensions; + for (size_t i = 1; i <= segs.size(); ++i) { + dimensions.push_back(std::accumulate( + shape.dimensions().begin() + segs[i - 1], + shape.dimensions().begin() + + (segs.size() == i ? shape.dimensions().size() : segs[i]), + 1, std::multiplies())); + } + return ShapeUtil::MakeShapeWithDescendingLayout(shape.element_type(), + dimensions); +} +} // namespace + +tensorflow::gtl::optional > FindTranspose021( + const Shape& a, const Shape& b) { + if (!ShapeUtil::CompatibleIgnoringElementType(a, b)) { + return tensorflow::gtl::nullopt; + } + + std::vector perm(a.dimensions().size()); + { + auto layout_a_orig = LayoutUtil::MinorToMajor(a); + std::vector layout_a(layout_a_orig.rbegin(), layout_a_orig.rend()); + auto layout_b_orig = LayoutUtil::MinorToMajor(b); + std::vector layout_b(layout_b_orig.rbegin(), layout_b_orig.rend()); + for (size_t i = 0; i < perm.size(); ++i) { + perm[i] = PositionInContainer(layout_b, layout_a[i]); + } + } + auto segs = ConsecutiveSegments(perm); + if ((3 == segs.size() && 0 == perm[0]) || 2 == segs.size()) { + Shape norm_a = + ShapeUtil::MakeShapeWithDescendingLayoutAndSamePhysicalLayout(a); + Shape reduced_a = MergeDimensions(segs, norm_a); + auto reduced_a_dims = reduced_a.dimensions(); + std::vector dims_021; + if (2 == segs.size()) { + // The logical component-0 is of size one. + dims_021 = {1, reduced_a_dims[1], reduced_a_dims[0]}; + } else { + dims_021 = {reduced_a_dims[0], reduced_a_dims[2], reduced_a_dims[1]}; + } + + return dims_021; + } + + return tensorflow::gtl::nullopt; +} + +IrArray::Index GetUnreducedOutputIndex( + const IrArray::Index& reduced_output_index, + const Shape& reduced_output_shape, const Shape& unreduced_output_shape, + llvm::IRBuilder<>* ir_builder) { + auto bounds = reduced_output_shape.dimensions(); + auto minor_to_major = reduced_output_shape.layout().minor_to_major(); + llvm::Value* linear_index = reduced_output_index.GetConstantWithIndexType(0); + int64 multiplier = 1; + for (int i = 0; i < reduced_output_index.size(); ++i) { + int64 dim = minor_to_major[i]; + llvm::Value* addend = ir_builder->CreateMul( + reduced_output_index[dim], + reduced_output_index.GetConstantWithIndexType(multiplier), + "linearizing", + /*HasNUW=*/true, /*HasNSW=*/true); + linear_index = ir_builder->CreateAdd(linear_index, addend, "", + /*HasNUW=*/true, /*HasNSW=*/true); + multiplier *= bounds[dim]; + } + + return IrArray::Index(linear_index, unreduced_output_shape, ir_builder); +} + +} // namespace llvm_ir +} // namespace xla diff --git a/tensorflow/compiler/xla/service/llvm_ir/kernel_tiling.h b/tensorflow/compiler/xla/service/llvm_ir/kernel_tiling.h new file mode 100644 index 0000000000000000000000000000000000000000..6f1268fffbe5425d35142512d89871c6fb35db41 --- /dev/null +++ b/tensorflow/compiler/xla/service/llvm_ir/kernel_tiling.h @@ -0,0 +1,80 @@ +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#ifndef TENSORFLOW_COMPILER_XLA_SERVICE_LLVM_IR_KERNEL_TILING_H_ +#define TENSORFLOW_COMPILER_XLA_SERVICE_LLVM_IR_KERNEL_TILING_H_ + +#include "llvm/IR/Value.h" +#include "tensorflow/compiler/xla/service/hlo_instruction.h" +#include "tensorflow/compiler/xla/service/llvm_ir/ir_array.h" + +namespace xla { +namespace llvm_ir { + +// About 0-2-1 transpose: +// +// If a shape can be viewed as three logical components 0-1-2 in the order of +// major to minor, a 0-2-1-transpose changes the order of such logical +// components to 0-2-1. We call the shape being transposed the input shape and +// the transposed shape the output shape. The logical view of the input and +// output shapes for the transpose are called the 0-1-2 shape or reduced input +// shape and the 0-2-1 shape or the reduced output shape respectively. The +// original input and output shapes are called the unreduced input and output +// shapes. + +// If `b` is a 0-2-1 transpose of `a` in 0-1-2, return the dimensions for the +// reduced shape of `b` or the 0-2-1 shape. +tensorflow::gtl::optional > FindTranspose021(const Shape& a, + const Shape& b); + +// Return the unreduced output index corresponding to the given reduced output +// index. +IrArray::Index GetUnreducedOutputIndex( + const IrArray::Index& reduced_output_index, + const Shape& reduced_output_shape, const Shape& unreduced_output_shape, + llvm::IRBuilder<>* ir_builder); + +// A class to represent information for tiled parameters to support IR emission +// for 021 transpose. +class TiledParameterInfo { + public: + TiledParameterInfo(tensorflow::gtl::ArraySlice param_buffers, + llvm::Value* y, llvm::Value* x) + : param_buffers_(param_buffers), y_(y), x_(x) {} + + llvm::Value* x() const { return x_; } + llvm::Value* y() const { return y_; } + + void set_x(llvm::Value* x) { x_ = x; } + void set_y(llvm::Value* y) { y_ = y; } + + llvm::Value* GetBufferForParameter(int64 index) const { + return param_buffers_[index]; + } + + private: + // Param_buffers_[i] stores the tile buffer for the ith parameter or nullptr + // if the parameter is not tiled. + tensorflow::gtl::ArraySlice param_buffers_; + // The y coordinate within a tile. + llvm::Value* y_; + // The x coordinate within a tile. + llvm::Value* x_; +}; + +} // namespace llvm_ir +} // namespace xla + +#endif // TENSORFLOW_COMPILER_XLA_SERVICE_LLVM_IR_KERNEL_TILING_H_ diff --git a/tensorflow/compiler/xla/service/llvm_ir/llvm_util.cc b/tensorflow/compiler/xla/service/llvm_ir/llvm_util.cc index e61a2fd12de71709dfb1b5a3b736c461d6072c1e..6c55361b44b51d49cb47ea0690ff4954d4e93f84 100644 --- a/tensorflow/compiler/xla/service/llvm_ir/llvm_util.cc +++ b/tensorflow/compiler/xla/service/llvm_ir/llvm_util.cc @@ -26,7 +26,7 @@ limitations under the License. #include "llvm/Target/TargetOptions.h" #include "llvm/Transforms/Utils/Cloning.h" #include "tensorflow/compiler/xla/layout_util.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/service/name_uniquer.h" #include "tensorflow/compiler/xla/shape_util.h" #include "tensorflow/compiler/xla/types.h" @@ -36,6 +36,7 @@ limitations under the License. #include "tensorflow/core/lib/io/path.h" #include "tensorflow/core/lib/strings/str_util.h" #include "tensorflow/core/lib/strings/strcat.h" +#include "tensorflow/core/platform/byte_order.h" #include "tensorflow/core/platform/env.h" #include "tensorflow/core/platform/logging.h" #include "tensorflow/core/platform/types.h" @@ -251,14 +252,12 @@ StatusOr DecodeSelfDescribingShapeConstant(const void* shape_ptr, llvm::Constant* ConvertLiteralToIrConstant(const Literal& literal, llvm::Module* module) { - const Shape& shape = literal.shape(); - llvm::Type* type = shape.element_type() == C64 - ? llvm::Type::getFloatTy(module->getContext()) - : PrimitiveTypeToIrType(shape.element_type(), module); const char* data = static_cast(literal.untyped_data()); - uint64 num_elements = literal.size_bytes() * 8 / GetSizeInBits(type); - return llvm::ConstantDataArray::getRaw( - llvm::StringRef(data, literal.size_bytes()), num_elements, type); + CHECK_EQ(module->getDataLayout().isLittleEndian(), + tensorflow::port::kLittleEndian); + return llvm::ConstantDataArray::getString( + module->getContext(), llvm::StringRef(data, literal.size_bytes()), + /*AddNull=*/false); } llvm::AllocaInst* EmitAllocaAtFunctionEntry(llvm::Type* type, diff --git a/tensorflow/compiler/xla/service/llvm_ir/llvm_util.h b/tensorflow/compiler/xla/service/llvm_ir/llvm_util.h index 4a10ec466dae6fdb56546fb8d8b353dcff6a5b8d..9c51861eacaafcbc120e4b5f3301fe208d4c7bff 100644 --- a/tensorflow/compiler/xla/service/llvm_ir/llvm_util.h +++ b/tensorflow/compiler/xla/service/llvm_ir/llvm_util.h @@ -27,7 +27,7 @@ limitations under the License. #include "llvm/IR/Module.h" #include "llvm/IR/Value.h" #include "llvm/Support/raw_ostream.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/service/hlo_instruction.h" #include "tensorflow/compiler/xla/service/hlo_module_config.h" #include "tensorflow/compiler/xla/types.h" diff --git a/tensorflow/compiler/xla/service/logical_buffer_analysis.cc b/tensorflow/compiler/xla/service/logical_buffer_analysis.cc index f410921b4b5337192bdeae5924631d9c06b7d5a5..d631fb5ee42df6525681a5cd1fe1a8241824121d 100644 --- a/tensorflow/compiler/xla/service/logical_buffer_analysis.cc +++ b/tensorflow/compiler/xla/service/logical_buffer_analysis.cc @@ -131,18 +131,23 @@ Status LogicalBufferAnalysis::HandleDomain(HloInstruction*) { return Status::OK(); } -Status LogicalBufferAnalysis::HandleRecvDone(HloInstruction*) { - // RecvDone doesn't create a new buffer but rather aliases its input (Recv) - // tuple element at {0} to its output. +Status LogicalBufferAnalysis::HandleRecvDone(HloInstruction* recv_done) { + // RecvDone produces a two-element tuple containing the data value (which + // aliases part of its operand) and a token. Only the tuple index table and + // the token are defined by the RecvDone. + NewLogicalBuffer(recv_done, /*index=*/{}); + NewLogicalBuffer(recv_done, /*index=*/{1}); return Status::OK(); } Status LogicalBufferAnalysis::HandleSend(HloInstruction* send) { - // Send creates new buffers for the top-level tuple and the context (tuple - // element at {1}). Tuple element at {0} is an alias of the Send operand, so - // we don't need to create a new Logical Buffer for that. + // Send creates new buffers for the top-level tuple, the context (tuple + // element at {1}), and the token (tuple element at {2}). Tuple element at {0} + // is an alias of the Send operand, so we don't need to create a new Logical + // Buffer for that. NewLogicalBuffer(send, /*index=*/{}); NewLogicalBuffer(send, /*index=*/{1}); + NewLogicalBuffer(send, /*index=*/{2}); return Status::OK(); } @@ -152,10 +157,10 @@ Status LogicalBufferAnalysis::HandleTuple(HloInstruction* tuple) { return Status::OK(); } -Status LogicalBufferAnalysis::HandleSelect(HloInstruction* select) { +Status LogicalBufferAnalysis::HandleTupleSelect(HloInstruction* tuple_select) { // Select allocates a new buffer and then shallow copies the on_true or // on_false buffer into this new buffer. - NewLogicalBuffer(select, /*index=*/{}); + NewLogicalBuffer(tuple_select, /*index=*/{}); return Status::OK(); } diff --git a/tensorflow/compiler/xla/service/logical_buffer_analysis.h b/tensorflow/compiler/xla/service/logical_buffer_analysis.h index b5ef3967875a58b35631d5f69c210f5cbcd91250..81f524d84a8091e1fff13dc7c55b401143a02753 100644 --- a/tensorflow/compiler/xla/service/logical_buffer_analysis.h +++ b/tensorflow/compiler/xla/service/logical_buffer_analysis.h @@ -63,7 +63,7 @@ class LogicalBufferAnalysis : public DfsHloVisitorWithDefault { Status HandleCopy(HloInstruction* copy) override; Status HandleRecvDone(HloInstruction* recv_done) override; Status HandleSend(HloInstruction* send) override; - Status HandleSelect(HloInstruction* select) override; + Status HandleTupleSelect(HloInstruction* tuple_select) override; // A map from the buffer ID to the logical buffer std::vector> logical_buffers_; diff --git a/tensorflow/compiler/xla/service/multi_output_fusion.cc b/tensorflow/compiler/xla/service/multi_output_fusion.cc index 79b5a442aa0ecd0f67ffe4dad50465627d8913fd..4166ef5baf9c891968b584a0c498005e9ae87784 100644 --- a/tensorflow/compiler/xla/service/multi_output_fusion.cc +++ b/tensorflow/compiler/xla/service/multi_output_fusion.cc @@ -115,39 +115,18 @@ HloInstruction* MultiOutputFusion::Fuse(HloInstruction* instr1, HloInstruction* fused = instr2; // Make sure that if only one of the instructions is a fusion, or if only one // of the instructions is a multi-output fusion, it's what will be fused into. - // - // An invariant is that no bitcast nodes will show up in the middle of a - // fusion node. This invariant must hold in order for us to lower it. Given - // that, we require that during multi-output fusion, a fusion node ending with - // bitcast to preserve its structure as a nested fusion instead being - // merged and flattened. - if (fused->opcode() == HloOpcode::kFusion && - fused->fused_expression_root()->opcode() != HloOpcode::kBitcast) { + if (fused->opcode() == HloOpcode::kFusion) { std::swap(remaining, fused); } if (fused->IsMultiOutputFusion()) { std::swap(remaining, fused); } - if (fused->opcode() == HloOpcode::kFusion && - fused->fused_expression_root()->opcode() != HloOpcode::kBitcast) { + if (fused->opcode() == HloOpcode::kFusion) { remaining->MergeFusionInstructionIntoMultiOutput(fused); } else { - if (remaining->opcode() == HloOpcode::kFusion && - remaining->fused_expression_root()->opcode() == HloOpcode::kBitcast) { - auto parent_computation = remaining->parent(); - // Create a nested fusion node. - auto remaining_nested_fused = - parent_computation->AddInstruction(HloInstruction::CreateFusion( - remaining->shape(), HloInstruction::FusionKind::kLoop, - remaining)); - TF_CHECK_OK(parent_computation->ReplaceInstruction( - remaining, remaining_nested_fused)); - remaining = remaining_nested_fused; - } remaining->FuseInstructionIntoMultiOutput(fused); } - return remaining; } diff --git a/tensorflow/compiler/xla/service/multi_output_fusion.h b/tensorflow/compiler/xla/service/multi_output_fusion.h index d23822e33e11ede0c5cac97e9fe2b0c3dc88cf3d..0019cd725417d81900974b462c3b05075ce3e893 100644 --- a/tensorflow/compiler/xla/service/multi_output_fusion.h +++ b/tensorflow/compiler/xla/service/multi_output_fusion.h @@ -78,6 +78,10 @@ class MultiOutputFusion : public HloPassInterface { // Test if it's legal to fuse instr1 and instr2 into one fusion instruction. virtual bool LegalToFuse(HloInstruction* instr1, HloInstruction* instr2); + // Fuse HloInstrctuion instr1 and instr2 and return the fused instruction. + // The other instruction is removed from its parent computation. + virtual HloInstruction* Fuse(HloInstruction* instr1, HloInstruction* instr2); + // Recompute reachability for the current computation. void RecomputeReachability(); @@ -101,10 +105,6 @@ class MultiOutputFusion : public HloPassInterface { virtual bool DoProducerConsumerMultiOutputFusion(); private: - // Fuse HloInstrctuion instr1 and instr2 and return the fused instruction. - // The other instruction is removed from its parent computation. - HloInstruction* Fuse(HloInstruction* instr1, HloInstruction* instr2); - // Update the internal data structures after instr1 and instr2 are fused into // one fusion instruction. void Update(HloInstruction* instr1, HloInstruction* instr2); diff --git a/tensorflow/compiler/xla/service/name_uniquer.cc b/tensorflow/compiler/xla/service/name_uniquer.cc index 3a6a7c25f4b727c7112dbcbcb4f3d892679a0011..f6e7578a89551ec2f23d4d8c8b488c3c10e0bf1c 100644 --- a/tensorflow/compiler/xla/service/name_uniquer.cc +++ b/tensorflow/compiler/xla/service/name_uniquer.cc @@ -67,22 +67,17 @@ string NameUniquer::GetUniqueName(tensorflow::StringPiece prefix) { has_numeric_suffix = true; // Remove numeric suffix from root. root = root.substr(0, separator_index); - // Update count to at least the numeric suffix value to avoid future - // colisions with this name. - generated_names_[root] = std::max(generated_names_[root], numeric_suffix); } } - int64* count = &(generated_names_[root]); - if (*count == 0) { - *count = 1; + + SequentialIdGenerator& id_generator = generated_names_[root]; + numeric_suffix = id_generator.RegisterId(numeric_suffix); + if (numeric_suffix == 0) { return has_numeric_suffix ? tensorflow::strings::StrCat(root, separator_, 0) : root; - } else { - tensorflow::strings::StrAppend(&root, separator_, *count); - // Increment lookup under old 'root' name. - (*count)++; - return root; } + tensorflow::strings::StrAppend(&root, separator_, numeric_suffix); + return root; } } // namespace xla diff --git a/tensorflow/compiler/xla/service/name_uniquer.h b/tensorflow/compiler/xla/service/name_uniquer.h index 4139c2700b25e8600182a034a8ac6f4f041c12e6..4423d6106920eaeab830bd9dc08529ff409a5161 100644 --- a/tensorflow/compiler/xla/service/name_uniquer.h +++ b/tensorflow/compiler/xla/service/name_uniquer.h @@ -17,10 +17,11 @@ limitations under the License. #define TENSORFLOW_COMPILER_XLA_SERVICE_NAME_UNIQUER_H_ #include -#include #include "tensorflow/compiler/xla/types.h" #include "tensorflow/core/lib/core/stringpiece.h" +#include "tensorflow/core/lib/gtl/flatmap.h" +#include "tensorflow/core/lib/gtl/flatset.h" #include "tensorflow/core/platform/macros.h" namespace xla { @@ -44,13 +45,40 @@ class NameUniquer { static string GetSanitizedName(const string& name); private: + // Used to track and generate new identifiers for the same instruction name + // root. + class SequentialIdGenerator { + public: + SequentialIdGenerator() = default; + + // Tries to register id as used identifier. If id is not already used, the + // id itself will be returned. Otherwise a new one will be generated, and + // returned. + int64 RegisterId(int64 id) { + if (used_.insert(id).second) { + return id; + } + while (!used_.insert(next_).second) { + ++next_; + } + return next_++; + } + + private: + // The next identifier to be tried. + int64 next_ = 0; + + // Set of all the identifiers which has been used. + tensorflow::gtl::FlatSet used_; + }; + // The string to use to separate the prefix of the name from the uniquing // integer value. string separator_; - // Map from name prefix to the number of names generated using that prefix - // so far. - std::unordered_map generated_names_; + // Map from name prefix to the generator data structure which tracks used + // identifiers and generates new ones. + tensorflow::gtl::FlatMap generated_names_; TF_DISALLOW_COPY_AND_ASSIGN(NameUniquer); }; diff --git a/tensorflow/compiler/xla/service/name_uniquer_test.cc b/tensorflow/compiler/xla/service/name_uniquer_test.cc index 2ec255558c4ed3695ec6c824458cbedac44dc297..3e2592c6ac626143f1421e545a31d9be91e376bc 100644 --- a/tensorflow/compiler/xla/service/name_uniquer_test.cc +++ b/tensorflow/compiler/xla/service/name_uniquer_test.cc @@ -54,12 +54,13 @@ TEST_F(NameUniquerTest, NumericSuffixes) { EXPECT_EQ("foo", uniquer.GetUniqueName("foo")); EXPECT_EQ("foo.54", uniquer.GetUniqueName("foo.54")); - EXPECT_EQ("foo.55", uniquer.GetUniqueName("foo")); + EXPECT_EQ("foo.1", uniquer.GetUniqueName("foo")); EXPECT_EQ("foo.55.1", uniquer.GetUniqueName("foo.55.1")); - EXPECT_EQ("foo.55.2", uniquer.GetUniqueName("foo.55.1")); - EXPECT_EQ("bar.0", uniquer.GetUniqueName("bar.-1000")); - EXPECT_EQ("bar.1", uniquer.GetUniqueName("bar.-2000")); - EXPECT_EQ("bar.2", uniquer.GetUniqueName("bar.1")); + EXPECT_EQ("foo.55.0", uniquer.GetUniqueName("foo.55.1")); + EXPECT_EQ("bar.1000", uniquer.GetUniqueName("bar.1000")); + EXPECT_EQ("bar.2000", uniquer.GetUniqueName("bar.2000")); + EXPECT_EQ("bar.-2000", uniquer.GetUniqueName("bar.-2000")); + EXPECT_EQ("bar.1", uniquer.GetUniqueName("bar.1")); } TEST_F(NameUniquerTest, PrefixHasSuffix) { @@ -77,12 +78,12 @@ TEST_F(NameUniquerTest, Sanitize) { EXPECT_EQ("foo.54", uniquer.GetUniqueName("foo.54")); EXPECT_EQ("foo_54", uniquer.GetUniqueName("foo_54")); EXPECT_EQ("foo_54.1", uniquer.GetUniqueName("foo_54.1")); - EXPECT_EQ("foo_55", uniquer.GetUniqueName("foo")); + EXPECT_EQ("foo_2", uniquer.GetUniqueName("foo")); // Invalid characters will be replaced with '_'. - EXPECT_EQ("bar_0", uniquer.GetUniqueName("bar<-1000")); - EXPECT_EQ("bar_1", uniquer.GetUniqueName("bar<-2000")); - EXPECT_EQ("bar_2", uniquer.GetUniqueName("bar_1")); + EXPECT_EQ("bar_1000", uniquer.GetUniqueName("bar<1000")); + EXPECT_EQ("bar_2000", uniquer.GetUniqueName("bar<2000")); + EXPECT_EQ("bar_1", uniquer.GetUniqueName("bar_1")); // Separator is only recognized in the middle of the prefix. EXPECT_EQ("_10", uniquer.GetUniqueName( @@ -93,5 +94,15 @@ TEST_F(NameUniquerTest, Sanitize) { EXPECT_EQ("foobar__1", uniquer.GetUniqueName("foobar_")); } +TEST_F(NameUniquerTest, KeepNamesInRandomOrder) { + NameUniquer uniquer("."); + + EXPECT_EQ("foo.11", uniquer.GetUniqueName("foo.11")); + EXPECT_EQ("foo.10", uniquer.GetUniqueName("foo.10")); + EXPECT_EQ("foo.1", uniquer.GetUniqueName("foo.1")); + EXPECT_EQ("foo.12", uniquer.GetUniqueName("foo.12")); + EXPECT_EQ("foo.3", uniquer.GetUniqueName("foo.3")); +} + } // namespace } // namespace xla diff --git a/tensorflow/compiler/xla/service/pattern_matcher.h b/tensorflow/compiler/xla/service/pattern_matcher.h index 2515222cf2db3d9699c85c13f4fe72b3488fa217..ac6ea4c72f61a47726b3ae7dd000837d3fba1b93 100644 --- a/tensorflow/compiler/xla/service/pattern_matcher.h +++ b/tensorflow/compiler/xla/service/pattern_matcher.h @@ -86,8 +86,8 @@ namespace xla { // are provided below. // // Example nullary instruction: -// Recv() == Op().WithOpcode(HloOpcode::kRecv) -// Recv(&a) == Op(&a).WithOpcode(HloOpcode::kRecv) +// Param() == Op().WithOpcode(HloOpcode::kParam) +// Param(&a) == Op(&a).WithOpcode(HloOpcode::kParam) // // Example unary instruction: // Abs() == Op().WithOpcode(HloOpcode::kAbs) @@ -726,6 +726,32 @@ class HloInstructionPatternFusionKindImpl { ::xla::HloInstruction::FusionKind kind_; }; +// An HloInstructionPattern implementation that matches only if the instruction +// is a kGetTupleElement with a particular tuple index. +template +class HloInstructionPatternTupleIndexImpl { + public: + explicit constexpr HloInstructionPatternTupleIndexImpl( + const Previous& previous, int64 tuple_index) + : previous_(previous), tuple_index_(tuple_index) {} + + bool Match(const ::xla::HloInstruction* inst) const { + return previous_.Match(inst) && + inst->opcode() == HloOpcode::kGetTupleElement && + inst->tuple_index() == tuple_index_; + } + + bool Match(::xla::HloInstruction* inst) const { + return previous_.Match(inst) && + inst->opcode() == HloOpcode::kGetTupleElement && + inst->tuple_index() == tuple_index_; + } + + private: + Previous previous_; + int64 tuple_index_; +}; + // A pattern that matches HloInstructions. template class HloInstructionPattern { @@ -841,6 +867,17 @@ class HloInstructionPattern { HloInstructionPatternFusionKindImpl(impl_, kind), matched_inst_); } + // Modifies the pattern to match only if the instruction is a + // get-tuple-element with the given tuple index. + constexpr HloInstructionPattern> + WithTupleIndex(int64 tuple_index) const { + return HloInstructionPattern>( + HloInstructionPatternTupleIndexImpl(impl_, tuple_index), + matched_inst_); + } + private: Impl impl_; HloInstructionType** matched_inst_; @@ -880,9 +917,7 @@ Op(::xla::HloInstruction** matched_inst) { return Op(matched_inst).WithOpcode(HloOpcode::k##NAME); \ } XLA_NULLOP_PATTERN(Constant) -XLA_NULLOP_PATTERN(Infeed) XLA_NULLOP_PATTERN(Parameter) -XLA_NULLOP_PATTERN(Recv) #undef XLA_NULLOP_PATTERN // Helpers for unary instructions. @@ -919,18 +954,21 @@ XLA_UNOP_PATTERN(Cos) XLA_UNOP_PATTERN(Exp) XLA_UNOP_PATTERN(Fft) XLA_UNOP_PATTERN(Floor) +XLA_UNOP_PATTERN(GetTupleElement) XLA_UNOP_PATTERN(Imag) +XLA_UNOP_PATTERN(Infeed) XLA_UNOP_PATTERN(IsFinite) XLA_UNOP_PATTERN(Log) XLA_UNOP_PATTERN(Not) XLA_UNOP_PATTERN(Negate) -XLA_UNOP_PATTERN(Outfeed) XLA_UNOP_PATTERN(Real) +XLA_UNOP_PATTERN(Recv) +XLA_UNOP_PATTERN(RecvDone) XLA_UNOP_PATTERN(Reduce) XLA_UNOP_PATTERN(ReducePrecision) XLA_UNOP_PATTERN(Reshape) XLA_UNOP_PATTERN(Reverse) -XLA_UNOP_PATTERN(Send) +XLA_UNOP_PATTERN(SendDone) XLA_UNOP_PATTERN(Sign) XLA_UNOP_PATTERN(Sin) XLA_UNOP_PATTERN(Sort) @@ -981,8 +1019,10 @@ XLA_BINOP_PATTERN(Maximum) XLA_BINOP_PATTERN(Minimum) XLA_BINOP_PATTERN(Multiply) XLA_BINOP_PATTERN(Ne) +XLA_BINOP_PATTERN(Outfeed) XLA_BINOP_PATTERN(Power) XLA_BINOP_PATTERN(Remainder) +XLA_BINOP_PATTERN(Send) XLA_BINOP_PATTERN(Subtract) XLA_BINOP_PATTERN(And) XLA_BINOP_PATTERN(Or) @@ -1040,6 +1080,32 @@ inline auto NonConstant(HloInstructionType** matched_inst) return Op(matched_inst).IsNonConstant(); } +// Add overloads for GetTupleElement which take a int64 specifying which tuple +// element is selected. +template +inline auto GetTupleElement(Arg&& arg, int64 tuple_index) + -> decltype(Op().WithOpcode(HloOpcode::kGetTupleElement) + .WithOperand(0, std::forward(arg)) + .WithTupleIndex(tuple_index)) { + return Op() + .WithOpcode(HloOpcode::kGetTupleElement) + .WithOperand(0, std::forward(arg)) + .WithTupleIndex(tuple_index); +} + +template +inline auto GetTupleElement(HloInstructionType** matched_inst, Arg&& arg, + int64 tuple_index) + -> decltype(Op(matched_inst) + .WithOpcode(HloOpcode::kGetTupleElement) + .WithOperand(0, std::forward(arg)) + .WithTupleIndex(tuple_index)) { + return Op(matched_inst) + .WithOpcode(HloOpcode::kGetTupleElement) + .WithOperand(0, std::forward(arg)) + .WithTupleIndex(tuple_index); +} + } // namespace match } // namespace xla diff --git a/tensorflow/compiler/xla/service/pattern_matcher_test.cc b/tensorflow/compiler/xla/service/pattern_matcher_test.cc index fef3c132b0f3467a01b02f2be88b419459179277..a530581c34bf1d699eae3c53203c197f7943cc53 100644 --- a/tensorflow/compiler/xla/service/pattern_matcher_test.cc +++ b/tensorflow/compiler/xla/service/pattern_matcher_test.cc @@ -193,5 +193,23 @@ TEST(PatternMatcherTest, FusionKind) { HloInstruction::FusionKind::kLoop))); } +TEST(PatternMatcherTest, GetTupleElement) { + constexpr char kModuleStr[] = R"( + HloModule test_module + + ENTRY while.v11 { + p0 = (f32[], f32[], f32[]) parameter(0) + ROOT gte = f32[] get-tuple-element(p0), index=1 + })"; + TF_ASSERT_OK_AND_ASSIGN(auto hlo_module, ParseHloString(kModuleStr)); + + auto* root = hlo_module->entry_computation()->root_instruction(); + EXPECT_FALSE(Match(root, match::Op().WithTupleIndex(0))); + EXPECT_TRUE(Match(root, match::Op().WithTupleIndex(1))); + EXPECT_FALSE(Match(root, match::Op().WithTupleIndex(2))); + EXPECT_FALSE(Match(root, match::GetTupleElement(match::Op(), 0))); + EXPECT_TRUE(Match(root, match::GetTupleElement(match::Op(), 1))); +} + } // namespace } // namespace xla diff --git a/tensorflow/compiler/xla/service/reshape_mover.cc b/tensorflow/compiler/xla/service/reshape_mover.cc index 49ec38eb62c7b51c7a2d301d882cef032b288036..ca86c5d13e98a98c62d0c9e8e32e28fe99e0fa1f 100644 --- a/tensorflow/compiler/xla/service/reshape_mover.cc +++ b/tensorflow/compiler/xla/service/reshape_mover.cc @@ -38,7 +38,7 @@ limitations under the License. #include "tensorflow/compiler/xla/service/reshape_mover.h" #include -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/shape_util.h" #include "tensorflow/compiler/xla/status_macros.h" #include "tensorflow/compiler/xla/util.h" diff --git a/tensorflow/compiler/xla/service/reshape_mover_test.cc b/tensorflow/compiler/xla/service/reshape_mover_test.cc index 13e2d3258e3b92f52320201c382594962c0e3b2b..ad3b662c20ac53b0a6d634b16b3b908f730f3d2d 100644 --- a/tensorflow/compiler/xla/service/reshape_mover_test.cc +++ b/tensorflow/compiler/xla/service/reshape_mover_test.cc @@ -16,7 +16,7 @@ limitations under the License. #include "tensorflow/compiler/xla/service/reshape_mover.h" #include "tensorflow/compiler/xla/layout_util.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/ptr_util.h" #include "tensorflow/compiler/xla/service/hlo_computation.h" #include "tensorflow/compiler/xla/service/hlo_instruction.h" @@ -175,8 +175,9 @@ TEST_F(ReshapeMoverTest, EquivalentReshapesMoved) { TEST_F(ReshapeMoverTest, 1ConstantAnd2ReshapesMoved) { HloComputation::Builder builder(TestName()); auto root_shape = ShapeUtil::MakeShape(F32, {2, 3}); - auto const0 = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR2({{true, true, false}, {false, false, true}}))); + auto const0 = builder.AddInstruction( + HloInstruction::CreateConstant(LiteralUtil::CreateR2( + {{true, true, false}, {false, false, true}}))); auto param1 = builder.AddInstruction(HloInstruction::CreateParameter( 0, ShapeUtil::MakeShape(F32, {1, 3, 1, 2}), "param1")); @@ -255,12 +256,12 @@ TEST_F(ReshapeMoverTest, 2TrivialConstantReshapeNotMoved) { HloComputation::Builder builder(TestName()); auto root_shape = ShapeUtil::MakeShape(F32, {3, 2}); auto const0 = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR2({{1, 2, 3}, {4, 5, 6}}))); + LiteralUtil::CreateR2({{1, 2, 3}, {4, 5, 6}}))); auto reshape0 = builder.AddInstruction(HloInstruction::CreateReshape(root_shape, const0)); auto const1 = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR2({{1, 2, 3}, {4, 5, 6}}))); + LiteralUtil::CreateR2({{1, 2, 3}, {4, 5, 6}}))); auto reshape1 = builder.AddInstruction(HloInstruction::CreateReshape(root_shape, const1)); @@ -309,7 +310,7 @@ TEST_F(ReshapeMoverTest, 1NonTrivialReshapeMoved) { auto param0 = builder.AddInstruction(HloInstruction::CreateParameter( 0, ShapeUtil::MakeShape(F32, {1, 3, 1, 2}), "param0")); auto const1 = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR2({{1, 2, 3}, {4, 5, 6}}))); + LiteralUtil::CreateR2({{1, 2, 3}, {4, 5, 6}}))); auto reshape0 = builder.AddInstruction(HloInstruction::CreateReshape(root_shape, param0)); builder.AddInstruction(HloInstruction::CreateBinary( @@ -348,7 +349,7 @@ TEST_F(ReshapeMoverTest, 1NonTrivialReshapeWith1ReshapedConstNotMoved) { auto param0 = builder.AddInstruction(HloInstruction::CreateParameter( 0, ShapeUtil::MakeShape(F32, {1, 3}), "param0")); auto const1 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR1({9, 8, 7}))); + HloInstruction::CreateConstant(LiteralUtil::CreateR1({9, 8, 7}))); auto reshape0 = builder.AddInstruction(HloInstruction::CreateReshape(root_shape, param0)); auto reshape1 = diff --git a/tensorflow/compiler/xla/service/shape_inference.cc b/tensorflow/compiler/xla/service/shape_inference.cc index bbc95f86302e628bf45dae6bd720b75420e45543..70edf7883f91a0112a9576b639eb0e75b7f471e4 100644 --- a/tensorflow/compiler/xla/service/shape_inference.cc +++ b/tensorflow/compiler/xla/service/shape_inference.cc @@ -69,11 +69,11 @@ Status VerifyReducerShape(const ProgramShape& reducer_shape, } const Shape& accumulator_shape = reducer_shape.result(); - if (ShapeUtil::Rank(accumulator_shape) != 0) { + if (!ShapeUtil::IsArray(accumulator_shape) || + ShapeUtil::Rank(accumulator_shape) != 0) { return InvalidArgument( - "Reduction function must have rank 0 (rank %lld reduction function " - "given).", - ShapeUtil::Rank(accumulator_shape)); + "Reduction function must produce a scalar but has shape: %s", + ShapeUtil::HumanString(accumulator_shape).c_str()); } // Check that the accumulator can be passed in as the first argument. @@ -239,7 +239,6 @@ StatusOr InferWindowOutputShape(const Shape& base_shape, case HloOpcode::kNegate: case HloOpcode::kRoundNearestAfz: case HloOpcode::kSign: - case HloOpcode::kSort: return shape; case HloOpcode::kNot: @@ -329,7 +328,7 @@ StatusOr InferWindowOutputShape(const Shape& base_shape, return ShapeUtil::MakeShape(element_type, new_dimensions); } -/* static */ StatusOr ShapeInference::InferGenerateTokenShape( +/* static */ StatusOr ShapeInference::InferAfterAllShape( tensorflow::gtl::ArraySlice arg_shapes) { for (const Shape* arg_shape : arg_shapes) { if (arg_shape->element_type() != TOKEN) { @@ -930,6 +929,8 @@ ShapeInference::InferDegenerateDimensionBroadcastShape(HloOpcode operation, return InferClampShape(lhs, rhs, ehs); case HloOpcode::kSelect: return InferSelectShape(lhs, rhs, ehs); + case HloOpcode::kTupleSelect: + return InferTupleSelectShape(lhs, rhs, ehs); default: return InvalidArgument("Unknown operation %s.", HloOpcodeString(opcode).c_str()); @@ -962,6 +963,15 @@ ShapeInference::InferDegenerateDimensionBroadcastShape(HloOpcode operation, } return result; } + case HloOpcode::kSort: { + if (operand_shapes.size() == 1) { + return *operand_shapes[0]; + } else if (operand_shapes.size() == 2) { + return ShapeUtil::MakeTupleShape( + {*operand_shapes[0], *operand_shapes[1]}); + } + return InvalidArgument("Unexpected number of operands for sort"); + } default: return InvalidArgument("Unknown operation %s.", HloOpcodeString(opcode).c_str()); @@ -2259,15 +2269,7 @@ ShapeInference::InferDegenerateDimensionBroadcastShape(HloOpcode operation, // broadcast from all operands, not just the predicate. /* static */ StatusOr ShapeInference::InferSelectShape( const Shape& pred, const Shape& on_true, const Shape& on_false) { - bool compatible; - if (ShapeUtil::IsTuple(on_true)) { - // Select only defines the top-level buffer, so if it's a tuple, the two - // input must match exactly. - compatible = ShapeUtil::Compatible(on_true, on_false); - } else { - compatible = ShapeUtil::CompatibleIgnoringFpPrecision(on_true, on_false); - } - if (!compatible) { + if (!ShapeUtil::CompatibleIgnoringFpPrecision(on_true, on_false)) { return InvalidArgument( "Operands to select must be the same shape; got %s and %s.", ShapeUtil::HumanString(on_true).c_str(), @@ -2279,7 +2281,7 @@ ShapeInference::InferDegenerateDimensionBroadcastShape(HloOpcode operation, ShapeUtil::HumanString(pred).c_str()); } if (ShapeUtil::CompatibleIgnoringElementType(pred, on_true) || - ShapeUtil::Rank(pred) == 0) { + ShapeUtil::IsScalar(pred)) { // By this stage we know that pred's element type is PRED. Therefore, this // check restricts pred to be a PRED scalar, or a PRED array with the same // dimensions as on_true and on_false. @@ -2293,6 +2295,29 @@ ShapeInference::InferDegenerateDimensionBroadcastShape(HloOpcode operation, } } +/* static */ StatusOr ShapeInference::InferTupleSelectShape( + const Shape& pred, const Shape& on_true, const Shape& on_false) { + // Select only defines the top-level buffer, so if it's a tuple, the two + // input must match exactly. + if (!ShapeUtil::Compatible(on_true, on_false)) { + return InvalidArgument( + "Operands to tuple-select must be the same shape; got %s and %s.", + ShapeUtil::HumanString(on_true).c_str(), + ShapeUtil::HumanString(on_false).c_str()); + } + if (pred.element_type() != PRED) { + return InvalidArgument( + "TupleSelect's pred operand must have PRED element type; got %s.", + ShapeUtil::HumanString(pred).c_str()); + } + if (!ShapeUtil::IsScalar(pred)) { + return InvalidArgument( + "TupleSelect operation with non-scalar predicate: %s.", + ShapeUtil::HumanString(pred).c_str()); + } + return on_true; +} + /* static */ StatusOr ShapeInference::InferCallShape( tensorflow::gtl::ArraySlice arg_shapes, const ProgramShape& to_apply) { diff --git a/tensorflow/compiler/xla/service/shape_inference.h b/tensorflow/compiler/xla/service/shape_inference.h index eef6e62fc8d933452ebc3f9a5b8bc49828455be5..1a5684e3c306eef90fd1bfdf4565b0dcde2fbab6 100644 --- a/tensorflow/compiler/xla/service/shape_inference.h +++ b/tensorflow/compiler/xla/service/shape_inference.h @@ -216,11 +216,11 @@ class ShapeInference { static StatusOr InferConcatOpShape( tensorflow::gtl::ArraySlice arg_shapes, int64 dimension); - // Infers the shape produced by a kGenerateToken operation. Trivially this - // shape is always a TOKEN shape. However, ShapeInference serves two purposes: - // inferring shapes and checking operand shapes. This method verifies that the - // operand shapes are all TOKENs. - static StatusOr InferGenerateTokenShape( + // Infers the shape produced by a kAfterAll. Trivially this shape is always a + // TOKEN shape. However, ShapeInference serves two purposes: inferring shapes + // and checking operand shapes. This method verifies that the operand shapes + // are all TOKENs. + static StatusOr InferAfterAllShape( tensorflow::gtl::ArraySlice arg_shapes); // Helper that validates the given operand shape can be converted to the @@ -286,6 +286,10 @@ class ShapeInference { static StatusOr InferSelectShape(const Shape& pred, const Shape& on_true, const Shape& on_false); + // Helper for inferring the shape of TupleSelect ops. + static StatusOr InferTupleSelectShape(const Shape& pred, + const Shape& on_true, + const Shape& on_false); // Helper for inferring shapes of binary operations which use degenerate // dimension broadcasting (a dimension of size 1 in one operand is broadcast diff --git a/tensorflow/compiler/xla/service/transfer_manager.cc b/tensorflow/compiler/xla/service/transfer_manager.cc index 4c5038a009ba5da4172129980014913f3f4418f4..7232c658b3f0687ac93a83e46a200f88bf202084 100644 --- a/tensorflow/compiler/xla/service/transfer_manager.cc +++ b/tensorflow/compiler/xla/service/transfer_manager.cc @@ -44,6 +44,7 @@ StatusOr> TransferManager::TransferLiteralFromDevice( se::Stream* stream, const ShapedBuffer& device_buffer) { StatusOr> ret; se::Stream* substream = stream->GetOrCreateSubStream(); + substream->ThenWaitFor(stream); auto cleanup = tensorflow::gtl::MakeCleanup( [&]() { stream->ReturnSubStream(substream); }); @@ -64,6 +65,7 @@ Status TransferManager::TransferLiteralToDevice( // Use a substream so that if we are called from a HostCallback we don't // deadlock. se::Stream* substream = stream->GetOrCreateSubStream(); + substream->ThenWaitFor(stream); auto cleanup = tensorflow::gtl::MakeCleanup( [&]() { stream->ReturnSubStream(substream); }); TF_RETURN_IF_ERROR( diff --git a/tensorflow/compiler/xla/service/transfer_manager.h b/tensorflow/compiler/xla/service/transfer_manager.h index e384359642a8fe09e0b8516e342a56259912922a..82c599e482d85fc5bbe5a5a48c6c6b053186803b 100644 --- a/tensorflow/compiler/xla/service/transfer_manager.h +++ b/tensorflow/compiler/xla/service/transfer_manager.h @@ -20,7 +20,7 @@ limitations under the License. #include #include -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/service/shaped_buffer.h" #include "tensorflow/compiler/xla/statusor.h" #include "tensorflow/compiler/xla/types.h" @@ -167,16 +167,6 @@ class TransferManager { const se::Platform* platform); protected: - // Transfer a memory block of the given size from 'source' buffer to the - // Infeed interface of the device using the given executor. - // - // size is the size to transfer from source in bytes. - // - // source is the source data that must be in the target-dependent layout that - // the Infeed HLO used in the computation expects. - virtual Status TransferBufferToInfeed(se::StreamExecutor* executor, - int64 size, const void* source) = 0; - // Transfer a memory block of the given size from the device source into the // 'destination' buffer. // diff --git a/tensorflow/compiler/xla/service/transpose_folding_test.cc b/tensorflow/compiler/xla/service/transpose_folding_test.cc index cccb8f2fbb0266bbf1f40b09170938a1e5d3e78d..7051a4cf51749d294478cf9a34d4700cb52ae312 100644 --- a/tensorflow/compiler/xla/service/transpose_folding_test.cc +++ b/tensorflow/compiler/xla/service/transpose_folding_test.cc @@ -20,7 +20,7 @@ limitations under the License. #include #include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/service/gpu/ir_emission_utils.h" #include "tensorflow/compiler/xla/service/hlo_computation.h" #include "tensorflow/compiler/xla/service/hlo_instruction.h" @@ -160,11 +160,11 @@ TEST_F(TransposeFoldingTest, FuseDotWithConstantOperands) { auto builder = HloComputation::Builder("entry"); // (1.0 + 2.0) * (2.0 - 3.0) HloInstruction* const1 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0))); HloInstruction* const2 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(2.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(2.0))); HloInstruction* const3 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(3.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(3.0))); HloInstruction* add = builder.AddInstruction(HloInstruction::CreateBinary( const1->shape(), HloOpcode::kAdd, const1, const2)); HloInstruction* sub = builder.AddInstruction(HloInstruction::CreateBinary( diff --git a/tensorflow/compiler/xla/service/tuple_points_to_analysis.cc b/tensorflow/compiler/xla/service/tuple_points_to_analysis.cc index d1e174464759dbc2c0d84c4ddac27cb21635e131..990dfc410ccf6ab84af00f4a16dc783c11985844 100644 --- a/tensorflow/compiler/xla/service/tuple_points_to_analysis.cc +++ b/tensorflow/compiler/xla/service/tuple_points_to_analysis.cc @@ -292,22 +292,29 @@ Status TuplePointsToAnalysis::HandleSlice(HloInstruction* slice) { } Status TuplePointsToAnalysis::HandleRecvDone(HloInstruction* recv_done) { - // RecvDone aliases its input (Recv) tuple element {0} to its output. + // RecvDone aliases its input (Recv) tuple element {0} to element {0} of its + // output. The other indices ({} and {1}) define their own buffers. PointsToSet& points_to_set = CreateEmptyPointsToSet(recv_done); + points_to_set.AddPointedToBuffer( + logical_buffer_analysis_->GetBuffer(recv_done, /*index=*/{}), + /*index=*/{}); + points_to_set.AddPointedToBuffer( + logical_buffer_analysis_->GetBuffer(recv_done, /*index=*/{1}), + /*index=*/{1}); + const PointsToSet& operand_points_to_set = GetPointsToSet(recv_done->operand(0)); - // Recursively copy the points to set of the operand tuple {0}. + // Recursively copy the points to set of the operand tuple {0} to the output + // element {0}. points_to_set.ForEachMutableElement( [this, &points_to_set, &operand_points_to_set]( const ShapeIndex& index, PointsToSet::BufferList* buffers) { - ShapeIndex src_index({0}); - for (auto element : index) { - src_index.push_back(element); + if (index.empty() || index[0] != 0) { + return; } - *buffers = operand_points_to_set.element(src_index); - for (auto& tuple_source : - operand_points_to_set.tuple_sources(src_index)) { + *buffers = operand_points_to_set.element(index); + for (auto& tuple_source : operand_points_to_set.tuple_sources(index)) { points_to_set.add_tuple_source(index, tuple_source); } }); @@ -315,7 +322,7 @@ Status TuplePointsToAnalysis::HandleRecvDone(HloInstruction* recv_done) { } Status TuplePointsToAnalysis::HandleSend(HloInstruction* send) { - // Send creates a tuple of {aliased operand, U32 context}. + // Send creates a tuple of {aliased operand, U32 context, token}. PointsToSet& points_to_set = CreateEmptyPointsToSet(send); // Creates the points to set for the tuple and its element at {1}. @@ -328,6 +335,10 @@ Status TuplePointsToAnalysis::HandleSend(HloInstruction* send) { context_buffer->push_back( &logical_buffer_analysis_->GetBuffer(send, ShapeIndex({1}))); + auto token_buffer = points_to_set.mutable_element(ShapeIndex({2})); + token_buffer->push_back( + &logical_buffer_analysis_->GetBuffer(send, ShapeIndex({2}))); + // Recursively copy the points to set of the operand to output tuple {0}. const PointsToSet& operand_points_to_set = GetPointsToSet(send->operand(0)); operand_points_to_set.ForEachElement( @@ -388,7 +399,7 @@ Status TuplePointsToAnalysis::HandleTuple(HloInstruction* tuple) { return Status::OK(); } -Status TuplePointsToAnalysis::HandleSelect(HloInstruction* select) { +Status TuplePointsToAnalysis::HandleTupleSelect(HloInstruction* tuple_select) { // Select allocates a new buffer and then shallow copies the on_true or // on_false buffer into this new buffer. Which side is chosen cannot be // determined statically so conservatively set the points-to set to the union @@ -396,9 +407,9 @@ Status TuplePointsToAnalysis::HandleSelect(HloInstruction* select) { // // First create a copy of the on_true points-to set (and tuple sources), then // add in elements of the on_false points-to set (tuple sources). - auto on_true = select->operand(1); - auto on_false = select->operand(2); - PointsToSet& points_to_set = CreateCopiedPointsToSet(select, on_true); + auto on_true = tuple_select->operand(1); + auto on_false = tuple_select->operand(2); + PointsToSet& points_to_set = CreateCopiedPointsToSet(tuple_select, on_true); const PointsToSet& false_points_to_set = *PerInst(on_false)->points_to_set; points_to_set.ForEachMutableElement( [&](const ShapeIndex& index, PointsToSet::BufferList* buffers) { @@ -416,7 +427,7 @@ Status TuplePointsToAnalysis::HandleSelect(HloInstruction* select) { // respective element in the points-to set should contain only itself. points_to_set.mutable_element({})->clear(); points_to_set.AddPointedToBuffer( - logical_buffer_analysis_->GetBuffer(select, /*index=*/{}), + logical_buffer_analysis_->GetBuffer(tuple_select, /*index=*/{}), /*index=*/{}); return Status::OK(); } diff --git a/tensorflow/compiler/xla/service/tuple_points_to_analysis.h b/tensorflow/compiler/xla/service/tuple_points_to_analysis.h index c0d82414806d9a6ff57aec59d077f444137fec9a..686bb053288fbd6a46ca50a2c65c739354fd2678 100644 --- a/tensorflow/compiler/xla/service/tuple_points_to_analysis.h +++ b/tensorflow/compiler/xla/service/tuple_points_to_analysis.h @@ -253,7 +253,7 @@ class TuplePointsToAnalysis : public DfsHloVisitorWithDefault { Status HandleCopy(HloInstruction* copy) override; Status HandleRecvDone(HloInstruction* recv_done) override; Status HandleSend(HloInstruction* send) override; - Status HandleSelect(HloInstruction* select) override; + Status HandleTupleSelect(HloInstruction* tuple_select) override; string ToString() const; diff --git a/tensorflow/compiler/xla/service/tuple_points_to_analysis_test.cc b/tensorflow/compiler/xla/service/tuple_points_to_analysis_test.cc index 5734f284071944bc22011405898cf86f33dc48d7..0ac8df42714a1550d36560cbff901f6a8a4b3a8d 100644 --- a/tensorflow/compiler/xla/service/tuple_points_to_analysis_test.cc +++ b/tensorflow/compiler/xla/service/tuple_points_to_analysis_test.cc @@ -124,9 +124,9 @@ class TuplePointsToAnalysisTest : public HloTestBase { TEST_F(TuplePointsToAnalysisTest, SimpleTuple) { auto builder = HloComputation::Builder(TestName()); auto constant1 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0))); auto constant2 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(2.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(2.0))); auto tuple = builder.AddInstruction( HloInstruction::CreateTuple({constant1, constant2})); @@ -177,14 +177,14 @@ TEST_F(TuplePointsToAnalysisTest, NestedTuple) { // tuple. auto builder = HloComputation::Builder(TestName()); auto constant1 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0))); auto constant2 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(2.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(2.0))); auto inner_tuple = builder.AddInstruction( HloInstruction::CreateTuple({constant1, constant2})); auto constant3 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(3.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(3.0))); auto tuple = builder.AddInstruction( HloInstruction::CreateTuple({inner_tuple, constant3})); @@ -238,14 +238,14 @@ TEST_F(TuplePointsToAnalysisTest, GetTupleElement) { // tuple. auto builder = HloComputation::Builder(TestName()); auto constant1 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0))); auto constant2 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(2.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(2.0))); auto inner_tuple = builder.AddInstruction( HloInstruction::CreateTuple({constant1, constant2})); auto constant3 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(3.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(3.0))); auto tuple = builder.AddInstruction( HloInstruction::CreateTuple({inner_tuple, constant3})); @@ -270,7 +270,7 @@ TEST_F(TuplePointsToAnalysisTest, DuplicatedElement) { // Create a tuple which contains duplicate elements. auto builder = HloComputation::Builder(TestName()); auto constant = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0))); auto tuple = builder.AddInstruction( HloInstruction::CreateTuple({constant, constant, constant})); @@ -291,9 +291,9 @@ TEST_F(TuplePointsToAnalysisTest, TupleCopy) { // the same. auto builder = HloComputation::Builder(TestName()); auto constant1 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0))); auto constant2 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(2.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(2.0))); auto tuple = builder.AddInstruction( HloInstruction::CreateTuple({constant1, constant2})); auto copy = builder.AddInstruction( @@ -317,9 +317,10 @@ TEST_F(TuplePointsToAnalysisTest, SendAndSendDone) { // Send forwards its operand to the output tuple at {0}. auto builder = HloComputation::Builder(TestName()); auto constant = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0))); + auto token = builder.AddInstruction(HloInstruction::CreateToken()); auto send = builder.AddInstruction( - HloInstruction::CreateSend(constant, /*channel_id=*/0)); + HloInstruction::CreateSend(constant, token, /*channel_id=*/0)); auto send_done = builder.AddInstruction(HloInstruction::CreateSendDone(send)); BuildModuleAndRunAnalysis(builder.Build()); @@ -342,8 +343,9 @@ TEST_F(TuplePointsToAnalysisTest, SendAndSendDone) { TEST_F(TuplePointsToAnalysisTest, RecvAndRecvDone) { // RecvDone forwards its operand tuple element at {0} to the output. auto builder = HloComputation::Builder(TestName()); + auto token = builder.AddInstruction(HloInstruction::CreateToken()); auto recv = builder.AddInstruction(HloInstruction::CreateRecv( - ShapeUtil::MakeShape(F32, {1, 2, 3}), /*channel_id=*/0)); + ShapeUtil::MakeShape(F32, {1, 2, 3}), token, /*channel_id=*/0)); auto recv_done = builder.AddInstruction(HloInstruction::CreateRecvDone(recv)); BuildModuleAndRunAnalysis(builder.Build()); @@ -355,7 +357,7 @@ TEST_F(TuplePointsToAnalysisTest, RecvAndRecvDone) { ExpectHasTopLevelBuffers( points_to_analysis_->GetPointsToSet(recv).element({}), {recv}); - ExpectHasBufferAliases(recv, {0}, {{recv, {0}}, {recv_done, {}}}); + ExpectHasBufferAliases(recv, {0}, {{recv, {0}}, {recv_done, {0}}}); } TEST_F(TuplePointsToAnalysisTest, TupleSelect) { @@ -363,18 +365,18 @@ TEST_F(TuplePointsToAnalysisTest, TupleSelect) { // set containing the union of both sides. auto builder = HloComputation::Builder(TestName()); auto constant1 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0))); auto constant2 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(2.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(2.0))); auto tuple1 = builder.AddInstruction( HloInstruction::CreateTuple({constant1, constant2})); auto tuple2 = builder.AddInstruction( HloInstruction::CreateTuple({constant2, constant2})); auto pred = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(false))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(false))); auto select = builder.AddInstruction(HloInstruction::CreateTernary( - tuple1->shape(), HloOpcode::kSelect, pred, tuple1, tuple2)); + tuple1->shape(), HloOpcode::kTupleSelect, pred, tuple1, tuple2)); BuildModuleAndRunAnalysis(builder.Build()); @@ -401,9 +403,9 @@ TEST_F(TuplePointsToAnalysisTest, SelectTupleParameters) { auto param1 = builder.AddInstruction( HloInstruction::CreateParameter(1, tuple_shape, "param1")); auto pred = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(false))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(false))); auto select = builder.AddInstruction(HloInstruction::CreateTernary( - tuple_shape, HloOpcode::kSelect, pred, param0, param1)); + tuple_shape, HloOpcode::kTupleSelect, pred, param0, param1)); auto copy = builder.AddInstruction( HloInstruction::CreateUnary(tuple_shape, HloOpcode::kCopy, select)); @@ -441,18 +443,18 @@ TEST_F(TuplePointsToAnalysisTest, UnambiguousTupleSelect) { // Select from two identical tuples. The result should not be ambiguous. auto builder = HloComputation::Builder(TestName()); auto constant1 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0))); auto constant2 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(2.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(2.0))); auto tuple1 = builder.AddInstruction( HloInstruction::CreateTuple({constant1, constant2})); auto tuple2 = builder.AddInstruction( HloInstruction::CreateTuple({constant1, constant2})); auto pred = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(false))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(false))); auto select = builder.AddInstruction(HloInstruction::CreateTernary( - tuple1->shape(), HloOpcode::kSelect, pred, tuple1, tuple2)); + tuple1->shape(), HloOpcode::kTupleSelect, pred, tuple1, tuple2)); BuildModuleAndRunAnalysis(builder.Build()); @@ -472,9 +474,9 @@ TEST_F(TuplePointsToAnalysisTest, NestedTupleSelect) { // the right values. auto builder = HloComputation::Builder(TestName()); auto constant1 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0))); auto constant2 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(2.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(2.0))); auto inner_tuple1 = builder.AddInstruction( HloInstruction::CreateTuple({constant1, constant2})); auto inner_tuple2 = builder.AddInstruction( @@ -486,9 +488,9 @@ TEST_F(TuplePointsToAnalysisTest, NestedTupleSelect) { builder.AddInstruction(HloInstruction::CreateTuple({inner_tuple2})); auto pred = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(false))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(false))); auto select = builder.AddInstruction(HloInstruction::CreateTernary( - tuple1->shape(), HloOpcode::kSelect, pred, tuple1, tuple2)); + tuple1->shape(), HloOpcode::kTupleSelect, pred, tuple1, tuple2)); BuildModuleAndRunAnalysis(builder.Build()); @@ -519,9 +521,9 @@ TEST_F(TuplePointsToAnalysisTest, TupleWithBitcast) { // have the operand of the bitcast in its points-to set. auto builder = HloComputation::Builder(TestName()); auto constant1 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0))); auto constant2 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(2.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(2.0))); auto bitcast = builder.AddInstruction(HloInstruction::CreateUnary( constant2->shape(), HloOpcode::kBitcast, constant2)); auto tuple = @@ -555,9 +557,10 @@ TEST_F(TuplePointsToAnalysisTest, PointsToTupleConstantElements) { // Construct a tuple constant and kCopy it. Verify the points-to set of the // copy correctly correctly points into the nested elements of the constant. auto builder = HloComputation::Builder(TestName()); - auto tuple_constant = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::MakeTuple({Literal::CreateR2({{1.0}, {2.0}}).get(), - Literal::CreateR1({2.0, 42}).get()}))); + auto tuple_constant = builder.AddInstruction( + HloInstruction::CreateConstant(LiteralUtil::MakeTuple( + {LiteralUtil::CreateR2({{1.0}, {2.0}}).get(), + LiteralUtil::CreateR1({2.0, 42}).get()}))); auto copy = builder.AddInstruction(HloInstruction::CreateUnary( tuple_constant->shape(), HloOpcode::kCopy, tuple_constant)); @@ -577,9 +580,9 @@ TEST_F(TuplePointsToAnalysisTest, BufferAliases) { // times. Verify buffer alias sets. auto builder = HloComputation::Builder(TestName()); auto constant1 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0))); auto constant2 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(2.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(2.0))); auto inner_tuple = builder.AddInstruction( HloInstruction::CreateTuple({constant1, constant2})); auto tuple = builder.AddInstruction( @@ -618,7 +621,7 @@ class FusionPointsToAnalysisTest : public TuplePointsToAnalysisTest { auto tuple_element1 = builder.AddInstruction( HloInstruction::CreateGetTupleElement(update_shape, tuple_param0, 1)); auto ones = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR1({1.f, 1.f, 1.f, 1.f}))); + LiteralUtil::CreateR1({1.f, 1.f, 1.f, 1.f}))); // Create 'update' = Add(GetTupleElement(tuple_param0, 1), ones) auto update = builder.AddInstruction(HloInstruction::CreateBinary( update_shape, HloOpcode::kAdd, tuple_element1, ones)); @@ -866,9 +869,9 @@ TEST_F(DoesNotUseOperandBufferTest, FusedDynamicUpdateSlice) { // Create a DynamicUpdateSlice instruction of tuple element 1. auto starts = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR1({2}))); + HloInstruction::CreateConstant(LiteralUtil::CreateR1({2}))); auto update = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR1({2.f, 2.f, 2.f}))); + LiteralUtil::CreateR1({2.f, 2.f, 2.f}))); auto dynamic_update_slice = builder.AddInstruction(HloInstruction::CreateDynamicUpdateSlice( data_shape, gte1, update, starts)); @@ -960,9 +963,9 @@ TEST_F(CanShareOperandBufferWithUserTest, FusedDynamicUpdateSlice) { // Create a DynamicUpdateSlice instruction of tuple element 1. auto starts = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR1({2}))); + HloInstruction::CreateConstant(LiteralUtil::CreateR1({2}))); auto update = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR1({2.f, 2.f, 2.f}))); + LiteralUtil::CreateR1({2.f, 2.f, 2.f}))); auto dynamic_update_slice = builder.AddInstruction(HloInstruction::CreateDynamicUpdateSlice( data_shape, gte1, update, starts)); @@ -1014,9 +1017,9 @@ TEST_F(CanShareOperandBufferWithUserTest, FusedDotAdd) { Shape data_shape = ShapeUtil::MakeShape(F32, {2, 2}); auto a = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR2({{1.0, 0.0}, {0.0, 1.0}}))); + LiteralUtil::CreateR2({{1.0, 0.0}, {0.0, 1.0}}))); auto b = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR2({{2.0, 2.0}, {2.0, 2.0}}))); + LiteralUtil::CreateR2({{2.0, 2.0}, {2.0, 2.0}}))); DotDimensionNumbers dot_dnums; dot_dnums.add_lhs_contracting_dimensions(1); @@ -1025,7 +1028,7 @@ TEST_F(CanShareOperandBufferWithUserTest, FusedDotAdd) { HloInstruction::CreateDot(data_shape, a, b, dot_dnums)); auto one = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0))); auto add_operand = builder.AddInstruction( HloInstruction::CreateBroadcast(data_shape, one, {1})); @@ -1047,7 +1050,7 @@ TEST_F(CanShareOperandBufferWithUserTest, OutputFusionCantAliasOperandBuffer) { Shape data_shape = ShapeUtil::MakeShape(F32, {2, 2}); auto one = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0))); auto operand = builder.AddInstruction( HloInstruction::CreateBroadcast(data_shape, one, {1})); @@ -1055,7 +1058,7 @@ TEST_F(CanShareOperandBufferWithUserTest, OutputFusionCantAliasOperandBuffer) { HloInstruction::CreateReverse(data_shape, operand, {0, 1})); auto two = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR2({{2.0, 2.0}, {2.0, 2.0}}))); + LiteralUtil::CreateR2({{2.0, 2.0}, {2.0, 2.0}}))); auto add = builder.AddInstruction( HloInstruction::CreateBinary(data_shape, HloOpcode::kAdd, reverse, two)); @@ -1120,7 +1123,7 @@ TEST_F(CanShareOperandBufferWithUserTest, CallToComputationWithFusionRoot) { auto sub_param = sub_builder.AddInstruction( HloInstruction::CreateParameter(0, shape, "sub_param")); auto one = sub_builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0))); auto ones = sub_builder.AddInstruction( HloInstruction::CreateBroadcast(shape, one, {1})); auto add = sub_builder.AddInstruction( diff --git a/tensorflow/compiler/xla/service/tuple_simplifier_test.cc b/tensorflow/compiler/xla/service/tuple_simplifier_test.cc index d3635eae81ec7017f9bf6a69250d10716309c9ec..39b693872da6bd985d95c2abc9519662c838a3f5 100644 --- a/tensorflow/compiler/xla/service/tuple_simplifier_test.cc +++ b/tensorflow/compiler/xla/service/tuple_simplifier_test.cc @@ -18,7 +18,7 @@ limitations under the License. #include #include -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/service/hlo_computation.h" #include "tensorflow/compiler/xla/service/hlo_instruction.h" #include "tensorflow/compiler/xla/service/hlo_matchers.h" diff --git a/tensorflow/compiler/xla/service/while_loop_invariant_code_motion_test.cc b/tensorflow/compiler/xla/service/while_loop_invariant_code_motion_test.cc index 8831c513eee66e36163135b732f833d46cb7eb03..32e69c335b713c438bd7fcb2053709b0624f58ed 100644 --- a/tensorflow/compiler/xla/service/while_loop_invariant_code_motion_test.cc +++ b/tensorflow/compiler/xla/service/while_loop_invariant_code_motion_test.cc @@ -53,7 +53,7 @@ HloComputation* WhileLoopInvariantCodeMotionTest::MakeAlwaysTrueComputation( builder.AddInstruction( HloInstruction::CreateParameter(0, param_shape, "param")); builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(true))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(true))); return module->AddEmbeddedComputation(builder.Build()); } @@ -125,7 +125,7 @@ TEST_F(WhileLoopInvariantCodeMotionTest, HoistInvariantOperationTree) { builder.AddInstruction(HloInstruction::CreateUnary( scalar_s32, HloOpcode::kNegate, mul_result)); HloInstruction* constant = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(4))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(4))); HloInstruction* sub_result = builder.AddInstruction(HloInstruction::CreateBinary( scalar_s32, HloOpcode::kSubtract, negate_result, constant)); @@ -248,7 +248,9 @@ TEST_F(WhileLoopInvariantCodeMotionTest, TEST_F(WhileLoopInvariantCodeMotionTest, DontHoistInstructionWithSideEffects) { auto scalar_s32 = ShapeUtil::MakeShape(S32, {}); - Shape while_shape = ShapeUtil::MakeTupleShape({scalar_s32, scalar_s32}); + auto token_shape = ShapeUtil::MakeTokenShape(); + Shape while_shape = + ShapeUtil::MakeTupleShape({scalar_s32, scalar_s32, token_shape}); HloComputation* while_body = [&]() { HloComputation::Builder builder(TestName() + ".while_body"); @@ -258,25 +260,32 @@ TEST_F(WhileLoopInvariantCodeMotionTest, DontHoistInstructionWithSideEffects) { HloInstruction::CreateGetTupleElement(scalar_s32, param, 0)); HloInstruction* gte_1 = builder.AddInstruction( HloInstruction::CreateGetTupleElement(scalar_s32, param, 1)); + HloInstruction* in_token = builder.AddInstruction( + HloInstruction::CreateGetTupleElement(token_shape, param, 2)); + HloInstruction* out_token = builder.AddInstruction( + HloInstruction::CreateOutfeed(scalar_s32, gte_0, in_token, "")); builder.AddInstruction( - HloInstruction::CreateOutfeed(scalar_s32, gte_0, "")); - builder.AddInstruction(HloInstruction::CreateTuple({gte_0, gte_1})); + HloInstruction::CreateTuple({gte_0, gte_1, out_token})); return module().AddEmbeddedComputation(builder.Build()); }(); HloComputation::Builder builder(TestName()); + auto* scalar_param = builder.AddInstruction( + HloInstruction::CreateParameter(0, scalar_s32, "param")); + auto* token = builder.AddInstruction(HloInstruction::CreateToken()); auto* init_value = builder.AddInstruction( - HloInstruction::CreateParameter(0, while_shape, "init_value")); + HloInstruction::CreateTuple({scalar_param, scalar_param, token})); auto* while_inst = builder.AddInstruction(HloInstruction::CreateWhile( while_shape, MakeAlwaysTrueComputation(while_shape, &module()), while_body, init_value)); - + builder.AddInstruction( + HloInstruction::CreateGetTupleElement(scalar_s32, while_inst, 0)); module().AddEntryComputation(builder.Build()); TF_ASSERT_OK_AND_ASSIGN(bool simplified_loop, WhileLoopInvariantCodeMotion{}.Run(&module())); - EXPECT_FALSE(simplified_loop); + ASSERT_FALSE(simplified_loop); EXPECT_THAT(while_inst->while_body()->instructions(), Contains(op::Outfeed())); @@ -287,7 +296,9 @@ TEST_F(WhileLoopInvariantCodeMotionTest, DontHoistBitcastAlone) { // bitcast either. auto scalar_s32 = ShapeUtil::MakeShape(S32, {}); auto scalar_f32 = ShapeUtil::MakeShape(F32, {}); - Shape while_shape = ShapeUtil::MakeTupleShape({scalar_s32, scalar_s32}); + auto token_shape = ShapeUtil::MakeTokenShape(); + Shape while_shape = + ShapeUtil::MakeTupleShape({scalar_s32, scalar_s32, token_shape}); HloComputation* while_body = [&]() { HloComputation::Builder builder(TestName() + ".while_body"); @@ -297,21 +308,29 @@ TEST_F(WhileLoopInvariantCodeMotionTest, DontHoistBitcastAlone) { HloInstruction::CreateGetTupleElement(scalar_s32, param, 0)); HloInstruction* gte_1 = builder.AddInstruction( HloInstruction::CreateGetTupleElement(scalar_s32, param, 1)); + HloInstruction* in_token = builder.AddInstruction( + HloInstruction::CreateGetTupleElement(token_shape, param, 2)); HloInstruction* bitcast_inst = builder.AddInstruction( HloInstruction::CreateUnary(scalar_f32, HloOpcode::kBitcast, gte_0)); + HloInstruction* out_token = builder.AddInstruction( + HloInstruction::CreateOutfeed(scalar_f32, bitcast_inst, in_token, "")); builder.AddInstruction( - HloInstruction::CreateOutfeed(scalar_f32, bitcast_inst, "")); - builder.AddInstruction(HloInstruction::CreateTuple({gte_0, gte_1})); + HloInstruction::CreateTuple({gte_0, gte_1, out_token})); return module().AddEmbeddedComputation(builder.Build()); }(); HloComputation::Builder builder(TestName()); + auto* scalar_param = builder.AddInstruction( + HloInstruction::CreateParameter(0, scalar_s32, "param")); + auto* token = builder.AddInstruction(HloInstruction::CreateToken()); auto* init_value = builder.AddInstruction( - HloInstruction::CreateParameter(0, while_shape, "init_value")); + HloInstruction::CreateTuple({scalar_param, scalar_param, token})); auto* while_inst = builder.AddInstruction(HloInstruction::CreateWhile( while_shape, MakeAlwaysTrueComputation(while_shape, &module()), while_body, init_value)); + builder.AddInstruction( + HloInstruction::CreateGetTupleElement(scalar_s32, while_inst, 0)); module().AddEntryComputation(builder.Build()); diff --git a/tensorflow/compiler/xla/service/while_loop_simplifier_test.cc b/tensorflow/compiler/xla/service/while_loop_simplifier_test.cc index 619e87caa5b6d0f6ec3c3b1489b0d4f50ef29963..2e1571943e537f772ee7dcd95c80ba540445b76e 100644 --- a/tensorflow/compiler/xla/service/while_loop_simplifier_test.cc +++ b/tensorflow/compiler/xla/service/while_loop_simplifier_test.cc @@ -157,7 +157,7 @@ TEST_F(WhileLoopSimplifierTest, auto* while_op = computation->root_instruction(); ASSERT_EQ(while_op->opcode(), HloOpcode::kWhile); auto* true_op = while_op->while_body()->AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(true))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(true))); TF_ASSERT_OK(true_op->AddControlDependencyTo( while_op->while_body()->root_instruction())); ASSERT_TRUE(WhileLoopSimplifier().Run(the_module).ValueOrDie()); @@ -175,9 +175,11 @@ TEST_F(WhileLoopSimplifierTest, LoopWithSendNotSimplified) { auto* while_op = computation->root_instruction(); ASSERT_EQ(while_op->opcode(), HloOpcode::kWhile); auto* while_body = while_op->while_body(); + auto* token = while_body->AddInstruction(HloInstruction::CreateToken()); auto* send = while_body->AddInstruction(HloInstruction::CreateSend( while_body->AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(true))), + HloInstruction::CreateConstant(LiteralUtil::CreateR0(true))), + token, /*channel_id=*/0)); while_body->AddInstruction(HloInstruction::CreateSendDone(send)); EXPECT_FALSE(WhileLoopSimplifier().Run(the_module).ValueOrDie()); @@ -190,8 +192,9 @@ TEST_F(WhileLoopSimplifierTest, LoopWithRecvNotSimplified) { auto* while_op = computation->root_instruction(); ASSERT_EQ(while_op->opcode(), HloOpcode::kWhile); auto* while_body = while_op->while_body(); + auto* token = while_body->AddInstruction(HloInstruction::CreateToken()); auto* recv = while_body->AddInstruction( - HloInstruction::CreateRecv(ShapeUtil::MakeShape(F32, {1}), + HloInstruction::CreateRecv(ShapeUtil::MakeShape(F32, {1}), token, /*channel_id=*/0)); while_body->AddInstruction(HloInstruction::CreateRecvDone(recv)); EXPECT_FALSE(WhileLoopSimplifier().Run(the_module).ValueOrDie()); @@ -208,8 +211,9 @@ TEST_F(WhileLoopSimplifierTest, LoopWithInfeedNotSimplified) { auto* while_op = computation->root_instruction(); ASSERT_EQ(while_op->opcode(), HloOpcode::kWhile); auto* while_body = while_op->while_body(); - while_body->AddInstruction( - HloInstruction::CreateInfeed(ShapeUtil::MakeShape(F32, {1}), "config")); + auto token = while_body->AddInstruction(HloInstruction::CreateToken()); + while_body->AddInstruction(HloInstruction::CreateInfeed( + ShapeUtil::MakeShape(F32, {1}), token, "config")); EXPECT_FALSE(WhileLoopSimplifier().Run(the_module).ValueOrDie()); } diff --git a/tensorflow/compiler/xla/service/while_util.cc b/tensorflow/compiler/xla/service/while_util.cc index 473eab2ea84eb8faf745cbe299bc80bcc1b62a35..1ef17b9d7d2e769aadf39f8a70f78200b88e9d2c 100644 --- a/tensorflow/compiler/xla/service/while_util.cc +++ b/tensorflow/compiler/xla/service/while_util.cc @@ -14,6 +14,7 @@ limitations under the License. ==============================================================================*/ #include "tensorflow/compiler/xla/service/while_util.h" +#include "tensorflow/compiler/xla/literal_util.h" #include "tensorflow/compiler/xla/service/hlo_computation.h" #include "tensorflow/compiler/xla/service/hlo_creation_utils.h" #include "tensorflow/compiler/xla/service/tuple_util.h" @@ -38,7 +39,7 @@ static StatusOr WidenWhileCondition( // the root instruction later. We later change the root instruction to // something more appropriate. builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(false))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(false))); return narrow_condition->parent()->AddEmbeddedComputation(builder.Build()); }(); @@ -154,7 +155,7 @@ MakeCountedLoopConditionComputation(const Shape& loop_state_shape, {&loop_state_shape}, scalar_pred, "while_cond")); HloInstruction* trip_count_constant = cond_computation->AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(trip_count))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(trip_count))); HloInstruction* param = cond_computation->parameter_instruction(0); TF_ASSIGN_OR_RETURN(HloInstruction * indvar, @@ -175,7 +176,7 @@ static StatusOr> MakeCountedLoopBodyComputation( CreateComputationWithSignature( {&loop_state_shape}, loop_state_shape, "while_body")); HloInstruction* one = body_computation->AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1))); HloInstruction* param = body_computation->parameter_instruction(0); TF_ASSIGN_OR_RETURN(HloInstruction * indvar, MakeGetTupleElementHlo(param, 0)); @@ -203,7 +204,7 @@ static StatusOr MakeInitTupleFromInitValues( std::vector init_values_with_indvar; init_values_with_indvar.reserve(init_values.size() + 1); HloInstruction* zero = computation->AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(0))); init_values_with_indvar.push_back(zero); c_copy(init_values, std::back_inserter(init_values_with_indvar)); return computation->AddInstruction( diff --git a/tensorflow/compiler/xla/service/while_util_test.cc b/tensorflow/compiler/xla/service/while_util_test.cc index d79d3297213e832306ea4726483b0f215df0f5d3..2ccb919acf9c4e7c59a1ebaf36f42a6781068b5e 100644 --- a/tensorflow/compiler/xla/service/while_util_test.cc +++ b/tensorflow/compiler/xla/service/while_util_test.cc @@ -179,7 +179,9 @@ body { cond { param.c = (s32[], s32[]) parameter(0) - ROOT condition = pred[] infeed() + token = token[] after-all() + infeed = (pred[], token[]) infeed(token) + ROOT condition = pred[] get-tuple-element(infeed), index=0 } ENTRY main { diff --git a/tensorflow/compiler/xla/service/zero_sized_hlo_elimination.cc b/tensorflow/compiler/xla/service/zero_sized_hlo_elimination.cc index 44b0ec5cd4c1d406467007fcc530e919d602c438..83d696fe0915086c3c98b6d7cbdaeaeb4d9d0bdb 100644 --- a/tensorflow/compiler/xla/service/zero_sized_hlo_elimination.cc +++ b/tensorflow/compiler/xla/service/zero_sized_hlo_elimination.cc @@ -15,7 +15,7 @@ limitations under the License. #include "tensorflow/compiler/xla/service/zero_sized_hlo_elimination.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/service/hlo_computation.h" #include "tensorflow/compiler/xla/service/hlo_instruction.h" #include "tensorflow/compiler/xla/shape_util.h" @@ -32,7 +32,8 @@ StatusOr ZeroSizedHloElimination::Run(HloModule* module) { for (HloComputation* comp : module->MakeNonfusionComputations()) { for (HloInstruction* instruction : comp->MakeInstructionPostOrder()) { if (instruction->HasSideEffect() || - !ShapeUtil::IsArray(instruction->shape())) { + !ShapeUtil::IsArray(instruction->shape()) || + instruction->opcode() == HloOpcode::kConstant) { continue; } if (comp->IsRemovable(instruction) && diff --git a/tensorflow/compiler/xla/service/zero_sized_hlo_elimination_test.cc b/tensorflow/compiler/xla/service/zero_sized_hlo_elimination_test.cc index f5331280ee9f252aa5717baab88f2c203be5c372..b9ef18892d7aa859f6b0b505db4c004e4f5c5066 100644 --- a/tensorflow/compiler/xla/service/zero_sized_hlo_elimination_test.cc +++ b/tensorflow/compiler/xla/service/zero_sized_hlo_elimination_test.cc @@ -19,7 +19,7 @@ limitations under the License. #include #include -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/service/hlo_computation.h" #include "tensorflow/compiler/xla/service/hlo_instruction.h" #include "tensorflow/compiler/xla/service/hlo_module.h" @@ -67,7 +67,16 @@ TEST_F(ZeroSizedHloEliminationTest, DoesNotEliminateParameter) { } TEST_F(ZeroSizedHloEliminationTest, DoesNotEliminateSideEffects) { - builder_.AddInstruction(HloInstruction::CreateSend(zero_sized_param_, 0)); + auto token = builder_.AddInstruction(HloInstruction::CreateToken()); + builder_.AddInstruction( + HloInstruction::CreateSend(zero_sized_param_, token, 0)); + TF_ASSERT_OK_AND_ASSIGN(bool changed, RunZeroSizedElimination()); + EXPECT_FALSE(changed); +} + +TEST_F(ZeroSizedHloEliminationTest, DoesNotEliminateConstant) { + builder_.AddInstruction( + HloInstruction::CreateConstant(LiteralUtil::CreateR1({}))); TF_ASSERT_OK_AND_ASSIGN(bool changed, RunZeroSizedElimination()); EXPECT_FALSE(changed); } diff --git a/tensorflow/compiler/xla/shape_layout.cc b/tensorflow/compiler/xla/shape_layout.cc index 7ee366b27a82bdbcb7a63a57ea80194db8ca7df4..caad31d6ce7ce35fa362ec364b0d7f1d95973715 100644 --- a/tensorflow/compiler/xla/shape_layout.cc +++ b/tensorflow/compiler/xla/shape_layout.cc @@ -67,6 +67,14 @@ void ShapeLayout::ResetLayout(const Layout& layout) { TF_CHECK_OK(ShapeUtil::ValidateShape(shape_)); } +void ShapeLayout::ResetLayout(const Layout& layout, + ShapeIndexView shape_index) { + CHECK(ShapeUtil::IsTuple(shape_)); + *ShapeUtil::GetMutableSubshape(&shape_, shape_index)->mutable_layout() = + layout; + TF_CHECK_OK(ShapeUtil::ValidateShape(shape_)); +} + bool ShapeLayout::operator==(const ShapeLayout& other) const { return ShapeUtil::Equal(shape_, other.shape_); } diff --git a/tensorflow/compiler/xla/shape_layout.h b/tensorflow/compiler/xla/shape_layout.h index 36806da599cc9b27286e67c128bb7f496f29c105..214cf98854938414c23c5031f4114016140ae9a7 100644 --- a/tensorflow/compiler/xla/shape_layout.h +++ b/tensorflow/compiler/xla/shape_layout.h @@ -72,6 +72,10 @@ class ShapeLayout { // tuple. void ResetLayout(const Layout& layout); + // Resets the layout on the shape at the provided ShapeIndex to the provided + // layout. Shape must be a tuple. + void ResetLayout(const Layout& layout, ShapeIndexView shape_index); + // Returns a string representation of this object. string ToString() const { return ShapeUtil::HumanStringWithLayout(shape_); } diff --git a/tensorflow/compiler/xla/shape_util.cc b/tensorflow/compiler/xla/shape_util.cc index 98c3095499f23a816722430846da7c7cbe2ece67..f4668c0f559acd7b1301499aaf71d5c6925424b3 100644 --- a/tensorflow/compiler/xla/shape_util.cc +++ b/tensorflow/compiler/xla/shape_util.cc @@ -24,6 +24,7 @@ limitations under the License. #include "tensorflow/compiler/xla/index_util.h" #include "tensorflow/compiler/xla/layout_util.h" +#include "tensorflow/compiler/xla/overflow_util.h" #include "tensorflow/compiler/xla/primitive_util.h" #include "tensorflow/compiler/xla/status_macros.h" #include "tensorflow/compiler/xla/types.h" @@ -45,28 +46,14 @@ namespace xla { using ::tensorflow::strings::StrAppend; using ::tensorflow::strings::StrCat; -string ShapeIndex::ToString() const { - return StrCat("{", tensorflow::str_util::Join(indices_, ","), "}"); -} +string ShapeIndex::ToString() const { return ShapeIndexView(*this).ToString(); } string ShapeIndexView::ToString() const { - return StrCat("{", - tensorflow::str_util::Join( - tensorflow::gtl::make_range(begin_, end_), ","), - "}"); + return StrCat("{", tensorflow::str_util::Join(indices_, ","), "}"); } bool ShapeIndexView::operator==(const ShapeIndexView& other) const { - if (size() != other.size()) { - return false; - } - for (auto it = begin(), other_it = other.begin(); it != end(); - ++it, ++other_it) { - if (*it != *other_it) { - return false; - } - } - return true; + return indices_ == other.indices_; } bool ShapeIndexView::operator!=(const ShapeIndexView& other) const { @@ -94,8 +81,11 @@ bool IsArrayPrimitiveType(PrimitiveType primitive_type) { // Recursive helper for comparing the equality of two shapes. Returns true if // the shapes are the same. If compare_layouts is true, then layouts must also // match. -bool CompareShapes(const Shape& lhs, const Shape& rhs, bool compare_layouts) { - if (!ShapeUtil::SameElementType(lhs, rhs)) { +bool CompareShapes(const Shape& lhs, const Shape& rhs, bool compare_layouts, + bool ignore_fp_precision) { + if ((ignore_fp_precision && + !ShapeUtil::SameElementTypeIgnoringFpPrecision(lhs, rhs)) || + (!ignore_fp_precision && !ShapeUtil::SameElementType(lhs, rhs))) { VLOG(3) << "CompareShapes: lhs element type != rhs element type"; return false; } @@ -103,7 +93,8 @@ bool CompareShapes(const Shape& lhs, const Shape& rhs, bool compare_layouts) { if (ShapeUtil::IsTuple(lhs)) { return ContainersEqual(lhs.tuple_shapes(), rhs.tuple_shapes(), [=](const Shape& l, const Shape& r) { - return CompareShapes(l, r, compare_layouts); + return CompareShapes(l, r, compare_layouts, + ignore_fp_precision); }); } else if (!ShapeUtil::IsArray(lhs)) { // Non-tuple, non-array tupes such as opaque and token types are trivially @@ -170,7 +161,8 @@ StatusOr MakeShapeWithLayoutInternal( } // namespace /* static */ bool ShapeUtil::Equal(const Shape& lhs, const Shape& rhs) { - bool equal = CompareShapes(lhs, rhs, /*compare_layouts=*/true); + bool equal = CompareShapes(lhs, rhs, /*compare_layouts=*/true, + /*ignore_fp_precision=*/false); if (!equal && VLOG_IS_ON(3)) { VLOG(3) << "ShapeUtil::Equal differ: lhs = " << lhs.ShortDebugString() << ", rhs = " << rhs.ShortDebugString(); @@ -179,6 +171,18 @@ StatusOr MakeShapeWithLayoutInternal( return equal; } +/* static */ bool ShapeUtil::EqualIgnoringFpPrecision(const Shape& lhs, + const Shape& rhs) { + bool equal = CompareShapes(lhs, rhs, /*compare_layouts=*/true, + /*ignore_fp_precision=*/true); + if (!equal && VLOG_IS_ON(3)) { + VLOG(3) << "ShapeUtil::EqualIgnoringFpPrecision differ: lhs = " + << lhs.ShortDebugString() << ", rhs = " << rhs.ShortDebugString(); + } + + return equal; +} + /* static */ int64 ShapeUtil::Rank(const Shape& shape) { CHECK(ShapeUtil::IsArray(shape)) << "Non-arrays do not have a rank, shape: " << shape; @@ -574,12 +578,11 @@ StatusOr ParseShapeStringInternal(tensorflow::StringPiece* s) { // tensorflow::StringPiece is not compatible with internal RE2 StringPiece, so // we convert in to the RE2-consumable type and then consume the corresponding // amount from our StringPiece type. + static LazyRE2 shape_pattern = { + "^(\\w*\\d*)\\[([\\d,]*)\\](?:\\s*(dense|sparse)?\\s*{([\\d,]+)})?"}; tensorflow::RegexpStringPiece s_consumable(s->data(), s->size()); - if (RE2::Consume( - &s_consumable, - "^(\\w*\\d*)\\[([\\d,]*)\\](?:\\s*(dense|sparse)?\\s*{([\\d,]+)})?", - &element_type_string, &dimensions_string, &format_string, - &layout_string)) { + if (RE2::Consume(&s_consumable, *shape_pattern, &element_type_string, + &dimensions_string, &format_string, &layout_string)) { size_t consumed = s->size() - s_consumable.size(); s->remove_prefix(consumed); auto string_to_int64 = [&s](const string& input) -> StatusOr { @@ -665,7 +668,8 @@ StatusOr ParseShapeStringInternal(tensorflow::StringPiece* s) { } /* static */ bool ShapeUtil::Compatible(const Shape& lhs, const Shape& rhs) { - return CompareShapes(lhs, rhs, /*compare_layouts=*/false); + return CompareShapes(lhs, rhs, /*compare_layouts=*/false, + /*ignore_fp_precision=*/false); } /* static */ bool ShapeUtil::CompatibleIgnoringElementType(const Shape& lhs, @@ -867,6 +871,60 @@ StatusOr ParseShapeStringInternal(tensorflow::StringPiece* s) { } } + TF_RETURN_IF_ERROR(ValidateShapeSize(shape)); + return Status::OK(); +} + +/* static */ Status ShapeUtil::ValidateShapeSize(const Shape& shape) { + VLOG(3) << "Validating shape size: " << ShapeUtil::HumanString(shape); + + if (!IsArray(shape)) { + return Status::OK(); + } + + int64 shape_size = [&shape]() { + int64 shape_size; + if (LayoutUtil::IsSparseArray(shape)) { + shape_size = LayoutUtil::MaxSparseElements(shape.layout()); + if (shape_size < 0) { + return shape_size; + } + shape_size = MultiplyWithoutOverflow(shape_size, ShapeUtil::Rank(shape)); + if (shape_size < 0) { + return shape_size; + } + shape_size = MultiplyWithoutOverflow(shape_size, sizeof(int64)); + if (shape_size < 0) { + return shape_size; + } + } + + shape_size = 1; + + // This is intentionally unconditional: even if the shape is sparse, we want + // to verify the densified version has a reasonable size. + if (shape.dimensions().empty()) { + return shape_size; + } + + for (int64 dim : shape.dimensions()) { + shape_size = MultiplyWithoutOverflow(shape_size, dim); + if (shape_size < 0) { + return shape_size; + } + } + shape_size = MultiplyWithoutOverflow( + shape_size, ByteSizeOfPrimitiveType(shape.element_type())); + + return shape_size; + }(); + + if (shape_size < 0) { + return InvalidArgument("Shape %s size may overflow int64.", + ShapeUtil::HumanString(shape).c_str()); + } + + VLOG(3) << "Shape size is valid: " << shape_size; return Status::OK(); } @@ -1054,12 +1112,41 @@ Status ForEachMutableSubshapeHelper( for (auto dim : Permute(permutation, shape.dimensions())) { new_shape.add_dimensions(dim); } + + // If `shape` has a layout, by contract we choose a new layout such that the + // transpose defined by this permutation is a bitcast. + // + // Some formalism helps to understand the correct way to do this. We're going + // to do algebra in the group of permutations of the dimensions of `shape`. + // + // Since the order of `shape`'s dimensions is not permuted relative to itself, + // `shape`'s list of dimensions is isomorphic to the identity I. + // + // Let `shape`'s layout be L. A layout is a permutation which maps a + // minor-to-major physical layout to the order of a shape's logical dims. + // Therefore inverse of a layout maps from logical to physical dims, and so + // the physical layout of I is simply L'.I = L', where L' is the inverse of L. + // + // Let the argument `permutation` be P. This is a permutation over `shape`'s + // dimensions, so our return value will be a shape with dims P.I = P. Our + // goal is to construct a layout permutation L* that we can apply to P such + // that that the physical dimension ordering of the returned shape is the same + // as that of the original shape, namely L'. + // + // Our returned shape has dims P and layout L*, so its in-memory layout is + // L*'.P. Setting this equal to L' and solving for L*, we get: + // + // L*'.P = L' => + // L*' = L'P' => + // L* = P.L + // if (shape.has_layout()) { CHECK(LayoutUtil::IsDenseArray(shape)); Layout* new_layout = new_shape.mutable_layout(); new_layout->set_format(DENSE); new_layout->clear_minor_to_major(); - for (auto index : Permute(permutation, shape.layout().minor_to_major())) { + for (auto index : ComposePermutations( + permutation, AsInt64Slice(shape.layout().minor_to_major()))) { new_layout->add_minor_to_major(index); } if (shape.layout().padded_dimensions_size() > 0) { @@ -1069,6 +1156,13 @@ Status ForEachMutableSubshapeHelper( new_layout->add_padded_dimensions(dim); } } + // The permutation accepted by TransposeIsBitcast is the inverse of the + // permutation here. + CHECK(TransposeIsBitcast(shape, new_shape, InversePermutation(permutation))) + << "shape=" << HumanStringWithLayout(shape) + << ", new_shape=" << HumanStringWithLayout(new_shape) + << ", permutation={" << tensorflow::str_util::Join(permutation, ",") + << "}"; } return new_shape; } diff --git a/tensorflow/compiler/xla/shape_util.h b/tensorflow/compiler/xla/shape_util.h index 02e4f41505f16de7369e1dbd712dd0756c3f28e7..17c1d7b10a70bece62cd4815e55138e49f182e58 100644 --- a/tensorflow/compiler/xla/shape_util.h +++ b/tensorflow/compiler/xla/shape_util.h @@ -110,31 +110,33 @@ class ShapeIndex { class ShapeIndexView { public: ShapeIndexView(const ShapeIndex& shape_index, int64 offset = 0) - : ShapeIndexView(shape_index.data() + offset, - shape_index.data() + shape_index.size()) { + : indices_(shape_index.data() + offset, shape_index.size() - offset) { CHECK_LE(offset, shape_index.size()); } - ShapeIndexView(std::initializer_list indices) - : ShapeIndexView(indices.begin(), indices.end()) {} + ShapeIndexView(std::initializer_list indices) : indices_(indices) {} ShapeIndexView(const ShapeIndexView& other) = default; using iterator = const int64*; - iterator begin() const { return begin_; } - iterator end() const { return end_; } - int64 size() const { return std::distance(begin_, end_); } - bool empty() const { return begin_ == end_; } + iterator begin() const { return indices_.begin(); } + iterator end() const { return indices_.end(); } + int64 size() const { return indices_.size(); } + bool empty() const { return indices_.empty(); } int64 front() const { CHECK(!empty()); - return *begin_; + return indices_.front(); } ShapeIndexView ConsumeFront() const { - CHECK(!empty()); - auto new_begin = begin_; - ++new_begin; - return ShapeIndexView(new_begin, end_); + ShapeIndexView result = *this; + result.indices_.pop_front(); + return result; } - ShapeIndex ToShapeIndex() const { return ShapeIndex(begin_, end_); } + ShapeIndexView ConsumeBack() const { + ShapeIndexView result = *this; + result.indices_.pop_back(); + return result; + } + ShapeIndex ToShapeIndex() const { return ShapeIndex(begin(), end()); } bool operator==(const ShapeIndexView& other) const; bool operator!=(const ShapeIndexView& other) const; @@ -142,10 +144,7 @@ class ShapeIndexView { string ToString() const; private: - ShapeIndexView(iterator begin, iterator end) : begin_(begin), end_(end) {} - - iterator begin_; - iterator end_; + tensorflow::gtl::ArraySlice indices_; }; std::ostream& operator<<(std::ostream& out, const ShapeIndex& shape_index); @@ -280,6 +279,9 @@ class ShapeUtil { // Returns whether the lhs and rhs shapes are identical protobufs. static bool Equal(const Shape& lhs, const Shape& rhs); + // As Equal, but allow one of lhs and rhs to be F16 while the other is F32. + static bool EqualIgnoringFpPrecision(const Shape& lhs, const Shape& rhs); + // Returns the rank (number of dimensions) of the given shape. // Precondition: !IsTuple(shape) static int64 Rank(const Shape& shape); @@ -527,7 +529,13 @@ class ShapeUtil { static bool HasDegenerateDimensions(const Shape& shape); // Permutes the dimensions by the given permutation, so - // return_value.dimensions[permutation[i]] = argument.dimensions[i] + // return_value.dimensions[permutation[i]] = argument.dimensions[i]. + // + // Postcondition: For any valid permutation, + // + // !HasLayout(shape) || + // TransposeIsBitcast(shape, PermuteDimensions(permutation, shape), + // InversePermutation(permutation)). static Shape PermuteDimensions(tensorflow::gtl::ArraySlice permutation, const Shape& shape); @@ -699,6 +707,10 @@ class ShapeUtil { static size_t Hash(const Shape& shape); private: + // Validates the shape size is sane. This makes sure it's safe to do + // calculations in int64 without overflowing. + static Status ValidateShapeSize(const Shape& shape); + // Validates all of the non-layout properties of the shape -- this is a helper // used by both the layout-optional and layout-required public method. static Status ValidateShapeWithOptionalLayoutInternal(const Shape& shape); diff --git a/tensorflow/compiler/xla/shape_util_test.cc b/tensorflow/compiler/xla/shape_util_test.cc index 606f7492cead5c3b6772625612fec67296740c7f..ed2d16c0e90685d6cb2603eba0a4cf880046aa18 100644 --- a/tensorflow/compiler/xla/shape_util_test.cc +++ b/tensorflow/compiler/xla/shape_util_test.cc @@ -15,6 +15,7 @@ limitations under the License. #include "tensorflow/compiler/xla/shape_util.h" +#include #include "tensorflow/compiler/xla/layout_util.h" #include "tensorflow/compiler/xla/status_macros.h" #include "tensorflow/compiler/xla/test.h" @@ -22,12 +23,23 @@ limitations under the License. #include "tensorflow/compiler/xla/types.h" #include "tensorflow/compiler/xla/util.h" #include "tensorflow/compiler/xla/xla_data.pb.h" +#include "tensorflow/core/lib/strings/str_util.h" +#include "tensorflow/core/lib/strings/strcat.h" namespace xla { namespace { using ::testing::ElementsAre; +TEST(ShapeUtilTest, ShapeIndexViewTest) { + ShapeIndex index = {1, 2, 3, 4}; + ShapeIndexView index_view(index, 1); + EXPECT_EQ(3, index_view.size()); + EXPECT_EQ(ShapeIndexView({2, 3, 4}), index_view); + EXPECT_EQ(ShapeIndexView({3, 4}), index_view.ConsumeFront()); + EXPECT_EQ(ShapeIndexView({2, 3}), index_view.ConsumeBack()); +} + TEST(ShapeUtilTest, GetDimensionHelperCanNegativeIndex) { Shape matrix = ShapeUtil::MakeShape(F32, {2, 3}); EXPECT_EQ(3, ShapeUtil::GetDimension(matrix, -1)); @@ -242,6 +254,24 @@ TEST(ShapeUtilTest, IncompatibleDifferentElementShapes) { EXPECT_FALSE(ShapeUtil::Compatible(shape_1, shape_2)); } +TEST(ShapeUtilTest, EqualIgnoringFpPrecision) { + EXPECT_TRUE(ShapeUtil::EqualIgnoringFpPrecision( + ShapeUtil::MakeShapeWithLayout(F32, {4, 3}, {0, 1}), + ShapeUtil::MakeShapeWithLayout(F16, {4, 3}, {0, 1}))); +} + +TEST(ShapeUtilTest, UnequalIgnoringFpPrecision) { + EXPECT_FALSE(ShapeUtil::EqualIgnoringFpPrecision( + ShapeUtil::MakeShapeWithLayout(F32, {4, 3}, {0, 1}), + ShapeUtil::MakeShapeWithLayout(F16, {3, 4}, {0, 1}))); + EXPECT_FALSE(ShapeUtil::EqualIgnoringFpPrecision( + ShapeUtil::MakeShapeWithLayout(F32, {3, 4}, {0, 1}), + ShapeUtil::MakeShapeWithLayout(F16, {3, 4}, {1, 0}))); + EXPECT_FALSE(ShapeUtil::EqualIgnoringFpPrecision( + ShapeUtil::MakeShapeWithLayout(F32, {4, 3}, {0, 1}), + ShapeUtil::MakeShapeWithLayout(PRED, {4, 3}, {0, 1}))); +} + TEST(ShapeUtilTest, CompatibleTuples) { Shape tuple1 = ShapeUtil::MakeTupleShape( {ShapeUtil::MakeShape(F32, {3, 2}), ShapeUtil::MakeShape(PRED, {4, 5})}); @@ -803,6 +833,28 @@ TEST(ShapeUtilTest, HasDegenerateDimensions) { ShapeUtil::HasDegenerateDimensions(ShapeUtil::MakeShape(F32, {3, 0, 5}))); } +TEST(ShapeUtilTest, PermuteDimensionsLayout) { + std::vector layout(3); + std::iota(layout.begin(), layout.end(), 0); + do { + Shape s = ShapeUtil::MakeShapeWithLayout(F32, {10, 100, 1000}, layout); + SCOPED_TRACE(tensorflow::strings::StrCat("s=", ShapeUtil::HumanString(s))); + + std::vector permutation(3); + std::iota(permutation.begin(), permutation.end(), 0); + do { + SCOPED_TRACE(tensorflow::strings::StrCat( + "permutation=", tensorflow::str_util::Join(permutation, ","))); + + // TransposeIsBitcast takes the inverse of the permutation that + // PermuteDimensions takes. + EXPECT_TRUE(ShapeUtil::TransposeIsBitcast( + s, ShapeUtil::PermuteDimensions(permutation, s), + InversePermutation(permutation))); + } while (std::next_permutation(permutation.begin(), permutation.end())); + } while (std::next_permutation(layout.begin(), layout.end())); +} + TEST(AlgebraicSimplifierTest, ReshapeIsBitcast_3x2x2_6x2_Dim0IsMostMinor) { EXPECT_FALSE(ShapeUtil::ReshapeIsBitcast( ShapeUtil::MakeShapeWithLayout(F32, {3, 2, 2}, {0, 1, 2}), diff --git a/tensorflow/compiler/xla/statusor.h b/tensorflow/compiler/xla/statusor.h index 0e1387c93938fa520562fcd63ac107a82b089a51..a32e2ad9851b0b5644f7e6f0f9ead6c438934c07 100644 --- a/tensorflow/compiler/xla/statusor.h +++ b/tensorflow/compiler/xla/statusor.h @@ -12,297 +12,17 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ - -// StatusOr is the union of a Status object and a T object. StatusOr models -// the concept of an object that is either a value, or an error Status -// explaining why such a value is not present. To this end, StatusOr does not -// allow its Status value to be Status::OK. -// -// The primary use-case for StatusOr is as the return value of a -// function which may fail. -// -// Example client usage for a StatusOr, where T is not a pointer: -// -// StatusOr result = DoBigCalculationThatCouldFail(); -// if (result.ok()) { -// float answer = result.ValueOrDie(); -// printf("Big calculation yielded: %f", answer); -// } else { -// LOG(ERROR) << result.status(); -// } -// -// Example client usage for a StatusOr: -// -// StatusOr result = FooFactory::MakeNewFoo(arg); -// if (result.ok()) { -// std::unique_ptr foo(result.ValueOrDie()); -// foo->DoSomethingCool(); -// } else { -// LOG(ERROR) << result.status(); -// } -// -// Example client usage for a StatusOr>: -// -// StatusOr> result = FooFactory::MakeNewFoo(arg); -// if (result.ok()) { -// std::unique_ptr foo = std::move(result.ValueOrDie()); -// foo->DoSomethingCool(); -// } else { -// LOG(ERROR) << result.status(); -// } -// -// Example factory implementation returning StatusOr: -// -// StatusOr FooFactory::MakeNewFoo(int arg) { -// if (arg <= 0) { -// return tensorflow::InvalidArgument("Arg must be positive"); -// } else { -// return new Foo(arg); -// } -// } -// -// Note that the assignment operators require that destroying the currently -// stored value cannot invalidate the argument; in other words, the argument -// cannot be an alias for the current value, or anything owned by the current -// value. #ifndef TENSORFLOW_COMPILER_XLA_STATUSOR_H_ #define TENSORFLOW_COMPILER_XLA_STATUSOR_H_ #include "tensorflow/compiler/xla/status.h" -#include "tensorflow/compiler/xla/statusor_internals.h" -#include "tensorflow/core/platform/macros.h" +#include "tensorflow/stream_executor/lib/statusor.h" namespace xla { -#if defined(__clang__) -// Only clang supports warn_unused_result as a type annotation. -template -class TF_MUST_USE_RESULT StatusOr; -#endif - -template -class StatusOr : private internal_statusor::StatusOrData, - private internal_statusor::TraitsBase< - std::is_copy_constructible::value, - std::is_move_constructible::value> { - template - friend class StatusOr; - - typedef internal_statusor::StatusOrData Base; - - public: - typedef T element_type; - - // Constructs a new StatusOr with Status::UNKNOWN status. This is marked - // 'explicit' to try to catch cases like 'return {};', where people think - // StatusOr> will be initialized with an empty vector, - // instead of a Status::UNKNOWN status. - explicit StatusOr(); - - // StatusOr will be copy constructible/assignable if T is copy - // constructible. - StatusOr(const StatusOr&) = default; - StatusOr& operator=(const StatusOr&) = default; - - // StatusOr will be move constructible/assignable if T is move - // constructible. - StatusOr(StatusOr&&) = default; - StatusOr& operator=(StatusOr&&) = default; - - // Conversion copy/move constructor, T must be convertible from U. - template ::value>::type* = nullptr> - StatusOr(const StatusOr& other); - template ::value>::type* = nullptr> - StatusOr(StatusOr&& other); - - // Conversion copy/move assignment operator, T must be convertible from U. - template ::value>::type* = nullptr> - StatusOr& operator=(const StatusOr& other); - template ::value>::type* = nullptr> - StatusOr& operator=(StatusOr&& other); - - // Constructs a new StatusOr with the given value. After calling this - // constructor, calls to ValueOrDie() will succeed, and calls to status() will - // return OK. - // - // NOTE: Not explicit - we want to use StatusOr as a return type - // so it is convenient and sensible to be able to do 'return T()' - // when the return type is StatusOr. - // - // REQUIRES: T is copy constructible. - StatusOr(const T& value); - - // Constructs a new StatusOr with the given non-ok status. After calling - // this constructor, calls to ValueOrDie() will CHECK-fail. - // - // NOTE: Not explicit - we want to use StatusOr as a return - // value, so it is convenient and sensible to be able to do 'return - // Status()' when the return type is StatusOr. - // - // REQUIRES: !status.ok(). This requirement is DCHECKed. - // In optimized builds, passing Status::OK() here will have the effect - // of passing tensorflow::error::INTERNAL as a fallback. - StatusOr(const Status& status); - StatusOr& operator=(const Status& status); - - // TODO(b/62186997): Add operator=(T) overloads. - - // Similar to the `const T&` overload. - // - // REQUIRES: T is move constructible. - StatusOr(T&& value); - - // RValue versions of the operations declared above. - StatusOr(Status&& status); - StatusOr& operator=(Status&& status); - - // Returns this->status().ok() - bool ok() const { return this->status_.ok(); } - - // Returns a reference to our status. If this contains a T, then - // returns Status::OK(). - const Status& status() const &; - Status status() &&; - - // Returns a reference to our current value, or CHECK-fails if !this->ok(). - // - // Note: for value types that are cheap to copy, prefer simple code: - // - // T value = statusor.ValueOrDie(); - // - // Otherwise, if the value type is expensive to copy, but can be left - // in the StatusOr, simply assign to a reference: - // - // T& value = statusor.ValueOrDie(); // or `const T&` - // - // Otherwise, if the value type supports an efficient move, it can be - // used as follows: - // - // T value = std::move(statusor).ValueOrDie(); - // - // The std::move on statusor instead of on the whole expression enables - // warnings about possible uses of the statusor object after the move. - // C++ style guide waiver for ref-qualified overloads granted in cl/143176389 - // See go/ref-qualifiers for more details on such overloads. - const T& ValueOrDie() const &; - T& ValueOrDie() &; - const T&& ValueOrDie() const &&; - T&& ValueOrDie() &&; - - T ConsumeValueOrDie() { return std::move(ValueOrDie()); } - - // Ignores any errors. This method does nothing except potentially suppress - // complaints from any tools that are checking that errors are not dropped on - // the floor. - void IgnoreError() const; -}; - -//////////////////////////////////////////////////////////////////////////////// -// Implementation details for StatusOr - -template -StatusOr::StatusOr() : Base(Status(tensorflow::error::UNKNOWN, "")) {} - -template -StatusOr::StatusOr(const T& value) : Base(value) {} - -template -StatusOr::StatusOr(const Status& status) : Base(status) {} - -template -StatusOr& StatusOr::operator=(const Status& status) { - this->Assign(status); - return *this; -} - -template -StatusOr::StatusOr(T&& value) : Base(std::move(value)) {} - -template -StatusOr::StatusOr(Status&& status) : Base(std::move(status)) {} - -template -StatusOr& StatusOr::operator=(Status&& status) { - this->Assign(std::move(status)); - return *this; -} - -template -template ::value>::type*> -inline StatusOr::StatusOr(const StatusOr& other) - : Base(static_cast::Base&>(other)) {} - -template -template ::value>::type*> -inline StatusOr& StatusOr::operator=(const StatusOr& other) { - if (other.ok()) - this->Assign(other.ValueOrDie()); - else - this->Assign(other.status()); - return *this; -} - -template -template ::value>::type*> -inline StatusOr::StatusOr(StatusOr&& other) - : Base(static_cast::Base&&>(other)) {} - -template -template ::value>::type*> -inline StatusOr& StatusOr::operator=(StatusOr&& other) { - if (other.ok()) { - this->Assign(std::move(other).ValueOrDie()); - } else { - this->Assign(std::move(other).status()); - } - return *this; -} - -template -const Status& StatusOr::status() const & { - return this->status_; -} -template -Status StatusOr::status() && { - return ok() ? Status::OK() : std::move(this->status_); -} - -template -const T& StatusOr::ValueOrDie() const & { - this->EnsureOk(); - return this->data_; -} - -template -T& StatusOr::ValueOrDie() & { - this->EnsureOk(); - return this->data_; -} - -template -const T&& StatusOr::ValueOrDie() const && { - this->EnsureOk(); - return std::move(this->data_); -} - -template -T&& StatusOr::ValueOrDie() && { - this->EnsureOk(); - return std::move(this->data_); -} - +// Use steam_executor's StatusOr so we don't duplicate code. template -void StatusOr::IgnoreError() const { - // no-op -} +using StatusOr = ::stream_executor::port::StatusOr; } // namespace xla diff --git a/tensorflow/compiler/xla/tests/BUILD b/tensorflow/compiler/xla/tests/BUILD index b76830f6662ba3bda2b0c64c88cd74f0a13f75f0..6a75aa6794e617e57e40b9a3cef74a9af74b91d3 100644 --- a/tensorflow/compiler/xla/tests/BUILD +++ b/tensorflow/compiler/xla/tests/BUILD @@ -65,6 +65,7 @@ cc_library( srcs = ["test_utils.cc"], hdrs = ["test_utils.h"], deps = [ + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:literal_util", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:util", @@ -88,6 +89,7 @@ cc_library( "//tensorflow/compiler/xla:array3d", "//tensorflow/compiler/xla:array4d", "//tensorflow/compiler/xla:error_spec", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:literal_comparison", "//tensorflow/compiler/xla:literal_util", "//tensorflow/compiler/xla:test", @@ -179,6 +181,7 @@ cc_library( "//tensorflow/compiler/xla:array3d", "//tensorflow/compiler/xla:array4d", "//tensorflow/compiler/xla:execution_options_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:literal_util", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:status_macros", @@ -209,6 +212,7 @@ cc_library( deps = [ ":codegen_test_base", ":filecheck", + "//tensorflow/compiler/xla/service:hlo_parser", "//tensorflow/compiler/xla/service:llvm_compiler", "//tensorflow/compiler/xla/service/llvm_ir:llvm_util", "//tensorflow/core:test", @@ -302,7 +306,7 @@ xla_test( "enable_for_xla_interpreter", ], deps = [ - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:statusor", "//tensorflow/compiler/xla:test", @@ -345,7 +349,7 @@ xla_test( "enable_for_xla_interpreter", ], deps = [ - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:status_macros", "//tensorflow/compiler/xla:statusor", @@ -406,7 +410,7 @@ xla_test( tags = ["enable_for_xla_interpreter"], deps = [ "//tensorflow/compiler/xla:array2d", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:statusor", "//tensorflow/compiler/xla:test", @@ -435,7 +439,7 @@ xla_test( tags = ["optonly"], deps = [ "//tensorflow/compiler/xla:array2d", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:statusor", "//tensorflow/compiler/xla:xla_data_proto", @@ -531,6 +535,7 @@ xla_test( srcs = ["scalar_computations_test.cc"], shard_count = 32, deps = [ + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:literal_util", "//tensorflow/compiler/xla:status_macros", "//tensorflow/compiler/xla:statusor", @@ -573,7 +578,7 @@ xla_test( "enable_for_xla_interpreter", ], deps = [ - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:statusor", "//tensorflow/compiler/xla:test", @@ -599,7 +604,7 @@ xla_test( "//tensorflow/compiler/xla:array2d", "//tensorflow/compiler/xla:array3d", "//tensorflow/compiler/xla:array4d", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:statusor", "//tensorflow/compiler/xla:test", @@ -645,7 +650,7 @@ xla_test( tags = ["enable_for_xla_interpreter"], deps = [ "//tensorflow/compiler/xla:array2d", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:statusor", "//tensorflow/compiler/xla:test", @@ -697,6 +702,7 @@ xla_test( "//tensorflow/compiler/xla:execution_options_util", "//tensorflow/compiler/xla:status_macros", "//tensorflow/compiler/xla:test", + "//tensorflow/compiler/xla/client/xla_client:xla_builder", "//tensorflow/compiler/xla/service:hlo_parser", "//tensorflow/compiler/xla/tests:xla_internal_test_main", ], @@ -763,6 +769,7 @@ xla_test( "//tensorflow/compiler/xla:array2d", "//tensorflow/compiler/xla:array3d", "//tensorflow/compiler/xla:array4d", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:literal_util", "//tensorflow/compiler/xla:xla_data_proto", "//tensorflow/compiler/xla/client:local_client", @@ -779,7 +786,7 @@ xla_test( CONVOLUTION_TEST_DEPS = [ "//tensorflow/compiler/xla:array2d", "//tensorflow/compiler/xla:array4d", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:reference_util", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:statusor", @@ -826,7 +833,7 @@ xla_test( deps = [ "//tensorflow/compiler/xla:array3d", "//tensorflow/compiler/xla:array4d", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:reference_util", "//tensorflow/compiler/xla:xla_data_proto", "//tensorflow/compiler/xla/client:local_client", @@ -873,7 +880,7 @@ xla_test( ":test_utils", "//tensorflow/compiler/xla:array2d", "//tensorflow/compiler/xla:array4d", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:reference_util", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:status_macros", @@ -885,6 +892,7 @@ xla_test( "//tensorflow/compiler/xla/client:global_data", "//tensorflow/compiler/xla/client:local_client", "//tensorflow/compiler/xla/client/lib:arithmetic", + "//tensorflow/compiler/xla/client/lib:math", "//tensorflow/compiler/xla/client/xla_client:xla_builder", "//tensorflow/compiler/xla/client/xla_client:xla_computation", "//tensorflow/compiler/xla/service:hlo", @@ -905,7 +913,7 @@ xla_test( ":test_utils", "//tensorflow/compiler/xla:array2d", "//tensorflow/compiler/xla:array4d", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:reference_util", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:status_macros", @@ -938,7 +946,7 @@ xla_test( ], deps = [ ":test_utils", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:statusor", "//tensorflow/compiler/xla:test", "//tensorflow/compiler/xla:test_helpers", @@ -1029,6 +1037,7 @@ xla_test( ], deps = [ "//tensorflow/compiler/xla:array2d", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:literal_util", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:statusor", @@ -1077,6 +1086,7 @@ xla_test( deps = [ "//tensorflow/compiler/xla:array2d", "//tensorflow/compiler/xla:array4d", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:literal_util", "//tensorflow/compiler/xla:reference_util", "//tensorflow/compiler/xla:shape_util", @@ -1147,7 +1157,7 @@ xla_test( ], deps = [ "//tensorflow/compiler/xla:array2d", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:reference_util", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:status_macros", @@ -1174,7 +1184,7 @@ xla_test( deps = [ ":client_library_test_base", "//tensorflow/compiler/xla:array2d", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:util", "//tensorflow/compiler/xla:xla_data_proto", "//tensorflow/compiler/xla/client/xla_client:xla_builder", @@ -1226,6 +1236,7 @@ xla_test( "enable_for_xla_interpreter", ], deps = [ + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:literal_util", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:test_helpers", @@ -1244,10 +1255,12 @@ xla_test( name = "custom_call_test", srcs = ["custom_call_test.cc"], deps = [ + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:literal_util", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:util", "//tensorflow/compiler/xla:xla_data_proto", + "//tensorflow/compiler/xla/client/xla_client:xla_builder", "//tensorflow/compiler/xla/service:hlo", "//tensorflow/compiler/xla/service/cpu:custom_call_target_registry", "//tensorflow/compiler/xla/tests:client_library_test_base", @@ -1288,6 +1301,7 @@ xla_test( deps = [ "//tensorflow/compiler/xla:array2d", "//tensorflow/compiler/xla:array4d", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:literal_util", "//tensorflow/compiler/xla:statusor", "//tensorflow/compiler/xla:test", @@ -1365,7 +1379,7 @@ xla_test( ], deps = [ "//tensorflow/compiler/xla:array2d", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:reference_util", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:statusor", @@ -1388,7 +1402,7 @@ xla_test( name = "prng_test", srcs = ["prng_test.cc"], deps = [ - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:test", "//tensorflow/compiler/xla:util", @@ -1413,6 +1427,7 @@ xla_test( deps = [ "//tensorflow/compiler/xla:array2d", "//tensorflow/compiler/xla:array4d", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:literal_util", "//tensorflow/compiler/xla:reference_util", "//tensorflow/compiler/xla:shape_util", @@ -1527,7 +1542,7 @@ xla_test( name = "cross_replica_sum_test", srcs = ["cross_replica_sum_test.cc"], deps = [ - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:test", "//tensorflow/compiler/xla:test_helpers", @@ -1571,7 +1586,7 @@ xla_test( name = "compilation_cache_test", srcs = ["compilation_cache_test.cc"], deps = [ - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:statusor", "//tensorflow/compiler/xla:xla_data_proto", @@ -1611,7 +1626,7 @@ xla_test( name = "compute_constant_test", srcs = ["compute_constant_test.cc"], deps = [ - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:status_macros", "//tensorflow/compiler/xla:statusor", @@ -1686,7 +1701,7 @@ xla_test( "enable_for_xla_interpreter", ], deps = [ - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:protobuf_util", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:statusor", @@ -1711,7 +1726,7 @@ xla_test( "enable_for_xla_interpreter", ], deps = [ - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:util", "//tensorflow/compiler/xla:xla_data_proto", @@ -1728,6 +1743,7 @@ tf_cc_test( srcs = ["llvm_compiler_test.cc"], tags = ["requires-gpu-sm35"], deps = [ + "//tensorflow/compiler/xla:literal_util", "//tensorflow/compiler/xla:test_helpers", "//tensorflow/compiler/xla/service:backend", "//tensorflow/compiler/xla/service:cpu_plugin", @@ -1748,7 +1764,7 @@ xla_test( name = "round_trip_packed_literal_test", srcs = ["round_trip_packed_literal_test.cc"], deps = [ - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:packed_literal_reader", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:statusor", @@ -1771,7 +1787,7 @@ xla_test( ], deps = [ "//tensorflow/compiler/xla:array2d", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:util", "//tensorflow/compiler/xla:xla_data_proto", @@ -1780,6 +1796,7 @@ xla_test( "//tensorflow/compiler/xla/client/xla_client:xla_builder", "//tensorflow/compiler/xla/client/xla_client:xla_computation", "//tensorflow/compiler/xla/service:hlo", + "//tensorflow/compiler/xla/service:hlo_parser", "//tensorflow/compiler/xla/service:hlo_runner", "//tensorflow/compiler/xla/service:platform_util", "//tensorflow/compiler/xla/tests:client_library_test_base", @@ -1798,7 +1815,7 @@ xla_test( srcs = ["multioutput_fusion_test.cc"], deps = [ "//tensorflow/compiler/xla:array2d", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:util", "//tensorflow/compiler/xla:xla_data_proto", @@ -1838,7 +1855,7 @@ xla_test( name = "local_client_allocation_test", srcs = ["local_client_allocation_test.cc"], deps = [ - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:statusor", "//tensorflow/compiler/xla/client:local_client", "//tensorflow/compiler/xla/client/xla_client:xla_builder", @@ -1861,7 +1878,7 @@ xla_test( shard_count = 30, tags = ["optonly"], deps = [ - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:statusor", "//tensorflow/compiler/xla:test", @@ -1907,7 +1924,7 @@ xla_test( srcs = ["round_trip_transfer_test.cc"], deps = [ "//tensorflow/compiler/xla:array4d", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:statusor", "//tensorflow/compiler/xla:xla_data_proto", @@ -1928,7 +1945,7 @@ xla_test( deps = [ "//tensorflow/compiler/xla:array2d", "//tensorflow/compiler/xla:array4d", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:reference_util", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:status_macros", @@ -1976,7 +1993,7 @@ xla_test( ":literal_test_util", ":local_client_test_base", ":xla_internal_test_main", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:statusor", "//tensorflow/compiler/xla:types", @@ -2038,6 +2055,7 @@ xla_test( "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla/client/xla_client:xla_builder", "//tensorflow/compiler/xla/client/xla_client:xla_computation", + "//tensorflow/compiler/xla/service:hlo_parser", "//tensorflow/compiler/xla/tests:xla_internal_test_main", "//tensorflow/core:test", ], diff --git a/tensorflow/compiler/xla/tests/array_elementwise_ops_test.cc b/tensorflow/compiler/xla/tests/array_elementwise_ops_test.cc index 8ac771ae5a01487a556f007418a6254c61d56ba1..3ae96fa1bcb1057653a75db62def5556ae37f886 100644 --- a/tensorflow/compiler/xla/tests/array_elementwise_ops_test.cc +++ b/tensorflow/compiler/xla/tests/array_elementwise_ops_test.cc @@ -26,7 +26,7 @@ limitations under the License. #include "tensorflow/compiler/xla/client/local_client.h" #include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/compiler/xla/layout_util.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/statusor.h" #include "tensorflow/compiler/xla/test.h" #include "tensorflow/compiler/xla/tests/client_library_test_base.h" @@ -51,16 +51,16 @@ class ArrayElementwiseOpTestParamCount XLA_TEST_F(ArrayElementwiseOpTest, NegConstantZeroElementF32) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({}); - builder.Neg(a); + auto a = ConstantR1(&builder, {}); + Neg(a); ComputeAndCompareR1(&builder, {}, {}, error_spec_); } XLA_TEST_F(ArrayElementwiseOpTest, NegConstantF32) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({-2.5f, 3.14f, 2.25f, -10.0f, 6.0f}); - builder.Neg(a); + auto a = ConstantR1(&builder, {-2.5f, 3.14f, 2.25f, -10.0f, 6.0f}); + Neg(a); ComputeAndCompareR1(&builder, {2.5f, -3.14f, -2.25f, 10.0f, -6.0f}, {}, error_spec_); @@ -68,10 +68,10 @@ XLA_TEST_F(ArrayElementwiseOpTest, NegConstantF32) { XLA_TEST_F(ArrayElementwiseOpTest, NegConstantS32) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({-1, 0, 1, 324, - std::numeric_limits::min(), - std::numeric_limits::max()}); - builder.Neg(a); + auto a = ConstantR1(&builder, + {-1, 0, 1, 324, std::numeric_limits::min(), + std::numeric_limits::max()}); + Neg(a); // -min == min for int32 due to an overflow. In C++ it is undefined behavior // to do this calculation. For XLA we have not specified that, so it @@ -84,17 +84,17 @@ XLA_TEST_F(ArrayElementwiseOpTest, NegConstantS32) { XLA_TEST_F(ArrayElementwiseOpTest, NegConstantZeroElementC64) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({}); - builder.Neg(a); + auto a = ConstantR1(&builder, {}); + Neg(a); ComputeAndCompareR1(&builder, {}, {}, error_spec_); } XLA_TEST_F(ArrayElementwiseOpTest, NegConstantC64) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1( - {{-2.5f, 1.0f}, {0.0f, 3.14f}, {2.25f, -1.0f}, {-10.0f, 0.0f}}); - builder.Neg(a); + auto a = ConstantR1( + &builder, {{-2.5f, 1.0f}, {0.0f, 3.14f}, {2.25f, -1.0f}, {-10.0f, 0.0f}}); + Neg(a); ComputeAndCompareR1( &builder, {{2.5f, -1.0f}, {0.0f, -3.14f}, {-2.25f, 1.0f}, {10.0f, 0.0f}}, @@ -103,16 +103,17 @@ XLA_TEST_F(ArrayElementwiseOpTest, NegConstantC64) { XLA_TEST_F(ArrayElementwiseOpTest, NegConstantS64) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({ - -1, - 1, - 0, - 0x12345678, - static_cast(0xffffffff12345678l), - static_cast(0x8000000000000000LL), - static_cast(0x8000000000000001LL), - }); - builder.Neg(a); + auto a = + ConstantR1(&builder, { + -1, + 1, + 0, + 0x12345678, + static_cast(0xffffffff12345678l), + static_cast(0x8000000000000000LL), + static_cast(0x8000000000000001LL), + }); + Neg(a); LOG(INFO) << -static_cast(0x7FFFFFFFFFFFFFFFLL); ComputeAndCompareR1(&builder, @@ -130,8 +131,8 @@ XLA_TEST_F(ArrayElementwiseOpTest, NegConstantS64) { XLA_TEST_F(ArrayElementwiseOpTest, IsFiniteZeroElementF32s) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({}); - builder.IsFinite(a); + auto a = ConstantR1(&builder, {}); + IsFinite(a); ComputeAndCompareR1(&builder, {}, {}); } @@ -141,21 +142,21 @@ static const float kNonCanonicalNaN = tensorflow::bit_cast(0x7FD01234); XLA_TEST_F(ArrayElementwiseOpTest, IsFiniteScalarF32) { XlaBuilder builder(TestName()); - builder.IsFinite(builder.ConstantR0(NAN)); + IsFinite(ConstantR0(&builder, NAN)); ComputeAndCompareR0(&builder, false, {}); EXPECT_TRUE(std::isnan(kNonCanonicalNaN)); - builder.IsFinite(builder.ConstantR0(kNonCanonicalNaN)); + IsFinite(ConstantR0(&builder, kNonCanonicalNaN)); ComputeAndCompareR0(&builder, false, {}); const float inf = std::numeric_limits::infinity(); - builder.IsFinite(builder.ConstantR0(inf)); + IsFinite(ConstantR0(&builder, inf)); ComputeAndCompareR0(&builder, false, {}); - builder.IsFinite(builder.ConstantR0(-inf)); + IsFinite(ConstantR0(&builder, -inf)); ComputeAndCompareR0(&builder, false, {}); - builder.IsFinite(builder.ConstantR0(0.0f)); + IsFinite(ConstantR0(&builder, 0.0f)); ComputeAndCompareR0(&builder, true, {}); } @@ -163,9 +164,9 @@ XLA_TEST_F(ArrayElementwiseOpTest, IsFiniteR1F32s) { XlaBuilder builder(TestName()); const float inf = std::numeric_limits::infinity(); EXPECT_TRUE(std::isnan(kNonCanonicalNaN)); - auto a = builder.ConstantR1( - {{NAN, 7.0f, kNonCanonicalNaN, -1.0f, inf, -inf}}); - builder.IsFinite(a); + auto a = ConstantR1(&builder, + {{NAN, 7.0f, kNonCanonicalNaN, -1.0f, inf, -inf}}); + IsFinite(a); ComputeAndCompareR1(&builder, {false, true, false, true, false, false}, {}); @@ -173,9 +174,9 @@ XLA_TEST_F(ArrayElementwiseOpTest, IsFiniteR1F32s) { XLA_TEST_F(ArrayElementwiseOpTest, AddTwoConstantF32s) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({-2.5f, 3.14f, 2.25f, -10.0f, 6.0f}); - auto b = builder.ConstantR1({100.0f, 3.13f, 2.75f, 10.5f, -999.0f}); - builder.Add(a, b); + auto a = ConstantR1(&builder, {-2.5f, 3.14f, 2.25f, -10.0f, 6.0f}); + auto b = ConstantR1(&builder, {100.0f, 3.13f, 2.75f, 10.5f, -999.0f}); + Add(a, b); ComputeAndCompareR1(&builder, {97.5f, 6.27f, 5.0f, 0.5f, -993.0f}, {}, error_spec_); @@ -183,20 +184,20 @@ XLA_TEST_F(ArrayElementwiseOpTest, AddTwoConstantF32s) { XLA_TEST_F(ArrayElementwiseOpTest, AddTwoConstantZeroElementF32s) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({}); - auto b = builder.ConstantR1({}); - builder.Add(a, b); + auto a = ConstantR1(&builder, {}); + auto b = ConstantR1(&builder, {}); + Add(a, b); ComputeAndCompareR1(&builder, {}, {}, error_spec_); } XLA_TEST_F(ArrayElementwiseOpTest, AddTwoConstantC64s) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1( - {{-2.5f, 0.0f}, {0.0f, 3.14f}, {2.25f, 0.0f}, {1.0f, -10.0f}}); - auto b = builder.ConstantR1( - {{100.0f, 0.0f}, {3.13f, 0.0f}, {2.75f, 1.0f}, {-2.0f, 10.5f}}); - builder.Add(a, b); + auto a = ConstantR1( + &builder, {{-2.5f, 0.0f}, {0.0f, 3.14f}, {2.25f, 0.0f}, {1.0f, -10.0f}}); + auto b = ConstantR1( + &builder, {{100.0f, 0.0f}, {3.13f, 0.0f}, {2.75f, 1.0f}, {-2.0f, 10.5f}}); + Add(a, b); ComputeAndCompareR1( &builder, {97.5f, {3.13f, 3.14f}, {5.0f, 1.0f}, {-1.0f, 0.5f}}, {}, @@ -205,9 +206,9 @@ XLA_TEST_F(ArrayElementwiseOpTest, AddTwoConstantC64s) { XLA_TEST_F(ArrayElementwiseOpTest, AddTwoConstantZeroElementC64s) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({}); - auto b = builder.ConstantR1({}); - builder.Add(a, b); + auto a = ConstantR1(&builder, {}); + auto b = ConstantR1(&builder, {}); + Add(a, b); ComputeAndCompareR1(&builder, {}, {}, error_spec_); } @@ -224,8 +225,8 @@ XLA_TEST_F(ArrayElementwiseOpTest, AddTwoConstantU64s) { 0x8000000000000000LL, 0x8000000000000000LL, 1}; - std::unique_ptr lhs_literal = Literal::CreateR1({lhs}); - auto lhs_param = b.Parameter(0, lhs_literal->shape(), "lhs_param"); + std::unique_ptr lhs_literal = LiteralUtil::CreateR1({lhs}); + auto lhs_param = Parameter(&b, 0, lhs_literal->shape(), "lhs_param"); std::unique_ptr lhs_data = client_->TransferToServer(*lhs_literal).ConsumeValueOrDie(); @@ -238,12 +239,12 @@ XLA_TEST_F(ArrayElementwiseOpTest, AddTwoConstantU64s) { 0, 1, 0x8000000000000000LL}; - std::unique_ptr rhs_literal = Literal::CreateR1({rhs}); - auto rhs_param = b.Parameter(1, rhs_literal->shape(), "rhs_param"); + std::unique_ptr rhs_literal = LiteralUtil::CreateR1({rhs}); + auto rhs_param = Parameter(&b, 1, rhs_literal->shape(), "rhs_param"); std::unique_ptr rhs_data = client_->TransferToServer(*rhs_literal).ConsumeValueOrDie(); - b.Add(lhs_param, rhs_param); + Add(lhs_param, rhs_param); std::vector expected(lhs.size()); for (int64 i = 0; i < lhs.size(); ++i) { @@ -264,8 +265,8 @@ XLA_TEST_F(ArrayElementwiseOpTest, SubTwoConstantS64s) { 1, 0, -1}; - std::unique_ptr lhs_literal = Literal::CreateR1({lhs}); - auto lhs_param = b.Parameter(0, lhs_literal->shape(), "lhs_param"); + std::unique_ptr lhs_literal = LiteralUtil::CreateR1({lhs}); + auto lhs_param = Parameter(&b, 0, lhs_literal->shape(), "lhs_param"); std::unique_ptr lhs_data = client_->TransferToServer(*lhs_literal).ConsumeValueOrDie(); @@ -277,12 +278,12 @@ XLA_TEST_F(ArrayElementwiseOpTest, SubTwoConstantS64s) { 0x7FFFFFFFFFFFFFFLL, 0x7FFFFFFFFFFFFFFFLL, 0x7FFFFFFFFFFFFFFFLL}; - std::unique_ptr rhs_literal = Literal::CreateR1({rhs}); - auto rhs_param = b.Parameter(1, rhs_literal->shape(), "rhs_param"); + std::unique_ptr rhs_literal = LiteralUtil::CreateR1({rhs}); + auto rhs_param = Parameter(&b, 1, rhs_literal->shape(), "rhs_param"); std::unique_ptr rhs_data = client_->TransferToServer(*rhs_literal).ConsumeValueOrDie(); - auto sub = b.Sub(lhs_param, rhs_param); + Sub(lhs_param, rhs_param); std::vector expected(lhs.size()); for (int64 i = 0; i < lhs.size(); ++i) { @@ -302,26 +303,26 @@ TEST_P(ArrayElementwiseOpTestParamCount, AddManyValues) { b_values.push_back(2 * i / static_cast(count + 2)); } - std::unique_ptr a_literal = Literal::CreateR1({a_values}); + std::unique_ptr a_literal = LiteralUtil::CreateR1({a_values}); std::unique_ptr a_data = client_->TransferToServer(*a_literal).ConsumeValueOrDie(); - auto a_constant = builder.ConstantR1(a_values); - auto a_param = builder.Parameter(0, a_literal->shape(), "a_param"); + auto a_constant = ConstantR1(&builder, a_values); + auto a_param = Parameter(&builder, 0, a_literal->shape(), "a_param"); - std::unique_ptr b_literal = Literal::CreateR1({b_values}); + std::unique_ptr b_literal = LiteralUtil::CreateR1({b_values}); std::unique_ptr b_data = client_->TransferToServer(*b_literal).ConsumeValueOrDie(); - auto b_constant = builder.Parameter(1, a_literal->shape(), "b_param"); - auto b_param = builder.ConstantR1(b_values); + auto b_constant = Parameter(&builder, 1, a_literal->shape(), "b_param"); + auto b_param = ConstantR1(&builder, b_values); - auto sum1 = builder.Add(a_constant, b_constant); - auto sum2 = builder.Add(a_constant, b_param); - auto sum3 = builder.Add(a_param, b_constant); - auto sum4 = builder.Add(a_param, b_param); + auto sum1 = Add(a_constant, b_constant); + auto sum2 = Add(a_constant, b_param); + auto sum3 = Add(a_param, b_constant); + auto sum4 = Add(a_param, b_param); - auto sum = builder.Add(sum1, sum2); - sum = builder.Add(sum, sum3); - sum = builder.Add(sum, sum4); + auto sum = Add(sum1, sum2); + sum = Add(sum, sum3); + sum = Add(sum, sum4); std::vector expected; for (int64 i = 0; i < count; ++i) { @@ -334,9 +335,9 @@ TEST_P(ArrayElementwiseOpTestParamCount, AddManyValues) { XLA_TEST_F(ArrayElementwiseOpTest, SubTwoConstantF32s) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({-2.5f, 3.14f, 2.25f, -10.0f, 6.0f}); - auto b = builder.ConstantR1({100.0f, 3.13f, 2.75f, 10.5f, -999.0f}); - builder.Sub(a, b); + auto a = ConstantR1(&builder, {-2.5f, 3.14f, 2.25f, -10.0f, 6.0f}); + auto b = ConstantR1(&builder, {100.0f, 3.13f, 2.75f, 10.5f, -999.0f}); + Sub(a, b); ComputeAndCompareR1(&builder, {-102.5f, 0.01f, -0.5f, -20.5f, 1005.0f}, {}, error_spec_); @@ -344,38 +345,38 @@ XLA_TEST_F(ArrayElementwiseOpTest, SubTwoConstantF32s) { XLA_TEST_F(ArrayElementwiseOpTest, SubTwoConstantZeroElementF32s) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({}); - auto b = builder.ConstantR1({}); - builder.Sub(a, b); + auto a = ConstantR1(&builder, {}); + auto b = ConstantR1(&builder, {}); + Sub(a, b); ComputeAndCompareR1(&builder, {}, {}, error_spec_); } XLA_TEST_F(ArrayElementwiseOpTest, SubTwoConstantS32s) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({-1, 0, 2, 1000000000}); - auto b = builder.ConstantR1({-1, 2, 1, -1}); - builder.Sub(a, b); + auto a = ConstantR1(&builder, {-1, 0, 2, 1000000000}); + auto b = ConstantR1(&builder, {-1, 2, 1, -1}); + Sub(a, b); ComputeAndCompareR1(&builder, {0, -2, 1, 1000000001}, {}); } XLA_TEST_F(ArrayElementwiseOpTest, SubTwoConstantZeroElementS32s) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({}); - auto b = builder.ConstantR1({}); - builder.Sub(a, b); + auto a = ConstantR1(&builder, {}); + auto b = ConstantR1(&builder, {}); + Sub(a, b); ComputeAndCompareR1(&builder, {}, {}); } XLA_TEST_F(ArrayElementwiseOpTest, SubTwoConstantC64s) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1( - {{-2.5f, 0.0f}, {0.0f, 3.14f}, {3.0f, 2.25f}}); - auto b = builder.ConstantR1( - {{0.0f, 10.0f}, {3.13f, 0.0f}, {2.75f, -0.25f}}); - builder.Sub(a, b); + auto a = ConstantR1(&builder, + {{-2.5f, 0.0f}, {0.0f, 3.14f}, {3.0f, 2.25f}}); + auto b = ConstantR1( + &builder, {{0.0f, 10.0f}, {3.13f, 0.0f}, {2.75f, -0.25f}}); + Sub(a, b); ComputeAndCompareR1( &builder, {{-2.5f, -10.0f}, {-3.13f, 3.14f}, {0.25f, 2.5f}}, {}, @@ -384,18 +385,18 @@ XLA_TEST_F(ArrayElementwiseOpTest, SubTwoConstantC64s) { XLA_TEST_F(ArrayElementwiseOpTest, SubTwoConstantZeroElementC64s) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({}); - auto b = builder.ConstantR1({}); - builder.Sub(a, b); + auto a = ConstantR1(&builder, {}); + auto b = ConstantR1(&builder, {}); + Sub(a, b); ComputeAndCompareR1(&builder, {}, {}, error_spec_); } XLA_TEST_F(ArrayElementwiseOpTest, DivTwoConstantF32s) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({-2.5f, 25.5f, 2.25f, -10.0f, 6.0f}); - auto b = builder.ConstantR1({10.0f, 5.1f, 1.0f, 10.0f, -6.0f}); - builder.Div(a, b); + auto a = ConstantR1(&builder, {-2.5f, 25.5f, 2.25f, -10.0f, 6.0f}); + auto b = ConstantR1(&builder, {10.0f, 5.1f, 1.0f, 10.0f, -6.0f}); + Div(a, b); ComputeAndCompareR1(&builder, {-0.25f, 5.0f, 2.25f, -1.0f, -1.0f}, {}, error_spec_); @@ -403,9 +404,9 @@ XLA_TEST_F(ArrayElementwiseOpTest, DivTwoConstantF32s) { XLA_TEST_F(ArrayElementwiseOpTest, DivTwoConstantZeroElementF32s) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({}); - auto b = builder.ConstantR1({}); - builder.Div(a, b); + auto a = ConstantR1(&builder, {}); + auto b = ConstantR1(&builder, {}); + Div(a, b); ComputeAndCompareR1(&builder, {}, {}, error_spec_); } @@ -442,7 +443,7 @@ XLA_TEST_F(ArrayElementwiseOpTest, DivS32s) { CreateR1Parameter(dividends, 0, "dividend", &builder, ÷nd); auto divisor_data = CreateR1Parameter(divisors, 1, "divisor", &builder, &divisor); - builder.Div(dividend, divisor); + Div(dividend, divisor); ComputeAndCompareR1(&builder, quotients, {dividend_data.get(), divisor_data.get()}); @@ -454,7 +455,7 @@ XLA_TEST_F(ArrayElementwiseOpTest, DivS32s) { XlaOp dividend; auto dividend_data = CreateR1Parameter(dividends, 0, "dividend", &builder, ÷nd); - builder.Div(dividend, builder.ConstantR1(divisors)); + Div(dividend, ConstantR1(&builder, divisors)); ComputeAndCompareR1(&builder, quotients, {dividend_data.get()}); } @@ -467,7 +468,7 @@ XLA_TEST_F(ArrayElementwiseOpTest, DivS32s) { CreateR1Parameter(dividends, 0, "dividend", &builder, ÷nd); auto divisor_data = CreateR1Parameter(divisors, 1, "divisor", &builder, &divisor); - builder.Rem(dividend, divisor); + Rem(dividend, divisor); ComputeAndCompareR1(&builder, remainders, {dividend_data.get(), divisor_data.get()}); @@ -479,7 +480,7 @@ XLA_TEST_F(ArrayElementwiseOpTest, DivS32s) { XlaOp dividend; auto dividend_data = CreateR1Parameter(dividends, 0, "dividend", &builder, ÷nd); - builder.Rem(dividend, builder.ConstantR1(divisors)); + Rem(dividend, ConstantR1(&builder, divisors)); ComputeAndCompareR1(&builder, remainders, {dividend_data.get()}); } @@ -513,7 +514,7 @@ XLA_TEST_F(ArrayElementwiseOpTest, DivU32s) { &builder, ÷nd); auto divisor_data = CreateR1Parameter(divisors, 1, "divisor", &builder, &divisor); - builder.Div(dividend, divisor); + Div(dividend, divisor); ComputeAndCompareR1(&builder, quotients, {dividend_data.get(), divisor_data.get()}); @@ -524,7 +525,7 @@ XLA_TEST_F(ArrayElementwiseOpTest, DivU32s) { XlaOp dividend; auto dividend_data = CreateR1Parameter(dividends, 0, "dividend", &builder, ÷nd); - builder.Div(dividend, builder.ConstantR1(divisors)); + Div(dividend, ConstantR1(&builder, divisors)); ComputeAndCompareR1(&builder, quotients, {dividend_data.get()}); } @@ -537,7 +538,7 @@ XLA_TEST_F(ArrayElementwiseOpTest, DivU32s) { &builder, ÷nd); auto divisor_data = CreateR1Parameter(divisors, 1, "divisor", &builder, &divisor); - builder.Rem(dividend, divisor); + Rem(dividend, divisor); ComputeAndCompareR1(&builder, remainders, {dividend_data.get(), divisor_data.get()}); @@ -548,7 +549,7 @@ XLA_TEST_F(ArrayElementwiseOpTest, DivU32s) { XlaOp dividend; auto dividend_data = CreateR1Parameter(dividends, 0, "dividend", &builder, ÷nd); - builder.Rem(dividend, builder.ConstantR1(divisors)); + Rem(dividend, ConstantR1(&builder, divisors)); ComputeAndCompareR1(&builder, remainders, {dividend_data.get()}); } @@ -556,11 +557,11 @@ XLA_TEST_F(ArrayElementwiseOpTest, DivU32s) { XLA_TEST_F(ArrayElementwiseOpTest, DivTwoConstantC64s) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1( - {{-2.5f, 1.0f}, {-25.5f, 0.0f}, {2.0f, -1.0f}}); - auto b = builder.ConstantR1( - {{10.0f, 0.0f}, {0.0f, 1.0f}, {2.0f, -1.0f}}); - builder.Div(a, b); + auto a = ConstantR1( + &builder, {{-2.5f, 1.0f}, {-25.5f, 0.0f}, {2.0f, -1.0f}}); + auto b = ConstantR1(&builder, + {{10.0f, 0.0f}, {0.0f, 1.0f}, {2.0f, -1.0f}}); + Div(a, b); ComputeAndCompareR1( &builder, {{-0.25f, 0.1f}, {0.0f, 25.5f}, {1.0f, 0.0f}}, {}, error_spec_); @@ -568,20 +569,20 @@ XLA_TEST_F(ArrayElementwiseOpTest, DivTwoConstantC64s) { XLA_TEST_F(ArrayElementwiseOpTest, DivTwoConstantZeroElementC64s) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({}); - auto b = builder.ConstantR1({}); - builder.Div(a, b); + auto a = ConstantR1(&builder, {}); + auto b = ConstantR1(&builder, {}); + Div(a, b); ComputeAndCompareR1(&builder, {}, {}, error_spec_); } XLA_TEST_F(ArrayElementwiseOpTest, RemF32s) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1( - {-2.5f, 25.5f, 2.25f, -10.0f, 6.0f, 3.0f, 3.0f, -1.0f, -8.0f}); - auto b = builder.ConstantR1( - {10.0f, 5.1f, 1.0f, 10.0f, -6.0f, 2.0f, -2.0f, 7.0f, -4.0f}); - builder.Rem(a, b); + auto a = ConstantR1( + &builder, {-2.5f, 25.5f, 2.25f, -10.0f, 6.0f, 3.0f, 3.0f, -1.0f, -8.0f}); + auto b = ConstantR1( + &builder, {10.0f, 5.1f, 1.0f, 10.0f, -6.0f, 2.0f, -2.0f, 7.0f, -4.0f}); + Rem(a, b); ComputeAndCompareR1( &builder, {-2.5f, 0.0f, 0.25f, 0.0f, -0.0f, 1.0f, 1.0f, -1.0f, -0.0f}, {}, @@ -590,20 +591,20 @@ XLA_TEST_F(ArrayElementwiseOpTest, RemF32s) { XLA_TEST_F(ArrayElementwiseOpTest, RemZeroElementF32s) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({}); - auto b = builder.ConstantR1({}); - builder.Rem(a, b); + auto a = ConstantR1(&builder, {}); + auto b = ConstantR1(&builder, {}); + Rem(a, b); ComputeAndCompareR1(&builder, {}, {}, error_spec_); } XLA_TEST_F(ArrayElementwiseOpTest, RemF64s) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1( - {-2.5, 25.5, 2.25, -10.0, 6.0, 3.0, 3.0, -1.0, -8.0}); - auto b = builder.ConstantR1( - {10.0, 5.1, 1.0, 10.0, -6.0, 2.0, -2.0, 7.0, -4.0}); - builder.Rem(a, b); + auto a = ConstantR1( + &builder, {-2.5, 25.5, 2.25, -10.0, 6.0, 3.0, 3.0, -1.0, -8.0}); + auto b = ConstantR1( + &builder, {10.0, 5.1, 1.0, 10.0, -6.0, 2.0, -2.0, 7.0, -4.0}); + Rem(a, b); ComputeAndCompareR1( &builder, {-2.5, 0.0, 0.25, 0.0, -0.0, 1.0, 1.0, -1.0, -0.0}, {}, @@ -612,9 +613,9 @@ XLA_TEST_F(ArrayElementwiseOpTest, RemF64s) { XLA_TEST_F(ArrayElementwiseOpTest, MulTwoConstantF32s) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({-2.5f, 25.5f, 2.25f, -10.0f, 6.0f}); - auto b = builder.ConstantR1({10.0f, 5.0f, 1.0f, 10.0f, -6.0f}); - builder.Mul(a, b); + auto a = ConstantR1(&builder, {-2.5f, 25.5f, 2.25f, -10.0f, 6.0f}); + auto b = ConstantR1(&builder, {10.0f, 5.0f, 1.0f, 10.0f, -6.0f}); + Mul(a, b); ComputeAndCompareR1(&builder, {-25.0f, 127.5f, 2.25f, -100.0f, -36.0f}, {}, error_spec_); @@ -622,9 +623,9 @@ XLA_TEST_F(ArrayElementwiseOpTest, MulTwoConstantF32s) { XLA_TEST_F(ArrayElementwiseOpTest, MulTwoConstantZeroElementF32s) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({}); - auto b = builder.ConstantR1({}); - builder.Mul(a, b); + auto a = ConstantR1(&builder, {}); + auto b = ConstantR1(&builder, {}); + Mul(a, b); ComputeAndCompareR1(&builder, {}, {}, error_spec_); } @@ -648,18 +649,18 @@ XLA_TEST_F(ArrayElementwiseOpTest, MulTwoConstantS32s) { } XlaBuilder builder(TestName()); - auto a = builder.ConstantR1(a_data); - auto b = builder.ConstantR1(b_data); - builder.Mul(a, b); + auto a = ConstantR1(&builder, a_data); + auto b = ConstantR1(&builder, b_data); + Mul(a, b); ComputeAndCompareR1(&builder, expected, {}); } XLA_TEST_F(ArrayElementwiseOpTest, MulTwoConstantZeroElementS32s) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({}); - auto b = builder.ConstantR1({}); - builder.Mul(a, b); + auto a = ConstantR1(&builder, {}); + auto b = ConstantR1(&builder, {}); + Mul(a, b); ComputeAndCompareR1(&builder, {}, {}); } @@ -679,20 +680,20 @@ XLA_TEST_F(ArrayElementwiseOpTest, MulTwoConstantU32s) { } XlaBuilder builder(TestName()); - auto a = builder.ConstantR1(a_data); - auto b = builder.ConstantR1(b_data); - builder.Mul(a, b); + auto a = ConstantR1(&builder, a_data); + auto b = ConstantR1(&builder, b_data); + Mul(a, b); ComputeAndCompareR1(&builder, expected, {}); } XLA_TEST_F(ArrayElementwiseOpTest, MulTwoConstantC64s) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1( - {{-2.5f, 0.0f}, {0.0f, 25.5f}, {2.0f, -10.0f}}); - auto b = builder.ConstantR1( - {{0.0f, 10.0f}, {5.0f, 1.0f}, {10.0f, -6.0f}}); - builder.Mul(a, b); + auto a = ConstantR1( + &builder, {{-2.5f, 0.0f}, {0.0f, 25.5f}, {2.0f, -10.0f}}); + auto b = ConstantR1(&builder, + {{0.0f, 10.0f}, {5.0f, 1.0f}, {10.0f, -6.0f}}); + Mul(a, b); ComputeAndCompareR1( &builder, {{0.0f, -25.0f}, {-25.5f, 127.5f}, {-40.0f, -112.0}}, {}, @@ -701,27 +702,27 @@ XLA_TEST_F(ArrayElementwiseOpTest, MulTwoConstantC64s) { XLA_TEST_F(ArrayElementwiseOpTest, MulTwoConstantZeroElementC64s) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({}); - auto b = builder.ConstantR1({}); - builder.Mul(a, b); + auto a = ConstantR1(&builder, {}); + auto b = ConstantR1(&builder, {}); + Mul(a, b); ComputeAndCompareR1(&builder, {}, {}, error_spec_); } XLA_TEST_F(ArrayElementwiseOpTest, AndPredR1) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({false, false, true, true}); - auto b = builder.ConstantR1({false, true, false, true}); - builder.And(a, b); + auto a = ConstantR1(&builder, {false, false, true, true}); + auto b = ConstantR1(&builder, {false, true, false, true}); + And(a, b); ComputeAndCompareR1(&builder, {false, false, false, true}, {}); } XLA_TEST_F(ArrayElementwiseOpTest, AndPredR2) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR2({{false, false}, {true, true}}); - auto b = builder.ConstantR2({{false, true}, {false, true}}); - builder.And(a, b); + auto a = ConstantR2(&builder, {{false, false}, {true, true}}); + auto b = ConstantR2(&builder, {{false, true}, {false, true}}); + And(a, b); Array2D expected_array({{false, false}, {false, true}}); ComputeAndCompareR2(&builder, expected_array, {}); @@ -729,27 +730,27 @@ XLA_TEST_F(ArrayElementwiseOpTest, AndPredR2) { XLA_TEST_F(ArrayElementwiseOpTest, AndZeroElementPredR1) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({}); - auto b = builder.ConstantR1({}); - builder.And(a, b); + auto a = ConstantR1(&builder, {}); + auto b = ConstantR1(&builder, {}); + And(a, b); ComputeAndCompareR1(&builder, {}, {}); } XLA_TEST_F(ArrayElementwiseOpTest, AndS32R1) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({0, -1, -8}); - auto b = builder.ConstantR1({5, -7, 12}); - builder.And(a, b); + auto a = ConstantR1(&builder, {0, -1, -8}); + auto b = ConstantR1(&builder, {5, -7, 12}); + And(a, b); ComputeAndCompareR1(&builder, {0, -7, 8}, {}); } XLA_TEST_F(ArrayElementwiseOpTest, AndS32R2) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR2({{0, -5}, {-1, 5}}); - auto b = builder.ConstantR2({{1, -6}, {4, 5}}); - builder.And(a, b); + auto a = ConstantR2(&builder, {{0, -5}, {-1, 5}}); + auto b = ConstantR2(&builder, {{1, -6}, {4, 5}}); + And(a, b); Array2D expected_array({{0, -6}, {4, 5}}); ComputeAndCompareR2(&builder, expected_array, {}); @@ -757,27 +758,27 @@ XLA_TEST_F(ArrayElementwiseOpTest, AndS32R2) { XLA_TEST_F(ArrayElementwiseOpTest, AndZeroElementS32R1) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({}); - auto b = builder.ConstantR1({}); - builder.And(a, b); + auto a = ConstantR1(&builder, {}); + auto b = ConstantR1(&builder, {}); + And(a, b); ComputeAndCompareR1(&builder, {}, {}); } XLA_TEST_F(ArrayElementwiseOpTest, AndU32R1) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({0, 1, 8}); - auto b = builder.ConstantR1({5, 7, 12}); - builder.And(a, b); + auto a = ConstantR1(&builder, {0, 1, 8}); + auto b = ConstantR1(&builder, {5, 7, 12}); + And(a, b); ComputeAndCompareR1(&builder, {0, 1, 8}, {}); } XLA_TEST_F(ArrayElementwiseOpTest, AndU32R2) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR2({{0, 1}, {3, 8}}); - auto b = builder.ConstantR2({{1, 0}, {7, 6}}); - builder.And(a, b); + auto a = ConstantR2(&builder, {{0, 1}, {3, 8}}); + auto b = ConstantR2(&builder, {{1, 0}, {7, 6}}); + And(a, b); Array2D expected_array({{0, 0}, {3, 0}}); ComputeAndCompareR2(&builder, expected_array, {}); @@ -785,27 +786,27 @@ XLA_TEST_F(ArrayElementwiseOpTest, AndU32R2) { XLA_TEST_F(ArrayElementwiseOpTest, AndZeroElementU32R1) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({}); - auto b = builder.ConstantR1({}); - builder.And(a, b); + auto a = ConstantR1(&builder, {}); + auto b = ConstantR1(&builder, {}); + And(a, b); ComputeAndCompareR1(&builder, {}, {}); } XLA_TEST_F(ArrayElementwiseOpTest, OrPredR1) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({false, false, true, true}); - auto b = builder.ConstantR1({false, true, false, true}); - builder.Or(a, b); + auto a = ConstantR1(&builder, {false, false, true, true}); + auto b = ConstantR1(&builder, {false, true, false, true}); + Or(a, b); ComputeAndCompareR1(&builder, {false, true, true, true}, {}); } XLA_TEST_F(ArrayElementwiseOpTest, OrPredR2) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR2({{false, false}, {true, true}}); - auto b = builder.ConstantR2({{false, true}, {false, true}}); - builder.Or(a, b); + auto a = ConstantR2(&builder, {{false, false}, {true, true}}); + auto b = ConstantR2(&builder, {{false, true}, {false, true}}); + Or(a, b); Array2D expected_array({{false, true}, {true, true}}); ComputeAndCompareR2(&builder, expected_array, {}); @@ -813,27 +814,27 @@ XLA_TEST_F(ArrayElementwiseOpTest, OrPredR2) { XLA_TEST_F(ArrayElementwiseOpTest, OrZeroElementPredR1) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({}); - auto b = builder.ConstantR1({}); - builder.Or(a, b); + auto a = ConstantR1(&builder, {}); + auto b = ConstantR1(&builder, {}); + Or(a, b); ComputeAndCompareR1(&builder, {}, {}); } XLA_TEST_F(ArrayElementwiseOpTest, OrS32R1) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({0, -1, 8}); - auto b = builder.ConstantR1({5, -7, 4}); - builder.Or(a, b); + auto a = ConstantR1(&builder, {0, -1, 8}); + auto b = ConstantR1(&builder, {5, -7, 4}); + Or(a, b); ComputeAndCompareR1(&builder, {5, -1, 12}, {}); } XLA_TEST_F(ArrayElementwiseOpTest, OrS32R2) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR2({{0, -1}, {8, 8}}); - auto b = builder.ConstantR2({{5, -7}, {4, 1}}); - builder.Or(a, b); + auto a = ConstantR2(&builder, {{0, -1}, {8, 8}}); + auto b = ConstantR2(&builder, {{5, -7}, {4, 1}}); + Or(a, b); Array2D expected_array({{5, -1}, {12, 9}}); ComputeAndCompareR2(&builder, expected_array, {}); @@ -841,27 +842,27 @@ XLA_TEST_F(ArrayElementwiseOpTest, OrS32R2) { XLA_TEST_F(ArrayElementwiseOpTest, OrZeroElementS32R1) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({}); - auto b = builder.ConstantR1({}); - builder.Or(a, b); + auto a = ConstantR1(&builder, {}); + auto b = ConstantR1(&builder, {}); + Or(a, b); ComputeAndCompareR1(&builder, {}, {}); } XLA_TEST_F(ArrayElementwiseOpTest, OrU32R1) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({0, 1, 8}); - auto b = builder.ConstantR1({5, 7, 4}); - builder.Or(a, b); + auto a = ConstantR1(&builder, {0, 1, 8}); + auto b = ConstantR1(&builder, {5, 7, 4}); + Or(a, b); ComputeAndCompareR1(&builder, {5, 7, 12}, {}); } XLA_TEST_F(ArrayElementwiseOpTest, OrU32R2) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR2({{0, 1}, {8, 8}}); - auto b = builder.ConstantR2({{5, 7}, {4, 1}}); - builder.Or(a, b); + auto a = ConstantR2(&builder, {{0, 1}, {8, 8}}); + auto b = ConstantR2(&builder, {{5, 7}, {4, 1}}); + Or(a, b); Array2D expected_array({{5, 7}, {12, 9}}); ComputeAndCompareR2(&builder, expected_array, {}); @@ -869,27 +870,27 @@ XLA_TEST_F(ArrayElementwiseOpTest, OrU32R2) { XLA_TEST_F(ArrayElementwiseOpTest, OrZeroElementU32R1) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({}); - auto b = builder.ConstantR1({}); - builder.Or(a, b); + auto a = ConstantR1(&builder, {}); + auto b = ConstantR1(&builder, {}); + Or(a, b); ComputeAndCompareR1(&builder, {}, {}); } XLA_TEST_F(ArrayElementwiseOpTest, XorPredR1) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({false, false, true, true}); - auto b = builder.ConstantR1({false, true, false, true}); - builder.Xor(a, b); + auto a = ConstantR1(&builder, {false, false, true, true}); + auto b = ConstantR1(&builder, {false, true, false, true}); + Xor(a, b); ComputeAndCompareR1(&builder, {false, true, true, false}, {}); } XLA_TEST_F(ArrayElementwiseOpTest, XorPredR2) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR2({{false, false}, {true, true}}); - auto b = builder.ConstantR2({{false, true}, {false, true}}); - builder.Xor(a, b); + auto a = ConstantR2(&builder, {{false, false}, {true, true}}); + auto b = ConstantR2(&builder, {{false, true}, {false, true}}); + Xor(a, b); Array2D expected_array({{false, true}, {true, false}}); ComputeAndCompareR2(&builder, expected_array, {}); @@ -897,27 +898,27 @@ XLA_TEST_F(ArrayElementwiseOpTest, XorPredR2) { XLA_TEST_F(ArrayElementwiseOpTest, XorZeroElementPredR1) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({}); - auto b = builder.ConstantR1({}); - builder.Xor(a, b); + auto a = ConstantR1(&builder, {}); + auto b = ConstantR1(&builder, {}); + Xor(a, b); ComputeAndCompareR1(&builder, {}, {}); } XLA_TEST_F(ArrayElementwiseOpTest, XorS32R1) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({0, -1, 8}); - auto b = builder.ConstantR1({5, -7, 4}); - builder.Xor(a, b); + auto a = ConstantR1(&builder, {0, -1, 8}); + auto b = ConstantR1(&builder, {5, -7, 4}); + Xor(a, b); ComputeAndCompareR1(&builder, {5, 6, 12}, {}); } XLA_TEST_F(ArrayElementwiseOpTest, XorS32R2) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR2({{0, -1}, {8, 8}}); - auto b = builder.ConstantR2({{5, -7}, {4, 1}}); - builder.Xor(a, b); + auto a = ConstantR2(&builder, {{0, -1}, {8, 8}}); + auto b = ConstantR2(&builder, {{5, -7}, {4, 1}}); + Xor(a, b); Array2D expected_array({{5, 6}, {12, 9}}); ComputeAndCompareR2(&builder, expected_array, {}); @@ -925,27 +926,27 @@ XLA_TEST_F(ArrayElementwiseOpTest, XorS32R2) { XLA_TEST_F(ArrayElementwiseOpTest, XorZeroElementS32R1) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({}); - auto b = builder.ConstantR1({}); - builder.Xor(a, b); + auto a = ConstantR1(&builder, {}); + auto b = ConstantR1(&builder, {}); + Xor(a, b); ComputeAndCompareR1(&builder, {}, {}); } XLA_TEST_F(ArrayElementwiseOpTest, XorU32R1) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({0, 1, 8}); - auto b = builder.ConstantR1({5, 7, 4}); - builder.Xor(a, b); + auto a = ConstantR1(&builder, {0, 1, 8}); + auto b = ConstantR1(&builder, {5, 7, 4}); + Xor(a, b); ComputeAndCompareR1(&builder, {5, 6, 12}, {}); } XLA_TEST_F(ArrayElementwiseOpTest, XorU32R2) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR2({{0, 1}, {8, 8}}); - auto b = builder.ConstantR2({{5, 7}, {4, 1}}); - builder.Xor(a, b); + auto a = ConstantR2(&builder, {{0, 1}, {8, 8}}); + auto b = ConstantR2(&builder, {{5, 7}, {4, 1}}); + Xor(a, b); Array2D expected_array({{5, 6}, {12, 9}}); ComputeAndCompareR2(&builder, expected_array, {}); @@ -953,24 +954,24 @@ XLA_TEST_F(ArrayElementwiseOpTest, XorU32R2) { XLA_TEST_F(ArrayElementwiseOpTest, XorZeroElementU32R1) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({}); - auto b = builder.ConstantR1({}); - builder.Xor(a, b); + auto a = ConstantR1(&builder, {}); + auto b = ConstantR1(&builder, {}); + Xor(a, b); ComputeAndCompareR1(&builder, {}, {}); } XLA_TEST_F(ArrayElementwiseOpTest, NotPredR1) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({false, true, true, false}); - builder.Not(a); + auto a = ConstantR1(&builder, {false, true, true, false}); + Not(a); ComputeAndCompareR1(&builder, {true, false, false, true}, {}); } XLA_TEST_F(ArrayElementwiseOpTest, NotPredR2) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR2({{false, true}, {true, false}}); - builder.Not(a); + auto a = ConstantR2(&builder, {{false, true}, {true, false}}); + Not(a); Array2D expected_array({{true, false}, {false, true}}); ComputeAndCompareR2(&builder, expected_array, {}); @@ -978,24 +979,24 @@ XLA_TEST_F(ArrayElementwiseOpTest, NotPredR2) { XLA_TEST_F(ArrayElementwiseOpTest, NotZeroElementPredR1) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({}); - builder.Not(a); + auto a = ConstantR1(&builder, {}); + Not(a); ComputeAndCompareR1(&builder, {}, {}); } XLA_TEST_F(ArrayElementwiseOpTest, NotS32R1) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({-1, 0, 1}); - builder.Not(a); + auto a = ConstantR1(&builder, {-1, 0, 1}); + Not(a); ComputeAndCompareR1(&builder, {0, -1, -2}, {}); } XLA_TEST_F(ArrayElementwiseOpTest, NotS32R2) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR2({{-1, 0}, {1, 8}}); - builder.Not(a); + auto a = ConstantR2(&builder, {{-1, 0}, {1, 8}}); + Not(a); Array2D expected_array({{0, -1}, {-2, -9}}); ComputeAndCompareR2(&builder, expected_array, {}); @@ -1003,24 +1004,24 @@ XLA_TEST_F(ArrayElementwiseOpTest, NotS32R2) { XLA_TEST_F(ArrayElementwiseOpTest, NotZeroElementS32R1) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({}); - builder.Not(a); + auto a = ConstantR1(&builder, {}); + Not(a); ComputeAndCompareR1(&builder, {}, {}); } XLA_TEST_F(ArrayElementwiseOpTest, NotU32R1) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({0, 4294967295}); - builder.Not(a); + auto a = ConstantR1(&builder, {0, 4294967295}); + Not(a); ComputeAndCompareR1(&builder, {4294967295, 0}, {}); } XLA_TEST_F(ArrayElementwiseOpTest, NotU32R2) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR2({{0, 4294967295}, {1, 4294967294}}); - builder.Not(a); + auto a = ConstantR2(&builder, {{0, 4294967295}, {1, 4294967294}}); + Not(a); Array2D expected_array({{4294967295, 0}, {4294967294, 1}}); ComputeAndCompareR2(&builder, expected_array, {}); @@ -1028,19 +1029,19 @@ XLA_TEST_F(ArrayElementwiseOpTest, NotU32R2) { XLA_TEST_F(ArrayElementwiseOpTest, NotZeroElementU32R1) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({}); - builder.Not(a); + auto a = ConstantR1(&builder, {}); + Not(a); ComputeAndCompareR1(&builder, {}, {}); } XLA_TEST_F(ArrayElementwiseOpTest, ShiftLeftS32) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({static_cast(0x12345678), - static_cast(0xF0001000), 1, 3, 77, - 1, -3, 77}); - auto b = builder.ConstantR1({4, 8, 2, 7, 15, 32, 100, -1}); - builder.ShiftLeft(a, b); + auto a = ConstantR1( + &builder, {static_cast(0x12345678), static_cast(0xF0001000), + 1, 3, 77, 1, -3, 77}); + auto b = ConstantR1(&builder, {4, 8, 2, 7, 15, 32, 100, -1}); + ShiftLeft(a, b); ComputeAndCompareR1(&builder, {static_cast(0x23456780), 0x00100000, 0x4, @@ -1050,11 +1051,11 @@ XLA_TEST_F(ArrayElementwiseOpTest, ShiftLeftS32) { XLA_TEST_F(ArrayElementwiseOpTest, ShiftRightArithmeticS32) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({static_cast(0x92345678), - static_cast(0x10001000), 1, 3, 77, - 1, -3, 77}); - auto b = builder.ConstantR1({4, 8, 2, 7, 2, 32, 100, -1}); - builder.ShiftRightArithmetic(a, b); + auto a = ConstantR1( + &builder, {static_cast(0x92345678), static_cast(0x10001000), + 1, 3, 77, 1, -3, 77}); + auto b = ConstantR1(&builder, {4, 8, 2, 7, 2, 32, 100, -1}); + ShiftRightArithmetic(a, b); ComputeAndCompareR1( &builder, @@ -1065,11 +1066,11 @@ XLA_TEST_F(ArrayElementwiseOpTest, ShiftRightArithmeticS32) { XLA_TEST_F(ArrayElementwiseOpTest, ShiftRightLogicalS32) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({static_cast(0x92345678), - static_cast(0x10001000), 1, 3, 77, - 1, -3, 77}); - auto b = builder.ConstantR1({4, 8, 2, 7, 5, 32, 100, -1}); - builder.ShiftRightLogical(a, b); + auto a = ConstantR1( + &builder, {static_cast(0x92345678), static_cast(0x10001000), + 1, 3, 77, 1, -3, 77}); + auto b = ConstantR1(&builder, {4, 8, 2, 7, 5, 32, 100, -1}); + ShiftRightLogical(a, b); ComputeAndCompareR1(&builder, {0x09234567, 0x00100010, 0, 0, 2, 0, 0, 0}, {}); @@ -1077,10 +1078,10 @@ XLA_TEST_F(ArrayElementwiseOpTest, ShiftRightLogicalS32) { XLA_TEST_F(ArrayElementwiseOpTest, ShiftLeftU32) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1( - {0x12345678, 0xF0001000, 1, 3, 77, 1, ~3u, 77}); - auto b = builder.ConstantR1({4, 8, 2, 7, 15, 32, 100, ~0u}); - builder.ShiftLeft(a, b); + auto a = ConstantR1(&builder, + {0x12345678, 0xF0001000, 1, 3, 77, 1, ~3u, 77}); + auto b = ConstantR1(&builder, {4, 8, 2, 7, 15, 32, 100, ~0u}); + ShiftLeft(a, b); ComputeAndCompareR1( &builder, {0x23456780, 0x00100000, 0x4, 0x180, 2523136, 0, 0, 0}, {}); @@ -1088,10 +1089,10 @@ XLA_TEST_F(ArrayElementwiseOpTest, ShiftLeftU32) { XLA_TEST_F(ArrayElementwiseOpTest, ShiftRightArithmeticU32) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1( - {0x92345678, 0x10001000, 1, 3, 77, 1, ~3u, 77}); - auto b = builder.ConstantR1({4, 8, 2, 7, 2, 32, 100, ~0u}); - builder.ShiftRightArithmetic(a, b); + auto a = ConstantR1(&builder, + {0x92345678, 0x10001000, 1, 3, 77, 1, ~3u, 77}); + auto b = ConstantR1(&builder, {4, 8, 2, 7, 2, 32, 100, ~0u}); + ShiftRightArithmetic(a, b); ComputeAndCompareR1( &builder, {0xF9234567, 0x00100010, 0, 0, 19, 0, ~0u, 0}, {}); @@ -1099,10 +1100,10 @@ XLA_TEST_F(ArrayElementwiseOpTest, ShiftRightArithmeticU32) { XLA_TEST_F(ArrayElementwiseOpTest, ShiftRightLogicalU32) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1( - {0x92345678, 0x10001000, 1, 3, 77, 1, ~3u, 77}); - auto b = builder.ConstantR1({4, 8, 2, 7, 5, 32, 100, ~0u}); - builder.ShiftRightLogical(a, b); + auto a = ConstantR1(&builder, + {0x92345678, 0x10001000, 1, 3, 77, 1, ~3u, 77}); + auto b = ConstantR1(&builder, {4, 8, 2, 7, 5, 32, 100, ~0u}); + ShiftRightLogical(a, b); ComputeAndCompareR1(&builder, {0x09234567, 0x00100010, 0, 0, 2, 0, 0, 0}, {}); @@ -1111,18 +1112,18 @@ XLA_TEST_F(ArrayElementwiseOpTest, ShiftRightLogicalU32) { XLA_TEST_F(ArrayElementwiseOpTest, CompareEqF32s) { SetFastMathDisabled(true); XlaBuilder builder(TestName()); - auto lhs = builder.ConstantR1({-2.5f, 25.5f, 2.25f, NAN, 6.0f}); - auto rhs = builder.ConstantR1({10.0f, 5.0f, 2.25f, 10.0f, NAN}); - builder.Eq(lhs, rhs); + auto lhs = ConstantR1(&builder, {-2.5f, 25.5f, 2.25f, NAN, 6.0f}); + auto rhs = ConstantR1(&builder, {10.0f, 5.0f, 2.25f, 10.0f, NAN}); + Eq(lhs, rhs); ComputeAndCompareR1(&builder, {false, false, true, false, false}, {}); } XLA_TEST_F(ArrayElementwiseOpTest, CompareEqZeroElementF32s) { XlaBuilder builder(TestName()); - auto lhs = builder.ConstantR1({}); - auto rhs = builder.ConstantR1({}); - builder.Eq(lhs, rhs); + auto lhs = ConstantR1(&builder, {}); + auto rhs = ConstantR1(&builder, {}); + Eq(lhs, rhs); ComputeAndCompareR1(&builder, {}, {}); } @@ -1130,9 +1131,9 @@ XLA_TEST_F(ArrayElementwiseOpTest, CompareEqZeroElementF32s) { XLA_TEST_F(ArrayElementwiseOpTest, CompareGeF32s) { SetFastMathDisabled(true); XlaBuilder builder(TestName()); - auto lhs = builder.ConstantR1({-2.5f, 25.5f, 2.25f, NAN, 6.0f}); - auto rhs = builder.ConstantR1({10.0f, 5.0f, 1.0f, 10.0f, NAN}); - builder.Ge(lhs, rhs); + auto lhs = ConstantR1(&builder, {-2.5f, 25.5f, 2.25f, NAN, 6.0f}); + auto rhs = ConstantR1(&builder, {10.0f, 5.0f, 1.0f, 10.0f, NAN}); + Ge(lhs, rhs); ComputeAndCompareR1(&builder, {false, true, true, false, false}, {}); } @@ -1140,9 +1141,9 @@ XLA_TEST_F(ArrayElementwiseOpTest, CompareGeF32s) { XLA_TEST_F(ArrayElementwiseOpTest, CompareGtF32s) { SetFastMathDisabled(true); XlaBuilder builder(TestName()); - auto lhs = builder.ConstantR1({-2.5f, 25.5f, 2.25f, NAN, 6.0f}); - auto rhs = builder.ConstantR1({10.0f, 5.0f, 1.0f, 10.0f, NAN}); - builder.Gt(lhs, rhs); + auto lhs = ConstantR1(&builder, {-2.5f, 25.5f, 2.25f, NAN, 6.0f}); + auto rhs = ConstantR1(&builder, {10.0f, 5.0f, 1.0f, 10.0f, NAN}); + Gt(lhs, rhs); ComputeAndCompareR1(&builder, {false, true, true, false, false}, {}); } @@ -1150,9 +1151,9 @@ XLA_TEST_F(ArrayElementwiseOpTest, CompareGtF32s) { XLA_TEST_F(ArrayElementwiseOpTest, CompareLeF32s) { SetFastMathDisabled(true); XlaBuilder builder(TestName()); - auto lhs = builder.ConstantR1({-2.5f, 5.0f, 2.25f, NAN, 6.0f}); - auto rhs = builder.ConstantR1({10.0f, 5.0f, 1.0f, 10.0f, NAN}); - builder.Le(lhs, rhs); + auto lhs = ConstantR1(&builder, {-2.5f, 5.0f, 2.25f, NAN, 6.0f}); + auto rhs = ConstantR1(&builder, {10.0f, 5.0f, 1.0f, 10.0f, NAN}); + Le(lhs, rhs); ComputeAndCompareR1(&builder, {true, true, false, false, false}, {}); } @@ -1160,9 +1161,9 @@ XLA_TEST_F(ArrayElementwiseOpTest, CompareLeF32s) { XLA_TEST_F(ArrayElementwiseOpTest, CompareLtF32s) { SetFastMathDisabled(true); XlaBuilder builder(TestName()); - auto lhs = builder.ConstantR1({-2.5f, 25.5f, 2.25f, NAN, 6.0f}); - auto rhs = builder.ConstantR1({10.0f, 5.0f, 1.0f, 10.0f, NAN}); - builder.Lt(lhs, rhs); + auto lhs = ConstantR1(&builder, {-2.5f, 25.5f, 2.25f, NAN, 6.0f}); + auto rhs = ConstantR1(&builder, {10.0f, 5.0f, 1.0f, 10.0f, NAN}); + Lt(lhs, rhs); ComputeAndCompareR1(&builder, {true, false, false, false, false}, {}); } @@ -1171,9 +1172,10 @@ XLA_TEST_F(ArrayElementwiseOpTest, CompareEqS32s) { const int32 min = std::numeric_limits::min(); const int32 max = std::numeric_limits::max(); XlaBuilder builder(TestName()); - auto lhs = builder.ConstantR1({min, min, min, 0, 0, 0, max, max, max}); - auto rhs = builder.ConstantR1({min, 0, max, -1, 0, 1, min, 0, max}); - builder.Eq(lhs, rhs); + auto lhs = + ConstantR1(&builder, {min, min, min, 0, 0, 0, max, max, max}); + auto rhs = ConstantR1(&builder, {min, 0, max, -1, 0, 1, min, 0, max}); + Eq(lhs, rhs); ComputeAndCompareR1( &builder, {true, false, false, false, true, false, false, false, true}, @@ -1182,9 +1184,9 @@ XLA_TEST_F(ArrayElementwiseOpTest, CompareEqS32s) { XLA_TEST_F(ArrayElementwiseOpTest, CompareEqZeroElementS32s) { XlaBuilder builder(TestName()); - auto lhs = builder.ConstantR1({}); - auto rhs = builder.ConstantR1({}); - builder.Eq(lhs, rhs); + auto lhs = ConstantR1(&builder, {}); + auto rhs = ConstantR1(&builder, {}); + Eq(lhs, rhs); ComputeAndCompareR1(&builder, {}, {}); } @@ -1192,26 +1194,26 @@ XLA_TEST_F(ArrayElementwiseOpTest, CompareEqZeroElementS32s) { XLA_TEST_F(ArrayElementwiseOpTest, CompareEqC64s) { SetFastMathDisabled(true); XlaBuilder builder(TestName()); - auto lhs = builder.ConstantR1({{-2.5f, 10.0f}, - {1.0f, 25.5f}, - {2.25f, -3.0f}, - {NAN, 0.0f}, - {1.0f, 6.0f}}); - auto rhs = builder.ConstantR1({{0.0f, 10.0f}, - {1.0f, 5.0f}, - {2.25f, -3.0f}, - {10.0f, 0.0f}, - {1.0f, NAN}}); - builder.Eq(lhs, rhs); + auto lhs = ConstantR1(&builder, {{-2.5f, 10.0f}, + {1.0f, 25.5f}, + {2.25f, -3.0f}, + {NAN, 0.0f}, + {1.0f, 6.0f}}); + auto rhs = ConstantR1(&builder, {{0.0f, 10.0f}, + {1.0f, 5.0f}, + {2.25f, -3.0f}, + {10.0f, 0.0f}, + {1.0f, NAN}}); + Eq(lhs, rhs); ComputeAndCompareR1(&builder, {false, false, true, false, false}, {}); } XLA_TEST_F(ArrayElementwiseOpTest, CompareEqZeroElementC64s) { XlaBuilder builder(TestName()); - auto lhs = builder.ConstantR1({}); - auto rhs = builder.ConstantR1({}); - builder.Eq(lhs, rhs); + auto lhs = ConstantR1(&builder, {}); + auto rhs = ConstantR1(&builder, {}); + Eq(lhs, rhs); ComputeAndCompareR1(&builder, {}, {}); } @@ -1221,17 +1223,17 @@ XLA_TEST_F(ArrayElementwiseOpTest, CompareNeC64s) { SetFastMathDisabled(true); XlaBuilder builder(TestName()); - auto lhs = builder.ConstantR1({{-2.5f, 10.0f}, - {1.0f, 25.5f}, - {2.25f, -3.0f}, - {NAN, 0.0f}, - {1.0f, 6.0f}}); - auto rhs = builder.ConstantR1({{0.0f, 10.0f}, - {1.0f, 5.0f}, - {2.25f, -3.0f}, - {10.0f, 0.0f}, - {1.0f, NAN}}); - builder.Ne(lhs, rhs); + auto lhs = ConstantR1(&builder, {{-2.5f, 10.0f}, + {1.0f, 25.5f}, + {2.25f, -3.0f}, + {NAN, 0.0f}, + {1.0f, 6.0f}}); + auto rhs = ConstantR1(&builder, {{0.0f, 10.0f}, + {1.0f, 5.0f}, + {2.25f, -3.0f}, + {10.0f, 0.0f}, + {1.0f, NAN}}); + Ne(lhs, rhs); ComputeAndCompareR1(&builder, {true, true, false, true, true}, {}); } @@ -1241,9 +1243,9 @@ XLA_TEST_F(ArrayElementwiseOpTest, CompareNeF32s) { SetFastMathDisabled(true); XlaBuilder builder(TestName()); - auto lhs = builder.ConstantR1({-2.5f, 25.5f, 2.25f, NAN, 6.0f}); - auto rhs = builder.ConstantR1({10.0f, 25.5f, 1.0f, 10.0f, NAN}); - builder.Ne(lhs, rhs); + auto lhs = ConstantR1(&builder, {-2.5f, 25.5f, 2.25f, NAN, 6.0f}); + auto rhs = ConstantR1(&builder, {10.0f, 25.5f, 1.0f, 10.0f, NAN}); + Ne(lhs, rhs); ComputeAndCompareR1(&builder, {true, false, true, true, true}, {}); } @@ -1252,9 +1254,10 @@ XLA_TEST_F(ArrayElementwiseOpTest, CompareNeS32s) { const int32 min = std::numeric_limits::min(); const int32 max = std::numeric_limits::max(); XlaBuilder builder(TestName()); - auto lhs = builder.ConstantR1({min, min, min, 0, 0, 0, max, max, max}); - auto rhs = builder.ConstantR1({min, 0, max, -1, 0, 1, min, 0, max}); - builder.Ne(lhs, rhs); + auto lhs = + ConstantR1(&builder, {min, min, min, 0, 0, 0, max, max, max}); + auto rhs = ConstantR1(&builder, {min, 0, max, -1, 0, 1, min, 0, max}); + Ne(lhs, rhs); ComputeAndCompareR1( &builder, {false, true, true, true, false, true, true, true, false}, {}); @@ -1264,9 +1267,10 @@ XLA_TEST_F(ArrayElementwiseOpTest, CompareGeS32s) { const int32 min = std::numeric_limits::min(); const int32 max = std::numeric_limits::max(); XlaBuilder builder(TestName()); - auto lhs = builder.ConstantR1({min, min, min, 0, 0, 0, max, max, max}); - auto rhs = builder.ConstantR1({min, 0, max, -1, 0, 1, min, 0, max}); - builder.Ge(lhs, rhs); + auto lhs = + ConstantR1(&builder, {min, min, min, 0, 0, 0, max, max, max}); + auto rhs = ConstantR1(&builder, {min, 0, max, -1, 0, 1, min, 0, max}); + Ge(lhs, rhs); ComputeAndCompareR1( &builder, {true, false, false, true, true, false, true, true, true}, {}); @@ -1276,9 +1280,10 @@ XLA_TEST_F(ArrayElementwiseOpTest, CompareGtS32s) { const int32 min = std::numeric_limits::min(); const int32 max = std::numeric_limits::max(); XlaBuilder builder(TestName()); - auto lhs = builder.ConstantR1({min, min, min, 0, 0, 0, max, max, max}); - auto rhs = builder.ConstantR1({min, 0, max, -1, 0, 1, min, 0, max}); - builder.Gt(lhs, rhs); + auto lhs = + ConstantR1(&builder, {min, min, min, 0, 0, 0, max, max, max}); + auto rhs = ConstantR1(&builder, {min, 0, max, -1, 0, 1, min, 0, max}); + Gt(lhs, rhs); ComputeAndCompareR1( &builder, {false, false, false, true, false, false, true, true, false}, @@ -1289,9 +1294,10 @@ XLA_TEST_F(ArrayElementwiseOpTest, CompareLeS32s) { const int32 min = std::numeric_limits::min(); const int32 max = std::numeric_limits::max(); XlaBuilder builder(TestName()); - auto lhs = builder.ConstantR1({min, min, min, 0, 0, 0, max, max, max}); - auto rhs = builder.ConstantR1({min, 0, max, -1, 0, 1, min, 0, max}); - builder.Le(lhs, rhs); + auto lhs = + ConstantR1(&builder, {min, min, min, 0, 0, 0, max, max, max}); + auto rhs = ConstantR1(&builder, {min, 0, max, -1, 0, 1, min, 0, max}); + Le(lhs, rhs); ComputeAndCompareR1( &builder, {true, true, true, false, true, true, false, false, true}, {}); @@ -1301,9 +1307,10 @@ XLA_TEST_F(ArrayElementwiseOpTest, CompareLtS32s) { const int32 min = std::numeric_limits::min(); const int32 max = std::numeric_limits::max(); XlaBuilder builder(TestName()); - auto lhs = builder.ConstantR1({min, min, min, 0, 0, 0, max, max, max}); - auto rhs = builder.ConstantR1({min, 0, max, -1, 0, 1, min, 0, max}); - builder.Lt(lhs, rhs); + auto lhs = + ConstantR1(&builder, {min, min, min, 0, 0, 0, max, max, max}); + auto rhs = ConstantR1(&builder, {min, 0, max, -1, 0, 1, min, 0, max}); + Lt(lhs, rhs); ComputeAndCompareR1( &builder, {false, true, true, false, false, true, false, false, false}, @@ -1313,9 +1320,9 @@ XLA_TEST_F(ArrayElementwiseOpTest, CompareLtS32s) { XLA_TEST_F(ArrayElementwiseOpTest, CompareEqU32s) { const uint32 max = std::numeric_limits::max(); XlaBuilder builder(TestName()); - auto lhs = builder.ConstantR1({0, 0, 0, 5, 5, 5, max, max, max}); - auto rhs = builder.ConstantR1({0, 1, max, 4, 5, 6, 0, 1, max}); - builder.Eq(lhs, rhs); + auto lhs = ConstantR1(&builder, {0, 0, 0, 5, 5, 5, max, max, max}); + auto rhs = ConstantR1(&builder, {0, 1, max, 4, 5, 6, 0, 1, max}); + Eq(lhs, rhs); ComputeAndCompareR1( &builder, {true, false, false, false, true, false, false, false, true}, @@ -1325,9 +1332,9 @@ XLA_TEST_F(ArrayElementwiseOpTest, CompareEqU32s) { XLA_TEST_F(ArrayElementwiseOpTest, CompareNeU32s) { const uint32 max = std::numeric_limits::max(); XlaBuilder builder(TestName()); - auto lhs = builder.ConstantR1({0, 0, 0, 5, 5, 5, max, max, max}); - auto rhs = builder.ConstantR1({0, 1, max, 4, 5, 6, 0, 1, max}); - builder.Ne(lhs, rhs); + auto lhs = ConstantR1(&builder, {0, 0, 0, 5, 5, 5, max, max, max}); + auto rhs = ConstantR1(&builder, {0, 1, max, 4, 5, 6, 0, 1, max}); + Ne(lhs, rhs); ComputeAndCompareR1( &builder, {false, true, true, true, false, true, true, true, false}, {}); @@ -1336,9 +1343,9 @@ XLA_TEST_F(ArrayElementwiseOpTest, CompareNeU32s) { XLA_TEST_F(ArrayElementwiseOpTest, CompareGeU32s) { const uint32 max = std::numeric_limits::max(); XlaBuilder builder(TestName()); - auto lhs = builder.ConstantR1({0, 0, 0, 5, 5, 5, max, max, max}); - auto rhs = builder.ConstantR1({0, 1, max, 4, 5, 6, 0, 1, max}); - builder.Ge(lhs, rhs); + auto lhs = ConstantR1(&builder, {0, 0, 0, 5, 5, 5, max, max, max}); + auto rhs = ConstantR1(&builder, {0, 1, max, 4, 5, 6, 0, 1, max}); + Ge(lhs, rhs); ComputeAndCompareR1( &builder, {true, false, false, true, true, false, true, true, true}, {}); @@ -1347,9 +1354,9 @@ XLA_TEST_F(ArrayElementwiseOpTest, CompareGeU32s) { XLA_TEST_F(ArrayElementwiseOpTest, CompareGtU32s) { const uint32 max = std::numeric_limits::max(); XlaBuilder builder(TestName()); - auto lhs = builder.ConstantR1({0, 0, 0, 5, 5, 5, max, max, max}); - auto rhs = builder.ConstantR1({0, 1, max, 4, 5, 6, 0, 1, max}); - builder.Gt(lhs, rhs); + auto lhs = ConstantR1(&builder, {0, 0, 0, 5, 5, 5, max, max, max}); + auto rhs = ConstantR1(&builder, {0, 1, max, 4, 5, 6, 0, 1, max}); + Gt(lhs, rhs); ComputeAndCompareR1( &builder, {false, false, false, true, false, false, true, true, false}, @@ -1359,9 +1366,9 @@ XLA_TEST_F(ArrayElementwiseOpTest, CompareGtU32s) { XLA_TEST_F(ArrayElementwiseOpTest, CompareLeU32s) { const uint32 max = std::numeric_limits::max(); XlaBuilder builder(TestName()); - auto lhs = builder.ConstantR1({0, 0, 0, 5, 5, 5, max, max, max}); - auto rhs = builder.ConstantR1({0, 1, max, 4, 5, 6, 0, 1, max}); - builder.Le(lhs, rhs); + auto lhs = ConstantR1(&builder, {0, 0, 0, 5, 5, 5, max, max, max}); + auto rhs = ConstantR1(&builder, {0, 1, max, 4, 5, 6, 0, 1, max}); + Le(lhs, rhs); ComputeAndCompareR1( &builder, {true, true, true, false, true, true, false, false, true}, {}); @@ -1370,9 +1377,9 @@ XLA_TEST_F(ArrayElementwiseOpTest, CompareLeU32s) { XLA_TEST_F(ArrayElementwiseOpTest, CompareLtU32s) { const uint32 max = std::numeric_limits::max(); XlaBuilder builder(TestName()); - auto lhs = builder.ConstantR1({0, 0, 0, 5, 5, 5, max, max, max}); - auto rhs = builder.ConstantR1({0, 1, max, 4, 5, 6, 0, 1, max}); - builder.Lt(lhs, rhs); + auto lhs = ConstantR1(&builder, {0, 0, 0, 5, 5, 5, max, max, max}); + auto rhs = ConstantR1(&builder, {0, 1, max, 4, 5, 6, 0, 1, max}); + Lt(lhs, rhs); ComputeAndCompareR1( &builder, {false, true, true, false, false, true, false, false, false}, @@ -1383,10 +1390,10 @@ XLA_TEST_F(ArrayElementwiseOpTest, PowF32s) { SetFastMathDisabled(true); XlaBuilder builder(TestName()); auto lhs = - builder.ConstantR1({4.0f, 2.0f, 2.0f, NAN, 6.0f, -2.0f, -2.0f}); + ConstantR1(&builder, {4.0f, 2.0f, 2.0f, NAN, 6.0f, -2.0f, -2.0f}); auto rhs = - builder.ConstantR1({2.0f, -2.0f, 3.0f, 10.0f, NAN, 3.0f, 4.0f}); - builder.Pow(lhs, rhs); + ConstantR1(&builder, {2.0f, -2.0f, 3.0f, 10.0f, NAN, 3.0f, 4.0f}); + Pow(lhs, rhs); ComputeAndCompareR1( &builder, {16.0f, 0.25f, 8.0f, NAN, NAN, -8.0f, 16.0f}, {}, error_spec_); @@ -1395,9 +1402,9 @@ XLA_TEST_F(ArrayElementwiseOpTest, PowF32s) { XLA_TEST_F(ArrayElementwiseOpTest, PowNonIntegerF32s) { SetFastMathDisabled(true); XlaBuilder builder(TestName()); - auto lhs = builder.ConstantR1({-2.0f, -0.6f, -0.6f, 0.0f}); - auto rhs = builder.ConstantR1({0.5f, 0.6f, -0.6f, -0.6f}); - builder.Pow(lhs, rhs); + auto lhs = ConstantR1(&builder, {-2.0f, -0.6f, -0.6f, 0.0f}); + auto rhs = ConstantR1(&builder, {0.5f, 0.6f, -0.6f, -0.6f}); + Pow(lhs, rhs); ComputeAndCompareR1(&builder, {NAN, NAN, NAN, INFINITY}, {}, error_spec_); @@ -1405,9 +1412,9 @@ XLA_TEST_F(ArrayElementwiseOpTest, PowNonIntegerF32s) { XLA_TEST_F(ArrayElementwiseOpTest, PowZeroElementF32s) { XlaBuilder builder(TestName()); - auto lhs = builder.ConstantR1({}); - auto rhs = builder.ConstantR1({}); - builder.Pow(lhs, rhs); + auto lhs = ConstantR1(&builder, {}); + auto rhs = ConstantR1(&builder, {}); + Pow(lhs, rhs); ComputeAndCompareR1(&builder, {}, {}, error_spec_); } @@ -1419,14 +1426,14 @@ XLA_TEST_F(ArrayElementwiseOpTest, PowSpecialF32) { std::vector values = {1.0f, 2.0f, 3.2f, -4.0f}; std::vector exponents = {0.0f, 1.0f, 2.0f, 0.5f, -1.0f, -0.5f}; - std::unique_ptr param_literal = Literal::CreateR1(values); + std::unique_ptr param_literal = LiteralUtil::CreateR1(values); std::unique_ptr param_data = client_->TransferToServer(*param_literal).ConsumeValueOrDie(); - auto sum = b.ConstantR0(0.0f); - auto param = b.Parameter(0, param_literal->shape(), "param"); + auto sum = ConstantR0(&b, 0.0f); + auto param = Parameter(&b, 0, param_literal->shape(), "param"); for (float exponent : exponents) { - sum = b.Add(sum, b.Pow(param, b.ConstantR0(exponent))); + sum = Add(sum, Pow(param, ConstantR0(&b, exponent))); } std::vector expected; @@ -1447,15 +1454,15 @@ XLA_TEST_F(ArrayElementwiseOpTest, PowOfExpF32) { std::vector values0 = {1.0f, 2.0f, 3.2f, -4.0f, 0.0f, 5.7f}; std::vector values1 = {0.0f, 1.0f, 2.0f, 0.5f, -1.0f, -0.5f}; - std::unique_ptr literal0 = Literal::CreateR1(values0); + std::unique_ptr literal0 = LiteralUtil::CreateR1(values0); std::unique_ptr data0 = client_->TransferToServer(*literal0).ConsumeValueOrDie(); - std::unique_ptr literal1 = Literal::CreateR1(values1); + std::unique_ptr literal1 = LiteralUtil::CreateR1(values1); std::unique_ptr data1 = client_->TransferToServer(*literal1).ConsumeValueOrDie(); - auto param0 = b.Parameter(0, literal0->shape(), "param0"); - auto param1 = b.Parameter(1, literal1->shape(), "param1"); - b.Pow(b.Exp(param0), param1); + auto param0 = Parameter(&b, 0, literal0->shape(), "param0"); + auto param1 = Parameter(&b, 1, literal1->shape(), "param1"); + Pow(Exp(param0), param1); std::vector expected(values0.size()); for (int64 i = 0; i < values0.size(); ++i) { @@ -1472,15 +1479,15 @@ XLA_TEST_F(ArrayElementwiseOpTest, LogOfPowerF32) { std::vector values0 = {1.0f, 2.0f, 3.2f, 4.0f, 0.5f, 5.7f}; std::vector values1 = {0.0f, 1.0f, 2.0f, 0.5f, -1.0f, -0.5f}; - std::unique_ptr literal0 = Literal::CreateR1(values0); + std::unique_ptr literal0 = LiteralUtil::CreateR1(values0); std::unique_ptr data0 = client_->TransferToServer(*literal0).ConsumeValueOrDie(); - std::unique_ptr literal1 = Literal::CreateR1(values1); + std::unique_ptr literal1 = LiteralUtil::CreateR1(values1); std::unique_ptr data1 = client_->TransferToServer(*literal1).ConsumeValueOrDie(); - auto param0 = b.Parameter(0, literal0->shape(), "param0"); - auto param1 = b.Parameter(1, literal1->shape(), "param1"); - b.Log(b.Pow(param0, param1)); + auto param0 = Parameter(&b, 0, literal0->shape(), "param0"); + auto param1 = Parameter(&b, 1, literal1->shape(), "param1"); + Log(Pow(param0, param1)); std::vector expected(values0.size()); for (int64 i = 0; i < values0.size(); ++i) { @@ -1497,15 +1504,15 @@ XLA_TEST_F(ArrayElementwiseOpTest, MulOfExpF32) { std::vector values0 = {1.0f, 2.0f, 3.2f, -4.0f, 0.0f, 5.7f}; std::vector values1 = {0.0f, 1.0f, 2.0f, 0.5f, -1.0f, -0.5f}; - std::unique_ptr literal0 = Literal::CreateR1(values0); + std::unique_ptr literal0 = LiteralUtil::CreateR1(values0); std::unique_ptr data0 = client_->TransferToServer(*literal0).ConsumeValueOrDie(); - std::unique_ptr literal1 = Literal::CreateR1(values1); + std::unique_ptr literal1 = LiteralUtil::CreateR1(values1); std::unique_ptr data1 = client_->TransferToServer(*literal1).ConsumeValueOrDie(); - auto param0 = b.Parameter(0, literal0->shape(), "param0"); - auto param1 = b.Parameter(1, literal1->shape(), "param1"); - b.Mul(b.Exp(param0), b.Exp(param1)); + auto param0 = Parameter(&b, 0, literal0->shape(), "param0"); + auto param1 = Parameter(&b, 1, literal1->shape(), "param1"); + Mul(Exp(param0), Exp(param1)); std::vector expected(values0.size()); for (int64 i = 0; i < values0.size(); ++i) { @@ -1522,15 +1529,15 @@ XLA_TEST_F(ArrayElementwiseOpTest, DivOfExpF32) { std::vector values0 = {1.0f, 2.0f, 3.2f, -4.0f, 0.0f, 5.7f}; std::vector values1 = {0.0f, 1.0f, 2.0f, 0.5f, -1.0f, -0.5f}; - std::unique_ptr literal0 = Literal::CreateR1(values0); + std::unique_ptr literal0 = LiteralUtil::CreateR1(values0); std::unique_ptr data0 = client_->TransferToServer(*literal0).ConsumeValueOrDie(); - std::unique_ptr literal1 = Literal::CreateR1(values1); + std::unique_ptr literal1 = LiteralUtil::CreateR1(values1); std::unique_ptr data1 = client_->TransferToServer(*literal1).ConsumeValueOrDie(); - auto param0 = b.Parameter(0, literal0->shape(), "param0"); - auto param1 = b.Parameter(1, literal1->shape(), "param1"); - b.Div(param0, b.Exp(param1)); + auto param0 = Parameter(&b, 0, literal0->shape(), "param0"); + auto param1 = Parameter(&b, 1, literal1->shape(), "param1"); + Div(param0, Exp(param1)); std::vector expected(values0.size()); for (int64 i = 0; i < values0.size(); ++i) { @@ -1548,21 +1555,21 @@ XLA_TEST_F(ArrayElementwiseOpTest, Div3_lhs_F32) { std::vector values1 = {0.1f, 1.0f, 2.0f, 0.5f, -1.0f, -0.5f}; std::vector values2 = {0.1f, 1.1f, 6.9f, 12.5f, -15.0f, -0.5f}; - std::unique_ptr literal0 = Literal::CreateR1(values0); + std::unique_ptr literal0 = LiteralUtil::CreateR1(values0); std::unique_ptr data0 = client_->TransferToServer(*literal0).ConsumeValueOrDie(); - std::unique_ptr literal1 = Literal::CreateR1(values1); + std::unique_ptr literal1 = LiteralUtil::CreateR1(values1); std::unique_ptr data1 = client_->TransferToServer(*literal1).ConsumeValueOrDie(); - std::unique_ptr literal2 = Literal::CreateR1(values2); + std::unique_ptr literal2 = LiteralUtil::CreateR1(values2); std::unique_ptr data2 = client_->TransferToServer(*literal2).ConsumeValueOrDie(); - auto param0 = b.Parameter(0, literal0->shape(), "param0"); - auto param1 = b.Parameter(1, literal1->shape(), "param1"); - auto param2 = b.Parameter(2, literal2->shape(), "param2"); - b.Div(b.Div(param0, param1), param2); + auto param0 = Parameter(&b, 0, literal0->shape(), "param0"); + auto param1 = Parameter(&b, 1, literal1->shape(), "param1"); + auto param2 = Parameter(&b, 2, literal2->shape(), "param2"); + Div(Div(param0, param1), param2); std::vector expected(values0.size()); for (int64 i = 0; i < values0.size(); ++i) { @@ -1580,22 +1587,22 @@ XLA_TEST_F(ArrayElementwiseOpTest, Div3_rhs_F32) { std::vector values1 = {0.1f, 1.0f, 2.0f, 0.5f, -1.0f, -0.5f}; std::vector values2 = {0.1f, 1.1f, 6.9f, 12.5f, -15.0f, -0.5f}; - std::unique_ptr literal0 = Literal::CreateR1(values0); + std::unique_ptr literal0 = LiteralUtil::CreateR1(values0); std::unique_ptr data0 = client_->TransferToServer(*literal0).ConsumeValueOrDie(); - std::unique_ptr literal1 = Literal::CreateR1(values1); + std::unique_ptr literal1 = LiteralUtil::CreateR1(values1); std::unique_ptr data1 = client_->TransferToServer(*literal1).ConsumeValueOrDie(); - std::unique_ptr literal2 = Literal::CreateR1(values2); + std::unique_ptr literal2 = LiteralUtil::CreateR1(values2); std::unique_ptr data2 = client_->TransferToServer(*literal2).ConsumeValueOrDie(); - auto param0 = b.Parameter(0, literal0->shape(), "param0"); - auto param1 = b.Parameter(1, literal1->shape(), "param1"); - auto param2 = b.Parameter(2, literal2->shape(), "param2"); - b.Div(param0, b.Div(param1, param2)); + auto param0 = Parameter(&b, 0, literal0->shape(), "param0"); + auto param1 = Parameter(&b, 1, literal1->shape(), "param1"); + auto param2 = Parameter(&b, 2, literal2->shape(), "param2"); + Div(param0, Div(param1, param2)); std::vector expected(values0.size()); for (int64 i = 0; i < values0.size(); ++i) { @@ -1613,22 +1620,22 @@ XLA_TEST_F(ArrayElementwiseOpTest, DivOfPowerF32) { std::vector values1 = {0.1f, 1.0f, 2.0f, 0.5f, 1.0f, 0.5f}; std::vector values2 = {0.1f, 1.1f, 6.9f, 9.5f, -11.0f, -0.5f}; - std::unique_ptr literal0 = Literal::CreateR1(values0); + std::unique_ptr literal0 = LiteralUtil::CreateR1(values0); std::unique_ptr data0 = client_->TransferToServer(*literal0).ConsumeValueOrDie(); - std::unique_ptr literal1 = Literal::CreateR1(values1); + std::unique_ptr literal1 = LiteralUtil::CreateR1(values1); std::unique_ptr data1 = client_->TransferToServer(*literal1).ConsumeValueOrDie(); - std::unique_ptr literal2 = Literal::CreateR1(values2); + std::unique_ptr literal2 = LiteralUtil::CreateR1(values2); std::unique_ptr data2 = client_->TransferToServer(*literal2).ConsumeValueOrDie(); - auto param0 = b.Parameter(0, literal0->shape(), "param0"); - auto param1 = b.Parameter(1, literal1->shape(), "param1"); - auto param2 = b.Parameter(2, literal2->shape(), "param2"); - b.Div(param0, b.Pow(param1, param2)); + auto param0 = Parameter(&b, 0, literal0->shape(), "param0"); + auto param1 = Parameter(&b, 1, literal1->shape(), "param1"); + auto param2 = Parameter(&b, 2, literal2->shape(), "param2"); + Div(param0, Pow(param1, param2)); std::vector expected(values0.size()); for (int64 i = 0; i < values0.size(); ++i) { @@ -1647,27 +1654,27 @@ XLA_TEST_F(ArrayElementwiseOpTest, Div4F32) { std::vector values2 = {0.1f, 1.1f, 6.9f, 12.5f, -15.0f, -0.5f}; std::vector values3 = {2.1f, 3.1f, 9.9f, -4.5f, -11.0f, -21.5f}; - std::unique_ptr literal0 = Literal::CreateR1(values0); + std::unique_ptr literal0 = LiteralUtil::CreateR1(values0); std::unique_ptr data0 = client_->TransferToServer(*literal0).ConsumeValueOrDie(); - std::unique_ptr literal1 = Literal::CreateR1(values1); + std::unique_ptr literal1 = LiteralUtil::CreateR1(values1); std::unique_ptr data1 = client_->TransferToServer(*literal1).ConsumeValueOrDie(); - std::unique_ptr literal2 = Literal::CreateR1(values2); + std::unique_ptr literal2 = LiteralUtil::CreateR1(values2); std::unique_ptr data2 = client_->TransferToServer(*literal2).ConsumeValueOrDie(); - std::unique_ptr literal3 = Literal::CreateR1(values3); + std::unique_ptr literal3 = LiteralUtil::CreateR1(values3); std::unique_ptr data3 = client_->TransferToServer(*literal3).ConsumeValueOrDie(); - auto param0 = b.Parameter(0, literal0->shape(), "param0"); - auto param1 = b.Parameter(1, literal1->shape(), "param1"); - auto param2 = b.Parameter(2, literal2->shape(), "param2"); - auto param3 = b.Parameter(3, literal3->shape(), "param2"); - b.Div(b.Div(param0, param1), b.Div(param2, param3)); + auto param0 = Parameter(&b, 0, literal0->shape(), "param0"); + auto param1 = Parameter(&b, 1, literal1->shape(), "param1"); + auto param2 = Parameter(&b, 2, literal2->shape(), "param2"); + auto param3 = Parameter(&b, 3, literal3->shape(), "param2"); + Div(Div(param0, param1), Div(param2, param3)); std::vector expected(values0.size()); for (int64 i = 0; i < values0.size(); ++i) { @@ -1687,8 +1694,8 @@ TEST_P(ArrayElementwiseOpTestParamCount, SquareManyValues) { for (int i = 0; i < count; ++i) { values.push_back(i / static_cast(count)); } - auto x = builder.ConstantR1(values); - builder.Pow(x, builder.ConstantR0(2.0f)); + auto x = ConstantR1(&builder, values); + Pow(x, ConstantR0(&builder, 2.0f)); std::vector expected; expected.reserve(values.size()); @@ -1713,8 +1720,8 @@ XLA_TEST_F(ArrayElementwiseOpTest, SquareIn4D) { Array4D expected(2, 2, 2, 2, expected_vector); - auto x = builder.ConstantR4FromArray4D(values); - builder.Pow(x, builder.ConstantR0(2.0f)); + auto x = ConstantR4FromArray4D(&builder, values); + Pow(x, ConstantR0(&builder, 2.0f)); ComputeAndCompareR4(&builder, expected, {}, error_spec_); } @@ -1724,8 +1731,8 @@ XLA_TEST_F(ArrayElementwiseOpTest, SquareIn4DZeroElements) { Array4D values(2, 2, 0, 2); Array4D expected(2, 2, 0, 2); - auto x = builder.ConstantR4FromArray4D(values); - builder.Pow(x, builder.ConstantR0(2.0f)); + auto x = ConstantR4FromArray4D(&builder, values); + Pow(x, ConstantR0(&builder, 2.0f)); ComputeAndCompareR4(&builder, expected, {}, error_spec_); } @@ -1733,9 +1740,9 @@ XLA_TEST_F(ArrayElementwiseOpTest, SquareIn4DZeroElements) { XLA_TEST_F(ArrayElementwiseOpTest, MinF32s) { XlaBuilder builder(TestName()); SetFastMathDisabled(true); - auto lhs = builder.ConstantR1({1.0f, 1.0f, 2.25f, NAN, 6.0f}); - auto rhs = builder.ConstantR1({2.0f, -5.0f, 1.0f, 10.0f, NAN}); - builder.Min(lhs, rhs); + auto lhs = ConstantR1(&builder, {1.0f, 1.0f, 2.25f, NAN, 6.0f}); + auto rhs = ConstantR1(&builder, {2.0f, -5.0f, 1.0f, 10.0f, NAN}); + Min(lhs, rhs); ComputeAndCompareR1(&builder, {1.0f, -5.0f, 1.0f, NAN, NAN}, {}, error_spec_); @@ -1743,18 +1750,18 @@ XLA_TEST_F(ArrayElementwiseOpTest, MinF32s) { XLA_TEST_F(ArrayElementwiseOpTest, MinZeroElementF32s) { XlaBuilder builder(TestName()); - auto lhs = builder.ConstantR1({}); - auto rhs = builder.ConstantR1({}); - builder.Min(lhs, rhs); + auto lhs = ConstantR1(&builder, {}); + auto rhs = ConstantR1(&builder, {}); + Min(lhs, rhs); ComputeAndCompareR1(&builder, {}, {}, error_spec_); } XLA_TEST_F(ArrayElementwiseOpTest, MinF64s) { XlaBuilder builder(TestName()); SetFastMathDisabled(true); - auto lhs = builder.ConstantR1({1.0, 1.0, 2.25, NAN, 6.0}); - auto rhs = builder.ConstantR1({2.0, -5.0, 1.0, 10.0, NAN}); - builder.Min(lhs, rhs); + auto lhs = ConstantR1(&builder, {1.0, 1.0, 2.25, NAN, 6.0}); + auto rhs = ConstantR1(&builder, {2.0, -5.0, 1.0, 10.0, NAN}); + Min(lhs, rhs); ComputeAndCompareR1(&builder, {1.0, -5.0, 1.0, NAN, NAN}, {}, error_spec_); @@ -1763,9 +1770,9 @@ XLA_TEST_F(ArrayElementwiseOpTest, MinF64s) { XLA_TEST_F(ArrayElementwiseOpTest, MaxF32s) { XlaBuilder builder(TestName()); SetFastMathDisabled(true); - auto lhs = builder.ConstantR1({1.0f, 1.0f, 2.25f, NAN, 6.0f}); - auto rhs = builder.ConstantR1({2.0f, -5.0f, 1.0f, 10.0f, NAN}); - builder.Max(lhs, rhs); + auto lhs = ConstantR1(&builder, {1.0f, 1.0f, 2.25f, NAN, 6.0f}); + auto rhs = ConstantR1(&builder, {2.0f, -5.0f, 1.0f, 10.0f, NAN}); + Max(lhs, rhs); ComputeAndCompareR1(&builder, {2.0f, 1.0f, 2.25f, NAN, NAN}, {}, error_spec_); @@ -1773,18 +1780,18 @@ XLA_TEST_F(ArrayElementwiseOpTest, MaxF32s) { XLA_TEST_F(ArrayElementwiseOpTest, MaxZeroElementF32s) { XlaBuilder builder(TestName()); - auto lhs = builder.ConstantR1({}); - auto rhs = builder.ConstantR1({}); - builder.Max(lhs, rhs); + auto lhs = ConstantR1(&builder, {}); + auto rhs = ConstantR1(&builder, {}); + Max(lhs, rhs); ComputeAndCompareR1(&builder, {}, {}, error_spec_); } XLA_TEST_F(ArrayElementwiseOpTest, MaxF64s) { XlaBuilder builder(TestName()); SetFastMathDisabled(true); - auto lhs = builder.ConstantR1({1.0, 1.0, 2.25, NAN, 6.0}); - auto rhs = builder.ConstantR1({2.0, -5.0, 1.0, 10.0, NAN}); - builder.Max(lhs, rhs); + auto lhs = ConstantR1(&builder, {1.0, 1.0, 2.25, NAN, 6.0}); + auto rhs = ConstantR1(&builder, {2.0, -5.0, 1.0, 10.0, NAN}); + Max(lhs, rhs); ComputeAndCompareR1(&builder, {2.0, 1.0, 2.25, NAN, NAN}, {}, error_spec_); @@ -1794,11 +1801,11 @@ XLA_TEST_F(ArrayElementwiseOpTest, MaxS32s) { const int32 min = std::numeric_limits::min(); const int32 max = std::numeric_limits::max(); XlaBuilder builder(TestName()); - auto x = builder.ConstantR1( - {min, min, min, -1, -1, 0, 0, 0, 1, 1, max, max, max}); - auto y = builder.ConstantR1( - {min, max, 0, -10, 0, -1, 0, 1, 0, 10, 0, max, min}); - builder.Max(x, y); + auto x = ConstantR1( + &builder, {min, min, min, -1, -1, 0, 0, 0, 1, 1, max, max, max}); + auto y = ConstantR1( + &builder, {min, max, 0, -10, 0, -1, 0, 1, 0, 10, 0, max, min}); + Max(x, y); std::vector expected = {min, max, 0, -1, 0, 0, 0, 1, 1, 10, max, max, max}; @@ -1809,11 +1816,11 @@ XLA_TEST_F(ArrayElementwiseOpTest, MinS32s) { const int32 min = std::numeric_limits::min(); const int32 max = std::numeric_limits::max(); XlaBuilder builder(TestName()); - auto x = builder.ConstantR1( - {min, min, min, -1, -1, 0, 0, 0, 1, 1, max, max, max}); - auto y = builder.ConstantR1( - {min, max, 0, -10, 0, -1, 0, 1, 0, 10, 0, max, min}); - builder.Min(x, y); + auto x = ConstantR1( + &builder, {min, min, min, -1, -1, 0, 0, 0, 1, 1, max, max, max}); + auto y = ConstantR1( + &builder, {min, max, 0, -10, 0, -1, 0, 1, 0, 10, 0, max, min}); + Min(x, y); std::vector expected = {min, min, min, -10, -1, -1, 0, 0, 0, 1, 0, max, min}; @@ -1823,9 +1830,9 @@ XLA_TEST_F(ArrayElementwiseOpTest, MinS32s) { XLA_TEST_F(ArrayElementwiseOpTest, MaxU32s) { const uint32 max = std::numeric_limits::max(); XlaBuilder builder(TestName()); - auto x = builder.ConstantR1({0, 0, 1, 1, 1, max, max, max}); - auto y = builder.ConstantR1({0, 1, 0, 1, 10, 0, 234234, max}); - builder.Max(x, y); + auto x = ConstantR1(&builder, {0, 0, 1, 1, 1, max, max, max}); + auto y = ConstantR1(&builder, {0, 1, 0, 1, 10, 0, 234234, max}); + Max(x, y); std::vector expected = {0, 1, 1, 1, 10, max, max, max}; ComputeAndCompareR1(&builder, expected, {}); @@ -1834,9 +1841,9 @@ XLA_TEST_F(ArrayElementwiseOpTest, MaxU32s) { XLA_TEST_F(ArrayElementwiseOpTest, MinU32s) { const uint32 max = std::numeric_limits::max(); XlaBuilder builder(TestName()); - auto x = builder.ConstantR1({0, 0, 1, 1, 1, max, max, max}); - auto y = builder.ConstantR1({0, 1, 0, 1, 10, 0, 234234, max}); - builder.Min(x, y); + auto x = ConstantR1(&builder, {0, 0, 1, 1, 1, max, max, max}); + auto y = ConstantR1(&builder, {0, 1, 0, 1, 10, 0, 234234, max}); + Min(x, y); std::vector expected = {0, 0, 0, 1, 1, 0, 234234, max}; ComputeAndCompareR1(&builder, expected, {}); @@ -1844,11 +1851,11 @@ XLA_TEST_F(ArrayElementwiseOpTest, MinU32s) { XLA_TEST_F(ArrayElementwiseOpTest, MaxTenF32s) { XlaBuilder builder(TestName()); - auto x = builder.ConstantR1( - {-0.0, 1.0, 2.0, -3.0, -4.0, 5.0, 6.0, -7.0, -8.0, 9.0}); - auto y = builder.ConstantR1( - {-0.0, -1.0, -2.0, 3.0, 4.0, -5.0, -6.0, 7.0, 8.0, -9.0}); - builder.Max(x, y); + auto x = ConstantR1( + &builder, {-0.0, 1.0, 2.0, -3.0, -4.0, 5.0, 6.0, -7.0, -8.0, 9.0}); + auto y = ConstantR1( + &builder, {-0.0, -1.0, -2.0, 3.0, 4.0, -5.0, -6.0, 7.0, 8.0, -9.0}); + Max(x, y); std::vector expected = {-0.0, 1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0}; @@ -1857,9 +1864,9 @@ XLA_TEST_F(ArrayElementwiseOpTest, MaxTenF32s) { XLA_TEST_F(ArrayElementwiseOpTest, MaxR1S1AndR1S0F32s) { XlaBuilder builder(TestName()); - auto u = builder.ConstantR1({3.5}); - auto v = builder.ConstantR1({}); - builder.Max(u, v); + auto u = ConstantR1(&builder, {3.5}); + auto v = ConstantR1(&builder, {}); + Max(u, v); ComputeAndCompareR1(&builder, {}, {}, error_spec_); } @@ -1867,9 +1874,9 @@ XLA_TEST_F(ArrayElementwiseOpTest, MaxR1S1AndR1S0F32s) { XLA_TEST_F(ArrayElementwiseOpTest, MaxR1S0AndR2S0x2F32s) { for (int broadcast_dim : {0, 1}) { XlaBuilder builder(TestName()); - auto u = builder.ConstantR1({3.5}); - auto v = builder.ConstantR2FromArray2D(Array2D(0, 2)); - builder.Max(u, v, /*broadcast_dimensions=*/{broadcast_dim}); + auto u = ConstantR1(&builder, {3.5}); + auto v = ConstantR2FromArray2D(&builder, Array2D(0, 2)); + Max(u, v, /*broadcast_dimensions=*/{broadcast_dim}); ComputeAndCompareR2(&builder, Array2D(0, 2), {}, error_spec_); } @@ -1877,10 +1884,10 @@ XLA_TEST_F(ArrayElementwiseOpTest, MaxR1S0AndR2S0x2F32s) { XLA_TEST_F(ArrayElementwiseOpTest, Max1DAnd2DF32s) { XlaBuilder builder(TestName()); - auto v = builder.ConstantR1({2.0f, 3.0f, 4.0f}); - auto m = - builder.ConstantR2({{-2.5f, 3.14f, 1.0f}, {2.25f, -10.0f, 3.33f}}); - builder.Max(v, m, /*broadcast_dimensions=*/{1}); + auto v = ConstantR1(&builder, {2.0f, 3.0f, 4.0f}); + auto m = ConstantR2(&builder, + {{-2.5f, 3.14f, 1.0f}, {2.25f, -10.0f, 3.33f}}); + Max(v, m, /*broadcast_dimensions=*/{1}); Array2D expected({{2.0f, 3.14f, 4.0f}, {2.25f, 3.0f, 4.0f}}); ComputeAndCompareR2(&builder, expected, {}, error_spec_); @@ -1888,9 +1895,9 @@ XLA_TEST_F(ArrayElementwiseOpTest, Max1DAnd2DF32s) { XLA_TEST_F(ArrayElementwiseOpTest, Max1DAnd2DZeroElementF32s) { XlaBuilder builder(TestName()); - auto v = builder.ConstantR1({}); - auto m = builder.ConstantR2({{}, {}}); - builder.Max(v, m, /*broadcast_dimensions=*/{1}); + auto v = ConstantR1(&builder, {}); + auto m = ConstantR2(&builder, {{}, {}}); + Max(v, m, /*broadcast_dimensions=*/{1}); Array2D expected({{}, {}}); ComputeAndCompareR2(&builder, expected, {}, error_spec_); @@ -1898,10 +1905,10 @@ XLA_TEST_F(ArrayElementwiseOpTest, Max1DAnd2DZeroElementF32s) { XLA_TEST_F(ArrayElementwiseOpTest, Max3DAndScalarS32s) { XlaBuilder builder(TestName()); - auto scalar = builder.ConstantR0(2); + auto scalar = ConstantR0(&builder, 2); Array3D a_3d({{{3, 9, -1}, {2, -10, 3}}, {{-2, 2, 8}, {12, 10, 4}}}); - auto array = builder.ConstantR3FromArray3D(a_3d); - builder.Max(array, scalar, /*broadcast_dimensions=*/{}); + auto array = ConstantR3FromArray3D(&builder, a_3d); + Max(array, scalar, /*broadcast_dimensions=*/{}); Array3D expected({{{3, 9, 2}, {2, 2, 3}}, {{2, 2, 8}, {12, 10, 4}}}); ComputeAndCompareR3(&builder, expected, {}); @@ -1909,10 +1916,10 @@ XLA_TEST_F(ArrayElementwiseOpTest, Max3DAndScalarS32s) { XLA_TEST_F(ArrayElementwiseOpTest, Max3DAndScalarZeroElementS32s) { XlaBuilder builder(TestName()); - auto scalar = builder.ConstantR0(2); + auto scalar = ConstantR0(&builder, 2); Array3D a_3d(2, 0, 3); - auto array = builder.ConstantR3FromArray3D(a_3d); - builder.Max(array, scalar, /*broadcast_dimensions=*/{}); + auto array = ConstantR3FromArray3D(&builder, a_3d); + Max(array, scalar, /*broadcast_dimensions=*/{}); Array3D expected(2, 0, 3); ComputeAndCompareR3(&builder, expected, {}); @@ -1920,10 +1927,10 @@ XLA_TEST_F(ArrayElementwiseOpTest, Max3DAndScalarZeroElementS32s) { XLA_TEST_F(ArrayElementwiseOpTest, Min2DTo1DF32s) { XlaBuilder builder(TestName()); - auto m = - builder.ConstantR2({{-10.4f, 64.0f, 6.0f}, {0.1f, 32.0f, 16.1f}}); - auto v = builder.ConstantR1({-10.2f, 16.4f}); - builder.Min(m, v, /*broadcast_dimensions=*/{0}); + auto m = ConstantR2(&builder, + {{-10.4f, 64.0f, 6.0f}, {0.1f, 32.0f, 16.1f}}); + auto v = ConstantR1(&builder, {-10.2f, 16.4f}); + Min(m, v, /*broadcast_dimensions=*/{0}); Array2D expected({{-10.4f, -10.2f, -10.2f}, {0.1f, 16.4f, 16.1f}}); ComputeAndCompareR2(&builder, expected, {}, error_spec_); @@ -1931,9 +1938,9 @@ XLA_TEST_F(ArrayElementwiseOpTest, Min2DTo1DF32s) { XLA_TEST_F(ArrayElementwiseOpTest, Min2DTo1DZeroElementF32s) { XlaBuilder builder(TestName()); - auto m = builder.ConstantR2({{}, {}}); - auto v = builder.ConstantR1({-10.2f, 16.4f}); - builder.Min(m, v, /*broadcast_dimensions=*/{0}); + auto m = ConstantR2(&builder, {{}, {}}); + auto v = ConstantR1(&builder, {-10.2f, 16.4f}); + Min(m, v, /*broadcast_dimensions=*/{0}); Array2D expected({{}, {}}); ComputeAndCompareR2(&builder, expected, {}, error_spec_); @@ -1942,11 +1949,11 @@ XLA_TEST_F(ArrayElementwiseOpTest, Min2DTo1DZeroElementF32s) { XLA_TEST_F(ArrayElementwiseOpTest, Min2DTo4DF32s) { XlaBuilder builder(TestName()); auto array2d = - builder.ConstantR2({{-12.2f, 64.3f, 6.1f}, {0.0f, 32.2f, 2.5f}}); - auto array4d = builder.ConstantR4FromArray4D( - {{{{-12.1f, 32.3f, 6.2f}}, {{0.0f, 32.5f, 3.0f}}}, - {{{-2.5f, 64.29f, 6.5f}}, {{-0.01f, 32.25f, 2.6f}}}}); - builder.Min(array2d, array4d, /*broadcast_dimensions=*/{1, 3}); + ConstantR2(&builder, {{-12.2f, 64.3f, 6.1f}, {0.0f, 32.2f, 2.5f}}); + auto array4d = ConstantR4FromArray4D( + &builder, {{{{-12.1f, 32.3f, 6.2f}}, {{0.0f, 32.5f, 3.0f}}}, + {{{-2.5f, 64.29f, 6.5f}}, {{-0.01f, 32.25f, 2.6f}}}}); + Min(array2d, array4d, /*broadcast_dimensions=*/{1, 3}); Array4D expected( {{{{-12.2f, 32.3f, 6.1f}}, {{0.0f, 32.2f, 2.5f}}}, @@ -1957,10 +1964,10 @@ XLA_TEST_F(ArrayElementwiseOpTest, Min2DTo4DF32s) { XLA_TEST_F(ArrayElementwiseOpTest, Min2DTo4DZeroElementF32s) { XlaBuilder builder(TestName()); auto array2d = - builder.ConstantR2({{-12.2f, 64.3f, 6.1f}, {0.0f, 32.2f, 2.5f}}); + ConstantR2(&builder, {{-12.2f, 64.3f, 6.1f}, {0.0f, 32.2f, 2.5f}}); Array4D arg(2, 2, 0, 3); - auto array4d = builder.ConstantR4FromArray4D(arg); - builder.Min(array2d, array4d, /*broadcast_dimensions=*/{1, 3}); + auto array4d = ConstantR4FromArray4D(&builder, arg); + Min(array2d, array4d, /*broadcast_dimensions=*/{1, 3}); Array4D expected(2, 2, 0, 3); ComputeAndCompareR4(&builder, expected, {}, error_spec_); @@ -1968,9 +1975,9 @@ XLA_TEST_F(ArrayElementwiseOpTest, Min2DTo4DZeroElementF32s) { XLA_TEST_F(ArrayElementwiseOpTest, MinTenS32s) { XlaBuilder builder(TestName()); - auto x = builder.ConstantR1({0, 1, 2, 3, 4, 5, 6, 7, 8, 9}); - auto y = builder.ConstantR1({9, 8, 7, 6, 5, 4, 3, 2, 1, 0}); - builder.Min(x, y); + auto x = ConstantR1(&builder, {0, 1, 2, 3, 4, 5, 6, 7, 8, 9}); + auto y = ConstantR1(&builder, {9, 8, 7, 6, 5, 4, 3, 2, 1, 0}); + Min(x, y); std::vector expected = {0, 1, 2, 3, 4, 4, 3, 2, 1, 0}; ComputeAndCompareR1(&builder, expected, {}); @@ -1978,9 +1985,9 @@ XLA_TEST_F(ArrayElementwiseOpTest, MinTenS32s) { XLA_TEST_F(ArrayElementwiseOpTest, MaxTenS32s) { XlaBuilder builder(TestName()); - auto x = builder.ConstantR1({0, 1, 2, 3, 4, 5, 6, 7, 8, 9}); - auto y = builder.ConstantR1({9, 8, 7, 6, 5, 4, 3, 2, 1, 0}); - builder.Max(x, y); + auto x = ConstantR1(&builder, {0, 1, 2, 3, 4, 5, 6, 7, 8, 9}); + auto y = ConstantR1(&builder, {9, 8, 7, 6, 5, 4, 3, 2, 1, 0}); + Max(x, y); std::vector expected = {9, 8, 7, 6, 5, 5, 6, 7, 8, 9}; ComputeAndCompareR1(&builder, expected, {}); @@ -1988,19 +1995,20 @@ XLA_TEST_F(ArrayElementwiseOpTest, MaxTenS32s) { XLA_TEST_F(ArrayElementwiseOpTest, RemTwoConstantS32s) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({-3, 26, 2, -1, 1}); - auto b = builder.ConstantR1({10, 5, 1, 10, -10}); - builder.Rem(a, b); + auto a = ConstantR1(&builder, {-3, 26, 2, -1, 1}); + auto b = ConstantR1(&builder, {10, 5, 1, 10, -10}); + Rem(a, b); ComputeAndCompareR1(&builder, {-3, 1, 0, -1, 1}, {}); } XLA_TEST_F(ArrayElementwiseOpTest, NonNanClampF32) { XlaBuilder builder(TestName()); - auto minimum = builder.ConstantR1({1.0f, -6.5f, 1.0f, 2.25f, 0.0f}); - auto argument = builder.ConstantR1({2.0f, 10.0f, -5.0f, 1.0f, 10.0f}); - auto maximum = builder.ConstantR1({3.0f, 0.5f, 25.5f, 5.0f, 123.0}); - builder.Clamp(minimum, argument, maximum); + auto minimum = ConstantR1(&builder, {1.0f, -6.5f, 1.0f, 2.25f, 0.0f}); + auto argument = + ConstantR1(&builder, {2.0f, 10.0f, -5.0f, 1.0f, 10.0f}); + auto maximum = ConstantR1(&builder, {3.0f, 0.5f, 25.5f, 5.0f, 123.0}); + Clamp(minimum, argument, maximum); ComputeAndCompareR1(&builder, {2.0f, 0.5f, 1.0f, 2.25f, 10.0f}, {}, error_spec_); @@ -2008,10 +2016,10 @@ XLA_TEST_F(ArrayElementwiseOpTest, NonNanClampF32) { XLA_TEST_F(ArrayElementwiseOpTest, ClampF32Scalar) { XlaBuilder builder(TestName()); - auto minimum = builder.ConstantR0(0.0f); - auto argument = builder.ConstantR1({2.0f, 10.0f, -5.0f, 1.0f, 4.0f}); - auto maximum = builder.ConstantR0(5.0f); - builder.Clamp(minimum, argument, maximum); + auto minimum = ConstantR0(&builder, 0.0f); + auto argument = ConstantR1(&builder, {2.0f, 10.0f, -5.0f, 1.0f, 4.0f}); + auto maximum = ConstantR0(&builder, 5.0f); + Clamp(minimum, argument, maximum); ComputeAndCompareR1(&builder, {2.0f, 5.0f, 0.0f, 1.0f, 4.0f}, {}, error_spec_); @@ -2019,16 +2027,19 @@ XLA_TEST_F(ArrayElementwiseOpTest, ClampF32Scalar) { XLA_TEST_F(ArrayElementwiseOpTest, ClampF32ScalarVector) { XlaBuilder builder(TestName()); - auto min_scalar = builder.ConstantR0(0.0f); - auto min_vector = builder.ConstantR1({1.0f, -6.5f, 1.0f, 2.25f, 0.0f}); - auto arg_vector = builder.ConstantR1({2.0f, 10.0f, -5.0f, 1.0f, 4.0f}); - auto max_scalar = builder.ConstantR0(3.0f); - auto max_vector = builder.ConstantR1({3.0f, 0.5f, 25.5f, 5.0f, 123.0}); + auto min_scalar = ConstantR0(&builder, 0.0f); + auto min_vector = + ConstantR1(&builder, {1.0f, -6.5f, 1.0f, 2.25f, 0.0f}); + auto arg_vector = + ConstantR1(&builder, {2.0f, 10.0f, -5.0f, 1.0f, 4.0f}); + auto max_scalar = ConstantR0(&builder, 3.0f); + auto max_vector = + ConstantR1(&builder, {3.0f, 0.5f, 25.5f, 5.0f, 123.0}); // Perform clamp with broadcasted scalar and vector. - builder.Add(builder.Add(builder.Clamp(min_vector, arg_vector, max_scalar), - builder.Clamp(min_scalar, arg_vector, max_vector)), - builder.Add(builder.Clamp(min_vector, arg_vector, max_vector), - builder.Clamp(min_scalar, arg_vector, max_scalar))); + Add(Add(Clamp(min_vector, arg_vector, max_scalar), + Clamp(min_scalar, arg_vector, max_vector)), + Add(Clamp(min_vector, arg_vector, max_vector), + Clamp(min_scalar, arg_vector, max_scalar))); ComputeAndCompareR1(&builder, {8.0f, 7.0f, 2.0f, 6.5f, 14.0f}, {}, error_spec_); @@ -2036,52 +2047,52 @@ XLA_TEST_F(ArrayElementwiseOpTest, ClampF32ScalarVector) { XLA_TEST_F(ArrayElementwiseOpTest, ClampS32Vector) { XlaBuilder builder(TestName()); - auto min_vector = builder.ConstantR1({1, -6, 1, 2, 0, -5}); - auto arg_vector = builder.ConstantR1({2, 10, -5, 1, 4, 10}); - auto max_vector = builder.ConstantR1({3, 0, 25, 5, 123, -1}); - builder.Clamp(min_vector, arg_vector, max_vector); + auto min_vector = ConstantR1(&builder, {1, -6, 1, 2, 0, -5}); + auto arg_vector = ConstantR1(&builder, {2, 10, -5, 1, 4, 10}); + auto max_vector = ConstantR1(&builder, {3, 0, 25, 5, 123, -1}); + Clamp(min_vector, arg_vector, max_vector); ComputeAndCompareR1(&builder, {2, 0, 1, 2, 4, -1}, {}); } XLA_TEST_F(ArrayElementwiseOpTest, ClampS32ScalarVector) { XlaBuilder builder(TestName()); - auto min_scalar = builder.ConstantR0(0); - auto min_vector = builder.ConstantR1({1, -6, 1, 2, 0}); - auto arg_vector = builder.ConstantR1({2, 10, -5, 1, 4}); - auto max_scalar = builder.ConstantR0(3); - auto max_vector = builder.ConstantR1({3, 1, 25, 5, 123}); + auto min_scalar = ConstantR0(&builder, 0); + auto min_vector = ConstantR1(&builder, {1, -6, 1, 2, 0}); + auto arg_vector = ConstantR1(&builder, {2, 10, -5, 1, 4}); + auto max_scalar = ConstantR0(&builder, 3); + auto max_vector = ConstantR1(&builder, {3, 1, 25, 5, 123}); // Perform clamp with broadcasted scalar and vector. - builder.Add(builder.Add(builder.Clamp(min_vector, arg_vector, max_scalar), - builder.Clamp(min_scalar, arg_vector, max_vector)), - builder.Add(builder.Clamp(min_vector, arg_vector, max_vector), - builder.Clamp(min_scalar, arg_vector, max_scalar))); + Add(Add(Clamp(min_vector, arg_vector, max_scalar), + Clamp(min_scalar, arg_vector, max_vector)), + Add(Clamp(min_vector, arg_vector, max_vector), + Clamp(min_scalar, arg_vector, max_scalar))); ComputeAndCompareR1(&builder, {8, 8, 2, 6, 14}, {}); } XLA_TEST_F(ArrayElementwiseOpTest, ClampU32Vector) { XlaBuilder builder(TestName()); - auto min_vector = builder.ConstantR1({1, 2, 1, 2, 0, ~0u - 4}); - auto arg_vector = builder.ConstantR1({2, 10, 5, 1, 4, 10}); - auto max_vector = builder.ConstantR1({3, 5, 25, 5, 123, ~0u}); - builder.Clamp(min_vector, arg_vector, max_vector); + auto min_vector = ConstantR1(&builder, {1, 2, 1, 2, 0, ~0u - 4}); + auto arg_vector = ConstantR1(&builder, {2, 10, 5, 1, 4, 10}); + auto max_vector = ConstantR1(&builder, {3, 5, 25, 5, 123, ~0u}); + Clamp(min_vector, arg_vector, max_vector); ComputeAndCompareR1(&builder, {2, 5, 5, 2, 4, ~0u - 4}, {}); } XLA_TEST_F(ArrayElementwiseOpTest, ClampU32ScalarVector) { XlaBuilder builder(TestName()); - auto min_scalar = builder.ConstantR0(0); - auto min_vector = builder.ConstantR1({1, 0, 1, 2, 0}); - auto arg_vector = builder.ConstantR1({2, 10, 0, 1, 4}); - auto max_scalar = builder.ConstantR0(3); - auto max_vector = builder.ConstantR1({3, 1, 25, 5, 123}); + auto min_scalar = ConstantR0(&builder, 0); + auto min_vector = ConstantR1(&builder, {1, 0, 1, 2, 0}); + auto arg_vector = ConstantR1(&builder, {2, 10, 0, 1, 4}); + auto max_scalar = ConstantR0(&builder, 3); + auto max_vector = ConstantR1(&builder, {3, 1, 25, 5, 123}); // Perform clamp with broadcasted scalar and vector. - builder.Add(builder.Add(builder.Clamp(min_vector, arg_vector, max_scalar), - builder.Clamp(min_scalar, arg_vector, max_vector)), - builder.Add(builder.Clamp(min_vector, arg_vector, max_vector), - builder.Clamp(min_scalar, arg_vector, max_scalar))); + Add(Add(Clamp(min_vector, arg_vector, max_scalar), + Clamp(min_scalar, arg_vector, max_vector)), + Add(Clamp(min_vector, arg_vector, max_vector), + Clamp(min_scalar, arg_vector, max_scalar))); ComputeAndCompareR1(&builder, {8, 8, 2, 6, 14}, {}); } @@ -2090,18 +2101,18 @@ XLA_TEST_F(ArrayElementwiseOpTest, AddTwoParametersF32s) { XlaBuilder builder(TestName()); std::unique_ptr param0_literal = - Literal::CreateR1({1.1f, 2.2f, 3.3f, 5.5f}); + LiteralUtil::CreateR1({1.1f, 2.2f, 3.3f, 5.5f}); std::unique_ptr param0_data = client_->TransferToServer(*param0_literal).ConsumeValueOrDie(); std::unique_ptr param1_literal = - Literal::CreateR1({7.2f, 2.3f, 3.4f, 5.6f}); + LiteralUtil::CreateR1({7.2f, 2.3f, 3.4f, 5.6f}); std::unique_ptr param1_data = client_->TransferToServer(*param1_literal).ConsumeValueOrDie(); - auto p0 = builder.Parameter(0, param0_literal->shape(), "param0"); - auto p1 = builder.Parameter(1, param1_literal->shape(), "param1"); - builder.Add(p0, p1); + auto p0 = Parameter(&builder, 0, param0_literal->shape(), "param0"); + auto p1 = Parameter(&builder, 1, param1_literal->shape(), "param1"); + Add(p0, p1); ComputeAndCompareR1(&builder, {8.3f, 4.5f, 6.7f, 11.1f}, {param0_data.get(), param1_data.get()}, @@ -2112,18 +2123,18 @@ XLA_TEST_F(ArrayElementwiseOpTest, AddTwoParametersZeroElementF32s) { XlaBuilder builder(TestName()); std::unique_ptr param0_literal = - Literal::CreateR3FromArray3D(Array3D(0, 7, 0)); + LiteralUtil::CreateR3FromArray3D(Array3D(0, 7, 0)); std::unique_ptr param0_data = client_->TransferToServer(*param0_literal).ConsumeValueOrDie(); std::unique_ptr param1_literal = - Literal::CreateR3FromArray3D(Array3D(0, 7, 0)); + LiteralUtil::CreateR3FromArray3D(Array3D(0, 7, 0)); std::unique_ptr param1_data = client_->TransferToServer(*param1_literal).ConsumeValueOrDie(); - auto p0 = builder.Parameter(0, param0_literal->shape(), "param0"); - auto p1 = builder.Parameter(1, param1_literal->shape(), "param1"); - builder.Add(p0, p1); + auto p0 = Parameter(&builder, 0, param0_literal->shape(), "param0"); + auto p1 = Parameter(&builder, 1, param1_literal->shape(), "param1"); + Add(p0, p1); Array3D expected(0, 7, 0); ComputeAndCompareR3( @@ -2134,13 +2145,13 @@ XLA_TEST_F(ArrayElementwiseOpTest, AddParameterToConstantF32s) { XlaBuilder builder(TestName()); std::unique_ptr param0_literal = - Literal::CreateR1({1.1f, 2.2f, 3.3f, 5.5f}); + LiteralUtil::CreateR1({1.1f, 2.2f, 3.3f, 5.5f}); std::unique_ptr param0_data = client_->TransferToServer(*param0_literal).ConsumeValueOrDie(); - auto a = builder.ConstantR1({1.1f, 2.2f, 3.3f, 4.4f}); - auto p = builder.Parameter(0, param0_literal->shape(), "param0"); - builder.Add(a, p); + auto a = ConstantR1(&builder, {1.1f, 2.2f, 3.3f, 4.4f}); + auto p = Parameter(&builder, 0, param0_literal->shape(), "param0"); + Add(a, p); ComputeAndCompareR1(&builder, {2.2f, 4.4f, 6.6f, 9.9f}, {param0_data.get()}, error_spec_); @@ -2148,8 +2159,8 @@ XLA_TEST_F(ArrayElementwiseOpTest, AddParameterToConstantF32s) { XLA_TEST_F(ArrayElementwiseOpTest, CosF32s) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({3.14159f, 0.0f, 1.570796f, -0.78539f}); - builder.Cos(a); + auto a = ConstantR1(&builder, {3.14159f, 0.0f, 1.570796f, -0.78539f}); + Cos(a); ComputeAndCompareR1(&builder, {-1.0f, 1.0f, 0.0f, 0.707107f}, {}, error_spec_); @@ -2157,8 +2168,8 @@ XLA_TEST_F(ArrayElementwiseOpTest, CosF32s) { XLA_TEST_F(ArrayElementwiseOpTest, SinF32s) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({3.14159f, 0.0f, 1.570796f, -0.78539f}); - builder.Sin(a); + auto a = ConstantR1(&builder, {3.14159f, 0.0f, 1.570796f, -0.78539f}); + Sin(a); ComputeAndCompareR1(&builder, {0.0f, 0.0f, 1.0f, -0.707107f}, {}, error_spec_); @@ -2166,9 +2177,9 @@ XLA_TEST_F(ArrayElementwiseOpTest, SinF32s) { XLA_TEST_F(ArrayElementwiseOpTest, Atan2F32s) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({0.0f, 5.0f, 0.0f, -3.0f, 2.0f, -8.0f}); - auto b = builder.ConstantR1({6.0f, 0.0f, -4.0f, 0.0f, 2.0f, 8.0f}); - builder.Atan2(a, b); + auto a = ConstantR1(&builder, {0.0f, 5.0f, 0.0f, -3.0f, 2.0f, -8.0f}); + auto b = ConstantR1(&builder, {6.0f, 0.0f, -4.0f, 0.0f, 2.0f, 8.0f}); + Atan2(a, b); ComputeAndCompareR1( &builder, @@ -2178,8 +2189,8 @@ XLA_TEST_F(ArrayElementwiseOpTest, Atan2F32s) { XLA_TEST_F(ArrayElementwiseOpTest, TanhF32s) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({-2.5f, 3.14f, 2.25f}); - builder.Tanh(a); + auto a = ConstantR1(&builder, {-2.5f, 3.14f, 2.25f}); + Tanh(a); ComputeAndCompareR1(&builder, {-0.986614f, 0.996260f, 0.978026}, {}, error_spec_); @@ -2190,7 +2201,7 @@ XLA_TEST_F(ArrayElementwiseOpTest, TanhF32sVector) { // the input tensor is large enough to exercise the vectorized tanh // implementation on XLA CPU. XlaBuilder builder(TestName()); - auto input_literal = Literal::CreateR1( + auto input_literal = LiteralUtil::CreateR1( {1.02, -0.32, 0.85, 0.90, 1.23, -0.91, -0.49, 0.80, -0.67, 0.16, -0.07, 0.39, -0.41, 0.04, 1.36, 1.25, 0.41, 0.65, -1.08, 0.32, -1.45, -0.77, -1.09, 0.91, -1.03, -0.30, -1.11, -1.17, 1.50, -0.85, @@ -2201,8 +2212,8 @@ XLA_TEST_F(ArrayElementwiseOpTest, TanhF32sVector) { TF_ASSERT_OK_AND_ASSIGN(auto input_data, client_->TransferToServer(*input_literal)); - auto input = builder.Parameter(0, input_literal->shape(), "input"); - builder.Tanh(input); + auto input = Parameter(&builder, 0, input_literal->shape(), "input"); + Tanh(input); ComputeAndCompareR1( &builder, @@ -2232,7 +2243,7 @@ XLA_TEST_F(ArrayElementwiseOpTest, ExpF32sVector) { // Just to help make sense of the scales here -- exp(89) saturates float32 and // exp(-10) is smaller than our error spec. - std::unique_ptr input_literal = Literal::CreateR1( + std::unique_ptr input_literal = LiteralUtil::CreateR1( {1.02, -0.32, 0.85, 0.9, 1.23, -0.91, -0.49, 0.8, -1.31, -1.44, -0.13, -1.31, -0.79, 1.41, 1.21, 1.05, -195.6, -194.5, -193.4, -192.3, -191.2, -190.1, -189.0, -187.9, -19.6, -18.5, -17.4, @@ -2247,8 +2258,8 @@ XLA_TEST_F(ArrayElementwiseOpTest, ExpF32sVector) { TF_ASSERT_OK_AND_ASSIGN(std::unique_ptr input_data, client_->TransferToServer(*input_literal)); - auto input = builder.Parameter(0, input_literal->shape(), "input"); - builder.Exp(input); + auto input = Parameter(&builder, 0, input_literal->shape(), "input"); + Exp(input); std::vector expected_result; int64 input_size = input_literal->shape().dimensions(0); @@ -2266,7 +2277,7 @@ XLA_TEST_F(ArrayElementwiseOpTest, LogF32sVector) { // implementation on XLA CPU. XlaBuilder builder(TestName()); - std::unique_ptr input_literal = Literal::CreateR1( + std::unique_ptr input_literal = LiteralUtil::CreateR1( {-1.29, -1.41, -1.25, -13.5, -11.7, -17.9, -198, -167, 1.29, 1.41, 1.25, 13.5, 11.7, 17.9, 198, 167, 1.27e+03, 1.33e+03, 1.74e+03, 1.6e+04, 1.84e+04, @@ -2285,8 +2296,8 @@ XLA_TEST_F(ArrayElementwiseOpTest, LogF32sVector) { TF_ASSERT_OK_AND_ASSIGN(std::unique_ptr input_data, client_->TransferToServer(*input_literal)); - auto input = builder.Parameter(0, input_literal->shape(), "input"); - builder.Log(input); + auto input = Parameter(&builder, 0, input_literal->shape(), "input"); + Log(input); std::vector expected_result; int64 input_size = input_literal->shape().dimensions(0); @@ -2301,9 +2312,9 @@ XLA_TEST_F(ArrayElementwiseOpTest, LogF32sVector) { XLA_TEST_F(ArrayElementwiseOpTest, ClzU32s) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1( - {0, 1, 0x10, 0x10000, 0x700000, 0x12345678, 0xF2345678}); - builder.Clz(a); + auto a = ConstantR1( + &builder, {0, 1, 0x10, 0x10000, 0x700000, 0x12345678, 0xF2345678}); + Clz(a); ComputeAndCompareR1(&builder, {32, 31, 27, 15, 9, 3, 0}, {}); } @@ -2311,8 +2322,8 @@ XLA_TEST_F(ArrayElementwiseOpTest, ClzU32s) { XLA_TEST_F(ArrayElementwiseOpTest, ClzS64s) { XlaBuilder builder(TestName()); auto a = - builder.ConstantR1({0, 1, 0x80000000, 0x7FFFFFFFF2345678ul, -1}); - builder.Clz(a); + ConstantR1(&builder, {0, 1, 0x80000000, 0x7FFFFFFFF2345678ul, -1}); + Clz(a); ComputeAndCompareR1(&builder, {64, 63, 32, 1, 0}, {}); } @@ -2324,12 +2335,12 @@ XLA_TEST_F(ArrayElementwiseOpTest, AddChainFoldLeft) { // c---------------------/ XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({1.1f, 2.2f, 3.3f, 4.4f}); - auto b = builder.ConstantR1({2.1f, 3.2f, 4.3f, 5.4f}); - auto c = builder.ConstantR1({-3.3f, -15.5f, -7.7f, -29.9f}); + auto a = ConstantR1(&builder, {1.1f, 2.2f, 3.3f, 4.4f}); + auto b = ConstantR1(&builder, {2.1f, 3.2f, 4.3f, 5.4f}); + auto c = ConstantR1(&builder, {-3.3f, -15.5f, -7.7f, -29.9f}); - auto add = builder.Add(a, b); - builder.Add(add, c); + auto add = Add(a, b); + Add(add, c); ComputeAndCompareR1(&builder, {-0.1f, -10.1f, -0.1f, -20.1f}, {}, error_spec_); @@ -2342,12 +2353,12 @@ XLA_TEST_F(ArrayElementwiseOpTest, AddChainFoldRight) { // a---------------------/ XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({91.1f, 2.2f, 3.3f, 4.4f}); - auto b = builder.ConstantR1({2.1f, 3.2f, 4.3f, 5.4f}); - auto c = builder.ConstantR1({-3.3f, -15.5f, -7.7f, -29.9f}); + auto a = ConstantR1(&builder, {91.1f, 2.2f, 3.3f, 4.4f}); + auto b = ConstantR1(&builder, {2.1f, 3.2f, 4.3f, 5.4f}); + auto c = ConstantR1(&builder, {-3.3f, -15.5f, -7.7f, -29.9f}); - auto add = builder.Add(b, c); - builder.Add(a, add); + auto add = Add(b, c); + Add(a, add); ComputeAndCompareR1(&builder, {89.9f, -10.1f, -0.1f, -20.1f}, {}, error_spec_); @@ -2359,12 +2370,12 @@ XLA_TEST_F(ArrayElementwiseOpTest, AddWithNeg) { // b ----- (neg) ----/ XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({91.1f, 2.2f, 3.3f, 4.4f}); - auto b = builder.ConstantR1({2.1f, 3.2f, 4.3f, 5.4f}); + auto a = ConstantR1(&builder, {91.1f, 2.2f, 3.3f, 4.4f}); + auto b = ConstantR1(&builder, {2.1f, 3.2f, 4.3f, 5.4f}); - auto neg_a = builder.Neg(a); - auto neg_b = builder.Neg(b); - builder.Add(neg_a, neg_b); + auto neg_a = Neg(a); + auto neg_b = Neg(b); + Add(neg_a, neg_b); ComputeAndCompareR1(&builder, {-93.2f, -5.4f, -7.6f, -9.8f}, {}, error_spec_); @@ -2380,14 +2391,14 @@ XLA_TEST_F(ArrayElementwiseOpTest, AddChainTwoSide) { // d -----/ XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({91.1f, 2.2f, 3.3f, 4.4f}); - auto b = builder.ConstantR1({2.1f, 3.2f, 4.3f, 5.4f}); - auto c = builder.ConstantR1({-3.3f, -15.5f, -7.7f, -29.9f}); - auto d = builder.ConstantR1({-19.0f, 10.0f, -40.0f, 20.2f}); + auto a = ConstantR1(&builder, {91.1f, 2.2f, 3.3f, 4.4f}); + auto b = ConstantR1(&builder, {2.1f, 3.2f, 4.3f, 5.4f}); + auto c = ConstantR1(&builder, {-3.3f, -15.5f, -7.7f, -29.9f}); + auto d = ConstantR1(&builder, {-19.0f, 10.0f, -40.0f, 20.2f}); - auto add_ab = builder.Add(a, b); - auto add_cd = builder.Add(c, d); - builder.Add(add_ab, add_cd); + auto add_ab = Add(a, b); + auto add_cd = Add(c, d); + Add(add_ab, add_cd); ComputeAndCompareR1(&builder, {70.9f, -0.1f, -40.1f, 0.1f}, {}, error_spec_); @@ -2395,11 +2406,11 @@ XLA_TEST_F(ArrayElementwiseOpTest, AddChainTwoSide) { XLA_TEST_F(ArrayElementwiseOpTest, 2DBinaryOpF32s) { XlaBuilder builder(TestName()); - auto a = - builder.ConstantR2({{-2.5f, 3.14f, 1.0f}, {2.25f, -10.0f, 3.33f}}); - auto b = - builder.ConstantR2({{-1.5f, 8.14f, 42.0}, {-1.0f, -4.0f, 5.55f}}); - builder.Add(a, b); + auto a = ConstantR2(&builder, + {{-2.5f, 3.14f, 1.0f}, {2.25f, -10.0f, 3.33f}}); + auto b = ConstantR2(&builder, + {{-1.5f, 8.14f, 42.0}, {-1.0f, -4.0f, 5.55f}}); + Add(a, b); Array2D expected_array( {{-4.0f, 11.28f, 43.0f}, {1.25f, -14.0f, 8.88f}}); @@ -2409,10 +2420,10 @@ XLA_TEST_F(ArrayElementwiseOpTest, 2DBinaryOpF32s) { XLA_TEST_F(ArrayElementwiseOpTest, ScalarPlus2DF32) { // Add a scalar + matrix. XlaBuilder builder(TestName()); - auto a = - builder.ConstantR2({{-2.5f, 3.14f, 1.0f}, {2.25f, -10.0f, 3.33f}}); - auto scalar = builder.ConstantR0(3.0f); - builder.Add(scalar, a); + auto a = ConstantR2(&builder, + {{-2.5f, 3.14f, 1.0f}, {2.25f, -10.0f, 3.33f}}); + auto scalar = ConstantR0(&builder, 3.0f); + Add(scalar, a); Array2D expected_array({{0.5f, 6.14f, 4.0f}, {5.25f, -7.0f, 6.33f}}); ComputeAndCompareR2(&builder, expected_array, {}, error_spec_); @@ -2421,10 +2432,10 @@ XLA_TEST_F(ArrayElementwiseOpTest, ScalarPlus2DF32) { XLA_TEST_F(ArrayElementwiseOpTest, 2DPlusScalarF32) { // Add a matrix + scalar. XlaBuilder builder(TestName()); - auto a = - builder.ConstantR2({{-2.5f, 3.14f, 1.0f}, {2.25f, -10.0f, 3.33f}}); - auto scalar = builder.ConstantR0(3.0f); - builder.Add(a, scalar); + auto a = ConstantR2(&builder, + {{-2.5f, 3.14f, 1.0f}, {2.25f, -10.0f, 3.33f}}); + auto scalar = ConstantR0(&builder, 3.0f); + Add(a, scalar); Array2D expected_array({{0.5f, 6.14f, 4.0f}, {5.25f, -7.0f, 6.33f}}); ComputeAndCompareR2(&builder, expected_array, {}, error_spec_); @@ -2434,13 +2445,13 @@ XLA_TEST_F(ArrayElementwiseOpTest, Add1DTo2DF32) { // Test simple broadcasting of a R1F32 over R2F32. The vector's size matches // only dim 0 of the matrix. XlaBuilder builder(TestName()); - auto v = builder.ConstantR1({20.0f, 40.0f, 60.0f}); + auto v = ConstantR1(&builder, {20.0f, 40.0f, 60.0f}); // clang-format off - auto m = builder.ConstantR2({ + auto m = ConstantR2(&builder, { {-2.5f, 3.14f, 1.0f}, {2.25f, -10.0f, 3.33f}}); // clang-format on - builder.Add(v, m, /*broadcast_dimensions=*/{1}); + Add(v, m, /*broadcast_dimensions=*/{1}); Array2D expected_array( {{17.5f, 43.14f, 61.0f}, {22.25f, 30.0f, 63.33f}}); ComputeAndCompareR2(&builder, expected_array, {}, error_spec_); @@ -2449,27 +2460,27 @@ XLA_TEST_F(ArrayElementwiseOpTest, Add1DTo2DF32) { XLA_TEST_F(ArrayElementwiseOpTest, Compare1DTo2DS32Eq) { // Test broadcasting in Eq comparison. XlaBuilder builder(TestName()); - auto v = builder.ConstantR1({42, 73}); - auto m = builder.ConstantR2({{42, 73}, {42, 52}}); + auto v = ConstantR1(&builder, {42, 73}); + auto m = ConstantR2(&builder, {{42, 73}, {42, 52}}); // This test exercises both possible broadcast dimensions for a vector/matrix // comparison. - auto cmp_dim_0 = builder.Eq(v, m, /*broadcast_dimensions=*/{1}); - auto cmp_dim_1 = builder.Eq(v, m, /*broadcast_dimensions=*/{0}); - auto result = builder.Tuple({cmp_dim_0, cmp_dim_1}); + auto cmp_dim_0 = Eq(v, m, /*broadcast_dimensions=*/{1}); + auto cmp_dim_1 = Eq(v, m, /*broadcast_dimensions=*/{0}); + Tuple(&builder, {cmp_dim_0, cmp_dim_1}); - auto expected = Literal::MakeTuple( - {Literal::CreateR2({{true, true}, {true, false}}).get(), - Literal::CreateR2({{true, false}, {false, false}}).get()}); + auto expected = LiteralUtil::MakeTuple( + {LiteralUtil::CreateR2({{true, true}, {true, false}}).get(), + LiteralUtil::CreateR2({{true, false}, {false, false}}).get()}); ComputeAndCompareTuple(&builder, *expected, {}, error_spec_); } XLA_TEST_F(ArrayElementwiseOpTest, Compare1DTo2DS32Ne) { // Test broadcasting in Ne comparison. XlaBuilder builder(TestName()); - auto v = builder.ConstantR1({42, 73}); - auto m = builder.ConstantR2({{42, 73}, {42, 52}}); - builder.Ne(v, m, /*broadcast_dimensions=*/{1}); + auto v = ConstantR1(&builder, {42, 73}); + auto m = ConstantR2(&builder, {{42, 73}, {42, 52}}); + Ne(v, m, /*broadcast_dimensions=*/{1}); const string expected = R"(pred[2,2] { { 00 }, @@ -2481,9 +2492,9 @@ XLA_TEST_F(ArrayElementwiseOpTest, Compare1DTo2DS32Ne) { XLA_TEST_F(ArrayElementwiseOpTest, Compare1DTo2DS32Ge) { // Test broadcasting in Ge comparison. XlaBuilder builder(TestName()); - auto v = builder.ConstantR1({1, 2, 3, 4}); - auto m = builder.ConstantR2({{1, 0, 5, 6}, {42, 52, 10, 4}}); - builder.Ge(v, m, /*broadcast_dimensions=*/{1}); + auto v = ConstantR1(&builder, {1, 2, 3, 4}); + auto m = ConstantR2(&builder, {{1, 0, 5, 6}, {42, 52, 10, 4}}); + Ge(v, m, /*broadcast_dimensions=*/{1}); const string expected = R"(pred[2,4] { { 1100 }, @@ -2495,9 +2506,9 @@ XLA_TEST_F(ArrayElementwiseOpTest, Compare1DTo2DS32Ge) { XLA_TEST_F(ArrayElementwiseOpTest, Compare1DTo2DS32Gt) { // Test broadcasting in Gt comparison. XlaBuilder builder(TestName()); - auto v = builder.ConstantR1({1, 2, 3, 4}); - auto m = builder.ConstantR2({{1, 0, 5, 6}, {42, 52, 10, 4}}); - builder.Gt(v, m, /*broadcast_dimensions=*/{1}); + auto v = ConstantR1(&builder, {1, 2, 3, 4}); + auto m = ConstantR2(&builder, {{1, 0, 5, 6}, {42, 52, 10, 4}}); + Gt(v, m, /*broadcast_dimensions=*/{1}); const string expected = R"(pred[2,4] { { 0100 }, @@ -2509,9 +2520,9 @@ XLA_TEST_F(ArrayElementwiseOpTest, Compare1DTo2DS32Gt) { XLA_TEST_F(ArrayElementwiseOpTest, Compare1DTo2DS32Le) { // Test broadcasting in Le comparison. XlaBuilder builder(TestName()); - auto v = builder.ConstantR1({1, 2, 3, 4}); - auto m = builder.ConstantR2({{1, 0, 5, 6}, {42, 52, 10, 4}}); - builder.Le(v, m, /*broadcast_dimensions=*/{1}); + auto v = ConstantR1(&builder, {1, 2, 3, 4}); + auto m = ConstantR2(&builder, {{1, 0, 5, 6}, {42, 52, 10, 4}}); + Le(v, m, /*broadcast_dimensions=*/{1}); const string expected = R"(pred[2,4] { { 1011 }, @@ -2523,9 +2534,9 @@ XLA_TEST_F(ArrayElementwiseOpTest, Compare1DTo2DS32Le) { XLA_TEST_F(ArrayElementwiseOpTest, Compare1DTo2DS32Lt) { // Test broadcasting in Lt comparison. XlaBuilder builder(TestName()); - auto v = builder.ConstantR1({1, 2, 3, 4}); - auto m = builder.ConstantR2({{1, 0, 5, 6}, {42, 52, 10, 4}}); - builder.Lt(v, m, /*broadcast_dimensions=*/{1}); + auto v = ConstantR1(&builder, {1, 2, 3, 4}); + auto m = ConstantR2(&builder, {{1, 0, 5, 6}, {42, 52, 10, 4}}); + Lt(v, m, /*broadcast_dimensions=*/{1}); const string expected = R"(pred[2,4] { { 0011 }, @@ -2538,9 +2549,10 @@ XLA_TEST_F(ArrayElementwiseOpTest, Mul2Dby1DF32) { // Test simple broadcasting of a R1F32 over R2F32 when the order of binary op // arguments is reversed. XlaBuilder builder(TestName()); - auto m = builder.ConstantR2({{1.5f, 2.5f, 3.5f}, {4.5f, 5.5f, 6.5f}}); - auto v = builder.ConstantR1({2.0f, 4.0f, 6.0f}); - builder.Mul(m, v, /*broadcast_dimensions=*/{1}); + auto m = + ConstantR2(&builder, {{1.5f, 2.5f, 3.5f}, {4.5f, 5.5f, 6.5f}}); + auto v = ConstantR1(&builder, {2.0f, 4.0f, 6.0f}); + Mul(m, v, /*broadcast_dimensions=*/{1}); Array2D expected_array({{3.0f, 10.0f, 21.0f}, {9.0f, 22.0f, 39.0f}}); ComputeAndCompareR2(&builder, expected_array, {}, error_spec_); } @@ -2551,10 +2563,10 @@ XLA_TEST_F(ArrayElementwiseOpTest, Add2DTo2DWithDegenerateDim1) { // m's shape in XLA notation is {3, 2} // md's shape in XLA notation is {3, 1} // The result has shape {3, 2}, where md is broadcast over m - auto m = - builder.ConstantR2({{-2.5f, 3.14f, 1.0f}, {2.25f, -10.0f, 3.33f}}); - auto md = builder.ConstantR2({{10.0f, 20.0f, 30.0f}}); - builder.Add(m, md); + auto m = ConstantR2(&builder, + {{-2.5f, 3.14f, 1.0f}, {2.25f, -10.0f, 3.33f}}); + auto md = ConstantR2(&builder, {{10.0f, 20.0f, 30.0f}}); + Add(m, md); Array2D expected_array( {{7.5f, 23.14f, 31.0f}, {12.25f, 10.0f, 33.33f}}); ComputeAndCompareR2(&builder, expected_array, {}, error_spec_); @@ -2566,10 +2578,10 @@ XLA_TEST_F(ArrayElementwiseOpTest, Add2DTo2DWithDegenerateDim0) { // m's shape in XLA notation is {3, 2} // md's shape in XLA notation is {1, 2} // The result has shape {3, 2}, where md is broadcast over m - auto m = - builder.ConstantR2({{-2.5f, 3.14f, 1.0f}, {2.25f, -10.0f, 3.33f}}); - auto md = builder.ConstantR2({{10.0f}, {20.0f}}); - builder.Add(m, md); + auto m = ConstantR2(&builder, + {{-2.5f, 3.14f, 1.0f}, {2.25f, -10.0f, 3.33f}}); + auto md = ConstantR2(&builder, {{10.0f}, {20.0f}}); + Add(m, md); Array2D expected_array( {{7.5f, 13.14f, 11.0f}, {22.25f, 10.0f, 23.33f}}); ComputeAndCompareR2(&builder, expected_array, {}, error_spec_); @@ -2584,9 +2596,9 @@ XLA_TEST_F(ArrayElementwiseOpTest, Add2DsWithDegenerateDimsOuterProduct) { // a's shape in XLA notation is {1, 4} // b's shape in XLA notation is {3, 1} // The result has shape {3, 4}. - auto a = builder.ConstantR2({{0.0f}, {10.0f}, {20.0f}, {30.0f}}); - auto b = builder.ConstantR2({{1.0f, 2.0f, 3.0f}}); - builder.Add(a, b); + auto a = ConstantR2(&builder, {{0.0f}, {10.0f}, {20.0f}, {30.0f}}); + auto b = ConstantR2(&builder, {{1.0f, 2.0f, 3.0f}}); + Add(a, b); Array2D expected_array({{1.0f, 2.0f, 3.0f}, {11.0f, 12.0f, 13.0f}, {21.0f, 22.0f, 23.0f}, @@ -2598,9 +2610,9 @@ XLA_TEST_F(ArrayElementwiseOpTest, Add1DTo2DF32TwoWaysOver1) { // Add together a (2,2) array and a (2) array, using dimension 0 for // broadcasting (though there are two ways to broadcast these shapes). XlaBuilder builder(TestName()); - auto v = builder.ConstantR1({20.0f, 40.0f}); - auto m = builder.ConstantR2({{10.0f, 50.0f}, {77.0f, 88.0f}}); - builder.Add(v, m, /*broadcast_dimensions=*/{1}); + auto v = ConstantR1(&builder, {20.0f, 40.0f}); + auto m = ConstantR2(&builder, {{10.0f, 50.0f}, {77.0f, 88.0f}}); + Add(v, m, /*broadcast_dimensions=*/{1}); Array2D expected_array({{30.0f, 90.0f}, {97.0f, 128.0f}}); ComputeAndCompareR2(&builder, expected_array, {}, error_spec_); } @@ -2609,9 +2621,9 @@ XLA_TEST_F(ArrayElementwiseOpTest, Add1DTo2DF32TwoWaysOver0) { // Add together a (2,2) array and a (2) array, using dimension 1 for // broadcasting (though there are two ways to broadcast these shapes). XlaBuilder builder(TestName()); - auto v = builder.ConstantR1({20.0f, 40.0f}); - auto m = builder.ConstantR2({{10.0f, 50.0f}, {77.0f, 88.0f}}); - builder.Add(v, m, /*broadcast_dimensions=*/{0}); + auto v = ConstantR1(&builder, {20.0f, 40.0f}); + auto m = ConstantR2(&builder, {{10.0f, 50.0f}, {77.0f, 88.0f}}); + Add(v, m, /*broadcast_dimensions=*/{0}); Array2D expected_array({{30.0f, 70.0f}, {117.0f, 128.0f}}); ComputeAndCompareR2(&builder, expected_array, {}, error_spec_); } @@ -2621,12 +2633,12 @@ XLA_TEST_F(ArrayElementwiseOpTest, 3DBinaryOpF32s) { XlaBuilder builder(TestName()); Array3D a_3d({{{1.0f, 2.0f}, {3.0f, 4.0f}, {5.0f, 6.0f}}, {{7.0f, 8.0f}, {9.0f, 10.0f}, {11.0f, 12.0f}}}); - auto a = builder.ConstantR3FromArray3D(a_3d); + auto a = ConstantR3FromArray3D(&builder, a_3d); Array3D b_3d({{{2.0f, 4.0f}, {6.0f, 8.0f}, {10.0f, 12.0f}}, {{14.0f, 16.0f}, {18.0f, 20.0f}, {22.0f, 24.0f}}}); - auto b = builder.ConstantR3FromArray3D(b_3d); - builder.Add(a, b); + auto b = ConstantR3FromArray3D(&builder, b_3d); + Add(a, b); Array3D expected_3d( {{{3.0f, 6.0f}, {9.0f, 12.0f}, {15.0f, 18.0f}}, @@ -2648,9 +2660,9 @@ XLA_TEST_F(ArrayElementwiseOpTest, Add1DTo3DTwoWaysOver2) { {11.0f, 12.0f}}, }); // clang-format on - auto a = builder.ConstantR3FromArray3D(a_3d); - auto v = builder.ConstantR1({10.0f, 20.0f}); - builder.Add(a, v, /*broadcast_dimensions=*/{2}); + auto a = ConstantR3FromArray3D(&builder, a_3d); + auto v = ConstantR1(&builder, {10.0f, 20.0f}); + Add(a, v, /*broadcast_dimensions=*/{2}); Array3D expected_3d( {{{11.0f, 22.0f}, {13.0f, 24.0f}, {15.0f, 26.0f}}, @@ -2672,9 +2684,9 @@ XLA_TEST_F(ArrayElementwiseOpTest, Add1DTo3DTwoWaysOver0) { {11.0f, 12.0f}}, }); // clang-format on - auto a = builder.ConstantR3FromArray3D(a_3d); - auto v = builder.ConstantR1({10.0f, 20.0f}); - builder.Add(a, v, /*broadcast_dimensions=*/{0}); + auto a = ConstantR3FromArray3D(&builder, a_3d); + auto v = ConstantR1(&builder, {10.0f, 20.0f}); + Add(a, v, /*broadcast_dimensions=*/{0}); // clang-format off Array3D expected_3d({ @@ -2702,12 +2714,12 @@ XLA_TEST_F(ArrayElementwiseOpTest, Add2DTo3D) { {9.0f, 10.0f}, {11.0f, 12.0f}}, }); - auto a = builder.ConstantR3FromArray3D(a_3d); - auto m = builder.ConstantR2({ + auto a = ConstantR3FromArray3D(&builder, a_3d); + auto m = ConstantR2(&builder, { {10.0f, 20.0f, 30.0f}, {40.0f, 50.0f, 60.0f}, }); - builder.Add(a, m, /*broadcast_dimensions=*/{0, 1}); + Add(a, m, /*broadcast_dimensions=*/{0, 1}); Array3D expected_3d({ {{11.0f, 12.0f}, @@ -2727,12 +2739,12 @@ XLA_TEST_F(ArrayElementwiseOpTest, CompareGtR3F32sWithDegenerateDim2) { XlaBuilder builder(TestName()); Array3D a_3d({{{1.0f, 2.0f}, {3.0f, 4.0f}, {5.0f, 6.0f}}, {{7.0f, 8.0f}, {9.0f, 10.0f}, {11.0f, 12.0f}}}); - auto a = builder.ConstantR3FromArray3D(a_3d); + auto a = ConstantR3FromArray3D(&builder, a_3d); Array3D b_3d({{{7.0f, 1.0f}, {3.0f, 10.0f}, {15.0f, 6.0f}}}); - auto b = builder.ConstantR3FromArray3D(b_3d); + auto b = ConstantR3FromArray3D(&builder, b_3d); - builder.Gt(a, b); + Gt(a, b); Array3D expected_3d( {{{0, 1}, {0, 0}, {0, 0}}, {{0, 1}, {1, 0}, {0, 1}}}); @@ -2767,9 +2779,9 @@ XLA_TEST_F(ArrayElementwiseOpTest, 4DBinaryOpF32s) { } } - auto a = builder.ConstantR4FromArray4D(*operand_a_4d); - auto b = builder.ConstantR4FromArray4D(*operand_b_4d); - builder.Add(a, b); + auto a = ConstantR4FromArray4D(&builder, *operand_a_4d); + auto b = ConstantR4FromArray4D(&builder, *operand_b_4d); + Add(a, b); ComputeAndCompareR4(&builder, *expected_4d, {}, error_spec_); } @@ -2795,9 +2807,9 @@ XLA_TEST_F(ArrayElementwiseOpTest, R4PlusR1InDim1) { } } - auto a = builder.ConstantR4FromArray4D(*operand_a_4d); - auto b = builder.ConstantR1(operand_b_1d); - builder.Add(a, b, {1}); + auto a = ConstantR4FromArray4D(&builder, *operand_a_4d); + auto b = ConstantR1(&builder, operand_b_1d); + Add(a, b, {1}); ComputeAndCompareR4(&builder, *expected_4d, {}, error_spec_); } @@ -2813,11 +2825,12 @@ XLA_TEST_F(ArrayElementwiseOpTest, R4_16x16x2x2_Plus_R1_16) { std::iota(r1.begin(), r1.end(), 1.0); XlaBuilder builder(TestName()); - std::unique_ptr a_literal = Literal::CreateR4FromArray4DWithLayout( - r4, LayoutUtil::MakeLayout({0, 1, 2, 3})); - auto a = builder.ConstantLiteral(*a_literal); - auto b = builder.ConstantR1(r1); - builder.Add(a, b, {1}); + std::unique_ptr a_literal = + LiteralUtil::CreateR4FromArray4DWithLayout( + r4, LayoutUtil::MakeLayout({0, 1, 2, 3})); + auto a = ConstantLiteral(&builder, *a_literal); + auto b = ConstantR1(&builder, r1); + Add(a, b, {1}); for (int i0 = 0; i0 < d0; ++i0) { for (int i1 = 0; i1 < d1; ++i1) { @@ -2835,8 +2848,8 @@ XLA_TEST_F(ArrayElementwiseOpTest, R4_16x16x2x2_Plus_R1_16) { XLA_TEST_F(ArrayElementwiseOpTest, CannotAddOpaques) { XlaBuilder builder(TestName()); auto shape = ShapeUtil::MakeOpaqueShape(); - auto x = builder.Parameter(0, shape, "x"); - builder.Add(x, x); + auto x = Parameter(&builder, 0, shape, "x"); + Add(x, x); auto computation_status = builder.Build(); ASSERT_FALSE(computation_status.ok()); EXPECT_THAT(computation_status.status().ToString(), @@ -2846,11 +2859,11 @@ XLA_TEST_F(ArrayElementwiseOpTest, CannotAddOpaques) { XLA_TEST_F(ArrayElementwiseOpTest, IdentityBroadcastOfSameRankIsAllowed) { XlaBuilder builder(TestName()); - auto a = - builder.ConstantR2({{-2.5f, 3.14f, 1.0f}, {2.25f, -10.0f, 3.33f}}); - auto b = - builder.ConstantR2({{-1.5f, 8.14f, 42.0}, {-1.0f, -4.0f, 5.55f}}); - builder.Add(a, b, /*broadcast_dimensions=*/{0, 1}); + auto a = ConstantR2(&builder, + {{-2.5f, 3.14f, 1.0f}, {2.25f, -10.0f, 3.33f}}); + auto b = ConstantR2(&builder, + {{-1.5f, 8.14f, 42.0}, {-1.0f, -4.0f, 5.55f}}); + Add(a, b, /*broadcast_dimensions=*/{0, 1}); Array2D expected_array( {{-4.0f, 11.28f, 43.0f}, {1.25f, -14.0f, 8.88f}}); @@ -2859,11 +2872,11 @@ XLA_TEST_F(ArrayElementwiseOpTest, IdentityBroadcastOfSameRankIsAllowed) { XLA_TEST_F(ArrayElementwiseOpTest, NonIdentityBroadcastOfSameRankIsDisallowed) { XlaBuilder builder(TestName()); - auto a = - builder.ConstantR2({{-2.5f, 3.14f, 1.0f}, {2.25f, -10.0f, 3.33f}}); - auto b = - builder.ConstantR2({{-1.5f, 8.14f, 42.0}, {-1.0f, -4.0f, 5.55f}}); - builder.Add(a, b, /*broadcast_dimensions=*/{1, 0}); + auto a = ConstantR2(&builder, + {{-2.5f, 3.14f, 1.0f}, {2.25f, -10.0f, 3.33f}}); + auto b = ConstantR2(&builder, + {{-1.5f, 8.14f, 42.0}, {-1.0f, -4.0f, 5.55f}}); + Add(a, b, /*broadcast_dimensions=*/{1, 0}); auto computation_status = builder.Build(); ASSERT_FALSE(computation_status.ok()); @@ -2875,15 +2888,15 @@ XLA_TEST_F(ArrayElementwiseOpTest, NonIdentityBroadcastOfSameRankIsDisallowed) { // broadcast. XLA_TEST_F(ArrayElementwiseOpTest, ImplictBroadcastInFusedExpressions) { XlaBuilder builder(TestName()); - auto x_literal = Literal::CreateR1({1, 2, 3}); - auto y_literal = Literal::CreateR1({4, 5}); + auto x_literal = LiteralUtil::CreateR1({1, 2, 3}); + auto y_literal = LiteralUtil::CreateR1({4, 5}); auto x_data = client_->TransferToServer(*x_literal).ConsumeValueOrDie(); auto y_data = client_->TransferToServer(*y_literal).ConsumeValueOrDie(); - auto x = builder.Parameter(0, x_literal->shape(), "x"); - auto y = builder.Parameter(1, y_literal->shape(), "y"); - auto slice = builder.Slice(x, {1}, {2}, {1}); - builder.Sub(slice, y); + auto x = Parameter(&builder, 0, x_literal->shape(), "x"); + auto y = Parameter(&builder, 1, y_literal->shape(), "y"); + auto slice = Slice(x, {1}, {2}, {1}); + Sub(slice, y); ComputeAndCompareR1(&builder, {-2, -3}, {x_data.get(), y_data.get()}, error_spec_); diff --git a/tensorflow/compiler/xla/tests/axpy_simple_test.cc b/tensorflow/compiler/xla/tests/axpy_simple_test.cc index fcd9ff55e393f64476ddd4754e0fa74427f1cb51..8d15b7841bc7298cd6865d8689cc496c0459e4b9 100644 --- a/tensorflow/compiler/xla/tests/axpy_simple_test.cc +++ b/tensorflow/compiler/xla/tests/axpy_simple_test.cc @@ -29,10 +29,10 @@ class AxpySimpleTest : public ClientLibraryTestBase {}; TEST_F(AxpySimpleTest, AxTenValues) { XlaBuilder builder("ax_10"); - auto alpha = builder.ConstantR0(3.1415926535); - auto x = builder.ConstantR1( - {-1.0, 1.0, 2.0, -2.0, -3.0, 3.0, 4.0, -4.0, -5.0, 5.0}); - builder.Mul(alpha, x); + auto alpha = ConstantR0(&builder, 3.1415926535); + auto x = ConstantR1( + &builder, {-1.0, 1.0, 2.0, -2.0, -3.0, 3.0, 4.0, -4.0, -5.0, 5.0}); + Mul(alpha, x); std::vector expected = { -3.14159265, 3.14159265, 6.28318531, -6.28318531, -9.42477796, @@ -42,11 +42,11 @@ TEST_F(AxpySimpleTest, AxTenValues) { XLA_TEST_F(AxpySimpleTest, AxpyZeroValues) { XlaBuilder builder("axpy_10"); - auto alpha = builder.ConstantR0(3.1415926535); - auto x = builder.ConstantR1({}); - auto y = builder.ConstantR1({}); - auto ax = builder.Mul(alpha, x); - builder.Add(ax, y); + auto alpha = ConstantR0(&builder, 3.1415926535); + auto x = ConstantR1(&builder, {}); + auto y = ConstantR1(&builder, {}); + auto ax = Mul(alpha, x); + Add(ax, y); std::vector expected = {}; ComputeAndCompareR1(&builder, expected, {}, ErrorSpec(0.0001)); @@ -54,13 +54,13 @@ XLA_TEST_F(AxpySimpleTest, AxpyZeroValues) { TEST_F(AxpySimpleTest, AxpyTenValues) { XlaBuilder builder("axpy_10"); - auto alpha = builder.ConstantR0(3.1415926535); - auto x = builder.ConstantR1( - {-1.0, 1.0, 2.0, -2.0, -3.0, 3.0, 4.0, -4.0, -5.0, 5.0}); - auto y = builder.ConstantR1( - {5.0, -5.0, -4.0, 4.0, 3.0, -3.0, -2.0, 2.0, 1.0, -1.0}); - auto ax = builder.Mul(alpha, x); - builder.Add(ax, y); + auto alpha = ConstantR0(&builder, 3.1415926535); + auto x = ConstantR1( + &builder, {-1.0, 1.0, 2.0, -2.0, -3.0, 3.0, 4.0, -4.0, -5.0, 5.0}); + auto y = ConstantR1( + &builder, {5.0, -5.0, -4.0, 4.0, 3.0, -3.0, -2.0, 2.0, 1.0, -1.0}); + auto ax = Mul(alpha, x); + Add(ax, y); TF_ASSERT_OK_AND_ASSIGN(ProgramShape shape, builder.GetProgramShape()); diff --git a/tensorflow/compiler/xla/tests/bad_rng_shape_validation_test.cc b/tensorflow/compiler/xla/tests/bad_rng_shape_validation_test.cc index 22c3394e6f34bd018ffaaaa4d9d68339673c3764..8c227df7f04e79ccc332062d0889d282c0f5e40f 100644 --- a/tensorflow/compiler/xla/tests/bad_rng_shape_validation_test.cc +++ b/tensorflow/compiler/xla/tests/bad_rng_shape_validation_test.cc @@ -35,10 +35,10 @@ class BadRngShapeValidationTest : public ClientLibraryTestBase {}; TEST_F(BadRngShapeValidationTest, DefaultConstructedShapeCreatesError) { XlaBuilder builder(TestName()); - auto zero = builder.ConstantR0(0.0); - auto one = builder.ConstantR0(1.0); + auto zero = ConstantR0(&builder, 0.0); + auto one = ConstantR0(&builder, 1.0); Shape default_constructed; - builder.RngUniform(zero, one, default_constructed); + RngUniform(zero, one, default_constructed); StatusOr computation = builder.Build(); EXPECT_FALSE(computation.ok()); @@ -49,13 +49,13 @@ TEST_F(BadRngShapeValidationTest, DefaultConstructedShapeCreatesError) { TEST_F(BadRngShapeValidationTest, ShapeWithoutLayoutIsOk) { XlaBuilder builder(TestName()); - auto zero = builder.ConstantR0(0.0); - auto one = builder.ConstantR0(1.0); + auto zero = ConstantR0(&builder, 0.0); + auto one = ConstantR0(&builder, 1.0); Shape sans_layout; sans_layout.set_element_type(F32); sans_layout.add_dimensions(1); - builder.RngUniform(zero, one, sans_layout); + RngUniform(zero, one, sans_layout); StatusOr computation = builder.Build(); ASSERT_TRUE(computation.ok()); diff --git a/tensorflow/compiler/xla/tests/batch_normalization_test.cc b/tensorflow/compiler/xla/tests/batch_normalization_test.cc index f3dac75a44b948c4b45b80b93e7462073010979e..6a024798f9e3faa5164b4dce6f43e517f6ab8eca 100644 --- a/tensorflow/compiler/xla/tests/batch_normalization_test.cc +++ b/tensorflow/compiler/xla/tests/batch_normalization_test.cc @@ -20,10 +20,11 @@ limitations under the License. #include "tensorflow/compiler/xla/array2d.h" #include "tensorflow/compiler/xla/array4d.h" #include "tensorflow/compiler/xla/client/lib/arithmetic.h" +#include "tensorflow/compiler/xla/client/lib/math.h" #include "tensorflow/compiler/xla/client/local_client.h" #include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/compiler/xla/client/xla_client/xla_computation.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/reference_util.h" #include "tensorflow/compiler/xla/service/hlo_computation.h" #include "tensorflow/compiler/xla/service/hlo_instruction.h" @@ -62,7 +63,7 @@ class BatchNormalizationTest {5.0f, 4.4f}, // p2 }); input_array_.FillWithPZ(pz); - input_literal_ = std::move(*Literal::CreateR4FromArray4D(input_array_)); + input_literal_ = std::move(*LiteralUtil::CreateR4FromArray4D(input_array_)); CHECK_EQ(kSamples, input_array_.planes()); CHECK_EQ(kZ, input_array_.depth()); CHECK_EQ(kY, input_array_.height()); @@ -101,9 +102,9 @@ INSTANTIATE_TEST_CASE_P(BatchNormalizationTestInstance, BatchNormalizationTest, XLA_TEST_P(BatchNormalizationTest, SubtractInZ) { XlaBuilder builder("subtract_in_z_one_sample"); - auto x = builder.ConstantLiteral(input_literal_); - auto y = builder.ConstantR1({3.14, 4.25}); - builder.Sub(x, y, /*broadcast_dimensions=*/{1}); + auto x = ConstantLiteral(&builder, input_literal_); + auto y = ConstantR1(&builder, {3.14, 4.25}); + Sub(x, y, /*broadcast_dimensions=*/{1}); Array4D expected(kSamples, kZ, kY, kX); Array2D pz({ @@ -117,8 +118,8 @@ XLA_TEST_P(BatchNormalizationTest, SubtractInZ) { XLA_TEST_P(BatchNormalizationTest, SquareTesseractElementwise) { XlaBuilder builder("square_tesseract_elementwise"); - auto x = builder.ConstantLiteral(input_literal_); - builder.SquareF32(x); + auto x = ConstantLiteral(&builder, input_literal_); + Square(x); using tensorflow::MathUtil; @@ -134,11 +135,10 @@ XLA_TEST_P(BatchNormalizationTest, SquareTesseractElementwise) { XLA_TEST_P(BatchNormalizationTest, SumToZ) { XlaBuilder builder("sum_to_z"); - auto input_activations = builder.ConstantLiteral(input_literal_); + auto input_activations = ConstantLiteral(&builder, input_literal_); XlaComputation add = CreateScalarAddComputation(F32, &builder); // Reduce all but the Z dimension. - builder.Reduce(input_activations, builder.ConstantR0(0.0f), add, - {0, 2, 3}); + Reduce(input_activations, ConstantR0(&builder, 0.0f), add, {0, 2, 3}); std::vector expected = {6, 12.6}; ComputeAndCompareR1(&builder, expected, {}, error_spec_); @@ -146,13 +146,13 @@ XLA_TEST_P(BatchNormalizationTest, SumToZ) { XLA_TEST_P(BatchNormalizationTest, SquareAndReduce) { XlaBuilder builder("square_and_reduce"); - auto input_activations = builder.ConstantLiteral(input_literal_); - auto set_means = builder.ConstantR1({2.f, 4.2f}); - auto activation_deviations = builder.Sub(input_activations, set_means, - /*broadcast_dimensions=*/{1}); + auto input_activations = ConstantLiteral(&builder, input_literal_); + auto set_means = ConstantR1(&builder, {2.f, 4.2f}); + auto activation_deviations = Sub(input_activations, set_means, + /*broadcast_dimensions=*/{1}); XlaComputation add = CreateScalarAddComputation(F32, &builder); - auto dev_squares = builder.SquareF32(activation_deviations); - builder.Reduce(dev_squares, builder.ConstantR0(0.0f), add, {0, 2, 3}); + auto dev_squares = Square(activation_deviations); + Reduce(dev_squares, ConstantR0(&builder, 0.0f), add, {0, 2, 3}); std::vector expected = {18, 0.06}; ComputeAndCompareR1(&builder, expected, {}, error_spec_); @@ -160,8 +160,8 @@ XLA_TEST_P(BatchNormalizationTest, SquareAndReduce) { XLA_TEST_P(BatchNormalizationTest, VarianceToStddev) { XlaBuilder builder("variance_to_stddev"); - auto variance = builder.ConstantR1({6.f, .02f}); - builder.SqrtF32(variance); + auto variance = ConstantR1(&builder, {6.f, .02f}); + Sqrt(variance); std::vector expected = {2.44948974f, 0.14142136f}; ComputeAndCompareR1(&builder, expected, {}, error_spec_); @@ -172,50 +172,50 @@ XLA_TEST_P(BatchNormalizationTest, VarianceToStddev) { XLA_TEST_P(BatchNormalizationTest, SpecComparisonForward) { XlaBuilder builder("batch_normalize_per_spec"); auto input_activations = - CheckShape(&builder, builder.ConstantLiteral(input_literal_), + CheckShape(&builder, ConstantLiteral(&builder, input_literal_), ShapeUtil::MakeShape(F32, {3, 2, 1, 1})); - auto gamma = builder.ConstantR1({1.0, 1.0}); - auto beta = builder.ConstantR1({0.0, 0.0}); + auto gamma = ConstantR1(&builder, {1.0, 1.0}); + auto beta = ConstantR1(&builder, {0.0, 0.0}); XlaComputation add = CreateScalarAddComputation(F32, &builder); // Reduce all dimensions except dimension 1. Shape TwoElementVectorF32 = ShapeUtil::MakeShape(F32, {2}); auto sum = CheckShape( &builder, - builder.Reduce(input_activations, builder.ConstantR0(0.0f), add, - /*dimensions_to_reduce=*/{0, 2, 3}), + Reduce(input_activations, ConstantR0(&builder, 0.0f), add, + /*dimensions_to_reduce=*/{0, 2, 3}), TwoElementVectorF32); auto input_shape = builder.GetShape(input_activations).ConsumeValueOrDie(); auto sum_shape = builder.GetShape(sum).ConsumeValueOrDie(); - auto count = builder.ConstantR0(ShapeUtil::ElementsIn(input_shape) / - ShapeUtil::ElementsIn(sum_shape)); - auto set_means = builder.Div(sum, count); + auto count = + ConstantR0(&builder, ShapeUtil::ElementsIn(input_shape) / + ShapeUtil::ElementsIn(sum_shape)); + auto set_means = Div(sum, count); const float kEpsilon = 1e-9f; - auto epsilon = builder.ConstantR0(kEpsilon); - auto epsilon2 = builder.ConstantR1({kEpsilon, kEpsilon}); - auto activation_deviations = builder.Sub(input_activations, set_means, - /*broadcast_dimensions=*/{1}); - auto dev_squares = builder.SquareF32(activation_deviations); - auto sum_of_squares = CheckShape( - &builder, - builder.Reduce(dev_squares, builder.ConstantR0(0.0f), add, - /*dimensions_to_reduce=*/{0, 2, 3}), - TwoElementVectorF32); - auto variance = builder.Div(sum_of_squares, count); - auto standard_deviation = builder.SqrtF32(variance); + auto epsilon = ConstantR0(&builder, kEpsilon); + auto epsilon2 = ConstantR1(&builder, {kEpsilon, kEpsilon}); + auto activation_deviations = Sub(input_activations, set_means, + /*broadcast_dimensions=*/{1}); + auto dev_squares = Square(activation_deviations); + auto sum_of_squares = + CheckShape(&builder, + Reduce(dev_squares, ConstantR0(&builder, 0.0f), add, + /*dimensions_to_reduce=*/{0, 2, 3}), + TwoElementVectorF32); + auto variance = Div(sum_of_squares, count); + auto standard_deviation = Sqrt(variance); auto standard_deviation_above_epsilon = - CheckShape(&builder, builder.Gt(standard_deviation, epsilon), + CheckShape(&builder, Gt(standard_deviation, epsilon), ShapeUtil::MakeShape(PRED, {2})); - auto gt_eps = builder.Select(standard_deviation_above_epsilon, - standard_deviation, epsilon2); - auto normalization_factors = builder.ReciprocalF32(gt_eps); + auto gt_eps = + Select(standard_deviation_above_epsilon, standard_deviation, epsilon2); + auto normalization_factors = Reciprocal(gt_eps); auto normalized_input_activations = - builder.Mul(activation_deviations, normalization_factors, - /*broadcast_dimensions=*/{1}); - /* auto output_activations = */ builder.Add( - builder.Mul(normalized_input_activations, gamma, - /*broadcast_dimensions=*/{1}), - beta, /*broadcast_dimensions=*/{1}); + Mul(activation_deviations, normalization_factors, + /*broadcast_dimensions=*/{1}); + /* auto output_activations = */ Add(Mul(normalized_input_activations, gamma, + /*broadcast_dimensions=*/{1}), + beta, /*broadcast_dimensions=*/{1}); Array4D expected(kSamples, kZ, kY, kX); Array2D pz({ @@ -232,46 +232,47 @@ XLA_TEST_P(BatchNormalizationTest, BasicTraining) { const int kFeatureIndex = 3; XlaBuilder builder(TestName()); - auto operand = builder.ConstantR4FromArray4D( - {{{{1.f, 2.f}}, {{3.f, 4.f}}}, {{{5.f, 6.f}}, {{7.f, 8.f}}}}); + auto operand = ConstantR4FromArray4D( + &builder, {{{{1.f, 2.f}}, {{3.f, 4.f}}}, {{{5.f, 6.f}}, {{7.f, 8.f}}}}); - auto scale = builder.ConstantR1({2.0f, 3.0f}); + auto scale = ConstantR1(&builder, {2.0f, 3.0f}); - auto offset = builder.ConstantR1({1.0f, 2.0f}); + auto offset = ConstantR1(&builder, {1.0f, 2.0f}); - builder.BatchNormTraining(operand, scale, offset, - /*epsilon=*/0.001, kFeatureIndex); + BatchNormTraining(operand, scale, offset, + /*epsilon=*/0.001, kFeatureIndex); - auto expected = Literal::MakeTuple( - {Literal::CreateR4({{{{-1.6f, -2.0f}}, {{0.1f, 0.6f}}}, - {{{1.9f, 3.3f}}, {{3.7f, 6.0f}}}}) + auto expected = LiteralUtil::MakeTuple( + {LiteralUtil::CreateR4({{{{-1.6f, -2.0f}}, {{0.1f, 0.6f}}}, + {{{1.9f, 3.3f}}, {{3.7f, 6.0f}}}}) .get(), - Literal::CreateR1({4, 5}).get(), - Literal::CreateR1({5, 5}).get()}); + LiteralUtil::CreateR1({4, 5}).get(), + LiteralUtil::CreateR1({5, 5}).get()}); ComputeAndCompareTuple(&builder, *expected, {}, ErrorSpec(0.1)); } -XLA_TEST_P(BatchNormalizationTest, BasicTrainingOnSublane) { +XLA_TEST_P(BatchNormalizationTest, BasicTrainingOnDimension2) { const int kFeatureIndex = 2; XlaBuilder builder(TestName()); - auto operand = builder.ConstantR4FromArray4D( + auto operand = ConstantR4FromArray4D( + &builder, {{{{1.f}, {2.f}}, {{3.f}, {4.f}}}, {{{5.f}, {6.f}}, {{7.f}, {8.f}}}}); - auto scale = builder.ConstantR1({2.0f, 3.0f}); + auto scale = ConstantR1(&builder, {2.0f, 3.0f}); - auto offset = builder.ConstantR1({1.0f, 2.0f}); + auto offset = ConstantR1(&builder, {1.0f, 2.0f}); - builder.BatchNormTraining(operand, scale, offset, - /*epsilon=*/0.001, kFeatureIndex); + BatchNormTraining(operand, scale, offset, + /*epsilon=*/0.001, kFeatureIndex); - auto expected = Literal::MakeTuple( - {Literal::CreateR4({{{{-1.6f}, {-2.0f}}, {{0.1f}, {0.6f}}}, - {{{1.9f}, {3.3f}}, {{3.7f}, {6.0f}}}}) + auto expected = LiteralUtil::MakeTuple( + {LiteralUtil::CreateR4({{{{-1.6f}, {-2.0f}}, {{0.1f}, {0.6f}}}, + {{{1.9f}, {3.3f}}, {{3.7f}, {6.0f}}}}) .get(), - Literal::CreateR1({4, 5}).get(), - Literal::CreateR1({5, 5}).get()}); + LiteralUtil::CreateR1({4, 5}).get(), + LiteralUtil::CreateR1({5, 5}).get()}); ComputeAndCompareTuple(&builder, *expected, {}, ErrorSpec(0.1)); } @@ -294,14 +295,14 @@ XLA_TEST_P(BatchNormalizationTest, TrainingWithFeatureOnLowDimension) { CreateR1Parameter(std::vector(260, 1.0f), /*parameter_number=*/2, "offset", &builder, &h2); - builder.BatchNormTraining(h0, h1, h2, - /*epsilon=*/1, kFeatureIndex); + BatchNormTraining(h0, h1, h2, + /*epsilon=*/1, kFeatureIndex); - auto expected = Literal::MakeTuple( - {Literal::CreateR3FromArray3D(Array3D(260, 2, 2, 1.0f)) + auto expected = LiteralUtil::MakeTuple( + {LiteralUtil::CreateR3FromArray3D(Array3D(260, 2, 2, 1.0f)) .get(), - Literal::CreateR1(std::vector(260, 1.0f)).get(), - Literal::CreateR1(std::vector(260, 0.0f)).get()}); + LiteralUtil::CreateR1(std::vector(260, 1.0f)).get(), + LiteralUtil::CreateR1(std::vector(260, 0.0f)).get()}); ComputeAndCompareTuple(&builder, *expected, {operand.get(), scale.get(), offset.get()}, @@ -327,14 +328,15 @@ XLA_TEST_P(BatchNormalizationTest, LargeEpsilonTest) { /*parameter_number=*/2, "offset", &builder, &h2); // var = 125, mean = 15, epsilon = -100 - builder.BatchNormTraining(h0, h1, h2, - /*epsilon=*/-100, kFeatureIndex); + BatchNormTraining(h0, h1, h2, + /*epsilon=*/-100, kFeatureIndex); - auto expected = Literal::MakeTuple( - {Literal::CreateR3FromArray3D({{{-3.0f}, {-1.0f}, {1.0f}, {3.0f}}}) + auto expected = LiteralUtil::MakeTuple( + {LiteralUtil::CreateR3FromArray3D( + {{{-3.0f}, {-1.0f}, {1.0f}, {3.0f}}}) .get(), - Literal::CreateR1(std::vector(1, 15.0f)).get(), - Literal::CreateR1(std::vector(1, 125.0f)).get()}); + LiteralUtil::CreateR1(std::vector(1, 15.0f)).get(), + LiteralUtil::CreateR1(std::vector(1, 125.0f)).get()}); ComputeAndCompareTuple(&builder, *expected, {operand.get(), scale.get(), offset.get()}, @@ -346,26 +348,27 @@ XLA_TEST_P(BatchNormalizationTest, BatchNormGradBasic) { XlaBuilder builder(TestName()); auto operand = - builder.ConstantR4FromArray4D(Array4D(2, 2, 2, 1, 0.0f)); + ConstantR4FromArray4D(&builder, Array4D(2, 2, 2, 1, 0.0f)); - auto scale = builder.ConstantR1({1.0f, 1.0f}); + auto scale = ConstantR1(&builder, {1.0f, 1.0f}); - auto mean = builder.ConstantR1({0.0f, 0.0f}); + auto mean = ConstantR1(&builder, {0.0f, 0.0f}); - auto var = builder.ConstantR1({1.0f, 1.0f}); + auto var = ConstantR1(&builder, {1.0f, 1.0f}); - auto grad_output = builder.ConstantR4FromArray4D( + auto grad_output = ConstantR4FromArray4D( + &builder, {{{{1.f}, {2.f}}, {{3.f}, {4.f}}}, {{{5.f}, {6.f}}, {{7.f}, {8.f}}}}); - builder.BatchNormGrad(operand, scale, mean, var, grad_output, - /*epsilon=*/0.0, kFeatureIndex); + BatchNormGrad(operand, scale, mean, var, grad_output, + /*epsilon=*/0.0, kFeatureIndex); - auto expected = Literal::MakeTuple( - {Literal::CreateR4({{{{-3.f}, {-3.f}}, {{-1.f}, {-1.f}}}, - {{{1.f}, {1.f}}, {{3.f}, {3.f}}}}) + auto expected = LiteralUtil::MakeTuple( + {LiteralUtil::CreateR4({{{{-3.f}, {-3.f}}, {{-1.f}, {-1.f}}}, + {{{1.f}, {1.f}}, {{3.f}, {3.f}}}}) .get(), - Literal::CreateR1({0, 0}).get(), - Literal::CreateR1({16, 20}).get()}); + LiteralUtil::CreateR1({0, 0}).get(), + LiteralUtil::CreateR1({16, 20}).get()}); ComputeAndCompareTuple(&builder, *expected, {}, ErrorSpec(0.1)); } @@ -511,22 +514,23 @@ XLA_TEST_P(BatchNormTestManySizes, RandomizedTrainingTests) { auto normalized = *ReferenceUtil::BatchNorm4D(input_array, mean4D, var4D, scale4D, offset4D, epsilon); - auto expected_normalized = Literal::CreateR4FromArray4D(normalized); + auto expected_normalized = + LiteralUtil::CreateR4FromArray4D(normalized); - auto offset_literal = Literal::CreateR1(offset); - auto scale_literal = Literal::CreateR1(scale); - auto input_literal = Literal::CreateR4FromArray4D(input_array); + auto offset_literal = LiteralUtil::CreateR1(offset); + auto scale_literal = LiteralUtil::CreateR1(scale); + auto input_literal = LiteralUtil::CreateR4FromArray4D(input_array); auto input_activations = - builder.Parameter(0, input_literal->shape(), "input"); + Parameter(&builder, 0, input_literal->shape(), "input"); auto scale_activations = - builder.Parameter(1, scale_literal->shape(), "offset"); + Parameter(&builder, 1, scale_literal->shape(), "offset"); auto offset_activations = - builder.Parameter(2, offset_literal->shape(), "scale"); + Parameter(&builder, 2, offset_literal->shape(), "scale"); - auto expected = Literal::MakeTuple({expected_normalized.get(), - Literal::CreateR1(mean).get(), - Literal::CreateR1(var).get()}); + auto expected = LiteralUtil::MakeTuple( + {expected_normalized.get(), LiteralUtil::CreateR1(mean).get(), + LiteralUtil::CreateR1(var).get()}); std::unique_ptr input_data = client_->TransferToServer(*input_literal).ConsumeValueOrDie(); @@ -535,8 +539,8 @@ XLA_TEST_P(BatchNormTestManySizes, RandomizedTrainingTests) { std::unique_ptr offset_data = client_->TransferToServer(*offset_literal).ConsumeValueOrDie(); - builder.BatchNormTraining(input_activations, scale_activations, - offset_activations, epsilon, feature_index); + BatchNormTraining(input_activations, scale_activations, offset_activations, + epsilon, feature_index); // Run all HLO passes during this test. In particular, ClientLibraryTestBase // disables constant folding, but we want it enabled for our zero-sized tensor @@ -611,21 +615,21 @@ XLA_TEST_P(BatchNormTestManySizes, RandomizedInferencingTests) { auto normalized = *ReferenceUtil::BatchNorm4D(input_array, mean4D, var4D, scale4D, offset4D, epsilon); - auto offset_literal = Literal::CreateR1(offset); - auto scale_literal = Literal::CreateR1(scale); - auto mean_literal = Literal::CreateR1(mean); - auto var_literal = Literal::CreateR1(var); - auto input_literal = Literal::CreateR4FromArray4D(input_array); + auto offset_literal = LiteralUtil::CreateR1(offset); + auto scale_literal = LiteralUtil::CreateR1(scale); + auto mean_literal = LiteralUtil::CreateR1(mean); + auto var_literal = LiteralUtil::CreateR1(var); + auto input_literal = LiteralUtil::CreateR4FromArray4D(input_array); auto input_activations = - builder.Parameter(0, input_literal->shape(), "input"); + Parameter(&builder, 0, input_literal->shape(), "input"); auto scale_activations = - builder.Parameter(1, scale_literal->shape(), "offset"); + Parameter(&builder, 1, scale_literal->shape(), "offset"); auto offset_activations = - builder.Parameter(2, offset_literal->shape(), "scale"); - auto mean_activations = builder.Parameter(3, mean_literal->shape(), "mean"); + Parameter(&builder, 2, offset_literal->shape(), "scale"); + auto mean_activations = Parameter(&builder, 3, mean_literal->shape(), "mean"); auto variance_activations = - builder.Parameter(4, var_literal->shape(), "variance"); + Parameter(&builder, 4, var_literal->shape(), "variance"); Array4D expected = normalized; @@ -640,9 +644,9 @@ XLA_TEST_P(BatchNormTestManySizes, RandomizedInferencingTests) { std::unique_ptr variance_data = client_->TransferToServer(*var_literal).ConsumeValueOrDie(); - builder.BatchNormInference(input_activations, scale_activations, - offset_activations, mean_activations, - variance_activations, epsilon, feature_index); + BatchNormInference(input_activations, scale_activations, offset_activations, + mean_activations, variance_activations, epsilon, + feature_index); // Run all HLO passes during this test. In particular, ClientLibraryTestBase // disables constant folding, but we want it enabled for our zero-sized tensor @@ -798,21 +802,23 @@ XLA_TEST_P(BatchNormTestManySizes, RandomizedGradTests) { }); auto expected_grad_activation = - Literal::CreateR4FromArray4D(grad_activation); + LiteralUtil::CreateR4FromArray4D(grad_activation); - auto input_literal = Literal::CreateR4FromArray4D(input_array); - auto scale_literal = Literal::CreateR1(scale); - auto mean_literal = Literal::CreateR1(mean); - auto var_literal = Literal::CreateR1(var); + auto input_literal = LiteralUtil::CreateR4FromArray4D(input_array); + auto scale_literal = LiteralUtil::CreateR1(scale); + auto mean_literal = LiteralUtil::CreateR1(mean); + auto var_literal = LiteralUtil::CreateR1(var); auto grad_output_literal = - Literal::CreateR4FromArray4D(grad_output_array); - - auto input_parameter = builder.Parameter(0, input_literal->shape(), "input"); - auto scale_parameter = builder.Parameter(1, scale_literal->shape(), "scale"); - auto mean_parameter = builder.Parameter(2, mean_literal->shape(), "mean"); - auto var_parameter = builder.Parameter(3, var_literal->shape(), "variance"); + LiteralUtil::CreateR4FromArray4D(grad_output_array); + + auto input_parameter = + Parameter(&builder, 0, input_literal->shape(), "input"); + auto scale_parameter = + Parameter(&builder, 1, scale_literal->shape(), "scale"); + auto mean_parameter = Parameter(&builder, 2, mean_literal->shape(), "mean"); + auto var_parameter = Parameter(&builder, 3, var_literal->shape(), "variance"); auto grad_output_parameter = - builder.Parameter(4, grad_output_literal->shape(), "grad_output"); + Parameter(&builder, 4, grad_output_literal->shape(), "grad_output"); std::unique_ptr input_data = client_->TransferToServer(*input_literal).ConsumeValueOrDie(); @@ -825,14 +831,13 @@ XLA_TEST_P(BatchNormTestManySizes, RandomizedGradTests) { std::unique_ptr grad_output_data = client_->TransferToServer(*grad_output_literal).ConsumeValueOrDie(); - builder.BatchNormGrad(input_parameter, scale_parameter, mean_parameter, - var_parameter, grad_output_parameter, epsilon, - feature_index); + BatchNormGrad(input_parameter, scale_parameter, mean_parameter, var_parameter, + grad_output_parameter, epsilon, feature_index); auto expected = - Literal::MakeTuple({expected_grad_activation.get(), - Literal::CreateR1(grad_scale).get(), - Literal::CreateR1(grad_offset).get()}); + LiteralUtil::MakeTuple({expected_grad_activation.get(), + LiteralUtil::CreateR1(grad_scale).get(), + LiteralUtil::CreateR1(grad_offset).get()}); // Run all HLO passes during this test. In particular, ClientLibraryTestBase // disables constant folding, but we want it enabled for our zero-sized tensor diff --git a/tensorflow/compiler/xla/tests/bfloat16_test.cc b/tensorflow/compiler/xla/tests/bfloat16_test.cc index ca337e78840e77377719636cd4cf33af2578210d..747c82b502c8ec9f8121641382d9fd3c9552b010 100644 --- a/tensorflow/compiler/xla/tests/bfloat16_test.cc +++ b/tensorflow/compiler/xla/tests/bfloat16_test.cc @@ -22,7 +22,7 @@ limitations under the License. #include "tensorflow/compiler/xla/client/lib/arithmetic.h" #include "tensorflow/compiler/xla/client/local_client.h" #include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/reference_util.h" #include "tensorflow/compiler/xla/service/hlo_computation.h" #include "tensorflow/compiler/xla/service/hlo_instruction.h" @@ -51,9 +51,9 @@ class Bfloat16Test : public ClientLibraryTestBase { XLA_TEST_F(Bfloat16Test, ScalarOperation) { XlaBuilder builder(TestName()); - auto x = builder.ConstantR0(static_cast(2.0f)); - auto y = builder.ConstantR0(static_cast(1.0f)); - builder.Add(x, y); + auto x = ConstantR0(&builder, static_cast(2.0f)); + auto y = ConstantR0(&builder, static_cast(1.0f)); + Add(x, y); ComputeAndCompareR0(&builder, static_cast(3.0f), {}, error_spec_); @@ -61,8 +61,8 @@ XLA_TEST_F(Bfloat16Test, ScalarOperation) { XLA_TEST_F(Bfloat16Test, LogOperation) { XlaBuilder builder(TestName()); - auto x = builder.ConstantR0(static_cast(4.0f)); - builder.Log(x); + auto x = ConstantR0(&builder, static_cast(4.0f)); + Log(x); ComputeAndCompareR0(&builder, static_cast(1.387f), {}, error_spec_); @@ -70,7 +70,7 @@ XLA_TEST_F(Bfloat16Test, LogOperation) { XLA_TEST_F(Bfloat16Test, NegateScalarF16) { XlaBuilder builder(TestName()); - builder.Neg(builder.ConstantR0(static_cast(2.1f))); + Neg(ConstantR0(&builder, static_cast(2.1f))); ComputeAndCompareR0(&builder, static_cast(-2.1f), {}, error_spec_); @@ -80,33 +80,33 @@ XLA_TEST_F(Bfloat16Test, BatchNormTraining) { const int kFeatureIndex = 2; XlaBuilder builder(TestName()); - auto operand = builder.ConstantR4FromArray4D( + auto operand = ConstantR4FromArray4D( + &builder, {{{{static_cast(1.f)}, {static_cast(2.f)}}, {{static_cast(3.f)}, {static_cast(4.f)}}}, {{{static_cast(5.f)}, {static_cast(6.f)}}, {{static_cast(7.f)}, {static_cast(8.f)}}}}); - auto scale = builder.ConstantR1( - {static_cast(2.0f), static_cast(3.0f)}); + auto scale = ConstantR1( + &builder, {static_cast(2.0f), static_cast(3.0f)}); - auto offset = builder.ConstantR1( - {static_cast(1.0f), static_cast(2.0f)}); + auto offset = ConstantR1( + &builder, {static_cast(1.0f), static_cast(2.0f)}); - auto tuple = builder.BatchNormTraining(operand, scale, offset, - /*epsilon=*/0.001, kFeatureIndex); + BatchNormTraining(operand, scale, offset, /*epsilon=*/0.001, kFeatureIndex); - auto expected = Literal::MakeTuple( - {Literal::CreateR4( + auto expected = LiteralUtil::MakeTuple( + {LiteralUtil::CreateR4( {{{{static_cast(-1.6875f)}, {static_cast(-2.04f)}}, {{static_cast(0.105f)}, {static_cast(0.66f)}}}, {{{static_cast(1.89f)}, {static_cast(3.35f)}}, {{static_cast(3.7f)}, {static_cast(6.04f)}}}}) .get(), - Literal::CreateR1( + LiteralUtil::CreateR1( {static_cast(4), static_cast(5)}) .get(), - Literal::CreateR1( + LiteralUtil::CreateR1( {static_cast(5), static_cast(5)}) .get()}); @@ -117,38 +117,39 @@ XLA_TEST_F(Bfloat16Test, BatchNormGrad) { const int kFeatureIndex = 2; XlaBuilder builder(TestName()); - auto operand = builder.ConstantR4FromArray4D( - Array4D(2, 2, 2, 1, static_cast(0.0f))); + auto operand = ConstantR4FromArray4D( + &builder, Array4D(2, 2, 2, 1, static_cast(0.0f))); - auto scale = builder.ConstantR1( - {static_cast(1.0f), static_cast(1.0f)}); + auto scale = ConstantR1( + &builder, {static_cast(1.0f), static_cast(1.0f)}); - auto mean = builder.ConstantR1( - {static_cast(0.0f), static_cast(0.0f)}); + auto mean = ConstantR1( + &builder, {static_cast(0.0f), static_cast(0.0f)}); - auto var = builder.ConstantR1( - {static_cast(1.0f), static_cast(1.0f)}); + auto var = ConstantR1( + &builder, {static_cast(1.0f), static_cast(1.0f)}); - auto grad_output = builder.ConstantR4FromArray4D( + auto grad_output = ConstantR4FromArray4D( + &builder, {{{{static_cast(1.f)}, {static_cast(2.f)}}, {{static_cast(3.f)}, {static_cast(4.f)}}}, {{{static_cast(5.f)}, {static_cast(6.f)}}, {{static_cast(7.f)}, {static_cast(8.f)}}}}); - builder.BatchNormGrad(operand, scale, mean, var, grad_output, - /*epsilon=*/0.0, kFeatureIndex); + BatchNormGrad(operand, scale, mean, var, grad_output, + /*epsilon=*/0.0, kFeatureIndex); - auto expected = Literal::MakeTuple( - {Literal::CreateR4( + auto expected = LiteralUtil::MakeTuple( + {LiteralUtil::CreateR4( {{{{static_cast(-3.f)}, {static_cast(-3.f)}}, {{static_cast(-1.f)}, {static_cast(-1.f)}}}, {{{static_cast(1.f)}, {static_cast(1.f)}}, {{static_cast(3.f)}, {static_cast(3.f)}}}}) .get(), - Literal::CreateR1( + LiteralUtil::CreateR1( {static_cast(0), static_cast(0)}) .get(), - Literal::CreateR1( + LiteralUtil::CreateR1( {static_cast(16), static_cast(20)}) .get()}); diff --git a/tensorflow/compiler/xla/tests/binop_scaling_test.cc b/tensorflow/compiler/xla/tests/binop_scaling_test.cc index 48203b1d40ea69ff00a57c2c9e42620739b23d59..20cb989751ad69e2f3cf97c87c43293951f599ab 100644 --- a/tensorflow/compiler/xla/tests/binop_scaling_test.cc +++ b/tensorflow/compiler/xla/tests/binop_scaling_test.cc @@ -33,9 +33,9 @@ TEST_F(BinopScalingTest, MatrixPlusPseudoMatrixRowVector_32x4) { auto arhs = MakeLinspaceArray2D(0.0, 1.0, 1, 4); XlaBuilder builder(TestName()); - auto lhs = builder.ConstantR2FromArray2D(*alhs); - auto rhs = builder.ConstantR2FromArray2D(*arhs); - builder.Add(lhs, rhs); + auto lhs = ConstantR2FromArray2D(&builder, *alhs); + auto rhs = ConstantR2FromArray2D(&builder, *arhs); + Add(lhs, rhs); auto aexpected = ReferenceUtil::MapWithIndexArray2D( *alhs, [&](float lhs_value, int64 row, int64 col) { @@ -49,9 +49,9 @@ TEST_F(BinopScalingTest, MatrixPlusPseudoMatrixRowVector_129x129) { auto arhs = MakeLinspaceArray2D(0.0, 1.0, 1, 129); XlaBuilder builder(TestName()); - auto lhs = builder.ConstantR2FromArray2D(*alhs); - auto rhs = builder.ConstantR2FromArray2D(*arhs); - builder.Add(lhs, rhs); + auto lhs = ConstantR2FromArray2D(&builder, *alhs); + auto rhs = ConstantR2FromArray2D(&builder, *arhs); + Add(lhs, rhs); auto aexpected = ReferenceUtil::MapWithIndexArray2D( *alhs, [&](float lhs_value, int64 row, int64 col) { @@ -65,9 +65,9 @@ TEST_F(BinopScalingTest, MatrixPlusPseudoMatrixColVector_9x5) { auto arhs = MakeLinspaceArray2D(0.0, 1.0, 9, 1); XlaBuilder builder(TestName()); - auto lhs = builder.ConstantR2FromArray2D(*alhs); - auto rhs = builder.ConstantR2FromArray2D(*arhs); - builder.Add(lhs, rhs); + auto lhs = ConstantR2FromArray2D(&builder, *alhs); + auto rhs = ConstantR2FromArray2D(&builder, *arhs); + Add(lhs, rhs); auto aexpected = ReferenceUtil::MapWithIndexArray2D( *alhs, [&](float lhs_value, int64 row, int64 col) { @@ -81,9 +81,9 @@ TEST_F(BinopScalingTest, MatrixPlusPseudoMatrixColVector_129x257) { auto arhs = MakeLinspaceArray2D(0.0, 1.0, 129, 1); XlaBuilder builder(TestName()); - auto lhs = builder.ConstantR2FromArray2D(*alhs); - auto rhs = builder.ConstantR2FromArray2D(*arhs); - builder.Add(lhs, rhs); + auto lhs = ConstantR2FromArray2D(&builder, *alhs); + auto rhs = ConstantR2FromArray2D(&builder, *arhs); + Add(lhs, rhs); auto aexpected = ReferenceUtil::MapWithIndexArray2D( *alhs, [&](float lhs_value, int64 row, int64 col) { @@ -94,11 +94,12 @@ TEST_F(BinopScalingTest, MatrixPlusPseudoMatrixColVector_129x257) { TEST_F(BinopScalingTest, R0PlusR2F32) { XlaBuilder builder(TestName()); - auto lhs = builder.ConstantR0(42.0); - auto rhs = builder.ConstantR2({ - {1.0, 2.0}, {3.0, 4.0}, - }); - builder.Add(lhs, rhs); + auto lhs = ConstantR0(&builder, 42.0); + auto rhs = ConstantR2(&builder, { + {1.0, 2.0}, + {3.0, 4.0}, + }); + Add(lhs, rhs); Array2D expected(2, 2); expected(0, 0) = 42.0 + 1.0; @@ -129,9 +130,9 @@ TEST_F(BinopScalingTest, R4PlusR0S32) { }); // clang-format on - auto lhs = builder.ConstantR4FromArray4D(lhs_array); - auto rhs = builder.ConstantR0(42); - builder.Add(lhs, rhs); + auto lhs = ConstantR4FromArray4D(&builder, lhs_array); + auto rhs = ConstantR0(&builder, 42); + Add(lhs, rhs); ComputeAndCompareR4(&builder, expected, {}); } diff --git a/tensorflow/compiler/xla/tests/bitcast_convert_test.cc b/tensorflow/compiler/xla/tests/bitcast_convert_test.cc index bff60f25ec8f15d372d251ac313200301a04f20f..d531e8fa82e47f7bcd278f10da2c205e44db0ac1 100644 --- a/tensorflow/compiler/xla/tests/bitcast_convert_test.cc +++ b/tensorflow/compiler/xla/tests/bitcast_convert_test.cc @@ -43,8 +43,8 @@ class BitcastConvertTest : public ClientLibraryTestBase { TEST_F(BitcastConvertTest, ConvertR1S32ToR1S32) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({42, 64}); - builder.BitcastConvertType(a, S32); + auto a = ConstantR1(&builder, {42, 64}); + BitcastConvertType(a, S32); std::vector expected = {42, 64}; ComputeAndCompareR1(&builder, expected, {}); @@ -52,8 +52,8 @@ TEST_F(BitcastConvertTest, ConvertR1S32ToR1S32) { TEST_F(BitcastConvertTest, ConvertR1F32ToR1F32) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({42.0f, 64.0f}); - builder.BitcastConvertType(a, F32); + auto a = ConstantR1(&builder, {42.0f, 64.0f}); + BitcastConvertType(a, F32); std::vector expected = {42.0f, 64.0f}; ComputeAndCompareR1(&builder, expected, {}); @@ -62,10 +62,10 @@ TEST_F(BitcastConvertTest, ConvertR1F32ToR1F32) { TEST_F(BitcastConvertTest, BitcastR1S32ToR1F32) { XlaBuilder builder(TestName()); auto a = - builder.ConstantR1({0, static_cast(0x80000000), 0x3F800000, - static_cast(0xBF800000), 0x3F000000, - static_cast(0xBF000000)}); - builder.BitcastConvertType(a, F32); + ConstantR1(&builder, {0, static_cast(0x80000000), + 0x3F800000, static_cast(0xBF800000), + 0x3F000000, static_cast(0xBF000000)}); + BitcastConvertType(a, F32); std::vector expected = {0.0f, -0.0f, 1.0f, -1.0f, 0.5f, -0.5f}; ComputeAndCompareR1(&builder, expected, {}); @@ -73,8 +73,8 @@ TEST_F(BitcastConvertTest, BitcastR1S32ToR1F32) { XLA_TEST_F(BitcastConvertTest, ConvertR1S0S32ToR1S0F32) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({}); - builder.BitcastConvertType(a, F32); + auto a = ConstantR1(&builder, {}); + BitcastConvertType(a, F32); std::vector expected = {}; ComputeAndCompareR1(&builder, expected, {}); @@ -82,8 +82,8 @@ XLA_TEST_F(BitcastConvertTest, ConvertR1S0S32ToR1S0F32) { TEST_F(BitcastConvertTest, ConvertR1F32ToR1S32) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({42.6, 64.4}); - builder.BitcastConvertType(a, S32); + auto a = ConstantR1(&builder, {42.6, 64.4}); + BitcastConvertType(a, S32); std::vector expected = {0x422a6666, 0x4280cccd}; ComputeAndCompareR1(&builder, expected, {}); @@ -91,9 +91,9 @@ TEST_F(BitcastConvertTest, ConvertR1F32ToR1S32) { TEST_F(BitcastConvertTest, ConvertS32Extremes) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1( - {std::numeric_limits::min(), std::numeric_limits::max()}); - builder.BitcastConvertType(a, F32); + auto a = ConstantR1(&builder, {std::numeric_limits::min(), + std::numeric_limits::max()}); + BitcastConvertType(a, F32); std::vector expected = {-0.0f, NAN}; ComputeAndCompareR1(&builder, expected, {}, ErrorSpec(0, 0)); @@ -102,10 +102,10 @@ TEST_F(BitcastConvertTest, ConvertS32Extremes) { TEST_F(BitcastConvertTest, ConvertMapToS32) { XlaBuilder builder(TestName()); auto b = builder.CreateSubBuilder("convert"); - auto param = b->Parameter(0, ShapeUtil::MakeShape(F32, {}), "in"); - b->BitcastConvertType(param, S32); - auto a = builder.ConstantR1({42.0f, 64.0f}); - builder.Map({a}, b->BuildAndNoteError(), {0}); + auto param = Parameter(b.get(), 0, ShapeUtil::MakeShape(F32, {}), "in"); + BitcastConvertType(param, S32); + auto a = ConstantR1(&builder, {42.0f, 64.0f}); + Map(&builder, {a}, b->BuildAndNoteError(), {0}); std::vector expected = {0x42280000, 0x42800000}; ComputeAndCompareR1(&builder, expected, {}); @@ -114,10 +114,10 @@ TEST_F(BitcastConvertTest, ConvertMapToS32) { TEST_F(BitcastConvertTest, ConvertMapToF32) { XlaBuilder builder(TestName()); auto b = builder.CreateSubBuilder("convert"); - auto param = b->Parameter(0, ShapeUtil::MakeShape(S32, {}), "in"); - b->BitcastConvertType(param, F32); - auto a = builder.ConstantR1({0x42280000, 0x42800000}); - builder.Map({a}, b->BuildAndNoteError(), {0}); + auto param = Parameter(b.get(), 0, ShapeUtil::MakeShape(S32, {}), "in"); + BitcastConvertType(param, F32); + auto a = ConstantR1(&builder, {0x42280000, 0x42800000}); + Map(&builder, {a}, b->BuildAndNoteError(), {0}); std::vector expected = {42.0f, 64.0f}; ComputeAndCompareR1(&builder, expected, {}); @@ -130,9 +130,9 @@ TEST_F(BitcastConvertTest, ConvertMapToF32) { // the new convert should have the same element type as the old convert. TEST_F(BitcastConvertTest, ConvertReshape) { XlaBuilder builder(TestName()); - auto input = builder.ConstantR1({0x42280000}); - auto reshape = builder.Reshape(input, /*dimensions=*/{0}, /*new_sizes=*/{}); - builder.BitcastConvertType(reshape, F32); + auto input = ConstantR1(&builder, {0x42280000}); + auto reshape = Reshape(input, /*dimensions=*/{0}, /*new_sizes=*/{}); + BitcastConvertType(reshape, F32); ComputeAndCompareR0(&builder, 42.0f, {}); } diff --git a/tensorflow/compiler/xla/tests/broadcast_simple_test.cc b/tensorflow/compiler/xla/tests/broadcast_simple_test.cc index 3a0f51fc66d65c8684bd607b9e8103559cd4d8d4..50dd574624bb3874e682be5a272fb5bdefa4adc4 100644 --- a/tensorflow/compiler/xla/tests/broadcast_simple_test.cc +++ b/tensorflow/compiler/xla/tests/broadcast_simple_test.cc @@ -21,6 +21,7 @@ limitations under the License. #include "tensorflow/compiler/xla/array4d.h" #include "tensorflow/compiler/xla/client/local_client.h" #include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/literal_util.h" #include "tensorflow/compiler/xla/statusor.h" #include "tensorflow/compiler/xla/test.h" @@ -37,17 +38,17 @@ class BroadcastSimpleTest : public ClientLibraryTestBase { XlaBuilder* builder) { switch (op) { case HloOpcode::kMinimum: { - return builder->Min(lhs, rhs); + return Min(lhs, rhs); } case HloOpcode::kMaximum: { - return builder->Max(lhs, rhs); + return Max(lhs, rhs); } case HloOpcode::kMultiply: { - return builder->Mul(lhs, rhs); + return Mul(lhs, rhs); } default: { // Default to Add - return builder->Add(lhs, rhs); + return Add(lhs, rhs); } } } @@ -58,7 +59,7 @@ class BroadcastSimpleTest : public ClientLibraryTestBase { Array3D* r3_array, float start, float end, int seed) { *r3_shape = ShapeUtil::MakeShapeWithLayout(F32, bounds, minor_to_major); r3_array->FillRandom(start, end, seed); - auto r3_data = Literal::CreateR3FromArray3D(*r3_array)->Relayout( + auto r3_data = LiteralUtil::CreateR3FromArray3D(*r3_array)->Relayout( LayoutUtil::MakeLayout(minor_to_major)); std::unique_ptr r3_global_data = client_->TransferToServer(*r3_data).ConsumeValueOrDie(); @@ -71,7 +72,7 @@ class BroadcastSimpleTest : public ClientLibraryTestBase { Array2D* r2_array, float start, float end, int seed) { *r2_shape = ShapeUtil::MakeShapeWithLayout(F32, bounds, minor_to_major); r2_array->FillRandom(start, end, seed); - auto r2_data = Literal::CreateR2FromArray2D(*r2_array)->Relayout( + auto r2_data = LiteralUtil::CreateR2FromArray2D(*r2_array)->Relayout( LayoutUtil::MakeLayout(minor_to_major)); std::unique_ptr r2_global_data = client_->TransferToServer(*r2_data).ConsumeValueOrDie(); @@ -104,13 +105,13 @@ using ::testing::HasSubstr; XLA_TEST_F(BroadcastSimpleTest, ScalarNoOpBroadcast) { XlaBuilder b(TestName()); - b.Broadcast(b.ConstantR0(1.5), {}); + Broadcast(ConstantR0(&b, 1.5), {}); ComputeAndCompareR0(&b, 1.5, {}, ErrorSpec(0.0001)); } XLA_TEST_F(BroadcastSimpleTest, ScalarTo2D_2x3) { XlaBuilder b(TestName()); - b.Broadcast(b.ConstantR0(2.25), {2, 3}); + Broadcast(ConstantR0(&b, 2.25), {2, 3}); Array2D expected(2, 3, 2.25); ComputeAndCompareR2(&b, expected, {}, ErrorSpec(0.0001)); } @@ -122,7 +123,7 @@ XLA_TEST_F(BroadcastSimpleTest, ScalarParamTo2D_2x3) { CreateR0Parameter(2.25f, /*parameter_number=*/0, /*name=*/"src", /*builder=*/&b, /*data_handle=*/&src); - b.Broadcast(src, {2, 3}); + Broadcast(src, {2, 3}); Array2D expected(2, 3, 2.25); ComputeAndCompareR2(&b, expected, {param_data.get()}, ErrorSpec(0.0001)); @@ -130,21 +131,21 @@ XLA_TEST_F(BroadcastSimpleTest, ScalarParamTo2D_2x3) { XLA_TEST_F(BroadcastSimpleTest, ScalarTo2D_2x0) { XlaBuilder b(TestName()); - b.Broadcast(b.ConstantR0(2.25), {2, 0}); + Broadcast(ConstantR0(&b, 2.25), {2, 0}); Array2D expected(2, 0); ComputeAndCompareR2(&b, expected, {}, ErrorSpec(0.0001)); } XLA_TEST_F(BroadcastSimpleTest, ScalarTo2D_0x2) { XlaBuilder b(TestName()); - b.Broadcast(b.ConstantR0(2.25), {0, 2}); + Broadcast(ConstantR0(&b, 2.25), {0, 2}); Array2D expected(0, 2); ComputeAndCompareR2(&b, expected, {}, ErrorSpec(0.0001)); } XLA_TEST_F(BroadcastSimpleTest, 1DTo2D) { XlaBuilder b(TestName()); - b.Broadcast(b.ConstantR1({1, 2, 3}), {2}); + Broadcast(ConstantR1(&b, {1, 2, 3}), {2}); Array2D expected(2, 3); expected(0, 0) = 1; @@ -156,6 +157,86 @@ XLA_TEST_F(BroadcastSimpleTest, 1DTo2D) { ComputeAndCompareR2(&b, expected, {}, ErrorSpec(0.0001)); } +XLA_TEST_F(BroadcastSimpleTest, 1DTo2D_WithDimsUsual) { + XlaBuilder b(TestName()); + BroadcastInDim(ConstantR1(&b, {1, 2}), + ShapeUtil::MakeShape(F32, {2, 2}), {1}); + + Array2D expected(2, 2); + expected(0, 0) = 1; + expected(0, 1) = 2; + expected(1, 0) = 1; + expected(1, 1) = 2; + + ComputeAndCompareR2(&b, expected, {}, ErrorSpec(0.0001)); +} + +XLA_TEST_F(BroadcastSimpleTest, 1DTo2D_WithDimsTranspose) { + XlaBuilder b(TestName()); + BroadcastInDim(ConstantR1(&b, {1, 2}), + ShapeUtil::MakeShape(F32, {2, 2}), {0}); + + Array2D expected(2, 2); + expected(0, 0) = 1; + expected(0, 1) = 1; + expected(1, 0) = 2; + expected(1, 1) = 2; + + ComputeAndCompareR2(&b, expected, {}, ErrorSpec(0.0001)); +} + +XLA_TEST_F(BroadcastSimpleTest, 2DTo3D_WithDims) { + XlaBuilder b(TestName()); + BroadcastInDim(ConstantR2(&b, {{1.0, 5.0}, {2.0, 6.0}}), + ShapeUtil::MakeShape(F32, {2, 2, 2}), {0, 1}); + + Array3D expected(2, 2, 2); + expected(0, 0, 0) = 1.0; + expected(1, 0, 0) = 2.0; + expected(0, 0, 1) = 1.0; + expected(1, 0, 1) = 2.0; + expected(0, 1, 0) = 5.0; + expected(1, 1, 0) = 6.0; + expected(1, 1, 1) = 6.0; + expected(0, 1, 1) = 5.0; + + ComputeAndCompareR3(&b, expected, {}, ErrorSpec(0.0001)); +} + +XLA_TEST_F(BroadcastSimpleTest, 2DTo3D_WithDimsNotPossibleWithBroadCast) { + XlaBuilder b(TestName()); + BroadcastInDim(ConstantR2(&b, {{1.0, 5.0}, {2.0, 6.0}}), + ShapeUtil::MakeShape(F32, {2, 2, 2}), {0, 2}); + + Array3D expected(2, 2, 2); + expected(0, 0, 0) = 1.0; + expected(1, 0, 0) = 2.0; + expected(0, 0, 1) = 5.0; + expected(1, 0, 1) = 6.0; + expected(0, 1, 0) = 1.0; + expected(1, 1, 0) = 2.0; + expected(1, 1, 1) = 6.0; + expected(0, 1, 1) = 5.0; + + ComputeAndCompareR3(&b, expected, {}, ErrorSpec(0.0001)); +} + +XLA_TEST_F(BroadcastSimpleTest, 1DTo2D_WithDimsNotPossibleWithBroadCast) { + XlaBuilder b(TestName()); + BroadcastInDim(ConstantR1(&b, {1, 2}), + ShapeUtil::MakeShape(F32, {3, 2}), {1}); + + Array2D expected(3, 2); + expected(0, 0) = 1; + expected(0, 1) = 2; + expected(1, 0) = 1; + expected(1, 1) = 2; + expected(2, 0) = 1; + expected(2, 1) = 2; + + ComputeAndCompareR2(&b, expected, {}, ErrorSpec(0.0001)); +} + // Tests implicit broadcasting of PREDs. XLA_TEST_F(BroadcastSimpleTest, BooleanAnd2DTo3D_Pred) { XlaBuilder b(TestName()); @@ -172,7 +253,7 @@ XLA_TEST_F(BroadcastSimpleTest, BooleanAnd2DTo3D_Pred) { XlaOp x, y; auto x_data = CreateR2Parameter(x_vals, 0, "x", &b, &x); auto y_data = CreateR3Parameter(y_vals, 1, "y", &b, &y); - b.And(x, y, /*broadcast_dimensions=*/{1, 2}); + And(x, y, /*broadcast_dimensions=*/{1, 2}); Array3D expected(2, 2, 1); expected(0, 0, 0) = false; @@ -185,7 +266,7 @@ XLA_TEST_F(BroadcastSimpleTest, BooleanAnd2DTo3D_Pred) { XLA_TEST_F(BroadcastSimpleTest, ZeroElement_1DTo2D) { XlaBuilder b(TestName()); - b.Broadcast(b.ConstantR1({}), {2}); + Broadcast(ConstantR1(&b, {}), {2}); Array2D expected(2, 0); ComputeAndCompareR2(&b, expected, {}, ErrorSpec(0.0001)); @@ -193,7 +274,7 @@ XLA_TEST_F(BroadcastSimpleTest, ZeroElement_1DTo2D) { XLA_TEST_F(BroadcastSimpleTest, 1DToZeroElement2D) { XlaBuilder b(TestName()); - b.Broadcast(b.ConstantR1({1, 2, 3}), {0}); + Broadcast(ConstantR1(&b, {1, 2, 3}), {0}); Array2D expected(0, 3); ComputeAndCompareR2(&b, expected, {}, ErrorSpec(0.0001)); @@ -209,14 +290,14 @@ XLA_TEST_F(BroadcastSimpleTest, InDimensionAndDegenerateBroadcasting) { // dimensions. XlaBuilder b(TestName()); - b.Add(b.ConstantR2({{1.0, 5.0}}), - b.ConstantLiteral(*Literal::CreateR3( - {{{2.0}, {3.0}, {4.0}}, {{5.0}, {6.0}, {7.0}}})), - /*broadcast_dimensions=*/{1, 2}); + Add(ConstantR2(&b, {{1.0, 5.0}}), + ConstantLiteral(&b, *LiteralUtil::CreateR3( + {{{2.0}, {3.0}, {4.0}}, {{5.0}, {6.0}, {7.0}}})), + /*broadcast_dimensions=*/{1, 2}); auto expected = - Literal::CreateR3({{{3.0, 7.0}, {4.0, 8.0}, {5.0, 9.0}}, - {{6.0, 10.0}, {7.0, 11.0}, {8.0, 12.0}}}); + LiteralUtil::CreateR3({{{3.0, 7.0}, {4.0, 8.0}, {5.0, 9.0}}, + {{6.0, 10.0}, {7.0, 11.0}, {8.0, 12.0}}}); ComputeAndCompareLiteral(&b, *expected, {}, ErrorSpec(0.0001)); } @@ -260,9 +341,10 @@ XLA_TEST_P(BroadcastR3ImplicitTest, Doit) { MakeR3Data(spec.input_bounds, spec.minor2major_layout, &r3_implicit_shape, &r3_implicit_array, 1.0, 0.2, 56789); - auto r3_implicit_parameter = builder.Parameter(0, r3_implicit_shape, "input"); - auto r3_parameter = builder.Parameter(1, r3_shape, "input"); - XlaOp op = BuildBinOp(spec.op, r3_implicit_parameter, r3_parameter, &builder); + auto r3_implicit_parameter = + Parameter(&builder, 0, r3_implicit_shape, "input"); + auto r3_parameter = Parameter(&builder, 1, r3_shape, "input"); + BuildBinOp(spec.op, r3_implicit_parameter, r3_parameter, &builder); Array3D expected_array(spec.output_bounds[0], spec.output_bounds[1], spec.output_bounds[2]); @@ -284,7 +366,7 @@ XLA_TEST_P(BroadcastR3ImplicitTest, Doit) { } } } - auto expected = Literal::CreateR3FromArray3D(expected_array); + auto expected = LiteralUtil::CreateR3FromArray3D(expected_array); ComputeAndCompareLiteral( &builder, *expected, {r3_implicit_global_data.get(), r3_global_data.get()}, @@ -306,10 +388,10 @@ XLA_TEST_F(BroadcastSimpleTest, Add3DTo3DDegenerate_1_2) { auto r1 = CreateR3Parameter(r1d, 1, "r1", &b, &r1h); auto r3 = CreateR3Parameter(r3d, 0, "r3", &b, &r3h); - b.Add(r3h, r1h); + Add(r3h, r1h); auto expected = - Literal::CreateR3({{{2, 3}, {4, 5}}, {{7, 8}, {9, 10}}}); + LiteralUtil::CreateR3({{{2, 3}, {4, 5}}, {{7, 8}, {9, 10}}}); ComputeAndCompareLiteral(&b, *expected, {r3.get(), r1.get()}, ErrorSpec(0.0001)); @@ -317,79 +399,81 @@ XLA_TEST_F(BroadcastSimpleTest, Add3DTo3DDegenerate_1_2) { XLA_TEST_F(BroadcastSimpleTest, Add3DTo3DDegenerate_0_1) { XlaBuilder b(TestName()); - auto r1 = b.ConstantLiteral(*Literal::CreateR3({{{1, 2}}})); - auto r3 = b.ConstantLiteral( - *Literal::CreateR3({{{1, 2}, {3, 4}}, {{5, 6}, {7, 8}}})); - b.Add(r3, r1); + auto r1 = ConstantLiteral(&b, *LiteralUtil::CreateR3({{{1, 2}}})); + auto r3 = ConstantLiteral( + &b, *LiteralUtil::CreateR3({{{1, 2}, {3, 4}}, {{5, 6}, {7, 8}}})); + Add(r3, r1); auto expected = - Literal::CreateR3({{{2, 4}, {4, 6}}, {{6, 8}, {8, 10}}}); + LiteralUtil::CreateR3({{{2, 4}, {4, 6}}, {{6, 8}, {8, 10}}}); ComputeAndCompareLiteral(&b, *expected, {}, ErrorSpec(0.0001)); } XLA_TEST_F(BroadcastSimpleTest, Add3DTo3DDegenerate_0_2) { XlaBuilder b(TestName()); - auto r1 = b.ConstantLiteral(*Literal::CreateR3({{{1}, {2}}})); - auto r3 = b.ConstantLiteral( - *Literal::CreateR3({{{1, 2}, {3, 4}}, {{5, 6}, {7, 8}}})); - b.Add(r3, r1); + auto r1 = ConstantLiteral(&b, *LiteralUtil::CreateR3({{{1}, {2}}})); + auto r3 = ConstantLiteral( + &b, *LiteralUtil::CreateR3({{{1, 2}, {3, 4}}, {{5, 6}, {7, 8}}})); + Add(r3, r1); auto expected = - Literal::CreateR3({{{2, 3}, {5, 6}}, {{6, 7}, {9, 10}}}); + LiteralUtil::CreateR3({{{2, 3}, {5, 6}}, {{6, 7}, {9, 10}}}); ComputeAndCompareLiteral(&b, *expected, {}, ErrorSpec(0.0001)); } XLA_TEST_F(BroadcastSimpleTest, Add3DTo3DDegenerate_0) { XlaBuilder b(TestName()); - auto r1 = b.ConstantLiteral(*Literal::CreateR3({{{1, 2}, {3, 4}}})); - auto r3 = b.ConstantLiteral( - *Literal::CreateR3({{{1, 2}, {3, 4}}, {{5, 6}, {7, 8}}})); - b.Add(r3, r1); + auto r1 = + ConstantLiteral(&b, *LiteralUtil::CreateR3({{{1, 2}, {3, 4}}})); + auto r3 = ConstantLiteral( + &b, *LiteralUtil::CreateR3({{{1, 2}, {3, 4}}, {{5, 6}, {7, 8}}})); + Add(r3, r1); auto expected = - Literal::CreateR3({{{2, 4}, {6, 8}}, {{6, 8}, {10, 12}}}); + LiteralUtil::CreateR3({{{2, 4}, {6, 8}}, {{6, 8}, {10, 12}}}); ComputeAndCompareLiteral(&b, *expected, {}, ErrorSpec(0.0001)); } XLA_TEST_F(BroadcastSimpleTest, Add3DTo3DDegenerate_1) { XlaBuilder b(TestName()); - auto r1 = b.ConstantLiteral(*Literal::CreateR3({{{1, 2}}, {{3, 4}}})); - auto r3 = b.ConstantLiteral( - *Literal::CreateR3({{{1, 2}, {3, 4}}, {{5, 6}, {7, 8}}})); - b.Add(r3, r1); + auto r1 = + ConstantLiteral(&b, *LiteralUtil::CreateR3({{{1, 2}}, {{3, 4}}})); + auto r3 = ConstantLiteral( + &b, *LiteralUtil::CreateR3({{{1, 2}, {3, 4}}, {{5, 6}, {7, 8}}})); + Add(r3, r1); auto expected = - Literal::CreateR3({{{2, 4}, {4, 6}}, {{8, 10}, {10, 12}}}); + LiteralUtil::CreateR3({{{2, 4}, {4, 6}}, {{8, 10}, {10, 12}}}); ComputeAndCompareLiteral(&b, *expected, {}, ErrorSpec(0.0001)); } XLA_TEST_F(BroadcastSimpleTest, Add3DTo3DDegenerate_2) { XlaBuilder b(TestName()); - auto r1 = - b.ConstantLiteral(*Literal::CreateR3({{{1}, {2}}, {{3}, {4}}})); - auto r3 = b.ConstantLiteral( - *Literal::CreateR3({{{1, 2}, {3, 4}}, {{5, 6}, {7, 8}}})); - b.Add(r3, r1); + auto r1 = ConstantLiteral( + &b, *LiteralUtil::CreateR3({{{1}, {2}}, {{3}, {4}}})); + auto r3 = ConstantLiteral( + &b, *LiteralUtil::CreateR3({{{1, 2}, {3, 4}}, {{5, 6}, {7, 8}}})); + Add(r3, r1); auto expected = - Literal::CreateR3({{{2, 3}, {5, 6}}, {{8, 9}, {11, 12}}}); + LiteralUtil::CreateR3({{{2, 3}, {5, 6}}, {{8, 9}, {11, 12}}}); ComputeAndCompareLiteral(&b, *expected, {}, ErrorSpec(0.0001)); } XLA_TEST_F(BroadcastSimpleTest, Add3DTo3DDegenerate_0_1_2) { XlaBuilder b(TestName()); - auto r1 = b.ConstantLiteral(*Literal::CreateR3({{{1}}})); - auto r3 = b.ConstantLiteral( - *Literal::CreateR3({{{1, 2}, {3, 4}}, {{5, 6}, {7, 8}}})); - b.Add(r3, r1); + auto r1 = ConstantLiteral(&b, *LiteralUtil::CreateR3({{{1}}})); + auto r3 = ConstantLiteral( + &b, *LiteralUtil::CreateR3({{{1, 2}, {3, 4}}, {{5, 6}, {7, 8}}})); + Add(r3, r1); auto expected = - Literal::CreateR3({{{2, 3}, {4, 5}}, {{6, 7}, {8, 9}}}); + LiteralUtil::CreateR3({{{2, 3}, {4, 5}}, {{6, 7}, {8, 9}}}); ComputeAndCompareLiteral(&b, *expected, {}, ErrorSpec(0.0001)); } @@ -509,14 +593,14 @@ XLA_TEST_P(BroadcastR2ImplicitTest, Doit) { &r2_implicit_shape2, &r2_implicit_array2, 0.8, 0.4, 56789); auto r2_implicit_parameter1 = - builder.Parameter(0, r2_implicit_shape1, "input0"); - auto r2_parameter = builder.Parameter(1, r2_shape, "input1"); + Parameter(&builder, 0, r2_implicit_shape1, "input0"); + auto r2_parameter = Parameter(&builder, 1, r2_shape, "input1"); auto r2_implicit_parameter2 = - builder.Parameter(2, r2_implicit_shape2, "input2"); + Parameter(&builder, 2, r2_implicit_shape2, "input2"); XlaOp op1 = BuildBinOp(spec.op1, r2_implicit_parameter1, r2_parameter, &builder); - XlaOp op2 = BuildBinOp(spec.op2, op1, r2_implicit_parameter2, &builder); + BuildBinOp(spec.op2, op1, r2_implicit_parameter2, &builder); Array2D expected_array(spec.output_bounds[0], spec.output_bounds[1]); @@ -530,7 +614,7 @@ XLA_TEST_P(BroadcastR2ImplicitTest, Doit) { *v = ApplyOpToFloats(spec.op2, tmp, v3); }); - auto expected = Literal::CreateR2FromArray2D(expected_array); + auto expected = LiteralUtil::CreateR2FromArray2D(expected_array); ComputeAndCompareLiteral( &builder, *expected, {r2_implicit_global_data1.get(), r2_global_data.get(), @@ -544,80 +628,82 @@ INSTANTIATE_TEST_CASE_P(BroadcastR2ImplicitTestInstances, XLA_TEST_F(BroadcastSimpleTest, Add2DTo2DDegenerate_0) { XlaBuilder b(TestName()); - auto r1 = b.ConstantLiteral(*Literal::CreateR2({{1, 2}})); - auto r2 = b.ConstantLiteral(*Literal::CreateR2({{1, 2}, {3, 4}})); - b.Add(r2, r1); + auto r1 = ConstantLiteral(&b, *LiteralUtil::CreateR2({{1, 2}})); + auto r2 = + ConstantLiteral(&b, *LiteralUtil::CreateR2({{1, 2}, {3, 4}})); + Add(r2, r1); - auto expected = Literal::CreateR2({{2, 4}, {4, 6}}); + auto expected = LiteralUtil::CreateR2({{2, 4}, {4, 6}}); ComputeAndCompareLiteral(&b, *expected, {}, ErrorSpec(0.0001)); } XLA_TEST_F(BroadcastSimpleTest, Add2DTo2DDegenerate_1) { XlaBuilder b(TestName()); - auto r1 = b.ConstantLiteral(*Literal::CreateR2({{1}, {2}})); - auto r2 = b.ConstantLiteral(*Literal::CreateR2({{1, 2}, {3, 4}})); - b.Add(r2, r1); + auto r1 = ConstantLiteral(&b, *LiteralUtil::CreateR2({{1}, {2}})); + auto r2 = + ConstantLiteral(&b, *LiteralUtil::CreateR2({{1, 2}, {3, 4}})); + Add(r2, r1); - auto expected = Literal::CreateR2({{2, 3}, {5, 6}}); + auto expected = LiteralUtil::CreateR2({{2, 3}, {5, 6}}); ComputeAndCompareLiteral(&b, *expected, {}, ErrorSpec(0.0001)); } XLA_TEST_F(BroadcastSimpleTest, Add1DTo3DInDim0) { XlaBuilder b(TestName()); - auto r1 = b.ConstantR1({10, 20}); - auto r3 = b.ConstantLiteral( - *Literal::CreateR3({{{1, 2}, {3, 4}}, {{5, 6}, {7, 8}}})); - b.Add(r3, r1, {0}); + auto r1 = ConstantR1(&b, {10, 20}); + auto r3 = ConstantLiteral( + &b, *LiteralUtil::CreateR3({{{1, 2}, {3, 4}}, {{5, 6}, {7, 8}}})); + Add(r3, r1, {0}); - auto expected = - Literal::CreateR3({{{11, 12}, {13, 14}}, {{25, 26}, {27, 28}}}); + auto expected = LiteralUtil::CreateR3( + {{{11, 12}, {13, 14}}, {{25, 26}, {27, 28}}}); ComputeAndCompareLiteral(&b, *expected, {}, ErrorSpec(0.0001)); } XLA_TEST_F(BroadcastSimpleTest, Add1DTo3DInDim1) { XlaBuilder b(TestName()); - auto r1 = b.ConstantR1({10, 20}); - auto r3 = b.ConstantLiteral( - *Literal::CreateR3({{{1, 2}, {3, 4}}, {{5, 6}, {7, 8}}})); - b.Add(r1, r3, {1}); + auto r1 = ConstantR1(&b, {10, 20}); + auto r3 = ConstantLiteral( + &b, *LiteralUtil::CreateR3({{{1, 2}, {3, 4}}, {{5, 6}, {7, 8}}})); + Add(r1, r3, {1}); - auto expected = - Literal::CreateR3({{{11, 12}, {23, 24}}, {{15, 16}, {27, 28}}}); + auto expected = LiteralUtil::CreateR3( + {{{11, 12}, {23, 24}}, {{15, 16}, {27, 28}}}); ComputeAndCompareLiteral(&b, *expected, {}, ErrorSpec(0.0001)); } XLA_TEST_F(BroadcastSimpleTest, Add1DTo3DInDim2) { XlaBuilder b(TestName()); - auto r1 = b.ConstantR1({10, 20}); - auto r3 = b.ConstantLiteral( - *Literal::CreateR3({{{1, 2}, {3, 4}}, {{5, 6}, {7, 8}}})); - b.Add(r1, r3, {2}); + auto r1 = ConstantR1(&b, {10, 20}); + auto r3 = ConstantLiteral( + &b, *LiteralUtil::CreateR3({{{1, 2}, {3, 4}}, {{5, 6}, {7, 8}}})); + Add(r1, r3, {2}); - auto expected = - Literal::CreateR3({{{11, 22}, {13, 24}}, {{15, 26}, {17, 28}}}); + auto expected = LiteralUtil::CreateR3( + {{{11, 22}, {13, 24}}, {{15, 26}, {17, 28}}}); ComputeAndCompareLiteral(&b, *expected, {}, ErrorSpec(0.0001)); } XLA_TEST_F(BroadcastSimpleTest, Add1DTo3DInDimAll) { XlaBuilder b(TestName()); - auto r1_0 = b.ConstantR1({1000, 2000}); - auto r1_1 = b.ConstantR1({100, 200}); - auto r1_2 = b.ConstantR1({10, 20}); - auto r3 = b.ConstantLiteral( - *Literal::CreateR3({{{1, 2}, {3, 4}}, {{5, 6}, {7, 8}}})); + auto r1_0 = ConstantR1(&b, {1000, 2000}); + auto r1_1 = ConstantR1(&b, {100, 200}); + auto r1_2 = ConstantR1(&b, {10, 20}); + auto r3 = ConstantLiteral( + &b, *LiteralUtil::CreateR3({{{1, 2}, {3, 4}}, {{5, 6}, {7, 8}}})); for (int i = 0; i < 3; ++i) { - r3 = b.Add(r1_0, r3, {0}); - r3 = b.Add(r3, r1_1, {1}); - r3 = b.Add(r1_2, r3, {2}); + r3 = Add(r1_0, r3, {0}); + r3 = Add(r3, r1_1, {1}); + r3 = Add(r1_2, r3, {2}); } - r3 = b.Mul(r3, b.ConstantR0(-2)); + r3 = Mul(r3, ConstantR0(&b, -2)); - auto expected = Literal::CreateR3( + auto expected = LiteralUtil::CreateR3( {{{-6 * 1110 - 2, -6 * 1120 - 4}, {-6 * 1210 - 6, -6 * 1220 - 8}}, {{-6 * 2110 - 10, -6 * 2120 - 12}, {-6 * 2210 - 14, -6 * 2220 - 16}}}); @@ -626,19 +712,19 @@ XLA_TEST_F(BroadcastSimpleTest, Add1DTo3DInDimAll) { XLA_TEST_F(BroadcastSimpleTest, Add1DTo3DInDimAllWithScalarBroadcast) { XlaBuilder b(TestName()); - auto r1_0 = b.ConstantR1({1000, 2000}); - auto r1_1 = b.ConstantR1({100, 200}); - auto r1_2 = b.ConstantR1({10, 20}); - auto r0 = b.ConstantR0(3); - auto r3 = b.Broadcast(r0, {2, 2, 2}); + auto r1_0 = ConstantR1(&b, {1000, 2000}); + auto r1_1 = ConstantR1(&b, {100, 200}); + auto r1_2 = ConstantR1(&b, {10, 20}); + auto r0 = ConstantR0(&b, 3); + auto r3 = Broadcast(r0, {2, 2, 2}); for (int i = 0; i < 3; ++i) { - r3 = b.Add(r1_0, r3, {0}); - r3 = b.Add(r3, r1_1, {1}); - r3 = b.Add(r1_2, r3, {2}); + r3 = Add(r1_0, r3, {0}); + r3 = Add(r3, r1_1, {1}); + r3 = Add(r1_2, r3, {2}); } - r3 = b.Mul(r3, b.ConstantR0(-1)); + r3 = Mul(r3, ConstantR0(&b, -1)); - auto expected = Literal::CreateR3( + auto expected = LiteralUtil::CreateR3( {{{-3 * 1110 - 3, -3 * 1120 - 3}, {-3 * 1210 - 3, -3 * 1220 - 3}}, {{-3 * 2110 - 3, -3 * 2120 - 3}, {-3 * 2210 - 3, -3 * 2220 - 3}}}); @@ -650,10 +736,10 @@ XLA_TEST_F(BroadcastSimpleTest, InvalidBinaryAndDegenerateBroadcasting) { // results in a shape incompatible with the lhs [2, 3, 1]. XlaBuilder b(TestName()); - b.Add(b.ConstantR2({{1.0, 5.0}, {1.0, 5.0}}), - b.ConstantLiteral(*Literal::CreateR3( - {{{2.0}, {3.0}, {4.0}}, {{5.0}, {6.0}, {7.0}}})), - /*broadcast_dimensions=*/{1, 2}); + Add(ConstantR2(&b, {{1.0, 5.0}, {1.0, 5.0}}), + ConstantLiteral(&b, *LiteralUtil::CreateR3( + {{{2.0}, {3.0}, {4.0}}, {{5.0}, {6.0}, {7.0}}})), + /*broadcast_dimensions=*/{1, 2}); auto result_status = Execute(&b, {}); EXPECT_FALSE(result_status.ok()); @@ -665,8 +751,8 @@ XLA_TEST_F(BroadcastSimpleTest, InvalidInDimensionBroadcasting) { // Test invalid broadcasting with [1, 2] and [2, 3] inputs. XlaBuilder b(TestName()); - b.Add(b.ConstantR2({{1.0, 2.0}}), - b.ConstantR2({{1.0, 2.0, 3.0}, {4.0, 5.0, 6.0}})); + Add(ConstantR2(&b, {{1.0, 2.0}}), + ConstantR2(&b, {{1.0, 2.0, 3.0}, {4.0, 5.0, 6.0}})); auto result_status = Execute(&b, {}); EXPECT_FALSE(result_status.ok()); @@ -678,8 +764,8 @@ XLA_TEST_F(BroadcastSimpleTest, InvalidDegenerateBroadcasting) { // Test invalid broadcasting with [1, 2] and [2, 3] inputs. XlaBuilder b(TestName()); - b.Add(b.ConstantR2({{1.0, 2.0}}), - b.ConstantR2({{1.0, 2.0, 3.0}, {4.0, 5.0, 6.0}})); + Add(ConstantR2(&b, {{1.0, 2.0}}), + ConstantR2(&b, {{1.0, 2.0, 3.0}, {4.0, 5.0, 6.0}})); auto result_status = Execute(&b, {}); EXPECT_FALSE(result_status.ok()); diff --git a/tensorflow/compiler/xla/tests/broadcast_test.cc b/tensorflow/compiler/xla/tests/broadcast_test.cc index 51b9f0d3e330e73f5d110f0a62f824179d5c7cf7..c7b94b5bbaaa512ad36056f9e68a87cc706c24b1 100644 --- a/tensorflow/compiler/xla/tests/broadcast_test.cc +++ b/tensorflow/compiler/xla/tests/broadcast_test.cc @@ -16,7 +16,7 @@ limitations under the License. #include #include -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/ptr_util.h" #include "tensorflow/compiler/xla/service/hlo_computation.h" #include "tensorflow/compiler/xla/service/hlo_instruction.h" @@ -37,7 +37,7 @@ XLA_TEST_F(BroadcastTest, BroadcastScalarToScalar) { // Test degenerate case of broadcasting a scalar into a scalar. auto builder = HloComputation::Builder(TestName()); auto input = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(42.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(42.0))); builder.AddInstruction(HloInstruction::CreateBroadcast( ShapeUtil::MakeShape(F32, {}), input, {})); @@ -46,14 +46,14 @@ XLA_TEST_F(BroadcastTest, BroadcastScalarToScalar) { hlo_module->AddEntryComputation(builder.Build()); auto result = ExecuteAndTransfer(std::move(hlo_module), {}); - EXPECT_TRUE(LiteralTestUtil::Near(*Literal::CreateR0(42.0), *result, - error_spec_)); + EXPECT_TRUE(LiteralTestUtil::Near(*LiteralUtil::CreateR0(42.0), + *result, error_spec_)); } XLA_TEST_F(BroadcastTest, BroadcastScalarTo2D) { auto builder = HloComputation::Builder(TestName()); auto input = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(42.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(42.0))); builder.AddInstruction(HloInstruction::CreateBroadcast( ShapeUtil::MakeShape(F32, {2, 2}), input, {})); @@ -63,14 +63,14 @@ XLA_TEST_F(BroadcastTest, BroadcastScalarTo2D) { auto result = ExecuteAndTransfer(std::move(hlo_module), {}); EXPECT_TRUE(LiteralTestUtil::Near( - *Literal::CreateR2({{42.0, 42.0}, {42.0, 42.0}}), *result, + *LiteralUtil::CreateR2({{42.0, 42.0}, {42.0, 42.0}}), *result, error_spec_)); } XLA_TEST_F(BroadcastTest, BroadcastVectorTo2D) { auto builder = HloComputation::Builder(TestName()); auto input = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR1({1.0, 2.0, 3.0}))); + LiteralUtil::CreateR1({1.0, 2.0, 3.0}))); // Broadcast vector in both dimension 0 and dimension 1. Join them in a tuple // to enable testing of the results. @@ -86,18 +86,18 @@ XLA_TEST_F(BroadcastTest, BroadcastVectorTo2D) { auto result = ExecuteAndTransfer(std::move(hlo_module), {}); EXPECT_TRUE(LiteralTestUtil::Near( - *Literal::CreateR2({{1.0, 1.0}, {2.0, 2.0}, {3.0, 3.0}}), + *LiteralUtil::CreateR2({{1.0, 1.0}, {2.0, 2.0}, {3.0, 3.0}}), LiteralSlice(*result, {0}), error_spec_)); EXPECT_TRUE(LiteralTestUtil::Near( - *Literal::CreateR2({{1.0, 2.0, 3.0}, {1.0, 2.0, 3.0}}), + *LiteralUtil::CreateR2({{1.0, 2.0, 3.0}, {1.0, 2.0, 3.0}}), LiteralSlice(*result, {1}), error_spec_)); } XLA_TEST_F(BroadcastTest, Broadcast2DTo2D) { auto builder = HloComputation::Builder(TestName()); auto input = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR2({{1.0, 2.0}, {3.0, 4.0}}))); + LiteralUtil::CreateR2({{1.0, 2.0}, {3.0, 4.0}}))); builder.AddInstruction(HloInstruction::CreateBroadcast( ShapeUtil::MakeShape(F32, {2, 2}), input, {0, 1})); @@ -106,9 +106,9 @@ XLA_TEST_F(BroadcastTest, Broadcast2DTo2D) { hlo_module->AddEntryComputation(builder.Build()); auto result = ExecuteAndTransfer(std::move(hlo_module), {}); - EXPECT_TRUE( - LiteralTestUtil::Near(*Literal::CreateR2({{1.0, 2.0}, {3.0, 4.0}}), - *result, error_spec_)); + EXPECT_TRUE(LiteralTestUtil::Near( + *LiteralUtil::CreateR2({{1.0, 2.0}, {3.0, 4.0}}), *result, + error_spec_)); } XLA_TEST_F(BroadcastTest, Broadcast2DTo2DTranspose) { @@ -116,7 +116,7 @@ XLA_TEST_F(BroadcastTest, Broadcast2DTo2DTranspose) { // the dimensions, ie transpose. auto builder = HloComputation::Builder(TestName()); auto input = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR2({{1.0, 2.0}, {3.0, 4.0}}))); + LiteralUtil::CreateR2({{1.0, 2.0}, {3.0, 4.0}}))); builder.AddInstruction(HloInstruction::CreateBroadcast( ShapeUtil::MakeShape(F32, {2, 2}), input, {1, 0})); @@ -125,15 +125,15 @@ XLA_TEST_F(BroadcastTest, Broadcast2DTo2DTranspose) { hlo_module->AddEntryComputation(builder.Build()); auto result = ExecuteAndTransfer(std::move(hlo_module), {}); - EXPECT_TRUE( - LiteralTestUtil::Near(*Literal::CreateR2({{1.0, 3.0}, {2.0, 4.0}}), - *result, error_spec_)); + EXPECT_TRUE(LiteralTestUtil::Near( + *LiteralUtil::CreateR2({{1.0, 3.0}, {2.0, 4.0}}), *result, + error_spec_)); } XLA_TEST_F(BroadcastTest, Broadcast2DTo3D) { auto builder = HloComputation::Builder(TestName()); auto input = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR2({{1.0, 2.0}, {3.0, 4.0}}))); + LiteralUtil::CreateR2({{1.0, 2.0}, {3.0, 4.0}}))); builder.AddInstruction(HloInstruction::CreateBroadcast( ShapeUtil::MakeShape(F32, {2, 3, 2}), input, {0, 2})); @@ -143,15 +143,15 @@ XLA_TEST_F(BroadcastTest, Broadcast2DTo3D) { auto result = ExecuteAndTransfer(std::move(hlo_module), {}); EXPECT_TRUE(LiteralTestUtil::Near( - *Literal::CreateR3({{{1.0, 2.0}, {1.0, 2.0}, {1.0, 2.0}}, - {{3.0, 4.0}, {3.0, 4.0}, {3.0, 4.0}}}), + *LiteralUtil::CreateR3({{{1.0, 2.0}, {1.0, 2.0}, {1.0, 2.0}}, + {{3.0, 4.0}, {3.0, 4.0}, {3.0, 4.0}}}), *result, error_spec_)); } TEST_F(BroadcastTest, Broadcast_R1_2_To_R4_2x2x3x3) { auto builder = HloComputation::Builder(TestName()); auto input = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR1({1.0, 2.0}))); + HloInstruction::CreateConstant(LiteralUtil::CreateR1({1.0, 2.0}))); // Broadcast vector in dimension 1. builder.AddInstruction(HloInstruction::CreateBroadcast( @@ -166,8 +166,9 @@ TEST_F(BroadcastTest, Broadcast_R1_2_To_R4_2x2x3x3) { Array2D pz({{1, 2}, {1, 2}}); expected.FillWithPZ(pz); - EXPECT_TRUE(LiteralTestUtil::Near( - *Literal::CreateR4FromArray4D(expected), *result, error_spec_)); + EXPECT_TRUE( + LiteralTestUtil::Near(*LiteralUtil::CreateR4FromArray4D(expected), + *result, error_spec_)); } TEST_F(BroadcastTest, Broadcast_R1_1025_To_R4_3x3x3x1025) { @@ -176,7 +177,7 @@ TEST_F(BroadcastTest, Broadcast_R1_1025_To_R4_3x3x3x1025) { int64 r1_size = input_data.size(); std::iota(input_data.begin(), input_data.end(), 0.0f); auto input = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR1(input_data))); + HloInstruction::CreateConstant(LiteralUtil::CreateR1(input_data))); // Broadcast vector in dimension 3. builder.AddInstruction(HloInstruction::CreateBroadcast( @@ -196,8 +197,9 @@ TEST_F(BroadcastTest, Broadcast_R1_1025_To_R4_3x3x3x1025) { } expected.FillWithYX(yx); - EXPECT_TRUE(LiteralTestUtil::Near( - *Literal::CreateR4FromArray4D(expected), *result, error_spec_)); + EXPECT_TRUE( + LiteralTestUtil::Near(*LiteralUtil::CreateR4FromArray4D(expected), + *result, error_spec_)); } XLA_TEST_F(BroadcastTest, Broadcast_R1_64_To_R4_32x64x7x7) { @@ -207,7 +209,7 @@ XLA_TEST_F(BroadcastTest, Broadcast_R1_64_To_R4_32x64x7x7) { std::vector r1_array(64, 42.0); auto input = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR1(r1_array))); + HloInstruction::CreateConstant(LiteralUtil::CreateR1(r1_array))); // Broadcast vector in dimension 1. builder.AddInstruction(HloInstruction::CreateBroadcast( @@ -218,14 +220,14 @@ XLA_TEST_F(BroadcastTest, Broadcast_R1_64_To_R4_32x64x7x7) { hlo_module->AddEntryComputation(builder.Build()); auto result = ExecuteAndTransfer(std::move(hlo_module), {}); - EXPECT_TRUE(LiteralTestUtil::Near(*Literal::CreateR4FromArray4D(r4_array), + EXPECT_TRUE(LiteralTestUtil::Near(*LiteralUtil::CreateR4FromArray4D(r4_array), *result, error_spec_)); } TEST_F(BroadcastTest, Broadcast_R0_to_R4_64x64x3x3) { auto builder = HloComputation::Builder(TestName()); auto input = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1.0f))); builder.AddInstruction(HloInstruction::CreateBroadcast( ShapeUtil::MakeShape(F32, {64, 64, 3, 3}), input, {})); @@ -238,15 +240,16 @@ TEST_F(BroadcastTest, Broadcast_R0_to_R4_64x64x3x3) { Array4D expected(64, 64, 3, 3); expected.Fill(1.0f); - EXPECT_TRUE(LiteralTestUtil::Near( - *Literal::CreateR4FromArray4D(expected), *result, error_spec_)); + EXPECT_TRUE( + LiteralTestUtil::Near(*LiteralUtil::CreateR4FromArray4D(expected), + *result, error_spec_)); } TEST_F(BroadcastTest, Broadcast_R2_2x2_To_R4_3x3x2x2) { auto builder = HloComputation::Builder(TestName()); Array2D to_broadcast({{1.0f, 2.0f}, {3.0f, 4.0f}}); auto input = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR2FromArray2D(to_broadcast))); + LiteralUtil::CreateR2FromArray2D(to_broadcast))); // Broadcast vector in dimensions 2 and 3. builder.AddInstruction(HloInstruction::CreateBroadcast( @@ -260,8 +263,9 @@ TEST_F(BroadcastTest, Broadcast_R2_2x2_To_R4_3x3x2x2) { Array4D expected(3, 3, 2, 2); expected.FillWithYX(to_broadcast); - EXPECT_TRUE(LiteralTestUtil::Near( - *Literal::CreateR4FromArray4D(expected), *result, error_spec_)); + EXPECT_TRUE( + LiteralTestUtil::Near(*LiteralUtil::CreateR4FromArray4D(expected), + *result, error_spec_)); } TEST_F(BroadcastTest, Broadcast_R3_2x3x4_to_R4_2x3x4x5) { @@ -280,7 +284,7 @@ TEST_F(BroadcastTest, Broadcast_R3_2x3x4_to_R4_2x3x4x5) { } } auto input = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR3FromArray3D(input_vals))); + LiteralUtil::CreateR3FromArray3D(input_vals))); // Broadcast vector in dimensions 2 and 3. builder.AddInstruction(HloInstruction::CreateBroadcast( @@ -291,8 +295,9 @@ TEST_F(BroadcastTest, Broadcast_R3_2x3x4_to_R4_2x3x4x5) { hlo_module->AddEntryComputation(builder.Build()); auto result = ExecuteAndTransfer(std::move(hlo_module), {}); - EXPECT_TRUE(LiteralTestUtil::Near( - *Literal::CreateR4FromArray4D(expected), *result, error_spec_)); + EXPECT_TRUE( + LiteralTestUtil::Near(*LiteralUtil::CreateR4FromArray4D(expected), + *result, error_spec_)); } } // namespace diff --git a/tensorflow/compiler/xla/tests/call_test.cc b/tensorflow/compiler/xla/tests/call_test.cc index 5fd33b50c94356839bbed58acd43b7d0286f4a7e..2086e38b91955b23ab11af73acd7faf46ca4bb18 100644 --- a/tensorflow/compiler/xla/tests/call_test.cc +++ b/tensorflow/compiler/xla/tests/call_test.cc @@ -18,6 +18,7 @@ limitations under the License. #include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/compiler/xla/client/xla_client/xla_computation.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/literal_util.h" #include "tensorflow/compiler/xla/shape_util.h" #include "tensorflow/compiler/xla/test_helpers.h" @@ -34,7 +35,7 @@ class CallOpTest : public ClientLibraryTestBase { protected: XlaComputation CreateR0F32IdentityComputation() { XlaBuilder builder("Identity"); - builder.Parameter(0, r0f32_, "x"); + Parameter(&builder, 0, r0f32_, "x"); auto build_status = builder.Build(); EXPECT_IS_OK(build_status.status()); return build_status.ConsumeValueOrDie(); @@ -42,9 +43,9 @@ class CallOpTest : public ClientLibraryTestBase { XlaComputation CreateR1S0F32AdditionComputation() { XlaBuilder builder("Addition"); - auto x = builder.Parameter(0, r1s0f32_, "x"); - auto y = builder.Parameter(1, r1s0f32_, "y"); - builder.Add(x, y); + auto x = Parameter(&builder, 0, r1s0f32_, "x"); + auto y = Parameter(&builder, 1, r1s0f32_, "y"); + Add(x, y); auto build_status = builder.Build(); EXPECT_IS_OK(build_status.status()); return build_status.ConsumeValueOrDie(); @@ -52,9 +53,9 @@ class CallOpTest : public ClientLibraryTestBase { XlaComputation CreateR1S2F32AdditionComputation() { XlaBuilder builder("Addition"); - auto x = builder.Parameter(0, r1s2f32_, "x"); - auto y = builder.Parameter(1, r1s2f32_, "y"); - builder.Add(x, y); + auto x = Parameter(&builder, 0, r1s2f32_, "x"); + auto y = Parameter(&builder, 1, r1s2f32_, "y"); + Add(x, y); auto build_status = builder.Build(); EXPECT_IS_OK(build_status.status()); return build_status.ConsumeValueOrDie(); @@ -62,7 +63,7 @@ class CallOpTest : public ClientLibraryTestBase { XlaComputation CreateR0F32TupleComputation() { XlaBuilder builder("Tuple"); - builder.Tuple({builder.Parameter(0, r0f32_, "x")}); + Tuple(&builder, {Parameter(&builder, 0, r0f32_, "x")}); auto build_status = builder.Build(); EXPECT_IS_OK(build_status.status()); return build_status.ConsumeValueOrDie(); @@ -76,8 +77,9 @@ class CallOpTest : public ClientLibraryTestBase { XLA_TEST_F(CallOpTest, CallR0F32IdentityScalar) { XlaBuilder builder(TestName()); XlaComputation callee = CreateR0F32IdentityComputation(); - auto constant = builder.ConstantLiteral(*Literal::CreateR0(42.0)); - builder.Call(callee, {constant}); + auto constant = + ConstantLiteral(&builder, *LiteralUtil::CreateR0(42.0)); + Call(&builder, callee, {constant}); ComputeAndCompareR0(&builder, 42.0, {}, ErrorSpec(0.01f)); } @@ -85,9 +87,9 @@ XLA_TEST_F(CallOpTest, CallR0F32IdentityScalar) { XLA_TEST_F(CallOpTest, CallR1S0F32AddArray) { XlaBuilder builder(TestName()); XlaComputation callee = CreateR1S0F32AdditionComputation(); - auto x = builder.ConstantLiteral(*Literal::CreateR1({})); - auto y = builder.ConstantLiteral(*Literal::CreateR1({})); - builder.Call(callee, {x, y}); + auto x = ConstantLiteral(&builder, *LiteralUtil::CreateR1({})); + auto y = ConstantLiteral(&builder, *LiteralUtil::CreateR1({})); + Call(&builder, callee, {x, y}); ComputeAndCompareR1(&builder, {}, {}, ErrorSpec(0.01f)); } @@ -95,9 +97,11 @@ XLA_TEST_F(CallOpTest, CallR1S0F32AddArray) { XLA_TEST_F(CallOpTest, CallR1S2F32AddArray) { XlaBuilder builder(TestName()); XlaComputation callee = CreateR1S2F32AdditionComputation(); - auto x = builder.ConstantLiteral(*Literal::CreateR1({1.0f, 2.0f})); - auto y = builder.ConstantLiteral(*Literal::CreateR1({2.0f, 3.0f})); - builder.Call(callee, {x, y}); + auto x = + ConstantLiteral(&builder, *LiteralUtil::CreateR1({1.0f, 2.0f})); + auto y = + ConstantLiteral(&builder, *LiteralUtil::CreateR1({2.0f, 3.0f})); + Call(&builder, callee, {x, y}); ComputeAndCompareR1(&builder, {3.0f, 5.0f}, {}, ErrorSpec(0.01f)); } @@ -105,40 +109,40 @@ XLA_TEST_F(CallOpTest, CallR1S2F32AddArray) { XLA_TEST_F(CallOpTest, CallTreeTwoDeepBranchFactorThree) { XlaBuilder builder("inner"); { - auto x = builder.Parameter(0, r0f32_, "x"); - builder.Add(x, builder.ConstantR0(1.0)); + auto x = Parameter(&builder, 0, r0f32_, "x"); + Add(x, ConstantR0(&builder, 1.0)); } TF_ASSERT_OK_AND_ASSIGN(XlaComputation inner, builder.Build()); XlaBuilder builder2("outer"); { - auto x = builder2.Parameter(0, r0f32_, "x"); - x = builder2.Call(inner, {x}); - x = builder2.Call(inner, {x}); - x = builder2.Call(inner, {x}); + auto x = Parameter(&builder2, 0, r0f32_, "x"); + x = Call(&builder2, inner, {x}); + x = Call(&builder2, inner, {x}); + x = Call(&builder2, inner, {x}); } TF_ASSERT_OK_AND_ASSIGN(XlaComputation outer, builder2.Build()); XlaBuilder builder3("outermost"); { - auto x = builder3.Parameter(0, r0f32_, "x"); - x = builder3.Call(outer, {x}); - x = builder3.Call(outer, {x}); - x = builder3.Call(outer, {x}); + auto x = Parameter(&builder3, 0, r0f32_, "x"); + x = Call(&builder3, outer, {x}); + x = Call(&builder3, outer, {x}); + x = Call(&builder3, outer, {x}); } TF_ASSERT_OK_AND_ASSIGN( std::unique_ptr start, - client_->TransferToServer(*Literal::CreateR0(1.0f))); + client_->TransferToServer(*LiteralUtil::CreateR0(1.0f))); ComputeAndCompareR0(&builder3, 10.0f, {start.get()}, ErrorSpec(0.0f)); } XLA_TEST_F(CallOpTest, CallR0F32Tuple) { XlaBuilder builder(TestName()); XlaComputation callee = CreateR0F32TupleComputation(); - auto elem = Literal::CreateR0(42.0); - auto tuple = Literal::MakeTuple({elem.get()}); - builder.Call(callee, {builder.ConstantLiteral(*elem)}); + auto elem = LiteralUtil::CreateR0(42.0); + auto tuple = LiteralUtil::MakeTuple({elem.get()}); + Call(&builder, callee, {ConstantLiteral(&builder, *elem)}); ComputeAndCompareTuple(&builder, *tuple, {}, ErrorSpec(0.01f)); } diff --git a/tensorflow/compiler/xla/tests/check_execution_arity_test.cc b/tensorflow/compiler/xla/tests/check_execution_arity_test.cc index 660ff0cad5666219a4a7cb1eedbed03f06e651ba..0bc8facfe2cfcfab094f483137f6d8e241c6aaf9 100644 --- a/tensorflow/compiler/xla/tests/check_execution_arity_test.cc +++ b/tensorflow/compiler/xla/tests/check_execution_arity_test.cc @@ -18,7 +18,7 @@ limitations under the License. #include "tensorflow/compiler/xla/client/global_data.h" #include "tensorflow/compiler/xla/client/local_client.h" #include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/shape_util.h" #include "tensorflow/compiler/xla/statusor.h" #include "tensorflow/compiler/xla/test.h" @@ -36,11 +36,11 @@ class CheckExecutionArityTest : public ClientLibraryTestBase {}; TEST_F(CheckExecutionArityTest, TwoParamComputationNumArguments) { XlaBuilder builder("add_two_params"); - auto param_literal = Literal::CreateR1({1.1f, 2.2f}); + auto param_literal = LiteralUtil::CreateR1({1.1f, 2.2f}); - auto p0 = builder.Parameter(0, param_literal->shape(), "param0"); - auto p1 = builder.Parameter(1, param_literal->shape(), "param1"); - auto add = builder.Add(p0, p1); + auto p0 = Parameter(&builder, 0, param_literal->shape(), "param0"); + auto p1 = Parameter(&builder, 1, param_literal->shape(), "param1"); + Add(p0, p1); auto param0_data = client_->TransferToServer(*param_literal).ConsumeValueOrDie(); @@ -77,20 +77,20 @@ TEST_F(CheckExecutionArityTest, TwoParamComputationNumArguments) { XLA_TEST_F(CheckExecutionArityTest, CheckArgumentShapes) { XlaBuilder builder("add_two_params"); - auto p0 = builder.Parameter(0, ShapeUtil::MakeShape(F32, {}), "param0"); - auto p1 = builder.Parameter(1, ShapeUtil::MakeShape(F32, {4}), "param1"); - auto add = builder.Mul(p0, p1); + auto p0 = Parameter(&builder, 0, ShapeUtil::MakeShape(F32, {}), "param0"); + auto p1 = Parameter(&builder, 1, ShapeUtil::MakeShape(F32, {4}), "param1"); + Mul(p0, p1); auto computation_status = builder.Build(); ASSERT_IS_OK(computation_status.status()); auto computation = computation_status.ConsumeValueOrDie(); - auto f32_literal = Literal::CreateR0(1.1f); + auto f32_literal = LiteralUtil::CreateR0(1.1f); auto f32_data = client_->TransferToServer(*f32_literal).ConsumeValueOrDie(); - auto f32_4_literal = Literal::CreateR1({1.0f, 2.0f, 3.0f, 4.0f}); + auto f32_4_literal = LiteralUtil::CreateR1({1.0f, 2.0f, 3.0f, 4.0f}); auto f32_4_data = client_->TransferToServer(*f32_4_literal).ConsumeValueOrDie(); - auto u8_4_literal = Literal::CreateR1U8("hola"); + auto u8_4_literal = LiteralUtil::CreateR1U8("hola"); auto u8_4_data = client_->TransferToServer(*u8_4_literal).ConsumeValueOrDie(); // Match diff --git a/tensorflow/compiler/xla/tests/client_library_test_base.cc b/tensorflow/compiler/xla/tests/client_library_test_base.cc index bf8ed4d9fb0bc61b86ef0b5872711a122a3d416b..ef784da457be608de63eac478ef61c3df8627036 100644 --- a/tensorflow/compiler/xla/tests/client_library_test_base.cc +++ b/tensorflow/compiler/xla/tests/client_library_test_base.cc @@ -19,6 +19,7 @@ limitations under the License. #include "tensorflow/compiler/xla/client/client_library.h" #include "tensorflow/compiler/xla/client/local_client.h" +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/compiler/xla/execution_options_util.h" #include "tensorflow/compiler/xla/literal_util.h" #include "tensorflow/compiler/xla/ptr_util.h" @@ -156,7 +157,7 @@ string ClientLibraryTestBase::ExecuteToString( void ClientLibraryTestBase::ComputeAndCompareR1( XlaBuilder* builder, const tensorflow::core::Bitmap& expected, tensorflow::gtl::ArraySlice arguments) { - std::unique_ptr expected_literal = Literal::CreateR1(expected); + std::unique_ptr expected_literal = LiteralUtil::CreateR1(expected); ClientLibraryTestBase::ComputeAndCompareLiteral(builder, *expected_literal, arguments); } @@ -294,7 +295,7 @@ Status ClientLibraryTestBase::ComputeAndCompareLiteralWithStatus( std::unique_ptr converted_expected; Shape layout_shape; if (use_bfloat16_) { - converted_expected = Literal::ConvertF32ToBF16(expected); + converted_expected = LiteralUtil::ConvertF32ToBF16(expected); expected_ptr = converted_expected.get(); if (shape_with_layout != nullptr) { layout_shape = *shape_with_layout; @@ -346,7 +347,7 @@ Status ClientLibraryTestBase::ComputeAndCompareLiteralWithStatus( std::unique_ptr converted_expected; Shape layout_shape; if (use_bfloat16_) { - converted_expected = Literal::ConvertF32ToBF16(expected); + converted_expected = LiteralUtil::ConvertF32ToBF16(expected); expected_ptr = converted_expected.get(); if (shape_with_layout != nullptr) { layout_shape = *shape_with_layout; @@ -388,7 +389,7 @@ void ClientLibraryTestBase::ComputeAndCompareR1U8( auto actual = actual_status.ConsumeValueOrDie(); // Turn the expected value into a literal. - std::unique_ptr expected_literal = Literal::CreateR1U8(expected); + std::unique_ptr expected_literal = LiteralUtil::CreateR1U8(expected); VLOG(1) << "expected: " << expected_literal->ToString(); VLOG(1) << "actual: " << actual->ToString(); @@ -486,11 +487,11 @@ ClientLibraryTestBase::ComputeValueAndReference( XlaComputation ClientLibraryTestBase::CreateScalarRelu() { XlaBuilder builder("relu"); auto shape = ShapeUtil::MakeShape(use_bfloat16_ ? BF16 : F32, {}); - auto z_value = builder.Parameter(0, shape, "z_value"); + auto z_value = Parameter(&builder, 0, shape, "z_value"); auto zero = use_bfloat16_ - ? builder.ConstantR0(static_cast(0.0f)) - : builder.ConstantR0(0.0f); - builder.Max(z_value, zero); + ? ConstantR0(&builder, static_cast(0.0f)) + : ConstantR0(&builder, 0.0f); + Max(z_value, zero); auto computation_status = builder.Build(); TF_CHECK_OK(computation_status.status()); return computation_status.ConsumeValueOrDie(); @@ -499,9 +500,9 @@ XlaComputation ClientLibraryTestBase::CreateScalarRelu() { XlaComputation ClientLibraryTestBase::CreateScalarMax() { XlaBuilder builder("max"); auto shape = ShapeUtil::MakeShape(use_bfloat16_ ? BF16 : F32, {}); - auto x = builder.Parameter(0, shape, "x"); - auto y = builder.Parameter(1, shape, "y"); - builder.Max(x, y); + auto x = Parameter(&builder, 0, shape, "x"); + auto y = Parameter(&builder, 1, shape, "y"); + Max(x, y); auto computation_status = builder.Build(); TF_CHECK_OK(computation_status.status()); return computation_status.ConsumeValueOrDie(); @@ -510,13 +511,13 @@ XlaComputation ClientLibraryTestBase::CreateScalarMax() { XlaComputation ClientLibraryTestBase::CreateScalarReluSensitivity() { XlaBuilder builder("relu_sensitivity"); auto shape = ShapeUtil::MakeShape(use_bfloat16_ ? BF16 : F32, {}); - auto activation = builder.Parameter(0, shape, "activation"); - auto backprop = builder.Parameter(1, shape, "backprop"); + auto activation = Parameter(&builder, 0, shape, "activation"); + auto backprop = Parameter(&builder, 1, shape, "backprop"); auto zero = use_bfloat16_ - ? builder.ConstantR0(static_cast(0.0f)) - : builder.ConstantR0(0.0f); - auto activation_gtz = builder.Gt(activation, zero); - builder.Select(activation_gtz, /*on_true=*/backprop, /*on_false=*/zero); + ? ConstantR0(&builder, static_cast(0.0f)) + : ConstantR0(&builder, 0.0f); + auto activation_gtz = Gt(activation, zero); + Select(activation_gtz, /*on_true=*/backprop, /*on_false=*/zero); auto computation_status = builder.Build(); TF_CHECK_OK(computation_status.status()); @@ -559,8 +560,9 @@ XlaOp ClientLibraryTestBase::AddParam(const Literal& argument, XlaOp ClientLibraryTestBase::CreateConstantFromLiteral(const Literal& literal, XlaBuilder* builder) { - return builder->ConstantLiteral( - use_bfloat16_ ? *Literal::ConvertF32ToBF16(literal) : literal); + return ConstantLiteral(builder, use_bfloat16_ + ? *LiteralUtil::ConvertF32ToBF16(literal) + : literal); } std::unique_ptr @@ -581,14 +583,14 @@ ClientLibraryTestBase::CreateParameterAndTransferLiteral( const Literal* param_literal = &literal; std::unique_ptr converted_literal; if (use_bfloat16_) { - converted_literal = Literal::ConvertF32ToBF16(literal); + converted_literal = LiteralUtil::ConvertF32ToBF16(literal); param_literal = converted_literal.get(); } std::unique_ptr data = client_->TransferToServer(*param_literal, device_handle) .ConsumeValueOrDie(); *data_handle = - builder->Parameter(parameter_number, param_literal->shape(), name); + Parameter(builder, parameter_number, param_literal->shape(), name); return data; } diff --git a/tensorflow/compiler/xla/tests/client_library_test_base.h b/tensorflow/compiler/xla/tests/client_library_test_base.h index 0499fec5898a42affa0e0a712dee10187355c13e..fcc9347db519c2fe5b7804d381c7b4d14a85b8cc 100644 --- a/tensorflow/compiler/xla/tests/client_library_test_base.h +++ b/tensorflow/compiler/xla/tests/client_library_test_base.h @@ -28,6 +28,7 @@ limitations under the License. #include "tensorflow/compiler/xla/client/global_data.h" #include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/compiler/xla/client/xla_client/xla_computation.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/literal_util.h" #include "tensorflow/compiler/xla/ptr_util.h" #include "tensorflow/compiler/xla/statusor.h" @@ -284,7 +285,7 @@ class ClientLibraryTestBase : public ::testing::Test { template XlaOp AddParam(const Array& argument, XlaBuilder* builder) { - return AddParam(*Literal::CreateFromArray(argument), builder); + return AddParam(*LiteralUtil::CreateFromArray(argument), builder); } // Creates a constant instruction with the given literal. When the @@ -299,13 +300,14 @@ class ClientLibraryTestBase : public ::testing::Test { template XlaOp CreateConstantFromArray(const Array& array, XlaBuilder* builder) { - return CreateConstantFromLiteral(*Literal::CreateFromArray(array), builder); + return CreateConstantFromLiteral(*LiteralUtil::CreateFromArray(array), + builder); } // Same as CreateConstantFromArray, but for scalars. template XlaOp CreateConstantFromScalar(NativeT value, XlaBuilder* builder) { - return CreateConstantFromLiteral(*Literal::CreateR0(value), + return CreateConstantFromLiteral(*LiteralUtil::CreateR0(value), builder); } @@ -373,6 +375,13 @@ class ClientLibraryTestBase : public ::testing::Test { // The float type used in this test, BF16 or F32 according to use_bfloat16. PrimitiveType FloatType() const { return use_bfloat16_ ? BF16 : F32; } + // Executes the computation and calculates the expected reference value using + // the reference client. Returns two literals in the order of (expected, + // actual). + StatusOr, std::unique_ptr>> + ComputeValueAndReference(XlaBuilder* builder, + tensorflow::gtl::ArraySlice arguments); + Client* client_; Client* ref_client_; // To compute reference result. ExecutionOptions execution_options_; @@ -390,13 +399,6 @@ class ClientLibraryTestBase : public ::testing::Test { const string& error_message)>& verify_output, const Shape* output_with_layout = nullptr); - // Executes the computation and calculates the expected reference value using - // the reference client. Returns two literals in the order of (expected, - // actual). - StatusOr, std::unique_ptr>> - ComputeValueAndReference(XlaBuilder* builder, - tensorflow::gtl::ArraySlice arguments); - // Whether to run tests with all float-type input/output converted to // bfloat16. bool use_bfloat16_ = false; @@ -410,7 +412,7 @@ void ClientLibraryTestBase::ComputeAndCompareR0( XlaBuilder* builder, NativeT expected, tensorflow::gtl::ArraySlice arguments) { std::unique_ptr expected_literal = - Literal::CreateR0(expected); + LiteralUtil::CreateR0(expected); ClientLibraryTestBase::ComputeAndCompareLiteral(builder, *expected_literal, arguments); } @@ -426,7 +428,7 @@ void ClientLibraryTestBase::ComputeAndCompareR0( std::is_same::value, "Float or complex type required when specifying an ErrorSpec"); std::unique_ptr expected_literal = - Literal::CreateR0(expected); + LiteralUtil::CreateR0(expected); ClientLibraryTestBase::ComputeAndCompareLiteral(builder, *expected_literal, arguments, error); } @@ -436,7 +438,7 @@ void ClientLibraryTestBase::ComputeAndCompareR1( XlaBuilder* builder, tensorflow::gtl::ArraySlice expected, tensorflow::gtl::ArraySlice arguments) { std::unique_ptr expected_literal = - Literal::CreateR1(expected); + LiteralUtil::CreateR1(expected); ClientLibraryTestBase::ComputeAndCompareLiteral(builder, *expected_literal, arguments); } @@ -452,7 +454,7 @@ void ClientLibraryTestBase::ComputeAndCompareR1( std::is_same::value, "Float or complex type required when specifying an ErrorSpec"); std::unique_ptr expected_literal = - Literal::CreateR1(expected); + LiteralUtil::CreateR1(expected); ClientLibraryTestBase::ComputeAndCompareLiteral(builder, *expected_literal, arguments, error); } @@ -462,7 +464,7 @@ void ClientLibraryTestBase::ComputeAndCompareR2( XlaBuilder* builder, const Array2D& expected, tensorflow::gtl::ArraySlice arguments) { std::unique_ptr expected_literal = - Literal::CreateR2FromArray2D(expected); + LiteralUtil::CreateR2FromArray2D(expected); ClientLibraryTestBase::ComputeAndCompareLiteral(builder, *expected_literal, arguments); } @@ -478,7 +480,7 @@ void ClientLibraryTestBase::ComputeAndCompareR2( std::is_same::value, "Float or complex type required when specifying an ErrorSpec"); std::unique_ptr expected_literal = - Literal::CreateR2FromArray2D(expected); + LiteralUtil::CreateR2FromArray2D(expected); ClientLibraryTestBase::ComputeAndCompareLiteral(builder, *expected_literal, arguments, error); } @@ -488,7 +490,7 @@ void ClientLibraryTestBase::ComputeAndCompareR3( XlaBuilder* builder, const Array3D& expected, tensorflow::gtl::ArraySlice arguments) { std::unique_ptr expected_literal = - Literal::CreateR3FromArray3D(expected); + LiteralUtil::CreateR3FromArray3D(expected); ClientLibraryTestBase::ComputeAndCompareLiteral(builder, *expected_literal, arguments); } @@ -504,7 +506,7 @@ void ClientLibraryTestBase::ComputeAndCompareR3( std::is_same::value, "Float or complex type required when specifying an ErrorSpec"); std::unique_ptr expected_literal = - Literal::CreateR3FromArray3D(expected); + LiteralUtil::CreateR3FromArray3D(expected); ClientLibraryTestBase::ComputeAndCompareLiteral(builder, *expected_literal, arguments, error); } @@ -514,7 +516,7 @@ void ClientLibraryTestBase::ComputeAndCompareR4( XlaBuilder* builder, const Array4D& expected, tensorflow::gtl::ArraySlice arguments) { std::unique_ptr expected_literal = - Literal::CreateR4FromArray4D(expected); + LiteralUtil::CreateR4FromArray4D(expected); ClientLibraryTestBase::ComputeAndCompareLiteral(builder, *expected_literal, arguments); } @@ -530,7 +532,7 @@ void ClientLibraryTestBase::ComputeAndCompareR4( std::is_same::value, "Float or complex type required when specifying an ErrorSpec"); std::unique_ptr expected_literal = - Literal::CreateR4FromArray4D(expected); + LiteralUtil::CreateR4FromArray4D(expected); ClientLibraryTestBase::ComputeAndCompareLiteral(builder, *expected_literal, arguments, error); } @@ -539,13 +541,13 @@ template std::unique_ptr ClientLibraryTestBase::CreateR0Parameter( NativeT value, int64 parameter_number, const string& name, XlaBuilder* builder, XlaOp* data_handle) { - std::unique_ptr literal = Literal::CreateR0(value); + std::unique_ptr literal = LiteralUtil::CreateR0(value); if (use_bfloat16_ && literal->shape().element_type() == F32) { - literal = Literal::ConvertF32ToBF16(*literal); + literal = LiteralUtil::ConvertF32ToBF16(*literal); } std::unique_ptr data = client_->TransferToServer(*literal).ConsumeValueOrDie(); - *data_handle = builder->Parameter(parameter_number, literal->shape(), name); + *data_handle = Parameter(builder, parameter_number, literal->shape(), name); return data; } @@ -553,13 +555,13 @@ template std::unique_ptr ClientLibraryTestBase::CreateR1Parameter( tensorflow::gtl::ArraySlice values, int64 parameter_number, const string& name, XlaBuilder* builder, XlaOp* data_handle) { - std::unique_ptr literal = Literal::CreateR1(values); + std::unique_ptr literal = LiteralUtil::CreateR1(values); if (use_bfloat16_ && literal->shape().element_type() == F32) { - literal = Literal::ConvertF32ToBF16(*literal); + literal = LiteralUtil::ConvertF32ToBF16(*literal); } std::unique_ptr data = client_->TransferToServer(*literal).ConsumeValueOrDie(); - *data_handle = builder->Parameter(parameter_number, literal->shape(), name); + *data_handle = Parameter(builder, parameter_number, literal->shape(), name); return data; } @@ -567,13 +569,13 @@ template std::unique_ptr ClientLibraryTestBase::CreateR2Parameter( const Array2D& array_2d, int64 parameter_number, const string& name, XlaBuilder* builder, XlaOp* data_handle) { - std::unique_ptr literal = Literal::CreateR2FromArray2D(array_2d); + std::unique_ptr literal = LiteralUtil::CreateR2FromArray2D(array_2d); if (use_bfloat16_ && literal->shape().element_type() == F32) { - literal = Literal::ConvertF32ToBF16(*literal); + literal = LiteralUtil::ConvertF32ToBF16(*literal); } std::unique_ptr data = client_->TransferToServer(*literal).ConsumeValueOrDie(); - *data_handle = builder->Parameter(parameter_number, literal->shape(), name); + *data_handle = Parameter(builder, parameter_number, literal->shape(), name); return data; } @@ -581,13 +583,13 @@ template std::unique_ptr ClientLibraryTestBase::CreateR3Parameter( const Array3D& array_3d, int64 parameter_number, const string& name, XlaBuilder* builder, XlaOp* data_handle) { - std::unique_ptr literal = Literal::CreateR3FromArray3D(array_3d); + std::unique_ptr literal = LiteralUtil::CreateR3FromArray3D(array_3d); if (use_bfloat16_ && literal->shape().element_type() == F32) { - literal = Literal::ConvertF32ToBF16(*literal); + literal = LiteralUtil::ConvertF32ToBF16(*literal); } std::unique_ptr data = client_->TransferToServer(*literal).ConsumeValueOrDie(); - *data_handle = builder->Parameter(parameter_number, literal->shape(), name); + *data_handle = Parameter(builder, parameter_number, literal->shape(), name); return data; } diff --git a/tensorflow/compiler/xla/tests/client_test.cc b/tensorflow/compiler/xla/tests/client_test.cc index 08671cf62445826649b5c97003f998ae98a59d97..6ce2f844a34b01cd07df97a9bb12842490838be6 100644 --- a/tensorflow/compiler/xla/tests/client_test.cc +++ b/tensorflow/compiler/xla/tests/client_test.cc @@ -43,8 +43,8 @@ XLA_TEST_F(ClientTest, ExecuteWithLayout) { std::vector> layouts = {{0, 1}, {1, 0}}; for (const std::vector& execute_layout : layouts) { for (const std::vector& transfer_layout : layouts) { - b.Add(b.ConstantR2({{1, 2}, {3, 4}}), - b.ConstantR2({{10, 20}, {30, 40}})); + Add(ConstantR2(&b, {{1, 2}, {3, 4}}), + ConstantR2(&b, {{10, 20}, {30, 40}})); TF_ASSERT_OK_AND_ASSIGN(auto computation, b.Build()); ExecutionOptions execution_options = execution_options_; @@ -56,7 +56,7 @@ XLA_TEST_F(ClientTest, ExecuteWithLayout) { client_->Execute(computation, {}, &execution_options)); std::unique_ptr expected_literal = - Literal::CreateR2WithLayout( + LiteralUtil::CreateR2WithLayout( {{11, 22}, {33, 44}}, LayoutUtil::MakeLayout(transfer_layout)); TF_ASSERT_OK_AND_ASSIGN( @@ -72,8 +72,8 @@ XLA_TEST_F(ClientTest, ExecuteWithLayout) { XLA_TEST_F(ClientTest, ExecuteWithTupleLayout) { XlaBuilder b(TestName()); - b.Tuple({b.ConstantR2({{1, 2}, {3, 4}}), - b.ConstantR2({{10, 20}, {30, 40}})}); + Tuple(&b, {ConstantR2(&b, {{1, 2}, {3, 4}}), + ConstantR2(&b, {{10, 20}, {30, 40}})}); TF_ASSERT_OK_AND_ASSIGN(auto computation, b.Build()); @@ -112,13 +112,13 @@ XLA_TEST_F(ClientTest, DISABLED_ON_GPU(ExecuteParallel)) { XlaComputation add_with_one_arg, mul_with_two_args, dot_with_one_arg; Shape shape = ShapeUtil::MakeShape(S32, {2, 2}); - TF_ASSERT_OK_AND_ASSIGN( - std::unique_ptr const_arg, - client_->TransferToServer(*Literal::CreateR2({{5, 6}, {7, 8}}))); + TF_ASSERT_OK_AND_ASSIGN(std::unique_ptr const_arg, + client_->TransferToServer( + *LiteralUtil::CreateR2({{5, 6}, {7, 8}}))); XlaBuilder b(TestName() + ".add"); - b.Add(b.Parameter(0, shape, "param_0"), - b.ConstantR2({{1, 2}, {3, 4}})); + Add(Parameter(&b, 0, shape, "param_0"), + ConstantR2(&b, {{1, 2}, {3, 4}})); TF_ASSERT_OK_AND_ASSIGN(add_with_one_arg, b.Build()); // We can't really test parallel execution on CPU since all of the cores in a @@ -136,7 +136,7 @@ XLA_TEST_F(ClientTest, DISABLED_ON_GPU(ExecuteParallel)) { TF_ASSERT_OK_AND_ASSIGN(auto results, client_->ExecuteParallel(computation_instances)); - auto expected_result = Literal::CreateR2({{6, 8}, {10, 12}}); + auto expected_result = LiteralUtil::CreateR2({{6, 8}, {10, 12}}); TF_ASSERT_OK_AND_ASSIGN( auto result_literal, diff --git a/tensorflow/compiler/xla/tests/compilation_cache_test.cc b/tensorflow/compiler/xla/tests/compilation_cache_test.cc index 50a006964869b3e5dce431d441f7cd81af9df910..ff3824628676df8a37f3a98742d9423fd42928e4 100644 --- a/tensorflow/compiler/xla/tests/compilation_cache_test.cc +++ b/tensorflow/compiler/xla/tests/compilation_cache_test.cc @@ -21,7 +21,7 @@ limitations under the License. #include "tensorflow/compiler/xla/client/local_client.h" #include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/compiler/xla/client/xla_client/xla_computation.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/shape_util.h" #include "tensorflow/compiler/xla/statusor.h" #include "tensorflow/compiler/xla/tests/client_library_test_base.h" @@ -50,7 +50,7 @@ class CompilationCacheTest : public ClientLibraryTestBase { &execution_profile) .ConsumeValueOrDie(); EXPECT_TRUE(LiteralTestUtil::Near( - *Literal::CreateR0(expected_result), *result, error_spec_)); + *LiteralUtil::CreateR0(expected_result), *result, error_spec_)); EXPECT_EQ(expect_cache_hit, execution_profile.compilation_cache_hit()); } @@ -67,7 +67,7 @@ class CompilationCacheTest : public ClientLibraryTestBase { std::unique_ptr result = client_->Transfer(*data_handle).ConsumeValueOrDie(); EXPECT_TRUE(LiteralTestUtil::Near( - *Literal::CreateR2(expected_result), *result, error_spec_)); + *LiteralUtil::CreateR2(expected_result), *result, error_spec_)); EXPECT_EQ(expect_cache_hit, execution_profile.compilation_cache_hit()); } @@ -77,7 +77,7 @@ class CompilationCacheTest : public ClientLibraryTestBase { // TODO(b/74197823): Disabled because there is no cache in the new design. XLA_TEST_F(CompilationCacheTest, DISABLED_ComputationCalledMultipleTimes) { XlaBuilder builder(TestName()); - builder.Neg(builder.ConstantR0(42.0)); + Neg(ConstantR0(&builder, 42.0)); XlaComputation computation = builder.Build().ConsumeValueOrDie(); ExecuteComputationR0F32(computation, {}, -42.0, /*expect_cache_hit=*/false); @@ -89,17 +89,17 @@ XLA_TEST_F(CompilationCacheTest, DISABLED_ComputationCalledMultipleTimes) { XLA_TEST_F(CompilationCacheTest, DISABLED_ComputationCalledWithDifferentParameters) { std::unique_ptr data_42 = - client_->TransferToServer(*Literal::CreateR0(42.0f)) + client_->TransferToServer(*LiteralUtil::CreateR0(42.0f)) .ConsumeValueOrDie(); std::unique_ptr data_123 = - client_->TransferToServer(*Literal::CreateR0(123.0f)) + client_->TransferToServer(*LiteralUtil::CreateR0(123.0f)) .ConsumeValueOrDie(); std::unique_ptr data_456 = - client_->TransferToServer(*Literal::CreateR0(456.0f)) + client_->TransferToServer(*LiteralUtil::CreateR0(456.0f)) .ConsumeValueOrDie(); XlaBuilder builder(TestName()); - builder.Neg(builder.Parameter(0, ShapeUtil::MakeShape(F32, {}), "param")); + Neg(Parameter(&builder, 0, ShapeUtil::MakeShape(F32, {}), "param")); XlaComputation computation = builder.Build().ConsumeValueOrDie(); ExecuteComputationR0F32(computation, {data_42.get()}, -42.0, @@ -115,16 +115,16 @@ XLA_TEST_F(CompilationCacheTest, // TODO(b/74197823): Disabled because there is no cache in the new design. XLA_TEST_F(CompilationCacheTest, DISABLED_MultipleComputations) { XlaBuilder builder_neg(TestName() + "_neg"); - builder_neg.Neg(builder_neg.ConstantR0(42.0)); + Neg(ConstantR0(&builder_neg, 42.0)); XlaComputation computation_neg = builder_neg.Build().ConsumeValueOrDie(); XlaBuilder builder_exp(TestName() + "_exp"); - builder_exp.Exp(builder_exp.ConstantR0(1.0)); + Exp(ConstantR0(&builder_exp, 1.0)); XlaComputation computation_exp = builder_exp.Build().ConsumeValueOrDie(); XlaBuilder builder_add(TestName() + "_add"); - builder_add.Add(builder_add.ConstantR0(2.0), - builder_add.ConstantR0(3.0)); + Add(ConstantR0(&builder_add, 2.0), + ConstantR0(&builder_add, 3.0)); XlaComputation computation_add = builder_add.Build().ConsumeValueOrDie(); ExecuteComputationR0F32(computation_neg, {}, -42.0, @@ -143,18 +143,18 @@ XLA_TEST_F(CompilationCacheTest, DISABLED_DifferentParameterLayouts) { // layouts. Use these arrays as parameters to a simple computation. If the // layout of the array changes then computation should be recompiled (cache // miss). - auto rowmaj_array = Literal::CreateR2WithLayout( + auto rowmaj_array = LiteralUtil::CreateR2WithLayout( {{1.0f, 2.0f}, {3.0f, 4.0f}}, LayoutUtil::MakeLayout({1, 0})); auto rowmaj_handle = client_->TransferToServer(*rowmaj_array).ConsumeValueOrDie(); - auto colmaj_array = Literal::CreateR2WithLayout( + auto colmaj_array = LiteralUtil::CreateR2WithLayout( {{1.0f, 2.0f}, {3.0f, 4.0f}}, LayoutUtil::MakeLayout({0, 1})); auto colmaj_handle = client_->TransferToServer(*colmaj_array).ConsumeValueOrDie(); XlaBuilder builder(TestName()); - builder.Parameter(0, ShapeUtil::MakeShape(F32, {2, 2}), "param0"); + Parameter(&builder, 0, ShapeUtil::MakeShape(F32, {2, 2}), "param0"); XlaComputation computation = builder.Build().ConsumeValueOrDie(); ExecuteComputationR2F32(computation, {colmaj_handle.get()}, diff --git a/tensorflow/compiler/xla/tests/compute_constant_test.cc b/tensorflow/compiler/xla/tests/compute_constant_test.cc index ba22530f1cfee56337f862c25122d399dbf0f1e4..64bf8b3b3878f74e0557afc520c9ae342bd07c4a 100644 --- a/tensorflow/compiler/xla/tests/compute_constant_test.cc +++ b/tensorflow/compiler/xla/tests/compute_constant_test.cc @@ -22,7 +22,7 @@ limitations under the License. #include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/compiler/xla/client/xla_client/xla_computation.h" #include "tensorflow/compiler/xla/layout_util.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/shape_util.h" #include "tensorflow/compiler/xla/status_macros.h" #include "tensorflow/compiler/xla/statusor.h" @@ -99,7 +99,7 @@ TEST_F(ComputeConstantTest, ScalarInt32Literal) { for (ClientType client_type : client_types) { Client* client = ClientOrDie(platform_, client_type); XlaBuilder b(TestName()); - auto computation = b.ConstantR0(42); + auto computation = ConstantR0(&b, 42); EXPECT_TRUE(IsConstant(computation, &b)); auto value = ComputeConstantScalar(client, computation, &b); @@ -113,7 +113,7 @@ TEST_F(ComputeConstantTest, ScalarFloatAdd) { Client* client = ClientOrDie(platform_, client_type); XlaBuilder b(TestName()); auto computation = - b.Add(b.ConstantR0(42.5f), b.ConstantR0(1.5f)); + Add(ConstantR0(&b, 42.5f), ConstantR0(&b, 1.5f)); EXPECT_TRUE(IsConstant(computation, &b)); auto value = ComputeConstantScalar(client, computation, &b); @@ -127,8 +127,8 @@ TEST_F(ComputeConstantTest, ScalarRng) { Client* client = ClientOrDie(platform_, client_type); XlaBuilder b(TestName()); auto computation = - b.RngUniform(b.ConstantR0(1.1f), b.ConstantR0(2.1f), - ShapeUtil::MakeShape(F32, {})); + RngUniform(ConstantR0(&b, 1.1f), ConstantR0(&b, 2.1f), + ShapeUtil::MakeShape(F32, {})); EXPECT_FALSE(IsConstant(computation, &b)); auto value = ComputeConstantScalar(client, computation, &b); @@ -141,7 +141,7 @@ TEST_F(ComputeConstantTest, DirectParamMissing) { for (ClientType client_type : client_types) { Client* client = ClientOrDie(platform_, client_type); XlaBuilder b(TestName()); - auto computation = b.Parameter(0, ShapeUtil::MakeShape(F32, {}), "param"); + auto computation = Parameter(&b, 0, ShapeUtil::MakeShape(F32, {}), "param"); EXPECT_FALSE(IsConstant(computation, &b)); auto value = ComputeConstantScalar(client, computation, &b); @@ -156,8 +156,8 @@ TEST_F(ComputeConstantTest, IndirectParamMissing) { Client* client = ClientOrDie(platform_, client_type); XlaBuilder b(TestName()); auto computation = - b.Add(b.ConstantR0(1.0f), - b.Parameter(0, ShapeUtil::MakeShape(F32, {}), "param")); + Add(ConstantR0(&b, 1.0f), + Parameter(&b, 0, ShapeUtil::MakeShape(F32, {}), "param")); EXPECT_FALSE(IsConstant(computation, &b)); auto value = ComputeConstantScalar(client, computation, &b); @@ -174,18 +174,18 @@ TEST_F(ComputeConstantTest, UnrelatedParam) { Client* client = ClientOrDie(platform_, client_type); XlaBuilder b(TestName()); - auto param_a = b.Parameter(10, ShapeUtil::MakeShape(F32, {}), "param0"); + auto param_a = Parameter(&b, 10, ShapeUtil::MakeShape(F32, {}), "param0"); auto constant_4 = - b.Add(b.ConstantR0(2.5f), b.ConstantR0(1.5f)); - auto not_constant_a = b.Add(constant_4, param_a); + Add(ConstantR0(&b, 2.5f), ConstantR0(&b, 1.5f)); + auto not_constant_a = Add(constant_4, param_a); - auto param_b = b.Parameter(1, ShapeUtil::MakeShape(F32, {}), "param1"); + auto param_b = Parameter(&b, 1, ShapeUtil::MakeShape(F32, {}), "param1"); auto constant_9 = - b.Mul(b.ConstantR0(2.0f), b.ConstantR0(4.5f)); - auto not_constant_b = b.Add(param_b, constant_9); + Mul(ConstantR0(&b, 2.0f), ConstantR0(&b, 4.5f)); + auto not_constant_b = Add(param_b, constant_9); - auto constant_13 = b.Add(constant_4, constant_9); - b.Add(not_constant_b, b.Add(constant_13, not_constant_a)); + auto constant_13 = Add(constant_4, constant_9); + Add(not_constant_b, Add(constant_13, not_constant_a)); EXPECT_TRUE(IsConstant(constant_13, &b)); @@ -201,13 +201,13 @@ TEST_F(ComputeConstantTest, NonScalarAdd) { XlaBuilder b(TestName()); auto computation = - b.Add(b.ConstantR1({1, 2}), b.ConstantR1({3, 4})); + Add(ConstantR1(&b, {1, 2}), ConstantR1(&b, {3, 4})); EXPECT_TRUE(IsConstant(computation, &b)); TF_ASSERT_OK_AND_ASSIGN(auto computed, ComputeConstantLiteral(client, computation, &b)); std::unique_ptr expected_literal = - Literal::CreateR1({4, 6}); + LiteralUtil::CreateR1({4, 6}); EXPECT_TRUE(LiteralTestUtil::Equal(*expected_literal, *computed)); } } @@ -216,12 +216,12 @@ TEST_F(ComputeConstantTest, IntegerDivide) { for (ClientType client_type : client_types) { Client* client = ClientOrDie(platform_, client_type); XlaBuilder b(TestName()); - auto computation = b.Div(b.ConstantR0(15), b.ConstantR0(3)); + auto computation = Div(ConstantR0(&b, 15), ConstantR0(&b, 3)); EXPECT_TRUE(IsConstant(computation, &b)); TF_ASSERT_OK_AND_ASSIGN(auto computed, ComputeConstantLiteral(client, computation, &b)); - std::unique_ptr expected_literal = Literal::CreateR0(5); + std::unique_ptr expected_literal = LiteralUtil::CreateR0(5); EXPECT_TRUE(LiteralTestUtil::Equal(*expected_literal, *computed)); } } @@ -237,13 +237,13 @@ XLA_TEST_F(ComputeConstantTest, Layout) { TF_ASSERT_OK_AND_ASSIGN( auto computed, ComputeConstantLiteral( client, - b.Add(b.ConstantR2({{1, 2}, {3, 4}}), - b.ConstantR2({{10, 20}, {30, 40}})), + Add(ConstantR2(&b, {{1, 2}, {3, 4}}), + ConstantR2(&b, {{10, 20}, {30, 40}})), &b, &layout_proto)); std::unique_ptr expected_literal = - Literal::CreateR2WithLayout({{11, 22}, {33, 44}}, - LayoutUtil::MakeLayout(layout)); + LiteralUtil::CreateR2WithLayout( + {{11, 22}, {33, 44}}, LayoutUtil::MakeLayout(layout)); ASSERT_TRUE(LiteralTestUtil::EqualShapesAndLayouts( expected_literal->shape(), computed->shape())); EXPECT_TRUE(LiteralTestUtil::Equal(*expected_literal, *computed)); diff --git a/tensorflow/compiler/xla/tests/concat_test.cc b/tensorflow/compiler/xla/tests/concat_test.cc index 352864502a184237fde600330836fe471a5444f2..9f288634c0fa7d7dffa7f9c1af3a0752996cdec2 100644 --- a/tensorflow/compiler/xla/tests/concat_test.cc +++ b/tensorflow/compiler/xla/tests/concat_test.cc @@ -39,7 +39,7 @@ using ::testing::HasSubstr; // Concatenate expects at least one argument. XLA_TEST_F(ConcatTest, Concat_Nothing) { XlaBuilder builder(TestName()); - builder.ConcatInDim({}, 0); + ConcatInDim(&builder, {}, 0); StatusOr computation_status = builder.Build(); ASSERT_FALSE(computation_status.ok()); EXPECT_THAT(computation_status.status().ToString(), @@ -49,8 +49,8 @@ XLA_TEST_F(ConcatTest, Concat_Nothing) { // Concatenate with one argument works. XLA_TEST_F(ConcatTest, Concat_R1_With_Nothing) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({42.0, 64.0}); - builder.ConcatInDim({a}, 0); + auto a = ConstantR1(&builder, {42.0, 64.0}); + ConcatInDim(&builder, {a}, 0); std::vector expected = {42, 64}; ComputeAndCompareR1(&builder, expected, {}, ErrorSpec(0.0001)); @@ -58,8 +58,8 @@ XLA_TEST_F(ConcatTest, Concat_R1_With_Nothing) { XLA_TEST_F(ConcatTest, Concat_R1_L0_With_Nothing) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({}); - builder.ConcatInDim({a}, 0); + auto a = ConstantR1(&builder, {}); + ConcatInDim(&builder, {a}, 0); std::vector expected = {}; ComputeAndCompareR1(&builder, expected, {}, ErrorSpec(0.0001)); @@ -69,9 +69,9 @@ XLA_TEST_F(ConcatTest, Concat_R1_L0_With_Nothing) { // to concatenate on. XLA_TEST_F(ConcatTest, CannotConcatR0WithR0) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR0(42.0); - auto b = builder.ConstantR0(64.0); - builder.ConcatInDim({a, b}, 0); + auto a = ConstantR0(&builder, 42.0); + auto b = ConstantR0(&builder, 64.0); + ConcatInDim(&builder, {a, b}, 0); StatusOr computation_status = builder.Build(); ASSERT_FALSE(computation_status.ok()); EXPECT_THAT(computation_status.status().ToString(), @@ -80,9 +80,9 @@ XLA_TEST_F(ConcatTest, CannotConcatR0WithR0) { XLA_TEST_F(ConcatTest, Concat_R1_L0_With_R1_L0) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({}); - auto b = builder.ConstantR1({}); - builder.ConcatInDim({a, b}, 0); + auto a = ConstantR1(&builder, {}); + auto b = ConstantR1(&builder, {}); + ConcatInDim(&builder, {a, b}, 0); std::vector expected = {}; ComputeAndCompareR1(&builder, expected, {}, ErrorSpec(0.0001)); @@ -90,9 +90,9 @@ XLA_TEST_F(ConcatTest, Concat_R1_L0_With_R1_L0) { XLA_TEST_F(ConcatTest, Concat_R1_L0_With_R1_L1) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({}); - auto b = builder.ConstantR1({256.0}); - builder.ConcatInDim({a, b}, 0); + auto a = ConstantR1(&builder, {}); + auto b = ConstantR1(&builder, {256.0}); + ConcatInDim(&builder, {a, b}, 0); std::vector expected = {256}; ComputeAndCompareR1(&builder, expected, {}, ErrorSpec(0.0001)); @@ -100,9 +100,9 @@ XLA_TEST_F(ConcatTest, Concat_R1_L0_With_R1_L1) { XLA_TEST_F(ConcatTest, Concat_R1_L2_With_R1_L0) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({42.0, 64.0}); - auto b = builder.ConstantR1({}); - builder.ConcatInDim({a, b}, 0); + auto a = ConstantR1(&builder, {42.0, 64.0}); + auto b = ConstantR1(&builder, {}); + ConcatInDim(&builder, {a, b}, 0); std::vector expected = {42, 64}; ComputeAndCompareR1(&builder, expected, {}, ErrorSpec(0.0001)); @@ -110,9 +110,9 @@ XLA_TEST_F(ConcatTest, Concat_R1_L2_With_R1_L0) { XLA_TEST_F(ConcatTest, Concat_R1_L2_With_R1_L1) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({42.0, 64.0}); - auto b = builder.ConstantR1({256.0}); - builder.ConcatInDim({a, b}, 0); + auto a = ConstantR1(&builder, {42.0, 64.0}); + auto b = ConstantR1(&builder, {256.0}); + ConcatInDim(&builder, {a, b}, 0); std::vector expected = {42, 64, 256}; ComputeAndCompareR1(&builder, expected, {}, ErrorSpec(0.0001)); @@ -130,9 +130,9 @@ XLA_TEST_F(ConcatTest, Concat_R1_L253_With_R1_L7) { } XlaBuilder builder(TestName()); - auto a = builder.ConstantR1(lhs); - auto b = builder.ConstantR1(rhs); - builder.ConcatInDim({a, b}, 0); + auto a = ConstantR1(&builder, lhs); + auto b = ConstantR1(&builder, rhs); + ConcatInDim(&builder, {a, b}, 0); ComputeAndCompareR1(&builder, expected, {}, ErrorSpec(0.0001)); } @@ -140,9 +140,9 @@ XLA_TEST_F(ConcatTest, Concat_R1_L253_With_R1_L7) { XLA_TEST_F(ConcatTest, Concat_0x0_With_0x0) { for (int dim : {0, 1}) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR2FromArray2D(Array2D(0, 0)); - auto b = builder.ConstantR2FromArray2D(Array2D(0, 0)); - builder.ConcatInDim({a, b}, dim); + auto a = ConstantR2FromArray2D(&builder, Array2D(0, 0)); + auto b = ConstantR2FromArray2D(&builder, Array2D(0, 0)); + ConcatInDim(&builder, {a, b}, dim); ComputeAndCompareR2(&builder, Array2D(0, 0), {}, ErrorSpec(0.0001)); @@ -153,9 +153,9 @@ XLA_TEST_F(ConcatTest, Concat_1x1_With_1x1_InDim0) { XlaBuilder builder(TestName()); auto a_array = CreatePatternedMatrix(1, 1); auto b_array = CreatePatternedMatrix(1, 1, /*offset=*/64.0); - auto a = builder.ConstantR2FromArray2D(*a_array); - auto b = builder.ConstantR2FromArray2D(*b_array); - builder.ConcatInDim({a, b}, 0); + auto a = ConstantR2FromArray2D(&builder, *a_array); + auto b = ConstantR2FromArray2D(&builder, *b_array); + ConcatInDim(&builder, {a, b}, 0); Array2D expected({ {0}, @@ -168,9 +168,9 @@ XLA_TEST_F(ConcatTest, Concat_1x1_With_1x1_InDim1) { XlaBuilder builder(TestName()); auto a_array = CreatePatternedMatrix(1, 1); auto b_array = CreatePatternedMatrix(1, 1, /*offset=*/64.0); - auto a = builder.ConstantR2FromArray2D(*a_array); - auto b = builder.ConstantR2FromArray2D(*b_array); - builder.ConcatInDim({a, b}, 1); + auto a = ConstantR2FromArray2D(&builder, *a_array); + auto b = ConstantR2FromArray2D(&builder, *b_array); + ConcatInDim(&builder, {a, b}, 1); Array2D expected({ {0, 64}, @@ -181,9 +181,9 @@ XLA_TEST_F(ConcatTest, Concat_1x1_With_1x1_InDim1) { XLA_TEST_F(ConcatTest, Concat2x0With2x5) { XlaBuilder builder(TestName()); auto b_array = CreatePatternedMatrix(2, 5, /*offset=*/64.0); - auto a = builder.ConstantR2FromArray2D(Array2D(2, 0)); - auto b = builder.ConstantR2FromArray2D(*b_array); - builder.ConcatInDim({a, b}, 1); + auto a = ConstantR2FromArray2D(&builder, Array2D(2, 0)); + auto b = ConstantR2FromArray2D(&builder, *b_array); + ConcatInDim(&builder, {a, b}, 1); ComputeAndCompareR2(&builder, *b_array, {}, ErrorSpec(0.0001)); } @@ -192,9 +192,9 @@ XLA_TEST_F(ConcatTest, Concat2x3With2x5) { XlaBuilder builder(TestName()); auto a_array = CreatePatternedMatrix(2, 3); auto b_array = CreatePatternedMatrix(2, 5, /*offset=*/64.0); - auto a = builder.ConstantR2FromArray2D(*a_array); - auto b = builder.ConstantR2FromArray2D(*b_array); - builder.ConcatInDim({a, b}, 1); + auto a = ConstantR2FromArray2D(&builder, *a_array); + auto b = ConstantR2FromArray2D(&builder, *b_array); + ConcatInDim(&builder, {a, b}, 1); Array2D expected({ {0, 1, 2, 64, 65, 66, 67, 68}, @@ -206,9 +206,9 @@ XLA_TEST_F(ConcatTest, Concat2x3With2x5) { XLA_TEST_F(ConcatTest, Concat3x2With0x2) { XlaBuilder builder(TestName()); auto a_array = CreatePatternedMatrix(3, 2); - auto a = builder.ConstantR2FromArray2D(*a_array); - auto b = builder.ConstantR2FromArray2D(Array2D(0, 2)); - builder.ConcatInDim({a, b}, 0); + auto a = ConstantR2FromArray2D(&builder, *a_array); + auto b = ConstantR2FromArray2D(&builder, Array2D(0, 2)); + ConcatInDim(&builder, {a, b}, 0); ComputeAndCompareR2(&builder, *a_array, {}, ErrorSpec(0.0001)); } @@ -217,9 +217,9 @@ XLA_TEST_F(ConcatTest, Concat3x2With5x2) { XlaBuilder builder(TestName()); auto a_array = CreatePatternedMatrix(3, 2); auto b_array = CreatePatternedMatrix(5, 2, /*offset=*/64.0); - auto a = builder.ConstantR2FromArray2D(*a_array); - auto b = builder.ConstantR2FromArray2D(*b_array); - builder.ConcatInDim({a, b}, 0); + auto a = ConstantR2FromArray2D(&builder, *a_array); + auto b = ConstantR2FromArray2D(&builder, *b_array); + ConcatInDim(&builder, {a, b}, 0); Array2D expected({ {0, 1}, @@ -236,9 +236,9 @@ XLA_TEST_F(ConcatTest, Concat3x2With5x2) { XLA_TEST_F(ConcatTest, Concat_R3_3x0x2_3x0x1) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR3FromArray3D(Array3D(3, 0, 2)); - auto b = builder.ConstantR3FromArray3D(Array3D(3, 0, 1)); - builder.ConcatInDim({a, b}, 2); + auto a = ConstantR3FromArray3D(&builder, Array3D(3, 0, 2)); + auto b = ConstantR3FromArray3D(&builder, Array3D(3, 0, 1)); + ConcatInDim(&builder, {a, b}, 2); ComputeAndCompareR3(&builder, Array3D(3, 0, 3), {}, ErrorSpec(0.0001)); } @@ -257,9 +257,9 @@ XLA_TEST_F(ConcatTest, Concat_R3_3x1x2_3x1x1) { {{7}}, {{8}}, }); - auto a = builder.ConstantR3FromArray3D(a_array); - auto b = builder.ConstantR3FromArray3D(b_array); - builder.ConcatInDim({a, b}, 2); + auto a = ConstantR3FromArray3D(&builder, a_array); + auto b = ConstantR3FromArray3D(&builder, b_array); + ConcatInDim(&builder, {a, b}, 2); Array3D expected({ {{0, 1, 6}}, @@ -271,10 +271,10 @@ XLA_TEST_F(ConcatTest, Concat_R3_3x1x2_3x1x1) { XLA_TEST_F(ConcatTest, Concat_R1_1x1_1x1_1x1) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({42.0}); - auto b = builder.ConstantR1({64.0}); - auto c = builder.ConstantR1({256.0}); - builder.ConcatInDim({a, b, c}, 0); + auto a = ConstantR1(&builder, {42.0}); + auto b = ConstantR1(&builder, {64.0}); + auto c = ConstantR1(&builder, {256.0}); + ConcatInDim(&builder, {a, b, c}, 0); std::vector expected = {42, 64, 256}; ComputeAndCompareR1(&builder, expected, {}, ErrorSpec(0.0001)); @@ -300,10 +300,10 @@ XLA_TEST_F(ConcatTest, Concat_R3_3x1x2_3x1x1_3x1x1) { {{7}}, {{11}}, }); - auto a = builder.ConstantR3FromArray3D(a_array); - auto b = builder.ConstantR3FromArray3D(b_array); - auto c = builder.ConstantR3FromArray3D(c_array); - builder.ConcatInDim({a, b, c}, 2); + auto a = ConstantR3FromArray3D(&builder, a_array); + auto b = ConstantR3FromArray3D(&builder, b_array); + auto c = ConstantR3FromArray3D(&builder, c_array); + ConcatInDim(&builder, {a, b, c}, 2); Array3D expected({ {{0, 1, 2, 3}}, @@ -315,11 +315,11 @@ XLA_TEST_F(ConcatTest, Concat_R3_3x1x2_3x1x1_3x1x1) { XLA_TEST_F(ConcatTest, DoubleConcatLeftAssociative) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({42.0}); - auto b = builder.ConstantR1({64.0}); - auto c = builder.ConstantR1({256.0}); + auto a = ConstantR1(&builder, {42.0}); + auto b = ConstantR1(&builder, {64.0}); + auto c = ConstantR1(&builder, {256.0}); // concatenated = (a concat b) concat c - builder.ConcatInDim({builder.ConcatInDim({a, b}, 0), c}, 0); + ConcatInDim(&builder, {ConcatInDim(&builder, {a, b}, 0), c}, 0); std::vector expected = {42, 64, 256}; ComputeAndCompareR1(&builder, expected, {}, ErrorSpec(0.0001)); @@ -327,11 +327,11 @@ XLA_TEST_F(ConcatTest, DoubleConcatLeftAssociative) { XLA_TEST_F(ConcatTest, DoubleConcatRightAssociative) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({42.0}); - auto b = builder.ConstantR1({64.0}); - auto c = builder.ConstantR1({256.0}); + auto a = ConstantR1(&builder, {42.0}); + auto b = ConstantR1(&builder, {64.0}); + auto c = ConstantR1(&builder, {256.0}); // concatenated = a concat (b concat c) - builder.ConcatInDim({a, builder.ConcatInDim({b, c}, 0)}, 0); + ConcatInDim(&builder, {a, ConcatInDim(&builder, {b, c}, 0)}, 0); std::vector expected = {42, 64, 256}; ComputeAndCompareR1(&builder, expected, {}, ErrorSpec(0.0001)); @@ -346,9 +346,9 @@ XLA_TEST_F(ConcatTest, Concat_1x1024_With_1x1024_InDim0) { } XlaBuilder builder(TestName()); - auto a = builder.ConstantR2FromArray2D(lhs); - auto b = builder.ConstantR2FromArray2D(rhs); - builder.ConcatInDim({a, b}, 0); + auto a = ConstantR2FromArray2D(&builder, lhs); + auto b = ConstantR2FromArray2D(&builder, rhs); + ConcatInDim(&builder, {a, b}, 0); Array2D expected(2, 1024); for (int i = 0; i < 1024; ++i) { @@ -367,9 +367,9 @@ XLA_TEST_F(ConcatTest, Concat_1x1024_With_1x1024_InDim1) { } XlaBuilder builder(TestName()); - auto a = builder.ConstantR2FromArray2D(lhs); - auto b = builder.ConstantR2FromArray2D(rhs); - builder.ConcatInDim({a, b}, 1); + auto a = ConstantR2FromArray2D(&builder, lhs); + auto b = ConstantR2FromArray2D(&builder, rhs); + ConcatInDim(&builder, {a, b}, 1); Array2D expected(1, 2048); for (int i = 0; i < 1024; ++i) { @@ -392,9 +392,9 @@ XLA_TEST_F(ConcatTest, Concat_64x64_With_64x2) { } XlaBuilder builder(TestName()); - auto a = builder.ConstantR2FromArray2D(lhs); - auto b = builder.ConstantR2FromArray2D(rhs); - builder.ConcatInDim({a, b}, 1); + auto a = ConstantR2FromArray2D(&builder, lhs); + auto b = ConstantR2FromArray2D(&builder, rhs); + ConcatInDim(&builder, {a, b}, 1); Array2D expected(64, 66); for (int i0 = 0; i0 < 64; ++i0) { @@ -410,9 +410,9 @@ XLA_TEST_F(ConcatTest, CannotConcatOpaques) { XlaBuilder builder(TestName()); auto opaque_shape = ShapeUtil::MakeOpaqueShape(); auto r1f32 = xla::ShapeUtil::MakeShape(xla::F32, {1}); - auto x = builder.Parameter(0, r1f32, "x"); - auto y = builder.Parameter(1, opaque_shape, "y"); - builder.ConcatInDim({x, y}, 0); + auto x = Parameter(&builder, 0, r1f32, "x"); + auto y = Parameter(&builder, 1, opaque_shape, "y"); + ConcatInDim(&builder, {x, y}, 0); StatusOr computation_status = builder.Build(); ASSERT_FALSE(computation_status.ok()); EXPECT_THAT( @@ -425,9 +425,9 @@ XLA_TEST_F(ConcatTest, CannotConcatTokens) { XlaBuilder builder(TestName()); auto token_shape = ShapeUtil::MakeTokenShape(); auto r1f32 = xla::ShapeUtil::MakeShape(xla::F32, {1}); - auto x = builder.Parameter(0, r1f32, "x"); - auto y = builder.Parameter(1, token_shape, "y"); - builder.ConcatInDim({x, y}, 0); + auto x = Parameter(&builder, 0, r1f32, "x"); + auto y = Parameter(&builder, 1, token_shape, "y"); + ConcatInDim(&builder, {x, y}, 0); StatusOr computation_status = builder.Build(); ASSERT_FALSE(computation_status.ok()); EXPECT_THAT( @@ -437,10 +437,10 @@ XLA_TEST_F(ConcatTest, CannotConcatTokens) { XLA_TEST_F(ConcatTest, ConcatSeveralBoxedPredicates) { XlaBuilder builder(TestName()); - auto p0 = builder.ConstantR1({true}); - auto p1 = builder.ConstantR1({false}); - auto p2 = builder.ConstantR1({true}); - builder.ConcatInDim({p0, p1, p2}, 0); + auto p0 = ConstantR1(&builder, {true}); + auto p1 = ConstantR1(&builder, {false}); + auto p2 = ConstantR1(&builder, {true}); + ConcatInDim(&builder, {p0, p1, p2}, 0); bool expected[] = {true, false, true}; ComputeAndCompareR1(&builder, expected, {}); @@ -448,11 +448,11 @@ XLA_TEST_F(ConcatTest, ConcatSeveralBoxedPredicates) { XLA_TEST_F(ConcatTest, ConcatSeveralR1S32s) { XlaBuilder builder(TestName()); - auto a0 = builder.ConstantR1({1}); - auto a1 = builder.ConstantR1({2, 3}); - auto a2 = builder.ConstantR1({4, 5, 6}); - auto a3 = builder.ConstantR1({7, 8, 9, 10}); - builder.ConcatInDim({a0, a1, a2, a3}, 0); + auto a0 = ConstantR1(&builder, {1}); + auto a1 = ConstantR1(&builder, {2, 3}); + auto a2 = ConstantR1(&builder, {4, 5, 6}); + auto a3 = ConstantR1(&builder, {7, 8, 9, 10}); + ConcatInDim(&builder, {a0, a1, a2, a3}, 0); std::vector expected(10); std::iota(expected.begin(), expected.end(), 1); @@ -487,7 +487,7 @@ XLA_TEST_F(ConcatTest, ConcatR3WeirdDims) { auto p1 = CreateR3Parameter(arr1, /*parameter_number=*/1, "p1", &builder, &h1); - builder.ConcatInDim({h0, h1}, 2); + ConcatInDim(&builder, {h0, h1}, 2); ComputeAndCompareR3(&builder, expected, {p0.get(), p1.get()}); } @@ -514,9 +514,9 @@ TEST_P(ConcatR2BinaryTest, DoIt) { rhs.FillUnique(1000); XlaBuilder builder(TestName()); - auto a0 = builder.ConstantR2FromArray2D(lhs); - auto a1 = builder.ConstantR2FromArray2D(rhs); - builder.ConcatInDim({a0, a1}, spec.concat_dimension); + auto a0 = ConstantR2FromArray2D(&builder, lhs); + auto a1 = ConstantR2FromArray2D(&builder, rhs); + ConcatInDim(&builder, {a0, a1}, spec.concat_dimension); std::unique_ptr> expected = ReferenceUtil::Concat2D(lhs, rhs, spec.concat_dimension); @@ -534,19 +534,19 @@ TEST_P(ConcatR2BinaryTest, DoIt) { // concat XLA_TEST_F(ConcatTest, ConcatOperandsOfSameOperand) { auto f32_scalar = ShapeUtil::MakeShape(xla::F32, {}); - auto x_literal = Literal::CreateR0(2.f); - auto y_literal = Literal::CreateR0(3.f); + auto x_literal = LiteralUtil::CreateR0(2.f); + auto y_literal = LiteralUtil::CreateR0(3.f); auto x_data = client_->TransferToServer(*x_literal).ConsumeValueOrDie(); auto y_data = client_->TransferToServer(*y_literal).ConsumeValueOrDie(); XlaBuilder builder(TestName()); - auto x = builder.Parameter(0, f32_scalar, "x"); - auto y = builder.Parameter(1, f32_scalar, "y"); - auto mul = builder.Mul(x, y); - auto add1 = builder.Add(mul, builder.ConstantR1({1.f, 2.f})); - auto add2 = builder.Add(mul, builder.ConstantR1({3.f, 4.f})); - auto add3 = builder.Add(mul, builder.ConstantR1({5.f, 6.f})); - builder.ConcatInDim({add1, add2, add3}, /*dimension=*/0); + auto x = Parameter(&builder, 0, f32_scalar, "x"); + auto y = Parameter(&builder, 1, f32_scalar, "y"); + auto mul = Mul(x, y); + auto add1 = Add(mul, ConstantR1(&builder, {1.f, 2.f})); + auto add2 = Add(mul, ConstantR1(&builder, {3.f, 4.f})); + auto add3 = Add(mul, ConstantR1(&builder, {5.f, 6.f})); + ConcatInDim(&builder, {add1, add2, add3}, /*dimension=*/0); ComputeAndCompareR1(&builder, {7., 8., 9., 10., 11., 12.}, {x_data.get(), y_data.get()}, ErrorSpec(1e-4)); @@ -556,21 +556,21 @@ XLA_TEST_F(ConcatTest, ConcatOperandsOfSameOperand) { // produces the correct result in rank 1. XLA_TEST_F(ConcatTest, ConcatBroadcastArgument) { auto f32_scalar = ShapeUtil::MakeShape(xla::F32, {}); - auto x_literal = Literal::CreateR1({2.0f, 3.0f, 5.0f, 6.0f}); - auto y_literal = Literal::CreateR0(1.5f); - auto z_literal = Literal::CreateR0(5.5f); + auto x_literal = LiteralUtil::CreateR1({2.0f, 3.0f, 5.0f, 6.0f}); + auto y_literal = LiteralUtil::CreateR0(1.5f); + auto z_literal = LiteralUtil::CreateR0(5.5f); auto x_data = client_->TransferToServer(*x_literal).ConsumeValueOrDie(); auto y_data = client_->TransferToServer(*y_literal).ConsumeValueOrDie(); auto z_data = client_->TransferToServer(*z_literal).ConsumeValueOrDie(); XlaBuilder builder(TestName()); - auto x = builder.Parameter(0, x_literal->shape(), "x"); - auto y = builder.Parameter(1, f32_scalar, "y"); - auto z = builder.Parameter(2, f32_scalar, "z"); - auto bcast = builder.Broadcast(y, {5}); - auto bcast2 = builder.Broadcast(z, {3}); - auto concat = builder.ConcatInDim({bcast, x}, /*dimension=*/0); - builder.ConcatInDim({concat, bcast2}, /*dimension=*/0); + auto x = Parameter(&builder, 0, x_literal->shape(), "x"); + auto y = Parameter(&builder, 1, f32_scalar, "y"); + auto z = Parameter(&builder, 2, f32_scalar, "z"); + auto bcast = Broadcast(y, {5}); + auto bcast2 = Broadcast(z, {3}); + auto concat = ConcatInDim(&builder, {bcast, x}, /*dimension=*/0); + ConcatInDim(&builder, {concat, bcast2}, /*dimension=*/0); ComputeAndCompareR1( &builder, @@ -584,21 +584,21 @@ XLA_TEST_F(ConcatTest, ConcatBroadcastArgument) { XLA_TEST_F(ConcatTest, ConcatBroadcastArgumentR3) { auto f32_scalar = ShapeUtil::MakeShape(xla::F32, {}); Array3D x3d(3, 5, 7, 3.14f); - auto x_literal = Literal::CreateR3FromArray3D(x3d); - auto y_literal = Literal::CreateR0(1.5f); - auto z_literal = Literal::CreateR0(5.5f); + auto x_literal = LiteralUtil::CreateR3FromArray3D(x3d); + auto y_literal = LiteralUtil::CreateR0(1.5f); + auto z_literal = LiteralUtil::CreateR0(5.5f); auto x_data = client_->TransferToServer(*x_literal).ConsumeValueOrDie(); auto y_data = client_->TransferToServer(*y_literal).ConsumeValueOrDie(); auto z_data = client_->TransferToServer(*z_literal).ConsumeValueOrDie(); XlaBuilder builder(TestName()); - auto x = builder.Parameter(0, x_literal->shape(), "x"); - auto y = builder.Parameter(1, f32_scalar, "y"); - auto z = builder.Parameter(2, f32_scalar, "y"); - auto y_bcast = builder.Broadcast(y, {1, 5, 7}); - auto z_bcast = builder.Broadcast(z, {4, 1, 7}); - auto concat = builder.ConcatInDim({y_bcast, x}, /*dimension=*/0); - builder.ConcatInDim({concat, z_bcast}, /*dimension=*/1); + auto x = Parameter(&builder, 0, x_literal->shape(), "x"); + auto y = Parameter(&builder, 1, f32_scalar, "y"); + auto z = Parameter(&builder, 2, f32_scalar, "y"); + auto y_bcast = Broadcast(y, {1, 5, 7}); + auto z_bcast = Broadcast(z, {4, 1, 7}); + auto concat = ConcatInDim(&builder, {y_bcast, x}, /*dimension=*/0); + ConcatInDim(&builder, {concat, z_bcast}, /*dimension=*/1); Array3D y_bcast3d(1, 5, 7, 1.5f); Array3D z_bcast3d(4, 1, 7, 5.5f); auto concat0 = ReferenceUtil::Concat3D(y_bcast3d, x3d, 0); diff --git a/tensorflow/compiler/xla/tests/conditional_test.cc b/tensorflow/compiler/xla/tests/conditional_test.cc index 7ff6706935740c7d76ee5cd03eae292386760397..35f1400fb2a6494f0c8b7f92a782b5936fe558b6 100644 --- a/tensorflow/compiler/xla/tests/conditional_test.cc +++ b/tensorflow/compiler/xla/tests/conditional_test.cc @@ -26,8 +26,8 @@ class ConditionalOpTest : public ClientLibraryTestBase { protected: XlaComputation CreateR0ConstantComputation(float value) { XlaBuilder builder("Constant"); - builder.Parameter(0, empty_tuple_, "tuple"); - builder.ConstantR0(value); + Parameter(&builder, 0, empty_tuple_, "tuple"); + ConstantR0(&builder, value); auto build_status = builder.Build(); EXPECT_IS_OK(build_status.status()); return build_status.ConsumeValueOrDie(); @@ -35,7 +35,7 @@ class ConditionalOpTest : public ClientLibraryTestBase { XlaComputation CreateR0IdentityComputation() { XlaBuilder builder("Identity"); - builder.Parameter(0, r0f32_, "x"); + Parameter(&builder, 0, r0f32_, "x"); auto build_status = builder.Build(); EXPECT_IS_OK(build_status.status()); return build_status.ConsumeValueOrDie(); @@ -43,8 +43,8 @@ class ConditionalOpTest : public ClientLibraryTestBase { XlaComputation CreateCeilComputation(const Shape& shape) { XlaBuilder builder("Ceil"); - auto param = builder.Parameter(0, shape, "param"); - builder.Ceil(param); + auto param = Parameter(&builder, 0, shape, "param"); + Ceil(param); auto build_status = builder.Build(); EXPECT_IS_OK(build_status.status()); return build_status.ConsumeValueOrDie(); @@ -60,8 +60,8 @@ class ConditionalOpTest : public ClientLibraryTestBase { XlaComputation CreateFloorComputation(const Shape& shape) { XlaBuilder builder("Floor"); - auto param = builder.Parameter(0, shape, "param"); - builder.Floor(param); + auto param = Parameter(&builder, 0, shape, "param"); + Floor(param); auto build_status = builder.Build(); EXPECT_IS_OK(build_status.status()); return build_status.ConsumeValueOrDie(); @@ -78,12 +78,12 @@ class ConditionalOpTest : public ClientLibraryTestBase { XlaComputation CreateTupleCeilComputation(const string& computation_name, const Shape& tuple_shape) { XlaBuilder builder(computation_name); - auto tuple = builder.Parameter(0, tuple_shape, "tuple"); - auto x = builder.GetTupleElement(tuple, 0); - auto y = builder.GetTupleElement(tuple, 1); - auto x_ceil = builder.Ceil(x); - auto y_ceil = builder.Ceil(y); - builder.Tuple({x_ceil, y_ceil}); + auto tuple = Parameter(&builder, 0, tuple_shape, "tuple"); + auto x = GetTupleElement(tuple, 0); + auto y = GetTupleElement(tuple, 1); + auto x_ceil = Ceil(x); + auto y_ceil = Ceil(y); + Tuple(&builder, {x_ceil, y_ceil}); auto build_status = builder.Build(); EXPECT_IS_OK(build_status.status()); return build_status.ConsumeValueOrDie(); @@ -100,12 +100,12 @@ class ConditionalOpTest : public ClientLibraryTestBase { XlaComputation CreateTupleFloorComputation(const string& computation_name, const Shape& tuple_shape) { XlaBuilder builder(computation_name); - auto tuple = builder.Parameter(0, tuple_shape, "tuple"); - auto x = builder.GetTupleElement(tuple, 0); - auto y = builder.GetTupleElement(tuple, 1); - auto x_floor = builder.Floor(x); - auto y_floor = builder.Floor(y); - builder.Tuple({x_floor, y_floor}); + auto tuple = Parameter(&builder, 0, tuple_shape, "tuple"); + auto x = GetTupleElement(tuple, 0); + auto y = GetTupleElement(tuple, 1); + auto x_floor = Floor(x); + auto y_floor = Floor(y); + Tuple(&builder, {x_floor, y_floor}); auto build_status = builder.Build(); EXPECT_IS_OK(build_status.status()); return build_status.ConsumeValueOrDie(); @@ -122,10 +122,10 @@ class ConditionalOpTest : public ClientLibraryTestBase { XlaComputation CreateTupleAddComputation(const string& computation_name, const Shape& tuple_shape) { XlaBuilder builder(computation_name); - auto tuple = builder.Parameter(0, tuple_shape, "tuple"); - auto x = builder.GetTupleElement(tuple, 0); - auto y = builder.GetTupleElement(tuple, 1); - builder.Add(x, y); + auto tuple = Parameter(&builder, 0, tuple_shape, "tuple"); + auto x = GetTupleElement(tuple, 0); + auto y = GetTupleElement(tuple, 1); + Add(x, y); auto build_status = builder.Build(); EXPECT_IS_OK(build_status.status()); return build_status.ConsumeValueOrDie(); @@ -142,10 +142,10 @@ class ConditionalOpTest : public ClientLibraryTestBase { XlaComputation CreateTupleSubComputation(const string& computation_name, const Shape& tuple_shape) { XlaBuilder builder(computation_name); - auto tuple = builder.Parameter(0, tuple_shape, "tuple"); - auto x = builder.GetTupleElement(tuple, 0); - auto y = builder.GetTupleElement(tuple, 1); - builder.Sub(x, y); + auto tuple = Parameter(&builder, 0, tuple_shape, "tuple"); + auto x = GetTupleElement(tuple, 0); + auto y = GetTupleElement(tuple, 1); + Sub(x, y); auto build_status = builder.Build(); EXPECT_IS_OK(build_status.status()); return build_status.ConsumeValueOrDie(); @@ -172,12 +172,11 @@ class ConditionalOpTest : public ClientLibraryTestBase { // Test true and false computations that do not take any parameters. XLA_TEST_F(ConditionalOpTest, Parameters0) { XlaBuilder builder(TestName()); - auto pred = builder.ConstantR0(true); - auto operands = builder.Tuple({}); + auto pred = ConstantR0(&builder, true); + auto operands = Tuple(&builder, {}); auto true_computation = CreateR0ConstantComputation(56.0f); auto false_computation = CreateR0ConstantComputation(12.0f); - builder.Conditional(pred, operands, true_computation, operands, - false_computation); + Conditional(pred, operands, true_computation, operands, false_computation); ComputeAndCompareR0(&builder, 56.0f, {}, error_spec_); } @@ -185,11 +184,11 @@ XLA_TEST_F(ConditionalOpTest, Parameters0) { // Test true and false computations that take in 1 parameter. XLA_TEST_F(ConditionalOpTest, Parameters1) { XlaBuilder builder(TestName()); - auto pred = builder.ConstantR0(false); - auto operand1 = builder.ConstantR0(56.0f); - auto operand2 = builder.ConstantR0(12.0f); + auto pred = ConstantR0(&builder, false); + auto operand1 = ConstantR0(&builder, 56.0f); + auto operand2 = ConstantR0(&builder, 12.0f); auto identity = CreateR0IdentityComputation(); - builder.Conditional(pred, operand1, identity, operand2, identity); + Conditional(pred, operand1, identity, operand2, identity); ComputeAndCompareR0(&builder, 12.0f, {}, error_spec_); } @@ -198,11 +197,11 @@ XLA_TEST_F(ConditionalOpTest, Parameters1) { // that take in different arguments. XLA_TEST_F(ConditionalOpTest, DiffComputationsDiffArgs) { XlaBuilder builder(TestName()); - auto pred = builder.ConstantR0(false); - auto operand1 = builder.ConstantR0(56.4f); - auto operand2 = builder.ConstantR0(12.6f); - builder.Conditional(pred, operand1, CreateR0CeilComputation(), operand2, - CreateR0FloorComputation()); + auto pred = ConstantR0(&builder, false); + auto operand1 = ConstantR0(&builder, 56.4f); + auto operand2 = ConstantR0(&builder, 12.6f); + Conditional(pred, operand1, CreateR0CeilComputation(), operand2, + CreateR0FloorComputation()); ComputeAndCompareR0(&builder, 12.0f, {}, error_spec_); } @@ -211,10 +210,10 @@ XLA_TEST_F(ConditionalOpTest, DiffComputationsDiffArgs) { // that take in the same arguments. XLA_TEST_F(ConditionalOpTest, DiffComputationsSameArg) { XlaBuilder builder(TestName()); - auto pred = builder.ConstantR0(false); - auto operand = builder.ConstantR0(12.6f); - builder.Conditional(pred, operand, CreateR0CeilComputation(), operand, - CreateR0FloorComputation()); + auto pred = ConstantR0(&builder, false); + auto operand = ConstantR0(&builder, 12.6f); + Conditional(pred, operand, CreateR0CeilComputation(), operand, + CreateR0FloorComputation()); ComputeAndCompareR0(&builder, 12.0f, {}, error_spec_); } @@ -223,11 +222,11 @@ XLA_TEST_F(ConditionalOpTest, DiffComputationsSameArg) { // take in different arguments. XLA_TEST_F(ConditionalOpTest, SameComputationDiffArgs) { XlaBuilder builder(TestName()); - auto pred = builder.ConstantR0(false); - auto operand1 = builder.ConstantR0(56.4f); - auto operand2 = builder.ConstantR0(12.6f); + auto pred = ConstantR0(&builder, false); + auto operand1 = ConstantR0(&builder, 56.4f); + auto operand2 = ConstantR0(&builder, 12.6f); auto floor = CreateR0FloorComputation(); - builder.Conditional(pred, operand1, floor, operand2, floor); + Conditional(pred, operand1, floor, operand2, floor); ComputeAndCompareR0(&builder, 12.0f, {}, error_spec_); } @@ -236,10 +235,10 @@ XLA_TEST_F(ConditionalOpTest, SameComputationDiffArgs) { // take in the same arguments. XLA_TEST_F(ConditionalOpTest, SameComputationSameArg) { XlaBuilder builder(TestName()); - auto pred = builder.ConstantR0(false); - auto operand = builder.ConstantR0(12.6f); + auto pred = ConstantR0(&builder, false); + auto operand = ConstantR0(&builder, 12.6f); auto floor = CreateR0FloorComputation(); - builder.Conditional(pred, operand, floor, operand, floor); + Conditional(pred, operand, floor, operand, floor); ComputeAndCompareR0(&builder, 12.0f, {}, error_spec_); } @@ -248,11 +247,11 @@ XLA_TEST_F(ConditionalOpTest, SameComputationSameArg) { // and false cases. XLA_TEST_F(ConditionalOpTest, SameComputationDiffInstances) { XlaBuilder builder(TestName()); - auto pred = builder.ConstantR0(false); - auto operand1 = builder.ConstantR0(56.4f); - auto operand2 = builder.ConstantR0(12.6f); - builder.Conditional(pred, operand1, CreateR0FloorComputation(), operand2, - CreateR0FloorComputation()); + auto pred = ConstantR0(&builder, false); + auto operand1 = ConstantR0(&builder, 56.4f); + auto operand2 = ConstantR0(&builder, 12.6f); + Conditional(pred, operand1, CreateR0FloorComputation(), operand2, + CreateR0FloorComputation()); ComputeAndCompareR0(&builder, 12.0f, {}, error_spec_); } @@ -261,19 +260,19 @@ XLA_TEST_F(ConditionalOpTest, SameComputationDiffInstances) { XLA_TEST_F(ConditionalOpTest, ConditionalWithCall) { Shape r0bool = ShapeUtil::MakeShape(PRED, {}); XlaBuilder inner_builder(TestName() + ".inner_conditional"); - auto pred_cond = inner_builder.Parameter(0, r0bool, "param0"); - auto true_operand = inner_builder.Parameter(1, r0f32_, "param1"); - auto false_operand = inner_builder.Parameter(2, r0f32_, "param2"); - inner_builder.Conditional(pred_cond, true_operand, CreateR0CeilComputation(), - false_operand, CreateR0FloorComputation()); + auto pred_cond = Parameter(&inner_builder, 0, r0bool, "param0"); + auto true_operand = Parameter(&inner_builder, 1, r0f32_, "param1"); + auto false_operand = Parameter(&inner_builder, 2, r0f32_, "param2"); + Conditional(pred_cond, true_operand, CreateR0CeilComputation(), false_operand, + CreateR0FloorComputation()); auto inner_builder_result = inner_builder.Build(); XlaBuilder builder(TestName()); - auto pred = builder.ConstantR0(false); - auto operand1 = builder.ConstantR0(56.4f); - auto operand2 = builder.ConstantR0(12.6f); - builder.Call(inner_builder_result.ConsumeValueOrDie(), - {pred, operand1, operand2}); + auto pred = ConstantR0(&builder, false); + auto operand1 = ConstantR0(&builder, 56.4f); + auto operand2 = ConstantR0(&builder, 12.6f); + Call(&builder, inner_builder_result.ConsumeValueOrDie(), + {pred, operand1, operand2}); ComputeAndCompareR0(&builder, 12.0f, {}, error_spec_); } @@ -282,12 +281,12 @@ XLA_TEST_F(ConditionalOpTest, ConditionalWithCall) { // true. XLA_TEST_F(ConditionalOpTest, Parameters2TrueBranch) { XlaBuilder builder(TestName()); - auto pred = builder.ConstantR0(true); - auto operand1 = builder.ConstantR0(56.0f); - auto operand2 = builder.ConstantR0(12.0f); - auto operands = builder.Tuple({operand1, operand2}); - builder.Conditional(pred, operands, CreateR0TupleAddComputation(), operands, - CreateR0TupleSubComputation()); + auto pred = ConstantR0(&builder, true); + auto operand1 = ConstantR0(&builder, 56.0f); + auto operand2 = ConstantR0(&builder, 12.0f); + auto operands = Tuple(&builder, {operand1, operand2}); + Conditional(pred, operands, CreateR0TupleAddComputation(), operands, + CreateR0TupleSubComputation()); ComputeAndCompareR0(&builder, 68.0f, {}, error_spec_); } @@ -296,12 +295,12 @@ XLA_TEST_F(ConditionalOpTest, Parameters2TrueBranch) { // false. XLA_TEST_F(ConditionalOpTest, Parameters2FalseBranch) { XlaBuilder builder(TestName()); - auto pred = builder.ConstantR0(false); - auto operand1 = builder.ConstantR0(56.0f); - auto operand2 = builder.ConstantR0(12.0f); - auto operands = builder.Tuple({operand1, operand2}); - builder.Conditional(pred, operands, CreateR0TupleAddComputation(), operands, - CreateR0TupleSubComputation()); + auto pred = ConstantR0(&builder, false); + auto operand1 = ConstantR0(&builder, 56.0f); + auto operand2 = ConstantR0(&builder, 12.0f); + auto operands = Tuple(&builder, {operand1, operand2}); + Conditional(pred, operands, CreateR0TupleAddComputation(), operands, + CreateR0TupleSubComputation()); ComputeAndCompareR0(&builder, 44.0f, {}, error_spec_); } @@ -310,12 +309,12 @@ XLA_TEST_F(ConditionalOpTest, Parameters2FalseBranch) { // predicate is true. XLA_TEST_F(ConditionalOpTest, Parameters2ArrayTrueBranch) { XlaBuilder builder(TestName()); - auto pred = builder.ConstantR0(true); - auto operand1 = builder.ConstantR1({24.0f, 56.0f}); - auto operand2 = builder.ConstantR1({10.0f, 11.0f}); - auto operands = builder.Tuple({operand1, operand2}); - builder.Conditional(pred, operands, CreateR1TupleAddComputation(), operands, - CreateR1TupleSubComputation()); + auto pred = ConstantR0(&builder, true); + auto operand1 = ConstantR1(&builder, {24.0f, 56.0f}); + auto operand2 = ConstantR1(&builder, {10.0f, 11.0f}); + auto operands = Tuple(&builder, {operand1, operand2}); + Conditional(pred, operands, CreateR1TupleAddComputation(), operands, + CreateR1TupleSubComputation()); ComputeAndCompareR1(&builder, {34.0f, 67.0f}, {}, error_spec_); } @@ -324,12 +323,12 @@ XLA_TEST_F(ConditionalOpTest, Parameters2ArrayTrueBranch) { // predicate is false. XLA_TEST_F(ConditionalOpTest, Parameters2ArrayFalseBranch) { XlaBuilder builder(TestName()); - auto pred = builder.ConstantR0(false); - auto operand1 = builder.ConstantR1({24.0f, 56.0f}); - auto operand2 = builder.ConstantR1({10.0f, 11.0f}); - auto operands = builder.Tuple({operand1, operand2}); - builder.Conditional(pred, operands, CreateR1TupleAddComputation(), operands, - CreateR1TupleSubComputation()); + auto pred = ConstantR0(&builder, false); + auto operand1 = ConstantR1(&builder, {24.0f, 56.0f}); + auto operand2 = ConstantR1(&builder, {10.0f, 11.0f}); + auto operands = Tuple(&builder, {operand1, operand2}); + Conditional(pred, operands, CreateR1TupleAddComputation(), operands, + CreateR1TupleSubComputation()); ComputeAndCompareR1(&builder, {14.0f, 45.0f}, {}, error_spec_); } @@ -337,32 +336,34 @@ XLA_TEST_F(ConditionalOpTest, Parameters2ArrayFalseBranch) { // Test true and false computations that return a tuple of scalars. XLA_TEST_F(ConditionalOpTest, ReturnTupleOfScalars) { XlaBuilder builder(TestName()); - auto pred = builder.ConstantR0(false); - auto operands = builder.Tuple( - {builder.ConstantR0(12.2f), builder.ConstantR0(25.6f)}); - builder.Conditional(pred, operands, CreateR0TupleCeilComputation(), operands, - CreateR0TupleFloorComputation()); + auto pred = ConstantR0(&builder, false); + auto operands = Tuple(&builder, {ConstantR0(&builder, 12.2f), + ConstantR0(&builder, 25.6f)}); + Conditional(pred, operands, CreateR0TupleCeilComputation(), operands, + CreateR0TupleFloorComputation()); ComputeAndCompareTuple( &builder, - *Literal::MakeTuple({Literal::CreateR0(12.0f).get(), - Literal::CreateR0(25.0f).get()}), + *LiteralUtil::MakeTuple({LiteralUtil::CreateR0(12.0f).get(), + LiteralUtil::CreateR0(25.0f).get()}), {}, error_spec_); } // Test true and false computations that return a tuple of arrays. XLA_TEST_F(ConditionalOpTest, ReturnTupleOfArrays) { XlaBuilder builder(TestName()); - auto pred = builder.ConstantR0(true); - auto operands = builder.Tuple({builder.ConstantR1({12.2f, 15.8f}), - builder.ConstantR1({25.6f, 29.2f})}); - builder.Conditional(pred, operands, CreateR1TupleCeilComputation(), operands, - CreateR1TupleFloorComputation()); + auto pred = ConstantR0(&builder, true); + auto operands = + Tuple(&builder, {ConstantR1(&builder, {12.2f, 15.8f}), + ConstantR1(&builder, {25.6f, 29.2f})}); + Conditional(pred, operands, CreateR1TupleCeilComputation(), operands, + CreateR1TupleFloorComputation()); ComputeAndCompareTuple( &builder, - *Literal::MakeTuple({Literal::CreateR1({13.0f, 16.0f}).get(), - Literal::CreateR1({26.0f, 30.0f}).get()}), + *LiteralUtil::MakeTuple( + {LiteralUtil::CreateR1({13.0f, 16.0f}).get(), + LiteralUtil::CreateR1({26.0f, 30.0f}).get()}), {}, error_spec_); } @@ -371,37 +372,38 @@ XLA_TEST_F(ConditionalOpTest, ReturnTupleOfArrays) { XLA_TEST_F(ConditionalOpTest, ReturnTupleofPredicateScalarArray) { XlaBuilder true_builder(TestName() + ".true"); { - true_builder.Parameter(0, empty_tuple_, "tuple"); - auto true_pred = true_builder.ConstantR0(true); - auto true_scalar = true_builder.ConstantR0(12.2f); - auto true_array = true_builder.ConstantR1({12.8f, 14.6f}); - true_builder.Tuple({true_pred, true_scalar, true_array}); + Parameter(&true_builder, 0, empty_tuple_, "tuple"); + auto true_pred = ConstantR0(&true_builder, true); + auto true_scalar = ConstantR0(&true_builder, 12.2f); + auto true_array = ConstantR1(&true_builder, {12.8f, 14.6f}); + Tuple(&true_builder, {true_pred, true_scalar, true_array}); } auto true_builder_result = true_builder.Build(); EXPECT_IS_OK(true_builder_result.status()); XlaBuilder false_builder(TestName() + ".false"); { - false_builder.Parameter(0, empty_tuple_, "tuple"); - auto false_pred = false_builder.ConstantR0(false); - auto false_scalar = false_builder.ConstantR0(25.6f); - auto false_array = false_builder.ConstantR1({26.4f, 32.6f}); - false_builder.Tuple({false_pred, false_scalar, false_array}); + Parameter(&false_builder, 0, empty_tuple_, "tuple"); + auto false_pred = ConstantR0(&false_builder, false); + auto false_scalar = ConstantR0(&false_builder, 25.6f); + auto false_array = ConstantR1(&false_builder, {26.4f, 32.6f}); + Tuple(&false_builder, {false_pred, false_scalar, false_array}); } auto false_builder_result = false_builder.Build(); EXPECT_IS_OK(false_builder_result.status()); XlaBuilder builder(TestName()); - auto pred = builder.ConstantR0(true); - auto operands = builder.Tuple({}); - builder.Conditional(pred, operands, true_builder_result.ConsumeValueOrDie(), - operands, false_builder_result.ConsumeValueOrDie()); + auto pred = ConstantR0(&builder, true); + auto operands = Tuple(&builder, {}); + Conditional(pred, operands, true_builder_result.ConsumeValueOrDie(), operands, + false_builder_result.ConsumeValueOrDie()); ComputeAndCompareTuple( &builder, - *Literal::MakeTuple({Literal::CreateR0(true).get(), - Literal::CreateR0(12.2f).get(), - Literal::CreateR1({12.8f, 14.6f}).get()}), + *LiteralUtil::MakeTuple( + {LiteralUtil::CreateR0(true).get(), + LiteralUtil::CreateR0(12.2f).get(), + LiteralUtil::CreateR1({12.8f, 14.6f}).get()}), {}, error_spec_); } @@ -409,45 +411,48 @@ XLA_TEST_F(ConditionalOpTest, ReturnTupleofPredicateScalarArray) { XLA_TEST_F(ConditionalOpTest, ReturnNestedTuple) { XlaBuilder true_builder(TestName() + ".true"); { - true_builder.Parameter(0, empty_tuple_, "tuple"); - auto true_constant1 = true_builder.ConstantR0(12.2f); - auto true_constant2 = true_builder.ConstantR1({12.8f, 14.6f}); - auto true_constant3 = true_builder.ConstantR1({25.4f, 29.8f}); - auto true_constant4 = true_builder.ConstantR0(35.6f); - true_builder.Tuple({true_builder.Tuple({true_constant1, true_constant2}), - true_builder.Tuple({true_constant3, true_constant4})}); + Parameter(&true_builder, 0, empty_tuple_, "tuple"); + auto true_constant1 = ConstantR0(&true_builder, 12.2f); + auto true_constant2 = ConstantR1(&true_builder, {12.8f, 14.6f}); + auto true_constant3 = ConstantR1(&true_builder, {25.4f, 29.8f}); + auto true_constant4 = ConstantR0(&true_builder, 35.6f); + Tuple(&true_builder, + {Tuple(&true_builder, {true_constant1, true_constant2}), + Tuple(&true_builder, {true_constant3, true_constant4})}); } auto true_builder_result = true_builder.Build(); EXPECT_IS_OK(true_builder_result.status()); XlaBuilder false_builder(TestName() + ".false"); { - false_builder.Parameter(0, empty_tuple_, "tuple"); - auto false_constant1 = false_builder.ConstantR0(46.6f); - auto false_constant2 = false_builder.ConstantR1({54.4f, 58.4f}); - auto false_constant3 = false_builder.ConstantR1({62.1f, 67.4f}); - auto false_constant4 = false_builder.ConstantR0(9.3f); - false_builder.Tuple( - {false_builder.Tuple({false_constant1, false_constant2}), - false_builder.Tuple({false_constant3, false_constant4})}); + Parameter(&false_builder, 0, empty_tuple_, "tuple"); + auto false_constant1 = ConstantR0(&false_builder, 46.6f); + auto false_constant2 = ConstantR1(&false_builder, {54.4f, 58.4f}); + auto false_constant3 = ConstantR1(&false_builder, {62.1f, 67.4f}); + auto false_constant4 = ConstantR0(&false_builder, 9.3f); + Tuple(&false_builder, + {Tuple(&false_builder, {false_constant1, false_constant2}), + Tuple(&false_builder, {false_constant3, false_constant4})}); } auto false_builder_result = false_builder.Build(); EXPECT_IS_OK(false_builder_result.status()); XlaBuilder builder(TestName()); - auto pred = builder.ConstantR0(false); - auto operands = builder.Tuple({}); - builder.Conditional(pred, operands, true_builder_result.ConsumeValueOrDie(), - operands, false_builder_result.ConsumeValueOrDie()); + auto pred = ConstantR0(&builder, false); + auto operands = Tuple(&builder, {}); + Conditional(pred, operands, true_builder_result.ConsumeValueOrDie(), operands, + false_builder_result.ConsumeValueOrDie()); ComputeAndCompareTuple( &builder, - *Literal::MakeTuple( - {Literal::MakeTuple({Literal::CreateR0(46.6f).get(), - Literal::CreateR1({54.4f, 58.4f}).get()}) + *LiteralUtil::MakeTuple( + {LiteralUtil::MakeTuple( + {LiteralUtil::CreateR0(46.6f).get(), + LiteralUtil::CreateR1({54.4f, 58.4f}).get()}) .get(), - Literal::MakeTuple({Literal::CreateR1({62.1f, 67.4f}).get(), - Literal::CreateR0(9.3f).get()}) + LiteralUtil::MakeTuple( + {LiteralUtil::CreateR1({62.1f, 67.4f}).get(), + LiteralUtil::CreateR0(9.3f).get()}) .get()}), {}, error_spec_); } @@ -464,8 +469,8 @@ XLA_TEST_F(ConditionalOpTest, ScalarOperandsFromExternalParams) { CreateR0Parameter(56.3f, 1, "operand1", &builder, &operand1); auto operand2_param = CreateR0Parameter(12.7f, 2, "operand2", &builder, &operand2); - builder.Conditional(pred, operand1, CreateR0CeilComputation(), operand2, - CreateR0FloorComputation()); + Conditional(pred, operand1, CreateR0CeilComputation(), operand2, + CreateR0FloorComputation()); ComputeAndCompareR0( &builder, 57.0f, @@ -484,8 +489,8 @@ XLA_TEST_F(ConditionalOpTest, ArrayOperandsFromExternalParams) { &builder, &operand1); auto operand2_param = CreateR1Parameter({10.2f, 11.6f}, 2, "operand2", &builder, &operand2); - builder.Conditional(pred, operand1, CreateR1CeilComputation(), operand2, - CreateR1FloorComputation()); + Conditional(pred, operand1, CreateR1CeilComputation(), operand2, + CreateR1FloorComputation()); ComputeAndCompareR1( &builder, {10.0f, 11.0f}, @@ -499,27 +504,25 @@ XLA_TEST_F(ConditionalOpTest, NestedConditionals) { { Shape r0bool = ShapeUtil::MakeShape(PRED, {}); Shape tuple_shape = ShapeUtil::MakeTupleShape({r0bool, r0f32_, r0f32_}); - auto param0 = inner_builder.Parameter(0, tuple_shape, "param0"); - auto pred_cond = inner_builder.GetTupleElement(param0, 0); - auto true_operand = inner_builder.GetTupleElement(param0, 1); - auto false_operand = inner_builder.GetTupleElement(param0, 2); - inner_builder.Conditional(pred_cond, true_operand, - CreateR0CeilComputation(), false_operand, - CreateR0FloorComputation()); + auto param0 = Parameter(&inner_builder, 0, tuple_shape, "param0"); + auto pred_cond = GetTupleElement(param0, 0); + auto true_operand = GetTupleElement(param0, 1); + auto false_operand = GetTupleElement(param0, 2); + Conditional(pred_cond, true_operand, CreateR0CeilComputation(), + false_operand, CreateR0FloorComputation()); } auto inner_builder_result = inner_builder.Build(); EXPECT_IS_OK(inner_builder_result.status()); XlaBuilder builder(TestName()); - auto pred1 = builder.ConstantR0(true); - auto pred2 = builder.ConstantR0(false); - auto operand1 = builder.ConstantR0(1.1f); - auto operand2 = builder.ConstantR0(12.2f); - auto operand3 = builder.ConstantR0(43.3f); - auto tuple_operand = builder.Tuple({pred2, operand1, operand2}); - builder.Conditional(pred1, tuple_operand, - inner_builder_result.ConsumeValueOrDie(), operand3, - CreateR0IdentityComputation()); + auto pred1 = ConstantR0(&builder, true); + auto pred2 = ConstantR0(&builder, false); + auto operand1 = ConstantR0(&builder, 1.1f); + auto operand2 = ConstantR0(&builder, 12.2f); + auto operand3 = ConstantR0(&builder, 43.3f); + auto tuple_operand = Tuple(&builder, {pred2, operand1, operand2}); + Conditional(pred1, tuple_operand, inner_builder_result.ConsumeValueOrDie(), + operand3, CreateR0IdentityComputation()); ComputeAndCompareR0(&builder, 12.0f, {}, error_spec_); } @@ -529,23 +532,22 @@ XLA_TEST_F(ConditionalOpTest, ConditionalInNestedComputation) { { Shape r0bool = ShapeUtil::MakeShape(PRED, {}); Shape tuple_shape = ShapeUtil::MakeTupleShape({r0bool, r0f32_, r0f32_}); - auto param0 = inner_builder.Parameter(0, tuple_shape, "param0"); - auto pred_cond = inner_builder.GetTupleElement(param0, 0); - auto true_operand = inner_builder.GetTupleElement(param0, 1); - auto false_operand = inner_builder.GetTupleElement(param0, 2); - inner_builder.Conditional(pred_cond, true_operand, - CreateR0CeilComputation(), false_operand, - CreateR0FloorComputation()); + auto param0 = Parameter(&inner_builder, 0, tuple_shape, "param0"); + auto pred_cond = GetTupleElement(param0, 0); + auto true_operand = GetTupleElement(param0, 1); + auto false_operand = GetTupleElement(param0, 2); + Conditional(pred_cond, true_operand, CreateR0CeilComputation(), + false_operand, CreateR0FloorComputation()); } auto inner_builder_result = inner_builder.Build(); EXPECT_IS_OK(inner_builder_result.status()); XlaBuilder builder(TestName()); - auto pred2 = builder.ConstantR0(false); - auto operand1 = builder.ConstantR0(1.1f); - auto operand2 = builder.ConstantR0(12.2f); - auto tuple_operand = builder.Tuple({pred2, operand1, operand2}); - builder.Call(inner_builder_result.ConsumeValueOrDie(), {tuple_operand}); + auto pred2 = ConstantR0(&builder, false); + auto operand1 = ConstantR0(&builder, 1.1f); + auto operand2 = ConstantR0(&builder, 12.2f); + auto tuple_operand = Tuple(&builder, {pred2, operand1, operand2}); + Call(&builder, inner_builder_result.ConsumeValueOrDie(), {tuple_operand}); ComputeAndCompareR0(&builder, 12.0f, {}, error_spec_); } @@ -553,12 +555,12 @@ XLA_TEST_F(ConditionalOpTest, ConditionalInNestedComputation) { // Test a mismatch in the shape of the true operand and true computation. XLA_TEST_F(ConditionalOpTest, ShapeMismatch) { XlaBuilder builder(TestName()); - auto pred = builder.ConstantR0(true); - auto operand1 = builder.ConstantR0(56.0f); - auto operand2 = builder.ConstantR0(12.0f); - auto operands = builder.Tuple({operand1, operand2}); - builder.Conditional(pred, operands, CreateR1TupleAddComputation(), operands, - CreateR0TupleSubComputation()); + auto pred = ConstantR0(&builder, true); + auto operand1 = ConstantR0(&builder, 56.0f); + auto operand2 = ConstantR0(&builder, 12.0f); + auto operands = Tuple(&builder, {operand1, operand2}); + Conditional(pred, operands, CreateR1TupleAddComputation(), operands, + CreateR0TupleSubComputation()); auto result = builder.Build(); EXPECT_FALSE(result.ok()); @@ -572,45 +574,45 @@ XLA_TEST_F(ConditionalOpTest, SwappedInputsInSequentialConditionals) { XlaComputation swapper; { XlaBuilder builder(TestName() + ".swapper"); - auto param0 = builder.Parameter(0, tuple_shape, "sp0"); - auto x = builder.GetTupleElement(param0, 0); - auto y = builder.GetTupleElement(param0, 1); - builder.Tuple({y, x}); + auto param0 = Parameter(&builder, 0, tuple_shape, "sp0"); + auto x = GetTupleElement(param0, 0); + auto y = GetTupleElement(param0, 1); + Tuple(&builder, {y, x}); swapper = builder.Build().ConsumeValueOrDie(); } XlaComputation forwarder; { XlaBuilder builder(TestName() + ".forwarder"); - auto param0 = builder.Parameter(0, tuple_shape, "fp0"); - auto x = builder.GetTupleElement(param0, 0); - auto y = builder.GetTupleElement(param0, 1); - builder.Tuple({x, y}); + auto param0 = Parameter(&builder, 0, tuple_shape, "fp0"); + auto x = GetTupleElement(param0, 0); + auto y = GetTupleElement(param0, 1); + Tuple(&builder, {x, y}); forwarder = builder.Build().ConsumeValueOrDie(); } XlaComputation main; { XlaBuilder builder(TestName() + ".main"); - auto param0 = builder.Parameter(0, tuple_shape, "mp0"); - auto x = builder.GetTupleElement(param0, 0); - auto y = builder.GetTupleElement(param0, 1); - auto lt_pred = builder.Lt(x, y); - auto res = builder.Conditional(lt_pred, param0, forwarder, param0, swapper); - auto ge_pred = builder.Ge(x, y); - builder.Conditional(ge_pred, res, swapper, res, forwarder); + auto param0 = Parameter(&builder, 0, tuple_shape, "mp0"); + auto x = GetTupleElement(param0, 0); + auto y = GetTupleElement(param0, 1); + auto lt_pred = Lt(x, y); + auto res = Conditional(lt_pred, param0, forwarder, param0, swapper); + auto ge_pred = Ge(x, y); + Conditional(ge_pred, res, swapper, res, forwarder); main = builder.Build().ConsumeValueOrDie(); } auto test_swap = [&](float a, float b) { XlaBuilder builder(TestName()); - auto x = builder.ConstantR0(a); - auto y = builder.ConstantR0(b); - auto tuple_operand = builder.Tuple({x, y}); - builder.Call(main, {tuple_operand}); + auto x = ConstantR0(&builder, a); + auto y = ConstantR0(&builder, b); + auto tuple_operand = Tuple(&builder, {x, y}); + Call(&builder, main, {tuple_operand}); ComputeAndCompareTuple( &builder, - *Literal::MakeTuple({Literal::CreateR0(a).get(), - Literal::CreateR0(b).get()}), + *LiteralUtil::MakeTuple({LiteralUtil::CreateR0(a).get(), + LiteralUtil::CreateR0(b).get()}), {}, error_spec_); }; diff --git a/tensorflow/compiler/xla/tests/constants_test.cc b/tensorflow/compiler/xla/tests/constants_test.cc index 916ffadbc798ec0dd016f45b0bc4c36233455ee7..71d72a9828c5445be2cb1f559cf31363507bcd8d 100644 --- a/tensorflow/compiler/xla/tests/constants_test.cc +++ b/tensorflow/compiler/xla/tests/constants_test.cc @@ -26,6 +26,7 @@ limitations under the License. #include "tensorflow/compiler/xla/literal_util.h" #include "tensorflow/compiler/xla/tests/client_library_test_base.h" #include "tensorflow/compiler/xla/tests/literal_test_util.h" +#include "tensorflow/core/lib/core/status_test_util.h" #include "tensorflow/core/platform/test.h" #include "tensorflow/core/platform/types.h" @@ -39,7 +40,7 @@ class ConstantsTest : public ClientLibraryTestBase { TEST_F(ConstantsTest, ZeroCellF32) { XlaBuilder builder(TestName()); - builder.ConstantR1({}); + ConstantR1(&builder, {}); ComputeAndCompareR1(&builder, {}, {}, error_spec_); } @@ -48,7 +49,7 @@ TEST_F(ConstantsTest, OneCellF32) { std::vector constant = {2.0}; XlaBuilder builder(TestName()); - builder.ConstantR1(constant); + ConstantR1(&builder, constant); ComputeAndCompareR1(&builder, constant, {}, error_spec_); } @@ -57,7 +58,7 @@ TEST_F(ConstantsTest, OneCellS32) { std::vector constant = {2}; XlaBuilder builder(TestName()); - builder.ConstantR1(constant); + ConstantR1(&builder, constant); ComputeAndCompareR1(&builder, constant, {}); } @@ -66,7 +67,7 @@ TEST_F(ConstantsTest, OneCellU32) { std::vector constant = {2}; XlaBuilder builder(TestName()); - builder.ConstantR1(constant); + ConstantR1(&builder, constant); ComputeAndCompareR1(&builder, constant, {}); } @@ -75,7 +76,7 @@ TEST_F(ConstantsTest, EightCells) { std::vector constant = {0.0, 1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0}; XlaBuilder builder(TestName()); - builder.ConstantR1(constant); + ConstantR1(&builder, constant); ComputeAndCompareR1(&builder, constant, {}, error_spec_); } @@ -85,14 +86,14 @@ TEST_F(ConstantsTest, SixteenCells) { 8.0, 9.0, 10.0, 11.0, 12.0, 13.0, 14.0, 15.0}; XlaBuilder builder(TestName()); - builder.ConstantR1(constant); + ConstantR1(&builder, constant); ComputeAndCompareR1(&builder, constant, {}, error_spec_); } TEST_F(ConstantsTest, Empty_0x2) { XlaBuilder builder(TestName()); - builder.ConstantR2FromArray2D(Array2D(0, 2)); + ConstantR2FromArray2D(&builder, Array2D(0, 2)); ComputeAndCompareR2(&builder, Array2D(0, 2), {}, error_spec_); } @@ -102,15 +103,15 @@ TEST_F(ConstantsTest, Small_2x2) { MakeLinspaceArray2D(100.0, 200.0, 2, 2); XlaBuilder builder(TestName()); - builder.ConstantR2FromArray2D(*constant); + ConstantR2FromArray2D(&builder, *constant); ComputeAndCompareR2(&builder, *constant, {}, error_spec_); } TEST_F(ConstantsTest, Empty_3x0x2) { XlaBuilder builder(TestName()); - auto constant = builder.ConstantLiteral( - *Literal::CreateR3FromArray3D(Array3D(3, 0, 2))); + ConstantLiteral(&builder, *LiteralUtil::CreateR3FromArray3D( + Array3D(3, 0, 2))); ComputeAndCompareR3(&builder, Array3D(3, 0, 2), {}); } @@ -125,8 +126,7 @@ TEST_F(ConstantsTest, Small_2x2x2) { {{5.f, 6.f}, // y0 {7.f, 8.f}}, // y1 }); - auto constant = - builder.ConstantLiteral(*Literal::CreateR3FromArray3D(array3d)); + ConstantLiteral(&builder, *LiteralUtil::CreateR3FromArray3D(array3d)); ComputeAndCompareR3(&builder, array3d, {}); } @@ -141,17 +141,17 @@ TEST_F(ConstantsTest, Small_3x2x1x1) { }); input_array.FillWithPZ(pz); std::unique_ptr input_literal = - Literal::CreateR4FromArray4D(input_array); + LiteralUtil::CreateR4FromArray4D(input_array); { XlaBuilder builder(TestName()); - builder.ConstantLiteral(*input_literal); + ConstantLiteral(&builder, *input_literal); ComputeAndCompareR4(&builder, input_array, {}, error_spec_); } { XlaBuilder builder(TestName()); - builder.ConstantR4FromArray4D(input_array); + ConstantR4FromArray4D(&builder, input_array); ComputeAndCompareR4(&builder, input_array, {}, error_spec_); } } @@ -159,17 +159,26 @@ TEST_F(ConstantsTest, Small_3x2x1x1) { // TODO(b/29263943): Support tuple constants. TEST_F(ConstantsTest, DISABLED_TupleConstant) { XlaBuilder builder(TestName()); - builder.ConstantLiteral( - *Literal::MakeTuple({Literal::CreateR2({{1.0}, {2.0}}).get(), - Literal::CreateR1({2.0, 42}).get()})); + ConstantLiteral(&builder, + *LiteralUtil::MakeTuple( + {LiteralUtil::CreateR2({{1.0}, {2.0}}).get(), + LiteralUtil::CreateR1({2.0, 42}).get()})); std::unique_ptr result = ExecuteAndTransfer(&builder, {}).ConsumeValueOrDie(); - LiteralTestUtil::ExpectR2Near( - {{1.0}, {2.0}}, LiteralSlice(*result, {0}), error_spec_); - LiteralTestUtil::ExpectR1Near( - {2.0, 42.0}, LiteralSlice(*result, {1}), error_spec_); + LiteralTestUtil::ExpectR2Near({{1.0}, {2.0}}, + LiteralSlice(*result, {0}), error_spec_); + LiteralTestUtil::ExpectR1Near({2.0, 42.0}, LiteralSlice(*result, {1}), + error_spec_); +} + +TEST_F(ConstantsTest, Token) { + XlaBuilder builder(TestName()); + ConstantLiteral(&builder, *LiteralUtil::CreateToken()); + // TODO(b/80000000): tokens cannot be returned from computations. + Tuple(&builder, {}); + TF_ASSERT_OK(Execute(&builder, {}).status()); } } // namespace diff --git a/tensorflow/compiler/xla/tests/convert_test.cc b/tensorflow/compiler/xla/tests/convert_test.cc index 3a885b43893f84fb331572343308130bb06f7e86..dca57fd1c705da758ad0305b7799d6e806bbf72f 100644 --- a/tensorflow/compiler/xla/tests/convert_test.cc +++ b/tensorflow/compiler/xla/tests/convert_test.cc @@ -45,8 +45,8 @@ class ConvertTest : public ClientLibraryTestBase { TEST_F(ConvertTest, ConvertR1S32ToR1S32) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({42, 64}); - builder.ConvertElementType(a, S32); + auto a = ConstantR1(&builder, {42, 64}); + ConvertElementType(a, S32); std::vector expected = {42, 64}; ComputeAndCompareR1(&builder, expected, {}); @@ -54,8 +54,8 @@ TEST_F(ConvertTest, ConvertR1S32ToR1S32) { TEST_F(ConvertTest, ConvertR1F32ToR1F32) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({42.0f, 64.0f}); - builder.ConvertElementType(a, F32); + auto a = ConstantR1(&builder, {42.0f, 64.0f}); + ConvertElementType(a, F32); std::vector expected = {42.0f, 64.0f}; ComputeAndCompareR1(&builder, expected, {}, ErrorSpec(0.0001)); @@ -63,8 +63,8 @@ TEST_F(ConvertTest, ConvertR1F32ToR1F32) { TEST_F(ConvertTest, ConvertR1S32ToR1F32) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({42, 64}); - builder.ConvertElementType(a, F32); + auto a = ConstantR1(&builder, {42, 64}); + ConvertElementType(a, F32); std::vector expected = {42.0f, 64.0f}; ComputeAndCompareR1(&builder, expected, {}, ErrorSpec(0.0001)); @@ -72,8 +72,8 @@ TEST_F(ConvertTest, ConvertR1S32ToR1F32) { TEST_F(ConvertTest, ConvertR1PREDToR1S32) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({true, false, true}); - builder.ConvertElementType(a, S32); + auto a = ConstantR1(&builder, {true, false, true}); + ConvertElementType(a, S32); std::vector expected = {1, 0, 1}; ComputeAndCompareR1(&builder, expected, {}); @@ -81,8 +81,8 @@ TEST_F(ConvertTest, ConvertR1PREDToR1S32) { TEST_F(ConvertTest, ConvertR1PREDToR1F32) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({true, false, true}); - builder.ConvertElementType(a, F32); + auto a = ConstantR1(&builder, {true, false, true}); + ConvertElementType(a, F32); std::vector expected = {1., 0., 1.}; ComputeAndCompareR1(&builder, expected, {}); @@ -90,8 +90,8 @@ TEST_F(ConvertTest, ConvertR1PREDToR1F32) { XLA_TEST_F(ConvertTest, ConvertR1S0S32ToR1S0F32) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({}); - builder.ConvertElementType(a, F32); + auto a = ConstantR1(&builder, {}); + ConvertElementType(a, F32); std::vector expected = {}; ComputeAndCompareR1(&builder, expected, {}, ErrorSpec(0.0001)); @@ -99,8 +99,8 @@ XLA_TEST_F(ConvertTest, ConvertR1S0S32ToR1S0F32) { TEST_F(ConvertTest, ConvertR1F32ToR1S32) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({42.6, 64.4}); - builder.ConvertElementType(a, S32); + auto a = ConstantR1(&builder, {42.6, 64.4}); + ConvertElementType(a, S32); std::vector expected = {42, 64}; ComputeAndCompareR1(&builder, expected, {}); @@ -145,12 +145,12 @@ XLA_TEST_F(ConvertTest, ConvertR1S64ToR1F32) { static_cast(0x8000008000000000LL), static_cast(0x8000010000000000LL), }; - std::unique_ptr arg_literal = Literal::CreateR1({arg}); - auto arg_param = builder.Parameter(0, arg_literal->shape(), "arg_param"); + std::unique_ptr arg_literal = LiteralUtil::CreateR1({arg}); + auto arg_param = Parameter(&builder, 0, arg_literal->shape(), "arg_param"); std::unique_ptr arg_data = client_->TransferToServer(*arg_literal).ConsumeValueOrDie(); - builder.ConvertElementType(arg_param, F32); + ConvertElementType(arg_param, F32); std::vector expected(arg.size()); for (int64 i = 0; i < arg.size(); ++i) { @@ -164,12 +164,12 @@ XLA_TEST_F(ConvertTest, ConvertR1U32ToR1F32) { std::vector arg{0, 1, 0x1000, 0x7fffffff, 0x80000000, 0x80000001, 0x80000002, 0x80000003, 0x80000080, 0x80000081, 0x80000082, 0xFFFFFFFF}; - std::unique_ptr arg_literal = Literal::CreateR1({arg}); - auto arg_param = builder.Parameter(0, arg_literal->shape(), "arg_param"); + std::unique_ptr arg_literal = LiteralUtil::CreateR1({arg}); + auto arg_param = Parameter(&builder, 0, arg_literal->shape(), "arg_param"); std::unique_ptr arg_data = client_->TransferToServer(*arg_literal).ConsumeValueOrDie(); - builder.ConvertElementType(arg_param, F32); + ConvertElementType(arg_param, F32); std::vector expected(arg.size()); for (int64 i = 0; i < arg.size(); ++i) { @@ -182,12 +182,12 @@ XLA_TEST_F(ConvertTest, ConvertR1F32ToR1U32) { XlaBuilder builder(TestName()); std::vector arg{0.0f, 1.0f, 16777216.0f, 16777218.0f, 2147483647.0f, 4294967040.0f}; - std::unique_ptr arg_literal = Literal::CreateR1({arg}); - auto arg_param = builder.Parameter(0, arg_literal->shape(), "arg_param"); + std::unique_ptr arg_literal = LiteralUtil::CreateR1({arg}); + auto arg_param = Parameter(&builder, 0, arg_literal->shape(), "arg_param"); std::unique_ptr arg_data = client_->TransferToServer(*arg_literal).ConsumeValueOrDie(); - builder.ConvertElementType(arg_param, U32); + ConvertElementType(arg_param, U32); std::vector expected(arg.size()); for (int64 i = 0; i < arg.size(); ++i) { @@ -199,12 +199,12 @@ XLA_TEST_F(ConvertTest, ConvertR1F32ToR1U32) { XLA_TEST_F(ConvertTest, ConvertR1U32ToR1S64) { XlaBuilder builder(TestName()); std::vector arg{0, 1, 0x1000, 0x7fffffff, 0x80000082, 0xFFFFFFFF}; - std::unique_ptr arg_literal = Literal::CreateR1({arg}); - auto arg_param = builder.Parameter(0, arg_literal->shape(), "arg_param"); + std::unique_ptr arg_literal = LiteralUtil::CreateR1({arg}); + auto arg_param = Parameter(&builder, 0, arg_literal->shape(), "arg_param"); std::unique_ptr arg_data = client_->TransferToServer(*arg_literal).ConsumeValueOrDie(); - builder.ConvertElementType(arg_param, S64); + ConvertElementType(arg_param, S64); std::vector expected(arg.size()); for (int64 i = 0; i < arg.size(); ++i) { @@ -216,12 +216,12 @@ XLA_TEST_F(ConvertTest, ConvertR1U32ToR1S64) { XLA_TEST_F(ConvertTest, ConvertR1S32ToR1S64) { XlaBuilder builder(TestName()); std::vector arg{0, 1, 0x1000, -1, -0x1000}; - std::unique_ptr arg_literal = Literal::CreateR1({arg}); - auto arg_param = builder.Parameter(0, arg_literal->shape(), "arg_param"); + std::unique_ptr arg_literal = LiteralUtil::CreateR1({arg}); + auto arg_param = Parameter(&builder, 0, arg_literal->shape(), "arg_param"); std::unique_ptr arg_data = client_->TransferToServer(*arg_literal).ConsumeValueOrDie(); - builder.ConvertElementType(arg_param, S64); + ConvertElementType(arg_param, S64); std::vector expected(arg.size()); for (int64 i = 0; i < arg.size(); ++i) { @@ -253,12 +253,12 @@ XLA_TEST_F(ConvertTest, ConvertR1F32ToR1S64) { 9223370937343148032.f, -9223371487098961920.f, -9223370937343148032.f}; - std::unique_ptr arg_literal = Literal::CreateR1({arg}); - auto arg_param = builder.Parameter(0, arg_literal->shape(), "arg_param"); + std::unique_ptr arg_literal = LiteralUtil::CreateR1({arg}); + auto arg_param = Parameter(&builder, 0, arg_literal->shape(), "arg_param"); std::unique_ptr arg_data = client_->TransferToServer(*arg_literal).ConsumeValueOrDie(); - builder.ConvertElementType(arg_param, S64); + ConvertElementType(arg_param, S64); std::vector expected(arg.size()); for (int64 i = 0; i < arg.size(); ++i) { @@ -269,8 +269,8 @@ XLA_TEST_F(ConvertTest, ConvertR1F32ToR1S64) { XLA_TEST_F(ConvertTest, ConvertR1U8ToR1F32) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({32, 64}); - builder.ConvertElementType(a, F32); + auto a = ConstantR1(&builder, {32, 64}); + ConvertElementType(a, F32); std::vector expected = {32.0, 64.0}; ComputeAndCompareR1(&builder, expected, {}); @@ -278,8 +278,8 @@ XLA_TEST_F(ConvertTest, ConvertR1U8ToR1F32) { XLA_TEST_F(ConvertTest, ConvertR1U8ToR1S32) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({32, 64}); - builder.ConvertElementType(a, S32); + auto a = ConstantR1(&builder, {32, 64}); + ConvertElementType(a, S32); std::vector expected = {32, 64}; ComputeAndCompareR1(&builder, expected, {}); @@ -287,8 +287,8 @@ XLA_TEST_F(ConvertTest, ConvertR1U8ToR1S32) { XLA_TEST_F(ConvertTest, ConvertR1U8ToR1U32) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({32, 64}); - builder.ConvertElementType(a, U32); + auto a = ConstantR1(&builder, {32, 64}); + ConvertElementType(a, U32); std::vector expected = {32, 64}; ComputeAndCompareR1(&builder, expected, {}); @@ -296,8 +296,8 @@ XLA_TEST_F(ConvertTest, ConvertR1U8ToR1U32) { XLA_TEST_F(ConvertTest, ConvertR1F32ToR1F64) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({32.0f, 64.0f}); - builder.ConvertElementType(a, F64); + auto a = ConstantR1(&builder, {32.0f, 64.0f}); + ConvertElementType(a, F64); std::vector expected = {32.0, 64.0}; ComputeAndCompareR1(&builder, expected, {}); @@ -305,8 +305,8 @@ XLA_TEST_F(ConvertTest, ConvertR1F32ToR1F64) { XLA_TEST_F(ConvertTest, ConvertR1F64ToR1F32) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({32.0, 64.0}); - builder.ConvertElementType(a, F32); + auto a = ConstantR1(&builder, {32.0, 64.0}); + ConvertElementType(a, F32); std::vector expected = {32.0f, 64.0f}; ComputeAndCompareR1(&builder, expected, {}); @@ -314,9 +314,9 @@ XLA_TEST_F(ConvertTest, ConvertR1F64ToR1F32) { TEST_F(ConvertTest, ConvertS32Extremes) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1( - {std::numeric_limits::min(), std::numeric_limits::max()}); - builder.ConvertElementType(a, F32); + auto a = ConstantR1(&builder, {std::numeric_limits::min(), + std::numeric_limits::max()}); + ConvertElementType(a, F32); std::vector expected = { static_cast(std::numeric_limits::min()), @@ -327,10 +327,10 @@ TEST_F(ConvertTest, ConvertS32Extremes) { TEST_F(ConvertTest, ConvertMapToS32) { XlaBuilder builder(TestName()); auto b = builder.CreateSubBuilder("convert"); - auto param = b->Parameter(0, ShapeUtil::MakeShape(F32, {}), "in"); - b->ConvertElementType(param, S32); - auto a = builder.ConstantR1({42.0f, 64.0f}); - builder.Map({a}, b->BuildAndNoteError(), {0}); + auto param = Parameter(b.get(), 0, ShapeUtil::MakeShape(F32, {}), "in"); + ConvertElementType(param, S32); + auto a = ConstantR1(&builder, {42.0f, 64.0f}); + Map(&builder, {a}, b->BuildAndNoteError(), {0}); std::vector expected = {42, 64}; ComputeAndCompareR1(&builder, expected, {}); @@ -339,10 +339,10 @@ TEST_F(ConvertTest, ConvertMapToS32) { TEST_F(ConvertTest, ConvertMapToF32) { XlaBuilder builder(TestName()); auto b = builder.CreateSubBuilder("convert"); - auto param = b->Parameter(0, ShapeUtil::MakeShape(S32, {}), "in"); - b->ConvertElementType(param, F32); - auto a = builder.ConstantR1({42, 64}); - builder.Map({a}, b->BuildAndNoteError(), {0}); + auto param = Parameter(b.get(), 0, ShapeUtil::MakeShape(S32, {}), "in"); + ConvertElementType(param, F32); + auto a = ConstantR1(&builder, {42, 64}); + Map(&builder, {a}, b->BuildAndNoteError(), {0}); std::vector expected = {42.0f, 64.0f}; ComputeAndCompareR1(&builder, expected, {}, ErrorSpec(0.0001)); @@ -355,9 +355,9 @@ TEST_F(ConvertTest, ConvertMapToF32) { // the new convert should have the same element type as the old convert. TEST_F(ConvertTest, ConvertReshape) { XlaBuilder builder(TestName()); - auto input = builder.ConstantR1({42}); - auto reshape = builder.Reshape(input, /*dimensions=*/{0}, /*new_sizes=*/{}); - builder.ConvertElementType(reshape, F32); + auto input = ConstantR1(&builder, {42}); + auto reshape = Reshape(input, /*dimensions=*/{0}, /*new_sizes=*/{}); + ConvertElementType(reshape, F32); ComputeAndCompareR0(&builder, 42.0f, {}, ErrorSpec(0.0001)); } @@ -391,13 +391,13 @@ XLA_TEST_F(ConvertTest, ConvertR1F16ToR1F32) { TF_ASSERT_OK_AND_ASSIGN( std::unique_ptr dot_lhs_handle, - client_->TransferToServer(*Literal::CreateR1(input))); + client_->TransferToServer(*LiteralUtil::CreateR1(input))); XlaBuilder builder(TestName()); - builder.ConvertElementType( - builder.Parameter( - 0, ShapeUtil::MakeShape(F16, {static_cast(input.size())}), - "param"), + ConvertElementType( + Parameter(&builder, 0, + ShapeUtil::MakeShape(F16, {static_cast(input.size())}), + "param"), F32); ComputeAndCompareR1(&builder, expected_output, {dot_lhs_handle.get()}); @@ -411,13 +411,13 @@ XLA_TEST_F(ConvertTest, ConvertR1F32ToR1F16) { TF_ASSERT_OK_AND_ASSIGN( std::unique_ptr dot_lhs_handle, - client_->TransferToServer(*Literal::CreateR1(input))); + client_->TransferToServer(*LiteralUtil::CreateR1(input))); XlaBuilder builder(TestName()); - builder.ConvertElementType( - builder.Parameter( - 0, ShapeUtil::MakeShape(F32, {static_cast(input.size())}), - "param"), + ConvertElementType( + Parameter(&builder, 0, + ShapeUtil::MakeShape(F32, {static_cast(input.size())}), + "param"), F16); ComputeAndCompareR1(&builder, expected_output, {dot_lhs_handle.get()}); @@ -426,28 +426,28 @@ XLA_TEST_F(ConvertTest, ConvertR1F32ToR1F16) { XLA_TEST_F(ConvertTest, ConvertC64ToC64) { XlaBuilder builder(TestName()); std::vector x = {{42.0f, 64.0f}}; - builder.ConvertElementType(builder.ConstantR1(x), C64); + ConvertElementType(ConstantR1(&builder, x), C64); ComputeAndCompareR1(&builder, x, {}, ErrorSpec(0.0001)); } XLA_TEST_F(ConvertTest, ConvertS64S64) { XlaBuilder builder(TestName()); std::vector x = {{-42, 64}}; - builder.ConvertElementType(builder.ConstantR1(x), S64); + ConvertElementType(ConstantR1(&builder, x), S64); ComputeAndCompareR1(&builder, x, {}); } XLA_TEST_F(ConvertTest, ConvertU64U64) { XlaBuilder builder(TestName()); std::vector x = {{42, 64}}; - builder.ConvertElementType(builder.ConstantR1(x), U64); + ConvertElementType(ConstantR1(&builder, x), U64); ComputeAndCompareR1(&builder, x, {}); } XLA_TEST_F(ConvertTest, ConvertU64S64) { XlaBuilder builder(TestName()); std::vector unsigned_x = {{42, UINT64_MAX}}; - builder.ConvertElementType(builder.ConstantR1(unsigned_x), S64); + ConvertElementType(ConstantR1(&builder, unsigned_x), S64); std::vector signed_x = {{42, -1}}; ComputeAndCompareR1(&builder, signed_x, {}); } @@ -455,7 +455,7 @@ XLA_TEST_F(ConvertTest, ConvertU64S64) { XLA_TEST_F(ConvertTest, ConvertS64U64) { XlaBuilder builder(TestName()); std::vector signed_x = {{42, -1, INT64_MIN}}; - builder.ConvertElementType(builder.ConstantR1(signed_x), U64); + ConvertElementType(ConstantR1(&builder, signed_x), U64); std::vector unsigned_x = { {42, UINT64_MAX, tensorflow::MathUtil::IPow(2, 63)}}; ComputeAndCompareR1(&builder, unsigned_x, {}); @@ -475,10 +475,9 @@ XLA_TEST_F(ConvertTest, ConvertBF16F32) { } // Exhaustively test all bf16 to f32 conversions. - xla::XlaOp all_bfloats_bf16 = builder.ConstantR1(all_bfloats); - xla::XlaOp all_bfloats_f32 = - builder.ConvertElementType(all_bfloats_bf16, F32); - xla::XlaOp all_bfloats_u32 = builder.BitcastConvertType(all_bfloats_f32, U32); + xla::XlaOp all_bfloats_bf16 = ConstantR1(&builder, all_bfloats); + xla::XlaOp all_bfloats_f32 = ConvertElementType(all_bfloats_bf16, F32); + BitcastConvertType(all_bfloats_f32, U32); ComputeAndCompareR1(&builder, expected, {}); } diff --git a/tensorflow/compiler/xla/tests/convolution_dimension_numbers_test.cc b/tensorflow/compiler/xla/tests/convolution_dimension_numbers_test.cc index b5a42e305987df030c15d089f5877f73bb61de1b..944366410b14439aa33999185525f1029735e95b 100644 --- a/tensorflow/compiler/xla/tests/convolution_dimension_numbers_test.cc +++ b/tensorflow/compiler/xla/tests/convolution_dimension_numbers_test.cc @@ -93,14 +93,15 @@ XLA_TEST_F(ConvolutionDimensionNumbersTest, auto weight_array = MakeUnique>(4, 3, 1, 1); weight_array->FillWithMultiples(0.2); auto weight_data = - client_->TransferToServer(*Literal::CreateR4FromArray4D(*weight_array)) + client_ + ->TransferToServer(*LiteralUtil::CreateR4FromArray4D(*weight_array)) .ConsumeValueOrDie(); XlaBuilder builder(TestName()); - auto input = builder.ConstantR4FromArray4D(*input_array); + auto input = ConstantR4FromArray4D(&builder, *input_array); auto weight = - builder.Parameter(0, ShapeUtil::MakeShape(F32, {4, 3, 1, 1}), "weight"); - auto conv1 = builder.Conv(input, weight, {1, 1}, Padding::kValid); + Parameter(&builder, 0, ShapeUtil::MakeShape(F32, {4, 3, 1, 1}), "weight"); + auto conv1 = Conv(input, weight, {1, 1}, Padding::kValid); ConvolutionDimensionNumbers dim_nums = XlaBuilder::CreateDefaultConvDimensionNumbers(); @@ -117,8 +118,7 @@ XLA_TEST_F(ConvolutionDimensionNumbersTest, dim_nums.set_kernel_input_feature_dimension( dim_nums.kernel_output_feature_dimension()); dim_nums.set_kernel_output_feature_dimension(old_kernel_input_feature_dim); - builder.ConvWithGeneralDimensions(input, conv1, {1, 1}, Padding::kValid, - dim_nums); + ConvWithGeneralDimensions(input, conv1, {1, 1}, Padding::kValid, dim_nums); auto expected_conv1 = ReferenceUtil::ConvArray4D(*input_array, *weight_array, {1, 1}, Padding::kValid); diff --git a/tensorflow/compiler/xla/tests/convolution_test.cc b/tensorflow/compiler/xla/tests/convolution_test.cc index 346bb3a3996ee5bf662b0f74dd0c2096efbf5295..a8b8f74ca9603a71acefc0be2141d7b9caf2b73b 100644 --- a/tensorflow/compiler/xla/tests/convolution_test.cc +++ b/tensorflow/compiler/xla/tests/convolution_test.cc @@ -25,7 +25,7 @@ limitations under the License. #include "tensorflow/compiler/xla/client/padding.h" #include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/compiler/xla/layout_util.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/ptr_util.h" #include "tensorflow/compiler/xla/reference_util.h" #include "tensorflow/compiler/xla/shape_util.h" @@ -89,9 +89,9 @@ class ForwardPassConvolution_3x3x256_256_OutputZ_Iota : public ConvolutionTest { ASSERT_EQ(2, arhs->height()); XlaBuilder builder(TestName()); - auto lhs = builder.ConstantR4FromArray4D(*alhs); - auto rhs = builder.ConstantR4FromArray4D(*arhs); - builder.Conv(lhs, rhs, {1, 1}, Padding::kValid); + auto lhs = ConstantR4FromArray4D(&builder, *alhs); + auto rhs = ConstantR4FromArray4D(&builder, *arhs); + Conv(lhs, rhs, {1, 1}, Padding::kValid); ComputeAndCompare(&builder, {}, error_spec_); } @@ -109,9 +109,9 @@ class Convolve_1x1x1x2_1x1x1x2_Valid : public ConvolutionTest { XlaBuilder builder(TestName()); Shape input_shape = ShapeUtil::MakeShapeWithType({1, 1, 1, 2}); Shape filter_shape = ShapeUtil::MakeShapeWithType({1, 1, 1, 2}); - auto input = builder.Parameter(0, input_shape, "input"); - auto filter = builder.Parameter(1, filter_shape, "filter"); - builder.Conv(input, filter, {1, 1}, Padding::kValid); + auto input = Parameter(&builder, 0, input_shape, "input"); + auto filter = Parameter(&builder, 1, filter_shape, "filter"); + Conv(input, filter, {1, 1}, Padding::kValid); Array4D input_data(1, 1, 1, 2); input_data.FillWithYX(Array2D({ @@ -123,8 +123,8 @@ class Convolve_1x1x1x2_1x1x1x2_Valid : public ConvolutionTest { })); ComputeAndCompare(&builder, - {std::move(*Literal::CreateFromArray(input_data)), - std::move(*Literal::CreateFromArray(filter_data))}, + {std::move(*LiteralUtil::CreateFromArray(input_data)), + std::move(*LiteralUtil::CreateFromArray(filter_data))}, error_spec_); } }; @@ -140,9 +140,9 @@ class Convolve_1x1x4x4_1x1x2x2_Valid : public ConvolutionTest { XlaBuilder builder(TestName()); Shape input_shape = ShapeUtil::MakeShapeWithType({1, 1, 4, 4}); Shape filter_shape = ShapeUtil::MakeShapeWithType({1, 1, 2, 2}); - auto input = builder.Parameter(0, input_shape, "input"); - auto filter = builder.Parameter(1, filter_shape, "filter"); - builder.Conv(input, filter, {1, 1}, Padding::kValid); + auto input = Parameter(&builder, 0, input_shape, "input"); + auto filter = Parameter(&builder, 1, filter_shape, "filter"); + Conv(input, filter, {1, 1}, Padding::kValid); Array4D input_data(1, 1, 4, 4); input_data.FillWithYX(Array2D({ @@ -157,8 +157,8 @@ class Convolve_1x1x4x4_1x1x2x2_Valid : public ConvolutionTest { {7.0f, 8.0f}, })); ComputeAndCompare(&builder, - {std::move(*Literal::CreateFromArray(input_data)), - std::move(*Literal::CreateFromArray(filter_data))}, + {std::move(*LiteralUtil::CreateFromArray(input_data)), + std::move(*LiteralUtil::CreateFromArray(filter_data))}, error_spec_); } }; @@ -174,9 +174,9 @@ class Convolve_1x1x4x4_1x1x2x2_Same : public ConvolutionTest { XlaBuilder builder(TestName()); Shape input_shape = ShapeUtil::MakeShapeWithType({1, 1, 4, 4}); Shape filter_shape = ShapeUtil::MakeShapeWithType({1, 1, 2, 2}); - auto input = builder.Parameter(0, input_shape, "input"); - auto filter = builder.Parameter(1, filter_shape, "filter"); - builder.Conv(input, filter, {1, 1}, Padding::kSame); + auto input = Parameter(&builder, 0, input_shape, "input"); + auto filter = Parameter(&builder, 1, filter_shape, "filter"); + Conv(input, filter, {1, 1}, Padding::kSame); Array4D input_data(1, 1, 4, 4); input_data.FillWithYX(Array2D({ @@ -192,8 +192,8 @@ class Convolve_1x1x4x4_1x1x2x2_Same : public ConvolutionTest { })); ComputeAndCompare(&builder, - {std::move(*Literal::CreateFromArray(input_data)), - std::move(*Literal::CreateFromArray(filter_data))}, + {std::move(*LiteralUtil::CreateFromArray(input_data)), + std::move(*LiteralUtil::CreateFromArray(filter_data))}, error_spec_); } }; @@ -210,9 +210,9 @@ class Convolve_1x1x4x4_1x1x3x3_Same : public ConvolutionTest { XlaBuilder builder(TestName()); Shape input_shape = ShapeUtil::MakeShapeWithType({1, 1, 4, 4}); Shape filter_shape = ShapeUtil::MakeShapeWithType({1, 1, 3, 3}); - auto input = builder.Parameter(0, input_shape, "input"); - auto filter = builder.Parameter(1, filter_shape, "filter"); - builder.Conv(input, filter, {1, 1}, Padding::kSame); + auto input = Parameter(&builder, 0, input_shape, "input"); + auto filter = Parameter(&builder, 1, filter_shape, "filter"); + Conv(input, filter, {1, 1}, Padding::kSame); Array4D input_data(1, 1, 4, 4); input_data.FillWithYX(Array2D({{1.0f, 2.0f, 3.0f, 4.0f}, @@ -224,8 +224,8 @@ class Convolve_1x1x4x4_1x1x3x3_Same : public ConvolutionTest { {{5.0f, 6.0f, 7.0f}, {8.0f, 9.0f, 10.0f}, {11.0f, 12.0f, 13.0f}})); // clang-format on ComputeAndCompare(&builder, - {std::move(*Literal::CreateFromArray(input_data)), - std::move(*Literal::CreateFromArray(filter_data))}, + {std::move(*LiteralUtil::CreateFromArray(input_data)), + std::move(*LiteralUtil::CreateFromArray(filter_data))}, error_spec_); } }; @@ -238,9 +238,9 @@ XLA_TEST_F(ConvolutionTest, Convolve1D_1x2x5_1x2x2_Valid) { { Shape input_shape = ShapeUtil::MakeShape(F32, {1, 2, 5}); Shape filter_shape = ShapeUtil::MakeShape(F32, {1, 2, 2}); - auto input = builder.Parameter(0, input_shape, "input"); - auto filter = builder.Parameter(1, filter_shape, "filter"); - builder.Conv(input, filter, {1}, Padding::kValid); + auto input = Parameter(&builder, 0, input_shape, "input"); + auto filter = Parameter(&builder, 1, filter_shape, "filter"); + Conv(input, filter, {1}, Padding::kValid); } Array3D input({{{1, 2, 3, 4, 5}, {6, 7, 8, 9, 10}}}); @@ -249,10 +249,10 @@ XLA_TEST_F(ConvolutionTest, Convolve1D_1x2x5_1x2x2_Valid) { Array3D expected({{{510, 610, 710, 810}}}); auto input_literal = - client_->TransferToServer(*Literal::CreateR3FromArray3D(input)) + client_->TransferToServer(*LiteralUtil::CreateR3FromArray3D(input)) .ConsumeValueOrDie(); auto filter_literal = - client_->TransferToServer(*Literal::CreateR3FromArray3D(filter)) + client_->TransferToServer(*LiteralUtil::CreateR3FromArray3D(filter)) .ConsumeValueOrDie(); ComputeAndCompareR3(&builder, expected, @@ -268,10 +268,10 @@ class Convolve1D_1x2x5_1x2x2_WithRHSDilation : public ConvolutionTest { { Shape input_shape = ShapeUtil::MakeShapeWithType({1, 2, 5}); Shape filter_shape = ShapeUtil::MakeShapeWithType({1, 2, 2}); - auto input = builder.Parameter(0, input_shape, "input"); - auto filter = builder.Parameter(1, filter_shape, "filter"); + auto input = Parameter(&builder, 0, input_shape, "input"); + auto filter = Parameter(&builder, 1, filter_shape, "filter"); // Convolution dimensions are bf0_oi0->bo0. - builder.ConvGeneralDilated( + ConvGeneralDilated( input, filter, /*window_strides=*/{1}, /*padding=*/{{0, 0}}, /*lhs_dilation=*/{1}, /*rhs_dilation=*/{2}, /*dimension_numbers=*/builder.CreateDefaultConvDimensionNumbers(1)); @@ -284,10 +284,10 @@ class Convolve1D_1x2x5_1x2x2_WithRHSDilation : public ConvolutionTest { Array3D expected({{{570.0f, 670.0f, 770.0f}}}); auto input_literal = - client_->TransferToServer(*Literal::CreateR3FromArray3D(input)) + client_->TransferToServer(*LiteralUtil::CreateR3FromArray3D(input)) .ConsumeValueOrDie(); auto filter_literal = - client_->TransferToServer(*Literal::CreateR3FromArray3D(filter)) + client_->TransferToServer(*LiteralUtil::CreateR3FromArray3D(filter)) .ConsumeValueOrDie(); ComputeAndCompareR3(&builder, expected, @@ -304,10 +304,10 @@ XLA_TEST_F(ConvolutionTest, Convolve1D_1x2x5_1x2x2_WithLHSDilation) { { Shape input_shape = ShapeUtil::MakeShape(F32, {1, 2, 5}); Shape filter_shape = ShapeUtil::MakeShape(F32, {1, 2, 2}); - auto input = builder.Parameter(0, input_shape, "input"); - auto filter = builder.Parameter(1, filter_shape, "filter"); + auto input = Parameter(&builder, 0, input_shape, "input"); + auto filter = Parameter(&builder, 1, filter_shape, "filter"); // Convolution dimensions are bf0_oi0->bo0. - builder.ConvGeneralDilated( + ConvGeneralDilated( input, filter, /*window_strides=*/{1}, /*padding=*/{{0, 0}}, /*lhs_dilation=*/{2}, /*rhs_dilation=*/{1}, /*dimension_numbers=*/builder.CreateDefaultConvDimensionNumbers(1)); @@ -319,10 +319,10 @@ XLA_TEST_F(ConvolutionTest, Convolve1D_1x2x5_1x2x2_WithLHSDilation) { Array3D expected({{{190, 320, 230, 380, 270, 440, 310, 500}}}); auto input_literal = - client_->TransferToServer(*Literal::CreateR3FromArray3D(input)) + client_->TransferToServer(*LiteralUtil::CreateR3FromArray3D(input)) .ConsumeValueOrDie(); auto filter_literal = - client_->TransferToServer(*Literal::CreateR3FromArray3D(filter)) + client_->TransferToServer(*LiteralUtil::CreateR3FromArray3D(filter)) .ConsumeValueOrDie(); ComputeAndCompareR3(&builder, expected, @@ -335,10 +335,10 @@ XLA_TEST_F(ConvolutionTest, Convolve1D_1x2x5_1x2x2_WithLHSAndRHSDilation) { { Shape input_shape = ShapeUtil::MakeShape(F32, {1, 2, 5}); Shape filter_shape = ShapeUtil::MakeShape(F32, {1, 2, 2}); - auto input = builder.Parameter(0, input_shape, "input"); - auto filter = builder.Parameter(1, filter_shape, "filter"); + auto input = Parameter(&builder, 0, input_shape, "input"); + auto filter = Parameter(&builder, 1, filter_shape, "filter"); // Convolution dimensions are bf0_oi0->bo0. - builder.ConvGeneralDilated( + ConvGeneralDilated( input, filter, /*window_strides=*/{1}, /*padding=*/{{0, 0}}, /*lhs_dilation=*/{2}, /*rhs_dilation=*/{2}, /*dimension_numbers=*/builder.CreateDefaultConvDimensionNumbers(1)); @@ -350,10 +350,10 @@ XLA_TEST_F(ConvolutionTest, Convolve1D_1x2x5_1x2x2_WithLHSAndRHSDilation) { Array3D expected({{{510, 0, 610, 0, 710, 0, 810}}}); auto input_literal = - client_->TransferToServer(*Literal::CreateR3FromArray3D(input)) + client_->TransferToServer(*LiteralUtil::CreateR3FromArray3D(input)) .ConsumeValueOrDie(); auto filter_literal = - client_->TransferToServer(*Literal::CreateR3FromArray3D(filter)) + client_->TransferToServer(*LiteralUtil::CreateR3FromArray3D(filter)) .ConsumeValueOrDie(); ComputeAndCompareR3(&builder, expected, @@ -369,10 +369,10 @@ class Convolve1D_1x2x5_1x2x2_WithPadding : public ConvolutionTest { { Shape input_shape = ShapeUtil::MakeShapeWithType({1, 2, 5}); Shape filter_shape = ShapeUtil::MakeShapeWithType({1, 2, 2}); - auto input = builder.Parameter(0, input_shape, "input"); - auto filter = builder.Parameter(1, filter_shape, "filter"); + auto input = Parameter(&builder, 0, input_shape, "input"); + auto filter = Parameter(&builder, 1, filter_shape, "filter"); // Convolution dimensions are bf0_oi0->bo0. - builder.ConvGeneralDilated( + ConvGeneralDilated( input, filter, /*window_strides=*/{1}, /*padding=*/{{2, 2}}, /*lhs_dilation=*/{1}, /*rhs_dilation=*/{1}, /*dimension_numbers=*/builder.CreateDefaultConvDimensionNumbers(1)); @@ -386,10 +386,10 @@ class Convolve1D_1x2x5_1x2x2_WithPadding : public ConvolutionTest { {{{0.0f, 260.0f, 510.0f, 610.0f, 710.0f, 810.0f, 350.0f, 0.0f}}}); auto input_literal = - client_->TransferToServer(*Literal::CreateR3FromArray3D(input)) + client_->TransferToServer(*LiteralUtil::CreateR3FromArray3D(input)) .ConsumeValueOrDie(); auto filter_literal = - client_->TransferToServer(*Literal::CreateR3FromArray3D(filter)) + client_->TransferToServer(*LiteralUtil::CreateR3FromArray3D(filter)) .ConsumeValueOrDie(); ComputeAndCompareR3(&builder, expected, @@ -408,8 +408,8 @@ XLA_TEST_F(ConvolutionTest, Convolve3D_1x4x2x3x3_2x2x2x3x3_Valid) { Shape input_shape = ShapeUtil::MakeShape(F32, input_dims); Shape filter_shape = ShapeUtil::MakeShape(F32, filter_dims); { - auto input = builder.Parameter(0, input_shape, "input"); - auto filter = builder.Parameter(1, filter_shape, "filter"); + auto input = Parameter(&builder, 0, input_shape, "input"); + auto filter = Parameter(&builder, 1, filter_shape, "filter"); // Tensorflow dimension numbers for 3D convolution. ConvolutionDimensionNumbers dnums; @@ -429,21 +429,20 @@ XLA_TEST_F(ConvolutionTest, Convolve3D_1x4x2x3x3_2x2x2x3x3_Valid) { dnums.set_kernel_input_feature_dimension(3); dnums.set_kernel_output_feature_dimension(4); - builder.ConvWithGeneralDimensions(input, filter, {1, 1, 1}, Padding::kValid, - dnums); + ConvWithGeneralDimensions(input, filter, {1, 1, 1}, Padding::kValid, dnums); } std::vector input_elems(ShapeUtil::ElementsIn(input_shape)); iota(input_elems.begin(), input_elems.end(), 1.0f); - auto input_r1 = Literal::CreateR1(input_elems); + auto input_r1 = LiteralUtil::CreateR1(input_elems); auto input_r5 = input_r1->Reshape(input_dims).ConsumeValueOrDie(); std::vector filter_elems(ShapeUtil::ElementsIn(filter_shape)); iota(filter_elems.begin(), filter_elems.end(), 1.0f); - auto filter_r1 = Literal::CreateR1(filter_elems); + auto filter_r1 = LiteralUtil::CreateR1(filter_elems); auto filter_r5 = filter_r1->Reshape(filter_dims).ConsumeValueOrDie(); - auto expected_r1 = Literal::CreateR1( + auto expected_r1 = LiteralUtil::CreateR1( {19554, 19962, 20370, 22110, 22590, 23070, 34890, 35730, 36570, 37446, 38358, 39270, 50226, 51498, 52770, 52782, 54126, 55470}); auto expected_r5 = expected_r1->Reshape({1, 3, 1, 2, 3}).ConsumeValueOrDie(); @@ -475,8 +474,8 @@ class Convolve2D_1x3x3x5_3x3x5x5_Valid : public ConvolutionTest { Shape input_shape = ShapeUtil::MakeShapeWithType(input_dims); Shape filter_shape = ShapeUtil::MakeShapeWithType(filter_dims); { - auto input = builder.Parameter(0, input_shape, "input"); - auto filter = builder.Parameter(1, filter_shape, "filter"); + auto input = Parameter(&builder, 0, input_shape, "input"); + auto filter = Parameter(&builder, 1, filter_shape, "filter"); // Tensorflow dimension numbers for 2D convolution. ConvolutionDimensionNumbers dnums; @@ -493,21 +492,20 @@ class Convolve2D_1x3x3x5_3x3x5x5_Valid : public ConvolutionTest { dnums.set_kernel_input_feature_dimension(2); dnums.set_kernel_output_feature_dimension(3); - builder.ConvWithGeneralDimensions(input, filter, {1, 1}, Padding::kValid, - dnums); + ConvWithGeneralDimensions(input, filter, {1, 1}, Padding::kValid, dnums); } std::vector input_elems(ShapeUtil::ElementsIn(input_shape)); iota_int_init_value(input_elems, 1); - auto input_r1 = Literal::CreateR1(input_elems); + auto input_r1 = LiteralUtil::CreateR1(input_elems); auto input_r4 = input_r1->Reshape(input_dims).ConsumeValueOrDie(); std::vector filter_elems(ShapeUtil::ElementsIn(filter_shape)); iota_int_init_value(filter_elems, 1); - auto filter_r1 = Literal::CreateR1(filter_elems); + auto filter_r1 = LiteralUtil::CreateR1(filter_elems); auto filter_r4 = filter_r1->Reshape(filter_dims).ConsumeValueOrDie(); - auto expected_r1 = Literal::CreateR1( + auto expected_r1 = LiteralUtil::CreateR1( {static_cast(92115), static_cast(93150), static_cast(94185)}); auto expected_r4 = expected_r1->Reshape({1, 1, 1, 3}).ConsumeValueOrDie(); @@ -541,8 +539,8 @@ XLA_TEST_P(ConvolveWithAndWithoutCanonicalization, Shape input_shape = ShapeUtil::MakeShape(F32, {4, 29}); Shape filter_shape = ShapeUtil::MakeShape(F32, {4, 10}); - auto input = builder.Parameter(0, input_shape, "input"); - auto filter = builder.Parameter(1, filter_shape, "filter"); + auto input = Parameter(&builder, 0, input_shape, "input"); + auto filter = Parameter(&builder, 1, filter_shape, "filter"); ConvolutionDimensionNumbers dnums; dnums.set_input_feature_dimension(0); @@ -551,7 +549,7 @@ XLA_TEST_P(ConvolveWithAndWithoutCanonicalization, dnums.set_kernel_output_feature_dimension(1); dnums.set_output_batch_dimension(0); dnums.set_output_feature_dimension(1); - builder.ConvWithGeneralDimensions(input, filter, {}, Padding::kValid, dnums); + ConvWithGeneralDimensions(input, filter, {}, Padding::kValid, dnums); Array2D param0(4, 29); param0.FillUnique(); @@ -563,8 +561,8 @@ XLA_TEST_P(ConvolveWithAndWithoutCanonicalization, expected_result.Fill(0); ComputeAndCompare(&builder, - {std::move(*Literal::CreateFromArray(param0)), - std::move(*Literal::CreateFromArray(param1))}, + {std::move(*LiteralUtil::CreateFromArray(param0)), + std::move(*LiteralUtil::CreateFromArray(param1))}, error_spec_); } @@ -599,8 +597,8 @@ class Convolve1D1WindowTestBase Shape input_shape = ShapeUtil::MakeShapeWithType(input_dims); Shape filter_shape = ShapeUtil::MakeShapeWithType(filter_dims); { - auto input = builder.Parameter(0, input_shape, "input"); - auto filter = builder.Parameter(1, filter_shape, "filter"); + auto input = Parameter(&builder, 0, input_shape, "input"); + auto filter = Parameter(&builder, 1, filter_shape, "filter"); // Tensorflow dimension numbers for 1D convolution. ConvolutionDimensionNumbers dnums; @@ -614,24 +612,23 @@ class Convolve1D1WindowTestBase dnums.set_kernel_input_feature_dimension(1); dnums.set_kernel_output_feature_dimension(2); - builder.ConvWithGeneralDimensions(input, filter, {1}, Padding::kValid, - dnums); + ConvWithGeneralDimensions(input, filter, {1}, Padding::kValid, dnums); } std::vector input_elems(ShapeUtil::ElementsIn(input_shape), static_cast(1.0f)); - auto input_r1 = Literal::CreateR1(input_elems); + auto input_r1 = LiteralUtil::CreateR1(input_elems); auto input_r3 = input_r1->Reshape(input_dims).ConsumeValueOrDie(); std::vector filter_elems(ShapeUtil::ElementsIn(filter_shape), static_cast(1.0f)); - auto filter_r1 = Literal::CreateR1(filter_elems); + auto filter_r1 = LiteralUtil::CreateR1(filter_elems); auto filter_r3 = filter_r1->Reshape(filter_dims).ConsumeValueOrDie(); std::vector expect_elems(batch * output_feature * num_windows, static_cast(window_size * input_feature)); - auto expected_r1 = Literal::CreateR1(expect_elems); + auto expected_r1 = LiteralUtil::CreateR1(expect_elems); auto expected_r3 = expected_r1->Reshape({batch, num_windows, output_feature}) .ConsumeValueOrDie(); @@ -726,9 +723,9 @@ XLA_TEST_F(ConvolutionTest, Convolve_bf16_1x1x1x2_1x1x1x2_Valid) { XlaBuilder builder(TestName()); Shape input_shape = ShapeUtil::MakeShape(BF16, {1, 1, 1, 2}); Shape filter_shape = ShapeUtil::MakeShape(BF16, {1, 1, 1, 2}); - auto input = builder.Parameter(0, input_shape, "input"); - auto filter = builder.Parameter(1, filter_shape, "filter"); - builder.Conv(input, filter, {1, 1}, Padding::kValid); + auto input = Parameter(&builder, 0, input_shape, "input"); + auto filter = Parameter(&builder, 1, filter_shape, "filter"); + Conv(input, filter, {1, 1}, Padding::kValid); Array4D input_data(1, 1, 1, 2); input_data.FillWithYX(Array2D({ @@ -740,8 +737,8 @@ XLA_TEST_F(ConvolutionTest, Convolve_bf16_1x1x1x2_1x1x1x2_Valid) { })); ComputeAndCompare(&builder, - {std::move(*Literal::CreateFromArray(input_data)), - std::move(*Literal::CreateFromArray(filter_data))}, + {std::move(*LiteralUtil::CreateFromArray(input_data)), + std::move(*LiteralUtil::CreateFromArray(filter_data))}, error_spec_); } @@ -754,9 +751,9 @@ XLA_TEST_F(ConvolutionTest, NoCudnnAlgorithmPicker) { XlaBuilder builder(TestName()); Shape input_shape = ShapeUtil::MakeShape(F32, {1, 1, 1, 2}); Shape filter_shape = ShapeUtil::MakeShape(F32, {1, 1, 1, 2}); - auto input = builder.Parameter(0, input_shape, "input"); - auto filter = builder.Parameter(1, filter_shape, "filter"); - builder.Conv(input, filter, {1, 1}, Padding::kValid); + auto input = Parameter(&builder, 0, input_shape, "input"); + auto filter = Parameter(&builder, 1, filter_shape, "filter"); + Conv(input, filter, {1, 1}, Padding::kValid); Array4D input_data(1, 1, 1, 2); input_data.FillIota(0); @@ -764,8 +761,8 @@ XLA_TEST_F(ConvolutionTest, NoCudnnAlgorithmPicker) { filter_data.FillIota(10); ComputeAndCompare(&builder, - {std::move(*Literal::CreateFromArray(input_data)), - std::move(*Literal::CreateFromArray(filter_data))}); + {std::move(*LiteralUtil::CreateFromArray(input_data)), + std::move(*LiteralUtil::CreateFromArray(filter_data))}); } } // namespace diff --git a/tensorflow/compiler/xla/tests/convolution_variants_test.cc b/tensorflow/compiler/xla/tests/convolution_variants_test.cc index fea850dc135e33fe098aa755c6fdd93319cd2837..8792e7781b17465d94ae8ac8375a4523f368d720 100644 --- a/tensorflow/compiler/xla/tests/convolution_variants_test.cc +++ b/tensorflow/compiler/xla/tests/convolution_variants_test.cc @@ -28,7 +28,7 @@ limitations under the License. #include "tensorflow/compiler/xla/client/local_client.h" #include "tensorflow/compiler/xla/client/padding.h" #include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/reference_util.h" #include "tensorflow/compiler/xla/tests/client_library_test_base.h" #include "tensorflow/compiler/xla/tests/literal_test_util.h" @@ -55,12 +55,12 @@ XLA_TEST_F(ConvolutionVariantsTest, Minimal) { XlaBuilder builder(TestName()); const Array4D input_array(1, 1, 1, 1, {2}); - auto input = builder.ConstantR4FromArray4D(input_array); + auto input = ConstantR4FromArray4D(&builder, input_array); const Array4D filter_array(1, 1, 1, 1, {3}); - auto filter = builder.ConstantR4FromArray4D(filter_array); + auto filter = ConstantR4FromArray4D(&builder, filter_array); - builder.Conv(input, filter, {1, 1}, Padding::kValid); + Conv(input, filter, {1, 1}, Padding::kValid); const Array4D expected(1, 1, 1, 1, {6}); ComputeAndCompareR4(&builder, expected, {}, error_spec_); @@ -70,12 +70,12 @@ XLA_TEST_F(ConvolutionVariantsTest, MinimalWithBatch) { XlaBuilder builder(TestName()); const Array4D input_array(5, 1, 1, 1, {1, 2, 3, 4, 5}); - auto input = builder.ConstantR4FromArray4D(input_array); + auto input = ConstantR4FromArray4D(&builder, input_array); const Array4D filter_array(1, 1, 1, 1, {2}); - auto filter = builder.ConstantR4FromArray4D(filter_array); + auto filter = ConstantR4FromArray4D(&builder, filter_array); - builder.Conv(input, filter, {1, 1}, Padding::kValid); + Conv(input, filter, {1, 1}, Padding::kValid); const Array4D expected(5, 1, 1, 1, {2, 4, 6, 8, 10}); ComputeAndCompareR4(&builder, expected, {}, error_spec_); @@ -86,12 +86,12 @@ XLA_TEST_F(ConvolutionVariantsTest, Flat1x1) { Array4D input_array(2, 1, 3, 4); input_array.FillWithMultiples(1); - auto input = builder.ConstantR4FromArray4D(input_array); + auto input = ConstantR4FromArray4D(&builder, input_array); const Array4D filter_array(1, 1, 1, 1, {2.3}); - auto filter = builder.ConstantR4FromArray4D(filter_array); + auto filter = ConstantR4FromArray4D(&builder, filter_array); - builder.Conv(input, filter, {1, 1}, Padding::kValid); + Conv(input, filter, {1, 1}, Padding::kValid); Array4D expected(2, 1, 3, 4); expected.FillWithMultiples(2.3); @@ -102,12 +102,12 @@ XLA_TEST_F(ConvolutionVariantsTest, Deep1x1) { XlaBuilder builder(TestName()); Array4D input_array(1, 2, 1, 1, {10, 1}); - auto input = builder.ConstantR4FromArray4D(input_array); + auto input = ConstantR4FromArray4D(&builder, input_array); const Array4D filter_array(3, 2, 1, 1, {1, 2, 3, 4, 5, 6}); - auto filter = builder.ConstantR4FromArray4D(filter_array); + auto filter = ConstantR4FromArray4D(&builder, filter_array); - builder.Conv(input, filter, {1, 1}, Padding::kValid); + Conv(input, filter, {1, 1}, Padding::kValid); Array4D expected(1, 3, 1, 1, {12, 34, 56}); ComputeAndCompareR4(&builder, expected, {}, error_spec_); @@ -117,12 +117,12 @@ XLA_TEST_F(ConvolutionVariantsTest, Filter1x2in1x2) { XlaBuilder builder(TestName()); Array4D input_array(1, 1, 1, 2, {1, 2}); - auto input = builder.ConstantR4FromArray4D(input_array); + auto input = ConstantR4FromArray4D(&builder, input_array); const Array4D filter_array(1, 1, 1, 2, {10, 1}); - auto filter = builder.ConstantR4FromArray4D(filter_array); + auto filter = ConstantR4FromArray4D(&builder, filter_array); - builder.Conv(input, filter, {1, 1}, Padding::kValid); + Conv(input, filter, {1, 1}, Padding::kValid); Array4D expected(1, 1, 1, 1, {12}); ComputeAndCompareR4(&builder, expected, {}, error_spec_); @@ -132,12 +132,12 @@ XLA_TEST_F(ConvolutionVariantsTest, Filter1x2in1x3) { XlaBuilder builder(TestName()); Array4D input_array(1, 1, 1, 3, {1, 2, 3}); - auto input = builder.ConstantR4FromArray4D(input_array); + auto input = ConstantR4FromArray4D(&builder, input_array); const Array4D filter_array(1, 1, 1, 2, {10, 1}); - auto filter = builder.ConstantR4FromArray4D(filter_array); + auto filter = ConstantR4FromArray4D(&builder, filter_array); - builder.Conv(input, filter, {1, 1}, Padding::kValid); + Conv(input, filter, {1, 1}, Padding::kValid); Array4D expected(1, 1, 1, 2, {12, 23}); ComputeAndCompareR4(&builder, expected, {}, error_spec_); @@ -147,12 +147,12 @@ XLA_TEST_F(ConvolutionVariantsTest, Filter1x2in2x2) { XlaBuilder builder(TestName()); Array4D input_array(1, 1, 2, 2, {1, 2, 3, 4}); - auto input = builder.ConstantR4FromArray4D(input_array); + auto input = ConstantR4FromArray4D(&builder, input_array); const Array4D filter_array(1, 1, 1, 2, {10, 1}); - auto filter = builder.ConstantR4FromArray4D(filter_array); + auto filter = ConstantR4FromArray4D(&builder, filter_array); - builder.Conv(input, filter, {1, 1}, Padding::kValid); + Conv(input, filter, {1, 1}, Padding::kValid); Array4D expected(1, 1, 2, 1, {12, 34}); ComputeAndCompareR4(&builder, expected, {}, error_spec_); @@ -162,12 +162,12 @@ XLA_TEST_F(ConvolutionVariantsTest, Filter2x1in2x2) { XlaBuilder builder(TestName()); Array4D input_array(1, 1, 2, 2, {1, 2, 3, 4}); - auto input = builder.ConstantR4FromArray4D(input_array); + auto input = ConstantR4FromArray4D(&builder, input_array); const Array4D filter_array(1, 1, 2, 1, {10, 1}); - auto filter = builder.ConstantR4FromArray4D(filter_array); + auto filter = ConstantR4FromArray4D(&builder, filter_array); - builder.Conv(input, filter, {1, 1}, Padding::kValid); + Conv(input, filter, {1, 1}, Padding::kValid); Array4D expected(1, 1, 1, 2, {13, 24}); ComputeAndCompareR4(&builder, expected, {}, error_spec_); @@ -177,12 +177,12 @@ XLA_TEST_F(ConvolutionVariantsTest, Filter2x2in2x2) { XlaBuilder builder(TestName()); Array4D input_array(1, 1, 2, 2, {1, 2, 3, 4}); - auto input = builder.ConstantR4FromArray4D(input_array); + auto input = ConstantR4FromArray4D(&builder, input_array); const Array4D filter_array(1, 1, 2, 2, {1000, 100, 10, 1}); - auto filter = builder.ConstantR4FromArray4D(filter_array); + auto filter = ConstantR4FromArray4D(&builder, filter_array); - builder.Conv(input, filter, {1, 1}, Padding::kValid); + Conv(input, filter, {1, 1}, Padding::kValid); Array4D expected(1, 1, 1, 1, {1234}); ComputeAndCompareR4(&builder, expected, {}, error_spec_); @@ -194,13 +194,13 @@ XLA_TEST_F(ConvolutionVariantsTest, Filter1x2in2x3WithDepthAndBatch) { Array4D input_array( 2, 2, 2, 3, {0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 0, 0, // plane 0 0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 0, 0}); // plane 1 - auto input = builder.ConstantR4FromArray4D(input_array); + auto input = ConstantR4FromArray4D(&builder, input_array); const Array4D filter_array( 2, 2, 1, 2, {1000, 100, 10, 1, 0.1, 0.01, 0.001, 0.0001}); - auto filter = builder.ConstantR4FromArray4D(filter_array); + auto filter = ConstantR4FromArray4D(&builder, filter_array); - builder.Conv(input, filter, {1, 1}, Padding::kValid); + Conv(input, filter, {1, 1}, Padding::kValid); Array4D expected( 2, 2, 2, 2, @@ -213,12 +213,12 @@ XLA_TEST_F(ConvolutionVariantsTest, Filter1x1stride1x2in1x4) { XlaBuilder builder(TestName()); Array4D input_array(1, 1, 1, 4, {1, 2, 3, 4}); - auto input = builder.ConstantR4FromArray4D(input_array); + auto input = ConstantR4FromArray4D(&builder, input_array); const Array4D filter_array(1, 1, 1, 1, {10}); - auto filter = builder.ConstantR4FromArray4D(filter_array); + auto filter = ConstantR4FromArray4D(&builder, filter_array); - builder.Conv(input, filter, {1, 2}, Padding::kValid); + Conv(input, filter, {1, 2}, Padding::kValid); Array4D expected(1, 1, 1, 2, {10, 30}); ComputeAndCompareR4(&builder, expected, {}, error_spec_); @@ -228,12 +228,12 @@ XLA_TEST_F(ConvolutionVariantsTest, Filter1x1stride1x2in1x5) { XlaBuilder builder(TestName()); Array4D input_array(1, 1, 1, 5, {1, 2, 3, 4, 5}); - auto input = builder.ConstantR4FromArray4D(input_array); + auto input = ConstantR4FromArray4D(&builder, input_array); const Array4D filter_array(1, 1, 1, 1, {10}); - auto filter = builder.ConstantR4FromArray4D(filter_array); + auto filter = ConstantR4FromArray4D(&builder, filter_array); - builder.Conv(input, filter, {1, 2}, Padding::kValid); + Conv(input, filter, {1, 2}, Padding::kValid); Array4D expected(1, 1, 1, 3, {10, 30, 50}); ComputeAndCompareR4(&builder, expected, {}, error_spec_); @@ -243,12 +243,12 @@ XLA_TEST_F(ConvolutionVariantsTest, Filter1x3stride1x2in1x4) { XlaBuilder builder(TestName()); Array4D input_array(1, 1, 1, 4, {1, 2, 3, 4}); - auto input = builder.ConstantR4FromArray4D(input_array); + auto input = ConstantR4FromArray4D(&builder, input_array); const Array4D filter_array(1, 1, 1, 3, {100, 10, 1}); - auto filter = builder.ConstantR4FromArray4D(filter_array); + auto filter = ConstantR4FromArray4D(&builder, filter_array); - builder.Conv(input, filter, {1, 2}, Padding::kValid); + Conv(input, filter, {1, 2}, Padding::kValid); Array4D expected(1, 1, 1, 1, {123}); ComputeAndCompareR4(&builder, expected, {}, error_spec_); @@ -258,12 +258,12 @@ XLA_TEST_F(ConvolutionVariantsTest, Filter1x3stride1x2in1x5) { XlaBuilder builder(TestName()); Array4D input_array(1, 1, 1, 5, {1, 2, 3, 4, 5}); - auto input = builder.ConstantR4FromArray4D(input_array); + auto input = ConstantR4FromArray4D(&builder, input_array); const Array4D filter_array(1, 1, 1, 3, {100, 10, 1}); - auto filter = builder.ConstantR4FromArray4D(filter_array); + auto filter = ConstantR4FromArray4D(&builder, filter_array); - builder.Conv(input, filter, {1, 2}, Padding::kValid); + Conv(input, filter, {1, 2}, Padding::kValid); Array4D expected(1, 1, 1, 2, {123, 345}); ComputeAndCompareR4(&builder, expected, {}, error_spec_); @@ -273,12 +273,12 @@ XLA_TEST_F(ConvolutionVariantsTest, Filter1x1stride2x2in3x3) { XlaBuilder builder(TestName()); Array4D input_array(1, 1, 3, 3, {1, 2, 3, 4, 5, 6, 7, 8, 9}); - auto input = builder.ConstantR4FromArray4D(input_array); + auto input = ConstantR4FromArray4D(&builder, input_array); const Array4D filter_array(1, 1, 1, 1, {10}); - auto filter = builder.ConstantR4FromArray4D(filter_array); + auto filter = ConstantR4FromArray4D(&builder, filter_array); - builder.Conv(input, filter, {2, 2}, Padding::kValid); + Conv(input, filter, {2, 2}, Padding::kValid); Array4D expected(1, 1, 2, 2, {10, 30, 70, 90}); ComputeAndCompareR4(&builder, expected, {}, error_spec_); @@ -288,12 +288,12 @@ XLA_TEST_F(ConvolutionVariantsTest, Filter3x1in1x1Padded) { XlaBuilder builder(TestName()); Array4D input_array(1, 1, 1, 1, {1}); - auto input = builder.ConstantR4FromArray4D(input_array); + auto input = ConstantR4FromArray4D(&builder, input_array); const Array4D filter_array(1, 1, 1, 3, {10, 20, 30}); - auto filter = builder.ConstantR4FromArray4D(filter_array); + auto filter = ConstantR4FromArray4D(&builder, filter_array); - builder.Conv(input, filter, {1, 1}, Padding::kSame); + Conv(input, filter, {1, 1}, Padding::kSame); Array4D expected(1, 1, 1, 1, {20}); ComputeAndCompareR4(&builder, expected, {}, error_spec_); @@ -303,12 +303,12 @@ XLA_TEST_F(ConvolutionVariantsTest, Filter5x1in3x1Padded) { XlaBuilder builder(TestName()); Array4D input_array(1, 1, 1, 3, {1, 2, 3}); - auto input = builder.ConstantR4FromArray4D(input_array); + auto input = ConstantR4FromArray4D(&builder, input_array); const Array4D filter_array(1, 1, 1, 5, {10000, 1000, 100, 10, 1}); - auto filter = builder.ConstantR4FromArray4D(filter_array); + auto filter = ConstantR4FromArray4D(&builder, filter_array); - builder.Conv(input, filter, {1, 1}, Padding::kSame); + Conv(input, filter, {1, 1}, Padding::kSame); Array4D expected(1, 1, 1, 3, {123, 1230, 12300}); ComputeAndCompareR4(&builder, expected, {}, error_spec_); @@ -318,15 +318,15 @@ XLA_TEST_F(ConvolutionVariantsTest, Filter3x3in2x2Padded) { XlaBuilder builder(TestName()); Array4D input_array(1, 1, 2, 2, {1, 2, 3, 4}); - auto input = builder.ConstantR4FromArray4D(input_array); + auto input = ConstantR4FromArray4D(&builder, input_array); const Array4D filter_array(1, 1, 3, 3, {10000, 0, 1000, // row 0 0, 100, 0, // row 1 10, 0, 1}); // row 2 - auto filter = builder.ConstantR4FromArray4D(filter_array); + auto filter = ConstantR4FromArray4D(&builder, filter_array); - builder.Conv(input, filter, {1, 1}, Padding::kSame); + Conv(input, filter, {1, 1}, Padding::kSame); Array4D expected(1, 1, 2, 2, {104, 230, 2300, 10400}); ComputeAndCompareR4(&builder, expected, {}, error_spec_); @@ -336,12 +336,12 @@ XLA_TEST_F(ConvolutionVariantsTest, Filter1x1in2x1WithPaddingAndDepth) { XlaBuilder builder(TestName()); Array4D input_array(1, 2, 1, 2, {1, 2, 3, 4}); - auto input = builder.ConstantR4FromArray4D(input_array); + auto input = ConstantR4FromArray4D(&builder, input_array); const Array4D filter_array(1, 2, 1, 1, {10, 1}); - auto filter = builder.ConstantR4FromArray4D(filter_array); + auto filter = ConstantR4FromArray4D(&builder, filter_array); - builder.Conv(input, filter, {1, 1}, Padding::kSame); + Conv(input, filter, {1, 1}, Padding::kSame); Array4D expected(1, 1, 1, 2, {13, 24}); ComputeAndCompareR4(&builder, expected, {}, error_spec_); @@ -351,12 +351,12 @@ XLA_TEST_F(ConvolutionVariantsTest, Filter2x2Stride1x1Input3x3) { XlaBuilder builder(TestName()); Array4D input_array(1, 1, 3, 3, {1, 2, 3, 4, 5, 6, 7, 8, 9}); - auto input = builder.ConstantR4FromArray4D(input_array); + auto input = ConstantR4FromArray4D(&builder, input_array); const Array4D filter_array(1, 1, 2, 2, {7, 13, 17, 23}); - auto filter = builder.ConstantR4FromArray4D(filter_array); + auto filter = ConstantR4FromArray4D(&builder, filter_array); - builder.Conv(input, filter, {1, 1}, Padding::kValid); + Conv(input, filter, {1, 1}, Padding::kValid); Array4D expected(1, 1, 2, 2, {216, 276, 396, 456}); ComputeAndCompareR4(&builder, expected, {}, error_spec_); @@ -366,12 +366,12 @@ XLA_TEST_F(ConvolutionVariantsTest, Filter1x2Stride1x1Input1x3) { XlaBuilder builder(TestName()); Array4D input_array(1, 1, 1, 3, {1, 2, 3}); - auto input = builder.ConstantR4FromArray4D(input_array); + auto input = ConstantR4FromArray4D(&builder, input_array); const Array4D filter_array(1, 1, 1, 2, {7, 13}); - auto filter = builder.ConstantR4FromArray4D(filter_array); + auto filter = ConstantR4FromArray4D(&builder, filter_array); - builder.Conv(input, filter, {1, 1}, Padding::kValid); + Conv(input, filter, {1, 1}, Padding::kValid); Array4D expected(1, 1, 1, 2, {33, 53}); ComputeAndCompareR4(&builder, expected, {}, error_spec_); @@ -383,15 +383,15 @@ XLA_TEST_F(ConvolutionVariantsTest, Filter2x1x8x8Input1x1x8x8) { std::vector input_data(64); std::iota(input_data.begin(), input_data.end(), 0.0); Array4D input_array(1, 1, 8, 8, input_data); - auto input = builder.ConstantR4FromArray4D(input_array); + auto input = ConstantR4FromArray4D(&builder, input_array); std::vector filter_data(128); std::fill(filter_data.begin(), filter_data.begin() + 64, 1.0); std::fill(filter_data.begin() + 64, filter_data.begin() + 128, 2.0); const Array4D filter_array(2, 1, 8, 8, filter_data); - auto filter = builder.ConstantR4FromArray4D(filter_array); + auto filter = ConstantR4FromArray4D(&builder, filter_array); - builder.Conv(input, filter, {1, 1}, Padding::kValid); + Conv(input, filter, {1, 1}, Padding::kValid); Array4D expected(1, 2, 1, 1, {2016, 4032}); ComputeAndCompareR4(&builder, expected, {}, error_spec_); @@ -403,14 +403,14 @@ XLA_TEST_F(ConvolutionVariantsTest, Filter1x1x1x1Input16x1x1x1) { std::vector input_data(16 * 1 * 1 * 1); std::iota(input_data.begin(), input_data.end(), 1.0); Array4D input_array(16, 1, 1, 1, input_data); - auto input = builder.ConstantR4FromArray4D(input_array); + auto input = ConstantR4FromArray4D(&builder, input_array); std::vector filter_data(1 * 1 * 1 * 1); std::iota(filter_data.begin(), filter_data.end(), 1.0); const Array4D filter_array(1, 1, 1, 1, filter_data); - auto filter = builder.ConstantR4FromArray4D(filter_array); + auto filter = ConstantR4FromArray4D(&builder, filter_array); - builder.Conv(input, filter, {1, 1}, Padding::kValid); + Conv(input, filter, {1, 1}, Padding::kValid); std::vector expected_data = {1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16}; @@ -432,14 +432,14 @@ XLA_TEST_F(ConvolutionVariantsTest, Filter1x1x2x2Input16x1x2x2) { } } } - auto input = builder.ConstantR4FromArray4D(input_array); + auto input = ConstantR4FromArray4D(&builder, input_array); std::vector filter_data(1 * 1 * ky * kx); std::iota(filter_data.begin(), filter_data.end(), 1.0); const Array4D filter_array(1, 1, ky, kx, filter_data); - auto filter = builder.ConstantR4FromArray4D(filter_array); + auto filter = ConstantR4FromArray4D(&builder, filter_array); - builder.Conv(input, filter, {1, 1}, Padding::kValid); + Conv(input, filter, {1, 1}, Padding::kValid); std::vector expected_data(bs); for (int i = 0; i < bs; ++i) { @@ -463,14 +463,14 @@ XLA_TEST_F(ConvolutionVariantsTest, Filter1x1x2x2Input3x1x2x2) { } } } - auto input = builder.ConstantR4FromArray4D(input_array); + auto input = ConstantR4FromArray4D(&builder, input_array); std::vector filter_data(1 * 1 * ky * kx); std::iota(filter_data.begin(), filter_data.end(), 1.0); const Array4D filter_array(1, 1, ky, kx, filter_data); - auto filter = builder.ConstantR4FromArray4D(filter_array); + auto filter = ConstantR4FromArray4D(&builder, filter_array); - builder.Conv(input, filter, {1, 1}, Padding::kValid); + Conv(input, filter, {1, 1}, Padding::kValid); std::vector expected_data = { 23, @@ -492,14 +492,14 @@ XLA_TEST_F(ConvolutionVariantsTest, Filter1x1x8x8Input16x1x8x8) { } } } - auto input = builder.ConstantR4FromArray4D(input_array); + auto input = ConstantR4FromArray4D(&builder, input_array); std::vector filter_data(1 * 1 * 8 * 8); std::iota(filter_data.begin(), filter_data.end(), 1.0); const Array4D filter_array(1, 1, 8, 8, filter_data); - auto filter = builder.ConstantR4FromArray4D(filter_array); + auto filter = ConstantR4FromArray4D(&builder, filter_array); - builder.Conv(input, filter, {1, 1}, Padding::kValid); + Conv(input, filter, {1, 1}, Padding::kValid); std::vector expected_data = { 19664, 21744, 23824, 25904, 27984, 30064, 32144, 34224, @@ -515,7 +515,7 @@ XLA_TEST_F(ConvolutionVariantsTest, Filter2x2x8x8Input1x2x8x8) { std::vector input_data(2 * 8 * 8); std::iota(input_data.begin(), input_data.end(), 0.0); Array4D input_array(1, 2, 8, 8, input_data); - auto input = builder.ConstantR4FromArray4D(input_array); + auto input = ConstantR4FromArray4D(&builder, input_array); std::vector filter_data(2 * 2 * 8 * 8); std::fill(filter_data.begin(), filter_data.begin() + filter_data.size() / 4, @@ -527,9 +527,9 @@ XLA_TEST_F(ConvolutionVariantsTest, Filter2x2x8x8Input1x2x8x8) { std::fill(filter_data.begin() + 3 * filter_data.size() / 4, filter_data.end(), 4.0); const Array4D filter_array(2, 2, 8, 8, filter_data); - auto filter = builder.ConstantR4FromArray4D(filter_array); + auto filter = ConstantR4FromArray4D(&builder, filter_array); - builder.Conv(input, filter, {1, 1}, Padding::kValid); + Conv(input, filter, {1, 1}, Padding::kValid); Array4D expected(1, 2, 1, 1, {14240, 30496}); ComputeAndCompareR4(&builder, expected, {}, error_spec_); @@ -541,7 +541,7 @@ XLA_TEST_F(ConvolutionVariantsTest, Filter2x2x8x8Input2x2x8x8) { std::vector input_data(2 * 2 * 8 * 8); std::iota(input_data.begin(), input_data.end(), 0.0); Array4D input_array(2, 2, 8, 8, input_data); - auto input = builder.ConstantR4FromArray4D(input_array); + auto input = ConstantR4FromArray4D(&builder, input_array); std::vector filter_data(2 * 2 * 8 * 8); std::fill(filter_data.begin(), filter_data.begin() + filter_data.size() / 4, @@ -553,9 +553,9 @@ XLA_TEST_F(ConvolutionVariantsTest, Filter2x2x8x8Input2x2x8x8) { std::fill(filter_data.begin() + 3 * filter_data.size() / 4, filter_data.end(), 4.0); const Array4D filter_array(2, 2, 8, 8, filter_data); - auto filter = builder.ConstantR4FromArray4D(filter_array); + auto filter = ConstantR4FromArray4D(&builder, filter_array); - builder.Conv(input, filter, {1, 1}, Padding::kValid); + Conv(input, filter, {1, 1}, Padding::kValid); Array4D expected(2, 2, 1, 1, {14240, 30496, 38816, 87840}); ComputeAndCompareR4(&builder, expected, {}, error_spec_); @@ -567,7 +567,7 @@ XLA_TEST_F(ConvolutionVariantsTest, Filter2x2x8x8Input32x2x8x8) { std::vector input_data(32 * 2 * 8 * 8); std::iota(input_data.begin(), input_data.end(), 0.0); Array4D input_array(32, 2, 8, 8, input_data); - auto input = builder.ConstantR4FromArray4D(input_array); + auto input = ConstantR4FromArray4D(&builder, input_array); std::vector filter_data(2 * 2 * 8 * 8); std::fill(filter_data.begin(), filter_data.begin() + filter_data.size() / 4, @@ -579,9 +579,9 @@ XLA_TEST_F(ConvolutionVariantsTest, Filter2x2x8x8Input32x2x8x8) { std::fill(filter_data.begin() + 3 * filter_data.size() / 4, filter_data.end(), 4.0); const Array4D filter_array(2, 2, 8, 8, filter_data); - auto filter = builder.ConstantR4FromArray4D(filter_array); + auto filter = ConstantR4FromArray4D(&builder, filter_array); - builder.Conv(input, filter, {1, 1}, Padding::kValid); + Conv(input, filter, {1, 1}, Padding::kValid); std::vector expected_data = { 14240, 30496, 38816, 87840, 63392, 145184, 87968, @@ -613,9 +613,9 @@ XLA_TEST_F(ConvolutionVariantsTest, Filter16x16x1x1Input16x16x1x1) { } } - auto input = builder.ConstantR4FromArray4D(input_array); - auto filter = builder.ConstantR4FromArray4D(filter_array); - builder.Conv(input, filter, {1, 1}, Padding::kValid); + auto input = ConstantR4FromArray4D(&builder, input_array); + auto filter = ConstantR4FromArray4D(&builder, filter_array); + Conv(input, filter, {1, 1}, Padding::kValid); Array4D expected(16, 16, 1, 1); for (int i0 = 0; i0 < 16; ++i0) { @@ -635,9 +635,9 @@ XLA_TEST_F(ConvolutionVariantsTest, FlatRhsDilation) { Array4D input_array(1, 1, 4, 6, input_data); Array4D filter_array(1, 1, 2, 3, {1, 10, 100, 2, 20, 200}); - auto input = builder.ConstantR4FromArray4D(input_array); - auto filter = builder.ConstantR4FromArray4D(filter_array); - builder.ConvGeneralDilated( + auto input = ConstantR4FromArray4D(&builder, input_array); + auto filter = ConstantR4FromArray4D(&builder, filter_array); + ConvGeneralDilated( /*lhs=*/input, /*rhs=*/filter, /*window_strides=*/{}, /*padding=*/{}, /*lhs_dilation=*/{}, /*rhs_dilation=*/{2, 2}, XlaBuilder::CreateDefaultConvDimensionNumbers()); @@ -654,9 +654,9 @@ XLA_TEST_F(ConvolutionVariantsTest, FlatLhsDilation1D) { Array4D input_array(1, 1, 1, 5, input_data); Array4D filter_array(1, 1, 1, 2, {10, 1}); - auto input = builder.ConstantR4FromArray4D(input_array); - auto filter = builder.ConstantR4FromArray4D(filter_array); - builder.ConvGeneralDilated( + auto input = ConstantR4FromArray4D(&builder, input_array); + auto filter = ConstantR4FromArray4D(&builder, filter_array); + ConvGeneralDilated( /*lhs=*/input, /*rhs=*/filter, /*window_strides=*/{}, /*padding=*/{}, /*lhs_dilation=*/{1, 2}, /*rhs_dilation=*/{}, XlaBuilder::CreateDefaultConvDimensionNumbers()); @@ -677,9 +677,9 @@ XLA_TEST_F(ConvolutionVariantsTest, FlatLhsDilation) { 200, 20, 2, // 300, 30, 3, // 400, 40, 4}); - auto input = builder.ConstantR4FromArray4D(input_array); - auto filter = builder.ConstantR4FromArray4D(filter_array); - builder.ConvGeneralDilated( + auto input = ConstantR4FromArray4D(&builder, input_array); + auto filter = ConstantR4FromArray4D(&builder, filter_array); + ConvGeneralDilated( /*lhs=*/input, /*rhs=*/filter, /*window_strides=*/{2, 1}, /*padding=*/{{1, 0}, {0, 0}}, /*lhs_dilation=*/{3, 2}, /*rhs_dilation=*/{}, XlaBuilder::CreateDefaultConvDimensionNumbers()); @@ -699,9 +699,9 @@ XLA_TEST_F(ConvolutionVariantsTest, NegativePaddingOnBothEnds) { Array4D input_array(1, 1, 1, 5, input_data); Array4D filter_array(1, 1, 1, 2, {10, 1}); - auto input = builder.ConstantR4FromArray4D(input_array); - auto filter = builder.ConstantR4FromArray4D(filter_array); - builder.ConvGeneral( + auto input = ConstantR4FromArray4D(&builder, input_array); + auto filter = ConstantR4FromArray4D(&builder, filter_array); + ConvGeneral( /*lhs=*/input, /*rhs=*/filter, /*window_strides=*/{}, /*padding=*/{{0, 0}, {-1, -1}}, XlaBuilder::CreateDefaultConvDimensionNumbers()); @@ -718,9 +718,9 @@ XLA_TEST_F(ConvolutionVariantsTest, NegativePaddingLowAndPositivePaddingHigh) { Array4D input_array(1, 1, 1, 5, input_data); Array4D filter_array(1, 1, 1, 2, {10, 1}); - auto input = builder.ConstantR4FromArray4D(input_array); - auto filter = builder.ConstantR4FromArray4D(filter_array); - builder.ConvGeneral( + auto input = ConstantR4FromArray4D(&builder, input_array); + auto filter = ConstantR4FromArray4D(&builder, filter_array); + ConvGeneral( /*lhs=*/input, /*rhs=*/filter, /*window_strides=*/{}, /*padding=*/{{0, 0}, {-1, 2}}, XlaBuilder::CreateDefaultConvDimensionNumbers()); @@ -737,9 +737,9 @@ XLA_TEST_F(ConvolutionVariantsTest, PositivePaddingLowAndNegativePaddingHigh) { Array4D input_array(1, 1, 1, 5, input_data); Array4D filter_array(1, 1, 1, 2, {10, 1}); - auto input = builder.ConstantR4FromArray4D(input_array); - auto filter = builder.ConstantR4FromArray4D(filter_array); - builder.ConvGeneral( + auto input = ConstantR4FromArray4D(&builder, input_array); + auto filter = ConstantR4FromArray4D(&builder, filter_array); + ConvGeneral( /*lhs=*/input, /*rhs=*/filter, /*window_strides=*/{}, /*padding=*/{{0, 0}, {2, -1}}, XlaBuilder::CreateDefaultConvDimensionNumbers()); @@ -756,9 +756,9 @@ XLA_TEST_F(ConvolutionVariantsTest, PositivePaddingAndDilation) { Array4D input_array(1, 1, 1, 5, input_data); Array4D filter_array(1, 1, 1, 2, {10, 1}); - auto input = builder.ConstantR4FromArray4D(input_array); - auto filter = builder.ConstantR4FromArray4D(filter_array); - builder.ConvGeneralDilated( + auto input = ConstantR4FromArray4D(&builder, input_array); + auto filter = ConstantR4FromArray4D(&builder, filter_array); + ConvGeneralDilated( /*lhs=*/input, /*rhs=*/filter, /*window_strides=*/{}, /*padding=*/{{0, 0}, {3, 2}}, /*lhs_dilation=*/{1, 2}, /*rhs_dilation=*/{1, 2}, @@ -781,9 +781,9 @@ XLA_TEST_F(ConvolutionVariantsTest, NegativePaddingAndDilation) { Array4D input_array(1, 1, 1, 5, input_data); Array4D filter_array(1, 1, 1, 2, {10, 1}); - auto input = builder.ConstantR4FromArray4D(input_array); - auto filter = builder.ConstantR4FromArray4D(filter_array); - builder.ConvGeneralDilated( + auto input = ConstantR4FromArray4D(&builder, input_array); + auto filter = ConstantR4FromArray4D(&builder, filter_array); + ConvGeneralDilated( /*lhs=*/input, /*rhs=*/filter, /*window_strides=*/{}, /*padding=*/{{0, 0}, {-3, -2}}, /*lhs_dilation=*/{1, 2}, /*rhs_dilation=*/{1, 2}, @@ -821,9 +821,9 @@ XLA_TEST_F(ConvolutionVariantsTest, RandomData_Input1x1x2x3_Filter2x1x1x2) { Array4D filter_array(oz, iz, ky, kx, kernel_data); XlaBuilder builder(TestName()); - auto input = builder.ConstantR4FromArray4D(input_array); - auto filter = builder.ConstantR4FromArray4D(filter_array); - builder.Conv(input, filter, {1, 1}, Padding::kValid); + auto input = ConstantR4FromArray4D(&builder, input_array); + auto filter = ConstantR4FromArray4D(&builder, filter_array); + Conv(input, filter, {1, 1}, Padding::kValid); std::unique_ptr> expected = ReferenceUtil::ConvArray4D( input_array, filter_array, {1, 1}, Padding::kValid); @@ -854,9 +854,9 @@ XLA_TEST_F(ConvolutionVariantsTest, RandomData_Input1x16x1x1_Filter1x16x1x1) { Array4D filter_array(oz, iz, ky, kx, kernel_data); XlaBuilder builder(TestName()); - auto input = builder.ConstantR4FromArray4D(input_array); - auto filter = builder.ConstantR4FromArray4D(filter_array); - builder.Conv(input, filter, {1, 1}, Padding::kValid); + auto input = ConstantR4FromArray4D(&builder, input_array); + auto filter = ConstantR4FromArray4D(&builder, filter_array); + Conv(input, filter, {1, 1}, Padding::kValid); std::unique_ptr> expected = ReferenceUtil::ConvArray4D( input_array, filter_array, {1, 1}, Padding::kValid); @@ -887,9 +887,9 @@ XLA_TEST_F(ConvolutionVariantsTest, RandomData_Input16x16x1x1_Filter1x16x1x1) { Array4D filter_array(oz, iz, ky, kx, kernel_data); XlaBuilder builder(TestName()); - auto input = builder.ConstantR4FromArray4D(input_array); - auto filter = builder.ConstantR4FromArray4D(filter_array); - builder.Conv(input, filter, {1, 1}, Padding::kValid); + auto input = ConstantR4FromArray4D(&builder, input_array); + auto filter = ConstantR4FromArray4D(&builder, filter_array); + Conv(input, filter, {1, 1}, Padding::kValid); std::unique_ptr> expected = ReferenceUtil::ConvArray4D( input_array, filter_array, {1, 1}, Padding::kValid); @@ -920,9 +920,9 @@ XLA_TEST_F(ConvolutionVariantsTest, RandomData_Input16x16x1x1_Filter16x16x1x1) { Array4D filter_array(oz, iz, ky, kx, kernel_data); XlaBuilder builder(TestName()); - auto input = builder.ConstantR4FromArray4D(input_array); - auto filter = builder.ConstantR4FromArray4D(filter_array); - builder.Conv(input, filter, {1, 1}, Padding::kValid); + auto input = ConstantR4FromArray4D(&builder, input_array); + auto filter = ConstantR4FromArray4D(&builder, filter_array); + Conv(input, filter, {1, 1}, Padding::kValid); std::unique_ptr> expected = ReferenceUtil::ConvArray4D( input_array, filter_array, {1, 1}, Padding::kValid); @@ -954,9 +954,9 @@ XLA_TEST_F(ConvolutionVariantsTest, Array4D filter_array(oz, iz, ky, kx, kernel_data); XlaBuilder builder(TestName()); - auto input = builder.ConstantR4FromArray4D(input_array); - auto filter = builder.ConstantR4FromArray4D(filter_array); - builder.Conv(input, filter, {1, 1}, Padding::kValid); + auto input = ConstantR4FromArray4D(&builder, input_array); + auto filter = ConstantR4FromArray4D(&builder, filter_array); + Conv(input, filter, {1, 1}, Padding::kValid); std::unique_ptr> expected = ReferenceUtil::ConvArray4D( input_array, filter_array, {1, 1}, Padding::kValid); @@ -970,12 +970,12 @@ XLA_TEST_F(ConvolutionVariantsTest, Filter1x2x1x1Input1x2x3x1GeneralPadding) { std::vector input_data(1 * 2 * 3 * 1); std::iota(input_data.begin(), input_data.end(), 1.0); Array4D input_array(1, 2, 3, 1, input_data); - auto input = builder.ConstantR4FromArray4D(input_array); + auto input = ConstantR4FromArray4D(&builder, input_array); std::vector filter_data(1 * 2 * 1 * 1); std::iota(filter_data.begin(), filter_data.end(), 1.0); Array4D filter_array(1, 2, 1, 1, filter_data); - auto filter = builder.ConstantR4FromArray4D(filter_array); + auto filter = ConstantR4FromArray4D(&builder, filter_array); ConvolutionDimensionNumbers dnums; // NHWC input format. @@ -995,7 +995,7 @@ XLA_TEST_F(ConvolutionVariantsTest, Filter1x2x1x1Input1x2x3x1GeneralPadding) { dnums.set_kernel_output_feature_dimension(3); // Tests padding sizes that don't correspond either to SAME or VALID padding. - builder.ConvGeneral(input, filter, {1, 1}, {{2, 1}, {2, 3}}, dnums); + ConvGeneral(input, filter, {1, 1}, {{2, 1}, {2, 3}}, dnums); std::vector expected_data = { 0, 0, 0, 0, 0, 0, 0, // @@ -1014,12 +1014,12 @@ XLA_TEST_F(ConvolutionVariantsTest, Filter1x1x1x1Input1x2x3x1GeneralPadding) { std::vector input_data(1 * 2 * 3 * 1); std::iota(input_data.begin(), input_data.end(), 1.0); Array4D input_array(1, 2, 3, 1, input_data); - auto input = builder.ConstantR4FromArray4D(input_array); + auto input = ConstantR4FromArray4D(&builder, input_array); std::vector filter_data(1 * 1 * 1 * 1); std::iota(filter_data.begin(), filter_data.end(), 2.0); Array4D filter_array(1, 1, 1, 1, filter_data); - auto filter = builder.ConstantR4FromArray4D(filter_array); + auto filter = ConstantR4FromArray4D(&builder, filter_array); ConvolutionDimensionNumbers dnums; // NHWC input format. @@ -1039,7 +1039,7 @@ XLA_TEST_F(ConvolutionVariantsTest, Filter1x1x1x1Input1x2x3x1GeneralPadding) { dnums.set_kernel_output_feature_dimension(3); // Tests padding sizes that don't correspond either to SAME or VALID padding. - builder.ConvGeneral(input, filter, {1, 1}, {{2, 1}, {2, 3}}, dnums); + ConvGeneral(input, filter, {1, 1}, {{2, 1}, {2, 3}}, dnums); std::vector expected_data = { 0, 0, 0, 0, 0, 0, 0, 0, // @@ -1058,12 +1058,12 @@ XLA_TEST_F(ConvolutionVariantsTest, Filter1x1x1x1Input1x2x3x1NoPadding) { std::vector input_data(1 * 2 * 3 * 1); std::iota(input_data.begin(), input_data.end(), 1.0); Array4D input_array(1, 2, 3, 1, input_data); - auto input = builder.ConstantR4FromArray4D(input_array); + auto input = ConstantR4FromArray4D(&builder, input_array); std::vector filter_data(1 * 1 * 1 * 1); std::iota(filter_data.begin(), filter_data.end(), 2.0); Array4D filter_array(1, 1, 1, 1, filter_data); - auto filter = builder.ConstantR4FromArray4D(filter_array); + auto filter = ConstantR4FromArray4D(&builder, filter_array); ConvolutionDimensionNumbers dnums; // NHWC input format. @@ -1083,7 +1083,7 @@ XLA_TEST_F(ConvolutionVariantsTest, Filter1x1x1x1Input1x2x3x1NoPadding) { dnums.set_kernel_output_feature_dimension(3); // Tests zero padding sizes. This can use matmul for computation. - builder.ConvGeneral(input, filter, {1, 1}, {{0, 0}, {0, 0}}, dnums); + ConvGeneral(input, filter, {1, 1}, {{0, 0}, {0, 0}}, dnums); std::vector expected_data = { 2, 4, 6, // @@ -1099,12 +1099,12 @@ XLA_TEST_F(ConvolutionVariantsTest, Filter1x1x2x3Input1x2x3x2NoPadding) { std::vector input_data(1 * 2 * 3 * 2); std::iota(input_data.begin(), input_data.end(), 1.0); Array4D input_array(1, 2, 3, 2, input_data); - auto input = builder.ConstantR4FromArray4D(input_array); + auto input = ConstantR4FromArray4D(&builder, input_array); std::vector filter_data(1 * 1 * 2 * 3); std::iota(filter_data.begin(), filter_data.end(), 2.0); Array4D filter_array(1, 1, 2, 3, filter_data); - auto filter = builder.ConstantR4FromArray4D(filter_array); + auto filter = ConstantR4FromArray4D(&builder, filter_array); ConvolutionDimensionNumbers dnums; // NHWC input format. @@ -1124,7 +1124,7 @@ XLA_TEST_F(ConvolutionVariantsTest, Filter1x1x2x3Input1x2x3x2NoPadding) { dnums.set_kernel_output_feature_dimension(3); // Tests zero padding sizes. This can use matmul for computation. - builder.ConvGeneral(input, filter, {1, 1}, {{0, 0}, {0, 0}}, dnums); + ConvGeneral(input, filter, {1, 1}, {{0, 0}, {0, 0}}, dnums); std::vector expected_data = { 12, 15, 18, // @@ -1148,14 +1148,14 @@ XLA_TEST_F(ConvolutionVariantsTest, BackwardInputLowPaddingLessThanHighPadding) { XlaBuilder builder(TestName()); - auto gradients = builder.ConstantR4FromArray4D( - Array4D(1, 1, 1, 3, /*values=*/{1, 2, 3})); - auto weights = builder.ConstantR4FromArray4D( - Array4D(1, 1, 1, 2, /*values=*/{5, 6})); - auto mirrored_weights = builder.Rev(weights, {2, 3}); - builder.ConvWithGeneralPadding(gradients, mirrored_weights, - /*window_strides=*/{1, 1}, - /*padding=*/{{0, 0}, {1, 0}}); + auto gradients = ConstantR4FromArray4D( + &builder, Array4D(1, 1, 1, 3, /*values=*/{1, 2, 3})); + auto weights = ConstantR4FromArray4D( + &builder, Array4D(1, 1, 1, 2, /*values=*/{5, 6})); + auto mirrored_weights = Rev(weights, {2, 3}); + ConvWithGeneralPadding(gradients, mirrored_weights, + /*window_strides=*/{1, 1}, + /*padding=*/{{0, 0}, {1, 0}}); ComputeAndCompareR4(&builder, {{{{5, 16, 27}}}}, {}, error_spec_); } @@ -1167,16 +1167,16 @@ XLA_TEST_F(ConvolutionVariantsTest, BackwardInputLowPaddingGreaterThanHighPadding) { XlaBuilder builder(TestName()); - auto gradients = builder.ConstantR4FromArray4D( - Array4D(1, 1, 1, 1, /*values=*/{1})); - auto weights = builder.ConstantR4FromArray4D( - Array4D(1, 1, 1, 3, /*values=*/{1, 10, 100})); - auto mirrored_weights = builder.Rev(weights, {2, 3}); - builder.ConvGeneralDilated(gradients, mirrored_weights, - /*window_strides=*/{1, 1}, - /*padding=*/{{0, 0}, {0, 3}}, - /*lhs_dilation=*/{1, 3}, /*rhs_dilation=*/{}, - XlaBuilder::CreateDefaultConvDimensionNumbers()); + auto gradients = ConstantR4FromArray4D( + &builder, Array4D(1, 1, 1, 1, /*values=*/{1})); + auto weights = ConstantR4FromArray4D( + &builder, Array4D(1, 1, 1, 3, /*values=*/{1, 10, 100})); + auto mirrored_weights = Rev(weights, {2, 3}); + ConvGeneralDilated(gradients, mirrored_weights, + /*window_strides=*/{1, 1}, + /*padding=*/{{0, 0}, {0, 3}}, + /*lhs_dilation=*/{1, 3}, /*rhs_dilation=*/{}, + XlaBuilder::CreateDefaultConvDimensionNumbers()); ComputeAndCompareR4(&builder, {{{{100, 0}}}}, {}, error_spec_); } @@ -1187,14 +1187,14 @@ XLA_TEST_F(ConvolutionVariantsTest, XLA_TEST_F(ConvolutionVariantsTest, BackwardInputEvenPadding) { XlaBuilder builder(TestName()); - auto gradients = builder.ConstantR4FromArray4D( - Array4D(1, 1, 1, 1, /*values=*/{1})); - auto weights = builder.ConstantR4FromArray4D( - Array4D(1, 1, 1, 3, /*values=*/{1, 10, 100})); - auto mirrored_weights = builder.Rev(weights, {2, 3}); - builder.ConvWithGeneralPadding(gradients, mirrored_weights, - /*window_strides=*/{1, 1}, - /*padding=*/{{0, 0}, {1, 1}}); + auto gradients = ConstantR4FromArray4D( + &builder, Array4D(1, 1, 1, 1, /*values=*/{1})); + auto weights = ConstantR4FromArray4D( + &builder, Array4D(1, 1, 1, 3, /*values=*/{1, 10, 100})); + auto mirrored_weights = Rev(weights, {2, 3}); + ConvWithGeneralPadding(gradients, mirrored_weights, + /*window_strides=*/{1, 1}, + /*padding=*/{{0, 0}, {1, 1}}); ComputeAndCompareR4(&builder, {{{{10}}}}, {}, error_spec_); } @@ -1208,14 +1208,14 @@ XLA_TEST_F(ConvolutionVariantsTest, BackwardInputEvenPadding) { XLA_TEST_F(ConvolutionVariantsTest, BackwardInputWithNegativePaddingHigh) { XlaBuilder builder(TestName()); - auto gradients = builder.ConstantR4FromArray4D( - Array4D(1, 1, 1, 3, /*values=*/{1, 2, 3})); - auto weights = builder.ConstantR4FromArray4D( - Array4D(1, 1, 1, 2, /*values=*/{1, 10})); - auto mirrored_weights = builder.Rev(weights, {2, 3}); - builder.ConvWithGeneralPadding(gradients, mirrored_weights, - /*window_strides=*/{1, 1}, - /*padding=*/{{0, 0}, {0, 2}}); + auto gradients = ConstantR4FromArray4D( + &builder, Array4D(1, 1, 1, 3, /*values=*/{1, 2, 3})); + auto weights = ConstantR4FromArray4D( + &builder, Array4D(1, 1, 1, 2, /*values=*/{1, 10})); + auto mirrored_weights = Rev(weights, {2, 3}); + ConvWithGeneralPadding(gradients, mirrored_weights, + /*window_strides=*/{1, 1}, + /*padding=*/{{0, 0}, {0, 2}}); ComputeAndCompareR4(&builder, {{{{12, 23, 30, 0}}}}, {}, error_spec_); } @@ -1229,17 +1229,17 @@ XLA_TEST_F(ConvolutionVariantsTest, // weight gradients: 24,130,240 // // This pattern will be fused to backward convolution with padding=(1,2). - auto activations = builder.ConstantR4FromArray4D( - Array4D(1, 1, 1, 4, /*values=*/{1, 2, 3, 4})); - auto gradients = builder.ConstantR4FromArray4D( - Array4D(1, 1, 1, 3, /*values=*/{100, 10, 1})); - auto forward_conv = builder.ConvGeneralDilated( - activations, gradients, - /*window_strides=*/{1, 1}, - /*padding=*/{{0, 0}, {1, 2}}, - /*lhs_dilation=*/{}, /*rhs_dilation=*/{1, 2}, - XlaBuilder::CreateDefaultConvDimensionNumbers()); - builder.Transpose(forward_conv, {0, 1, 2, 3}); + auto activations = ConstantR4FromArray4D( + &builder, Array4D(1, 1, 1, 4, /*values=*/{1, 2, 3, 4})); + auto gradients = ConstantR4FromArray4D( + &builder, Array4D(1, 1, 1, 3, /*values=*/{100, 10, 1})); + auto forward_conv = + ConvGeneralDilated(activations, gradients, + /*window_strides=*/{1, 1}, + /*padding=*/{{0, 0}, {1, 2}}, + /*lhs_dilation=*/{}, /*rhs_dilation=*/{1, 2}, + XlaBuilder::CreateDefaultConvDimensionNumbers()); + Transpose(forward_conv, {0, 1, 2, 3}); ComputeAndCompareR4(&builder, {{{{24, 130, 240}}}}, {}, error_spec_); } @@ -1255,17 +1255,17 @@ XLA_TEST_F(ConvolutionVariantsTest, // This pattern will be fused to backward convolution with padding=(2,1). // Note: both (2,1) and (2,0) are valid padding for the backward convolution // because the stride is 2. - auto activations = builder.ConstantR4FromArray4D( - Array4D(1, 1, 1, 4, /*values=*/{1, 2, 3, 4})); - auto gradients = builder.ConstantR4FromArray4D( - Array4D(1, 1, 1, 3, /*values=*/{100, 10, 1})); - auto forward_conv = builder.ConvGeneralDilated( - activations, gradients, - /*window_strides=*/{1, 1}, - /*padding=*/{{0, 0}, {2, 0}}, - /*lhs_dilation=*/{}, /*rhs_dilation=*/{1, 2}, - XlaBuilder::CreateDefaultConvDimensionNumbers()); - builder.Transpose(forward_conv, {0, 1, 2, 3}); + auto activations = ConstantR4FromArray4D( + &builder, Array4D(1, 1, 1, 4, /*values=*/{1, 2, 3, 4})); + auto gradients = ConstantR4FromArray4D( + &builder, Array4D(1, 1, 1, 3, /*values=*/{100, 10, 1})); + auto forward_conv = + ConvGeneralDilated(activations, gradients, + /*window_strides=*/{1, 1}, + /*padding=*/{{0, 0}, {2, 0}}, + /*lhs_dilation=*/{}, /*rhs_dilation=*/{1, 2}, + XlaBuilder::CreateDefaultConvDimensionNumbers()); + Transpose(forward_conv, {0, 1, 2, 3}); ComputeAndCompareR4(&builder, {{{{13, 24}}}}, {}, error_spec_); } @@ -1282,17 +1282,17 @@ XLA_TEST_F(ConvolutionVariantsTest, BackwardFilterEvenPadding) { // because the stride is 2. ConvolutionFolding prefers (2,2) because cuDNN // supports even padding only -- using (2,1) would need extra effort of // canonicalization. - auto activations = builder.ConstantR4FromArray4D( - Array4D(1, 1, 1, 4, /*values=*/{1, 2, 3, 4})); - auto gradients = builder.ConstantR4FromArray4D( - Array4D(1, 1, 1, 3, /*values=*/{100, 10, 1})); - auto forward_conv = builder.ConvGeneralDilated( - activations, gradients, - /*window_strides=*/{1, 1}, - /*padding=*/{{0, 0}, {2, 1}}, - /*lhs_dilation=*/{}, /*rhs_dilation=*/{1, 2}, - XlaBuilder::CreateDefaultConvDimensionNumbers()); - builder.Transpose(forward_conv, {0, 1, 2, 3}); + auto activations = ConstantR4FromArray4D( + &builder, Array4D(1, 1, 1, 4, /*values=*/{1, 2, 3, 4})); + auto gradients = ConstantR4FromArray4D( + &builder, Array4D(1, 1, 1, 3, /*values=*/{100, 10, 1})); + auto forward_conv = + ConvGeneralDilated(activations, gradients, + /*window_strides=*/{1, 1}, + /*padding=*/{{0, 0}, {2, 1}}, + /*lhs_dilation=*/{}, /*rhs_dilation=*/{1, 2}, + XlaBuilder::CreateDefaultConvDimensionNumbers()); + Transpose(forward_conv, {0, 1, 2, 3}); ComputeAndCompareR4(&builder, {{{{13, 24, 130}}}}, {}, error_spec_); } @@ -1300,14 +1300,14 @@ XLA_TEST_F(ConvolutionVariantsTest, BackwardFilterEvenPadding) { XLA_TEST_F(ConvolutionVariantsTest, BackwardInputEvenPadding1D) { XlaBuilder builder(TestName()); - auto gradients = builder.ConstantR3FromArray3D( - Array3D(1, 1, 1, /*value=*/1)); + auto gradients = ConstantR3FromArray3D( + &builder, Array3D(1, 1, 1, /*value=*/1)); auto weights = - builder.ConstantR3FromArray3D(Array3D({{{1, 10, 100}}})); - auto mirrored_weights = builder.Rev(weights, {2}); - builder.ConvWithGeneralPadding(gradients, mirrored_weights, - /*window_strides=*/{1}, - /*padding=*/{{1, 1}}); + ConstantR3FromArray3D(&builder, Array3D({{{1, 10, 100}}})); + auto mirrored_weights = Rev(weights, {2}); + ConvWithGeneralPadding(gradients, mirrored_weights, + /*window_strides=*/{1}, + /*padding=*/{{1, 1}}); ComputeAndCompareR3(&builder, {{{10}}}, {}, error_spec_); } @@ -1315,17 +1315,17 @@ XLA_TEST_F(ConvolutionVariantsTest, BackwardFilterEvenPadding1D) { XlaBuilder builder(TestName()); auto activations = - builder.ConstantR3FromArray3D(Array3D({{{1, 2, 3, 4}}})); + ConstantR3FromArray3D(&builder, Array3D({{{1, 2, 3, 4}}})); auto gradients = - builder.ConstantR3FromArray3D(Array3D({{{100, 10, 1}}})); + ConstantR3FromArray3D(&builder, Array3D({{{100, 10, 1}}})); auto forward_conv = - builder.ConvGeneralDilated(activations, gradients, - /*window_strides=*/{1}, - /*padding=*/{{2, 1}}, - /*lhs_dilation=*/{}, /*rhs_dilation=*/{2}, - XlaBuilder::CreateDefaultConvDimensionNumbers( - /*num_spatial_dims=*/1)); - builder.Transpose(forward_conv, {0, 1, 2}); + ConvGeneralDilated(activations, gradients, + /*window_strides=*/{1}, + /*padding=*/{{2, 1}}, + /*lhs_dilation=*/{}, /*rhs_dilation=*/{2}, + XlaBuilder::CreateDefaultConvDimensionNumbers( + /*num_spatial_dims=*/1)); + Transpose(forward_conv, {0, 1, 2}); ComputeAndCompareR3(&builder, {{{13, 24, 130}}}, {}, error_spec_); } @@ -1333,52 +1333,52 @@ XLA_TEST_F(ConvolutionVariantsTest, BackwardFilterEvenPadding1D) { XLA_TEST_F(ConvolutionVariantsTest, BackwardInputEvenPadding3D) { XlaBuilder builder(TestName()); - auto gradients_flat = Literal::CreateR1({1}); + auto gradients_flat = LiteralUtil::CreateR1({1}); auto gradients_literal = gradients_flat->Reshape({1, 1, 1, 1, 1}).ConsumeValueOrDie(); - auto gradients = builder.ConstantLiteral(*gradients_literal); + auto gradients = ConstantLiteral(&builder, *gradients_literal); - auto weights_flat = Literal::CreateR1({1, 10, 100}); + auto weights_flat = LiteralUtil::CreateR1({1, 10, 100}); auto weights_literal = weights_flat->Reshape({1, 1, 1, 1, 3}).ConsumeValueOrDie(); - auto weights = builder.ConstantLiteral(*weights_literal); + auto weights = ConstantLiteral(&builder, *weights_literal); - auto expected_flat = Literal::CreateR1({10}); + auto expected_flat = LiteralUtil::CreateR1({10}); auto expected_literal = expected_flat->Reshape({1, 1, 1, 1, 1}).ConsumeValueOrDie(); - auto mirrored_weights = builder.Rev(weights, {2, 3, 4}); - builder.ConvWithGeneralPadding(gradients, mirrored_weights, - /*window_strides=*/{1, 1, 1}, - /*padding=*/{{0, 0}, {0, 0}, {1, 1}}); + auto mirrored_weights = Rev(weights, {2, 3, 4}); + ConvWithGeneralPadding(gradients, mirrored_weights, + /*window_strides=*/{1, 1, 1}, + /*padding=*/{{0, 0}, {0, 0}, {1, 1}}); ComputeAndCompareLiteral(&builder, *expected_literal, {}, error_spec_); } XLA_TEST_F(ConvolutionVariantsTest, BackwardFilterEvenPadding3D) { XlaBuilder builder(TestName()); - auto activations_flat = Literal::CreateR1({1, 2, 3, 4}); + auto activations_flat = LiteralUtil::CreateR1({1, 2, 3, 4}); auto activations_literal = activations_flat->Reshape({1, 1, 1, 1, 4}).ConsumeValueOrDie(); - auto activations = builder.ConstantLiteral(*activations_literal); + auto activations = ConstantLiteral(&builder, *activations_literal); - auto gradients_flat = Literal::CreateR1({100, 10, 1}); + auto gradients_flat = LiteralUtil::CreateR1({100, 10, 1}); auto gradients_literal = gradients_flat->Reshape({1, 1, 1, 1, 3}).ConsumeValueOrDie(); - auto gradients = builder.ConstantLiteral(*gradients_literal); + auto gradients = ConstantLiteral(&builder, *gradients_literal); - auto expected_flat = Literal::CreateR1({13, 24, 130}); + auto expected_flat = LiteralUtil::CreateR1({13, 24, 130}); auto expected_literal = expected_flat->Reshape({1, 1, 1, 1, 3}).ConsumeValueOrDie(); - auto forward_conv = builder.ConvGeneralDilated( - activations, gradients, - /*window_strides=*/{1, 1, 1}, - /*padding=*/{{0, 0}, {0, 0}, {2, 1}}, - /*lhs_dilation=*/{}, /*rhs_dilation=*/{1, 1, 2}, - XlaBuilder::CreateDefaultConvDimensionNumbers( - /*num_spatial_dims=*/3)); - builder.Transpose(forward_conv, {0, 1, 2, 3, 4}); + auto forward_conv = + ConvGeneralDilated(activations, gradients, + /*window_strides=*/{1, 1, 1}, + /*padding=*/{{0, 0}, {0, 0}, {2, 1}}, + /*lhs_dilation=*/{}, /*rhs_dilation=*/{1, 1, 2}, + XlaBuilder::CreateDefaultConvDimensionNumbers( + /*num_spatial_dims=*/3)); + Transpose(forward_conv, {0, 1, 2, 3, 4}); ComputeAndCompareLiteral(&builder, *expected_literal, {}, error_spec_); } diff --git a/tensorflow/compiler/xla/tests/copy_test.cc b/tensorflow/compiler/xla/tests/copy_test.cc index 2b3390ca98cb2922410d451c06811aa9d4ff8c0b..1dc6ff0f4f51b51002cfb868a51457c08a259a80 100644 --- a/tensorflow/compiler/xla/tests/copy_test.cc +++ b/tensorflow/compiler/xla/tests/copy_test.cc @@ -18,7 +18,7 @@ limitations under the License. #include "tensorflow/compiler/xla/array2d.h" #include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/ptr_util.h" #include "tensorflow/compiler/xla/service/hlo_computation.h" #include "tensorflow/compiler/xla/service/hlo_instruction.h" @@ -58,37 +58,38 @@ class CopyOpTest : public HloTestBase { }; XLA_TEST_F(CopyOpTest, CopyR0Bool) { - TestCopyOp(*Literal::CreateR0(true)); + TestCopyOp(*LiteralUtil::CreateR0(true)); } XLA_TEST_F(CopyOpTest, CopyR1S0U32) { - TestCopyOp(*Literal::CreateR1({})); + TestCopyOp(*LiteralUtil::CreateR1({})); } XLA_TEST_F(CopyOpTest, CopyR1S3U32) { - TestCopyOp(*Literal::CreateR1({1, 2, 3})); + TestCopyOp(*LiteralUtil::CreateR1({1, 2, 3})); } XLA_TEST_F(CopyOpTest, CopyR3F32_2x2x3) { - TestCopyOp(*Literal::CreateR3({{{1.0f, 2.0f, 3.0f}, {4.0f, 5.0f, 6.0f}}, - {{1.1f, 2.1f, 3.1f}, {6.1f, 3.5f, 2.8f}}})); + TestCopyOp( + *LiteralUtil::CreateR3({{{1.0f, 2.0f, 3.0f}, {4.0f, 5.0f, 6.0f}}, + {{1.1f, 2.1f, 3.1f}, {6.1f, 3.5f, 2.8f}}})); } XLA_TEST_F(CopyOpTest, CopyR4S32_2x2x3x2) { - TestCopyOp(*Literal::CreateR4( + TestCopyOp(*LiteralUtil::CreateR4( {{{{1, -2}, {-4, 5}, {6, 7}}, {{8, 9}, {10, 11}, {12, 13}}}, {{{10, 3}, {7, -2}, {3, 6}}, {{2, 5}, {-11, 5}, {-2, -5}}}})); } XLA_TEST_F(CopyOpTest, CopyR4S32_0x2x3x2) { - TestCopyOp(*Literal::CreateR4FromArray4D(Array4D(0, 2, 3, 2))); + TestCopyOp(*LiteralUtil::CreateR4FromArray4D(Array4D(0, 2, 3, 2))); } XLA_TEST_F(CopyOpTest, CopyParameterScalar) { auto builder = HloComputation::Builder(TestName()); // Copy literal to device to use as parameter. - auto literal = Literal::CreateR0(42.0); + auto literal = LiteralUtil::CreateR0(42.0); Shape shape = literal->shape(); auto param0 = builder.AddInstruction( @@ -109,7 +110,7 @@ XLA_TEST_F(CopyOpTest, CopyParameterScalar) { XLA_TEST_F(CopyOpTest, CopyConstantR2Twice) { auto builder = HloComputation::Builder(TestName()); - auto literal = Literal::CreateR2({{1.0, 2.0}, {3.0, 4.0}}); + auto literal = LiteralUtil::CreateR2({{1.0, 2.0}, {3.0, 4.0}}); auto constant = builder.AddInstruction( HloInstruction::CreateConstant(std::move(literal))); @@ -131,7 +132,7 @@ XLA_TEST_F(CopyOpTest, CopyConstantR2DifferentLayouts) { HloComputation::Builder builder(TestName()); std::unique_ptr literal = - Literal::CreateR2({{1.0, 2.0}, {3.0, 4.0}}); + LiteralUtil::CreateR2({{1.0, 2.0}, {3.0, 4.0}}); // Reverse the minor-to-major order of the literal. Layout* literal_layout = literal->mutable_shape_do_not_use()->mutable_layout(); @@ -168,7 +169,7 @@ void CopyOpTest::TestCopyConstantLayout021(size_t n1, size_t n2, size_t n3) { HloComputation::Builder builder(TestName()); - std::unique_ptr literal = Literal::CreateR3FromArray3D(a); + std::unique_ptr literal = LiteralUtil::CreateR3FromArray3D(a); HloInstruction* constant = builder.AddInstruction( HloInstruction::CreateConstant(std::move(literal))); @@ -202,7 +203,7 @@ void CopyOpTest::TestCopyConstantLayoutR4( HloComputation::Builder builder(TestName()); - std::unique_ptr literal = Literal::CreateR4FromArray4D(a); + std::unique_ptr literal = LiteralUtil::CreateR4FromArray4D(a); HloInstruction* constant = builder.AddInstruction( HloInstruction::CreateConstant(std::move(literal))); @@ -248,7 +249,7 @@ XLA_TEST_F(CopyOpClientTest, Copy0x0) { auto empty = Literal::CreateFromShape(in_shape); XlaBuilder builder(TestName()); - auto param0 = builder.Parameter(0, in_shape, "input"); + Parameter(&builder, 0, in_shape, "input"); auto input_data = client_->TransferToServer(*empty).ConsumeValueOrDie(); auto actual = ExecuteAndTransfer(&builder, {input_data.get()}, &out_shape) diff --git a/tensorflow/compiler/xla/tests/cross_replica_sum_test.cc b/tensorflow/compiler/xla/tests/cross_replica_sum_test.cc index b151187c4b8f01c5b46ccadf27d2e22a7c902e98..d12a4e7fcd7813775a81677bcaa07af60ff9b477 100644 --- a/tensorflow/compiler/xla/tests/cross_replica_sum_test.cc +++ b/tensorflow/compiler/xla/tests/cross_replica_sum_test.cc @@ -13,7 +13,7 @@ See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/service/hlo_parser.h" #include "tensorflow/compiler/xla/shape_util.h" #include "tensorflow/compiler/xla/test.h" @@ -45,7 +45,7 @@ XLA_TEST_F(TrivialCrossReplicaSumTest, OneOperand) { })"; auto module = ParseHloString(module_str, GetModuleConfigForTest()).ValueOrDie(); - auto literal = Literal::CreateR1({1, 2, 3}); + auto literal = LiteralUtil::CreateR1({1, 2, 3}); EXPECT_EQ(*literal, *ExecuteAndTransfer(std::move(module), {literal.get()})); } @@ -66,10 +66,10 @@ XLA_TEST_F(TrivialCrossReplicaSumTest, MultipleOperands) { })"; auto module = ParseHloString(module_str, GetModuleConfigForTest()).ValueOrDie(); - auto literal0 = Literal::CreateR1({1, 2, 3}); - auto literal1 = Literal::CreateR1({10, 20}); + auto literal0 = LiteralUtil::CreateR1({1, 2, 3}); + auto literal1 = LiteralUtil::CreateR1({10, 20}); EXPECT_EQ( - *Literal::MakeTuple({literal0.get(), literal1.get()}), + *LiteralUtil::MakeTuple({literal0.get(), literal1.get()}), *ExecuteAndTransfer(std::move(module), {literal0.get(), literal1.get()})); } @@ -93,9 +93,9 @@ XLA_TEST_F(TrivialCrossReplicaSumTest, ConstantOperand) { })"; auto module = ParseHloString(module_str, GetModuleConfigForTest()).ValueOrDie(); - auto literal0 = Literal::CreateR1({1, 2, 3}); - auto literal1 = Literal::CreateR1({10, 20}); - EXPECT_EQ(*Literal::MakeTuple({literal0.get(), literal1.get()}), + auto literal0 = LiteralUtil::CreateR1({1, 2, 3}); + auto literal1 = LiteralUtil::CreateR1({10, 20}); + EXPECT_EQ(*LiteralUtil::MakeTuple({literal0.get(), literal1.get()}), *ExecuteAndTransfer(std::move(module), {literal0.get()})); } diff --git a/tensorflow/compiler/xla/tests/custom_call_test.cc b/tensorflow/compiler/xla/tests/custom_call_test.cc index b43d5c9ff5d75ee0e1b3c9ceb2bc295e631ac107..90f3d1b874f4da09104dc066c6642db1d2e77997 100644 --- a/tensorflow/compiler/xla/tests/custom_call_test.cc +++ b/tensorflow/compiler/xla/tests/custom_call_test.cc @@ -16,6 +16,7 @@ limitations under the License. #include #include +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/compiler/xla/literal_util.h" #include "tensorflow/compiler/xla/ptr_util.h" #include "tensorflow/compiler/xla/service/cpu/custom_call_target_registry.h" @@ -73,7 +74,7 @@ XLA_TEST_F(CustomCallTest, DISABLED_ON_GPU(CustomCallR0F32Add2)) { auto builder = HloComputation::Builder(TestName()); auto constant = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(42.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(42.0f))); builder.AddInstruction( HloInstruction::CreateCustomCall(r0f32_, {constant}, "R0F32Add2")); @@ -94,7 +95,7 @@ XLA_TEST_F(CustomCallTest, DISABLED_ON_GPU(CustomCallR2F32Reduce)) { array(1, 1) = 4.0f; auto constant = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR2FromArray2D(array))); + HloInstruction::CreateConstant(LiteralUtil::CreateR2FromArray2D(array))); builder.AddInstruction( HloInstruction::CreateCustomCall(r0f32_, {constant}, "R2F32ReduceSum")); @@ -110,7 +111,7 @@ XLA_TEST_F(CustomCallTest, auto b = HloComputation::Builder(TestName()); auto input = b.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR2FromArray2D( + HloInstruction::CreateConstant(LiteralUtil::CreateR2FromArray2D( Array2D{{1.0f, 2.0f}, {3.0f, 4.0f}}))); auto incremented = b.AddInstruction(HloInstruction::CreateCustomCall( ShapeUtil::MakeShape(F32, {1, 2, 2}), {input}, "Add1ToValues")); @@ -135,8 +136,8 @@ class CustomCallClientAPITest : public ClientLibraryTestBase {}; // are reserved for internal use. XLA_TEST_F(CustomCallClientAPITest, IllegalCustomCallTarget) { XlaBuilder builder(TestName()); - builder.CustomCall("$illegal", /*operands=*/{}, - ShapeUtil::MakeShape(F32, {1})); + CustomCall(&builder, "$illegal", /*operands=*/{}, + ShapeUtil::MakeShape(F32, {1})); StatusOr> result = Execute(&builder, /*arguments=*/{}); diff --git a/tensorflow/compiler/xla/tests/deallocation_test.cc b/tensorflow/compiler/xla/tests/deallocation_test.cc index bfe688e20d182d581c3e3b545ac2289413deef7c..d4b3aac85bff283515088f6e61c9d2bad11f60d3 100644 --- a/tensorflow/compiler/xla/tests/deallocation_test.cc +++ b/tensorflow/compiler/xla/tests/deallocation_test.cc @@ -48,7 +48,7 @@ class DeallocationTest : public ClientLibraryTestBase { TEST_F(DeallocationTest, DeallocateScalar) { XlaBuilder builder(TestName()); - builder.ConstantR0(42.0); + ConstantR0(&builder, 42.0); auto global_data = ExecuteAndCheckTransfer(&builder, {}); // A result can be transferred an arbitrary number of times. Add an extra @@ -66,7 +66,7 @@ TEST_F(DeallocationTest, DeallocateScalar) { TEST_F(DeallocationTest, DeallocateVector) { XlaBuilder builder(TestName()); - builder.ConstantR1({1.0, 2.0, 3.0, 4.0}); + ConstantR1(&builder, {1.0, 2.0, 3.0, 4.0}); auto global_data = ExecuteAndCheckTransfer(&builder, {}); ASSERT_IS_OK(client_->Unregister(*global_data)); @@ -79,7 +79,7 @@ TEST_F(DeallocationTest, DeallocateVector) { TEST_F(DeallocationTest, DeallocateEmptyVector) { XlaBuilder builder(TestName()); - builder.ConstantR1({}); + ConstantR1(&builder, {}); auto global_data = ExecuteAndCheckTransfer(&builder, {}); ASSERT_IS_OK(client_->Unregister(*global_data)); @@ -92,8 +92,8 @@ TEST_F(DeallocationTest, DeallocateEmptyVector) { XLA_TEST_F(DeallocationTest, DeallocateTuple) { XlaBuilder builder(TestName()); - builder.Tuple({builder.ConstantR0(42.0), - builder.ConstantR1({1.0, 2.0, 3.0})}); + Tuple(&builder, {ConstantR0(&builder, 42.0), + ConstantR1(&builder, {1.0, 2.0, 3.0})}); auto global_data = ExecuteAndCheckTransfer(&builder, {}); ASSERT_IS_OK(client_->Unregister(*global_data)); @@ -106,9 +106,10 @@ XLA_TEST_F(DeallocationTest, DeallocateTuple) { XLA_TEST_F(DeallocationTest, DeallocateTupleWithRepeatedElements) { XlaBuilder builder(TestName()); - auto element = builder.ConstantR0(42.0); - auto inner_tuple = builder.Tuple({builder.ConstantR0(42.0), element}); - builder.Tuple({element, inner_tuple, element}); + auto element = ConstantR0(&builder, 42.0); + auto inner_tuple = + Tuple(&builder, {ConstantR0(&builder, 42.0), element}); + Tuple(&builder, {element, inner_tuple, element}); auto global_data = ExecuteAndCheckTransfer(&builder, {}); ASSERT_IS_OK(client_->Unregister(*global_data)); @@ -122,9 +123,9 @@ XLA_TEST_F(DeallocationTest, DeallocateTupleWithRepeatedElements) { XLA_TEST_F(DeallocationTest, DeallocateNestedTuple) { XlaBuilder builder(TestName()); auto inner_tuple = - builder.Tuple({builder.ConstantR0(42.0), - builder.ConstantR1({1.0, 2.0, 3.0})}); - builder.Tuple({inner_tuple, builder.ConstantR1({0.123, 0.456})}); + Tuple(&builder, {ConstantR0(&builder, 42.0), + ConstantR1(&builder, {1.0, 2.0, 3.0})}); + Tuple(&builder, {inner_tuple, ConstantR1(&builder, {0.123, 0.456})}); auto global_data = ExecuteAndCheckTransfer(&builder, {}); ASSERT_IS_OK(client_->Unregister(*global_data)); diff --git a/tensorflow/compiler/xla/tests/deconstruct_tuple_test.cc b/tensorflow/compiler/xla/tests/deconstruct_tuple_test.cc index 12789fe66530fe03eb33316eda652336f29971ab..a6a233e71aabc47c78ea291b71f8b831f1c60323 100644 --- a/tensorflow/compiler/xla/tests/deconstruct_tuple_test.cc +++ b/tensorflow/compiler/xla/tests/deconstruct_tuple_test.cc @@ -20,7 +20,7 @@ limitations under the License. #include "tensorflow/compiler/xla/client/local_client.h" #include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/compiler/xla/client/xla_client/xla_computation.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/shape_util.h" #include "tensorflow/compiler/xla/statusor.h" #include "tensorflow/compiler/xla/test.h" @@ -54,9 +54,9 @@ class DeconstructTupleTest : public ClientLibraryTestBase { TEST_F(DeconstructTupleTest, DeconstructTuple) { XlaBuilder builder(TestName()); - auto const1 = builder.ConstantR1({1.0, 2.0, 3.0, 4.0}); - auto const2 = builder.ConstantR1({2.0, 4.0, 6.0, 8.0}); - builder.Tuple({const1, const2}); + auto const1 = ConstantR1(&builder, {1.0, 2.0, 3.0, 4.0}); + auto const2 = ConstantR1(&builder, {2.0, 4.0, 6.0, 8.0}); + Tuple(&builder, {const1, const2}); auto global_data = ExecuteAndCheckTransfer(&builder, {}); auto result_status = client_->DeconstructTuple(*global_data); @@ -73,9 +73,9 @@ TEST_F(DeconstructTupleTest, DeconstructTuple) { TEST_F(DeconstructTupleTest, DeconstructTupleTwice) { XlaBuilder builder(TestName()); - auto const1 = builder.ConstantR1({1.0, 2.0, 3.0, 4.0}); - auto const2 = builder.ConstantR1({2.0, 4.0, 6.0, 8.0}); - builder.Tuple({const1, const2}); + auto const1 = ConstantR1(&builder, {1.0, 2.0, 3.0, 4.0}); + auto const2 = ConstantR1(&builder, {2.0, 4.0, 6.0, 8.0}); + Tuple(&builder, {const1, const2}); auto global_data = ExecuteAndCheckTransfer(&builder, {}); auto result_status1 = client_->DeconstructTuple(*global_data); @@ -103,9 +103,9 @@ TEST_F(DeconstructTupleTest, DeconstructTupleTwice) { XLA_TEST_F(DeconstructTupleTest, DeconstructTupleRepeatedElement) { XlaBuilder builder(TestName()); - auto const1 = builder.ConstantR1({1.0, 2.0, 3.0, 4.0}); - auto const2 = builder.ConstantR1({2.0, 4.0, 6.0, 8.0}); - builder.Tuple({const1, const2, const2, const1}); + auto const1 = ConstantR1(&builder, {1.0, 2.0, 3.0, 4.0}); + auto const2 = ConstantR1(&builder, {2.0, 4.0, 6.0, 8.0}); + Tuple(&builder, {const1, const2, const2, const1}); auto global_data = ExecuteAndCheckTransfer(&builder, {}); auto result_status = client_->DeconstructTuple(*global_data); @@ -129,9 +129,9 @@ XLA_TEST_F(DeconstructTupleTest, DeconstructTupleRepeatedElement) { TEST_F(DeconstructTupleTest, DeconstructTupleThenDeallocate) { XlaBuilder builder(TestName()); - auto const1 = builder.ConstantR1({1.0, 2.0, 3.0, 4.0}); - auto const2 = builder.ConstantR1({2.0, 4.0, 6.0, 8.0}); - builder.Tuple({const1, const2, const1}); + auto const1 = ConstantR1(&builder, {1.0, 2.0, 3.0, 4.0}); + auto const2 = ConstantR1(&builder, {2.0, 4.0, 6.0, 8.0}); + Tuple(&builder, {const1, const2, const1}); auto global_data = ExecuteAndCheckTransfer(&builder, {}); auto result_status = client_->DeconstructTuple(*global_data); @@ -159,7 +159,7 @@ TEST_F(DeconstructTupleTest, DeconstructTupleThenDeallocate) { TEST_F(DeconstructTupleTest, DeconstructNonTuple) { XlaBuilder builder(TestName()); - builder.ConstantR1({1.0, 2.0, 3.0, 4.0}); + ConstantR1(&builder, {1.0, 2.0, 3.0, 4.0}); auto global_data = ExecuteAndCheckTransfer(&builder, {}); auto result_status = client_->DeconstructTuple(*global_data); @@ -171,11 +171,11 @@ TEST_F(DeconstructTupleTest, DeconstructNonTuple) { XLA_TEST_F(DeconstructTupleTest, DeconstructTupleFromParam) { XlaBuilder builder(TestName()); std::unique_ptr param0_literal = - Literal::CreateR1({3.14f, -100.25f}); + LiteralUtil::CreateR1({3.14f, -100.25f}); std::unique_ptr param0_data = client_->TransferToServer(*param0_literal).ConsumeValueOrDie(); - auto p = builder.Parameter(0, ShapeUtil::MakeShape(F32, {2}), "param0"); - builder.Tuple({p}); + auto p = Parameter(&builder, 0, ShapeUtil::MakeShape(F32, {2}), "param0"); + Tuple(&builder, {p}); auto global_data = ExecuteAndCheckTransfer(&builder, {param0_data.get()}); auto result_status = client_->DeconstructTuple(*global_data); @@ -186,9 +186,9 @@ XLA_TEST_F(DeconstructTupleTest, DeconstructTupleFromParam) { XLA_TEST_F(DeconstructTupleTest, DeconstructNestedTuple) { XlaBuilder builder(TestName()); - auto const1 = builder.ConstantR1({1.0, 2.0, 3.0, 4.0}); - auto const2 = builder.ConstantR1({2.0, 4.0, 6.0, 8.0}); - builder.Tuple({builder.Tuple({const1, const2}), const1}); + auto const1 = ConstantR1(&builder, {1.0, 2.0, 3.0, 4.0}); + auto const2 = ConstantR1(&builder, {2.0, 4.0, 6.0, 8.0}); + Tuple(&builder, {Tuple(&builder, {const1, const2}), const1}); auto global_data = ExecuteAndCheckTransfer(&builder, {}); auto result_status = client_->DeconstructTuple(*global_data); diff --git a/tensorflow/compiler/xla/tests/deep_graph_test.cc b/tensorflow/compiler/xla/tests/deep_graph_test.cc index 085a5105aca1c173a7cbc211aebbeb5b254b0753..810947ab01b69b10b6ae60c551bd7aba10a6313d 100644 --- a/tensorflow/compiler/xla/tests/deep_graph_test.cc +++ b/tensorflow/compiler/xla/tests/deep_graph_test.cc @@ -30,7 +30,7 @@ TEST_F(ClientLibraryTestBase, DeepGraph) { auto y_data = CreateR0Parameter(1, 1, "y", &b, &y); XlaOp z = x; for (int i = 0; i < kDepth; ++i) { - z = b.Add(z, y); + z = Add(z, y); } ComputeAndCompareR0(&b, /*expected=*/kDepth + 3, {x_data.get(), y_data.get()}); diff --git a/tensorflow/compiler/xla/tests/dot_operation_test.cc b/tensorflow/compiler/xla/tests/dot_operation_test.cc index 0fd846cef8095a857dd7b2c12d8afdf409e2bd66..d86fd7cc2d4da10ed726ca11a6d9f86287a5d11e 100644 --- a/tensorflow/compiler/xla/tests/dot_operation_test.cc +++ b/tensorflow/compiler/xla/tests/dot_operation_test.cc @@ -67,15 +67,16 @@ XLA_TEST_F(DotOperationTest, DotOfInputTupleElem) { XlaOp param; auto param_data = CreateParameterAndTransferLiteral( 0, - *Literal::MakeTuple({Literal::CreateR2({{1, 2}, {3, 4}}).get(), - Literal::CreateR2({{5, 6}, {7, 8}}).get()}), + *LiteralUtil::MakeTuple( + {LiteralUtil::CreateR2({{1, 2}, {3, 4}}).get(), + LiteralUtil::CreateR2({{5, 6}, {7, 8}}).get()}), "arg0", &builder, ¶m); - auto lhs = builder.GetTupleElement(param, 0); - auto rhs = builder.GetTupleElement(param, 1); - builder.Dot(lhs, rhs); + auto lhs = GetTupleElement(param, 0); + auto rhs = GetTupleElement(param, 1); + Dot(lhs, rhs); ComputeAndCompareLiteral(&builder, - *Literal::CreateR2({{19, 22}, {43, 50}}), + *LiteralUtil::CreateR2({{19, 22}, {43, 50}}), {param_data.get()}); } @@ -87,9 +88,9 @@ XLA_TYPED_TEST(DotOperationTest_F16F32F64CF64, ZeroElementVectorDot) { using T = TypeParam; XlaBuilder builder(this->TestName()); - auto lhs = builder.ConstantR1({}); - auto rhs = builder.ConstantR1({}); - auto result = builder.Dot(lhs, rhs); + auto lhs = ConstantR1(&builder, {}); + auto rhs = ConstantR1(&builder, {}); + Dot(lhs, rhs); this->template ComputeAndCompareR0(&builder, static_cast(0.0), {}, this->error_spec_); @@ -102,9 +103,9 @@ TYPED_TEST_CASE(DotOperationTest_F16F32F64, TypesF16F32F64); XLA_TYPED_TEST(DotOperationTest_F16F32F64, TrivialMatrixVectorDot) { using T = TypeParam; XlaBuilder builder(this->TestName()); - auto lhs = builder.ConstantR2FromArray2D({{3.0f, 4.0f}}); - auto rhs = builder.ConstantFromArray({3.0f, 4.0f}); - auto result = builder.Dot(lhs, rhs); + auto lhs = ConstantR2FromArray2D(&builder, {{3.0f, 4.0f}}); + auto rhs = ConstantFromArray(&builder, {3.0f, 4.0f}); + Dot(lhs, rhs); this->template ComputeAndCompareR1(&builder, {static_cast(25.0f)}, {}, this->error_spec_); @@ -113,9 +114,9 @@ XLA_TYPED_TEST(DotOperationTest_F16F32F64, TrivialMatrixVectorDot) { XLA_TYPED_TEST(DotOperationTest_F16F32F64, OneElementVectorDot) { using T = TypeParam; XlaBuilder builder(this->TestName()); - auto lhs = builder.ConstantR1({static_cast(2.0f)}); - auto rhs = builder.ConstantR1({static_cast(3.0f)}); - auto result = builder.Dot(lhs, rhs); + auto lhs = ConstantR1(&builder, {static_cast(2.0f)}); + auto rhs = ConstantR1(&builder, {static_cast(3.0f)}); + Dot(lhs, rhs); this->template ComputeAndCompareR0(&builder, static_cast(6.0f), {}, this->error_spec_); @@ -124,9 +125,9 @@ XLA_TYPED_TEST(DotOperationTest_F16F32F64, OneElementVectorDot) { XLA_TYPED_TEST(DotOperationTest_F16F32F64, VectorDot) { using T = TypeParam; XlaBuilder builder(this->TestName()); - auto lhs = builder.ConstantFromArray({1.0f, 2.5f, 42.0f}); - auto rhs = builder.ConstantFromArray({11.0f, -1.0f, 0.5f}); - auto result = builder.Dot(lhs, rhs); + auto lhs = ConstantFromArray(&builder, {1.0f, 2.5f, 42.0f}); + auto rhs = ConstantFromArray(&builder, {11.0f, -1.0f, 0.5f}); + Dot(lhs, rhs); this->template ComputeAndCompareR0(&builder, static_cast(29.5f), {}, this->error_spec_); @@ -139,9 +140,9 @@ std::vector MinorToMajorForIsRowMajor(bool row_major) { XLA_TYPED_TEST(DotOperationTest_F16F32F64, Dot_0x2_2x0) { using T = TypeParam; XlaBuilder builder(this->TestName()); - auto lhs = builder.ConstantR2FromArray2D(Array2D(0, 2)); - auto rhs = builder.ConstantR2FromArray2D(Array2D(2, 0)); - auto result = builder.Dot(lhs, rhs); + auto lhs = ConstantR2FromArray2D(&builder, Array2D(0, 2)); + auto rhs = ConstantR2FromArray2D(&builder, Array2D(2, 0)); + Dot(lhs, rhs); this->template ComputeAndCompareR2(&builder, Array2D(0, 0), {}, this->error_spec_); @@ -150,10 +151,10 @@ XLA_TYPED_TEST(DotOperationTest_F16F32F64, Dot_0x2_2x0) { XLA_TYPED_TEST(DotOperationTest_F16F32F64, Dot_0x2_2x3) { using T = TypeParam; XlaBuilder builder(this->TestName()); - auto lhs = builder.ConstantR2FromArray2D(Array2D(0, 2)); - auto rhs = builder.ConstantR2FromArray2D( - {{7.0f, 8.0f, 9.0f}, {42.0f, 77.0f, 101.0f}}); - auto result = builder.Dot(lhs, rhs); + auto lhs = ConstantR2FromArray2D(&builder, Array2D(0, 2)); + auto rhs = ConstantR2FromArray2D( + &builder, {{7.0f, 8.0f, 9.0f}, {42.0f, 77.0f, 101.0f}}); + Dot(lhs, rhs); this->template ComputeAndCompareR2(&builder, Array2D(0, 3), {}, this->error_spec_); @@ -162,10 +163,10 @@ XLA_TYPED_TEST(DotOperationTest_F16F32F64, Dot_0x2_2x3) { XLA_TYPED_TEST(DotOperationTest_F16F32F64, Dot_3x2_2x0) { using T = TypeParam; XlaBuilder builder(this->TestName()); - auto lhs = builder.ConstantR2FromArray2D( - {{7.0f, 8.0f}, {9.0f, 42.0f}, {77.0f, 101.0f}}); - auto rhs = builder.ConstantR2FromArray2D(Array2D(2, 0)); - auto result = builder.Dot(lhs, rhs); + auto lhs = ConstantR2FromArray2D( + &builder, {{7.0f, 8.0f}, {9.0f, 42.0f}, {77.0f, 101.0f}}); + auto rhs = ConstantR2FromArray2D(&builder, Array2D(2, 0)); + Dot(lhs, rhs); this->template ComputeAndCompareR2(&builder, Array2D(3, 0), {}, this->error_spec_); @@ -174,9 +175,9 @@ XLA_TYPED_TEST(DotOperationTest_F16F32F64, Dot_3x2_2x0) { XLA_TYPED_TEST(DotOperationTest_F16F32F64, Dot_2x0_0x2) { using T = TypeParam; XlaBuilder builder(this->TestName()); - auto lhs = builder.ConstantR2FromArray2D(Array2D(2, 0)); - auto rhs = builder.ConstantR2FromArray2D(Array2D(0, 2)); - auto result = builder.Dot(lhs, rhs); + auto lhs = ConstantR2FromArray2D(&builder, Array2D(2, 0)); + auto rhs = ConstantR2FromArray2D(&builder, Array2D(0, 2)); + Dot(lhs, rhs); this->template ComputeAndCompareR2( &builder, Array2D(2, 2, static_cast(0.0f)), {}, this->error_spec_); @@ -186,19 +187,19 @@ XLA_TYPED_TEST(DotOperationTest_F16F32F64, FusedDot) { using T = TypeParam; XlaBuilder builder(this->TestName()); auto param0 = - builder.Parameter(0, ShapeUtil::MakeShapeWithType({2, 4}), "arg0"); + Parameter(&builder, 0, ShapeUtil::MakeShapeWithType({2, 4}), "arg0"); auto param1 = - builder.Parameter(1, ShapeUtil::MakeShapeWithType({4, 1}), "arg1"); - auto exp0 = builder.Exp(param0); - auto result = builder.Dot(exp0, param1); + Parameter(&builder, 1, ShapeUtil::MakeShapeWithType({4, 1}), "arg1"); + auto exp0 = Exp(param0); + Dot(exp0, param1); auto lhs_handle = this->client_ - ->TransferToServer(*Literal::CreateR2FromArray2D( + ->TransferToServer(*LiteralUtil::CreateR2FromArray2D( {{1.0f, 2.0f, 3.0f, 4.0f}, {-1.0f, -2.0f, -3.0f, -4.0f}})) .ConsumeValueOrDie(); auto rhs_handle = this->client_ - ->TransferToServer(*Literal::CreateR2FromArray2D( + ->TransferToServer(*LiteralUtil::CreateR2FromArray2D( {{1.0f}, {2.0f}, {3.0f}, {4.0f}})) .ConsumeValueOrDie(); @@ -217,23 +218,22 @@ class SquareMatrixDot : public DotOperationTest { void TestImpl(bool lhs_row_major, bool rhs_row_major) { auto lhs_handle = client_ - ->TransferToServer(*Literal::CreateFromArrayWithLayout( + ->TransferToServer(*LiteralUtil::CreateFromArrayWithLayout( {{1.0f, 2.0f}, {3.0f, -4.0f}}, LayoutUtil::MakeLayout( MinorToMajorForIsRowMajor(lhs_row_major)))) .ConsumeValueOrDie(); auto rhs_handle = client_ - ->TransferToServer(*Literal::CreateFromArrayWithLayout( + ->TransferToServer(*LiteralUtil::CreateFromArrayWithLayout( {{1.0f, 6.0f}, {7.0f, -4.0f}}, LayoutUtil::MakeLayout( MinorToMajorForIsRowMajor(rhs_row_major)))) .ConsumeValueOrDie(); XlaBuilder builder(TestName()); auto prim_type = primitive_util::NativeToPrimitiveType(); - auto result = builder.Dot( - builder.Parameter(0, ShapeUtil::MakeShape(prim_type, {2, 2}), "lhs"), - builder.Parameter(1, ShapeUtil::MakeShape(prim_type, {2, 2}), "rhs")); + Dot(Parameter(&builder, 0, ShapeUtil::MakeShape(prim_type, {2, 2}), "lhs"), + Parameter(&builder, 1, ShapeUtil::MakeShape(prim_type, {2, 2}), "rhs")); Array2D expected({{15.0f, -2.0f}, {-25.0f, 34.0f}}); ComputeAndCompareR2(&builder, expected, @@ -287,9 +287,10 @@ void ParametricDotTest::TestImpl() { std::unique_ptr> dot_lhs_data = MakeLinspaceArray2D(0.0, 1.0, param.m, param.k); - std::unique_ptr dot_lhs_lit = Literal::CreateR2FromArray2DWithLayout( - *dot_lhs_data, LayoutUtil::MakeLayout( - MinorToMajorForIsRowMajor(param.dot_lhs_row_major))); + std::unique_ptr dot_lhs_lit = + LiteralUtil::CreateR2FromArray2DWithLayout( + *dot_lhs_data, LayoutUtil::MakeLayout(MinorToMajorForIsRowMajor( + param.dot_lhs_row_major))); std::unique_ptr dot_lhs_handle = client_->TransferToServer(*dot_lhs_lit).ConsumeValueOrDie(); @@ -298,7 +299,7 @@ void ParametricDotTest::TestImpl() { Layout rhs_layout = LayoutUtil::MakeLayout( MinorToMajorForIsRowMajor(param.dot_rhs_row_major)); std::unique_ptr dot_rhs_lit = - Literal::CreateR2FromArray2DWithLayout(*dot_rhs_data, rhs_layout); + LiteralUtil::CreateR2FromArray2DWithLayout(*dot_rhs_data, rhs_layout); std::unique_ptr dot_rhs_handle = client_->TransferToServer(*dot_rhs_lit).ConsumeValueOrDie(); @@ -308,7 +309,7 @@ void ParametricDotTest::TestImpl() { if (param.has_addend) { addend_data = MakeLinspaceArray2D(0.0, 1.0, param.m, param.n); - addend_lit = Literal::CreateR2FromArray2DWithLayout( + addend_lit = LiteralUtil::CreateR2FromArray2DWithLayout( *addend_data, LayoutUtil::MakeLayout( MinorToMajorForIsRowMajor(param.addend_row_major))); addend_handle = client_->TransferToServer(*addend_lit).ConsumeValueOrDie(); @@ -316,26 +317,26 @@ void ParametricDotTest::TestImpl() { XlaBuilder builder(TestName()); auto prim_type = primitive_util::NativeToPrimitiveType(); - auto result = builder.Dot( - builder.Parameter(0, - ShapeUtil::MakeShapeWithLayout( - prim_type, {param.m, param.k}, - MinorToMajorForIsRowMajor(param.dot_lhs_row_major)), - "dot_lhs"), - builder.Parameter(1, - ShapeUtil::MakeShapeWithLayout( - prim_type, {param.k, param.n}, - MinorToMajorForIsRowMajor(param.dot_rhs_row_major)), - "dot_rhs")); + auto result = + Dot(Parameter(&builder, 0, + ShapeUtil::MakeShapeWithLayout( + prim_type, {param.m, param.k}, + MinorToMajorForIsRowMajor(param.dot_lhs_row_major)), + "dot_lhs"), + Parameter(&builder, 1, + ShapeUtil::MakeShapeWithLayout( + prim_type, {param.k, param.n}, + MinorToMajorForIsRowMajor(param.dot_rhs_row_major)), + "dot_rhs")); if (param.has_addend) { - result = builder.Add( - result, builder.Parameter( - 2, - ShapeUtil::MakeShapeWithLayout( - prim_type, {param.m, param.n}, - MinorToMajorForIsRowMajor(param.addend_row_major)), - "addend")); + result = + Add(result, + Parameter(&builder, 2, + ShapeUtil::MakeShapeWithLayout( + prim_type, {param.m, param.n}, + MinorToMajorForIsRowMajor(param.addend_row_major)), + "addend")); } std::unique_ptr> expected; @@ -477,14 +478,14 @@ class NonsquareMatrixDot : public DotOperationTest { void TestImpl(bool lhs_row_major, bool rhs_row_major) { auto lhs_handle = client_ - ->TransferToServer(*Literal::CreateFromArrayWithLayout( + ->TransferToServer(*LiteralUtil::CreateFromArrayWithLayout( {{1.0f, 2.0f, 3.0f}, {3.0f, -4.0f, -1.0f}}, LayoutUtil::MakeLayout( MinorToMajorForIsRowMajor(lhs_row_major)))) .ConsumeValueOrDie(); auto rhs_handle = client_ - ->TransferToServer(*Literal::CreateFromArrayWithLayout( + ->TransferToServer(*LiteralUtil::CreateFromArrayWithLayout( {{1.0f, 6.0f}, {2.0f, 3.0f}, {7.0f, -4.0f}}, LayoutUtil::MakeLayout( MinorToMajorForIsRowMajor(rhs_row_major)))) @@ -492,9 +493,8 @@ class NonsquareMatrixDot : public DotOperationTest { XlaBuilder builder(TestName()); auto prim_type = primitive_util::NativeToPrimitiveType(); - auto result = builder.Dot( - builder.Parameter(0, ShapeUtil::MakeShape(prim_type, {2, 3}), "lhs"), - builder.Parameter(1, ShapeUtil::MakeShape(prim_type, {3, 2}), "rhs")); + Dot(Parameter(&builder, 0, ShapeUtil::MakeShape(prim_type, {2, 3}), "lhs"), + Parameter(&builder, 1, ShapeUtil::MakeShape(prim_type, {3, 2}), "rhs")); Array2D expected({{26.0f, 0.0f}, {-12.0f, 10.0f}}); @@ -512,21 +512,20 @@ XLA_TYPED_TEST(NonsquareMatrixDot, TestTT) { this->TestImpl(true, true); } XLA_TEST_F(DotOperationTest, MatrixVectorC64) { auto lhs_handle = client_ - ->TransferToServer(*Literal::CreateR2WithLayout( + ->TransferToServer(*LiteralUtil::CreateR2WithLayout( {{1.0, 2.0, 3.0, -4.0}}, LayoutUtil::MakeLayout({1, 0}))) .ConsumeValueOrDie(); auto rhs_handle = client_ - ->TransferToServer(*Literal::CreateR2WithLayout( + ->TransferToServer(*LiteralUtil::CreateR2WithLayout( {{1.0, 1.0}, {2.0, 2.0}, {3.0, 3.0}, {-4.0, 4.0}}, LayoutUtil::MakeLayout({1, 0}))) .ConsumeValueOrDie(); XlaBuilder builder(TestName()); auto prim_type = primitive_util::NativeToPrimitiveType(); - auto result = builder.Dot( - builder.Parameter(0, ShapeUtil::MakeShape(prim_type, {1, 4}), "lhs"), - builder.Parameter(1, ShapeUtil::MakeShape(prim_type, {4, 2}), "rhs")); + Dot(Parameter(&builder, 0, ShapeUtil::MakeShape(prim_type, {1, 4}), "lhs"), + Parameter(&builder, 1, ShapeUtil::MakeShape(prim_type, {4, 2}), "rhs")); Array2D expected({{30.0, -2.0}}); @@ -538,11 +537,13 @@ XLA_TYPED_TEST(DotOperationTest_F16F32F64, ConcurrentMatMult) { using T = TypeParam; XlaBuilder builder(this->TestName()); - auto matrix1 = builder.ConstantR2FromArray2D({{1.0f, 2.0f}, {3.0f, 4.0f}}); - auto matrix2 = builder.ConstantR2FromArray2D({{5.0f, 6.0f}, {7.0f, 8.0f}}); - auto matrix12 = builder.Dot(matrix1, matrix2); - auto matrix21 = builder.Dot(matrix2, matrix1); - builder.Add(matrix12, matrix21); + auto matrix1 = + ConstantR2FromArray2D(&builder, {{1.0f, 2.0f}, {3.0f, 4.0f}}); + auto matrix2 = + ConstantR2FromArray2D(&builder, {{5.0f, 6.0f}, {7.0f, 8.0f}}); + auto matrix12 = Dot(matrix1, matrix2); + auto matrix21 = Dot(matrix2, matrix1); + Add(matrix12, matrix21); Array2D expected({{42.0f, 56.0f}, {74.0f, 96.0f}}); this->template ComputeAndCompareR2(&builder, expected, {}, @@ -559,32 +560,32 @@ TYPED_TEST_CASE(DotOperationTestForBatchMatMul, TypesF16F32F64); XLA_TYPED_TEST(DotOperationTestForBatchMatMul, Types) { using T = TypeParam; XlaBuilder builder(this->TestName()); - auto x = - builder.Parameter(0, ShapeUtil::MakeShapeWithType({2, 2, 2, 2}), "x"); - auto y = - builder.Parameter(1, ShapeUtil::MakeShapeWithType({2, 2, 2, 2}), "y"); + auto x = Parameter(&builder, 0, ShapeUtil::MakeShapeWithType({2, 2, 2, 2}), + "x"); + auto y = Parameter(&builder, 1, ShapeUtil::MakeShapeWithType({2, 2, 2, 2}), + "y"); - auto x_flat = builder.Reshape(x, {0, 1, 2, 3}, {4, 2, 2}); - auto y_flat = builder.Reshape(y, {0, 1, 2, 3}, {4, 2, 2}); + auto x_flat = Reshape(x, {0, 1, 2, 3}, {4, 2, 2}); + auto y_flat = Reshape(y, {0, 1, 2, 3}, {4, 2, 2}); // Slice batches into individual matrices and multiply them. std::vector out_slices; for (int i = 0; i < 4; ++i) { // Slice off individual matrices and reshape to 2D tensors. - auto x_slice = builder.Slice(x_flat, {i, 0, 0}, {i + 1, 2, 2}, {1, 1, 1}); - x_slice = builder.Reshape(x_slice, {0, 1, 2}, {2, 2}); - auto y_slice = builder.Slice(y_flat, {i, 0, 0}, {i + 1, 2, 2}, {1, 1, 1}); - y_slice = builder.Reshape(y_slice, {0, 1, 2}, {2, 2}); + auto x_slice = Slice(x_flat, {i, 0, 0}, {i + 1, 2, 2}, {1, 1, 1}); + x_slice = Reshape(x_slice, {0, 1, 2}, {2, 2}); + auto y_slice = Slice(y_flat, {i, 0, 0}, {i + 1, 2, 2}, {1, 1, 1}); + y_slice = Reshape(y_slice, {0, 1, 2}, {2, 2}); - auto out = builder.Dot(x_slice, y_slice); - out = builder.Reshape(out, {0, 1}, {1, 2, 2}); + auto out = Dot(x_slice, y_slice); + out = Reshape(out, {0, 1}, {1, 2, 2}); out_slices.push_back(out); } - auto out_flat = builder.ConcatInDim(out_slices, 0); - builder.Reshape(out_flat, {0, 1, 2}, {2, 2, 2, 2}); + auto out_flat = ConcatInDim(&builder, out_slices, 0); + Reshape(out_flat, {0, 1, 2}, {2, 2, 2, 2}); auto x_data = this->client_ - ->TransferToServer(*Literal::CreateR4FromArray4D( + ->TransferToServer(*LiteralUtil::CreateR4FromArray4D( {{{{1000.0f, 100.0f}, {10.0f, 1.0f}}, {{2000.0f, 200.0f}, {20.0f, 2.0f}}}, {{{3000.0f, 300.0f}, {30.0f, 3.0f}}, @@ -592,7 +593,7 @@ XLA_TYPED_TEST(DotOperationTestForBatchMatMul, Types) { .ConsumeValueOrDie(); auto y_data = this->client_ - ->TransferToServer(*Literal::CreateR4FromArray4D( + ->TransferToServer(*LiteralUtil::CreateR4FromArray4D( {{{{1.0f, 2.0f}, {3.0f, 4.0f}}, {{5.0f, 6.0f}, {7.0f, 8.0f}}}, {{{11.0f, 22.0f}, {33.0f, 44.0f}}, {{55.0f, 66.0f}, {77.0f, 88.0f}}}})) @@ -616,9 +617,9 @@ XLA_TYPED_TEST(DotOperationTest_F16F32F64, GeneralMatMul) { XlaBuilder builder(this->TestName()); auto x = - builder.Parameter(0, ShapeUtil::MakeShapeWithType({2, 2, 2}), "x"); + Parameter(&builder, 0, ShapeUtil::MakeShapeWithType({2, 2, 2}), "x"); auto y = - builder.Parameter(1, ShapeUtil::MakeShapeWithType({2, 2, 2}), "y"); + Parameter(&builder, 1, ShapeUtil::MakeShapeWithType({2, 2, 2}), "y"); DotDimensionNumbers dnums; dnums.add_lhs_contracting_dimensions(2); @@ -626,17 +627,17 @@ XLA_TYPED_TEST(DotOperationTest_F16F32F64, GeneralMatMul) { dnums.add_lhs_batch_dimensions(0); dnums.add_rhs_batch_dimensions(0); - auto out = builder.DotGeneral(x, y, dnums); + DotGeneral(x, y, dnums); auto x_data = this->client_ - ->TransferToServer(*Literal::CreateR3FromArray3D( + ->TransferToServer(*LiteralUtil::CreateR3FromArray3D( {{{1.0f, 2.0f}, {3.0f, 4.0f}}, {{5.0f, 6.0f}, {7.0f, 8.0f}}})) .ConsumeValueOrDie(); auto y_data = this->client_ - ->TransferToServer(*Literal::CreateR3FromArray3D( + ->TransferToServer(*LiteralUtil::CreateR3FromArray3D( {{{1.0f, 0.0f}, {0.0f, 1.0f}}, {{1.0f, 0.0f}, {0.0f, 1.0f}}})) .ConsumeValueOrDie(); @@ -665,32 +666,36 @@ XLA_TYPED_TEST(DotOperationTest_F16F32F64, TransposeFolding) { } auto lhs_handle = this->client_ - ->TransferToServer(*Literal::CreateR2FromArray2DWithLayout( - *lhs, LayoutUtil::MakeLayout( - MinorToMajorForIsRowMajor(row_major)))) + ->TransferToServer( + *LiteralUtil::CreateR2FromArray2DWithLayout( + *lhs, LayoutUtil::MakeLayout( + MinorToMajorForIsRowMajor(row_major)))) .ConsumeValueOrDie(); auto rhs_handle = this->client_ - ->TransferToServer(*Literal::CreateR2FromArray2DWithLayout( - *rhs, LayoutUtil::MakeLayout( - MinorToMajorForIsRowMajor(row_major)))) + ->TransferToServer( + *LiteralUtil::CreateR2FromArray2DWithLayout( + *rhs, LayoutUtil::MakeLayout( + MinorToMajorForIsRowMajor(row_major)))) .ConsumeValueOrDie(); XlaBuilder builder(this->TestName()); auto prim_type = primitive_util::NativeToPrimitiveType(); - auto lhs_arg = builder.Parameter( - 0, ShapeUtil::MakeShape(prim_type, {lhs->height(), lhs->width()}), + auto lhs_arg = Parameter( + &builder, 0, + ShapeUtil::MakeShape(prim_type, {lhs->height(), lhs->width()}), "lhs"); - auto rhs_arg = builder.Parameter( - 1, ShapeUtil::MakeShape(prim_type, {rhs->height(), rhs->width()}), + auto rhs_arg = Parameter( + &builder, 1, + ShapeUtil::MakeShape(prim_type, {rhs->height(), rhs->width()}), "rhs"); if (transpose_lhs) { - lhs_arg = builder.Transpose(lhs_arg, {1, 0}); + lhs_arg = Transpose(lhs_arg, {1, 0}); } if (transpose_rhs) { - rhs_arg = builder.Transpose(rhs_arg, {1, 0}); + rhs_arg = Transpose(rhs_arg, {1, 0}); } - auto result = builder.Dot(lhs_arg, rhs_arg); + Dot(lhs_arg, rhs_arg); Array2D expected({{26.0f, 0.0f}, {-12.0f, 10.0f}}); VLOG(1) << "TestTransposeFolding " << transpose_lhs << " " @@ -713,15 +718,15 @@ XLA_TYPED_TEST(DotOperationTest_F16F32F64, {6.0f, 5.0f, 4.0f, 3.0f, 2.0f, 1.0f}})); XlaBuilder builder(this->TestName()); - auto lhs_constant = builder.ConstantR2FromArray2D(*constant_lhs_array); - auto rhs_arg_0 = builder.Parameter(0, ShapeUtil::MakeShape(prim_type, {2, 2}), - "rhs_arg_0"); - auto rhs_arg_1 = builder.Parameter(1, ShapeUtil::MakeShape(prim_type, {3, 2}), - "rhs_arg_1"); - auto rhs_arg_2 = builder.Parameter(2, ShapeUtil::MakeShape(prim_type, {1, 2}), - "rhs_arg_2"); - auto result = builder.Dot( - lhs_constant, builder.ConcatInDim({rhs_arg_0, rhs_arg_1, rhs_arg_2}, 0)); + auto lhs_constant = ConstantR2FromArray2D(&builder, *constant_lhs_array); + auto rhs_arg_0 = Parameter( + &builder, 0, ShapeUtil::MakeShape(prim_type, {2, 2}), "rhs_arg_0"); + auto rhs_arg_1 = Parameter( + &builder, 1, ShapeUtil::MakeShape(prim_type, {3, 2}), "rhs_arg_1"); + auto rhs_arg_2 = Parameter( + &builder, 2, ShapeUtil::MakeShape(prim_type, {1, 2}), "rhs_arg_2"); + Dot(lhs_constant, + ConcatInDim(&builder, {rhs_arg_0, rhs_arg_1, rhs_arg_2}, 0)); std::unique_ptr> arg_0_value_array( new Array2D({{1.0f, 2.0f}, {3.0f, 4.0f}})); @@ -732,15 +737,15 @@ XLA_TYPED_TEST(DotOperationTest_F16F32F64, TF_ASSERT_OK_AND_ASSIGN( auto arg_0_value, this->client_->TransferToServer( - *Literal::CreateR2FromArray2D(*arg_0_value_array))); + *LiteralUtil::CreateR2FromArray2D(*arg_0_value_array))); TF_ASSERT_OK_AND_ASSIGN( auto arg_1_value, this->client_->TransferToServer( - *Literal::CreateR2FromArray2D(*arg_1_value_array))); + *LiteralUtil::CreateR2FromArray2D(*arg_1_value_array))); TF_ASSERT_OK_AND_ASSIGN( auto arg_2_value, this->client_->TransferToServer( - *Literal::CreateR2FromArray2D(*arg_2_value_array))); + *LiteralUtil::CreateR2FromArray2D(*arg_2_value_array))); Array2D expected({{53.0f, 74.0f}, {45.0f, 66.0f}}); this->template ComputeAndCompareR2( @@ -761,15 +766,15 @@ XLA_TYPED_TEST(DotOperationTest_F16F32F64, {2.0f, 1.0f}})); XlaBuilder builder(this->TestName()); - auto rhs_constant = builder.ConstantR2FromArray2D(*constant_rhs_array); - auto lhs_arg_0 = builder.Parameter(0, ShapeUtil::MakeShapeWithType({2, 2}), - "lhs_arg_0"); - auto lhs_arg_1 = builder.Parameter(1, ShapeUtil::MakeShapeWithType({2, 3}), - "lhs_arg_1"); - auto lhs_arg_2 = builder.Parameter(2, ShapeUtil::MakeShapeWithType({2, 1}), - "lhs_arg_2"); - auto result = builder.Dot( - builder.ConcatInDim({lhs_arg_0, lhs_arg_1, lhs_arg_2}, 1), rhs_constant); + auto rhs_constant = ConstantR2FromArray2D(&builder, *constant_rhs_array); + auto lhs_arg_0 = Parameter( + &builder, 0, ShapeUtil::MakeShapeWithType({2, 2}), "lhs_arg_0"); + auto lhs_arg_1 = Parameter( + &builder, 1, ShapeUtil::MakeShapeWithType({2, 3}), "lhs_arg_1"); + auto lhs_arg_2 = Parameter( + &builder, 2, ShapeUtil::MakeShapeWithType({2, 1}), "lhs_arg_2"); + Dot(ConcatInDim(&builder, {lhs_arg_0, lhs_arg_1, lhs_arg_2}, 1), + rhs_constant); std::unique_ptr> arg_0_value_array( new Array2D({{1.0f, 2.0f}, {3.0f, 4.0f}})); @@ -781,15 +786,15 @@ XLA_TYPED_TEST(DotOperationTest_F16F32F64, TF_ASSERT_OK_AND_ASSIGN( auto arg_0_value, this->client_->TransferToServer( - *Literal::CreateR2FromArray2D(*arg_0_value_array))); + *LiteralUtil::CreateR2FromArray2D(*arg_0_value_array))); TF_ASSERT_OK_AND_ASSIGN( auto arg_1_value, this->client_->TransferToServer( - *Literal::CreateR2FromArray2D(*arg_1_value_array))); + *LiteralUtil::CreateR2FromArray2D(*arg_1_value_array))); TF_ASSERT_OK_AND_ASSIGN( auto arg_2_value, this->client_->TransferToServer( - *Literal::CreateR2FromArray2D(*arg_2_value_array))); + *LiteralUtil::CreateR2FromArray2D(*arg_2_value_array))); Array2D expected({{38.0f, 36.0f}, {93.0f, 91.0f}}); this->template ComputeAndCompareR2( @@ -811,16 +816,15 @@ XLA_TEST_F(DotOperationTest, DotOfGatherOptimizationWithConstRHSClassicMM) { // Dot result to slice from: {{114, 105, 96}, {96, 105, 114}} XlaBuilder builder(TestName()); - auto lhs_constant = builder.ConstantR2FromArray2D(*constant_lhs_array); - auto rhs_constant = builder.ConstantR2FromArray2D(*constant_rhs_array); - auto start_constant = builder.ConstantR1({1, 0}); - auto dynamic_slice = - builder.DynamicSlice(lhs_constant, start_constant, {1, 6}); + auto lhs_constant = ConstantR2FromArray2D(&builder, *constant_lhs_array); + auto rhs_constant = ConstantR2FromArray2D(&builder, *constant_rhs_array); + auto start_constant = ConstantR1(&builder, {1, 0}); + auto dynamic_slice = DynamicSlice(lhs_constant, start_constant, {1, 6}); DotDimensionNumbers dot_dnums; dot_dnums.add_lhs_contracting_dimensions(1); dot_dnums.add_rhs_contracting_dimensions(0); - auto result = builder.DotGeneral(dynamic_slice, rhs_constant, dot_dnums); + DotGeneral(dynamic_slice, rhs_constant, dot_dnums); Array2D expected({{96.0, 105.0, 114.0}}); ComputeAndCompareR2(&builder, expected, {}, error_spec_); @@ -839,25 +843,23 @@ XLA_TEST_F(DotOperationTest, DotOfGatherOptimizationWithConstLHSClassicMM) { // Dot result to slice from: {{114, 105, 96}, {96, 105, 114}} XlaBuilder builder(TestName()); - auto lhs_constant = builder.ConstantR2FromArray2D(*constant_lhs_array); - auto rhs_constant = builder.ConstantR2FromArray2D(*constant_rhs_array); - auto start_constant = builder.ConstantR1({0, 1}); - auto dynamic_slice = - builder.DynamicSlice(rhs_constant, start_constant, {6, 1}); + auto lhs_constant = ConstantR2FromArray2D(&builder, *constant_lhs_array); + auto rhs_constant = ConstantR2FromArray2D(&builder, *constant_rhs_array); + auto start_constant = ConstantR1(&builder, {0, 1}); + auto dynamic_slice = DynamicSlice(rhs_constant, start_constant, {6, 1}); DotDimensionNumbers dot_dnums; dot_dnums.add_lhs_contracting_dimensions(1); dot_dnums.add_rhs_contracting_dimensions(0); - auto result = builder.DotGeneral(lhs_constant, dynamic_slice, dot_dnums); + DotGeneral(lhs_constant, dynamic_slice, dot_dnums); Array2D expected({{105.0}, {105.0}}); ComputeAndCompareR2(&builder, expected, {}, error_spec_); } -// TODO (b/69062148) Enable when Dot implements general contracting dimensions. XLA_TEST_F(DotOperationTest, - DISABLED_ON_CPU(DISABLED_ON_GPU(DISABLED_ON_INTERPRETER( - DotOfGatherOptimizationWithConstRHSReverseMM)))) { + + DotOfGatherOptimizationWithConstRHSReverseMM) { std::unique_ptr> constant_lhs_array( new Array2D({{1.0, 2.0, 3.0}, {4.0, 5.0, 6.0}, @@ -870,25 +872,21 @@ XLA_TEST_F(DotOperationTest, // Dot result to slice from: {{114, 96}, {105, 105}, {96, 114}} XlaBuilder builder(TestName()); - auto lhs_constant = builder.ConstantR2FromArray2D(*constant_lhs_array); - auto rhs_constant = builder.ConstantR2FromArray2D(*constant_rhs_array); - auto start_constant = builder.ConstantR1({0, 1}); - auto dynamic_slice = - builder.DynamicSlice(lhs_constant, start_constant, {6, 1}); + auto lhs_constant = ConstantR2FromArray2D(&builder, *constant_lhs_array); + auto rhs_constant = ConstantR2FromArray2D(&builder, *constant_rhs_array); + auto start_constant = ConstantR1(&builder, {0, 1}); + auto dynamic_slice = DynamicSlice(lhs_constant, start_constant, {6, 1}); DotDimensionNumbers dot_dnums; dot_dnums.add_lhs_contracting_dimensions(0); dot_dnums.add_rhs_contracting_dimensions(1); - auto result = builder.DotGeneral(dynamic_slice, rhs_constant, dot_dnums); + DotGeneral(dynamic_slice, rhs_constant, dot_dnums); Array2D expected({{105.0, 105.0}}); ComputeAndCompareR2(&builder, expected, {}, error_spec_); } -// TODO (b/69062148) Enable when Dot implements general contracting dimensions. -XLA_TEST_F(DotOperationTest, - DISABLED_ON_CPU(DISABLED_ON_GPU(DISABLED_ON_INTERPRETER( - DotOfGatherOptimizationWithConstLHSReverseMM)))) { +XLA_TEST_F(DotOperationTest, DotOfGatherOptimizationWithConstLHSReverseMM) { std::unique_ptr> constant_lhs_array( new Array2D({{1.0, 2.0, 3.0}, {4.0, 5.0, 6.0}, @@ -901,25 +899,21 @@ XLA_TEST_F(DotOperationTest, // Dot result to slice from: {{114, 96}, {105, 105}, {96, 114}} XlaBuilder builder(TestName()); - auto lhs_constant = builder.ConstantR2FromArray2D(*constant_lhs_array); - auto rhs_constant = builder.ConstantR2FromArray2D(*constant_rhs_array); - auto start_constant = builder.ConstantR1({1, 0}); - auto dynamic_slice = - builder.DynamicSlice(rhs_constant, start_constant, {1, 6}); + auto lhs_constant = ConstantR2FromArray2D(&builder, *constant_lhs_array); + auto rhs_constant = ConstantR2FromArray2D(&builder, *constant_rhs_array); + auto start_constant = ConstantR1(&builder, {1, 0}); + auto dynamic_slice = DynamicSlice(rhs_constant, start_constant, {1, 6}); DotDimensionNumbers dot_dnums; dot_dnums.add_lhs_contracting_dimensions(0); dot_dnums.add_rhs_contracting_dimensions(1); - auto result = builder.DotGeneral(lhs_constant, dynamic_slice, dot_dnums); + DotGeneral(lhs_constant, dynamic_slice, dot_dnums); Array2D expected({{96.0}, {105.0}, {114.0}}); ComputeAndCompareR2(&builder, expected, {}, error_spec_); } -// TODO (b/69062148) Enable when Dot implements general contracting dimensions. -XLA_TEST_F(DotOperationTest, - DISABLED_ON_CPU(DISABLED_ON_GPU( - DISABLED_ON_INTERPRETER(DotOfGatherOptimizationWithConstRHSRows)))) { +XLA_TEST_F(DotOperationTest, DotOfGatherOptimizationWithConstRHSRows) { std::unique_ptr> constant_lhs_array( new Array2D({{1.0, 2.0}, {3.0, 4.0}, @@ -937,25 +931,21 @@ XLA_TEST_F(DotOperationTest, // Dot result to slice from: {{132, 129, 126}, {126, 129, 132}} XlaBuilder builder(TestName()); - auto lhs_constant = builder.ConstantR2FromArray2D(*constant_lhs_array); - auto rhs_constant = builder.ConstantR2FromArray2D(*constant_rhs_array); - auto start_constant = builder.ConstantR1({0, 1}); - auto dynamic_slice = - builder.DynamicSlice(lhs_constant, start_constant, {6, 1}); + auto lhs_constant = ConstantR2FromArray2D(&builder, *constant_lhs_array); + auto rhs_constant = ConstantR2FromArray2D(&builder, *constant_rhs_array); + auto start_constant = ConstantR1(&builder, {0, 1}); + auto dynamic_slice = DynamicSlice(lhs_constant, start_constant, {6, 1}); DotDimensionNumbers dot_dnums; dot_dnums.add_lhs_contracting_dimensions(0); dot_dnums.add_rhs_contracting_dimensions(0); - auto result = builder.DotGeneral(dynamic_slice, rhs_constant, dot_dnums); + DotGeneral(dynamic_slice, rhs_constant, dot_dnums); Array2D expected({{126.0, 129.0, 132.0}}); ComputeAndCompareR2(&builder, expected, {}, error_spec_); } -// TODO (b/69062148) Enable when Dot implements general contracting dimensions. -XLA_TEST_F(DotOperationTest, - DISABLED_ON_CPU(DISABLED_ON_GPU( - DISABLED_ON_INTERPRETER(DotOfGatherOptimizationWithConstLHSRows)))) { +XLA_TEST_F(DotOperationTest, DotOfGatherOptimizationWithConstLHSRows) { std::unique_ptr> constant_lhs_array( new Array2D({{1.0, 2.0}, {3.0, 4.0}, @@ -973,25 +963,21 @@ XLA_TEST_F(DotOperationTest, // Dot result to slice from: {{132, 129, 126}, {126, 129, 132}} XlaBuilder builder(TestName()); - auto lhs_constant = builder.ConstantR2FromArray2D(*constant_lhs_array); - auto rhs_constant = builder.ConstantR2FromArray2D(*constant_rhs_array); - auto start_constant = builder.ConstantR1({0, 1}); - auto dynamic_slice = - builder.DynamicSlice(rhs_constant, start_constant, {6, 1}); + auto lhs_constant = ConstantR2FromArray2D(&builder, *constant_lhs_array); + auto rhs_constant = ConstantR2FromArray2D(&builder, *constant_rhs_array); + auto start_constant = ConstantR1(&builder, {0, 1}); + auto dynamic_slice = DynamicSlice(rhs_constant, start_constant, {6, 1}); DotDimensionNumbers dot_dnums; dot_dnums.add_lhs_contracting_dimensions(0); dot_dnums.add_rhs_contracting_dimensions(0); - auto result = builder.DotGeneral(lhs_constant, dynamic_slice, dot_dnums); + DotGeneral(lhs_constant, dynamic_slice, dot_dnums); Array2D expected({{129.0}, {129.0}}); ComputeAndCompareR2(&builder, expected, {}, error_spec_); } -// TODO (b/69062148) Enable when Dot implements general contracting dimensions. -XLA_TEST_F(DotOperationTest, - DISABLED_ON_CPU(DISABLED_ON_GPU( - DISABLED_ON_INTERPRETER(DotOfGatherOptimizationWithConstRHSCols)))) { +XLA_TEST_F(DotOperationTest, DotOfGatherOptimizationWithConstRHSCols) { std::unique_ptr> constant_lhs_array(new Array2D( {{1.0, 2.0, 3.0, 4.0, 5.0, 6.0}, {6.0, 5.0, 4.0, 3.0, 2.0, 1.0}})); std::unique_ptr> constant_rhs_array( @@ -1001,25 +987,21 @@ XLA_TEST_F(DotOperationTest, // Dot result to slice from: {{91, 168, 56}, {56, 168, 91}} XlaBuilder builder(TestName()); - auto lhs_constant = builder.ConstantR2FromArray2D(*constant_lhs_array); - auto rhs_constant = builder.ConstantR2FromArray2D(*constant_rhs_array); - auto start_constant = builder.ConstantR1({1, 0}); - auto dynamic_slice = - builder.DynamicSlice(lhs_constant, start_constant, {1, 6}); + auto lhs_constant = ConstantR2FromArray2D(&builder, *constant_lhs_array); + auto rhs_constant = ConstantR2FromArray2D(&builder, *constant_rhs_array); + auto start_constant = ConstantR1(&builder, {1, 0}); + auto dynamic_slice = DynamicSlice(lhs_constant, start_constant, {1, 6}); DotDimensionNumbers dot_dnums; dot_dnums.add_lhs_contracting_dimensions(1); dot_dnums.add_rhs_contracting_dimensions(1); - auto result = builder.DotGeneral(dynamic_slice, rhs_constant, dot_dnums); + DotGeneral(dynamic_slice, rhs_constant, dot_dnums); Array2D expected({{56.0, 168.0, 91.0}}); ComputeAndCompareR2(&builder, expected, {}, error_spec_); } -// TODO (b/69062148) Enable when Dot implements general contracting dimensions. -XLA_TEST_F(DotOperationTest, - DISABLED_ON_CPU(DISABLED_ON_GPU( - DISABLED_ON_INTERPRETER(DotOfGatherOptimizationWithConstLHSCols)))) { +XLA_TEST_F(DotOperationTest, DotOfGatherOptimizationWithConstLHSCols) { std::unique_ptr> constant_lhs_array(new Array2D( {{1.0, 2.0, 3.0, 4.0, 5.0, 6.0}, {6.0, 5.0, 4.0, 3.0, 2.0, 1.0}})); std::unique_ptr> constant_rhs_array( @@ -1029,19 +1011,41 @@ XLA_TEST_F(DotOperationTest, // Dot result to slice from: {{91, 168, 56}, {56, 168, 91}} XlaBuilder builder(TestName()); - auto lhs_constant = builder.ConstantR2FromArray2D(*constant_lhs_array); - auto rhs_constant = builder.ConstantR2FromArray2D(*constant_rhs_array); - auto start_constant = builder.ConstantR1({1, 0}); - auto dynamic_slice = - builder.DynamicSlice(rhs_constant, start_constant, {1, 6}); + auto lhs_constant = ConstantR2FromArray2D(&builder, *constant_lhs_array); + auto rhs_constant = ConstantR2FromArray2D(&builder, *constant_rhs_array); + auto start_constant = ConstantR1(&builder, {1, 0}); + auto dynamic_slice = DynamicSlice(rhs_constant, start_constant, {1, 6}); DotDimensionNumbers dot_dnums; dot_dnums.add_lhs_contracting_dimensions(1); dot_dnums.add_rhs_contracting_dimensions(1); - auto result = builder.DotGeneral(lhs_constant, dynamic_slice, dot_dnums); + DotGeneral(lhs_constant, dynamic_slice, dot_dnums); Array2D expected({{168.0}, {168.0}}); ComputeAndCompareR2(&builder, expected, {}, error_spec_); } + +XLA_TEST_F(DotOperationTest, DotRank2AndRank2NonDefaultContractionDims) { + XlaBuilder builder(TestName()); + + Array2D lhs_array({{1.0f, 2.0f}, {3.0f, 4.0f}}); + auto lhs_constant = ConstantR2FromArray2D(&builder, lhs_array); + + Array2D rhs_array({{5.0f, 6.0f}, {7.0f, 8.0f}}); + auto rhs_constant = ConstantR2FromArray2D(&builder, rhs_array); + + Shape shape = ShapeUtil::MakeShape(F32, {2, 2}); + DotDimensionNumbers dot_dnums; + dot_dnums.add_lhs_contracting_dimensions(0); + dot_dnums.add_rhs_contracting_dimensions(0); + DotGeneral(lhs_constant, rhs_constant, dot_dnums); + + Array2D expected({ + {26.f, 30.f}, + {38.f, 44.f}, + }); + + ComputeAndCompareR2(&builder, expected, {}, error_spec_); +} } // namespace } // namespace xla diff --git a/tensorflow/compiler/xla/tests/dynamic_ops_test.cc b/tensorflow/compiler/xla/tests/dynamic_ops_test.cc index a918c91f07ff0241845df4ef99334020859d8311..88ac96d6b0f9206ef1ed0e4135495d7903ebf3f4 100644 --- a/tensorflow/compiler/xla/tests/dynamic_ops_test.cc +++ b/tensorflow/compiler/xla/tests/dynamic_ops_test.cc @@ -124,11 +124,11 @@ class DynamicSliceTest : public ClientLibraryTestBase { // vector is special so that it cannot be an ArraySlice, which // is what the code below wants. So instead we do this. Literal input_values = - std::move(*Literal::CreateR1(input_values_int) + std::move(*LiteralUtil::CreateR1(input_values_int) ->Convert(primitive_util::NativeToPrimitiveType()) .ValueOrDie()); Literal expected_values = - std::move(*Literal::CreateR1(expected_values_int) + std::move(*LiteralUtil::CreateR1(expected_values_int) ->Convert(primitive_util::NativeToPrimitiveType()) .ValueOrDie()); @@ -138,8 +138,8 @@ class DynamicSliceTest : public ClientLibraryTestBase { std::unique_ptr start_data = CreateR1Parameter( slice_starts, 0, "slice_starts", &builder, &starts); // Build dynamic slice computation. - auto input = builder.ConstantLiteral(input_values); - builder.DynamicSlice(input, starts, slice_sizes); + auto input = ConstantLiteral(&builder, input_values); + DynamicSlice(input, starts, slice_sizes); // Run computation and compare against expected values. ComputeAndCompareLiteral(&builder, expected_values, {start_data.get()}); } @@ -150,11 +150,11 @@ class DynamicSliceTest : public ClientLibraryTestBase { const std::vector& slice_sizes, const Array2D& expected_values_int) { Literal input_values = - std::move(*Literal::CreateR2FromArray2D(input_values_int) + std::move(*LiteralUtil::CreateR2FromArray2D(input_values_int) ->Convert(primitive_util::NativeToPrimitiveType()) .ValueOrDie()); Literal expected_values = - std::move(*Literal::CreateR2FromArray2D(expected_values_int) + std::move(*LiteralUtil::CreateR2FromArray2D(expected_values_int) ->Convert(primitive_util::NativeToPrimitiveType()) .ValueOrDie()); @@ -164,8 +164,8 @@ class DynamicSliceTest : public ClientLibraryTestBase { std::unique_ptr start_data = CreateR1Parameter( slice_starts, 0, "slice_starts", &builder, &starts); // Build dynamic slice computation. - auto input = builder.ConstantLiteral(input_values); - builder.DynamicSlice(input, starts, slice_sizes); + auto input = ConstantLiteral(&builder, input_values); + DynamicSlice(input, starts, slice_sizes); // Run computation and compare against expected values. ComputeAndCompareLiteral(&builder, expected_values, {start_data.get()}); } @@ -176,11 +176,11 @@ class DynamicSliceTest : public ClientLibraryTestBase { const std::vector& slice_sizes, const Array3D& expected_values_int) { Literal input_values = - std::move(*Literal::CreateR3FromArray3D(input_values_int) + std::move(*LiteralUtil::CreateR3FromArray3D(input_values_int) ->Convert(primitive_util::NativeToPrimitiveType()) .ValueOrDie()); Literal expected_values = - std::move(*Literal::CreateR3FromArray3D(expected_values_int) + std::move(*LiteralUtil::CreateR3FromArray3D(expected_values_int) ->Convert(primitive_util::NativeToPrimitiveType()) .ValueOrDie()); @@ -190,8 +190,8 @@ class DynamicSliceTest : public ClientLibraryTestBase { std::unique_ptr start_data = CreateR1Parameter( slice_starts, 0, "slice_starts", &builder, &starts); // Build dynamic slice computation. - auto input = builder.ConstantLiteral(input_values); - builder.DynamicSlice(input, starts, slice_sizes); + auto input = ConstantLiteral(&builder, input_values); + DynamicSlice(input, starts, slice_sizes); // Run computation and compare against expected values. ComputeAndCompareLiteral(&builder, expected_values, {start_data.get()}); } @@ -202,18 +202,28 @@ XLA_TEST_F(DynamicSliceTest, Int32R1) { TestR1(); } XLA_TEST_F(DynamicSliceTest, Int32R1OOB) { TestR1OOB(); } XLA_TEST_F(DynamicSliceTest, Int64R1) { TestR1(); } XLA_TEST_F(DynamicSliceTest, UInt64R1) { TestR1(); } +XLA_TEST_F(DynamicSliceTest, UInt32R1OOB) { + RunR1({0, 1, 2, 3, 4}, {2147483648u}, {2}, {3, 4}); +} XLA_TEST_F(DynamicSliceTest, Int32R2BF16) { TestR2(); } XLA_TEST_F(DynamicSliceTest, Int32R2) { TestR2(); } XLA_TEST_F(DynamicSliceTest, Int32R2OOB) { TestR2OOB(); } XLA_TEST_F(DynamicSliceTest, Int64R2) { TestR2(); } XLA_TEST_F(DynamicSliceTest, UInt64R2) { TestR2(); } +XLA_TEST_F(DynamicSliceTest, UInt32R2OOB) { + RunR2({{0, 1}, {2, 3}}, {2147483648u, 0}, {1, 1}, {{2}}); +} XLA_TEST_F(DynamicSliceTest, Int32R3BF16) { TestR3(); } XLA_TEST_F(DynamicSliceTest, Int32R3) { TestR3(); } XLA_TEST_F(DynamicSliceTest, Int32R3OOB) { TestR3OOB(); } XLA_TEST_F(DynamicSliceTest, Int64R3) { TestR3(); } XLA_TEST_F(DynamicSliceTest, UInt64R3) { TestR3(); } +XLA_TEST_F(DynamicSliceTest, UInt32R3OOB) { + RunR3({{{0, 1}, {2, 3}}, {{4, 5}, {6, 7}}}, + {2147483648u, 0, 2147483648u}, {1, 1, 1}, {{{5}}}); +} XLA_TEST_F(DynamicSliceTest, Int32R1Pred) { // Slice at dimension start. @@ -349,15 +359,15 @@ class DynamicUpdateSliceTest : public ClientLibraryTestBase { void RunR0(int input_value_int, int update_value_int, const std::vector slice_starts, int expected_value_int) { Literal input_value = - std::move(*Literal::CreateR0(input_value_int) + std::move(*LiteralUtil::CreateR0(input_value_int) ->Convert(primitive_util::NativeToPrimitiveType()) .ValueOrDie()); Literal update_value = - std::move(*Literal::CreateR0(update_value_int) + std::move(*LiteralUtil::CreateR0(update_value_int) ->Convert(primitive_util::NativeToPrimitiveType()) .ValueOrDie()); Literal expected_value = - std::move(*Literal::CreateR0(expected_value_int) + std::move(*LiteralUtil::CreateR0(expected_value_int) ->Convert(primitive_util::NativeToPrimitiveType()) .ValueOrDie()); @@ -367,9 +377,9 @@ class DynamicUpdateSliceTest : public ClientLibraryTestBase { std::unique_ptr start_data = CreateR1Parameter( slice_starts, 0, "slice_starts", &builder, &starts); // Build dynamic slice computation. - auto input = builder.ConstantLiteral(input_value); - auto update = builder.ConstantLiteral(update_value); - builder.DynamicUpdateSlice(input, update, starts); + auto input = ConstantLiteral(&builder, input_value); + auto update = ConstantLiteral(&builder, update_value); + DynamicUpdateSlice(input, update, starts); // Run computation and compare against expected values. ComputeAndCompareLiteral(&builder, expected_value, {start_data.get()}); } @@ -380,15 +390,15 @@ class DynamicUpdateSliceTest : public ClientLibraryTestBase { const std::vector slice_starts, tensorflow::gtl::ArraySlice expected_values_int) { Literal input_values = - std::move(*Literal::CreateR1(input_values_int) + std::move(*LiteralUtil::CreateR1(input_values_int) ->Convert(primitive_util::NativeToPrimitiveType()) .ValueOrDie()); Literal update_values = - std::move(*Literal::CreateR1(update_values_int) + std::move(*LiteralUtil::CreateR1(update_values_int) ->Convert(primitive_util::NativeToPrimitiveType()) .ValueOrDie()); Literal expected_values = - std::move(*Literal::CreateR1(expected_values_int) + std::move(*LiteralUtil::CreateR1(expected_values_int) ->Convert(primitive_util::NativeToPrimitiveType()) .ValueOrDie()); @@ -398,9 +408,9 @@ class DynamicUpdateSliceTest : public ClientLibraryTestBase { std::unique_ptr start_data = CreateR1Parameter( slice_starts, 0, "slice_starts", &builder, &starts); // Build dynamic slice computation. - auto input = builder.ConstantLiteral(input_values); - auto update = builder.ConstantLiteral(update_values); - builder.DynamicUpdateSlice(input, update, starts); + auto input = ConstantLiteral(&builder, input_values); + auto update = ConstantLiteral(&builder, update_values); + DynamicUpdateSlice(input, update, starts); // Run computation and compare against expected values. ComputeAndCompareLiteral(&builder, expected_values, {start_data.get()}); } @@ -411,15 +421,15 @@ class DynamicUpdateSliceTest : public ClientLibraryTestBase { const std::vector slice_starts, const Array2D& expected_values_int) { Literal input_values = - std::move(*Literal::CreateR2FromArray2D(input_values_int) + std::move(*LiteralUtil::CreateR2FromArray2D(input_values_int) ->Convert(primitive_util::NativeToPrimitiveType()) .ValueOrDie()); Literal update_values = - std::move(*Literal::CreateR2FromArray2D(update_values_int) + std::move(*LiteralUtil::CreateR2FromArray2D(update_values_int) ->Convert(primitive_util::NativeToPrimitiveType()) .ValueOrDie()); Literal expected_values = - std::move(*Literal::CreateR2FromArray2D(expected_values_int) + std::move(*LiteralUtil::CreateR2FromArray2D(expected_values_int) ->Convert(primitive_util::NativeToPrimitiveType()) .ValueOrDie()); @@ -429,9 +439,9 @@ class DynamicUpdateSliceTest : public ClientLibraryTestBase { std::unique_ptr start_data = CreateR1Parameter( slice_starts, 0, "slice_starts", &builder, &starts); // Build dynamic slice computation. - auto input = builder.ConstantLiteral(input_values); - auto update = builder.ConstantLiteral(update_values); - builder.DynamicUpdateSlice(input, update, starts); + auto input = ConstantLiteral(&builder, input_values); + auto update = ConstantLiteral(&builder, update_values); + DynamicUpdateSlice(input, update, starts); // Run computation and compare against expected values. ComputeAndCompareLiteral(&builder, expected_values, {start_data.get()}); } @@ -442,15 +452,15 @@ class DynamicUpdateSliceTest : public ClientLibraryTestBase { const std::vector slice_starts, const Array3D& expected_values_int) { Literal input_values = - std::move(*Literal::CreateR3FromArray3D(input_values_int) + std::move(*LiteralUtil::CreateR3FromArray3D(input_values_int) ->Convert(primitive_util::NativeToPrimitiveType()) .ValueOrDie()); Literal update_values = - std::move(*Literal::CreateR3FromArray3D(update_values_int) + std::move(*LiteralUtil::CreateR3FromArray3D(update_values_int) ->Convert(primitive_util::NativeToPrimitiveType()) .ValueOrDie()); Literal expected_values = - std::move(*Literal::CreateR3FromArray3D(expected_values_int) + std::move(*LiteralUtil::CreateR3FromArray3D(expected_values_int) ->Convert(primitive_util::NativeToPrimitiveType()) .ValueOrDie()); @@ -460,9 +470,9 @@ class DynamicUpdateSliceTest : public ClientLibraryTestBase { std::unique_ptr start_data = CreateR1Parameter( slice_starts, 0, "slice_starts", &builder, &starts); // Build dynamic slice computation. - auto input = builder.ConstantLiteral(input_values); - auto update = builder.ConstantLiteral(update_values); - builder.DynamicUpdateSlice(input, update, starts); + auto input = ConstantLiteral(&builder, input_values); + auto update = ConstantLiteral(&builder, update_values); + DynamicUpdateSlice(input, update, starts); // Run computation and compare against expected values. ComputeAndCompareLiteral(&builder, expected_values, {start_data.get()}); } @@ -508,8 +518,8 @@ class DynamicUpdateSliceTest : public ClientLibraryTestBase { XlaOp update; std::unique_ptr update_data = CreateR3Parameter( update_values, 1, "update_values", &builder, &update); - auto starts = builder.ConstantR1({index, 0, 0}); - builder.DynamicUpdateSlice(input, update, starts); + auto starts = ConstantR1(&builder, {index, 0, 0}); + DynamicUpdateSlice(input, update, starts); // Run computation and compare against expected values. ComputeAndCompareR3(&builder, expected_values, @@ -520,7 +530,7 @@ class DynamicUpdateSliceTest : public ClientLibraryTestBase { template void DumpArray(const string& name, const Array3D values) { std::unique_ptr literal = - Literal::CreateR3FromArray3D(values); + LiteralUtil::CreateR3FromArray3D(values); LOG(INFO) << name << ":" << literal->ToString(); } }; @@ -530,21 +540,32 @@ XLA_TEST_F(DynamicUpdateSliceTest, Int32R0) { TestR0(); } XLA_TEST_F(DynamicUpdateSliceTest, Int64R0) { TestR0(); } XLA_TEST_F(DynamicUpdateSliceTest, UInt64R0) { TestR0(); } -// TODO(b/71820067): The CPU parallel backend failed for this on 2018-01-10. XLA_TEST_F(DynamicUpdateSliceTest, Int32R1BF16) { TestR1(); } XLA_TEST_F(DynamicUpdateSliceTest, Int32R1) { TestR1(); } XLA_TEST_F(DynamicUpdateSliceTest, Int64R1) { TestR1(); } XLA_TEST_F(DynamicUpdateSliceTest, UInt64R1) { TestR1(); } +XLA_TEST_F(DynamicUpdateSliceTest, UInt32R1OOB) { + RunR1({0, 1, 2, 3, 4}, {5, 6}, {2147483648u}, {0, 1, 2, 5, 6}); +} XLA_TEST_F(DynamicUpdateSliceTest, Int32R2BF16) { TestR2(); } XLA_TEST_F(DynamicUpdateSliceTest, Int32R2) { TestR2(); } XLA_TEST_F(DynamicUpdateSliceTest, Int64R2) { TestR2(); } XLA_TEST_F(DynamicUpdateSliceTest, UInt64R2) { TestR2(); } +XLA_TEST_F(DynamicUpdateSliceTest, UInt32R2OOB) { + RunR2({{0, 1}, {2, 3}}, {{4}}, {2147483648u, 0}, + {{0, 1}, {4, 3}}); +} XLA_TEST_F(DynamicUpdateSliceTest, Int32R3BF16) { TestR3(); } XLA_TEST_F(DynamicUpdateSliceTest, Int32R3) { TestR3(); } XLA_TEST_F(DynamicUpdateSliceTest, Int64R3) { TestR3(); } XLA_TEST_F(DynamicUpdateSliceTest, UInt64R3) { TestR3(); } +XLA_TEST_F(DynamicUpdateSliceTest, UInt32R3OOB) { + RunR3({{{0, 1}, {2, 3}}, {{4, 5}, {6, 7}}}, {{{8}}}, + {2147483648u, 0, 2147483648u}, + {{{0, 1}, {2, 3}}, {{4, 8}, {6, 7}}}); +} XLA_TEST_F(DynamicUpdateSliceTest, Int32OOBBF16) { TestOOB(); } XLA_TEST_F(DynamicUpdateSliceTest, Int32OOB) { TestOOB(); } @@ -695,17 +716,17 @@ void BM_DynamicSlice(int num_iters) { XlaBuilder builder("DynamicSlice"); // Create input as a constant: shape [1, 2, 3, 4] - auto input_literal = Literal::CreateR4( + auto input_literal = LiteralUtil::CreateR4( {{{{1, 2, 3, 4}, {5, 6, 7, 8}, {9, 10, 11, 12}}, {{13, 14, 15, 16}, {17, 18, 19, 20}, {21, 22, 23, 24}}}}); - auto input = builder.ConstantLiteral(*input_literal); + auto input = ConstantLiteral(&builder, *input_literal); // Create dynamic slice start indices as a parameter: shape [4] auto start_indices_shape = ShapeUtil::MakeShape(S32, {4}); auto start_indices = - builder.Parameter(0, start_indices_shape, "start_indices"); + Parameter(&builder, 0, start_indices_shape, "start_indices"); // Add DynamicSlice op to the computatation. - builder.DynamicSlice(input, start_indices, {1, 1, 1, 1}); + DynamicSlice(input, start_indices, {1, 1, 1, 1}); auto computation = builder.Build().ConsumeValueOrDie(); // Initialize and transfer parameter buffer. @@ -715,7 +736,7 @@ void BM_DynamicSlice(int num_iters) { start_indices_shape, &allocator, /*device_ordinal=*/0) .ConsumeValueOrDie(); - auto start_indices_literal = Literal::CreateR1({0, 1, 2, 3}); + auto start_indices_literal = LiteralUtil::CreateR1({0, 1, 2, 3}); auto stream = client->mutable_backend()->BorrowStream(device_ordinal).ValueOrDie(); ASSERT_IS_OK(transfer_manager->TransferLiteralToDevice( diff --git a/tensorflow/compiler/xla/tests/execution_profile_test.cc b/tensorflow/compiler/xla/tests/execution_profile_test.cc index a6ba6db5d3bf86de91f6fda022c46afee01281c2..ebba13c5b39e241ff01c9ddcba8a1c04180b4bd0 100644 --- a/tensorflow/compiler/xla/tests/execution_profile_test.cc +++ b/tensorflow/compiler/xla/tests/execution_profile_test.cc @@ -31,10 +31,10 @@ XLA_TEST_F(ExecutionProfileTest, ExecuteWithExecutionProfile) { TF_ASSERT_OK_AND_ASSIGN( std::unique_ptr input, client_->TransferToServer( - *Literal::CreateR2F32Linspace(1e0, 1e5, 256, 256))); + *LiteralUtil::CreateR2F32Linspace(1e0, 1e5, 256, 256))); XlaBuilder b(TestName() + ".add"); - b.Dot(b.Parameter(0, shape, "param_0"), b.Parameter(1, shape, "param_1")); + Dot(Parameter(&b, 0, shape, "param_0"), Parameter(&b, 1, shape, "param_1")); TF_ASSERT_OK_AND_ASSIGN(XlaComputation dot_product, b.Build()); ExecutionProfile execution_profile; diff --git a/tensorflow/compiler/xla/tests/exhaustive_f32_elementwise_op_test.cc b/tensorflow/compiler/xla/tests/exhaustive_f32_elementwise_op_test.cc index 0a37e4d423620122f2e109343a86a964f46d778f..86bfaea4ef43ad382e497fd281ec5439f001b56f 100644 --- a/tensorflow/compiler/xla/tests/exhaustive_f32_elementwise_op_test.cc +++ b/tensorflow/compiler/xla/tests/exhaustive_f32_elementwise_op_test.cc @@ -39,7 +39,7 @@ class ExhaustiveF32ElementwiseOpTest XlaBuilder builder(TestName()); std::unique_ptr input_literal = - Literal::CreateFromDimensions(F32, {input_size}); + LiteralUtil::CreateFromDimensions(F32, {input_size}); for (int64 i = begin; i < end; i++) { if (i >= known_incorrect_range.first && i < known_incorrect_range.second) { @@ -54,7 +54,7 @@ class ExhaustiveF32ElementwiseOpTest TF_ASSERT_OK_AND_ASSIGN(std::unique_ptr input_data, client_->TransferToServer(*input_literal)); - auto input = builder.Parameter(0, input_literal->shape(), "input"); + auto input = Parameter(&builder, 0, input_literal->shape(), "input"); enqueue_op(&builder, input); std::vector expected_result; @@ -79,8 +79,8 @@ XLA_TEST_P(ExhaustiveF32ElementwiseOpTest, LogF32) { #endif ExhaustivelyTestF32Op( - [](XlaBuilder* builder, const XlaOp& input) { builder->Log(input); }, - std::log, known_incorrect_range); + [](XlaBuilder* builder, const XlaOp& input) { Log(input); }, std::log, + known_incorrect_range); } XLA_TEST_P(ExhaustiveF32ElementwiseOpTest, ExpF32) { @@ -95,14 +95,14 @@ XLA_TEST_P(ExhaustiveF32ElementwiseOpTest, ExpF32) { #endif ExhaustivelyTestF32Op( - [](XlaBuilder* builder, const XlaOp& input) { builder->Exp(input); }, - std::exp, known_incorrect_range); + [](XlaBuilder* builder, const XlaOp& input) { Exp(input); }, std::exp, + known_incorrect_range); } XLA_TEST_P(ExhaustiveF32ElementwiseOpTest, TanhF32) { ExhaustivelyTestF32Op( - [](XlaBuilder* builder, const XlaOp& input) { builder->Tanh(input); }, - std::tanh, /*known_incorrect_range=*/{0, 0}); + [](XlaBuilder* builder, const XlaOp& input) { Tanh(input); }, std::tanh, + /*known_incorrect_range=*/{0, 0}); } std::vector> CreateExhaustiveParameters() { diff --git a/tensorflow/compiler/xla/tests/filecheck.cc b/tensorflow/compiler/xla/tests/filecheck.cc index 93d1c921c4a138cda55ed7338b8e3aa82518d114..dcb469087e0064d17ce3b04fdeaf0b6136069a55 100644 --- a/tensorflow/compiler/xla/tests/filecheck.cc +++ b/tensorflow/compiler/xla/tests/filecheck.cc @@ -76,6 +76,11 @@ StatusOr RunFileCheck(const string& input, const string& pattern) { XLA_LOG_LINES(tensorflow::WARNING, input); LOG(WARNING) << "FileCheck pattern was:"; XLA_LOG_LINES(tensorflow::WARNING, pattern); + } else if (!standard_error.empty()) { + LOG(INFO) << "FileCheck stderr:"; + XLA_LOG_LINES(tensorflow::INFO, standard_error); + LOG(INFO) << "FileCheck input was:"; + XLA_LOG_LINES(tensorflow::INFO, input); } return succeeded; } diff --git a/tensorflow/compiler/xla/tests/floor_ceil_test.cc b/tensorflow/compiler/xla/tests/floor_ceil_test.cc index 71eb914a8e5eaef2e38b9e6e7d45b8a10ce1bd7a..30dc639f117b9871238f0bf1628502cf8bef2e0c 100644 --- a/tensorflow/compiler/xla/tests/floor_ceil_test.cc +++ b/tensorflow/compiler/xla/tests/floor_ceil_test.cc @@ -42,12 +42,12 @@ class FloorCeilTest : public ClientLibraryTestBase { LOG(INFO) << "input: {" << tensorflow::str_util::Join(expected, ", ") << "}"; XlaBuilder builder(TestName()); - auto c = builder.ConstantR1(input); + auto c = ConstantR1(&builder, input); if (f == kCeil) { - builder.Ceil(c); + Ceil(c); } else { ASSERT_EQ(kFloor, f); - builder.Floor(c); + Floor(c); } ComputeAndCompareR1(&builder, expected, /*arguments=*/{}); } @@ -55,12 +55,12 @@ class FloorCeilTest : public ClientLibraryTestBase { void TestR0F32(float input, float expected, Function f) { LOG(INFO) << "input: " << expected; XlaBuilder builder(TestName()); - auto c = builder.ConstantR0(input); + auto c = ConstantR0(&builder, input); if (f == kCeil) { - builder.Ceil(c); + Ceil(c); } else { ASSERT_EQ(kFloor, f); - builder.Floor(c); + Floor(c); } ComputeAndCompareR0(&builder, expected, /*arguments=*/{}); } diff --git a/tensorflow/compiler/xla/tests/fmax_test.cc b/tensorflow/compiler/xla/tests/fmax_test.cc index 73f029b59bc56aa6c3e86200a49fcae0fd177101..0254ae1baaa864b38c3b217a5c2026d34b7f7d12 100644 --- a/tensorflow/compiler/xla/tests/fmax_test.cc +++ b/tensorflow/compiler/xla/tests/fmax_test.cc @@ -28,11 +28,11 @@ class FmaxSimpleTest : public ClientLibraryTestBase {}; TEST_F(FmaxSimpleTest, FmaxTenValues) { XlaBuilder builder(TestName()); - auto x = builder.ConstantR1( - {-0.0, 1.0, 2.0, -3.0, -4.0, 5.0, 6.0, -7.0, -8.0, 9.0}); - auto y = builder.ConstantR1( - {-0.0, -1.0, -2.0, 3.0, 4.0, -5.0, -6.0, 7.0, 8.0, -9.0}); - builder.Max(x, y); + auto x = ConstantR1( + &builder, {-0.0, 1.0, 2.0, -3.0, -4.0, 5.0, 6.0, -7.0, -8.0, 9.0}); + auto y = ConstantR1( + &builder, {-0.0, -1.0, -2.0, 3.0, 4.0, -5.0, -6.0, 7.0, 8.0, -9.0}); + Max(x, y); std::vector expected = {-0.0, 1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0}; diff --git a/tensorflow/compiler/xla/tests/fusion_test.cc b/tensorflow/compiler/xla/tests/fusion_test.cc index e6f79b5ac55dddfbb213a36cadbee53bc9443d9d..dc6447793575c757fdd68baea03f416a951e45fc 100644 --- a/tensorflow/compiler/xla/tests/fusion_test.cc +++ b/tensorflow/compiler/xla/tests/fusion_test.cc @@ -26,13 +26,14 @@ limitations under the License. #include "tensorflow/compiler/xla/array2d.h" #include "tensorflow/compiler/xla/client/client_library.h" #include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/primitive_util.h" #include "tensorflow/compiler/xla/ptr_util.h" #include "tensorflow/compiler/xla/service/hlo_computation.h" #include "tensorflow/compiler/xla/service/hlo_instruction.h" #include "tensorflow/compiler/xla/service/hlo_module.h" #include "tensorflow/compiler/xla/service/hlo_opcode.h" +#include "tensorflow/compiler/xla/service/hlo_parser.h" #include "tensorflow/compiler/xla/service/platform_util.h" #include "tensorflow/compiler/xla/shape_util.h" #include "tensorflow/compiler/xla/tests/client_library_test_base.h" @@ -89,7 +90,7 @@ class FusionTest : public HloTestBase { HloInstruction* hlos[4]; for (int i = 0; i < Arity; ++i) { hlos[i + 1] = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR2FromArray2D(operand_data[i]))); + LiteralUtil::CreateR2FromArray2D(operand_data[i]))); } auto answer_shape = ShapeUtil::MakeShape(prim_type, {test_width, test_height}); @@ -115,7 +116,7 @@ class FusionTest : public HloTestBase { ArraySlice(hlos, 0, Arity + 1), HloInstruction::FusionKind::kLoop); - auto expected = Literal::CreateR2FromArray2D(answer_data); + auto expected = LiteralUtil::CreateR2FromArray2D(answer_data); auto actual = ExecuteAndTransfer(std::move(hlo_module), {}); if (primitive_util::IsFloatingPointType(prim_type)) { EXPECT_TRUE(LiteralTestUtil::Near(*expected, *actual, ErrorSpec(1e-4))); @@ -186,27 +187,28 @@ XLA_TEST_F(FusionTest, Test) { auto builder = HloComputation::Builder(TestName()); auto hlo_module = CreateNewModule(); auto const0 = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR2({{1.0}, {2.0}, {3.0}}))); + LiteralUtil::CreateR2({{1.0}, {2.0}, {3.0}}))); auto const1 = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR2({{-1.0}, {-1.0}, {-1.0}}))); + LiteralUtil::CreateR2({{-1.0}, {-1.0}, {-1.0}}))); auto add2 = builder.AddInstruction(HloInstruction::CreateBinary( ShapeUtil::MakeShape(F32, {3, 1}), HloOpcode::kAdd, const0, const1)); auto reshape3 = builder.AddInstruction(HloInstruction::CreateTranspose( ShapeUtil::MakeShape(F32, {1, 3}), add2, {1, 0})); auto const4 = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR2({{1.62, 2.72, 3.14}}))); + LiteralUtil::CreateR2({{1.62, 2.72, 3.14}}))); auto concat5 = builder.AddInstruction(HloInstruction::CreateConcatenate( ShapeUtil::MakeShape(F32, {2, 3}), {reshape3, const4}, 0)); auto const6 = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR2({{1.0, 1.0, 1.0}, {0.0, 0.0, 0.0}}))); + LiteralUtil::CreateR2({{1.0, 1.0, 1.0}, {0.0, 0.0, 0.0}}))); auto negate7 = builder.AddInstruction(HloInstruction::CreateUnary( ShapeUtil::MakeShape(F32, {2, 3}), HloOpcode::kNegate, const6)); auto add8 = builder.AddInstruction(HloInstruction::CreateBinary( ShapeUtil::MakeShape(F32, {2, 3}), HloOpcode::kAdd, concat5, negate7)); auto const9 = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR2({{0.5, 0.5, 0.5}, {0.5, 0.5, 0.5}}))); - auto const10 = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR2({{true, false, true}, {false, true, false}}))); + LiteralUtil::CreateR2({{0.5, 0.5, 0.5}, {0.5, 0.5, 0.5}}))); + auto const10 = builder.AddInstruction( + HloInstruction::CreateConstant(LiteralUtil::CreateR2( + {{true, false, true}, {false, true, false}}))); auto select11 = builder.AddInstruction( HloInstruction::CreateTernary(ShapeUtil::MakeShape(F32, {2, 3}), HloOpcode::kSelect, const10, add8, const9)); @@ -222,7 +224,7 @@ XLA_TEST_F(FusionTest, Test) { HloInstruction::FusionKind::kLoop); EXPECT_TRUE(LiteralTestUtil::Near( - *Literal::CreateR2({{0.5}, {2.72}}), + *LiteralUtil::CreateR2({{0.5}, {2.72}}), *ExecuteAndTransfer(std::move(hlo_module), {}), ErrorSpec(1e-4))); } @@ -233,11 +235,11 @@ XLA_TEST_F(FusionTest, Parameter) { auto builder = HloComputation::Builder(TestName()); auto hlo_module = CreateNewModule(); auto const0 = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR2({{1.0, 2.0, 3.0}}))); + LiteralUtil::CreateR2({{1.0, 2.0, 3.0}}))); auto copy1 = builder.AddInstruction(HloInstruction::CreateUnary( ShapeUtil::MakeShape(F32, {1, 3}), HloOpcode::kCopy, const0)); auto const2 = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR2({{-2.0, -2.0, -2.0}}))); + LiteralUtil::CreateR2({{-2.0, -2.0, -2.0}}))); // add3 = copy1 + const2 = const0 + const2 = {1,2,3} + {-2,-2,-2} = {-1,0,+1} auto add3 = builder.AddInstruction(HloInstruction::CreateBinary( ShapeUtil::MakeShape(F32, {1, 3}), HloOpcode::kAdd, copy1, const2)); @@ -248,7 +250,7 @@ XLA_TEST_F(FusionTest, Parameter) { HloInstruction::FusionKind::kLoop); EXPECT_TRUE(LiteralTestUtil::Near( - *Literal::CreateR2({{-1.0, 0.0, 1.0}}), + *LiteralUtil::CreateR2({{-1.0, 0.0, 1.0}}), *ExecuteAndTransfer(std::move(hlo_module), {}), ErrorSpec(1e-4))); } @@ -269,7 +271,7 @@ XLA_TEST_F(FusionTest, RandomizedParallelPartition) { auto hlo_module = CreateNewModule(); auto two = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(2.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(2.0))); auto x = builder.AddInstruction(HloInstruction::CreateBroadcast(shape, two, {})); auto y = builder.AddInstruction( @@ -292,9 +294,9 @@ XLA_TEST_F(FusionTest, BroadcastIntoBinaryOp) { auto builder = HloComputation::Builder(TestName()); auto hlo_module = CreateNewModule(); auto const_vector = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR1({1.0, 2.0, 3.0}))); + LiteralUtil::CreateR1({1.0, 2.0, 3.0}))); auto const_array = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR2({{-1.0, -2.0, -4.0}, {10.0, 20.0, 30.0}}))); + LiteralUtil::CreateR2({{-1.0, -2.0, -4.0}, {10.0, 20.0, 30.0}}))); auto broadcast = builder.AddInstruction( HloInstruction::CreateBroadcast(const_array->shape(), const_vector, {1})); // add2 = broadcast(const_vector) + const_array @@ -308,7 +310,7 @@ XLA_TEST_F(FusionTest, BroadcastIntoBinaryOp) { HloInstruction::FusionKind::kLoop); EXPECT_TRUE(LiteralTestUtil::Near( - *Literal::CreateR2({{0.0, 0.0, -1.0}, {11.0, 22.0, 33.0}}), + *LiteralUtil::CreateR2({{0.0, 0.0, -1.0}, {11.0, 22.0, 33.0}}), *ExecuteAndTransfer(std::move(hlo_module), {}), ErrorSpec(1e-4))); } @@ -316,14 +318,14 @@ XLA_TEST_F(FusionTest, ReshapeToScalar) { auto builder = HloComputation::Builder(TestName()); auto hlo_module = CreateNewModule(); auto single_element_array = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR2({{5}}))); + HloInstruction::CreateConstant(LiteralUtil::CreateR2({{5}}))); auto reshape = builder.AddInstruction(HloInstruction::CreateReshape( ShapeUtil::MakeShape(S32, {}), single_element_array)); hlo_module->AddEntryComputation(builder.Build()) ->CreateFusionInstruction(/*instructions_to_fuse=*/{reshape}, HloInstruction::FusionKind::kLoop); EXPECT_TRUE( - LiteralTestUtil::Equal(*Literal::CreateR0(5), + LiteralTestUtil::Equal(*LiteralUtil::CreateR0(5), *ExecuteAndTransfer(std::move(hlo_module), {}))); } @@ -331,14 +333,14 @@ XLA_TEST_F(FusionTest, Reshape_3by2_1by2by3) { auto builder = HloComputation::Builder(TestName()); auto hlo_module = CreateNewModule(); auto const0 = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR2({{1, 2}, {3, 4}, {5, 6}}))); + LiteralUtil::CreateR2({{1, 2}, {3, 4}, {5, 6}}))); auto reshape1 = builder.AddInstruction(HloInstruction::CreateReshape( ShapeUtil::MakeShape(S32, {1, 2, 3}), const0)); hlo_module->AddEntryComputation(builder.Build()) ->CreateFusionInstruction(/*instructions_to_fuse=*/{reshape1}, HloInstruction::FusionKind::kLoop); EXPECT_TRUE(LiteralTestUtil::Equal( - *Literal::CreateR3({{{1, 2, 3}, {4, 5, 6}}}), + *LiteralUtil::CreateR3({{{1, 2, 3}, {4, 5, 6}}}), *ExecuteAndTransfer(std::move(hlo_module), {}))); } @@ -346,14 +348,14 @@ XLA_TEST_F(FusionTest, Reshape_1by2by3_3by2) { auto builder = HloComputation::Builder(TestName()); auto hlo_module = CreateNewModule(); auto const0 = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR3({{{1, 2, 3}, {4, 5, 6}}}))); + LiteralUtil::CreateR3({{{1, 2, 3}, {4, 5, 6}}}))); auto reshape1 = builder.AddInstruction( HloInstruction::CreateReshape(ShapeUtil::MakeShape(S32, {3, 2}), const0)); hlo_module->AddEntryComputation(builder.Build()) ->CreateFusionInstruction(/*instructions_to_fuse=*/{reshape1}, HloInstruction::FusionKind::kLoop); EXPECT_TRUE(LiteralTestUtil::Equal( - *Literal::CreateR2({{1, 2}, {3, 4}, {5, 6}}), + *LiteralUtil::CreateR2({{1, 2}, {3, 4}, {5, 6}}), *ExecuteAndTransfer(std::move(hlo_module), {}))); } @@ -361,14 +363,14 @@ XLA_TEST_F(FusionTest, Reshape_1by1by1_) { auto builder = HloComputation::Builder(TestName()); auto hlo_module = CreateNewModule(); auto const0 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR3({{{7}}}))); + HloInstruction::CreateConstant(LiteralUtil::CreateR3({{{7}}}))); auto reshape1 = builder.AddInstruction( HloInstruction::CreateReshape(ShapeUtil::MakeShape(S32, {}), const0)); hlo_module->AddEntryComputation(builder.Build()) ->CreateFusionInstruction(/*instructions_to_fuse=*/{reshape1}, HloInstruction::FusionKind::kLoop); EXPECT_TRUE( - LiteralTestUtil::Equal(*Literal::CreateR0(7), + LiteralTestUtil::Equal(*LiteralUtil::CreateR0(7), *ExecuteAndTransfer(std::move(hlo_module), {}))); } @@ -376,14 +378,14 @@ XLA_TEST_F(FusionTest, Reshape__1by1by1) { auto builder = HloComputation::Builder(TestName()); auto hlo_module = CreateNewModule(); auto const0 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(7))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(7))); auto reshape1 = builder.AddInstruction(HloInstruction::CreateReshape( ShapeUtil::MakeShape(S32, {1, 1, 1}), const0)); hlo_module->AddEntryComputation(builder.Build()) ->CreateFusionInstruction(/*instructions_to_fuse=*/{reshape1}, HloInstruction::FusionKind::kLoop); EXPECT_TRUE( - LiteralTestUtil::Equal(*Literal::CreateR3({{{7}}}), + LiteralTestUtil::Equal(*LiteralUtil::CreateR3({{{7}}}), *ExecuteAndTransfer(std::move(hlo_module), {}))); } @@ -391,14 +393,14 @@ XLA_TEST_F(FusionTest, Reshape__) { auto builder = HloComputation::Builder(TestName()); auto hlo_module = CreateNewModule(); auto const0 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(7))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(7))); auto reshape1 = builder.AddInstruction( HloInstruction::CreateReshape(ShapeUtil::MakeShape(S32, {}), const0)); hlo_module->AddEntryComputation(builder.Build()) ->CreateFusionInstruction(/*instructions_to_fuse=*/{reshape1}, HloInstruction::FusionKind::kLoop); EXPECT_TRUE( - LiteralTestUtil::Equal(*Literal::CreateR0(7), + LiteralTestUtil::Equal(*LiteralUtil::CreateR0(7), *ExecuteAndTransfer(std::move(hlo_module), {}))); } @@ -406,14 +408,14 @@ XLA_TEST_F(FusionTest, Reshape_3by3_3by3) { auto builder = HloComputation::Builder(TestName()); auto hlo_module = CreateNewModule(); auto const0 = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR2({{1, 2, 3}, {4, 5, 6}, {7, 8, 9}}))); + LiteralUtil::CreateR2({{1, 2, 3}, {4, 5, 6}, {7, 8, 9}}))); auto reshape1 = builder.AddInstruction( HloInstruction::CreateReshape(ShapeUtil::MakeShape(S32, {3, 3}), const0)); hlo_module->AddEntryComputation(builder.Build()) ->CreateFusionInstruction(/*instructions_to_fuse=*/{reshape1}, HloInstruction::FusionKind::kLoop); EXPECT_TRUE(LiteralTestUtil::Equal( - *Literal::CreateR2({{1, 2, 3}, {4, 5, 6}, {7, 8, 9}}), + *LiteralUtil::CreateR2({{1, 2, 3}, {4, 5, 6}, {7, 8, 9}}), *ExecuteAndTransfer(std::move(hlo_module), {}))); } @@ -421,14 +423,14 @@ XLA_TEST_F(FusionTest, Transpose_2by3) { auto builder = HloComputation::Builder(TestName()); auto hlo_module = CreateNewModule(); auto const0 = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR2({{1, 2, 3}, {4, 5, 6}}))); + LiteralUtil::CreateR2({{1, 2, 3}, {4, 5, 6}}))); auto reshape1 = builder.AddInstruction(HloInstruction::CreateTranspose( ShapeUtil::MakeShape(S32, {3, 2}), const0, {1, 0})); hlo_module->AddEntryComputation(builder.Build()) ->CreateFusionInstruction(/*instructions_to_fuse=*/{reshape1}, HloInstruction::FusionKind::kLoop); EXPECT_TRUE(LiteralTestUtil::Equal( - *Literal::CreateR2({{1, 4}, {2, 5}, {3, 6}}), + *LiteralUtil::CreateR2({{1, 4}, {2, 5}, {3, 6}}), *ExecuteAndTransfer(std::move(hlo_module), {}))); } @@ -436,14 +438,14 @@ XLA_TEST_F(FusionTest, Transpose_3by3) { auto builder = HloComputation::Builder(TestName()); auto hlo_module = CreateNewModule(); auto const0 = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR2({{1, 2, 3}, {4, 5, 6}, {7, 8, 9}}))); + LiteralUtil::CreateR2({{1, 2, 3}, {4, 5, 6}, {7, 8, 9}}))); auto reshape1 = builder.AddInstruction(HloInstruction::CreateTranspose( ShapeUtil::MakeShape(S32, {3, 3}), const0, {1, 0})); hlo_module->AddEntryComputation(builder.Build()) ->CreateFusionInstruction(/*instructions_to_fuse=*/{reshape1}, HloInstruction::FusionKind::kLoop); EXPECT_TRUE(LiteralTestUtil::Equal( - *Literal::CreateR2({{1, 4, 7}, {2, 5, 8}, {3, 6, 9}}), + *LiteralUtil::CreateR2({{1, 4, 7}, {2, 5, 8}, {3, 6, 9}}), *ExecuteAndTransfer(std::move(hlo_module), {}))); } @@ -451,7 +453,7 @@ XLA_TEST_F(FusionTest, Reverse) { auto builder = HloComputation::Builder(TestName()); auto hlo_module = CreateNewModule(); auto const0 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR1({1, 2, 3}))); + HloInstruction::CreateConstant(LiteralUtil::CreateR1({1, 2, 3}))); auto reverse1 = builder.AddInstruction(HloInstruction::CreateReverse( ShapeUtil::MakeShape(S32, {3}), const0, {0})); hlo_module->AddEntryComputation(builder.Build()) @@ -459,7 +461,7 @@ XLA_TEST_F(FusionTest, Reverse) { HloInstruction::FusionKind::kLoop); EXPECT_TRUE( - LiteralTestUtil::Equal(*Literal::CreateR1({3, 2, 1}), + LiteralTestUtil::Equal(*LiteralUtil::CreateR1({3, 2, 1}), *ExecuteAndTransfer(std::move(hlo_module), {}))); } @@ -467,7 +469,7 @@ XLA_TEST_F(FusionTest, ReverseNegate) { auto builder = HloComputation::Builder(TestName()); auto hlo_module = CreateNewModule(); auto const0 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR1({1, 2, 3}))); + HloInstruction::CreateConstant(LiteralUtil::CreateR1({1, 2, 3}))); auto reverse1 = builder.AddInstruction(HloInstruction::CreateReverse( ShapeUtil::MakeShape(S32, {3}), const0, {0})); auto negate2 = builder.AddInstruction(HloInstruction::CreateUnary( @@ -477,7 +479,7 @@ XLA_TEST_F(FusionTest, ReverseNegate) { HloInstruction::FusionKind::kLoop); EXPECT_TRUE( - LiteralTestUtil::Equal(*Literal::CreateR1({-3, -2, -1}), + LiteralTestUtil::Equal(*LiteralUtil::CreateR1({-3, -2, -1}), *ExecuteAndTransfer(std::move(hlo_module), {}))); } @@ -485,7 +487,7 @@ XLA_TEST_F(FusionTest, BroadcastNegate) { auto builder = HloComputation::Builder(TestName()); auto hlo_module = CreateNewModule(); auto const0 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1))); auto broadcast1 = builder.AddInstruction(HloInstruction::CreateBroadcast( ShapeUtil::MakeShape(S32, {2}), const0, {})); auto negate2 = builder.AddInstruction(HloInstruction::CreateUnary( @@ -495,15 +497,15 @@ XLA_TEST_F(FusionTest, BroadcastNegate) { HloInstruction::FusionKind::kLoop); EXPECT_TRUE( - LiteralTestUtil::Equal(*Literal::CreateR1({-1, -1}), + LiteralTestUtil::Equal(*LiteralUtil::CreateR1({-1, -1}), *ExecuteAndTransfer(std::move(hlo_module), {}))); } XLA_TEST_F(FusionTest, SliceNegate) { auto builder = HloComputation::Builder(TestName()); auto hlo_module = CreateNewModule(); - auto const0 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR1({1, 2, 3, 4}))); + auto const0 = builder.AddInstruction(HloInstruction::CreateConstant( + LiteralUtil::CreateR1({1, 2, 3, 4}))); auto slice1 = builder.AddInstruction(HloInstruction::CreateSlice( ShapeUtil::MakeShape(S32, {2}), const0, {0}, {4}, {2})); auto negate2 = builder.AddInstruction(HloInstruction::CreateUnary( @@ -513,17 +515,17 @@ XLA_TEST_F(FusionTest, SliceNegate) { HloInstruction::FusionKind::kLoop); EXPECT_TRUE( - LiteralTestUtil::Equal(*Literal::CreateR1({-1, -3}), + LiteralTestUtil::Equal(*LiteralUtil::CreateR1({-1, -3}), *ExecuteAndTransfer(std::move(hlo_module), {}))); } XLA_TEST_F(FusionTest, DynamicSliceNegate) { auto builder = HloComputation::Builder(TestName()); auto hlo_module = CreateNewModule(); - auto const0 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR1({1, 2, 3, 4}))); + auto const0 = builder.AddInstruction(HloInstruction::CreateConstant( + LiteralUtil::CreateR1({1, 2, 3, 4}))); auto const1 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR1({1}))); + HloInstruction::CreateConstant(LiteralUtil::CreateR1({1}))); auto dynamic_slice2 = builder.AddInstruction(HloInstruction::CreateDynamicSlice( ShapeUtil::MakeShape(S32, {2}), const0, const1, {2})); @@ -535,15 +537,15 @@ XLA_TEST_F(FusionTest, DynamicSliceNegate) { HloInstruction::FusionKind::kLoop); EXPECT_TRUE( - LiteralTestUtil::Equal(*Literal::CreateR1({-2, -3}), + LiteralTestUtil::Equal(*LiteralUtil::CreateR1({-2, -3}), *ExecuteAndTransfer(std::move(hlo_module), {}))); } XLA_TEST_F(FusionTest, ReshapeNegate) { auto builder = HloComputation::Builder(TestName()); auto hlo_module = CreateNewModule(); - auto const0 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR1({1, 2, 3, 4}))); + auto const0 = builder.AddInstruction(HloInstruction::CreateConstant( + LiteralUtil::CreateR1({1, 2, 3, 4}))); auto reshape1 = builder.AddInstruction( HloInstruction::CreateReshape(ShapeUtil::MakeShape(S32, {2, 2}), const0)); auto negate2 = builder.AddInstruction(HloInstruction::CreateUnary( @@ -552,17 +554,16 @@ XLA_TEST_F(FusionTest, ReshapeNegate) { ->CreateFusionInstruction(/*instructions_to_fuse=*/{negate2, reshape1}, HloInstruction::FusionKind::kLoop); - EXPECT_TRUE( - LiteralTestUtil::Equal(*Literal::CreateR2({{-1, -2}, {-3, -4}}), - *ExecuteAndTransfer(std::move(hlo_module), {}))); + EXPECT_TRUE(LiteralTestUtil::Equal( + *LiteralUtil::CreateR2({{-1, -2}, {-3, -4}}), + *ExecuteAndTransfer(std::move(hlo_module), {}))); } -// TODO(b/64070202): Investigate failure. -XLA_TEST_F(FusionTest, DISABLED_ON_GPU(TransposeNegate)) { +XLA_TEST_F(FusionTest, TransposeNegate) { auto builder = HloComputation::Builder(TestName()); auto hlo_module = CreateNewModule(); auto const0 = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR2({{1, 2}, {3, 4}}))); + LiteralUtil::CreateR2({{1, 2}, {3, 4}}))); auto transpose1 = builder.AddInstruction(HloInstruction::CreateTranspose( ShapeUtil::MakeShape(S32, {2, 2}), const0, {1, 0})); auto negate2 = builder.AddInstruction(HloInstruction::CreateUnary( @@ -571,9 +572,9 @@ XLA_TEST_F(FusionTest, DISABLED_ON_GPU(TransposeNegate)) { ->CreateFusionInstruction(/*instructions_to_fuse=*/{negate2, transpose1}, HloInstruction::FusionKind::kLoop); - EXPECT_TRUE( - LiteralTestUtil::Equal(*Literal::CreateR2({{-1, -3}, {-2, -4}}), - *ExecuteAndTransfer(std::move(hlo_module), {}))); + EXPECT_TRUE(LiteralTestUtil::Equal( + *LiteralUtil::CreateR2({{-1, -3}, {-2, -4}}), + *ExecuteAndTransfer(std::move(hlo_module), {}))); } std::unique_ptr MakeReduceTestComputation() { @@ -591,10 +592,10 @@ XLA_TEST_F(FusionTest, DISABLED_ON_CPU(Reduce)) { auto hlo_module = CreateNewModule(); auto builder = HloComputation::Builder(TestName()); - auto const0 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR1({1, 2, 4, 8}))); + auto const0 = builder.AddInstruction(HloInstruction::CreateConstant( + LiteralUtil::CreateR1({1, 2, 4, 8}))); auto const1 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(0))); auto reduce2 = builder.AddInstruction(HloInstruction::CreateReduce( ShapeUtil::MakeShape(S32, {}), const0, const1, {0}, hlo_module->AddEmbeddedComputation(MakeReduceTestComputation()))); @@ -603,7 +604,7 @@ XLA_TEST_F(FusionTest, DISABLED_ON_CPU(Reduce)) { HloInstruction::FusionKind::kLoop); EXPECT_TRUE( - LiteralTestUtil::Equal(*Literal::CreateR0(15), + LiteralTestUtil::Equal(*LiteralUtil::CreateR0(15), *ExecuteAndTransfer(std::move(hlo_module), {}))); } @@ -611,10 +612,10 @@ XLA_TEST_F(FusionTest, DISABLED_ON_CPU(ReduceImplicitBroadcast)) { auto hlo_module = CreateNewModule(); auto builder = HloComputation::Builder(TestName()); - auto const0 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR1({1, 2, 4, 8}))); + auto const0 = builder.AddInstruction(HloInstruction::CreateConstant( + LiteralUtil::CreateR1({1, 2, 4, 8}))); auto const1 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(0))); auto reduce2 = builder.AddInstruction(HloInstruction::CreateReduce( ShapeUtil::MakeShape(S32, {}), const0, const1, {0}, hlo_module->AddEmbeddedComputation(MakeReduceTestComputation()))); @@ -625,7 +626,7 @@ XLA_TEST_F(FusionTest, DISABLED_ON_CPU(ReduceImplicitBroadcast)) { HloInstruction::FusionKind::kLoop); EXPECT_TRUE( - LiteralTestUtil::Equal(*Literal::CreateR0(-15), + LiteralTestUtil::Equal(*LiteralUtil::CreateR0(-15), *ExecuteAndTransfer(std::move(hlo_module), {}))); } @@ -633,9 +634,9 @@ XLA_TEST_F(FusionTest, DISABLED_ON_CPU(ReduceWindow)) { auto builder = HloComputation::Builder(TestName()); auto hlo_module = CreateNewModule(); auto const0 = builder.AddInstruction(HloInstruction::CreateConstant( - Literal::CreateR2({{2, 3, 5}, {7, 11, 13}, {17, 19, 23}}))); + LiteralUtil::CreateR2({{2, 3, 5}, {7, 11, 13}, {17, 19, 23}}))); auto const1 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(1))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(1))); Window window; ASSERT_TRUE( tensorflow::protobuf::TextFormat::ParseFromString("dimensions:{\n" @@ -675,7 +676,7 @@ XLA_TEST_F(FusionTest, DISABLED_ON_CPU(ReduceWindow)) { HloInstruction::FusionKind::kLoop); EXPECT_TRUE(LiteralTestUtil::Equal( - *Literal::CreateR2({{462, 2145}, {24871, 62491}}), + *LiteralUtil::CreateR2({{462, 2145}, {24871, 62491}}), *ExecuteAndTransfer(std::move(hlo_module), {}))); } @@ -687,9 +688,9 @@ XLA_TEST_F(FusionTest, SharedConstant) { auto builder = HloComputation::Builder(TestName()); auto const0 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR1({0}))); + HloInstruction::CreateConstant(LiteralUtil::CreateR1({0}))); auto const1 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR1({2}))); + HloInstruction::CreateConstant(LiteralUtil::CreateR1({2}))); auto add1 = builder.AddInstruction(HloInstruction::CreateBinary( ShapeUtil::MakeShape(S32, {1}), HloOpcode::kAdd, const1, const0)); auto add2 = builder.AddInstruction(HloInstruction::CreateBinary( @@ -711,7 +712,7 @@ XLA_TEST_F(FusionTest, SharedConstant) { EXPECT_EQ(entry_comp->root_instruction()->fused_instruction_count(), 6); EXPECT_TRUE( - LiteralTestUtil::Equal(*Literal::CreateR1({8}), + LiteralTestUtil::Equal(*LiteralUtil::CreateR1({8}), *ExecuteAndTransfer(std::move(hlo_module), {}))); } @@ -765,6 +766,39 @@ XLA_TEST_F(FusionTest, Clamp2D) { TestElementwise2D(HloOpcode::kClamp); } +// TODO(b/73903144): Enable on interpreter once interpreter supports bitcast. +XLA_TEST_F(FusionTest, DISABLED_ON_INTERPRETER(FusionWithLayout)) { + const string hlo_text = R"( +HloModule Cluster + +fusion_c { + fusion.arg = f32[2,2]{1,0} parameter(0) + bitcast.0 = f32[2,2,1]{2,1,0} bitcast(fusion.arg) + tanh.0 = f32[2,2,1]{0,2,1} tanh(bitcast.0) + ROOT bitcast.2 = f32[2,2,1]{1,2,0} bitcast(tanh.0) +} + +ENTRY main { + arg = f32[2,2]{1,0} parameter(0) + ROOT fusion = f32[2,2,1]{1,2,0} fusion(arg), kind=kLoop, calls=fusion_c +} +)"; + + std::unique_ptr operand = + LiteralUtil::CreateR2({{0., 0.}, {1., 0.}}); + HloModuleConfig config; + config.set_debug_options(GetDebugOptionsForTest()); + TF_ASSERT_OK_AND_ASSIGN(std::unique_ptr module, + ParseHloString(hlo_text, config)); + TF_ASSERT_OK_AND_ASSIGN( + std::unique_ptr result, + test_runner_.Execute(std::move(module), {operand.get()}, + /*run_hlo_passes=*/false)); + EXPECT_TRUE(LiteralTestUtil::Equal( + *LiteralUtil::CreateR3({{{0.}, {0.76159415595}}, {{0.}, {0.}}}), + *result)); +} + void BM_ParallelFusion(int num_iters) { // Simple element-wise computation to benchmark parallel task partitioning. tensorflow::testing::StopTiming(); @@ -793,31 +827,31 @@ void BM_ParallelFusion(int num_iters) { // Create computation. XlaBuilder builder("ParallelFusion"); Shape shape0 = ShapeUtil::MakeShape(F32, {param0_dim0, param0_dim1}); - auto param0 = builder.Parameter(0, shape0, "param0"); + auto param0 = Parameter(&builder, 0, shape0, "param0"); Shape shape1 = ShapeUtil::MakeShape(F32, {param1_dim0, param1_dim1}); - auto param1 = builder.Parameter(1, shape1, "param1"); + auto param1 = Parameter(&builder, 1, shape1, "param1"); Shape shape2 = ShapeUtil::MakeShape(F32, {param2_dim0, param2_dim1}); - auto param2 = builder.Parameter(2, shape2, "param2"); + auto param2 = Parameter(&builder, 2, shape2, "param2"); - auto x = builder.Mul(param0, param1); - auto y = builder.Add(x, param2); + auto x = Mul(param0, param1); + Add(x, param2); auto computation = builder.Build().ConsumeValueOrDie(); // Transfer literals to device. auto param0_literal = - Literal::CreateR2F32Linspace(1.0, 2.0, param0_dim0, param0_dim1); + LiteralUtil::CreateR2F32Linspace(1.0, 2.0, param0_dim0, param0_dim1); ScopedShapedBuffer buffer0 = client->LiteralToShapedBuffer(*param0_literal, device_ordinal) .ConsumeValueOrDie(); auto param1_literal = - Literal::CreateR2F32Linspace(1.0, 2.0, param1_dim0, param1_dim1); + LiteralUtil::CreateR2F32Linspace(1.0, 2.0, param1_dim0, param1_dim1); ScopedShapedBuffer buffer1 = client->LiteralToShapedBuffer(*param1_literal, device_ordinal) .ConsumeValueOrDie(); auto param2_literal = - Literal::CreateR2F32Linspace(1.0, 2.0, param2_dim0, param2_dim1); + LiteralUtil::CreateR2F32Linspace(1.0, 2.0, param2_dim0, param2_dim1); ScopedShapedBuffer buffer2 = client->LiteralToShapedBuffer(*param2_literal, device_ordinal) .ConsumeValueOrDie(); diff --git a/tensorflow/compiler/xla/tests/gather_operation_test.cc b/tensorflow/compiler/xla/tests/gather_operation_test.cc index 6fefae36958011c918cedc6703289551b00acc80..c5ca64fa3f0fc70542c577828d53eeecbd05067b 100644 --- a/tensorflow/compiler/xla/tests/gather_operation_test.cc +++ b/tensorflow/compiler/xla/tests/gather_operation_test.cc @@ -13,6 +13,7 @@ See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ +#include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/compiler/xla/execution_options_util.h" #include "tensorflow/compiler/xla/service/hlo_parser.h" #include "tensorflow/compiler/xla/status_macros.h" @@ -21,9 +22,6 @@ limitations under the License. #include "tensorflow/compiler/xla/tests/hlo_test_base.h" #include "tensorflow/compiler/xla/tests/test_macros.h" -// NB! TODO(b/74360564): These tests do not test out of bounds behavior since -// that hasn't been specced yet. - namespace xla { namespace { @@ -62,8 +60,9 @@ ENTRY main { } )"; std::unique_ptr operand = - Literal::CreateR2({{1, 2, 3}, {4, 5, 6}, {7, 8, 9}}); - std::unique_ptr gather_indices = Literal::CreateR1({0, 2}); + LiteralUtil::CreateR2({{1, 2, 3}, {4, 5, 6}, {7, 8, 9}}); + std::unique_ptr gather_indices = + LiteralUtil::CreateR1({0, 2}); RunTest(hlo_text, operand.get(), gather_indices.get()); } @@ -83,8 +82,9 @@ ENTRY main { } )"; std::unique_ptr operand = - Literal::CreateR2({{1, 2, 3}, {4, 5, 6}, {7, 8, 9}}); - std::unique_ptr gather_indices = Literal::CreateR1({0, 2}); + LiteralUtil::CreateR2({{1, 2, 3}, {4, 5, 6}, {7, 8, 9}}); + std::unique_ptr gather_indices = + LiteralUtil::CreateR1({0, 2}); RunTest(hlo_text, operand.get(), gather_indices.get()); } @@ -104,9 +104,9 @@ ENTRY main { } )"; std::unique_ptr operand = - Literal::CreateR2({{1, 2, 3}, {4, 5, 6}, {7, 8, 9}}); + LiteralUtil::CreateR2({{1, 2, 3}, {4, 5, 6}, {7, 8, 9}}); std::unique_ptr gather_indices = - Literal::CreateR2({{0, 2}, {2, 1}}); + LiteralUtil::CreateR2({{0, 2}, {2, 1}}); RunTest(hlo_text, operand.get(), gather_indices.get()); } @@ -126,9 +126,9 @@ ENTRY main { } )"; std::unique_ptr operand = - Literal::CreateR2({{1, 2, 3}, {4, 5, 6}, {7, 8, 9}}); + LiteralUtil::CreateR2({{1, 2, 3}, {4, 5, 6}, {7, 8, 9}}); std::unique_ptr gather_indices = - Literal::CreateR3({{{0, 2}, {2, 1}}, {{1, 2}, {2, 0}}}); + LiteralUtil::CreateR3({{{0, 2}, {2, 1}}, {{1, 2}, {2, 0}}}); RunTest(hlo_text, operand.get(), gather_indices.get()); } @@ -148,9 +148,9 @@ ENTRY main { } )"; std::unique_ptr operand = - Literal::CreateR2({{1, 2, 3}, {4, 5, 6}, {7, 8, 9}}); + LiteralUtil::CreateR2({{1, 2, 3}, {4, 5, 6}, {7, 8, 9}}); std::unique_ptr gather_indices = - Literal::CreateR3({{{0, 2}, {2, 1}}, {{1, 2}, {2, 0}}}); + LiteralUtil::CreateR3({{{0, 2}, {2, 1}}, {{1, 2}, {2, 0}}}); RunTest(hlo_text, operand.get(), gather_indices.get()); } @@ -170,11 +170,11 @@ ENTRY main { } )"; std::unique_ptr operand = - Literal::CreateR3({{{-1, 1}, {-2, 2}, {-3, 3}}, // - {{-4, 4}, {-5, 5}, {-6, 6}}, // - {{-7, 7}, {-8, 8}, {-9, 9}}}); + LiteralUtil::CreateR3({{{-1, 1}, {-2, 2}, {-3, 3}}, // + {{-4, 4}, {-5, 5}, {-6, 6}}, // + {{-7, 7}, {-8, 8}, {-9, 9}}}); std::unique_ptr gather_indices = - Literal::CreateR2({{0, 0}, {1, 0}}); + LiteralUtil::CreateR2({{0, 0}, {1, 0}}); RunTest(hlo_text, operand.get(), gather_indices.get()); } @@ -194,11 +194,11 @@ ENTRY main { } )"; std::unique_ptr operand = - Literal::CreateR3({{{-1, 1}, {-2, 2}, {-3, 3}}, // - {{-4, 4}, {-5, 5}, {-6, 6}}, // - {{-7, 7}, {-8, 8}, {-9, 9}}}); + LiteralUtil::CreateR3({{{-1, 1}, {-2, 2}, {-3, 3}}, // + {{-4, 4}, {-5, 5}, {-6, 6}}, // + {{-7, 7}, {-8, 8}, {-9, 9}}}); std::unique_ptr gather_indices = - Literal::CreateR2({{0, 0}, {1, 0}}); + LiteralUtil::CreateR2({{0, 0}, {1, 0}}); RunTest(hlo_text, operand.get(), gather_indices.get()); } @@ -218,8 +218,9 @@ ENTRY main { } )"; std::unique_ptr operand = - Literal::CreateR2({{1, 2, 3}, {4, 5, 6}, {7, 8, 9}}); - std::unique_ptr gather_indices = Literal::CreateR1({1, 1}); + LiteralUtil::CreateR2({{1, 2, 3}, {4, 5, 6}, {7, 8, 9}}); + std::unique_ptr gather_indices = + LiteralUtil::CreateR1({1, 1}); RunTest(hlo_text, operand.get(), gather_indices.get()); } @@ -239,9 +240,9 @@ ENTRY main { } )"; std::unique_ptr operand = - Literal::CreateR2({{1, 2, 3}, {4, 5, 6}, {7, 8, 9}}); + LiteralUtil::CreateR2({{1, 2, 3}, {4, 5, 6}, {7, 8, 9}}); std::unique_ptr gather_indices = - Literal::CreateR2({{2, 1}, {1, 1}}); + LiteralUtil::CreateR2({{2, 1}, {1, 1}}); RunTest(hlo_text, operand.get(), gather_indices.get()); } @@ -260,18 +261,15 @@ ENTRY main { window_bounds={1, 0} } )"; - std::unique_ptr operand = Literal::CreateR2({{}, {}, {}}); - std::unique_ptr gather_indices = Literal::CreateR1({0, 2}); + std::unique_ptr operand = LiteralUtil::CreateR2({{}, {}, {}}); + std::unique_ptr gather_indices = + LiteralUtil::CreateR1({0, 2}); RunTest(hlo_text, operand.get(), gather_indices.get()); } XLA_TEST_F(GatherOperationTest, OutOfBoundsIndex) { // Out of bounds indices must not crash, and the indices in range should // produce the same values across all backends. - // - // TODO(b/74360564): Once we have a well defined semantics for OOB accesses, - // we should get rid of the mask and check that backends produce the same - // value for OOB indices too. const string hlo_text = R"( HloModule BatchDynamicSlice @@ -285,29 +283,45 @@ ENTRY main { gather_dims_to_operand_dims={0,1}, index_vector_dim=1, window_bounds={1,1} - gather_reshaped = s32[6]{0} reshape(gather) - in_bounds_mask = s32[6]{0} parameter(2) - ROOT result = s32[6]{0} multiply(gather_reshaped, in_bounds_mask) + ROOT result = s32[6]{0} reshape(gather) } )"; std::unique_ptr operand = - Literal::CreateR2({{1, 2, 3}, {4, 5, 6}, {7, 8, 9}}); - std::unique_ptr gather_indices = Literal::CreateR2( + LiteralUtil::CreateR2({{1, 2, 3}, {4, 5, 6}, {7, 8, 9}}); + std::unique_ptr gather_indices = LiteralUtil::CreateR2( {{2, 7}, {2, 1}, {1, 1}, {5, 1}, {2147483647, 1}, {1, 2}}); - std::unique_ptr in_bounds_mask = - Literal::CreateR1({0, 1, 1, 0, 0, 1}); + RunTest(hlo_text, operand.get(), gather_indices.get()); +} + +XLA_TEST_F(GatherOperationTest, OutOfBoundsUnsignedIndex) { + // Out of bounds indices must not crash, and the indices in range should + // produce the same values across all backends. - RunTest(hlo_text, - {operand.get(), gather_indices.get(), in_bounds_mask.get()}); + const string hlo_text = R"( +HloModule BatchDynamicSlice + +ENTRY main { + operand = s32[3,3]{1,0} parameter(0) + indices = u32[6,2]{1,0} parameter(1) + gather = s32[6,1,1]{2,1,0} gather(operand, indices), + output_window_dims={1,2}, + elided_window_dims={}, + gather_dims_to_operand_dims={0,1}, + index_vector_dim=1, + window_bounds={1,1} + ROOT result = s32[6]{0} reshape(gather) +} +)"; + std::unique_ptr operand = + LiteralUtil::CreateR2({{1, 2, 3}, {4, 5, 6}, {7, 8, 9}}); + std::unique_ptr gather_indices = LiteralUtil::CreateR2( + {{2, 7}, {2, 1}, {1, 1}, {5, 1}, {2147483648u, 1}, {1, 2}}); + RunTest(hlo_text, operand.get(), gather_indices.get()); } XLA_TEST_F(GatherOperationTest, NegativeIndex) { // Negative indices must not crash, and the indices in range should produce // the same values across all backends. - // - // TODO(b/74360564): Once we have a well defined semantics for negative - // accesses, we should get rid of the mask and check that backends produce the - // same value for negative indices too. const string hlo_text = R"( HloModule BatchDynamicSlice @@ -321,20 +335,40 @@ ENTRY main { gather_dims_to_operand_dims={0,1}, index_vector_dim=1, window_bounds={1,1} - gather_reshaped = s32[6]{0} reshape(gather) - in_bounds_mask = s32[6]{0} parameter(2) - ROOT result = s32[6]{0} multiply(gather_reshaped, in_bounds_mask) + ROOT result = s32[6]{0} reshape(gather) } )"; std::unique_ptr operand = - Literal::CreateR2({{1, 2, 3}, {4, 5, 6}, {7, 8, 9}}); - std::unique_ptr gather_indices = Literal::CreateR2( + LiteralUtil::CreateR2({{1, 2, 3}, {4, 5, 6}, {7, 8, 9}}); + std::unique_ptr gather_indices = LiteralUtil::CreateR2( {{2, -1}, {2, 1}, {1, 1}, {-500, 1}, {-2147483648, 1}, {1, 2}}); - std::unique_ptr in_bounds_mask = - Literal::CreateR1({0, 1, 1, 0, 0, 1}); + RunTest(hlo_text, operand.get(), gather_indices.get()); +} + +XLA_TEST_F(GatherOperationTest, NegativeIndexIntoUnsignedOperand) { + // Negative indices must not crash, and the indices in range should produce + // the same values across all backends. - RunTest(hlo_text, - {operand.get(), gather_indices.get(), in_bounds_mask.get()}); + const string hlo_text = R"( +HloModule BatchDynamicSlice + +ENTRY main { + operand = u32[3,3]{1,0} parameter(0) + indices = s32[6,2]{1,0} parameter(1) + gather = u32[6,1,1]{2,1,0} gather(operand, indices), + output_window_dims={1,2}, + elided_window_dims={}, + gather_dims_to_operand_dims={0,1}, + index_vector_dim=1, + window_bounds={1,1} + ROOT result = u32[6]{0} reshape(gather) +} +)"; + std::unique_ptr operand = + LiteralUtil::CreateR2({{1, 2, 3}, {4, 5, 6}, {7, 8, 9}}); + std::unique_ptr gather_indices = LiteralUtil::CreateR2( + {{2, -1}, {2, 1}, {1, 1}, {-500, 1}, {-2147483648, 1}, {1, 2}}); + RunTest(hlo_text, operand.get(), gather_indices.get()); } XLA_TEST_F(GatherOperationTest, OneScalarIndex) { @@ -352,9 +386,9 @@ ENTRY main { window_bounds={1,3,2} } )"; - std::unique_ptr operand = Literal::CreateR3( + std::unique_ptr operand = LiteralUtil::CreateR3( {{{1, 2}, {3, 4}, {5, 6}}, {{7, 8}, {9, 10}, {11, 12}}}); - std::unique_ptr gather_indices = Literal::CreateR0(1); + std::unique_ptr gather_indices = LiteralUtil::CreateR0(1); RunTest(hlo_text, operand.get(), gather_indices.get()); } @@ -373,8 +407,8 @@ ENTRY main { window_bounds={1} } )"; - std::unique_ptr operand = Literal::CreateR1({1, 2, 3, 4}); - std::unique_ptr gather_indices = Literal::CreateR0(1); + std::unique_ptr operand = LiteralUtil::CreateR1({1, 2, 3, 4}); + std::unique_ptr gather_indices = LiteralUtil::CreateR0(1); RunTest(hlo_text, operand.get(), gather_indices.get()); } @@ -394,8 +428,8 @@ ENTRY main { } )"; std::unique_ptr operand = - Literal::CreateR2({{1, 2, 3}, {4, 5, 6}, {7, 8, 9}}); - std::unique_ptr gather_indices = Literal::CreateR1({}); + LiteralUtil::CreateR2({{1, 2, 3}, {4, 5, 6}, {7, 8, 9}}); + std::unique_ptr gather_indices = LiteralUtil::CreateR1({}); RunTest(hlo_text, operand.get(), gather_indices.get()); } @@ -418,8 +452,9 @@ ENTRY main { } )"; std::unique_ptr operand = - Literal::CreateR2({{1, 2, 3}, {4, 5, 6}, {7, 8, 9}}); - std::unique_ptr gather_indices = Literal::CreateR1({0, 2}); + LiteralUtil::CreateR2({{1, 2, 3}, {4, 5, 6}, {7, 8, 9}}); + std::unique_ptr gather_indices = + LiteralUtil::CreateR1({0, 2}); RunTest(hlo_text, operand.get(), gather_indices.get()); } @@ -442,9 +477,9 @@ ENTRY main { } )"; std::unique_ptr operand = - Literal::CreateR2({{1, 2, 3}, {4, 5, 6}, {7, 8, 9}}); + LiteralUtil::CreateR2({{1, 2, 3}, {4, 5, 6}, {7, 8, 9}}); std::unique_ptr gather_indices = - Literal::CreateR2({{0, 2}, {2, 1}}); + LiteralUtil::CreateR2({{0, 2}, {2, 1}}); RunTest(hlo_text, operand.get(), gather_indices.get()); } @@ -467,9 +502,9 @@ ENTRY main { } )"; std::unique_ptr operand = - Literal::CreateR2({{1, 2, 3}, {4, 5, 6}, {7, 8, 9}}); + LiteralUtil::CreateR2({{1, 2, 3}, {4, 5, 6}, {7, 8, 9}}); std::unique_ptr gather_indices = - Literal::CreateR3({{{0, 2}, {2, 1}}, {{1, 2}, {2, 0}}}); + LiteralUtil::CreateR3({{{0, 2}, {2, 1}}, {{1, 2}, {2, 0}}}); RunTest(hlo_text, operand.get(), gather_indices.get()); } @@ -492,11 +527,11 @@ ENTRY main { } )"; std::unique_ptr operand = - Literal::CreateR3({{{-1, 1}, {-2, 2}, {-3, 3}}, // - {{-4, 4}, {-5, 5}, {-6, 6}}, // - {{-7, 7}, {-8, 8}, {-9, 9}}}); + LiteralUtil::CreateR3({{{-1, 1}, {-2, 2}, {-3, 3}}, // + {{-4, 4}, {-5, 5}, {-6, 6}}, // + {{-7, 7}, {-8, 8}, {-9, 9}}}); std::unique_ptr gather_indices = - Literal::CreateR2({{0, 0}, {1, 0}}); + LiteralUtil::CreateR2({{0, 0}, {1, 0}}); RunTest(hlo_text, operand.get(), gather_indices.get()); } @@ -520,11 +555,11 @@ ENTRY main { } )"; std::unique_ptr operand = - Literal::CreateR3({{{-1, 1}, {-2, 2}, {-3, 3}}, // - {{-4, 4}, {-5, 5}, {-6, 6}}, // - {{-7, 7}, {-8, 8}, {-9, 9}}}); + LiteralUtil::CreateR3({{{-1, 1}, {-2, 2}, {-3, 3}}, // + {{-4, 4}, {-5, 5}, {-6, 6}}, // + {{-7, 7}, {-8, 8}, {-9, 9}}}); std::unique_ptr gather_indices = - Literal::CreateR2({{0, 0}, {1, 0}}); + LiteralUtil::CreateR2({{0, 0}, {1, 0}}); RunTest(hlo_text, operand.get(), gather_indices.get()); } @@ -547,8 +582,9 @@ ENTRY main { } )"; std::unique_ptr operand = - Literal::CreateR2({{1, 2, 3}, {4, 5, 6}, {7, 8, 9}}); - std::unique_ptr gather_indices = Literal::CreateR1({1, 1}); + LiteralUtil::CreateR2({{1, 2, 3}, {4, 5, 6}, {7, 8, 9}}); + std::unique_ptr gather_indices = + LiteralUtil::CreateR1({1, 1}); RunTest(hlo_text, operand.get(), gather_indices.get()); } @@ -571,9 +607,9 @@ ENTRY main { } )"; std::unique_ptr operand = - Literal::CreateR2({{1, 2, 3}, {4, 5, 6}, {7, 8, 9}}); + LiteralUtil::CreateR2({{1, 2, 3}, {4, 5, 6}, {7, 8, 9}}); std::unique_ptr gather_indices = - Literal::CreateR2({{2, 1}, {1, 1}}); + LiteralUtil::CreateR2({{2, 1}, {1, 1}}); RunTest(hlo_text, operand.get(), gather_indices.get()); } @@ -598,22 +634,23 @@ XLA_TEST_F(GatherClientLibraryTest, DISABLED_ON_GPU(Basic)) { Shape operand_shape = ShapeUtil::MakeShape(S32, {3, 3}); Shape indices_shape = ShapeUtil::MakeShape(S32, {2}); - auto operand = builder.Parameter(0, operand_shape, "operand"); - auto indices = builder.Parameter(1, indices_shape, "indices"); + auto operand = Parameter(&builder, 0, operand_shape, "operand"); + auto indices = Parameter(&builder, 1, indices_shape, "indices"); GatherDimensionNumbers dim_numbers; dim_numbers.add_output_window_dims(1); dim_numbers.add_elided_window_dims(0); dim_numbers.add_gather_dims_to_operand_dims(0); dim_numbers.set_index_vector_dim(1); - builder.Gather(operand, indices, dim_numbers, {1, 3}); + Gather(operand, indices, dim_numbers, {1, 3}); std::vector expected = {}; - TF_ASSERT_OK_AND_ASSIGN(std::unique_ptr operand_arg, - client_->TransferToServer(*Literal::CreateR2( - {{1, 2, 3}, {4, 5, 6}, {7, 8, 9}}))); + TF_ASSERT_OK_AND_ASSIGN( + std::unique_ptr operand_arg, + client_->TransferToServer( + *LiteralUtil::CreateR2({{1, 2, 3}, {4, 5, 6}, {7, 8, 9}}))); TF_ASSERT_OK_AND_ASSIGN( std::unique_ptr indices_arg, - client_->TransferToServer(*Literal::CreateR1({0, 2}))); + client_->TransferToServer(*LiteralUtil::CreateR1({0, 2}))); TF_ASSERT_OK_AND_ASSIGN(std::vector devices, client_->GetDeviceHandles(1)); xla::ExecutionOptions execution_options = CreateDefaultExecutionOptions(); diff --git a/tensorflow/compiler/xla/tests/half_test.cc b/tensorflow/compiler/xla/tests/half_test.cc index 76bf47845ca045b4eede9a3b47ae5c2ce93ce577..73a47eda721971c75f61109787844c40be0b7080 100644 --- a/tensorflow/compiler/xla/tests/half_test.cc +++ b/tensorflow/compiler/xla/tests/half_test.cc @@ -17,7 +17,7 @@ limitations under the License. #include #include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/statusor.h" #include "tensorflow/compiler/xla/test.h" #include "tensorflow/compiler/xla/test_helpers.h" @@ -37,8 +37,7 @@ class HalfTestBase : public ClientLibraryTestBase { static const int kNumElements = 4; }; -using UnaryBuildFuncTy = - std::function; +using UnaryBuildFuncTy = std::function; struct UnaryOpTestParam { std::function compute_func; @@ -62,7 +61,7 @@ XLA_TEST_P(UnaryOpTest, Ops) { } UnaryBuildFuncTy build_func = GetParam().build_func; - build_func(&builder, x_opnd); + build_func(x_opnd); ComputeAndCompareR1(&builder, expected, {x_data.get()}, error_spec_); } @@ -79,18 +78,17 @@ half round_imp(half value) { INSTANTIATE_TEST_CASE_P( half, UnaryOpTest, ::testing::Values( - UnaryOpTestParam{[](half x) { return abs(x); }, &XlaBuilder::Abs}, - UnaryOpTestParam{[](half x) { return round_imp(x); }, - &XlaBuilder::Round}, - UnaryOpTestParam{[](half x) { return ceil(x); }, &XlaBuilder::Ceil}, - UnaryOpTestParam{[](half x) { return cos(x); }, &XlaBuilder::Cos}, - UnaryOpTestParam{[](half x) { return exp(x); }, &XlaBuilder::Exp}, - UnaryOpTestParam{[](half x) { return floor(x); }, &XlaBuilder::Floor}, - UnaryOpTestParam{[](half x) { return log(x); }, &XlaBuilder::Log}, - UnaryOpTestParam{[](half x) { return -x; }, &XlaBuilder::Neg}, - UnaryOpTestParam{[](half x) { return sign_imp(x); }, &XlaBuilder::Sign}, - UnaryOpTestParam{[](half x) { return sin(x); }, &XlaBuilder::Sin}, - UnaryOpTestParam{[](half x) { return tanh(x); }, &XlaBuilder::Tanh} + UnaryOpTestParam{[](half x) { return abs(x); }, &Abs}, + UnaryOpTestParam{[](half x) { return round_imp(x); }, &Round}, + UnaryOpTestParam{[](half x) { return ceil(x); }, &Ceil}, + UnaryOpTestParam{[](half x) { return cos(x); }, &Cos}, + UnaryOpTestParam{[](half x) { return exp(x); }, &Exp}, + UnaryOpTestParam{[](half x) { return floor(x); }, &Floor}, + UnaryOpTestParam{[](half x) { return log(x); }, &Log}, + UnaryOpTestParam{[](half x) { return -x; }, &Neg}, + UnaryOpTestParam{[](half x) { return sign_imp(x); }, &Sign}, + UnaryOpTestParam{[](half x) { return sin(x); }, &Sin}, + UnaryOpTestParam{[](half x) { return tanh(x); }, &Tanh} )); @@ -118,19 +116,18 @@ XLA_TEST_P(UnaryPredTest, Ops) { } UnaryBuildFuncTy build_func = GetParam().build_func; - build_func(&builder, x_opnd); + build_func(x_opnd); ComputeAndCompareR1(&builder, expected, {x_data.get()}); } INSTANTIATE_TEST_CASE_P(half, UnaryPredTest, ::testing::Values(UnaryPredTestParam{ - [](half x) { return isfinite(x); }, - &XlaBuilder::IsFinite})); + [](half x) { return isfinite(x); }, &IsFinite})); -using BinaryBuildFuncTy = std::function)>; +using BinaryBuildFuncTy = + std::function)>; struct BinaryOpTestParam { std::function compute_func; @@ -159,7 +156,7 @@ XLA_TEST_P(BinaryOpTest, Ops) { } BinaryBuildFuncTy build_func = GetParam().build_func; - build_func(&builder, x_opnd, y_opnd, {}); + build_func(x_opnd, y_opnd, {}); ComputeAndCompareR1(&builder, expected, {x_data.get(), y_data.get()}, error_spec_); @@ -173,22 +170,15 @@ half atan2_imp(half x, half y) { INSTANTIATE_TEST_CASE_P( half, BinaryOpTest, ::testing::Values( - BinaryOpTestParam{[](half x, half y) { return x + y; }, - &XlaBuilder::Add}, + BinaryOpTestParam{[](half x, half y) { return x + y; }, &Add}, BinaryOpTestParam{[](half x, half y) { return atan2_imp(x, y); }, - &XlaBuilder::Atan2}, - BinaryOpTestParam{[](half x, half y) { return x / y; }, - &XlaBuilder::Div}, - BinaryOpTestParam{[](half x, half y) { return max(x, y); }, - &XlaBuilder::Max}, - BinaryOpTestParam{[](half x, half y) { return min(x, y); }, - &XlaBuilder::Min}, - BinaryOpTestParam{[](half x, half y) { return x * y; }, - &XlaBuilder::Mul}, - BinaryOpTestParam{[](half x, half y) { return pow(x, y); }, - &XlaBuilder::Pow}, - BinaryOpTestParam{[](half x, half y) { return x - y; }, - &XlaBuilder::Sub} + &Atan2}, + BinaryOpTestParam{[](half x, half y) { return x / y; }, &Div}, + BinaryOpTestParam{[](half x, half y) { return max(x, y); }, &Max}, + BinaryOpTestParam{[](half x, half y) { return min(x, y); }, &Min}, + BinaryOpTestParam{[](half x, half y) { return x * y; }, &Mul}, + BinaryOpTestParam{[](half x, half y) { return pow(x, y); }, &Pow}, + BinaryOpTestParam{[](half x, half y) { return x - y; }, &Sub} )); @@ -221,27 +211,22 @@ XLA_TEST_P(BinaryPredTest, Ops) { } BinaryBuildFuncTy build_func = GetParam().build_func; - build_func(&builder, x_opnd, y_opnd, {}); + build_func(x_opnd, y_opnd, {}); ComputeAndCompareR1(&builder, expected, {x_data.get(), y_data.get()}); } INSTANTIATE_TEST_CASE_P( half, BinaryPredTest, - ::testing::Values(BinaryPredTestParam{[](half x, half y) { return x == y; }, - &XlaBuilder::Eq}, - BinaryPredTestParam{[](half x, half y) { return x != y; }, - &XlaBuilder::Ne}, - BinaryPredTestParam{[](half x, half y) { return x >= y; }, - &XlaBuilder::Ge}, - BinaryPredTestParam{[](half x, half y) { return x > y; }, - &XlaBuilder::Gt}, - BinaryPredTestParam{[](half x, half y) { return x <= y; }, - &XlaBuilder::Le}, - BinaryPredTestParam{[](half x, half y) { return x < y; }, - &XlaBuilder::Lt} - - )); + ::testing::Values( + BinaryPredTestParam{[](half x, half y) { return x == y; }, &Eq}, + BinaryPredTestParam{[](half x, half y) { return x != y; }, &Ne}, + BinaryPredTestParam{[](half x, half y) { return x >= y; }, &Ge}, + BinaryPredTestParam{[](half x, half y) { return x > y; }, &Gt}, + BinaryPredTestParam{[](half x, half y) { return x <= y; }, &Le}, + BinaryPredTestParam{[](half x, half y) { return x < y; }, &Lt} + + )); } // namespace } // namespace xla diff --git a/tensorflow/compiler/xla/tests/hlo_metadata_test.cc b/tensorflow/compiler/xla/tests/hlo_metadata_test.cc index cf971dd61b71ad329b20b0bb7c16166126562681..4d82442f7e3630c115eff1f17544e2b892c5e7eb 100644 --- a/tensorflow/compiler/xla/tests/hlo_metadata_test.cc +++ b/tensorflow/compiler/xla/tests/hlo_metadata_test.cc @@ -30,9 +30,9 @@ class HloMetadataTest : public LocalClientTestBase { } void BuildAddComputation(XlaBuilder* builder) { - auto x = builder->Parameter(0, ShapeUtil::MakeShape(F32, {}), "x"); - auto y = builder->Parameter(1, ShapeUtil::MakeShape(F32, {}), "y"); - builder->Add(x, y); + auto x = Parameter(builder, 0, ShapeUtil::MakeShape(F32, {}), "x"); + auto y = Parameter(builder, 1, ShapeUtil::MakeShape(F32, {}), "y"); + Add(x, y); } OpMetadata metadata_; diff --git a/tensorflow/compiler/xla/tests/hlo_test_base.cc b/tensorflow/compiler/xla/tests/hlo_test_base.cc index 242cc5db11ff2bdf69209df7537216573d8afbf3..b662e837168c8b16daea0181786be19fa0237a8c 100644 --- a/tensorflow/compiler/xla/tests/hlo_test_base.cc +++ b/tensorflow/compiler/xla/tests/hlo_test_base.cc @@ -276,9 +276,10 @@ StatusOr<::testing::AssertionResult> HloTestBase::RunAndCompareInternal( HloComputation* HloTestBase::FindComputation(HloModule* module, tensorflow::StringPiece name) { - auto it = c_find_if(module->computations(), + auto computations = module->computations(); + auto it = c_find_if(computations, [&](HloComputation* c) { return c->name() == name; }); - if (it == module->computations().end()) { + if (it == computations.end()) { return nullptr; } return *it; @@ -287,9 +288,10 @@ HloComputation* HloTestBase::FindComputation(HloModule* module, HloInstruction* HloTestBase::FindInstruction(HloModule* module, tensorflow::StringPiece name) { for (const HloComputation* c : module->computations()) { - auto it = c_find_if(c->instructions(), + auto instructions = c->instructions(); + auto it = c_find_if(instructions, [&](HloInstruction* i) { return i->name() == name; }); - if (it != c->instructions().end()) { + if (it != instructions.end()) { return *it; } } diff --git a/tensorflow/compiler/xla/tests/hlo_test_base.h b/tensorflow/compiler/xla/tests/hlo_test_base.h index 9009d67cea6840235d63724ef76d777c8f693d33..66719b1460063a61541535ff7507468ae0ca1ada 100644 --- a/tensorflow/compiler/xla/tests/hlo_test_base.h +++ b/tensorflow/compiler/xla/tests/hlo_test_base.h @@ -200,6 +200,13 @@ class HloTestBase : public ::testing::Test { ->ResetLayout(layout); } + void ForceResultLayout(HloModule* module, const Layout& layout, + ShapeIndexView shape_index) { + module->mutable_entry_computation_layout() + ->mutable_result_layout() + ->ResetLayout(layout, shape_index); + } + // Convenience method to clear the layout of the computation result in // 'module'. void ForceClearResultLayout(HloModule* module) { diff --git a/tensorflow/compiler/xla/tests/literal_test_util.h b/tensorflow/compiler/xla/tests/literal_test_util.h index d1b8a6cf0b2552f1b7d95a2560d502da14ddc39a..31a099c15f1f20457c90de97054f68a31eb49011 100644 --- a/tensorflow/compiler/xla/tests/literal_test_util.h +++ b/tensorflow/compiler/xla/tests/literal_test_util.h @@ -25,6 +25,7 @@ limitations under the License. #include "tensorflow/compiler/xla/array3d.h" #include "tensorflow/compiler/xla/array4d.h" #include "tensorflow/compiler/xla/error_spec.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/literal_util.h" #include "tensorflow/compiler/xla/test.h" #include "tensorflow/compiler/xla/test_helpers.h" @@ -154,20 +155,20 @@ class LiteralTestUtil { template /* static */ void LiteralTestUtil::ExpectR0Equal(NativeT expected, const LiteralSlice& actual) { - EXPECT_TRUE(Equal(*Literal::CreateR0(expected), actual)); + EXPECT_TRUE(Equal(*LiteralUtil::CreateR0(expected), actual)); } template /* static */ void LiteralTestUtil::ExpectR1Equal( tensorflow::gtl::ArraySlice expected, const LiteralSlice& actual) { - EXPECT_TRUE(Equal(*Literal::CreateR1(expected), actual)); + EXPECT_TRUE(Equal(*LiteralUtil::CreateR1(expected), actual)); } template /* static */ void LiteralTestUtil::ExpectR2Equal( std::initializer_list> expected, const LiteralSlice& actual) { - EXPECT_TRUE(Equal(*Literal::CreateR2(expected), actual)); + EXPECT_TRUE(Equal(*LiteralUtil::CreateR2(expected), actual)); } template @@ -175,46 +176,46 @@ template std::initializer_list>> expected, const LiteralSlice& actual) { - EXPECT_TRUE(Equal(*Literal::CreateR3(expected), actual)); + EXPECT_TRUE(Equal(*LiteralUtil::CreateR3(expected), actual)); } template /* static */ void LiteralTestUtil::ExpectR2EqualArray2D( const Array2D& expected, const LiteralSlice& actual) { - EXPECT_TRUE(Equal(*Literal::CreateR2FromArray2D(expected), actual)); + EXPECT_TRUE(Equal(*LiteralUtil::CreateR2FromArray2D(expected), actual)); } template /* static */ void LiteralTestUtil::ExpectR3EqualArray3D( const Array3D& expected, const LiteralSlice& actual) { - EXPECT_TRUE(Equal(*Literal::CreateR3FromArray3D(expected), actual)); + EXPECT_TRUE(Equal(*LiteralUtil::CreateR3FromArray3D(expected), actual)); } template /* static */ void LiteralTestUtil::ExpectR4EqualArray4D( const Array4D& expected, const LiteralSlice& actual) { - EXPECT_TRUE(Equal(*Literal::CreateR4FromArray4D(expected), actual)); + EXPECT_TRUE(Equal(*LiteralUtil::CreateR4FromArray4D(expected), actual)); } template /* static */ void LiteralTestUtil::ExpectR0Near(NativeT expected, const LiteralSlice& actual, const ErrorSpec& error) { - EXPECT_TRUE(Near(*Literal::CreateR0(expected), actual, error)); + EXPECT_TRUE(Near(*LiteralUtil::CreateR0(expected), actual, error)); } template /* static */ void LiteralTestUtil::ExpectR1Near( tensorflow::gtl::ArraySlice expected, const LiteralSlice& actual, const ErrorSpec& error) { - EXPECT_TRUE(Near(*Literal::CreateR1(expected), actual, error)); + EXPECT_TRUE(Near(*LiteralUtil::CreateR1(expected), actual, error)); } template /* static */ void LiteralTestUtil::ExpectR2Near( std::initializer_list> expected, const LiteralSlice& actual, const ErrorSpec& error) { - EXPECT_TRUE(Near(*Literal::CreateR2(expected), actual, error)); + EXPECT_TRUE(Near(*LiteralUtil::CreateR2(expected), actual, error)); } template @@ -222,7 +223,7 @@ template std::initializer_list>> expected, const LiteralSlice& actual, const ErrorSpec& error) { - EXPECT_TRUE(Near(*Literal::CreateR3(expected), actual, error)); + EXPECT_TRUE(Near(*LiteralUtil::CreateR3(expected), actual, error)); } template @@ -231,28 +232,28 @@ template std::initializer_list>>> expected, const LiteralSlice& actual, const ErrorSpec& error) { - EXPECT_TRUE(Near(*Literal::CreateR4(expected), actual, error)); + EXPECT_TRUE(Near(*LiteralUtil::CreateR4(expected), actual, error)); } template /* static */ void LiteralTestUtil::ExpectR2NearArray2D( const Array2D& expected, const LiteralSlice& actual, const ErrorSpec& error) { - EXPECT_TRUE(Near(*Literal::CreateR2FromArray2D(expected), actual, error)); + EXPECT_TRUE(Near(*LiteralUtil::CreateR2FromArray2D(expected), actual, error)); } template /* static */ void LiteralTestUtil::ExpectR3NearArray3D( const Array3D& expected, const LiteralSlice& actual, const ErrorSpec& error) { - EXPECT_TRUE(Near(*Literal::CreateR3FromArray3D(expected), actual, error)); + EXPECT_TRUE(Near(*LiteralUtil::CreateR3FromArray3D(expected), actual, error)); } template /* static */ void LiteralTestUtil::ExpectR4NearArray4D( const Array4D& expected, const LiteralSlice& actual, const ErrorSpec& error) { - EXPECT_TRUE(Near(*Literal::CreateR4FromArray4D(expected), actual, error)); + EXPECT_TRUE(Near(*LiteralUtil::CreateR4FromArray4D(expected), actual, error)); } } // namespace xla diff --git a/tensorflow/compiler/xla/tests/literal_test_util_test.cc b/tensorflow/compiler/xla/tests/literal_test_util_test.cc index bbac7285aefbb1f028fad152e4b7fe6af01e9f6d..f297b2b847f570d26e71ddcd8e34bc626f982e1f 100644 --- a/tensorflow/compiler/xla/tests/literal_test_util_test.cc +++ b/tensorflow/compiler/xla/tests/literal_test_util_test.cc @@ -31,8 +31,9 @@ namespace xla { namespace { TEST(LiteralTestUtilTest, ComparesEqualTuplesEqual) { - std::unique_ptr literal = Literal::MakeTuple({ - Literal::CreateR0(42).get(), Literal::CreateR0(64).get(), + std::unique_ptr literal = LiteralUtil::MakeTuple({ + LiteralUtil::CreateR0(42).get(), + LiteralUtil::CreateR0(64).get(), }); EXPECT_TRUE(LiteralTestUtil::Equal(*literal, *literal)); } @@ -42,11 +43,13 @@ TEST(LiteralTestUtilTest, ComparesUnequalTuplesUnequal) { // un-fail an assertion failure. The CHECK-failure is death, so we can make a // death assertion. auto unequal_things_are_equal = [] { - std::unique_ptr lhs = Literal::MakeTuple({ - Literal::CreateR0(42).get(), Literal::CreateR0(64).get(), + std::unique_ptr lhs = LiteralUtil::MakeTuple({ + LiteralUtil::CreateR0(42).get(), + LiteralUtil::CreateR0(64).get(), }); - std::unique_ptr rhs = Literal::MakeTuple({ - Literal::CreateR0(64).get(), Literal::CreateR0(42).get(), + std::unique_ptr rhs = LiteralUtil::MakeTuple({ + LiteralUtil::CreateR0(64).get(), + LiteralUtil::CreateR0(42).get(), }); CHECK(LiteralTestUtil::Equal(*lhs, *rhs)) << "LHS and RHS are unequal"; }; @@ -55,8 +58,8 @@ TEST(LiteralTestUtilTest, ComparesUnequalTuplesUnequal) { TEST(LiteralTestUtilTest, ExpectNearFailurePlacesResultsInTemporaryDirectory) { auto dummy_lambda = [] { - auto two = Literal::CreateR0(2); - auto four = Literal::CreateR0(4); + auto two = LiteralUtil::CreateR0(2); + auto four = LiteralUtil::CreateR0(4); ErrorSpec error(0.001); CHECK(LiteralTestUtil::Near(*two, *four, error)) << "two is not near four"; }; @@ -98,8 +101,8 @@ TEST(LiteralTestUtilTest, ExpectNearFailurePlacesResultsInTemporaryDirectory) { } TEST(LiteralTestUtilTest, NotEqualHasValuesInMessage) { - auto expected = Literal::CreateR1({1, 2, 3}); - auto actual = Literal::CreateR1({4, 5, 6}); + auto expected = LiteralUtil::CreateR1({1, 2, 3}); + auto actual = LiteralUtil::CreateR1({4, 5, 6}); ::testing::AssertionResult result = LiteralTestUtil::Equal(*expected, *actual); EXPECT_THAT(result.message(), ::testing::HasSubstr("expected: {1, 2, 3}")); @@ -107,25 +110,26 @@ TEST(LiteralTestUtilTest, NotEqualHasValuesInMessage) { } TEST(LiteralTestUtilTest, NearComparatorR1) { - auto a = - Literal::CreateR1({0.0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8}); - auto b = - Literal::CreateR1({0.0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8}); + auto a = LiteralUtil::CreateR1( + {0.0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8}); + auto b = LiteralUtil::CreateR1( + {0.0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8}); EXPECT_TRUE(LiteralTestUtil::Near(*a, *b, ErrorSpec{0.0001})); } TEST(LiteralTestUtilTest, NearComparatorR1Nan) { - auto a = - Literal::CreateR1({0.0, 0.1, 0.2, 0.3, NAN, 0.5, 0.6, 0.7, 0.8}); - auto b = - Literal::CreateR1({0.0, 0.1, 0.2, 0.3, NAN, 0.5, 0.6, 0.7, 0.8}); + auto a = LiteralUtil::CreateR1( + {0.0, 0.1, 0.2, 0.3, NAN, 0.5, 0.6, 0.7, 0.8}); + auto b = LiteralUtil::CreateR1( + {0.0, 0.1, 0.2, 0.3, NAN, 0.5, 0.6, 0.7, 0.8}); EXPECT_TRUE(LiteralTestUtil::Near(*a, *b, ErrorSpec{0.0001})); } TEST(LiteralTestUtil, NearComparatorDifferentLengths) { - auto a = - Literal::CreateR1({0.0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8}); - auto b = Literal::CreateR1({0.0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7}); + auto a = LiteralUtil::CreateR1( + {0.0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8}); + auto b = + LiteralUtil::CreateR1({0.0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7}); EXPECT_FALSE(LiteralTestUtil::Near(*a, *b, ErrorSpec{0.0001})); EXPECT_FALSE(LiteralTestUtil::Near(*b, *a, ErrorSpec{0.0001})); } diff --git a/tensorflow/compiler/xla/tests/llvm_compiler_test.cc b/tensorflow/compiler/xla/tests/llvm_compiler_test.cc index 082bc34136e004795ce300c66591758f47c665fe..13df83fffff0851dda7615dfa48f5629ed6f22d0 100644 --- a/tensorflow/compiler/xla/tests/llvm_compiler_test.cc +++ b/tensorflow/compiler/xla/tests/llvm_compiler_test.cc @@ -14,6 +14,7 @@ limitations under the License. ==============================================================================*/ #include "tensorflow/compiler/xla/service/llvm_compiler.h" +#include "tensorflow/compiler/xla/literal_util.h" #include "tensorflow/compiler/xla/service/backend.h" #include "tensorflow/compiler/xla/service/cpu/cpu_compiler.h" #include "tensorflow/compiler/xla/service/gpu/gpu_compiler.h" @@ -64,7 +65,7 @@ class LLVMCompilerTest : public ::testing::Test { // Create HLO module, and run the compiler. auto builder = HloComputation::Builder(TestName()); builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(42.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(42.0))); auto hlo_module = CreateNewModule(); hlo_module->AddEntryComputation(builder.Build()); @@ -86,7 +87,7 @@ class LLVMCompilerTest : public ::testing::Test { void TestMultiModuleCompilation(LLVMCompiler *compiler) { HloComputation::Builder builder(TestName()); builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(42.0))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(42.0))); std::unique_ptr hlo_module = CreateNewModule(); hlo_module->AddEntryComputation(builder.Build()); diff --git a/tensorflow/compiler/xla/tests/llvm_irgen_test_base.cc b/tensorflow/compiler/xla/tests/llvm_irgen_test_base.cc index 2c45f19c090d2690878430363bf0d20252b2f3df..6fc11150978931f980349799372872f9fb68f292 100644 --- a/tensorflow/compiler/xla/tests/llvm_irgen_test_base.cc +++ b/tensorflow/compiler/xla/tests/llvm_irgen_test_base.cc @@ -18,6 +18,7 @@ limitations under the License. #include #include +#include "tensorflow/compiler/xla/service/hlo_parser.h" #include "tensorflow/compiler/xla/service/llvm_ir/llvm_util.h" #include "tensorflow/compiler/xla/tests/filecheck.h" #include "tensorflow/core/lib/core/status_test_util.h" @@ -25,28 +26,28 @@ limitations under the License. namespace xla { -void LLVMIRGenTestBase::SetIrHook(bool match_optimized_ir) { +void LlvmIrGenTestBase::SetIrHook(bool match_optimized_ir) { auto llvm_compiler = GetLLVMCompiler(); using std::placeholders::_1; // Add the IR inspection hook to the LLVM compiler. if (match_optimized_ir) { llvm_compiler->SetPostOptimizationHook( - std::bind(&LLVMIRGenTestBase::IrHook, this, _1)); + std::bind(&LlvmIrGenTestBase::IrHook, this, _1)); } else { llvm_compiler->SetPreOptimizationHook( - std::bind(&LLVMIRGenTestBase::IrHook, this, _1)); + std::bind(&LlvmIrGenTestBase::IrHook, this, _1)); } } -void LLVMIRGenTestBase::ResetIrHook() { +void LlvmIrGenTestBase::ResetIrHook() { auto llvm_compiler = GetLLVMCompiler(); llvm_compiler->RemovePreOptimizationHook(); llvm_compiler->RemovePostOptimizationHook(); } -void LLVMIRGenTestBase::CompileAndVerifyIr( +void LlvmIrGenTestBase::CompileAndVerifyIr( std::unique_ptr hlo_module, const string& pattern, bool match_optimized_ir) { SetIrHook(match_optimized_ir); @@ -58,7 +59,17 @@ void LLVMIRGenTestBase::CompileAndVerifyIr( EXPECT_TRUE(filecheck_result.ValueOrDie()); } -void LLVMIRGenTestBase::CompileAheadOfTimeAndVerifyIr( +void LlvmIrGenTestBase::CompileAndVerifyIr(const string& hlo_text, + const string& expected_llvm_ir, + bool match_optimized_ir) { + HloModuleConfig config; + config.set_debug_options(GetDebugOptionsForTest()); + TF_ASSERT_OK_AND_ASSIGN(std::unique_ptr module, + ParseHloString(hlo_text, config)); + CompileAndVerifyIr(std::move(module), expected_llvm_ir, match_optimized_ir); +} + +void LlvmIrGenTestBase::CompileAheadOfTimeAndVerifyIr( std::unique_ptr hlo_module, const AotCompilationOptions& options, const string& pattern, bool match_optimized_ir) { SetIrHook(match_optimized_ir); @@ -71,11 +82,11 @@ void LLVMIRGenTestBase::CompileAheadOfTimeAndVerifyIr( EXPECT_TRUE(filecheck_result.ValueOrDie()); } -LLVMCompiler* LLVMIRGenTestBase::GetLLVMCompiler() { +LLVMCompiler* LlvmIrGenTestBase::GetLLVMCompiler() { return static_cast(backend().compiler()); } -Status LLVMIRGenTestBase::IrHook(const llvm::Module& module) { +Status LlvmIrGenTestBase::IrHook(const llvm::Module& module) { ir_ = llvm_ir::DumpModuleToString(module); return Status::OK(); } diff --git a/tensorflow/compiler/xla/tests/llvm_irgen_test_base.h b/tensorflow/compiler/xla/tests/llvm_irgen_test_base.h index 74cbb5f5df662992046a5b0f9a31e52879f375ad..018f9546afc3e408686a9ac75a74320a05b27182 100644 --- a/tensorflow/compiler/xla/tests/llvm_irgen_test_base.h +++ b/tensorflow/compiler/xla/tests/llvm_irgen_test_base.h @@ -24,7 +24,7 @@ limitations under the License. namespace xla { // Tests that verify IR emitted by the CPU/GPU backend is as expected. -class LLVMIRGenTestBase : public CodegenTestBase { +class LlvmIrGenTestBase : public CodegenTestBase { protected: // Compiles the given HLO module to LLVM IR and verifies the IR matches the // given pattern. `pattern` is in the FileCheck pattern matching syntax @@ -38,6 +38,12 @@ class LLVMIRGenTestBase : public CodegenTestBase { void CompileAndVerifyIr(std::unique_ptr hlo_module, const string& pattern, bool match_optimized_ir); + // A thin wrapper around CompileAndVerifyIr that parses `hlo_text` to create + // an HLO module. + void CompileAndVerifyIr(const string& hlo_text, + const string& expected_llvm_ir, + bool match_optimized_ir = false); + // Compiles the given HLO module to LLVM IR and verifies the IR matches the // given pattern. `pattern` is in the FileCheck pattern matching syntax // (http://llvm.org/docs/CommandGuide/FileCheck.html). diff --git a/tensorflow/compiler/xla/tests/local_client_allocation_test.cc b/tensorflow/compiler/xla/tests/local_client_allocation_test.cc index f21f83992ffb7c07dff31c68a7e9e3f7944bf512..0df50150aee69749beea79ff522fb6f820d1945d 100644 --- a/tensorflow/compiler/xla/tests/local_client_allocation_test.cc +++ b/tensorflow/compiler/xla/tests/local_client_allocation_test.cc @@ -17,7 +17,7 @@ limitations under the License. #include "tensorflow/compiler/xla/client/local_client.h" #include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/service/local_service.h" #include "tensorflow/compiler/xla/service/shaped_buffer.h" #include "tensorflow/compiler/xla/statusor.h" @@ -38,14 +38,14 @@ class LocalClientAllocationTest : public LocalClientTestBase { XLA_TEST_F(LocalClientAllocationTest, AddVectors) { XlaBuilder builder(TestName()); - auto x = builder.ConstantR1({0.0f, 1.0f, 2.0f}); - auto y = builder.ConstantR1({2.0f, 3.0f, 4.0f}); - builder.Add(x, y); + auto x = ConstantR1(&builder, {0.0f, 1.0f, 2.0f}); + auto y = ConstantR1(&builder, {2.0f, 3.0f, 4.0f}); + Add(x, y); TestAllocator* allocator = GetOrCreateAllocator(local_client_->platform()); auto x_array = - LiteralToShapedBuffer(*Literal::CreateR1({0.0f, 1.0f, 2.0f})); + LiteralToShapedBuffer(*LiteralUtil::CreateR1({0.0f, 1.0f, 2.0f})); int64 allocation_count_before = allocator_->allocation_count(); @@ -74,9 +74,9 @@ XLA_TEST_F(LocalClientAllocationTest, RunOnDevices) { // Run a computation on every device on the system. Verify that allocation // occurs on the proper device. XlaBuilder builder(TestName()); - auto x = builder.ConstantR1({0.0f, 1.0f, 2.0f}); - auto y = builder.ConstantR1({2.0f, 3.0f, 4.0f}); - builder.Add(x, y); + auto x = ConstantR1(&builder, {0.0f, 1.0f, 2.0f}); + auto y = ConstantR1(&builder, {2.0f, 3.0f, 4.0f}); + Add(x, y); auto computation = builder.Build().ConsumeValueOrDie(); TestAllocator* allocator = GetOrCreateAllocator(local_client_->platform()); diff --git a/tensorflow/compiler/xla/tests/local_client_aot_test_helper.cc b/tensorflow/compiler/xla/tests/local_client_aot_test_helper.cc index a366afe8262e1f537b225e395bba9cb2fc22683a..70612e7c49d2815096cc54fd6ae796148249b4db 100644 --- a/tensorflow/compiler/xla/tests/local_client_aot_test_helper.cc +++ b/tensorflow/compiler/xla/tests/local_client_aot_test_helper.cc @@ -37,8 +37,8 @@ using xla::string; xla::XlaComputation Doubler() { xla::XlaBuilder builder("doubler"); auto r0f32 = xla::ShapeUtil::MakeShape(xla::F32, {}); - auto x = builder.Parameter(0, r0f32, "x"); - builder.Mul(x, builder.ConstantR0(2.0)); + auto x = xla::Parameter(&builder, 0, r0f32, "x"); + xla::Mul(x, xla::ConstantR0(&builder, 2.0)); return std::move(builder.Build().ValueOrDie()); } @@ -51,10 +51,10 @@ int main(int argc, char** argv) { xla::XlaBuilder builder("aot_test_helper"); auto opaque_shape = xla::ShapeUtil::MakeOpaqueShape(); - auto opaque_param = builder.Parameter(0, opaque_shape, "x"); + auto opaque_param = Parameter(&builder, 0, opaque_shape, "x"); auto r0f32 = xla::ShapeUtil::MakeShape(xla::F32, {}); - auto sum = builder.CustomCall("SumStructElements", {opaque_param}, r0f32); - builder.Call(Doubler(), {sum}); + auto sum = CustomCall(&builder, "SumStructElements", {opaque_param}, r0f32); + Call(&builder, Doubler(), {sum}); if (argc != 2) { LOG(FATAL) << "local_client_aot_test_helper TARGET_CPU"; diff --git a/tensorflow/compiler/xla/tests/local_client_execute_test.cc b/tensorflow/compiler/xla/tests/local_client_execute_test.cc index 5a70c2a9ae5f32f27ec012d554c183159a63576c..2f4d197ae632c08cb80b5d09ab4918f018e992ef 100644 --- a/tensorflow/compiler/xla/tests/local_client_execute_test.cc +++ b/tensorflow/compiler/xla/tests/local_client_execute_test.cc @@ -21,7 +21,7 @@ limitations under the License. #include "tensorflow/compiler/xla/client/local_client.h" #include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/compiler/xla/layout_util.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/service/device_memory_allocator.h" #include "tensorflow/compiler/xla/service/local_service.h" #include "tensorflow/compiler/xla/service/platform_util.h" @@ -54,7 +54,7 @@ class LocalClientExecuteTest : public LocalClientTestBase { XLA_TEST_F(LocalClientExecuteTest, Constant) { XlaBuilder builder(TestName()); - auto y = builder.ConstantR0(123.0f); + ConstantR0(&builder, 123.0f); ScopedShapedBuffer result = ExecuteLocallyOrDie(builder.Build().ValueOrDie(), {}); @@ -64,11 +64,11 @@ XLA_TEST_F(LocalClientExecuteTest, Constant) { XLA_TEST_F(LocalClientExecuteTest, AddScalars) { XlaBuilder builder(TestName()); - auto x = builder.Parameter(0, ShapeUtil::MakeShape(F32, {}), "x"); - auto y = builder.ConstantR0(123.0f); - builder.Add(x, y); + auto x = Parameter(&builder, 0, ShapeUtil::MakeShape(F32, {}), "x"); + auto y = ConstantR0(&builder, 123.0f); + Add(x, y); - auto x_value = LiteralToShapedBuffer(*Literal::CreateR0(42.0f)); + auto x_value = LiteralToShapedBuffer(*LiteralUtil::CreateR0(42.0f)); ScopedShapedBuffer result = ExecuteLocallyOrDie(builder.Build().ValueOrDie(), {&x_value}); LiteralTestUtil::ExpectR0Near(165.f, *ShapedBufferToLiteral(result), @@ -77,11 +77,11 @@ XLA_TEST_F(LocalClientExecuteTest, AddScalars) { XLA_TEST_F(LocalClientExecuteTest, AddZeroElementVectors) { XlaBuilder builder(TestName()); - auto x = builder.Parameter(0, ShapeUtil::MakeShape(F32, {0}), "x"); - auto y = builder.ConstantR1({}); - builder.Add(x, y); + auto x = Parameter(&builder, 0, ShapeUtil::MakeShape(F32, {0}), "x"); + auto y = ConstantR1(&builder, {}); + Add(x, y); - auto x_array = LiteralToShapedBuffer(*Literal::CreateR1({})); + auto x_array = LiteralToShapedBuffer(*LiteralUtil::CreateR1({})); ScopedShapedBuffer result = ExecuteLocallyOrDie(builder.Build().ValueOrDie(), {&x_array}); LiteralTestUtil::ExpectR1Near({}, *ShapedBufferToLiteral(result), @@ -90,12 +90,12 @@ XLA_TEST_F(LocalClientExecuteTest, AddZeroElementVectors) { XLA_TEST_F(LocalClientExecuteTest, AddVectors) { XlaBuilder builder(TestName()); - auto x = builder.Parameter(0, ShapeUtil::MakeShape(F32, {3}), "x"); - auto y = builder.ConstantR1({2.0f, 3.0f, 4.0f}); - builder.Add(x, y); + auto x = Parameter(&builder, 0, ShapeUtil::MakeShape(F32, {3}), "x"); + auto y = ConstantR1(&builder, {2.0f, 3.0f, 4.0f}); + Add(x, y); auto x_array = - LiteralToShapedBuffer(*Literal::CreateR1({0.0f, 1.0f, 2.0f})); + LiteralToShapedBuffer(*LiteralUtil::CreateR1({0.0f, 1.0f, 2.0f})); ScopedShapedBuffer result = ExecuteLocallyOrDie(builder.Build().ValueOrDie(), {&x_array}); LiteralTestUtil::ExpectR1Near( @@ -104,12 +104,12 @@ XLA_TEST_F(LocalClientExecuteTest, AddVectors) { XLA_TEST_F(LocalClientExecuteTest, AddVectorsWithProfile) { XlaBuilder builder(TestName()); - auto x = builder.Parameter(0, ShapeUtil::MakeShape(F32, {3}), "x"); - auto y = builder.ConstantR1({2.0f, 3.0f, 4.0f}); - builder.Add(x, y); + auto x = Parameter(&builder, 0, ShapeUtil::MakeShape(F32, {3}), "x"); + auto y = ConstantR1(&builder, {2.0f, 3.0f, 4.0f}); + Add(x, y); auto x_array = - LiteralToShapedBuffer(*Literal::CreateR1({0.0f, 1.0f, 2.0f})); + LiteralToShapedBuffer(*LiteralUtil::CreateR1({0.0f, 1.0f, 2.0f})); ExecutionProfile profile; ScopedShapedBuffer result = ExecuteLocallyOrDie( builder.Build().ValueOrDie(), {&x_array}, DefaultExecutableBuildOptions(), @@ -122,19 +122,19 @@ XLA_TEST_F(LocalClientExecuteTest, AddVectorsWithProfile) { XLA_TEST_F(LocalClientExecuteTest, AddArraysWithDifferentInputLayouts) { XlaBuilder builder(TestName()); - auto x = builder.Parameter(0, ShapeUtil::MakeShape(F32, {2, 2}), "x"); - auto y = builder.Parameter(1, ShapeUtil::MakeShape(F32, {2, 2}), "y"); - builder.Add(x, y); + auto x = Parameter(&builder, 0, ShapeUtil::MakeShape(F32, {2, 2}), "x"); + auto y = Parameter(&builder, 1, ShapeUtil::MakeShape(F32, {2, 2}), "y"); + Add(x, y); auto computation = builder.Build().ConsumeValueOrDie(); // Create x as a col-major array. - auto x_array = LiteralToShapedBuffer(*Literal::CreateR2WithLayout( + auto x_array = LiteralToShapedBuffer(*LiteralUtil::CreateR2WithLayout( {{1.0f, 2.0f}, {3.0f, 4.0f}}, LayoutUtil::MakeLayout({0, 1}))); EXPECT_TRUE(LayoutUtil::Equal(x_array.on_device_shape().layout(), LayoutUtil::MakeLayout({0, 1}))); // Create y as a row-major array. - auto y_array = LiteralToShapedBuffer(*Literal::CreateR2WithLayout( + auto y_array = LiteralToShapedBuffer(*LiteralUtil::CreateR2WithLayout( {{10.0f, 20.0f}, {30.0f, 40.0f}}, LayoutUtil::MakeLayout({1, 0}))); EXPECT_TRUE(LayoutUtil::Equal(y_array.on_device_shape().layout(), LayoutUtil::MakeLayout({1, 0}))); @@ -155,15 +155,15 @@ XLA_TEST_F(LocalClientExecuteTest, AddArraysWithDifferentInputLayouts) { XLA_TEST_F(LocalClientExecuteTest, AddArraysWithDifferentOutputLayouts) { XlaBuilder builder(TestName()); - auto x = builder.Parameter(0, ShapeUtil::MakeShape(F32, {2, 2}), "x"); - auto y = builder.Parameter(1, ShapeUtil::MakeShape(F32, {2, 2}), "y"); - builder.Add(x, y); + auto x = Parameter(&builder, 0, ShapeUtil::MakeShape(F32, {2, 2}), "x"); + auto y = Parameter(&builder, 1, ShapeUtil::MakeShape(F32, {2, 2}), "y"); + Add(x, y); auto computation = builder.Build().ConsumeValueOrDie(); auto x_array = LiteralToShapedBuffer( - *Literal::CreateR2({{1.0f, 2.0f}, {3.0f, 4.0f}})); + *LiteralUtil::CreateR2({{1.0f, 2.0f}, {3.0f, 4.0f}})); auto y_array = LiteralToShapedBuffer( - *Literal::CreateR2({{10.0f, 20.0f}, {30.0f, 40.0f}})); + *LiteralUtil::CreateR2({{10.0f, 20.0f}, {30.0f, 40.0f}})); // Run with col-major result layout. ScopedShapedBuffer result_colmaj = ExecuteLocallyOrDie( @@ -192,15 +192,15 @@ XLA_TEST_F(LocalClientExecuteTest, AddArraysWithDifferentOutputLayouts) { XLA_TEST_F(LocalClientExecuteTest, TupleResult) { XlaBuilder builder(TestName()); - auto x = builder.Parameter(0, ShapeUtil::MakeShape(F32, {2, 2}), "x"); - auto y = builder.Parameter(1, ShapeUtil::MakeShape(F32, {2, 2}), "y"); - builder.Tuple({x, y, x}); + auto x = Parameter(&builder, 0, ShapeUtil::MakeShape(F32, {2, 2}), "x"); + auto y = Parameter(&builder, 1, ShapeUtil::MakeShape(F32, {2, 2}), "y"); + Tuple(&builder, {x, y, x}); auto computation = builder.Build().ConsumeValueOrDie(); auto x_array = LiteralToShapedBuffer( - *Literal::CreateR2({{1.0f, 2.0f}, {3.0f, 4.0f}})); + *LiteralUtil::CreateR2({{1.0f, 2.0f}, {3.0f, 4.0f}})); auto y_array = LiteralToShapedBuffer( - *Literal::CreateR2({{10.0f, 20.0f}, {30.0f, 40.0f}})); + *LiteralUtil::CreateR2({{10.0f, 20.0f}, {30.0f, 40.0f}})); ScopedShapedBuffer result = ExecuteLocallyOrDie(computation, {&x_array, &y_array}); @@ -219,16 +219,16 @@ XLA_TEST_F(LocalClientExecuteTest, TupleResult) { XLA_TEST_F(LocalClientExecuteTest, NestedTupleResult) { XlaBuilder builder(TestName()); - auto x = builder.Parameter(0, ShapeUtil::MakeShape(F32, {2, 2}), "x"); - auto y = builder.Parameter(1, ShapeUtil::MakeShape(F32, {2, 2}), "y"); - auto inner_tuple = builder.Tuple({x, y, x}); - builder.Tuple({inner_tuple, x}); + auto x = Parameter(&builder, 0, ShapeUtil::MakeShape(F32, {2, 2}), "x"); + auto y = Parameter(&builder, 1, ShapeUtil::MakeShape(F32, {2, 2}), "y"); + auto inner_tuple = Tuple(&builder, {x, y, x}); + Tuple(&builder, {inner_tuple, x}); auto computation = builder.Build().ConsumeValueOrDie(); auto x_array = LiteralToShapedBuffer( - *Literal::CreateR2({{1.0f, 2.0f}, {3.0f, 4.0f}})); + *LiteralUtil::CreateR2({{1.0f, 2.0f}, {3.0f, 4.0f}})); auto y_array = LiteralToShapedBuffer( - *Literal::CreateR2({{10.0f, 20.0f}, {30.0f, 40.0f}})); + *LiteralUtil::CreateR2({{10.0f, 20.0f}, {30.0f, 40.0f}})); ScopedShapedBuffer result = ExecuteLocallyOrDie(computation, {&x_array, &y_array}); @@ -250,12 +250,12 @@ XLA_TEST_F(LocalClientExecuteTest, NestedTupleResult) { XLA_TEST_F(LocalClientExecuteTest, TupleResultWithLayout) { // Verify setting the result layout of a computation with a tuple output. XlaBuilder builder(TestName()); - auto x = builder.Parameter(0, ShapeUtil::MakeShape(F32, {2, 2}), "x"); - auto y = builder.Parameter(1, ShapeUtil::MakeShape(F32, {2, 2}), "y"); - builder.Tuple({x, y}); + auto x = Parameter(&builder, 0, ShapeUtil::MakeShape(F32, {2, 2}), "x"); + auto y = Parameter(&builder, 1, ShapeUtil::MakeShape(F32, {2, 2}), "y"); + Tuple(&builder, {x, y}); auto array = LiteralToShapedBuffer( - *Literal::CreateR2({{1.0f, 2.0f}, {3.0f, 4.0f}})); + *LiteralUtil::CreateR2({{1.0f, 2.0f}, {3.0f, 4.0f}})); ExecutableBuildOptions options = DefaultExecutableBuildOptions(); Shape shape_with_layout = ShapeUtil::MakeTupleShape( @@ -287,23 +287,23 @@ XLA_TEST_F(LocalClientExecuteTest, TupleArguments) { // Computation adds the respective array and vector elements from each tuple // argument and returns the results as a tuple. XlaBuilder builder(TestName()); - auto x = builder.Parameter(0, tuple_shape0, "x"); - auto y = builder.Parameter(1, tuple_shape1, "y"); - auto x_0 = builder.GetTupleElement(x, 0); - auto x_1 = builder.GetTupleElement(x, 1); - auto y_0 = builder.GetTupleElement(y, 0); - auto y_1 = builder.GetTupleElement(y, 1); - auto array_sum = builder.Add(x_0, y_1); - auto vector_diff = builder.Sub(x_1, y_0); - builder.Tuple({array_sum, vector_diff}); + auto x = Parameter(&builder, 0, tuple_shape0, "x"); + auto y = Parameter(&builder, 1, tuple_shape1, "y"); + auto x_0 = GetTupleElement(x, 0); + auto x_1 = GetTupleElement(x, 1); + auto y_0 = GetTupleElement(y, 0); + auto y_1 = GetTupleElement(y, 1); + auto array_sum = Add(x_0, y_1); + auto vector_diff = Sub(x_1, y_0); + Tuple(&builder, {array_sum, vector_diff}); auto computation = builder.Build().ConsumeValueOrDie(); - auto x_literal = Literal::MakeTuple( - {Literal::CreateR2({{1.0, 2.0}, {3.0, 4.0}}).get(), - Literal::CreateR1({42.0, 75.0, 123.0}).get()}); - auto y_literal = Literal::MakeTuple( - {Literal::CreateR1({2.0, 4.0, 6.0}).get(), - Literal::CreateR2({{55.0, 44.0}, {33.0, 22.0}}).get()}); + auto x_literal = LiteralUtil::MakeTuple( + {LiteralUtil::CreateR2({{1.0, 2.0}, {3.0, 4.0}}).get(), + LiteralUtil::CreateR1({42.0, 75.0, 123.0}).get()}); + auto y_literal = LiteralUtil::MakeTuple( + {LiteralUtil::CreateR1({2.0, 4.0, 6.0}).get(), + LiteralUtil::CreateR2({{55.0, 44.0}, {33.0, 22.0}}).get()}); auto x_buffer = LiteralToShapedBuffer(*x_literal); auto y_buffer = LiteralToShapedBuffer(*y_literal); @@ -333,23 +333,23 @@ XLA_TEST_F(LocalClientExecuteTest, NestedTupleArgument) { // Computation negates the array element and sums the two vector elements in // the nested tuple. The resulting array and vector are returned as a tuple. XlaBuilder builder(TestName()); - auto param = builder.Parameter(0, nested_tuple_shape, "param"); - auto inner_tuple = builder.GetTupleElement(param, 0); - auto inner_array = builder.GetTupleElement(inner_tuple, 0); - auto inner_vector = builder.GetTupleElement(inner_tuple, 1); - auto outer_vector = builder.GetTupleElement(param, 1); - - auto negate_array = builder.Neg(inner_array); - auto vector_sum = builder.Add(inner_vector, outer_vector); - builder.Tuple({negate_array, vector_sum}); + auto param = Parameter(&builder, 0, nested_tuple_shape, "param"); + auto inner_tuple = GetTupleElement(param, 0); + auto inner_array = GetTupleElement(inner_tuple, 0); + auto inner_vector = GetTupleElement(inner_tuple, 1); + auto outer_vector = GetTupleElement(param, 1); + + auto negate_array = Neg(inner_array); + auto vector_sum = Add(inner_vector, outer_vector); + Tuple(&builder, {negate_array, vector_sum}); auto computation = builder.Build().ConsumeValueOrDie(); - auto arg_literal = Literal::MakeTuple( - {Literal::MakeTuple( - {Literal::CreateR2({{1.0, 2.0}, {3.0, 4.0}}).get(), - Literal::CreateR1({42.0, 75.0, 123.0}).get()}) + auto arg_literal = LiteralUtil::MakeTuple( + {LiteralUtil::MakeTuple( + {LiteralUtil::CreateR2({{1.0, 2.0}, {3.0, 4.0}}).get(), + LiteralUtil::CreateR1({42.0, 75.0, 123.0}).get()}) .get(), - Literal::CreateR1({222.0, -2.0, 10.0}).get()}); + LiteralUtil::CreateR1({222.0, -2.0, 10.0}).get()}); auto arg_buffer = LiteralToShapedBuffer(*arg_literal); ScopedShapedBuffer result = ExecuteLocallyOrDie(computation, {&arg_buffer}); @@ -371,15 +371,15 @@ XLA_TEST_F(LocalClientExecuteTest, PassingTupleResultBackIntoComputation) { ShapeUtil::MakeTupleShape({array_shape, array_shape}); XlaBuilder builder(TestName()); - auto param = builder.Parameter(0, tuple_shape, "param"); - auto element_0 = builder.GetTupleElement(param, 0); - auto element_1 = builder.GetTupleElement(param, 1); - builder.Tuple({builder.Neg(element_0), builder.Add(element_1, element_1)}); + auto param = Parameter(&builder, 0, tuple_shape, "param"); + auto element_0 = GetTupleElement(param, 0); + auto element_1 = GetTupleElement(param, 1); + Tuple(&builder, {Neg(element_0), Add(element_1, element_1)}); auto computation = builder.Build().ConsumeValueOrDie(); - auto arg_literal = Literal::MakeTuple( - {Literal::CreateR2({{1.0, 2.0}, {3.0, 4.0}}).get(), - Literal::CreateR2({{11.0, 3.0}, {4.0, 5.0}}).get()}); + auto arg_literal = LiteralUtil::MakeTuple( + {LiteralUtil::CreateR2({{1.0, 2.0}, {3.0, 4.0}}).get(), + LiteralUtil::CreateR2({{11.0, 3.0}, {4.0, 5.0}}).get()}); auto arg_buffer = LiteralToShapedBuffer(*arg_literal); ScopedShapedBuffer result_0 = ExecuteLocallyOrDie(computation, {&arg_buffer}); @@ -414,26 +414,25 @@ XLA_TEST_F(LocalClientExecuteTest, LargeTuple) { const Shape tuple_shape = ShapeUtil::MakeTupleShape(element_shapes); XlaBuilder builder(TestName()); - auto param = builder.Parameter(0, tuple_shape, "param"); + auto param = Parameter(&builder, 0, tuple_shape, "param"); // Add each element's tuple index value to every element. std::vector result_elements; for (int i = 0; i < kElementCount; ++i) { - auto element = builder.GetTupleElement(param, i); - result_elements.push_back( - builder.Add(element, builder.ConstantR0(i))); + auto element = GetTupleElement(param, i); + result_elements.push_back(Add(element, ConstantR0(&builder, i))); } - builder.Tuple(result_elements); + Tuple(&builder, result_elements); auto computation = builder.Build().ConsumeValueOrDie(); // Feed in a tuple where each two-element vector element is {tuple_index, // -tuple_index}. std::vector> arg_elements; for (int i = 0; i < kElementCount; ++i) { - arg_elements.push_back(Literal::CreateR1({1.0f * i, -1.0f * i})); + arg_elements.push_back(LiteralUtil::CreateR1({1.0f * i, -1.0f * i})); } std::unique_ptr arg_literal = - Literal::MakeTupleOwned(std::move(arg_elements)); + LiteralUtil::MakeTupleOwned(std::move(arg_elements)); auto arg_buffer = LiteralToShapedBuffer(*arg_literal); ScopedShapedBuffer result = ExecuteLocallyOrDie(computation, {&arg_buffer}); @@ -458,22 +457,22 @@ XLA_TEST_F(LocalClientExecuteTest, LargeNestedTuple) { const Shape tuple_shape = ShapeUtil::MakeTupleShape(inner_tuple_shapes); XlaBuilder builder(TestName()); - auto param = builder.Parameter(0, tuple_shape, "param"); + auto param = Parameter(&builder, 0, tuple_shape, "param"); // The computation increments each leaf value by an amount equal to the leaf's // ordinal position in a traversal of the tuple. std::vector result_elements; for (int i = 0; i < kFanout; ++i) { - auto outer_element = builder.GetTupleElement(param, i); + auto outer_element = GetTupleElement(param, i); std::vector inner_result_elements; for (int j = 0; j < kFanout; ++j) { - auto inner_element = builder.GetTupleElement(outer_element, j); - inner_result_elements.push_back(builder.Add( - inner_element, builder.ConstantR0(i * kFanout + j))); + auto inner_element = GetTupleElement(outer_element, j); + inner_result_elements.push_back( + Add(inner_element, ConstantR0(&builder, i * kFanout + j))); } - result_elements.push_back(builder.Tuple(inner_result_elements)); + result_elements.push_back(Tuple(&builder, inner_result_elements)); } - builder.Tuple(result_elements); + Tuple(&builder, result_elements); auto computation = builder.Build().ConsumeValueOrDie(); // Construct the argument to pass to the computation. @@ -481,12 +480,13 @@ XLA_TEST_F(LocalClientExecuteTest, LargeNestedTuple) { for (int i = 0; i < kFanout; ++i) { std::vector> inner_tuple_elements; for (int j = 0; j < kFanout; ++j) { - inner_tuple_elements.push_back(Literal::CreateR0(i + j)); + inner_tuple_elements.push_back(LiteralUtil::CreateR0(i + j)); } outer_tuple_elements.push_back( - Literal::MakeTupleOwned(std::move(inner_tuple_elements))); + LiteralUtil::MakeTupleOwned(std::move(inner_tuple_elements))); } - auto arg_literal = Literal::MakeTupleOwned(std::move(outer_tuple_elements)); + auto arg_literal = + LiteralUtil::MakeTupleOwned(std::move(outer_tuple_elements)); auto arg_buffer = LiteralToShapedBuffer(*arg_literal); ScopedShapedBuffer result = ExecuteLocallyOrDie(computation, {&arg_buffer}); @@ -513,23 +513,23 @@ XLA_TEST_F(LocalClientExecuteTest, DeepTuple) { } XlaBuilder builder(TestName()); - auto element = builder.Parameter(0, shape, "param"); + auto element = Parameter(&builder, 0, shape, "param"); for (int i = 0; i < kTupleDepth; ++i) { - element = builder.GetTupleElement(element, 0); + element = GetTupleElement(element, 0); } - auto output = builder.Add(element, builder.ConstantR0(42.0)); + auto output = Add(element, ConstantR0(&builder, 42.0)); for (int i = 0; i < kTupleDepth; ++i) { - output = builder.Tuple({output}); + output = Tuple(&builder, {output}); } auto computation = builder.Build().ConsumeValueOrDie(); // Construct the argument to pass to the computation. - std::unique_ptr arg_literal = Literal::CreateR0(123.0); + std::unique_ptr arg_literal = LiteralUtil::CreateR0(123.0); for (int i = 0; i < kTupleDepth; ++i) { std::vector> arg_vector; arg_vector.push_back(std::move(arg_literal)); - arg_literal = Literal::MakeTupleOwned(std::move(arg_vector)); + arg_literal = LiteralUtil::MakeTupleOwned(std::move(arg_vector)); } auto arg_buffer = LiteralToShapedBuffer(*arg_literal); @@ -547,12 +547,12 @@ XLA_TEST_F(LocalClientExecuteTest, DeepTuple) { XLA_TEST_F(LocalClientExecuteTest, InvalidNumberOfArguments) { // Test passing in an invalid number of arguments. XlaBuilder builder(TestName()); - auto x = builder.Parameter(0, ShapeUtil::MakeShape(F32, {3}), "x"); - auto y = builder.Parameter(1, ShapeUtil::MakeShape(F32, {3}), "y"); - builder.Add(x, y); + auto x = Parameter(&builder, 0, ShapeUtil::MakeShape(F32, {3}), "x"); + auto y = Parameter(&builder, 1, ShapeUtil::MakeShape(F32, {3}), "y"); + Add(x, y); auto x_array = - LiteralToShapedBuffer(*Literal::CreateR1({1.0f, 2.0f, 3.0f})); + LiteralToShapedBuffer(*LiteralUtil::CreateR1({1.0f, 2.0f, 3.0f})); auto execute_status = ExecuteLocally(builder.Build().ValueOrDie(), {&x_array}); @@ -564,11 +564,11 @@ XLA_TEST_F(LocalClientExecuteTest, InvalidNumberOfArguments) { XLA_TEST_F(LocalClientExecuteTest, IncorrectArgumentShape) { // Test passing in an argument with the wrong shape. XlaBuilder builder(TestName()); - auto x = builder.Parameter(0, ShapeUtil::MakeShape(F32, {3}), "x"); - builder.Neg(x); + auto x = Parameter(&builder, 0, ShapeUtil::MakeShape(F32, {3}), "x"); + Neg(x); auto x_array = LiteralToShapedBuffer( - *Literal::CreateR2({{0.0f, 1.0f}, {2.0f, 3.0f}})); + *LiteralUtil::CreateR2({{0.0f, 1.0f}, {2.0f, 3.0f}})); auto execute_status = ExecuteLocally(builder.Build().ValueOrDie(), {&x_array}); @@ -581,11 +581,11 @@ XLA_TEST_F(LocalClientExecuteTest, IncorrectArgumentShape) { XLA_TEST_F(LocalClientExecuteTest, InvalidResultLayout) { // Test passing in an invalid result layout parameter. XlaBuilder builder(TestName()); - auto x = builder.Parameter(0, ShapeUtil::MakeShape(F32, {2, 2}), "x"); - builder.Neg(x); + auto x = Parameter(&builder, 0, ShapeUtil::MakeShape(F32, {2, 2}), "x"); + Neg(x); auto x_array = LiteralToShapedBuffer( - *Literal::CreateR2({{0.0f, 1.0f}, {2.0f, 3.0f}})); + *LiteralUtil::CreateR2({{0.0f, 1.0f}, {2.0f, 3.0f}})); auto execute_status = ExecuteLocally( builder.Build().ValueOrDie(), {&x_array}, DefaultExecutableBuildOptions().set_result_layout( @@ -604,7 +604,7 @@ XLA_TEST_F(LocalClientExecuteTest, RunOnAllDeviceOrdinals) { // Try to run a trivial computation on every device on the system. If a // specific device is not supported, check that the right error is returned. XlaBuilder builder(TestName()); - builder.ConstantR0(42.0f); + ConstantR0(&builder, 42.0f); auto computation = builder.Build().ConsumeValueOrDie(); for (int d = 0; d < local_client_->device_count(); ++d) { if (!local_client_->device_ordinal_supported(d)) { @@ -631,7 +631,7 @@ XLA_TEST_F(LocalClientExecuteTest, InvalidDeviceOrdinalValues) { // Try running computations on devices with device ordinal values which do not // exist. XlaBuilder builder(TestName()); - builder.ConstantR0(42.0f); + ConstantR0(&builder, 42.0f); auto computation = builder.Build().ConsumeValueOrDie(); auto execute_status = @@ -648,7 +648,7 @@ XLA_TEST_F(LocalClientExecuteTest, InvalidDeviceOrdinalValues) { XLA_TEST_F(LocalClientExecuteTest, RunOnStream) { // Run a computation on a specific stream on each device on the system. XlaBuilder builder(TestName()); - builder.ConstantR0(42.0f); + ConstantR0(&builder, 42.0f); auto computation = builder.Build().ConsumeValueOrDie(); for (int d = 0; d < local_client_->device_count(); ++d) { @@ -684,7 +684,7 @@ XLA_TEST_F(LocalClientExecuteTest, wrong_stream.Init(); XlaBuilder builder(TestName()); - builder.ConstantR0(42.0f); + ConstantR0(&builder, 42.0f); auto execute_status = ExecuteLocally( builder.Build().ValueOrDie(), {}, DefaultExecutableBuildOptions(), DefaultExecutableRunOptions().set_stream(&wrong_stream)); @@ -701,7 +701,7 @@ XLA_TEST_F(LocalClientExecuteTest, TestAllocator allocator(wrong_platform); XlaBuilder builder(TestName()); - auto y = builder.ConstantR0(123.0f); + ConstantR0(&builder, 123.0f); auto execute_status = ExecuteLocally( builder.Build().ValueOrDie(), {}, DefaultExecutableBuildOptions(), @@ -714,7 +714,7 @@ XLA_TEST_F(LocalClientExecuteTest, XLA_TEST_F(LocalClientExecuteTest, RunOnUninitializedStream) { // Try to run a computation on a stream that has not been initialized. XlaBuilder builder(TestName()); - builder.ConstantR0(42.0f); + ConstantR0(&builder, 42.0f); LOG(INFO) << "default device = " << local_client_->default_device_ordinal(); se::StreamExecutor* executor = @@ -737,11 +737,11 @@ XLA_TEST_F(LocalClientExecuteTest, SelectBetweenTuples) { std::initializer_list vec1 = {1.f, 2.f, 3.f}; std::initializer_list vec2 = {2.f, 4.f, 6.f}; - auto tuple12 = builder.Tuple( - {builder.ConstantR1(vec1), builder.ConstantR1(vec2)}); - auto tuple21 = builder.Tuple( - {builder.ConstantR1(vec2), builder.ConstantR1(vec1)}); - builder.Select(builder.ConstantR0(false), tuple12, tuple21); + auto tuple12 = Tuple(&builder, {ConstantR1(&builder, vec1), + ConstantR1(&builder, vec2)}); + auto tuple21 = Tuple(&builder, {ConstantR1(&builder, vec2), + ConstantR1(&builder, vec1)}); + Select(ConstantR0(&builder, false), tuple12, tuple21); ScopedShapedBuffer result = ExecuteLocallyOrDie(builder.Build().ValueOrDie(), {}); @@ -754,9 +754,9 @@ XLA_TEST_F(LocalClientExecuteTest, SelectBetweenTuples) { XLA_TEST_F(LocalClientExecuteTest, CompileExecutable) { XlaBuilder builder(TestName()); - auto x = builder.Parameter(0, ShapeUtil::MakeShape(F32, {3}), "x"); - auto y = builder.ConstantR1({2.0f, 3.0f, 4.0f}); - builder.Add(x, y); + auto x = Parameter(&builder, 0, ShapeUtil::MakeShape(F32, {3}), "x"); + auto y = ConstantR1(&builder, {2.0f, 3.0f, 4.0f}); + Add(x, y); Shape argument_layout = ShapeUtil::MakeShapeWithLayout(F32, /*dimensions=*/{3}, {0}); @@ -768,7 +768,7 @@ XLA_TEST_F(LocalClientExecuteTest, CompileExecutable) { executable_status.ConsumeValueOrDie(); auto x_array = - LiteralToShapedBuffer(*Literal::CreateR1({0.0f, 1.0f, 2.0f})); + LiteralToShapedBuffer(*LiteralUtil::CreateR1({0.0f, 1.0f, 2.0f})); ScopedShapedBuffer result = executable->Run({&x_array}, DefaultExecutableRunOptions()) .ConsumeValueOrDie(); @@ -792,29 +792,29 @@ XLA_TEST_F(LocalClientExecuteTest, ShapeBufferToLiteralConversion) { }; // Array shapes. - test_to_device_and_back(*Literal::CreateR0(42.0)); - test_to_device_and_back(*Literal::CreateR0(true)); - test_to_device_and_back(*Literal::CreateR1({1.0, 42.0, 744.4})); + test_to_device_and_back(*LiteralUtil::CreateR0(42.0)); + test_to_device_and_back(*LiteralUtil::CreateR0(true)); + test_to_device_and_back(*LiteralUtil::CreateR1({1.0, 42.0, 744.4})); test_to_device_and_back( - *Literal::CreateR2({{1.0, 2.0, 3.0}, {44.0, 0.1, -3}})); - test_to_device_and_back(*Literal::CreateR2({{2, 1}, {4444, 56}})); + *LiteralUtil::CreateR2({{1.0, 2.0, 3.0}, {44.0, 0.1, -3}})); + test_to_device_and_back(*LiteralUtil::CreateR2({{2, 1}, {4444, 56}})); // Null shape (empty tuple). - test_to_device_and_back(*Literal::MakeTuple({})); + test_to_device_and_back(*LiteralUtil::MakeTuple({})); // Non-nested tuples. test_to_device_and_back( - *Literal::MakeTuple({Literal::CreateR0(12223.0).get()})); + *LiteralUtil::MakeTuple({LiteralUtil::CreateR0(12223.0).get()})); test_to_device_and_back( - *Literal::MakeTuple({Literal::CreateR1({1.0, -42.0}).get(), - Literal::CreateR0(123456.0).get()})); + *LiteralUtil::MakeTuple({LiteralUtil::CreateR1({1.0, -42.0}).get(), + LiteralUtil::CreateR0(123456.0).get()})); // Nested tuple. - test_to_device_and_back(*Literal::MakeTuple( - {Literal::MakeTuple({Literal::CreateR1({1.0, -42.0}).get(), - Literal::CreateR0(123456.0).get()}) + test_to_device_and_back(*LiteralUtil::MakeTuple( + {LiteralUtil::MakeTuple({LiteralUtil::CreateR1({1.0, -42.0}).get(), + LiteralUtil::CreateR0(123456.0).get()}) .get(), - Literal::CreateR0(false).get()})); + LiteralUtil::CreateR0(false).get()})); } XLA_TEST_F(LocalClientExecuteTest, ShapeBufferToLiteralConversion64bit) { @@ -832,24 +832,47 @@ XLA_TEST_F(LocalClientExecuteTest, ShapeBufferToLiteralConversion64bit) { }; test_to_device_and_back( - *Literal::CreateR2({{1.0, 2.0, 3.0}, {44.0, 0.1, -3}})); - test_to_device_and_back(*Literal::CreateR2({{2, 1}, {4444, 56}})); + *LiteralUtil::CreateR2({{1.0, 2.0, 3.0}, {44.0, 0.1, -3}})); + test_to_device_and_back(*LiteralUtil::CreateR2({{2, 1}, {4444, 56}})); test_to_device_and_back( - *Literal::CreateR2({{20000000000ULL, 1}, {4444, 56}})); - test_to_device_and_back( - *Literal::MakeTuple({Literal::CreateR1({1.0, -42.0}).get(), - Literal::CreateR0(123456789000LL).get()})); + *LiteralUtil::CreateR2({{20000000000ULL, 1}, {4444, 56}})); + test_to_device_and_back(*LiteralUtil::MakeTuple( + {LiteralUtil::CreateR1({1.0, -42.0}).get(), + LiteralUtil::CreateR0(123456789000LL).get()})); +} + +XLA_TEST_F(LocalClientExecuteTest, InfeedTest) { + XlaBuilder builder(TestName()); + const Shape shape = ShapeUtil::MakeShape(F32, {3}); + auto in = Infeed(&builder, shape); + auto constant = ConstantR1(&builder, {1.0f, 2.0f, 3.0f}); + Add(in, constant); + + std::unique_ptr result; + std::unique_ptr thread( + tensorflow::Env::Default()->StartThread( + tensorflow::ThreadOptions(), "execute_thread", [&] { + result = ShapedBufferToLiteral(ExecuteLocallyOrDie( + builder.Build().ValueOrDie(), /*arguments=*/{})); + })); + + ASSERT_IS_OK(local_client_->TransferToInfeedLocal( + *LiteralUtil::CreateR1({-5.0, 123.0, 42.0}), + local_client_->default_device_ordinal())); + + // Join the thread. + thread.reset(); + + LiteralTestUtil::ExpectR1Equal({-4.0, 125.0, 45.0}, *result); } -// TODO(b/34359662): Support infeed/outfeed on GPU and CPU parallel. -// 2017-10-18. -XLA_TEST_F(LocalClientExecuteTest, DISABLED_ON_GPU(InfeedOutfeedTest)) { +XLA_TEST_F(LocalClientExecuteTest, InfeedOutfeedTest) { XlaBuilder builder(TestName()); const Shape shape = ShapeUtil::MakeShape(F32, {3}); - auto in = builder.Infeed(shape); - auto constant = builder.ConstantR1({1.0f, 2.0f, 3.0f}); - auto sum = builder.Add(in, constant); - builder.Outfeed(sum, shape, /*outfeed_config=*/""); + auto in = Infeed(&builder, shape); + auto constant = ConstantR1(&builder, {1.0f, 2.0f, 3.0f}); + auto sum = Add(in, constant); + Outfeed(sum, shape, /*outfeed_config=*/""); std::unique_ptr thread( tensorflow::Env::Default()->StartThread( @@ -857,7 +880,7 @@ XLA_TEST_F(LocalClientExecuteTest, DISABLED_ON_GPU(InfeedOutfeedTest)) { [&] { ExecuteLocallyOrDie(builder.Build().ValueOrDie(), {}); })); ASSERT_IS_OK(local_client_->TransferToInfeedLocal( - *Literal::CreateR1({-5.0, 123.0, 42.0}), + *LiteralUtil::CreateR1({-5.0, 123.0, 42.0}), local_client_->default_device_ordinal())); TF_ASSERT_OK_AND_ASSIGN(std::unique_ptr result, @@ -884,15 +907,15 @@ void BM_LocalClientOverhead(int num_iters) { // Use a tiny add operation as the computation. XlaBuilder builder("Add"); auto shape = ShapeUtil::MakeShape(F32, {2, 3}); - auto x = builder.Parameter(0, shape, "x"); - builder.Add(x, x); + auto x = Parameter(&builder, 0, shape, "x"); + Add(x, x); auto computation = builder.Build().ConsumeValueOrDie(); auto buffer = transfer_manager ->AllocateScopedShapedBuffer(shape, &allocator, /*device_ordinal=*/0) .ConsumeValueOrDie(); - auto literal = Literal::CreateR2({{0, 0, 0}, {0, 0, 0}}); + auto literal = LiteralUtil::CreateR2({{0, 0, 0}, {0, 0, 0}}); auto stream = client->mutable_backend()->BorrowStream(device_ordinal).ValueOrDie(); ASSERT_IS_OK(transfer_manager->TransferLiteralToDevice(stream.get(), *literal, diff --git a/tensorflow/compiler/xla/tests/log_test.cc b/tensorflow/compiler/xla/tests/log_test.cc index c0c02e584c2348f64a9d7d0800038f5ca67a2171..cdf70ee4185be2ecd9dcb2d21fbd98c2ab6cc0ad 100644 --- a/tensorflow/compiler/xla/tests/log_test.cc +++ b/tensorflow/compiler/xla/tests/log_test.cc @@ -30,8 +30,8 @@ class LogTest : public ClientLibraryTestBase {}; XLA_TEST_F(LogTest, LogZeroValues) { XlaBuilder builder(TestName()); - auto x = builder.ConstantR3FromArray3D(Array3D(3, 0, 0)); - builder.Log(x); + auto x = ConstantR3FromArray3D(&builder, Array3D(3, 0, 0)); + Log(x); ComputeAndCompareR3(&builder, Array3D(3, 0, 0), {}, ErrorSpec(0.0001)); @@ -42,8 +42,8 @@ TEST_F(LogTest, LogTenValues) { 5.0, 6.0, -7.0, -8.0, 9.0}; XlaBuilder builder(TestName()); - auto x = builder.ConstantR1(input); - builder.Log(x); + auto x = ConstantR1(&builder, input); + Log(x); std::vector expected; expected.reserve(input.size()); diff --git a/tensorflow/compiler/xla/tests/map_test.cc b/tensorflow/compiler/xla/tests/map_test.cc index 3975e9125703ee081d4e84fa8bd27fcbe483ac34..7ddc6369319810c0806afa161bc00f51caea2072 100644 --- a/tensorflow/compiler/xla/tests/map_test.cc +++ b/tensorflow/compiler/xla/tests/map_test.cc @@ -21,7 +21,7 @@ limitations under the License. #include "tensorflow/compiler/xla/client/local_client.h" #include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/compiler/xla/client/xla_client/xla_computation.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/shape_util.h" #include "tensorflow/compiler/xla/statusor.h" #include "tensorflow/compiler/xla/test.h" @@ -52,9 +52,9 @@ class MapTest : public ClientLibraryTestBase { // 1.0f ---------/ XlaComputation CreateAdderToOne() { XlaBuilder mapped_builder(TestName()); - auto x = mapped_builder.Parameter(0, ShapeUtil::MakeShape(F32, {}), "x"); - auto one = mapped_builder.ConstantR0(1.0); - mapped_builder.Add(x, one); + auto x = Parameter(&mapped_builder, 0, ShapeUtil::MakeShape(F32, {}), "x"); + auto one = ConstantR0(&mapped_builder, 1.0); + Add(x, one); auto computation_status = mapped_builder.Build(); TF_CHECK_OK(computation_status.status()); return computation_status.ConsumeValueOrDie(); @@ -62,9 +62,9 @@ class MapTest : public ClientLibraryTestBase { XlaComputation CreateMax() { XlaBuilder b(TestName()); - auto lhs = b.Parameter(0, ShapeUtil::MakeShape(F32, {}), "x"); - auto rhs = b.Parameter(1, ShapeUtil::MakeShape(F32, {}), "y"); - b.Max(lhs, rhs); + auto lhs = Parameter(&b, 0, ShapeUtil::MakeShape(F32, {}), "x"); + auto rhs = Parameter(&b, 1, ShapeUtil::MakeShape(F32, {}), "y"); + Max(lhs, rhs); auto computation_status = b.Build(); TF_CHECK_OK(computation_status.status()); return computation_status.ConsumeValueOrDie(); @@ -75,8 +75,8 @@ class MapTest : public ClientLibraryTestBase { template XlaComputation CreateScalarOne() { XlaBuilder mapped_builder("scalar_one"); - (void)mapped_builder.Parameter(0, ShapeUtil::MakeShape(F32, {}), "x"); - mapped_builder.ConstantR0(1); + (void)Parameter(&mapped_builder, 0, ShapeUtil::MakeShape(F32, {}), "x"); + ConstantR0(&mapped_builder, 1); auto computation_status = mapped_builder.Build(); TF_CHECK_OK(computation_status.status()); return computation_status.ConsumeValueOrDie(); @@ -89,9 +89,9 @@ class MapTest : public ClientLibraryTestBase { // 2.0f ---------/ XlaComputation CreateMulByTwo() { XlaBuilder mapped_builder(TestName()); - auto x = mapped_builder.Parameter(0, ShapeUtil::MakeShape(F32, {}), "x"); - auto two = mapped_builder.ConstantR0(2.0); - mapped_builder.Mul(x, two); + auto x = Parameter(&mapped_builder, 0, ShapeUtil::MakeShape(F32, {}), "x"); + auto two = ConstantR0(&mapped_builder, 2.0); + Mul(x, two); auto computation_status = mapped_builder.Build(); TF_CHECK_OK(computation_status.status()); return computation_status.ConsumeValueOrDie(); @@ -107,10 +107,10 @@ class MapTest : public ClientLibraryTestBase { // 1.0f ---------/ XlaComputation CreateAdderToOneTimesItself() { XlaBuilder mapped_builder(TestName()); - auto x = mapped_builder.Parameter(0, ShapeUtil::MakeShape(F32, {}), "x"); - auto one = mapped_builder.ConstantR0(1.0); - auto adder_to_one = mapped_builder.Add(x, one); - mapped_builder.Mul(x, adder_to_one); + auto x = Parameter(&mapped_builder, 0, ShapeUtil::MakeShape(F32, {}), "x"); + auto one = ConstantR0(&mapped_builder, 1.0); + auto adder_to_one = Add(x, one); + Mul(x, adder_to_one); auto computation_status = mapped_builder.Build(); TF_CHECK_OK(computation_status.status()); return computation_status.ConsumeValueOrDie(); @@ -125,10 +125,10 @@ class MapTest : public ClientLibraryTestBase { XlaComputation CreateMapPlusN(const XlaComputation& embedded_computation, float n) { XlaBuilder builder(TestName()); - auto x = builder.Parameter(0, ShapeUtil::MakeShape(F32, {}), "x"); - auto map = builder.Map({x}, embedded_computation, {}); - auto constant_n = builder.ConstantR0(n); - builder.Add(map, constant_n); + auto x = Parameter(&builder, 0, ShapeUtil::MakeShape(F32, {}), "x"); + auto map = Map(&builder, {x}, embedded_computation, {}); + auto constant_n = ConstantR0(&builder, n); + Add(map, constant_n); auto computation_status = builder.Build(); TF_CHECK_OK(computation_status.status()); return computation_status.ConsumeValueOrDie(); @@ -138,9 +138,9 @@ class MapTest : public ClientLibraryTestBase { // defined by (x, y) -> x > y. XlaComputation CreateGt() { XlaBuilder b("Gt"); - auto x = b.Parameter(0, ShapeUtil::MakeShape(F32, {}), "x"); - auto y = b.Parameter(1, ShapeUtil::MakeShape(F32, {}), "y"); - b.Gt(x, y); + auto x = Parameter(&b, 0, ShapeUtil::MakeShape(F32, {}), "x"); + auto y = Parameter(&b, 1, ShapeUtil::MakeShape(F32, {}), "y"); + Gt(x, y); auto computation_status = b.Build(); TF_CHECK_OK(computation_status.status()); return computation_status.ConsumeValueOrDie(); @@ -155,11 +155,11 @@ class MapTest : public ClientLibraryTestBase { // z {R0F32} ---------------/ XlaComputation CreateTernaryAdder() { XlaBuilder mapped_builder("TernaryAdder"); - auto x = mapped_builder.Parameter(0, ShapeUtil::MakeShape(F32, {}), "x"); - auto y = mapped_builder.Parameter(1, ShapeUtil::MakeShape(F32, {}), "y"); - auto z = mapped_builder.Parameter(2, ShapeUtil::MakeShape(F32, {}), "z"); - auto xy = mapped_builder.Add(x, y); - mapped_builder.Add(xy, z); + auto x = Parameter(&mapped_builder, 0, ShapeUtil::MakeShape(F32, {}), "x"); + auto y = Parameter(&mapped_builder, 1, ShapeUtil::MakeShape(F32, {}), "y"); + auto z = Parameter(&mapped_builder, 2, ShapeUtil::MakeShape(F32, {}), "z"); + auto xy = Add(x, y); + Add(xy, z); auto computation_status = mapped_builder.Build(); TF_CHECK_OK(computation_status.status()); return computation_status.ConsumeValueOrDie(); @@ -169,12 +169,12 @@ class MapTest : public ClientLibraryTestBase { TEST_F(MapTest, MapEachElemPlusOneR0) { // Applies lambda (x) (+ x 1)) to an input scalar. XlaBuilder builder(TestName()); - std::unique_ptr param0_literal = Literal::CreateR0(42.0); + std::unique_ptr param0_literal = LiteralUtil::CreateR0(42.0); std::unique_ptr param0_data = client_->TransferToServer(*param0_literal).ConsumeValueOrDie(); - auto param = builder.Parameter(0, param0_literal->shape(), "param0"); - builder.Map({param}, CreateAdderToOne(), {}); + auto param = Parameter(&builder, 0, param0_literal->shape(), "param0"); + Map(&builder, {param}, CreateAdderToOne(), {}); ComputeAndCompareR0(&builder, 43.0, {param0_data.get()}, ErrorSpec(0.01f)); @@ -183,12 +183,12 @@ TEST_F(MapTest, MapEachElemPlusOneR0) { XLA_TEST_F(MapTest, MapEachElemPlusOneR1S0) { // Maps (lambda (x) (+ x 1)) onto an input R1F32 vector of length 0. XlaBuilder builder(TestName()); - std::unique_ptr param0_literal = Literal::CreateR1({}); + std::unique_ptr param0_literal = LiteralUtil::CreateR1({}); std::unique_ptr param0_data = client_->TransferToServer(*param0_literal).ConsumeValueOrDie(); - auto param = builder.Parameter(0, param0_literal->shape(), "param0"); - builder.Map({param}, CreateAdderToOne(), {0}); + auto param = Parameter(&builder, 0, param0_literal->shape(), "param0"); + Map(&builder, {param}, CreateAdderToOne(), {0}); ComputeAndCompareR1(&builder, {}, {param0_data.get()}, ErrorSpec(0.01f)); @@ -198,12 +198,12 @@ TEST_F(MapTest, MapEachElemPlusOneR1S4) { // Maps (lambda (x) (+ x 1)) onto an input R1F32 vector of length 4. XlaBuilder builder(TestName()); std::unique_ptr param0_literal = - Literal::CreateR1({2.2f, 3.3f, 4.4f, 5.5f}); + LiteralUtil::CreateR1({2.2f, 3.3f, 4.4f, 5.5f}); std::unique_ptr param0_data = client_->TransferToServer(*param0_literal).ConsumeValueOrDie(); - auto param = builder.Parameter(0, param0_literal->shape(), "param0"); - builder.Map({param}, CreateAdderToOne(), {0}); + auto param = Parameter(&builder, 0, param0_literal->shape(), "param0"); + Map(&builder, {param}, CreateAdderToOne(), {0}); ComputeAndCompareR1(&builder, {3.2f, 4.3f, 5.4f, 6.5f}, {param0_data.get()}, ErrorSpec(0.01f)); @@ -212,12 +212,12 @@ TEST_F(MapTest, MapEachElemPlusOneR1S4) { TEST_F(MapTest, MapEachF32ElementToS32Constant) { XlaBuilder builder(TestName()); std::unique_ptr param0_literal = - Literal::CreateR1({2.2f, 3.3f, 4.4f, 5.5f}); + LiteralUtil::CreateR1({2.2f, 3.3f, 4.4f, 5.5f}); std::unique_ptr param0_data = client_->TransferToServer(*param0_literal).ConsumeValueOrDie(); - auto param = builder.Parameter(0, param0_literal->shape(), "param0"); - builder.Map({param}, CreateScalarOne(), {0}); + auto param = Parameter(&builder, 0, param0_literal->shape(), "param0"); + Map(&builder, {param}, CreateScalarOne(), {0}); ComputeAndCompareR1(&builder, {1, 1, 1, 1}, {param0_data.get()}); } @@ -225,12 +225,12 @@ TEST_F(MapTest, MapEachF32ElementToS32Constant) { TEST_F(MapTest, MapEachF32ElementToU32Constant) { XlaBuilder builder(TestName()); std::unique_ptr param0_literal = - Literal::CreateR1({2.2f, 3.3f, 4.4f, 5.5f}); + LiteralUtil::CreateR1({2.2f, 3.3f, 4.4f, 5.5f}); std::unique_ptr param0_data = client_->TransferToServer(*param0_literal).ConsumeValueOrDie(); - auto param = builder.Parameter(0, param0_literal->shape(), "param0"); - builder.Map({param}, CreateScalarOne(), {0}); + auto param = Parameter(&builder, 0, param0_literal->shape(), "param0"); + Map(&builder, {param}, CreateScalarOne(), {0}); ComputeAndCompareR1(&builder, {1, 1, 1, 1}, {param0_data.get()}); } @@ -239,12 +239,12 @@ TEST_F(MapTest, MapEachElemLongerChainR1) { // Maps (lambda (x) (* (+ x 1) x)) onto an input R1F32 vector. XlaBuilder builder(TestName()); std::unique_ptr param0_literal = - Literal::CreateR1({2.6f, -5.1f, 0.1f, 0.2f, 999.0f, 255.5f}); + LiteralUtil::CreateR1({2.6f, -5.1f, 0.1f, 0.2f, 999.0f, 255.5f}); std::unique_ptr param0_data = client_->TransferToServer(*param0_literal).ConsumeValueOrDie(); - auto param = builder.Parameter(0, param0_literal->shape(), "param0"); - builder.Map({param}, CreateAdderToOneTimesItself(), {0}); + auto param = Parameter(&builder, 0, param0_literal->shape(), "param0"); + Map(&builder, {param}, CreateAdderToOneTimesItself(), {0}); ComputeAndCompareR1( &builder, {9.36f, 20.91f, 0.11f, 0.24f, 999000.0f, 65535.75f}, @@ -255,13 +255,13 @@ XLA_TEST_F(MapTest, MapMultipleMapsR1S0) { // Maps (lambda (x) (+ x 1)) onto an input R1F32 vector of length 0, and then // maps (lambda (x) (* x 2)) on the result. XlaBuilder builder(TestName()); - std::unique_ptr param0_literal = Literal::CreateR1({}); + std::unique_ptr param0_literal = LiteralUtil::CreateR1({}); std::unique_ptr param0_data = client_->TransferToServer(*param0_literal).ConsumeValueOrDie(); - auto param = builder.Parameter(0, param0_literal->shape(), "param0"); - auto map1 = builder.Map({param}, CreateAdderToOne(), {0}); - builder.Map({map1}, CreateMulByTwo(), {0}); + auto param = Parameter(&builder, 0, param0_literal->shape(), "param0"); + auto map1 = Map(&builder, {param}, CreateAdderToOne(), {0}); + Map(&builder, {map1}, CreateMulByTwo(), {0}); ComputeAndCompareR1(&builder, {}, {param0_data.get()}, ErrorSpec(0.01f)); @@ -272,13 +272,13 @@ TEST_F(MapTest, MapMultipleMapsR1S4) { // maps (lambda (x) (* x 2)) on the result. XlaBuilder builder(TestName()); std::unique_ptr param0_literal = - Literal::CreateR1({2.2f, 3.3f, 4.4f, 5.5f}); + LiteralUtil::CreateR1({2.2f, 3.3f, 4.4f, 5.5f}); std::unique_ptr param0_data = client_->TransferToServer(*param0_literal).ConsumeValueOrDie(); - auto param = builder.Parameter(0, param0_literal->shape(), "param0"); - auto map1 = builder.Map({param}, CreateAdderToOne(), {0}); - builder.Map({map1}, CreateMulByTwo(), {0}); + auto param = Parameter(&builder, 0, param0_literal->shape(), "param0"); + auto map1 = Map(&builder, {param}, CreateAdderToOne(), {0}); + Map(&builder, {map1}, CreateMulByTwo(), {0}); ComputeAndCompareR1(&builder, {6.4f, 8.6f, 10.8f, 13.0f}, {param0_data.get()}, ErrorSpec(0.01f)); @@ -287,13 +287,13 @@ TEST_F(MapTest, MapMultipleMapsR1S4) { TEST_F(MapTest, MapEachElemPlusOneR2) { // Maps (lambda (x) (+ x 1)) onto an input R2F32 vector. XlaBuilder builder(TestName()); - std::unique_ptr param0_literal = Literal::CreateR2( + std::unique_ptr param0_literal = LiteralUtil::CreateR2( {{13.25f, 14.0f}, {-7.1f, -7.2f}, {-8.8f, 8.8f}}); std::unique_ptr param0_data = client_->TransferToServer(*param0_literal).ConsumeValueOrDie(); - auto param = builder.Parameter(0, param0_literal->shape(), "param0"); - builder.Map({param}, CreateAdderToOne(), {0, 1}); + auto param = Parameter(&builder, 0, param0_literal->shape(), "param0"); + Map(&builder, {param}, CreateAdderToOne(), {0, 1}); Array2D expected_array( {{14.25f, 15.0f}, {-6.1f, -6.2f}, {-7.8f, 9.8f}}); @@ -319,10 +319,10 @@ XLA_TEST_F(MapTest, ComplexNestedMaps) { auto embed3 = CreateMapPlusN(embed1, 4.0); XlaBuilder embed4_builder("embed4"); - auto embed4_param = embed4_builder.Parameter(0, scalar_shape, "x"); - auto embed4_map_lhs = embed4_builder.Map({embed4_param}, embed2, {}); - auto embed4_map_rhs = embed4_builder.Map({embed4_param}, embed3, {}); - embed4_builder.Add(embed4_map_lhs, embed4_map_rhs); + auto embed4_param = Parameter(&embed4_builder, 0, scalar_shape, "x"); + auto embed4_map_lhs = Map(&embed4_builder, {embed4_param}, embed2, {}); + auto embed4_map_rhs = Map(&embed4_builder, {embed4_param}, embed3, {}); + Add(embed4_map_lhs, embed4_map_rhs); auto embed4_status = embed4_builder.Build(); ASSERT_IS_OK(embed4_status.status()); auto embed4 = embed4_status.ConsumeValueOrDie(); @@ -330,11 +330,11 @@ XLA_TEST_F(MapTest, ComplexNestedMaps) { auto embed5 = CreateMapPlusN(embed2, 6.0); XlaBuilder builder(TestName()); - auto constant_42 = builder.ConstantR0(42.0); - auto constant_7 = builder.ConstantR0(7.0); - auto map_42 = builder.Map({constant_42}, embed5, {}); - auto map_7 = builder.Map({constant_7}, embed4, {}); - builder.Add(map_42, map_7); + auto constant_42 = ConstantR0(&builder, 42.0); + auto constant_7 = ConstantR0(&builder, 7.0); + auto map_42 = Map(&builder, {constant_42}, embed5, {}); + auto map_7 = Map(&builder, {constant_7}, embed4, {}); + Add(map_42, map_7); ComputeAndCompareR0(&builder, 73.0, {}, ErrorSpec(0.01f)); } @@ -343,17 +343,18 @@ TEST_F(MapTest, MapBinaryAdder) { // Maps (lambda (x y) (+ x y)) onto two R1F32 vectors. XlaBuilder builder(TestName()); std::unique_ptr param0_literal = - Literal::CreateR1({2.2f, 3.3f, 4.4f, 5.5f}); + LiteralUtil::CreateR1({2.2f, 3.3f, 4.4f, 5.5f}); std::unique_ptr param0_data = client_->TransferToServer(*param0_literal).ConsumeValueOrDie(); std::unique_ptr param1_literal = - Literal::CreateR1({5.1f, 4.4f, -0.1f, -5.5f}); + LiteralUtil::CreateR1({5.1f, 4.4f, -0.1f, -5.5f}); std::unique_ptr param1_data = client_->TransferToServer(*param1_literal).ConsumeValueOrDie(); - auto param0 = builder.Parameter(0, param0_literal->shape(), "param0"); - auto param1 = builder.Parameter(1, param1_literal->shape(), "param1"); - builder.Map({param0, param1}, CreateScalarAddComputation(F32, &builder), {0}); + auto param0 = Parameter(&builder, 0, param0_literal->shape(), "param0"); + auto param1 = Parameter(&builder, 1, param1_literal->shape(), "param1"); + Map(&builder, {param0, param1}, CreateScalarAddComputation(F32, &builder), + {0}); ComputeAndCompareR1(&builder, {7.3f, 7.7, 4.3f, 0}, {param0_data.get(), param1_data.get()}, @@ -364,20 +365,20 @@ TEST_F(MapTest, MapBinaryAdder) { // for Map that used to fail in shape inference (b/28989438). XLA_TEST_F(MapTest, AddWithMixedLayouts) { XlaBuilder builder(TestName()); - std::unique_ptr param0_literal = Literal::CreateR2WithLayout( + std::unique_ptr param0_literal = LiteralUtil::CreateR2WithLayout( {{1, 2}, {3, 4}}, LayoutUtil::MakeLayout({1, 0})); std::unique_ptr param0_data = client_->TransferToServer(*param0_literal).ConsumeValueOrDie(); - std::unique_ptr param1_literal = Literal::CreateR2WithLayout( + std::unique_ptr param1_literal = LiteralUtil::CreateR2WithLayout( {{10, 20}, {30, 40}}, LayoutUtil::MakeLayout({0, 1})); std::unique_ptr param1_data = client_->TransferToServer(*param1_literal).ConsumeValueOrDie(); - auto param0 = builder.Parameter(0, param0_literal->shape(), "param0"); - auto param1 = builder.Parameter(1, param1_literal->shape(), "param1"); - builder.Map({param0, param1}, CreateScalarAddComputation(S32, &builder), - {0, 1}); + auto param0 = Parameter(&builder, 0, param0_literal->shape(), "param0"); + auto param1 = Parameter(&builder, 1, param1_literal->shape(), "param1"); + Map(&builder, {param0, param1}, CreateScalarAddComputation(S32, &builder), + {0, 1}); Array2D expected(2, 2); expected(0, 0) = 11; @@ -391,19 +392,19 @@ XLA_TEST_F(MapTest, AddWithMixedLayouts) { XLA_TEST_F(MapTest, AddR3_3x0x2) { XlaBuilder builder(TestName()); std::unique_ptr param0_literal = - Literal::CreateR3FromArray3D(Array3D(3, 0, 2)); + LiteralUtil::CreateR3FromArray3D(Array3D(3, 0, 2)); std::unique_ptr param0_data = client_->TransferToServer(*param0_literal).ConsumeValueOrDie(); std::unique_ptr param1_literal = - Literal::CreateR3FromArray3D(Array3D(3, 0, 2)); + LiteralUtil::CreateR3FromArray3D(Array3D(3, 0, 2)); std::unique_ptr param1_data = client_->TransferToServer(*param1_literal).ConsumeValueOrDie(); - auto param0 = builder.Parameter(0, param0_literal->shape(), "param0"); - auto param1 = builder.Parameter(1, param1_literal->shape(), "param1"); - builder.Map({param0, param1}, CreateScalarAddComputation(S32, &builder), - {0, 1, 2}); + auto param0 = Parameter(&builder, 0, param0_literal->shape(), "param0"); + auto param1 = Parameter(&builder, 1, param1_literal->shape(), "param1"); + Map(&builder, {param0, param1}, CreateScalarAddComputation(S32, &builder), + {0, 1, 2}); ComputeAndCompareR3(&builder, Array3D(3, 0, 2), {param0_data.get(), param1_data.get()}); @@ -413,22 +414,22 @@ TEST_F(MapTest, MapTernaryAdder) { // Maps (lambda (x y z) (+ x y z)) onto three R1F32 vectors. XlaBuilder builder(TestName()); std::unique_ptr param0_literal = - Literal::CreateR1({2.2f, 3.3f, 4.4f, 5.5f}); + LiteralUtil::CreateR1({2.2f, 3.3f, 4.4f, 5.5f}); std::unique_ptr param0_data = client_->TransferToServer(*param0_literal).ConsumeValueOrDie(); std::unique_ptr param1_literal = - Literal::CreateR1({5.1f, 4.4f, -0.1f, -5.5f}); + LiteralUtil::CreateR1({5.1f, 4.4f, -0.1f, -5.5f}); std::unique_ptr param1_data = client_->TransferToServer(*param1_literal).ConsumeValueOrDie(); std::unique_ptr param2_literal = - Literal::CreateR1({-10.0f, -100.0f, -900.0f, -400.0f}); + LiteralUtil::CreateR1({-10.0f, -100.0f, -900.0f, -400.0f}); std::unique_ptr param2_data = client_->TransferToServer(*param2_literal).ConsumeValueOrDie(); - auto param0 = builder.Parameter(0, param0_literal->shape(), "param0"); - auto param1 = builder.Parameter(1, param1_literal->shape(), "param1"); - auto param2 = builder.Parameter(2, param2_literal->shape(), "param2"); - builder.Map({param0, param1, param2}, CreateTernaryAdder(), {0}); + auto param0 = Parameter(&builder, 0, param0_literal->shape(), "param0"); + auto param1 = Parameter(&builder, 1, param1_literal->shape(), "param1"); + auto param2 = Parameter(&builder, 2, param2_literal->shape(), "param2"); + Map(&builder, {param0, param1, param2}, CreateTernaryAdder(), {0}); ComputeAndCompareR1( &builder, {-2.7f, -92.3f, -895.7f, -400.0f}, @@ -440,7 +441,8 @@ TEST_F(MapTest, MapGt) { // Maps (x,y) -> x > y onto two R1F32 vectors. XlaBuilder b(TestName()); auto gt = CreateGt(); - b.Map({b.ConstantR1({1, 20}), b.ConstantR1({10, 2})}, gt, {0}); + Map(&b, {ConstantR1(&b, {1, 20}), ConstantR1(&b, {10, 2})}, gt, + {0}); ComputeAndCompareR1(&b, {false, true}, {}); } @@ -449,15 +451,15 @@ TEST_F(MapTest, NestedBinaryMap) { { // max_with_square(x) = do max(x, x^2) via a map. XlaBuilder b("max_with_square"); - auto x = b.Parameter(0, ShapeUtil::MakeShape(F32, {}), "x"); - b.Map({x, b.Mul(x, x)}, CreateMax(), {}); + auto x = Parameter(&b, 0, ShapeUtil::MakeShape(F32, {}), "x"); + Map(&b, {x, Mul(x, x)}, CreateMax(), {}); auto computation_status = b.Build(); ASSERT_IS_OK(computation_status.status()); max_with_square = computation_status.ConsumeValueOrDie(); } XlaBuilder b(TestName()); - auto input = b.ConstantR1({0.1f, 0.5f, -0.5f, 1.0f, 2.0f}); - b.Map({input}, max_with_square, {0}); + auto input = ConstantR1(&b, {0.1f, 0.5f, -0.5f, 1.0f, 2.0f}); + Map(&b, {input}, max_with_square, {0}); ComputeAndCompareR1(&b, {0.1f, 0.5f, 0.25f, 1.0f, 4.0f}, {}); } @@ -468,23 +470,23 @@ TEST_F(MapTest, MapOperantionWithBuildError) { XlaBuilder builder(TestName()); auto sub_builder = builder.CreateSubBuilder("ErrorAdd"); - auto x = sub_builder->Parameter(0, ShapeUtil::MakeShape(F32, {}), "x"); - auto y = sub_builder->Parameter(1, ShapeUtil::MakeShape(U16, {}), "y"); - sub_builder->Add(x, y); + auto x = Parameter(sub_builder.get(), 0, ShapeUtil::MakeShape(F32, {}), "x"); + auto y = Parameter(sub_builder.get(), 1, ShapeUtil::MakeShape(U16, {}), "y"); + Add(x, y); auto error_add = sub_builder->BuildAndNoteError(); std::unique_ptr param0_literal = - Literal::CreateR1({2.2f, 3.3f, 4.4f, 5.5f}); + LiteralUtil::CreateR1({2.2f, 3.3f, 4.4f, 5.5f}); std::unique_ptr param0_data = client_->TransferToServer(*param0_literal).ConsumeValueOrDie(); std::unique_ptr param1_literal = - Literal::CreateR1({5.1f, 4.4f, -0.1f, -5.5f}); + LiteralUtil::CreateR1({5.1f, 4.4f, -0.1f, -5.5f}); std::unique_ptr param1_data = client_->TransferToServer(*param1_literal).ConsumeValueOrDie(); - auto param0 = builder.Parameter(0, param0_literal->shape(), "param0"); - auto param1 = builder.Parameter(1, param1_literal->shape(), "param1"); - builder.Map({param0, param1}, error_add, {0}); + auto param0 = Parameter(&builder, 0, param0_literal->shape(), "param0"); + auto param1 = Parameter(&builder, 1, param1_literal->shape(), "param1"); + Map(&builder, {param0, param1}, error_add, {0}); StatusOr computation_status = builder.Build(); ASSERT_TRUE(!computation_status.ok()); @@ -506,21 +508,21 @@ TEST_F(MapTestWithFullOpt, MapScalarPower) { XlaBuilder builder(TestName()); auto sub_builder = builder.CreateSubBuilder("power"); - auto x = sub_builder->Parameter(0, ShapeUtil::MakeShape(F32, {}), "x"); - auto y = sub_builder->Parameter(1, ShapeUtil::MakeShape(F32, {}), "y"); - sub_builder->Pow(x, y); + auto x = Parameter(sub_builder.get(), 0, ShapeUtil::MakeShape(F32, {}), "x"); + auto y = Parameter(sub_builder.get(), 1, ShapeUtil::MakeShape(F32, {}), "y"); + Pow(x, y); auto power = sub_builder->BuildAndNoteError(); - std::unique_ptr param0_literal = Literal::CreateR0(2.0f); - std::unique_ptr param1_literal = Literal::CreateR0(5.0f); + std::unique_ptr param0_literal = LiteralUtil::CreateR0(2.0f); + std::unique_ptr param1_literal = LiteralUtil::CreateR0(5.0f); std::unique_ptr param0_data = client_->TransferToServer(*param0_literal).ConsumeValueOrDie(); std::unique_ptr param1_data = client_->TransferToServer(*param1_literal).ConsumeValueOrDie(); - auto param0 = builder.Parameter(0, param0_literal->shape(), "param0"); - auto param1 = builder.Parameter(1, param1_literal->shape(), "param1"); - builder.Map({param0, param1}, power, {}); + auto param0 = Parameter(&builder, 0, param0_literal->shape(), "param0"); + auto param1 = Parameter(&builder, 1, param1_literal->shape(), "param1"); + Map(&builder, {param0, param1}, power, {}); ComputeAndCompareR0(&builder, 32.0f, {param0_data.get(), param1_data.get()}, @@ -533,21 +535,21 @@ TEST_F(MapTestWithFullOpt, MapSubtractOppositeOrder) { XlaBuilder builder(TestName()); auto sub_builder = builder.CreateSubBuilder("power"); - auto x = sub_builder->Parameter(0, ShapeUtil::MakeShape(F32, {}), "x"); - auto y = sub_builder->Parameter(1, ShapeUtil::MakeShape(F32, {}), "y"); - sub_builder->Sub(y, x); // note that this is y - x, not x - y + auto x = Parameter(sub_builder.get(), 0, ShapeUtil::MakeShape(F32, {}), "x"); + auto y = Parameter(sub_builder.get(), 1, ShapeUtil::MakeShape(F32, {}), "y"); + Sub(y, x); // note that this is y - x, not x - y auto sub_opposite = sub_builder->BuildAndNoteError(); - std::unique_ptr param0_literal = Literal::CreateR0(2.0f); - std::unique_ptr param1_literal = Literal::CreateR0(5.0f); + std::unique_ptr param0_literal = LiteralUtil::CreateR0(2.0f); + std::unique_ptr param1_literal = LiteralUtil::CreateR0(5.0f); std::unique_ptr param0_data = client_->TransferToServer(*param0_literal).ConsumeValueOrDie(); std::unique_ptr param1_data = client_->TransferToServer(*param1_literal).ConsumeValueOrDie(); - auto param0 = builder.Parameter(0, param0_literal->shape(), "param0"); - auto param1 = builder.Parameter(1, param1_literal->shape(), "param1"); - builder.Map({param0, param1}, sub_opposite, {}); + auto param0 = Parameter(&builder, 0, param0_literal->shape(), "param0"); + auto param1 = Parameter(&builder, 1, param1_literal->shape(), "param1"); + Map(&builder, {param0, param1}, sub_opposite, {}); ComputeAndCompareR0( &builder, 3.0f, {param0_data.get(), param1_data.get()}, ErrorSpec(0.01f)); @@ -559,16 +561,16 @@ TEST_F(MapTestWithFullOpt, MapSquare) { XlaBuilder builder(TestName()); auto sub_builder = builder.CreateSubBuilder("power"); - auto x = sub_builder->Parameter(0, ShapeUtil::MakeShape(F32, {}), "x"); - sub_builder->Mul(x, x); + auto x = Parameter(sub_builder.get(), 0, ShapeUtil::MakeShape(F32, {}), "x"); + Mul(x, x); auto square = sub_builder->BuildAndNoteError(); - std::unique_ptr param0_literal = Literal::CreateR0(10.0f); + std::unique_ptr param0_literal = LiteralUtil::CreateR0(10.0f); std::unique_ptr param0_data = client_->TransferToServer(*param0_literal).ConsumeValueOrDie(); - auto param0 = builder.Parameter(0, param0_literal->shape(), "param0"); - builder.Map({param0}, square, {}); + auto param0 = Parameter(&builder, 0, param0_literal->shape(), "param0"); + Map(&builder, {param0}, square, {}); ComputeAndCompareR0(&builder, 100.0f, {param0_data.get()}, ErrorSpec(0.01f)); diff --git a/tensorflow/compiler/xla/tests/matrix_ops_simple_test.cc b/tensorflow/compiler/xla/tests/matrix_ops_simple_test.cc index 27fd36e06acdc589f3a84ad561164e4a33b93506..069b8a881f4be0c05b19bb1f323bdc13c7222ceb 100644 --- a/tensorflow/compiler/xla/tests/matrix_ops_simple_test.cc +++ b/tensorflow/compiler/xla/tests/matrix_ops_simple_test.cc @@ -21,7 +21,7 @@ limitations under the License. #include "tensorflow/compiler/xla/client/local_client.h" #include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/compiler/xla/client/xla_client/xla_computation.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/ptr_util.h" #include "tensorflow/compiler/xla/reference_util.h" #include "tensorflow/compiler/xla/shape_util.h" @@ -56,15 +56,15 @@ TYPED_TEST_CASE(MatOpsSimpleTest_F16F32, TypesF16F32); XLA_TYPED_TEST(MatOpsSimpleTest_F16F32, ExpTwoByTwoValues) { using T = TypeParam; XlaBuilder builder("exp_2x2"); - auto data = builder.ConstantR2FromArray2D({ - {1.0f, 0.0f}, // row 0 - {-1.0f, 0.5f}, // row 1 - }); - builder.Exp(data); + auto data = ConstantR2FromArray2D(&builder, { + {1.0f, 0.0f}, // row 0 + {-1.0f, 0.5f}, // row 1 + }); + Exp(data); std::unique_ptr expected = - Literal::CreateR2FromArray2D({{2.71828f, 1.00000f}, // row 0 - {0.36788f, 1.64872f}}); // row 1 + LiteralUtil::CreateR2FromArray2D({{2.71828f, 1.00000f}, // row 0 + {0.36788f, 1.64872f}}); // row 1 this->ComputeAndCompareLiteral(&builder, *expected, {}, ErrorSpec(1e-5)); } @@ -76,43 +76,43 @@ XLA_TYPED_TEST(MatOpsSimpleTest_F16F32, MapTwoByTwo) { // add_half(x) = x + 0.5 XlaBuilder builder("add_half"); auto x_value = - builder.Parameter(0, ShapeUtil::MakeShapeWithType({}), "x_value"); - auto half = builder.ConstantR0(static_cast(0.5)); - builder.Add(x_value, half); + Parameter(&builder, 0, ShapeUtil::MakeShapeWithType({}), "x_value"); + auto half = ConstantR0(&builder, static_cast(0.5)); + Add(x_value, half); auto computation_status = builder.Build(); ASSERT_IS_OK(computation_status.status()); add_half = computation_status.ConsumeValueOrDie(); } XlaBuilder builder("map_2x2"); - auto data = builder.ConstantR2FromArray2D({ - {1.0f, 0.0f}, // row 0 - {-1.0f, 0.5f}, // row 1 - }); - auto map = builder.Map({data}, add_half, {0, 1}); + auto data = ConstantR2FromArray2D(&builder, { + {1.0f, 0.0f}, // row 0 + {-1.0f, 0.5f}, // row 1 + }); + Map(&builder, {data}, add_half, {0, 1}); std::unique_ptr expected = - Literal::CreateR2FromArray2D({{1.5f, 0.5f}, // row 0 - {-0.5f, 1.0f}}); // row 1 + LiteralUtil::CreateR2FromArray2D({{1.5f, 0.5f}, // row 0 + {-0.5f, 1.0f}}); // row 1 this->ComputeAndCompareLiteral(&builder, *expected, {}, ErrorSpec(1e-5)); } XLA_TYPED_TEST(MatOpsSimpleTest_F16F32, MaxTwoByTwoValues) { using T = TypeParam; XlaBuilder builder("max_2x2"); - auto lhs = builder.ConstantR2FromArray2D({ - {7.0f, 2.0f}, // row 0 - {3.0f, -4.0f}, // row 1 - }); - auto rhs = builder.ConstantR2FromArray2D({ - {5.0f, 6.0f}, // row 0 - {1.0f, -8.0f}, // row 1 - }); - auto max = builder.Max(lhs, rhs); + auto lhs = ConstantR2FromArray2D(&builder, { + {7.0f, 2.0f}, // row 0 + {3.0f, -4.0f}, // row 1 + }); + auto rhs = ConstantR2FromArray2D(&builder, { + {5.0f, 6.0f}, // row 0 + {1.0f, -8.0f}, // row 1 + }); + Max(lhs, rhs); std::unique_ptr expected = - Literal::CreateR2FromArray2D({{7.0f, 6.0f}, // row 0 - {3.0f, -4.0f}}); // row 1 + LiteralUtil::CreateR2FromArray2D({{7.0f, 6.0f}, // row 0 + {3.0f, -4.0f}}); // row 1 this->ComputeAndCompareLiteral(&builder, *expected, {}, ErrorSpec(1e-6)); } @@ -137,9 +137,9 @@ class TestLinspaceMaxParametric XlaBuilder builder( tensorflow::strings::Printf("max_%lldx%lld_linspace", rows, cols)); - auto lhs = builder.ConstantR2FromArray2D(*alhs); - auto rhs = builder.ConstantR2FromArray2D(*arhs); - auto max = builder.Max(lhs, rhs); + auto lhs = ConstantR2FromArray2D(&builder, *alhs); + auto rhs = ConstantR2FromArray2D(&builder, *arhs); + Max(lhs, rhs); Array2D expected(rows, cols); for (int row = 0; row < rows; ++row) { @@ -200,31 +200,33 @@ class MatOpsDotAddTest TF_ASSERT_OK_AND_ASSIGN( auto lhs_handle, - client_->TransferToServer(*Literal::CreateR2FromArray2DWithLayout( - lhs, LayoutUtil::MakeLayout(minor_to_major(row_major))))); + client_->TransferToServer( + *LiteralUtil::CreateR2FromArray2DWithLayout( + lhs, LayoutUtil::MakeLayout(minor_to_major(row_major))))); TF_ASSERT_OK_AND_ASSIGN( auto rhs_handle, - client_->TransferToServer(*Literal::CreateR2FromArray2DWithLayout( - rhs, LayoutUtil::MakeLayout(minor_to_major(row_major))))); + client_->TransferToServer( + *LiteralUtil::CreateR2FromArray2DWithLayout( + rhs, LayoutUtil::MakeLayout(minor_to_major(row_major))))); XlaBuilder builder(TestName()); - auto lhs_arg = builder.Parameter(0, lhs_shape, "lhs"); + auto lhs_arg = Parameter(&builder, 0, lhs_shape, "lhs"); auto lhs_mat_arg = lhs_arg; if (transpose) { - lhs_mat_arg = builder.Transpose(lhs_mat_arg, {1, 0}); + lhs_mat_arg = Transpose(lhs_mat_arg, {1, 0}); } - auto rhs_arg = builder.Parameter(1, rhs_shape, "rhs"); - auto result = builder.Dot(lhs_mat_arg, rhs_arg); + auto rhs_arg = Parameter(&builder, 1, rhs_shape, "rhs"); + auto result = Dot(lhs_mat_arg, rhs_arg); Array2D expected; if (add_lhs) { - result = builder.Add(result, lhs_arg); + result = Add(result, lhs_arg); if (transpose) { expected = Array2D({{47.0f, 52.0f}, {71.0f, 78.0f}}); } else { expected = Array2D({{35.0f, 39.0f}, {81.0f, 89.0f}}); } } else { - result = builder.Add(result, rhs_arg); + result = Add(result, rhs_arg); if (transpose) { expected = Array2D({{56.0f, 61.0f}, {80.0f, 87.0f}}); } else { diff --git a/tensorflow/compiler/xla/tests/multidimensional_slice_test.cc b/tensorflow/compiler/xla/tests/multidimensional_slice_test.cc index 0791a71aacf7614286fe964623a3172a174d4722..e576f000ef23e761d6fa818457eec2144d4bcb00 100644 --- a/tensorflow/compiler/xla/tests/multidimensional_slice_test.cc +++ b/tensorflow/compiler/xla/tests/multidimensional_slice_test.cc @@ -33,9 +33,10 @@ class SliceTest : public ClientLibraryTestBase {}; XLA_TEST_F(SliceTest, Slice2D) { XlaBuilder builder("slice_2d"); - auto original = builder.ConstantR2( + auto original = ConstantR2( + &builder, {{1.0, 2.0, 3.0}, {4.0, 5.0, 6.0}, {7.0, 8.0, 9.0}, {10.0, 11.0, 12.0}}); - builder.Slice(original, {2, 1}, {4, 3}, {1, 1}); + Slice(original, {2, 1}, {4, 3}, {1, 1}); Array2D expected({{8.0f, 9.0f}, {11.0f, 12.0f}}); ComputeAndCompareR2(&builder, expected, {}, ErrorSpec(0.000001)); @@ -45,8 +46,8 @@ XLA_TEST_F(SliceTest, Slice3D) { XlaBuilder builder("slice_3d"); Array3D array_3d( {{{1.0f, 2.0f}, {3.0f, 4.0f}}, {{5.0f, 6.0f}, {7.0f, 8.0f}}}); - auto original = builder.ConstantR3FromArray3D(array_3d); - builder.Slice(original, {0, 0, 1}, {2, 1, 2}, {1, 1, 1}); + auto original = ConstantR3FromArray3D(&builder, array_3d); + Slice(original, {0, 0, 1}, {2, 1, 2}, {1, 1, 1}); Array3D expected_3d({{{2.0f}}, {{6.0f}}}); ComputeAndCompareR3(&builder, expected_3d, {}, ErrorSpec(0.000001)); diff --git a/tensorflow/compiler/xla/tests/multioutput_fusion_test.cc b/tensorflow/compiler/xla/tests/multioutput_fusion_test.cc index a42a19af15e87bd58d16294d012ec4db31e90070..eb06b115daa96bccd73de30bb7fa30733a6fd947 100644 --- a/tensorflow/compiler/xla/tests/multioutput_fusion_test.cc +++ b/tensorflow/compiler/xla/tests/multioutput_fusion_test.cc @@ -20,7 +20,7 @@ limitations under the License. #include #include "tensorflow/compiler/xla/client/local_client.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/primitive_util.h" #include "tensorflow/compiler/xla/ptr_util.h" #include "tensorflow/compiler/xla/service/hlo_computation.h" @@ -60,7 +60,7 @@ class MultiOutputFusionTest : public HloTestBase { const Shape elem_shape2 = ShapeUtil::MakeShape(F32, {size, size}); auto const0 = builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(8.0f))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(8.0f))); auto param0 = builder.AddInstruction( HloInstruction::CreateParameter(0, elem_shape0, "0")); @@ -105,8 +105,9 @@ class MultiOutputFusionTest : public HloTestBase { Literal expect(ShapeUtil::MakeShape(F32, {size, size})); expect.PopulateWithValue(size * 1.5f * 3.5f); - auto actual = ExecuteAndTransfer( - std::move(hlo_module), {Literal::CreateR0(-9.0f).get(), &arg1}); + auto actual = + ExecuteAndTransfer(std::move(hlo_module), + {LiteralUtil::CreateR0(-9.0f).get(), &arg1}); EXPECT_TRUE(LiteralTestUtil::Near(expect, *actual, error_spec_)); } @@ -165,7 +166,8 @@ class MultiOutputFusionTest : public HloTestBase { Literal input1(ShapeUtil::MakeShape(F64, {size})); input1.PopulateWithValue(1.); - Literal expect = std::move(*Literal::CreateR1({size * 1.5f * 3.5f})); + Literal expect = + std::move(*LiteralUtil::CreateR1({size * 1.5f * 3.5f})); auto actual = ExecuteAndTransfer(std::move(hlo_module), {&input0, &input1}); EXPECT_TRUE(LiteralTestUtil::Near(expect, *actual, error_spec_)); } @@ -198,16 +200,16 @@ XLA_TEST_F(MultiOutputFusionTest, FusionNodeIsRoot) { auto module = HloRunner::CreateModuleFromString(testcase, GetDebugOptionsForTest()) .ValueOrDie(); - auto param = Literal::MakeTupleOwned( - Literal::MakeTupleOwned( - Literal::MakeTupleOwned(Literal::CreateR0(42)), - Literal::CreateR0(1.0)), - Literal::MakeTupleOwned(Literal::CreateR0(3.0), - Literal::CreateR0(4))); + auto param = LiteralUtil::MakeTupleOwned( + LiteralUtil::MakeTupleOwned( + LiteralUtil::MakeTupleOwned(LiteralUtil::CreateR0(42)), + LiteralUtil::CreateR0(1.0)), + LiteralUtil::MakeTupleOwned(LiteralUtil::CreateR0(3.0), + LiteralUtil::CreateR0(4))); std::unique_ptr result = ExecuteNoHloPasses(std::move(module), {param.get()}); EXPECT_TRUE(LiteralTestUtil::Equal( - *Literal::MakeTupleOwned(Literal::CreateR0(42)), *result)); + *LiteralUtil::MakeTupleOwned(LiteralUtil::CreateR0(42)), *result)); } XLA_TEST_F(MultiOutputFusionTest, MultiOutputLoopFusion) { @@ -232,7 +234,7 @@ XLA_TEST_F(MultiOutputFusionTest, MultiOutputLoopFusion) { auto module = HloRunner::CreateModuleFromString(testcase, GetDebugOptionsForTest()) .ValueOrDie(); - auto param = Literal::CreateR1({1.0, 2.0, 3.0, -1.0}); + auto param = LiteralUtil::CreateR1({1.0, 2.0, 3.0, -1.0}); std::unique_ptr result = ExecuteNoHloPasses(std::move(module), {param.get()}); LiteralTestUtil::ExpectR1Equal({0.0, 4.0, 9.0, 1.0}, *result); @@ -265,7 +267,7 @@ XLA_TEST_F(MultiOutputFusionTest, MultiOutputLoopFeedingMap) { auto module = HloRunner::CreateModuleFromString(testcase, GetDebugOptionsForTest()) .ValueOrDie(); - auto param = Literal::CreateR1({1.0, 2.0, 3.0}); + auto param = LiteralUtil::CreateR1({1.0, 2.0, 3.0}); std::unique_ptr result = ExecuteNoHloPasses(std::move(module), {param.get()}); LiteralTestUtil::ExpectR1Equal({0.0, 4.0, 9.0}, *result); @@ -308,12 +310,14 @@ XLA_TEST_F(MultiOutputFusionTest, auto module = HloRunner::CreateModuleFromString(testcase, GetDebugOptionsForTest()) .ValueOrDie(); - auto param = Literal::CreateR3({{{1, 2}, {3, 4}}, {{5, 6}, {7, 8}}}); + auto param = + LiteralUtil::CreateR3({{{1, 2}, {3, 4}}, {{5, 6}, {7, 8}}}); std::unique_ptr result = ExecuteNoHloPasses(std::move(module), {param.get()}); EXPECT_TRUE(LiteralTestUtil::Equal( - *Literal::MakeTupleOwned(Literal::CreateR2({{3, 7}, {11, 15}}), - Literal::CreateR2({{5, 16}, {36, 64}})), + *LiteralUtil::MakeTupleOwned( + LiteralUtil::CreateR2({{3, 7}, {11, 15}}), + LiteralUtil::CreateR2({{5, 16}, {36, 64}})), *result)); } @@ -338,12 +342,14 @@ XLA_TEST_F(MultiOutputFusionTest, auto module = HloRunner::CreateModuleFromString(testcase, GetDebugOptionsForTest()) .ValueOrDie(); - auto param = Literal::CreateR3({{{1, 2}, {3, 4}}, {{5, 6}, {7, 8}}}); + auto param = + LiteralUtil::CreateR3({{{1, 2}, {3, 4}}, {{5, 6}, {7, 8}}}); std::unique_ptr result = ExecuteNoHloPasses(std::move(module), {param.get()}); EXPECT_TRUE(LiteralTestUtil::Equal( - *Literal::MakeTupleOwned(Literal::CreateR2({{6, 8}, {10, 12}}), - Literal::CreateR2({{25, 36}, {49, 64}})), + *LiteralUtil::MakeTupleOwned( + LiteralUtil::CreateR2({{6, 8}, {10, 12}}), + LiteralUtil::CreateR2({{25, 36}, {49, 64}})), *result)); } @@ -369,13 +375,14 @@ XLA_TEST_F(MultiOutputFusionTest, auto module = HloRunner::CreateModuleFromString(testcase, GetDebugOptionsForTest()) .ValueOrDie(); - auto param = Literal::CreateR3({{{1, 2}, {3, 4}}, {{5, 6}, {7, 8}}}); + auto param = + LiteralUtil::CreateR3({{{1, 2}, {3, 4}}, {{5, 6}, {7, 8}}}); std::unique_ptr result = ExecuteNoHloPasses(std::move(module), {param.get()}); EXPECT_TRUE(LiteralTestUtil::Equal( - *Literal::MakeTupleOwned(Literal::CreateR1({14, 22}), - Literal::CreateR1({36, 64}), - Literal::CreateR1({66, 138})), + *LiteralUtil::MakeTupleOwned(LiteralUtil::CreateR1({14, 22}), + LiteralUtil::CreateR1({36, 64}), + LiteralUtil::CreateR1({66, 138})), *result)); } @@ -401,14 +408,15 @@ XLA_TEST_F(MultiOutputFusionTest, auto module = HloRunner::CreateModuleFromString(testcase, GetDebugOptionsForTest()) .ValueOrDie(); - auto param = Literal::CreateR3({{{1, 2}, {3, 4}}, {{5, 6}, {7, 8}}}); + auto param = + LiteralUtil::CreateR3({{{1, 2}, {3, 4}}, {{5, 6}, {7, 8}}}); std::unique_ptr result = ExecuteNoHloPasses(std::move(module), {param.get()}); EXPECT_TRUE(LiteralTestUtil::Equal( - *Literal::MakeTupleOwned( - Literal::CreateR3({{{1, 2}, {3, 4}}, {{5, 6}, {7, 8}}}), - Literal::CreateR2({{3, 7}, {11, 15}}), - Literal::CreateR2({{5, 16}, {36, 64}})), + *LiteralUtil::MakeTupleOwned( + LiteralUtil::CreateR3({{{1, 2}, {3, 4}}, {{5, 6}, {7, 8}}}), + LiteralUtil::CreateR2({{3, 7}, {11, 15}}), + LiteralUtil::CreateR2({{5, 16}, {36, 64}})), *result)); } @@ -434,14 +442,16 @@ XLA_TEST_F(MultiOutputFusionTest, auto module = HloRunner::CreateModuleFromString(testcase, GetDebugOptionsForTest()) .ValueOrDie(); - auto param = Literal::CreateR3({{{1, 2}, {3, 4}}, {{5, 6}, {7, 8}}}); + auto param = + LiteralUtil::CreateR3({{{1, 2}, {3, 4}}, {{5, 6}, {7, 8}}}); std::unique_ptr result = ExecuteNoHloPasses(std::move(module), {param.get()}); EXPECT_TRUE(LiteralTestUtil::Equal( - *Literal::MakeTupleOwned( - Literal::CreateR2({{6, 8}, {10, 12}}), - Literal::CreateR3({{{1, 4}, {9, 16}}, {{25, 36}, {49, 64}}}), - Literal::CreateR2({{25, 36}, {49, 64}})), + *LiteralUtil::MakeTupleOwned( + LiteralUtil::CreateR2({{6, 8}, {10, 12}}), + LiteralUtil::CreateR3( + {{{1, 4}, {9, 16}}, {{25, 36}, {49, 64}}}), + LiteralUtil::CreateR2({{25, 36}, {49, 64}})), *result)); } @@ -454,7 +464,8 @@ XLA_TEST_F(MultiOutputFusionTest, r1 = f32[2]{0} reduce(p0, c0), dimensions={0,2}, to_apply=Add mul = f32[2,2,2]{2,1,0} multiply(p0, p0) c1 = f32[] constant(5) - mul2 = f32[2,2,2]{2,1,0} multiply(p0, c1) + b1 = f32[2,2,2]{2,1,0} broadcast(c1), dimensions={} + mul2 = f32[2,2,2]{2,1,0} multiply(p0, b1) ROOT tuple = (f32[2]{0}, f32[2,2,2]{2,1,0}, f32[2,2,2]{2,1,0}) tuple(r1, mul, mul2) } @@ -467,14 +478,16 @@ XLA_TEST_F(MultiOutputFusionTest, auto module = HloRunner::CreateModuleFromString(testcase, GetDebugOptionsForTest()) .ValueOrDie(); - auto param = Literal::CreateR3({{{1, 2}, {3, 4}}, {{5, 6}, {7, 8}}}); + auto param = + LiteralUtil::CreateR3({{{1, 2}, {3, 4}}, {{5, 6}, {7, 8}}}); std::unique_ptr result = ExecuteNoHloPasses(std::move(module), {param.get()}); EXPECT_TRUE(LiteralTestUtil::Equal( - *Literal::MakeTupleOwned( - Literal::CreateR1({14, 22}), - Literal::CreateR3({{{1, 4}, {9, 16}}, {{25, 36}, {49, 64}}}), - Literal::CreateR3( + *LiteralUtil::MakeTupleOwned( + LiteralUtil::CreateR1({14, 22}), + LiteralUtil::CreateR3( + {{{1, 4}, {9, 16}}, {{25, 36}, {49, 64}}}), + LiteralUtil::CreateR3( {{{5, 10}, {15, 20}}, {{25, 30}, {35, 40}}})), *result)); } @@ -501,15 +514,16 @@ XLA_TEST_F(MultiOutputFusionTest, auto module = HloRunner::CreateModuleFromString(testcase, GetDebugOptionsForTest()) .ValueOrDie(); - auto param = Literal::CreateR3({{{0, 2}, {3, 4}}, {{5, 6}, {7, 8}}}); - auto init1 = Literal::CreateR0(5); - auto init2 = Literal::CreateR0(6); + auto param = + LiteralUtil::CreateR3({{{0, 2}, {3, 4}}, {{5, 6}, {7, 8}}}); + auto init1 = LiteralUtil::CreateR0(5); + auto init2 = LiteralUtil::CreateR0(6); std::unique_ptr result = ExecuteNoHloPasses( std::move(module), {param.get(), init1.get(), init2.get()}); EXPECT_TRUE(LiteralTestUtil::Equal( - *Literal::MakeTupleOwned( - Literal::CreateR2({{167, 172}, {176, 180}}), - Literal::CreateR2({{6, 6}, {6, 8}})), + *LiteralUtil::MakeTupleOwned( + LiteralUtil::CreateR2({{167, 172}, {176, 180}}), + LiteralUtil::CreateR2({{6, 6}, {6, 8}})), *result)); } @@ -536,19 +550,20 @@ XLA_TEST_F(MultiOutputFusionTest, auto module = HloRunner::CreateModuleFromString(testcase, GetDebugOptionsForTest()) .ValueOrDie(); - auto param = Literal::CreateR3( + auto param = LiteralUtil::CreateR3( {{{Eigen::half(1), Eigen::half(2)}, {Eigen::half(3), Eigen::half(4)}}, {{Eigen::half(5), Eigen::half(6)}, {Eigen::half(7), Eigen::half(8)}}}); std::unique_ptr result = ExecuteNoHloPasses(std::move(module), {param.get()}); EXPECT_TRUE(LiteralTestUtil::Equal( - *Literal::MakeTupleOwned( - Literal::CreateR2({{3, 7}, {11, 15}}), - Literal::CreateR2({{5, 16}, {36, 64}}), - Literal::CreateR3({{{Eigen::half(1), Eigen::half(2)}, - {Eigen::half(3), Eigen::half(4)}}, - {{Eigen::half(5), Eigen::half(6)}, - {Eigen::half(7), Eigen::half(8)}}})), + *LiteralUtil::MakeTupleOwned( + LiteralUtil::CreateR2({{3, 7}, {11, 15}}), + LiteralUtil::CreateR2({{5, 16}, {36, 64}}), + LiteralUtil::CreateR3( + {{{Eigen::half(1), Eigen::half(2)}, + {Eigen::half(3), Eigen::half(4)}}, + {{Eigen::half(5), Eigen::half(6)}, + {Eigen::half(7), Eigen::half(8)}}})), *result)); } diff --git a/tensorflow/compiler/xla/tests/pad_test.cc b/tensorflow/compiler/xla/tests/pad_test.cc index ce295b832d79e4f00656f2893c2ba1162693dd73..e428fa9b5e14d0cb6e5610a1b69b07c6b0c9952a 100644 --- a/tensorflow/compiler/xla/tests/pad_test.cc +++ b/tensorflow/compiler/xla/tests/pad_test.cc @@ -93,8 +93,8 @@ XLA_TEST_P(PadTestFloat, Pad1DS0ToS0Array) { dimension->set_edge_padding_high(0); dimension->set_interior_padding(0); - b.Pad(AddParam(*Literal::CreateR1({}), &b), - AddParam(*Literal::CreateR0(0.1), &b), padding_config); + Pad(AddParam(*LiteralUtil::CreateR1({}), &b), + AddParam(*LiteralUtil::CreateR0(0.1), &b), padding_config); ComputeAndCompareR1(&b, {}, {}, DefaultErrorSpec()); } @@ -108,8 +108,8 @@ XLA_TEST_P(PadTestFloat, Pad1DS0ToS5Array) { dimension->set_edge_padding_high(4); dimension->set_interior_padding(7); - b.Pad(AddParam(*Literal::CreateR1({}), &b), - AddParam(*Literal::CreateR0(0.1), &b), padding_config); + Pad(AddParam(*LiteralUtil::CreateR1({}), &b), + AddParam(*LiteralUtil::CreateR0(0.1), &b), padding_config); ComputeAndCompareR1(&b, std::vector(5, 0.1), {}, DefaultErrorSpec()); } @@ -123,16 +123,17 @@ XLA_TEST_P(PadTestFloat, Pad1DS3Array) { dimension->set_edge_padding_high(0); dimension->set_interior_padding(1); - b.Pad(AddParam(*Literal::CreateR1({1, 2, 3}), &b), - AddParam(*Literal::CreateR0(0.1), &b), padding_config); + Pad(AddParam(*LiteralUtil::CreateR1({1, 2, 3}), &b), + AddParam(*LiteralUtil::CreateR0(0.1), &b), padding_config); std::vector expected({0.1, 0.1, 0.1, 1, 0.1, 2, 0.1, 3}); ComputeAndCompareR1(&b, expected, {}, DefaultErrorSpec()); } XLA_TEST_P(PadTestFloat, Pad4D_2x0x3x2_FloatArray) { XlaBuilder b(TestName()); - b.Pad(AddParam(Array4D(2, 0, 3, 2), &b), - AddParam(*Literal::CreateR0(1.5), &b), r4_padding_on_dim0_dim1_); + Pad(AddParam(Array4D(2, 0, 3, 2), &b), + AddParam(*LiteralUtil::CreateR0(1.5), &b), + r4_padding_on_dim0_dim1_); ComputeAndCompareR4(&b, Array4D(5, 2, 3, 2, 1.5f), {}, DefaultErrorSpec()); } @@ -147,8 +148,8 @@ TEST_P(PadTestFloat, Pad4DFloat_1x1x3x2_Array) { }); input->FillWithYX(input_xy); - b.Pad(AddParam(*input, &b), AddParam(*Literal::CreateR0(1.5), &b), - r4_padding_on_dim0_dim1_); + Pad(AddParam(*input, &b), AddParam(*LiteralUtil::CreateR0(1.5), &b), + r4_padding_on_dim0_dim1_); auto expected = MakeUnique>(2, 3, 3, 2); expected->Fill(1.5); @@ -166,8 +167,9 @@ TEST_P(PadTestFloat, Pad4DFloatArrayWithInteriorPadding) { const float pad_value = 1.5f; Array4D input(3, 2, 1, 1, {1, 2, 3, 4, 5, 6}); - b.Pad(AddParam(input, &b), AddParam(*Literal::CreateR0(pad_value), &b), - r4_padding_on_dim0_dim1_); + Pad(AddParam(input, &b), + AddParam(*LiteralUtil::CreateR0(pad_value), &b), + r4_padding_on_dim0_dim1_); auto expected = MakeUnique>(8, 5, 1, 1); expected->Fill(pad_value); @@ -205,11 +207,11 @@ TEST_P(PadTestFloat, Pad4DFloatArrayMinorFirstSmall) { const float pad_value = -5.123f; Array4D input_array(1, 1, 2, 3, {1, 2, 3, 4, 5, 6}); - auto input = Literal::CreateR4FromArray4D(input_array); + auto input = LiteralUtil::CreateR4FromArray4D(input_array); input = input->Relayout(layout); - b.Pad(AddParam(*input, &b), - AddParam(*Literal::CreateR0(pad_value), &b), padding_config); + Pad(AddParam(*input, &b), + AddParam(*LiteralUtil::CreateR0(pad_value), &b), padding_config); Array4D expected_array(1, 1, 5, 8); expected_array.Fill(pad_value); @@ -251,11 +253,11 @@ XLA_TEST_P(PadTestFloat, Pad4DFloatArrayMinorFirstNonTrivialMinorDimensions) { input_array(0, 0, 0, 0) = 1.0f; input_array(0, 24, 6, 6) = 2.0f; input_array(0, 17, 2, 5) = 3.0f; - auto input = Literal::CreateR4FromArray4D(input_array); + auto input = LiteralUtil::CreateR4FromArray4D(input_array); input = input->Relayout(layout); - b.Pad(AddParam(*input, &b), - AddParam(*Literal::CreateR0(pad_value), &b), padding_config); + Pad(AddParam(*input, &b), + AddParam(*LiteralUtil::CreateR0(pad_value), &b), padding_config); Array4D expected_array(1, 25, 17, 11); expected_array.Fill(pad_value); @@ -275,8 +277,8 @@ XLA_TEST_F(PadTest, Pad4DU8Array) { }); input->FillWithYX(input_xy); - b.Pad(AddParam(*input, &b), b.ConstantR0(35), - r4_padding_on_dim0_dim1_); + Pad(AddParam(*input, &b), ConstantR0(&b, 35), + r4_padding_on_dim0_dim1_); auto expected = MakeUnique>(2, 3, 3, 2); expected->Fill(35); @@ -294,16 +296,16 @@ XLA_TEST_F(PadTest, Pad4DPredArray) { // Since bool is currently not well supported, use Broadcast operation to // create the operand for Pad. - auto input = b.Broadcast(b.ConstantR0(true), {1, 1, 3, 2}); + auto input = Broadcast(ConstantR0(&b, true), {1, 1, 3, 2}); auto padded = - b.Pad(input, b.ConstantR0(false), r4_padding_on_dim0_dim1_); + Pad(input, ConstantR0(&b, false), r4_padding_on_dim0_dim1_); // For the same reason, use Select to convert boolean values to int32. auto zeros = MakeUnique>(2, 3, 3, 2); auto ones = MakeUnique>(2, 3, 3, 2); zeros->Fill(0); ones->Fill(1); - b.Select(padded, AddParam(*ones, &b), AddParam(*zeros, &b)); + Select(padded, AddParam(*ones, &b), AddParam(*zeros, &b)); auto expected = MakeUnique>(2, 3, 3, 2); expected->Fill(0); @@ -329,7 +331,7 @@ XLA_TEST_P(PadTestFloat, Large2DPad) { padding_config.mutable_dimensions(dim)->set_edge_padding_high(58 + 100 * dim); } - b.Pad(input, AddParam(*Literal::CreateR0(0.0f), &b), padding_config); + Pad(input, AddParam(*LiteralUtil::CreateR0(0.0f), &b), padding_config); auto expected = ReferenceUtil::PadArray2D(*ones, padding_config, 0.0f); ComputeAndCompareR2(&b, *expected, {}, DefaultErrorSpec()); @@ -351,7 +353,8 @@ XLA_TEST_P(PadTestFloat, AllTypes2DPad) { padding_config.mutable_dimensions(1)->set_edge_padding_low(6); padding_config.mutable_dimensions(1)->set_edge_padding_high(4); padding_config.mutable_dimensions(1)->set_interior_padding(2); - b.Pad(input, AddParam(*Literal::CreateR0(3.14f), &b), padding_config); + Pad(input, AddParam(*LiteralUtil::CreateR0(3.14f), &b), + padding_config); auto expected = ReferenceUtil::PadArray2D(*operand, padding_config, 3.14f); ComputeAndCompareR2(&b, *expected, {}, DefaultErrorSpec()); @@ -376,7 +379,8 @@ XLA_TEST_P(PadTestFloat, High2DPad) { padding_config.mutable_dimensions(dim)->set_interior_padding( interior_padding); } - b.Pad(input, AddParam(*Literal::CreateR0(2.718f), &b), padding_config); + Pad(input, AddParam(*LiteralUtil::CreateR0(2.718f), &b), + padding_config); auto expected = ReferenceUtil::PadArray2D(*operand, padding_config, 2.718f); @@ -403,7 +407,8 @@ XLA_TEST_P(PadTestFloat, NegativePadding2D) { padding_config.mutable_dimensions(dim)->set_interior_padding( interior_padding); } - b.Pad(input, AddParam(*Literal::CreateR0(2.718f), &b), padding_config); + Pad(input, AddParam(*LiteralUtil::CreateR0(2.718f), &b), + padding_config); auto expected = ReferenceUtil::PadArray2D(*operand, padding_config, 2.718f); @@ -430,7 +435,8 @@ XLA_TEST_P(PadTestFloat, NegativeAndInteriorPadding2D) { padding_config.mutable_dimensions(dim)->set_interior_padding( interior_padding[dim]); } - b.Pad(input, AddParam(*Literal::CreateR0(2.718f), &b), padding_config); + Pad(input, AddParam(*LiteralUtil::CreateR0(2.718f), &b), + padding_config); auto expected = ReferenceUtil::PadArray2D(*operand, padding_config, 2.718f); @@ -446,12 +452,13 @@ XLA_TEST_P(PadTestFloat, ReducePad) { XlaComputation add = CreateScalarAddComputation(FloatType(), &b); auto reduce = - b.Reduce(input, AddParam(*Literal::CreateR0(0.0), &b), add, {0}); + Reduce(input, AddParam(*LiteralUtil::CreateR0(0.0), &b), add, {0}); PaddingConfig padding_config = MakeNoPaddingConfig(3); padding_config.mutable_dimensions(0)->set_edge_padding_low(1); padding_config.mutable_dimensions(0)->set_edge_padding_high(1); - b.Pad(reduce, AddParam(*Literal::CreateR0(0.0f), &b), padding_config); + Pad(reduce, AddParam(*LiteralUtil::CreateR0(0.0f), &b), + padding_config); Array3D expected({{{0.0, 0.0}, {0.0, 0.0}}, {{2.0, 2.0}, {2.0, 2.0}}, diff --git a/tensorflow/compiler/xla/tests/params_test.cc b/tensorflow/compiler/xla/tests/params_test.cc index 838f1b4e2f0f0e0871ec717bdeefcbbc653397e3..8ba1d11b33344463ffcb059f453754f31e177184 100644 --- a/tensorflow/compiler/xla/tests/params_test.cc +++ b/tensorflow/compiler/xla/tests/params_test.cc @@ -24,7 +24,7 @@ limitations under the License. #include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/compiler/xla/client/xla_client/xla_computation.h" #include "tensorflow/compiler/xla/layout_util.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/shape_util.h" #include "tensorflow/compiler/xla/statusor.h" #include "tensorflow/compiler/xla/tests/client_library_test_base.h" @@ -42,11 +42,12 @@ class ParamsTest : public ClientLibraryTestBase {}; XLA_TEST_F(ParamsTest, ConstantR0F32Param) { XlaBuilder builder(TestName()); - std::unique_ptr param0_literal = Literal::CreateR0(3.14159f); + std::unique_ptr param0_literal = + LiteralUtil::CreateR0(3.14159f); std::unique_ptr param0_data = client_->TransferToServer(*param0_literal).ConsumeValueOrDie(); - auto p = builder.Parameter(0, ShapeUtil::MakeShape(F32, {}), "param0"); + Parameter(&builder, 0, ShapeUtil::MakeShape(F32, {}), "param0"); ComputeAndCompareR0(&builder, 3.14159f, {param0_data.get()}, ErrorSpec(0.0001f)); @@ -54,11 +55,11 @@ XLA_TEST_F(ParamsTest, ConstantR0F32Param) { XLA_TEST_F(ParamsTest, ConstantR1S0F32Param) { XlaBuilder builder(TestName()); - std::unique_ptr param0_literal = Literal::CreateR1({}); + std::unique_ptr param0_literal = LiteralUtil::CreateR1({}); std::unique_ptr param0_data = client_->TransferToServer(*param0_literal).ConsumeValueOrDie(); - auto p = builder.Parameter(0, ShapeUtil::MakeShape(F32, {0}), "param0"); + Parameter(&builder, 0, ShapeUtil::MakeShape(F32, {0}), "param0"); ComputeAndCompareR1(&builder, {}, {param0_data.get()}, ErrorSpec(0.01f)); @@ -67,11 +68,11 @@ XLA_TEST_F(ParamsTest, ConstantR1S0F32Param) { XLA_TEST_F(ParamsTest, ConstantR1S2F32Param) { XlaBuilder builder(TestName()); std::unique_ptr param0_literal = - Literal::CreateR1({3.14f, -100.25f}); + LiteralUtil::CreateR1({3.14f, -100.25f}); std::unique_ptr param0_data = client_->TransferToServer(*param0_literal).ConsumeValueOrDie(); - auto p = builder.Parameter(0, ShapeUtil::MakeShape(F32, {2}), "param0"); + Parameter(&builder, 0, ShapeUtil::MakeShape(F32, {2}), "param0"); ComputeAndCompareR1(&builder, {3.14f, -100.25f}, {param0_data.get()}, ErrorSpec(0.01f)); @@ -80,12 +81,13 @@ XLA_TEST_F(ParamsTest, ConstantR1S2F32Param) { XLA_TEST_F(ParamsTest, ConstantR1U8Param) { XlaBuilder builder(TestName()); string str("hello world"); - std::unique_ptr param0_literal = Literal::CreateR1U8(str); + std::unique_ptr param0_literal = LiteralUtil::CreateR1U8(str); std::unique_ptr param0_data = client_->TransferToServer(*param0_literal).ConsumeValueOrDie(); - auto p = builder.Parameter( - 0, ShapeUtil::MakeShape(U8, {static_cast(str.size())}), "param0"); + Parameter(&builder, 0, + ShapeUtil::MakeShape(U8, {static_cast(str.size())}), + "param0"); ComputeAndCompareR1U8(&builder, str, {param0_data.get()}); } @@ -93,11 +95,11 @@ XLA_TEST_F(ParamsTest, ConstantR1U8Param) { XLA_TEST_F(ParamsTest, ConstantR2_3x0_F32Param) { XlaBuilder builder(TestName()); std::unique_ptr param0_literal = - Literal::CreateR2FromArray2D(Array2D(3, 0)); + LiteralUtil::CreateR2FromArray2D(Array2D(3, 0)); std::unique_ptr param0_data = client_->TransferToServer(*param0_literal).ConsumeValueOrDie(); - auto p = builder.Parameter(0, ShapeUtil::MakeShape(F32, {3, 0}), "param0"); + Parameter(&builder, 0, ShapeUtil::MakeShape(F32, {3, 0}), "param0"); ComputeAndCompareR2(&builder, Array2D(3, 0), {param0_data.get()}, ErrorSpec(0.01f)); @@ -105,12 +107,12 @@ XLA_TEST_F(ParamsTest, ConstantR2_3x0_F32Param) { XLA_TEST_F(ParamsTest, ConstantR2F32Param) { XlaBuilder builder(TestName()); - std::unique_ptr param0_literal = Literal::CreateR2( + std::unique_ptr param0_literal = LiteralUtil::CreateR2( {{3.14f, -100.25f}, {7e8f, 7e-9f}, {30.3f, -100.0f}}); std::unique_ptr param0_data = client_->TransferToServer(*param0_literal).ConsumeValueOrDie(); - auto p = builder.Parameter(0, ShapeUtil::MakeShape(F32, {3, 2}), "param0"); + Parameter(&builder, 0, ShapeUtil::MakeShape(F32, {3, 2}), "param0"); Array2D expected_array( {{3.14f, -100.25f}, {7e8f, 7e-9f}, {30.3f, -100.0f}}); @@ -121,28 +123,28 @@ XLA_TEST_F(ParamsTest, ConstantR2F32Param) { XLA_TEST_F(ParamsTest, TwoParameters) { XlaBuilder builder(TestName()); - std::unique_ptr literal0 = Literal::CreateR1({1, 2}); + std::unique_ptr literal0 = LiteralUtil::CreateR1({1, 2}); std::unique_ptr param0_data = client_->TransferToServer(*literal0).ConsumeValueOrDie(); - auto param0 = builder.Parameter(0, literal0->shape(), "param0"); + auto param0 = Parameter(&builder, 0, literal0->shape(), "param0"); - std::unique_ptr literal1 = Literal::CreateR1({10, 20}); + std::unique_ptr literal1 = LiteralUtil::CreateR1({10, 20}); std::unique_ptr param1_data = client_->TransferToServer(*literal1).ConsumeValueOrDie(); - auto param1 = builder.Parameter(1, literal1->shape(), "param1"); + auto param1 = Parameter(&builder, 1, literal1->shape(), "param1"); // Use both parameters // // {1, 2} + {10, 20} = {11, 22} - auto sum = builder.Add(param0, param1); - sum = builder.Add(param0, param1); + auto sum = Add(param0, param1); + sum = Add(param0, param1); // Use only the second parameter again, to show that it can be used // twice and to make the computation asymmetric in the two // parameters to test that the parameters are not swapped. // // {11, 22} * {10, 20} = {110, 440} - auto prod = builder.Mul(sum, param1); + Mul(sum, param1); ComputeAndCompareR1(&builder, {110, 440}, {param0_data.get(), param1_data.get()}, @@ -152,12 +154,12 @@ XLA_TEST_F(ParamsTest, TwoParameters) { XLA_TEST_F(ParamsTest, MissingParameter) { // Test that an error is returned when a computation with an incomplete set of // parameters (parameter numbers not contiguous from 0) is executed. - std::unique_ptr literal = Literal::CreateR0(3.14159f); + std::unique_ptr literal = LiteralUtil::CreateR0(3.14159f); std::unique_ptr data = client_->TransferToServer(*literal).ConsumeValueOrDie(); XlaBuilder builder(TestName()); - auto p = builder.Parameter(2, ShapeUtil::MakeShape(F32, {}), "param2"); + Parameter(&builder, 2, ShapeUtil::MakeShape(F32, {}), "param2"); auto computation_status = builder.Build(); ASSERT_NE(computation_status.status(), Status::OK()); @@ -166,15 +168,15 @@ XLA_TEST_F(ParamsTest, MissingParameter) { XLA_TEST_F(ParamsTest, UnusedParameter) { XlaBuilder builder(TestName()); - std::unique_ptr literal0 = Literal::CreateR1({1, 2}); + std::unique_ptr literal0 = LiteralUtil::CreateR1({1, 2}); std::unique_ptr param0_data = client_->TransferToServer(*literal0).ConsumeValueOrDie(); - auto param0 = builder.Parameter(0, literal0->shape(), "param0"); + Parameter(&builder, 0, literal0->shape(), "param0"); - std::unique_ptr literal1 = Literal::CreateR1({10, 20}); + std::unique_ptr literal1 = LiteralUtil::CreateR1({10, 20}); std::unique_ptr param1_data = client_->TransferToServer(*literal1).ConsumeValueOrDie(); - auto param1 = builder.Parameter(1, literal1->shape(), "param1"); + Parameter(&builder, 1, literal1->shape(), "param1"); ComputeAndCompareR1(&builder, {10, 20}, {param0_data.get(), param1_data.get()}, @@ -186,22 +188,23 @@ XLA_TEST_F(ParamsTest, UnusedParametersInUnusedExpression) { // unused expression. XlaBuilder builder(TestName()); - std::unique_ptr literal0 = Literal::CreateR1({1, 2}); + std::unique_ptr literal0 = LiteralUtil::CreateR1({1, 2}); std::unique_ptr param0_data = client_->TransferToServer(*literal0).ConsumeValueOrDie(); - std::unique_ptr literal1 = Literal::CreateR1({10, 20, 30}); + std::unique_ptr literal1 = + LiteralUtil::CreateR1({10, 20, 30}); std::unique_ptr param1_data = client_->TransferToServer(*literal1).ConsumeValueOrDie(); - auto param0 = builder.Parameter(0, literal0->shape(), "param0"); - auto param1 = builder.Parameter(1, literal1->shape(), "param1"); - auto param2 = builder.Parameter(2, literal1->shape(), "param2"); + auto param0 = Parameter(&builder, 0, literal0->shape(), "param0"); + auto param1 = Parameter(&builder, 1, literal1->shape(), "param1"); + auto param2 = Parameter(&builder, 2, literal1->shape(), "param2"); // This add is unused. - builder.Add(param1, param2); + Add(param1, param2); - builder.Neg(param0); + Neg(param0); ComputeAndCompareR1( &builder, {-1, -2}, @@ -215,7 +218,7 @@ XLA_TEST_F(ParamsTest, HundredLargeR1Parameters) { std::vector init_value = {{0, 1}}; init_value.resize(size); - XlaOp sum_handle = builder.ConstantR1(init_value); + XlaOp sum_handle = ConstantR1(&builder, init_value); std::vector sum = {{0, 1}}; sum.resize(size); @@ -230,11 +233,11 @@ XLA_TEST_F(ParamsTest, HundredLargeR1Parameters) { std::vector sum_value = {{entry0, entry1}}; sum_value.resize(size); - std::unique_ptr literal = Literal::CreateR1(sum_value); + std::unique_ptr literal = LiteralUtil::CreateR1(sum_value); param_data_owner.push_back( client_->TransferToServer(*literal).ConsumeValueOrDie()); - XlaOp param = builder.Parameter(i, literal->shape(), "param"); - sum_handle = builder.Add(sum_handle, param); + XlaOp param = Parameter(&builder, i, literal->shape(), "param"); + sum_handle = Add(sum_handle, param); } std::vector param_data; @@ -260,16 +263,16 @@ XLA_TEST_F(ParamsTest, XlaBuilder builder(TestName()); std::vector> param_data_owner; - XlaOp sum_handle = builder.ConstantR0(0.0f); + XlaOp sum_handle = ConstantR0(&builder, 0.0f); float target = 0.0; constexpr int kParamCount = 3000; for (int i = 0; i < kParamCount; ++i) { target += i; - std::unique_ptr literal = Literal::CreateR0(i); + std::unique_ptr literal = LiteralUtil::CreateR0(i); param_data_owner.push_back( std::move(client_->TransferToServer(*literal)).ValueOrDie()); - XlaOp param = builder.Parameter(i, literal->shape(), "param"); - sum_handle = builder.Add(sum_handle, param); + XlaOp param = Parameter(&builder, i, literal->shape(), "param"); + sum_handle = Add(sum_handle, param); } std::vector param_data; @@ -291,26 +294,26 @@ XLA_TEST_F(ParamsTest, DISABLED_ON_CPU(DISABLED_ON_GPU( XlaBuilder builder(TestName()); std::vector> param_data_owner; - XlaOp sum_handle = builder.ConstantR1({0, 0}); + XlaOp sum_handle = ConstantR1(&builder, {0, 0}); int32 target = 0; constexpr int kParamCount = 3000; std::vector params; for (int i = 0; i < kParamCount; ++i) { target += i; - std::unique_ptr literal = Literal::CreateR1({i, i}); + std::unique_ptr literal = LiteralUtil::CreateR1({i, i}); param_data_owner.push_back( std::move(client_->TransferToServer(*literal)).ValueOrDie()); - XlaOp param = builder.Parameter(i, literal->shape(), "param"); + XlaOp param = Parameter(&builder, i, literal->shape(), "param"); params.push_back(param); - sum_handle = builder.Add(sum_handle, param); + sum_handle = Add(sum_handle, param); } std::vector outputs; for (int i = 0; i < kParamCount; ++i) { - outputs.push_back(builder.Add(params[i], sum_handle)); + outputs.push_back(Add(params[i], sum_handle)); } - builder.Tuple(outputs); + Tuple(&builder, outputs); std::vector param_data; param_data.reserve(param_data_owner.size()); @@ -321,10 +324,10 @@ XLA_TEST_F(ParamsTest, DISABLED_ON_CPU(DISABLED_ON_GPU( std::vector> elements; std::vector ptrs; for (int i = 0; i < kParamCount; ++i) { - elements.push_back(Literal::CreateR1({target + i, target + i})); + elements.push_back(LiteralUtil::CreateR1({target + i, target + i})); ptrs.push_back(elements.back().get()); } - ComputeAndCompareTuple(&builder, *Literal::MakeTuple(ptrs), param_data); + ComputeAndCompareTuple(&builder, *LiteralUtil::MakeTuple(ptrs), param_data); } // Test large number of parameters flowing into a while-loop. @@ -353,25 +356,25 @@ XLA_TEST_F(ParamsTest, std::vector params; std::vector parameter_shapes; for (int i = 0; i < kParamCount; ++i) { - std::unique_ptr literal = Literal::CreateR1({i, i}); + std::unique_ptr literal = LiteralUtil::CreateR1({i, i}); param_data_owner.push_back( std::move(client_->TransferToServer(*literal)).ValueOrDie()); - XlaOp param = builder.Parameter(i, literal->shape(), "param"); + XlaOp param = Parameter(&builder, i, literal->shape(), "param"); params.push_back(param); parameter_shapes.push_back(literal->shape()); } // Add bool parameter for the loop condition. Use a parameter HLO instead of a // constant because DCE may eliminate the while-body otherwise. - std::unique_ptr bool_literal = Literal::CreateR0(false); + std::unique_ptr bool_literal = LiteralUtil::CreateR0(false); param_data_owner.push_back( std::move(client_->TransferToServer(*bool_literal)).ValueOrDie()); XlaOp bool_param = - builder.Parameter(kParamCount, bool_literal->shape(), "bool_param"); + Parameter(&builder, kParamCount, bool_literal->shape(), "bool_param"); params.push_back(bool_param); parameter_shapes.push_back(bool_literal->shape()); - auto init = builder.Tuple(params); + auto init = Tuple(&builder, params); // Create a computation for the condition: while(bool_param). Shape while_shape = ShapeUtil::MakeTupleShape(parameter_shapes); @@ -379,8 +382,8 @@ XLA_TEST_F(ParamsTest, { XlaBuilder builder("condition"); auto condition_parameter = - builder.Parameter(0, while_shape, "condition_parameter"); - builder.GetTupleElement(condition_parameter, kParamCount); + Parameter(&builder, 0, while_shape, "condition_parameter"); + GetTupleElement(condition_parameter, kParamCount); condition = builder.Build().ConsumeValueOrDie(); } @@ -389,27 +392,27 @@ XLA_TEST_F(ParamsTest, XlaComputation body; { XlaBuilder builder("body"); - auto body_parameter = builder.Parameter(0, while_shape, "body_parameter"); + auto body_parameter = Parameter(&builder, 0, while_shape, "body_parameter"); std::vector updates; for (int i = 0; i < kParamCount; ++i) { - auto add = builder.Add(builder.GetTupleElement(body_parameter, i), - builder.ConstantR1({1, 1})); + auto add = Add(GetTupleElement(body_parameter, i), + ConstantR1(&builder, {1, 1})); updates.push_back(add); } // Add bool parameter. - updates.push_back(builder.GetTupleElement(body_parameter, kParamCount)); + updates.push_back(GetTupleElement(body_parameter, kParamCount)); - builder.Tuple(updates); + Tuple(&builder, updates); body = builder.Build().ConsumeValueOrDie(); } - auto loop = builder.While(condition, body, init); + auto loop = While(condition, body, init); std::vector outputs; for (int i = 0; i < kParamCount; ++i) { - outputs.push_back(builder.GetTupleElement(loop, i)); + outputs.push_back(GetTupleElement(loop, i)); } - builder.Tuple(outputs); + Tuple(&builder, outputs); std::vector param_data; param_data.reserve(param_data_owner.size()); @@ -420,10 +423,10 @@ XLA_TEST_F(ParamsTest, std::vector> elements; std::vector ptrs; for (int i = 0; i < kParamCount; ++i) { - elements.push_back(Literal::CreateR1({i, i})); + elements.push_back(LiteralUtil::CreateR1({i, i})); ptrs.push_back(elements.back().get()); } - ComputeAndCompareTuple(&builder, *Literal::MakeTuple(ptrs), param_data); + ComputeAndCompareTuple(&builder, *LiteralUtil::MakeTuple(ptrs), param_data); } #endif @@ -433,16 +436,16 @@ XLA_TEST_F(ParamsTest, TupleOfR1ParametersAddedTogether) { Shape r1f32_3 = ShapeUtil::MakeShape(F32, {3}); Shape tuple_shape = ShapeUtil::MakeTupleShape({r1f32_3, r1f32_3}); - auto input = builder.Parameter(0, tuple_shape, "input"); - auto lhs = builder.GetTupleElement(input, 0); - auto rhs = builder.GetTupleElement(input, 1); - builder.Add(lhs, rhs); + auto input = Parameter(&builder, 0, tuple_shape, "input"); + auto lhs = GetTupleElement(input, 0); + auto rhs = GetTupleElement(input, 1); + Add(lhs, rhs); std::unique_ptr data = client_ - ->TransferToServer(*Literal::MakeTuple({ - Literal::CreateR1({1, 2, 3}).get(), - Literal::CreateR1({4, 5, 6}).get(), + ->TransferToServer(*LiteralUtil::MakeTuple({ + LiteralUtil::CreateR1({1, 2, 3}).get(), + LiteralUtil::CreateR1({4, 5, 6}).get(), })) .ConsumeValueOrDie(); @@ -454,10 +457,10 @@ XLA_TEST_F(ParamsTest, TupleOfR1ParametersAddedTogether) { // Verifies that passing a 2x2 with {0, 1} layout returns the same value back // when (transferred to the server and) passed through a parameter. XLA_TEST_F(ParamsTest, R2_2x2_Layout_01) { - std::unique_ptr literal = Literal::CreateR2WithLayout( + std::unique_ptr literal = LiteralUtil::CreateR2WithLayout( {{1, 2}, {3, 4}}, LayoutUtil::MakeLayout({0, 1})); XlaBuilder builder(TestName()); - builder.Parameter(0, literal->shape(), "input"); + Parameter(&builder, 0, literal->shape(), "input"); std::unique_ptr data = client_->TransferToServer(*literal).ConsumeValueOrDie(); @@ -466,10 +469,10 @@ XLA_TEST_F(ParamsTest, R2_2x2_Layout_01) { // As above, but for {1, 0} layout. XLA_TEST_F(ParamsTest, R2_2x2_Layout_10) { - std::unique_ptr literal = Literal::CreateR2WithLayout( + std::unique_ptr literal = LiteralUtil::CreateR2WithLayout( {{1, 3}, {2, 4}}, LayoutUtil::MakeLayout({1, 0})); XlaBuilder builder(TestName()); - builder.Parameter(0, literal->shape(), "input"); + Parameter(&builder, 0, literal->shape(), "input"); std::unique_ptr data = client_->TransferToServer(*literal).ConsumeValueOrDie(); @@ -477,8 +480,9 @@ XLA_TEST_F(ParamsTest, R2_2x2_Layout_10) { } XLA_TEST_F(ParamsTest, R2_2x2_TryToPassReverseLayoutToParameter) { - std::unique_ptr literal = Literal::CreateR2({ - {1, 3}, {2, 4}, + std::unique_ptr literal = LiteralUtil::CreateR2({ + {1, 3}, + {2, 4}, }); const Shape original = literal->shape(); { @@ -494,9 +498,9 @@ XLA_TEST_F(ParamsTest, R2_2x2_TryToPassReverseLayoutToParameter) { } // Use the original shape in building the computation. XlaBuilder builder(TestName()); - auto input = builder.Parameter(0, original, "input"); + auto input = Parameter(&builder, 0, original, "input"); // Use the slice operator to get an off-diagonal element. - builder.Slice(input, {0, 1}, {1, 2}, {1, 1}); + Slice(input, {0, 1}, {1, 2}, {1, 1}); std::unique_ptr data = client_->TransferToServer(*literal).ConsumeValueOrDie(); diff --git a/tensorflow/compiler/xla/tests/pred_test.cc b/tensorflow/compiler/xla/tests/pred_test.cc index 77159efb26f3b7dd4918f24305f7269a2d6ff647..5c351b2d113709105244de4aafa49d7cc535ced1 100644 --- a/tensorflow/compiler/xla/tests/pred_test.cc +++ b/tensorflow/compiler/xla/tests/pred_test.cc @@ -29,64 +29,63 @@ namespace { class PredTest : public ClientLibraryTestBase { protected: - void TestCompare( - bool lhs, bool rhs, bool expected, - XlaOp (XlaBuilder::*op)(const xla::XlaOp&, const xla::XlaOp&, - tensorflow::gtl::ArraySlice)) { + void TestCompare(bool lhs, bool rhs, bool expected, + std::function)> + op) { XlaBuilder builder(TestName()); - XlaOp lhs_op = builder.ConstantR0(lhs); - XlaOp rhs_op = builder.ConstantR0(rhs); - XlaOp result = (builder.*op)(lhs_op, rhs_op, {}); + XlaOp lhs_op = ConstantR0(&builder, lhs); + XlaOp rhs_op = ConstantR0(&builder, rhs); + op(lhs_op, rhs_op, {}); ComputeAndCompareR0(&builder, expected, {}); } }; TEST_F(PredTest, ConstantR0PredTrue) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR0(true); + ConstantR0(&builder, true); ComputeAndCompareR0(&builder, true, {}); } TEST_F(PredTest, ConstantR0PredFalse) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR0(false); + ConstantR0(&builder, false); ComputeAndCompareR0(&builder, false, {}); } TEST_F(PredTest, ConstantR0PredCompareEq) { - TestCompare(true, false, false, &XlaBuilder::Eq); + TestCompare(true, false, false, &Eq); } TEST_F(PredTest, ConstantR0PredCompareNe) { - TestCompare(true, false, true, &XlaBuilder::Ne); + TestCompare(true, false, true, &Ne); } TEST_F(PredTest, ConstantR0PredCompareLe) { - TestCompare(true, false, false, &XlaBuilder::Le); + TestCompare(true, false, false, &Le); } TEST_F(PredTest, ConstantR0PredCompareLt) { - TestCompare(true, false, false, &XlaBuilder::Lt); + TestCompare(true, false, false, &Lt); } TEST_F(PredTest, ConstantR0PredCompareGe) { - TestCompare(true, false, true, &XlaBuilder::Ge); + TestCompare(true, false, true, &Ge); } TEST_F(PredTest, ConstantR0PredCompareGt) { - TestCompare(true, false, true, &XlaBuilder::Gt); + TestCompare(true, false, true, &Gt); } TEST_F(PredTest, ConstantR1Pred) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({true, false, false, true}); + ConstantR1(&builder, {true, false, false, true}); ComputeAndCompareR1(&builder, {true, false, false, true}, {}); } TEST_F(PredTest, ConstantR2Pred) { XlaBuilder builder(TestName()); - auto a = - builder.ConstantR2({{false, true, true}, {true, false, false}}); + ConstantR2(&builder, {{false, true, true}, {true, false, false}}); const string expected = R"(pred[2,3] { { 011 }, { 100 } @@ -96,44 +95,44 @@ TEST_F(PredTest, ConstantR2Pred) { TEST_F(PredTest, AnyR1True) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({true, false}); - TF_ASSERT_OK(Any(a, &builder).status()); + auto a = ConstantR1(&builder, {true, false}); + Any(a); ComputeAndCompareR0(&builder, true, {}); } TEST_F(PredTest, AnyR1False) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({false, false}); - TF_ASSERT_OK(Any(a, &builder).status()); + auto a = ConstantR1(&builder, {false, false}); + Any(a); ComputeAndCompareR0(&builder, false, {}); } TEST_F(PredTest, AnyR1VacuouslyFalse) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR1({}); - TF_ASSERT_OK(Any(a, &builder).status()); + auto a = ConstantR1(&builder, {}); + Any(a); ComputeAndCompareR0(&builder, false, {}); } TEST_F(PredTest, AnyR2True) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR2({ - {false, false, false}, - {false, false, false}, - {false, false, true}, - }); - TF_ASSERT_OK(Any(a, &builder).status()); + auto a = ConstantR2(&builder, { + {false, false, false}, + {false, false, false}, + {false, false, true}, + }); + Any(a); ComputeAndCompareR0(&builder, true, {}); } TEST_F(PredTest, AnyR2False) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR2({ - {false, false, false}, - {false, false, false}, - {false, false, false}, - }); - TF_ASSERT_OK(Any(a, &builder).status()); + auto a = ConstantR2(&builder, { + {false, false, false}, + {false, false, false}, + {false, false, false}, + }); + Any(a); ComputeAndCompareR0(&builder, false, {}); } diff --git a/tensorflow/compiler/xla/tests/prng_test.cc b/tensorflow/compiler/xla/tests/prng_test.cc index 1a2de6937c3e134852a730f62f7b56417cf49b28..5ebf8344d2b113b15f049b001c044c29c21c9004 100644 --- a/tensorflow/compiler/xla/tests/prng_test.cc +++ b/tensorflow/compiler/xla/tests/prng_test.cc @@ -18,7 +18,7 @@ limitations under the License. #include "tensorflow/compiler/xla/client/local_client.h" #include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/primitive_util.h" #include "tensorflow/compiler/xla/shape_util.h" #include "tensorflow/compiler/xla/test.h" @@ -53,8 +53,8 @@ template std::unique_ptr PrngTest::UniformTest( T a, T b, tensorflow::gtl::ArraySlice dims, int64 seed) { XlaBuilder builder(TestName()); - builder.RngUniform( - builder.ConstantR0(a), builder.ConstantR0(b), + RngUniform( + ConstantR0(&builder, a), ConstantR0(&builder, b), ShapeUtil::MakeShape(primitive_util::NativeToPrimitiveType(), dims)); SetSeed(seed); @@ -141,9 +141,9 @@ double PrngTest::UniformChiSquared(int32 range_size, int32 expected_count, int32 sample_size = range_size * expected_count; XlaBuilder builder(TestName()); - builder.RngUniform(builder.ConstantR0(0), - builder.ConstantR0(range_size), - ShapeUtil::MakeShape(S32, {sample_size})); + RngUniform(ConstantR0(&builder, 0), + ConstantR0(&builder, range_size), + ShapeUtil::MakeShape(S32, {sample_size})); SetSeed(seed); auto actual = @@ -184,21 +184,22 @@ XLA_TEST_F(PrngTest, MapUsingRng) { // Build a x -> (x + U[0,1)) computation. auto build_sum_rng = [this](XlaBuilder& builder) { auto b = builder.CreateSubBuilder("sum_with_rng"); - auto x = b->Parameter(0, ShapeUtil::MakeShape(F32, {}), "input"); - b->Add(x, b->RngUniform(b->ConstantR0(0), b->ConstantR0(1), - ShapeUtil::MakeShape(F32, {}))); + auto x = Parameter(b.get(), 0, ShapeUtil::MakeShape(F32, {}), "input"); + Add(x, + RngUniform(ConstantR0(b.get(), 0), ConstantR0(b.get(), 1), + ShapeUtil::MakeShape(F32, {}))); return b->BuildAndNoteError(); }; XlaBuilder builder(TestName()); std::unique_ptr param0_literal = - Literal::CreateR1({2.2f, 5.3f, 4.4f, 5.5f}); + LiteralUtil::CreateR1({2.2f, 5.3f, 4.4f, 5.5f}); TF_ASSERT_OK_AND_ASSIGN(std::unique_ptr param0_data, client_->TransferToServer(*param0_literal)); - auto param0 = builder.Parameter(0, param0_literal->shape(), "param0"); + auto param0 = Parameter(&builder, 0, param0_literal->shape(), "param0"); auto fn = build_sum_rng(builder); - builder.Map({param0}, fn, {0}); + Map(&builder, {param0}, fn, {0}); TF_ASSERT_OK_AND_ASSIGN(auto computation, builder.Build()); @@ -226,9 +227,8 @@ XLA_TEST_F(PrngTest, PassInGlobalRngSeed) { // Build a U[0,1) computation. auto build_computation = [this]() { XlaBuilder builder(TestName()); - builder.RngUniform(builder.ConstantR0(0), - builder.ConstantR0(1), - ShapeUtil::MakeShape(F32, {10})); + RngUniform(ConstantR0(&builder, 0), ConstantR0(&builder, 1), + ShapeUtil::MakeShape(F32, {10})); return builder.Build(); }; @@ -282,8 +282,8 @@ XLA_TEST_F(PrngTest, PassInGlobalRngSeed) { XLA_TEST_F(PrngTest, TenValuesN01) { XlaBuilder builder(TestName()); - builder.RngNormal(builder.ConstantR0(0), builder.ConstantR0(1), - ShapeUtil::MakeShape(F32, {10})); + RngNormal(ConstantR0(&builder, 0), ConstantR0(&builder, 1), + ShapeUtil::MakeShape(F32, {10})); SetSeed(42); ExecuteAndTransfer(&builder, /*arguments=*/{}).ConsumeValueOrDie(); @@ -294,9 +294,9 @@ XLA_TEST_F(PrngTest, RngUniformCrash) { XlaBuilder builder(TestName()); // This used to crash XLA during LLVM IR generation for CPUs. - auto rng_uniform = builder.RngUniform(builder.ConstantR0(0), - builder.ConstantR0(1000 * 1000), - ShapeUtil::MakeShape(S32, {})); + RngUniform(ConstantR0(&builder, 0), + ConstantR0(&builder, 1000 * 1000), + ShapeUtil::MakeShape(S32, {})); SetSeed(0); ExecuteAndTransfer(&builder, /*arguments=*/{}).ConsumeValueOrDie(); } diff --git a/tensorflow/compiler/xla/tests/query_inferred_shape_test.cc b/tensorflow/compiler/xla/tests/query_inferred_shape_test.cc index f95e75648343aa88bd7c39de4ee9f387f2b60506..526a38e8d1dbed9cdd4a31bfbec49bc5c6bb174b 100644 --- a/tensorflow/compiler/xla/tests/query_inferred_shape_test.cc +++ b/tensorflow/compiler/xla/tests/query_inferred_shape_test.cc @@ -31,8 +31,8 @@ class QueryInferredShapeTest : public ClientLibraryTestBase {}; TEST_F(QueryInferredShapeTest, OnePlusOneShape) { XlaBuilder builder("one_plus_one"); - auto one = builder.ConstantR0(1.0); - auto result = builder.Add(one, one); + auto one = ConstantR0(&builder, 1.0); + auto result = Add(one, one); StatusOr shape_status = builder.GetShape(result); ASSERT_IS_OK(shape_status.status()); auto shape = shape_status.ConsumeValueOrDie(); diff --git a/tensorflow/compiler/xla/tests/reduce_hlo_test.cc b/tensorflow/compiler/xla/tests/reduce_hlo_test.cc index 9052b188ed09a715b6ad7c3a40dc853d02cdd70c..a080dd1732bde21712cf47b4b57538cf4040f30e 100644 --- a/tensorflow/compiler/xla/tests/reduce_hlo_test.cc +++ b/tensorflow/compiler/xla/tests/reduce_hlo_test.cc @@ -95,21 +95,21 @@ XLA_TEST_P(ReduceWithLayoutTest, DISABLED_ON_GPU(Reduce)) { *reduce_input_shape->mutable_layout() = LayoutUtil::MakeLayout(reduce_layout.input_minor_to_major); - std::unique_ptr reduce_input = - Literal::CreateR4({{ /*i0=0*/ - {/*i1=0*/ - {-0.246092796, -0.179497838, -0.161181688}, - {-0.151643038, -0.240213156, -0.198156}}, - {/*i1=1*/ - {-0.14222312, -0.162200093, -0.193907976}, - {-0.239411, -0.198166847, -0.172471642}}}, - { /*i0=1*/ - {/*i1=0*/ - {-0.22965157, -0.218723893, -0.129257083}, - {-0.188762426, -0.16123569, -0.181166649}}, - {/*i1=1*/ - {-0.241772294, -0.245131493, -0.160247207}, - {-0.179881215, -0.23383224, -0.121976733}}}}); + std::unique_ptr reduce_input = LiteralUtil::CreateR4( + {{ /*i0=0*/ + {/*i1=0*/ + {-0.246092796, -0.179497838, -0.161181688}, + {-0.151643038, -0.240213156, -0.198156}}, + {/*i1=1*/ + {-0.14222312, -0.162200093, -0.193907976}, + {-0.239411, -0.198166847, -0.172471642}}}, + { /*i0=1*/ + {/*i1=0*/ + {-0.22965157, -0.218723893, -0.129257083}, + {-0.188762426, -0.16123569, -0.181166649}}, + {/*i1=1*/ + {-0.241772294, -0.245131493, -0.160247207}, + {-0.179881215, -0.23383224, -0.121976733}}}}); EXPECT_TRUE(RunAndCompareNoHloPasses(std::move(module), ErrorSpec(1e-5))); } diff --git a/tensorflow/compiler/xla/tests/reduce_precision_test.cc b/tensorflow/compiler/xla/tests/reduce_precision_test.cc index b311785449f1774c3bc1e4d7ad35c2866e3b4061..04c7f316463441d1bd458393b29ea5eb2acb9c9b 100644 --- a/tensorflow/compiler/xla/tests/reduce_precision_test.cc +++ b/tensorflow/compiler/xla/tests/reduce_precision_test.cc @@ -24,7 +24,7 @@ limitations under the License. #include "tensorflow/compiler/xla/client/local_client.h" #include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/compiler/xla/layout_util.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/service/reduce_precision_insertion.h" #include "tensorflow/compiler/xla/statusor.h" #include "tensorflow/compiler/xla/test.h" @@ -230,12 +230,13 @@ XLA_TEST_P(ReducePrecisionAccuracyTest, ReducePrecisionF32) { XlaBuilder builder(TestName()); - std::unique_ptr a_literal = Literal::CreateR1({input_values}); + std::unique_ptr a_literal = + LiteralUtil::CreateR1({input_values}); std::unique_ptr a_data = client_->TransferToServer(*a_literal).ConsumeValueOrDie(); - auto a = builder.Parameter(0, a_literal->shape(), "a"); + auto a = Parameter(&builder, 0, a_literal->shape(), "a"); - builder.ReducePrecision(a, exponent_bits, mantissa_bits); + ReducePrecision(a, exponent_bits, mantissa_bits); ComputeAndCompareR1(&builder, expected_values, {a_data.get()}); } @@ -253,18 +254,18 @@ XLA_TEST_F(ReducePrecisionInsertionTest, DISABLED_ON_INTERPRETER(ReducePrecisionBeforeFusion)) { XlaBuilder builder(TestName()); - std::unique_ptr a_literal = Literal::CreateR1({1.00001}); + std::unique_ptr a_literal = LiteralUtil::CreateR1({1.00001}); std::unique_ptr a_data = client_->TransferToServer(*a_literal).ConsumeValueOrDie(); - auto a = builder.Parameter(0, a_literal->shape(), "a"); + auto a = Parameter(&builder, 0, a_literal->shape(), "a"); // Abs doesn't affect resolution. - auto abs = builder.Abs(a); + auto abs = Abs(a); // Near 1.0, Log(x) approximates x - 1; this lets us confirm that the // reduce-precision operation showed up in the correct place in the // graph. - builder.Log(abs); + Log(abs); // Insert precision-reduction after the Abs(x) operation, rounding that // result to exactly 1.0f. @@ -282,14 +283,14 @@ XLA_TEST_F(ReducePrecisionInsertionTest, DISABLED_ON_INTERPRETER(ReducePrecisionSkippedAfterFusion)) { XlaBuilder builder(TestName()); - std::unique_ptr a_literal = Literal::CreateR1({1.00001}); + std::unique_ptr a_literal = LiteralUtil::CreateR1({1.00001}); std::unique_ptr a_data = client_->TransferToServer(*a_literal).ConsumeValueOrDie(); - auto a = builder.Parameter(0, a_literal->shape(), "a"); + auto a = Parameter(&builder, 0, a_literal->shape(), "a"); // These two operations should be fused by any reasonable backend. - auto abs = builder.Abs(a); - builder.Neg(abs); + auto abs = Abs(a); + Neg(abs); // Add a pass after operation fusion, suffixing kAbs operations. This // should not see into the fusion nodes and thus should not affect the @@ -308,14 +309,14 @@ XLA_TEST_F(ReducePrecisionInsertionTest, DISABLED_ON_INTERPRETER(ReducePrecisionAddedAfterFusion)) { XlaBuilder builder(TestName()); - std::unique_ptr a_literal = Literal::CreateR1({1.00001}); + std::unique_ptr a_literal = LiteralUtil::CreateR1({1.00001}); std::unique_ptr a_data = client_->TransferToServer(*a_literal).ConsumeValueOrDie(); - auto a = builder.Parameter(0, a_literal->shape(), "a"); + auto a = Parameter(&builder, 0, a_literal->shape(), "a"); // These two operations should be fused by any reasonable backend. - auto abs = builder.Abs(a); - builder.Neg(abs); + auto abs = Abs(a); + Neg(abs); // Add a pass after operation fusion, suffixing kFusion operations. auto reduce_precision_pass = execution_options_.mutable_debug_options() @@ -332,14 +333,14 @@ XLA_TEST_F(ReducePrecisionInsertionTest, DISABLED_ON_INTERPRETER(ReducePrecisionSkippedFusionContains)) { XlaBuilder builder(TestName()); - std::unique_ptr a_literal = Literal::CreateR1({1.00001}); + std::unique_ptr a_literal = LiteralUtil::CreateR1({1.00001}); std::unique_ptr a_data = client_->TransferToServer(*a_literal).ConsumeValueOrDie(); - auto a = builder.Parameter(0, a_literal->shape(), "a"); + auto a = Parameter(&builder, 0, a_literal->shape(), "a"); // These two operations should be fused by any reasonable backend. - auto abs = builder.Abs(a); - builder.Neg(abs); + auto abs = Abs(a); + Neg(abs); // Add a pass suffixing fusion nodes containing kCos operations. This // should have no effect. @@ -357,14 +358,14 @@ XLA_TEST_F(ReducePrecisionInsertionTest, DISABLED_ON_INTERPRETER(ReducePrecisionAddedFusionContains)) { XlaBuilder builder(TestName()); - std::unique_ptr a_literal = Literal::CreateR1({1.00001}); + std::unique_ptr a_literal = LiteralUtil::CreateR1({1.00001}); std::unique_ptr a_data = client_->TransferToServer(*a_literal).ConsumeValueOrDie(); - auto a = builder.Parameter(0, a_literal->shape(), "a"); + auto a = Parameter(&builder, 0, a_literal->shape(), "a"); // These two operations should be fused by any reasonable backend. - auto abs = builder.Abs(a); - builder.Neg(abs); + auto abs = Abs(a); + Neg(abs); // Add a pass suffixing fusion nodes containing kAbs operations. This // should see the kAbs operation within the above fusion node. diff --git a/tensorflow/compiler/xla/tests/reduce_test.cc b/tensorflow/compiler/xla/tests/reduce_test.cc index d671d40456a276a44b462f390c95aa4af301263a..1407fca72fd494c27fa999e67d69ecf36cbff81b 100644 --- a/tensorflow/compiler/xla/tests/reduce_test.cc +++ b/tensorflow/compiler/xla/tests/reduce_test.cc @@ -67,12 +67,12 @@ class ReduceTest : public ClientLibraryTestBase { ReduceTest() { // Implementation note: laid out z >> y >> x by default. // clang-format off - literal_2d_ = Literal::CreateR2({ + literal_2d_ = LiteralUtil::CreateR2({ // x0 x1 x2 { 1.f, 2.f, 3.f}, // y0 { 4.f, 5.f, 6.f}, // y1 }); - literal_3d_ = Literal::CreateR3Projected({ + literal_3d_ = LiteralUtil::CreateR3Projected({ // x0 x1 x2 { 1.f, 2.f, 3.f}, // y0 { 4.f, 5.f, 6.f}, // y1 @@ -89,9 +89,9 @@ class ReduceTest : public ClientLibraryTestBase { XlaBuilder builder(TestName()); XlaComputation add_f32 = CreateScalarAddComputation(F32, &builder); const Shape input_shape = ShapeUtil::MakeShape(F32, {element_count}); - auto input = builder.Parameter(0, input_shape, "input"); - auto zero = builder.ConstantR0(0.0); - builder.Reduce(input, zero, add_f32, /*dimensions_to_reduce=*/{0}); + auto input = Parameter(&builder, 0, input_shape, "input"); + auto zero = ConstantR0(&builder, 0.0); + Reduce(input, zero, add_f32, /*dimensions_to_reduce=*/{0}); std::vector input_data(element_count); for (int64 i = 0; i < element_count; ++i) { @@ -101,7 +101,7 @@ class ReduceTest : public ClientLibraryTestBase { } } std::unique_ptr input_literal = - Literal::CreateR1(AsSlice(input_data)); + LiteralUtil::CreateR1(AsSlice(input_data)); std::unique_ptr input_global_data = client_->TransferToServer(*input_literal).ConsumeValueOrDie(); @@ -118,22 +118,22 @@ class ReduceTest : public ClientLibraryTestBase { const int element_count = input_data.size(); XlaBuilder builder(TestName()); const Shape input_shape = ShapeUtil::MakeShape(S32, {element_count}); - auto input_par = builder.Parameter(0, input_shape, "input"); + auto input_par = Parameter(&builder, 0, input_shape, "input"); auto pred_values = - builder.Eq(input_par, builder.ConstantR1(element_count, 1)); + Eq(input_par, ConstantR1(&builder, element_count, 1)); XlaOp init_value; XlaComputation reduce; if (and_reduce) { - init_value = builder.ConstantR0(true); + init_value = ConstantR0(&builder, true); reduce = CreateScalarAndComputation(&builder); } else { - init_value = builder.ConstantR0(false); + init_value = ConstantR0(&builder, false); reduce = CreateScalarOrComputation(&builder); } - builder.Reduce(pred_values, init_value, reduce, - /*dimensions_to_reduce=*/{0}); + Reduce(pred_values, init_value, reduce, + /*dimensions_to_reduce=*/{0}); - std::unique_ptr input_literal = Literal::CreateR1(input_data); + std::unique_ptr input_literal = LiteralUtil::CreateR1(input_data); std::unique_ptr input_global_data = client_->TransferToServer(*input_literal).ConsumeValueOrDie(); @@ -156,26 +156,26 @@ class ReduceTest : public ClientLibraryTestBase { int64 major = 0) { XlaBuilder builder(TestName()); const Shape input_shape = ShapeUtil::MakeShape(U8, {rows, cols}); - auto input = builder.Parameter(0, input_shape, "input"); - auto input_pred = builder.Eq(input, builder.ConstantR0(1)); + auto input = Parameter(&builder, 0, input_shape, "input"); + auto input_pred = Eq(input, ConstantR0(&builder, 1)); XlaOp init_value; XlaComputation reduce_op; if (and_reduce) { - init_value = builder.ConstantR0(true); + init_value = ConstantR0(&builder, true); reduce_op = CreateScalarAndComputation(&builder); } else { - init_value = builder.ConstantR0(false); + init_value = ConstantR0(&builder, false); reduce_op = CreateScalarOrComputation(&builder); } - builder.Reduce(input_pred, init_value, reduce_op, - /*dimensions_to_reduce=*/{0}); + Reduce(input_pred, init_value, reduce_op, + /*dimensions_to_reduce=*/{0}); Array2D input_data(rows, cols); input_data.FillRandom(0, 1); std::unique_ptr input_literal = - Literal::CreateR2FromArray2D(input_data); + LiteralUtil::CreateR2FromArray2D(input_data); input_literal = input_literal->Relayout(LayoutUtil::MakeLayout({minor, major})); std::unique_ptr input_global_data = @@ -202,14 +202,14 @@ class ReduceTest : public ClientLibraryTestBase { XlaBuilder builder(TestName()); XlaComputation add_f32 = CreateScalarAddComputation(F32, &builder); const Shape input_shape = ShapeUtil::MakeShape(F32, {rows, cols}); - auto input = builder.Parameter(0, input_shape, "input"); - auto zero = builder.ConstantR0(0.0); - builder.Reduce(input, zero, add_f32, /*dimensions_to_reduce=*/{0, 1}); + auto input = Parameter(&builder, 0, input_shape, "input"); + auto zero = ConstantR0(&builder, 0.0); + Reduce(input, zero, add_f32, /*dimensions_to_reduce=*/{0, 1}); Array2D input_data(rows, cols); input_data.FillRandom(3.14f, 0.04); std::unique_ptr input_literal = - Literal::CreateR2FromArray2D(input_data); + LiteralUtil::CreateR2FromArray2D(input_data); input_literal = input_literal->Relayout(LayoutUtil::MakeLayout({minor, major})); std::unique_ptr input_global_data = @@ -230,14 +230,14 @@ class ReduceTest : public ClientLibraryTestBase { XlaBuilder builder(TestName()); XlaComputation add_f32 = CreateScalarAddComputation(F32, &builder); const Shape input_shape = ShapeUtil::MakeShape(F32, {rows, cols}); - auto input = builder.Parameter(0, input_shape, "input"); - auto zero = builder.ConstantR0(0.0); - builder.Reduce(input, zero, add_f32, /*dimensions_to_reduce=*/{0}); + auto input = Parameter(&builder, 0, input_shape, "input"); + auto zero = ConstantR0(&builder, 0.0); + Reduce(input, zero, add_f32, /*dimensions_to_reduce=*/{0}); Array2D input_data(rows, cols); input_data.FillRandom(3.14f, 0.04); std::unique_ptr input_literal = - Literal::CreateR2FromArray2D(input_data); + LiteralUtil::CreateR2FromArray2D(input_data); input_literal = input_literal->Relayout(LayoutUtil::MakeLayout({minor, major})); std::unique_ptr input_global_data = @@ -287,15 +287,15 @@ class ReduceTest : public ClientLibraryTestBase { XlaComputation reduction_function = reduction_function_generator(&builder); const Shape input_shape = ShapeUtil::MakeShape( xla::primitive_util::NativeToPrimitiveType(), {rows, cols}); - auto input = builder.Parameter(0, input_shape, "input"); - auto zero = builder.ConstantR0(initial_value); - builder.Reduce(input, zero, reduction_function, - /*dimensions_to_reduce=*/{0}); + auto input = Parameter(&builder, 0, input_shape, "input"); + auto zero = ConstantR0(&builder, initial_value); + Reduce(input, zero, reduction_function, + /*dimensions_to_reduce=*/{0}); Array2D input_data(rows, cols); input_data.FillUnique(initial_value); std::unique_ptr input_literal = - Literal::CreateR2FromArray2D(input_data); + LiteralUtil::CreateR2FromArray2D(input_data); input_literal = input_literal->Relayout(LayoutUtil::MakeLayout({minor, major})); std::unique_ptr input_global_data = @@ -442,15 +442,15 @@ XLA_TEST_F(ReduceTest, ReduceElementwiseR2_111x50_To_R1) { XlaBuilder builder(TestName()); XlaComputation add_f32 = CreateScalarAddComputation(F32, &builder); const Shape input_shape = ShapeUtil::MakeShape(F32, {rows, cols}); - auto input = builder.Parameter(0, input_shape, "input"); - auto zero = builder.ConstantR0(0.0); - auto log_ = builder.Log(input); - builder.Reduce(log_, zero, add_f32, /*dimensions_to_reduce=*/{0}); + auto input = Parameter(&builder, 0, input_shape, "input"); + auto zero = ConstantR0(&builder, 0.0); + auto log_ = Log(input); + Reduce(log_, zero, add_f32, /*dimensions_to_reduce=*/{0}); Array2D input_data(rows, cols); input_data.FillRandom(3.14f, 0.04); std::unique_ptr input_literal = - Literal::CreateR2FromArray2D(input_data); + LiteralUtil::CreateR2FromArray2D(input_data); input_literal = input_literal->Relayout(LayoutUtil::MakeLayout({0, 1})); std::unique_ptr input_global_data = client_->TransferToServer(*input_literal).ConsumeValueOrDie(); @@ -473,16 +473,16 @@ XLA_TEST_F(ReduceTest, TransposeAndReduceElementwiseR2_111x50_To_R1) { XlaBuilder builder(TestName()); XlaComputation add_f32 = CreateScalarAddComputation(F32, &builder); const Shape input_shape = ShapeUtil::MakeShape(F32, {rows, cols}); - auto input = builder.Parameter(0, input_shape, "input"); - auto zero = builder.ConstantR0(0.0); - auto log_ = builder.Log(input); - auto transpose = builder.Transpose(log_, {1, 0}); - builder.Reduce(transpose, zero, add_f32, /*dimensions_to_reduce=*/{1}); + auto input = Parameter(&builder, 0, input_shape, "input"); + auto zero = ConstantR0(&builder, 0.0); + auto log_ = Log(input); + auto transpose = Transpose(log_, {1, 0}); + Reduce(transpose, zero, add_f32, /*dimensions_to_reduce=*/{1}); Array2D input_data(rows, cols); input_data.FillRandom(3.14f, 0.04); std::unique_ptr input_literal = - Literal::CreateR2FromArray2D(input_data); + LiteralUtil::CreateR2FromArray2D(input_data); input_literal = input_literal->Relayout(LayoutUtil::MakeLayout({0, 1})); std::unique_ptr input_global_data = client_->TransferToServer(*input_literal).ConsumeValueOrDie(); @@ -505,10 +505,10 @@ XLA_TEST_F(ReduceTest, TransposeAndReduceR3_12x111x50_To_R2) { XlaBuilder builder(TestName()); XlaComputation add_f32 = CreateScalarAddComputation(F32, &builder); const Shape input_shape = ShapeUtil::MakeShape(F32, {12, 111, 50}); - XlaOp input = builder.Parameter(0, input_shape, "input"); - XlaOp zero = builder.ConstantR0(0.0); - XlaOp transpose = builder.Transpose(input, /*permutation=*/{1, 0, 2}); - builder.Reduce(transpose, zero, add_f32, /*dimensions_to_reduce=*/{0}); + XlaOp input = Parameter(&builder, 0, input_shape, "input"); + XlaOp zero = ConstantR0(&builder, 0.0); + XlaOp transpose = Transpose(input, /*permutation=*/{1, 0, 2}); + Reduce(transpose, zero, add_f32, /*dimensions_to_reduce=*/{0}); TF_ASSERT_OK_AND_ASSIGN(std::unique_ptr input_data, MakeFakeLiteral(input_shape)); @@ -522,16 +522,16 @@ XLA_TEST_F(ReduceTest, Reshape_111x2x25Reduce_111x50_To_R1) { XlaBuilder builder(TestName()); XlaComputation add_f32 = CreateScalarAddComputation(F32, &builder); const Shape input_shape = ShapeUtil::MakeShape(F32, {rows, 2, cols / 2}); - auto input = builder.Parameter(0, input_shape, "input"); - auto zero = builder.ConstantR0(0.0); - auto log_ = builder.Tanh(input); - auto reshape = builder.Reshape(log_, {rows, cols}); - builder.Reduce(reshape, zero, add_f32, /*dimensions_to_reduce=*/{0}); + auto input = Parameter(&builder, 0, input_shape, "input"); + auto zero = ConstantR0(&builder, 0.0); + auto log_ = Tanh(input); + auto reshape = Reshape(log_, {rows, cols}); + Reduce(reshape, zero, add_f32, /*dimensions_to_reduce=*/{0}); Array3D input_data(rows, 2, cols / 2); input_data.FillRandom(3.14f, 0.04); std::unique_ptr input_literal = - Literal::CreateR3FromArray3D(input_data); + LiteralUtil::CreateR3FromArray3D(input_data); std::unique_ptr input_global_data = client_->TransferToServer(*input_literal).ConsumeValueOrDie(); @@ -568,9 +568,9 @@ void PrintTo(const BoundsLayout& spec, std::ostream* os) { XLA_TEST_F(ReduceTest, AddReduce2DScalarToR0) { XlaBuilder builder(TestName()); auto add = CreateScalarAddComputation(F32, &builder); - auto scalar = builder.ConstantR0(42.0); - auto broadcasted = builder.Broadcast(scalar, {500, 500}); - builder.Reduce(broadcasted, builder.ConstantR0(0.0f), add, {0, 1}); + auto scalar = ConstantR0(&builder, 42.0); + auto broadcasted = Broadcast(scalar, {500, 500}); + Reduce(broadcasted, ConstantR0(&builder, 0.0f), add, {0, 1}); float expected = 42.0f * static_cast(500 * 500); ComputeAndCompareR0(&builder, expected, {}, ErrorSpec(0.0001)); @@ -580,9 +580,9 @@ XLA_TEST_F(ReduceTest, AddReduce2DScalarToR0) { XLA_TEST_F(ReduceTest, MaxReduce2DScalarToR0) { XlaBuilder builder(TestName()); auto max = CreateScalarMaxComputation(F32, &builder); - auto scalar = builder.ConstantR0(42.0); - auto broadcasted = builder.Broadcast(scalar, {500, 500}); - builder.Reduce(broadcasted, builder.ConstantR0(0.0f), max, {0, 1}); + auto scalar = ConstantR0(&builder, 42.0); + auto broadcasted = Broadcast(scalar, {500, 500}); + Reduce(broadcasted, ConstantR0(&builder, 0.0f), max, {0, 1}); float expected = 42.0f; ComputeAndCompareR0(&builder, expected, {}, ErrorSpec(0.0001)); @@ -594,9 +594,9 @@ XLA_TEST_F(ReduceTest, MaxReduce2DToR0) { auto max = CreateScalarMaxComputation(F32, &builder); Array2D input(300, 250); input.FillRandom(214.0f); - auto input_literal = Literal::CreateR2FromArray2D(input); - builder.Reduce(builder.ConstantLiteral(*input_literal), - builder.ConstantR0(FLT_MIN), max, {0, 1}); + auto input_literal = LiteralUtil::CreateR2FromArray2D(input); + Reduce(ConstantLiteral(&builder, *input_literal), + ConstantR0(&builder, FLT_MIN), max, {0, 1}); auto input_max = FLT_MIN; input.Each( [&](int64, int64, float* v) { input_max = std::max(input_max, *v); }); @@ -609,9 +609,9 @@ XLA_TEST_F(ReduceTest, MinReduce2DToR0) { auto min = CreateScalarMinComputation(F32, &builder); Array2D input(150, 130); input.FillRandom(214.0f); - auto input_literal = Literal::CreateR2FromArray2D(input); - builder.Reduce(builder.ConstantLiteral(*input_literal), - builder.ConstantR0(FLT_MAX), min, {0, 1}); + auto input_literal = LiteralUtil::CreateR2FromArray2D(input); + Reduce(ConstantLiteral(&builder, *input_literal), + ConstantR0(&builder, FLT_MAX), min, {0, 1}); auto input_min = FLT_MAX; input.Each( @@ -623,12 +623,11 @@ XLA_TEST_F(ReduceTest, UnsignedInt_MinReduce) { XlaBuilder builder(TestName()); Array2D input({{1}, {2}}); auto min = CreateScalarMinComputation(U32, &builder); - auto input_literal = Literal::CreateR2FromArray2D(input); + auto input_literal = LiteralUtil::CreateR2FromArray2D(input); auto initial_value = - builder.ConstantR0(std::numeric_limits::max()); + ConstantR0(&builder, std::numeric_limits::max()); - builder.Reduce(builder.ConstantLiteral(*input_literal), initial_value, min, - {0, 1}); + Reduce(ConstantLiteral(&builder, *input_literal), initial_value, min, {0, 1}); ComputeAndCompareR0(&builder, 1, {}); } @@ -636,21 +635,20 @@ XLA_TEST_F(ReduceTest, UnsignedInt_MaxReduce) { XlaBuilder builder(TestName()); Array2D input({{1}, {2}}); auto max = CreateScalarMaxComputation(U32, &builder); - auto input_literal = Literal::CreateR2FromArray2D(input); + auto input_literal = LiteralUtil::CreateR2FromArray2D(input); auto initial_value = - builder.ConstantR0(std::numeric_limits::min()); + ConstantR0(&builder, std::numeric_limits::min()); - builder.Reduce(builder.ConstantLiteral(*input_literal), initial_value, max, - {0, 1}); + Reduce(ConstantLiteral(&builder, *input_literal), initial_value, max, {0, 1}); ComputeAndCompareR0(&builder, 2, {}); } // Reduces a matrix among dimension 1. XLA_TEST_F(ReduceTest, Reduce2DAmong1) { XlaBuilder builder(TestName()); - auto m = builder.ConstantLiteral(*literal_2d_); + auto m = ConstantLiteral(&builder, *literal_2d_); auto add = CreateScalarAddComputation(F32, &builder); - builder.Reduce(m, builder.ConstantR0(0.0f), add, {1}); + Reduce(m, ConstantR0(&builder, 0.0f), add, {1}); std::vector expected = {6.f, 15.f}; ComputeAndCompareR1(&builder, expected, {}, ErrorSpec(0.0001)); @@ -659,9 +657,9 @@ XLA_TEST_F(ReduceTest, Reduce2DAmong1) { XLA_TEST_F(ReduceTest, Reduce2DAmong0and1) { // Reduce a matrix among dimensions 0 and 1 (sum it up to a scalar). XlaBuilder builder(TestName()); - auto m = builder.ConstantLiteral(*literal_2d_); + auto m = ConstantLiteral(&builder, *literal_2d_); auto add = CreateScalarAddComputation(F32, &builder); - builder.Reduce(m, builder.ConstantR0(0.0f), add, {0, 1}); + Reduce(m, ConstantR0(&builder, 0.0f), add, {0, 1}); ComputeAndCompareR0(&builder, 21.0f, {}, ErrorSpec(0.0001, 1e-4)); } @@ -669,9 +667,9 @@ XLA_TEST_F(ReduceTest, Reduce2DAmong0and1) { // Tests 2D matrix ReduceToRow operation. XLA_TEST_F(ReduceTest, Reduce2DAmongY) { XlaBuilder builder("reduce_among_y"); - auto m = builder.ConstantLiteral(*literal_2d_); + auto m = ConstantLiteral(&builder, *literal_2d_); auto add = CreateScalarAddComputation(F32, &builder); - builder.Reduce(m, builder.ConstantR0(0.0f), add, {0}); + Reduce(m, ConstantR0(&builder, 0.0f), add, {0}); std::vector expected = {5.f, 7.f, 9.f}; ComputeAndCompareR1(&builder, expected, {}, ErrorSpec(0.0001)); @@ -679,9 +677,9 @@ XLA_TEST_F(ReduceTest, Reduce2DAmongY) { XLA_TEST_F(ReduceTest, ReduceR3AmongDims_1_2) { XlaBuilder builder(TestName()); - auto m = builder.ConstantLiteral(*literal_3d_); + auto m = ConstantLiteral(&builder, *literal_3d_); auto add = CreateScalarAddComputation(F32, &builder); - builder.Reduce(m, builder.ConstantR0(0.0f), add, {1, 2}); + Reduce(m, ConstantR0(&builder, 0.0f), add, {1, 2}); std::vector expected = {21.f, 21.f, 21.f, 21.f}; ComputeAndCompareR1(&builder, expected, {}, ErrorSpec(0.0001)); @@ -689,9 +687,9 @@ XLA_TEST_F(ReduceTest, ReduceR3AmongDims_1_2) { XLA_TEST_F(ReduceTest, ReduceR3AmongDims_0_1) { XlaBuilder builder(TestName()); - auto m = builder.ConstantLiteral(*literal_3d_); + auto m = ConstantLiteral(&builder, *literal_3d_); auto add = CreateScalarAddComputation(F32, &builder); - builder.Reduce(m, builder.ConstantR0(0.0f), add, {0, 1}); + Reduce(m, ConstantR0(&builder, 0.0f), add, {0, 1}); std::vector expected = {20.f, 28.f, 36.f}; ComputeAndCompareR1(&builder, expected, {}, ErrorSpec(0.0001)); @@ -699,9 +697,9 @@ XLA_TEST_F(ReduceTest, ReduceR3AmongDims_0_1) { XLA_TEST_F(ReduceTest, ReduceR3ToR0) { XlaBuilder builder(TestName()); - auto m = builder.ConstantLiteral(*literal_3d_); + auto m = ConstantLiteral(&builder, *literal_3d_); auto add = CreateScalarAddComputation(F32, &builder); - builder.Reduce(m, builder.ConstantR0(0.0f), add, {0, 1, 2}); + Reduce(m, ConstantR0(&builder, 0.0f), add, {0, 1, 2}); float expected = 21.0f * 4.0; ComputeAndCompareR0(&builder, expected, {}, ErrorSpec(0.0001)); @@ -709,9 +707,9 @@ XLA_TEST_F(ReduceTest, ReduceR3ToR0) { XLA_TEST_F(ReduceTest, ReduceR3AmongDim0) { XlaBuilder builder(TestName()); - auto m = builder.ConstantLiteral(*literal_3d_); + auto m = ConstantLiteral(&builder, *literal_3d_); auto add = CreateScalarAddComputation(F32, &builder); - builder.Reduce(m, builder.ConstantR0(0.0f), add, {0}); + Reduce(m, ConstantR0(&builder, 0.0f), add, {0}); // clang-format off Array2D expected({ @@ -724,9 +722,9 @@ XLA_TEST_F(ReduceTest, ReduceR3AmongDim0) { XLA_TEST_F(ReduceTest, ReduceR3AmongDim1) { XlaBuilder builder(TestName()); - auto m = builder.ConstantLiteral(*literal_3d_); + auto m = ConstantLiteral(&builder, *literal_3d_); auto add = CreateScalarAddComputation(F32, &builder); - builder.Reduce(m, builder.ConstantR0(0.0f), add, {1}); + Reduce(m, ConstantR0(&builder, 0.0f), add, {1}); // clang-format off Array2D expected({ @@ -741,9 +739,9 @@ XLA_TEST_F(ReduceTest, ReduceR3AmongDim1) { XLA_TEST_F(ReduceTest, ReduceR3AmongDim2) { XlaBuilder builder(TestName()); - auto m = builder.ConstantLiteral(*literal_3d_); + auto m = ConstantLiteral(&builder, *literal_3d_); auto add = CreateScalarAddComputation(F32, &builder); - builder.Reduce(m, builder.ConstantR0(0.0f), add, {2}); + Reduce(m, ConstantR0(&builder, 0.0f), add, {2}); // clang-format off Array2D expected({ @@ -820,17 +818,17 @@ XLA_TEST_P(ReduceR3ToR2Test, ReduceR3ToR2) { // input_array.FillRandom(3.14f, 0.05); input_array.Fill(1.0f); - auto input_literal = Literal::CreateR3FromArray3D(input_array); + auto input_literal = LiteralUtil::CreateR3FromArray3D(input_array); input_literal = input_literal->Relayout(LayoutUtil::MakeLayout(GetParam().layout)); std::unique_ptr input_data = client_->TransferToServer(*input_literal).ConsumeValueOrDie(); auto input_activations = - builder.Parameter(0, input_literal->shape(), "input"); + Parameter(&builder, 0, input_literal->shape(), "input"); XlaComputation add = CreateScalarAddComputation(F32, &builder); - auto sum = builder.Reduce(input_activations, builder.ConstantR0(0.0f), - add, GetParam().reduce_dims); + Reduce(input_activations, ConstantR0(&builder, 0.0f), add, + GetParam().reduce_dims); auto expected = ReferenceUtil::Reduce3DTo2D(input_array, 0.0f, GetParam().reduce_dims, @@ -871,14 +869,15 @@ XLA_TEST_F(ReduceTest, DISABLED_ON_GPU(OperationOnConstantAsInitValue)) { XlaBuilder builder(TestName()); XlaComputation max_f32 = CreateScalarMaxComputation(F32, &builder); - auto a = builder.ConstantR0(2.0f); - auto a2 = builder.Abs(a); + auto a = ConstantR0(&builder, 2.0f); + auto a2 = Abs(a); - std::unique_ptr b_literal = Literal::CreateR1({1.0f, 4.0f}); + std::unique_ptr b_literal = + LiteralUtil::CreateR1({1.0f, 4.0f}); std::unique_ptr b_data = client_->TransferToServer(*b_literal).ConsumeValueOrDie(); - auto b = builder.Parameter(0, b_literal->shape(), "b"); - auto max = builder.Reduce(b, a2, max_f32, {0}); + auto b = Parameter(&builder, 0, b_literal->shape(), "b"); + Reduce(b, a2, max_f32, {0}); ComputeAndCompareR0(&builder, 4.0f, {b_data.get()}); } @@ -900,13 +899,13 @@ class ReduceInitializerTest : public ReduceTest { XlaComputation max_fn = CreateScalarMaxComputation( primitive_util::NativeToPrimitiveType(), &builder); - auto init = builder.ConstantR0(initializer); + auto init = ConstantR0(&builder, initializer); std::vector input_arr(num_elems, std::numeric_limits::lowest()); - auto input_literal = Literal::CreateR1(input_arr); + auto input_literal = LiteralUtil::CreateR1(input_arr); auto input_data = client_->TransferToServer(*input_literal).ConsumeValueOrDie(); - builder.Reduce(builder.Parameter(0, input_literal->shape(), "input"), init, - max_fn, {0}); + Reduce(Parameter(&builder, 0, input_literal->shape(), "input"), init, + max_fn, {0}); ComputeAndCompareR0(&builder, initializer, {input_data.get()}); } @@ -939,23 +938,24 @@ XLA_TEST_F(ReduceInitializerTest, U64InitializerBigValue) { XLA_TEST_F(ReduceTest, ReduceIdentity) { XlaBuilder builder(TestName()); Shape single_float = ShapeUtil::MakeShape(F32, {}); - builder.Parameter(0, single_float, "lhs-unused"); - builder.Parameter(1, single_float, "rhs-used"); + Parameter(&builder, 0, single_float, "lhs-unused"); + Parameter(&builder, 1, single_float, "rhs-used"); auto computation_status = builder.Build(); TF_ASSERT_OK(computation_status.status()); Shape operand_shape = ShapeUtil::MakeShape(F32, {1}); - builder.Reduce(builder.Parameter(0, operand_shape, "operand"), - builder.Parameter(1, single_float, "init"), - computation_status.ValueOrDie(), {0}); + Reduce(Parameter(&builder, 0, operand_shape, "operand"), + Parameter(&builder, 1, single_float, "init"), + computation_status.ValueOrDie(), {0}); float operand[] = {42.0f}; float init = 58.5f; float expected = 42.0f; - std::unique_ptr input_literal = Literal::CreateR1(operand); + std::unique_ptr input_literal = + LiteralUtil::CreateR1(operand); std::unique_ptr input_global_data = client_->TransferToServer(*input_literal).ConsumeValueOrDie(); - std::unique_ptr input_literal2 = Literal::CreateR0(init); + std::unique_ptr input_literal2 = LiteralUtil::CreateR0(init); std::unique_ptr input_global_data2 = client_->TransferToServer(*input_literal2).ConsumeValueOrDie(); ComputeAndCompareR0( diff --git a/tensorflow/compiler/xla/tests/reduce_window_test.cc b/tensorflow/compiler/xla/tests/reduce_window_test.cc index 266760e8202fddc48792ac66dda334255e428808..c2681f70f7e5727462c20d5eb3120bd34fd75765 100644 --- a/tensorflow/compiler/xla/tests/reduce_window_test.cc +++ b/tensorflow/compiler/xla/tests/reduce_window_test.cc @@ -70,31 +70,33 @@ class ReduceWindowTest : public ::testing::WithParamInterface, tensorflow::gtl::ArraySlice window_dimensions, tensorflow::gtl::ArraySlice window_strides, Padding padding) { - auto init = - CreateConstantFromLiteral(*Literal::CreateR0(0.0f), &builder_); - builder_.ReduceWindow(input, init, - CreateScalarAddComputation(FloatType(), &builder_), - window_dimensions, window_strides, padding); + auto init = CreateConstantFromLiteral(*LiteralUtil::CreateR0(0.0f), + &builder_); + ReduceWindow(input, init, + CreateScalarAddComputation(FloatType(), &builder_), + window_dimensions, window_strides, padding); } void ReduceWindowMax(const XlaOp& input, tensorflow::gtl::ArraySlice window_dimensions, tensorflow::gtl::ArraySlice window_strides, Padding padding) { - auto init = CreateConstantFromLiteral(Literal::MinValue(F32), &builder_); - builder_.ReduceWindow(input, init, - CreateScalarMaxComputation(FloatType(), &builder_), - window_dimensions, window_strides, padding); + auto init = + CreateConstantFromLiteral(LiteralUtil::MinValue(F32), &builder_); + ReduceWindow(input, init, + CreateScalarMaxComputation(FloatType(), &builder_), + window_dimensions, window_strides, padding); } void ReduceWindowMin(const XlaOp& input, tensorflow::gtl::ArraySlice window_dimensions, tensorflow::gtl::ArraySlice window_strides, Padding padding) { - auto init = CreateConstantFromLiteral(Literal::MaxValue(F32), &builder_); - builder_.ReduceWindow(input, init, - CreateScalarMinComputation(FloatType(), &builder_), - window_dimensions, window_strides, padding); + auto init = + CreateConstantFromLiteral(LiteralUtil::MaxValue(F32), &builder_); + ReduceWindow(input, init, + CreateScalarMinComputation(FloatType(), &builder_), + window_dimensions, window_strides, padding); } XlaBuilder builder_; @@ -102,14 +104,14 @@ class ReduceWindowTest : public ::testing::WithParamInterface, TEST_P(ReduceWindowTest, MismatchedRanksGivesErrorStatus) { const auto input = CreateConstantFromLiteral( - *Literal::CreateR1({1, 1, 1, 1}), &builder_); + *LiteralUtil::CreateR1({1, 1, 1, 1}), &builder_); const auto init_value = - CreateConstantFromLiteral(*Literal::CreateR0(0), &builder_); + CreateConstantFromLiteral(*LiteralUtil::CreateR0(0), &builder_); TF_ASSERT_OK(builder_.first_error()); - builder_.ReduceWindow(input, init_value, - CreateScalarAddComputation(FloatType(), &builder_), - /*window_dimensions=*/{1, 2}, - /*window_strides=*/{1}, Padding::kValid); + ReduceWindow(input, init_value, + CreateScalarAddComputation(FloatType(), &builder_), + /*window_dimensions=*/{1, 2}, + /*window_strides=*/{1}, Padding::kValid); ASSERT_EQ(builder_.first_error().code(), tensorflow::error::INVALID_ARGUMENT) << builder_.first_error(); ASSERT_THAT(builder_.first_error().error_message(), @@ -119,33 +121,32 @@ TEST_P(ReduceWindowTest, MismatchedRanksGivesErrorStatus) { // Regression test for b/68964348. TEST_P(ReduceWindowTest, R0ReduceWindow) { const auto input = - CreateConstantFromLiteral(*Literal::CreateR0(42.0), &builder_); + CreateConstantFromLiteral(*LiteralUtil::CreateR0(42.0), &builder_); const auto init = - CreateConstantFromLiteral(*Literal::CreateR0(1.0), &builder_); - builder_.ReduceWindow(input, init, - CreateScalarAddComputation(FloatType(), &builder_), - /*window_dimensions=*/{}, - /*window_strides=*/{}, Padding::kSame); - ComputeAndCompareLiteral(&builder_, *Literal::CreateR0(43.0), {}, + CreateConstantFromLiteral(*LiteralUtil::CreateR0(1.0), &builder_); + ReduceWindow(input, init, CreateScalarAddComputation(FloatType(), &builder_), + /*window_dimensions=*/{}, + /*window_strides=*/{}, Padding::kSame); + ComputeAndCompareLiteral(&builder_, *LiteralUtil::CreateR0(43.0), {}, ErrorSpec(0.00001)); } TEST_P(ReduceWindowTest, Min3In5Stride2) { const auto input = CreateConstantFromLiteral( - *Literal::CreateR1({10000, 1000, 100, 10, 1}), &builder_); + *LiteralUtil::CreateR1({10000, 1000, 100, 10, 1}), &builder_); ReduceWindowMin(input, {3}, {2}, Padding::kValid); - ComputeAndCompareLiteral(&builder_, *Literal::CreateR1({100, 1}), {}, - ErrorSpec(0.00001)); + ComputeAndCompareLiteral(&builder_, *LiteralUtil::CreateR1({100, 1}), + {}, ErrorSpec(0.00001)); } TEST_P(ReduceWindowTest, Min3In5Stride1WithSamePadding) { const auto input = CreateConstantFromLiteral( - *Literal::CreateR1({10000, 1000, 100, 10, 1}), &builder_); + *LiteralUtil::CreateR1({10000, 1000, 100, 10, 1}), &builder_); ReduceWindowMin(input, /*window_dimensions=*/{3}, /*window_strides=*/{1}, Padding::kSame); ComputeAndCompareLiteral(&builder_, - *Literal::CreateR1({1000, 100, 10, 1, 1}), {}, - ErrorSpec(0.00001)); + *LiteralUtil::CreateR1({1000, 100, 10, 1, 1}), + {}, ErrorSpec(0.00001)); } XLA_TEST_P(ReduceWindowTest, ZeroElementSmall) { @@ -157,7 +158,7 @@ XLA_TEST_P(ReduceWindowTest, ZeroElementSmall) { auto res = ReferenceUtil::ReduceWindow4DAdd(input_array, 0.0f, {1, 1, 2, 1}, {1, 1, 1, 1}, padding); - ComputeAndCompareLiteral(&builder_, *Literal::CreateFromArray(*res), {}, + ComputeAndCompareLiteral(&builder_, *LiteralUtil::CreateFromArray(*res), {}, DefaultErrorSpec()); } @@ -172,7 +173,7 @@ TEST_P(ReduceWindowTest, NonSquareSmall) { auto res = ReferenceUtil::ReduceWindow4DAdd(input_array, 0.0f, {1, 1, 2, 1}, {1, 1, 1, 1}, padding); - ComputeAndCompareLiteral(&builder_, *Literal::CreateFromArray(*res), {}, + ComputeAndCompareLiteral(&builder_, *LiteralUtil::CreateFromArray(*res), {}, DefaultErrorSpec()); } @@ -186,7 +187,7 @@ TEST_P(ReduceWindowTest, MiddleDimsSmall) { auto res = ReferenceUtil::ReduceWindow4DAdd(input_array, 0.0f, {1, 1, 1, 1}, {1, 2, 2, 1}, padding); - ComputeAndCompareLiteral(&builder_, *Literal::CreateFromArray(*res), {}, + ComputeAndCompareLiteral(&builder_, *LiteralUtil::CreateFromArray(*res), {}, DefaultErrorSpec()); } @@ -203,7 +204,7 @@ TEST_P(ReduceWindowTest, Along2ndMinorDim) { auto res = ReferenceUtil::ReduceWindow4DAdd( input_array, 0.0f, {1, 1, lrn_diameter, 1}, {1, 1, 1, 1}, padding); - ComputeAndCompareLiteral(&builder_, *Literal::CreateFromArray(*res), {}, + ComputeAndCompareLiteral(&builder_, *LiteralUtil::CreateFromArray(*res), {}, DefaultErrorSpec()); } @@ -225,8 +226,8 @@ TEST_P(ReduceWindowTest, AmongMajor2Dims) { input_array, 0.0f, {win_len, win_len, 1, 1}, {win_stride, win_stride, 1, 1}, padding); - ComputeAndCompareLiteral(&builder_, *Literal::CreateFromArray(*result), {}, - DefaultErrorSpec()); + ComputeAndCompareLiteral(&builder_, *LiteralUtil::CreateFromArray(*result), + {}, DefaultErrorSpec()); } TEST_P(ReduceWindowTest, AmongMajor2DimsMediumSize) { @@ -248,8 +249,8 @@ TEST_P(ReduceWindowTest, AmongMajor2DimsMediumSize) { input_array, 0.0f, {win_len, win_len, 1, 1}, {win_stride, win_stride, 1, 1}, padding); - ComputeAndCompareLiteral(&builder_, *Literal::CreateFromArray(*result), {}, - DefaultErrorSpec()); + ComputeAndCompareLiteral(&builder_, *LiteralUtil::CreateFromArray(*result), + {}, DefaultErrorSpec()); } // Tests the super windowing logic w.r.t handling prime number of windows in a @@ -273,8 +274,8 @@ TEST_P(ReduceWindowTest, PrimeWindowsInReductionDimension) { input_array, 0.0f, {win_len, win_len, 1, 1}, {win_stride, win_stride, 1, 1}, padding); - ComputeAndCompareLiteral(&builder_, *Literal::CreateFromArray(*result), {}, - DefaultErrorSpec()); + ComputeAndCompareLiteral(&builder_, *LiteralUtil::CreateFromArray(*result), + {}, DefaultErrorSpec()); } TEST_P(ReduceWindowTest, ReduceAlongLaneDimension) { @@ -290,8 +291,8 @@ TEST_P(ReduceWindowTest, ReduceAlongLaneDimension) { auto result = ReferenceUtil::ReduceWindow4DAdd( input_array, 0.0f, {1, 1, 1, 11}, {1, 1, 1, 1}, padding); - ComputeAndCompareLiteral(&builder_, *Literal::CreateFromArray(*result), {}, - DefaultErrorSpec()); + ComputeAndCompareLiteral(&builder_, *LiteralUtil::CreateFromArray(*result), + {}, DefaultErrorSpec()); } // Tests a reduction function that is not a simple add/min/max/etc. @@ -306,15 +307,15 @@ XLA_TEST_P(ReduceWindowTest, NonstandardReduceFunction) { Padding padding = Padding::kValid; const Shape scalar = ShapeUtil::MakeShape(FloatType(), {}); auto b = builder_.CreateSubBuilder("unusual"); - auto lhs = b->Parameter(0, scalar, "lhs"); - auto rhs = b->Parameter(1, scalar, "rhs"); - b->Min(b->Add(lhs, rhs), - CreateConstantFromLiteral(*Literal::CreateR0(8.0f), b.get())); + auto lhs = Parameter(b.get(), 0, scalar, "lhs"); + auto rhs = Parameter(b.get(), 1, scalar, "rhs"); + Min(Add(lhs, rhs), + CreateConstantFromLiteral(*LiteralUtil::CreateR0(8.0f), b.get())); XlaComputation reduce_fn = b->BuildAndNoteError(); - builder_.ReduceWindow( + ReduceWindow( input, - CreateConstantFromLiteral(*Literal::CreateR0(0.0f), &builder_), + CreateConstantFromLiteral(*LiteralUtil::CreateR0(0.0f), &builder_), reduce_fn, /*window_dimensions=*/{1, 1, 2, 1}, /*window_strides=*/{1, 1, 1, 1}, padding); @@ -328,15 +329,15 @@ XLA_TEST_P(ReduceWindowTest, NonstandardReduceFunction) { /*window=*/{1, 1, 2, 1}, /*stride=*/{1, 1, 1, 1}, padding); - ComputeAndCompareLiteral(&builder_, *Literal::CreateFromArray(*expected), {}, - DefaultErrorSpec()); + ComputeAndCompareLiteral(&builder_, *LiteralUtil::CreateFromArray(*expected), + {}, DefaultErrorSpec()); } TEST_P(ReduceWindowTest, R4UnitWindow) { Array4D input_array(13, 12, 8, 15); input_array.FillRandom(2.f, 2.f); std::unique_ptr input_literal = - Literal::CreateR4FromArray4DWithLayout( + LiteralUtil::CreateR4FromArray4DWithLayout( input_array, LayoutUtil::MakeLayout({0, 3, 2, 1})); XlaOp input; auto input_data = CreateParameterAndTransferLiteral( @@ -348,7 +349,7 @@ TEST_P(ReduceWindowTest, R4UnitWindow) { auto res = ReferenceUtil::ReduceWindow4DAdd(input_array, 0.0f, {1, 1, 7, 1}, {1, 4, 1, 1}, padding); - ComputeAndCompareLiteral(&builder_, *Literal::CreateFromArray(*res), + ComputeAndCompareLiteral(&builder_, *LiteralUtil::CreateFromArray(*res), {input_data.get()}, DefaultErrorSpec()); } @@ -377,7 +378,7 @@ XLA_TEST_P(ReduceWindowTest, R6Add) { auto shape = ShapeUtil::MakeShape(F32, input_dims); std::unique_ptr arg_literal = - Literal::CreateFullWithDescendingLayout(input_dims, 1.0f); + LiteralUtil::CreateFullWithDescendingLayout(input_dims, 1.0f); const auto input = CreateConstantFromLiteral(*arg_literal, &builder_); @@ -386,7 +387,7 @@ XLA_TEST_P(ReduceWindowTest, R6Add) { std::vector output_dims = {8, 8, 6, 6, 8, 8}; std::unique_ptr expected = - Literal::CreateFullWithDescendingLayout(output_dims, 9.0f); + LiteralUtil::CreateFullWithDescendingLayout(output_dims, 9.0f); ComputeAndCompareLiteral(&builder_, *expected, {}, DefaultErrorSpec()); } @@ -395,7 +396,7 @@ XLA_TEST_P(ReduceWindowTest, R4SecondMinorStride) { Array4D input_array(2, 1, 27, 119); input_array.FillRandom(2.0f); std::unique_ptr input_literal = - Literal::CreateR4FromArray4DWithLayout( + LiteralUtil::CreateR4FromArray4DWithLayout( input_array, LayoutUtil::MakeLayout({3, 2, 1, 0})); XlaOp input; auto input_data = CreateParameterAndTransferLiteral( @@ -409,7 +410,7 @@ XLA_TEST_P(ReduceWindowTest, R4SecondMinorStride) { auto res = ReferenceUtil::ReduceWindow4DAdd( input_array, 0.0f, {1, 1, win_len, 1}, {1, 1, stride, 1}, padding); - ComputeAndCompareLiteral(&builder_, *Literal::CreateFromArray(*res), + ComputeAndCompareLiteral(&builder_, *LiteralUtil::CreateFromArray(*res), {input_data.get()}, DefaultErrorSpec()); } @@ -417,7 +418,7 @@ XLA_TEST_P(ReduceWindowTest, R4SecondMinorUnitStride) { Array4D input_array(3, 2, 4, 64); input_array.FillRandom(2.0f); std::unique_ptr input_literal = - Literal::CreateR4FromArray4DWithLayout( + LiteralUtil::CreateR4FromArray4DWithLayout( input_array, LayoutUtil::MakeLayout({3, 2, 1, 0})); XlaOp input; auto input_data = CreateParameterAndTransferLiteral( @@ -431,7 +432,7 @@ XLA_TEST_P(ReduceWindowTest, R4SecondMinorUnitStride) { auto res = ReferenceUtil::ReduceWindow4DAdd( input_array, 0.0f, {1, 1, win_len, 1}, {1, 1, stride, 1}, padding); - ComputeAndCompareLiteral(&builder_, *Literal::CreateFromArray(*res), + ComputeAndCompareLiteral(&builder_, *LiteralUtil::CreateFromArray(*res), {input_data.get()}, DefaultErrorSpec()); } @@ -439,7 +440,7 @@ XLA_TEST_P(ReduceWindowTest, R4SecondMinorWin) { Array4D input_array(1, 3, 12, 200); input_array.FillRandom(2.0f); std::unique_ptr input_literal = - Literal::CreateR4FromArray4DWithLayout( + LiteralUtil::CreateR4FromArray4DWithLayout( input_array, LayoutUtil::MakeLayout({3, 2, 1, 0})); XlaOp input; auto input_data = CreateParameterAndTransferLiteral( @@ -453,7 +454,7 @@ XLA_TEST_P(ReduceWindowTest, R4SecondMinorWin) { auto res = ReferenceUtil::ReduceWindow4DAdd( input_array, 0.0f, {1, 1, win_len, 1}, {1, 1, stride, 1}, padding); - ComputeAndCompareLiteral(&builder_, *Literal::CreateFromArray(*res), + ComputeAndCompareLiteral(&builder_, *LiteralUtil::CreateFromArray(*res), {input_data.get()}, DefaultErrorSpec()); } @@ -474,18 +475,18 @@ TEST_P(ReduceWindowTest, AmongMajor2DimsMultipleMinor) { auto result = ReferenceUtil::ReduceWindow4DAdd( input_array, 0.0f, {win_len, win_len, 1, 1}, {win_stride, win_stride, 1, 1}, padding); - ComputeAndCompareLiteral(&builder_, *Literal::CreateFromArray(*result), {}, - DefaultErrorSpec()); + ComputeAndCompareLiteral(&builder_, *LiteralUtil::CreateFromArray(*result), + {}, DefaultErrorSpec()); } XLA_TEST_P(ReduceWindowTest, Add24In1152_NoOverlap) { std::vector input_vector(128 * 9, 1); const auto input = CreateConstantFromLiteral( - *Literal::CreateR1(input_vector), &builder_); + *LiteralUtil::CreateR1(input_vector), &builder_); ReduceWindowAdd(input, {32}, {128}, Padding::kValid); ComputeAndCompareLiteral( &builder_, - *Literal::CreateR1({32, 32, 32, 32, 32, 32, 32, 32, 32}), {}, + *LiteralUtil::CreateR1({32, 32, 32, 32, 32, 32, 32, 32, 32}), {}, DefaultErrorSpec()); } @@ -500,9 +501,9 @@ XLA_TEST_P(ReduceWindowTest, Add128In128Stride128) { 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16}; const auto input = CreateConstantFromLiteral( - *Literal::CreateR1(input_vector), &builder_); + *LiteralUtil::CreateR1(input_vector), &builder_); ReduceWindowAdd(input, {128}, {128}, Padding::kValid); - ComputeAndCompareLiteral(&builder_, *Literal::CreateR1({1088}), {}, + ComputeAndCompareLiteral(&builder_, *LiteralUtil::CreateR1({1088}), {}, DefaultErrorSpec()); } @@ -517,9 +518,9 @@ XLA_TEST_P(ReduceWindowTest, Add128In128) { 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16}; const auto input = CreateConstantFromLiteral( - *Literal::CreateR1(input_vector), &builder_); + *LiteralUtil::CreateR1(input_vector), &builder_); ReduceWindowAdd(input, {128}, {1}, Padding::kValid); - ComputeAndCompareLiteral(&builder_, *Literal::CreateR1({1088}), {}, + ComputeAndCompareLiteral(&builder_, *LiteralUtil::CreateR1({1088}), {}, DefaultErrorSpec()); } @@ -536,14 +537,15 @@ TEST_P(ReduceWindowTest, R2ReduceWindowInceptionFromBroadcast) { auto res = ReferenceUtil::ReduceWindow2DAdd( input_array, 0.0f, {win_len, win_len}, {stride, stride}, padding); - ComputeAndCompareLiteral(&builder_, *Literal::CreateFromArray(*res), - {}, DefaultErrorSpec()); + ComputeAndCompareLiteral(&builder_, + *LiteralUtil::CreateFromArray(*res), {}, + DefaultErrorSpec()); } TEST_P(ReduceWindowTest, R2ReduceWindowNonOverlappingFromBroadcast) { Array2D input_array(6, 4, 1.0f); - XlaOp input = builder_.Broadcast( - CreateConstantFromLiteral(Literal::One(F32), &builder_), {6, 4}); + XlaOp input = Broadcast( + CreateConstantFromLiteral(LiteralUtil::One(F32), &builder_), {6, 4}); Padding padding = Padding::kSame; ReduceWindowAdd(input, {4, 2}, {3, 3}, padding); @@ -551,8 +553,9 @@ TEST_P(ReduceWindowTest, R2ReduceWindowNonOverlappingFromBroadcast) { auto res = ReferenceUtil::ReduceWindow2DAdd(input_array, 0.0f, {4, 2}, {3, 3}, padding); - ComputeAndCompareLiteral(&builder_, *Literal::CreateFromArray(*res), - {}, DefaultErrorSpec()); + ComputeAndCompareLiteral(&builder_, + *LiteralUtil::CreateFromArray(*res), {}, + DefaultErrorSpec()); } INSTANTIATE_TEST_CASE_P(ReduceWindowTestInstance, ReduceWindowTest, @@ -610,7 +613,7 @@ class R4ReduceWindowTest : public ReduceWindowTestBase, param.base_bounds[2], param.base_bounds[3]); input.FillIota(1); std::unique_ptr input_literal = - Literal::CreateR4FromArray4DWithLayout( + LiteralUtil::CreateR4FromArray4DWithLayout( input, LayoutUtil::MakeLayout(param.layout)); XlaOp parameter; auto input_arg = CreateParameterAndTransferLiteral(0, *input_literal, "p0", @@ -622,12 +625,12 @@ class R4ReduceWindowTest : public ReduceWindowTestBase, } auto init_value = - CreateConstantFromLiteral(*Literal::CreateR0(kInitValue), &b); + CreateConstantFromLiteral(*LiteralUtil::CreateR0(kInitValue), &b); CHECK(param.reducer == kAdd || param.reducer == kMax); auto computation = param.reducer == kAdd ? CreateScalarAddComputation(FloatType(), &b) : CreateScalarMaxComputation(FloatType(), &b); - b.ReduceWindowWithGeneralPadding( + ReduceWindowWithGeneralPadding( /*operand=*/parameter, /*init_value=*/init_value, /*computation=*/computation, @@ -648,7 +651,7 @@ class R4ReduceWindowTest : public ReduceWindowTestBase, /*stride=*/param.strides, /*padding=*/padding); std::unique_ptr expected_literal = - Literal::CreateFromArray(*expected); + LiteralUtil::CreateFromArray(*expected); const Shape& expected_shape_with_layout = ShapeUtil::MakeShapeWithLayout( input_literal->shape().element_type(), AsInt64Slice(expected_literal->shape().dimensions()), param.layout); @@ -960,25 +963,25 @@ TEST_P(R3ReduceWindowTest, Add) { Array3D input(param.base_bounds[0], param.base_bounds[1], param.base_bounds[2], 1.0f); std::unique_ptr input_literal = - Literal::CreateR3FromArray3DWithLayout( + LiteralUtil::CreateR3FromArray3DWithLayout( input, LayoutUtil::MakeLayout(param.layout)); XlaOp parameter; auto input_arg = CreateParameterAndTransferLiteral(0, *input_literal, "p0", &b, ¶meter); auto init_value = - CreateConstantFromLiteral(*Literal::CreateR0(kInitValue), &b); - b.ReduceWindow(/*operand=*/parameter, - /*init_value=*/init_value, - /*computation=*/CreateScalarAddComputation(FloatType(), &b), - /*window_dimensions=*/param.window_bounds, - /*window_strides=*/param.strides, /*padding=*/param.padding); + CreateConstantFromLiteral(*LiteralUtil::CreateR0(kInitValue), &b); + ReduceWindow(/*operand=*/parameter, + /*init_value=*/init_value, + /*computation=*/CreateScalarAddComputation(FloatType(), &b), + /*window_dimensions=*/param.window_bounds, + /*window_strides=*/param.strides, /*padding=*/param.padding); auto expected = ReferenceUtil::ReduceWindow3DAdd( /*operand=*/input, /*init=*/kInitValue, /*window=*/param.window_bounds, /*stride=*/param.strides, /*padding=*/param.padding); - ComputeAndCompareLiteral(&b, *Literal::CreateFromArray(*expected), + ComputeAndCompareLiteral(&b, *LiteralUtil::CreateFromArray(*expected), {input_arg.get()}, DefaultErrorSpec()); } @@ -1094,7 +1097,7 @@ class R2ReduceWindowTest : public ReduceWindowTestBase, const float kInitValue = 0.0f; Array2D input(param.base_bounds[0], param.base_bounds[1], 1.0f); std::unique_ptr input_literal = - Literal::CreateR2FromArray2DWithLayout( + LiteralUtil::CreateR2FromArray2DWithLayout( input, LayoutUtil::MakeLayout(param.layout)); XlaOp parameter; @@ -1108,8 +1111,8 @@ class R2ReduceWindowTest : public ReduceWindowTestBase, ? CreateScalarAddComputation(FloatType(), &b) : CreateScalarMaxComputation(FloatType(), &b); auto init_value = - CreateConstantFromLiteral(*Literal::CreateR0(kInitValue), &b); - b.ReduceWindowWithGeneralPadding( + CreateConstantFromLiteral(*LiteralUtil::CreateR0(kInitValue), &b); + ReduceWindowWithGeneralPadding( /*operand=*/parameter, /*init_value=*/init_value, /*computation=*/computation, @@ -1124,7 +1127,7 @@ class R2ReduceWindowTest : public ReduceWindowTestBase, /*window=*/param.window_bounds, /*stride=*/param.strides, /*padding=*/padding); - ComputeAndCompareLiteral(&b, *Literal::CreateFromArray(*expected), + ComputeAndCompareLiteral(&b, *LiteralUtil::CreateFromArray(*expected), {input_arg.get()}, DefaultErrorSpec()); } }; @@ -1293,7 +1296,7 @@ TEST_P(R1ReduceWindowTest, DoIt) { std::vector input_vector(param.base_bounds[0]); std::iota(std::begin(input_vector), std::end(input_vector), 0); std::unique_ptr input_literal = - Literal::CreateR1(tensorflow::gtl::ArraySlice(input_vector)); + LiteralUtil::CreateR1(tensorflow::gtl::ArraySlice(input_vector)); XlaOp parameter; auto input_arg = CreateParameterAndTransferLiteral(0, *input_literal, "p0", &b, ¶meter); @@ -1305,8 +1308,8 @@ TEST_P(R1ReduceWindowTest, DoIt) { ? CreateScalarAddComputation(FloatType(), &b) : CreateScalarMaxComputation(FloatType(), &b); auto init_value = - CreateConstantFromLiteral(*Literal::CreateR0(kInitValue), &b); - b.ReduceWindowWithGeneralPadding( + CreateConstantFromLiteral(*LiteralUtil::CreateR0(kInitValue), &b); + ReduceWindowWithGeneralPadding( /*operand=*/parameter, /*init_value=*/init_value, /*computation=*/computation, @@ -1324,7 +1327,7 @@ TEST_P(R1ReduceWindowTest, DoIt) { /*stride=*/param.strides, /*padding=*/padding); - ComputeAndCompareLiteral(&b, *Literal::CreateR1(*expected), + ComputeAndCompareLiteral(&b, *LiteralUtil::CreateR1(*expected), {input_arg.get()}, DefaultErrorSpec()); } diff --git a/tensorflow/compiler/xla/tests/replay_test.cc b/tensorflow/compiler/xla/tests/replay_test.cc index 36d763b0f7f4267ede076c0b25cfaf9654e96e0d..d544968648d7602464bd141a12c75eeb8c1678da 100644 --- a/tensorflow/compiler/xla/tests/replay_test.cc +++ b/tensorflow/compiler/xla/tests/replay_test.cc @@ -19,7 +19,7 @@ limitations under the License. #include "tensorflow/compiler/xla/client/local_client.h" #include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/compiler/xla/client/xla_client/xla_computation.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/protobuf_util.h" #include "tensorflow/compiler/xla/service/hlo.pb.h" #include "tensorflow/compiler/xla/shape_util.h" @@ -39,8 +39,8 @@ class ReplayTest : public ClientLibraryTestBase {}; TEST_F(ReplayTest, TwoPlusTwoReplay) { // Make 2+2 computation. XlaBuilder builder(TestName()); - auto two = builder.ConstantR0(2); - builder.Add(two, two); + auto two = ConstantR0(&builder, 2); + Add(two, two); XlaComputation computation = builder.Build().ConsumeValueOrDie(); // Serialize it out. @@ -70,9 +70,9 @@ TEST_F(ReplayTest, TwoPlusTwoReplay) { XLA_TEST_F(ReplayTest, XPlusYReplayWithParameters) { // Make computation. XlaBuilder builder(TestName()); - auto x = builder.Parameter(0, ShapeUtil::MakeShape(S32, {}), "x"); - auto y = builder.Parameter(1, ShapeUtil::MakeShape(S32, {}), "y"); - builder.Add(x, y); + auto x = Parameter(&builder, 0, ShapeUtil::MakeShape(S32, {}), "x"); + auto y = Parameter(&builder, 1, ShapeUtil::MakeShape(S32, {}), "y"); + Add(x, y); XlaComputation computation = builder.Build().ConsumeValueOrDie(); // Serialize it out. @@ -91,10 +91,10 @@ XLA_TEST_F(ReplayTest, XPlusYReplayWithParameters) { // Run it. std::unique_ptr x_data = - client_->TransferToServer(*Literal::CreateR0(2)) + client_->TransferToServer(*LiteralUtil::CreateR0(2)) .ConsumeValueOrDie(); std::unique_ptr y_data = - client_->TransferToServer(*Literal::CreateR0(3)) + client_->TransferToServer(*LiteralUtil::CreateR0(3)) .ConsumeValueOrDie(); std::unique_ptr literal = client_ @@ -111,13 +111,13 @@ TEST_F(ReplayTest, MapPlusTwoOverR1) { // As above, but with map(+2) over some constant array. XlaBuilder plus_two_builder("plus two"); auto input = - plus_two_builder.Parameter(0, ShapeUtil::MakeShape(S32, {}), "input"); - plus_two_builder.Add(input, plus_two_builder.ConstantR0(2)); + Parameter(&plus_two_builder, 0, ShapeUtil::MakeShape(S32, {}), "input"); + Add(input, ConstantR0(&plus_two_builder, 2)); XlaComputation plus_two = plus_two_builder.Build().ConsumeValueOrDie(); XlaBuilder mapper_builder(TestName()); - auto original = mapper_builder.ConstantR1({1, 2, 3}); - mapper_builder.Map({original}, plus_two, {0}); + auto original = ConstantR1(&mapper_builder, {1, 2, 3}); + Map(&mapper_builder, {original}, plus_two, {0}); XlaComputation computation = mapper_builder.Build().ConsumeValueOrDie(); diff --git a/tensorflow/compiler/xla/tests/reshape_motion_test.cc b/tensorflow/compiler/xla/tests/reshape_motion_test.cc index da1b588ec41cef711412367e89b2a9b1029bca71..7c0389cfa3251a6b62f83a78e986d870177d4d91 100644 --- a/tensorflow/compiler/xla/tests/reshape_motion_test.cc +++ b/tensorflow/compiler/xla/tests/reshape_motion_test.cc @@ -24,7 +24,7 @@ limitations under the License. #include "tensorflow/compiler/xla/client/local_client.h" #include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/compiler/xla/layout_util.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/reference_util.h" #include "tensorflow/compiler/xla/shape_util.h" #include "tensorflow/compiler/xla/status_macros.h" @@ -44,11 +44,11 @@ using ReshapeMotionTest = ClientLibraryTestBase; TEST_F(ReshapeMotionTest, ElementwiseOfReshapesWithNonSameInputShapes) { XlaBuilder builder(TestName()); - auto a = builder.ConstantR2({{2, 3, 5}, {7, 11, 13}}); - auto b = builder.ConstantR2({{17, 19}, {23, 29}, {31, 37}}); - auto c = builder.Reshape(a, {6}); - auto d = builder.Reshape(b, {6}); - auto e = builder.Mul(c, d); + auto a = ConstantR2(&builder, {{2, 3, 5}, {7, 11, 13}}); + auto b = ConstantR2(&builder, {{17, 19}, {23, 29}, {31, 37}}); + auto c = Reshape(a, {6}); + auto d = Reshape(b, {6}); + Mul(c, d); ComputeAndCompareR1(&builder, {34, 57, 115, 203, 341, 481}, {}); } diff --git a/tensorflow/compiler/xla/tests/reshape_test.cc b/tensorflow/compiler/xla/tests/reshape_test.cc index a4580cd71d46ad0a0186eddd51291f9c322b6f49..46d91711a55c5aa0ad906bc9ba0265fab194cf1a 100644 --- a/tensorflow/compiler/xla/tests/reshape_test.cc +++ b/tensorflow/compiler/xla/tests/reshape_test.cc @@ -55,39 +55,39 @@ XLA_TEST_P(ReshapeTest, CollapseTrivial1x1) { XlaBuilder builder(TestName()); Array2D input_array(1, 1); input_array.Fill(1.0f); - auto input_literal = Literal::CreateR2FromArray2D(input_array); + auto input_literal = LiteralUtil::CreateR2FromArray2D(input_array); XlaOp parameter; auto input = CreateParameterAndTransferLiteral(0, *input_literal, "parameter", &builder, ¶meter); - builder.Collapse(/*operand=*/parameter, /*dimensions=*/{0, 1}); + Collapse(/*operand=*/parameter, /*dimensions=*/{0, 1}); - auto expected_literal = Literal::CreateR1({1.0f}); + auto expected_literal = LiteralUtil::CreateR1({1.0f}); ComputeAndCompareLiteral(&builder, *expected_literal, {input.get()}, zero_error_spec_); } XLA_TEST_P(ReshapeTest, CollapseTrivialR1EmptyDims) { XlaBuilder builder(TestName()); - auto input_literal = Literal::CreateR1({1.0f}); + auto input_literal = LiteralUtil::CreateR1({1.0f}); XlaOp parameter; auto input = CreateParameterAndTransferLiteral(0, *input_literal, "parameter", &builder, ¶meter); - builder.Collapse(/*operand=*/parameter, /*dimensions=*/{}); + Collapse(/*operand=*/parameter, /*dimensions=*/{}); - auto expected_literal = Literal::CreateR1({1.0f}); + auto expected_literal = LiteralUtil::CreateR1({1.0f}); ComputeAndCompareLiteral(&builder, *expected_literal, {input.get()}, zero_error_spec_); } XLA_TEST_P(ReshapeTest, CollapseTrivialR1OnlyDim) { XlaBuilder builder(TestName()); - auto input_literal = Literal::CreateR1({1.0f}); + auto input_literal = LiteralUtil::CreateR1({1.0f}); XlaOp parameter; auto input = CreateParameterAndTransferLiteral(0, *input_literal, "parameter", &builder, ¶meter); - builder.Collapse(/*operand=*/parameter, /*dimensions=*/{0}); + Collapse(/*operand=*/parameter, /*dimensions=*/{0}); - auto expected_literal = Literal::CreateR1({1.0f}); + auto expected_literal = LiteralUtil::CreateR1({1.0f}); ComputeAndCompareLiteral(&builder, *expected_literal, {input.get()}, zero_error_spec_); } @@ -97,15 +97,15 @@ XLA_TEST_P(ReshapeTest, SingleElementArrayToScalar) { XlaBuilder builder(TestName()); Array2D input_array(1, 1); input_array.Fill(1.0f); - auto input_literal = Literal::CreateR2FromArray2D(input_array); + auto input_literal = LiteralUtil::CreateR2FromArray2D(input_array); XlaOp parameter; auto input = CreateParameterAndTransferLiteral(0, *input_literal, "parameter", &builder, ¶meter); - auto reshape = builder.Reshape(/*operand=*/parameter, /*dimensions=*/{0, 1}, - /*new_sizes=*/{}); + auto reshape = Reshape(/*operand=*/parameter, /*dimensions=*/{0, 1}, + /*new_sizes=*/{}); auto new_shape = builder.GetShape(reshape).ConsumeValueOrDie(); - auto expected_literal = Literal::CreateR0(1.0f); + auto expected_literal = LiteralUtil::CreateR0(1.0f); ComputeAndCompareLiteral(&builder, *expected_literal, {input.get()}, zero_error_spec_); } @@ -113,63 +113,54 @@ XLA_TEST_P(ReshapeTest, SingleElementArrayToScalar) { XLA_TEST_P(ReshapeTest, ScalarToSingleElementArray) { XlaBuilder builder(TestName()); - std::unique_ptr param0_literal = Literal::CreateR0(1.0f); + std::unique_ptr param0_literal = LiteralUtil::CreateR0(1.0f); XlaOp parameter; auto input = CreateParameterAndTransferLiteral(0, *param0_literal, "param0", &builder, ¶meter); - auto a = builder.Neg(parameter); - builder.Reshape(/*operand=*/a, /*dimensions=*/{}, /*new_sizes=*/{1}); + auto a = Neg(parameter); + Reshape(/*operand=*/a, /*dimensions=*/{}, /*new_sizes=*/{1}); - auto expected_literal = Literal::CreateR1({-1.0f}); + auto expected_literal = LiteralUtil::CreateR1({-1.0f}); ComputeAndCompareLiteral(&builder, *expected_literal, {input.get()}, zero_error_spec_); } -// TODO(b/29185393): Make this work with the GPU backend. The GPU backend -// does not handle zero-sized shapes correctly. Failed last on 2017-11-30 -// with an incorrect result rank. -XLA_TEST_P(ReshapeTest, DISABLED_ON_GPU(Trivial0x3)) { +XLA_TEST_P(ReshapeTest, Trivial0x3) { XlaBuilder builder(TestName()); Array2D input_array(0, 3); - auto input_literal = Literal::CreateR2FromArray2D(input_array); + auto input_literal = LiteralUtil::CreateR2FromArray2D(input_array); XlaOp parameter; auto input = CreateParameterAndTransferLiteral(0, *input_literal, "input", &builder, ¶meter); - builder.Collapse(/*operand=*/parameter, /*dimensions=*/{0, 1}); - auto expected_literal = Literal::CreateR1({}); + Collapse(/*operand=*/parameter, /*dimensions=*/{0, 1}); + auto expected_literal = LiteralUtil::CreateR1({}); ComputeAndCompareLiteral(&builder, *expected_literal, {input.get()}, zero_error_spec_); } -// TODO(b/29185393): Make this work with the GPU backend. The GPU backend -// does not handle zero-sized shapes correctly. Failed last on 2017-05-15 -// with an incorrect result rank. -XLA_TEST_P(ReshapeTest, DISABLED_ON_GPU(Trivial0x3WithParameter)) { +XLA_TEST_P(ReshapeTest, Trivial0x3WithParameter) { XlaBuilder builder(TestName()); std::unique_ptr param0_literal = - Literal::CreateR2FromArray2D(Array2D(0, 3)); + LiteralUtil::CreateR2FromArray2D(Array2D(0, 3)); XlaOp parameter; auto input = CreateParameterAndTransferLiteral(0, *param0_literal, "param0", &builder, ¶meter); - builder.Collapse(/*operand=*/parameter, /*dimensions=*/{0, 1}); - auto expected_literal = Literal::CreateR1({}); + Collapse(/*operand=*/parameter, /*dimensions=*/{0, 1}); + auto expected_literal = LiteralUtil::CreateR1({}); ComputeAndCompareLiteral(&builder, *expected_literal, {input.get()}, zero_error_spec_); } -// TODO(b/29185393): Make this work with the GPU backend. The GPU backend -// does not handle zero-sized shapes correctly. Failed last on 2017-11-30 -// with an incorrect result rank. -XLA_TEST_P(ReshapeTest, DISABLED_ON_GPU(Trivial3x0)) { +XLA_TEST_P(ReshapeTest, Trivial3x0) { XlaBuilder builder(TestName()); Array2D input_array(3, 0); - auto input_literal = Literal::CreateR2FromArray2D(input_array); + auto input_literal = LiteralUtil::CreateR2FromArray2D(input_array); XlaOp parameter; auto input = CreateParameterAndTransferLiteral(0, *input_literal, "input", &builder, ¶meter); - builder.Collapse(/*operand=*/parameter, /*dimensions=*/{0, 1}); - auto expected_literal = Literal::CreateR1({}); + Collapse(/*operand=*/parameter, /*dimensions=*/{0, 1}); + auto expected_literal = LiteralUtil::CreateR1({}); ComputeAndCompareLiteral(&builder, *expected_literal, {input.get()}, zero_error_spec_); } @@ -177,12 +168,12 @@ XLA_TEST_P(ReshapeTest, DISABLED_ON_GPU(Trivial3x0)) { // Collapses a 2-dimensional row vector to 1 dimension. XLA_TEST_P(ReshapeTest, Trivial1x3) { XlaBuilder builder(TestName()); - auto input_literal = Literal::CreateR2({{1.0f, 2.0f, 3.0f}}); + auto input_literal = LiteralUtil::CreateR2({{1.0f, 2.0f, 3.0f}}); XlaOp parameter; auto input = CreateParameterAndTransferLiteral(0, *input_literal, "input", &builder, ¶meter); - builder.Collapse(/*operand=*/parameter, /*dimensions=*/{0, 1}); - auto expected_literal = Literal::CreateR1({1.0f, 2.0f, 3.0f}); + Collapse(/*operand=*/parameter, /*dimensions=*/{0, 1}); + auto expected_literal = LiteralUtil::CreateR1({1.0f, 2.0f, 3.0f}); ComputeAndCompareLiteral(&builder, *expected_literal, {input.get()}, zero_error_spec_); } @@ -190,30 +181,26 @@ XLA_TEST_P(ReshapeTest, Trivial1x3) { // Collapses a 2-dimensional column vector to 1 dimension. XLA_TEST_P(ReshapeTest, Trivial3x1) { XlaBuilder builder(TestName()); - auto input_literal = Literal::CreateR2({{1.0f}, {2.0f}, {3.0f}}); + auto input_literal = LiteralUtil::CreateR2({{1.0f}, {2.0f}, {3.0f}}); XlaOp parameter; auto input = CreateParameterAndTransferLiteral(0, *input_literal, "input", &builder, ¶meter); - builder.Collapse(/*operand=*/parameter, /*dimensions=*/{0, 1}); - auto expected_literal = Literal::CreateR1({1.0f, 2.0f, 3.0f}); + Collapse(/*operand=*/parameter, /*dimensions=*/{0, 1}); + auto expected_literal = LiteralUtil::CreateR1({1.0f, 2.0f, 3.0f}); ComputeAndCompareLiteral(&builder, *expected_literal, {input.get()}, zero_error_spec_); } -// TODO(b/29185393): Make this work with the GPU backend. The GPU backend -// does not handle zero-sized shapes correctly. Failed last on 2017-11-30 -// with an incorrect result rank. -// // Splits an empty vector into an empty matrix. -XLA_TEST_P(ReshapeTest, DISABLED_ON_GPU(R1ToR2_0_To_2x0)) { +XLA_TEST_P(ReshapeTest, R1ToR2_0_To_2x0) { XlaBuilder builder(TestName()); - auto input_literal = Literal::CreateR1({}); + auto input_literal = LiteralUtil::CreateR1({}); XlaOp parameter; auto input = CreateParameterAndTransferLiteral(0, *input_literal, "input", &builder, ¶meter); - builder.Reshape(/*operand=*/parameter, /*dimensions=*/{0}, - /*new_sizes=*/{2, 0}); - auto expected_literal = Literal::CreateR2({{}, {}}); + Reshape(/*operand=*/parameter, /*dimensions=*/{0}, + /*new_sizes=*/{2, 0}); + auto expected_literal = LiteralUtil::CreateR2({{}, {}}); ComputeAndCompareLiteral(&builder, *expected_literal, {input.get()}, zero_error_spec_); } @@ -222,32 +209,28 @@ XLA_TEST_P(ReshapeTest, DISABLED_ON_GPU(R1ToR2_0_To_2x0)) { XLA_TEST_P(ReshapeTest, R1ToR2_6_To_2x3) { XlaBuilder builder(TestName()); auto input_literal = - Literal::CreateR1({1.0f, 2.0f, 3.0f, 4.0f, 5.0f, 6.0f}); + LiteralUtil::CreateR1({1.0f, 2.0f, 3.0f, 4.0f, 5.0f, 6.0f}); XlaOp parameter; auto input = CreateParameterAndTransferLiteral(0, *input_literal, "input", &builder, ¶meter); - builder.Reshape(/*operand=*/parameter, /*dimensions=*/{0}, - /*new_sizes=*/{2, 3}); + Reshape(/*operand=*/parameter, /*dimensions=*/{0}, + /*new_sizes=*/{2, 3}); auto expected_literal = - Literal::CreateR2({{1.0f, 2.0f, 3.0f}, {4.0f, 5.0f, 6.0f}}); + LiteralUtil::CreateR2({{1.0f, 2.0f, 3.0f}, {4.0f, 5.0f, 6.0f}}); ComputeAndCompareLiteral(&builder, *expected_literal, {input.get()}, zero_error_spec_); } -// TODO(b/29185393): Make this work with the GPU backend. The GPU backend -// does not handle zero-sized shapes correctly. Failed last on 2017-11-30 -// with an incorrect result rank. -// // Transposes a 2x0 array to a 0x2 array. -XLA_TEST_P(ReshapeTest, DISABLED_ON_GPU(Reshape0x2To2x0)) { +XLA_TEST_P(ReshapeTest, Reshape0x2To2x0) { XlaBuilder builder(TestName()); - auto input_literal = Literal::CreateFromArray(Array2D(0, 2)); + auto input_literal = LiteralUtil::CreateFromArray(Array2D(0, 2)); XlaOp parameter; auto input = CreateParameterAndTransferLiteral(0, *input_literal, "input", &builder, ¶meter); - builder.Reshape(/*operand=*/parameter, /*dimensions=*/{0, 1}, - /*new_sizes=*/{2, 0}); - auto expected_literal = Literal::CreateR2({{}, {}}); + Reshape(/*operand=*/parameter, /*dimensions=*/{0, 1}, + /*new_sizes=*/{2, 0}); + auto expected_literal = LiteralUtil::CreateR2({{}, {}}); ComputeAndCompareLiteral(&builder, *expected_literal, {input.get()}, zero_error_spec_); } @@ -256,15 +239,15 @@ XLA_TEST_P(ReshapeTest, DISABLED_ON_GPU(Reshape0x2To2x0)) { XLA_TEST_P(ReshapeTest, ReshapeRowToCol) { XlaBuilder builder(TestName()); auto simple = MakeLinspaceArray2D(1.0f, 3.0f, 1, 3); - auto input_literal = Literal::CreateFromArray(*simple); + auto input_literal = LiteralUtil::CreateFromArray(*simple); XlaOp parameter; auto input = CreateParameterAndTransferLiteral(0, *input_literal, "input", &builder, ¶meter); - builder.Reshape(/*operand=*/parameter, /*dimensions=*/{0, 1}, - /*new_sizes=*/{3, 1}); + Reshape(/*operand=*/parameter, /*dimensions=*/{0, 1}, + /*new_sizes=*/{3, 1}); auto expected = ReferenceUtil::TransposeArray2D(*simple); - auto expected_literal = Literal::CreateFromArray(*expected); + auto expected_literal = LiteralUtil::CreateFromArray(*expected); ComputeAndCompareLiteral(&builder, *expected_literal, {input.get()}, zero_error_spec_); } @@ -273,32 +256,28 @@ XLA_TEST_P(ReshapeTest, ReshapeRowToCol) { XLA_TEST_P(ReshapeTest, TransposeAsReshape) { XlaBuilder builder(TestName()); auto a4x3 = MakeLinspaceArray2D(1.0f, 12.0f, 4, 3); - auto input_literal = Literal::CreateFromArray(*a4x3); + auto input_literal = LiteralUtil::CreateFromArray(*a4x3); XlaOp parameter; auto input = CreateParameterAndTransferLiteral(0, *input_literal, "input", &builder, ¶meter); - builder.Reshape(/*operand=*/parameter, /*dimensions=*/{1, 0}, - /*new_sizes=*/{3, 4}); + Reshape(/*operand=*/parameter, /*dimensions=*/{1, 0}, + /*new_sizes=*/{3, 4}); auto expected = ReferenceUtil::TransposeArray2D(*a4x3); - auto expected_literal = Literal::CreateFromArray(*expected); + auto expected_literal = LiteralUtil::CreateFromArray(*expected); ComputeAndCompareLiteral(&builder, *expected_literal, {input.get()}, zero_error_spec_); } -// TODO(b/29185393): Make this work with the GPU backend. The GPU backend -// does not handle zero-sized shapes correctly. Failed last on 2017-11-30 -// with an incorrect result rank. -// // Transposes a 0x4 array with XlaBuilder::Transpose. -XLA_TEST_P(ReshapeTest, DISABLED_ON_GPU(Transpose0x4)) { +XLA_TEST_P(ReshapeTest, Transpose0x4) { XlaBuilder builder(TestName()); - auto input_literal = Literal::CreateFromArray(Array2D(0, 4)); + auto input_literal = LiteralUtil::CreateFromArray(Array2D(0, 4)); XlaOp parameter; auto input = CreateParameterAndTransferLiteral(0, *input_literal, "input", &builder, ¶meter); - builder.Transpose(parameter, {1, 0}); - auto expected_literal = Literal::CreateR2({{}, {}, {}, {}}); + Transpose(parameter, {1, 0}); + auto expected_literal = LiteralUtil::CreateR2({{}, {}, {}, {}}); ComputeAndCompareLiteral(&builder, *expected_literal, {input.get()}, zero_error_spec_); } @@ -307,49 +286,43 @@ XLA_TEST_P(ReshapeTest, DISABLED_ON_GPU(Transpose0x4)) { XLA_TEST_P(ReshapeTest, Transpose4x3) { XlaBuilder builder(TestName()); auto a4x3 = MakeLinspaceArray2D(1.0f, 12.0f, 4, 3); - auto input_literal = Literal::CreateFromArray(*a4x3); + auto input_literal = LiteralUtil::CreateFromArray(*a4x3); XlaOp parameter; auto input = CreateParameterAndTransferLiteral(0, *input_literal, "input", &builder, ¶meter); - builder.Transpose(parameter, {1, 0}); + Transpose(parameter, {1, 0}); auto expected = ReferenceUtil::TransposeArray2D(*a4x3); - auto expected_literal = Literal::CreateFromArray(*expected); + auto expected_literal = LiteralUtil::CreateFromArray(*expected); ComputeAndCompareLiteral(&builder, *expected_literal, {input.get()}, zero_error_spec_); } -// TODO(b/29185393): Make this work with the GPU backend. The GPU backend -// does not handle zero-sized shapes correctly. Failed last on 2017-11-30 -// with an incorrect result rank. -// // Reshapes an empty 2-dimensional array with dimensions that are not just a // rearrangement of the originals (split), but no reordering (no shuffle). -XLA_TEST_P(ReshapeTest, DISABLED_ON_GPU(ReshapeSplitNoShuffleZeroElements)) { +XLA_TEST_P(ReshapeTest, ReshapeSplitNoShuffleZeroElements) { XlaBuilder builder(TestName()); - auto input_literal = Literal::CreateFromArray(Array2D(6, 0)); + auto input_literal = LiteralUtil::CreateFromArray(Array2D(6, 0)); XlaOp parameter; auto input = CreateParameterAndTransferLiteral(0, *input_literal, "input", &builder, ¶meter); - builder.Reshape(/*operand=*/parameter, /*dimensions=*/{0, 1}, - /*new_sizes=*/{2, 3, 0, 0}); - auto expected_literal = Literal::CreateFromArray(Array4D(2, 3, 0, 0)); + Reshape(/*operand=*/parameter, /*dimensions=*/{0, 1}, + /*new_sizes=*/{2, 3, 0, 0}); + auto expected_literal = + LiteralUtil::CreateFromArray(Array4D(2, 3, 0, 0)); ComputeAndCompareLiteral(&builder, *expected_literal, {input.get()}, zero_error_spec_); } -// TODO(b/29185393): Make this work with the GPU backend. The GPU backend -// does not handle zero-sized shapes correctly. Failed last on 2017-11-30 -// with an incorrect result rank. -XLA_TEST_P(ReshapeTest, DISABLED_ON_GPU(ReshapeR4ToR2ZeroElements)) { +XLA_TEST_P(ReshapeTest, ReshapeR4ToR2ZeroElements) { XlaBuilder builder(TestName()); - auto input_literal = Literal::CreateFromArray(Array4D(2, 3, 4, 0)); + auto input_literal = LiteralUtil::CreateFromArray(Array4D(2, 3, 4, 0)); XlaOp parameter; auto input = CreateParameterAndTransferLiteral(0, *input_literal, "input", &builder, ¶meter); - builder.Reshape(/*operand=*/parameter, /*dimensions=*/{0, 1, 2, 3}, - /*new_sizes=*/{24, 0}); - auto expected_literal = Literal::CreateFromArray(Array2D(24, 0)); + Reshape(/*operand=*/parameter, /*dimensions=*/{0, 1, 2, 3}, + /*new_sizes=*/{24, 0}); + auto expected_literal = LiteralUtil::CreateFromArray(Array2D(24, 0)); ComputeAndCompareLiteral(&builder, *expected_literal, {input.get()}, zero_error_spec_); } @@ -359,32 +332,28 @@ XLA_TEST_P(ReshapeTest, DISABLED_ON_GPU(ReshapeR4ToR2ZeroElements)) { XLA_TEST_P(ReshapeTest, ReshapeSplitNoShuffle) { XlaBuilder builder(TestName()); auto a4x3 = MakeLinspaceArray2D(1.0f, 12.0f, 4, 3); - auto input_literal = Literal::CreateFromArray(*a4x3); + auto input_literal = LiteralUtil::CreateFromArray(*a4x3); XlaOp parameter; auto input = CreateParameterAndTransferLiteral(0, *input_literal, "input", &builder, ¶meter); - builder.Reshape(/*operand=*/parameter, /*dimensions=*/{0, 1}, - /*new_sizes=*/{2, 6}); + Reshape(/*operand=*/parameter, /*dimensions=*/{0, 1}, + /*new_sizes=*/{2, 6}); auto expected = MakeLinspaceArray2D(1.0f, 12.0f, 2, 6); - auto expected_literal = Literal::CreateFromArray(*expected); + auto expected_literal = LiteralUtil::CreateFromArray(*expected); ComputeAndCompareLiteral(&builder, *expected_literal, {input.get()}, zero_error_spec_); } -// TODO(b/29185393): Make this work with the GPU backend. The GPU backend -// does not handle zero-sized shapes correctly. Failed last on 2017-11-30 -// with an incorrect result rank. -// -XLA_TEST_P(ReshapeTest, DISABLED_ON_GPU(ReshapeSplitAndShuffleZeroElements)) { +XLA_TEST_P(ReshapeTest, ReshapeSplitAndShuffleZeroElements) { XlaBuilder builder(TestName()); - auto input_literal = Literal::CreateFromArray(Array2D(0, 6)); + auto input_literal = LiteralUtil::CreateFromArray(Array2D(0, 6)); XlaOp parameter; auto input = CreateParameterAndTransferLiteral(0, *input_literal, "input", &builder, ¶meter); - builder.Reshape(/*operand=*/parameter, /*dimensions=*/{1, 0}, - /*new_sizes=*/{3, 0}); - auto expected_literal = Literal::CreateFromArray(Array2D(3, 0)); + Reshape(/*operand=*/parameter, /*dimensions=*/{1, 0}, + /*new_sizes=*/{3, 0}); + auto expected_literal = LiteralUtil::CreateFromArray(Array2D(3, 0)); ComputeAndCompareLiteral(&builder, *expected_literal, {input.get()}, zero_error_spec_); } @@ -394,15 +363,15 @@ XLA_TEST_P(ReshapeTest, DISABLED_ON_GPU(ReshapeSplitAndShuffleZeroElements)) { XLA_TEST_P(ReshapeTest, ReshapeSplitAndShuffle) { XlaBuilder builder(TestName()); auto a4x3 = MakeLinspaceArray2D(1.0f, 12.0f, 4, 3); - auto input_literal = Literal::CreateFromArray(*a4x3); + auto input_literal = LiteralUtil::CreateFromArray(*a4x3); XlaOp parameter; auto input = CreateParameterAndTransferLiteral(0, *input_literal, "input", &builder, ¶meter); - builder.Reshape(/*operand=*/parameter, /*dimensions=*/{1, 0}, - /*new_sizes=*/{2, 6}); + Reshape(/*operand=*/parameter, /*dimensions=*/{1, 0}, + /*new_sizes=*/{2, 6}); Array2D expected({{1.0f, 4.0f, 7.0f, 10.0f, 2.0f, 5.0f}, {8.0f, 11.0f, 3.0f, 6.0f, 9.0f, 12.0f}}); - auto expected_literal = Literal::CreateFromArray(expected); + auto expected_literal = LiteralUtil::CreateFromArray(expected); ComputeAndCompareLiteral(&builder, *expected_literal, {input.get()}, zero_error_spec_); } @@ -420,13 +389,13 @@ static Array3D ArrayForDocR3Tests() { XLA_TEST_P(ReshapeTest, DocR3_R1_Collapse_012) { XlaBuilder builder(TestName()); - auto input_literal = Literal::CreateFromArray(ArrayForDocR3Tests()); + auto input_literal = LiteralUtil::CreateFromArray(ArrayForDocR3Tests()); XlaOp parameter; auto input = CreateParameterAndTransferLiteral(0, *input_literal, "input", &builder, ¶meter); - builder.Reshape(/*operand=*/parameter, /*dimensions=*/{0, 1, 2}, - /*new_sizes=*/{24}); - auto expected_literal = Literal::CreateR1( + Reshape(/*operand=*/parameter, /*dimensions=*/{0, 1, 2}, + /*new_sizes=*/{24}); + auto expected_literal = LiteralUtil::CreateR1( {10, 11, 12, 15, 16, 17, 20, 21, 22, 25, 26, 27, 30, 31, 32, 35, 36, 37, 40, 41, 42, 45, 46, 47}); ComputeAndCompareLiteral(&builder, *expected_literal, {input.get()}, @@ -435,33 +404,33 @@ XLA_TEST_P(ReshapeTest, DocR3_R1_Collapse_012) { XLA_TEST_P(ReshapeTest, DocR3_R2_Collapse_012_Refine_83) { XlaBuilder builder(TestName()); - auto input_literal = Literal::CreateFromArray(ArrayForDocR3Tests()); + auto input_literal = LiteralUtil::CreateFromArray(ArrayForDocR3Tests()); XlaOp parameter; auto input = CreateParameterAndTransferLiteral(0, *input_literal, "input", &builder, ¶meter); - builder.Reshape(/*operand=*/parameter, /*dimensions=*/{0, 1, 2}, - /*new_sizes=*/{8, 3}); - auto expected_literal = Literal::CreateR2({{10, 11, 12}, - {15, 16, 17}, - {20, 21, 22}, - {25, 26, 27}, - {30, 31, 32}, - {35, 36, 37}, - {40, 41, 42}, - {45, 46, 47}}); + Reshape(/*operand=*/parameter, /*dimensions=*/{0, 1, 2}, + /*new_sizes=*/{8, 3}); + auto expected_literal = LiteralUtil::CreateR2({{10, 11, 12}, + {15, 16, 17}, + {20, 21, 22}, + {25, 26, 27}, + {30, 31, 32}, + {35, 36, 37}, + {40, 41, 42}, + {45, 46, 47}}); ComputeAndCompareLiteral(&builder, *expected_literal, {input.get()}, zero_error_spec_); } XLA_TEST_P(ReshapeTest, DocR3_R1_Collapse_120) { XlaBuilder builder(TestName()); - auto input_literal = Literal::CreateFromArray(ArrayForDocR3Tests()); + auto input_literal = LiteralUtil::CreateFromArray(ArrayForDocR3Tests()); XlaOp parameter; auto input = CreateParameterAndTransferLiteral(0, *input_literal, "input", &builder, ¶meter); - builder.Reshape(/*operand=*/parameter, /*dimensions=*/{1, 2, 0}, - /*new_sizes=*/{24}); - auto expected_literal = Literal::CreateR1( + Reshape(/*operand=*/parameter, /*dimensions=*/{1, 2, 0}, + /*new_sizes=*/{24}); + auto expected_literal = LiteralUtil::CreateR1( {10, 20, 30, 40, 11, 21, 31, 41, 12, 22, 32, 42, 15, 25, 35, 45, 16, 26, 36, 46, 17, 27, 37, 47}); ComputeAndCompareLiteral(&builder, *expected_literal, {input.get()}, @@ -470,33 +439,33 @@ XLA_TEST_P(ReshapeTest, DocR3_R1_Collapse_120) { XLA_TEST_P(ReshapeTest, DocR3_R2_Collapse_120_Refine_83) { XlaBuilder builder(TestName()); - auto input_literal = Literal::CreateFromArray(ArrayForDocR3Tests()); + auto input_literal = LiteralUtil::CreateFromArray(ArrayForDocR3Tests()); XlaOp parameter; auto input = CreateParameterAndTransferLiteral(0, *input_literal, "input", &builder, ¶meter); - builder.Reshape(/*operand=*/parameter, /*dimensions=*/{1, 2, 0}, - /*new_sizes=*/{8, 3}); - auto expected_literal = Literal::CreateR2({{10, 20, 30}, - {40, 11, 21}, - {31, 41, 12}, - {22, 32, 42}, - {15, 25, 35}, - {45, 16, 26}, - {36, 46, 17}, - {27, 37, 47}}); + Reshape(/*operand=*/parameter, /*dimensions=*/{1, 2, 0}, + /*new_sizes=*/{8, 3}); + auto expected_literal = LiteralUtil::CreateR2({{10, 20, 30}, + {40, 11, 21}, + {31, 41, 12}, + {22, 32, 42}, + {15, 25, 35}, + {45, 16, 26}, + {36, 46, 17}, + {27, 37, 47}}); ComputeAndCompareLiteral(&builder, *expected_literal, {input.get()}, zero_error_spec_); } XLA_TEST_P(ReshapeTest, DocR3_R3_Collapse_120_Refine_262) { XlaBuilder builder(TestName()); - auto input_literal = Literal::CreateFromArray(ArrayForDocR3Tests()); + auto input_literal = LiteralUtil::CreateFromArray(ArrayForDocR3Tests()); XlaOp parameter; auto input = CreateParameterAndTransferLiteral(0, *input_literal, "input", &builder, ¶meter); - builder.Reshape(/*operand=*/parameter, /*dimensions=*/{1, 2, 0}, - /*new_sizes=*/{2, 6, 2}); - auto expected_literal = Literal::CreateR3( + Reshape(/*operand=*/parameter, /*dimensions=*/{1, 2, 0}, + /*new_sizes=*/{2, 6, 2}); + auto expected_literal = LiteralUtil::CreateR3( {{{10, 20}, {30, 40}, {11, 21}, {31, 41}, {12, 22}, {32, 42}}, {{15, 25}, {35, 45}, {16, 26}, {36, 46}, {17, 27}, {37, 47}}}); ComputeAndCompareLiteral(&builder, *expected_literal, {input.get()}, @@ -523,12 +492,12 @@ XLA_TEST_P(ReshapeTest, FullyConnectedCollapse) { Array4D t2x2x2x3(2, 2, 2, 3); auto filler2x3 = MakeLinspaceArray2D(1.0f, 6.0f, 2, 3); t2x2x2x3.FillWithYX(*filler2x3); - auto input_literal = Literal::CreateFromArray(t2x2x2x3); + auto input_literal = LiteralUtil::CreateFromArray(t2x2x2x3); XlaOp parameter; auto input = CreateParameterAndTransferLiteral(0, *input_literal, "input", &builder, ¶meter); - builder.Collapse(/*operand=*/parameter, /*dimensions=*/{1, 2, 3}); - auto expected_literal = Literal::CreateR2( + Collapse(/*operand=*/parameter, /*dimensions=*/{1, 2, 3}); + auto expected_literal = LiteralUtil::CreateR2( {{1.0f, 2.0f, 3.0f, 4.0f, 5.0f, 6.0f, 1.0f, 2.0f, 3.0f, 4.0f, 5.0f, 6.0f}, {1.0f, 2.0f, 3.0f, 4.0f, 5.0f, 6.0f, 1.0f, 2.0f, 3.0f, 4.0f, 5.0f, 6.0f}}); @@ -548,15 +517,15 @@ XLA_TEST_P(ReshapeTest, FullyConnectedCollapseDesugared) { t(1, 0, 0, 1) = 5; t(1, 0, 1, 0) = 6; t(1, 0, 1, 1) = 7; - auto input_literal = Literal::CreateFromArray(t); + auto input_literal = LiteralUtil::CreateFromArray(t); XlaOp parameter; auto input = CreateParameterAndTransferLiteral(0, *input_literal, "input", &builder, ¶meter); - builder.Reshape(/*operand=*/parameter, /*dimensions=*/{0, 1, 2, 3}, - /*new_sizes=*/{2, 4}); + Reshape(/*operand=*/parameter, /*dimensions=*/{0, 1, 2, 3}, + /*new_sizes=*/{2, 4}); auto expected_literal = - Literal::CreateR2({{0, 1, 2, 3}, {4, 5, 6, 7}}); + LiteralUtil::CreateR2({{0, 1, 2, 3}, {4, 5, 6, 7}}); ComputeAndCompareLiteral(&builder, *expected_literal, {input.get()}, zero_error_spec_); } @@ -575,9 +544,9 @@ XLA_TEST_P(ReshapeTest, ToScalar) { XlaOp parameter; auto input = CreateParameterAndTransferLiteral(0, input_literal, "input", &b, ¶meter); - b.Reshape(parameter, dimensions, {}); + Reshape(parameter, dimensions, {}); - auto expected_literal = Literal::CreateR0(83.0f); + auto expected_literal = LiteralUtil::CreateR0(83.0f); ComputeAndCompareLiteral(&b, *expected_literal, {input.get()}, zero_error_spec_); } @@ -585,11 +554,11 @@ XLA_TEST_P(ReshapeTest, ToScalar) { XLA_TEST_P(ReshapeTest, BadDimensions) { XlaBuilder b(TestName()); - auto input_literal = Literal::CreateR1({1.0f}); + auto input_literal = LiteralUtil::CreateR1({1.0f}); XlaOp parameter; auto input = CreateParameterAndTransferLiteral(0, *input_literal, "input", &b, ¶meter); - b.Reshape(parameter, {}, {}); + Reshape(parameter, {}, {}); EXPECT_THAT( ExecuteToString(&b, {}), ::testing::HasSubstr("not a permutation of the operand dimensions")); @@ -597,11 +566,11 @@ XLA_TEST_P(ReshapeTest, BadDimensions) { XLA_TEST_P(ReshapeTest, BadNewSizes) { XlaBuilder b(TestName()); - auto input_literal = Literal::CreateR1({1.0f, 2.0f}); + auto input_literal = LiteralUtil::CreateR1({1.0f, 2.0f}); XlaOp parameter; auto input = CreateParameterAndTransferLiteral(0, *input_literal, "input", &b, ¶meter); - b.Reshape(parameter, {1}, {}); + Reshape(parameter, {1}, {}); EXPECT_THAT(ExecuteToString(&b, {}), ::testing::HasSubstr("mismatched element counts")); } @@ -609,7 +578,8 @@ XLA_TEST_P(ReshapeTest, BadNewSizes) { XLA_TEST_P(ReshapeTest, R4Dim0MinorLayoutToR2Dim0MajorLayout) { XlaBuilder builder(TestName()); // clang-format off - auto input_literal = Literal::CreateR4FromArray4DWithLayout(Array4D{ + auto input_literal = LiteralUtil::CreateR4FromArray4DWithLayout( + Array4D{ { { {0, 1}, @@ -637,7 +607,7 @@ XLA_TEST_P(ReshapeTest, R4Dim0MinorLayoutToR2Dim0MajorLayout) { auto input = CreateParameterAndTransferLiteral(0, *input_literal, "input", &builder, ¶meter); - builder.Reshape(parameter, /*dimensions=*/{0, 1, 2, 3}, /*new_sizes=*/{2, 8}); + Reshape(parameter, /*dimensions=*/{0, 1, 2, 3}, /*new_sizes=*/{2, 8}); Array2D expected_array({ {0, 1, 2, 3, 100, 101, 102, 103}, @@ -654,16 +624,16 @@ XLA_TEST_P(ReshapeTest, R4Dim0MinorLayoutToR2Dim0MajorLayout) { ->ExecuteAndTransfer(computation, {input.get()}, &execution_options) .ConsumeValueOrDie(); std::unique_ptr expected = - Literal::CreateR2FromArray2D(expected_array); + LiteralUtil::CreateR2FromArray2D(expected_array); if (use_bfloat16()) { - expected = Literal::ConvertF32ToBF16(*expected); + expected = LiteralUtil::ConvertF32ToBF16(*expected); } EXPECT_TRUE(LiteralTestUtil::Equal(*expected, *actual)); } XLA_TEST_P(ReshapeTest, R2ToR4_3x8_To_3x2x1x4) { XlaBuilder builder(TestName()); - std::unique_ptr input_literal = Literal::CreateR2({ + std::unique_ptr input_literal = LiteralUtil::CreateR2({ {0, 1, 2, 3, 4, 5, 6, 7}, {100, 101, 102, 103, 104, 105, 106, 107}, {200, 201, 202, 203, 204, 205, 206, 207}, @@ -671,10 +641,10 @@ XLA_TEST_P(ReshapeTest, R2ToR4_3x8_To_3x2x1x4) { XlaOp parameter; auto input = CreateParameterAndTransferLiteral(0, *input_literal, "input", &builder, ¶meter); - builder.Reshape(parameter, /*dimensions=*/{0, 1}, /*new_sizes=*/{3, 2, 1, 4}); + Reshape(parameter, /*dimensions=*/{0, 1}, /*new_sizes=*/{3, 2, 1, 4}); // clang-format off - auto expected_literal = Literal::CreateR4({ + auto expected_literal = LiteralUtil::CreateR4({ {{{0, 1, 2, 3}}, {{4, 5, 6, 7}}}, {{{100, 101, 102, 103}}, @@ -690,7 +660,7 @@ XLA_TEST_P(ReshapeTest, R2ToR4_3x8_To_3x2x1x4) { // Tests R2->R4 reshape with the reshape dimensions {1, 0}. XLA_TEST_P(ReshapeTest, R2ToR4_3x8_To_3x2x1x4_Dimensions_10) { XlaBuilder builder(TestName()); - std::unique_ptr input_literal = Literal::CreateR2({ + std::unique_ptr input_literal = LiteralUtil::CreateR2({ {0, 1, 2, 3, 4, 5, 6, 7}, {100, 101, 102, 103, 104, 105, 106, 107}, {200, 201, 202, 203, 204, 205, 206, 207}, @@ -698,10 +668,10 @@ XLA_TEST_P(ReshapeTest, R2ToR4_3x8_To_3x2x1x4_Dimensions_10) { XlaOp parameter; auto input = CreateParameterAndTransferLiteral(0, *input_literal, "input", &builder, ¶meter); - builder.Reshape(parameter, /*dimensions=*/{1, 0}, /*new_sizes=*/{3, 2, 1, 4}); + Reshape(parameter, /*dimensions=*/{1, 0}, /*new_sizes=*/{3, 2, 1, 4}); // clang-format off - auto expected_literal = Literal::CreateR4({ + auto expected_literal = LiteralUtil::CreateR4({ {{{0, 100, 200, 1}}, {{101, 201, 2, 102}}}, {{{202, 3, 103, 203}}, @@ -723,15 +693,15 @@ XLA_TEST_P(ReshapeTest, R4ToR2_2x1x1x1_To_2x1) { [&rng, &distribution](tensorflow::gtl::ArraySlice /* indices */, float* cell) { *cell = distribution(rng); }); std::unique_ptr input_literal = - Literal::CreateR4FromArray4DWithLayout( + LiteralUtil::CreateR4FromArray4DWithLayout( input, LayoutUtil::MakeLayout({3, 2, 1, 0})); XlaOp parameter; auto input_data = CreateParameterAndTransferLiteral( 0, *input_literal, "input", &builder, ¶meter); - builder.Reshape(parameter, /*dimensions=*/{0, 1, 2, 3}, /*new_sizes=*/{2, 1}); + Reshape(parameter, /*dimensions=*/{0, 1, 2, 3}, /*new_sizes=*/{2, 1}); std::unique_ptr expected = - Literal::ReshapeSlice({2, 1}, {1, 0}, *input_literal); + LiteralUtil::ReshapeSlice({2, 1}, {1, 0}, *input_literal); ComputeAndCompareLiteral(&builder, *expected, {input_data.get()}, zero_error_spec_); } @@ -745,15 +715,15 @@ XLA_TEST_P(ReshapeTest, R4ToR2_2x1x4x1_To_4x2) { [&rng, &distribution](tensorflow::gtl::ArraySlice /* indices */, float* cell) { *cell = distribution(rng); }); std::unique_ptr input_literal = - Literal::CreateR4FromArray4DWithLayout( + LiteralUtil::CreateR4FromArray4DWithLayout( input, LayoutUtil::MakeLayout({3, 2, 1, 0})); XlaOp parameter; auto input_data = CreateParameterAndTransferLiteral( 0, *input_literal, "input", &builder, ¶meter); - builder.Reshape(parameter, /*dimensions=*/{0, 1, 2, 3}, /*new_sizes=*/{4, 2}); + Reshape(parameter, /*dimensions=*/{0, 1, 2, 3}, /*new_sizes=*/{4, 2}); std::unique_ptr expected = - Literal::ReshapeSlice({4, 2}, {1, 0}, *input_literal); + LiteralUtil::ReshapeSlice({4, 2}, {1, 0}, *input_literal); ComputeAndCompareLiteral(&builder, *expected, {input_data.get()}, zero_error_spec_); } @@ -768,20 +738,20 @@ XLA_TEST_P(ReshapeTest, R4ToR2_5x10x2x3_To_5x60_Dimensions_0213) { [&rng, &distribution](tensorflow::gtl::ArraySlice /* indices */, float* cell) { *cell = distribution(rng); }); std::unique_ptr input_literal = - Literal::CreateR4FromArray4DWithLayout( + LiteralUtil::CreateR4FromArray4DWithLayout( input, LayoutUtil::MakeLayout({3, 2, 1, 0})); XlaOp parameter; auto input_data = CreateParameterAndTransferLiteral( 0, *input_literal, "input", &builder, ¶meter); - builder.Reshape(parameter, /*dimensions=*/{0, 2, 1, 3}, - /*new_sizes=*/{5, 60}); + Reshape(parameter, /*dimensions=*/{0, 2, 1, 3}, + /*new_sizes=*/{5, 60}); Array2D expected_array(5, 60); input.Each([&](tensorflow::gtl::ArraySlice indices, float* cell) { expected_array(indices[0], indices[2] * 30 + indices[1] * 3 + indices[3]) = *cell; }); - auto expected = Literal::CreateR2FromArray2D(expected_array); + auto expected = LiteralUtil::CreateR2FromArray2D(expected_array); ComputeAndCompareLiteral(&builder, *expected, {input_data.get()}, zero_error_spec_); } @@ -795,13 +765,13 @@ XLA_TEST_P(ReshapeTest, NoopReshape) { [&rng, &distribution](tensorflow::gtl::ArraySlice /* indices */, float* cell) { *cell = distribution(rng); }); std::unique_ptr input_literal = - Literal::CreateR4FromArray4DWithLayout( + LiteralUtil::CreateR4FromArray4DWithLayout( input_array, LayoutUtil::MakeLayout({1, 2, 3, 0})); XlaOp parameter; auto input_data = CreateParameterAndTransferLiteral( 0, *input_literal, "input", &builder, ¶meter); - builder.Reshape(parameter, /*dimensions=*/{3, 0, 1, 2}, - /*new_sizes=*/{7, 2, 3, 5}); + Reshape(parameter, /*dimensions=*/{3, 0, 1, 2}, + /*new_sizes=*/{7, 2, 3, 5}); XlaComputation computation = builder.Build().ConsumeValueOrDie(); ExecutionOptions execution_options = execution_options_; @@ -817,7 +787,7 @@ XLA_TEST_P(ReshapeTest, NoopReshape) { // Since the reshape is a no-op, verify that it does not change the underlying // data. if (use_bfloat16()) { - auto expected = Literal::ConvertF32ToBF16(*input_literal); + auto expected = LiteralUtil::ConvertF32ToBF16(*input_literal); EXPECT_EQ(expected->data(), output_literal->data()); } else { EXPECT_EQ(input_literal->data(), output_literal->data()); @@ -826,21 +796,21 @@ XLA_TEST_P(ReshapeTest, NoopReshape) { XLA_TEST_P(ReshapeTest, R4ToR4Reshape_Trivial) { XlaBuilder builder(TestName()); - auto literal_1x2x3x4 = Literal::CreateR4( + auto literal_1x2x3x4 = LiteralUtil::CreateR4( {{{{1, 2, 3, 4}, {5, 6, 7, 8}, {9, 10, 11, 12}}, {{13, 14, 15, 16}, {17, 18, 19, 20}, {21, 22, 23, 24}}}}); XlaOp parameter; auto input = CreateParameterAndTransferLiteral(0, *literal_1x2x3x4, "input", &builder, ¶meter); - builder.Reshape(parameter, /*dimensions=*/{0, 1, 2, 3}, - /*new_sizes=*/{1, 2, 3, 4}); + Reshape(parameter, /*dimensions=*/{0, 1, 2, 3}, + /*new_sizes=*/{1, 2, 3, 4}); ComputeAndCompareLiteral(&builder, *literal_1x2x3x4, {input.get()}); } XLA_TEST_P(ReshapeTest, R4ToR4Reshape) { - auto literal_1x2x3x4 = Literal::CreateR4( + auto literal_1x2x3x4 = LiteralUtil::CreateR4( {{{{1, 2, 3, 4}, {5, 6, 7, 8}, {9, 10, 11, 12}}, {{13, 14, 15, 16}, {17, 18, 19, 20}, {21, 22, 23, 24}}}}); @@ -848,11 +818,11 @@ XLA_TEST_P(ReshapeTest, R4ToR4Reshape) { XlaOp parameter; auto input = CreateParameterAndTransferLiteral(0, *literal_1x2x3x4, "input", &builder, ¶meter); - builder.Reshape(parameter, /*dimensions=*/{1, 3, 2, 0}, - /*new_sizes=*/{2, 4, 3, 1}); + Reshape(parameter, /*dimensions=*/{1, 3, 2, 0}, + /*new_sizes=*/{2, 4, 3, 1}); // clang-format off - auto expected_2x4x3x1 = Literal::CreateR4( + auto expected_2x4x3x1 = LiteralUtil::CreateR4( {{{{1}, {5}, {9}}, {{2}, {6}, {10}}, {{3}, {7}, {11}}, @@ -876,17 +846,17 @@ XLA_TEST_P(ReshapeTest, R4TwoMinorTransposeSimple) { [&rng, &distribution](tensorflow::gtl::ArraySlice /* indices */, float* cell) { *cell = distribution(rng); }); std::unique_ptr input_literal = - Literal::CreateR4FromArray4DWithLayout( + LiteralUtil::CreateR4FromArray4DWithLayout( input, LayoutUtil::MakeLayout({3, 2, 1, 0})); XlaBuilder builder(TestName()); XlaOp parameter; auto input_data = CreateParameterAndTransferLiteral( 0, *input_literal, "input", &builder, ¶meter); - builder.Reshape(parameter, /*dimensions=*/{0, 1, 3, 2}, - /*new_sizes=*/new_bounds); + Reshape(parameter, /*dimensions=*/{0, 1, 3, 2}, + /*new_sizes=*/new_bounds); std::unique_ptr expected = - Literal::ReshapeSlice(new_bounds, {2, 3, 1, 0}, *input_literal) + LiteralUtil::ReshapeSlice(new_bounds, {2, 3, 1, 0}, *input_literal) ->Relayout(LayoutUtil::MakeLayout({3, 2, 1, 0})); // Specify the requested output shape explicitly to ensure that this reshape @@ -905,17 +875,17 @@ XLA_TEST_P(ReshapeTest, R4TwoMinorTransposeMajorFirstEffectiveR2) { [&rng, &distribution](tensorflow::gtl::ArraySlice /* indices */, float* cell) { *cell = distribution(rng); }); std::unique_ptr input_literal = - Literal::CreateR4FromArray4DWithLayout( + LiteralUtil::CreateR4FromArray4DWithLayout( input, LayoutUtil::MakeLayout({3, 2, 1, 0})); XlaBuilder builder(TestName()); XlaOp parameter; auto input_data = CreateParameterAndTransferLiteral( 0, *input_literal, "input", &builder, ¶meter); - builder.Reshape(parameter, /*dimensions=*/{0, 1, 3, 2}, - /*new_sizes=*/new_bounds); + Reshape(parameter, /*dimensions=*/{0, 1, 3, 2}, + /*new_sizes=*/new_bounds); std::unique_ptr expected = - Literal::ReshapeSlice(new_bounds, {2, 3, 1, 0}, *input_literal) + LiteralUtil::ReshapeSlice(new_bounds, {2, 3, 1, 0}, *input_literal) ->Relayout(LayoutUtil::MakeLayout({3, 2, 1, 0})); // Specify the requested output shape explicitly to ensure that this reshape @@ -934,17 +904,17 @@ XLA_TEST_P(ReshapeTest, R4TwoMinorTransposeMajorFirstMinorEffectiveR1) { [&rng, &distribution](tensorflow::gtl::ArraySlice /* indices */, float* cell) { *cell = distribution(rng); }); std::unique_ptr input_literal = - Literal::CreateR4FromArray4DWithLayout( + LiteralUtil::CreateR4FromArray4DWithLayout( input, LayoutUtil::MakeLayout({3, 2, 1, 0})); XlaBuilder builder(TestName()); XlaOp parameter; auto input_data = CreateParameterAndTransferLiteral( 0, *input_literal, "input", &builder, ¶meter); - builder.Reshape(parameter, /*dimensions=*/{0, 1, 3, 2}, - /*new_sizes=*/new_bounds); + Reshape(parameter, /*dimensions=*/{0, 1, 3, 2}, + /*new_sizes=*/new_bounds); std::unique_ptr expected = - Literal::ReshapeSlice(new_bounds, {2, 3, 1, 0}, *input_literal) + LiteralUtil::ReshapeSlice(new_bounds, {2, 3, 1, 0}, *input_literal) ->Relayout(LayoutUtil::MakeLayout({3, 2, 1, 0})); // Specify the requested output shape explicitly to ensure that this reshape @@ -964,17 +934,17 @@ XLA_TEST_P(ReshapeTest, R4TwoMinorTransposeMajorFirstMinorEffectiveR1InR2) { [&rng, &distribution](tensorflow::gtl::ArraySlice /* indices */, float* cell) { *cell = distribution(rng); }); std::unique_ptr input_literal = - Literal::CreateR4FromArray4DWithLayout( + LiteralUtil::CreateR4FromArray4DWithLayout( input, LayoutUtil::MakeLayout({3, 2, 1, 0})); XlaBuilder builder(TestName()); XlaOp parameter; auto input_data = CreateParameterAndTransferLiteral( 0, *input_literal, "input", &builder, ¶meter); - builder.Reshape(parameter, /*dimensions=*/{0, 1, 3, 2}, - /*new_sizes=*/new_bounds); + Reshape(parameter, /*dimensions=*/{0, 1, 3, 2}, + /*new_sizes=*/new_bounds); std::unique_ptr expected = - Literal::ReshapeSlice(new_bounds, {2, 3, 1, 0}, *input_literal) + LiteralUtil::ReshapeSlice(new_bounds, {2, 3, 1, 0}, *input_literal) ->Relayout(LayoutUtil::MakeLayout({3, 2, 1, 0})); // Specify the requested output shape explicitly to ensure that this reshape @@ -993,17 +963,17 @@ XLA_TEST_P(ReshapeTest, R4TwoMinorTransposeTrivialR2) { [&rng, &distribution](tensorflow::gtl::ArraySlice /* indices */, float* cell) { *cell = distribution(rng); }); std::unique_ptr input_literal = - Literal::CreateR4FromArray4DWithLayout( + LiteralUtil::CreateR4FromArray4DWithLayout( input, LayoutUtil::MakeLayout({0, 1, 2, 3})); XlaBuilder builder(TestName()); XlaOp parameter; auto input_data = CreateParameterAndTransferLiteral( 0, *input_literal, "input", &builder, ¶meter); - builder.Reshape(parameter, /*dimensions=*/{1, 0, 2, 3}, - /*new_sizes=*/new_bounds); + Reshape(parameter, /*dimensions=*/{1, 0, 2, 3}, + /*new_sizes=*/new_bounds); std::unique_ptr expected = - Literal::ReshapeSlice(new_bounds, {1, 0, 2, 3}, *input_literal) + LiteralUtil::ReshapeSlice(new_bounds, {1, 0, 2, 3}, *input_literal) ->Relayout(input_literal->shape().layout()); // Specify the requested output shape explicitly to ensure that this reshape diff --git a/tensorflow/compiler/xla/tests/reverse_test.cc b/tensorflow/compiler/xla/tests/reverse_test.cc index e7bd142dc9ddefbd8bebfb77d72218d662645c31..23f0d26d93bf979970d112993c0a945fb4fe7d53 100644 --- a/tensorflow/compiler/xla/tests/reverse_test.cc +++ b/tensorflow/compiler/xla/tests/reverse_test.cc @@ -82,12 +82,12 @@ TEST_P(FloatReverseTest, Reverses) { std::vector input_vector( ShapeUtil::ElementsIn(ShapeUtil::MakeShape(F32, spec.input_dims))); std::iota(input_vector.begin(), input_vector.end(), 0.0); - auto r1_literal = Literal::CreateR1(input_vector); + auto r1_literal = LiteralUtil::CreateR1(input_vector); auto input_literal = r1_literal->Reshape(spec.input_dims).ConsumeValueOrDie(); XlaBuilder builder(TestName()); auto a = AddParam(*input_literal, &builder); - builder.Rev(a, spec.reversal); + Rev(a, spec.reversal); std::unique_ptr expected = input_literal->CloneToUnique(); std::vector output_indices(spec.input_dims.size()); @@ -127,7 +127,7 @@ XLA_TEST_F(ReverseTest, Reverse4DU8ArrayOnDim23) { }}); // clang-format on - b.Rev(b.ConstantR4FromArray4D(input), {0, 3}); + Rev(ConstantR4FromArray4D(&b, input), {0, 3}); // clang-format off Array4D expected({{ @@ -163,7 +163,7 @@ TEST_F(ReverseTest, Reverse4DFloatArrayOnDim01) { }); // clang-format on - b.Rev(b.ConstantR4FromArray4D(input), {0, 1}); + Rev(ConstantR4FromArray4D(&b, input), {0, 1}); // clang-format off Array4D expected({ diff --git a/tensorflow/compiler/xla/tests/round_trip_packed_literal_test.cc b/tensorflow/compiler/xla/tests/round_trip_packed_literal_test.cc index 7cfca781acda15879075f4386c2096e537877aac..a620fe19085d98c8b6642b25b159d6c2308bdae2 100644 --- a/tensorflow/compiler/xla/tests/round_trip_packed_literal_test.cc +++ b/tensorflow/compiler/xla/tests/round_trip_packed_literal_test.cc @@ -18,7 +18,7 @@ limitations under the License. #include "tensorflow/compiler/xla/client/global_data.h" #include "tensorflow/compiler/xla/client/local_client.h" #include "tensorflow/compiler/xla/layout_util.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/packed_literal_reader.h" #include "tensorflow/compiler/xla/shape_util.h" #include "tensorflow/compiler/xla/statusor.h" diff --git a/tensorflow/compiler/xla/tests/round_trip_transfer_test.cc b/tensorflow/compiler/xla/tests/round_trip_transfer_test.cc index f334a8c1318a59bbfdd27dd1a63ed162600089ce..a8193c2eac05ba4f0df339909f3e82a28ac35253 100644 --- a/tensorflow/compiler/xla/tests/round_trip_transfer_test.cc +++ b/tensorflow/compiler/xla/tests/round_trip_transfer_test.cc @@ -23,7 +23,7 @@ limitations under the License. #include "tensorflow/compiler/xla/array4d.h" #include "tensorflow/compiler/xla/client/global_data.h" #include "tensorflow/compiler/xla/client/local_client.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/statusor.h" #include "tensorflow/compiler/xla/tests/client_library_test_base.h" #include "tensorflow/compiler/xla/tests/literal_test_util.h" @@ -46,61 +46,62 @@ class RoundTripTransferTest : public ClientLibraryTestBase { }; TEST_F(RoundTripTransferTest, R0S32) { - RoundTripTest(*Literal::CreateR0(42)); + RoundTripTest(*LiteralUtil::CreateR0(42)); } TEST_F(RoundTripTransferTest, R0F32) { - RoundTripTest(*Literal::CreateR0(42.0)); + RoundTripTest(*LiteralUtil::CreateR0(42.0)); } TEST_F(RoundTripTransferTest, R1F32_Len0) { - RoundTripTest(*Literal::CreateR1({})); + RoundTripTest(*LiteralUtil::CreateR1({})); } TEST_F(RoundTripTransferTest, R1F32_Len2) { - RoundTripTest(*Literal::CreateR1({42.0, 64.0})); + RoundTripTest(*LiteralUtil::CreateR1({42.0, 64.0})); } TEST_F(RoundTripTransferTest, R1F32_Len256) { std::vector values(256); std::iota(values.begin(), values.end(), 1.0); - RoundTripTest(*Literal::CreateR1(values)); + RoundTripTest(*LiteralUtil::CreateR1(values)); } TEST_F(RoundTripTransferTest, R1F32_Len1024) { std::vector values(1024); std::iota(values.begin(), values.end(), 1.0); - RoundTripTest(*Literal::CreateR1(values)); + RoundTripTest(*LiteralUtil::CreateR1(values)); } TEST_F(RoundTripTransferTest, R1F32_Len1025) { std::vector values(1025); std::iota(values.begin(), values.end(), 1.0); - RoundTripTest(*Literal::CreateR1(values)); + RoundTripTest(*LiteralUtil::CreateR1(values)); } TEST_F(RoundTripTransferTest, R1F32_Len4096) { std::vector values(4096); std::iota(values.begin(), values.end(), 1.0); - RoundTripTest(*Literal::CreateR1(values)); + RoundTripTest(*LiteralUtil::CreateR1(values)); } TEST_F(RoundTripTransferTest, R2F32_Len10x0) { - RoundTripTest(*Literal::CreateR2FromArray2D(Array2D(10, 0))); + RoundTripTest( + *LiteralUtil::CreateR2FromArray2D(Array2D(10, 0))); } TEST_F(RoundTripTransferTest, R2F32_Len2x2) { - RoundTripTest(*Literal::CreateR2({{42.0, 64.0}, {77.0, 88.0}})); + RoundTripTest(*LiteralUtil::CreateR2({{42.0, 64.0}, {77.0, 88.0}})); } TEST_F(RoundTripTransferTest, R3F32) { RoundTripTest( - *Literal::CreateR3({{{1.0, 2.0}, {1.0, 2.0}, {1.0, 2.0}}, - {{3.0, 4.0}, {3.0, 4.0}, {3.0, 4.0}}})); + *LiteralUtil::CreateR3({{{1.0, 2.0}, {1.0, 2.0}, {1.0, 2.0}}, + {{3.0, 4.0}, {3.0, 4.0}, {3.0, 4.0}}})); } TEST_F(RoundTripTransferTest, R4F32) { - RoundTripTest(*Literal::CreateR4({{ + RoundTripTest(*LiteralUtil::CreateR4({{ {{10, 11, 12, 13}, {14, 15, 16, 17}}, {{18, 19, 20, 21}, {22, 23, 24, 25}}, {{26, 27, 28, 29}, {30, 31, 32, 33}}, @@ -108,33 +109,36 @@ TEST_F(RoundTripTransferTest, R4F32) { } TEST_F(RoundTripTransferTest, EmptyTuple) { - RoundTripTest(*Literal::MakeTuple({})); + RoundTripTest(*LiteralUtil::MakeTuple({})); } TEST_F(RoundTripTransferTest, TupleOfR1F32) { - RoundTripTest(*Literal::MakeTuple({Literal::CreateR1({1, 2}).get(), - Literal::CreateR1({3, 4}).get()})); + RoundTripTest( + *LiteralUtil::MakeTuple({LiteralUtil::CreateR1({1, 2}).get(), + LiteralUtil::CreateR1({3, 4}).get()})); } TEST_F(RoundTripTransferTest, TupleOfR1F32_Len0_Len2) { - RoundTripTest(*Literal::MakeTuple({Literal::CreateR1({}).get(), - Literal::CreateR1({3, 4}).get()})); + RoundTripTest( + *LiteralUtil::MakeTuple({LiteralUtil::CreateR1({}).get(), + LiteralUtil::CreateR1({3, 4}).get()})); } TEST_F(RoundTripTransferTest, TupleOfR0F32AndR1S32) { - RoundTripTest(*Literal::MakeTuple({Literal::CreateR0(1.0).get(), - Literal::CreateR1({2, 3}).get()})); + RoundTripTest( + *LiteralUtil::MakeTuple({LiteralUtil::CreateR0(1.0).get(), + LiteralUtil::CreateR1({2, 3}).get()})); } // Below two tests are added to identify the cost of large data transfers. TEST_F(RoundTripTransferTest, R2F32_Large) { - RoundTripTest(*Literal::CreateR2F32Linspace(-1.0f, 1.0f, 512, 512)); + RoundTripTest(*LiteralUtil::CreateR2F32Linspace(-1.0f, 1.0f, 512, 512)); } TEST_F(RoundTripTransferTest, R4F32_Large) { Array4D array4d(2, 2, 256, 256); array4d.FillWithMultiples(1.0f); - RoundTripTest(*Literal::CreateR4FromArray4D(array4d)); + RoundTripTest(*LiteralUtil::CreateR4FromArray4D(array4d)); } } // namespace diff --git a/tensorflow/compiler/xla/tests/scalar_computations_test.cc b/tensorflow/compiler/xla/tests/scalar_computations_test.cc index 308d3fc78a51e63c0e3db8c0cda18caf11f665bd..3b603c0d31565c20ff4e43c3ffd6001e9d54612e 100644 --- a/tensorflow/compiler/xla/tests/scalar_computations_test.cc +++ b/tensorflow/compiler/xla/tests/scalar_computations_test.cc @@ -21,6 +21,7 @@ limitations under the License. #include "tensorflow/compiler/xla/client/local_client.h" #include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/compiler/xla/client/xla_client/xla_computation.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/literal_util.h" #include "tensorflow/compiler/xla/status_macros.h" #include "tensorflow/compiler/xla/statusor.h" @@ -44,74 +45,75 @@ class ScalarComputationsTest : public ClientLibraryTestBase { protected: // A template for building and running a binary comparison test. template - void TestCompare( - NativeT lhs, NativeT rhs, bool expected, - XlaOp (XlaBuilder::*op)(const XlaOp&, const XlaOp&, - tensorflow::gtl::ArraySlice)) { + void TestCompare(NativeT lhs, NativeT rhs, bool expected, + std::function)> + op) { XlaBuilder builder(TestName()); - XlaOp lhs_op = builder.ConstantR0(lhs); - XlaOp rhs_op = builder.ConstantR0(rhs); - XlaOp result = (builder.*op)(lhs_op, rhs_op, {}); + XlaOp lhs_op = ConstantR0(&builder, lhs); + XlaOp rhs_op = ConstantR0(&builder, rhs); + op(lhs_op, rhs_op, {}); ComputeAndCompareR0(&builder, expected, {}); } template void TestMinMax(NativeT lhs, NativeT rhs, NativeT expected, - XlaOp (XlaBuilder::*op)(const XlaOp&, const XlaOp&, - tensorflow::gtl::ArraySlice)) { + std::function)> + op) { XlaBuilder builder(TestName()); - XlaOp lhs_op = builder.ConstantR0(lhs); - XlaOp rhs_op = builder.ConstantR0(rhs); - XlaOp result = (builder.*op)(lhs_op, rhs_op, {}); + XlaOp lhs_op = ConstantR0(&builder, lhs); + XlaOp rhs_op = ConstantR0(&builder, rhs); + op(lhs_op, rhs_op, {}); ComputeAndCompareR0(&builder, expected, {}); } }; XLA_TEST_F(ScalarComputationsTest, ReturnScalarF32) { XlaBuilder builder(TestName()); - builder.ConstantR0(2.1f); + ConstantR0(&builder, 2.1f); ComputeAndCompareR0(&builder, 2.1f, {}, error_spec_); } XLA_TEST_F(ScalarComputationsTest, NegateScalarF32) { XlaBuilder builder(TestName()); - builder.Neg(builder.ConstantR0(2.1f)); + Neg(ConstantR0(&builder, 2.1f)); ComputeAndCompareR0(&builder, -2.1f, {}, error_spec_); } XLA_TEST_F(ScalarComputationsTest, NegateScalarS32) { XlaBuilder builder(TestName()); - builder.Neg(builder.ConstantR0(2)); + Neg(ConstantR0(&builder, 2)); ComputeAndCompareR0(&builder, -2, {}); } XLA_TEST_F(ScalarComputationsTest, AddTwoScalarsF32) { XlaBuilder builder(TestName()); - builder.Add(builder.ConstantR0(2.1f), builder.ConstantR0(5.5f)); + Add(ConstantR0(&builder, 2.1f), ConstantR0(&builder, 5.5f)); ComputeAndCompareR0(&builder, 7.6f, {}, error_spec_); } XLA_TEST_F(ScalarComputationsTest, AddTwoScalarsS32) { XlaBuilder builder(TestName()); - builder.Add(builder.ConstantR0(2), builder.ConstantR0(5)); + Add(ConstantR0(&builder, 2), ConstantR0(&builder, 5)); ComputeAndCompareR0(&builder, 7, {}); } XLA_TEST_F(ScalarComputationsTest, AddTwoScalarsU32) { XlaBuilder builder(TestName()); - builder.Add(builder.ConstantR0(35), builder.ConstantR0(57)); + Add(ConstantR0(&builder, 35), ConstantR0(&builder, 57)); ComputeAndCompareR0(&builder, 92, {}); } XLA_TEST_F(ScalarComputationsTest, AddTwoScalarsU8) { XlaBuilder builder(TestName()); - builder.Add(builder.ConstantR0(35), builder.ConstantR0(57)); + Add(ConstantR0(&builder, 35), ConstantR0(&builder, 57)); ComputeAndCompareR0(&builder, 92, {}); } @@ -120,7 +122,7 @@ XLA_TEST_F(ScalarComputationsTest, AddTwoScalarsU64) { XlaBuilder builder(TestName()); const uint64 a = static_cast(1) << 63; const uint64 b = a + 1; - builder.Add(builder.ConstantR0(a), builder.ConstantR0(b)); + Add(ConstantR0(&builder, a), ConstantR0(&builder, b)); ComputeAndCompareR0(&builder, a + b, {}); } @@ -129,40 +131,39 @@ XLA_TEST_F(ScalarComputationsTest, AddTwoScalarsS64) { XlaBuilder builder(TestName()); const int64 a = static_cast(1) << 62; const int64 b = a - 1; - builder.Add(builder.ConstantR0(a), builder.ConstantR0(b)); + Add(ConstantR0(&builder, a), ConstantR0(&builder, b)); ComputeAndCompareR0(&builder, a + b, {}); } XLA_TEST_F(ScalarComputationsTest, AddTwoScalarsF64) { XlaBuilder builder(TestName()); - builder.Add(builder.ConstantR0(0.25), - builder.ConstantR0(3.5)); + Add(ConstantR0(&builder, 0.25), ConstantR0(&builder, 3.5)); ComputeAndCompareR0(&builder, 3.75, {}); } XLA_TEST_F(ScalarComputationsTest, SubtractTwoScalarsF32) { XlaBuilder builder(TestName()); - builder.Sub(builder.ConstantR0(2.1f), builder.ConstantR0(5.5f)); + Sub(ConstantR0(&builder, 2.1f), ConstantR0(&builder, 5.5f)); ComputeAndCompareR0(&builder, -3.4f, {}, error_spec_); } XLA_TEST_F(ScalarComputationsTest, SubtractTwoScalarsS32) { XlaBuilder builder(TestName()); - builder.Sub(builder.ConstantR0(2), builder.ConstantR0(5)); + Sub(ConstantR0(&builder, 2), ConstantR0(&builder, 5)); ComputeAndCompareR0(&builder, -3, {}); } XLA_TEST_F(ScalarComputationsTest, CastS64ToF32) { XlaBuilder builder(TestName()); - auto a = builder.Parameter(0, ShapeUtil::MakeShape(S64, {}), "a"); - builder.ConvertElementType(a, F32); + auto a = Parameter(&builder, 0, ShapeUtil::MakeShape(S64, {}), "a"); + ConvertElementType(a, F32); int64 value = 3LL << 35; - std::unique_ptr a_literal = Literal::CreateR0(value); + std::unique_ptr a_literal = LiteralUtil::CreateR0(value); std::unique_ptr a_data = client_->TransferToServer(*a_literal).ConsumeValueOrDie(); ComputeAndCompareR0(&builder, static_cast(value), @@ -171,9 +172,8 @@ XLA_TEST_F(ScalarComputationsTest, CastS64ToF32) { XLA_TEST_F(ScalarComputationsTest, MulThreeScalarsF32) { XlaBuilder builder(TestName()); - builder.Mul(builder.Mul(builder.ConstantR0(2.1f), - builder.ConstantR0(5.5f)), - builder.ConstantR0(0.5f)); + Mul(Mul(ConstantR0(&builder, 2.1f), ConstantR0(&builder, 5.5f)), + ConstantR0(&builder, 0.5f)); ComputeAndCompareR0(&builder, 5.775f, {}, error_spec_); } @@ -190,7 +190,7 @@ XLA_TEST_F(ScalarComputationsTest, MulTwoScalarsS32) { for (int32 x : data) { for (int32 y : data) { XlaBuilder builder(TestName()); - builder.Mul(builder.ConstantR0(x), builder.ConstantR0(y)); + Mul(ConstantR0(&builder, x), ConstantR0(&builder, y)); // Signed integer overflow is undefined behavior in C++. Convert the input // integers to unsigned, perform the multiplication unsigned, and convert @@ -209,7 +209,7 @@ XLA_TEST_F(ScalarComputationsTest, MulTwoScalarsU32) { for (uint32 x : data) { for (uint32 y : data) { XlaBuilder builder(TestName()); - builder.Mul(builder.ConstantR0(x), builder.ConstantR0(y)); + Mul(ConstantR0(&builder, x), ConstantR0(&builder, y)); uint32 expected = x * y; ComputeAndCompareR0(&builder, expected, {}); @@ -219,18 +219,17 @@ XLA_TEST_F(ScalarComputationsTest, MulTwoScalarsU32) { XLA_TEST_F(ScalarComputationsTest, MulThreeScalarsS32) { XlaBuilder builder(TestName()); - builder.Mul( - builder.Mul(builder.ConstantR0(2), builder.ConstantR0(5)), - builder.ConstantR0(1)); + Mul(Mul(ConstantR0(&builder, 2), ConstantR0(&builder, 5)), + ConstantR0(&builder, 1)); ComputeAndCompareR0(&builder, 10, {}); } XLA_TEST_F(ScalarComputationsTest, MulThreeScalarsF32Params) { XlaBuilder builder(TestName()); - std::unique_ptr a_literal = Literal::CreateR0(2.1f); - std::unique_ptr b_literal = Literal::CreateR0(5.5f); - std::unique_ptr c_literal = Literal::CreateR0(0.5f); + std::unique_ptr a_literal = LiteralUtil::CreateR0(2.1f); + std::unique_ptr b_literal = LiteralUtil::CreateR0(5.5f); + std::unique_ptr c_literal = LiteralUtil::CreateR0(0.5f); std::unique_ptr a_data = client_->TransferToServer(*a_literal).ConsumeValueOrDie(); @@ -239,10 +238,10 @@ XLA_TEST_F(ScalarComputationsTest, MulThreeScalarsF32Params) { std::unique_ptr c_data = client_->TransferToServer(*c_literal).ConsumeValueOrDie(); - XlaOp a = builder.Parameter(0, a_literal->shape(), "a"); - XlaOp b = builder.Parameter(1, b_literal->shape(), "b"); - XlaOp c = builder.Parameter(2, c_literal->shape(), "c"); - builder.Mul(builder.Mul(a, b), c); + XlaOp a = Parameter(&builder, 0, a_literal->shape(), "a"); + XlaOp b = Parameter(&builder, 1, b_literal->shape(), "b"); + XlaOp c = Parameter(&builder, 2, c_literal->shape(), "c"); + Mul(Mul(a, b), c); ComputeAndCompareR0(&builder, 5.775f, {a_data.get(), b_data.get(), c_data.get()}, @@ -251,14 +250,14 @@ XLA_TEST_F(ScalarComputationsTest, MulThreeScalarsF32Params) { XLA_TEST_F(ScalarComputationsTest, DivideTwoScalarsF32) { XlaBuilder builder(TestName()); - builder.Div(builder.ConstantR0(5.0f), builder.ConstantR0(2.5f)); + Div(ConstantR0(&builder, 5.0f), ConstantR0(&builder, 2.5f)); ComputeAndCompareR0(&builder, 2.0f, {}, error_spec_); } XLA_TEST_F(ScalarComputationsTest, RemTwoScalarsF32) { XlaBuilder builder(TestName()); - builder.Rem(builder.ConstantR0(2.5f), builder.ConstantR0(5.0f)); + Rem(ConstantR0(&builder, 2.5f), ConstantR0(&builder, 5.0f)); ComputeAndCompareR0(&builder, 2.5f, {}, error_spec_); } @@ -281,8 +280,8 @@ class DivS32Test : public ClientLibraryTestBase, XLA_TEST_P(DivS32Test, DivideTwoScalarsS32) { DivS32Params p = GetParam(); XlaBuilder builder(TestName()); - builder.Div(builder.ConstantR0(p.dividend), - builder.ConstantR0(p.divisor)); + Div(ConstantR0(&builder, p.dividend), + ConstantR0(&builder, p.divisor)); ComputeAndCompareR0(&builder, p.quotient, {}); } @@ -290,8 +289,8 @@ XLA_TEST_P(DivS32Test, DivideTwoScalarsS32) { XLA_TEST_P(DivS32Test, RemainderTwoScalarsS32) { DivS32Params p = GetParam(); XlaBuilder builder(TestName()); - builder.Rem(builder.ConstantR0(p.dividend), - builder.ConstantR0(p.divisor)); + Rem(ConstantR0(&builder, p.dividend), + ConstantR0(&builder, p.divisor)); ComputeAndCompareR0(&builder, p.remainder, {}); } @@ -305,7 +304,7 @@ XLA_TEST_P(DivS32Test, DivideTwoScalarsNonConstS32) { CreateR0Parameter(p.dividend, 0, "dividend", &builder, ÷nd); auto divisord = CreateR0Parameter(p.divisor, 1, "divisor", &builder, &divisor); - builder.Div(dividend, divisor); + Div(dividend, divisor); ComputeAndCompareR0(&builder, p.quotient, {dividendd.get(), divisord.get()}); @@ -320,7 +319,7 @@ XLA_TEST_P(DivS32Test, RemainderTwoScalarsNonConstDivisorS32) { CreateR0Parameter(p.dividend, 0, "dividend", &builder, ÷nd); auto divisord = CreateR0Parameter(p.divisor, 1, "divisor", &builder, &divisor); - builder.Rem(dividend, divisor); + Rem(dividend, divisor); ComputeAndCompareR0(&builder, p.remainder, {dividendd.get(), divisord.get()}); @@ -367,18 +366,18 @@ XLA_TEST_F(ScalarComputationsTest, DivU32s) { XlaBuilder builder(TestName()); XlaOp dividend = - builder.Parameter(0, ShapeUtil::MakeShape(U32, {}), "dividend"); + Parameter(&builder, 0, ShapeUtil::MakeShape(U32, {}), "dividend"); XlaOp divisor = - builder.Parameter(1, ShapeUtil::MakeShape(U32, {}), "divisor"); - builder.Div(dividend, divisor); + Parameter(&builder, 1, ShapeUtil::MakeShape(U32, {}), "divisor"); + Div(dividend, divisor); TF_ASSERT_OK_AND_ASSIGN(div_computation, builder.Build()); } for (uint32 divisor : vals) { if (divisor != 0) { for (uint32 dividend : vals) { - auto dividend_literal = Literal::CreateR0(dividend); - auto divisor_literal = Literal::CreateR0(divisor); + auto dividend_literal = LiteralUtil::CreateR0(dividend); + auto divisor_literal = LiteralUtil::CreateR0(divisor); TF_ASSERT_OK_AND_ASSIGN(auto dividend_data, client_->TransferToServer(*dividend_literal)); TF_ASSERT_OK_AND_ASSIGN(auto divisor_data, @@ -389,7 +388,8 @@ XLA_TEST_F(ScalarComputationsTest, DivU32s) { {dividend_data.get(), divisor_data.get()}, &execution_options_) .ConsumeValueOrDie(); - auto expected_literal = Literal::CreateR0(dividend / divisor); + auto expected_literal = + LiteralUtil::CreateR0(dividend / divisor); EXPECT_TRUE(LiteralTestUtil::Equal(*expected_literal, *actual_literal)); } } @@ -408,18 +408,18 @@ XLA_TEST_F(ScalarComputationsTest, RemU32s) { XlaBuilder builder(TestName()); XlaOp dividend = - builder.Parameter(0, ShapeUtil::MakeShape(U32, {}), "dividend"); + Parameter(&builder, 0, ShapeUtil::MakeShape(U32, {}), "dividend"); XlaOp divisor = - builder.Parameter(1, ShapeUtil::MakeShape(U32, {}), "divisor"); - builder.Rem(dividend, divisor); + Parameter(&builder, 1, ShapeUtil::MakeShape(U32, {}), "divisor"); + Rem(dividend, divisor); TF_ASSERT_OK_AND_ASSIGN(rem_computation, builder.Build()); } for (uint32 divisor : vals) { if (divisor != 0) { for (uint32 dividend : vals) { - auto dividend_literal = Literal::CreateR0(dividend); - auto divisor_literal = Literal::CreateR0(divisor); + auto dividend_literal = LiteralUtil::CreateR0(dividend); + auto divisor_literal = LiteralUtil::CreateR0(divisor); TF_ASSERT_OK_AND_ASSIGN(auto dividend_data, client_->TransferToServer(*dividend_literal)); TF_ASSERT_OK_AND_ASSIGN(auto divisor_data, @@ -430,7 +430,8 @@ XLA_TEST_F(ScalarComputationsTest, RemU32s) { {dividend_data.get(), divisor_data.get()}, &execution_options_) .ConsumeValueOrDie(); - auto expected_literal = Literal::CreateR0(dividend % divisor); + auto expected_literal = + LiteralUtil::CreateR0(dividend % divisor); EXPECT_TRUE(LiteralTestUtil::Equal(*expected_literal, *actual_literal)); } } @@ -439,10 +440,10 @@ XLA_TEST_F(ScalarComputationsTest, RemU32s) { XLA_TEST_F(ScalarComputationsTest, RemainderTwoScalarsNonConstDividendS32) { XlaBuilder builder(TestName()); - auto x = builder.Parameter(0, ShapeUtil::MakeShape(S32, {}), "x"); - builder.Rem(x, builder.ConstantR0(80000)); + auto x = Parameter(&builder, 0, ShapeUtil::MakeShape(S32, {}), "x"); + Rem(x, ConstantR0(&builder, 80000)); - std::unique_ptr literal = Literal::CreateR0(87919); + std::unique_ptr literal = LiteralUtil::CreateR0(87919); TF_ASSERT_OK_AND_ASSIGN(auto input_data, client_->TransferToServer(*literal)); ComputeAndCompareR0(&builder, 7919, {input_data.get()}); } @@ -451,15 +452,15 @@ XLA_TEST_F(ScalarComputationsTest, DivideTwoScalarsU32) { XlaBuilder builder(TestName()); // This verifies 0xFFFFFFFE / 2 = 0x7FFFFFFF. If XLA incorrectly treated U32 // as S32, it would output -2 / 2 = -1 (0xFFFFFFFF). - builder.Div(builder.ConstantR0(0xFFFFFFFE), - builder.ConstantR0(2)); + Div(ConstantR0(&builder, 0xFFFFFFFE), + ConstantR0(&builder, 2)); ComputeAndCompareR0(&builder, 0x7FFFFFFF, {}); } XLA_TEST_F(ScalarComputationsTest, RemTwoScalarsU32) { XlaBuilder builder(TestName()); - builder.Rem(builder.ConstantR0(11), builder.ConstantR0(3)); + Rem(ConstantR0(&builder, 11), ConstantR0(&builder, 3)); ComputeAndCompareR0(&builder, 2, {}); } @@ -468,7 +469,7 @@ XLA_TEST_F(ScalarComputationsTest, AndBool) { for (bool x : {false, true}) { for (bool y : {false, true}) { XlaBuilder builder(TestName()); - builder.And(builder.ConstantR0(x), builder.ConstantR0(y)); + And(ConstantR0(&builder, x), ConstantR0(&builder, y)); ComputeAndCompareR0(&builder, x && y, {}); } @@ -479,7 +480,7 @@ XLA_TEST_F(ScalarComputationsTest, AndS32) { for (int32 x : {0, 8}) { for (int32 y : {1, -16}) { XlaBuilder builder(TestName()); - builder.And(builder.ConstantR0(x), builder.ConstantR0(y)); + And(ConstantR0(&builder, x), ConstantR0(&builder, y)); ComputeAndCompareR0(&builder, x & y, {}); } @@ -490,7 +491,7 @@ XLA_TEST_F(ScalarComputationsTest, AndU32) { for (uint32 x : {0, 8}) { for (uint32 y : {1, 16}) { XlaBuilder builder(TestName()); - builder.And(builder.ConstantR0(x), builder.ConstantR0(y)); + And(ConstantR0(&builder, x), ConstantR0(&builder, y)); ComputeAndCompareR0(&builder, x & y, {}); } @@ -501,7 +502,7 @@ XLA_TEST_F(ScalarComputationsTest, OrBool) { for (bool x : {false, true}) { for (bool y : {false, true}) { XlaBuilder builder(TestName()); - builder.Or(builder.ConstantR0(x), builder.ConstantR0(y)); + Or(ConstantR0(&builder, x), ConstantR0(&builder, y)); ComputeAndCompareR0(&builder, x || y, {}); } @@ -512,7 +513,7 @@ XLA_TEST_F(ScalarComputationsTest, OrS32) { for (int32 x : {0, 8}) { for (int32 y : {1, -16}) { XlaBuilder builder(TestName()); - builder.Or(builder.ConstantR0(x), builder.ConstantR0(y)); + Or(ConstantR0(&builder, x), ConstantR0(&builder, y)); ComputeAndCompareR0(&builder, x | y, {}); } @@ -523,7 +524,7 @@ XLA_TEST_F(ScalarComputationsTest, OrU32) { for (uint32 x : {0, 8}) { for (uint32 y : {1, 16}) { XlaBuilder builder(TestName()); - builder.Or(builder.ConstantR0(x), builder.ConstantR0(y)); + Or(ConstantR0(&builder, x), ConstantR0(&builder, y)); ComputeAndCompareR0(&builder, x | y, {}); } @@ -533,7 +534,7 @@ XLA_TEST_F(ScalarComputationsTest, OrU32) { XLA_TEST_F(ScalarComputationsTest, NotBool) { for (bool x : {false, true}) { XlaBuilder builder(TestName()); - builder.Not(builder.ConstantR0(x)); + Not(ConstantR0(&builder, x)); ComputeAndCompareR0(&builder, !x, {}); } @@ -542,7 +543,7 @@ XLA_TEST_F(ScalarComputationsTest, NotBool) { XLA_TEST_F(ScalarComputationsTest, NotS32) { for (int32 x : {-1, 0, 1}) { XlaBuilder builder(TestName()); - builder.Not(builder.ConstantR0(x)); + Not(ConstantR0(&builder, x)); ComputeAndCompareR0(&builder, ~x, {}); } @@ -551,7 +552,7 @@ XLA_TEST_F(ScalarComputationsTest, NotS32) { XLA_TEST_F(ScalarComputationsTest, NotU32) { for (uint32 x : {0, 1, 2}) { XlaBuilder builder(TestName()); - builder.Not(builder.ConstantR0(x)); + Not(ConstantR0(&builder, x)); ComputeAndCompareR0(&builder, ~x, {}); } @@ -559,18 +560,18 @@ XLA_TEST_F(ScalarComputationsTest, NotU32) { XLA_TEST_F(ScalarComputationsTest, SelectScalarTrue) { XlaBuilder builder(TestName()); - builder.Select(builder.ConstantR0(true), // The predicate. - builder.ConstantR0(123.0f), // The value on true. - builder.ConstantR0(42.0f)); // The value on false. + Select(ConstantR0(&builder, true), // The predicate. + ConstantR0(&builder, 123.0f), // The value on true. + ConstantR0(&builder, 42.0f)); // The value on false. ComputeAndCompareR0(&builder, 123.0f, {}, error_spec_); } XLA_TEST_F(ScalarComputationsTest, SelectScalarFalse) { XlaBuilder builder(TestName()); - builder.Select(builder.ConstantR0(false), // The predicate. - builder.ConstantR0(123.0f), // The value on true. - builder.ConstantR0(42.0f)); // The value on false. + Select(ConstantR0(&builder, false), // The predicate. + ConstantR0(&builder, 123.0f), // The value on true. + ConstantR0(&builder, 42.0f)); // The value on false. ComputeAndCompareR0(&builder, 42.0f, {}, error_spec_); } @@ -579,313 +580,311 @@ XLA_TEST_F(ScalarComputationsTest, SelectScalarFalse) { // templatized comparison tests. XLA_TEST_F(ScalarComputationsTest, CompareGtScalar) { XlaBuilder builder(TestName()); - builder.Gt(builder.ConstantR0(2.0f), builder.ConstantR0(1.0f)); + Gt(ConstantR0(&builder, 2.0f), ConstantR0(&builder, 1.0f)); ComputeAndCompareR0(&builder, true, {}); } // S32 comparisons. XLA_TEST_F(ScalarComputationsTest, CompareEqS32Greater) { - TestCompare(2, 1, false, &XlaBuilder::Eq); + TestCompare(2, 1, false, &Eq); } XLA_TEST_F(ScalarComputationsTest, CompareEqS32Equal) { - TestCompare(3, 3, true, &XlaBuilder::Eq); + TestCompare(3, 3, true, &Eq); } XLA_TEST_F(ScalarComputationsTest, CompareNeS32) { - TestCompare(2, 1, true, &XlaBuilder::Ne); + TestCompare(2, 1, true, &Ne); } XLA_TEST_F(ScalarComputationsTest, CompareGeS32) { - TestCompare(2, 1, true, &XlaBuilder::Ge); + TestCompare(2, 1, true, &Ge); } XLA_TEST_F(ScalarComputationsTest, CompareGtS32) { - TestCompare(1, 5, false, &XlaBuilder::Gt); + TestCompare(1, 5, false, &Gt); } XLA_TEST_F(ScalarComputationsTest, CompareLeS32) { - TestCompare(2, 1, false, &XlaBuilder::Le); + TestCompare(2, 1, false, &Le); } XLA_TEST_F(ScalarComputationsTest, CompareLtS32) { - TestCompare(9, 7, false, &XlaBuilder::Lt); + TestCompare(9, 7, false, &Lt); TestCompare(std::numeric_limits::min(), - std::numeric_limits::max(), true, &XlaBuilder::Lt); + std::numeric_limits::max(), true, &Lt); } // U32 comparisons. XLA_TEST_F(ScalarComputationsTest, CompareEqU32False) { - TestCompare(2, 1, false, &XlaBuilder::Eq); + TestCompare(2, 1, false, &Eq); } XLA_TEST_F(ScalarComputationsTest, CompareNeU32) { - TestCompare(2, 1, true, &XlaBuilder::Ne); + TestCompare(2, 1, true, &Ne); } XLA_TEST_F(ScalarComputationsTest, CompareGeU32Greater) { - TestCompare(2, 1, true, &XlaBuilder::Ge); + TestCompare(2, 1, true, &Ge); } XLA_TEST_F(ScalarComputationsTest, CompareGeU32Equal) { - TestCompare(3, 3, true, &XlaBuilder::Ge); + TestCompare(3, 3, true, &Ge); } XLA_TEST_F(ScalarComputationsTest, CompareGtU32) { - TestCompare(1, 5, false, &XlaBuilder::Gt); - TestCompare(5, 5, false, &XlaBuilder::Gt); - TestCompare(5, 1, true, &XlaBuilder::Gt); + TestCompare(1, 5, false, &Gt); + TestCompare(5, 5, false, &Gt); + TestCompare(5, 1, true, &Gt); } XLA_TEST_F(ScalarComputationsTest, CompareLeU32) { - TestCompare(2, 1, false, &XlaBuilder::Le); + TestCompare(2, 1, false, &Le); } XLA_TEST_F(ScalarComputationsTest, CompareLtU32) { - TestCompare(9, 7, false, &XlaBuilder::Lt); - TestCompare(0, std::numeric_limits::max(), true, - &XlaBuilder::Lt); + TestCompare(9, 7, false, &Lt); + TestCompare(0, std::numeric_limits::max(), true, &Lt); } // F32 comparisons. XLA_TEST_F(ScalarComputationsTest, CompareEqF32False) { - TestCompare(2.0, 1.3, false, &XlaBuilder::Eq); + TestCompare(2.0, 1.3, false, &Eq); } XLA_TEST_F(ScalarComputationsTest, CompareNeF32) { - TestCompare(2.0, 1.3, true, &XlaBuilder::Ne); + TestCompare(2.0, 1.3, true, &Ne); } XLA_TEST_F(ScalarComputationsTest, CompareGeF32Greater) { - TestCompare(2.0, 1.9, true, &XlaBuilder::Ge); + TestCompare(2.0, 1.9, true, &Ge); } XLA_TEST_F(ScalarComputationsTest, CompareGeF32Equal) { - TestCompare(3.5, 3.5, true, &XlaBuilder::Ge); + TestCompare(3.5, 3.5, true, &Ge); } XLA_TEST_F(ScalarComputationsTest, CompareGtF32) { - TestCompare(1.0, 5.2, false, &XlaBuilder::Gt); + TestCompare(1.0, 5.2, false, &Gt); } XLA_TEST_F(ScalarComputationsTest, CompareLeF32) { - TestCompare(2.0, 1.2, false, &XlaBuilder::Le); + TestCompare(2.0, 1.2, false, &Le); } XLA_TEST_F(ScalarComputationsTest, CompareLtF32) { - TestCompare(9.0, 7.2, false, &XlaBuilder::Lt); + TestCompare(9.0, 7.2, false, &Lt); } // F32 comparisons with exceptional values. The test names encode the // left/right operands at the end, and use Minf and Mzero for -inf and -0.0. XLA_TEST_F(ScalarComputationsTest, CompareLtF32MinfMzero) { - TestCompare(-INFINITY, -0.0, true, &XlaBuilder::Lt); + TestCompare(-INFINITY, -0.0, true, &Lt); } XLA_TEST_F(ScalarComputationsTest, CompareLtF32MzeroZero) { // Comparisons of 0.0 to -0.0 consider them equal in IEEE 754. - TestCompare(-0.0, 0.0, false, &XlaBuilder::Lt); + TestCompare(-0.0, 0.0, false, &Lt); } XLA_TEST_F(ScalarComputationsTest, CompareLtF32ZeroInf) { - TestCompare(0.0, INFINITY, true, &XlaBuilder::Lt); + TestCompare(0.0, INFINITY, true, &Lt); } XLA_TEST_F(ScalarComputationsTest, CompareGeF32MinfMzero) { - TestCompare(-INFINITY, -0.0, false, &XlaBuilder::Ge); + TestCompare(-INFINITY, -0.0, false, &Ge); } XLA_TEST_F(ScalarComputationsTest, CompareGeF32MzeroZero) { // Comparisons of 0.0 to -0.0 consider them equal in IEEE 754. - TestCompare(-0.0, 0.0, true, &XlaBuilder::Ge); + TestCompare(-0.0, 0.0, true, &Ge); } XLA_TEST_F(ScalarComputationsTest, CompareGeF32ZeroInf) { - TestCompare(0.0, INFINITY, false, &XlaBuilder::Ge); + TestCompare(0.0, INFINITY, false, &Ge); } XLA_TEST_F(ScalarComputationsTest, ExpScalar) { XlaBuilder builder(TestName()); - builder.Exp(builder.ConstantR0(2.0f)); + Exp(ConstantR0(&builder, 2.0f)); ComputeAndCompareR0(&builder, 7.3890562, {}, error_spec_); } XLA_TEST_F(ScalarComputationsTest, LogScalar) { XlaBuilder builder("log"); - builder.Log(builder.ConstantR0(2.0f)); + Log(ConstantR0(&builder, 2.0f)); ComputeAndCompareR0(&builder, 0.6931471, {}, error_spec_); } XLA_TEST_F(ScalarComputationsTest, TanhScalar) { XlaBuilder builder(TestName()); - builder.Tanh(builder.ConstantR0(2.0f)); + Tanh(ConstantR0(&builder, 2.0f)); ComputeAndCompareR0(&builder, 0.96402758, {}, error_spec_); } XLA_TEST_F(ScalarComputationsTest, TanhDoubleScalar) { XlaBuilder builder(TestName()); - builder.Tanh(builder.ConstantR0(2.0)); + Tanh(ConstantR0(&builder, 2.0)); ComputeAndCompareR0(&builder, 0.96402758, {}, error_spec_); } XLA_TEST_F(ScalarComputationsTest, PowScalar) { XlaBuilder builder(TestName()); - builder.Pow(builder.ConstantR0(2.0f), builder.ConstantR0(3.0f)); + Pow(ConstantR0(&builder, 2.0f), ConstantR0(&builder, 3.0f)); ComputeAndCompareR0(&builder, 8.0, {}, error_spec_); } XLA_TEST_F(ScalarComputationsTest, ClampScalarHighS32) { XlaBuilder builder(TestName()); - builder.Clamp(builder.ConstantR0(-1), // The lower bound. - builder.ConstantR0(5), // The operand to be clamped. - builder.ConstantR0(3)); // The upper bound. + Clamp(ConstantR0(&builder, -1), // The lower bound. + ConstantR0(&builder, 5), // The operand to be clamped. + ConstantR0(&builder, 3)); // The upper bound. ComputeAndCompareR0(&builder, 3, {}); } XLA_TEST_F(ScalarComputationsTest, ClampScalarMiddleS32) { XlaBuilder builder(TestName()); - builder.Clamp(builder.ConstantR0(-1), // The lower bound. - builder.ConstantR0(2), // The operand to be clamped. - builder.ConstantR0(3)); // The upper bound. + Clamp(ConstantR0(&builder, -1), // The lower bound. + ConstantR0(&builder, 2), // The operand to be clamped. + ConstantR0(&builder, 3)); // The upper bound. ComputeAndCompareR0(&builder, 2, {}); } XLA_TEST_F(ScalarComputationsTest, ClampScalarLowS32) { XlaBuilder builder(TestName()); - builder.Clamp(builder.ConstantR0(-1), // The lower bound. - builder.ConstantR0(-5), // The operand to be clamped. - builder.ConstantR0(3)); // The upper bound. + Clamp(ConstantR0(&builder, -1), // The lower bound. + ConstantR0(&builder, -5), // The operand to be clamped. + ConstantR0(&builder, 3)); // The upper bound. ComputeAndCompareR0(&builder, -1, {}); } XLA_TEST_F(ScalarComputationsTest, ClampScalarHighU32) { XlaBuilder builder(TestName()); - builder.Clamp(builder.ConstantR0(1), // The lower bound. - builder.ConstantR0(5), // The operand to be clamped. - builder.ConstantR0(3)); // The upper bound. + Clamp(ConstantR0(&builder, 1), // The lower bound. + ConstantR0(&builder, 5), // The operand to be clamped. + ConstantR0(&builder, 3)); // The upper bound. ComputeAndCompareR0(&builder, 3, {}); } XLA_TEST_F(ScalarComputationsTest, ClampScalarMiddleU32) { XlaBuilder builder(TestName()); - builder.Clamp(builder.ConstantR0(1), // The lower bound. - builder.ConstantR0(2), // The operand to be clamped. - builder.ConstantR0(3)); // The upper bound. + Clamp(ConstantR0(&builder, 1), // The lower bound. + ConstantR0(&builder, 2), // The operand to be clamped. + ConstantR0(&builder, 3)); // The upper bound. ComputeAndCompareR0(&builder, 2, {}); } XLA_TEST_F(ScalarComputationsTest, ClampScalarLowU32) { XlaBuilder builder(TestName()); - builder.Clamp(builder.ConstantR0(1), // The lower bound. - builder.ConstantR0(0), // The operand to be clamped. - builder.ConstantR0(3)); // The upper bound. + Clamp(ConstantR0(&builder, 1), // The lower bound. + ConstantR0(&builder, 0), // The operand to be clamped. + ConstantR0(&builder, 3)); // The upper bound. ComputeAndCompareR0(&builder, 1, {}); } XLA_TEST_F(ScalarComputationsTest, ClampScalarHighF32) { XlaBuilder builder(TestName()); - builder.Clamp(builder.ConstantR0(2.0f), // The lower bound. - builder.ConstantR0(5.0f), // The operand to be clamped. - builder.ConstantR0(3.0f)); // The upper bound. + Clamp(ConstantR0(&builder, 2.0f), // The lower bound. + ConstantR0(&builder, 5.0f), // The operand to be clamped. + ConstantR0(&builder, 3.0f)); // The upper bound. ComputeAndCompareR0(&builder, 3.0, {}, error_spec_); } XLA_TEST_F(ScalarComputationsTest, ClampScalarMiddleF32) { XlaBuilder builder(TestName()); - builder.Clamp(builder.ConstantR0(2.0f), // The lower bound. - builder.ConstantR0(2.5f), // The operand to be clamped. - builder.ConstantR0(3.0f)); // The upper bound. + Clamp(ConstantR0(&builder, 2.0f), // The lower bound. + ConstantR0(&builder, 2.5f), // The operand to be clamped. + ConstantR0(&builder, 3.0f)); // The upper bound. ComputeAndCompareR0(&builder, 2.5, {}, error_spec_); } XLA_TEST_F(ScalarComputationsTest, ClampScalarLowF32) { XlaBuilder builder(TestName()); - builder.Clamp(builder.ConstantR0(2.0f), // The lower bound. - builder.ConstantR0(-5.0f), // The operand to be clamped. - builder.ConstantR0(3.0f)); // The upper bound. + Clamp(ConstantR0(&builder, 2.0f), // The lower bound. + ConstantR0(&builder, -5.0f), // The operand to be clamped. + ConstantR0(&builder, 3.0f)); // The upper bound. ComputeAndCompareR0(&builder, 2.0, {}, error_spec_); } XLA_TEST_F(ScalarComputationsTest, MinS32Above) { - TestMinMax(10, 3, 3, &XlaBuilder::Min); + TestMinMax(10, 3, 3, &Min); } XLA_TEST_F(ScalarComputationsTest, MinS32Below) { - TestMinMax(-100, 3, -100, &XlaBuilder::Min); + TestMinMax(-100, 3, -100, &Min); } XLA_TEST_F(ScalarComputationsTest, MaxS32Above) { - TestMinMax(10, 3, 10, &XlaBuilder::Max); + TestMinMax(10, 3, 10, &Max); } XLA_TEST_F(ScalarComputationsTest, MaxS32Below) { - TestMinMax(-100, 3, 3, &XlaBuilder::Max); + TestMinMax(-100, 3, 3, &Max); } XLA_TEST_F(ScalarComputationsTest, MinU32Above) { const uint32 large = std::numeric_limits::max(); - TestMinMax(large, 3, 3, &XlaBuilder::Min); + TestMinMax(large, 3, 3, &Min); } XLA_TEST_F(ScalarComputationsTest, MinU32Below) { - TestMinMax(0, 5, 0, &XlaBuilder::Min); + TestMinMax(0, 5, 0, &Min); } XLA_TEST_F(ScalarComputationsTest, MaxU32Above) { const uint32 large = std::numeric_limits::max(); - TestMinMax(large, 3, large, &XlaBuilder::Max); + TestMinMax(large, 3, large, &Max); } XLA_TEST_F(ScalarComputationsTest, MaxU32Below) { - TestMinMax(0, 5, 5, &XlaBuilder::Max); + TestMinMax(0, 5, 5, &Max); } XLA_TEST_F(ScalarComputationsTest, MinF32Above) { - TestMinMax(10.1f, 3.1f, 3.1f, &XlaBuilder::Min); + TestMinMax(10.1f, 3.1f, 3.1f, &Min); } XLA_TEST_F(ScalarComputationsTest, MinF32Below) { - TestMinMax(-100.1f, 3.1f, -100.1f, &XlaBuilder::Min); + TestMinMax(-100.1f, 3.1f, -100.1f, &Min); } XLA_TEST_F(ScalarComputationsTest, MinPropagatesNan) { SetFastMathDisabled(true); - TestMinMax(NAN, 3.1f, NAN, &XlaBuilder::Min); - TestMinMax(-3.1f, NAN, NAN, &XlaBuilder::Min); + TestMinMax(NAN, 3.1f, NAN, &Min); + TestMinMax(-3.1f, NAN, NAN, &Min); } XLA_TEST_F(ScalarComputationsTest, MaxF32Above) { - TestMinMax(10.1f, 3.1f, 10.1f, &XlaBuilder::Max); + TestMinMax(10.1f, 3.1f, 10.1f, &Max); } XLA_TEST_F(ScalarComputationsTest, MaxF32Below) { - TestMinMax(-100.1f, 3.1f, 3.1f, &XlaBuilder::Max); + TestMinMax(-100.1f, 3.1f, 3.1f, &Max); } XLA_TEST_F(ScalarComputationsTest, MaxPropagatesNan) { SetFastMathDisabled(true); - TestMinMax(NAN, 3.1f, NAN, &XlaBuilder::Max); - TestMinMax(-3.1f, NAN, NAN, &XlaBuilder::Max); + TestMinMax(NAN, 3.1f, NAN, &Max); + TestMinMax(-3.1f, NAN, NAN, &Max); } XLA_TEST_F(ScalarComputationsTest, ComplicatedArithmeticExpressionF32) { // Compute the expression (1 * (3 - 1) * (7 + 0) - 4) / 20. XlaBuilder b(TestName()); - b.Div( - b.Sub(b.Mul(b.ConstantR0(1), - b.Mul(b.Sub(b.ConstantR0(3), b.ConstantR0(1)), - b.Add(b.ConstantR0(7), b.ConstantR0(0)))), - b.ConstantR0(4)), - b.ConstantR0(20)); + Div(Sub(Mul(ConstantR0(&b, 1), + Mul(Sub(ConstantR0(&b, 3), ConstantR0(&b, 1)), + Add(ConstantR0(&b, 7), ConstantR0(&b, 0)))), + ConstantR0(&b, 4)), + ConstantR0(&b, 20)); ComputeAndCompareR0(&b, 0.5, {}, error_spec_); } @@ -893,30 +892,18 @@ XLA_TEST_F(ScalarComputationsTest, ComplicatedArithmeticExpressionF32) { XLA_TEST_F(ScalarComputationsTest, ComplicatedArithmeticExpressionS32) { // Compute the expression 1 * (3 - 1) * (7 + 0) - 4. XlaBuilder b(TestName()); - b.Sub(b.Mul(b.ConstantR0(1), - b.Mul(b.Sub(b.ConstantR0(3), b.ConstantR0(1)), - b.Add(b.ConstantR0(7), b.ConstantR0(0)))), - b.ConstantR0(4)); + Sub(Mul(ConstantR0(&b, 1), + Mul(Sub(ConstantR0(&b, 3), ConstantR0(&b, 1)), + Add(ConstantR0(&b, 7), ConstantR0(&b, 0)))), + ConstantR0(&b, 4)); ComputeAndCompareR0(&b, 10, {}); } -XLA_TEST_F(ScalarComputationsTest, SqrtF320) { - XlaBuilder builder(TestName()); - Literal zero_literal = Literal::Zero(PrimitiveType::F32); - - std::unique_ptr zero_data = - client_->TransferToServer(zero_literal).ConsumeValueOrDie(); - - XlaOp zero = builder.Parameter(0, zero_literal.shape(), "zero"); - builder.SqrtF32(zero); - - ComputeAndCompareR0(&builder, 0.0f, {zero_data.get()}, error_spec_); -} XLA_TEST_F(ScalarComputationsTest, RoundScalar) { XlaBuilder builder(TestName()); - builder.Round(builder.ConstantR0(1.4f)); + Round(ConstantR0(&builder, 1.4f)); ComputeAndCompareR0(&builder, 1.0f, {}, error_spec_); } diff --git a/tensorflow/compiler/xla/tests/select_and_scatter_test.cc b/tensorflow/compiler/xla/tests/select_and_scatter_test.cc index 7015e5a6a31f506d30c2629d7735482cf354455a..b1f1e69d3cdb9398a2ebc929a344a69adc6c2223 100644 --- a/tensorflow/compiler/xla/tests/select_and_scatter_test.cc +++ b/tensorflow/compiler/xla/tests/select_and_scatter_test.cc @@ -25,7 +25,7 @@ limitations under the License. #include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/compiler/xla/client/xla_client/xla_computation.h" #include "tensorflow/compiler/xla/layout_util.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/reference_util.h" #include "tensorflow/compiler/xla/status_macros.h" #include "tensorflow/compiler/xla/tests/client_library_test_base.h" @@ -73,16 +73,16 @@ XLA_TEST_P(SelectAndScatterTest, ParamTest) { auto operand_shape = GetParam().operand_shape; Array o(operand_shape); o.FillRandom(1.5f); - auto operand = builder_.ConstantFromArray(o); + auto operand = ConstantFromArray(&builder_, o); auto source_shape = GetParam().source_shape; Array s(source_shape); s.FillRandom(12.0f); - auto source = builder_.ConstantFromArray(s); + auto source = ConstantFromArray(&builder_, s); - builder_.SelectAndScatter(operand, ge_f32_, GetParam().window_dimensions, - GetParam().window_strides, GetParam().padding_type, - source, builder_.ConstantR0(0.0f), add_f32_); + SelectAndScatter(operand, ge_f32_, GetParam().window_dimensions, + GetParam().window_strides, GetParam().padding_type, source, + ConstantR0(&builder_, 0.0f), add_f32_); ComputeAndCompare(&builder_, {}, ErrorSpec(1e-5)); } @@ -197,110 +197,110 @@ INSTANTIATE_TEST_CASE_P( // Test for F32 1D array, with a zero-element input. XLA_TEST_F(SelectAndScatterTest, R1S0F32) { - const auto operand = builder_.ConstantR1({}); - const auto source = builder_.ConstantR1({}); - builder_.SelectAndScatter(operand, ge_f32_, /*window_dimensions=*/{3}, - /*window_strides=*/{3}, Padding::kValid, source, - builder_.ConstantR0(0.0f), add_f32_); + const auto operand = ConstantR1(&builder_, {}); + const auto source = ConstantR1(&builder_, {}); + SelectAndScatter(operand, ge_f32_, /*window_dimensions=*/{3}, + /*window_strides=*/{3}, Padding::kValid, source, + ConstantR0(&builder_, 0.0f), add_f32_); ComputeAndCompareR1(&builder_, {}, {}, ErrorSpec(1e-7)); } // Test for F32 1D array, when windows do not overlap. XLA_TEST_F(SelectAndScatterTest, R1F32) { const auto operand = - builder_.ConstantR1({1.f, 9.f, 3.f, 7.f, 5.f, 6.f}); - const auto source = builder_.ConstantR1({34.f, 42.f}); + ConstantR1(&builder_, {1.f, 9.f, 3.f, 7.f, 5.f, 6.f}); + const auto source = ConstantR1(&builder_, {34.f, 42.f}); const std::vector expected = {0.f, 34.f, 0.f, 42.f, 0.f, 0.f}; - builder_.SelectAndScatter(operand, ge_f32_, /*window_dimensions=*/{3}, - /*window_strides=*/{3}, Padding::kValid, source, - builder_.ConstantR0(0.0f), add_f32_); + SelectAndScatter(operand, ge_f32_, /*window_dimensions=*/{3}, + /*window_strides=*/{3}, Padding::kValid, source, + ConstantR0(&builder_, 0.0f), add_f32_); ComputeAndCompareR1(&builder_, expected, {}, ErrorSpec(1e-7)); } // Test for S32 1D array, when windows do not overlap and the init value is 1. XLA_TEST_F(SelectAndScatterTest, R1S32) { - const auto operand = builder_.ConstantR1({-1, 0, 6, 4, -4, 10}); - const auto source = builder_.ConstantR1({-10, 20}); + const auto operand = ConstantR1(&builder_, {-1, 0, 6, 4, -4, 10}); + const auto source = ConstantR1(&builder_, {-10, 20}); const std::vector expected = {1, 1, -9, 1, 1, 21}; - builder_.SelectAndScatter(operand, ge_s32_, /*window_dimensions=*/{3}, - /*window_strides=*/{3}, Padding::kValid, source, - builder_.ConstantR0(1), add_s32_); + SelectAndScatter(operand, ge_s32_, /*window_dimensions=*/{3}, + /*window_strides=*/{3}, Padding::kValid, source, + ConstantR0(&builder_, 1), add_s32_); ComputeAndCompareR1(&builder_, expected, {}); } // Test for S32 1D array, when windows overlap with each other. XLA_TEST_F(SelectAndScatterTest, R1S32OverlappingWindow) { - const auto operand = builder_.ConstantR1({1, 9, 3, 7, 5, 6}); - const auto source = builder_.ConstantR1({34, 42, 53, 19}); + const auto operand = ConstantR1(&builder_, {1, 9, 3, 7, 5, 6}); + const auto source = ConstantR1(&builder_, {34, 42, 53, 19}); const std::vector expected = {0, 76, 0, 72, 0, 0}; - builder_.SelectAndScatter(operand, ge_s32_, /*window_dimensions=*/{3}, - /*window_strides=*/{1}, Padding::kValid, source, - builder_.ConstantR0(0), add_s32_); + SelectAndScatter(operand, ge_s32_, /*window_dimensions=*/{3}, + /*window_strides=*/{1}, Padding::kValid, source, + ConstantR0(&builder_, 0), add_s32_); ComputeAndCompareR1(&builder_, expected, {}); } // Test for S32 2D array, when windows do not overlap. XLA_TEST_F(SelectAndScatterTest, R2S32) { const auto operand = - builder_.ConstantR2({{7, 2, 5, 3, 10, 2}, {3, 8, 9, 3, 4, 2}}); - const auto source = builder_.ConstantR2({{2, 6}}); + ConstantR2(&builder_, {{7, 2, 5, 3, 10, 2}, {3, 8, 9, 3, 4, 2}}); + const auto source = ConstantR2(&builder_, {{2, 6}}); Array2D expected({{0, 0, 0, 0, 6, 0}, {0, 0, 2, 0, 0, 0}}); - builder_.SelectAndScatter(operand, ge_s32_, /*window_dimensions=*/{2, 3}, - /*window_strides=*/{2, 3}, Padding::kValid, source, - builder_.ConstantR0(0), add_s32_); + SelectAndScatter(operand, ge_s32_, /*window_dimensions=*/{2, 3}, + /*window_strides=*/{2, 3}, Padding::kValid, source, + ConstantR0(&builder_, 0), add_s32_); ComputeAndCompareR2(&builder_, expected, {}); } // Test for tie breaking rule in ge_f32_. When a tie is present, the operand // that has the lower lexicographical order (smaller index) should be chosen. XLA_TEST_F(SelectAndScatterTest, R2F32Tie) { - const auto operand = builder_.ConstantR2( - {{0.f, 0.f, 0.f}, {0.f, 0.f, 0.f}, {0.f, 0.f, 0.f}}); - const auto source = builder_.ConstantR2( - {{1.0f, 2.0f, 3.0f}, {4.f, 5.0f, 6.0f}, {7.0f, 8.0f, 9.0f}}); + const auto operand = ConstantR2( + &builder_, {{0.f, 0.f, 0.f}, {0.f, 0.f, 0.f}, {0.f, 0.f, 0.f}}); + const auto source = ConstantR2( + &builder_, {{1.0f, 2.0f, 3.0f}, {4.f, 5.0f, 6.0f}, {7.0f, 8.0f, 9.0f}}); Array2D expected( {{12.f, 9.f, 0.f}, {15.f, 9.f, 0.f}, {0.f, 0.f, 0.f}}); - builder_.SelectAndScatter(operand, ge_f32_, /*window_dimensions=*/{3, 3}, - /*window_strides=*/{1, 1}, Padding::kSame, source, - builder_.ConstantR0(0.0f), add_f32_); + SelectAndScatter(operand, ge_f32_, /*window_dimensions=*/{3, 3}, + /*window_strides=*/{1, 1}, Padding::kSame, source, + ConstantR0(&builder_, 0.0f), add_f32_); ComputeAndCompareR2(&builder_, expected, {}, ErrorSpec(1e-7)); } // Similar to SelectAndScatterTest.R2S32 but the input is transposed. XLA_TEST_F(SelectAndScatterTest, ReshapeR2S32) { - const auto operand = builder_.ConstantR2( - {{7, 3}, {2, 8}, {5, 9}, {3, 3}, {10, 4}, {2, 2}}); + const auto operand = ConstantR2( + &builder_, {{7, 3}, {2, 8}, {5, 9}, {3, 3}, {10, 4}, {2, 2}}); const auto reshape = - builder_.Reshape(operand, /*dimensions=*/{1, 0}, /*new_sizes=*/{2, 6}); - const auto source = builder_.ConstantR2({{2, 6}}); + Reshape(operand, /*dimensions=*/{1, 0}, /*new_sizes=*/{2, 6}); + const auto source = ConstantR2(&builder_, {{2, 6}}); Array2D expected({{0, 0, 0, 0, 6, 0}, {0, 0, 2, 0, 0, 0}}); - builder_.SelectAndScatter(reshape, ge_s32_, /*window_dimensions=*/{2, 3}, - /*window_strides=*/{2, 3}, Padding::kValid, source, - builder_.ConstantR0(0), add_s32_); + SelectAndScatter(reshape, ge_s32_, /*window_dimensions=*/{2, 3}, + /*window_strides=*/{2, 3}, Padding::kValid, source, + ConstantR0(&builder_, 0), add_s32_); ComputeAndCompareR2(&builder_, expected, {}); } // Test for S32 2D array, when windows overlap with each other. XLA_TEST_F(SelectAndScatterTest, R2S32OverlappingWindow) { const auto operand = - builder_.ConstantR2({{7, 2, 5, 3, 8}, {3, 8, 9, 3, 4}}); - const auto source = builder_.ConstantR2({{2, 6, 4}}); + ConstantR2(&builder_, {{7, 2, 5, 3, 8}, {3, 8, 9, 3, 4}}); + const auto source = ConstantR2(&builder_, {{2, 6, 4}}); Array2D expected({{0, 0, 0, 0, 0}, {0, 0, 12, 0, 0}}); - builder_.SelectAndScatter(operand, ge_s32_, /*window_dimensions=*/{2, 3}, - /*window_strides=*/{1, 1}, Padding::kValid, source, - builder_.ConstantR0(0), add_s32_); + SelectAndScatter(operand, ge_s32_, /*window_dimensions=*/{2, 3}, + /*window_strides=*/{1, 1}, Padding::kValid, source, + ConstantR0(&builder_, 0), add_s32_); ComputeAndCompareR2(&builder_, expected, {}); } // Test for S32 2D array, when the padding is Padding::kSAME. XLA_TEST_F(SelectAndScatterTest, R2S32SamePadding) { const auto operand = - builder_.ConstantR2({{7, 2, 5, 3, 8}, {3, 8, 9, 3, 4}}); - const auto source = builder_.ConstantR2({{2, 6, 4}}); + ConstantR2(&builder_, {{7, 2, 5, 3, 8}, {3, 8, 9, 3, 4}}); + const auto source = ConstantR2(&builder_, {{2, 6, 4}}); Array2D expected({{0, 0, 0, 0, 4}, {0, 2, 6, 0, 0}}); - builder_.SelectAndScatter(operand, ge_s32_, /*window_dimensions=*/{2, 2}, - /*window_strides=*/{2, 2}, Padding::kSame, source, - builder_.ConstantR0(0), add_s32_); + SelectAndScatter(operand, ge_s32_, /*window_dimensions=*/{2, 2}, + /*window_strides=*/{2, 2}, Padding::kSame, source, + ConstantR0(&builder_, 0), add_s32_); ComputeAndCompareR2(&builder_, expected, {}); } @@ -308,25 +308,26 @@ XLA_TEST_F(SelectAndScatterTest, R2S32SamePadding) { // with each other. XLA_TEST_F(SelectAndScatterTest, R2S32SamePaddingOverlappingWindow) { const auto operand = - builder_.ConstantR2({{7, 2, 5, 3, 8}, {3, 8, 9, 3, 4}}); + ConstantR2(&builder_, {{7, 2, 5, 3, 8}, {3, 8, 9, 3, 4}}); const auto source = - builder_.ConstantR2({{2, 6, 4, 7, 1}, {3, 5, 8, 9, 10}}); + ConstantR2(&builder_, {{2, 6, 4, 7, 1}, {3, 5, 8, 9, 10}}); Array2D expected({{0, 0, 0, 0, 8}, {0, 5, 23, 0, 19}}); - builder_.SelectAndScatter(operand, ge_s32_, /*window_dimensions=*/{2, 2}, - /*window_strides=*/{1, 1}, Padding::kSame, source, - builder_.ConstantR0(0), add_s32_); + SelectAndScatter(operand, ge_s32_, /*window_dimensions=*/{2, 2}, + /*window_strides=*/{1, 1}, Padding::kSame, source, + ConstantR0(&builder_, 0), add_s32_); ComputeAndCompareR2(&builder_, expected, {}); } XLA_TEST_F(SelectAndScatterTest, R2F32OverlappingR2Source) { - const auto operand = builder_.ConstantR2( - {{1.5f, 2.5f, 1.5f}, {3.5f, 1.5f, 3.5f}, {4.5f, 2.5f, 4.5f}}); - const auto source = builder_.ConstantR2({{1.0f, 2.0f}, {3.0f, 4.0f}}); + const auto operand = ConstantR2( + &builder_, {{1.5f, 2.5f, 1.5f}, {3.5f, 1.5f, 3.5f}, {4.5f, 2.5f, 4.5f}}); + const auto source = + ConstantR2(&builder_, {{1.0f, 2.0f}, {3.0f, 4.0f}}); Array2D expected( {{0.0f, 0.0f, 0.0f}, {1.0f, 0.0f, 2.0f}, {3.0f, 0.0f, 4.0f}}); - builder_.SelectAndScatter(operand, ge_f32_, /*window_dimensions=*/{2, 2}, - /*window_strides=*/{1, 1}, Padding::kValid, source, - builder_.ConstantR0(0.0f), add_f32_); + SelectAndScatter(operand, ge_f32_, /*window_dimensions=*/{2, 2}, + /*window_strides=*/{1, 1}, Padding::kValid, source, + ConstantR0(&builder_, 0.0f), add_f32_); ComputeAndCompareR2(&builder_, expected, {}, ErrorSpec(1e-7)); } @@ -342,16 +343,16 @@ TEST_F(SelectAndScatterTest, R4F32Valid) { {0.0f, 0.0f, 0.0f, 0.0f, 0.0f, 1.0f}}; Array4D o(4, 6, 15, 220); o.FillWithPZ(pzo); - auto operand = builder_.ConstantR4FromArray4D(o); + auto operand = ConstantR4FromArray4D(&builder_, o); Array4D e(4, 6, 15, 220); e.FillWithPZ(pze); Array4D s(2, 2, 15, 220); s.FillWithPZ(pzs); - auto source = builder_.ConstantR4FromArray4D(s); + auto source = ConstantR4FromArray4D(&builder_, s); s.FillWithPZ(pzs); - builder_.SelectAndScatter(operand, ge_f32_, {2, 3, 1, 1}, {2, 3, 1, 1}, - Padding::kValid, source, - builder_.ConstantR0(0.0f), add_f32_); + SelectAndScatter(operand, ge_f32_, {2, 3, 1, 1}, {2, 3, 1, 1}, + Padding::kValid, source, ConstantR0(&builder_, 0.0f), + add_f32_); ComputeAndCompareR4(&builder_, e, {}, ErrorSpec(1e-7)); } @@ -367,16 +368,16 @@ TEST_F(SelectAndScatterTest, R4F32Overlap) { {0.0f, 0.0f, 0.0f, 1.0f, 0.0f}}; Array4D o(4, 5, 17, 128); o.FillWithPZ(pzo); - auto operand = builder_.ConstantR4FromArray4D(o); + auto operand = ConstantR4FromArray4D(&builder_, o); Array4D e(4, 5, 17, 128); e.FillWithPZ(pze); Array4D s(2, 2, 17, 128); s.FillWithPZ(pzs); - auto source = builder_.ConstantR4FromArray4D(s); + auto source = ConstantR4FromArray4D(&builder_, s); s.FillWithPZ(pzs); - builder_.SelectAndScatter(operand, ge_f32_, {2, 3, 1, 1}, {2, 2, 1, 1}, - Padding::kValid, source, - builder_.ConstantR0(0.0f), add_f32_); + SelectAndScatter(operand, ge_f32_, {2, 3, 1, 1}, {2, 2, 1, 1}, + Padding::kValid, source, ConstantR0(&builder_, 0.0f), + add_f32_); ComputeAndCompareR4(&builder_, e, {}, ErrorSpec(1e-7)); } @@ -392,16 +393,16 @@ TEST_F(SelectAndScatterTest, R4F32OverlapSmall) { {0.0f, 0.0f, 0.0f, 1.0f, 0.0f}}; Array4D o(4, 5, 1, 1); o.FillWithPZ(pzo); - auto operand = builder_.ConstantR4FromArray4D(o); + auto operand = ConstantR4FromArray4D(&builder_, o); Array4D e(4, 5, 1, 1); e.FillWithPZ(pze); Array4D s(2, 2, 1, 1); s.FillWithPZ(pzs); - auto source = builder_.ConstantR4FromArray4D(s); + auto source = ConstantR4FromArray4D(&builder_, s); s.FillWithPZ(pzs); - builder_.SelectAndScatter(operand, ge_f32_, {2, 3, 1, 1}, {2, 2, 1, 1}, - Padding::kValid, source, - builder_.ConstantR0(0.0f), add_f32_); + SelectAndScatter(operand, ge_f32_, {2, 3, 1, 1}, {2, 2, 1, 1}, + Padding::kValid, source, ConstantR0(&builder_, 0.0f), + add_f32_); ComputeAndCompareR4(&builder_, e, {}, ErrorSpec(1e-7)); } @@ -414,39 +415,39 @@ TEST_F(SelectAndScatterTest, R4F32RefValidFixedSmall) { Array2D pzs = {{2.0f, 6.0f}, {3.0f, 1.0f}}; Array4D o(4, 6, 4, 4); o.FillWithPZ(pzo); - auto operand = builder_.ConstantR4FromArray4D(o); + auto operand = ConstantR4FromArray4D(&builder_, o); Array4D s(2, 2, 4, 4); s.FillWithPZ(pzs); - auto source = builder_.ConstantR4FromArray4D(s); + auto source = ConstantR4FromArray4D(&builder_, s); s.FillWithPZ(pzs); - builder_.SelectAndScatter(operand, ge_f32_, {2, 3, 1, 1}, {2, 3, 1, 1}, - Padding::kValid, source, - builder_.ConstantR0(0.0f), add_f32_); + SelectAndScatter(operand, ge_f32_, {2, 3, 1, 1}, {2, 3, 1, 1}, + Padding::kValid, source, ConstantR0(&builder_, 0.0f), + add_f32_); auto e = ReferenceUtil::SelectAndScatter4DGePlus(o, s, 0.0f, {2, 3, 1, 1}, {2, 3, 1, 1}, false); ComputeAndCompareR4(&builder_, *e, {}, ErrorSpec(1e-7)); } XLA_TEST_F(SelectAndScatterTest, R1F32OverlappingWindowMaxScatter) { - const auto operand = builder_.ConstantR1({1, 2, 3, 100, 3, 2, 1}); - const auto source = builder_.ConstantR1({34, 42, 53, 19}); + const auto operand = ConstantR1(&builder_, {1, 2, 3, 100, 3, 2, 1}); + const auto source = ConstantR1(&builder_, {34, 42, 53, 19}); const std::vector expected = {0, 0, 0, 53, 0, 0, 0}; - builder_.SelectAndScatter(operand, ge_f32_, /*window_dimensions=*/{4}, - /*window_strides=*/{1}, Padding::kValid, source, - builder_.ConstantR0(0), max_f32_); + SelectAndScatter(operand, ge_f32_, /*window_dimensions=*/{4}, + /*window_strides=*/{1}, Padding::kValid, source, + ConstantR0(&builder_, 0), max_f32_); ComputeAndCompareR1(&builder_, expected, {}, ErrorSpec(1e-7)); } XLA_TEST_F(SelectAndScatterTest, R1F32OverlappingWindowMinScatter) { - const auto operand = builder_.ConstantR1({1, 2, 3, 100, 3, 2, 1}); - const auto source = builder_.ConstantR1({34, 42, 53, 19}); + const auto operand = ConstantR1(&builder_, {1, 2, 3, 100, 3, 2, 1}); + const auto source = ConstantR1(&builder_, {34, 42, 53, 19}); const float max_float = std::numeric_limits::max(); const std::vector expected = {max_float, max_float, max_float, 19, max_float, max_float, max_float}; - builder_.SelectAndScatter(operand, ge_f32_, /*window_dimensions=*/{4}, - /*window_strides=*/{1}, Padding::kValid, source, - builder_.ConstantR0(max_float), min_f32_); + SelectAndScatter(operand, ge_f32_, /*window_dimensions=*/{4}, + /*window_strides=*/{1}, Padding::kValid, source, + ConstantR0(&builder_, max_float), min_f32_); ComputeAndCompareR1(&builder_, expected, {}, ErrorSpec(1e-7)); } diff --git a/tensorflow/compiler/xla/tests/select_test.cc b/tensorflow/compiler/xla/tests/select_test.cc index 72707f224446c7585d1d90ac6681a7b38c41d5f1..59409ab26e1c19a8271318c18e19caa7b8ddc3b7 100644 --- a/tensorflow/compiler/xla/tests/select_test.cc +++ b/tensorflow/compiler/xla/tests/select_test.cc @@ -35,50 +35,52 @@ class SelectTest : public ClientLibraryTestBase { TEST_F(SelectTest, SelectScalarF32True) { XlaBuilder builder(TestName()); - auto pred = builder.ConstantR0(true); - auto on_true = builder.ConstantR0(123.0f); - auto on_false = builder.ConstantR0(42.0f); - auto result = builder.Select(pred, on_true, on_false); + auto pred = ConstantR0(&builder, true); + auto on_true = ConstantR0(&builder, 123.0f); + auto on_false = ConstantR0(&builder, 42.0f); + Select(pred, on_true, on_false); ComputeAndCompareR0(&builder, 123.0f, {}, error_spec_); } TEST_F(SelectTest, SelectScalarS32True) { XlaBuilder builder(TestName()); - auto pred = builder.ConstantR0(true); - auto on_true = builder.ConstantR0(-42); - auto on_false = builder.ConstantR0(42); - auto result = builder.Select(pred, on_true, on_false); + auto pred = ConstantR0(&builder, true); + auto on_true = ConstantR0(&builder, -42); + auto on_false = ConstantR0(&builder, 42); + Select(pred, on_true, on_false); ComputeAndCompareR0(&builder, -42, {}); } TEST_F(SelectTest, SelectScalarF32False) { XlaBuilder builder(TestName()); - auto pred = builder.ConstantR0(false); - auto on_true = builder.ConstantR0(123.0f); - auto on_false = builder.ConstantR0(42.0f); - auto result = builder.Select(pred, on_true, on_false); + auto pred = ConstantR0(&builder, false); + auto on_true = ConstantR0(&builder, 123.0f); + auto on_false = ConstantR0(&builder, 42.0f); + Select(pred, on_true, on_false); ComputeAndCompareR0(&builder, 42.0f, {}, error_spec_); } XLA_TEST_F(SelectTest, SelectR1S0F32WithConstantR1S0PRED) { XlaBuilder builder(TestName()); - auto pred = builder.ConstantR1({}); - auto on_true = builder.ConstantR1({}); - auto on_false = builder.ConstantR1({}); - auto select = builder.Select(pred, on_true, on_false); + auto pred = ConstantR1(&builder, {}); + auto on_true = ConstantR1(&builder, {}); + auto on_false = ConstantR1(&builder, {}); + Select(pred, on_true, on_false); ComputeAndCompareR1(&builder, {}, {}, error_spec_); } TEST_F(SelectTest, SelectR1F32WithConstantR1PRED) { XlaBuilder builder(TestName()); - auto pred = builder.ConstantR1({false, true, false, true, false}); - auto on_true = builder.ConstantR1({-2.5f, 25.5f, 2.25f, -10.0f, 6.0f}); - auto on_false = builder.ConstantR1({10.0f, 5.0f, 1.0f, 10.0f, -6.0f}); - auto select = builder.Select(pred, on_true, on_false); + auto pred = ConstantR1(&builder, {false, true, false, true, false}); + auto on_true = + ConstantR1(&builder, {-2.5f, 25.5f, 2.25f, -10.0f, 6.0f}); + auto on_false = + ConstantR1(&builder, {10.0f, 5.0f, 1.0f, 10.0f, -6.0f}); + Select(pred, on_true, on_false); ComputeAndCompareR1(&builder, {10.0f, 25.5f, 1.0f, -10.0f, -6.0f}, {}, error_spec_); @@ -88,12 +90,12 @@ XLA_TEST_F(SelectTest, SelectR1S0F32WithCmpR1S0S32s) { // Similar to SelectR1S0F32WithConstantR1S0PRED, except that the pred vector // is not a constant, but rather the result of comparing two other vectors. XlaBuilder builder(TestName()); - auto v1 = builder.ConstantR1({}); - auto v2 = builder.ConstantR1({}); - auto cmp = builder.Eq(v1, v2); - auto on_true = builder.ConstantR1({}); - auto on_false = builder.ConstantR1({}); - auto select = builder.Select(cmp, on_true, on_false); + auto v1 = ConstantR1(&builder, {}); + auto v2 = ConstantR1(&builder, {}); + auto cmp = Eq(v1, v2); + auto on_true = ConstantR1(&builder, {}); + auto on_false = ConstantR1(&builder, {}); + Select(cmp, on_true, on_false); ComputeAndCompareR1(&builder, {}, {}, error_spec_); } @@ -102,12 +104,14 @@ TEST_F(SelectTest, SelectR1F32WithCmpR1S32s) { // Similar to SelectR1F32WithConstantR1PRED, except that the pred vector is // not a constant, but rather the result of comparing two other vectors. XlaBuilder builder(TestName()); - auto v1 = builder.ConstantR1({1, 2, 3, 4, 5}); - auto v2 = builder.ConstantR1({9, 2, 9, 4, 9}); - auto cmp = builder.Eq(v1, v2); - auto on_true = builder.ConstantR1({-2.5f, 25.5f, 2.25f, -10.0f, 6.0f}); - auto on_false = builder.ConstantR1({10.0f, 5.0f, 1.0f, 10.0f, -6.0f}); - auto select = builder.Select(cmp, on_true, on_false); + auto v1 = ConstantR1(&builder, {1, 2, 3, 4, 5}); + auto v2 = ConstantR1(&builder, {9, 2, 9, 4, 9}); + auto cmp = Eq(v1, v2); + auto on_true = + ConstantR1(&builder, {-2.5f, 25.5f, 2.25f, -10.0f, 6.0f}); + auto on_false = + ConstantR1(&builder, {10.0f, 5.0f, 1.0f, 10.0f, -6.0f}); + Select(cmp, on_true, on_false); ComputeAndCompareR1(&builder, {10.0f, 25.5f, 1.0f, -10.0f, -6.0f}, {}, error_spec_); @@ -116,12 +120,14 @@ TEST_F(SelectTest, SelectR1F32WithCmpR1S32s) { TEST_F(SelectTest, SelectR1F32WithCmpR1F32s) { // Similar to SelectR1F32WithCmpR1S32s, except "gt"-comparing two R1F32s. XlaBuilder builder(TestName()); - auto v1 = builder.ConstantR1({1.0f, 2.0f, 3.0f, 4.0f, 5.0f}); - auto v2 = builder.ConstantR1({-1.0f, -2.0f, 13.0f, 14.0f, 4.4f}); - auto cmp = builder.Gt(v1, v2); - auto on_true = builder.ConstantR1({-2.5f, 25.5f, 2.25f, -10.0f, 6.0f}); - auto on_false = builder.ConstantR1({10.0f, 5.0f, 1.0f, 10.0f, -6.0f}); - auto select = builder.Select(cmp, on_true, on_false); + auto v1 = ConstantR1(&builder, {1.0f, 2.0f, 3.0f, 4.0f, 5.0f}); + auto v2 = ConstantR1(&builder, {-1.0f, -2.0f, 13.0f, 14.0f, 4.4f}); + auto cmp = Gt(v1, v2); + auto on_true = + ConstantR1(&builder, {-2.5f, 25.5f, 2.25f, -10.0f, 6.0f}); + auto on_false = + ConstantR1(&builder, {10.0f, 5.0f, 1.0f, 10.0f, -6.0f}); + Select(cmp, on_true, on_false); ComputeAndCompareR1(&builder, {-2.5f, 25.5f, 1.0f, 10.0f, 6.0f}, {}, error_spec_); @@ -140,8 +146,8 @@ TEST_F(SelectTest, SelectR1F32WithCmpR1F32sFromParamsSmall) { {21.0f, 22.0f, 23.0f, 24.0f}, /*parameter_number=*/1, /*name=*/"v2", /*builder=*/&builder, /*data_handle=*/&v2); - auto cmp = builder.Gt(v1, v2); - auto select = builder.Select(cmp, v1, v2); + auto cmp = Gt(v1, v2); + Select(cmp, v1, v2); ComputeAndCompareR1(&builder, {41.0f, 22.0f, 23.0f, 84.0f}, {param0_data.get(), param1_data.get()}, error_spec_); @@ -181,8 +187,8 @@ TEST_F(SelectTest, SelectR1F32WithCmpR1F32sFromParamsLarge) { CreateR1Parameter(v2vec, /*parameter_number=*/1, /*name=*/"v2", /*builder=*/&builder, /*data_handle=*/&v2); - auto cmp = builder.Gt(v1, v2); - auto select = builder.Select(cmp, v1, v2); + auto cmp = Gt(v1, v2); + Select(cmp, v1, v2); ComputeAndCompareR1(&builder, expected_vec, {param0_data.get(), param1_data.get()}, error_spec_); @@ -192,14 +198,14 @@ TEST_F(SelectTest, SelectR1F32WithCmpR1S32ToScalar) { // "gt"-compares a R1S32 with a S32 scalar, and uses the resulting R1PRED to // select between two R1F32s. XlaBuilder builder(TestName()); - auto v = builder.ConstantR1({1, -1, 2, -2}); - auto s = builder.ConstantR0(0); - auto cmp = builder.Gt(v, s); + auto v = ConstantR1(&builder, {1, -1, 2, -2}); + auto s = ConstantR0(&builder, 0); + auto cmp = Gt(v, s); - auto on_true = builder.ConstantR1({11.0f, 22.0f, 33.0f, 44.0f}); + auto on_true = ConstantR1(&builder, {11.0f, 22.0f, 33.0f, 44.0f}); auto on_false = - builder.ConstantR1({-111.0f, -222.0f, -333.0f, -444.0f}); - auto select = builder.Select(cmp, on_true, on_false); + ConstantR1(&builder, {-111.0f, -222.0f, -333.0f, -444.0f}); + Select(cmp, on_true, on_false); ComputeAndCompareR1(&builder, {11.0f, -222.0f, 33.0f, -444.0f}, {}, error_spec_); @@ -209,14 +215,14 @@ TEST_F(SelectTest, SelectR1F32WithCmpR1F32ToScalar) { // "gt"-compares a R1F32 with a F32 scalar, and uses the resulting R1PRED to // select between two R1F32s. XlaBuilder builder(TestName()); - auto v = builder.ConstantR1({1.0f, 2.0f, 3.0f, 4.0f}); - auto s = builder.ConstantR0(2.5f); - auto cmp = builder.Gt(v, s); + auto v = ConstantR1(&builder, {1.0f, 2.0f, 3.0f, 4.0f}); + auto s = ConstantR0(&builder, 2.5f); + auto cmp = Gt(v, s); - auto on_true = builder.ConstantR1({11.0f, 22.0f, 33.0f, 44.0f}); + auto on_true = ConstantR1(&builder, {11.0f, 22.0f, 33.0f, 44.0f}); auto on_false = - builder.ConstantR1({-111.0f, -222.0f, -333.0f, -444.0f}); - auto select = builder.Select(cmp, on_true, on_false); + ConstantR1(&builder, {-111.0f, -222.0f, -333.0f, -444.0f}); + Select(cmp, on_true, on_false); ComputeAndCompareR1(&builder, {-111.0f, -222.0f, 33.0f, 44.0f}, {}, error_spec_); @@ -225,10 +231,10 @@ TEST_F(SelectTest, SelectR1F32WithCmpR1F32ToScalar) { XLA_TEST_F(SelectTest, SelectR1S0F32WithScalarPredicate) { for (bool which : {false, true}) { XlaBuilder builder(TestName()); - auto pred = builder.ConstantR0(which); - auto on_true = builder.ConstantR1({}); - auto on_false = builder.ConstantR1({}); - auto select = builder.Select(pred, on_true, on_false); + auto pred = ConstantR0(&builder, which); + auto on_true = ConstantR1(&builder, {}); + auto on_false = ConstantR1(&builder, {}); + Select(pred, on_true, on_false); ComputeAndCompareR1(&builder, {}, {}, error_spec_); } @@ -236,20 +242,20 @@ XLA_TEST_F(SelectTest, SelectR1S0F32WithScalarPredicate) { TEST_F(SelectTest, SelectR1F32WithScalarPredicateTrue) { XlaBuilder builder(TestName()); - auto pred = builder.ConstantR0(true); - auto on_true = builder.ConstantR1({-2.5f, 25.5f}); - auto on_false = builder.ConstantR1({10.0f, 5.0f}); - auto select = builder.Select(pred, on_true, on_false); + auto pred = ConstantR0(&builder, true); + auto on_true = ConstantR1(&builder, {-2.5f, 25.5f}); + auto on_false = ConstantR1(&builder, {10.0f, 5.0f}); + Select(pred, on_true, on_false); ComputeAndCompareR1(&builder, {-2.5f, 25.5f}, {}, error_spec_); } TEST_F(SelectTest, SelectR1F32WithScalarPredicateFalse) { XlaBuilder builder(TestName()); - auto pred = builder.ConstantR0(false); - auto on_true = builder.ConstantR1({-2.5f, 25.5f}); - auto on_false = builder.ConstantR1({10.0f, 5.0f}); - auto select = builder.Select(pred, on_true, on_false); + auto pred = ConstantR0(&builder, false); + auto on_true = ConstantR1(&builder, {-2.5f, 25.5f}); + auto on_false = ConstantR1(&builder, {10.0f, 5.0f}); + Select(pred, on_true, on_false); ComputeAndCompareR1(&builder, {10.0f, 5.0f}, {}, error_spec_); } diff --git a/tensorflow/compiler/xla/tests/slice_test.cc b/tensorflow/compiler/xla/tests/slice_test.cc index 5653bf11a7364bf9ed79bcb6b53f7db31f454803..48138e7b076af7ff206ea78c93bd99425323c8ba 100644 --- a/tensorflow/compiler/xla/tests/slice_test.cc +++ b/tensorflow/compiler/xla/tests/slice_test.cc @@ -42,8 +42,8 @@ TEST_F(SliceTest, Slice3x3x3_To_3x3x1_F32) { values.FillIota(0); XlaBuilder builder(TestName()); - auto original = builder.ConstantR3FromArray3D(values); - builder.Slice(original, {0, 0, 0}, {3, 3, 1}, {1, 1, 1}); + auto original = ConstantR3FromArray3D(&builder, values); + Slice(original, {0, 0, 0}, {3, 3, 1}, {1, 1, 1}); Array3D expected{ {{0.0}, {3.0}, {6.0}}, {{9.0}, {12.0}, {15.0}}, {{18.0}, {21.0}, {24.0}}}; @@ -55,8 +55,8 @@ TEST_F(SliceTest, Slice3x3x3_To_3x1x3_F32) { values.FillIota(0); XlaBuilder builder(TestName()); - auto original = builder.ConstantR3FromArray3D(values); - builder.Slice(original, {0, 0, 0}, {3, 1, 3}, {1, 1, 1}); + auto original = ConstantR3FromArray3D(&builder, values); + Slice(original, {0, 0, 0}, {3, 1, 3}, {1, 1, 1}); Array3D expected{ {{0.0, 1.0, 2.0}}, {{9.0, 10.0, 11.0}}, {{18.0, 19.0, 20.0}}}; @@ -68,8 +68,8 @@ TEST_F(SliceTest, Slice3x3x3_To_1x3x3_F32) { values.FillIota(0); XlaBuilder builder(TestName()); - auto original = builder.ConstantR3FromArray3D(values); - builder.Slice(original, {0, 0, 0}, {1, 3, 3}, {1, 1, 1}); + auto original = ConstantR3FromArray3D(&builder, values); + Slice(original, {0, 0, 0}, {1, 3, 3}, {1, 1, 1}); Array3D expected{ {{{0.0, 1.0, 2.0}, {3.0, 4.0, 5.0}, {6.0, 7.0, 8.0}}}}; @@ -78,24 +78,24 @@ TEST_F(SliceTest, Slice3x3x3_To_1x3x3_F32) { XLA_TEST_F(SliceTest, Slice0x0to0x0F32) { XlaBuilder builder(TestName()); - auto original = builder.ConstantR2FromArray2D(Array2D(0, 0)); - builder.Slice(original, {0, 0}, {0, 0}, {1, 1}); + auto original = ConstantR2FromArray2D(&builder, Array2D(0, 0)); + Slice(original, {0, 0}, {0, 0}, {1, 1}); ComputeAndCompareR2(&builder, Array2D(0, 0), {}); } XLA_TEST_F(SliceTest, Slice0x20to0x5F32) { XlaBuilder builder(TestName()); - auto original = builder.ConstantR2FromArray2D(Array2D(0, 20)); - builder.Slice(original, {0, 15}, {0, 20}, {1, 1}); + auto original = ConstantR2FromArray2D(&builder, Array2D(0, 20)); + Slice(original, {0, 15}, {0, 20}, {1, 1}); ComputeAndCompareR2(&builder, Array2D(0, 5), {}); } XLA_TEST_F(SliceTest, Slice3x0to2x0F32) { XlaBuilder builder(TestName()); - auto original = builder.ConstantR2FromArray2D(Array2D(3, 0)); - builder.Slice(original, {1, 0}, {3, 0}, {1, 1}); + auto original = ConstantR2FromArray2D(&builder, Array2D(3, 0)); + Slice(original, {1, 0}, {3, 0}, {1, 1}); ComputeAndCompareR2(&builder, Array2D(2, 0), {}); } @@ -109,8 +109,8 @@ XLA_TEST_F(SliceTest, SliceQuadrantOf256x256) { } XlaBuilder builder(TestName()); - auto original = builder.ConstantR2FromArray2D(values); - builder.Slice(original, {128, 128}, {256, 256}, {1, 1}); + auto original = ConstantR2FromArray2D(&builder, values); + Slice(original, {128, 128}, {256, 256}, {1, 1}); Array2D expected(128, 128); for (int row = 0; row < 128; ++row) { @@ -127,8 +127,8 @@ TEST_F(SliceTest, Slice_1x4096_To_1x1024) { std::iota(values.data(), values.data() + 4096, 0.0); XlaBuilder builder(TestName()); - auto original = builder.ConstantR2FromArray2D(values); - builder.Slice(original, {0, 3072}, {1, 4096}, {1, 1}); + auto original = ConstantR2FromArray2D(&builder, values); + Slice(original, {0, 3072}, {1, 4096}, {1, 1}); Array2D expected(1, 1024); std::iota(expected.data(), expected.data() + 1024, 3072.0); @@ -148,8 +148,8 @@ TEST_F(SliceTest, Slice_16x4_To_16x2) { } } XlaBuilder builder(TestName()); - auto original = builder.ConstantR2FromArray2D(values); - builder.Slice(original, {0, 0}, {16, 2}, {1, 1}); + auto original = ConstantR2FromArray2D(&builder, values); + Slice(original, {0, 0}, {16, 2}, {1, 1}); ComputeAndCompareR2(&builder, expected, {}, ErrorSpec(0.000001)); } @@ -160,8 +160,8 @@ TEST_F(SliceTest, SliceR4ThreeDimsMiddleMinor) { auto expected = ReferenceUtil::Slice4D( values, {{1, 0, 8, 0}}, {{2, 2, 16, 128}}, /*strides=*/{{1, 1, 1, 1}}); XlaBuilder builder(TestName()); - auto original = builder.ConstantR4FromArray4D(values); - builder.Slice(original, {1, 0, 8, 0}, {2, 2, 16, 128}, {1, 1, 1, 1}); + auto original = ConstantR4FromArray4D(&builder, values); + Slice(original, {1, 0, 8, 0}, {2, 2, 16, 128}, {1, 1, 1, 1}); ComputeAndCompareR4(&builder, *expected, {}, ErrorSpec(0.000001)); } @@ -170,11 +170,11 @@ XLA_TEST_F(SliceTest, StridedSliceR4WithOutputLayout) { values.FillRandom(3.14f); auto expected = ReferenceUtil::Slice4D(values, {{0, 0, 0, 0}}, {{2, 4, 6, 8}}, /*strides=*/{{1, 1, 2, 1}}); - auto expected_literal = Literal::CreateR4FromArray4DWithLayout( + auto expected_literal = LiteralUtil::CreateR4FromArray4DWithLayout( *expected, LayoutUtil::MakeLayout({0, 1, 2, 3})); XlaBuilder builder(TestName()); - auto original = builder.ConstantR4FromArray4D(values); - builder.Slice(original, {0, 0, 0, 0}, {2, 4, 6, 8}, {1, 1, 2, 1}); + auto original = ConstantR4FromArray4D(&builder, values); + Slice(original, {0, 0, 0, 0}, {2, 4, 6, 8}, {1, 1, 2, 1}); ComputeAndCompareLiteral(&builder, *expected_literal, {}, ErrorSpec(0.000001), &expected_literal->shape()); } @@ -197,12 +197,12 @@ class SliceR1Test : public ClientLibraryTestBase, // vector. tensorflow::gtl::InlinedVector input(spec.input_dim0); std::iota(input.begin(), input.end(), NativeT()); - auto literal = Literal::CreateR1(input); + auto literal = LiteralUtil::CreateR1(input); XlaBuilder builder(TestName()); - auto original = builder.Parameter(0, literal->shape(), "p0"); - builder.Slice(original, {spec.slice_start}, {spec.slice_limit}, - {spec.slice_stride}); + auto original = Parameter(&builder, 0, literal->shape(), "p0"); + Slice(original, {spec.slice_start}, {spec.slice_limit}, + {spec.slice_stride}); // Ditto. tensorflow::gtl::InlinedVector expected; @@ -368,12 +368,12 @@ XLA_TEST_P(SliceR2Test, DoIt) { const R2Spec& spec = GetParam(); Array2D input(spec.input_dim0, spec.input_dim1); input.FillUnique(); - auto literal = Literal::CreateR2FromArray2DWithLayout( + auto literal = LiteralUtil::CreateR2FromArray2DWithLayout( input, LayoutUtil::MakeLayout(spec.layout)); XlaBuilder builder(TestName()); - auto a = builder.Parameter(0, literal->shape(), "p0"); - builder.Slice(a, spec.slice_starts, spec.slice_limits, spec.slice_strides); + auto a = Parameter(&builder, 0, literal->shape(), "p0"); + Slice(a, spec.slice_starts, spec.slice_limits, spec.slice_strides); TF_ASSERT_OK_AND_ASSIGN(std::unique_ptr arg, client_->TransferToServer(*literal)); @@ -463,13 +463,12 @@ class SliceR4Test : public ClientLibraryTestBase, auto expected = ReferenceUtil::Slice4D( values, spec.slice_starts, spec.slice_limits, spec.slice_strides); XlaBuilder builder(TestName()); - auto literal = Literal::CreateR4FromArray4DWithLayout( + auto literal = LiteralUtil::CreateR4FromArray4DWithLayout( values, LayoutUtil::MakeLayout(spec.input_layout)); - auto parameter = builder.Parameter(0, literal->shape(), "p0"); + auto parameter = Parameter(&builder, 0, literal->shape(), "p0"); TF_ASSERT_OK_AND_ASSIGN(std::unique_ptr arg, client_->TransferToServer(*literal)); - builder.Slice(parameter, spec.slice_starts, spec.slice_limits, - spec.slice_strides); + Slice(parameter, spec.slice_starts, spec.slice_limits, spec.slice_strides); ComputeAndCompareR4(&builder, *expected, {arg.get()}, ErrorSpec(0.000001)); } }; diff --git a/tensorflow/compiler/xla/tests/test_utils.cc b/tensorflow/compiler/xla/tests/test_utils.cc index 000535a982fb08af69e7b317501f82ba7f402fb9..2647937013222ccfdae98b0c1d141f461020b5c9 100644 --- a/tensorflow/compiler/xla/tests/test_utils.cc +++ b/tensorflow/compiler/xla/tests/test_utils.cc @@ -14,6 +14,7 @@ limitations under the License. ==============================================================================*/ #include "tensorflow/compiler/xla/tests/test_utils.h" +#include "tensorflow/compiler/xla/literal_util.h" #include "tensorflow/compiler/xla/primitive_util.h" #include "tensorflow/compiler/xla/service/hlo_dataflow_analysis.h" #include "tensorflow/compiler/xla/service/hlo_verifier.h" @@ -110,7 +111,7 @@ StatusOr> MakeFakeLiteralInternal( MakeFakeLiteralInternal(element_shape, engine)); elements.push_back(std::move(element)); } - return Literal::MakeTupleOwned(std::move(elements)); + return LiteralUtil::MakeTupleOwned(std::move(elements)); } if (engine == nullptr) { return Literal::CreateFromShape(shape); @@ -161,6 +162,9 @@ StatusOr> MakeFakeLiteralInternal( })); break; } + // Token requires no data. + case TOKEN: + break; default: return Unimplemented("Unsupported type for fake literal generation: %s", ShapeUtil::HumanString(shape).c_str()); @@ -217,7 +221,7 @@ std::unique_ptr MakeRandomNonwrappingSliceIndex( start_indices[i] = generator(*engine); } } - return Literal::CreateR1(start_indices); + return LiteralUtil::CreateR1(start_indices); } // Use dataflow analysis on each parameter to see if there are uses that would @@ -315,9 +319,9 @@ StatusOr> CreateLiteralForConstrainedUses( } else if (needs_constant != nullptr) { switch (constant_type) { case ConstantType::kZero: - return Literal::Zero(param.shape().element_type()).CloneToUnique(); + return LiteralUtil::Zero(param.shape().element_type()).CloneToUnique(); case ConstantType::kOne: - return Literal::One(param.shape().element_type()).CloneToUnique(); + return LiteralUtil::One(param.shape().element_type()).CloneToUnique(); case ConstantType::kUnknown: // We want the identity element for the computation, but we don't really // know what it is - so any value we generate will be just as wrong. diff --git a/tensorflow/compiler/xla/tests/test_utils.h b/tensorflow/compiler/xla/tests/test_utils.h index a8689f64981569ceb7c8a712f8ece00c99e8cf2d..e59f215a9a3ace80d7a23e1bbc40970c7a63ea0d 100644 --- a/tensorflow/compiler/xla/tests/test_utils.h +++ b/tensorflow/compiler/xla/tests/test_utils.h @@ -21,7 +21,7 @@ limitations under the License. #include #include "tensorflow/compiler/xla/layout_util.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/ptr_util.h" #include "tensorflow/compiler/xla/service/hlo_module.h" #include "tensorflow/compiler/xla/xla_data.pb.h" diff --git a/tensorflow/compiler/xla/tests/test_utils_test.cc b/tensorflow/compiler/xla/tests/test_utils_test.cc index 59afd28a80c0fbf3df38457cd05961c883769856..8f424ae81f592bfd8accd8decb8fc363f7561c73 100644 --- a/tensorflow/compiler/xla/tests/test_utils_test.cc +++ b/tensorflow/compiler/xla/tests/test_utils_test.cc @@ -16,6 +16,7 @@ limitations under the License. #include "tensorflow/compiler/xla/tests/test_utils.h" #include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" +#include "tensorflow/compiler/xla/service/hlo_parser.h" #include "tensorflow/compiler/xla/shape_util.h" #include "tensorflow/compiler/xla/tests/local_client_test_base.h" #include "tensorflow/compiler/xla/tests/test_macros.h" @@ -31,16 +32,16 @@ XLA_TEST_F(TestUtilsTest, UnusedParam) { XlaBuilder builder(TestName()); // Make the reduction lambda. Shape single_float = ShapeUtil::MakeShape(F32, {}); - builder.Parameter(0, single_float, "unused"); - builder.Parameter(1, single_float, "used"); + Parameter(&builder, 0, single_float, "unused"); + Parameter(&builder, 1, single_float, "used"); auto computation_status = builder.Build(); TF_ASSERT_OK(computation_status.status()); // Make the reduction. Shape pair_float = ShapeUtil::MakeShape(F32, {2}); - builder.Reduce(builder.Parameter(0, pair_float, "operand"), - builder.Parameter(1, single_float, "init"), - computation_status.ValueOrDie(), {0}); + Reduce(Parameter(&builder, 0, pair_float, "operand"), + Parameter(&builder, 1, single_float, "init"), + computation_status.ValueOrDie(), {0}); computation_status = builder.Build(); TF_ASSERT_OK(computation_status.status()); @@ -53,5 +54,23 @@ XLA_TEST_F(TestUtilsTest, UnusedParam) { TF_ASSERT_OK(MakeFakeArguments(&module).status()); } +XLA_TEST_F(TestUtilsTest, Token) { + auto module = ParseHloString( + R"(HloModule outfeed_module + + ENTRY InfeedToOutfeed { + token = token[] parameter(0) + infeed = ((u32[3]{0}, pred[]), token[]) infeed(token) + infeed.data = (u32[3]{0}, pred[]) get-tuple-element(infeed), index=0 + outfeed = token[] outfeed(infeed.data, token) + ROOT infeed.1 = ((u32[3]{0}, pred[]), token[]) infeed(token) + infeed.1.data = (u32[3]{0}, pred[]) get-tuple-element(infeed.1), index=0 + infeed.1.token = token[] get-tuple-element(infeed.1), index=1 + outfeed.1 = token[] outfeed(infeed.1.data, infeed.1.token) + })") + .ValueOrDie(); + TF_ASSERT_OK(MakeFakeArguments(module.get()).status()); +} + } // namespace } // namespace xla diff --git a/tensorflow/compiler/xla/tests/token_hlo_test.cc b/tensorflow/compiler/xla/tests/token_hlo_test.cc index 8541698576fb8aae1e3528cb618b367f843b8d53..2bdbd08309a81b201fc224110805549f7fb5bb55 100644 --- a/tensorflow/compiler/xla/tests/token_hlo_test.cc +++ b/tensorflow/compiler/xla/tests/token_hlo_test.cc @@ -31,27 +31,29 @@ class TokenHloTest : public HloTestBase {}; XLA_TEST_F(TokenHloTest, SingleTokenInstruction) { std::unique_ptr module = CreateNewModule(); auto builder = HloComputation::Builder(TestName()); - builder.AddInstruction(HloInstruction::CreateGenerateToken({})); - builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(42))); + builder.AddInstruction(HloInstruction::CreateToken()); module->AddEntryComputation(builder.Build()); - EXPECT_IS_OK(HloVerifier().Run(module.get()).status()); + + TF_ASSERT_OK_AND_ASSIGN(std::unique_ptr result, + Execute(std::move(module), {})); + EXPECT_TRUE(LiteralTestUtil::Equal(*result, *LiteralUtil::CreateToken())); } XLA_TEST_F(TokenHloTest, TokenTree) { std::unique_ptr module = CreateNewModule(); auto builder = HloComputation::Builder(TestName()); - auto token0 = builder.AddInstruction(HloInstruction::CreateGenerateToken({})); - auto token1 = builder.AddInstruction(HloInstruction::CreateGenerateToken({})); - auto token2 = builder.AddInstruction(HloInstruction::CreateGenerateToken({})); + auto token0 = builder.AddInstruction(HloInstruction::CreateToken()); + auto token1 = builder.AddInstruction(HloInstruction::CreateToken()); + auto token2 = builder.AddInstruction(HloInstruction::CreateToken()); builder.AddInstruction( - HloInstruction::CreateGenerateToken({token0, token0, token1, token2})); - builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(42))); + HloInstruction::CreateAfterAll({token0, token0, token1, token2})); module->AddEntryComputation(builder.Build()); - EXPECT_IS_OK(HloVerifier().Run(module.get()).status()); + + TF_ASSERT_OK_AND_ASSIGN(std::unique_ptr result, + Execute(std::move(module), {})); + EXPECT_TRUE(LiteralTestUtil::Equal(*result, *LiteralUtil::CreateToken())); } XLA_TEST_F(TokenHloTest, InvalidTokenShapedEntryParameter) { @@ -62,7 +64,7 @@ XLA_TEST_F(TokenHloTest, InvalidTokenShapedEntryParameter) { builder.AddInstruction( HloInstruction::CreateParameter(1, ShapeUtil::MakeTokenShape(), "p1")); builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(42))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(42))); module->AddEntryComputation(builder.Build()); Status status = HloVerifier().Run(module.get()).status(); @@ -89,26 +91,14 @@ XLA_TEST_F(TokenHloTest, InvalidTupleTokenShapedEntryParameter) { ::testing::HasSubstr("Entry parameter 0 is or contains a token shape")); } -XLA_TEST_F(TokenHloTest, InvalidTokenRoot) { - std::unique_ptr module = CreateNewModule(); - auto builder = HloComputation::Builder(TestName()); - builder.AddInstruction(HloInstruction::CreateGenerateToken({})); - module->AddEntryComputation(builder.Build()); - - Status status = HloVerifier().Run(module.get()).status(); - ASSERT_IS_NOT_OK(status); - EXPECT_THAT(status.error_message(), - ::testing::HasSubstr("Entry root is or contains a token shape")); -} - XLA_TEST_F(TokenHloTest, InvalidOperandToTokenInstruction) { std::unique_ptr module = CreateNewModule(); auto builder = HloComputation::Builder(TestName()); auto param = builder.AddInstruction( HloInstruction::CreateParameter(0, ShapeUtil::MakeShape(F32, {}), "p0")); - builder.AddInstruction(HloInstruction::CreateGenerateToken({param})); + builder.AddInstruction(HloInstruction::CreateAfterAll({param})); builder.AddInstruction( - HloInstruction::CreateConstant(Literal::CreateR0(123))); + HloInstruction::CreateConstant(LiteralUtil::CreateR0(123))); module->AddEntryComputation(builder.Build()); Status status = HloVerifier().Run(module.get()).status(); @@ -120,7 +110,7 @@ XLA_TEST_F(TokenHloTest, InvalidOperandToTokenInstruction) { XLA_TEST_F(TokenHloTest, TokenInWhileLoop) { // Thread a token around a while loop. Token is created and consumed by a - // GenerateToken instruction in the while body. + // AfterAll instruction in the while body. string module_string = R"( HloModule TokenInWhileLoop @@ -130,8 +120,8 @@ HloModule TokenInWhileLoop %constant.1 = s32[] constant(1) %add = s32[] add(s32[] %get-tuple-element.1, s32[] %constant.1) %get-tuple-element.2 = token[] get-tuple-element((s32[], token[]) %param.1), index=1 - %generate-token = token[] generate-token(token[] %get-tuple-element.2) - ROOT %tuple = (s32[], token[]) tuple(s32[] %add, token[] %generate-token) + %after-all = token[] after-all(token[] %get-tuple-element.2) + ROOT %tuple = (s32[], token[]) tuple(s32[] %add, token[] %after-all) } %Cond (param: (s32[], token[])) -> pred[] { @@ -143,7 +133,7 @@ HloModule TokenInWhileLoop ENTRY %TokenInWhileLoop () -> s32[] { %zero = s32[] constant(0) - %init_token = token[] generate-token() + %init_token = token[] after-all() %init_tuple = (s32[], token[]) tuple(s32[] %zero, token[] %init_token) %while = (s32[], token[]) while((s32[], token[]) %init_tuple), condition=%Cond, body=%Body ROOT %root = s32[] get-tuple-element((s32[], token[]) %while), index=0 @@ -172,13 +162,13 @@ HloModule TokenInConditional %False (param.2: s32[]) -> (s32[], token[]) { %param.2 = s32[] parameter(0) - %new_token = token[] generate-token() + %new_token = token[] after-all() ROOT %tuple = (s32[], token[]) tuple(s32[] %param.2, token[] %new_token) } ENTRY %TokenInConditional (param.3: pred[]) -> s32[] { %param.3 = pred[] parameter(0) - %init_token = token[] generate-token() + %init_token = token[] after-all() %seven = s32[] constant(7) %cond = (s32[], token[]) conditional(pred[] %param.3, token[] %init_token, s32[] %seven), true_computation=True, false_computation=False ROOT %root = s32[] get-tuple-element((s32[], token[]) %cond), index=0 @@ -194,7 +184,7 @@ ENTRY %TokenInConditional (param.3: pred[]) -> s32[] { TF_ASSERT_OK_AND_ASSIGN( std::unique_ptr module, HloRunner::CreateModuleFromString(module_string, debug_options)); - auto arg = Literal::CreateR0(true); + auto arg = LiteralUtil::CreateR0(true); TF_ASSERT_OK_AND_ASSIGN(std::unique_ptr result, Execute(std::move(module), {arg.get()})); EXPECT_EQ(42, result->Get({})); @@ -205,7 +195,7 @@ ENTRY %TokenInConditional (param.3: pred[]) -> s32[] { TF_ASSERT_OK_AND_ASSIGN( std::unique_ptr module, HloRunner::CreateModuleFromString(module_string, debug_options)); - auto arg = Literal::CreateR0(false); + auto arg = LiteralUtil::CreateR0(false); TF_ASSERT_OK_AND_ASSIGN(std::unique_ptr result, Execute(std::move(module), {arg.get()})); EXPECT_EQ(7, result->Get({})); diff --git a/tensorflow/compiler/xla/tests/transfer_manager_test.cc b/tensorflow/compiler/xla/tests/transfer_manager_test.cc index 85799d4cfb4838d91bd51c8d24d7ca70b41e6df1..0f86b7f20f9bd7597ece713626ee0e9c23509e05 100644 --- a/tensorflow/compiler/xla/tests/transfer_manager_test.cc +++ b/tensorflow/compiler/xla/tests/transfer_manager_test.cc @@ -18,7 +18,7 @@ limitations under the License. #include #include "tensorflow/compiler/xla/layout_util.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/service/device_memory_allocator.h" #include "tensorflow/compiler/xla/service/generic_transfer_manager.h" #include "tensorflow/compiler/xla/service/shaped_buffer.h" @@ -68,7 +68,7 @@ class TransferManagerTest : public LocalClientTestBase { }; XLA_TEST_F(TransferManagerTest, TransferR0U32) { - std::unique_ptr literal = Literal::CreateR0(42); + std::unique_ptr literal = LiteralUtil::CreateR0(42); const Shape& shape = literal->shape(); auto device_buffer = AllocateDeviceBuffer(shape); @@ -84,7 +84,7 @@ XLA_TEST_F(TransferManagerTest, TransferR0U32) { XLA_TEST_F(TransferManagerTest, TransferR1F32) { std::unique_ptr literal = - Literal::CreateR1({1.25f, 2.5f, -17.0f, -20.125f}); + LiteralUtil::CreateR1({1.25f, 2.5f, -17.0f, -20.125f}); const Shape& shape = literal->shape(); auto device_buffer = AllocateDeviceBuffer(shape); @@ -102,7 +102,7 @@ XLA_TEST_F(TransferManagerTest, TransferR1F32) { XLA_TEST_F(TransferManagerTest, TransferR1LargeF32) { std::vector test_vector(1024 * 1024); std::iota(test_vector.begin(), test_vector.end(), 0); - std::unique_ptr literal = Literal::CreateR1(test_vector); + std::unique_ptr literal = LiteralUtil::CreateR1(test_vector); const Shape& shape = literal->shape(); auto device_buffer = AllocateDeviceBuffer(shape); @@ -118,7 +118,7 @@ XLA_TEST_F(TransferManagerTest, TransferR1LargeF32) { XLA_TEST_F(TransferManagerTest, TransferR1U8) { const char* test_string = "0123456789abcdef"; - std::unique_ptr literal = Literal::CreateR1U8(test_string); + std::unique_ptr literal = LiteralUtil::CreateR1U8(test_string); const Shape& shape = literal->shape(); auto device_buffer = AllocateDeviceBuffer(shape); @@ -134,7 +134,7 @@ XLA_TEST_F(TransferManagerTest, TransferR1U8) { XLA_TEST_F(TransferManagerTest, TransferR2F32) { std::unique_ptr literal = - Literal::CreateR2({{1.0f, 2.0f, 3.0f}, {4.0f, 5.0f, 6.0f}}); + LiteralUtil::CreateR2({{1.0f, 2.0f, 3.0f}, {4.0f, 5.0f, 6.0f}}); const Shape& shape = literal->shape(); auto device_buffer = AllocateDeviceBuffer(shape); @@ -151,7 +151,7 @@ XLA_TEST_F(TransferManagerTest, TransferR2F32) { XLA_TEST_F(TransferManagerTest, TransferR2F32AndChangeLayoutTransferringToDevice) { - std::unique_ptr literal = Literal::CreateR2WithLayout( + std::unique_ptr literal = LiteralUtil::CreateR2WithLayout( {{1.0f, 2.0f, 3.0f}, {4.0f, 5.0f, 6.0f}}, LayoutUtil::MakeLayout({0, 1})); const Shape ondevice_shape = ShapeUtil::MakeShapeWithLayout(F32, {2, 3}, {1, 0}); @@ -172,10 +172,10 @@ XLA_TEST_F(TransferManagerTest, } XLA_TEST_F(TransferManagerTest, TransferTuple) { - std::unique_ptr literal = Literal::MakeTuple( - {Literal::CreateR0(123.0f).get(), - Literal::CreateR2({{1.0f, 2.0f}, {4.0f, 5.0f}}).get(), - Literal::CreateR1({44.0f, -10.0f, 3333333.3f}).get()}); + std::unique_ptr literal = LiteralUtil::MakeTuple( + {LiteralUtil::CreateR0(123.0f).get(), + LiteralUtil::CreateR2({{1.0f, 2.0f}, {4.0f, 5.0f}}).get(), + LiteralUtil::CreateR1({44.0f, -10.0f, 3333333.3f}).get()}); auto device_buffer = AllocateDeviceBuffer(literal->shape()); // Round trip literal through device. @@ -189,7 +189,7 @@ XLA_TEST_F(TransferManagerTest, TransferTuple) { } XLA_TEST_F(TransferManagerTest, TransferEmptyTuple) { - std::unique_ptr literal = Literal::MakeTuple({}); + std::unique_ptr literal = LiteralUtil::MakeTuple({}); auto device_buffer = AllocateDeviceBuffer(literal->shape()); // Round trip literal through device. @@ -203,13 +203,13 @@ XLA_TEST_F(TransferManagerTest, TransferEmptyTuple) { } XLA_TEST_F(TransferManagerTest, TransferNestedTuple) { - std::unique_ptr literal = Literal::MakeTuple( - {Literal::CreateR0(123.0f).get(), - Literal::MakeTuple( - {Literal::CreateR2({{1.0f, 2.0f}, {4.0f, 5.0f}}).get(), - Literal::CreateR1({44.0f, -10.0f, 3333333.3f}).get()}) + std::unique_ptr literal = LiteralUtil::MakeTuple( + {LiteralUtil::CreateR0(123.0f).get(), + LiteralUtil::MakeTuple( + {LiteralUtil::CreateR2({{1.0f, 2.0f}, {4.0f, 5.0f}}).get(), + LiteralUtil::CreateR1({44.0f, -10.0f, 3333333.3f}).get()}) .get(), - Literal::CreateR1({-10.0f, 123.0f}).get()}); + LiteralUtil::CreateR1({-10.0f, 123.0f}).get()}); auto device_buffer = AllocateDeviceBuffer(literal->shape()); // Round trip literal through device. @@ -223,7 +223,7 @@ XLA_TEST_F(TransferManagerTest, TransferNestedTuple) { } XLA_TEST_F(TransferManagerTest, TransferComplexValue) { - std::unique_ptr literal = Literal::CreateR1( + std::unique_ptr literal = LiteralUtil::CreateR1( {complex64(1.0f, 2.0f), complex64(42.0f, -123.4f)}); auto device_buffer = AllocateDeviceBuffer(literal->shape()); @@ -238,12 +238,12 @@ XLA_TEST_F(TransferManagerTest, TransferComplexValue) { } XLA_TEST_F(TransferManagerTest, TransferComplexValueInTuple) { - std::unique_ptr literal = Literal::MakeTuple( - {Literal::CreateR1( + std::unique_ptr literal = LiteralUtil::MakeTuple( + {LiteralUtil::CreateR1( {complex64(1.0f, 2.0f), complex64(42.0f, -123.4f)}) .get(), - Literal::CreateR1({1, 2, 3, 4, 5, 6}).get(), - Literal::CreateR0(complex64(0.3f, -0.4f)).get()}); + LiteralUtil::CreateR1({1, 2, 3, 4, 5, 6}).get(), + LiteralUtil::CreateR0(complex64(0.3f, -0.4f)).get()}); auto device_buffer = AllocateDeviceBuffer(literal->shape()); // Round trip literal through device. @@ -256,22 +256,34 @@ XLA_TEST_F(TransferManagerTest, TransferComplexValueInTuple) { EXPECT_TRUE(LiteralTestUtil::Equal(*literal, *result)); } +XLA_TEST_F(TransferManagerTest, TransferTokenFromDevice) { + // "Copy" a token from the device. The token has no physical representation so + // no copying is actually performed, but it shouldn't fail. + // TODO(b/110532604): Add transferring the token to device when this is + // supported. + auto device_buffer = AllocateDeviceBuffer(ShapeUtil::MakeTokenShape()); + TF_ASSERT_OK_AND_ASSIGN( + std::unique_ptr result, + transfer_manager_->TransferLiteralFromDevice(stream_, device_buffer)); + EXPECT_TRUE(LiteralTestUtil::Equal(*LiteralUtil::CreateToken(), *result)); +} + XLA_TEST_F(TransferManagerTest, MultiStreamRoundTripSoak) { const int64 kIterationCount = 5000; - std::unique_ptr literal1 = Literal::MakeTuple( - {Literal::CreateR0(123.0f).get(), - Literal::MakeTuple( - {Literal::CreateR2({{1.0f, 2.0f}, {4.0f, 5.0f}}).get(), - Literal::CreateR1({44.0f, -10.0f, 3333333.3f}).get()}) + std::unique_ptr literal1 = LiteralUtil::MakeTuple( + {LiteralUtil::CreateR0(123.0f).get(), + LiteralUtil::MakeTuple( + {LiteralUtil::CreateR2({{1.0f, 2.0f}, {4.0f, 5.0f}}).get(), + LiteralUtil::CreateR1({44.0f, -10.0f, 3333333.3f}).get()}) .get(), - Literal::CreateR1({-10.0f, 123.0f}).get()}); - std::unique_ptr literal2 = Literal::MakeTuple( - {Literal::CreateR0(456.0f).get(), - Literal::MakeTuple( - {Literal::CreateR2({{5.0f, 7.0f}, {9.0f, 4.0f}}).get(), - Literal::CreateR1({44.0f, -11.0f, 3333333.3f}).get()}) + LiteralUtil::CreateR1({-10.0f, 123.0f}).get()}); + std::unique_ptr literal2 = LiteralUtil::MakeTuple( + {LiteralUtil::CreateR0(456.0f).get(), + LiteralUtil::MakeTuple( + {LiteralUtil::CreateR2({{5.0f, 7.0f}, {9.0f, 4.0f}}).get(), + LiteralUtil::CreateR1({44.0f, -11.0f, 3333333.3f}).get()}) .get(), - Literal::CreateR1({-98.0f, 153.0f}).get()}); + LiteralUtil::CreateR1({-98.0f, 153.0f}).get()}); auto device_buffer1 = AllocateDeviceBuffer(literal1->shape()); auto device_buffer2 = AllocateDeviceBuffer(literal2->shape()); @@ -313,10 +325,10 @@ class TransferDeviceToHostBenchmark : public TransferManagerTest { std::vector> tuple_elements; for (int i = 0; i < num_tuple_elements; ++i) { tuple_elements.push_back( - Literal::CreateR2F32Linspace(0.0f, 1.0f, array_size, array_size)); + LiteralUtil::CreateR2F32Linspace(0.0f, 1.0f, array_size, array_size)); } std::unique_ptr literal = - Literal::MakeTupleOwned(std::move(tuple_elements)); + LiteralUtil::MakeTupleOwned(std::move(tuple_elements)); auto device_buffer = AllocateDeviceBuffer(literal->shape()); TF_CHECK_OK(transfer_manager_->TransferLiteralToDevice(stream_, *literal, device_buffer)); @@ -345,10 +357,10 @@ class TransferHostToDeviceBenchmark : public TransferManagerTest { std::vector> tuple_elements; for (int i = 0; i < num_tuple_elements; ++i) { tuple_elements.push_back( - Literal::CreateR2F32Linspace(0.0f, 1.0f, array_size, array_size)); + LiteralUtil::CreateR2F32Linspace(0.0f, 1.0f, array_size, array_size)); } std::unique_ptr literal = - Literal::MakeTupleOwned(std::move(tuple_elements)); + LiteralUtil::MakeTupleOwned(std::move(tuple_elements)); auto device_buffer = AllocateDeviceBuffer(literal->shape()); tensorflow::testing::StartTiming(); for (int i = 0; i < iters; ++i) { diff --git a/tensorflow/compiler/xla/tests/transpose_test.cc b/tensorflow/compiler/xla/tests/transpose_test.cc index fe1e3da7eca00e128377e6e56af877868aafa836..6ebb4324f8d20ed9f8886d92b0513441685ed19b 100644 --- a/tensorflow/compiler/xla/tests/transpose_test.cc +++ b/tensorflow/compiler/xla/tests/transpose_test.cc @@ -38,34 +38,35 @@ class TransposeTest : public ClientLibraryTestBase { XLA_TEST_F(TransposeTest, Transpose0x0) { XlaBuilder builder("Transpose"); - auto lhs = builder.ConstantR2FromArray2D(Array2D(0, 0)); - auto result = builder.Transpose(lhs, {1, 0}); + auto lhs = ConstantR2FromArray2D(&builder, Array2D(0, 0)); + Transpose(lhs, {1, 0}); ComputeAndCompareR2(&builder, Array2D(0, 0), {}, error_spec_); } XLA_TEST_F(TransposeTest, Transpose0x42) { XlaBuilder builder("Transpose"); - auto lhs = builder.ConstantR2FromArray2D(Array2D(0, 42)); - auto result = builder.Transpose(lhs, {1, 0}); + auto lhs = ConstantR2FromArray2D(&builder, Array2D(0, 42)); + Transpose(lhs, {1, 0}); ComputeAndCompareR2(&builder, Array2D(42, 0), {}, error_spec_); } XLA_TEST_F(TransposeTest, Transpose7x0) { XlaBuilder builder("Transpose"); - auto lhs = builder.ConstantR2FromArray2D(Array2D(7, 0)); - auto result = builder.Transpose(lhs, {1, 0}); + auto lhs = ConstantR2FromArray2D(&builder, Array2D(7, 0)); + Transpose(lhs, {1, 0}); ComputeAndCompareR2(&builder, Array2D(0, 7), {}, error_spec_); } TEST_F(TransposeTest, Transpose2x2) { XlaBuilder builder("Transpose"); - auto lhs = builder.ConstantR2({ - {1.0, 2.0}, {3.0, 4.0}, - }); - auto result = builder.Transpose(lhs, {1, 0}); + auto lhs = ConstantR2(&builder, { + {1.0, 2.0}, + {3.0, 4.0}, + }); + Transpose(lhs, {1, 0}); Array2D expected({{1.0f, 3.0f}, {2.0f, 4.0f}}); @@ -74,16 +75,18 @@ TEST_F(TransposeTest, Transpose2x2) { XLA_TEST_F(TransposeTest, Transpose0x2x3_2x3x0) { XlaBuilder builder("Transpose"); - auto operand = builder.ConstantR3FromArray3D(Array3D(0, 2, 3)); - auto result = builder.Transpose(operand, {1, 2, 0}); + auto operand = + ConstantR3FromArray3D(&builder, Array3D(0, 2, 3)); + Transpose(operand, {1, 2, 0}); ComputeAndCompareR3(&builder, Array3D(2, 3, 0), {}); } TEST_F(TransposeTest, Transpose1x2x3_2x3x1) { XlaBuilder builder("Transpose"); - auto operand = builder.ConstantR3FromArray3D({{{1, 2, 3}, {4, 5, 6}}}); - auto result = builder.Transpose(operand, {1, 2, 0}); + auto operand = + ConstantR3FromArray3D(&builder, {{{1, 2, 3}, {4, 5, 6}}}); + Transpose(operand, {1, 2, 0}); Array3D expected({{{1}, {2}, {3}}, {{4}, {5}, {6}}}); @@ -92,8 +95,9 @@ TEST_F(TransposeTest, Transpose1x2x3_2x3x1) { TEST_F(TransposeTest, Transpose1x2x3_3x2x1) { XlaBuilder builder("Transpose"); - auto operand = builder.ConstantR3FromArray3D({{{1, 2, 3}, {4, 5, 6}}}); - auto result = builder.Transpose(operand, {2, 1, 0}); + auto operand = + ConstantR3FromArray3D(&builder, {{{1, 2, 3}, {4, 5, 6}}}); + Transpose(operand, {2, 1, 0}); Array3D expected({{{1}, {4}}, {{2}, {5}}, {{3}, {6}}}); @@ -102,8 +106,9 @@ TEST_F(TransposeTest, Transpose1x2x3_3x2x1) { TEST_F(TransposeTest, Transpose1x2x3_1x2x3) { XlaBuilder builder("Transpose"); - auto operand = builder.ConstantR3FromArray3D({{{1, 2, 3}, {4, 5, 6}}}); - auto result = builder.Transpose(operand, {0, 1, 2}); + auto operand = + ConstantR3FromArray3D(&builder, {{{1, 2, 3}, {4, 5, 6}}}); + Transpose(operand, {0, 1, 2}); Array3D expected({{{1, 2, 3}, {4, 5, 6}}}); @@ -116,9 +121,9 @@ TEST_F(TransposeTest, MultiTranspose3x2) { for (int transposes = 0; transposes <= 10; ++transposes) { XlaBuilder builder("Transpose"); - auto computed = builder.ConstantR2FromArray2D(input); + auto computed = ConstantR2FromArray2D(&builder, input); for (int i = 0; i < transposes; ++i) { - computed = builder.Transpose(computed, {1, 0}); + computed = Transpose(computed, {1, 0}); } const Array2D& expected = transposes % 2 == 0 ? input : transposed; ComputeAndCompareR2(&builder, expected, {}, error_spec_); @@ -130,8 +135,8 @@ TEST_F(TransposeTest, Small_1x1) { auto aoperand = MakeLinspaceArray2D(0.0, 1.0, 1, 1); XlaBuilder builder("transpose_1x1"); - auto operand = builder.ConstantR2FromArray2D(*aoperand); - builder.Transpose(operand, {1, 0}); + auto operand = ConstantR2FromArray2D(&builder, *aoperand); + Transpose(operand, {1, 0}); auto expected = ReferenceUtil::TransposeArray2D(*aoperand); ComputeAndCompareR2(&builder, *expected, {}, ErrorSpec(1e-4)); @@ -142,8 +147,8 @@ TEST_F(TransposeTest, Small_2x2) { auto aoperand = MakeLinspaceArray2D(0.0, 4.0, 2, 2); XlaBuilder builder("transpose_2x2"); - auto operand = builder.ConstantR2FromArray2D(*aoperand); - builder.Transpose(operand, {1, 0}); + auto operand = ConstantR2FromArray2D(&builder, *aoperand); + Transpose(operand, {1, 0}); auto expected = ReferenceUtil::TransposeArray2D(*aoperand); ComputeAndCompareR2(&builder, *expected, {}, ErrorSpec(1e-4)); @@ -162,8 +167,8 @@ void TransposeTest::TestTransposeConstant021(size_t n1, size_t n2, size_t n3) { } XlaBuilder builder(TestName()); - auto operand = builder.ConstantR3FromArray3D(aoperand); - builder.Transpose(operand, {0, 2, 1}); + auto operand = ConstantR3FromArray3D(&builder, aoperand); + Transpose(operand, {0, 2, 1}); ComputeAndCompareR3(&builder, expected, {}); } diff --git a/tensorflow/compiler/xla/tests/tuple_test.cc b/tensorflow/compiler/xla/tests/tuple_test.cc index 220d9f6320632cae2c51f71cc7c568120bda1f04..bf86c5dfb6a1caf2c8574a5a4f7b77d982039bde 100644 --- a/tensorflow/compiler/xla/tests/tuple_test.cc +++ b/tensorflow/compiler/xla/tests/tuple_test.cc @@ -49,12 +49,12 @@ XLA_TEST_F(TupleTest, TupleConstant) { {1.1f, 2.2f, 3.5f}, // row 0 {4.8f, 5.0f, 6.7f}, // row 1 }; - auto value = - Literal::MakeTuple({Literal::CreateR0(constant_scalar).get(), - Literal::CreateR1(constant_vector).get(), - Literal::CreateR2(constant_matrix).get()}); + auto value = LiteralUtil::MakeTuple( + {LiteralUtil::CreateR0(constant_scalar).get(), + LiteralUtil::CreateR1(constant_vector).get(), + LiteralUtil::CreateR2(constant_matrix).get()}); - builder.ConstantLiteral(*value); + ConstantLiteral(&builder, *value); ComputeAndCompareTuple(&builder, *value, {}, error_spec_); } @@ -64,11 +64,11 @@ XLA_TEST_F(TupleTest, TupleScalarConstant) { const float constant_scalar1 = 7.3f; const float constant_scalar2 = 1.2f; - auto value = - Literal::MakeTuple({Literal::CreateR0(constant_scalar1).get(), - Literal::CreateR0(constant_scalar2).get()}); + auto value = LiteralUtil::MakeTuple( + {LiteralUtil::CreateR0(constant_scalar1).get(), + LiteralUtil::CreateR0(constant_scalar2).get()}); - builder.ConstantLiteral(*value); + ConstantLiteral(&builder, *value); ComputeAndCompareTuple(&builder, *value, {}, error_spec_); } @@ -82,14 +82,14 @@ XLA_TEST_F(TupleTest, TupleCreate) { {1.1f, 2.2f, 3.5f}, // row 0 {4.8f, 5.0f, 6.7f}, // row 1 }; - builder.Tuple({builder.ConstantR0(constant_scalar), - builder.ConstantR1(constant_vector), - builder.ConstantR2(constant_matrix)}); - - auto expected = - Literal::MakeTuple({Literal::CreateR0(constant_scalar).get(), - Literal::CreateR1(constant_vector).get(), - Literal::CreateR2(constant_matrix).get()}); + Tuple(&builder, {ConstantR0(&builder, constant_scalar), + ConstantR1(&builder, constant_vector), + ConstantR2(&builder, constant_matrix)}); + + auto expected = LiteralUtil::MakeTuple( + {LiteralUtil::CreateR0(constant_scalar).get(), + LiteralUtil::CreateR1(constant_vector).get(), + LiteralUtil::CreateR2(constant_matrix).get()}); ComputeAndCompareTuple(&builder, *expected, {}, error_spec_); } @@ -97,19 +97,20 @@ XLA_TEST_F(TupleTest, TupleCreate) { XLA_TEST_F(TupleTest, TupleCreateWithZeroElementEntry) { XlaBuilder builder(TestName()); - builder.Tuple( - {builder.ConstantR0(7.0), builder.ConstantR1({})}); + Tuple(&builder, + {ConstantR0(&builder, 7.0), ConstantR1(&builder, {})}); - auto expected = Literal::MakeTuple({Literal::CreateR0(7.0).get(), - Literal::CreateR1({}).get()}); + auto expected = + LiteralUtil::MakeTuple({LiteralUtil::CreateR0(7.0).get(), + LiteralUtil::CreateR1({}).get()}); ComputeAndCompareTuple(&builder, *expected, {}, error_spec_); } // Tests the creation of an empty tuple. XLA_TEST_F(TupleTest, EmptyTupleCreate) { XlaBuilder builder(TestName()); - builder.Tuple({}); - auto expected = Literal::MakeTuple({}); + Tuple(&builder, {}); + auto expected = LiteralUtil::MakeTuple({}); ComputeAndCompareTuple(&builder, *expected, {}, error_spec_); } @@ -121,9 +122,10 @@ XLA_TEST_F(TupleTest, GetTupleElement) { {1.f, 2.f, 3.f}, // row 0 {4.f, 5.f, 6.f}, // row 1 }; - auto tuple_data = builder.Tuple({builder.ConstantR1(constant_vector), - builder.ConstantR2(constant_matrix)}); - builder.GetTupleElement(tuple_data, 1); + auto tuple_data = + Tuple(&builder, {ConstantR1(&builder, constant_vector), + ConstantR2(&builder, constant_matrix)}); + GetTupleElement(tuple_data, 1); ComputeAndCompareR2(&builder, Array2D(constant_matrix), {}, error_spec_); } @@ -131,17 +133,18 @@ XLA_TEST_F(TupleTest, GetTupleElement) { // Trivial test for extracting a tuple element with GetTupleElement. XLA_TEST_F(TupleTest, GetTupleElementWithZeroElements) { XlaBuilder builder(TestName()); - auto tuple_data = builder.Tuple( - {builder.ConstantR1({}), - builder.ConstantR2FromArray2D(Array2D(0, 101))}); - builder.GetTupleElement(tuple_data, 1); + auto tuple_data = + Tuple(&builder, + {ConstantR1(&builder, {}), + ConstantR2FromArray2D(&builder, Array2D(0, 101))}); + GetTupleElement(tuple_data, 1); ComputeAndCompareR2(&builder, Array2D(0, 101), {}, error_spec_); } XLA_TEST_F(TupleTest, GetTupleElementOfNonTupleFailsGracefully) { XlaBuilder builder(TestName()); - auto value = builder.ConstantR1({4.5f}); - builder.GetTupleElement(value, 1); + auto value = ConstantR1(&builder, {4.5f}); + GetTupleElement(value, 1); auto result_status = builder.Build(); EXPECT_FALSE(result_status.ok()); EXPECT_THAT( @@ -158,14 +161,15 @@ XLA_TEST_F(TupleTest, AddTupleElements) { {1.f, 2.f, 3.f}, // row 0 {4.f, 5.f, 6.f}, // row 1 }; - auto tuple_data = builder.Tuple({builder.ConstantR1(constant_vector), - builder.ConstantR2(constant_matrix)}); - auto vector_element = builder.GetTupleElement(tuple_data, 0); - auto matrix_element = builder.GetTupleElement(tuple_data, 1); + auto tuple_data = + Tuple(&builder, {ConstantR1(&builder, constant_vector), + ConstantR2(&builder, constant_matrix)}); + auto vector_element = GetTupleElement(tuple_data, 0); + auto matrix_element = GetTupleElement(tuple_data, 1); auto vector_shape = builder.GetShape(vector_element).ConsumeValueOrDie(); auto matrix_shape = builder.GetShape(matrix_element).ConsumeValueOrDie(); - builder.Add(matrix_element, vector_element, - /*broadcast_dimensions=*/{1}); + Add(matrix_element, vector_element, + /*broadcast_dimensions=*/{1}); Array2D expected({ {2.f, 4.f, 6.f}, // row 0 @@ -185,13 +189,14 @@ XLA_TEST_F(TupleTest, TupleGTEToTuple) { {1.f, 2.f, 3.f}, // row 0 {4.f, 5.f, 6.f}, // row 1 }; - auto tuple_data = builder.Tuple({builder.ConstantR1(constant_vector), - builder.ConstantR2(constant_matrix)}); - builder.Tuple({builder.GetTupleElement(tuple_data, 1), - builder.GetTupleElement(tuple_data, 0)}); - auto expected = - Literal::MakeTuple({Literal::CreateR2(constant_matrix).get(), - Literal::CreateR1(constant_vector).get()}); + auto tuple_data = + Tuple(&builder, {ConstantR1(&builder, constant_vector), + ConstantR2(&builder, constant_matrix)}); + Tuple(&builder, + {GetTupleElement(tuple_data, 1), GetTupleElement(tuple_data, 0)}); + auto expected = LiteralUtil::MakeTuple( + {LiteralUtil::CreateR2(constant_matrix).get(), + LiteralUtil::CreateR1(constant_vector).get()}); ComputeAndCompareTuple(&builder, *expected, {}, error_spec_); } @@ -206,14 +211,14 @@ XLA_TEST_F(TupleTest, SelectBetweenPredTuples) { std::unique_ptr v2_data = CreateR0Parameter(1.0f, /*parameter_number=*/1, /*name=*/"v2", /*builder=*/&b, /*data_handle=*/&v2); - auto v1_gt = b.Gt(v1, v2); // false - auto v2_gt = b.Gt(v2, v1); // true - auto v1_v2 = b.Tuple({v1_gt, v2_gt}); // {false, true} - auto v2_v1 = b.Tuple({v2_gt, v1_gt}); // {true, false} - b.Select(direction ? v1_gt : v2_gt, v1_v2, v2_v1); + auto v1_gt = Gt(v1, v2); // false + auto v2_gt = Gt(v2, v1); // true + auto v1_v2 = Tuple(&b, {v1_gt, v2_gt}); // {false, true} + auto v2_v1 = Tuple(&b, {v2_gt, v1_gt}); // {true, false} + Select(direction ? v1_gt : v2_gt, v1_v2, v2_v1); auto expected = - Literal::MakeTuple({Literal::CreateR0(direction).get(), - Literal::CreateR0(!direction).get()}); + LiteralUtil::MakeTuple({LiteralUtil::CreateR0(direction).get(), + LiteralUtil::CreateR0(!direction).get()}); ComputeAndCompareTuple(&b, *expected, {v1_data.get(), v2_data.get()}, error_spec_); @@ -243,22 +248,23 @@ XLA_TEST_F(TupleTest, TupleGTEToTupleToGTEAdd) { {1.f, 2.f, 3.f}, // row 0 {4.f, 5.f, 6.f}, // row 1 }; - auto tuple_data = builder.Tuple({builder.ConstantR1(constant_vector), - builder.ConstantR2(constant_matrix)}); - auto new_tuple01 = builder.Tuple({builder.GetTupleElement(tuple_data, 0), - builder.GetTupleElement(tuple_data, 1)}); - auto new_tuple10 = builder.Tuple({builder.GetTupleElement(tuple_data, 1), - builder.GetTupleElement(tuple_data, 0)}); - auto vector_from_01 = builder.GetTupleElement(new_tuple01, 0); - auto vector_from_10 = builder.GetTupleElement(new_tuple10, 1); - auto matrix_from_01 = builder.GetTupleElement(new_tuple01, 1); - auto matrix_from_10 = builder.GetTupleElement(new_tuple10, 0); - - auto addvectors = builder.Add(vector_from_01, vector_from_10); - auto addmatrices = builder.Add(matrix_from_01, matrix_from_10); - - builder.Add(addmatrices, addvectors, - /*broadcast_dimensions=*/{1}); + auto tuple_data = + Tuple(&builder, {ConstantR1(&builder, constant_vector), + ConstantR2(&builder, constant_matrix)}); + auto new_tuple01 = Tuple(&builder, {GetTupleElement(tuple_data, 0), + GetTupleElement(tuple_data, 1)}); + auto new_tuple10 = Tuple(&builder, {GetTupleElement(tuple_data, 1), + GetTupleElement(tuple_data, 0)}); + auto vector_from_01 = GetTupleElement(new_tuple01, 0); + auto vector_from_10 = GetTupleElement(new_tuple10, 1); + auto matrix_from_01 = GetTupleElement(new_tuple01, 1); + auto matrix_from_10 = GetTupleElement(new_tuple10, 0); + + auto addvectors = Add(vector_from_01, vector_from_10); + auto addmatrices = Add(matrix_from_01, matrix_from_10); + + Add(addmatrices, addvectors, + /*broadcast_dimensions=*/{1}); Array2D expected({ {4.f, 8.f, 12.f}, // row 0 @@ -273,14 +279,15 @@ XLA_TEST_F(TupleTest, SelectBetweenTuplesOnFalse) { std::initializer_list vec1 = {1.f, 2.f, 3.f}; std::initializer_list vec2 = {2.f, 4.f, 6.f}; - auto tuple12 = builder.Tuple( - {builder.ConstantR1(vec1), builder.ConstantR1(vec2)}); - auto tuple21 = builder.Tuple( - {builder.ConstantR1(vec2), builder.ConstantR1(vec1)}); - - builder.Select(builder.ConstantR0(false), tuple12, tuple21); - auto expected = Literal::MakeTuple({Literal::CreateR1(vec2).get(), - Literal::CreateR1(vec1).get()}); + auto tuple12 = Tuple(&builder, {ConstantR1(&builder, vec1), + ConstantR1(&builder, vec2)}); + auto tuple21 = Tuple(&builder, {ConstantR1(&builder, vec2), + ConstantR1(&builder, vec1)}); + + Select(ConstantR0(&builder, false), tuple12, tuple21); + auto expected = + LiteralUtil::MakeTuple({LiteralUtil::CreateR1(vec2).get(), + LiteralUtil::CreateR1(vec1).get()}); ComputeAndCompareTuple(&builder, *expected, {}, error_spec_); } @@ -292,22 +299,22 @@ XLA_TEST_F(TupleTest, TuplesInAMap) { // Need to put a select in there to prevent HLO-level optimizations from // optimizing out the tuples. XlaBuilder b("sort_square"); - auto x = b.Parameter(0, ShapeUtil::MakeShape(F32, {}), "x"); - auto x2 = b.Mul(x, x); - auto x_smaller_tuple = b.Tuple({x, x2}); - auto x2_smaller_tuple = b.Tuple({x2, x}); - auto sorted = b.Select(b.Lt(x, x2), x_smaller_tuple, x2_smaller_tuple); - auto smaller = b.GetTupleElement(sorted, 0); - auto greater = b.GetTupleElement(sorted, 1); - b.Add(greater, b.Mul(b.ConstantR0(100.0f), smaller)); + auto x = Parameter(&b, 0, ShapeUtil::MakeShape(F32, {}), "x"); + auto x2 = Mul(x, x); + auto x_smaller_tuple = Tuple(&b, {x, x2}); + auto x2_smaller_tuple = Tuple(&b, {x2, x}); + auto sorted = Select(Lt(x, x2), x_smaller_tuple, x2_smaller_tuple); + auto smaller = GetTupleElement(sorted, 0); + auto greater = GetTupleElement(sorted, 1); + Add(greater, Mul(ConstantR0(&b, 100.0f), smaller)); auto computation_status = b.Build(); ASSERT_IS_OK(computation_status.status()); tuple_computation = computation_status.ConsumeValueOrDie(); } XlaBuilder b(TestName()); - auto input = b.ConstantR1({-1.0f, 1.0f, 2.1f}); - b.Map({input}, tuple_computation, {0}); + auto input = ConstantR1(&b, {-1.0f, 1.0f, 2.1f}); + Map(&b, {input}, tuple_computation, {0}); ComputeAndCompareR1(&b, {-99.0f, 101.0f, 214.41f}, {}, error_spec_); } @@ -317,14 +324,15 @@ XLA_TEST_F(TupleTest, SelectBetweenTuplesOnTrue) { std::initializer_list vec1 = {1.f, 2.f, 3.f}; std::initializer_list vec2 = {2.f, 4.f, 6.f}; - auto tuple12 = builder.Tuple( - {builder.ConstantR1(vec1), builder.ConstantR1(vec2)}); - auto tuple21 = builder.Tuple( - {builder.ConstantR1(vec2), builder.ConstantR1(vec1)}); - - builder.Select(builder.ConstantR0(true), tuple12, tuple21); - auto expected = Literal::MakeTuple({Literal::CreateR1(vec1).get(), - Literal::CreateR1(vec2).get()}); + auto tuple12 = Tuple(&builder, {ConstantR1(&builder, vec1), + ConstantR1(&builder, vec2)}); + auto tuple21 = Tuple(&builder, {ConstantR1(&builder, vec2), + ConstantR1(&builder, vec1)}); + + Select(ConstantR0(&builder, true), tuple12, tuple21); + auto expected = + LiteralUtil::MakeTuple({LiteralUtil::CreateR1(vec1).get(), + LiteralUtil::CreateR1(vec2).get()}); ComputeAndCompareTuple(&builder, *expected, {}, error_spec_); } @@ -335,14 +343,13 @@ XLA_TEST_F(TupleTest, SelectBetweenTuplesElementResult) { std::initializer_list vec1 = {1.f, 2.f, 3.f}; std::initializer_list vec2 = {2.f, 4.f, 6.f}; - auto tuple12 = builder.Tuple( - {builder.ConstantR1(vec1), builder.ConstantR1(vec2)}); - auto tuple21 = builder.Tuple( - {builder.ConstantR1(vec2), builder.ConstantR1(vec1)}); + auto tuple12 = Tuple(&builder, {ConstantR1(&builder, vec1), + ConstantR1(&builder, vec2)}); + auto tuple21 = Tuple(&builder, {ConstantR1(&builder, vec2), + ConstantR1(&builder, vec1)}); - auto select = - builder.Select(builder.ConstantR0(false), tuple12, tuple21); - builder.GetTupleElement(select, 0); + auto select = Select(ConstantR0(&builder, false), tuple12, tuple21); + GetTupleElement(select, 0); ComputeAndCompareR1(&builder, vec2, {}, error_spec_); } @@ -371,19 +378,16 @@ XLA_TEST_F(TupleTest, SelectBetweenTuplesCascaded) { std::initializer_list vec1 = {1.f, 2.f, 3.f}; std::initializer_list vec2 = {2.f, 4.f, 6.f}; - auto pred_tuple = builder.Tuple( - {builder.ConstantR0(true), builder.ConstantR0(false)}); - auto tuple12 = builder.Tuple( - {builder.ConstantR1(vec1), builder.ConstantR1(vec2)}); - auto tuple21 = builder.Tuple( - {builder.ConstantR1(vec2), builder.ConstantR1(vec1)}); + auto pred_tuple = Tuple(&builder, {ConstantR0(&builder, true), + ConstantR0(&builder, false)}); + auto tuple12 = Tuple(&builder, {ConstantR1(&builder, vec1), + ConstantR1(&builder, vec2)}); + auto tuple21 = Tuple(&builder, {ConstantR1(&builder, vec2), + ConstantR1(&builder, vec1)}); - auto select1 = - builder.Select(builder.GetTupleElement(pred_tuple, 0), tuple12, tuple21); - auto select2 = - builder.Select(builder.GetTupleElement(pred_tuple, 1), tuple21, select1); - builder.Add(builder.GetTupleElement(select2, 0), - builder.GetTupleElement(select2, 1)); + auto select1 = Select(GetTupleElement(pred_tuple, 0), tuple12, tuple21); + auto select2 = Select(GetTupleElement(pred_tuple, 1), tuple21, select1); + Add(GetTupleElement(select2, 0), GetTupleElement(select2, 1)); ComputeAndCompareR1(&builder, {3.f, 6.f, 9.f}, {}, error_spec_); } @@ -395,31 +399,32 @@ XLA_TEST_F(TupleTest, SelectBetweenTuplesReuseConstants) { std::initializer_list vec1 = {1.f, 2.f, 3.f}; std::initializer_list vec2 = {2.f, 4.f, 6.f}; - auto c1 = builder.ConstantR1(vec1); - auto c2 = builder.ConstantR1(vec2); - auto tuple12 = builder.Tuple({c1, c2}); - auto tuple21 = builder.Tuple({c2, c1}); + auto c1 = ConstantR1(&builder, vec1); + auto c2 = ConstantR1(&builder, vec2); + auto tuple12 = Tuple(&builder, {c1, c2}); + auto tuple21 = Tuple(&builder, {c2, c1}); - builder.Select(builder.ConstantR0(false), tuple12, tuple21); + Select(ConstantR0(&builder, false), tuple12, tuple21); - auto expected = Literal::MakeTuple({Literal::CreateR1(vec2).get(), - Literal::CreateR1(vec1).get()}); + auto expected = + LiteralUtil::MakeTuple({LiteralUtil::CreateR1(vec2).get(), + LiteralUtil::CreateR1(vec1).get()}); ComputeAndCompareTuple(&builder, *expected, {}, error_spec_); } XLA_TEST_F(TupleTest, NestedTuples) { XlaBuilder builder(TestName()); - auto inner_tuple = builder.Tuple( - {builder.ConstantR1({1.0, 2.0}), builder.ConstantR0(42.0)}); - builder.Tuple({inner_tuple, builder.ConstantR1({22.0, 44.0})}); + auto inner_tuple = Tuple(&builder, {ConstantR1(&builder, {1.0, 2.0}), + ConstantR0(&builder, 42.0)}); + Tuple(&builder, {inner_tuple, ConstantR1(&builder, {22.0, 44.0})}); - auto expected_v1 = Literal::CreateR1({1.0, 2.0}); - auto expected_s = Literal::CreateR0(42.0); + auto expected_v1 = LiteralUtil::CreateR1({1.0, 2.0}); + auto expected_s = LiteralUtil::CreateR0(42.0); auto expected_inner_tuple = - Literal::MakeTuple({expected_v1.get(), expected_s.get()}); - auto expected_v2 = Literal::CreateR1({22.0, 44.0}); + LiteralUtil::MakeTuple({expected_v1.get(), expected_s.get()}); + auto expected_v2 = LiteralUtil::CreateR1({22.0, 44.0}); auto expected = - Literal::MakeTuple({expected_inner_tuple.get(), expected_v2.get()}); + LiteralUtil::MakeTuple({expected_inner_tuple.get(), expected_v2.get()}); ComputeAndCompareTuple(&builder, *expected, {}, error_spec_); } @@ -432,21 +437,21 @@ XLA_TEST_F(TupleTest, GetTupleElementOfNestedTuple) { Shape outer_tuple_shape = ShapeUtil::MakeTupleShape({inner_tuple_shape, data_shape}); - auto input = builder.Parameter(0, outer_tuple_shape, "input"); - auto gte0 = builder.GetTupleElement(input, 0); - auto gte1 = builder.GetTupleElement(gte0, 1); - builder.Add(gte1, builder.ConstantR1({10.0, 11.0, 12.0})); + auto input = Parameter(&builder, 0, outer_tuple_shape, "input"); + auto gte0 = GetTupleElement(input, 0); + auto gte1 = GetTupleElement(gte0, 1); + Add(gte1, ConstantR1(&builder, {10.0, 11.0, 12.0})); std::unique_ptr data = client_ - ->TransferToServer(*Literal::MakeTuple({ - Literal::MakeTuple( + ->TransferToServer(*LiteralUtil::MakeTuple({ + LiteralUtil::MakeTuple( { - Literal::CreateR1({1.0, 2.0, 3.0}).get(), - Literal::CreateR1({4.0, 5.0, 6.0}).get(), + LiteralUtil::CreateR1({1.0, 2.0, 3.0}).get(), + LiteralUtil::CreateR1({4.0, 5.0, 6.0}).get(), }) .get(), - Literal::CreateR1({7.0, 8.0, 9.0}).get(), + LiteralUtil::CreateR1({7.0, 8.0, 9.0}).get(), })) .ConsumeValueOrDie(); @@ -463,25 +468,26 @@ XLA_TEST_F(TupleTest, ComplexTuples) { Shape c64r2 = ShapeUtil::MakeShape(C64, {3, 2}); Shape arg0_shape = ShapeUtil::MakeTupleShape( {c64r0, ShapeUtil::MakeTupleShape({c64r1, c64r2})}); - auto input0 = builder.Parameter(0, arg0_shape, "input0"); - auto t0 = builder.GetTupleElement(input0, 0); - auto t1 = builder.GetTupleElement(input0, 1); - auto t10 = builder.GetTupleElement(t1, 0); - auto t11 = builder.GetTupleElement(t1, 1); - auto sum = builder.Add(builder.Add(t10, t11, {1}), t0); - auto input1 = builder.Parameter(1, c64r1, "input1"); - auto prod = builder.Mul(input1, sum, {1}); - builder.Tuple({builder.Tuple({prod, sum}), - builder.ConstantR0({123, 456})}); + auto input0 = Parameter(&builder, 0, arg0_shape, "input0"); + auto t0 = GetTupleElement(input0, 0); + auto t1 = GetTupleElement(input0, 1); + auto t10 = GetTupleElement(t1, 0); + auto t11 = GetTupleElement(t1, 1); + auto sum = Add(Add(t10, t11, {1}), t0); + auto input1 = Parameter(&builder, 1, c64r1, "input1"); + auto prod = Mul(input1, sum, {1}); + Tuple(&builder, {Tuple(&builder, {prod, sum}), + ConstantR0(&builder, {123, 456})}); } std::unique_ptr arg0 = client_ - ->TransferToServer(*Literal::MakeTuple( - {Literal::CreateR0({1, 2}).get(), - Literal::MakeTuple( - {Literal::CreateR1({{10, 20}, {30, 40}}).get(), - Literal::CreateR2( + ->TransferToServer(*LiteralUtil::MakeTuple( + {LiteralUtil::CreateR0({1, 2}).get(), + LiteralUtil::MakeTuple( + {LiteralUtil::CreateR1({{10, 20}, {30, 40}}) + .get(), + LiteralUtil::CreateR2( {{{100, 200}, {300, 400}}, {{1000, 2000}, {3000, 4000}}, {{10000, 20000}, {30000, 40000}}}) @@ -490,11 +496,13 @@ XLA_TEST_F(TupleTest, ComplexTuples) { .ConsumeValueOrDie(); std::unique_ptr arg1 = client_ - ->TransferToServer(*Literal::CreateR1({{1, 2}, {1, -2}})) + ->TransferToServer( + *LiteralUtil::CreateR1({{1, 2}, {1, -2}})) .ConsumeValueOrDie(); - auto sum = Literal::CreateR2({{{111, 222}, {331, 442}}, - {{1011, 2022}, {3031, 4042}}, - {{10011, 20022}, {30031, 40042}}}); + auto sum = + LiteralUtil::CreateR2({{{111, 222}, {331, 442}}, + {{1011, 2022}, {3031, 4042}}, + {{10011, 20022}, {30031, 40042}}}); auto prod = MakeUnique(sum->shape()); ASSERT_TRUE(prod->Populate( [&sum](tensorflow::gtl::ArraySlice indexes) { @@ -504,9 +512,9 @@ XLA_TEST_F(TupleTest, ComplexTuples) { : complex64(1, -2)); }) .ok()); - auto expected = - Literal::MakeTuple({Literal::MakeTuple({prod.get(), sum.get()}).get(), - Literal::CreateR0({123, 456}).get()}); + auto expected = LiteralUtil::MakeTuple( + {LiteralUtil::MakeTuple({prod.get(), sum.get()}).get(), + LiteralUtil::CreateR0({123, 456}).get()}); ComputeAndCompareTuple(&builder, *expected, {arg0.get(), arg1.get()}, error_spec_); } @@ -529,10 +537,11 @@ XLA_TEST_F(TupleHloTest, DISABLED_ON_INTERPRETER(BitcastAfterGTE)) { auto module = HloRunner::CreateModuleFromString(testcase, GetDebugOptionsForTest()) .ValueOrDie(); - auto param = Literal::MakeTupleOwned(Literal::CreateR1({1, 2, 3})); + auto param = + LiteralUtil::MakeTupleOwned(LiteralUtil::CreateR1({1, 2, 3})); auto result = ExecuteNoHloPasses(std::move(module), {param.get()}); EXPECT_TRUE(LiteralTestUtil::Equal( - *Literal::MakeTupleOwned(Literal::CreateR2({{1, 2, 3}})), + *LiteralUtil::MakeTupleOwned(LiteralUtil::CreateR2({{1, 2, 3}})), *result)); } diff --git a/tensorflow/compiler/xla/tests/unary_op_test.cc b/tensorflow/compiler/xla/tests/unary_op_test.cc index c3abe22797f5eaa76ced2ad8534bd68c32983e60..a90a6fb0a5b5bb5119eee93c9c6a1377e3461b46 100644 --- a/tensorflow/compiler/xla/tests/unary_op_test.cc +++ b/tensorflow/compiler/xla/tests/unary_op_test.cc @@ -38,8 +38,8 @@ class UnaryOpTest : public ClientLibraryTestBase { template void AbsSize0TestHelper() { XlaBuilder builder(TestName()); - auto arg = builder.ConstantR1({}); - auto abs = builder.Abs(arg); + auto arg = ConstantR1(&builder, {}); + Abs(arg); if (primitive_util::NativeToPrimitiveType() == C64) { ComputeAndCompareR1(&builder, {}, {}); @@ -51,8 +51,8 @@ class UnaryOpTest : public ClientLibraryTestBase { template void AbsTestHelper() { XlaBuilder builder(TestName()); - auto arg = builder.ConstantR1({-2, 25, 0, -123, inf(), -inf()}); - auto abs = builder.Abs(arg); + auto arg = ConstantR1(&builder, {-2, 25, 0, -123, inf(), -inf()}); + Abs(arg); ComputeAndCompareR1(&builder, {2, 25, 0, 123, inf(), inf()}, {}); } @@ -60,9 +60,9 @@ class UnaryOpTest : public ClientLibraryTestBase { template void SignTestHelper() { XlaBuilder builder(TestName()); - auto arg = builder.ConstantR1( - {-2, 25, 0, static_cast(-0.0), -123, inf(), -inf()}); - auto sign = builder.Sign(arg); + auto arg = ConstantR1( + &builder, {-2, 25, 0, static_cast(-0.0), -123, inf(), -inf()}); + Sign(arg); ComputeAndCompareR1(&builder, {-1, 1, 0, 0, -1, 1, -1}, {}); } @@ -70,10 +70,10 @@ class UnaryOpTest : public ClientLibraryTestBase { template void SignAbsTestHelper() { XlaBuilder builder(TestName()); - auto arg = builder.ConstantR1({-2, 25, 0, -123}); - auto sign = builder.Sign(arg); - auto abs = builder.Abs(arg); - builder.Sub(builder.Mul(sign, abs), arg); + auto arg = ConstantR1(&builder, {-2, 25, 0, -123}); + auto sign = Sign(arg); + auto abs = Abs(arg); + Sub(Mul(sign, abs), arg); ComputeAndCompareR1(&builder, {0, 0, 0, 0}, {}); } @@ -92,27 +92,28 @@ int64 UnaryOpTest::inf() { template <> void UnaryOpTest::AbsTestHelper() { XlaBuilder builder(TestName()); - auto arg = builder.ConstantR1({{-2, 0}, - {0, 25}, - {0, 0}, - {-0.3f, 0.4f}, - {0, inf()}, - {-inf(), 0}}); - auto abs = builder.Abs(arg); + auto arg = ConstantR1(&builder, {{-2, 0}, + {0, 25}, + {0, 0}, + {-0.3f, 0.4f}, + {0, inf()}, + {-inf(), 0}}); + Abs(arg); std::unique_ptr expected = - Literal::CreateR1({2, 25, 0, 0.5, inf(), inf()}); + LiteralUtil::CreateR1({2, 25, 0, 0.5, inf(), inf()}); ComputeAndCompareLiteral(&builder, *expected, {}, ErrorSpec(1e-6f)); } template <> void UnaryOpTest::SignTestHelper() { XlaBuilder builder(TestName()); - auto arg = builder.ConstantR1( + auto arg = ConstantR1( + &builder, {{-2, 0}, {0, 25}, {0, 0}, {static_cast(-0.0), 0}, {-1, 1}}); - auto sign = builder.Sign(arg); + Sign(arg); - std::unique_ptr expected = Literal::CreateR1( + std::unique_ptr expected = LiteralUtil::CreateR1( {{-1, 0}, {0, 1}, {0, 0}, {0, 0}, {-std::sqrt(0.5f), std::sqrt(0.5f)}}); ComputeAndCompareLiteral(&builder, *expected, {}, ErrorSpec(1e-6f)); } @@ -121,13 +122,13 @@ template <> void UnaryOpTest::SignAbsTestHelper() { XlaBuilder builder(TestName()); auto arg = - builder.ConstantR1({{-2, 0}, {0, 25}, {0, 0}, {-0.4, 0.3}}); - auto sign = builder.Sign(arg); - auto abs = builder.Abs(arg); - builder.Sub(builder.Mul(sign, builder.ConvertElementType(abs, C64)), arg); + ConstantR1(&builder, {{-2, 0}, {0, 25}, {0, 0}, {-0.4, 0.3}}); + auto sign = Sign(arg); + auto abs = Abs(arg); + Sub(Mul(sign, ConvertElementType(abs, C64)), arg); std::unique_ptr expected = - Literal::CreateR1({0, 0, 0, 0}); + LiteralUtil::CreateR1({0, 0, 0, 0}); ComputeAndCompareLiteral(&builder, *expected, {}, ErrorSpec(1e-6f)); } @@ -145,37 +146,34 @@ XLA_TEST_F(UnaryOpTest, AbsTestR1) { XLA_TEST_F(UnaryOpTest, AbsTestR0) { XlaBuilder builder(TestName()); - auto argi = builder.ConstantR0(-5); - auto absi = builder.Abs(argi); - auto argf = builder.ConstantR0(-3.0f); - auto absf = builder.Abs(argf); - auto argf0 = builder.ConstantR0(-0.0f); - auto absf0 = builder.Abs(argf0); - auto argc = builder.ConstantR0({-0.3f, 0.4f}); - auto absc = builder.Abs(argc); - builder.Add(builder.Add(absc, absf0), - builder.Add(absf, builder.ConvertElementType(absi, F32))); + auto argi = ConstantR0(&builder, -5); + auto absi = Abs(argi); + auto argf = ConstantR0(&builder, -3.0f); + auto absf = Abs(argf); + auto argf0 = ConstantR0(&builder, -0.0f); + auto absf0 = Abs(argf0); + auto argc = ConstantR0(&builder, {-0.3f, 0.4f}); + auto absc = Abs(argc); + Add(Add(absc, absf0), Add(absf, ConvertElementType(absi, F32))); ComputeAndCompareR0(&builder, 8.5f, {}); } XLA_TEST_F(UnaryOpTest, SignTestR0) { XlaBuilder builder(TestName()); - auto argi = builder.ConstantR0(-5); - auto sgni = builder.Sign(argi); // -1 - auto argf = builder.ConstantR0(-4.0f); - auto sgnf = builder.Sign(argf); // -1 - auto argf0 = builder.ConstantR0(-0.0f); - auto sgnf0 = builder.Sign(argf0); // 0 - auto argc = builder.ConstantR0({-.3, .4}); - auto sgnc = builder.Sign(argc); // (-.6, .8) - builder.Add(sgnc, builder.ConvertElementType( - builder.Add(builder.Add(sgnf0, sgnf), - builder.ConvertElementType(sgni, F32)), - C64)); + auto argi = ConstantR0(&builder, -5); + auto sgni = Sign(argi); // -1 + auto argf = ConstantR0(&builder, -4.0f); + auto sgnf = Sign(argf); // -1 + auto argf0 = ConstantR0(&builder, -0.0f); + auto sgnf0 = Sign(argf0); // 0 + auto argc = ConstantR0(&builder, {-.3, .4}); + auto sgnc = Sign(argc); // (-.6, .8) + Add(sgnc, ConvertElementType( + Add(Add(sgnf0, sgnf), ConvertElementType(sgni, F32)), C64)); std::unique_ptr expected = - Literal::CreateR0({-2.6f, 0.8f}); + LiteralUtil::CreateR0({-2.6f, 0.8f}); ComputeAndCompareLiteral(&builder, *expected, {}, ErrorSpec(1e-6f)); } @@ -194,9 +192,9 @@ XLA_TEST_F(UnaryOpTest, SignAbsTestR1) { XLA_TEST_F(UnaryOpTest, UnsignedAbsTestR1) { XlaBuilder builder(TestName()); - auto arg = builder.ConstantR1( - {2, 25, 0, 123, std::numeric_limits::max()}); - auto abs = builder.Abs(arg); + auto arg = ConstantR1( + &builder, {2, 25, 0, 123, std::numeric_limits::max()}); + Abs(arg); ComputeAndCompareR1( &builder, {2, 25, 0, 123, std::numeric_limits::max()}, {}); @@ -204,37 +202,37 @@ XLA_TEST_F(UnaryOpTest, UnsignedAbsTestR1) { XLA_TEST_F(UnaryOpTest, UnsignedSignTestR1) { XlaBuilder builder(TestName()); - auto arg = builder.ConstantR1( - {2, 25, 0, 123, std::numeric_limits::max()}); - auto sign = builder.Sign(arg); + auto arg = ConstantR1( + &builder, {2, 25, 0, 123, std::numeric_limits::max()}); + Sign(arg); ComputeAndCompareR1(&builder, {1, 1, 0, 1, 1}, {}); } XLA_TEST_F(UnaryOpTest, SignAbsTestR2) { XlaBuilder builder(TestName()); - auto arg = builder.ConstantR2({{1.0, -2.0}, {-3.0, 4.0}}); - auto sign = builder.Sign(arg); - auto abs = builder.Abs(arg); - builder.Sub(builder.Mul(sign, abs), arg); + auto arg = ConstantR2(&builder, {{1.0, -2.0}, {-3.0, 4.0}}); + auto sign = Sign(arg); + auto abs = Abs(arg); + Sub(Mul(sign, abs), arg); ComputeAndCompareR2(&builder, {{0, 0}, {0, 0}}, {}); } XLA_TEST_F(UnaryOpTest, ConvertElementTypePredToS32) { XlaBuilder builder(TestName()); - auto lhs = builder.ConstantR1({0, 1}); - auto rhs = builder.ConstantR1({1, 1}); - builder.ConvertElementType(builder.Eq(lhs, rhs), S32); + auto lhs = ConstantR1(&builder, {0, 1}); + auto rhs = ConstantR1(&builder, {1, 1}); + ConvertElementType(Eq(lhs, rhs), S32); ComputeAndCompareR1(&builder, {0, 1}, {}); } XLA_TEST_F(UnaryOpTest, ConvertElementTypePredToF32) { XlaBuilder builder(TestName()); - auto lhs = builder.ConstantR1({0, 1}); - auto rhs = builder.ConstantR1({1, 1}); - builder.ConvertElementType(builder.Eq(lhs, rhs), F32); + auto lhs = ConstantR1(&builder, {0, 1}); + auto rhs = ConstantR1(&builder, {1, 1}); + ConvertElementType(Eq(lhs, rhs), F32); ComputeAndCompareR1(&builder, {0.0, 1.0}, {}); } diff --git a/tensorflow/compiler/xla/tests/vector_ops_reduce_test.cc b/tensorflow/compiler/xla/tests/vector_ops_reduce_test.cc index 82d301983fc7885ef5c1c1ed05b74fc017bb7727..ea3aba6df1d3fbd492a23b280309322b8524c0bf 100644 --- a/tensorflow/compiler/xla/tests/vector_ops_reduce_test.cc +++ b/tensorflow/compiler/xla/tests/vector_ops_reduce_test.cc @@ -46,7 +46,7 @@ class VecOpsReduceTest : public ClientLibraryTestBase { {{1.0, 2.0, 3.0}, // } plane 2 in dim 0 {4.0, 5.0, 6.0}}}); // clang-format on - return builder_.ConstantR3FromArray3D(x3d); + return ConstantR3FromArray3D(&builder_, x3d); } XlaBuilder builder_; @@ -56,11 +56,10 @@ class VecOpsReduceTest : public ClientLibraryTestBase { TEST_F(VecOpsReduceTest, AddReduceR1F32) { auto sum_reducer = CreateScalarAddComputation(F32, &builder_); - auto x = builder_.ConstantR1( - {2.1, -2.6, 2.6, -4.0, 2.1, 2.3, -5.0, -0.9, -2.4, 1.6}); - auto add_reduce = - builder_.Reduce(x, builder_.ConstantR0(0.0f), sum_reducer, - /*dimensions_to_reduce=*/{0}); + auto x = ConstantR1( + &builder_, {2.1, -2.6, 2.6, -4.0, 2.1, 2.3, -5.0, -0.9, -2.4, 1.6}); + Reduce(x, ConstantR0(&builder_, 0.0f), sum_reducer, + /*dimensions_to_reduce=*/{0}); ComputeAndCompareR0(&builder_, -4.2f, {}, errspec_); } @@ -71,10 +70,9 @@ TEST_F(VecOpsReduceTest, AddReduceBigR1F32) { std::vector input(3000); std::iota(input.begin(), input.end(), 100.0f); - auto x = builder_.ConstantR1(input); - auto add_reduce = - builder_.Reduce(x, builder_.ConstantR0(0.0f), sum_reducer, - /*dimensions_to_reduce=*/{0}); + auto x = ConstantR1(&builder_, input); + Reduce(x, ConstantR0(&builder_, 0.0f), sum_reducer, + /*dimensions_to_reduce=*/{0}); float expected = std::accumulate(input.begin(), input.end(), 0.0f); ComputeAndCompareR0(&builder_, expected, {}, errspec_); @@ -83,11 +81,10 @@ TEST_F(VecOpsReduceTest, AddReduceBigR1F32) { TEST_F(VecOpsReduceTest, MaxReduceR1F32) { auto max_reducer = CreateScalarMax(); - auto x = builder_.ConstantR1( - {2.1, -2.6, 2.6, -4.0, 2.1, 2.3, -5.0, -0.9, -2.4, 1.6}); - auto max_reduce = - builder_.Reduce(x, builder_.ConstantR0(0.0f), max_reducer, - /*dimensions_to_reduce=*/{0}); + auto x = ConstantR1( + &builder_, {2.1, -2.6, 2.6, -4.0, 2.1, 2.3, -5.0, -0.9, -2.4, 1.6}); + Reduce(x, ConstantR0(&builder_, 0.0f), max_reducer, + /*dimensions_to_reduce=*/{0}); ComputeAndCompareR0(&builder_, 2.6f, {}, errspec_); } @@ -95,11 +92,10 @@ TEST_F(VecOpsReduceTest, MaxReduceR1F32) { TEST_F(VecOpsReduceTest, MaxReduceR1F32WithNontrivialInit) { auto max_reducer = CreateScalarMax(); - auto x = builder_.ConstantR1( - {2.1, -2.6, 2.6, -4.0, 2.1, 2.3, -5.0, -0.9, -2.4, 1.6}); - auto max_reduce = - builder_.Reduce(x, builder_.ConstantR0(4.0f), max_reducer, - /*dimensions_to_reduce=*/{0}); + auto x = ConstantR1( + &builder_, {2.1, -2.6, 2.6, -4.0, 2.1, 2.3, -5.0, -0.9, -2.4, 1.6}); + Reduce(x, ConstantR0(&builder_, 4.0f), max_reducer, + /*dimensions_to_reduce=*/{0}); ComputeAndCompareR0(&builder_, 4.0f, {}, errspec_); } @@ -108,15 +104,14 @@ TEST_F(VecOpsReduceTest, AddReduceR2F32Dim1) { auto sum_reducer = CreateScalarAddComputation(F32, &builder_); // clang-format off - auto x = builder_.ConstantR2({ + auto x = ConstantR2(&builder_, { {1.0, 2.0, 3.0}, // | dim 0 {4.0, 5.0, 6.0}}); // | // ------ dim 1 ---------- // clang-format on - auto add_reduce = - builder_.Reduce(x, builder_.ConstantR0(0.0f), sum_reducer, - /*dimensions_to_reduce=*/{1}); + Reduce(x, ConstantR0(&builder_, 0.0f), sum_reducer, + /*dimensions_to_reduce=*/{1}); ComputeAndCompareR1(&builder_, {6.0, 15.0}, {}, errspec_); } @@ -125,13 +120,12 @@ TEST_F(VecOpsReduceTest, AddReduceR2F32Dim0) { auto sum_reducer = CreateScalarAddComputation(F32, &builder_); // clang-format off - auto x = builder_.ConstantR2({ + auto x = ConstantR2(&builder_, { {1.0, 2.0, 3.0}, {4.0, 5.0, 6.0}}); // clang-format on - auto add_reduce = - builder_.Reduce(x, builder_.ConstantR0(0.0f), sum_reducer, - /*dimensions_to_reduce=*/{0}); + Reduce(x, ConstantR0(&builder_, 0.0f), sum_reducer, + /*dimensions_to_reduce=*/{0}); ComputeAndCompareR1(&builder_, {5.0, 7.0, 9.0}, {}, errspec_); } @@ -139,9 +133,8 @@ TEST_F(VecOpsReduceTest, AddReduceR2F32Dim0) { TEST_F(VecOpsReduceTest, AddReduceR3F32Dim2) { auto sum_reducer = CreateScalarAddComputation(F32, &builder_); auto x = BuildSampleConstantCube(); - auto add_reduce = - builder_.Reduce(x, builder_.ConstantR0(0.0f), sum_reducer, - /*dimensions_to_reduce=*/{2}); + Reduce(x, ConstantR0(&builder_, 0.0f), sum_reducer, + /*dimensions_to_reduce=*/{2}); Array2D expected_array({{6.0f, 15.0f}, {6.0f, 15.0f}, {6.0f, 15.0f}}); @@ -151,9 +144,8 @@ TEST_F(VecOpsReduceTest, AddReduceR3F32Dim2) { TEST_F(VecOpsReduceTest, AddReduceR3F32Dim1) { auto sum_reducer = CreateScalarAddComputation(F32, &builder_); auto x = BuildSampleConstantCube(); - auto add_reduce = - builder_.Reduce(x, builder_.ConstantR0(0.0f), sum_reducer, - /*dimensions_to_reduce=*/{1}); + Reduce(x, ConstantR0(&builder_, 0.0f), sum_reducer, + /*dimensions_to_reduce=*/{1}); Array2D expected_array( {{5.0f, 7.0f, 9.0f}, {5.0f, 7.0f, 9.0f}, {5.0f, 7.0f, 9.0f}}); @@ -164,9 +156,8 @@ TEST_F(VecOpsReduceTest, AddReduceR3F32Dim1) { TEST_F(VecOpsReduceTest, AddReduceR3F32Dim0) { auto sum_reducer = CreateScalarAddComputation(F32, &builder_); auto x = BuildSampleConstantCube(); - auto add_reduce = - builder_.Reduce(x, builder_.ConstantR0(0.0f), sum_reducer, - /*dimensions_to_reduce=*/{0}); + Reduce(x, ConstantR0(&builder_, 0.0f), sum_reducer, + /*dimensions_to_reduce=*/{0}); Array2D expected_array({{3.0f, 6.0f, 9.0f}, {12.0f, 15.0f, 18.0f}}); @@ -176,9 +167,8 @@ TEST_F(VecOpsReduceTest, AddReduceR3F32Dim0) { TEST_F(VecOpsReduceTest, AddReduceR3F32Dims1and2) { auto sum_reducer = CreateScalarAddComputation(F32, &builder_); auto x = BuildSampleConstantCube(); - auto add_reduce = - builder_.Reduce(x, builder_.ConstantR0(0.0f), sum_reducer, - /*dimensions_to_reduce=*/{1, 2}); + Reduce(x, ConstantR0(&builder_, 0.0f), sum_reducer, + /*dimensions_to_reduce=*/{1, 2}); ComputeAndCompareR1(&builder_, {21.0, 21.0, 21.0}, {}, errspec_); } @@ -186,9 +176,8 @@ TEST_F(VecOpsReduceTest, AddReduceR3F32Dims1and2) { XLA_TEST_F(VecOpsReduceTest, AddReduceR3F32Dims0and2) { auto sum_reducer = CreateScalarAddComputation(F32, &builder_); auto x = BuildSampleConstantCube(); - auto add_reduce = - builder_.Reduce(x, builder_.ConstantR0(0.0f), sum_reducer, - /*dimensions_to_reduce=*/{0, 2}); + Reduce(x, ConstantR0(&builder_, 0.0f), sum_reducer, + /*dimensions_to_reduce=*/{0, 2}); ComputeAndCompareR1(&builder_, {18.0, 45.0}, {}, errspec_); } @@ -196,9 +185,8 @@ XLA_TEST_F(VecOpsReduceTest, AddReduceR3F32Dims0and2) { TEST_F(VecOpsReduceTest, AddReduceR3F32Dims0and1) { auto sum_reducer = CreateScalarAddComputation(F32, &builder_); auto x = BuildSampleConstantCube(); - auto add_reduce = - builder_.Reduce(x, builder_.ConstantR0(0.0f), sum_reducer, - /*dimensions_to_reduce=*/{0, 1}); + Reduce(x, ConstantR0(&builder_, 0.0f), sum_reducer, + /*dimensions_to_reduce=*/{0, 1}); ComputeAndCompareR1(&builder_, {15.0, 21.0, 27.0}, {}, errspec_); } @@ -206,9 +194,8 @@ TEST_F(VecOpsReduceTest, AddReduceR3F32Dims0and1) { TEST_F(VecOpsReduceTest, AddReduceR3F32AllDims) { auto sum_reducer = CreateScalarAddComputation(F32, &builder_); auto x = BuildSampleConstantCube(); - auto add_reduce = - builder_.Reduce(x, builder_.ConstantR0(0.0f), sum_reducer, - /*dimensions_to_reduce=*/{0, 1, 2}); + Reduce(x, ConstantR0(&builder_, 0.0f), sum_reducer, + /*dimensions_to_reduce=*/{0, 1, 2}); ComputeAndCompareR0(&builder_, 63.0, {}, errspec_); } diff --git a/tensorflow/compiler/xla/tests/vector_ops_simple_test.cc b/tensorflow/compiler/xla/tests/vector_ops_simple_test.cc index 5cce7a2bf82c1a8403536a91e67910f949ef185a..79bae22dac9599a38c73ea1dc2e6b4856395ff79 100644 --- a/tensorflow/compiler/xla/tests/vector_ops_simple_test.cc +++ b/tensorflow/compiler/xla/tests/vector_ops_simple_test.cc @@ -50,9 +50,9 @@ class VecOpsSimpleTest : public ClientLibraryTestBase { XLA_TEST_F(VecOpsSimpleTest, ExpTenValues) { XlaBuilder builder(TestName()); - auto x = builder.ConstantR1( - {2.1, -2.6, 2.6, -4.0, 2.1, 2.3, -5.0, -0.9, -2.4, 1.6}); - auto exp = builder.Exp(x); + auto x = ConstantR1( + &builder, {2.1, -2.6, 2.6, -4.0, 2.1, 2.3, -5.0, -0.9, -2.4, 1.6}); + Exp(x); std::vector expected = {8.1662, 7.4274e-02, 13.4637, 1.8316e-02, 8.1662, 9.9742, 6.7379e-03, 4.0657e-01, @@ -69,8 +69,8 @@ XLA_TEST_F(VecOpsSimpleTest, ExpManyValues) { for (int i = 0; i < count; ++i) { exponents.push_back(i / static_cast(count)); } - auto x = builder.ConstantR1(exponents); - auto exp = builder.Exp(x); + auto x = ConstantR1(&builder, exponents); + Exp(x); std::vector expected; expected.reserve(exponents.size()); @@ -98,8 +98,8 @@ XLA_TEST_F(VecOpsSimpleTest, ExpIn4D) { Array4D expected(2, 2, 2, 2, expected_vector); - auto x = builder.ConstantR4FromArray4D(exponents); - auto exp = builder.Exp(x); + auto x = ConstantR4FromArray4D(&builder, exponents); + Exp(x); ComputeAndCompareR4(&builder, expected, {}, ErrorSpec(/*aabs=*/1e-2, /*arel=*/1e-3)); @@ -107,9 +107,9 @@ XLA_TEST_F(VecOpsSimpleTest, ExpIn4D) { XLA_TEST_F(VecOpsSimpleTest, NegateTenFloatValues) { XlaBuilder builder(TestName()); - auto x = builder.ConstantR1( - {2.1, -2.6, 2.6, -4.0, 2.1, 2.3, -5.0, -0.9, -2.4, 1.6}); - builder.Neg(x); + auto x = ConstantR1( + &builder, {2.1, -2.6, 2.6, -4.0, 2.1, 2.3, -5.0, -0.9, -2.4, 1.6}); + Neg(x); std::vector expected = {-2.1, 2.6, -2.6, 4.0, -2.1, -2.3, 5.0, 0.9, 2.4, -1.6}; @@ -118,8 +118,8 @@ XLA_TEST_F(VecOpsSimpleTest, NegateTenFloatValues) { XLA_TEST_F(VecOpsSimpleTest, NegateTenInt32Values) { XlaBuilder builder(TestName()); - auto x = builder.ConstantR1({2, -2, 12, -4, 5, 20, -15, 0, -2, 1}); - builder.Neg(x); + auto x = ConstantR1(&builder, {2, -2, 12, -4, 5, 20, -15, 0, -2, 1}); + Neg(x); std::vector expected = {-2, 2, -12, 4, -5, -20, 15, 0, 2, -1}; ComputeAndCompareR1(&builder, expected, {}); @@ -127,59 +127,19 @@ XLA_TEST_F(VecOpsSimpleTest, NegateTenInt32Values) { XLA_TEST_F(VecOpsSimpleTest, NegateUint32Values) { XlaBuilder builder(TestName()); - auto x = builder.ConstantR1( - {0, 1, 42, static_cast(-1), static_cast(-12)}); - builder.Neg(x); + auto x = ConstantR1( + &builder, {0, 1, 42, static_cast(-1), static_cast(-12)}); + Neg(x); std::vector expected = {0, static_cast(-1), static_cast(-42), 1, 12}; ComputeAndCompareR1(&builder, expected, {}); } -XLA_TEST_F(VecOpsSimpleTest, SquareTenValues) { - XlaBuilder builder(TestName()); - auto x = builder.ConstantR1( - {2.1, -2.6, 2.6, -4.0, 2.1, 2.3, -5.0, -0.9, -2.4, 1.6}); - builder.SquareF32(x); - - std::vector expected = {4.41, 6.76, 6.76, 16., 4.41, - 5.29, 25., 0.81, 5.76, 2.56}; - ComputeAndCompareR1(&builder, expected, {}, error_spec_); -} - -XLA_TEST_F(VecOpsSimpleTest, ReciprocalTenValues) { - XlaBuilder builder(TestName()); - auto x = builder.ConstantR1( - {2.1, -2.6, 2.6, -4.0, 2.1, 2.3, -5.0, -0.9, -2.4, 1.6}); - builder.ReciprocalF32(x); - - std::vector expected = { - 0.47619048, -0.38461538, 0.38461538, -0.25, 0.47619048, - 0.43478261, -0.2, -1.11111111, -0.41666667, 0.625}; - ComputeAndCompareR1(&builder, expected, {}, error_spec_); -} - -XLA_TEST_F(VecOpsSimpleTest, SqrtZeroes) { - XlaBuilder builder(TestName()); - auto x = builder.ConstantR1({0.0, -0.0}); - auto exp = builder.SqrtF32(x); - - ComputeAndCompareR1(&builder, {0, 0}, {}, error_spec_); -} - -XLA_TEST_F(VecOpsSimpleTest, SqrtSixValues) { - XlaBuilder builder(TestName()); - auto x = builder.ConstantR1({16.0, 1.0, 1024.0, 0.16, 0.2, 12345}); - auto exp = builder.SqrtF32(x); - - std::vector expected = {4, 1, 32, 0.4, 0.4472, 111.1080}; - ComputeAndCompareR1(&builder, expected, {}, error_spec_); -} - XLA_TEST_F(VecOpsSimpleTest, InvSqrtSevenValues) { XlaBuilder builder(TestName()); - auto x = - builder.ConstantR1({16.0, 1.0, 1024.0, 0.16, 0.2, 12345, 1.2345}); - auto exp = builder.Pow(x, builder.ConstantR0(-.5f)); + auto x = ConstantR1(&builder, + {16.0, 1.0, 1024.0, 0.16, 0.2, 12345, 1.2345}); + Pow(x, ConstantR0(&builder, -.5f)); std::vector expected = {.25, 1, .03125, 2.5, 2.23607, .009000, .900025}; @@ -191,11 +151,11 @@ XLA_TEST_F(VecOpsSimpleTest, AddTenValuesViaMap) { XlaBuilder builder(TestName()); auto add = CreateScalarAddComputation(F32, &builder); - auto x = builder.ConstantR1( - {2.1, -2.6, 2.6, -4.0, 2.1, 2.3, -5.0, -0.9, -2.4, 1.6}); - auto y = builder.ConstantR1( - {-0.4, -0.6, -3.0, 0.2, 3.8, -2.2, -1.8, 4.9, 1.4, 0.6}); - auto max = builder.Map({x, y}, add, {0}); + auto x = ConstantR1( + &builder, {2.1, -2.6, 2.6, -4.0, 2.1, 2.3, -5.0, -0.9, -2.4, 1.6}); + auto y = ConstantR1( + &builder, {-0.4, -0.6, -3.0, 0.2, 3.8, -2.2, -1.8, 4.9, 1.4, 0.6}); + Map(&builder, {x, y}, add, {0}); std::vector expected = {1.7, -3.2, -0.4, -3.8, 5.9, 0.1, -6.8, 4., -1., 2.2}; @@ -204,11 +164,11 @@ XLA_TEST_F(VecOpsSimpleTest, AddTenValuesViaMap) { XLA_TEST_F(VecOpsSimpleTest, MaxTenValues) { XlaBuilder builder(TestName()); - auto x = builder.ConstantR1( - {2.1, -2.6, 2.6, -4.0, 2.1, 2.3, -5.0, -0.9, -2.4, 1.6}); - auto y = builder.ConstantR1( - {-0.4, -0.6, -3.0, 0.2, 3.8, -2.2, -1.8, 4.9, 1.4, 0.6}); - auto max = builder.Max(x, y); + auto x = ConstantR1( + &builder, {2.1, -2.6, 2.6, -4.0, 2.1, 2.3, -5.0, -0.9, -2.4, 1.6}); + auto y = ConstantR1( + &builder, {-0.4, -0.6, -3.0, 0.2, 3.8, -2.2, -1.8, 4.9, 1.4, 0.6}); + Max(x, y); std::vector expected = {2.1, -0.6, 2.6, 0.2, 3.8, 2.3, -1.8, 4.9, 1.4, 1.6}; @@ -227,7 +187,7 @@ XLA_TEST_F(VecOpsSimpleTest, MaxTenValuesFromParams) { {21.0f, 22.0f, 23.0f, 24.0f}, /*parameter_number=*/1, /*name=*/"v2", /*builder=*/&builder, /*data_handle=*/&v2); - auto max = builder.Max(v1, v2); + Max(v1, v2); ComputeAndCompareR1(&builder, {41.0f, 22.0f, 23.0f, 84.0f}, {param0_data.get(), param1_data.get()}, error_spec_); @@ -267,7 +227,7 @@ XLA_TEST_F(VecOpsSimpleTest, Max15000ValuesFromParams) { CreateR1Parameter(v2vec, /*parameter_number=*/1, /*name=*/"v2", /*builder=*/&builder, /*data_handle=*/&v2); - auto max = builder.Max(v1, v2); + Max(v1, v2); ComputeAndCompareR1(&builder, expected_vec, {param0_data.get(), param1_data.get()}, error_spec_); @@ -275,10 +235,10 @@ XLA_TEST_F(VecOpsSimpleTest, Max15000ValuesFromParams) { XLA_TEST_F(VecOpsSimpleTest, MaxTenValuesWithScalar) { XlaBuilder builder(TestName()); - auto x = builder.ConstantR1( - {2.1, -2.6, 2.6, -4.0, 2.1, 2.3, -5.0, -0.9, -2.4, 1.6}); - auto y = builder.ConstantR0(0); - auto max = builder.Max(x, y); + auto x = ConstantR1( + &builder, {2.1, -2.6, 2.6, -4.0, 2.1, 2.3, -5.0, -0.9, -2.4, 1.6}); + auto y = ConstantR0(&builder, 0); + Max(x, y); std::vector expected = {2.1, 0.0, 2.6, 0.0, 2.1, 2.3, 0.0, 0.0, 0.0, 1.6}; @@ -287,11 +247,11 @@ XLA_TEST_F(VecOpsSimpleTest, MaxTenValuesWithScalar) { XLA_TEST_F(VecOpsSimpleTest, MinTenValues) { XlaBuilder builder(TestName()); - auto x = builder.ConstantR1( - {2.1, -2.6, 2.6, -4.0, 2.1, 2.3, -5.0, -0.9, -2.4, 1.6}); - auto y = builder.ConstantR1( - {-0.4, -0.6, -3.0, 0.2, 3.8, -2.2, -1.8, 4.9, 1.4, 0.6}); - auto min = builder.Min(x, y); + auto x = ConstantR1( + &builder, {2.1, -2.6, 2.6, -4.0, 2.1, 2.3, -5.0, -0.9, -2.4, 1.6}); + auto y = ConstantR1( + &builder, {-0.4, -0.6, -3.0, 0.2, 3.8, -2.2, -1.8, 4.9, 1.4, 0.6}); + Min(x, y); std::vector expected = {-0.4, -2.6, -3.0, -4.0, 2.1, -2.2, -5.0, -0.9, -2.4, 0.6}; @@ -300,11 +260,11 @@ XLA_TEST_F(VecOpsSimpleTest, MinTenValues) { XLA_TEST_F(VecOpsSimpleTest, MinMaxTenValues) { XlaBuilder builder(TestName()); - auto zero = builder.ConstantR0(0); - auto one = builder.ConstantR0(1); - auto x = builder.ConstantR1( - {2.1, -2.6, 2.6, 0.3, 3.1, 0.9, -5.0, 0.1, -2.4, 0.6}); - auto clamp = builder.Min(builder.Max(x, zero), one); + auto zero = ConstantR0(&builder, 0); + auto one = ConstantR0(&builder, 1); + auto x = ConstantR1( + &builder, {2.1, -2.6, 2.6, 0.3, 3.1, 0.9, -5.0, 0.1, -2.4, 0.6}); + Min(Max(x, zero), one); std::vector expected = {1.0, 0.0, 1.0, 0.3, 1.0, 0.9, 0.0, 0.1, 0.0, 0.6}; @@ -313,11 +273,11 @@ XLA_TEST_F(VecOpsSimpleTest, MinMaxTenValues) { XLA_TEST_F(VecOpsSimpleTest, ClampTenValuesConstant) { XlaBuilder builder(TestName()); - auto zero = builder.ConstantR0(0); - auto one = builder.ConstantR0(1); - auto x = builder.ConstantR1( - {2.1, -2.6, 2.6, 0.3, 3.1, 0.9, -5.0, 0.1, -2.4, 0.6}); - auto clamp = builder.Clamp(zero, x, one); + auto zero = ConstantR0(&builder, 0); + auto one = ConstantR0(&builder, 1); + auto x = ConstantR1( + &builder, {2.1, -2.6, 2.6, 0.3, 3.1, 0.9, -5.0, 0.1, -2.4, 0.6}); + Clamp(zero, x, one); std::vector expected = {1.0, 0.0, 1.0, 0.3, 1.0, 0.9, 0.0, 0.1, 0.0, 0.6}; @@ -326,10 +286,10 @@ XLA_TEST_F(VecOpsSimpleTest, ClampTenValuesConstant) { XLA_TEST_F(VecOpsSimpleTest, ClampTwoValuesConstant) { XlaBuilder builder(TestName()); - auto zero = builder.ConstantR1({0.0f, 0.0f}); - auto one = builder.ConstantR1({1.0f, 1.0f}); - auto x = builder.ConstantR1({2.1, -2.6}); - auto clamp = builder.Clamp(zero, x, one); + auto zero = ConstantR1(&builder, {0.0f, 0.0f}); + auto one = ConstantR1(&builder, {1.0f, 1.0f}); + auto x = ConstantR1(&builder, {2.1, -2.6}); + Clamp(zero, x, one); std::vector expected = {1.0, 0.0}; ComputeAndCompareR1(&builder, expected, {}); @@ -337,11 +297,11 @@ XLA_TEST_F(VecOpsSimpleTest, ClampTwoValuesConstant) { XLA_TEST_F(VecOpsSimpleTest, ClampTenValuesConstantNonzeroLower) { XlaBuilder builder(TestName()); - auto one = builder.ConstantR0(1); - auto two = builder.ConstantR0(2); - auto x = builder.ConstantR1( - {2.1, -2.6, 2.6, 0.3, 3.1, 0.9, -5.0, 0.1, -2.4, 0.6}); - auto clamp = builder.Clamp(one, x, two); + auto one = ConstantR0(&builder, 1); + auto two = ConstantR0(&builder, 2); + auto x = ConstantR1( + &builder, {2.1, -2.6, 2.6, 0.3, 3.1, 0.9, -5.0, 0.1, -2.4, 0.6}); + Clamp(one, x, two); std::vector expected = {2.0, 1.0, 2.0, 1.0, 2.0, 1.0, 1.0, 1.0, 1.0, 1.0}; @@ -350,10 +310,10 @@ XLA_TEST_F(VecOpsSimpleTest, ClampTenValuesConstantNonzeroLower) { XLA_TEST_F(VecOpsSimpleTest, ClampValuesConstantS64) { XlaBuilder builder(TestName()); - auto zero = builder.ConstantR0(0); - auto one = builder.ConstantR0(10); - auto x = builder.ConstantR1({-3, 3, 9, 13}); - auto clamp = builder.Clamp(zero, x, one); + auto zero = ConstantR0(&builder, 0); + auto one = ConstantR0(&builder, 10); + auto x = ConstantR1(&builder, {-3, 3, 9, 13}); + Clamp(zero, x, one); std::vector expected = {0, 3, 9, 10}; ComputeAndCompareR1(&builder, expected, {}); @@ -365,9 +325,9 @@ XLA_TEST_F(VecOpsSimpleTest, MapTenValues) { // add_half(x) = x + 0.5 XlaBuilder builder("add_half"); auto x_value = - builder.Parameter(0, ShapeUtil::MakeShape(F32, {}), "x_value"); - auto half = builder.ConstantR0(0.5); - builder.Add(x_value, half); + Parameter(&builder, 0, ShapeUtil::MakeShape(F32, {}), "x_value"); + auto half = ConstantR0(&builder, 0.5); + Add(x_value, half); auto computation_status = builder.Build(); ASSERT_IS_OK(computation_status.status()); add_half = computation_status.ConsumeValueOrDie(); @@ -378,9 +338,9 @@ XLA_TEST_F(VecOpsSimpleTest, MapTenValues) { // clamp(y) = clamp<0,5>(y) XlaBuilder builder("clamp"); auto y_value = - builder.Parameter(0, ShapeUtil::MakeShape(F32, {}), "y_value"); - auto zero = builder.ConstantR0(0.0); - auto clamped = builder.Clamp(zero, y_value, builder.ConstantR0(5)); + Parameter(&builder, 0, ShapeUtil::MakeShape(F32, {}), "y_value"); + auto zero = ConstantR0(&builder, 0.0); + Clamp(zero, y_value, ConstantR0(&builder, 5)); auto computation_status = builder.Build(); ASSERT_IS_OK(computation_status.status()); clamp = computation_status.ConsumeValueOrDie(); @@ -391,13 +351,13 @@ XLA_TEST_F(VecOpsSimpleTest, MapTenValues) { // mult_relu_add(z) = clamp(add_half(2 * max(z, 0))) XlaBuilder builder("mult_relu_add"); auto z_value = - builder.Parameter(0, ShapeUtil::MakeShape(F32, {}), "z_value"); - auto zero = builder.ConstantR0(0.0); - auto two = builder.ConstantR0(2.0); - auto max = builder.Max(z_value, zero); - auto mult = builder.Mul(two, max); - auto inner = builder.Map({mult}, add_half, {}); - builder.Map({inner}, clamp, {}); + Parameter(&builder, 0, ShapeUtil::MakeShape(F32, {}), "z_value"); + auto zero = ConstantR0(&builder, 0.0); + auto two = ConstantR0(&builder, 2.0); + auto max = Max(z_value, zero); + auto mult = Mul(two, max); + auto inner = Map(&builder, {mult}, add_half, {}); + Map(&builder, {inner}, clamp, {}); auto computation_status = builder.Build(); ASSERT_IS_OK(computation_status.status()); mult_relu_add = computation_status.ConsumeValueOrDie(); @@ -405,9 +365,9 @@ XLA_TEST_F(VecOpsSimpleTest, MapTenValues) { XlaBuilder builder("map10"); { - auto x = builder.ConstantR1( - {2.1, -21.6, 2.6, -4.0, 2.1, 2.3, -5.0, -0.9, -2.4, 1.6}); - auto activations = builder.Map({x}, mult_relu_add, {0}); + auto x = ConstantR1( + &builder, {2.1, -21.6, 2.6, -4.0, 2.1, 2.3, -5.0, -0.9, -2.4, 1.6}); + Map(&builder, {x}, mult_relu_add, {0}); } std::vector expected = {4.7, 0.5, 5.0, 0.5, 4.7, @@ -417,9 +377,9 @@ XLA_TEST_F(VecOpsSimpleTest, MapTenValues) { XLA_TEST_F(VecOpsSimpleTest, RemainderTenValuesS32) { XlaBuilder builder(TestName()); - auto x = builder.ConstantR1({-5, -4, -3, -2, -1, 0, 1, 2, 3, 4}); - auto y = builder.ConstantR0(3); - builder.Rem(x, y); + auto x = ConstantR1(&builder, {-5, -4, -3, -2, -1, 0, 1, 2, 3, 4}); + auto y = ConstantR0(&builder, 3); + Rem(x, y); std::vector expected = {-2, -1, 0, -2, -1, 0, 1, 2, 0, 1}; ComputeAndCompareR1(&builder, expected, {}); @@ -427,9 +387,9 @@ XLA_TEST_F(VecOpsSimpleTest, RemainderTenValuesS32) { XLA_TEST_F(VecOpsSimpleTest, VectorPredicateEqual) { XlaBuilder builder(TestName()); - auto x = builder.ConstantR1({false, true}); - auto y = builder.ConstantR1({true, false}); - builder.Eq(x, y); + auto x = ConstantR1(&builder, {false, true}); + auto y = ConstantR1(&builder, {true, false}); + Eq(x, y); std::array expected = {{false, false}}; ComputeAndCompareR1(&builder, expected, {}); @@ -437,9 +397,9 @@ XLA_TEST_F(VecOpsSimpleTest, VectorPredicateEqual) { XLA_TEST_F(VecOpsSimpleTest, VectorPredicateNotEqual) { XlaBuilder builder(TestName()); - auto x = builder.ConstantR1({false, true}); - auto y = builder.ConstantR1({true, false}); - builder.Ne(x, y); + auto x = ConstantR1(&builder, {false, true}); + auto y = ConstantR1(&builder, {true, false}); + Ne(x, y); std::array expected = {{true, true}}; ComputeAndCompareR1(&builder, expected, {}); diff --git a/tensorflow/compiler/xla/tests/while_test.cc b/tensorflow/compiler/xla/tests/while_test.cc index c463f3eac55e5b8ab32dc52d5a38e7840241bc58..29befef92e44c4f7673e0c7153efad31d2bbc2b1 100644 --- a/tensorflow/compiler/xla/tests/while_test.cc +++ b/tensorflow/compiler/xla/tests/while_test.cc @@ -22,7 +22,7 @@ limitations under the License. #include "tensorflow/compiler/xla/client/local_client.h" #include "tensorflow/compiler/xla/client/xla_client/xla_builder.h" #include "tensorflow/compiler/xla/client/xla_client/xla_computation.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/service/platform_util.h" #include "tensorflow/compiler/xla/shape_util.h" #include "tensorflow/compiler/xla/status_macros.h" @@ -55,8 +55,8 @@ TEST_F(WhileTest, WhileWithScalarS32Result) { XlaComputation condition; { XlaBuilder builder("condition"); - auto prev = builder.Parameter(0, result_shape, "prev"); - builder.Gt(builder.ConstantR0(5), prev); + auto prev = Parameter(&builder, 0, result_shape, "prev"); + Gt(ConstantR0(&builder, 5), prev); condition = builder.Build().ConsumeValueOrDie(); } @@ -64,16 +64,16 @@ TEST_F(WhileTest, WhileWithScalarS32Result) { XlaComputation body; { XlaBuilder builder("body"); - auto prev = builder.Parameter(0, result_shape, "prev"); - auto input = builder.ConstantR0(1); - builder.Add(input, prev); + auto prev = Parameter(&builder, 0, result_shape, "prev"); + auto input = ConstantR0(&builder, 1); + Add(input, prev); body = builder.Build().ConsumeValueOrDie(); } // Create a While node with computations for the condition and the body. XlaBuilder builder(TestName()); - auto init = builder.ConstantR0(0); - builder.While(condition, body, init); + auto init = ConstantR0(&builder, 0); + While(condition, body, init); ComputeAndCompareR0(&builder, 5, {}); } @@ -91,8 +91,8 @@ TEST_F(WhileTest, WhileWithScalarS64Result) { XlaComputation condition; { XlaBuilder builder("condition"); - auto prev = builder.Parameter(0, result_shape, "prev"); - builder.Gt(builder.ConstantR0(5), prev); + auto prev = Parameter(&builder, 0, result_shape, "prev"); + Gt(ConstantR0(&builder, 5), prev); condition = builder.Build().ConsumeValueOrDie(); } @@ -100,16 +100,16 @@ TEST_F(WhileTest, WhileWithScalarS64Result) { XlaComputation body; { XlaBuilder builder("body"); - auto prev = builder.Parameter(0, result_shape, "prev"); - auto input = builder.ConstantR0(1); - builder.Add(input, prev); + auto prev = Parameter(&builder, 0, result_shape, "prev"); + auto input = ConstantR0(&builder, 1); + Add(input, prev); body = builder.Build().ConsumeValueOrDie(); } // Create a While node with computations for the condition and the body. XlaBuilder builder(TestName()); - auto init = builder.ConstantR0(0); - builder.While(condition, body, init); + auto init = ConstantR0(&builder, 0); + While(condition, body, init); ComputeAndCompareR0(&builder, 5, {}); } @@ -122,8 +122,8 @@ TEST_F(WhileTest, WhileWithScalarResultNonConstInit) { XlaComputation condition; { XlaBuilder builder("condition"); - auto prev = builder.Parameter(0, result_shape, "prev"); - builder.Gt(builder.ConstantR0(5), prev); + auto prev = Parameter(&builder, 0, result_shape, "prev"); + Gt(ConstantR0(&builder, 5), prev); condition = builder.Build().ConsumeValueOrDie(); } @@ -131,18 +131,18 @@ TEST_F(WhileTest, WhileWithScalarResultNonConstInit) { XlaComputation body; { XlaBuilder builder("body"); - auto prev = builder.Parameter(0, result_shape, "prev"); - auto input = builder.ConstantR0(1); - builder.Add(input, prev); + auto prev = Parameter(&builder, 0, result_shape, "prev"); + auto input = ConstantR0(&builder, 1); + Add(input, prev); body = builder.Build().ConsumeValueOrDie(); } // Create a While node with computations for the condition and the body. XlaBuilder builder(TestName()); - auto init = builder.Reduce(builder.ConstantR1(2, 1), - builder.ConstantR0(0), - CreateScalarAddComputation(S32, &builder), {0}); - builder.While(condition, body, init); + auto init = + Reduce(ConstantR1(&builder, 2, 1), ConstantR0(&builder, 0), + CreateScalarAddComputation(S32, &builder), {0}); + While(condition, body, init); ComputeAndCompareR0(&builder, 5, {}); } @@ -154,8 +154,8 @@ TEST_F(WhileTest, WhileWithPredicateResult) { XlaComputation condition; { XlaBuilder builder("condition"); - auto prev = builder.Parameter(0, result_shape, "prev"); - builder.Ne(builder.ConstantR0(true), prev); + auto prev = Parameter(&builder, 0, result_shape, "prev"); + Ne(ConstantR0(&builder, true), prev); condition = builder.Build().ConsumeValueOrDie(); } @@ -163,16 +163,16 @@ TEST_F(WhileTest, WhileWithPredicateResult) { XlaComputation body; { XlaBuilder builder("body"); - auto prev = builder.Parameter(0, result_shape, "prev"); - builder.Or(prev, builder.ConstantR0(true)); + auto prev = Parameter(&builder, 0, result_shape, "prev"); + Or(prev, ConstantR0(&builder, true)); body = builder.Build().ConsumeValueOrDie(); } // Create a While node with computations for the condition and the body. XlaBuilder builder(TestName()); - auto init = builder.Ne(builder.ConstantR0(false), - builder.ConstantR0(true)); - builder.While(condition, body, init); + auto init = + Ne(ConstantR0(&builder, false), ConstantR0(&builder, true)); + While(condition, body, init); ComputeAndCompareR0(&builder, true, {}); } @@ -184,17 +184,16 @@ TEST_F(WhileTest, WhileWithPredicateResult) { // while (result.sum() < 15.5f) { // result = result + vector(0); // } -// TODO(b/29185393): does not terminate on CPU. -TEST_F(WhileTest, DISABLED_WhileWithEmptyVectorResult) { +TEST_F(WhileTest, DISABLED_ON_INTERPRETER(WhileWithEmptyVectorResult)) { Shape result_shape = ShapeUtil::MakeShape(F32, {0}); // Create a computation for the reduction. XlaComputation add; { XlaBuilder builder("add"); - auto x = builder.Parameter(0, ShapeUtil::MakeShape(F32, {}), "x"); - auto y = builder.Parameter(1, ShapeUtil::MakeShape(F32, {}), "y"); - builder.Add(x, y); + auto x = Parameter(&builder, 0, ShapeUtil::MakeShape(F32, {}), "x"); + auto y = Parameter(&builder, 1, ShapeUtil::MakeShape(F32, {}), "y"); + Add(x, y); add = builder.Build().ConsumeValueOrDie(); } @@ -203,10 +202,10 @@ TEST_F(WhileTest, DISABLED_WhileWithEmptyVectorResult) { XlaComputation condition; { XlaBuilder builder("condition"); - auto prev = builder.Parameter(0, result_shape, "prev"); - auto sum = builder.Reduce(prev, builder.ConstantR0(0.0f), add, - /*dimensions_to_reduce=*/{0}); - builder.Gt(builder.ConstantR0(15.5f), sum); + auto prev = Parameter(&builder, 0, result_shape, "prev"); + auto sum = Reduce(prev, ConstantR0(&builder, 0.0f), add, + /*dimensions_to_reduce=*/{0}); + Gt(ConstantR0(&builder, 15.5f), sum); condition = builder.Build().ConsumeValueOrDie(); } @@ -215,16 +214,16 @@ TEST_F(WhileTest, DISABLED_WhileWithEmptyVectorResult) { XlaComputation body; { XlaBuilder builder("body"); - auto prev = builder.Parameter(0, result_shape, "prev"); - auto input = builder.ConstantR1({}); - builder.Add(input, prev); + auto prev = Parameter(&builder, 0, result_shape, "prev"); + auto input = ConstantR1(&builder, {}); + Add(input, prev); body = builder.Build().ConsumeValueOrDie(); } // Create a While node with computations for the condition and the body. XlaBuilder builder("while"); - auto init = builder.ConstantR1({}); - auto result = builder.While(condition, body, init); + auto init = ConstantR1(&builder, {}); + auto result = While(condition, body, init); VLOG(2) << "while = " << ShapeUtil::HumanString( builder.GetShape(result).ConsumeValueOrDie()); @@ -246,9 +245,9 @@ TEST_F(WhileTest, WhileWithVectorResult) { XlaComputation add; { XlaBuilder builder("add"); - auto x = builder.Parameter(0, ShapeUtil::MakeShape(F32, {}), "x"); - auto y = builder.Parameter(1, ShapeUtil::MakeShape(F32, {}), "y"); - builder.Add(x, y); + auto x = Parameter(&builder, 0, ShapeUtil::MakeShape(F32, {}), "x"); + auto y = Parameter(&builder, 1, ShapeUtil::MakeShape(F32, {}), "y"); + Add(x, y); add = builder.Build().ConsumeValueOrDie(); } @@ -257,10 +256,10 @@ TEST_F(WhileTest, WhileWithVectorResult) { XlaComputation condition; { XlaBuilder builder("condition"); - auto prev = builder.Parameter(0, result_shape, "prev"); - auto sum = builder.Reduce(prev, builder.ConstantR0(0.0f), add, - /*dimensions_to_reduce=*/{0}); - builder.Gt(builder.ConstantR0(15.5f), sum); + auto prev = Parameter(&builder, 0, result_shape, "prev"); + auto sum = Reduce(prev, ConstantR0(&builder, 0.0f), add, + /*dimensions_to_reduce=*/{0}); + Gt(ConstantR0(&builder, 15.5f), sum); condition = builder.Build().ConsumeValueOrDie(); } @@ -269,16 +268,16 @@ TEST_F(WhileTest, WhileWithVectorResult) { XlaComputation body; { XlaBuilder builder("body"); - auto prev = builder.Parameter(0, result_shape, "prev"); - auto input = builder.ConstantR1(8, 0.125f); - builder.Add(input, prev); + auto prev = Parameter(&builder, 0, result_shape, "prev"); + auto input = ConstantR1(&builder, 8, 0.125f); + Add(input, prev); body = builder.Build().ConsumeValueOrDie(); } // Create a While node with computations for the condition and the body. XlaBuilder builder("while"); - auto init = builder.ConstantR1(8, 0.f); - auto result = builder.While(condition, body, init); + auto init = ConstantR1(&builder, 8, 0.f); + auto result = While(condition, body, init); VLOG(2) << "while = " << ShapeUtil::HumanString( builder.GetShape(result).ConsumeValueOrDie()); @@ -306,9 +305,9 @@ TEST_F(WhileTest, WhileWithVectorResultIntoTuple) { XlaComputation add; { XlaBuilder builder("add"); - auto x = builder.Parameter(0, ShapeUtil::MakeShape(F32, {}), "x"); - auto y = builder.Parameter(1, ShapeUtil::MakeShape(F32, {}), "y"); - builder.Add(x, y); + auto x = Parameter(&builder, 0, ShapeUtil::MakeShape(F32, {}), "x"); + auto y = Parameter(&builder, 1, ShapeUtil::MakeShape(F32, {}), "y"); + Add(x, y); add = builder.Build().ConsumeValueOrDie(); } @@ -317,10 +316,10 @@ TEST_F(WhileTest, WhileWithVectorResultIntoTuple) { XlaComputation condition; { XlaBuilder builder("condition"); - auto prev = builder.Parameter(0, result_shape, "prev"); - auto sum = builder.Reduce(prev, builder.ConstantR0(0.0f), add, - /*dimensions_to_reduce=*/{0}); - builder.Gt(builder.ConstantR0(15.5f), sum); + auto prev = Parameter(&builder, 0, result_shape, "prev"); + auto sum = Reduce(prev, ConstantR0(&builder, 0.0f), add, + /*dimensions_to_reduce=*/{0}); + Gt(ConstantR0(&builder, 15.5f), sum); condition = builder.Build().ConsumeValueOrDie(); } @@ -329,27 +328,27 @@ TEST_F(WhileTest, WhileWithVectorResultIntoTuple) { XlaComputation body; { XlaBuilder builder("body"); - auto prev = builder.Parameter(0, result_shape, "prev"); - auto input = builder.ConstantR1(8, 0.125f); - builder.Add(input, prev); + auto prev = Parameter(&builder, 0, result_shape, "prev"); + auto input = ConstantR1(&builder, 8, 0.125f); + Add(input, prev); body = builder.Build().ConsumeValueOrDie(); } // Create a While node with computations for the condition and the body. XlaBuilder builder("while"); - auto init = builder.ConstantR1(8, 0.f); - auto result = builder.While(condition, body, init); + auto init = ConstantR1(&builder, 8, 0.f); + auto result = While(condition, body, init); VLOG(2) << "while = " << ShapeUtil::HumanString( builder.GetShape(result).ConsumeValueOrDie()); - builder.Tuple({result}); + Tuple(&builder, {result}); // Individual elements with increase by 1/8 each time through the loop, so // the sum will increase by 1.0. It will first be >15.5 when the elements // have all reached 2.0. auto expected_data = - Literal::CreateR1({2.f, 2.f, 2.f, 2.f, 2.f, 2.f, 2.f, 2.f}); - auto expected = Literal::MakeTuple({expected_data.get()}); + LiteralUtil::CreateR1({2.f, 2.f, 2.f, 2.f, 2.f, 2.f, 2.f, 2.f}); + auto expected = LiteralUtil::MakeTuple({expected_data.get()}); VLOG(2) << "expected = " << ShapeUtil::HumanString(expected->shape()); ComputeAndCompareTuple(&builder, *expected, {}, ErrorSpec(0.0001)); } @@ -366,9 +365,9 @@ TEST_F(WhileTest, WhileWithPermutationAndTupleResult) { XlaComputation condition; { XlaBuilder builder("condition"); - auto prev = builder.Parameter(0, result_shape, "prev"); - auto iteration = builder.GetTupleElement(prev, 0); - builder.Gt(builder.ConstantR0(N), iteration); + auto prev = Parameter(&builder, 0, result_shape, "prev"); + auto iteration = GetTupleElement(prev, 0); + Gt(ConstantR0(&builder, N), iteration); condition = builder.Build().ConsumeValueOrDie(); } @@ -377,32 +376,34 @@ TEST_F(WhileTest, WhileWithPermutationAndTupleResult) { XlaComputation body; { XlaBuilder builder("body"); - auto prev = builder.Parameter(0, result_shape, "prev"); - auto iteration = builder.GetTupleElement(prev, 0); - auto w1 = builder.GetTupleElement(prev, 1); - auto w2 = builder.GetTupleElement(prev, 2); - auto w3 = builder.GetTupleElement(prev, 3); - builder.Tuple( - {builder.Add(iteration, builder.ConstantR0(1)), w3, w1, w2}); + auto prev = Parameter(&builder, 0, result_shape, "prev"); + auto iteration = GetTupleElement(prev, 0); + auto w1 = GetTupleElement(prev, 1); + auto w2 = GetTupleElement(prev, 2); + auto w3 = GetTupleElement(prev, 3); + Tuple(&builder, + {Add(iteration, ConstantR0(&builder, 1)), w3, w1, w2}); body = builder.Build().ConsumeValueOrDie(); } // Create a While node with computations for the condition and the body. XlaBuilder builder("while"); - auto init = builder.Tuple( - {builder.ConstantR0(0), builder.ConstantR1(3, 1.f), - builder.ConstantR1(3, 2.f), builder.ConstantR1(3, 3.f)}); - auto result = builder.While(condition, body, init); + auto init = Tuple(&builder, {ConstantR0(&builder, 0), + ConstantR1(&builder, 3, 1.f), + ConstantR1(&builder, 3, 2.f), + ConstantR1(&builder, 3, 3.f)}); + auto result = While(condition, body, init); VLOG(2) << "result = " << ShapeUtil::HumanString( builder.GetShape(result).ConsumeValueOrDie()); - auto expected_counter = Literal::CreateR0(N); - auto expected_w1 = Literal::CreateR1({1.0f, 1.0f, 1.0f}); - auto expected_w2 = Literal::CreateR1({2.0f, 2.0f, 2.0f}); - auto expected_w3 = Literal::CreateR1({3.0f, 3.0f, 3.0f}); - auto expected = Literal::MakeTuple({expected_counter.get(), expected_w2.get(), - expected_w3.get(), expected_w1.get()}); + auto expected_counter = LiteralUtil::CreateR0(N); + auto expected_w1 = LiteralUtil::CreateR1({1.0f, 1.0f, 1.0f}); + auto expected_w2 = LiteralUtil::CreateR1({2.0f, 2.0f, 2.0f}); + auto expected_w3 = LiteralUtil::CreateR1({3.0f, 3.0f, 3.0f}); + auto expected = + LiteralUtil::MakeTuple({expected_counter.get(), expected_w2.get(), + expected_w3.get(), expected_w1.get()}); VLOG(2) << "expected = " << ShapeUtil::HumanString(expected->shape()); ComputeAndCompareTuple(&builder, *expected, {}, ErrorSpec(0.0001)); } @@ -419,9 +420,9 @@ TEST_F(WhileTest, WhileWithPermutationAndVectorResult) { XlaComputation condition; { XlaBuilder builder("condition"); - auto prev = builder.Parameter(0, result_shape, "prev"); - auto iteration = builder.GetTupleElement(prev, 0); - builder.Gt(builder.ConstantR0(N), iteration); + auto prev = Parameter(&builder, 0, result_shape, "prev"); + auto iteration = GetTupleElement(prev, 0); + Gt(ConstantR0(&builder, N), iteration); condition = builder.Build().ConsumeValueOrDie(); } @@ -430,26 +431,27 @@ TEST_F(WhileTest, WhileWithPermutationAndVectorResult) { XlaComputation body; { XlaBuilder builder("body"); - auto prev = builder.Parameter(0, result_shape, "prev"); - auto iteration = builder.GetTupleElement(prev, 0); - auto w1 = builder.GetTupleElement(prev, 1); - auto w2 = builder.GetTupleElement(prev, 2); - auto w3 = builder.GetTupleElement(prev, 3); - builder.Tuple( - {builder.Add(iteration, builder.ConstantR0(1)), w3, w1, w2}); + auto prev = Parameter(&builder, 0, result_shape, "prev"); + auto iteration = GetTupleElement(prev, 0); + auto w1 = GetTupleElement(prev, 1); + auto w2 = GetTupleElement(prev, 2); + auto w3 = GetTupleElement(prev, 3); + Tuple(&builder, + {Add(iteration, ConstantR0(&builder, 1)), w3, w1, w2}); body = builder.Build().ConsumeValueOrDie(); } // Create a While node with computations for the condition and the body. XlaBuilder builder("while"); - auto init = builder.Tuple( - {builder.ConstantR0(0), builder.ConstantR1(3, 1.f), - builder.ConstantR1(3, 2.f), builder.ConstantR1(3, 3.f)}); - auto xla_while = builder.While(condition, body, init); - - auto add12 = builder.Add(builder.GetTupleElement(xla_while, 1), - builder.GetTupleElement(xla_while, 2)); - auto result = builder.Add(add12, builder.GetTupleElement(xla_while, 3)); + auto init = Tuple(&builder, {ConstantR0(&builder, 0), + ConstantR1(&builder, 3, 1.f), + ConstantR1(&builder, 3, 2.f), + ConstantR1(&builder, 3, 3.f)}); + auto xla_while = While(condition, body, init); + + auto add12 = + Add(GetTupleElement(xla_while, 1), GetTupleElement(xla_while, 2)); + auto result = Add(add12, GetTupleElement(xla_while, 3)); VLOG(2) << "result = " << ShapeUtil::HumanString( builder.GetShape(result).ConsumeValueOrDie()); @@ -474,9 +476,9 @@ TEST_F(WhileTest, WhileWithTupleResult) { XlaComputation condition; { XlaBuilder builder("condition"); - auto prev = builder.Parameter(0, result_shape, "prev"); - auto iteration = builder.GetTupleElement(prev, 0); - builder.Gt(builder.ConstantR0(5), iteration); + auto prev = Parameter(&builder, 0, result_shape, "prev"); + auto iteration = GetTupleElement(prev, 0); + Gt(ConstantR0(&builder, 5), iteration); condition = builder.Build().ConsumeValueOrDie(); } @@ -486,30 +488,30 @@ TEST_F(WhileTest, WhileWithTupleResult) { XlaComputation body; { XlaBuilder builder("body"); - auto prev = builder.Parameter(0, result_shape, "prev"); - auto iteration = builder.GetTupleElement(prev, 0); - auto weights = builder.GetTupleElement(prev, 1); - auto input = builder.ConstantR1(10, 1.f); - auto new_weights = builder.Add(weights, input); - builder.Tuple( - {builder.Add(iteration, builder.ConstantR0(1)), new_weights}); + auto prev = Parameter(&builder, 0, result_shape, "prev"); + auto iteration = GetTupleElement(prev, 0); + auto weights = GetTupleElement(prev, 1); + auto input = ConstantR1(&builder, 10, 1.f); + auto new_weights = Add(weights, input); + Tuple(&builder, + {Add(iteration, ConstantR0(&builder, 1)), new_weights}); body = builder.Build().ConsumeValueOrDie(); } // Create a While node with computations for the condition and the body. XlaBuilder builder("while"); - auto init = builder.Tuple( - {builder.ConstantR0(0), builder.ConstantR1(10, 0.f)}); - auto result = builder.While(condition, body, init); + auto init = Tuple(&builder, {ConstantR0(&builder, 0), + ConstantR1(&builder, 10, 0.f)}); + auto result = While(condition, body, init); VLOG(2) << "while = " << ShapeUtil::HumanString( builder.GetShape(result).ConsumeValueOrDie()); - auto expected_counter = Literal::CreateR0(5); - auto expected_data = Literal::CreateR1( + auto expected_counter = LiteralUtil::CreateR0(5); + auto expected_data = LiteralUtil::CreateR1( {5.0f, 5.0f, 5.0f, 5.0f, 5.0f, 5.0f, 5.0f, 5.0f, 5.0f, 5.0f}); auto expected = - Literal::MakeTuple({expected_counter.get(), expected_data.get()}); + LiteralUtil::MakeTuple({expected_counter.get(), expected_data.get()}); VLOG(2) << "expected = " << ShapeUtil::HumanString(expected->shape()); ComputeAndCompareTuple(&builder, *expected, {}, ErrorSpec(0.0001)); } @@ -524,9 +526,9 @@ TEST_F(WhileTest, WhileWithPredicateTupleResult) { XlaComputation condition; { XlaBuilder builder("condition"); - auto prev = builder.Parameter(0, result_shape, "prev"); - auto iteration = builder.GetTupleElement(prev, 0); - builder.Gt(builder.ConstantR0(5), iteration); + auto prev = Parameter(&builder, 0, result_shape, "prev"); + auto iteration = GetTupleElement(prev, 0); + Gt(ConstantR0(&builder, 5), iteration); condition = builder.Build().ConsumeValueOrDie(); } @@ -535,29 +537,28 @@ TEST_F(WhileTest, WhileWithPredicateTupleResult) { XlaComputation body; { XlaBuilder builder("body"); - auto prev = builder.Parameter(0, result_shape, "prev"); - auto iteration = builder.GetTupleElement(prev, 0); - auto pred = builder.GetTupleElement(prev, 1); - auto new_pred = builder.Or(pred, builder.ConstantR0(true)); - builder.Tuple( - {builder.Add(iteration, builder.ConstantR0(1)), new_pred}); + auto prev = Parameter(&builder, 0, result_shape, "prev"); + auto iteration = GetTupleElement(prev, 0); + auto pred = GetTupleElement(prev, 1); + auto new_pred = Or(pred, ConstantR0(&builder, true)); + Tuple(&builder, {Add(iteration, ConstantR0(&builder, 1)), new_pred}); body = builder.Build().ConsumeValueOrDie(); } // Create a While node with computations for the condition and the body. XlaBuilder builder("while"); - auto init = builder.Tuple({builder.ConstantR0(0), - builder.Ne(builder.ConstantR0(false), - builder.ConstantR0(true))}); - auto result = builder.While(condition, body, init); + auto init = Tuple(&builder, {ConstantR0(&builder, 0), + Ne(ConstantR0(&builder, false), + ConstantR0(&builder, true))}); + auto result = While(condition, body, init); VLOG(2) << "while = " << ShapeUtil::HumanString( builder.GetShape(result).ConsumeValueOrDie()); - auto expected_counter = Literal::CreateR0(5); - auto expected_predicate = Literal::CreateR0(true); - auto expected = - Literal::MakeTuple({expected_counter.get(), expected_predicate.get()}); + auto expected_counter = LiteralUtil::CreateR0(5); + auto expected_predicate = LiteralUtil::CreateR0(true); + auto expected = LiteralUtil::MakeTuple( + {expected_counter.get(), expected_predicate.get()}); ComputeAndCompareTuple(&builder, *expected, {}, ErrorSpec(0)); } @@ -571,9 +572,9 @@ TEST_F(WhileTest, WhileWithTupleConstantScalarResult) { XlaComputation condition; { XlaBuilder builder("condition"); - auto prev = builder.Parameter(0, result_shape, "prev"); - auto iteration = builder.GetTupleElement(prev, 0); - builder.Gt(builder.ConstantR0(5), iteration); + auto prev = Parameter(&builder, 0, result_shape, "prev"); + auto iteration = GetTupleElement(prev, 0); + Gt(ConstantR0(&builder, 5), iteration); condition = builder.Build().ConsumeValueOrDie(); } @@ -583,26 +584,26 @@ TEST_F(WhileTest, WhileWithTupleConstantScalarResult) { XlaComputation body; { XlaBuilder builder("body"); - auto prev = builder.Parameter(0, result_shape, "prev"); - auto iteration = builder.GetTupleElement(prev, 0); - builder.Tuple({builder.Add(iteration, builder.ConstantR0(1)), - builder.ConstantR0(7)}); + auto prev = Parameter(&builder, 0, result_shape, "prev"); + auto iteration = GetTupleElement(prev, 0); + Tuple(&builder, {Add(iteration, ConstantR0(&builder, 1)), + ConstantR0(&builder, 7)}); body = builder.Build().ConsumeValueOrDie(); } // Create a While node with computations for the condition and the body. XlaBuilder builder("while"); - auto init = builder.Tuple( - {builder.ConstantR0(0), builder.ConstantR0(7)}); - auto result = builder.While(condition, body, init); + auto init = Tuple(&builder, {ConstantR0(&builder, 0), + ConstantR0(&builder, 7)}); + auto result = While(condition, body, init); VLOG(2) << "while = " << ShapeUtil::HumanString( builder.GetShape(result).ConsumeValueOrDie()); - auto expected_counter = Literal::CreateR0(5); - auto expected_data = Literal::CreateR0(7); + auto expected_counter = LiteralUtil::CreateR0(5); + auto expected_data = LiteralUtil::CreateR0(7); auto expected = - Literal::MakeTuple({expected_counter.get(), expected_data.get()}); + LiteralUtil::MakeTuple({expected_counter.get(), expected_data.get()}); VLOG(2) << "expected = " << ShapeUtil::HumanString(expected->shape()); ComputeAndCompareTuple(&builder, *expected, {}, ErrorSpec(0.0001)); } @@ -632,9 +633,9 @@ TEST_F(WhileTest, TwoWhileWithTupleResult) { const int c1 = 5; { XlaBuilder builder("condition"); - auto prev = builder.Parameter(0, result_shape, "prev"); - auto iteration = builder.GetTupleElement(prev, 0); - builder.Lt(iteration, builder.ConstantR0(c1)); + auto prev = Parameter(&builder, 0, result_shape, "prev"); + auto iteration = GetTupleElement(prev, 0); + Lt(iteration, ConstantR0(&builder, c1)); TF_ASSERT_OK_AND_ASSIGN(condition, builder.Build()); } @@ -642,9 +643,9 @@ TEST_F(WhileTest, TwoWhileWithTupleResult) { const int c2 = 7; { XlaBuilder builder("condition2"); - auto prev = builder.Parameter(0, result_shape, "prev"); - auto iteration = builder.GetTupleElement(prev, 0); - builder.Lt(iteration, builder.ConstantR0(c2)); + auto prev = Parameter(&builder, 0, result_shape, "prev"); + auto iteration = GetTupleElement(prev, 0); + Lt(iteration, ConstantR0(&builder, c2)); TF_ASSERT_OK_AND_ASSIGN(condition2, builder.Build()); } @@ -654,43 +655,43 @@ TEST_F(WhileTest, TwoWhileWithTupleResult) { XlaComputation body; { XlaBuilder builder("body"); - auto prev = builder.Parameter(0, result_shape, "prev"); - auto iteration = builder.GetTupleElement(prev, 0); - auto weights = builder.GetTupleElement(prev, 1); - auto input = builder.ConstantR1(10, 1.f); - auto new_weights = builder.Add(weights, input); - builder.Tuple( - {builder.Add(iteration, builder.ConstantR0(1)), new_weights}); + auto prev = Parameter(&builder, 0, result_shape, "prev"); + auto iteration = GetTupleElement(prev, 0); + auto weights = GetTupleElement(prev, 1); + auto input = ConstantR1(&builder, 10, 1.f); + auto new_weights = Add(weights, input); + Tuple(&builder, + {Add(iteration, ConstantR0(&builder, 1)), new_weights}); TF_ASSERT_OK_AND_ASSIGN(body, builder.Build()); } XlaComputation body2; { XlaBuilder builder("body"); - auto prev = builder.Parameter(0, result_shape, "prev"); - auto iteration = builder.GetTupleElement(prev, 0); - auto weights = builder.GetTupleElement(prev, 1); - auto input = builder.ConstantR1(10, 1.f); - auto new_weights = builder.Add(weights, input); - builder.Tuple( - {builder.Add(iteration, builder.ConstantR0(1)), new_weights}); + auto prev = Parameter(&builder, 0, result_shape, "prev"); + auto iteration = GetTupleElement(prev, 0); + auto weights = GetTupleElement(prev, 1); + auto input = ConstantR1(&builder, 10, 1.f); + auto new_weights = Add(weights, input); + Tuple(&builder, + {Add(iteration, ConstantR0(&builder, 1)), new_weights}); TF_ASSERT_OK_AND_ASSIGN(body2, builder.Build()); } // Create a While node with computations for the condition and the body. XlaBuilder builder("while"); - auto init = builder.Tuple( - {builder.ConstantR0(0), builder.ConstantR1(10, 0.f)}); - auto while1 = builder.While(condition, body, init); + auto init = Tuple(&builder, {ConstantR0(&builder, 0), + ConstantR1(&builder, 10, 0.f)}); + auto while1 = While(condition, body, init); - auto while2 = builder.While(condition2, body2, while1); + auto while2 = While(condition2, body2, while1); - auto while_result1 = builder.GetTupleElement(while1, 1); - auto while_result2 = builder.GetTupleElement(while2, 1); + auto while_result1 = GetTupleElement(while1, 1); + auto while_result2 = GetTupleElement(while2, 1); VLOG(2) << "while_result2 = " << ShapeUtil::HumanString( builder.GetShape(while_result2).ConsumeValueOrDie()); - auto result = builder.Add(while_result1, while_result2); + auto result = Add(while_result1, while_result2); VLOG(2) << "result = " << ShapeUtil::HumanString( builder.GetShape(result).ConsumeValueOrDie()); @@ -711,9 +712,9 @@ TEST_F(WhileTest, TwoWhileLoopsAndSharedBody) { const int c1 = 5; { XlaBuilder builder("condition"); - auto prev = builder.Parameter(0, result_shape, "prev"); - auto iteration = builder.GetTupleElement(prev, 0); - builder.Lt(iteration, builder.ConstantR0(c1)); + auto prev = Parameter(&builder, 0, result_shape, "prev"); + auto iteration = GetTupleElement(prev, 0); + Lt(iteration, ConstantR0(&builder, c1)); TF_ASSERT_OK_AND_ASSIGN(condition, builder.Build()); } @@ -721,9 +722,9 @@ TEST_F(WhileTest, TwoWhileLoopsAndSharedBody) { const int c2 = 7; { XlaBuilder builder("condition2"); - auto prev = builder.Parameter(0, result_shape, "prev"); - auto iteration = builder.GetTupleElement(prev, 0); - builder.Lt(iteration, builder.ConstantR0(c2)); + auto prev = Parameter(&builder, 0, result_shape, "prev"); + auto iteration = GetTupleElement(prev, 0); + Lt(iteration, ConstantR0(&builder, c2)); TF_ASSERT_OK_AND_ASSIGN(condition2, builder.Build()); } @@ -733,30 +734,30 @@ TEST_F(WhileTest, TwoWhileLoopsAndSharedBody) { XlaComputation body; { XlaBuilder builder("body"); - auto prev = builder.Parameter(0, result_shape, "prev"); - auto iteration = builder.GetTupleElement(prev, 0); - auto weights = builder.GetTupleElement(prev, 1); - auto input = builder.ConstantR1(10, 1.f); - auto new_weights = builder.Add(weights, input); - builder.Tuple( - {builder.Add(iteration, builder.ConstantR0(1)), new_weights}); + auto prev = Parameter(&builder, 0, result_shape, "prev"); + auto iteration = GetTupleElement(prev, 0); + auto weights = GetTupleElement(prev, 1); + auto input = ConstantR1(&builder, 10, 1.f); + auto new_weights = Add(weights, input); + Tuple(&builder, + {Add(iteration, ConstantR0(&builder, 1)), new_weights}); TF_ASSERT_OK_AND_ASSIGN(body, builder.Build()); } // Create a While node with computations for the condition and the body. XlaBuilder builder("while"); - auto init = builder.Tuple( - {builder.ConstantR0(0), builder.ConstantR1(10, 0.f)}); - auto while1 = builder.While(condition, body, init); + auto init = Tuple(&builder, {ConstantR0(&builder, 0), + ConstantR1(&builder, 10, 0.f)}); + auto while1 = While(condition, body, init); - auto while2 = builder.While(condition2, body, while1); + auto while2 = While(condition2, body, while1); - auto while_result1 = builder.GetTupleElement(while1, 1); - auto while_result2 = builder.GetTupleElement(while2, 1); + auto while_result1 = GetTupleElement(while1, 1); + auto while_result2 = GetTupleElement(while2, 1); VLOG(2) << "while_result2 = " << ShapeUtil::HumanString( builder.GetShape(while_result2).ConsumeValueOrDie()); - auto result = builder.Add(while_result1, while_result2); + auto result = Add(while_result1, while_result2); VLOG(2) << "result = " << ShapeUtil::HumanString( builder.GetShape(result).ConsumeValueOrDie()); @@ -778,9 +779,9 @@ TEST_F(WhileTest, DISABLED_ON_GPU(WhileLoopsWithSharedBodyAndInit)) { const int c1 = 5; { XlaBuilder builder("condition"); - auto prev = builder.Parameter(0, result_shape, "prev"); - auto iteration = builder.GetTupleElement(prev, 0); - builder.Lt(iteration, builder.ConstantR0(c1)); + auto prev = Parameter(&builder, 0, result_shape, "prev"); + auto iteration = GetTupleElement(prev, 0); + Lt(iteration, ConstantR0(&builder, c1)); TF_ASSERT_OK_AND_ASSIGN(condition, builder.Build()); } @@ -788,9 +789,9 @@ TEST_F(WhileTest, DISABLED_ON_GPU(WhileLoopsWithSharedBodyAndInit)) { const int c2 = 7; { XlaBuilder builder("condition2"); - auto prev = builder.Parameter(0, result_shape, "prev"); - auto iteration = builder.GetTupleElement(prev, 0); - builder.Lt(iteration, builder.ConstantR0(c2)); + auto prev = Parameter(&builder, 0, result_shape, "prev"); + auto iteration = GetTupleElement(prev, 0); + Lt(iteration, ConstantR0(&builder, c2)); TF_ASSERT_OK_AND_ASSIGN(condition2, builder.Build()); } @@ -800,29 +801,29 @@ TEST_F(WhileTest, DISABLED_ON_GPU(WhileLoopsWithSharedBodyAndInit)) { XlaComputation body; { XlaBuilder builder("body"); - auto prev = builder.Parameter(0, result_shape, "prev"); - auto iteration = builder.GetTupleElement(prev, 0); - auto weights = builder.GetTupleElement(prev, 1); - auto input = builder.ConstantR1(10, 1.f); - auto new_weights = builder.Add(weights, input); - builder.Tuple( - {builder.Add(iteration, builder.ConstantR0(1)), new_weights}); + auto prev = Parameter(&builder, 0, result_shape, "prev"); + auto iteration = GetTupleElement(prev, 0); + auto weights = GetTupleElement(prev, 1); + auto input = ConstantR1(&builder, 10, 1.f); + auto new_weights = Add(weights, input); + Tuple(&builder, + {Add(iteration, ConstantR0(&builder, 1)), new_weights}); TF_ASSERT_OK_AND_ASSIGN(body, builder.Build()); } // Create a While node with computations for the condition and the body. XlaBuilder builder("while"); - auto init = builder.Tuple( - {builder.ConstantR0(0), builder.ConstantR1(10, 0.f)}); - auto while1 = builder.While(condition, body, init); - auto while2 = builder.While(condition2, body, init); + auto init = Tuple(&builder, {ConstantR0(&builder, 0), + ConstantR1(&builder, 10, 0.f)}); + auto while1 = While(condition, body, init); + auto while2 = While(condition2, body, init); - auto while_result1 = builder.GetTupleElement(while1, 1); - auto while_result2 = builder.GetTupleElement(while2, 1); + auto while_result1 = GetTupleElement(while1, 1); + auto while_result2 = GetTupleElement(while2, 1); VLOG(2) << "while_result2 = " << ShapeUtil::HumanString( builder.GetShape(while_result2).ConsumeValueOrDie()); - auto result = builder.Add(while_result1, while_result2); + auto result = Add(while_result1, while_result2); VLOG(2) << "result = " << ShapeUtil::HumanString( builder.GetShape(result).ConsumeValueOrDie()); @@ -844,9 +845,9 @@ XLA_TEST_F(WhileTest, WhileWithDynamicUpdateSlice) { XlaComputation condition; { XlaBuilder builder("condition"); - auto prev = builder.Parameter(0, result_shape, "prev"); - auto iteration = builder.GetTupleElement(prev, 0); - builder.Gt(builder.ConstantR0(5), iteration); + auto prev = Parameter(&builder, 0, result_shape, "prev"); + auto iteration = GetTupleElement(prev, 0); + Gt(ConstantR0(&builder, 5), iteration); condition = builder.Build().ConsumeValueOrDie(); } @@ -856,38 +857,37 @@ XLA_TEST_F(WhileTest, WhileWithDynamicUpdateSlice) { XlaComputation body; { XlaBuilder builder("body"); - auto prev = builder.Parameter(0, result_shape, "prev"); + auto prev = Parameter(&builder, 0, result_shape, "prev"); // TupleElement 0 - auto iteration = builder.GetTupleElement(prev, 0); - auto out0 = builder.Add(iteration, builder.ConstantR0(1)); + auto iteration = GetTupleElement(prev, 0); + auto out0 = Add(iteration, ConstantR0(&builder, 1)); // TupleElement 1 - auto input = builder.GetTupleElement(prev, 1); + auto input = GetTupleElement(prev, 1); // Update. - auto update = builder.ConvertElementType(builder.Broadcast(out0, {2}), F32); + auto update = ConvertElementType(Broadcast(out0, {2}), F32); // Starts = iteration * 2; - auto starts = builder.Reshape( - builder.Mul(iteration, builder.ConstantR0(2)), {1}); + auto starts = Reshape(Mul(iteration, ConstantR0(&builder, 2)), {1}); // UpdateSlice. - auto out1 = builder.DynamicUpdateSlice(input, update, starts); + auto out1 = DynamicUpdateSlice(input, update, starts); - builder.Tuple({out0, out1}); + Tuple(&builder, {out0, out1}); body = builder.Build().ConsumeValueOrDie(); } // Create a While node with computations for the condition and the body. XlaBuilder builder("while"); - auto init = builder.Tuple( - {builder.ConstantR0(0), builder.ConstantR1(10, 0.f)}); - auto result = builder.While(condition, body, init); + auto init = Tuple(&builder, {ConstantR0(&builder, 0), + ConstantR1(&builder, 10, 0.f)}); + auto result = While(condition, body, init); VLOG(2) << "while = " << ShapeUtil::HumanString( builder.GetShape(result).ConsumeValueOrDie()); - auto expected_counter = Literal::CreateR0(5); - auto expected_data = Literal::CreateR1( + auto expected_counter = LiteralUtil::CreateR0(5); + auto expected_data = LiteralUtil::CreateR1( {1.0f, 1.0f, 2.0f, 2.0f, 3.0f, 3.0f, 4.0f, 4.0f, 5.0f, 5.0f}); auto expected = - Literal::MakeTuple({expected_counter.get(), expected_data.get()}); + LiteralUtil::MakeTuple({expected_counter.get(), expected_data.get()}); VLOG(2) << "expected = " << ShapeUtil::HumanString(expected->shape()); ComputeAndCompareTuple(&builder, *expected, {}, ErrorSpec(0.0001)); } @@ -913,10 +913,9 @@ TEST_F(WhileTest, DISABLED_ON_INTERPRETER(WhileWithPrngScalarResult)) { // Create a computation for the condition: repeat for count iterations. auto build_condition = [this, v6s32](int count) { XlaBuilder builder(TestName()); - auto prev = builder.Reshape( - builder.Slice(builder.Parameter(0, v6s32, "prev"), {0}, {1}, {1}), {0}, - {}); - builder.Gt(builder.ConstantR0(count), prev); + auto prev = Reshape( + Slice(Parameter(&builder, 0, v6s32, "prev"), {0}, {1}, {1}), {0}, {}); + Gt(ConstantR0(&builder, count), prev); return builder.Build().ConsumeValueOrDie(); }; @@ -924,22 +923,22 @@ TEST_F(WhileTest, DISABLED_ON_INTERPRETER(WhileWithPrngScalarResult)) { XlaComputation body; { XlaBuilder builder("body"); - auto prev = builder.Parameter(0, v6s32, "prev"); - auto inc = builder.ConcatInDim( - {builder.ConstantR1({1}), - builder.RngUniform(builder.ConstantR0(0), - builder.ConstantR0(100), - ShapeUtil::MakeShape(S32, {5}))}, - 0); - builder.Add(inc, prev); + auto prev = Parameter(&builder, 0, v6s32, "prev"); + auto inc = ConcatInDim(&builder, + {ConstantR1(&builder, {1}), + RngUniform(ConstantR0(&builder, 0), + ConstantR0(&builder, 100), + ShapeUtil::MakeShape(S32, {5}))}, + 0); + Add(inc, prev); body = builder.Build().ConsumeValueOrDie(); } // Create a While node with computations for the condition and the body. auto while_loop = [this, &body, build_condition](int count) { XlaBuilder builder(TestName()); - auto init = builder.ConstantR1({0, 0, 0, 0, 0, 0}); - builder.While(build_condition(count), body, init); + auto init = ConstantR1(&builder, {0, 0, 0, 0, 0, 0}); + While(build_condition(count), body, init); return builder.Build(); }; @@ -958,33 +957,30 @@ TEST_F(WhileTest, WhileThatSwapsParameterWithTupleElement) { auto element_shape = ShapeUtil::MakeShape(F32, {2}); XlaBuilder outer("outer"); - auto p = outer.Parameter(0, element_shape, "param"); - auto t = outer.Tuple({p, outer.ConstantR1({1, 1})}); + auto p = Parameter(&outer, 0, element_shape, "param"); + auto t = Tuple(&outer, {p, ConstantR1(&outer, {1, 1})}); TF_ASSERT_OK_AND_ASSIGN(Shape tuple_shape, outer.GetShape(t)); XlaBuilder cond("cond"); - auto cond_t = cond.Parameter(0, tuple_shape, "t"); - TF_ASSERT_OK(Any(cond.Eq(cond.GetTupleElement(cond_t, 0), - cond.ConstantR1({42, 42})), - &cond) - .status()); + auto cond_t = Parameter(&cond, 0, tuple_shape, "t"); + Any(Eq(GetTupleElement(cond_t, 0), ConstantR1(&cond, {42, 42}))); XlaBuilder body("body"); - auto body_t = body.Parameter(0, tuple_shape, "t"); - auto e = body.GetTupleElement(body_t, 1); - body.Tuple({e, e}); + auto body_t = Parameter(&body, 0, tuple_shape, "t"); + auto e = GetTupleElement(body_t, 1); + Tuple(&body, {e, e}); TF_ASSERT_OK_AND_ASSIGN(auto cond_computation, cond.Build()); TF_ASSERT_OK_AND_ASSIGN(auto body_computation, body.Build()); - outer.While(cond_computation, body_computation, t); + While(cond_computation, body_computation, t); - auto expected_element = Literal::CreateR1({1, 1}); + auto expected_element = LiteralUtil::CreateR1({1, 1}); auto expected = - Literal::MakeTuple({expected_element.get(), expected_element.get()}); + LiteralUtil::MakeTuple({expected_element.get(), expected_element.get()}); TF_ASSERT_OK_AND_ASSIGN( std::unique_ptr parameter_data, - client_->TransferToServer(*Literal::CreateR1({42, 42}))); + client_->TransferToServer(*LiteralUtil::CreateR1({42, 42}))); ComputeAndCompareTuple(&outer, *expected, {parameter_data.get()}, ErrorSpec(1e-6)); } @@ -993,24 +989,23 @@ TEST_F(WhileTest, WhileThatSwapsParameterWithBroadcast) { auto element_shape = ShapeUtil::MakeShape(F32, {2}); XlaBuilder outer("outer"); - auto p = outer.Parameter(0, element_shape, "param"); + auto p = Parameter(&outer, 0, element_shape, "param"); XlaBuilder cond("cond"); - auto cond_t = cond.Parameter(0, element_shape, "t"); - TF_ASSERT_OK( - Any(cond.Eq(cond_t, cond.ConstantR1({42, 42})), &cond).status()); + auto cond_t = Parameter(&cond, 0, element_shape, "t"); + Any(Eq(cond_t, ConstantR1(&cond, {42, 42}))); XlaBuilder body("body"); - auto body_t = body.Parameter(0, element_shape, "t"); - auto e = body.Broadcast(body.ConstantR0(1.0), {2}); + Parameter(&body, 0, element_shape, "t"); + Broadcast(ConstantR0(&body, 1.0), {2}); TF_ASSERT_OK_AND_ASSIGN(auto cond_computation, cond.Build()); TF_ASSERT_OK_AND_ASSIGN(auto body_computation, body.Build()); - outer.While(cond_computation, body_computation, p); + While(cond_computation, body_computation, p); TF_ASSERT_OK_AND_ASSIGN( std::unique_ptr parameter_data, - client_->TransferToServer(*Literal::CreateR1({42, 42}))); + client_->TransferToServer(*LiteralUtil::CreateR1({42, 42}))); ComputeAndCompareR1(&outer, {1.0f, 1.0f}, {parameter_data.get()}, ErrorSpec(1e-6)); } @@ -1019,25 +1014,24 @@ TEST_F(WhileTest, WhileThatTurnsScalarParameterToTupleElement) { auto element_shape = ShapeUtil::MakeShape(F32, {}); XlaBuilder outer("outer"); - auto p = outer.Parameter(0, element_shape, "param"); + auto p = Parameter(&outer, 0, element_shape, "param"); XlaBuilder cond("cond"); - auto cond_t = cond.Parameter(0, element_shape, "t"); - cond.Eq(cond_t, cond.ConstantR0(42)); + auto cond_t = Parameter(&cond, 0, element_shape, "t"); + Eq(cond_t, ConstantR0(&cond, 42)); XlaBuilder body("body"); - auto body_t = body.Parameter(0, element_shape, "t"); - auto tuple = - body.Tuple({body_t, body.Add(body_t, body.ConstantR0(1))}); - auto e = body.GetTupleElement(tuple, 1); + auto body_t = Parameter(&body, 0, element_shape, "t"); + auto tuple = Tuple(&body, {body_t, Add(body_t, ConstantR0(&body, 1))}); + GetTupleElement(tuple, 1); TF_ASSERT_OK_AND_ASSIGN(auto cond_computation, cond.Build()); TF_ASSERT_OK_AND_ASSIGN(auto body_computation, body.Build()); - outer.While(cond_computation, body_computation, p); + While(cond_computation, body_computation, p); TF_ASSERT_OK_AND_ASSIGN( std::unique_ptr parameter_data, - client_->TransferToServer(*Literal::CreateR0(42))); + client_->TransferToServer(*LiteralUtil::CreateR0(42))); ComputeAndCompareR0(&outer, 43.0f, {parameter_data.get()}, ErrorSpec(1e-6)); } @@ -1056,33 +1050,31 @@ TEST_F(WhileTest, WhileWithMixedTupleElements) { XlaBuilder outer("outer"); auto p = - outer.Tuple({outer.ConstantR0(0), - outer.Parameter(0, ShapeUtil::MakeShape(S32, {}), "t")}); + Tuple(&outer, {ConstantR0(&outer, 0), + Parameter(&outer, 0, ShapeUtil::MakeShape(S32, {}), "t")}); XlaBuilder cond("cond"); - auto params = cond.Parameter(0, result_shape, "prev"); - auto cond_t = cond.Add(cond.GetTupleElement(params, 1), - cond.GetTupleElement(params, 0)); - cond.Lt(cond_t, cond.ConstantR0(30)); + auto params = Parameter(&cond, 0, result_shape, "prev"); + auto cond_t = Add(GetTupleElement(params, 1), GetTupleElement(params, 0)); + Lt(cond_t, ConstantR0(&cond, 30)); XlaBuilder body("body"); - auto body_t = body.Parameter(0, result_shape, "t"); + auto body_t = Parameter(&body, 0, result_shape, "t"); - auto tuple = body.Tuple( - {body.Add(body.GetTupleElement(body_t, 0), body.ConstantR0(1)), - body.Add(body.GetTupleElement(body_t, 1), body.ConstantR0(1))}); + Tuple(&body, {Add(GetTupleElement(body_t, 0), ConstantR0(&body, 1)), + Add(GetTupleElement(body_t, 1), ConstantR0(&body, 1))}); TF_ASSERT_OK_AND_ASSIGN(auto cond_computation, cond.Build()); TF_ASSERT_OK_AND_ASSIGN(auto body_computation, body.Build()); - outer.While(cond_computation, body_computation, p); + While(cond_computation, body_computation, p); TF_ASSERT_OK_AND_ASSIGN( std::unique_ptr parameter_data, - client_->TransferToServer(*Literal::CreateR0(1))); + client_->TransferToServer(*LiteralUtil::CreateR0(1))); - auto add1 = Literal::CreateR0(15); - auto add2 = Literal::CreateR0(16); - auto expected = Literal::MakeTuple({add1.get(), add2.get()}); + auto add1 = LiteralUtil::CreateR0(15); + auto add2 = LiteralUtil::CreateR0(16); + auto expected = LiteralUtil::MakeTuple({add1.get(), add2.get()}); ComputeAndCompareTuple(&outer, *expected, {parameter_data.get()}, ErrorSpec(1e-6)); } @@ -1105,9 +1097,9 @@ XLA_TEST_F(WhileTest, NestedWhileWithScalarResult) { XlaComputation inner_condition; { XlaBuilder builder("inner_condition"); - auto params = builder.Parameter(0, inner_result_shape, "prev"); - auto i = builder.GetTupleElement(params, 0); - builder.Lt(i, builder.ConstantR0(7)); + auto params = Parameter(&builder, 0, inner_result_shape, "prev"); + auto i = GetTupleElement(params, 0); + Lt(i, ConstantR0(&builder, 7)); inner_condition = builder.Build().ConsumeValueOrDie(); } @@ -1116,8 +1108,8 @@ XLA_TEST_F(WhileTest, NestedWhileWithScalarResult) { XlaComputation outer_condition; { XlaBuilder builder("outer_condition"); - auto prev = builder.Parameter(0, outer_result_shape, "prev"); - builder.Lt(prev, builder.ConstantR0(30)); + auto prev = Parameter(&builder, 0, outer_result_shape, "prev"); + Lt(prev, ConstantR0(&builder, 30)); outer_condition = builder.Build().ConsumeValueOrDie(); } @@ -1126,12 +1118,12 @@ XLA_TEST_F(WhileTest, NestedWhileWithScalarResult) { XlaComputation inner_body; { XlaBuilder builder("inner_body"); - auto params = builder.Parameter(0, inner_result_shape, "prev"); - auto i = builder.GetTupleElement(params, 0); - auto result = builder.GetTupleElement(params, 1); - i = builder.Add(builder.ConstantR0(1), i); - result = builder.Add(builder.ConstantR0(2), result); - builder.Tuple({i, result}); + auto params = Parameter(&builder, 0, inner_result_shape, "prev"); + auto i = GetTupleElement(params, 0); + auto result = GetTupleElement(params, 1); + i = Add(ConstantR0(&builder, 1), i); + result = Add(ConstantR0(&builder, 2), result); + Tuple(&builder, {i, result}); inner_body = builder.Build().ConsumeValueOrDie(); } @@ -1139,17 +1131,17 @@ XLA_TEST_F(WhileTest, NestedWhileWithScalarResult) { XlaComputation outer_body; { XlaBuilder builder("outer_body"); - auto prev = builder.Parameter(0, outer_result_shape, "prev"); - auto init = builder.Tuple({builder.ConstantR0(0), prev}); - auto result = builder.While(inner_condition, inner_body, init); - builder.GetTupleElement(result, 1); + auto prev = Parameter(&builder, 0, outer_result_shape, "prev"); + auto init = Tuple(&builder, {ConstantR0(&builder, 0), prev}); + auto result = While(inner_condition, inner_body, init); + GetTupleElement(result, 1); outer_body = builder.Build().ConsumeValueOrDie(); } // Create a While node with computations for the condition and the body. XlaBuilder builder(TestName()); - auto init = builder.ConstantR0(0); - builder.While(outer_condition, outer_body, init); + auto init = ConstantR0(&builder, 0); + While(outer_condition, outer_body, init); ComputeAndCompareR0(&builder, 42, {}); } @@ -1167,8 +1159,8 @@ TEST_F(WhileTest, DISABLED_ON_INTERPRETER(WhileWithCallInsideCondition)) { XlaComputation condition_callee; { XlaBuilder builder("condition_callee"); - auto prev = builder.Parameter(0, result_shape, "prev"); - builder.Tuple({builder.Gt(builder.ConstantR0(5), prev)}); + auto prev = Parameter(&builder, 0, result_shape, "prev"); + Tuple(&builder, {Gt(ConstantR0(&builder, 5), prev)}); condition_callee = builder.Build().ConsumeValueOrDie(); } @@ -1176,9 +1168,9 @@ TEST_F(WhileTest, DISABLED_ON_INTERPRETER(WhileWithCallInsideCondition)) { XlaComputation condition; { XlaBuilder builder("condition"); - auto prev = builder.Parameter(0, result_shape, "prev"); - auto result = builder.Call(condition_callee, {prev}); - builder.GetTupleElement(result, 0); + auto prev = Parameter(&builder, 0, result_shape, "prev"); + auto result = Call(&builder, condition_callee, {prev}); + GetTupleElement(result, 0); condition = builder.Build().ConsumeValueOrDie(); } @@ -1186,16 +1178,16 @@ TEST_F(WhileTest, DISABLED_ON_INTERPRETER(WhileWithCallInsideCondition)) { XlaComputation body; { XlaBuilder builder("body"); - auto prev = builder.Parameter(0, result_shape, "prev"); - auto input = builder.ConstantR0(1); - builder.Add(input, prev); + auto prev = Parameter(&builder, 0, result_shape, "prev"); + auto input = ConstantR0(&builder, 1); + Add(input, prev); body = builder.Build().ConsumeValueOrDie(); } // Create a While node with computations for the condition and the body. XlaBuilder builder(TestName()); - auto init = builder.ConstantR0(0); - builder.While(condition, body, init); + auto init = ConstantR0(&builder, 0); + While(condition, body, init); ComputeAndCompareR0(&builder, 5, {}); } @@ -1210,34 +1202,34 @@ TEST_F(WhileTest, WhileWithLoopInvariantOperation) { XlaComputation condition; { XlaBuilder builder("condition"); - auto state = builder.Parameter(0, while_shape, "state"); - builder.Gt(builder.ConstantR0(5), builder.GetTupleElement(state, 0)); + auto state = Parameter(&builder, 0, while_shape, "state"); + Gt(ConstantR0(&builder, 5), GetTupleElement(state, 0)); TF_ASSERT_OK_AND_ASSIGN(condition, builder.Build()); } XlaComputation body; { XlaBuilder builder("body"); - auto state = builder.Parameter(0, while_shape, "state"); - auto indvar = builder.GetTupleElement(state, 0); - auto input_0 = builder.GetTupleElement(state, 1); - auto input_1 = builder.GetTupleElement(state, 2); - auto output = builder.Tanh(builder.Dot(input_0, input_1)); - auto indvar_next = builder.Add(indvar, builder.ConstantR0(1)); - builder.Tuple({indvar_next, input_0, input_1, output}); + auto state = Parameter(&builder, 0, while_shape, "state"); + auto indvar = GetTupleElement(state, 0); + auto input_0 = GetTupleElement(state, 1); + auto input_1 = GetTupleElement(state, 2); + auto output = Tanh(Dot(input_0, input_1)); + auto indvar_next = Add(indvar, ConstantR0(&builder, 1)); + Tuple(&builder, {indvar_next, input_0, input_1, output}); TF_ASSERT_OK_AND_ASSIGN(body, builder.Build()); } XlaBuilder builder(TestName()); - auto matrix_input = builder.Parameter(0, matrix_shape, "matrix"); - auto init = builder.Tuple( - {builder.ConstantR0(0), matrix_input, matrix_input, matrix_input}); - auto while_instruction = builder.While(condition, body, init); - builder.GetTupleElement(while_instruction, 3); + auto matrix_input = Parameter(&builder, 0, matrix_shape, "matrix"); + auto init = Tuple(&builder, {ConstantR0(&builder, 0), matrix_input, + matrix_input, matrix_input}); + auto while_instruction = While(condition, body, init); + GetTupleElement(while_instruction, 3); - TF_ASSERT_OK_AND_ASSIGN(auto param_value, - client_->TransferToServer(*Literal::CreateR2( - {{1.0, 2.0}, {-1.0, -2.0}}))); + TF_ASSERT_OK_AND_ASSIGN( + auto param_value, client_->TransferToServer(*LiteralUtil::CreateR2( + {{1.0, 2.0}, {-1.0, -2.0}}))); ComputeAndCompareR2( &builder, {{-0.76159416, -0.96402758}, {0.76159416, 0.96402758}}, @@ -1264,9 +1256,9 @@ void BM_WhileLoop(int num_iters) { XlaComputation condition; { XlaBuilder builder("condition"); - auto prev = builder.Parameter(0, loop_state_shape, "prev"); - auto iteration = builder.GetTupleElement(prev, 0); - builder.Lt(iteration, builder.ConstantR0(loop_limit)); + auto prev = Parameter(&builder, 0, loop_state_shape, "prev"); + auto iteration = GetTupleElement(prev, 0); + Lt(iteration, ConstantR0(&builder, loop_limit)); condition = builder.Build().ConsumeValueOrDie(); } @@ -1274,29 +1266,29 @@ void BM_WhileLoop(int num_iters) { XlaComputation body; { XlaBuilder builder("body"); - auto prev = builder.Parameter(0, loop_state_shape, "prev"); + auto prev = Parameter(&builder, 0, loop_state_shape, "prev"); // TupleElement 0 - auto iteration = builder.GetTupleElement(prev, 0); - auto out0 = builder.Add(iteration, builder.ConstantR0(1)); + auto iteration = GetTupleElement(prev, 0); + auto out0 = Add(iteration, ConstantR0(&builder, 1)); // TupleElement 1 - auto input = builder.GetTupleElement(prev, 1); + auto input = GetTupleElement(prev, 1); // Update. - auto one = builder.ConstantR0(1.0); - auto update = builder.Broadcast(one, {1, 1024, 1024}); + auto one = ConstantR0(&builder, 1.0); + auto update = Broadcast(one, {1, 1024, 1024}); // Starts = iteration * 2; - auto starts = builder.ConstantR1({0, 0, 0}); + auto starts = ConstantR1(&builder, {0, 0, 0}); // UpdateSlice. - auto out1 = builder.DynamicUpdateSlice(input, update, starts); - builder.Tuple({out0, out1}); + auto out1 = DynamicUpdateSlice(input, update, starts); + Tuple(&builder, {out0, out1}); body = builder.Build().ConsumeValueOrDie(); } // Create a While instruction. XlaBuilder builder("while"); - auto zero = builder.ConstantR0(0.0); - auto input = builder.Broadcast(zero, {seq_len, 1024, 1024}); - auto init = builder.Tuple({builder.ConstantR0(0), input}); - builder.While(condition, body, init); + auto zero = ConstantR0(&builder, 0.0); + auto input = Broadcast(zero, {seq_len, 1024, 1024}); + auto init = Tuple(&builder, {ConstantR0(&builder, 0), input}); + While(condition, body, init); auto computation = builder.Build().ConsumeValueOrDie(); std::unique_ptr executable = diff --git a/tensorflow/compiler/xla/tests/xla_hlo_profile_test.cc b/tensorflow/compiler/xla/tests/xla_hlo_profile_test.cc index 0be950cacbf07eece9aff9ffe1d0e571e9b25038..4d4dd62a3f0426342012b4999c73891c0c601052 100644 --- a/tensorflow/compiler/xla/tests/xla_hlo_profile_test.cc +++ b/tensorflow/compiler/xla/tests/xla_hlo_profile_test.cc @@ -79,7 +79,9 @@ struct ParsedProfileOutputLine { Status ParseOneProfileOutputLine( const string& line, bool expect_hlo, - gtl::FlatMap* parsed_results) { + gtl::FlatMap* parsed_results, + tensorflow::gtl::ArraySlice opcodes_to_ignore = + {}) { string separator = "[^:]*:: +"; string match_percentage = "\\d+\\.\\d\\d%"; string match_cycles = "(\\d+) cycles +\\( *(" + match_percentage + ")\\)"; @@ -113,7 +115,9 @@ Status ParseOneProfileOutputLine( ", Regexp: ", regexp_pattern); } - InsertOrDie(parsed_results, parsed_line.opcode, parsed_line); + if (!c_linear_search(opcodes_to_ignore, parsed_line.opcode)) { + InsertOrDie(parsed_results, parsed_line.opcode, parsed_line); + } return Status::OK(); } @@ -187,9 +191,9 @@ XLA_TEST_F(HloProfileTest, ProfileSingleComputation) { ClientLibrary::GetOrCreateLocalClient(platform)); XlaBuilder builder(TestName()); - auto result = builder.Tanh(builder.Add( - builder.Parameter(0, ShapeUtil::MakeShape(F32, {m, k}), "dot_lhs"), - builder.Parameter(1, ShapeUtil::MakeShape(F32, {k, n}), "dot_rhs"))); + Tanh(Add( + Parameter(&builder, 0, ShapeUtil::MakeShape(F32, {m, k}), "dot_lhs"), + Parameter(&builder, 1, ShapeUtil::MakeShape(F32, {k, n}), "dot_rhs"))); TF_ASSERT_OK_AND_ASSIGN(auto computation, builder.Build()); @@ -239,9 +243,7 @@ XLA_TEST_F(HloProfileTest, ProfileSingleComputation) { EXPECT_TRUE(HasTrops(tanh_profile)); } -// TODO(b/71544591): The GPU backend does not record cycles spent in on Hlo -// instructions "interior" to while nodes. -XLA_TEST_F(HloProfileTest, DISABLED_ON_GPU(ProfileWhileComputation)) { +XLA_TEST_F(HloProfileTest, ProfileWhileComputation) { const int64 size = 256; Shape matrix_shape = ShapeUtil::MakeShape(F32, {size, size}); Shape while_result_shape = @@ -255,30 +257,30 @@ XLA_TEST_F(HloProfileTest, DISABLED_ON_GPU(ProfileWhileComputation)) { XlaComputation condition; { XlaBuilder builder("condition"); - auto state = builder.Parameter(0, while_result_shape, "state"); - auto iteration = builder.GetTupleElement(state, 0); - builder.Gt(builder.ConstantR0(5), iteration); + auto state = Parameter(&builder, 0, while_result_shape, "state"); + auto iteration = GetTupleElement(state, 0); + Gt(ConstantR0(&builder, 5), iteration); TF_ASSERT_OK_AND_ASSIGN(condition, builder.Build()); } XlaComputation body; { XlaBuilder builder("body"); - auto state = builder.Parameter(0, while_result_shape, "state"); - auto matrix = builder.GetTupleElement(state, 1); - auto next_iteration = builder.Add(builder.GetTupleElement(state, 0), - builder.ConstantR0(1)); - builder.Tuple({next_iteration, builder.Add(matrix, matrix)}); + auto state = Parameter(&builder, 0, while_result_shape, "state"); + auto matrix = GetTupleElement(state, 1); + auto next_iteration = + Add(GetTupleElement(state, 0), ConstantR0(&builder, 1)); + Tuple(&builder, {next_iteration, Mul(matrix, matrix)}); TF_ASSERT_OK_AND_ASSIGN(body, builder.Build()); } XlaBuilder builder(TestName()); auto initial_while_state = - builder.Tuple({builder.ConstantR0(0), - builder.Parameter(0, matrix_shape, "initial_value")}); - auto while_result = builder.While(condition, body, initial_while_state); - builder.Add(builder.GetTupleElement(while_result, 1), - builder.Parameter(1, matrix_shape, "other_value")); + Tuple(&builder, {ConstantR0(&builder, 0), + Parameter(&builder, 0, matrix_shape, "initial_value")}); + auto while_result = While(condition, body, initial_while_state); + Add(GetTupleElement(while_result, 1), + Parameter(&builder, 1, matrix_shape, "other_value")); TF_ASSERT_OK_AND_ASSIGN(auto computation, builder.Build()); @@ -290,36 +292,50 @@ XLA_TEST_F(HloProfileTest, DISABLED_ON_GPU(ProfileWhileComputation)) { tensorflow::str_util::Split(profile_output, '\n'); auto while_body_profile_start = - std::find_if(profile_output_lines.begin(), profile_output_lines.end(), + c_find_if(profile_output_lines, [](tensorflow::StringPiece s) { + return tensorflow::str_util::StartsWith(s, + "Execution profile for body"); + }); + + ASSERT_NE(while_body_profile_start, profile_output_lines.cend()); + + auto while_body_profile_end = + std::find_if(while_body_profile_start, profile_output_lines.end(), [](tensorflow::StringPiece s) { return tensorflow::str_util::StartsWith( - s, "Execution profile for body"); + s, "********** microseconds report **********"); }); - ASSERT_NE(while_body_profile_start, profile_output_lines.end()); + // We emit a blank line before the "********** microseconds report **********" + // line. + while_body_profile_end--; - gtl::FlatMap parsed_profile_lines; + ASSERT_NE(while_body_profile_end, profile_output_lines.end()); - TF_ASSERT_OK( - ParseOneProfileOutputLine(*std::next(while_body_profile_start, 1), - /*expect_hlo=*/false, &parsed_profile_lines)); + gtl::FlatMap parsed_profile_lines; - TF_ASSERT_OK( - ParseOneProfileOutputLine(*std::next(while_body_profile_start, 2), - /*expect_hlo=*/true, &parsed_profile_lines)); + for (auto while_body_profile_i = while_body_profile_start + 1; + while_body_profile_i != while_body_profile_end; while_body_profile_i++) { + // There are multiple "get-tuple-element" instructions in the while body so + // we ignore them -- we don't want parsed_profile_lines to be a multi-map. + TF_ASSERT_OK(ParseOneProfileOutputLine( + *while_body_profile_i, + /*expect_hlo=*/while_body_profile_i != (while_body_profile_start + 1), + &parsed_profile_lines, {"get-tuple-element"})); + } TF_ASSERT_OK_AND_ASSIGN(ParsedProfileOutputLine total_while_body_profile, MaybeFind(parsed_profile_lines, "[total]")); - TF_ASSERT_OK_AND_ASSIGN(ParsedProfileOutputLine dot_profile, - MaybeFind(parsed_profile_lines, "add")); + TF_ASSERT_OK_AND_ASSIGN(ParsedProfileOutputLine multiply_profile, + MaybeFind(parsed_profile_lines, "multiply")); EXPECT_GT(total_while_body_profile.cycles, 0); EXPECT_EQ(total_while_body_profile.opcode, "[total]"); EXPECT_EQ(total_while_body_profile.cycles_percentage, "100.00%"); - EXPECT_GT(total_while_body_profile.cycles, dot_profile.cycles); - EXPECT_NE(dot_profile.cycles_percentage, "0.00%"); - EXPECT_NE(dot_profile.cycles_percentage, "100.00%"); + EXPECT_GT(total_while_body_profile.cycles, multiply_profile.cycles); + EXPECT_NE(multiply_profile.cycles_percentage, "0.00%"); + EXPECT_NE(multiply_profile.cycles_percentage, "100.00%"); } } // namespace } // namespace xla @@ -336,8 +352,11 @@ static std::pair AddXlaHloProfileFlag(int argc, char** argv) { new_argv[argc] = strdup("--xla_hlo_profile"); // Fusion can change the Hlo instructions that show up in the final Hlo - // executable, so block it here. - new_argv[argc + 1] = strdup("--xla_disable_hlo_passes=fusion"); + // executable, so block it here. Also block the WhileLoopInvariantCodeMotion + // pass, otherwise a while loop is transformed and we could not match the + // original name in the ProfileWhileComputation test. + new_argv[argc + 1] = strdup( + "--xla_disable_hlo_passes=fusion,while-loop-invariant-code-motion"); return {argc + 2, new_argv}; } diff --git a/tensorflow/compiler/xla/text_literal_reader.cc b/tensorflow/compiler/xla/text_literal_reader.cc index 56702feab9a4e8d00df3a165ab994aef2d42d830..897123d7606db60abc1105b03beb3f23ab249579 100644 --- a/tensorflow/compiler/xla/text_literal_reader.cc +++ b/tensorflow/compiler/xla/text_literal_reader.cc @@ -20,7 +20,7 @@ limitations under the License. #include #include -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/ptr_util.h" #include "tensorflow/compiler/xla/shape_util.h" #include "tensorflow/compiler/xla/status_macros.h" diff --git a/tensorflow/compiler/xla/text_literal_reader.h b/tensorflow/compiler/xla/text_literal_reader.h index e45e5291c9b10803f5e5008b72c7dd0116a0dea0..708e8c80d8b5c09454eb64d4e12df51a5b7ea628 100644 --- a/tensorflow/compiler/xla/text_literal_reader.h +++ b/tensorflow/compiler/xla/text_literal_reader.h @@ -18,7 +18,7 @@ limitations under the License. #include -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/statusor.h" #include "tensorflow/compiler/xla/types.h" #include "tensorflow/compiler/xla/xla_data.pb.h" diff --git a/tensorflow/compiler/xla/text_literal_reader_test.cc b/tensorflow/compiler/xla/text_literal_reader_test.cc index 23070b663870a2b78b38663e09a32fcb28d9c2dc..92f9b4f9f0efa2dc08287bdcbefc88f879164308 100644 --- a/tensorflow/compiler/xla/text_literal_reader_test.cc +++ b/tensorflow/compiler/xla/text_literal_reader_test.cc @@ -17,7 +17,7 @@ limitations under the License. #include -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/shape_util.h" #include "tensorflow/compiler/xla/test.h" #include "tensorflow/compiler/xla/types.h" diff --git a/tensorflow/compiler/xla/text_literal_writer.cc b/tensorflow/compiler/xla/text_literal_writer.cc index 373c0d2d8d8ab05dec11e51f265d41b91e7920bf..24e0784741a4c9779b0adb7a7740c3d6e2fb033a 100644 --- a/tensorflow/compiler/xla/text_literal_writer.cc +++ b/tensorflow/compiler/xla/text_literal_writer.cc @@ -17,7 +17,7 @@ limitations under the License. #include -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/shape_util.h" #include "tensorflow/compiler/xla/status_macros.h" #include "tensorflow/compiler/xla/types.h" diff --git a/tensorflow/compiler/xla/text_literal_writer.h b/tensorflow/compiler/xla/text_literal_writer.h index 0a1235b5e04675da0f412bafab6c4ecf04367787..159ac1b7e1b6f9c07dac795fb640cd0b2d284bcb 100644 --- a/tensorflow/compiler/xla/text_literal_writer.h +++ b/tensorflow/compiler/xla/text_literal_writer.h @@ -16,7 +16,7 @@ limitations under the License. #ifndef TENSORFLOW_COMPILER_XLA_TEXT_LITERAL_WRITER_H_ #define TENSORFLOW_COMPILER_XLA_TEXT_LITERAL_WRITER_H_ -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/types.h" #include "tensorflow/compiler/xla/xla_data.pb.h" #include "tensorflow/core/lib/core/status.h" diff --git a/tensorflow/compiler/xla/text_literal_writer_test.cc b/tensorflow/compiler/xla/text_literal_writer_test.cc index 70cf2fb1b8a1b4f2ecfdaeaef3a00ddc974e2652..4ea02faffcd52065b05c0444202bd1a3d9d87ee6 100644 --- a/tensorflow/compiler/xla/text_literal_writer_test.cc +++ b/tensorflow/compiler/xla/text_literal_writer_test.cc @@ -18,6 +18,7 @@ limitations under the License. #include #include +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/literal_util.h" #include "tensorflow/compiler/xla/test.h" #include "tensorflow/compiler/xla/test_helpers.h" @@ -30,8 +31,9 @@ namespace xla { namespace { TEST(TextLiteralWriterTest, WritesFloatLiteral) { - auto literal = Literal::CreateR2({ - {3.14, 2.17}, {1.23, 4.56}, + auto literal = LiteralUtil::CreateR2({ + {3.14, 2.17}, + {1.23, 4.56}, }); string path = tensorflow::io::JoinPath(tensorflow::testing::TmpDir(), "/whatever"); diff --git a/tensorflow/compiler/xla/tools/BUILD b/tensorflow/compiler/xla/tools/BUILD index e4a052c8f1c0009619c3a94606f6384d04006e4e..55501827f29582324ce3308f2a7d96bc20b65760 100644 --- a/tensorflow/compiler/xla/tools/BUILD +++ b/tensorflow/compiler/xla/tools/BUILD @@ -74,7 +74,7 @@ cc_library( srcs = ["replay_computation.cc"], deps = [ "//tensorflow/compiler/xla:execution_options_util", - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:shape_util", "//tensorflow/compiler/xla:status_macros", "//tensorflow/compiler/xla:statusor", @@ -123,7 +123,7 @@ tf_cc_binary( name = "show_literal", srcs = ["show_literal.cc"], deps = [ - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:types", "//tensorflow/compiler/xla:xla_data_proto", "//tensorflow/core:lib", @@ -145,7 +145,7 @@ tf_cc_binary( name = "show_text_literal", srcs = ["show_text_literal.cc"], deps = [ - "//tensorflow/compiler/xla:literal_util", + "//tensorflow/compiler/xla:literal", "//tensorflow/compiler/xla:statusor", "//tensorflow/compiler/xla:text_literal_reader", "//tensorflow/compiler/xla:types", diff --git a/tensorflow/compiler/xla/tools/replay_computation.cc b/tensorflow/compiler/xla/tools/replay_computation.cc index f7574e0b1cc95daee6d6743ba4e2e490ee87e7c6..854e797ec2e31d32d98f46a75c31ff89caac613b 100644 --- a/tensorflow/compiler/xla/tools/replay_computation.cc +++ b/tensorflow/compiler/xla/tools/replay_computation.cc @@ -43,7 +43,7 @@ limitations under the License. #include "tensorflow/compiler/xla/client/lib/testing.h" #include "tensorflow/compiler/xla/client/local_client.h" #include "tensorflow/compiler/xla/execution_options_util.h" -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/service/gpu/infeed_manager.h" #include "tensorflow/compiler/xla/service/hlo.pb.h" #include "tensorflow/compiler/xla/service/hlo_parser.h" @@ -174,6 +174,11 @@ StatusOr ReplayComputation(const HloSnapshot& module, client->Compile(computation, argument_layouts, ExecutableBuildOptions()) .ValueOrDie(); + // Do not attmept to run the executable, if num_runs is less than 1. + if (opts.num_runs < 1) { + return Cancelled("Cancelled after compilation since --num_runs < 1."); + } + // Run the computation num_runs times, and return the result from the last // execution. StreamExecutorMemoryAllocator allocator( @@ -191,9 +196,6 @@ StatusOr ReplayComputation(const HloSnapshot& module, << static_cast(profile.compute_time_ns()) / 1e9 << "s"; } - // Check that --num_runs > 0, otherwise *result below will fail with an - // unhelpful error (because the loop didn't run any iterations). - CHECK_GT(opts.num_runs, 0) << "--num_runs must be > 0"; TF_ASSIGN_OR_RETURN(std::unique_ptr result_literal, client->ShapedBufferToLiteral(*result)); return std::move(*result_literal); diff --git a/tensorflow/compiler/xla/tools/show_literal.cc b/tensorflow/compiler/xla/tools/show_literal.cc index fe8e72ba32bb4493b2751cfdfeb977f271092f9c..51909190a3ef20c3df78d08796e88bdbb650609d 100644 --- a/tensorflow/compiler/xla/tools/show_literal.cc +++ b/tensorflow/compiler/xla/tools/show_literal.cc @@ -21,7 +21,7 @@ limitations under the License. #include #include -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/types.h" #include "tensorflow/compiler/xla/xla_data.pb.h" #include "tensorflow/core/lib/core/status.h" diff --git a/tensorflow/compiler/xla/tools/show_text_literal.cc b/tensorflow/compiler/xla/tools/show_text_literal.cc index 8525873e913185554d18df8c8c3584bfcdcdcabe..48c837481181f6ad8f864569fd62e0e23fa02ecd 100644 --- a/tensorflow/compiler/xla/tools/show_text_literal.cc +++ b/tensorflow/compiler/xla/tools/show_text_literal.cc @@ -20,7 +20,7 @@ limitations under the License. #include #include -#include "tensorflow/compiler/xla/literal_util.h" +#include "tensorflow/compiler/xla/literal.h" #include "tensorflow/compiler/xla/statusor.h" #include "tensorflow/compiler/xla/text_literal_reader.h" #include "tensorflow/compiler/xla/types.h" diff --git a/tensorflow/compiler/xla/util.h b/tensorflow/compiler/xla/util.h index 6041fae1595dacb309008857f1c758ee96a646bb..5ae099a4622bb7116c7a17f93060b699ead6e3a6 100644 --- a/tensorflow/compiler/xla/util.h +++ b/tensorflow/compiler/xla/util.h @@ -500,17 +500,17 @@ bool c_is_sorted(const C& c, Compare&& comp) { } template -auto c_adjacent_find(const C& c) -> decltype(std::begin(c)) { +auto c_adjacent_find(C& c) -> decltype(std::begin(c)) { return std::adjacent_find(std::begin(c), std::end(c)); } template -auto c_find_if(const C& c, Pred&& pred) -> decltype(std::begin(c)) { +auto c_find_if(C& c, Pred&& pred) -> decltype(std::begin(c)) { return std::find_if(std::begin(c), std::end(c), std::forward(pred)); } template -auto c_find(const C& c, Value&& value) -> decltype(std::begin(c)) { +auto c_find(C& c, Value&& value) -> decltype(std::begin(c)) { return std::find(std::begin(c), std::end(c), std::forward(value)); } @@ -534,6 +534,13 @@ c_count_if(const C& c, Pred&& pred) { return std::count_if(std::begin(c), std::end(c), std::forward(pred)); } +// Determines whether `value` is present in `c`. +template +bool c_linear_search(const C& c, T&& value) { + auto last = std::end(c); + return std::find(std::begin(c), last, std::forward(value)) != last; +} + template int64 FindIndex(const C& c, Value&& value) { auto it = c_find(c, std::forward(value)); @@ -555,6 +562,11 @@ void EraseAt(C* c, int64 index) { c->erase(c->begin() + index); } +template +std::vector ArraySliceToVector(tensorflow::gtl::ArraySlice slice) { + return std::vector(slice.begin(), slice.end()); +} + template std::vector InlinedVectorToVector( const tensorflow::gtl::InlinedVector& inlined_vector) { diff --git a/tensorflow/contrib/BUILD b/tensorflow/contrib/BUILD index fffab5a79549bfa2d74bea227b6e0245834a84c2..60be9db2638d539b4982493d50b533fee9b26bfe 100644 --- a/tensorflow/contrib/BUILD +++ b/tensorflow/contrib/BUILD @@ -9,6 +9,7 @@ load("//third_party/mpi:mpi.bzl", "if_mpi") load("@local_config_cuda//cuda:build_defs.bzl", "if_cuda") load("@local_config_tensorrt//:build_defs.bzl", "if_tensorrt") load("//tensorflow:tensorflow.bzl", "if_not_windows") +load("//tensorflow:tensorflow.bzl", "if_not_windows_cuda") py_library( name = "contrib_py", @@ -26,8 +27,6 @@ py_library( "//tensorflow/contrib/bayesflow:bayesflow_py", "//tensorflow/contrib/boosted_trees:init_py", "//tensorflow/contrib/checkpoint/python:checkpoint", - "//tensorflow/contrib/cloud:cloud_py", - "//tensorflow/contrib/cluster_resolver:cluster_resolver_pip", "//tensorflow/contrib/cluster_resolver:cluster_resolver_py", "//tensorflow/contrib/coder:coder_py", "//tensorflow/contrib/compiler:compiler_py", @@ -45,7 +44,6 @@ py_library( "//tensorflow/contrib/factorization:factorization_py", "//tensorflow/contrib/feature_column:feature_column_py", "//tensorflow/contrib/framework:framework_py", - "//tensorflow/contrib/fused_conv:fused_conv_py", "//tensorflow/contrib/gan", "//tensorflow/contrib/graph_editor:graph_editor_py", "//tensorflow/contrib/grid_rnn:grid_rnn_py", @@ -123,7 +121,17 @@ py_library( "//tensorflow/contrib/kafka", ], "//conditions:default": [], - }) + if_not_windows([ + }) + select({ + "//tensorflow:with_aws_support_windows_override": [], + "//tensorflow:with_aws_support": [ + "//tensorflow/contrib/kinesis", + ], + "//conditions:default": [], + }) + if_not_windows_cuda([ + "//tensorflow/contrib/fused_conv:fused_conv_py", # unresolved symbols, need to export more symbols + ]) + if_not_windows([ + "//tensorflow/contrib/bigtable", # depends on bigtable + "//tensorflow/contrib/cloud:cloud_py", # doesn't compile on Windows "//tensorflow/contrib/ffmpeg:ffmpeg_ops_py", "//tensorflow/contrib/lite/python:lite", # unix dependency, need to fix code ]), @@ -154,6 +162,12 @@ cc_library( "//tensorflow/contrib/kafka:dataset_kernels", ], "//conditions:default": [], + }) + select({ + "//tensorflow:with_aws_support_windows_override": [], + "//tensorflow:with_aws_support": [ + "//tensorflow/contrib/kinesis:dataset_kernels", + ], + "//conditions:default": [], }), ) @@ -183,5 +197,11 @@ cc_library( "//tensorflow/contrib/kafka:dataset_ops_op_lib", ], "//conditions:default": [], + }) + select({ + "//tensorflow:with_aws_support_windows_override": [], + "//tensorflow:with_aws_support": [ + "//tensorflow/contrib/kinesis:dataset_ops_op_lib", + ], + "//conditions:default": [], }), ) diff --git a/tensorflow/contrib/__init__.py b/tensorflow/contrib/__init__.py index 9aad772f0acd941d50d6ba238d345616195a6939..ded05da71877566781a5fb6d0c21e1c8d43de9ed 100644 --- a/tensorflow/contrib/__init__.py +++ b/tensorflow/contrib/__init__.py @@ -25,7 +25,8 @@ import os from tensorflow.contrib import batching from tensorflow.contrib import bayesflow from tensorflow.contrib import checkpoint -from tensorflow.contrib import cloud +if os.name != "nt": + from tensorflow.contrib import cloud from tensorflow.contrib import cluster_resolver from tensorflow.contrib import coder from tensorflow.contrib import compiler diff --git a/tensorflow/contrib/autograph/README.md b/tensorflow/contrib/autograph/README.md index 7e26f4711851138c1834f881621ebfa227a85821..679ab48e5cf46a601f8b550773ca2b3f6c04957d 100644 --- a/tensorflow/contrib/autograph/README.md +++ b/tensorflow/contrib/autograph/README.md @@ -4,7 +4,7 @@ IMPORTANT: AutoGraph is alpha software, and under active development. Expect rou AutoGraph is a Python to TensorFlow compiler. -With AutoGraph, you can write [Eager style](https://www.tensorflow.org/guide/eager) code in a concise manner, and run it as a TensorFlow graph. AutoGraph uses source code transformation and partial evaluation to generate Python code that builds an equivalent TensorFlow subgraph. The result is code that behaves like ops and can be freely combined with other TensorFlow ops. +With AutoGraph, you can write [Eager style](https://www.tensorflow.org/guide/eager) code in a concise manner, and run it as a TensorFlow graph. AutoGraph uses source code transformation and partial evaluation to generate Python code that builds an equivalent TensorFlow subgraph. The result is code that behaves like ops and can be freely combined with other TensorFlow ops. [Please see this file for which parts of the Python language we currently support](LIMITATIONS.md). For example, this Python function: diff --git a/tensorflow/contrib/autograph/__init__.py b/tensorflow/contrib/autograph/__init__.py index 361cf2d77c7e46912d5bff5881df2ffa897c5179..4f8ef2d8a152eb50dfe970e7200dd1f563395262 100644 --- a/tensorflow/contrib/autograph/__init__.py +++ b/tensorflow/contrib/autograph/__init__.py @@ -29,6 +29,10 @@ from tensorflow.contrib.autograph.impl.api import converted_call from tensorflow.contrib.autograph.impl.api import do_not_convert from tensorflow.contrib.autograph.impl.api import RunMode from tensorflow.contrib.autograph.impl.api import to_code +from tensorflow.contrib.autograph.core.errors import improved_errors +from tensorflow.contrib.autograph.core.errors import rewrite_graph_construction_error +from tensorflow.contrib.autograph.core.errors import GraphConstructionError +from tensorflow.contrib.autograph.core.errors import TfRuntimeError from tensorflow.contrib.autograph.impl.api import to_graph from tensorflow.contrib.autograph.lang.directives import set_element_type from tensorflow.contrib.autograph.lang.directives import set_loop_options @@ -42,10 +46,12 @@ _allowed_symbols = [ 'convert', 'converted_call', 'do_not_convert', + 'improved_errors', 'to_code', 'to_graph', # Overloaded operators 'operators', + 'rewrite_graph_construction_error', # Python language "extensions" 'set_element_type', 'set_loop_options', diff --git a/tensorflow/contrib/autograph/converters/__init__.py b/tensorflow/contrib/autograph/converters/__init__.py index e4e8eda42f655e204310eaa9defdd5c90bf06e15..6325ac78dc3a08d14c1abf5e0f1ae60258639162 100644 --- a/tensorflow/contrib/autograph/converters/__init__.py +++ b/tensorflow/contrib/autograph/converters/__init__.py @@ -18,5 +18,15 @@ from __future__ import absolute_import from __future__ import division from __future__ import print_function -# TODO(mdan): Define a base transformer class that can recognize skip_processing -# TODO(mdan): All converters are incomplete, especially those that change blocks +# Naming conventions: +# * each converter should specialize on a single idiom; be consistent with +# the Python reference for naming +# * all converters inherit core.converter.Base +# * module names describe the idiom that the converter covers, plural +# * the converter class is named consistent with the module, singular and +# includes the word Transformer +# +# Example: +# +# lists.py +# class ListTransformer(converter.Base) diff --git a/tensorflow/contrib/autograph/core/BUILD b/tensorflow/contrib/autograph/core/BUILD index 833f9dced81bd651244d281322c830bb1c88b259..1873045a921f8af6068d8fccca6a5625b2aedcf8 100644 --- a/tensorflow/contrib/autograph/core/BUILD +++ b/tensorflow/contrib/autograph/core/BUILD @@ -19,6 +19,7 @@ py_library( srcs = [ "config.py", "converter.py", + "errors.py", "naming.py", ], srcs_version = "PY2AND3", @@ -30,6 +31,31 @@ py_library( ], ) +py_test( + name = "errors_test", + srcs = ["errors_test.py"], + srcs_version = "PY2AND3", + deps = [ + ":core", + "//tensorflow/python:array_ops", + "//tensorflow/python:client_testlib", + "//tensorflow/python:control_flow_ops", + "//tensorflow/python:dtypes", + "//tensorflow/python:math_ops", + "//tensorflow/python:random_ops", + ], +) + +py_test( + name = "naming_test", + srcs = ["naming_test.py"], + srcs_version = "PY2AND3", + deps = [ + ":core", + "//tensorflow/python:client_testlib", + ], +) + py_library( name = "test_lib", srcs = [ @@ -47,13 +73,3 @@ py_library( "@six_archive//:six", ], ) - -py_test( - name = "naming_test", - srcs = ["naming_test.py"], - srcs_version = "PY2AND3", - deps = [ - ":core", - "//tensorflow/python:client_testlib", - ], -) diff --git a/tensorflow/contrib/autograph/core/annos.py b/tensorflow/contrib/autograph/core/annos.py new file mode 100644 index 0000000000000000000000000000000000000000..b8937ce36a9631739ab3d7e65a4dad4124406a00 --- /dev/null +++ b/tensorflow/contrib/autograph/core/annos.py @@ -0,0 +1,39 @@ +# Copyright 2017 The TensorFlow Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================== +"""Annotations specific to AutoGraph.""" + +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + +from enum import Enum + + +class NoValue(Enum): + + def __repr__(self): + return self.name + + +class NodeAnno(NoValue): + """Additional annotations used by AutoGraph converters. + + These are in addition to the basic annotations declared in pyct/anno.py and + pyct/static_analysis/annos.py. + """ + + # The directives collection - see directives.py + DIRECTIVES = ( + 'Dict depicting static directive calls. See the directives converter.') diff --git a/tensorflow/contrib/autograph/core/errors.py b/tensorflow/contrib/autograph/core/errors.py new file mode 100644 index 0000000000000000000000000000000000000000..e58745337a3faac5e9f351174465443fa52fd6bc --- /dev/null +++ b/tensorflow/contrib/autograph/core/errors.py @@ -0,0 +1,272 @@ +# Copyright 2018 The TensorFlow Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================== +"""Error rewriting logic. + +Contains the functions responsible for rewriting tracebacks of errors raised +in AutoGraph (AG) code to refer to user written code, so that errors only refer +to the original user code. + +When 'user code' is used in comments it refers to the original source code that +the user wrote and is converting using AutoGraph. +""" + +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + +import contextlib +import logging +import sys +import traceback + +from tensorflow.contrib.autograph.pyct.origin_info import CodeLocation +from tensorflow.python.framework import errors_impl +from tensorflow.python.util import tf_inspect + + +class GraphConstructionError(Exception): + """Error for graph construction errors from AutoGraph generated code.""" + + def __init__(self, original_error, custom_traceback): + self.original_error = original_error + self.custom_traceback = custom_traceback + super(GraphConstructionError, self).__init__() + + def __str__(self): + traceback_str = ''.join(traceback.format_list(self.custom_traceback)) + return ('Traceback (most recent call last):\n' + traceback_str + '\n' + str( + self.original_error) + '\n') + + +class TfRuntimeError(Exception): + """Error wrapper for runtime errors raised by AutoGraph generated code.""" + + def __init__(self, op_name, op_message, custom_traceback): + self.op_name = op_name + self.op_message = op_message + self.custom_traceback = custom_traceback + super(TfRuntimeError, self).__init__() + + def __str__(self): + message = '%s\n\nCaused by op %r, defined at:\n' % (self.op_message, + self.op_name) + return message + ''.join(traceback.format_list(self.custom_traceback)) + + +def _rewrite_frame(source_map, cleaned_traceback, stack_frame_indices): + """Rewrites the stack frames at the given indices using the given source map. + + Args: + source_map: Dict[CodeLocation, OriginInfo], a mapping between the user and + AG generated code. + cleaned_traceback: List[Tuple[text, text, text, text]], the current + traceback. + stack_frame_indices: Iterable[Int], frame indices to possibly rewrite if + there are matching source mapping keys. + + Returns: + None + """ + for frame_index in stack_frame_indices: + # (file_path, line number, function name, code) + file_path, line_number, _, _ = cleaned_traceback[frame_index] + source_map_key = CodeLocation(file_path=file_path, line_number=line_number) + found_mapping = source_map_key in source_map + if found_mapping: + cleaned_traceback[frame_index] = source_map[source_map_key].as_frame() + + +# TODO(znado): Make more robust to name changes in the rewriting logic. +def _remove_rewrite_frames(tb): + """Remove stack frames containing the error rewriting logic.""" + cleaned_tb = [] + for f in tb: + if 'ag__.rewrite_graph_construction_error' not in f[3]: + cleaned_tb.append(f) + return cleaned_tb + + +def rewrite_graph_construction_error(source_map): + """Rewrites errors raised by non-AG APIs inside AG generated code. + + Meant to be called from the try/except block inside each AutoGraph generated + function. Only rewrites the traceback frames corresponding to the function + that this is called from. When we raise a GraphConstructionError at the end + it is then caught by calling functions, where they can be responsible for + rewriting their own frames. + + Args: + source_map: Dict[CodeLocation, OriginInfo], a mapping between the user and + AG generated code. + + Raises: + GraphConstructionError: The rewritten underlying error. + Exception: The underlying error, if it could not be rewritten. + """ + error_info = sys.exc_info() + _, original_error, e_traceback = error_info + assert original_error is not None + try: + _, _, _, func_name, _, _ = tf_inspect.stack()[1] + # The latest function call is added to the beginning of a traceback, but + # when rewriting the traceback of multiple function calls the previous + # functions' except blocks may have already rewritten their own frames so + # we want to copy over all of the previous frames. We may have rewritten + # previous frames only if the error is a GraphConstructionError. + if isinstance(original_error, GraphConstructionError): + cleaned_traceback = traceback.extract_tb(e_traceback) + previous_traceback = original_error.custom_traceback + cleaned_traceback = [cleaned_traceback[0]] + previous_traceback + else: + cleaned_traceback = traceback.extract_tb(e_traceback) + cleaned_traceback = _remove_rewrite_frames(cleaned_traceback) + + current_frame_indices = [] + # This code is meant to be called from the try/except block that wraps a + # function body. Here we look for all frames that came from the function + # that this wraps, look for any matching line numbers in the source + # mapping, and then rewrite them if matches are found. + for fi, frame in enumerate(cleaned_traceback): + _, _, frame_func_name, _ = frame + if frame_func_name == func_name: + current_frame_indices.append(fi) + break + if current_frame_indices: + _rewrite_frame(source_map, cleaned_traceback, current_frame_indices) + + if isinstance(original_error, GraphConstructionError): + original_error.custom_traceback = cleaned_traceback + new_error = original_error + else: + new_error = GraphConstructionError(original_error, cleaned_traceback) + except Exception: + logging.exception('Error while rewriting AutoGraph error:') + raise original_error + else: + raise new_error + finally: + # Addresses warning https://docs.python.org/2/library/sys.html#sys.exc_info. + del e_traceback + + +def rewrite_tf_runtime_error(error, source_map): + """Rewrites TensorFlow runtime errors raised by ops created in AG code. + + Args: + error: error_impl.OpError, an TensorFlow error that will have its traceback + rewritten. + source_map: Dict[CodeLocation, OriginInfo], a mapping between the user and + AG generated code. + + Returns: + A TfRuntimeError with a traceback rewritten according to the given + source mapping. + """ + # Check for cases where we leave a user method and re-enter it in the + # traceback. This is done by looking at the function names when the + # filenames are from any files the user code is in. If we find a case where + # we return to a user method after leaving it then we cut out the frames in + # between because we assume this means these in between frames are from + # internal AutoGraph code that shouldn't be included. + # + # An example of this is: + # + # File "file1.py", line 57, in my_func + # ... + # File "control_flow_ops.py", line 231, in cond + # ... + # File "control_flow_ops.py", line 1039, in inner_cond + # ... + # File "file1.py", line 68, in my_func + # ... + # + # Where we would remove the control_flow_ops.py frames because we re-enter + # my_func in file1.py. + # + # The source map keys are (file_path, line_number) so get the set of all user + # file_paths. + try: + all_user_files = set(k.file_path for k in source_map) + cleaned_traceback = [] + last_user_frame_index = None + last_user_user_file_path = None + last_user_user_fn_name = None + for fi, frame in enumerate(error.op.traceback): + frame_file_path, frame_line_number, _, _ = frame + src_map_key = CodeLocation( + file_path=frame_file_path, line_number=frame_line_number) + if frame_file_path in all_user_files: + if src_map_key in source_map: + original_fn_name = source_map[src_map_key].function_name + if (last_user_frame_index is not None and + last_user_user_file_path == frame_file_path): + if last_user_user_fn_name == original_fn_name: + cleaned_traceback = cleaned_traceback[:last_user_frame_index] + else: + cleaned_traceback = cleaned_traceback[:last_user_frame_index + 1] + last_user_user_fn_name = original_fn_name + else: + last_user_user_fn_name = None + last_user_frame_index = fi + last_user_user_file_path = frame_file_path + cleaned_traceback.append(frame) + + for fi in range(len(cleaned_traceback)): + _rewrite_frame(source_map, cleaned_traceback, [fi]) + op_name = error.op.name + op_message = error.message + rewritten_error = TfRuntimeError(op_name, op_message, cleaned_traceback) + return rewritten_error + except Exception: # pylint: disable=broad-except + logging.exception('Error while rewriting AutoGraph error:') + return error + + +# TODO(znado): Add arg to enable different levels of error rewriting. +@contextlib.contextmanager +def improved_errors(converted_function): + """Context manager that rewrites runtime errors. + + This context manager will rewrite runtime errors so that their traceback + is relative to the original code before conversion. + + Use with the output of to_graph, and wrap the execution of respective ops. + Example: + + converted_my_func = ag.to_graph(my_func) + ops = converted_my_func(...) + + with ag.improved_errors(converted_my_func): + sess.run(ops) + + Args: + converted_function: Callable[..., Any], the output of a to_graph call + + Yields: + None + + Raises: + TfRuntimeError: if any OpError originates in the converted code, it will + be wrapped into a TfRuntimeError + ValueError: If converted_function is not generated by AutoGraph + """ + if (getattr(converted_function, 'ag_source_map', None) is None or + not converted_function.ag_source_map): + raise ValueError( + 'converted_function must be the result of an autograph.to_graph call') + try: + yield + except errors_impl.OpError as e: + raise rewrite_tf_runtime_error(e, converted_function.ag_source_map) diff --git a/tensorflow/contrib/autograph/core/errors_test.py b/tensorflow/contrib/autograph/core/errors_test.py new file mode 100644 index 0000000000000000000000000000000000000000..7be54563a1a86a56437f4da2941bf5187ce813a9 --- /dev/null +++ b/tensorflow/contrib/autograph/core/errors_test.py @@ -0,0 +1,116 @@ +# Copyright 2018 The TensorFlow Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================== +"""Tests for errors module.""" + +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + +from tensorflow.contrib.autograph.core import errors +from tensorflow.contrib.autograph.pyct import origin_info +from tensorflow.python.framework import dtypes +from tensorflow.python.framework import errors as tf_errors +from tensorflow.python.ops import array_ops +from tensorflow.python.platform import test +from tensorflow.python.util import tf_inspect + + +def zero_div(): + return array_ops.constant(10, dtype=dtypes.int32) // 0 + + +def zero_div_caller(): + a = zero_div() + 2 + return a + + +class RuntimeErrorsTest(test.TestCase): + + def setUp(self): + self._fake_origin = origin_info.OriginInfo('new file', 'new func', 96, 0, + 'print("hello world!")') + + def test_error_replacement(self): + _, zero_div_lineno = tf_inspect.getsourcelines(zero_div) + src_map = { + errors.CodeLocation( + file_path=__file__, line_number=zero_div_lineno + 1): + self._fake_origin + } + with self.assertRaises(errors.TfRuntimeError) as cm: + z = zero_div_caller() + zero_div_caller.ag_source_map = src_map + with errors.improved_errors(zero_div_caller): + with self.test_session() as sess: + sess.run(z) + expected = cm.exception + current_traceback = expected.custom_traceback + for frame in current_traceback: + self.assertNotEqual('zero_div', frame[2]) + self.assertTrue( + any(self._fake_origin.as_frame() == frame + for frame in current_traceback)) + + def test_error_not_found(self): + src_map = { + errors.CodeLocation(file_path=__file__, line_number=-1): + self._fake_origin + } + with self.assertRaises(errors.TfRuntimeError) as cm: + z = zero_div_caller() + zero_div_caller.ag_source_map = src_map + with errors.improved_errors(zero_div_caller): + with self.test_session() as sess: + sess.run(z) + expected = cm.exception + current_traceback = expected.custom_traceback + self.assertTrue(any('zero_div' in frame[2] for frame in current_traceback)) + for frame in current_traceback: + self.assertNotEqual(frame, self._fake_origin.as_frame()) + + def test_rewriting_error(self): + _, zero_div_lineno = tf_inspect.getsourcelines(zero_div) + src_map = { + errors.CodeLocation( + file_path=__file__, line_number=zero_div_lineno + 1): + None + } + with self.assertRaisesRegexp(tf_errors.InvalidArgumentError, + 'Integer division by zero'): + z = zero_div_caller() + zero_div_caller.ag_source_map = src_map + with errors.improved_errors(zero_div_caller): + with self.test_session() as sess: + sess.run(z) + + def test_no_ag_source_map(self): + with self.assertRaisesRegexp( + ValueError, + 'converted_function must be the result of an autograph.to_graph call'): + with errors.improved_errors(None): + pass + + def test_bad_ag_source_map(self): + with self.assertRaisesRegexp( + ValueError, + 'converted_function must be the result of an autograph.to_graph call'): + src_map = None + zero_div_caller.ag_source_map = src_map + with errors.improved_errors(None): + pass + + +if __name__ == '__main__': + test.main() diff --git a/tensorflow/contrib/autograph/examples/integration_tests/BUILD b/tensorflow/contrib/autograph/examples/integration_tests/BUILD new file mode 100644 index 0000000000000000000000000000000000000000..1368ce244c26a95d1ee7cc4708891badc7981db7 --- /dev/null +++ b/tensorflow/contrib/autograph/examples/integration_tests/BUILD @@ -0,0 +1,29 @@ +licenses(["notice"]) # Apache 2.0 + +exports_files(["LICENSE"]) + +load("//tensorflow:tensorflow.bzl", "py_test") + +filegroup( + name = "all_files", + srcs = glob( + ["**/*"], + exclude = [ + "**/METADATA", + "**/OWNERS", + ], + ), + visibility = ["//tensorflow:__subpackages__"], +) + +py_test( + name = "keras_test", + srcs = [ + "keras_test.py", + ], + srcs_version = "PY2AND3", + visibility = ["//visibility:public"], + deps = [ + "//tensorflow:tensorflow_py", + ], +) diff --git a/tensorflow/contrib/autograph/examples/integration_tests/keras_test.py b/tensorflow/contrib/autograph/examples/integration_tests/keras_test.py new file mode 100644 index 0000000000000000000000000000000000000000..a2fc7c550efe829a104dc3931f29cc4f8fcf60d4 --- /dev/null +++ b/tensorflow/contrib/autograph/examples/integration_tests/keras_test.py @@ -0,0 +1,37 @@ +# Copyright 2017 The TensorFlow Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================== +"""Keras integration tests.""" + +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + +import tensorflow as tf + + +class MinimalKeras(tf.keras.Model): + + def call(self, x): + return x * 3 + + +class KerasTest(tf.test.TestCase): + + def test_basic(self): + MinimalKeras() + + +if __name__ == '__main__': + tf.test.main() diff --git a/tensorflow/contrib/autograph/examples/notebooks/autograph_vs_eager_mnist_benchmark.ipynb b/tensorflow/contrib/autograph/examples/notebooks/autograph_vs_eager_mnist_benchmark.ipynb new file mode 100644 index 0000000000000000000000000000000000000000..d6a29ea1ec72d1720f8f4c46347fe7fba3ac25b5 --- /dev/null +++ b/tensorflow/contrib/autograph/examples/notebooks/autograph_vs_eager_mnist_benchmark.ipynb @@ -0,0 +1,668 @@ +{ + "cells": [ + { + "cell_type": "code", + "execution_count": 0, + "metadata": { + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + }, + "colab_type": "code", + "id": "Pa2qpEmoVOGe" + }, + "outputs": [], + "source": [ + "from __future__ import absolute_import\n", + "from __future__ import division\n", + "from __future__ import print_function\n", + "\n", + "import os\n", + "import time\n", + "\n", + "import tensorflow as tf\n", + "\n", + "import matplotlib.pyplot as plt\n", + "import numpy as np\n", + "import six\n", + "\n", + "from tensorflow.contrib import autograph\n", + "from tensorflow.contrib.eager.python import tfe\n", + "from tensorflow.python.eager import context\n" + ] + }, + { + "cell_type": "code", + "execution_count": 0, + "metadata": { + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + }, + "colab_type": "code", + "id": "YfnHJbBOBKae" + }, + "outputs": [], + "source": [ + "import gzip\n", + "import shutil\n", + "\n", + "from six.moves import urllib\n", + "\n", + "\n", + "def download(directory, filename):\n", + " filepath = os.path.join(directory, filename)\n", + " if tf.gfile.Exists(filepath):\n", + " return filepath\n", + " if not tf.gfile.Exists(directory):\n", + " tf.gfile.MakeDirs(directory)\n", + " url = 'https://storage.googleapis.com/cvdf-datasets/mnist/' + filename + '.gz'\n", + " zipped_filepath = filepath + '.gz'\n", + " print('Downloading %s to %s' % (url, zipped_filepath))\n", + " urllib.request.urlretrieve(url, zipped_filepath)\n", + " with gzip.open(zipped_filepath, 'rb') as f_in, open(filepath, 'wb') as f_out:\n", + " shutil.copyfileobj(f_in, f_out)\n", + " os.remove(zipped_filepath)\n", + " return filepath\n", + "\n", + "\n", + "def dataset(directory, images_file, labels_file):\n", + " images_file = download(directory, images_file)\n", + " labels_file = download(directory, labels_file)\n", + "\n", + " def decode_image(image):\n", + " # Normalize from [0, 255] to [0.0, 1.0]\n", + " image = tf.decode_raw(image, tf.uint8)\n", + " image = tf.cast(image, tf.float32)\n", + " image = tf.reshape(image, [784])\n", + " return image / 255.0\n", + "\n", + " def decode_label(label):\n", + " label = tf.decode_raw(label, tf.uint8)\n", + " label = tf.reshape(label, [])\n", + " return tf.to_int32(label)\n", + "\n", + " images = tf.data.FixedLengthRecordDataset(\n", + " images_file, 28 * 28, header_bytes=16).map(decode_image)\n", + " labels = tf.data.FixedLengthRecordDataset(\n", + " labels_file, 1, header_bytes=8).map(decode_label)\n", + " return tf.data.Dataset.zip((images, labels))\n", + "\n", + "\n", + "def mnist_train(directory):\n", + " return dataset(directory, 'train-images-idx3-ubyte',\n", + " 'train-labels-idx1-ubyte')\n", + "\n", + "def mnist_test(directory):\n", + " return dataset(directory, 't10k-images-idx3-ubyte', 't10k-labels-idx1-ubyte')\n", + "\n", + "def setup_mnist_data(is_training, hp, batch_size):\n", + " if is_training:\n", + " ds = mnist_train('/tmp/autograph_mnist_data')\n", + " ds = ds.cache()\n", + " ds = ds.shuffle(batch_size * 10)\n", + " else:\n", + " ds = mnist_test('/tmp/autograph_mnist_data')\n", + " ds = ds.cache()\n", + " ds = ds.repeat()\n", + " ds = ds.batch(batch_size)\n", + " return ds\n" + ] + }, + { + "cell_type": "code", + "execution_count": 0, + "metadata": { + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + }, + "colab_type": "code", + "id": "x_MU13boiok2" + }, + "outputs": [], + "source": [ + "def mlp_model(input_shape):\n", + " model = tf.keras.Sequential((\n", + " tf.keras.layers.Dense(100, activation='relu', input_shape=input_shape),\n", + " tf.keras.layers.Dense(100, activation='relu'),\n", + " tf.keras.layers.Dense(10, activation='softmax')))\n", + " model.build()\n", + " return model\n" + ] + }, + { + "cell_type": "code", + "execution_count": 0, + "metadata": { + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + }, + "colab_type": "code", + "id": "kfZk9EFZ5TeQ" + }, + "outputs": [], + "source": [ + "# Test-only parameters. Test checks successful completion not correctness. \n", + "burn_ins = 1\n", + "trials = 1\n", + "max_steps = 2" + ] + }, + { + "cell_type": "code", + "execution_count": 0, + "metadata": { + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + }, + "colab_type": "code", + "id": "gWXV8WHn43iZ" + }, + "outputs": [], + "source": [ + "#@test {\"skip\": true} \n", + "burn_ins = 3\n", + "trials = 10\n", + "max_steps = 500" + ] + }, + { + "cell_type": "markdown", + "metadata": { + "colab_type": "text", + "id": "DXt4GoTxtvn2" + }, + "source": [ + "# Autograph" + ] + }, + { + "cell_type": "code", + "execution_count": 0, + "metadata": { + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + }, + "colab_type": "code", + "id": "W51sfbONiz_5" + }, + "outputs": [], + "source": [ + "def predict(m, x, y):\n", + " y_p = m(x)\n", + " losses = tf.keras.losses.categorical_crossentropy(y, y_p)\n", + " l = tf.reduce_mean(losses)\n", + " accuracies = tf.keras.metrics.categorical_accuracy(y, y_p)\n", + " accuracy = tf.reduce_mean(accuracies)\n", + " return l, accuracy\n" + ] + }, + { + "cell_type": "code", + "execution_count": 0, + "metadata": { + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + }, + "colab_type": "code", + "id": "CsAD0ajbi9iZ" + }, + "outputs": [], + "source": [ + "def fit(m, x, y, opt):\n", + " l, accuracy = predict(m, x, y)\n", + " opt.minimize(l)\n", + " return l, accuracy\n" + ] + }, + { + "cell_type": "code", + "execution_count": 0, + "metadata": { + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + }, + "colab_type": "code", + "id": "RVw57HdTjPzi" + }, + "outputs": [], + "source": [ + "def get_next_batch(ds):\n", + " itr = ds.make_one_shot_iterator()\n", + " image, label = itr.get_next()\n", + " x = tf.to_float(tf.reshape(image, (-1, 28 * 28)))\n", + " y = tf.one_hot(tf.squeeze(label), 10)\n", + " return x, y\n" + ] + }, + { + "cell_type": "code", + "execution_count": 0, + "metadata": { + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + }, + "colab_type": "code", + "id": "UUI0566FjZPx" + }, + "outputs": [], + "source": [ + "def train(train_ds, test_ds, hp):\n", + " m = mlp_model((28 * 28,))\n", + " opt = tf.train.MomentumOptimizer(hp.learning_rate, 0.9)\n", + " train_losses = []\n", + " train_losses = autograph.utils.set_element_type(train_losses, tf.float32)\n", + " test_losses = []\n", + " test_losses = autograph.utils.set_element_type(test_losses, tf.float32)\n", + " train_accuracies = []\n", + " train_accuracies = autograph.utils.set_element_type(train_accuracies,\n", + " tf.float32)\n", + " test_accuracies = []\n", + " test_accuracies = autograph.utils.set_element_type(test_accuracies,\n", + " tf.float32)\n", + " i = tf.constant(0)\n", + " while i \u003c hp.max_steps:\n", + " train_x, train_y = get_next_batch(train_ds)\n", + " test_x, test_y = get_next_batch(test_ds)\n", + " step_train_loss, step_train_accuracy = fit(m, train_x, train_y, opt)\n", + " step_test_loss, step_test_accuracy = predict(m, test_x, test_y)\n", + "\n", + " train_losses.append(step_train_loss)\n", + " test_losses.append(step_test_loss)\n", + " train_accuracies.append(step_train_accuracy)\n", + " test_accuracies.append(step_test_accuracy)\n", + " i += 1\n", + " return (autograph.stack(train_losses), autograph.stack(test_losses), autograph.stack(train_accuracies),\n", + " autograph.stack(test_accuracies))\n" + ] + }, + { + "cell_type": "code", + "execution_count": 0, + "metadata": { + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + }, + "height": 789 + }, + "colab_type": "code", + "executionInfo": { + "elapsed": 11529, + "status": "ok", + "timestamp": 1531163743912, + "user": { + "displayName": "", + "photoUrl": "", + "userId": "" + }, + "user_tz": 240 + }, + "id": "K1m8TwOKjdNd", + "outputId": "59db8f19-23a5-413a-e9d0-fb756b0e4757" + }, + "outputs": [ + { + "name": "stdout", + "output_type": "stream", + "text": [ + "Duration: 0.592790126801\n", + "Duration: 0.594069957733\n", + "Duration: 0.591835975647\n", + "Duration: 0.592386007309\n", + "Duration: 0.595040082932\n", + "Duration: 0.594245910645\n", + "Duration: 0.624264001846\n", + "Duration: 0.6021900177\n", + "Duration: 0.592960119247\n", + "Duration: 0.599496841431\n", + "Mean duration: 0.597927904129 +/- 0.0093268291102\n" + ] + }, + { + "data": { + "image/png": "iVBORw0KGgoAAAANSUhEUgAAAYkAAAEcCAYAAAAydkhNAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nDAkiInKKIUFERE4xJIiIyCmGBBEROfV/smX5vm0Z6kkAAAAASUVORK5CYII=\n", + "text/plain": [ + "\u003cmatplotlib.figure.Figure at 0x7f970d490590\u003e" + ] + }, + "metadata": { + "tags": [] + }, + "output_type": "display_data" + }, + { + "name": "stdout", + "output_type": "stream", + "text": [ + "test_accuracy 0.1\n" + ] + }, + { + "data": { + "image/png": "iVBORw0KGgoAAAANSUhEUgAAAYwAAAEcCAYAAADUX4MJAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nkUgkEquQCkMikUgkVvH/AcQ/YGad+SX7AAAAAElFTkSuQmCC\n", + "text/plain": [ + "\u003cmatplotlib.figure.Figure at 0x7f971b401110\u003e" + ] + }, + "metadata": { + "tags": [] + }, + "output_type": "display_data" + } + ], + "source": [ + "#@test {\"timeout\": 90} \n", + "with tf.Graph().as_default():\n", + " hp = tf.contrib.training.HParams(\n", + " learning_rate=0.05,\n", + " max_steps=max_steps,\n", + " )\n", + " train_ds = setup_mnist_data(True, hp, 500)\n", + " test_ds = setup_mnist_data(False, hp, 100)\n", + " tf_train = autograph.to_graph(train)\n", + " (train_losses_, test_losses_, train_accuracies_,\n", + " test_accuracies_) = tf_train(train_ds, test_ds, hp)\n", + "\n", + " with tf.Session() as sess:\n", + " durations = []\n", + " for t in range(burn_ins + trials):\n", + " sess.run(tf.global_variables_initializer())\n", + " start = time.time()\n", + " (train_losses, test_losses, train_accuracies,\n", + " test_accuracies) = sess.run([train_losses_, \n", + " test_losses_, \n", + " train_accuracies_,\n", + " test_accuracies_])\n", + " if t \u003c burn_ins:\n", + " continue\n", + " duration = time.time() - start\n", + " durations.append(duration)\n", + " print('Duration:', duration)\n", + "\n", + " print('Mean duration:', np.mean(durations), '+/-', np.std(durations))\n", + " plt.title('MNIST train/test losses')\n", + " plt.plot(train_losses, label='train loss')\n", + " plt.plot(test_losses, label='test loss')\n", + " plt.legend()\n", + " plt.xlabel('Training step')\n", + " plt.ylabel('Loss')\n", + " plt.show()\n", + " plt.title('MNIST train/test accuracies')\n", + " plt.plot(train_accuracies, label='train accuracy')\n", + " plt.plot(test_accuracies, label='test accuracy')\n", + " print('test_accuracy', test_accuracies[-1])\n", + " plt.legend(loc='lower right')\n", + " plt.xlabel('Training step')\n", + " plt.ylabel('Accuracy')\n", + " plt.show()\n" + ] + }, + { + "cell_type": "markdown", + "metadata": { + "colab_type": "text", + "id": "A06kdgtZtlce" + }, + "source": [ + "# Eager" + ] + }, + { + "cell_type": "code", + "execution_count": 0, + "metadata": { + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + }, + "colab_type": "code", + "id": "hBKOKGrWty4e" + }, + "outputs": [], + "source": [ + "def predict(m, x, y):\n", + " y_p = m(x)\n", + " losses = tf.keras.losses.categorical_crossentropy(tf.cast(y, tf.float32), y_p)\n", + " l = tf.reduce_mean(losses)\n", + " accuracies = tf.keras.metrics.categorical_accuracy(y, y_p)\n", + " accuracy = tf.reduce_mean(accuracies)\n", + " return l, accuracy\n" + ] + }, + { + "cell_type": "code", + "execution_count": 0, + "metadata": { + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + }, + "colab_type": "code", + "id": "HCgTZ0MTt6vt" + }, + "outputs": [], + "source": [ + "def train(ds, hp):\n", + " m = mlp_model((28 * 28,))\n", + " opt = tf.train.MomentumOptimizer(hp.learning_rate, 0.9)\n", + " train_losses = []\n", + " test_losses = []\n", + " train_accuracies = []\n", + " test_accuracies = []\n", + " i = 0\n", + " train_test_itr = tfe.Iterator(ds)\n", + " for (train_x, train_y), (test_x, test_y) in train_test_itr:\n", + " train_x = tf.to_float(tf.reshape(train_x, (-1, 28 * 28)))\n", + " train_y = tf.one_hot(tf.squeeze(train_y), 10)\n", + " test_x = tf.to_float(tf.reshape(test_x, (-1, 28 * 28)))\n", + " test_y = tf.one_hot(tf.squeeze(test_y), 10)\n", + " if i \u003e hp.max_steps:\n", + " break\n", + " with tf.GradientTape() as tape:\n", + " step_train_loss, step_train_accuracy = predict(m, train_x, train_y)\n", + " grad = tape.gradient(step_train_loss, m.variables)\n", + " opt.apply_gradients(zip(grad, m.variables))\n", + " step_test_loss, step_test_accuracy = predict(m, test_x, test_y)\n", + "\n", + " train_losses.append(step_train_loss)\n", + " test_losses.append(step_test_loss)\n", + " train_accuracies.append(step_train_accuracy)\n", + " test_accuracies.append(step_test_accuracy)\n", + " i += 1\n", + " return train_losses, test_losses, train_accuracies, test_accuracies\n" + ] + }, + { + "cell_type": "code", + "execution_count": 0, + "metadata": { + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + }, + "height": 789 + }, + "colab_type": "code", + "executionInfo": { + "elapsed": 56025, + "status": "ok", + "timestamp": 1531163800231, + "user": { + "displayName": "", + "photoUrl": "", + "userId": "" + }, + "user_tz": 240 + }, + "id": "plv_yrn_t8Dy", + "outputId": "68be955d-61dd-43e4-b540-3794e3c8f990" + }, + "outputs": [ + { + "name": "stdout", + "output_type": "stream", + "text": [ + "Duration: 4.2232978344\n", + "Duration: 4.2386469841\n", + "Duration: 4.24286484718\n", + "Duration: 4.24036884308\n", + "Duration: 4.25758385658\n", + "Duration: 4.23242998123\n", + "Duration: 4.4213449955\n", + "Duration: 4.29613113403\n", + "Duration: 4.28209114075\n", + "Duration: 4.24192905426\n", + "Mean duration: 4.26766886711 +/- 0.055508619589\n" + ] + }, + { + "data": { + "image/png": "iVBORw0KGgoAAAANSUhEUgAAAYwAAAEcCAYAAADUX4MJAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz\nAAALEgAACxIB0t1+/AAAIABJREFUeJzsnXdgFGX6x78zW7KbTSE9JIA0pQkIooCgqBx2qiK/O0XU\n8zyFAw/w8MSCFcuJCHqKoFiwIHIgIgooaGjSCU1aaCEJ6W1btszM74/ZmZ2tWchuQjbP55/s7szO\nvDPZeb/vU97nZQRBEEAQBEEQ9cA2dQMIgiCI5gEJBkEQBBESJBgEQRBESJBgEARBECFBgkEQBEGE\nBAkGQRAEERIkGARBEERIkGAQRIjs3r0bt99+e1M3o14KCwvRtWtX8Dzf1E0hogwSDKLB3HzzzejZ\nsyeqq6s9Ph85ciS6du2KoqIiAMC///1vdO3aFQcPHpT3yc/PR9euXeX348ePx/Lly+X3CxYswNCh\nQ9G3b1/ceOONmDZtGgDgrrvuQt++fdG3b190794dvXr1Qp8+fdC3b18sXLjQp43vvfceZsyY0aDr\n7NevH3766acL+s6HH36IuXPnYufOnRgyZEiDzi/hfY/8wTBMWM5FEErUTd0AIjpo06YN1qxZg/vu\nuw8AcPz4cdhsNo+Oi2EYtGrVCu+88w4+/vhjj8/9sXLlSqxevRqfffYZ2rRpg4qKCmzcuBEA8MMP\nP8j7jR8/HqNGjcLdd9/doGsQBCHsHW1OTg6efPJJOBwO6sSJZg9ZGERYGDlyJFauXCm/X7lyJUaP\nHu2z3+jRo3Hs2DHs3r273mMeOnQIgwcPRps2bQAAKSkpGDt2rN99g1W42bx5MxYsWIAff/wRffr0\nwahRowCIQjN37lz8+c9/xlVXXYWCggKsWLECd9xxB/r27Ythw4bhm2++kY/jbSXcfPPNWLx4MUaM\nGIFrrrkG06ZNg91ul7fX1tbi7Nmz6N69Ox599FGUlpbKVlBZWRkEQcDChQsxbNgwDBgwAFOnTkVt\nbS0AwG6341//+hf69++Pa665BmPHjkVlZSXmzp2LPXv24OWXX0bfvn3xyiuv1HsfS0tL8fjjj6N/\n//649dZb8e2338rbDhw4gLvvvhtXX301Bg8ejDfeeCPo+QHAZDLhmWeeweDBgzFkyBC888478v3P\nz8/H+PHj0a9fPwwcOFC2CInogCwMIiz07t0bq1atwqlTp9C+fXusXbsWX331FebOneuxn06nw2OP\nPYa3334bX331Vb3HfPXVV5Geno7+/fuje/fuYNkLH+Ncf/31eOyxx5Cfn48333zTY9vq1auxaNEi\ndOjQATzPIyUlBQsXLkSbNm2we/duPPLII+jVqxe6desGwNcaWrt2LRYvXgytVov/+7//w8qVKzFu\n3DgAwJYtWzBgwADodDosWrQIM2bMwG+//SZ/99NPP8XGjRvx5ZdfIikpCa+88gpefPFFzJkzBytX\nroTJZMLmzZuh0Whw5MgRxMTEYOrUqdi7dy9GjhyJe+65J6TrnzZtGrp06YL58+fj5MmTeOihh9C2\nbVsMGDAAs2fPxoQJEzBixAhYrVacOHECAAKeHwBmzJiB9PR0bNiwAWazGY899hiysrJw7733Yt68\neRg8eDCWLFkCu92OQ4cOXfD/i7h0IQuDCBsjR47Ed999h61bt6Jjx45IT0/3u9+9996L8+fPY/Pm\nzUGPN2LECDz33HPYunUrxo8fj+uuu85vfKIhjB49Gp06dQLLslCr1RgyZIhs0fTr1w+DBg0Kag09\n8MADSE1NRUJCAm666SYcOXJE3vbbb78FjVssW7YM//znP5Geng6NRoNJkyZh3bp14HkearUa1dXV\nOH36NBiGQffu3WEwGC74+s6fP499+/bhySefhEajQdeuXTF27FisWrUKAKBWq5Gfn4+qqiro9Xr0\n6tVL/tzf+SsqKrB582bMnDkTMTExSE5OxoQJE7BmzRr5e4WFhSgpKYFWq0Xfvn0vuM3EpQtZGETY\nGDFiBO6//34UFBRg5MiRAffTarWYOHEi5s2bhzlz5gQ95l133YW77roLHMfhl19+wfTp09GjRw8M\nGjQoLG3OzMz0eJ+Tk4P3338fZ86cAc/zqKurQ5cuXQJ+PyUlRX6t1+tRVlYGQHSRbdu2DU8//XTA\n7xYVFeEf//iHbDUJggC1Wo3y8nKMHDkSxcXFmDZtGoxGI4YPH45p06ZBpVJd0PWVlZUhMTERer1e\n/iwrKwuHDx8GAMyePRvz5s3D7bffjrZt22LSpEm48cYbfc4/YsQITJ06FYWFhXA6nRg8eLDcZkEQ\n0Lp1awCi9fHOO+/gnnvuQatWrfDggw82OLZEXDqQYBBhIysrC9nZ2di0aRNmz54ddN8xY8bgo48+\nws8//xzSsVUqFW699VYsXLgQJ06cCJtgKF1MdrsdTzzxBP7zn/9g6NChYFkWkyZNChofCcTBgwfR\npk0bJCUl+ZxHonXr1pg9ezb69Onj9xiTJk3CpEmTUFRUhL/97W/o2LEj7r777gsKnqenp6OmpgYW\niwWxsbEARKtDsv7atWsni/a6deswZcoU7Ny5Ezqdzuf8HTp0wA033ICYmBjs2LHDbztSUlLw8ssv\nAwD27NmDhx56CNdeey3atm0bcpuJSxdySRFhZfbs2fjss8+g0+mC7qdSqfCPf/wDixYtCrjPypUr\nkZOTA7PZDEEQkJOTg5MnT8pukwshNTUVhYWFQTt/h8MBh8OBpKQksCyLnJwcbN269YLPBYjuqBtu\nuEF+n5KSgurqaphMJvmzcePG4e2335bTjisrK7FhwwYAwI4dO3D8+HHwPI/Y2Fio1WrZukhNTcW5\nc+eCnl+6zszMTPTp0wdvv/027HY7jh49iuXLl2PEiBEAgO+//14OZsfHx4NhGLAsG/D8aWlpGDRo\nEGbPng2TyQRBEHDu3Dns2rULgBjTKSkpAQAkJCSAZdmLijsRlyYRtTCKi4sxY8YMlJeXQ6VSYezY\nsXjggQc89tm5cycmTpwoj0CGDRuGiRMnRrJZRJhRjjS9R5LBRsN33XUXFi5cCKPR6Hf/uLg4LFiw\nAKdOnQLHccjKysILL7zg4xcPZcR922234fvvv0f//v3Rpk0brFixwud7BoMBzzzzDJ544gk4HA7c\ndNNNGDp0aMBjBjtvTk4OXnrpJfl9x44dceedd2Lo0KEQBAFr1qzBhAkTAAAPP/wwysrKkJKSgttv\nvx1Dhw5FeXk5Zs2ahZKSEhgMBtxxxx1yJ//AAw/gqaeewtKlSzFixAg888wzQds2Z84czJo1C9df\nfz0SExPxxBNPYODAgQDEDLLXX38ddXV1yM7Oxty5c6HVaoOe/4033sBbb72FO++8ExaLBW3btsUj\njzwCQLSsJDFJTU3FM888g+zs7KD/G6L5wERyxb2ysjKUl5ejW7duMJvNGDNmDN5//3106tRJ3mfn\nzp1YvHgxFixYEKlmEESjUlFRgVGjRtUb1CeI5kZEbcW0tDQ5HdFgMKBTp04oLS2N5CkJoskxGo1B\ng90E0VxptKB3QUEBjh496tf/nJubi1GjRiE9PR0zZsxA586dG6tZBBF22rdvj/bt2zd1Mwgi7ETU\nJSVhNpsxfvx4TJw4EX/60598trEsC71ej5ycHMyePRvr1q2LdJMIgiCICyTi6QtOpxNTpkzByJEj\nfcQCEF1VUo74kCFD4HA4fIrYedMIGkcQBEF4EXGX1MyZM9G5c2c5I8Sb8vJypKamAhDr2gBAq1at\ngh6TYRiUlRmD7tNSSEuLp3vhgu6FG7oXbuheuElLi2/Q9yMqGHv27MHq1atxxRVXYNSoUWAYBlOn\nTkVRUREYhsG4ceOwbt06fP3111Cr1dDpdD61hwiCIIhLg0aJYUQCGjGI0OjJDd0LN3Qv3NC9cNNQ\nC4OmYBIEQRAhQYJBEARBhAQJBkEQBBESJBgEQRBESJBgEARBECFBgkEQBKHAZDJh5crlF/XdGTP+\nCbPZVP+OLhYvXoilS7+4qHM1BSQYBEEQCozGWqxc+a3fbTzPB/3um2++A4MhLhLNuiSgFfcIgiAU\nLFjwHoqKCvHww/ehX7/+GDhwED75ZBFSUlKRl3ccS5Ysw9NPP4myslLY7TaMHftnDB8+CgAwduwI\nfPzxElgsFjz55BT07HkVDh3aj7S0DLz++hxotdqA5z1x4hjeeut12Gw2ZGdn4+mnZyEuLg7ffrsU\nq1atgFqtRvv2HfDCC69i3749mD9/jmvdEwb//e8ij2V4IwUJBkEQlyzLNuZh19GGLYmgUjHgOPf8\n5Gu6puPemwNXxH788ck4c+YUFi/+EgCwb98eHDnyB5YsWSavAT9z5izEx8fDZrPhb397AEOG3IyE\nhAQA7oWrCgrO4cUXX8NTTz2D559/Gr/9thG33HJbwPO+8soLmDbtKfTufRU+/vhDfPLJQkyePA1f\nfvkZli9fDbVaLbu7li79AtOn/xtXXtkLdXV1QYUonDRLl9T3m042dRMIgmhBdO/eQxYLAFi27Cs8\n+OBf8Pe/P4TS0lIUFOS7triFqXXrLHTqJApTly5dUVxcFPD4ZrMJZrMJvXtfBQC47bY7kZu7DwDQ\nufPleOGFZ7B+/U9gWXGZ3p49e2P+/LexfPlSGI21jbYMbrO0MBatOoTOrQcirVXkTTCCIJqOe2/u\nHNQaCIVwlAZRrlG/b98e7N27GwsXfgqtVovJk/8Ou93u8x3lqJ9lVX73URKoStN//jMPubl7sWVL\nDj799CN88cW3uP/+B3Hdddfj99+34O9/fwjvvPM+2rW77CKvLnSapYUBAAVloWciEARBhEpsbCws\nFkvA7WazCfHx8dBqtTh79gwOHz7kd78LKdNnMMQhISEBBw7kAgDWrfsRV10lrl1fUlKMPn2uxuOP\nT4HZbILVakFhYQE6duyE++6bgC5duiE//0zoF9gAmqWFAQAFZWb0uTytqZtBEESUkZCQiJ49e2PC\nhP9D//7XYeDAQR7b+/e/Dt999z88+OBf0K7dZbjyyp6Kre4YhhiQDp2ZM1/AW2+9BpvNhqysbMyc\nOQtOpxMvvfQczGYzAAHjxt0HgyEOixZ9gL17d0OlUqF9+44YMGBQvccPB82yWu3w6atwbbd0PDby\nyqZuSpNDlTjd0L1wQ/fCDd0LNy2yWq1Wo0JJpbWpm0EQBNGiaJYuqYTsMpTXNk4aGUEQBCHSLC0M\nc/pO8J23wGbnmropBEEQLYZmKRgSFbV1Td0EgiCIFgMJBkEQBBESzVIwkrViOm1JTW0Tt4QgCKLl\n0CwFIzsuGwBQYa5p4pYQBBFtNKS8OQAsW/Y1bDab322TJ/8dx44dvehjNzXNUjCS9YkAgCorWRgE\nQYSXYOXNQ+Hbb7+GzRad7vJmmVabFp8EAKi1U3kQgiDCi3d584kTp+Crr5bg119/hsPhxA033IiH\nH34UdXV1eP75f6OsrBQ8z2PChEdQWVmO8vIyTJ78GFq1aoV58z4IeJ6ff16LL774FAAwYMAgPP74\nZPA8j9dffxnHjh0BwODOO0fg3nv/7LfEeVPQLAWjdWIyAMDoJJcUQUQzK/J+wL7Sgw06hoplwPHu\nghZ90ntiTOe7Au7vXd58167tKCjIx6JFn0MQBDz11DTs35+L6upKpKam4c033wEAWCxmxMYa8M03\nX+Pddz90lTv3T3l5ORYseA+ffPIl4uLiMXXqJGzZkoO0tAyUlZXis8+WAoBcztxfifOmoFm6pDol\ntwUAmJmKJm4JQRDRzs6dO7Br1048/PB9ePjh+5CffxYFBfno2LEzdu/eiQUL3sP+/bmIjTW4viFA\nWebcH0ePHkbfvv2QkJAIlmUxbNhtyM3dh6ysbJw/X4R33nkLO3b8Lh/TX4nzpqBZWhiZ8ekAp4Zd\nU9nUTSEIIoKM6XxXUGsgFBpaS0oQBIwf/yBGjBjts+3jj7/A779vxYcfvodrrx2ABx98JORj+ivj\nFx8fj08//Ro7dvyOFSuWYePGn/H008/7LXHeWGtgKGmWFgbLsNA6kiHEmGB1RGdwiSCIpsG7vHn/\n/gOwZs33sFrF+nXl5WWoqqpCeXk5YmJicMstt+HPf74fx48fc33f4KouG5ju3a/E/v37UFtbA47j\n8Msv63DVVX1RU1MNnucwZMhNeOSRx3HihHhMfyXOm4JmaWEAgA7xsKMUJaZKtE/KaurmEAQRJXiX\nN584cQrOnDmDxx57CIAoKM899zIKCs7hv/+dB5ZloFZr8OSTTwMARowYhSefnILU1DSfoLdU8jwl\nJRV///skTJ78dwDAwIGDMXjwDcjLO4HZs1+EIPBgGAaPPTY5YInzpqBZljcHgH9+tQBFqv24r8ME\nXNehR1M3p8mg0s1u6F64oXvhhu6FmxZZ3hwAWunECy+urWrilhAEQbQMmq1gpBrEyXtlJkqtJQiC\naAyarWBkJoiT96rqyNQkCIJoDJqtYGS3EgXDaCfBIAiCaAyarWBkuMqDWLmmSS8jCIJoaTRbwTBo\nYwEBcIDW9iYIgmgMmq1gsAwLhteCY2x+Z0wSBEEQ4SWiglFcXIwHHngAd9xxB4YPH47PP//c736v\nvPIKbrnlFowcORJHjhwJ+fhq6AC1A1Ybre1NEAQRaSI601ulUuHpp59Gt27dYDabMWbMGAwaNAid\nOnWS98nJyUF+fj7Wr1+P/fv3Y9asWVi2bFlIx9cyMbCrjKgx2xCra7aT1gmCIJoFEbUw0tLS0K1b\nNwCAwWBAp06dUFpa6rHPhg0bMGrUKABA7969YTQaUV5eHtLxdaweDCug3EiZUgRBEJGm0WIYBQUF\nOHr0KHr16uXxeWlpKTIzM+X3GRkZKCkpCemYsepYAEC5iVbeIwiCiDSN4scxm82YMmUKZs6cCYPB\n4LHNX8BaKtAVjLS0eCTHJeBcNWCFrcE1UpozLfnavaF74YbuhRu6F+Eh4oLhdDoxZcoUjBw5En/6\n0598tmdkZKC4uFh+X1xcjPT09HqPW1ZmRCyrAwAUlFe02OJiVFjNDd0LN3Qv3NC9cHPJFx+cOXMm\nOnfujAkTJvjdPnToUHz33XcAgNzcXCQkJCA1NTWkYycbxCUQq620tjdBEESkiaiFsWfPHqxevRpX\nXHEFRo0aBYZhMHXqVBQVFYFhGIwbNw5DhgxBTk4Ohg0bBr1ej9deey3k46fFiYJRS+VBCIIgIk5E\nBePqq68OaV7F888/f1HHz4gTLRETRxVrCYIgIk2znekNAGl6UTDsrJFmexMEQUSYZi0YerUOKl4H\nQWuGxeZs6uYQBEFENc1aMABAJySAibHCaLU1dVMIgiCimuYvGEwcGEZAhYUm7xEEQUSSZi8YWlYL\nALDY6pq4JQRBENFN8xcMlQYAYCLBIAiCiCjNXjBiVC4Lw06CQRAEEUmav2CoYwAAFoe9iVtCEAQR\n3TR7wdCpRQvD6iALgyAIIpI0e8HQa0QLo85JabUEQRCRpNkLRqwsGOSSIgiCiCTNXjAkC8NGgkEQ\nBBFRmr1gGGLENTFsPAkGQRBEJIkawbBzJBgEQRCRpNkLRrxWFAwH72jilhAEQUQ3zV4w9FoxhkEW\nBkEQRGRp9oKhkyfu2cDTmhgEQRARo9kLhlR8UGCcqKihyXsEQRCRotkLhlRLCiyP8xXmpm0MQRBE\nFNPsBUPFqsBCBUblREmltambQxAEEbU0e8EAgBiVDlA7YKVlWgmCICJGVAiGXqUDo3LQut4EQRAR\nJCoEI1YTC6idsNhoLgZBEESkiArBMGhiwTACzDaKYRAEQUSKqBCMeK0eAGB2UlotQRBEpIgKwTBo\nDQAAq9PSxC0hCIKIXqJCMGLVooVhJQuDIAgiYkSFYBg0sQAAG08xDIIgiEgRFYIhWRh2gZZpJQiC\niBRRIRgJMfEAAE5lAcfzTdwagiCI6CQqBCPLkAkAYPUmWG1cE7eGIAgiOokKwUjQxkMlxICJNVJ5\nEIIgiAgRFYLBMAwMSAITY0GNlVJrCYIgIkFUCAYAxLFJYBig1FTR1E0hCIKISqJGMOLVCQCACmt1\nE7eEIAgiOomoYMycORPXXXcdhg8f7nf7zp070a9fP4wePRqjR4/G+++/f9HnStSKglFlI8EgCIKI\nBOpIHnzMmDEYP348ZsyYEXCffv36YcGCBQ0+V6uYVoAZqLHXNvhYBEEQhC8RtTD69euHhISESJ5C\nJlnfCgBgcpJgEARBRIImj2Hk5uZi1KhRePTRR5GXl3fRx0mLTQIAmDljuJpGEARBKIioS6o+evTo\ngV9//RV6vR45OTmYNGkS1q1bd1HHitfpIDjVsKmonhRBEEQkaFLBMBgM8ushQ4bgxRdfRHV1NVq1\nalXvd9PS4j3eMxo1hF0aOFQ2n23RTku73mDQvXBD98IN3YvwEHHBEAQh4Lby8nKkpqYCAA4cOAAA\nIYkFAJSVebqerDYnwGnghNlnWzSTlhbfoq43GHQv3NC9cEP3wk1DhTOigjF9+nTs2LED1dXVuPHG\nGzF58mQ4HA4wDINx48Zh3bp1+Prrr6FWq6HT6TB37tyLPpdOq4Lg1EBgONg5B7QqTRivhCAIgoio\nYMyZMyfo9vvuuw/33XdfWM7FMAzUgg4CAIvTAq0qMSzHJQiCIESaPEsqnGjZGACA2UH1pAiCIMJN\nVAmGjhVX3qutMzVxSwiCIKKPqBIMaeW99w4sgpOnMucEQRDhJLoEQxsjv66xUVYEQRBEOIkqwcjW\nXSa/5gRaeY8gCCKcRJVgZMalwlnSFgAJBkEQRLiJKsGI02sBgQEAcDwJBkEQRDiJKsHQaVUQBPGS\nyMIgCIIIL1ElGDEaldvCIMEgCIIIK1ElGFoNC0gWBrmkCIIgwkpIgvHjjz/CZBInw82bNw9//etf\ncejQoYg27GIQLQzxkpxkYRAEQYSVkATjgw8+QFxcHA4cOIAtW7Zg1KhReOWVVyLdtgtGq1EBPAW9\nCYIgIkFIgqFWizUKt27dirFjx2L48OGw2WwRbdjFEKNhKehNEAQRIUISDIZh8P3332PNmjUYOHAg\nAMDhcES0YReD1iPozTdxawiCIKKLkATj2Wefxdq1azF27Fi0bdsWZ86cQf/+/SPdtgtGrWLBgoLe\nBEEQkSCk9TD69u2L999/X37fvn17PPfccxFrVENQs+IlUdCbIAgivIRkYbz++uswGo1wOp34y1/+\ngquuugqrVq2KdNsuCjWrAgDwZGEQBEGElZAEY9u2bYiPj8eWLVuQkZGBdevWYfHixZFu20WhUYmC\nQRYGQRBEeLmgiXu7du3CsGHDkJGRAYZhItWmBqFWiS4pypIiCIIILyEJRkpKCp599ln8+OOPGDRo\nEJxOJzju0uyQta4YBgW9CYIgwktIgjFnzhx07twZc+fORWJiIoqLi/HQQw9Fum0XhUYtuqTsHK24\nRxAEEU5CEozk5GTcf//9MBgMyMvLQ2ZmJsaMGRPptl0UGpdLyu4kwSAIgggnIaXVHjx4EFOmTIFW\nq4UgCHA6nXj33XfRo0ePSLfvgolRawAANuelN7GQIAiiOROSYLz66quYPXu2PMt7+/btePnll7F0\n6dKINu5i0GlEwSALgyAIIryE5JKyWq2yWADAgAEDYLVaI9aohhCjcbmkKIZBEAQRVkISDL1ej+3b\nt8vvd+7cCb1eH7FGNQS9RgsAsJGFQRAEEVZCcknNnDkTTzzxBLRasTN2OByYP39+RBt2scRqtYAd\ncJBgEARBhJWQBKNXr15Yv349Tp8+DUEQ0KFDB9xyyy347bffIty8C0cfowZMgIMnwSAIgggnIQkG\nAGg0GlxxxRXye0EQItKghhKrjQEAOCiGQRAEEVYuek3vS7U0iCHG5Ta7RGeiEwRBNFeCWhh5eXkB\ntzkv0RhBrEswnFQahCAIIqwEFYxHH3004LaYmJiwNyYcGHTiPAwnxTAIgiDCSlDB2LhxY2O1I2wY\nXEJG1WoJgiDCy0XHMC5VpBgGVaslCIIIL1EnGFqNCgLPws6aYeeonhRBEES4iDrBAABVdTvwagt2\nFu9p6qYQBEFEDREVjJkzZ+K6667D8OHDA+7zyiuv4JZbbsHIkSNx5MiRsJzXYGsDAKi1G8NyPIIg\nCCLCgjFmzBh8/PHHAbfn5OQgPz8f69evx0svvYRZs2aF5bx6jQ4AYHPaw3I8giAIIsKC0a9fPyQk\nJATcvmHDBowaNQoA0Lt3bxiNRpSXlzf4vAaXYJjtdQ0+FkEQBCHSpDGM0tJSZGZmyu8zMjJQUlLS\n4OPG6UgwCIIgwk2TCoa/elThKDkSFyOWXrc6bA0+FkEQBCEScvHBSJCRkYHi4mL5fXFxMdLT00P6\nblpafMBtrVNbAeWAE46g+wXCzjnwxf4VGNbperRNzLrg7zc2F3ON0QrdCzd0L9zQvQgPEReMYFVt\nhw4dii+//BJ33HEHcnNzkZCQgNTU1JCOW1YWOANKzbMQBMBkswbdLxCbCrZh7YnfsOXMLrxxfXgC\n8ZEiLS3+oq4xGqF74YbuhRu6F24aKpwRFYzp06djx44dqK6uxo033ojJkyfD4XCAYRiMGzcOQ4YM\nQU5ODoYNGwa9Xo/XXnstLOeN02sAXgU7d3FZUmaHuPysyWEOS3sIgiCigYgKxpw5c+rd5/nnnw/7\neRMMWoBTw666OMHgqQ4VQRCED1E50zsxLgYCr4JDcJcGKbdWYMf50GZ+8wIPAGCZqLw9BEEQF0WT\nBr0jRaJBC3AqcIJV/mzW728AANrGZyMrLjPQVwEAnCQYuDQXiSIIgmgKonIIHaNRgRXUEBgnzhkL\nPSrXStZDMHiQhUEQBOFNVFoYAKBmtHAywOu75uGWy26SPw9lnQy3S0oVsfYRBEE0N6J2CK1hNfLr\nXcX75NehLN0qCYaKLAyCIAiZqO0RdSq9/Fq5XKuDr3+NDCmGEY5Z5wRBENFC1ApGK1WK/FoZtwhl\nrW+BLAyCIAgforZHTNG6S4ywrPsyQxEMjmIYBEEQPkStYGTo3amzVoc7vdYRimDwlCVFEAThTdT2\niK1iDXCWtwYAOBWZUaEIhtMV5yCXFEEQhJuo7RENOg0c57r4fB6KS0oKjDMkGARBEDJR2yMaXAUI\nvXGGkCXCp1jIAAAgAElEQVRld4kKS1lSBEEQMtErGDo1IPheXiguKYerym3gwuwEQRAtj+gVDL0G\n4C9OMOwuK4T3M8nvaOUJvLN3AaxOq882giCIaCZ6BUOnBsAAgqdb6UJiGP7KiLybuwgnqk9hx/m9\nYWknQRBEcyFqBUPFstDHqHzcUiEJBud07Ru4jIhADiuCIFoYUSsYgJgp5R34DqU0iCQqoRQqJAiC\naClEvWAIvLdLKvTig1yQUuhkYRAE0dKIbsHQqyF4Bb5DKz7Iefz1i0CCQRBEyyK6BUOnkWMYrOtS\nL6SWlL8sKQmSC4IgWhrRLRiK1NoYVgcgxFpSLsvCKXAQAlgS5JIiCKKlEd2CoZy8x4kLKtVnYQiC\n4FEOPZQlXRsbk92M8+aSpm4GQRAtjKgWDH2MGoJLMBw2FipGVa+F4S0Qmwp/xz9/m4kamzFi7bxQ\nntn2Kl7ZMScka4kgCCJcRLVg1Nk5MBobAMBm0ULNqOsNentnRi0/8T0cvBMHyw97fB7IVdUYSFaS\ng6s/gE8QBBEuolowEmI1YLRiCQ/epocKGtQ564J+hw+QGcXg0itESPNECIJoTKJaMG7qmw2GFS0B\nwa6H4FTDytXBZuewcPVhFJabfb4TbO6Fkksh6B1KxhdBEES4iGrBUCmWZtVy8bDZWFiddVi78yy2\nHy7Bf77yrQcVapC7KV1SEqFMQiQIgggXUS0YADD96okY1u5GXNHqcjhsLHiBh91Vvtxo8YwBcDyH\n1afWhXTcxnIH8QKPUkuZX4FyCmRhEATReES9YHRMbI9Rne9AWis9BE4NAMgXDoBtVeLjVPr9/C5s\nLdoR0nG5Rhrdrz/7K17c/h/8fn63zzZySQG1diNKzKVN3QyCaBFEvWBIxMaoAac4F+MUvxsxV+zz\n2cdo941pSHjHLEKNdTSUPSX7AQCHKo74bCPBAJ7e8jJe2vFWUzeDIFoELUYwDDqNbGHIsKFbCd7x\ngkshQ4kEw82lEFMiiGinxQhGrE4NeAkGE+NtUQTudLznbzS6YPiLYVDQW+ZSEHCCiHZalGAIrvIg\nEqw+sAvKG+/OubE6a4YJPP+Dgt5uSDwJIvK0HMGIUfsYEIzOgnOlJuUn8ist6ykuzqa2MPzQWKVB\nmoO7J9CES4IgwkeLEQxlqXMZlQOzFu9UfODuGPVqnceuDq/RPMc3blFCf112fTGMrYU78N/cj33m\nlhypOI5lx1eFJARFpmL849ensOP8ngtpbqPTWEkIBNGSaTGCEatTgyvPgiP/Ctj+6A8AYNSi1eCv\n89ep9R7vvTtnrpHdQf5mltfnhsktP4Q/Ko/B7LB4fP7e/o+QU7AVxZb601G3nRcF9ZvjKy+gtY0P\nJQAQROSJuGBs2rQJt912G2699VYsXLjQZ/vKlSsxcOBAjB49GqNHj8by5csj0o5YnRoAC2dxR/B1\nBgAAoxI7mYpam+/+3hYG5y0YTT+ira+TtDnFCYoNcZ9J1gnLqOrZs2m5FP4fBBHtqOvf5eLheR4v\nv/wyPv30U6Snp+Oee+7B0KFD0alTJ4/97rzzTjz77LORbApiNIoOz5UtpTaYwPTKwYdbq9Cv9ZUo\nV1nlXXReguEdYG6siXvBqC/obeNEIQxkiYRyDbzLbaViLm1j9FKIKRFEtBNRwThw4AAuu+wyZGdn\nAxCFYcOGDT6C0RhBVYZhMLhnazh5HtsPl0DgVECMGSyAYuzAD9U74CjqCE2WuL9PDMPHJdVIWVJy\nIF68R8p4RL0WhkswuAD72UNY31wKJrOXumBcAgJOENFORHuBkpIStG7dWn6fkZGB0lJfv/n69esx\ncuRIPPHEEyguLo5Yex6+sxsevqOb+MZ7Eh8ANq5Kfu0tGD4xjAvooOqcdTAFmEXu5J04VXMmYNFD\n76Ra7gIEo04SjADHDmU9DU52SV3igkEWBkFEnIhaGKFYDjfffDPuuusuaDQaLF26FE899RQ+++yz\ner+XlhbfoLapoAUPz9iFKsEtGMnxCR7bzJwJKSkG+b2d57D5UDEcTh7jhnUJeq57v5kBAFg27gOf\nbR/vWYp1eTl4/JrxuKnjdT7b1WrRlabRqpGWFo86h3s9D61O3BboXkgWRHxiDNKSxX2U/xN9vLre\n+6g96Tq/StXgex5JEhJFgb+U29jY0L1wQ/ciPERUMDIzM1FUVCS/LykpQXp6usc+iYmJ8ut7770X\nb70VWl2gsrKLXzJ1wfQhmL//CE7XBj6GYPcM8p6qysdHO5a53xdV4cgffwAAbr4qK6Tz+mvzznNi\nrah9547gyviePtudnDjCt9ucKCszwuJwx1lWHlmLUd1uhana11LgBR42pyiI5ZW1iOfEc9tclXoB\noLyqFmWa4Pex1iJaRoLANOieR5qyylp0TmnY76K5wws8BEGAihXFvSXfCyV0L9w0VDgj6mfo2bMn\n8vPzUVhYCLvdjjVr1mDo0KEe+5SVlcmvN2zYgM6dO0eySQAArUaFWI0+6D7eLikA+CU/x/2GcY/U\nrTYnVm05jYIyk8936kNy9YSa5ePtevnm0Gq/+9kVwqAMeludVr/7BMLqWqHwUgx6K914NHEPeGPX\nfEzNiWzyCNGyiaiFoVKp8Nxzz+Hhhx+GIAi455570KlTJ8yfPx89e/bETTfdhCVLlmDjxo1Qq9VI\nTEzEa6+9FskmyfgTBCU6dUzwAzDuzmrN72fx4/az+P1wMV7/+8CAXymrsiAtKdbjM6kjrr/DEwXK\nWzACxSGk+IX3d6yKJWrrW99cuf+lOM9BKRj1TaS0OutgcViRok+KdLOajAJTUf07EUQDiKhgAMAN\nN9yAG264weOzKVOmyK+nTZuGadOmRboZPsRpDEG3/7b3PBDMCFFYGL/uKwQAlFZZfXZTjuif+nAb\npo/rix4dkuXPWFkwgge9pbN575cS678DVLqeLE4rquqqkaRr5SEY9hCC3lL7bSFYI42NRwJAPSnG\nL/7+JowOE+bdOBtq1v2zLzGXotRajp6p3SPWzsamOZRyIZonl56foZFI0rUKuj2voB6fp8LCsNrc\nnVVRdSU2F/4OO+dAta0GT26apfiOgEOnKzwOU59geOM9klZ2fkqk+AUAfHzoCzy7bTZMDvNFWxj+\nBEMQBNQpjtfYKDPV6nPpGR2iu9A7PfqlHW9hwYFPfWbD+8PqtOLzP75BhbXyIlrbeDjJPUdEiBYr\nGCm65OA7CAycFZngqlMx54aX8UC3cR6bGdZ/B/W/wxux9NhKzN37Ac6bSzw3MoJP9VlVkBhGjckG\nh9Pzc2/XVSBXkY3znb2+pXAHamw18vtQ0mqlOIeDd/iI2rLjqzB90/Mot1b4+2rEUbraQk1zDnS/\nQonnfH9yHXYU78HHh74MrYFNxKXoPiSig4i7pC5Vkr0sDN6mAxujHC0zcJy8CgCgZbXINHhmd0Hl\n+VBe2y0duXnlyCs7DyQA+cYC6FRecRKWh3e1cqnkhj8LY+p7WxHT3Qw2zl1LyltYvEuWSNT5EYzV\np9Z6vA9l4p5ytGrn7B4z4DcVbgMAnKw+g1R9Sr3HCjceMYwQR9UNEQyTy0qxOOu3RpoSEgwiUrRY\nCyNZ5+n7F2zeAQu3H7isxoo4lWc6GqPiPPbpnJ2I7NQ42AWl6HjXU/f1LdfvkmI8DiUJRvuEdgAC\nlzgPJeZQn2BwPOfRrkDHbKpJc6FaGMoONFCZFBtf//2S7r3qEq+rVZ9gVFirkF9b0EitIZqaalsN\nqhWehYbQYgXDO+gt2Dyzl27sm4Ur2ohzRJ7+cDvmfn3U9yAKK6NVXAwMejWgltJQVT6dE8PwqLN7\nfsa6TI5QO13JJRWj0gIAnAHcSsoYRiCCuaQsDisWH/Z0vfhzcwFNt3iRMp4T7P4p4xPK4LiHGDpD\nEAzXdarYS08wlIHu+mJTXx/7H+bt+zDSTWpSjlXm4VD5kaZuxiXBM1tfxTNbXw3LsVqsYDAMg8d6\nPSi/97YwDDo1+nV1u6GKyn3dEIzKgZf+ei1u6puN3p1TEafXgNG4C/6dr/QKnDMCrHWeoz+3heFp\nfQRKE5U6Rq1LMCQLw8bZsf7Mr6hzCUWgzl1JMAtjw7lNyC075PFZiaUMB8oOB2xTY6M8b7BAr1Iw\njledxObC3wF4phjbQ7IwXIJxCVoYSvGrb2GtKlsN6jibj1X2R8UxbMjfFJH2NTbzcxfigwOfNHUz\noo4WKxgAPFIp777OM62ydWoskhM8YxBclVccQ+1Em7Q4jL+lCzRqFgadCtCIHQ/DAHvzvIPePCw2\nz4dZFSCGYbJ4duZyDMMlJLKF4Xrolx9fhVWnfsL3p34C4D+GIZERmwYAcChcTA7OgV/yc2C0i356\nf4Kz4MCn+PDgZzhZfcbj86bymXtM3Ati5Zgd7jpey45/h6XHVsLssHgISSguPOmeMD4Vvpoeh4fb\nLfj/w+K6bm9h+e/+j7Ei7wd50EEQ3rRowVCSluBZO0rNAu3S46BWsejdKQWPDu8OR35X8DYdeKvo\nzrqqS6LHdzQ6p0dQ21jnNS+DEXDwVAXyCmtQWi1uqzWLwuAtGEYvwTBZXYs9ebmkpIf+ZM0ZAJBT\nPoN1gFL8xq7oMH49twUr89bg08Nfi00N0imWeC28FErAOBJ4xDCCpNWa/KTMztj8AnaV7JPfhyIY\nUgFJK+c736bYXIKcgm1NNgdC6WoLZmEIgiALRqC5K6GkGEcLds7hMVcq2gj377HFZkn5Y1BWf2wt\n2gEAiFHHILWVHv+degM0alFX+1xxO5Zu6AxD+3P4teRnDL4q1eP7lapTgKKfL60xAcpYOcNDEIDZ\nS/YgwaDFv/7cB+fKjFAleqbLWm1O5Jd6urOksiOSsGjlGIYTJocZJRaxxIqaVcPisOCcsTDgdcZp\n4qBlNXLHAQCVtmoAwDmT7/c0rMbDL27xesCCWTNKzhmLkBGbKre9oXi4pIJYGJYAHeCPp3+WX4ci\netJcDmU9L4k3d78LG2dHRmwauiZfXu+xwo0zRAvDwTtk951yP2XHYnKYomZGfH3p1s9ufRVmpwX/\nvfnNRmpR46J8RsIhHi3ewojXxgEAErTx+EvXu/HMtdNwW/uh6JHSFQBksQDERZgm3NYV2UliSm4d\n79lRnnEcgsCpwBld2707IUWWlIk9j+NFJfJndo7DkvXHsOdYGWYu3I6PfvAM2Dk5Hk6Od1sYrGRh\nOPDt8VXyflW2Gry+az6OVeUFvOZYjR7JuiRU1rmr80rBd3/ZWglazwyxyjpRXCQrxBrC5L2ztefw\n+q53sPDg5363VxltKKu+sJGeMs7zw+l1KLf4n1Bn5epvX7CYjyAIKLGUyddpddb5PHyShVJsrn/Z\n20igFMxgguGRAKDYz6xIFfZnkYWTs7XnsOb0zxGzxpTHrS8T0NyAFOkySwUWHVyCKtfzcCmiHOiF\nI9bY4i2MGf0m40TVKXRKbA8AyIrLRFZcZtDv6F2FCyVTluM5LD22AkauGoI1EXCIdagY1hWgFmJh\nZyxol2HA2VPA4GvisYdZi2+LDgCMmJ1VXmvGmX2F+HVvYMvgx9/PIqOD+ANgoQEgPvSVio6y0lol\nj4SVZBkyUWQW1xqJVeuRrE9CsaUU5dZKrMxbA6tD7PAkwVC6pBK0caioc5+jwjVRT8Wq4OSdIc32\nliygI5XH/W6f/t+tAIDF/7653mNJeE9iXH74R9zdfqTPfqH45INZGHtKcvHJH1/L7zmBg513yG5B\nQExe4AUelbYqf4eIOMpFsoK5pJTW4cZzm+HknfhL13tQY6uVP1fGfCLBm7vfBQB0T+6CDontwn58\np8e9qH+uESA+wxea/fbl0W9xovoUVAyLh6+874K+21jYudB+F6HS4i2MZF0S+re+2mcGdjCk9b6l\nEefRqjxsO78LACDYdRAE1211CUasVtz/nps64L1/3oD4BHEExGhtYFwlRjiNCWCD/UMFfLflNL76\n5RgAYOfhcgDixD2T3YxkXRK6Jl3uVywA4OqM3vLrU/kWJKpFK2jx4S+RW3YQx6rF43I8L8dLJOK9\nLAyp85dmqQeyMEwOM/aUiOXbA5Uw8b7GMkvgWeMlljIsObJMdjF5xy10AVxddSFZGIEF41DFMZ/P\njHZPl2FyjHg/KwOMNo9UHsdHB5eIAl9XhUkbZ2BL4fZ62xUqTiE0C0Ppnssp2IatRTvh4J0egmEK\nQTA2FfyOY5WBrdhQECDAyTuxMX8TamzhKz+uFIlgqeOemWWe+31z7Du8vef9oOepC1I2JxR4gceB\nssMRTRpRXlc4ztPiBeNi0KtFC0N6+DSKkYmejQMEl/i4BEOvliwOAbE6NZLjFSm8CjeVpp17rsdl\nGfEY3Ku1z34Wm/jjLCqzQsWo4OCdMNqNiNfGITkmcLmTGJW7+u7BE0YcPyUe52ztOY/9HByHqe9u\n8VjqL8HltpMos1bAZDfL0xKVMYw6uxN1dvGHOW/vh1h8+EscrTwBdZBUVMmFoG57DC9sfyOgFfLR\nwSXYfn431p39FYCviR2r9V8tMpQ5KcqHnuM5HCz/Qy7O6G8sccbrvsVqREuxIkCZlPdyP8K+soPI\nNxZgvys1+etjK+ptV6g4L8LCkKhz1qFaaWEEWB0SEDvg38/vxjfHV2J+7sKQyssEghd4bC3aif/l\n/YBFAVyVF4PSDRXMwlC6IVfk/YBFB5fI7w+UH8bJmjNBxUD6/VucVvzvxOoLDp7nFGzDhwc/w/IT\n/pcoCAeOEO9FqJBgXARS3ENKQVVWfb217+W4tovo0hJngwN6jdhZSx1cWqJSMNyjHFXKeQzskYGJ\no67ErIeuwd1DFGufS8Ii/RVYsGBRW2eEU+CQoI2Dw+w5+VBJnMbd6QucBtZa/6NxhhXAXnbAY/Tl\nbWEAwKmaM3JnoXRJTXx7E6bM2wxe4GUXmNlhCVpcUSreqE4XO+HD5X4mScI98pUsGu+AZqBg+rmK\nwD7m7DhRlJWdx8Zzm7HgwKdYmbcGgKd7Ttr/lCsrTULqpPy5v5QWmCAAWpUmYHsuFs+02sAdg9lP\nwP7fW17C/rKD8vtgFsaPZ37BF0fcC4ntU3zvQrFxdnkGcr4xfDPPleVygsUwlMkLW4t2IrfsIIrN\npR7t8rYk/XGq5gw2ntss/15C5axrtv3hCv+/93CgFHRySTUR0ixx95wF9yhEzaoQqxM7LrVa7CSl\ntTUkF4pe5+6ADLHukTdvTMLfhveQJwwmGrRITxbdWR2yXJ22LBgMnE5GDvSWlfPYsivwjztGUMxs\nd2qgsicG3FedVugRRFXOio9jxJpRO4r3yHNDvF1STk7A3L0L5PecwHn8WJXBdgAwSi4wTrwXoUyi\nA3wD9NJs7RqTDUvWHYOlzoFqkw1nSgPHFeQUY8WDddbVeW0q3IbcskMe7spOiR2gYlTIry2A1WmV\nvyfNafE3Ij1ZfVp+beftcsJCOAk1SypQHaxDik4rmGB4lxTJU1zbhWLjbBdcrTkUPF1SF2Ztvbzj\nLfx6bov8vjaIYAhepX9MfiyzIlMx9pX6F1XJMyH9vywOK7459l2DqiGXmEux/syv8v20k0uq6VGz\nasSq9XK8QBkwbROXJfv2e14udsrSyFcaESs7z1idCq1ixP3aZvq6VGK00r+IR0KsBnBZLe0zWoHn\nGFmEzhU5INTF+XxfQnC4JyEKPAsVFxd0EakCxahco1LLbazKF91eylngUozA5nCLjHIE/sXGQ/hh\n+0n5/XPbPBfJkiYpCrxLMAK4Obw9QzUWT6GycXYsP/49nt3+MjYVbcXKLSfF+SyqwA9Ksi4JDBiY\nHWbwAo8jFZ7usEUHPwerOHOsRo9WMYk4XZuPJzfNwqeuYLjUZn9ip3Rf2TmHd4WxsBCyS8qPhSEh\n/U4tF1Cy/nTNWZ/PVp9ah8m//tuvi0b5v7VxdlkwvDvfhhCqGybQvZAqAQBA7QXEVjR+LMdXd76N\njw75z6RSueJ60v9u7dkN2FS4DYsPfxXyOb15ffd8rDr1E/5wxd3IJXWJEK+N97EwhrW7EV2TL5dn\nb0sTo6RsGs5P/jsncFCzamhYNTR+PBXSSIETeKQk6sGoxXON6H8FILj/fYJDi9v6XBGwvTaz4uBO\nDVgwaBvfJuD+J8vd2Vqck8Hw9L+g7vBAOM939D02Zwcv8O7Z6V7B+zrOiqLKwMvXykF2WTA8O90q\now3Lfs0DL/UprpjH4TOe8QK7045DFUfAs3ZoLzuKU/wu1JrtYLwE4/FeD8mv4zVxSNYlodRajg35\nm/De/o+wr/SAx/5KC0PLajwqHR921SuSihfaOLtPuqgyTnS2tAq/7PXtZC+UIlMxlhxZJrvAQg16\nB0sjzTZkQqvSXpAv/ry5xCdLbu2ZDeAFHqf8iIkyA8vO2SNSZsXOhSgYAa5TmQAQzMLwRs0ETuyo\nqPO1ciWxlP5fVod4H6v87Bsq0rMjXZuHS6oB8SYJEoyLJF5rgNlhAcdz8j/p8iQx5iA9BNLnsoXh\n6vyVI0CO56BiVNCyWr8/bqnGlCAI6N05Ra5VlRrXCjq1wrXhiEHrFIOniHDuhzG/xIwXBjyFbNMQ\nCPZYgAF6p/YIeH2M3v1gf/5THhb+7zQEcyJ8x/kitVYL3l1xANDYoO/3i+ex1A45xVjim40n8NbS\nfdifV+4jGJVmz07tvysPYu2OfNSaxftZXiM+WBVG8aEQHKIYHj5bCqvdAYFTga+LRbHqEEqM1T4W\nhkHjjvXEavRIj02F0W7yqZ2luAL5lValRasYt2BwAg+O5+SHkRd4n7pWysmQq7efxKlid4fgLS5O\njkdFTf0j/PdyP8L287uxqUAsMa8UiR/P/AKjzb9AW4NYGMm6JMSq9bA4rDhaeQKTNs6o1+UkQEB5\nABeKt+sRgFc5FltY1op38k6PtWeUz1GwVSUDueeU1s6FCEawisf+3ExS3Ez6vbBs4LVxLhZPC4Nc\nUk1GvDYeAgSYHBb5Hy9ZEtJD4BYMsUOT6h15+JsFJ1QMC41K43cugKCwMG7vfxlSUljX+eOg07qt\nBsGuQ2ZyLFSu+RlcTQrq9ruXxt19tBSp+mTE2kSrwmx14I+9Bqh5/5lFUsBePLjnz8Rxzncm88b9\nZ5FfYoIqsVz+7NrMvuILlcMjuA8A63adxR9nqjBv+QG5DIrkkjpTKprvuSfK8d6KgzhVVOvxXXOd\nuH+VSez8JKsnv7QKRrsJgtUAriwbYAScNp72sTCUgqFX6eTaWkWm837vhTK4rlVpEKd1f1/sMCs8\nOhnl/5EXeE/fNst5LL7lHfP46ufj+NcH23D6vOc1eyN1ZFKGmrdVsTV/N+ycHW/smo/NivRdY5AM\nqDitAVomBhanFd+d/BEAsO7sxoD7S6nSyjk6Skot5T6f1Sg6YBtnv6B09kB8/sc3eGXHHLnGWUNd\nUkqCxjC8xD5QRQEAKLX63gvJOpS8CNJAM5yCYQ/RVRkqJBgXSbxGypQyyg+9JBisK5gljTiklFZ/\nJRmszjqoWRW0XuU3JKQfDy9w0KhZpCQzYBkWsWo9DFq3hSHYY5CRrIeaEQVDcMQAzhjYT/cA8gai\ntNqKFZtO4dAp8eGuNtmx+0gljLuvB28LHMsQD8bg0eHd8ehwsUCj83xHOMtbe+xSXF3jOq9bxDJi\nxeA9o3YA3isUKiyOo/mukSgv/hwZlsO+E2WY/78D2Hu8zKc5VpuYumu0ivdXcLrOqXaAUfEQnFpw\ntWLZlmO2XWDUTgic+6fOcu77tvC7E+CtogAEyqhR/l9ES9DzwSu2eLZRKRhWZx0ECLK7gmF5D/H0\nniT3W26R2O780GYPS0LlLRh6tQ7Hq04i31iApcdWYHdJLowWO44WBp6JrmV0OF/mgNXhDuaXmMvw\n780v+V0/o7Xr/+s9ek5yWWD+Zr1vK9opv7b5qZh7MewpFef6nDWKrj9HiC6pGntgUZYyIWvtRtTY\narH65Np6Kxr4EwzJ7VTmRzyVrjzR0yAlAISv+rPTI+hNLqkmQ6q1U2otlwVDK1sYomBIIyzvGIZ3\nh6NiVNCoNKixG7Hx3GaPbbwsGOJfo92EOI0BLMNCrxCMB4f1RnysFipB7DylUTVX1hYPDL4O+hgV\n1vx+1qeMOsACXPBJdXcO6IABPTKRKqcDM4hTeWZZ5Ve4On2lMNh0EHgWsQbBQyDE/dzvD5wtAiC4\nM8BYDut2es5zUFJSZcX+vAq54xVc7We04ohRx+qRrE4DBAZmeI1+BeDpD/a43zq1yDsWPGvpWKH7\nYeccLK5OFydBSpaJdzFGKY7x+dqjWL5dLHAYKy3AxXIe114TYATr5HhY6hx44ZOd+GW3773wHpl7\n19Kqc9o8XEKfHP4Kp4trRGsvAJxdDTjVAAOYXPG5irpKGB0mfHlEnDOitKQyDRkAgJ/zc7DkyDL5\nNyp1kiavSaS8wONg+R/y82F12MI7ac3121aOqpceW4kKi/+YQHVd4EWFLotvCzWrRq3NhC+OfIu1\nZzdi+Ynvsfrk2oBVAcyKmAgvCHj3fweghvjb8mepKEvWmJ2WiFgYlFZ7idAmLgsAUGAsCuiSktCy\n7iypM7X5+PbEKo/tKpeFAQD/O7Haw+8qPaBSR2+0m+XRj+QSiFFpcX3PtgDcJUOULqWMZD06tPas\nxqtEUAhG2/g2MDCey9cmxYsWkkHv3q+Dqg+cFa3hrBDnnJQbxc5BGatYsek04NRAq3P6rIF+x3Vt\nkJKgAxNjgb7vr8joc0zuSBmVE8fPBR9hf/j9YberidNA4FkwWvEBTI1LRFK8HoLdbTk5Czu7rlUD\nTtkUpxrxbBKuSOoc8FzKTBmjmccVSZ3wxuBZGNnxDgDAqpM/eexv42z442wVck4cwU6H+L82Vrvi\nSSznIarKAKdUYBIA6uwczhYbkV9iwle/nMD5Cv+uJOlzKcGiZ2o3AKJlU+iaByNxrqpCtPYCYLOq\nIHDi78c7OJ5fVo3P1h7FiQL3/0XKyqu1G7H9/G6UWsphcVhlq8nG2cHxPH7dWwCrzYlauxGcwKF9\ngp5oOOIAACAASURBVPhb3ZtXHHQdk/rILzHi87W+cxi8rYrdhQd89gEQdBW6VrpEJGjjUWs3oszl\nTtp+fjfWnt0ou+u8UVoYNSY79p0oh80h/l/8xVKUc3bqnHXyIMDJcXByDRMNwY94kmA0IZJgnDMV\nyqmUsmB41aSJUQS939nru9KZaGG4R7nVNvdDyStcUg7OgTquTnaHSSM8pR/8xvbXAABu63a1/Jle\n67kYlA+u2IGW1eDf10xB70zPGEVGsmhZGHRud9O9N3ZFRu110NSJrh9G5cSgnplIT3Ffh7MqHQKn\nhsBy0Lg+7p7cBQDQ/8pUvPH4QDx0t3gfazVnkJHsmhGvtfmMhNtlxEGjEn+ukoBJGWOJOgPAq8Cw\n4kPSp0MWurZLguBw3ffaJDgrxPPI7isXAqfG+UoLLk8ILBjSGicAUFhqxQ/bzmDOV39g8Xdn/O5+\ntrQam/cXQRXvtm4cVnd9MaV4StkzgiDg+Y/d7poft5/F5+vcJUmKK7zcHa6B/t7jZeAFQR5JsjZR\n7MtrjSg0esZkfij6Fow6SMqthREtDH+onMjJLYLD6bYwtu2rggbuCgKv75qHf22eJcdV6pw2/LQ9\nH0vWH8dHPx3Egv3igkaJajE12+qou6BUT++YwSuf75FdeEq8j2nQxsLisOD1XfOw8ODn2FYklvGp\nCiIYBnUsErTxqLJV+0zGlFxw3u2p42zILTuE/NoCVyKHIKfB+0u3Vrqk6jibey4KI3gMHi4Gd4IN\nzcO4JIjTGtAqJhGFxiJ5wphkSXinCmoVLil/D4hoYbgfVGU9ImVareRzTYgR3RvXZw8AALR2CQcA\n3NbxBjzVbwqGX+Eu4hejVWFI7yz8c6y7nhQA9OjgKiXisjCkH1m7BK90W9dzEatztzG9lR4v/bU/\n+nfJdl2EExNu64qh/cTYRrZpiChEvAoO3o6+XcRzSQFnB28HyzAex9QoPEOsoQa9O6Wg7xVpmP/E\n9XjhoWsRFyvu0L2Dq/S2a7TMcFqPjLAEXRwG9cyUG945qxX6XS7eIz2rR5u0OGRX34K72t6FyzPT\nUVJpwffrA1s0jEIwNueWYMWmUzhbbISxxn/n+vnPf+Do2SpZsADI1k5KksYjhiH9rw+f9g0cl1S5\nXRw1Fs8OR9lXlVVbsf246LbatV+0hn7cfgJnazw7U9bg6xaRKisDgMkICLz/a2Jj6qBuc8xzXXpe\nBdbhntTp/du2cXbkl4jnPG07iHMmsT0qh8s9p+LgVEysyyuowRfrj+G3fYU+9czeWroP73zraSl4\nj8Klxcm800ctDisOVRzFOWMh9pcdwpdHv8X6M7/6WBgjO90uv47V6OUqzd712Q5VHMU3x1bKMTQA\ncvHSRQc/xxu756OwugxgBDByNWpfwVC6pGzO8LrnpGOFGs8JFRKMBtAmLgs1diPK6yqhYdWyZeHj\nkvLKkvJGxag8ROb387vk0YskGIIgyJN/pKBia0MGXrp5Oib2flj+LsMwaJfQBizDYtq9vXHnwMuQ\naNCCYRj06pSCuf8YhFZxYkd2W3+xUqjUmUkxlnZe8zPaxouioFb5/lx6tBU74lv6t4ZaxcqB41v6\ndUC7jDhkJsXDzjnkY0uCIVlFyrIbyh83a6jBwCsz8Y8xPRGn97IKwGPGn/ugXZY4ur3vph5gBPf9\ni9fGISMpFhkpomWk02rwyJ09wTIsumW3xkt/vRYzx/wJt19+Azq71m231/qWP5HvqZc7rX/3DDw3\noR9iNQGSBVgOtRbPQL+0BHDHbIOHu7DCWonCcjPeXrY/4PkBYEXOKZRWiVZGrcXunpMCYMfhEhTX\n1Hich9GbwTMBgvjn28uv7Sf6yq/PFtqgjg2c6aPJOu3hchR4Vs7K80ed04bjLheWco7Clj01oguR\n5WB1uNs4+8td2Li3EJ+vO4YPvjsEQRDwyY9HsGTdMfxxpgoHT1WgpDJw+77fKqYAe5ezt9itOFzu\nWUBylWtlSo+qzGp3XM6gjgXnDNw9bir83cPF0ye9F2IYd8ZhbuU+j1iVOFdJQG5eOTieh51zeIhI\nHWfzOJ7VduGuui0H3BalFOBWpjF7u04vBhKMBtA2XnRzVNZVedQx8rEwWLdLyh8qRgWHYvWz3SW5\n2OvK/JBiF5zAyyZ0kmLiWNe0znJ5C2+u7JiCu4d08giQJsbFYOb4qzH13t7o0T4ZN/fNxg0dPS2P\nrDh3BtT8G19DnNY9ipw0+kpMH3eV/D5BL25jY0SzXfLVJsXG4oWHrkVKnAECBNn8lgRD2i/QDGVG\nZ8aVHVL8XpeTd6LrZUnQ6DioGRWu6pSJtqnuhz1VJ1ozIy+/BQBwU9vroVVpMfmqRzCm83CPY/W5\nXAxcQ1BhSpfpSFSL91LgGXBVab4n59T489DL0aF1AjKT3fdFa0uFI190t0mxlaREdycpxYlsvN1j\nguahgkJ89pPSDy8AfmY9m6wOvP7lXgDAb/sK3fswYhVjyT0nCQYbJ7q6HIWd4Ch01yRL4LPAlSkG\nBAoXXU0N0DnG/b/1hxQnAgDwKhhNQWZoMwJqXbPxVayiq3HEiGVgWA4HTisyzBQd7JGzVfjw+8PY\nfOA8ck7lQttlF8A68dOOfJQHWTfF5uDkOSh3d74LALD7RAF2FfiuRQ8AYy6/S3596A/Fb9GmwoFz\ngRMvAICHYrKknYWzzu2eO2M+6XE9ds6OX3YXYP7yA/jqtwN4dturHseqc9o8LACr7cKsDUEQsPhH\n9xo6UhJErSN8VYABEowGIcUxAPeoH3BniUhIMYxApcfVrMrHjD5SeQIAwMMdw3BbGIHrQIVCaqIe\nPTuKnfH9t3TB2GuuFdvhEjoNq8bjvR7C9Ksn+cRjru6S7nZlQbRGtCotDpQdhiAIsq9WsqqkYL4U\nRI2VBUOaGe32Dzt5p3xt6rQiLM37xm/7pYfB7LDAoIkFwzAea1Ok6sX29UnviblDXkGPFLEjvyKp\ns89Kch2zEtCvSxr+b+jl6JKdgSSdmBwg2HXgje7rFATAfrInBJsBCQat6z66LYwJfYaDd3XWOh3Q\nKk6L63q5V2S8vpM4SdLBOaBz1RLjbTowGjvyCsWBABNjgf7adVBnncTrfx/gc93VJjtKq61Yt/Oc\nPDKWrB9GbYeKUePqTqIYMCrX78acgPuvHiofQy3o5fRlEQa8yfV74tS4odOVuFn1NzhL2vqcH3DF\nlyR4Vly22BK4JI00adIuKOam1BmgZrSAygmrXXE8ZSYd60Su8ycwhhrEdNkDVWIFVClF2LS/CDMW\nuEt3eMByWPDdIew4kQ8A6JjYAQBwpCwPjNbXJZSh6oCuSe543bY97s71fKkDzoLgKydyikHemUIr\nHA7FhD+hwmP+j5134Gi+6HbcVrJNHvnzZvH3dryoHFa7u42WOv/WYbG5FL8VbAXnyqKTz+e1pPNJ\n1+RQk90ENatGliH4Gj+hQoLRAC5LcD9UbeLd4uHdyUrvt5/f7fc48Zo4H/+lVE5CDnpDkJdRVVoY\n4UCr0uLpa/6J5wf8S/7sytRu6Jh4WQjf1aBnSjeU11Uip3CbbDlIQiFZXmaHBQwYxLpKw0uCoSyN\n7uAdiNfGyYK7p3S/h59Z6iTdxdosMLgKI2YrrCKdokZWfcvBsgyDiaN74pZrxP9lgk48HqOxe8RF\nBEsCuIps/N/N7uB4crwOzhLRrdcpuS2GDxBH8tdfnYSX/tpf7gCnXz0RE/7UGxpWDTvngJQNLdh1\ngNoOyVro3Uf8q2mTh/SkWEwfdxWm3NPLo70vf7oLVpsTLCvei/atXZMI1Q7EaWLx+IjeHgMWwRaL\nHm2y5ffdsjIxqKf7t9o+Mx4JRTfCunsYAAadshPx/+2deVwV57nHfzNzVg5nAQ77JqsiKosKLkQR\nCbihUEEbkza9as1iNKJZDPfT2BtTc29MbZO0ualNW5PWW1vbmn760U+allSjDcFoJGpQEzSKGAHZ\nZD/bvPePOTPMcEBRIQq833/kzHZmXs+8z/u8z/P+nonR/nBcSgSxCyNm4lAjtKNnEah0bcKB2LyQ\nZ30Yhro02M5NBukVNBen37odwv9znCMLcGqg5TTCPlk8R+69qAIvgfO5Bm1Cec9+WYbX1cYOT80B\nlsdn5xtBVN0gThWOfCp00Jyx7/gU59LDSyW0n2A0e674/se14Nv84LjiKYUjIvcwyk8rZWoYlgdr\nUK7zOPWVkH7tsvd4daLBOHSyGqdk3laHree9IITg9FeNuN5uw39/8ir2fvFX7PzgQzzx08NobhOO\n612t8kRVvbBWyd4OL9YAa5fyd3S7UINxB8g77iCvniyk3lNS/ckf6DihY5sTniHN/U+0jke0eQyu\ndtTBxbsU6zDEvHq5NzNYhBlD4Kfvv57GjciLnge9So/3L37gIYciehodjg6oWZW0XXxeeYaXg3dC\nzaolowIA+y/8Q1rcJV+kxhMeXc5ueLmrH4qyLHdKsv9EAO6OThYAJg4tCu6LQk5aT4U4J8/DcSkB\ntmM5MGq8kZuYBL1Kj+MNn0KrgSx7Tuh4NawGNt4OTkVAeAZwaIRaGyoHOJZBcoxyCiw0hMMvq18G\nF9ijydTR7cS89AhpiifIXweDTgVG7YC3xuD2tnqmRhKCQ+Bn6mlPs16PlfenwnI9Gd2fT0eo1YA1\niydImXI+Ri1iQk1ISwiAUSecp2I0eHC2ctoSALzcWQrTxgfC4owCf90ffKcyfXvt0gSEWg1wQmiL\nU+e63OfqBM0xWZxHrhIgBtcZlhfaCnAnORCwxkb856/+7TFxJ8ZXGLUdxKHBB5/UKfbbKtNhvzRO\n+nypxoGNPz2G7opZsJ1JUxwrZdO5biBFL4/nuFRwNgiG2KoR+gLWolyf4xINjCxxQBrocE5FfZc/\nHDqHr6624vi5a1i/cz/+96v/wVN7/iBNW50i74PRt+HVsv/D2x8cx4ef9coWY3gcOXkVzV1taGkB\nPjndf0bYrUANxh2SEiBYbrm30dtA9Bdj2Dj5MZSkFcNP7yv9EERxOwKC6/ZWRfC7svEcfLQWqZO8\nV/D38kOUKQLX7W3SQjRRuVM0EDaXHQQ9noetjykpAgI1q1ZU5/vo6lH84tTbwn638XQSJzocnSAg\nkocxzicWLMNibrjnSPhWmBqUgimBybBfHA9Xqy/CtNGwXxoH+4UJCPBR1htJTwgEwGBFttAJ6VQ6\nTAlMRrujA1931ErZc+J0mU6lQ7ezG3odAxWrgr9RmAoK8ldhTLARKq5noOHiXSit/lBow8gz7pgE\nD4Bgycwo6Tgn78SP1qSB4ZySDL3ObTC81QY8tUxIsxYHGeK6hycz85EcEoOiObEIdD9XsJ/wL8ey\neHTJBPga3OnbZnOfv+HiolQ8/70p8DXpoFa5f/O8crBkNnGIDjH1dK7u/RaDl5AGLXbyYBA/vu9p\nGDUjPA+r6wQXcBnahE+gDvvS80CGF9pIZQfH6wGeAyE9XgPf5Q1X3RjpsxhXInYvmLz0eHZFirTP\n18uIZx5I8UjD7o8N30rF+sw8lKQVY1PaI9CwGqj8lOtgGLcop+gpWfV+SAoRPFaGcyoMEKPtxNa3\nj+Hn+07BbrootIOswBqjckI38d+4pjqDcvu7ioC38F08/lZeBbA8DGoDHsicOKDnuBnUYNwhDycs\nx6bJjyNeNsKVT0k9NXktNJwGP83c5nFugN4qjTASfYVOJ94nBmatMEprsV33kH0eihrIg0GgQRgd\n17QJQntioF/DypMBWMmA/O3Ce/jHpYNSpyiiYlWSYRCpbDwHF++SYhdO3ikV3BFXW3upvbC78DUU\nxC68o+dgGRb/kbgCrvoIwKHDipgHhU7GqUVyrFVxbEyoGT8vnoU5KT1TPqKnea2rUbagU+u+Rz06\nHZ1w8k7o1RpMiRXiDQ8tiMJTy1MUWTPvnj+gWPWvHV8Ofdr7CJh6AloNJ02e2HkHCCecJxoMsY1N\nssJXa5NXIdFvHLLC7wMABPh4Yd3SSTAZNPDWq7FtzTT853d61u4I1xE6S51KA2+1AepeZXYtBh3G\nBAm/VdFgsHaD4hiby4as1DBpPQJxqcCxjJRhpvfiwTEcfHUWNNtasLZgAnpj1rsNmU89NGMqhe/x\nEbyHmRN65ubjI41YlhMBhgGCzT4AGCllnCVqyVtI9he+g3Tr8eO1M1GYGYONy5IwNsJHkrbZtCwV\n8eEWxaJWAODbTXDWek7VhvlZMDFaeJ9NGiMWx8zzOCY8SI+XH50uxXXWJa9GpNVtiDnl4lbOKFud\n7u4COK5v3S1GY4M1uhaKZAmGl1brR/pZMXPczaeXBwI1GHeImlMj2p2DLSKfkopyxwF6v2yA0rAs\niV2ADSmPYEZImlR7QszRH+sTi0luZdlYS/9zqncTUTdKlFUWn1deXW7FuEJp9AugzxWzalYlZZPN\nDElHSsAkt8hjh7Sa2cm7cNadFDBWtkKbY7lBEbMDgJhQoSO0WnSYkxqKB7LjoNV4SnHrtSrFd4oB\n94auRpmWmNCBe6n0sPMOdDm7oWbVUgfvQDe0Gk6hVdRbIkakjalX1qJ2OaQ4j2ggwtyDEPkUY7Ah\nEI8nrZRUAnoT5OsFL51yND0teCqMam+M840DwzAeK4XlnuDksYLhnheRg3mRWZg/Rgi021x2RAYZ\nERUidPrfnjMO6wsnSW2i0bmgZlUwaUxotbchJd4q1H2RxTbkXqgIq+1GTFq1lBoOACZvFcbHCt/j\nbxA8Ko4I3yMmU0QGGvG9xBVYFLgMq+6bDR+jFgumRSIiUGi7ZMcyeF9YiCBfL7Asg4mRymCxr7cR\n8yZ7LvIUi6SJzAnP8DhmyewIWC16RITo3OfoEC4aDFYZz2Hdiz6FMs3C70vFMSB83122JvQr7Fg3\nQ/qs1fasHwo0WqBTaW8azxsINxYRotwW/cUsfpC+CTaXHS8fex2AMptKzaqkeXipWJHbYIij3s+u\nnUZqwOAErwYbeQxHw6qlTlT+I51gTYCaVSEveh4qG8/ifK8ypwCgYtWSweAYVqon3mpvV3gYVS0X\noGI4D2M9WDz17RR02Zww6NT4Ts7YAZ/nrxeyz75ur8Wl60I2k9o9DSfGZq7bWxFkCJQMhqi51OXq\nCVwa1F6KHPpoc6RUX6Ld0aGQyhc1y/y9BA/oO+OXY1xQNCJ1Y275ueVMD56C6cFTpM8+WsELKIpf\nAgaMwoPJmBiMyEAjwgK8wTKxKHOvphZrS+j1AGzA3ORIcCyH02c10rN4qw0waY3gW3l0ODqx+Tsp\n2PX5WVx2O1ydzi6EegfjSi814a9RiRBrj0fjghPN7sSQaP9ATP3WePyr9Qucb70AG+lC8bIkRAYa\noWZVmJ84BX2xZpEyVrN4egxe6ZEeQ0yAFeGWQMCtWB9likSzrUURNxJ5dNL3sPPUO4izRONccxXO\nd55BCuLgbQDQIigla9zTUHERBrQ4HHAQb7R1OMB4CVO7S2ZG4WKZHo0QastzLAO5/23RmkEIQYej\nE1rZLXAckZIIArwFoxRqUAqG3g7UYHyDiFIeWeH39VtDAAAs7ikp0cNgGAYaTo2pQSn9nnO3iTZH\nSh2KXaHu2jNqFUeV88ZkwaI19Wkw1CwnqXWyDCtTDW2VgoI2lx1NthYEewcNSX1sANCqOWjVt17c\nx1fvCwaMpKAK9AgFymNPk6zjpfUtf/zir4qYBeAp4xBhDEOEMQwHa/6N5u4WmZClQ5LOFo2VmlVh\nSUIOrl0b3Bz8J5JXobajHskBnvPhDMNIo3QAUjXHvV/+FVMCk9HtErwq0auWDyTaHR0wu41Pq70N\nH1w9jMv2nhgFT3iYtSbJYCRZE/FZw+fSuSJO3olmt6Cgj9aMlAh/uOqm4fznFxBhDJNSyW8F+X3O\nH5ONtKAUWPV+aOpuRrfThrzoXOn5ezPROh6vZm5DafWHONdchX9dPoLJAUm40n4VWk4DjuXAEhYc\nw4FwDrAuHmqiAt+tAWtswv88NhV+Zh3iIr3ReFWs1kjAdxjBaLvAqJyIMIbBwTtwpukLhVy70ZtD\nu7t2jtifPJmy5pafvzfUYAwBNxNUWxqXd8P9ooch1hlgh8HMIcdyyI6cjb1fKIUV+3ODk/0n4lhd\nBc40KUuiBhoCEGWKxNnmLxHo5S9Ne8hLXIoSKfJU2nsFNatCiHeQx2gYgJTCCQi1QkRj6uSd+MMX\n7yqOtbnsYMBIMSwtp5WmPeTV29rs7ah3y6sHeCljLINNkCFQGvTcjLG+cbBozWixXcf2Y6+jobtJ\n4VH3HpGbNEKndt3Wio9rPdPP5ZlzsZYoWHQWHKr5N2plhZPsLofkYfi4g/STA5Lga/GGH26gpXYD\nQgxBWBh1P8b6xCHGMkbanhM5Z0Dn916T9crxnwPomc5jGAZmrQkttuvgCQ8dp0VSaDhOtzXBxrYB\nMEqGQPwtJIWNwXV7C6o7LiPMGIKGLiGlV66N5WfRYoyPBcebAbO7P+mrhOytMuQ90Ycffoh58+Yh\nNzcXO3fu9Nhvt9tRXFyMnJwcLF++HF9/7SkmNty4U41/i9YMHafD541CVkR/dRruNWaHzsCCMdko\nilsibZN7GHJ0Ki2eSF7tsT3ZfwJWTXgQK8YuRUboNMnD6KsmcvAAO69vmm+P/Vaf2+WdXoDeCj+9\nLzZP3dDvdWaGpkt/q1gOZnen2tDZk/Pfam/DsboKsAw7JOnWt4tepcP65O8DABrcAx957EXbayBh\n1vZ4GH1fr6fttCot/NwGQV6LxME7PBa3MgyD9LAUKZHkVmEYBgui7lcYi1ul3eGpNCz3IEXD2uHo\nhIpVYVyQkHFZ11mP67ZWnGxQrlL3NRjh6yU8X5h3iCRGKn9HCFzSlJTlNp+9L4bUw+B5Hlu3bsWu\nXbsQEBCAwsJCzJ07FzExPRlFf/rTn2A2m/H+++/jwIED2L59O37yk58M5W0NOao+Aty3AsuwCDT4\nS4v36js9iwjdizAMg4XROYptfB8yF3JK0orh5J041XAGHY5OWN3TKmJnKc6Tn7h2yuPcCX7jPLbd\nC0SbI/Fa5kt45vAPFenWetmUlDg1E24MQV50Lg5f+VgKXmtYNUxaE+aEZUDHafHP6kNI9Bsnqab+\nq+aIx3fGW2I8FozebQINAZgbPgttjnYcrf1UsU9uMNSsWurQ93/1jz6vZVB7warzRUN3E8waEwxq\nYVBWJyvSVN1Wg5r2r8Ey7G0biKFgbsQs1HXW41TDGagYzmMGQjRuLuJCp7NLSlr49ef/1+f1vFR6\n+GgtONd8HtHmSKl/+MMX+6Rjvmy5AAAesaY7ZUgNxsmTJxEZGYnQUCHtcOHChSgtLVUYjNLSUqxf\nvx4AkJubixdeeGEob+kbId4nBvMis5B8BwFqcdQAAMvi8wfjtu4KvUuQ9kacVpJ3rHLE/P86mdFc\nHp+PWWEz+jz+XoFjObx83w8VUxJsP/XQ542Zi/sjMrH+4HMAgB/P3goGDBiGwZKY+cgKvw9mrUmK\ne8lH4WaNCQ7eccfpxEPFt+IWgSc8jtZ+ijhZhp98hP1f05+FQe2FsT6xONdc5XENHadDRkg67gud\nhsrGL5DoN06q4d3bePKExxhThMdU0N3EpDHi0Un/geu2Vmg5DT68UqaQ6rDIpH4SfOMRY4mCXqVH\nl7NvzayxPnGItUQhK/w+cCwneeF9VQQ0qL0GdSAxpAajrq4OwcE988yBgYE4dUo5Uqyvr0dQkNB4\nHMfBZDKhpaUFFsu9417fKizDIq+PPOxboTBuMbScBkvjFkvu+nAkziIsMhvonG9veqeBzh+Tfc8b\nC5H+XlT59Ir8WLPGBC2nURoZ2WjZr9fiudzIrD7z/e81WIbFj2dthUrWHvLqdOLzPZG8GicbKlHb\nUY+/XXhP2r9ywoOSqsKMEGEhYn/TkfMis5DZR0rrvYD4nDd6Fx4cVwiGYfB40kqcvPY5ciIz8fTh\nHwIQ4ikTrAmI8xEMr5i+31+qNADMCp0+SHcvMKQGo3eBkYEcQwgZtFz64Yy/lx9WTnjwbt/GHWPV\n++GnmdskYcPbYWlcHv785d+wZuJ3pfUow5G04Mm40lGL2f28xFtnPHfD3z7DMJgckITj9Z9hScx8\nzAm7NzvGvui9TiHRbyzeu1gqZRkBgmFJ9p8A+ANTA1PgIi64iKtP48AwDMb5xOFss5BNtSw+H+lB\nqQodseGCWH9mTliG9P8fbY6UtNz+O+N5qFiVlHnWmyiTclGeWWPCjJCpmBqYgkDD7QX7+4MhA+nV\nb5OKigq8/vrr+NWvfgUAUtB7zZqe9K7Vq1dj3bp1SEpKgsvlQkZGBsrK+lGjpFAoFMpdY0gn+iZO\nnIjq6mpcuXIFdrsd+/fvx9y5cxXHzJkzB/v2CcGa9957D9Omeco6UygUCuXuM6QeBiCk1f7oRz8C\nIQSFhYVYs2YNXnvtNUycOBFz5syB3W7H008/jTNnzsBisWDHjh0ICwu7+YUpFAqF8o0y5AaDQqFQ\nKCODeyf3jEKhUCj3NNRgUCgUCmVAUINBoVAolAEx7AzGzbSpRholJSWYMWMG8vJ6BAuvX7+OlStX\nIjc3F6tWrUJbW8/K3xdffBE5OTlYsmQJzpw5czdueUiora3Fd7/7XSxYsAB5eXl45513AIzOtrDb\n7SgqKkJ+fj7y8vLws5/9DABQU1ODZcuWITc3Fxs3boTT6ZSOH2l6bb3heR4FBQV49NFHAYzetsjK\nysLixYuRn5+PwsJCAIP8jpBhhMvlItnZ2aSmpobY7XayePFiUlVVdbdva0j55JNPSGVlJVm0aJG0\n7eWXXyY7d+4khBDyi1/8gmzfvp0QQsjBgwfJ97//fUIIIRUVFaSoqOibv+Ehor6+nlRWVhJCCGlv\nbyc5OTmkqqpqVLYFIYR0dnYSQghxOp2kqKiIVFRUkCeffJIcOHCAEELI888/T37/+98TQgjZvXs3\n2bJlCyGEkP3795MNGzbclXseSn7zm9+QTZs2kUceeYQQQkZtW2RlZZGWlhbFtsF8R4aVhyHXT7us\nDgAACDZJREFUplKr1ZI21UhmypQpMJmUQmqlpaUoKCgAABQUFEhtUFpaivx8QXcqKSkJbW1taGho\n+GZveIjw9/dHQkICAMBgMCAmJgZ1dXWjsi0AQK8X5EXsdjucTicYhkF5eTlyc4WV0wUFBfjnP/8J\nQPl7yc3NHXELY2tra3Ho0CEUFRVJ2z7++ONR2RaEEPC8ssTxYL4jw8pg9KVNVV9ff4MzRiZNTU2w\nWoXaB/7+/mhqEkTp5LpcgNA+dXV1fV5jOFNTU4OzZ88iKSkJjY2No7IteJ5Hfn4+Zs6ciZkzZyI8\nPBwmkwksK7zSQUFB0vP2p9c2Uti2bRueeeYZSVajubkZZrN5VLYFwzBYtWoVli5dir179wLAoL4j\nw6qAEqFLRm5IX+0z0nS5Ojo6sH79epSUlMBgMPT7fCO9LViWxbvvvov29nasXbsW58+f9zhGfN7e\nbUFGkF7bwYMHYbVakZCQgPLycgDC8/V+5tHQFgCwZ88eySisXLkSUVFRg/qODCuDERQUpAhS1dXV\nISBgcMW1hgN+fn5oaGiA1WrFtWvX4OvrC0AYIdTW1krH1dbWjqj2cTqdWL9+PZYsWYLs7GwAo7ct\nRLy9vTF16lR89tlnaG1tBc/zYFlW8bxiWwQGBsLlcqG9vR1ms/kmVx4efPrpp/jggw9w6NAh2Gw2\ndHR0YNu2bWhraxt1bQEIHgQA+Pr6Ijs7GydPnhzUd2RYTUkNRJtqJNJ7JJCVlYW//OUvAIB9+/ZJ\nbTB37ly8+65Q6rOiogImk0lyRUcCJSUliI2NxcMPPyxtG41t0dTUJGW6dHd3o6ysDLGxsUhPT8d7\n7wmy4PK2yMrKGrF6bRs3bsTBgwdRWlqKHTt2ID09Ha+88sqobIuuri50dAjV/To7O3HkyBHEx8cP\n6jsy7KRB+tKmGsls2rQJ5eXlaGlpgdVqxbp165CdnY0nn3wSV69eRUhICF599VUpMP7CCy/g8OHD\n0Ov1eOmll5CYOHzlwOUcP34cDz30EOLj48EwQnGh4uJiTJo0CRs2bBhVbXHu3Dls3rwZPM+D53ks\nWLAAjz32GC5fvoyNGzeitbUVCQkJ2L59O9Rq9ajRazt69Ch+/etf48033xyVbXH58mU88cQTYBgG\nLpcLeXl5WLNmDVpaWgbtHRl2BoNCoVAod4dhNSVFoVAolLsHNRgUCoVCGRDUYFAoFAplQFCDQaFQ\nKJQBQQ0GhUKhUAYENRgUCoVCGRDUYFCGNcuWLUNBQQEWLlyIxMREFBQUoKCgACUlJbd8rdWrVw9I\n7vq5555DRUXF7dzuLVFZWYm///3vQ/49FMpAoeswKCOCK1euoLCw8Ibqo6JUxHBh7969KCsrw44d\nO+72rVAoAIaZlhSFciuUlZVh+/btSE5ORmVlJdauXYumpibs3r1bKqizefNmpKWlAQBmz56NXbt2\nISoqCitWrEBKSgpOnDiB+vp6LFq0CBs2bAAArFixAo8//jgyMjLw9NNPw9vbG+fPn0ddXR1SU1Px\n0ksvARC0eZ555hk0NzcjPDwcLpcLWVlZWL58ueI+GxoasGnTJjQ3NwMAMjIysHr1arzxxhvo7OxE\nQUEB0tPTsXnzZpw4cQI7duxAV1cXAGD9+vWYNWsWqqursWLFCixatAjHjx+H3W7Hli1bkJqa+o20\nNWWUcCfFOiiUe4Wamhoybdo0xbaPPvqIjB8/npw6dUraJi8uU1VVRTIzM6XPs2bNIhcuXCCEEPLA\nAw+QTZs2EUIIaW1tJWlpaaSmpkbad/jwYUIIIU899RR56KGHiMPhIDabjcybN4+Ul5cTQgh57LHH\nyC9/+UtCCCGXL18mKSkpZM+ePR73/tZbb5Hnn39e+tza2koIIeSPf/wj2bhxo+Le8/PzSWNjIyGE\nkNraWjJr1izS3t5OLl26RMaOHUv2798vPXtmZiZxOp0Db0QK5SZQD4MyoomOjsaECROkzxcvXsRr\nr72G+vp6cByH+vp6tLS0wGKxeJw7f/58AIDRaERUVBSqq6sRGhrqcdz9998PlUp4lcaPH4/q6mqk\npaWhvLwcL774IgAgLCxM8mR6k5ycjN/97nd45ZVXMHXqVGRkZPR53PHjx1FTU4NVq1ZJgpQcx+Hy\n5cvw8vKCXq/HggULAADTp08Hx3G4ePEiYmJiBtpcFMoNoQaDMqIxGAyKz8XFxdiyZQtmz54Nnucx\nadIk2Gy2Ps/VarXS3yzLwuVy3dJxA62zMHnyZOzbtw8fffQR/vznP+Ott97Cb3/7W4/jCCFITEzE\nrl27PPZVV1d7bON5fkTVeqDcfYZPBJBCuQlkAPkb7e3tkjrpnj17+jUCg0FaWpokK33lyhUcPXq0\nz+Nqamrg7e2NBQsWYPPmzTh9+jQAodaFKGMOAKmpqaiqqsKxY8ekbSdPnpT+7urqwoEDBwAIJUoB\nIDIycnAfijKqoR4GZcQwkNF0SUkJ1qxZg+DgYKSnp8NoNPZ5fu9r9bfvRsf94Ac/wLPPPov9+/cj\nOjoaqampiu8TKSsrwzvvvAOO40AIwdatWwEAM2fOxNtvv438/HxMmzYNmzdvxhtvvIHt27ejra0N\nDocD4eHhePPNNwEAVqsVX375JYqKimC327Fjxw5wHHfTNqFQBgpNq6VQhgibzQa1Wg2WZVFXV4ei\noiLs3r0b4eHhg/5dYpbUkSNHBv3aFIoI9TAolCHiwoULeO6550AIAc/zKC4uHhJjQaF8U1APg0Kh\nUCgDgga9KRQKhTIgqMGgUCgUyoCgBoNCoVAoA4IaDAqFQqEMCGowKBQKhTIgqMGgUCgUyoD4f001\n1ZxdsABYAAAAAElFTkSuQmCC\n", + "text/plain": [ + "\u003cmatplotlib.figure.Figure at 0x7f96f1241810\u003e" + ] + }, + "metadata": { + "tags": [] + }, + "output_type": "display_data" + }, + { + "name": "stdout", + "output_type": "stream", + "text": [ + "test_accuracy tf.Tensor(0.99, shape=(), dtype=float32)\n" + ] + }, + { + "data": { + "image/png": "iVBORw0KGgoAAAANSUhEUgAAAYwAAAEcCAYAAADUX4MJAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nAAAASUVORK5CYII=\n", + "text/plain": [ + "\u003cmatplotlib.figure.Figure at 0x7f96f7389490\u003e" + ] + }, + "metadata": { + "tags": [] + }, + "output_type": "display_data" + } + ], + "source": [ + "#@test {\"timeout\": 90} \n", + "with context.eager_mode():\n", + " durations = []\n", + " for t in range(burn_ins + trials):\n", + " hp = tf.contrib.training.HParams(\n", + " learning_rate=0.05,\n", + " max_steps=max_steps,\n", + " )\n", + " train_ds = setup_mnist_data(True, hp, 500)\n", + " test_ds = setup_mnist_data(False, hp, 100)\n", + " ds = tf.data.Dataset.zip((train_ds, test_ds))\n", + " start = time.time()\n", + " (train_losses, test_losses, train_accuracies,\n", + " test_accuracies) = train(ds, hp)\n", + " if t \u003c burn_ins:\n", + " continue\n", + " train_losses[-1].numpy()\n", + " test_losses[-1].numpy()\n", + " train_accuracies[-1].numpy()\n", + " test_accuracies[-1].numpy()\n", + " duration = time.time() - start\n", + " durations.append(duration)\n", + " print('Duration:', duration)\n", + "\n", + "\n", + " print('Mean duration:', np.mean(durations), '+/-', np.std(durations))\n", + " plt.title('MNIST train/test losses')\n", + " plt.plot(train_losses, label='train loss')\n", + " plt.plot(test_losses, label='test loss')\n", + " plt.legend()\n", + " plt.xlabel('Training step')\n", + " plt.ylabel('Loss')\n", + " plt.show()\n", + " plt.title('MNIST train/test accuracies')\n", + " plt.plot(train_accuracies, label='train accuracy')\n", + " plt.plot(test_accuracies, label='test accuracy')\n", + " print('test_accuracy', test_accuracies[-1])\n", + " plt.legend(loc='lower right')\n", + " plt.xlabel('Training step')\n", + " plt.ylabel('Accuracy')\n", + " plt.show()\n" + ] + } + ], + "metadata": { + "colab": { + "collapsed_sections": [], + "default_view": {}, + "last_runtime": { + "build_target": "", + "kind": "local" + }, + "name": "Autograph vs. Eager MNIST benchmark", + "provenance": [ + { + "file_id": "1tAQW5tHUgAc8M4-iwwJm6Xs6dV9nEqtD", + "timestamp": 1530297010607 + }, + { + "file_id": "18dCjshrmHiPTIe1CNsL8tnpdGkuXgpM9", + "timestamp": 1530289467317 + }, + { + "file_id": "1DcfimonWU11tmyivKBGVrbpAl3BIOaRG", + "timestamp": 1522272821237 + }, + { + "file_id": "1wCZUh73zTNs1jzzYjqoxMIdaBWCdKJ2K", + "timestamp": 1522238054357 + }, + { + "file_id": "1_HpC-RrmIv4lNaqeoslUeWaX8zH5IXaJ", + "timestamp": 1521743157199 + }, + { + "file_id": "1mjO2fQ2F9hxpAzw2mnrrUkcgfb7xSGW-", + "timestamp": 1520522344607 + } + ], + "version": "0.3.2", + "views": {} + } + }, + "nbformat": 4, + "nbformat_minor": 0 +} diff --git a/tensorflow/contrib/autograph/examples/notebooks/rnn_keras_estimator.ipynb b/tensorflow/contrib/autograph/examples/notebooks/rnn_keras_estimator.ipynb index 324b23c24b5a7970d7f20ed955839ba1cf1774fc..44532cb078f9bd1578172f8a7d8a4b55cd21a7cb 100644 --- a/tensorflow/contrib/autograph/examples/notebooks/rnn_keras_estimator.ipynb +++ b/tensorflow/contrib/autograph/examples/notebooks/rnn_keras_estimator.ipynb @@ -190,7 +190,6 @@ " self.upper_cell = tf.contrib.rnn.LSTMBlockCell(128)\n", " self.relu_layer = tf.layers.Dense(3, activation=tf.nn.relu)\n", "\n", - "\n", " def _rnn_layer(self, chars, cell, batch_size, training):\n", " \"\"\"A single RNN layer.\n", "\n", @@ -203,13 +202,12 @@ " Returns:\n", " A Tensor of shape (max_sequence_length, batch_size, output_size).\n", " \"\"\"\n", - " hidden_outputs = []\n", - " autograph.utils.set_element_type(hidden_outputs, tf.float32)\n", + " hidden_outputs = tf.TensorArray(tf.float32, 0, True)\n", " state, output = cell.zero_state(batch_size, tf.float32)\n", " for ch in chars:\n", " cell_output, (state, output) = cell.call(ch, (state, output))\n", " hidden_outputs.append(cell_output)\n", - " hidden_outputs = hidden_outputs.stack()\n", + " hidden_outputs = autograph.stack(hidden_outputs)\n", " if training:\n", " hidden_outputs = tf.nn.dropout(hidden_outputs, 0.5)\n", " return hidden_outputs\n", @@ -223,7 +221,7 @@ "\n", "\n", " def call(self, inputs, training=False):\n", - " \"\"\"The RNN model code. Uses Eager and \n", + " \"\"\"The RNN model code. Uses Eager.\n", "\n", " The model consists of two RNN layers (made by lower_cell and upper_cell),\n", " followed by a fully connected layer with ReLU activation.\n", @@ -243,7 +241,8 @@ " seq = self._rnn_layer(seq, self.upper_cell, batch_size, training)\n", "\n", " # Grab just the end-of-sequence from each output.\n", - " indices = tf.stack([length - 1, range(batch_size)], axis=1)\n", + " indices = (length - 1, range(batch_size))\n", + " indices = tf.stack(indices, 1)\n", " sequence_ends = tf.gather_nd(seq, indices)\n", " return self.relu_layer(sequence_ends)\n", "\n", @@ -381,7 +380,7 @@ }, { "cell_type": "code", - "execution_count": 7, + "execution_count": 107, "metadata": { "colab": { "autoexec": { @@ -392,9 +391,9 @@ }, "colab_type": "code", "executionInfo": { - "elapsed": 10604, + "elapsed": 5454, "status": "ok", - "timestamp": 1524095272039, + "timestamp": 1529952160455, "user": { "displayName": "", "photoUrl": "", @@ -403,7 +402,7 @@ "user_tz": 240 }, "id": "2pg1AfbxBJQq", - "outputId": "9c924b4f-06e1-4538-976c-a3e1ddac5660", + "outputId": "4aef3052-f7c7-4bb1-a0a2-73fef2e96efb", "slideshow": { "slide_type": "-" } @@ -413,7 +412,7 @@ "name": "stdout", "output_type": "stream", "text": [ - "Eval loss at step 100: 0.0674834\n" + "Eval loss at step 100: 0.0705221\n" ] } ], @@ -423,8 +422,8 @@ " 'learning_rate': 0.01,\n", "}\n", "\n", - "train_url = \"https://raw.githubusercontent.com/random-forests/tensorflow-workshop/master/extras/colorbot/data/train.csv\"\n", - "test_url = \"https://raw.githubusercontent.com/random-forests/tensorflow-workshop/master/extras/colorbot/data/test.csv\"\n", + "train_url = \"https://raw.githubusercontent.com/random-forests/tensorflow-workshop/master/archive/extras/colorbot/data/train.csv\"\n", + "test_url = \"https://raw.githubusercontent.com/random-forests/tensorflow-workshop/master/archive/extras/colorbot/data/test.csv\"\n", "data_dir = \"tmp/rnn/data\"\n", "\n", "regressor = tf.estimator.Estimator(\n", @@ -457,7 +456,7 @@ }, { "cell_type": "code", - "execution_count": 8, + "execution_count": 108, "metadata": { "colab": { "autoexec": { @@ -468,9 +467,9 @@ }, "colab_type": "code", "executionInfo": { - "elapsed": 7990, + "elapsed": 3432, "status": "ok", - "timestamp": 1524095280105, + "timestamp": 1529952163923, "user": { "displayName": "", "photoUrl": "", @@ -479,7 +478,7 @@ "user_tz": 240 }, "id": "dxHex2tUN_10", - "outputId": "2b889e5a-b9ed-4645-bf03-d98f26c72101", + "outputId": "1ff438f2-b045-4f4e-86a0-4dae7503f6b2", "slideshow": { "slide_type": "slide" } @@ -491,12 +490,12 @@ "\u003clink rel=stylesheet type=text/css href='/nbextensions/google.colab/tabbar.css'\u003e\u003c/link\u003e" ], "text/plain": [ - "\u003cIPython.core.display.HTML at 0x7f3f36aa6cd0\u003e" + "\u003cIPython.core.display.HTML at 0x7fcd7222a110\u003e" ] }, "metadata": { "tags": [ - "outputarea_id1" + "outputarea_id3" ] }, "output_type": "display_data" @@ -507,12 +506,12 @@ "\u003cscript src='/nbextensions/google.colab/tabbar_main.min.js'\u003e\u003c/script\u003e" ], "text/plain": [ - "\u003cIPython.core.display.HTML at 0x7f3eca67f7d0\u003e" + "\u003cIPython.core.display.HTML at 0x7fcd7222a8d0\u003e" ] }, "metadata": { "tags": [ - "outputarea_id1" + "outputarea_id3" ] }, "output_type": "display_data" @@ -520,15 +519,15 @@ { "data": { "text/html": [ - "\u003cdiv id=\"id1\"\u003e\u003c/div\u003e" + "\u003cdiv id=\"id3\"\u003e\u003c/div\u003e" ], "text/plain": [ - "\u003cIPython.core.display.HTML at 0x7f3eca67f8d0\u003e" + "\u003cIPython.core.display.HTML at 0x7fcd7222a050\u003e" ] }, "metadata": { "tags": [ - "outputarea_id1" + "outputarea_id3" ] }, "output_type": "display_data" @@ -536,16 +535,16 @@ { "data": { "application/javascript": [ - "window[\"e8ddfa22-4362-11e8-91ec-c8d3ffb5fbe0\"] = colab_lib.createTabBar({\"contentBorder\": [\"0px\"], \"elementId\": \"id1\", \"borderColor\": [\"#a7a7a7\"], \"contentHeight\": [\"initial\"], \"tabNames\": [\"RNN Colorbot\"], \"location\": \"top\", \"initialSelection\": 0});\n", - "//# sourceURL=js_71b9087b6d" + "window[\"8a03307e-78a7-11e8-99f9-c8d3ffb5fbe0\"] = colab_lib.createTabBar({\"contentBorder\": [\"0px\"], \"elementId\": \"id3\", \"contentHeight\": [\"initial\"], \"tabNames\": [\"RNN Colorbot\"], \"location\": \"top\", \"initialSelection\": 0, \"borderColor\": [\"#a7a7a7\"]});\n", + "//# sourceURL=js_dc5d7f2784" ], "text/plain": [ - "\u003cIPython.core.display.Javascript at 0x7f3eca67f950\u003e" + "\u003cIPython.core.display.Javascript at 0x7fcd7222a190\u003e" ] }, "metadata": { "tags": [ - "outputarea_id1" + "outputarea_id3" ] }, "output_type": "display_data" @@ -553,16 +552,16 @@ { "data": { "application/javascript": [ - "window[\"e8ddfa23-4362-11e8-91ec-c8d3ffb5fbe0\"] = window[\"id1\"].setSelectedTabIndex(0);\n", - "//# sourceURL=js_e390445f33" + "window[\"8a03307f-78a7-11e8-99f9-c8d3ffb5fbe0\"] = window[\"id3\"].setSelectedTabIndex(0);\n", + "//# sourceURL=js_be7950150b" ], "text/plain": [ - "\u003cIPython.core.display.Javascript at 0x7f3eca67f990\u003e" + "\u003cIPython.core.display.Javascript at 0x7fcd7222ac90\u003e" ] }, "metadata": { "tags": [ - "outputarea_id1" + "outputarea_id3" ] }, "output_type": "display_data" @@ -570,17 +569,17 @@ { "data": { "application/javascript": [ - "window[\"e8ddfa24-4362-11e8-91ec-c8d3ffb5fbe0\"] = google.colab.output.getActiveOutputArea();\n", - "//# sourceURL=js_241dd76d85" + "window[\"8a033080-78a7-11e8-99f9-c8d3ffb5fbe0\"] = google.colab.output.getActiveOutputArea();\n", + "//# sourceURL=js_d0c3bd4eaa" ], "text/plain": [ - "\u003cIPython.core.display.Javascript at 0x7f3eca67fc50\u003e" + "\u003cIPython.core.display.Javascript at 0x7fcd7222aad0\u003e" ] }, "metadata": { "tags": [ - "id1_content_0", - "outputarea_id1" + "id3_content_0", + "outputarea_id3" ] }, "output_type": "display_data" @@ -588,17 +587,17 @@ { "data": { "application/javascript": [ - "window[\"e8ddfa25-4362-11e8-91ec-c8d3ffb5fbe0\"] = document.querySelector(\"#id1_content_0\");\n", - "//# sourceURL=js_60c64e3d50" + "window[\"8a033081-78a7-11e8-99f9-c8d3ffb5fbe0\"] = document.querySelector(\"#id3_content_0\");\n", + "//# sourceURL=js_f10f6eba86" ], "text/plain": [ - "\u003cIPython.core.display.Javascript at 0x7f3eca67fd90\u003e" + "\u003cIPython.core.display.Javascript at 0x7fcd7222aed0\u003e" ] }, "metadata": { "tags": [ - "id1_content_0", - "outputarea_id1" + "id3_content_0", + "outputarea_id3" ] }, "output_type": "display_data" @@ -606,17 +605,17 @@ { "data": { "application/javascript": [ - "window[\"e8ddfa26-4362-11e8-91ec-c8d3ffb5fbe0\"] = google.colab.output.setActiveOutputArea(window[\"e8ddfa25-4362-11e8-91ec-c8d3ffb5fbe0\"]);\n", - "//# sourceURL=js_14ea437cbd" + "window[\"8a033082-78a7-11e8-99f9-c8d3ffb5fbe0\"] = google.colab.output.setActiveOutputArea(window[\"8a033081-78a7-11e8-99f9-c8d3ffb5fbe0\"]);\n", + "//# sourceURL=js_ff29697179" ], "text/plain": [ - "\u003cIPython.core.display.Javascript at 0x7f3eca67fe10\u003e" + "\u003cIPython.core.display.Javascript at 0x7fcd7222abd0\u003e" ] }, "metadata": { "tags": [ - "id1_content_0", - "outputarea_id1" + "id3_content_0", + "outputarea_id3" ] }, "output_type": "display_data" @@ -624,17 +623,17 @@ { "data": { "application/javascript": [ - "window[\"e8ddfa27-4362-11e8-91ec-c8d3ffb5fbe0\"] = window[\"id1\"].setSelectedTabIndex(0);\n", - "//# sourceURL=js_09294c2226" + "window[\"8a033083-78a7-11e8-99f9-c8d3ffb5fbe0\"] = window[\"id3\"].setSelectedTabIndex(0);\n", + "//# sourceURL=js_ff85295dc7" ], "text/plain": [ - "\u003cIPython.core.display.Javascript at 0x7f3eca67fcd0\u003e" + "\u003cIPython.core.display.Javascript at 0x7fcd7222ab90\u003e" ] }, "metadata": { "tags": [ - "id1_content_0", - "outputarea_id1" + "id3_content_0", + "outputarea_id3" ] }, "output_type": "display_data" @@ -642,17 +641,17 @@ { "data": { "application/javascript": [ - "window[\"ec965514-4362-11e8-91ec-c8d3ffb5fbe0\"] = google.colab.output.setActiveOutputArea(window[\"e8ddfa24-4362-11e8-91ec-c8d3ffb5fbe0\"]);\n", - "//# sourceURL=js_e5e8266997" + "window[\"8b18d8dc-78a7-11e8-99f9-c8d3ffb5fbe0\"] = google.colab.output.setActiveOutputArea(window[\"8a033080-78a7-11e8-99f9-c8d3ffb5fbe0\"]);\n", + "//# sourceURL=js_ed7aabfedb" ], "text/plain": [ - "\u003cIPython.core.display.Javascript at 0x7f3eca67fe10\u003e" + "\u003cIPython.core.display.Javascript at 0x7fcd7222a110\u003e" ] }, "metadata": { "tags": [ - "id1_content_0", - "outputarea_id1" + "id3_content_0", + "outputarea_id3" ] }, "output_type": "display_data" @@ -660,17 +659,17 @@ { "data": { "application/javascript": [ - "window[\"ec965515-4362-11e8-91ec-c8d3ffb5fbe0\"] = google.colab.output.getActiveOutputArea();\n", - "//# sourceURL=js_07a097f0ee" + "window[\"8b18d8dd-78a7-11e8-99f9-c8d3ffb5fbe0\"] = google.colab.output.getActiveOutputArea();\n", + "//# sourceURL=js_c86f8feaf4" ], "text/plain": [ - "\u003cIPython.core.display.Javascript at 0x7f3eca67fc90\u003e" + "\u003cIPython.core.display.Javascript at 0x7fcd7222acd0\u003e" ] }, "metadata": { "tags": [ - "id1_content_0", - "outputarea_id1" + "id3_content_0", + "outputarea_id3" ] }, "output_type": "display_data" @@ -678,17 +677,17 @@ { "data": { "application/javascript": [ - "window[\"ec965516-4362-11e8-91ec-c8d3ffb5fbe0\"] = document.querySelector(\"#id1_content_0\");\n", - "//# sourceURL=js_790d669ca8" + "window[\"8b18d8de-78a7-11e8-99f9-c8d3ffb5fbe0\"] = document.querySelector(\"#id3_content_0\");\n", + "//# sourceURL=js_4d0fde6662" ], "text/plain": [ - "\u003cIPython.core.display.Javascript at 0x7f3eca67f8d0\u003e" + "\u003cIPython.core.display.Javascript at 0x7fcd7222ae50\u003e" ] }, "metadata": { "tags": [ - "id1_content_0", - "outputarea_id1" + "id3_content_0", + "outputarea_id3" ] }, "output_type": "display_data" @@ -696,17 +695,17 @@ { "data": { "application/javascript": [ - "window[\"ec965517-4362-11e8-91ec-c8d3ffb5fbe0\"] = google.colab.output.setActiveOutputArea(window[\"ec965516-4362-11e8-91ec-c8d3ffb5fbe0\"]);\n", - "//# sourceURL=js_d30df771f0" + "window[\"8b18d8df-78a7-11e8-99f9-c8d3ffb5fbe0\"] = google.colab.output.setActiveOutputArea(window[\"8b18d8de-78a7-11e8-99f9-c8d3ffb5fbe0\"]);\n", + "//# sourceURL=js_3f66d52720" ], "text/plain": [ - "\u003cIPython.core.display.Javascript at 0x7f3eca67fd90\u003e" + "\u003cIPython.core.display.Javascript at 0x7fcd7222a210\u003e" ] }, "metadata": { "tags": [ - "id1_content_0", - "outputarea_id1" + "id3_content_0", + "outputarea_id3" ] }, "output_type": "display_data" @@ -714,32 +713,32 @@ { "data": { "application/javascript": [ - "window[\"ec965518-4362-11e8-91ec-c8d3ffb5fbe0\"] = window[\"id1\"].setSelectedTabIndex(0);\n", - "//# sourceURL=js_8a43a2da4b" + "window[\"8b18d8e0-78a7-11e8-99f9-c8d3ffb5fbe0\"] = window[\"id3\"].setSelectedTabIndex(0);\n", + "//# sourceURL=js_375f5ae6d7" ], "text/plain": [ - "\u003cIPython.core.display.Javascript at 0x7f3eca67fc50\u003e" + "\u003cIPython.core.display.Javascript at 0x7fcd7222a310\u003e" ] }, "metadata": { "tags": [ - "id1_content_0", - "outputarea_id1" + "id3_content_0", + "outputarea_id3" ] }, "output_type": "display_data" }, { "data": { - "image/png": "iVBORw0KGgoAAAANSUhEUgAAAQwAAAENCAYAAAD60Fs2AAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\n", + "image/png": "iVBORw0KGgoAAAANSUhEUgAAAQwAAAENCAYAAAD60Fs2AAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\n", "text/plain": [ - "\u003cmatplotlib.figure.Figure at 0x7f3ecc00bf10\u003e" + "\u003cmatplotlib.figure.Figure at 0x7fcd0d02dc90\u003e" ] }, "metadata": { "tags": [ - "id1_content_0", - "outputarea_id1", + "id3_content_0", + "outputarea_id3", "user_output" ] }, @@ -748,17 +747,17 @@ { "data": { "application/javascript": [ - "window[\"ec965519-4362-11e8-91ec-c8d3ffb5fbe0\"] = google.colab.output.setActiveOutputArea(window[\"ec965515-4362-11e8-91ec-c8d3ffb5fbe0\"]);\n", - "//# sourceURL=js_893ad561f4" + "window[\"8b18d8e1-78a7-11e8-99f9-c8d3ffb5fbe0\"] = google.colab.output.setActiveOutputArea(window[\"8b18d8dd-78a7-11e8-99f9-c8d3ffb5fbe0\"]);\n", + "//# sourceURL=js_34b0509660" ], "text/plain": [ - "\u003cIPython.core.display.Javascript at 0x7f3f31b55c90\u003e" + "\u003cIPython.core.display.Javascript at 0x7fcd08e6e850\u003e" ] }, "metadata": { "tags": [ - "id1_content_0", - "outputarea_id1" + "id3_content_0", + "outputarea_id3" ] }, "output_type": "display_data" @@ -766,17 +765,17 @@ { "data": { "application/javascript": [ - "window[\"ec96551a-4362-11e8-91ec-c8d3ffb5fbe0\"] = google.colab.output.getActiveOutputArea();\n", - "//# sourceURL=js_2d99e0ac17" + "window[\"8b18d8e2-78a7-11e8-99f9-c8d3ffb5fbe0\"] = google.colab.output.getActiveOutputArea();\n", + "//# sourceURL=js_518a0f26fe" ], "text/plain": [ - "\u003cIPython.core.display.Javascript at 0x7f3eca67fe50\u003e" + "\u003cIPython.core.display.Javascript at 0x7fcd08e6ec90\u003e" ] }, "metadata": { "tags": [ - "id1_content_0", - "outputarea_id1" + "id3_content_0", + "outputarea_id3" ] }, "output_type": "display_data" @@ -784,17 +783,17 @@ { "data": { "application/javascript": [ - "window[\"ec96551b-4362-11e8-91ec-c8d3ffb5fbe0\"] = document.querySelector(\"#id1_content_0\");\n", - "//# sourceURL=js_5c19462e32" + "window[\"8b18d8e3-78a7-11e8-99f9-c8d3ffb5fbe0\"] = document.querySelector(\"#id3_content_0\");\n", + "//# sourceURL=js_17eb3ff612" ], "text/plain": [ - "\u003cIPython.core.display.Javascript at 0x7f3f31b55dd0\u003e" + "\u003cIPython.core.display.Javascript at 0x7fcd08e6eb50\u003e" ] }, "metadata": { "tags": [ - "id1_content_0", - "outputarea_id1" + "id3_content_0", + "outputarea_id3" ] }, "output_type": "display_data" @@ -802,17 +801,17 @@ { "data": { "application/javascript": [ - "window[\"ec96551c-4362-11e8-91ec-c8d3ffb5fbe0\"] = google.colab.output.setActiveOutputArea(window[\"ec96551b-4362-11e8-91ec-c8d3ffb5fbe0\"]);\n", - "//# sourceURL=js_b9c8b7567b" + "window[\"8b18d8e4-78a7-11e8-99f9-c8d3ffb5fbe0\"] = google.colab.output.setActiveOutputArea(window[\"8b18d8e3-78a7-11e8-99f9-c8d3ffb5fbe0\"]);\n", + "//# sourceURL=js_99da807c8e" ], "text/plain": [ - "\u003cIPython.core.display.Javascript at 0x7f3f31b55a50\u003e" + "\u003cIPython.core.display.Javascript at 0x7fcd08e6eb90\u003e" ] }, "metadata": { "tags": [ - "id1_content_0", - "outputarea_id1" + "id3_content_0", + "outputarea_id3" ] }, "output_type": "display_data" @@ -820,17 +819,17 @@ { "data": { "application/javascript": [ - "window[\"ec96551d-4362-11e8-91ec-c8d3ffb5fbe0\"] = window[\"id1\"].setSelectedTabIndex(0);\n", - "//# sourceURL=js_fd05186348" + "window[\"8b18d8e5-78a7-11e8-99f9-c8d3ffb5fbe0\"] = window[\"id3\"].setSelectedTabIndex(0);\n", + "//# sourceURL=js_dee01cb4b6" ], "text/plain": [ - "\u003cIPython.core.display.Javascript at 0x7f3f31b55810\u003e" + "\u003cIPython.core.display.Javascript at 0x7fcd08e6e610\u003e" ] }, "metadata": { "tags": [ - "id1_content_0", - "outputarea_id1" + "id3_content_0", + "outputarea_id3" ] }, "output_type": "display_data" @@ -838,16 +837,16 @@ { "data": { "text/html": [ - "\u003cdiv class=id_888646481 style=\"margin-right:10px; display:flex;align-items:center;\"\u003e\u003cspan style=\"margin-right: 3px;\"\u003e\u003c/span\u003e\u003c/div\u003e" + "\u003cdiv class=id_853612217 style=\"margin-right:10px; display:flex;align-items:center;\"\u003e\u003cspan style=\"margin-right: 3px;\"\u003e\u003c/span\u003e\u003c/div\u003e" ], "text/plain": [ - "\u003cIPython.core.display.HTML at 0x7f3f32414810\u003e" + "\u003cIPython.core.display.HTML at 0x7fcd7222aa10\u003e" ] }, "metadata": { "tags": [ - "id1_content_0", - "outputarea_id1", + "id3_content_0", + "outputarea_id3", "user_output" ] }, @@ -856,17 +855,17 @@ { "data": { "application/javascript": [ - "window[\"ec96551e-4362-11e8-91ec-c8d3ffb5fbe0\"] = jQuery(\".id_888646481 span\");\n", - "//# sourceURL=js_efef96e882" + "window[\"8b18d8e6-78a7-11e8-99f9-c8d3ffb5fbe0\"] = jQuery(\".id_853612217 span\");\n", + "//# sourceURL=js_8c378be329" ], "text/plain": [ - "\u003cIPython.core.display.Javascript at 0x7f3f31b55710\u003e" + "\u003cIPython.core.display.Javascript at 0x7fcd08e6e990\u003e" ] }, "metadata": { "tags": [ - "id1_content_0", - "outputarea_id1", + "id3_content_0", + "outputarea_id3", "user_output" ] }, @@ -875,17 +874,17 @@ { "data": { "application/javascript": [ - "window[\"ec96551f-4362-11e8-91ec-c8d3ffb5fbe0\"] = window[\"ec96551e-4362-11e8-91ec-c8d3ffb5fbe0\"].text(\"Give me a color name (or press 'enter' to exit): \");\n", - "//# sourceURL=js_6eca889864" + "window[\"8b18d8e7-78a7-11e8-99f9-c8d3ffb5fbe0\"] = window[\"8b18d8e6-78a7-11e8-99f9-c8d3ffb5fbe0\"].text(\"Give me a color name (or press 'enter' to exit): \");\n", + "//# sourceURL=js_f0b946600c" ], "text/plain": [ - "\u003cIPython.core.display.Javascript at 0x7f3eca67f990\u003e" + "\u003cIPython.core.display.Javascript at 0x7fcd08e6e310\u003e" ] }, "metadata": { "tags": [ - "id1_content_0", - "outputarea_id1", + "id3_content_0", + "outputarea_id3", "user_output" ] }, @@ -894,17 +893,17 @@ { "data": { "application/javascript": [ - "window[\"ed8ea972-4362-11e8-91ec-c8d3ffb5fbe0\"] = jQuery(\".id_888646481 input\");\n", - "//# sourceURL=js_f02070cc60" + "window[\"8b18d8e9-78a7-11e8-99f9-c8d3ffb5fbe0\"] = jQuery(\".id_853612217 input\");\n", + "//# sourceURL=js_9e21b1373a" ], "text/plain": [ - "\u003cIPython.core.display.Javascript at 0x7f3f31b553d0\u003e" + "\u003cIPython.core.display.Javascript at 0x7fcd08e6ea90\u003e" ] }, "metadata": { "tags": [ - "id1_content_0", - "outputarea_id1", + "id3_content_0", + "outputarea_id3", "user_output" ] }, @@ -913,17 +912,17 @@ { "data": { "application/javascript": [ - "window[\"ed8ea973-4362-11e8-91ec-c8d3ffb5fbe0\"] = window[\"ed8ea972-4362-11e8-91ec-c8d3ffb5fbe0\"].remove();\n", - "//# sourceURL=js_ed9faba660" + "window[\"8b18d8ea-78a7-11e8-99f9-c8d3ffb5fbe0\"] = window[\"8b18d8e9-78a7-11e8-99f9-c8d3ffb5fbe0\"].remove();\n", + "//# sourceURL=js_a7764968c6" ], "text/plain": [ - "\u003cIPython.core.display.Javascript at 0x7f3f31a95450\u003e" + "\u003cIPython.core.display.Javascript at 0x7fcd08e6e5d0\u003e" ] }, "metadata": { "tags": [ - "id1_content_0", - "outputarea_id1", + "id3_content_0", + "outputarea_id3", "user_output" ] }, @@ -932,17 +931,17 @@ { "data": { "application/javascript": [ - "window[\"ed8ea974-4362-11e8-91ec-c8d3ffb5fbe0\"] = jQuery(\".id_888646481 span\");\n", - "//# sourceURL=js_f3458d7074" + "window[\"8b18d8eb-78a7-11e8-99f9-c8d3ffb5fbe0\"] = jQuery(\".id_853612217 span\");\n", + "//# sourceURL=js_74279d3ff0" ], "text/plain": [ - "\u003cIPython.core.display.Javascript at 0x7f3f31a95250\u003e" + "\u003cIPython.core.display.Javascript at 0x7fcd08e6e890\u003e" ] }, "metadata": { "tags": [ - "id1_content_0", - "outputarea_id1", + "id3_content_0", + "outputarea_id3", "user_output" ] }, @@ -951,17 +950,17 @@ { "data": { "application/javascript": [ - "window[\"ed8ea975-4362-11e8-91ec-c8d3ffb5fbe0\"] = window[\"ed8ea974-4362-11e8-91ec-c8d3ffb5fbe0\"].text(\"Give me a color name (or press 'enter' to exit): \");\n", - "//# sourceURL=js_3ffd97bd6f" + "window[\"8b18d8ec-78a7-11e8-99f9-c8d3ffb5fbe0\"] = window[\"8b18d8eb-78a7-11e8-99f9-c8d3ffb5fbe0\"].text(\"Give me a color name (or press 'enter' to exit): \");\n", + "//# sourceURL=js_82b6c34cdb" ], "text/plain": [ - "\u003cIPython.core.display.Javascript at 0x7f3f31a953d0\u003e" + "\u003cIPython.core.display.Javascript at 0x7fcd08e6e8d0\u003e" ] }, "metadata": { "tags": [ - "id1_content_0", - "outputarea_id1", + "id3_content_0", + "outputarea_id3", "user_output" ] }, @@ -970,17 +969,17 @@ { "data": { "application/javascript": [ - "window[\"ed8ea976-4362-11e8-91ec-c8d3ffb5fbe0\"] = google.colab.output.setActiveOutputArea(window[\"ec96551a-4362-11e8-91ec-c8d3ffb5fbe0\"]);\n", - "//# sourceURL=js_7f73e8bcca" + "window[\"8b18d8ed-78a7-11e8-99f9-c8d3ffb5fbe0\"] = google.colab.output.setActiveOutputArea(window[\"8b18d8e2-78a7-11e8-99f9-c8d3ffb5fbe0\"]);\n", + "//# sourceURL=js_ff6144734a" ], "text/plain": [ - "\u003cIPython.core.display.Javascript at 0x7f3f31b55710\u003e" + "\u003cIPython.core.display.Javascript at 0x7fcd08e6e8d0\u003e" ] }, "metadata": { "tags": [ - "id1_content_0", - "outputarea_id1" + "id3_content_0", + "outputarea_id3" ] }, "output_type": "display_data" @@ -1043,28 +1042,6 @@ "kind": "local" }, "name": "RNN Colorbot using Keras and Estimators", - "provenance": [ - { - "file_id": "1CtzefX39ffFibX_BqE6cRbT0UW_DdVKl", - "timestamp": 1523579810961 - }, - { - "file_id": "1DcfimonWU11tmyivKBGVrbpAl3BIOaRG", - "timestamp": 1523016192637 - }, - { - "file_id": "1wCZUh73zTNs1jzzYjqoxMIdaBWCdKJ2K", - "timestamp": 1522238054357 - }, - { - "file_id": "1_HpC-RrmIv4lNaqeoslUeWaX8zH5IXaJ", - "timestamp": 1521743157199 - }, - { - "file_id": "1mjO2fQ2F9hxpAzw2mnrrUkcgfb7xSGW-", - "timestamp": 1520522344607 - } - ], "version": "0.3.2", "views": {} }, diff --git a/tensorflow/contrib/autograph/examples/notebooks/workshop.ipynb b/tensorflow/contrib/autograph/examples/notebooks/workshop.ipynb new file mode 100644 index 0000000000000000000000000000000000000000..e8f16b431df189b0da5630be69f546a9743468cc --- /dev/null +++ b/tensorflow/contrib/autograph/examples/notebooks/workshop.ipynb @@ -0,0 +1,1093 @@ +{ + "cells": [ + { + "cell_type": "code", + "execution_count": 0, + "metadata": { + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + }, + "colab_type": "code", + "id": "qWUV0FYjDSKj" + }, + "outputs": [], + "source": [ + "import tensorflow as tf\n", + "from tensorflow.contrib import autograph\n", + "\n", + "import matplotlib.pyplot as plt" + ] + }, + { + "cell_type": "markdown", + "metadata": { + "colab_type": "text", + "id": "kGXS3UWBBNoc" + }, + "source": [ + "# 1. AutoGraph writes graph code for you\n", + "\n", + "[AutoGraph](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/contrib/autograph/README.md) helps you write complicated graph code using just plain Python -- behind the scenes, AutoGraph automatically transforms your code into the equivalent TF graph code. We support a large chunk of the Python language, which is growing. [Please see this document for what we currently support, and what we're working on](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/contrib/autograph/LIMITATIONS.md).\n", + "\n", + "Here's a quick example of how it works:\n", + "\n" + ] + }, + { + "cell_type": "code", + "execution_count": 0, + "metadata": { + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + }, + "colab_type": "code", + "id": "aA3gOodCBkOw" + }, + "outputs": [], + "source": [ + "# Autograph can convert functions like this...\n", + "def g(x):\n", + " if x \u003e 0:\n", + " x = x * x\n", + " else:\n", + " x = 0.0\n", + " return x\n", + "\n", + "# ...into graph-building functions like this:\n", + "def tf_g(x):\n", + " with tf.name_scope('g'):\n", + " \n", + " def if_true():\n", + " with tf.name_scope('if_true'):\n", + " x_1, = x,\n", + " x_1 = x_1 * x_1\n", + " return x_1,\n", + "\n", + " def if_false():\n", + " with tf.name_scope('if_false'):\n", + " x_1, = x,\n", + " x_1 = 0.0\n", + " return x_1,\n", + "\n", + " x = autograph_utils.run_cond(tf.greater(x, 0), if_true, if_false)\n", + " return x\n" + ] + }, + { + "cell_type": "code", + "execution_count": 0, + "metadata": { + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + }, + "colab_type": "code", + "id": "I1RtBvoKBxq5" + }, + "outputs": [], + "source": [ + "# You can run your plain-Python code in graph mode,\n", + "# and get the same results out, but with all the benfits of graphs:\n", + "print('Original value: %2.2f' % g(9.0))\n", + "\n", + "# Generate a graph-version of g and call it:\n", + "tf_g = autograph.to_graph(g)\n", + "\n", + "with tf.Graph().as_default(): \n", + " # The result works like a regular op: takes tensors in, returns tensors.\n", + " # You can inspect the graph using tf.get_default_graph().as_graph_def()\n", + " g_ops = tf_g(tf.constant(9.0))\n", + " with tf.Session() as sess:\n", + " print('Autograph value: %2.2f\\n' % sess.run(g_ops))\n", + " \n", + " \n", + "# You can view, debug and tweak the generated code:\n", + "print(autograph.to_code(g))" + ] + }, + { + "cell_type": "markdown", + "metadata": { + "colab_type": "text", + "id": "m-jWmsCmByyw" + }, + "source": [ + "#### Automatically converting complex control flow\n", + "\n", + "AutoGraph can convert a large chunk of the Python language into equivalent graph-construction code, and we're adding new supported language features all the time. In this section, we'll give you a taste of some of the functionality in AutoGraph.\n", + "AutoGraph will automatically convert most Python control flow statements into their correct graph equivalent. \n", + " \n", + "We support common statements like `while`, `for`, `if`, `break`, `return` and more. You can even nest them as much as you like. Imagine trying to write the graph version of this code by hand:\n" + ] + }, + { + "cell_type": "code", + "execution_count": 0, + "metadata": { + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + }, + "colab_type": "code", + "id": "toxKBOXbB1ro" + }, + "outputs": [], + "source": [ + "# Continue in a loop\n", + "def f(l):\n", + " s = 0\n", + " for c in l:\n", + " if c % 2 \u003e 0:\n", + " continue\n", + " s += c\n", + " return s\n", + "\n", + "print('Original value: %d' % f([10,12,15,20]))\n", + "\n", + "tf_f = autograph.to_graph(f)\n", + "with tf.Graph().as_default(): \n", + " with tf.Session():\n", + " print('Graph value: %d\\n\\n' % tf_f(tf.constant([10,12,15,20])).eval())\n", + " \n", + "print(autograph.to_code(f))" + ] + }, + { + "cell_type": "markdown", + "metadata": { + "colab_type": "text", + "id": "FUJJ-WTdCGeq" + }, + "source": [ + "Try replacing the `continue` in the above code with `break` -- AutoGraph supports that as well! \n", + " \n", + "Let's try some other useful Python constructs, like `print` and `assert`. We automatically convert Python `assert` statements into the equivalent `tf.Assert` code. " + ] + }, + { + "cell_type": "code", + "execution_count": 0, + "metadata": { + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + }, + "colab_type": "code", + "id": "IAOgh62zCPZ4" + }, + "outputs": [], + "source": [ + "def f(x):\n", + " assert x != 0, 'Do not pass zero!'\n", + " return x * x\n", + "\n", + "tf_f = autograph.to_graph(f)\n", + "with tf.Graph().as_default(): \n", + " with tf.Session():\n", + " try:\n", + " print(tf_f(tf.constant(0)).eval())\n", + " except tf.errors.InvalidArgumentError as e:\n", + " print('Got error message:\\n%s' % e.message)" + ] + }, + { + "cell_type": "markdown", + "metadata": { + "colab_type": "text", + "id": "KRu8iIPBCQr5" + }, + "source": [ + "You can also use plain Python `print` functions in in-graph" + ] + }, + { + "cell_type": "code", + "execution_count": 0, + "metadata": { + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + }, + "colab_type": "code", + "id": "ySTsuxnqCTQi" + }, + "outputs": [], + "source": [ + "def f(n):\n", + " if n \u003e= 0:\n", + " while n \u003c 5:\n", + " n += 1\n", + " print(n)\n", + " return n\n", + " \n", + "tf_f = autograph.to_graph(f)\n", + "with tf.Graph().as_default():\n", + " with tf.Session():\n", + " tf_f(tf.constant(0)).eval()" + ] + }, + { + "cell_type": "markdown", + "metadata": { + "colab_type": "text", + "id": "NqF0GT-VCVFh" + }, + "source": [ + "Appending to lists in loops also works (we create a `TensorArray` for you behind the scenes)" + ] + }, + { + "cell_type": "code", + "execution_count": 0, + "metadata": { + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + }, + "colab_type": "code", + "id": "ABX070KwCczR" + }, + "outputs": [], + "source": [ + "def f(n):\n", + " z = []\n", + " # We ask you to tell us the element dtype of the list\n", + " z = autograph.utils.set_element_type(z, tf.int32)\n", + " for i in range(n):\n", + " z.append(i)\n", + " # when you're done with the list, stack it\n", + " # (this is just like np.stack)\n", + " return autograph.stack(z) \n", + "\n", + "tf_f = autograph.to_graph(f)\n", + "with tf.Graph().as_default(): \n", + " with tf.Session():\n", + " print(tf_f(tf.constant(3)).eval())\n", + "\n", + "print('\\n\\n'+autograph.to_code(f))" + ] + }, + { + "cell_type": "code", + "execution_count": 0, + "metadata": { + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + }, + "colab_type": "code", + "id": "iu5IF7n2Df7C" + }, + "outputs": [], + "source": [ + "def fizzbuzz(num):\n", + " if num % 3 == 0 and num % 5 == 0:\n", + " print('FizzBuzz')\n", + " elif num % 3 == 0:\n", + " print('Fizz')\n", + " elif num % 5 == 0:\n", + " print('Buzz')\n", + " else:\n", + " print(num)\n", + " return num" + ] + }, + { + "cell_type": "code", + "execution_count": 0, + "metadata": { + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + }, + "colab_type": "code", + "id": "EExAjWuwDPpR" + }, + "outputs": [], + "source": [ + "tf_g = autograph.to_graph(fizzbuzz)\n", + "\n", + "with tf.Graph().as_default(): \n", + " # The result works like a regular op: takes tensors in, returns tensors.\n", + " # You can inspect the graph using tf.get_default_graph().as_graph_def()\n", + " g_ops = tf_g(tf.constant(15))\n", + " with tf.Session() as sess:\n", + " sess.run(g_ops) \n", + " \n", + "# You can view, debug and tweak the generated code:\n", + "print('\\n')\n", + "print(autograph.to_code(fizzbuzz))" + ] + }, + { + "cell_type": "markdown", + "metadata": { + "colab_type": "text", + "id": "SzpKGzVpBkph" + }, + "source": [ + "# De-graphify Exercises\n" + ] + }, + { + "cell_type": "markdown", + "metadata": { + "colab_type": "text", + "id": "8k23dxcSmmXq" + }, + "source": [ + "#### Easy print statements" + ] + }, + { + "cell_type": "code", + "execution_count": 0, + "metadata": { + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + }, + "colab_type": "code", + "id": "dE1Vsmp-mlpK" + }, + "outputs": [], + "source": [ + "# See what happens when you turn AutoGraph off.\n", + "# Do you see the type or the value of x when you print it?\n", + "\n", + "# @autograph.convert()\n", + "def square_log(x):\n", + " x = x * x\n", + " print('Squared value of x =', x)\n", + " return x\n", + "\n", + "\n", + "with tf.Graph().as_default(): \n", + " with tf.Session() as sess:\n", + " print(sess.run(square_log(tf.constant(4))))" + ] + }, + { + "cell_type": "markdown", + "metadata": { + "colab_type": "text", + "id": "_R-Q7BbxmkBF" + }, + "source": [ + "#### Now some exercises. Convert the TensorFlow code into AutoGraph'd Python code." + ] + }, + { + "cell_type": "code", + "execution_count": 0, + "metadata": { + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + }, + "colab_type": "code", + "id": "SwA11tO-yCvg" + }, + "outputs": [], + "source": [ + "def square_if_positive(x):\n", + " x = tf.cond(tf.greater(x, 0), lambda: x * x, lambda: x)\n", + " return x\n", + "\n", + "with tf.Session() as sess:\n", + " print(sess.run(square_if_positive(tf.constant(4))))" + ] + }, + { + "cell_type": "code", + "execution_count": 0, + "metadata": { + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + }, + "colab_type": "code", + "id": "GPmx4CNhyPI_" + }, + "outputs": [], + "source": [ + "@autograph.convert()\n", + "def square_if_positive(x):\n", + " ... # \u003c\u003c\u003c fill it in!\n", + " \n", + "with tf.Session() as sess:\n", + " print(sess.run(square_if_positive(tf.constant(4))))" + ] + }, + { + "cell_type": "markdown", + "metadata": { + "colab_type": "text", + "id": "qqsjik-QyA9R" + }, + "source": [ + "#### Uncollapse to see answer" + ] + }, + { + "cell_type": "code", + "execution_count": 0, + "metadata": { + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + }, + "colab_type": "code", + "id": "DaSmaWUEvMRv" + }, + "outputs": [], + "source": [ + "# Simple cond\n", + "@autograph.convert()\n", + "def square_if_positive(x):\n", + " if x \u003e 0:\n", + " x = x * x\n", + " return x\n", + "\n", + "with tf.Graph().as_default(): \n", + " with tf.Session() as sess:\n", + " print(sess.run(square_if_positive(tf.constant(4))))" + ] + }, + { + "cell_type": "markdown", + "metadata": { + "colab_type": "text", + "id": "qj7am2I_xvTJ" + }, + "source": [ + "#### Nested If statement" + ] + }, + { + "cell_type": "code", + "execution_count": 0, + "metadata": { + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + }, + "colab_type": "code", + "id": "4yyNOf-Twr6s" + }, + "outputs": [], + "source": [ + "def nearest_odd_square(x):\n", + "\n", + " def if_positive():\n", + " x1 = x * x\n", + " x1 = tf.cond(tf.equal(x1 % 2, 0), lambda: x1 + 1, lambda: x1)\n", + " return x1,\n", + "\n", + " x = tf.cond(tf.greater(x, 0), if_positive, lambda: x)\n", + " return x\n", + "\n", + "with tf.Graph().as_default(): \n", + " with tf.Session() as sess:\n", + " print(sess.run(nearest_odd_square(tf.constant(4))))" + ] + }, + { + "cell_type": "code", + "execution_count": 0, + "metadata": { + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + }, + "colab_type": "code", + "id": "hqmh5b2VyU9w" + }, + "outputs": [], + "source": [ + "@autograph.convert()\n", + "def nearest_odd_square(x):\n", + " ... # \u003c\u003c\u003c fill it in!\n", + " \n", + "with tf.Session() as sess:\n", + " print(sess.run(nearest_odd_square(tf.constant(4))))" + ] + }, + { + "cell_type": "markdown", + "metadata": { + "colab_type": "text", + "id": "b9AXIkNLxp6J" + }, + "source": [ + "#### Uncollapse to reveal answer" + ] + }, + { + "cell_type": "code", + "execution_count": 0, + "metadata": { + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + }, + "colab_type": "code", + "id": "8RlCVEpNxD91" + }, + "outputs": [], + "source": [ + "@autograph.convert()\n", + "def nearest_odd_square(x):\n", + " if x \u003e 0:\n", + " x = x * x\n", + " if x % 2 == 0:\n", + " x = x + 1\n", + " return x\n", + "\n", + "with tf.Graph().as_default(): \n", + " with tf.Session() as sess:\n", + " print(sess.run(nearest_odd_square(tf.constant(4))))" + ] + }, + { + "cell_type": "markdown", + "metadata": { + "colab_type": "text", + "id": "jXAxjeBr1qWK" + }, + "source": [ + "#### Convert a while loop" + ] + }, + { + "cell_type": "code", + "execution_count": 0, + "metadata": { + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + }, + "colab_type": "code", + "id": "kWkv7anlxoee" + }, + "outputs": [], + "source": [ + "# Convert a while loop\n", + "def square_until_stop(x, y):\n", + " x = tf.while_loop(lambda x: tf.less(x, y), lambda x: x * x, [x])\n", + " return x\n", + " \n", + "with tf.Graph().as_default(): \n", + " with tf.Session() as sess:\n", + " print(sess.run(square_until_stop(tf.constant(4), tf.constant(100))))" + ] + }, + { + "cell_type": "code", + "execution_count": 0, + "metadata": { + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + }, + "colab_type": "code", + "id": "zVUsc1eA1u2K" + }, + "outputs": [], + "source": [ + "@autograph.convert()\n", + "def square_until_stop(x, y):\n", + " ... # fill it in!\n", + " \n", + "with tf.Graph().as_default(): \n", + " with tf.Session() as sess:\n", + " print(sess.run(square_until_stop(tf.constant(4), tf.constant(100))))" + ] + }, + { + "cell_type": "markdown", + "metadata": { + "colab_type": "text", + "id": "L2psuzPI02S9" + }, + "source": [ + "#### Uncollapse for the answer\n" + ] + }, + { + "cell_type": "code", + "execution_count": 0, + "metadata": { + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + }, + "colab_type": "code", + "id": "ucmZyQVL03bF" + }, + "outputs": [], + "source": [ + "@autograph.convert()\n", + "def square_until_stop(x, y):\n", + " while x \u003c y:\n", + " x = x * x\n", + " return x\n", + " \n", + "with tf.Graph().as_default(): \n", + " with tf.Session() as sess:\n", + " print(sess.run(square_until_stop(tf.constant(4), tf.constant(100))))" + ] + }, + { + "cell_type": "markdown", + "metadata": { + "colab_type": "text", + "id": "FXB0Zbwl13PY" + }, + "source": [ + "#### Nested loop and conditional" + ] + }, + { + "cell_type": "code", + "execution_count": 0, + "metadata": { + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + }, + "colab_type": "code", + "id": "clGymxdf15Ig" + }, + "outputs": [], + "source": [ + "@autograph.convert()\n", + "def argwhere_cumsum(x, threshold):\n", + " current_sum = 0.0\n", + " idx = 0\n", + " \n", + " for i in range(len(x)):\n", + " idx = i\n", + " if current_sum \u003e= threshold:\n", + " break\n", + " current_sum += x[i]\n", + " return idx\n", + "\n", + "N = 10\n", + "with tf.Graph().as_default(): \n", + " with tf.Session() as sess:\n", + " idx = argwhere_cumsum(tf.ones(N), tf.constant(float(N/2)))\n", + " print(sess.run(idx))" + ] + }, + { + "cell_type": "code", + "execution_count": 0, + "metadata": { + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + }, + "colab_type": "code", + "id": "i7PF-uId9lp5" + }, + "outputs": [], + "source": [ + "@autograph.convert()\n", + "def argwhere_cumsum(x, threshold):\n", + " ...\n", + "\n", + "N = 10\n", + "with tf.Graph().as_default(): \n", + " with tf.Session() as sess:\n", + " idx = argwhere_cumsum(tf.ones(N), tf.constant(float(N/2)))\n", + " print(sess.run(idx))" + ] + }, + { + "cell_type": "markdown", + "metadata": { + "colab_type": "text", + "id": "weKFXAb615Vp" + }, + "source": [ + "#### Uncollapse to see answer" + ] + }, + { + "cell_type": "code", + "execution_count": 0, + "metadata": { + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + }, + "colab_type": "code", + "id": "1sjaFcL717Ig" + }, + "outputs": [], + "source": [ + "@autograph.convert()\n", + "def argwhere_cumsum(x, threshold):\n", + " current_sum = 0.0\n", + " idx = 0\n", + " for i in range(len(x)):\n", + " idx = i\n", + " if current_sum \u003e= threshold:\n", + " break\n", + " current_sum += x[i]\n", + " return idx\n", + "\n", + "N = 10\n", + "with tf.Graph().as_default(): \n", + " with tf.Session() as sess:\n", + " idx = argwhere_cumsum(tf.ones(N), tf.constant(float(N/2)))\n", + " print(sess.run(idx))" + ] + }, + { + "cell_type": "markdown", + "metadata": { + "colab_type": "text", + "id": "4LfnJjm0Bm0B" + }, + "source": [ + "# 3. Training MNIST in-graph\n", + "\n", + "Writing control flow in AutoGraph is easy, so running a training loop in a TensorFlow graph should be easy as well! \n", + "\n", + "Here, we show an example of training a simple Keras model on MNIST, where the entire training process -- loading batches, calculating gradients, updating parameters, calculating validation accuracy, and repeating until convergence -- is done in-graph." + ] + }, + { + "cell_type": "markdown", + "metadata": { + "colab_type": "text", + "id": "Em5dzSUOtLRP" + }, + "source": [ + "#### Download data" + ] + }, + { + "cell_type": "code", + "execution_count": 0, + "metadata": { + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + }, + "colab_type": "code", + "id": "xqoxumv0ssQW" + }, + "outputs": [], + "source": [ + "import gzip\n", + "import os\n", + "import shutil\n", + "\n", + "from six.moves import urllib\n", + "\n", + "\n", + "def download(directory, filename):\n", + " filepath = os.path.join(directory, filename)\n", + " if tf.gfile.Exists(filepath):\n", + " return filepath\n", + " if not tf.gfile.Exists(directory):\n", + " tf.gfile.MakeDirs(directory)\n", + " url = 'https://storage.googleapis.com/cvdf-datasets/mnist/' + filename + '.gz'\n", + " zipped_filepath = filepath + '.gz'\n", + " print('Downloading %s to %s' % (url, zipped_filepath))\n", + " urllib.request.urlretrieve(url, zipped_filepath)\n", + " with gzip.open(zipped_filepath, 'rb') as f_in, open(filepath, 'wb') as f_out:\n", + " shutil.copyfileobj(f_in, f_out)\n", + " os.remove(zipped_filepath)\n", + " return filepath\n", + "\n", + "\n", + "def dataset(directory, images_file, labels_file):\n", + " images_file = download(directory, images_file)\n", + " labels_file = download(directory, labels_file)\n", + "\n", + " def decode_image(image):\n", + " # Normalize from [0, 255] to [0.0, 1.0]\n", + " image = tf.decode_raw(image, tf.uint8)\n", + " image = tf.cast(image, tf.float32)\n", + " image = tf.reshape(image, [784])\n", + " return image / 255.0\n", + "\n", + " def decode_label(label):\n", + " label = tf.decode_raw(label, tf.uint8)\n", + " label = tf.reshape(label, [])\n", + " return tf.to_int32(label)\n", + "\n", + " images = tf.data.FixedLengthRecordDataset(\n", + " images_file, 28 * 28, header_bytes=16).map(decode_image)\n", + " labels = tf.data.FixedLengthRecordDataset(\n", + " labels_file, 1, header_bytes=8).map(decode_label)\n", + " return tf.data.Dataset.zip((images, labels))\n", + "\n", + "\n", + "def mnist_train(directory):\n", + " return dataset(directory, 'train-images-idx3-ubyte',\n", + " 'train-labels-idx1-ubyte')\n", + "\n", + "def mnist_test(directory):\n", + " return dataset(directory, 't10k-images-idx3-ubyte', 't10k-labels-idx1-ubyte')" + ] + }, + { + "cell_type": "markdown", + "metadata": { + "colab_type": "text", + "id": "znmy4l8ntMvW" + }, + "source": [ + "#### Define the model" + ] + }, + { + "cell_type": "code", + "execution_count": 0, + "metadata": { + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + }, + "colab_type": "code", + "id": "Pe-erWQdBoC5" + }, + "outputs": [], + "source": [ + "def mlp_model(input_shape):\n", + " model = tf.keras.Sequential((\n", + " tf.keras.layers.Dense(100, activation='relu', input_shape=input_shape),\n", + " tf.keras.layers.Dense(100, activation='relu'),\n", + " tf.keras.layers.Dense(10, activation='softmax')))\n", + " model.build()\n", + " return model\n", + "\n", + "\n", + "def predict(m, x, y):\n", + " y_p = m(x)\n", + " losses = tf.keras.losses.categorical_crossentropy(y, y_p)\n", + " l = tf.reduce_mean(losses)\n", + " accuracies = tf.keras.metrics.categorical_accuracy(y, y_p)\n", + " accuracy = tf.reduce_mean(accuracies)\n", + " return l, accuracy\n", + "\n", + "\n", + "def fit(m, x, y, opt):\n", + " l, accuracy = predict(m, x, y)\n", + " opt.minimize(l)\n", + " return l, accuracy\n", + "\n", + "\n", + "def setup_mnist_data(is_training, hp, batch_size):\n", + " if is_training:\n", + " ds = mnist_train('/tmp/autograph_mnist_data')\n", + " ds = ds.shuffle(batch_size * 10)\n", + " else:\n", + " ds = mnist_test('/tmp/autograph_mnist_data')\n", + " ds = ds.repeat()\n", + " ds = ds.batch(batch_size)\n", + " return ds\n", + "\n", + "\n", + "def get_next_batch(ds):\n", + " itr = ds.make_one_shot_iterator()\n", + " image, label = itr.get_next()\n", + " x = tf.to_float(tf.reshape(image, (-1, 28 * 28)))\n", + " y = tf.one_hot(tf.squeeze(label), 10)\n", + " return x, y" + ] + }, + { + "cell_type": "markdown", + "metadata": { + "colab_type": "text", + "id": "oeYV6mKnJGMr" + }, + "source": [ + "#### Define the training loop" + ] + }, + { + "cell_type": "code", + "execution_count": 0, + "metadata": { + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + }, + "colab_type": "code", + "id": "3xtg_MMhJETd" + }, + "outputs": [], + "source": [ + "def train(train_ds, test_ds, hp):\n", + " m = mlp_model((28 * 28,))\n", + " opt = tf.train.MomentumOptimizer(hp.learning_rate, 0.9)\n", + " \n", + " # We'd like to save our losses to a list. In order for AutoGraph\n", + " # to convert these lists into their graph equivalent,\n", + " # we need to specify the element type of the lists.\n", + " train_losses = []\n", + " train_losses = autograph.utils.set_element_type(train_losses, tf.float32)\n", + " test_losses = []\n", + " test_losses = autograph.utils.set_element_type(test_losses, tf.float32)\n", + " train_accuracies = []\n", + " train_accuracies = autograph.utils.set_element_type(train_accuracies, tf.float32)\n", + " test_accuracies = []\n", + " test_accuracies = autograph.utils.set_element_type(test_accuracies, tf.float32)\n", + " \n", + " # This entire training loop will be run in-graph.\n", + " i = tf.constant(0)\n", + " while i \u003c hp.max_steps:\n", + " train_x, train_y = get_next_batch(train_ds)\n", + " test_x, test_y = get_next_batch(test_ds)\n", + " # add get next\n", + " step_train_loss, step_train_accuracy = fit(m, train_x, train_y, opt)\n", + " step_test_loss, step_test_accuracy = predict(m, test_x, test_y)\n", + " if i % (hp.max_steps // 10) == 0:\n", + " print('Step', i, 'train loss:', step_train_loss, 'test loss:',\n", + " step_test_loss, 'train accuracy:', step_train_accuracy,\n", + " 'test accuracy:', step_test_accuracy)\n", + " train_losses.append(step_train_loss)\n", + " test_losses.append(step_test_loss)\n", + " train_accuracies.append(step_train_accuracy)\n", + " test_accuracies.append(step_test_accuracy)\n", + " i += 1\n", + " \n", + " # We've recorded our loss values and accuracies \n", + " # to a list in a graph with AutoGraph's help.\n", + " # In order to return the values as a Tensor, \n", + " # we need to stack them before returning them.\n", + " return (autograph.stack(train_losses), autograph.stack(test_losses), autograph.stack(train_accuracies),\n", + " autograph.stack(test_accuracies))" + ] + }, + { + "cell_type": "code", + "execution_count": 0, + "metadata": { + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + }, + "colab_type": "code", + "id": "HYh6MSZyJOag" + }, + "outputs": [], + "source": [ + "with tf.Graph().as_default():\n", + " hp = tf.contrib.training.HParams(\n", + " learning_rate=0.05,\n", + " max_steps=500,\n", + " )\n", + " train_ds = setup_mnist_data(True, hp, 50)\n", + " test_ds = setup_mnist_data(False, hp, 1000)\n", + " tf_train = autograph.to_graph(train)\n", + " (train_losses, test_losses, train_accuracies,\n", + " test_accuracies) = tf_train(train_ds, test_ds, hp)\n", + "\n", + " with tf.Session() as sess:\n", + " sess.run(tf.global_variables_initializer())\n", + " (train_losses, test_losses, train_accuracies,\n", + " test_accuracies) = sess.run([train_losses, test_losses, train_accuracies,\n", + " test_accuracies])\n", + " plt.title('MNIST train/test losses')\n", + " plt.plot(train_losses, label='train loss')\n", + " plt.plot(test_losses, label='test loss')\n", + " plt.legend()\n", + " plt.xlabel('Training step')\n", + " plt.ylabel('Loss')\n", + " plt.show()\n", + " plt.title('MNIST train/test accuracies')\n", + " plt.plot(train_accuracies, label='train accuracy')\n", + " plt.plot(test_accuracies, label='test accuracy')\n", + " plt.legend(loc='lower right')\n", + " plt.xlabel('Training step')\n", + " plt.ylabel('Accuracy')\n", + " plt.show()" + ] + } + ], + "metadata": { + "colab": { + "collapsed_sections": [ + "qqsjik-QyA9R", + "b9AXIkNLxp6J", + "L2psuzPI02S9", + "weKFXAb615Vp", + "Em5dzSUOtLRP" + ], + "default_view": {}, + "name": "AutoGraph Workshop.ipynb", + "provenance": [ + { + "file_id": "1kE2gz_zuwdYySL4K2HQSz13uLCYi-fYP", + "timestamp": 1530563781803 + } + ], + "version": "0.3.2", + "views": {} + } + }, + "nbformat": 4, + "nbformat_minor": 0 +} diff --git a/tensorflow/contrib/autograph/operators/__init__.py b/tensorflow/contrib/autograph/operators/__init__.py index c900fd6af2ea5dfb419f731ee8d8822d68424b27..392cb60bcc44c0f554defcddc50c4afbdaa25067 100644 --- a/tensorflow/contrib/autograph/operators/__init__.py +++ b/tensorflow/contrib/autograph/operators/__init__.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. # ============================================================================== -"""This module implements operators that we overload. +"""This module implements operators that AutoGraph overloads. Note that "operator" is used loosely here, and includes control structures like conditionals and loops, implemented in functional form, using for example diff --git a/tensorflow/contrib/autograph/pyct/BUILD b/tensorflow/contrib/autograph/pyct/BUILD index 8f09689fe9b33bec03dc8b5370633c3a953fa322..f77a6ab3928e4f933f4d21abba2030d4d6f8ec0a 100644 --- a/tensorflow/contrib/autograph/pyct/BUILD +++ b/tensorflow/contrib/autograph/pyct/BUILD @@ -22,8 +22,10 @@ py_library( "__init__.py", "anno.py", "ast_util.py", + "cfg.py", "compiler.py", "inspect_utils.py", + "origin_info.py", "parser.py", "pretty_printer.py", "qual_names.py", @@ -63,6 +65,17 @@ py_test( ], ) +py_test( + name = "cfg_test", + srcs = ["cfg_test.py"], + srcs_version = "PY2AND3", + deps = [ + ":pyct", + "//tensorflow/python:client_testlib", + "@gast_archive//:gast", + ], +) + py_test( name = "compiler_test", srcs = ["compiler_test.py"], diff --git a/tensorflow/contrib/autograph/pyct/anno.py b/tensorflow/contrib/autograph/pyct/anno.py index ae861627fd65cca057e7bf1af41424e605d4b7a1..1a52110ef36bbc0888e03cc25b3717822cb75c16 100644 --- a/tensorflow/contrib/autograph/pyct/anno.py +++ b/tensorflow/contrib/autograph/pyct/anno.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. # ============================================================================== -"""Handling annotations on AST nodes. +"""AST node annotation support. Adapted from Tangent. """ @@ -21,37 +21,90 @@ from __future__ import absolute_import from __future__ import division from __future__ import print_function -from enum import Enum +import enum +# pylint:disable=g-bad-import-order +import gast +# pylint:enable=g-bad-import-order -class NoValue(Enum): + +# TODO(mdan): Shorten the names. +# These names are heavily used, and anno.blaa +# TODO(mdan): Replace the attr-dict mechanism with a more typed solution. + + +class NoValue(enum.Enum): def __repr__(self): return self.name class Basic(NoValue): - """Container for annotation keys. + """Container for basic annotation keys. The enum values are used strictly for documentation purposes. """ - QN = 'Qualified name, as it appeared in the code.' + QN = 'Qualified name, as it appeared in the code. See qual_names.py.' SKIP_PROCESSING = ( 'This node should be preserved as is and not processed any further.') INDENT_BLOCK_REMAINDER = ( - 'When a node is annotated with this, the remainder of the block should ' - 'be indented below it. The annotation contains a tuple ' - '(new_body, name_map), where `new_body` is the new indented block and ' - '`name_map` allows renaming symbols.') + 'When a node is annotated with this, the remainder of the block should' + ' be indented below it. The annotation contains a tuple' + ' (new_body, name_map), where `new_body` is the new indented block and' + ' `name_map` allows renaming symbols.') + ORIGIN = ('Information about the source code that converted code originated' + ' from. See origin_information.py.') + + +class Static(NoValue): + """Container for static analysis annotation keys. + + The enum values are used strictly for documentation purposes. + """ + + # Deprecated - use reaching definitions instead. + # Symbols + # These flags are boolean. + IS_LOCAL = 'Symbol is local to the function scope being analyzed.' + IS_PARAM = 'Symbol is a parameter to the function being analyzed.' + + # Scopes + # Scopes are represented by objects of type activity.Scope. + SCOPE = 'The scope for the annotated node. See activity.py.' + # TODO(mdan): Drop these in favor of accessing the child's SCOPE. + ARGS_SCOPE = 'The scope for the argument list of a function call.' + COND_SCOPE = 'The scope for the test node of a conditional statement.' + BODY_SCOPE = ( + 'The scope for the main body of a statement (True branch for if ' + 'statements, main body for loops).') + ORELSE_SCOPE = ( + 'The scope for the orelse body of a statement (False branch for if ' + 'statements, orelse body for loops).') + + # Static analysis annotations. + DEFINITIONS = ( + 'Reaching definition information. See reaching_definitions.py.') + ORIG_DEFINITIONS = ( + 'The value of DEFINITIONS that applied to the original code before any' + ' conversion.') + DEFINED_VARS_IN = ( + 'Symbols defined when entering the node. See reaching_definitions.py.') + LIVE_VARS_OUT = ('Symbols live when exiting the node. See liveness.py.') FAIL = object() +def keys(node, field_name='___pyct_anno'): + if not hasattr(node, field_name): + return frozenset() + return frozenset(getattr(node, field_name).keys()) + + def getanno(node, key, default=FAIL, field_name='___pyct_anno'): - if (default is FAIL or - (hasattr(node, field_name) and (key in getattr(node, field_name)))): + if (default is FAIL or (hasattr(node, field_name) and + (key in getattr(node, field_name)))): return getattr(node, field_name)[key] else: return default @@ -86,3 +139,19 @@ def copyanno(from_node, to_node, key, field_name='___pyct_anno'): key, getanno(from_node, key, field_name=field_name), field_name=field_name) + + +def dup(node, copy_map, field_name='___pyct_anno'): + """Recursively copies annotations in an AST tree. + + Args: + node: ast.AST + copy_map: Dict[Hashable, Hashable], maps a source anno key to a destination + key. All annotations with the source key will be copied to identical + annotations with the destination key. + field_name: str + """ + for n in gast.walk(node): + for k in copy_map: + if hasanno(n, k, field_name): + setanno(n, copy_map[k], getanno(n, k, field_name), field_name) diff --git a/tensorflow/contrib/autograph/pyct/anno_test.py b/tensorflow/contrib/autograph/pyct/anno_test.py index f2c0c8cf05ca4b3671eb653ce56f6da61de54aee..5ef4da61a3627f9c0bc615ce5cb56052a28c64d1 100644 --- a/tensorflow/contrib/autograph/pyct/anno_test.py +++ b/tensorflow/contrib/autograph/pyct/anno_test.py @@ -32,22 +32,27 @@ class AnnoTest(test.TestCase): def test_basic(self): node = ast.Name() + self.assertEqual(anno.keys(node), set()) self.assertFalse(anno.hasanno(node, 'foo')) with self.assertRaises(AttributeError): anno.getanno(node, 'foo') anno.setanno(node, 'foo', 3) + + self.assertEqual(anno.keys(node), {'foo'}) self.assertTrue(anno.hasanno(node, 'foo')) self.assertEqual(anno.getanno(node, 'foo'), 3) self.assertEqual(anno.getanno(node, 'bar', default=7), 7) anno.delanno(node, 'foo') + + self.assertEqual(anno.keys(node), set()) self.assertFalse(anno.hasanno(node, 'foo')) with self.assertRaises(AttributeError): anno.getanno(node, 'foo') self.assertIsNone(anno.getanno(node, 'foo', default=None)) - def test_copyanno(self): + def test_copy(self): node_1 = ast.Name() anno.setanno(node_1, 'foo', 3) @@ -58,6 +63,22 @@ class AnnoTest(test.TestCase): self.assertTrue(anno.hasanno(node_2, 'foo')) self.assertFalse(anno.hasanno(node_2, 'bar')) + def test_duplicate(self): + node = ast.If( + test=ast.Num(1), + body=[ast.Expr(ast.Name('bar', ast.Load()))], + orelse=[]) + anno.setanno(node, 'spam', 1) + anno.setanno(node, 'ham', 1) + anno.setanno(node.body[0], 'ham', 1) + + anno.dup(node, {'spam': 'eggs'}) + + self.assertTrue(anno.hasanno(node, 'spam')) + self.assertTrue(anno.hasanno(node, 'ham')) + self.assertTrue(anno.hasanno(node, 'eggs')) + self.assertFalse(anno.hasanno(node.body[0], 'eggs')) + if __name__ == '__main__': test.main() diff --git a/tensorflow/contrib/autograph/pyct/ast_util.py b/tensorflow/contrib/autograph/pyct/ast_util.py index c4f82d11708393a6029d3f17be428b47eb9342ff..0cf87dd8d3b07b818b18a82d5b8d971f393751bc 100644 --- a/tensorflow/contrib/autograph/pyct/ast_util.py +++ b/tensorflow/contrib/autograph/pyct/ast_util.py @@ -12,7 +12,7 @@ # See the License for the specific language governing permissions and # limitations under the License. # ============================================================================== -"""Copy an AST tree, discarding annotations.""" +"""AST manipulation utilities.""" from __future__ import absolute_import from __future__ import division @@ -26,47 +26,53 @@ from tensorflow.contrib.autograph.pyct import anno from tensorflow.contrib.autograph.pyct import parser -class CleanCopier(gast.NodeVisitor): - """Copies AST nodes. +class CleanCopier(object): + """NodeTransformer-like visitor that copies an AST.""" - The copied nodes will ignore almost all fields that are prefixed by '__'. - Exceptions make some annotations. - """ + def __init__(self, preserve_annos): + super(CleanCopier, self).__init__() + self.preserve_annos = preserve_annos - # TODO(mdan): Parametrize which annotations get carried over. + def copy(self, node): + """Returns a deep copy of node (excluding some fields, see copy_clean).""" + + if isinstance(node, list): + return [self.copy(n) for n in node] + elif isinstance(node, tuple): + return tuple(self.copy(n) for n in node) + elif not isinstance(node, (gast.AST, ast.AST)): + # Assuming everything that's not an AST, list or tuple is a value type + # and may simply be assigned. + return node + + assert isinstance(node, (gast.AST, ast.AST)) - def generic_visit(self, node): new_fields = {} for f in node._fields: - if f.startswith('__'): - continue - if not hasattr(node, f): - continue - v = getattr(node, f) - if isinstance(v, list): - v = [self.generic_visit(n) for n in v] - elif isinstance(v, tuple): - v = tuple(self.generic_visit(n) for n in v) - elif isinstance(v, (gast.AST, ast.AST)): - v = self.generic_visit(v) - else: - # Assume everything else is a value type. - pass - new_fields[f] = v + if not f.startswith('__') and hasattr(node, f): + new_fields[f] = self.copy(getattr(node, f)) new_node = type(node)(**new_fields) - if anno.hasanno(node, anno.Basic.SKIP_PROCESSING): - anno.setanno(new_node, anno.Basic.SKIP_PROCESSING, True) + + if self.preserve_annos: + for k in self.preserve_annos: + anno.copyanno(node, new_node, k) return new_node -def copy_clean(node): - copier = CleanCopier() - if isinstance(node, list): - return [copier.visit(n) for n in node] - elif isinstance(node, tuple): - return tuple(copier.visit(n) for n in node) - else: - return copier.visit(node) +def copy_clean(node, preserve_annos=None): + """Creates a deep copy of an AST. + + The copy will not include fields that are prefixed by '__', with the + exception of user-specified annotations. + + Args: + node: ast.AST + preserve_annos: Optional[Set[Hashable]], annotation keys to include in the + copy + Returns: + ast.AST + """ + return CleanCopier(preserve_annos).copy(node) class SymbolRenamer(gast.NodeTransformer): @@ -78,7 +84,11 @@ class SymbolRenamer(gast.NodeTransformer): def _process(self, node): qn = anno.getanno(node, anno.Basic.QN) if qn in self.name_map: - return gast.Name(str(self.name_map[qn]), node.ctx, None) + new_node = gast.Name(str(self.name_map[qn]), node.ctx, None) + # All annotations get carried over. + for k in anno.keys(node): + anno.copyanno(node, new_node, k) + return new_node return self.generic_visit(node) def visit_Name(self, node): @@ -92,6 +102,7 @@ class SymbolRenamer(gast.NodeTransformer): def rename_symbols(node, name_map): + """Renames symbols in an AST. Requires qual_names annotations.""" renamer = SymbolRenamer(name_map) if isinstance(node, list): return [renamer.visit(n) for n in node] @@ -101,6 +112,7 @@ def rename_symbols(node, name_map): def keywords_to_dict(keywords): + """Converts a list of ast.keyword objects to a dict.""" keys = [] values = [] for kw in keywords: @@ -110,10 +122,7 @@ def keywords_to_dict(keywords): class PatternMatcher(gast.NodeVisitor): - """Matches a node against a pattern represented by a node. - - The pattern may contain wildcards represented by the symbol '_'. - """ + """Matches a node against a pattern represented by a node.""" def __init__(self, pattern): self.pattern = pattern @@ -177,9 +186,68 @@ class PatternMatcher(gast.NodeVisitor): def matches(node, pattern): + """Basic pattern matcher for AST. + + The pattern may contain wildcards represented by the symbol '_'. A node + matches a pattern if for every node in the tree, either there is a node of + the same type in pattern, or a Name node with id='_'. + + Args: + node: ast.AST + pattern: ast.AST + Returns: + bool + """ if isinstance(pattern, str): pattern = parser.parse_expression(pattern) matcher = PatternMatcher(pattern) matcher.visit(node) return matcher.matches + +# TODO(mdan): Once we have error tracing, we may be able to just go to SSA. +def apply_to_single_assignments(targets, values, apply_fn): + """Applies a function to each individual assignment. + + This function can process a possibly-unpacked (e.g. a, b = c, d) assignment. + It tries to break down the unpacking if possible. In effect, it has the same + effect as passing the assigned values in SSA form to apply_fn. + + Examples: + + The following will result in apply_fn(a, c), apply_fn(b, d): + + a, b = c, d + + The following will result in apply_fn(a, c[0]), apply_fn(b, c[1]): + + a, b = c + + The following will result in apply_fn(a, (b, c)): + + a = b, c + + It uses the visitor pattern to allow subclasses to process single + assignments individually. + + Args: + targets: Union[List[ast.AST, ...], Tuple[ast.AST, ...], ast.AST, should be + used with the targets field of an ast.Assign node + values: ast.AST + apply_fn: Callable[[ast.AST, ast.AST], None], called with the + respective nodes of each single assignment + """ + if not isinstance(targets, (list, tuple)): + targets = (targets,) + for target in targets: + if isinstance(target, (gast.Tuple, gast.List)): + for i in range(len(target.elts)): + target_el = target.elts[i] + if isinstance(values, (gast.Tuple, gast.List)): + value_el = values.elts[i] + else: + idx = parser.parse_expression(str(i)) + value_el = gast.Subscript(values, gast.Index(idx), ctx=gast.Load()) + apply_to_single_assignments(target_el, value_el, apply_fn) + else: + apply_fn(target, values) diff --git a/tensorflow/contrib/autograph/pyct/ast_util_test.py b/tensorflow/contrib/autograph/pyct/ast_util_test.py index 3afa04a50685d19c90944c14ed39f9d3ad35e486..bd546c7f48b9adcda672897d04c8d243582148fd 100644 --- a/tensorflow/contrib/autograph/pyct/ast_util_test.py +++ b/tensorflow/contrib/autograph/pyct/ast_util_test.py @@ -19,7 +19,10 @@ from __future__ import division from __future__ import print_function import ast +import collections +import textwrap +from tensorflow.contrib.autograph.pyct import anno from tensorflow.contrib.autograph.pyct import ast_util from tensorflow.contrib.autograph.pyct import compiler from tensorflow.contrib.autograph.pyct import parser @@ -29,53 +32,65 @@ from tensorflow.python.platform import test class AstUtilTest(test.TestCase): - def test_rename_symbols(self): - node = ast.Tuple([ - ast.Name('a', ast.Load()), - ast.Name('b', ast.Load()), - ast.Attribute(ast.Name('b', None), 'c', ast.Store()), - ast.Attribute( - ast.Attribute(ast.Name('b', None), 'c', ast.Load()), 'd', None) - ], None) + def setUp(self): + super(AstUtilTest, self).setUp() + self._invocation_counts = collections.defaultdict(lambda: 0) + + def test_rename_symbols_basic(self): + node = parser.parse_str('a + b') + node = qual_names.resolve(node) + + node = ast_util.rename_symbols( + node, {qual_names.QN('a'): qual_names.QN('renamed_a')}) + + self.assertIsInstance(node.body[0].value.left.id, str) + self.assertEqual(compiler.ast_to_source(node).strip(), 'renamed_a + b') + + def test_rename_symbols_attributes(self): + node = parser.parse_str('b.c = b.c.d') node = qual_names.resolve(node) + node = ast_util.rename_symbols( - node, { - qual_names.QN('a'): - qual_names.QN('renamed_a'), - qual_names.QN(qual_names.QN('b'), attr='c'): - qual_names.QN('renamed_b_c'), - }) - - self.assertEqual(node.elts[0].id, 'renamed_a') - self.assertTrue(isinstance(node.elts[0].ctx, ast.Load)) - self.assertEqual(node.elts[1].id, 'b') - self.assertEqual(node.elts[2].id, 'renamed_b_c') - self.assertTrue(isinstance(node.elts[2].ctx, ast.Store)) - self.assertEqual(node.elts[3].value.id, 'renamed_b_c') - self.assertTrue(isinstance(node.elts[3].value.ctx, ast.Load)) + node, {qual_names.from_str('b.c'): qual_names.QN('renamed_b_c')}) + + self.assertEqual( + compiler.ast_to_source(node).strip(), 'renamed_b_c = renamed_b_c.d') + + def test_rename_symbols_annotations(self): + node = parser.parse_str('a[i]') + node = qual_names.resolve(node) + anno.setanno(node, 'foo', 'bar') + orig_anno = anno.getanno(node, 'foo') + + node = ast_util.rename_symbols(node, + {qual_names.QN('a'): qual_names.QN('b')}) + + self.assertIs(anno.getanno(node, 'foo'), orig_anno) def test_copy_clean(self): - ret = ast.Return( - ast.BinOp( - op=ast.Add(), - left=ast.Name(id='a', ctx=ast.Load()), - right=ast.Num(1))) - setattr(ret, '__foo', 'bar') - node = ast.FunctionDef( - name='f', - args=ast.arguments( - args=[ast.Name(id='a', ctx=ast.Param())], - vararg=None, - kwarg=None, - defaults=[]), - body=[ret], - decorator_list=[], - returns=None) + node = parser.parse_str( + textwrap.dedent(""" + def f(a): + return a + 1 + """)) + setattr(node.body[0], '__foo', 'bar') new_node = ast_util.copy_clean(node) - self.assertFalse(node is new_node) - self.assertFalse(ret is new_node.body[0]) + self.assertIsNot(new_node, node) + self.assertIsNot(new_node.body[0], node.body[0]) self.assertFalse(hasattr(new_node.body[0], '__foo')) + def test_copy_clean_preserves_annotations(self): + node = parser.parse_str( + textwrap.dedent(""" + def f(a): + return a + 1 + """)) + anno.setanno(node.body[0], 'foo', 'bar') + anno.setanno(node.body[0], 'baz', 1) + new_node = ast_util.copy_clean(node, preserve_annos={'foo'}) + self.assertEqual(anno.getanno(new_node.body[0], 'foo'), 'bar') + self.assertFalse(anno.hasanno(new_node.body[0], 'baz')) + def test_keywords_to_dict(self): keywords = parser.parse_expression('f(a=b, c=1, d=\'e\')').keywords d = ast_util.keywords_to_dict(keywords) @@ -113,6 +128,33 @@ class AstUtilTest(test.TestCase): self.assertNoMatch('super(Foo, self).__init__()', 'super(Bar, _).__init__(_)') + def _mock_apply_fn(self, target, source): + target = compiler.ast_to_source(target).strip() + source = compiler.ast_to_source(source).strip() + self._invocation_counts[(target, source)] += 1 + + def test_apply_to_single_assignments_dynamic_unpack(self): + node = parser.parse_str('a, b, c = d') + node = node.body[0] + ast_util.apply_to_single_assignments(node.targets, node.value, + self._mock_apply_fn) + self.assertDictEqual(self._invocation_counts, { + ('a', 'd[0]'): 1, + ('b', 'd[1]'): 1, + ('c', 'd[2]'): 1, + }) + + def test_apply_to_single_assignments_static_unpack(self): + node = parser.parse_str('a, b, c = d, e, f') + node = node.body[0] + ast_util.apply_to_single_assignments(node.targets, node.value, + self._mock_apply_fn) + self.assertDictEqual(self._invocation_counts, { + ('a', 'd'): 1, + ('b', 'e'): 1, + ('c', 'f'): 1, + }) + if __name__ == '__main__': test.main() diff --git a/tensorflow/contrib/autograph/pyct/cfg.py b/tensorflow/contrib/autograph/pyct/cfg.py new file mode 100644 index 0000000000000000000000000000000000000000..8ef234745cce60595d3ab78aee8a223d5ca60376 --- /dev/null +++ b/tensorflow/contrib/autograph/pyct/cfg.py @@ -0,0 +1,812 @@ +# Copyright 2017 The TensorFlow Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================== +"""Control flow graph (CFG) structure for Python AST representation. + +The CFG is a digraph with edges representing valid control flow. Each +node is associated with exactly one AST node, but not all AST nodes may have +a corresponding CFG counterpart. + +Once built, the CFG itself is immutable, but the values it holds need not be; +they are usually annotated with information extracted by walking the graph. +""" + +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + +import collections +from enum import Enum + +# pylint:disable=g-bad-import-order +import gast +# pylint:enable=g-bad-import-order + +from tensorflow.contrib.autograph.pyct import compiler + + +class Node(object): + """A node in the CFG. + + Although new instances of this class are mutable, the objects that a user + finds in the CFG are typically not. + + The nodes represent edges in the CFG graph, and maintain pointers to allow + efficient walking in both forward and reverse order. The following property + holds for all nodes: "child in node.next" iff "node in child.prev". + + Attributes: + next: FrozenSet[Node, ...], the nodes that follow this node, in control + flow order + prev: FrozenSet[Node, ...], the nodes that precede this node, in reverse + control flow order + ast_node: ast.AST, the AST node corresponding to this CFG node + """ + + def __init__(self, next_, prev, ast_node): + self.next = next_ + self.prev = prev + self.ast_node = ast_node + + def freeze(self): + self.next = frozenset(self.next) + self.prev = frozenset(self.prev) + + def __repr__(self): + if isinstance(self.ast_node, gast.FunctionDef): + return 'def %s' % self.ast_node.name + elif isinstance(self.ast_node, gast.withitem): + return compiler.ast_to_source(self.ast_node.context_expr).strip() + return compiler.ast_to_source(self.ast_node).strip() + + +class Graph( + collections.namedtuple( + 'Graph', + ['entry', 'exit', 'error', 'index', 'stmt_prev', 'stmt_next'])): + """A Control Flow Graph. + + The CFG maintains an index to allow looking up a CFG node by the AST node to + which it is associated. The index can also be enumerated in top-down, depth + first order. + + Walking the graph in forward or reverse order is supported by double + parent-child links. + + Note: the error nodes are not wired to their corresponding finally guards, + because these are shared, and wiring them would create a reverse path from + normal control flow into the error nodes, which we want to avoid. + + The graph also maintains edges corresponding to higher level statements + like for-else loops. A node is considered successor of a statement if there + is an edge from a node that is lexically a child of that statement to a node + that is not. Statement predecessors are analogously defined. + + Attributes: + entry: Node, the entry node + exit: FrozenSet[Node, ...], the exit nodes + error: FrozenSet[Node, ...], nodes that exit due to an explicitly raised + error (errors propagated from function calls are not accounted) + index: Dict[ast.Node, Node], mapping AST nodes to the respective CFG + node + stmt_prev: Dict[ast.Node, FrozenSet[Node, ...]], mapping statement AST + nodes to their predecessor CFG nodes + stmt_next: Dict[ast.Node, FrozenSet[Node, ...]], mapping statement AST + nodes to their successor CFG nodes + """ + + def __repr__(self): + result = 'digraph CFG {\n' + for node in self.index.values(): + result += ' %s [label="%s"];\n' % (id(node), node) + for node in self.index.values(): + for next_ in node.next: + result += ' %s -> %s;\n' % (id(node), id(next_)) + result += '}' + return result + + +class _WalkMode(Enum): + FORWARD = 1 + REVERSE = 2 + + +class GraphVisitor(object): + """Base class for a CFG visitors. + + This implementation is not thread safe. + + The visitor has some facilities to simplify dataflow analyses. In particular, + it allows revisiting the nodes at the decision of the subclass. This can be + used to visit the graph until the state reaches a fixed point. + + For more details on dataflow analysis, see + https://www.seas.harvard.edu/courses/cs252/2011sp/slides/Lec02-Dataflow.pdf + + Note: the literature generally suggests visiting successor nodes only when the + state of the current node changed, regardless of whether that successor has + ever been visited. This implementation visits every successor at least once. + + Attributes: + graph: Graph + in_: Dict[Node, Any], stores node-keyed state during a visit + out: Dict[Node, Any], stores node-keyed state during a visit + """ + + def __init__(self, graph): + self.graph = graph + self.reset() + + def init_state(self, node): + """State initialization function. Optional to overload. + + An in/out state slot will be created for each node in the graph. Subclasses + must overload this to control what that is initialized to. + + Args: + node: Node + """ + raise NotImplementedError('Subclasses must implement this.') + + def visit_node(self, node): + """Visitor function. + + Args: + node: Node + Returns: + bool, whether the node should be revisited; subclasses can visit every + reachable node exactly once by always returning False + """ + raise NotImplementedError('Subclasses must implement this.') + + def reset(self): + self.in_ = { + node: self.init_state(node) for node in self.graph.index.values() + } + self.out = { + node: self.init_state(node) for node in self.graph.index.values() + } + + def _visit_internal(self, mode): + """Visits the CFG, depth-first.""" + assert mode in (_WalkMode.FORWARD, _WalkMode.REVERSE) + if mode == _WalkMode.FORWARD: + open_ = [self.graph.entry] + elif mode == _WalkMode.REVERSE: + open_ = list(self.graph.exit) + closed = set() + + while open_: + node = open_.pop(0) + closed.add(node) + + should_revisit = self.visit_node(node) + + if mode == _WalkMode.FORWARD: + children = node.next + elif mode == _WalkMode.REVERSE: + children = node.prev + + for next_ in children: + if should_revisit or next_ not in closed: + open_.append(next_) + + def visit_forward(self): + self._visit_internal(_WalkMode.FORWARD) + + def visit_reverse(self): + self._visit_internal(_WalkMode.REVERSE) + + +class GraphBuilder(object): + """Builder that constructs a CFG from a given AST. + + This GraphBuilder facilitates constructing the DAG that forms the CFG when + nodes + are supplied in lexical order (i.e., top-down, depth first). Under these + conditions, it supports building patterns found in typical structured + programs. + + This builder ignores the flow generated by exceptions, which are assumed to + always be catastrophic and present purely for diagnostic purposes (e.g. to + print debug information). Statements like raise and try/catch sections are + allowed and will generate control flow edges, but ordinaty statements are + assumed not to raise exceptions. + + Finally sections are also correctly interleaved between break/continue/return + nodes and their subsequent statements. + + Important concepts: + * nodes - nodes refer refer to CFG nodes; AST nodes are qualified explicitly + * leaf set - since the graph is constructed gradually, a leaf set maintains + the CFG nodes that will precede the node that the builder expects to + receive next; when an ordinary node is added, it is connected to the + existing leaves and it in turn becomes the new leaf + * jump nodes - nodes that should generate edges other than what + ordinary nodes would; these correspond to break, continue and return + statements + * sections - logical delimiters for subgraphs that require special + edges; there are various types of nodes, each admitting various + types of jump nodes; sections are identified by their corresponding AST + node + """ + + # TODO(mdan): Perhaps detail this in a markdown doc. + # TODO(mdan): Add exception support. + + def __init__(self, parent_ast_node): + self.reset() + self.parent = parent_ast_node + + def reset(self): + """Resets the state of this factory.""" + self.head = None + self.errors = set() + self.node_index = collections.OrderedDict() + + # TODO(mdan): Too many primitives. Use classes. + self.leaves = set() + + # Note: This mechanism requires that nodes are added in lexical order (top + # to bottom, depth first). + self.active_stmts = set() + self.owners = {} # type: Set[any] + self.forward_edges = set() # type: Tuple[Node, Node] # (from, to) + + self.finally_sections = {} + # Dict values represent (entry, exits) + self.finally_section_subgraphs = { + } # type: Dict[ast.AST, Tuple[Node, Set[Node]]] + # Whether the guard section can be reached from the statement that precedes + # it. + self.finally_section_has_direct_flow = {} + # Finally sections that await their first node. + self.pending_finally_sections = set() + + # Exit jumps keyed by the section they affect. + self.exits = {} + + # The entry of loop sections, keyed by the section. + self.section_entry = {} + # Continue jumps keyed by the section they affect. + self.continues = {} + + # The entry of conditional sections, keyed by the section. + self.cond_entry = {} + # Lists of leaf nodes corresponding to each branch in the section. + self.cond_leaves = {} + + def _connect_nodes(self, first, second): + """Connects nodes to signify that control flows from first to second. + + Args: + first: Union[Set[Node, ...], Node] + second: Node + """ + if isinstance(first, Node): + first.next.add(second) + second.prev.add(first) + self.forward_edges.add((first, second)) + else: + for node in first: + self._connect_nodes(node, second) + + def _add_new_node(self, ast_node): + """Grows the graph by adding a CFG node following the current leaves.""" + if ast_node is self.node_index: + raise ValueError('%s added twice' % ast_node) + node = Node(next_=set(), prev=set(), ast_node=ast_node) + self.node_index[ast_node] = node + self.owners[node] = frozenset(self.active_stmts) + + if self.head is None: + self.head = node + + for leaf in self.leaves: + self._connect_nodes(leaf, node) + + # If any finally section awaits its first node, populate it. + for section_id in self.pending_finally_sections: + self.finally_section_subgraphs[section_id][0] = node + self.pending_finally_sections = set() + + return node + + def begin_statement(self, stmt): + """Marks the beginning of a statement. + + Args: + stmt: Hashable, a key by which the statement can be identified in + the CFG's stmt_prev and stmt_next attributes + """ + self.active_stmts.add(stmt) + + def end_statement(self, stmt): + """Marks the end of a statement. + + Args: + stmt: Hashable, a key by which the statement can be identified in + the CFG's stmt_prev and stmt_next attributes; must match a key + previously passed to begin_statement. + """ + self.active_stmts.remove(stmt) + + def add_ordinary_node(self, ast_node): + """Grows the graph by adding an ordinary CFG node. + + Ordinary nodes are followed by the next node, in lexical order, that is, + they become the new leaf set. + + Args: + ast_node: ast.AST + Returns: + Node + """ + node = self._add_new_node(ast_node) + self.leaves = set((node,)) + return node + + def _add_jump_node(self, ast_node, guards): + """Grows the graph by adding a jump node. + + Jump nodes are added to the current leaf set, and the leaf set becomes + empty. If the jump node is the last in a cond section, then it may be added + back to the leaf set by a separate mechanism. + + Args: + ast_node: ast.AST + guards: Tuple[ast.AST, ...], the finally sections active for this node + Returns: + Node + """ + node = self._add_new_node(ast_node) + self.leaves = set() + # The guards themselves may not yet be complete, and will be wired later. + self.finally_sections[node] = guards + return node + + def _connect_jump_to_finally_sections(self, node): + """Connects a jump node to the finally sections protecting it.""" + cursor = set((node,)) + for guard_section_id in self.finally_sections[node]: + guard_begin, guard_ends = self.finally_section_subgraphs[guard_section_id] + self._connect_nodes(cursor, guard_begin) + cursor = guard_ends + del self.finally_sections[node] + # TODO(mdan): Should garbage-collect finally_section_subgraphs. + return cursor + + def add_exit_node(self, ast_node, section_id, guards): + """Grows the graph by adding an exit node. + + This node becomes an exit for the current section. + + Args: + ast_node: ast.AST + section_id: Hashable, the node for which ast_node should be considered + to be an exit node + guards: Tuple[ast.AST, ...], the finally sections that guard ast_node + """ + node = self._add_jump_node(ast_node, guards) + self.exits[section_id].add(node) + + def add_continue_node(self, ast_node, section_id, guards): + """Grows the graph by adding a reentry node. + + This node causes control flow to go back to the loop section's entry. + + Args: + ast_node: ast.AST + section_id: Hashable, the node for which ast_node should be considered + to be an exit node + guards: Tuple[ast.AST, ...], the finally sections that guard ast_node + """ + node = self._add_jump_node(ast_node, guards) + self.continues[section_id].add(node) + + def add_error_node(self, ast_node, guards): + """Grows the graph by adding an error node. + + This node becomes an exit for the entire graph. + + Args: + ast_node: ast.AST + guards: Tuple[ast.AST, ...], the finally sections that guard ast_node + """ + node = self._add_jump_node(ast_node, guards) + self.errors.add(node) + self.leaves = set() + + def enter_section(self, section_id): + """Enters a regular section. + + Regular sections admit exit jumps, which end the section. + + Args: + section_id: Hashable, the same node that will be used in calls to the + ast_node arg passed to add_exit_node + """ + assert section_id not in self.exits + self.exits[section_id] = set() + + def exit_section(self, section_id): + """Exits a regular section.""" + + # Exits are jump nodes, which may be protected. + for exit_ in self.exits[section_id]: + self.leaves |= self._connect_jump_to_finally_sections(exit_) + + del self.exits[section_id] + + def enter_loop_section(self, section_id, entry_node): + """Enters a loop section. + + Loop sections define an entry node. The end of the section always flows back + to the entry node. These admit continue jump nodes which also flow to the + entry node. + + Args: + section_id: Hashable, the same node that will be used in calls to the + ast_node arg passed to add_continue_node + entry_node: ast.AST, the entry node into the loop (e.g. the test node + for while loops) + """ + assert section_id not in self.section_entry + assert section_id not in self.continues + self.continues[section_id] = set() + node = self.add_ordinary_node(entry_node) + self.section_entry[section_id] = node + + def exit_loop_section(self, section_id): + """Exits a loop section.""" + self._connect_nodes(self.leaves, self.section_entry[section_id]) + + # continues are jump nodes, which may be protected. + for reentry in self.continues[section_id]: + guard_ends = self._connect_jump_to_finally_sections(reentry) + self._connect_nodes(guard_ends, self.section_entry[section_id]) + + # Loop nodes always loop back. + self.leaves = set((self.section_entry[section_id],)) + + del self.continues[section_id] + del self.section_entry[section_id] + + def enter_cond_section(self, section_id): + """Enters a conditional section. + + Conditional sections define an entry node, and one or more branches. + + Args: + section_id: Hashable, the same node that will be used in calls to the + section_id arg passed to new_cond_branch + """ + + assert section_id not in self.cond_entry + assert section_id not in self.cond_leaves + self.cond_leaves[section_id] = [] + + def new_cond_branch(self, section_id): + """Begins a new branch in a cond section.""" + assert section_id in self.cond_leaves + + if section_id in self.cond_entry: + # Subsequent splits move back to the split point, and memorize the + # current leaves. + self.cond_leaves[section_id].append(self.leaves) + self.leaves = self.cond_entry[section_id] + else: + # If this is the first time we split a section, just remember the split + # point. + self.cond_entry[section_id] = self.leaves + + def exit_cond_section(self, section_id): + """Exits a conditional section.""" + for split in self.cond_leaves[section_id]: + self.leaves |= split + del self.cond_entry[section_id] + del self.cond_leaves[section_id] + + def enter_finally_section(self, section_id): + """Enters a finally section.""" + # TODO(mdan): This, not the caller, should track the active sections. + self.finally_section_subgraphs[section_id] = [None, None] + if self.leaves: + self.finally_section_has_direct_flow[section_id] = True + else: + self.finally_section_has_direct_flow[section_id] = False + self.pending_finally_sections.add(section_id) + + def exit_finally_section(self, section_id): + """Exits a finally section.""" + assert section_id not in self.pending_finally_sections, 'Empty finally?' + self.finally_section_subgraphs[section_id][1] = self.leaves + # If the guard can only be reached by a jump, then it will not flow + # into the statement that follows it. + if not self.finally_section_has_direct_flow[section_id]: + self.leaves = set() + del self.finally_section_has_direct_flow[section_id] + + def build(self): + """Returns the CFG accumulated so far and resets the builder. + + Returns: + Graph + """ + # Freeze the nodes. + for node in self.node_index.values(): + node.freeze() + + # Build the statement edges. + stmt_next = {} + stmt_prev = {} + for node, _ in self.forward_edges: + for stmt in self.owners[node]: + if stmt not in stmt_next: + stmt_next[stmt] = set() + if stmt not in stmt_prev: + stmt_prev[stmt] = set() + for first, second in self.forward_edges: + stmts_exited = self.owners[first] - self.owners[second] + for stmt in stmts_exited: + stmt_next[stmt].add(second) + stmts_entered = self.owners[second] - self.owners[first] + for stmt in stmts_entered: + stmt_prev[stmt].add(first) + for stmt in stmt_next: + stmt_next[stmt] = frozenset(stmt_next[stmt]) + for stmt in stmt_prev: + stmt_prev[stmt] = frozenset(stmt_prev[stmt]) + + # Construct the final graph object. + result = Graph( + entry=self.head, + exit=self.leaves, + error=self.errors, + index=self.node_index, + stmt_prev=stmt_prev, + stmt_next=stmt_next) + + # Reset the state. + self.reset() + + return result + + +class AstToCfg(gast.NodeVisitor): + """Converts an AST to CFGs. + + A separate CFG will be constructed for each function. + """ + + def __init__(self): + super(AstToCfg, self).__init__() + + self.builder_stack = [] + self.builder = None + self.cfgs = {} + + self.lexical_scopes = [] + + def _enter_lexical_scope(self, node): + self.lexical_scopes.append(node) + + def _exit_lexical_scope(self, node): + leaving_node = self.lexical_scopes.pop() + assert node == leaving_node + + def _get_enclosing_scopes(self, include, stop_at): + included = [] + for node in reversed(self.lexical_scopes): + if isinstance(node, include): + included.append(node) + if isinstance(node, stop_at): + return node, included + return None, included + + def _process_basic_statement(self, node): + self.generic_visit(node) + self.builder.add_ordinary_node(node) + + def _process_exit_statement(self, node, *exits_nodes_of_type): + # Note: this is safe because we process functions separately. + try_node, guards = self._get_enclosing_scopes( + include=(gast.Try,), + stop_at=tuple(exits_nodes_of_type), + ) + if try_node is None: + raise ValueError( + '%s that is not enclosed by any of %s' % (node, exits_nodes_of_type)) + self.builder.add_exit_node(node, try_node, guards) + + def _process_continue_statement(self, node, *loops_to_nodes_of_type): + # Note: this is safe because we process functions separately. + try_node, guards = self._get_enclosing_scopes( + include=(gast.Try,), + stop_at=tuple(loops_to_nodes_of_type), + ) + if try_node is None: + raise ValueError('%s that is not enclosed by any of %s' % + (node, loops_to_nodes_of_type)) + self.builder.add_continue_node(node, try_node, guards) + + def visit_FunctionDef(self, node): + # We also keep the FunctionDef node in the CFG. This allows us to determine + # things like reaching definitions via closure. Note that the function body + # will be stored in a separate graph, because function definitions are not + # the same as function calls. + if self.builder is not None: + self.builder.add_ordinary_node(node) + + self.builder_stack.append(self.builder) + self.builder = GraphBuilder(node) + + self._enter_lexical_scope(node) + self.builder.enter_section(node) + + self._process_basic_statement(node.args) + for stmt in node.body: + self.visit(stmt) + + self.builder.exit_section(node) + self._exit_lexical_scope(node) + + self.cfgs[node] = self.builder.build() + self.builder = self.builder_stack.pop() + + def visit_Lambda(self, node): + # TODO(mdan): Treat like FunctionDef? That would be a separate CFG. + raise NotImplementedError() + + def visit_Return(self, node): + self._process_exit_statement(node, gast.FunctionDef) + + def visit_Expr(self, node): + self._process_basic_statement(node) + + def visit_Assign(self, node): + self._process_basic_statement(node) + + def visit_AnnAssign(self, node): + self._process_basic_statement(node) + + def visit_AugAssign(self, node): + self._process_basic_statement(node) + + def visit_Print(self, node): + self._process_basic_statement(node) + + def visit_Raise(self, node): + try_node, guards = self._get_enclosing_scopes( + include=(gast.Try,), + stop_at=(gast.FunctionDef,), + ) + if try_node is None: + raise ValueError('%s that is not enclosed by any FunctionDef' % node) + self.builder.add_error_node(node, try_node, guards) + + def visit_Assert(self, node): + # Ignoring the effect of exceptions. + self._process_basic_statement(node) + + def visit_Delete(self, node): + self._process_basic_statement(node) + + def visit_If(self, node): + # No need to track ifs as lexical scopes, for now. + # Lexical scopes are generally tracked in order to be able to resolve the + # targets of jump statements like break/continue/etc. Since there is no + # statement that can interrupt a conditional, we don't need to track their + # lexical scope. That may change in the future. + self.builder.begin_statement(node) + + self.builder.enter_cond_section(node) + self._process_basic_statement(node.test) + + self.builder.new_cond_branch(node) + for stmt in node.body: + self.visit(stmt) + + self.builder.new_cond_branch(node) + for stmt in node.orelse: + self.visit(stmt) + + self.builder.exit_cond_section(node) + self.builder.end_statement(node) + + def visit_While(self, node): + self.builder.begin_statement(node) + self._enter_lexical_scope(node) + + self.builder.enter_section(node) + + self.builder.enter_loop_section(node, node.test) + for stmt in node.body: + self.visit(stmt) + self.builder.exit_loop_section(node) + + # Note: although the orelse is technically part of the loop node, + # the statements inside it don't affect the loop itself. For example, a + # break in the loop's orelse will not affect the loop itself. + self._exit_lexical_scope(node) + + for stmt in node.orelse: + self.visit(stmt) + + self.builder.exit_section(node) + self.builder.end_statement(node) + + def visit_For(self, node): + self.builder.begin_statement(node) + self._enter_lexical_scope(node) + + self.builder.enter_section(node) + + # TODO(mdan): Strictly speaking, this should be node.target + node.iter. + # A blind dataflow analysis would have to process both node.target and + # node.iter to properly process read and write access. + self.builder.enter_loop_section(node, node.iter) + for stmt in node.body: + self.visit(stmt) + self.builder.exit_loop_section(node) + + # Note: although the orelse is technically part of the loop node, + # they don't count as loop bodies. For example, a break in the loop's + # orelse will affect the parent loop, not the current one. + self._exit_lexical_scope(node) + + for stmt in node.orelse: + self.visit(stmt) + + self.builder.exit_section(node) + self.builder.end_statement(node) + + def visit_Break(self, node): + self._process_exit_statement(node, gast.While, gast.For) + + def visit_Continue(self, node): + self._process_continue_statement(node, gast.While, gast.For) + + def visit_Try(self, node): + self._enter_lexical_scope(node) + + for stmt in node.body: + self.visit(stmt) + # Unlike loops, the orelse is a simple continuation of the body. + for stmt in node.orelse: + self.visit(stmt) + + if node.handlers: + # TODO(mdan): Should we still support bare try/except? Might be confusing. + raise NotImplementedError('exceptions are not yet supported') + + self._exit_lexical_scope(node) + + self.builder.enter_finally_section(node) + for stmt in node.finalbody: + self.visit(stmt) + self.builder.exit_finally_section(node) + + def visit_With(self, node): + # TODO(mdan): Mark the context manager's exit call as exit guard. + for item in node.items: + self._process_basic_statement(item) + for stmt in node.body: + self.visit(stmt) + + +def build(node): + visitor = AstToCfg() + visitor.visit(node) + return visitor.cfgs diff --git a/tensorflow/contrib/autograph/pyct/cfg_test.py b/tensorflow/contrib/autograph/pyct/cfg_test.py new file mode 100644 index 0000000000000000000000000000000000000000..9d0a85d615cc5a7dcebf405aebdbfe409be0b5cf --- /dev/null +++ b/tensorflow/contrib/autograph/pyct/cfg_test.py @@ -0,0 +1,969 @@ +# Copyright 2017 The TensorFlow Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================== +"""Tests for cfg module.""" + +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + +from tensorflow.contrib.autograph.pyct import cfg +from tensorflow.contrib.autograph.pyct import parser +from tensorflow.python.platform import test + + +class CountingVisitor(cfg.GraphVisitor): + + def __init__(self, graph): + super(CountingVisitor, self).__init__(graph) + self.counts = {} + + def init_state(self, _): + return None + + def visit_node(self, node): + self.counts[node.ast_node] = self.counts.get(node.ast_node, 0) + 1 + return False # visit only once + + +class GraphVisitorTest(test.TestCase): + + def _build_cfg(self, fn): + node, _ = parser.parse_entity(fn) + cfgs = cfg.build(node) + return cfgs, node + + def test_basic_coverage_forward(self): + + def test_fn(a): + while a > 0: + a = 1 + break + return a # pylint:disable=unreachable + a = 2 + + graphs, node = self._build_cfg(test_fn) + graph, = graphs.values() + visitor = CountingVisitor(graph) + visitor.visit_forward() + fn_node = node.body[0] + + self.assertEqual(visitor.counts[fn_node.args], 1) + self.assertEqual(visitor.counts[fn_node.body[0].test], 1) + self.assertEqual(visitor.counts[fn_node.body[0].body[0]], 1) + self.assertEqual(visitor.counts[fn_node.body[0].body[1]], 1) + # The return node should be unreachable in forward direction. + self.assertTrue(fn_node.body[0].body[2] not in visitor.counts) + self.assertEqual(visitor.counts[fn_node.body[1]], 1) + + def test_basic_coverage_reverse(self): + + def test_fn(a): + while a > 0: + a = 1 + break + return a # pylint:disable=unreachable + a = 2 + + graphs, node = self._build_cfg(test_fn) + graph, = graphs.values() + visitor = CountingVisitor(graph) + visitor.visit_reverse() + fn_node = node.body[0] + + self.assertEqual(visitor.counts[fn_node.args], 1) + self.assertEqual(visitor.counts[fn_node.body[0].test], 1) + self.assertEqual(visitor.counts[fn_node.body[0].body[0]], 1) + self.assertEqual(visitor.counts[fn_node.body[0].body[1]], 1) + self.assertTrue(visitor.counts[fn_node.body[0].body[2]], 1) + self.assertEqual(visitor.counts[fn_node.body[1]], 1) + + +class AstToCfgTest(test.TestCase): + + def _build_cfg(self, fn): + node, _ = parser.parse_entity(fn) + cfgs = cfg.build(node) + return cfgs + + def _repr_set(self, node_set): + return frozenset(repr(n) for n in node_set) + + def _as_set(self, elements): + if elements is None: + return frozenset() + elif isinstance(elements, str): + return frozenset((elements,)) + else: + return frozenset(elements) + + def assertGraphMatches(self, graph, edges): + """Tests whether the CFG contains the specified edges.""" + for prev, node_repr, next_ in edges: + matched = False + for cfg_node in graph.index.values(): + if repr(cfg_node) == node_repr: + if (self._as_set(prev) == frozenset(map(repr, cfg_node.prev)) and + self._as_set(next_) == frozenset(map(repr, cfg_node.next))): + matched = True + break + if not matched: + self.fail( + 'match failed for node "%s" in graph:\n%s' % (node_repr, graph)) + + def assertStatementEdges(self, graph, edges): + """Tests whether the CFG contains the specified statement edges.""" + for prev_node_reprs, node_repr, next_node_reprs in edges: + matched = False + partial_matches = [] + self.assertSetEqual( + frozenset(graph.stmt_next.keys()), frozenset(graph.stmt_prev.keys())) + for stmt_ast_node in graph.stmt_next: + ast_repr = '%s:%s' % (stmt_ast_node.__class__.__name__, + stmt_ast_node.lineno) + if ast_repr == node_repr: + actual_next = frozenset(map(repr, graph.stmt_next[stmt_ast_node])) + actual_prev = frozenset(map(repr, graph.stmt_prev[stmt_ast_node])) + partial_matches.append((actual_prev, node_repr, actual_next)) + if (self._as_set(prev_node_reprs) == actual_prev and + self._as_set(next_node_reprs) == actual_next): + matched = True + break + if not matched: + self.fail('edges mismatch for %s: %s' % (node_repr, partial_matches)) + + def test_straightline(self): + + def test_fn(a): + a += 1 + a = 2 + a = 3 + return + + graph, = self._build_cfg(test_fn).values() + + self.assertGraphMatches( + graph, + ( + (None, 'a', 'a += 1'), + ('a += 1', 'a = 2', 'a = 3'), + ('a = 2', 'a = 3', 'return'), + ('a = 3', 'return', None), + ), + ) + + def test_straightline_no_return(self): + + def test_fn(a, b): + a = b + 1 + a += max(a) + + graph, = self._build_cfg(test_fn).values() + + self.assertGraphMatches( + graph, + ( + (None, 'a, b', 'a = b + 1'), + ('a = b + 1', 'a += max(a)', None), + ), + ) + + def test_unreachable_code(self): + + def test_fn(a): + return + a += 1 # pylint:disable=unreachable + + graph, = self._build_cfg(test_fn).values() + + self.assertGraphMatches( + graph, + ( + (None, 'a', 'return'), + ('a', 'return', None), + (None, 'a += 1', None), + ), + ) + + def test_if_straightline(self): + + def test_fn(a): + if a > 0: + a = 1 + else: + a += -1 + + graph, = self._build_cfg(test_fn).values() + + self.assertGraphMatches( + graph, + ( + (None, 'a', '(a > 0)'), + ('(a > 0)', 'a = 1', None), + ('(a > 0)', 'a += -1', None), + ), + ) + self.assertStatementEdges( + graph, + (('a', 'If:2', None),), + ) + + def test_branch_nested(self): + + def test_fn(a): + if a > 0: + if a > 1: + a = 1 + else: + a = 2 + else: + if a > 2: + a = 3 + else: + a = 4 + + graph, = self._build_cfg(test_fn).values() + + self.assertGraphMatches( + graph, + ( + (None, 'a', '(a > 0)'), + ('a', '(a > 0)', ('(a > 1)', '(a > 2)')), + ('(a > 0)', '(a > 1)', ('a = 1', 'a = 2')), + ('(a > 1)', 'a = 1', None), + ('(a > 1)', 'a = 2', None), + ('(a > 0)', '(a > 2)', ('a = 3', 'a = 4')), + ('(a > 2)', 'a = 3', None), + ('(a > 2)', 'a = 4', None), + ), + ) + self.assertStatementEdges( + graph, + ( + ('a', 'If:2', None), + ('(a > 0)', 'If:3', None), + ('(a > 0)', 'If:8', None), + ), + ) + + def test_branch_straightline_semi(self): + + def test_fn(a): + if a > 0: + a = 1 + + graph, = self._build_cfg(test_fn).values() + + self.assertGraphMatches( + graph, + ( + (None, 'a', '(a > 0)'), + ('a', '(a > 0)', 'a = 1'), + ('(a > 0)', 'a = 1', None), + ), + ) + self.assertStatementEdges( + graph, + (('a', 'If:2', None),), + ) + + def test_branch_return(self): + + def test_fn(a): + if a > 0: + return + else: + a = 1 + a = 2 + + graph, = self._build_cfg(test_fn).values() + + self.assertGraphMatches( + graph, + ( + ('a', '(a > 0)', ('return', 'a = 1')), + ('(a > 0)', 'a = 1', 'a = 2'), + ('(a > 0)', 'return', None), + ('a = 1', 'a = 2', None), + ), + ) + self.assertStatementEdges( + graph, + (('a', 'If:2', 'a = 2'),), + ) + + def test_branch_return_minimal(self): + + def test_fn(a): + if a > 0: + return + + graph, = self._build_cfg(test_fn).values() + + self.assertGraphMatches( + graph, + ( + ('a', '(a > 0)', 'return'), + ('(a > 0)', 'return', None), + ), + ) + self.assertStatementEdges( + graph, + (('a', 'If:2', None),), + ) + + def test_while_straightline(self): + + def test_fn(a): + while a > 0: + a = 1 + a = 2 + + graph, = self._build_cfg(test_fn).values() + + self.assertGraphMatches( + graph, + ( + (('a', 'a = 1'), '(a > 0)', ('a = 1', 'a = 2')), + ('(a > 0)', 'a = 1', '(a > 0)'), + ('(a > 0)', 'a = 2', None), + ), + ) + self.assertStatementEdges( + graph, + (('a', 'While:2', 'a = 2'),), + ) + + def test_while_else_straightline(self): + + def test_fn(a): + while a > 0: + a = 1 + else: # pylint:disable=useless-else-on-loop + a = 2 + a = 3 + + graph, = self._build_cfg(test_fn).values() + + self.assertGraphMatches( + graph, + ( + (('a', 'a = 1'), '(a > 0)', ('a = 1', 'a = 2')), + ('(a > 0)', 'a = 1', '(a > 0)'), + ('(a > 0)', 'a = 2', 'a = 3'), + ('a = 2', 'a = 3', None), + ), + ) + self.assertStatementEdges( + graph, + (('a', 'While:2', 'a = 3'),), + ) + + def test_while_else_continue(self): + + def test_fn(a): + while a > 0: + if a > 1: + continue + else: + a = 0 + a = 1 + else: # pylint:disable=useless-else-on-loop + a = 2 + a = 3 + + graph, = self._build_cfg(test_fn).values() + + self.assertGraphMatches( + graph, + ( + (('a', 'continue', 'a = 1'), '(a > 0)', ('(a > 1)', 'a = 2')), + ('(a > 0)', '(a > 1)', ('continue', 'a = 0')), + ('(a > 1)', 'continue', '(a > 0)'), + ('a = 0', 'a = 1', '(a > 0)'), + ('(a > 0)', 'a = 2', 'a = 3'), + ('a = 2', 'a = 3', None), + ), + ) + self.assertStatementEdges( + graph, + ( + ('a', 'While:2', 'a = 3'), + ('(a > 0)', 'If:3', ('a = 1', '(a > 0)')), + ), + ) + + def test_while_else_break(self): + + def test_fn(a): + while a > 0: + if a > 1: + break + a = 1 + else: + a = 2 + a = 3 + + graph, = self._build_cfg(test_fn).values() + + self.assertGraphMatches( + graph, + ( + (('a', 'a = 1'), '(a > 0)', ('(a > 1)', 'a = 2')), + ('(a > 0)', '(a > 1)', ('break', 'a = 1')), + ('(a > 1)', 'break', 'a = 3'), + ('(a > 1)', 'a = 1', '(a > 0)'), + ('(a > 0)', 'a = 2', 'a = 3'), + (('break', 'a = 2'), 'a = 3', None), + ), + ) + self.assertStatementEdges( + graph, + ( + ('a', 'While:2', 'a = 3'), + ('(a > 0)', 'If:3', ('a = 1', 'a = 3')), + ), + ) + + def test_while_else_return(self): + + def test_fn(a): + while a > 0: + if a > 1: + return + a = 1 + else: # pylint:disable=useless-else-on-loop + a = 2 + a = 3 + + graph, = self._build_cfg(test_fn).values() + + self.assertGraphMatches( + graph, + ( + (('a', 'a = 1'), '(a > 0)', ('(a > 1)', 'a = 2')), + ('(a > 0)', '(a > 1)', ('return', 'a = 1')), + ('(a > 1)', 'return', None), + ('(a > 1)', 'a = 1', '(a > 0)'), + ('(a > 0)', 'a = 2', 'a = 3'), + ('a = 2', 'a = 3', None), + ), + ) + self.assertStatementEdges( + graph, + ( + ('a', 'While:2', 'a = 3'), + ('(a > 0)', 'If:3', 'a = 1'), + ), + ) + + def test_while_nested_straightline(self): + + def test_fn(a): + while a > 0: + while a > 1: + a = 1 + a = 2 + a = 3 + + graph, = self._build_cfg(test_fn).values() + + self.assertGraphMatches( + graph, + ( + (('a', 'a = 2'), '(a > 0)', ('(a > 1)', 'a = 3')), + (('(a > 0)', 'a = 1'), '(a > 1)', ('a = 1', 'a = 2')), + ('(a > 1)', 'a = 1', '(a > 1)'), + ('(a > 1)', 'a = 2', '(a > 0)'), + ('(a > 0)', 'a = 3', None), + ), + ) + self.assertStatementEdges( + graph, + ( + ('a', 'While:2', 'a = 3'), + ('(a > 0)', 'While:3', 'a = 2'), + ), + ) + + def test_while_nested_continue(self): + + def test_fn(a): + while a > 0: + while a > 1: + if a > 3: + continue + a = 1 + a = 2 + a = 3 + + graph, = self._build_cfg(test_fn).values() + + self.assertGraphMatches( + graph, + ( + (('a', 'a = 2'), '(a > 0)', ('(a > 1)', 'a = 3')), + (('(a > 0)', 'continue', 'a = 1'), '(a > 1)', ('(a > 3)', 'a = 2')), + ('(a > 1)', '(a > 3)', ('continue', 'a = 1')), + ('(a > 3)', 'continue', '(a > 1)'), + ('(a > 3)', 'a = 1', '(a > 1)'), + ('(a > 1)', 'a = 2', '(a > 0)'), + ('(a > 0)', 'a = 3', None), + ), + ) + self.assertStatementEdges( + graph, + ( + ('a', 'While:2', 'a = 3'), + ('(a > 0)', 'While:3', 'a = 2'), + ('(a > 1)', 'If:4', ('a = 1', '(a > 1)')), + ), + ) + + def test_while_nested_break(self): + + def test_fn(a): + while a > 0: + while a > 1: + if a > 2: + break + a = 1 + a = 2 + a = 3 + + graph, = self._build_cfg(test_fn).values() + + self.assertGraphMatches(graph, ( + (('a', 'a = 2'), '(a > 0)', ('(a > 1)', 'a = 3')), + (('(a > 0)', 'a = 1'), '(a > 1)', ('(a > 2)', 'a = 2')), + ('(a > 1)', '(a > 2)', ('break', 'a = 1')), + ('(a > 2)', 'break', 'a = 2'), + ('(a > 2)', 'a = 1', '(a > 1)'), + (('(a > 1)', 'break'), 'a = 2', '(a > 0)'), + ('(a > 0)', 'a = 3', None), + )) + self.assertStatementEdges( + graph, + ( + ('a', 'While:2', 'a = 3'), + ('(a > 0)', 'While:3', 'a = 2'), + ('(a > 1)', 'If:4', ('a = 1', 'a = 2')), + ), + ) + + def test_for_straightline(self): + + def test_fn(a): + for a in range(0, a): + a = 1 + a = 2 + + graph, = self._build_cfg(test_fn).values() + + self.assertGraphMatches( + graph, + ( + (('a', 'a = 1'), 'range(0, a)', ('a = 1', 'a = 2')), + ('range(0, a)', 'a = 1', 'range(0, a)'), + ('range(0, a)', 'a = 2', None), + ), + ) + self.assertStatementEdges( + graph, + (('a', 'For:2', 'a = 2'),), + ) + + def test_for_else_straightline(self): + + def test_fn(a): + for a in range(0, a): + a = 1 + else: # pylint:disable=useless-else-on-loop + a = 2 + a = 3 + + graph, = self._build_cfg(test_fn).values() + + self.assertGraphMatches( + graph, + ( + (('a', 'a = 1'), 'range(0, a)', ('a = 1', 'a = 2')), + ('range(0, a)', 'a = 1', 'range(0, a)'), + ('range(0, a)', 'a = 2', 'a = 3'), + ('a = 2', 'a = 3', None), + ), + ) + self.assertStatementEdges( + graph, + (('a', 'For:2', 'a = 3'),), + ) + + def test_for_else_continue(self): + + def test_fn(a): + for a in range(0, a): + if a > 1: + continue + else: + a = 0 + a = 1 + else: # pylint:disable=useless-else-on-loop + a = 2 + a = 3 + + graph, = self._build_cfg(test_fn).values() + + self.assertGraphMatches( + graph, + ( + (('a', 'continue', 'a = 1'), 'range(0, a)', ('(a > 1)', 'a = 2')), + ('range(0, a)', '(a > 1)', ('continue', 'a = 0')), + ('(a > 1)', 'continue', 'range(0, a)'), + ('(a > 1)', 'a = 0', 'a = 1'), + ('a = 0', 'a = 1', 'range(0, a)'), + ('range(0, a)', 'a = 2', 'a = 3'), + ('a = 2', 'a = 3', None), + ), + ) + self.assertStatementEdges( + graph, + ( + ('a', 'For:2', 'a = 3'), + ('range(0, a)', 'If:3', ('a = 1', 'range(0, a)')), + ), + ) + + def test_for_else_break(self): + + def test_fn(a): + for a in range(0, a): + if a > 1: + break + a = 1 + else: + a = 2 + a = 3 + + graph, = self._build_cfg(test_fn).values() + + self.assertGraphMatches( + graph, + ( + (('a', 'a = 1'), 'range(0, a)', ('(a > 1)', 'a = 2')), + ('range(0, a)', '(a > 1)', ('break', 'a = 1')), + ('(a > 1)', 'break', 'a = 3'), + ('(a > 1)', 'a = 1', 'range(0, a)'), + ('range(0, a)', 'a = 2', 'a = 3'), + (('break', 'a = 2'), 'a = 3', None), + ), + ) + self.assertStatementEdges( + graph, + ( + ('a', 'For:2', 'a = 3'), + ('range(0, a)', 'If:3', ('a = 1', 'a = 3')), + ), + ) + + def test_for_else_return(self): + + def test_fn(a): + for a in range(0, a): + if a > 1: + return + a = 1 + else: # pylint:disable=useless-else-on-loop + a = 2 + a = 3 + + graph, = self._build_cfg(test_fn).values() + + self.assertGraphMatches( + graph, + ( + (('a', 'a = 1'), 'range(0, a)', ('(a > 1)', 'a = 2')), + ('range(0, a)', '(a > 1)', ('return', 'a = 1')), + ('(a > 1)', 'return', None), + ('(a > 1)', 'a = 1', 'range(0, a)'), + ('range(0, a)', 'a = 2', 'a = 3'), + ('a = 2', 'a = 3', None), + ), + ) + self.assertStatementEdges( + graph, + ( + ('a', 'For:2', 'a = 3'), + ('range(0, a)', 'If:3', 'a = 1'), + ), + ) + + def test_for_nested_straightline(self): + + def test_fn(a): + for a in range(0, a): + for b in range(1, a): + b += 1 + a = 2 + a = 3 + + graph, = self._build_cfg(test_fn).values() + + self.assertGraphMatches( + graph, + ( + (('a', 'a = 2'), 'range(0, a)', ('range(1, a)', 'a = 3')), + (('range(0, a)', 'b += 1'), 'range(1, a)', ('b += 1', 'a = 2')), + ('range(1, a)', 'b += 1', 'range(1, a)'), + ('range(1, a)', 'a = 2', 'range(0, a)'), + ('range(0, a)', 'a = 3', None), + ), + ) + self.assertStatementEdges( + graph, + ( + ('a', 'For:2', 'a = 3'), + ('range(0, a)', 'For:3', 'a = 2'), + ), + ) + + def test_for_nested_continue(self): + + def test_fn(a): + for a in range(0, a): + for b in range(1, a): + if a > 3: + continue + b += 1 + a = 2 + a = 3 + + graph, = self._build_cfg(test_fn).values() + + self.assertGraphMatches( + graph, + ( + (('a', 'a = 2'), 'range(0, a)', ('range(1, a)', 'a = 3')), + (('range(0, a)', 'continue', 'b += 1'), 'range(1, a)', + ('(a > 3)', 'a = 2')), + ('range(1, a)', '(a > 3)', ('continue', 'b += 1')), + ('(a > 3)', 'continue', 'range(1, a)'), + ('(a > 3)', 'b += 1', 'range(1, a)'), + ('range(1, a)', 'a = 2', 'range(0, a)'), + ('range(0, a)', 'a = 3', None), + ), + ) + self.assertStatementEdges( + graph, + ( + ('a', 'For:2', 'a = 3'), + ('range(0, a)', 'For:3', 'a = 2'), + ('range(1, a)', 'If:4', ('b += 1', 'range(1, a)')), + ), + ) + + def test_for_nested_break(self): + + def test_fn(a): + for a in range(0, a): + for b in range(1, a): + if a > 2: + break + b += 1 + a = 2 + a = 3 + + graph, = self._build_cfg(test_fn).values() + + self.assertGraphMatches( + graph, + ( + (('a', 'a = 2'), 'range(0, a)', ('range(1, a)', 'a = 3')), + (('range(0, a)', 'b += 1'), 'range(1, a)', ('(a > 2)', 'a = 2')), + ('range(1, a)', '(a > 2)', ('break', 'b += 1')), + ('(a > 2)', 'break', 'a = 2'), + ('(a > 2)', 'b += 1', 'range(1, a)'), + (('range(1, a)', 'break'), 'a = 2', 'range(0, a)'), + ('range(0, a)', 'a = 3', None), + ), + ) + self.assertStatementEdges( + graph, + ( + ('a', 'For:2', 'a = 3'), + ('range(0, a)', 'For:3', 'a = 2'), + ('range(1, a)', 'If:4', ('b += 1', 'a = 2')), + ), + ) + + def test_complex(self): + + def test_fn(a): + b = 0 + while a > 0: + for b in range(0, a): + if a > 2: + break + if a > 3: + if a > 4: + continue + else: + max(a) + break + b += 1 + else: # for b in range(0, a): + return a + a = 2 + for a in range(1, a): + return b + a = 3 + + graph, = self._build_cfg(test_fn).values() + + self.assertGraphMatches( + graph, + ( + (('b = 0', 'a = 2'), '(a > 0)', ('range(0, a)', 'range(1, a)')), + ( + ('(a > 0)', 'continue', 'b += 1'), + 'range(0, a)', + ('(a > 2)', 'return a'), + ), + ('range(0, a)', '(a > 2)', ('(a > 3)', 'break')), + ('(a > 2)', 'break', 'a = 2'), + ('(a > 2)', '(a > 3)', ('(a > 4)', 'b += 1')), + ('(a > 3)', '(a > 4)', ('continue', 'max(a)')), + ('(a > 4)', 'max(a)', 'break'), + ('max(a)', 'break', 'a = 2'), + ('(a > 4)', 'continue', 'range(0, a)'), + ('(a > 3)', 'b += 1', 'range(0, a)'), + ('range(0, a)', 'return a', None), + ('break', 'a = 2', '(a > 0)'), + ('(a > 0)', 'range(1, a)', ('return b', 'a = 3')), + ('range(1, a)', 'return b', None), + ('range(1, a)', 'a = 3', None), + ), + ) + self.assertStatementEdges( + graph, + ( + ('b = 0', 'While:3', 'range(1, a)'), + ('(a > 0)', 'For:4', 'a = 2'), + ('range(0, a)', 'If:5', ('(a > 3)', 'a = 2')), + ('(a > 2)', 'If:7', ('b += 1', 'a = 2', 'range(0, a)')), + ('(a > 3)', 'If:8', ('a = 2', 'range(0, a)')), + ('(a > 0)', 'For:17', 'a = 3'), + ), + ) + + def test_finally_straightline(self): + + def test_fn(a): + try: + a += 1 + finally: + a = 2 + a = 3 + + graph, = self._build_cfg(test_fn).values() + + self.assertGraphMatches( + graph, + ( + ('a', 'a += 1', 'a = 2'), + ('a += 1', 'a = 2', 'a = 3'), + ('a = 2', 'a = 3', None), + ), + ) + + def test_return_finally(self): + + def test_fn(a): + try: + return a + finally: + a = 1 + a = 2 + + graph, = self._build_cfg(test_fn).values() + + self.assertGraphMatches( + graph, + ( + ('a', 'return a', 'a = 1'), + ('return a', 'a = 1', None), + (None, 'a = 2', None), + ), + ) + + def test_break_finally(self): + + def test_fn(a): + while a > 0: + try: + break + finally: + a = 1 + + graph, = self._build_cfg(test_fn).values() + + self.assertGraphMatches( + graph, + ( + ('a', '(a > 0)', 'break'), + ('(a > 0)', 'break', 'a = 1'), + ('break', 'a = 1', None), + ), + ) + + def test_continue_finally(self): + + def test_fn(a): + while a > 0: + try: + continue + finally: + a = 1 + + graph, = self._build_cfg(test_fn).values() + + self.assertGraphMatches( + graph, + ( + (('a', 'a = 1'), '(a > 0)', 'continue'), + ('(a > 0)', 'continue', 'a = 1'), + ('continue', 'a = 1', '(a > 0)'), + ), + ) + + def test_with_straightline(self): + + def test_fn(a): + with max(a) as b: + a = 0 + return b + + graph, = self._build_cfg(test_fn).values() + + self.assertGraphMatches( + graph, + ( + ('a', 'max(a)', 'a = 0'), + ('max(a)', 'a = 0', 'return b'), + ('a = 0', 'return b', None), + ), + ) + + +if __name__ == '__main__': + test.main() diff --git a/tensorflow/contrib/autograph/pyct/origin_info.py b/tensorflow/contrib/autograph/pyct/origin_info.py new file mode 100644 index 0000000000000000000000000000000000000000..b3c6a43d37fce87fc2e359a0b053e8f5dc608713 --- /dev/null +++ b/tensorflow/contrib/autograph/pyct/origin_info.py @@ -0,0 +1,52 @@ +# Copyright 2018 The TensorFlow Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================== +"""Container for origin source code information before AutoGraph compilation.""" +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + +from collections import namedtuple + + +class CodeLocation(namedtuple('CodeLocation', ('file_path', 'line_number'))): + """Location of a line of code. + + Attributes: + file_path: text, the full path to the file containing the code. + line_number: Int, the 1-based line number of the code in its file. + """ + pass + + +class OriginInfo( + namedtuple('OriginInfo', ('file_path', 'function_name', 'line_number', + 'column_offset', 'source_code_line'))): + """Container for information about the source code before conversion. + + Instances of this class contain information about the source code that + transformed code originated from. Examples include: + * line number + * file name + * original user code + """ + + def as_frame(self): + """Makes a traceback frame tuple. + + Returns: + A tuple of (file_path, line_number, function_name, source_code_line). + """ + return (self.file_path, self.line_number, self.function_name, + self.source_code_line) diff --git a/tensorflow/contrib/autograph/pyct/qual_names.py b/tensorflow/contrib/autograph/pyct/qual_names.py index da07013cf4f4309c0e24adda3017575d942861b7..fb81404edc1994309f5108fc7e7ba368a1ea3ccb 100644 --- a/tensorflow/contrib/autograph/pyct/qual_names.py +++ b/tensorflow/contrib/autograph/pyct/qual_names.py @@ -30,6 +30,7 @@ import collections import gast from tensorflow.contrib.autograph.pyct import anno +from tensorflow.contrib.autograph.pyct import parser class Symbol(collections.namedtuple('Symbol', ['name'])): @@ -89,7 +90,8 @@ class QN(object): if not isinstance(base, (str, StringLiteral, NumberLiteral)): # TODO(mdan): Require Symbol instead of string. raise ValueError( - 'For simple QNs, base must be a string or a Literal object.') + 'for simple QNs, base must be a string or a Literal object;' + ' got instead "%s"' % type(base)) assert '.' not in base and '[' not in base and ']' not in base self._parent = None self.qn = (base,) @@ -112,6 +114,22 @@ class QN(object): raise ValueError('Cannot get parent of simple name "%s".' % self.qn[0]) return self._parent + @property + def owner_set(self): + """Returns all the symbols (simple or composite) that own this QN. + + In other words, if this symbol was modified, the symbols in the owner set + may also be affected. + + Examples: + 'a.b[c.d]' has two owners, 'a' and 'a.b' + """ + owners = set() + if self.has_attr() or self.has_subscript(): + owners.add(self.parent) + owners.update(self.parent.owner_set) + return owners + @property def support_set(self): """Returns the set of simple symbols that this QN relies on. @@ -122,7 +140,7 @@ class QN(object): Examples: 'a.b' has only one support symbol, 'a' - 'a[i]' has two roots, 'a' and 'i' + 'a[i]' has two support symbols, 'a' and 'i' """ # TODO(mdan): This might be the set of Name nodes in the AST. Track those? roots = set() @@ -231,3 +249,9 @@ class QnResolver(gast.NodeTransformer): def resolve(node): return QnResolver().visit(node) + + +def from_str(qn_str): + node = parser.parse_expression(qn_str) + node = resolve(node) + return anno.getanno(node, anno.Basic.QN) diff --git a/tensorflow/contrib/autograph/pyct/qual_names_test.py b/tensorflow/contrib/autograph/pyct/qual_names_test.py index 264afd508cdb847315c486806b531dc1483ef622..c793c2bb39df19f1af9b74f33323dbd4c985ee0d 100644 --- a/tensorflow/contrib/autograph/pyct/qual_names_test.py +++ b/tensorflow/contrib/autograph/pyct/qual_names_test.py @@ -30,6 +30,15 @@ from tensorflow.python.platform import test class QNTest(test.TestCase): + def test_from_str(self): + a = QN('a') + b = QN('b') + a_dot_b = QN(a, attr='b') + a_sub_b = QN(a, subscript=b) + self.assertEqual(qual_names.from_str('a.b'), a_dot_b) + self.assertEqual(qual_names.from_str('a'), a) + self.assertEqual(qual_names.from_str('a[b]'), a_sub_b) + def test_basic(self): a = QN('a') self.assertEqual(a.qn, ('a',)) diff --git a/tensorflow/contrib/autograph/pyct/static_analysis/BUILD b/tensorflow/contrib/autograph/pyct/static_analysis/BUILD index bcf2dacec2062704805f1d72ec27a243159d13c1..25f78536e08c70080247a28586790f70b520c8a9 100644 --- a/tensorflow/contrib/autograph/pyct/static_analysis/BUILD +++ b/tensorflow/contrib/autograph/pyct/static_analysis/BUILD @@ -19,8 +19,10 @@ py_library( srcs = [ "activity.py", "annos.py", - "cfg.py", + "cfg.py", # TODO(mdan): Remove. "live_values.py", + "liveness.py", + "reaching_definitions.py", "type_info.py", ], srcs_version = "PY2AND3", @@ -28,6 +30,7 @@ py_library( deps = [ "//tensorflow/contrib/autograph/pyct", "//tensorflow/contrib/autograph/utils", + "//tensorflow/python:util", "@gast_archive//:gast", ], ) @@ -70,6 +73,37 @@ py_test( ], ) +# TODO(mdan): Enable these tests once child change is in. +py_test( + name = "liveness_test", + srcs = ["liveness_test.py"], + srcs_version = "PY2AND3", + tags = [ + "manual", + "notap", + ], + deps = [ + ":static_analysis", + "//tensorflow/contrib/autograph/pyct", + "//tensorflow/python:client_testlib", + ], +) + +py_test( + name = "reaching_definitions_test", + srcs = ["reaching_definitions_test.py"], + srcs_version = "PY2AND3", + tags = [ + "manual", + "notap", + ], + deps = [ + ":static_analysis", + "//tensorflow/contrib/autograph/pyct", + "//tensorflow/python:client_testlib", + ], +) + py_test( name = "type_info_test", srcs = ["type_info_test.py"], diff --git a/tensorflow/contrib/autograph/pyct/static_analysis/__init__.py b/tensorflow/contrib/autograph/pyct/static_analysis/__init__.py index c325e19f28376da3be6db4b00b9f664eac047af2..9a82de735dc663f6a824488e4c5864943cecc3d4 100644 --- a/tensorflow/contrib/autograph/pyct/static_analysis/__init__.py +++ b/tensorflow/contrib/autograph/pyct/static_analysis/__init__.py @@ -18,10 +18,14 @@ This module contains utilities to help annotate AST nodes with as much runtime information as can be possibly extracted without actually executing the code, under that assumption that the context in which the code will run is known. -Note: It's a fair bet that this analysis cannot be reused across contexts -without re-running it. In most cases, the context usually means referenced -modules, which should be static enough to allow reuse, but that is not being -reliably verified. +Overall, the different analyses have the functions listed below: + + * activity: inventories symbols read, written to, params, etc. at different + levels + * liveness, reaching_definitions: dataflow analyses based on the program's CFG + and using the symbol information gathered by activity analysis + * live_values, type_info: type and value inference based on dataflow + analysis and context information """ from __future__ import absolute_import diff --git a/tensorflow/contrib/autograph/pyct/static_analysis/annos.py b/tensorflow/contrib/autograph/pyct/static_analysis/annos.py index b929b35b79200b0968c9c4f26b10cda28763773a..5eefecf278992f73464817585a3498de4c031978 100644 --- a/tensorflow/contrib/autograph/pyct/static_analysis/annos.py +++ b/tensorflow/contrib/autograph/pyct/static_analysis/annos.py @@ -21,6 +21,9 @@ from __future__ import print_function from enum import Enum +# TODO(mdan): Remove. + + class NoValue(Enum): def __repr__(self): @@ -50,10 +53,3 @@ class NodeAnno(NoValue): ORELSE_SCOPE = ( 'The scope for the orelse body of a statement (False branch for if ' 'statements, orelse body for loops).') - - # Type and Value annotations - # Type annotations are represented by objects of type type_info.Type. - STATIC_INFO = ( - 'The type or value information that should be asserted about the entity ' - 'referenced by the symbol holding this annotation, irrespective of the ' - 'execution context.') diff --git a/tensorflow/contrib/autograph/pyct/static_analysis/cfg.py b/tensorflow/contrib/autograph/pyct/static_analysis/cfg.py index 39eca6e44441cc28e565d383759cc796d57d6438..4acc4ed66a62b0ccd407d39b1abda00c4c88a9a1 100644 --- a/tensorflow/contrib/autograph/pyct/static_analysis/cfg.py +++ b/tensorflow/contrib/autograph/pyct/static_analysis/cfg.py @@ -286,7 +286,7 @@ class Forward(object): # TODO(alexbw): see if we can simplify by visiting breadth-first def visit(self, node): - """Depth-first walking the CFG, applying dataflow information propagation.""" + """Depth-first walking the CFG, applying dataflow info propagation.""" # node.value is None only for the exit CfgNode. if not node.value: return diff --git a/tensorflow/contrib/autograph/pyct/static_analysis/liveness.py b/tensorflow/contrib/autograph/pyct/static_analysis/liveness.py new file mode 100644 index 0000000000000000000000000000000000000000..bf29d868a2e4d2a4c7dd1057c0ed93e54d01d750 --- /dev/null +++ b/tensorflow/contrib/autograph/pyct/static_analysis/liveness.py @@ -0,0 +1,200 @@ +# Copyright 2018 The TensorFlow Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================== +"""Live variable analysis. + +This analysis attaches a set containing the live symbols that are live at the +exit of control flow statements. + +Requires activity analysis. +""" + +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + +import gast + +from tensorflow.contrib.autograph.pyct import anno +from tensorflow.contrib.autograph.pyct import cfg +from tensorflow.contrib.autograph.pyct import transformer +from tensorflow.contrib.autograph.pyct.static_analysis import annos + + +class Analyzer(cfg.GraphVisitor): + """CFG visitor that performs liveness analysis at statement level.""" + + def __init__(self, graph): + super(Analyzer, self).__init__(graph) + # This allows communicating that nodes generate extra symbols, + # e.g. those that a function definition closes over. + self.extra_gen = {} + + def init_state(self, _): + return set() + + def visit_node(self, node): + prev_live_in = self.in_[node] + + if anno.hasanno(node.ast_node, anno.Static.SCOPE): + node_scope = anno.getanno(node.ast_node, anno.Static.SCOPE) + + gen = node_scope.used | self.extra_gen.get(node.ast_node, frozenset()) + # TODO(mdan): verify whether composites' parents need to be added. + # E.g. if x.y is live whether x needs to be added. Theoretically the + # activity analysis should have both so that wouldn't be needed. + kill = node_scope.modified + + live_out = set() + for n in node.next: + live_out |= self.in_[n] + live_in = gen | (live_out - kill) + + else: + # Nodes that don't have a scope annotation are assumed not to touch any + # symbols. + # This Name node below is a literal name, e.g. False + assert isinstance(node.ast_node, + (gast.Name, gast.Continue, gast.Break)), type( + node.ast_node) + live_in = prev_live_in + live_out = live_in + + self.in_[node] = live_in + self.out[node] = live_out + + # TODO(mdan): Move this to the superclass? + return prev_live_in != live_in + + +class WholeTreeAnalyzer(transformer.Base): + """Runs liveness analysis on each of the functions defined in the AST. + + If a function defined other local functions, those will have separate CFGs. + However, dataflow analysis needs to tie up these CFGs to properly emulate the + effect of closures. In the case of liveness, the parent function's live + variables must account for the variables that are live at the entry of each + subfunction. For example: + + def foo(): + # baz is live here + def bar(): + print(baz) + + This analyzer runs liveness analysis on each individual function, accounting + for the effect above. + """ + + def __init__(self, source_info, graphs): + super(WholeTreeAnalyzer, self).__init__(source_info) + self.graphs = graphs + self.current_analyzer = None + self.analyzers = {} + + def visit_FunctionDef(self, node): + parent_analyzer = self.current_analyzer + subgraph = self.graphs[node] + + # Postorder tree processing makes this a bit complicated: + # 1. construct an analyzer object and put it on stack + # 2. recursively walk the subtree; this will initialize the analyzer's + # in_ state properly (done in a block below) + # 3. run the final analysis + analyzer = Analyzer(subgraph) + self.current_analyzer = analyzer + node = self.generic_visit(node) + analyzer.visit_reverse() + + if parent_analyzer is not None: + # Wire the state between the two subgraphs' analyzers. + child_in_state = analyzer.in_[subgraph.entry] + # Exception: symbols modified in the child function are local to it + body_scope = anno.getanno(node, annos.NodeAnno.BODY_SCOPE) + for qn in body_scope.modified: + # Note: a function modifying the symbol doesn't make that symbol + # live at the function's entry. In fact when that happens it is + # probably a case of undefined assignment, like this: + # + # bar = 0 + # def foo(): + # print(bar) # bar is undefined here! + # bar = 1 + # + # Hence we use discard and not remove below. + child_in_state.discard(qn) + parent_analyzer.extra_gen[node] = frozenset(child_in_state,) + + self.analyzers[node] = analyzer + self.current_analyzer = parent_analyzer + return node + + def visit_nonlocal(self, node): + raise NotImplementedError() + + def visit_global(self, node): + raise NotImplementedError() + + +class Annotator(transformer.Base): + """AST visitor that annotates each control flow block with live symbols.""" + + # Note: additional nodes may be added as needed. + + def __init__(self, source_info, cross_function_analyzer): + super(Annotator, self).__init__(source_info) + self.cross_function_analyzer = cross_function_analyzer + self.current_analyzer = None + + def visit_FunctionDef(self, node): + parent_analyzer = self.current_analyzer + self.current_analyzer = self.cross_function_analyzer.analyzers[node] + + node = self.generic_visit(node) + self.current_analyzer = parent_analyzer + return node + + def _aggregate_successors_live_in(self, node): + successors = self.current_analyzer.graph.stmt_next[node] + node_live_out = set() + for s in successors: + node_live_out.update(self.current_analyzer.in_[s]) + anno.setanno(node, anno.Static.LIVE_VARS_OUT, frozenset(node_live_out)) + node = self.generic_visit(node) + return node + + def visit_If(self, node): + return self._aggregate_successors_live_in(node) + + def visit_For(self, node): + return self._aggregate_successors_live_in(node) + + def visit_While(self, node): + return self._aggregate_successors_live_in(node) + + +def resolve(node, source_info, graphs): + """Resolves the live symbols at the exit of control flow statements. + + Args: + node: ast.AST + source_info: transformer.SourceInfo + graphs: Dict[ast.FunctionDef, cfg.Graph] + Returns: + ast.AST + """ + cross_function_analyzer = WholeTreeAnalyzer(source_info, graphs) + node = cross_function_analyzer.visit(node) + visitor = Annotator(source_info, cross_function_analyzer) + node = visitor.visit(node) + return node diff --git a/tensorflow/contrib/autograph/pyct/static_analysis/liveness_test.py b/tensorflow/contrib/autograph/pyct/static_analysis/liveness_test.py new file mode 100644 index 0000000000000000000000000000000000000000..d53adb28af03f0de14f319f642ee82928a480e3a --- /dev/null +++ b/tensorflow/contrib/autograph/pyct/static_analysis/liveness_test.py @@ -0,0 +1,149 @@ +# Copyright 2018 The TensorFlow Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================== +"""Tests for liveness module.""" + +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + +from tensorflow.contrib.autograph.pyct import anno +from tensorflow.contrib.autograph.pyct import cfg +from tensorflow.contrib.autograph.pyct import parser +from tensorflow.contrib.autograph.pyct import qual_names +from tensorflow.contrib.autograph.pyct import transformer +from tensorflow.contrib.autograph.pyct.static_analysis import activity +from tensorflow.contrib.autograph.pyct.static_analysis import liveness +from tensorflow.python.platform import test + + +class LivenessTest(test.TestCase): + + def _parse_and_analyze(self, test_fn): + node, source = parser.parse_entity(test_fn) + entity_info = transformer.EntityInfo( + source_code=source, + source_file=None, + namespace={}, + arg_values=None, + arg_types=None, + owner_type=None) + node = qual_names.resolve(node) + node = activity.resolve(node, entity_info) + graphs = cfg.build(node) + liveness.resolve(node, entity_info, graphs) + return node + + def assertHasLiveOut(self, node, expected): + live_out = anno.getanno(node, anno.Static.LIVE_VARS_OUT) + live_out_str = set(str(v) for v in live_out) + if not expected: + expected = () + if not isinstance(expected, tuple): + expected = (expected,) + self.assertSetEqual(live_out_str, set(expected)) + + def test_stacked_if(self): + + def test_fn(x, a): + if a > 0: + x = 0 + if a > 1: + x = 1 + return x + + node = self._parse_and_analyze(test_fn) + fn_body = node.body[0].body + + self.assertHasLiveOut(fn_body[0], ('a', 'x')) + self.assertHasLiveOut(fn_body[1], 'x') + + def test_stacked_if_else(self): + + def test_fn(x, a): + if a > 0: + x = 0 + if a > 1: + x = 1 + else: + x = 2 + return x + + node = self._parse_and_analyze(test_fn) + fn_body = node.body[0].body + + self.assertHasLiveOut(fn_body[0], 'a') + self.assertHasLiveOut(fn_body[1], 'x') + + def test_for_basic(self): + + def test_fn(x, a): + for i in range(a): + x += i + return x + + node = self._parse_and_analyze(test_fn) + fn_body = node.body[0].body + + self.assertHasLiveOut(fn_body[0], 'x') + + def test_attributes(self): + + def test_fn(x, a): + if a > 0: + x.y = 0 + return x.y + + node = self._parse_and_analyze(test_fn) + fn_body = node.body[0].body + + self.assertHasLiveOut(fn_body[0], ('x.y', 'x')) + + def test_nested_functions(self): + + def test_fn(a, b): + if b: + a = [] + + def foo(): + return a + + foo() + + node = self._parse_and_analyze(test_fn) + fn_body = node.body[0].body + + self.assertHasLiveOut(fn_body[0], 'a') + + def test_nested_functions_isolation(self): + + def test_fn(b): + if b: + a = 0 # pylint:disable=unused-variable + + def child(): + max(a) # pylint:disable=used-before-assignment + a = 1 + return a + + child() + + node = self._parse_and_analyze(test_fn) + fn_body = node.body[0].body + + self.assertHasLiveOut(fn_body[0], 'max') + + +if __name__ == '__main__': + test.main() diff --git a/tensorflow/contrib/autograph/pyct/static_analysis/reaching_definitions.py b/tensorflow/contrib/autograph/pyct/static_analysis/reaching_definitions.py new file mode 100644 index 0000000000000000000000000000000000000000..4d79b0a56af292f49c1576bfaafb1ce8ca685188 --- /dev/null +++ b/tensorflow/contrib/autograph/pyct/static_analysis/reaching_definitions.py @@ -0,0 +1,273 @@ +# Copyright 2018 The TensorFlow Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================== +"""Reaching definition analysis. + +This analysis attaches a set of a Definition objects to each symbol, one +for each distinct definition that may reach it. The Definition objects are +mutable and may be used by subsequent analyses to further annotate data like +static type and value information. +The analysis also attaches the set of the symbols defined at the entry of +control flow statements. + +Requires activity analysis. +""" + +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + +import gast + +from tensorflow.contrib.autograph.pyct import anno +from tensorflow.contrib.autograph.pyct import cfg +from tensorflow.contrib.autograph.pyct import transformer +from tensorflow.contrib.autograph.pyct.static_analysis import annos + + +class Definition(object): + """Definition objects describe a unique definition of a variable. + + Subclasses of this may be used by passing an appropriate factory fuction to + resolve. + + Attributes: + param_of: Optional[ast.AST] + """ + + def __init__(self): + self.param_of = None + + def __repr__(self): + return '%s[%d]' % (self.__class__.__name__, id(self)) + + +class _NodeState(object): + """Abstraction for the state of the CFG walk for reaching definition analysis. + + This is a value type. Only implements the strictly necessary operators. + + Attributes: + value: Dict[qual_names.QN, Set[Definition, ...]], the defined symbols and + their possible definitions + """ + + def __init__(self, init_from=None): + if init_from: + if isinstance(init_from, _NodeState): + self.value = { + s: set(other_infos) for s, other_infos in init_from.value.items() + } + elif isinstance(init_from, dict): + self.value = {s: set((init_from[s],)) for s in init_from} + else: + assert False, init_from + else: + self.value = {} + + def __eq__(self, other): + if frozenset(self.value.keys()) != frozenset(other.value.keys()): + return False + ret = all(self.value[s] == other.value[s] for s in self.value) + return ret + + def __ne__(self, other): + return not self.__eq__(other) + + def __or__(self, other): + assert isinstance(other, _NodeState) + result = _NodeState(self) + for s, other_infos in other.value.items(): + if s in result.value: + result.value[s].update(other_infos) + else: + result.value[s] = set(other_infos) + return result + + def __sub__(self, other): + assert isinstance(other, set) + result = _NodeState(self) + for s in other: + result.value.pop(s, None) + return result + + def __repr__(self): + return 'NodeState[%s]=%s' % (id(self), repr(self.value)) + + +class Analyzer(cfg.GraphVisitor): + """CFG visitor that determines reaching definitions at statement level.""" + + def __init__(self, graph, definition_factory): + self._definition_factory = definition_factory + super(Analyzer, self).__init__(graph) + self.defs_by_ast_node = {} + # This allows communicating that nodes have extra reaching definitions, + # e.g. those that a function closes over. + self.extra_in = {} + + self.gen_map = {} + + def init_state(self, _): + return _NodeState() + + def visit_node(self, node): + prev_defs_out = self.out[node] + + defs_in = _NodeState(self.extra_in.get(node.ast_node, None)) + for n in node.prev: + defs_in |= self.out[n] + + if anno.hasanno(node.ast_node, anno.Static.SCOPE): + node_scope = anno.getanno(node.ast_node, anno.Static.SCOPE) + # The definition objects created by each node must be singletons because + # their ids are used in equality checks. + if node not in self.gen_map: + node_symbols = {} + for s in node_scope.modified: + def_ = self._definition_factory() + if s in node_scope.params: + def_.param_of = node_scope.params[s] + node_symbols[s] = def_ + self.gen_map[node] = _NodeState(node_symbols) + + gen = self.gen_map[node] + kill = node_scope.modified + defs_out = gen | (defs_in - kill) + + else: + # Nodes that don't have a scope annotation are assumed not to touch any + # symbols. + # This Name node below is a literal name, e.g. False + # This can also happen if activity.py forgot to annotate the node with a + # scope object. + assert isinstance(node.ast_node, + (gast.Name, gast.Break, gast.Continue)), (node.ast_node, + node) + defs_out = defs_in + + self.in_[node] = defs_in + self.out[node] = defs_out + self.defs_by_ast_node[node.ast_node] = defs_out.value + + # TODO(mdan): Move this to the superclass? + return prev_defs_out != defs_out + + +class WholeTreeAnalyzer(transformer.Base): + """AST visitor that annotates each symbol name with its reaching definitions. + + Simultaneously, the visitor runs the dataflow analysis on each function node, + accounting for the effect of closures. For example: + + def foo(): + bar = 1 + def baz(): + # bar = 1 reaches here + """ + + def __init__(self, source_info, graphs, definition_factory): + super(WholeTreeAnalyzer, self).__init__(source_info) + self.stmt_reaching_defs_info = None + self.graphs = graphs + self.current_analyzer = None + self.definition_factory = definition_factory + self.current_stmt_defs = None + + def visit_FunctionDef(self, node): + parent_analyzer = self.current_analyzer + subgraph = self.graphs[node] + + # Preorder tree processing: + # 1. if this is a child function, the parent was already analyzed and it + # has the proper state value for the subgraph's entry + # 2. analyze the current function body + # 2. recursively walk the subtree; child functions will be processed + analyzer = Analyzer(subgraph, self.definition_factory) + if parent_analyzer is not None: + # Wire the state between the two subgraphs' analyzers. + parent_out_state = parent_analyzer.out[parent_analyzer.graph.index[node]] + # Exception: symbols modified in the child function are local to it + body_scope = anno.getanno(node, annos.NodeAnno.BODY_SCOPE) + parent_out_state -= body_scope.modified + analyzer.extra_in[node.args] = parent_out_state + + # Complete the analysis for the local function and annotate its body. + analyzer.visit_forward() + + # Recursively process any remaining subfunctions. + self.current_analyzer = analyzer + node = self.generic_visit(node) + self.current_analyzer = parent_analyzer + + return node + + def visit_nonlocal(self, node): + raise NotImplementedError() + + def visit_global(self, node): + raise NotImplementedError() + + def visit_Name(self, node): + if self.current_analyzer is None: + # Names may appear outside function defs - for example in class + # definitions. + return node + + qn = anno.getanno(node, anno.Basic.QN) + assert self.current_stmt_defs is not None, ( + 'name node outside of any statement?') + anno.setanno(node, anno.Static.DEFINITIONS, + tuple(self.current_stmt_defs.get(qn, ()))) + return node + + def _aggregate_predecessors_defined_in(self, node): + preds = self.current_analyzer.graph.stmt_prev[node] + node_defined_in = set() + for p in preds: + node_defined_in |= set(self.current_analyzer.out[p].value.keys()) + anno.setanno(node, anno.Static.DEFINED_VARS_IN, frozenset(node_defined_in)) + node = self.generic_visit(node) + return node + + def visit_If(self, node): + return self._aggregate_predecessors_defined_in(node) + + def visit_For(self, node): + return self._aggregate_predecessors_defined_in(node) + + def visit_While(self, node): + return self._aggregate_predecessors_defined_in(node) + + def visit(self, node): + if (self.current_analyzer is not None and + node in self.current_analyzer.defs_by_ast_node): + self.current_stmt_defs = self.current_analyzer.defs_by_ast_node[node] + return super(WholeTreeAnalyzer, self).visit(node) + + +def resolve(node, source_info, graphs, definition_factory): + """Resolves reaching definitions for each symbol. + + Args: + node: ast.AST + source_info: transformer.SourceInfo + graphs: Dict[ast.FunctionDef, cfg.Graph] + definition_factory: Callable[[], Definition] + Returns: + ast.AST + """ + visitor = WholeTreeAnalyzer(source_info, graphs, definition_factory) + node = visitor.visit(node) + return node diff --git a/tensorflow/contrib/autograph/pyct/static_analysis/reaching_definitions_test.py b/tensorflow/contrib/autograph/pyct/static_analysis/reaching_definitions_test.py new file mode 100644 index 0000000000000000000000000000000000000000..0410bb2a35135a34ba9d1d11fa3a5916f3e831e6 --- /dev/null +++ b/tensorflow/contrib/autograph/pyct/static_analysis/reaching_definitions_test.py @@ -0,0 +1,221 @@ +# Copyright 2018 The TensorFlow Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================== +"""Tests for reaching_definitions module.""" + +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + +from tensorflow.contrib.autograph.pyct import anno +from tensorflow.contrib.autograph.pyct import cfg +from tensorflow.contrib.autograph.pyct import parser +from tensorflow.contrib.autograph.pyct import qual_names +from tensorflow.contrib.autograph.pyct import transformer +from tensorflow.contrib.autograph.pyct.static_analysis import activity +from tensorflow.contrib.autograph.pyct.static_analysis import reaching_definitions +from tensorflow.python.platform import test + + +class DefinitionInfoTest(test.TestCase): + + def _parse_and_analyze(self, test_fn): + node, source = parser.parse_entity(test_fn) + entity_info = transformer.EntityInfo( + source_code=source, + source_file=None, + namespace={}, + arg_values=None, + arg_types=None, + owner_type=None) + node = qual_names.resolve(node) + node = activity.resolve(node, entity_info) + graphs = cfg.build(node) + node = reaching_definitions.resolve(node, entity_info, graphs, + reaching_definitions.Definition) + return node + + def assertHasDefs(self, node, num): + defs = anno.getanno(node, anno.Static.DEFINITIONS) + self.assertEqual(len(defs), num) + for r in defs: + self.assertIsInstance(r, reaching_definitions.Definition) + + def assertHasDefinedIn(self, node, expected): + defined_in = anno.getanno(node, anno.Static.DEFINED_VARS_IN) + defined_in_str = set(str(v) for v in defined_in) + if not expected: + expected = () + if not isinstance(expected, tuple): + expected = (expected,) + self.assertSetEqual(defined_in_str, set(expected)) + + def test_conditional(self): + + def test_fn(a, b): + a = [] + if b: + a = [] + return a + + node = self._parse_and_analyze(test_fn) + fn_body = node.body[0].body + + self.assertHasDefs(fn_body[0].targets[0], 1) + self.assertHasDefs(fn_body[1].test, 1) + self.assertHasDefs(fn_body[1].body[0].targets[0], 1) + self.assertHasDefs(fn_body[2].value, 2) + + self.assertHasDefinedIn(fn_body[1], ('a', 'b')) + + def test_while(self): + + def test_fn(a): + max(a) + while True: + a = a + a = a + return a + + node = self._parse_and_analyze(test_fn) + fn_body = node.body[0].body + + self.assertHasDefs(fn_body[0].value.args[0], 1) + self.assertHasDefs(fn_body[1].body[0].targets[0], 1) + self.assertHasDefs(fn_body[1].body[0].value, 1) + self.assertHasDefs(fn_body[1].body[1].targets[0], 1) + self.assertHasDefs(fn_body[1].body[1].value, 1) + # The loop does have an invariant test, but the CFG doesn't know that. + self.assertHasDefs(fn_body[2].value, 2) + + def test_while_else(self): + + def test_fn(x, i): + y = 0 + while x: + x += i + if i: + break + else: + y = 1 + return x, y + + node = self._parse_and_analyze(test_fn) + fn_body = node.body[0].body + + self.assertHasDefs(fn_body[0].targets[0], 1) + self.assertHasDefs(fn_body[1].test, 2) + self.assertHasDefs(fn_body[1].body[0].target, 1) + self.assertHasDefs(fn_body[1].body[1].test, 1) + self.assertHasDefs(fn_body[1].orelse[0].targets[0], 1) + self.assertHasDefs(fn_body[2].value.elts[0], 2) + self.assertHasDefs(fn_body[2].value.elts[1], 2) + + def test_for_else(self): + + def test_fn(x, i): + y = 0 + for i in x: + x += i + if i: + break + else: + continue + else: + y = 1 + return x, y + + node = self._parse_and_analyze(test_fn) + fn_body = node.body[0].body + + self.assertHasDefs(fn_body[0].targets[0], 1) + self.assertHasDefs(fn_body[1].target, 1) + self.assertHasDefs(fn_body[1].body[0].target, 1) + self.assertHasDefs(fn_body[1].body[1].test, 1) + self.assertHasDefs(fn_body[1].orelse[0].targets[0], 1) + self.assertHasDefs(fn_body[2].value.elts[0], 2) + self.assertHasDefs(fn_body[2].value.elts[1], 2) + + def test_nested_functions(self): + + def test_fn(a, b): + a = [] + if b: + a = [] + + def foo(): + return a + + foo() + + return a + + node = self._parse_and_analyze(test_fn) + fn_body = node.body[0].body + def_of_a_in_if = fn_body[1].body[0].targets[0] + + self.assertHasDefs(fn_body[0].targets[0], 1) + self.assertHasDefs(fn_body[1].test, 1) + self.assertHasDefs(def_of_a_in_if, 1) + self.assertHasDefs(fn_body[2].value, 2) + + inner_fn_body = fn_body[1].body[1].body + self.assertHasDefs(inner_fn_body[0].value, 1) + self.assertTrue( + anno.getanno(inner_fn_body[0].value, anno.Static.DEFINITIONS)[0] is + anno.getanno(def_of_a_in_if, anno.Static.DEFINITIONS)[0]) + + def test_nested_functions_isolation(self): + + def test_fn(a): + a = 0 + + def child(): + a = 1 + return a + + child() + return a + + node = self._parse_and_analyze(test_fn) + fn_body = node.body[0].body + + self.assertHasDefs(fn_body[3].value, 1) + self.assertHasDefs(fn_body[1].body[1].value, 1) + + parent_return = fn_body[3] + child_return = fn_body[1].body[1] + # The assignment `a = 1` makes `a` local to `child`. + self.assertFalse( + anno.getanno(parent_return.value, anno.Static.DEFINITIONS)[0] is + anno.getanno(child_return.value, anno.Static.DEFINITIONS)[0]) + + def test_debug(self): + + def foo(_): + pass + + def test_fn(a): + with foo(a): + return a + + node = self._parse_and_analyze(test_fn) + fn_body = node.body[0].body + + self.assertHasDefs(fn_body[0].items[0].context_expr.func, 0) + self.assertHasDefs(fn_body[0].items[0].context_expr.args[0], 1) + + +if __name__ == '__main__': + test.main() diff --git a/tensorflow/contrib/autograph/pyct/templates.py b/tensorflow/contrib/autograph/pyct/templates.py index 9c479ebc2fa83d27dc363ae306daedb556734a1f..9001e54e463e8db77671b92e60de2c68e38e3029 100644 --- a/tensorflow/contrib/autograph/pyct/templates.py +++ b/tensorflow/contrib/autograph/pyct/templates.py @@ -26,6 +26,7 @@ import textwrap import gast +from tensorflow.contrib.autograph.pyct import anno from tensorflow.contrib.autograph.pyct import ast_util from tensorflow.contrib.autograph.pyct import parser from tensorflow.contrib.autograph.pyct import qual_names @@ -43,39 +44,64 @@ class ReplaceTransformer(gast.NodeTransformer): """ self.replacements = replacements self.in_replacements = False + self.preserved_annos = { + anno.Basic.SKIP_PROCESSING, + anno.Static.ORIG_DEFINITIONS, + } + + def _prepare_replacement(self, replaced, key): + """Prepares a replacement AST that's safe to swap in for a node. + + Args: + replaced: ast.AST, the node being replaced + key: Hashable, the key of the replacement AST + Returns: + ast.AST, the replacement AST + """ + repl = self.replacements[key] + + new_nodes = ast_util.copy_clean(repl, preserve_annos=self.preserved_annos) + if isinstance(new_nodes, gast.AST): + new_nodes = [new_nodes] + + return new_nodes def visit_Expr(self, node): - if (isinstance(node.value, gast.Name) and - node.value.id in self.replacements): - return self.visit(node.value) - self.generic_visit(node) - return node + # When replacing a placeholder with an entire statement, the replacement + # must stand on its own and not be wrapped in an Expr. + new_value = self.visit(node.value) + if new_value is node.value: + return node + return new_value def visit_keyword(self, node): - if node.arg in self.replacements: - repl = self.replacements[node.arg] - if isinstance(repl, gast.keyword): - return repl - elif (isinstance(repl, (list, tuple)) and repl and - all(isinstance(r, gast.keyword) for r in repl)): - return repl - # TODO(mdan): We may allow replacing with a string as well. - # For example, if one wanted to replace foo with bar in foo=baz, then - # we could allow changing just node arg, so that we end up with bar=baz. - raise ValueError( - 'a keyword argument may only be replaced by another keyword or a ' - 'non-empty list of keywords. Found: %s' % repl) - return self.generic_visit(node) + if node.arg not in self.replacements: + return self.generic_visit(node) + + repl = self._prepare_replacement(node, node.arg) + if isinstance(repl, gast.keyword): + return repl + elif (repl and isinstance(repl, (list, tuple)) and + all(isinstance(r, gast.keyword) for r in repl)): + return repl + # TODO(mdan): We may allow replacing with a string as well. + # For example, if one wanted to replace foo with bar in foo=baz, then + # we could allow changing just node arg, so that we end up with bar=baz. + raise ValueError( + 'a keyword argument may only be replaced by another keyword or a ' + 'non-empty list of keywords. Found: %s' % repl) def visit_FunctionDef(self, node): node = self.generic_visit(node) - if node.name in self.replacements: - repl = self.replacements[node.name] - if not isinstance(repl, (gast.Name, ast.Name)): - raise ValueError( - 'a function name can only be replaced by a Name node. Found: %s' % - repl) - node.name = repl.id + if node.name not in self.replacements: + return node + + repl = self.replacements[node.name] + if not isinstance(repl, (gast.Name, ast.Name)): + raise ValueError( + 'a function name can only be replaced by a Name node. Found: %s' % + repl) + node.name = repl.id return node def _check_has_context(self, node): @@ -148,6 +174,7 @@ class ReplaceTransformer(gast.NodeTransformer): node = self.generic_visit(node) if node.attr not in self.replacements: return node + repl = self.replacements[node.attr] if not isinstance(repl, gast.Name): raise ValueError( @@ -159,9 +186,7 @@ class ReplaceTransformer(gast.NodeTransformer): if node.id not in self.replacements: return node - new_nodes = ast_util.copy_clean(self.replacements[node.id]) - if isinstance(new_nodes, gast.AST): - new_nodes = [new_nodes] + new_nodes = self._prepare_replacement(node, node.id) # Preserve the target context. for n in new_nodes: @@ -182,7 +207,7 @@ class ReplaceTransformer(gast.NodeTransformer): def _convert_to_ast(n): - """Convert from a known data type to AST.""" + """Converts from a known data type to AST.""" if isinstance(n, str): # Note: the node will receive the ctx value from the template, see # ReplaceTransformer.visit_Name. @@ -197,7 +222,7 @@ def _convert_to_ast(n): def replace(template, **replacements): - """Replace placeholders in a Python template. + """Replaces placeholders in a Python template. AST Name and Tuple nodes always receive the context that inferred from the template. However, when replacing more complex nodes (that can potentially diff --git a/tensorflow/contrib/autograph/pyct/transformer.py b/tensorflow/contrib/autograph/pyct/transformer.py index 76558118308c31a2c1a770cad814e96abd6a6063..d9a157aead071ff76f66f920850a87966d20ffb4 100644 --- a/tensorflow/contrib/autograph/pyct/transformer.py +++ b/tensorflow/contrib/autograph/pyct/transformer.py @@ -59,6 +59,103 @@ class EntityInfo(object): self.owner_type = owner_type +class _StateStack(object): + """Typed stack abstraction. + + This class provides syntactic sugar for a stack of objects of known + type. It allows accessing attributes of the object at the top of the stack + directly against this object, which allows for very terse syntax. + + For example, this code: + + stack = _StateStack(Foo) + stack.enter() + stack.bar + + Is equivalent to: + + stack = [] + stack.append(Foo()) + foo = stack[-1] + foo.bar + + See _State for more on how this is used. + + Attributes: + type: Any, the type of objects that this stack holds + level: int, the current stack depth + value: Any, the instance of the object at the top of the stack + """ + + def __init__(self, type_): + # Because we override __setattr__, we need to attach these attributes using + # the superclass' setattr. + object.__setattr__(self, 'type', type_) + object.__setattr__(self, '_stack', []) + self.enter() + + def enter(self): + self._stack.append(self.type()) + + def exit(self): + return self._stack.pop() + + @property + def level(self): + return len(self._stack) + + @property + def value(self): + return self._stack[-1] + + def __getattr__(self, key): + return getattr(self._stack[-1], key) + + def __setattr__(self, key, value): + setattr(self._stack[-1], key, value) + + +class _State(object): + """Supporting class for nested scope variable space for converter.Base. + + This structure offers syntactic sugar over a dict of stacks of objects + of known type. These structures are useful to keep state during AST walks. + Multiple different scopes can be tracked in parallel. For example: + + s = _State() + + s[foo].enter() + s[bar].enter() # this will not affect s[foo] + + Element access has special semantics: + * keys are a data type + * element values are _StateStack(type=key) objects + * missing elements are automatically added, similarly to defaultdict + + For example, the following block : + + _State s + s[Foo] + + Is equivalent to: + + s = {} + if Foo not in s: + s[Foo] = Foo() + s[Foo] + + See Base for how it's used. + """ + + def __init__(self): + self._value = {} + + def __getitem__(self, key): + if key not in self._value: + self._value[key] = _StateStack(key) + return self._value[key] + + class Base(gast.NodeTransformer): """Base class for general-purpose code transformers transformers. @@ -71,6 +168,27 @@ class Base(gast.NodeTransformer): (possibly nested) scopes, use enter/exit_local_scope and set/get_local. You must call enter/exit_local_scope manually, but the transformer detects when they are not properly paired. + + The transformer allows keeping state across calls to visit_* that is local to + arbitrary nodes and their descendants, using the self.state attribute. + Multiple independent scopes are allowed and automatically constructed. + + For example, to keep track of the If node that encloses any Name node, one can + write: + + class FooType(object): + + def __init__(self): + self.foo_property = None + + class DummyTransformer(Base): + + def visit_If(self, node): + self.state[FooType].enter() + self.state[FooType].foo_property = node + + def visit_Name(self, node): + self.state[FooType].foo_property # will hold the innermost enclosing if """ # TODO(mdan): Document all extra features. @@ -92,6 +210,12 @@ class Base(gast.NodeTransformer): self._local_scope_state = [] self.enter_local_scope() + # Allows scoping of local variables to keep state across calls to visit_* + # methods. Multiple scope hierchies may exist and are keyed by tag. A scope + # is valid at one or more nodes and all its children. Scopes created in + # child nodes supersede their parent. Scopes are isolated from one another. + self.state = _State() + @property def enclosing_entities(self): return tuple(self._enclosing_entities) @@ -101,7 +225,9 @@ class Base(gast.NodeTransformer): return len(self._local_scope_state) def enter_local_scope(self, inherit=None): - """Marks entry into a new local scope. + """Deprecated. Use self.state instead. + + Marks entry into a new local scope. Args: inherit: Optional enumerable of variable names to copy from the @@ -116,7 +242,9 @@ class Base(gast.NodeTransformer): self._local_scope_state.append(scope_entered) def exit_local_scope(self, keep=None): - """Marks exit from the current local scope. + """Deprecated. Use self.state instead. + + Marks exit from the current local scope. Args: keep: Optional enumerable of variable names to copy into the @@ -133,9 +261,11 @@ class Base(gast.NodeTransformer): return scope_left def set_local(self, name, value): + """Deprecated. Use self.state instead.""" self._local_scope_state[-1][name] = value def get_local(self, name, default=None): + """Deprecated. Use self.state instead.""" return self._local_scope_state[-1].get(name, default) def debug_print(self, node): @@ -216,7 +346,7 @@ class Base(gast.NodeTransformer): node_destination = new_destination return results - # TODO(mdan): Once we have error tracing, we may be able to just go to SSA. + # TODO(mdan): Remove. def apply_to_single_assignments(self, targets, values, apply_fn): """Applies a function to each individual assignment. diff --git a/tensorflow/contrib/autograph/pyct/transformer_test.py b/tensorflow/contrib/autograph/pyct/transformer_test.py index baf04653ae862b0159fb50a1c67fa675ceb74b9a..19b80b09ac748999cac9e21a60b2da6c2a23faa5 100644 --- a/tensorflow/contrib/autograph/pyct/transformer_test.py +++ b/tensorflow/contrib/autograph/pyct/transformer_test.py @@ -93,6 +93,83 @@ class TransformerTest(test.TestCase): inner_function, lambda_node), anno.getanno(lambda_expr, 'enclosing_entities')) + def assertSameAnno(self, first, second, key): + self.assertIs(anno.getanno(first, key), anno.getanno(second, key)) + + def assertDifferentAnno(self, first, second, key): + self.assertIsNot(anno.getanno(first, key), anno.getanno(second, key)) + + def test_state_tracking(self): + + class LoopState(object): + pass + + class CondState(object): + pass + + class TestTransformer(transformer.Base): + + def visit(self, node): + anno.setanno(node, 'loop_state', self.state[LoopState].value) + anno.setanno(node, 'cond_state', self.state[CondState].value) + return super(TestTransformer, self).visit(node) + + def visit_While(self, node): + self.state[LoopState].enter() + node = self.generic_visit(node) + self.state[LoopState].exit() + return node + + def visit_If(self, node): + self.state[CondState].enter() + node = self.generic_visit(node) + self.state[CondState].exit() + return node + + tr = TestTransformer(self._simple_source_info()) + + def test_function(a): + a = 1 + while a: + _ = 'a' + if a > 2: + _ = 'b' + while True: + raise '1' + if a > 3: + _ = 'c' + while True: + raise '1' + + node, _ = parser.parse_entity(test_function) + node = tr.visit(node) + + fn_body = node.body[0].body + outer_while_body = fn_body[1].body + self.assertSameAnno(fn_body[0], outer_while_body[0], 'cond_state') + self.assertDifferentAnno(fn_body[0], outer_while_body[0], 'loop_state') + + first_if_body = outer_while_body[1].body + self.assertDifferentAnno(outer_while_body[0], first_if_body[0], + 'cond_state') + self.assertSameAnno(outer_while_body[0], first_if_body[0], 'loop_state') + + first_inner_while_body = first_if_body[1].body + self.assertSameAnno(first_if_body[0], first_inner_while_body[0], + 'cond_state') + self.assertDifferentAnno(first_if_body[0], first_inner_while_body[0], + 'loop_state') + + second_if_body = outer_while_body[2].body + self.assertDifferentAnno(first_if_body[0], second_if_body[0], 'cond_state') + self.assertSameAnno(first_if_body[0], second_if_body[0], 'loop_state') + + second_inner_while_body = second_if_body[1].body + self.assertDifferentAnno(first_inner_while_body[0], + second_inner_while_body[0], 'cond_state') + self.assertDifferentAnno(first_inner_while_body[0], + second_inner_while_body[0], 'loop_state') + def test_local_scope_info_stack(self): class TestTransformer(transformer.Base): diff --git a/tensorflow/contrib/batching/python/ops/batch_ops.py b/tensorflow/contrib/batching/python/ops/batch_ops.py index 47b80bdf4ad88ebce3603a14ea2aa3cbe5bd345f..55faad983f2bcf2f3fa633669bd371608e2e925b 100644 --- a/tensorflow/contrib/batching/python/ops/batch_ops.py +++ b/tensorflow/contrib/batching/python/ops/batch_ops.py @@ -58,8 +58,6 @@ def batch_function(num_batch_threads, max_batch_size, batch_timeout_micros, allowed_batch_sizes=None, - grad_timeout_micros=60 * 1000 * 1000, - unbatch_timeout_micros=60 * 1000 * 1000, max_enqueued_batches=10): """Batches the computation done by the decorated function. @@ -94,10 +92,6 @@ def batch_function(num_batch_threads, does nothing. Otherwise, supplies a list of batch sizes, causing the op to pad batches up to one of those sizes. The entries must increase monotonically, and the final entry must equal max_batch_size. - grad_timeout_micros: The timeout to use for the gradient. See the - documentation of the unbatch op for more details. Defaults to 60s. - unbatch_timeout_micros: The timeout to use for unbatching. See the - documentation of the unbatch op for more details. Defaults to 60s. max_enqueued_batches: The maximum depth of the batch queue. Defaults to 10. Returns: diff --git a/tensorflow/contrib/bayesflow/python/ops/monte_carlo_impl.py b/tensorflow/contrib/bayesflow/python/ops/monte_carlo_impl.py index 032b859d469ee5039e08e4af4c2f4ebf35c2ff19..68ead2f7609ca987180fe8973cf902f1e56b8388 100644 --- a/tensorflow/contrib/bayesflow/python/ops/monte_carlo_impl.py +++ b/tensorflow/contrib/bayesflow/python/ops/monte_carlo_impl.py @@ -192,7 +192,7 @@ def _logspace_mean(log_values): def expectation(f, samples, log_prob=None, use_reparametrization=True, axis=0, keep_dims=False, name=None): - """Computes the Monte-Carlo approximation of \\(E_p[f(X)]\\). + r"""Computes the Monte-Carlo approximation of \\(E_p[f(X)]\\). This function computes the Monte-Carlo approximation of an expectation, i.e., diff --git a/tensorflow/contrib/bigtable/BUILD b/tensorflow/contrib/bigtable/BUILD new file mode 100644 index 0000000000000000000000000000000000000000..71538e0770dcb436c8ff1571c22e950336328357 --- /dev/null +++ b/tensorflow/contrib/bigtable/BUILD @@ -0,0 +1,213 @@ +# Cloud Bigtable client for TensorFlow + +package( + default_visibility = ["//tensorflow:internal"], +) + +licenses(["notice"]) # Apache 2.0 + +load("//tensorflow:tensorflow.bzl", "tf_custom_op_py_library") +load( + "//tensorflow:tensorflow.bzl", + "tf_copts", + "tf_custom_op_library", + "tf_gen_op_libs", + "tf_gen_op_wrapper_py", + "tf_kernel_library", + "tf_cc_test", + "tf_py_test", +) + +tf_custom_op_py_library( + name = "bigtable", + srcs = ["__init__.py"] + glob(["python/ops/*.py"]), + dso = [ + ":python/ops/_bigtable.so", + ], + kernels = [ + ":bigtable_kernels", + ":bigtable_ops_op_lib", + ], + srcs_version = "PY2AND3", + deps = [ + ":bigtable_ops", + "//tensorflow/contrib/data/python/ops:interleave_ops", + "//tensorflow/contrib/util:util_py", + "//tensorflow/python:framework_for_generated_wrappers", + "//tensorflow/python:platform", + "//tensorflow/python:util", + "//tensorflow/python/data", + ], +) + +KERNEL_FILES = [ + "kernels/bigtable_kernels.cc", + "kernels/bigtable_lookup_dataset_op.cc", + "kernels/bigtable_prefix_key_dataset_op.cc", + "kernels/bigtable_range_key_dataset_op.cc", + "kernels/bigtable_sample_keys_dataset_op.cc", + "kernels/bigtable_sample_key_pairs_dataset_op.cc", + "kernels/bigtable_scan_dataset_op.cc", +] + +tf_custom_op_library( + name = "python/ops/_bigtable.so", + srcs = KERNEL_FILES + [ + "ops/bigtable_ops.cc", + ], + deps = [ + ":bigtable_lib_cc", + ":bigtable_range_helpers", + "@com_github_googlecloudplatform_google_cloud_cpp//google/cloud/bigtable:bigtable_client", + ], +) + +tf_gen_op_wrapper_py( + name = "bigtable_ops", + deps = [":bigtable_ops_op_lib"], +) + +tf_gen_op_libs( + op_lib_names = [ + "bigtable_ops", + "bigtable_test_ops", + ], +) + +tf_kernel_library( + name = "bigtable_kernels", + srcs = KERNEL_FILES, + deps = [ + ":bigtable_lib_cc", + ":bigtable_range_helpers", + "//tensorflow/core:framework_headers_lib", + "//third_party/eigen3", + "@com_github_googlecloudplatform_google_cloud_cpp//google/cloud/bigtable:bigtable_client", + ], +) + +# A library for use in the bigtable kernels. +cc_library( + name = "bigtable_lib_cc", + srcs = ["kernels/bigtable_lib.cc"], + hdrs = ["kernels/bigtable_lib.h"], + deps = [ + "//tensorflow/core:framework_headers_lib", + "//third_party/eigen3", + "@com_github_googlecloudplatform_google_cloud_cpp//google/cloud/bigtable:bigtable_client", + ], +) + +cc_library( + name = "bigtable_range_helpers", + srcs = ["kernels/bigtable_range_helpers.cc"], + hdrs = ["kernels/bigtable_range_helpers.h"], + deps = [ + "//tensorflow/core:framework_headers_lib", + ], +) + +cc_library( + name = "bigtable_test_client", + srcs = ["kernels/test_kernels/bigtable_test_client.cc"], + hdrs = ["kernels/test_kernels/bigtable_test_client.h"], + deps = [ + "//tensorflow/core:framework_headers_lib", + "@com_github_googleapis_googleapis//:bigtable_protos", + "@com_github_googlecloudplatform_google_cloud_cpp//google/cloud/bigtable:bigtable_client", + "@com_googlesource_code_re2//:re2", + ], +) + +tf_cc_test( + name = "bigtable_test_client_test", + srcs = ["kernels/test_kernels/bigtable_test_client_test.cc"], + tags = ["manual"], + deps = [ + ":bigtable_test_client", + "//tensorflow/core:test", + "//tensorflow/core:test_main", + "@com_github_googlecloudplatform_google_cloud_cpp//google/cloud/bigtable:bigtable_client", + ], +) + +tf_cc_test( + name = "bigtable_range_helpers_test", + size = "small", + srcs = ["kernels/bigtable_range_helpers_test.cc"], + deps = [ + ":bigtable_range_helpers", + "//tensorflow/core:test", + "//tensorflow/core:test_main", + ], +) + +tf_gen_op_wrapper_py( + name = "bigtable_test_ops", + deps = [":bigtable_test_ops_op_lib"], +) + +tf_custom_op_library( + name = "python/kernel_tests/_bigtable_test.so", + srcs = [ + "kernels/test_kernels/bigtable_test_client_op.cc", + "ops/bigtable_test_ops.cc", + ], + deps = [ + ":bigtable_lib_cc", + ":bigtable_test_client", + "@com_googlesource_code_re2//:re2", + ], +) + +# Don't use tf_kernel_library because it prevents access to strings/stringprintf.h +cc_library( + name = "bigtable_test_kernels", + srcs = [ + "kernels/test_kernels/bigtable_test_client_op.cc", + ], + copts = tf_copts(), + linkstatic = 1, + deps = [ + ":bigtable_lib_cc", + ":bigtable_test_client", + "//tensorflow/core:framework_headers_lib", + "//third_party/eigen3", + "@com_googlesource_code_re2//:re2", + ], + alwayslink = 1, +) + +tf_custom_op_py_library( + name = "bigtable_test_py", + dso = [ + ":python/kernel_tests/_bigtable_test.so", + ], + kernels = [ + ":bigtable_test_kernels", + ":bigtable_test_ops_op_lib", + ], + srcs_version = "PY2AND3", + deps = [ + ":bigtable_test_ops", + ], +) + +tf_py_test( + name = "bigtable_ops_test", + size = "small", + srcs = ["python/kernel_tests/bigtable_ops_test.py"], + additional_deps = [ + ":bigtable", + ":bigtable_test_py", + "//tensorflow/core:protos_all_py", + "//tensorflow/contrib/util:util_py", + "//tensorflow/python:array_ops", + "//tensorflow/python:client_testlib", + "//tensorflow/python:dtypes", + "//tensorflow/python:framework_for_generated_wrappers", + "//tensorflow/python:platform", + "//tensorflow/python:util", + ], + tags = ["manual"], +) diff --git a/tensorflow/contrib/bigtable/README.md b/tensorflow/contrib/bigtable/README.md new file mode 100644 index 0000000000000000000000000000000000000000..ef3c60069e8a97f7a13457156d20f3f7a4f7eccb --- /dev/null +++ b/tensorflow/contrib/bigtable/README.md @@ -0,0 +1,10 @@ +# Bigtable # + +[Google Cloud Bigtable](https://cloud.google.com/bigtable/) is a high +performance storage system that can store and serve training data. This contrib +package contains an experimental integration with TensorFlow. + +> **Status: Highly experimental.** The current implementation is very much in +> flux. Please use at your own risk! :-) + + diff --git a/tensorflow/contrib/proto/python/kernel_tests/test_case.py b/tensorflow/contrib/bigtable/__init__.py similarity index 58% rename from tensorflow/contrib/proto/python/kernel_tests/test_case.py rename to tensorflow/contrib/bigtable/__init__.py index b95202c5df654cfc02339477b242b2c58575a4d5..7df054637cdab32f2dd6201dd3488a90495e1cf5 100644 --- a/tensorflow/contrib/proto/python/kernel_tests/test_case.py +++ b/tensorflow/contrib/bigtable/__init__.py @@ -1,4 +1,3 @@ -# ============================================================================= # Copyright 2018 The TensorFlow Authors. All Rights Reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); @@ -12,24 +11,29 @@ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. -# ============================================================================= -"""Test case base for testing proto operations.""" +# ============================================================================== +"""Cloud Bigtable Client for TensorFlow. + +This contrib package allows TensorFlow to interface directly with Cloud Bigtable +for high-speed data loading. + +@@BigtableClient +@@BigTable + +""" -# Python3 preparedness imports. from __future__ import absolute_import from __future__ import division from __future__ import print_function -import ctypes as ct -import os - -from tensorflow.python.platform import test +from tensorflow.contrib.bigtable.python.ops.bigtable_api import BigTable +from tensorflow.contrib.bigtable.python.ops.bigtable_api import BigtableClient +from tensorflow.python.util.all_util import remove_undocumented -class ProtoOpTestCase(test.TestCase): +_allowed_symbols = [ + 'BigTable', + 'BigtableClient', +] - def __init__(self, methodName='runTest'): # pylint: disable=invalid-name - super(ProtoOpTestCase, self).__init__(methodName) - lib = os.path.join(os.path.dirname(__file__), 'libtestexample.so') - if os.path.isfile(lib): - ct.cdll.LoadLibrary(lib) +remove_undocumented(__name__, _allowed_symbols) diff --git a/tensorflow/contrib/bigtable/kernels/bigtable_kernels.cc b/tensorflow/contrib/bigtable/kernels/bigtable_kernels.cc new file mode 100644 index 0000000000000000000000000000000000000000..70923e6287018c5ecdf8849be99da4ffa68d9bd2 --- /dev/null +++ b/tensorflow/contrib/bigtable/kernels/bigtable_kernels.cc @@ -0,0 +1,355 @@ +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow/contrib/bigtable/kernels/bigtable_lib.h" + +#include "tensorflow/core/framework/op_kernel.h" +#include "tensorflow/core/lib/core/threadpool.h" + +namespace tensorflow { + +namespace { + +class BigtableClientOp : public OpKernel { + public: + explicit BigtableClientOp(OpKernelConstruction* ctx) : OpKernel(ctx) { + OP_REQUIRES_OK(ctx, ctx->GetAttr("project_id", &project_id_)); + OP_REQUIRES_OK(ctx, ctx->GetAttr("instance_id", &instance_id_)); + OP_REQUIRES(ctx, !project_id_.empty(), + errors::InvalidArgument("project_id must be non-empty")); + OP_REQUIRES(ctx, !instance_id_.empty(), + errors::InvalidArgument("instance_id must be non-empty")); + + OP_REQUIRES_OK( + ctx, ctx->GetAttr("connection_pool_size", &connection_pool_size_)); + // If left unset by the client code, set it to a default of 100. Note: the + // cloud-cpp default of 4 concurrent connections is far too low for high + // performance streaming. + if (connection_pool_size_ == -1) { + connection_pool_size_ = 100; + } + + OP_REQUIRES_OK(ctx, ctx->GetAttr("max_receive_message_size", + &max_receive_message_size_)); + // If left unset by the client code, set it to a default of 100. Note: the + // cloud-cpp default of 4 concurrent connections is far too low for high + // performance streaming. + if (max_receive_message_size_ == -1) { + max_receive_message_size_ = 1 << 24; // 16 MBytes + } + OP_REQUIRES(ctx, max_receive_message_size_ > 0, + errors::InvalidArgument("connection_pool_size must be > 0")); + } + + ~BigtableClientOp() override { + if (cinfo_.resource_is_private_to_kernel()) { + if (!cinfo_.resource_manager() + ->Delete(cinfo_.container(), + cinfo_.name()) + .ok()) { + // Do nothing; the resource can have been deleted by session resets. + } + } + } + + void Compute(OpKernelContext* ctx) override LOCKS_EXCLUDED(mu_) { + mutex_lock l(mu_); + if (!initialized_) { + ResourceMgr* mgr = ctx->resource_manager(); + OP_REQUIRES_OK(ctx, cinfo_.Init(mgr, def())); + BigtableClientResource* resource; + OP_REQUIRES_OK( + ctx, + mgr->LookupOrCreate( + cinfo_.container(), cinfo_.name(), &resource, + [this, ctx]( + BigtableClientResource** ret) EXCLUSIVE_LOCKS_REQUIRED(mu_) { + auto client_options = + google::cloud::bigtable::ClientOptions() + .set_connection_pool_size(connection_pool_size_) + .set_data_endpoint("batch-bigtable.googleapis.com"); + auto channel_args = client_options.channel_arguments(); + channel_args.SetMaxReceiveMessageSize( + max_receive_message_size_); + channel_args.SetUserAgentPrefix("tensorflow"); + client_options.set_channel_arguments(channel_args); + std::shared_ptr client = + google::cloud::bigtable::CreateDefaultDataClient( + project_id_, instance_id_, std::move(client_options)); + *ret = new BigtableClientResource(project_id_, instance_id_, + std::move(client)); + return Status::OK(); + })); + core::ScopedUnref resource_cleanup(resource); + initialized_ = true; + } + OP_REQUIRES_OK(ctx, MakeResourceHandleToOutput( + ctx, 0, cinfo_.container(), cinfo_.name(), + MakeTypeIndex())); + } + + private: + string project_id_; + string instance_id_; + int64 connection_pool_size_; + int32 max_receive_message_size_; + + mutex mu_; + ContainerInfo cinfo_ GUARDED_BY(mu_); + bool initialized_ GUARDED_BY(mu_) = false; +}; + +REGISTER_KERNEL_BUILDER(Name("BigtableClient").Device(DEVICE_CPU), + BigtableClientOp); + +class BigtableTableOp : public OpKernel { + public: + explicit BigtableTableOp(OpKernelConstruction* ctx) : OpKernel(ctx) { + OP_REQUIRES_OK(ctx, ctx->GetAttr("table_name", &table_)); + OP_REQUIRES(ctx, !table_.empty(), + errors::InvalidArgument("table_name must be non-empty")); + } + + ~BigtableTableOp() override { + if (cinfo_.resource_is_private_to_kernel()) { + if (!cinfo_.resource_manager() + ->Delete(cinfo_.container(), + cinfo_.name()) + .ok()) { + // Do nothing; the resource can have been deleted by session resets. + } + } + } + + void Compute(OpKernelContext* ctx) override LOCKS_EXCLUDED(mu_) { + mutex_lock l(mu_); + if (!initialized_) { + ResourceMgr* mgr = ctx->resource_manager(); + OP_REQUIRES_OK(ctx, cinfo_.Init(mgr, def())); + + BigtableClientResource* client_resource; + OP_REQUIRES_OK( + ctx, LookupResource(ctx, HandleFromInput(ctx, 0), &client_resource)); + core::ScopedUnref unref_client(client_resource); + + BigtableTableResource* resource; + OP_REQUIRES_OK( + ctx, mgr->LookupOrCreate( + cinfo_.container(), cinfo_.name(), &resource, + [this, client_resource](BigtableTableResource** ret) { + *ret = new BigtableTableResource(client_resource, table_); + return Status::OK(); + })); + initialized_ = true; + } + OP_REQUIRES_OK(ctx, MakeResourceHandleToOutput( + ctx, 0, cinfo_.container(), cinfo_.name(), + MakeTypeIndex())); + } + + private: + string table_; // Note: this is const after construction. + + mutex mu_; + ContainerInfo cinfo_ GUARDED_BY(mu_); + bool initialized_ GUARDED_BY(mu_) = false; +}; + +REGISTER_KERNEL_BUILDER(Name("BigtableTable").Device(DEVICE_CPU), + BigtableTableOp); + +class ToBigtableOp : public AsyncOpKernel { + public: + explicit ToBigtableOp(OpKernelConstruction* ctx) + : AsyncOpKernel(ctx), + thread_pool_(new thread::ThreadPool( + ctx->env(), ThreadOptions(), + strings::StrCat("to_bigtable_op_", SanitizeThreadSuffix(name())), + /* num_threads = */ 1, /* low_latency_hint = */ false)) {} + + void ComputeAsync(OpKernelContext* ctx, DoneCallback done) override { + // The call to `iterator->GetNext()` may block and depend on an + // inter-op thread pool thread, so we issue the call from the + // owned thread pool. + thread_pool_->Schedule([this, ctx, done]() { + const Tensor* column_families_tensor; + OP_REQUIRES_OK_ASYNC( + ctx, ctx->input("column_families", &column_families_tensor), done); + OP_REQUIRES_ASYNC( + ctx, column_families_tensor->dims() == 1, + errors::InvalidArgument("`column_families` must be a vector."), done); + + const Tensor* columns_tensor; + OP_REQUIRES_OK_ASYNC(ctx, ctx->input("columns", &columns_tensor), done); + OP_REQUIRES_ASYNC(ctx, columns_tensor->dims() == 1, + errors::InvalidArgument("`columns` must be a vector."), + done); + OP_REQUIRES_ASYNC( + ctx, + columns_tensor->NumElements() == + column_families_tensor->NumElements(), + errors::InvalidArgument("len(column_families) != len(columns)"), + done); + + std::vector column_families; + column_families.reserve(column_families_tensor->NumElements()); + std::vector columns; + columns.reserve(column_families_tensor->NumElements()); + for (uint64 i = 0; i < column_families_tensor->NumElements(); ++i) { + column_families.push_back(column_families_tensor->flat()(i)); + columns.push_back(columns_tensor->flat()(i)); + } + + DatasetBase* dataset; + OP_REQUIRES_OK_ASYNC( + ctx, GetDatasetFromVariantTensor(ctx->input(1), &dataset), done); + + IteratorContext iter_ctx = dataset::MakeIteratorContext(ctx); + std::unique_ptr iterator; + OP_REQUIRES_OK_ASYNC( + ctx, + dataset->MakeIterator(&iter_ctx, "ToBigtableOpIterator", &iterator), + done); + + int64 timestamp_int; + OP_REQUIRES_OK_ASYNC( + ctx, ParseScalarArgument(ctx, "timestamp", ×tamp_int), + done); + OP_REQUIRES_ASYNC(ctx, timestamp_int >= -1, + errors::InvalidArgument("timestamp must be >= -1"), + done); + + BigtableTableResource* resource; + OP_REQUIRES_OK_ASYNC( + ctx, LookupResource(ctx, HandleFromInput(ctx, 0), &resource), done); + core::ScopedUnref resource_cleanup(resource); + + std::vector components; + components.reserve(dataset->output_dtypes().size()); + bool end_of_sequence = false; + do { + ::google::cloud::bigtable::BulkMutation mutation; + // TODO(saeta): Make # of mutations configurable. + for (uint64 i = 0; i < 100 && !end_of_sequence; ++i) { + OP_REQUIRES_OK_ASYNC( + ctx, iterator->GetNext(&iter_ctx, &components, &end_of_sequence), + done); + if (!end_of_sequence) { + OP_REQUIRES_OK_ASYNC( + ctx, + CreateMutation(std::move(components), column_families, columns, + timestamp_int, &mutation), + done); + } + components.clear(); + } + grpc::Status mutation_status; + std::vector<::google::cloud::bigtable::FailedMutation> failures = + resource->table().BulkApply(std::move(mutation), mutation_status); + if (!mutation_status.ok()) { + LOG(ERROR) << "Failure applying mutation: " + << mutation_status.error_code() << " - " + << mutation_status.error_message() << " (" + << mutation_status.error_details() << ")."; + } + if (!failures.empty()) { + for (const auto& failure : failures) { + LOG(ERROR) << "Failure applying mutation on row (" + << failure.original_index() + << "): " << failure.mutation().row_key() + << " - error: " << failure.status().error_message() + << " (Details: " << failure.status().error_details() + << ")."; + } + } + OP_REQUIRES_ASYNC( + ctx, failures.empty() && mutation_status.ok(), + errors::Unknown("Failure while writing to BigTable: ", + mutation_status.error_code(), " - ", + mutation_status.error_message(), " (", + mutation_status.error_details(), + "), # of mutation failures: ", failures.size(), + ". See the log for the specific error details."), + done); + } while (!end_of_sequence); + done(); + }); + } + + private: + static string SanitizeThreadSuffix(string suffix) { + string clean; + for (int i = 0; i < suffix.size(); ++i) { + const char ch = suffix[i]; + if ((ch >= 'a' && ch <= 'z') || (ch >= 'A' && ch <= 'Z') || + (ch >= '0' && ch <= '9') || ch == '_' || ch == '-') { + clean += ch; + } else { + clean += '_'; + } + } + return clean; + } + + Status CreateMutation( + std::vector tensors, const std::vector& column_families, + const std::vector& columns, int64 timestamp_int, + ::google::cloud::bigtable::BulkMutation* bulk_mutation) { + if (tensors.size() != column_families.size() + 1) { + return errors::InvalidArgument( + "Iterator produced a set of Tensors shorter than expected"); + } + ::google::cloud::bigtable::SingleRowMutation mutation( + std::move(tensors[0].scalar()())); + std::chrono::milliseconds timestamp(timestamp_int); + for (size_t i = 1; i < tensors.size(); ++i) { + if (!TensorShapeUtils::IsScalar(tensors[i].shape())) { + return errors::Internal("Output tensor ", i, " was not a scalar"); + } + if (timestamp_int == -1) { + mutation.emplace_back(::google::cloud::bigtable::SetCell( + column_families[i - 1], columns[i - 1], + std::move(tensors[i].scalar()()))); + } else { + mutation.emplace_back(::google::cloud::bigtable::SetCell( + column_families[i - 1], columns[i - 1], timestamp, + std::move(tensors[i].scalar()()))); + } + } + bulk_mutation->emplace_back(std::move(mutation)); + return Status::OK(); + } + + template + Status ParseScalarArgument(OpKernelContext* ctx, + const StringPiece& argument_name, T* output) { + const Tensor* argument_t; + TF_RETURN_IF_ERROR(ctx->input(argument_name, &argument_t)); + if (!TensorShapeUtils::IsScalar(argument_t->shape())) { + return errors::InvalidArgument(argument_name, " must be a scalar"); + } + *output = argument_t->scalar()(); + return Status::OK(); + } + + std::unique_ptr thread_pool_; +}; + +REGISTER_KERNEL_BUILDER(Name("DatasetToBigtable").Device(DEVICE_CPU), + ToBigtableOp); + +} // namespace + +} // namespace tensorflow diff --git a/tensorflow/contrib/bigtable/kernels/bigtable_lib.cc b/tensorflow/contrib/bigtable/kernels/bigtable_lib.cc new file mode 100644 index 0000000000000000000000000000000000000000..2514575f30831bdcfab87eba07511fd309e8b1c2 --- /dev/null +++ b/tensorflow/contrib/bigtable/kernels/bigtable_lib.cc @@ -0,0 +1,45 @@ +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow/contrib/bigtable/kernels/bigtable_lib.h" + +namespace tensorflow { + +Status GrpcStatusToTfStatus(const ::grpc::Status& status) { + if (status.ok()) { + return Status::OK(); + } + auto grpc_code = status.error_code(); + if (status.error_code() == ::grpc::StatusCode::ABORTED || + status.error_code() == ::grpc::StatusCode::UNAVAILABLE || + status.error_code() == ::grpc::StatusCode::OUT_OF_RANGE) { + grpc_code = ::grpc::StatusCode::INTERNAL; + } + return Status( + static_cast<::tensorflow::error::Code>(status.error_code()), + strings::StrCat("Error reading from BigTable: ", status.error_message(), + " (Details: ", status.error_details(), ")")); +} + +string RegexFromStringSet(const std::vector& strs) { + CHECK(!strs.empty()) << "The list of strings to turn into a regex was empty."; + std::unordered_set uniq(strs.begin(), strs.end()); + if (uniq.size() == 1) { + return *uniq.begin(); + } + return str_util::Join(uniq, "|"); +} + +} // namespace tensorflow diff --git a/tensorflow/contrib/bigtable/kernels/bigtable_lib.h b/tensorflow/contrib/bigtable/kernels/bigtable_lib.h new file mode 100644 index 0000000000000000000000000000000000000000..a2a5df1037a00ccfdff1910dd950d7b012e684e2 --- /dev/null +++ b/tensorflow/contrib/bigtable/kernels/bigtable_lib.h @@ -0,0 +1,143 @@ +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#ifndef TENSORFLOW_CONTRIB_BIGTABLE_KERNELS_BIGTABLE_LIB_H_ +#define TENSORFLOW_CONTRIB_BIGTABLE_KERNELS_BIGTABLE_LIB_H_ + +// Note: we use bigtable/client/internal/table.h as this is the no-exception API + +#include "google/cloud/bigtable/data_client.h" +#include "google/cloud/bigtable/internal/table.h" +#include "tensorflow/core/framework/dataset.h" +#include "tensorflow/core/framework/resource_mgr.h" + +namespace tensorflow { + +Status GrpcStatusToTfStatus(const ::grpc::Status& status); + +string RegexFromStringSet(const std::vector& strs); + +class BigtableClientResource : public ResourceBase { + public: + BigtableClientResource( + string project_id, string instance_id, + std::shared_ptr client) + : project_id_(std::move(project_id)), + instance_id_(std::move(instance_id)), + client_(std::move(client)) {} + + std::shared_ptr get_client() { + return client_; + } + + string DebugString() override { + return strings::StrCat("BigtableClientResource(project_id: ", project_id_, + ", instance_id: ", instance_id_, ")"); + } + + private: + const string project_id_; + const string instance_id_; + std::shared_ptr client_; +}; + +class BigtableTableResource : public ResourceBase { + public: + BigtableTableResource(BigtableClientResource* client, string table_name) + : client_(client), + table_name_(std::move(table_name)), + table_(client->get_client(), table_name_, + google::cloud::bigtable::AlwaysRetryMutationPolicy()) { + client_->Ref(); + } + + ~BigtableTableResource() override { client_->Unref(); } + + ::google::cloud::bigtable::noex::Table& table() { return table_; } + + string DebugString() override { + return strings::StrCat( + "BigtableTableResource(client: ", client_->DebugString(), + ", table: ", table_name_, ")"); + } + + private: + BigtableClientResource* client_; // Ownes one ref. + const string table_name_; + ::google::cloud::bigtable::noex::Table table_; +}; + +// BigtableReaderDatasetIterator is an abstract class for iterators from +// datasets that are "readers" (source datasets, not transformation datasets) +// that read from Bigtable. +template +class BigtableReaderDatasetIterator : public DatasetIterator { + public: + explicit BigtableReaderDatasetIterator( + const typename DatasetIterator::Params& params) + : DatasetIterator(params), iterator_(nullptr, false) {} + + Status GetNextInternal(IteratorContext* ctx, std::vector* out_tensors, + bool* end_of_sequence) override { + mutex_lock l(mu_); + TF_RETURN_IF_ERROR(EnsureIteratorInitialized()); + if (iterator_ == reader_->end()) { + grpc::Status status = reader_->Finish(); + if (status.ok()) { + *end_of_sequence = true; + return Status::OK(); + } + return GrpcStatusToTfStatus(status); + } + *end_of_sequence = false; + google::cloud::bigtable::Row& row = *iterator_; + Status s = ParseRow(ctx, row, out_tensors); + // Ensure we always advance. + ++iterator_; + return s; + } + + protected: + virtual ::google::cloud::bigtable::RowRange MakeRowRange() = 0; + virtual ::google::cloud::bigtable::Filter MakeFilter() = 0; + virtual Status ParseRow(IteratorContext* ctx, + const ::google::cloud::bigtable::Row& row, + std::vector* out_tensors) = 0; + + private: + Status EnsureIteratorInitialized() EXCLUSIVE_LOCKS_REQUIRED(mu_) { + if (reader_) { + return Status::OK(); + } + + auto rows = MakeRowRange(); + auto filter = MakeFilter(); + + // Note: the this in `this->dataset()` below is necessary due to namespace + // name conflicts. + reader_.reset(new ::google::cloud::bigtable::RowReader( + this->dataset()->table()->table().ReadRows(rows, filter))); + iterator_ = reader_->begin(); + return Status::OK(); + } + + mutex mu_; + std::unique_ptr<::google::cloud::bigtable::RowReader> reader_ GUARDED_BY(mu_); + ::google::cloud::bigtable::RowReader::iterator iterator_ GUARDED_BY(mu_); +}; + +} // namespace tensorflow + +#endif // TENSORFLOW_CONTRIB_BIGTABLE_KERNELS_BIGTABLE_LIB_H_ diff --git a/tensorflow/contrib/bigtable/kernels/bigtable_lookup_dataset_op.cc b/tensorflow/contrib/bigtable/kernels/bigtable_lookup_dataset_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..9e49fa35db4b2cd2c8991100a28a5b9c55f01ffe --- /dev/null +++ b/tensorflow/contrib/bigtable/kernels/bigtable_lookup_dataset_op.cc @@ -0,0 +1,221 @@ +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow/contrib/bigtable/kernels/bigtable_lib.h" +#include "tensorflow/core/framework/op_kernel.h" + +namespace tensorflow { +namespace { + +class BigtableLookupDatasetOp : public UnaryDatasetOpKernel { + public: + using UnaryDatasetOpKernel::UnaryDatasetOpKernel; + + void MakeDataset(OpKernelContext* ctx, DatasetBase* input, + DatasetBase** output) override { + BigtableTableResource* table; + OP_REQUIRES_OK(ctx, LookupResource(ctx, HandleFromInput(ctx, 1), &table)); + + std::vector column_families; + std::vector columns; + OP_REQUIRES_OK(ctx, ParseVectorArgument(ctx, "column_families", + &column_families)); + OP_REQUIRES_OK(ctx, ParseVectorArgument(ctx, "columns", &columns)); + OP_REQUIRES( + ctx, column_families.size() == columns.size(), + errors::InvalidArgument("len(columns) != len(column_families)")); + + const uint64 num_outputs = columns.size() + 1; + std::vector output_shapes; + output_shapes.reserve(num_outputs); + DataTypeVector output_types; + output_types.reserve(num_outputs); + for (uint64 i = 0; i < num_outputs; ++i) { + output_shapes.push_back({}); + output_types.push_back(DT_STRING); + } + + *output = + new Dataset(ctx, input, table, std::move(column_families), + std::move(columns), output_types, std::move(output_shapes)); + } + + private: + class Dataset : public GraphDatasetBase { + public: + explicit Dataset(OpKernelContext* ctx, const DatasetBase* input, + BigtableTableResource* table, + std::vector column_families, + std::vector columns, + const DataTypeVector& output_types, + std::vector output_shapes) + : GraphDatasetBase(ctx), + input_(input), + table_(table), + column_families_(std::move(column_families)), + columns_(std::move(columns)), + output_types_(output_types), + output_shapes_(std::move(output_shapes)), + filter_(MakeFilter(column_families_, columns_)) { + table_->Ref(); + input_->Ref(); + } + + ~Dataset() override { + table_->Unref(); + input_->Unref(); + } + + std::unique_ptr MakeIteratorInternal( + const string& prefix) const override { + return std::unique_ptr(new Iterator( + {this, strings::StrCat(prefix, "::BigtableLookupDataset")})); + } + + const DataTypeVector& output_dtypes() const override { + return output_types_; + } + + const std::vector& output_shapes() const override { + return output_shapes_; + } + + string DebugString() const override { + return "BigtableLookupDatasetOp::Dataset"; + } + + private: + static ::google::cloud::bigtable::Filter MakeFilter( + const std::vector& column_families, + const std::vector& columns) { + string column_family_regex = RegexFromStringSet(column_families); + string column_regex = RegexFromStringSet(columns); + + return ::google::cloud::bigtable::Filter::Chain( + ::google::cloud::bigtable::Filter::Latest(1), + ::google::cloud::bigtable::Filter::FamilyRegex(column_family_regex), + ::google::cloud::bigtable::Filter::ColumnRegex(column_regex)); + } + + class Iterator : public DatasetIterator { + public: + explicit Iterator(const Params& params) + : DatasetIterator(params) {} + + Status Initialize(IteratorContext* ctx) override { + return dataset()->input_->MakeIterator(ctx, prefix(), &input_impl_); + } + + Status GetNextInternal(IteratorContext* ctx, + std::vector* out_tensors, + bool* end_of_sequence) override { + mutex_lock l(mu_); // Sequence requests. + std::vector input_tensors; + TF_RETURN_IF_ERROR( + input_impl_->GetNext(ctx, &input_tensors, end_of_sequence)); + if (*end_of_sequence) { + return Status::OK(); + } + if (input_tensors.size() != 1) { + return errors::InvalidArgument( + "Upstream iterator (", dataset()->input_->DebugString(), + ") did not produce a single `tf.string` `tf.Tensor`. It " + "produced ", + input_tensors.size(), " tensors."); + } + if (input_tensors[0].NumElements() == 0) { + return errors::InvalidArgument("Upstream iterator (", + dataset()->input_->DebugString(), + ") return an empty set of keys."); + } + if (input_tensors[0].NumElements() == 1) { + // Single key lookup. + ::grpc::Status status; + auto pair = dataset()->table_->table().ReadRow( + input_tensors[0].scalar()(), dataset()->filter_, status); + if (!status.ok()) { + return GrpcStatusToTfStatus(status); + } + if (!pair.first) { + return errors::DataLoss("Row key '", + input_tensors[0].scalar()(), + "' not found."); + } + TF_RETURN_IF_ERROR(ParseRow(ctx, pair.second, out_tensors)); + } else { + // Batched get. + return errors::Unimplemented( + "BigtableLookupDataset doesn't yet support batched retrieval."); + } + return Status::OK(); + } + + private: + Status ParseRow(IteratorContext* ctx, + const ::google::cloud::bigtable::Row& row, + std::vector* out_tensors) { + out_tensors->reserve(dataset()->columns_.size() + 1); + Tensor row_key_tensor(ctx->allocator({}), DT_STRING, {}); + row_key_tensor.scalar()() = string(row.row_key()); + out_tensors->emplace_back(std::move(row_key_tensor)); + + if (row.cells().size() > 2 * dataset()->columns_.size()) { + LOG(WARNING) << "An excessive number of columns (" + << row.cells().size() + << ") were retrieved when reading row: " + << row.row_key(); + } + + for (uint64 i = 0; i < dataset()->columns_.size(); ++i) { + Tensor col_tensor(ctx->allocator({}), DT_STRING, {}); + bool found_column = false; + for (auto cell_itr = row.cells().begin(); + !found_column && cell_itr != row.cells().end(); ++cell_itr) { + if (cell_itr->family_name() == dataset()->column_families_[i] && + string(cell_itr->column_qualifier()) == + dataset()->columns_[i]) { + col_tensor.scalar()() = string(cell_itr->value()); + found_column = true; + } + } + if (!found_column) { + return errors::DataLoss("Column ", dataset()->column_families_[i], + ":", dataset()->columns_[i], + " not found in row: ", row.row_key()); + } + out_tensors->emplace_back(std::move(col_tensor)); + } + return Status::OK(); + } + + mutex mu_; + std::unique_ptr input_impl_ GUARDED_BY(mu_); + }; + + const DatasetBase* const input_; + BigtableTableResource* table_; + const std::vector column_families_; + const std::vector columns_; + const DataTypeVector output_types_; + const std::vector output_shapes_; + const ::google::cloud::bigtable::Filter filter_; + }; +}; + +REGISTER_KERNEL_BUILDER(Name("BigtableLookupDataset").Device(DEVICE_CPU), + BigtableLookupDatasetOp); + +} // namespace +} // namespace tensorflow diff --git a/tensorflow/contrib/bigtable/kernels/bigtable_prefix_key_dataset_op.cc b/tensorflow/contrib/bigtable/kernels/bigtable_prefix_key_dataset_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..e960719614a1c7c6c4af53ea924aef214a09b24d --- /dev/null +++ b/tensorflow/contrib/bigtable/kernels/bigtable_prefix_key_dataset_op.cc @@ -0,0 +1,104 @@ +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow/contrib/bigtable/kernels/bigtable_lib.h" +#include "tensorflow/core/framework/op_kernel.h" + +namespace tensorflow { +namespace { + +class BigtablePrefixKeyDatasetOp : public DatasetOpKernel { + public: + using DatasetOpKernel::DatasetOpKernel; + + void MakeDataset(OpKernelContext* ctx, DatasetBase** output) override { + string prefix; + OP_REQUIRES_OK(ctx, ParseScalarArgument(ctx, "prefix", &prefix)); + + BigtableTableResource* resource; + OP_REQUIRES_OK(ctx, + LookupResource(ctx, HandleFromInput(ctx, 0), &resource)); + + *output = new Dataset(ctx, resource, std::move(prefix)); + } + + private: + class Dataset : public GraphDatasetBase { + public: + explicit Dataset(OpKernelContext* ctx, BigtableTableResource* table, + string prefix) + : GraphDatasetBase(ctx), table_(table), prefix_(std::move(prefix)) { + table_->Ref(); + } + + ~Dataset() override { table_->Unref(); } + + std::unique_ptr MakeIteratorInternal( + const string& prefix) const override { + return std::unique_ptr(new Iterator( + {this, strings::StrCat(prefix, "::BigtablePrefixKeyDataset")})); + } + + const DataTypeVector& output_dtypes() const override { + static DataTypeVector* dtypes = new DataTypeVector({DT_STRING}); + return *dtypes; + } + + const std::vector& output_shapes() const override { + static std::vector* shapes = + new std::vector({{}}); + return *shapes; + } + + string DebugString() const override { + return "BigtablePrefixKeyDatasetOp::Dataset"; + } + + BigtableTableResource* table() const { return table_; } + + private: + class Iterator : public BigtableReaderDatasetIterator { + public: + explicit Iterator(const Params& params) + : BigtableReaderDatasetIterator(params) {} + + ::google::cloud::bigtable::RowRange MakeRowRange() override { + return ::google::cloud::bigtable::RowRange::Prefix(dataset()->prefix_); + } + ::google::cloud::bigtable::Filter MakeFilter() override { + return ::google::cloud::bigtable::Filter::Chain( + ::google::cloud::bigtable::Filter::CellsRowLimit(1), + ::google::cloud::bigtable::Filter::StripValueTransformer()); + } + Status ParseRow(IteratorContext* ctx, + const ::google::cloud::bigtable::Row& row, + std::vector* out_tensors) override { + Tensor output_tensor(ctx->allocator({}), DT_STRING, {}); + output_tensor.scalar()() = string(row.row_key()); + out_tensors->emplace_back(std::move(output_tensor)); + return Status::OK(); + } + }; + + BigtableTableResource* const table_; + const string prefix_; + }; +}; + +REGISTER_KERNEL_BUILDER(Name("BigtablePrefixKeyDataset").Device(DEVICE_CPU), + BigtablePrefixKeyDatasetOp); + +} // namespace +} // namespace tensorflow diff --git a/tensorflow/contrib/bigtable/kernels/bigtable_range_helpers.cc b/tensorflow/contrib/bigtable/kernels/bigtable_range_helpers.cc new file mode 100644 index 0000000000000000000000000000000000000000..51965f6214413c08453473e71c30eecbd8925a64 --- /dev/null +++ b/tensorflow/contrib/bigtable/kernels/bigtable_range_helpers.cc @@ -0,0 +1,68 @@ +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow/contrib/bigtable/kernels/bigtable_range_helpers.h" + +#include "tensorflow/core/platform/logging.h" + +namespace tensorflow { + +namespace { + +string MakePrefixEndKey(const string& prefix) { + string end = prefix; + while (true) { + if (end.empty()) { + return end; + } + ++end[end.size() - 1]; + if (end[end.size() - 1] == 0) { + // Handle wraparound case. + end = end.substr(0, end.size() - 1); + } else { + return end; + } + } +} + +} // namespace + +/* static */ MultiModeKeyRange MultiModeKeyRange::FromPrefix(string prefix) { + string end = MakePrefixEndKey(prefix); + VLOG(1) << "Creating MultiModeKeyRange from Prefix: " << prefix + << ", with end key: " << end; + return MultiModeKeyRange(std::move(prefix), std::move(end)); +} + +/* static */ MultiModeKeyRange MultiModeKeyRange::FromRange(string begin, + string end) { + return MultiModeKeyRange(std::move(begin), std::move(end)); +} + +const string& MultiModeKeyRange::begin_key() const { return begin_; } + +const string& MultiModeKeyRange::end_key() const { return end_; } + +bool MultiModeKeyRange::contains_key(StringPiece key) const { + if (StringPiece(begin_) > key) { + return false; + } + if (StringPiece(end_) <= key && !end_.empty()) { + return false; + } + return true; +} + +} // namespace tensorflow diff --git a/tensorflow/contrib/bigtable/kernels/bigtable_range_helpers.h b/tensorflow/contrib/bigtable/kernels/bigtable_range_helpers.h new file mode 100644 index 0000000000000000000000000000000000000000..44c628e366c26b88011642f1e8e8d8e74b4698fd --- /dev/null +++ b/tensorflow/contrib/bigtable/kernels/bigtable_range_helpers.h @@ -0,0 +1,67 @@ +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#ifndef TENSORFLOW_CONTRIB_BIGTABLE_KERNELS_BIGTABLE_RANGE_HELPERS_H_ +#define TENSORFLOW_CONTRIB_BIGTABLE_KERNELS_BIGTABLE_RANGE_HELPERS_H_ + +#include + +#include "tensorflow/core/lib/core/stringpiece.h" +#include "tensorflow/core/platform/types.h" + +namespace tensorflow { + +// Represents a continuous range of keys defined by either a prefix or a range. +// +// Ranges are represented as "half-open", where the beginning key is included +// in the range, and the end_key is the first excluded key after the range. +// +// The range of keys can be specified either by a key prefix, or by an explicit +// begin key and end key. All methods on this class are valid no matter which +// way the range was specified. +// +// Example: +// MultiModeKeyRange range = MultiModeKeyRange::FromPrefix("myPrefix"); +// if (range.contains_key("myPrefixedKey")) { +// LOG(INFO) << "range from " << range.begin_key() << " to " +// << range.end_key() << "contains \"myPrefixedKey\""; +// } +// if (!range.contains_key("randomKey")) { +// LOG(INFO) << "range does not contain \"randomKey\""; +// } +// range = MultiModeKeyRange::FromRange("a_start_key", "z_end_key"); +class MultiModeKeyRange { + public: + static MultiModeKeyRange FromPrefix(string prefix); + static MultiModeKeyRange FromRange(string begin, string end); + + // The first valid key in the range. + const string& begin_key() const; + // The first invalid key after the valid range. + const string& end_key() const; + // Returns true if the provided key is a part of the range, false otherwise. + bool contains_key(StringPiece key) const; + + private: + MultiModeKeyRange(string begin, string end) + : begin_(std::move(begin)), end_(std::move(end)) {} + + const string begin_; + const string end_; +}; + +} // namespace tensorflow + +#endif // TENSORFLOW_CONTRIB_BIGTABLE_KERNELS_BIGTABLE_RANGE_HELPERS_H_ diff --git a/tensorflow/contrib/bigtable/kernels/bigtable_range_helpers_test.cc b/tensorflow/contrib/bigtable/kernels/bigtable_range_helpers_test.cc new file mode 100644 index 0000000000000000000000000000000000000000..1bfc547271d5e58a9145b73356b2b558dc1af9f1 --- /dev/null +++ b/tensorflow/contrib/bigtable/kernels/bigtable_range_helpers_test.cc @@ -0,0 +1,107 @@ +/* Copyright 2016 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow/contrib/bigtable/kernels/bigtable_range_helpers.h" +#include "tensorflow/core/platform/test.h" + +namespace tensorflow { +namespace { + +TEST(MultiModeKeyRangeTest, SimplePrefix) { + MultiModeKeyRange r = MultiModeKeyRange::FromPrefix("prefix"); + EXPECT_EQ("prefix", r.begin_key()); + EXPECT_EQ("prefiy", r.end_key()); + EXPECT_TRUE(r.contains_key("prefixed_key")); + EXPECT_FALSE(r.contains_key("not-prefixed-key")); + EXPECT_FALSE(r.contains_key("prefi")); + EXPECT_FALSE(r.contains_key("prefiy")); + EXPECT_FALSE(r.contains_key("early")); + EXPECT_FALSE(r.contains_key("")); +} + +TEST(MultiModeKeyRangeTest, Range) { + MultiModeKeyRange r = MultiModeKeyRange::FromRange("a", "b"); + EXPECT_EQ("a", r.begin_key()); + EXPECT_EQ("b", r.end_key()); + EXPECT_TRUE(r.contains_key("a")); + EXPECT_TRUE(r.contains_key("ab")); + EXPECT_FALSE(r.contains_key("b")); + EXPECT_FALSE(r.contains_key("bc")); + EXPECT_FALSE(r.contains_key("A")); + EXPECT_FALSE(r.contains_key("B")); + EXPECT_FALSE(r.contains_key("")); +} + +TEST(MultiModeKeyRangeTest, InvertedRange) { + MultiModeKeyRange r = MultiModeKeyRange::FromRange("b", "a"); + EXPECT_FALSE(r.contains_key("a")); + EXPECT_FALSE(r.contains_key("b")); + EXPECT_FALSE(r.contains_key("")); +} + +TEST(MultiModeKeyRangeTest, EmptyPrefix) { + MultiModeKeyRange r = MultiModeKeyRange::FromPrefix(""); + EXPECT_EQ("", r.begin_key()); + EXPECT_EQ("", r.end_key()); + EXPECT_TRUE(r.contains_key("")); + EXPECT_TRUE(r.contains_key("a")); + EXPECT_TRUE(r.contains_key("z")); + EXPECT_TRUE(r.contains_key("A")); + EXPECT_TRUE(r.contains_key("ZZZZZZ")); +} + +TEST(MultiModeKeyRangeTest, HalfRange) { + MultiModeKeyRange r = MultiModeKeyRange::FromRange("start", ""); + EXPECT_EQ("start", r.begin_key()); + EXPECT_EQ("", r.end_key()); + EXPECT_TRUE(r.contains_key("start")); + EXPECT_TRUE(r.contains_key("starting")); + EXPECT_TRUE(r.contains_key("z-end")); + EXPECT_FALSE(r.contains_key("")); + EXPECT_FALSE(r.contains_key("early")); +} + +TEST(MultiModeKeyRangeTest, PrefixWrapAround) { + string prefix = "abc\xff"; + MultiModeKeyRange r = MultiModeKeyRange::FromPrefix(prefix); + EXPECT_EQ(prefix, r.begin_key()); + EXPECT_EQ("abd", r.end_key()); + + EXPECT_TRUE(r.contains_key("abc\xff\x07")); + EXPECT_TRUE(r.contains_key("abc\xff\x15")); + EXPECT_TRUE(r.contains_key("abc\xff\x61")); + EXPECT_TRUE(r.contains_key("abc\xff\xff")); + EXPECT_FALSE(r.contains_key("abc\0")); + EXPECT_FALSE(r.contains_key("abd")); +} + +TEST(MultiModeKeyRangeTest, PrefixSignedWrapAround) { + string prefix = "abc\x7f"; + MultiModeKeyRange r = MultiModeKeyRange::FromPrefix(prefix); + EXPECT_EQ(prefix, r.begin_key()); + EXPECT_EQ("abc\x80", r.end_key()); + + EXPECT_TRUE(r.contains_key("abc\x7f\x07")); + EXPECT_TRUE(r.contains_key("abc\x7f\x15")); + EXPECT_TRUE(r.contains_key("abc\x7f\x61")); + EXPECT_TRUE(r.contains_key("abc\x7f\xff")); + EXPECT_FALSE(r.contains_key("abc\0")); + EXPECT_FALSE(r.contains_key("abc\x01")); + EXPECT_FALSE(r.contains_key("abd")); + EXPECT_FALSE(r.contains_key("ab\x80")); +} + +} // namespace +} // namespace tensorflow diff --git a/tensorflow/contrib/bigtable/kernels/bigtable_range_key_dataset_op.cc b/tensorflow/contrib/bigtable/kernels/bigtable_range_key_dataset_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..96d3565d9b90e72f9e25e69e91f1931c982714cd --- /dev/null +++ b/tensorflow/contrib/bigtable/kernels/bigtable_range_key_dataset_op.cc @@ -0,0 +1,112 @@ +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow/contrib/bigtable/kernels/bigtable_lib.h" +#include "tensorflow/core/framework/op_kernel.h" + +namespace tensorflow { +namespace { + +class BigtableRangeKeyDatasetOp : public DatasetOpKernel { + public: + using DatasetOpKernel::DatasetOpKernel; + + void MakeDataset(OpKernelContext* ctx, DatasetBase** output) override { + string start_key; + OP_REQUIRES_OK(ctx, + ParseScalarArgument(ctx, "start_key", &start_key)); + string end_key; + OP_REQUIRES_OK(ctx, ParseScalarArgument(ctx, "end_key", &end_key)); + + BigtableTableResource* resource; + OP_REQUIRES_OK(ctx, + LookupResource(ctx, HandleFromInput(ctx, 0), &resource)); + + *output = + new Dataset(ctx, resource, std::move(start_key), std::move(end_key)); + } + + private: + class Dataset : public GraphDatasetBase { + public: + explicit Dataset(OpKernelContext* ctx, BigtableTableResource* table, + string start_key, string end_key) + : GraphDatasetBase(ctx), + table_(table), + start_key_(std::move(start_key)), + end_key_(std::move(end_key)) { + table_->Ref(); + } + + ~Dataset() override { table_->Unref(); } + + std::unique_ptr MakeIteratorInternal( + const string& prefix) const override { + return std::unique_ptr(new Iterator( + {this, strings::StrCat(prefix, "::BigtableRangeKeyDataset")})); + } + + const DataTypeVector& output_dtypes() const override { + static DataTypeVector* dtypes = new DataTypeVector({DT_STRING}); + return *dtypes; + } + + const std::vector& output_shapes() const override { + static std::vector* shapes = + new std::vector({{}}); + return *shapes; + } + + string DebugString() const override { + return "BigtableRangeKeyDatasetOp::Dataset"; + } + + BigtableTableResource* table() const { return table_; } + + private: + class Iterator : public BigtableReaderDatasetIterator { + public: + explicit Iterator(const Params& params) + : BigtableReaderDatasetIterator(params) {} + + ::google::cloud::bigtable::RowRange MakeRowRange() override { + return ::google::cloud::bigtable::RowRange::Range(dataset()->start_key_, + dataset()->end_key_); + } + ::google::cloud::bigtable::Filter MakeFilter() override { + return ::google::cloud::bigtable::Filter::Chain( + ::google::cloud::bigtable::Filter::CellsRowLimit(1), + ::google::cloud::bigtable::Filter::StripValueTransformer()); + } + Status ParseRow(IteratorContext* ctx, + const ::google::cloud::bigtable::Row& row, + std::vector* out_tensors) override { + Tensor output_tensor(ctx->allocator({}), DT_STRING, {}); + output_tensor.scalar()() = string(row.row_key()); + out_tensors->emplace_back(std::move(output_tensor)); + return Status::OK(); + } + }; + + BigtableTableResource* const table_; + const string start_key_; + const string end_key_; + }; +}; + +REGISTER_KERNEL_BUILDER(Name("BigtableRangeKeyDataset").Device(DEVICE_CPU), + BigtableRangeKeyDatasetOp); +} // namespace +} // namespace tensorflow diff --git a/tensorflow/contrib/bigtable/kernels/bigtable_sample_key_pairs_dataset_op.cc b/tensorflow/contrib/bigtable/kernels/bigtable_sample_key_pairs_dataset_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..a1a63a975afd62325e01586542006058fa2c83bc --- /dev/null +++ b/tensorflow/contrib/bigtable/kernels/bigtable_sample_key_pairs_dataset_op.cc @@ -0,0 +1,200 @@ +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow/contrib/bigtable/kernels/bigtable_lib.h" +#include "tensorflow/contrib/bigtable/kernels/bigtable_range_helpers.h" +#include "tensorflow/core/framework/op_kernel.h" + +namespace tensorflow { +namespace { + +class BigtableSampleKeyPairsDatasetOp : public DatasetOpKernel { + public: + using DatasetOpKernel::DatasetOpKernel; + + void MakeDataset(OpKernelContext* ctx, DatasetBase** output) override { + string prefix; + OP_REQUIRES_OK(ctx, ParseScalarArgument(ctx, "prefix", &prefix)); + + string start_key; + OP_REQUIRES_OK(ctx, + ParseScalarArgument(ctx, "start_key", &start_key)); + string end_key; + OP_REQUIRES_OK(ctx, ParseScalarArgument(ctx, "end_key", &end_key)); + + BigtableTableResource* resource; + OP_REQUIRES_OK(ctx, + LookupResource(ctx, HandleFromInput(ctx, 0), &resource)); + + OP_REQUIRES(ctx, prefix.empty() || start_key.empty(), + errors::InvalidArgument( + "Only one of prefix and start_key can be provided")); + if (!prefix.empty()) { + OP_REQUIRES(ctx, end_key.empty(), + errors::InvalidArgument( + "If prefix is specified, end_key must be empty.")); + } + + *output = new Dataset(ctx, resource, std::move(prefix), + std::move(start_key), std::move(end_key)); + } + + private: + class Dataset : public GraphDatasetBase { + public: + explicit Dataset(OpKernelContext* ctx, BigtableTableResource* table, + string prefix, string start_key, string end_key) + : GraphDatasetBase(ctx), + table_(table), + key_range_(MakeMultiModeKeyRange( + std::move(prefix), std::move(start_key), std::move(end_key))) { + table_->Ref(); + } + + ~Dataset() override { table_->Unref(); } + + std::unique_ptr MakeIteratorInternal( + const string& prefix) const override { + return std::unique_ptr(new Iterator( + {this, strings::StrCat(prefix, "::BigtableSampleKeyPairsDataset")})); + } + + const DataTypeVector& output_dtypes() const override { + static DataTypeVector* dtypes = + new DataTypeVector({DT_STRING, DT_STRING}); + return *dtypes; + } + + const std::vector& output_shapes() const override { + static std::vector* shapes = + new std::vector({{}, {}}); + return *shapes; + } + + string DebugString() const override { + return "BigtableSampleKeyPairsDatasetOp::Dataset"; + } + + private: + static MultiModeKeyRange MakeMultiModeKeyRange(string prefix, + string start_key, + string end_key) { + if (!start_key.empty()) { + return MultiModeKeyRange::FromRange(std::move(start_key), + std::move(end_key)); + } + return MultiModeKeyRange::FromPrefix(std::move(prefix)); + } + + BigtableTableResource& table() const { return *table_; } + + class Iterator : public DatasetIterator { + public: + explicit Iterator(const Params& params) + : DatasetIterator(params) {} + + // Computes split points (`keys_`) to use when scanning the table. + // + // Initialize first retrieves the sample keys from the table (`row_keys`), + // as these often form good split points within the table. We then iterate + // over them, and copy them to `keys_` if they fall within the requested + // range to scan (`dataset()->key_range_`). Because the requested range + // might start between elements of the sampled keys list, care is taken to + // ensure we don't accidentally miss any subsets of the requested range by + // including `begin_key()` and `end_key()` as appropriate. + Status Initialize(IteratorContext* ctx) override { + grpc::Status status; + std::vector row_keys = + dataset()->table().table().SampleRows(status); + if (!status.ok()) { + return GrpcStatusToTfStatus(status); + } + + for (size_t i = 0; i < row_keys.size(); ++i) { + string row_key(row_keys[i].row_key); + if (dataset()->key_range_.contains_key(row_key)) { + // First key: check to see if we need to add the begin_key. + if (keys_.empty() && dataset()->key_range_.begin_key() != row_key) { + keys_.push_back(dataset()->key_range_.begin_key()); + } + keys_.push_back(std::move(row_key)); + } else if (!keys_.empty()) { + // If !keys_.empty(), then we have found at least one element of + // `row_keys` that is within our requested range + // (`dataset()->key_range_`). Because `row_keys` is sorted, if we + // have found an element that's not within our key range, then we + // are after our requested range (ranges are contiguous) and can end + // iteration early. + break; + } + } + + // Handle the case where we skip over the selected range entirely. + if (keys_.empty()) { + keys_.push_back(dataset()->key_range_.begin_key()); + } + + // Last key: check to see if we need to add the end_key. + if (keys_.back() != dataset()->key_range_.end_key()) { + keys_.push_back(dataset()->key_range_.end_key()); + } + return Status::OK(); + } + + Status GetNextInternal(IteratorContext* ctx, + std::vector* out_tensors, + bool* end_of_sequence) override { + mutex_lock l(mu_); + if (index_ > keys_.size() - 2) { + *end_of_sequence = true; + return Status::OK(); + } + + *end_of_sequence = false; + out_tensors->emplace_back(ctx->allocator({}), DT_STRING, + TensorShape({})); + out_tensors->back().scalar()() = keys_[index_]; + + out_tensors->emplace_back(ctx->allocator({}), DT_STRING, + TensorShape({})); + out_tensors->back().scalar()() = keys_[index_ + 1]; + ++index_; + + return Status::OK(); + } + + private: + mutex mu_; + size_t index_ GUARDED_BY(mu_) = 0; + // Note: we store the keys_ on the iterator instead of the dataset + // because we want to re-sample the row keys in case there have been + // tablet rebalancing operations since the dataset was created. + // + // Note: keys_ is readonly after Initialize, and thus does not need a + // guarding lock. + std::vector keys_; + }; + + BigtableTableResource* const table_; + const MultiModeKeyRange key_range_; + }; +}; + +REGISTER_KERNEL_BUILDER( + Name("BigtableSampleKeyPairsDataset").Device(DEVICE_CPU), + BigtableSampleKeyPairsDatasetOp); + +} // namespace +} // namespace tensorflow diff --git a/tensorflow/contrib/bigtable/kernels/bigtable_sample_keys_dataset_op.cc b/tensorflow/contrib/bigtable/kernels/bigtable_sample_keys_dataset_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..a5a47cfe2dcf7c4034e0d5bc7d9a73ef9c1dc94e --- /dev/null +++ b/tensorflow/contrib/bigtable/kernels/bigtable_sample_keys_dataset_op.cc @@ -0,0 +1,113 @@ +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow/contrib/bigtable/kernels/bigtable_lib.h" +#include "tensorflow/core/framework/op_kernel.h" + +namespace tensorflow { +namespace { + +class BigtableSampleKeysDatasetOp : public DatasetOpKernel { + public: + using DatasetOpKernel::DatasetOpKernel; + + void MakeDataset(OpKernelContext* ctx, DatasetBase** output) override { + BigtableTableResource* resource; + OP_REQUIRES_OK(ctx, + LookupResource(ctx, HandleFromInput(ctx, 0), &resource)); + *output = new Dataset(ctx, resource); + } + + private: + class Dataset : public GraphDatasetBase { + public: + explicit Dataset(OpKernelContext* ctx, BigtableTableResource* table) + : GraphDatasetBase(ctx), table_(table) { + table_->Ref(); + } + + ~Dataset() override { table_->Unref(); } + + std::unique_ptr MakeIteratorInternal( + const string& prefix) const override { + return std::unique_ptr(new Iterator( + {this, strings::StrCat(prefix, "::BigtableSampleKeysDataset")})); + } + + const DataTypeVector& output_dtypes() const override { + static DataTypeVector* dtypes = new DataTypeVector({DT_STRING}); + return *dtypes; + } + + const std::vector& output_shapes() const override { + static std::vector* shapes = + new std::vector({{}}); + return *shapes; + } + + string DebugString() const override { + return "BigtableRangeKeyDatasetOp::Dataset"; + } + + BigtableTableResource* table() const { return table_; } + + private: + class Iterator : public DatasetIterator { + public: + explicit Iterator(const Params& params) + : DatasetIterator(params) {} + + Status Initialize(IteratorContext* ctx) override { + ::grpc::Status status; + row_keys_ = dataset()->table()->table().SampleRows(status); + if (!status.ok()) { + row_keys_.clear(); + return GrpcStatusToTfStatus(status); + } + return Status::OK(); + } + + Status GetNextInternal(IteratorContext* ctx, + std::vector* out_tensors, + bool* end_of_sequence) override { + mutex_lock l(mu_); + if (index_ < row_keys_.size()) { + out_tensors->emplace_back(ctx->allocator({}), DT_STRING, + TensorShape({})); + out_tensors->back().scalar()() = + string(row_keys_[index_].row_key); + *end_of_sequence = false; + index_++; + } else { + *end_of_sequence = true; + } + return Status::OK(); + } + + private: + mutex mu_; + size_t index_ = 0; + std::vector<::google::cloud::bigtable::RowKeySample> row_keys_; + }; + + BigtableTableResource* const table_; + }; +}; + +REGISTER_KERNEL_BUILDER(Name("BigtableSampleKeysDataset").Device(DEVICE_CPU), + BigtableSampleKeysDatasetOp); + +} // namespace +} // namespace tensorflow diff --git a/tensorflow/contrib/bigtable/kernels/bigtable_scan_dataset_op.cc b/tensorflow/contrib/bigtable/kernels/bigtable_scan_dataset_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..13cb8681679ec1541b74a20474665f770790201f --- /dev/null +++ b/tensorflow/contrib/bigtable/kernels/bigtable_scan_dataset_op.cc @@ -0,0 +1,219 @@ +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow/contrib/bigtable/kernels/bigtable_lib.h" +#include "tensorflow/core/framework/op_kernel.h" + +namespace tensorflow { +namespace { + +class BigtableScanDatasetOp : public DatasetOpKernel { + public: + using DatasetOpKernel::DatasetOpKernel; + + void MakeDataset(OpKernelContext* ctx, DatasetBase** output) override { + string prefix; + OP_REQUIRES_OK(ctx, ParseScalarArgument(ctx, "prefix", &prefix)); + string start_key; + OP_REQUIRES_OK(ctx, + ParseScalarArgument(ctx, "start_key", &start_key)); + string end_key; + OP_REQUIRES_OK(ctx, ParseScalarArgument(ctx, "end_key", &end_key)); + + OP_REQUIRES(ctx, !(prefix.empty() && start_key.empty()), + errors::InvalidArgument( + "Either prefix or start_key must be specified")); + OP_REQUIRES(ctx, prefix.empty() || start_key.empty(), + errors::InvalidArgument( + "Only one of prefix and start_key can be provided")); + if (!prefix.empty()) { + OP_REQUIRES(ctx, end_key.empty(), + errors::InvalidArgument( + "If prefix is specified, end_key must be empty.")); + } + + std::vector column_families; + std::vector columns; + OP_REQUIRES_OK(ctx, ParseVectorArgument(ctx, "column_families", + &column_families)); + OP_REQUIRES_OK(ctx, ParseVectorArgument(ctx, "columns", &columns)); + OP_REQUIRES( + ctx, column_families.size() == columns.size(), + errors::InvalidArgument("len(columns) != len(column_families)")); + OP_REQUIRES(ctx, !column_families.empty(), + errors::InvalidArgument("`column_families` is empty")); + + float probability = 0; + OP_REQUIRES_OK( + ctx, ParseScalarArgument(ctx, "probability", &probability)); + OP_REQUIRES( + ctx, probability > 0 && probability <= 1, + errors::InvalidArgument( + "Probability outside the range of (0, 1]. Got: ", probability)); + + BigtableTableResource* resource; + OP_REQUIRES_OK(ctx, + LookupResource(ctx, HandleFromInput(ctx, 0), &resource)); + + const uint64 num_outputs = columns.size() + 1; + std::vector output_shapes; + output_shapes.reserve(num_outputs); + DataTypeVector output_types; + output_types.reserve(num_outputs); + for (uint64 i = 0; i < num_outputs; ++i) { + output_shapes.push_back({}); + output_types.push_back(DT_STRING); + } + + *output = new Dataset(ctx, resource, std::move(prefix), + std::move(start_key), std::move(end_key), + std::move(column_families), std::move(columns), + probability, output_types, std::move(output_shapes)); + } + + private: + class Dataset : public GraphDatasetBase { + public: + explicit Dataset(OpKernelContext* ctx, BigtableTableResource* table, + string prefix, string start_key, string end_key, + std::vector column_families, + std::vector columns, float probability, + const DataTypeVector& output_types, + std::vector output_shapes) + : GraphDatasetBase(ctx), + table_(table), + prefix_(std::move(prefix)), + start_key_(std::move(start_key)), + end_key_(std::move(end_key)), + column_families_(std::move(column_families)), + columns_(std::move(columns)), + column_family_regex_(RegexFromStringSet(column_families_)), + column_regex_(RegexFromStringSet(columns_)), + probability_(probability), + output_types_(output_types), + output_shapes_(std::move(output_shapes)) { + table_->Ref(); + } + + ~Dataset() override { table_->Unref(); } + + std::unique_ptr MakeIteratorInternal( + const string& prefix) const override { + return std::unique_ptr(new Iterator( + {this, strings::StrCat(prefix, "::BigtableScanDataset")})); + } + + const DataTypeVector& output_dtypes() const override { + return output_types_; + } + + const std::vector& output_shapes() const override { + return output_shapes_; + } + + string DebugString() const override { + return "BigtableScanDatasetOp::Dataset"; + } + + BigtableTableResource* table() const { return table_; } + + private: + class Iterator : public BigtableReaderDatasetIterator { + public: + explicit Iterator(const Params& params) + : BigtableReaderDatasetIterator(params) {} + + ::google::cloud::bigtable::RowRange MakeRowRange() override { + if (!dataset()->prefix_.empty()) { + DCHECK(dataset()->start_key_.empty()); + return ::google::cloud::bigtable::RowRange::Prefix( + dataset()->prefix_); + } else { + DCHECK(!dataset()->start_key_.empty()) + << "Both prefix and start_key were empty!"; + return ::google::cloud::bigtable::RowRange::Range( + dataset()->start_key_, dataset()->end_key_); + } + } + ::google::cloud::bigtable::Filter MakeFilter() override { + // TODO(saeta): Investigate optimal ordering here. + return ::google::cloud::bigtable::Filter::Chain( + ::google::cloud::bigtable::Filter::Latest(1), + ::google::cloud::bigtable::Filter::FamilyRegex( + dataset()->column_family_regex_), + ::google::cloud::bigtable::Filter::ColumnRegex( + dataset()->column_regex_), + dataset()->probability_ != 1.0 + ? ::google::cloud::bigtable::Filter::RowSample( + dataset()->probability_) + : ::google::cloud::bigtable::Filter::PassAllFilter()); + } + Status ParseRow(IteratorContext* ctx, + const ::google::cloud::bigtable::Row& row, + std::vector* out_tensors) override { + out_tensors->reserve(dataset()->columns_.size() + 1); + Tensor row_key_tensor(ctx->allocator({}), DT_STRING, {}); + row_key_tensor.scalar()() = string(row.row_key()); + out_tensors->emplace_back(std::move(row_key_tensor)); + + if (row.cells().size() > 2 * dataset()->columns_.size()) { + LOG(WARNING) << "An excessive number of columns (" + << row.cells().size() + << ") were retrieved when reading row: " + << row.row_key(); + } + + for (uint64 i = 0; i < dataset()->columns_.size(); ++i) { + Tensor col_tensor(ctx->allocator({}), DT_STRING, {}); + bool found_column = false; + for (auto cell_itr = row.cells().begin(); + !found_column && cell_itr != row.cells().end(); ++cell_itr) { + if (cell_itr->family_name() == dataset()->column_families_[i] && + string(cell_itr->column_qualifier()) == + dataset()->columns_[i]) { + col_tensor.scalar()() = string(cell_itr->value()); + found_column = true; + } + } + if (!found_column) { + return errors::InvalidArgument( + "Column ", dataset()->column_families_[i], ":", + dataset()->columns_[i], " not found in row: ", row.row_key()); + } + out_tensors->emplace_back(std::move(col_tensor)); + } + return Status::OK(); + } + }; + + BigtableTableResource* table_; + const string prefix_; + const string start_key_; + const string end_key_; + const std::vector column_families_; + const std::vector columns_; + const string column_family_regex_; + const string column_regex_; + const float probability_; + const DataTypeVector output_types_; + const std::vector output_shapes_; + }; +}; + +REGISTER_KERNEL_BUILDER(Name("BigtableScanDataset").Device(DEVICE_CPU), + BigtableScanDatasetOp); + +} // namespace +} // namespace tensorflow diff --git a/tensorflow/contrib/bigtable/kernels/test_kernels/bigtable_test_client.cc b/tensorflow/contrib/bigtable/kernels/test_kernels/bigtable_test_client.cc new file mode 100644 index 0000000000000000000000000000000000000000..f083ce6f44b3c2a83d9b5d3235056eb94c4be4a8 --- /dev/null +++ b/tensorflow/contrib/bigtable/kernels/test_kernels/bigtable_test_client.cc @@ -0,0 +1,374 @@ +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow/contrib/bigtable/kernels/test_kernels/bigtable_test_client.h" + +#include "google/bigtable/v2/data.pb.h" +#include "google/protobuf/wrappers.pb.h" +#include "re2/re2.h" +#include "tensorflow/core/lib/strings/stringprintf.h" +#include "tensorflow/core/util/ptr_util.h" +// #include "util/task/codes.pb.h" + +namespace tensorflow { +namespace { + +void UpdateRow(const ::google::bigtable::v2::Mutation& mut, + std::map* row) { + if (mut.has_set_cell()) { + CHECK(mut.set_cell().timestamp_micros() >= -1) + << "Timestamp_micros: " << mut.set_cell().timestamp_micros(); + auto col = + strings::Printf("%s:%s", mut.set_cell().family_name().c_str(), + string(mut.set_cell().column_qualifier()).c_str()); + (*row)[col] = string(mut.set_cell().value()); + } else if (mut.has_delete_from_column()) { + auto col = strings::Printf( + "%s:%s", mut.delete_from_column().family_name().c_str(), + string(mut.delete_from_column().column_qualifier()).c_str()); + row->erase(col); + } else if (mut.has_delete_from_family()) { + auto itr = row->lower_bound(mut.delete_from_family().family_name()); + auto prefix = + strings::Printf("%s:", mut.delete_from_family().family_name().c_str()); + while (itr != row->end() && itr->first.substr(0, prefix.size()) == prefix) { + row->erase(itr); + } + } else if (mut.has_delete_from_row()) { + row->clear(); + } else { + LOG(ERROR) << "Unknown mutation: " << mut.ShortDebugString(); + } +} + +} // namespace + +class SampleRowKeysResponse : public grpc::ClientReaderInterface< + google::bigtable::v2::SampleRowKeysResponse> { + public: + explicit SampleRowKeysResponse(BigtableTestClient* client) + : client_(client) {} + + bool NextMessageSize(uint32_t* sz) override { + mutex_lock l(mu_); + mutex_lock l2(client_->mu_); + if (num_messages_sent_ * 2 < client_->table_.rows.size()) { + *sz = 10000; // A sufficiently high enough value to not worry about. + return true; + } + return false; + } + + bool Read(google::bigtable::v2::SampleRowKeysResponse* resp) override { + // Send every other key from the table. + mutex_lock l(mu_); + mutex_lock l2(client_->mu_); + *resp = google::bigtable::v2::SampleRowKeysResponse(); + auto itr = client_->table_.rows.begin(); + for (uint64 i = 0; i < 2 * num_messages_sent_; ++i) { + ++itr; + if (itr == client_->table_.rows.end()) { + return false; + } + } + resp->set_row_key(itr->first); + resp->set_offset_bytes(100 * num_messages_sent_); + num_messages_sent_++; + return true; + } + + grpc::Status Finish() override { return grpc::Status::OK; } + + void WaitForInitialMetadata() override {} // Do nothing. + + private: + mutex mu_; + int64 num_messages_sent_ GUARDED_BY(mu_) = 0; + BigtableTestClient* client_; // Not owned. +}; + +class ReadRowsResponse : public grpc::ClientReaderInterface< + google::bigtable::v2::ReadRowsResponse> { + public: + ReadRowsResponse(BigtableTestClient* client, + google::bigtable::v2::ReadRowsRequest const& request) + : client_(client), request_(request) {} + + bool NextMessageSize(uint32_t* sz) override { + mutex_lock l(mu_); + if (sent_first_message_) { + return false; + } + *sz = 10000000; // A sufficiently high enough value to not worry about. + return true; + } + + bool Read(google::bigtable::v2::ReadRowsResponse* resp) override { + mutex_lock l(mu_); + if (sent_first_message_) { + return false; + } + sent_first_message_ = true; + RowFilter filter = MakeRowFilter(); + + mutex_lock l2(client_->mu_); + *resp = google::bigtable::v2::ReadRowsResponse(); + // Send all contents in first response. + for (auto itr = client_->table_.rows.begin(); + itr != client_->table_.rows.end(); ++itr) { + if (filter.AllowRow(itr->first)) { + ::google::bigtable::v2::ReadRowsResponse_CellChunk* chunk = nullptr; + bool sent_first = false; + for (auto col_itr = itr->second.columns.begin(); + col_itr != itr->second.columns.end(); ++col_itr) { + if (filter.AllowColumn(col_itr->first)) { + chunk = resp->add_chunks(); + if (!sent_first) { + sent_first = true; + chunk->set_row_key(itr->first); + } + auto colon_idx = col_itr->first.find(":"); + CHECK(colon_idx != string::npos) + << "No ':' found in: " << col_itr->first; + chunk->mutable_family_name()->set_value( + string(col_itr->first, 0, colon_idx)); + chunk->mutable_qualifier()->set_value( + string(col_itr->first, ++colon_idx)); + if (!filter.strip_values) { + chunk->set_value(col_itr->second); + } + if (filter.only_one_column) { + break; + } + } + } + if (sent_first) { + // We are sending this row, so set the commit flag on the last chunk. + chunk->set_commit_row(true); + } + } + } + return true; + } + + grpc::Status Finish() override { return grpc::Status::OK; } + + void WaitForInitialMetadata() override {} // Do nothing. + + private: + struct RowFilter { + std::set row_set; + std::vector> row_ranges; + double row_sample = 0.0; // Note: currently ignored. + std::unique_ptr col_filter; + bool strip_values = false; + bool only_one_column = false; + + bool AllowRow(const string& row) { + if (row_set.find(row) != row_set.end()) { + return true; + } + for (const auto& range : row_ranges) { + if (range.first <= row && range.second > row) { + return true; + } + } + return false; + } + + bool AllowColumn(const string& col) { + if (col_filter) { + return RE2::FullMatch(col, *col_filter); + } else { + return true; + } + } + }; + + RowFilter MakeRowFilter() { + RowFilter filter; + for (auto i = request_.rows().row_keys().begin(); + i != request_.rows().row_keys().end(); ++i) { + filter.row_set.insert(string(*i)); + } + for (auto i = request_.rows().row_ranges().begin(); + i != request_.rows().row_ranges().end(); ++i) { + if (i->start_key_case() != + google::bigtable::v2::RowRange::kStartKeyClosed || + i->end_key_case() != google::bigtable::v2::RowRange::kEndKeyOpen) { + LOG(WARNING) << "Skipping row range that cannot be processed: " + << i->ShortDebugString(); + continue; + } + filter.row_ranges.emplace_back(std::make_pair( + string(i->start_key_closed()), string(i->end_key_open()))); + } + if (request_.filter().has_chain()) { + string family_filter; + string qualifier_filter; + for (auto i = request_.filter().chain().filters().begin(); + i != request_.filter().chain().filters().end(); ++i) { + switch (i->filter_case()) { + case google::bigtable::v2::RowFilter::kFamilyNameRegexFilter: + family_filter = i->family_name_regex_filter(); + break; + case google::bigtable::v2::RowFilter::kColumnQualifierRegexFilter: + qualifier_filter = i->column_qualifier_regex_filter(); + break; + case google::bigtable::v2::RowFilter::kCellsPerColumnLimitFilter: + if (i->cells_per_column_limit_filter() != 1) { + LOG(ERROR) << "Unexpected cells_per_column_limit_filter: " + << i->cells_per_column_limit_filter(); + } + break; + case google::bigtable::v2::RowFilter::kStripValueTransformer: + filter.strip_values = i->strip_value_transformer(); + break; + case google::bigtable::v2::RowFilter::kRowSampleFilter: + LOG(INFO) << "Ignoring row sample directive."; + break; + case google::bigtable::v2::RowFilter::kPassAllFilter: + break; + case google::bigtable::v2::RowFilter::kCellsPerRowLimitFilter: + filter.only_one_column = true; + break; + default: + LOG(WARNING) << "Ignoring unknown filter type: " + << i->ShortDebugString(); + } + } + if (family_filter.empty() || qualifier_filter.empty()) { + LOG(WARNING) << "Missing regex!"; + } else { + string regex = strings::Printf("%s:%s", family_filter.c_str(), + qualifier_filter.c_str()); + filter.col_filter.reset(new RE2(regex)); + } + } else { + LOG(WARNING) << "Read request did not have a filter chain specified: " + << request_.filter().DebugString(); + } + return filter; + } + + mutex mu_; + bool sent_first_message_ GUARDED_BY(mu_) = false; + BigtableTestClient* client_; // Not owned. + const google::bigtable::v2::ReadRowsRequest request_; +}; + +class MutateRowsResponse : public grpc::ClientReaderInterface< + google::bigtable::v2::MutateRowsResponse> { + public: + explicit MutateRowsResponse(size_t num_successes) + : num_successes_(num_successes) {} + + bool NextMessageSize(uint32_t* sz) override { + mutex_lock l(mu_); + if (sent_first_message_) { + return false; + } + *sz = 10000000; // A sufficiently high enough value to not worry about. + return true; + } + + bool Read(google::bigtable::v2::MutateRowsResponse* resp) override { + mutex_lock l(mu_); + if (sent_first_message_) { + return false; + } + sent_first_message_ = true; + *resp = google::bigtable::v2::MutateRowsResponse(); + for (size_t i = 0; i < num_successes_; ++i) { + auto entry = resp->add_entries(); + entry->set_index(i); + } + return true; + } + + grpc::Status Finish() override { return grpc::Status::OK; } + + void WaitForInitialMetadata() override {} // Do nothing. + + private: + const size_t num_successes_; + + mutex mu_; + bool sent_first_message_ = false; +}; + +grpc::Status BigtableTestClient::MutateRow( + grpc::ClientContext* context, + google::bigtable::v2::MutateRowRequest const& request, + google::bigtable::v2::MutateRowResponse* response) { + mutex_lock l(mu_); + auto* row = &table_.rows[string(request.row_key())]; + for (int i = 0; i < request.mutations_size(); ++i) { + UpdateRow(request.mutations(i), &row->columns); + } + *response = google::bigtable::v2::MutateRowResponse(); + return grpc::Status::OK; +} +grpc::Status BigtableTestClient::CheckAndMutateRow( + grpc::ClientContext* context, + google::bigtable::v2::CheckAndMutateRowRequest const& request, + google::bigtable::v2::CheckAndMutateRowResponse* response) { + return grpc::Status(grpc::StatusCode::UNIMPLEMENTED, + "CheckAndMutateRow not implemented."); +} +grpc::Status BigtableTestClient::ReadModifyWriteRow( + grpc::ClientContext* context, + google::bigtable::v2::ReadModifyWriteRowRequest const& request, + google::bigtable::v2::ReadModifyWriteRowResponse* response) { + return grpc::Status(grpc::StatusCode::UNIMPLEMENTED, + "ReadModifyWriteRow not implemented."); +} +std::unique_ptr< + grpc::ClientReaderInterface> +BigtableTestClient::ReadRows( + grpc::ClientContext* context, + google::bigtable::v2::ReadRowsRequest const& request) { + return MakeUnique(this, request); +} + +std::unique_ptr< + grpc::ClientReaderInterface> +BigtableTestClient::SampleRowKeys( + grpc::ClientContext* context, + google::bigtable::v2::SampleRowKeysRequest const& request) { + return MakeUnique(this); +} +std::unique_ptr< + grpc::ClientReaderInterface> +BigtableTestClient::MutateRows( + grpc::ClientContext* context, + google::bigtable::v2::MutateRowsRequest const& request) { + mutex_lock l(mu_); + for (auto i = request.entries().begin(); i != request.entries().end(); ++i) { + auto* row = &table_.rows[string(i->row_key())]; + for (auto mut = i->mutations().begin(); mut != i->mutations().end(); + ++mut) { + UpdateRow(*mut, &row->columns); + } + } + return MakeUnique(request.entries_size()); +} + +std::shared_ptr BigtableTestClient::Channel() { + LOG(WARNING) << "Call to InMemoryDataClient::Channel(); this will likely " + "cause a crash!"; + return nullptr; +} +} // namespace tensorflow diff --git a/tensorflow/contrib/bigtable/kernels/test_kernels/bigtable_test_client.h b/tensorflow/contrib/bigtable/kernels/test_kernels/bigtable_test_client.h new file mode 100644 index 0000000000000000000000000000000000000000..dac2b16a216d26f02684c7401ed2ddaa4b7baddb --- /dev/null +++ b/tensorflow/contrib/bigtable/kernels/test_kernels/bigtable_test_client.h @@ -0,0 +1,87 @@ +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#ifndef TENSORFLOW_CONTRIB_BIGTABLE_KERNELS_TEST_KERNELS_BIGTABLE_TEST_CLIENT_H_ +#define TENSORFLOW_CONTRIB_BIGTABLE_KERNELS_TEST_KERNELS_BIGTABLE_TEST_CLIENT_H_ + +#include "google/cloud/bigtable/data_client.h" +#include "tensorflow/core/platform/logging.h" +#include "tensorflow/core/platform/mutex.h" + +namespace tensorflow { + +class BigtableTestClient : public ::google::cloud::bigtable::DataClient { + public: + std::string const& project_id() const override { return project_id_; } + std::string const& instance_id() const override { return instance_id_; } + void reset() override { + mutex_lock l(mu_); + table_ = Table(); + } + + grpc::Status MutateRow( + grpc::ClientContext* context, + google::bigtable::v2::MutateRowRequest const& request, + google::bigtable::v2::MutateRowResponse* response) override; + + grpc::Status CheckAndMutateRow( + grpc::ClientContext* context, + google::bigtable::v2::CheckAndMutateRowRequest const& request, + google::bigtable::v2::CheckAndMutateRowResponse* response) override; + + grpc::Status ReadModifyWriteRow( + grpc::ClientContext* context, + google::bigtable::v2::ReadModifyWriteRowRequest const& request, + google::bigtable::v2::ReadModifyWriteRowResponse* response) override; + + std::unique_ptr< + grpc::ClientReaderInterface> + ReadRows(grpc::ClientContext* context, + google::bigtable::v2::ReadRowsRequest const& request) override; + std::unique_ptr< + grpc::ClientReaderInterface> + SampleRowKeys( + grpc::ClientContext* context, + google::bigtable::v2::SampleRowKeysRequest const& request) override; + + std::unique_ptr< + grpc::ClientReaderInterface> + MutateRows(grpc::ClientContext* context, + google::bigtable::v2::MutateRowsRequest const& request) override; + + std::shared_ptr Channel() override; + + private: + friend class SampleRowKeysResponse; + friend class ReadRowsResponse; + friend class MutateRowsResponse; + + struct Row { + string row_key; + std::map columns; + }; + struct Table { + std::map rows; + }; + + mutex mu_; + const std::string project_id_ = "testproject"; + const std::string instance_id_ = "testinstance"; + Table table_ GUARDED_BY(mu_); +}; + +} // namespace tensorflow + +#endif // TENSORFLOW_CONTRIB_BIGTABLE_KERNELS_TEST_KERNELS_BIGTABLE_TEST_CLIENT_H_ diff --git a/tensorflow/contrib/bigtable/kernels/test_kernels/bigtable_test_client_op.cc b/tensorflow/contrib/bigtable/kernels/test_kernels/bigtable_test_client_op.cc new file mode 100644 index 0000000000000000000000000000000000000000..fa3e587b90147bd519586eef0cfb5e048b1b75be --- /dev/null +++ b/tensorflow/contrib/bigtable/kernels/test_kernels/bigtable_test_client_op.cc @@ -0,0 +1,78 @@ +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow/contrib/bigtable/kernels/bigtable_lib.h" +#include "tensorflow/contrib/bigtable/kernels/test_kernels/bigtable_test_client.h" +#include "tensorflow/core/framework/op_kernel.h" +#include "tensorflow/core/lib/strings/stringprintf.h" + +namespace tensorflow { + +namespace { + +class BigtableTestClientOp : public OpKernel { + public: + explicit BigtableTestClientOp(OpKernelConstruction* ctx) : OpKernel(ctx) {} + ~BigtableTestClientOp() override { + if (cinfo_.resource_is_private_to_kernel()) { + if (!cinfo_.resource_manager() + ->Delete(cinfo_.container(), + cinfo_.name()) + .ok()) { + // Do nothing; the resource can have been deleted by session resets. + } + } + } + void Compute(OpKernelContext* ctx) override LOCKS_EXCLUDED(mu_) { + mutex_lock l(mu_); + if (!initialized_) { + ResourceMgr* mgr = ctx->resource_manager(); + OP_REQUIRES_OK(ctx, cinfo_.Init(mgr, def())); + BigtableClientResource* resource; + OP_REQUIRES_OK( + ctx, + mgr->LookupOrCreate( + cinfo_.container(), cinfo_.name(), &resource, + [this, ctx](BigtableClientResource** ret) + EXCLUSIVE_LOCKS_REQUIRED(mu_) { + std::shared_ptr client( + new BigtableTestClient()); + // Note: must make explicit copies to sequence + // them before the move of client. + string project_id = client->project_id(); + string instance_id = client->instance_id(); + *ret = new BigtableClientResource(std::move(project_id), + std::move(instance_id), + std::move(client)); + return Status::OK(); + })); + initialized_ = true; + } + OP_REQUIRES_OK(ctx, MakeResourceHandleToOutput( + ctx, 0, cinfo_.container(), cinfo_.name(), + MakeTypeIndex())); + } + + private: + mutex mu_; + ContainerInfo cinfo_ GUARDED_BY(mu_); + bool initialized_ GUARDED_BY(mu_) = false; +}; + +REGISTER_KERNEL_BUILDER(Name("BigtableTestClient").Device(DEVICE_CPU), + BigtableTestClientOp); + +} // namespace +} // namespace tensorflow diff --git a/tensorflow/contrib/bigtable/kernels/test_kernels/bigtable_test_client_test.cc b/tensorflow/contrib/bigtable/kernels/test_kernels/bigtable_test_client_test.cc new file mode 100644 index 0000000000000000000000000000000000000000..32611e2590d9a81f46d0b9dfc09fe7e0068e9671 --- /dev/null +++ b/tensorflow/contrib/bigtable/kernels/test_kernels/bigtable_test_client_test.cc @@ -0,0 +1,345 @@ +/* Copyright 2016 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow/contrib/bigtable/kernels/test_kernels/bigtable_test_client.h" +#include "google/cloud/bigtable/internal/table.h" +#include "tensorflow/core/platform/test.h" + +namespace tensorflow { +namespace { + +void WriteCell(const string& row, const string& family, const string& column, + const string& value, + ::google::cloud::bigtable::noex::Table* table) { + ::google::cloud::bigtable::SingleRowMutation mut(row); + mut.emplace_back(::google::cloud::bigtable::SetCell(family, column, value)); + table->Apply(std::move(mut)); +} + +TEST(BigtableTestClientTest, EmptyRowRead) { + std::shared_ptr<::google::cloud::bigtable::DataClient> client_ptr = + std::make_shared(); + ::google::cloud::bigtable::noex::Table table(client_ptr, "test_table"); + + ::google::cloud::bigtable::RowSet rowset; + rowset.Append("r1"); + auto filter = ::google::cloud::bigtable::Filter::Chain( + ::google::cloud::bigtable::Filter::Latest(1)); + auto rows = table.ReadRows(std::move(rowset), filter); + EXPECT_EQ(rows.begin(), rows.end()) << "Some rows were returned in response!"; + EXPECT_TRUE(rows.Finish().ok()) << "Error reading rows."; +} + +TEST(BigtableTestClientTest, SingleRowWriteAndRead) { + std::shared_ptr<::google::cloud::bigtable::DataClient> client_ptr = + std::make_shared(); + ::google::cloud::bigtable::noex::Table table(client_ptr, "test_table"); + + WriteCell("r1", "f1", "c1", "v1", &table); + + ::google::cloud::bigtable::RowSet rowset("r1"); + auto filter = ::google::cloud::bigtable::Filter::Chain( + ::google::cloud::bigtable::Filter::Latest(1)); + auto rows = table.ReadRows(std::move(rowset), filter); + auto itr = rows.begin(); + EXPECT_NE(itr, rows.end()) << "No rows were returned in response!"; + EXPECT_EQ(itr->row_key(), "r1"); + EXPECT_EQ(itr->cells().size(), 1); + EXPECT_EQ(itr->cells()[0].family_name(), "f1"); + EXPECT_EQ(itr->cells()[0].column_qualifier(), "c1"); + EXPECT_EQ(itr->cells()[0].value(), "v1"); + + ++itr; + EXPECT_EQ(itr, rows.end()); + EXPECT_TRUE(rows.Finish().ok()); +} + +TEST(BigtableTestClientTest, MultiRowWriteAndSingleRowRead) { + std::shared_ptr<::google::cloud::bigtable::DataClient> client_ptr = + std::make_shared(); + ::google::cloud::bigtable::noex::Table table(client_ptr, "test_table"); + + WriteCell("r1", "f1", "c1", "v1", &table); + WriteCell("r2", "f1", "c1", "v2", &table); + WriteCell("r3", "f1", "c1", "v3", &table); + + ::google::cloud::bigtable::RowSet rowset("r1"); + auto filter = ::google::cloud::bigtable::Filter::Chain( + ::google::cloud::bigtable::Filter::Latest(1)); + auto rows = table.ReadRows(std::move(rowset), filter); + auto itr = rows.begin(); + + EXPECT_NE(itr, rows.end()) << "Missing rows"; + EXPECT_EQ(itr->row_key(), "r1"); + EXPECT_EQ(itr->cells().size(), 1); + EXPECT_EQ(itr->cells()[0].family_name(), "f1"); + EXPECT_EQ(itr->cells()[0].column_qualifier(), "c1"); + EXPECT_EQ(itr->cells()[0].value(), "v1"); + + ++itr; + EXPECT_EQ(itr, rows.end()) << "Extra rows in the response."; + EXPECT_TRUE(rows.Finish().ok()); +} + +TEST(BigtableTestClientTest, MultiRowWriteAndRead) { + std::shared_ptr<::google::cloud::bigtable::DataClient> client_ptr = + std::make_shared(); + ::google::cloud::bigtable::noex::Table table(client_ptr, "test_table"); + + WriteCell("r1", "f1", "c1", "v1", &table); + WriteCell("r2", "f1", "c1", "v2", &table); + WriteCell("r3", "f1", "c1", "v3", &table); + + ::google::cloud::bigtable::RowSet rowset("r1", "r2", "r3"); + auto filter = ::google::cloud::bigtable::Filter::Chain( + ::google::cloud::bigtable::Filter::Latest(1)); + auto rows = table.ReadRows(std::move(rowset), filter); + auto itr = rows.begin(); + + EXPECT_NE(itr, rows.end()) << "Missing rows"; + EXPECT_EQ(itr->row_key(), "r1"); + EXPECT_EQ(itr->cells().size(), 1); + EXPECT_EQ(itr->cells()[0].family_name(), "f1"); + EXPECT_EQ(itr->cells()[0].column_qualifier(), "c1"); + EXPECT_EQ(itr->cells()[0].value(), "v1"); + + ++itr; + + EXPECT_NE(itr, rows.end()) << "Missing rows"; + EXPECT_EQ(itr->row_key(), "r2"); + EXPECT_EQ(itr->cells().size(), 1); + EXPECT_EQ(itr->cells()[0].family_name(), "f1"); + EXPECT_EQ(itr->cells()[0].column_qualifier(), "c1"); + EXPECT_EQ(itr->cells()[0].value(), "v2"); + + ++itr; + + EXPECT_NE(itr, rows.end()) << "Missing rows"; + EXPECT_EQ(itr->row_key(), "r3"); + EXPECT_EQ(itr->cells().size(), 1); + EXPECT_EQ(itr->cells()[0].family_name(), "f1"); + EXPECT_EQ(itr->cells()[0].column_qualifier(), "c1"); + EXPECT_EQ(itr->cells()[0].value(), "v3"); + + ++itr; + EXPECT_EQ(itr, rows.end()) << "Extra rows in the response."; + EXPECT_TRUE(rows.Finish().ok()); +} + +TEST(BigtableTestClientTest, MultiRowWriteAndPrefixRead) { + std::shared_ptr<::google::cloud::bigtable::DataClient> client_ptr = + std::make_shared(); + ::google::cloud::bigtable::noex::Table table(client_ptr, "test_table"); + + WriteCell("r1", "f1", "c1", "v1", &table); + WriteCell("r2", "f1", "c1", "v2", &table); + WriteCell("r3", "f1", "c1", "v3", &table); + + auto filter = ::google::cloud::bigtable::Filter::Chain( + ::google::cloud::bigtable::Filter::Latest(1)); + auto rows = + table.ReadRows(::google::cloud::bigtable::RowRange::Prefix("r"), filter); + auto itr = rows.begin(); + + EXPECT_NE(itr, rows.end()) << "Missing rows"; + EXPECT_EQ(itr->row_key(), "r1"); + EXPECT_EQ(itr->cells().size(), 1); + EXPECT_EQ(itr->cells()[0].family_name(), "f1"); + EXPECT_EQ(itr->cells()[0].column_qualifier(), "c1"); + EXPECT_EQ(itr->cells()[0].value(), "v1"); + + ++itr; + + EXPECT_NE(itr, rows.end()) << "Missing rows"; + EXPECT_EQ(itr->row_key(), "r2"); + EXPECT_EQ(itr->cells().size(), 1); + EXPECT_EQ(itr->cells()[0].family_name(), "f1"); + EXPECT_EQ(itr->cells()[0].column_qualifier(), "c1"); + EXPECT_EQ(itr->cells()[0].value(), "v2"); + + ++itr; + + EXPECT_NE(itr, rows.end()) << "Missing rows"; + EXPECT_EQ(itr->row_key(), "r3"); + EXPECT_EQ(itr->cells().size(), 1); + EXPECT_EQ(itr->cells()[0].family_name(), "f1"); + EXPECT_EQ(itr->cells()[0].column_qualifier(), "c1"); + EXPECT_EQ(itr->cells()[0].value(), "v3"); + + ++itr; + EXPECT_EQ(itr, rows.end()) << "Extra rows in the response."; + EXPECT_TRUE(rows.Finish().ok()); +} + +TEST(BigtableTestClientTest, ColumnFiltering) { + std::shared_ptr<::google::cloud::bigtable::DataClient> client_ptr = + std::make_shared(); + ::google::cloud::bigtable::noex::Table table(client_ptr, "test_table"); + + WriteCell("r1", "f1", "c1", "v1", &table); + WriteCell("r2", "f1", "c1", "v2", &table); + WriteCell("r3", "f1", "c1", "v3", &table); + + // Extra cells + WriteCell("r1", "f2", "c1", "v1", &table); + WriteCell("r2", "f2", "c1", "v2", &table); + WriteCell("r3", "f1", "c2", "v3", &table); + + auto filter = ::google::cloud::bigtable::Filter::Chain( + ::google::cloud::bigtable::Filter::Latest(1), + ::google::cloud::bigtable::Filter::FamilyRegex("f1"), + ::google::cloud::bigtable::Filter::ColumnRegex("c1")); + auto rows = + table.ReadRows(::google::cloud::bigtable::RowRange::Prefix("r"), filter); + auto itr = rows.begin(); + + EXPECT_NE(itr, rows.end()) << "Missing rows"; + EXPECT_EQ(itr->row_key(), "r1"); + EXPECT_EQ(itr->cells().size(), 1); + EXPECT_EQ(itr->cells()[0].family_name(), "f1"); + EXPECT_EQ(itr->cells()[0].column_qualifier(), "c1"); + EXPECT_EQ(itr->cells()[0].value(), "v1"); + + ++itr; + + EXPECT_NE(itr, rows.end()) << "Missing rows"; + EXPECT_EQ(itr->row_key(), "r2"); + EXPECT_EQ(itr->cells().size(), 1); + EXPECT_EQ(itr->cells()[0].family_name(), "f1"); + EXPECT_EQ(itr->cells()[0].column_qualifier(), "c1"); + EXPECT_EQ(itr->cells()[0].value(), "v2"); + + ++itr; + + EXPECT_NE(itr, rows.end()) << "Missing rows"; + EXPECT_EQ(itr->row_key(), "r3"); + EXPECT_EQ(itr->cells().size(), 1); + EXPECT_EQ(itr->cells()[0].family_name(), "f1"); + EXPECT_EQ(itr->cells()[0].column_qualifier(), "c1"); + EXPECT_EQ(itr->cells()[0].value(), "v3"); + + ++itr; + EXPECT_EQ(itr, rows.end()) << "Extra rows in the response."; + EXPECT_TRUE(rows.Finish().ok()); +} + +TEST(BigtableTestClientTest, RowKeys) { + std::shared_ptr<::google::cloud::bigtable::DataClient> client_ptr = + std::make_shared(); + ::google::cloud::bigtable::noex::Table table(client_ptr, "test_table"); + + WriteCell("r1", "f1", "c1", "v1", &table); + WriteCell("r2", "f1", "c1", "v2", &table); + WriteCell("r3", "f1", "c1", "v3", &table); + + // Extra cells + WriteCell("r1", "f2", "c1", "v1", &table); + WriteCell("r2", "f2", "c1", "v2", &table); + WriteCell("r3", "f1", "c2", "v3", &table); + + auto filter = ::google::cloud::bigtable::Filter::Chain( + ::google::cloud::bigtable::Filter::Latest(1), + ::google::cloud::bigtable::Filter::CellsRowLimit(1), + ::google::cloud::bigtable::Filter::StripValueTransformer()); + auto rows = + table.ReadRows(::google::cloud::bigtable::RowRange::Prefix("r"), filter); + auto itr = rows.begin(); + EXPECT_NE(itr, rows.end()) << "Missing rows"; + EXPECT_EQ(itr->row_key(), "r1"); + EXPECT_EQ(itr->cells().size(), 1); + EXPECT_EQ(itr->cells()[0].family_name(), "f1"); + EXPECT_EQ(itr->cells()[0].column_qualifier(), "c1"); + EXPECT_EQ(itr->cells()[0].value(), ""); + + ++itr; + + EXPECT_NE(itr, rows.end()) << "Missing rows"; + EXPECT_EQ(itr->row_key(), "r2"); + EXPECT_EQ(itr->cells().size(), 1); + EXPECT_EQ(itr->cells()[0].family_name(), "f1"); + EXPECT_EQ(itr->cells()[0].column_qualifier(), "c1"); + EXPECT_EQ(itr->cells()[0].value(), ""); + + ++itr; + + EXPECT_NE(itr, rows.end()) << "Missing rows"; + EXPECT_EQ(itr->row_key(), "r3"); + EXPECT_EQ(itr->cells().size(), 1); + EXPECT_EQ(itr->cells()[0].family_name(), "f1"); + EXPECT_EQ(itr->cells()[0].column_qualifier(), "c1"); + EXPECT_EQ(itr->cells()[0].value(), ""); + + ++itr; + EXPECT_EQ(itr, rows.end()) << "Extra rows in the response."; + EXPECT_TRUE(rows.Finish().ok()); +} + +TEST(BigtableTestClientTest, SampleKeys) { + std::shared_ptr<::google::cloud::bigtable::DataClient> client_ptr = + std::make_shared(); + ::google::cloud::bigtable::noex::Table table(client_ptr, "test_table"); + + WriteCell("r1", "f1", "c1", "v1", &table); + WriteCell("r2", "f1", "c1", "v2", &table); + WriteCell("r3", "f1", "c1", "v3", &table); + WriteCell("r4", "f1", "c1", "v4", &table); + WriteCell("r5", "f1", "c1", "v5", &table); + + grpc::Status status; + auto resp = table.SampleRows(status); + EXPECT_TRUE(status.ok()); + EXPECT_EQ(3, resp.size()); + EXPECT_EQ("r1", string(resp[0].row_key)); + EXPECT_EQ(0, resp[0].offset_bytes); + EXPECT_EQ("r3", string(resp[1].row_key)); + EXPECT_EQ(100, resp[1].offset_bytes); + EXPECT_EQ("r5", string(resp[2].row_key)); + EXPECT_EQ(200, resp[2].offset_bytes); +} + +TEST(BigtableTestClientTest, SampleKeysShort) { + std::shared_ptr<::google::cloud::bigtable::DataClient> client_ptr = + std::make_shared(); + ::google::cloud::bigtable::noex::Table table(client_ptr, "test_table"); + + WriteCell("r1", "f1", "c1", "v1", &table); + + grpc::Status status; + auto resp = table.SampleRows(status); + EXPECT_TRUE(status.ok()); + EXPECT_EQ(1, resp.size()); + EXPECT_EQ("r1", string(resp[0].row_key)); +} + +TEST(BigtableTestClientTest, SampleKeysEvenNumber) { + std::shared_ptr<::google::cloud::bigtable::DataClient> client_ptr = + std::make_shared(); + ::google::cloud::bigtable::noex::Table table(client_ptr, "test_table"); + + WriteCell("r1", "f1", "c1", "v1", &table); + WriteCell("r2", "f1", "c1", "v2", &table); + WriteCell("r3", "f1", "c1", "v3", &table); + WriteCell("r4", "f1", "c1", "v4", &table); + + grpc::Status status; + auto resp = table.SampleRows(status); + EXPECT_TRUE(status.ok()); + EXPECT_EQ(2, resp.size()); + EXPECT_EQ("r1", string(resp[0].row_key)); + EXPECT_EQ("r3", string(resp[1].row_key)); +} + +} // namespace +} // namespace tensorflow diff --git a/tensorflow/contrib/bigtable/ops/bigtable_ops.cc b/tensorflow/contrib/bigtable/ops/bigtable_ops.cc new file mode 100644 index 0000000000000000000000000000000000000000..416b719e30aa5f2504449d151a48e95c9105c68b --- /dev/null +++ b/tensorflow/contrib/bigtable/ops/bigtable_ops.cc @@ -0,0 +1,107 @@ +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow/core/framework/common_shape_fns.h" +#include "tensorflow/core/framework/op.h" + +namespace tensorflow { + +// TODO(saeta): Add support for setting ClientOptions values. +REGISTER_OP("BigtableClient") + .Attr("project_id: string") + .Attr("instance_id: string") + .Attr("connection_pool_size: int") + .Attr("max_receive_message_size: int = -1") + .Attr("container: string = ''") + .Attr("shared_name: string = ''") + .Output("client: resource") + .SetShapeFn(shape_inference::ScalarShape); + +// TODO(saeta): Add support for Application Profiles. +// See https://cloud.google.com/bigtable/docs/app-profiles for more info. +REGISTER_OP("BigtableTable") + .Input("client: resource") + .Attr("table_name: string") + .Attr("container: string = ''") + .Attr("shared_name: string = ''") + .Output("table: resource") + .SetShapeFn(shape_inference::ScalarShape); + +REGISTER_OP("DatasetToBigtable") + .Input("table: resource") + .Input("input_dataset: variant") + .Input("column_families: string") + .Input("columns: string") + .Input("timestamp: int64") + .SetShapeFn(shape_inference::NoOutputs); + +REGISTER_OP("BigtableLookupDataset") + .Input("keys_dataset: variant") + .Input("table: resource") + .Input("column_families: string") + .Input("columns: string") + .Output("handle: variant") + .SetShapeFn(shape_inference::ScalarShape); + +REGISTER_OP("BigtablePrefixKeyDataset") + .Input("table: resource") + .Input("prefix: string") + .Output("handle: variant") + .SetIsStateful() // TODO(b/65524810): Source dataset ops must be marked + // stateful to inhibit constant folding. + .SetShapeFn(shape_inference::ScalarShape); + +REGISTER_OP("BigtableRangeKeyDataset") + .Input("table: resource") + .Input("start_key: string") + .Input("end_key: string") + .Output("handle: variant") + .SetIsStateful() // TODO(b/65524810): Source dataset ops must be marked + // stateful to inhibit constant folding. + .SetShapeFn(shape_inference::ScalarShape); + +REGISTER_OP("BigtableSampleKeysDataset") + .Input("table: resource") + .Output("handle: variant") + .SetIsStateful() // TODO(b/65524810): Source dataset ops must be marked + // stateful to inhibit constant folding. + .SetShapeFn(shape_inference::ScalarShape); + +REGISTER_OP("BigtableSampleKeyPairsDataset") + .Input("table: resource") + .Input("prefix: string") + .Input("start_key: string") + .Input("end_key: string") + .Output("handle: variant") + .SetIsStateful() // TODO(b/65524810): Source dataset ops must be marked + // stateful to inhibit constant folding. + .SetShapeFn(shape_inference::ScalarShape); + +// TODO(saeta): Support continuing despite bad data (e.g. empty string, or +// skip incomplete row.) +REGISTER_OP("BigtableScanDataset") + .Input("table: resource") + .Input("prefix: string") + .Input("start_key: string") + .Input("end_key: string") + .Input("column_families: string") + .Input("columns: string") + .Input("probability: float") + .Output("handle: variant") + .SetIsStateful() // TODO(b/65524810): Source dataset ops must be marked + // stateful to inhibit constant folding. + .SetShapeFn(shape_inference::ScalarShape); + +} // namespace tensorflow diff --git a/tensorflow/contrib/bigtable/ops/bigtable_test_ops.cc b/tensorflow/contrib/bigtable/ops/bigtable_test_ops.cc new file mode 100644 index 0000000000000000000000000000000000000000..f7d02458f63d547000f00b184b3d5e3c5007fb72 --- /dev/null +++ b/tensorflow/contrib/bigtable/ops/bigtable_test_ops.cc @@ -0,0 +1,27 @@ +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow/core/framework/common_shape_fns.h" +#include "tensorflow/core/framework/op.h" + +namespace tensorflow { + +REGISTER_OP("BigtableTestClient") + .Attr("container: string = ''") + .Attr("shared_name: string = ''") + .Output("client: resource") + .SetShapeFn(shape_inference::ScalarShape); + +} // namespace tensorflow diff --git a/tensorflow/contrib/bigtable/python/kernel_tests/__init__.py b/tensorflow/contrib/bigtable/python/kernel_tests/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..292d8f4e51abbbd89d68b47febd86b7297bb8ed2 --- /dev/null +++ b/tensorflow/contrib/bigtable/python/kernel_tests/__init__.py @@ -0,0 +1,20 @@ +# Copyright 2018 The TensorFlow Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================== + +"""This module contains tests for the bigtable integration.""" + +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function diff --git a/tensorflow/contrib/bigtable/python/kernel_tests/bigtable_ops_test.py b/tensorflow/contrib/bigtable/python/kernel_tests/bigtable_ops_test.py new file mode 100644 index 0000000000000000000000000000000000000000..2f2006461926c11ea1c150cf6dd8219e776b7dd1 --- /dev/null +++ b/tensorflow/contrib/bigtable/python/kernel_tests/bigtable_ops_test.py @@ -0,0 +1,272 @@ +# Copyright 2016 The TensorFlow Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================== +"""Tests for Bigtable Ops.""" + +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + +from tensorflow.contrib import bigtable +from tensorflow.contrib.bigtable.ops import gen_bigtable_ops +from tensorflow.contrib.bigtable.ops import gen_bigtable_test_ops +from tensorflow.contrib.bigtable.python.ops import bigtable_api +from tensorflow.contrib.util import loader +from tensorflow.python.data.ops import dataset_ops +from tensorflow.python.framework import errors +from tensorflow.python.platform import resource_loader +from tensorflow.python.platform import test +from tensorflow.python.util import compat + +_bigtable_so = loader.load_op_library( + resource_loader.get_path_to_datafile("_bigtable_test.so")) + + +def _ListOfTuplesOfStringsToBytes(values): + return [(compat.as_bytes(i[0]), compat.as_bytes(i[1])) for i in values] + + +class BigtableOpsTest(test.TestCase): + COMMON_ROW_KEYS = ["r1", "r2", "r3"] + COMMON_VALUES = ["v1", "v2", "v3"] + + def setUp(self): + self._client = gen_bigtable_test_ops.bigtable_test_client() + table = gen_bigtable_ops.bigtable_table(self._client, "testtable") + self._table = bigtable.BigTable("testtable", None, table) + + def _makeSimpleDataset(self): + output_rows = dataset_ops.Dataset.from_tensor_slices(self.COMMON_ROW_KEYS) + output_values = dataset_ops.Dataset.from_tensor_slices(self.COMMON_VALUES) + return dataset_ops.Dataset.zip((output_rows, output_values)) + + def _writeCommonValues(self, sess): + output_ds = self._makeSimpleDataset() + write_op = self._table.write(output_ds, ["cf1"], ["c1"]) + sess.run(write_op) + + def runReadKeyTest(self, read_ds): + itr = read_ds.make_initializable_iterator() + n = itr.get_next() + expected = list(self.COMMON_ROW_KEYS) + expected.reverse() + with self.test_session() as sess: + self._writeCommonValues(sess) + sess.run(itr.initializer) + for i in range(3): + output = sess.run(n) + want = expected.pop() + self.assertEqual( + compat.as_bytes(want), compat.as_bytes(output), + "Unequal at step %d: want: %s, got: %s" % (i, want, output)) + + def testReadPrefixKeys(self): + self.runReadKeyTest(self._table.keys_by_prefix_dataset("r")) + + def testReadRangeKeys(self): + self.runReadKeyTest(self._table.keys_by_range_dataset("r1", "r4")) + + def runScanTest(self, read_ds): + itr = read_ds.make_initializable_iterator() + n = itr.get_next() + expected_keys = list(self.COMMON_ROW_KEYS) + expected_keys.reverse() + expected_values = list(self.COMMON_VALUES) + expected_values.reverse() + with self.test_session() as sess: + self._writeCommonValues(sess) + sess.run(itr.initializer) + for i in range(3): + output = sess.run(n) + want = expected_keys.pop() + self.assertEqual( + compat.as_bytes(want), compat.as_bytes(output[0]), + "Unequal keys at step %d: want: %s, got: %s" % (i, want, output[0])) + want = expected_values.pop() + self.assertEqual( + compat.as_bytes(want), compat.as_bytes(output[1]), + "Unequal values at step: %d: want: %s, got: %s" % (i, want, + output[1])) + + def testScanPrefixStringCol(self): + self.runScanTest(self._table.scan_prefix("r", cf1="c1")) + + def testScanPrefixListCol(self): + self.runScanTest(self._table.scan_prefix("r", cf1=["c1"])) + + def testScanPrefixTupleCol(self): + self.runScanTest(self._table.scan_prefix("r", columns=("cf1", "c1"))) + + def testScanRangeStringCol(self): + self.runScanTest(self._table.scan_range("r1", "r4", cf1="c1")) + + def testScanRangeListCol(self): + self.runScanTest(self._table.scan_range("r1", "r4", cf1=["c1"])) + + def testScanRangeTupleCol(self): + self.runScanTest(self._table.scan_range("r1", "r4", columns=("cf1", "c1"))) + + def testLookup(self): + ds = self._table.keys_by_prefix_dataset("r") + ds = ds.apply(self._table.lookup_columns(cf1="c1")) + itr = ds.make_initializable_iterator() + n = itr.get_next() + expected_keys = list(self.COMMON_ROW_KEYS) + expected_values = list(self.COMMON_VALUES) + expected_tuples = zip(expected_keys, expected_values) + with self.test_session() as sess: + self._writeCommonValues(sess) + sess.run(itr.initializer) + for i, elem in enumerate(expected_tuples): + output = sess.run(n) + self.assertEqual( + compat.as_bytes(elem[0]), compat.as_bytes(output[0]), + "Unequal keys at step %d: want: %s, got: %s" % + (i, compat.as_bytes(elem[0]), compat.as_bytes(output[0]))) + self.assertEqual( + compat.as_bytes(elem[1]), compat.as_bytes(output[1]), + "Unequal values at step %d: want: %s, got: %s" % + (i, compat.as_bytes(elem[1]), compat.as_bytes(output[1]))) + + def testSampleKeys(self): + ds = self._table.sample_keys() + itr = ds.make_initializable_iterator() + n = itr.get_next() + expected_key = self.COMMON_ROW_KEYS[0] + with self.test_session() as sess: + self._writeCommonValues(sess) + sess.run(itr.initializer) + output = sess.run(n) + self.assertEqual( + compat.as_bytes(self.COMMON_ROW_KEYS[0]), compat.as_bytes(output), + "Unequal keys: want: %s, got: %s" % (compat.as_bytes( + self.COMMON_ROW_KEYS[0]), compat.as_bytes(output))) + output = sess.run(n) + self.assertEqual( + compat.as_bytes(self.COMMON_ROW_KEYS[2]), compat.as_bytes(output), + "Unequal keys: want: %s, got: %s" % (compat.as_bytes( + self.COMMON_ROW_KEYS[2]), compat.as_bytes(output))) + with self.assertRaises(errors.OutOfRangeError): + sess.run(n) + + def runSampleKeyPairsTest(self, ds, expected_key_pairs): + itr = ds.make_initializable_iterator() + n = itr.get_next() + with self.test_session() as sess: + self._writeCommonValues(sess) + sess.run(itr.initializer) + for i, elems in enumerate(expected_key_pairs): + output = sess.run(n) + self.assertEqual( + compat.as_bytes(elems[0]), compat.as_bytes(output[0]), + "Unequal key pair (first element) at step %d; want: %s, got %s" % + (i, compat.as_bytes(elems[0]), compat.as_bytes(output[0]))) + self.assertEqual( + compat.as_bytes(elems[1]), compat.as_bytes(output[1]), + "Unequal key pair (second element) at step %d; want: %s, got %s" % + (i, compat.as_bytes(elems[1]), compat.as_bytes(output[1]))) + with self.assertRaises(errors.OutOfRangeError): + sess.run(n) + + def testSampleKeyPairsSimplePrefix(self): + ds = bigtable_api._BigtableSampleKeyPairsDataset( + self._table, prefix="r", start="", end="") + expected_key_pairs = [("r", "r1"), ("r1", "r3"), ("r3", "s")] + self.runSampleKeyPairsTest(ds, expected_key_pairs) + + def testSampleKeyPairsSimpleRange(self): + ds = bigtable_api._BigtableSampleKeyPairsDataset( + self._table, prefix="", start="r1", end="r3") + expected_key_pairs = [("r1", "r3")] + self.runSampleKeyPairsTest(ds, expected_key_pairs) + + def testSampleKeyPairsSkipRangePrefix(self): + ds = bigtable_api._BigtableSampleKeyPairsDataset( + self._table, prefix="r2", start="", end="") + expected_key_pairs = [("r2", "r3")] + self.runSampleKeyPairsTest(ds, expected_key_pairs) + + def testSampleKeyPairsSkipRangeRange(self): + ds = bigtable_api._BigtableSampleKeyPairsDataset( + self._table, prefix="", start="r2", end="r3") + expected_key_pairs = [("r2", "r3")] + self.runSampleKeyPairsTest(ds, expected_key_pairs) + + def testSampleKeyPairsOffsetRanges(self): + ds = bigtable_api._BigtableSampleKeyPairsDataset( + self._table, prefix="", start="r2", end="r4") + expected_key_pairs = [("r2", "r3"), ("r3", "r4")] + self.runSampleKeyPairsTest(ds, expected_key_pairs) + + def testSampleKeyPairEverything(self): + ds = bigtable_api._BigtableSampleKeyPairsDataset( + self._table, prefix="", start="", end="") + expected_key_pairs = [("", "r1"), ("r1", "r3"), ("r3", "")] + self.runSampleKeyPairsTest(ds, expected_key_pairs) + + def testSampleKeyPairsPrefixAndStartKey(self): + ds = bigtable_api._BigtableSampleKeyPairsDataset( + self._table, prefix="r", start="r1", end="") + itr = ds.make_initializable_iterator() + with self.test_session() as sess: + with self.assertRaises(errors.InvalidArgumentError): + sess.run(itr.initializer) + + def testSampleKeyPairsPrefixAndEndKey(self): + ds = bigtable_api._BigtableSampleKeyPairsDataset( + self._table, prefix="r", start="", end="r3") + itr = ds.make_initializable_iterator() + with self.test_session() as sess: + with self.assertRaises(errors.InvalidArgumentError): + sess.run(itr.initializer) + + def testParallelScanPrefix(self): + ds = self._table.parallel_scan_prefix(prefix="r", cf1="c1") + itr = ds.make_initializable_iterator() + n = itr.get_next() + with self.test_session() as sess: + self._writeCommonValues(sess) + sess.run(itr.initializer) + expected_values = list(zip(self.COMMON_ROW_KEYS, self.COMMON_VALUES)) + actual_values = [] + for _ in range(len(expected_values)): + output = sess.run(n) + actual_values.append(output) + with self.assertRaises(errors.OutOfRangeError): + sess.run(n) + self.assertItemsEqual( + _ListOfTuplesOfStringsToBytes(expected_values), + _ListOfTuplesOfStringsToBytes(actual_values)) + + def testParallelScanRange(self): + ds = self._table.parallel_scan_range(start="r1", end="r4", cf1="c1") + itr = ds.make_initializable_iterator() + n = itr.get_next() + with self.test_session() as sess: + self._writeCommonValues(sess) + sess.run(itr.initializer) + expected_values = list(zip(self.COMMON_ROW_KEYS, self.COMMON_VALUES)) + actual_values = [] + for _ in range(len(expected_values)): + output = sess.run(n) + actual_values.append(output) + with self.assertRaises(errors.OutOfRangeError): + sess.run(n) + self.assertItemsEqual( + _ListOfTuplesOfStringsToBytes(expected_values), + _ListOfTuplesOfStringsToBytes(actual_values)) + + +if __name__ == "__main__": + test.main() diff --git a/tensorflow/contrib/bigtable/python/ops/__init__.py b/tensorflow/contrib/bigtable/python/ops/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..36d75b0d7068a650347a5e17f4727a5432d8752f --- /dev/null +++ b/tensorflow/contrib/bigtable/python/ops/__init__.py @@ -0,0 +1,20 @@ +# Copyright 2017 The TensorFlow Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================== + +"""This module contains the Python API for the Cloud Bigtable integration.""" + +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function diff --git a/tensorflow/contrib/bigtable/python/ops/bigtable_api.py b/tensorflow/contrib/bigtable/python/ops/bigtable_api.py new file mode 100644 index 0000000000000000000000000000000000000000..9f73b7223c64a23bd570ef342fe9cf89bdf8832c --- /dev/null +++ b/tensorflow/contrib/bigtable/python/ops/bigtable_api.py @@ -0,0 +1,741 @@ +# Copyright 2018 The TensorFlow Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================== +"""The Python API for TensorFlow's Bigtable integration. + +TensorFlow has support for reading from and writing to Cloud Bigtable. To use +the Bigtable TensorFlow integration, first create a BigtableClient (which +configures your connection to Cloud Bigtable), and then open a Table. The Table +object then allows you to create numerous @{tf.data.Dataset}s to read data, or +write a @{tf.data.Dataset} object to the underlying Bigtable Table. + +For background on Google Cloud Bigtable, see: https://cloud.google.com/bigtable. +""" + +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + +from six import iteritems +from six import string_types + +from tensorflow.contrib.bigtable.ops import gen_bigtable_ops +from tensorflow.contrib.data.python.ops import interleave_ops +from tensorflow.contrib.util import loader +from tensorflow.python.data.ops import dataset_ops +from tensorflow.python.data.util import nest +from tensorflow.python.framework import dtypes +from tensorflow.python.framework import ops +from tensorflow.python.framework import tensor_shape +from tensorflow.python.platform import resource_loader + +_bigtable_so = loader.load_op_library( + resource_loader.get_path_to_datafile("_bigtable.so")) + + +class BigtableClient(object): + """BigtableClient is the entrypoint for interacting with Cloud Bigtable in TF. + + BigtableClient encapsulates a connection to Cloud Bigtable, and exposes the + `table` method to open a Bigtable Table. + """ + + def __init__(self, + project_id, + instance_id, + connection_pool_size=None, + max_receive_message_size=None): + """Creates a BigtableClient that can be used to open connections to tables. + + Args: + project_id: A string representing the GCP project id to connect to. + instance_id: A string representing the Bigtable instance to connect to. + connection_pool_size: (Optional.) A number representing the number of + concurrent connections to the Cloud Bigtable service to make. + max_receive_message_size: (Optional.) The maximum bytes received in a + single gRPC response. + + Raises: + ValueError: if the arguments are invalid (e.g. wrong type, or out of + expected ranges (e.g. negative).) + """ + if not isinstance(project_id, str): + raise ValueError("`project_id` must be a string") + self._project_id = project_id + + if not isinstance(instance_id, str): + raise ValueError("`instance_id` must be a string") + self._instance_id = instance_id + + if connection_pool_size is None: + connection_pool_size = -1 + elif connection_pool_size < 1: + raise ValueError("`connection_pool_size` must be positive") + + if max_receive_message_size is None: + max_receive_message_size = -1 + elif max_receive_message_size < 1: + raise ValueError("`max_receive_message_size` must be positive") + + self._connection_pool_size = connection_pool_size + + self._resource = gen_bigtable_ops.bigtable_client( + project_id, instance_id, connection_pool_size, max_receive_message_size) + + def table(self, name, snapshot=None): + """Opens a table and returns a `BigTable` object. + + Args: + name: A `tf.string` `tf.Tensor` name of the table to open. + snapshot: Either a `tf.string` `tf.Tensor` snapshot id, or `True` to + request the creation of a snapshot. (Note: currently unimplemented.) + + Returns: + A `BigTable` python object representing the operations available on the + table. + """ + # TODO(saeta): Implement snapshot functionality. + table = gen_bigtable_ops.bigtable_table(self._resource, name) + return BigTable(name, snapshot, table) + + +class BigTable(object): + """BigTable is the entrypoint for reading and writing data in Cloud Bigtable. + + This BigTable class is the python representation of the Cloud Bigtable table + within TensorFlow. Methods on this class allow data to be read from and + written to the Cloud Bigtable service in flexible and high performance + manners. + """ + + # TODO(saeta): Investigate implementing tf.contrib.lookup.LookupInterface. + # TODO(saeta): Consider variant tensors instead of resources (while supporting + # connection pooling). + + def __init__(self, name, snapshot, resource): + self._name = name + self._snapshot = snapshot + self._resource = resource + + def lookup_columns(self, *args, **kwargs): + """Retrieves the values of columns for a dataset of keys. + + Example usage: + ``` + table = bigtable_client.table("my_table") + key_dataset = table.get_keys_prefix("imagenet") + images = key_dataset.apply(table.lookup_columns(("cf1", "image"), + ("cf2", "label"), + ("cf2", "boundingbox"))) + training_data = images.map(parse_and_crop, num_parallel_calls=64).batch(128) + ``` + + Alternatively, you can use keyword arguments to specify the columns to + capture. Example (same as above, rewritten): + ``` + table = bigtable_client.table("my_table") + key_dataset = table.get_keys_prefix("imagenet") + images = key_dataset.apply(table.lookup_columns( + cf1="image", cf2=("label", "boundingbox"))) + training_data = images.map(parse_and_crop, num_parallel_calls=64).batch(128) + ``` + + Note: certain kwargs keys are reserved, and thus some column families cannot + be identified using the kwargs syntax. Instead, please use the args syntax. + This list includes: + - 'name' + This list can change at any time. + + Args: + *args: A list of tuples containing (column family, column name) pairs. + **kwargs: Column families and + + Returns: + A function that can be passed to `tf.data.Dataset.apply` to retrieve the + values of columns for the rows. + """ + table = self # Capture self + normalized = args + if normalized is None: + normalized = [] + if isinstance(normalized, tuple): + normalized = list(normalized) + for key, value in iteritems(kwargs): + if key == "name": + continue + if isinstance(value, str): + normalized.append((key, value)) + continue + for col in value: + normalized.append((key, col)) + + def _apply_fn(dataset): + # TODO(saeta): Verify dataset's types are correct! + return _BigtableLookupDataset(dataset, table, normalized) + + return _apply_fn + + def keys_by_range_dataset(self, start, end): + """Retrieves all row keys between start and end. + + Note: it does NOT retrieve the values of columns. + + Args: + start: The start row key. The row keys for rows after start (inclusive) + will be retrieved. + end: (Optional.) The end row key. Rows up to (but not including) end will + be retrieved. If end is None, all subsequent row keys will be retrieved. + + Returns: + A @{tf.data.Dataset} containing `tf.string` Tensors corresponding to all + of the row keys between `start` and `end`. + """ + # TODO(saeta): Make inclusive / exclusive configurable? + if end is None: + end = "" + return _BigtableRangeKeyDataset(self, start, end) + + def keys_by_prefix_dataset(self, prefix): + """Retrieves the row keys matching a given prefix. + + Args: + prefix: All row keys that begin with `prefix` in the table will be + retrieved. + + Returns: + A @{tf.data.Dataset}. containing `tf.string` Tensors corresponding to all + of the row keys matching that prefix. + """ + return _BigtablePrefixKeyDataset(self, prefix) + + def sample_keys(self): + """Retrieves a sampling of row keys from the Bigtable table. + + This dataset is most often used in conjunction with + @{tf.contrib.data.parallel_interleave} to construct a set of ranges for + scanning in parallel. + + Returns: + A @{tf.data.Dataset} returning string row keys. + """ + return _BigtableSampleKeysDataset(self) + + def scan_prefix(self, prefix, probability=None, columns=None, **kwargs): + """Retrieves row (including values) from the Bigtable service. + + Rows with row-key prefixed by `prefix` will be retrieved. + + Specifying the columns to retrieve for each row is done by either using + kwargs or in the columns parameter. To retrieve values of the columns "c1", + and "c2" from the column family "cfa", and the value of the column "c3" + from column family "cfb", the following datasets (`ds1`, and `ds2`) are + equivalent: + + ``` + table = # ... + ds1 = table.scan_prefix("row_prefix", columns=[("cfa", "c1"), + ("cfa", "c2"), + ("cfb", "c3")]) + ds2 = table.scan_prefix("row_prefix", cfa=["c1", "c2"], cfb="c3") + ``` + + Note: only the latest value of a cell will be retrieved. + + Args: + prefix: The prefix all row keys must match to be retrieved for prefix- + based scans. + probability: (Optional.) A float between 0 (exclusive) and 1 (inclusive). + A non-1 value indicates to probabilistically sample rows with the + provided probability. + columns: The columns to read. Note: most commonly, they are expressed as + kwargs. Use the columns value if you are using column families that are + reserved. The value of columns and kwargs are merged. Columns is a list + of tuples of strings ("column_family", "column_qualifier"). + **kwargs: The column families and columns to read. Keys are treated as + column_families, and values can be either lists of strings, or strings + that are treated as the column qualifier (column name). + + Returns: + A @{tf.data.Dataset} returning the row keys and the cell contents. + + Raises: + ValueError: If the configured probability is unexpected. + """ + probability = _normalize_probability(probability) + normalized = _normalize_columns(columns, kwargs) + return _BigtableScanDataset(self, prefix, "", "", normalized, probability) + + def scan_range(self, start, end, probability=None, columns=None, **kwargs): + """Retrieves rows (including values) from the Bigtable service. + + Rows with row-keys between `start` and `end` will be retrieved. + + Specifying the columns to retrieve for each row is done by either using + kwargs or in the columns parameter. To retrieve values of the columns "c1", + and "c2" from the column family "cfa", and the value of the column "c3" + from column family "cfb", the following datasets (`ds1`, and `ds2`) are + equivalent: + + ``` + table = # ... + ds1 = table.scan_range("row_start", "row_end", columns=[("cfa", "c1"), + ("cfa", "c2"), + ("cfb", "c3")]) + ds2 = table.scan_range("row_start", "row_end", cfa=["c1", "c2"], cfb="c3") + ``` + + Note: only the latest value of a cell will be retrieved. + + Args: + start: The start of the range when scanning by range. + end: (Optional.) The end of the range when scanning by range. + probability: (Optional.) A float between 0 (exclusive) and 1 (inclusive). + A non-1 value indicates to probabilistically sample rows with the + provided probability. + columns: The columns to read. Note: most commonly, they are expressed as + kwargs. Use the columns value if you are using column families that are + reserved. The value of columns and kwargs are merged. Columns is a list + of tuples of strings ("column_family", "column_qualifier"). + **kwargs: The column families and columns to read. Keys are treated as + column_families, and values can be either lists of strings, or strings + that are treated as the column qualifier (column name). + + Returns: + A @{tf.data.Dataset} returning the row keys and the cell contents. + + Raises: + ValueError: If the configured probability is unexpected. + """ + probability = _normalize_probability(probability) + normalized = _normalize_columns(columns, kwargs) + return _BigtableScanDataset(self, "", start, end, normalized, probability) + + def parallel_scan_prefix(self, + prefix, + num_parallel_scans=None, + probability=None, + columns=None, + **kwargs): + """Retrieves row (including values) from the Bigtable service at high speed. + + Rows with row-key prefixed by `prefix` will be retrieved. This method is + similar to `scan_prefix`, but by constrast performs multiple sub-scans in + parallel in order to achieve higher performance. + + Note: The dataset produced by this method is not deterministic! + + Specifying the columns to retrieve for each row is done by either using + kwargs or in the columns parameter. To retrieve values of the columns "c1", + and "c2" from the column family "cfa", and the value of the column "c3" + from column family "cfb", the following datasets (`ds1`, and `ds2`) are + equivalent: + + ``` + table = # ... + ds1 = table.parallel_scan_prefix("row_prefix", columns=[("cfa", "c1"), + ("cfa", "c2"), + ("cfb", "c3")]) + ds2 = table.parallel_scan_prefix("row_prefix", cfa=["c1", "c2"], cfb="c3") + ``` + + Note: only the latest value of a cell will be retrieved. + + Args: + prefix: The prefix all row keys must match to be retrieved for prefix- + based scans. + num_parallel_scans: (Optional.) The number of concurrent scans against the + Cloud Bigtable instance. + probability: (Optional.) A float between 0 (exclusive) and 1 (inclusive). + A non-1 value indicates to probabilistically sample rows with the + provided probability. + columns: The columns to read. Note: most commonly, they are expressed as + kwargs. Use the columns value if you are using column families that are + reserved. The value of columns and kwargs are merged. Columns is a list + of tuples of strings ("column_family", "column_qualifier"). + **kwargs: The column families and columns to read. Keys are treated as + column_families, and values can be either lists of strings, or strings + that are treated as the column qualifier (column name). + + Returns: + A @{tf.data.Dataset} returning the row keys and the cell contents. + + Raises: + ValueError: If the configured probability is unexpected. + """ + probability = _normalize_probability(probability) + normalized = _normalize_columns(columns, kwargs) + ds = _BigtableSampleKeyPairsDataset(self, prefix, "", "") + return self._make_parallel_scan_dataset(ds, num_parallel_scans, probability, + normalized) + + def parallel_scan_range(self, + start, + end, + num_parallel_scans=None, + probability=None, + columns=None, + **kwargs): + """Retrieves rows (including values) from the Bigtable service. + + Rows with row-keys between `start` and `end` will be retrieved. This method + is similar to `scan_range`, but by constrast performs multiple sub-scans in + parallel in order to achieve higher performance. + + Note: The dataset produced by this method is not deterministic! + + Specifying the columns to retrieve for each row is done by either using + kwargs or in the columns parameter. To retrieve values of the columns "c1", + and "c2" from the column family "cfa", and the value of the column "c3" + from column family "cfb", the following datasets (`ds1`, and `ds2`) are + equivalent: + + ``` + table = # ... + ds1 = table.parallel_scan_range("row_start", + "row_end", + columns=[("cfa", "c1"), + ("cfa", "c2"), + ("cfb", "c3")]) + ds2 = table.parallel_scan_range("row_start", "row_end", + cfa=["c1", "c2"], cfb="c3") + ``` + + Note: only the latest value of a cell will be retrieved. + + Args: + start: The start of the range when scanning by range. + end: (Optional.) The end of the range when scanning by range. + num_parallel_scans: (Optional.) The number of concurrent scans against the + Cloud Bigtable instance. + probability: (Optional.) A float between 0 (exclusive) and 1 (inclusive). + A non-1 value indicates to probabilistically sample rows with the + provided probability. + columns: The columns to read. Note: most commonly, they are expressed as + kwargs. Use the columns value if you are using column families that are + reserved. The value of columns and kwargs are merged. Columns is a list + of tuples of strings ("column_family", "column_qualifier"). + **kwargs: The column families and columns to read. Keys are treated as + column_families, and values can be either lists of strings, or strings + that are treated as the column qualifier (column name). + + Returns: + A @{tf.data.Dataset} returning the row keys and the cell contents. + + Raises: + ValueError: If the configured probability is unexpected. + """ + probability = _normalize_probability(probability) + normalized = _normalize_columns(columns, kwargs) + ds = _BigtableSampleKeyPairsDataset(self, "", start, end) + return self._make_parallel_scan_dataset(ds, num_parallel_scans, probability, + normalized) + + def write(self, dataset, column_families, columns, timestamp=None): + """Writes a dataset to the table. + + Args: + dataset: A @{tf.data.Dataset} to be written to this table. It must produce + a list of number-of-columns+1 elements, all of which must be strings. + The first value will be used as the row key, and subsequent values will + be used as cell values for the corresponding columns from the + corresponding column_families and columns entries. + column_families: A @{tf.Tensor} of `tf.string`s corresponding to the + column names to store the dataset's elements into. + columns: A `tf.Tensor` of `tf.string`s corresponding to the column names + to store the dataset's elements into. + timestamp: (Optional.) An int64 timestamp to write all the values at. + Leave as None to use server-provided timestamps. + + Returns: + A @{tf.Operation} that can be run to perform the write. + + Raises: + ValueError: If there are unexpected or incompatible types, or if the + number of columns and column_families does not match the output of + `dataset`. + """ + if timestamp is None: + timestamp = -1 # Bigtable server provided timestamp. + for tensor_type in nest.flatten(dataset.output_types): + if tensor_type != dtypes.string: + raise ValueError("Not all elements of the dataset were `tf.string`") + for shape in nest.flatten(dataset.output_shapes): + if not shape.is_compatible_with(tensor_shape.scalar()): + raise ValueError("Not all elements of the dataset were scalars") + if len(column_families) != len(columns): + raise ValueError("len(column_families) != len(columns)") + if len(nest.flatten(dataset.output_types)) != len(columns) + 1: + raise ValueError("A column name must be specified for every component of " + "the dataset elements. (e.g.: len(columns) != " + "len(dataset.output_types))") + return gen_bigtable_ops.dataset_to_bigtable( + self._resource, + dataset._as_variant_tensor(), # pylint: disable=protected-access + column_families, + columns, + timestamp) + + def _make_parallel_scan_dataset(self, ds, num_parallel_scans, + normalized_probability, normalized_columns): + """Builds a parallel dataset from a given range. + + Args: + ds: A `_BigtableSampleKeyPairsDataset` returning ranges of keys to use. + num_parallel_scans: The number of concurrent parallel scans to use. + normalized_probability: A number between 0 and 1 for the keep probability. + normalized_columns: The column families and column qualifiers to retrieve. + + Returns: + A @{tf.data.Dataset} representing the result of the parallel scan. + """ + if num_parallel_scans is None: + num_parallel_scans = 50 + + ds = ds.shuffle(buffer_size=10000) # TODO(saeta): Make configurable. + + def _interleave_fn(start, end): + return _BigtableScanDataset( + self, + prefix="", + start=start, + end=end, + normalized=normalized_columns, + probability=normalized_probability) + + # Note prefetch_input_elements must be set in order to avoid rpc timeouts. + ds = ds.apply( + interleave_ops.parallel_interleave( + _interleave_fn, + cycle_length=num_parallel_scans, + sloppy=True, + prefetch_input_elements=1)) + return ds + + +def _normalize_probability(probability): + if probability is None: + probability = 1.0 + if isinstance(probability, float) and (probability <= 0.0 or + probability > 1.0): + raise ValueError("probability must be in the range (0, 1].") + return probability + + +def _normalize_columns(columns, provided_kwargs): + """Converts arguments (columns, and kwargs dict) to C++ representation. + + Args: + columns: a datastructure containing the column families and qualifier to + retrieve. Valid types include (1) None, (2) list of tuples, (3) a tuple of + strings. + provided_kwargs: a dictionary containing the column families and qualifiers + to retrieve + + Returns: + A list of pairs of column family+qualifier to retrieve. + + Raises: + ValueError: If there are no cells to retrieve or the columns are in an + incorrect format. + """ + normalized = columns + if normalized is None: + normalized = [] + if isinstance(normalized, tuple): + if len(normalized) == 2: + normalized = [normalized] + else: + raise ValueError("columns was a tuple of inappropriate length") + for key, value in iteritems(provided_kwargs): + if key == "name": + continue + if isinstance(value, string_types): + normalized.append((key, value)) + continue + for col in value: + normalized.append((key, col)) + if not normalized: + raise ValueError("At least one column + column family must be specified.") + return normalized + + +class _BigtableKeyDataset(dataset_ops.Dataset): + """_BigtableKeyDataset is an abstract class representing the keys of a table. + """ + + def __init__(self, table): + """Constructs a _BigtableKeyDataset. + + Args: + table: a Bigtable class. + """ + super(_BigtableKeyDataset, self).__init__() + self._table = table + + @property + def output_classes(self): + return ops.Tensor + + @property + def output_shapes(self): + return tensor_shape.TensorShape([]) + + @property + def output_types(self): + return dtypes.string + + +class _BigtablePrefixKeyDataset(_BigtableKeyDataset): + """_BigtablePrefixKeyDataset represents looking up keys by prefix. + """ + + def __init__(self, table, prefix): + super(_BigtablePrefixKeyDataset, self).__init__(table) + self._prefix = prefix + + def _as_variant_tensor(self): + return gen_bigtable_ops.bigtable_prefix_key_dataset( + table=self._table._resource, # pylint: disable=protected-access + prefix=self._prefix) + + +class _BigtableRangeKeyDataset(_BigtableKeyDataset): + """_BigtableRangeKeyDataset represents looking up keys by range. + """ + + def __init__(self, table, start, end): + super(_BigtableRangeKeyDataset, self).__init__(table) + self._start = start + self._end = end + + def _as_variant_tensor(self): + return gen_bigtable_ops.bigtable_range_key_dataset( + table=self._table._resource, # pylint: disable=protected-access + start_key=self._start, + end_key=self._end) + + +class _BigtableSampleKeysDataset(_BigtableKeyDataset): + """_BigtableSampleKeysDataset represents a sampling of row keys. + """ + + # TODO(saeta): Expose the data size offsets into the keys. + + def __init__(self, table): + super(_BigtableSampleKeysDataset, self).__init__(table) + + def _as_variant_tensor(self): + return gen_bigtable_ops.bigtable_sample_keys_dataset( + table=self._table._resource) # pylint: disable=protected-access + + +class _BigtableLookupDataset(dataset_ops.Dataset): + """_BigtableLookupDataset represents a dataset that retrieves values for keys. + """ + + def __init__(self, dataset, table, normalized): + self._num_outputs = len(normalized) + 1 # 1 for row key + self._dataset = dataset + self._table = table + self._normalized = normalized + self._column_families = [i[0] for i in normalized] + self._columns = [i[1] for i in normalized] + + @property + def output_classes(self): + return tuple([ops.Tensor] * self._num_outputs) + + @property + def output_shapes(self): + return tuple([tensor_shape.TensorShape([])] * self._num_outputs) + + @property + def output_types(self): + return tuple([dtypes.string] * self._num_outputs) + + def _as_variant_tensor(self): + # pylint: disable=protected-access + return gen_bigtable_ops.bigtable_lookup_dataset( + keys_dataset=self._dataset._as_variant_tensor(), + table=self._table._resource, + column_families=self._column_families, + columns=self._columns) + + +class _BigtableScanDataset(dataset_ops.Dataset): + """_BigtableScanDataset represents a dataset that retrieves keys and values. + """ + + def __init__(self, table, prefix, start, end, normalized, probability): + self._table = table + self._prefix = prefix + self._start = start + self._end = end + self._column_families = [i[0] for i in normalized] + self._columns = [i[1] for i in normalized] + self._probability = probability + self._num_outputs = len(normalized) + 1 # 1 for row key + + @property + def output_classes(self): + return tuple([ops.Tensor] * self._num_outputs) + + @property + def output_shapes(self): + return tuple([tensor_shape.TensorShape([])] * self._num_outputs) + + @property + def output_types(self): + return tuple([dtypes.string] * self._num_outputs) + + def _as_variant_tensor(self): + return gen_bigtable_ops.bigtable_scan_dataset( + table=self._table._resource, # pylint: disable=protected-access + prefix=self._prefix, + start_key=self._start, + end_key=self._end, + column_families=self._column_families, + columns=self._columns, + probability=self._probability) + + +class _BigtableSampleKeyPairsDataset(dataset_ops.Dataset): + """_BigtableKeyRangeDataset returns key pairs from the Bigtable. + """ + + def __init__(self, table, prefix, start, end): + self._table = table + self._prefix = prefix + self._start = start + self._end = end + + @property + def output_classes(self): + return (ops.Tensor, ops.Tensor) + + @property + def output_shapes(self): + return (tensor_shape.TensorShape([]), tensor_shape.TensorShape([])) + + @property + def output_types(self): + return (dtypes.string, dtypes.string) + + def _as_variant_tensor(self): + # pylint: disable=protected-access + return gen_bigtable_ops.bigtable_sample_key_pairs_dataset( + table=self._table._resource, + prefix=self._prefix, + start_key=self._start, + end_key=self._end) diff --git a/tensorflow/contrib/boosted_trees/estimator_batch/BUILD b/tensorflow/contrib/boosted_trees/estimator_batch/BUILD index 8cff1a3bb1d11aff6a264636291a7149b40de516..ef0e80cd0997bc0e95cd0d150e87db144a2dde44 100644 --- a/tensorflow/contrib/boosted_trees/estimator_batch/BUILD +++ b/tensorflow/contrib/boosted_trees/estimator_batch/BUILD @@ -15,8 +15,9 @@ py_library( srcs = ["__init__.py"], srcs_version = "PY2AND3", deps = [ - "custom_export_strategy", + ":custom_export_strategy", ":custom_loss_head", + ":distillation_loss", ":estimator", ":model", ":trainer_hooks", @@ -144,6 +145,7 @@ py_library( srcs = ["dnn_tree_combined_estimator.py"], srcs_version = "PY2AND3", deps = [ + ":distillation_loss", ":estimator_utils", ":trainer_hooks", "//tensorflow/contrib/boosted_trees:gbdt_batch", @@ -156,6 +158,17 @@ py_library( ], ) +py_library( + name = "distillation_loss", + srcs = ["distillation_loss.py"], + srcs_version = "PY2AND3", + deps = [ + "//tensorflow/contrib/learn", + "//tensorflow/python:math_ops", + "//tensorflow/python:nn", + ], +) + py_test( name = "dnn_tree_combined_estimator_test", size = "medium", diff --git a/tensorflow/contrib/boosted_trees/estimator_batch/distillation_loss.py b/tensorflow/contrib/boosted_trees/estimator_batch/distillation_loss.py new file mode 100644 index 0000000000000000000000000000000000000000..9aacc5534329d1302b25dcfab678f9adb8f773f6 --- /dev/null +++ b/tensorflow/contrib/boosted_trees/estimator_batch/distillation_loss.py @@ -0,0 +1,75 @@ +# Copyright 2018 The TensorFlow Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================== +"""Utill functions for distillation loss. + +The distillation loss_fn will be called with the following: + +Args: + dnn_logits: Tensor of logits from the dnn, treated as the "target". This will + be the output of a call to tf.stop_gradient(). + tree_logits: Tensor of logits from the tree, treated as the "predictions". + example_weights: Tensor of example weights, or a single scalar. + +Returns: + A scalar indicating the reduced loss for that batch of examples. + +Note: we calls the loss_fn defined in contrib head, which is computing two +losses, first one for training and second one for reporting. We only take the +first one here. +""" +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + +from tensorflow.contrib.learn.python.learn.estimators import head as head_lib +from tensorflow.python.ops import math_ops +from tensorflow.python.ops import nn + + +def _logits_to_label_for_tree(logits, n_classes): + if n_classes == 2: + return math_ops.sigmoid(logits) + else: + return nn.softmax(logits) + + +def create_dnn_to_tree_squared_loss_fn(n_classes): + """Returns a squared loss function for dnn to tree distillation.""" + + def _dnn_to_tree_squared_loss(dnn_logits, tree_logits, example_weights): + return head_lib._mean_squared_loss( # pylint: disable=protected-access + labels=_logits_to_label_for_tree(dnn_logits, n_classes), + logits=_logits_to_label_for_tree(tree_logits, n_classes), + weights=example_weights)[0] + + return _dnn_to_tree_squared_loss + + +def create_dnn_to_tree_cross_entropy_loss_fn(n_classes): + """Returns a cross entropy loss function for dnn to tree distillation.""" + + def _dnn_to_tree_cross_entropy_loss(dnn_logits, tree_logits, example_weights): + if n_classes == 2: + return head_lib._log_loss_with_two_classes( # pylint: disable=protected-access + labels=_logits_to_label_for_tree(dnn_logits, n_classes), + logits=tree_logits, + weights=example_weights)[0] + else: + return head_lib._softmax_cross_entropy_loss( # pylint: disable=protected-access + labels=_logits_to_label_for_tree(dnn_logits, n_classes), + logits=tree_logits, + weights=example_weights)[0] + + return _dnn_to_tree_cross_entropy_loss diff --git a/tensorflow/contrib/boosted_trees/estimator_batch/dnn_tree_combined_estimator.py b/tensorflow/contrib/boosted_trees/estimator_batch/dnn_tree_combined_estimator.py index 911d87fa10570382ee5f03edfc1bfd1d116c8360..7eb429b636a5193a124dd9b0c020dae6cac910cb 100644 --- a/tensorflow/contrib/boosted_trees/estimator_batch/dnn_tree_combined_estimator.py +++ b/tensorflow/contrib/boosted_trees/estimator_batch/dnn_tree_combined_estimator.py @@ -24,7 +24,9 @@ from __future__ import division from __future__ import print_function import six + from tensorflow.contrib import layers +from tensorflow.contrib.boosted_trees.estimator_batch import distillation_loss from tensorflow.contrib.boosted_trees.estimator_batch import estimator_utils from tensorflow.contrib.boosted_trees.estimator_batch import trainer_hooks from tensorflow.contrib.boosted_trees.python.ops import model_ops @@ -35,11 +37,13 @@ from tensorflow.contrib.learn.python.learn.estimators import head as head_lib from tensorflow.contrib.learn.python.learn.estimators import model_fn from tensorflow.python.feature_column import feature_column as feature_column_lib from tensorflow.python.framework import ops +from tensorflow.python.ops import array_ops from tensorflow.python.ops import control_flow_ops from tensorflow.python.ops import nn from tensorflow.python.ops import partitioned_variables from tensorflow.python.ops import state_ops from tensorflow.python.ops import variable_scope +from tensorflow.python.platform import tf_logging as logging from tensorflow.python.summary import summary from tensorflow.python.training import training_util @@ -77,6 +81,7 @@ def _dnn_tree_combined_model_fn(features, predict_with_tree_only=False, tree_feature_columns=None, tree_center_bias=False, + dnn_to_tree_distillation_param=None, use_core_versions=False): """DNN and GBDT combined model_fn. @@ -117,6 +122,13 @@ def _dnn_tree_combined_model_fn(features, set to True, these features are in addition to dnn_feature_columns. tree_center_bias: Whether a separate tree should be created for first fitting the bias. + dnn_to_tree_distillation_param: A Tuple of (float, loss_fn), where the + float defines the weight of the distillation loss, and the loss_fn, for + computing distillation loss, takes dnn_logits, tree_logits and weight + tensor. If the entire tuple is None, no distillation will be applied. If + only the loss_fn is None, we will take the sigmoid/softmax cross entropy + loss be default. When distillation is applied, `predict_with_tree_only` + will be set to True. use_core_versions: Whether feature columns and loss are from the core (as opposed to contrib) version of tensorflow. @@ -132,6 +144,12 @@ def _dnn_tree_combined_model_fn(features, if not dnn_feature_columns: raise ValueError("dnn_feature_columns must be specified") + if dnn_to_tree_distillation_param: + if not predict_with_tree_only: + logging.warning("update predict_with_tree_only to True since distillation" + "is specified.") + predict_with_tree_only = True + # Build DNN Logits. dnn_parent_scope = "dnn" dnn_partitioner = dnn_input_layer_partitioner or ( @@ -225,6 +243,25 @@ def _dnn_tree_combined_model_fn(features, def _tree_train_op_fn(loss): """Returns the op to optimize the loss.""" + if dnn_to_tree_distillation_param: + loss_weight, loss_fn = dnn_to_tree_distillation_param + weight_tensor = head_lib._weight_tensor( # pylint: disable=protected-access + features, head.weight_column_name) + dnn_logits_fixed = array_ops.stop_gradient(dnn_logits) + + if loss_fn is None: + # we create the loss_fn similar to the head loss_fn for + # multi_class_head used previously as the default one. + n_classes = 2 if head.logits_dimension == 1 else head.logits_dimension + loss_fn = distillation_loss.create_dnn_to_tree_cross_entropy_loss_fn( + n_classes) + + dnn_to_tree_distillation_loss = loss_weight * loss_fn( + dnn_logits_fixed, tree_logits, weight_tensor) + summary.scalar("dnn_to_tree_distillation_loss", + dnn_to_tree_distillation_loss) + loss += dnn_to_tree_distillation_loss + update_op = gbdt_model.train(loss, predictions_dict, labels) with ops.control_dependencies( [update_op]), (ops.colocate_with(global_step)): @@ -232,7 +269,7 @@ def _dnn_tree_combined_model_fn(features, return update_op if predict_with_tree_only: - if mode == model_fn.ModeKeys.TRAIN or mode == model_fn.ModeKeys.PREDICT: + if mode == model_fn.ModeKeys.TRAIN or mode == model_fn.ModeKeys.INFER: tree_train_logits = tree_logits else: tree_train_logits = control_flow_ops.cond( @@ -331,6 +368,7 @@ class DNNBoostedTreeCombinedClassifier(estimator.Estimator): predict_with_tree_only=False, tree_feature_columns=None, tree_center_bias=False, + dnn_to_tree_distillation_param=None, use_core_versions=False): """Initializes a DNNBoostedTreeCombinedClassifier instance. @@ -378,6 +416,13 @@ class DNNBoostedTreeCombinedClassifier(estimator.Estimator): set to True, these features are in addition to dnn_feature_columns. tree_center_bias: Whether a separate tree should be created for first fitting the bias. + dnn_to_tree_distillation_param: A Tuple of (float, loss_fn), where the + float defines the weight of the distillation loss, and the loss_fn, for + computing distillation loss, takes dnn_logits, tree_logits and weight + tensor. If the entire tuple is None, no distillation will be applied. If + only the loss_fn is None, we will take the sigmoid/softmax cross entropy + loss be default. When distillation is applied, `predict_with_tree_only` + will be set to True. use_core_versions: Whether feature columns and loss are from the core (as opposed to contrib) version of tensorflow. """ @@ -409,6 +454,7 @@ class DNNBoostedTreeCombinedClassifier(estimator.Estimator): predict_with_tree_only=predict_with_tree_only, tree_feature_columns=tree_feature_columns, tree_center_bias=tree_center_bias, + dnn_to_tree_distillation_param=dnn_to_tree_distillation_param, use_core_versions=use_core_versions) super(DNNBoostedTreeCombinedClassifier, self).__init__( @@ -442,6 +488,7 @@ class DNNBoostedTreeCombinedRegressor(estimator.Estimator): predict_with_tree_only=False, tree_feature_columns=None, tree_center_bias=False, + dnn_to_tree_distillation_param=None, use_core_versions=False): """Initializes a DNNBoostedTreeCombinedRegressor instance. @@ -489,6 +536,13 @@ class DNNBoostedTreeCombinedRegressor(estimator.Estimator): set to True, these features are in addition to dnn_feature_columns. tree_center_bias: Whether a separate tree should be created for first fitting the bias. + dnn_to_tree_distillation_param: A Tuple of (float, loss_fn), where the + float defines the weight of the distillation loss, and the loss_fn, for + computing distillation loss, takes dnn_logits, tree_logits and weight + tensor. If the entire tuple is None, no distillation will be applied. If + only the loss_fn is None, we will take the sigmoid/softmax cross entropy + loss be default. When distillation is applied, `predict_with_tree_only` + will be set to True. use_core_versions: Whether feature columns and loss are from the core (as opposed to contrib) version of tensorflow. """ @@ -525,6 +579,7 @@ class DNNBoostedTreeCombinedRegressor(estimator.Estimator): predict_with_tree_only=predict_with_tree_only, tree_feature_columns=tree_feature_columns, tree_center_bias=tree_center_bias, + dnn_to_tree_distillation_param=dnn_to_tree_distillation_param, use_core_versions=use_core_versions) super(DNNBoostedTreeCombinedRegressor, self).__init__( @@ -559,6 +614,7 @@ class DNNBoostedTreeCombinedEstimator(estimator.Estimator): predict_with_tree_only=False, tree_feature_columns=None, tree_center_bias=False, + dnn_to_tree_distillation_param=None, use_core_versions=False): """Initializes a DNNBoostedTreeCombinedEstimator instance. @@ -601,6 +657,13 @@ class DNNBoostedTreeCombinedEstimator(estimator.Estimator): set to True, these features are in addition to dnn_feature_columns. tree_center_bias: Whether a separate tree should be created for first fitting the bias. + dnn_to_tree_distillation_param: A Tuple of (float, loss_fn), where the + float defines the weight of the distillation loss, and the loss_fn, for + computing distillation loss, takes dnn_logits, tree_logits and weight + tensor. If the entire tuple is None, no distillation will be applied. If + only the loss_fn is None, we will take the sigmoid/softmax cross entropy + loss be default. When distillation is applied, `predict_with_tree_only` + will be set to True. use_core_versions: Whether feature columns and loss are from the core (as opposed to contrib) version of tensorflow. """ @@ -626,6 +689,7 @@ class DNNBoostedTreeCombinedEstimator(estimator.Estimator): predict_with_tree_only=predict_with_tree_only, tree_feature_columns=tree_feature_columns, tree_center_bias=tree_center_bias, + dnn_to_tree_distillation_param=dnn_to_tree_distillation_param, use_core_versions=use_core_versions) super(DNNBoostedTreeCombinedEstimator, self).__init__( diff --git a/tensorflow/contrib/boosted_trees/estimator_batch/dnn_tree_combined_estimator_test.py b/tensorflow/contrib/boosted_trees/estimator_batch/dnn_tree_combined_estimator_test.py index f495edc62f0909880c170ccb4cf5d11e3f20f55c..9b7acfa664b0398216b5a7fb904960d8363929d6 100644 --- a/tensorflow/contrib/boosted_trees/estimator_batch/dnn_tree_combined_estimator_test.py +++ b/tensorflow/contrib/boosted_trees/estimator_batch/dnn_tree_combined_estimator_test.py @@ -131,6 +131,30 @@ class DNNBoostedTreeCombinedTest(test_util.TensorFlowTestCase): classifier.fit(input_fn=_train_input_fn, steps=15) classifier.evaluate(input_fn=_eval_input_fn, steps=1) + def testFitAndEvaluateWithDistillation(self): + learner_config = learner_pb2.LearnerConfig() + learner_config.num_classes = 2 + learner_config.constraints.max_tree_depth = 1 + model_dir = tempfile.mkdtemp() + config = run_config.RunConfig() + + classifier = estimator.DNNBoostedTreeCombinedClassifier( + dnn_hidden_units=[1], + dnn_feature_columns=[feature_column.real_valued_column("x")], + tree_learner_config=learner_config, + num_trees=1, + tree_examples_per_layer=3, + n_classes=2, + model_dir=model_dir, + config=config, + dnn_steps_to_train=10, + dnn_input_layer_to_tree=False, + tree_feature_columns=[feature_column.real_valued_column("x")], + dnn_to_tree_distillation_param=(1, None)) + + classifier.fit(input_fn=_train_input_fn, steps=15) + classifier.evaluate(input_fn=_eval_input_fn, steps=1) + if __name__ == "__main__": googletest.main() diff --git a/tensorflow/contrib/boosted_trees/estimator_batch/estimator.py b/tensorflow/contrib/boosted_trees/estimator_batch/estimator.py index 9c36c302210185bc390751a0229a61f2f8cd91b8..59a78515c6c1fc98c879e590491433695bdd3445 100644 --- a/tensorflow/contrib/boosted_trees/estimator_batch/estimator.py +++ b/tensorflow/contrib/boosted_trees/estimator_batch/estimator.py @@ -269,3 +269,88 @@ class GradientBoostedDecisionTreeEstimator(estimator.Estimator): model_dir=model_dir, config=config, feature_engineering_fn=feature_engineering_fn) + + +class GradientBoostedDecisionTreeRanker(estimator.Estimator): + """A ranking estimator using gradient boosted decision trees.""" + + def __init__( + self, + learner_config, + examples_per_layer, + head, + ranking_model_pair_keys, + num_trees=None, + feature_columns=None, + weight_column_name=None, + model_dir=None, + config=None, + label_keys=None, + feature_engineering_fn=None, + logits_modifier_function=None, + center_bias=False, + use_core_libs=False, + output_leaf_index=False, + ): + """Initializes a GradientBoostedDecisionTreeRanker instance. + + This is an estimator that can be trained off the pairwise data and can be + used for inference on non-paired data. This is essentially LambdaMart. + Args: + learner_config: A config for the learner. + examples_per_layer: Number of examples to accumulate before growing a + layer. It can also be a function that computes the number of examples + based on the depth of the layer that's being built. + head: `Head` instance. + ranking_model_pair_keys: Keys to distinguish between features + for left and right part of the training pairs for ranking. For example, + for an Example with features "a.f1" and "b.f1", the keys would be + ("a", "b"). + num_trees: An int, number of trees to build. + feature_columns: A list of feature columns. + weight_column_name: Name of the column for weights, or None if not + weighted. + model_dir: Directory for model exports, etc. + config: `RunConfig` object to configure the runtime settings. + label_keys: Optional list of strings with size `[n_classes]` defining the + label vocabulary. Only supported for `n_classes` > 2. + feature_engineering_fn: Feature engineering function. Takes features and + labels which are the output of `input_fn` and returns features and + labels which will be fed into the model. + logits_modifier_function: A modifier function for the logits. + center_bias: Whether a separate tree should be created for first fitting + the bias. + use_core_libs: Whether feature columns and loss are from the core (as + opposed to contrib) version of tensorflow. + output_leaf_index: whether to output leaf indices along with predictions + during inference. The leaf node indexes are available in predictions + dict by the key 'leaf_index'. It is a Tensor of rank 2 and its shape is + [batch_size, num_trees]. + For example, + result_iter = classifier.predict(...) + for result_dict in result_iter: + # access leaf index list by result_dict["leaf_index"] + # which contains one leaf index per tree + + Raises: + ValueError: If learner_config is not valid. + """ + super(GradientBoostedDecisionTreeRanker, self).__init__( + model_fn=model.ranking_model_builder, + params={ + 'head': head, + 'n_classes': 2, + 'feature_columns': feature_columns, + 'learner_config': learner_config, + 'num_trees': num_trees, + 'weight_column_name': weight_column_name, + 'examples_per_layer': examples_per_layer, + 'center_bias': center_bias, + 'logits_modifier_function': logits_modifier_function, + 'use_core_libs': use_core_libs, + 'output_leaf_index': output_leaf_index, + 'ranking_model_pair_keys': ranking_model_pair_keys, + }, + model_dir=model_dir, + config=config, + feature_engineering_fn=feature_engineering_fn) diff --git a/tensorflow/contrib/boosted_trees/estimator_batch/estimator_test.py b/tensorflow/contrib/boosted_trees/estimator_batch/estimator_test.py index 75ef1b050028b6462b255827c06e836e5c481844..2c2dcb039d98c4793996800e73d7bb9c4d6e6b89 100644 --- a/tensorflow/contrib/boosted_trees/estimator_batch/estimator_test.py +++ b/tensorflow/contrib/boosted_trees/estimator_batch/estimator_test.py @@ -37,12 +37,31 @@ def _train_input_fn(): return features, label +def _ranking_train_input_fn(): + features = { + "a.f1": constant_op.constant([[3.], [0.3], [1.]]), + "a.f2": constant_op.constant([[0.1], [3.], [1.]]), + "b.f1": constant_op.constant([[13.], [0.4], [5.]]), + "b.f2": constant_op.constant([[1.], [3.], [0.01]]), + } + label = constant_op.constant([[0], [0], [1]], dtype=dtypes.int32) + return features, label + + def _eval_input_fn(): features = {"x": constant_op.constant([[1.], [2.], [2.]])} label = constant_op.constant([[0], [1], [1]], dtype=dtypes.int32) return features, label +def _infer_ranking_train_input_fn(): + features = { + "f1": constant_op.constant([[3.], [2], [1.]]), + "f2": constant_op.constant([[0.1], [3.], [1.]]) + } + return features, None + + class BoostedTreeEstimatorTest(test_util.TensorFlowTestCase): def setUp(self): @@ -155,6 +174,34 @@ class BoostedTreeEstimatorTest(test_util.TensorFlowTestCase): regressor.evaluate(input_fn=_eval_input_fn, steps=1) regressor.export(self._export_dir_base) + def testRankingDontThrowExceptionForForEstimator(self): + learner_config = learner_pb2.LearnerConfig() + learner_config.num_classes = 2 + learner_config.constraints.max_tree_depth = 1 + model_dir = tempfile.mkdtemp() + config = run_config.RunConfig() + + head_fn = head_lib._binary_logistic_head_with_sigmoid_cross_entropy_loss( + loss_reduction=losses.Reduction.SUM_OVER_BATCH_SIZE) + + model = estimator.GradientBoostedDecisionTreeRanker( + head=head_fn, + learner_config=learner_config, + num_trees=1, + examples_per_layer=3, + model_dir=model_dir, + config=config, + use_core_libs=True, + feature_columns=[ + core_feature_column.numeric_column("f1"), + core_feature_column.numeric_column("f2") + ], + ranking_model_pair_keys=("a", "b")) + + model.fit(input_fn=_ranking_train_input_fn, steps=1000) + model.evaluate(input_fn=_ranking_train_input_fn, steps=1) + model.predict(input_fn=_infer_ranking_train_input_fn) + if __name__ == "__main__": googletest.main() diff --git a/tensorflow/contrib/boosted_trees/estimator_batch/model.py b/tensorflow/contrib/boosted_trees/estimator_batch/model.py index 1ee891198939e53fc5913104b2c2e65dc977823f..0e8a56e6e9ec0e4b6f8e3cebd15d72fbf68dad32 100644 --- a/tensorflow/contrib/boosted_trees/estimator_batch/model.py +++ b/tensorflow/contrib/boosted_trees/estimator_batch/model.py @@ -20,6 +20,7 @@ from __future__ import print_function import copy +from tensorflow.contrib import learn from tensorflow.contrib.boosted_trees.estimator_batch import estimator_utils from tensorflow.contrib.boosted_trees.estimator_batch import trainer_hooks from tensorflow.contrib.boosted_trees.python.ops import model_ops @@ -28,7 +29,6 @@ from tensorflow.python.framework import ops from tensorflow.python.ops import state_ops from tensorflow.python.training import training_util - def model_builder(features, labels, mode, params, config): """Multi-machine batch gradient descent tree model. @@ -141,3 +141,184 @@ def model_builder(features, labels, mode, params, config): trainer_hooks.StopAfterNTrees(num_trees, attempted_trees, finalized_trees)) return model_fn_ops + + +def ranking_model_builder(features, labels, mode, params, config): + """Multi-machine batch gradient descent tree model for ranking. + + Args: + features: `Tensor` or `dict` of `Tensor` objects. + labels: Labels used to train on. + mode: Mode we are in. (TRAIN/EVAL/INFER) + params: A dict of hyperparameters. + The following hyperparameters are expected: + * head: A `Head` instance. + * learner_config: A config for the learner. + * feature_columns: An iterable containing all the feature columns used by + the model. + * examples_per_layer: Number of examples to accumulate before growing a + layer. It can also be a function that computes the number of examples + based on the depth of the layer that's being built. + * weight_column_name: The name of weight column. + * center_bias: Whether a separate tree should be created for first fitting + the bias. + * ranking_model_pair_keys (Optional): Keys to distinguish between features + for left and right part of the training pairs for ranking. For example, + for an Example with features "a.f1" and "b.f1", the keys would be + ("a", "b"). + config: `RunConfig` of the estimator. + + Returns: + A `ModelFnOps` object. + Raises: + ValueError: if inputs are not valid. + """ + head = params["head"] + learner_config = params["learner_config"] + examples_per_layer = params["examples_per_layer"] + feature_columns = params["feature_columns"] + weight_column_name = params["weight_column_name"] + num_trees = params["num_trees"] + use_core_libs = params["use_core_libs"] + logits_modifier_function = params["logits_modifier_function"] + output_leaf_index = params["output_leaf_index"] + ranking_model_pair_keys = params["ranking_model_pair_keys"] + + if features is None: + raise ValueError("At least one feature must be specified.") + + if config is None: + raise ValueError("Missing estimator RunConfig.") + + center_bias = params["center_bias"] + + if isinstance(features, ops.Tensor): + features = {features.name: features} + + # Make a shallow copy of features to ensure downstream usage + # is unaffected by modifications in the model function. + training_features = copy.copy(features) + training_features.pop(weight_column_name, None) + global_step = training_util.get_global_step() + with ops.device(global_step.device): + ensemble_handle = model_ops.tree_ensemble_variable( + stamp_token=0, + tree_ensemble_config="", # Initialize an empty ensemble. + name="ensemble_model") + + # Extract the features. + if mode == learn.ModeKeys.TRAIN or mode == learn.ModeKeys.EVAL: + # For ranking pairwise training, we extract two sets of features. + if len(ranking_model_pair_keys) != 2: + raise ValueError("You must provide keys for ranking.") + left_pair_key = ranking_model_pair_keys[0] + right_pair_key = ranking_model_pair_keys[1] + if left_pair_key is None or right_pair_key is None: + raise ValueError("Both pair keys should be provided for ranking.") + + features_1 = {} + features_2 = {} + for name in training_features: + feature = training_features[name] + new_name = name[2:] + if name.startswith(left_pair_key + "."): + features_1[new_name] = feature + else: + assert name.startswith(right_pair_key + ".") + features_2[new_name] = feature + + main_features = features_1 + supplementary_features = features_2 + else: + # For non-ranking or inference ranking, we have only 1 set of features. + main_features = training_features + + # Create GBDT model. + gbdt_model_main = gbdt_batch.GradientBoostedDecisionTreeModel( + is_chief=config.is_chief, + num_ps_replicas=config.num_ps_replicas, + ensemble_handle=ensemble_handle, + center_bias=center_bias, + examples_per_layer=examples_per_layer, + learner_config=learner_config, + feature_columns=feature_columns, + logits_dimension=head.logits_dimension, + features=main_features, + use_core_columns=use_core_libs, + output_leaf_index=output_leaf_index) + + with ops.name_scope("gbdt", "gbdt_optimizer"): + # Logits for inference. + if mode == learn.ModeKeys.INFER: + predictions_dict = gbdt_model_main.predict(mode) + logits = predictions_dict[gbdt_batch.PREDICTIONS] + if logits_modifier_function: + logits = logits_modifier_function(logits, features, mode) + else: + gbdt_model_supplementary = gbdt_batch.GradientBoostedDecisionTreeModel( + is_chief=config.is_chief, + num_ps_replicas=config.num_ps_replicas, + ensemble_handle=ensemble_handle, + center_bias=center_bias, + examples_per_layer=examples_per_layer, + learner_config=learner_config, + feature_columns=feature_columns, + logits_dimension=head.logits_dimension, + features=supplementary_features, + use_core_columns=use_core_libs, + output_leaf_index=output_leaf_index) + + # Logits for train and eval. + if not supplementary_features: + raise ValueError("Features for ranking must be specified.") + + predictions_dict_1 = gbdt_model_main.predict(mode) + predictions_1 = predictions_dict_1[gbdt_batch.PREDICTIONS] + + predictions_dict_2 = gbdt_model_supplementary.predict(mode) + predictions_2 = predictions_dict_2[gbdt_batch.PREDICTIONS] + + logits = predictions_1 - predictions_2 + if logits_modifier_function: + logits = logits_modifier_function(logits, features, mode) + + predictions_dict = predictions_dict_1 + predictions_dict[gbdt_batch.PREDICTIONS] = logits + + def _train_op_fn(loss): + """Returns the op to optimize the loss.""" + update_op = gbdt_model_main.train(loss, predictions_dict, labels) + with ops.control_dependencies( + [update_op]), (ops.colocate_with(global_step)): + update_op = state_ops.assign_add(global_step, 1).op + return update_op + + create_estimator_spec_op = getattr(head, "create_estimator_spec", None) + if use_core_libs and callable(create_estimator_spec_op): + model_fn_ops = head.create_estimator_spec( + features=features, + mode=mode, + labels=labels, + train_op_fn=_train_op_fn, + logits=logits) + model_fn_ops = estimator_utils.estimator_spec_to_model_fn_ops(model_fn_ops) + else: + model_fn_ops = head.create_model_fn_ops( + features=features, + mode=mode, + labels=labels, + train_op_fn=_train_op_fn, + logits=logits) + + if output_leaf_index and gbdt_batch.LEAF_INDEX in predictions_dict: + model_fn_ops.predictions[gbdt_batch.LEAF_INDEX] = predictions_dict[ + gbdt_batch.LEAF_INDEX] + if num_trees: + if center_bias: + num_trees += 1 + finalized_trees, attempted_trees = ( + gbdt_model_main.get_number_of_trees_tensor()) + model_fn_ops.training_hooks.append( + trainer_hooks.StopAfterNTrees(num_trees, attempted_trees, + finalized_trees)) + return model_fn_ops diff --git a/tensorflow/contrib/boosted_trees/lib/learner/batch/base_split_handler.py b/tensorflow/contrib/boosted_trees/lib/learner/batch/base_split_handler.py index 56ff00b39062d57c813633c98c765e077dd4c262..1b7f59ea4218355a13f1df7264352bd68503bd19 100644 --- a/tensorflow/contrib/boosted_trees/lib/learner/batch/base_split_handler.py +++ b/tensorflow/contrib/boosted_trees/lib/learner/batch/base_split_handler.py @@ -37,6 +37,7 @@ class BaseSplitHandler(object): gradient_shape, hessian_shape, multiclass_strategy, + loss_uses_sum_reduction=False, name=None): """Constructor for BaseSplitHandler. @@ -51,6 +52,8 @@ class BaseSplitHandler(object): gradient_shape: A TensorShape, containing shape of gradients. hessian_shape: A TensorShape, containing shape of hessians. multiclass_strategy: Strategy describing how to treat multiclass problems. + loss_uses_sum_reduction: A scalar boolean tensor that specifies whether + SUM or MEAN reduction was used for the loss. name: An optional handler name. """ self._l1_regularization = l1_regularization @@ -62,6 +65,7 @@ class BaseSplitHandler(object): self._multiclass_strategy = multiclass_strategy self._hessian_shape = hessian_shape self._gradient_shape = gradient_shape + self._loss_uses_sum_reduction = loss_uses_sum_reduction def scheduled_reads(self): """Returns the list of `ScheduledOp`s required for update_stats.""" diff --git a/tensorflow/contrib/boosted_trees/lib/learner/batch/categorical_split_handler.py b/tensorflow/contrib/boosted_trees/lib/learner/batch/categorical_split_handler.py index 9f78ab20242800fd8af7ad049d5970fbe26ec0ea..bf686237ff696dadad9713d26bf784d7442b80d0 100644 --- a/tensorflow/contrib/boosted_trees/lib/learner/batch/categorical_split_handler.py +++ b/tensorflow/contrib/boosted_trees/lib/learner/batch/categorical_split_handler.py @@ -23,6 +23,7 @@ from tensorflow.contrib.boosted_trees.python.ops import split_handler_ops from tensorflow.contrib.boosted_trees.python.ops import stats_accumulator_ops from tensorflow.python.framework import constant_op from tensorflow.python.framework import dtypes +from tensorflow.python.framework import ops from tensorflow.python.ops import array_ops from tensorflow.python.ops import control_flow_ops from tensorflow.python.ops import math_ops @@ -44,6 +45,7 @@ class EqualitySplitHandler(base_split_handler.BaseSplitHandler): hessian_shape, multiclass_strategy, init_stamp_token=0, + loss_uses_sum_reduction=False, name=None): """Initialize the internal state for this split handler. @@ -62,6 +64,8 @@ class EqualitySplitHandler(base_split_handler.BaseSplitHandler): multiclass_strategy: Strategy describing how to treat multiclass problems. init_stamp_token: A tensor containing an scalar for initial stamp of the stamped objects. + loss_uses_sum_reduction: A scalar boolean tensor that specifies whether + SUM or MEAN reduction was used for the loss. name: An optional handler name. """ super(EqualitySplitHandler, self).__init__( @@ -73,6 +77,7 @@ class EqualitySplitHandler(base_split_handler.BaseSplitHandler): gradient_shape=gradient_shape, hessian_shape=hessian_shape, multiclass_strategy=multiclass_strategy, + loss_uses_sum_reduction=loss_uses_sum_reduction, name=name) self._stats_accumulator = stats_accumulator_ops.StatsAccumulator( init_stamp_token, @@ -173,6 +178,11 @@ class EqualitySplitHandler(base_split_handler.BaseSplitHandler): # pair. num_minibatches, partition_ids, feature_ids, gradients, hessians = ( self._stats_accumulator.flush(stamp_token, next_stamp_token)) + # For sum_reduction, we don't need to divide by number of minibatches. + + num_minibatches = control_flow_ops.cond( + ops.convert_to_tensor(self._loss_uses_sum_reduction), + lambda: math_ops.to_int64(1), lambda: num_minibatches) partition_ids, gains, split_infos = ( split_handler_ops.build_categorical_equality_splits( num_minibatches=num_minibatches, @@ -187,7 +197,7 @@ class EqualitySplitHandler(base_split_handler.BaseSplitHandler): tree_complexity_regularization=self._tree_complexity_regularization, min_node_weight=self._min_node_weight, bias_feature_id=_BIAS_FEATURE_ID, - multiclass_strategy=self._multiclass_strategy,)) + multiclass_strategy=self._multiclass_strategy)) # There are no warm-up rounds needed in the equality column handler. So we # always return ready. are_splits_ready = constant_op.constant(True) diff --git a/tensorflow/contrib/boosted_trees/lib/learner/batch/categorical_split_handler_test.py b/tensorflow/contrib/boosted_trees/lib/learner/batch/categorical_split_handler_test.py index 0b65eba2a76273a81f1464ed7639f0c0760e0050..ef253e7cec4e8a96b360ced32b59398c2e2c9680 100644 --- a/tensorflow/contrib/boosted_trees/lib/learner/batch/categorical_split_handler_test.py +++ b/tensorflow/contrib/boosted_trees/lib/learner/batch/categorical_split_handler_test.py @@ -90,7 +90,17 @@ class EqualitySplitHandlerTest(test_util.TensorFlowTestCase): empty_hessians, example_weights, is_active=array_ops.constant([True, True])) - with ops.control_dependencies([update_1]): + update_2 = split_handler.update_stats_sync( + 0, + partition_ids, + gradients, + hessians, + empty_gradients, + empty_hessians, + example_weights, + is_active=array_ops.constant([True, True])) + + with ops.control_dependencies([update_1, update_2]): are_splits_ready, partitions, gains, splits = ( split_handler.make_splits(0, 1, class_id)) are_splits_ready, partitions, gains, splits = (sess.run( @@ -159,6 +169,129 @@ class EqualitySplitHandlerTest(test_util.TensorFlowTestCase): self.assertEqual(1, split_node.feature_id) + def testGenerateFeatureSplitCandidatesSumReduction(self): + with self.test_session() as sess: + # The data looks like the following: + # Example | Gradients | Partition | Feature ID | + # i0 | (0.2, 0.12) | 0 | 1,2 | + # i1 | (-0.5, 0.07) | 0 | | + # i2 | (1.2, 0.2) | 0 | 2 | + # i3 | (4.0, 0.13) | 1 | 1 | + gradients = array_ops.constant([0.2, -0.5, 1.2, 4.0]) + hessians = array_ops.constant([0.12, 0.07, 0.2, 0.13]) + partition_ids = [0, 0, 0, 1] + indices = [[0, 0], [0, 1], [2, 0], [3, 0]] + values = array_ops.constant([1, 2, 2, 1], dtype=dtypes.int64) + + gradient_shape = tensor_shape.scalar() + hessian_shape = tensor_shape.scalar() + class_id = -1 + + split_handler = categorical_split_handler.EqualitySplitHandler( + l1_regularization=0.1, + l2_regularization=1, + tree_complexity_regularization=0, + min_node_weight=0, + sparse_int_column=sparse_tensor.SparseTensor(indices, values, [4, 1]), + feature_column_group_id=0, + gradient_shape=gradient_shape, + hessian_shape=hessian_shape, + multiclass_strategy=learner_pb2.LearnerConfig.TREE_PER_CLASS, + init_stamp_token=0, + loss_uses_sum_reduction=True) + resources.initialize_resources(resources.shared_resources()).run() + + empty_gradients, empty_hessians = get_empty_tensors( + gradient_shape, hessian_shape) + example_weights = array_ops.ones([4, 1], dtypes.float32) + + update_1 = split_handler.update_stats_sync( + 0, + partition_ids, + gradients, + hessians, + empty_gradients, + empty_hessians, + example_weights, + is_active=array_ops.constant([True, True])) + update_2 = split_handler.update_stats_sync( + 0, + partition_ids, + gradients, + hessians, + empty_gradients, + empty_hessians, + example_weights, + is_active=array_ops.constant([True, True])) + with ops.control_dependencies([update_1, update_2]): + are_splits_ready, partitions, gains, splits = ( + split_handler.make_splits(0, 1, class_id)) + are_splits_ready, partitions, gains, splits = ( + sess.run([are_splits_ready, partitions, gains, splits])) + self.assertTrue(are_splits_ready) + self.assertAllEqual([0, 1], partitions) + + # Check the split on partition 0. + # -(0.4 + 2.4 - 0.1) / (0.24 + 0.4 + 1) + expected_left_weight = -1.6463414634146338 + + # (0.4 + 2.4 - 0.1) ** 2 / (0.24 + 0.4 + 1) + expected_left_gain = 4.445121951219511 + + # -(-1 + 0.1) / (0.14 + 1) + expected_right_weight = 0.789473684211 + + # (-1 + 0.1) ** 2 / (0.14 + 1) + expected_right_gain = 0.710526315789 + + # (0.4 + -1 + 2.4 - 0.1) ** 2 / (0.24 + 0.14 + 0.4 + 1) + expected_bias_gain = 1.6235955056179772 + + split_info = split_info_pb2.SplitInfo() + split_info.ParseFromString(splits[0]) + left_child = split_info.left_child.vector + right_child = split_info.right_child.vector + split_node = split_info.split_node.categorical_id_binary_split + + self.assertEqual(0, split_node.feature_column) + + self.assertEqual(2, split_node.feature_id) + + self.assertAllClose( + expected_left_gain + expected_right_gain - expected_bias_gain, gains[0], + 0.00001) + + self.assertAllClose([expected_left_weight], left_child.value, 0.00001) + + self.assertAllClose([expected_right_weight], right_child.value, 0.00001) + + # Check the split on partition 1. + # (-8 + 0.1) / (0.26 + 1) + expected_left_weight = -6.26984126984 + # (-8 + 0.1) ** 2 / (0.26 + 1) + expected_left_gain = 49.5317460317 + expected_right_weight = 0 + expected_right_gain = 0 + # (-8 + 0.1) ** 2 / (0.26 + 1) + expected_bias_gain = 49.5317460317 + + # Verify candidate for partition 1, there's only one active feature here + # so zero gain is expected. + split_info = split_info_pb2.SplitInfo() + split_info.ParseFromString(splits[1]) + left_child = split_info.left_child.vector + right_child = split_info.right_child.vector + split_node = split_info.split_node.categorical_id_binary_split + self.assertAllClose(0.0, gains[1], 0.00001) + + self.assertAllClose([expected_left_weight], left_child.value, 0.00001) + + self.assertAllClose([expected_right_weight], right_child.value, 0.00001) + + self.assertEqual(0, split_node.feature_column) + + self.assertEqual(1, split_node.feature_id) + def testGenerateFeatureSplitCandidatesMulticlass(self): with self.test_session() as sess: # Batch size is 4, 2 gradients per each instance. diff --git a/tensorflow/contrib/boosted_trees/lib/learner/batch/ordinal_split_handler.py b/tensorflow/contrib/boosted_trees/lib/learner/batch/ordinal_split_handler.py index 409a2d8f46c331c13aec10542c4967d50575e94a..df0bec1fe363e07bbff6b059e86076239bd605e9 100644 --- a/tensorflow/contrib/boosted_trees/lib/learner/batch/ordinal_split_handler.py +++ b/tensorflow/contrib/boosted_trees/lib/learner/batch/ordinal_split_handler.py @@ -99,6 +99,7 @@ class InequalitySplitHandler(base_split_handler.BaseSplitHandler): hessian_shape, multiclass_strategy, init_stamp_token=0, + loss_uses_sum_reduction=False, name=None): """Initialize the internal state for this split handler. @@ -117,6 +118,8 @@ class InequalitySplitHandler(base_split_handler.BaseSplitHandler): multiclass_strategy: Strategy describing how to treat multiclass problems. init_stamp_token: A tensor containing an scalar for initial stamp of the stamped objects. + loss_uses_sum_reduction: A scalar boolean tensor that specifies whether + SUM or MEAN reduction was used for the loss. name: An optional handler name. """ super(InequalitySplitHandler, self).__init__( @@ -128,7 +131,8 @@ class InequalitySplitHandler(base_split_handler.BaseSplitHandler): feature_column_group_id=feature_column_group_id, gradient_shape=gradient_shape, hessian_shape=hessian_shape, - multiclass_strategy=multiclass_strategy) + multiclass_strategy=multiclass_strategy, + loss_uses_sum_reduction=loss_uses_sum_reduction) self._stats_accumulator = stats_accumulator_ops.StatsAccumulator( init_stamp_token, gradient_shape, @@ -160,6 +164,7 @@ class DenseSplitHandler(InequalitySplitHandler): hessian_shape, multiclass_strategy, init_stamp_token=0, + loss_uses_sum_reduction=False, name=None): """Initialize the internal state for this split handler. @@ -179,6 +184,8 @@ class DenseSplitHandler(InequalitySplitHandler): multiclass_strategy: Strategy describing how to treat multiclass problems. init_stamp_token: A tensor containing an scalar for initial stamp of the stamped objects. + loss_uses_sum_reduction: A scalar boolean tensor that specifies whether + SUM or MEAN reduction was used for the loss. name: An optional handler name. """ super(DenseSplitHandler, self).__init__( @@ -193,7 +200,8 @@ class DenseSplitHandler(InequalitySplitHandler): name=name, gradient_shape=gradient_shape, hessian_shape=hessian_shape, - multiclass_strategy=multiclass_strategy) + multiclass_strategy=multiclass_strategy, + loss_uses_sum_reduction=loss_uses_sum_reduction) self._dense_float_column = dense_float_column # Register dense_make_stats_update function as an Op to the graph. g = ops.get_default_graph() @@ -255,15 +263,15 @@ class DenseSplitHandler(InequalitySplitHandler): next_stamp_token, self._multiclass_strategy, class_id, self._feature_column_group_id, self._l1_regularization, self._l2_regularization, self._tree_complexity_regularization, - self._min_node_weight)) + self._min_node_weight, self._loss_uses_sum_reduction)) return are_splits_ready, partition_ids, gains, split_infos -def _make_dense_split(quantile_accumulator_handle, stats_accumulator_handle, - stamp_token, next_stamp_token, multiclass_strategy, - class_id, feature_column_id, l1_regularization, - l2_regularization, tree_complexity_regularization, - min_node_weight, is_multi_dimentional): +def _make_dense_split( + quantile_accumulator_handle, stats_accumulator_handle, stamp_token, + next_stamp_token, multiclass_strategy, class_id, feature_column_id, + l1_regularization, l2_regularization, tree_complexity_regularization, + min_node_weight, is_multi_dimentional, loss_uses_sum_reduction): """Function that builds splits for a dense feature column.""" # Get the bucket boundaries are_splits_ready, buckets = ( @@ -291,7 +299,10 @@ def _make_dense_split(quantile_accumulator_handle, stats_accumulator_handle, num_minibatches, partition_ids, bucket_ids, gradients, hessians = ( gen_stats_accumulator_ops.stats_accumulator_scalar_flush( stats_accumulator_handle, stamp_token, next_stamp_token)) - + # For sum_reduction, we don't need to divide by number of minibatches. + num_minibatches = control_flow_ops.cond(loss_uses_sum_reduction, + lambda: math_ops.to_int64(1), + lambda: num_minibatches) # Put quantile and stats accumulator flushing in the dependency path. with ops.control_dependencies([flush_quantiles, partition_ids]): are_splits_ready = array_ops.identity(are_splits_ready) @@ -329,6 +340,7 @@ class SparseSplitHandler(InequalitySplitHandler): hessian_shape, multiclass_strategy, init_stamp_token=0, + loss_uses_sum_reduction=False, name=None): """Initialize the internal state for this split handler. @@ -348,6 +360,8 @@ class SparseSplitHandler(InequalitySplitHandler): multiclass_strategy: Strategy describing how to treat multiclass problems. init_stamp_token: A tensor containing an scalar for initial stamp of the stamped objects. + loss_uses_sum_reduction: A scalar boolean tensor that specifies whether + SUM or MEAN reduction was used for the loss. name: An optional handler name. """ super(SparseSplitHandler, self).__init__( @@ -362,6 +376,7 @@ class SparseSplitHandler(InequalitySplitHandler): hessian_shape=hessian_shape, multiclass_strategy=multiclass_strategy, init_stamp_token=init_stamp_token, + loss_uses_sum_reduction=loss_uses_sum_reduction, name=name) self._sparse_float_column = sparse_float_column @@ -424,15 +439,15 @@ class SparseSplitHandler(InequalitySplitHandler): next_stamp_token, self._multiclass_strategy, class_id, self._feature_column_group_id, self._l1_regularization, self._l2_regularization, self._tree_complexity_regularization, - self._min_node_weight)) + self._min_node_weight, self._loss_uses_sum_reduction)) return are_splits_ready, partition_ids, gains, split_infos -def _make_sparse_split(quantile_accumulator_handle, stats_accumulator_handle, - stamp_token, next_stamp_token, multiclass_strategy, - class_id, feature_column_id, l1_regularization, - l2_regularization, tree_complexity_regularization, - min_node_weight, is_multi_dimentional): +def _make_sparse_split( + quantile_accumulator_handle, stats_accumulator_handle, stamp_token, + next_stamp_token, multiclass_strategy, class_id, feature_column_id, + l1_regularization, l2_regularization, tree_complexity_regularization, + min_node_weight, is_multi_dimentional, loss_uses_sum_reduction): """Function that builds splits for a sparse feature column.""" # Get the bucket boundaries are_splits_ready, buckets = ( @@ -460,7 +475,9 @@ def _make_sparse_split(quantile_accumulator_handle, stats_accumulator_handle, num_minibatches, partition_ids, bucket_ids, gradients, hessians = ( gen_stats_accumulator_ops.stats_accumulator_scalar_flush( stats_accumulator_handle, stamp_token, next_stamp_token)) - + num_minibatches = control_flow_ops.cond(loss_uses_sum_reduction, + lambda: math_ops.to_int64(1), + lambda: num_minibatches) # Put quantile and stats accumulator flushing in the dependency path. with ops.control_dependencies([flush_quantiles, partition_ids]): are_splits_ready = array_ops.identity(are_splits_ready) @@ -498,17 +515,18 @@ def _specialize_make_split(func, is_multi_dimentional): dtypes.float32, dtypes.float32, dtypes.float32, + dtypes.bool, noinline=True) def f(quantile_accumulator_handle, stats_accumulator_handle, stamp_token, next_stamp_token, multiclass_strategy, class_id, feature_column_id, l1_regularization, l2_regularization, tree_complexity_regularization, - min_node_weight): + min_node_weight, loss_uses_sum_reduction): """Function that builds splits for a sparse feature column.""" - return func( - quantile_accumulator_handle, stats_accumulator_handle, stamp_token, - next_stamp_token, multiclass_strategy, class_id, feature_column_id, - l1_regularization, l2_regularization, tree_complexity_regularization, - min_node_weight, is_multi_dimentional) + return func(quantile_accumulator_handle, stats_accumulator_handle, + stamp_token, next_stamp_token, multiclass_strategy, class_id, + feature_column_id, l1_regularization, l2_regularization, + tree_complexity_regularization, min_node_weight, + is_multi_dimentional, loss_uses_sum_reduction) return f diff --git a/tensorflow/contrib/boosted_trees/lib/learner/batch/ordinal_split_handler_test.py b/tensorflow/contrib/boosted_trees/lib/learner/batch/ordinal_split_handler_test.py index 2f2c2302113bf59d6a065d5005c934dc76c2148d..d59732cf92eb85e88732ac5a17dccf475ae5342f 100644 --- a/tensorflow/contrib/boosted_trees/lib/learner/batch/ordinal_split_handler_test.py +++ b/tensorflow/contrib/boosted_trees/lib/learner/batch/ordinal_split_handler_test.py @@ -182,6 +182,144 @@ class DenseSplitHandlerTest(test_util.TensorFlowTestCase): self.assertAllClose(0.52, split_node.threshold, 0.00001) + def testGenerateFeatureSplitCandidatesLossUsesSumReduction(self): + with self.test_session() as sess: + # The data looks like the following: + # Example | Gradients | Partition | Dense Quantile | + # i0 | (0.2, 0.12) | 0 | 1 | + # i1 | (-0.5, 0.07) | 0 | 1 | + # i2 | (1.2, 0.2) | 0 | 0 | + # i3 | (4.0, 0.13) | 1 | 1 | + dense_column = array_ops.constant([0.52, 0.52, 0.3, 0.52]) + gradients = array_ops.constant([0.2, -0.5, 1.2, 4.0]) + hessians = array_ops.constant([0.12, 0.07, 0.2, 0.13]) + partition_ids = array_ops.constant([0, 0, 0, 1], dtype=dtypes.int32) + class_id = -1 + + gradient_shape = tensor_shape.scalar() + hessian_shape = tensor_shape.scalar() + split_handler = ordinal_split_handler.DenseSplitHandler( + l1_regularization=0.2, + l2_regularization=2., + tree_complexity_regularization=0., + min_node_weight=0., + epsilon=0.001, + num_quantiles=10, + feature_column_group_id=0, + dense_float_column=dense_column, + init_stamp_token=0, + gradient_shape=gradient_shape, + hessian_shape=hessian_shape, + multiclass_strategy=learner_pb2.LearnerConfig.TREE_PER_CLASS, + loss_uses_sum_reduction=True) + resources.initialize_resources(resources.shared_resources()).run() + + empty_gradients, empty_hessians = get_empty_tensors( + gradient_shape, hessian_shape) + example_weights = array_ops.ones([4, 1], dtypes.float32) + + update_1 = split_handler.update_stats_sync( + 0, + partition_ids, + gradients, + hessians, + empty_gradients, + empty_hessians, + example_weights, + is_active=array_ops.constant([True, True])) + with ops.control_dependencies([update_1]): + are_splits_ready = split_handler.make_splits( + np.int64(0), np.int64(1), class_id)[0] + + with ops.control_dependencies([are_splits_ready]): + update_2 = split_handler.update_stats_sync( + 1, + partition_ids, + gradients, + hessians, + empty_gradients, + empty_hessians, + example_weights, + is_active=array_ops.constant([True, True])) + update_3 = split_handler.update_stats_sync( + 1, + partition_ids, + gradients, + hessians, + empty_gradients, + empty_hessians, + example_weights, + is_active=array_ops.constant([True, True])) + with ops.control_dependencies([update_2, update_3]): + are_splits_ready2, partitions, gains, splits = ( + split_handler.make_splits(np.int64(1), np.int64(2), class_id)) + are_splits_ready, are_splits_ready2, partitions, gains, splits = ( + sess.run([ + are_splits_ready, are_splits_ready2, partitions, gains, splits + ])) + + # During the first iteration, inequality split handlers are not going to + # have any splits. Make sure that we return not_ready in that case. + self.assertFalse(are_splits_ready) + self.assertTrue(are_splits_ready2) + + self.assertAllEqual([0, 1], partitions) + + # Check the split on partition 0. + # -(2.4 - 0.2) / (0.4 + 2) + expected_left_weight = -0.91666 + + # expected_left_weight * -(2.4 - 0.2) + expected_left_gain = 2.016666666666666 + + # -(-1 + 0.4 + 0.2) / (0.38 + 2) + expected_right_weight = 0.1680672 + + # expected_right_weight * -(-1 + 0.4 + 0.2) + expected_right_gain = 0.0672268907563025 + + # (0.2 + -0.5 + 1.2 - 0.1) ** 2 / (0.12 + 0.07 + 0.2 + 1) + expected_bias_gain = 0.9208633093525178 + + split_info = split_info_pb2.SplitInfo() + split_info.ParseFromString(splits[0]) + left_child = split_info.left_child.vector + right_child = split_info.right_child.vector + split_node = split_info.split_node.dense_float_binary_split + self.assertAllClose( + expected_left_gain + expected_right_gain - expected_bias_gain, gains[0], + 0.00001) + + self.assertAllClose([expected_left_weight], left_child.value, 0.00001) + + self.assertAllClose([expected_right_weight], right_child.value, 0.00001) + + self.assertEqual(0, split_node.feature_column) + + self.assertAllClose(0.3, split_node.threshold, 0.00001) + + # Check the split on partition 1. + # (-8 + 0.2) / (0.26 + 2) + expected_left_weight = -3.4513274336283186 + expected_right_weight = 0 + + # Verify candidate for partition 1, there's only one active bucket here + # so zero gain is expected. + split_info = split_info_pb2.SplitInfo() + split_info.ParseFromString(splits[1]) + left_child = split_info.left_child.vector + right_child = split_info.right_child.vector + split_node = split_info.split_node.dense_float_binary_split + self.assertAllClose(0.0, gains[1], 0.00001) + + self.assertAllClose([expected_left_weight], left_child.value, 0.00001) + + self.assertAllClose([expected_right_weight], right_child.value, 0.00001) + + self.assertEqual(0, split_node.feature_column) + + self.assertAllClose(0.52, split_node.threshold, 0.00001) + def testGenerateFeatureSplitCandidatesMulticlassFullHessian(self): with self.test_session() as sess: dense_column = array_ops.constant([0.52, 0.52, 0.3, 0.52]) @@ -798,6 +936,139 @@ class SparseSplitHandlerTest(test_util.TensorFlowTestCase): self.assertAllClose(0.52, split_node.split.threshold) + def testGenerateFeatureSplitCandidatesLossUsesSumReduction(self): + with self.test_session() as sess: + # The data looks like the following: + # Example | Gradients | Partition | Sparse Quantile | + # i0 | (0.2, 0.12) | 0 | 1 | + # i1 | (-0.5, 0.07) | 0 | N/A | + # i2 | (1.2, 0.2) | 0 | 0 | + # i3 | (4.0, 0.13) | 1 | 1 | + gradients = array_ops.constant([0.2, -0.5, 1.2, 4.0]) + hessians = array_ops.constant([0.12, 0.07, 0.2, 0.13]) + example_partitions = array_ops.constant([0, 0, 0, 1], dtype=dtypes.int32) + indices = array_ops.constant([[0, 0], [2, 0], [3, 0]], dtype=dtypes.int64) + values = array_ops.constant([0.52, 0.3, 0.52]) + sparse_column = sparse_tensor.SparseTensor(indices, values, [4, 1]) + + gradient_shape = tensor_shape.scalar() + hessian_shape = tensor_shape.scalar() + class_id = -1 + + split_handler = ordinal_split_handler.SparseSplitHandler( + l1_regularization=0.0, + l2_regularization=4.0, + tree_complexity_regularization=0.0, + min_node_weight=0.0, + epsilon=0.01, + num_quantiles=2, + feature_column_group_id=0, + sparse_float_column=sparse_column, + init_stamp_token=0, + gradient_shape=gradient_shape, + hessian_shape=hessian_shape, + multiclass_strategy=learner_pb2.LearnerConfig.TREE_PER_CLASS, + loss_uses_sum_reduction=True) + resources.initialize_resources(resources.shared_resources()).run() + + empty_gradients, empty_hessians = get_empty_tensors( + gradient_shape, hessian_shape) + example_weights = array_ops.ones([4, 1], dtypes.float32) + + update_1 = split_handler.update_stats_sync( + 0, + example_partitions, + gradients, + hessians, + empty_gradients, + empty_hessians, + example_weights, + is_active=array_ops.constant([True, True])) + with ops.control_dependencies([update_1]): + are_splits_ready = split_handler.make_splits( + np.int64(0), np.int64(1), class_id)[0] + with ops.control_dependencies([are_splits_ready]): + update_2 = split_handler.update_stats_sync( + 1, + example_partitions, + gradients, + hessians, + empty_gradients, + empty_hessians, + example_weights, + is_active=array_ops.constant([True, True])) + update_3 = split_handler.update_stats_sync( + 1, + example_partitions, + gradients, + hessians, + empty_gradients, + empty_hessians, + example_weights, + is_active=array_ops.constant([True, True])) + with ops.control_dependencies([update_2, update_3]): + are_splits_ready2, partitions, gains, splits = ( + split_handler.make_splits(np.int64(1), np.int64(2), class_id)) + are_splits_ready, are_splits_ready2, partitions, gains, splits = ( + sess.run([ + are_splits_ready, are_splits_ready2, partitions, gains, splits + ])) + + # During the first iteration, inequality split handlers are not going to + # have any splits. Make sure that we return not_ready in that case. + self.assertFalse(are_splits_ready) + self.assertTrue(are_splits_ready2) + + self.assertAllEqual([0, 1], partitions) + # Check the split on partition 0. + # -(0.4 + 2.4) / (0.24 + 0.4 + 4) + expected_left_weight = -0.603448275862069 + # (0.4 + 2.4) ** 2 / (0.24 + 0.4 + 4) + expected_left_gain = 1.689655172413793 + # 1 / (0.14 + 4) + expected_right_weight = 0.24154589371980678 + # 1 ** 2 / (0.14 + 4) + expected_right_gain = 0.24154589371980678 + # (0.4 + 2.4 - 1) ** 2 / (0.24 + 0.4 + 0.14 + 4) + expected_bias_gain = 0.6778242677824265 + + split_info = split_info_pb2.SplitInfo() + split_info.ParseFromString(splits[0]) + left_child = split_info.left_child.vector + right_child = split_info.right_child.vector + split_node = split_info.split_node.sparse_float_binary_split_default_right + self.assertAllClose( + expected_left_gain + expected_right_gain - expected_bias_gain, gains[0]) + + self.assertAllClose([expected_left_weight], left_child.value) + + self.assertAllClose([expected_right_weight], right_child.value) + + self.assertEqual(0, split_node.split.feature_column) + + self.assertAllClose(0.52, split_node.split.threshold) + + # Check the split on partition 1. + expected_left_weight = -1.8779342723004695 + expected_right_weight = 0 + + # Verify candidate for partition 1, there's only one active bucket here + # so zero gain is expected. + split_info.ParseFromString(splits[1]) + left_child = split_info.left_child.vector + right_child = split_info.right_child.vector + split_node = split_info.split_node.sparse_float_binary_split_default_left + + self.assertAllClose(0.0, gains[1]) + + self.assertAllClose([expected_left_weight], left_child.value) + + self.assertAllClose([expected_right_weight], right_child.value) + + self.assertEqual(0, split_node.split.feature_column) + + self.assertAllClose(0.52, split_node.split.threshold) + def testGenerateFeatureSplitCandidatesMulticlassFullHessian(self): with self.test_session() as sess: # Batch is 4, 2 classes diff --git a/tensorflow/contrib/boosted_trees/lib/quantiles/weighted_quantiles_summary.h b/tensorflow/contrib/boosted_trees/lib/quantiles/weighted_quantiles_summary.h index a7e7bfc13cadcea4d29d33e0dbd955bdad6ffcb9..69bb8fd4ada861a42a0ccc3f287a47d91be5c879 100644 --- a/tensorflow/contrib/boosted_trees/lib/quantiles/weighted_quantiles_summary.h +++ b/tensorflow/contrib/boosted_trees/lib/quantiles/weighted_quantiles_summary.h @@ -51,7 +51,7 @@ class WeightedQuantilesSummary { SummaryEntry() { memset(this, 0, sizeof(*this)); - value = 0; + value = ValueType(); weight = 0; min_rank = 0; max_rank = 0; diff --git a/tensorflow/contrib/boosted_trees/python/training/functions/gbdt_batch.py b/tensorflow/contrib/boosted_trees/python/training/functions/gbdt_batch.py index bc8651ba92a4a4ea22e6eeec1013367eabc03ac7..1ee7f2395ea2ad71a7d380a1cc8f9a77bd4782b3 100644 --- a/tensorflow/contrib/boosted_trees/python/training/functions/gbdt_batch.py +++ b/tensorflow/contrib/boosted_trees/python/training/functions/gbdt_batch.py @@ -46,6 +46,7 @@ from tensorflow.python.ops import gradients_impl from tensorflow.python.ops import math_ops from tensorflow.python.ops import variable_scope from tensorflow.python.ops import variables +from tensorflow.python.ops.losses import losses from tensorflow.python.platform import tf_logging as logging from tensorflow.python.summary import summary from tensorflow.python.training import device_setter @@ -62,15 +63,11 @@ LEAF_INDEX = "leaf_index" _FEATURE_NAME_TEMPLATE = "%s_%d" # Keys in Training state. -_NUM_LAYER_EXAMPLES = "num_layer_examples" -_NUM_LAYER_STEPS = "num_layer_steps" -_NUM_LAYERS = "num_layers" -_ACTIVE_TREE = "active_tree" -_ACTIVE_LAYER = "active_layer" -_CONTINUE_CENTERING = "continue_centering" -_BIAS_STATS_ACCUMULATOR = "bias_stats_accumulator" -_STEPS_ACCUMULATOR = "steps_accumulator" -_HANDLERS = "handlers" +GBDTTrainingState = collections.namedtuple("GBDTTrainingState", [ + "num_layer_examples", "num_layer_steps", "num_layers", "active_tree", + "active_layer", "continue_centering", "bias_stats_accumulator", + "steps_accumulator", "handlers" +]) def _get_column_by_index(tensor, indices): @@ -287,6 +284,7 @@ class GradientBoostedDecisionTreeModel(object): learner_config, features, logits_dimension, + loss_reduction=losses.Reduction.SUM_OVER_NONZERO_WEIGHTS, feature_columns=None, use_core_columns=False, output_leaf_index=False): @@ -303,7 +301,10 @@ class GradientBoostedDecisionTreeModel(object): learner_config: A learner config. features: `dict` of `Tensor` objects. logits_dimension: An int, the dimension of logits. + loss_reduction: Either `SUM_OVER_NONZERO_WEIGHTS` (mean) or `SUM`. feature_columns: A list of feature columns. + use_core_columns: A boolean specifying whether core feature columns are + used. output_leaf_index: A boolean variable indicating whether to output leaf index into predictions dictionary. @@ -326,6 +327,13 @@ class GradientBoostedDecisionTreeModel(object): self._center_bias = center_bias self._examples_per_layer = examples_per_layer + # Check loss reduction value. + if (loss_reduction != losses.Reduction.SUM and + loss_reduction != losses.Reduction.SUM_OVER_NONZERO_WEIGHTS): + raise ValueError( + "Invalid loss reduction is provided: %s." % loss_reduction) + self._loss_reduction = loss_reduction + # Fill in the defaults. if (learner_config.multi_class_strategy == learner_pb2.LearnerConfig.MULTI_CLASS_STRATEGY_UNSPECIFIED): @@ -383,6 +391,7 @@ class GradientBoostedDecisionTreeModel(object): sparse_int_values, sparse_int_shapes) = extract_features( features, self._feature_columns, use_core_columns) logging.info("Active Feature Columns: " + str(fc_names)) + logging.info("Learner config: " + str(learner_config)) self._fc_names = fc_names self._dense_floats = dense_floats self._sparse_float_indices = sparse_float_indices @@ -565,7 +574,11 @@ class GradientBoostedDecisionTreeModel(object): about predictions per example. Returns: - An op that adds a new tree to the ensemble. + Three values: + - An op that adds a new tree to the ensemble, and + - An op that increments the stamp but removes all the trees and resets + the handlers. This can be used to reset the state of the ensemble. + - A dict containing the training state. Raises: ValueError: if inputs are not valid. @@ -642,6 +655,8 @@ class GradientBoostedDecisionTreeModel(object): self._learner_config.regularization.tree_complexity, dtypes.float32) min_node_weight = constant_op.constant( self._learner_config.constraints.min_node_weight, dtypes.float32) + loss_uses_sum_reduction = self._loss_reduction == losses.Reduction.SUM + loss_uses_sum_reduction = constant_op.constant(loss_uses_sum_reduction) epsilon = 0.01 num_quantiles = 100 strategy_tensor = constant_op.constant(strategy) @@ -655,7 +670,8 @@ class GradientBoostedDecisionTreeModel(object): l2_regularization=l2_regularization, tree_complexity_regularization=tree_complexity_regularization, min_node_weight=min_node_weight, - feature_column_group_id=dense_float_column_idx, + feature_column_group_id=constant_op.constant( + dense_float_column_idx), epsilon=epsilon, num_quantiles=num_quantiles, dense_float_column=self._dense_floats[dense_float_column_idx], @@ -663,7 +679,9 @@ class GradientBoostedDecisionTreeModel(object): gradient_shape=self._gradient_shape, hessian_shape=self._hessian_shape, multiclass_strategy=strategy_tensor, - init_stamp_token=init_stamp_token)) + init_stamp_token=init_stamp_token, + loss_uses_sum_reduction=loss_uses_sum_reduction, + )) fc_name_idx += 1 # Create handlers for sparse float columns. @@ -675,7 +693,8 @@ class GradientBoostedDecisionTreeModel(object): l2_regularization=l2_regularization, tree_complexity_regularization=tree_complexity_regularization, min_node_weight=min_node_weight, - feature_column_group_id=sparse_float_column_idx, + feature_column_group_id=constant_op.constant( + sparse_float_column_idx), epsilon=epsilon, num_quantiles=num_quantiles, sparse_float_column=sparse_tensor.SparseTensor( @@ -686,7 +705,8 @@ class GradientBoostedDecisionTreeModel(object): gradient_shape=self._gradient_shape, hessian_shape=self._hessian_shape, multiclass_strategy=strategy_tensor, - init_stamp_token=init_stamp_token)) + init_stamp_token=init_stamp_token, + loss_uses_sum_reduction=loss_uses_sum_reduction)) fc_name_idx += 1 # Create handlers for sparse int columns. @@ -698,7 +718,8 @@ class GradientBoostedDecisionTreeModel(object): l2_regularization=l2_regularization, tree_complexity_regularization=tree_complexity_regularization, min_node_weight=min_node_weight, - feature_column_group_id=sparse_int_column_idx, + feature_column_group_id=constant_op.constant( + sparse_int_column_idx), sparse_int_column=sparse_tensor.SparseTensor( self._sparse_int_indices[sparse_int_column_idx], self._sparse_int_values[sparse_int_column_idx], @@ -707,7 +728,8 @@ class GradientBoostedDecisionTreeModel(object): gradient_shape=self._gradient_shape, hessian_shape=self._hessian_shape, multiclass_strategy=strategy_tensor, - init_stamp_token=init_stamp_token)) + init_stamp_token=init_stamp_token, + loss_uses_sum_reduction=loss_uses_sum_reduction)) fc_name_idx += 1 # Create ensemble stats variables. @@ -843,21 +865,45 @@ class GradientBoostedDecisionTreeModel(object): for update in update_results.values(): stats_update_ops += update - training_state = { - _NUM_LAYER_EXAMPLES: num_layer_examples, - _NUM_LAYER_STEPS: num_layer_steps, - _NUM_LAYERS: num_layers, - _ACTIVE_TREE: active_tree, - _ACTIVE_LAYER: active_layer, - _CONTINUE_CENTERING: continue_centering, - _BIAS_STATS_ACCUMULATOR: bias_stats_accumulator, - _STEPS_ACCUMULATOR: steps_accumulator, - _HANDLERS: handlers - } - return stats_update_ops, training_state - - def increment_step_counter_and_maybe_update_ensemble( - self, predictions_dict, batch_size, training_state): + training_state = GBDTTrainingState( + num_layer_examples=num_layer_examples, + num_layer_steps=num_layer_steps, + num_layers=num_layers, + active_tree=active_tree, + active_layer=active_layer, + continue_centering=continue_centering, + bias_stats_accumulator=bias_stats_accumulator, + steps_accumulator=steps_accumulator, + handlers=handlers) + + reset_op = control_flow_ops.no_op() + if self._is_chief: + # Advance the ensemble stamp to throw away staggered workers. + stamp_token, _ = model_ops.tree_ensemble_serialize(self._ensemble_handle) + next_stamp_token = stamp_token + 1 + + reset_ops = [] + for handler in handlers: + reset_ops.append(handler.make_splits(stamp_token, next_stamp_token, 0)) + if self._center_bias: + reset_ops.append( + bias_stats_accumulator.flush(stamp_token, next_stamp_token)) + reset_ops.append(steps_accumulator.flush(stamp_token, next_stamp_token)) + reset_ops.append(self._finalized_trees.assign(0).op) + reset_ops.append(self._attempted_trees.assign(0).op) + reset_ops.append( + model_ops.tree_ensemble_deserialize( + self._ensemble_handle, + stamp_token=next_stamp_token, + tree_ensemble_config="", + name="reset_gbdt")) + + reset_op = control_flow_ops.group([reset_ops]) + + return stats_update_ops, reset_op, training_state + + def increment_step_counter_and_maybe_update_ensemble(self, predictions_dict, + training_state): """Increments number of visited examples and grows the ensemble. If the number of visited examples reaches the target examples_per_layer, @@ -866,24 +912,20 @@ class GradientBoostedDecisionTreeModel(object): Args: predictions_dict: Dictionary of Rank 2 `Tensor` representing information about predictions per example. - batch_size: Number of examples in the batch. training_state: `dict` returned by update_stats. Returns: An op that updates the counters and potientially grows the ensemble. """ + batch_size = math_ops.cast( + array_ops.shape(predictions_dict[PREDICTIONS])[0], dtypes.float32) ensemble_stamp = predictions_dict[ENSEMBLE_STAMP] # Accumulate a step after updating stats. - num_layer_examples = training_state[_NUM_LAYER_EXAMPLES] - num_layer_steps = training_state[_NUM_LAYER_STEPS] - num_layers = training_state[_NUM_LAYERS] - active_tree = training_state[_ACTIVE_TREE] - active_layer = training_state[_ACTIVE_LAYER] - continue_centering = training_state[_CONTINUE_CENTERING] - bias_stats_accumulator = training_state[_BIAS_STATS_ACCUMULATOR] - steps_accumulator = training_state[_STEPS_ACCUMULATOR] - handlers = training_state[_HANDLERS] + steps_accumulator = training_state.steps_accumulator + num_layer_examples = training_state.num_layer_examples + num_layer_steps = training_state.num_layer_steps + active_layer = training_state.active_layer add_step_op = steps_accumulator.add( ensemble_stamp, [0], [[0, 0]], [batch_size], [1.0]) @@ -910,11 +952,8 @@ class GradientBoostedDecisionTreeModel(object): ensemble_update_ops.append( control_flow_ops.cond( acc_examples >= examples_per_layer, - self.make_update_ensemble_fn( - ensemble_stamp, steps_accumulator, - bias_stats_accumulator, continue_centering, - handlers, num_layers, active_tree, - active_layer, dropout_seed, class_id), + self.make_update_ensemble_fn(ensemble_stamp, training_state, + dropout_seed, class_id), control_flow_ops.no_op)) # Note, the loss is calculated from the prediction considering dropouts, so @@ -922,9 +961,7 @@ class GradientBoostedDecisionTreeModel(object): # high. eval_loss might be referred instead in the aspect of convergence. return control_flow_ops.group(*ensemble_update_ops) - def make_update_ensemble_fn(self, ensemble_stamp, steps_accumulator, - bias_stats_accumulator, continue_centering, - handlers, num_layers, active_tree, active_layer, + def make_update_ensemble_fn(self, ensemble_stamp, training_state, dropout_seed, class_id): """A method to create the function which updates the tree ensemble.""" # Determine learning rate. @@ -943,8 +980,9 @@ class GradientBoostedDecisionTreeModel(object): # Get next stamp token. next_ensemble_stamp = ensemble_stamp + 1 # Finalize bias stats. - _, _, _, bias_grads, bias_hess = bias_stats_accumulator.flush( - ensemble_stamp, next_ensemble_stamp) + _, _, _, bias_grads, bias_hess = ( + training_state.bias_stats_accumulator.flush(ensemble_stamp, + next_ensemble_stamp)) # Finalize handler splits. are_splits_ready_list = [] @@ -952,7 +990,7 @@ class GradientBoostedDecisionTreeModel(object): gains_list = [] split_info_list = [] - for handler in handlers: + for handler in training_state.handlers: (are_splits_ready, partition_ids, gains, split_info) = handler.make_splits( ensemble_stamp, next_ensemble_stamp, class_id) @@ -985,7 +1023,7 @@ class GradientBoostedDecisionTreeModel(object): next_stamp_token=next_ensemble_stamp, delta_updates=delta_updates, learner_config=self._learner_config_serialized) - return continue_centering.assign(center_bias) + return training_state.continue_centering.assign(center_bias) # Define ensemble growing operations. def _grow_ensemble_ready_fn(): @@ -1030,7 +1068,7 @@ class GradientBoostedDecisionTreeModel(object): # Update ensemble. update_ops = [are_all_splits_ready] if self._center_bias: - update_model = control_flow_ops.cond(continue_centering, + update_model = control_flow_ops.cond(training_state.continue_centering, _center_bias_fn, _grow_ensemble_fn) else: update_model = _grow_ensemble_fn() @@ -1042,13 +1080,15 @@ class GradientBoostedDecisionTreeModel(object): self._ensemble_handle, stamp_token=next_ensemble_stamp) update_ops.append(self._finalized_trees.assign(stats.num_trees)) update_ops.append(self._attempted_trees.assign(stats.attempted_trees)) - update_ops.append(num_layers.assign(stats.num_layers)) - update_ops.append(active_tree.assign(stats.active_tree)) - update_ops.append(active_layer.assign(stats.active_layer)) + update_ops.append(training_state.num_layers.assign(stats.num_layers)) + update_ops.append(training_state.active_tree.assign(stats.active_tree)) + update_ops.append( + training_state.active_layer.assign(stats.active_layer)) # Flush step stats. update_ops.extend( - steps_accumulator.flush(ensemble_stamp, next_ensemble_stamp)) + training_state.steps_accumulator.flush(ensemble_stamp, + next_ensemble_stamp)) return control_flow_ops.group(*update_ops, name="update_ensemble") return _update_ensemble @@ -1063,7 +1103,8 @@ class GradientBoostedDecisionTreeModel(object): loss: A scalar tensor representing average loss of examples. predictions_dict: Dictionary of Rank 2 `Tensor` representing information about predictions per example. - labels: Rank 2 `Tensor` representing labels per example. + labels: Rank 2 `Tensor` representing labels per example. Has no effect + on the training and is only kept for backward compatibility. Returns: An op that adds a new tree to the ensemble. @@ -1071,11 +1112,11 @@ class GradientBoostedDecisionTreeModel(object): Raises: ValueError: if inputs are not valid. """ - batch_size = math_ops.cast(array_ops.shape(labels)[0], dtypes.float32) - update_op, handlers = self.update_stats(loss, predictions_dict) + del labels # unused; kept for backward compatibility. + update_op, _, training_state = self.update_stats(loss, predictions_dict) with ops.control_dependencies(update_op): return self.increment_step_counter_and_maybe_update_ensemble( - predictions_dict, batch_size, handlers) + predictions_dict, training_state) def _get_weights(self, hessian_shape, hessians): """Derives weights to be used based on hessians and multiclass strategy.""" diff --git a/tensorflow/contrib/boosted_trees/python/training/functions/gbdt_batch_test.py b/tensorflow/contrib/boosted_trees/python/training/functions/gbdt_batch_test.py index e3d4397fadcbaf148f7f6cfaca13e850639786cf..f7867d882d6813a8701065ad0ce8d27f8bb9c301 100644 --- a/tensorflow/contrib/boosted_trees/python/training/functions/gbdt_batch_test.py +++ b/tensorflow/contrib/boosted_trees/python/training/functions/gbdt_batch_test.py @@ -29,6 +29,7 @@ from tensorflow.contrib.layers.python.layers import feature_column as feature_co from tensorflow.contrib.learn.python.learn.estimators import model_fn from tensorflow.python.feature_column import feature_column_lib as core_feature_column from tensorflow.python.framework import dtypes +from tensorflow.python.framework import ops from tensorflow.python.framework import sparse_tensor from tensorflow.python.framework import test_util from tensorflow.python.ops import array_ops @@ -1560,6 +1561,301 @@ class GbdtTest(test_util.TensorFlowTestCase): self.assertEquals(output.growing_metadata.num_layers_attempted, 2) + def testResetModelBeforeAndAfterSplit(self): + """Tests whether resetting works.""" + with self.test_session(): + # First build a small tree and train it to verify training works. + ensemble_handle = model_ops.tree_ensemble_variable( + stamp_token=0, tree_ensemble_config="", name="tree_ensemble") + learner_config = learner_pb2.LearnerConfig() + learner_config.learning_rate_tuner.fixed.learning_rate = 0.1 + learner_config.num_classes = 2 + learner_config.constraints.max_tree_depth = 1 + features = {} + features["dense_float"] = array_ops.ones([4, 1], dtypes.float32) + + gbdt_model = gbdt_batch.GradientBoostedDecisionTreeModel( + is_chief=True, + num_ps_replicas=0, + center_bias=False, + ensemble_handle=ensemble_handle, + examples_per_layer=1, + learner_config=learner_config, + logits_dimension=1, + features=features) + + predictions = array_ops.constant( + [[0.0], [1.0], [0.0], [2.0]], dtype=dtypes.float32) + partition_ids = array_ops.zeros([4], dtypes.int32) + ensemble_stamp = model_ops.tree_ensemble_stamp_token(ensemble_handle) + + predictions_dict = { + "predictions": predictions, + "predictions_no_dropout": predictions, + "partition_ids": partition_ids, + "ensemble_stamp": ensemble_stamp, + "num_trees": 12, + "max_tree_depth": 4, + } + + labels = array_ops.ones([4, 1], dtypes.float32) + weights = array_ops.ones([4, 1], dtypes.float32) + loss = math_ops.reduce_mean(_squared_loss(labels, weights, predictions)) + + # Create train op. + update_op, reset_op, training_state = gbdt_model.update_stats( + loss, predictions_dict) + with ops.control_dependencies(update_op): + train_op = gbdt_model.increment_step_counter_and_maybe_update_ensemble( + predictions_dict, training_state) + + variables.global_variables_initializer().run() + resources.initialize_resources(resources.shared_resources()).run() + + original_stamp = ensemble_stamp.eval() + expected_tree = """ + nodes { + dense_float_binary_split { + threshold: 1.0 + left_id: 1 + right_id: 2 + } + node_metadata { + gain: 0 + } + } + nodes { + leaf { + vector { + value: 0.25 + } + } + } + nodes { + leaf { + vector { + value: 0.0 + } + } + }""" + + def _train_once_and_check(expect_split): + stamp = ensemble_stamp.eval() + train_op.run() + stamp_token, serialized = model_ops.tree_ensemble_serialize( + ensemble_handle) + output = tree_config_pb2.DecisionTreeEnsembleConfig() + output.ParseFromString(serialized.eval()) + self.assertEquals(stamp_token.eval(), stamp + 1) + if expect_split: + # State of the ensemble after a split occurs. + self.assertEquals(len(output.trees), 1) + self.assertProtoEquals(expected_tree, output.trees[0]) + else: + # State of the ensemble after a single accumulation but before any + # splitting occurs + self.assertEquals(len(output.trees), 0) + self.assertProtoEquals(""" + growing_metadata { + num_trees_attempted: 1 + num_layers_attempted: 1 + }""", output) + + def _run_reset(): + stamp_before_reset = ensemble_stamp.eval() + reset_op.run() + stamp_after_reset = ensemble_stamp.eval() + self.assertNotEquals(stamp_after_reset, stamp_before_reset) + + _, serialized = model_ops.tree_ensemble_serialize( + ensemble_handle) + output = tree_config_pb2.DecisionTreeEnsembleConfig() + output.ParseFromString(serialized.eval()) + self.assertProtoEquals("", output) + + return stamp_after_reset + + # Exit after one train_op, so no new layer are created but the handlers + # contain enough information to split on the next call to train. + _train_once_and_check(expect_split=False) + self.assertEquals(ensemble_stamp.eval(), original_stamp + 1) + + # Reset the handlers so it still requires two training calls to split. + stamp_after_reset = _run_reset() + + _train_once_and_check(expect_split=False) + _train_once_and_check(expect_split=True) + self.assertEquals(ensemble_stamp.eval(), stamp_after_reset + 2) + + # This time, test that the reset_op works right after splitting. + stamp_after_reset = _run_reset() + + # Test that after resetting, the tree can be trained as normal. + _train_once_and_check(expect_split=False) + _train_once_and_check(expect_split=True) + self.assertEquals(ensemble_stamp.eval(), stamp_after_reset + 2) + + def testResetModelNonChief(self): + """Tests the reset function on a non-chief worker.""" + with self.test_session(): + # Create ensemble with one bias node. + ensemble_config = tree_config_pb2.DecisionTreeEnsembleConfig() + text_format.Merge( + """ + trees { + nodes { + leaf { + vector { + value: 0.25 + } + } + } + } + tree_weights: 1.0 + tree_metadata { + num_tree_weight_updates: 1 + num_layers_grown: 1 + is_finalized: false + }""", ensemble_config) + ensemble_handle = model_ops.tree_ensemble_variable( + stamp_token=0, + tree_ensemble_config=ensemble_config.SerializeToString(), + name="tree_ensemble") + learner_config = learner_pb2.LearnerConfig() + learner_config.learning_rate_tuner.fixed.learning_rate = 0.1 + learner_config.num_classes = 2 + learner_config.constraints.max_tree_depth = 1 + features = {} + features["dense_float"] = array_ops.ones([4, 1], dtypes.float32) + + gbdt_model = gbdt_batch.GradientBoostedDecisionTreeModel( + is_chief=False, + num_ps_replicas=0, + center_bias=False, + ensemble_handle=ensemble_handle, + examples_per_layer=1, + learner_config=learner_config, + logits_dimension=1, + features=features) + + predictions = array_ops.constant( + [[0.0], [1.0], [0.0], [2.0]], dtype=dtypes.float32) + partition_ids = array_ops.zeros([4], dtypes.int32) + ensemble_stamp = model_ops.tree_ensemble_stamp_token(ensemble_handle) + + predictions_dict = { + "predictions": predictions, + "predictions_no_dropout": predictions, + "partition_ids": partition_ids, + "ensemble_stamp": ensemble_stamp + } + + labels = array_ops.ones([4, 1], dtypes.float32) + weights = array_ops.ones([4, 1], dtypes.float32) + loss = math_ops.reduce_mean(_squared_loss(labels, weights, predictions)) + + # Create reset op. + _, reset_op, _ = gbdt_model.update_stats( + loss, predictions_dict) + + variables.global_variables_initializer().run() + resources.initialize_resources(resources.shared_resources()).run() + + # Reset op doesn't do anything because this is a non-chief worker. + reset_op.run() + stamp_token, serialized = model_ops.tree_ensemble_serialize( + ensemble_handle) + output = tree_config_pb2.DecisionTreeEnsembleConfig() + output.ParseFromString(serialized.eval()) + self.assertEquals(len(output.trees), 1) + self.assertEquals(len(output.tree_weights), 1) + self.assertEquals(stamp_token.eval(), 0) + + def testResetModelWithCenterBias(self): + """Tests the reset function running on chief with bias centering.""" + with self.test_session(): + ensemble_handle = model_ops.tree_ensemble_variable( + stamp_token=0, tree_ensemble_config="", name="tree_ensemble") + learner_config = learner_pb2.LearnerConfig() + learner_config.learning_rate_tuner.fixed.learning_rate = 0.1 + learner_config.num_classes = 2 + learner_config.regularization.l1 = 0 + learner_config.regularization.l2 = 0 + learner_config.constraints.max_tree_depth = 1 + learner_config.constraints.min_node_weight = 0 + features = {} + features["dense_float"] = array_ops.ones([4, 1], dtypes.float32) + + gbdt_model = gbdt_batch.GradientBoostedDecisionTreeModel( + is_chief=True, + num_ps_replicas=0, + center_bias=True, + ensemble_handle=ensemble_handle, + examples_per_layer=1, + learner_config=learner_config, + logits_dimension=1, + features=features) + + predictions = array_ops.constant( + [[0.0], [1.0], [0.0], [2.0]], dtype=dtypes.float32) + partition_ids = array_ops.zeros([4], dtypes.int32) + ensemble_stamp = model_ops.tree_ensemble_stamp_token(ensemble_handle) + + predictions_dict = { + "predictions": predictions, + "predictions_no_dropout": predictions, + "partition_ids": partition_ids, + "ensemble_stamp": ensemble_stamp, + "num_trees": 12, + } + + labels = array_ops.ones([4, 1], dtypes.float32) + weights = array_ops.ones([4, 1], dtypes.float32) + loss = math_ops.reduce_mean(_squared_loss(labels, weights, predictions)) + + # Create train op. + update_op, reset_op, training_state = gbdt_model.update_stats( + loss, predictions_dict) + with ops.control_dependencies(update_op): + train_op = gbdt_model.increment_step_counter_and_maybe_update_ensemble( + predictions_dict, training_state) + + variables.global_variables_initializer().run() + resources.initialize_resources(resources.shared_resources()).run() + + # On first run, expect bias to be centered. + def train_and_check(): + train_op.run() + _, serialized = model_ops.tree_ensemble_serialize(ensemble_handle) + output = tree_config_pb2.DecisionTreeEnsembleConfig() + output.ParseFromString(serialized.eval()) + expected_tree = """ + nodes { + leaf { + vector { + value: 0.25 + } + } + }""" + self.assertEquals(len(output.trees), 1) + self.assertAllEqual(output.tree_weights, [1.0]) + self.assertProtoEquals(expected_tree, output.trees[0]) + + train_and_check() + self.assertEquals(ensemble_stamp.eval(), 1) + + reset_op.run() + stamp_token, serialized = model_ops.tree_ensemble_serialize( + ensemble_handle) + output = tree_config_pb2.DecisionTreeEnsembleConfig() + output.ParseFromString(serialized.eval()) + self.assertEquals(len(output.trees), 0) + self.assertEquals(len(output.tree_weights), 0) + self.assertEquals(stamp_token.eval(), 2) + + train_and_check() + self.assertEquals(ensemble_stamp.eval(), 3) + if __name__ == "__main__": googletest.main() diff --git a/tensorflow/contrib/boosted_trees/python/utils/losses.py b/tensorflow/contrib/boosted_trees/python/utils/losses.py index ab7ac2aba605db22a8ed370049b27d55cf1d413a..b5ebaf1999519f65110e8164fa20bace5ecc3ef6 100644 --- a/tensorflow/contrib/boosted_trees/python/utils/losses.py +++ b/tensorflow/contrib/boosted_trees/python/utils/losses.py @@ -23,6 +23,12 @@ from tensorflow.python.ops import array_ops from tensorflow.python.ops import control_flow_ops from tensorflow.python.ops import math_ops from tensorflow.python.ops import nn +from tensorflow.python.ops.losses import losses + + +def per_example_squared_hinge_loss(labels, weights, predictions): + loss = losses.hinge_loss(labels=labels, logits=predictions, weights=weights) + return math_ops.square(loss), control_flow_ops.no_op() def per_example_logistic_loss(labels, weights, predictions): @@ -126,7 +132,7 @@ def per_example_squared_loss(labels, weights, predictions): def per_example_exp_loss(labels, weights, predictions, name=None, eps=0.1): - """Exponential loss given labels, example weights and predictions. + """Trimmed exponential loss given labels, example weights and predictions. Note that this is only for binary classification. If logistic loss tries to make sure that the classifier is certain of its @@ -211,3 +217,62 @@ def per_example_exp_loss(labels, weights, predictions, name=None, eps=0.1): unweighted_loss = exp_with_logits( name=name, eps=eps, labels=labels, logits=predictions) return unweighted_loss * weights, control_flow_ops.no_op() + + +def per_example_full_exp_loss(labels, weights, predictions, name=None): + """Full exponential loss given labels, example weights and predictions. + + Note that this is only for binary classification. + The loss returns is exp(-targets*logits), where targets are converted to -1 + and 1. + + Args: + labels: Rank 2 (N, D) tensor of per-example labels. + weights: Rank 2 (N, 1) tensor of per-example weights. + predictions: Rank 2 (N, D) tensor of per-example predictions. + name: A name for the operation (optional). + + Returns: + loss: A Rank 2 (N, 1) tensor of per-example exp loss + update_op: An update operation to update the loss's internal state. + """ + + def full_exp_with_logits(name, labels=None, logits=None): + """Computes exponential loss given `logits`. + + Args: + name: A name for the operation (optional). + labels: A `Tensor` of the same type and shape as `logits`. + logits: A `Tensor` of type `float32` or `float64`. + + Returns: + A `Tensor` of the same shape as `logits` with the componentwise + exponential losses. + + Raises: + ValueError: If `logits` and `labels` do not have the same shape. + """ + with ops.name_scope(name, "exp_loss", [logits, labels]) as name: + logits = ops.convert_to_tensor(logits, name="logits") + labels = ops.convert_to_tensor(labels, name="labels") + try: + labels.get_shape().merge_with(logits.get_shape()) + except ValueError: + raise ValueError("logits and labels must have the same shape (%s vs %s)" + % (logits.get_shape(), labels.get_shape())) + + # Default threshold of 0 to switch between classes + zeros = array_ops.zeros_like(logits, dtype=logits.dtype) + ones = array_ops.ones_like(logits, dtype=logits.dtype) + neg_ones = -array_ops.ones_like(logits, dtype=logits.dtype) + + # Convert labels to 1 and -1 + cond_labels = (labels > zeros) + labels_converted = array_ops.where(cond_labels, ones, neg_ones) + + return math_ops.exp(-1.0 * logits * labels_converted) + + labels = math_ops.to_float(labels) + unweighted_loss = full_exp_with_logits( + name=name, labels=labels, logits=predictions) + return unweighted_loss * weights, control_flow_ops.no_op() diff --git a/tensorflow/contrib/checkpoint/__init__.py b/tensorflow/contrib/checkpoint/__init__.py index 8c1ce5c2a2d552e30d3b676e3ac8b5fc7c74a917..2fbaa31d5e19b58c335cd0a894e1db9af2c34d08 100644 --- a/tensorflow/contrib/checkpoint/__init__.py +++ b/tensorflow/contrib/checkpoint/__init__.py @@ -44,8 +44,8 @@ from tensorflow.core.protobuf.checkpointable_object_graph_pb2 import Checkpointa from tensorflow.python.training.checkpointable.base import CheckpointableBase from tensorflow.python.training.checkpointable.data_structures import List from tensorflow.python.training.checkpointable.data_structures import Mapping +from tensorflow.python.training.checkpointable.data_structures import NoDependency from tensorflow.python.training.checkpointable.tracking import Checkpointable -from tensorflow.python.training.checkpointable.tracking import NoDependency from tensorflow.python.training.checkpointable.util import capture_dependencies from tensorflow.python.training.checkpointable.util import list_objects from tensorflow.python.training.checkpointable.util import object_metadata diff --git a/tensorflow/contrib/checkpoint/python/containers_test.py b/tensorflow/contrib/checkpoint/python/containers_test.py index 64d056bd689a14c0c58d7a0f75c833c71b00a5c3..ac85c7be803cd4c2f8ba19d3ef887a3c65a15933 100644 --- a/tensorflow/contrib/checkpoint/python/containers_test.py +++ b/tensorflow/contrib/checkpoint/python/containers_test.py @@ -26,6 +26,7 @@ from tensorflow.python.keras import layers from tensorflow.python.ops import array_ops from tensorflow.python.ops import resource_variable_ops from tensorflow.python.platform import test +from tensorflow.python.training.checkpointable import data_structures from tensorflow.python.training.checkpointable import tracking from tensorflow.python.training.checkpointable import util @@ -79,7 +80,7 @@ class UniqueNameTrackerTests(test.TestCase): resource_variable_ops.ResourceVariable(4.), "y")) slots.append(slotdeps.track( resource_variable_ops.ResourceVariable(5.), "x")) - self.slots = slots + self.slots = data_structures.NoDependency(slots) manager = SlotManager() self.evaluate([v.initializer for v in manager.slots]) diff --git a/tensorflow/contrib/cloud/BUILD b/tensorflow/contrib/cloud/BUILD index 1a7a3759baa4a5559b4b70ff4f7467c41da9111f..523a9efcf05f5d32589f6e1734f866bf8b4b9cdc 100644 --- a/tensorflow/contrib/cloud/BUILD +++ b/tensorflow/contrib/cloud/BUILD @@ -50,6 +50,7 @@ py_library( deps = [ ":gen_bigquery_reader_ops", ":gen_gcs_config_ops", + "//tensorflow/contrib/bigtable", "//tensorflow/python:framework_for_generated_wrappers", "//tensorflow/python:io_ops", "//tensorflow/python:util", diff --git a/tensorflow/contrib/cloud/README.md b/tensorflow/contrib/cloud/README.md new file mode 100644 index 0000000000000000000000000000000000000000..134ce057f4334096b4fbbec29cc85f0ea42c9f86 --- /dev/null +++ b/tensorflow/contrib/cloud/README.md @@ -0,0 +1,18 @@ +# Cloud # + +## BigTable ## + +[Google Cloud BigTable](https://cloud.google.com/bigtable/) is a high +performance storage system that can store and serve training data. This contrib +package contains an experimental integration with TensorFlow. + +> **Status: Highly experimental.** The current implementation is very much in +> flux. Please use at your own risk! :-) + + + +## Cloud Storage (GCS) ## + +The Google Cloud Storage ops allow the user to configure the GCS File System. + + diff --git a/tensorflow/contrib/cloud/__init__.py b/tensorflow/contrib/cloud/__init__.py index ef7aa7624ce7b9b6480c4d088a2fb7678a7acc76..af81106a6848bfd8c91108b56c8150d47c3eb501 100644 --- a/tensorflow/contrib/cloud/__init__.py +++ b/tensorflow/contrib/cloud/__init__.py @@ -18,15 +18,24 @@ from __future__ import absolute_import from __future__ import division from __future__ import print_function -# pylint: disable=line-too-long,wildcard-import +import os + +# pylint: disable=line-too-long,wildcard-import,g-import-not-at-top from tensorflow.contrib.cloud.python.ops.bigquery_reader_ops import * from tensorflow.contrib.cloud.python.ops.gcs_config_ops import * -# pylint: enable=line-too-long,wildcard-import + +if os.name != 'nt': + from tensorflow.contrib.bigtable.python.ops.bigtable_api import BigTable + from tensorflow.contrib.bigtable.python.ops.bigtable_api import BigtableClient + +del os from tensorflow.python.util.all_util import remove_undocumented _allowed_symbols = [ 'BigQueryReader', + 'BigTable', + 'BigtableClient', 'BlockCacheParams', 'configure_colab_session', 'configure_gcs', diff --git a/tensorflow/contrib/cluster_resolver/BUILD b/tensorflow/contrib/cluster_resolver/BUILD index c239e6f8f960910cee14e1df7c4678c643496f54..707f6211846ca0310bde297603928e9ec5bb471c 100644 --- a/tensorflow/contrib/cluster_resolver/BUILD +++ b/tensorflow/contrib/cluster_resolver/BUILD @@ -12,6 +12,15 @@ licenses(["notice"]) # Apache 2.0 py_library( name = "cluster_resolver_pip", + srcs_version = "PY2AND3", + visibility = ["//visibility:public"], + deps = [ + ":cluster_resolver_py", + ], +) + +py_library( + name = "cluster_resolver_py", srcs = [ "__init__.py", "python/training/__init__.py", @@ -19,7 +28,7 @@ py_library( srcs_version = "PY2AND3", visibility = ["//visibility:public"], deps = [ - ":cluster_resolver_py", + ":base_cluster_resolver_py", ":gce_cluster_resolver_py", ":tpu_cluster_resolver_py", "//tensorflow/python:util", @@ -27,7 +36,7 @@ py_library( ) py_library( - name = "cluster_resolver_py", + name = "base_cluster_resolver_py", srcs = ["python/training/cluster_resolver.py"], srcs_version = "PY2AND3", deps = [ @@ -40,7 +49,7 @@ py_library( srcs = ["python/training/gce_cluster_resolver.py"], srcs_version = "PY2AND3", deps = [ - ":cluster_resolver_py", + ":base_cluster_resolver_py", "//tensorflow/python:training", ], ) @@ -50,13 +59,13 @@ py_library( srcs = ["python/training/tpu_cluster_resolver.py"], srcs_version = "PY2AND3", deps = [ - ":cluster_resolver_py", + ":base_cluster_resolver_py", "//tensorflow/python:training", ], ) tf_py_test( - name = "cluster_resolver_py_test", + name = "base_cluster_resolver_py_test", srcs = ["python/training/cluster_resolver_test.py"], additional_deps = [ ":cluster_resolver_py", diff --git a/tensorflow/contrib/cmake/external/nsync.cmake b/tensorflow/contrib/cmake/external/nsync.cmake index 6d50a4956b8b525b231d4344b83481f3ab2699e9..eba3bcfc79efe87d0a45c979c5accfa1b6511ed0 100644 --- a/tensorflow/contrib/cmake/external/nsync.cmake +++ b/tensorflow/contrib/cmake/external/nsync.cmake @@ -16,7 +16,7 @@ include (ExternalProject) set(nsync_INCLUDE_DIR ${CMAKE_CURRENT_BINARY_DIR}/external/nsync/public) set(nsync_URL https://github.com/google/nsync) -set(nsync_TAG 5e8b19a81e5729922629dd505daa651f6ffdf107) +set(nsync_TAG 1.20.0) set(nsync_BUILD ${CMAKE_CURRENT_BINARY_DIR}/nsync/src/nsync) set(nsync_INSTALL ${CMAKE_CURRENT_BINARY_DIR}/nsync/install) diff --git a/tensorflow/contrib/cmake/python_modules.txt b/tensorflow/contrib/cmake/python_modules.txt index d530572e91825ed88d09c26a10693288878d09ed..75e00f32675df1b7e523bc7e8bb44fa584b79347 100644 --- a/tensorflow/contrib/cmake/python_modules.txt +++ b/tensorflow/contrib/cmake/python_modules.txt @@ -14,6 +14,7 @@ tensorflow/examples/tutorials tensorflow/examples/tutorials/mnist tensorflow/python tensorflow/python/client +tensorflow/python/compat tensorflow/python/data tensorflow/python/data/ops tensorflow/python/data/util @@ -61,6 +62,8 @@ tensorflow/python/saved_model tensorflow/python/summary tensorflow/python/summary/writer tensorflow/python/tools +tensorflow/python/tools/api +tensorflow/python/tools/api/generator tensorflow/python/training tensorflow/python/training/checkpointable tensorflow/python/user_ops @@ -68,7 +71,6 @@ tensorflow/python/util tensorflow/python/util/protobuf tensorflow/tools tensorflow/tools/api -tensorflow/tools/api/generator tensorflow/tools/graph_transforms tensorflow/contrib tensorflow/contrib/all_reduce @@ -86,6 +88,8 @@ tensorflow/contrib/batching/python/ops tensorflow/contrib/bayesflow tensorflow/contrib/bayesflow/python tensorflow/contrib/bayesflow/python/ops +# tensorflow/contrib/bigtable/python +# tensorflow/contrib/bigtable/python/ops tensorflow/contrib/boosted_trees tensorflow/contrib/boosted_trees/estimator_batch tensorflow/contrib/boosted_trees/kernels @@ -238,6 +242,8 @@ tensorflow/contrib/keras/api/keras/wrappers/scikit_learn tensorflow/contrib/kernel_methods tensorflow/contrib/kernel_methods/python tensorflow/contrib/kernel_methods/python/mappers +tensorflow/contrib/kinesis/python +tensorflow/contrib/kinesis/python/ops tensorflow/contrib/kfac tensorflow/contrib/kfac/examples tensorflow/contrib/kfac/python diff --git a/tensorflow/contrib/cmake/tf_c.cmake b/tensorflow/contrib/cmake/tf_c.cmake index 2e0a2fcef4cbdc50f0521296c4a25a864dbd8b77..7a30eb94f54b18a2a517615a315e23e09e1170d0 100644 --- a/tensorflow/contrib/cmake/tf_c.cmake +++ b/tensorflow/contrib/cmake/tf_c.cmake @@ -36,16 +36,3 @@ add_dependencies( tf_cc_while_loop tf_core_lib tf_protos_cc) - -if(tensorflow_BUILD_PYTHON_BINDINGS) - add_library(tf_c_python_api OBJECT - "${tensorflow_source_dir}/tensorflow/c/python_api.cc" - "${tensorflow_source_dir}/tensorflow/c/python_api.h" - ) - add_dependencies( - tf_c_python_api - tf_c - tf_core_lib - tf_core_framework - tf_protos_cc) -endif() diff --git a/tensorflow/contrib/cmake/tf_core_framework.cmake b/tensorflow/contrib/cmake/tf_core_framework.cmake index d044ac75ae55dcfd6aa1ff3890a71d35e775cb9b..067c299a71cd4ac96878bcf27b4453466785e4ba 100644 --- a/tensorflow/contrib/cmake/tf_core_framework.cmake +++ b/tensorflow/contrib/cmake/tf_core_framework.cmake @@ -125,6 +125,7 @@ endfunction() file(GLOB_RECURSE tf_protos_cc_srcs RELATIVE ${tensorflow_source_dir} "${tensorflow_source_dir}/tensorflow/core/*.proto" + "${tensorflow_source_dir}/tensorflow/compiler/xla/*.proto" "${tensorflow_source_dir}/tensorflow/contrib/boosted_trees/proto/*.proto" "${tensorflow_source_dir}/tensorflow/contrib/tpu/proto/*.proto" ) diff --git a/tensorflow/contrib/cmake/tf_python.cmake b/tensorflow/contrib/cmake/tf_python.cmake index 786ea05c744167ad52d52cc73328bd8c25d78c3e..32b185f07b6ba836ffb47e85beff6fb2481fdc3e 100755 --- a/tensorflow/contrib/cmake/tf_python.cmake +++ b/tensorflow/contrib/cmake/tf_python.cmake @@ -456,6 +456,18 @@ add_custom_command( COMMENT "Running SWIG to generate Python wrappers" VERBATIM ) +add_library(tf_c_python_api OBJECT + "${tensorflow_source_dir}/tensorflow/c/python_api.cc" + "${tensorflow_source_dir}/tensorflow/c/python_api.h" +) +add_dependencies( + tf_c_python_api + tf_c + tf_core_lib + tf_core_framework + tf_protos_cc + tf_python_protos_cc) + set (pywrap_tensorflow_internal_src "${tensorflow_source_dir}/tensorflow/core/profiler/internal/print_model_analysis.h" "${tensorflow_source_dir}/tensorflow/core/profiler/internal/print_model_analysis.cc" @@ -724,8 +736,8 @@ endif() # Generate API __init__.py files. ######################################################## -# Parse tensorflow/tools/api/generator/BUILD to get list of generated files. -FILE(READ ${tensorflow_source_dir}/tensorflow/tools/api/generator/api_gen.bzl api_generator_BUILD_text) +# Parse tensorflow/python/tools/api/generator/BUILD to get list of generated files. +FILE(READ ${tensorflow_source_dir}/tensorflow/python/tools/api/generator/api_gen.bzl api_generator_BUILD_text) STRING(REGEX MATCH "# BEGIN GENERATED FILES.*# END GENERATED FILES" api_init_files_text ${api_generator_BUILD_text}) string(REPLACE "# BEGIN GENERATED FILES" "" api_init_files_text ${api_init_files_text}) string(REPLACE "# END GENERATED FILES" "" api_init_files_text ${api_init_files_text}) @@ -769,7 +781,7 @@ if (tensorflow_ENABLE_MKL_SUPPORT) # Run create_python_api.py to generate API init files. COMMAND ${CMAKE_COMMAND} -E env PYTHONPATH=${CMAKE_CURRENT_BINARY_DIR}/tf_python PATH=${PY_RUNTIME_ENV} ${PYTHON_EXECUTABLE} - "${CMAKE_CURRENT_BINARY_DIR}/tf_python/tensorflow/tools/api/generator/create_python_api.py" + "${CMAKE_CURRENT_BINARY_DIR}/tf_python/tensorflow/python/tools/api/generator/create_python_api.py" "--root_init_template=${CMAKE_CURRENT_BINARY_DIR}/tf_python/tensorflow/api_template.__init__.py" "--apidir=${CMAKE_CURRENT_BINARY_DIR}/tf_python/tensorflow" "--package=tensorflow.python" @@ -791,7 +803,7 @@ else (tensorflow_ENABLE_MKL_SUPPORT) # Run create_python_api.py to generate API init files. COMMAND ${CMAKE_COMMAND} -E env PYTHONPATH=${CMAKE_CURRENT_BINARY_DIR}/tf_python ${PYTHON_EXECUTABLE} - "${CMAKE_CURRENT_BINARY_DIR}/tf_python/tensorflow/tools/api/generator/create_python_api.py" + "${CMAKE_CURRENT_BINARY_DIR}/tf_python/tensorflow/python/tools/api/generator/create_python_api.py" "--root_init_template=${CMAKE_CURRENT_BINARY_DIR}/tf_python/tensorflow/api_template.__init__.py" "--apidir=${CMAKE_CURRENT_BINARY_DIR}/tf_python/tensorflow" "--package=tensorflow.python" @@ -812,8 +824,8 @@ add_dependencies(tf_python_api tf_python_ops) # Generate API __init__.py files for tf.estimator. ######################################################## -# Parse tensorflow/tools/api/generator/BUILD to get list of generated files. -FILE(READ ${tensorflow_source_dir}/tensorflow/tools/api/generator/api_gen.bzl api_generator_BUILD_text) +# Parse tensorflow/python/tools/api/generator/BUILD to get list of generated files. +FILE(READ ${tensorflow_source_dir}/tensorflow/python/tools/api/generator/api_gen.bzl api_generator_BUILD_text) STRING(REGEX MATCH "# BEGIN GENERATED ESTIMATOR FILES.*# END GENERATED ESTIMATOR FILES" api_init_files_text ${api_generator_BUILD_text}) string(REPLACE "# BEGIN GENERATED ESTIMATOR FILES" "" api_init_files_text ${api_init_files_text}) string(REPLACE "# END GENERATED ESTIMATOR FILES" "" api_init_files_text ${api_init_files_text}) @@ -837,10 +849,11 @@ add_custom_command( # Run create_python_api.py to generate API init files. COMMAND ${CMAKE_COMMAND} -E env PYTHONPATH=${CMAKE_CURRENT_BINARY_DIR}/tf_python ${PYTHON_EXECUTABLE} - "${CMAKE_CURRENT_BINARY_DIR}/tf_python/tensorflow/tools/api/generator/create_python_api.py" + "${CMAKE_CURRENT_BINARY_DIR}/tf_python/tensorflow/python/tools/api/generator/create_python_api.py" "--apidir=${CMAKE_CURRENT_BINARY_DIR}/tf_python/tensorflow/python/estimator/api" "--package=tensorflow.python.estimator" "--apiname=estimator" + "--output_package=tensorflow.python.estimator.api" "${estimator_api_init_list_file}" COMMENT "Generating __init__.py files for Python API." diff --git a/tensorflow/contrib/cmake/tf_stream_executor.cmake b/tensorflow/contrib/cmake/tf_stream_executor.cmake index 9a37b681194d4ef82b27a0160dd969f733ecad67..6d634cb1709910f366c7ca538d28bd802b2a7c63 100644 --- a/tensorflow/contrib/cmake/tf_stream_executor.cmake +++ b/tensorflow/contrib/cmake/tf_stream_executor.cmake @@ -64,8 +64,6 @@ file(GLOB tf_stream_executor_srcs if (tensorflow_ENABLE_GPU) file(GLOB tf_stream_executor_gpu_srcs "${tensorflow_source_dir}/tensorflow/stream_executor/cuda/*.cc" - "${tensorflow_source_dir}/tensorflow/compiler/xla/statusor.h" - "${tensorflow_source_dir}/tensorflow/compiler/xla/statusor.cc" ) if (NOT tensorflow_BUILD_CC_TESTS) file(GLOB tf_stream_executor_gpu_tests @@ -76,11 +74,11 @@ if (tensorflow_ENABLE_GPU) list(APPEND tf_stream_executor_srcs ${tf_stream_executor_gpu_srcs}) endif() -#file(GLOB_RECURSE tf_stream_executor_test_srcs -# "${tensorflow_source_dir}/tensorflow/stream_executor/*_test.cc" -# "${tensorflow_source_dir}/tensorflow/stream_executor/*_test.h" -#) -#list(REMOVE_ITEM tf_stream_executor_srcs ${tf_stream_executor_test_srcs}) +file(GLOB_RECURSE tf_stream_executor_test_srcs + "${tensorflow_source_dir}/tensorflow/stream_executor/*test.cc" + "${tensorflow_source_dir}/tensorflow/stream_executor/lib/*test.h" +) +list(REMOVE_ITEM tf_stream_executor_srcs ${tf_stream_executor_test_srcs}) if (NOT WIN32) set (CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -lgomp") diff --git a/tensorflow/contrib/copy_graph/python/util/copy_elements.py b/tensorflow/contrib/copy_graph/python/util/copy_elements.py index a0dd3881a86c19e47ccb65f84a2477a55626b81c..5931c8a27996534cca80797e8b840559c124297c 100644 --- a/tensorflow/contrib/copy_graph/python/util/copy_elements.py +++ b/tensorflow/contrib/copy_graph/python/util/copy_elements.py @@ -18,7 +18,7 @@ These functions allow for recursive copying of elements (ops and variables) from one graph to another. The copied elements are initialized inside a user-specified scope in the other graph. There are separate functions to copy ops and variables. -There is also a function to retrive the copied version of an op from the +There is also a function to retrieve the copied version of an op from the first graph inside a scope in the second graph. @@copy_op_to_graph @@ -77,7 +77,7 @@ def copy_variable_to_graph(org_instance, to_graph, scope=''): else: collections.append(scope + '/' + name) - #See if its trainable. + #See if it's trainable. trainable = ( org_instance in org_instance.graph.get_collection( ops.GraphKeys.TRAINABLE_VARIABLES)) @@ -162,7 +162,7 @@ def copy_op_to_graph(org_instance, to_graph, variables, scope=''): if isinstance(org_instance, ops.Tensor): - #If its a Tensor, it is one of the outputs of the underlying + #If it's a Tensor, it is one of the outputs of the underlying #op. Therefore, copy the op itself and return the appropriate #output. op = org_instance.op diff --git a/tensorflow/contrib/data/__init__.py b/tensorflow/contrib/data/__init__.py index 156538b4e01bf1a1ccca0fca1e309b1d37b6dbc0..675330716b2f53edabb61f3ecb37aaed20c4eb90 100644 --- a/tensorflow/contrib/data/__init__.py +++ b/tensorflow/contrib/data/__init__.py @@ -34,6 +34,7 @@ See @{$guide/datasets$Importing Data} for an overview. @@batch_and_drop_remainder @@bucket_by_sequence_length @@choose_from_datasets +@@copy_to_device @@dense_to_sparse_batch @@enumerate_dataset @@ -86,6 +87,7 @@ from tensorflow.contrib.data.python.ops.interleave_ops import sample_from_datase from tensorflow.contrib.data.python.ops.interleave_ops import sloppy_interleave from tensorflow.contrib.data.python.ops.iterator_ops import CheckpointInputPipelineHook from tensorflow.contrib.data.python.ops.iterator_ops import make_saveable_from_iterator +from tensorflow.contrib.data.python.ops.prefetching_ops import copy_to_device from tensorflow.contrib.data.python.ops.prefetching_ops import prefetch_to_device from tensorflow.contrib.data.python.ops.random_ops import RandomDataset from tensorflow.contrib.data.python.ops.readers import CsvDataset diff --git a/tensorflow/contrib/data/kernels/prefetching_kernels.cc b/tensorflow/contrib/data/kernels/prefetching_kernels.cc index a2bfce03620a1482f5b21cbf23c66833bc5cd480..b3d464d7165d53cf198072e06214f7d5e982073d 100644 --- a/tensorflow/contrib/data/kernels/prefetching_kernels.cc +++ b/tensorflow/contrib/data/kernels/prefetching_kernels.cc @@ -40,7 +40,8 @@ class FunctionBufferingResource : public ResourceBase { const NameAttrList& func, int64 buffer_size, const string& source_device, const string& target_device, - const std::vector& func_args) + const std::vector& func_args, + const DataTypeVector& output_types) : lib_(lib), pflr_(std::move(pflr)), func_(func), @@ -48,6 +49,7 @@ class FunctionBufferingResource : public ResourceBase { source_device_(source_device), target_device_(target_device), func_args_(func_args), + output_types_(output_types), handle_(kInvalidHandle), is_buffering_(false), end_of_sequence_(false), @@ -176,6 +178,13 @@ class FunctionBufferingResource : public ResourceBase { AllocatorAttributes arg_alloc_attr; arg_alloc_attr.set_on_host(true); opts.args_alloc_attrs.push_back(arg_alloc_attr); + for (const auto& dtype : output_types_) { + AllocatorAttributes ret_alloc_attrs; + if (DataTypeAlwaysOnHost(dtype)) { + ret_alloc_attrs.set_on_host(true); + } + opts.rets_alloc_attrs.push_back(ret_alloc_attrs); + } if (opts.source_device != target_device_) { opts.remote_execution = true; } @@ -233,6 +242,7 @@ class FunctionBufferingResource : public ResourceBase { const string source_device_; const string target_device_; const std::vector func_args_; + const DataTypeVector output_types_; FunctionLibraryRuntime::Handle handle_ GUARDED_BY(mu_); std::deque buffer_ GUARDED_BY(mu_); std::deque requests_ GUARDED_BY(mu_); @@ -250,6 +260,7 @@ class FunctionBufferResourceHandleOp : public OpKernel { OP_REQUIRES_OK(ctx, ctx->GetAttr("buffer_size", &buffer_size_)); OP_REQUIRES_OK(ctx, ctx->GetAttr("container", &container_)); OP_REQUIRES_OK(ctx, ctx->GetAttr("shared_name", &name_)); + OP_REQUIRES_OK(ctx, ctx->GetAttr("output_types", &output_types_)); } ~FunctionBufferResourceHandleOp() override { @@ -269,18 +280,20 @@ class FunctionBufferResourceHandleOp : public OpKernel { std::vector func_args; func_args.push_back(*string_arg); + const string& source_device = ctx->device()->name(); + // Obtain and canonicalize target_device. const Tensor* target_arg; OP_REQUIRES_OK(ctx, ctx->input("target_device", &target_arg)); - const string& target_device = - DeviceNameUtils::CanonicalizeDeviceName(target_arg->scalar()()); + string target_device; + OP_REQUIRES_OK(ctx, DeviceNameUtils::CanonicalizeDeviceName( + target_arg->scalar()(), source_device, + &target_device)); FunctionLibraryRuntime* lib = ctx->function_library(); OP_REQUIRES(ctx, lib != nullptr, errors::Internal("No function library is provided.")); - const string& source_device = ctx->device()->name(); - mutex_lock l(mu_); if (!initialized_) { OP_REQUIRES_OK(ctx, cinfo_.Init(ctx->resource_manager(), def())); @@ -297,7 +310,7 @@ class FunctionBufferResourceHandleOp : public OpKernel { this](FunctionBufferingResource** ptr) { *ptr = new FunctionBufferingResource( clone_lib, std::move(pflr), func_, buffer_size_, - source_device, target_device, func_args); + source_device, target_device, func_args, output_types_); return Status::OK(); })); core::ScopedUnref s(buffer); @@ -319,6 +332,7 @@ class FunctionBufferResourceHandleOp : public OpKernel { int64 buffer_size_; string container_; string name_; + DataTypeVector output_types_; }; REGISTER_KERNEL_BUILDER(Name("FunctionBufferingResource") diff --git a/tensorflow/contrib/data/ops/dataset_ops.cc b/tensorflow/contrib/data/ops/dataset_ops.cc index f48e96509a193266d5d43453291c5e463f088117..8413fcaf872f49f654c6a1327a14d5c44bdd815a 100644 --- a/tensorflow/contrib/data/ops/dataset_ops.cc +++ b/tensorflow/contrib/data/ops/dataset_ops.cc @@ -104,6 +104,7 @@ REGISTER_OP("FunctionBufferingResource") .Attr("container: string") .Attr("f: func") .Attr("buffer_size: int") + .Attr("output_types: list(type)") .SetShapeFn(shape_inference::UnknownShape) .Doc(R"doc( Creates a resource that fills up a buffer by making function calls. @@ -117,6 +118,7 @@ container: If non-empty, this resource is placed in the given container. Otherwise, a default container is used. shared_name: If non-empty, this resource will be shared under the given name across multiple sessions. +output_types: The type list for the return values. )doc"); REGISTER_OP("FunctionBufferingResourceGetNext") diff --git a/tensorflow/contrib/data/python/kernel_tests/BUILD b/tensorflow/contrib/data/python/kernel_tests/BUILD index d81654e039c53e5b9434288352ef1b2416a4b7e8..9a454efc4ca68336180213bd812aeebf5189498c 100644 --- a/tensorflow/contrib/data/python/kernel_tests/BUILD +++ b/tensorflow/contrib/data/python/kernel_tests/BUILD @@ -188,6 +188,7 @@ py_test( "optonly", ], deps = [ + "//tensorflow/contrib/data/python/ops:batching", "//tensorflow/contrib/data/python/ops:error_ops", "//tensorflow/python:array_ops", "//tensorflow/python:client_testlib", @@ -228,9 +229,11 @@ cuda_py_test( "//tensorflow/python:framework_test_lib", "//tensorflow/python:function", "//tensorflow/python:resource_variable_ops", + "//tensorflow/python/compat:compat", "//tensorflow/python/data/ops:dataset_ops", "//tensorflow/python/data/ops:iterator_ops", ], + tags = ["no_windows_gpu"], ) py_test( @@ -465,6 +468,28 @@ py_test( ], ) +py_test( + name = "window_dataset_op_test", + size = "medium", + srcs = ["window_dataset_op_test.py"], + srcs_version = "PY2AND3", + tags = [ + "no_pip", + ], + deps = [ + "//tensorflow/contrib/data/python/ops:batching", + "//tensorflow/contrib/data/python/ops:grouping", + "//tensorflow/python:array_ops", + "//tensorflow/python:client_testlib", + "//tensorflow/python:dtypes", + "//tensorflow/python:math_ops", + "//tensorflow/python:sparse_tensor", + "//tensorflow/python/data/ops:dataset_ops", + "//third_party/py/numpy", + "@absl_py//absl/testing:parameterized", + ], +) + py_test( name = "writer_ops_test", size = "small", diff --git a/tensorflow/contrib/data/python/kernel_tests/batch_dataset_op_test.py b/tensorflow/contrib/data/python/kernel_tests/batch_dataset_op_test.py index af97fbf87aee5f7005f9d266ba9b1b6cf109a2ec..42adfd17f07e508f25d8b351c791fa519eca8bd9 100644 --- a/tensorflow/contrib/data/python/kernel_tests/batch_dataset_op_test.py +++ b/tensorflow/contrib/data/python/kernel_tests/batch_dataset_op_test.py @@ -293,7 +293,7 @@ class BatchDatasetTest(test.TestCase, parameterized.TestCase): ph2: np.arange(8).astype(np.int32) }) with self.assertRaises(errors.InvalidArgumentError): - print(sess.run(next_element)) + sess.run(next_element) # No 0th dimension (i.e. scalar value) for one component. sess.run( @@ -303,7 +303,7 @@ class BatchDatasetTest(test.TestCase, parameterized.TestCase): ph2: 7 }) with self.assertRaises(errors.InvalidArgumentError): - print(sess.run(next_element)) + sess.run(next_element) def testBatchAndDropRemainder(self): components = (np.arange(7), diff --git a/tensorflow/contrib/data/python/kernel_tests/bucketing_test.py b/tensorflow/contrib/data/python/kernel_tests/bucketing_test.py index 5fc7e51d814901985d33525b782434386c3ad18a..2022c1f2bdd09cdf43a993b3666335ce468a40ba 100644 --- a/tensorflow/contrib/data/python/kernel_tests/bucketing_test.py +++ b/tensorflow/contrib/data/python/kernel_tests/bucketing_test.py @@ -616,7 +616,44 @@ class BucketBySequenceLength(test.TestCase): batch_sizes = batch_sizes[:-1] self.assertEqual(sum(batch_sizes_val), sum(batch_sizes)) self.assertEqual(sorted(batch_sizes), sorted(batch_sizes_val)) - self.assertEqual(sorted(boundaries), sorted(lengths_val)) + self.assertEqual([boundary - 1 for boundary in sorted(boundaries)], + sorted(lengths_val)) + + def testPadToBoundaryNoExtraneousPadding(self): + + boundaries = [3, 7, 11] + batch_sizes = [2, 2, 2, 2] + lengths = range(1, 11) + + def element_gen(): + for length in lengths: + yield ([1] * length,) + + element_len = lambda element: array_ops.shape(element)[0] + dataset = dataset_ops.Dataset.from_generator( + element_gen, (dtypes.int64,), ([None],)).apply( + grouping.bucket_by_sequence_length( + element_len, boundaries, batch_sizes, + pad_to_bucket_boundary=True)) + batch, = dataset.make_one_shot_iterator().get_next() + + with self.test_session() as sess: + batches = [] + for _ in range(5): + batches.append(sess.run(batch)) + with self.assertRaises(errors.OutOfRangeError): + sess.run(batch) + + self.assertAllEqual(batches[0], [[1, 0], + [1, 1]]) + self.assertAllEqual(batches[1], [[1, 1, 1, 0, 0, 0], + [1, 1, 1, 1, 0, 0]]) + self.assertAllEqual(batches[2], [[1, 1, 1, 1, 1, 0], + [1, 1, 1, 1, 1, 1]]) + self.assertAllEqual(batches[3], [[1, 1, 1, 1, 1, 1, 1, 0, 0, 0], + [1, 1, 1, 1, 1, 1, 1, 1, 0, 0]]) + self.assertAllEqual(batches[4], [[1, 1, 1, 1, 1, 1, 1, 1, 1, 0], + [1, 1, 1, 1, 1, 1, 1, 1, 1, 1]]) def testTupleElements(self): diff --git a/tensorflow/contrib/data/python/kernel_tests/map_dataset_op_test.py b/tensorflow/contrib/data/python/kernel_tests/map_dataset_op_test.py index 270a2297b4d7b4fc44e3d1fa0aea8c9dfa5f39d3..b7025f3802c0c280981df20c86747e49fdf2274f 100644 --- a/tensorflow/contrib/data/python/kernel_tests/map_dataset_op_test.py +++ b/tensorflow/contrib/data/python/kernel_tests/map_dataset_op_test.py @@ -17,19 +17,28 @@ from __future__ import absolute_import from __future__ import division from __future__ import print_function +import hashlib +import itertools import os +import time import numpy as np +from tensorflow.contrib.data.python.ops import batching from tensorflow.contrib.data.python.ops import error_ops +from tensorflow.core.protobuf import config_pb2 +from tensorflow.python.client import session from tensorflow.python.data.ops import dataset_ops from tensorflow.python.framework import errors from tensorflow.python.framework import ops from tensorflow.python.ops import array_ops from tensorflow.python.ops import io_ops +from tensorflow.python.ops import math_ops from tensorflow.python.platform import test from tensorflow.python.util import compat +_NUMPY_RANDOM_SEED = 42 + class MapDatasetTest(test.TestCase): @@ -135,5 +144,125 @@ class MapDatasetTest(test.TestCase): sess.run(get_next) +class MapDatasetBenchmark(test.Benchmark): + + # The purpose of this benchmark is to compare the performance of chaining vs + # fusing of the map and batch transformations across various configurations. + # + # NOTE: It is recommended to build the benchmark with + # `-c opt --copt=-mavx --copt=-mavx2 --copt=-mfma --copt=-gmlt` + # and execute it on a machine with at least 32 CPU cores. + def benchmarkMapAndBatch(self): + + # Sequential pipeline configurations. + seq_elem_size_series = itertools.product([1], [1], [1, 2, 4, 8], [16]) + seq_batch_size_series = itertools.product([1], [1], [1], [8, 16, 32, 64]) + + # Parallel pipeline configuration. + par_elem_size_series = itertools.product([32], [32], [1, 2, 4, 8], [256]) + par_batch_size_series = itertools.product([32], [32], [1], + [128, 256, 512, 1024]) + par_num_calls_series = itertools.product([8, 16, 32, 64], [32], [1], [512]) + par_inter_op_series = itertools.product([32], [8, 16, 32, 64], [1], [512]) + + def name(method, label, num_calls, inter_op, element_size, batch_size): + return ("%s_id_%s_num_calls_%d_inter_op_%d_elem_size_%d_batch_size_%d" % ( + method, + hashlib.sha1(label).hexdigest(), + num_calls, + inter_op, + element_size, + batch_size, + )) + + def benchmark(label, series): + + print("%s:" % label) + for num_calls, inter_op, element_size, batch_size in series: + + num_iters = 1024 // ( + (element_size * batch_size) // min(num_calls, inter_op)) + k = 1024 * 1024 + dataset = dataset_ops.Dataset.from_tensors((np.random.rand( + element_size, 4 * k), np.random.rand(4 * k, 1))).repeat() + + chained_dataset = dataset.map( + math_ops.matmul, + num_parallel_calls=num_calls).batch(batch_size=batch_size) + chained_iterator = chained_dataset.make_one_shot_iterator() + chained_get_next = chained_iterator.get_next() + + chained_deltas = [] + with session.Session( + config=config_pb2.ConfigProto( + inter_op_parallelism_threads=inter_op, + use_per_session_threads=True)) as sess: + for _ in range(5): + sess.run(chained_get_next.op) + for _ in range(num_iters): + start = time.time() + sess.run(chained_get_next.op) + end = time.time() + chained_deltas.append(end - start) + + fused_dataset = dataset = dataset.apply( + batching.map_and_batch( + math_ops.matmul, + num_parallel_calls=num_calls, + batch_size=batch_size)) + fused_iterator = fused_dataset.make_one_shot_iterator() + fused_get_next = fused_iterator.get_next() + + fused_deltas = [] + with session.Session( + config=config_pb2.ConfigProto( + inter_op_parallelism_threads=inter_op, + use_per_session_threads=True)) as sess: + + for _ in range(5): + sess.run(fused_get_next.op) + for _ in range(num_iters): + start = time.time() + sess.run(fused_get_next.op) + end = time.time() + fused_deltas.append(end - start) + + print( + "batch size: %d, num parallel calls: %d, inter-op parallelism: %d, " + "element size: %d, num iters: %d\nchained wall time: %f (median), " + "%f (mean), %f (stddev), %f (min), %f (max)\n fused wall time: " + "%f (median), %f (mean), %f (stddev), %f (min), %f (max)\n " + "chained/fused: %.2fx (median), %.2fx (mean)" % + (batch_size, num_calls, inter_op, element_size, num_iters, + np.median(chained_deltas), np.mean(chained_deltas), + np.std(chained_deltas), np.min(chained_deltas), + np.max(chained_deltas), np.median(fused_deltas), + np.mean(fused_deltas), np.std(fused_deltas), np.min(fused_deltas), + np.max(fused_deltas), + np.median(chained_deltas) / np.median(fused_deltas), + np.mean(chained_deltas) / np.mean(fused_deltas))) + + self.report_benchmark( + iters=num_iters, + wall_time=np.median(chained_deltas), + name=name("chained", label, num_calls, inter_op, element_size, + batch_size)) + + self.report_benchmark( + iters=num_iters, + wall_time=np.median(fused_deltas), + name=name("fused", label, num_calls, inter_op, element_size, + batch_size)) + + print("") + + np.random.seed(_NUMPY_RANDOM_SEED) + benchmark("Sequential element size evaluation", seq_elem_size_series) + benchmark("Sequential batch size evaluation", seq_batch_size_series) + benchmark("Parallel element size evaluation", par_elem_size_series) + benchmark("Parallel batch size evaluation", par_batch_size_series) + benchmark("Transformation parallelism evaluation", par_num_calls_series) + benchmark("Threadpool size evaluation", par_inter_op_series) + if __name__ == "__main__": test.main() diff --git a/tensorflow/contrib/data/python/kernel_tests/optimize_dataset_op_test.py b/tensorflow/contrib/data/python/kernel_tests/optimize_dataset_op_test.py index e35be8a23f3706bd170c09b967b4f419fc9a626e..21eebccd1113f56fccad7951ea97369152b1f698 100644 --- a/tensorflow/contrib/data/python/kernel_tests/optimize_dataset_op_test.py +++ b/tensorflow/contrib/data/python/kernel_tests/optimize_dataset_op_test.py @@ -35,8 +35,6 @@ class OptimizeDatasetTest(test.TestCase): with self.test_session() as sess: graph = graph_pb2.GraphDef().FromString( sess.run(dataset._as_serialized_graph())) - self.assertTrue( - all([node.op != "MapAndBatchDatasetV2" for node in graph.node])) self.assertAllEqual([x * x for x in range(10)], sess.run(get_next)) with self.assertRaises(errors.OutOfRangeError): sess.run(get_next) @@ -50,8 +48,6 @@ class OptimizeDatasetTest(test.TestCase): with self.test_session() as sess: graph = graph_pb2.GraphDef().FromString( sess.run(dataset._as_serialized_graph())) - self.assertTrue( - all([node.op != "MapAndBatchDatasetV2" for node in graph.node])) self.assertAllEqual([x * x for x in range(10)], sess.run(get_next)) with self.assertRaises(errors.OutOfRangeError): sess.run(get_next) @@ -65,12 +61,21 @@ class OptimizeDatasetTest(test.TestCase): with self.test_session() as sess: graph = graph_pb2.GraphDef().FromString( sess.run(dataset._as_serialized_graph())) - self.assertTrue( - any([node.op == "MapAndBatchDatasetV2" for node in graph.node])) self.assertAllEqual([x * x for x in range(10)], sess.run(get_next)) with self.assertRaises(errors.OutOfRangeError): sess.run(get_next) + def testFunctionLibraryDefinitionModification(self): + dataset = dataset_ops.Dataset.from_tensors(0).map(lambda x: x).apply( + optimization.optimize(["_test_only_function_rename"])) + iterator = dataset.make_one_shot_iterator() + get_next = iterator.get_next() + + with self.test_session() as sess: + with self.assertRaisesRegexp(errors.NotFoundError, + "Function .* is not defined."): + sess.run(get_next) + if __name__ == "__main__": test.main() diff --git a/tensorflow/contrib/data/python/kernel_tests/prefetching_ops_test.py b/tensorflow/contrib/data/python/kernel_tests/prefetching_ops_test.py index b08132cd72254326d965907a1fdafb8a820926a1..82543b10395116a273954bd71bd5e5fde6679585 100644 --- a/tensorflow/contrib/data/python/kernel_tests/prefetching_ops_test.py +++ b/tensorflow/contrib/data/python/kernel_tests/prefetching_ops_test.py @@ -21,6 +21,7 @@ import threading from tensorflow.contrib.data.python.ops import prefetching_ops from tensorflow.core.protobuf import config_pb2 +from tensorflow.python.compat import compat from tensorflow.python.data.ops import dataset_ops from tensorflow.python.data.ops import iterator_ops from tensorflow.python.framework import constant_op @@ -68,6 +69,7 @@ class PrefetchingKernelsOpsTest(test.TestCase): with ops.device(device1): buffer_resource_handle = prefetching_ops.function_buffering_resource( f=_remote_fn, + output_types=[dtypes.float32], target_device=target, string_arg=ds_iterator_handle, buffer_size=3, @@ -85,8 +87,7 @@ class PrefetchingKernelsOpsTest(test.TestCase): return (prefetch_op, reset_op, destroy_op) def _prefetch_fn_helper_one_shot(self, buffer_name, device0, device1): - worker_config = config_pb2.ConfigProto() - worker_config.device_count["CPU"] = 2 + worker_config = config_pb2.ConfigProto(device_count={"CPU": 2}) ds, ds_iterator = self._create_ds_and_iterator(device0, initializable=False) prefetch_op, _, destroy_op = self._create_ops(ds, ds_iterator, buffer_name, @@ -125,8 +126,7 @@ class PrefetchingKernelsOpsTest(test.TestCase): "/job:localhost/replica:0/task:0/gpu:0") def testReinitialization(self): - worker_config = config_pb2.ConfigProto() - worker_config.device_count["CPU"] = 2 + worker_config = config_pb2.ConfigProto(device_count={"CPU": 2}) device0 = "/job:localhost/replica:0/task:0/cpu:0" device1 = "/job:localhost/replica:0/task:0/cpu:1" @@ -166,8 +166,7 @@ class PrefetchingKernelsOpsTest(test.TestCase): sess.run(destroy_op) def testReinitializationOutOfRange(self): - worker_config = config_pb2.ConfigProto() - worker_config.device_count["CPU"] = 2 + worker_config = config_pb2.ConfigProto(device_count={"CPU": 2}) device0 = "/job:localhost/replica:0/task:0/cpu:0" device1 = "/job:localhost/replica:0/task:0/cpu:1" @@ -201,6 +200,49 @@ class PrefetchingKernelsOpsTest(test.TestCase): sess.run(destroy_op) + def testStringsGPU(self): + if not test_util.is_gpu_available(): + self.skipTest("No GPU available") + + device0 = "/job:localhost/replica:0/task:0/cpu:0" + device1 = "/job:localhost/replica:0/task:0/gpu:0" + + ds = dataset_ops.Dataset.from_tensor_slices(["a", "b", "c"]) + ds_iterator = ds.make_one_shot_iterator() + ds_iterator_handle = ds_iterator.string_handle() + + @function.Defun(dtypes.string) + def _remote_fn(h): + remote_iterator = iterator_ops.Iterator.from_string_handle( + h, ds.output_types, ds.output_shapes) + return remote_iterator.get_next() + + target = constant_op.constant(device0) + with ops.device(device1): + buffer_resource_handle = prefetching_ops.function_buffering_resource( + f=_remote_fn, + output_types=[dtypes.string], + target_device=target, + string_arg=ds_iterator_handle, + buffer_size=3, + shared_name="strings") + + with ops.device(device1): + prefetch_op = prefetching_ops.function_buffering_resource_get_next( + function_buffer_resource=buffer_resource_handle, + output_types=[dtypes.string]) + destroy_op = resource_variable_ops.destroy_resource_op( + buffer_resource_handle, ignore_lookup_error=True) + + with self.test_session() as sess: + self.assertEqual([b"a"], sess.run(prefetch_op)) + self.assertEqual([b"b"], sess.run(prefetch_op)) + self.assertEqual([b"c"], sess.run(prefetch_op)) + with self.assertRaises(errors.OutOfRangeError): + sess.run(prefetch_op) + + sess.run(destroy_op) + class PrefetchToDeviceTest(test.TestCase): @@ -227,14 +269,43 @@ class PrefetchToDeviceTest(test.TestCase): self.assertEqual(dtypes.int64, next_element.dtype) self.assertEqual([], next_element.shape) - worker_config = config_pb2.ConfigProto() - worker_config.device_count["CPU"] = 2 + worker_config = config_pb2.ConfigProto(device_count={"CPU": 2}) with self.test_session(config=worker_config) as sess: for i in range(10): self.assertEqual(i, sess.run(next_element)) with self.assertRaises(errors.OutOfRangeError): sess.run(next_element) + def testPrefetchToSameDevice(self): + host_dataset = dataset_ops.Dataset.range(10) + device_dataset = host_dataset.apply( + prefetching_ops.prefetch_to_device( + "/job:localhost/replica:0/task:0/device:CPU:0")) + + # NOTE(mrry): This device block creates the "host" dataset and iterator on + # /cpu:0, and ensures that the prefetching is across devices. In typical use + # this would not be necessary, because the GPU device would not support any + # of the dataset-related ops. + with ops.device("/cpu:0"): + iterator = device_dataset.make_one_shot_iterator() + + self.assertEqual(host_dataset.output_types, device_dataset.output_types) + self.assertEqual(host_dataset.output_types, iterator.output_types) + self.assertEqual(host_dataset.output_shapes, device_dataset.output_shapes) + self.assertEqual(host_dataset.output_shapes, iterator.output_shapes) + self.assertEqual(host_dataset.output_classes, device_dataset.output_classes) + self.assertEqual(host_dataset.output_classes, iterator.output_classes) + + next_element = iterator.get_next() + self.assertEqual(dtypes.int64, next_element.dtype) + self.assertEqual([], next_element.shape) + + with self.test_session() as sess: + for i in range(10): + self.assertEqual(i, sess.run(next_element)) + with self.assertRaises(errors.OutOfRangeError): + sess.run(next_element) + def testPrefetchDictToDevice(self): host_dataset = dataset_ops.Dataset.range(10).map(lambda x: {"a": x}) device_dataset = host_dataset.apply( @@ -258,8 +329,7 @@ class PrefetchToDeviceTest(test.TestCase): self.assertEqual(dtypes.int64, next_element["a"].dtype) self.assertEqual([], next_element["a"].shape) - worker_config = config_pb2.ConfigProto() - worker_config.device_count["CPU"] = 2 + worker_config = config_pb2.ConfigProto(device_count={"CPU": 2}) with self.test_session(config=worker_config) as sess: for i in range(10): self.assertEqual({"a": i}, sess.run(next_element)) @@ -292,8 +362,7 @@ class PrefetchToDeviceTest(test.TestCase): next_element = iterator.get_next() self.assertEqual(dtypes.int64, next_element.dtype) - worker_config = config_pb2.ConfigProto() - worker_config.device_count["CPU"] = 2 + worker_config = config_pb2.ConfigProto(device_count={"CPU": 2}) with self.test_session(config=worker_config) as sess: for i in range(10): actual = sess.run(next_element) @@ -343,8 +412,7 @@ class PrefetchToDeviceTest(test.TestCase): self.assertEqual(dtypes.int64, next_element.dtype) self.assertEqual([], next_element.shape) - worker_config = config_pb2.ConfigProto() - worker_config.device_count["CPU"] = 2 + worker_config = config_pb2.ConfigProto(device_count={"CPU": 2}) with self.test_session(config=worker_config) as sess: sess.run(iterator.initializer) for i in range(5): @@ -377,5 +445,467 @@ class PrefetchToDeviceTest(test.TestCase): sess.run(next_element) +class CopyToDeviceTest(test.TestCase): + + def testCopyToDevice(self): + host_dataset = dataset_ops.Dataset.range(10) + device_dataset = host_dataset.apply( + prefetching_ops.copy_to_device("/cpu:1")) + + with ops.device("/cpu:1"): + iterator = device_dataset.make_one_shot_iterator() + next_element = iterator.get_next() + + self.assertEqual(host_dataset.output_types, device_dataset.output_types) + self.assertEqual(host_dataset.output_types, iterator.output_types) + self.assertEqual(host_dataset.output_shapes, device_dataset.output_shapes) + self.assertEqual(host_dataset.output_shapes, iterator.output_shapes) + self.assertEqual(host_dataset.output_classes, device_dataset.output_classes) + self.assertEqual(host_dataset.output_classes, iterator.output_classes) + + self.assertEqual(dtypes.int64, next_element.dtype) + self.assertEqual([], next_element.shape) + + worker_config = config_pb2.ConfigProto(device_count={"CPU": 2}) + with self.test_session(config=worker_config) as sess: + for i in range(10): + self.assertEqual(i, sess.run(next_element)) + with self.assertRaises(errors.OutOfRangeError): + sess.run(next_element) + + def testCopyToDeviceInt32(self): + host_dataset = dataset_ops.Dataset.from_tensors([0, 1, 2, 3]) + device_dataset = host_dataset.apply( + prefetching_ops.copy_to_device("/cpu:1")) + + with ops.device("/cpu:1"): + iterator = device_dataset.make_one_shot_iterator() + next_element = iterator.get_next() + + self.assertEqual(host_dataset.output_types, device_dataset.output_types) + self.assertEqual(host_dataset.output_types, iterator.output_types) + self.assertEqual(host_dataset.output_shapes, device_dataset.output_shapes) + self.assertEqual(host_dataset.output_shapes, iterator.output_shapes) + self.assertEqual(host_dataset.output_classes, device_dataset.output_classes) + self.assertEqual(host_dataset.output_classes, iterator.output_classes) + + self.assertEqual(dtypes.int32, next_element.dtype) + self.assertEqual((4,), next_element.shape) + + worker_config = config_pb2.ConfigProto(device_count={"CPU": 2}) + with self.test_session(config=worker_config) as sess: + self.assertAllEqual([0, 1, 2, 3], sess.run(next_element)) + with self.assertRaises(errors.OutOfRangeError): + sess.run(next_element) + + def testCopyToSameDevice(self): + host_dataset = dataset_ops.Dataset.range(10) + device_dataset = host_dataset.apply( + prefetching_ops.copy_to_device("/cpu:0")) + + with ops.device("/cpu:0"): + iterator = device_dataset.make_one_shot_iterator() + next_element = iterator.get_next() + + self.assertEqual(host_dataset.output_types, device_dataset.output_types) + self.assertEqual(host_dataset.output_types, iterator.output_types) + self.assertEqual(host_dataset.output_shapes, device_dataset.output_shapes) + self.assertEqual(host_dataset.output_shapes, iterator.output_shapes) + self.assertEqual(host_dataset.output_classes, device_dataset.output_classes) + self.assertEqual(host_dataset.output_classes, iterator.output_classes) + + self.assertEqual(dtypes.int64, next_element.dtype) + self.assertEqual([], next_element.shape) + + worker_config = config_pb2.ConfigProto(device_count={"CPU": 2}) + with self.test_session(config=worker_config) as sess: + for i in range(10): + self.assertEqual(i, sess.run(next_element)) + with self.assertRaises(errors.OutOfRangeError): + sess.run(next_element) + + def testCopyToDeviceWithPrefetch(self): + host_dataset = dataset_ops.Dataset.range(10) + device_dataset = host_dataset.apply( + prefetching_ops.copy_to_device("/cpu:1")).prefetch(1) + + with ops.device("/cpu:1"): + iterator = device_dataset.make_one_shot_iterator() + next_element = iterator.get_next() + + self.assertEqual(host_dataset.output_types, device_dataset.output_types) + self.assertEqual(host_dataset.output_types, iterator.output_types) + self.assertEqual(host_dataset.output_shapes, device_dataset.output_shapes) + self.assertEqual(host_dataset.output_shapes, iterator.output_shapes) + self.assertEqual(host_dataset.output_classes, device_dataset.output_classes) + self.assertEqual(host_dataset.output_classes, iterator.output_classes) + + self.assertEqual(dtypes.int64, next_element.dtype) + self.assertEqual([], next_element.shape) + + worker_config = config_pb2.ConfigProto(device_count={"CPU": 2}) + with self.test_session(config=worker_config) as sess: + for i in range(10): + self.assertEqual(i, sess.run(next_element)) + with self.assertRaises(errors.OutOfRangeError): + sess.run(next_element) + + def testCopyDictToDevice(self): + host_dataset = dataset_ops.Dataset.range(10).map(lambda x: {"a": x}) + device_dataset = host_dataset.apply( + prefetching_ops.copy_to_device("/cpu:1")) + + with ops.device("/cpu:1"): + iterator = device_dataset.make_one_shot_iterator() + next_element = iterator.get_next() + + self.assertEqual(host_dataset.output_types, device_dataset.output_types) + self.assertEqual(host_dataset.output_types, iterator.output_types) + self.assertEqual(host_dataset.output_shapes, device_dataset.output_shapes) + self.assertEqual(host_dataset.output_shapes, iterator.output_shapes) + self.assertEqual(host_dataset.output_classes, device_dataset.output_classes) + self.assertEqual(host_dataset.output_classes, iterator.output_classes) + + self.assertEqual(dtypes.int64, next_element["a"].dtype) + self.assertEqual([], next_element["a"].shape) + + worker_config = config_pb2.ConfigProto(device_count={"CPU": 2}) + with self.test_session(config=worker_config) as sess: + for i in range(10): + self.assertEqual({"a": i}, sess.run(next_element)) + with self.assertRaises(errors.OutOfRangeError): + sess.run(next_element) + + def testCopyDictToDeviceWithPrefetch(self): + host_dataset = dataset_ops.Dataset.range(10).map(lambda x: {"a": x}) + device_dataset = host_dataset.apply( + prefetching_ops.copy_to_device("/cpu:1")).prefetch(1) + + with ops.device("/cpu:1"): + iterator = device_dataset.make_one_shot_iterator() + next_element = iterator.get_next() + + self.assertEqual(host_dataset.output_types, device_dataset.output_types) + self.assertEqual(host_dataset.output_types, iterator.output_types) + self.assertEqual(host_dataset.output_shapes, device_dataset.output_shapes) + self.assertEqual(host_dataset.output_shapes, iterator.output_shapes) + self.assertEqual(host_dataset.output_classes, device_dataset.output_classes) + self.assertEqual(host_dataset.output_classes, iterator.output_classes) + + self.assertEqual(dtypes.int64, next_element["a"].dtype) + self.assertEqual([], next_element["a"].shape) + + worker_config = config_pb2.ConfigProto(device_count={"CPU": 2}) + with self.test_session(config=worker_config) as sess: + for i in range(10): + self.assertEqual({"a": i}, sess.run(next_element)) + with self.assertRaises(errors.OutOfRangeError): + sess.run(next_element) + + def testCopySparseTensorsToDevice(self): + + def make_tensor(i): + return sparse_tensor.SparseTensorValue( + indices=[[0, 0]], values=(i * [1]), dense_shape=[2, 2]) + + host_dataset = dataset_ops.Dataset.range(10).map(make_tensor) + + device_dataset = host_dataset.apply( + prefetching_ops.copy_to_device("/cpu:1")) + + with ops.device("/cpu:1"): + iterator = device_dataset.make_one_shot_iterator() + next_element = iterator.get_next() + + self.assertEqual(host_dataset.output_types, device_dataset.output_types) + self.assertEqual(host_dataset.output_types, iterator.output_types) + self.assertEqual(host_dataset.output_shapes, device_dataset.output_shapes) + self.assertEqual(host_dataset.output_shapes, iterator.output_shapes) + self.assertEqual(host_dataset.output_classes, device_dataset.output_classes) + self.assertEqual(host_dataset.output_classes, iterator.output_classes) + + self.assertEqual(dtypes.int64, next_element.dtype) + + worker_config = config_pb2.ConfigProto(device_count={"CPU": 2}) + with self.test_session(config=worker_config) as sess: + for i in range(10): + actual = sess.run(next_element) + self.assertAllEqual([i], actual.values) + self.assertAllEqual([[0, 0]], actual.indices) + self.assertAllEqual([2, 2], actual.dense_shape) + with self.assertRaises(errors.OutOfRangeError): + sess.run(next_element) + + def testCopySparseTensorsToDeviceWithPrefetch(self): + + def make_tensor(i): + return sparse_tensor.SparseTensorValue( + indices=[[0, 0]], values=(i * [1]), dense_shape=[2, 2]) + + host_dataset = dataset_ops.Dataset.range(10).map(make_tensor) + + device_dataset = host_dataset.apply( + prefetching_ops.copy_to_device("/cpu:1")).prefetch(1) + + with ops.device("/cpu:1"): + iterator = device_dataset.make_one_shot_iterator() + next_element = iterator.get_next() + + self.assertEqual(host_dataset.output_types, device_dataset.output_types) + self.assertEqual(host_dataset.output_types, iterator.output_types) + self.assertEqual(host_dataset.output_shapes, device_dataset.output_shapes) + self.assertEqual(host_dataset.output_shapes, iterator.output_shapes) + self.assertEqual(host_dataset.output_classes, device_dataset.output_classes) + self.assertEqual(host_dataset.output_classes, iterator.output_classes) + + self.assertEqual(dtypes.int64, next_element.dtype) + + worker_config = config_pb2.ConfigProto(device_count={"CPU": 2}) + with self.test_session(config=worker_config) as sess: + for i in range(10): + actual = sess.run(next_element) + self.assertAllEqual([i], actual.values) + self.assertAllEqual([[0, 0]], actual.indices) + self.assertAllEqual([2, 2], actual.dense_shape) + with self.assertRaises(errors.OutOfRangeError): + sess.run(next_element) + + def testCopyToDeviceGpu(self): + if not test_util.is_gpu_available(): + self.skipTest("No GPU available") + + host_dataset = dataset_ops.Dataset.range(10) + device_dataset = host_dataset.apply( + prefetching_ops.copy_to_device("/gpu:0")) + + with ops.device("/gpu:0"): + iterator = device_dataset.make_initializable_iterator() + next_element = iterator.get_next() + + with self.test_session() as sess: + sess.run(iterator.initializer) + for i in range(10): + self.assertEqual(i, sess.run(next_element)) + with self.assertRaises(errors.OutOfRangeError): + sess.run(next_element) + + def testCopyToDeviceGpuWithPrefetch(self): + if not test_util.is_gpu_available(): + self.skipTest("No GPU available") + + host_dataset = dataset_ops.Dataset.range(10) + device_dataset = host_dataset.apply( + prefetching_ops.copy_to_device("/gpu:0")).prefetch(1) + + with ops.device("/gpu:0"): + iterator = device_dataset.make_initializable_iterator() + next_element = iterator.get_next() + + with self.test_session() as sess: + sess.run(iterator.initializer) + for i in range(10): + self.assertEqual(i, sess.run(next_element)) + with self.assertRaises(errors.OutOfRangeError): + sess.run(next_element) + + def testCopyToDeviceGpuInt32(self): + if not test_util.is_gpu_available(): + self.skipTest("No GPU available") + + host_dataset = dataset_ops.Dataset.from_tensors([0, 1, 2, 3]) + device_dataset = host_dataset.apply( + prefetching_ops.copy_to_device("/gpu:0")) + + with ops.device("/gpu:0"): + iterator = device_dataset.make_initializable_iterator() + next_element = iterator.get_next() + + with self.test_session() as sess: + sess.run(iterator.initializer) + self.assertAllEqual([0, 1, 2, 3], sess.run(next_element)) + with self.assertRaises(errors.OutOfRangeError): + sess.run(next_element) + + def testCopyToDeviceGpuInt32AndPrefetch(self): + if not test_util.is_gpu_available(): + self.skipTest("No GPU available") + + host_dataset = dataset_ops.Dataset.from_tensors([0, 1, 2, 3]) + device_dataset = host_dataset.apply( + prefetching_ops.copy_to_device("/gpu:0")).prefetch(1) + + with ops.device("/gpu:0"): + iterator = device_dataset.make_initializable_iterator() + next_element = iterator.get_next() + + with self.test_session() as sess: + sess.run(iterator.initializer) + self.assertAllEqual([0, 1, 2, 3], sess.run(next_element)) + with self.assertRaises(errors.OutOfRangeError): + sess.run(next_element) + + def testCopyToDeviceGpuStrings(self): + if not test_util.is_gpu_available(): + self.skipTest("No GPU available") + + host_dataset = dataset_ops.Dataset.from_tensors(["a", "b", "c"]) + device_dataset = host_dataset.apply( + prefetching_ops.copy_to_device("/gpu:0")) + + with ops.device("/gpu:0"): + iterator = device_dataset.make_initializable_iterator() + next_element = iterator.get_next() + + with self.test_session() as sess: + sess.run(iterator.initializer) + self.assertAllEqual([b"a", b"b", b"c"], sess.run(next_element)) + with self.assertRaises(errors.OutOfRangeError): + sess.run(next_element) + + def testCopyToDeviceGpuStringsAndPrefetch(self): + if not test_util.is_gpu_available(): + self.skipTest("No GPU available") + + host_dataset = dataset_ops.Dataset.from_tensors(["a", "b", "c"]) + device_dataset = host_dataset.apply( + prefetching_ops.copy_to_device("/gpu:0")) + + with ops.device("/gpu:0"): + iterator = device_dataset.make_initializable_iterator() + next_element = iterator.get_next() + + with self.test_session() as sess: + sess.run(iterator.initializer) + self.assertAllEqual([b"a", b"b", b"c"], sess.run(next_element)) + with self.assertRaises(errors.OutOfRangeError): + sess.run(next_element) + + def testCopyToDevicePingPongCPUGPU(self): + if not test_util.is_gpu_available(): + self.skipTest("No GPU available") + + with compat.forward_compatibility_horizon(2018, 8, 4): + host_dataset = dataset_ops.Dataset.range(10) + device_dataset = host_dataset.apply( + prefetching_ops.copy_to_device("/gpu:0", source_device="/cpu:0")) + back_to_cpu_dataset = device_dataset.apply( + prefetching_ops.copy_to_device("/cpu:0", source_device="/gpu:0")) + + with ops.device("/cpu:0"): + iterator = back_to_cpu_dataset.make_initializable_iterator() + next_element = iterator.get_next() + + with self.test_session() as sess: + sess.run(iterator.initializer) + for i in range(10): + self.assertEqual(i, sess.run(next_element)) + with self.assertRaises(errors.OutOfRangeError): + sess.run(next_element) + + def testCopyToDeviceWithReInit(self): + host_dataset = dataset_ops.Dataset.range(10) + device_dataset = host_dataset.apply( + prefetching_ops.copy_to_device("/cpu:1")) + + with ops.device("/cpu:1"): + iterator = device_dataset.make_initializable_iterator() + next_element = iterator.get_next() + + self.assertEqual(host_dataset.output_types, device_dataset.output_types) + self.assertEqual(host_dataset.output_types, iterator.output_types) + self.assertEqual(host_dataset.output_shapes, device_dataset.output_shapes) + self.assertEqual(host_dataset.output_shapes, iterator.output_shapes) + self.assertEqual(host_dataset.output_classes, device_dataset.output_classes) + self.assertEqual(host_dataset.output_classes, iterator.output_classes) + + self.assertEqual(dtypes.int64, next_element.dtype) + self.assertEqual([], next_element.shape) + + worker_config = config_pb2.ConfigProto(device_count={"CPU": 2}) + with self.test_session(config=worker_config) as sess: + sess.run(iterator.initializer) + for i in range(5): + self.assertEqual(i, sess.run(next_element)) + sess.run(iterator.initializer) + for i in range(10): + self.assertEqual(i, sess.run(next_element)) + with self.assertRaises(errors.OutOfRangeError): + sess.run(next_element) + + def testCopyToDeviceWithReInitAndPrefetch(self): + host_dataset = dataset_ops.Dataset.range(10) + device_dataset = host_dataset.apply( + prefetching_ops.copy_to_device("/cpu:1")).prefetch(1) + + with ops.device("/cpu:1"): + iterator = device_dataset.make_initializable_iterator() + next_element = iterator.get_next() + + self.assertEqual(host_dataset.output_types, device_dataset.output_types) + self.assertEqual(host_dataset.output_types, iterator.output_types) + self.assertEqual(host_dataset.output_shapes, device_dataset.output_shapes) + self.assertEqual(host_dataset.output_shapes, iterator.output_shapes) + self.assertEqual(host_dataset.output_classes, device_dataset.output_classes) + self.assertEqual(host_dataset.output_classes, iterator.output_classes) + + self.assertEqual(dtypes.int64, next_element.dtype) + self.assertEqual([], next_element.shape) + + worker_config = config_pb2.ConfigProto(device_count={"CPU": 2}) + with self.test_session(config=worker_config) as sess: + sess.run(iterator.initializer) + for i in range(5): + self.assertEqual(i, sess.run(next_element)) + sess.run(iterator.initializer) + for i in range(10): + self.assertEqual(i, sess.run(next_element)) + with self.assertRaises(errors.OutOfRangeError): + sess.run(next_element) + + def testCopyToDeviceGpuWithReInit(self): + if not test_util.is_gpu_available(): + self.skipTest("No GPU available") + + host_dataset = dataset_ops.Dataset.range(10) + device_dataset = host_dataset.apply( + prefetching_ops.copy_to_device("/gpu:0")) + + with ops.device("/gpu:0"): + iterator = device_dataset.make_initializable_iterator() + next_element = iterator.get_next() + + with self.test_session() as sess: + sess.run(iterator.initializer) + for i in range(5): + self.assertEqual(i, sess.run(next_element)) + sess.run(iterator.initializer) + for i in range(10): + self.assertEqual(i, sess.run(next_element)) + with self.assertRaises(errors.OutOfRangeError): + sess.run(next_element) + + def testCopyToDeviceGpuWithReInitAndPrefetch(self): + if not test_util.is_gpu_available(): + self.skipTest("No GPU available") + + host_dataset = dataset_ops.Dataset.range(10) + device_dataset = host_dataset.apply( + prefetching_ops.copy_to_device("/gpu:0")).prefetch(1) + + with ops.device("/gpu:0"): + iterator = device_dataset.make_initializable_iterator() + next_element = iterator.get_next() + + with self.test_session() as sess: + sess.run(iterator.initializer) + for i in range(5): + self.assertEqual(i, sess.run(next_element)) + sess.run(iterator.initializer) + for i in range(10): + self.assertEqual(i, sess.run(next_element)) + with self.assertRaises(errors.OutOfRangeError): + sess.run(next_element) + + if __name__ == "__main__": test.main() diff --git a/tensorflow/contrib/data/python/kernel_tests/window_dataset_op_test.py b/tensorflow/contrib/data/python/kernel_tests/window_dataset_op_test.py new file mode 100644 index 0000000000000000000000000000000000000000..33d95d67549e1c8d1d9af578fcebbb4f939c418a --- /dev/null +++ b/tensorflow/contrib/data/python/kernel_tests/window_dataset_op_test.py @@ -0,0 +1,523 @@ +# Copyright 2017 The TensorFlow Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================== +"""Tests for the experimental input pipeline ops.""" +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + +from absl.testing import parameterized +import numpy as np + +from tensorflow.contrib.data.python.ops import batching +from tensorflow.contrib.data.python.ops import grouping +from tensorflow.python.data.ops import dataset_ops +from tensorflow.python.framework import dtypes +from tensorflow.python.framework import errors +from tensorflow.python.framework import sparse_tensor +from tensorflow.python.ops import array_ops +from tensorflow.python.ops import sparse_ops +from tensorflow.python.platform import test + + +class WindowDatasetTest(test.TestCase, parameterized.TestCase): + + def _structuredDataset(self, structure, shape, dtype): + if structure is None: + return dataset_ops.Dataset.from_tensors( + array_ops.zeros(shape, dtype=dtype)) + else: + return dataset_ops.Dataset.zip( + tuple([ + self._structuredDataset(substructure, shape, dtype) + for substructure in structure + ])) + + def _structuredElement(self, structure, shape, dtype): + if structure is None: + return array_ops.zeros(shape, dtype=dtype) + else: + return tuple([ + self._structuredElement(substructure, shape, dtype) + for substructure in structure + ]) + + def _assertEqual(self, xs, ys): + self.assertEqual(type(xs), type(ys)) + if isinstance(xs, tuple) and isinstance(ys, tuple): + self.assertEqual(len(xs), len(ys)) + for x, y in zip(xs, ys): + self._assertEqual(x, y) + elif isinstance(xs, np.ndarray) and isinstance(ys, np.ndarray): + self.assertAllEqual(xs, ys) + else: + self.assertEqual(xs, ys) + + @parameterized.parameters( + (None, np.int32([]), dtypes.bool), + (None, np.int32([]), dtypes.int32), + (None, np.int32([]), dtypes.float32), + (None, np.int32([]), dtypes.string), + (None, np.int32([2]), dtypes.int32), + (None, np.int32([2, 2]), dtypes.int32), + ((None, None, None), np.int32([]), dtypes.int32), + ((None, (None, None)), np.int32([]), dtypes.int32), + ) + def testWindowDatasetFlatMap(self, structure, shape, dtype): + """Tests windowing by chaining it with flat map. + + Args: + structure: the input structure + shape: the input shape + dtype: the input data type + """ + + def fn(*args): + if len(args) == 1 and not isinstance(args[0], tuple): + return args[0] + return dataset_ops.Dataset.zip( + tuple([fn(*arg) if isinstance(arg, tuple) else arg for arg in args])) + + dataset = self._structuredDataset(structure, shape, dtype).apply( + grouping.window_dataset(5)).flat_map(fn) + get_next = dataset.make_one_shot_iterator().get_next() + with self.test_session() as sess: + expected = sess.run(self._structuredElement(structure, shape, dtype)) + actual = sess.run(get_next) + self._assertEqual(expected, actual) + + @parameterized.parameters( + (None, np.int32([]), dtypes.bool), + (None, np.int32([]), dtypes.int32), + (None, np.int32([]), dtypes.float32), + (None, np.int32([]), dtypes.string), + (None, np.int32([2]), dtypes.int32), + (None, np.int32([2, 2]), dtypes.int32), + ((None, None, None), np.int32([]), dtypes.int32), + ((None, (None, None)), np.int32([]), dtypes.int32), + ) + def testWindowDatasetBatchDense(self, structure, shape, dtype): + """Tests batching of dense tensor windows. + + Args: + structure: the input structure + shape: the input shape + dtype: the input data type + """ + + def fn(*args): + if len(args) == 1 and not isinstance(args[0], tuple): + return batching.batch_window(args[0]) + + return tuple([ + fn(*arg) if isinstance(arg, tuple) else batching.batch_window(arg) + for arg in args + ]) + + dataset = self._structuredDataset(structure, shape, dtype).repeat(5).apply( + grouping.window_dataset(5)).apply(grouping._map_x_dataset(fn)) + get_next = dataset.make_one_shot_iterator().get_next() + with self.test_session() as sess: + expected = sess.run( + self._structuredElement(structure, np.concatenate( + ([5], shape), axis=0), dtype)) + actual = sess.run(get_next) + self._assertEqual(expected, actual) + + @parameterized.parameters( + (np.int32([]),), + (np.int32([1]),), + (np.int32([1, 2, 3]),), + ) + def testWindowDatasetBatchDenseDynamicShape(self, shape): + """Tests batching of dynamically shaped dense tensor windows. + + Args: + shape: the input shape + """ + + shape_t = array_ops.placeholder(dtypes.int32) + dataset = dataset_ops.Dataset.from_tensors( + array_ops.zeros(shape_t)).repeat(5).apply( + grouping.window_dataset(5)).apply( + grouping._map_x_dataset(batching.batch_window)) + iterator = dataset.make_initializable_iterator() + init_op = iterator.initializer + get_next = iterator.get_next() + with self.test_session() as sess: + sess.run(init_op, {shape_t: shape}) + expected = sess.run( + self._structuredElement(None, np.concatenate(([5], shape), axis=0), + dtypes.int32)) + actual = sess.run(get_next) + self._assertEqual(expected, actual) + + def _make_dense_to_sparse_fn(self, is_scalar): + + def dense_to_sparse_scalar(tensor): + indices = [[]] + values = array_ops.expand_dims(tensor, 0) + shape = [] + return sparse_tensor.SparseTensorValue(indices, values, shape) + + def dense_to_sparse_non_scalar(tensor): + indices = array_ops.where(array_ops.ones_like(tensor, dtype=dtypes.bool)) + values = array_ops.gather_nd(tensor, indices) + shape = array_ops.shape(tensor, out_type=dtypes.int64) + return sparse_tensor.SparseTensorValue(indices, values, shape) + + if is_scalar: + return dense_to_sparse_scalar + return dense_to_sparse_non_scalar + + def _structuredSparseDataset(self, structure, shape, dtype): + dense_to_sparse = self._make_dense_to_sparse_fn(len(shape) == 0) # pylint: disable=g-explicit-length-test + if structure is None: + return dataset_ops.Dataset.from_tensors( + dense_to_sparse(array_ops.zeros(shape, dtype=dtype))) + else: + return dataset_ops.Dataset.zip( + tuple([ + self._structuredSparseDataset(substructure, shape, dtype) + for substructure in structure + ])) + + def _structuredSparseElement(self, structure, shape, dtype): + dense_to_sparse = self._make_dense_to_sparse_fn(len(shape) == 0) # pylint: disable=g-explicit-length-test + if structure is None: + return dense_to_sparse(array_ops.zeros(shape, dtype=dtype)) + else: + return tuple([ + self._structuredSparseElement(substructure, shape, dtype) + for substructure in structure + ]) + + @parameterized.parameters( + (None, np.int32([]), dtypes.bool), + (None, np.int32([]), dtypes.int32), + (None, np.int32([]), dtypes.float32), + (None, np.int32([]), dtypes.string), + (None, np.int32([2]), dtypes.int32), + (None, np.int32([2, 2]), dtypes.int32), + ((None, None, None), np.int32([]), dtypes.int32), + ((None, (None, None)), np.int32([]), dtypes.int32), + ) + def testWindowDatasetBatchSparse(self, structure, shape, dtype): + """Tests batching of sparse tensor windows. + + Args: + structure: the input structure + shape: the input shape + dtype: the input data type + """ + + def fn(*args): + if len(args) == 1 and not isinstance(args[0], tuple): + return batching.batch_window(args[0]) + + return tuple([ + fn(*arg) if isinstance(arg, tuple) else batching.batch_window(arg) + for arg in args + ]) + + dataset = self._structuredSparseDataset( + structure, shape, dtype).repeat(5).apply( + grouping.window_dataset(5)).apply(grouping._map_x_dataset(fn)) + get_next = dataset.make_one_shot_iterator().get_next() + with self.test_session() as sess: + expected = sess.run( + self._structuredSparseElement(structure, + np.concatenate(([5], shape), axis=0), + dtype)) + actual = sess.run(get_next) + self._assertEqual(expected, actual) + + @parameterized.parameters( + (np.int32([]),), + (np.int32([1]),), + (np.int32([1, 2, 3]),), + ) + def testWindowDatasetBatchSparseDynamicShape(self, shape): + """Tests batching of dynamically shaped sparse tensor windows. + + Args: + shape: the input shape + """ + + shape_t = array_ops.placeholder(dtypes.int32) + dataset = dataset_ops.Dataset.from_tensors(array_ops.zeros(shape_t)).map( + self._make_dense_to_sparse_fn(len(shape) == 0)).repeat(5).apply( # pylint: disable=g-explicit-length-test + grouping.window_dataset(5)).apply( + grouping._map_x_dataset(batching.batch_window)) + iterator = dataset.make_initializable_iterator() + init_op = iterator.initializer + get_next = iterator.get_next() + with self.test_session() as sess: + sess.run(init_op, {shape_t: shape}) + expected = sess.run( + self._structuredSparseElement(None, + np.concatenate(([5], shape), axis=0), + dtypes.int32)) + actual = sess.run(get_next) + self._assertEqual(expected, actual) + + def _structuredRaggedDataset(self, structure, shapes, dtype): + + if structure is None: + return dataset_ops.Dataset.from_tensor_slices(shapes).map( + lambda shape: array_ops.zeros(shape, dtype=dtype)) + else: + return dataset_ops.Dataset.zip( + tuple([ + self._structuredRaggedDataset(substructure, shapes, dtype) + for substructure in structure + ])) + + @parameterized.parameters( + (None, np.int32([[1], [2], [3]]), dtypes.bool, [-1]), + (None, np.int32([[1], [2], [3]]), dtypes.int32, [-1]), + (None, np.int32([[1], [2], [3]]), dtypes.float32, [-1]), + (None, np.int32([[1], [2], [3]]), dtypes.string, [-1]), + (None, np.int32([[1, 3], [2, 2], [3, 1]]), dtypes.int32, [-1, -1]), + (None, np.int32([[3, 1, 3], [1, 3, 1]]), dtypes.int32, [-1, -1, -1]), + ((None, None, None), np.int32([[1], [2], [3]]), dtypes.int32, [-1]), + ((None, (None, None)), np.int32([[1], [2], [3]]), dtypes.int32, [-1]), + (None, np.int32([[1], [2], [3]]), dtypes.int32, [-1]), + (None, np.int32([[1], [2], [3]]), dtypes.int32, np.int32([10])), + ) + def testWindowDatasetPaddedBatchDense(self, structure, shapes, dtype, + padded_shape): + """Tests padded batching of dense tensor windows. + + Args: + structure: the input structure + shapes: the input shapes + dtype: the input data type + padded_shape: the shape to pad the output to + """ + + def fn(*args): + if len(args) == 1 and not isinstance(args[0], tuple): + return batching.padded_batch_window(args[0], padded_shape) + + return tuple([ + fn(*arg) if isinstance(arg, tuple) else batching.padded_batch_window( + arg, padded_shape) for arg in args + ]) + + dataset = self._structuredRaggedDataset(structure, shapes, dtype).apply( + grouping.window_dataset(len(shapes))).apply( + grouping._map_x_dataset(fn)) + get_next = dataset.make_one_shot_iterator().get_next() + with self.test_session() as sess: + expected_shape = np.maximum(np.amax(shapes, axis=0), padded_shape) + expected = sess.run( + self._structuredElement( + structure, + np.concatenate((np.int32([len(shapes)]), expected_shape)), dtype)) + actual = sess.run(get_next) + self._assertEqual(expected, actual) + + @parameterized.parameters( + (np.int32([[1], [2], [3]]), [-1]), + (np.int32([[1, 3], [2, 2], [3, 1]]), [-1, -1]), + (np.int32([[3, 1, 3], [1, 3, 1]]), [-1, -1, -1]), + ) + def testWindowDatasetPaddedBatchDenseDynamicShape(self, shapes, padded_shape): + """Tests padded batching of dynamically shaped dense tensor windows. + + Args: + shapes: the input shapes + padded_shape: the shape to pad the output to + """ + + shapes_t = array_ops.placeholder(dtypes.int32) + dataset = dataset_ops.Dataset.from_tensor_slices(shapes_t).map( + lambda shape: array_ops.zeros(shape, dtype=dtypes.int32)).apply( + grouping.window_dataset(len(shapes))).apply( + grouping._map_x_dataset( + lambda x: batching.padded_batch_window(x, padded_shape))) + iterator = dataset.make_initializable_iterator() + init_op = iterator.initializer + get_next = iterator.get_next() + with self.test_session() as sess: + sess.run(init_op, {shapes_t: shapes}) + expected_shape = np.maximum(np.amax(shapes, axis=0), padded_shape) + expected = sess.run( + self._structuredElement( + None, np.concatenate((np.int32([len(shapes)]), expected_shape)), + dtypes.int32)) + actual = sess.run(get_next) + self._assertEqual(expected, actual) + + @parameterized.parameters( + (np.int32([[1]]), np.int32([0])), + (np.int32([[10], [20]]), np.int32([15])), + ) + def testWindowDatasetPaddedBatchDenseInvalid(self, shapes, padded_shape): + """Tests invalid padded batching of dense tensor windows. + + Args: + shapes: the input shapes + padded_shape: the shape to pad the output to + """ + + dataset = dataset_ops.Dataset.from_tensor_slices(shapes).map( + lambda shape: array_ops.zeros(shape, dtype=dtypes.int32)).apply( + grouping.window_dataset(len(shapes))).apply( + grouping._map_x_dataset( + lambda x: batching.padded_batch_window(x, padded_shape))) + get_next = dataset.make_one_shot_iterator().get_next() + with self.test_session() as sess: + with self.assertRaises(errors.InvalidArgumentError): + sess.run(get_next) + + def _structuredRaggedSparseDataset(self, structure, shapes, dtype): + + def map_fn(shape): + dense_to_sparse = self._make_dense_to_sparse_fn(False) + return dense_to_sparse(array_ops.zeros(shape, dtype=dtype)) + + if structure is None: + return dataset_ops.Dataset.from_tensor_slices(shapes).map(map_fn) + else: + return dataset_ops.Dataset.zip( + tuple([ + self._structuredRaggedSparseDataset(substructure, shapes, dtype) + for substructure in structure + ])) + + def _structuredRaggedSparseElement(self, structure, shapes, dtype, + padded_shape): + if structure is None: + dense_shape = np.maximum(np.amax(shapes, axis=0), padded_shape) + values = [] + for shape in shapes: + dense_to_sparse = self._make_dense_to_sparse_fn(len(shape) == 0) # pylint: disable=g-explicit-length-test + sparse = dense_to_sparse(array_ops.zeros(shape, dtype=dtype)) + padded_sparse = sparse_tensor.SparseTensor(sparse.indices, + sparse.values, dense_shape) + reshaped_sparse = sparse_ops.sparse_reshape( + padded_sparse, + array_ops.concat([np.array([1], dtype=np.int64), dense_shape], 0)) + values.append(reshaped_sparse) + return sparse_ops.sparse_concat(0, values) + else: + return tuple([ + self._structuredRaggedSparseElement(substructure, shapes, dtype, + padded_shape) + for substructure in structure + ]) + + @parameterized.parameters( + (None, np.int64([[1], [2], [3]]), dtypes.bool, [-1]), + (None, np.int64([[1], [2], [3]]), dtypes.int32, [-1]), + (None, np.int64([[1], [2], [3]]), dtypes.float32, [-1]), + (None, np.int64([[1], [2], [3]]), dtypes.string, [-1]), + (None, np.int64([[1, 3], [2, 2], [3, 1]]), dtypes.int32, [-1, -1]), + (None, np.int64([[1, 3, 1], [3, 1, 3]]), dtypes.int32, [-1, -1, -1]), + ((None, None, None), np.int64([[1], [2], [3]]), dtypes.int32, [-1]), + ((None, (None, None)), np.int64([[1], [2], [3]]), dtypes.int32, [-1]), + (None, np.int64([[1], [2], [3]]), dtypes.int32, [-1]), + (None, np.int64([[1], [2], [3]]), dtypes.int32, np.int64([10])), + ) + def testWindowDatasetPaddedBatchSparse(self, structure, shapes, dtype, + padded_shape): + """Tests padded batching of sparse tensor windows. + + Args: + structure: the input structure + shapes: the input shapes + dtype: the input data type + padded_shape: the shape to pad the output to + """ + + def fn(*args): + if len(args) == 1 and not isinstance(args[0], tuple): + return batching.padded_batch_window(args[0], padded_shape) + + return tuple([ + fn(*arg) if isinstance(arg, tuple) else batching.padded_batch_window( + arg, padded_shape) for arg in args + ]) + + dataset = self._structuredRaggedSparseDataset( + structure, shapes, dtype).apply(grouping.window_dataset( + len(shapes))).apply(grouping._map_x_dataset(fn)) + get_next = dataset.make_one_shot_iterator().get_next() + with self.test_session() as sess: + expected = sess.run( + self._structuredRaggedSparseElement(structure, shapes, dtype, + padded_shape)) + actual = sess.run(get_next) + self._assertEqual(expected, actual) + + @parameterized.parameters( + (np.int64([[1], [2], [3]]), [-1]), + (np.int64([[1, 3], [2, 2], [3, 1]]), [-1, -1]), + (np.int64([[3, 1, 3], [1, 3, 1]]), [-1, -1, -1]), + ) + def testWindowDatasetPaddedBatchSparseDynamicShape(self, shapes, + padded_shape): + """Tests padded batching of dynamically shaped sparse tensor windows. + + Args: + shapes: the input shapes + padded_shape: the shape to pad the output to + """ + + shapes_t = array_ops.placeholder(dtypes.int32) + dataset = dataset_ops.Dataset.from_tensor_slices(shapes_t).map( + lambda shape: array_ops.zeros(shape, dtype=dtypes.int32)).map( + self._make_dense_to_sparse_fn(False) + ).apply(grouping.window_dataset(len(shapes))).apply( + grouping._map_x_dataset( + lambda x: batching.padded_batch_window(x, padded_shape))) + iterator = dataset.make_initializable_iterator() + init_op = iterator.initializer + get_next = iterator.get_next() + with self.test_session() as sess: + sess.run(init_op, {shapes_t: shapes}) + expected = sess.run( + self._structuredRaggedSparseElement(None, shapes, dtypes.int32, + padded_shape)) + actual = sess.run(get_next) + self._assertEqual(expected, actual) + + @parameterized.parameters( + (np.int64([[1]]), [0]), + (np.int64([[10], [20]]), [15]), + ) + def testWindowDatasetPaddedBatchSparseInvalid(self, shapes, padded_shape): + """Tests invalid padded batching of sparse tensor windows. + + Args: + shapes: the input shapes + padded_shape: the shape to pad the output to + """ + + dataset = dataset_ops.Dataset.from_tensor_slices(shapes).map( + lambda shape: array_ops.zeros(shape, dtype=dtypes.int32)).map( + self._make_dense_to_sparse_fn(False) + ).apply(grouping.window_dataset(len(shapes))).apply( + grouping._map_x_dataset( + lambda x: batching.padded_batch_window(x, padded_shape))) + get_next = dataset.make_one_shot_iterator().get_next() + with self.test_session() as sess: + with self.assertRaises(errors.InvalidArgumentError): + sess.run(get_next) + + +if __name__ == "__main__": + test.main() diff --git a/tensorflow/contrib/data/python/ops/BUILD b/tensorflow/contrib/data/python/ops/BUILD index 02408145625b7e751541e7b87dc4fd5da4f7cad9..160d7fe22a9f127f7ee23d7a988c22cc4430ce11 100644 --- a/tensorflow/contrib/data/python/ops/BUILD +++ b/tensorflow/contrib/data/python/ops/BUILD @@ -115,6 +115,8 @@ py_library( srcs = ["batching.py"], srcs_version = "PY2AND3", deps = [ + ":get_single_element", + ":grouping", "//tensorflow/contrib/framework:framework_py", "//tensorflow/python:array_ops", "//tensorflow/python:dataset_ops_gen", diff --git a/tensorflow/contrib/data/python/ops/batching.py b/tensorflow/contrib/data/python/ops/batching.py index 5708d47c2081976f82722018adf30523c091416a..a4914f4cde71925af477636c91d98b54ce0cce0e 100644 --- a/tensorflow/contrib/data/python/ops/batching.py +++ b/tensorflow/contrib/data/python/ops/batching.py @@ -17,22 +17,135 @@ from __future__ import absolute_import from __future__ import division from __future__ import print_function +import numpy as np + +from tensorflow.contrib.data.python.ops import get_single_element +from tensorflow.contrib.data.python.ops import grouping from tensorflow.contrib.framework import with_shape from tensorflow.python.data.ops import dataset_ops from tensorflow.python.data.util import convert from tensorflow.python.data.util import nest from tensorflow.python.data.util import sparse +from tensorflow.python.framework import constant_op from tensorflow.python.framework import dtypes from tensorflow.python.framework import ops from tensorflow.python.framework import sparse_tensor from tensorflow.python.framework import tensor_shape from tensorflow.python.framework import tensor_util from tensorflow.python.ops import array_ops +from tensorflow.python.ops import check_ops +from tensorflow.python.ops import control_flow_ops +from tensorflow.python.ops import gen_array_ops from tensorflow.python.ops import gen_dataset_ops from tensorflow.python.ops import math_ops +from tensorflow.python.ops import sparse_ops from tensorflow.python.util import deprecation +def batch_window(dataset): + """Batches a window of tensors. + + Args: + dataset: the input dataset. + + Returns: + A `Tensor` representing the batch of the entire input dataset. + """ + if isinstance(dataset.output_classes, tuple): + raise TypeError("Input dataset expected to have a single component") + if dataset.output_classes is ops.Tensor: + return _batch_dense_window(dataset) + elif dataset.output_classes is sparse_tensor.SparseTensor: + return _batch_sparse_window(dataset) + else: + raise TypeError("Unsupported dataset type: %s" % dataset.output_classes) + + +def _batch_dense_window(dataset): + """Batches a window of dense tensors.""" + + def key_fn(_): + return np.int64(0) + + def shape_init_fn(_): + return array_ops.shape(first_element) + + def shape_reduce_fn(state, value): + check_ops.assert_equal(state, array_ops.shape(value)) + return state + + def finalize_fn(state): + return state + + if dataset.output_shapes.is_fully_defined(): + shape = dataset.output_shapes + else: + first_element = get_single_element.get_single_element(dataset.take(1)) + shape_reducer = grouping.Reducer(shape_init_fn, shape_reduce_fn, + finalize_fn) + shape = get_single_element.get_single_element( + dataset.apply(grouping.group_by_reducer(key_fn, shape_reducer))) + + def batch_init_fn(_): + batch_shape = array_ops.concat([[0], shape], 0) + return gen_array_ops.empty(batch_shape, dtype=dataset.output_types) + + def batch_reduce_fn(state, value): + return array_ops.concat([state, [value]], 0) + + batch_reducer = grouping.Reducer(batch_init_fn, batch_reduce_fn, finalize_fn) + return get_single_element.get_single_element( + dataset.apply(grouping.group_by_reducer(key_fn, batch_reducer))) + + +def _batch_sparse_window(dataset): + """Batches a window of sparse tensors.""" + + def key_fn(_): + return np.int64(0) + + def shape_init_fn(_): + return first_element.dense_shape + + def shape_reduce_fn(state, value): + check_ops.assert_equal(state, value.dense_shape) + return state + + def finalize_fn(state): + return state + + if dataset.output_shapes.is_fully_defined(): + shape = dataset.output_shapes + else: + first_element = get_single_element.get_single_element(dataset.take(1)) + shape_reducer = grouping.Reducer(shape_init_fn, shape_reduce_fn, + finalize_fn) + shape = get_single_element.get_single_element( + dataset.apply(grouping.group_by_reducer(key_fn, shape_reducer))) + + def batch_init_fn(_): + indices_shape = array_ops.concat([[0], [array_ops.size(shape) + 1]], 0) + return sparse_tensor.SparseTensor( + indices=gen_array_ops.empty(indices_shape, dtype=dtypes.int64), + values=constant_op.constant([], shape=[0], dtype=dataset.output_types), + dense_shape=array_ops.concat( + [np.array([0], dtype=np.int64), + math_ops.cast(shape, dtypes.int64)], 0)) + + def batch_reduce_fn(state, value): + return sparse_ops.sparse_concat(0, [state, value]) + + def reshape_fn(value): + return sparse_ops.sparse_reshape( + value, + array_ops.concat([np.array([1], dtype=np.int64), value.dense_shape], 0)) + + batch_reducer = grouping.Reducer(batch_init_fn, batch_reduce_fn, finalize_fn) + return get_single_element.get_single_element( + dataset.map(reshape_fn).apply( + grouping.group_by_reducer(key_fn, batch_reducer))) + + def dense_to_sparse_batch(batch_size, row_shape): """A transformation that batches ragged elements into `tf.SparseTensor`s. @@ -82,6 +195,157 @@ def dense_to_sparse_batch(batch_size, row_shape): return _apply_fn +def padded_batch_window(dataset, padded_shape, padding_value=None): + """Batches a window of tensors with padding. + + Args: + dataset: the input dataset. + padded_shape: (Optional.) `tf.TensorShape` or `tf.int64` vector tensor-like + object representing the shape to which the input elements should be padded + prior to batching. Any unknown dimensions (e.g. `tf.Dimension(None)` in a + `tf.TensorShape` or `-1` in a tensor-like object) will be padded to the + maximum size of that dimension in each batch. + padding_value: (Optional.) A scalar-shaped `tf.Tensor`, representing the + padding value to use. Defaults are `0` for numeric types and the empty + string for string types. If `dataset` contains `tf.SparseTensor`, this + value is ignored. + + Returns: + A `Tensor` representing the batch of the entire input dataset. + + Raises: + ValueError: if invalid arguments are provided. + """ + if not issubclass(dataset.output_classes, + (ops.Tensor, sparse_tensor.SparseTensor)): + raise TypeError("Input dataset expected to have a single tensor component") + if issubclass(dataset.output_classes, (ops.Tensor)): + return _padded_batch_dense_window(dataset, padded_shape, padding_value) + elif issubclass(dataset.output_classes, (sparse_tensor.SparseTensor)): + if padding_value is not None: + raise ValueError("Padding value not allowed for sparse tensors") + return _padded_batch_sparse_window(dataset, padded_shape) + else: + raise TypeError("Unsupported dataset type: %s" % dataset.output_classes) + + +def _padded_batch_dense_window(dataset, padded_shape, padding_value=None): + """Batches a window of dense tensors with padding.""" + + padded_shape = math_ops.cast( + convert.partial_shape_to_tensor(padded_shape), dtypes.int32) + + def key_fn(_): + return np.int64(0) + + def max_init_fn(_): + return padded_shape + + def max_reduce_fn(state, value): + """Computes the maximum shape to pad to.""" + condition = math_ops.reduce_all( + math_ops.logical_or( + math_ops.less_equal(array_ops.shape(value), padded_shape), + math_ops.equal(padded_shape, -1))) + assert_op = control_flow_ops.Assert(condition, [ + "Actual shape greater than padded shape: ", + array_ops.shape(value), padded_shape + ]) + with ops.control_dependencies([assert_op]): + return math_ops.maximum(state, array_ops.shape(value)) + + def finalize_fn(state): + return state + + # Compute the padded shape. + max_reducer = grouping.Reducer(max_init_fn, max_reduce_fn, finalize_fn) + padded_shape = get_single_element.get_single_element( + dataset.apply(grouping.group_by_reducer(key_fn, max_reducer))) + + if padding_value is None: + if dataset.output_types == dtypes.string: + padding_value = "" + elif dataset.output_types == dtypes.bool: + padding_value = False + elif dataset.output_types == dtypes.variant: + raise TypeError("Unable to create padding for field of type 'variant'") + else: + padding_value = 0 + + def batch_init_fn(_): + return array_ops.fill( + array_ops.concat([np.array([0], dtype=np.int32), padded_shape], 0), + constant_op.constant(padding_value, dtype=dataset.output_types)) + + def batch_reduce_fn(state, value): + return array_ops.concat([state, [value]], 0) + + def pad_fn(value): + shape = array_ops.shape(value) + left = array_ops.zeros_like(shape) + right = padded_shape - shape + return array_ops.pad( + value, array_ops.stack([left, right], 1), constant_values=padding_value) + + batch_reducer = grouping.Reducer(batch_init_fn, batch_reduce_fn, finalize_fn) + return get_single_element.get_single_element( + dataset.map(pad_fn).apply( + grouping.group_by_reducer(key_fn, batch_reducer))) + + +def _padded_batch_sparse_window(dataset, padded_shape): + """Batches a window of sparse tensors with padding.""" + + def key_fn(_): + return np.int64(0) + + def max_init_fn(_): + return convert.partial_shape_to_tensor(padded_shape) + + def max_reduce_fn(state, value): + """Computes the maximum shape to pad to.""" + condition = math_ops.reduce_all( + math_ops.logical_or( + math_ops.less_equal(value.dense_shape, padded_shape), + math_ops.equal(padded_shape, -1))) + assert_op = control_flow_ops.Assert(condition, [ + "Actual shape greater than padded shape: ", value.dense_shape, + padded_shape + ]) + with ops.control_dependencies([assert_op]): + return math_ops.maximum(state, value.dense_shape) + + def finalize_fn(state): + return state + + # Compute the padded shape. + max_reducer = grouping.Reducer(max_init_fn, max_reduce_fn, finalize_fn) + padded_shape = get_single_element.get_single_element( + dataset.apply(grouping.group_by_reducer(key_fn, max_reducer))) + + def batch_init_fn(_): + indices_shape = array_ops.concat([[0], [array_ops.size(padded_shape) + 1]], + 0) + return sparse_tensor.SparseTensor( + indices=gen_array_ops.empty(indices_shape, dtype=dtypes.int64), + values=constant_op.constant([], shape=[0], dtype=dataset.output_types), + dense_shape=array_ops.concat( + [np.array([0], dtype=np.int64), padded_shape], 0)) + + def batch_reduce_fn(state, value): + padded_value = sparse_tensor.SparseTensor( + indices=value.indices, values=value.values, dense_shape=padded_shape) + reshaped_value = sparse_ops.sparse_reshape( + padded_value, + array_ops.concat( + [np.array([1], dtype=np.int64), padded_value.dense_shape], 0)) + return sparse_ops.sparse_concat(0, [state, reshaped_value]) + + reducer = grouping.Reducer(batch_init_fn, batch_reduce_fn, finalize_fn) + return get_single_element.get_single_element( + dataset.apply(grouping.group_by_reducer(key_fn, reducer))) + + class _UnbatchDataset(dataset_ops.Dataset): """A dataset that splits the elements of its input into multiple elements.""" @@ -175,7 +439,7 @@ def unbatch(): return _apply_fn -def filter_irregular_batches(batch_size): +def _filter_irregular_batches(batch_size): """Transformation that filters out batches that are not of size batch_size.""" def _apply_fn(dataset): @@ -254,7 +518,7 @@ def batch_and_drop_remainder(batch_size): # TODO(jsimsa): Switch to using `batch(..., drop_remainder=True)` any time # after 6/30/2018. batched = dataset.batch(batch_size) - return filter_irregular_batches(batch_size)(batched) + return _filter_irregular_batches(batch_size)(batched) return _apply_fn @@ -293,7 +557,7 @@ def padded_batch_and_drop_remainder(batch_size, # any time after 6/30/2018. batched = dataset.padded_batch( batch_size, padded_shapes=padded_shapes, padding_values=padding_values) - return filter_irregular_batches(batch_size)(batched) + return _filter_irregular_batches(batch_size)(batched) return _apply_fn diff --git a/tensorflow/contrib/data/python/ops/grouping.py b/tensorflow/contrib/data/python/ops/grouping.py index ca9540bf136a5028c4321319bdfacaf8a16484c7..bd8d398c58cc1825616c1ab5337cf6668c66697e 100644 --- a/tensorflow/contrib/data/python/ops/grouping.py +++ b/tensorflow/contrib/data/python/ops/grouping.py @@ -149,9 +149,9 @@ def bucket_by_sequence_length(element_length_func, @{tf.data.Dataset.padded_batch}. Defaults to padding with 0. pad_to_bucket_boundary: bool, if `False`, will pad dimensions with unknown size to maximum length in batch. If `True`, will pad dimensions with - unknown size to bucket boundary, and caller must ensure that the source - `Dataset` does not contain any elements with length longer than - `max(bucket_boundaries)`. + unknown size to bucket boundary minus 1 (i.e., the maximum length in each + bucket), and caller must ensure that the source `Dataset` does not contain + any elements with length longer than `max(bucket_boundaries)`. Returns: A `Dataset` transformation function, which can be passed to @@ -203,7 +203,7 @@ def bucket_by_sequence_length(element_length_func, none_filler = None if pad_to_bucket_boundary: err_msg = ("When pad_to_bucket_boundary=True, elements must have " - "length <= max(bucket_boundaries).") + "length < max(bucket_boundaries).") check = check_ops.assert_less( bucket_id, constant_op.constant(len(bucket_batch_sizes) - 1, @@ -213,7 +213,7 @@ def bucket_by_sequence_length(element_length_func, boundaries = constant_op.constant(bucket_boundaries, dtype=dtypes.int64) bucket_boundary = boundaries[bucket_id] - none_filler = bucket_boundary + none_filler = bucket_boundary - 1 shapes = make_padded_shapes( padded_shapes or grouped_dataset.output_shapes, none_filler=none_filler) @@ -227,6 +227,50 @@ def bucket_by_sequence_length(element_length_func, return _apply_fn +def _map_x_dataset(map_func): + """A transformation that maps `map_func` across its input. + + This transformation is similar to `tf.data.Dataset.map`, but in addition to + supporting dense and sparse tensor inputs, it also supports dataset inputs. + + Args: + map_func: A function mapping a nested structure of tensors and/or datasets + (having shapes and types defined by `self.output_shapes` and + `self.output_types`) to another nested structure of tensors and/or + datasets. + + Returns: + Dataset: A `Dataset`. + """ + + def _apply_fn(dataset): + """Function from `Dataset` to `Dataset` that applies the transformation.""" + return _MapXDataset(dataset, map_func) + + return _apply_fn + + +def window_dataset(window_size): + """A transformation that creates window datasets from the input dataset. + + The resulting datasets will contain `window_size` elements (or + `N % window_size` for the last dataset if `window_size` does not divide the + number of input elements `N` evenly). + + Args: + window_size: A `tf.int64` scalar `tf.Tensor`, representing the number of + consecutive elements of the input dataset to combine into a window. + + Returns: + Dataset: A `Dataset`. + """ + + def _apply_fn(dataset): + return _WindowDataset(dataset, window_size) + + return _apply_fn + + class _GroupByReducerDataset(dataset_ops.Dataset): """A `Dataset` that groups its input and performs a reduction.""" @@ -468,3 +512,85 @@ class Reducer(object): @property def finalize_func(self): return self._finalize_func + + +class _MapXDataset(dataset_ops.Dataset): + """A `Dataset` that maps a function over elements in its input.""" + + def __init__(self, input_dataset, map_func): + """See `map_x_dataset()` for details.""" + super(_MapXDataset, self).__init__() + self._input_dataset = input_dataset + + wrapped_func = dataset_ops.StructuredFunctionWrapper( + map_func, + "tf.contrib.data.map_x_dataset()", + input_dataset, + experimental_nested_dataset_support=True) + self._output_classes = wrapped_func.output_classes + self._output_shapes = wrapped_func.output_shapes + self._output_types = wrapped_func.output_types + self._map_func = wrapped_func.function + + def _as_variant_tensor(self): + input_t = self._input_dataset._as_variant_tensor() # pylint: disable=protected-access + return gen_dataset_ops.map_dataset( + input_t, + self._map_func.captured_inputs, + f=self._map_func, + **dataset_ops.flat_structure(self)) + + @property + def output_classes(self): + return self._output_classes + + @property + def output_shapes(self): + return self._output_shapes + + @property + def output_types(self): + return self._output_types + + +class _WindowDataset(dataset_ops.Dataset): + """A dataset that creates window datasets from the input elements.""" + + def __init__(self, input_dataset, window_size): + """See `window_dataset()` for more details.""" + super(_WindowDataset, self).__init__() + self._input_dataset = input_dataset + self._window_size = ops.convert_to_tensor( + window_size, dtype=dtypes.int64, name="window_size") + self._output_classes = nest.pack_sequence_as( + input_dataset.output_classes, + [ + dataset_ops._NestedDatasetComponent( # pylint: disable=protected-access + output_classes=output_class, + output_shapes=output_shape, + output_types=output_type) + for output_class, output_shape, output_type in zip( + nest.flatten(input_dataset.output_classes), + nest.flatten(input_dataset.output_shapes), + nest.flatten(input_dataset.output_types)) + ]) + self._output_shapes = self._output_classes + self._output_types = self._output_classes + + def _as_variant_tensor(self): + return gen_dataset_ops.window_dataset( + self._input_dataset._as_variant_tensor(), # pylint: disable=protected-access + self._window_size, + **dataset_ops.flat_structure(self)) + + @property + def output_classes(self): + return self._output_classes + + @property + def output_shapes(self): + return self._output_shapes + + @property + def output_types(self): + return self._output_types diff --git a/tensorflow/contrib/data/python/ops/prefetching_ops.py b/tensorflow/contrib/data/python/ops/prefetching_ops.py index e4c9f8b58a2a4390004b0ad318163526b443d44f..50212d3b523dda2d74523b83e910f1e8f2991cdd 100644 --- a/tensorflow/contrib/data/python/ops/prefetching_ops.py +++ b/tensorflow/contrib/data/python/ops/prefetching_ops.py @@ -26,21 +26,42 @@ from tensorflow.python.data.ops import iterator_ops from tensorflow.python.data.util import nest from tensorflow.python.data.util import sparse from tensorflow.python.eager import context +from tensorflow.python.framework import device as framework_device from tensorflow.python.framework import dtypes from tensorflow.python.framework import function from tensorflow.python.framework import ops +from tensorflow.python.ops import array_ops +from tensorflow.python.ops import functional_ops from tensorflow.python.ops import gen_dataset_ops as core_gen_dataset_ops +from tensorflow.python.ops import resource_variable_ops -# TODO(rohanj): Add a python class that constructs resource in the __init__ -# method and provides a get_next() that calls the prefetch op. def function_buffering_resource(string_arg, target_device, f, buffer_size, + output_types, container="", shared_name=None, name=None): + """Creates a FunctionBufferingResource. + + A FunctionBufferingResource fills up a buffer by calling a function `f` on + `target_device`. `f` should take in only a single string argument as input. + + Args: + string_arg: The single string argument to the function. + target_device: The device to run `f` on. + f: The function to be executed. + buffer_size: Size of the buffer to be populated. + output_types: The output types generated by the function. + container: (Optional) string. Defaults to "". + shared_name: (Optional) string. + name: (Optional) string to name the op. + + Returns: + Handle to a FunctionBufferingResource. + """ if shared_name is None: shared_name = "" return gen_dataset_ops.function_buffering_resource( @@ -50,7 +71,8 @@ def function_buffering_resource(string_arg, f=f, buffer_size=buffer_size, container=container, - name=name) + name=name, + output_types=output_types) def function_buffering_resource_get_next(function_buffer_resource, @@ -123,7 +145,10 @@ class _PrefetchToDeviceIterator(object): target_device=iterator_device, string_arg=input_iterator_handle, buffer_size=buffer_size, - shared_name=shared_name) + shared_name=shared_name, + output_types=nest.flatten( + sparse.as_dense_types(self._input_dataset.output_types, + self._input_dataset.output_classes))) if not self._one_shot: reset_op = function_buffering_resource_reset(self._buffering_resource) @@ -212,6 +237,7 @@ class _PrefetchToDeviceEagerIterator(iterator_ops.EagerIterator): with ops.device(device): self._buffering_resource = function_buffering_resource( f=_prefetch_fn, + output_types=self._flat_output_types, target_device=gen_dataset_ops.iterator_get_device(self._resource), string_arg=input_iterator_handle, buffer_size=buffer_size, @@ -323,3 +349,172 @@ def prefetch_to_device(device, buffer_size=None): return _PrefetchToDeviceDataset(dataset, device, buffer_size) return _apply_fn + + +def copy_to_device(target_device, source_device="/cpu:0"): + """A transformation that copies dataset elements to the given `target_device`. + + Args: + target_device: The name of a device to which elements will be copied. + source_device: The original device on which `input_dataset` will be placed. + + Returns: + A `Dataset` transformation function, which can be passed to + @{tf.data.Dataset.apply}. + """ + + def _apply_fn(dataset): + return _CopyToDeviceDataset( + dataset, target_device=target_device, source_device=source_device) + + return _apply_fn + + +# TODO(rohanj): Use the _input_hostmem attr on the RemoteCall ops to indicate +# all inputs to the Op are in host memory, thereby avoiding some unnecessary +# Sends and Recvs. +class _CopyToDeviceDataset(dataset_ops.Dataset): + """A `Dataset` that copies elements to another device.""" + + def __init__(self, input_dataset, target_device, source_device="/cpu:0"): + """Constructs a _CopyToDeviceDataset. + + Args: + input_dataset: `Dataset` to be copied + target_device: The name of the device to which elements would be copied. + source_device: Device where input_dataset would be placed. + """ + self._input_dataset = input_dataset + self._target_device = target_device + spec = framework_device.DeviceSpec().from_string(self._target_device) + self._is_gpu_target = (spec.device_type == "GPU") + self._source_device_string = source_device + self._source_device = ops.convert_to_tensor(source_device) + + self._flat_output_shapes = nest.flatten( + sparse.as_dense_shapes(self._input_dataset.output_shapes, + self._input_dataset.output_classes)) + self._flat_output_types = nest.flatten( + sparse.as_dense_types(self._input_dataset.output_types, + self._input_dataset.output_classes)) + + @function.Defun() + def _init_func(): + """Creates an iterator for the input dataset. + + Returns: + A `string` tensor that encapsulates the iterator created. + """ + # pylint: disable=protected-access + ds_variant = self._input_dataset._as_variant_tensor() + resource = core_gen_dataset_ops.anonymous_iterator( + output_types=self._flat_output_types, + output_shapes=self._flat_output_shapes) + with ops.control_dependencies( + [core_gen_dataset_ops.make_iterator(ds_variant, resource)]): + return core_gen_dataset_ops.iterator_to_string_handle(resource) + + @function.Defun() + def _remote_init_func(): + return functional_ops.remote_call( + target=self._source_device, + args=_init_func.captured_inputs, + Tout=[dtypes.string], + f=_init_func) + + self._init_func = _remote_init_func + self._init_captured_args = _remote_init_func.captured_inputs + + @function.Defun(dtypes.string) + def _next_func(string_handle): + """Calls get_next for created iterator. + + Args: + string_handle: An iterator string handle created by _init_func + Returns: + The elements generated from `input_dataset` + """ + with ops.device(self._source_device_string): + iterator = iterator_ops.Iterator.from_string_handle( + string_handle, self.output_types, self.output_shapes, + self.output_classes) + ret = iterator.get_next() + return nest.flatten(sparse.serialize_sparse_tensors(ret)) + + @function.Defun(dtypes.string) + def _remote_next_func(string_handle): + return functional_ops.remote_call( + target=self._source_device, + args=[string_handle] + _next_func.captured_inputs, + Tout=self._flat_output_types, + f=_next_func) + + self._next_func = _remote_next_func + self._next_captured_args = _remote_next_func.captured_inputs + + @function.Defun(dtypes.string) + def _finalize_func(string_handle): + """Destroys the iterator resource created. + + Args: + string_handle: An iterator string handle created by _init_func + Returns: + Tensor constant 0 + """ + iterator_resource = core_gen_dataset_ops.iterator_from_string_handle_v2( + string_handle, + output_types=self._flat_output_types, + output_shapes=self._flat_output_shapes) + with ops.control_dependencies([ + resource_variable_ops.destroy_resource_op( + iterator_resource, ignore_lookup_error=True)]): + return array_ops.constant(0, dtypes.int64) + + @function.Defun(dtypes.string) + def _remote_finalize_func(string_handle): + return functional_ops.remote_call( + target=self._source_device, + args=[string_handle] + _finalize_func.captured_inputs, + Tout=[dtypes.int64], + f=_finalize_func) + + self._finalize_func = _remote_finalize_func + self._finalize_captured_args = _remote_finalize_func.captured_inputs + # pylint: enable=protected-scope + + # The one_shot_iterator implementation needs a 0 arg _make_dataset function + # that thereby captures all the inputs required to create the dataset. Since + # there are strings that are inputs to the GeneratorDataset which can't be + # placed on a GPU, this fails for the GPU case. Therefore, disabling it for + # GPU + def make_one_shot_iterator(self): + if self._is_gpu_target: + raise ValueError("Cannot create a one shot iterator when using " + "`tf.contrib.data.copy_to_device()` on GPU. Please use " + "`Dataset.make_initializable_iterator()` instead.") + else: + return super(_CopyToDeviceDataset, self).make_one_shot_iterator() + + def _as_variant_tensor(self): + with ops.device(self._target_device): + return core_gen_dataset_ops.generator_dataset( + self._init_captured_args, + self._next_captured_args, + self._finalize_captured_args, + init_func=self._init_func, + next_func=self._next_func, + finalize_func=self._finalize_func, + output_types=self._flat_output_types, + output_shapes=self._flat_output_shapes) + + @property + def output_types(self): + return self._input_dataset.output_types + + @property + def output_shapes(self): + return self._input_dataset.output_shapes + + @property + def output_classes(self): + return self._input_dataset.output_classes diff --git a/tensorflow/contrib/data/python/ops/readers.py b/tensorflow/contrib/data/python/ops/readers.py index 83095c7ba1c6465d18490e5197f71bf7f1fe2497..9373e37f5fe79a482cbe098b6bdb9e42650e9b3b 100644 --- a/tensorflow/contrib/data/python/ops/readers.py +++ b/tensorflow/contrib/data/python/ops/readers.py @@ -540,11 +540,11 @@ class CsvDataset(dataset_ops.Dataset): The expected output of its iterations is: ```python - next = dataset.make_one_shot_iterator().get_next() + next_element = dataset.make_one_shot_iterator().get_next() with tf.Session() as sess: while True: try: - print(sess.run(nxt)) + print(sess.run(next_element)) except tf.errors.OutOfRangeError: break diff --git a/tensorflow/contrib/distribute/BUILD b/tensorflow/contrib/distribute/BUILD index 74b2cd90a187159fd2da8ce236c14e813cc43c49..1126f76f5854932bcb6a9550c100768069bbd1cc 100644 --- a/tensorflow/contrib/distribute/BUILD +++ b/tensorflow/contrib/distribute/BUILD @@ -30,6 +30,7 @@ py_library( "//tensorflow/contrib/distribute/python:monitor", "//tensorflow/contrib/distribute/python:one_device_strategy", "//tensorflow/contrib/distribute/python:step_fn", + "//tensorflow/contrib/distribute/python:tpu_strategy", "//tensorflow/python:training", "//tensorflow/python:util", ], diff --git a/tensorflow/contrib/distribute/__init__.py b/tensorflow/contrib/distribute/__init__.py index 76711baf3a11c8978fbb5770ec173ff74a153158..2e2c3be853cc5503c86121c142394d49e5037405 100644 --- a/tensorflow/contrib/distribute/__init__.py +++ b/tensorflow/contrib/distribute/__init__.py @@ -24,6 +24,7 @@ from tensorflow.contrib.distribute.python.mirrored_strategy import MirroredStrat from tensorflow.contrib.distribute.python.monitor import Monitor from tensorflow.contrib.distribute.python.one_device_strategy import OneDeviceStrategy from tensorflow.contrib.distribute.python.step_fn import * +from tensorflow.contrib.distribute.python.tpu_strategy import TPUStrategy from tensorflow.python.training.distribute import * from tensorflow.python.util.all_util import remove_undocumented @@ -41,6 +42,7 @@ _allowed_symbols = [ 'StandardInputStep', 'StandardSingleLossStep', 'TowerContext', + 'TPUStrategy', 'get_cross_tower_context', 'get_distribution_strategy', 'get_loss_reduction', diff --git a/tensorflow/contrib/distribute/python/BUILD b/tensorflow/contrib/distribute/python/BUILD index eba0dd0ea330e29db0ea8e68ee14767fcb8ddad0..40dbfa3dd2f6783536fff44d7c499d9a14e378df 100644 --- a/tensorflow/contrib/distribute/python/BUILD +++ b/tensorflow/contrib/distribute/python/BUILD @@ -587,6 +587,7 @@ cuda_py_test( ], tags = [ "multi_and_single_gpu", + "no_windows_gpu", "notsan", ], ) diff --git a/tensorflow/contrib/distribute/python/cross_tower_ops.py b/tensorflow/contrib/distribute/python/cross_tower_ops.py index 1009c3c0124c254ee2b69ccc161c9a108bfb855c..b0baf0dad1d55eafac5338d1eb43465927e428a1 100644 --- a/tensorflow/contrib/distribute/python/cross_tower_ops.py +++ b/tensorflow/contrib/distribute/python/cross_tower_ops.py @@ -28,11 +28,12 @@ from tensorflow.python.eager import context from tensorflow.python.framework import ops from tensorflow.python.ops import array_ops from tensorflow.python.ops import math_ops +from tensorflow.python.ops import variable_scope as vs from tensorflow.python.platform import tf_logging as logging from tensorflow.python.training import device_util -def _validate_destinations(destinations): +def validate_destinations(destinations): if not isinstance(destinations, (value_lib.DistributedValues, six.string_types, list)): raise ValueError("destinations must be one of a `DistributedValues` object," @@ -55,7 +56,7 @@ def _validate_value_destination_pairs(value_destination_pairs): # TODO(yuefengz): consider calling this function in the caller of CrossTowerOps. -def _get_devices_from(destinations): +def get_devices_from(destinations): if isinstance(destinations, value_lib.DistributedValues): return list(destinations.devices) elif isinstance(destinations, six.string_types): @@ -65,7 +66,7 @@ def _get_devices_from(destinations): def _devices_match(left, right): - return set(_get_devices_from(left)) == set(_get_devices_from(right)) + return set(get_devices_from(left)) == set(get_devices_from(right)) def _all_devices_match(value_destination_pairs): @@ -80,7 +81,7 @@ def _all_devices_match(value_destination_pairs): def _simple_broadcast(value, destinations): index = {} - devices = _get_devices_from(destinations) + devices = get_devices_from(destinations) for d in devices: index[d] = cross_tower_utils.copy_tensor_or_indexed_slices_to_device( value, d) @@ -88,7 +89,7 @@ def _simple_broadcast(value, destinations): def _simple_reduce(per_device_value, reduce_to_device, accumulation_fn, - method_string): + aggregation): # pylint: disable=g-missing-docstring all_values = [] count = 0 @@ -112,11 +113,12 @@ def _simple_reduce(per_device_value, reduce_to_device, accumulation_fn, with context.context().device_policy(context.DEVICE_PLACEMENT_SILENT): reduced = cross_tower_utils.aggregate_tensors_or_indexed_slices( all_values, accumulation_fn) - if method_string == "mean": + if aggregation == vs.VariableAggregation.MEAN: reduced = cross_tower_utils.divide_by_n_tensors_or_indexed_slices( reduced, count) - elif method_string != "sum": - raise ValueError("`method_string` must be 'sum' or 'mean'") + elif aggregation != vs.VariableAggregation.SUM: + raise ValueError("`aggregation` must be VariableAggregation.SUM " + "or VariableAggregation.MEAN.") return reduced @@ -126,14 +128,15 @@ class CrossTowerOps(object): def __init__(self): pass - def reduce(self, method_string, per_device_value, destinations=None): + def reduce(self, aggregation, per_device_value, destinations=None): """Reduce `per_device_value` to `destinations`. - It runs the reduction operation defined by `method_string` and put the + It runs the reduction operation defined by `aggregation` and put the result on `destinations`. Args: - method_string: either 'sum' or 'mean' specifying the reduction method. + aggregation: Indicates how a variable will be aggregated. Accepted values + are @{tf.VariableAggregation.SUM}, @{tf.VariableAggregation.MEAN}. per_device_value: a PerDevice object. destinations: the reduction destinations. @@ -146,17 +149,18 @@ class CrossTowerOps(object): if not isinstance(per_device_value, value_lib.PerDevice): raise ValueError("`per_device_value` must be a `PerDevice` object.") if destinations is not None: - _validate_destinations(destinations) - return self._reduce(method_string, per_device_value, destinations) + validate_destinations(destinations) + return self._reduce(aggregation, per_device_value, destinations) - def batch_reduce(self, method_string, value_destination_pairs): + def batch_reduce(self, aggregation, value_destination_pairs): """Reduce PerDevice objects in a batch. Reduce each first element in `value_destination_pairs` to each second element which indicates the destinations. Args: - method_string: either 'sum' or 'mean' specifying the reduction method. + aggregation: Indicates how a variable will be aggregated. Accepted values + are @{tf.VariableAggregation.SUM}, @{tf.VariableAggregation.MEAN}. value_destination_pairs: a list or a tuple of tuples of PerDevice objects and destinations. If a destination is None, then the destinations are set to match the devices of the input PerDevice object. @@ -173,9 +177,9 @@ class CrossTowerOps(object): "tuples of PerDevice objects and destinations") for _, d in value_destination_pairs: if d is not None: - _validate_destinations(d) + validate_destinations(d) - return self._batch_reduce(method_string, value_destination_pairs) + return self._batch_reduce(aggregation, value_destination_pairs) def broadcast(self, tensor, destinations): """Broadcast the `tensor` to destinations. @@ -187,14 +191,14 @@ class CrossTowerOps(object): Returns: a Mirrored object. """ - _validate_destinations(destinations) + validate_destinations(destinations) return self._broadcast(tensor, destinations) - def _reduce(self, method_string, per_device_value, destinations): + def _reduce(self, aggregation, per_device_value, destinations): raise NotImplementedError( "_reduce method must be implemented in descendants.") - def _batch_reduce(self, method_string, value_destination_pairs): + def _batch_reduce(self, aggregation, value_destination_pairs): raise NotImplementedError( "_batch_reduce method must be implemented in descendants.") @@ -220,16 +224,18 @@ class ReductionToOneDeviceCrossTowerOps(CrossTowerOps): self.accumulation_fn = accumulation_fn super(ReductionToOneDeviceCrossTowerOps, self).__init__() - def _reduce(self, method_string, per_device_value, destinations): - devices = _get_devices_from(destinations or per_device_value) + def _reduce(self, aggregation, per_device_value, destinations): + devices = get_devices_from(destinations or per_device_value) reduce_to_device = self.reduce_to_device or devices[0] reduced = _simple_reduce(per_device_value, reduce_to_device, - self.accumulation_fn, method_string) + self.accumulation_fn, aggregation) return self.broadcast(reduced, devices) - def _batch_reduce(self, method_string, value_destination_pairs): - return [self._reduce(method_string, t, destinations=v) - for t, v in value_destination_pairs] + def _batch_reduce(self, aggregation, value_destination_pairs): + return [ + self._reduce(aggregation, t, destinations=v) + for t, v in value_destination_pairs + ] def _group_value_by_device(per_device_values): @@ -260,18 +266,19 @@ def _group_value_by_device(per_device_values): return grouped -def _ungroup_and_make_mirrored(grouped_reduced, destinations, method_string): +def _ungroup_and_make_mirrored(grouped_reduced, destinations, aggregation): """Ungroup results from all-reduce and make Mirrored objects. Each all-reduce result will be divided by the number of destinations before - Mirrored objects are created if method_string is "mean". + Mirrored objects are created if aggregation is "mean". Args: grouped_reduced: a list of lists, each sublist has components for each device, paired with a None. It is the result from cross_tower_utils.aggregate_gradients_using*. destinations: a list of device strings for returned Mirrored objects. - method_string: "mean" or "sum". + aggregation: Indicates how a variable will be aggregated. Accepted values + are @{tf.VariableAggregation.SUM}, @{tf.VariableAggregation.MEAN}. Returns: a list of Mirrored objects. @@ -279,7 +286,7 @@ def _ungroup_and_make_mirrored(grouped_reduced, destinations, method_string): index = [{} for _ in range(len(grouped_reduced[0]))] for d, per_device_reduced in enumerate(grouped_reduced): for i, (v, _) in enumerate(per_device_reduced): - if method_string == "mean": + if aggregation == vs.VariableAggregation.MEAN: index[i][destinations[d]] = v / len(destinations) else: index[i][destinations[d]] = v @@ -488,32 +495,32 @@ class AllReduceCrossTowerOps(CrossTowerOps): self._agg_small_grads_max_group = agg_small_grads_max_group super(AllReduceCrossTowerOps, self).__init__() - def _reduce(self, method_string, per_device_value, destinations): + def _reduce(self, aggregation, per_device_value, destinations): contains_indexed_slices = cross_tower_utils.contains_indexed_slices( per_device_value) if ((destinations is None or _devices_match(per_device_value, destinations)) and not context.executing_eagerly() and not contains_indexed_slices): - return self._batch_all_reduce(method_string, [per_device_value])[0] + return self._batch_all_reduce(aggregation, [per_device_value])[0] else: if contains_indexed_slices: logging.log_first_n( logging.WARN, "Efficient allreduce is not supported for IndexedSlices.", 10) - devices = _get_devices_from(destinations or per_device_value) + devices = get_devices_from(destinations or per_device_value) reduce_to_device = devices[0] reduced = _simple_reduce(per_device_value, reduce_to_device, - math_ops.add_n, method_string) + math_ops.add_n, aggregation) return self.broadcast(reduced, devices) - def _batch_reduce(self, method_string, value_destination_pairs): + def _batch_reduce(self, aggregation, value_destination_pairs): all_devices_match = _all_devices_match(value_destination_pairs) contains_indexed_slices = cross_tower_utils.contains_indexed_slices( value_destination_pairs) if (all_devices_match and not context.executing_eagerly() and not contains_indexed_slices): - return self._batch_all_reduce(method_string, + return self._batch_all_reduce(aggregation, [v[0] for v in value_destination_pairs]) else: if not all_devices_match: @@ -521,11 +528,11 @@ class AllReduceCrossTowerOps(CrossTowerOps): "destinations are different.") return [ - self._reduce(method_string, t, destinations=v) + self._reduce(aggregation, t, destinations=v) for t, v in value_destination_pairs ] - def _batch_all_reduce(self, method_string, per_device_values): + def _batch_all_reduce(self, aggregation, per_device_values): """All reduce algorithm in a batch.""" logging.info( "batch_all_reduce invoked for batches size = %d with " @@ -556,7 +563,7 @@ class AllReduceCrossTowerOps(CrossTowerOps): reduced = _unpack_tensors(reduced, tensor_packer) return _ungroup_and_make_mirrored(reduced, per_device_values[0].devices, - method_string) + aggregation) AllReduceSpecTuple = collections.namedtuple("AllReduceSpecTuple", @@ -635,7 +642,7 @@ class MultiWorkerAllReduce(AllReduceCrossTowerOps): validate_and_complete_spec(spec) for spec in all_reduce_spec ] - def _batch_all_reduce(self, method_string, per_device_values): + def _batch_all_reduce(self, aggregation, per_device_values): """All reduce algorithm in a batch.""" logging.info( "distributed batch_all_reduce invoked for batches size = %d with " @@ -682,7 +689,7 @@ class MultiWorkerAllReduce(AllReduceCrossTowerOps): assert not remaining_grads return _ungroup_and_make_mirrored(aggregated_grads, destinations, - method_string) + aggregation) _dgx1_links = [[1, 2, 3, 4], [0, 2, 3, 5], [0, 1, 3, 6], [0, 1, 2, 7], diff --git a/tensorflow/contrib/distribute/python/cross_tower_ops_test.py b/tensorflow/contrib/distribute/python/cross_tower_ops_test.py index fed5505d92ef2544215069736c166a67d6141708..6a780ff60ffcd59d416278bfde6d005d7ad37a68 100644 --- a/tensorflow/contrib/distribute/python/cross_tower_ops_test.py +++ b/tensorflow/contrib/distribute/python/cross_tower_ops_test.py @@ -32,11 +32,12 @@ from tensorflow.python.framework import constant_op from tensorflow.python.framework import ops from tensorflow.python.ops import array_ops from tensorflow.python.ops import math_ops +from tensorflow.python.ops import variable_scope as vs from tensorflow.python.training import device_util def _make_per_device(values, devices): - devices = cross_tower_ops_lib._get_devices_from(devices) + devices = cross_tower_ops_lib.get_devices_from(devices) assert len(values) == len(devices) index = {} for d, v in zip(devices, values): @@ -53,7 +54,7 @@ def _fake_mirrored(value, devices): All components of the returned Mirrored have the same objects, which is not true in reality. """ - devices = cross_tower_ops_lib._get_devices_from(devices) + devices = cross_tower_ops_lib.get_devices_from(devices) return value_lib.Mirrored( {d: v for d, v in zip(devices, [value] * len(devices))}) @@ -93,7 +94,7 @@ class CrossTowerOpsTestBase(test.TestCase, parameterized.TestCase): self._assert_values_equal(l, r) else: self.assertEqual(type(left), type(right)) - self.assertEqual(left.devices, right.devices) + self.assertEqual(set(left.devices), set(right.devices)) if isinstance(list(left._index.values())[0], ops.IndexedSlices): for (d, v) in left._index.items(): self._assert_indexed_slices_equal(v, right._index[d]) @@ -129,32 +130,45 @@ class CrossTowerOpsTestBase(test.TestCase, parameterized.TestCase): # test reduce() for destinations in all_destinations: self._assert_values_equal( - cross_tower_ops.reduce("mean", per_device, destinations=destinations), + cross_tower_ops.reduce( + vs.VariableAggregation.MEAN, + per_device, + destinations=destinations), _fake_mirrored(mean, destinations or per_device)) self._assert_values_equal( cross_tower_ops.reduce( - "mean", per_device_2, destinations=destinations), + vs.VariableAggregation.MEAN, + per_device_2, + destinations=destinations), _fake_mirrored(mean_2, destinations or per_device)) self._assert_values_equal( - cross_tower_ops.reduce("sum", per_device, destinations=destinations), + cross_tower_ops.reduce( + vs.VariableAggregation.SUM, per_device, + destinations=destinations), _fake_mirrored(mean * len(devices), destinations or per_device)) self._assert_values_equal( cross_tower_ops.reduce( - "sum", per_device_2, destinations=destinations), + vs.VariableAggregation.SUM, + per_device_2, + destinations=destinations), _fake_mirrored(mean_2 * len(devices), destinations or per_device)) # test batch_reduce() for d1, d2 in itertools.product(all_destinations, all_destinations): self._assert_values_equal( - cross_tower_ops.batch_reduce( - "mean", [(per_device, d1), (per_device_2, d2)]), - [_fake_mirrored(mean, d1 or per_device), - _fake_mirrored(mean_2, d2 or per_device_2)]) + cross_tower_ops.batch_reduce(vs.VariableAggregation.MEAN, + [(per_device, d1), (per_device_2, d2)]), + [ + _fake_mirrored(mean, d1 or per_device), + _fake_mirrored(mean_2, d2 or per_device_2) + ]) self._assert_values_equal( - cross_tower_ops.batch_reduce( - "sum", [(per_device, d1), (per_device_2, d2)]), - [_fake_mirrored(mean * len(devices), d1 or per_device), - _fake_mirrored(mean_2 * len(devices), d2 or per_device_2)]) + cross_tower_ops.batch_reduce(vs.VariableAggregation.SUM, + [(per_device, d1), (per_device_2, d2)]), + [ + _fake_mirrored(mean * len(devices), d1 or per_device), + _fake_mirrored(mean_2 * len(devices), d2 or per_device_2) + ]) # test broadcast() for destinations in all_destinations: @@ -255,8 +269,8 @@ class SingleWorkerCrossTowerOpsTest(CrossTowerOpsTestBase): t0 = _make_indexed_slices([[1., 2.]], [1], [5, 2], devices[0]) t1 = _make_indexed_slices([[3., 4.], [5., 6.]], [1, 3], [5, 2], devices[1]) per_device = value_lib.PerDevice({devices[0]: t0, devices[1]: t1}) - result = cross_tower_ops_lib._simple_reduce(per_device, devices[0], - math_ops.add_n, "sum") + result = cross_tower_ops_lib._simple_reduce( + per_device, devices[0], math_ops.add_n, vs.VariableAggregation.SUM) # Test that the result is semantically equal to both the concatenated # IndexedSlices with and without duplicate indices. @@ -267,21 +281,22 @@ class SingleWorkerCrossTowerOpsTest(CrossTowerOpsTestBase): self._assert_indexed_slices_equal(total_with_dups, result) self._assert_indexed_slices_equal(total_without_dups, result) - @combinations.generate(combinations.combine( - cross_tower_ops_instance=[ - combinations.NamedObject( - "ReductionToOneDeviceCrossTowerOps", - cross_tower_ops_lib.ReductionToOneDeviceCrossTowerOps()), - combinations.NamedObject( - "AllReduceCrossTowerOps", - cross_tower_ops_lib.AllReduceCrossTowerOps()) - ], - method_string=["sum", "mean"], - batch_reduce=[True, False], - mode=["graph", "eager"], - required_gpus=1)) - def testIndexedSlicesAllReduce(self, cross_tower_ops_instance, - method_string, batch_reduce): + @combinations.generate( + combinations.combine( + cross_tower_ops_instance=[ + combinations.NamedObject( + "ReductionToOneDeviceCrossTowerOps", + cross_tower_ops_lib.ReductionToOneDeviceCrossTowerOps()), + combinations.NamedObject( + "AllReduceCrossTowerOps", + cross_tower_ops_lib.AllReduceCrossTowerOps()) + ], + aggregation=[vs.VariableAggregation.SUM, vs.VariableAggregation.MEAN], + batch_reduce=[True, False], + mode=["graph", "eager"], + required_gpus=1)) + def testIndexedSlicesAllReduce(self, cross_tower_ops_instance, aggregation, + batch_reduce): devices = ["/cpu:0", "/gpu:0"] dense_shape = [5, 2] t0 = _make_indexed_slices([[1., 2.]], [1], dense_shape, devices[0]) @@ -290,20 +305,19 @@ class SingleWorkerCrossTowerOpsTest(CrossTowerOpsTestBase): per_device = value_lib.PerDevice({devices[0]: t0, devices[1]: t1}) if batch_reduce: - result = cross_tower_ops_instance.batch_reduce(method_string, + result = cross_tower_ops_instance.batch_reduce(aggregation, [(per_device, devices)]) else: - result = cross_tower_ops_instance.reduce(method_string, per_device, - devices) + result = cross_tower_ops_instance.reduce(aggregation, per_device, devices) total_indices_with_dups = [1, 1, 3] total_indices_without_dups = [1, 3] - if method_string == "sum": + if aggregation == vs.VariableAggregation.SUM: total_values_with_dups = [[1., 2.], [3., 4.], [5., 6.]] total_values_without_dups = [[4., 6.], [5., 6.]] else: - assert method_string == "mean" + assert aggregation == vs.VariableAggregation.MEAN total_values_with_dups = [[0.5, 1.], [1.5, 2.], [2.5, 3.]] total_values_without_dups = [[2., 3.], [2.5, 3.]] diff --git a/tensorflow/contrib/distribute/python/mirrored_strategy.py b/tensorflow/contrib/distribute/python/mirrored_strategy.py index 98fea76b3d5cc4a634b1787cae89f879e2c2af01..dcbc6b0878b89cbb5b9779de315429e6f9478d15 100644 --- a/tensorflow/contrib/distribute/python/mirrored_strategy.py +++ b/tensorflow/contrib/distribute/python/mirrored_strategy.py @@ -104,9 +104,36 @@ class MirroredStrategy(distribute_lib.DistributionStrategy): colocate_with = kwargs.pop("colocate_with", None) devices = self._get_devices_from(colocate_with) - tower_local = kwargs.pop("tower_local_reduce_method", None) - if tower_local is not None: + # Get synchronization value + synchronization = kwargs.get( + "synchronization", variable_scope.VariableSynchronization.ON_WRITE) + if synchronization == variable_scope.VariableSynchronization.NONE: + raise ValueError("`NONE` variable synchronization mode is not " + "supported with `Mirrored` distribution strategy. Please" + " change the `synchronization` for variable: " + + kwargs["name"]) + elif synchronization == variable_scope.VariableSynchronization.ON_READ: + # Variables that are to be synced on read are tower local. + is_tower_local = True kwargs["trainable"] = False + elif (synchronization == variable_scope.VariableSynchronization.ON_WRITE or + synchronization == variable_scope.VariableSynchronization.AUTO): + # `AUTO` synchronization for `MirroredStrategy` is `ON_WRITE`. + is_tower_local = False + else: + raise ValueError("Invalid variable synchronization mode: " + + synchronization + " for variable: " + kwargs["name"]) + + # Get aggregation value + aggregation = kwargs.pop("aggregation", + variable_scope.VariableAggregation.NONE) + if aggregation not in [ + variable_scope.VariableAggregation.NONE, + variable_scope.VariableAggregation.SUM, + variable_scope.VariableAggregation.MEAN + ]: + raise ValueError("Invalid variable aggregation mode: " + aggregation + + " for variable: " + kwargs["name"]) # Ignore user-specified caching device, not needed for mirrored variables. kwargs.pop("caching_device", None) @@ -139,11 +166,11 @@ class MirroredStrategy(distribute_lib.DistributionStrategy): assert not isinstance(v, values.DistributedVariable) index[d] = v - if tower_local is None: - result = values.MirroredVariable(index, index[devices[0]]) + if is_tower_local: + result = values.TowerLocalVariable(index, index[devices[0]], + aggregation) else: - result = values.TowerLocalVariable( - index, index[devices[0]], tower_local) + result = values.MirroredVariable(index, index[devices[0]], aggregation) if not context.executing_eagerly(): g = ops.get_default_graph() @@ -308,16 +335,36 @@ class MirroredStrategy(distribute_lib.DistributionStrategy): cross_tower_ops_lib.ReductionToOneDeviceCrossTowerOps()) return self._cross_tower_ops - def _reduce(self, method_string, value, destinations): - if len(self._devices) == 1 and not isinstance(value, values.PerDevice): - value = values.PerDevice({self._devices[0]: value}) - assert isinstance(value, values.PerDevice) + def _reduce(self, aggregation, value, destinations): + assert not isinstance(value, values.Mirrored) + if not isinstance(value, values.PerDevice): + if value == 0: + return 0 + if aggregation == variable_scope.VariableAggregation.MEAN: + return self._broadcast(value, destinations) + + cross_tower_ops_lib.validate_destinations(destinations) + if len(self._devices) == 1: + if destinations: + # TODO(anjalisridhar): Moves these methods to a device utility file? + devices = cross_tower_ops_lib.get_devices_from(destinations) + if len(devices) == 1: + with ops.device(devices[0]): + return array_ops.identity(value) + else: + value_updates = {} + for d in devices: + with ops.device(d): + value_updates[d] = array_ops.identity(value) + return values.Mirrored(value_updates) + raise ValueError("A non PerDevice value cannot be reduced with the given " + "aggregation.") return self._get_cross_tower_ops().reduce( - method_string, value, destinations=destinations) + aggregation, value, destinations=destinations) - def _batch_reduce(self, method_string, value_destination_pairs): - return self._get_cross_tower_ops().batch_reduce(method_string, + def _batch_reduce(self, aggregation, value_destination_pairs): + return self._get_cross_tower_ops().batch_reduce(aggregation, value_destination_pairs) def _update(self, var, fn, *args, **kwargs): diff --git a/tensorflow/contrib/distribute/python/mirrored_strategy_multigpu_test.py b/tensorflow/contrib/distribute/python/mirrored_strategy_multigpu_test.py index 647cf953d73aea92dfe7662e1d2264c8505c75df..6a14b833d2485e1c672f6cb3e066c56cc19e699c 100644 --- a/tensorflow/contrib/distribute/python/mirrored_strategy_multigpu_test.py +++ b/tensorflow/contrib/distribute/python/mirrored_strategy_multigpu_test.py @@ -32,12 +32,14 @@ from tensorflow.python.framework import dtypes from tensorflow.python.framework import ops from tensorflow.python.framework import test_util from tensorflow.python.layers import core +from tensorflow.python.ops import math_ops from tensorflow.python.ops import rnn from tensorflow.python.ops import rnn_cell_impl from tensorflow.python.ops import variable_scope from tensorflow.python.ops import variables from tensorflow.python.training import distribute as distribute_lib + GPU_TEST = "test_gpu" in sys.argv[0] @@ -112,12 +114,35 @@ class MirroredTwoDeviceDistributionTest(strategy_test_lib.DistributionTestBase): dist = self._get_distribution_strategy() with dist.scope(): result = dist.call_for_each_tower(run_fn, dist.worker_device_index) - reduced = dist.reduce("sum", result, destinations="/device:CPU:0") + reduced = dist.reduce( + variable_scope.VariableAggregation.SUM, + result, + destinations="/device:CPU:0") unwrapped = dist.unwrap(reduced) self.assertEqual(1, len(unwrapped)) expected = sum(range(len(dist.worker_devices))) self.assertEqual(expected, self.evaluate(unwrapped[0])) + @test_util.run_in_graph_and_eager_modes() + def testReduceToMultipleDestinations(self): + if not GPU_TEST: + self.skipTest("Not GPU test") + + devices = ["/device:GPU:0"] + if GPU_TEST: + self.assertGreater(context.num_gpus(), 0) + print(self.id().split(".")[-1], "devices:", ", ".join(devices)) + + dist = mirrored_strategy.MirroredStrategy(devices) + with dist.scope(): + reduced = dist.reduce( + variable_scope.VariableAggregation.SUM, + 1.0, + destinations=["/device:CPU:0", "/device:GPU:0"]) + unwrapped = dist.unwrap(reduced) + self.assertEqual(2, len(unwrapped)) + self.assertEqual(1.0, self.evaluate(unwrapped[0])) + class MirroredStrategyVariableCreationTest(test.TestCase): @@ -263,19 +288,69 @@ class MirroredStrategyVariableCreationTest(test.TestCase): self.assertIsInstance(bias, values.MirroredVariable) self.assertEquals("common/dense" + suffix + "/bias:0", bias.name) + @test_util.run_in_graph_and_eager_modes(config=config) + def testWithVariableAndVariableScope(self): + self._skip_eager_if_gpus_less_than(1) + + def model_fn(): + v0 = variable_scope.variable(1.0, name="var0", aggregation=None) + with variable_scope.variable_scope("common"): + v1 = variable_scope.variable(1.0, name="var1") + # This will pause the current thread, and execute the other thread. + distribute_lib.get_tower_context().merge_call(lambda _: _) + v2 = variable_scope.variable( + 1.0, + name="var2", + synchronization=variable_scope.VariableSynchronization.ON_READ, + aggregation=variable_scope.VariableAggregation.SUM) + v3 = variable_scope.variable( + 1.0, + name="var3", + synchronization=variable_scope.VariableSynchronization.ON_WRITE, + aggregation=variable_scope.VariableAggregation.MEAN) + + return v0, v1, v2, v3 + + devices = ["/device:CPU:0", "/device:GPU:0"] + dist = mirrored_strategy.MirroredStrategy(devices) + with dist.scope(): + v = variable_scope.variable(1.0, name="var-main0") + self.assertEquals("var-main0:0", v.name) + + result = dist.call_for_each_tower(model_fn, run_concurrently=False) + self.assertEquals(4, len(result)) + v0, v1, v2, v3 = result + self.assertIsInstance(v0, values.MirroredVariable) + self.assertEquals("var0:0", v0.name) + self.assertIsInstance(v1, values.MirroredVariable) + self.assertEquals("common/var1:0", v1.name) + self.assertIsInstance(v2, values.TowerLocalVariable) + self.assertEquals("common/var2:0", v2.name) + self.assertEquals(variable_scope.VariableAggregation.SUM, v2.aggregation) + self.assertIsInstance(v3, values.MirroredVariable) + self.assertEquals("common/var3:0", v3.name) + self.assertEquals(variable_scope.VariableAggregation.MEAN, v3.aggregation) + @test_util.run_in_graph_and_eager_modes(config=config) def testWithGetVariableAndVariableScope(self): self._skip_eager_if_gpus_less_than(1) def model_fn(): - v0 = variable_scope.get_variable("var-thread0", [1]) + v0 = variable_scope.get_variable("var0", [1]) with variable_scope.variable_scope("common"): - v1 = variable_scope.get_variable("var-thread1", [1]) + v1 = variable_scope.get_variable("var1", [1]) # This will pause the current thread, and execute the other thread. distribute_lib.get_tower_context().merge_call(lambda _: _) - v2 = variable_scope.get_variable("var-thread2", [1]) + v2 = variable_scope.get_variable( + "var2", [1], + synchronization=variable_scope.VariableSynchronization.ON_READ, + aggregation=variable_scope.VariableAggregation.SUM) + v3 = variable_scope.get_variable( + "var3", [1], + synchronization=variable_scope.VariableSynchronization.ON_WRITE, + aggregation=variable_scope.VariableAggregation.MEAN) - return v0, v1, v2 + return v0, v1, v2, v3 devices = ["/device:CPU:0", "/device:GPU:0"] dist = mirrored_strategy.MirroredStrategy(devices) @@ -285,14 +360,89 @@ class MirroredStrategyVariableCreationTest(test.TestCase): self.assertEquals("main/var-main0:0", v.name) result = dist.call_for_each_tower(model_fn, run_concurrently=False) - self.assertEquals(3, len(result)) - v0, v1, v2 = result + self.assertEquals(4, len(result)) + v0, v1, v2, v3 = result self.assertIsInstance(v0, values.MirroredVariable) - self.assertEquals("main/var-thread0:0", v0.name) + self.assertEquals("main/var0:0", v0.name) self.assertIsInstance(v1, values.MirroredVariable) - self.assertEquals("main/common/var-thread1:0", v1.name) - self.assertIsInstance(v2, values.MirroredVariable) - self.assertEquals("main/common/var-thread2:0", v2.name) + self.assertEquals("main/common/var1:0", v1.name) + self.assertIsInstance(v2, values.TowerLocalVariable) + self.assertEquals("main/common/var2:0", v2.name) + self.assertEquals(variable_scope.VariableAggregation.SUM, + v2.aggregation) + self.assertIsInstance(v3, values.MirroredVariable) + self.assertEquals("main/common/var3:0", v3.name) + self.assertEquals(variable_scope.VariableAggregation.MEAN, + v3.aggregation) + + @test_util.run_in_graph_and_eager_modes(config=config) + def testNoneSynchronizationWithGetVariable(self): + self._skip_eager_if_gpus_less_than(1) + devices = ["/device:CPU:0", "/device:GPU:0"] + dist = mirrored_strategy.MirroredStrategy(devices) + with dist.scope(): + with self.assertRaisesRegexp( + ValueError, "`NONE` variable synchronization mode is not " + "supported with `Mirrored` distribution strategy. Please change " + "the `synchronization` for variable: v"): + variable_scope.get_variable( + "v", [1], + synchronization=variable_scope.VariableSynchronization.NONE) + + @test_util.run_in_graph_and_eager_modes(config=config) + def testNoneSynchronizationWithVariable(self): + self._skip_eager_if_gpus_less_than(1) + devices = ["/device:CPU:0", "/device:GPU:0"] + dist = mirrored_strategy.MirroredStrategy(devices) + with dist.scope(): + with self.assertRaisesRegexp( + ValueError, "`NONE` variable synchronization mode is not " + "supported with `Mirrored` distribution strategy. Please change " + "the `synchronization` for variable: v"): + variable_scope.variable( + 1.0, + name="v", + synchronization=variable_scope.VariableSynchronization.NONE) + + @test_util.run_in_graph_and_eager_modes(config=config) + def testInvalidSynchronizationWithVariable(self): + self._skip_eager_if_gpus_less_than(1) + devices = ["/device:CPU:0", "/device:GPU:0"] + dist = mirrored_strategy.MirroredStrategy(devices) + with dist.scope(): + with self.assertRaisesRegexp( + ValueError, "Invalid variable synchronization mode: Invalid for " + "variable: v"): + variable_scope.variable(1.0, name="v", synchronization="Invalid") + + @test_util.run_in_graph_and_eager_modes(config=config) + def testInvalidAggregationWithGetVariable(self): + self._skip_eager_if_gpus_less_than(1) + devices = ["/device:CPU:0", "/device:GPU:0"] + dist = mirrored_strategy.MirroredStrategy(devices) + with dist.scope(): + with self.assertRaisesRegexp( + ValueError, "Invalid variable aggregation mode: invalid for " + "variable: v"): + variable_scope.get_variable( + "v", [1], + synchronization=variable_scope.VariableSynchronization.ON_WRITE, + aggregation="invalid") + + @test_util.run_in_graph_and_eager_modes(config=config) + def testInvalidAggregationWithVariable(self): + self._skip_eager_if_gpus_less_than(1) + devices = ["/device:CPU:0", "/device:GPU:0"] + dist = mirrored_strategy.MirroredStrategy(devices) + with dist.scope(): + with self.assertRaisesRegexp( + ValueError, "Invalid variable aggregation mode: invalid for " + "variable: v"): + variable_scope.variable( + 1.0, + name="v", + synchronization=variable_scope.VariableSynchronization.ON_WRITE, + aggregation="invalid") @test_util.run_in_graph_and_eager_modes(config=config) def testThreeDevices(self): @@ -341,11 +491,14 @@ class MirroredStrategyVariableCreationTest(test.TestCase): components_mean = {} def model_fn(device_id): - tower_context = distribute_lib.get_tower_context() - with tower_context.tower_local_var_scope("sum"): - v_sum = variable_scope.variable(1.0) - with tower_context.tower_local_var_scope("mean"): - v_mean = variable_scope.variable(4.0) + v_sum = variable_scope.variable( + 1.0, + synchronization=variable_scope.VariableSynchronization.ON_READ, + aggregation=variable_scope.VariableAggregation.SUM) + v_mean = variable_scope.variable( + 4.0, + synchronization=variable_scope.VariableSynchronization.ON_READ, + aggregation=variable_scope.VariableAggregation.MEAN) self.assertTrue(isinstance(v_sum, values.TowerLocalVariable)) self.assertTrue(isinstance(v_mean, values.TowerLocalVariable)) updates = [v_sum.assign_add(2.0 + device_id), @@ -548,9 +701,10 @@ class MirroredStrategyVariableCreationTest(test.TestCase): with context.graph_mode(): def model_fn(): - tower_context = distribute_lib.get_tower_context() - with tower_context.tower_local_var_scope("sum"): - v_sum = variable_scope.variable(1.0) + v_sum = variable_scope.variable( + 1.0, + synchronization=variable_scope.VariableSynchronization.ON_READ, + aggregation=variable_scope.VariableAggregation.SUM) self.assertTrue(isinstance(v_sum, values.TowerLocalVariable)) return v_sum @@ -581,5 +735,237 @@ class MirroredStrategyVariableCreationTest(test.TestCase): self.assertEquals(10.0, self.evaluate(ret_v_sum)) +class MirroredVariableUpdateTest(test.TestCase): + # The following tests check assign, assign_add and assign_sub on Mirrored + # variables in tower and cross tower context. + config = config_pb2.ConfigProto() + config.allow_soft_placement = True + + def _skip_eager_if_gpus_less_than(self, num_gpus): + if context.num_gpus() < num_gpus and context.executing_eagerly(): + self.skipTest("Enough GPUs not available for this test in eager mode.") + + @test_util.run_in_graph_and_eager_modes(config=config) + def testAssignMirroredVarTowerContextWithoutAggregationType(self): + # Test that we always have an aggregation type set on the mirrored variable + # if we assign to it in tower mode. + self._skip_eager_if_gpus_less_than(1) + def var_fn(): + v = variable_scope.variable(1.0, name="foo") + return v + + dist = mirrored_strategy.MirroredStrategy( + ["/device:GPU:0", "/device:CPU:0"]) + + with dist.scope(): + mirrored_var = dist.call_for_each_tower(var_fn, run_concurrently=False) + self.assertIsInstance(mirrored_var, values.MirroredVariable) + self.evaluate(variables.global_variables_initializer()) + + def model_fn(): + return mirrored_var.assign(5.0) + + with self.assertRaisesRegexp( + ValueError, "You must specify an aggregation method to update a " + "MirroredVariable in Tower Context."): + self.evaluate(dist.unwrap(dist.call_for_each_tower(model_fn))) + + @test_util.run_in_graph_and_eager_modes(config=config) + def testAssignMirroredVarTowerContextWithSum(self): + # Test that we don't reduce a non-per-device value with the "sum" + # aggregation type. + self._skip_eager_if_gpus_less_than(1) + def var_fn(): + v = variable_scope.variable( + 1.0, name="foo", aggregation=variable_scope.VariableAggregation.SUM) + return v + + dist = mirrored_strategy.MirroredStrategy( + ["/device:GPU:0", "/device:CPU:0"]) + + with dist.scope(): + mirrored_var = dist.call_for_each_tower(var_fn, run_concurrently=False) + self.assertIsInstance(mirrored_var, values.MirroredVariable) + self.evaluate(variables.global_variables_initializer()) + + def model_fn(): + return mirrored_var.assign(5.0) + + with self.assertRaisesRegexp( + ValueError, "A non PerDevice value cannot be reduced with the given " + "aggregation."): + self.evaluate(dist.unwrap(dist.call_for_each_tower(model_fn))) + + @test_util.run_in_graph_and_eager_modes(config=config) + def testAssignMirroredVarCrossTowerContext(self): + self._skip_eager_if_gpus_less_than(1) + def var_fn(): + return variable_scope.variable(1.0, name="foo") + + dist = mirrored_strategy.MirroredStrategy( + ["/device:GPU:0", "/device:CPU:0"]) + + with dist.scope(): + mirrored_var = dist.call_for_each_tower(var_fn, run_concurrently=False) + self.assertIsInstance(mirrored_var, values.MirroredVariable) + self.evaluate(variables.global_variables_initializer()) + self.assertEquals(1.0, self.evaluate(mirrored_var)) + mirrored_var_result = self.evaluate(mirrored_var.assign(6.0)) + self.assertEquals(6.0, mirrored_var_result) + + @test_util.run_in_graph_and_eager_modes(config=config) + def testAssignMirroredVarTowerContext(self): + self._skip_eager_if_gpus_less_than(1) + def var_fn(): + return variable_scope.variable( + 1.0, name="foo", aggregation=variable_scope.VariableAggregation.MEAN) + + dist = mirrored_strategy.MirroredStrategy( + ["/device:GPU:0", "/device:CPU:0"]) + + with dist.scope(): + mirrored_var = dist.call_for_each_tower(var_fn, run_concurrently=False) + self.assertIsInstance(mirrored_var, values.MirroredVariable) + self.evaluate(variables.global_variables_initializer()) + self.assertEquals(1.0, self.evaluate(mirrored_var)) + + def model_fn(): + value = math_ops.cast(distribute_lib.get_tower_context().tower_id, + mirrored_var.dtype) + return mirrored_var.assign(value) + + self.evaluate(dist.unwrap(dist.call_for_each_tower( + model_fn, run_concurrently=False))) + self.assertEquals(0.5, self.evaluate(mirrored_var)) + + @test_util.run_in_graph_and_eager_modes(config=config) + def testAssignAddMirroredVarCrossTowerContext(self): + self._skip_eager_if_gpus_less_than(1) + def var_fn(): + return variable_scope.variable(1.0, name="foo") + + dist = mirrored_strategy.MirroredStrategy( + ["/device:GPU:0", "/device:CPU:0"]) + + with dist.scope(): + mirrored_var = dist.call_for_each_tower(var_fn, run_concurrently=False) + self.assertIsInstance(mirrored_var, values.MirroredVariable) + self.evaluate(variables.global_variables_initializer()) + self.assertEquals(1.0, self.evaluate(mirrored_var)) + mirrored_var_result = self.evaluate(mirrored_var.assign_add(6.0)) + self.assertEquals(7.0, mirrored_var_result) + + @test_util.run_in_graph_and_eager_modes(config=config) + def testAssignAddMirroredVarTowerContext(self): + self._skip_eager_if_gpus_less_than(1) + def var_fn(): + return variable_scope.variable( + 1.0, name="foo", aggregation=variable_scope.VariableAggregation.MEAN) + + dist = mirrored_strategy.MirroredStrategy( + ["/device:GPU:0", "/device:CPU:0"]) + + with dist.scope(): + mirrored_var = dist.call_for_each_tower(var_fn, run_concurrently=False) + self.assertIsInstance(mirrored_var, values.MirroredVariable) + self.evaluate(variables.global_variables_initializer()) + self.assertEquals(1.0, self.evaluate(mirrored_var)) + + def model_fn(): + value = math_ops.cast(distribute_lib.get_tower_context().tower_id, + mirrored_var.dtype) + return mirrored_var.assign_add(value) + + self.evaluate(dist.unwrap(dist.call_for_each_tower( + model_fn, run_concurrently=False))) + self.assertEquals(1.5, self.evaluate(mirrored_var)) + + @test_util.run_in_graph_and_eager_modes(config=config) + def testAssignSubMirroredVarCrossTowerContext(self): + self._skip_eager_if_gpus_less_than(1) + def var_fn(): + return variable_scope.variable(5.0, name="foo") + + dist = mirrored_strategy.MirroredStrategy( + ["/device:GPU:0", "/device:CPU:0"]) + + with dist.scope(): + mirrored_var = dist.call_for_each_tower(var_fn, run_concurrently=False) + self.assertIsInstance(mirrored_var, values.MirroredVariable) + self.evaluate(variables.global_variables_initializer()) + self.assertEquals(5.0, self.evaluate(mirrored_var)) + mirrored_var_result = self.evaluate(mirrored_var.assign_sub(2.0)) + self.assertEquals(3.0, mirrored_var_result) + + @test_util.run_in_graph_and_eager_modes(config=config) + def testAssignSubMirroredVarTowerContext(self): + self._skip_eager_if_gpus_less_than(1) + def var_fn(): + return variable_scope.variable( + 5.0, name="foo", aggregation=variable_scope.VariableAggregation.MEAN) + + dist = mirrored_strategy.MirroredStrategy( + ["/device:GPU:0", "/device:CPU:0"]) + + with dist.scope(): + mirrored_var = dist.call_for_each_tower(var_fn, run_concurrently=False) + self.assertIsInstance(mirrored_var, values.MirroredVariable) + self.evaluate(variables.global_variables_initializer()) + self.assertEquals(5.0, self.evaluate(mirrored_var)) + + def model_fn(): + value = math_ops.cast(distribute_lib.get_tower_context().tower_id, + mirrored_var.dtype) + return mirrored_var.assign_sub(value) + + self.evaluate(dist.unwrap(dist.call_for_each_tower( + model_fn, run_concurrently=False))) + self.assertEquals(4.5, self.evaluate(mirrored_var)) + + +class MirroredAndTowerLocalVariableInitializerTest(test.TestCase): + config = config_pb2.ConfigProto() + config.allow_soft_placement = True + + def testAssignMirroredVarInitializer(self): + # This test is not eager compatible since in eager variables are initialized + # upon construction instead of once the initialization op is run. + with context.graph_mode(): + def var_fn(): + v = variable_scope.variable(1.0, name="foo") + return v + + dist = mirrored_strategy.MirroredStrategy( + ["/device:GPU:0", "/device:CPU:0"]) + + with dist.scope(): + mirrored_var = dist.call_for_each_tower(var_fn) + self.assertIsInstance(mirrored_var, values.MirroredVariable) + self.assertFalse(self.evaluate(mirrored_var.is_initialized())) + self.evaluate(mirrored_var.initializer) + self.assertTrue(self.evaluate(mirrored_var.is_initialized())) + + def testAssignTowerLocalVarInitializer(self): + # This test is not eager compatible since in eager variables are initialized + # upon construction instead of once the initialization op is run. + with context.graph_mode(): + def model_fn(): + v_sum = variable_scope.variable( + 1.0, + synchronization=variable_scope.VariableSynchronization.ON_READ, + aggregation=variable_scope.VariableAggregation.SUM) + self.assertTrue(isinstance(v_sum, values.TowerLocalVariable)) + return v_sum + + dist = mirrored_strategy.MirroredStrategy( + ["/device:GPU:0", "/device:CPU:0"]) + + with dist.scope(): + tower_local_var = dist.call_for_each_tower(model_fn) + self.assertTrue(isinstance(tower_local_var, values.TowerLocalVariable)) + self.assertFalse(self.evaluate(tower_local_var.is_initialized())) + self.evaluate(tower_local_var.initializer) + self.assertTrue(self.evaluate(tower_local_var.is_initialized())) + if __name__ == "__main__": test.main() diff --git a/tensorflow/contrib/distribute/python/multi_worker_strategy.py b/tensorflow/contrib/distribute/python/multi_worker_strategy.py index 0f21a427320510635279f80c11711e81715ec37c..cbfe5df61d1ee6fa1eb9275b715b0721d678a46f 100644 --- a/tensorflow/contrib/distribute/python/multi_worker_strategy.py +++ b/tensorflow/contrib/distribute/python/multi_worker_strategy.py @@ -46,7 +46,7 @@ class MultiWorkerMirroredStrategy(MirroredStrategy): * **In-graph replication**: the `client` creates a single `tf.Graph` that specifies tasks for devices on all workers. The `client` then creates a client session which will talk to the `master` service of a `worker`. Then - the `master` will parition the graph and distribute the work to all + the `master` will partition the graph and distribute the work to all participating workers. * **Worker**: A `worker` is a TensorFlow `task` that usually maps to one physical machine. We will have multiple `worker`s with different `task` diff --git a/tensorflow/contrib/distribute/python/one_device_strategy.py b/tensorflow/contrib/distribute/python/one_device_strategy.py index a580dac96c5e6c6c8790aa6af7309988bf7a6477..dbd3514aec7d40d9a04dba4bcbc5c14be639aa33 100644 --- a/tensorflow/contrib/distribute/python/one_device_strategy.py +++ b/tensorflow/contrib/distribute/python/one_device_strategy.py @@ -24,6 +24,7 @@ from tensorflow.contrib.distribute.python import values from tensorflow.python.framework import ops from tensorflow.python.ops import array_ops from tensorflow.python.ops import math_ops +from tensorflow.python.ops import variable_scope as vs from tensorflow.python.training import distribute as distribute_lib @@ -43,11 +44,6 @@ class OneDeviceStrategy(distribute_lib.DistributionStrategy): self._default_device = device def _create_variable(self, next_creator, *args, **kwargs): - # No need to distinguish tower-local variables when not mirroring, - # we just enforce that they are not trainable. - if kwargs.pop("tower_local_reduce_method", None) is not None: - kwargs["trainable"] = False - colocate_with = kwargs.pop("colocate_with", None) if colocate_with is None: with ops.device(self._device): @@ -80,15 +76,15 @@ class OneDeviceStrategy(distribute_lib.DistributionStrategy): with ops.device(self._device): return values.MapOutput([fn(m, *args, **kwargs) for m in map_over]) - def _reduce(self, method_string, value, destinations): + def _reduce(self, aggregation, value, destinations): if not isinstance(value, values.MapOutput): return value l = value.get() assert l with ops.device(self._device): - if method_string == "sum": + if aggregation == vs.VariableAggregation.SUM: return math_ops.add_n(l) - elif method_string == "mean": + elif aggregation == vs.VariableAggregation.MEAN: return math_ops.add_n(l) / len(l) else: assert False diff --git a/tensorflow/contrib/distribute/python/prefetching_ops_v2.py b/tensorflow/contrib/distribute/python/prefetching_ops_v2.py index 7b3670b45aba801cf8c18e04bfea03e23eb67184..24cdc627a35f4455cb92484566dc13fa1bbaf2cc 100644 --- a/tensorflow/contrib/distribute/python/prefetching_ops_v2.py +++ b/tensorflow/contrib/distribute/python/prefetching_ops_v2.py @@ -89,6 +89,9 @@ class _PrefetchToDeviceIterator(object): with ops.device(device): buffer_resource_handle = prefetching_ops.function_buffering_resource( f=_prefetch_fn, + output_types=data_nest.flatten( + sparse.as_dense_types(self._input_dataset.output_types, + self._input_dataset.output_classes)), target_device=target_device, string_arg=input_iterator_handle, buffer_size=buffer_size, diff --git a/tensorflow/contrib/distribute/python/strategy_test_lib.py b/tensorflow/contrib/distribute/python/strategy_test_lib.py index d2fe8b3b1efabf7b35c070a82d01595f3fa51bf9..baed0ebaae8a3f41c55f309d28203b363336dd16 100644 --- a/tensorflow/contrib/distribute/python/strategy_test_lib.py +++ b/tensorflow/contrib/distribute/python/strategy_test_lib.py @@ -26,6 +26,7 @@ from tensorflow.python.framework import constant_op from tensorflow.python.framework import ops from tensorflow.python.layers import core from tensorflow.python.ops import array_ops +from tensorflow.python.ops import variable_scope from tensorflow.python.ops import variables from tensorflow.python.training import distribute as distribute_lib from tensorflow.python.training import optimizer @@ -110,7 +111,8 @@ class DistributionTestBase(test.TestCase): before_list.append(fetched) # control_dependencies irrelevant but harmless in eager execution with ops.control_dependencies([fetched]): - g = d.reduce("sum", g, destinations=v) + g = d.reduce( + variable_scope.VariableAggregation.SUM, g, destinations=v) with ops.control_dependencies(d.unwrap(d.update(v, update, g))): after_list.append(d.read_var(v)) return before_list, after_list @@ -162,7 +164,8 @@ class DistributionTestBase(test.TestCase): fetched = d.read_var(v) before_list.append(fetched) with ops.control_dependencies([fetched]): - g = d.reduce("sum", g, destinations=v) + g = d.reduce( + variable_scope.VariableAggregation.SUM, g, destinations=v) with ops.control_dependencies(d.unwrap(d.update(v, update, g))): after_list.append(d.read_var(v)) return before_list, after_list @@ -184,7 +187,7 @@ class DistributionTestBase(test.TestCase): with d.scope(): map_in = [constant_op.constant(i) for i in range(10)] map_out = d.map(map_in, lambda x, y: x * y, 2) - observed = d.reduce("sum", map_out) + observed = d.reduce(variable_scope.VariableAggregation.SUM, map_out) expected = 90 # 2 * (0 + 1 + ... + 9) self.assertEqual(expected, observed.numpy()) diff --git a/tensorflow/contrib/distribute/python/tpu_strategy.py b/tensorflow/contrib/distribute/python/tpu_strategy.py index b177e09adbc89684ff885d2903a30cc3696a2140..bc53898539d76320e331784f9a717be9491365e1 100644 --- a/tensorflow/contrib/distribute/python/tpu_strategy.py +++ b/tensorflow/contrib/distribute/python/tpu_strategy.py @@ -23,10 +23,13 @@ from __future__ import print_function from tensorflow.contrib import tpu from tensorflow.contrib.distribute.python import one_device_strategy +from tensorflow.contrib.distribute.python import values from tensorflow.contrib.tpu.python.ops import tpu_ops from tensorflow.python.framework import constant_op from tensorflow.python.framework import ops +from tensorflow.python.ops import array_ops from tensorflow.python.ops import control_flow_ops +from tensorflow.python.ops import variable_scope as vs from tensorflow.python.util import nest @@ -47,7 +50,10 @@ class TPUStrategy(one_device_strategy.OneDeviceStrategy): return self._call_dataset_fn(dataset_fn) # TODO(priyag): Deal with OutOfRange errors. - def run_steps_on_dataset(self, fn, iterator, iterations): + # TODO(sourabhbajaj): Remove the initial_loop_values parameter when we have + # a mechanism to infer the outputs of `fn`. Pending b/110550782. + def _run_steps_on_dataset(self, fn, iterator, iterations, + initial_loop_values=None): # Enqueue ops shapes = nest.flatten(iterator.output_shapes) if any([not s.is_fully_defined() for s in shapes]): @@ -93,26 +99,48 @@ class TPUStrategy(one_device_strategy.OneDeviceStrategy): return nest.pack_sequence_as(iterator.output_shapes, dequeued) # Wrap `fn` for repeat. - run_fn = lambda: fn(dequeue_fn()) + if initial_loop_values is None: + initial_loop_values = [] + ctx = values.MultiStepContext(initial_loop_values) + def run_fn(*args, **kwargs): + del args, kwargs + fn_result = fn(ctx, dequeue_fn()) + if ctx.last_step_outputs is None: + ctx.last_step_outputs = [] + with ops.control_dependencies([fn_result]): + return array_ops.identity(ctx.last_step_outputs) # Repeat + # TODO(sourabhbajaj): The input to while loop should be based on the output + # type of the step_fn def iterate_on_tpu(): - return tpu.repeat(iterations, run_fn, []) + return tpu.repeat(iterations, run_fn, [initial_loop_values]) # Re-write and distribute computation. - tpu_result = tpu.batch_parallel( + # TODO(sourabhbajaj): Convert the output to PerDevice variable and + # implement support for that in reduce. + last_step_tensor_outputs = tpu.batch_parallel( iterate_on_tpu, [], num_shards=self._num_cores_per_host) - return control_flow_ops.group(tpu_result, enqueue_ops) + # Take index [0] of last_step_tensor_outputs as we wrapped + # initial_loop_values in a list in the `repeat` call. + return (control_flow_ops.group(last_step_tensor_outputs, enqueue_ops), + last_step_tensor_outputs[0], ctx) def _call_for_each_tower(self, fn, *args, **kwargs): kwargs.pop('run_concurrently', None) with one_device_strategy._OneDeviceTowerContext(self): # pylint: disable=protected-access return fn(*args, **kwargs) - def _reduce(self, method_string, value, destinations): + def get_initialization_ops(self): + return [tpu.initialize_system()] + + def get_finalize_ops(self): + return [tpu.shutdown_system()] + + def _reduce(self, aggregation, value, destinations): del destinations # TPU is graph mode only. Rely on implicit Send/Recv. - if method_string == 'mean': + if aggregation == vs.VariableAggregation.MEAN: # TODO(jhseu): Revisit once we support model-parallelism. value *= (1. / self._num_cores_per_host) return tpu_ops.cross_replica_sum(value) diff --git a/tensorflow/contrib/distribute/python/values.py b/tensorflow/contrib/distribute/python/values.py index 9a48928a9530c3b24b603acf8b0b7584f5294b9e..1b5e00bc793242faea65e0be23cebeb0deb55525 100644 --- a/tensorflow/contrib/distribute/python/values.py +++ b/tensorflow/contrib/distribute/python/values.py @@ -23,7 +23,6 @@ from __future__ import print_function import collections import weakref - import six from tensorflow.contrib.distribute.python import input_ops @@ -34,6 +33,8 @@ from tensorflow.python.framework import ops from tensorflow.python.ops import array_ops from tensorflow.python.ops import control_flow_ops from tensorflow.python.ops import math_ops +from tensorflow.python.ops import state_ops +from tensorflow.python.ops import variable_scope as vs from tensorflow.python.training import device_util from tensorflow.python.training import distribute as distribute_lib from tensorflow.python.training import saver @@ -251,21 +252,6 @@ class DistributedVariable(DistributedDelegate): ops.register_dense_tensor_like_type(DistributedVariable) -class _MirroredSaveable(saver.BaseSaverBuilder.ResourceVariableSaveable): - """Class for defining how to restore a MirroredVariable.""" - - def __init__(self, mirrored_variable, primary_variable, name): - self._mirrored_variable = mirrored_variable - super(_MirroredSaveable, self).__init__(primary_variable, "", name) - - def restore(self, restored_tensors, restored_shapes): - """Restore the same value into all variables.""" - tensor, = restored_tensors - return control_flow_ops.group([ - _assign_on_device(d, v, tensor) - for d, v in six.iteritems(self._mirrored_variable._index)]) # pylint: disable=protected-access - - def _get_update_device(): """Validate we are in update/update_non_slot() and return current device. @@ -286,30 +272,113 @@ def _get_update_device(): return device +class _MirroredSaveable(saver.BaseSaverBuilder.ResourceVariableSaveable): + """Class for defining how to restore a MirroredVariable.""" + + def __init__(self, mirrored_variable, primary_variable, name): + self._mirrored_variable = mirrored_variable + super(_MirroredSaveable, self).__init__(primary_variable, "", name) + + def restore(self, restored_tensors, restored_shapes): + """Restore the same value into all variables.""" + tensor, = restored_tensors + return control_flow_ops.group([ + _assign_on_device(d, v, tensor) + for d, v in six.iteritems(self._mirrored_variable._index)]) # pylint: disable=protected-access + + class MirroredVariable(DistributedVariable, Mirrored, checkpointable.CheckpointableBase): """Holds a map from device to variables whose values are kept in sync.""" - def __init__(self, index, primary_var): + def __init__(self, index, primary_var, aggregation): + # Use a weakref to make it easy to map from the contained values + # to the container without introducing a reference cycle. + for v in six.itervalues(index): + v._mirrored_container = weakref.ref(self) # pylint: disable=protected-access self._primary_var = primary_var + # tf.keras keeps track of variables initialized using this attribute. When + # tf.keras gets the default session, it initializes all uninitialized vars. + # We need to make _keras_initialized a member of MirroredVariable because + # without this it will use `__getattr__` which will delegate to a component + # variable. + self._keras_initialized = False + self._aggregation = aggregation super(MirroredVariable, self).__init__(index) - # We use _get_update_device() for the assign* methods to enforce - # that we are in an update() function. The arguments to update() are - # automatically unwrapped so the update() function would normally - # see regular variables, not MirroredVariables. However, the update - # function can still operate on wrapped MirroredVariables through - # object members, captured arguments, etc. This is more likely in an + # The arguments to update() are automatically unwrapped so the update() + # function would normally see regular variables, not MirroredVariables. + # However, the update function can still operate on wrapped MirroredVariables + # through object members, captured arguments, etc. This is more likely in an # update_non_slot() function (like OptimizerV2._finish), which can # update several non-slot variables in one call. + def _assign_func(self, *args, **kwargs): + f = kwargs.pop("f") + if distribute_lib.get_cross_tower_context(): + update_device = distribute_lib.get_update_device() + # We are calling update on the mirrored variable in cross tower context. + if update_device is not None: + # We are calling an assign function on the mirrored variable in cross + # tower context. + v = self.get(device=update_device) + return f(v, *args, **kwargs) + + return distribute_lib.get_distribution_strategy().update( + self, f, *args, **kwargs) + else: + # We are calling an assign function on the mirrored variable in tower + # context. + # We reduce the value we want to assign/add/sub. More details about how we + # handle the different use cases can be found in the _reduce method. + # We call the function on each of the mirrored variables with the reduced + # value. + if self._aggregation == vs.VariableAggregation.NONE: + raise ValueError("You must specify an aggregation method to update a " + "MirroredVariable in Tower Context.") + + def merge_fn(strategy, value): + return strategy.update( + self, f, + strategy.reduce( + aggregation=self._aggregation, value=value, destinations=self)) + + return distribute_lib.get_tower_context().merge_call(merge_fn, *args, + **kwargs) + def assign_sub(self, *args, **kwargs): - return self.get(device=_get_update_device()).assign_sub(*args, **kwargs) + return self._assign_func(f=state_ops.assign_sub, *args, **kwargs) def assign_add(self, *args, **kwargs): - return self.get(device=_get_update_device()).assign_add(*args, **kwargs) + return self._assign_func(f=state_ops.assign_add, *args, **kwargs) def assign(self, *args, **kwargs): - return self.get(device=_get_update_device()).assign(*args, **kwargs) + return self._assign_func(f=state_ops.assign, *args, **kwargs) + + def is_initialized(self, name=None): + # We have to cast the self._index.values() to a `list` because when we + # use `model_to_estimator` to run tf.keras models, self._index.values() is + # of type `dict_values` and not `list`. + values_list = list(self._index.values()) + result = values_list[0].is_initialized() + # We iterate through the list of values except the last one to allow us to + # name the final `logical_and` op the same name that is passed by the user + # to the `is_initialized` op. For mirrored variables, the `is_initialized` + # op is a `logical_and` op. + for v in values_list[1:-1]: + result = math_ops.logical_and(result, v.is_initialized()) + result = math_ops.logical_and(result, values_list[-1].is_initialized(), + name=name) + return result + + @property + def initializer(self): + # return grouped ops of all the var initializations of component values of + # the mirrored variable + return control_flow_ops.group([v.initializer for v in self._index.values()]) + + @property + def aggregation(self): + return self._aggregation def _get_cross_tower(self): device = device_util.canonicalize(device_util.current()) @@ -374,7 +443,7 @@ class _TowerLocalSaveable(saver.BaseSaverBuilder.SaveableObject): # To preserve the sum across save and restore, we have to divide the # total across all devices when restoring a variable that was summed # when saving. - if self._tower_local_variable.reduce_method == "sum": + if self._tower_local_variable.aggregation == vs.VariableAggregation.SUM: tensor *= 1. / len(self._tower_local_variable.devices) return control_flow_ops.group([ _assign_on_device(d, v, tensor) @@ -391,9 +460,15 @@ class TowerLocalVariable(DistributedVariable, PerDevice, checkpointable.CheckpointableBase): """Holds a map from device to variables whose values are reduced on save.""" - def __init__(self, index, primary_var, reduce_method): + def __init__(self, index, primary_var, aggregation): self._primary_var = primary_var - self._reduce_method = reduce_method + self._aggregation = aggregation + # tf.keras keeps track of variables initialized using this attribute. When + # tf.keras gets the default session, it initializes all uninitialized vars. + # We need to make _keras_initialized a member of TowerLocalVariable because + # without this it will use `__getattr__` which will delegate to a component + # variable. + self._keras_initialized = False super(TowerLocalVariable, self).__init__(index) def assign_sub(self, *args, **kwargs): @@ -408,15 +483,37 @@ class TowerLocalVariable(DistributedVariable, PerDevice, _assert_tower_context() return self.get().assign(*args, **kwargs) + def is_initialized(self, name=None): + # We have to cast the self._index.values() to a `list` because when we + # use `model_to_estimator` to run tf.keras models, self._index.values() is + # of type `dict_values` and not `list`. + values_list = list(self._index.values()) + result = values_list[0].is_initialized() + # We iterate through the list of values except the last one to allow us to + # name the final `logical_and` op the same name that is passed by the user + # to the `is_initialized` op. For tower local variables, the + # `is_initialized` op is a `logical_and` op. + for v in values_list[1:-1]: + result = math_ops.logical_and(result, v.is_initialized()) + result = math_ops.logical_and(result, values_list[-1].is_initialized(), + name=name) + return result + @property - def reduce_method(self): - return self._reduce_method + def initializer(self): + # return grouped ops of all the var initializations of component values of + # the tower local variable + return control_flow_ops.group([v.initializer for v in self._index.values()]) + + @property + def aggregation(self): + return self._aggregation def _get_cross_tower(self): all_components = tuple(self._index.values()) # TODO(josh11b): Use a strategy-specific method. total = math_ops.add_n(all_components) - if self._reduce_method == "mean": + if self._aggregation == vs.VariableAggregation.MEAN: return total * (1./ len(all_components)) return total @@ -824,3 +921,71 @@ class MapOutput(object): def get(self): return self._l + + +class MultiStepContext(object): + """A context object that can be used to capture things when running steps. + + This context object is useful when running multiple steps at a time using the + `run_steps_on_dataset` API. For e.g. it allows the user's step function to + specify which outputs to emit at what frequency. Currently it only supports + capturing output from the last step, but will soon be augmented to support + other use cases such as output each N steps. + """ + + def __init__(self, initial_loop_values=None): + """Initializes an output context. + + Args: + initial_loop_values: Initial values passed to the run steps + while loop. The only purpose is to verify the shapes and types + when the actual output is set. This will be removed once we + automatically infer the output shapes and types (and do not need to + check for user error in specifying them manually). + Returns: + A context object. + """ + self._last_step_outputs = None + self._non_tensor_outputs = None + self._initial_loop_values = initial_loop_values + + @property + def last_step_outputs(self): + """Return the last step's outputs.""" + return self._last_step_outputs + + @last_step_outputs.setter + def last_step_outputs(self, outputs): + """Set the last step's outputs.""" + self._verify_structure_shapes_types(outputs, self._initial_loop_values) + self._last_step_outputs = outputs + + @property + def non_tensor_outputs(self): + """Return the non tensor outputs.""" + return self._non_tensor_outputs + + @non_tensor_outputs.setter + def non_tensor_outputs(self, outputs): + """Set any non tensor outputs.""" + self._non_tensor_outputs = outputs + + def _verify_structure_shapes_types(self, left, right): + """Verify that the structure, shapes and types of left are same as right.""" + nest.assert_same_structure(left, right) + flat_left = nest.flatten(left) + flat_right = nest.flatten(right) + assert len(flat_left) == len(flat_right), ( + "Length of left {} and right {} should be same.". + format(len(flat_left), len(flat_right))) + + for o, i in zip(flat_left, flat_right): + # TODO(priyag): Add checks for other types like IndexedSlices. + if isinstance(o, ops.Tensor): + assert isinstance(i, ops.Tensor) + assert o.shape == i.shape, ( + "Shape {} of left {} doesn't match shape {} of right {}.". + format(o.shape, o, i.shape, i)) + assert o.dtype == i.dtype, ( + "Dtype {} of left {} doesn't match dtype {} of right {}.". + format(o.dtype, o, i.dtype, i)) diff --git a/tensorflow/contrib/distribute/python/values_test.py b/tensorflow/contrib/distribute/python/values_test.py index c5b246e8041500e478478d1bb1527c3fe752b377..8e44f2fea16ac851c124b573948ee14ec0640556 100644 --- a/tensorflow/contrib/distribute/python/values_test.py +++ b/tensorflow/contrib/distribute/python/values_test.py @@ -158,7 +158,8 @@ def _make_mirrored(): v.append(variable_scope.get_variable( name=n, initializer=init, use_resource=True)) index[d] = v[-1] - mirrored = values.MirroredVariable(index, v[0]) + mirrored = values.MirroredVariable(index, v[0], + variable_scope.VariableAggregation.SUM) return v, devices, mirrored @@ -277,7 +278,8 @@ class RegroupAndSelectDeviceTest(test.TestCase): v = variable_scope.get_variable( name="v", initializer=1., use_resource=True) index = {d: v} - mirrored = values.MirroredVariable(index, v) + mirrored = values.MirroredVariable(index, v, + variable_scope.VariableAggregation.SUM) result = values.regroup(index) self.assertIs(mirrored, result) @@ -581,7 +583,8 @@ class MirroredVariableTest(test.TestCase): v = variable_scope.get_variable( name="v", initializer=[1.], use_resource=True) index = {"/job:foo/device:CPU:0": v} - mirrored = values.MirroredVariable(index, v) + mirrored = values.MirroredVariable(index, v, + variable_scope.VariableAggregation.MEAN) self.assertEquals(v.name, mirrored.name) self.assertEquals(v.dtype, mirrored.dtype) @@ -716,7 +719,9 @@ class MirroredVariableTest(test.TestCase): with ops.device("/device:GPU:0"): v = variable_scope.get_variable( name="v", initializer=1., use_resource=True) - mirrored = values.MirroredVariable({"/device:GPU:0": v}, v) + mirrored = values.MirroredVariable({ + "/device:GPU:0": v + }, v, variable_scope.VariableAggregation.MEAN) sess.run(variables_lib.global_variables_initializer()) sess.run({"complicated": mirrored}) @@ -746,24 +751,27 @@ class TowerLocalVariableTest(test.TestCase): if context.num_gpus() < 1 and context.executing_eagerly(): self.skipTest("A GPU is not available for this test in eager mode.") - v, tower_local = _make_tower_local("sum") + v, tower_local = _make_tower_local(variable_scope.VariableAggregation.SUM) self.assertEquals(v[0].name, tower_local.name) self.assertEquals(v[0].dtype, tower_local.dtype) self.assertEquals(v[0].shape, tower_local.shape) - self.assertEquals("sum", tower_local.reduce_method) + self.assertEquals(variable_scope.VariableAggregation.SUM, + tower_local.aggregation) @test_util.run_in_graph_and_eager_modes(config=config) def testVariableOnAnotherDevice(self): v = variable_scope.get_variable( name="v", initializer=[1.], use_resource=True) index = {"/job:foo/device:CPU:0": v} - tower_local = values.TowerLocalVariable(index, v, "mean") + tower_local = values.TowerLocalVariable( + index, v, variable_scope.VariableAggregation.MEAN) self.assertEquals(v.name, tower_local.name) self.assertEquals(v.dtype, tower_local.dtype) self.assertEquals(v.shape, tower_local.shape) - self.assertEquals("mean", tower_local.reduce_method) + self.assertEquals(variable_scope.VariableAggregation.MEAN, + tower_local.aggregation) def _assign_tower_local(self, devices, v, new): for d, var, n in zip(devices, v, new): @@ -789,7 +797,7 @@ class TowerLocalVariableTest(test.TestCase): self.skipTest("A GPU is not available for this test in eager mode.") with self.test_session() as sess: - v, tower_local = _make_tower_local("sum") + v, tower_local = _make_tower_local(variable_scope.VariableAggregation.SUM) # Overwrite the initial values. self._assign_tower_local(_devices, v, [3., 4.]) @@ -812,7 +820,8 @@ class TowerLocalVariableTest(test.TestCase): self.skipTest("A GPU is not available for this test in eager mode.") with self.test_session() as sess: - v, tower_local = _make_tower_local("mean") + v, tower_local = _make_tower_local( + variable_scope.VariableAggregation.MEAN) # Overwrite the initial values. self._assign_tower_local(_devices, v, [3., 4.]) @@ -831,7 +840,8 @@ class TowerLocalVariableTest(test.TestCase): def _save_tower_local_mean(self): """Save variables with mirroring, returns save_path.""" with self.test_session(graph=ops.Graph()) as sess: - v, tower_local = _make_tower_local("mean") + v, tower_local = _make_tower_local( + variable_scope.VariableAggregation.MEAN) # Overwrite the initial values. self._assign_tower_local(_devices, v, [3., 4.]) @@ -893,7 +903,8 @@ class TowerLocalVariableTest(test.TestCase): def _restore_tower_local_mean(self, save_path): """Restore to variables with mirroring in a fresh graph.""" with self.test_session(graph=ops.Graph()) as sess: - v, tower_local = _make_tower_local("mean") + v, tower_local = _make_tower_local( + variable_scope.VariableAggregation.MEAN) # Overwrite the initial values. self._assign_tower_local(_devices, v, [7., 8.]) @@ -907,7 +918,7 @@ class TowerLocalVariableTest(test.TestCase): def _restore_tower_local_sum(self, save_path): """Restore to variables with mirroring in a fresh graph.""" with self.test_session(graph=ops.Graph()) as sess: - v, tower_local = _make_tower_local("sum") + v, tower_local = _make_tower_local(variable_scope.VariableAggregation.SUM) # Overwrite the initial values. self._assign_tower_local(_devices, v, [7., 8.]) @@ -968,7 +979,7 @@ class TowerLocalVariableTest(test.TestCase): def testTensorConversion(self): with context.graph_mode(): - _, tower_local = _make_tower_local("sum") + _, tower_local = _make_tower_local(variable_scope.VariableAggregation.SUM) converted = ops.internal_convert_to_tensor(tower_local, as_ref=False) self.assertIsInstance(converted, ops.Tensor) self.assertEqual(converted.dtype, tower_local.dtype) diff --git a/tensorflow/contrib/eager/python/datasets.py b/tensorflow/contrib/eager/python/datasets.py index adf92c27ea0a27c5741bcdd175b277462cb28d02..58c548d798178a2848006cbf301f7d5cb2143f24 100644 --- a/tensorflow/contrib/eager/python/datasets.py +++ b/tensorflow/contrib/eager/python/datasets.py @@ -102,6 +102,7 @@ class Iterator(iterator_ops.EagerIterator, checkpointable.CheckpointableBase): with ops.device(self._device): self._buffer_resource_handle = prefetching_ops.function_buffering_resource( # pylint: disable=line-too-long string_arg=iter_string_handle, + output_types=self._flat_output_types, f=remote_fn, target_device=target, buffer_size=10, diff --git a/tensorflow/contrib/eager/python/examples/densenet/BUILD b/tensorflow/contrib/eager/python/examples/densenet/BUILD new file mode 100644 index 0000000000000000000000000000000000000000..de2a817d173b9d07b101e0e1dc959fe9c4baf768 --- /dev/null +++ b/tensorflow/contrib/eager/python/examples/densenet/BUILD @@ -0,0 +1,29 @@ +licenses(["notice"]) # Apache 2.0 + +package(default_visibility = ["//tensorflow:internal"]) + +load("//tensorflow:tensorflow.bzl", "cuda_py_test") + +py_binary( + name = "densenet", + srcs = ["densenet.py"], + srcs_version = "PY2AND3", + deps = [ + "//tensorflow:tensorflow_py", + "//tensorflow/contrib/eager/python:tfe", + ], +) + +cuda_py_test( + name = "densenet_test", + srcs = ["densenet_test.py"], + additional_deps = [ + ":densenet", + "//tensorflow/contrib/eager/python:tfe", + "//tensorflow:tensorflow_py", + ], + tags = [ + "no_pip", + "optonly", + ], +) diff --git a/tensorflow/contrib/eager/python/examples/densenet/densenet.py b/tensorflow/contrib/eager/python/examples/densenet/densenet.py new file mode 100644 index 0000000000000000000000000000000000000000..3a2b2de25023868a374f984a69fa4a65aecc1cbd --- /dev/null +++ b/tensorflow/contrib/eager/python/examples/densenet/densenet.py @@ -0,0 +1,274 @@ +# Copyright 2017 The TensorFlow Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================== +"""Densely Connected Convolutional Networks. + +Reference [ +Densely Connected Convolutional Networks](https://arxiv.org/abs/1608.06993) + +""" + +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + +import tensorflow as tf +l2 = tf.keras.regularizers.l2 + + +class ConvBlock(tf.keras.Model): + """Convolutional Block consisting of (batchnorm->relu->conv). + + Arguments: + num_filters: number of filters passed to a convolutional layer. + bottleneck: if True, then a 1x1 Conv is performed followed by 3x3 Conv. + weight_decay: weight decay + dropout_rate: dropout rate. + """ + + def __init__(self, num_filters, bottleneck, weight_decay=1e-4, + dropout_rate=0): + super(ConvBlock, self).__init__() + self.bottleneck = bottleneck + inter_filter = num_filters * 4 + # don't forget to set use_bias=False when using batchnorm + self.conv2 = tf.keras.layers.Conv2D(num_filters, + (3, 3), + padding="same", + use_bias=False, + kernel_initializer="he_normal", + kernel_regularizer=l2(weight_decay)) + self.batchnorm1 = tf.keras.layers.BatchNormalization() + self.dropout = tf.keras.layers.Dropout(dropout_rate) + + if self.bottleneck: + self.conv1 = tf.keras.layers.Conv2D(inter_filter, + (1, 1), + padding="same", + use_bias=False, + kernel_initializer="he_normal", + kernel_regularizer=l2(weight_decay)) + self.batchnorm2 = tf.keras.layers.BatchNormalization() + + def call(self, x, training=True): + output = self.batchnorm1(x, training=training) + + if self.bottleneck: + output = self.conv1(tf.nn.relu(output)) + output = self.batchnorm2(output, training=training) + + output = self.conv2(tf.nn.relu(output)) + output = self.dropout(output, training=training) + + return output + + +class TransitionBlock(tf.keras.Model): + """Transition Block to reduce the number of features. + + Arguments: + num_filters: number of filters passed to a convolutional layer. + weight_decay: weight decay + dropout_rate: dropout rate. + """ + + def __init__(self, num_filters, weight_decay=1e-4, dropout_rate=0): + super(TransitionBlock, self).__init__() + self.batchnorm = tf.keras.layers.BatchNormalization() + self.conv = tf.keras.layers.Conv2D(num_filters, + (1, 1), + padding="same", + use_bias=False, + kernel_initializer="he_normal", + kernel_regularizer=l2(weight_decay)) + self.avg_pool = tf.keras.layers.AveragePooling2D() + + def call(self, x, training=True): + output = self.batchnorm(x, training=training) + output = self.conv(tf.nn.relu(output)) + output = self.avg_pool(output) + return output + + +class DenseBlock(tf.keras.Model): + """Dense Block consisting of ConvBlocks where each block's + output is concatenated with its input. + + Arguments: + num_layers: Number of layers in each block. + growth_rate: number of filters to add per conv block. + bottleneck: boolean, that decides which part of ConvBlock to call. + weight_decay: weight decay + dropout_rate: dropout rate. + """ + + def __init__(self, num_layers, growth_rate, bottleneck, + weight_decay=1e-4, dropout_rate=0): + super(DenseBlock, self).__init__() + self.num_layers = num_layers + + self.blocks = [] + for _ in range(int(self.num_layers)): + self.blocks.append(ConvBlock(growth_rate, + bottleneck, + weight_decay, + dropout_rate)) + + def call(self, x, training=True): + for i in range(int(self.num_layers)): + output = self.blocks[i](x, training=training) + x = tf.concat([x, output], axis=-1) + + return x + + +class DenseNet(tf.keras.Model): + """Creating the Densenet Architecture. + + Arguments: + depth_of_model: number of layers in the model. + growth_rate: number of filters to add per conv block. + num_of_blocks: number of dense blocks. + output_classes: number of output classes. + num_layers_in_each_block: number of layers in each block. + If -1, then we calculate this by (depth-3)/4. + If positive integer, then the it is used as the + number of layers per block. + If list or tuple, then this list is used directly. + bottleneck: boolean, to decide which part of conv block to call. + compression: reducing the number of inputs(filters) to the transition block. + weight_decay: weight decay + rate: dropout rate. + pool_initial: If True add a 7x7 conv with stride 2 followed by 3x3 maxpool + else, do a 3x3 conv with stride 1. + include_top: If true, GlobalAveragePooling Layer and Dense layer are + included. + """ + + def __init__(self, depth_of_model, growth_rate, num_of_blocks, + output_classes, num_layers_in_each_block, + bottleneck=True, compression=0.5, weight_decay=1e-4, + dropout_rate=0, pool_initial=False, include_top=True): + super(DenseNet, self).__init__() + self.depth_of_model = depth_of_model + self.growth_rate = growth_rate + self.num_of_blocks = num_of_blocks + self.output_classes = output_classes + self.num_layers_in_each_block = num_layers_in_each_block + self.bottleneck = bottleneck + self.compression = compression + self.weight_decay = weight_decay + self.dropout_rate = dropout_rate + self.pool_initial = pool_initial + self.include_top = include_top + + # deciding on number of layers in each block + if isinstance(self.num_layers_in_each_block, list) or isinstance( + self.num_layers_in_each_block, tuple): + self.num_layers_in_each_block = list(self.num_layers_in_each_block) + else: + if self.num_layers_in_each_block == -1: + if self.num_of_blocks != 3: + raise ValueError( + "Number of blocks must be 3 if num_layers_in_each_block is -1") + if (self.depth_of_model - 4) % 3 == 0: + num_layers = (self.depth_of_model - 4) / 3 + if self.bottleneck: + num_layers //= 2 + self.num_layers_in_each_block = [num_layers] * self.num_of_blocks + else: + raise ValueError("Depth must be 3N+4 if num_layer_in_each_block=-1") + else: + self.num_layers_in_each_block = [ + self.num_layers_in_each_block] * self.num_of_blocks + + # setting the filters and stride of the initial covn layer. + if self.pool_initial: + init_filters = (7, 7) + stride = (2, 2) + else: + init_filters = (3, 3) + stride = (1, 1) + + self.num_filters = 2 * self.growth_rate + + # first conv and pool layer + self.conv1 = tf.keras.layers.Conv2D(self.num_filters, + init_filters, + strides=stride, + padding="same", + use_bias=False, + kernel_initializer="he_normal", + kernel_regularizer=l2( + self.weight_decay)) + if self.pool_initial: + self.pool1 = tf.keras.layers.MaxPooling2D(pool_size=(3, 3), + strides=(2, 2), + padding="same") + self.batchnorm1 = tf.keras.layers.BatchNormalization() + + self.batchnorm2 = tf.keras.layers.BatchNormalization() + + # last pooling and fc layer + if self.include_top: + self.last_pool = tf.keras.layers.GlobalAveragePooling2D() + self.classifier = tf.keras.layers.Dense(self.output_classes) + + # calculating the number of filters after each block + num_filters_after_each_block = [self.num_filters] + for i in range(1, self.num_of_blocks): + temp_num_filters = num_filters_after_each_block[i-1] + ( + self.growth_rate * self.num_layers_in_each_block[i-1]) + # using compression to reduce the number of inputs to the + # transition block + temp_num_filters = int(temp_num_filters * compression) + num_filters_after_each_block.append(temp_num_filters) + + # dense block initialization + self.dense_blocks = [] + self.transition_blocks = [] + for i in range(self.num_of_blocks): + self.dense_blocks.append(DenseBlock(self.num_layers_in_each_block[i], + self.growth_rate, + self.bottleneck, + self.weight_decay, + self.dropout_rate)) + if i+1 < self.num_of_blocks: + self.transition_blocks.append( + TransitionBlock(num_filters_after_each_block[i+1], + self.weight_decay, + self.dropout_rate)) + + def call(self, x, training=True): + output = self.conv1(x) + + if self.pool_initial: + output = self.batchnorm1(output, training=training) + output = tf.nn.relu(output) + output = self.pool1(output) + + for i in range(self.num_of_blocks - 1): + output = self.dense_blocks[i](output, training=training) + output = self.transition_blocks[i](output, training=training) + + output = self.dense_blocks[ + self.num_of_blocks - 1](output, training=training) + output = self.batchnorm2(output, training=training) + output = tf.nn.relu(output) + + if self.include_top: + output = self.last_pool(output) + output = self.classifier(output) + + return output diff --git a/tensorflow/contrib/eager/python/examples/densenet/densenet_test.py b/tensorflow/contrib/eager/python/examples/densenet/densenet_test.py new file mode 100644 index 0000000000000000000000000000000000000000..56d3362f3ba91011aff88a905d9afdf371561007 --- /dev/null +++ b/tensorflow/contrib/eager/python/examples/densenet/densenet_test.py @@ -0,0 +1,83 @@ +# Copyright 2018 The TensorFlow Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================== +"""Tests for various Densenet architectures.""" + +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + +import tensorflow as tf +from tensorflow.contrib.eager.python.examples.densenet import densenet + + +class DensenetTest(tf.test.TestCase): + + def test_bottleneck_true(self): + depth = 7 + growth_rate = 2 + num_blocks = 3 + output_classes = 10 + num_layers_in_each_block = -1 + batch_size = 1 + + model = densenet.DenseNet(depth, growth_rate, num_blocks, + output_classes, num_layers_in_each_block, + bottleneck=True, compression=0.5, + weight_decay=1e-4, dropout_rate=0, + pool_initial=False, include_top=True) + + rand_input = tf.random_uniform((batch_size, 32, 32, 3)) + output_shape = model(rand_input).shape + self.assertEqual(output_shape, (batch_size, output_classes)) + + def test_bottleneck_false(self): + depth = 7 + growth_rate = 2 + num_blocks = 3 + output_classes = 10 + num_layers_in_each_block = -1 + batch_size = 1 + + model = densenet.DenseNet(depth, growth_rate, num_blocks, + output_classes, num_layers_in_each_block, + bottleneck=False, compression=0.5, + weight_decay=1e-4, dropout_rate=0, + pool_initial=False, include_top=True) + + rand_input = tf.random_uniform((batch_size, 32, 32, 3)) + output_shape = model(rand_input).shape + self.assertEqual(output_shape, (batch_size, output_classes)) + + def test_pool_initial_true(self): + depth = 7 + growth_rate = 2 + num_blocks = 4 + output_classes = 10 + num_layers_in_each_block = [1, 2, 2, 1] + batch_size = 1 + + model = densenet.DenseNet(depth, growth_rate, num_blocks, + output_classes, num_layers_in_each_block, + bottleneck=True, compression=0.5, + weight_decay=1e-4, dropout_rate=0, + pool_initial=True, include_top=True) + + rand_input = tf.random_uniform((batch_size, 32, 32, 3)) + output_shape = model(rand_input).shape + self.assertEqual(output_shape, (batch_size, output_classes)) + +if __name__ == '__main__': + tf.enable_eager_execution() + tf.test.main() diff --git a/tensorflow/contrib/eager/python/examples/gan/mnist.py b/tensorflow/contrib/eager/python/examples/gan/mnist.py index cc9cf53410f641cc3303b4450e9eaa1301904a64..b33243021bb56492a3f5543c3e2a5f07c66f021f 100644 --- a/tensorflow/contrib/eager/python/examples/gan/mnist.py +++ b/tensorflow/contrib/eager/python/examples/gan/mnist.py @@ -214,7 +214,7 @@ def train_one_epoch(generator, discriminator, generator_optimizer, total_generator_loss = 0.0 total_discriminator_loss = 0.0 - for (batch_index, images) in enumerate(tfe.Iterator(dataset)): + for (batch_index, images) in enumerate(dataset): with tf.device('/cpu:0'): tf.assign_add(step_counter, 1) @@ -227,7 +227,10 @@ def train_one_epoch(generator, discriminator, generator_optimizer, maxval=1., seed=batch_index) - with tf.GradientTape(persistent=True) as g: + # we can use 2 tapes or a single persistent tape. + # Using two tapes is memory efficient since intermediate tensors can be + # released between the two .gradient() calls below + with tf.GradientTape() as gen_tape, tf.GradientTape() as disc_tape: generated_images = generator(noise) tf.contrib.summary.image( 'generated_images', @@ -243,9 +246,10 @@ def train_one_epoch(generator, discriminator, generator_optimizer, generator_loss_val = generator_loss(discriminator_gen_outputs) total_generator_loss += generator_loss_val - generator_grad = g.gradient(generator_loss_val, generator.variables) - discriminator_grad = g.gradient(discriminator_loss_val, - discriminator.variables) + generator_grad = gen_tape.gradient(generator_loss_val, + generator.variables) + discriminator_grad = disc_tape.gradient(discriminator_loss_val, + discriminator.variables) generator_optimizer.apply_gradients( zip(generator_grad, generator.variables)) diff --git a/tensorflow/contrib/eager/python/examples/generative_examples/image_captioning_with_attention.ipynb b/tensorflow/contrib/eager/python/examples/generative_examples/image_captioning_with_attention.ipynb new file mode 100644 index 0000000000000000000000000000000000000000..1a5a186e7a3e456cc43f8091370d3eeb795d5e0e --- /dev/null +++ b/tensorflow/contrib/eager/python/examples/generative_examples/image_captioning_with_attention.ipynb @@ -0,0 +1,1184 @@ +{ + "nbformat": 4, + "nbformat_minor": 0, + "metadata": { + "colab": { + "name": "image_captioning_with_attention.ipynb", + "version": "0.3.2", + "views": {}, + "default_view": {}, + "provenance": [ + { + "file_id": "1HI8OK2sMjcx9CTWVn0122QAHOuXaOaMg", + "timestamp": 1530222436922 + } + ], + "private_outputs": true, + "collapsed_sections": [], + "toc_visible": true + }, + "kernelspec": { + "display_name": "Python 3", + "language": "python", + "name": "python3" + }, + "accelerator": "GPU" + }, + "cells": [ + { + "metadata": { + "id": "K2s1A9eLRPEj", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "##### Copyright 2018 The TensorFlow Authors.\n", + "\n", + "Licensed under the Apache License, Version 2.0 (the \"License\").\n" + ] + }, + { + "metadata": { + "id": "Cffg2i257iMS", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "# Image Captioning with Attention\n", + "\n", + "
\n", + "\n", + " Run in Google Colab \n", + "\n", + "View source on GitHub
" + ] + }, + { + "metadata": { + "id": "QASbY_HGo4Lq", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "Image captioning is the task of generating a caption for an image. Given an image like this:\n", + "\n", + "![Man Surfing](https://tensorflow.org/images/surf.jpg) \n", + "\n", + "[Image Source](https://commons.wikimedia.org/wiki/Surfing#/media/File:Surfing_in_Hawaii.jpg), License: Public Domain\n", + "\n", + "Our goal is generate a caption, such as \"a surfer riding on a wave\". Here, we'll use an attention based model. This enables us to see which parts of the image the model focuses on as it generates a caption.\n", + "\n", + "![Prediction](https://tensorflow.org/images/imcap_prediction.png)\n", + "\n", + "This model architecture below is similar to [Show, Attend and Tell: Neural Image Caption Generation with Visual Attention](https://arxiv.org/abs/1502.03044). \n", + "\n", + "The code uses [tf.keras](https://www.tensorflow.org/programmers_guide/keras) and [eager execution](https://www.tensorflow.org/programmers_guide/eager), which you can learn more about in the linked guides.\n", + "\n", + "This notebook is an end-to-end example. If you run it, it will download the [MS-COCO](http://cocodataset.org/#home) dataset, preprocess and cache a subset of the images using Inception V3, train an encoder-decoder model, and use it to generate captions on new images.\n", + "\n", + "The code requires TensorFlow version >=1.9. If you're running this in [Colab]()\n", + "\n", + "In this example, we're training on a relatively small amount of data as an example. On a single P100 GPU, this example will take about ~2 hours to train. We train on the first 30,000 captions (corresponding to about ~20,000 images depending on shuffling, as there are multiple captions per image in the dataset)\n" + ] + }, + { + "metadata": { + "id": "U8l4RJ0XRPEm", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + } + }, + "cell_type": "code", + "source": [ + "# Import TensorFlow and enable eager execution\n", + "# This code requires TensorFlow version >=1.9\n", + "import tensorflow as tf\n", + "tf.enable_eager_execution()\n", + "\n", + "# We'll generate plots of attention in order to see which parts of an image\n", + "# our model focuses on during captioning\n", + "import matplotlib.pyplot as plt\n", + "\n", + "# Scikit-learn includes many helpful utilities\n", + "from sklearn.model_selection import train_test_split\n", + "from sklearn.utils import shuffle\n", + "\n", + "import re\n", + "import numpy as np\n", + "import os\n", + "import time\n", + "import json\n", + "from glob import glob\n", + "from PIL import Image\n", + "import pickle" + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "b6qbGw8MRPE5", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "## Download and prepare the MS-COCO dataset\n", + "\n", + "We will use the [MS-COCO dataset](http://cocodataset.org/#home) to train our model. This dataset contains >82,000 images, each of which has been annotated with at least 5 different captions. The code code below will download and extract the dataset automatically. \n", + "\n", + "**Caution: large download ahead**. We'll use the training set, it's a 13GB file." + ] + }, + { + "metadata": { + "id": "krQuPYTtRPE7", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + } + }, + "cell_type": "code", + "source": [ + "annotation_zip = tf.keras.utils.get_file('captions.zip', \n", + " cache_subdir=os.path.abspath('.'),\n", + " origin = 'http://images.cocodataset.org/annotations/annotations_trainval2014.zip',\n", + " extract = True)\n", + "annotation_file = os.path.dirname(annotation_zip)+'/annotations/captions_train2014.json'\n", + "\n", + "name_of_zip = 'train2014.zip'\n", + "if not os.path.exists(os.path.abspath('.') + '/' + name_of_zip):\n", + " image_zip = tf.keras.utils.get_file(name_of_zip, \n", + " cache_subdir=os.path.abspath('.'),\n", + " origin = 'http://images.cocodataset.org/zips/train2014.zip',\n", + " extract = True)\n", + " PATH = os.path.dirname(image_zip)+'/train2014/'\n", + "else:\n", + " PATH = os.path.abspath('.')+'/train2014/'" + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "aANEzb5WwSzg", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "## Optionally, limit the size of the training set for faster training\n", + "For this example, we'll select a subset of 30,000 captions and use these and the corresponding images to train our model. As always, captioning quality will improve if you choose to use more data." + ] + }, + { + "metadata": { + "id": "4G3b8x8_RPFD", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + } + }, + "cell_type": "code", + "source": [ + "# read the json file\n", + "with open(annotation_file, 'r') as f:\n", + " annotations = json.load(f)\n", + "\n", + "# storing the captions and the image name in vectors\n", + "all_captions = []\n", + "all_img_name_vector = []\n", + "\n", + "for annot in annotations['annotations']:\n", + " caption = ' ' + annot['caption'] + ' '\n", + " image_id = annot['image_id']\n", + " full_coco_image_path = PATH + 'COCO_train2014_' + '%012d.jpg' % (image_id)\n", + " \n", + " all_img_name_vector.append(full_coco_image_path)\n", + " all_captions.append(caption)\n", + "\n", + "# shuffling the captions and image_names together\n", + "# setting a random state\n", + "train_captions, img_name_vector = shuffle(all_captions,\n", + " all_img_name_vector,\n", + " random_state=1)\n", + "\n", + "# selecting the first 30000 captions from the shuffled set\n", + "num_examples = 30000\n", + "train_captions = train_captions[:num_examples]\n", + "img_name_vector = img_name_vector[:num_examples]" + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "mPBMgK34RPFL", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + } + }, + "cell_type": "code", + "source": [ + "len(train_captions), len(all_captions)" + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "8cSW4u-ORPFQ", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "## Preprocess the images using InceptionV3\n", + "Next, we will use InceptionV3 (pretrained on Imagenet) to classify each image. We will extract features from the last convolutional layer. \n", + "\n", + "First, we will need to convert the images into the format inceptionV3 expects by:\n", + "* Resizing the image to (299, 299)\n", + "* Using the [preprocess_input](https://www.tensorflow.org/api_docs/python/tf/keras/applications/inception_v3/preprocess_input) method to place the pixels in the range of -1 to 1 (to match the format of the images used to train InceptionV3)." + ] + }, + { + "metadata": { + "id": "zXR0217aRPFR", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + } + }, + "cell_type": "code", + "source": [ + "def load_image(image_path):\n", + " img = tf.read_file(image_path)\n", + " img = tf.image.decode_jpeg(img, channels=3)\n", + " img = tf.image.resize_images(img, (299, 299))\n", + " img = tf.keras.applications.inception_v3.preprocess_input(img)\n", + " return img, image_path" + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "MDvIu4sXRPFV", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "## Initialize InceptionV3 and load the pretrained Imagenet weights\n", + "\n", + "To do so, we'll create a tf.keras model where the output layer is the last convolutional layer in the InceptionV3 architecture. \n", + "* Each image is forwarded through the network and the vector that we get at the end is stored in a dictionary (image_name --> feature_vector). \n", + "* We use the last convolutional layer because we are using attention in this example. The shape of the output of this layer is ```8x8x2048```. \n", + "* We avoid doing this during training so it does not become a bottleneck. \n", + "* After all the images are passed through the network, we pickle the dictionary and save it to disk." + ] + }, + { + "metadata": { + "id": "RD3vW4SsRPFW", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + } + }, + "cell_type": "code", + "source": [ + "image_model = tf.keras.applications.InceptionV3(include_top=False, \n", + " weights='imagenet')\n", + "new_input = image_model.input\n", + "hidden_layer = image_model.layers[-1].output\n", + "\n", + "image_features_extract_model = tf.keras.Model(new_input, hidden_layer)" + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "rERqlR3WRPGO", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "## Caching the features extracted from InceptionV3\n", + "\n", + "We will pre-process each image with InceptionV3 and cache the output to disk. Caching the output in RAM would be faster but memory intensive, requiring 8 \\* 8 \\* 2048 floats per image. At the time of writing, this would exceed the memory limitations of Colab (although these may change, an instance appears to have about 12GB of memory currently). \n", + "\n", + "Performance could be improved with a more sophisticated caching strategy (e.g., by sharding the images to reduce random access disk I/O) at the cost of more code.\n", + "\n", + "This will take about 10 minutes to run in Colab with a GPU. If you'd like to see a progress bar, you could: install [tqdm](https://github.com/tqdm/tqdm) (```!pip install tqdm```), then change this line: \n", + "\n", + "```for img, path in image_dataset:``` \n", + "\n", + "to:\n", + "\n", + "```for img, path in tqdm(image_dataset):```." + ] + }, + { + "metadata": { + "id": "Dx_fvbVgRPGQ", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + } + }, + "cell_type": "code", + "source": [ + "# getting the unique images\n", + "encode_train = sorted(set(img_name_vector))\n", + "\n", + "# feel free to change the batch_size according to your system configuration\n", + "image_dataset = tf.data.Dataset.from_tensor_slices(\n", + " encode_train).map(load_image).batch(16)\n", + "\n", + "for img, path in image_dataset:\n", + " batch_features = image_features_extract_model(img)\n", + " batch_features = tf.reshape(batch_features, \n", + " (batch_features.shape[0], -1, batch_features.shape[3]))\n", + "\n", + " for bf, p in zip(batch_features, path):\n", + " path_of_feature = p.numpy().decode(\"utf-8\")\n", + " np.save(path_of_feature, bf.numpy())" + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "nyqH3zFwRPFi", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "## Preprocess and tokenize the captions\n", + "\n", + "* First, we'll tokenize the captions (e.g., by splitting on spaces). This will give us a vocabulary of all the unique words in the data (e.g., \"surfing\", \"football\", etc).\n", + "* Next, we'll limit the vocabulary size to the top 5,000 words to save memory. We'll replace all other words with the token \"UNK\" (for unknown).\n", + "* Finally, we create a word --> index mapping and vice-versa.\n", + "* We will then pad all sequences to the be same length as the longest one. " + ] + }, + { + "metadata": { + "id": "HZfK8RhQRPFj", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + } + }, + "cell_type": "code", + "source": [ + "# This will find the maximum length of any caption in our dataset\n", + "def calc_max_length(tensor):\n", + " return max(len(t) for t in tensor)" + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "oJGE34aiRPFo", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + } + }, + "cell_type": "code", + "source": [ + "# The steps above is a general process of dealing with text processing\n", + "\n", + "# choosing the top 5000 words from the vocabulary\n", + "top_k = 5000\n", + "tokenizer = tf.keras.preprocessing.text.Tokenizer(num_words=top_k, \n", + " oov_token=\"\", \n", + " filters='!\"#$%&()*+.,-/:;=?@[\\]^_`{|}~ ')\n", + "tokenizer.fit_on_texts(train_captions)\n", + "train_seqs = tokenizer.texts_to_sequences(train_captions)" + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "8Q44tNQVRPFt", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + } + }, + "cell_type": "code", + "source": [ + "tokenizer.word_index = {key:value for key, value in tokenizer.word_index.items() if value <= top_k}\n", + "# putting token in the word2idx dictionary\n", + "tokenizer.word_index[tokenizer.oov_token] = top_k + 1\n", + "tokenizer.word_index[''] = 0" + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "0fpJb5ojRPFv", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + } + }, + "cell_type": "code", + "source": [ + "# creating the tokenized vectors\n", + "train_seqs = tokenizer.texts_to_sequences(train_captions)" + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "olQArbgbRPF1", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + } + }, + "cell_type": "code", + "source": [ + "# creating a reverse mapping (index -> word)\n", + "index_word = {value:key for key, value in tokenizer.word_index.items()}" + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "AidglIZVRPF4", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + } + }, + "cell_type": "code", + "source": [ + "# padding each vector to the max_length of the captions\n", + "# if the max_length parameter is not provided, pad_sequences calculates that automatically\n", + "cap_vector = tf.keras.preprocessing.sequence.pad_sequences(train_seqs, padding='post')" + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "gL0wkttkRPGA", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + } + }, + "cell_type": "code", + "source": [ + "# calculating the max_length \n", + "# used to store the attention weights\n", + "max_length = calc_max_length(train_seqs)" + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "M3CD75nDpvTI", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "## Split the data into training and testing" + ] + }, + { + "metadata": { + "id": "iS7DDMszRPGF", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + } + }, + "cell_type": "code", + "source": [ + "# Create training and validation sets using 80-20 split\n", + "img_name_train, img_name_val, cap_train, cap_val = train_test_split(img_name_vector, \n", + " cap_vector, \n", + " test_size=0.2, \n", + " random_state=0)" + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "XmViPkRFRPGH", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + } + }, + "cell_type": "code", + "source": [ + "len(img_name_train), len(cap_train), len(img_name_val), len(cap_val)" + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "uEWM9xrYcg45", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "## Our images and captions are ready! Next, let's create a tf.data dataset to use for training our model.\n", + "\n" + ] + }, + { + "metadata": { + "id": "Q3TnZ1ToRPGV", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + } + }, + "cell_type": "code", + "source": [ + "# feel free to change these parameters according to your system's configuration\n", + "\n", + "BATCH_SIZE = 64\n", + "BUFFER_SIZE = 1000\n", + "embedding_dim = 256\n", + "units = 512\n", + "vocab_size = len(tokenizer.word_index)\n", + "# shape of the vector extracted from InceptionV3 is (64, 2048)\n", + "# these two variables represent that\n", + "features_shape = 2048\n", + "attention_features_shape = 64" + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "SmZS2N0bXG3T", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + } + }, + "cell_type": "code", + "source": [ + "# loading the numpy files \n", + "def map_func(img_name, cap):\n", + " img_tensor = np.load(img_name.decode('utf-8')+'.npy')\n", + " return img_tensor, cap" + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "FDF_Nm3tRPGZ", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + } + }, + "cell_type": "code", + "source": [ + "dataset = tf.data.Dataset.from_tensor_slices((img_name_train, cap_train))\n", + "\n", + "# using map to load the numpy files in parallel\n", + "# NOTE: Be sure to set num_parallel_calls to the number of CPU cores you have\n", + "# https://www.tensorflow.org/api_docs/python/tf/py_func\n", + "dataset = dataset.map(lambda item1, item2: tf.py_func(\n", + " map_func, [item1, item2], [tf.float32, tf.int32]), num_parallel_calls=8)\n", + "\n", + "# shuffling and batching\n", + "dataset = dataset.shuffle(BUFFER_SIZE)\n", + "# https://www.tensorflow.org/api_docs/python/tf/contrib/data/batch_and_drop_remainder\n", + "dataset = dataset.batch(BATCH_SIZE)\n", + "dataset = dataset.prefetch(1)" + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "nrvoDphgRPGd", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "## Model\n", + "\n", + "Fun fact, the decoder below is identical to the one in the example for [Neural Machine Translation with Attention]( https://github.com/tensorflow/tensorflow/blob/master/tensorflow/contrib/eager/python/examples/nmt_with_attention/nmt_with_attention.ipynb).\n", + "\n", + "The model architecture is inspired by the [Show, Attend and Tell](https://arxiv.org/pdf/1502.03044.pdf) paper.\n", + "\n", + "* In this example, we extract the features from the lower convolutional layer of InceptionV3 giving us a vector of shape (8, 8, 2048). \n", + "* We squash that to a shape of (64, 2048).\n", + "* This vector is then passed through the CNN Encoder(which consists of a single Fully connected layer).\n", + "* The RNN(here GRU) attends over the image to predict the next word." + ] + }, + { + "metadata": { + "id": "AAppCGLKRPGd", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + } + }, + "cell_type": "code", + "source": [ + "def gru(units):\n", + " # If you have a GPU, we recommend using the CuDNNGRU layer (it provides a \n", + " # significant speedup).\n", + " if tf.test.is_gpu_available():\n", + " return tf.keras.layers.CuDNNGRU(units, \n", + " return_sequences=True, \n", + " return_state=True, \n", + " recurrent_initializer='glorot_uniform')\n", + " else:\n", + " return tf.keras.layers.GRU(units, \n", + " return_sequences=True, \n", + " return_state=True, \n", + " recurrent_activation='sigmoid', \n", + " recurrent_initializer='glorot_uniform')" + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "ja2LFTMSdeV3", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + } + }, + "cell_type": "code", + "source": [ + "class BahdanauAttention(tf.keras.Model):\n", + " def __init__(self, units):\n", + " super(BahdanauAttention, self).__init__()\n", + " self.W1 = tf.keras.layers.Dense(units)\n", + " self.W2 = tf.keras.layers.Dense(units)\n", + " self.V = tf.keras.layers.Dense(1)\n", + " \n", + " def call(self, features, hidden):\n", + " # features(CNN_encoder output) shape == (batch_size, 64, embedding_dim)\n", + " \n", + " # hidden shape == (batch_size, hidden_size)\n", + " # hidden_with_time_axis shape == (batch_size, 1, hidden_size)\n", + " hidden_with_time_axis = tf.expand_dims(hidden, 1)\n", + " \n", + " # score shape == (batch_size, 64, hidden_size)\n", + " score = tf.nn.tanh(self.W1(features) + self.W2(hidden_with_time_axis))\n", + " \n", + " # attention_weights shape == (batch_size, 64, 1)\n", + " # we get 1 at the last axis because we are applying score to self.V\n", + " attention_weights = tf.nn.softmax(self.V(score), axis=1)\n", + " \n", + " # context_vector shape after sum == (batch_size, hidden_size)\n", + " context_vector = attention_weights * features\n", + " context_vector = tf.reduce_sum(context_vector, axis=1)\n", + " \n", + " return context_vector, attention_weights" + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "AZ7R1RxHRPGf", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + } + }, + "cell_type": "code", + "source": [ + "class CNN_Encoder(tf.keras.Model):\n", + " # Since we have already extracted the features and dumped it using pickle\n", + " # This encoder passes those features through a Fully connected layer\n", + " def __init__(self, embedding_dim):\n", + " super(CNN_Encoder, self).__init__()\n", + " # shape after fc == (batch_size, 64, embedding_dim)\n", + " self.fc = tf.keras.layers.Dense(embedding_dim)\n", + " \n", + " def call(self, x):\n", + " x = self.fc(x)\n", + " x = tf.nn.relu(x)\n", + " return x" + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "V9UbGQmERPGi", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + } + }, + "cell_type": "code", + "source": [ + "class RNN_Decoder(tf.keras.Model):\n", + " def __init__(self, embedding_dim, units, vocab_size):\n", + " super(RNN_Decoder, self).__init__()\n", + " self.units = units\n", + "\n", + " self.embedding = tf.keras.layers.Embedding(vocab_size, embedding_dim)\n", + " self.gru = gru(self.units)\n", + " self.fc1 = tf.keras.layers.Dense(self.units)\n", + " self.fc2 = tf.keras.layers.Dense(vocab_size)\n", + " \n", + " self.attention = BahdanauAttention(self.units)\n", + " \n", + " def call(self, x, features, hidden):\n", + " # defining attention as a separate model\n", + " context_vector, attention_weights = self.attention(features, hidden)\n", + " \n", + " # x shape after passing through embedding == (batch_size, 1, embedding_dim)\n", + " x = self.embedding(x)\n", + " \n", + " # x shape after concatenation == (batch_size, 1, embedding_dim + hidden_size)\n", + " x = tf.concat([tf.expand_dims(context_vector, 1), x], axis=-1)\n", + " \n", + " # passing the concatenated vector to the GRU\n", + " output, state = self.gru(x)\n", + " \n", + " # shape == (batch_size, max_length, hidden_size)\n", + " x = self.fc1(output)\n", + " \n", + " # x shape == (batch_size * max_length, hidden_size)\n", + " x = tf.reshape(x, (-1, x.shape[2]))\n", + " \n", + " # output shape == (batch_size * max_length, vocab)\n", + " x = self.fc2(x)\n", + "\n", + " return x, state, attention_weights\n", + "\n", + " def reset_state(self, batch_size):\n", + " return tf.zeros((batch_size, self.units))" + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "Qs_Sr03wRPGk", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + } + }, + "cell_type": "code", + "source": [ + "encoder = CNN_Encoder(embedding_dim)\n", + "decoder = RNN_Decoder(embedding_dim, units, vocab_size)" + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "-bYN7xA0RPGl", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + } + }, + "cell_type": "code", + "source": [ + "optimizer = tf.train.AdamOptimizer()\n", + "\n", + "# We are masking the loss calculated for padding\n", + "def loss_function(real, pred):\n", + " mask = 1 - np.equal(real, 0)\n", + " loss_ = tf.nn.sparse_softmax_cross_entropy_with_logits(labels=real, logits=pred) * mask\n", + " return tf.reduce_mean(loss_)" + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "PHod7t72RPGn", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "## Training\n", + "\n", + "* We extract the features stored in the respective `.npy` files and then pass those features through the encoder.\n", + "* The encoder output, hidden state(initialized to 0) and the decoder input (which is the start token) is passed to the decoder.\n", + "* The decoder returns the predictions and the decoder hidden state.\n", + "* The decoder hidden state is then passed back into the model and the predictions are used to calculate the loss.\n", + "* Use teacher forcing to decide the next input to the decoder.\n", + "* Teacher forcing is the technique where the target word is passed as the next input to the decoder.\n", + "* The final step is to calculate the gradients and apply it to the optimizer and backpropagate.\n" + ] + }, + { + "metadata": { + "id": "Vt4WZ5mhJE-E", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + } + }, + "cell_type": "code", + "source": [ + "# adding this in a separate cell because if you run the training cell \n", + "# many times, the loss_plot array will be reset\n", + "loss_plot = []" + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "UlA4VIQpRPGo", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + } + }, + "cell_type": "code", + "source": [ + "EPOCHS = 20\n", + "\n", + "for epoch in range(EPOCHS):\n", + " start = time.time()\n", + " total_loss = 0\n", + " \n", + " for (batch, (img_tensor, target)) in enumerate(dataset):\n", + " loss = 0\n", + " \n", + " # initializing the hidden state for each batch\n", + " # because the captions are not related from image to image\n", + " hidden = decoder.reset_state(batch_size=target.shape[0])\n", + "\n", + " dec_input = tf.expand_dims([tokenizer.word_index['']] * BATCH_SIZE, 1)\n", + " \n", + " with tf.GradientTape() as tape:\n", + " features = encoder(img_tensor)\n", + " \n", + " for i in range(1, target.shape[1]):\n", + " # passing the features through the decoder\n", + " predictions, hidden, _ = decoder(dec_input, features, hidden)\n", + "\n", + " loss += loss_function(target[:, i], predictions)\n", + " \n", + " # using teacher forcing\n", + " dec_input = tf.expand_dims(target[:, i], 1)\n", + " \n", + " total_loss += (loss / int(target.shape[1]))\n", + " \n", + " variables = encoder.variables + decoder.variables\n", + " \n", + " gradients = tape.gradient(loss, variables) \n", + " \n", + " optimizer.apply_gradients(zip(gradients, variables), tf.train.get_or_create_global_step())\n", + " \n", + " if batch % 100 == 0:\n", + " print ('Epoch {} Batch {} Loss {:.4f}'.format(epoch + 1, \n", + " batch, \n", + " loss.numpy() / int(target.shape[1])))\n", + " # storing the epoch end loss value to plot later\n", + " loss_plot.append(total_loss / len(cap_vector))\n", + " \n", + " print ('Epoch {} Loss {:.6f}'.format(epoch + 1, \n", + " total_loss/len(cap_vector)))\n", + " print ('Time taken for 1 epoch {} sec\\n'.format(time.time() - start))" + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "1Wm83G-ZBPcC", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + } + }, + "cell_type": "code", + "source": [ + "plt.plot(loss_plot)\n", + "plt.xlabel('Epochs')\n", + "plt.ylabel('Loss')\n", + "plt.title('Loss Plot')\n", + "plt.show()" + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "xGvOcLQKghXN", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "## Caption!\n", + "\n", + "* The evaluate function is similar to the training loop, except we don't use teacher forcing here. The input to the decoder at each time step is its previous predictions along with the hidden state and the encoder output.\n", + "* Stop predicting when the model predicts the end token.\n", + "* And store the attention weights for every time step." + ] + }, + { + "metadata": { + "id": "RCWpDtyNRPGs", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + } + }, + "cell_type": "code", + "source": [ + "def evaluate(image):\n", + " attention_plot = np.zeros((max_length, attention_features_shape))\n", + "\n", + " hidden = decoder.reset_state(batch_size=1)\n", + "\n", + " temp_input = tf.expand_dims(load_image(image)[0], 0)\n", + " img_tensor_val = image_features_extract_model(temp_input)\n", + " img_tensor_val = tf.reshape(img_tensor_val, (img_tensor_val.shape[0], -1, img_tensor_val.shape[3]))\n", + "\n", + " features = encoder(img_tensor_val)\n", + "\n", + " dec_input = tf.expand_dims([tokenizer.word_index['']], 0)\n", + " result = []\n", + "\n", + " for i in range(max_length):\n", + " predictions, hidden, attention_weights = decoder(dec_input, features, hidden)\n", + "\n", + " attention_plot[i] = tf.reshape(attention_weights, (-1, )).numpy()\n", + "\n", + " predicted_id = tf.multinomial(tf.exp(predictions), num_samples=1)[0][0].numpy()\n", + " result.append(index_word[predicted_id])\n", + "\n", + " if index_word[predicted_id] == '':\n", + " return result, attention_plot\n", + "\n", + " dec_input = tf.expand_dims([predicted_id], 0)\n", + "\n", + " attention_plot = attention_plot[:len(result), :]\n", + " return result, attention_plot" + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "fD_y7PD6RPGt", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + } + }, + "cell_type": "code", + "source": [ + "def plot_attention(image, result, attention_plot):\n", + " temp_image = np.array(Image.open(image))\n", + "\n", + " fig = plt.figure(figsize=(10, 10))\n", + " \n", + " len_result = len(result)\n", + " for l in range(len_result):\n", + " temp_att = np.resize(attention_plot[l], (8, 8))\n", + " ax = fig.add_subplot(len_result//2, len_result//2, l+1)\n", + " ax.set_title(result[l])\n", + " img = ax.imshow(temp_image)\n", + " ax.imshow(temp_att, cmap='gray', alpha=0.6, extent=img.get_extent())\n", + "\n", + " plt.tight_layout()\n", + " plt.show()" + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "io7ws3ReRPGv", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + } + }, + "cell_type": "code", + "source": [ + "# captions on the validation set\n", + "rid = np.random.randint(0, len(img_name_val))\n", + "image = img_name_val[rid]\n", + "real_caption = ' '.join([index_word[i] for i in cap_val[rid] if i not in [0]])\n", + "result, attention_plot = evaluate(image)\n", + "\n", + "print ('Real Caption:', real_caption)\n", + "print ('Prediction Caption:', ' '.join(result))\n", + "plot_attention(image, result, attention_plot)\n", + "# opening the image\n", + "Image.open(img_name_val[rid])" + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "Rprk3HEvZuxb", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "## Try it on your own images\n", + "For fun, below we've provided a method you can use to caption your own images with the model we've just trained. Keep in mind, it was trained on a relatively small amount of data, and your images may be different from the training data (so be prepared for weird results!)\n" + ] + }, + { + "metadata": { + "id": "9Psd1quzaAWg", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + } + }, + "cell_type": "code", + "source": [ + "image_url = 'https://tensorflow.org/images/surf.jpg'\n", + "image_extension = image_url[-4:]\n", + "image_path = tf.keras.utils.get_file('image'+image_extension, \n", + " origin=image_url)\n", + "\n", + "result, attention_plot = evaluate(image_path)\n", + "print ('Prediction Caption:', ' '.join(result))\n", + "plot_attention(image_path, result, attention_plot)\n", + "# opening the image\n", + "Image.open(image_path)" + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "VJZXyJco6uLO", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "# Next steps\n", + "\n", + "Congrats! You've just trained an image captioning model with attention. Next, we recommend taking a look at this example [Neural Machine Translation with Attention]( https://github.com/tensorflow/tensorflow/blob/master/tensorflow/contrib/eager/python/examples/nmt_with_attention/nmt_with_attention.ipynb). It uses a similar architecture to translate between Spanish and English sentences. You can also experiment with training the code in this notebook on a different dataset." + ] + } + ] +} diff --git a/tensorflow/contrib/eager/python/examples/generative_examples/text_generation.ipynb b/tensorflow/contrib/eager/python/examples/generative_examples/text_generation.ipynb new file mode 100644 index 0000000000000000000000000000000000000000..6be09f98dff6627d9bdcc8056eed14e2a621be4b --- /dev/null +++ b/tensorflow/contrib/eager/python/examples/generative_examples/text_generation.ipynb @@ -0,0 +1,689 @@ +{ + "nbformat": 4, + "nbformat_minor": 0, + "metadata": { + "colab": { + "name": "text_generation.ipynb", + "version": "0.3.2", + "views": {}, + "default_view": {}, + "provenance": [], + "private_outputs": true, + "collapsed_sections": [], + "toc_visible": true + }, + "kernelspec": { + "display_name": "Python 3", + "language": "python", + "name": "python3" + }, + "accelerator": "GPU" + }, + "cells": [ + { + "metadata": { + "id": "hcD2nPQvPOFM", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "##### Copyright 2018 The TensorFlow Authors.\n", + "\n", + "Licensed under the Apache License, Version 2.0 (the \"License\").\n", + "\n", + "# Text Generation using a RNN\n", + "\n", + "
\n", + "\n", + " Run in Google Colab \n", + "\n", + "View source on Github
" + ] + }, + { + "metadata": { + "id": "BwpJ5IffzRG6", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "This notebook demonstrates how to generate text using an RNN using [tf.keras](https://www.tensorflow.org/programmers_guide/keras) and [eager execution](https://www.tensorflow.org/programmers_guide/eager). If you like, you can write a similar [model](https://github.com/fchollet/deep-learning-with-python-notebooks/blob/master/8.1-text-generation-with-lstm.ipynb) using less code. Here, we show a lower-level impementation that's useful to understand as prework before diving in to deeper examples in a similar, like [Neural Machine Translation with Attention](https://colab.research.google.com/github/tensorflow/tensorflow/blob/master/tensorflow/contrib/eager/python/examples/nmt_with_attention/nmt_with_attention.ipynb).\n", + "\n", + "This notebook is an end-to-end example. When you run it, it will download a dataset of Shakespeare's writing. We'll use a collection of plays, borrowed from Andrej Karpathy's excellent [The Unreasonable Effectiveness of Recurrent Neural Networks](http://karpathy.github.io/2015/05/21/rnn-effectiveness/). The notebook will train a model, and use it to generate sample output.\n", + " \n", + "Here is the output(with start string='w') after training a single layer GRU for 30 epochs with the default settings below:\n", + "\n", + "```\n", + "were to the death of him\n", + "And nothing of the field in the view of hell,\n", + "When I said, banish him, I will not burn thee that would live.\n", + "\n", + "HENRY BOLINGBROKE:\n", + "My gracious uncle--\n", + "\n", + "DUKE OF YORK:\n", + "As much disgraced to the court, the gods them speak,\n", + "And now in peace himself excuse thee in the world.\n", + "\n", + "HORTENSIO:\n", + "Madam, 'tis not the cause of the counterfeit of the earth,\n", + "And leave me to the sun that set them on the earth\n", + "And leave the world and are revenged for thee.\n", + "\n", + "GLOUCESTER:\n", + "I would they were talking with the very name of means\n", + "To make a puppet of a guest, and therefore, good Grumio,\n", + "Nor arm'd to prison, o' the clouds, of the whole field,\n", + "With the admire\n", + "With the feeding of thy chair, and we have heard it so,\n", + "I thank you, sir, he is a visor friendship with your silly your bed.\n", + "\n", + "SAMPSON:\n", + "I do desire to live, I pray: some stand of the minds, make thee remedies\n", + "With the enemies of my soul.\n", + "\n", + "MENENIUS:\n", + "I'll keep the cause of my mistress.\n", + "\n", + "POLIXENES:\n", + "My brother Marcius!\n", + "\n", + "Second Servant:\n", + "Will't ple\n", + "```\n", + "\n", + "Of course, while some of the sentences are grammatical, most do not make sense. But, consider:\n", + "\n", + "* Our model is character based (when we began training, it did not yet know how to spell a valid English word, or that words were even a unit of text).\n", + "\n", + "* The structure of the output resembles a play (blocks begin with a speaker name, in all caps similar to the original text). Sentences generally end with a period. If you look at the text from a distance (or don't read the invididual words too closely, it appears as if it's an excerpt from a play).\n", + "\n", + "As a next step, you can experiment training the model on a different dataset - any large text file(ASCII) will do, and you can modify a single line of code below to make that change. Have fun!\n" + ] + }, + { + "metadata": { + "id": "R3p22DBDsaCA", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "## Install unidecode library\n", + "A helpful library to convert unicode to ASCII." + ] + }, + { + "metadata": { + "id": "wZ6LOM12wKGH", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + } + }, + "cell_type": "code", + "source": [ + "!pip install unidecode" + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "WGyKZj3bzf9p", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "## Import tensorflow and enable eager execution." + ] + }, + { + "metadata": { + "id": "yG_n40gFzf9s", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + } + }, + "cell_type": "code", + "source": [ + "# Import TensorFlow >= 1.9 and enable eager execution\n", + "import tensorflow as tf\n", + "\n", + "# Note: Once you enable eager execution, it cannot be disabled. \n", + "tf.enable_eager_execution()\n", + "\n", + "import numpy as np\n", + "import re\n", + "import random\n", + "import unidecode\n", + "import time" + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "EHDoRoc5PKWz", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "## Download the dataset\n", + "\n", + "In this example, we will use the [shakespeare dataset](https://raw.githubusercontent.com/karpathy/char-rnn/master/data/tinyshakespeare/input.txt). You can use any other dataset that you like.\n", + "\n" + ] + }, + { + "metadata": { + "id": "pD_55cOxLkAb", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + } + }, + "cell_type": "code", + "source": [ + "path_to_file = tf.keras.utils.get_file('shakespeare.txt', 'https://storage.googleapis.com/yashkatariya/shakespeare.txt')" + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "UHjdCjDuSvX_", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "## Read the dataset\n", + "\n" + ] + }, + { + "metadata": { + "id": "-E5JvY3wzf94", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + } + }, + "cell_type": "code", + "source": [ + "text = unidecode.unidecode(open(path_to_file).read())\n", + "# length of text is the number of characters in it\n", + "print (len(text))" + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "Il9ww98izf-D", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "Creating dictionaries to map from characters to their indices and vice-versa, which will be used to vectorize the inputs" + ] + }, + { + "metadata": { + "id": "IalZLbvOzf-F", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + } + }, + "cell_type": "code", + "source": [ + "# unique contains all the unique characters in the file\n", + "unique = sorted(set(text))\n", + "\n", + "# creating a mapping from unique characters to indices\n", + "char2idx = {u:i for i, u in enumerate(unique)}\n", + "idx2char = {i:u for i, u in enumerate(unique)}" + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "1v_qUYfAzf-I", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + } + }, + "cell_type": "code", + "source": [ + "# setting the maximum length sentence we want for a single input in characters\n", + "max_length = 100\n", + "\n", + "# length of the vocabulary in chars\n", + "vocab_size = len(unique)\n", + "\n", + "# the embedding dimension \n", + "embedding_dim = 256\n", + "\n", + "# number of RNN (here GRU) units\n", + "units = 1024\n", + "\n", + "# batch size \n", + "BATCH_SIZE = 64\n", + "\n", + "# buffer size to shuffle our dataset\n", + "BUFFER_SIZE = 10000" + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "LFjSVAlWzf-N", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "## Creating the input and output tensors\n", + "\n", + "Vectorizing the input and the target text because our model cannot understand strings only numbers.\n", + "\n", + "But first, we need to create the input and output vectors.\n", + "Remember the max_length we set above, we will use it here. We are creating **max_length** chunks of input, where each input vector is all the characters in that chunk except the last and the target vector is all the characters in that chunk except the first.\n", + "\n", + "For example, consider that the string = 'tensorflow' and the max_length is 9\n", + "\n", + "So, the `input = 'tensorflo'` and `output = 'ensorflow'`\n", + "\n", + "After creating the vectors, we convert each character into numbers using the **char2idx** dictionary we created above." + ] + }, + { + "metadata": { + "id": "0UHJDA39zf-O", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + } + }, + "cell_type": "code", + "source": [ + "input_text = []\n", + "target_text = []\n", + "\n", + "for f in range(0, len(text)-max_length, max_length):\n", + " inps = text[f:f+max_length]\n", + " targ = text[f+1:f+1+max_length]\n", + "\n", + " input_text.append([char2idx[i] for i in inps])\n", + " target_text.append([char2idx[t] for t in targ])\n", + " \n", + "print (np.array(input_text).shape)\n", + "print (np.array(target_text).shape)" + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "MJdfPmdqzf-R", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "## Creating batches and shuffling them using tf.data" + ] + }, + { + "metadata": { + "id": "p2pGotuNzf-S", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + } + }, + "cell_type": "code", + "source": [ + "dataset = tf.data.Dataset.from_tensor_slices((input_text, target_text)).shuffle(BUFFER_SIZE)\n", + "dataset = dataset.apply(tf.contrib.data.batch_and_drop_remainder(BATCH_SIZE))" + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "m8gPwEjRzf-Z", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "## Creating the model\n", + "\n", + "We use the Model Subclassing API which gives us full flexibility to create the model and change it however we like. We use 3 layers to define our model.\n", + "\n", + "* Embedding layer\n", + "* GRU layer (you can use an LSTM layer here)\n", + "* Fully connected layer" + ] + }, + { + "metadata": { + "id": "P3KTiiInzf-a", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + } + }, + "cell_type": "code", + "source": [ + "class Model(tf.keras.Model):\n", + " def __init__(self, vocab_size, embedding_dim, units, batch_size):\n", + " super(Model, self).__init__()\n", + " self.units = units\n", + " self.batch_sz = batch_size\n", + "\n", + " self.embedding = tf.keras.layers.Embedding(vocab_size, embedding_dim)\n", + "\n", + " if tf.test.is_gpu_available():\n", + " self.gru = tf.keras.layers.CuDNNGRU(self.units, \n", + " return_sequences=True, \n", + " return_state=True, \n", + " recurrent_initializer='glorot_uniform')\n", + " else:\n", + " self.gru = tf.keras.layers.GRU(self.units, \n", + " return_sequences=True, \n", + " return_state=True, \n", + " recurrent_activation='sigmoid', \n", + " recurrent_initializer='glorot_uniform')\n", + "\n", + " self.fc = tf.keras.layers.Dense(vocab_size)\n", + " \n", + " def call(self, x, hidden):\n", + " x = self.embedding(x)\n", + "\n", + " # output shape == (batch_size, max_length, hidden_size) \n", + " # states shape == (batch_size, hidden_size)\n", + "\n", + " # states variable to preserve the state of the model\n", + " # this will be used to pass at every step to the model while training\n", + " output, states = self.gru(x, initial_state=hidden)\n", + "\n", + "\n", + " # reshaping the output so that we can pass it to the Dense layer\n", + " # after reshaping the shape is (batch_size * max_length, hidden_size)\n", + " output = tf.reshape(output, (-1, output.shape[2]))\n", + "\n", + " # The dense layer will output predictions for every time_steps(max_length)\n", + " # output shape after the dense layer == (max_length * batch_size, vocab_size)\n", + " x = self.fc(output)\n", + "\n", + " return x, states" + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "trpqTWyvk0nr", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "## Call the model and set the optimizer and the loss function" + ] + }, + { + "metadata": { + "id": "7t2XrzEOzf-e", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + } + }, + "cell_type": "code", + "source": [ + "model = Model(vocab_size, embedding_dim, units, BATCH_SIZE)" + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "dkjWIATszf-h", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + } + }, + "cell_type": "code", + "source": [ + "optimizer = tf.train.AdamOptimizer()\n", + "\n", + "# using sparse_softmax_cross_entropy so that we don't have to create one-hot vectors\n", + "def loss_function(real, preds):\n", + " return tf.losses.sparse_softmax_cross_entropy(labels=real, logits=preds)" + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "lPrP0XMUzf-p", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "## Train the model\n", + "\n", + "Here we will use a custom training loop with the help of GradientTape()\n", + "\n", + "* We initialize the hidden state of the model with zeros and shape == (batch_size, number of rnn units). We do this by calling the function defined while creating the model.\n", + "\n", + "* Next, we iterate over the dataset(batch by batch) and calculate the **predictions and the hidden states** associated with that input.\n", + "\n", + "* There are a lot of interesting things happening here.\n", + " * The model gets hidden state(initialized with 0), lets call that **H0** and the first batch of input, lets call that **I0**.\n", + " * The model then returns the predictions **P1** and **H1**.\n", + " * For the next batch of input, the model receives **I1** and **H1**.\n", + " * The interesting thing here is that we pass **H1** to the model with **I1** which is how the model learns. The context learned from batch to batch is contained in the **hidden state**.\n", + " * We continue doing this until the dataset is exhausted and then we start a new epoch and repeat this.\n", + "\n", + "* After calculating the predictions, we calculate the **loss** using the loss function defined above. Then we calculate the gradients of the loss with respect to the model variables(input)\n", + "\n", + "* Finally, we take a step in that direction with the help of the optimizer using the apply_gradients function.\n", + "\n", + "Note:- If you are running this notebook in Colab which has a **Tesla K80 GPU** it takes about 23 seconds per epoch.\n" + ] + }, + { + "metadata": { + "id": "d4tSNwymzf-q", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + } + }, + "cell_type": "code", + "source": [ + "# Training step\n", + "\n", + "EPOCHS = 30\n", + "\n", + "for epoch in range(EPOCHS):\n", + " start = time.time()\n", + " \n", + " # initializing the hidden state at the start of every epoch\n", + " hidden = model.reset_states()\n", + " \n", + " for (batch, (inp, target)) in enumerate(dataset):\n", + " with tf.GradientTape() as tape:\n", + " # feeding the hidden state back into the model\n", + " # This is the interesting step\n", + " predictions, hidden = model(inp, hidden)\n", + " \n", + " # reshaping the target because that's how the \n", + " # loss function expects it\n", + " target = tf.reshape(target, (-1,))\n", + " loss = loss_function(target, predictions)\n", + " \n", + " grads = tape.gradient(loss, model.variables)\n", + " optimizer.apply_gradients(zip(grads, model.variables), global_step=tf.train.get_or_create_global_step())\n", + "\n", + " if batch % 100 == 0:\n", + " print ('Epoch {} Batch {} Loss {:.4f}'.format(epoch+1,\n", + " batch,\n", + " loss))\n", + " \n", + " print ('Epoch {} Loss {:.4f}'.format(epoch+1, loss))\n", + " print('Time taken for 1 epoch {} sec\\n'.format(time.time() - start))" + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "DjGz1tDkzf-u", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "## Predicting using our trained model\n", + "\n", + "The below code block is used to generated the text\n", + "\n", + "* We start by choosing a start string and initializing the hidden state and setting the number of characters we want to generate.\n", + "\n", + "* We get predictions using the start_string and the hidden state\n", + "\n", + "* Then we use a multinomial distribution to calculate the index of the predicted word. **We use this predicted word as our next input to the model**\n", + "\n", + "* **The hidden state returned by the model is fed back into the model so that it now has more context rather than just one word.** After we predict the next word, the modified hidden states are again fed back into the model, which is how it learns as it gets more context from the previously predicted words.\n", + "\n", + "* If you see the predictions, the model knows when to capitalize, make paragraphs and the text follows a shakespeare style of writing which is pretty awesome!" + ] + }, + { + "metadata": { + "id": "WvuwZBX5Ogfd", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + } + }, + "cell_type": "code", + "source": [ + "# Evaluation step(generating text using the model learned)\n", + "\n", + "# number of characters to generate\n", + "num_generate = 1000\n", + "\n", + "# You can change the start string to experiment\n", + "start_string = 'Q'\n", + "# converting our start string to numbers(vectorizing!) \n", + "input_eval = [char2idx[s] for s in start_string]\n", + "input_eval = tf.expand_dims(input_eval, 0)\n", + "\n", + "# empty string to store our results\n", + "text_generated = ''\n", + "\n", + "# low temperatures results in more predictable text.\n", + "# higher temperatures results in more surprising text\n", + "# experiment to find the best setting\n", + "temperature = 1.0\n", + "\n", + "# hidden state shape == (batch_size, number of rnn units); here batch size == 1\n", + "hidden = [tf.zeros((1, units))]\n", + "for i in range(num_generate):\n", + " predictions, hidden = model(input_eval, hidden)\n", + "\n", + " # using a multinomial distribution to predict the word returned by the model\n", + " predictions = predictions / temperature\n", + " predicted_id = tf.multinomial(tf.exp(predictions), num_samples=1)[0][0].numpy()\n", + " \n", + " # We pass the predicted word as the next input to the model\n", + " # along with the previous hidden state\n", + " input_eval = tf.expand_dims([predicted_id], 0)\n", + " \n", + " text_generated += idx2char[predicted_id]\n", + "\n", + "print (start_string + text_generated)" + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "AM2Uma_-yVIq", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "## Next steps\n", + "\n", + "* Change the start string to a different character, or the start of a sentence.\n", + "* Experiment with training on a different, or with different parameters. [Project Gutenberg](http://www.gutenberg.org/ebooks/100), for example, contains a large collection of books.\n", + "* Experiment with the temperature parameter.\n", + "* Add another RNN layer.\n" + ] + }, + { + "metadata": { + "id": "gtEd86sX5cB2", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + } + }, + "cell_type": "code", + "source": [ + "" + ], + "execution_count": 0, + "outputs": [] + } + ] +} diff --git a/tensorflow/contrib/eager/python/examples/nmt_with_attention/nmt_with_attention.ipynb b/tensorflow/contrib/eager/python/examples/nmt_with_attention/nmt_with_attention.ipynb index 54ebcad8e929c3195099121a290dd7c0651e5c9f..1f66d7e75299df0c7db9bc8ec67cb6c0b5d4de40 100644 --- a/tensorflow/contrib/eager/python/examples/nmt_with_attention/nmt_with_attention.ipynb +++ b/tensorflow/contrib/eager/python/examples/nmt_with_attention/nmt_with_attention.ipynb @@ -41,11 +41,11 @@ "\n", "# Neural Machine Translation with Attention\n", "\n", - "
\n", - "\n", + "
\n", + "\n", " Run in Google Colab \n", "\n", - "View source on Github
" + "View source on GitHub
" ] }, { diff --git a/tensorflow/contrib/eager/python/examples/notebooks/2_gradients.ipynb b/tensorflow/contrib/eager/python/examples/notebooks/2_gradients.ipynb deleted file mode 100644 index 9c1af9c2084bac7ae6369babeaa13720e6199097..0000000000000000000000000000000000000000 --- a/tensorflow/contrib/eager/python/examples/notebooks/2_gradients.ipynb +++ /dev/null @@ -1,323 +0,0 @@ -{ - "cells": [ - { - "cell_type": "markdown", - "metadata": { - "colab_type": "text", - "id": "vDJ4XzMqodTy" - }, - "source": [ - "# Automatic Differentiation\n", - "\n", - "In the previous tutorial we introduced `Tensor`s and operations on them. In this tutorial we will cover [automatic differentiation](https://en.wikipedia.org/wiki/Automatic_differentiation), a key technique for optimizing machine learning models." - ] - }, - { - "cell_type": "markdown", - "metadata": { - "colab_type": "text", - "id": "GQJysDM__Qb0" - }, - "source": [ - "## Setup\n" - ] - }, - { - "cell_type": "code", - "execution_count": 0, - "metadata": { - "colab": { - "autoexec": { - "startup": false, - "wait_interval": 0 - } - }, - "colab_type": "code", - "id": "OiMPZStlibBv" - }, - "outputs": [], - "source": [ - "import tensorflow as tf\n", - "tf.enable_eager_execution()\n", - "\n", - "tfe = tf.contrib.eager # Shorthand for some symbols" - ] - }, - { - "cell_type": "markdown", - "metadata": { - "colab_type": "text", - "id": "1CLWJl0QliB0" - }, - "source": [ - "## Derivatives of a function\n", - "\n", - "TensorFlow provides APIs for automatic differentiation - computing the derivative of a function. The way that more closely mimics the math is to encapsulate the computation in a Python function, say `f`, and use `tfe.gradients_function` to create a function that computes the derivatives of `f` with respect to its arguments. If you're familiar with [autograd](https://github.com/HIPS/autograd) for differentiating numpy functions, this will be familiar. For example: " - ] - }, - { - "cell_type": "code", - "execution_count": 0, - "metadata": { - "colab": { - "autoexec": { - "startup": false, - "wait_interval": 0 - } - }, - "colab_type": "code", - "id": "9FViq92UX7P8" - }, - "outputs": [], - "source": [ - "from math import pi\n", - "\n", - "def f(x):\n", - " return tf.square(tf.sin(x))\n", - "\n", - "assert f(pi/2).numpy() == 1.0\n", - "\n", - "\n", - "# grad_f will return a list of derivatives of f\n", - "# with respect to its arguments. Since f() has a single argument,\n", - "# grad_f will return a list with a single element.\n", - "grad_f = tfe.gradients_function(f)\n", - "assert tf.abs(grad_f(pi/2)[0]).numpy() \u003c 1e-7" - ] - }, - { - "cell_type": "markdown", - "metadata": { - "colab_type": "text", - "id": "v9fPs8RyopCf" - }, - "source": [ - "### Higher-order gradients\n", - "\n", - "The same API can be used to differentiate as many times as you like:\n" - ] - }, - { - "cell_type": "code", - "execution_count": 0, - "metadata": { - "colab": { - "autoexec": { - "startup": false, - "wait_interval": 0 - }, - "height": 276 - }, - "colab_type": "code", - "executionInfo": { - "elapsed": 730, - "status": "ok", - "timestamp": 1527005655565, - "user": { - "displayName": "", - "photoUrl": "", - "userId": "" - }, - "user_tz": 420 - }, - "id": "3D0ZvnGYo0rW", - "outputId": "e23f8cc6-6813-4944-f20f-825b8a03c2ff" - }, - "outputs": [ - { - "data": { - "image/png": "iVBORw0KGgoAAAANSUhEUgAAAXYAAAEDCAYAAAAhsS8XAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nF3bGGIswXNgZYyzC/Be68EGj7hfMcwAAAABJRU5ErkJggg==\n", - "text/plain": [ - "\u003cmatplotlib.figure.Figure at 0x7f385e198650\u003e" - ] - }, - "metadata": { - "tags": [] - }, - "output_type": "display_data" - } - ], - "source": [ - "def f(x):\n", - " return tf.square(tf.sin(x))\n", - "\n", - "def grad(f):\n", - " return lambda x: tfe.gradients_function(f)(x)[0]\n", - "\n", - "x = tf.lin_space(-2*pi, 2*pi, 100) # 100 points between -2Ļ€ and +2Ļ€\n", - "\n", - "import matplotlib.pyplot as plt\n", - "\n", - "plt.plot(x, f(x), label=\"f\")\n", - "plt.plot(x, grad(f)(x), label=\"first derivative\")\n", - "plt.plot(x, grad(grad(f))(x), label=\"second derivative\")\n", - "plt.plot(x, grad(grad(grad(f)))(x), label=\"third derivative\")\n", - "plt.legend()\n", - "plt.show()" - ] - }, - { - "cell_type": "markdown", - "metadata": { - "colab_type": "text", - "id": "-39gouo7mtgu" - }, - "source": [ - "## Gradient tapes\n", - "\n", - "Every differentiable TensorFlow operation has an associated gradient function. For example, the gradient function of `tf.square(x)` would be a function that returns `2.0 * x`. To compute the gradient of a user-defined function (like `f(x)` in the example above), TensorFlow first \"records\" all the operations applied to compute the output of the function. We call this record a \"tape\". It then uses that tape and the gradients functions associated with each primitive operation to compute the gradients of the user-defined function using [reverse mode differentiation](https://en.wikipedia.org/wiki/Automatic_differentiation).\n", - "\n", - "Since operations are recorded as they are executed, Python control flow (using `if`s and `while`s for example) is naturally handled:\n", - "\n" - ] - }, - { - "cell_type": "code", - "execution_count": 0, - "metadata": { - "colab": { - "autoexec": { - "startup": false, - "wait_interval": 0 - } - }, - "colab_type": "code", - "id": "MH0UfjympWf7" - }, - "outputs": [], - "source": [ - "def f(x, y):\n", - " output = 1\n", - " for i in range(y):\n", - " output = tf.multiply(output, x)\n", - " return output\n", - "\n", - "def g(x, y):\n", - " # Return the gradient of `f` with respect to it's first parameter\n", - " return tfe.gradients_function(f)(x, y)[0]\n", - "\n", - "assert f(3.0, 2).numpy() == 9.0 # f(x, 2) is essentially x * x\n", - "assert g(3.0, 2).numpy() == 6.0 # And its gradient will be 2 * x\n", - "assert f(4.0, 3).numpy() == 64.0 # f(x, 3) is essentially x * x * x\n", - "assert g(4.0, 3).numpy() == 48.0 # And its gradient will be 3 * x * x" - ] - }, - { - "cell_type": "markdown", - "metadata": { - "colab_type": "text", - "id": "aNmR5-jhpX2t" - }, - "source": [ - "At times it may be inconvenient to encapsulate computation of interest into a function. For example, if you want the gradient of the output with respect to intermediate values computed in the function. In such cases, the slightly more verbose but explicit [tf.GradientTape](https://www.tensorflow.org/api_docs/python/tf/GradientTape) context is useful. All computation inside the context of a `tf.GradientTape` is \"recorded\".\n", - "\n", - "For example:" - ] - }, - { - "cell_type": "code", - "execution_count": 0, - "metadata": { - "colab": { - "autoexec": { - "startup": false, - "wait_interval": 0 - } - }, - "colab_type": "code", - "id": "bAFeIE8EuVIq" - }, - "outputs": [], - "source": [ - "x = tf.ones((2, 2))\n", - " \n", - "# TODO(b/78880779): Remove the 'persistent=True' argument and use\n", - "# a single t.gradient() call when the bug is resolved.\n", - "with tf.GradientTape(persistent=True) as t:\n", - " # TODO(ashankar): Explain with \"watch\" argument better?\n", - " t.watch(x)\n", - " y = tf.reduce_sum(x)\n", - " z = tf.multiply(y, y)\n", - "\n", - "# Use the same tape to compute the derivative of z with respect to the\n", - "# intermediate value y.\n", - "dz_dy = t.gradient(z, y)\n", - "assert dz_dy.numpy() == 8.0\n", - "\n", - "# Derivative of z with respect to the original input tensor x\n", - "dz_dx = t.gradient(z, x)\n", - "for i in [0, 1]:\n", - " for j in [0, 1]:\n", - " assert dz_dx[i][j].numpy() == 8.0" - ] - }, - { - "cell_type": "markdown", - "metadata": { - "colab_type": "text", - "id": "DK05KXrAAld3" - }, - "source": [ - "### Higher-order gradients\n", - "\n", - "Operations inside of the `GradientTape` context manager are recorded for automatic differentiation. If gradients are computed in that context, then the gradient computation is recorded as well. As a result, the exact same API works for higher-order gradients as well. For example:" - ] - }, - { - "cell_type": "code", - "execution_count": 0, - "metadata": { - "colab": { - "autoexec": { - "startup": false, - "wait_interval": 0 - } - }, - "colab_type": "code", - "id": "cPQgthZ7ugRJ" - }, - "outputs": [], - "source": [ - "# TODO(ashankar): Should we use the persistent tape here instead? Follow up on Tom and Alex's discussion\n", - "\n", - "x = tf.constant(1.0) # Convert the Python 1.0 to a Tensor object\n", - "\n", - "with tf.GradientTape() as t:\n", - " with tf.GradientTape() as t2:\n", - " t2.watch(x)\n", - " y = x * x * x\n", - " # Compute the gradient inside the 't' context manager\n", - " # which means the gradient computation is differentiable as well.\n", - " dy_dx = t2.gradient(y, x)\n", - "d2y_dx2 = t.gradient(dy_dx, x)\n", - "\n", - "assert dy_dx.numpy() == 3.0\n", - "assert d2y_dx2.numpy() == 6.0" - ] - }, - { - "cell_type": "markdown", - "metadata": { - "colab_type": "text", - "id": "4U1KKzUpNl58" - }, - "source": [ - "## Next Steps\n", - "\n", - "In this tutorial we covered gradient computation in TensorFlow. With that we have enough of the primitives required to build an train neural networks, which we will cover in the [next tutorial](https://github.com/tensorflow/models/tree/master/official/contrib/eager/python/examples/notebooks/3_neural_networks.ipynb)." - ] - } - ], - "metadata": { - "colab": { - "collapsed_sections": [], - "default_view": {}, - "name": "Automatic Differentiation", - "provenance": [], - "version": "0.3.2", - "views": {} - } - }, - "nbformat": 4, - "nbformat_minor": 0 -} diff --git a/tensorflow/contrib/eager/python/examples/notebooks/3_datasets.ipynb b/tensorflow/contrib/eager/python/examples/notebooks/3_datasets.ipynb deleted file mode 100644 index d268cbcd9171b0f4a4f2ab27ad958374e521685b..0000000000000000000000000000000000000000 --- a/tensorflow/contrib/eager/python/examples/notebooks/3_datasets.ipynb +++ /dev/null @@ -1,209 +0,0 @@ -{ - "cells": [ - { - "cell_type": "markdown", - "metadata": { - "colab_type": "text", - "id": "U9i2Dsh-ziXr" - }, - "source": [ - "# Eager Execution Tutorial: Importing Data\n", - "\n", - "This notebook demonstrates the use of the [`tf.data.Dataset` API](https://www.tensorflow.org/guide/datasets) to build pipelines to feed data to your program. It covers:\n", - "\n", - "* Creating a `Dataset`.\n", - "* Iteration over a `Dataset` with eager execution enabled.\n", - "\n", - "We recommend using the `Dataset`s API for building performant, complex input pipelines from simple, re-usable pieces that will feed your model's training or evaluation loops.\n", - "\n", - "If you're familiar with TensorFlow graphs, the API for constructing the `Dataset` object remains exactly the same when eager execution is enabled, but the process of iterating over elements of the dataset is slightly simpler.\n", - "You can use Python iteration over the `tf.data.Dataset` object and do not need to explicitly create an `tf.data.Iterator` object.\n", - "As a result, the discussion on iterators in the [TensorFlow Guide](https://www.tensorflow.org/guide/datasets) is not relevant when eager execution is enabled." - ] - }, - { - "cell_type": "markdown", - "metadata": { - "colab_type": "text", - "id": "z1JcS5iBXMRO" - }, - "source": [ - "# Setup: Enable eager execution\n" - ] - }, - { - "cell_type": "code", - "execution_count": 0, - "metadata": { - "cellView": "code", - "colab": { - "autoexec": { - "startup": false, - "wait_interval": 0 - } - }, - "colab_type": "code", - "id": "RlIWhyeLoYnG" - }, - "outputs": [], - "source": [ - "# Import TensorFlow.\n", - "import tensorflow as tf\n", - "\n", - "# Enable eager execution\n", - "tf.enable_eager_execution()" - ] - }, - { - "cell_type": "markdown", - "metadata": { - "colab_type": "text", - "id": "H9UySOPLXdaw" - }, - "source": [ - "# Step 1: Create a source `Dataset`\n", - "\n", - "Create a _source_ dataset using one of the factory functions like [`Dataset.from_tensors`](https://www.tensorflow.org/api_docs/python/tf/data/Dataset#from_tensors), [`Dataset.from_tensor_slices`](https://www.tensorflow.org/api_docs/python/tf/data/Dataset#from_tensor_slices) or using objects that read from files like [`TextLineDataset`](https://www.tensorflow.org/api_docs/python/tf/data/TextLineDataset) or [`TFRecordDataset`](https://www.tensorflow.org/api_docs/python/tf/data/TFRecordDataset). See the [TensorFlow Guide](https://www.tensorflow.org/guide/datasets#reading_input_data) for more information." - ] - }, - { - "cell_type": "code", - "execution_count": 0, - "metadata": { - "cellView": "code", - "colab": { - "autoexec": { - "startup": false, - "wait_interval": 0 - } - }, - "colab_type": "code", - "id": "WPTUfGq6kJ5w" - }, - "outputs": [], - "source": [ - "ds_tensors = tf.data.Dataset.from_tensor_slices([1, 2, 3, 4, 5, 6])\n", - "\n", - "# Create a CSV file\n", - "import tempfile\n", - "_, filename = tempfile.mkstemp()\n", - "with open(filename, 'w') as f:\n", - " f.write(\"\"\"Line 1\n", - "Line 2\n", - "Line 3\n", - " \"\"\")\n", - "ds_file = tf.data.TextLineDataset(filename)\n" - ] - }, - { - "cell_type": "markdown", - "metadata": { - "colab_type": "text", - "id": "twBfWd5xyu_d" - }, - "source": [ - "# Step 2: Apply transformations\n", - "\n", - "Use the transformations functions like [`map`](https://www.tensorflow.org/api_docs/python/tf/data/Dataset#map), [`batch`](https://www.tensorflow.org/api_docs/python/tf/data/Dataset#batch), [`shuffle`](https://www.tensorflow.org/api_docs/python/tf/data/Dataset#shuffle) etc. to apply transformations to the records of the dataset. See the [API documentation for `tf.data.Dataset`](https://www.tensorflow.org/api_docs/python/tf/data/Dataset) for details." - ] - }, - { - "cell_type": "code", - "execution_count": 0, - "metadata": { - "cellView": "code", - "colab": { - "autoexec": { - "startup": false, - "wait_interval": 0 - } - }, - "colab_type": "code", - "id": "ngUe237Wt48W" - }, - "outputs": [], - "source": [ - "ds_tensors = ds_tensors.map(tf.square).shuffle(2).batch(2)\n", - "ds_file = ds_file.batch(2)" - ] - }, - { - "cell_type": "markdown", - "metadata": { - "colab_type": "text", - "id": "IDY4WsYRhP81" - }, - "source": [ - "# Step 3: Iterate\n", - "\n", - "When eager execution is enabled `Dataset` objects support iteration.\n", - "If you're familiar with the use of `Dataset`s in TensorFlow graphs, note that there is no need for calls to `Dataset.make_one_shot_iterator()` or `get_next()` calls." - ] - }, - { - "cell_type": "code", - "execution_count": 0, - "metadata": { - "colab": { - "autoexec": { - "startup": false, - "wait_interval": 0 - }, - "base_uri": "https://localhost:8080/", - "height": 153 - }, - "colab_type": "code", - "executionInfo": { - "elapsed": 388, - "status": "ok", - "timestamp": 1525154629129, - "user": { - "displayName": "", - "photoUrl": "", - "userId": "" - }, - "user_tz": 420 - }, - "id": "lCUWzso6mbqR", - "outputId": "8e4b0298-d27d-4ac7-e26a-ef94af0594ec" - }, - "outputs": [ - { - "name": "stdout", - "output_type": "stream", - "text": [ - "Elements of ds_tensors:\n", - "tf.Tensor([1 9], shape=(2,), dtype=int32)\n", - "tf.Tensor([16 25], shape=(2,), dtype=int32)\n", - "tf.Tensor([ 4 36], shape=(2,), dtype=int32)\n", - "\n", - "Elements in ds_file:\n", - "tf.Tensor(['Line 1' 'Line 2'], shape=(2,), dtype=string)\n", - "tf.Tensor(['Line 3' ' '], shape=(2,), dtype=string)\n" - ] - } - ], - "source": [ - "print('Elements of ds_tensors:')\n", - "for x in ds_tensors:\n", - " print(x)\n", - "\n", - "print('\\nElements in ds_file:')\n", - "for x in ds_file:\n", - " print(x)" - ] - } - ], - "metadata": { - "colab": { - "collapsed_sections": [], - "default_view": {}, - "name": "Eager Execution Tutorial: Importing Data", - "provenance": [], - "version": "0.3.2", - "views": {} - } - }, - "nbformat": 4, - "nbformat_minor": 0 -} diff --git a/tensorflow/contrib/eager/python/examples/notebooks/3_training_models.ipynb b/tensorflow/contrib/eager/python/examples/notebooks/3_training_models.ipynb deleted file mode 100644 index 84f1d031d40604ae029e8a8347474950ee01b38a..0000000000000000000000000000000000000000 --- a/tensorflow/contrib/eager/python/examples/notebooks/3_training_models.ipynb +++ /dev/null @@ -1,485 +0,0 @@ -{ - "cells": [ - { - "cell_type": "markdown", - "metadata": { - "colab_type": "text", - "id": "k2o3TTG4TFpt" - }, - "source": [ - "# Training Models\n", - "\n", - "In the previous tutorial we covered the TensorFlow APIs for automatic differentiation, a basic building block for machine learning.\n", - "In this tutorial we will use the TensorFlow primitives introduced in the prior tutorials to do some simple machine learning.\n", - "\n", - "TensorFlow also includes a higher-level neural networks API (`tf.keras`) which provides useful abstractions to reduce boilerplate. We strongly recommend those higher level APIs for people working with neural networks. However, in this short tutorial we cover neural network training from first principles to establish a strong foundation." - ] - }, - { - "cell_type": "markdown", - "metadata": { - "colab_type": "text", - "id": "3LXMVuV0VhDr" - }, - "source": [ - "## Setup" - ] - }, - { - "cell_type": "code", - "execution_count": 0, - "metadata": { - "colab": { - "autoexec": { - "startup": false, - "wait_interval": 0 - } - }, - "colab_type": "code", - "id": "PJ64L90aVir3" - }, - "outputs": [], - "source": [ - "import tensorflow as tf\n", - "tf.enable_eager_execution()\n", - "tfe = tf.contrib.eager # Shorthand for some symbols" - ] - }, - { - "cell_type": "markdown", - "metadata": { - "colab_type": "text", - "id": "eMAWbDJFVmMk" - }, - "source": [ - "## Variables\n", - "\n", - "Tensors in TensorFlow are immutable stateless objects. Machine learning models, however, need to have changing state: as your model trains, the same code to compute predictions should behave differently over time (hopefully with a lower loss!). To represent this state which needs to change over the course of your computation, you can choose to rely on the fact that Python is a stateful programming language:\n" - ] - }, - { - "cell_type": "code", - "execution_count": 0, - "metadata": { - "colab": { - "autoexec": { - "startup": false, - "wait_interval": 0 - } - }, - "colab_type": "code", - "id": "VkJwtLS_Jbn8" - }, - "outputs": [], - "source": [ - "# Using python state\n", - "x = tf.zeros([10, 10])\n", - "x += 2 # This is equivalent to x = x + 2, which does not mutate the original\n", - " # value of x\n", - "print(x)" - ] - }, - { - "cell_type": "markdown", - "metadata": { - "colab_type": "text", - "id": "wfneTXy7JcUz" - }, - "source": [ - "TensorFlow, however, has stateful operations built in, and these are often more pleasant to use than low-level Python representations of your state. To represent weights in a model, for example, it's often convenient and efficient to use TensorFlow variables.\n", - "\n", - "A Variable is an object which stores a value and, when used in a TensorFlow computation, will implicitly read from this stored value. There are operations (`tf.assign_sub`, `tf.scatter_update`, etc) which manipulate the value stored in a TensorFlow variable." - ] - }, - { - "cell_type": "code", - "execution_count": 0, - "metadata": { - "colab": { - "autoexec": { - "startup": false, - "wait_interval": 0 - } - }, - "colab_type": "code", - "id": "itxmrMil6DQi" - }, - "outputs": [], - "source": [ - "v = tfe.Variable(1.0)\n", - "assert v.numpy() == 1.0\n", - "\n", - "# Re-assign the value\n", - "v.assign(3.0)\n", - "assert v.numpy() == 3.0\n", - "\n", - "# Use `v` in a TensorFlow operation like tf.square() and reassign\n", - "v.assign(tf.square(v))\n", - "assert v.numpy() == 9.0" - ] - }, - { - "cell_type": "markdown", - "metadata": { - "colab_type": "text", - "id": "-paSaeq1JzwC" - }, - "source": [ - "Computations using Variables are automatically traced when computing gradients. For Variables representing embeddings TensorFlow will do sparse updates by default, which are more computation and memory efficient.\n", - "\n", - "Using Variables is also a way to quickly let a reader of your code know that this piece of state is mutable." - ] - }, - { - "cell_type": "markdown", - "metadata": { - "colab_type": "text", - "id": "BMiFcDzE7Qu3" - }, - "source": [ - "## Example: Fitting a linear model\n", - "\n", - "Let's now put the few concepts we have so far ---`Tensor`, `GradientTape`, `Variable` --- to build and train a simple model. This typically involves a few steps:\n", - "\n", - "1. Define the model.\n", - "2. Define a loss function.\n", - "3. Obtain training data.\n", - "4. Run through the training data and use an \"optimizer\" to adjust the variables to fit the data.\n", - "\n", - "In this tutorial, we'll walk through a trivial example of a simple linear model: `f(x) = x * W + b`, which has two variables - `W` and `b`. Furthermore, we'll synthesize data such that a well trained model would have `W = 3.0` and `b = 2.0`." - ] - }, - { - "cell_type": "markdown", - "metadata": { - "colab_type": "text", - "id": "gFzH64Jn9PIm" - }, - "source": [ - "### Define the model\n", - "\n", - "Let's define a simple class to encapsulate the variables and the computation." - ] - }, - { - "cell_type": "code", - "execution_count": 0, - "metadata": { - "colab": { - "autoexec": { - "startup": false, - "wait_interval": 0 - } - }, - "colab_type": "code", - "id": "_WRu7Pze7wk8" - }, - "outputs": [], - "source": [ - "class Model(object):\n", - " def __init__(self):\n", - " # Initialize variable to (5.0, 0.0)\n", - " # In practice, these should be initialized to random values.\n", - " self.W = tfe.Variable(5.0)\n", - " self.b = tfe.Variable(0.0)\n", - " \n", - " def __call__(self, x):\n", - " return self.W * x + self.b\n", - " \n", - "model = Model()\n", - "\n", - "assert model(3.0).numpy() == 15.0" - ] - }, - { - "cell_type": "markdown", - "metadata": { - "colab_type": "text", - "id": "xa6j_yXa-j79" - }, - "source": [ - "### Define a loss function\n", - "\n", - "A loss function measures how well the output of a model for a given input matches the desired output. Let's use the standard L2 loss." - ] - }, - { - "cell_type": "code", - "execution_count": 0, - "metadata": { - "colab": { - "autoexec": { - "startup": false, - "wait_interval": 0 - } - }, - "colab_type": "code", - "id": "Y0ysUFGY924U" - }, - "outputs": [], - "source": [ - "def loss(predicted_y, desired_y):\n", - " return tf.reduce_mean(tf.square(predicted_y - desired_y))" - ] - }, - { - "cell_type": "markdown", - "metadata": { - "colab_type": "text", - "id": "qutT_fkl_CBc" - }, - "source": [ - "### Obtain training data\n", - "\n", - "Let's synthesize the training data with some noise." - ] - }, - { - "cell_type": "code", - "execution_count": 0, - "metadata": { - "colab": { - "autoexec": { - "startup": false, - "wait_interval": 0 - } - }, - "colab_type": "code", - "id": "gxPTb-kt_N5m" - }, - "outputs": [], - "source": [ - "TRUE_W = 3.0\n", - "TRUE_b = 2.0\n", - "NUM_EXAMPLES = 1000\n", - "\n", - "inputs = tf.random_normal(shape=[NUM_EXAMPLES])\n", - "noise = tf.random_normal(shape=[NUM_EXAMPLES])\n", - "outputs = inputs * TRUE_W + TRUE_b + noise" - ] - }, - { - "cell_type": "markdown", - "metadata": { - "colab_type": "text", - "id": "-50nq-wPBsAW" - }, - "source": [ - "Before we train the model let's visualize where the model stands right now. We'll plot the model's predictions in red and the training data in blue." - ] - }, - { - "cell_type": "code", - "execution_count": 0, - "metadata": { - "colab": { - "autoexec": { - "startup": false, - "wait_interval": 0 - }, - "height": 293 - }, - "colab_type": "code", - "executionInfo": { - "elapsed": 1210, - "status": "ok", - "timestamp": 1527005898290, - "user": { - "displayName": "", - "photoUrl": "", - "userId": "" - }, - "user_tz": 420 - }, - "id": "_eb83LtrB4nt", - "outputId": "3873f508-72fb-41e7-a7f5-3f513deefe38" - }, - "outputs": [ - { - "data": { - "image/png": "iVBORw0KGgoAAAANSUhEUgAAAXwAAAEDCAYAAAA2k7/eAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\noZkggi8IgtBMEMEXBEFoJvw//5K32R/vBHAAAAAASUVORK5CYII=\n", - "text/plain": [ - "\u003cmatplotlib.figure.Figure at 0x7f5be3c99f50\u003e" - ] - }, - "metadata": { - "tags": [] - }, - "output_type": "display_data" - }, - { - "name": "stdout", - "output_type": "stream", - "text": [ - "Current loss: 9.48636\n" - ] - } - ], - "source": [ - "import matplotlib.pyplot as plt\n", - "\n", - "plt.scatter(inputs, outputs, c='b')\n", - "plt.scatter(inputs, model(inputs), c='r')\n", - "plt.show()\n", - "\n", - "print('Current loss: '),\n", - "print(loss(model(inputs), outputs).numpy())" - ] - }, - { - "cell_type": "markdown", - "metadata": { - "colab_type": "text", - "id": "sSDP-yeq_4jE" - }, - "source": [ - "### Define a training loop\n", - "\n", - "We now have our network and our training data. Let's train it, i.e., use the training data to update the model's variables (`W` and `b`) so that the loss goes down using [gradient descent](https://en.wikipedia.org/wiki/Gradient_descent). There are many variants of the gradient descent scheme that are captured in `tf.train.Optimizer` implementations. We'd highly recommend using those implementations, but in the spirit of building from first principles, in this particular example we will implement the basic math ourselves." - ] - }, - { - "cell_type": "code", - "execution_count": 0, - "metadata": { - "colab": { - "autoexec": { - "startup": false, - "wait_interval": 0 - } - }, - "colab_type": "code", - "id": "MBIACgdnA55X" - }, - "outputs": [], - "source": [ - "def train(model, inputs, outputs, learning_rate):\n", - " with tf.GradientTape() as t:\n", - " current_loss = loss(model(inputs), outputs)\n", - " dW, db = t.gradient(current_loss, [model.W, model.b])\n", - " model.W.assign_sub(learning_rate * dW)\n", - " model.b.assign_sub(learning_rate * db)" - ] - }, - { - "cell_type": "markdown", - "metadata": { - "colab_type": "text", - "id": "RwWPaJryD2aN" - }, - "source": [ - "Finally, let's repeatedly run through the training data and see how `W` and `b` evolve." - ] - }, - { - "cell_type": "code", - "execution_count": 0, - "metadata": { - "colab": { - "autoexec": { - "startup": false, - "wait_interval": 0 - }, - "height": 446 - }, - "colab_type": "code", - "executionInfo": { - "elapsed": 569, - "status": "ok", - "timestamp": 1527005915434, - "user": { - "displayName": "", - "photoUrl": "", - "userId": "" - }, - "user_tz": 420 - }, - "id": "XdfkR223D9dW", - "outputId": "c43591ae-d5ac-4f2b-a8e7-bfce607e0919" - }, - "outputs": [ - { - "name": "stdout", - "output_type": "stream", - "text": [ - "Epoch 0: W=5.00 b=0.00, loss=9.48636\n", - "Epoch 1: W=4.58 b=0.42, loss=6.28101\n", - "Epoch 2: W=4.24 b=0.76, loss=4.29357\n", - "Epoch 3: W=3.98 b=1.02, loss=3.06128\n", - "Epoch 4: W=3.78 b=1.23, loss=2.29721\n", - "Epoch 5: W=3.61 b=1.39, loss=1.82345\n", - "Epoch 6: W=3.49 b=1.52, loss=1.52970\n", - "Epoch 7: W=3.38 b=1.62, loss=1.34756\n", - "Epoch 8: W=3.30 b=1.70, loss=1.23463\n", - "Epoch 9: W=3.24 b=1.76, loss=1.16460\n" - ] - }, - { - "data": { - "image/png": "iVBORw0KGgoAAAANSUhEUgAAAW0AAAEDCAYAAAD+/1UIAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\n", - "text/plain": [ - "\u003cmatplotlib.figure.Figure at 0x7f5be4b8ec50\u003e" - ] - }, - "metadata": { - "tags": [] - }, - "output_type": "display_data" - } - ], - "source": [ - "model = Model()\n", - "\n", - "# Collect the history of W-values and b-values to plot later\n", - "Ws, bs = [], []\n", - "epochs = range(10)\n", - "for epoch in epochs:\n", - " Ws.append(model.W.numpy())\n", - " bs.append(model.b.numpy())\n", - " current_loss = loss(model(inputs), outputs)\n", - "\n", - " train(model, inputs, outputs, learning_rate=0.1)\n", - " print('Epoch %2d: W=%1.2f b=%1.2f, loss=%2.5f' %\n", - " (epoch, Ws[-1], bs[-1], current_loss))\n", - "\n", - "# Let's plot it all\n", - "plt.plot(epochs, Ws, 'r',\n", - " epochs, bs, 'b')\n", - "plt.plot([TRUE_W] * len(epochs), 'r--',\n", - " [TRUE_b] * len(epochs), 'b--')\n", - "plt.legend(['W', 'b', 'true W', 'true_b'])\n", - "plt.show()\n", - " " - ] - }, - { - "cell_type": "markdown", - "metadata": { - "colab_type": "text", - "id": "vPnIVuaSJwWz" - }, - "source": [ - "## Next Steps\n", - "\n", - "In this tutorial we covered `Variable`s and built and trained a simple linear model using the TensorFlow primitives discussed so far.\n", - "\n", - "In theory, this is pretty much all you need to use TensorFlow for your machine learning research.\n", - "In practice, particularly for neural networks, the higher level APIs like `tf.keras` will be much more convenient since it provides higher level building blocks (called \"layers\"), utilities to save and restore state, a suite of loss functions, a suite of optimization strategies etc. \n", - "\n", - "The [next tutorial](TODO) will cover these higher level APIs." - ] - } - ], - "metadata": { - "colab": { - "collapsed_sections": [], - "default_view": {}, - "name": "Training Models", - "provenance": [], - "version": "0.3.2", - "views": {} - } - }, - "nbformat": 4, - "nbformat_minor": 0 -} diff --git a/tensorflow/contrib/eager/python/examples/notebooks/4_high_level.ipynb b/tensorflow/contrib/eager/python/examples/notebooks/4_high_level.ipynb deleted file mode 100644 index 5749f22ac58e0a012ed7e3fec4dfe2913d3f8273..0000000000000000000000000000000000000000 --- a/tensorflow/contrib/eager/python/examples/notebooks/4_high_level.ipynb +++ /dev/null @@ -1,551 +0,0 @@ -{ - "cells": [ - { - "cell_type": "code", - "execution_count": 0, - "metadata": { - "colab": { - "autoexec": { - "startup": false, - "wait_interval": 0 - } - }, - "colab_type": "code", - "id": "pwX7Fii1rwsJ" - }, - "outputs": [], - "source": [ - "import tensorflow as tf\n", - "tf.enable_eager_execution()\n", - "tfe = tf.contrib.eager\n" - ] - }, - { - "cell_type": "markdown", - "metadata": { - "colab_type": "text", - "id": "UEu3q4jmpKVT" - }, - "source": [ - "# High level API\n", - "\n", - "We recommend using `tf.keras` as a high-level API for building neural networks. That said, most TensorFlow APIs are usable with eager execution.\n", - "\n" - ] - }, - { - "cell_type": "markdown", - "metadata": { - "colab_type": "text", - "id": "zSFfVVjkrrsI" - }, - "source": [ - "## Layers: common sets of useful operations\n", - "\n", - "Most of the time when writing code for machine learning models you want to operate at a higher level of abstraction than individual operations and manipulation of individual variables.\n", - "\n", - "Many machine learning models are expressible as the composition and stacking of relatively simple layers, and TensorFlow provides both a set of many common layers as a well as easy ways for you to write your own application-specific layers either from scratch or as the composition of existing layers.\n", - "\n", - "TensorFlow includes the full [Keras](https://keras.io) API in the tf.keras package, and the Keras layers are very useful when building your own models.\n" - ] - }, - { - "cell_type": "code", - "execution_count": 0, - "metadata": { - "colab": { - "autoexec": { - "startup": false, - "wait_interval": 0 - } - }, - "colab_type": "code", - "id": "8PyXlPl-4TzQ" - }, - "outputs": [], - "source": [ - "# In the tf.keras.layers package, layers are objects. To construct a layer,\n", - "# simply construct the object. Most layers take as a first argument the number\n", - "# of output dimensions / channels.\n", - "layer = tf.keras.layers.Dense(100)\n", - "# The number of input dimensions is often unnecessary, as it can be inferred\n", - "# the first time the layer is used, but it can be provided if you want to \n", - "# specify it manually, which is useful in some complex models.\n", - "layer = tf.keras.layers.Dense(10, input_shape=(None, 5))" - ] - }, - { - "cell_type": "markdown", - "metadata": { - "colab_type": "text", - "id": "Fn69xxPO5Psr" - }, - "source": [ - "The full list of pre-existing layers can be seen in [the documentation](https://www.tensorflow.org/api_docs/python/tf/keras/layers). It includes Dense (a fully-connected layer),\n", - "Conv2D, LSTM, BatchNormalization, Dropout, and many others." - ] - }, - { - "cell_type": "code", - "execution_count": 3, - "metadata": { - "colab": { - "autoexec": { - "startup": false, - "wait_interval": 0 - }, - "height": 204 - }, - "colab_type": "code", - "executionInfo": { - "elapsed": 244, - "status": "ok", - "timestamp": 1527783641557, - "user": { - "displayName": "", - "photoUrl": "", - "userId": "" - }, - "user_tz": 420 - }, - "id": "E3XKNknP5Mhb", - "outputId": "c5d52434-d980-4488-efa7-5660819d0207" - }, - "outputs": [ - { - "data": { - "text/plain": [ - "\u003ctf.Tensor: id=30, shape=(10, 10), dtype=float32, numpy=\n", - "array([[ 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],\n", - " [ 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],\n", - " [ 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],\n", - " [ 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],\n", - " [ 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],\n", - " [ 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],\n", - " [ 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],\n", - " [ 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],\n", - " [ 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.],\n", - " [ 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.]], dtype=float32)\u003e" - ] - }, - "execution_count": 3, - "metadata": { - "tags": [] - }, - "output_type": "execute_result" - } - ], - "source": [ - "# To use a layer, simply call it.\n", - "layer(tf.zeros([10, 5]))" - ] - }, - { - "cell_type": "code", - "execution_count": 4, - "metadata": { - "colab": { - "autoexec": { - "startup": false, - "wait_interval": 0 - }, - "height": 221 - }, - "colab_type": "code", - "executionInfo": { - "elapsed": 320, - "status": "ok", - "timestamp": 1527783642457, - "user": { - "displayName": "", - "photoUrl": "", - "userId": "" - }, - "user_tz": 420 - }, - "id": "Wt_Nsv-L5t2s", - "outputId": "f0d96dce-0128-4080-bfe2-0ee6fbc0ad90" - }, - "outputs": [ - { - "data": { - "text/plain": [ - "[\u003ctf.Variable 'dense_1/kernel:0' shape=(5, 10) dtype=float32, numpy=\n", - " array([[ 0.43788117, -0.62099844, -0.30525017, -0.59352523, 0.1783089 ,\n", - " 0.47078604, -0.23620895, -0.30482283, 0.01366901, -0.1288507 ],\n", - " [ 0.18407935, -0.56550485, 0.54180616, -0.42254075, 0.3702994 ,\n", - " 0.36705834, -0.29678228, 0.36660975, 0.36717761, 0.46269661],\n", - " [ 0.1709305 , -0.11529458, 0.32710236, 0.46300393, -0.62802851,\n", - " 0.51641601, 0.39624029, 0.26918125, -0.25196898, 0.21353298],\n", - " [ 0.35752094, 0.44161648, 0.61500639, -0.12653333, 0.41629118,\n", - " 0.36193585, 0.066082 , -0.59253877, 0.47318751, 0.17115968],\n", - " [-0.22554061, -0.17727301, 0.5525015 , 0.3678053 , -0.00454676,\n", - " 0.24066836, -0.53640735, 0.13792562, -0.10727292, 0.59708995]], dtype=float32)\u003e,\n", - " \u003ctf.Variable 'dense_1/bias:0' shape=(10,) dtype=float32, numpy=array([ 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.], dtype=float32)\u003e]" - ] - }, - "execution_count": 4, - "metadata": { - "tags": [] - }, - "output_type": "execute_result" - } - ], - "source": [ - "# Layers have many useful methods. For example, you can inspect all variables\n", - "# in a layer by calling layer.variables. In this case a fully-connected layer\n", - "# will have variables for weights and biases.\n", - "layer.variables" - ] - }, - { - "cell_type": "code", - "execution_count": 5, - "metadata": { - "colab": { - "autoexec": { - "startup": false, - "wait_interval": 0 - }, - "height": 221 - }, - "colab_type": "code", - "executionInfo": { - "elapsed": 226, - "status": "ok", - "timestamp": 1527783643252, - "user": { - "displayName": "", - "photoUrl": "", - "userId": "" - }, - "user_tz": 420 - }, - "id": "6ilvKjz8_4MQ", - "outputId": "f647fced-c2d7-41a3-c237-242036784665" - }, - "outputs": [ - { - "data": { - "text/plain": [ - "(\u003ctf.Variable 'dense_1/kernel:0' shape=(5, 10) dtype=float32, numpy=\n", - " array([[ 0.43788117, -0.62099844, -0.30525017, -0.59352523, 0.1783089 ,\n", - " 0.47078604, -0.23620895, -0.30482283, 0.01366901, -0.1288507 ],\n", - " [ 0.18407935, -0.56550485, 0.54180616, -0.42254075, 0.3702994 ,\n", - " 0.36705834, -0.29678228, 0.36660975, 0.36717761, 0.46269661],\n", - " [ 0.1709305 , -0.11529458, 0.32710236, 0.46300393, -0.62802851,\n", - " 0.51641601, 0.39624029, 0.26918125, -0.25196898, 0.21353298],\n", - " [ 0.35752094, 0.44161648, 0.61500639, -0.12653333, 0.41629118,\n", - " 0.36193585, 0.066082 , -0.59253877, 0.47318751, 0.17115968],\n", - " [-0.22554061, -0.17727301, 0.5525015 , 0.3678053 , -0.00454676,\n", - " 0.24066836, -0.53640735, 0.13792562, -0.10727292, 0.59708995]], dtype=float32)\u003e,\n", - " \u003ctf.Variable 'dense_1/bias:0' shape=(10,) dtype=float32, numpy=array([ 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.], dtype=float32)\u003e)" - ] - }, - "execution_count": 5, - "metadata": { - "tags": [] - }, - "output_type": "execute_result" - } - ], - "source": [ - "# The variables are also accessible through nice accessors\n", - "layer.kernel, layer.bias" - ] - }, - { - "cell_type": "markdown", - "metadata": { - "colab_type": "text", - "id": "O0kDbE54-5VS" - }, - "source": [ - "## Implementing custom layers\n", - "The best way to implement your own layer is extending the tf.keras.Layer class and implementing:\n", - " * `__init__` , where you can do all input-independent initialization\n", - " * `build`, where you know the shapes of the input tensors and can do the rest of the initialization\n", - " * `call`, where you do the forward computation\n", - "\n", - "Note that you don't have to wait until `build` is called to create your variables, you can also create them in `__init__`. However, the advantage of creating them in `build` is that it enables late variable creation based on the shape of the inputs the layer will operate on. On the other hand, creating variables in `__init__` would mean that shapes required to create the variables will need to be explicitly specified." - ] - }, - { - "cell_type": "code", - "execution_count": 7, - "metadata": { - "colab": { - "autoexec": { - "startup": false, - "wait_interval": 0 - }, - "height": 391 - }, - "colab_type": "code", - "executionInfo": { - "elapsed": 251, - "status": "ok", - "timestamp": 1527783661512, - "user": { - "displayName": "", - "photoUrl": "", - "userId": "" - }, - "user_tz": 420 - }, - "id": "5Byl3n1k5kIy", - "outputId": "6e7f9285-649a-4132-82ce-73ea92f15862" - }, - "outputs": [ - { - "name": "stdout", - "output_type": "stream", - "text": [ - "tf.Tensor(\n", - "[[ 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]\n", - " [ 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]\n", - " [ 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]\n", - " [ 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]\n", - " [ 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]\n", - " [ 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]\n", - " [ 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]\n", - " [ 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]\n", - " [ 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]\n", - " [ 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]], shape=(10, 10), dtype=float32)\n", - "[\u003ctf.Variable 'my_dense_layer_1/kernel:0' shape=(5, 10) dtype=float32, numpy=\n", - "array([[-0.4011991 , 0.22458655, -0.33237562, -0.25117266, 0.33528614,\n", - " -0.01392961, 0.58580834, -0.16346583, 0.28465688, -0.47191954],\n", - " [-0.52922136, 0.22416979, -0.58209574, -0.60914612, 0.05226624,\n", - " -0.18325993, 0.5591442 , -0.24718609, 0.37148207, 0.40475875],\n", - " [ 0.16912812, -0.47618777, -0.38989353, 0.30105609, -0.08085585,\n", - " 0.44758242, 0.545829 , 0.51421839, 0.11063248, 0.20159996],\n", - " [ 0.34073615, -0.59835428, 0.06498981, -0.44489855, -0.34302285,\n", - " 0.20969599, 0.35527444, -0.03173476, -0.22227573, 0.09303057],\n", - " [ 0.41764337, -0.06435019, -0.52509922, -0.39957345, 0.56811184,\n", - " 0.23481232, -0.61666459, 0.31144124, -0.11532354, -0.42421889]], dtype=float32)\u003e]\n" - ] - } - ], - "source": [ - "class MyDenseLayer(tf.keras.layers.Layer):\n", - " def __init__(self, num_outputs):\n", - " super(MyDenseLayer, self).__init__()\n", - " self.num_outputs = num_outputs\n", - " \n", - " def build(self, input_shape):\n", - " self.kernel = self.add_variable(\"kernel\", \n", - " shape=[input_shape[-1].value, \n", - " self.num_outputs])\n", - " \n", - " def call(self, input):\n", - " return tf.matmul(input, self.kernel)\n", - " \n", - "layer = MyDenseLayer(10)\n", - "print(layer(tf.zeros([10, 5])))\n", - "print(layer.variables)" - ] - }, - { - "cell_type": "markdown", - "metadata": { - "colab_type": "text", - "id": "tk8E2vY0-z4Z" - }, - "source": [ - "Note that you don't have to wait until `build` is called to create your variables, you can also create them in `__init__`.\n", - "\n", - "Overall code is easier to read and maintain if it uses standard layers whenever possible, as other readers will be familiar with the behavior of standard layers. If you want to use a layer which is not present in tf.keras.layers or tf.contrib.layers, consider filing a [github issue](http://github.com/tensorflow/tensorflow/issues/new) or, even better, sending us a pull request!" - ] - }, - { - "cell_type": "markdown", - "metadata": { - "colab_type": "text", - "id": "Qhg4KlbKrs3G" - }, - "source": [ - "## Models: composing layers\n", - "\n", - "Many interesting layer-like things in machine learning models are implemented by composing existing layers. For example, each residual block in a resnet is a composition of convolutions, batch normalizations, and a shortcut.\n", - "\n", - "The main class used when creating a layer-like thing which contains other layers is tf.keras.Model. Implementing one is done by inheriting from tf.keras.Model." - ] - }, - { - "cell_type": "code", - "execution_count": 9, - "metadata": { - "colab": { - "autoexec": { - "startup": false, - "wait_interval": 0 - }, - "height": 190 - }, - "colab_type": "code", - "executionInfo": { - "elapsed": 420, - "status": "ok", - "timestamp": 1527783698512, - "user": { - "displayName": "", - "photoUrl": "", - "userId": "" - }, - "user_tz": 420 - }, - "id": "N30DTXiRASlb", - "outputId": "a8b23a8e-5cf9-4bbf-f93b-6c763d74e2b3" - }, - "outputs": [ - { - "name": "stdout", - "output_type": "stream", - "text": [ - "tf.Tensor(\n", - "[[[[ 0. 0. 0.]\n", - " [ 0. 0. 0.]\n", - " [ 0. 0. 0.]]\n", - "\n", - " [[ 0. 0. 0.]\n", - " [ 0. 0. 0.]\n", - " [ 0. 0. 0.]]]], shape=(1, 2, 3, 3), dtype=float32)\n", - "['resnet_identity_block_1/conv2d_3/kernel:0', 'resnet_identity_block_1/conv2d_3/bias:0', 'resnet_identity_block_1/batch_normalization_3/gamma:0', 'resnet_identity_block_1/batch_normalization_3/beta:0', 'resnet_identity_block_1/conv2d_4/kernel:0', 'resnet_identity_block_1/conv2d_4/bias:0', 'resnet_identity_block_1/batch_normalization_4/gamma:0', 'resnet_identity_block_1/batch_normalization_4/beta:0', 'resnet_identity_block_1/conv2d_5/kernel:0', 'resnet_identity_block_1/conv2d_5/bias:0', 'resnet_identity_block_1/batch_normalization_5/gamma:0', 'resnet_identity_block_1/batch_normalization_5/beta:0', 'resnet_identity_block_1/batch_normalization_3/moving_mean:0', 'resnet_identity_block_1/batch_normalization_3/moving_variance:0', 'resnet_identity_block_1/batch_normalization_4/moving_mean:0', 'resnet_identity_block_1/batch_normalization_4/moving_variance:0', 'resnet_identity_block_1/batch_normalization_5/moving_mean:0', 'resnet_identity_block_1/batch_normalization_5/moving_variance:0']\n" - ] - } - ], - "source": [ - "class ResnetIdentityBlock(tf.keras.Model):\n", - " def __init__(self, kernel_size, filters):\n", - " super(ResnetIdentityBlock, self).__init__(name='')\n", - " filters1, filters2, filters3 = filters\n", - "\n", - " self.conv2a = tf.keras.layers.Conv2D(filters1, (1, 1))\n", - " self.bn2a = tf.keras.layers.BatchNormalization()\n", - "\n", - " self.conv2b = tf.keras.layers.Conv2D(filters2, kernel_size, padding='same')\n", - " self.bn2b = tf.keras.layers.BatchNormalization()\n", - "\n", - " self.conv2c = tf.keras.layers.Conv2D(filters3, (1, 1))\n", - " self.bn2c = tf.keras.layers.BatchNormalization()\n", - "\n", - " def call(self, input_tensor, training=False):\n", - " x = self.conv2a(input_tensor)\n", - " x = self.bn2a(x, training=training)\n", - " x = tf.nn.relu(x)\n", - "\n", - " x = self.conv2b(x)\n", - " x = self.bn2b(x, training=training)\n", - " x = tf.nn.relu(x)\n", - "\n", - " x = self.conv2c(x)\n", - " x = self.bn2c(x, training=training)\n", - "\n", - " x += input_tensor\n", - " return tf.nn.relu(x)\n", - "\n", - " \n", - "block = ResnetIdentityBlock(1, [1, 2, 3])\n", - "print(block(tf.zeros([1, 2, 3, 3])))\n", - "print([x.name for x in block.variables])" - ] - }, - { - "cell_type": "markdown", - "metadata": { - "colab_type": "text", - "id": "wYfucVw65PMj" - }, - "source": [ - "Much of the time, however, models which compose many layers simply call one layer after the other. This can be done in very little code using tf.keras.Sequential" - ] - }, - { - "cell_type": "code", - "execution_count": 0, - "metadata": { - "colab": { - "autoexec": { - "startup": false, - "wait_interval": 0 - }, - "base_uri": "https://localhost:8080/", - "height": 153 - }, - "colab_type": "code", - "executionInfo": { - "elapsed": 361, - "status": "ok", - "timestamp": 1526674830777, - "user": { - "displayName": "Alexandre Passos", - "photoUrl": "//lh4.googleusercontent.com/-kmTTWXEgAPw/AAAAAAAAAAI/AAAAAAAAAC0/q_DoOzKGwds/s50-c-k-no/photo.jpg", - "userId": "108023195365833072773" - }, - "user_tz": 420 - }, - "id": "L9frk7Ur4uvJ", - "outputId": "882e9076-b6d9-4380-bb1e-7c6b57d54c39" - }, - "outputs": [ - { - "data": { - "text/plain": [ - "\u003ctf.Tensor: id=1423, shape=(1, 2, 3, 3), dtype=float32, numpy=\n", - "array([[[[0., 0., 0.],\n", - " [0., 0., 0.],\n", - " [0., 0., 0.]],\n", - "\n", - " [[0., 0., 0.],\n", - " [0., 0., 0.],\n", - " [0., 0., 0.]]]], dtype=float32)\u003e" - ] - }, - "execution_count": 26, - "metadata": { - "tags": [] - }, - "output_type": "execute_result" - } - ], - "source": [ - " my_seq = tf.keras.Sequential([tf.keras.layers.Conv2D(1, (1, 1)),\n", - " tf.keras.layers.BatchNormalization(),\n", - " tf.keras.layers.Conv2D(2, 1, \n", - " padding='same'),\n", - " tf.keras.layers.BatchNormalization(),\n", - " tf.keras.layers.Conv2D(3, (1, 1)),\n", - " tf.keras.layers.BatchNormalization()])\n", - "my_seq(tf.zeros([1, 2, 3, 3]))" - ] - }, - { - "cell_type": "markdown", - "metadata": { - "colab_type": "text", - "id": "c5YwYcnuK-wc" - }, - "source": [ - "# Next steps\n", - "\n", - "Now you can go back to the previous notebook and adapt the linear regression example to use layers and models to be better structured." - ] - } - ], - "metadata": { - "colab": { - "collapsed_sections": [], - "default_view": {}, - "name": "4 - High level API - TensorFlow Eager.ipynb", - "provenance": [], - "version": "0.3.2", - "views": {} - }, - "kernelspec": { - "display_name": "Python 3", - "name": "python3" - } - }, - "nbformat": 4, - "nbformat_minor": 0 -} diff --git a/tensorflow/contrib/eager/python/examples/notebooks/README.md b/tensorflow/contrib/eager/python/examples/notebooks/README.md new file mode 100644 index 0000000000000000000000000000000000000000..0d5ed848946d1eee643a57bf8c341520268c56b1 --- /dev/null +++ b/tensorflow/contrib/eager/python/examples/notebooks/README.md @@ -0,0 +1,11 @@ +## Research and experimentation + +Eager execution provides an imperative, define-by-run interface for advanced +operations. Write custom layers, forward passes, and training loops with auto +differentiation. Start with these notebooks, then read the +[eager execution guide](https://www.tensorflow.org/guide/eager). + +1. [Eager execution basics](./eager_basics.ipynb) +2. [Automatic differentiation and gradient tapes](./automatic_differentiation.ipynb) +3. [Custom training: basics](./custom_training.ipynb) +4. [Custom layers](./custom_layers.ipynb) diff --git a/tensorflow/contrib/eager/python/examples/notebooks/automatic_differentiation.ipynb b/tensorflow/contrib/eager/python/examples/notebooks/automatic_differentiation.ipynb new file mode 100644 index 0000000000000000000000000000000000000000..7c0f9b5b8161a763c4153ebdeece7e0d1b90b384 --- /dev/null +++ b/tensorflow/contrib/eager/python/examples/notebooks/automatic_differentiation.ipynb @@ -0,0 +1,364 @@ +{ + "nbformat": 4, + "nbformat_minor": 0, + "metadata": { + "colab": { + "name": "automatic_differentiation.ipynb", + "version": "0.3.2", + "views": {}, + "default_view": {}, + "provenance": [], + "private_outputs": true, + "collapsed_sections": [], + "toc_visible": true + }, + "kernelspec": { + "name": "python3", + "display_name": "Python 3" + } + }, + "cells": [ + { + "metadata": { + "id": "t09eeeR5prIJ", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "##### Copyright 2018 The TensorFlow Authors." + ] + }, + { + "metadata": { + "id": "GCCk8_dHpuNf", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + }, + "cellView": "form" + }, + "cell_type": "code", + "source": [ + "#@title Licensed under the Apache License, Version 2.0 (the \"License\");\n", + "# you may not use this file except in compliance with the License.\n", + "# You may obtain a copy of the License at\n", + "#\n", + "# https://www.apache.org/licenses/LICENSE-2.0\n", + "#\n", + "# Unless required by applicable law or agreed to in writing, software\n", + "# distributed under the License is distributed on an \"AS IS\" BASIS,\n", + "# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n", + "# See the License for the specific language governing permissions and\n", + "# limitations under the License." + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "xh8WkEwWpnm7", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "# Automatic differentiation and gradient tape" + ] + }, + { + "metadata": { + "id": "idv0bPeCp325", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "
\n", + "\n", + " Run in Google Colab\n", + "\n", + "View source on GitHub
" + ] + }, + { + "metadata": { + "id": "vDJ4XzMqodTy", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "In the previous tutorial we introduced `Tensor`s and operations on them. In this tutorial we will cover [automatic differentiation](https://en.wikipedia.org/wiki/Automatic_differentiation), a key technique for optimizing machine learning models." + ] + }, + { + "metadata": { + "id": "GQJysDM__Qb0", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "## Setup\n" + ] + }, + { + "metadata": { + "id": "OiMPZStlibBv", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + } + }, + "cell_type": "code", + "source": [ + "import tensorflow as tf\n", + "tf.enable_eager_execution()\n", + "\n", + "tfe = tf.contrib.eager # Shorthand for some symbols" + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "1CLWJl0QliB0", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "## Derivatives of a function\n", + "\n", + "TensorFlow provides APIs for automatic differentiation - computing the derivative of a function. The way that more closely mimics the math is to encapsulate the computation in a Python function, say `f`, and use `tfe.gradients_function` to create a function that computes the derivatives of `f` with respect to its arguments. If you're familiar with [autograd](https://github.com/HIPS/autograd) for differentiating numpy functions, this will be familiar. For example: " + ] + }, + { + "metadata": { + "id": "9FViq92UX7P8", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + } + }, + "cell_type": "code", + "source": [ + "from math import pi\n", + "\n", + "def f(x):\n", + " return tf.square(tf.sin(x))\n", + "\n", + "assert f(pi/2).numpy() == 1.0\n", + "\n", + "\n", + "# grad_f will return a list of derivatives of f\n", + "# with respect to its arguments. Since f() has a single argument,\n", + "# grad_f will return a list with a single element.\n", + "grad_f = tfe.gradients_function(f)\n", + "assert tf.abs(grad_f(pi/2)[0]).numpy() < 1e-7" + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "v9fPs8RyopCf", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "### Higher-order gradients\n", + "\n", + "The same API can be used to differentiate as many times as you like:\n" + ] + }, + { + "metadata": { + "id": "3D0ZvnGYo0rW", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + } + }, + "cell_type": "code", + "source": [ + "def f(x):\n", + " return tf.square(tf.sin(x))\n", + "\n", + "def grad(f):\n", + " return lambda x: tfe.gradients_function(f)(x)[0]\n", + "\n", + "x = tf.lin_space(-2*pi, 2*pi, 100) # 100 points between -2Ļ€ and +2Ļ€\n", + "\n", + "import matplotlib.pyplot as plt\n", + "\n", + "plt.plot(x, f(x), label=\"f\")\n", + "plt.plot(x, grad(f)(x), label=\"first derivative\")\n", + "plt.plot(x, grad(grad(f))(x), label=\"second derivative\")\n", + "plt.plot(x, grad(grad(grad(f)))(x), label=\"third derivative\")\n", + "plt.legend()\n", + "plt.show()" + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "-39gouo7mtgu", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "## Gradient tapes\n", + "\n", + "Every differentiable TensorFlow operation has an associated gradient function. For example, the gradient function of `tf.square(x)` would be a function that returns `2.0 * x`. To compute the gradient of a user-defined function (like `f(x)` in the example above), TensorFlow first \"records\" all the operations applied to compute the output of the function. We call this record a \"tape\". It then uses that tape and the gradients functions associated with each primitive operation to compute the gradients of the user-defined function using [reverse mode differentiation](https://en.wikipedia.org/wiki/Automatic_differentiation).\n", + "\n", + "Since operations are recorded as they are executed, Python control flow (using `if`s and `while`s for example) is naturally handled:\n", + "\n" + ] + }, + { + "metadata": { + "id": "MH0UfjympWf7", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + } + }, + "cell_type": "code", + "source": [ + "def f(x, y):\n", + " output = 1\n", + " for i in range(y):\n", + " output = tf.multiply(output, x)\n", + " return output\n", + "\n", + "def g(x, y):\n", + " # Return the gradient of `f` with respect to it's first parameter\n", + " return tfe.gradients_function(f)(x, y)[0]\n", + "\n", + "assert f(3.0, 2).numpy() == 9.0 # f(x, 2) is essentially x * x\n", + "assert g(3.0, 2).numpy() == 6.0 # And its gradient will be 2 * x\n", + "assert f(4.0, 3).numpy() == 64.0 # f(x, 3) is essentially x * x * x\n", + "assert g(4.0, 3).numpy() == 48.0 # And its gradient will be 3 * x * x" + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "aNmR5-jhpX2t", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "At times it may be inconvenient to encapsulate computation of interest into a function. For example, if you want the gradient of the output with respect to intermediate values computed in the function. In such cases, the slightly more verbose but explicit [tf.GradientTape](https://www.tensorflow.org/api_docs/python/tf/GradientTape) context is useful. All computation inside the context of a `tf.GradientTape` is \"recorded\".\n", + "\n", + "For example:" + ] + }, + { + "metadata": { + "id": "bAFeIE8EuVIq", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + } + }, + "cell_type": "code", + "source": [ + "x = tf.ones((2, 2))\n", + " \n", + "# TODO(b/78880779): Remove the 'persistent=True' argument and use\n", + "# a single t.gradient() call when the bug is resolved.\n", + "with tf.GradientTape(persistent=True) as t:\n", + " # TODO(ashankar): Explain with \"watch\" argument better?\n", + " t.watch(x)\n", + " y = tf.reduce_sum(x)\n", + " z = tf.multiply(y, y)\n", + "\n", + "# Use the same tape to compute the derivative of z with respect to the\n", + "# intermediate value y.\n", + "dz_dy = t.gradient(z, y)\n", + "assert dz_dy.numpy() == 8.0\n", + "\n", + "# Derivative of z with respect to the original input tensor x\n", + "dz_dx = t.gradient(z, x)\n", + "for i in [0, 1]:\n", + " for j in [0, 1]:\n", + " assert dz_dx[i][j].numpy() == 8.0" + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "DK05KXrAAld3", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "### Higher-order gradients\n", + "\n", + "Operations inside of the `GradientTape` context manager are recorded for automatic differentiation. If gradients are computed in that context, then the gradient computation is recorded as well. As a result, the exact same API works for higher-order gradients as well. For example:" + ] + }, + { + "metadata": { + "id": "cPQgthZ7ugRJ", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + } + }, + "cell_type": "code", + "source": [ + "# TODO(ashankar): Should we use the persistent tape here instead? Follow up on Tom and Alex's discussion\n", + "\n", + "x = tf.constant(1.0) # Convert the Python 1.0 to a Tensor object\n", + "\n", + "with tf.GradientTape() as t:\n", + " with tf.GradientTape() as t2:\n", + " t2.watch(x)\n", + " y = x * x * x\n", + " # Compute the gradient inside the 't' context manager\n", + " # which means the gradient computation is differentiable as well.\n", + " dy_dx = t2.gradient(y, x)\n", + "d2y_dx2 = t.gradient(dy_dx, x)\n", + "\n", + "assert dy_dx.numpy() == 3.0\n", + "assert d2y_dx2.numpy() == 6.0" + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "4U1KKzUpNl58", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "## Next Steps\n", + "\n", + "In this tutorial we covered gradient computation in TensorFlow. With that we have enough of the primitives required to build an train neural networks, which we will cover in the [next tutorial](https://github.com/tensorflow/models/tree/master/official/contrib/eager/python/examples/notebooks/3_neural_networks.ipynb)." + ] + } + ] +} \ No newline at end of file diff --git a/tensorflow/contrib/eager/python/examples/notebooks/custom_layers.ipynb b/tensorflow/contrib/eager/python/examples/notebooks/custom_layers.ipynb new file mode 100644 index 0000000000000000000000000000000000000000..a0bbbb612381c5eb386b04fd7bb9914eb01f4c8e --- /dev/null +++ b/tensorflow/contrib/eager/python/examples/notebooks/custom_layers.ipynb @@ -0,0 +1,399 @@ +{ + "nbformat": 4, + "nbformat_minor": 0, + "metadata": { + "colab": { + "name": "custom_layers.ipynb", + "version": "0.3.2", + "views": {}, + "default_view": {}, + "provenance": [], + "private_outputs": true, + "collapsed_sections": [], + "toc_visible": true + }, + "kernelspec": { + "display_name": "Python 3", + "name": "python3" + } + }, + "cells": [ + { + "metadata": { + "id": "tDnwEv8FtJm7", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "##### Copyright 2018 The TensorFlow Authors." + ] + }, + { + "metadata": { + "id": "JlknJBWQtKkI", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + }, + "cellView": "form" + }, + "cell_type": "code", + "source": [ + "#@title Licensed under the Apache License, Version 2.0 (the \"License\");\n", + "# you may not use this file except in compliance with the License.\n", + "# You may obtain a copy of the License at\n", + "#\n", + "# https://www.apache.org/licenses/LICENSE-2.0\n", + "#\n", + "# Unless required by applicable law or agreed to in writing, software\n", + "# distributed under the License is distributed on an \"AS IS\" BASIS,\n", + "# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n", + "# See the License for the specific language governing permissions and\n", + "# limitations under the License." + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "60RdWsg1tETW", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "# Custom layers" + ] + }, + { + "metadata": { + "id": "BcJg7Enms86w", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "
\n", + "\n", + " Run in Google Colab\n", + "\n", + "View source on GitHub
" + ] + }, + { + "metadata": { + "id": "UEu3q4jmpKVT", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "We recommend using `tf.keras` as a high-level API for building neural networks. That said, most TensorFlow APIs are usable with eager execution.\n" + ] + }, + { + "metadata": { + "id": "pwX7Fii1rwsJ", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + } + }, + "cell_type": "code", + "source": [ + "import tensorflow as tf\n", + "tfe = tf.contrib.eager\n", + "\n", + "tf.enable_eager_execution()" + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "zSFfVVjkrrsI", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "## Layers: common sets of useful operations\n", + "\n", + "Most of the time when writing code for machine learning models you want to operate at a higher level of abstraction than individual operations and manipulation of individual variables.\n", + "\n", + "Many machine learning models are expressible as the composition and stacking of relatively simple layers, and TensorFlow provides both a set of many common layers as a well as easy ways for you to write your own application-specific layers either from scratch or as the composition of existing layers.\n", + "\n", + "TensorFlow includes the full [Keras](https://keras.io) API in the tf.keras package, and the Keras layers are very useful when building your own models.\n" + ] + }, + { + "metadata": { + "id": "8PyXlPl-4TzQ", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + } + }, + "cell_type": "code", + "source": [ + "# In the tf.keras.layers package, layers are objects. To construct a layer,\n", + "# simply construct the object. Most layers take as a first argument the number\n", + "# of output dimensions / channels.\n", + "layer = tf.keras.layers.Dense(100)\n", + "# The number of input dimensions is often unnecessary, as it can be inferred\n", + "# the first time the layer is used, but it can be provided if you want to \n", + "# specify it manually, which is useful in some complex models.\n", + "layer = tf.keras.layers.Dense(10, input_shape=(None, 5))" + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "Fn69xxPO5Psr", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "The full list of pre-existing layers can be seen in [the documentation](https://www.tensorflow.org/api_docs/python/tf/keras/layers). It includes Dense (a fully-connected layer),\n", + "Conv2D, LSTM, BatchNormalization, Dropout, and many others." + ] + }, + { + "metadata": { + "id": "E3XKNknP5Mhb", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + } + }, + "cell_type": "code", + "source": [ + "# To use a layer, simply call it.\n", + "layer(tf.zeros([10, 5]))" + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "Wt_Nsv-L5t2s", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + } + }, + "cell_type": "code", + "source": [ + "# Layers have many useful methods. For example, you can inspect all variables\n", + "# in a layer by calling layer.variables. In this case a fully-connected layer\n", + "# will have variables for weights and biases.\n", + "layer.variables" + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "6ilvKjz8_4MQ", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + } + }, + "cell_type": "code", + "source": [ + "# The variables are also accessible through nice accessors\n", + "layer.kernel, layer.bias" + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "O0kDbE54-5VS", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "## Implementing custom layers\n", + "The best way to implement your own layer is extending the tf.keras.Layer class and implementing:\n", + " * `__init__` , where you can do all input-independent initialization\n", + " * `build`, where you know the shapes of the input tensors and can do the rest of the initialization\n", + " * `call`, where you do the forward computation\n", + "\n", + "Note that you don't have to wait until `build` is called to create your variables, you can also create them in `__init__`. However, the advantage of creating them in `build` is that it enables late variable creation based on the shape of the inputs the layer will operate on. On the other hand, creating variables in `__init__` would mean that shapes required to create the variables will need to be explicitly specified." + ] + }, + { + "metadata": { + "id": "5Byl3n1k5kIy", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + } + }, + "cell_type": "code", + "source": [ + "class MyDenseLayer(tf.keras.layers.Layer):\n", + " def __init__(self, num_outputs):\n", + " super(MyDenseLayer, self).__init__()\n", + " self.num_outputs = num_outputs\n", + " \n", + " def build(self, input_shape):\n", + " self.kernel = self.add_variable(\"kernel\", \n", + " shape=[input_shape[-1].value, \n", + " self.num_outputs])\n", + " \n", + " def call(self, input):\n", + " return tf.matmul(input, self.kernel)\n", + " \n", + "layer = MyDenseLayer(10)\n", + "print(layer(tf.zeros([10, 5])))\n", + "print(layer.variables)" + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "tk8E2vY0-z4Z", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "Note that you don't have to wait until `build` is called to create your variables, you can also create them in `__init__`.\n", + "\n", + "Overall code is easier to read and maintain if it uses standard layers whenever possible, as other readers will be familiar with the behavior of standard layers. If you want to use a layer which is not present in tf.keras.layers or tf.contrib.layers, consider filing a [github issue](http://github.com/tensorflow/tensorflow/issues/new) or, even better, sending us a pull request!" + ] + }, + { + "metadata": { + "id": "Qhg4KlbKrs3G", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "## Models: composing layers\n", + "\n", + "Many interesting layer-like things in machine learning models are implemented by composing existing layers. For example, each residual block in a resnet is a composition of convolutions, batch normalizations, and a shortcut.\n", + "\n", + "The main class used when creating a layer-like thing which contains other layers is tf.keras.Model. Implementing one is done by inheriting from tf.keras.Model." + ] + }, + { + "metadata": { + "id": "N30DTXiRASlb", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + } + }, + "cell_type": "code", + "source": [ + "class ResnetIdentityBlock(tf.keras.Model):\n", + " def __init__(self, kernel_size, filters):\n", + " super(ResnetIdentityBlock, self).__init__(name='')\n", + " filters1, filters2, filters3 = filters\n", + "\n", + " self.conv2a = tf.keras.layers.Conv2D(filters1, (1, 1))\n", + " self.bn2a = tf.keras.layers.BatchNormalization()\n", + "\n", + " self.conv2b = tf.keras.layers.Conv2D(filters2, kernel_size, padding='same')\n", + " self.bn2b = tf.keras.layers.BatchNormalization()\n", + "\n", + " self.conv2c = tf.keras.layers.Conv2D(filters3, (1, 1))\n", + " self.bn2c = tf.keras.layers.BatchNormalization()\n", + "\n", + " def call(self, input_tensor, training=False):\n", + " x = self.conv2a(input_tensor)\n", + " x = self.bn2a(x, training=training)\n", + " x = tf.nn.relu(x)\n", + "\n", + " x = self.conv2b(x)\n", + " x = self.bn2b(x, training=training)\n", + " x = tf.nn.relu(x)\n", + "\n", + " x = self.conv2c(x)\n", + " x = self.bn2c(x, training=training)\n", + "\n", + " x += input_tensor\n", + " return tf.nn.relu(x)\n", + "\n", + " \n", + "block = ResnetIdentityBlock(1, [1, 2, 3])\n", + "print(block(tf.zeros([1, 2, 3, 3])))\n", + "print([x.name for x in block.variables])" + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "wYfucVw65PMj", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "Much of the time, however, models which compose many layers simply call one layer after the other. This can be done in very little code using tf.keras.Sequential" + ] + }, + { + "metadata": { + "id": "L9frk7Ur4uvJ", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + } + }, + "cell_type": "code", + "source": [ + " my_seq = tf.keras.Sequential([tf.keras.layers.Conv2D(1, (1, 1)),\n", + " tf.keras.layers.BatchNormalization(),\n", + " tf.keras.layers.Conv2D(2, 1, \n", + " padding='same'),\n", + " tf.keras.layers.BatchNormalization(),\n", + " tf.keras.layers.Conv2D(3, (1, 1)),\n", + " tf.keras.layers.BatchNormalization()])\n", + "my_seq(tf.zeros([1, 2, 3, 3]))" + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "c5YwYcnuK-wc", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "# Next steps\n", + "\n", + "Now you can go back to the previous notebook and adapt the linear regression example to use layers and models to be better structured." + ] + } + ] +} \ No newline at end of file diff --git a/tensorflow/contrib/eager/python/examples/notebooks/custom_training.ipynb b/tensorflow/contrib/eager/python/examples/notebooks/custom_training.ipynb new file mode 100644 index 0000000000000000000000000000000000000000..591e2d0c852a1482e624e83769f2d8df49547cea --- /dev/null +++ b/tensorflow/contrib/eager/python/examples/notebooks/custom_training.ipynb @@ -0,0 +1,478 @@ +{ + "nbformat": 4, + "nbformat_minor": 0, + "metadata": { + "colab": { + "name": "Custom training: basics", + "version": "0.3.2", + "views": {}, + "default_view": {}, + "provenance": [], + "private_outputs": true, + "collapsed_sections": [], + "toc_visible": true + }, + "kernelspec": { + "name": "python3", + "display_name": "Python 3" + } + }, + "cells": [ + { + "metadata": { + "id": "5rmpybwysXGV", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "##### Copyright 2018 The TensorFlow Authors." + ] + }, + { + "metadata": { + "id": "m8y3rGtQsYP2", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + }, + "cellView": "form" + }, + "cell_type": "code", + "source": [ + "#@title Licensed under the Apache License, Version 2.0 (the \"License\");\n", + "# you may not use this file except in compliance with the License.\n", + "# You may obtain a copy of the License at\n", + "#\n", + "# https://www.apache.org/licenses/LICENSE-2.0\n", + "#\n", + "# Unless required by applicable law or agreed to in writing, software\n", + "# distributed under the License is distributed on an \"AS IS\" BASIS,\n", + "# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n", + "# See the License for the specific language governing permissions and\n", + "# limitations under the License." + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "hrXv0rU9sIma", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "# Custom training: basics" + ] + }, + { + "metadata": { + "id": "7S0BwJ_8sLu7", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "
\n", + "\n", + " Run in Google Colab\n", + "\n", + "View source on GitHub
" + ] + }, + { + "metadata": { + "id": "k2o3TTG4TFpt", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "In the previous tutorial we covered the TensorFlow APIs for automatic differentiation, a basic building block for machine learning.\n", + "In this tutorial we will use the TensorFlow primitives introduced in the prior tutorials to do some simple machine learning.\n", + "\n", + "TensorFlow also includes a higher-level neural networks API (`tf.keras`) which provides useful abstractions to reduce boilerplate. We strongly recommend those higher level APIs for people working with neural networks. However, in this short tutorial we cover neural network training from first principles to establish a strong foundation." + ] + }, + { + "metadata": { + "id": "3LXMVuV0VhDr", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "## Setup" + ] + }, + { + "metadata": { + "id": "PJ64L90aVir3", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + } + }, + "cell_type": "code", + "source": [ + "import tensorflow as tf\n", + "tfe = tf.contrib.eager # Shorthand for some symbols\n", + "\n", + "tf.enable_eager_execution()" + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "eMAWbDJFVmMk", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "## Variables\n", + "\n", + "Tensors in TensorFlow are immutable stateless objects. Machine learning models, however, need to have changing state: as your model trains, the same code to compute predictions should behave differently over time (hopefully with a lower loss!). To represent this state which needs to change over the course of your computation, you can choose to rely on the fact that Python is a stateful programming language:\n" + ] + }, + { + "metadata": { + "id": "VkJwtLS_Jbn8", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + } + }, + "cell_type": "code", + "source": [ + "# Using python state\n", + "x = tf.zeros([10, 10])\n", + "x += 2 # This is equivalent to x = x + 2, which does not mutate the original\n", + " # value of x\n", + "print(x)" + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "wfneTXy7JcUz", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "TensorFlow, however, has stateful operations built in, and these are often more pleasant to use than low-level Python representations of your state. To represent weights in a model, for example, it's often convenient and efficient to use TensorFlow variables.\n", + "\n", + "A Variable is an object which stores a value and, when used in a TensorFlow computation, will implicitly read from this stored value. There are operations (`tf.assign_sub`, `tf.scatter_update`, etc) which manipulate the value stored in a TensorFlow variable." + ] + }, + { + "metadata": { + "id": "itxmrMil6DQi", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + } + }, + "cell_type": "code", + "source": [ + "v = tfe.Variable(1.0)\n", + "assert v.numpy() == 1.0\n", + "\n", + "# Re-assign the value\n", + "v.assign(3.0)\n", + "assert v.numpy() == 3.0\n", + "\n", + "# Use `v` in a TensorFlow operation like tf.square() and reassign\n", + "v.assign(tf.square(v))\n", + "assert v.numpy() == 9.0" + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "-paSaeq1JzwC", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "Computations using Variables are automatically traced when computing gradients. For Variables representing embeddings TensorFlow will do sparse updates by default, which are more computation and memory efficient.\n", + "\n", + "Using Variables is also a way to quickly let a reader of your code know that this piece of state is mutable." + ] + }, + { + "metadata": { + "id": "BMiFcDzE7Qu3", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "## Example: Fitting a linear model\n", + "\n", + "Let's now put the few concepts we have so far ---`Tensor`, `GradientTape`, `Variable` --- to build and train a simple model. This typically involves a few steps:\n", + "\n", + "1. Define the model.\n", + "2. Define a loss function.\n", + "3. Obtain training data.\n", + "4. Run through the training data and use an \"optimizer\" to adjust the variables to fit the data.\n", + "\n", + "In this tutorial, we'll walk through a trivial example of a simple linear model: `f(x) = x * W + b`, which has two variables - `W` and `b`. Furthermore, we'll synthesize data such that a well trained model would have `W = 3.0` and `b = 2.0`." + ] + }, + { + "metadata": { + "id": "gFzH64Jn9PIm", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "### Define the model\n", + "\n", + "Let's define a simple class to encapsulate the variables and the computation." + ] + }, + { + "metadata": { + "id": "_WRu7Pze7wk8", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + } + }, + "cell_type": "code", + "source": [ + "class Model(object):\n", + " def __init__(self):\n", + " # Initialize variable to (5.0, 0.0)\n", + " # In practice, these should be initialized to random values.\n", + " self.W = tfe.Variable(5.0)\n", + " self.b = tfe.Variable(0.0)\n", + " \n", + " def __call__(self, x):\n", + " return self.W * x + self.b\n", + " \n", + "model = Model()\n", + "\n", + "assert model(3.0).numpy() == 15.0" + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "xa6j_yXa-j79", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "### Define a loss function\n", + "\n", + "A loss function measures how well the output of a model for a given input matches the desired output. Let's use the standard L2 loss." + ] + }, + { + "metadata": { + "id": "Y0ysUFGY924U", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + } + }, + "cell_type": "code", + "source": [ + "def loss(predicted_y, desired_y):\n", + " return tf.reduce_mean(tf.square(predicted_y - desired_y))" + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "qutT_fkl_CBc", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "### Obtain training data\n", + "\n", + "Let's synthesize the training data with some noise." + ] + }, + { + "metadata": { + "id": "gxPTb-kt_N5m", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + } + }, + "cell_type": "code", + "source": [ + "TRUE_W = 3.0\n", + "TRUE_b = 2.0\n", + "NUM_EXAMPLES = 1000\n", + "\n", + "inputs = tf.random_normal(shape=[NUM_EXAMPLES])\n", + "noise = tf.random_normal(shape=[NUM_EXAMPLES])\n", + "outputs = inputs * TRUE_W + TRUE_b + noise" + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "-50nq-wPBsAW", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "Before we train the model let's visualize where the model stands right now. We'll plot the model's predictions in red and the training data in blue." + ] + }, + { + "metadata": { + "id": "_eb83LtrB4nt", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + } + }, + "cell_type": "code", + "source": [ + "import matplotlib.pyplot as plt\n", + "\n", + "plt.scatter(inputs, outputs, c='b')\n", + "plt.scatter(inputs, model(inputs), c='r')\n", + "plt.show()\n", + "\n", + "print('Current loss: '),\n", + "print(loss(model(inputs), outputs).numpy())" + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "sSDP-yeq_4jE", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "### Define a training loop\n", + "\n", + "We now have our network and our training data. Let's train it, i.e., use the training data to update the model's variables (`W` and `b`) so that the loss goes down using [gradient descent](https://en.wikipedia.org/wiki/Gradient_descent). There are many variants of the gradient descent scheme that are captured in `tf.train.Optimizer` implementations. We'd highly recommend using those implementations, but in the spirit of building from first principles, in this particular example we will implement the basic math ourselves." + ] + }, + { + "metadata": { + "id": "MBIACgdnA55X", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + } + }, + "cell_type": "code", + "source": [ + "def train(model, inputs, outputs, learning_rate):\n", + " with tf.GradientTape() as t:\n", + " current_loss = loss(model(inputs), outputs)\n", + " dW, db = t.gradient(current_loss, [model.W, model.b])\n", + " model.W.assign_sub(learning_rate * dW)\n", + " model.b.assign_sub(learning_rate * db)" + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "RwWPaJryD2aN", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "Finally, let's repeatedly run through the training data and see how `W` and `b` evolve." + ] + }, + { + "metadata": { + "id": "XdfkR223D9dW", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + } + }, + "cell_type": "code", + "source": [ + "model = Model()\n", + "\n", + "# Collect the history of W-values and b-values to plot later\n", + "Ws, bs = [], []\n", + "epochs = range(10)\n", + "for epoch in epochs:\n", + " Ws.append(model.W.numpy())\n", + " bs.append(model.b.numpy())\n", + " current_loss = loss(model(inputs), outputs)\n", + "\n", + " train(model, inputs, outputs, learning_rate=0.1)\n", + " print('Epoch %2d: W=%1.2f b=%1.2f, loss=%2.5f' %\n", + " (epoch, Ws[-1], bs[-1], current_loss))\n", + "\n", + "# Let's plot it all\n", + "plt.plot(epochs, Ws, 'r',\n", + " epochs, bs, 'b')\n", + "plt.plot([TRUE_W] * len(epochs), 'r--',\n", + " [TRUE_b] * len(epochs), 'b--')\n", + "plt.legend(['W', 'b', 'true W', 'true_b'])\n", + "plt.show()\n", + " " + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "vPnIVuaSJwWz", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "## Next Steps\n", + "\n", + "In this tutorial we covered `Variable`s and built and trained a simple linear model using the TensorFlow primitives discussed so far.\n", + "\n", + "In theory, this is pretty much all you need to use TensorFlow for your machine learning research.\n", + "In practice, particularly for neural networks, the higher level APIs like `tf.keras` will be much more convenient since it provides higher level building blocks (called \"layers\"), utilities to save and restore state, a suite of loss functions, a suite of optimization strategies etc. \n", + "\n", + "The [next tutorial](TODO) will cover these higher level APIs." + ] + } + ] +} \ No newline at end of file diff --git a/tensorflow/contrib/eager/python/examples/notebooks/1_basics.ipynb b/tensorflow/contrib/eager/python/examples/notebooks/eager_basics.ipynb similarity index 50% rename from tensorflow/contrib/eager/python/examples/notebooks/1_basics.ipynb rename to tensorflow/contrib/eager/python/examples/notebooks/eager_basics.ipynb index 51d10a778413cfbb574b4e22e8adcb18bd731dee..f1e13de5dec2fbda126caeb355494875317e3373 100644 --- a/tensorflow/contrib/eager/python/examples/notebooks/1_basics.ipynb +++ b/tensorflow/contrib/eager/python/examples/notebooks/eager_basics.ipynb @@ -1,27 +1,107 @@ { + "nbformat": 4, + "nbformat_minor": 0, + "metadata": { + "colab": { + "name": "eager_basics.ipynb", + "version": "0.3.2", + "views": {}, + "default_view": {}, + "provenance": [], + "private_outputs": true, + "collapsed_sections": [], + "toc_visible": true + }, + "kernelspec": { + "name": "python3", + "display_name": "Python 3" + } + }, "cells": [ { + "metadata": { + "id": "iPpI7RaYoZuE", + "colab_type": "text" + }, "cell_type": "markdown", + "source": [ + "##### Copyright 2018 The TensorFlow Authors." + ] + }, + { "metadata": { - "colab_type": "text", - "id": "U9i2Dsh-ziXr" + "id": "hro2InpHobKk", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + }, + "cellView": "form" }, + "cell_type": "code", + "source": [ + "#@title Licensed under the Apache License, Version 2.0 (the \"License\");\n", + "# you may not use this file except in compliance with the License.\n", + "# You may obtain a copy of the License at\n", + "#\n", + "# https://www.apache.org/licenses/LICENSE-2.0\n", + "#\n", + "# Unless required by applicable law or agreed to in writing, software\n", + "# distributed under the License is distributed on an \"AS IS\" BASIS,\n", + "# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n", + "# See the License for the specific language governing permissions and\n", + "# limitations under the License." + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "U9i2Dsh-ziXr", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "# Eager execution basics" + ] + }, + { + "metadata": { + "id": "Hndw-YcxoOJK", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "
\n", + "\n", + " Run in Google Colab\n", + "\n", + "View source on GitHub
" + ] + }, + { + "metadata": { + "id": "6sILUVbHoSgH", + "colab_type": "text" + }, + "cell_type": "markdown", "source": [ - "# An introduction to TensorFlow\n", - "\n", "This is an introductory tutorial for using TensorFlow. It will cover:\n", "\n", "* Importing required packages\n", "* Creating and using Tensors\n", - "* Using GPU acceleration\n" + "* Using GPU acceleration\n", + "* Datasets" ] }, { - "cell_type": "markdown", "metadata": { - "colab_type": "text", - "id": "z1JcS5iBXMRO" + "id": "z1JcS5iBXMRO", + "colab_type": "text" }, + "cell_type": "markdown", "source": [ "## Import TensorFlow\n", "\n", @@ -30,32 +110,32 @@ ] }, { - "cell_type": "code", - "execution_count": 0, "metadata": { - "cellView": "code", + "id": "RlIWhyeLoYnG", + "colab_type": "code", "colab": { "autoexec": { "startup": false, "wait_interval": 0 } }, - "colab_type": "code", - "id": "RlIWhyeLoYnG" + "cellView": "code" }, - "outputs": [], + "cell_type": "code", "source": [ "import tensorflow as tf\n", "\n", "tf.enable_eager_execution()" - ] + ], + "execution_count": 0, + "outputs": [] }, { - "cell_type": "markdown", "metadata": { - "colab_type": "text", - "id": "H9UySOPLXdaw" + "id": "H9UySOPLXdaw", + "colab_type": "text" }, + "cell_type": "markdown", "source": [ "## Tensors\n", "\n", @@ -63,46 +143,18 @@ ] }, { - "cell_type": "code", - "execution_count": 0, "metadata": { - "cellView": "code", + "id": "ngUe237Wt48W", + "colab_type": "code", "colab": { "autoexec": { "startup": false, "wait_interval": 0 - }, - "height": 125 - }, - "colab_type": "code", - "executionInfo": { - "elapsed": 320, - "status": "ok", - "timestamp": 1526420535530, - "user": { - "displayName": "", - "photoUrl": "", - "userId": "" - }, - "user_tz": 420 + } }, - "id": "ngUe237Wt48W", - "outputId": "b1a1cd60-4eb3-443d-cd6b-68406390784e" + "cellView": "code" }, - "outputs": [ - { - "name": "stdout", - "output_type": "stream", - "text": [ - "tf.Tensor(3, shape=(), dtype=int32)\n", - "tf.Tensor([4 6], shape=(2,), dtype=int32)\n", - "tf.Tensor(25, shape=(), dtype=int32)\n", - "tf.Tensor(6, shape=(), dtype=int32)\n", - "tf.Tensor(aGVsbG8gd29ybGQ, shape=(), dtype=string)\n", - "tf.Tensor(13, shape=(), dtype=int32)\n" - ] - } - ], + "cell_type": "code", "source": [ "print(tf.add(1, 2))\n", "print(tf.add([1, 2], [3, 4]))\n", @@ -112,66 +164,46 @@ "\n", "# Operator overloading is also supported\n", "print(tf.square(2) + tf.square(3))" - ] + ], + "execution_count": 0, + "outputs": [] }, { - "cell_type": "markdown", "metadata": { - "colab_type": "text", - "id": "IDY4WsYRhP81" + "id": "IDY4WsYRhP81", + "colab_type": "text" }, + "cell_type": "markdown", "source": [ "Each Tensor has a shape and a datatype" ] }, { - "cell_type": "code", - "execution_count": 0, "metadata": { + "id": "srYWH1MdJNG7", + "colab_type": "code", "colab": { "autoexec": { "startup": false, "wait_interval": 0 - }, - "height": 53 - }, - "colab_type": "code", - "executionInfo": { - "elapsed": 215, - "status": "ok", - "timestamp": 1526420538162, - "user": { - "displayName": "", - "photoUrl": "", - "userId": "" - }, - "user_tz": 420 - }, - "id": "srYWH1MdJNG7", - "outputId": "5e4ac41c-5115-4e50-eba0-42e249c16561" - }, - "outputs": [ - { - "name": "stdout", - "output_type": "stream", - "text": [ - "(1, 2)\n", - "\u003cdtype: 'int32'\u003e\n" - ] + } } - ], + }, + "cell_type": "code", "source": [ "x = tf.matmul([[1]], [[2, 3]])\n", "print(x.shape)\n", "print(x.dtype)" - ] + ], + "execution_count": 0, + "outputs": [] }, { - "cell_type": "markdown", "metadata": { - "colab_type": "text", - "id": "eBPw8e8vrsom" + "id": "eBPw8e8vrsom", + "colab_type": "text" }, + "cell_type": "markdown", "source": [ "The most obvious differences between NumPy arrays and TensorFlow Tensors are:\n", "\n", @@ -180,11 +212,11 @@ ] }, { - "cell_type": "markdown", "metadata": { - "colab_type": "text", - "id": "Dwi1tdW3JBw6" + "id": "Dwi1tdW3JBw6", + "colab_type": "text" }, + "cell_type": "markdown", "source": [ "### NumPy Compatibility\n", "\n", @@ -197,52 +229,17 @@ ] }, { - "cell_type": "code", - "execution_count": 0, "metadata": { + "id": "lCUWzso6mbqR", + "colab_type": "code", "colab": { "autoexec": { "startup": false, "wait_interval": 0 - }, - "height": 251 - }, - "colab_type": "code", - "executionInfo": { - "elapsed": 238, - "status": "ok", - "timestamp": 1526420540562, - "user": { - "displayName": "", - "photoUrl": "", - "userId": "" - }, - "user_tz": 420 - }, - "id": "lCUWzso6mbqR", - "outputId": "fd0a22bc-8249-49dd-fcbd-63161cc47e46" - }, - "outputs": [ - { - "name": "stdout", - "output_type": "stream", - "text": [ - "TensorFlow operations convert numpy arrays to Tensors automatically\n", - "tf.Tensor(\n", - "[[ 42. 42. 42.]\n", - " [ 42. 42. 42.]\n", - " [ 42. 42. 42.]], shape=(3, 3), dtype=float64)\n", - "And NumPy operations convert Tensors to numpy arrays automatically\n", - "[[ 43. 43. 43.]\n", - " [ 43. 43. 43.]\n", - " [ 43. 43. 43.]]\n", - "The .numpy() method explicitly converts a Tensor to a numpy array\n", - "[[ 42. 42. 42.]\n", - " [ 42. 42. 42.]\n", - " [ 42. 42. 42.]]\n" - ] + } } - ], + }, + "cell_type": "code", "source": [ "import numpy as np\n", "\n", @@ -258,14 +255,16 @@ "\n", "print(\"The .numpy() method explicitly converts a Tensor to a numpy array\")\n", "print(tensor.numpy())" - ] + ], + "execution_count": 0, + "outputs": [] }, { - "cell_type": "markdown", "metadata": { - "colab_type": "text", - "id": "PBNP8yTRfu_X" + "id": "PBNP8yTRfu_X", + "colab_type": "text" }, + "cell_type": "markdown", "source": [ "## GPU acceleration\n", "\n", @@ -273,42 +272,18 @@ ] }, { - "cell_type": "code", - "execution_count": 0, "metadata": { - "cellView": "code", + "id": "3Twf_Rw-gQFM", + "colab_type": "code", "colab": { "autoexec": { "startup": false, "wait_interval": 0 - }, - "height": 53 - }, - "colab_type": "code", - "executionInfo": { - "elapsed": 340, - "status": "ok", - "timestamp": 1526420543562, - "user": { - "displayName": "", - "photoUrl": "", - "userId": "" - }, - "user_tz": 420 + } }, - "id": "3Twf_Rw-gQFM", - "outputId": "2239ae2b-adf3-4895-b1f3-464cf5361d1b" + "cellView": "code" }, - "outputs": [ - { - "name": "stdout", - "output_type": "stream", - "text": [ - "Is there a GPU available: False\n", - "Is the Tensor on GPU #0: False\n" - ] - } - ], + "cell_type": "code", "source": [ "x = tf.random_uniform([3, 3])\n", "\n", @@ -317,26 +292,28 @@ "\n", "print(\"Is the Tensor on GPU #0: \"),\n", "print(x.device.endswith('GPU:0'))" - ] + ], + "execution_count": 0, + "outputs": [] }, { - "cell_type": "markdown", "metadata": { - "colab_type": "text", - "id": "vpgYzgVXW2Ud" + "id": "vpgYzgVXW2Ud", + "colab_type": "text" }, + "cell_type": "markdown", "source": [ "### Device Names\n", "\n", - "The `Tensor.device` property provides a fully qualified string name of the device hosting the contents of the Tensor. This name encodes a bunch of details, such as an identifier of the network address of the host on which this program is executing and the device within that host. This is required for distributed execution of TensorFlow programs, but we'll skip that for now. The string will end with `GPU:\u003cN\u003e` if the tensor is placed on the `N`-th tensor on the host." + "The `Tensor.device` property provides a fully qualified string name of the device hosting the contents of the Tensor. This name encodes a bunch of details, such as an identifier of the network address of the host on which this program is executing and the device within that host. This is required for distributed execution of TensorFlow programs, but we'll skip that for now. The string will end with `GPU:` if the tensor is placed on the `N`-th tensor on the host." ] }, { - "cell_type": "markdown", "metadata": { - "colab_type": "text", - "id": "ZWZQCimzuqyP" + "id": "ZWZQCimzuqyP", + "colab_type": "text" }, + "cell_type": "markdown", "source": [ "\n", "\n", @@ -346,41 +323,17 @@ ] }, { - "cell_type": "code", - "execution_count": 0, "metadata": { + "id": "RjkNZTuauy-Q", + "colab_type": "code", "colab": { "autoexec": { "startup": false, "wait_interval": 0 - }, - "height": 53 - }, - "colab_type": "code", - "executionInfo": { - "elapsed": 1762, - "status": "ok", - "timestamp": 1526420547562, - "user": { - "displayName": "", - "photoUrl": "", - "userId": "" - }, - "user_tz": 420 - }, - "id": "RjkNZTuauy-Q", - "outputId": "2e613293-ccac-4db2-b793-8ceb5b5adcfd" - }, - "outputs": [ - { - "name": "stdout", - "output_type": "stream", - "text": [ - "On CPU:\n", - "10 loops, best of 3: 35.8 ms per loop\n" - ] + } } - ], + }, + "cell_type": "code", "source": [ "def time_matmul(x):\n", " %timeit tf.matmul(x, x)\n", @@ -398,32 +351,141 @@ " x = tf.random_uniform([1000, 1000])\n", " assert x.device.endswith(\"GPU:0\")\n", " time_matmul(x)" + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "o1K4dlhhHtQj", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "## Datasets\n", + "\n", + "This section demonstrates the use of the [`tf.data.Dataset` API](https://www.tensorflow.org/guide/datasets) to build pipelines to feed data to your model. It covers:\n", + "\n", + "* Creating a `Dataset`.\n", + "* Iteration over a `Dataset` with eager execution enabled.\n", + "\n", + "We recommend using the `Dataset`s API for building performant, complex input pipelines from simple, re-usable pieces that will feed your model's training or evaluation loops.\n", + "\n", + "If you're familiar with TensorFlow graphs, the API for constructing the `Dataset` object remains exactly the same when eager execution is enabled, but the process of iterating over elements of the dataset is slightly simpler.\n", + "You can use Python iteration over the `tf.data.Dataset` object and do not need to explicitly create an `tf.data.Iterator` object.\n", + "As a result, the discussion on iterators in the [TensorFlow Guide](https://www.tensorflow.org/guide/datasets) is not relevant when eager execution is enabled." ] }, { + "metadata": { + "id": "zI0fmOynH-Ne", + "colab_type": "text" + }, "cell_type": "markdown", + "source": [ + "### Create a source `Dataset`\n", + "\n", + "Create a _source_ dataset using one of the factory functions like [`Dataset.from_tensors`](https://www.tensorflow.org/api_docs/python/tf/data/Dataset#from_tensors), [`Dataset.from_tensor_slices`](https://www.tensorflow.org/api_docs/python/tf/data/Dataset#from_tensor_slices) or using objects that read from files like [`TextLineDataset`](https://www.tensorflow.org/api_docs/python/tf/data/TextLineDataset) or [`TFRecordDataset`](https://www.tensorflow.org/api_docs/python/tf/data/TFRecordDataset). See the [TensorFlow Guide](https://www.tensorflow.org/guide/datasets#reading_input_data) for more information." + ] + }, + { "metadata": { - "colab_type": "text", - "id": "YEOJTNiOvnpQ" + "id": "F04fVOHQIBiG", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + } }, + "cell_type": "code", "source": [ - "## Next Steps\n", + "ds_tensors = tf.data.Dataset.from_tensor_slices([1, 2, 3, 4, 5, 6])\n", "\n", - "In this tutorial we covered the most fundamental concepts in TensorFlow - `Tensor`s, operations, and devices.\n", - "In [the next tutorial](https://github.com/tensorflow/models/tree/master/official/contrib/eager/python/examples/notebooks/2_gradients.ipynb) we will cover automatic differentiation - a building block required for training many machine learning models like neural networks." + "# Create a CSV file\n", + "import tempfile\n", + "_, filename = tempfile.mkstemp()\n", + "\n", + "with open(filename, 'w') as f:\n", + " f.write(\"\"\"Line 1\n", + "Line 2\n", + "Line 3\n", + " \"\"\")\n", + "\n", + "ds_file = tf.data.TextLineDataset(filename)" + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "vbxIhC-5IPdf", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "### Apply transformations\n", + "\n", + "Use the transformations functions like [`map`](https://www.tensorflow.org/api_docs/python/tf/data/Dataset#map), [`batch`](https://www.tensorflow.org/api_docs/python/tf/data/Dataset#batch), [`shuffle`](https://www.tensorflow.org/api_docs/python/tf/data/Dataset#shuffle) etc. to apply transformations to the records of the dataset. See the [API documentation for `tf.data.Dataset`](https://www.tensorflow.org/api_docs/python/tf/data/Dataset) for details." ] + }, + { + "metadata": { + "id": "uXSDZWE-ISsd", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + } + }, + "cell_type": "code", + "source": [ + "ds_tensors = ds_tensors.map(tf.square).shuffle(2).batch(2)\n", + "\n", + "ds_file = ds_file.batch(2)" + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "A8X1GNfoIZKJ", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "### Iterate\n", + "\n", + "When eager execution is enabled `Dataset` objects support iteration.\n", + "If you're familiar with the use of `Dataset`s in TensorFlow graphs, note that there is no need for calls to `Dataset.make_one_shot_iterator()` or `get_next()` calls." + ] + }, + { + "metadata": { + "id": "ws-WKRk5Ic6-", + "colab_type": "code", + "colab": { + "autoexec": { + "startup": false, + "wait_interval": 0 + } + } + }, + "cell_type": "code", + "source": [ + "print('Elements of ds_tensors:')\n", + "for x in ds_tensors:\n", + " print(x)\n", + "\n", + "print('\\nElements in ds_file:')\n", + "for x in ds_file:\n", + " print(x)" + ], + "execution_count": 0, + "outputs": [] } - ], - "metadata": { - "colab": { - "collapsed_sections": [], - "default_view": {}, - "name": "TensorFlow: An introduction", - "provenance": [], - "version": "0.3.2", - "views": {} - } - }, - "nbformat": 4, - "nbformat_minor": 0 -} + ] +} \ No newline at end of file diff --git a/tensorflow/contrib/eager/python/examples/resnet50/resnet50_test.py b/tensorflow/contrib/eager/python/examples/resnet50/resnet50_test.py index b14ef1df8ff4c660b9b6f2abfd5df6572d10b1e8..07d8788882c2d831dfb041fe7409af51857190bf 100644 --- a/tensorflow/contrib/eager/python/examples/resnet50/resnet50_test.py +++ b/tensorflow/contrib/eager/python/examples/resnet50/resnet50_test.py @@ -29,6 +29,7 @@ import tensorflow.contrib.eager as tfe from tensorflow.contrib.eager.python.examples.resnet50 import resnet50 from tensorflow.contrib.summary import summary_test_util from tensorflow.python.client import device_lib +from tensorflow.python.eager import tape def device_and_data_format(): @@ -49,13 +50,21 @@ def random_batch(batch_size, data_format): return images, one_hot -def compute_gradients(model, images, labels): - with tf.GradientTape() as tape: +def compute_gradients(model, images, labels, num_replicas=1): + with tf.GradientTape() as grad_tape: logits = model(images, training=True) loss = tf.losses.softmax_cross_entropy( logits=logits, onehot_labels=labels) tf.contrib.summary.scalar(name='loss', tensor=loss) - return tape.gradient(loss, model.variables) + if num_replicas != 1: + loss /= num_replicas + + # TODO(b/110991947): We can mistakenly trace the gradient call in + # multi-threaded environment. Explicitly disable recording until + # this is fixed. + with tape.stop_recording(): + grads = grad_tape.gradient(loss, model.variables) + return grads def apply_gradients(model, optimizer, gradients): @@ -188,11 +197,14 @@ class ResNet50Benchmarks(tf.test.Benchmark): return (32,) return (16, 32) - def _report(self, label, start, num_iters, device, batch_size, data_format): + def _report(self, label, start, num_iters, device, batch_size, data_format, + num_replicas=1): avg_time = (time.time() - start) / num_iters dev = tf.DeviceSpec.from_string(device).device_type.lower() - name = '%s_%s_batch_%d_%s' % (label, dev, batch_size, data_format) - extras = {'examples_per_sec': batch_size / avg_time} + replica_str = '' if num_replicas == 1 else 'replicas_%d_' % num_replicas + name = '%s_%s_batch_%d_%s%s' % (label, dev, batch_size, + replica_str, data_format) + extras = {'examples_per_sec': (num_replicas * batch_size) / avg_time} self.report_benchmark( iters=num_iters, wall_time=avg_time, name=name, extras=extras) diff --git a/tensorflow/contrib/eager/python/examples/revnet/BUILD b/tensorflow/contrib/eager/python/examples/revnet/BUILD index 432bb546f83932d0e0a465d7af7c641b60d2e564..0c0e4c0eb9d1a97eaca0fafa5df2517ef644fc95 100644 --- a/tensorflow/contrib/eager/python/examples/revnet/BUILD +++ b/tensorflow/contrib/eager/python/examples/revnet/BUILD @@ -72,11 +72,13 @@ cuda_py_test( size = "large", srcs = ["revnet_test.py"], additional_deps = [ + ":blocks_test", ":config", ":revnet", "//tensorflow:tensorflow_py", ], tags = [ + "no_pip", # depends on blocks_test, which is not available in pip package "optonly", ], ) @@ -87,7 +89,6 @@ py_library( srcs = ["cifar_input.py"], srcs_version = "PY2AND3", deps = [ - ":revnet", "//tensorflow:tensorflow_py", ], ) diff --git a/tensorflow/contrib/eager/python/examples/revnet/README.md b/tensorflow/contrib/eager/python/examples/revnet/README.md new file mode 100644 index 0000000000000000000000000000000000000000..21fc44febc8abdc30daad1b35d8434b083360bdf --- /dev/null +++ b/tensorflow/contrib/eager/python/examples/revnet/README.md @@ -0,0 +1,45 @@ +# RevNet with TensorFlow eager execution + +This folder contains an TensorFlow eager implementation of the [Reversible Residual Network](https://arxiv.org/pdf/1707.04585.pdf) adapted from the released implementation by the authors. The presented implementation can be ran both in eager and graph mode. The code is considerably simplified with `tf.GradientTape`. Moreover, we reduce the step of reconstructing the outputs. This saves us from using `tf.stop_gradient` and makes the model run faster. + +## Content + +- `revnet.py`: The RevNet model. +- `blocks.py`: The relevant reversible blocks. +- `cifar_tfrecords.py`: Script to generate the TFRecords for both CIFAR-10 and CIFAR-100. +- `cifar_input.py`: Script to read from TFRecords and generate dataset objects with the `tf.data` API. +- `config.py`: Configuration file for network architectures and training hyperparameters. +- `main.py`: Main training and evaluation script. +- `ops.py`: Auxiliary downsampling operation. + +## To run +- Make sure you have installed TensorFlow 1.9+ or the latest `tf-nightly` +or `tf-nightly-gpu` pip package in order to access the eager execution feature. + +- First run + +```bash +python cifar_tfrecords.py --data_dir ${PWD}/cifar +``` +to download the cifar dataset and convert them +to TFRecords. This produces TFRecord files for both CIFAR-10 and CIFAR-100. + +- To train a model run + +```bash +python main.py --data_dir ${PWD}/cifar +``` + +- Optional arguments for `main.py` include + - `train_dir`: Directory to store eventfiles and checkpoints. + - `restore`: Restore the latest checkpoint. + - `validate`: Use validation set for training monitoring. + - `manual_grad`: Use the manually defined gradient map given by the authors. + - `dataset`: Use either `cifar-10` or `cifar-100` + +## Performance +- With the current implementation, RevNet-38 achieves >92% on CIFAR-10 and >71% on CIFAR-100. + +## Reference +The Reversible Residual Network: Backpropagation Without Storing Activations. +Aidan N. Gomez, Mengye Ren, Raquel Urtasun, Roger B. Grosse. Neural Information Processing Systems (NIPS), 2017. diff --git a/tensorflow/contrib/eager/python/examples/revnet/blocks.py b/tensorflow/contrib/eager/python/examples/revnet/blocks.py index af41f6428660dd6b80e1a28f7e70021fe260a9b5..306096e9f8c4da0ed7f893ae75067cd24e7274b1 100644 --- a/tensorflow/contrib/eager/python/examples/revnet/blocks.py +++ b/tensorflow/contrib/eager/python/examples/revnet/blocks.py @@ -43,7 +43,8 @@ class RevBlock(tf.keras.Model): batch_norm_first=False, data_format="channels_first", bottleneck=False, - fused=True): + fused=True, + dtype=tf.float32): """Initialize RevBlock. Args: @@ -55,6 +56,7 @@ class RevBlock(tf.keras.Model): data_format: tensor data format, "NCHW"/"NHWC" bottleneck: use bottleneck residual if True fused: use fused batch normalization if True + dtype: float16, float32, or float64 """ super(RevBlock, self).__init__() self.blocks = tf.contrib.checkpoint.List() @@ -68,7 +70,8 @@ class RevBlock(tf.keras.Model): batch_norm_first=curr_batch_norm_first, data_format=data_format, bottleneck=bottleneck, - fused=fused) + fused=fused, + dtype=dtype) self.blocks.append(block) if data_format == "channels_first": @@ -93,11 +96,23 @@ class RevBlock(tf.keras.Model): for i in reversed(range(len(self.blocks))): block = self.blocks[i] - y_inv = x if i == 0 else block.backward(y, training=training) - dy, grads, vars_ = block.backward_grads_and_vars( - y_inv, dy, training=training) - grads_all += grads - vars_all += vars_ + if i == 0: + # First block usually contains downsampling that can't be reversed + with tf.GradientTape() as tape: + x = tf.identity(x) + tape.watch(x) + y = block(x, training=training) + + grads_combined = tape.gradient( + y, [x] + block.trainable_variables, output_gradients=dy) + dy = grads_combined[0] + grads_all += grads_combined[1:] + vars_all += block.trainable_variables + else: + y, dy, grads, vars_ = block.backward_grads_and_vars( + y, dy, training=training) + grads_all += grads + vars_all += vars_ return dy, grads_all, vars_all @@ -115,6 +130,7 @@ class _Residual(tf.keras.Model): data_format: tensor data format, "NCHW"/"NHWC", bottleneck: use bottleneck residual if True fused: use fused batch normalization if True + dtype: float16, float32, or float64 """ def __init__(self, @@ -124,7 +140,8 @@ class _Residual(tf.keras.Model): batch_norm_first=True, data_format="channels_first", bottleneck=False, - fused=True): + fused=True, + dtype=tf.float32): super(_Residual, self).__init__() self.filters = filters @@ -146,75 +163,68 @@ class _Residual(tf.keras.Model): input_shape=f_input_shape, batch_norm_first=batch_norm_first, data_format=data_format, - fused=fused) + fused=fused, + dtype=dtype) self.g = factory( filters=filters // 2, strides=(1, 1), input_shape=g_input_shape, batch_norm_first=batch_norm_first, data_format=data_format, - fused=fused) + fused=fused, + dtype=dtype) def call(self, x, training=True, concat=True): """Apply residual block to inputs.""" x1, x2 = tf.split(x, num_or_size_splits=2, axis=self.axis) - f_x2 = self.f.call(x2, training=training) - # TODO(lxuechen): Replace with simpler downsampling + f_x2 = self.f(x2, training=training) x1_down = ops.downsample( x1, self.filters // 2, self.strides, axis=self.axis) x2_down = ops.downsample( x2, self.filters // 2, self.strides, axis=self.axis) y1 = f_x2 + x1_down - g_y1 = self.g.call(y1, training=training) # self.g(y1) gives pylint error + g_y1 = self.g(y1, training=training) y2 = g_y1 + x2_down - if not concat: # Concat option needed for correct backward grads + if not concat: # For correct backward grads return y1, y2 - return tf.concat([y1, y2], axis=self.axis) - - def backward(self, y, training=True): - """Reconstruct inputs from outputs; only valid when stride 1.""" - - assert self.strides == (1, 1) - - y1, y2 = tf.split(y, num_or_size_splits=2, axis=self.axis) - g_y1 = self.g.call(y1, training=training) - x2 = y2 - g_y1 - f_x2 = self.f.call(x2, training=training) - x1 = y1 - f_x2 - return tf.concat([x1, x2], axis=self.axis) + return tf.concat([y1, y2], axis=self.axis) - def backward_grads_and_vars(self, x, dy, training=True): + def backward_grads_and_vars(self, y, dy, training=True): """Manually compute backward gradients given input and output grads.""" + dy1, dy2 = tf.split(dy, num_or_size_splits=2, axis=self.axis) with tf.GradientTape(persistent=True) as tape: - x = tf.identity(x) # TODO(lxuechen): Remove after b/110264016 is fixed - x1, x2 = tf.split(x, num_or_size_splits=2, axis=self.axis) - tape.watch([x1, x2]) - # Stitch back x for `call` so tape records correct grads - x = tf.concat([x1, x2], axis=self.axis) - dy1, dy2 = tf.split(dy, num_or_size_splits=2, axis=self.axis) - y1, y2 = self.call(x, training=training, concat=False) - x2_down = ops.downsample( - x2, self.filters // 2, self.strides, axis=self.axis) + y = tf.identity(y) + tape.watch(y) + y1, y2 = tf.split(y, num_or_size_splits=2, axis=self.axis) + z1 = y1 + gz1 = self.g(z1, training=training) + x2 = y2 - gz1 + fx2 = self.f(x2, training=training) + x1 = z1 - fx2 grads_combined = tape.gradient( - y2, [y1] + self.g.trainable_variables, output_gradients=[dy2]) - dy2_y1, dg = grads_combined[0], grads_combined[1:] - dy1_plus = dy2_y1 + dy1 + gz1, [z1] + self.g.trainable_variables, output_gradients=dy2) + dz1 = dy1 + grads_combined[0] + dg = grads_combined[1:] + dx1 = dz1 grads_combined = tape.gradient( - y1, [x1, x2] + self.f.trainable_variables, output_gradients=[dy1_plus]) - dx1, dx2, df = grads_combined[0], grads_combined[1], grads_combined[2:] - dx2 += tape.gradient(x2_down, [x2], output_gradients=[dy2])[0] + fx2, [x2] + self.f.trainable_variables, output_gradients=dz1) + dx2 = dy2 + grads_combined[0] + df = grads_combined[1:] del tape grads = df + dg vars_ = self.f.trainable_variables + self.g.trainable_variables - return tf.concat([dx1, dx2], axis=self.axis), grads, vars_ + x = tf.concat([x1, x2], axis=self.axis) + dx = tf.concat([dx1, dx2], axis=self.axis) + + return x, dx, grads, vars_ def _BottleneckResidualInner(filters, @@ -222,7 +232,8 @@ def _BottleneckResidualInner(filters, input_shape, batch_norm_first=True, data_format="channels_first", - fused=True): + fused=True, + dtype=tf.float32): """Single bottleneck residual inner function contained in _Resdual. Corresponds to the `F`/`G` functions in the paper. @@ -235,6 +246,7 @@ def _BottleneckResidualInner(filters, batch_norm_first: whether to apply activation and batch norm before conv data_format: tensor data format, "NCHW"/"NHWC" fused: use fused batch normalization if True + dtype: float16, float32, or float64 Returns: A keras model @@ -245,7 +257,7 @@ def _BottleneckResidualInner(filters, if batch_norm_first: model.add( tf.keras.layers.BatchNormalization( - axis=axis, input_shape=input_shape, fused=fused)) + axis=axis, input_shape=input_shape, fused=fused, dtype=dtype)) model.add(tf.keras.layers.Activation("relu")) model.add( tf.keras.layers.Conv2D( @@ -255,9 +267,11 @@ def _BottleneckResidualInner(filters, input_shape=input_shape, data_format=data_format, use_bias=False, - padding="SAME")) + padding="SAME", + dtype=dtype)) - model.add(tf.keras.layers.BatchNormalization(axis=axis, fused=fused)) + model.add( + tf.keras.layers.BatchNormalization(axis=axis, fused=fused, dtype=dtype)) model.add(tf.keras.layers.Activation("relu")) model.add( tf.keras.layers.Conv2D( @@ -266,9 +280,11 @@ def _BottleneckResidualInner(filters, strides=(1, 1), data_format=data_format, use_bias=False, - padding="SAME")) + padding="SAME", + dtype=dtype)) - model.add(tf.keras.layers.BatchNormalization(axis=axis, fused=fused)) + model.add( + tf.keras.layers.BatchNormalization(axis=axis, fused=fused, dtype=dtype)) model.add(tf.keras.layers.Activation("relu")) model.add( tf.keras.layers.Conv2D( @@ -277,7 +293,8 @@ def _BottleneckResidualInner(filters, strides=(1, 1), data_format=data_format, use_bias=False, - padding="SAME")) + padding="SAME", + dtype=dtype)) return model @@ -287,7 +304,8 @@ def _ResidualInner(filters, input_shape, batch_norm_first=True, data_format="channels_first", - fused=True): + fused=True, + dtype=tf.float32): """Single residual inner function contained in _ResdualBlock. Corresponds to the `F`/`G` functions in the paper. @@ -299,6 +317,7 @@ def _ResidualInner(filters, batch_norm_first: whether to apply activation and batch norm before conv data_format: tensor data format, "NCHW"/"NHWC" fused: use fused batch normalization if True + dtype: float16, float32, or float64 Returns: A keras model @@ -309,7 +328,7 @@ def _ResidualInner(filters, if batch_norm_first: model.add( tf.keras.layers.BatchNormalization( - axis=axis, input_shape=input_shape, fused=fused)) + axis=axis, input_shape=input_shape, fused=fused, dtype=dtype)) model.add(tf.keras.layers.Activation("relu")) model.add( tf.keras.layers.Conv2D( @@ -319,9 +338,11 @@ def _ResidualInner(filters, input_shape=input_shape, data_format=data_format, use_bias=False, - padding="SAME")) + padding="SAME", + dtype=dtype)) - model.add(tf.keras.layers.BatchNormalization(axis=axis, fused=fused)) + model.add( + tf.keras.layers.BatchNormalization(axis=axis, fused=fused, dtype=dtype)) model.add(tf.keras.layers.Activation("relu")) model.add( tf.keras.layers.Conv2D( @@ -330,6 +351,7 @@ def _ResidualInner(filters, strides=(1, 1), data_format=data_format, use_bias=False, - padding="SAME")) + padding="SAME", + dtype=dtype)) return model diff --git a/tensorflow/contrib/eager/python/examples/revnet/blocks_test.py b/tensorflow/contrib/eager/python/examples/revnet/blocks_test.py index f4436fd92506d54f1206fbfd424b897f9835657d..d74785c8fe1c170ee95172974141c1cfe18b9502 100644 --- a/tensorflow/contrib/eager/python/examples/revnet/blocks_test.py +++ b/tensorflow/contrib/eager/python/examples/revnet/blocks_test.py @@ -22,6 +22,27 @@ import tensorflow as tf from tensorflow.contrib.eager.python.examples.revnet import blocks +def compute_degree(g1, g2, eps=1e-7): + """Compute the degree between two vectors using their usual inner product.""" + + def _dot(u, v): + return tf.reduce_sum(u * v) + + g1_norm = tf.sqrt(_dot(g1, g1)) + g2_norm = tf.sqrt(_dot(g2, g2)) + if g1_norm.numpy() == 0 and g2_norm.numpy() == 0: + cosine = 1. - eps + else: + g1_norm = 1. if g1_norm.numpy() == 0 else g1_norm + g2_norm = 1. if g2_norm.numpy() == 0 else g2_norm + cosine = _dot(g1, g2) / g1_norm / g2_norm + # Restrict to arccos range + cosine = tf.minimum(tf.maximum(cosine, eps - 1.), 1. - eps) + degree = tf.acos(cosine) * 180. / 3.141592653589793 + + return degree + + def _validate_block_call_channels_last(block_factory, test): """Generic testing function for `channels_last` data format. @@ -33,30 +54,30 @@ def _validate_block_call_channels_last(block_factory, test): test: tf.test.TestCase object """ with tf.device("/cpu:0"): # NHWC format - input_shape = (224, 224, 32) + input_shape = (8, 8, 128) data_shape = (16,) + input_shape x = tf.random_normal(shape=data_shape) # Stride 1 block = block_factory( - filters=64, + filters=128, strides=(1, 1), input_shape=input_shape, data_format="channels_last") y_tr, y_ev = block(x, training=True), block(x, training=False) test.assertEqual(y_tr.shape, y_ev.shape) - test.assertEqual(y_ev.shape, (16, 224, 224, 64)) + test.assertEqual(y_ev.shape, (16, 8, 8, 128)) test.assertNotAllClose(y_tr, y_ev) # Stride of 2 block = block_factory( - filters=64, + filters=128, strides=(2, 2), input_shape=input_shape, data_format="channels_last") y_tr, y_ev = block(x, training=True), block(x, training=False) test.assertEqual(y_tr.shape, y_ev.shape) - test.assertEqual(y_ev.shape, (16, 112, 112, 64)) + test.assertEqual(y_ev.shape, (16, 4, 4, 128)) test.assertNotAllClose(y_tr, y_ev) @@ -74,22 +95,22 @@ def _validate_block_call_channels_first(block_factory, test): test.skipTest("GPU not available") with tf.device("/gpu:0"): # Default NCHW format - input_shape = (32, 224, 224) + input_shape = (128, 8, 8) data_shape = (16,) + input_shape x = tf.random_normal(shape=data_shape) # Stride of 1 - block = block_factory(filters=64, strides=(1, 1), input_shape=input_shape) + block = block_factory(filters=128, strides=(1, 1), input_shape=input_shape) y_tr, y_ev = block(x, training=True), block(x, training=False) test.assertEqual(y_tr.shape, y_ev.shape) - test.assertEqual(y_ev.shape, (16, 64, 224, 224)) + test.assertEqual(y_ev.shape, (16, 128, 8, 8)) test.assertNotAllClose(y_tr, y_ev) # Stride of 2 - block = block_factory(filters=64, strides=(2, 2), input_shape=input_shape) + block = block_factory(filters=128, strides=(2, 2), input_shape=input_shape) y_tr, y_ev = block(x, training=True), block(x, training=False) test.assertEqual(y_tr.shape, y_ev.shape) - test.assertEqual(y_ev.shape, (16, 64, 112, 112)) + test.assertEqual(y_ev.shape, (16, 128, 4, 4)) test.assertNotAllClose(y_tr, y_ev) @@ -101,121 +122,116 @@ class RevBlockTest(tf.test.TestCase): self.skipTest("GPU not available") with tf.device("/gpu:0"): # Default NCHW format - input_shape = (32, 224, 224) + input_shape = (128, 8, 8) data_shape = (16,) + input_shape x = tf.random_normal(shape=data_shape) # Stride of 1 block = blocks.RevBlock( - n_res=3, filters=64, strides=(1, 1), input_shape=input_shape) + n_res=3, filters=128, strides=(1, 1), input_shape=input_shape) y_tr, y_ev = block(x, training=True), block(x, training=False) self.assertEqual(y_tr.shape, y_ev.shape) - self.assertEqual(y_ev.shape, (16, 64, 224, 224)) + self.assertEqual(y_ev.shape, (16, 128, 8, 8)) self.assertNotAllClose(y_tr, y_ev) # Stride of 2 block = blocks.RevBlock( - n_res=3, filters=64, strides=(2, 2), input_shape=input_shape) + n_res=3, filters=128, strides=(2, 2), input_shape=input_shape) y_tr, y_ev = block(x, training=True), block(x, training=False) self.assertEqual(y_tr.shape, y_ev.shape) - self.assertEqual(y_ev.shape, [16, 64, 112, 112]) + self.assertEqual(y_ev.shape, [16, 128, 4, 4]) self.assertNotAllClose(y_tr, y_ev) def test_call_channels_last(self): """Test `call` function with `channels_last` data format.""" with tf.device("/cpu:0"): # NHWC format - input_shape = (224, 224, 32) + input_shape = (8, 8, 128) data_shape = (16,) + input_shape x = tf.random_normal(shape=data_shape) # Stride 1 block = blocks.RevBlock( n_res=3, - filters=64, + filters=128, strides=(1, 1), input_shape=input_shape, data_format="channels_last") y_tr, y_ev = block(x, training=True), block(x, training=False) self.assertEqual(y_tr.shape, y_ev.shape) - self.assertEqual(y_ev.shape, (16, 224, 224, 64)) + self.assertEqual(y_ev.shape, (16, 8, 8, 128)) self.assertNotAllClose(y_tr, y_ev) # Stride of 2 block = blocks.RevBlock( n_res=3, - filters=64, + filters=128, strides=(2, 2), input_shape=input_shape, data_format="channels_last") y_tr, y_ev = block(x, training=True), block(x, training=False) self.assertEqual(y_tr.shape, y_ev.shape) - self.assertEqual(y_ev.shape, (16, 112, 112, 64)) + self.assertEqual(y_ev.shape, (16, 4, 4, 128)) self.assertNotAllClose(y_tr, y_ev) + def _check_grad_angle(self, grads, grads_true, atol=1e0): + """Check the angle between two list of vectors are all close.""" + for g1, g2 in zip(grads, grads_true): + degree = compute_degree(g1, g2) + self.assertLessEqual(degree, atol) + def test_backward_grads_and_vars_channels_first(self): """Test `backward` function with `channels_first` data format.""" if not tf.test.is_gpu_available(): self.skipTest("GPU not available") with tf.device("/gpu:0"): # Default NCHW format - input_shape = (32, 224, 224) - data_shape = (16,) + input_shape - x = tf.random_normal(shape=data_shape) - # Stride 1 - y = tf.random_normal(shape=data_shape) - dy = tf.random_normal(shape=data_shape) - block = blocks.RevBlock( - n_res=3, filters=32, strides=(1, 1), input_shape=input_shape) - dy, grads, vars_ = block.backward_grads_and_vars(x, y, dy) - self.assertEqual(dy.shape, x.shape) - self.assertTrue(isinstance(grads, list)) - self.assertTrue(isinstance(vars_, list)) - - # Stride 2 - y = tf.random_normal(shape=(16, 32, 112, 112)) - dy = tf.random_normal(shape=(16, 32, 112, 112)) - block = blocks.RevBlock( - n_res=3, filters=32, strides=(2, 2), input_shape=input_shape) - dy, grads, vars_ = block.backward_grads_and_vars(x, y, dy) - self.assertEqual(dy.shape, x.shape) - self.assertTrue(isinstance(grads, list)) - self.assertTrue(isinstance(vars_, list)) - - def test_backward_grads_and_vars_channels_last(self): - """Test `backward` function with `channels_last` data format.""" - with tf.device("/cpu:0"): # NHWC format - input_shape = (224, 224, 32) + input_shape = (128, 8, 8) data_shape = (16,) + input_shape - x = tf.random_normal(shape=data_shape) - - # Stride 1 - y = tf.random_normal(shape=data_shape) - dy = tf.random_normal(shape=data_shape) + x = tf.random_normal(shape=data_shape, dtype=tf.float64) + dy = tf.random_normal(shape=data_shape, dtype=tf.float64) block = blocks.RevBlock( n_res=3, - filters=32, + filters=128, strides=(1, 1), input_shape=input_shape, - data_format="channels_last") - dy, grads, vars_ = block.backward_grads_and_vars(x, y, dy) - self.assertEqual(dy.shape, x.shape) - self.assertTrue(isinstance(grads, list)) - self.assertTrue(isinstance(vars_, list)) + fused=False, + dtype=tf.float64) + with tf.GradientTape() as tape: + tape.watch(x) + y = block(x, training=True) + # Compute grads from reconstruction + dx, dw, vars_ = block.backward_grads_and_vars(x, y, dy, training=True) + # Compute true grads + grads = tape.gradient(y, [x] + vars_, output_gradients=dy) + dx_true, dw_true = grads[0], grads[1:] + self.assertAllClose(dx_true, dx) + self.assertAllClose(dw_true, dw) + self._check_grad_angle(dx_true, dx) + self._check_grad_angle(dw_true, dw) # Stride 2 - y = tf.random_normal(shape=(16, 112, 112, 32)) - dy = tf.random_normal(shape=(16, 112, 112, 32)) + x = tf.random_normal(shape=data_shape, dtype=tf.float64) + dy = tf.random_normal(shape=(16, 128, 4, 4), dtype=tf.float64) block = blocks.RevBlock( n_res=3, - filters=32, + filters=128, strides=(2, 2), input_shape=input_shape, - data_format="channels_last") - dy, grads, vars_ = block.backward_grads_and_vars(x, y, dy) - self.assertEqual(dy.shape, x.shape) - self.assertTrue(isinstance(grads, list)) - self.assertTrue(isinstance(vars_, list)) + fused=False, + dtype=tf.float64) + with tf.GradientTape() as tape: + tape.watch(x) + y = block(x, training=True) + # Compute grads from reconstruction + dx, dw, vars_ = block.backward_grads_and_vars(x, y, dy, training=True) + # Compute true grads + grads = tape.gradient(y, [x] + vars_, output_gradients=dy) + dx_true, dw_true = grads[0], grads[1:] + self.assertAllClose(dx_true, dx) + self.assertAllClose(dw_true, dw) + self._check_grad_angle(dx_true, dx) + self._check_grad_angle(dw_true, dw) class _ResidualTest(tf.test.TestCase): @@ -229,98 +245,40 @@ class _ResidualTest(tf.test.TestCase): _validate_block_call_channels_first(blocks._Residual, self) _validate_block_call_channels_last(blocks._Residual, self) - def test_backward_channels_first(self): - """Test `backward` function with `channels_first` data format.""" - if not tf.test.is_gpu_available(): - self.skipTest("GPU not available") - - with tf.device("/gpu:0"): # Default NCHW format - input_shape = (16, 224, 224) - data_shape = (16,) + input_shape - x = tf.random_normal(shape=data_shape) - residual = blocks._Residual( - filters=16, strides=(1, 1), input_shape=input_shape) - y_tr, y_ev = residual(x, training=True), residual(x, training=False) - x_ = residual.backward(y_tr, training=True) - # The numerical loss is alarming; reconstructed inputs could differ from - # the original inputs often by more than 1e-3 - self.assertAllClose(x, x_, rtol=1e-01, atol=1e-01) - x_ = residual.backward(y_ev, training=False) - self.assertAllClose(x, x_, rtol=1e-01, atol=1e-01) - - def test_backward_channels_last(self): - """Test `backward` function with `channels_last` data format.""" - with tf.device("/cpu:0"): # NHWC format - input_shape = (224, 224, 16) - data_shape = (16,) + input_shape - x = tf.random_normal(shape=data_shape) - residual = blocks._Residual( - filters=16, - strides=(1, 1), - input_shape=input_shape, - data_format="channels_last") - y_tr, y_ev = residual(x, training=True), residual(x, training=False) - x_ = residual.backward(y_tr, training=True) - # Egregious numerical error - self.assertAllClose(x, x_, rtol=1e-01, atol=1e-01) - x_ = residual.backward(y_ev, training=False) - self.assertAllClose(x, x_, rtol=1e-01, atol=1e-01) - def test_backward_grads_and_vars_channels_first(self): """Test `backward_grads` function with `channels_first` data format.""" if not tf.test.is_gpu_available(): self.skipTest("GPU not available") with tf.device("/gpu:0"): # Default NCHW format - input_shape = (16, 224, 224) - data_shape = (16,) + input_shape - x = tf.random_normal(shape=data_shape) - dy = tf.random_normal(shape=data_shape) - residual = blocks._Residual( - filters=16, strides=(1, 1), input_shape=input_shape) - dx_tr, grads_tr, vars_tr = residual.backward_grads_and_vars( - x, dy=dy, training=True) - dx_ev, grads_ev, vars_ev = residual.backward_grads_and_vars( - x, dy=dy, training=False) - self.assertNotAllClose(dx_tr, dx_ev) - self.assertTrue(isinstance(grads_tr, list)) - self.assertTrue(isinstance(grads_ev, list)) - self.assertTrue(isinstance(vars_tr, list)) - self.assertTrue(isinstance(vars_ev, list)) - for grad_tr, var_tr, grad_ev, var_ev in zip(grads_tr, vars_tr, grads_ev, - vars_ev): - if grad_tr is not None: # Batch norm moving mean, var gives None grad - self.assertEqual(grad_tr.shape, grad_ev.shape) - self.assertEqual(var_tr.shape, var_ev.shape) - self.assertEqual(grad_tr.shape, var_tr.shape) - - def test_backward_grads_and_vars_channels_last(self): - """Test `backward_grads` function with `channels_last` data format.""" - with tf.device("/cpu:0"): # NHWC format - input_shape = (224, 224, 16) + input_shape = (128, 8, 8) data_shape = (16,) + input_shape - x = tf.random_normal(shape=data_shape) - dy = tf.random_normal(shape=data_shape) + # Use double precision for testing + x_true = tf.random_normal(shape=data_shape, dtype=tf.float64) + dy = tf.random_normal(shape=data_shape, dtype=tf.float64) residual = blocks._Residual( - filters=16, + filters=128, strides=(1, 1), input_shape=input_shape, - data_format="channels_last") - dx_tr, grads_tr, vars_tr = residual.backward_grads_and_vars( - x, dy=dy, training=True) - dx_ev, grads_ev, vars_ev = residual.backward_grads_and_vars( - x, dy=dy, training=False) - self.assertNotAllClose(dx_tr, dx_ev) - self.assertTrue(isinstance(grads_tr, list)) - self.assertTrue(isinstance(grads_ev, list)) - self.assertTrue(isinstance(vars_tr, list)) - self.assertTrue(isinstance(vars_ev, list)) - for grad_tr, var_tr, grad_ev, var_ev in zip(grads_tr, vars_tr, grads_ev, - vars_ev): - if grad_tr is not None: # Batch norm moving mean, var gives None grad - self.assertEqual(grad_tr.shape, grad_ev.shape) - self.assertEqual(var_tr.shape, var_ev.shape) - self.assertEqual(grad_tr.shape, var_tr.shape) + fused=False, + dtype=tf.float64) + + with tf.GradientTape() as tape: + x_true = tf.identity(x_true) + tape.watch(x_true) + y = residual(x_true, training=True) + + # Gradients computed due to reversibility + x, dx, dw, vars_ = residual.backward_grads_and_vars( + y, dy=dy, training=True) + + # True gradients computed by the tape + grads = tape.gradient(y, [x_true] + vars_, output_gradients=dy) + dx_true, dw_true = grads[0], grads[1:] + + self.assertAllClose(x_true, x) + self.assertAllClose(dx_true, dx) + self.assertAllClose(dw_true, dw) class _ResidualInnerTest(tf.test.TestCase): diff --git a/tensorflow/contrib/eager/python/examples/revnet/cifar_input.py b/tensorflow/contrib/eager/python/examples/revnet/cifar_input.py index 3bc69da5adae29e6b6f43ef5045eb0256e680fa4..b6d4c35bfd21f9d651c4f059c019cf2e585da8b2 100644 --- a/tensorflow/contrib/eager/python/examples/revnet/cifar_input.py +++ b/tensorflow/contrib/eager/python/examples/revnet/cifar_input.py @@ -26,8 +26,6 @@ import tensorflow as tf IMAGE_HEIGHT = 32 IMAGE_WIDTH = 32 NUM_CHANNEL = 3 -NUM_TRAIN_IMG = 50000 -NUM_TEST_IMG = 10000 def get_ds_from_tfrecords(data_dir, @@ -37,8 +35,8 @@ def get_ds_from_tfrecords(data_dir, epochs=None, shuffle=True, data_format="channels_first", - num_parallel_calls=4, - prefetch=True, + num_parallel_calls=12, + prefetch=0, div255=True, dtype=tf.float32): """Returns a tf.train.Dataset object from reading tfrecords. @@ -48,11 +46,12 @@ def get_ds_from_tfrecords(data_dir, split: "train", "validation", or "test" data_aug: Apply data augmentation if True batch_size: Batch size of dataset object - epochs: Number of epochs to repeat the dataset + epochs: Number of epochs to repeat the dataset; default `None` means + repeating indefinitely shuffle: Shuffle the dataset if True data_format: `channels_first` or `channels_last` num_parallel_calls: Number of threads for dataset preprocess - prefetch: Apply prefetch for the dataset if True + prefetch: Buffer size for prefetch div255: Divide the images by 255 if True dtype: Data type of images Returns: @@ -62,7 +61,7 @@ def get_ds_from_tfrecords(data_dir, ValueError: Unknown split """ - if split not in ["train", "validation", "test"]: + if split not in ["train", "validation", "test", "train_all"]: raise ValueError("Unknown split {}".format(split)) def _parser(serialized_example): @@ -74,7 +73,11 @@ def get_ds_from_tfrecords(data_dir, "label": tf.FixedLenFeature([], tf.int64), }) image = tf.decode_raw(features["image"], tf.uint8) - image = tf.reshape(image, [IMAGE_HEIGHT, IMAGE_WIDTH, NUM_CHANNEL]) + # Initially reshaping to [H, W, C] does not work + image = tf.reshape(image, [NUM_CHANNEL, IMAGE_HEIGHT, IMAGE_WIDTH]) + # This is needed for `tf.image.resize_image_with_crop_or_pad` + image = tf.transpose(image, [1, 2, 0]) + image = tf.cast(image, dtype) label = tf.cast(features["label"], tf.int32) @@ -93,13 +96,21 @@ def get_ds_from_tfrecords(data_dir, return image, label filename = os.path.join(data_dir, split + ".tfrecords") - dataset = tf.data.TFRecordDataset(filename).repeat(epochs) + dataset = tf.data.TFRecordDataset(filename) + dataset = dataset.repeat(epochs) dataset = dataset.map(_parser, num_parallel_calls=num_parallel_calls) + dataset = dataset.prefetch(prefetch) - if prefetch: - dataset = dataset.prefetch(batch_size) if shuffle: - dataset = dataset.shuffle(NUM_TRAIN_IMG) + # Find the right size according to the split + size = { + "train": 40000, + "validation": 10000, + "test": 10000, + "train_all": 50000 + }[split] + dataset = dataset.shuffle(size) + dataset = dataset.batch(batch_size) return dataset diff --git a/tensorflow/contrib/eager/python/examples/revnet/cifar_tfrecords.py b/tensorflow/contrib/eager/python/examples/revnet/cifar_tfrecords.py index f79428b2a97f0ac2ce991f4c26b9123cddc24325..377844ad8fbca92629a4d71f5df2aab67b570c3c 100644 --- a/tensorflow/contrib/eager/python/examples/revnet/cifar_tfrecords.py +++ b/tensorflow/contrib/eager/python/examples/revnet/cifar_tfrecords.py @@ -12,10 +12,10 @@ # See the License for the specific language governing permissions and # limitations under the License. # ============================================================================== -"""Read CIFAR-10 data from pickled numpy arrays and writes TFRecords. +"""Read CIFAR data from pickled numpy arrays and writes TFRecords. Generates tf.train.Example protos and writes them to TFRecord files from the -python version of the CIFAR-10 dataset downloaded from +python version of the CIFAR dataset downloaded from https://www.cs.toronto.edu/~kriz/cifar.html. """ @@ -32,20 +32,22 @@ from six.moves import cPickle as pickle from six.moves import urllib import tensorflow as tf -CIFAR_FILENAME = 'cifar-10-python.tar.gz' -CIFAR_DOWNLOAD_URL = 'https://www.cs.toronto.edu/~kriz/' + CIFAR_FILENAME -CIFAR_LOCAL_FOLDER = 'cifar-10-batches-py' +BASE_URL = 'https://www.cs.toronto.edu/~kriz/' +CIFAR_FILE_NAMES = ['cifar-10-python.tar.gz', 'cifar-100-python.tar.gz'] +CIFAR_DOWNLOAD_URLS = [BASE_URL + name for name in CIFAR_FILE_NAMES] +CIFAR_LOCAL_FOLDERS = ['cifar-10', 'cifar-100'] +EXTRACT_FOLDERS = ['cifar-10-batches-py', 'cifar-100-python'] -def download_and_extract(data_dir): - """Download CIFAR-10 if not already downloaded.""" - filepath = os.path.join(data_dir, CIFAR_FILENAME) +def download_and_extract(data_dir, file_name, url): + """Download CIFAR if not already downloaded.""" + filepath = os.path.join(data_dir, file_name) if tf.gfile.Exists(filepath): return filepath if not tf.gfile.Exists(data_dir): tf.gfile.MakeDirs(data_dir) - urllib.request.urlretrieve(CIFAR_DOWNLOAD_URL, filepath) + urllib.request.urlretrieve(url, filepath) tarfile.open(os.path.join(filepath), 'r:gz').extractall(data_dir) return filepath @@ -58,12 +60,22 @@ def _bytes_feature(value): return tf.train.Feature(bytes_list=tf.train.BytesList(value=[value])) -def _get_file_names(): +def _get_file_names(folder): """Returns the file names expected to exist in the input_dir.""" + assert folder in ['cifar-10', 'cifar-100'] + file_names = {} - file_names['train'] = ['data_batch_%d' % i for i in range(1, 5)] - file_names['validation'] = ['data_batch_5'] - file_names['test'] = ['test_batch'] + if folder == 'cifar-10': + file_names['train'] = ['data_batch_%d' % i for i in range(1, 5)] + file_names['validation'] = ['data_batch_5'] + file_names['train_all'] = ['data_batch_%d' % i for i in range(1, 6)] + file_names['test'] = ['test_batch'] + else: + file_names['train_all'] = ['train'] + file_names['test'] = ['test'] + # Split in `convert_to_tfrecord` function + file_names['train'] = ['train'] + file_names['validation'] = ['train'] return file_names @@ -76,14 +88,28 @@ def read_pickle_from_file(filename): return data_dict -def convert_to_tfrecord(input_files, output_file): +def convert_to_tfrecord(input_files, output_file, folder): """Converts files with pickled data to TFRecords.""" + assert folder in ['cifar-10', 'cifar-100'] + print('Generating %s' % output_file) with tf.python_io.TFRecordWriter(output_file) as record_writer: for input_file in input_files: data_dict = read_pickle_from_file(input_file) data = data_dict[b'data'] - labels = data_dict[b'labels'] + try: + labels = data_dict[b'labels'] + except KeyError: + labels = data_dict[b'fine_labels'] + + if folder == 'cifar-100' and input_file.endswith('train.tfrecords'): + data = data[:40000] + labels = labels[:40000] + elif folder == 'cifar-100' and input_file.endswith( + 'validation.tfrecords'): + data = data[40000:] + labels = labels[40000:] + num_entries_in_batch = len(labels) for i in range(num_entries_in_batch): @@ -97,19 +123,24 @@ def convert_to_tfrecord(input_files, output_file): def main(_): - print('Download from {} and extract.'.format(CIFAR_DOWNLOAD_URL)) - download_and_extract(FLAGS.data_dir) - file_names = _get_file_names() - input_dir = os.path.join(FLAGS.data_dir, CIFAR_LOCAL_FOLDER) - - for mode, files in file_names.items(): - input_files = [os.path.join(input_dir, f) for f in files] - output_file = os.path.join(FLAGS.data_dir, mode + '.tfrecords') - try: - os.remove(output_file) - except OSError: - pass - convert_to_tfrecord(input_files, output_file) + for file_name, url, folder, extract_folder in zip( + CIFAR_FILE_NAMES, CIFAR_DOWNLOAD_URLS, CIFAR_LOCAL_FOLDERS, + EXTRACT_FOLDERS): + print('Download from {} and extract.'.format(url)) + data_dir = os.path.join(FLAGS.data_dir, folder) + download_and_extract(data_dir, file_name, url) + file_names = _get_file_names(folder) + input_dir = os.path.join(data_dir, extract_folder) + + for mode, files in file_names.items(): + input_files = [os.path.join(input_dir, f) for f in files] + output_file = os.path.join(data_dir, mode + '.tfrecords') + try: + os.remove(output_file) + except OSError: + pass + convert_to_tfrecord(input_files, output_file, folder) + print('Done!') @@ -118,6 +149,6 @@ if __name__ == '__main__': flags.DEFINE_string( 'data_dir', default=None, - help='Directory to download and extract CIFAR-10 to.') + help='Directory to download, extract and store TFRecords.') tf.app.run(main) diff --git a/tensorflow/contrib/eager/python/examples/revnet/config.py b/tensorflow/contrib/eager/python/examples/revnet/config.py index 263a65dc768f421ef39091af6a95033c3d83ac2b..3d93fa955a29718fdec52b04500c41f77351dd8d 100644 --- a/tensorflow/contrib/eager/python/examples/revnet/config.py +++ b/tensorflow/contrib/eager/python/examples/revnet/config.py @@ -61,18 +61,39 @@ def get_hparams_cifar_38(): config.add_hparam("max_train_iter", 80000) config.add_hparam("seed", 1234) config.add_hparam("shuffle", True) - config.add_hparam("prefetch", True) - config.add_hparam("log_every", 50) - config.add_hparam("save_every", 50) + config.add_hparam("log_every", 500) + config.add_hparam("save_every", 500) config.add_hparam("dtype", tf.float32) - config.add_hparam("eval_batch_size", 500) + config.add_hparam("eval_batch_size", 1000) config.add_hparam("div255", True) + # This is imprecise, when training with validation set, + # we only have 40k images in training data config.add_hparam("iters_per_epoch", 50000 // config.batch_size) config.add_hparam("epochs", config.max_train_iter // config.iters_per_epoch) return config +def get_hparams_cifar_110(): + config = get_hparams_cifar_38() + config.filters = [32, 64, 128] + config.n_res = [9, 9, 9] + + return config + + +def get_hparams_cifar_164(): + config = get_hparams_cifar_38() + config.filters = [32, 64, 128] + config.n_res = [9, 9, 9] + config.use_bottleneck = True + # Due to bottleneck residual blocks + filters = [f * 4 for f in config.filters] + config.filters = filters + + return config + + def get_hparams_imagenet_56(): """RevNet-56 configurations for ImageNet.""" @@ -104,18 +125,16 @@ def get_hparams_imagenet_56(): config.add_hparam("max_train_iter", 600000) config.add_hparam("seed", 1234) config.add_hparam("shuffle", True) - config.add_hparam("prefetch", True) config.add_hparam("log_every", 50) config.add_hparam("save_every", 50) config.add_hparam("dtype", tf.float32) - config.add_hparam("eval_batch_size", 500) + config.add_hparam("eval_batch_size", 1000) config.add_hparam("div255", True) # TODO(lxuechen): Update this according to ImageNet data config.add_hparam("iters_per_epoch", 50000 // config.batch_size) config.add_hparam("epochs", config.max_train_iter // config.iters_per_epoch) - - if config.bottleneck: - filters = [f * 4 for f in config.filters] - config.filters = filters + # Due to bottleneck residual blocks + filters = [f * 4 for f in config.filters] + config.filters = filters return config diff --git a/tensorflow/contrib/eager/python/examples/revnet/main.py b/tensorflow/contrib/eager/python/examples/revnet/main.py index 9ef11f8e9b470f3bae6b7cfec194774160fc2bd1..e2f43b03f90ef6db01db1f85943e10ce8c9b582a 100644 --- a/tensorflow/contrib/eager/python/examples/revnet/main.py +++ b/tensorflow/contrib/eager/python/examples/revnet/main.py @@ -19,6 +19,7 @@ from __future__ import division from __future__ import print_function import os +import sys from absl import flags import tensorflow as tf @@ -30,118 +31,226 @@ tfe = tf.contrib.eager def main(_): """Eager execution workflow with RevNet trained on CIFAR-10.""" + config = get_config() + ds_train, ds_train_one_shot, ds_validation, ds_test = get_datasets(config) + model = revnet.RevNet(config=config) + global_step = tf.train.get_or_create_global_step() # Ensure correct summary + global_step.assign(1) + learning_rate = tf.train.piecewise_constant( + global_step, config.lr_decay_steps, config.lr_list) + optimizer = tf.train.MomentumOptimizer( + learning_rate, momentum=config.momentum) + checkpointer = tf.train.Checkpoint( + optimizer=optimizer, model=model, optimizer_step=global_step) + + if FLAGS.train_dir: + summary_writer = tf.contrib.summary.create_file_writer(FLAGS.train_dir) + if FLAGS.restore: + latest_path = tf.train.latest_checkpoint(FLAGS.train_dir) + checkpointer.restore(latest_path) + print("Restored latest checkpoint at path:\"{}\" " + "with global_step: {}".format(latest_path, global_step.numpy())) + sys.stdout.flush() + + if FLAGS.manual_grad: + print("Using manual gradients.") + else: + print("Not using manual gradients.") + sys.stdout.flush() + + for x, y in ds_train: + train_one_iter(model, x, y, optimizer, global_step=global_step) + + if global_step.numpy() % config.log_every == 0: + it_train = ds_train_one_shot.make_one_shot_iterator() + it_test = ds_test.make_one_shot_iterator() + acc_train, loss_train = evaluate(model, it_train) + acc_test, loss_test = evaluate(model, it_test) + + if FLAGS.validate: + it_validation = ds_validation.make_one_shot_iterator() + acc_validation, loss_validation = evaluate(model, it_validation) + print("Iter {}, " + "training set accuracy {:.4f}, loss {:.4f}; " + "validation set accuracy {:.4f}, loss {:4.f}" + "test accuracy {:.4f}, loss {:.4f}".format( + global_step.numpy(), acc_train, loss_train, acc_validation, + loss_validation, acc_test, loss_test)) + else: + print("Iter {}, " + "training set accuracy {:.4f}, loss {:.4f}; " + "test accuracy {:.4f}, loss {:.4f}".format( + global_step.numpy(), acc_train, loss_train, acc_test, + loss_test)) + sys.stdout.flush() + + if FLAGS.train_dir: + with summary_writer.as_default(): + with tf.contrib.summary.always_record_summaries(): + tf.contrib.summary.scalar("Training accuracy", acc_train) + tf.contrib.summary.scalar("Test accuracy", acc_test) + tf.contrib.summary.scalar("Training loss", loss_train) + tf.contrib.summary.scalar("Test loss", loss_test) + if FLAGS.validate: + tf.contrib.summary.scalar("Validation accuracy", acc_validation) + tf.contrib.summary.scalar("Validation loss", loss_validation) + + if global_step.numpy() % config.save_every == 0 and FLAGS.train_dir: + saved_path = checkpointer.save( + file_prefix=os.path.join(FLAGS.train_dir, "ckpt")) + print("Saved checkpoint at path: \"{}\" " + "with global_step: {}".format(saved_path, global_step.numpy())) + sys.stdout.flush() + + +def get_config(): + """Return configuration.""" + print("Config: {}".format(FLAGS.config)) + sys.stdout.flush() + config = { + "revnet-38": config_.get_hparams_cifar_38(), + "revnet-110": config_.get_hparams_cifar_110(), + "revnet-164": config_.get_hparams_cifar_164(), + }[FLAGS.config] + + if FLAGS.dataset == "cifar-100": + config.n_classes = 100 + + return config + + +def get_datasets(config): + """Return dataset.""" if FLAGS.data_dir is None: raise ValueError("No supplied data directory") - if not os.path.exists(FLAGS.data_dir): raise ValueError("Data directory {} does not exist".format(FLAGS.data_dir)) + if FLAGS.dataset not in ["cifar-10", "cifar-100"]: + raise ValueError("Unknown dataset {}".format(FLAGS.dataset)) - tf.enable_eager_execution() - config = config_.get_hparams_cifar_38() - model = revnet.RevNet(config=config) - - ds_train = cifar_input.get_ds_from_tfrecords( - data_dir=FLAGS.data_dir, - split="train", - data_aug=True, - batch_size=config.batch_size, - epochs=config.epochs, - shuffle=config.shuffle, - data_format=config.data_format, - dtype=config.dtype, - prefetch=config.prefetch) + print("Training on {} dataset.".format(FLAGS.dataset)) + sys.stdout.flush() + data_dir = os.path.join(FLAGS.data_dir, FLAGS.dataset) + if FLAGS.validate: + # 40k Training set + ds_train = cifar_input.get_ds_from_tfrecords( + data_dir=data_dir, + split="train", + data_aug=True, + batch_size=config.batch_size, + epochs=config.epochs, + shuffle=config.shuffle, + data_format=config.data_format, + dtype=config.dtype, + prefetch=config.batch_size) + # 10k Training set + ds_validation = cifar_input.get_ds_from_tfrecords( + data_dir=data_dir, + split="validation", + data_aug=False, + batch_size=config.eval_batch_size, + epochs=1, + shuffle=False, + data_format=config.data_format, + dtype=config.dtype, + prefetch=config.eval_batch_size) + else: + # 50k Training set + ds_train = cifar_input.get_ds_from_tfrecords( + data_dir=data_dir, + split="train_all", + data_aug=True, + batch_size=config.batch_size, + epochs=config.epochs, + shuffle=config.shuffle, + data_format=config.data_format, + dtype=config.dtype, + prefetch=config.batch_size) + ds_validation = None - ds_validation = cifar_input.get_ds_from_tfrecords( - data_dir=FLAGS.data_dir, - split="validation", + # Always compute loss and accuracy on whole training and test set + ds_train_one_shot = cifar_input.get_ds_from_tfrecords( + data_dir=data_dir, + split="train_all", data_aug=False, batch_size=config.eval_batch_size, epochs=1, + shuffle=False, data_format=config.data_format, dtype=config.dtype, - prefetch=config.prefetch) + prefetch=config.eval_batch_size) ds_test = cifar_input.get_ds_from_tfrecords( - data_dir=FLAGS.data_dir, + data_dir=data_dir, split="test", data_aug=False, batch_size=config.eval_batch_size, epochs=1, + shuffle=False, data_format=config.data_format, dtype=config.dtype, - prefetch=config.prefetch) - - global_step = tfe.Variable(1, trainable=False) + prefetch=config.eval_batch_size) - def learning_rate(): # TODO(lxuechen): Remove once cl/201089859 is in place - return tf.train.piecewise_constant(global_step, config.lr_decay_steps, - config.lr_list) - - optimizer = tf.train.MomentumOptimizer(learning_rate, momentum=0.9) - checkpoint = tf.train.Checkpoint( - optimizer=optimizer, model=model, optimizer_step=global_step) - - if FLAGS.train_dir: - summary_writer = tf.contrib.summary.create_file_writer(FLAGS.train_dir) - if FLAGS.restore: - latest_path = tf.train.latest_checkpoint(FLAGS.train_dir) - checkpoint.restore(latest_path) - - for x, y in ds_train: - loss = train_one_iter(model, x, y, optimizer, global_step=global_step) - - if global_step % config.log_every == 0: - it_validation = ds_validation.make_one_shot_iterator() - it_test = ds_test.make_one_shot_iterator() - acc_validation = evaluate(model, it_validation) - acc_test = evaluate(model, it_test) - print("Iter {}, " - "train loss {}, " - "validation accuracy {}, " - "test accuracy {}".format(global_step.numpy(), loss, acc_validation, - acc_test)) - - if FLAGS.train_dir: - with summary_writer.as_default(): - with tf.contrib.summary.always_record_summaries(): - tf.contrib.summary.scalar("Validation accuracy", acc_validation) - tf.contrib.summary.scalar("Test accuracy", acc_test) - tf.contrib.summary.scalar("Training loss", loss) - - if global_step.numpy() % config.save_every == 0 and FLAGS.train_dir: - checkpoint.save(file_prefix=FLAGS.train_dir + "ckpt") + return ds_train, ds_train_one_shot, ds_validation, ds_test def train_one_iter(model, inputs, labels, optimizer, global_step=None): """Train for one iteration.""" - grads, vars_, loss = model.compute_gradients(inputs, labels, training=True) - optimizer.apply_gradients(zip(grads, vars_), global_step=global_step) + if FLAGS.manual_grad: + grads, vars_, loss = model.compute_gradients(inputs, labels, training=True) + optimizer.apply_gradients(zip(grads, vars_), global_step=global_step) + else: # For correctness validation + with tf.GradientTape() as tape: + logits, _ = model(inputs, training=True) + loss = model.compute_loss(logits=logits, labels=labels) + tf.logging.info("Logits are placed on device: {}".format(logits.device)) + grads = tape.gradient(loss, model.trainable_variables) + optimizer.apply_gradients( + zip(grads, model.trainable_variables), global_step=global_step) return loss.numpy() def evaluate(model, iterator): """Compute accuracy with the given dataset iterator.""" + mean_loss = tfe.metrics.Mean() accuracy = tfe.metrics.Accuracy() for x, y in iterator: logits, _ = model(x, training=False) + loss = model.compute_loss(logits=logits, labels=y) accuracy( labels=tf.cast(y, tf.int64), predictions=tf.argmax(logits, axis=1, output_type=tf.int64)) + mean_loss(loss) - return accuracy.result().numpy() + return accuracy.result().numpy(), mean_loss.result().numpy() if __name__ == "__main__": + flags.DEFINE_string( + "data_dir", default=None, help="Directory to load tfrecords") flags.DEFINE_string( "train_dir", default=None, help="[Optional] Directory to store the training information") - flags.DEFINE_string( - "data_dir", default=None, help="Directory to load tfrecords.") flags.DEFINE_boolean( "restore", - default=True, + default=False, help="[Optional] Restore the latest checkpoint from `train_dir` if True") + flags.DEFINE_boolean( + "validate", + default=False, + help="[Optional] Use the validation set or not for hyperparameter search") + flags.DEFINE_boolean( + "manual_grad", + default=False, + help="[Optional] Use manual gradient graph to save memory") + flags.DEFINE_string( + "dataset", + default="cifar-10", + help="[Optional] The dataset used; either `cifar-10` or `cifar-100`") + flags.DEFINE_string( + "config", default="revnet-38", help="[Optional] Architecture of network.") FLAGS = flags.FLAGS + tf.enable_eager_execution() tf.app.run(main) diff --git a/tensorflow/contrib/eager/python/examples/revnet/revnet.py b/tensorflow/contrib/eager/python/examples/revnet/revnet.py index b3b8c262b1517baa1e65c105db9882b6f7672439..af0d20fa729836b12036d5d54a9b5b0b68d719d2 100644 --- a/tensorflow/contrib/eager/python/examples/revnet/revnet.py +++ b/tensorflow/contrib/eager/python/examples/revnet/revnet.py @@ -27,6 +27,7 @@ from __future__ import print_function import functools import operator +import six import tensorflow as tf from tensorflow.contrib.eager.python.examples.revnet import blocks @@ -58,9 +59,12 @@ class RevNet(tf.keras.Model): data_format=self.config.data_format, use_bias=False, padding="SAME", - input_shape=self.config.input_shape), + input_shape=self.config.input_shape, + dtype=self.config.dtype), tf.keras.layers.BatchNormalization( - axis=self.axis, fused=self.config.fused), + axis=self.axis, + fused=self.config.fused, + dtype=self.config.dtype), tf.keras.layers.Activation("relu"), ], name="init") @@ -70,7 +74,8 @@ class RevNet(tf.keras.Model): pool_size=(3, 3), strides=(2, 2), padding="SAME", - data_format=self.config.data_format)) + data_format=self.config.data_format, + dtype=self.config.dtype)) return init_block def _construct_final_block(self): @@ -95,11 +100,13 @@ class RevNet(tf.keras.Model): tf.keras.layers.BatchNormalization( axis=self.axis, input_shape=input_shape, - fused=self.config.fused), + fused=self.config.fused, + dtype=self.config.dtype), tf.keras.layers.Activation("relu"), tf.keras.layers.GlobalAveragePooling2D( - data_format=self.config.data_format), - tf.keras.layers.Dense(self.config.n_classes) + data_format=self.config.data_format, dtype=self.config.dtype), + tf.keras.layers.Dense( + self.config.n_classes, dtype=self.config.dtype) ], name="final") return final_block @@ -137,7 +144,8 @@ class RevNet(tf.keras.Model): batch_norm_first=(i != 0), # Only skip on first block data_format=self.config.data_format, bottleneck=self.config.bottleneck, - fused=self.config.fused) + fused=self.config.fused, + dtype=self.config.dtype) block_list.append(rev_block) # Precompute input shape for the next block @@ -153,7 +161,6 @@ class RevNet(tf.keras.Model): def call(self, inputs, training=True): """Forward pass.""" - # Only store hidden states during training if training: saved_hidden = [inputs] @@ -173,25 +180,39 @@ class RevNet(tf.keras.Model): def compute_loss(self, logits, labels): """Compute cross entropy loss.""" - cross_ent = tf.nn.sparse_softmax_cross_entropy_with_logits( - logits=logits, labels=labels) + if self.config.dtype == tf.float32 or self.config.dtype == tf.float16: + cross_ent = tf.nn.sparse_softmax_cross_entropy_with_logits( + logits=logits, labels=labels) + else: + # `sparse_softmax_cross_entropy_with_logits` does not have a GPU kernel + # for float64, int32 pairs + labels = tf.one_hot( + labels, depth=self.config.n_classes, axis=1, dtype=self.config.dtype) + cross_ent = tf.nn.softmax_cross_entropy_with_logits( + logits=logits, labels=labels) return tf.reduce_mean(cross_ent) - def compute_gradients(self, inputs, labels, training=True): + def compute_gradients(self, inputs, labels, training=True, l2_reg=True): """Manually computes gradients. + When eager execution is enabled, this method also SILENTLY updates the + running averages of batch normalization when `training` is set to True. + Args: inputs: Image tensor, either NHWC or NCHW, conforming to `data_format` labels: One-hot labels for classification - training: for batch normalization + training: Use the mini-batch stats in batch norm if set to True + l2_reg: Apply l2 regularization Returns: - list of tuple each being (grad, var) for optimizer use + list of tuples each being (grad, var) for optimizer to use """ - # Forward pass record hidden states before downsampling + # Run forward pass to record hidden states; avoid updating running averages + vars_and_vals = self.get_moving_stats() _, saved_hidden = self.call(inputs, training=training) + self.restore_moving_stats(vars_and_vals) grads_all = [] vars_all = [] @@ -199,8 +220,9 @@ class RevNet(tf.keras.Model): # Manually backprop through last block x = saved_hidden[-1] with tf.GradientTape() as tape: - x = tf.identity(x) # TODO(lxuechen): Remove after b/110264016 is fixed + x = tf.identity(x) tape.watch(x) + # Running stats updated below logits = self._final_block(x, training=training) loss = self.compute_loss(logits, labels) @@ -225,17 +247,55 @@ class RevNet(tf.keras.Model): assert not saved_hidden # Cleared after backprop with tf.GradientTape() as tape: - x = tf.identity(x) # TODO(lxuechen): Remove after b/110264016 is fixed + x = tf.identity(x) + # Running stats updated below y = self._init_block(x, training=training) grads_all += tape.gradient( - y, self._init_block.trainable_variables, output_gradients=[dy]) + y, self._init_block.trainable_variables, output_gradients=dy) vars_all += self._init_block.trainable_variables - grads_all = self._apply_weight_decay(grads_all, vars_all) + # Apply weight decay + if l2_reg: + grads_all = self._apply_weight_decay(grads_all, vars_all) return grads_all, vars_all, loss def _apply_weight_decay(self, grads, vars_): """Update gradients to reflect weight decay.""" - return [g + self.config.weight_decay * v for g, v in zip(grads, vars_)] + # Don't decay bias + return [ + g + self.config.weight_decay * v if v.name.endswith("kernel:0") else g + for g, v in zip(grads, vars_) + ] + + def get_moving_stats(self): + """Get moving averages of batch normalization. + + This is needed to avoid updating the running average twice in one iteration. + + Returns: + A dictionary mapping variables for batch normalization moving averages + to their current values. + """ + vars_and_vals = {} + + def _is_moving_var(v): + n = v.name + return n.endswith("moving_mean:0") or n.endswith("moving_variance:0") + + for v in filter(_is_moving_var, self.variables): + vars_and_vals[v] = v.read_value() + + return vars_and_vals + + def restore_moving_stats(self, vars_and_vals): + """Restore moving averages of batch normalization. + + This is needed to avoid updating the running average twice in one iteration. + + Args: + vars_and_vals: The dictionary mapping variables to their previous values. + """ + for var_, val in six.iteritems(vars_and_vals): + var_.assign(val) diff --git a/tensorflow/contrib/eager/python/examples/revnet/revnet_test.py b/tensorflow/contrib/eager/python/examples/revnet/revnet_test.py index cb3bac13f96094a5b0af88bb99c8cc8169d12f95..b2ac4b67c926951672996df5564b9b57def0ea13 100644 --- a/tensorflow/contrib/eager/python/examples/revnet/revnet_test.py +++ b/tensorflow/contrib/eager/python/examples/revnet/revnet_test.py @@ -22,6 +22,7 @@ import gc import time import tensorflow as tf +from tensorflow.contrib.eager.python.examples.revnet import blocks_test from tensorflow.contrib.eager.python.examples.revnet import config as config_ from tensorflow.contrib.eager.python.examples.revnet import revnet from tensorflow.python.client import device_lib @@ -36,19 +37,22 @@ def train_one_iter(model, inputs, labels, optimizer, global_step=None): return loss -class RevnetTest(tf.test.TestCase): +class RevNetTest(tf.test.TestCase): def setUp(self): - super(RevnetTest, self).setUp() - config = config_.get_hparams_imagenet_56() + super(RevNetTest, self).setUp() + config = config_.get_hparams_cifar_38() + # Reconstruction could cause numerical error, use double precision for tests + config.dtype = tf.float64 + config.fused = False # Fused batch norm does not support tf.float64 shape = (config.batch_size,) + config.input_shape self.model = revnet.RevNet(config=config) - self.x = tf.random_normal(shape=shape) + self.x = tf.random_normal(shape=shape, dtype=tf.float64) self.t = tf.random_uniform( shape=[config.batch_size], minval=0, maxval=config.n_classes, - dtype=tf.int32) + dtype=tf.int64) self.config = config def tearDown(self): @@ -56,7 +60,7 @@ class RevnetTest(tf.test.TestCase): del self.x del self.t del self.config - super(RevnetTest, self).tearDown() + super(RevNetTest, self).tearDown() def test_call(self): """Test `call` function.""" @@ -64,27 +68,58 @@ class RevnetTest(tf.test.TestCase): y, _ = self.model(self.x, training=False) self.assertEqual(y.shape, [self.config.batch_size, self.config.n_classes]) + def _check_grad_angle_combined(self, grads, grads_true): + """Verify that the reconstructed gradients has correct direction. + + Due to numerical imprecision, the magnitude may be slightly different. + Yet according to the paper, the angle should be roughly the same. + + Args: + grads: list of gradients from reconstruction + grads_true: list of true gradients + """ + + def _combine(gs): + return [tf.reshape(g, [-1]) for g in gs] + + g1_all = tf.concat(_combine(grads), axis=0) + g2_all = tf.concat(_combine(grads_true), axis=0) + + self.assertEqual(len(g1_all.shape), 1) + self.assertEqual(len(g2_all.shape), 1) + + degree = blocks_test.compute_degree(g1_all, g2_all) + self.assertLessEqual(degree, 1e0) + def test_compute_gradients(self): """Test `compute_gradients` function.""" - - grads, vars_, _ = self.model.compute_gradients(inputs=self.x, labels=self.t) + self.model(self.x, training=False) # Initialize model + grads, vars_, loss = self.model.compute_gradients( + inputs=self.x, labels=self.t, training=True, l2_reg=True) self.assertTrue(isinstance(grads, list)) self.assertTrue(isinstance(vars_, list)) self.assertEqual(len(grads), len(vars_)) for grad, var in zip(grads, vars_): - if grad is not None: - self.assertEqual(grad.shape, var.shape) + self.assertEqual(grad.shape, var.shape) + + # Compare against the true gradient computed by the tape + with tf.GradientTape() as tape: + logits, _ = self.model(self.x, training=True) + loss_true = self.model.compute_loss(logits=logits, labels=self.t) + grads_true = tape.gradient(loss_true, vars_) + self.assertAllClose(loss, loss_true) + self.assertAllClose(grads, grads_true, rtol=1e-4, atol=1e-4) + self._check_grad_angle_combined(grads, grads_true) def test_call_defun(self): """Test `call` function with defun.""" - y, _ = tfe.defun(self.model.call)(self.x, training=False) self.assertEqual(y.shape, [self.config.batch_size, self.config.n_classes]) def test_compute_gradients_defun(self): """Test `compute_gradients` function with defun.""" compute_gradients = tfe.defun(self.model.compute_gradients) - grads, vars_, _ = compute_gradients(self.x, self.t) + grads, vars_, _ = compute_gradients(self.x, self.t, training=True) self.assertTrue(isinstance(grads, list)) self.assertTrue(isinstance(vars_, list)) self.assertEqual(len(grads), len(vars_)) @@ -94,8 +129,8 @@ class RevnetTest(tf.test.TestCase): def test_training_graph(self): """Test model training in graph mode.""" - with tf.Graph().as_default(): + config = config_.get_hparams_cifar_38() x = tf.random_normal( shape=(self.config.batch_size,) + self.config.input_shape) t = tf.random_uniform( @@ -104,12 +139,14 @@ class RevnetTest(tf.test.TestCase): maxval=self.config.n_classes, dtype=tf.int32) global_step = tfe.Variable(0., trainable=False) - model = revnet.RevNet(config=self.config) - grads_all, vars_all, _ = model.compute_gradients(x, t, training=True) - optimizer = tf.train.AdamOptimizer(learning_rate=1e-3) + model = revnet.RevNet(config=config) + model(x) updates = model.get_updates_for(x) - self.assertEqual(len(updates), 192) - with tf.control_dependencies(model.get_updates_for(x)): + + x_ = tf.identity(x) + grads_all, vars_all, _ = model.compute_gradients(x_, t, training=True) + optimizer = tf.train.AdamOptimizer(learning_rate=1e-3) + with tf.control_dependencies(updates): train_op = optimizer.apply_gradients( zip(grads_all, vars_all), global_step=global_step) @@ -144,7 +181,7 @@ class MockIterator(object): return self._tensors -class RevnetBenchmark(tf.test.Benchmark): +class RevNetBenchmark(tf.test.Benchmark): """Eager and graph benchmarks for RevNet.""" def _train_batch_sizes(self): diff --git a/tensorflow/contrib/eager/python/examples/workshop/1_basic.ipynb b/tensorflow/contrib/eager/python/examples/workshop/1_basic.ipynb new file mode 100644 index 0000000000000000000000000000000000000000..75cb3f8227fe90223734f422e458f15810b8089a --- /dev/null +++ b/tensorflow/contrib/eager/python/examples/workshop/1_basic.ipynb @@ -0,0 +1,282 @@ +{ + "nbformat": 4, + "nbformat_minor": 0, + "metadata": { + "colab": { + "name": "TFE Workshop: control flow", + "version": "0.3.2", + "provenance": [], + "include_colab_link": true + } + }, + "cells": [ + { + "cell_type": "markdown", + "metadata": { + "id": "view-in-github", + "colab_type": "text" + }, + "source": [ + "[View in Colaboratory](https://colab.research.google.com/gist/alextp/664b2f8700485ff6801f4d26293bd567/tfe-workshop-control-flow.ipynb)" + ] + }, + { + "metadata": { + "id": "9BpQzh9BvJlj", + "colab_type": "code", + "colab": { + "base_uri": "https://localhost:8080/", + "height": 37 + }, + "outputId": "0b336886-8204-4815-89fa-5291a49d5784" + }, + "cell_type": "code", + "source": [ + "import tensorflow as tf\n", + "import numpy as np\n", + "tf.enable_eager_execution()" + ], + "execution_count": 1, + "outputs": [] + }, + { + "metadata": { + "id": "0roIB19GvOjI", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "# Eager execution basics\n", + "\n", + "When eager execution is enabled TensorFlow immediately executes operations, and Tensors are always available. " + ] + }, + { + "metadata": { + "id": "jeO8F-V-vN24", + "colab_type": "code", + "colab": { + "base_uri": "https://localhost:8080/", + "height": 68 + }, + "outputId": "aeb3bdec-50b7-440d-93d8-5a171f091081" + }, + "cell_type": "code", + "source": [ + "t = tf.constant([[1, 2], [3, 4]])\n", + "t" + ], + "execution_count": 2, + "outputs": [ + { + "output_type": "execute_result", + "data": { + "text/plain": [ + "" + ] + }, + "metadata": { + "tags": [] + }, + "execution_count": 2 + } + ] + }, + { + "metadata": { + "id": "Y17RwSFxvlDL", + "colab_type": "code", + "colab": { + "base_uri": "https://localhost:8080/", + "height": 68 + }, + "outputId": "cfcc10c7-707b-4997-99b3-a5f382c5166b" + }, + "cell_type": "code", + "source": [ + "tf.matmul(t, t)" + ], + "execution_count": 3, + "outputs": [ + { + "output_type": "execute_result", + "data": { + "text/plain": [ + "" + ] + }, + "metadata": { + "tags": [] + }, + "execution_count": 3 + } + ] + }, + { + "metadata": { + "id": "Dab1bS3TvmRE", + "colab_type": "code", + "colab": { + "base_uri": "https://localhost:8080/", + "height": 34 + }, + "outputId": "8a624f3d-a658-4359-c586-1c5f6bf4c8b7" + }, + "cell_type": "code", + "source": [ + "# It's also possible to have Python control flow which depends on the value of tensors.\n", + "if t[0, 0] > 0.5:\n", + " print(\"T is bigger\")\n", + "else:\n", + " print(\"T is smaller\")" + ], + "execution_count": 4, + "outputs": [ + { + "output_type": "stream", + "text": [ + "T is bigger\n" + ], + "name": "stdout" + } + ] + }, + { + "metadata": { + "id": "dPgptJcGwIon", + "colab_type": "code", + "colab": { + "base_uri": "https://localhost:8080/", + "height": 34 + }, + "outputId": "c4f27f2b-0848-4475-dde5-2534dac65a5c" + }, + "cell_type": "code", + "source": [ + "# Tensors are also usable as numpy arrays\n", + "np.prod(t)" + ], + "execution_count": 6, + "outputs": [ + { + "output_type": "execute_result", + "data": { + "text/plain": [ + "24" + ] + }, + "metadata": { + "tags": [] + }, + "execution_count": 6 + } + ] + }, + { + "metadata": { + "id": "p3DTfQXnwXzj", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "# Exercise\n", + "\n", + "The algorithm for bisecting line search is a pretty simple way to find a zero of a continuous scalar function in an interval [a,b] where f(a) and f(b) have different signs. Simply evaluate f((a+b)/2), and narrow the interval by replacing either a or b with (a+b)/2 such that the function when applied on the boundary of the interval still has different signs.\n", + "\n", + "Implement a python function `bisecting_line_search(f, a, b, epsilon)` which returns a value such that `tf.abs(f(value)) < epsilon`.\n", + "\n", + "One thing to keep in mind: python's `==` opertor is not overloaded on Tensors, so you need to use `tf.equal` to compare for equality." + ] + }, + { + "metadata": { + "id": "6eq0YuI6ykm5", + "colab_type": "code", + "colab": {} + }, + "cell_type": "code", + "source": [ + "# Example test harness to get you going\n", + "\n", + "def test_f(x):\n", + " return x - 0.1234\n", + "def bisecting_line_search(f, a, b, epsilon):\n", + " # Return x such that f(x) <= epsilon.\n", + " pass\n", + "a = tf.constant(0.0)\n", + "b = tf.constant(1.0)\n", + "epsilon = tf.constant(0.001)\n", + "x = bisecting_line_search(test_f, a, b, epsilon)\n" + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "LcMmEfd_xvej", + "colab_type": "code", + "colab": { + "base_uri": "https://localhost:8080/", + "height": 170 + }, + "outputId": "f402aa50-8ce3-4416-f755-8bbcd1af7809" + }, + "cell_type": "code", + "source": [ + "#@title Double-click to see the solution\n", + "\n", + "def bisecting_line_search(f, a, b, epsilon):\n", + " f_a = f(a)\n", + " f_b = f(b)\n", + " probe = (a + b) / 2\n", + " f_probe = f(probe)\n", + " while tf.abs(f_probe) > epsilon:\n", + " if tf.equal(tf.sign(f_probe), tf.sign(f_a)):\n", + " a = probe\n", + " f_a = f_probe\n", + " else:\n", + " b = probe\n", + " f_b = f_probe\n", + " probe = (a + b) / 2\n", + " f_probe = f(probe)\n", + " print(\"new probe\", probe)\n", + " return probe\n", + "\n", + "bisecting_line_search(test_f, 0., 1., 0.001)" + ], + "execution_count": 8, + "outputs": [ + { + "output_type": "stream", + "text": [ + "('new probe', 0.25)\n", + "('new probe', 0.125)\n", + "('new probe', 0.0625)\n", + "('new probe', 0.09375)\n", + "('new probe', 0.109375)\n", + "('new probe', 0.1171875)\n", + "('new probe', 0.12109375)\n", + "('new probe', 0.123046875)\n" + ], + "name": "stdout" + }, + { + "output_type": "execute_result", + "data": { + "text/plain": [ + "0.123046875" + ] + }, + "metadata": { + "tags": [] + }, + "execution_count": 8 + } + ] + } + ] +} diff --git a/tensorflow/contrib/eager/python/examples/workshop/2_models.ipynb b/tensorflow/contrib/eager/python/examples/workshop/2_models.ipynb new file mode 100644 index 0000000000000000000000000000000000000000..4f1410e00bb986f68f3c4c8494aa97bf66284510 --- /dev/null +++ b/tensorflow/contrib/eager/python/examples/workshop/2_models.ipynb @@ -0,0 +1,1018 @@ +{ + "nbformat": 4, + "nbformat_minor": 0, + "metadata": { + "colab": { + "name": "TFE Workshop: Models.ipynb", + "version": "0.3.2", + "provenance": [], + "collapsed_sections": [], + "include_colab_link": true + }, + "kernelspec": { + "name": "python3", + "display_name": "Python 3" + } + }, + "cells": [ + { + "cell_type": "markdown", + "metadata": { + "id": "view-in-github", + "colab_type": "text" + }, + "source": [ + "[View in Colaboratory](https://colab.research.google.com/gist/alextp/5cfcffd408bd5103f5ae747bc97ab0b5/tfe-workshop-models.ipynb)" + ] + }, + { + "metadata": { + "id": "BMxv1O6Q0SJL", + "colab_type": "code", + "colab": { + "base_uri": "https://localhost:8080/", + "height": 17 + }, + "outputId": "8be9c556-ac7f-4142-e35e-19dc2b097121" + }, + "cell_type": "code", + "source": [ + "import tensorflow as tf\n", + "tf.enable_eager_execution()\n", + "tfe = tf.contrib.eager" + ], + "execution_count": 1, + "outputs": [] + }, + { + "metadata": { + "id": "lE1vJhxp0WR9", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "# Variables\n", + "\n", + "TensorFlow variables are useful to store the state in your program. They are integrated with other parts of the API (taking gradients, checkpointing, graph functions)." + ] + }, + { + "metadata": { + "id": "C4ztQNgc0VpW", + "colab_type": "code", + "colab": { + "base_uri": "https://localhost:8080/", + "height": 34 + }, + "outputId": "8b63ae1f-2670-49c0-a31b-8cf7fc4194a1" + }, + "cell_type": "code", + "source": [ + "# Creating variables\n", + "v = tfe.Variable(1.0)\n", + "v" + ], + "execution_count": 2, + "outputs": [ + { + "output_type": "execute_result", + "data": { + "text/plain": [ + "" + ] + }, + "metadata": { + "tags": [] + }, + "execution_count": 2 + } + ] + }, + { + "metadata": { + "id": "H0daItGg1IAp", + "colab_type": "code", + "colab": { + "base_uri": "https://localhost:8080/", + "height": 34 + }, + "outputId": "e47d5aab-16a1-4e29-c27d-7fbc0b94b5d3" + }, + "cell_type": "code", + "source": [ + "v.assign_add(1.0)\n", + "v" + ], + "execution_count": 3, + "outputs": [ + { + "output_type": "execute_result", + "data": { + "text/plain": [ + "" + ] + }, + "metadata": { + "tags": [] + }, + "execution_count": 3 + } + ] + }, + { + "metadata": { + "id": "BJvBzcIG1hyK", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "# Layers: common sets of useful operations\n", + "\n", + "Most of the time when writing code for machine learning models you want to operate at a higher level of abstraction than individual operations and manipulation of individual variables.\n", + "\n", + "Many machine learning models are expressible as the composition and stacking of relatively simple layers, and TensorFlow provides both a set of many common layers as a well as easy ways for you to write your own application-specific layers either from scratch or as the composition of existing layers.\n", + "\n", + "TensorFlow includes the full [Keras](https://keras.io) API in the tf.keras package, and the Keras layers are very useful when building your own models.\n" + ] + }, + { + "metadata": { + "id": "iSQTS3QW1YQQ", + "colab_type": "code", + "colab": { + "base_uri": "https://localhost:8080/", + "height": 17 + }, + "outputId": "c5d8aa10-dcad-44f7-f0eb-0faf5249fd7e" + }, + "cell_type": "code", + "source": [ + "# In the tf.keras.layers package, layers are objects. To construct a layer,\n", + "# simply construct the object. Most layers take as a first argument the number\n", + "# of output dimensions / channels.\n", + "layer = tf.keras.layers.Dense(100)\n", + "\n", + "# The number of input dimensions is often unnecessary, as it can be inferred\n", + "# the first time the layer is used, but it can be provided if you want to \n", + "# specify it manually, which is useful in some complex models.\n", + "layer = tf.keras.layers.Dense(10, input_shape=(None, 5))\n" + ], + "execution_count": 4, + "outputs": [] + }, + { + "metadata": { + "id": "nRuUogoS1liV", + "colab_type": "code", + "colab": { + "base_uri": "https://localhost:8080/", + "height": 68 + }, + "outputId": "c352ce79-d519-45e4-a12e-1eaba76871a2" + }, + "cell_type": "code", + "source": [ + "layer(tf.zeros([2, 2]))" + ], + "execution_count": 5, + "outputs": [ + { + "output_type": "execute_result", + "data": { + "text/plain": [ + "" + ] + }, + "metadata": { + "tags": [] + }, + "execution_count": 5 + } + ] + }, + { + "metadata": { + "id": "JH4Kf4ka1mht", + "colab_type": "code", + "colab": { + "base_uri": "https://localhost:8080/", + "height": 136 + }, + "outputId": "c34e2378-f83d-42c5-d30a-ebe55620368a" + }, + "cell_type": "code", + "source": [ + "layer.variables" + ], + "execution_count": 6, + "outputs": [ + { + "output_type": "execute_result", + "data": { + "text/plain": [ + "[,\n", + " ]" + ] + }, + "metadata": { + "tags": [] + }, + "execution_count": 6 + } + ] + }, + { + "metadata": { + "id": "DSI4NF0_1vn-", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "The full list of pre-existing layers can be seen in [the documentation](https://www.tensorflow.org/api_docs/python/tf/keras/layers). It includes Dense (a fully-connected layer),\n", + "Conv2D, LSTM, BatchNormalization, Dropout, and many others." + ] + }, + { + "metadata": { + "id": "hMgDBftJ12Bp", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "# Models: composing layers\n", + "\n", + "Many interesting layer-like things in machine learning models are implemented by composing existing layers. For example, each residual block in a resnet is a composition of convolutions, batch normalizations, and a shortcut.\n", + "\n", + "The main class used when creating a layer-like thing which contains other layers is tf.keras.Model. Implementing one is done by inheriting from tf.keras.Model.\n" + ] + }, + { + "metadata": { + "id": "K3gVY6gj1nbe", + "colab_type": "code", + "colab": { + "base_uri": "https://localhost:8080/", + "height": 190 + }, + "outputId": "6e9be0c4-960e-46c2-cdd9-7e94ad09d46b" + }, + "cell_type": "code", + "source": [ + "class ResnetIdentityBlock(tf.keras.Model):\n", + " def __init__(self, kernel_size, filters):\n", + " super(ResnetIdentityBlock, self).__init__(name='')\n", + " filters1, filters2, filters3 = filters\n", + "\n", + " self.conv2a = tf.keras.layers.Conv2D(filters1, (1, 1))\n", + " self.bn2a = tf.keras.layers.BatchNormalization()\n", + "\n", + " self.conv2b = tf.keras.layers.Conv2D(filters2, kernel_size, padding='same')\n", + " self.bn2b = tf.keras.layers.BatchNormalization()\n", + "\n", + " self.conv2c = tf.keras.layers.Conv2D(filters3, (1, 1))\n", + " self.bn2c = tf.keras.layers.BatchNormalization()\n", + "\n", + " def call(self, input_tensor, training=False):\n", + " x = self.conv2a(input_tensor)\n", + " x = self.bn2a(x, training=training)\n", + " x = tf.nn.relu(x)\n", + "\n", + " x = self.conv2b(x)\n", + " x = self.bn2b(x, training=training)\n", + " x = tf.nn.relu(x)\n", + "\n", + " x = self.conv2c(x)\n", + " x = self.bn2c(x, training=training)\n", + "\n", + " x += input_tensor\n", + " return tf.nn.relu(x)\n", + " \n", + "block = ResnetIdentityBlock(1, [1, 2, 3])\n", + "print(block(tf.zeros([1, 2, 3, 3])))\n", + "print([x.name for x in block.variables])" + ], + "execution_count": 7, + "outputs": [ + { + "output_type": "stream", + "text": [ + "tf.Tensor(\n", + "[[[[0. 0. 0.]\n", + " [0. 0. 0.]\n", + " [0. 0. 0.]]\n", + "\n", + " [[0. 0. 0.]\n", + " [0. 0. 0.]\n", + " [0. 0. 0.]]]], shape=(1, 2, 3, 3), dtype=float32)\n", + "['resnet_identity_block/conv2d/kernel:0', 'resnet_identity_block/conv2d/bias:0', 'resnet_identity_block/batch_normalization/gamma:0', 'resnet_identity_block/batch_normalization/beta:0', 'resnet_identity_block/conv2d_1/kernel:0', 'resnet_identity_block/conv2d_1/bias:0', 'resnet_identity_block/batch_normalization_1/gamma:0', 'resnet_identity_block/batch_normalization_1/beta:0', 'resnet_identity_block/conv2d_2/kernel:0', 'resnet_identity_block/conv2d_2/bias:0', 'resnet_identity_block/batch_normalization_2/gamma:0', 'resnet_identity_block/batch_normalization_2/beta:0', 'resnet_identity_block/batch_normalization/moving_mean:0', 'resnet_identity_block/batch_normalization/moving_variance:0', 'resnet_identity_block/batch_normalization_1/moving_mean:0', 'resnet_identity_block/batch_normalization_1/moving_variance:0', 'resnet_identity_block/batch_normalization_2/moving_mean:0', 'resnet_identity_block/batch_normalization_2/moving_variance:0']\n" + ], + "name": "stdout" + } + ] + }, + { + "metadata": { + "id": "LPXhHUIc1-sO", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "Much of the time, however, models which compose many layers simply call one layer after the other. This can be done in very little code using tf.keras.Sequential" + ] + }, + { + "metadata": { + "id": "5pXgzNAU17xk", + "colab_type": "code", + "colab": { + "base_uri": "https://localhost:8080/", + "height": 173 + }, + "outputId": "03b7eaf8-9b35-482b-bcf0-a99af6c2c6a4" + }, + "cell_type": "code", + "source": [ + " my_seq = tf.keras.Sequential([tf.keras.layers.Conv2D(1, (1, 1)),\n", + " tf.keras.layers.BatchNormalization(),\n", + " tf.keras.layers.Conv2D(2, 1, \n", + " padding='same'),\n", + " tf.keras.layers.BatchNormalization(),\n", + " tf.keras.layers.Conv2D(3, (1, 1)),\n", + " tf.keras.layers.BatchNormalization()])\n", + "my_seq(tf.zeros([1, 2, 3, 3]))\n" + ], + "execution_count": 8, + "outputs": [ + { + "output_type": "execute_result", + "data": { + "text/plain": [ + "" + ] + }, + "metadata": { + "tags": [] + }, + "execution_count": 8 + } + ] + }, + { + "metadata": { + "id": "MZrns6p22GEQ", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "## Exercise!\n", + "\n", + "Make a simple convolutional neural network model, useful for things such as MNIST which don't need too many parameters. A sequence of two or three convolutions with small output channels (say, 32 and 64) plus one or two fully connected layers is probably enough.\n", + "\n", + "The input shape should be [batch_size, 28, 28, 1]." + ] + }, + { + "metadata": { + "id": "8CAUa3KNN916", + "colab_type": "code", + "colab": { + "base_uri": "https://localhost:8080/", + "height": 17 + }, + "outputId": "97c0ff3c-c962-4c13-eee8-406101465761" + }, + "cell_type": "code", + "source": [ + "# TODO: Implement a convolutional model as described above, and assign it to\n", + "# model.\n", + "model = tf.keras.Sequential([\n", + " \n", + "])" + ], + "execution_count": 9, + "outputs": [] + }, + { + "metadata": { + "id": "vLDDduR32E82", + "colab_type": "code", + "colab": { + "base_uri": "https://localhost:8080/", + "height": 34 + }, + "outputId": "09bb1d43-b4c6-44b5-916e-0d2903d10cf4" + }, + "cell_type": "code", + "source": [ + "#@title Click to see the answer\n", + "\n", + "max_pool = tf.keras.layers.MaxPooling2D(\n", + " (2, 2), (2, 2), padding='same')\n", + " # The model consists of a sequential chain of layers, so tf.keras.Sequential\n", + " # (a subclass of tf.keras.Model) makes for a compact description.\n", + "model = tf.keras.Sequential(\n", + " [\n", + " tf.keras.layers.Conv2D(\n", + " 32,\n", + " 5,\n", + " padding='same',\n", + " activation=tf.nn.relu),\n", + " max_pool,\n", + " tf.keras.layers.Conv2D(\n", + " 64,\n", + " 5,\n", + " padding='same',\n", + " activation=tf.nn.relu),\n", + " max_pool,\n", + " tf.keras.layers.Flatten(),\n", + " tf.keras.layers.Dense(1024, activation=tf.nn.relu),\n", + " tf.keras.layers.Dropout(0.4),\n", + " tf.keras.layers.Dense(10)\n", + " ])\n", + "\n", + "model(tf.zeros([1, 28, 28, 1]))" + ], + "execution_count": 10, + "outputs": [ + { + "output_type": "execute_result", + "data": { + "text/plain": [ + "" + ] + }, + "metadata": { + "tags": [] + }, + "execution_count": 10 + } + ] + }, + { + "metadata": { + "id": "H_CKVBroik4M", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "# Stop here for now" + ] + }, + { + "metadata": { + "id": "_yRwuE6MMmzC", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "# Training\n", + "\n", + "When eager execution is enabled, you can write Pythonic training loops. Simply\n", + "\n", + "1. load your data into a `tf.data.Dataset`, which lets you construct functional pipelines for processing, shuffling, and batching your data,\n", + "2. iterate over the dataset using a Python `for` loop, and\n", + "3. perform an optimization step in the body of your `for` loop.\n", + "\n", + "This workflow is exemplified in the following exercise." + ] + }, + { + "metadata": { + "id": "gj0-EkTc_Xt1", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "\n", + "\n", + "## Exercise!\n", + "\n", + "In this exercise, you'll train the convolutional model you implemented for the previous exericse on the MNIST dataset. " + ] + }, + { + "metadata": { + "id": "WOGm9HHn_byR", + "colab_type": "code", + "colab": { + "base_uri": "https://localhost:8080/", + "height": 17 + }, + "outputId": "bbccc7ad-33cd-446e-bcda-f358c7547e1b" + }, + "cell_type": "code", + "source": [ + "#@title Utilities for downloading MNIST data (double-click to show code)\n", + "import gzip\n", + "import os\n", + "import tempfile\n", + "from six.moves import urllib\n", + "import shutil\n", + "\n", + "import numpy as np\n", + "\n", + "def read32(bytestream):\n", + " \"\"\"Read 4 bytes from bytestream as an unsigned 32-bit integer.\"\"\"\n", + " dt = np.dtype(np.uint32).newbyteorder('>')\n", + " return np.frombuffer(bytestream.read(4), dtype=dt)[0]\n", + "\n", + "\n", + "def check_image_file_header(filename):\n", + " \"\"\"Validate that filename corresponds to images for the MNIST dataset.\"\"\"\n", + " with tf.gfile.Open(filename, 'rb') as f:\n", + " magic = read32(f)\n", + " read32(f) # num_images, unused\n", + " rows = read32(f)\n", + " cols = read32(f)\n", + " if magic != 2051:\n", + " raise ValueError('Invalid magic number %d in MNIST file %s' % (magic,\n", + " f.name))\n", + " if rows != 28 or cols != 28:\n", + " raise ValueError(\n", + " 'Invalid MNIST file %s: Expected 28x28 images, found %dx%d' %\n", + " (f.name, rows, cols))\n", + "\n", + "\n", + "def check_labels_file_header(filename):\n", + " \"\"\"Validate that filename corresponds to labels for the MNIST dataset.\"\"\"\n", + " with tf.gfile.Open(filename, 'rb') as f:\n", + " magic = read32(f)\n", + " read32(f) # num_items, unused\n", + " if magic != 2049:\n", + " raise ValueError('Invalid magic number %d in MNIST file %s' % (magic,\n", + " f.name))\n", + " \n", + "def download(directory, filename):\n", + " \"\"\"Download (and unzip) a file from the MNIST dataset if not already done.\"\"\"\n", + " filepath = os.path.join(directory, filename)\n", + " if tf.gfile.Exists(filepath):\n", + " return filepath\n", + " if not tf.gfile.Exists(directory):\n", + " tf.gfile.MakeDirs(directory)\n", + " # CVDF mirror of http://yann.lecun.com/exdb/mnist/\n", + " url = 'https://storage.googleapis.com/cvdf-datasets/mnist/' + filename + '.gz'\n", + " _, zipped_filepath = tempfile.mkstemp(suffix='.gz')\n", + " print('Downloading %s to %s' % (url, zipped_filepath))\n", + " urllib.request.urlretrieve(url, zipped_filepath)\n", + " with gzip.open(zipped_filepath, 'rb') as f_in, \\\n", + " tf.gfile.Open(filepath, 'wb') as f_out:\n", + " shutil.copyfileobj(f_in, f_out)\n", + " os.remove(zipped_filepath)\n", + " return filepath\n", + "\n", + "\n", + "def dataset(directory, images_file, labels_file):\n", + " \"\"\"Download and parse MNIST dataset.\"\"\"\n", + "\n", + " images_file = download(directory, images_file)\n", + " labels_file = download(directory, labels_file)\n", + "\n", + " check_image_file_header(images_file)\n", + " check_labels_file_header(labels_file)\n", + "\n", + " def decode_image(image):\n", + " # Normalize from [0, 255] to [0.0, 1.0]\n", + " image = tf.decode_raw(image, tf.uint8)\n", + " image = tf.cast(image, tf.float32)\n", + " image = tf.reshape(image, [28, 28, 1])\n", + " return image / 255.0\n", + "\n", + " def decode_label(label):\n", + " label = tf.decode_raw(label, tf.uint8) # tf.string -> [tf.uint8]\n", + " label = tf.reshape(label, []) # label is a scalar\n", + " return tf.to_int32(label)\n", + "\n", + " images = tf.data.FixedLengthRecordDataset(\n", + " images_file, 28 * 28, header_bytes=16).map(decode_image)\n", + " labels = tf.data.FixedLengthRecordDataset(\n", + " labels_file, 1, header_bytes=8).map(decode_label)\n", + " return tf.data.Dataset.zip((images, labels))\n", + "\n", + "\n", + "def get_training_data(directory):\n", + " \"\"\"tf.data.Dataset object for MNIST training data.\"\"\"\n", + " return dataset(directory, 'train-images-idx3-ubyte',\n", + " 'train-labels-idx1-ubyte').take(1024)\n", + "\n", + "def get_test_data(directory):\n", + " \"\"\"tf.data.Dataset object for MNIST test data.\"\"\"\n", + " return dataset(directory, 't10k-images-idx3-ubyte', 't10k-labels-idx1-ubyte')" + ], + "execution_count": 11, + "outputs": [] + }, + { + "metadata": { + "id": "4ejmJ2dv_f0R", + "colab_type": "code", + "colab": { + "base_uri": "https://localhost:8080/", + "height": 85 + }, + "outputId": "274c0381-e505-4e69-f910-3def6f8572a7" + }, + "cell_type": "code", + "source": [ + "# Don't forget to run the cell above!\n", + "training_data = get_training_data(\"/tmp/mnist/train\")\n", + "test_data = get_test_data(\"/tmp/mnist/test\")" + ], + "execution_count": 12, + "outputs": [ + { + "output_type": "stream", + "text": [ + "Downloading https://storage.googleapis.com/cvdf-datasets/mnist/train-images-idx3-ubyte.gz to /tmp/tmp4ull1xwa.gz\n", + "Downloading https://storage.googleapis.com/cvdf-datasets/mnist/train-labels-idx1-ubyte.gz to /tmp/tmp1eikhj1v.gz\n", + "Downloading https://storage.googleapis.com/cvdf-datasets/mnist/t10k-images-idx3-ubyte.gz to /tmp/tmpcp8xah9c.gz\n", + "Downloading https://storage.googleapis.com/cvdf-datasets/mnist/t10k-labels-idx1-ubyte.gz to /tmp/tmpqww_1e74.gz\n" + ], + "name": "stdout" + } + ] + }, + { + "metadata": { + "id": "TANpFS6GKLMC", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "Fill in the implementation of `train_one_epoch` below and run the cell to train your model. " + ] + }, + { + "metadata": { + "id": "btKL0Ss9_rmC", + "colab_type": "code", + "colab": { + "base_uri": "https://localhost:8080/", + "height": 102 + }, + "outputId": "56858516-86fc-424a-f00d-6f088f98bf9b" + }, + "cell_type": "code", + "source": [ + "EPOCHS = 5\n", + "optimizer = tf.train.MomentumOptimizer(learning_rate=0.01, momentum=0.5)\n", + "\n", + "def loss_fn(logits, labels):\n", + " return tf.reduce_mean(\n", + " tf.nn.sparse_softmax_cross_entropy_with_logits(\n", + " logits=tf.squeeze(logits), labels=labels))\n", + "\n", + "def train_one_epoch(model, training_data, optimizer):\n", + " # TODO: Implement an optimization step and return the average loss.\n", + " #\n", + " # Hint: Use `tf.GradientTape` to compute the gradient of the loss, and use\n", + " # `optimizer.apply_gradients` to update the model's variables, which are\n", + " # accessible as `model.variables`\n", + " average_loss = tfe.metrics.Mean('loss')\n", + " for images, labels in training_data.shuffle(buffer_size=10000).batch(64):\n", + " pass\n", + " return average_loss.result()\n", + "\n", + "for epoch in range(EPOCHS):\n", + " loss = train_one_epoch(model, training_data, optimizer)\n", + " print(\"Average loss after epoch %d: %.4f\" % (epoch, loss))" + ], + "execution_count": 14, + "outputs": [ + { + "output_type": "stream", + "text": [ + "Average loss after epoch 0: 2.2847\n", + "Average loss after epoch 1: 2.2305\n", + "Average loss after epoch 2: 2.1334\n", + "Average loss after epoch 3: 1.9115\n", + "Average loss after epoch 4: 1.4285\n" + ], + "name": "stdout" + } + ] + }, + { + "metadata": { + "id": "yAOFupJN_htg", + "colab_type": "code", + "colab": { + "base_uri": "https://localhost:8080/", + "height": 102 + }, + "outputId": "67e711e4-76c9-4e3f-bb49-a14955dba03a" + }, + "cell_type": "code", + "source": [ + "#@title Double-click to see a solution.\n", + "EPOCHS = 5\n", + "optimizer = tf.train.MomentumOptimizer(learning_rate=0.01, momentum=0.5)\n", + "\n", + "def _loss_fn(logits, labels):\n", + " return tf.reduce_mean(\n", + " tf.nn.sparse_softmax_cross_entropy_with_logits(\n", + " logits=tf.squeeze(logits), labels=labels))\n", + "\n", + "def _train_one_epoch(model, training_data):\n", + " average_loss = tfe.metrics.Mean(\"loss\")\n", + " for images, labels in training_data.shuffle(buffer_size=10000).batch(64):\n", + " with tf.GradientTape() as tape:\n", + " logits = model(images, training=True)\n", + " loss = _loss_fn(logits, labels)\n", + " average_loss(loss)\n", + " gradients = tape.gradient(loss, model.variables)\n", + " optimizer.apply_gradients(zip(gradients, model.variables))\n", + " return average_loss.result()\n", + " \n", + "for epoch in range(EPOCHS):\n", + " loss = _train_one_epoch(model, training_data)\n", + " print(\"Average loss after epoch %d: %.4f\" % (epoch, loss))" + ], + "execution_count": 15, + "outputs": [ + { + "output_type": "stream", + "text": [ + "Average loss after epoch 0: 1.0563\n", + "Average loss after epoch 1: 0.8013\n", + "Average loss after epoch 2: 0.6306\n", + "Average loss after epoch 3: 0.5543\n", + "Average loss after epoch 4: 0.5037\n" + ], + "name": "stdout" + } + ] + }, + { + "metadata": { + "id": "uDy1DrYA_2Jz", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "Run the below cell to qualitatively evaluate your model. Note how eager execution interoperates seamlessly with `matplotlib`." + ] + }, + { + "metadata": { + "id": "vR7rMtpu_3nB", + "colab_type": "code", + "colab": { + "base_uri": "https://localhost:8080/", + "height": 1752 + }, + "outputId": "b212aefa-f4b3-425c-f34d-2491429fa521" + }, + "cell_type": "code", + "source": [ + "import matplotlib.pyplot as plt\n", + "\n", + "sampled_data = test_data.batch(1).shuffle(buffer_size=10000).take(5)\n", + "for image, label in sampled_data:\n", + " plt.figure()\n", + " plt.imshow(tf.reshape(image, (28, 28)))\n", + " plt.show()\n", + " logits = model(image, training=False)\n", + " prediction = tf.argmax(logits, axis=1, output_type=tf.int64)\n", + " print(\"Prediction: %d\" % prediction)" + ], + "execution_count": 16, + "outputs": [ + { + "output_type": "display_data", + "data": { + "image/png": "iVBORw0KGgoAAAANSUhEUgAAAUsAAAFKCAYAAACU6307AAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nAKCzFeFbFn4BAAAAAElFTkSuQmCC\n", + "text/plain": [ + "" + ] + }, + "metadata": { + "tags": [] + } + }, + { + "output_type": "stream", + "text": [ + "Prediction: 5\n" + ], + "name": "stdout" + }, + { + "output_type": "display_data", + "data": { + "image/png": "iVBORw0KGgoAAAANSUhEUgAAAUsAAAFKCAYAAACU6307AAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nlgBg8H/nb4OLnfGqVAAAAABJRU5ErkJggg==\n", + "text/plain": [ + "" + ] + }, + "metadata": { + "tags": [] + } + }, + { + "output_type": "stream", + "text": [ + "Prediction: 1\n" + ], + "name": "stdout" + }, + { + "output_type": "display_data", + "data": { + "image/png": "iVBORw0KGgoAAAANSUhEUgAAAUsAAAFKCAYAAACU6307AAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nwIBYAoABsQQAA2IJAAbEEgAM/gepgR0uaefKmwAAAABJRU5ErkJggg==\n", + "text/plain": [ + "" + ] + }, + "metadata": { + "tags": [] + } + }, + { + "output_type": "stream", + "text": [ + "Prediction: 4\n" + ], + "name": "stdout" + }, + { + "output_type": "display_data", + "data": { + "image/png": "iVBORw0KGgoAAAANSUhEUgAAAUsAAAFKCAYAAACU6307AAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nN3rmU3AAAAAASUVORK5CYII=\n", + "text/plain": [ + "" + ] + }, + "metadata": { + "tags": [] + } + }, + { + "output_type": "stream", + "text": [ + "Prediction: 1\n" + ], + "name": "stdout" + }, + { + "output_type": "display_data", + "data": { + "image/png": "iVBORw0KGgoAAAANSUhEUgAAAUsAAAFKCAYAAACU6307AAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\niCUAGBBLADAglgBg8B9OkjtgR8VvdgAAAABJRU5ErkJggg==\n", + "text/plain": [ + "" + ] + }, + "metadata": { + "tags": [] + } + }, + { + "output_type": "stream", + "text": [ + "Prediction: 6\n" + ], + "name": "stdout" + } + ] + }, + { + "metadata": { + "id": "4SJizeJtNaAs", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "# Profiling\n", + "\n", + "If you want to drill down into the performance characteristics of your code, you can use native Python profilers like [`cProfile`](https://docs.python.org/3/library/profile.html). In the next exercise, you'll do just that." + ] + }, + { + "metadata": { + "id": "_2v0QnG8__PJ", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "## Exercise!\n", + "\n", + "This exercise does not require coding. If you have not completed the training exercise, replace `train_one_epoch` below with `_train_one_epoch`.\n", + "\n", + "Run the below cell and inspect the printed profiles. What parts of the code appear to be hotspots or\n", + "bottlenecks? How does sorting the profile by total time compare to sorting it\n", + "by cumulative time?\n", + "\n" + ] + }, + { + "metadata": { + "id": "IFypaYbG_9fB", + "colab_type": "code", + "colab": { + "base_uri": "https://localhost:8080/", + "height": 714 + }, + "outputId": "d9c3596b-a165-4edd-fc6b-53ccd0d01d19" + }, + "cell_type": "code", + "source": [ + "import cProfile\n", + "import pstats\n", + "\n", + "cProfile.run(\"train_one_epoch(model, training_data, optimizer)\", \"training_profile\")\n", + "\n", + "stats = pstats.Stats(\"training_profile\").strip_dirs().sort_stats(\"tottime\")\n", + "stats.print_stats(10)\n", + "\n", + "stats.sort_stats(\"cumtime\").print_stats(10)" + ], + "execution_count": 17, + "outputs": [ + { + "output_type": "stream", + "text": [ + "Thu Jun 7 12:25:04 2018 training_profile\n", + "\n", + " 92209 function calls (91817 primitive calls) in 3.446 seconds\n", + "\n", + " Ordered by: internal time\n", + " List reduced from 672 to 10 due to restriction <10>\n", + "\n", + " ncalls tottime percall cumtime percall filename:lineno(function)\n", + " 1080 2.552 0.002 2.552 0.002 {built-in method _pywrap_tensorflow_internal.TFE_Py_FastPathExecute}\n", + " 83 0.753 0.009 0.753 0.009 {built-in method _pywrap_tensorflow_internal.TFE_Py_Execute}\n", + " 16 0.006 0.000 1.019 0.064 network.py:736(_run_internal_graph)\n", + " 16 0.005 0.000 2.253 0.141 {built-in method _pywrap_tensorflow_internal.TFE_Py_TapeGradient}\n", + " 2321 0.004 0.000 0.007 0.000 abc.py:178(__instancecheck__)\n", + " 288 0.004 0.000 0.009 0.000 inspect.py:2092(_signature_from_function)\n", + " 878 0.004 0.000 0.005 0.000 ops.py:5936(__enter__)\n", + " 288 0.004 0.000 0.016 0.000 inspect.py:1079(getfullargspec)\n", + " 11006 0.003 0.000 0.005 0.000 {built-in method builtins.isinstance}\n", + " 768 0.003 0.000 0.008 0.000 {built-in method _pywrap_tensorflow_internal.Flatten}\n", + "\n", + "\n", + "Thu Jun 7 12:25:04 2018 training_profile\n", + "\n", + " 92209 function calls (91817 primitive calls) in 3.446 seconds\n", + "\n", + " Ordered by: cumulative time\n", + " List reduced from 672 to 10 due to restriction <10>\n", + "\n", + " ncalls tottime percall cumtime percall filename:lineno(function)\n", + " 1 0.000 0.000 3.446 3.446 {built-in method builtins.exec}\n", + " 1 0.000 0.000 3.446 3.446 :1()\n", + " 1 0.001 0.001 3.446 3.446 :9(train_one_epoch)\n", + " 1080 2.552 0.002 2.552 0.002 {built-in method _pywrap_tensorflow_internal.TFE_Py_FastPathExecute}\n", + " 16 0.000 0.000 2.255 0.141 backprop.py:739(gradient)\n", + " 16 0.000 0.000 2.253 0.141 imperative_grad.py:31(imperative_grad)\n", + " 16 0.005 0.000 2.253 0.141 {built-in method _pywrap_tensorflow_internal.TFE_Py_TapeGradient}\n", + " 400 0.002 0.000 2.246 0.006 backprop.py:145(grad_fn)\n", + " 400 0.002 0.000 2.239 0.006 backprop.py:95(_magic_gradient_function)\n", + " 32 0.001 0.000 1.601 0.050 nn_grad.py:497(_Conv2DGrad)\n", + "\n", + "\n" + ], + "name": "stdout" + }, + { + "output_type": "execute_result", + "data": { + "text/plain": [ + "" + ] + }, + "metadata": { + "tags": [] + }, + "execution_count": 17 + } + ] + }, + { + "metadata": { + "id": "8ixpnyCNNTI4", + "colab_type": "code", + "colab": {} + }, + "cell_type": "code", + "source": [ + "" + ], + "execution_count": 0, + "outputs": [] + } + ] +} \ No newline at end of file diff --git a/tensorflow/contrib/eager/python/examples/workshop/3_inspecting.ipynb b/tensorflow/contrib/eager/python/examples/workshop/3_inspecting.ipynb new file mode 100644 index 0000000000000000000000000000000000000000..64d19ec5c9bfccd07eabb21ce8fbb62b21f23efa --- /dev/null +++ b/tensorflow/contrib/eager/python/examples/workshop/3_inspecting.ipynb @@ -0,0 +1,443 @@ +{ + "nbformat": 4, + "nbformat_minor": 0, + "metadata": { + "colab": { + "name": "Debugging \"graph-first\" models with eager execution", + "version": "0.3.2", + "provenance": [], + "include_colab_link": true + } + }, + "cells": [ + { + "cell_type": "markdown", + "metadata": { + "id": "view-in-github", + "colab_type": "text" + }, + "source": [ + "[View in Colaboratory](https://colab.research.google.com/gist/alextp/9568ab40f6ed6f9a3ba4736f6aef6127/debugging-graph-first-models-with-eager-execution.ipynb)" + ] + }, + { + "metadata": { + "id": "mm-t0GuIu1Dt", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "This colab uses eager execution and the Python debugger to modify the execution of a translation model. This combination lets you quickly explore counterfactuals when researching and designing modifications to a model.\n", + "\n", + "The model, Transformer from [Tensor2Tensor](https://github.com/tensorflow/tensor2tensor), was originally written with graph building in mind. Executing it eagerly can still be helpful!" + ] + }, + { + "metadata": { + "id": "gxb1DvIDg4sv", + "colab_type": "code", + "colab": {} + }, + "cell_type": "code", + "source": [ + "#@title License (double click to show)\n", + "# Copyright 2018 The TensorFlow Authors.\n", + "\n", + "# Licensed under the Apache License, Version 2.0 (the \"License\");\n", + "# you may not use this file except in compliance with the License.\n", + "# You may obtain a copy of the License at\n", + "\n", + "# https://www.apache.org/licenses/LICENSE-2.0\n", + "\n", + "# Unless required by applicable law or agreed to in writing, software\n", + "# distributed under the License is distributed on an \"AS IS\" BASIS,\n", + "# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n", + "# See the License for the specific language governing permissions and\n", + "# limitations under the License." + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "Gx3HA9N1ui64", + "colab_type": "code", + "colab": { + "base_uri": "https://localhost:8080/", + "height": 37 + }, + "outputId": "f6986f34-f3e1-44e1-c902-2eb33081acad" + }, + "cell_type": "code", + "source": [ + "import tensorflow as tf\n", + "import pdb\n", + "tfe = tf.contrib.eager\n", + "\n", + "tf.enable_eager_execution()" + ], + "execution_count": 1, + "outputs": [] + }, + { + "metadata": { + "id": "3LkOm2ct-Lmc", + "colab_type": "code", + "colab": { + "base_uri": "https://localhost:8080/", + "height": 37 + }, + "outputId": "2edc74d9-6bc0-4e78-ab4e-83bf96099ef4" + }, + "cell_type": "code", + "source": [ + "!pip install -q -U tensor2tensor\n", + "from tensor2tensor.models import transformer" + ], + "execution_count": 2, + "outputs": [] + }, + { + "metadata": { + "id": "1Z3oMsqV0zB6", + "colab_type": "code", + "colab": { + "base_uri": "https://localhost:8080/", + "height": 170 + }, + "outputId": "0a8186ee-c688-457f-c9f6-9a6c1477a93b" + }, + "cell_type": "code", + "source": [ + "#@title Create a tensor2tensor translation model, fetch a checkpoint (double click to show)\n", + "from tensor2tensor import problems\n", + "from tensor2tensor.utils import trainer_lib\n", + "from tensor2tensor.utils import registry\n", + "\n", + "import numpy as np\n", + "import os\n", + "\n", + "# Setup some directories\n", + "data_dir = os.path.expanduser(\"~/t2t/data\")\n", + "tmp_dir = os.path.expanduser(\"~/t2t/tmp\")\n", + "train_dir = os.path.expanduser(\"~/t2t/train\")\n", + "checkpoint_dir = os.path.expanduser(\"~/t2t/checkpoints\")\n", + "tf.gfile.MakeDirs(data_dir)\n", + "tf.gfile.MakeDirs(tmp_dir)\n", + "tf.gfile.MakeDirs(train_dir)\n", + "tf.gfile.MakeDirs(checkpoint_dir)\n", + "gs_data_dir = \"gs://tensor2tensor-data\"\n", + "gs_ckpt_dir = \"gs://tensor2tensor-checkpoints/\"\n", + "\n", + "# Fetch the problem\n", + "ende_problem = problems.problem(\"translate_ende_wmt32k\")\n", + "\n", + "# Copy the vocab file locally so we can encode inputs and decode model outputs\n", + "# All vocabs are stored on GCS\n", + "vocab_name = \"vocab.ende.32768\"\n", + "vocab_file = os.path.join(gs_data_dir, vocab_name)\n", + "!gsutil cp {vocab_file} {data_dir}\n", + "\n", + "# Get the encoders from the problem\n", + "encoders = ende_problem.feature_encoders(data_dir)\n", + "\n", + "# Setup helper functions for encoding and decoding\n", + "def encode(input_str, output_str=None):\n", + " \"\"\"Input str to features dict, ready for inference\"\"\"\n", + " inputs = encoders[\"inputs\"].encode(input_str) + [1] # add EOS id\n", + " batch_inputs = tf.reshape(inputs, [1, -1, 1]) # Make it 3D.\n", + " return {\"inputs\": batch_inputs}\n", + "\n", + "def decode(integers):\n", + " \"\"\"List of ints to str\"\"\"\n", + " integers = list(np.squeeze(integers))\n", + " if 1 in integers:\n", + " integers = integers[:integers.index(1)]\n", + " return encoders[\"inputs\"].decode(np.squeeze(integers))\n", + "\n", + "# Copy the pretrained checkpoint locally\n", + "ckpt_name = \"transformer_ende_test\"\n", + "gs_ckpt = os.path.join(gs_ckpt_dir, ckpt_name)\n", + "!gsutil -q cp -R {gs_ckpt} {checkpoint_dir}\n", + "checkpoint_path = tf.train.latest_checkpoint(\n", + " os.path.join(checkpoint_dir, ckpt_name))\n", + "\n", + "# Create hparams and the model\n", + "model_name = \"transformer\"\n", + "hparams_set = \"transformer_base\"\n", + "\n", + "hparams = trainer_lib.create_hparams(hparams_set, data_dir=data_dir, problem_name=\"translate_ende_wmt32k\")\n", + "\n", + "# NOTE: Only create the model once when restoring from a checkpoint; it's a\n", + "# Layer and so subsequent instantiations will have different variable scopes\n", + "# that will not match the checkpoint.\n", + "translate_model = registry.model(model_name)(hparams, tf.estimator.ModeKeys.EVAL)" + ], + "execution_count": 3, + "outputs": [ + { + "output_type": "stream", + "text": [ + "Copying gs://tensor2tensor-data/vocab.ende.32768...\n", + "/ [1 files][316.4 KiB/316.4 KiB] \n", + "Operation completed over 1 objects/316.4 KiB. \n", + "INFO:tensorflow:Setting T2TModel mode to 'eval'\n", + "INFO:tensorflow:Setting hparams.layer_prepostprocess_dropout to 0.0\n", + "INFO:tensorflow:Setting hparams.symbol_dropout to 0.0\n", + "INFO:tensorflow:Setting hparams.attention_dropout to 0.0\n", + "INFO:tensorflow:Setting hparams.dropout to 0.0\n", + "INFO:tensorflow:Setting hparams.relu_dropout to 0.0\n" + ], + "name": "stdout" + } + ] + }, + { + "metadata": { + "id": "4IblPXLGjuCl", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "We've created a Transformer model and fetched an existing training checkpoint. It hasn't created variables yet, and we want to load them from the checkpoint before they're used (restore-on-create) so the first run of the model outputs the correct value. The `tfe.restore_variables_on_create` API looks up variables by name on creation and restores their values." + ] + }, + { + "metadata": { + "id": "o3MWxcAqJoqG", + "colab_type": "code", + "colab": { + "base_uri": "https://localhost:8080/", + "height": 51 + }, + "outputId": "fbc1b1bf-ffbe-4621-b3cb-5eb855fec3a8" + }, + "cell_type": "code", + "source": [ + "with tfe.restore_variables_on_create(checkpoint_path):\n", + " model_output = translate_model.infer(encode(\"Eager execution\"))\n", + "print(decode(model_output[\"outputs\"]))" + ], + "execution_count": 4, + "outputs": [ + { + "output_type": "stream", + "text": [ + "INFO:tensorflow:Greedy Decoding\n", + "Hinrichtung\n" + ], + "name": "stdout" + } + ] + }, + { + "metadata": { + "id": "xk5HV9Hhu9zO", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "Using global variable names can get somewhat fragile, so for new code we recommend the object-based `tf.keras.Model.save_weights` or `tf.train.Checkpoint`. However, these require some small code changes to work with existing graph building code.\n", + "\n", + "The Transformer model translates \"Eager execution\" in English to \"Hinrichtung\" in German, which refers to capital punishment rather than getting things done. Transformer first encodes the English, then decodes to German. We'll add a debugging hook at the start of the decode phase (once the encodings have been finalized) and see if we can correct the translation." + ] + }, + { + "metadata": { + "id": "GUGwbYvXZ9-7", + "colab_type": "code", + "colab": {} + }, + "cell_type": "code", + "source": [ + "previous_fast_decode = transformer.fast_decode\n", + "def debug_fn(*args, **kwargs):\n", + " pdb.set_trace()\n", + " return previous_fast_decode(*args, **kwargs) # \"step\" in pdb to step in\n", + "transformer.fast_decode = debug_fn # Add our debugging hook to Transformer" + ], + "execution_count": 0, + "outputs": [] + }, + { + "metadata": { + "id": "f61HlvECxJn0", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "Now that we've \"monkey patched\" the model, we'll drop into a debugger just before decoding starts. In most cases it'd be simpler to add the `pdb.set_trace()` call to the code directly, but in this case we're working with prepackaged library code.\n", + "\n", + "First, let's find an encoding which represents the correct sense of \"execution\". Then we'll patch part of that encoding into the encoding of \"Eager execution\" to fix the translation. Feel free to poke around with the debugger (e.g. print a Tensor's value), but your main task is to save the encodings by assigning them to an attribute of the function:\n", + "\n", + "```\n", + "(running the next cell drops you into a pdb shell)\n", + "step\n", + "fast_decode.previous_encoding = encoder_output\n", + "continue\n", + "\n", + "```\n", + "\n", + "You can type `next` (or `n`) a few times before `continue` to watch the decoding ops run." + ] + }, + { + "metadata": { + "id": "dX4CPOGSpZrb", + "colab_type": "code", + "colab": { + "base_uri": "https://localhost:8080/", + "height": 179 + }, + "outputId": "6de38c31-836f-40ef-b701-e42908172619" + }, + "cell_type": "code", + "source": [ + "model_output = translate_model.infer(encode(\"Immediate running\"))\n", + "print(decode(model_output[\"outputs\"]))" + ], + "execution_count": 7, + "outputs": [ + { + "output_type": "stream", + "text": [ + "> (4)debug_fn()\n", + "-> return previous_fast_decode(*args, **kwargs) # \"step\" in pdb to step in\n", + "(Pdb) step\n", + "--Call--\n", + "> /usr/local/lib/python2.7/dist-packages/tensor2tensor/models/transformer.py(427)fast_decode()\n", + "-> def fast_decode(encoder_output,\n", + "(Pdb) fast_decode.previous_encoding = encoder_output\n", + "(Pdb) continue\n", + "Sofortige DurchfĆ¼hrung\n" + ], + "name": "stdout" + } + ] + }, + { + "metadata": { + "id": "-ZEZciV4FpLo", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "Now we have an encoding saved which gets the correct sense for \"execution\"." + ] + }, + { + "metadata": { + "id": "QeC_oDVqHD_v", + "colab_type": "code", + "colab": { + "base_uri": "https://localhost:8080/", + "height": 179 + }, + "outputId": "253c9af1-003e-46bd-8bf5-db968cf6a8cf" + }, + "cell_type": "code", + "source": [ + "# Assumes you followed the pdb instructions above!\n", + "transformer.fast_decode.previous_encoding" + ], + "execution_count": 8, + "outputs": [ + { + "output_type": "execute_result", + "data": { + "text/plain": [ + "" + ] + }, + "metadata": { + "tags": [] + }, + "execution_count": 8 + } + ] + }, + { + "metadata": { + "id": "bC9JjeDcHEav", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "Let's replace part of the encoding for \"Eager execution\" with the encoding of \"Immediate running\".\n", + "\n", + "Again we'll drop into a pdb shell. This time we'll run some TensorFlow operations to patch the encodings while the model is running.\n", + "\n", + "```\n", + "(running the next cell again drops you into a pdb shell)\n", + "step\n", + "encoder_output = tf.concat([fast_decode.previous_encoding[:, :3], encoder_output[:, 3:]], axis=1)\n", + "continue\n", + "```" + ] + }, + { + "metadata": { + "id": "t2as_Kn1h65G", + "colab_type": "code", + "colab": { + "base_uri": "https://localhost:8080/", + "height": 179 + }, + "outputId": "5b4e546e-3bb4-4761-c545-467b631e3ffe" + }, + "cell_type": "code", + "source": [ + "model_output = translate_model.infer(encode(\"Eager execution\"))\n", + "print(decode(model_output[\"outputs\"]))" + ], + "execution_count": 9, + "outputs": [ + { + "output_type": "stream", + "text": [ + "> (4)debug_fn()\n", + "-> return previous_fast_decode(*args, **kwargs) # \"step\" in pdb to step in\n", + "(Pdb) step\n", + "--Call--\n", + "> /usr/local/lib/python2.7/dist-packages/tensor2tensor/models/transformer.py(427)fast_decode()\n", + "-> def fast_decode(encoder_output,\n", + "(Pdb) encoder_output = tf.concat([fast_decode.previous_encoding[:, :3], encoder_output[:, 3:]], axis=1)\n", + "(Pdb) continue\n", + "sofortige AusfĆ¼hrung\n" + ], + "name": "stdout" + } + ] + }, + { + "metadata": { + "id": "rK6tYZ23I2cm", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "We get a different decoding, with the correct sense of \"execution\". Likely we're keeping just the encoding of \"tion\" from \"Eager execution\", so no great breakthrough in translation modeling.\n", + "\n", + "Similarly it's possible to modify attention vectors, or change words during decoding to help debug a beam search." + ] + }, + { + "metadata": { + "id": "Nb-4ipYNRWxA", + "colab_type": "text" + }, + "cell_type": "markdown", + "source": [ + "This colab was adapted from the [Tensor2Tensor colab](https://colab.research.google.com/github/tensorflow/tensor2tensor/blob/master/tensor2tensor/notebooks/hello_t2t.ipynb). Credit to Ankur Taly for its concept." + ] + } + ] +} \ No newline at end of file diff --git a/tensorflow/contrib/estimator/BUILD b/tensorflow/contrib/estimator/BUILD index 30d297a5fb2dd2f844093d790d051a79105984dd..11d40f5982ac81d7f3d32cada3457037280801cb 100644 --- a/tensorflow/contrib/estimator/BUILD +++ b/tensorflow/contrib/estimator/BUILD @@ -18,6 +18,7 @@ py_library( ":boosted_trees", ":dnn", ":dnn_linear_combined", + ":early_stopping", ":export", ":extenders", ":head", @@ -590,3 +591,31 @@ py_test( "@six_archive//:six", ], ) + +py_library( + name = "early_stopping", + srcs = ["python/estimator/early_stopping.py"], + srcs_version = "PY2AND3", + deps = [ + "//tensorflow/python:dtypes", + "//tensorflow/python:framework_ops", + "//tensorflow/python:init_ops", + "//tensorflow/python:platform", + "//tensorflow/python:state_ops", + "//tensorflow/python:summary", + "//tensorflow/python:training", + "//tensorflow/python/estimator", + ], +) + +py_test( + name = "early_stopping_test", + srcs = ["python/estimator/early_stopping_test.py"], + srcs_version = "PY2AND3", + deps = [ + ":early_stopping", + "//tensorflow/python:client_testlib", + "//tensorflow/python/estimator", + "@absl_py//absl/testing:parameterized", + ], +) diff --git a/tensorflow/contrib/estimator/__init__.py b/tensorflow/contrib/estimator/__init__.py index 788ac5ca7046d6dd30a3d5520b243944532622fa..09fcfd66a1f90d5a323b09472c4b7f5b1234ee35 100644 --- a/tensorflow/contrib/estimator/__init__.py +++ b/tensorflow/contrib/estimator/__init__.py @@ -23,6 +23,7 @@ from tensorflow.contrib.estimator.python.estimator.baseline import * from tensorflow.contrib.estimator.python.estimator.boosted_trees import * from tensorflow.contrib.estimator.python.estimator.dnn import * from tensorflow.contrib.estimator.python.estimator.dnn_linear_combined import * +from tensorflow.contrib.estimator.python.estimator.early_stopping import * from tensorflow.contrib.estimator.python.estimator.export import * from tensorflow.contrib.estimator.python.estimator.extenders import * from tensorflow.contrib.estimator.python.estimator.head import * @@ -63,6 +64,12 @@ _allowed_symbols = [ 'RNNEstimator', 'export_saved_model_for_mode', 'export_all_saved_models', + 'make_early_stopping_hook', + 'read_eval_metrics', + 'stop_if_lower_hook', + 'stop_if_higher_hook', + 'stop_if_no_increase_hook', + 'stop_if_no_decrease_hook', ] remove_undocumented(__name__, allowed_exception_list=_allowed_symbols) diff --git a/tensorflow/contrib/estimator/python/estimator/baseline_test.py b/tensorflow/contrib/estimator/python/estimator/baseline_test.py index d0e3e670f7332811c1bfdaea65b0308ce59ade59..505c94e97192afdd4e2ce9af2abb9825320751f2 100644 --- a/tensorflow/contrib/estimator/python/estimator/baseline_test.py +++ b/tensorflow/contrib/estimator/python/estimator/baseline_test.py @@ -113,6 +113,8 @@ class BaselineEstimatorEvaluationTest(test.TestCase): self.assertDictEqual({ metric_keys.MetricKeys.LOSS: 18., metric_keys.MetricKeys.LOSS_MEAN: 9., + metric_keys.MetricKeys.PREDICTION_MEAN: 13., + metric_keys.MetricKeys.LABEL_MEAN: 10., ops.GraphKeys.GLOBAL_STEP: 100 }, eval_metrics) @@ -141,6 +143,8 @@ class BaselineEstimatorEvaluationTest(test.TestCase): self.assertDictEqual({ metric_keys.MetricKeys.LOSS: 27., metric_keys.MetricKeys.LOSS_MEAN: 9., + metric_keys.MetricKeys.PREDICTION_MEAN: 13., + metric_keys.MetricKeys.LABEL_MEAN: 10., ops.GraphKeys.GLOBAL_STEP: 100 }, eval_metrics) @@ -166,7 +170,9 @@ class BaselineEstimatorEvaluationTest(test.TestCase): self.assertItemsEqual( (metric_keys.MetricKeys.LOSS, metric_keys.MetricKeys.LOSS_MEAN, - ops.GraphKeys.GLOBAL_STEP), eval_metrics.keys()) + metric_keys.MetricKeys.PREDICTION_MEAN, + metric_keys.MetricKeys.LABEL_MEAN, ops.GraphKeys.GLOBAL_STEP), + eval_metrics.keys()) # Logit is bias which is [46, 58] self.assertAlmostEqual(0, eval_metrics[metric_keys.MetricKeys.LOSS]) diff --git a/tensorflow/contrib/estimator/python/estimator/boosted_trees.py b/tensorflow/contrib/estimator/python/estimator/boosted_trees.py index bd641014e9eec6623d66574bccd08ff03ebc28ac..43bfcffd790e7b3c716c3f70820851a8819af225 100644 --- a/tensorflow/contrib/estimator/python/estimator/boosted_trees.py +++ b/tensorflow/contrib/estimator/python/estimator/boosted_trees.py @@ -49,7 +49,8 @@ class _BoostedTreesEstimator(estimator.Estimator): l2_regularization=0., tree_complexity=0., min_node_weight=0., - config=None): + config=None, + center_bias=False): """Initializes a `BoostedTreesEstimator` instance. Args: @@ -82,17 +83,30 @@ class _BoostedTreesEstimator(estimator.Estimator): considered. The value will be compared with sum(leaf_hessian)/ (batch_size * n_batches_per_layer). config: `RunConfig` object to configure the runtime settings. + center_bias: Whether bias centering needs to occur. Bias centering refers + to the first node in the very first tree returning the prediction that + is aligned with the original labels distribution. For example, for + regression problems, the first node will return the mean of the labels. + For binary classification problems, it will return a logit for a prior + probability of label 1. + """ # pylint:disable=protected-access # HParams for the model. tree_hparams = canned_boosted_trees._TreeHParams( n_trees, max_depth, learning_rate, l1_regularization, l2_regularization, - tree_complexity, min_node_weight) + tree_complexity, min_node_weight, center_bias) def _model_fn(features, labels, mode, config): return canned_boosted_trees._bt_model_fn( - features, labels, mode, head, feature_columns, tree_hparams, - n_batches_per_layer, config) + features, + labels, + mode, + head, + feature_columns, + tree_hparams, + n_batches_per_layer, + config=config) super(_BoostedTreesEstimator, self).__init__( model_fn=_model_fn, model_dir=model_dir, config=config) @@ -114,7 +128,8 @@ def boosted_trees_classifier_train_in_memory( tree_complexity=0., min_node_weight=0., config=None, - train_hooks=None): + train_hooks=None, + center_bias=False): """Trains a boosted tree classifier with in memory dataset. Example: @@ -186,7 +201,13 @@ def boosted_trees_classifier_train_in_memory( considered. The value will be compared with sum(leaf_hessian)/ (batch_size * n_batches_per_layer). config: `RunConfig` object to configure the runtime settings. - train_hooks: a list of Hook instances to be passed to estimator.train(). + train_hooks: a list of Hook instances to be passed to estimator.train() + center_bias: Whether bias centering needs to occur. Bias centering refers + to the first node in the very first tree returning the prediction that + is aligned with the original labels distribution. For example, for + regression problems, the first node will return the mean of the labels. + For binary classification problems, it will return a logit for a prior + probability of label 1. Returns: a `BoostedTreesClassifier` instance created with the given arguments and @@ -207,7 +228,7 @@ def boosted_trees_classifier_train_in_memory( # HParams for the model. tree_hparams = canned_boosted_trees._TreeHParams( n_trees, max_depth, learning_rate, l1_regularization, l2_regularization, - tree_complexity, min_node_weight) + tree_complexity, min_node_weight, center_bias) def _model_fn(features, labels, mode, config): return canned_boosted_trees._bt_model_fn( @@ -247,7 +268,8 @@ def boosted_trees_regressor_train_in_memory( tree_complexity=0., min_node_weight=0., config=None, - train_hooks=None): + train_hooks=None, + center_bias=False): """Trains a boosted tree regressor with in memory dataset. Example: @@ -313,6 +335,12 @@ def boosted_trees_regressor_train_in_memory( (batch_size * n_batches_per_layer). config: `RunConfig` object to configure the runtime settings. train_hooks: a list of Hook instances to be passed to estimator.train(). + center_bias: Whether bias centering needs to occur. Bias centering refers + to the first node in the very first tree returning the prediction that + is aligned with the original labels distribution. For example, for + regression problems, the first node will return the mean of the labels. + For binary classification problems, it will return a logit for a prior + probability of label 1. Returns: a `BoostedTreesClassifier` instance created with the given arguments and @@ -332,7 +360,7 @@ def boosted_trees_regressor_train_in_memory( # HParams for the model. tree_hparams = canned_boosted_trees._TreeHParams( n_trees, max_depth, learning_rate, l1_regularization, l2_regularization, - tree_complexity, min_node_weight) + tree_complexity, min_node_weight, center_bias) def _model_fn(features, labels, mode, config): return canned_boosted_trees._bt_model_fn( diff --git a/tensorflow/contrib/estimator/python/estimator/boosted_trees_test.py b/tensorflow/contrib/estimator/python/estimator/boosted_trees_test.py index 76cbefe5e94502188388df6fc2816d130ac896d5..999c2aa5e28242f996e12da3807a74c6acf31df9 100644 --- a/tensorflow/contrib/estimator/python/estimator/boosted_trees_test.py +++ b/tensorflow/contrib/estimator/python/estimator/boosted_trees_test.py @@ -115,6 +115,27 @@ class BoostedTreesEstimatorTest(test_util.TensorFlowTestCase): eval_res = est.evaluate(input_fn=input_fn, steps=1) self.assertAllClose(eval_res['average_loss'], 1.008551) + def testTrainAndEvaluateEstimatorWithCenterBias(self): + input_fn = _make_train_input_fn(is_classification=False) + + est = boosted_trees._BoostedTreesEstimator( + feature_columns=self._feature_columns, + n_batches_per_layer=1, + n_trees=2, + head=self._head, + max_depth=5, + center_bias=True) + + # It will stop after 11 steps because of the max depth and num trees. + num_steps = 100 + # Train for a few steps, and validate final checkpoint. + est.train(input_fn, steps=num_steps) + # 10 steps for training and 2 step for bias centering. + self._assert_checkpoint( + est.model_dir, global_step=12, finalized_trees=2, attempted_layers=10) + eval_res = est.evaluate(input_fn=input_fn, steps=1) + self.assertAllClose(eval_res['average_loss'], 0.614642) + def testInferEstimator(self): train_input_fn = _make_train_input_fn(is_classification=False) predict_input_fn = numpy_io.numpy_input_fn( @@ -139,6 +160,33 @@ class BoostedTreesEstimatorTest(test_util.TensorFlowTestCase): [[0.571619], [0.262821], [0.124549], [0.956801], [1.769801]], [pred['predictions'] for pred in predictions]) + def testInferEstimatorWithCenterBias(self): + train_input_fn = _make_train_input_fn(is_classification=False) + predict_input_fn = numpy_io.numpy_input_fn( + x=FEATURES_DICT, y=None, batch_size=1, num_epochs=1, shuffle=False) + + est = boosted_trees._BoostedTreesEstimator( + feature_columns=self._feature_columns, + n_batches_per_layer=1, + n_trees=1, + max_depth=5, + center_bias=True, + head=self._head) + + # It will stop after 6 steps because of the max depth and num trees (5 for + # training and 2 for bias centering). + num_steps = 100 + # Train for a few steps, and validate final checkpoint. + est.train(train_input_fn, steps=num_steps) + self._assert_checkpoint( + est.model_dir, global_step=7, finalized_trees=1, attempted_layers=5) + # Validate predictions. + predictions = list(est.predict(input_fn=predict_input_fn)) + + self.assertAllClose( + [[1.634501], [1.325703], [1.187431], [2.019683], [2.832683]], + [pred['predictions'] for pred in predictions]) + def testBinaryClassifierTrainInMemoryAndEvalAndInfer(self): train_input_fn = _make_train_input_fn(is_classification=True) predict_input_fn = numpy_io.numpy_input_fn( @@ -159,14 +207,40 @@ class BoostedTreesEstimatorTest(test_util.TensorFlowTestCase): self.assertAllClose([[0], [1], [1], [0], [0]], [pred['class_ids'] for pred in predictions]) + def testBinaryClassifierTrainInMemoryAndEvalAndInferWithCenterBias(self): + train_input_fn = _make_train_input_fn(is_classification=True) + predict_input_fn = numpy_io.numpy_input_fn( + x=FEATURES_DICT, y=None, batch_size=1, num_epochs=1, shuffle=False) + + est = boosted_trees.boosted_trees_classifier_train_in_memory( + train_input_fn=train_input_fn, + feature_columns=self._feature_columns, + n_trees=1, + max_depth=5, + center_bias=True) + # It will stop after 5 steps + 3 for bias, because of the max depth and num + # trees. + self._assert_checkpoint( + est.model_dir, global_step=8, finalized_trees=1, attempted_layers=5) + + # Check evaluate and predict. + eval_res = est.evaluate(input_fn=train_input_fn, steps=1) + self.assertAllClose(eval_res['accuracy'], 1.0) + # Validate predictions. + predictions = list(est.predict(input_fn=predict_input_fn)) + self.assertAllClose([[0], [1], [1], [0], [0]], + [pred['class_ids'] for pred in predictions]) + def testBinaryClassifierTrainInMemoryWithDataset(self): train_input_fn = _make_train_input_fn_dataset(is_classification=True) predict_input_fn = numpy_io.numpy_input_fn( x=FEATURES_DICT, y=None, batch_size=1, num_epochs=1, shuffle=False) est = boosted_trees.boosted_trees_classifier_train_in_memory( - train_input_fn=train_input_fn, feature_columns=self._feature_columns, - n_trees=1, max_depth=5) + train_input_fn=train_input_fn, + feature_columns=self._feature_columns, + n_trees=1, + max_depth=5) # It will stop after 5 steps because of the max depth and num trees. self._assert_checkpoint( est.model_dir, global_step=5, finalized_trees=1, attempted_layers=5) diff --git a/tensorflow/contrib/estimator/python/estimator/dnn.py b/tensorflow/contrib/estimator/python/estimator/dnn.py index f1c60a912c8b1daa7db34f46e92bcc36ab300716..9efa8f474d865a36788cba40a15404bf0b30a17e 100644 --- a/tensorflow/contrib/estimator/python/estimator/dnn.py +++ b/tensorflow/contrib/estimator/python/estimator/dnn.py @@ -53,6 +53,18 @@ class DNNEstimator(estimator.Estimator): l1_regularization_strength=0.001 )) + # Or estimator using an optimizer with a learning rate decay. + estimator = DNNEstimator( + head=tf.contrib.estimator.multi_label_head(n_classes=3), + feature_columns=[sparse_feature_a_emb, sparse_feature_b_emb], + hidden_units=[1024, 512, 256], + optimizer=lambda: tf.AdamOptimizer( + learning_rate=tf.exponential_decay( + learning_rate=0.1, + global_step=tf.get_global_step(), + decay_steps=10000, + decay_rate=0.96)) + # Or estimator with warm-starting from a previous checkpoint. estimator = DNNEstimator( head=tf.contrib.estimator.multi_label_head(n_classes=3), @@ -100,7 +112,8 @@ class DNNEstimator(estimator.Estimator): dropout=None, input_layer_partitioner=None, config=None, - warm_start_from=None): + warm_start_from=None, + batch_norm=False): """Initializes a `DNNEstimator` instance. Args: @@ -115,8 +128,9 @@ class DNNEstimator(estimator.Estimator): model_dir: Directory to save model parameters, graph and etc. This can also be used to load checkpoints from the directory into a estimator to continue training a previously saved model. - optimizer: An instance of `tf.Optimizer` used to train the model. Defaults - to Adagrad optimizer. + optimizer: An instance of `tf.Optimizer` used to train the model. Can also + be a string (one of 'Adagrad', 'Adam', 'Ftrl', 'RMSProp', 'SGD'), or + callable. Defaults to Adagrad optimizer. activation_fn: Activation function applied to each layer. If `None`, will use `tf.nn.relu`. dropout: When not `None`, the probability we will drop out a given @@ -129,6 +143,7 @@ class DNNEstimator(estimator.Estimator): string filepath is provided instead of a `WarmStartSettings`, then all weights are warm-started, and it is assumed that vocabularies and Tensor names are unchanged. + batch_norm: Whether to use batch normalization after each hidden layer. """ def _model_fn(features, labels, mode, config): return dnn_lib._dnn_model_fn( # pylint: disable=protected-access @@ -142,7 +157,8 @@ class DNNEstimator(estimator.Estimator): activation_fn=activation_fn, dropout=dropout, input_layer_partitioner=input_layer_partitioner, - config=config) + config=config, + batch_norm=batch_norm) super(DNNEstimator, self).__init__( model_fn=_model_fn, model_dir=model_dir, config=config, warm_start_from=warm_start_from) diff --git a/tensorflow/contrib/estimator/python/estimator/dnn_linear_combined.py b/tensorflow/contrib/estimator/python/estimator/dnn_linear_combined.py index ccaf1128bf23af734f7a5722a4dd8c1f0304fab7..2eef60c39f54bfb464b7da0eb57a47e9eee9b800 100644 --- a/tensorflow/contrib/estimator/python/estimator/dnn_linear_combined.py +++ b/tensorflow/contrib/estimator/python/estimator/dnn_linear_combined.py @@ -53,12 +53,19 @@ class DNNLinearCombinedEstimator(estimator.Estimator): dnn_hidden_units=[1000, 500, 100], dnn_optimizer=tf.train.ProximalAdagradOptimizer(...)) - # To apply L1 and L2 regularization, you can set optimizers as follows: + # To apply L1 and L2 regularization, you can set dnn_optimizer to: tf.train.ProximalAdagradOptimizer( learning_rate=0.1, l1_regularization_strength=0.001, l2_regularization_strength=0.001) - # It is same for FtrlOptimizer. + # To apply learning rate decay, you can set dnn_optimizer to a callable: + lambda: tf.AdamOptimizer( + learning_rate=tf.exponential_decay( + learning_rate=0.1, + global_step=tf.get_global_step(), + decay_steps=10000, + decay_rate=0.96) + # It is the same for linear_optimizer. # Input builders def input_fn_train: # returns x, y @@ -103,7 +110,8 @@ class DNNLinearCombinedEstimator(estimator.Estimator): dnn_activation_fn=nn.relu, dnn_dropout=None, input_layer_partitioner=None, - config=None): + config=None, + linear_sparse_combiner='sum'): """Initializes a DNNLinearCombinedEstimator instance. Args: @@ -116,12 +124,16 @@ class DNNLinearCombinedEstimator(estimator.Estimator): used by linear part of the model. All items in the set must be instances of classes derived from `FeatureColumn`. linear_optimizer: An instance of `tf.Optimizer` used to apply gradients to - the linear part of the model. Defaults to FTRL optimizer. + the linear part of the model. Can also be a string (one of 'Adagrad', + 'Adam', 'Ftrl', 'RMSProp', 'SGD'), or callable. Defaults to FTRL + optimizer. dnn_feature_columns: An iterable containing all the feature columns used by deep part of the model. All items in the set must be instances of classes derived from `FeatureColumn`. dnn_optimizer: An instance of `tf.Optimizer` used to apply gradients to - the deep part of the model. Defaults to Adagrad optimizer. + the deep part of the model. Can also be a string (one of 'Adagrad', + 'Adam', 'Ftrl', 'RMSProp', 'SGD'), or callable. Defaults to Adagrad + optimizer. dnn_hidden_units: List of hidden units per layer. All layers are fully connected. dnn_activation_fn: Activation function applied to each layer. If None, @@ -131,6 +143,11 @@ class DNNLinearCombinedEstimator(estimator.Estimator): input_layer_partitioner: Partitioner for input layer. Defaults to `min_max_variable_partitioner` with `min_slice_size` 64 << 20. config: RunConfig object to configure the runtime settings. + linear_sparse_combiner: A string specifying how to reduce the linear model + if a categorical column is multivalent. One of "mean", "sqrtn", and + "sum" -- these are effectively different ways to do example-level + normalization, which can be useful for bag-of-words features. For more + details, see @{tf.feature_column.linear_model$linear_model}. Raises: ValueError: If both linear_feature_columns and dnn_features_columns are @@ -158,7 +175,8 @@ class DNNLinearCombinedEstimator(estimator.Estimator): dnn_activation_fn=dnn_activation_fn, dnn_dropout=dnn_dropout, input_layer_partitioner=input_layer_partitioner, - config=config) + config=config, + linear_sparse_combiner=linear_sparse_combiner) super(DNNLinearCombinedEstimator, self).__init__( model_fn=_model_fn, model_dir=model_dir, config=config) diff --git a/tensorflow/contrib/estimator/python/estimator/dnn_linear_combined_test.py b/tensorflow/contrib/estimator/python/estimator/dnn_linear_combined_test.py index dd009a6753f3231638f93e50fc8f19eae8820139..51b9ce7005cec3910ba73db62a674e4628ca30a2 100644 --- a/tensorflow/contrib/estimator/python/estimator/dnn_linear_combined_test.py +++ b/tensorflow/contrib/estimator/python/estimator/dnn_linear_combined_test.py @@ -100,7 +100,8 @@ def _linear_only_estimator_fn( weight_column=None, optimizer='Ftrl', config=None, - partitioner=None): + partitioner=None, + sparse_combiner='sum'): return dnn_linear_combined.DNNLinearCombinedEstimator( head=head_lib.regression_head( weight_column=weight_column, label_dimension=label_dimension, @@ -110,7 +111,8 @@ def _linear_only_estimator_fn( linear_feature_columns=feature_columns, linear_optimizer=optimizer, input_layer_partitioner=partitioner, - config=config) + config=config, + linear_sparse_combiner=sparse_combiner) class LinearOnlyEstimatorEvaluateTest( diff --git a/tensorflow/contrib/estimator/python/estimator/early_stopping.py b/tensorflow/contrib/estimator/python/estimator/early_stopping.py new file mode 100644 index 0000000000000000000000000000000000000000..af4855e91e530b4b2815c8039e4d482e8cef485d --- /dev/null +++ b/tensorflow/contrib/estimator/python/estimator/early_stopping.py @@ -0,0 +1,468 @@ +# Copyright 2018 The TensorFlow Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================== +"""Utilities for early stopping.""" + +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + +import operator +import os + +from tensorflow.python.estimator import estimator as estimator_lib +from tensorflow.python.framework import dtypes +from tensorflow.python.framework import ops +from tensorflow.python.ops import init_ops +from tensorflow.python.ops import state_ops +from tensorflow.python.ops import variable_scope +from tensorflow.python.platform import gfile +from tensorflow.python.platform import tf_logging +from tensorflow.python.summary import summary_iterator +from tensorflow.python.training import basic_session_run_hooks +from tensorflow.python.training import session_run_hook +from tensorflow.python.training import training_util + +_EVENT_FILE_GLOB_PATTERN = 'events.out.tfevents.*' + + +def make_early_stopping_hook(estimator, + should_stop_fn, + run_every_secs=60, + run_every_steps=None): + """Creates early-stopping hook. + + Returns a `SessionRunHook` that stops training when `should_stop_fn` returns + `True`. + + Usage example: + + ```python + estimator = ... + hook = early_stopping.make_early_stopping_hook( + estimator, should_stop_fn=make_stop_fn(...)) + train_spec = tf.estimator.TrainSpec(..., hooks=[hook]) + tf.estimator.train_and_evaluate(estimator, train_spec, ...) + ``` + + Args: + estimator: A `tf.estimator.Estimator` instance. + should_stop_fn: `callable`, function that takes no arguments and returns a + `bool`. If the function returns `True`, stopping will be initiated by the + chief. + run_every_secs: If specified, calls `should_stop_fn` at an interval of + `run_every_secs` seconds. Defaults to 60 seconds. Either this or + `run_every_steps` must be set. + run_every_steps: If specified, calls `should_stop_fn` every + `run_every_steps` steps. Either this or `run_every_secs` must be set. + + Returns: + A `SessionRunHook` that periodically executes `should_stop_fn` and initiates + early stopping if the function returns `True`. + + Raises: + TypeError: If `estimator` is not of type `tf.estimator.Estimator`. + ValueError: If both `run_every_secs` and `run_every_steps` are set. + """ + if not isinstance(estimator, estimator_lib.Estimator): + raise TypeError('`estimator` must have type `tf.estimator.Estimator`. ' + 'Got: {}'.format(type(estimator))) + + if run_every_secs is not None and run_every_steps is not None: + raise ValueError('Only one of `run_every_secs` and `run_every_steps` must ' + 'be set.') + + if estimator.config.is_chief: + return _StopOnPredicateHook(should_stop_fn, run_every_secs, run_every_steps) + else: + return _CheckForStoppingHook() + + +def stop_if_higher_hook(estimator, + metric_name, + threshold, + eval_dir=None, + min_steps=0, + run_every_secs=60, + run_every_steps=None): + """Creates hook to stop if the given metric is higher than the threshold. + + Usage example: + + ```python + estimator = ... + # Hook to stop training if accuracy becomes higher than 0.9. + hook = early_stopping.stop_if_higher_hook(estimator, "accuracy", 0.9) + train_spec = tf.estimator.TrainSpec(..., hooks=[hook]) + tf.estimator.train_and_evaluate(estimator, train_spec, ...) + ``` + + Args: + estimator: A `tf.estimator.Estimator` instance. + metric_name: `str`, metric to track. "loss", "accuracy", etc. + threshold: Numeric threshold for the given metric. + eval_dir: If set, directory containing summary files with eval metrics. By + default, `estimator.eval_dir()` will be used. + min_steps: `int`, stop is never requested if global step is less than this + value. Defaults to 0. + run_every_secs: If specified, calls `should_stop_fn` at an interval of + `run_every_secs` seconds. Defaults to 60 seconds. Either this or + `run_every_steps` must be set. + run_every_steps: If specified, calls `should_stop_fn` every + `run_every_steps` steps. Either this or `run_every_secs` must be set. + + Returns: + An early-stopping hook of type `SessionRunHook` that periodically checks + if the given metric is higher than specified threshold and initiates + early stopping if true. + """ + return _stop_if_threshold_crossed_hook( + estimator=estimator, + metric_name=metric_name, + threshold=threshold, + higher_is_better=True, + eval_dir=eval_dir, + min_steps=min_steps, + run_every_secs=run_every_secs, + run_every_steps=run_every_steps) + + +def stop_if_lower_hook(estimator, + metric_name, + threshold, + eval_dir=None, + min_steps=0, + run_every_secs=60, + run_every_steps=None): + """Creates hook to stop if the given metric is lower than the threshold. + + Usage example: + + ```python + estimator = ... + # Hook to stop training if loss becomes lower than 100. + hook = early_stopping.stop_if_lower_hook(estimator, "loss", 100) + train_spec = tf.estimator.TrainSpec(..., hooks=[hook]) + tf.estimator.train_and_evaluate(estimator, train_spec, ...) + ``` + + Args: + estimator: A `tf.estimator.Estimator` instance. + metric_name: `str`, metric to track. "loss", "accuracy", etc. + threshold: Numeric threshold for the given metric. + eval_dir: If set, directory containing summary files with eval metrics. By + default, `estimator.eval_dir()` will be used. + min_steps: `int`, stop is never requested if global step is less than this + value. Defaults to 0. + run_every_secs: If specified, calls `should_stop_fn` at an interval of + `run_every_secs` seconds. Defaults to 60 seconds. Either this or + `run_every_steps` must be set. + run_every_steps: If specified, calls `should_stop_fn` every + `run_every_steps` steps. Either this or `run_every_secs` must be set. + + Returns: + An early-stopping hook of type `SessionRunHook` that periodically checks + if the given metric is lower than specified threshold and initiates + early stopping if true. + """ + return _stop_if_threshold_crossed_hook( + estimator=estimator, + metric_name=metric_name, + threshold=threshold, + higher_is_better=False, + eval_dir=eval_dir, + min_steps=min_steps, + run_every_secs=run_every_secs, + run_every_steps=run_every_steps) + + +def stop_if_no_increase_hook(estimator, + metric_name, + max_steps_without_increase, + eval_dir=None, + min_steps=0, + run_every_secs=60, + run_every_steps=None): + """Creates hook to stop if metric does not increase within given max steps. + + Usage example: + + ```python + estimator = ... + # Hook to stop training if accuracy does not increase in over 100000 steps. + hook = early_stopping.stop_if_no_increase_hook(estimator, "accuracy", 100000) + train_spec = tf.estimator.TrainSpec(..., hooks=[hook]) + tf.estimator.train_and_evaluate(estimator, train_spec, ...) + ``` + + Args: + estimator: A `tf.estimator.Estimator` instance. + metric_name: `str`, metric to track. "loss", "accuracy", etc. + max_steps_without_increase: `int`, maximum number of training steps with no + increase in the given metric. + eval_dir: If set, directory containing summary files with eval metrics. By + default, `estimator.eval_dir()` will be used. + min_steps: `int`, stop is never requested if global step is less than this + value. Defaults to 0. + run_every_secs: If specified, calls `should_stop_fn` at an interval of + `run_every_secs` seconds. Defaults to 60 seconds. Either this or + `run_every_steps` must be set. + run_every_steps: If specified, calls `should_stop_fn` every + `run_every_steps` steps. Either this or `run_every_secs` must be set. + + Returns: + An early-stopping hook of type `SessionRunHook` that periodically checks + if the given metric shows no increase over given maximum number of + training steps, and initiates early stopping if true. + """ + return _stop_if_no_metric_improvement_hook( + estimator=estimator, + metric_name=metric_name, + max_steps_without_improvement=max_steps_without_increase, + higher_is_better=True, + eval_dir=eval_dir, + min_steps=min_steps, + run_every_secs=run_every_secs, + run_every_steps=run_every_steps) + + +def stop_if_no_decrease_hook(estimator, + metric_name, + max_steps_without_decrease, + eval_dir=None, + min_steps=0, + run_every_secs=60, + run_every_steps=None): + """Creates hook to stop if metric does not decrease within given max steps. + + Usage example: + + ```python + estimator = ... + # Hook to stop training if loss does not decrease in over 100000 steps. + hook = early_stopping.stop_if_no_decrease_hook(estimator, "loss", 100000) + train_spec = tf.estimator.TrainSpec(..., hooks=[hook]) + tf.estimator.train_and_evaluate(estimator, train_spec, ...) + ``` + + Args: + estimator: A `tf.estimator.Estimator` instance. + metric_name: `str`, metric to track. "loss", "accuracy", etc. + max_steps_without_decrease: `int`, maximum number of training steps with no + decrease in the given metric. + eval_dir: If set, directory containing summary files with eval metrics. By + default, `estimator.eval_dir()` will be used. + min_steps: `int`, stop is never requested if global step is less than this + value. Defaults to 0. + run_every_secs: If specified, calls `should_stop_fn` at an interval of + `run_every_secs` seconds. Defaults to 60 seconds. Either this or + `run_every_steps` must be set. + run_every_steps: If specified, calls `should_stop_fn` every + `run_every_steps` steps. Either this or `run_every_secs` must be set. + + Returns: + An early-stopping hook of type `SessionRunHook` that periodically checks + if the given metric shows no decrease over given maximum number of + training steps, and initiates early stopping if true. + """ + return _stop_if_no_metric_improvement_hook( + estimator=estimator, + metric_name=metric_name, + max_steps_without_improvement=max_steps_without_decrease, + higher_is_better=False, + eval_dir=eval_dir, + min_steps=min_steps, + run_every_secs=run_every_secs, + run_every_steps=run_every_steps) + + +def read_eval_metrics(eval_dir): + """Helper to read eval metrics from eval summary files. + + Args: + eval_dir: Directory containing summary files with eval metrics. + + Returns: + A `dict` with global steps mapping to `dict` of metric names and values. + """ + eval_metrics_dict = {} + for event in _summaries(eval_dir): + if not event.HasField('summary'): + continue + metrics = {} + for value in event.summary.value: + if value.HasField('simple_value'): + metrics[value.tag] = value.simple_value + if metrics: + eval_metrics_dict[event.step] = metrics + return eval_metrics_dict + + +def _stop_if_threshold_crossed_hook(estimator, metric_name, threshold, + higher_is_better, eval_dir, min_steps, + run_every_secs, run_every_steps): + """Creates early-stopping hook to stop training if threshold is crossed.""" + + if eval_dir is None: + eval_dir = estimator.eval_dir() + + is_lhs_better = operator.gt if higher_is_better else operator.lt + greater_or_lesser = 'greater than' if higher_is_better else 'less than' + + def stop_if_threshold_crossed_fn(): + """Returns `True` if the given metric crosses specified threshold.""" + + eval_results = read_eval_metrics(eval_dir) + + for step, metrics in eval_results.items(): + if step < min_steps: + continue + val = metrics[metric_name] + if is_lhs_better(val, threshold): + tf_logging.info( + 'At step %s, metric "%s" has value %s which is %s the configured ' + 'threshold (%s) for early stopping.', step, metric_name, val, + greater_or_lesser, threshold) + return True + return False + + return make_early_stopping_hook( + estimator=estimator, + should_stop_fn=stop_if_threshold_crossed_fn, + run_every_secs=run_every_secs, + run_every_steps=run_every_steps) + + +def _stop_if_no_metric_improvement_hook( + estimator, metric_name, max_steps_without_improvement, higher_is_better, + eval_dir, min_steps, run_every_secs, run_every_steps): + """Returns hook to stop training if given metric shows no improvement.""" + + if eval_dir is None: + eval_dir = estimator.eval_dir() + + is_lhs_better = operator.gt if higher_is_better else operator.lt + increase_or_decrease = 'increase' if higher_is_better else 'decrease' + + def stop_if_no_metric_improvement_fn(): + """Returns `True` if metric does not improve within max steps.""" + + eval_results = read_eval_metrics(eval_dir) + + best_val = None + best_val_step = None + for step, metrics in eval_results.items(): + if step < min_steps: + continue + val = metrics[metric_name] + if best_val is None or is_lhs_better(val, best_val): + best_val = val + best_val_step = step + if step - best_val_step >= max_steps_without_improvement: + tf_logging.info( + 'No %s in metric "%s" for %s steps, which is greater than or equal ' + 'to max steps (%s) configured for early stopping.', + increase_or_decrease, metric_name, step - best_val_step, + max_steps_without_improvement) + return True + return False + + return make_early_stopping_hook( + estimator=estimator, + should_stop_fn=stop_if_no_metric_improvement_fn, + run_every_secs=run_every_secs, + run_every_steps=run_every_steps) + + +def _summaries(eval_dir): + """Yields `tensorflow.Event` protos from event files in the eval dir. + + Args: + eval_dir: Directory containing summary files with eval metrics. + + Yields: + `tensorflow.Event` object read from the event files. + """ + for event_file in gfile.Glob( + os.path.join(eval_dir, _EVENT_FILE_GLOB_PATTERN)): + for event in summary_iterator.summary_iterator(event_file): + yield event + + +def _get_or_create_stop_var(): + with variable_scope.variable_scope( + name_or_scope='signal_early_stopping', + values=[], + reuse=variable_scope.AUTO_REUSE): + return variable_scope.get_variable( + name='STOP', + shape=[], + dtype=dtypes.bool, + initializer=init_ops.constant_initializer(False), + collections=[ops.GraphKeys.GLOBAL_VARIABLES], + trainable=False) + + +class _StopOnPredicateHook(session_run_hook.SessionRunHook): + """Hook that requests stop when `should_stop_fn` returns `True`.""" + + def __init__(self, should_stop_fn, run_every_secs=60, run_every_steps=None): + if not callable(should_stop_fn): + raise TypeError('`should_stop_fn` must be callable.') + + self._should_stop_fn = should_stop_fn + self._timer = basic_session_run_hooks.SecondOrStepTimer( + every_secs=run_every_secs, every_steps=run_every_steps) + self._global_step_tensor = None + self._stop_var = None + self._stop_op = None + + def begin(self): + self._global_step_tensor = training_util.get_global_step() + self._stop_var = _get_or_create_stop_var() + self._stop_op = state_ops.assign(self._stop_var, True) + + def before_run(self, run_context): + del run_context + return session_run_hook.SessionRunArgs(self._global_step_tensor) + + def after_run(self, run_context, run_values): + global_step = run_values.results + if self._timer.should_trigger_for_step(global_step): + self._timer.update_last_triggered_step(global_step) + if self._should_stop_fn(): + tf_logging.info('Requesting early stopping at global step %d', + global_step) + run_context.session.run(self._stop_op) + run_context.request_stop() + + +class _CheckForStoppingHook(session_run_hook.SessionRunHook): + """Hook that requests stop if stop is requested by `_StopOnPredicateHook`.""" + + def __init__(self): + self._stop_var = None + + def begin(self): + self._stop_var = _get_or_create_stop_var() + + def before_run(self, run_context): + del run_context + return session_run_hook.SessionRunArgs(self._stop_var) + + def after_run(self, run_context, run_values): + should_early_stop = run_values.results + if should_early_stop: + tf_logging.info('Early stopping requested, suspending run.') + run_context.request_stop() diff --git a/tensorflow/contrib/estimator/python/estimator/early_stopping_test.py b/tensorflow/contrib/estimator/python/estimator/early_stopping_test.py new file mode 100644 index 0000000000000000000000000000000000000000..b5eee818fa0bce068597a7fb8d99dd86f43d396a --- /dev/null +++ b/tensorflow/contrib/estimator/python/estimator/early_stopping_test.py @@ -0,0 +1,233 @@ +# Copyright 2018 The TensorFlow Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================== +"""Tests for early_stopping.""" + +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + +import os +import tempfile + +from absl.testing import parameterized +from tensorflow.contrib.estimator.python.estimator import early_stopping +from tensorflow.python.estimator import estimator +from tensorflow.python.estimator import run_config +from tensorflow.python.framework import ops +from tensorflow.python.ops import control_flow_ops +from tensorflow.python.ops import state_ops +from tensorflow.python.platform import test +from tensorflow.python.training import monitored_session +from tensorflow.python.training import training_util + + +class _FakeRunConfig(run_config.RunConfig): + + def __init__(self, is_chief): + super(_FakeRunConfig, self).__init__() + self._is_chief = is_chief + + @property + def is_chief(self): + return self._is_chief + + +def _dummy_model_fn(features, labels, params): + _, _, _ = features, labels, params + + +class _FakeEstimator(estimator.Estimator): + """Fake estimator for testing.""" + + def __init__(self, config): + super(_FakeEstimator, self).__init__( + model_fn=_dummy_model_fn, config=config) + + +def _write_events(eval_dir, params): + """Test helper to write events to summary files.""" + for steps, loss, accuracy in params: + estimator._write_dict_to_summary(eval_dir, { + 'loss': loss, + 'accuracy': accuracy, + }, steps) + + +class ReadEvalMetricsTest(test.TestCase): + + def test_read_eval_metrics(self): + eval_dir = tempfile.mkdtemp() + _write_events( + eval_dir, + [ + # steps, loss, accuracy + (1000, 1, 2), + (2000, 3, 4), + (3000, 5, 6), + ]) + self.assertEqual({ + 1000: { + 'loss': 1, + 'accuracy': 2 + }, + 2000: { + 'loss': 3, + 'accuracy': 4 + }, + 3000: { + 'loss': 5, + 'accuracy': 6 + }, + }, early_stopping.read_eval_metrics(eval_dir)) + + +class EarlyStoppingHooksTest(test.TestCase, parameterized.TestCase): + + def setUp(self): + config = _FakeRunConfig(is_chief=True) + self._estimator = _FakeEstimator(config=config) + eval_dir = self._estimator.eval_dir() + os.makedirs(eval_dir) + _write_events( + eval_dir, + [ + # steps, loss, accuracy + (1000, 0.8, 0.5), + (2000, 0.7, 0.6), + (3000, 0.4, 0.7), + (3500, 0.41, 0.68), + ]) + + def run_session(self, hooks, should_stop): + hooks = hooks if isinstance(hooks, list) else [hooks] + with ops.Graph().as_default(): + training_util.create_global_step() + no_op = control_flow_ops.no_op() + with monitored_session.SingularMonitoredSession(hooks=hooks) as mon_sess: + mon_sess.run(no_op) + self.assertEqual(mon_sess.should_stop(), should_stop) + + @parameterized.parameters((0.8, 0, False), (0.6, 4000, False), (0.6, 0, True)) + def test_stop_if_higher_hook(self, threshold, min_steps, should_stop): + self.run_session( + early_stopping.stop_if_higher_hook( + self._estimator, + metric_name='accuracy', + threshold=threshold, + min_steps=min_steps), should_stop) + + @parameterized.parameters((0.3, 0, False), (0.5, 4000, False), (0.5, 0, True)) + def test_stop_if_lower_hook(self, threshold, min_steps, should_stop): + self.run_session( + early_stopping.stop_if_lower_hook( + self._estimator, + metric_name='loss', + threshold=threshold, + min_steps=min_steps), should_stop) + + @parameterized.parameters((1500, 0, False), (500, 4000, False), + (500, 0, True)) + def test_stop_if_no_increase_hook(self, max_steps, min_steps, should_stop): + self.run_session( + early_stopping.stop_if_no_increase_hook( + self._estimator, + metric_name='accuracy', + max_steps_without_increase=max_steps, + min_steps=min_steps), should_stop) + + @parameterized.parameters((1500, 0, False), (500, 4000, False), + (500, 0, True)) + def test_stop_if_no_decrease_hook(self, max_steps, min_steps, should_stop): + self.run_session( + early_stopping.stop_if_no_decrease_hook( + self._estimator, + metric_name='loss', + max_steps_without_decrease=max_steps, + min_steps=min_steps), should_stop) + + @parameterized.parameters((1500, 0.3, False), (1500, 0.5, True), + (500, 0.3, True)) + def test_multiple_hooks(self, max_steps, loss_threshold, should_stop): + self.run_session([ + early_stopping.stop_if_no_decrease_hook( + self._estimator, + metric_name='loss', + max_steps_without_decrease=max_steps), + early_stopping.stop_if_lower_hook( + self._estimator, metric_name='loss', threshold=loss_threshold) + ], should_stop) + + @parameterized.parameters(False, True) + def test_make_early_stopping_hook(self, should_stop): + self.run_session([ + early_stopping.make_early_stopping_hook( + self._estimator, should_stop_fn=lambda: should_stop) + ], should_stop) + + def test_make_early_stopping_hook_typeerror(self): + with self.assertRaises(TypeError): + early_stopping.make_early_stopping_hook( + estimator=object(), should_stop_fn=lambda: True) + + def test_make_early_stopping_hook_valueerror(self): + with self.assertRaises(ValueError): + early_stopping.make_early_stopping_hook( + self._estimator, + should_stop_fn=lambda: True, + run_every_secs=60, + run_every_steps=100) + + +class StopOnPredicateHookTest(test.TestCase): + + def test_stop(self): + hook = early_stopping._StopOnPredicateHook( + should_stop_fn=lambda: False, run_every_secs=0) + with ops.Graph().as_default(): + training_util.create_global_step() + no_op = control_flow_ops.no_op() + with monitored_session.SingularMonitoredSession(hooks=[hook]) as mon_sess: + mon_sess.run(no_op) + self.assertFalse(mon_sess.should_stop()) + self.assertFalse(mon_sess.raw_session().run(hook._stop_var)) + + hook = early_stopping._StopOnPredicateHook( + should_stop_fn=lambda: True, run_every_secs=0) + with ops.Graph().as_default(): + training_util.create_global_step() + no_op = control_flow_ops.no_op() + with monitored_session.SingularMonitoredSession(hooks=[hook]) as mon_sess: + mon_sess.run(no_op) + self.assertTrue(mon_sess.should_stop()) + self.assertTrue(mon_sess.raw_session().run(hook._stop_var)) + + +class CheckForStoppingHookTest(test.TestCase): + + def test_stop(self): + hook = early_stopping._CheckForStoppingHook() + with ops.Graph().as_default(): + no_op = control_flow_ops.no_op() + assign_op = state_ops.assign(early_stopping._get_or_create_stop_var(), + True) + with monitored_session.SingularMonitoredSession(hooks=[hook]) as mon_sess: + mon_sess.run(no_op) + self.assertFalse(mon_sess.should_stop()) + mon_sess.run(assign_op) + self.assertTrue(mon_sess.should_stop()) + + +if __name__ == '__main__': + test.main() diff --git a/tensorflow/contrib/estimator/python/estimator/head.py b/tensorflow/contrib/estimator/python/estimator/head.py index 9594e5132fd20dadea118fd1dd6768feb7fd7fff..c9d86ef4ab89950b0c7b0414ba60d9e0a1cbe476 100644 --- a/tensorflow/contrib/estimator/python/estimator/head.py +++ b/tensorflow/contrib/estimator/python/estimator/head.py @@ -534,7 +534,8 @@ def multi_label_head(n_classes, * An integer `SparseTensor` of class indices. The `dense_shape` must be `[D0, D1, ... DN, ?]` and the values within `[0, n_classes)`. * If `label_vocabulary` is given, a string `SparseTensor`. The `dense_shape` - must be `[D0, D1, ... DN, ?]` and the values within `label_vocabulary`. + must be `[D0, D1, ... DN, ?]` and the values within `label_vocabulary` or a + multi-hot tensor of shape `[D0, D1, ... DN, n_classes]`. If `weight_column` is specified, weights must be of shape `[D0, D1, ... DN]`, or `[D0, D1, ... DN, 1]`. diff --git a/tensorflow/contrib/estimator/python/estimator/head_test.py b/tensorflow/contrib/estimator/python/estimator/head_test.py index b2b57fa06ba818d4455871fe57dde5ce287b39a2..7b884402d4650636bc9fe053994246aabb9c312d 100644 --- a/tensorflow/contrib/estimator/python/estimator/head_test.py +++ b/tensorflow/contrib/estimator/python/estimator/head_test.py @@ -568,6 +568,33 @@ class MultiLabelHead(test.TestCase): expected_loss=expected_loss, expected_metrics=expected_metrics) + def test_eval_with_label_vocabulary_with_multi_hot_input(self): + n_classes = 2 + head = head_lib.multi_label_head( + n_classes, label_vocabulary=['class0', 'class1']) + logits = np.array([[-1., 1.], [-1.5, 1.5]], dtype=np.float32) + labels_multi_hot = np.array([[1, 0], [1, 1]], dtype=np.int64) + # loss = labels * -log(sigmoid(logits)) + + # (1 - labels) * -log(1 - sigmoid(logits)) + # Sum over examples, divide by batch_size. + expected_loss = 0.5 * np.sum( + _sigmoid_cross_entropy(labels=labels_multi_hot, logits=logits)) + keys = metric_keys.MetricKeys + expected_metrics = { + # Average loss over examples. + keys.LOSS_MEAN: expected_loss, + # auc and auc_pr cannot be reliably calculated for only 4 samples, but + # this assert tests that the algorithm remains consistent. + keys.AUC: 0.3333, + keys.AUC_PR: 0.7639, + } + self._test_eval( + head=head, + logits=logits, + labels=labels_multi_hot, + expected_loss=expected_loss, + expected_metrics=expected_metrics) + def test_eval_with_thresholds(self): n_classes = 2 thresholds = [0.25, 0.5, 0.75] diff --git a/tensorflow/contrib/estimator/python/estimator/linear.py b/tensorflow/contrib/estimator/python/estimator/linear.py index 3bf4abe83d54504d55de73b63f369cceaf149dd2..62a37abefb1f6ed291df1df3da6de35bfd2b6c52 100644 --- a/tensorflow/contrib/estimator/python/estimator/linear.py +++ b/tensorflow/contrib/estimator/python/estimator/linear.py @@ -39,6 +39,18 @@ class LinearEstimator(estimator.Estimator): feature_columns=[categorical_column_a, categorical_feature_a_x_categorical_feature_b]) + # Or estimator using an optimizer with a learning rate decay. + estimator = LinearEstimator( + head=tf.contrib.estimator.multi_label_head(n_classes=3), + feature_columns=[categorical_column_a, + categorical_feature_a_x_categorical_feature_b], + optimizer=lambda: tf.train.FtrlOptimizer( + learning_rate=tf.exponential_decay( + learning_rate=0.1, + global_step=tf.get_global_step(), + decay_steps=10000, + decay_rate=0.96)) + # Or estimator using the FTRL optimizer with regularization. estimator = LinearEstimator( head=tf.contrib.estimator.multi_label_head(n_classes=3), @@ -87,7 +99,8 @@ class LinearEstimator(estimator.Estimator): model_dir=None, optimizer='Ftrl', config=None, - partitioner=None): + partitioner=None, + sparse_combiner='sum'): """Initializes a `LinearEstimator` instance. Args: @@ -99,10 +112,16 @@ class LinearEstimator(estimator.Estimator): model_dir: Directory to save model parameters, graph and etc. This can also be used to load checkpoints from the directory into a estimator to continue training a previously saved model. - optimizer: An instance of `tf.Optimizer` used to train the model. Defaults - to FTRL optimizer. + optimizer: An instance of `tf.Optimizer` used to train the model. Can also + be a string (one of 'Adagrad', 'Adam', 'Ftrl', 'RMSProp', 'SGD'), or + callable. Defaults to FTRL optimizer. config: `RunConfig` object to configure the runtime settings. partitioner: Optional. Partitioner for input layer. + sparse_combiner: A string specifying how to reduce if a categorical column + is multivalent. One of "mean", "sqrtn", and "sum" -- these are + effectively different ways to do example-level normalization, which can + be useful for bag-of-words features. for more details, see + @{tf.feature_column.linear_model$linear_model}. """ def _model_fn(features, labels, mode, config): return linear_lib._linear_model_fn( # pylint: disable=protected-access @@ -113,6 +132,7 @@ class LinearEstimator(estimator.Estimator): feature_columns=tuple(feature_columns or []), optimizer=optimizer, partitioner=partitioner, - config=config) + config=config, + sparse_combiner=sparse_combiner) super(LinearEstimator, self).__init__( model_fn=_model_fn, model_dir=model_dir, config=config) diff --git a/tensorflow/contrib/factorization/kernels/wals_solver_ops.cc b/tensorflow/contrib/factorization/kernels/wals_solver_ops.cc index bb9b835889b1b5e36d6f470b51834d4c6bb3d493..7fcae5ad8e1536530e2d039e1d14df4e192c4fa3 100644 --- a/tensorflow/contrib/factorization/kernels/wals_solver_ops.cc +++ b/tensorflow/contrib/factorization/kernels/wals_solver_ops.cc @@ -62,10 +62,11 @@ class WALSComputePartialLhsAndRhsOp : public OpKernel { public: explicit WALSComputePartialLhsAndRhsOp(OpKernelConstruction* context) : OpKernel(context) { - OP_REQUIRES_OK(context, context->MatchSignature( - {DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_FLOAT, - DT_INT64, DT_FLOAT, DT_INT64, DT_BOOL}, - {DT_FLOAT, DT_FLOAT})); + OP_REQUIRES_OK(context, + context->MatchSignature( + {DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_INT64, + DT_FLOAT, DT_FLOAT, DT_INT64, DT_BOOL}, + {DT_FLOAT, DT_FLOAT})); } void Compute(OpKernelContext* context) override { @@ -75,8 +76,9 @@ class WALSComputePartialLhsAndRhsOp : public OpKernel { const Tensor& input_weights = context->input(3); const Tensor& input_indices = context->input(4); const Tensor& input_values = context->input(5); - const Tensor& input_block_size = context->input(6); - const Tensor& input_is_transpose = context->input(7); + const Tensor& entry_weights = context->input(6); + const Tensor& input_block_size = context->input(7); + const Tensor& input_is_transpose = context->input(8); OP_REQUIRES(context, TensorShapeUtils::IsMatrix(factors.shape()), InvalidArgument("Input factors should be a matrix.")); @@ -89,13 +91,33 @@ class WALSComputePartialLhsAndRhsOp : public OpKernel { InvalidArgument("Input input_weights should be a vector.")); OP_REQUIRES(context, TensorShapeUtils::IsMatrix(input_indices.shape()), InvalidArgument("Input input_indices should be a matrix.")); + OP_REQUIRES( + context, input_indices.dim_size(1) == 2, + InvalidArgument("Input input_indices should have shape (?, 2).")); OP_REQUIRES(context, TensorShapeUtils::IsVector(input_values.shape()), InvalidArgument("Input input_values should be a vector")); + OP_REQUIRES(context, TensorShapeUtils::IsVector(entry_weights.shape()), + InvalidArgument("Input entry_weights should be a vector")); + OP_REQUIRES(context, input_indices.dim_size(0) == input_values.dim_size(0), + InvalidArgument("Input input_values' length should match the " + "first dimension of Input input_indices ")); OP_REQUIRES(context, TensorShapeUtils::IsScalar(input_block_size.shape()), InvalidArgument("Input input_block_size should be a scalar.")); OP_REQUIRES( context, TensorShapeUtils::IsScalar(input_is_transpose.shape()), InvalidArgument("Input input_is_transpose should be a scalar.")); + OP_REQUIRES( + context, + ((input_weights.dim_size(0) > 0 && + factor_weights.dim_size(0) == factors.dim_size(0) && + entry_weights.dim_size(0) == 0) || + (input_weights.dim_size(0) == 0 && factor_weights.dim_size(0) == 0 && + entry_weights.dim_size(0) == input_indices.dim_size(0))), + InvalidArgument("To specify the weights for observed entries, either " + "(1) entry_weights must be set or (2) input_weights " + "and factor_weights must be set, but not both.")); + // TODO(yifanchen): Deprecate the support of input_weights and + // factor_weights. const int64 factor_dim = factors.dim_size(1); const int64 factors_size = factors.dim_size(0); @@ -105,6 +127,7 @@ class WALSComputePartialLhsAndRhsOp : public OpKernel { const auto& input_weights_vec = input_weights.vec(); const float w_0 = unobserved_weights.scalar()(); const auto& input_values_vec = input_values.vec(); + const auto& entry_weights_vec = entry_weights.vec(); ConstEigenMatrixFloatMap factors_mat(factors.matrix().data(), factor_dim, factors_size); @@ -134,6 +157,8 @@ class WALSComputePartialLhsAndRhsOp : public OpKernel { return is_transpose ? indices_mat(0, i) : indices_mat(1, i); }; + const bool use_entry_weights = entry_weights_vec.size() > 0; + // TODO(rmlarsen): In principle, we should be using the SparseTensor class // and machinery for iterating over groups, but the fact that class // SparseTensor makes a complete copy of the matrix makes me reluctant to @@ -195,6 +220,8 @@ class WALSComputePartialLhsAndRhsOp : public OpKernel { // map using the hash of the thread id as the key. // // TODO(jpoulson): Switch to try_emplace once C++17 is supported + // TODO(b/72952120): Check whether the 3 lock-unlock pairs can be + // consolidated into just one. map_mutex.lock(); const auto key_count = factor_batch_map.count(id_hash); map_mutex.unlock(); @@ -213,6 +240,8 @@ class WALSComputePartialLhsAndRhsOp : public OpKernel { CHECK_LE(shard.second, perm.size()); CHECK_LE(shard.first, shard.second); const int64 input_index = get_input_index(perm[shard.first]); + const float input_weight = + use_entry_weights ? 1.0 : input_weights_vec(input_index); // Accumulate the rhs and lhs terms in the normal equations // for the non-zero elements in the row or column of the sparse matrix // corresponding to input_index. @@ -228,7 +257,8 @@ class WALSComputePartialLhsAndRhsOp : public OpKernel { const int64 factor_index = get_factor_index(i); const float input_value = input_values_vec(i); const float weight = - input_weights_vec(input_index) * factor_weights_vec(factor_index); + use_entry_weights ? entry_weights_vec(i) + : input_weight * factor_weights_vec(factor_index); CHECK_GE(weight, 0); factor_batch.col(num_batched) = factors_mat.col(factor_index) * std::sqrt(weight); diff --git a/tensorflow/contrib/factorization/ops/factorization_ops.cc b/tensorflow/contrib/factorization/ops/factorization_ops.cc index 11ea36946e92769cd6901eb998a20148250ef7ce..1d31bd38c824f24e9a70c0f69da129f5ddc18985 100644 --- a/tensorflow/contrib/factorization/ops/factorization_ops.cc +++ b/tensorflow/contrib/factorization/ops/factorization_ops.cc @@ -25,20 +25,33 @@ REGISTER_OP("WALSComputePartialLhsAndRhs") .Input("input_weights: float32") .Input("input_indices: int64") .Input("input_values: float32") + .Input("entry_weights: float32") .Input("input_block_size: int64") .Input("input_is_transpose: bool") .Output("partial_lhs: float32") .Output("partial_rhs: float32") .SetShapeFn(shape_inference::UnknownShape) .Doc(R"( -Computes the partial left-hand side and right-hand side of WALS update. +Computes the partial left-hand side and right-hand side of WALS update. For +observed entry input_indices[i]=[m, n] with value input_values[i]=v, the weight +should be specified either through (1) entry_weights[i] or (2) through +input_weights[m] * factor_weights[n] (if input_is_transpose is false) or +input_weights[n] * factor_weights[m] (if input_is_transpose is true). Note it is +not allowed to have both (1) and (2) specified at the same time: when one +approach is used, the input tensors related to the other approach must be kept +completely empty. factors: Matrix of size m * k. -factor_weights: Vector of size m. Corresponds to column weights +factor_weights: Vector of size m. Corresponds to column weights. Should be empty + if entry_weights is used. unobserved_weights: Scalar. Weight for unobserved input entries. -input_weights: Vector of size n. Corresponds to row weights. +input_weights: Vector of size n. Corresponds to row weights. Should be empty if + entry_weights is used. input_indices: Indices for the input SparseTensor. input_values: Values for the input SparseTensor. +entry_weights: If not empty, this must be same length as input_vaues and is used + as the per-entry non-zero weight. If this is used, input_weights and + factor_weights must be empty. input_block_size: Scalar. Number of rows spanned by input. input_is_transpose: If true, logically transposes the input for processing. partial_lhs: 3-D tensor with size input_block_size x k x k. diff --git a/tensorflow/contrib/factorization/python/kernel_tests/wals_solver_ops_test.py b/tensorflow/contrib/factorization/python/kernel_tests/wals_solver_ops_test.py index ba30fd997700f461b6afffa13cf371c598d3332e..6c2f1d46084d701beac1e3a99e3ad66bae57eda5 100644 --- a/tensorflow/contrib/factorization/python/kernel_tests/wals_solver_ops_test.py +++ b/tensorflow/contrib/factorization/python/kernel_tests/wals_solver_ops_test.py @@ -55,7 +55,41 @@ class WalsSolverOpsTest(test.TestCase): rhs_matrix] = gen_factorization_ops.wals_compute_partial_lhs_and_rhs( self._column_factors, self._column_weights, self._unobserved_weights, self._row_weights, sparse_block.indices, sparse_block.values, - sparse_block.dense_shape[0], False) + [], + input_block_size=sparse_block.dense_shape[0], + input_is_transpose=False) + self.assertAllClose(lhs_tensor.eval(), [[ + [0.014800, 0.017000, 0.019200], + [0.017000, 0.019600, 0.022200], + [0.019200, 0.022200, 0.025200], + ], [ + [0.0064000, 0.0080000, 0.0096000], + [0.0080000, 0.0100000, 0.0120000], + [0.0096000, 0.0120000, 0.0144000], + ], [ + [0.0099000, 0.0126000, 0.0153000], + [0.0126000, 0.0162000, 0.0198000], + [0.0153000, 0.0198000, 0.0243000], + ], [ + [0.058800, 0.067200, 0.075600], + [0.067200, 0.076800, 0.086400], + [0.075600, 0.086400, 0.097200], + ]]) + self.assertAllClose(rhs_matrix.eval(), [[0.019300, 0.023000, 0.026700], + [0.061600, 0.077000, 0.092400], + [0.160400, 0.220000, 0.279600], + [0.492800, 0.563200, 0.633600]]) + + def testWalsSolverLhsEntryWeights(self): + sparse_block = SparseBlock3x3() + with self.test_session(): + [lhs_tensor, + rhs_matrix] = gen_factorization_ops.wals_compute_partial_lhs_and_rhs( + self._column_factors, [], self._unobserved_weights, + [], sparse_block.indices, sparse_block.values, + [0.01, 0.03, 0.04, 0.03, 0.06, 0.12], + input_block_size=sparse_block.dense_shape[0], + input_is_transpose=False) self.assertAllClose(lhs_tensor.eval(), [[ [0.014800, 0.017000, 0.019200], [0.017000, 0.019600, 0.022200], diff --git a/tensorflow/contrib/factorization/python/ops/factorization_ops.py b/tensorflow/contrib/factorization/python/ops/factorization_ops.py index 8f73274c2a0ebbdc41ce6a647a8a5650694c9a23..7ab70fbcfd7324961b61526a08daab7e393630e9 100644 --- a/tensorflow/contrib/factorization/python/ops/factorization_ops.py +++ b/tensorflow/contrib/factorization/python/ops/factorization_ops.py @@ -943,6 +943,7 @@ class WALSModel(object): row_weights_slice, new_sp_input.indices, new_sp_input.values, + [], num_rows, transpose_input, name="wals_compute_partial_lhs_rhs")) diff --git a/tensorflow/contrib/gan/BUILD b/tensorflow/contrib/gan/BUILD index b305f37791d71f5a6edeada2bb710a2e5f23087d..10a8796bcb7a813241cc472bd017e1399ea6dc7b 100644 --- a/tensorflow/contrib/gan/BUILD +++ b/tensorflow/contrib/gan/BUILD @@ -45,6 +45,7 @@ py_library( "//tensorflow/python:framework_ops", "//tensorflow/python:init_ops", "//tensorflow/python:training", + "//tensorflow/python:training_util", "//tensorflow/python:variable_scope", "//tensorflow/python/ops/distributions", "//tensorflow/python/ops/losses", @@ -59,6 +60,7 @@ py_test( deps = [ ":features", ":namedtuples", + ":random_tensor_pool", ":train", "//tensorflow/contrib/framework:framework_py", "//tensorflow/contrib/slim:learning", @@ -70,6 +72,7 @@ py_test( "//tensorflow/python:random_ops", "//tensorflow/python:random_seed", "//tensorflow/python:training", + "//tensorflow/python:training_util", "//tensorflow/python:variable_scope", "//tensorflow/python:variables", "//tensorflow/python/ops/distributions", @@ -188,6 +191,7 @@ py_test( srcs = ["python/losses/python/tuple_losses_test.py"], srcs_version = "PY2AND3", deps = [ + ":namedtuples", ":tuple_losses", "//tensorflow/python:client_testlib", "//tensorflow/python:constant_op", @@ -344,9 +348,11 @@ py_library( "//tensorflow/python:image_ops", "//tensorflow/python:linalg_ops", "//tensorflow/python:math_ops", + "//tensorflow/python:nn", "//tensorflow/python:nn_ops", "//tensorflow/python:platform", "//tensorflow/python:util", + "@six_archive//:six", ], ) @@ -470,12 +476,12 @@ py_library( ], srcs_version = "PY2AND3", deps = [ - ":head", ":namedtuples", ":summaries", ":train", "//tensorflow/contrib/framework:framework_py", "//tensorflow/python:framework_ops", + "//tensorflow/python:metrics", "//tensorflow/python:util", "//tensorflow/python:variable_scope", "//tensorflow/python/estimator", @@ -498,16 +504,19 @@ py_test( "//tensorflow/core:protos_all_py", "//tensorflow/python:array_ops", "//tensorflow/python:client_testlib", - "//tensorflow/python:control_flow_ops", "//tensorflow/python:dtypes", "//tensorflow/python:framework_ops", + "//tensorflow/python:math_ops", + "//tensorflow/python:metrics", "//tensorflow/python:parsing_ops", "//tensorflow/python:summary", "//tensorflow/python:training", - "//tensorflow/python/estimator:head", + "//tensorflow/python:training_util", + "//tensorflow/python:variable_scope", "//tensorflow/python/estimator:model_fn", "//tensorflow/python/estimator:numpy_io", "//third_party/py/numpy", + "@absl_py//absl/testing:parameterized", "@six_archive//:six", ], ) diff --git a/tensorflow/contrib/gan/python/estimator/python/gan_estimator_impl.py b/tensorflow/contrib/gan/python/estimator/python/gan_estimator_impl.py index 4092b320042162e4eb4c5f4879c2c3ea5dc14fc9..8e4affb9b4f95bf5afab0f50c86954e60a942279 100644 --- a/tensorflow/contrib/gan/python/estimator/python/gan_estimator_impl.py +++ b/tensorflow/contrib/gan/python/estimator/python/gan_estimator_impl.py @@ -24,11 +24,11 @@ import enum from tensorflow.contrib.framework.python.ops import variables as variable_lib from tensorflow.contrib.gan.python import namedtuples as tfgan_tuples from tensorflow.contrib.gan.python import train as tfgan_train -from tensorflow.contrib.gan.python.estimator.python import head as head_lib from tensorflow.contrib.gan.python.eval.python import summaries as tfgan_summaries from tensorflow.python.estimator import estimator from tensorflow.python.estimator import model_fn as model_fn_lib from tensorflow.python.framework import ops +from tensorflow.python.ops import metrics as metrics_lib from tensorflow.python.ops import variable_scope from tensorflow.python.util import tf_inspect as inspect @@ -154,94 +154,93 @@ class GANEstimator(estimator.Estimator): use_loss_summaries: If `True`, add loss summaries. If `False`, does not. If `None`, uses defaults. config: `RunConfig` object to configure the runtime settings. + + Raises: + ValueError: If loss functions aren't callable. + ValueError: If `use_loss_summaries` isn't boolean or `None`. + ValueError: If `get_hooks_fn` isn't callable or `None`. """ - # TODO(joelshor): Explicitly validate inputs. + if not callable(generator_loss_fn): + raise ValueError('generator_loss_fn must be callable.') + if not callable(discriminator_loss_fn): + raise ValueError('discriminator_loss_fn must be callable.') + if use_loss_summaries not in [True, False, None]: + raise ValueError('use_loss_summaries must be True, False or None.') + if get_hooks_fn is not None and not callable(get_hooks_fn): + raise TypeError('get_hooks_fn must be callable.') def _model_fn(features, labels, mode): - gopt = (generator_optimizer() if callable(generator_optimizer) else - generator_optimizer) - dopt = (discriminator_optimizer() if callable(discriminator_optimizer) - else discriminator_optimizer) - gan_head = head_lib.gan_head( - generator_loss_fn, discriminator_loss_fn, gopt, dopt, - use_loss_summaries, get_hooks_fn=get_hooks_fn, - get_eval_metric_ops_fn=get_eval_metric_ops_fn) - return _gan_model_fn( - features, labels, mode, generator_fn, discriminator_fn, gan_head, + """GANEstimator model function.""" + if mode not in [model_fn_lib.ModeKeys.TRAIN, model_fn_lib.ModeKeys.EVAL, + model_fn_lib.ModeKeys.PREDICT]: + raise ValueError('Mode not recognized: %s' % mode) + real_data = labels # rename inputs for clarity + generator_inputs = features # rename inputs for clarity + + # Make GANModel, which encapsulates the GAN model architectures. + gan_model = _get_gan_model( + mode, generator_fn, discriminator_fn, real_data, generator_inputs, add_summaries) + # Make the EstimatorSpec, which incorporates the GANModel, losses, eval + # metrics, and optimizers (if required). + return _get_estimator_spec( + mode, gan_model, generator_loss_fn, discriminator_loss_fn, + get_eval_metric_ops_fn, generator_optimizer, discriminator_optimizer, + get_hooks_fn) + super(GANEstimator, self).__init__( model_fn=_model_fn, model_dir=model_dir, config=config) -def _gan_model_fn( - features, - labels, - mode, - generator_fn, - discriminator_fn, - head, - add_summaries=None, - generator_scope_name='Generator'): - """The `model_fn` for the GAN estimator. - - We make the following convention: - features -> TFGAN's `generator_inputs` - labels -> TFGAN's `real_data` - - Args: - features: A dictionary to feed to generator. In the unconditional case, - this might be just `noise`. In the conditional GAN case, this - might be the generator's conditioning. The `generator_fn` determines - what the required keys are. - labels: Real data. Can be any structure, as long as `discriminator_fn` - can accept it for the first argument. - mode: Defines whether this is training, evaluation or prediction. - See `ModeKeys`. - generator_fn: A python lambda that takes `generator_inputs` as inputs and - returns the outputs of the GAN generator. - discriminator_fn: A python lambda that takes `real_data`/`generated data` - and `generator_inputs`. Outputs a Tensor in the range [-inf, inf]. - head: A `Head` instance suitable for GANs. - add_summaries: `None`, a single `SummaryType`, or a list of `SummaryType`. - generator_scope_name: The name of the generator scope. We need this to be - the same for GANModels produced by TFGAN's `train.gan_model` and the - manually constructed ones for predictions. - - Returns: - `ModelFnOps` - - Raises: - ValueError: If `labels` isn't `None` during prediction. - """ - real_data = labels - generator_inputs = features - - if mode == model_fn_lib.ModeKeys.TRAIN: - gan_model = _make_train_gan_model( - generator_fn, discriminator_fn, real_data, generator_inputs, - generator_scope_name, add_summaries) - elif mode == model_fn_lib.ModeKeys.EVAL: - gan_model = _make_eval_gan_model( - generator_fn, discriminator_fn, real_data, generator_inputs, - generator_scope_name, add_summaries) - else: +def _get_gan_model( + mode, generator_fn, discriminator_fn, real_data, generator_inputs, + add_summaries, generator_scope='Generator'): + """Makes the GANModel tuple, which encapsulates the GAN model architecture.""" + if mode == model_fn_lib.ModeKeys.PREDICT: if real_data is not None: raise ValueError('`labels` must be `None` when mode is `predict`. ' 'Instead, found %s' % real_data) gan_model = _make_prediction_gan_model( - generator_inputs, generator_fn, generator_scope_name) + generator_inputs, generator_fn, generator_scope) + else: # model_fn_lib.ModeKeys.TRAIN or model_fn_lib.ModeKeys.EVAL + gan_model = _make_gan_model( + generator_fn, discriminator_fn, real_data, generator_inputs, + generator_scope, add_summaries, mode) - return head.create_estimator_spec( - features=None, - mode=mode, - logits=gan_model, - labels=None) + return gan_model + + +def _get_estimator_spec( + mode, gan_model, generator_loss_fn, discriminator_loss_fn, + get_eval_metric_ops_fn, generator_optimizer, discriminator_optimizer, + get_hooks_fn=None): + """Get the EstimatorSpec for the current mode.""" + if mode == model_fn_lib.ModeKeys.PREDICT: + estimator_spec = model_fn_lib.EstimatorSpec( + mode=mode, predictions=gan_model.generated_data) + else: + gan_loss = tfgan_tuples.GANLoss( + generator_loss=generator_loss_fn(gan_model), + discriminator_loss=discriminator_loss_fn(gan_model)) + if mode == model_fn_lib.ModeKeys.EVAL: + estimator_spec = _get_eval_estimator_spec( + gan_model, gan_loss, get_eval_metric_ops_fn) + else: # model_fn_lib.ModeKeys.TRAIN: + gopt = (generator_optimizer() if callable(generator_optimizer) else + generator_optimizer) + dopt = (discriminator_optimizer() if callable(discriminator_optimizer) + else discriminator_optimizer) + get_hooks_fn = get_hooks_fn or tfgan_train.get_sequential_train_hooks() + estimator_spec = _get_train_estimator_spec( + gan_model, gan_loss, gopt, dopt, get_hooks_fn) + + return estimator_spec def _make_gan_model(generator_fn, discriminator_fn, real_data, generator_inputs, generator_scope, add_summaries, mode): - """Make a `GANModel`, and optionally pass in `mode`.""" + """Construct a `GANModel`, and optionally pass in `mode`.""" # If network functions have an argument `mode`, pass mode to it. if 'mode' in inspect.getargspec(generator_fn).args: generator_fn = functools.partial(generator_fn, mode=mode) @@ -264,22 +263,6 @@ def _make_gan_model(generator_fn, discriminator_fn, real_data, return gan_model -def _make_train_gan_model(generator_fn, discriminator_fn, real_data, - generator_inputs, generator_scope, add_summaries): - """Make a `GANModel` for training.""" - return _make_gan_model(generator_fn, discriminator_fn, real_data, - generator_inputs, generator_scope, add_summaries, - model_fn_lib.ModeKeys.TRAIN) - - -def _make_eval_gan_model(generator_fn, discriminator_fn, real_data, - generator_inputs, generator_scope, add_summaries): - """Make a `GANModel` for evaluation.""" - return _make_gan_model(generator_fn, discriminator_fn, real_data, - generator_inputs, generator_scope, add_summaries, - model_fn_lib.ModeKeys.EVAL) - - def _make_prediction_gan_model(generator_inputs, generator_fn, generator_scope): """Make a `GANModel` from just the generator.""" # If `generator_fn` has an argument `mode`, pass mode to it. @@ -303,3 +286,46 @@ def _make_prediction_gan_model(generator_inputs, generator_fn, generator_scope): discriminator_variables=None, discriminator_scope=None, discriminator_fn=None) + + +def _get_eval_estimator_spec(gan_model, gan_loss, get_eval_metric_ops_fn=None, + name=None): + """Return an EstimatorSpec for the eval case.""" + scalar_loss = gan_loss.generator_loss + gan_loss.discriminator_loss + with ops.name_scope(None, 'metrics', + [gan_loss.generator_loss, + gan_loss.discriminator_loss]): + def _summary_key(head_name, val): + return '%s/%s' % (val, head_name) if head_name else val + eval_metric_ops = { + _summary_key(name, 'generator_loss'): + metrics_lib.mean(gan_loss.generator_loss), + _summary_key(name, 'discriminator_loss'): + metrics_lib.mean(gan_loss.discriminator_loss) + } + if get_eval_metric_ops_fn is not None: + custom_eval_metric_ops = get_eval_metric_ops_fn(gan_model) + if not isinstance(custom_eval_metric_ops, dict): + raise TypeError('get_eval_metric_ops_fn must return a dict, ' + 'received: {}'.format(custom_eval_metric_ops)) + eval_metric_ops.update(custom_eval_metric_ops) + return model_fn_lib.EstimatorSpec( + mode=model_fn_lib.ModeKeys.EVAL, + predictions=gan_model.generated_data, + loss=scalar_loss, + eval_metric_ops=eval_metric_ops) + + +def _get_train_estimator_spec( + gan_model, gan_loss, generator_optimizer, discriminator_optimizer, + get_hooks_fn, train_op_fn=tfgan_train.gan_train_ops): + """Return an EstimatorSpec for the train case.""" + scalar_loss = gan_loss.generator_loss + gan_loss.discriminator_loss + train_ops = train_op_fn(gan_model, gan_loss, generator_optimizer, + discriminator_optimizer) + training_hooks = get_hooks_fn(train_ops) + return model_fn_lib.EstimatorSpec( + loss=scalar_loss, + mode=model_fn_lib.ModeKeys.TRAIN, + train_op=train_ops.global_step_inc_op, + training_hooks=training_hooks) diff --git a/tensorflow/contrib/gan/python/estimator/python/gan_estimator_test.py b/tensorflow/contrib/gan/python/estimator/python/gan_estimator_test.py index 955482599b372be3f0d0cbc81451c514958d0eb1..9ac9c6ca9ca86a8a9abe9c0f6ebc4cdf5dd2cfb1 100644 --- a/tensorflow/contrib/gan/python/estimator/python/gan_estimator_test.py +++ b/tensorflow/contrib/gan/python/estimator/python/gan_estimator_test.py @@ -21,30 +21,30 @@ from __future__ import print_function import shutil import tempfile +from absl.testing import parameterized import numpy as np import six from tensorflow.contrib import layers -from tensorflow.contrib.gan.python import namedtuples +from tensorflow.contrib.gan.python import namedtuples as tfgan_tuples from tensorflow.contrib.gan.python.estimator.python import gan_estimator_impl as estimator from tensorflow.contrib.gan.python.losses.python import tuple_losses as losses from tensorflow.contrib.learn.python.learn.learn_io import graph_io from tensorflow.core.example import example_pb2 from tensorflow.core.example import feature_pb2 from tensorflow.python.estimator import model_fn as model_fn_lib -from tensorflow.python.estimator.canned import head as head_lib from tensorflow.python.estimator.inputs import numpy_io from tensorflow.python.framework import dtypes from tensorflow.python.framework import ops from tensorflow.python.ops import array_ops -from tensorflow.python.ops import control_flow_ops +from tensorflow.python.ops import math_ops from tensorflow.python.ops import metrics as metrics_lib from tensorflow.python.ops import parsing_ops +from tensorflow.python.ops import variable_scope from tensorflow.python.platform import test from tensorflow.python.summary.writer import writer_cache from tensorflow.python.training import input as input_lib from tensorflow.python.training import learning_rate_decay -from tensorflow.python.training import monitored_session from tensorflow.python.training import training from tensorflow.python.training import training_util @@ -60,120 +60,109 @@ def discriminator_fn(data, unused_conditioning, mode): return layers.fully_connected(data, 1) -def mock_head(testcase, expected_generator_inputs, expected_real_data, - generator_scope_name): - """Returns a mock head that validates logits values and variable names.""" - discriminator_scope_name = 'Discriminator' # comes from TFGAN defaults - generator_var_names = set([ - '%s/fully_connected/weights:0' % generator_scope_name, - '%s/fully_connected/biases:0' % generator_scope_name]) - discriminator_var_names = set([ - '%s/fully_connected/weights:0' % discriminator_scope_name, - '%s/fully_connected/biases:0' % discriminator_scope_name]) - - def _create_estimator_spec(features, mode, logits, labels): - gan_model = logits # renaming for clarity - is_predict = mode == model_fn_lib.ModeKeys.PREDICT - testcase.assertIsNone(features) - testcase.assertIsNone(labels) - testcase.assertIsInstance(gan_model, namedtuples.GANModel) - - trainable_vars = ops.get_collection(ops.GraphKeys.TRAINABLE_VARIABLES) - expected_var_names = (generator_var_names if is_predict else - generator_var_names | discriminator_var_names) - testcase.assertItemsEqual(expected_var_names, - [var.name for var in trainable_vars]) - - assertions = [] - def _or_none(x): - return None if is_predict else x - testcase.assertEqual(expected_generator_inputs, gan_model.generator_inputs) - # TODO(joelshor): Add check on `generated_data`. - testcase.assertItemsEqual( - generator_var_names, - set([x.name for x in gan_model.generator_variables])) - testcase.assertEqual(generator_scope_name, gan_model.generator_scope.name) - testcase.assertEqual(_or_none(expected_real_data), gan_model.real_data) - # TODO(joelshor): Add check on `discriminator_real_outputs`. - # TODO(joelshor): Add check on `discriminator_gen_outputs`. - if is_predict: - testcase.assertIsNone(gan_model.discriminator_scope) - else: - testcase.assertEqual(discriminator_scope_name, - gan_model.discriminator_scope.name) - - with ops.control_dependencies(assertions): - if mode == model_fn_lib.ModeKeys.TRAIN: - return model_fn_lib.EstimatorSpec( - mode=mode, loss=array_ops.zeros([]), - train_op=control_flow_ops.no_op(), training_hooks=[]) - elif mode == model_fn_lib.ModeKeys.EVAL: - return model_fn_lib.EstimatorSpec( - mode=mode, predictions=gan_model.generated_data, - loss=array_ops.zeros([])) - elif mode == model_fn_lib.ModeKeys.PREDICT: - return model_fn_lib.EstimatorSpec( - mode=mode, predictions=gan_model.generated_data) - else: - testcase.fail('Invalid mode: {}'.format(mode)) - - head = test.mock.NonCallableMagicMock(spec=head_lib._Head) - head.create_estimator_spec = test.mock.MagicMock( - wraps=_create_estimator_spec) - - return head - - -class GANModelFnTest(test.TestCase): - """Tests that _gan_model_fn passes expected logits to mock head.""" - - def setUp(self): - self._model_dir = tempfile.mkdtemp() - - def tearDown(self): - if self._model_dir: - writer_cache.FileWriterCache.clear() - shutil.rmtree(self._model_dir) +class GetGANModelTest(test.TestCase, parameterized.TestCase): + """Tests that `GetGANModel` produces the correct model.""" - def _test_logits_helper(self, mode): - """Tests that the expected logits are passed to mock head.""" + @parameterized.named_parameters( + ('train', model_fn_lib.ModeKeys.TRAIN), + ('eval', model_fn_lib.ModeKeys.EVAL), + ('predict', model_fn_lib.ModeKeys.PREDICT)) + def test_get_gan_model(self, mode): with ops.Graph().as_default(): - training_util.get_or_create_global_step() - generator_inputs = {'x': array_ops.zeros([5, 4])} - real_data = (None if mode == model_fn_lib.ModeKeys.PREDICT else - array_ops.zeros([5, 4])) - generator_scope_name = 'generator' - head = mock_head(self, - expected_generator_inputs=generator_inputs, - expected_real_data=real_data, - generator_scope_name=generator_scope_name) - estimator_spec = estimator._gan_model_fn( - features=generator_inputs, - labels=real_data, - mode=mode, - generator_fn=generator_fn, - discriminator_fn=discriminator_fn, - generator_scope_name=generator_scope_name, - head=head) - with monitored_session.MonitoredTrainingSession( - checkpoint_dir=self._model_dir) as sess: - if mode == model_fn_lib.ModeKeys.TRAIN: - sess.run(estimator_spec.train_op) - elif mode == model_fn_lib.ModeKeys.EVAL: - sess.run(estimator_spec.loss) - elif mode == model_fn_lib.ModeKeys.PREDICT: - sess.run(estimator_spec.predictions) - else: - self.fail('Invalid mode: {}'.format(mode)) - - def test_logits_predict(self): - self._test_logits_helper(model_fn_lib.ModeKeys.PREDICT) - - def test_logits_eval(self): - self._test_logits_helper(model_fn_lib.ModeKeys.EVAL) - - def test_logits_train(self): - self._test_logits_helper(model_fn_lib.ModeKeys.TRAIN) + generator_inputs = {'x': array_ops.ones([3, 4])} + real_data = (array_ops.zeros([3, 4]) if + mode != model_fn_lib.ModeKeys.PREDICT else None) + gan_model = estimator._get_gan_model( + mode, generator_fn, discriminator_fn, real_data, generator_inputs, + add_summaries=False) + + self.assertEqual(generator_inputs, gan_model.generator_inputs) + self.assertIsNotNone(gan_model.generated_data) + self.assertEqual(2, len(gan_model.generator_variables)) # 1 FC layer + self.assertIsNotNone(gan_model.generator_fn) + if mode == model_fn_lib.ModeKeys.PREDICT: + self.assertIsNone(gan_model.real_data) + self.assertIsNone(gan_model.discriminator_real_outputs) + self.assertIsNone(gan_model.discriminator_gen_outputs) + self.assertIsNone(gan_model.discriminator_variables) + self.assertIsNone(gan_model.discriminator_scope) + self.assertIsNone(gan_model.discriminator_fn) + else: + self.assertIsNotNone(gan_model.real_data) + self.assertIsNotNone(gan_model.discriminator_real_outputs) + self.assertIsNotNone(gan_model.discriminator_gen_outputs) + self.assertEqual(2, len(gan_model.discriminator_variables)) # 1 FC layer + self.assertIsNotNone(gan_model.discriminator_scope) + self.assertIsNotNone(gan_model.discriminator_fn) + + +def get_dummy_gan_model(): + # TODO(joelshor): Find a better way of creating a variable scope. + with variable_scope.variable_scope('generator') as gen_scope: + gen_var = variable_scope.get_variable('dummy_var', initializer=0.0) + with variable_scope.variable_scope('discriminator') as dis_scope: + dis_var = variable_scope.get_variable('dummy_var', initializer=0.0) + return tfgan_tuples.GANModel( + generator_inputs=None, + generated_data=array_ops.ones([3, 4]), + generator_variables=[gen_var], + generator_scope=gen_scope, + generator_fn=None, + real_data=array_ops.zeros([3, 4]), + discriminator_real_outputs=array_ops.ones([1, 2, 3]) * dis_var, + discriminator_gen_outputs=array_ops.ones([1, 2, 3]) * gen_var * dis_var, + discriminator_variables=[dis_var], + discriminator_scope=dis_scope, + discriminator_fn=None) + + +def dummy_loss_fn(gan_model): + return math_ops.reduce_sum(gan_model.discriminator_real_outputs - + gan_model.discriminator_gen_outputs) + + +def get_metrics(gan_model): + return { + 'mse_custom_metric': metrics_lib.mean_squared_error( + gan_model.real_data, gan_model.generated_data) + } + + +class GetEstimatorSpecTest(test.TestCase, parameterized.TestCase): + """Tests that the EstimatorSpec is constructed appropriately.""" + + @classmethod + def setUpClass(cls): + cls._generator_optimizer = training.GradientDescentOptimizer(1.0) + cls._discriminator_optimizer = training.GradientDescentOptimizer(1.0) + + @parameterized.named_parameters( + ('train', model_fn_lib.ModeKeys.TRAIN), + ('eval', model_fn_lib.ModeKeys.EVAL), + ('predict', model_fn_lib.ModeKeys.PREDICT)) + def test_get_estimator_spec(self, mode): + with ops.Graph().as_default(): + self._gan_model = get_dummy_gan_model() + spec = estimator._get_estimator_spec( + mode, + self._gan_model, + generator_loss_fn=dummy_loss_fn, + discriminator_loss_fn=dummy_loss_fn, + get_eval_metric_ops_fn=get_metrics, + generator_optimizer=self._generator_optimizer, + discriminator_optimizer=self._discriminator_optimizer) + + self.assertEqual(mode, spec.mode) + if mode == model_fn_lib.ModeKeys.PREDICT: + self.assertEqual(self._gan_model.generated_data, spec.predictions) + elif mode == model_fn_lib.ModeKeys.TRAIN: + self.assertShapeEqual(np.array(0), spec.loss) # must be a scalar + self.assertIsNotNone(spec.train_op) + self.assertIsNotNone(spec.training_hooks) + elif mode == model_fn_lib.ModeKeys.EVAL: + self.assertEqual(self._gan_model.generated_data, spec.predictions) + self.assertShapeEqual(np.array(0), spec.loss) # must be a scalar + self.assertIsNotNone(spec.eval_metric_ops) # TODO(joelshor): Add pandas test. @@ -195,12 +184,6 @@ class GANEstimatorIntegrationTest(test.TestCase): lr = learning_rate_decay.exponential_decay(1.0, gstep, 10, 0.9) return training.GradientDescentOptimizer(lr) - def get_metrics(gan_model): - return { - 'mse_custom_metric': metrics_lib.mean_squared_error( - gan_model.real_data, gan_model.generated_data) - } - gopt = make_opt if lr_decay else training.GradientDescentOptimizer(1.0) dopt = make_opt if lr_decay else training.GradientDescentOptimizer(1.0) est = estimator.GANEstimator( diff --git a/tensorflow/contrib/gan/python/estimator/python/head_impl.py b/tensorflow/contrib/gan/python/estimator/python/head_impl.py index 5b5557bd8f12b4d42e508f185cb8561eaebea84e..1a0ee6dfc498eb6dc8c97411589d9e35bc352062 100644 --- a/tensorflow/contrib/gan/python/estimator/python/head_impl.py +++ b/tensorflow/contrib/gan/python/estimator/python/head_impl.py @@ -27,16 +27,21 @@ from tensorflow.python.estimator.canned import head from tensorflow.python.estimator.export import export_output from tensorflow.python.framework import ops from tensorflow.python.ops import metrics as metrics_lib +from tensorflow.python.util import deprecation __all__ = [ 'GANHead', 'gan_head', ] + def _summary_key(head_name, val): return '%s/%s' % (val, head_name) if head_name else val +@deprecation.deprecated( + None, 'Please use tf.contrib.gan.GANEstimator without explicitly making a ' + 'GANHead.') def gan_head(generator_loss_fn, discriminator_loss_fn, generator_optimizer, discriminator_optimizer, use_loss_summaries=True, get_hooks_fn=tfgan_train.get_sequential_train_hooks(), @@ -77,6 +82,9 @@ def gan_head(generator_loss_fn, discriminator_loss_fn, generator_optimizer, class GANHead(head._Head): # pylint: disable=protected-access """`Head` for a GAN.""" + @deprecation.deprecated( + None, 'Please use tf.contrib.gan.GANEstimator without explicitly making ' + 'a GANHead.') def __init__(self, generator_loss_fn, discriminator_loss_fn, generator_optimizer, discriminator_optimizer, use_loss_summaries=True, @@ -103,9 +111,20 @@ class GANHead(head._Head): # pylint: disable=protected-access name: name of the head. If provided, summary and metrics keys will be suffixed by `"/" + name`. """ + + if not callable(generator_loss_fn): + raise TypeError('generator_loss_fn must be callable.') + if not callable(discriminator_loss_fn): + raise TypeError('discriminator_loss_fn must be callable.') + if use_loss_summaries not in [True, False, None]: + raise ValueError('use_loss_summaries must be True, False or None.') + if get_hooks_fn is not None and not callable(get_hooks_fn): + raise TypeError('get_hooks_fn must be callable.') + if name is not None and not isinstance(name, str): + raise TypeError('name must be string.') + if get_hooks_fn is None: get_hooks_fn = tfgan_train.get_sequential_train_hooks() - # TODO(joelshor): Validate inputs. if use_loss_summaries in [True, False]: generator_loss_fn = functools.partial( diff --git a/tensorflow/contrib/gan/python/estimator/python/head_test.py b/tensorflow/contrib/gan/python/estimator/python/head_test.py index 5309d87765694fa476dae006105e842420a7c437..8205bc889dc01c8680e2139393d65723280cfbd0 100644 --- a/tensorflow/contrib/gan/python/estimator/python/head_test.py +++ b/tensorflow/contrib/gan/python/estimator/python/head_test.py @@ -67,7 +67,7 @@ class GANHeadTest(test.TestCase): generator_optimizer=training.GradientDescentOptimizer(1.0), discriminator_optimizer=training.GradientDescentOptimizer(1.0), get_eval_metric_ops_fn=self.get_metrics) - self.assertTrue(isinstance(self.gan_head, head.GANHead)) + self.assertIsInstance(self.gan_head, head.GANHead) def get_metrics(self, gan_model): self.assertTrue(isinstance(gan_model, tfgan_tuples.GANModel)) diff --git a/tensorflow/contrib/gdr/gdr_memory_manager.cc b/tensorflow/contrib/gdr/gdr_memory_manager.cc index 81e70ae30a4c72dbcedd1aabfe758ecca4c8b366..1435e19109ca2f3bbd6ce70e6e5f26a92dfc2713 100644 --- a/tensorflow/contrib/gdr/gdr_memory_manager.cc +++ b/tensorflow/contrib/gdr/gdr_memory_manager.cc @@ -34,8 +34,9 @@ limitations under the License. #include "tensorflow/core/common_runtime/device.h" #include "tensorflow/core/common_runtime/dma_helper.h" #if GOOGLE_CUDA +#include "tensorflow/core/common_runtime/gpu/gpu_process_state.h" #include "tensorflow/core/common_runtime/gpu/gpu_util.h" -#include "tensorflow/core/common_runtime/gpu/process_state.h" +#include "tensorflow/core/common_runtime/process_state.h" #endif // GOOGLE_CUDA #include "tensorflow/core/framework/allocator_registry.h" #include "tensorflow/core/lib/core/status.h" @@ -274,7 +275,7 @@ Status GdrMemoryManager::Init() { Allocator* allocators[] = { #if GOOGLE_CUDA - ProcessState::singleton()->GetCUDAHostAllocator(0), + GPUProcessState::singleton()->GetCUDAHostAllocator(0), ProcessState::singleton()->GetCPUAllocator(0), #endif // GOOGLE_CUDA cpu_allocator(), @@ -308,7 +309,8 @@ Status GdrMemoryManager::Init() { if (IsGDRAvailable()) { // Note we don't free allocated GPU memory so there is no free visitor int32_t bus_id = TryToReadNumaNode(listening_->verbs->device) + 1; - ProcessState::singleton()->AddGPUAllocVisitor(bus_id, cuda_alloc_visitor); + GPUProcessState::singleton()->AddGPUAllocVisitor(bus_id, + cuda_alloc_visitor); LOG(INFO) << "Instrumenting GPU allocator with bus_id " << bus_id; } #endif // GOOGLE_CUDA @@ -430,7 +432,7 @@ void GdrMemoryManager::TransportOptionsFromTensor( #if GOOGLE_CUDA if (!on_host) { - Allocator* alloc = ProcessState::singleton()->GetCUDAHostAllocator(0); + Allocator* alloc = GPUProcessState::singleton()->GetCUDAHostAllocator(0); Tensor* host_copy = new Tensor(alloc, tensor.dtype(), tensor.shape()); GPUUtil::CopyGPUTensorToCPU( device, device_context, &tensor, host_copy, @@ -532,7 +534,7 @@ void GdrMemoryManager::TensorFromTransportOptions( Tensor host_copy; #if GOOGLE_CUDA if (mr == nullptr && !on_host) { - Allocator* alloc = ProcessState::singleton()->GetCUDAHostAllocator(0); + Allocator* alloc = GPUProcessState::singleton()->GetCUDAHostAllocator(0); host_copy = Tensor(alloc, tensor->dtype(), tensor->shape()); buffer = DMAHelper::buffer(&host_copy); addr = buffer->data(); diff --git a/tensorflow/contrib/image/kernels/image_ops.cc b/tensorflow/contrib/image/kernels/image_ops.cc index c2e32da133b32c8fe169302668031af8bace2c22..022e17d13963a14f81d76e683d13060d1f3f8a7e 100644 --- a/tensorflow/contrib/image/kernels/image_ops.cc +++ b/tensorflow/contrib/image/kernels/image_ops.cc @@ -35,6 +35,7 @@ typedef Eigen::ThreadPoolDevice CPUDevice; template struct FillProjectiveTransform; template struct FillProjectiveTransform; template struct FillProjectiveTransform; +template struct FillProjectiveTransform; template struct FillProjectiveTransform; template struct FillProjectiveTransform; @@ -99,6 +100,7 @@ class ImageProjectiveTransform : public OpKernel { TF_CALL_uint8(REGISTER); TF_CALL_int32(REGISTER); TF_CALL_int64(REGISTER); +TF_CALL_half(REGISTER); TF_CALL_float(REGISTER); TF_CALL_double(REGISTER); diff --git a/tensorflow/contrib/image/kernels/image_ops.h b/tensorflow/contrib/image/kernels/image_ops.h index ad501330617be89c87a0e94ab6e8773a6e1eecf6..209aa24548443bb10c13cd506b8c93c23cfff4a4 100644 --- a/tensorflow/contrib/image/kernels/image_ops.h +++ b/tensorflow/contrib/image/kernels/image_ops.h @@ -21,6 +21,7 @@ limitations under the License. #define EIGEN_USE_THREADS #include "third_party/eigen3/unsupported/Eigen/CXX11/Tensor" + #include "tensorflow/core/framework/tensor_types.h" #include "tensorflow/core/platform/types.h" @@ -58,6 +59,11 @@ class ProjectiveGenerator { ? transforms_.data() : &transforms_.data()[transforms_.dimension(1) * coords[0]]; float projection = transform[6] * output_x + transform[7] * output_y + 1.f; + if (projection == 0) { + // Return the fill value (0) for infinite coordinates, + // which are outside the input image + return T(0); + } const float input_x = (transform[0] * output_x + transform[1] * output_y + transform[2]) / projection; @@ -105,21 +111,21 @@ class ProjectiveGenerator { // f(x, y_floor) = (x_ceil - x) / (x_ceil - x_floor) * f(x_floor, y_floor) // + (x - x_floor) / (x_ceil - x_floor) * f(x_ceil, y_floor) const float value_yfloor = - (x_ceil - x) * read_with_fill_value(batch, DenseIndex(y_floor), - DenseIndex(x_floor), channel, - fill_value) + - (x - x_floor) * read_with_fill_value(batch, DenseIndex(y_floor), - DenseIndex(x_ceil), channel, - fill_value); + (x_ceil - x) * static_cast(read_with_fill_value( + batch, DenseIndex(y_floor), DenseIndex(x_floor), + channel, fill_value)) + + (x - x_floor) * static_cast(read_with_fill_value( + batch, DenseIndex(y_floor), DenseIndex(x_ceil), + channel, fill_value)); // f(x, y_ceil) = (x_ceil - x) / (x_ceil - x_floor) * f(x_floor, y_ceil) // + (x - x_floor) / (x_ceil - x_floor) * f(x_ceil, y_ceil) const float value_yceil = - (x_ceil - x) * read_with_fill_value(batch, DenseIndex(y_ceil), - DenseIndex(x_floor), channel, - fill_value) + - (x - x_floor) * read_with_fill_value(batch, DenseIndex(y_ceil), - DenseIndex(x_ceil), channel, - fill_value); + (x_ceil - x) * static_cast(read_with_fill_value( + batch, DenseIndex(y_ceil), DenseIndex(x_floor), + channel, fill_value)) + + (x - x_floor) * static_cast(read_with_fill_value( + batch, DenseIndex(y_ceil), DenseIndex(x_ceil), + channel, fill_value)); // f(x, y) = (y_ceil - y) / (y_ceil - y_floor) * f(x, y_floor) // + (y - y_floor) / (y_ceil - y_floor) * f(x, y_ceil) return T((y_ceil - y) * value_yfloor + (y - y_floor) * value_yceil); diff --git a/tensorflow/contrib/image/ops/image_ops.cc b/tensorflow/contrib/image/ops/image_ops.cc index ebdcaea7abae2a967786831b62b331897aa3f6a3..e59f1bf8443732a4b84fe7461439e3d0ee7dd158 100644 --- a/tensorflow/contrib/image/ops/image_ops.cc +++ b/tensorflow/contrib/image/ops/image_ops.cc @@ -29,7 +29,7 @@ using shape_inference::ShapeHandle; REGISTER_OP("ImageProjectiveTransform") .Input("images: dtype") .Input("transforms: float32") - .Attr("dtype: {uint8, int32, int64, float32, float64}") + .Attr("dtype: {uint8, int32, int64, float16, float32, float64}") .Attr("interpolation: string") .Output("transformed_images: dtype") .SetShapeFn([](InferenceContext* c) { diff --git a/tensorflow/contrib/image/python/kernel_tests/image_ops_test.py b/tensorflow/contrib/image/python/kernel_tests/image_ops_test.py index b50177ae5651fbc15f292e11031411c2074357ec..62a22dcf3411fb160b3c432bbdd67303697f7262 100644 --- a/tensorflow/contrib/image/python/kernel_tests/image_ops_test.py +++ b/tensorflow/contrib/image/python/kernel_tests/image_ops_test.py @@ -30,7 +30,8 @@ from tensorflow.python.ops import math_ops from tensorflow.python.platform import googletest _DTYPES = set( - [dtypes.uint8, dtypes.int32, dtypes.int64, dtypes.float32, dtypes.float64]) + [dtypes.uint8, dtypes.int32, dtypes.int64, + dtypes.float16, dtypes.float32, dtypes.float64]) class ImageOpsTest(test_util.TensorFlowTestCase): @@ -127,6 +128,23 @@ class ImageOpsTest(test_util.TensorFlowTestCase): [0, 1, 0, 1], [0, 1, 1, 1]]) + def test_extreme_projective_transform(self): + for dtype in _DTYPES: + with self.test_session(): + image = constant_op.constant( + [[1, 0, 1, 0], + [0, 1, 0, 1], + [1, 0, 1, 0], + [0, 1, 0, 1]], dtype=dtype) + transformation = constant_op.constant([1, 0, 0, 0, 1, 0, -1, 0], + dtypes.float32) + image_transformed = image_ops.transform(image, transformation) + self.assertAllEqual(image_transformed.eval(), + [[1, 0, 0, 0], + [0, 0, 0, 0], + [1, 0, 0, 0], + [0, 0, 0, 0]]) + def test_bilinear(self): with self.test_session(): image = constant_op.constant( diff --git a/tensorflow/contrib/image/python/ops/image_ops.py b/tensorflow/contrib/image/python/ops/image_ops.py index cd984c80543886be1f682933e2e003bd3374e425..86b0ffe9a0f2236d5ac7d5f846e7b5d2615c9b09 100644 --- a/tensorflow/contrib/image/python/ops/image_ops.py +++ b/tensorflow/contrib/image/python/ops/image_ops.py @@ -33,7 +33,8 @@ _image_ops_so = loader.load_op_library( resource_loader.get_path_to_datafile("_image_ops.so")) _IMAGE_DTYPES = set( - [dtypes.uint8, dtypes.int32, dtypes.int64, dtypes.float32, dtypes.float64]) + [dtypes.uint8, dtypes.int32, dtypes.int64, + dtypes.float16, dtypes.float32, dtypes.float64]) ops.RegisterShape("ImageConnectedComponents")(common_shapes.call_cpp_shape_fn) ops.RegisterShape("ImageProjectiveTransform")(common_shapes.call_cpp_shape_fn) diff --git a/tensorflow/contrib/kafka/ops/kafka_ops.cc b/tensorflow/contrib/kafka/ops/kafka_ops.cc new file mode 100644 index 0000000000000000000000000000000000000000..8cdf16103bab2b22d51c144d21a589e1e39f2f0b --- /dev/null +++ b/tensorflow/contrib/kafka/ops/kafka_ops.cc @@ -0,0 +1,44 @@ +/* Copyright 2017 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow/core/framework/common_shape_fns.h" +#include "tensorflow/core/framework/op.h" +#include "tensorflow/core/framework/shape_inference.h" + +namespace tensorflow { + +REGISTER_OP("KafkaDataset") + .Input("topics: string") + .Input("servers: string") + .Input("group: string") + .Input("eof: bool") + .Input("timeout: int64") + .Output("handle: variant") + .SetIsStateful() + .SetShapeFn(shape_inference::ScalarShape) + .Doc(R"doc( +Creates a dataset that emits the messages of one or more Kafka topics. + +topics: A `tf.string` tensor containing one or more subscriptions, + in the format of [topic:partition:offset:length], + by default length is -1 for unlimited. +servers: A list of bootstrap servers. +group: The consumer group id. +eof: If True, the kafka reader will stop on EOF. +timeout: The timeout value for the Kafka Consumer to wait + (in millisecond). +)doc"); + +} // namespace tensorflow diff --git a/tensorflow/contrib/kinesis/BUILD b/tensorflow/contrib/kinesis/BUILD new file mode 100644 index 0000000000000000000000000000000000000000..25443d0ad47aa7d503f905eb34000488b62f22c6 --- /dev/null +++ b/tensorflow/contrib/kinesis/BUILD @@ -0,0 +1,113 @@ +package(default_visibility = ["//tensorflow:internal"]) + +licenses(["notice"]) # Apache 2.0 + +exports_files(["LICENSE"]) + +load( + "//tensorflow:tensorflow.bzl", + "tf_custom_op_library", + "tf_custom_op_py_library", + "tf_gen_op_libs", + "tf_gen_op_wrapper_py", + "tf_kernel_library", + "tf_py_test", +) + +py_library( + name = "kinesis", + srcs = ["__init__.py"], + srcs_version = "PY2AND3", + deps = [ + ":dataset_ops", + ], +) + +tf_custom_op_library( + name = "_dataset_ops.so", + srcs = ["ops/dataset_ops.cc"], + deps = [":dataset_kernels"], +) + +tf_gen_op_libs( + op_lib_names = ["dataset_ops"], +) + +cc_library( + name = "dataset_kernels", + srcs = [ + "kernels/kinesis_dataset_ops.cc", + ], + deps = [ + "//tensorflow/core:framework_headers_lib", + "//tensorflow/core/platform/s3:aws_crypto", + "//third_party/eigen3", + "@aws", + "@protobuf_archive//:protobuf_headers", + ], + alwayslink = 1, +) + +py_library( + name = "dataset_ops", + srcs = [ + "python/ops/kinesis_dataset_ops.py", + ], + srcs_version = "PY2AND3", + deps = [ + ":kinesis_op_loader", + "//tensorflow/python:dataset_ops_gen", + "//tensorflow/python:util", + "//tensorflow/python/data/ops:dataset_ops", + "//tensorflow/python/data/util:nest", + ], +) + +tf_gen_op_wrapper_py( + name = "gen_dataset_ops", + out = "python/ops/gen_dataset_ops.py", + deps = ["//tensorflow/contrib/kinesis:dataset_ops_op_lib"], +) + +tf_kernel_library( + name = "dataset_ops_kernels", + deps = [ + ":dataset_kernels", + "//tensorflow/core:framework", + ], + alwayslink = 1, +) + +tf_custom_op_py_library( + name = "kinesis_op_loader", + srcs = ["python/ops/kinesis_op_loader.py"], + dso = ["//tensorflow/contrib/kinesis:_dataset_ops.so"], + kernels = [ + ":dataset_ops_kernels", + "//tensorflow/contrib/kinesis:dataset_ops_op_lib", + ], + srcs_version = "PY2AND3", + deps = [ + ":gen_dataset_ops", + "//tensorflow/contrib/util:util_py", + "//tensorflow/python:platform", + ], +) + +tf_py_test( + name = "kinesis_test", + srcs = ["python/kernel_tests/kinesis_test.py"], + additional_deps = [ + ":kinesis", + "//third_party/py/numpy", + "//tensorflow/python:client_testlib", + "//tensorflow/python:framework", + "//tensorflow/python:framework_test_lib", + "//tensorflow/python:platform_test", + ], + tags = [ + "manual", + "no_windows", + "notap", + ], +) diff --git a/tensorflow/contrib/kinesis/__init__.py b/tensorflow/contrib/kinesis/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..3824b8ae7532ab97a5ebf01ab66ece6476c87d42 --- /dev/null +++ b/tensorflow/contrib/kinesis/__init__.py @@ -0,0 +1,32 @@ +# Copyright 2018 The TensorFlow Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================== +"""Kinesis Dataset. + +@@KinesisDataset +""" + +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + +from tensorflow.contrib.kinesis.python.ops.kinesis_dataset_ops import KinesisDataset + +from tensorflow.python.util.all_util import remove_undocumented + +_allowed_symbols = [ + "KinesisDataset", +] + +remove_undocumented(__name__) diff --git a/tensorflow/contrib/kinesis/kernels/kinesis_dataset_ops.cc b/tensorflow/contrib/kinesis/kernels/kinesis_dataset_ops.cc new file mode 100644 index 0000000000000000000000000000000000000000..3212279c4c50efb92acc712b82cb3e1a22c76870 --- /dev/null +++ b/tensorflow/contrib/kinesis/kernels/kinesis_dataset_ops.cc @@ -0,0 +1,359 @@ +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include "tensorflow/core/framework/dataset.h" +#include "tensorflow/core/platform/s3/aws_crypto.h" + +namespace tensorflow { +namespace { + +Aws::Client::ClientConfiguration* InitializeDefaultClientConfig() { + static Aws::Client::ClientConfiguration config; + const char* endpoint = getenv("KINESIS_ENDPOINT"); + if (endpoint) { + config.endpointOverride = Aws::String(endpoint); + } + const char* region = getenv("AWS_REGION"); + if (region) { + config.region = Aws::String(region); + } else { + // Load config file (e.g., ~/.aws/config) only if AWS_SDK_LOAD_CONFIG + // is set with a truthy value. + const char* load_config_env = getenv("AWS_SDK_LOAD_CONFIG"); + string load_config = + load_config_env ? str_util::Lowercase(load_config_env) : ""; + if (load_config == "true" || load_config == "1") { + Aws::String config_file; + // If AWS_CONFIG_FILE is set then use it, otherwise use ~/.aws/config. + const char* config_file_env = getenv("AWS_CONFIG_FILE"); + if (config_file_env) { + config_file = config_file_env; + } else { + const char* home_env = getenv("HOME"); + if (home_env) { + config_file = home_env; + config_file += "/.aws/config"; + } + } + Aws::Config::AWSConfigFileProfileConfigLoader loader(config_file); + // Load the configuration. If successful, get the region. + // If the load is not successful, then generate a warning. + if (loader.Load()) { + auto profiles = loader.GetProfiles(); + if (!profiles["default"].GetRegion().empty()) { + config.region = profiles["default"].GetRegion(); + } + } else { + LOG(WARNING) << "Failed to load the profile in " << config_file << "."; + } + } + } + const char* use_https = getenv("KINESIS_USE_HTTPS"); + if (use_https) { + if (use_https[0] == '0') { + config.scheme = Aws::Http::Scheme::HTTP; + } else { + config.scheme = Aws::Http::Scheme::HTTPS; + } + } + const char* verify_ssl = getenv("KINESIS_VERIFY_SSL"); + if (verify_ssl) { + if (verify_ssl[0] == '0') { + config.verifySSL = false; + } else { + config.verifySSL = true; + } + } + const char* connect_timeout = getenv("KINESIS_CONNECT_TIMEOUT_MSEC"); + if (connect_timeout) { + int64 timeout; + + if (strings::safe_strto64(connect_timeout, &timeout)) { + config.connectTimeoutMs = timeout; + } + } + const char* request_timeout = getenv("KINESIS_REQUEST_TIMEOUT_MSEC"); + if (request_timeout) { + int64 timeout; + + if (strings::safe_strto64(request_timeout, &timeout)) { + config.requestTimeoutMs = timeout; + } + } + + return &config; +} + +Aws::Client::ClientConfiguration& GetDefaultClientConfig() { + static Aws::Client::ClientConfiguration* config = + InitializeDefaultClientConfig(); + return *config; +} + +static mutex mu(LINKER_INITIALIZED); +static unsigned count(0); +void AwsInitAPI() { + mutex_lock lock(mu); + count++; + if (count == 1) { + Aws::SDKOptions options; + options.cryptoOptions.sha256Factory_create_fn = []() { + return Aws::MakeShared(AWSCryptoAllocationTag); + }; + options.cryptoOptions.sha256HMACFactory_create_fn = []() { + return Aws::MakeShared(AWSCryptoAllocationTag); + }; + Aws::InitAPI(options); + } +} +void AwsShutdownAPI() { + mutex_lock lock(mu); + count--; + if (count == 0) { + Aws::SDKOptions options; + Aws::ShutdownAPI(options); + } +} +void ShutdownClient(Aws::Kinesis::KinesisClient* client) { + if (client != nullptr) { + delete client; + AwsShutdownAPI(); + } +} +} +class KinesisDatasetOp : public DatasetOpKernel { + public: + using DatasetOpKernel::DatasetOpKernel; + + void MakeDataset(OpKernelContext* ctx, DatasetBase** output) override { + std::string stream = ""; + OP_REQUIRES_OK(ctx, + ParseScalarArgument(ctx, "stream", &stream)); + std::string shard = ""; + OP_REQUIRES_OK(ctx, ParseScalarArgument(ctx, "shard", &shard)); + bool read_indefinitely = true; + OP_REQUIRES_OK(ctx, ParseScalarArgument(ctx, "read_indefinitely", + &read_indefinitely)); + int64 interval = -1; + OP_REQUIRES_OK(ctx, ParseScalarArgument(ctx, "interval", &interval)); + OP_REQUIRES(ctx, (interval > 0), + errors::InvalidArgument( + "Interval value should be large than 0, got ", interval)); + *output = new Dataset(ctx, stream, shard, read_indefinitely, interval); + } + + private: + class Dataset : public GraphDatasetBase { + public: + Dataset(OpKernelContext* ctx, const string& stream, const string& shard, + const bool read_indefinitely, const int64 interval) + : GraphDatasetBase(ctx), + stream_(stream), + shard_(shard), + read_indefinitely_(read_indefinitely), + interval_(interval) {} + + std::unique_ptr MakeIteratorInternal( + const string& prefix) const override { + return std::unique_ptr( + new Iterator({this, strings::StrCat(prefix, "::Kinesis")})); + } + + const DataTypeVector& output_dtypes() const override { + static DataTypeVector* dtypes = new DataTypeVector({DT_STRING}); + return *dtypes; + } + + const std::vector& output_shapes() const override { + static std::vector* shapes = + new std::vector({{}}); + return *shapes; + } + + string DebugString() const override { return "KinesisDatasetOp::Dataset"; } + + protected: + Status AsGraphDefInternal(DatasetGraphDefBuilder* b, + Node** output) const override { + Node* stream = nullptr; + TF_RETURN_IF_ERROR(b->AddScalar(stream_, &stream)); + Node* shard = nullptr; + TF_RETURN_IF_ERROR(b->AddScalar(shard_, &shard)); + Node* read_indefinitely = nullptr; + TF_RETURN_IF_ERROR(b->AddScalar(read_indefinitely_, &read_indefinitely)); + Node* interval = nullptr; + TF_RETURN_IF_ERROR(b->AddScalar(interval_, &interval)); + TF_RETURN_IF_ERROR(b->AddDataset( + this, {stream, shard, read_indefinitely, interval}, output)); + return Status::OK(); + } + + private: + class Iterator : public DatasetIterator { + public: + explicit Iterator(const Params& params) + : DatasetIterator(params), + client_(nullptr, ShutdownClient) {} + + Status GetNextInternal(IteratorContext* ctx, + std::vector* out_tensors, + bool* end_of_sequence) override { + mutex_lock l(mu_); + if (iterator_ == "") { + TF_RETURN_IF_ERROR(SetupStreamsLocked()); + } + do { + Aws::Kinesis::Model::GetRecordsRequest request; + auto outcome = client_->GetRecords( + request.WithShardIterator(iterator_).WithLimit(1)); + if (!outcome.IsSuccess()) { + return errors::Unknown(outcome.GetError().GetExceptionName(), ": ", + outcome.GetError().GetMessage()); + } + if (outcome.GetResult().GetRecords().size() == 0) { + // If no records were returned then nothing is available at the + // moment. + if (!dataset()->read_indefinitely_) { + *end_of_sequence = true; + return Status::OK(); + } + // Continue the loop after a period of time. + ctx->env()->SleepForMicroseconds(dataset()->interval_); + continue; + } + if (outcome.GetResult().GetRecords().size() != 1) { + return errors::Unknown("invalid number of records ", + outcome.GetResult().GetRecords().size(), + " returned"); + } + + iterator_ = outcome.GetResult().GetNextShardIterator(); + + const auto& data = outcome.GetResult().GetRecords()[0].GetData(); + StringPiece value( + reinterpret_cast(data.GetUnderlyingData()), + data.GetLength()); + Tensor value_tensor(ctx->allocator({}), DT_STRING, {}); + value_tensor.scalar()() = std::string(value); + out_tensors->emplace_back(std::move(value_tensor)); + + *end_of_sequence = false; + return Status::OK(); + } while (true); + } + + protected: + Status SaveInternal(IteratorStateWriter* writer) override { + return errors::Unimplemented("SaveInternal is currently not supported"); + } + + Status RestoreInternal(IteratorContext* ctx, + IteratorStateReader* reader) override { + return errors::Unimplemented( + "RestoreInternal is currently not supported"); + } + + private: + // Sets up Kinesis streams to read from. + Status SetupStreamsLocked() EXCLUSIVE_LOCKS_REQUIRED(mu_) { + AwsInitAPI(); + client_.reset( + new Aws::Kinesis::KinesisClient(GetDefaultClientConfig())); + + Aws::Kinesis::Model::DescribeStreamRequest request; + auto outcome = client_->DescribeStream( + request.WithStreamName(dataset()->stream_.c_str())); + if (!outcome.IsSuccess()) { + return errors::Unknown(outcome.GetError().GetExceptionName(), ": ", + outcome.GetError().GetMessage()); + } + Aws::String shard; + Aws::String sequence; + if (dataset()->shard_ == "") { + if (outcome.GetResult().GetStreamDescription().GetShards().size() != + 1) { + return errors::InvalidArgument( + "shard has to be provided unless the stream only have one " + "shard, there are ", + outcome.GetResult().GetStreamDescription().GetShards().size(), + " shards in stream ", dataset()->stream_); + } + shard = outcome.GetResult() + .GetStreamDescription() + .GetShards()[0] + .GetShardId(); + sequence = outcome.GetResult() + .GetStreamDescription() + .GetShards()[0] + .GetSequenceNumberRange() + .GetStartingSequenceNumber(); + } else { + for (const auto& entry : + outcome.GetResult().GetStreamDescription().GetShards()) { + if (entry.GetShardId() == dataset()->shard_.c_str()) { + shard = entry.GetShardId(); + sequence = + entry.GetSequenceNumberRange().GetStartingSequenceNumber(); + break; + } + } + if (shard == "") { + return errors::InvalidArgument("no shard ", dataset()->shard_, + " in stream ", dataset()->stream_); + } + } + + Aws::Kinesis::Model::GetShardIteratorRequest iterator_request; + auto iterator_outcome = client_->GetShardIterator( + iterator_request.WithStreamName(dataset()->stream_.c_str()) + .WithShardId(shard) + .WithShardIteratorType( + Aws::Kinesis::Model::ShardIteratorType::AT_SEQUENCE_NUMBER) + .WithStartingSequenceNumber(sequence)); + if (!iterator_outcome.IsSuccess()) { + return errors::Unknown(iterator_outcome.GetError().GetExceptionName(), + ": ", + iterator_outcome.GetError().GetMessage()); + } + iterator_ = iterator_outcome.GetResult().GetShardIterator(); + return Status::OK(); + } + + mutex mu_; + Aws::String iterator_ GUARDED_BY(mu_); + std::unique_ptr + client_ GUARDED_BY(mu_); + }; + + const std::string stream_; + const std::string shard_; + const bool read_indefinitely_; + const int64 interval_; + }; +}; + +REGISTER_KERNEL_BUILDER(Name("KinesisDataset").Device(DEVICE_CPU), + KinesisDatasetOp); + +} // namespace tensorflow diff --git a/tensorflow/contrib/kinesis/ops/dataset_ops.cc b/tensorflow/contrib/kinesis/ops/dataset_ops.cc new file mode 100644 index 0000000000000000000000000000000000000000..54204513cf22519ecfb5fa45748250ee0f4aac7a --- /dev/null +++ b/tensorflow/contrib/kinesis/ops/dataset_ops.cc @@ -0,0 +1,42 @@ +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#include "tensorflow/core/framework/common_shape_fns.h" +#include "tensorflow/core/framework/op.h" +#include "tensorflow/core/framework/shape_inference.h" + +namespace tensorflow { + +REGISTER_OP("KinesisDataset") + .Input("stream: string") + .Input("shard: string") + .Input("read_indefinitely: bool") + .Input("interval: int64") + .Output("handle: variant") + .SetIsStateful() + .SetShapeFn(shape_inference::ScalarShape) + .Doc(R"doc( +Creates a dataset that emits the messages of one or more Kinesis topics. + +stream: A `tf.string` tensor containing the name of the stream. +shard: A `tf.string` tensor containing the id of the shard. +read_indefinitely: If `True`, the Kinesis dataset will keep retry + again on `EOF` after the `interval` period. If `False`, then + the dataset will stop on `EOF`. The default value is `True`. +interval: The interval for the Kinesis Client to wait before + it tries to get records again (in millisecond). +)doc"); + +} // namespace tensorflow diff --git a/tensorflow/contrib/kinesis/python/kernel_tests/kinesis_test.py b/tensorflow/contrib/kinesis/python/kernel_tests/kinesis_test.py new file mode 100644 index 0000000000000000000000000000000000000000..7289b45c50fa92455b4c317b8a039ca414fa585e --- /dev/null +++ b/tensorflow/contrib/kinesis/python/kernel_tests/kinesis_test.py @@ -0,0 +1,139 @@ +# Copyright 2018 The TensorFlow Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); you may not +# use this file except in compliance with the License. You may obtain a copy of +# the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, WITHOUT +# WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the +# License for the specific language governing permissions and limitations under +# the License. +# ============================================================================== +"""Tests for KinesisDataset. +NOTE: boto3 is needed and the test has to be invoked manually: +``` +$ bazel test -s --verbose_failures --config=opt \ + --action_env=AWS_ACCESS_KEY_ID=XXXXXX \ + --action_env=AWS_SECRET_ACCESS_KEY=XXXXXX \ + //tensorflow/contrib/kinesis:kinesis_test +``` +""" + +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + +import boto3 + +from tensorflow.contrib.kinesis.python.ops import kinesis_dataset_ops +from tensorflow.python.data.ops import iterator_ops +from tensorflow.python.framework import dtypes +from tensorflow.python.framework import errors +from tensorflow.python.ops import array_ops +from tensorflow.python.platform import test + + +class KinesisDatasetTest(test.TestCase): + + def testKinesisDatasetOneShard(self): + client = boto3.client('kinesis', region_name='us-east-1') + + # Setup the Kinesis with 1 shard. + stream_name = "tf_kinesis_test_1" + client.create_stream(StreamName=stream_name, ShardCount=1) + # Wait until stream exists, default is 10 * 18 seconds. + client.get_waiter('stream_exists').wait(StreamName=stream_name) + for i in range(10): + data = "D" + str(i) + client.put_record( + StreamName=stream_name, Data=data, PartitionKey="TensorFlow" + str(i)) + + stream = array_ops.placeholder(dtypes.string, shape=[]) + num_epochs = array_ops.placeholder(dtypes.int64, shape=[]) + batch_size = array_ops.placeholder(dtypes.int64, shape=[]) + + repeat_dataset = kinesis_dataset_ops.KinesisDataset( + stream, read_indefinitely=False).repeat(num_epochs) + batch_dataset = repeat_dataset.batch(batch_size) + + iterator = iterator_ops.Iterator.from_structure(batch_dataset.output_types) + init_op = iterator.make_initializer(repeat_dataset) + init_batch_op = iterator.make_initializer(batch_dataset) + get_next = iterator.get_next() + + with self.test_session() as sess: + # Basic test: read from shard 0 of stream 1. + sess.run(init_op, feed_dict={stream: stream_name, num_epochs: 1}) + for i in range(10): + self.assertEqual("D" + str(i), sess.run(get_next)) + with self.assertRaises(errors.OutOfRangeError): + sess.run(get_next) + + client.delete_stream(StreamName=stream_name) + # Wait until stream deleted, default is 10 * 18 seconds. + client.get_waiter('stream_not_exists').wait(StreamName=stream_name) + + def testKinesisDatasetTwoShards(self): + client = boto3.client('kinesis', region_name='us-east-1') + + # Setup the Kinesis with 2 shards. + stream_name = "tf_kinesis_test_2" + client.create_stream(StreamName=stream_name, ShardCount=2) + # Wait until stream exists, default is 10 * 18 seconds. + client.get_waiter('stream_exists').wait(StreamName=stream_name) + + for i in range(10): + data = "D" + str(i) + client.put_record( + StreamName=stream_name, Data=data, PartitionKey="TensorFlow" + str(i)) + response = client.describe_stream(StreamName=stream_name) + shard_id_0 = response["StreamDescription"]["Shards"][0]["ShardId"] + shard_id_1 = response["StreamDescription"]["Shards"][1]["ShardId"] + + stream = array_ops.placeholder(dtypes.string, shape=[]) + shard = array_ops.placeholder(dtypes.string, shape=[]) + num_epochs = array_ops.placeholder(dtypes.int64, shape=[]) + batch_size = array_ops.placeholder(dtypes.int64, shape=[]) + + repeat_dataset = kinesis_dataset_ops.KinesisDataset( + stream, shard, read_indefinitely=False).repeat(num_epochs) + batch_dataset = repeat_dataset.batch(batch_size) + + iterator = iterator_ops.Iterator.from_structure(batch_dataset.output_types) + init_op = iterator.make_initializer(repeat_dataset) + init_batch_op = iterator.make_initializer(batch_dataset) + get_next = iterator.get_next() + + data = list() + with self.test_session() as sess: + # Basic test: read from shard 0 of stream 2. + sess.run( + init_op, feed_dict={ + stream: stream_name, shard: shard_id_0, num_epochs: 1}) + with self.assertRaises(errors.OutOfRangeError): + # Use range(11) to guarantee the OutOfRangeError. + for i in range(11): + data.append(sess.run(get_next)) + + # Basic test: read from shard 1 of stream 2. + sess.run( + init_op, feed_dict={ + stream: stream_name, shard: shard_id_1, num_epochs: 1}) + with self.assertRaises(errors.OutOfRangeError): + # Use range(11) to guarantee the OutOfRangeError. + for i in range(11): + data.append(sess.run(get_next)) + + data.sort() + self.assertEqual(data, ["D" + str(i) for i in range(10)]) + + client.delete_stream(StreamName=stream_name) + # Wait until stream deleted, default is 10 * 18 seconds. + client.get_waiter('stream_not_exists').wait(StreamName=stream_name) + + +if __name__ == "__main__": + test.main() diff --git a/tensorflow/contrib/kinesis/python/ops/kinesis_dataset_ops.py b/tensorflow/contrib/kinesis/python/ops/kinesis_dataset_ops.py new file mode 100644 index 0000000000000000000000000000000000000000..ca2df95ba4f20ec5fa58ff13530096e6e065f4fe --- /dev/null +++ b/tensorflow/contrib/kinesis/python/ops/kinesis_dataset_ops.py @@ -0,0 +1,96 @@ +# Copyright 2018 The TensorFlow Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================== +"""Kinesis Dataset.""" +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + +from tensorflow.contrib.kinesis.python.ops import kinesis_op_loader # pylint: disable=unused-import +from tensorflow.contrib.kinesis.python.ops import gen_dataset_ops +from tensorflow.python.data.ops.dataset_ops import Dataset +from tensorflow.python.framework import dtypes +from tensorflow.python.framework import ops +from tensorflow.python.framework import tensor_shape + + +class KinesisDataset(Dataset): + """A Kinesis Dataset that consumes the message. + + Kinesis is a managed service provided by AWS for data streaming. + This dataset reads messages from Kinesis with each message presented + as a `tf.string`. + + For example, we can construct and use the KinesisDataset as follows: + ```python + dataset = tf.contrib.kinesis.KinesisDataset( + "kinesis_stream_name", read_indefinitely=False) + next = dataset.make_one_shot_iterator().get_next() + with tf.Session() as sess: + while True: + try: + print(sess.run(nxt)) + except tf.errors.OutOfRangeError: + break + ``` + + Since Kinesis is a data streaming service, data may not be available + at the time it is being read. The argument `read_indefinitely` is + used to control the behavior in this situation. If `read_indefinitely` + is `True`, then `KinesisDataset` will keep retrying to retrieve data + from the stream. If `read_indefinitely` is `False`, an `OutOfRangeError` + is returned immediately instead. + """ + + def __init__(self, + stream, + shard="", + read_indefinitely=True, + interval=100000): + """Create a KinesisDataset. + + Args: + stream: A `tf.string` tensor containing the name of the stream. + shard: A `tf.string` tensor containing the id of the shard. + read_indefinitely: If `True`, the Kinesis dataset will keep retry + again on `EOF` after the `interval` period. If `False`, then + the dataset will stop on `EOF`. The default value is `True`. + interval: The interval for the Kinesis Client to wait before + it tries to get records again (in millisecond). + """ + super(KinesisDataset, self).__init__() + self._stream = ops.convert_to_tensor( + stream, dtype=dtypes.string, name="stream") + self._shard = ops.convert_to_tensor( + shard, dtype=dtypes.string, name="shard") + self._read_indefinitely = ops.convert_to_tensor( + read_indefinitely, dtype=dtypes.bool, name="read_indefinitely") + self._interval = ops.convert_to_tensor( + interval, dtype=dtypes.int64, name="interval") + + def _as_variant_tensor(self): + return gen_dataset_ops.kinesis_dataset( + self._stream, self._shard, self._read_indefinitely, self._interval) + + @property + def output_classes(self): + return ops.Tensor + + @property + def output_shapes(self): + return tensor_shape.scalar() + + @property + def output_types(self): + return dtypes.string diff --git a/tensorflow/contrib/kinesis/python/ops/kinesis_op_loader.py b/tensorflow/contrib/kinesis/python/ops/kinesis_op_loader.py new file mode 100644 index 0000000000000000000000000000000000000000..c9ce9f3646200a777cdbdf34b37626154ca730bb --- /dev/null +++ b/tensorflow/contrib/kinesis/python/ops/kinesis_op_loader.py @@ -0,0 +1,24 @@ +# Copyright 2018 The TensorFlow Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================== +"""Python helper for loading kinesis ops and kernels.""" +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + +from tensorflow.contrib.util import loader +from tensorflow.python.platform import resource_loader + +_dataset_ops = loader.load_op_library( + resource_loader.get_path_to_datafile("../../_dataset_ops.so")) diff --git a/tensorflow/contrib/layers/python/layers/embedding_ops_test.py b/tensorflow/contrib/layers/python/layers/embedding_ops_test.py index dd2395f8c9748dadbecfe47df5511874d5f848ea..7ede193029d2d95fa4953b4c417a1e86ebb4a42e 100644 --- a/tensorflow/contrib/layers/python/layers/embedding_ops_test.py +++ b/tensorflow/contrib/layers/python/layers/embedding_ops_test.py @@ -21,7 +21,6 @@ from __future__ import print_function import itertools import math -import sys import numpy as np diff --git a/tensorflow/contrib/layers/python/layers/rev_block_lib.py b/tensorflow/contrib/layers/python/layers/rev_block_lib.py index 0e35b1aa8bf682c1b4f7e8d974d3e8fad69e33cb..dad3da3748097c26e07b4abe0495f62a18aad369 100644 --- a/tensorflow/contrib/layers/python/layers/rev_block_lib.py +++ b/tensorflow/contrib/layers/python/layers/rev_block_lib.py @@ -514,15 +514,15 @@ def _recompute_grad(fn, args, use_data_dep=_USE_DEFAULT, tupleize_grads=False): original_vars = set(tape.watched_variables()) # Backward pass - def grad_fn(*output_grads, **kwargs): + def _grad_fn(output_grads, variables=None): """Recompute outputs for gradient computation.""" - variables = [] + variables = variables or [] if original_vars: - variables = kwargs["variables"] - if set(variables) != original_vars: - raise ValueError(_WRONG_VARS_ERR) - del kwargs - inputs = list(args) + assert variables, ("Fn created variables but the variables were not " + "passed to the gradient fn.") + if set(variables) != original_vars: + raise ValueError(_WRONG_VARS_ERR) + inputs = [array_ops.identity(x) for x in list(args)] # Recompute outputs with framework_ops.control_dependencies(output_grads): if use_data_dep_: @@ -538,7 +538,7 @@ def _recompute_grad(fn, args, use_data_dep=_USE_DEFAULT, tupleize_grads=False): if original_vars != recompute_vars: raise ValueError(_WRONG_VARS_ERR) - if not (isinstance(outputs, list) or isinstance(outputs, tuple)): + if not isinstance(outputs, (list, tuple)): outputs = [outputs] outputs = list(outputs) grads = gradients_impl.gradients(outputs, inputs + variables, @@ -554,6 +554,16 @@ def _recompute_grad(fn, args, use_data_dep=_USE_DEFAULT, tupleize_grads=False): grad_vars = grads[len(inputs):] return grad_inputs, grad_vars + # custom_gradient inspects the signature of the function to determine + # whether the user expects variables passed in the grad_fn. If the function + # created variables, the grad_fn should accept the "variables" kwarg. + if original_vars: + def grad_fn(*output_grads, **kwargs): + return _grad_fn(output_grads, kwargs["variables"]) + else: + def grad_fn(*output_grads): + return _grad_fn(output_grads) + return outputs, grad_fn return fn_with_recompute(*args) diff --git a/tensorflow/contrib/layers/python/layers/rev_block_lib_test.py b/tensorflow/contrib/layers/python/layers/rev_block_lib_test.py index bc09ba8d439808c1582f207a99504012afcf33a6..d5971fb9d8e2fbc1e14fd24fc79e7981a284a418 100644 --- a/tensorflow/contrib/layers/python/layers/rev_block_lib_test.py +++ b/tensorflow/contrib/layers/python/layers/rev_block_lib_test.py @@ -372,6 +372,26 @@ class RecomputeTest(test.TestCase): self.assertEqual(2, len(update_ops)) self.assertEqual([False, True], kwarg_values) + def testWithoutVariables(self): + + def concat_n(layer_list, num_inputs): + return math_ops.reduce_sum( + array_ops.concat([x for x in layer_list[-num_inputs:]], axis=-1), + axis=1, keepdims=True) + + @rev_block_lib.recompute_grad + def concat_n_wrap(*args): + return concat_n(args, 3) + + # DenseNet-style layers + layer_list = [random_ops.random_uniform((4, 8))] + for _ in range(5): + layer_list.append(math_ops.sqrt(concat_n_wrap(*layer_list))) + + grads = gradients_impl.gradients(layer_list[-1], layer_list[0]) + with self.test_session() as sess: + sess.run(grads) + if __name__ == "__main__": test.main() diff --git a/tensorflow/contrib/legacy_seq2seq/python/ops/seq2seq.py b/tensorflow/contrib/legacy_seq2seq/python/ops/seq2seq.py index 5e7b422e3cc368a22eb94ed470297ae78293c4eb..e74244720896a835174f54bb97049c1d9b1c92f8 100644 --- a/tensorflow/contrib/legacy_seq2seq/python/ops/seq2seq.py +++ b/tensorflow/contrib/legacy_seq2seq/python/ops/seq2seq.py @@ -625,11 +625,13 @@ def attention_decoder(decoder_inputs, v = [] attention_vec_size = attn_size # Size of query vectors for attention. for a in xrange(num_heads): - k = variable_scope.get_variable("AttnW_%d" % a, - [1, 1, attn_size, attention_vec_size]) + k = variable_scope.get_variable( + "AttnW_%d" % a, [1, 1, attn_size, attention_vec_size], + dtype=dtype) hidden_features.append(nn_ops.conv2d(hidden, k, [1, 1, 1, 1], "SAME")) v.append( - variable_scope.get_variable("AttnV_%d" % a, [attention_vec_size])) + variable_scope.get_variable( + "AttnV_%d" % a, [attention_vec_size], dtype=dtype)) state = initial_state @@ -647,11 +649,13 @@ def attention_decoder(decoder_inputs, with variable_scope.variable_scope("Attention_%d" % a): y = Linear(query, attention_vec_size, True)(query) y = array_ops.reshape(y, [-1, 1, 1, attention_vec_size]) + y = math_ops.cast(y, dtype) # Attention mask is a softmax of v^T * tanh(...). s = math_ops.reduce_sum(v[a] * math_ops.tanh(hidden_features[a] + y), [2, 3]) - a = nn_ops.softmax(s) + a = nn_ops.softmax(math_ops.cast(s, dtype=dtypes.float32)) # Now calculate the attention-weighted vector d. + a = math_ops.cast(a, dtype) d = math_ops.reduce_sum( array_ops.reshape(a, [-1, attn_length, 1, 1]) * hidden, [1, 2]) ds.append(array_ops.reshape(d, [-1, attn_size])) @@ -681,6 +685,7 @@ def attention_decoder(decoder_inputs, raise ValueError("Could not infer input size from input: %s" % inp.name) inputs = [inp] + attns + inputs = [math_ops.cast(e, dtype) for e in inputs] x = Linear(inputs, input_size, True)(inputs) # Run the RNN. cell_output, state = cell(x, state) @@ -693,6 +698,7 @@ def attention_decoder(decoder_inputs, attns = attention(state) with variable_scope.variable_scope("AttnOutputProjection"): + cell_output = math_ops.cast(cell_output, dtype) inputs = [cell_output] + attns output = Linear(inputs, output_size, True)(inputs) if loop_function is not None: diff --git a/tensorflow/contrib/linear_optimizer/BUILD b/tensorflow/contrib/linear_optimizer/BUILD index 5b89c6cef9fa9fdef7c26ddee1efa03f3056d881..fe0ba19fcbe90edbeb1445e1fea77c36cf3ba170 100644 --- a/tensorflow/contrib/linear_optimizer/BUILD +++ b/tensorflow/contrib/linear_optimizer/BUILD @@ -41,6 +41,7 @@ py_test( size = "medium", srcs = ["python/kernel_tests/sdca_ops_test.py"], srcs_version = "PY2AND3", + tags = ["no_windows_gpu"], deps = [ ":sdca_ops_py", ":sparse_feature_column_py", diff --git a/tensorflow/contrib/lite/BUILD b/tensorflow/contrib/lite/BUILD index 8c17c65fcc0dbd58e2b3e9042a983e400cd6c2b9..b95d4d0fce312b351abc714bbda2438240037874 100644 --- a/tensorflow/contrib/lite/BUILD +++ b/tensorflow/contrib/lite/BUILD @@ -128,6 +128,7 @@ cc_library( hdrs = [ "allocation.h", "context.h", + "context_util.h", "error_reporter.h", "graph_info.h", "interpreter.h", @@ -145,6 +146,7 @@ cc_library( ":memory_planner", ":schema_fbs_version", ":simple_memory_arena", + ":string", ":util", "//tensorflow/contrib/lite/kernels:eigen_support", "//tensorflow/contrib/lite/kernels:gemm_support", diff --git a/tensorflow/contrib/lite/Makefile b/tensorflow/contrib/lite/Makefile index 2b6997146e1e5a3873ed0f94a9221b34bed7621d..a616138d3321d43f66a2b430f7df609a13b9caf6 100644 --- a/tensorflow/contrib/lite/Makefile +++ b/tensorflow/contrib/lite/Makefile @@ -17,7 +17,29 @@ else endif endif -ARCH := $(shell if [[ $(shell uname -m) =~ i[345678]86 ]]; then echo x86_32; else echo $(shell uname -m); fi) +HOST_ARCH := $(shell if [[ $(shell uname -m) =~ i[345678]86 ]]; then echo x86_32; else echo $(shell uname -m); fi) + +# Self-hosting +TARGET_ARCH := ${HOST_ARCH} + +# Cross compiling +ifeq ($(CROSS),rpi) + TARGET_ARCH := armv7l + TARGET_TOOLCHAIN_PREFIX := arm-linux-gnueabihf- +endif + +ifeq ($(CROSS),riscv) + TARGET_ARCH := riscv + TARGET_TOOLCHAIN_PREFIX := riscv32-unknown-elf- +endif +ifeq ($(CROSS),stm32f7) + TARGET_ARCH := armf7 + TARGET_TOOLCHAIN_PREFIX := arm-none-eabi- +endif +ifeq ($(CROSS),stm32f1) + TARGET_ARCH := armm1 + TARGET_TOOLCHAIN_PREFIX := arm-none-eabi- +endif # Where compiled objects are stored. OBJDIR := $(MAKEFILE_DIR)/gen/obj/ @@ -25,11 +47,46 @@ BINDIR := $(MAKEFILE_DIR)/gen/bin/ LIBDIR := $(MAKEFILE_DIR)/gen/lib/ GENDIR := $(MAKEFILE_DIR)/gen/obj/ +LIBS := +ifeq ($(TARGET_ARCH),x86_64) + CXXFLAGS += -fPIC -DGEMMLOWP_ALLOW_SLOW_SCALAR_FALLBACK -pthread # -msse4.2 +endif + +ifeq ($(TARGET_ARCH),armv7l) + CXXFLAGS += -mfpu=neon -pthread -fPIC + LIBS += -ldl +endif + +ifeq ($(TARGET_ARCH),riscv) +# CXXFLAGS += -march=gap8 + CXXFLAGS += -DTFLITE_MCU + LIBS += -ldl + BUILD_TYPE := micro +endif + +ifeq ($(TARGET_ARCH),armf7) + CXXFLAGS += -DGEMMLOWP_ALLOW_SLOW_SCALAR_FALLBACK -DTFLITE_MCU + CXXFLAGS += -fno-rtti -fmessage-length=0 -fno-exceptions -fno-builtin -ffunction-sections -fdata-sections + CXXFLAGS += -funsigned-char -MMD + CXXFLAGS += -mcpu=cortex-m7 -mthumb -mfpu=fpv5-sp-d16 -mfloat-abi=softfp + CXXFLAGS += '-std=gnu++11' '-fno-rtti' '-Wvla' '-c' '-Wall' '-Wextra' '-Wno-unused-parameter' '-Wno-missing-field-initializers' '-fmessage-length=0' '-fno-exceptions' '-fno-builtin' '-ffunction-sections' '-fdata-sections' '-funsigned-char' '-MMD' '-fno-delete-null-pointer-checks' '-fomit-frame-pointer' '-Os' + LIBS += -ldl + BUILD_TYPE := micro +endif +ifeq ($(TARGET_ARCH),armm1) + CXXFLAGS += -DGEMMLOWP_ALLOW_SLOW_SCALAR_FALLBACK -mcpu=cortex-m1 -mthumb -DTFLITE_MCU + CXXFLAGS += -fno-rtti -fmessage-length=0 -fno-exceptions -fno-builtin -ffunction-sections -fdata-sections + CXXFLAGS += -funsigned-char -MMD + LIBS += -ldl +endif + # Settings for the host compiler. -CXX := $(CC_PREFIX)gcc -CXXFLAGS := --std=c++11 -O3 -DNDEBUG -CC := $(CC_PREFIX)gcc -CCFLAGS := -O3 -DNDEBUG +CXX := $(CC_PREFIX) ${TARGET_TOOLCHAIN_PREFIX}g++ +CXXFLAGS += --std=c++11 -O3 -DNDEBUG +CCFLAGS := ${CXXFLAGS} +CC := $(CC_PREFIX) ${TARGET_TOOLCHAIN_PREFIX}gcc +AR := $(CC_PREFIX) ${TARGET_TOOLCHAIN_PREFIX}ar +CFLAGS := LDOPTS := LDOPTS += -L/usr/local/lib ARFLAGS := -r @@ -48,7 +105,7 @@ INCLUDES := \ # override local versions in the source tree. INCLUDES += -I/usr/local/include -LIBS := \ +LIBS += \ -lstdc++ \ -lpthread \ -lm \ @@ -92,18 +149,21 @@ PROFILE_SUMMARIZER_SRCS := \ CORE_CC_ALL_SRCS := \ $(wildcard tensorflow/contrib/lite/*.cc) \ +$(wildcard tensorflow/contrib/lite/*.c) +ifneq ($(BUILD_TYPE),micro) +CORE_CC_ALL_SRCS += \ $(wildcard tensorflow/contrib/lite/kernels/*.cc) \ $(wildcard tensorflow/contrib/lite/kernels/internal/*.cc) \ $(wildcard tensorflow/contrib/lite/kernels/internal/optimized/*.cc) \ $(wildcard tensorflow/contrib/lite/kernels/internal/reference/*.cc) \ $(PROFILER_SRCS) \ -$(wildcard tensorflow/contrib/lite/*.c) \ $(wildcard tensorflow/contrib/lite/kernels/*.c) \ $(wildcard tensorflow/contrib/lite/kernels/internal/*.c) \ $(wildcard tensorflow/contrib/lite/kernels/internal/optimized/*.c) \ $(wildcard tensorflow/contrib/lite/kernels/internal/reference/*.c) \ $(wildcard tensorflow/contrib/lite/downloads/farmhash/src/farmhash.cc) \ $(wildcard tensorflow/contrib/lite/downloads/fft2d/fftsg.c) +endif # Remove any duplicates. CORE_CC_ALL_SRCS := $(sort $(CORE_CC_ALL_SRCS)) CORE_CC_EXCLUDE_SRCS := \ @@ -113,6 +173,11 @@ $(wildcard tensorflow/contrib/lite/*/*/*test.cc) \ $(wildcard tensorflow/contrib/lite/*/*/*/*test.cc) \ $(wildcard tensorflow/contrib/lite/kernels/test_util.cc) \ $(MINIMAL_SRCS) +ifeq ($(BUILD_TYPE),micro) +CORE_CC_EXCLUDE_SRCS += \ +tensorflow/contrib/lite/model.cc \ +tensorflow/contrib/lite/nnapi_delegate.cc +endif # Filter out all the excluded files. TF_LITE_CC_SRCS := $(filter-out $(CORE_CC_EXCLUDE_SRCS), $(CORE_CC_ALL_SRCS)) # File names of the intermediate files target compilation generates. @@ -120,7 +185,6 @@ TF_LITE_CC_OBJS := $(addprefix $(OBJDIR), \ $(patsubst %.cc,%.o,$(patsubst %.c,%.o,$(TF_LITE_CC_SRCS)))) LIB_OBJS := $(TF_LITE_CC_OBJS) - # Benchmark sources BENCHMARK_SRCS_DIR := tensorflow/contrib/lite/tools/benchmark BENCHMARK_ALL_SRCS := $(TFLITE_CC_SRCS) \ @@ -146,6 +210,9 @@ $(OBJDIR)%.o: %.c # The target that's compiled if there's no command-line arguments. all: $(LIB_PATH) $(MINIMAL_PATH) $(BENCHMARK_BINARY) +# The target that's compiled for micro-controllers +micro: $(LIB_PATH) + # Gathers together all the objects we've compiled into a single '.a' archive. $(LIB_PATH): $(LIB_OBJS) @mkdir -p $(dir $@) diff --git a/tensorflow/contrib/lite/allocation.cc b/tensorflow/contrib/lite/allocation.cc index a4772731ecda92431c412672610a39c188dabf27..c42622ff02fc2837b61b35f19e834276c0518d1e 100644 --- a/tensorflow/contrib/lite/allocation.cc +++ b/tensorflow/contrib/lite/allocation.cc @@ -14,7 +14,9 @@ limitations under the License. ==============================================================================*/ #include +#ifndef TFLITE_MCU #include +#endif #include #include #include @@ -27,10 +29,13 @@ limitations under the License. #include "tensorflow/contrib/lite/allocation.h" #include "tensorflow/contrib/lite/context.h" #include "tensorflow/contrib/lite/error_reporter.h" +#ifndef TFLITE_MCU #include "tensorflow/contrib/lite/nnapi_delegate.h" +#endif namespace tflite { +#ifndef TFLITE_MCU MMAPAllocation::MMAPAllocation(const char* filename, ErrorReporter* error_reporter) : Allocation(error_reporter), mmapped_buffer_(MAP_FAILED) { @@ -111,6 +116,7 @@ MemoryAllocation::MemoryAllocation(const void* ptr, size_t num_bytes, buffer_ = ptr; buffer_size_bytes_ = num_bytes; } +#endif MemoryAllocation::~MemoryAllocation() {} diff --git a/tensorflow/contrib/lite/allocation.h b/tensorflow/contrib/lite/allocation.h index 68aee2e64473320c461ec8b3f194904e7b8da43c..827ea86503f910714971e2b138295b9a5809dfd5 100644 --- a/tensorflow/contrib/lite/allocation.h +++ b/tensorflow/contrib/lite/allocation.h @@ -23,6 +23,7 @@ limitations under the License. #include "tensorflow/contrib/lite/context.h" #include "tensorflow/contrib/lite/error_reporter.h" #include "tensorflow/contrib/lite/simple_memory_arena.h" +#include "tensorflow/contrib/lite/string.h" namespace tflite { diff --git a/tensorflow/contrib/lite/arena_planner.cc b/tensorflow/contrib/lite/arena_planner.cc index 22be64d6ff649b4bff45a5e5680984d688a8cf38..4257e754ad5c30e17ec8ba8d5c6e69b5c5bcd728 100644 --- a/tensorflow/contrib/lite/arena_planner.cc +++ b/tensorflow/contrib/lite/arena_planner.cc @@ -35,12 +35,13 @@ struct AllocationInfo { }; ArenaPlanner::ArenaPlanner(TfLiteContext* context, - std::unique_ptr graph_info) + std::unique_ptr graph_info, + bool preserve_inputs) : context_(context), graph_info_(std::move(graph_info)), arena_(kDefaultArenaAlignment), - persistent_arena_(kDefaultArenaAlignment) {} - + persistent_arena_(kDefaultArenaAlignment), + preserve_inputs_(preserve_inputs) {} ArenaPlanner::~ArenaPlanner() {} int64_t ArenaPlanner::BasePointer(TfLiteAllocationType type) { @@ -112,9 +113,13 @@ TfLiteStatus ArenaPlanner::PlanAllocations() { refcounts[tensor_index]++; } - // Queue all graph inputs for allocation. + // Queue all graph inputs for allocation. If preserve_inputs_ is true, make + // sure they never be overwritten. for (int tensor_index : graph_info_->inputs()) { if (tensor_index != kOptionalTensor) { + if (preserve_inputs_) { + refcounts[tensor_index]++; + } TF_LITE_ENSURE_STATUS(allocate(0, tensor_index)); } } diff --git a/tensorflow/contrib/lite/arena_planner.h b/tensorflow/contrib/lite/arena_planner.h index e9d0fbc5a9b5aec06e28da8757466b25f40da2f5..1d84950e91bc48fd1c1a7e5b2d9063e20dea0718 100644 --- a/tensorflow/contrib/lite/arena_planner.h +++ b/tensorflow/contrib/lite/arena_planner.h @@ -43,8 +43,11 @@ struct AllocationInfo; class ArenaPlanner : public MemoryPlanner { public: // Ownership of 'context' is not taken and it must remain util the - // ArenaPlanner is destroyed. - ArenaPlanner(TfLiteContext* context, std::unique_ptr graph_info); + // ArenaPlanner is destroyed. If 'preserve_inputs' is true the inputs to the + // graph will not share memory with any other tensor, effectively preserving + // them until the end of inference. + ArenaPlanner(TfLiteContext* context, std::unique_ptr graph_info, + bool preserve_inputs); ~ArenaPlanner() override; ArenaPlanner(const ArenaPlanner&) = delete; ArenaPlanner& operator=(const ArenaPlanner&) = delete; @@ -100,6 +103,8 @@ class ArenaPlanner : public MemoryPlanner { // Raw memory buffer that is allocated for persistent tensors that are // declared as kTfLiteArenaRwPersistent. SimpleMemoryArena persistent_arena_; + + bool preserve_inputs_; }; } // namespace tflite diff --git a/tensorflow/contrib/lite/arena_planner_test.cc b/tensorflow/contrib/lite/arena_planner_test.cc index f0fd35216f645df59b03340e00daca9322721b1b..f5bd1932f976f5c7d0f0d14bbaf9ca3807dfd3b0 100644 --- a/tensorflow/contrib/lite/arena_planner_test.cc +++ b/tensorflow/contrib/lite/arena_planner_test.cc @@ -151,11 +151,12 @@ void ReportError(TfLiteContext* context, const char* format, ...) { class ArenaPlannerTest : public ::testing::Test { protected: - void SetGraph(TestGraph* graph) { + void SetGraph(TestGraph* graph, bool preserve_inputs = false) { graph_ = graph; context_.ReportError = ReportError; planner_.reset(new ArenaPlanner( - &context_, std::unique_ptr(new TestGraphInfo(graph)))); + &context_, std::unique_ptr(new TestGraphInfo(graph)), + preserve_inputs)); CHECK(planner_->ResetAllocations() == kTfLiteOk); CHECK(planner_->PlanAllocations() == kTfLiteOk); } @@ -243,6 +244,30 @@ TEST_F(ArenaPlannerTest, SimpleGraph) { EXPECT_EQ(GetOffset(3), 0); } +TEST_F(ArenaPlannerTest, SimpleGraphInputsPreserved) { + TestGraph graph({0, 1}, + { + /* in, out, tmp */ + {{0, 1}, {2}, {}}, // First op + {{2, 0}, {4, 5}, {}}, // Second op + {{4, 5}, {3}, {}} // Third op + }, + {3}); + SetGraph(&graph, /*preserve_inputs=*/true); + Execute(0, 10); + + // Alloc(+) and dealloc(-) order: +0 +1 +2 +4 +5 -2 +3 -4 -5 + EXPECT_EQ(GetOffset(0), 0); + EXPECT_EQ(GetOffset(1), GetOffsetAfter(0)); + EXPECT_EQ(GetOffset(2), GetOffsetAfter(1)); + EXPECT_EQ(GetOffset(4), GetOffsetAfter(2)); + EXPECT_EQ(GetOffset(5), GetOffsetAfter(4)); + // Because we are keeping the inputs alive until the end (due to + // preserve_inputs=true), the output tensor will not be able to use that + // space. It will end up using the same are as tensor #2. + EXPECT_EQ(GetOffset(3), GetOffsetAfter(1)); +} + TEST_F(ArenaPlannerTest, SimpleGraphWithTemporary) { TestGraph graph({0, 1}, { diff --git a/tensorflow/contrib/lite/build_def.bzl b/tensorflow/contrib/lite/build_def.bzl index 81883ba1fd5a2b0bde62b49d67d50dc5a3e281a0..b735d08b4b27be50c96fd5e93c2ae2e9676c1013 100644 --- a/tensorflow/contrib/lite/build_def.bzl +++ b/tensorflow/contrib/lite/build_def.bzl @@ -195,7 +195,7 @@ def json_to_tflite(name, src, out): def generated_test_models(): return [ "add", - "arg_max", + "arg_min_max", "avg_pool", "batch_to_space_nd", "concat", @@ -232,7 +232,8 @@ def generated_test_models(): "not_equal", "pad", "padv2", - # "prelu", + "prelu", + "pow", "relu", "relu1", "relu6", @@ -256,7 +257,7 @@ def generated_test_models(): "tile", "topk", "transpose", - "transpose_conv", + #"transpose_conv", # disabled due to b/111213074 "where", ] diff --git a/tensorflow/contrib/lite/builtin_op_data.h b/tensorflow/contrib/lite/builtin_op_data.h index 1b1b8b2985afda669c950eb1284d99d903e95455..a24aaad7dda56a2ece059cced4a3ca2cf13c26b9 100644 --- a/tensorflow/contrib/lite/builtin_op_data.h +++ b/tensorflow/contrib/lite/builtin_op_data.h @@ -92,8 +92,17 @@ typedef struct { TfLiteFusedActivation activation; } TfLiteSequenceRNNParams; +typedef enum { + kTfLiteFullyConnectedWeightsFormatDefault = 0, + kTfLiteFullyConnectedWeightsFormatShuffled4x16Int8 = 1, +} TfLiteFullyConnectedWeightsFormat; + typedef struct { + // Parameters for FullyConnected version 1 or above. TfLiteFusedActivation activation; + + // Parameters for FullyConnected version 2 or above. + TfLiteFullyConnectedWeightsFormat weights_format; } TfLiteFullyConnectedParams; typedef enum { @@ -240,6 +249,10 @@ typedef struct { TfLiteType output_type; } TfLiteArgMaxParams; +typedef struct { + TfLiteType output_type; +} TfLiteArgMinParams; + typedef struct { TfLitePadding padding; int stride_width; @@ -254,6 +267,16 @@ typedef struct { TfLiteType out_type; } TfLiteShapeParams; +typedef struct { + // Parameters supported by version 1: + float min; + float max; + int num_bits; + + // Parameters supported by version 2: + bool narrow_range; +} TfLiteFakeQuantParams; + #ifdef __cplusplus } // extern "C" #endif // __cplusplus diff --git a/tensorflow/contrib/lite/builtin_ops.h b/tensorflow/contrib/lite/builtin_ops.h index 7a78206ebf5f7a5e88e56723e874b9d552df05bd..6bde5d2e6d750a1382fc241d8dec7999cfae8425 100644 --- a/tensorflow/contrib/lite/builtin_ops.h +++ b/tensorflow/contrib/lite/builtin_ops.h @@ -103,6 +103,9 @@ typedef enum { kTfLiteBuiltinSqrt = 75, kTfLiteBuiltinRsqrt = 76, kTfLiteBuiltinShape = 77, + kTfLiteBuiltinPow = 78, + kTfLiteBuiltinArgMin = 79, + kTfLiteBuiltinFakeQuant = 80, } TfLiteBuiltinOperator; #ifdef __cplusplus diff --git a/tensorflow/contrib/lite/context.h b/tensorflow/contrib/lite/context.h index 6434e265b1f4a2e11b102e6ad112a6c8a53f5d93..1ff8843fa78f48fc74b4d7e7d0cc4ae2a0d255af 100644 --- a/tensorflow/contrib/lite/context.h +++ b/tensorflow/contrib/lite/context.h @@ -39,6 +39,26 @@ extern "C" { typedef enum { kTfLiteOk = 0, kTfLiteError = 1 } TfLiteStatus; +// The list of external context types known to TF Lite. This list exists solely +// to avoid conflicts and to ensure ops can share the external contexts they +// need. Access to the external contexts is controled by one of the +// corresponding support files. +typedef enum { + kTfLiteEigenContext = 0, // include eigen_support.h to use. + kTfLiteGemmLowpContext = 1, // include gemm_support.h to use. + kTfLiteMaxExternalContexts = 2 +} TfLiteExternalContextType; + +// An external context is a collection of information unrelated to the TF Lite +// framework, but useful to a subset of the ops. TF Lite knows very little +// about about the actual contexts, but it keeps a list of them, and is able to +// refresh them if configurations like the number of recommended threads +// change. +typedef struct { + TfLiteExternalContextType type; + TfLiteStatus (*Refresh)(struct TfLiteContext* context); +} TfLiteExternalContext; + // Forward declare so GetNode can use this is in Context. typedef struct _TfLiteRegistration TfLiteRegistration; typedef struct _TfLiteDelegate TfLiteDelegate; @@ -139,6 +159,7 @@ typedef enum { kTfLiteString = 5, kTfLiteBool = 6, kTfLiteInt16 = 7, + kTfLiteComplex64 = 8, } TfLiteType; // Parameters for asymmetric quantization. Quantized values can be converted @@ -159,6 +180,7 @@ typedef union { uint8_t* uint8; bool* b; int16_t* i16; + _Complex float* c64; } TfLitePtrUnion; // Memory allocation strategies. kTfLiteMmapRo is for read-only memory-mapped @@ -243,7 +265,8 @@ void TfLiteTensorReset(TfLiteType type, const char* name, TfLiteIntArray* dims, const void* allocation, bool is_variable, TfLiteTensor* tensor); -// Resize the allocated data of a (dynamic) tensor. +// Resize the allocated data of a (dynamic) tensor. Tensors with allocation +// types other than kTfLiteDynamic will be ignored. void TfLiteTensorRealloc(size_t num_bytes, TfLiteTensor* tensor); // A structure representing an instance of a node. @@ -336,10 +359,15 @@ typedef struct TfLiteContext { // eigen. int recommended_num_threads; - // TODO(ahentz): we should create a more general mechanism for this sort of - // library-global objects. - void* gemm_context; - void* eigen_context; + // Access external contexts by type. + // WARNING: This is an experimental interface that is subject to change. + TfLiteExternalContext* (*GetExternalContext)(struct TfLiteContext*, + TfLiteExternalContextType); + // Set the value of a external context. Does not take ownership of the + // pointer. + // WARNING: This is an experimental interface that is subject to change. + void (*SetExternalContext)(struct TfLiteContext*, TfLiteExternalContextType, + TfLiteExternalContext*); } TfLiteContext; typedef struct _TfLiteRegistration { diff --git a/tensorflow/contrib/lite/delegates/eager/BUILD b/tensorflow/contrib/lite/delegates/eager/BUILD new file mode 100644 index 0000000000000000000000000000000000000000..066b1062158ca1506ca1719ab8f027302e621adf --- /dev/null +++ b/tensorflow/contrib/lite/delegates/eager/BUILD @@ -0,0 +1,35 @@ +# +# This is a TF Lite delegate that is powered by TensorFlow's Eager. +# +package(default_visibility = [ + "//visibility:public", +]) + +licenses(["notice"]) # Apache 2.0 + +cc_library( + name = "util", + srcs = ["util.cc"], + hdrs = ["util.h"], + deps = [ + "//tensorflow/contrib/lite:framework", + "//tensorflow/contrib/lite:kernel_api", + "//tensorflow/core:framework", + "//tensorflow/core:lib", + ], +) + +cc_test( + name = "util_test", + size = "small", + srcs = ["util_test.cc"], + tags = [ + "tflite_not_portable", + ], + deps = [ + ":util", + "//tensorflow/contrib/lite/testing:util", + "//tensorflow/core:lib", + "@com_google_googletest//:gtest", + ], +) diff --git a/tensorflow/contrib/lite/delegates/eager/util.cc b/tensorflow/contrib/lite/delegates/eager/util.cc new file mode 100644 index 0000000000000000000000000000000000000000..04a852e515ec4358aecb3fd064d842745354390c --- /dev/null +++ b/tensorflow/contrib/lite/delegates/eager/util.cc @@ -0,0 +1,47 @@ +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#include "tensorflow/contrib/lite/delegates/eager/util.h" + +namespace tflite { + +TfLiteStatus ConvertStatus(TfLiteContext* context, + const tensorflow::Status& status) { + if (!status.ok()) { + context->ReportError(context, "%s", status.error_message().c_str()); + return kTfLiteError; + } + return kTfLiteOk; +} + +TfLiteStatus CopyShape(TfLiteContext* context, const tensorflow::Tensor& src, + TfLiteTensor* tensor) { + int num_dims = src.dims(); + TfLiteIntArray* shape = TfLiteIntArrayCreate(num_dims); + for (int j = 0; j < num_dims; ++j) { + // We need to cast from TensorFlow's int64 to TF Lite's int32. Let's + // make sure there's no overflow. + if (src.dim_size(j) >= std::numeric_limits::max()) { + context->ReportError(context, + "Dimension value in TensorFlow shape is larger than " + "supported by TF Lite"); + TfLiteIntArrayFree(shape); + return kTfLiteError; + } + shape->data[j] = static_cast(src.dim_size(j)); + } + return context->ResizeTensor(context, tensor, shape); +} + +} // namespace tflite diff --git a/tensorflow/contrib/lite/delegates/eager/util.h b/tensorflow/contrib/lite/delegates/eager/util.h new file mode 100644 index 0000000000000000000000000000000000000000..2696ca8d0d31de1f1dd8ddabc71c80217955a896 --- /dev/null +++ b/tensorflow/contrib/lite/delegates/eager/util.h @@ -0,0 +1,35 @@ +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#ifndef TENSORFLOW_CONTRIB_LITE_DELEGATES_EAGER_UTIL_H_ +#define TENSORFLOW_CONTRIB_LITE_DELEGATES_EAGER_UTIL_H_ + +#include "tensorflow/contrib/lite/context.h" +#include "tensorflow/core/framework/tensor.h" +#include "tensorflow/core/lib/core/status.h" + +namespace tflite { + +// Converts a tensorflow:Status into a TfLiteStatus. If the original status +// represented an error, reports it using the given 'context'. +TfLiteStatus ConvertStatus(TfLiteContext* context, + const tensorflow::Status& status); + +// Copies the given shape of the given 'src' into a TF Lite 'tensor'. Logs an +// error and returns kTfLiteError if the shape can't be converted. +TfLiteStatus CopyShape(TfLiteContext* context, const tensorflow::Tensor& src, + TfLiteTensor* tensor); +} // namespace tflite + +#endif // TENSORFLOW_CONTRIB_LITE_DELEGATES_EAGER_UTIL_H_ diff --git a/tensorflow/contrib/lite/delegates/eager/util_test.cc b/tensorflow/contrib/lite/delegates/eager/util_test.cc new file mode 100644 index 0000000000000000000000000000000000000000..563f82dec382720fd3c6593596a97961b9cdfc45 --- /dev/null +++ b/tensorflow/contrib/lite/delegates/eager/util_test.cc @@ -0,0 +1,100 @@ +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#include "tensorflow/contrib/lite/delegates/eager/util.h" + +#include + +#include +#include +#include "tensorflow/contrib/lite/testing/util.h" + +namespace tflite { +namespace { + +using ::testing::ElementsAre; + +struct TestContext : public TfLiteContext { + string error; + std::vector new_size; +}; + +void ReportError(TfLiteContext* context, const char* format, ...) { + TestContext* c = static_cast(context); + const size_t kBufferSize = 1024; + char temp_buffer[kBufferSize]; + + va_list args; + va_start(args, format); + vsnprintf(temp_buffer, kBufferSize, format, args); + va_end(args); + + c->error = temp_buffer; +} + +TfLiteStatus ResizeTensor(TfLiteContext* context, TfLiteTensor* tensor, + TfLiteIntArray* new_size) { + TestContext* c = static_cast(context); + c->new_size.clear(); + for (int i = 0; i < new_size->size; ++i) { + c->new_size.push_back(new_size->data[i]); + } + TfLiteIntArrayFree(new_size); + return kTfLiteOk; +} + +TEST(UtilTest, ConvertStatus) { + TestContext context; + context.ReportError = ReportError; + + EXPECT_EQ(ConvertStatus(&context, tensorflow::errors::Internal("Some Error")), + kTfLiteError); + EXPECT_EQ(context.error, "Some Error"); + + context.error.clear(); + EXPECT_EQ(ConvertStatus(&context, tensorflow::Status()), kTfLiteOk); + EXPECT_TRUE(context.error.empty()); +} + +TEST(UtilTest, CopyShape) { + TestContext context; + context.ReportError = ReportError; + context.ResizeTensor = ResizeTensor; + + using tensorflow::DT_FLOAT; + using tensorflow::Tensor; + + TfLiteTensor dst; + + EXPECT_EQ(CopyShape(&context, Tensor(), &dst), kTfLiteOk); + EXPECT_THAT(context.new_size, ElementsAre(0)); + + EXPECT_EQ(CopyShape(&context, Tensor(DT_FLOAT, {1, 2}), &dst), kTfLiteOk); + EXPECT_THAT(context.new_size, ElementsAre(1, 2)); + + EXPECT_EQ(CopyShape(&context, Tensor(DT_FLOAT, {1LL << 44, 2}), &dst), + kTfLiteError); + EXPECT_EQ(context.error, + "Dimension value in TensorFlow shape is larger than supported by " + "TF Lite"); +} + +} // namespace +} // namespace tflite + +int main(int argc, char** argv) { + ::tflite::LogToStderr(); + ::testing::InitGoogleTest(&argc, argv); + return RUN_ALL_TESTS(); +} diff --git a/tensorflow/contrib/lite/delegates/nnapi/nnapi_delegate.cc b/tensorflow/contrib/lite/delegates/nnapi/nnapi_delegate.cc index e96ee92376901a341a1f739d0d79727deeb443eb..f0d16575ec5a0f9db799b8be44907e8a999a348c 100644 --- a/tensorflow/contrib/lite/delegates/nnapi/nnapi_delegate.cc +++ b/tensorflow/contrib/lite/delegates/nnapi/nnapi_delegate.cc @@ -61,7 +61,10 @@ int32_t GetAndroidSdkVersion() { return 0; } +constexpr int32_t kMinSdkVersionForNNAPI = 27; +constexpr int32_t kMinSdkVersionForNNAPI11 = 28; static const int32_t kAndroidSdkVersion = GetAndroidSdkVersion(); + } // namespace // RAII NN API Model Destructor for use with std::unique_ptr @@ -133,6 +136,12 @@ class NNAPIOpBuilder { return AddScalarOperand(value, ANEURALNETWORKS_FLOAT32); } + TfLiteStatus AddVectorInt32Operand(const int32_t* values, + uint32_t num_values) { + return AddVectorOperand(values, num_values, + ANEURALNETWORKS_TENSOR_INT32); + } + TfLiteStatus AddPoolingParams(void* data) { auto builtin = reinterpret_cast(data); AddScalarInt32Operand(builtin->padding); @@ -244,6 +253,21 @@ class NNAPIOpBuilder { return kTfLiteOk; } + template + TfLiteStatus AddVectorOperand(const T* values, uint32_t num_values, + int32_t nn_type) { + ANeuralNetworksOperandType operand_type{ + .type = nn_type, .dimensionCount = 1, .dimensions = &num_values}; + CHECK_NN(context_, + ANeuralNetworksModel_addOperand(nn_model_, &operand_type)); + int ann_operand = operand_mapping_->add_new_non_tensor_operand(); + CHECK_NN(context_, + ANeuralNetworksModel_setOperandValue( + nn_model_, ann_operand, values, sizeof(T) * num_values)); + augmented_inputs_.push_back(ann_operand); + return kTfLiteOk; + } + // TfLiteContext for error handling. Must be named context for macros to // work. TfLiteContext* context_; @@ -411,6 +435,40 @@ class NNAPIDelegateKernel { return nullptr; } break; + case kTfLiteBuiltinSqueeze: + // Squeeze requires NNAPI1.1. + if (version == 1 && kAndroidSdkVersion >= kMinSdkVersionForNNAPI11) { + return [](TfLiteContext* context, NNAPIOpBuilder* builder, + TfLiteNode* node) -> ANeuralNetworksOperationType { + auto builtin = + reinterpret_cast(node->builtin_data); + // Note that we add the squeeze dimensions even if the dimensions + // were unspecified (empty), as NNAPI requires the operand. + builder->AddVectorInt32Operand( + builtin->squeeze_dims, + static_cast(builtin->num_squeeze_dims)); + return ANEURALNETWORKS_SQUEEZE; + }; + } else { + return nullptr; + } + case kTfLiteBuiltinTranspose: + // Transpose requires NNAPI1.1. Also note that the permutation input + // tensor value dictates the output dimensions. + // TODO(b/110888333): Support dynamically-sized tensors in delegates. + if ((version == 1) && + (kAndroidSdkVersion >= kMinSdkVersionForNNAPI11) && + (node->inputs->size > 1) && + (context->tensors[node->inputs->data[1]].allocation_type == + kTfLiteMmapRo)) { + return [](TfLiteContext* context, NNAPIOpBuilder* builder, + TfLiteNode* node) -> ANeuralNetworksOperationType { + return ANEURALNETWORKS_TRANSPOSE; + }; + } else { + return nullptr; + } + break; default: return nullptr; } @@ -560,8 +618,9 @@ TfLiteDelegate* NnApiDelegate() { .Prepare = [](TfLiteContext* context, TfLiteDelegate* delegate) -> TfLiteStatus { // Do not check nodes_ if NN API is unavailable. - // NN API is only available since Android O-MR1 (API 27). - if (kAndroidSdkVersion < 27 || !NNAPIExists()) return kTfLiteOk; + if (kAndroidSdkVersion < kMinSdkVersionForNNAPI || !NNAPIExists()) { + return kTfLiteOk; + } std::vector supported_nodes(1); // We don't care about all nodes_, we only care about ones in the diff --git a/tensorflow/contrib/lite/delegates/nnapi/nnapi_delegate_test.cc b/tensorflow/contrib/lite/delegates/nnapi/nnapi_delegate_test.cc index 799e3efe0bb09b242d8e5b1d15d7a9646965a85d..ab2181e8ff7b5b4e73dc0264a6e93591d2ee225a 100644 --- a/tensorflow/contrib/lite/delegates/nnapi/nnapi_delegate_test.cc +++ b/tensorflow/contrib/lite/delegates/nnapi/nnapi_delegate_test.cc @@ -27,14 +27,20 @@ using ::testing::ElementsAreArray; // TODO(b/110368244): figure out how to share the existing tests in kernels/ but // with the delegation on. Also, add more unit tests to improve code coverage. -class FloatAddOpModel : public SingleOpModel { +class SingleOpModelWithNNAPI : public SingleOpModel { + public: + SingleOpModelWithNNAPI() { + this->SetApplyDelegate([](Interpreter* interpreter) { + interpreter->ModifyGraphWithDelegate(NnApiDelegate(), false); + }); + } +}; + +class FloatAddOpModel : public SingleOpModelWithNNAPI { public: FloatAddOpModel(const TensorData& input1, const TensorData& input2, const TensorData& output, ActivationFunctionType activation_type) { - this->SetApplyDelegate([](Interpreter* interpreter) { - interpreter->ModifyGraphWithDelegate(NnApiDelegate()); - }); input1_ = AddInput(input1); input2_ = AddInput(input2); output_ = AddOutput(output); @@ -81,9 +87,6 @@ class FloatMulOpModel : public SingleOpModel { FloatMulOpModel(const TensorData& input1, const TensorData& input2, const TensorData& output, ActivationFunctionType activation_type) { - this->SetApplyDelegate([](Interpreter* interpreter) { - interpreter->ModifyGraphWithDelegate(NnApiDelegate()); - }); input1_ = AddInput(input1); input2_ = AddInput(input2); output_ = AddOutput(output); @@ -114,15 +117,11 @@ TEST(NNAPIDelegate, MulWithNoActivation) { ElementsAreArray(ArrayFloatNear({-0.2, 0.04, 0.21, 0.4}))); } -class FloatPoolingOpModel : public SingleOpModel { +class FloatPoolingOpModel : public SingleOpModelWithNNAPI { public: FloatPoolingOpModel(BuiltinOperator type, const TensorData& input, int filter_width, int filter_height, const TensorData& output) { - this->SetApplyDelegate([](Interpreter* interpreter) { - interpreter->ModifyGraphWithDelegate(NnApiDelegate()); - }); - input_ = AddInput(input); output_ = AddOutput(output); @@ -193,10 +192,6 @@ class BaseConvolutionOpModel : public SingleOpModel { enum Padding padding = Padding_VALID, enum ActivationFunctionType activation = ActivationFunctionType_NONE, int dilation_width_factor = 1, int dilation_height_factor = 1) { - this->SetApplyDelegate([](Interpreter* interpreter) { - interpreter->ModifyGraphWithDelegate(NnApiDelegate()); - }); - input_ = AddInput(input); filter_ = AddInput(filter); @@ -344,14 +339,10 @@ TEST(NNAPIDelegate, Conv2DWithNoActivation) { })); } -class DepthwiseConvolutionOpModel : public SingleOpModel { +class DepthwiseConvolutionOpModel : public SingleOpModelWithNNAPI { public: DepthwiseConvolutionOpModel(const TensorData& input, const TensorData& filter, const TensorData& output) { - this->SetApplyDelegate([](Interpreter* interpreter) { - interpreter->ModifyGraphWithDelegate(NnApiDelegate()); - }); - input_ = AddInput(input); filter_ = AddInput(filter); @@ -426,15 +417,11 @@ TEST(NNAPIDelegate, DepthwiseConv2DWithNoActivation) { })); } -class FloatFullyConnectedOpModel : public SingleOpModel { +class FloatFullyConnectedOpModel : public SingleOpModelWithNNAPI { public: FloatFullyConnectedOpModel(int units, int batches, const TensorData& input, const TensorData& output = {TensorType_FLOAT32}) : batches_(batches), units_(units) { - this->SetApplyDelegate([](Interpreter* interpreter) { - interpreter->ModifyGraphWithDelegate(NnApiDelegate()); - }); - int total_input_size = 1; for (int i = 0; i < input.shape.size(); ++i) { total_input_size *= input.shape[i]; @@ -515,14 +502,10 @@ TEST(NNAPIDelegate, FullyConnectedSimpleTest) { EXPECT_THAT(m.GetOutput(), ElementsAre(24, 25, 26, 58, 59, 60)); } -class SoftmaxOpModel : public SingleOpModel { +class SoftmaxOpModel : public SingleOpModelWithNNAPI { public: SoftmaxOpModel(int batches, int size, float beta) : batches_(batches), input_size_(size), beta_(beta) { - this->SetApplyDelegate([](Interpreter* interpreter) { - interpreter->ModifyGraphWithDelegate(NnApiDelegate()); - }); - input_ = AddInput(TensorType_FLOAT32); output_ = AddOutput(TensorType_FLOAT32); SetBuiltinOp(BuiltinOperator_SOFTMAX, BuiltinOptions_SoftmaxOptions, @@ -566,14 +549,10 @@ TEST(NNAPIDelegate, SoftmaxSimpleTest) { 1e-6))); } -class ReshapeOpModel : public SingleOpModel { +class ReshapeOpModel : public SingleOpModelWithNNAPI { public: ReshapeOpModel(std::initializer_list input_shape, std::initializer_list new_shape) { - this->SetApplyDelegate([](Interpreter* interpreter) { - interpreter->ModifyGraphWithDelegate(NnApiDelegate()); - }); - input_ = AddInput(TensorType_FLOAT32); new_shape_ = AddInput(TensorType_INT32); output_ = AddOutput(TensorType_FLOAT32); @@ -605,6 +584,100 @@ TEST(NNAPIDelegate, ReshapeSimpleTest) { EXPECT_THAT(m.GetOutputShape(), ElementsAreArray({2, 2, 2})); } +class SqueezeOpModel : public SingleOpModelWithNNAPI { + public: + SqueezeOpModel(const TensorData& input, const TensorData& output, + std::initializer_list axis) { + input_ = AddInput(input); + output_ = AddOutput(output); + SetBuiltinOp( + BuiltinOperator_SQUEEZE, BuiltinOptions_SqueezeOptions, + CreateSqueezeOptions(builder_, builder_.CreateVector(axis)) + .Union()); + BuildInterpreter({GetShape(input_)}); + } + + void SetInput(std::initializer_list data) { + PopulateTensor(input_, data); + } + std::vector GetOutput() { return ExtractVector(output_); } + std::vector GetOutputShape() { return GetTensorShape(output_); } + + private: + int input_; + int new_shape_; + int output_; +}; + +TEST(NNAPIDelegate, SqueezeSimpleTest) { + std::initializer_list data = { + 1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0, 10.0, 11.0, 12.0, + 13.0, 14.0, 15.0, 16.0, 17.0, 18.0, 19.0, 20.0, 21.0, 22.0, 23.0, 24.0}; + SqueezeOpModel m({TensorType_FLOAT32, {1, 24, 1}}, {TensorType_FLOAT32, {24}}, + {}); + m.SetInput(data); + m.Invoke(); + EXPECT_THAT(m.GetOutputShape(), ElementsAreArray({24})); + EXPECT_THAT( + m.GetOutput(), + ElementsAreArray({1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, + 9.0, 10.0, 11.0, 12.0, 13.0, 14.0, 15.0, 16.0, + 17.0, 18.0, 19.0, 20.0, 21.0, 22.0, 23.0, 24.0})); +} + +TEST(NNAPIDelegate, SqueezeWithAxisTest) { + std::initializer_list data = { + 1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0, 10.0, 11.0, 12.0, + 13.0, 14.0, 15.0, 16.0, 17.0, 18.0, 19.0, 20.0, 21.0, 22.0, 23.0, 24.0}; + SqueezeOpModel m({TensorType_FLOAT32, {1, 24, 1}}, {TensorType_FLOAT32, {24}}, + {2}); + m.SetInput(data); + m.Invoke(); + EXPECT_THAT(m.GetOutputShape(), ElementsAreArray({1, 24})); + EXPECT_THAT( + m.GetOutput(), + ElementsAreArray({1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, + 9.0, 10.0, 11.0, 12.0, 13.0, 14.0, 15.0, 16.0, + 17.0, 18.0, 19.0, 20.0, 21.0, 22.0, 23.0, 24.0})); +} + +class TransposeSimpleModel : public SingleOpModelWithNNAPI { + public: + TransposeSimpleModel(std::initializer_list input_shape, + std::initializer_list perm_shape, + std::initializer_list perm) { + input_ = AddInput(TensorType_FLOAT32); + perm_ = AddConstInput(TensorType_INT32, perm, perm_shape); + output_ = AddOutput(TensorType_FLOAT32); + SetBuiltinOp(BuiltinOperator_TRANSPOSE, BuiltinOptions_TransposeOptions, + CreateTransposeOptions(builder_).Union()); + BuildInterpreter({input_shape, perm_shape}); + } + + void SetInput(std::initializer_list data) { + PopulateTensor(input_, data); + } + + std::vector GetOutput() { return ExtractVector(output_); } + std::vector GetOutputShape() { return GetTensorShape(output_); } + + private: + int input_; + int perm_; + int output_; +}; + +TEST(NNAPIDelegate, TransposeSimpleTest) { + TransposeSimpleModel m({2, 3, 4}, {3}, {2, 0, 1}); + m.SetInput({0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, + 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23}); + m.Invoke(); + EXPECT_THAT(m.GetOutputShape(), ElementsAreArray({4, 2, 3})); + EXPECT_THAT(m.GetOutput(), + ElementsAreArray({0, 4, 8, 12, 16, 20, 1, 5, 9, 13, 17, 21, + 2, 6, 10, 14, 18, 22, 3, 7, 11, 15, 19, 23})); +} + } // namespace } // namespace tflite diff --git a/tensorflow/contrib/lite/examples/android/BUILD b/tensorflow/contrib/lite/examples/android/BUILD index dd2cd173246719976d7cd6e52d65f63125b5b2db..4d2437e7d3714e1b8b427b0c6197b295c0355b07 100644 --- a/tensorflow/contrib/lite/examples/android/BUILD +++ b/tensorflow/contrib/lite/examples/android/BUILD @@ -37,6 +37,7 @@ android_binary( "@tflite_conv_actions_frozen//:conv_actions_frozen.tflite", "//tensorflow/contrib/lite/examples/android/app/src/main/assets:conv_actions_labels.txt", "@tflite_mobilenet_ssd//:mobilenet_ssd.tflite", + "@tflite_mobilenet_ssd_quant//:detect.tflite", "//tensorflow/contrib/lite/examples/android/app/src/main/assets:box_priors.txt", "//tensorflow/contrib/lite/examples/android/app/src/main/assets:coco_labels_list.txt", ], diff --git a/tensorflow/contrib/lite/examples/android/app/README.md b/tensorflow/contrib/lite/examples/android/app/README.md new file mode 100644 index 0000000000000000000000000000000000000000..8e12bd04dd0517b229265bec15981b6eea2345df --- /dev/null +++ b/tensorflow/contrib/lite/examples/android/app/README.md @@ -0,0 +1,19 @@ +# TF Lite Android App Example + +## Building from Source with Bazel + +1. Follow the [Bazel steps for the TF Demo App](https://github.com/tensorflow/tensorflow/tree/master/tensorflow/examples/android#bazel). + +2. Build the app with Bazel. The demo needs C++11. We configure the fat_apk_cpu flag to package support for 4 hardware variants. You may replace it with --config=android_arm64 on a 64-bit device and --config=android_arm for 32-bit device: + + ```shell + bazel build -c opt --cxxopt='--std=c++11' --fat_apk_cpu=x86,x86_64,arm64-v8a,armeabi-v7a \ + //tensorflow/contrib/lite/examples/android:tflite_demo + ``` + +3. Install the demo on a + [debug-enabled device](https://github.com/tensorflow/tensorflow/tree/master/tensorflow/examples/android#install): + + ```shell + adb install bazel-bin/tensorflow/contrib/lite/examples/android/tflite_demo.apk + ``` diff --git a/tensorflow/contrib/lite/examples/android/app/build.gradle b/tensorflow/contrib/lite/examples/android/app/build.gradle index 8e0a98ed63f99b7477cdb2f851a19cd31b45f314..1ffb9dd377730bb3dc872cbf1548fa29ffaa0949 100644 --- a/tensorflow/contrib/lite/examples/android/app/build.gradle +++ b/tensorflow/contrib/lite/examples/android/app/build.gradle @@ -9,7 +9,7 @@ android { targetSdkVersion 26 versionCode 1 versionName "1.0" - testInstrumentationRunner "android.support.test.runner.AndroidJUnitRunner" + testInstrumentationRunner "androidx.test.runner.AndroidJUnitRunner" // Remove this block. jackOptions { @@ -51,7 +51,7 @@ apply from: "download-models.gradle" dependencies { compile fileTree(dir: 'libs', include: ['*.jar']) - androidTestCompile('com.android.support.test.espresso:espresso-core:2.2.2', { + androidTestCompile('com.androidx.test.espresso:espresso-core:2.2.2', { exclude group: 'com.android.support', module: 'support-annotations' }) compile 'org.tensorflow:tensorflow-lite:0.0.0-nightly' diff --git a/tensorflow/contrib/lite/examples/android/app/download-models.gradle b/tensorflow/contrib/lite/examples/android/app/download-models.gradle index 8e65dc076f2a8daaddf01ceab6796b8ed1127af3..c100e37c16f38a65f7b1f64a3f6e3eaa1477e8eb 100644 --- a/tensorflow/contrib/lite/examples/android/app/download-models.gradle +++ b/tensorflow/contrib/lite/examples/android/app/download-models.gradle @@ -12,8 +12,9 @@ def models = ['conv_actions_tflite.zip', 'mobilenet_ssd_tflite_v1.zip', - 'mobilenet_v1_224_android_quant_2017_11_08.zip'] -// LINT.ThenChange(//tensorflow/examples/android/BUILD) + 'mobilenet_v1_224_android_quant_2017_11_08.zip', + 'coco_ssd_mobilenet_v1_1.0_quant_2018_06_29.zip'] +// LINT.ThenChange(//tensorflow/contrib/lite/examples/android/BUILD) // Root URL for model archives def MODEL_URL = 'https://storage.googleapis.com/download.tensorflow.org/models/tflite' diff --git a/tensorflow/contrib/lite/examples/android/app/src/main/assets/pets_labels_list.txt b/tensorflow/contrib/lite/examples/android/app/src/main/assets/pets_labels_list.txt new file mode 100644 index 0000000000000000000000000000000000000000..d581f733e48ff8c2ba88162ee56b5e9d12aec7de --- /dev/null +++ b/tensorflow/contrib/lite/examples/android/app/src/main/assets/pets_labels_list.txt @@ -0,0 +1,38 @@ +??? +Abyssinian +american_bulldog +american_pit_bull_terrier +basset_hound +beagle +Bengal +Birman +Bombay +boxer +British_Shorthair +chihuahua +Egyptian_Mau +english_cocker_spaniel +english_setter +german_shorthaired +great_pyrenees +havanese +japanese_chin +keeshond +leonberger +Maine_Coon +miniature_pinscher +newfoundland +Persian +pomeranian +pug +Ragdoll +Russian_Blue +saint_bernard +samoyed +scottish_terrier +shiba_inu +Siamese +Sphynx +staffordshire_bull_terrier +wheaten_terrier +yorkshire_terrier diff --git a/tensorflow/contrib/lite/examples/android/app/src/main/java/org/tensorflow/demo/DetectorActivity.java b/tensorflow/contrib/lite/examples/android/app/src/main/java/org/tensorflow/demo/DetectorActivity.java index de997e454a1e33254cb7c2c932ca79d0072539fa..87160f6b3fb8c0d24e5df131d9becbb3eb6e2980 100644 --- a/tensorflow/contrib/lite/examples/android/app/src/main/java/org/tensorflow/demo/DetectorActivity.java +++ b/tensorflow/contrib/lite/examples/android/app/src/main/java/org/tensorflow/demo/DetectorActivity.java @@ -1,5 +1,5 @@ /* - * Copyright 2016 The TensorFlow Authors. All Rights Reserved. + * Copyright 2018 The TensorFlow Authors. All Rights Reserved. * * Licensed under the Apache License, Version 2.0 (the "License"); * you may not use this file except in compliance with the License. @@ -50,9 +50,10 @@ public class DetectorActivity extends CameraActivity implements OnImageAvailable // Configuration values for the prepackaged SSD model. private static final int TF_OD_API_INPUT_SIZE = 300; - private static final String TF_OD_API_MODEL_FILE = "mobilenet_ssd.tflite"; + private static final boolean TF_OD_API_IS_QUANTIZED = true; + private static final String TF_OD_API_MODEL_FILE = "detect.tflite"; private static final String TF_OD_API_LABELS_FILE = "file:///android_asset/coco_labels_list.txt"; - + // Which detection model to use: by default uses Tensorflow Object Detection API frozen // checkpoints. private enum DetectorMode { @@ -107,7 +108,11 @@ public class DetectorActivity extends CameraActivity implements OnImageAvailable try { detector = TFLiteObjectDetectionAPIModel.create( - getAssets(), TF_OD_API_MODEL_FILE, TF_OD_API_LABELS_FILE, TF_OD_API_INPUT_SIZE); + getAssets(), + TF_OD_API_MODEL_FILE, + TF_OD_API_LABELS_FILE, + TF_OD_API_INPUT_SIZE, + TF_OD_API_IS_QUANTIZED); cropSize = TF_OD_API_INPUT_SIZE; } catch (final IOException e) { LOGGER.e("Exception initializing classifier!", e); diff --git a/tensorflow/contrib/lite/examples/android/app/src/main/java/org/tensorflow/demo/TFLiteObjectDetectionAPIModel.java b/tensorflow/contrib/lite/examples/android/app/src/main/java/org/tensorflow/demo/TFLiteObjectDetectionAPIModel.java index bfb4a0a04bc90566736864bf62340d1032961858..9eb21de9d03e387d3c25b38171e154a358dc81ce 100644 --- a/tensorflow/contrib/lite/examples/android/app/src/main/java/org/tensorflow/demo/TFLiteObjectDetectionAPIModel.java +++ b/tensorflow/contrib/lite/examples/android/app/src/main/java/org/tensorflow/demo/TFLiteObjectDetectionAPIModel.java @@ -25,15 +25,14 @@ import java.io.FileInputStream; import java.io.IOException; import java.io.InputStream; import java.io.InputStreamReader; +import java.nio.ByteBuffer; +import java.nio.ByteOrder; import java.nio.MappedByteBuffer; import java.nio.channels.FileChannel; import java.util.ArrayList; -import java.util.Comparator; import java.util.HashMap; import java.util.List; import java.util.Map; -import java.util.PriorityQueue; -import java.util.StringTokenizer; import java.util.Vector; import org.tensorflow.demo.env.Logger; import org.tensorflow.lite.Interpreter; @@ -46,32 +45,35 @@ public class TFLiteObjectDetectionAPIModel implements Classifier { private static final Logger LOGGER = new Logger(); // Only return this many results. - private static final int NUM_RESULTS = 1917; - private static final int NUM_CLASSES = 91; - - private static final float Y_SCALE = 10.0f; - private static final float X_SCALE = 10.0f; - private static final float H_SCALE = 5.0f; - private static final float W_SCALE = 5.0f; - + private static final int NUM_DETECTIONS = 10; + private boolean isModelQuantized; + // Float model + private static final float IMAGE_MEAN = 128.0f; + private static final float IMAGE_STD = 128.0f; + // Number of threads in the java app + private static final int NUM_THREADS = 4; // Config values. private int inputSize; - - private final float[][] boxPriors = new float[4][NUM_RESULTS]; - // Pre-allocated buffers. private Vector labels = new Vector(); private int[] intValues; + // outputLocations: array of shape [Batchsize, NUM_DETECTIONS,4] + // contains the location of detected boxes private float[][][] outputLocations; - private float[][][] outputClasses; - - float[][][][] img; + // outputClasses: array of shape [Batchsize, NUM_DETECTIONS] + // contains the classes of detected boxes + private float[][] outputClasses; + // outputScores: array of shape [Batchsize, NUM_DETECTIONS] + // contains the scores of detected boxes + private float[][] outputScores; + // numDetections: array of shape [Batchsize] + // contains the number of detected boxes + private float[] numDetections; + + private ByteBuffer imgData; private Interpreter tfLite; - private float expit(final float x) { - return (float) (1. / (1. + Math.exp(-x))); - } /** Memory-map the model file in Assets. */ private static MappedByteBuffer loadModelFile(AssetManager assets, String modelFilename) @@ -84,77 +86,24 @@ public class TFLiteObjectDetectionAPIModel implements Classifier { return fileChannel.map(FileChannel.MapMode.READ_ONLY, startOffset, declaredLength); } - private void loadCoderOptions( - final AssetManager assetManager, final String locationFilename, final float[][] boxPriors) - throws IOException { - // Try to be intelligent about opening from assets or sdcard depending on prefix. - final String assetPrefix = "file:///android_asset/"; - InputStream is; - if (locationFilename.startsWith(assetPrefix)) { - is = assetManager.open(locationFilename.split(assetPrefix, -1)[1]); - } else { - is = new FileInputStream(locationFilename); - } - - final BufferedReader reader = new BufferedReader(new InputStreamReader(is)); - - for (int lineNum = 0; lineNum < 4; ++lineNum) { - String line = reader.readLine(); - final StringTokenizer st = new StringTokenizer(line, ", "); - int priorIndex = 0; - while (st.hasMoreTokens()) { - final String token = st.nextToken(); - try { - final float number = Float.parseFloat(token); - boxPriors[lineNum][priorIndex++] = number; - } catch (final NumberFormatException e) { - // Silently ignore. - } - } - if (priorIndex != NUM_RESULTS) { - throw new RuntimeException( - "BoxPrior length mismatch: " + priorIndex + " vs " + NUM_RESULTS); - } - } - - LOGGER.i("Loaded box priors!"); - } - - void decodeCenterSizeBoxes(float[][][] predictions) { - for (int i = 0; i < NUM_RESULTS; ++i) { - float ycenter = predictions[0][i][0] / Y_SCALE * boxPriors[2][i] + boxPriors[0][i]; - float xcenter = predictions[0][i][1] / X_SCALE * boxPriors[3][i] + boxPriors[1][i]; - float h = (float) Math.exp(predictions[0][i][2] / H_SCALE) * boxPriors[2][i]; - float w = (float) Math.exp(predictions[0][i][3] / W_SCALE) * boxPriors[3][i]; - - float ymin = ycenter - h / 2.f; - float xmin = xcenter - w / 2.f; - float ymax = ycenter + h / 2.f; - float xmax = xcenter + w / 2.f; - - predictions[0][i][0] = ymin; - predictions[0][i][1] = xmin; - predictions[0][i][2] = ymax; - predictions[0][i][3] = xmax; - } - } - /** * Initializes a native TensorFlow session for classifying images. * * @param assetManager The asset manager to be used to load assets. * @param modelFilename The filepath of the model GraphDef protocol buffer. * @param labelFilename The filepath of label file for classes. + * @param inputSize The size of image input + * @param isQuantized Boolean representing model is quantized or not */ public static Classifier create( final AssetManager assetManager, final String modelFilename, final String labelFilename, - final int inputSize) throws IOException { + final int inputSize, + final boolean isQuantized) + throws IOException { final TFLiteObjectDetectionAPIModel d = new TFLiteObjectDetectionAPIModel(); - d.loadCoderOptions(assetManager, "file:///android_asset/box_priors.txt", d.boxPriors); - InputStream labelsInput = null; String actualFilename = labelFilename.split("file:///android_asset/")[1]; labelsInput = assetManager.open(actualFilename); @@ -175,12 +124,23 @@ public class TFLiteObjectDetectionAPIModel implements Classifier { throw new RuntimeException(e); } + d.isModelQuantized = isQuantized; // Pre-allocate buffers. - d.img = new float[1][inputSize][inputSize][3]; - + int numBytesPerChannel; + if (isQuantized) { + numBytesPerChannel = 1; // Quantized + } else { + numBytesPerChannel = 4; // Floating point + } + d.imgData = ByteBuffer.allocateDirect(1 * d.inputSize * d.inputSize * 3 * numBytesPerChannel); + d.imgData.order(ByteOrder.nativeOrder()); d.intValues = new int[d.inputSize * d.inputSize]; - d.outputLocations = new float[1][NUM_RESULTS][4]; - d.outputClasses = new float[1][NUM_RESULTS][NUM_CLASSES]; + + d.tfLite.setNumThreads(NUM_THREADS); + d.outputLocations = new float[1][NUM_DETECTIONS][4]; + d.outputClasses = new float[1][NUM_DETECTIONS]; + d.outputScores = new float[1][NUM_DETECTIONS]; + d.numDetections = new float[1]; return d; } @@ -196,25 +156,37 @@ public class TFLiteObjectDetectionAPIModel implements Classifier { // on the provided parameters. bitmap.getPixels(intValues, 0, bitmap.getWidth(), 0, 0, bitmap.getWidth(), bitmap.getHeight()); + imgData.rewind(); for (int i = 0; i < inputSize; ++i) { for (int j = 0; j < inputSize; ++j) { - int pixel = intValues[j * inputSize + i]; - img[0][j][i][2] = (float) (pixel & 0xFF) / 128.0f - 1.0f; - img[0][j][i][1] = (float) ((pixel >> 8) & 0xFF) / 128.0f - 1.0f; - img[0][j][i][0] = (float) ((pixel >> 16) & 0xFF) / 128.0f - 1.0f; + int pixelValue = intValues[i * inputSize + j]; + if (isModelQuantized) { + // Quantized model + imgData.put((byte) ((pixelValue >> 16) & 0xFF)); + imgData.put((byte) ((pixelValue >> 8) & 0xFF)); + imgData.put((byte) (pixelValue & 0xFF)); + } else { // Float model + imgData.putFloat((((pixelValue >> 16) & 0xFF) - IMAGE_MEAN) / IMAGE_STD); + imgData.putFloat((((pixelValue >> 8) & 0xFF) - IMAGE_MEAN) / IMAGE_STD); + imgData.putFloat(((pixelValue & 0xFF) - IMAGE_MEAN) / IMAGE_STD); + } } } Trace.endSection(); // preprocessBitmap // Copy the input data into TensorFlow. Trace.beginSection("feed"); - outputLocations = new float[1][NUM_RESULTS][4]; - outputClasses = new float[1][NUM_RESULTS][NUM_CLASSES]; + outputLocations = new float[1][NUM_DETECTIONS][4]; + outputClasses = new float[1][NUM_DETECTIONS]; + outputScores = new float[1][NUM_DETECTIONS]; + numDetections = new float[1]; - Object[] inputArray = {img}; + Object[] inputArray = {imgData}; Map outputMap = new HashMap<>(); outputMap.put(0, outputLocations); outputMap.put(1, outputClasses); + outputMap.put(2, outputScores); + outputMap.put(3, numDetections); Trace.endSection(); // Run the inference call. @@ -222,56 +194,26 @@ public class TFLiteObjectDetectionAPIModel implements Classifier { tfLite.runForMultipleInputsOutputs(inputArray, outputMap); Trace.endSection(); - decodeCenterSizeBoxes(outputLocations); - - // Find the best detections. - final PriorityQueue pq = - new PriorityQueue( - 1, - new Comparator() { - @Override - public int compare(final Recognition lhs, final Recognition rhs) { - // Intentionally reversed to put high confidence at the head of the queue. - return Float.compare(rhs.getConfidence(), lhs.getConfidence()); - } - }); - - // Scale them back to the input size. - for (int i = 0; i < NUM_RESULTS; ++i) { - float topClassScore = -1000f; - int topClassScoreIndex = -1; - - // Skip the first catch-all class. - for (int j = 1; j < NUM_CLASSES; ++j) { - float score = expit(outputClasses[0][i][j]); - - if (score > topClassScore) { - topClassScoreIndex = j; - topClassScore = score; - } - } - - if (topClassScore > 0.001f) { - final RectF detection = - new RectF( - outputLocations[0][i][1] * inputSize, - outputLocations[0][i][0] * inputSize, - outputLocations[0][i][3] * inputSize, - outputLocations[0][i][2] * inputSize); - - pq.add( - new Recognition( - "" + i, - labels.get(topClassScoreIndex), - outputClasses[0][i][topClassScoreIndex], - detection)); - } - } - - final ArrayList recognitions = new ArrayList(); - for (int i = 0; i < Math.min(pq.size(), 10); ++i) { - Recognition recog = pq.poll(); - recognitions.add(recog); + // Show the best detections. + // after scaling them back to the input size. + final ArrayList recognitions = new ArrayList<>(NUM_DETECTIONS); + for (int i = 0; i < NUM_DETECTIONS; ++i) { + final RectF detection = + new RectF( + outputLocations[0][i][1] * inputSize, + outputLocations[0][i][0] * inputSize, + outputLocations[0][i][3] * inputSize, + outputLocations[0][i][2] * inputSize); + // SSD Mobilenet V1 Model assumes class 0 is background class + // in label file and class labels start from 1 to number_of_classes+1, + // while outputClasses correspond to class index from 0 to number_of_classes + int labelOffset = 1; + recognitions.add( + new Recognition( + "" + i, + labels.get((int) outputClasses[0][i] + labelOffset), + outputScores[0][i], + detection)); } Trace.endSection(); // "recognizeImage" return recognitions; diff --git a/tensorflow/contrib/lite/g3doc/benchmarks.md b/tensorflow/contrib/lite/g3doc/benchmarks.md new file mode 100644 index 0000000000000000000000000000000000000000..96536cba271922f1bec51f915c2996c8d9de3b9b --- /dev/null +++ b/tensorflow/contrib/lite/g3doc/benchmarks.md @@ -0,0 +1,178 @@ +# Performance Benchmark numbers + +This document contains the performance benchmark numbers for running a few well +known models on some Android and iOS devices. + +The benchmark numbers were generated by running the [TFLite benchmark +binary](https://github.com/tensorflow/tensorflow/tree/master/tensorflow/contrib/lite/tools/benchmark) +on Android and running the [iOS benchmark +app](https://github.com/tensorflow/tensorflow/tree/master/tensorflow/contrib/lite/tools/benchmark/ios) +on iOS. + +# Android benchmarks + +When running Android benchmarks, the CPU affinity is set to use big cores on the +device to reduce variance (see +[details](https://github.com/tensorflow/tensorflow/tree/master/tensorflow/contrib/lite/tools/benchmark#reducing-variance-between-runs-on-android)). + +Models are assumed to have been downloaded from the link, unzipped and pushed to +`/data/local/tmp/tflite_models` folder. The benchmark binary is built according +to instructions listed +[here](https://github.com/tensorflow/tensorflow/tree/master/tensorflow/contrib/lite/tools/benchmark#on-android) +and is assumed to have been pushed to `/data/local/tmp`. + +The following command was used to run the benchmark: + +``` +adb shell taskset ${CPU_MASK} /data/local/tmp/benchmark_model \ + --num_threads=1 \ + --graph=/data/local/tmp/tflite_models/${GRAPH} \ + --warmup_runs=1 \ + --num_runs=50 \ + --use_nnapi=false +``` + +where `${GRAPH}` is the name of model and `${CPU_MASK}` is the CPU affinity +chosen according to the following table: + +Device | CPU_MASK | +-------| ---------- +Pixel 2 | f0 | +Pixel xl | 0c | + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +
Model NameDevice Mean inference time (std dev)
+ Mobilenet_1.0_224(float) + Pixel 2 166.5 ms (2.6 ms)
Pixel xl 122.9 ms (1.8 ms)
+ Mobilenet_1.0_224 (quant) + Pixel 2 69.5 ms (0.9 ms)
Pixel xl 78.9 ms (2.2 ms)
+ NASNet mobile + Pixel 2 273.8 ms (3.5 ms)
Pixel xl 210.8 ms (4.2 ms)
+ SqueezeNet + Pixel 2 234.0 ms (2.1 ms)
Pixel xl 158.0 ms (2.1 ms)
+ Inception_ResNet_V2 + Pixel 2 2846.0 ms (15.0 ms)
Pixel xl 1973.0 ms (15.0 ms)
+ Inception_V4 + Pixel 2 3180.0 ms (11.7 ms)
Pixel xl 2262.0 ms (21.0 ms)
+ +# iOS benchmarks + +For running iOS benchmarks, the [benchmark +app](https://github.com/tensorflow/tensorflow/tree/master/tensorflow/contrib/lite/tools/benchmark/ios) +was modified to include the appropriate model and `benchmark_params.json` was +modified to set `num_threads` to 1. + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +
Model NameDevice Mean inference time (std dev)
+ Mobilenet_1.0_224(float) + iPhone 8 32.2 ms (0.8 ms)
+ Mobilenet_1.0_224 (quant) + iPhone 8 24.4 ms (0.8 ms)
+ NASNet mobile + iPhone 8 60.3 ms (0.6 ms)
+ SqueezeNet + iPhone 8 44.3 (0.7 ms)
+ Inception_ResNet_V2 + iPhone 8562.4 ms (18.2 ms)
+ Inception_V4 + iPhone 8 661.0 ms (29.2 ms)
diff --git a/tensorflow/contrib/lite/g3doc/tf_ops_compatibility.md b/tensorflow/contrib/lite/g3doc/tf_ops_compatibility.md index 45104c141945a451351257e9bdbb43c0ad328258..49d00a66ba7d7e37a672de250a98f877e0cae75f 100644 --- a/tensorflow/contrib/lite/g3doc/tf_ops_compatibility.md +++ b/tensorflow/contrib/lite/g3doc/tf_ops_compatibility.md @@ -42,6 +42,7 @@ counterparts: *as long as the input tensor is 4D (1 batch + 2 spatial + 1 other) and the crops attribute is not used* * [tf.exp](https://www.tensorflow.org/api_docs/python/tf/exp) +* [tf.fake_quant*](https://www.tensorflow.org/api_docs/python/tf/fake_quant_with_min_max_args) * [tf.matmul](https://www.tensorflow.org/api_docs/python/tf/matmul) - *as long as the second argument is constant and transposition is not used* * [tf.nn.avg_pool](https://www.tensorflow.org/api_docs/python/tf/nn/avg_pool) @@ -778,6 +779,42 @@ Outputs { } ``` +**POW** + +``` +Inputs { + 0: a tensor + 1: a tensor +} +Outputs { + 0: elementwise pow of the input tensors +} +``` + +**ARG_MAX** + +``` +Inputs { + 0: a tensor + 1: a tensor +} +Outputs { + 0: A tensor of indices of maximum values. +} +``` + +**ARG_MIN** + +``` +Inputs { + 0: a tensor + 1: a tensor +} +Outputs { + 0: A tensor of indices of minium values. +} +``` + And these are TensorFlow Lite operations that are present but not ready for custom models yet: diff --git a/tensorflow/contrib/lite/interpreter.cc b/tensorflow/contrib/lite/interpreter.cc index 57b2c0f32b6c07083bb88aa9b81fcd7f71dbc672..0641a08636248826167fcde6bcae5f530e177ba3 100644 --- a/tensorflow/contrib/lite/interpreter.cc +++ b/tensorflow/contrib/lite/interpreter.cc @@ -22,17 +22,21 @@ limitations under the License. #include "tensorflow/contrib/lite/arena_planner.h" #include "tensorflow/contrib/lite/context.h" +#include "tensorflow/contrib/lite/context_util.h" #include "tensorflow/contrib/lite/error_reporter.h" #include "tensorflow/contrib/lite/graph_info.h" -#include "tensorflow/contrib/lite/kernels/eigen_support.h" -#include "tensorflow/contrib/lite/kernels/gemm_support.h" #include "tensorflow/contrib/lite/memory_planner.h" +#ifndef TFLITE_MCU #include "tensorflow/contrib/lite/nnapi_delegate.h" +#endif #include "tensorflow/contrib/lite/profiling/profiler.h" #include "tensorflow/contrib/lite/schema/schema_generated.h" #include "tensorflow/contrib/lite/util.h" namespace tflite { +#ifdef TFLITE_MCU +class NNAPIDelegate {}; +#endif namespace { @@ -53,6 +57,19 @@ void SetForbiddenContextFunction(FunctionType* func) { *func = reinterpret_cast(ForbiddenContextFunction); } +// Returns true if at least one tensor in the given list is kTfLiteDynamic. +template +bool HasDynamicTensorImpl(const TfLiteContext& context, + const TensorIntArray& int_array) { + for (int i : int_array) { + const TfLiteTensor& tensor = context.tensors[i]; + if (tensor.allocation_type == kTfLiteDynamic) { + return true; + } + } + return false; +} + } // namespace // A trivial implementation of GraphInfo around the Interpreter. @@ -99,9 +116,9 @@ Interpreter::Interpreter(ErrorReporter* error_reporter) context_.AddTensors = AddTensors; context_.tensors = nullptr; context_.tensors_size = 0; - context_.eigen_context = nullptr; - context_.gemm_context = nullptr; context_.recommended_num_threads = -1; + context_.GetExternalContext = GetExternalContext; + context_.SetExternalContext = SetExternalContext; // Invalid to call these these except from TfLiteDelegate SetForbiddenContextFunction(&context_.GetNodeAndRegistration); @@ -112,6 +129,11 @@ Interpreter::Interpreter(ErrorReporter* error_reporter) tensors_.reserve(kTensorsReservedCapacity); nodes_and_registration_.reserve(kTensorsReservedCapacity); next_execution_plan_index_to_prepare_ = 0; + + for (int i = 0; i < kTfLiteMaxExternalContexts; ++i) { + external_contexts_[i] = nullptr; + } + UseNNAPI(false); } @@ -269,6 +291,33 @@ TfLiteStatus Interpreter::ReplaceSubgraphsWithDelegateKernels( return kTfLiteOk; } +TfLiteExternalContext* Interpreter::GetExternalContext( + TfLiteExternalContextType type) { + if (type >= 0 && type < kTfLiteMaxExternalContexts) { + return external_contexts_[type]; + } + return nullptr; +} + +TfLiteExternalContext* Interpreter::GetExternalContext( + struct TfLiteContext* context, TfLiteExternalContextType type) { + return static_cast(context->impl_)->GetExternalContext(type); +} + +void Interpreter::SetExternalContext(TfLiteExternalContextType type, + TfLiteExternalContext* ctx) { + if (type >= 0 && type < kTfLiteMaxExternalContexts) { + external_contexts_[type] = ctx; + } +} + +void Interpreter::SetExternalContext(struct TfLiteContext* context, + TfLiteExternalContextType type, + TfLiteExternalContext* ctx) { + return static_cast(context->impl_) + ->SetExternalContext(type, ctx); +} + // Gets an TfLiteIntArray* representing the execution plan. The interpreter owns // this memory and it is only guaranteed to exist during the invocation of the // delegate prepare. @@ -359,33 +408,46 @@ TfLiteStatus Interpreter::BytesRequired(TfLiteType type, const int* dims, case kTfLiteBool: *bytes = sizeof(bool) * count; break; + case kTfLiteComplex64: + *bytes = sizeof(std::complex) * count; + break; default: ReportError(&context_, - "Only float32, int16, int32, int64, uint8, bool supported " - "currently."); + "Only float32, int16, int32, int64, uint8, bool, complex64 " + "supported currently."); return kTfLiteError; } return kTfLiteOk; } TfLiteStatus Interpreter::AllocateTensors() { - next_execution_plan_index_to_prepare_ = 0; - if (memory_planner_) { - TF_LITE_ENSURE_STATUS(memory_planner_->ResetAllocations()); - } - if (!consistent_) { ReportError(&context_, "AllocateTensors() called on inconsistent model."); return kTfLiteError; } - TF_LITE_ENSURE_STATUS(PrepareOpsAndTensors()); + // Explicit (re)allocation is necessary if nodes have been changed or tensors + // have been resized. For inputs marked as dynamic, we can't short-circuit the + // allocation as the client may have done the resize manually. + if (state_ != kStateUninvokable && !HasDynamicTensorImpl(context_, inputs_)) { + return kTfLiteOk; + } - if (state_ == kStateUninvokable) { - state_ = kStateInvokable; + next_execution_plan_index_to_prepare_ = 0; + if (memory_planner_) { + TF_LITE_ENSURE_STATUS(memory_planner_->ResetAllocations()); } - TF_LITE_ENSURE(&context_, state_ == kStateInvokable || - state_ == kStateInvokableAndImmutable); + + TF_LITE_ENSURE_STATUS(PrepareOpsAndTensors()); + + state_ = kStateInvokable; + + // Reset the variable tensors to zero after (re)allocating the tensors. + // Developers shouldn't rely on the side effect of this function to reset + // variable tesnsors. They should call `ResetVariableTensorsToZero` directly + // instead. + ResetVariableTensorsToZero(); + return kTfLiteOk; } @@ -478,26 +540,26 @@ TfLiteStatus Interpreter::ResizeInputTensor(int tensor_index, "ResizeInputTensor is disallowed when graph is immutable."); return kTfLiteError; } - state_ = kStateUninvokable; // TODO(aselle): All bounds checks can be implemented as one-sided bounds // checks by casting to unsigned for efficiency. Profile before doing this. TF_LITE_ENSURE(&context_, tensor_index < context_.tensors_size && tensor_index >= 0); - TfLiteIntArray* dims_lite = ConvertVectorToTfLiteIntArray(dims); - return ResizeTensorImpl(&context_.tensors[tensor_index], dims_lite); + TfLiteTensor* tensor = &context_.tensors[tensor_index]; + + // Short-circuit the state change if the dimensions don't change, avoiding + // unnecessary (re)allocations. + if (EqualArrayAndTfLiteIntArray(tensor->dims, dims.size(), dims.data())) { + return kTfLiteOk; + } + + state_ = kStateUninvokable; + return ResizeTensorImpl(tensor, ConvertVectorToTfLiteIntArray(dims)); } -// Returns true if at least one tensor in the given list is kTfLiteDynamic. bool HasDynamicTensor(const TfLiteContext& context, - const TfLiteIntArray* tensors) { - for (int i = 0; i < tensors->size; ++i) { - const TfLiteTensor& tensor = context.tensors[tensors->data[i]]; - if (tensor.allocation_type == kTfLiteDynamic) { - return true; - } - } - return false; + const TfLiteIntArray* int_array) { + return HasDynamicTensorImpl(context, TfLiteIntArrayView{int_array}); } TfLiteStatus Interpreter::PrepareOpsStartingAt( @@ -510,6 +572,8 @@ TfLiteStatus Interpreter::PrepareOpsStartingAt( nodes_and_registration_[node_index].second; EnsureTensorsVectorCapacity(); if (OpPrepare(registration, &node) == kTfLiteError) { + context_.ReportError(&context_, "Node %d failed to prepare.\n", + node_index); return kTfLiteError; } @@ -528,7 +592,8 @@ TfLiteStatus Interpreter::PrepareOpsStartingAt( TfLiteStatus Interpreter::PrepareOpsAndTensors() { if (!memory_planner_) { memory_planner_.reset(new ArenaPlanner( - &context_, std::unique_ptr(new InterpreterInfo(this)))); + &context_, std::unique_ptr(new InterpreterInfo(this)), + /*preserve_inputs=*/true)); memory_planner_->PlanAllocations(); } @@ -554,6 +619,7 @@ TfLiteStatus Interpreter::Invoke() { } TfLiteStatus status = kTfLiteOk; +#ifndef TFLITE_MCU if (nnapi_delegate_) { if (next_execution_plan_index_to_prepare_ == execution_plan_.size()) { TF_LITE_ENSURE_OK(&context_, nnapi_delegate_->Invoke(this)); @@ -567,6 +633,7 @@ TfLiteStatus Interpreter::Invoke() { return kTfLiteError; } } +#endif // Invocations are always done in node order. // Note that calling Invoke repeatedly will cause the original memory plan to @@ -607,6 +674,8 @@ TfLiteStatus Interpreter::Invoke() { EnsureTensorsVectorCapacity(); tensor_resized_since_op_invoke_ = false; if (OpInvoke(registration, &node) == kTfLiteError) { + context_.ReportError(&context_, "Node %d failed to invoke.\n", + node_index); status = kTfLiteError; } @@ -823,6 +892,7 @@ TfLiteStatus Interpreter::ResizeTensorImpl(TfLiteTensor* tensor, } void Interpreter::UseNNAPI(bool enable) { +#ifndef TFLITE_MCU // TODO(aselle): This is a workaround for finding if NNAPI exists. // We also need to make sure getLibraryHandle() is renamed to be NNAPI // prefixed. @@ -832,15 +902,18 @@ void Interpreter::UseNNAPI(bool enable) { } else if (!nnapi_delegate_) { nnapi_delegate_.reset(new NNAPIDelegate); } +#endif } void Interpreter::SetNumThreads(int num_threads) { context_.recommended_num_threads = num_threads; - // TODO(ahentz): find a way to avoid this. It causes gemmlowp and eigen to - // be required in order to compile the framework. - gemm_support::SetNumThreads(&context_, num_threads); - eigen_support::SetNumThreads(&context_, num_threads); + for (int i = 0; i < kTfLiteMaxExternalContexts; ++i) { + auto* c = external_contexts_[i]; + if (c && c->Refresh) { + c->Refresh(&context_); + } + } } TfLiteStatus Interpreter::ModifyGraphWithDelegate(TfLiteDelegate* delegate, @@ -884,9 +957,10 @@ TfLiteStatus Interpreter::ModifyGraphWithDelegate(TfLiteDelegate* delegate, TF_LITE_ENSURE_OK(&context_, status); if (!allow_dynamic_tensors) { + // Reset the state to force tensor/op reallocation. + state_ = kStateUninvokable; TF_LITE_ENSURE_OK(&context_, AllocateTensors()); - TF_LITE_ENSURE(&context_, state_ == kStateInvokable || - state_ == kStateInvokableAndImmutable); + TF_LITE_ENSURE_EQ(&context_, state_, kStateInvokable); // After using a delegate which doesn't support dynamic tensors, make the // entire graph immutable. state_ = kStateInvokableAndImmutable; diff --git a/tensorflow/contrib/lite/interpreter.h b/tensorflow/contrib/lite/interpreter.h index e67543671bbb0c9c53fae86a65d2cd7ae4280332..1a1c3e272b8d1c90cda2129461a29bd5c266cb21 100644 --- a/tensorflow/contrib/lite/interpreter.h +++ b/tensorflow/contrib/lite/interpreter.h @@ -17,6 +17,7 @@ limitations under the License. #ifndef TENSORFLOW_CONTRIB_LITE_INTERPRETER_H_ #define TENSORFLOW_CONTRIB_LITE_INTERPRETER_H_ +#include #include #include #include @@ -58,6 +59,14 @@ template <> constexpr TfLiteType typeToTfLiteType() { return kTfLiteBool; } +template <> +constexpr TfLiteType typeToTfLiteType>() { + return kTfLiteComplex64; +} +template <> +constexpr TfLiteType typeToTfLiteType() { + return kTfLiteString; +} // Forward declare since NNAPIDelegate uses Interpreter. class NNAPIDelegate; @@ -405,6 +414,8 @@ class Interpreter { } private: + friend class InterpreterTest; + // Give 'op_reg' a chance to initialize itself using the contents of // 'buffer'. void* OpInit(const TfLiteRegistration& op_reg, const char* buffer, @@ -517,6 +528,18 @@ class Interpreter { static TfLiteStatus GetExecutionPlan(struct TfLiteContext* context, TfLiteIntArray** execution_plan); + // Retrieve an existing external context by type. + TfLiteExternalContext* GetExternalContext(TfLiteExternalContextType type); + static TfLiteExternalContext* GetExternalContext( + struct TfLiteContext* context, TfLiteExternalContextType type); + + // Set the value of an external context. + void SetExternalContext(TfLiteExternalContextType type, + TfLiteExternalContext* ctx); + static void SetExternalContext(struct TfLiteContext* context, + TfLiteExternalContextType type, + TfLiteExternalContext* ctx); + // Ensures that `tensors_` has at least `kTensorsCapacityHeadroom` extra // capacity. Calling this function may invalidate existing pointers to // tensors. After calling this function, adding `kTensorsCapacityHeadroom` @@ -607,6 +630,9 @@ class Interpreter { // Profiler for this interpreter instance. profiling::Profiler* profiler_; + + // List of active external contexts. + TfLiteExternalContext* external_contexts_[kTfLiteMaxExternalContexts]; }; } // namespace tflite diff --git a/tensorflow/contrib/lite/interpreter_test.cc b/tensorflow/contrib/lite/interpreter_test.cc index 21cdf87d1e421868d1b62c5e23c2481cfbb4c989..10119903fed448bd44f408efb495831216dc594c 100644 --- a/tensorflow/contrib/lite/interpreter_test.cc +++ b/tensorflow/contrib/lite/interpreter_test.cc @@ -23,6 +23,15 @@ limitations under the License. #include "tensorflow/contrib/lite/testing/util.h" namespace tflite { + +// InterpreterTest is a friend of Interpreter, so it can access context_. +class InterpreterTest : public ::testing::Test { + protected: + TfLiteContext* GetInterpreterContext() { return &interpreter_.context_; } + + Interpreter interpreter_; +}; + namespace ops { namespace builtin { TfLiteRegistration* Register_PADV2(); @@ -48,6 +57,22 @@ TEST(BasicInterpreter, InvokeInvalidModel) { ASSERT_EQ(interpreter.Invoke(), kTfLiteOk); } +TEST(BasicInterpreter, TestAllocateTensorsResetVariableTensors) { + Interpreter interpreter; + int tensor_index; + ASSERT_EQ(interpreter.AddTensors(1, &tensor_index), kTfLiteOk); + constexpr int kTensorSize = 16; + interpreter.SetTensorParametersReadWrite(tensor_index, kTfLiteFloat32, "", + {kTensorSize}, {}, true); + interpreter.SetVariables({tensor_index}); + ASSERT_EQ(interpreter.AllocateTensors(), kTfLiteOk); + TfLiteTensor* tensor = interpreter.tensor(tensor_index); + // Ensure that variable tensors are reset to zero. + for (int i = 0; i < kTensorSize; ++i) { + ASSERT_EQ(tensor->data.f[i], 0.0f); + } +} + // Test size accessor functions. TEST(BasicInterpreter, TestSizeFunctions) { Interpreter interpreter; @@ -231,32 +256,16 @@ TEST(BasicInterpreter, CheckArenaAllocation) { ASSERT_EQ(interpreter.AllocateTensors(), kTfLiteOk); - ASSERT_EQ(interpreter.tensor(0)->data.raw, interpreter.tensor(4)->data.raw); - ASSERT_EQ(interpreter.tensor(1)->data.raw, interpreter.tensor(7)->data.raw); - ASSERT_EQ(interpreter.tensor(8)->data.raw, nullptr); - - ASSERT_LT(interpreter.tensor(4)->data.raw, interpreter.tensor(1)->data.raw); - ASSERT_LT(interpreter.tensor(6)->data.raw, interpreter.tensor(1)->data.raw); ASSERT_LT(interpreter.tensor(0)->data.raw, interpreter.tensor(1)->data.raw); - - ASSERT_LT(interpreter.tensor(0)->data.raw, interpreter.tensor(3)->data.raw); - ASSERT_LT(interpreter.tensor(1)->data.raw, interpreter.tensor(3)->data.raw); + ASSERT_LT(interpreter.tensor(1)->data.raw, interpreter.tensor(2)->data.raw); ASSERT_LT(interpreter.tensor(2)->data.raw, interpreter.tensor(3)->data.raw); - ASSERT_LT(interpreter.tensor(4)->data.raw, interpreter.tensor(3)->data.raw); - ASSERT_LT(interpreter.tensor(6)->data.raw, interpreter.tensor(3)->data.raw); - ASSERT_LT(interpreter.tensor(7)->data.raw, interpreter.tensor(3)->data.raw); - ASSERT_LT(interpreter.tensor(8)->data.raw, interpreter.tensor(3)->data.raw); - ASSERT_LT(interpreter.tensor(9)->data.raw, interpreter.tensor(3)->data.raw); - - ASSERT_LT(interpreter.tensor(0)->data.raw, interpreter.tensor(5)->data.raw); - ASSERT_LT(interpreter.tensor(1)->data.raw, interpreter.tensor(5)->data.raw); - ASSERT_LT(interpreter.tensor(2)->data.raw, interpreter.tensor(5)->data.raw); - ASSERT_LT(interpreter.tensor(3)->data.raw, interpreter.tensor(5)->data.raw); + ASSERT_LT(interpreter.tensor(3)->data.raw, interpreter.tensor(4)->data.raw); ASSERT_LT(interpreter.tensor(4)->data.raw, interpreter.tensor(5)->data.raw); - ASSERT_LT(interpreter.tensor(6)->data.raw, interpreter.tensor(5)->data.raw); - ASSERT_LT(interpreter.tensor(7)->data.raw, interpreter.tensor(5)->data.raw); - ASSERT_LT(interpreter.tensor(8)->data.raw, interpreter.tensor(5)->data.raw); - ASSERT_LT(interpreter.tensor(9)->data.raw, interpreter.tensor(5)->data.raw); + ASSERT_LT(interpreter.tensor(5)->data.raw, interpreter.tensor(7)->data.raw); + ASSERT_EQ(interpreter.tensor(6)->data.raw, interpreter.tensor(2)->data.raw); + // #7 is the one with the largest pointer. + ASSERT_EQ(interpreter.tensor(8)->data.raw, nullptr); + ASSERT_EQ(interpreter.tensor(9)->data.raw, interpreter.tensor(5)->data.raw); } TEST(BasicInterpreter, BufferAccess) { @@ -292,6 +301,57 @@ TEST(BasicInterpreter, NoOpInterpreter) { ASSERT_EQ(interpreter.Invoke(), kTfLiteOk); } +TEST(BasicInterpreter, RedundantAllocateTensors) { + Interpreter interpreter; + ASSERT_EQ(interpreter.AddTensors(1), kTfLiteOk); + ASSERT_EQ(interpreter.SetInputs({0}), kTfLiteOk); + + ASSERT_EQ(interpreter.SetTensorParametersReadWrite( + 0, kTfLiteFloat32, "", {3}, TfLiteQuantizationParams()), + kTfLiteOk); + + ASSERT_EQ(interpreter.AllocateTensors(), kTfLiteOk); + const auto data_raw = interpreter.tensor(0)->data.raw; + ASSERT_NE(data_raw, nullptr); + + // A redundant allocation request should have no impact. + ASSERT_EQ(interpreter.AllocateTensors(), kTfLiteOk); + ASSERT_EQ(interpreter.tensor(0)->data.raw, data_raw); +} + +TEST(BasicInterpreter, RedundantAllocateTensorsWithDynamicInputs) { + Interpreter interpreter; + TfLiteRegistration reg = {nullptr, nullptr, nullptr, nullptr}; + ASSERT_EQ(interpreter.AddTensors(2), kTfLiteOk); + interpreter.SetInputs({0}); + interpreter.SetOutputs({1}); + interpreter.AddNodeWithParameters({0}, {1}, nullptr, 0, nullptr, ®); + + ASSERT_EQ(interpreter.SetTensorParametersReadWrite( + 0, kTfLiteFloat32, "", {3}, TfLiteQuantizationParams()), + kTfLiteOk); + ASSERT_EQ(interpreter.SetTensorParametersReadWrite( + 1, kTfLiteFloat32, "", {3}, TfLiteQuantizationParams()), + kTfLiteOk); + + // Configure the input tensor as dynamic. + interpreter.tensor(0)->data.raw = nullptr; + interpreter.tensor(0)->allocation_type = kTfLiteDynamic; + + ASSERT_EQ(interpreter.ResizeInputTensor(interpreter.inputs()[0], {1, 2, 3}), + kTfLiteOk); + ASSERT_EQ(interpreter.AllocateTensors(), kTfLiteOk); + ASSERT_NE(interpreter.tensor(1)->data.raw, nullptr); + + // Reset the output tensor's buffer. + interpreter.tensor(1)->data.raw = nullptr; + + // A redundant allocation request should be honored, as the input tensor + // was marked dynamic. + ASSERT_EQ(interpreter.AllocateTensors(), kTfLiteOk); + ASSERT_NE(interpreter.tensor(1)->data.raw, nullptr); +} + TEST(BasicInterpreter, ResizingTensors) { Interpreter interpreter; ASSERT_EQ(interpreter.AddTensors(1), kTfLiteOk); @@ -349,6 +409,37 @@ TEST(BasicInterpreter, ResizingTensors) { tensor->data.f[15] = 0.123f; } +TEST(BasicInterpreter, NoopResizingTensors) { + Interpreter interpreter; + ASSERT_EQ(interpreter.AddTensors(1), kTfLiteOk); + ASSERT_EQ(interpreter.SetInputs({0}), kTfLiteOk); + ASSERT_EQ(interpreter.SetOutputs({0}), kTfLiteOk); + + ASSERT_EQ(interpreter.SetTensorParametersReadWrite( + 0, kTfLiteFloat32, "", {3}, TfLiteQuantizationParams()), + kTfLiteOk); + + int t = interpreter.inputs()[0]; + TfLiteTensor* tensor = interpreter.tensor(t); + + ASSERT_EQ(interpreter.ResizeInputTensor(t, {1, 2, 3}), kTfLiteOk); + EXPECT_EQ(tensor->bytes, 6 * sizeof(float)); + ASSERT_EQ(interpreter.AllocateTensors(), kTfLiteOk); + tensor->data.f[5] = 0.123f; + + // Resizing to the same size should not trigger re-allocation. + ASSERT_EQ(interpreter.ResizeInputTensor(t, {1, 2, 3}), kTfLiteOk); + EXPECT_EQ(tensor->bytes, 6 * sizeof(float)); + ASSERT_NE(tensor->data.raw, nullptr); + ASSERT_EQ(tensor->data.f[5], 0.123f); + + // Explicitly allocating should be a no-op, as no resize was performed. + ASSERT_EQ(interpreter.AllocateTensors(), kTfLiteOk); + EXPECT_EQ(tensor->bytes, 6 * sizeof(float)); + ASSERT_NE(tensor->data.raw, nullptr); + ASSERT_EQ(tensor->data.f[5], 0.123f); +} + TEST(BasicInterpreter, OneOpInterpreter) { Interpreter interpreter; ASSERT_EQ(interpreter.AddTensors(2), kTfLiteOk); @@ -714,6 +805,47 @@ TEST(InterpreterTensorsCapacityTest, TestExceedHeadroom) { ASSERT_EQ(interpreter.AllocateTensors(), kTfLiteOk); } +struct TestExternalContext : public TfLiteExternalContext { + static const TfLiteExternalContextType kType = kTfLiteGemmLowpContext; + + static TestExternalContext* Get(TfLiteContext* context) { + return reinterpret_cast( + context->GetExternalContext(context, kType)); + } + + static void Set(TfLiteContext* context, TestExternalContext* value) { + context->SetExternalContext(context, kType, value); + } + + int num_refreshes = 0; +}; + +TEST_F(InterpreterTest, GetSetResetExternalContexts) { + auto* context = GetInterpreterContext(); + + TestExternalContext external_context; + external_context.Refresh = [](TfLiteContext* context) { + auto* ptr = TestExternalContext::Get(context); + if (ptr != nullptr) { + ++ptr->num_refreshes; + } + return kTfLiteOk; + }; + + EXPECT_EQ(TestExternalContext::Get(context), nullptr); + interpreter_.SetNumThreads(4); + + TestExternalContext::Set(context, &external_context); + EXPECT_EQ(TestExternalContext::Get(context), &external_context); + interpreter_.SetNumThreads(4); + interpreter_.SetNumThreads(5); + EXPECT_EQ(external_context.num_refreshes, 2); + + TestExternalContext::Set(context, nullptr); + EXPECT_EQ(TestExternalContext::Get(context), nullptr); + interpreter_.SetNumThreads(4); +} + // Test fixture that allows playing with execution plans. It creates a two // node graph that can be executed in either [0,1] order or [1,0] order. // The CopyOp records when it is invoked in the class member run_order_ diff --git a/tensorflow/contrib/lite/java/demo/app/build.gradle b/tensorflow/contrib/lite/java/demo/app/build.gradle index 44ea2dcd908644bcfc637f71573ce722adaf6935..49868c5a7566c8c537ac2ae9e0a4acc2c872ecbf 100644 --- a/tensorflow/contrib/lite/java/demo/app/build.gradle +++ b/tensorflow/contrib/lite/java/demo/app/build.gradle @@ -5,11 +5,12 @@ android { buildToolsVersion "26.0.1" defaultConfig { applicationId "android.example.com.tflitecamerademo" - minSdkVersion 15 + // Required by Camera2 API. + minSdkVersion 21 targetSdkVersion 26 versionCode 1 versionName "1.0" - testInstrumentationRunner "android.support.test.runner.AndroidJUnitRunner" + testInstrumentationRunner "androidx.test.runner.AndroidJUnitRunner" // Remove this block. jackOptions { @@ -43,7 +44,7 @@ repositories { dependencies { compile fileTree(dir: 'libs', include: ['*.jar']) - androidTestCompile('com.android.support.test.espresso:espresso-core:2.2.2', { + androidTestCompile('com.androidx.test.espresso:espresso-core:2.2.2', { exclude group: 'com.android.support', module: 'support-annotations' }) compile 'com.android.support:appcompat-v7:25.2.0' @@ -91,4 +92,4 @@ class DownloadUrlTask extends DefaultTask { void download() { ant.get(src: sourceUrl, dest: target) } -} \ No newline at end of file +} diff --git a/tensorflow/contrib/lite/java/ovic/demo/app/build.gradle b/tensorflow/contrib/lite/java/ovic/demo/app/build.gradle index c5d19bad89a93988a6830a17fe2fb4a60e2fb00f..3f32d62e5c08419c6413fffe09b64356edcac836 100644 --- a/tensorflow/contrib/lite/java/ovic/demo/app/build.gradle +++ b/tensorflow/contrib/lite/java/ovic/demo/app/build.gradle @@ -9,7 +9,7 @@ android { targetSdkVersion 26 versionCode 1 versionName "1.0" - testInstrumentationRunner "android.support.test.runner.AndroidJUnitRunner" + testInstrumentationRunner "androidx.test.runner.AndroidJUnitRunner" // Remove this block. jackOptions { @@ -43,7 +43,7 @@ repositories { dependencies { compile fileTree(dir: 'libs', include: ['*.jar']) - androidTestCompile('com.android.support.test.espresso:espresso-core:2.2.2', { + androidTestCompile('com.androidx.test.espresso:espresso-core:2.2.2', { exclude group: 'com.android.support', module: 'support-annotations' }) compile 'com.android.support:appcompat-v7:25.2.0' diff --git a/tensorflow/contrib/lite/java/ovic/src/test/java/org/tensorflow/ovic/OvicClassifierTest.java b/tensorflow/contrib/lite/java/ovic/src/test/java/org/tensorflow/ovic/OvicClassifierTest.java index 56f3e7604a5b172e907edbe862b017957594397f..1587c3c56f45c0baddfa75286c979fe0c0edffcc 100644 --- a/tensorflow/contrib/lite/java/ovic/src/test/java/org/tensorflow/ovic/OvicClassifierTest.java +++ b/tensorflow/contrib/lite/java/ovic/src/test/java/org/tensorflow/ovic/OvicClassifierTest.java @@ -127,12 +127,8 @@ public final class OvicClassifierTest { try { testResult = classifier.classifyByteBuffer(testImage); fail(); - } catch (RuntimeException e) { - assertThat(e) - .hasMessageThat() - .contains( - "Failed to get input dimensions. 0-th input should have 49152 bytes, " - + "but found 150528 bytes."); + } catch (IllegalArgumentException e) { + // Success. } } diff --git a/tensorflow/contrib/lite/java/src/main/java/org/tensorflow/lite/DataType.java b/tensorflow/contrib/lite/java/src/main/java/org/tensorflow/lite/DataType.java index 75334cd96e8daadc356dadea063eee30ef6d5245..94a1ec65d64b6493cdb309fc0c19155eb9cb26cb 100644 --- a/tensorflow/contrib/lite/java/src/main/java/org/tensorflow/lite/DataType.java +++ b/tensorflow/contrib/lite/java/src/main/java/org/tensorflow/lite/DataType.java @@ -27,10 +27,7 @@ enum DataType { UINT8(3), /** 64-bit signed integer. */ - INT64(4), - - /** A {@link ByteBuffer}. */ - BYTEBUFFER(999); + INT64(4); private final int value; @@ -69,8 +66,6 @@ enum DataType { return 1; case INT64: return 8; - case BYTEBUFFER: - return 1; } throw new IllegalArgumentException( "DataType error: DataType " + this + " is not supported yet"); @@ -87,8 +82,6 @@ enum DataType { return "byte"; case INT64: return "long"; - case BYTEBUFFER: - return "ByteBuffer"; } throw new IllegalArgumentException( "DataType error: DataType " + this + " is not supported yet"); diff --git a/tensorflow/contrib/lite/java/src/main/java/org/tensorflow/lite/Interpreter.java b/tensorflow/contrib/lite/java/src/main/java/org/tensorflow/lite/Interpreter.java index fd1f0ffa68eeca7b5866b146ecaa1f9216ef377d..7002f826775b216e0a27ebe00f30680c9ce362bb 100644 --- a/tensorflow/contrib/lite/java/src/main/java/org/tensorflow/lite/Interpreter.java +++ b/tensorflow/contrib/lite/java/src/main/java/org/tensorflow/lite/Interpreter.java @@ -135,7 +135,8 @@ public final class Interpreter implements AutoCloseable { * including int, float, long, and byte. {@link ByteBuffer} is the preferred way to pass large * input data. When {@link ByteBuffer} is used, its content should remain unchanged until * model inference is done. - * @param output a multidimensional array of output data. + * @param output a multidimensional array of output data, or a {@link ByteBuffer} of primitive + * types including int, float, long, and byte. */ public void run(@NonNull Object input, @NonNull Object output) { Object[] inputs = {input}; @@ -155,28 +156,16 @@ public final class Interpreter implements AutoCloseable { * primitive types including int, float, long, and byte. {@link ByteBuffer} is the preferred * way to pass large input data. When {@link ByteBuffer} is used, its content should remain * unchanged until model inference is done. - * @param outputs a map mapping output indices to multidimensional arrays of output data. It only - * needs to keep entries for the outputs to be used. + * @param outputs a map mapping output indices to multidimensional arrays of output data or {@link + * ByteBuffer}s of primitive types including int, float, long, and byte. It only needs to keep + * entries for the outputs to be used. */ public void runForMultipleInputsOutputs( @NonNull Object[] inputs, @NonNull Map outputs) { if (wrapper == null) { throw new IllegalStateException("Internal error: The Interpreter has already been closed."); } - Tensor[] tensors = wrapper.run(inputs); - if (outputs == null || tensors == null || outputs.size() > tensors.length) { - throw new IllegalArgumentException("Output error: Outputs do not match with model outputs."); - } - final int size = tensors.length; - for (Integer idx : outputs.keySet()) { - if (idx == null || idx < 0 || idx >= size) { - throw new IllegalArgumentException( - String.format( - "Output error: Invalid index of output %d (should be in range [0, %d))", - idx, size)); - } - tensors[idx].copyTo(outputs.get(idx)); - } + wrapper.run(inputs, outputs); } /** @@ -249,8 +238,10 @@ public final class Interpreter implements AutoCloseable { /** Release resources associated with the {@code Interpreter}. */ @Override public void close() { - wrapper.close(); - wrapper = null; + if (wrapper != null) { + wrapper.close(); + wrapper = null; + } } @Override diff --git a/tensorflow/contrib/lite/java/src/main/java/org/tensorflow/lite/NativeInterpreterWrapper.java b/tensorflow/contrib/lite/java/src/main/java/org/tensorflow/lite/NativeInterpreterWrapper.java index 80de88b6a1cd75b033e116f76f5612ee66e48f03..767a220f8cd5381ce10e044553317b1cb05ba17b 100644 --- a/tensorflow/contrib/lite/java/src/main/java/org/tensorflow/lite/NativeInterpreterWrapper.java +++ b/tensorflow/contrib/lite/java/src/main/java/org/tensorflow/lite/NativeInterpreterWrapper.java @@ -15,10 +15,10 @@ limitations under the License. package org.tensorflow.lite; -import java.lang.reflect.Array; import java.nio.ByteBuffer; import java.nio.ByteOrder; import java.nio.MappedByteBuffer; +import java.util.Arrays; import java.util.HashMap; import java.util.Map; @@ -40,6 +40,8 @@ final class NativeInterpreterWrapper implements AutoCloseable { modelHandle = createModel(modelPath, errorHandle); interpreterHandle = createInterpreter(modelHandle, errorHandle, numThreads); isMemoryAllocated = true; + inputTensors = new Tensor[getInputCount(interpreterHandle)]; + outputTensors = new Tensor[getOutputCount(interpreterHandle)]; } /** @@ -72,6 +74,8 @@ final class NativeInterpreterWrapper implements AutoCloseable { modelHandle = createModelWithBuffer(modelByteBuffer, errorHandle); interpreterHandle = createInterpreter(modelHandle, errorHandle, numThreads); isMemoryAllocated = true; + inputTensors = new Tensor[getInputCount(interpreterHandle)]; + outputTensors = new Tensor[getOutputCount(interpreterHandle)]; } /** Releases resources associated with this {@code NativeInterpreterWrapper}. */ @@ -85,75 +89,63 @@ final class NativeInterpreterWrapper implements AutoCloseable { inputsIndexes = null; outputsIndexes = null; isMemoryAllocated = false; + Arrays.fill(inputTensors, null); + Arrays.fill(outputTensors, null); } /** Sets inputs, runs model inference and returns outputs. */ - Tensor[] run(Object[] inputs) { + void run(Object[] inputs, Map outputs) { + inferenceDurationNanoseconds = -1; if (inputs == null || inputs.length == 0) { throw new IllegalArgumentException("Input error: Inputs should not be null or empty."); } - int[] dataTypes = new int[inputs.length]; - Object[] sizes = new Object[inputs.length]; - int[] numsOfBytes = new int[inputs.length]; + if (outputs == null || outputs.isEmpty()) { + throw new IllegalArgumentException("Input error: Outputs should not be null or empty."); + } + + // TODO(b/80431971): Remove implicit resize after deprecating multi-dimensional array inputs. + // Rather than forcing an immediate resize + allocation if an input's shape differs, we first + // flush all resizes, avoiding redundant allocations. for (int i = 0; i < inputs.length; ++i) { - DataType dataType = dataTypeOf(inputs[i]); - dataTypes[i] = dataType.getNumber(); - if (dataType == DataType.BYTEBUFFER) { - ByteBuffer buffer = (ByteBuffer) inputs[i]; - if (buffer == null || !buffer.isDirect() || buffer.order() != ByteOrder.nativeOrder()) { - throw new IllegalArgumentException( - "Input error: ByteBuffer should be a direct ByteBuffer that uses " - + "ByteOrder.nativeOrder()."); - } - numsOfBytes[i] = buffer.limit(); - sizes[i] = getInputDims(interpreterHandle, i, numsOfBytes[i]); - } else if (isNonEmptyArray(inputs[i])) { - int[] dims = shapeOf(inputs[i]); - sizes[i] = dims; - numsOfBytes[i] = dataType.elemByteSize() * numElements(dims); - } else { - throw new IllegalArgumentException( - String.format( - "Input error: %d-th element of the %d inputs is not an array or a ByteBuffer.", - i, inputs.length)); + Tensor tensor = getInputTensor(i); + int[] newShape = tensor.getInputShapeIfDifferent(inputs[i]); + if (newShape != null) { + resizeInput(i, newShape); } } - inferenceDurationNanoseconds = -1; - long[] outputsHandles = - run( - interpreterHandle, - errorHandle, - sizes, - dataTypes, - numsOfBytes, - inputs, - this, - isMemoryAllocated); - if (outputsHandles == null || outputsHandles.length == 0) { - throw new IllegalStateException("Internal error: Interpreter has no outputs."); + + if (!isMemoryAllocated) { + allocateTensors(interpreterHandle, errorHandle); + isMemoryAllocated = true; + // Allocation can trigger dynamic resizing of output tensors, so clear the + // output tensor cache. + Arrays.fill(outputTensors, null); } - isMemoryAllocated = true; - Tensor[] outputs = new Tensor[outputsHandles.length]; - for (int i = 0; i < outputsHandles.length; ++i) { - outputs[i] = Tensor.fromHandle(outputsHandles[i]); + + for (int i = 0; i < inputs.length; ++i) { + getInputTensor(i).setTo(inputs[i]); + } + + long inferenceStartNanos = System.nanoTime(); + run(interpreterHandle, errorHandle); + long inferenceDurationNanoseconds = System.nanoTime() - inferenceStartNanos; + + for (Map.Entry output : outputs.entrySet()) { + getOutputTensor(output.getKey()).copyTo(output.getValue()); } - return outputs; + + // Only set if the entire operation succeeds. + this.inferenceDurationNanoseconds = inferenceDurationNanoseconds; } - private static native long[] run( - long interpreterHandle, - long errorHandle, - Object[] sizes, - int[] dtypes, - int[] numsOfBytes, - Object[] values, - NativeInterpreterWrapper wrapper, - boolean memoryAllocated); + private static native boolean run(long interpreterHandle, long errorHandle); /** Resizes dimensions of a specific input. */ void resizeInput(int idx, int[] dims) { if (resizeInput(interpreterHandle, errorHandle, idx, dims)) { isMemoryAllocated = false; + // Resizing will invalidate the Tensor's shape, so invalidate the Tensor handle. + inputTensors[idx] = null; } } @@ -212,78 +204,6 @@ final class NativeInterpreterWrapper implements AutoCloseable { } } - static int numElements(int[] shape) { - if (shape == null) { - return 0; - } - int n = 1; - for (int i = 0; i < shape.length; i++) { - n *= shape[i]; - } - return n; - } - - static boolean isNonEmptyArray(Object o) { - return (o != null && o.getClass().isArray() && Array.getLength(o) != 0); - } - - /** Returns the type of the data. */ - static DataType dataTypeOf(Object o) { - if (o != null) { - Class c = o.getClass(); - while (c.isArray()) { - c = c.getComponentType(); - } - if (float.class.equals(c)) { - return DataType.FLOAT32; - } else if (int.class.equals(c)) { - return DataType.INT32; - } else if (byte.class.equals(c)) { - return DataType.UINT8; - } else if (long.class.equals(c)) { - return DataType.INT64; - } else if (ByteBuffer.class.isInstance(o)) { - return DataType.BYTEBUFFER; - } - } - throw new IllegalArgumentException( - "DataType error: cannot resolve DataType of " + o.getClass().getName()); - } - - /** Returns the shape of an object as an int array. */ - static int[] shapeOf(Object o) { - int size = numDimensions(o); - int[] dimensions = new int[size]; - fillShape(o, 0, dimensions); - return dimensions; - } - - static int numDimensions(Object o) { - if (o == null || !o.getClass().isArray()) { - return 0; - } - if (Array.getLength(o) == 0) { - throw new IllegalArgumentException("Array lengths cannot be 0."); - } - return 1 + numDimensions(Array.get(o, 0)); - } - - static void fillShape(Object o, int dim, int[] shape) { - if (shape == null || dim == shape.length) { - return; - } - final int len = Array.getLength(o); - if (shape[dim] == 0) { - shape[dim] = len; - } else if (shape[dim] != len) { - throw new IllegalArgumentException( - String.format("Mismatched lengths (%d and %d) in dimension %d", shape[dim], len, dim)); - } - for (int i = 0; i < len; ++i) { - fillShape(Array.get(o, i), dim + 1, shape); - } - } - /** * Gets the last inference duration in nanoseconds. It returns null if there is no previous * inference run or the last inference run failed. @@ -293,40 +213,55 @@ final class NativeInterpreterWrapper implements AutoCloseable { } /** - * Gets the dimensions of an input. It throws IllegalArgumentException if input index is invalid. + * Gets the quantization zero point of an output. + * + * @throws IllegalArgumentException if the output index is invalid. */ - int[] getInputDims(int index) { - return getInputDims(interpreterHandle, index, -1); + int getOutputQuantizationZeroPoint(int index) { + return getOutputQuantizationZeroPoint(interpreterHandle, index); } /** - * Gets the dimensions of an input. If numBytes >= 0, it will check whether num of bytes match the - * input. + * Gets the quantization scale of an output. + * + * @throws IllegalArgumentException if the output index is invalid. */ - private static native int[] getInputDims(long interpreterHandle, int inputIdx, int numBytes); - - /** Gets the type of an output. It throws IllegalArgumentException if output index is invalid. */ - String getOutputDataType(int index) { - int type = getOutputDataType(interpreterHandle, index); - return DataType.fromNumber(type).toStringName(); + float getOutputQuantizationScale(int index) { + return getOutputQuantizationScale(interpreterHandle, index); } /** - * Gets the quantization zero point of an output. + * Gets the input {@link Tensor} for the provided input index. * - * @throws IllegalArgumentExeption if the output index is invalid. + * @throws IllegalArgumentException if the input index is invalid. */ - int getOutputQuantizationZeroPoint(int index) { - return getOutputQuantizationZeroPoint(interpreterHandle, index); + Tensor getInputTensor(int index) { + if (index < 0 || index >= inputTensors.length) { + throw new IllegalArgumentException("Invalid input Tensor index: " + index); + } + Tensor inputTensor = inputTensors[index]; + if (inputTensor == null) { + inputTensor = + inputTensors[index] = Tensor.fromHandle(getInputTensor(interpreterHandle, index)); + } + return inputTensor; } /** - * Gets the quantization scale of an output. + * Gets the output {@link Tensor} for the provided output index. * - * @throws IllegalArgumentExeption if the output index is invalid. + * @throws IllegalArgumentException if the output index is invalid. */ - float getOutputQuantizationScale(int index) { - return getOutputQuantizationScale(interpreterHandle, index); + Tensor getOutputTensor(int index) { + if (index < 0 || index >= outputTensors.length) { + throw new IllegalArgumentException("Invalid output Tensor index: " + index); + } + Tensor outputTensor = outputTensors[index]; + if (outputTensor == null) { + outputTensor = + outputTensors[index] = Tensor.fromHandle(getOutputTensor(interpreterHandle, index)); + } + return outputTensor; } private static native int getOutputDataType(long interpreterHandle, int outputIdx); @@ -343,18 +278,30 @@ final class NativeInterpreterWrapper implements AutoCloseable { private long modelHandle; - private int inputSize; - private long inferenceDurationNanoseconds = -1; private ByteBuffer modelByteBuffer; + // Lazily constructed maps of input and output names to input and output Tensor indexes. private Map inputsIndexes; - private Map outputsIndexes; + // Lazily constructed and populated arrays of input and output Tensor wrappers. + private final Tensor[] inputTensors; + private final Tensor[] outputTensors; + private boolean isMemoryAllocated = false; + private static native long allocateTensors(long interpreterHandle, long errorHandle); + + private static native long getInputTensor(long interpreterHandle, int inputIdx); + + private static native long getOutputTensor(long interpreterHandle, int outputIdx); + + private static native int getInputCount(long interpreterHandle); + + private static native int getOutputCount(long interpreterHandle); + private static native String[] getInputNames(long interpreterHandle); private static native String[] getOutputNames(long interpreterHandle); diff --git a/tensorflow/contrib/lite/java/src/main/java/org/tensorflow/lite/Tensor.java b/tensorflow/contrib/lite/java/src/main/java/org/tensorflow/lite/Tensor.java index 09e887aae3339e9f114c07d689c0d7b5e2fc384b..2403570c527e762f6782e313731e383feeeef46d 100644 --- a/tensorflow/contrib/lite/java/src/main/java/org/tensorflow/lite/Tensor.java +++ b/tensorflow/contrib/lite/java/src/main/java/org/tensorflow/lite/Tensor.java @@ -15,6 +15,9 @@ limitations under the License. package org.tensorflow.lite; +import java.lang.reflect.Array; +import java.nio.ByteBuffer; +import java.nio.ByteOrder; import java.util.Arrays; /** @@ -29,30 +32,179 @@ final class Tensor { return new Tensor(nativeHandle); } - /** Reads Tensor content into an array. */ - T copyTo(T dst) { - if (NativeInterpreterWrapper.dataTypeOf(dst) != dtype) { + /** Returns the {@link DataType} of elements stored in the Tensor. */ + public DataType dataType() { + return dtype; + } + + /** Returns the size, in bytes, of the tensor data. */ + public int numBytes() { + return numBytes(nativeHandle); + } + + /** + * Returns the shape of + * the Tensor, i.e., the sizes of each dimension. + * + * @return an array where the i-th element is the size of the i-th dimension of the tensor. + */ + public int[] shape() { + return shapeCopy; + } + + /** + * Copies the contents of the provided {@code src} object to the Tensor. + * + *

The {@code src} should either be a (multi-dimensional) array with a shape matching that of + * this tensor, or a {@link ByteByffer} of compatible primitive type with a matching flat size. + * + * @throws IllegalArgumentException if the tensor is a scalar or if {@code src} is not compatible + * with the tensor (for example, mismatched data types or shapes). + */ + void setTo(Object src) { + throwExceptionIfTypeIsIncompatible(src); + if (isByteBuffer(src)) { + ByteBuffer srcBuffer = (ByteBuffer) src; + // For direct ByteBuffer instances we support zero-copy. Note that this assumes the caller + // retains ownership of the source buffer until inference has completed. + if (srcBuffer.isDirect() && srcBuffer.order() == ByteOrder.nativeOrder()) { + writeDirectBuffer(nativeHandle, srcBuffer); + } else { + buffer().put(srcBuffer); + } + return; + } + writeMultiDimensionalArray(nativeHandle, src); + } + + /** + * Copies the contents of the tensor to {@code dst} and returns {@code dst}. + * + * @param dst the destination buffer, either an explicitly-typed array or a {@link ByteBuffer}. + * @throws IllegalArgumentException if {@code dst} is not compatible with the tensor (for example, + * mismatched data types or shapes). + */ + Object copyTo(Object dst) { + throwExceptionIfTypeIsIncompatible(dst); + if (dst instanceof ByteBuffer) { + ByteBuffer dstByteBuffer = (ByteBuffer) dst; + dstByteBuffer.put(buffer()); + return dst; + } + readMultiDimensionalArray(nativeHandle, dst); + return dst; + } + + /** Returns the provided buffer's shape if specified and different from this Tensor's shape. */ + // TODO(b/80431971): Remove this method after deprecating multi-dimensional array inputs. + int[] getInputShapeIfDifferent(Object input) { + // Implicit resizes based on ByteBuffer capacity isn't supported, so short-circuit that path. + // The ByteBuffer's size will be validated against this Tensor's size in {@link #setTo(Object)}. + if (isByteBuffer(input)) { + return null; + } + int[] inputShape = shapeOf(input); + if (Arrays.equals(shapeCopy, inputShape)) { + return null; + } + return inputShape; + } + + /** Returns the type of the data. */ + static DataType dataTypeOf(Object o) { + if (o != null) { + Class c = o.getClass(); + while (c.isArray()) { + c = c.getComponentType(); + } + if (float.class.equals(c)) { + return DataType.FLOAT32; + } else if (int.class.equals(c)) { + return DataType.INT32; + } else if (byte.class.equals(c)) { + return DataType.UINT8; + } else if (long.class.equals(c)) { + return DataType.INT64; + } + } + throw new IllegalArgumentException( + "DataType error: cannot resolve DataType of " + o.getClass().getName()); + } + + /** Returns the shape of an object as an int array. */ + static int[] shapeOf(Object o) { + int size = numDimensions(o); + int[] dimensions = new int[size]; + fillShape(o, 0, dimensions); + return dimensions; + } + + /** Returns the number of dimensions of a multi-dimensional array, otherwise 0. */ + static int numDimensions(Object o) { + if (o == null || !o.getClass().isArray()) { + return 0; + } + if (Array.getLength(o) == 0) { + throw new IllegalArgumentException("Array lengths cannot be 0."); + } + return 1 + numDimensions(Array.get(o, 0)); + } + + /** Recursively populates the shape dimensions for a given (multi-dimensional) array. */ + static void fillShape(Object o, int dim, int[] shape) { + if (shape == null || dim == shape.length) { + return; + } + final int len = Array.getLength(o); + if (shape[dim] == 0) { + shape[dim] = len; + } else if (shape[dim] != len) { + throw new IllegalArgumentException( + String.format("Mismatched lengths (%d and %d) in dimension %d", shape[dim], len, dim)); + } + for (int i = 0; i < len; ++i) { + fillShape(Array.get(o, i), dim + 1, shape); + } + } + + private void throwExceptionIfTypeIsIncompatible(Object o) { + if (isByteBuffer(o)) { + ByteBuffer oBuffer = (ByteBuffer) o; + if (oBuffer.capacity() != numBytes()) { + throw new IllegalArgumentException( + String.format( + "Cannot convert between a TensorFlowLite buffer with %d bytes and a " + + "ByteBuffer with %d bytes.", + numBytes(), oBuffer.capacity())); + } + return; + } + DataType oType = dataTypeOf(o); + if (oType != dtype) { throw new IllegalArgumentException( String.format( - "Output error: Cannot convert an TensorFlowLite tensor with type %s to a Java " - + "object of type %s (which is compatible with the TensorFlowLite type %s)", - dtype, dst.getClass().getName(), NativeInterpreterWrapper.dataTypeOf(dst))); + "Cannot convert between a TensorFlowLite tensor with type %s and a Java " + + "object of type %s (which is compatible with the TensorFlowLite type %s).", + dtype, o.getClass().getName(), oType)); } - int[] dstShape = NativeInterpreterWrapper.shapeOf(dst); - if (!Arrays.equals(dstShape, shapeCopy)) { + + int[] oShape = shapeOf(o); + if (!Arrays.equals(oShape, shapeCopy)) { throw new IllegalArgumentException( String.format( - "Output error: Shape of output target %s does not match with the shape of the " - + "Tensor %s.", - Arrays.toString(dstShape), Arrays.toString(shapeCopy))); + "Cannot copy between a TensorFlowLite tensor with shape %s and a Java object " + + "with shape %s.", + Arrays.toString(shapeCopy), Arrays.toString(oShape))); } - readMultiDimensionalArray(nativeHandle, dst); - return dst; } - final long nativeHandle; - final DataType dtype; - final int[] shapeCopy; + private static boolean isByteBuffer(Object o) { + return o instanceof ByteBuffer; + } + + private final long nativeHandle; + private final DataType dtype; + private final int[] shapeCopy; private Tensor(long nativeHandle) { this.nativeHandle = nativeHandle; @@ -60,11 +212,23 @@ final class Tensor { this.shapeCopy = shape(nativeHandle); } + private ByteBuffer buffer() { + return buffer(nativeHandle).order(ByteOrder.nativeOrder()); + } + + private static native ByteBuffer buffer(long handle); + + private static native void writeDirectBuffer(long handle, ByteBuffer src); + private static native int dtype(long handle); private static native int[] shape(long handle); - private static native void readMultiDimensionalArray(long handle, Object value); + private static native int numBytes(long handle); + + private static native void readMultiDimensionalArray(long handle, Object dst); + + private static native void writeMultiDimensionalArray(long handle, Object src); static { TensorFlowLite.init(); diff --git a/tensorflow/contrib/lite/java/src/main/native/BUILD b/tensorflow/contrib/lite/java/src/main/native/BUILD index 4399ed202597082fba36c04a744bf6378e4539a2..4b4e1c21d818dc56803ff31d83d19dea2ac08707 100644 --- a/tensorflow/contrib/lite/java/src/main/native/BUILD +++ b/tensorflow/contrib/lite/java/src/main/native/BUILD @@ -11,7 +11,6 @@ licenses(["notice"]) # Apache 2.0 cc_library( name = "native_framework_only", srcs = [ - "duration_utils_jni.cc", "exception_jni.cc", "nativeinterpreterwrapper_jni.cc", "tensor_jni.cc", diff --git a/tensorflow/contrib/lite/java/src/main/native/nativeinterpreterwrapper_jni.cc b/tensorflow/contrib/lite/java/src/main/native/nativeinterpreterwrapper_jni.cc index 31f7b58fbc30cab9e6cb813094ea4b2627ba5cba..e2c1edd9afbac9ca75262e1e41f7cf6334564dec 100644 --- a/tensorflow/contrib/lite/java/src/main/native/nativeinterpreterwrapper_jni.cc +++ b/tensorflow/contrib/lite/java/src/main/native/nativeinterpreterwrapper_jni.cc @@ -16,9 +16,6 @@ limitations under the License. #include "tensorflow/contrib/lite/java/src/main/native/nativeinterpreterwrapper_jni.h" namespace { -const int kByteBufferValue = 999; -const int kBufferSize = 256; - tflite::Interpreter* convertLongToInterpreter(JNIEnv* env, jlong handle) { if (handle == 0) { throwException(env, kIllegalArgumentException, @@ -62,22 +59,6 @@ std::vector convertJIntArrayToVector(JNIEnv* env, jintArray inputs) { return outputs; } -bool isByteBuffer(jint data_type) { return data_type == kByteBufferValue; } - -TfLiteType resolveDataType(jint data_type) { - switch (data_type) { - case 1: - return kTfLiteFloat32; - case 2: - return kTfLiteInt32; - case 3: - return kTfLiteUInt8; - case 4: - return kTfLiteInt64; - default: - return kTfLiteNoType; - } -} int getDataType(TfLiteType data_type) { switch (data_type) { @@ -108,64 +89,6 @@ void printDims(char* buffer, int max_size, int* dims, int num_dims) { } } -TfLiteStatus checkInputs(JNIEnv* env, tflite::Interpreter* interpreter, - const int input_size, jintArray data_types, - jintArray nums_of_bytes, jobjectArray values, - jobjectArray sizes) { - if (input_size != interpreter->inputs().size()) { - throwException(env, kIllegalArgumentException, - "Input error: Expected num of inputs is %d but got %d", - interpreter->inputs().size(), input_size); - return kTfLiteError; - } - if (input_size != env->GetArrayLength(data_types) || - input_size != env->GetArrayLength(nums_of_bytes) || - input_size != env->GetArrayLength(values)) { - throwException(env, kIllegalArgumentException, - "Internal error: Arrays in arguments should be of the same " - "length, but got %d sizes, %d data_types, %d nums_of_bytes, " - "and %d values", - input_size, env->GetArrayLength(data_types), - env->GetArrayLength(nums_of_bytes), - env->GetArrayLength(values)); - return kTfLiteError; - } - for (int i = 0; i < input_size; ++i) { - int input_idx = interpreter->inputs()[i]; - TfLiteTensor* target = interpreter->tensor(input_idx); - jintArray dims = - static_cast(env->GetObjectArrayElement(sizes, i)); - int num_dims = static_cast(env->GetArrayLength(dims)); - if (target->dims->size != num_dims) { - throwException(env, kIllegalArgumentException, - "Input error: %d-th input should have %d dimensions, but " - "found %d dimensions", - i, target->dims->size, num_dims); - return kTfLiteError; - } - jint* ptr = env->GetIntArrayElements(dims, nullptr); - for (int j = 1; j < num_dims; ++j) { - if (target->dims->data[j] != ptr[j]) { - std::unique_ptr expected_dims(new char[kBufferSize]); - std::unique_ptr obtained_dims(new char[kBufferSize]); - printDims(expected_dims.get(), kBufferSize, target->dims->data, - num_dims); - printDims(obtained_dims.get(), kBufferSize, ptr, num_dims); - throwException(env, kIllegalArgumentException, - "Input error: %d-th input dimension should be [%s], but " - "found [%s]", - i, expected_dims.get(), obtained_dims.get()); - env->ReleaseIntArrayElements(dims, ptr, JNI_ABORT); - return kTfLiteError; - } - } - env->ReleaseIntArrayElements(dims, ptr, JNI_ABORT); - env->DeleteLocalRef(dims); - if (env->ExceptionCheck()) return kTfLiteError; - } - return kTfLiteOk; -} - // Checks whether there is any difference between dimensions of a tensor and a // given dimensions. Returns true if there is difference, else false. bool areDimsDifferent(JNIEnv* env, TfLiteTensor* tensor, jintArray dims) { @@ -188,74 +111,6 @@ bool areDimsDifferent(JNIEnv* env, TfLiteTensor* tensor, jintArray dims) { return false; } -bool areInputDimensionsTheSame(JNIEnv* env, tflite::Interpreter* interpreter, - int input_size, jobjectArray sizes) { - if (interpreter->inputs().size() != input_size) { - return false; - } - for (int i = 0; i < input_size; ++i) { - int input_idx = interpreter->inputs()[i]; - jintArray dims = - static_cast(env->GetObjectArrayElement(sizes, i)); - TfLiteTensor* target = interpreter->tensor(input_idx); - if (areDimsDifferent(env, target, dims)) return false; - env->DeleteLocalRef(dims); - if (env->ExceptionCheck()) return false; - } - return true; -} - -TfLiteStatus resizeInputs(JNIEnv* env, tflite::Interpreter* interpreter, - int input_size, jobjectArray sizes) { - for (int i = 0; i < input_size; ++i) { - int input_idx = interpreter->inputs()[i]; - jintArray dims = - static_cast(env->GetObjectArrayElement(sizes, i)); - TfLiteStatus status = interpreter->ResizeInputTensor( - input_idx, convertJIntArrayToVector(env, dims)); - if (status != kTfLiteOk) { - return status; - } - env->DeleteLocalRef(dims); - if (env->ExceptionCheck()) return kTfLiteError; - } - return kTfLiteOk; -} - -TfLiteStatus setInputs(JNIEnv* env, tflite::Interpreter* interpreter, - int input_size, jintArray data_types, - jintArray nums_of_bytes, jobjectArray values) { - jint* data_type = env->GetIntArrayElements(data_types, nullptr); - jint* num_bytes = env->GetIntArrayElements(nums_of_bytes, nullptr); - for (int i = 0; i < input_size; ++i) { - int input_idx = interpreter->inputs()[i]; - TfLiteTensor* target = interpreter->tensor(input_idx); - jobject value = env->GetObjectArrayElement(values, i); - bool is_byte_buffer = isByteBuffer(data_type[i]); - if (is_byte_buffer) { - writeByteBuffer(env, value, &(target->data.raw), - static_cast(num_bytes[i])); - } else { - TfLiteType type = resolveDataType(data_type[i]); - if (type != target->type) { - throwException(env, kIllegalArgumentException, - "Input error: DataType (%d) of input data does not " - "match with the DataType (%d) of model inputs.", - type, target->type); - return kTfLiteError; - } - writeMultiDimensionalArray(env, value, target->type, target->dims->size, - &(target->data.raw), - static_cast(num_bytes[i])); - } - env->DeleteLocalRef(value); - if (env->ExceptionCheck()) return kTfLiteError; - } - env->ReleaseIntArrayElements(data_types, data_type, JNI_ABORT); - env->ReleaseIntArrayElements(nums_of_bytes, num_bytes, JNI_ABORT); - return kTfLiteOk; -} - // TODO(yichengfan): evaluate the benefit to use tflite verifier. bool VerifyModel(const void* buf, size_t len) { flatbuffers::Verifier verifier(static_cast(buf), len); @@ -287,6 +142,63 @@ Java_org_tensorflow_lite_NativeInterpreterWrapper_getInputNames(JNIEnv* env, return names; } +JNIEXPORT void JNICALL +Java_org_tensorflow_lite_NativeInterpreterWrapper_allocateTensors( + JNIEnv* env, jclass clazz, jlong handle, jlong error_handle) { + tflite::Interpreter* interpreter = convertLongToInterpreter(env, handle); + if (interpreter == nullptr) return; + BufferErrorReporter* error_reporter = + convertLongToErrorReporter(env, error_handle); + if (error_reporter == nullptr) return; + + if (interpreter->AllocateTensors() != kTfLiteOk) { + throwException(env, kNullPointerException, + "Internal error: Cannot allocate memory for the interpreter:" + " %s", + error_reporter->CachedErrorMessage()); + } +} + +JNIEXPORT jlong JNICALL +Java_org_tensorflow_lite_NativeInterpreterWrapper_getInputTensor(JNIEnv* env, + jclass clazz, + jlong handle, + jint index) { + tflite::Interpreter* interpreter = convertLongToInterpreter(env, handle); + if (interpreter == nullptr) return 0; + return reinterpret_cast( + interpreter->tensor(interpreter->inputs()[index])); +} + +JNIEXPORT jlong JNICALL +Java_org_tensorflow_lite_NativeInterpreterWrapper_getOutputTensor(JNIEnv* env, + jclass clazz, + jlong handle, + jint index) { + tflite::Interpreter* interpreter = convertLongToInterpreter(env, handle); + if (interpreter == nullptr) return 0; + return reinterpret_cast( + interpreter->tensor(interpreter->outputs()[index])); +} + +JNIEXPORT jint JNICALL +Java_org_tensorflow_lite_NativeInterpreterWrapper_getInputCount(JNIEnv* env, + jclass clazz, + jlong handle) { + tflite::Interpreter* interpreter = convertLongToInterpreter(env, handle); + if (interpreter == nullptr) return 0; + return static_cast(interpreter->inputs().size()); +} + +JNIEXPORT jint JNICALL +Java_org_tensorflow_lite_NativeInterpreterWrapper_getOutputCount(JNIEnv* env, + jclass clazz, + jlong handle) { + tflite::Interpreter* interpreter = convertLongToInterpreter(env, handle); + if (interpreter == nullptr) return 0; + return static_cast(interpreter->outputs().size()); +} + JNIEXPORT jobjectArray JNICALL Java_org_tensorflow_lite_NativeInterpreterWrapper_getOutputNames(JNIEnv* env, jclass clazz, @@ -434,114 +346,21 @@ Java_org_tensorflow_lite_NativeInterpreterWrapper_createInterpreter( } // Sets inputs, runs inference, and returns outputs as long handles. -JNIEXPORT jlongArray JNICALL -Java_org_tensorflow_lite_NativeInterpreterWrapper_run( - JNIEnv* env, jclass clazz, jlong interpreter_handle, jlong error_handle, - jobjectArray sizes, jintArray data_types, jintArray nums_of_bytes, - jobjectArray values, jobject wrapper, jboolean memory_allocated) { +JNIEXPORT void JNICALL Java_org_tensorflow_lite_NativeInterpreterWrapper_run( + JNIEnv* env, jclass clazz, jlong interpreter_handle, jlong error_handle) { tflite::Interpreter* interpreter = convertLongToInterpreter(env, interpreter_handle); - if (interpreter == nullptr) return nullptr; + if (interpreter == nullptr) return; BufferErrorReporter* error_reporter = convertLongToErrorReporter(env, error_handle); - if (error_reporter == nullptr) return nullptr; - const int input_size = env->GetArrayLength(sizes); - // validates inputs - TfLiteStatus status = checkInputs(env, interpreter, input_size, data_types, - nums_of_bytes, values, sizes); - if (status != kTfLiteOk) return nullptr; - if (!memory_allocated || - !areInputDimensionsTheSame(env, interpreter, input_size, sizes)) { - // resizes inputs - status = resizeInputs(env, interpreter, input_size, sizes); - if (status != kTfLiteOk) { - throwException(env, kNullPointerException, - "Internal error: Can not resize the input: %s", - error_reporter->CachedErrorMessage()); - return nullptr; - } - // allocates memory - status = interpreter->AllocateTensors(); - if (status != kTfLiteOk) { - throwException(env, kNullPointerException, - "Internal error: Can not allocate memory for the given " - "inputs: %s", - error_reporter->CachedErrorMessage()); - return nullptr; - } - } - // sets inputs - status = setInputs(env, interpreter, input_size, data_types, nums_of_bytes, - values); - if (status != kTfLiteOk) return nullptr; - timespec beforeInference = ::tflite::getCurrentTime(); - // runs inference + if (error_reporter == nullptr) return; + if (interpreter->Invoke() != kTfLiteOk) { throwException(env, kIllegalArgumentException, "Internal error: Failed to run on the given Interpreter: %s", error_reporter->CachedErrorMessage()); - return nullptr; - } - timespec afterInference = ::tflite::getCurrentTime(); - jclass wrapper_clazz = env->GetObjectClass(wrapper); - jfieldID fid = - env->GetFieldID(wrapper_clazz, "inferenceDurationNanoseconds", "J"); - if (env->ExceptionCheck()) { - env->ExceptionClear(); - } else if (fid != nullptr) { - env->SetLongField( - wrapper, fid, - ::tflite::timespec_diff_nanoseconds(&beforeInference, &afterInference)); - } - // returns outputs - const std::vector& results = interpreter->outputs(); - if (results.empty()) { - throwException( - env, kIllegalArgumentException, - "Internal error: The Interpreter does not have any outputs."); - return nullptr; - } - jlongArray outputs = env->NewLongArray(results.size()); - size_t size = results.size(); - for (int i = 0; i < size; ++i) { - TfLiteTensor* source = interpreter->tensor(results[i]); - jlong output = reinterpret_cast(source); - env->SetLongArrayRegion(outputs, i, 1, &output); - } - return outputs; -} - -JNIEXPORT jintArray JNICALL -Java_org_tensorflow_lite_NativeInterpreterWrapper_getInputDims( - JNIEnv* env, jclass clazz, jlong handle, jint input_idx, jint num_bytes) { - tflite::Interpreter* interpreter = convertLongToInterpreter(env, handle); - if (interpreter == nullptr) return nullptr; - const int idx = static_cast(input_idx); - if (input_idx < 0 || input_idx >= interpreter->inputs().size()) { - throwException(env, kIllegalArgumentException, - "Input error: Out of range: Failed to get %d-th input out of" - " %d inputs", - input_idx, interpreter->inputs().size()); - return nullptr; - } - TfLiteTensor* target = interpreter->tensor(interpreter->inputs()[idx]); - int size = target->dims->size; - if (num_bytes >= 0) { // verifies num of bytes matches if num_bytes if valid. - int expected_num_bytes = elementByteSize(target->type); - for (int i = 0; i < size; ++i) { - expected_num_bytes *= target->dims->data[i]; - } - if (num_bytes != expected_num_bytes) { - throwException(env, kIllegalArgumentException, - "Input error: Failed to get input dimensions. %d-th input " - "should have %d bytes, but found %d bytes.", - idx, expected_num_bytes, num_bytes); - return nullptr; - } + return; } - jintArray outputs = env->NewIntArray(size); - env->SetIntArrayRegion(outputs, 0, size, &(target->dims->data[0])); - return outputs; } JNIEXPORT jint JNICALL diff --git a/tensorflow/contrib/lite/java/src/main/native/nativeinterpreterwrapper_jni.h b/tensorflow/contrib/lite/java/src/main/native/nativeinterpreterwrapper_jni.h index 128ece49811a112684dac7b36810e920eeeb7351..618fba480e4a1c4a1ff8531cb3fbc29fcb8191d8 100644 --- a/tensorflow/contrib/lite/java/src/main/native/nativeinterpreterwrapper_jni.h +++ b/tensorflow/contrib/lite/java/src/main/native/nativeinterpreterwrapper_jni.h @@ -29,15 +29,63 @@ limitations under the License. namespace tflite { // This is to be provided at link-time by a library. extern std::unique_ptr CreateOpResolver(); -extern timespec getCurrentTime(); -extern jlong timespec_diff_nanoseconds(struct timespec* start, - struct timespec* stop); } // namespace tflite #ifdef __cplusplus extern "C" { #endif // __cplusplus +/* + * Class: org_tensorflow_lite_NativeInterpreterWrapper + * Method: allocateTensors + * Signature: (JJ)V + */ +JNIEXPORT void JNICALL +Java_org_tensorflow_lite_NativeInterpreterWrapper_allocateTensors( + JNIEnv* env, jclass clazz, jlong handle, jlong error_handle); + +/* + * Class: org_tensorflow_lite_NativeInterpreterWrapper + * Method: getInputTensor + * Signature: (JI)J + */ +JNIEXPORT jlong JNICALL +Java_org_tensorflow_lite_NativeInterpreterWrapper_getInputTensor(JNIEnv* env, + jclass clazz, + jlong handle, + jint index); + +/* + * Class: org_tensorflow_lite_NativeInterpreterWrapper + * Method: getOutputTensor + * Signature: (JI)J + */ +JNIEXPORT jlong JNICALL +Java_org_tensorflow_lite_NativeInterpreterWrapper_getOutputTensor(JNIEnv* env, + jclass clazz, + jlong handle, + jint index); + +/* + * Class: org_tensorflow_lite_NativeInterpreterWrapper + * Method: getInputCount + * Signature: (J)I + */ +JNIEXPORT jint JNICALL +Java_org_tensorflow_lite_NativeInterpreterWrapper_getInputCount(JNIEnv* env, + jclass clazz, + jlong handle); + +/* + * Class: org_tensorflow_lite_NativeInterpreterWrapper + * Method: getOutputCount + * Signature: (J)I + */ +JNIEXPORT jint JNICALL +Java_org_tensorflow_lite_NativeInterpreterWrapper_getOutputCount(JNIEnv* env, + jclass clazz, + jlong handle); + /* * Class: org_tensorflow_lite_NativeInterpreterWrapper * Method: @@ -118,28 +166,11 @@ Java_org_tensorflow_lite_NativeInterpreterWrapper_createInterpreter( /* * Class: org_tensorflow_lite_NativeInterpreterWrapper - * Method: - * Signature: - * (JJ[Ljava/lang/Object;[I[I[Ljava/lang/Object;Ljava/lang/Object;Z)[J - */ -JNIEXPORT jlongArray JNICALL -Java_org_tensorflow_lite_NativeInterpreterWrapper_run( - JNIEnv* env, jclass clazz, jlong interpreter_handle, jlong error_handle, - jobjectArray sizes, jintArray data_types, jintArray nums_of_bytes, - jobjectArray values, jobject wrapper, jboolean memory_allocated); - -/* - * Class: org_tensorflow_lite_NativeInterpreterWrapper - * Method: - * Signature: (JII)[I - * - * Gets input dimensions. If num_bytes is non-negative, it will check whether - * num_bytes matches num of bytes required by the input, and return null and - * throw IllegalArgumentException if not. + * Method: run + * Signature: (JJ)V */ -JNIEXPORT jintArray JNICALL -Java_org_tensorflow_lite_NativeInterpreterWrapper_getInputDims( - JNIEnv* env, jclass clazz, jlong handle, jint input_idx, jint num_bytes); +JNIEXPORT void JNICALL Java_org_tensorflow_lite_NativeInterpreterWrapper_run( + JNIEnv* env, jclass clazz, jlong interpreter_handle, jlong error_handle); /* * Class: org_tensorflow_lite_NativeInterpreterWrapper diff --git a/tensorflow/contrib/lite/java/src/main/native/tensor_jni.cc b/tensorflow/contrib/lite/java/src/main/native/tensor_jni.cc index 9e9387da86ebde7d711a7ce967461e370c95bc3e..7ff96a3172dcf020b34fcbe7491c9022fc7f51de 100644 --- a/tensorflow/contrib/lite/java/src/main/native/tensor_jni.cc +++ b/tensorflow/contrib/lite/java/src/main/native/tensor_jni.cc @@ -29,6 +29,35 @@ TfLiteTensor* convertLongToTensor(JNIEnv* env, jlong handle) { return reinterpret_cast(handle); } +size_t elementByteSize(TfLiteType data_type) { + // The code in this file makes the assumption that the + // TensorFlow TF_DataTypes and the Java primitive types + // have the same byte sizes. Validate that: + switch (data_type) { + case kTfLiteFloat32: + static_assert(sizeof(jfloat) == 4, + "Interal error: Java float not compatible with " + "kTfLiteFloat"); + return 4; + case kTfLiteInt32: + static_assert(sizeof(jint) == 4, + "Interal error: Java int not compatible with kTfLiteInt"); + return 4; + case kTfLiteUInt8: + static_assert(sizeof(jbyte) == 1, + "Interal error: Java byte not compatible with " + "kTfLiteUInt8"); + return 1; + case kTfLiteInt64: + static_assert(sizeof(jlong) == 8, + "Interal error: Java long not compatible with " + "kTfLiteInt64"); + return 8; + default: + return 0; + } +} + size_t writeOneDimensionalArray(JNIEnv* env, jobject object, TfLiteType type, void* dst, size_t dst_size) { jarray array = static_cast(object); @@ -141,48 +170,6 @@ size_t readMultiDimensionalArray(JNIEnv* env, TfLiteType data_type, char* src, } } -} // namespace - -size_t elementByteSize(TfLiteType data_type) { - // The code in this file makes the assumption that the - // TensorFlow TF_DataTypes and the Java primitive types - // have the same byte sizes. Validate that: - switch (data_type) { - case kTfLiteFloat32: - static_assert(sizeof(jfloat) == 4, - "Interal error: Java float not compatible with " - "kTfLiteFloat"); - return 4; - case kTfLiteInt32: - static_assert(sizeof(jint) == 4, - "Interal error: Java int not compatible with kTfLiteInt"); - return 4; - case kTfLiteUInt8: - static_assert(sizeof(jbyte) == 1, - "Interal error: Java byte not compatible with " - "kTfLiteUInt8"); - return 1; - case kTfLiteInt64: - static_assert(sizeof(jlong) == 8, - "Interal error: Java long not compatible with " - "kTfLiteInt64"); - return 8; - default: - return 0; - } -} - -size_t writeByteBuffer(JNIEnv* env, jobject object, char** dst, int dst_size) { - char* buf = static_cast(env->GetDirectBufferAddress(object)); - if (!buf) { - throwException(env, kIllegalArgumentException, - "Input ByteBuffer is not a direct buffer"); - return 0; - } - *dst = buf; - return dst_size; -} - size_t writeMultiDimensionalArray(JNIEnv* env, jobject src, TfLiteType type, int dims_left, char** dst, int dst_size) { if (dims_left <= 1) { @@ -203,6 +190,37 @@ size_t writeMultiDimensionalArray(JNIEnv* env, jobject src, TfLiteType type, } } +} // namespace + +JNIEXPORT jobject JNICALL Java_org_tensorflow_lite_Tensor_buffer(JNIEnv* env, + jclass clazz, + jlong handle) { + TfLiteTensor* tensor = convertLongToTensor(env, handle); + if (tensor == nullptr) return nullptr; + if (tensor->data.raw == nullptr) { + throwException(env, kIllegalArgumentException, + "Internal error: Tensor hasn't been allocated."); + return nullptr; + } + return env->NewDirectByteBuffer(static_cast(tensor->data.raw), + static_cast(tensor->bytes)); +} + +JNIEXPORT void JNICALL Java_org_tensorflow_lite_Tensor_writeDirectBuffer( + JNIEnv* env, jclass clazz, jlong handle, jobject src) { + TfLiteTensor* tensor = convertLongToTensor(env, handle); + if (tensor == nullptr) return; + + char* src_data_raw = static_cast(env->GetDirectBufferAddress(src)); + if (!src_data_raw) { + throwException(env, kIllegalArgumentException, + "Input ByteBuffer is not a direct buffer"); + return; + } + + tensor->data.raw = src_data_raw; +} + JNIEXPORT void JNICALL Java_org_tensorflow_lite_Tensor_readMultiDimensionalArray(JNIEnv* env, jclass clazz, @@ -220,6 +238,27 @@ Java_org_tensorflow_lite_Tensor_readMultiDimensionalArray(JNIEnv* env, num_dims, static_cast(value)); } +JNIEXPORT void JNICALL +Java_org_tensorflow_lite_Tensor_writeMultiDimensionalArray(JNIEnv* env, + jclass clazz, + jlong handle, + jobject src) { + TfLiteTensor* tensor = convertLongToTensor(env, handle); + if (tensor == nullptr) return; + if (tensor->data.raw == nullptr) { + throwException(env, kIllegalArgumentException, + "Internal error: Target Tensor hasn't been allocated."); + return; + } + if (tensor->dims->size == 0) { + throwException(env, kIllegalArgumentException, + "Internal error: Cannot copy empty/scalar Tensors."); + return; + } + writeMultiDimensionalArray(env, src, tensor->type, tensor->dims->size, + &tensor->data.raw, tensor->bytes); +} + JNIEXPORT jint JNICALL Java_org_tensorflow_lite_Tensor_dtype(JNIEnv* env, jclass clazz, jlong handle) { @@ -237,3 +276,11 @@ Java_org_tensorflow_lite_Tensor_shape(JNIEnv* env, jclass clazz, jlong handle) { env->SetIntArrayRegion(result, 0, num_dims, tensor->dims->data); return result; } + +JNIEXPORT jint JNICALL Java_org_tensorflow_lite_Tensor_numBytes(JNIEnv* env, + jclass clazz, + jlong handle) { + const TfLiteTensor* tensor = convertLongToTensor(env, handle); + if (tensor == nullptr) return 0; + return static_cast(tensor->bytes); +} diff --git a/tensorflow/contrib/lite/java/src/main/native/tensor_jni.h b/tensorflow/contrib/lite/java/src/main/native/tensor_jni.h index 3a4910dcc3a719fbb9f365dae693423de768349c..06e2546af8400de117ed6923a1d1bd67bcb998e2 100644 --- a/tensorflow/contrib/lite/java/src/main/native/tensor_jni.h +++ b/tensorflow/contrib/lite/java/src/main/native/tensor_jni.h @@ -24,8 +24,25 @@ extern "C" { #endif // __cplusplus /* - * Class: org_tensorflow_lite_TfLiteTensor - * Method: + * Class: org_tensorflow_lite_Tensor + * Method: buffer + * Signature: (J)Ljava/nio/ByteBuffer; + */ +JNIEXPORT jobject JNICALL Java_org_tensorflow_lite_Tensor_buffer(JNIEnv* env, + jclass clazz, + jlong handle); + +/* + * Class: org_tensorflow_lite_Tensor + * Method: writeDirectBuffer + * Signature: (JLjava/nio/ByteBuffer;) + */ +JNIEXPORT void JNICALL Java_org_tensorflow_lite_Tensor_writeDirectBuffer( + JNIEnv* env, jclass clazz, jlong handle, jobject src); + +/* + * Class: org_tensorflow_lite_Tensor + * Method: dtype * Signature: (J)I */ JNIEXPORT jint JNICALL Java_org_tensorflow_lite_Tensor_dtype(JNIEnv* env, @@ -33,8 +50,8 @@ JNIEXPORT jint JNICALL Java_org_tensorflow_lite_Tensor_dtype(JNIEnv* env, jlong handle); /* - * Class: org_tensorflow_lite_TfLiteTensor - * Method: + * Class: org_tensorflow_lite_Tensor + * Method: shape * Signature: (J)[I */ JNIEXPORT jintArray JNICALL Java_org_tensorflow_lite_Tensor_shape(JNIEnv* env, @@ -42,31 +59,35 @@ JNIEXPORT jintArray JNICALL Java_org_tensorflow_lite_Tensor_shape(JNIEnv* env, jlong handle); /* - * Class: org_tensorflow_lite_TfLiteTensor - * Method: + * Class: org_tensorflow_lite_Tensor + * Method: numBytes + * Signature: (J)I + */ +JNIEXPORT jint JNICALL Java_org_tensorflow_lite_Tensor_numBytes(JNIEnv* env, + jclass clazz, + jlong handle); + +/* + * Class: org_tensorflow_lite_Tensor + * Method: readMultiDimensionalArray * Signature: (JLjava/lang/Object;) */ JNIEXPORT void JNICALL Java_org_tensorflow_lite_Tensor_readMultiDimensionalArray(JNIEnv* env, jclass clazz, jlong handle, - jobject value); + jobject dst); /* - * Finds the size of each data type. - */ -size_t elementByteSize(TfLiteType data_type); - -/* - * Writes data of a ByteBuffer into dest. - */ -size_t writeByteBuffer(JNIEnv* env, jobject object, char** dst, int dst_size); - -/* - * Writes a multi-dimensional array into dest. + * Class: org_tensorflow_lite_Tensor + * Method: writeMultidimensionalArray + * Signature: (JLjava/lang/Object;) */ -size_t writeMultiDimensionalArray(JNIEnv* env, jobject src, TfLiteType type, - int dims_left, char** dst, int dst_size); +JNIEXPORT void JNICALL +Java_org_tensorflow_lite_Tensor_writeMultiDimensionalArray(JNIEnv* env, + jclass clazz, + jlong handle, + jobject src); #ifdef __cplusplus } // extern "C" diff --git a/tensorflow/contrib/lite/java/src/test/java/org/tensorflow/lite/InterpreterTest.java b/tensorflow/contrib/lite/java/src/test/java/org/tensorflow/lite/InterpreterTest.java index 82007a6ab5be3492495125b1c20ed155907ae5a0..d66a73db94f06776fe2a7310ed0837941aba87c4 100644 --- a/tensorflow/contrib/lite/java/src/test/java/org/tensorflow/lite/InterpreterTest.java +++ b/tensorflow/contrib/lite/java/src/test/java/org/tensorflow/lite/InterpreterTest.java @@ -164,6 +164,24 @@ public final class InterpreterTest { interpreter.close(); } + @Test + public void testRunWithByteBufferOutput() { + float[] oneD = {1.23f, 6.54f, 7.81f}; + float[][] twoD = {oneD, oneD, oneD, oneD, oneD, oneD, oneD, oneD}; + float[][][] threeD = {twoD, twoD, twoD, twoD, twoD, twoD, twoD, twoD}; + float[][][][] fourD = {threeD, threeD}; + ByteBuffer parsedOutput = + ByteBuffer.allocateDirect(2 * 8 * 8 * 3 * 4).order(ByteOrder.nativeOrder()); + try (Interpreter interpreter = new Interpreter(MODEL_FILE)) { + interpreter.run(fourD, parsedOutput); + } + float[] outputOneD = { + parsedOutput.getFloat(0), parsedOutput.getFloat(4), parsedOutput.getFloat(8) + }; + float[] expected = {3.69f, 19.62f, 23.43f}; + assertThat(outputOneD).usingTolerance(0.1f).containsExactly(expected).inOrder(); + } + @Test public void testMobilenetRun() { // Create a gray image. @@ -203,7 +221,9 @@ public final class InterpreterTest { assertThat(e) .hasMessageThat() .contains( - "DataType (2) of input data does not match with the DataType (1) of model inputs."); + "Cannot convert between a TensorFlowLite tensor with type " + + "FLOAT32 and a Java object of type [[[[I (which is compatible with the" + + " TensorFlowLite type INT32)"); } interpreter.close(); } @@ -223,8 +243,8 @@ public final class InterpreterTest { assertThat(e) .hasMessageThat() .contains( - "Cannot convert an TensorFlowLite tensor with type " - + "FLOAT32 to a Java object of type [[[[I (which is compatible with the" + "Cannot convert between a TensorFlowLite tensor with type " + + "FLOAT32 and a Java object of type [[[[I (which is compatible with the" + " TensorFlowLite type INT32)"); } interpreter.close(); @@ -311,4 +331,11 @@ public final class InterpreterTest { interpreter.close(); fileChannel.close(); } + + @Test + public void testRedundantClose() throws Exception { + Interpreter interpreter = new Interpreter(MODEL_FILE); + interpreter.close(); + interpreter.close(); + } } diff --git a/tensorflow/contrib/lite/java/src/test/java/org/tensorflow/lite/NativeInterpreterWrapperTest.java b/tensorflow/contrib/lite/java/src/test/java/org/tensorflow/lite/NativeInterpreterWrapperTest.java index 9e41cb132d8386748e24c46d846e04f158d8b4c6..9c4a5acd797ec3476f44fb203901c9ba0429ab26 100644 --- a/tensorflow/contrib/lite/java/src/test/java/org/tensorflow/lite/NativeInterpreterWrapperTest.java +++ b/tensorflow/contrib/lite/java/src/test/java/org/tensorflow/lite/NativeInterpreterWrapperTest.java @@ -20,6 +20,8 @@ import static org.junit.Assert.fail; import java.nio.ByteBuffer; import java.nio.ByteOrder; +import java.util.HashMap; +import java.util.Map; import org.junit.Test; import org.junit.runner.RunWith; import org.junit.runners.JUnit4; @@ -101,16 +103,37 @@ public final class NativeInterpreterWrapperTest { float[][][] threeD = {twoD, twoD, twoD, twoD, twoD, twoD, twoD, twoD}; float[][][][] fourD = {threeD, threeD}; Object[] inputs = {fourD}; - Tensor[] outputs = wrapper.run(inputs); - assertThat(outputs.length).isEqualTo(1); float[][][][] parsedOutputs = new float[2][8][8][3]; - outputs[0].copyTo(parsedOutputs); + Map outputs = new HashMap<>(); + outputs.put(0, parsedOutputs); + wrapper.run(inputs, outputs); float[] outputOneD = parsedOutputs[0][0][0]; float[] expected = {3.69f, -19.62f, 23.43f}; assertThat(outputOneD).usingTolerance(0.1f).containsExactly(expected).inOrder(); wrapper.close(); } + @Test + public void testRunWithBufferOutput() { + try (NativeInterpreterWrapper wrapper = new NativeInterpreterWrapper(FLOAT_MODEL_PATH)) { + float[] oneD = {1.23f, -6.54f, 7.81f}; + float[][] twoD = {oneD, oneD, oneD, oneD, oneD, oneD, oneD, oneD}; + float[][][] threeD = {twoD, twoD, twoD, twoD, twoD, twoD, twoD, twoD}; + float[][][][] fourD = {threeD, threeD}; + Object[] inputs = {fourD}; + ByteBuffer parsedOutput = + ByteBuffer.allocateDirect(2 * 8 * 8 * 3 * 4).order(ByteOrder.nativeOrder()); + Map outputs = new HashMap<>(); + outputs.put(0, parsedOutput); + wrapper.run(inputs, outputs); + float[] outputOneD = { + parsedOutput.getFloat(0), parsedOutput.getFloat(4), parsedOutput.getFloat(8) + }; + float[] expected = {3.69f, -19.62f, 23.43f}; + assertThat(outputOneD).usingTolerance(0.1f).containsExactly(expected).inOrder(); + } + } + @Test public void testRunWithInputsOfSameDims() { NativeInterpreterWrapper wrapper = new NativeInterpreterWrapper(FLOAT_MODEL_PATH); @@ -119,17 +142,16 @@ public final class NativeInterpreterWrapperTest { float[][][] threeD = {twoD, twoD, twoD, twoD, twoD, twoD, twoD, twoD}; float[][][][] fourD = {threeD, threeD}; Object[] inputs = {fourD}; - Tensor[] outputs = wrapper.run(inputs); - assertThat(outputs.length).isEqualTo(1); float[][][][] parsedOutputs = new float[2][8][8][3]; - outputs[0].copyTo(parsedOutputs); + Map outputs = new HashMap<>(); + outputs.put(0, parsedOutputs); + wrapper.run(inputs, outputs); float[] outputOneD = parsedOutputs[0][0][0]; float[] expected = {3.69f, -19.62f, 23.43f}; assertThat(outputOneD).usingTolerance(0.1f).containsExactly(expected).inOrder(); - outputs = wrapper.run(inputs); - assertThat(outputs.length).isEqualTo(1); parsedOutputs = new float[2][8][8][3]; - outputs[0].copyTo(parsedOutputs); + outputs.put(0, parsedOutputs); + wrapper.run(inputs, outputs); outputOneD = parsedOutputs[0][0][0]; assertThat(outputOneD).usingTolerance(0.1f).containsExactly(expected).inOrder(); wrapper.close(); @@ -143,10 +165,10 @@ public final class NativeInterpreterWrapperTest { int[][][] threeD = {twoD, twoD, twoD, twoD, twoD, twoD, twoD, twoD}; int[][][][] fourD = {threeD, threeD}; Object[] inputs = {fourD}; - Tensor[] outputs = wrapper.run(inputs); - assertThat(outputs.length).isEqualTo(1); int[][][][] parsedOutputs = new int[2][4][4][12]; - outputs[0].copyTo(parsedOutputs); + Map outputs = new HashMap<>(); + outputs.put(0, parsedOutputs); + wrapper.run(inputs, outputs); int[] outputOneD = parsedOutputs[0][0][0]; int[] expected = {3, 7, -4, 3, 7, -4, 3, 7, -4, 3, 7, -4}; assertThat(outputOneD).isEqualTo(expected); @@ -161,10 +183,10 @@ public final class NativeInterpreterWrapperTest { long[][][] threeD = {twoD, twoD, twoD, twoD, twoD, twoD, twoD, twoD}; long[][][][] fourD = {threeD, threeD}; Object[] inputs = {fourD}; - Tensor[] outputs = wrapper.run(inputs); - assertThat(outputs.length).isEqualTo(1); long[][][][] parsedOutputs = new long[2][4][4][12]; - outputs[0].copyTo(parsedOutputs); + Map outputs = new HashMap<>(); + outputs.put(0, parsedOutputs); + wrapper.run(inputs, outputs); long[] outputOneD = parsedOutputs[0][0][0]; long[] expected = {-892834092L, 923423L, 2123918239018L, -892834092L, 923423L, 2123918239018L, -892834092L, 923423L, 2123918239018L, -892834092L, 923423L, 2123918239018L}; @@ -182,10 +204,10 @@ public final class NativeInterpreterWrapperTest { Object[] inputs = {fourD}; int[] inputDims = {2, 8, 8, 3}; wrapper.resizeInput(0, inputDims); - Tensor[] outputs = wrapper.run(inputs); - assertThat(outputs.length).isEqualTo(1); byte[][][][] parsedOutputs = new byte[2][4][4][12]; - outputs[0].copyTo(parsedOutputs); + Map outputs = new HashMap<>(); + outputs.put(0, parsedOutputs); + wrapper.run(inputs, outputs); byte[] outputOneD = parsedOutputs[0][0][0]; byte[] expected = {(byte) 0xe0, 0x4f, (byte) 0xd0, (byte) 0xe0, 0x4f, (byte) 0xd0, (byte) 0xe0, 0x4f, (byte) 0xd0, (byte) 0xe0, 0x4f, (byte) 0xd0}; @@ -208,13 +230,14 @@ public final class NativeInterpreterWrapperTest { } } } + bbuf.rewind(); Object[] inputs = {bbuf}; int[] inputDims = {2, 8, 8, 3}; wrapper.resizeInput(0, inputDims); - Tensor[] outputs = wrapper.run(inputs); - assertThat(outputs.length).isEqualTo(1); byte[][][][] parsedOutputs = new byte[2][4][4][12]; - outputs[0].copyTo(parsedOutputs); + Map outputs = new HashMap<>(); + outputs.put(0, parsedOutputs); + wrapper.run(inputs, outputs); byte[] outputOneD = parsedOutputs[0][0][0]; byte[] expected = { (byte) 0xe0, 0x4f, (byte) 0xd0, (byte) 0xe0, 0x4f, (byte) 0xd0, @@ -240,21 +263,22 @@ public final class NativeInterpreterWrapperTest { } } Object[] inputs = {bbuf}; + float[][][][] parsedOutputs = new float[4][8][8][3]; + Map outputs = new HashMap<>(); + outputs.put(0, parsedOutputs); try { - wrapper.run(inputs); + wrapper.run(inputs, outputs); fail(); } catch (IllegalArgumentException e) { assertThat(e) .hasMessageThat() .contains( - "Failed to get input dimensions. 0-th input should have 768 bytes, but found 3072 bytes"); + "Cannot convert between a TensorFlowLite buffer with 768 bytes and a " + + "ByteBuffer with 3072 bytes."); } int[] inputDims = {4, 8, 8, 3}; wrapper.resizeInput(0, inputDims); - Tensor[] outputs = wrapper.run(inputs); - assertThat(outputs.length).isEqualTo(1); - float[][][][] parsedOutputs = new float[4][8][8][3]; - outputs[0].copyTo(parsedOutputs); + wrapper.run(inputs, outputs); float[] outputOneD = parsedOutputs[0][0][0]; float[] expected = {3.69f, -19.62f, 23.43f}; assertThat(outputOneD).usingTolerance(0.1f).containsExactly(expected).inOrder(); @@ -267,14 +291,18 @@ public final class NativeInterpreterWrapperTest { ByteBuffer bbuf = ByteBuffer.allocateDirect(2 * 7 * 8 * 3); bbuf.order(ByteOrder.nativeOrder()); Object[] inputs = {bbuf}; + Map outputs = new HashMap<>(); + ByteBuffer parsedOutput = ByteBuffer.allocateDirect(2 * 7 * 8 * 3); + outputs.put(0, parsedOutput); try { - wrapper.run(inputs); + wrapper.run(inputs, outputs); fail(); } catch (IllegalArgumentException e) { assertThat(e) .hasMessageThat() .contains( - "Failed to get input dimensions. 0-th input should have 192 bytes, but found 336 bytes."); + "Cannot convert between a TensorFlowLite buffer with 192 bytes and a " + + "ByteBuffer with 336 bytes."); } wrapper.close(); } @@ -287,14 +315,18 @@ public final class NativeInterpreterWrapperTest { int[][][] threeD = {twoD, twoD, twoD, twoD, twoD, twoD, twoD, twoD}; int[][][][] fourD = {threeD, threeD}; Object[] inputs = {fourD}; + int[][][][] parsedOutputs = new int[2][8][8][3]; + Map outputs = new HashMap<>(); + outputs.put(0, parsedOutputs); try { - wrapper.run(inputs); + wrapper.run(inputs, outputs); fail(); } catch (IllegalArgumentException e) { assertThat(e) .hasMessageThat() .contains( - "DataType (2) of input data does not match with the DataType (1) of model inputs."); + "Cannot convert between a TensorFlowLite tensor with type FLOAT32 and a Java object " + + "of type [[[[I (which is compatible with the TensorFlowLite type INT32)"); } wrapper.close(); } @@ -308,8 +340,11 @@ public final class NativeInterpreterWrapperTest { float[][][] threeD = {twoD, twoD, twoD, twoD, twoD, twoD, twoD, twoD}; float[][][][] fourD = {threeD, threeD}; Object[] inputs = {fourD}; + float[][][][] parsedOutputs = new float[2][8][8][3]; + Map outputs = new HashMap<>(); + outputs.put(0, parsedOutputs); try { - wrapper.run(inputs); + wrapper.run(inputs, outputs); fail(); } catch (IllegalArgumentException e) { assertThat(e).hasMessageThat().contains("Invalid handle to Interpreter."); @@ -321,7 +356,7 @@ public final class NativeInterpreterWrapperTest { NativeInterpreterWrapper wrapper = new NativeInterpreterWrapper(FLOAT_MODEL_PATH); try { Object[] inputs = {}; - wrapper.run(inputs); + wrapper.run(inputs, null); fail(); } catch (IllegalArgumentException e) { assertThat(e).hasMessageThat().contains("Inputs should not be null or empty."); @@ -337,11 +372,14 @@ public final class NativeInterpreterWrapperTest { float[][][] threeD = {twoD, twoD, twoD, twoD, twoD, twoD, twoD, twoD}; float[][][][] fourD = {threeD, threeD}; Object[] inputs = {fourD, fourD}; + float[][][][] parsedOutputs = new float[2][8][8][3]; + Map outputs = new HashMap<>(); + outputs.put(0, parsedOutputs); try { - wrapper.run(inputs); + wrapper.run(inputs, outputs); fail(); } catch (IllegalArgumentException e) { - assertThat(e).hasMessageThat().contains("Expected num of inputs is 1 but got 2"); + assertThat(e).hasMessageThat().contains("Invalid input Tensor index: 1"); } wrapper.close(); } @@ -353,13 +391,18 @@ public final class NativeInterpreterWrapperTest { float[][] twoD = {oneD, oneD, oneD, oneD, oneD, oneD, oneD}; float[][][] threeD = {twoD, twoD, twoD, twoD, twoD, twoD, twoD, twoD}; Object[] inputs = {threeD}; + float[][][][] parsedOutputs = new float[2][8][8][3]; + Map outputs = new HashMap<>(); + outputs.put(0, parsedOutputs); try { - wrapper.run(inputs); + wrapper.run(inputs, outputs); fail(); } catch (IllegalArgumentException e) { assertThat(e) .hasMessageThat() - .contains("0-th input should have 4 dimensions, but found 3 dimensions"); + .contains( + "Cannot copy between a TensorFlowLite tensor with shape [8, 7, 3] and a " + + "Java object with shape [2, 8, 8, 3]."); } wrapper.close(); } @@ -372,91 +415,22 @@ public final class NativeInterpreterWrapperTest { float[][][] threeD = {twoD, twoD, twoD, twoD, twoD, twoD, twoD, twoD}; float[][][][] fourD = {threeD, threeD}; Object[] inputs = {fourD}; + float[][][][] parsedOutputs = new float[2][8][8][3]; + Map outputs = new HashMap<>(); + outputs.put(0, parsedOutputs); try { - wrapper.run(inputs); + wrapper.run(inputs, outputs); fail(); } catch (IllegalArgumentException e) { assertThat(e) .hasMessageThat() - .contains("0-th input dimension should be [?,8,8,3], but found [?,8,7,3]"); + .contains( + "Cannot copy between a TensorFlowLite tensor with shape [2, 8, 7, 3] and a " + + "Java object with shape [2, 8, 8, 3]."); } wrapper.close(); } - @Test - public void testNumElements() { - int[] shape = {2, 3, 4}; - int num = NativeInterpreterWrapper.numElements(shape); - assertThat(num).isEqualTo(24); - shape = null; - num = NativeInterpreterWrapper.numElements(shape); - assertThat(num).isEqualTo(0); - } - - @Test - public void testIsNonEmtpyArray() { - assertThat(NativeInterpreterWrapper.isNonEmptyArray(null)).isFalse(); - assertThat(NativeInterpreterWrapper.isNonEmptyArray(3.2)).isFalse(); - int[] emptyArray = {}; - assertThat(NativeInterpreterWrapper.isNonEmptyArray(emptyArray)).isFalse(); - int[] validArray = {9, 5, 2, 1}; - assertThat(NativeInterpreterWrapper.isNonEmptyArray(validArray)).isTrue(); - } - - @Test - public void testDataTypeOf() { - float[] testEmtpyArray = {}; - DataType dataType = NativeInterpreterWrapper.dataTypeOf(testEmtpyArray); - assertThat(dataType).isEqualTo(DataType.FLOAT32); - float[] testFloatArray = {0.783f, 0.251f}; - dataType = NativeInterpreterWrapper.dataTypeOf(testFloatArray); - assertThat(dataType).isEqualTo(DataType.FLOAT32); - float[][] testMultiDimArray = {testFloatArray, testFloatArray, testFloatArray}; - dataType = NativeInterpreterWrapper.dataTypeOf(testFloatArray); - assertThat(dataType).isEqualTo(DataType.FLOAT32); - try { - double[] testDoubleArray = {0.783, 0.251}; - NativeInterpreterWrapper.dataTypeOf(testDoubleArray); - fail(); - } catch (IllegalArgumentException e) { - assertThat(e).hasMessageThat().contains("cannot resolve DataType of"); - } - try { - Float[] testBoxedArray = {0.783f, 0.251f}; - NativeInterpreterWrapper.dataTypeOf(testBoxedArray); - fail(); - } catch (IllegalArgumentException e) { - assertThat(e).hasMessageThat().contains("cannot resolve DataType of [Ljava.lang.Float;"); - } - } - - @Test - public void testNumDimensions() { - int scalar = 1; - assertThat(NativeInterpreterWrapper.numDimensions(scalar)).isEqualTo(0); - int[][] array = {{2, 4}, {1, 9}}; - assertThat(NativeInterpreterWrapper.numDimensions(array)).isEqualTo(2); - try { - int[] emptyArray = {}; - NativeInterpreterWrapper.numDimensions(emptyArray); - fail(); - } catch (IllegalArgumentException e) { - assertThat(e).hasMessageThat().contains("Array lengths cannot be 0."); - } - } - - @Test - public void testFillShape() { - int[][][] array = {{{23}, {14}, {87}}, {{12}, {42}, {31}}}; - int num = NativeInterpreterWrapper.numDimensions(array); - int[] shape = new int[num]; - NativeInterpreterWrapper.fillShape(array, 0, shape); - assertThat(num).isEqualTo(3); - assertThat(shape[0]).isEqualTo(2); - assertThat(shape[1]).isEqualTo(3); - assertThat(shape[2]).isEqualTo(1); - } - @Test public void testGetInferenceLatency() { NativeInterpreterWrapper wrapper = new NativeInterpreterWrapper(FLOAT_MODEL_PATH); @@ -465,8 +439,10 @@ public final class NativeInterpreterWrapperTest { float[][][] threeD = {twoD, twoD, twoD, twoD, twoD, twoD, twoD, twoD}; float[][][][] fourD = {threeD, threeD}; Object[] inputs = {fourD}; - Tensor[] outputs = wrapper.run(inputs); - assertThat(outputs.length).isEqualTo(1); + float[][][][] parsedOutputs = new float[2][8][8][3]; + Map outputs = new HashMap<>(); + outputs.put(0, parsedOutputs); + wrapper.run(inputs, outputs); assertThat(wrapper.getLastNativeInferenceDurationNanoseconds()).isGreaterThan(0L); wrapper.close(); } @@ -486,13 +462,14 @@ public final class NativeInterpreterWrapperTest { float[][][] threeD = {twoD, twoD, twoD, twoD, twoD, twoD, twoD, twoD}; float[][][][] fourD = {threeD, threeD}; Object[] inputs = {fourD}; + float[][][][] parsedOutputs = new float[2][8][8][3]; + Map outputs = new HashMap<>(); + outputs.put(0, parsedOutputs); try { - wrapper.run(inputs); + wrapper.run(inputs, outputs); fail(); } catch (IllegalArgumentException e) { - assertThat(e) - .hasMessageThat() - .contains("0-th input dimension should be [?,8,8,3], but found [?,8,7,3]"); + // Expected. } assertThat(wrapper.getLastNativeInferenceDurationNanoseconds()).isNull(); wrapper.close(); @@ -502,41 +479,7 @@ public final class NativeInterpreterWrapperTest { public void testGetInputDims() { NativeInterpreterWrapper wrapper = new NativeInterpreterWrapper(FLOAT_MODEL_PATH); int[] expectedDims = {1, 8, 8, 3}; - assertThat(wrapper.getInputDims(0)).isEqualTo(expectedDims); - wrapper.close(); - } - - @Test - public void testGetInputDimsOutOfRange() { - NativeInterpreterWrapper wrapper = new NativeInterpreterWrapper(FLOAT_MODEL_PATH); - try { - wrapper.getInputDims(-1); - fail(); - } catch (IllegalArgumentException e) { - assertThat(e).hasMessageThat().contains("Out of range"); - } - try { - wrapper.getInputDims(1); - fail(); - } catch (IllegalArgumentException e) { - assertThat(e).hasMessageThat().contains("Out of range"); - } - wrapper.close(); - } - - @Test - public void testGetOutputDataType() { - NativeInterpreterWrapper wrapper = new NativeInterpreterWrapper(FLOAT_MODEL_PATH); - assertThat(wrapper.getOutputDataType(0)).contains("float"); - wrapper.close(); - wrapper = new NativeInterpreterWrapper(LONG_MODEL_PATH); - assertThat(wrapper.getOutputDataType(0)).contains("long"); - wrapper.close(); - wrapper = new NativeInterpreterWrapper(INT_MODEL_PATH); - assertThat(wrapper.getOutputDataType(0)).contains("int"); - wrapper.close(); - wrapper = new NativeInterpreterWrapper(BYTE_MODEL_PATH); - assertThat(wrapper.getOutputDataType(0)).contains("byte"); + assertThat(wrapper.getInputTensor(0).shape()).isEqualTo(expectedDims); wrapper.close(); } diff --git a/tensorflow/contrib/lite/java/src/test/java/org/tensorflow/lite/TensorTest.java b/tensorflow/contrib/lite/java/src/test/java/org/tensorflow/lite/TensorTest.java index 94b6632bb8dd7117bf4074da1939bd23ce732efd..71ef04494357e8b951cbbbd2c68385b17c472736 100644 --- a/tensorflow/contrib/lite/java/src/test/java/org/tensorflow/lite/TensorTest.java +++ b/tensorflow/contrib/lite/java/src/test/java/org/tensorflow/lite/TensorTest.java @@ -18,6 +18,10 @@ package org.tensorflow.lite; import static com.google.common.truth.Truth.assertThat; import static org.junit.Assert.fail; +import java.nio.ByteBuffer; +import java.nio.ByteOrder; +import java.util.HashMap; +import java.util.Map; import org.junit.After; import org.junit.Before; import org.junit.Test; @@ -32,7 +36,7 @@ public final class TensorTest { "tensorflow/contrib/lite/java/src/testdata/add.bin"; private NativeInterpreterWrapper wrapper; - private long nativeHandle; + private Tensor tensor; @Before public void setUp() { @@ -42,8 +46,10 @@ public final class TensorTest { float[][][] threeD = {twoD, twoD, twoD, twoD, twoD, twoD, twoD, twoD}; float[][][][] fourD = {threeD, threeD}; Object[] inputs = {fourD}; - Tensor[] outputs = wrapper.run(inputs); - nativeHandle = outputs[0].nativeHandle; + Map outputs = new HashMap<>(); + outputs.put(0, new float[2][8][8][3]); + wrapper.run(inputs, outputs); + tensor = wrapper.getOutputTensor(0); } @After @@ -52,17 +58,16 @@ public final class TensorTest { } @Test - public void testFromHandle() throws Exception { - Tensor tensor = Tensor.fromHandle(nativeHandle); + public void testBasic() throws Exception { assertThat(tensor).isNotNull(); int[] expectedShape = {2, 8, 8, 3}; - assertThat(tensor.shapeCopy).isEqualTo(expectedShape); - assertThat(tensor.dtype).isEqualTo(DataType.FLOAT32); + assertThat(tensor.shape()).isEqualTo(expectedShape); + assertThat(tensor.dataType()).isEqualTo(DataType.FLOAT32); + assertThat(tensor.numBytes()).isEqualTo(2 * 8 * 8 * 3 * 4); } @Test public void testCopyTo() { - Tensor tensor = Tensor.fromHandle(nativeHandle); float[][][][] parsedOutputs = new float[2][8][8][3]; tensor.copyTo(parsedOutputs); float[] outputOneD = parsedOutputs[0][0][0]; @@ -70,9 +75,32 @@ public final class TensorTest { assertThat(outputOneD).usingTolerance(0.1f).containsExactly(expected).inOrder(); } + @Test + public void testCopyToByteBuffer() { + ByteBuffer parsedOutput = + ByteBuffer.allocateDirect(2 * 8 * 8 * 3 * 4).order(ByteOrder.nativeOrder()); + tensor.copyTo(parsedOutput); + assertThat(parsedOutput.position()).isEqualTo(2 * 8 * 8 * 3 * 4); + float[] outputOneD = { + parsedOutput.getFloat(0), parsedOutput.getFloat(4), parsedOutput.getFloat(8) + }; + float[] expected = {3.69f, 19.62f, 23.43f}; + assertThat(outputOneD).usingTolerance(0.1f).containsExactly(expected).inOrder(); + } + + @Test + public void testCopyToInvalidByteBuffer() { + ByteBuffer parsedOutput = ByteBuffer.allocateDirect(3 * 4).order(ByteOrder.nativeOrder()); + try { + tensor.copyTo(parsedOutput); + fail(); + } catch (IllegalArgumentException e) { + // Expected. + } + } + @Test public void testCopyToWrongType() { - Tensor tensor = Tensor.fromHandle(nativeHandle); int[][][][] parsedOutputs = new int[2][8][8][3]; try { tensor.copyTo(parsedOutputs); @@ -81,15 +109,13 @@ public final class TensorTest { assertThat(e) .hasMessageThat() .contains( - "Cannot convert an TensorFlowLite tensor with type " - + "FLOAT32 to a Java object of type [[[[I (which is compatible with the TensorFlowLite " - + "type INT32)"); + "Cannot convert between a TensorFlowLite tensor with type FLOAT32 and a Java object " + + "of type [[[[I (which is compatible with the TensorFlowLite type INT32)"); } } @Test public void testCopyToWrongShape() { - Tensor tensor = Tensor.fromHandle(nativeHandle); float[][][][] parsedOutputs = new float[1][8][8][3]; try { tensor.copyTo(parsedOutputs); @@ -98,8 +124,104 @@ public final class TensorTest { assertThat(e) .hasMessageThat() .contains( - "Shape of output target [1, 8, 8, 3] does not match " - + "with the shape of the Tensor [2, 8, 8, 3]."); + "Cannot copy between a TensorFlowLite tensor with shape [2, 8, 8, 3] " + + "and a Java object with shape [1, 8, 8, 3]."); + } + } + + @Test + public void testSetTo() { + float[][][][] input = new float[2][8][8][3]; + float[][][][] output = new float[2][8][8][3]; + ByteBuffer inputByteBuffer = + ByteBuffer.allocateDirect(2 * 8 * 8 * 3 * 4).order(ByteOrder.nativeOrder()); + + input[0][0][0][0] = 2.0f; + tensor.setTo(input); + tensor.copyTo(output); + assertThat(output[0][0][0][0]).isEqualTo(2.0f); + + inputByteBuffer.putFloat(0, 3.0f); + tensor.setTo(inputByteBuffer); + tensor.copyTo(output); + assertThat(output[0][0][0][0]).isEqualTo(3.0f); + } + + @Test + public void testSetToInvalidByteBuffer() { + ByteBuffer input = ByteBuffer.allocateDirect(3 * 4).order(ByteOrder.nativeOrder()); + try { + tensor.setTo(input); + fail(); + } catch (IllegalArgumentException e) { + // Success. + } + } + + @Test + public void testGetInputShapeIfDifferent() { + ByteBuffer bytBufferInput = ByteBuffer.allocateDirect(3 * 4).order(ByteOrder.nativeOrder()); + assertThat(tensor.getInputShapeIfDifferent(bytBufferInput)).isNull(); + + float[][][][] sameShapeInput = new float[2][8][8][3]; + assertThat(tensor.getInputShapeIfDifferent(sameShapeInput)).isNull(); + + float[][][][] differentShapeInput = new float[1][8][8][3]; + assertThat(tensor.getInputShapeIfDifferent(differentShapeInput)) + .isEqualTo(new int[] {1, 8, 8, 3}); + } + + @Test + public void testDataTypeOf() { + float[] testEmptyArray = {}; + DataType dataType = Tensor.dataTypeOf(testEmptyArray); + assertThat(dataType).isEqualTo(DataType.FLOAT32); + float[] testFloatArray = {0.783f, 0.251f}; + dataType = Tensor.dataTypeOf(testFloatArray); + assertThat(dataType).isEqualTo(DataType.FLOAT32); + float[][] testMultiDimArray = {testFloatArray, testFloatArray, testFloatArray}; + dataType = Tensor.dataTypeOf(testFloatArray); + assertThat(dataType).isEqualTo(DataType.FLOAT32); + try { + double[] testDoubleArray = {0.783, 0.251}; + Tensor.dataTypeOf(testDoubleArray); + fail(); + } catch (IllegalArgumentException e) { + assertThat(e).hasMessageThat().contains("cannot resolve DataType of"); } + try { + Float[] testBoxedArray = {0.783f, 0.251f}; + Tensor.dataTypeOf(testBoxedArray); + fail(); + } catch (IllegalArgumentException e) { + assertThat(e).hasMessageThat().contains("cannot resolve DataType of [Ljava.lang.Float;"); + } + } + + @Test + public void testNumDimensions() { + int scalar = 1; + assertThat(Tensor.numDimensions(scalar)).isEqualTo(0); + int[][] array = {{2, 4}, {1, 9}}; + assertThat(Tensor.numDimensions(array)).isEqualTo(2); + try { + int[] emptyArray = {}; + Tensor.numDimensions(emptyArray); + fail(); + } catch (IllegalArgumentException e) { + assertThat(e).hasMessageThat().contains("Array lengths cannot be 0."); + } + } + + @Test + public void testFillShape() { + int[][][] array = {{{23}, {14}, {87}}, {{12}, {42}, {31}}}; + int num = Tensor.numDimensions(array); + int[] shape = new int[num]; + Tensor.fillShape(array, 0, shape); + assertThat(num).isEqualTo(3); + assertThat(shape[0]).isEqualTo(2); + assertThat(shape[1]).isEqualTo(3); + assertThat(shape[2]).isEqualTo(1); } } diff --git a/tensorflow/contrib/lite/java/src/testhelper/java/org/tensorflow/lite/TestHelper.java b/tensorflow/contrib/lite/java/src/testhelper/java/org/tensorflow/lite/TestHelper.java index 3aef0c3bb6cc4748de0e55d31f0215a77320ae69..c23521c0774ebab01f38db8b416020ae5755cee9 100644 --- a/tensorflow/contrib/lite/java/src/testhelper/java/org/tensorflow/lite/TestHelper.java +++ b/tensorflow/contrib/lite/java/src/testhelper/java/org/tensorflow/lite/TestHelper.java @@ -58,7 +58,7 @@ public class TestHelper { */ public static int[] getInputDims(Interpreter interpreter, int index) { if (interpreter != null && interpreter.wrapper != null) { - return interpreter.wrapper.getInputDims(index); + return interpreter.wrapper.getInputTensor(index).shape(); } else { throw new IllegalArgumentException( "Interpreter has not initialized;" + " Failed to get input dimensions."); @@ -77,7 +77,7 @@ public class TestHelper { */ public static String getOutputDataType(Interpreter interpreter, int index) { if (interpreter != null && interpreter.wrapper != null) { - return interpreter.wrapper.getOutputDataType(index); + return interpreter.wrapper.getOutputTensor(index).dataType().toStringName(); } else { throw new IllegalArgumentException( "Interpreter has not initialized;" + " Failed to get output data type."); diff --git a/tensorflow/contrib/lite/kernels/BUILD b/tensorflow/contrib/lite/kernels/BUILD index a77897a173fc1bd9ceb63e6918ebbfb69f6d6af1..33594c138b0eed33dbc2f9f2bfc458477fc44c5c 100644 --- a/tensorflow/contrib/lite/kernels/BUILD +++ b/tensorflow/contrib/lite/kernels/BUILD @@ -46,11 +46,17 @@ cc_library( hdrs = [ "eigen_support.h", ], - copts = tflite_copts(), + copts = tflite_copts() + [ + "-Wno-error=reorder", + ] + select({ + "//tensorflow:ios": ["-Wno-error=invalid-partial-specialization"], + "//conditions:default": [ + ], + }), deps = [ ":op_macros", "//tensorflow/contrib/lite:context", - "//third_party/eigen3", + "//tensorflow/contrib/lite/kernels/internal:optimized", ], ) @@ -130,7 +136,7 @@ cc_library( srcs = [ "activations.cc", "add.cc", - "arg_max.cc", + "arg_min_max.cc", "audio_spectrogram.cc", "basic_rnn.cc", "batch_to_space_nd.cc", @@ -149,6 +155,7 @@ cc_library( "embedding_lookup_sparse.cc", "exp.cc", "expand_dims.cc", + "fake_quant.cc", "floor.cc", "fully_connected.cc", "gather.cc", @@ -163,6 +170,7 @@ cc_library( "neg.cc", "pad.cc", "pooling.cc", + "pow.cc", "reduce.cc", "register.cc", "reshape.cc", @@ -289,9 +297,9 @@ tf_cc_test( ) tf_cc_test( - name = "arg_max_test", + name = "arg_min_max_test", size = "small", - srcs = ["arg_max_test.cc"], + srcs = ["arg_min_max_test.cc"], tags = [ "tflite_not_portable_ios", ], @@ -556,6 +564,19 @@ tf_cc_test( ], ) +tf_cc_test( + name = "fake_quant_test", + size = "small", + srcs = ["fake_quant_test.cc"], + tags = ["tflite_not_portable_ios"], + deps = [ + ":builtin_ops", + "//tensorflow/contrib/lite:framework", + "//tensorflow/contrib/lite/kernels:test_util", + "@com_google_googletest//:gtest", + ], +) + tf_cc_test( name = "maximum_minimum_test", size = "small", @@ -1009,6 +1030,20 @@ tf_cc_test( ], ) +tf_cc_test( + name = "pow_test", + size = "small", + srcs = ["pow_test.cc"], + tags = ["tflite_not_portable_ios"], + deps = [ + ":builtin_ops", + "//tensorflow/contrib/lite:builtin_op_data", + "//tensorflow/contrib/lite:framework", + "//tensorflow/contrib/lite/kernels:test_util", + "@com_google_googletest//:gtest", + ], +) + filegroup( name = "all_files", srcs = glob( diff --git a/tensorflow/contrib/lite/kernels/add.cc b/tensorflow/contrib/lite/kernels/add.cc index ccb957ebc52e6ce9db3fbffb0c5beca9409edcc0..f44d531cbfa9ed41f881380752558555aab97b4d 100644 --- a/tensorflow/contrib/lite/kernels/add.cc +++ b/tensorflow/contrib/lite/kernels/add.cc @@ -170,29 +170,44 @@ TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) { } template -void EvalAddFloat(TfLiteContext* context, TfLiteNode* node, - TfLiteAddParams* params, const OpData* data, - const TfLiteTensor* input1, const TfLiteTensor* input2, - TfLiteTensor* output) { - float output_activation_min, output_activation_max; - CalculateActivationRangeFloat(params->activation, &output_activation_min, - &output_activation_max); -#define TF_LITE_ADD(type, opname) \ - type::opname(GetTensorData(input1), GetTensorDims(input1), \ - GetTensorData(input2), GetTensorDims(input2), \ - output_activation_min, output_activation_max, \ - GetTensorData(output), GetTensorDims(output)) - if (kernel_type == kReference) { - if (data->requires_broadcast) { - TF_LITE_ADD(reference_ops, BroadcastAdd); +void EvalAdd(TfLiteContext* context, TfLiteNode* node, TfLiteAddParams* params, + const OpData* data, const TfLiteTensor* input1, + const TfLiteTensor* input2, TfLiteTensor* output) { +#define TF_LITE_ADD(type, opname, data_type) \ + data_type output_activation_min, output_activation_max; \ + CalculateActivationRange(params->activation, &output_activation_min, \ + &output_activation_max); \ + type::opname(GetTensorData(input1), GetTensorDims(input1), \ + GetTensorData(input2), GetTensorDims(input2), \ + output_activation_min, output_activation_max, \ + GetTensorData(output), GetTensorDims(output)) + if (output->type == kTfLiteInt32) { + if (kernel_type == kReference) { + if (data->requires_broadcast) { + TF_LITE_ADD(reference_ops, BroadcastAdd, int32_t); + } else { + TF_LITE_ADD(reference_ops, Add, int32_t); + } } else { - TF_LITE_ADD(reference_ops, Add); + if (data->requires_broadcast) { + TF_LITE_ADD(optimized_ops, BroadcastAdd, int32_t); + } else { + TF_LITE_ADD(optimized_ops, Add, int32_t); + } } - } else { - if (data->requires_broadcast) { - TF_LITE_ADD(optimized_ops, BroadcastAdd); + } else if (output->type == kTfLiteFloat32) { + if (kernel_type == kReference) { + if (data->requires_broadcast) { + TF_LITE_ADD(reference_ops, BroadcastAdd, float); + } else { + TF_LITE_ADD(reference_ops, Add, float); + } } else { - TF_LITE_ADD(optimized_ops, Add); + if (data->requires_broadcast) { + TF_LITE_ADD(optimized_ops, BroadcastAdd, float); + } else { + TF_LITE_ADD(optimized_ops, Add, float); + } } } #undef TF_LITE_ADD @@ -251,9 +266,8 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { const TfLiteTensor* input2 = GetInput(context, node, kInputTensor2); TfLiteTensor* output = GetOutput(context, node, kOutputTensor); - if (output->type == kTfLiteFloat32) { - EvalAddFloat(context, node, params, data, input1, input2, - output); + if (output->type == kTfLiteFloat32 || output->type == kTfLiteInt32) { + EvalAdd(context, node, params, data, input1, input2, output); } else if (output->type == kTfLiteUInt8 || output->type == kTfLiteInt16) { TF_LITE_ENSURE_OK(context, EvalAddQuantized(context, node, params, data, diff --git a/tensorflow/contrib/lite/kernels/add_test.cc b/tensorflow/contrib/lite/kernels/add_test.cc index 456a754e7ee191fa74280da7af8fa844b2ef1923..0b5844321133de103919de76d367574f018a6698 100644 --- a/tensorflow/contrib/lite/kernels/add_test.cc +++ b/tensorflow/contrib/lite/kernels/add_test.cc @@ -52,6 +52,13 @@ class FloatAddOpModel : public BaseAddOpModel { std::vector GetOutput() { return ExtractVector(output_); } }; +class IntegerAddOpModel : public BaseAddOpModel { + public: + using BaseAddOpModel::BaseAddOpModel; + + std::vector GetOutput() { return ExtractVector(output_); } +}; + class QuantizedAddOpModel : public BaseAddOpModel { public: using BaseAddOpModel::BaseAddOpModel; @@ -133,6 +140,57 @@ TEST(FloatAddOpModel, WithBroadcast) { } } +TEST(IntegerAddOpModel, NoActivation) { + IntegerAddOpModel m({TensorType_INT32, {1, 2, 2, 1}}, + {TensorType_INT32, {1, 2, 2, 1}}, {TensorType_INT32, {}}, + ActivationFunctionType_NONE); + m.PopulateTensor(m.input1(), {-20, 2, 7, 8}); + m.PopulateTensor(m.input2(), {1, 2, 3, 5}); + m.Invoke(); + EXPECT_THAT(m.GetOutput(), ElementsAreArray({-19, 4, 10, 13})); +} + +TEST(IntegerAddOpModel, ActivationRELU_N1_TO_1) { + IntegerAddOpModel m({TensorType_INT32, {1, 2, 2, 1}}, + {TensorType_INT32, {1, 2, 2, 1}}, {TensorType_INT32, {}}, + ActivationFunctionType_RELU_N1_TO_1); + m.PopulateTensor(m.input1(), {-20, 2, 7, 8}); + m.PopulateTensor(m.input2(), {1, 2, 3, 5}); + m.Invoke(); + EXPECT_THAT(m.GetOutput(), ElementsAreArray({-1, 1, 1, 1})); +} + +TEST(IntegerAddOpModel, VariousInputShapes) { + std::vector> test_shapes = { + {6}, {2, 3}, {2, 1, 3}, {1, 3, 1, 2}}; + for (int i = 0; i < test_shapes.size(); ++i) { + IntegerAddOpModel m({TensorType_INT32, test_shapes[i]}, + {TensorType_INT32, test_shapes[i]}, + {TensorType_INT32, {}}, ActivationFunctionType_NONE); + m.PopulateTensor(m.input1(), {-20, 2, 7, 8, 11, 20}); + m.PopulateTensor(m.input2(), {1, 2, 3, 5, 11, 1}); + m.Invoke(); + EXPECT_THAT(m.GetOutput(), ElementsAreArray({-19, 04, 10, 13, 22, 21})) + << "With shape number " << i; + } +} + +TEST(IntegerAddOpModel, WithBroadcast) { + std::vector> test_shapes = { + {6}, {2, 3}, {2, 1, 3}, {1, 3, 1, 2}}; + for (int i = 0; i < test_shapes.size(); ++i) { + IntegerAddOpModel m({TensorType_INT32, test_shapes[i]}, + {TensorType_INT32, {}}, // always a scalar + {TensorType_INT32, {}}, ActivationFunctionType_NONE); + m.PopulateTensor(m.input1(), {-20, 2, 7, 8, 11, 20}); + m.PopulateTensor(m.input2(), {1}); + m.Invoke(); + EXPECT_THAT(m.GetOutput(), + ElementsAreArray(ArrayFloatNear({-19, 3, 8, 9, 12, 21}))) + << "With shape number " << i; + } +} + TEST(QuantizedAddOpModel, QuantizedTestsNoActivation) { float kQuantizedTolerance = GetTolerance(-1.0, 1.0); std::vector> inputs1 = { diff --git a/tensorflow/contrib/lite/kernels/arg_max.cc b/tensorflow/contrib/lite/kernels/arg_min_max.cc similarity index 70% rename from tensorflow/contrib/lite/kernels/arg_max.cc rename to tensorflow/contrib/lite/kernels/arg_min_max.cc index 26f57e88962116f446e72fbc164d2747e8b633b4..4f30d09030fb8d26c08090b180fdd352a967807f 100644 --- a/tensorflow/contrib/lite/kernels/arg_max.cc +++ b/tensorflow/contrib/lite/kernels/arg_min_max.cc @@ -23,7 +23,7 @@ limitations under the License. namespace tflite { namespace ops { namespace builtin { -namespace arg_max { +namespace arg_min_max { constexpr int kInputTensor = 0; constexpr int kAxis = 1; @@ -80,30 +80,39 @@ TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) { return context->ResizeTensor(context, output, output_size); } +template +std::function GetComparefunction(bool is_arg_max) { + if (is_arg_max) { + return std::greater(); + } else { + return std::less(); + } +} + // The current impl actually ignores the axis argument. // Only determine the index of the maximum value in the last dimension. -TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { +TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node, bool is_arg_max) { const TfLiteTensor* input = GetInput(context, node, kInputTensor); const TfLiteTensor* axis = GetInput(context, node, kAxis); TfLiteTensor* output = GetOutput(context, node, kOutputTensor); -#define TF_LITE_ARG_MAX(data_type, axis_type, output_type) \ - optimized_ops::ArgMax(GetTensorData(axis), \ - GetTensorData(input), GetTensorDims(input), \ - GetTensorData(output), \ - GetTensorDims(output)) +#define TF_LITE_ARG_MIN_MAX(data_type, axis_type, output_type) \ + optimized_ops::ArgMinMax( \ + GetTensorData(axis), GetTensorData(input), \ + GetTensorDims(input), GetTensorData(output), \ + GetTensorDims(output), GetComparefunction(is_arg_max)) if (axis->type == kTfLiteInt32) { switch (output->type) { case kTfLiteInt32: { switch (input->type) { case kTfLiteFloat32: - TF_LITE_ARG_MAX(float, int32_t, int32_t); + TF_LITE_ARG_MIN_MAX(float, int32_t, int32_t); break; case kTfLiteUInt8: - TF_LITE_ARG_MAX(uint8_t, int32_t, int32_t); + TF_LITE_ARG_MIN_MAX(uint8_t, int32_t, int32_t); break; case kTfLiteInt32: - TF_LITE_ARG_MAX(int32_t, int32_t, int32_t); + TF_LITE_ARG_MIN_MAX(int32_t, int32_t, int32_t); break; default: return kTfLiteError; @@ -112,13 +121,13 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { case kTfLiteInt64: { switch (input->type) { case kTfLiteFloat32: - TF_LITE_ARG_MAX(float, int32_t, int64_t); + TF_LITE_ARG_MIN_MAX(float, int32_t, int64_t); break; case kTfLiteUInt8: - TF_LITE_ARG_MAX(uint8_t, int32_t, int64_t); + TF_LITE_ARG_MIN_MAX(uint8_t, int32_t, int64_t); break; case kTfLiteInt32: - TF_LITE_ARG_MAX(int32_t, int32_t, int64_t); + TF_LITE_ARG_MIN_MAX(int32_t, int32_t, int64_t); break; default: return kTfLiteError; @@ -132,13 +141,13 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { case kTfLiteInt32: { switch (input->type) { case kTfLiteFloat32: - TF_LITE_ARG_MAX(float, int64_t, int32_t); + TF_LITE_ARG_MIN_MAX(float, int64_t, int32_t); break; case kTfLiteUInt8: - TF_LITE_ARG_MAX(uint8_t, int64_t, int32_t); + TF_LITE_ARG_MIN_MAX(uint8_t, int64_t, int32_t); break; case kTfLiteInt32: - TF_LITE_ARG_MAX(int32_t, int64_t, int32_t); + TF_LITE_ARG_MIN_MAX(int32_t, int64_t, int32_t); break; default: return kTfLiteError; @@ -147,13 +156,13 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { case kTfLiteInt64: { switch (input->type) { case kTfLiteFloat32: - TF_LITE_ARG_MAX(float, int64_t, int64_t); + TF_LITE_ARG_MIN_MAX(float, int64_t, int64_t); break; case kTfLiteUInt8: - TF_LITE_ARG_MAX(uint8_t, int64_t, int64_t); + TF_LITE_ARG_MIN_MAX(uint8_t, int64_t, int64_t); break; case kTfLiteInt32: - TF_LITE_ARG_MAX(int32_t, int64_t, int64_t); + TF_LITE_ARG_MIN_MAX(int32_t, int64_t, int64_t); break; default: return kTfLiteError; @@ -163,16 +172,30 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { return kTfLiteError; } } -#undef TF_LITE_ARG_MAX +#undef TF_LITE_ARG_MIN_MAX return kTfLiteOk; } -} // namespace arg_max +TfLiteStatus ArgMinEval(TfLiteContext* context, TfLiteNode* node) { + return Eval(context, node, false); +} + +TfLiteStatus ArgMaxEval(TfLiteContext* context, TfLiteNode* node) { + return Eval(context, node, true); +} + +} // namespace arg_min_max TfLiteRegistration* Register_ARG_MAX() { - static TfLiteRegistration r = {nullptr, nullptr, arg_max::Prepare, - arg_max::Eval}; + static TfLiteRegistration r = {nullptr, nullptr, arg_min_max::Prepare, + arg_min_max::ArgMaxEval}; + return &r; +} + +TfLiteRegistration* Register_ARG_MIN() { + static TfLiteRegistration r = {nullptr, nullptr, arg_min_max::Prepare, + arg_min_max::ArgMinEval}; return &r; } diff --git a/tensorflow/contrib/lite/kernels/arg_max_test.cc b/tensorflow/contrib/lite/kernels/arg_min_max_test.cc similarity index 52% rename from tensorflow/contrib/lite/kernels/arg_max_test.cc rename to tensorflow/contrib/lite/kernels/arg_min_max_test.cc index 31b15fe19ab87027c28bde9eaff7d88d03b2c213..90e5fdc532c821691aaeca6e6faa4c24919ca2c8 100644 --- a/tensorflow/contrib/lite/kernels/arg_max_test.cc +++ b/tensorflow/contrib/lite/kernels/arg_min_max_test.cc @@ -24,16 +24,13 @@ namespace { using ::testing::ElementsAreArray; template -class ArgMaxOpModel : public SingleOpModel { +class ArgBaseOpModel : public SingleOpModel { public: - ArgMaxOpModel(std::initializer_list input_shape, TensorType input_type, - TensorType output_type, TensorType index_output_type) { + ArgBaseOpModel(std::initializer_list input_shape, TensorType input_type, + TensorType output_type, TensorType index_output_type) { input_ = AddInput(input_type); axis_ = AddInput(TensorType_INT32); output_ = AddOutput(output_type); - SetBuiltinOp(BuiltinOperator_ARG_MAX, BuiltinOptions_ArgMaxOptions, - CreateArgMaxOptions(builder_, index_output_type).Union()); - BuildInterpreter({input_shape, {1, 1, 1, 1}}); } int input() { return input_; } @@ -42,12 +39,42 @@ class ArgMaxOpModel : public SingleOpModel { std::vector GetOutput() { return ExtractVector(output_); } std::vector GetOutputShape() { return GetTensorShape(output_); } - private: + protected: int input_; int axis_; int output_; }; +template +class ArgMaxOpModel : public ArgBaseOpModel { + public: + ArgMaxOpModel(std::initializer_list input_shape, TensorType input_type, + TensorType output_type, TensorType index_output_type) + : ArgBaseOpModel(input_shape, input_type, output_type, + index_output_type) { + ArgBaseOpModel::SetBuiltinOp( + BuiltinOperator_ARG_MAX, BuiltinOptions_ArgMaxOptions, + CreateArgMaxOptions(ArgBaseOpModel::builder_, index_output_type) + .Union()); + ArgBaseOpModel::BuildInterpreter({input_shape, {1, 1, 1, 1}}); + } +}; + +template +class ArgMinOpModel : public ArgBaseOpModel { + public: + ArgMinOpModel(std::initializer_list input_shape, TensorType input_type, + TensorType output_type, TensorType index_output_type) + : ArgBaseOpModel(input_shape, input_type, output_type, + index_output_type) { + ArgBaseOpModel::SetBuiltinOp( + BuiltinOperator_ARG_MIN, BuiltinOptions_ArgMinOptions, + CreateArgMinOptions(ArgBaseOpModel::builder_, index_output_type) + .Union()); + ArgBaseOpModel::BuildInterpreter({input_shape, {1, 1, 1, 1}}); + } +}; + TEST(ArgMaxOpTest, GetMaxArgFloat) { ArgMaxOpModel model({1, 1, 1, 4}, TensorType_FLOAT32, TensorType_INT32, TensorType_INT32); @@ -96,6 +123,54 @@ TEST(ArgMaxOpTest, GetMaxArgOutput64) { EXPECT_THAT(model.GetOutputShape(), ElementsAreArray({1, 1, 2, 1})); } +TEST(ArgMinOpTest, GetMinArgFloat) { + ArgMinOpModel model({1, 1, 1, 4}, TensorType_FLOAT32, + TensorType_INT32, TensorType_INT32); + model.PopulateTensor(model.input(), {0.1, 0.9, 0.7, 0.3}); + // Currently only support the last dimension. + model.PopulateTensor(model.axis(), {3}); + model.Invoke(); + + EXPECT_THAT(model.GetOutput(), ElementsAreArray({0})); + EXPECT_THAT(model.GetOutputShape(), ElementsAreArray({1, 1, 1, 1})); +} + +TEST(ArgMinOpTest, GetMinArgInt) { + ArgMinOpModel model({1, 1, 1, 4}, TensorType_INT32, TensorType_INT32, + TensorType_INT32); + model.PopulateTensor(model.input(), {1, 9, 7, 3}); + // Currently only support the last dimension. + model.PopulateTensor(model.axis(), {3}); + model.Invoke(); + + EXPECT_THAT(model.GetOutput(), ElementsAreArray({0})); + EXPECT_THAT(model.GetOutputShape(), ElementsAreArray({1, 1, 1, 1})); +} + +TEST(ArgMinOpTest, GetMinArgMulDimensions) { + ArgMinOpModel model({1, 1, 2, 4}, TensorType_INT32, TensorType_INT32, + TensorType_INT32); + model.PopulateTensor(model.input(), {1, 2, 7, 8, 1, 9, 7, 3}); + // Currently only support the last dimension. + model.PopulateTensor(model.axis(), {3}); + model.Invoke(); + + EXPECT_THAT(model.GetOutput(), ElementsAreArray({0, 0})); + EXPECT_THAT(model.GetOutputShape(), ElementsAreArray({1, 1, 2, 1})); +} + +TEST(ArgMinOpTest, GetMinArgOutput64) { + ArgMinOpModel model({1, 1, 2, 4}, TensorType_INT32, TensorType_INT64, + TensorType_INT64); + model.PopulateTensor(model.input(), {10, 2, 7, 8, 1, 9, 7, 3}); + // Currently only support the last dimension. + model.PopulateTensor(model.axis(), {3}); + model.Invoke(); + + EXPECT_THAT(model.GetOutput(), ElementsAreArray({1, 0})); + EXPECT_THAT(model.GetOutputShape(), ElementsAreArray({1, 1, 2, 1})); +} + } // namespace } // namespace tflite diff --git a/tensorflow/contrib/lite/kernels/bidirectional_sequence_lstm.cc b/tensorflow/contrib/lite/kernels/bidirectional_sequence_lstm.cc index 3425288f027a6fd9eb65f730bc7d039c832ace1c..14a19aeef390a71b99c4c9dff036746d102c9e9c 100644 --- a/tensorflow/contrib/lite/kernels/bidirectional_sequence_lstm.cc +++ b/tensorflow/contrib/lite/kernels/bidirectional_sequence_lstm.cc @@ -276,27 +276,33 @@ TfLiteStatus CheckLstmTensorDimensions( TfLiteStatus CheckInputTensorDimensions(TfLiteContext* context, TfLiteNode* node, int n_input, int n_output, int n_cell) { - CheckLstmTensorDimensions( - context, node, n_input, n_output, n_cell, kFwInputToInputWeightsTensor, - kFwInputToForgetWeightsTensor, kFwInputToCellWeightsTensor, - kFwInputToOutputWeightsTensor, kFwRecurrentToInputWeightsTensor, - kFwRecurrentToForgetWeightsTensor, kFwRecurrentToCellWeightsTensor, - kFwRecurrentToOutputWeightsTensor, kFwCellToInputWeightsTensor, - kFwCellToForgetWeightsTensor, kFwCellToOutputWeightsTensor, - kFwInputGateBiasTensor, kFwForgetGateBiasTensor, kFwCellGateBiasTensor, - kFwOutputGateBiasTensor, kFwProjectionWeightsTensor, - kFwProjectionBiasTensor); - - CheckLstmTensorDimensions( - context, node, n_input, n_output, n_cell, kBwInputToInputWeightsTensor, - kBwInputToForgetWeightsTensor, kBwInputToCellWeightsTensor, - kBwInputToOutputWeightsTensor, kBwRecurrentToInputWeightsTensor, - kBwRecurrentToForgetWeightsTensor, kBwRecurrentToCellWeightsTensor, - kBwRecurrentToOutputWeightsTensor, kBwCellToInputWeightsTensor, - kBwCellToForgetWeightsTensor, kBwCellToOutputWeightsTensor, - kBwInputGateBiasTensor, kBwForgetGateBiasTensor, kBwCellGateBiasTensor, - kBwOutputGateBiasTensor, kBwProjectionWeightsTensor, - kBwProjectionBiasTensor); + TF_LITE_ENSURE_OK( + context, + CheckLstmTensorDimensions( + context, node, n_input, n_output, n_cell, + kFwInputToInputWeightsTensor, kFwInputToForgetWeightsTensor, + kFwInputToCellWeightsTensor, kFwInputToOutputWeightsTensor, + kFwRecurrentToInputWeightsTensor, kFwRecurrentToForgetWeightsTensor, + kFwRecurrentToCellWeightsTensor, kFwRecurrentToOutputWeightsTensor, + kFwCellToInputWeightsTensor, kFwCellToForgetWeightsTensor, + kFwCellToOutputWeightsTensor, kFwInputGateBiasTensor, + kFwForgetGateBiasTensor, kFwCellGateBiasTensor, + kFwOutputGateBiasTensor, kFwProjectionWeightsTensor, + kFwProjectionBiasTensor)); + + TF_LITE_ENSURE_OK( + context, + CheckLstmTensorDimensions( + context, node, n_input, n_output, n_cell, + kBwInputToInputWeightsTensor, kBwInputToForgetWeightsTensor, + kBwInputToCellWeightsTensor, kBwInputToOutputWeightsTensor, + kBwRecurrentToInputWeightsTensor, kBwRecurrentToForgetWeightsTensor, + kBwRecurrentToCellWeightsTensor, kBwRecurrentToOutputWeightsTensor, + kBwCellToInputWeightsTensor, kBwCellToForgetWeightsTensor, + kBwCellToOutputWeightsTensor, kBwInputGateBiasTensor, + kBwForgetGateBiasTensor, kBwCellGateBiasTensor, + kBwOutputGateBiasTensor, kBwProjectionWeightsTensor, + kBwProjectionBiasTensor)); // Check if Forward and Backward tensors match along required dimensions. return kTfLiteOk; @@ -334,7 +340,9 @@ TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) { const int n_fw_output = fw_recurrent_to_output_weights->dims->data[1]; // Check that input tensor dimensions matches with each other. - CheckInputTensorDimensions(context, node, n_input, n_fw_output, n_fw_cell); + TF_LITE_ENSURE_OK( + context, CheckInputTensorDimensions(context, node, n_input, n_fw_output, + n_fw_cell)); // Get the pointer to output, state and scratch buffer tensors. TfLiteTensor* fw_output = GetOutput(context, node, kFwOutputTensor); @@ -404,7 +412,9 @@ TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) { const int n_bw_output = bw_recurrent_to_output_weights->dims->data[1]; // Check that input tensor dimensions matches with each other. - CheckInputTensorDimensions(context, node, n_input, n_bw_output, n_bw_cell); + TF_LITE_ENSURE_OK( + context, CheckInputTensorDimensions(context, node, n_input, n_bw_output, + n_bw_cell)); // Get the pointer to output, output_state and cell_state buffer tensors. TfLiteTensor* bw_output = GetOutput(context, node, kBwOutputTensor); diff --git a/tensorflow/contrib/lite/kernels/cast.cc b/tensorflow/contrib/lite/kernels/cast.cc index 60770ca0aa8b85d9710d26beca3d4d603da5db2f..8dd48af57fd1bd9ef21256410d6bede6b7baa566 100644 --- a/tensorflow/contrib/lite/kernels/cast.cc +++ b/tensorflow/contrib/lite/kernels/cast.cc @@ -14,6 +14,7 @@ limitations under the License. ==============================================================================*/ #include #include +#include #include "tensorflow/contrib/lite/builtin_op_data.h" #include "tensorflow/contrib/lite/context.h" #include "tensorflow/contrib/lite/kernels/internal/optimized/optimized_ops.h" @@ -53,6 +54,20 @@ void copyCast(const FromT* in, ToT* out, int num_elements) { [](FromT a) { return static_cast(a); }); } +template +void copyCast(const std::complex* in, ToT* out, int num_elements) { + std::transform(in, in + num_elements, out, [](std::complex a) { + return static_cast(std::real(a)); + }); +} + +template <> +void copyCast(const std::complex* in, std::complex* out, + int num_elements) { + std::transform(in, in + num_elements, out, + [](std::complex a) { return a; }); +} + template TfLiteStatus copyToTensor(const FromT* in, TfLiteTensor* out, int num_elements) { @@ -72,6 +87,10 @@ TfLiteStatus copyToTensor(const FromT* in, TfLiteTensor* out, case kTfLiteBool: copyCast(in, out->data.b, num_elements); break; + case kTfLiteComplex64: + copyCast(in, reinterpret_cast*>(out->data.c64), + num_elements); + break; default: // Unsupported type. return kTfLiteError; @@ -95,6 +114,10 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { return copyToTensor(input->data.f, output, num_elements); case kTfLiteBool: return copyToTensor(input->data.b, output, num_elements); + case kTfLiteComplex64: + return copyToTensor( + reinterpret_cast*>(input->data.c64), output, + num_elements); default: // Unsupported type. return kTfLiteError; diff --git a/tensorflow/contrib/lite/kernels/cast_test.cc b/tensorflow/contrib/lite/kernels/cast_test.cc index 53e20007378392467356ab29ecb8b217bb7a9e89..954f998206563a38c74a1382092851cfbee1013b 100644 --- a/tensorflow/contrib/lite/kernels/cast_test.cc +++ b/tensorflow/contrib/lite/kernels/cast_test.cc @@ -12,6 +12,8 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ +#include + #include #include "tensorflow/contrib/lite/interpreter.h" #include "tensorflow/contrib/lite/kernels/register.h" @@ -73,6 +75,71 @@ TEST(CastOpModel, CastBoolToFloat) { ElementsAreArray({1.f, 1.0f, 0.f, 1.0f, 0.0f, 1.0f})); } +TEST(CastOpModel, CastComplex64ToFloat) { + CastOpModel m({TensorType_COMPLEX64, {2, 3}}, {TensorType_FLOAT32, {2, 3}}); + m.PopulateTensor>( + m.input(), + {std::complex(1.0f, 11.0f), std::complex(2.0f, 12.0f), + std::complex(3.0f, 13.0f), std::complex(4.0f, 14.0f), + std::complex(5.0f, 15.0f), std::complex(6.0f, 16.0f)}); + m.Invoke(); + EXPECT_THAT(m.ExtractVector(m.output()), + ElementsAreArray({1.0f, 2.0f, 3.0f, 4.0f, 5.0f, 6.0f})); +} + +TEST(CastOpModel, CastFloatToComplex64) { + CastOpModel m({TensorType_FLOAT32, {2, 3}}, {TensorType_COMPLEX64, {2, 3}}); + m.PopulateTensor(m.input(), {1.0f, 2.0f, 3.0f, 4.0f, 5.0f, 6.0f}); + m.Invoke(); + EXPECT_THAT( + m.ExtractVector>(m.output()), + ElementsAreArray( + {std::complex(1.0f, 0.0f), std::complex(2.0f, 0.0f), + std::complex(3.0f, 0.0f), std::complex(4.0f, 0.0f), + std::complex(5.0f, 0.0f), std::complex(6.0f, 0.0f)})); +} + +TEST(CastOpModel, CastComplex64ToInt) { + CastOpModel m({TensorType_COMPLEX64, {2, 3}}, {TensorType_INT32, {2, 3}}); + m.PopulateTensor>( + m.input(), + {std::complex(1.0f, 11.0f), std::complex(2.0f, 12.0f), + std::complex(3.0f, 13.0f), std::complex(4.0f, 14.0f), + std::complex(5.0f, 15.0f), std::complex(6.0f, 16.0f)}); + m.Invoke(); + EXPECT_THAT(m.ExtractVector(m.output()), + ElementsAreArray({1, 2, 3, 4, 5, 6})); +} + +TEST(CastOpModel, CastIntToComplex64) { + CastOpModel m({TensorType_INT32, {2, 3}}, {TensorType_COMPLEX64, {2, 3}}); + m.PopulateTensor(m.input(), {1, 2, 3, 4, 5, 6}); + m.Invoke(); + EXPECT_THAT( + m.ExtractVector>(m.output()), + ElementsAreArray( + {std::complex(1.0f, 0.0f), std::complex(2.0f, 0.0f), + std::complex(3.0f, 0.0f), std::complex(4.0f, 0.0f), + std::complex(5.0f, 0.0f), std::complex(6.0f, 0.0f)})); +} + +TEST(CastOpModel, CastComplex64ToComplex64) { + CastOpModel m({TensorType_COMPLEX64, {2, 3}}, {TensorType_COMPLEX64, {2, 3}}); + m.PopulateTensor>( + m.input(), + {std::complex(1.0f, 11.0f), std::complex(2.0f, 12.0f), + std::complex(3.0f, 13.0f), std::complex(4.0f, 14.0f), + std::complex(5.0f, 15.0f), std::complex(6.0f, 16.0f)}); + m.Invoke(); + EXPECT_THAT( + m.ExtractVector>(m.output()), + ElementsAreArray( + {std::complex(1.0f, 11.0f), std::complex(2.0f, 12.0f), + std::complex(3.0f, 13.0f), std::complex(4.0f, 14.0f), + std::complex(5.0f, 15.0f), + std::complex(6.0f, 16.0f)})); +} + } // namespace } // namespace tflite int main(int argc, char** argv) { diff --git a/tensorflow/contrib/lite/kernels/conv.cc b/tensorflow/contrib/lite/kernels/conv.cc index 93267f9a4f1426e053f4132ab57aae024cbb1bf9..a4fe9e55506bd01ce6c9142777c2ac632b37d46a 100644 --- a/tensorflow/contrib/lite/kernels/conv.cc +++ b/tensorflow/contrib/lite/kernels/conv.cc @@ -309,18 +309,8 @@ TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) { TfLiteTensor* hwcn_weights = &context->tensors[node->temporaries->data[data->hwcn_weights_index]]; hwcn_weights->type = data_type; - hwcn_weights->allocation_type = kTfLiteDynamic; - // Make sure we release any previous allocations before we reallocate. - // TODO(petewarden): Persistent arenas would be a better fit for this, but - // they aren't fully implemented yet. - if (hwcn_weights->data.raw) { - free(hwcn_weights->data.raw); - hwcn_weights->data.raw = nullptr; - } + hwcn_weights->allocation_type = kTfLiteArenaRwPersistent; - // Note that hwcn_weights_status is a kTfLiteDynamic tensor, and - // ResizeTensor will actually allocate space for it. The would be more - // efficient if we placed hwcn_weights_status in the persistent arena. auto hwcn_weights_status = context->ResizeTensor(context, hwcn_weights, hwcn_weights_size); if (hwcn_weights_status != kTfLiteOk) return hwcn_weights_status; @@ -382,8 +372,8 @@ void EvalFloat(TfLiteContext* context, TfLiteNode* node, TfLiteTensor* filter, TfLiteTensor* bias, TfLiteTensor* im2col, TfLiteTensor* hwcn_weights, TfLiteTensor* output) { float output_activation_min, output_activation_max; - CalculateActivationRangeFloat(params->activation, &output_activation_min, - &output_activation_max); + CalculateActivationRange(params->activation, &output_activation_min, + &output_activation_max); KernelType effective_kernel_type; if (((kernel_type == kMultithreadOptimized) || (kernel_type == kCblasOptimized)) && @@ -428,6 +418,7 @@ void EvalFloat(TfLiteContext* context, TfLiteNode* node, filter_data = GetTensorData(filter); } multithreaded_ops::Conv( + *eigen_support::GetThreadPoolDevice(context), GetTensorData(input), GetTensorDims(input), filter_data, GetTensorDims(filter), GetTensorData(bias), GetTensorDims(bias), params->stride_width, params->stride_height, diff --git a/tensorflow/contrib/lite/kernels/depthwise_conv.cc b/tensorflow/contrib/lite/kernels/depthwise_conv.cc index a308de055f49eddba99d02e264fad11409a799f4..16e5f1d065d8ea6d187c5e368d6c9385fe62514b 100644 --- a/tensorflow/contrib/lite/kernels/depthwise_conv.cc +++ b/tensorflow/contrib/lite/kernels/depthwise_conv.cc @@ -173,8 +173,8 @@ void EvalFloat(TfLiteContext* context, TfLiteNode* node, const TfLiteTensor* input, const TfLiteTensor* filter, const TfLiteTensor* bias, TfLiteTensor* output) { float output_activation_min, output_activation_max; - CalculateActivationRangeFloat(params->activation, &output_activation_min, - &output_activation_max); + CalculateActivationRange(params->activation, &output_activation_min, + &output_activation_max); void (*depthwise_conv)(const float*, const Dims<4>&, const float*, const Dims<4>&, const float*, const Dims<4>&, int, int, diff --git a/tensorflow/contrib/lite/kernels/div.cc b/tensorflow/contrib/lite/kernels/div.cc index d264821e30cf622ff5d3d8ad513add46caa9e7ae..bc5c3783fd63451fd6d600df2d8e93f740c68e95 100644 --- a/tensorflow/contrib/lite/kernels/div.cc +++ b/tensorflow/contrib/lite/kernels/div.cc @@ -83,8 +83,8 @@ void EvalFloat(TfLiteContext* context, TfLiteNode* node, const TfLiteTensor* input1, const TfLiteTensor* input2, TfLiteTensor* output) { float output_activation_min, output_activation_max; - CalculateActivationRangeFloat(params->activation, &output_activation_min, - &output_activation_max); + CalculateActivationRange(params->activation, &output_activation_min, + &output_activation_max); #define TF_LITE_DIV(type, opname) \ type::opname(GetTensorData(input1), GetTensorDims(input1), \ GetTensorData(input2), GetTensorDims(input2), \ diff --git a/tensorflow/contrib/lite/kernels/eigen_support.cc b/tensorflow/contrib/lite/kernels/eigen_support.cc index f1fdb42624073717fb70423ff70dfad08e578ca6..4f0d020793eb4c62dfd1e02af883e10adbfab436 100644 --- a/tensorflow/contrib/lite/kernels/eigen_support.cc +++ b/tensorflow/contrib/lite/kernels/eigen_support.cc @@ -14,31 +14,89 @@ limitations under the License. ==============================================================================*/ #include "tensorflow/contrib/lite/kernels/eigen_support.h" -#include "third_party/eigen3/Eigen/Core" +#include + +#include "tensorflow/contrib/lite/kernels/internal/optimized/eigen_spatial_convolutions.h" #include "tensorflow/contrib/lite/kernels/op_macros.h" namespace tflite { namespace eigen_support { +namespace { + +// We have a single global threadpool for all convolution operations. This means +// that inferences started from different threads may block each other, but +// since the underlying resource of CPU cores should be consumed by the +// operations anyway, it shouldn't affect overall performance. +class EigenThreadPoolWrapper : public Eigen::ThreadPoolInterface { + public: + // Takes ownership of 'pool' + explicit EigenThreadPoolWrapper(Eigen::ThreadPool* pool) : pool_(pool) {} + ~EigenThreadPoolWrapper() override {} + + void Schedule(std::function fn) override { + pool_->Schedule(std::move(fn)); + } + int NumThreads() const override { return pool_->NumThreads(); } + int CurrentThreadId() const override { return pool_->CurrentThreadId(); } + + private: + std::unique_ptr pool_; +}; -struct RefCountedEigenContext { +struct RefCountedEigenContext : public TfLiteExternalContext { + std::unique_ptr thread_pool_wrapper; + std::unique_ptr device; int num_references = 0; }; +RefCountedEigenContext* GetEigenContext(TfLiteContext* context) { + return reinterpret_cast( + context->GetExternalContext(context, kTfLiteEigenContext)); +} + +void InitDevice(TfLiteContext* context, RefCountedEigenContext* ptr) { + int num_threads = 4; + if (context->recommended_num_threads != -1) { + num_threads = context->recommended_num_threads; + } + ptr->device.reset(); // destroy before we invalidate the thread pool + ptr->thread_pool_wrapper.reset( + new EigenThreadPoolWrapper(new Eigen::ThreadPool(num_threads))); + ptr->device.reset( + new Eigen::ThreadPoolDevice(ptr->thread_pool_wrapper.get(), num_threads)); +} + +TfLiteStatus Refresh(TfLiteContext* context) { + Eigen::setNbThreads(context->recommended_num_threads); + + auto* ptr = GetEigenContext(context); + if (ptr != nullptr) { + InitDevice(context, ptr); + } + + return kTfLiteOk; +} + +} // namespace + void IncrementUsageCounter(TfLiteContext* context) { - auto* ptr = reinterpret_cast(context->eigen_context); + auto* ptr = GetEigenContext(context); if (ptr == nullptr) { if (context->recommended_num_threads != -1) { Eigen::setNbThreads(context->recommended_num_threads); } ptr = new RefCountedEigenContext; + ptr->type = kTfLiteEigenContext; + ptr->Refresh = Refresh; ptr->num_references = 0; - context->eigen_context = ptr; + InitDevice(context, ptr); + context->SetExternalContext(context, kTfLiteEigenContext, ptr); } ptr->num_references++; } void DecrementUsageCounter(TfLiteContext* context) { - auto* ptr = reinterpret_cast(context->eigen_context); + auto* ptr = GetEigenContext(context); if (ptr == nullptr) { TF_LITE_FATAL( "Call to DecrementUsageCounter() not preceded by " @@ -46,14 +104,17 @@ void DecrementUsageCounter(TfLiteContext* context) { } if (--ptr->num_references == 0) { delete ptr; - context->eigen_context = nullptr; + context->SetExternalContext(context, kTfLiteEigenContext, nullptr); } } -void SetNumThreads(TfLiteContext* context, int num_threads) { - IncrementUsageCounter(context); - Eigen::setNbThreads(num_threads); - DecrementUsageCounter(context); +const Eigen::ThreadPoolDevice* GetThreadPoolDevice(TfLiteContext* context) { + auto* ptr = GetEigenContext(context); + if (ptr == nullptr) { + TF_LITE_FATAL( + "Call to GetFromContext() not preceded by IncrementUsageCounter()"); + } + return ptr->device.get(); } } // namespace eigen_support diff --git a/tensorflow/contrib/lite/kernels/eigen_support.h b/tensorflow/contrib/lite/kernels/eigen_support.h index aa8c351fd8e8dae45f7d4807ce24d80bb393c41c..ec77856b1054e85c405193c6f44dc6e74b58a645 100644 --- a/tensorflow/contrib/lite/kernels/eigen_support.h +++ b/tensorflow/contrib/lite/kernels/eigen_support.h @@ -17,6 +17,10 @@ limitations under the License. #include "tensorflow/contrib/lite/context.h" +namespace EigenForTFLite { +class ThreadPoolDevice; +} + namespace tflite { namespace eigen_support { @@ -28,8 +32,8 @@ void IncrementUsageCounter(TfLiteContext* context); // usages all temporary Eigen objects will be deleted. void DecrementUsageCounter(TfLiteContext* context); -// Set the number of threads that can be used by Eigen. -void SetNumThreads(TfLiteContext* context, int num_threads); +const EigenForTFLite::ThreadPoolDevice* GetThreadPoolDevice( + TfLiteContext* context); } // namespace eigen_support } // namespace tflite diff --git a/tensorflow/contrib/lite/kernels/embedding_lookup.cc b/tensorflow/contrib/lite/kernels/embedding_lookup.cc index 9410bead5e7a68363d034c22fb2c0eff9f060ef1..f550339d03c4c774ae5bc90fd81079d12efe69ae 100644 --- a/tensorflow/contrib/lite/kernels/embedding_lookup.cc +++ b/tensorflow/contrib/lite/kernels/embedding_lookup.cc @@ -94,7 +94,7 @@ TfLiteStatus EvalHybrid(TfLiteContext* context, TfLiteNode* node, const TfLiteTensor* lookup, const TfLiteTensor* value, TfLiteTensor* output) { const int row_size = SizeOfDimension(value, 0); - const double scaling_factor = 1.0 / value->params.scale; + const double scaling_factor = value->params.scale; // col_size after we flatten tensor into 2D. int col_size = 1; @@ -112,8 +112,9 @@ TfLiteStatus EvalHybrid(TfLiteContext* context, TfLiteNode* node, // TODO(alanchiao): refactor scalar multiply into separate function // for ease of adding a neon equivalent if ever necessary. for (int j = 0; j < col_size; j++) { + const int8_t* value_ptr = reinterpret_cast(value->data.uint8); output->data.f[j + i * col_size] = - value->data.uint8[j + idx * col_size] * scaling_factor; + value_ptr[j + idx * col_size] * scaling_factor; } } } diff --git a/tensorflow/contrib/lite/kernels/embedding_lookup_test.cc b/tensorflow/contrib/lite/kernels/embedding_lookup_test.cc index 04657fd86323ef1c58d069c06097c7665f55cc87..4a88d168c60203f10802e634def9b1d1316c9c6d 100644 --- a/tensorflow/contrib/lite/kernels/embedding_lookup_test.cc +++ b/tensorflow/contrib/lite/kernels/embedding_lookup_test.cc @@ -107,9 +107,9 @@ TEST(HybridEmbeddingLookupHybridOpTest, Simple2DTest) { HybridEmbeddingLookupOpModel m({3}, {3, 8}); m.SetInput({1, 0, 2}); m.SetWeight({ - 0.00, 0.01, 0.02, 0.03, 0.10, 0.11, 0.12, 0.13, // Row 0 - 1.00, 1.01, 1.02, 1.03, 1.10, 1.11, 1.12, 1.13, // Row 1 - 2.00, 2.01, 2.02, 2.03, 2.10, 2.11, 2.12, 2.13, // Row 2 + 0.00, 0.01, 0.02, 0.03, 0.10, 0.11, 0.12, 0.13, // Row 0 + 1.00, -1.01, 1.02, 1.03, 1.10, 1.11, 1.12, 1.13, // Row 1 + 2.00, 2.01, 2.02, 2.03, 2.10, 2.11, 2.12, 2.13, // Row 2 }); m.Invoke(); @@ -117,9 +117,9 @@ TEST(HybridEmbeddingLookupHybridOpTest, Simple2DTest) { EXPECT_THAT(m.GetOutput(), ElementsAreArray(ArrayFloatNear( { - 1.00, 1.01, 1.02, 1.03, 1.10, 1.11, 1.12, 1.13, // Row 1 - 0.00, 0.01, 0.02, 0.03, 0.10, 0.11, 0.12, 0.13, // Row 0 - 2.00, 2.01, 2.02, 2.03, 2.10, 2.11, 2.12, 2.13, // Row 2 + 1.00, -1.01, 1.02, 1.03, 1.10, 1.11, 1.12, 1.13, // Row 1 + 0.00, 0.01, 0.02, 0.03, 0.10, 0.11, 0.12, 0.13, // Row 0 + 2.00, 2.01, 2.02, 2.03, 2.10, 2.11, 2.12, 2.13, // Row 2 }, 7.41e-03))); } @@ -128,9 +128,9 @@ TEST(HybridEmbeddingLookupHybridOpTest, Simple3DTest) { HybridEmbeddingLookupOpModel m({3}, {3, 2, 4}); m.SetInput({1, 0, 2}); m.SetWeight({ - 0.00, 0.01, 0.02, 0.03, 0.10, 0.11, 0.12, 0.13, // Row 0 - 1.00, 1.01, 1.02, 1.03, 1.10, 1.11, 1.12, 1.13, // Row 1 - 2.00, 2.01, 2.02, 2.03, 2.10, 2.11, 2.12, 2.13, // Row 2 + 0.00, 0.01, 0.02, 0.03, 0.10, 0.11, 0.12, 0.13, // Row 0 + 1.00, -1.01, 1.02, 1.03, 1.10, 1.11, 1.12, 1.13, // Row 1 + 2.00, 2.01, 2.02, 2.03, 2.10, 2.11, 2.12, 2.13, // Row 2 }); m.Invoke(); @@ -138,9 +138,9 @@ TEST(HybridEmbeddingLookupHybridOpTest, Simple3DTest) { EXPECT_THAT(m.GetOutput(), ElementsAreArray(ArrayFloatNear( { - 1.00, 1.01, 1.02, 1.03, 1.10, 1.11, 1.12, 1.13, // Row 1 - 0.00, 0.01, 0.02, 0.03, 0.10, 0.11, 0.12, 0.13, // Row 0 - 2.00, 2.01, 2.02, 2.03, 2.10, 2.11, 2.12, 2.13, // Row 2 + 1.00, -1.01, 1.02, 1.03, 1.10, 1.11, 1.12, 1.13, // Row 1 + 0.00, 0.01, 0.02, 0.03, 0.10, 0.11, 0.12, 0.13, // Row 0 + 2.00, 2.01, 2.02, 2.03, 2.10, 2.11, 2.12, 2.13, // Row 2 }, 7.41e-03))); } @@ -149,9 +149,9 @@ TEST(HybridEmbeddingLookupHybridOpTest, Simple4DTest) { HybridEmbeddingLookupOpModel m({3}, {3, 2, 2, 2}); m.SetInput({1, 0, 2}); m.SetWeight({ - 0.00, 0.01, 0.02, 0.03, 0.10, 0.11, 0.12, 0.13, // Row 0 - 1.00, 1.01, 1.02, 1.03, 1.10, 1.11, 1.12, 1.13, // Row 1 - 2.00, 2.01, 2.02, 2.03, 2.10, 2.11, 2.12, 2.13, // Row 2 + 0.00, 0.01, 0.02, 0.03, 0.10, 0.11, 0.12, 0.13, // Row 0 + 1.00, -1.01, 1.02, 1.03, 1.10, 1.11, 1.12, 1.13, // Row 1 + 2.00, 2.01, 2.02, 2.03, 2.10, 2.11, 2.12, 2.13, // Row 2 }); m.Invoke(); @@ -159,9 +159,9 @@ TEST(HybridEmbeddingLookupHybridOpTest, Simple4DTest) { EXPECT_THAT(m.GetOutput(), ElementsAreArray(ArrayFloatNear( { - 1.00, 1.01, 1.02, 1.03, 1.10, 1.11, 1.12, 1.13, // Row 1 - 0.00, 0.01, 0.02, 0.03, 0.10, 0.11, 0.12, 0.13, // Row 0 - 2.00, 2.01, 2.02, 2.03, 2.10, 2.11, 2.12, 2.13, // Row 2 + 1.00, -1.01, 1.02, 1.03, 1.10, 1.11, 1.12, 1.13, // Row 1 + 0.00, 0.01, 0.02, 0.03, 0.10, 0.11, 0.12, 0.13, // Row 0 + 2.00, 2.01, 2.02, 2.03, 2.10, 2.11, 2.12, 2.13, // Row 2 }, 7.41e-03))); } diff --git a/tensorflow/contrib/lite/kernels/fake_quant.cc b/tensorflow/contrib/lite/kernels/fake_quant.cc new file mode 100644 index 0000000000000000000000000000000000000000..0ef1a50b308b2e8a781bc9ed7195c22e627ea2de --- /dev/null +++ b/tensorflow/contrib/lite/kernels/fake_quant.cc @@ -0,0 +1,92 @@ +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#include +#include +#include "tensorflow/contrib/lite/builtin_op_data.h" +#include "tensorflow/contrib/lite/context.h" +#include "tensorflow/contrib/lite/kernels/internal/reference/reference_ops.h" +#include "tensorflow/contrib/lite/kernels/internal/tensor.h" +#include "tensorflow/contrib/lite/kernels/kernel_util.h" +#include "tensorflow/contrib/lite/kernels/op_macros.h" + +namespace tflite { +namespace ops { +namespace builtin { +namespace fake_quant { + +// This file has reference implementation of FakeQuant. +enum KernelType { + kReference, +}; + +struct OpContext { + OpContext(TfLiteContext* context, TfLiteNode* node) { + input = GetInput(context, node, 0); + output = GetOutput(context, node, 0); + } + const TfLiteTensor* input; + TfLiteTensor* output; +}; + +TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) { + TF_LITE_ENSURE_EQ(context, NumInputs(node), 1); + TF_LITE_ENSURE_EQ(context, NumOutputs(node), 1); + + const auto* params = + reinterpret_cast(node->builtin_data); + + if (params->narrow_range) { + context->ReportError( + context, + "narrow_range FakeQuant is not currently supported at runtime. " + "narrow_range is only meant to be applied to weights, not activations"); + return kTfLiteError; + } + + OpContext op_context(context, node); + TfLiteIntArray* output_dims = TfLiteIntArrayCopy(op_context.input->dims); + op_context.output->type = op_context.input->type; + return context->ResizeTensor(context, op_context.output, output_dims); +} + +template +TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { + OpContext op_context(context, node); + + const auto* params = + reinterpret_cast(node->builtin_data); + + reference_ops::FakeQuant(GetTensorData(op_context.input), + GetTensorDims(op_context.input), params->min, + params->max, params->num_bits, + GetTensorData(op_context.output), + GetTensorDims(op_context.output)); + + return kTfLiteOk; +} + +} // namespace fake_quant + +TfLiteRegistration* Register_FAKE_QUANT_REF() { + static TfLiteRegistration r = {nullptr, nullptr, fake_quant::Prepare, + fake_quant::Eval}; + return &r; +} + +TfLiteRegistration* Register_FAKE_QUANT() { return Register_FAKE_QUANT_REF(); } + +} // namespace builtin +} // namespace ops +} // namespace tflite diff --git a/tensorflow/contrib/lite/kernels/fake_quant_test.cc b/tensorflow/contrib/lite/kernels/fake_quant_test.cc new file mode 100644 index 0000000000000000000000000000000000000000..11a02f7ed7474e05b887955c111179d2d403f0e6 --- /dev/null +++ b/tensorflow/contrib/lite/kernels/fake_quant_test.cc @@ -0,0 +1,112 @@ +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#include +#include "tensorflow/contrib/lite/interpreter.h" +#include "tensorflow/contrib/lite/kernels/register.h" +#include "tensorflow/contrib/lite/kernels/test_util.h" +#include "tensorflow/contrib/lite/model.h" + +namespace tflite { +namespace { + +using ::testing::ElementsAreArray; + +class FakeQuantOpModel : public SingleOpModel { + public: + FakeQuantOpModel(const TensorData& input, const TensorType& output, float min, + float max, int num_bits) { + input_ = AddInput(input); + output_ = AddOutput(output); + SetBuiltinOp(BuiltinOperator_FAKE_QUANT, BuiltinOptions_FakeQuantOptions, + CreateFakeQuantOptions(builder_, min, max, num_bits).Union()); + BuildInterpreter({GetShape(input_)}); + } + + template + void SetInput(std::initializer_list data) { + PopulateTensor(input_, data); + } + + template + std::vector GetOutput() { + return ExtractVector(output_); + } + std::vector GetOutputShape() { return GetTensorShape(output_); } + + protected: + int input_; + int output_; +}; + +TEST(FakeQuantOpTest, FloatPositiveRange8Test) { + std::initializer_list data = {0.0, 1.0, 0.25, + 0.50, 0.4444444, 0.00001}; + FakeQuantOpModel m({TensorType_FLOAT32, {3, 1, 2}}, TensorType_FLOAT32, 0.0f, + 1.0f, 8); + m.SetInput(data); + m.Invoke(); + EXPECT_THAT(m.GetOutputShape(), ElementsAreArray({3, 1, 2})); + EXPECT_THAT( + m.GetOutput(), + ElementsAreArray(ArrayFloatNear({0, 1, 0.25098, 0.498039, 0.443137, 0}))); +} + +TEST(FakeQuantOpTest, FloatNegativeRange8Test) { + std::initializer_list data = {0.0, -0.9, 0.25, + 0.50, 0.4444444, -0.00001}; + FakeQuantOpModel m({TensorType_FLOAT32, {3, 1, 2}}, TensorType_FLOAT32, -0.9f, + 0.9f, 8); + m.SetInput(data); + m.Invoke(); + EXPECT_THAT(m.GetOutputShape(), ElementsAreArray({3, 1, 2})); + EXPECT_THAT(m.GetOutput(), + ElementsAreArray(ArrayFloatNear( + {0, -0.896471, 0.247059, 0.501176, 0.444706, 0}))); +} + +TEST(FakeQuantOpTest, FloatPositiveRange16Test) { + std::initializer_list data = {0.0, 1.0, 0.25, + 0.50, 0.4444444, 0.00001}; + FakeQuantOpModel m({TensorType_FLOAT32, {3, 1, 2}}, TensorType_FLOAT32, 0.0f, + 1.0f, 16); + m.SetInput(data); + m.Invoke(); + EXPECT_THAT(m.GetOutputShape(), ElementsAreArray({3, 1, 2})); + EXPECT_THAT(m.GetOutput(), + ElementsAreArray(ArrayFloatNear( + {0, 1, 0.250004, 0.500008, 0.44445, 1.5259e-05}))); +} + +TEST(FakeQuantOpTest, FloatNegativeRange16Test) { + std::initializer_list data = {0.0, -0.9, 0.25, + 0.50, 0.4444444, -0.00001}; + FakeQuantOpModel m({TensorType_FLOAT32, {3, 1, 2}}, TensorType_FLOAT32, -0.9f, + 0.9f, 16); + m.SetInput(data); + m.Invoke(); + EXPECT_THAT(m.GetOutputShape(), ElementsAreArray({3, 1, 2})); + EXPECT_THAT(m.GetOutput(), + ElementsAreArray(ArrayFloatNear( + {0, -0.900014, 0.249998, 0.499995, 0.444431, 0}))); +} + +} // namespace +} // namespace tflite + +int main(int argc, char** argv) { + ::tflite::LogToStderr(); + ::testing::InitGoogleTest(&argc, argv); + return RUN_ALL_TESTS(); +} diff --git a/tensorflow/contrib/lite/kernels/fully_connected.cc b/tensorflow/contrib/lite/kernels/fully_connected.cc index f6fc0f5b6ad12d58c541efc6eae566ab4b8327f4..3b203dd480f95c5dc70a69aafce0bac6ab2cbc06 100644 --- a/tensorflow/contrib/lite/kernels/fully_connected.cc +++ b/tensorflow/contrib/lite/kernels/fully_connected.cc @@ -63,6 +63,7 @@ constexpr int kInputTensor = 0; constexpr int kWeightsTensor = 1; constexpr int kBiasTensor = 2; constexpr int kOutputTensor = 0; +constexpr int kShuffledInputWorkspaceTensor = 1; constexpr int kScratchBufferTensor = 1; void* Init(TfLiteContext* context, const char* buffer, size_t length) { @@ -87,7 +88,11 @@ TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) { // Check we have all the inputs and outputs we need. TF_LITE_ENSURE_EQ(context, node->inputs->size, 3); - TF_LITE_ENSURE_EQ(context, node->outputs->size, 1); + // Shuffled formats need a workspace to store the shuffled input activations. + const int expected_outputs_count = + params->weights_format == kTfLiteFullyConnectedWeightsFormatDefault ? 1 + : 2; + TF_LITE_ENSURE_EQ(context, node->outputs->size, expected_outputs_count); const TfLiteTensor* input = GetInput(context, node, kInputTensor); const TfLiteTensor* filter = GetInput(context, node, kWeightsTensor); @@ -121,9 +126,9 @@ TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) { QuantizeMultiplierSmallerThanOneExp( real_multiplier, &data->output_multiplier, &data->output_shift); data->output_shift *= -1; - CalculateActivationRangeUint8(params->activation, output, - &data->output_activation_min, - &data->output_activation_max); + TF_LITE_ENSURE_STATUS(CalculateActivationRangeQuantized( + context, params->activation, output, &data->output_activation_min, + &data->output_activation_max)); } // If we have to perform on-the-fly quantization (with quantized weights and @@ -278,44 +283,101 @@ TfLiteStatus EvalQuantized(TfLiteContext* context, TfLiteNode* node, int32_t input_offset = -input->params.zero_point; int32_t filter_offset = -filter->params.zero_point; int32_t output_offset = output->params.zero_point; -#define TF_LITE_FULLY_CONNECTED(type) \ +#define TF_LITE_FULLY_CONNECTED(type, output_data_type) \ type::FullyConnected( \ GetTensorData(input), GetTensorDims(input), input_offset, \ GetTensorData(filter), GetTensorDims(filter), filter_offset, \ GetTensorData(bias), GetTensorDims(bias), output_offset, \ data->output_multiplier, data->output_shift, \ data->output_activation_min, data->output_activation_max, \ - GetTensorData(output), GetTensorDims(output), gemm_context) + GetTensorData(output), GetTensorDims(output), \ + gemm_context) if (kernel_type == kReference) { - TF_LITE_FULLY_CONNECTED(reference_ops); - } else if (kernel_type == kPie) { - if (input->type == kTfLiteFloat32) { - // Pie currently only supports quantized models and float inputs/outputs. - TfLiteTensor* input_quantized = - &context->tensors[node->temporaries->data[0]]; - return EvalPieQuantized(context, node, params, data, input, filter, bias, - input_quantized, output); - } else { - // TODO(ahentz): we don't have a quantized version of the PIE kernels, so - // we just defer to the MINI ones. - TF_LITE_FULLY_CONNECTED(optimized_ops); + switch (output->type) { + case kTfLiteUInt8: + TF_LITE_FULLY_CONNECTED(reference_ops, uint8_t); + break; + case kTfLiteInt16: + TF_LITE_FULLY_CONNECTED(reference_ops, int16_t); + break; + default: + context->ReportError( + context, + "Quantized FullyConnected expects output data type uint8 or int16"); + return kTfLiteError; } + } else if (kernel_type == kPie && input->type == kTfLiteFloat32) { + // Pie currently only supports quantized models and float inputs/outputs. + TfLiteTensor* input_quantized = + &context->tensors[node->temporaries->data[0]]; + return EvalPieQuantized(context, node, params, data, input, filter, bias, + input_quantized, output); } else { - TF_LITE_FULLY_CONNECTED(optimized_ops); + switch (output->type) { + case kTfLiteUInt8: + TF_LITE_FULLY_CONNECTED(optimized_ops, uint8_t); + break; + case kTfLiteInt16: + TF_LITE_FULLY_CONNECTED(optimized_ops, int16_t); + break; + default: + context->ReportError( + context, + "Quantized FullyConnected expects output data type uint8 or int16"); + return kTfLiteError; + } } #undef TF_LITE_FULLY_CONNECTED return kTfLiteOk; } +template +TfLiteStatus EvalShuffledQuantized(TfLiteContext* context, TfLiteNode* node, + TfLiteFullyConnectedParams* params, + OpData* data, const TfLiteTensor* input, + const TfLiteTensor* filter, + const TfLiteTensor* bias, + TfLiteTensor* output, + TfLiteTensor* shuffled_input_workspace) { + gemmlowp::GemmContext* gemm_context = gemm_support::GetFromContext(context); + + // TODO(b/110697972) decide more consistently if / how / where we want + // to perform this kind of runtime data type checks. + if (input->type != kTfLiteUInt8 || filter->type != kTfLiteUInt8 || + bias->type != kTfLiteInt32 || output->type != kTfLiteInt16 || + shuffled_input_workspace->type != kTfLiteUInt8) { + context->ReportError(context, "Unexpected data type"); + return kTfLiteError; + } + +#define TF_LITE_SHUFFLED_FULLY_CONNECTED(type) \ + type::ShuffledFullyConnected( \ + GetTensorData(input), GetTensorDims(input), \ + GetTensorData(filter), GetTensorDims(filter), \ + GetTensorData(bias), GetTensorDims(bias), \ + data->output_multiplier, data->output_shift, \ + data->output_activation_min, data->output_activation_max, \ + GetTensorData(output), GetTensorDims(output), \ + GetTensorData(shuffled_input_workspace), gemm_context) + if (kernel_type == kReference) { + TF_LITE_SHUFFLED_FULLY_CONNECTED(reference_ops); + } else { + TF_LITE_SHUFFLED_FULLY_CONNECTED(optimized_ops); + } +#undef TF_LITE_SHUFFLED_FULLY_CONNECTED + + return kTfLiteOk; +} + template TfLiteStatus EvalFloat(TfLiteContext* context, TfLiteNode* node, TfLiteFullyConnectedParams* params, OpData* data, const TfLiteTensor* input, const TfLiteTensor* filter, const TfLiteTensor* bias, TfLiteTensor* output) { float output_activation_min, output_activation_max; - CalculateActivationRangeFloat(params->activation, &output_activation_min, - &output_activation_max); + CalculateActivationRange(params->activation, &output_activation_min, + &output_activation_max); #define TF_LITE_FULLY_CONNECTED(type) \ type::FullyConnected(GetTensorData(input), GetTensorDims(input), \ GetTensorData(filter), GetTensorDims(filter), \ @@ -352,8 +414,22 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { return EvalFloat(context, node, params, data, input, filter, bias, output); case kTfLiteUInt8: - return EvalQuantized(context, node, params, data, input, - filter, bias, output); + if (params->weights_format == + kTfLiteFullyConnectedWeightsFormatShuffled4x16Int8) { + TfLiteTensor* shuffled_input_workspace = + GetOutput(context, node, kShuffledInputWorkspaceTensor); + return EvalShuffledQuantized(context, node, params, data, + input, filter, bias, output, + shuffled_input_workspace); + } else if (params->weights_format == + kTfLiteFullyConnectedWeightsFormatDefault) { + return EvalQuantized(context, node, params, data, input, + filter, bias, output); + } else { + context->ReportError(context, + "Unhandled fully-connected weights format"); + return kTfLiteError; + } default: context->ReportError(context, "Type %d not currently supported.", filter->type); diff --git a/tensorflow/contrib/lite/kernels/fully_connected_test.cc b/tensorflow/contrib/lite/kernels/fully_connected_test.cc index 05dd028b484c09bdf90a09fab1238f48e8a9ddab..ec949056971ccb5f7a6f93fa9f236a93625ca6ad 100644 --- a/tensorflow/contrib/lite/kernels/fully_connected_test.cc +++ b/tensorflow/contrib/lite/kernels/fully_connected_test.cc @@ -15,6 +15,7 @@ limitations under the License. // Unit test for TFLite FULLY_CONNECTED op. #include +#include #include #include @@ -133,9 +134,12 @@ static float fully_connected_golden_output[] = { class BaseFullyConnectedOpModel : public SingleOpModel { public: // TODO(ahentz): test different activation types too. - BaseFullyConnectedOpModel(TfLiteRegistration* registration, int units, - int batches, const TensorData& input, - const TensorData& output = {TensorType_FLOAT32}) + BaseFullyConnectedOpModel( + TfLiteRegistration* registration, int units, int batches, + const TensorData& input, const TensorData& output = {TensorType_FLOAT32}, + ActivationFunctionType activation_func = ActivationFunctionType_RELU, + FullyConnectedOptionsWeightsFormat weights_format = + FullyConnectedOptionsWeightsFormat_DEFAULT) : batches_(batches), units_(units) { int total_input_size = 1; for (int i = 0; i < input.shape.size(); ++i) { @@ -159,10 +163,13 @@ class BaseFullyConnectedOpModel : public SingleOpModel { } output_ = AddOutput(output); + if (weights_format != FullyConnectedOptionsWeightsFormat_DEFAULT) { + AddOutput({TensorType_UINT8, input.shape}); + } SetBuiltinOp( BuiltinOperator_FULLY_CONNECTED, BuiltinOptions_FullyConnectedOptions, - CreateFullyConnectedOptions(builder_, ActivationFunctionType_RELU) + CreateFullyConnectedOptions(builder_, activation_func, weights_format) .Union()); resolver_ = absl::make_unique( BuiltinOperator_FULLY_CONNECTED, registration); @@ -188,13 +195,11 @@ class FloatFullyConnectedOpModel : public BaseFullyConnectedOpModel { public: using BaseFullyConnectedOpModel::BaseFullyConnectedOpModel; - void SetBias(std::initializer_list f) { PopulateTensor(bias_, f); } + void SetBias(const std::vector& f) { PopulateTensor(bias_, f); } - void SetWeights(std::initializer_list f) { - PopulateTensor(weights_, f); - } + void SetWeights(const std::vector& f) { PopulateTensor(weights_, f); } - void SetInput(std::initializer_list data) { + void SetInput(const std::vector& data) { PopulateTensor(input_, data); } void SetInput(int offset, float* begin, float* end) { @@ -208,20 +213,50 @@ class QuantizedFullyConnectedOpModel : public BaseFullyConnectedOpModel { public: using BaseFullyConnectedOpModel::BaseFullyConnectedOpModel; - void SetBias(std::initializer_list data) { + void SetBias(const std::vector& data) { QuantizeAndPopulate(bias_, data); } - void SetWeights(std::initializer_list data) { + void SetWeights(const std::vector& data) { QuantizeAndPopulate(weights_, data); } - void SetInput(std::initializer_list data) { + void ShuffleAndSetWeights(const std::vector& data, int input_depth, + int output_depth) { + std::vector shuffled_data(data.size()); + CHECK_EQ(input_depth % 16, 0); + CHECK_EQ(output_depth % 4, 0); + float* shuffled_data_ptr = shuffled_data.data(); + for (int block_o = 0; block_o < output_depth; block_o += 4) { + for (int block_i = 0; block_i < input_depth; block_i += 16) { + for (int o = 0; o < 4; o++) { + for (int i = 0; i < 16; i++) { + *shuffled_data_ptr++ = + data[(block_o + o) * input_depth + block_i + i]; + } + } + } + } + TfLiteTensor* t = interpreter_->tensor(weights_); + auto quantized_data = + Quantize(shuffled_data, t->params.scale, t->params.zero_point); + for (uint8_t& q : quantized_data) { + q ^= 0x80; + } + PopulateTensor(weights_, 0, quantized_data.data(), + quantized_data.data() + quantized_data.size()); + } + void SetInput(const std::vector& data) { QuantizeAndPopulate(input_, data); } - std::vector GetOutput() { return ExtractVector(output_); } + template + std::vector GetOutput() { + return ExtractVector(output_); + } + + template std::vector GetDequantizedOutput() { - return Dequantize(ExtractVector(output_), - GetScale(output_), GetZeroPoint(output_)); + return Dequantize(ExtractVector(output_), GetScale(output_), + GetZeroPoint(output_)); } }; @@ -256,12 +291,12 @@ class HybridFullyConnectedOpModel : public SingleOpModel { ops::builtin::Register_FULLY_CONNECTED_PIE()); BuildInterpreter({GetShape(input_), GetShape(weights_), GetShape(bias_)}); } - void SetBias(std::initializer_list f) { PopulateTensor(bias_, f); } - void SetWeights(std::initializer_list data) { + void SetBias(const std::vector& f) { PopulateTensor(bias_, f); } + void SetWeights(const std::vector& data) { SymmetricQuantizeAndPopulate(weights_, data); } - void SetInput(std::initializer_list f) { PopulateTensor(input_, f); } + void SetInput(const std::vector& f) { PopulateTensor(input_, f); } std::vector GetOutput() { return ExtractVector(output_); } int input_size() { return input_size_; } @@ -340,6 +375,24 @@ TEST_P(FloatFullyConnectedOpTest, SimpleTest) { EXPECT_THAT(m.GetOutput(), ElementsAre(24, 25, 26, 58, 59, 60)); } +TEST_P(FloatFullyConnectedOpTest, SimpleTest2) { + FloatFullyConnectedOpModel m(GetRegistration(), /*units=*/1, /*batches=*/2, + /*input=*/{TensorType_FLOAT32, {2, 2}}); + m.SetWeights({ + 2, 4, // u = 0 + }); + m.SetBias({1}); + + m.SetInput({ + 1, 2, // b = 0 + 2, 1, // b = 1 + }); + + m.Invoke(); + + EXPECT_THAT(m.GetOutput(), ElementsAre(11, 9)); +} + TEST_P(QuantizedFullyConnectedOpTest, SimpleTestQuantized) { QuantizedFullyConnectedOpModel m( GetRegistration(), /*units=*/3, /*batches*/ 2, @@ -350,7 +403,7 @@ TEST_P(QuantizedFullyConnectedOpTest, SimpleTestQuantized) { m.SetWeights({ 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, // u = 0 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, // u = 1 - 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, // u = 1 + 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, // u = 2 }); m.SetBias({1, 2, 3}); @@ -361,11 +414,136 @@ TEST_P(QuantizedFullyConnectedOpTest, SimpleTestQuantized) { m.Invoke(); - EXPECT_THAT(m.GetDequantizedOutput(), ElementsAreArray(ArrayFloatNear({ - 24, 25, 26, // - 58, 59, 60, // - }))); - EXPECT_THAT(m.GetOutput(), ElementsAre(151, 152, 153, 185, 186, 187)); + EXPECT_THAT(m.GetDequantizedOutput(), + ElementsAreArray(ArrayFloatNear({ + 24, 25, 26, // + 58, 59, 60, // + }))); + EXPECT_THAT(m.GetOutput(), + ElementsAre(151, 152, 153, 185, 186, 187)); +} + +void SimpleTestQuantizedInt16OutputCase( + TfLiteRegistration* registration, int input_depth, int output_depth, + int batches, FullyConnectedOptionsWeightsFormat weights_format) { + const uint8_t kWeightsZeroPoint = 128; + const float kWeightsScale = 1.f / 128.f; + const uint8_t kInputZeroPoint = 128; + const float kInputScale = 1.f / 128.f; + const float kInputMin = (0 - kInputZeroPoint) * kInputScale; + const float kInputMax = (255 - kInputZeroPoint) * kInputScale; + // Output ranges in [-8..8] encoded as int16 + const float kOutputScale = 8.f / 32768.f; + const float kOutputMin = -32768 * kOutputScale; + const float kOutputMax = 32767 * kOutputScale; + + QuantizedFullyConnectedOpModel m( + registration, output_depth, batches, + /*input=*/ + {TensorType_UINT8, {batches, input_depth}, kInputMin, kInputMax}, + /*output=*/{TensorType_INT16, {}, kOutputMin, kOutputMax}, + /*activation_func=*/ActivationFunctionType_NONE, weights_format); + + std::mt19937 random_engine; + std::uniform_int_distribution weights_dist; + + std::vector weights_data(input_depth * output_depth); + for (auto& w : weights_data) { + uint8_t q = weights_dist(random_engine); + w = (q - kWeightsZeroPoint) * kWeightsScale; + } + + // Based on weights_format, enforce any shape requirement for that format/path + // and set the (possibly shuffled) weights. + switch (weights_format) { + case FullyConnectedOptionsWeightsFormat_DEFAULT: + m.SetWeights(weights_data); + break; + case FullyConnectedOptionsWeightsFormat_SHUFFLED4x16INT8: + // The shuffled path currently supports only a restrictive subset of + // shapes, described by the following assertions: + CHECK_EQ(input_depth % 16, 0); + CHECK_EQ(output_depth % 4, 0); + CHECK(batches == 1 || batches == 4); + m.ShuffleAndSetWeights(weights_data, input_depth, output_depth); + break; + default: + LOG(FATAL) << "Unhandled weights format"; + } + + std::uniform_int_distribution input_dist; + std::vector input_data(input_depth * batches); + for (auto& i : input_data) { + uint8_t q = input_dist(random_engine); + i = (q - kInputZeroPoint) * kInputScale; + } + + std::vector bias_data(output_depth); + // As the output ranges in [-8, 8], it's reasonable to have bias values + // in [-1, 1], this won't result in too much saturation. + std::uniform_real_distribution bias_dist(-1.f, 1.f); + for (auto& b : bias_data) { + b = bias_dist(random_engine); + } + + m.SetBias(bias_data); + m.SetInput(input_data); + + m.Invoke(); + + std::vector expected_output_data(output_depth * batches); + for (int b = 0; b < batches; b++) { + for (int o = 0; o < output_depth; o++) { + float accum = bias_data[o]; + for (int i = 0; i < input_depth; i++) { + accum += + input_data[b * input_depth + i] * weights_data[o * input_depth + i]; + } + accum = std::min(accum, kOutputMax); + accum = std::max(accum, kOutputMin); + expected_output_data[b * output_depth + o] = accum; + } + } + + EXPECT_THAT(m.GetDequantizedOutput(), + ElementsAreArray(ArrayFloatNear(expected_output_data, 3e-4f))); +} + +TEST_P(QuantizedFullyConnectedOpTest, + SimpleTestQuantizedInt16OutputDefaultWeights) { + for (int input_depth : {1, 3, 10, 100}) { + for (int output_depth : {1, 3, 10, 100}) { + for (int batch : {1, 3, 10, 100}) { + SimpleTestQuantizedInt16OutputCase( + GetRegistration(), input_depth, output_depth, batch, + FullyConnectedOptionsWeightsFormat_DEFAULT); + } + } + } +} + +TEST_P(QuantizedFullyConnectedOpTest, + SimpleTestQuantizedInt16OutputShuffled4x16Int8Weights) { + // The shuffled weights block shape is 4x16. The shape of the weights matrix + // is: rows = output_depth, cols = input_depth. It must be a multiple of 4x16. + // This means that output_depth must be a multiple of 4, and input_deth must + // be a multiple of 16. + for (int input_depth_numblocks : {1, 3}) { + for (int output_depth_numblocks : {1, 3}) { + int input_depth = 16 * input_depth_numblocks; + int output_depth = 4 * output_depth_numblocks; + // The fast shuffled path is currently supporting only batch sizes of 1 + // and 4. The idea is that the whole point of that path is to go as fast + // as possible for small batch size, which requires fully specializing + // it for each batch size, and for larger batch sizes the generic + // gemmlowp-based implementation is fast enough. + for (int batch : {1, 4}) { + SimpleTestQuantizedInt16OutputCase( + GetRegistration(), input_depth, output_depth, batch, + FullyConnectedOptionsWeightsFormat_SHUFFLED4x16INT8); + } + } + } } TEST(HybridFullyConnectedOpTest, SimpleTestQuantized) { @@ -396,11 +574,11 @@ TEST(HybridFullyConnectedOpTest, SimpleTestQuantized) { /*max_abs_error=*/1.3f))); } -TEST(FloatFullyConnectedOpTest, SimpleTest4DInput) { +TEST_P(FloatFullyConnectedOpTest, SimpleTest4DInput) { // Note that it is not required that the first dimension be the number of // batches. All we care is that the input can be evenly distributed in // batches. In this case, we need the input to have multiples of '2'. - FloatFullyConnectedOpModel m(ops::builtin::Register_FULLY_CONNECTED_PIE(), + FloatFullyConnectedOpModel m(GetRegistration(), /*units=*/3, /*batches=*/2, /*input=*/{TensorType_FLOAT32, {4, 1, 5, 1}}); m.SetWeights({ @@ -444,11 +622,13 @@ TEST_P(QuantizedFullyConnectedOpTest, SimpleTest4dInputQuantized) { m.Invoke(); - EXPECT_THAT(m.GetDequantizedOutput(), ElementsAreArray(ArrayFloatNear({ - 24, 25, 26, // - 58, 59, 60, // - }))); - EXPECT_THAT(m.GetOutput(), ElementsAre(151, 152, 153, 185, 186, 187)); + EXPECT_THAT(m.GetDequantizedOutput(), + ElementsAreArray(ArrayFloatNear({ + 24, 25, 26, // + 58, 59, 60, // + }))); + EXPECT_THAT(m.GetOutput(), + ElementsAre(151, 152, 153, 185, 186, 187)); } INSTANTIATE_TEST_CASE_P( diff --git a/tensorflow/contrib/lite/kernels/gather.cc b/tensorflow/contrib/lite/kernels/gather.cc index 6a2341461f2c627c78bd4783ee27579b59b5fde3..2b2a9e662051287fd1e3dbe8978f4689dc731064 100644 --- a/tensorflow/contrib/lite/kernels/gather.cc +++ b/tensorflow/contrib/lite/kernels/gather.cc @@ -40,10 +40,8 @@ TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) { TfLiteTensor* output = GetOutput(context, node, kOutputTensor); // Only INT32 positions are supported. TF_LITE_ENSURE_EQ(context, positions->type, kTfLiteInt32); - // Check that input and output types match. - TF_LITE_ENSURE_EQ(context, input->type, output->type); - // TODO(mgubin): only 0D or 1D positions are currently supported. - TF_LITE_ENSURE(context, NumDimensions(positions) <= 1); + // Assign to output the input type. + output->type = input->type; // TODO(mgubin): Only default axis == 0 is supported. TF_LITE_ENSURE_EQ(context, params->axis, 0); // Check conditions for different types. @@ -102,6 +100,7 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { TF_LITE_GATHER(int32_t, int32_t); break; case kTfLiteString: { + // TODO(mgubin): Currently support only for 1D output tensors. DynamicBuffer buffer; const int32* indexes = positions->data.i32; const int num_strings = GetStringCount(input); diff --git a/tensorflow/contrib/lite/kernels/gather_test.cc b/tensorflow/contrib/lite/kernels/gather_test.cc index cdadbeda1884ba0186846826dd16be6ff69878d9..1d4292955cced59a47e0500833a86113cb9d3eb8 100644 --- a/tensorflow/contrib/lite/kernels/gather_test.cc +++ b/tensorflow/contrib/lite/kernels/gather_test.cc @@ -96,6 +96,15 @@ TEST(GatherOpTest, Test0DIndexWith0DResult) { EXPECT_TRUE(m.GetOutputShape().empty()); } +TEST(GatherOpTest, Test2DIndexWith2DResult) { + GatherOpModel m({3}, TensorType_FLOAT32, {1, 2}); + m.SetInputFloat({1.0, 2.0, 3.0}); + m.SetPositions({1, 0}); + m.Invoke(); + EXPECT_THAT(m.GetOutputFloat(), ElementsAreArray(ArrayFloatNear({2.0, 1.0}))); + EXPECT_THAT(m.GetOutputShape(), ElementsAreArray({1, 2})); +} + TEST(FloatGatherOpTest, Duplicate) { GatherOpModel m({1, 2, 2}, TensorType_FLOAT32, {2}); m.SetInputFloat({-2.0, 0.2, 0.7, 0.8}); diff --git a/tensorflow/contrib/lite/kernels/gemm_support.cc b/tensorflow/contrib/lite/kernels/gemm_support.cc index 95f45ea768be7f9bae9570563f161792afbff436..ed334af2da877edf9f591612478e22f04cf15931 100644 --- a/tensorflow/contrib/lite/kernels/gemm_support.cc +++ b/tensorflow/contrib/lite/kernels/gemm_support.cc @@ -14,57 +14,70 @@ limitations under the License. ==============================================================================*/ #include "tensorflow/contrib/lite/kernels/gemm_support.h" +#include + #include "tensorflow/contrib/lite/kernels/op_macros.h" namespace tflite { namespace gemm_support { +namespace { -struct RefCountedGemmContext { - gemmlowp::GemmContext* gemm_context_ = nullptr; - int num_references_ = 0; +struct RefCountedGemmContext : public TfLiteExternalContext { + std::unique_ptr gemm_context; + int num_references = 0; }; +RefCountedGemmContext* GetGemmLowpContext(TfLiteContext* context) { + return reinterpret_cast( + context->GetExternalContext(context, kTfLiteGemmLowpContext)); +} + +TfLiteStatus Refresh(TfLiteContext* context) { + auto* ptr = GetGemmLowpContext(context); + if (ptr != nullptr) { + ptr->gemm_context->set_max_num_threads(context->recommended_num_threads); + } + return kTfLiteOk; +} + +} // namespace + void IncrementUsageCounter(TfLiteContext* context) { - auto* ptr = reinterpret_cast(context->gemm_context); + auto* ptr = GetGemmLowpContext(context); if (ptr == nullptr) { ptr = new RefCountedGemmContext; - ptr->gemm_context_ = new gemmlowp::GemmContext(); + ptr->type = kTfLiteGemmLowpContext; + ptr->Refresh = Refresh; + ptr->gemm_context.reset(new gemmlowp::GemmContext()); if (context->recommended_num_threads != -1) { - ptr->gemm_context_->set_max_num_threads(context->recommended_num_threads); + ptr->gemm_context->set_max_num_threads(context->recommended_num_threads); } - ptr->num_references_ = 0; - context->gemm_context = ptr; + ptr->num_references = 0; + context->SetExternalContext(context, kTfLiteGemmLowpContext, ptr); } - ptr->num_references_++; + ptr->num_references++; } void DecrementUsageCounter(TfLiteContext* context) { - auto* ptr = reinterpret_cast(context->gemm_context); + auto* ptr = GetGemmLowpContext(context); if (ptr == nullptr) { TF_LITE_FATAL( "Call to DecrementUsageCounter() not preceded by " "IncrementUsageCounter()"); } - if (--ptr->num_references_ == 0) { - delete ptr->gemm_context_; + if (--ptr->num_references == 0) { delete ptr; - context->gemm_context = nullptr; + context->SetExternalContext(context, kTfLiteGemmLowpContext, nullptr); } } gemmlowp::GemmContext* GetFromContext(TfLiteContext* context) { - auto* ptr = reinterpret_cast(context->gemm_context); + auto* ptr = GetGemmLowpContext(context); if (ptr == nullptr) { TF_LITE_FATAL( "Call to GetFromContext() not preceded by IncrementUsageCounter()"); } - return ptr->gemm_context_; -} - -void SetNumThreads(TfLiteContext* context, int num_threads) { - IncrementUsageCounter(context); - GetFromContext(context)->set_max_num_threads(num_threads); - DecrementUsageCounter(context); + return ptr->gemm_context.get(); } } // namespace gemm_support diff --git a/tensorflow/contrib/lite/kernels/gemm_support.h b/tensorflow/contrib/lite/kernels/gemm_support.h index f033501cb6e341aa014fa4d956b531bd79aa555b..37af772c6846f2f8124faabf1a0f0987e2e9393d 100644 --- a/tensorflow/contrib/lite/kernels/gemm_support.h +++ b/tensorflow/contrib/lite/kernels/gemm_support.h @@ -45,9 +45,6 @@ void IncrementUsageCounter(TfLiteContext* context); // 'context'. If there are no more usages the GemmContext will be deleted. void DecrementUsageCounter(TfLiteContext* context); -// Set the number of threads that can be used by gemmlowp. -void SetNumThreads(TfLiteContext* context, int num_threads); - } // namespace gemm_support } // namespace tflite diff --git a/tensorflow/contrib/lite/kernels/internal/kernel_utils.cc b/tensorflow/contrib/lite/kernels/internal/kernel_utils.cc index 36c25388e8bde721d7644dc83d5b7c490d37b4d3..a0e382edb6efe467c7b16624cf1760b0d1c6d760 100644 --- a/tensorflow/contrib/lite/kernels/internal/kernel_utils.cc +++ b/tensorflow/contrib/lite/kernels/internal/kernel_utils.cc @@ -416,7 +416,7 @@ void LstmStep( if (!use_cifg) { if (use_peephole && !is_cell_state_all_zeros) { VectorMultiply(cell_to_input_weights_ptr, n_cell, - 1. / cell_to_input_weights_scale, recovered_cell_weights); + cell_to_input_weights_scale, recovered_cell_weights); tensor_utils::VectorBatchVectorCwiseProductAccumulate( recovered_cell_weights, n_cell, cell_state_ptr, n_batch, input_gate_scratch); @@ -428,7 +428,7 @@ void LstmStep( // For each batch and cell: update forget gate. if (use_peephole && !is_cell_state_all_zeros) { VectorMultiply(cell_to_forget_weights_ptr, n_cell, - 1. / cell_to_forget_weights_scale, recovered_cell_weights); + cell_to_forget_weights_scale, recovered_cell_weights); tensor_utils::VectorBatchVectorCwiseProductAccumulate( recovered_cell_weights, n_cell, cell_state_ptr, n_batch, forget_gate_scratch); @@ -460,7 +460,7 @@ void LstmStep( // For each batch and cell: update the output gate. if (use_peephole && !is_cell_state_all_zeros) { VectorMultiply(cell_to_output_weights_ptr, n_cell, - 1. / cell_to_output_weights_scale, recovered_cell_weights); + cell_to_output_weights_scale, recovered_cell_weights); tensor_utils::VectorBatchVectorCwiseProductAccumulate( recovered_cell_weights, n_cell, cell_state_ptr, n_batch, output_gate_scratch); diff --git a/tensorflow/contrib/lite/kernels/internal/optimized/depthwiseconv_uint8_3x3_filter.h b/tensorflow/contrib/lite/kernels/internal/optimized/depthwiseconv_uint8_3x3_filter.h index 4cfaa0f36defa9c1f7d4a51af243c416bf09e331..0ce64f8c70d76f970df610f47947580a1efde720 100644 --- a/tensorflow/contrib/lite/kernels/internal/optimized/depthwiseconv_uint8_3x3_filter.h +++ b/tensorflow/contrib/lite/kernels/internal/optimized/depthwiseconv_uint8_3x3_filter.h @@ -3242,6 +3242,7 @@ inline void DepthwiseConv3x3Filter( int32 output_shift, int32 output_activation_min, int32 output_activation_max, uint8* output_data, const Dims<4>& output_dims) { + gemmlowp::ScopedProfilingLabel label(__PRETTY_FUNCTION__); DepthwiseConvParams params; params.input_depth = ArraySize(input_dims, 0); params.input_width = ArraySize(input_dims, 1); diff --git a/tensorflow/contrib/lite/kernels/internal/optimized/legacy_optimized_ops.h b/tensorflow/contrib/lite/kernels/internal/optimized/legacy_optimized_ops.h index 7816752132761d9523ffc1f45b3740c0817ed402..6db41d796156827567adb3ba98698ebfd27b7dc4 100644 --- a/tensorflow/contrib/lite/kernels/internal/optimized/legacy_optimized_ops.h +++ b/tensorflow/contrib/lite/kernels/internal/optimized/legacy_optimized_ops.h @@ -61,9 +61,17 @@ inline void AveragePool(const float* input_data, const Dims<4>& input_dims, float output_activation_min, float output_activation_max, float* output_data, const Dims<4>& output_dims) { - AveragePool(input_data, DimsToShape(input_dims), stride_width, stride_height, - pad_width, pad_height, kwidth, kheight, output_activation_min, - output_activation_max, output_data, DimsToShape(output_dims)); + tflite::PoolParams params; + params.stride_height = stride_height; + params.stride_width = stride_width; + params.filter_height = kheight; + params.filter_width = kwidth; + params.padding_values.height = pad_height; + params.padding_values.width = pad_width; + params.float_activation_min = output_activation_min; + params.float_activation_max = output_activation_max; + AveragePool(params, DimsToShape(input_dims), input_data, + DimsToShape(output_dims), output_data); } // legacy, for compatibility with old checked-in code @@ -96,10 +104,17 @@ inline void AveragePool(const uint8* input_data, const Dims<4>& input_dims, int32 output_activation_min, int32 output_activation_max, uint8* output_data, const Dims<4>& output_dims) { - AveragePool(input_data, DimsToShape(input_dims), stride_width, stride_height, - pad_width, pad_height, filter_width, filter_height, - output_activation_min, output_activation_max, output_data, - DimsToShape(output_dims)); + tflite::PoolParams params; + params.stride_height = stride_height; + params.stride_width = stride_width; + params.filter_height = filter_height; + params.filter_width = filter_width; + params.padding_values.height = pad_height; + params.padding_values.width = pad_width; + params.quantized_activation_min = output_activation_min; + params.quantized_activation_max = output_activation_max; + AveragePool(params, DimsToShape(input_dims), input_data, + DimsToShape(output_dims), output_data); } // legacy, for compatibility with old checked-in code @@ -140,9 +155,17 @@ inline void MaxPool(const float* input_data, const Dims<4>& input_dims, int pad_height, int kwidth, int kheight, float output_activation_min, float output_activation_max, float* output_data, const Dims<4>& output_dims) { - MaxPool(input_data, DimsToShape(input_dims), stride_width, stride_height, - pad_width, pad_height, kwidth, kheight, output_activation_min, - output_activation_max, output_data, DimsToShape(output_dims)); + tflite::PoolParams params; + params.stride_height = stride_height; + params.stride_width = stride_width; + params.filter_height = kheight; + params.filter_width = kwidth; + params.padding_values.height = pad_height; + params.padding_values.width = pad_width; + params.float_activation_min = output_activation_min; + params.float_activation_max = output_activation_max; + MaxPool(params, DimsToShape(input_dims), input_data, DimsToShape(output_dims), + output_data); } // legacy, for compatibility with old checked-in code @@ -172,10 +195,17 @@ inline void MaxPool(const uint8* input_data, const Dims<4>& input_dims, int pad_height, int filter_width, int filter_height, int32 output_activation_min, int32 output_activation_max, uint8* output_data, const Dims<4>& output_dims) { - MaxPool(input_data, DimsToShape(input_dims), stride_width, stride_height, - pad_width, pad_height, filter_width, filter_height, - output_activation_min, output_activation_max, output_data, - DimsToShape(output_dims)); + PoolParams params; + params.stride_height = stride_height; + params.stride_width = stride_width; + params.filter_height = filter_height; + params.filter_width = filter_width; + params.padding_values.height = pad_height; + params.padding_values.width = pad_width; + params.quantized_activation_min = output_activation_min; + params.quantized_activation_max = output_activation_max; + MaxPool(params, DimsToShape(input_dims), input_data, DimsToShape(output_dims), + output_data); } // legacy, for compatibility with old checked-in code @@ -215,10 +245,17 @@ inline void L2Pool(const float* input_data, const Dims<4>& input_dims, int pad_height, int filter_width, int filter_height, float output_activation_min, float output_activation_max, float* output_data, const Dims<4>& output_dims) { - L2Pool(input_data, DimsToShape(input_dims), stride_width, stride_height, - pad_width, pad_height, filter_width, filter_height, - output_activation_min, output_activation_max, output_data, - DimsToShape(output_dims)); + PoolParams params; + params.stride_height = stride_height; + params.stride_width = stride_width; + params.filter_height = filter_height; + params.filter_width = filter_width; + params.padding_values.height = pad_height; + params.padding_values.width = pad_width; + params.float_activation_min = output_activation_min; + params.float_activation_max = output_activation_max; + L2Pool(params, DimsToShape(input_dims), input_data, DimsToShape(output_dims), + output_data); } // legacy, for compatibility with old checked-in code diff --git a/tensorflow/contrib/lite/kernels/internal/optimized/multithreaded_conv.h b/tensorflow/contrib/lite/kernels/internal/optimized/multithreaded_conv.h index 27d9224512a835ea58911031f1b4d6dcf5482ba9..4a3545d47aca7d649061d39cbc23fa7ddf208156 100644 --- a/tensorflow/contrib/lite/kernels/internal/optimized/multithreaded_conv.h +++ b/tensorflow/contrib/lite/kernels/internal/optimized/multithreaded_conv.h @@ -35,35 +35,6 @@ limitations under the License. namespace tflite { namespace multithreaded_ops { -class EigenThreadPoolWrapper : public Eigen::ThreadPoolInterface { - public: - explicit EigenThreadPoolWrapper(Eigen::ThreadPool* pool) : pool_(pool) {} - ~EigenThreadPoolWrapper() override {} - - void Schedule(std::function fn) override { - pool_->Schedule(std::move(fn)); - } - int NumThreads() const override { return pool_->NumThreads(); } - int CurrentThreadId() const override { return pool_->CurrentThreadId(); } - - private: - Eigen::ThreadPool* pool_ = nullptr; -}; - -// We have a single global threadpool for all convolution operations. This means -// that inferences started from different threads may block each other, but -// since the underlying resource of CPU cores should be consumed by the -// operations anyway, it shouldn't affect overall performance. -const Eigen::ThreadPoolDevice& GetThreadPoolDevice() { - const int thread_count = 4; - static Eigen::ThreadPool* tp = new Eigen::ThreadPool(thread_count); - static EigenThreadPoolWrapper* thread_pool_wrapper = - new EigenThreadPoolWrapper(tp); - static Eigen::ThreadPoolDevice* device = - new Eigen::ThreadPoolDevice(thread_pool_wrapper, thread_count); - return *device; -} - // Shorthands for the types we need when interfacing with the EigenTensor // library. typedef Eigen::TensorMap< @@ -113,14 +84,13 @@ class EigenTensorConvFunctor { } public: - void operator()(const T* input_data, T* im2col_buffer, int input_batches, - int input_height, int input_width, int input_depth, - const T* filter_data, int filter_height, int filter_width, - int filter_count, int stride_rows, int stride_cols, - int pad_width, int pad_height, TfLitePadding padding, - T* output_data, int output_height, int output_width) { - const Eigen::ThreadPoolDevice& device = GetThreadPoolDevice(); - + void operator()(const Eigen::ThreadPoolDevice& device, const T* input_data, + T* im2col_buffer, int input_batches, int input_height, + int input_width, int input_depth, const T* filter_data, + int filter_height, int filter_width, int filter_count, + int stride_rows, int stride_cols, int pad_width, + int pad_height, TfLitePadding padding, T* output_data, + int output_height, int output_width) { const bool is_1x1_kernel = (filter_height == 1 && filter_width == 1 && stride_rows == 1 && stride_cols == 1); if (is_1x1_kernel) { @@ -162,11 +132,11 @@ class EigenTensorConvFunctor { } }; -inline void Conv(const float* input_data, const Dims<4>& input_dims, - const float* filter_data, const Dims<4>& filter_dims, - const float* bias_data, const Dims<4>& bias_dims, - int stride_width, int stride_height, int pad_width, - int pad_height, TfLitePadding padding, +inline void Conv(const Eigen::ThreadPoolDevice& device, const float* input_data, + const Dims<4>& input_dims, const float* filter_data, + const Dims<4>& filter_dims, const float* bias_data, + const Dims<4>& bias_dims, int stride_width, int stride_height, + int pad_width, int pad_height, TfLitePadding padding, float output_activation_min, float output_activation_max, float* output_data, const Dims<4>& output_dims, float* im2col_data, const Dims<4>& im2col_dims) { @@ -180,10 +150,11 @@ inline void Conv(const float* input_data, const Dims<4>& input_dims, const int output_height = ArraySize(output_dims, 2); const int output_width = ArraySize(output_dims, 1); EigenTensorConvFunctor conv_functor; - conv_functor(input_data, im2col_data, batches, input_height, input_width, - input_depth, filter_data, filter_height, filter_width, - output_depth, stride_height, stride_width, pad_height, pad_width, - padding, output_data, output_height, output_width); + conv_functor(device, input_data, im2col_data, batches, input_height, + input_width, input_depth, filter_data, filter_height, + filter_width, output_depth, stride_height, stride_width, + pad_height, pad_width, padding, output_data, output_height, + output_width); optimized_ops::AddBiasAndEvalActivationFunction( bias_data, bias_dims, output_data, output_dims, output_activation_min, diff --git a/tensorflow/contrib/lite/kernels/internal/optimized/neon_tensor_utils.cc b/tensorflow/contrib/lite/kernels/internal/optimized/neon_tensor_utils.cc index 38ad32c734a2286c7d23162810625169a4d8df43..8c57c987d7b73b6082d06f0c64443c6c92bff1eb 100644 --- a/tensorflow/contrib/lite/kernels/internal/optimized/neon_tensor_utils.cc +++ b/tensorflow/contrib/lite/kernels/internal/optimized/neon_tensor_utils.cc @@ -55,83 +55,33 @@ void NeonMatrixBatchVectorMultiplyAccumulate(const float* matrix, int m_rows, const int postamble_start = m_cols - (m_cols & (kFloatWeightsPerNeonLane - 1)); - // The arrays used to cache the vector. - void* aligned_vector_cache_free = nullptr; - float32x4_t* vector_cache_float32x4 = - reinterpret_cast(aligned_alloc( - sizeof(float32x4_t), (postamble_start >> 2) * sizeof(float32x4_t), - &aligned_vector_cache_free)); - - const int kUnrollSize = 2; for (int b = 0; b < n_batch; b++) { float* result_in_batch = result + b * m_rows * result_stride; const float* vector_in_batch = vector + b * m_cols; + const float* matrix_row = matrix; - const float* matrix_ptr0 = matrix; - // If there is only 1 row, we don't want to assign an illegal pointer. - const float* matrix_ptr1 = nullptr; - if (m_rows > 1) { - matrix_ptr1 = matrix + m_cols; - } - - // Cache the vector. - for (int c = 0; c < postamble_start; c += kFloatWeightsPerNeonLane) { - vector_cache_float32x4[c >> 2] = vld1q_f32(vector_in_batch + c); - } - - // Main matrix by vector multiplication loop, which handles two rows of - // matrix by vector multiplication. - for (int r = 0; r < (m_rows & ~(kUnrollSize - 1)); r += kUnrollSize) { - float32x4_t acc0_32x4 = vmovq_n_f32(0.0); - float32x4_t acc1_32x4 = vmovq_n_f32(0.0); - for (int c = 0; c < postamble_start; c += kFloatWeightsPerNeonLane) { - float32x4_t temp = vector_cache_float32x4[c >> 2]; - // Load 4 float values from vector1 and vector2 and accumulator. - float32x4_t v0_f32x4 = vld1q_f32(matrix_ptr0 + c); - float32x4_t v1_f32x4 = vld1q_f32(matrix_ptr1 + c); - // Vector multiply-accumulate 4 float - acc0_32x4 = vmlaq_f32(acc0_32x4, v0_f32x4, temp); - acc1_32x4 = vmlaq_f32(acc1_32x4, v1_f32x4, temp); - } - // Add the 4 intermediate sum values to get the final dot-prod value for - // this column. - *result_in_batch += - (vgetq_lane_f32(acc0_32x4, 0) + vgetq_lane_f32(acc0_32x4, 1) + - vgetq_lane_f32(acc0_32x4, 2) + vgetq_lane_f32(acc0_32x4, 3)); - *(result_in_batch + result_stride) += - (vgetq_lane_f32(acc1_32x4, 0) + vgetq_lane_f32(acc1_32x4, 1) + - vgetq_lane_f32(acc1_32x4, 2) + vgetq_lane_f32(acc1_32x4, 3)); - for (int c = postamble_start; c < m_cols; c++) { - *result_in_batch += matrix_ptr0[c] * vector_in_batch[c]; - *(result_in_batch + result_stride) += - matrix_ptr1[c] * vector_in_batch[c]; - } - matrix_ptr0 += kUnrollSize * m_cols; - matrix_ptr1 += kUnrollSize * m_cols; - result_in_batch += kUnrollSize * result_stride; - } - for (int r = (m_rows & ~(kUnrollSize - 1)); r < m_rows; r++) { - float32x4_t acc0_32x4 = vmovq_n_f32(0.0); + // Main matrix by vector multiplication loop + for (int r = 0; r < m_rows; r++) { + float32x4_t acc_32x4 = vmovq_n_f32(0.0); for (int c = 0; c < postamble_start; c += kFloatWeightsPerNeonLane) { - float32x4_t temp = vector_cache_float32x4[c >> 2]; - // Load 4 float values from vector1 and vector2 and accumulator. - float32x4_t v0_f32x4 = vld1q_f32(matrix_ptr0 + c); - // Vector multiply-accumulate 4 float - acc0_32x4 = vmlaq_f32(acc0_32x4, v0_f32x4, temp); + // Load 4 float values from vector and matrix row. + float32x4_t vector_f32x4 = vld1q_f32(vector_in_batch + c); + float32x4_t matrix_f32x4 = vld1q_f32(matrix_row + c); + // Multiply the vector and matrix row and add to accumulator. + acc_32x4 = vmlaq_f32(acc_32x4, matrix_f32x4, vector_f32x4); } // Add the 4 intermediate sum values to get the final dot-prod value for // this column. *result_in_batch += - (vgetq_lane_f32(acc0_32x4, 0) + vgetq_lane_f32(acc0_32x4, 1) + - vgetq_lane_f32(acc0_32x4, 2) + vgetq_lane_f32(acc0_32x4, 3)); + (vgetq_lane_f32(acc_32x4, 0) + vgetq_lane_f32(acc_32x4, 1) + + vgetq_lane_f32(acc_32x4, 2) + vgetq_lane_f32(acc_32x4, 3)); for (int c = postamble_start; c < m_cols; c++) { - *result_in_batch += matrix_ptr0[c] * vector_in_batch[c]; + *result_in_batch += matrix_row[c] * vector_in_batch[c]; } - matrix_ptr0 += m_cols; + matrix_row += m_cols; result_in_batch += result_stride; } } - free(aligned_vector_cache_free); } void NeonMatrixBatchVectorMultiplyAccumulate( @@ -162,7 +112,7 @@ void NeonMatrixBatchVectorMultiplyAccumulate( int batch, row, col; for (batch = 0; batch < n_batch; ++batch) { - const float batch_scaling_factor_inv = 1.0 / scaling_factors[batch]; + const float batch_scaling_factor = scaling_factors[batch]; // Copy the vector data to an aligned vector. memcpy(aligned_vec, vectors + batch * m_cols, sizeof(int8) * m_cols); // Compute dot-product for every column. @@ -232,7 +182,7 @@ void NeonMatrixBatchVectorMultiplyAccumulate( int32 neon_sum = vgetq_lane_s64(pairwiseAdded, 0) + vgetq_lane_s64(pairwiseAdded, 1); - *result += ((neon_sum + postable_sum) * batch_scaling_factor_inv); + *result += ((neon_sum + postable_sum) * batch_scaling_factor); } // for row } // for batch @@ -296,17 +246,6 @@ void NeonVectorBatchVectorCwiseProductAccumulate(const float* vector, const int postamble_start = v_size - (v_size & (kFloatWeightsPerNeonLane - 1)); - // The arrays used to cache the vector. - void* aligned_vector_cache_free = nullptr; - float32x4_t* vector_cache_float32x4 = - reinterpret_cast(aligned_alloc( - sizeof(float32x4_t), (postamble_start >> 2) * sizeof(float32x4_t), - &aligned_vector_cache_free)); - - for (int v = 0; v < postamble_start; v += kFloatWeightsPerNeonLane) { - vector_cache_float32x4[v >> 2] = vld1q_f32(vector + v); - } - float* result_ptr = result; const float* batch_vector_ptr = batch_vector; for (int b = 0; b < n_batch; b++) { @@ -314,9 +253,9 @@ void NeonVectorBatchVectorCwiseProductAccumulate(const float* vector, // Load from memory to vectors. float32x4_t result_f32x4 = vld1q_f32(result_ptr + v); float32x4_t batch_vector_f32x4 = vld1q_f32(batch_vector_ptr + v); + float32x4_t vector_f32x4 = vld1q_f32(vector + v); // Multiply-accumulate. - result_f32x4 = vmlaq_f32(result_f32x4, batch_vector_f32x4, - vector_cache_float32x4[v >> 2]); + result_f32x4 = vmlaq_f32(result_f32x4, batch_vector_f32x4, vector_f32x4); // Store. vst1q_f32(result_ptr + v, result_f32x4); } @@ -328,7 +267,6 @@ void NeonVectorBatchVectorCwiseProductAccumulate(const float* vector, result_ptr += v_size; batch_vector_ptr += v_size; } - free(aligned_vector_cache_free); } void NeonSub1Vector(const float* vector, int v_size, float* result) { @@ -418,13 +356,14 @@ void NeonSymmetricQuantizeFloats(const float* values, const int size, *scaling_factor = 1; return; } - *scaling_factor = kScale / range; + *scaling_factor = range / kScale; + const float scaling_factor_inv = 1.0f / *scaling_factor; const int postamble_start = size - (size & (2 * kFloatWeightsPerNeonLane - 1)); // Vectorized constants. - const float32x4_t q_factor_f32x4 = vmovq_n_f32(*scaling_factor); + const float32x4_t q_factor_f32x4 = vmovq_n_f32(scaling_factor_inv); const float32x4_t point5_f32x4 = vmovq_n_f32(0.5); const float32x4_t zero_f32x4 = vmovq_n_f32(0.0); const int32x4_t scale_i32x4 = vmovq_n_s32(kScale); @@ -476,7 +415,7 @@ void NeonSymmetricQuantizeFloats(const float* values, const int size, for (int i = postamble_start; i < size; ++i) { const int32 quantized_value = - static_cast(TfLiteRound(*scaling_factor * values[i])); + static_cast(TfLiteRound(scaling_factor_inv * values[i])); quantized_values[i] = std::min(kScale, std::max(-kScale, quantized_value)); } } diff --git a/tensorflow/contrib/lite/kernels/internal/optimized/optimized_ops.h b/tensorflow/contrib/lite/kernels/internal/optimized/optimized_ops.h index 6b5d35f21ea1dd7c821f2d8efb410cbb62cd7d5f..c857fdf6995aef7d1ec09f1988411370ff483691 100644 --- a/tensorflow/contrib/lite/kernels/internal/optimized/optimized_ops.h +++ b/tensorflow/contrib/lite/kernels/internal/optimized/optimized_ops.h @@ -41,6 +41,7 @@ namespace optimized_ops { // Unoptimized reference ops: using reference_ops::ArgMax; +using reference_ops::ArgMinMax; using reference_ops::BroadcastGreater; using reference_ops::BroadcastGreaterEqual; using reference_ops::BroadcastLess; @@ -59,6 +60,7 @@ using reference_ops::Mean; using reference_ops::RankOneSelect; using reference_ops::Relu1; using reference_ops::Relu6; +using reference_ops::ReluX; using reference_ops::Select; using reference_ops::SpaceToBatchND; using reference_ops::StridedSlice; @@ -170,16 +172,9 @@ template MatrixMap MapAsMatrixWithGivenNumberOfRows(Scalar* data, const Dims& dims, int rows) { - int cols = 1; - bool matched_rows = false; - for (int d = 0; d < N; d++) { - cols *= dims.sizes[d]; - if (cols == rows) { - matched_rows = true; - cols = 1; - } - } - TFLITE_DCHECK(matched_rows); + const int flatsize = FlatSize(dims); + TFLITE_DCHECK((flatsize % rows) == 0); + const int cols = flatsize / rows; return MatrixMap(data, rows, cols); } @@ -2714,6 +2709,20 @@ inline void Add(const int16* input1_data, const Dims<4>& input1_dims, } } +inline void Add(const int32* input1_data, const Dims<4>& input1_dims, + const int32* input2_data, const Dims<4>& input2_dims, + int32 output_activation_min, int32 output_activation_max, + int32* output_data, const Dims<4>& output_dims) { + gemmlowp::ScopedProfilingLabel label("Add/int32"); + + const int flat_size = MatchingFlatSize(input1_dims, input2_dims, output_dims); + for (int i = 0; i < flat_size; ++i) { + output_data[i] = ActivationFunctionWithMinMax( + input1_data[i] + input2_data[i], output_activation_min, + output_activation_max); + } +} + template inline void Add(const int16* input1_data, const Dims<4>& input1_dims, int input1_shift, const int16* input2_data, @@ -3045,6 +3054,20 @@ void Mul(const float* input1_data, const Dims<4>& input1_dims, output_activation_max, output_data, output_dims); } +inline void Mul(const int32* input1_data, const Dims<4>& input1_dims, + const int32* input2_data, const Dims<4>& input2_dims, + int32 output_activation_min, int32 output_activation_max, + int32* output_data, const Dims<4>& output_dims) { + gemmlowp::ScopedProfilingLabel label("Mul/int32"); + + const int flat_size = MatchingFlatSize(input1_dims, input2_dims, output_dims); + for (int i = 0; i < flat_size; ++i) { + output_data[i] = ActivationFunctionWithMinMax( + input1_data[i] * input2_data[i], output_activation_min, + output_activation_max); + } +} + template void Mul(const int32* input1_data, const Dims<4>& input1_dims, const int32* input2_data, const Dims<4>& input2_dims, @@ -3763,21 +3786,20 @@ inline int NodeOffset(int b, int h, int w, int height, int width) { return (b * height + h) * width + w; } -inline void AveragePool(const float* input_data, - const RuntimeShape& input_shape, int stride_width, - int stride_height, int pad_width, int pad_height, - int kwidth, int kheight, float output_activation_min, - float output_activation_max, float* output_data, - const RuntimeShape& output_shape) { +inline void AveragePool(const PoolParams& params, + const RuntimeShape& input_shape, + const float* input_data, + const RuntimeShape& output_shape, float* output_data) { gemmlowp::ScopedProfilingLabel label("AveragePool"); TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4); TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4); const int batches = MatchingDim(input_shape, 0, output_shape, 0); - const int depth = MatchingDim(input_shape, 3, output_shape, 3); const int input_height = input_shape.Dims(1); const int input_width = input_shape.Dims(2); const int output_height = output_shape.Dims(1); const int output_width = output_shape.Dims(2); + const int stride_height = params.stride_height; + const int stride_width = params.stride_width; // TODO(benoitjacob) make this a proper reference impl without Eigen! const auto in_mat = MapAsMatrixWithLastDimAsRows(input_data, input_shape); @@ -3792,12 +3814,15 @@ inline void AveragePool(const float* input_data, for (int w = 0; w < input_width; ++w) { // (h_start, h_end) * (w_start, w_end) is the range that the input // vector projects to. - int hpad = h + pad_height; - int wpad = w + pad_width; - int h_start = - (hpad < kheight) ? 0 : (hpad - kheight) / stride_height + 1; + int hpad = h + params.padding_values.height; + int wpad = w + params.padding_values.width; + int h_start = (hpad < params.filter_height) + ? 0 + : (hpad - params.filter_height) / stride_height + 1; int h_end = std::min(hpad / stride_height + 1, output_height); - int w_start = (wpad < kwidth) ? 0 : (wpad - kwidth) / stride_width + 1; + int w_start = (wpad < params.filter_width) + ? 0 + : (wpad - params.filter_width) / stride_width + 1; int w_end = std::min(wpad / stride_width + 1, output_width); // compute elementwise sum for (int ph = h_start; ph < h_end; ++ph) { @@ -3815,29 +3840,21 @@ inline void AveragePool(const float* input_data, TFLITE_DCHECK_GT(out_count.minCoeff(), 0); out_mat.array().rowwise() /= out_count.transpose().array(); - for (int b = 0; b < batches; ++b) { - for (int y = 0; y < output_height; ++y) { - for (int x = 0; x < output_width; ++x) { - for (int c = 0; c < depth; ++c) { - output_data[Offset(output_shape, b, y, x, c)] = - ActivationFunctionWithMinMax( - output_data[Offset(output_shape, b, y, x, c)], - output_activation_min, output_activation_max); - } - } - } + const int flat_size = output_shape.FlatSize(); + for (int i = 0; i < flat_size; ++i) { + output_data[i] = ActivationFunctionWithMinMax(output_data[i], + params.float_activation_min, + params.float_activation_max); } } -inline void AveragePool(const uint8* input_data, - const RuntimeShape& input_shape, int stride_width, - int stride_height, int pad_width, int pad_height, - int filter_width, int filter_height, - int32 output_activation_min, - int32 output_activation_max, uint8* output_data, - const RuntimeShape& output_shape) { +inline void AveragePool(const PoolParams& params, + const RuntimeShape& input_shape, + const uint8* input_data, + const RuntimeShape& output_shape, uint8* output_data) { gemmlowp::ScopedProfilingLabel label("AveragePool/8bit"); - TFLITE_DCHECK_LE(output_activation_min, output_activation_max); + TFLITE_DCHECK_LE(params.quantized_activation_min, + params.quantized_activation_max); TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4); TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4); const int batches = MatchingDim(input_shape, 0, output_shape, 0); @@ -3846,17 +3863,21 @@ inline void AveragePool(const uint8* input_data, const int input_width = input_shape.Dims(2); const int output_height = output_shape.Dims(1); const int output_width = output_shape.Dims(2); + const int stride_height = params.stride_height; + const int stride_width = params.stride_width; for (int batch = 0; batch < batches; ++batch) { for (int out_y = 0; out_y < output_height; ++out_y) { for (int out_x = 0; out_x < output_width; ++out_x) { - const int in_x_origin = (out_x * stride_width) - pad_width; - const int in_y_origin = (out_y * stride_height) - pad_height; + const int in_x_origin = + (out_x * stride_width) - params.padding_values.width; + const int in_y_origin = + (out_y * stride_height) - params.padding_values.height; const int filter_x_start = std::max(0, -in_x_origin); const int filter_x_end = - std::min(filter_width, input_width - in_x_origin); + std::min(params.filter_width, input_width - in_x_origin); const int filter_y_start = std::max(0, -in_y_origin); const int filter_y_end = - std::min(filter_height, input_height - in_y_origin); + std::min(params.filter_height, input_height - in_y_origin); const int filter_count = (filter_x_end - filter_x_start) * (filter_y_end - filter_y_start); // 1280 required by Inception v3 @@ -3904,18 +3925,18 @@ inline void AveragePool(const uint8* input_data, output_data + Offset(output_shape, batch, out_y, out_x, 0); int channel = 0; #ifdef USE_NEON -#define AVGPOOL_DIVIDING_BY(FILTER_COUNT) \ - if (filter_count == FILTER_COUNT) { \ - for (; channel <= depth - 8; channel += 8) { \ - uint16 buf[8]; \ - for (int i = 0; i < 8; i++) { \ - buf[i] = (acc[channel + i] + FILTER_COUNT / 2) / FILTER_COUNT; \ - } \ - uint8x8_t buf8 = vqmovn_u16(vld1q_u16(buf)); \ - buf8 = vmin_u8(buf8, vdup_n_u8(output_activation_max)); \ - buf8 = vmax_u8(buf8, vdup_n_u8(output_activation_min)); \ - vst1_u8(output_ptr + channel, buf8); \ - } \ +#define AVGPOOL_DIVIDING_BY(FILTER_COUNT) \ + if (filter_count == FILTER_COUNT) { \ + for (; channel <= depth - 8; channel += 8) { \ + uint16 buf[8]; \ + for (int i = 0; i < 8; i++) { \ + buf[i] = (acc[channel + i] + FILTER_COUNT / 2) / FILTER_COUNT; \ + } \ + uint8x8_t buf8 = vqmovn_u16(vld1q_u16(buf)); \ + buf8 = vmin_u8(buf8, vdup_n_u8(params.quantized_activation_max)); \ + buf8 = vmax_u8(buf8, vdup_n_u8(params.quantized_activation_min)); \ + vst1_u8(output_ptr + channel, buf8); \ + } \ } AVGPOOL_DIVIDING_BY(9) AVGPOOL_DIVIDING_BY(15) @@ -3926,15 +3947,15 @@ inline void AveragePool(const uint8* input_data, buf[i] = (acc[channel + i] + filter_count / 2) / filter_count; } uint8x8_t buf8 = vqmovn_u16(vld1q_u16(buf)); - buf8 = vmin_u8(buf8, vdup_n_u8(output_activation_max)); - buf8 = vmax_u8(buf8, vdup_n_u8(output_activation_min)); + buf8 = vmin_u8(buf8, vdup_n_u8(params.quantized_activation_max)); + buf8 = vmax_u8(buf8, vdup_n_u8(params.quantized_activation_min)); vst1_u8(output_ptr + channel, buf8); } #endif for (; channel < depth; ++channel) { uint16 a = (acc[channel] + filter_count / 2) / filter_count; - a = std::max(a, output_activation_min); - a = std::min(a, output_activation_max); + a = std::max(a, params.quantized_activation_min); + a = std::min(a, params.quantized_activation_max); output_ptr[channel] = static_cast(a); } } @@ -3942,20 +3963,19 @@ inline void AveragePool(const uint8* input_data, } } -inline void MaxPool(const float* input_data, const RuntimeShape& input_shape, - int stride_width, int stride_height, int pad_width, - int pad_height, int kwidth, int kheight, - float output_activation_min, float output_activation_max, - float* output_data, const RuntimeShape& output_shape) { +inline void MaxPool(const PoolParams& params, const RuntimeShape& input_shape, + const float* input_data, const RuntimeShape& output_shape, + float* output_data) { gemmlowp::ScopedProfilingLabel label("MaxPool"); TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4); TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4); const int batches = MatchingDim(input_shape, 0, output_shape, 0); - const int depth = MatchingDim(input_shape, 3, output_shape, 3); const int input_height = input_shape.Dims(1); const int input_width = input_shape.Dims(2); const int output_height = output_shape.Dims(1); const int output_width = output_shape.Dims(2); + const int stride_height = params.stride_height; + const int stride_width = params.stride_width; const auto in_mat = MapAsMatrixWithLastDimAsRows(input_data, input_shape); auto out_mat = MapAsMatrixWithLastDimAsRows(output_data, output_shape); @@ -3966,12 +3986,15 @@ inline void MaxPool(const float* input_data, const RuntimeShape& input_shape, for (int w = 0; w < input_width; ++w) { // (h_start, h_end) * (w_start, w_end) is the range that the input // vector projects to. - int hpad = h + pad_height; - int wpad = w + pad_width; - int h_start = - (hpad < kheight) ? 0 : (hpad - kheight) / stride_height + 1; + int hpad = h + params.padding_values.height; + int wpad = w + params.padding_values.width; + int h_start = (hpad < params.filter_height) + ? 0 + : (hpad - params.filter_height) / stride_height + 1; int h_end = std::min(hpad / stride_height + 1, output_height); - int w_start = (wpad < kwidth) ? 0 : (wpad - kwidth) / stride_width + 1; + int w_start = (wpad < params.filter_width) + ? 0 + : (wpad - params.filter_width) / stride_width + 1; int w_end = std::min(wpad / stride_width + 1, output_width); // compute elementwise sum for (int ph = h_start; ph < h_end; ++ph) { @@ -3986,28 +4009,20 @@ inline void MaxPool(const float* input_data, const RuntimeShape& input_shape, } } } - - for (int b = 0; b < batches; ++b) { - for (int y = 0; y < output_height; ++y) { - for (int x = 0; x < output_width; ++x) { - for (int c = 0; c < depth; ++c) { - output_data[Offset(output_shape, b, y, x, c)] = - ActivationFunctionWithMinMax( - output_data[Offset(output_shape, b, y, x, c)], - output_activation_min, output_activation_max); - } - } - } + const int flat_size = output_shape.FlatSize(); + for (int i = 0; i < flat_size; ++i) { + output_data[i] = ActivationFunctionWithMinMax(output_data[i], + params.float_activation_min, + params.float_activation_max); } } -inline void MaxPool(const uint8* input_data, const RuntimeShape& input_shape, - int stride_width, int stride_height, int pad_width, - int pad_height, int filter_width, int filter_height, - int32 output_activation_min, int32 output_activation_max, - uint8* output_data, const RuntimeShape& output_shape) { +inline void MaxPool(const PoolParams& params, const RuntimeShape& input_shape, + const uint8* input_data, const RuntimeShape& output_shape, + uint8* output_data) { gemmlowp::ScopedProfilingLabel label("MaxPool/8bit"); - TFLITE_DCHECK_LE(output_activation_min, output_activation_max); + TFLITE_DCHECK_LE(params.quantized_activation_min, + params.quantized_activation_max); TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4); TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4); const int batches = MatchingDim(input_shape, 0, output_shape, 0); @@ -4016,17 +4031,21 @@ inline void MaxPool(const uint8* input_data, const RuntimeShape& input_shape, const int input_width = input_shape.Dims(2); const int output_height = output_shape.Dims(1); const int output_width = output_shape.Dims(2); + const int stride_height = params.stride_height; + const int stride_width = params.stride_width; for (int batch = 0; batch < batches; ++batch) { for (int out_y = 0; out_y < output_height; ++out_y) { for (int out_x = 0; out_x < output_width; ++out_x) { - const int in_x_origin = (out_x * stride_width) - pad_width; - const int in_y_origin = (out_y * stride_height) - pad_height; + const int in_x_origin = + (out_x * stride_width) - params.padding_values.width; + const int in_y_origin = + (out_y * stride_height) - params.padding_values.height; const int filter_x_start = std::max(0, -in_x_origin); const int filter_x_end = - std::min(filter_width, input_width - in_x_origin); + std::min(params.filter_width, input_width - in_x_origin); const int filter_y_start = std::max(0, -in_y_origin); const int filter_y_end = - std::min(filter_height, input_height - in_y_origin); + std::min(params.filter_height, input_height - in_y_origin); // 2048 required by Inception v3 static constexpr int kAccBufferMaxSize = 2048; TFLITE_DCHECK_LE(depth, kAccBufferMaxSize); @@ -4069,21 +4088,21 @@ inline void MaxPool(const uint8* input_data, const RuntimeShape& input_shape, #ifdef USE_NEON for (; channel <= depth - 16; channel += 16) { uint8x16_t a = vld1q_u8(acc + channel); - a = vminq_u8(a, vdupq_n_u8(output_activation_max)); - a = vmaxq_u8(a, vdupq_n_u8(output_activation_min)); + a = vminq_u8(a, vdupq_n_u8(params.quantized_activation_max)); + a = vmaxq_u8(a, vdupq_n_u8(params.quantized_activation_min)); vst1q_u8(output_ptr + channel, a); } for (; channel <= depth - 8; channel += 8) { uint8x8_t a = vld1_u8(acc + channel); - a = vmin_u8(a, vdup_n_u8(output_activation_max)); - a = vmax_u8(a, vdup_n_u8(output_activation_min)); + a = vmin_u8(a, vdup_n_u8(params.quantized_activation_max)); + a = vmax_u8(a, vdup_n_u8(params.quantized_activation_min)); vst1_u8(output_ptr + channel, a); } #endif for (; channel < depth; ++channel) { uint8 a = acc[channel]; - a = std::max(a, output_activation_min); - a = std::min(a, output_activation_max); + a = std::max(a, params.quantized_activation_min); + a = std::min(a, params.quantized_activation_max); output_ptr[channel] = static_cast(a); } } @@ -4091,11 +4110,9 @@ inline void MaxPool(const uint8* input_data, const RuntimeShape& input_shape, } } -inline void L2Pool(const float* input_data, const RuntimeShape& input_shape, - int stride_width, int stride_height, int pad_width, - int pad_height, int filter_width, int filter_height, - float output_activation_min, float output_activation_max, - float* output_data, const RuntimeShape& output_shape) { +inline void L2Pool(const PoolParams& params, const RuntimeShape& input_shape, + const float* input_data, const RuntimeShape& output_shape, + float* output_data) { gemmlowp::ScopedProfilingLabel label("L2Pool"); TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4); TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4); @@ -4104,6 +4121,8 @@ inline void L2Pool(const float* input_data, const RuntimeShape& input_shape, const int input_width = input_shape.Dims(2); const int output_height = output_shape.Dims(1); const int output_width = output_shape.Dims(2); + const int stride_height = params.stride_height; + const int stride_width = params.stride_width; // Actually carry out L2 Pool. Code is written in forward mode: we go through // the input values once, and write to all the pooled regions that it maps to. const auto in_mat = MapAsMatrixWithLastDimAsRows(input_data, input_shape); @@ -4118,15 +4137,17 @@ inline void L2Pool(const float* input_data, const RuntimeShape& input_shape, for (int w = 0; w < input_width; ++w) { // (h_start, h_end) * (w_start, w_end) is the range that the input // vector projects to. - const int hpad = h + pad_height; - const int wpad = w + pad_width; - const int h_start = (hpad < filter_height) - ? 0 - : (hpad - filter_height) / stride_height + 1; + const int hpad = h + params.padding_values.height; + const int wpad = w + params.padding_values.width; + const int h_start = + (hpad < params.filter_height) + ? 0 + : (hpad - params.filter_height) / stride_height + 1; const int h_end = std::min(hpad / stride_height + 1, output_height); - const int w_start = (wpad < filter_width) - ? 0 - : (wpad - filter_width) / stride_width + 1; + const int w_start = + (wpad < params.filter_width) + ? 0 + : (wpad - params.filter_width) / stride_width + 1; const int w_end = std::min(wpad / stride_width + 1, output_width); // pre-compute square const int in_offset = w + input_width * (h + input_height * b); @@ -4147,6 +4168,13 @@ inline void L2Pool(const float* input_data, const RuntimeShape& input_shape, out_count = out_count.array().inverse(); out_mat = (out_mat.array().rowwise() * out_count.transpose().array()).cwiseSqrt(); + + const int flat_size = output_shape.FlatSize(); + for (int i = 0; i < flat_size; ++i) { + output_data[i] = ActivationFunctionWithMinMax(output_data[i], + params.float_activation_min, + params.float_activation_max); + } } inline void LocalResponseNormalization(const float* input_data, diff --git a/tensorflow/contrib/lite/kernels/internal/quantization_util.h b/tensorflow/contrib/lite/kernels/internal/quantization_util.h index 525857a2e6f73276d0a6e64770947169033c7667..9b3f1823dc7e08562d8906346bc44e4478642ddc 100644 --- a/tensorflow/contrib/lite/kernels/internal/quantization_util.h +++ b/tensorflow/contrib/lite/kernels/internal/quantization_util.h @@ -28,8 +28,9 @@ namespace tflite { // Given the min and max values of a float array, return // reasonable quantization parameters to use for this array. template -QuantizationParams ChooseQuantizationParams(double rmin, double rmax) { - const T qmin = std::numeric_limits::min(); +QuantizationParams ChooseQuantizationParams(double rmin, double rmax, + bool narrow_range) { + const T qmin = std::numeric_limits::min() + (narrow_range ? 1 : 0); const T qmax = std::numeric_limits::max(); const double qmin_double = qmin; const double qmax_double = qmax; @@ -97,6 +98,11 @@ QuantizationParams ChooseQuantizationParams(double rmin, double rmax) { return quantization_params; } +template +QuantizationParams ChooseQuantizationParams(double rmin, double rmax) { + return ChooseQuantizationParams(rmin, rmax, false); +} + // Converts a floating-point number to an integer. For all inputs x where // static_cast(x) is legal according to the C++ standard, the result // is identical to that cast (i.e. the result is x with its fractional part diff --git a/tensorflow/contrib/lite/kernels/internal/reference/legacy_reference_ops.h b/tensorflow/contrib/lite/kernels/internal/reference/legacy_reference_ops.h index 878b2441b4f2828a014673f5bd80fb8aa29514db..f715d34bc1aab4fe483dc712018b5016ffef25f5 100644 --- a/tensorflow/contrib/lite/kernels/internal/reference/legacy_reference_ops.h +++ b/tensorflow/contrib/lite/kernels/internal/reference/legacy_reference_ops.h @@ -69,9 +69,17 @@ inline void AveragePool(const float* input_data, const Dims<4>& input_dims, float output_activation_min, float output_activation_max, float* output_data, const Dims<4>& output_dims) { - AveragePool(input_data, DimsToShape(input_dims), stride_width, stride_height, - pad_width, pad_height, kwidth, kheight, output_activation_min, - output_activation_max, output_data, DimsToShape(output_dims)); + tflite::PoolParams params; + params.stride_height = stride_height; + params.stride_width = stride_width; + params.filter_height = kheight; + params.filter_width = kwidth; + params.padding_values.height = pad_height; + params.padding_values.width = pad_width; + params.float_activation_min = output_activation_min; + params.float_activation_max = output_activation_max; + AveragePool(params, DimsToShape(input_dims), input_data, + DimsToShape(output_dims), output_data); } // legacy, for compatibility with old checked-in code @@ -104,10 +112,17 @@ inline void AveragePool(const uint8* input_data, const Dims<4>& input_dims, int32 output_activation_min, int32 output_activation_max, uint8* output_data, const Dims<4>& output_dims) { - AveragePool(input_data, DimsToShape(input_dims), stride_width, stride_height, - pad_width, pad_height, filter_width, filter_height, - output_activation_min, output_activation_max, output_data, - DimsToShape(output_dims)); + tflite::PoolParams params; + params.stride_height = stride_height; + params.stride_width = stride_width; + params.filter_height = filter_height; + params.filter_width = filter_width; + params.padding_values.height = pad_height; + params.padding_values.width = pad_width; + params.quantized_activation_min = output_activation_min; + params.quantized_activation_max = output_activation_max; + AveragePool(params, DimsToShape(input_dims), input_data, + DimsToShape(output_dims), output_data); } // legacy, for compatibility with old checked-in code @@ -148,9 +163,17 @@ inline void MaxPool(const float* input_data, const Dims<4>& input_dims, int pad_height, int kwidth, int kheight, float output_activation_min, float output_activation_max, float* output_data, const Dims<4>& output_dims) { - MaxPool(input_data, DimsToShape(input_dims), stride_width, stride_height, - pad_width, pad_height, kwidth, kheight, output_activation_min, - output_activation_max, output_data, DimsToShape(output_dims)); + tflite::PoolParams params; + params.stride_height = stride_height; + params.stride_width = stride_width; + params.filter_height = kheight; + params.filter_width = kwidth; + params.padding_values.height = pad_height; + params.padding_values.width = pad_width; + params.float_activation_min = output_activation_min; + params.float_activation_max = output_activation_max; + MaxPool(params, DimsToShape(input_dims), input_data, DimsToShape(output_dims), + output_data); } // legacy, for compatibility with old checked-in code @@ -180,10 +203,17 @@ inline void MaxPool(const uint8* input_data, const Dims<4>& input_dims, int pad_height, int filter_width, int filter_height, int32 output_activation_min, int32 output_activation_max, uint8* output_data, const Dims<4>& output_dims) { - MaxPool(input_data, DimsToShape(input_dims), stride_width, stride_height, - pad_width, pad_height, filter_width, filter_height, - output_activation_min, output_activation_max, output_data, - DimsToShape(output_dims)); + PoolParams params; + params.stride_height = stride_height; + params.stride_width = stride_width; + params.filter_height = filter_height; + params.filter_width = filter_width; + params.padding_values.height = pad_height; + params.padding_values.width = pad_width; + params.quantized_activation_min = output_activation_min; + params.quantized_activation_max = output_activation_max; + MaxPool(params, DimsToShape(input_dims), input_data, DimsToShape(output_dims), + output_data); } // legacy, for compatibility with old checked-in code @@ -223,10 +253,17 @@ inline void L2Pool(const float* input_data, const Dims<4>& input_dims, int pad_height, int filter_width, int filter_height, float output_activation_min, float output_activation_max, float* output_data, const Dims<4>& output_dims) { - L2Pool(input_data, DimsToShape(input_dims), stride_width, stride_height, - pad_width, pad_height, filter_width, filter_height, - output_activation_min, output_activation_max, output_data, - DimsToShape(output_dims)); + PoolParams params; + params.stride_height = stride_height; + params.stride_width = stride_width; + params.filter_height = filter_height; + params.filter_width = filter_width; + params.padding_values.height = pad_height; + params.padding_values.width = pad_width; + params.float_activation_min = output_activation_min; + params.float_activation_max = output_activation_max; + L2Pool(params, DimsToShape(input_dims), input_data, DimsToShape(output_dims), + output_data); } // legacy, for compatibility with old checked-in code diff --git a/tensorflow/contrib/lite/kernels/internal/reference/portable_tensor_utils.cc b/tensorflow/contrib/lite/kernels/internal/reference/portable_tensor_utils.cc index f8c6f341f7e61529bbbac592f9caf115f6121e0c..ccf112c990f3b5cba755a9b29aadd5aa82104849 100644 --- a/tensorflow/contrib/lite/kernels/internal/reference/portable_tensor_utils.cc +++ b/tensorflow/contrib/lite/kernels/internal/reference/portable_tensor_utils.cc @@ -51,10 +51,11 @@ void PortableSymmetricQuantizeFloats(const float* values, const int size, *scaling_factor = 1; return; } - *scaling_factor = kScale / range; + *scaling_factor = range / kScale; + const float scaling_factor_inv = 1.0f / *scaling_factor; for (int i = 0; i < size; ++i) { const int32_t quantized_value = - static_cast(TfLiteRound(*scaling_factor * values[i])); + static_cast(TfLiteRound(values[i] * scaling_factor_inv)); // Clamp: just in case some odd numeric offset. quantized_values[i] = std::min(kScale, std::max(-kScale, quantized_value)); } @@ -85,7 +86,7 @@ void PortableMatrixBatchVectorMultiplyAccumulate( float* __restrict__ result, int result_stride) { int batch, row, col; for (batch = 0; batch < n_batch; ++batch, vectors += m_cols) { - const float batch_scaling_factor_inv = 1.0 / scaling_factors[batch]; + const float batch_scaling_factor = scaling_factors[batch]; // Get the address of the first row. const int8_t* row_ptr = matrix; for (row = 0; row < m_rows; ++row, result += result_stride) { @@ -98,7 +99,7 @@ void PortableMatrixBatchVectorMultiplyAccumulate( for (col = 0; col < m_cols; ++col, ++row_ptr) { dotprod += (*row_ptr) * (vectors[col]); } // for col - *result += (dotprod * batch_scaling_factor_inv); + *result += (dotprod * batch_scaling_factor); } // for row } // for batch } diff --git a/tensorflow/contrib/lite/kernels/internal/reference/reference_ops.h b/tensorflow/contrib/lite/kernels/internal/reference/reference_ops.h index 7b8a56a524f45fd346c6d92e8e37a86621f28d4e..2d40f1769babc90dce45a31cb8526fa23787355a 100644 --- a/tensorflow/contrib/lite/kernels/internal/reference/reference_ops.h +++ b/tensorflow/contrib/lite/kernels/internal/reference/reference_ops.h @@ -951,6 +951,19 @@ inline void Relu6(const float* input_data, const RuntimeShape& input_shape, } } +inline void ReluX(uint8 min_value, uint8 max_value, const uint8* input_data, + const RuntimeShape& input_shape, uint8* output_data, + const RuntimeShape& output_shape) { + gemmlowp::ScopedProfilingLabel label("Quantized ReluX (not fused)"); + const int flat_size = MatchingFlatSize(input_shape, output_shape); + for (int i = 0; i < flat_size; ++i) { + const uint8 val = input_data[i]; + const uint8 clamped = + val > max_value ? max_value : val < min_value ? min_value : val; + output_data[i] = clamped; + } +} + template void L2Normalization(const float* input_data, const RuntimeShape& input_shape, float* output_data, const RuntimeShape& output_shape) { @@ -1051,10 +1064,11 @@ inline void L2Normalization(const uint8* input_data, } } -inline void Add(const float* input1_data, const Dims<4>& input1_dims, - const float* input2_data, const Dims<4>& input2_dims, - float output_activation_min, float output_activation_max, - float* output_data, const Dims<4>& output_dims) { +template +inline void Add(const T* input1_data, const Dims<4>& input1_dims, + const T* input2_data, const Dims<4>& input2_dims, + T output_activation_min, T output_activation_max, + T* output_data, const Dims<4>& output_dims) { const int flat_size = MatchingFlatSize(input1_dims, input2_dims, output_dims); for (int i = 0; i < flat_size; ++i) { output_data[i] = ActivationFunctionWithMinMax( @@ -1415,10 +1429,11 @@ inline void BroadcastAddFivefold( output_activation_max, output_data, output_dims); } -inline void Mul(const float* input1_data, const Dims<4>& input1_dims, - const float* input2_data, const Dims<4>& input2_dims, - float output_activation_min, float output_activation_max, - float* output_data, const Dims<4>& output_dims) { +template +inline void Mul(const T* input1_data, const Dims<4>& input1_dims, + const T* input2_data, const Dims<4>& input2_dims, + T output_activation_min, T output_activation_max, + T* output_data, const Dims<4>& output_dims) { const int flat_size = MatchingFlatSize(input1_dims, input2_dims, output_dims); for (int i = 0; i < flat_size; ++i) { output_data[i] = ActivationFunctionWithMinMax( @@ -2259,13 +2274,10 @@ inline int NodeOffset(int b, int h, int w, int height, int width) { return (b * height + h) * width + w; } -inline void AveragePool(const float* input_data, - const RuntimeShape& input_shape, int stride_width, - int stride_height, int pad_width, int pad_height, - int filter_width, int filter_height, - float output_activation_min, - float output_activation_max, float* output_data, - const RuntimeShape& output_shape) { +inline void AveragePool(const PoolParams& params, + const RuntimeShape& input_shape, + const float* input_data, + const RuntimeShape& output_shape, float* output_data) { TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4); TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4); const int batches = MatchingDim(input_shape, 0, output_shape, 0); @@ -2274,20 +2286,24 @@ inline void AveragePool(const float* input_data, const int input_width = input_shape.Dims(2); const int output_height = output_shape.Dims(1); const int output_width = output_shape.Dims(2); + const int stride_height = params.stride_height; + const int stride_width = params.stride_width; for (int batch = 0; batch < batches; ++batch) { for (int out_y = 0; out_y < output_height; ++out_y) { for (int out_x = 0; out_x < output_width; ++out_x) { for (int channel = 0; channel < depth; ++channel) { - const int in_x_origin = (out_x * stride_width) - pad_width; - const int in_y_origin = (out_y * stride_height) - pad_height; + const int in_x_origin = + (out_x * stride_width) - params.padding_values.width; + const int in_y_origin = + (out_y * stride_height) - params.padding_values.height; // Compute the boundaries of the filter region clamped so as to // ensure that the filter window fits in the input array. const int filter_x_start = std::max(0, -in_x_origin); const int filter_x_end = - std::min(filter_width, input_width - in_x_origin); + std::min(params.filter_width, input_width - in_x_origin); const int filter_y_start = std::max(0, -in_y_origin); const int filter_y_end = - std::min(filter_height, input_height - in_y_origin); + std::min(params.filter_height, input_height - in_y_origin); float total = 0.f; float filter_count = 0; for (int filter_y = filter_y_start; filter_y < filter_y_end; @@ -2303,22 +2319,20 @@ inline void AveragePool(const float* input_data, } const float average = total / filter_count; output_data[Offset(output_shape, batch, out_y, out_x, channel)] = - ActivationFunctionWithMinMax(average, output_activation_min, - output_activation_max); + ActivationFunctionWithMinMax(average, params.float_activation_min, + params.float_activation_max); } } } } } -inline void AveragePool(const uint8* input_data, - const RuntimeShape& input_shape, int stride_width, - int stride_height, int pad_width, int pad_height, - int filter_width, int filter_height, - int32 output_activation_min, - int32 output_activation_max, uint8* output_data, - const RuntimeShape& output_shape) { - TFLITE_DCHECK_LE(output_activation_min, output_activation_max); +inline void AveragePool(const PoolParams& params, + const RuntimeShape& input_shape, + const uint8* input_data, + const RuntimeShape& output_shape, uint8* output_data) { + TFLITE_DCHECK_LE(params.quantized_activation_min, + params.quantized_activation_max); TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4); TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4); const int batches = MatchingDim(input_shape, 0, output_shape, 0); @@ -2327,20 +2341,24 @@ inline void AveragePool(const uint8* input_data, const int input_width = input_shape.Dims(2); const int output_height = output_shape.Dims(1); const int output_width = output_shape.Dims(2); + const int stride_height = params.stride_height; + const int stride_width = params.stride_width; for (int batch = 0; batch < batches; ++batch) { for (int out_y = 0; out_y < output_height; ++out_y) { for (int out_x = 0; out_x < output_width; ++out_x) { for (int channel = 0; channel < depth; ++channel) { - const int in_x_origin = (out_x * stride_width) - pad_width; - const int in_y_origin = (out_y * stride_height) - pad_height; + const int in_x_origin = + (out_x * stride_width) - params.padding_values.width; + const int in_y_origin = + (out_y * stride_height) - params.padding_values.height; // Compute the boundaries of the filter region clamped so as to // ensure that the filter window fits in the input array. const int filter_x_start = std::max(0, -in_x_origin); const int filter_x_end = - std::min(filter_width, input_width - in_x_origin); + std::min(params.filter_width, input_width - in_x_origin); const int filter_y_start = std::max(0, -in_y_origin); const int filter_y_end = - std::min(filter_height, input_height - in_y_origin); + std::min(params.filter_height, input_height - in_y_origin); int32 acc = 0; int filter_count = 0; for (int filter_y = filter_y_start; filter_y < filter_y_end; @@ -2355,8 +2373,8 @@ inline void AveragePool(const uint8* input_data, } } acc = (acc + filter_count / 2) / filter_count; - acc = std::max(acc, output_activation_min); - acc = std::min(acc, output_activation_max); + acc = std::max(acc, params.quantized_activation_min); + acc = std::min(acc, params.quantized_activation_max); output_data[Offset(output_shape, batch, out_y, out_x, channel)] = static_cast(acc); } @@ -2365,11 +2383,9 @@ inline void AveragePool(const uint8* input_data, } } -inline void L2Pool(const float* input_data, const RuntimeShape& input_shape, - int stride_width, int stride_height, int pad_width, - int pad_height, int filter_width, int filter_height, - float output_activation_min, float output_activation_max, - float* output_data, const RuntimeShape& output_shape) { +inline void L2Pool(const PoolParams& params, const RuntimeShape& input_shape, + const float* input_data, const RuntimeShape& output_shape, + float* output_data) { TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4); TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4); const int batches = MatchingDim(input_shape, 0, output_shape, 0); @@ -2378,20 +2394,24 @@ inline void L2Pool(const float* input_data, const RuntimeShape& input_shape, const int input_width = input_shape.Dims(2); const int output_height = output_shape.Dims(1); const int output_width = output_shape.Dims(2); + const int stride_height = params.stride_height; + const int stride_width = params.stride_width; for (int batch = 0; batch < batches; ++batch) { for (int out_y = 0; out_y < output_height; ++out_y) { for (int out_x = 0; out_x < output_width; ++out_x) { for (int channel = 0; channel < depth; ++channel) { - const int in_x_origin = (out_x * stride_width) - pad_width; - const int in_y_origin = (out_y * stride_height) - pad_height; + const int in_x_origin = + (out_x * stride_width) - params.padding_values.width; + const int in_y_origin = + (out_y * stride_height) - params.padding_values.height; // Compute the boundaries of the filter region clamped so as to // ensure that the filter window fits in the input array. const int filter_x_start = std::max(0, -in_x_origin); const int filter_x_end = - std::min(filter_width, input_width - in_x_origin); + std::min(params.filter_width, input_width - in_x_origin); const int filter_y_start = std::max(0, -in_y_origin); const int filter_y_end = - std::min(filter_height, input_height - in_y_origin); + std::min(params.filter_height, input_height - in_y_origin); float sum_squares = 0.f; int filter_count = 0; for (int filter_y = filter_y_start; filter_y < filter_y_end; @@ -2408,19 +2428,18 @@ inline void L2Pool(const float* input_data, const RuntimeShape& input_shape, } const float l2pool_result = std::sqrt(sum_squares / filter_count); output_data[Offset(output_shape, batch, out_y, out_x, channel)] = - ActivationFunctionWithMinMax(l2pool_result, output_activation_min, - output_activation_max); + ActivationFunctionWithMinMax(l2pool_result, + params.float_activation_min, + params.float_activation_max); } } } } } -inline void MaxPool(const float* input_data, const RuntimeShape& input_shape, - int stride_width, int stride_height, int pad_width, - int pad_height, int filter_width, int filter_height, - float output_activation_min, float output_activation_max, - float* output_data, const RuntimeShape& output_shape) { +inline void MaxPool(const PoolParams& params, const RuntimeShape& input_shape, + const float* input_data, const RuntimeShape& output_shape, + float* output_data) { TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4); TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4); const int batches = MatchingDim(input_shape, 0, output_shape, 0); @@ -2429,20 +2448,24 @@ inline void MaxPool(const float* input_data, const RuntimeShape& input_shape, const int input_width = input_shape.Dims(2); const int output_height = output_shape.Dims(1); const int output_width = output_shape.Dims(2); + const int stride_height = params.stride_height; + const int stride_width = params.stride_width; for (int batch = 0; batch < batches; ++batch) { for (int out_y = 0; out_y < output_height; ++out_y) { for (int out_x = 0; out_x < output_width; ++out_x) { for (int channel = 0; channel < depth; ++channel) { - const int in_x_origin = (out_x * stride_width) - pad_width; - const int in_y_origin = (out_y * stride_height) - pad_height; + const int in_x_origin = + (out_x * stride_width) - params.padding_values.width; + const int in_y_origin = + (out_y * stride_height) - params.padding_values.height; // Compute the boundaries of the filter region clamped so as to // ensure that the filter window fits in the input array. const int filter_x_start = std::max(0, -in_x_origin); const int filter_x_end = - std::min(filter_width, input_width - in_x_origin); + std::min(params.filter_width, input_width - in_x_origin); const int filter_y_start = std::max(0, -in_y_origin); const int filter_y_end = - std::min(filter_height, input_height - in_y_origin); + std::min(params.filter_height, input_height - in_y_origin); float max = std::numeric_limits::lowest(); for (int filter_y = filter_y_start; filter_y < filter_y_end; ++filter_y) { @@ -2456,22 +2479,21 @@ inline void MaxPool(const float* input_data, const RuntimeShape& input_shape, } } output_data[Offset(output_shape, batch, out_y, out_x, channel)] = - ActivationFunctionWithMinMax(max, output_activation_min, - output_activation_max); + ActivationFunctionWithMinMax(max, params.float_activation_min, + params.float_activation_max); } } } } } -inline void MaxPool(const uint8* input_data, const RuntimeShape& input_shape, - int stride_width, int stride_height, int pad_width, - int pad_height, int filter_width, int filter_height, - int32 output_activation_min, int32 output_activation_max, - uint8* output_data, const RuntimeShape& output_shape) { - TFLITE_DCHECK_LE(output_activation_min, output_activation_max); - TFLITE_DCHECK_GE(output_activation_min, 0); - TFLITE_DCHECK_LE(output_activation_max, 255); +inline void MaxPool(const PoolParams& params, const RuntimeShape& input_shape, + const uint8* input_data, const RuntimeShape& output_shape, + uint8* output_data) { + TFLITE_DCHECK_LE(params.quantized_activation_min, + params.quantized_activation_max); + TFLITE_DCHECK_GE(params.quantized_activation_min, 0); + TFLITE_DCHECK_LE(params.quantized_activation_max, 255); TFLITE_DCHECK_EQ(input_shape.DimensionsCount(), 4); TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 4); const int batches = MatchingDim(input_shape, 0, output_shape, 0); @@ -2480,20 +2502,24 @@ inline void MaxPool(const uint8* input_data, const RuntimeShape& input_shape, const int input_width = input_shape.Dims(2); const int output_height = output_shape.Dims(1); const int output_width = output_shape.Dims(2); + const int stride_height = params.stride_height; + const int stride_width = params.stride_width; for (int batch = 0; batch < batches; ++batch) { for (int out_y = 0; out_y < output_height; ++out_y) { for (int out_x = 0; out_x < output_width; ++out_x) { for (int channel = 0; channel < depth; ++channel) { - const int in_x_origin = (out_x * stride_width) - pad_width; - const int in_y_origin = (out_y * stride_height) - pad_height; + const int in_x_origin = + (out_x * stride_width) - params.padding_values.width; + const int in_y_origin = + (out_y * stride_height) - params.padding_values.height; // Compute the boundaries of the filter region clamped so as to // ensure that the filter window fits in the input array. const int filter_x_start = std::max(0, -in_x_origin); const int filter_x_end = - std::min(filter_width, input_width - in_x_origin); + std::min(params.filter_width, input_width - in_x_origin); const int filter_y_start = std::max(0, -in_y_origin); const int filter_y_end = - std::min(filter_height, input_height - in_y_origin); + std::min(params.filter_height, input_height - in_y_origin); uint8 max = 0; for (int filter_y = filter_y_start; filter_y < filter_y_end; ++filter_y) { @@ -2506,8 +2532,8 @@ inline void MaxPool(const uint8* input_data, const RuntimeShape& input_shape, input_data[Offset(input_shape, batch, in_y, in_x, channel)]); } } - max = std::max(max, output_activation_min); - max = std::min(max, output_activation_max); + max = std::max(max, params.quantized_activation_min); + max = std::min(max, params.quantized_activation_max); output_data[Offset(output_shape, batch, out_y, out_x, channel)] = static_cast(max); } @@ -3342,7 +3368,7 @@ inline void Pad(const T* input_data, const Dims<4>& input_dims, template inline void StridedSlice(const T* input_data, const Dims<4>& input_dims, - int begin_mask, int end_mask, + int begin_mask, int end_mask, int shrink_axis_mask, const std::vector& start_indices, const std::vector& stop_indices, const std::vector& strides, T* output_data, @@ -3354,20 +3380,24 @@ inline void StridedSlice(const T* input_data, const Dims<4>& input_dims, TFLITE_DCHECK_EQ(strides.size(), 4); const int start_b = strided_slice::StartForAxis(begin_mask, start_indices, strides, input_dims.sizes, 3); - const int stop_b = strided_slice::StopForAxis(end_mask, stop_indices, strides, - input_dims.sizes, 3); + const int stop_b = + strided_slice::StopForAxis(end_mask, shrink_axis_mask, stop_indices, + strides, input_dims.sizes, 3, start_b); const int start_h = strided_slice::StartForAxis(begin_mask, start_indices, strides, input_dims.sizes, 2); - const int stop_h = strided_slice::StopForAxis(end_mask, stop_indices, strides, - input_dims.sizes, 2); + const int stop_h = + strided_slice::StopForAxis(end_mask, shrink_axis_mask, stop_indices, + strides, input_dims.sizes, 2, start_h); const int start_w = strided_slice::StartForAxis(begin_mask, start_indices, strides, input_dims.sizes, 1); - const int stop_w = strided_slice::StopForAxis(end_mask, stop_indices, strides, - input_dims.sizes, 1); + const int stop_w = + strided_slice::StopForAxis(end_mask, shrink_axis_mask, stop_indices, + strides, input_dims.sizes, 1, start_w); const int start_d = strided_slice::StartForAxis(begin_mask, start_indices, strides, input_dims.sizes, 0); - const int stop_d = strided_slice::StopForAxis(end_mask, stop_indices, strides, - input_dims.sizes, 0); + const int stop_d = + strided_slice::StopForAxis(end_mask, shrink_axis_mask, stop_indices, + strides, input_dims.sizes, 0, start_d); T* out_ptr = output_data; for (int in_b = start_b; @@ -3699,9 +3729,9 @@ void TensorFlowMaximumMinimum(const T* input1_data, const Dims<4>& input1_dims, } } -template -void ArgMax(const T3* axis, const T1* input_data, const Dims<4>& input_dims, - T2* output_data, const Dims<4>& output_dims) { +template +void ArgMinMax(const T3* axis, const T1* input_data, const Dims<4>& input_dims, + T2* output_data, const Dims<4>& output_dims, const Cmp& cmp) { // The current ArgMax implemention can only determine the index of the maximum // value in the last dimension. So the axis argument is ignored. @@ -3714,19 +3744,28 @@ void ArgMax(const T3* axis, const T1* input_data, const Dims<4>& input_dims, const int depth = ArraySize(input_dims, 0); for (int i = 0; i < outer_size; ++i) { - auto max_value = input_data[i * depth]; - int max_index = 0; + auto min_max_value = input_data[i * depth]; + int min_max_index = 0; for (int d = 1; d < depth; ++d) { const auto& curr_value = input_data[i * depth + d]; - if (curr_value > max_value) { - max_value = curr_value; - max_index = d; + if (cmp(curr_value, min_max_value)) { + min_max_value = curr_value; + min_max_index = d; } } - output_data[i] = max_index; + output_data[i] = min_max_index; } } +// TODO(renjieliu): Remove this one. +template +void ArgMax(const T3* axis, const T1* input_data, + const tflite::Dims<4>& input_dims, T2* output_data, + const tflite::Dims<4>& output_dims) { + ArgMinMax(axis, input_data, input_dims, output_data, output_dims, + std::greater()); +} + template void Transpose(const T* input, const Dims<4>& input_dims, T* output, const Dims<4>& output_dims, const int* permuted_axes) { @@ -4069,6 +4108,36 @@ inline void SparseToDense(const std::vector>& indices, } } +template +inline void Pow(const T* input1_data, const Dims<4>& input1_dims, + const T* input2_data, const Dims<4>& input2_dims, + T* output_data, const Dims<4>& output_dims) { + const int flat_size = MatchingFlatSize(input1_dims, input2_dims, output_dims); + for (int i = 0; i < flat_size; ++i) { + output_data[i] = std::pow(input1_data[i], input2_data[i]); + } +} + +template +inline void BroadcastPow(const T* input1_data, const Dims<4>& input1_dims, + const T* input2_data, const Dims<4>& input2_dims, + T* output_data, const Dims<4>& output_dims) { + NdArrayDesc<4> desc1; + NdArrayDesc<4> desc2; + NdArrayDescsForElementwiseBroadcast(input1_dims, input2_dims, &desc1, &desc2); + for (int b = 0; b < ArraySize(output_dims, 3); ++b) { + for (int y = 0; y < ArraySize(output_dims, 2); ++y) { + for (int x = 0; x < ArraySize(output_dims, 1); ++x) { + for (int c = 0; c < ArraySize(output_dims, 0); ++c) { + output_data[Offset(output_dims, c, x, y, b)] = + std::pow(input1_data[SubscriptToIndex(desc1, c, x, y, b)], + input2_data[SubscriptToIndex(desc2, c, x, y, b)]); + } + } + } + } +} + } // namespace reference_ops } // namespace tflite diff --git a/tensorflow/contrib/lite/kernels/internal/strided_slice_logic.h b/tensorflow/contrib/lite/kernels/internal/strided_slice_logic.h index ef77371bf65cc975dfa35275c8daa32de112a249..5994fad5c73df1dde6e33ba46dbd6e0802ea61be 100644 --- a/tensorflow/contrib/lite/kernels/internal/strided_slice_logic.h +++ b/tensorflow/contrib/lite/kernels/internal/strided_slice_logic.h @@ -74,12 +74,22 @@ inline int StartForAxis(int begin_mask, // size 4, this function would return 4 as the stop, because it is one past the // "real" indices of 0, 1, 2 & 3. template -inline int StopForAxis(int end_mask, std::vector const& stop_indices, +inline int StopForAxis(int end_mask, int shrink_axis_mask, + std::vector const& stop_indices, std::vector const& strides, - int const* input_shape, int axis) { + int const* input_shape, int axis, int start_for_axis) { // Begin with the specified index + const bool shrink_axis = shrink_axis_mask & (1 << axis); int stop = stop_indices[axis]; + // When shrinking an axis, the end position does not matter (and can be + // incorrect when negative indexing is used, see Issue #19260). Always use + // start_for_axis + 1 to generate a length 1 slice, since start_for_axis has + // already been adjusted for negative indices. + if (shrink_axis) { + stop = start_for_axis + 1; + } + // end_mask override if (end_mask & (1 << axis)) { if (strides[axis] > 0) { @@ -93,7 +103,7 @@ inline int StopForAxis(int end_mask, std::vector const& stop_indices, } // Handle negative indices - int axis_size = input_shape[axis]; + const int axis_size = input_shape[axis]; if (stop < 0) { stop += axis_size; } diff --git a/tensorflow/contrib/lite/kernels/internal/tensor.h b/tensorflow/contrib/lite/kernels/internal/tensor.h index 518bee1c6369d3ce93d1b98e19dba7615b5844dc..ee2af5b46046c9e8bdc5816d5b6e9e9100cdc240 100644 --- a/tensorflow/contrib/lite/kernels/internal/tensor.h +++ b/tensorflow/contrib/lite/kernels/internal/tensor.h @@ -15,6 +15,7 @@ limitations under the License. #ifndef TENSORFLOW_CONTRIB_LITE_KERNELS_INTERNAL_TENSOR_H_ #define TENSORFLOW_CONTRIB_LITE_KERNELS_INTERNAL_TENSOR_H_ +#include #include #include "tensorflow/contrib/lite/context.h" #include "tensorflow/contrib/lite/kernels/internal/types.h" @@ -54,6 +55,13 @@ inline bool* GetTensorData(TfLiteTensor* tensor) { return tensor != nullptr ? tensor->data.b : nullptr; } +template <> +inline std::complex* GetTensorData(TfLiteTensor* tensor) { + return tensor != nullptr + ? reinterpret_cast*>(tensor->data.c64) + : nullptr; +} + template inline const T* GetTensorData(const TfLiteTensor* tensor); @@ -87,6 +95,13 @@ inline const bool* GetTensorData(const TfLiteTensor* tensor) { return tensor != nullptr ? tensor->data.b : nullptr; } +template <> +inline const std::complex* GetTensorData(const TfLiteTensor* tensor) { + return tensor != nullptr + ? reinterpret_cast*>(tensor->data.c64) + : nullptr; +} + inline int RemapDim(int max_dimensions, int d) { return max_dimensions - d - 1; } diff --git a/tensorflow/contrib/lite/kernels/internal/tensor_utils_test.cc b/tensorflow/contrib/lite/kernels/internal/tensor_utils_test.cc index 14ee528394b6872d9e79969db0e431658277f56b..aa0d49ae4db6b4952b5864166f4a13459763cf44 100644 --- a/tensorflow/contrib/lite/kernels/internal/tensor_utils_test.cc +++ b/tensorflow/contrib/lite/kernels/internal/tensor_utils_test.cc @@ -63,7 +63,8 @@ TEST(uKernels, SymmetricQuantizeFloatsTest) { EXPECT_EQ(min, -640); EXPECT_EQ(max, 1000); - EXPECT_NEAR(scaling_factor, 0.127, 1e-6); // EQ won't work due to fpoint. + // EQ won't work due to fpoint. + EXPECT_NEAR(scaling_factor, 1000 / 127.0, 1e-6); EXPECT_THAT(output, testing::ElementsAreArray({-81, -81, -80, 1, 0, -1, -1, 0, 127})); } @@ -95,7 +96,7 @@ TEST(uKernels, SymmetricQuantizeFloatsAllAlmostZeroTest) { EXPECT_NEAR(min, -9e-05, 1e-6); EXPECT_NEAR(max, 0.0002, 1e-6); - EXPECT_EQ(scaling_factor, 635000); + EXPECT_NEAR(scaling_factor, 1.57e-6, 1e-6); EXPECT_THAT(output, testing::ElementsAreArray({-6, 19, -4, -57, 1, 25, 6, 127, 0})); } diff --git a/tensorflow/contrib/lite/kernels/internal/types.h b/tensorflow/contrib/lite/kernels/internal/types.h index fa2420713fea4faa3596251a95c2ed9606878b98..737cfb69c9e2ac48e87a6aa77e7f51a2098f8c41 100644 --- a/tensorflow/contrib/lite/kernels/internal/types.h +++ b/tensorflow/contrib/lite/kernels/internal/types.h @@ -23,7 +23,12 @@ limitations under the License. namespace tflite { enum class FusedActivationFunctionType : uint8 { kNone, kRelu6, kRelu1, kRelu }; -enum class PaddingType { kNone, kSame, kValid }; +enum class PaddingType : uint8 { kNone, kSame, kValid }; + +struct PaddingValues { + int8 width; + int8 height; +}; // This enumeration allows for non-default formats for the weights array // of a fully-connected operator, allowing the use of special optimized @@ -588,6 +593,22 @@ void ComputeStrides(Dims* dims) { } } +struct PoolParams { + FusedActivationFunctionType activation; + PaddingType padding_type; + PaddingValues padding_values; + int stride_height; + int stride_width; + int filter_height; + int filter_width; + // uint8, etc, inference params. + int32 quantized_activation_min; + int32 quantized_activation_max; + // float inference params. + float float_activation_min; + float float_activation_max; +}; + } // namespace tflite #endif // TENSORFLOW_CONTRIB_LITE_KERNELS_INTERNAL_TYPES_H_ diff --git a/tensorflow/contrib/lite/kernels/kernel_util.cc b/tensorflow/contrib/lite/kernels/kernel_util.cc index fdf9856912b9a0f4b6acf81db6ecb5f9c9385f0b..08f942c933552aa6ca7369550c928efba9e2e93e 100644 --- a/tensorflow/contrib/lite/kernels/kernel_util.cc +++ b/tensorflow/contrib/lite/kernels/kernel_util.cc @@ -103,24 +103,6 @@ void CalculateActivationRangeUint8(TfLiteFusedActivation activation, act_max); } -void CalculateActivationRangeFloat(TfLiteFusedActivation activation, - float* activation_min, - float* activation_max) { - if (activation == kTfLiteActRelu) { - *activation_min = 0.f; - *activation_max = std::numeric_limits::max(); - } else if (activation == kTfLiteActRelu6) { - *activation_min = 0.f; - *activation_max = 6.f; - } else if (activation == kTfLiteActRelu1) { - *activation_min = -1.f; - *activation_max = 1.f; - } else { - *activation_min = std::numeric_limits::lowest(); - *activation_max = std::numeric_limits::max(); - } -} - bool HaveSameShapes(const TfLiteTensor* input1, const TfLiteTensor* input2) { return TfLiteIntArrayEqual(input1->dims, input2->dims); } diff --git a/tensorflow/contrib/lite/kernels/kernel_util.h b/tensorflow/contrib/lite/kernels/kernel_util.h index 20058a5f6971ffc6c1763b0e98cd4a91fe7b6e44..c8ce3c917d5bf66e01fbae95c18dfe97b3c84bae 100644 --- a/tensorflow/contrib/lite/kernels/kernel_util.h +++ b/tensorflow/contrib/lite/kernels/kernel_util.h @@ -15,6 +15,8 @@ limitations under the License. #ifndef TENSORFLOW_CONTRIB_LITE_KERNELS_KERNEL_UTIL_H_ #define TENSORFLOW_CONTRIB_LITE_KERNELS_KERNEL_UTIL_H_ +#include + #include "tensorflow/contrib/lite/builtin_op_data.h" #include "tensorflow/contrib/lite/context.h" @@ -86,8 +88,8 @@ TfLiteStatus GetQuantizedConvolutionMultipler(TfLiteContext* context, TfLiteTensor* output, double* multiplier); -// Calculates the useful range of an activation layer given its activation -// tensor. +// Calculates the useful quantized range of an activation layer given its +// activation tensor. TfLiteStatus CalculateActivationRangeQuantized(TfLiteContext* context, TfLiteFusedActivation activation, TfLiteTensor* output, @@ -96,9 +98,25 @@ TfLiteStatus CalculateActivationRangeQuantized(TfLiteContext* context, void CalculateActivationRangeUint8(TfLiteFusedActivation activation, TfLiteTensor* output, int32_t* act_min, int32_t* act_max); -void CalculateActivationRangeFloat(TfLiteFusedActivation activation, - float* activation_min, - float* activation_max); +// Calculates the useful range of an activation layer given its activation +// tensor.a +template +void CalculateActivationRange(TfLiteFusedActivation activation, + T* activation_min, T* activation_max) { + if (activation == kTfLiteActRelu) { + *activation_min = 0; + *activation_max = std::numeric_limits::max(); + } else if (activation == kTfLiteActRelu6) { + *activation_min = 0; + *activation_max = 6; + } else if (activation == kTfLiteActRelu1) { + *activation_min = -1; + *activation_max = 1; + } else { + *activation_min = std::numeric_limits::lowest(); + *activation_max = std::numeric_limits::max(); + } +} // Return true if the given tensors have the same shape. bool HaveSameShapes(const TfLiteTensor* input1, const TfLiteTensor* input2); diff --git a/tensorflow/contrib/lite/kernels/lstm.cc b/tensorflow/contrib/lite/kernels/lstm.cc index 3577ae6caa1e02ce2e5db2e8054ba9c2fccbe93e..4dfc8915489f9dcf243b13bc10afcef278779a93 100644 --- a/tensorflow/contrib/lite/kernels/lstm.cc +++ b/tensorflow/contrib/lite/kernels/lstm.cc @@ -306,7 +306,8 @@ TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) { const int n_output = recurrent_to_output_weights->dims->data[1]; // Check that input tensor dimensions matches with each other. - CheckInputTensorDimensions(context, node, n_input, n_output, n_cell); + TF_LITE_ENSURE_OK(context, CheckInputTensorDimensions(context, node, n_input, + n_output, n_cell)); // Get the pointer to output, activation_state and cell_state tensors. TfLiteTensor* output = GetOutput(context, node, kOutputTensor); diff --git a/tensorflow/contrib/lite/kernels/lstm_test.cc b/tensorflow/contrib/lite/kernels/lstm_test.cc index 3f5c44a63ec328b23e11dc42428f7cd85a788509..0266f5fe57e6c60ea19ad5f8de05e879e7da9304 100644 --- a/tensorflow/contrib/lite/kernels/lstm_test.cc +++ b/tensorflow/contrib/lite/kernels/lstm_test.cc @@ -13,6 +13,9 @@ See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ // Unit test for TFLite LSTM op. +// +// TODO(alanchiao): add unit test with invalid input dimensions for this and its +// variants. #include #include @@ -360,14 +363,6 @@ class BaseLstmTest : public ::testing::Test { } EXPECT_THAT(lstm->GetOutput(), ElementsAreArray(ArrayFloatNear(expected, tolerance))); - for (int i = 0; i < num_outputs; ++i) { - std::cout << lstm->GetOutput()[i] << ", "; - } - std::cout << std::endl; - for (int i = 0; i < num_outputs; ++i) { - std::cout << expected[i] << ", "; - } - std::cout << std::endl; } } }; diff --git a/tensorflow/contrib/lite/kernels/mul.cc b/tensorflow/contrib/lite/kernels/mul.cc index 9e01b73c4933c34ce3fd549730080946674daaac..349f3e672611b76ba9eb0019bbd55a5881ed6535 100644 --- a/tensorflow/contrib/lite/kernels/mul.cc +++ b/tensorflow/contrib/lite/kernels/mul.cc @@ -100,29 +100,44 @@ TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) { } template -void EvalFloat(TfLiteContext* context, TfLiteNode* node, - TfLiteMulParams* params, const OpData* data, - const TfLiteTensor* input1, const TfLiteTensor* input2, - TfLiteTensor* output) { - float output_activation_min, output_activation_max; - CalculateActivationRangeFloat(params->activation, &output_activation_min, - &output_activation_max); -#define TF_LITE_MUL(type, opname) \ - type::opname(GetTensorData(input1), GetTensorDims(input1), \ - GetTensorData(input2), GetTensorDims(input2), \ - output_activation_min, output_activation_max, \ - GetTensorData(output), GetTensorDims(output)) - if (kernel_type == kReference) { - if (data->requires_broadcast) { - TF_LITE_MUL(reference_ops, BroadcastMul); +void EvalMul(TfLiteContext* context, TfLiteNode* node, TfLiteMulParams* params, + const OpData* data, const TfLiteTensor* input1, + const TfLiteTensor* input2, TfLiteTensor* output) { +#define TF_LITE_MUL(type, opname, data_type) \ + data_type output_activation_min, output_activation_max; \ + CalculateActivationRange(params->activation, &output_activation_min, \ + &output_activation_max); \ + type::opname(GetTensorData(input1), GetTensorDims(input1), \ + GetTensorData(input2), GetTensorDims(input2), \ + output_activation_min, output_activation_max, \ + GetTensorData(output), GetTensorDims(output)) + if (output->type == kTfLiteInt32) { + if (kernel_type == kReference) { + if (data->requires_broadcast) { + TF_LITE_MUL(reference_ops, BroadcastMul, int32_t); + } else { + TF_LITE_MUL(reference_ops, Mul, int32_t); + } } else { - TF_LITE_MUL(reference_ops, Mul); + if (data->requires_broadcast) { + TF_LITE_MUL(optimized_ops, BroadcastMul, int32_t); + } else { + TF_LITE_MUL(optimized_ops, Mul, int32_t); + } } - } else { - if (data->requires_broadcast) { - TF_LITE_MUL(optimized_ops, BroadcastMul); + } else if (output->type == kTfLiteFloat32) { + if (kernel_type == kReference) { + if (data->requires_broadcast) { + TF_LITE_MUL(reference_ops, BroadcastMul, float); + } else { + TF_LITE_MUL(reference_ops, Mul, float); + } } else { - TF_LITE_MUL(optimized_ops, Mul); + if (data->requires_broadcast) { + TF_LITE_MUL(optimized_ops, BroadcastMul, float); + } else { + TF_LITE_MUL(optimized_ops, Mul, float); + } } } #undef TF_LITE_MUL @@ -194,17 +209,17 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { const TfLiteTensor* input2 = GetInput(context, node, kInputTensor2); TfLiteTensor* output = GetOutput(context, node, kOutputTensor); - if (output->type == kTfLiteFloat32) { - EvalFloat(context, node, params, data, input1, input2, output); + if (output->type == kTfLiteFloat32 || output->type == kTfLiteInt32) { + EvalMul(context, node, params, data, input1, input2, output); } else if (output->type == kTfLiteUInt8 || output->type == kTfLiteInt16) { TF_LITE_ENSURE_OK( context, EvalQuantized(context, node, params, data, input1, input2, output)); } else { - context->ReportError( - context, - "Mul only supports FLOAT32 and quantized UINT8 and INT16 now, got %d.", - output->type); + context->ReportError(context, + "Mul only supports FLOAT32, INT32 and quantized UINT8 " + "and INT16 now, got %d.", + output->type); return kTfLiteError; } diff --git a/tensorflow/contrib/lite/kernels/mul_test.cc b/tensorflow/contrib/lite/kernels/mul_test.cc index 43d56e50d2686ff2624f36a0c5d8e43279a572cc..2807550a6b07f3f9f1f1e3f72acc9882c76d166a 100644 --- a/tensorflow/contrib/lite/kernels/mul_test.cc +++ b/tensorflow/contrib/lite/kernels/mul_test.cc @@ -52,6 +52,13 @@ class FloatMulOpModel : public BaseMulOpModel { std::vector GetOutput() { return ExtractVector(output_); } }; +class IntegerMulOpModel : public BaseMulOpModel { + public: + using BaseMulOpModel::BaseMulOpModel; + + std::vector GetOutput() { return ExtractVector(output_); } +}; + // For quantized Mul, the error shouldn't exceed (2*step + step^2). // The param min=-1.0 & max=1.0 is used in the following tests. // The tolerance value is ~0.0157. @@ -133,6 +140,57 @@ TEST(FloatMulOpTest, WithBroadcast) { } } +TEST(IntegerMulOpTest, NoActivation) { + IntegerMulOpModel m({TensorType_INT32, {1, 2, 2, 1}}, + {TensorType_INT32, {1, 2, 2, 1}}, {TensorType_INT32, {}}, + ActivationFunctionType_NONE); + m.PopulateTensor(m.input1(), {-20, 2, 7, 8}); + m.PopulateTensor(m.input2(), {1, 2, 3, 5}); + m.Invoke(); + EXPECT_THAT(m.GetOutput(), ElementsAreArray({-20, 4, 21, 40})); +} + +TEST(IntegerMulOpTest, ActivationRELU_N1_TO_1) { + IntegerMulOpModel m({TensorType_INT32, {1, 2, 2, 1}}, + {TensorType_INT32, {1, 2, 2, 1}}, {TensorType_INT32, {}}, + ActivationFunctionType_RELU_N1_TO_1); + m.PopulateTensor(m.input1(), {-20, 2, 7, 8}); + m.PopulateTensor(m.input2(), {1, 2, 3, 5}); + m.Invoke(); + EXPECT_THAT(m.GetOutput(), ElementsAreArray({-1, 1, 1, 1})); +} + +TEST(IntegerMulOpTest, VariousInputShapes) { + std::vector> test_shapes = { + {6}, {2, 3}, {2, 1, 3}, {1, 3, 1, 2}}; + for (int i = 0; i < test_shapes.size(); ++i) { + IntegerMulOpModel m({TensorType_INT32, test_shapes[i]}, + {TensorType_INT32, test_shapes[i]}, + {TensorType_INT32, {}}, ActivationFunctionType_NONE); + m.PopulateTensor(m.input1(), {-20, 2, 7, 8, 11, 20}); + m.PopulateTensor(m.input2(), {1, 2, 3, 5, 11, 1}); + m.Invoke(); + EXPECT_THAT(m.GetOutput(), ElementsAreArray({-20, 4, 21, 40, 121, 20})) + << "With shape number " << i; + } +} + +TEST(IntegerMulOpTest, WithBroadcast) { + std::vector> test_shapes = { + {6}, {2, 3}, {2, 1, 3}, {1, 3, 1, 2}}; + for (int i = 0; i < test_shapes.size(); ++i) { + IntegerMulOpModel m({TensorType_INT32, test_shapes[i]}, + {TensorType_INT32, {}}, // always a scalar + {TensorType_INT32, {}}, ActivationFunctionType_NONE); + m.PopulateTensor(m.input1(), {-20, 2, 7, 8, 11, 20}); + m.PopulateTensor(m.input2(), {1}); + m.Invoke(); + EXPECT_THAT(m.GetOutput(), + ElementsAreArray(ArrayFloatNear({-20, 2, 7, 8, 11, 20}))) + << "With shape number " << i; + } +} + TEST(QuantizedMulOpTest, NoActivation) { QuantizedMulOpModel m({TensorType_UINT8, {1, 2, 2, 1}, -1.0, 1.0}, {TensorType_UINT8, {1, 2, 2, 1}, -1.0, 1.0}, diff --git a/tensorflow/contrib/lite/kernels/pooling.cc b/tensorflow/contrib/lite/kernels/pooling.cc index 58d74c97a7a76c808c83ed51c8b984c103ce4cf0..9b0487ae16e6546ac39b5d75ecf9d14ae4e4e4cf 100644 --- a/tensorflow/contrib/lite/kernels/pooling.cc +++ b/tensorflow/contrib/lite/kernels/pooling.cc @@ -124,15 +124,21 @@ void AverageEvalFloat(TfLiteContext* context, TfLiteNode* node, TfLitePoolParams* params, OpData* data, const TfLiteTensor* input, TfLiteTensor* output) { float activation_min, activation_max; - CalculateActivationRangeFloat(params->activation, &activation_min, - &activation_max); -#define TF_LITE_AVERAGE_POOL(type) \ - type::AveragePool(GetTensorData(input), GetTensorShape(input), \ - params->stride_width, params->stride_height, \ - data->padding.width, data->padding.height, \ - params->filter_width, params->filter_height, \ - activation_min, activation_max, \ - GetTensorData(output), GetTensorShape(output)) + CalculateActivationRange(params->activation, &activation_min, + &activation_max); +#define TF_LITE_AVERAGE_POOL(type) \ + tflite::PoolParams op_params; \ + op_params.stride_height = params->stride_height; \ + op_params.stride_width = params->stride_width; \ + op_params.filter_height = params->filter_height; \ + op_params.filter_width = params->filter_width; \ + op_params.padding_values.height = data->padding.height; \ + op_params.padding_values.width = data->padding.width; \ + op_params.float_activation_min = activation_min; \ + op_params.float_activation_max = activation_max; \ + type::AveragePool(op_params, GetTensorShape(input), \ + GetTensorData(input), GetTensorShape(output), \ + GetTensorData(output)) if (kernel_type == kReference) { TF_LITE_AVERAGE_POOL(reference_ops); } else { @@ -149,13 +155,19 @@ void AverageEvalQuantized(TfLiteContext* context, TfLiteNode* node, int32_t activation_max; CalculateActivationRangeUint8(params->activation, output, &activation_min, &activation_max); -#define TF_LITE_AVERAGE_POOL(type) \ - type::AveragePool(GetTensorData(input), GetTensorShape(input), \ - params->stride_width, params->stride_height, \ - data->padding.width, data->padding.height, \ - params->filter_width, params->filter_height, \ - activation_min, activation_max, \ - GetTensorData(output), GetTensorShape(output)) +#define TF_LITE_AVERAGE_POOL(type) \ + tflite::PoolParams op_params; \ + op_params.stride_height = params->stride_height; \ + op_params.stride_width = params->stride_width; \ + op_params.filter_height = params->filter_height; \ + op_params.filter_width = params->filter_width; \ + op_params.padding_values.height = data->padding.height; \ + op_params.padding_values.width = data->padding.width; \ + op_params.quantized_activation_min = activation_min; \ + op_params.quantized_activation_max = activation_max; \ + type::AveragePool(op_params, GetTensorShape(input), \ + GetTensorData(input), GetTensorShape(output), \ + GetTensorData(output)) if (kernel_type == kReference) { TF_LITE_AVERAGE_POOL(reference_ops); } else { @@ -169,15 +181,20 @@ void MaxEvalFloat(TfLiteContext* context, TfLiteNode* node, TfLitePoolParams* params, OpData* data, const TfLiteTensor* input, TfLiteTensor* output) { float activation_min, activation_max; - CalculateActivationRangeFloat(params->activation, &activation_min, - &activation_max); -#define TF_LITE_MAX_POOL(type) \ - type::MaxPool(GetTensorData(input), GetTensorShape(input), \ - params->stride_width, params->stride_height, \ - data->padding.width, data->padding.height, \ - params->filter_width, params->filter_height, activation_min, \ - activation_max, GetTensorData(output), \ - GetTensorShape(output)) + CalculateActivationRange(params->activation, &activation_min, + &activation_max); +#define TF_LITE_MAX_POOL(type) \ + tflite::PoolParams op_params; \ + op_params.stride_height = params->stride_height; \ + op_params.stride_width = params->stride_width; \ + op_params.filter_height = params->filter_height; \ + op_params.filter_width = params->filter_width; \ + op_params.padding_values.height = data->padding.height; \ + op_params.padding_values.width = data->padding.width; \ + op_params.float_activation_min = activation_min; \ + op_params.float_activation_max = activation_max; \ + type::MaxPool(op_params, GetTensorShape(input), GetTensorData(input), \ + GetTensorShape(output), GetTensorData(output)) if (kernel_type == kReference) { TF_LITE_MAX_POOL(reference_ops); } else { @@ -194,13 +211,19 @@ void MaxEvalQuantized(TfLiteContext* context, TfLiteNode* node, int32_t activation_max; CalculateActivationRangeUint8(params->activation, output, &activation_min, &activation_max); -#define TF_LITE_MAX_POOL(type) \ - type::MaxPool(GetTensorData(input), GetTensorShape(input), \ - params->stride_width, params->stride_height, \ - data->padding.width, data->padding.height, \ - params->filter_width, params->filter_height, activation_min, \ - activation_max, GetTensorData(output), \ - GetTensorShape(output)) +#define TF_LITE_MAX_POOL(type) \ + tflite::PoolParams op_params; \ + op_params.stride_height = params->stride_height; \ + op_params.stride_width = params->stride_width; \ + op_params.filter_height = params->filter_height; \ + op_params.filter_width = params->filter_width; \ + op_params.padding_values.height = data->padding.height; \ + op_params.padding_values.width = data->padding.width; \ + op_params.quantized_activation_min = activation_min; \ + op_params.quantized_activation_max = activation_max; \ + type::MaxPool(op_params, GetTensorShape(input), \ + GetTensorData(input), GetTensorShape(output), \ + GetTensorData(output)) if (kernel_type == kReference) { TF_LITE_MAX_POOL(reference_ops); } else { @@ -214,15 +237,20 @@ void L2EvalFloat(TfLiteContext* context, TfLiteNode* node, TfLitePoolParams* params, OpData* data, const TfLiteTensor* input, TfLiteTensor* output) { float activation_min, activation_max; - CalculateActivationRangeFloat(params->activation, &activation_min, - &activation_max); -#define TF_LITE_L2_POOL(type) \ - type::L2Pool(GetTensorData(input), GetTensorShape(input), \ - params->stride_width, params->stride_height, \ - data->padding.width, data->padding.height, \ - params->filter_width, params->filter_height, activation_min, \ - activation_max, GetTensorData(output), \ - GetTensorShape(output)) + CalculateActivationRange(params->activation, &activation_min, + &activation_max); +#define TF_LITE_L2_POOL(type) \ + tflite::PoolParams op_params; \ + op_params.stride_height = params->stride_height; \ + op_params.stride_width = params->stride_width; \ + op_params.filter_height = params->filter_height; \ + op_params.filter_width = params->filter_width; \ + op_params.padding_values.height = data->padding.height; \ + op_params.padding_values.width = data->padding.width; \ + op_params.float_activation_min = activation_min; \ + op_params.float_activation_max = activation_max; \ + type::L2Pool(op_params, GetTensorShape(input), GetTensorData(input), \ + GetTensorShape(output), GetTensorData(output)) if (kernel_type == kReference) { TF_LITE_L2_POOL(reference_ops); } else { diff --git a/tensorflow/contrib/lite/kernels/pow.cc b/tensorflow/contrib/lite/kernels/pow.cc new file mode 100644 index 0000000000000000000000000000000000000000..4a539c47a8fbe392e0e6542ab8ffb9065b550485 --- /dev/null +++ b/tensorflow/contrib/lite/kernels/pow.cc @@ -0,0 +1,143 @@ +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#include "tensorflow/contrib/lite/context.h" +#include "tensorflow/contrib/lite/kernels/internal/reference/reference_ops.h" +#include "tensorflow/contrib/lite/kernels/internal/tensor.h" +#include "tensorflow/contrib/lite/kernels/kernel_util.h" +#include "tensorflow/contrib/lite/kernels/op_macros.h" + +namespace tflite { +namespace ops { +namespace builtin { +namespace pow { +namespace { + +// Input/output tensor index. +constexpr int kInputTensor1 = 0; +constexpr int kInputTensor2 = 1; +constexpr int kOutputTensor = 0; + +// Op data for pow op. +struct OpData { + bool requires_broadcast; +}; + +void* Init(TfLiteContext* context, const char* buffer, size_t length) { + auto* data = new OpData; + data->requires_broadcast = false; + return data; +} + +void Free(TfLiteContext* context, void* buffer) { + delete reinterpret_cast(buffer); +} + +TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) { + TF_LITE_ENSURE_EQ(context, NumInputs(node), 2); + TF_LITE_ENSURE_EQ(context, NumOutputs(node), 1); + + OpData* data = reinterpret_cast(node->user_data); + + const TfLiteTensor* input1 = GetInput(context, node, kInputTensor1); + const TfLiteTensor* input2 = GetInput(context, node, kInputTensor2); + TfLiteTensor* output = GetOutput(context, node, kOutputTensor); + + TF_LITE_ENSURE_EQ(context, input1->type, input2->type); + + const TfLiteType type = input1->type; + if (type != kTfLiteInt32 && type != kTfLiteFloat32) { + context->ReportError(context, "Unsupported data type %d.", type); + return kTfLiteError; + } + output->type = type; + + data->requires_broadcast = !HaveSameShapes(input1, input2); + + TfLiteIntArray* output_size = nullptr; + if (data->requires_broadcast) { + TF_LITE_ENSURE_OK(context, CalculateShapeForBroadcast( + context, input1, input2, &output_size)); + } else { + output_size = TfLiteIntArrayCopy(input1->dims); + } + + return context->ResizeTensor(context, output, output_size); +} + +template +void PowImpl(const TfLiteTensor* input1, const TfLiteTensor* input2, + TfLiteTensor* output, bool requires_broadcast) { + if (requires_broadcast) { + reference_ops::BroadcastPow(GetTensorData(input1), GetTensorDims(input1), + GetTensorData(input2), GetTensorDims(input2), + GetTensorData(output), + GetTensorDims(output)); + } else { + reference_ops::Pow(GetTensorData(input1), GetTensorDims(input1), + GetTensorData(input2), GetTensorDims(input2), + GetTensorData(output), GetTensorDims(output)); + } +} + +TfLiteStatus CheckValue(TfLiteContext* context, const TfLiteTensor* input) { + const int64_t num_elements = NumElements(input); + const int32_t* data = GetTensorData(input); + for (int i = 0; i < num_elements; ++i) { + if (data[i] < 0) { + context->ReportError(context, + "POW does not support negative value for int32."); + return kTfLiteError; + } + } + return kTfLiteOk; +} + +TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { + OpData* data = reinterpret_cast(node->user_data); + + const TfLiteTensor* input1 = GetInput(context, node, kInputTensor1); + const TfLiteTensor* input2 = GetInput(context, node, kInputTensor2); + TfLiteTensor* output = GetOutput(context, node, kOutputTensor); + + switch (output->type) { + case kTfLiteInt32: { + // TensorFlow does not support negative for int32. + TF_LITE_ENSURE_OK(context, CheckValue(context, input2)); + PowImpl(input1, input2, output, data->requires_broadcast); + break; + } + case kTfLiteFloat32: { + PowImpl(input1, input2, output, data->requires_broadcast); + break; + } + default: { + context->ReportError(context, "Unsupported data type: %d", output->type); + return kTfLiteError; + } + } + return kTfLiteOk; +} + +} // namespace +} // namespace pow + +TfLiteRegistration* Register_POW() { + static TfLiteRegistration r = {pow::Init, pow::Free, pow::Prepare, pow::Eval}; + return &r; +} + +} // namespace builtin +} // namespace ops +} // namespace tflite diff --git a/tensorflow/contrib/lite/kernels/pow_test.cc b/tensorflow/contrib/lite/kernels/pow_test.cc new file mode 100644 index 0000000000000000000000000000000000000000..474d323bc3a1a0f224aa0575a5bbd35394aa2f53 --- /dev/null +++ b/tensorflow/contrib/lite/kernels/pow_test.cc @@ -0,0 +1,117 @@ +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#include +#include "tensorflow/contrib/lite/interpreter.h" +#include "tensorflow/contrib/lite/kernels/register.h" +#include "tensorflow/contrib/lite/kernels/test_util.h" +#include "tensorflow/contrib/lite/model.h" + +namespace tflite { +namespace { + +using ::testing::ElementsAre; +using ::testing::ElementsAreArray; + +template +class PowOpModel : public SingleOpModel { + public: + PowOpModel(const TensorData& input1, const TensorData& input2, + const TensorData& output) { + input1_ = AddInput(input1); + input2_ = AddInput(input2); + output_ = AddOutput(output); + SetBuiltinOp(BuiltinOperator_POW, BuiltinOptions_PowOptions, + CreatePowOptions(builder_).Union()); + BuildInterpreter({GetShape(input1_), GetShape(input2_)}); + } + + int input1() { return input1_; } + int input2() { return input2_; } + + std::vector GetOutput() { return ExtractVector(output_); } + std::vector GetOutputShape() { return GetTensorShape(output_); } + + private: + int input1_; + int input2_; + int output_; +}; + +TEST(PowOpModel, Simple) { + PowOpModel model({TensorType_INT32, {1, 2, 2, 1}}, + {TensorType_INT32, {1, 2, 2, 1}}, + {TensorType_INT32, {}}); + model.PopulateTensor(model.input1(), {12, 2, 7, 8}); + model.PopulateTensor(model.input2(), {1, 2, 3, 1}); + model.Invoke(); + EXPECT_THAT(model.GetOutputShape(), ElementsAre(1, 2, 2, 1)); + EXPECT_THAT(model.GetOutput(), ElementsAre(12, 4, 343, 8)); +} + +TEST(PowOpModel, NegativeAndZeroValue) { + PowOpModel model({TensorType_INT32, {1, 2, 2, 1}}, + {TensorType_INT32, {1, 2, 2, 1}}, + {TensorType_INT32, {}}); + model.PopulateTensor(model.input1(), {0, 2, -7, 8}); + model.PopulateTensor(model.input2(), {1, 2, 3, 0}); + model.Invoke(); + EXPECT_THAT(model.GetOutputShape(), ElementsAre(1, 2, 2, 1)); + EXPECT_THAT(model.GetOutput(), ElementsAre(0, 4, -343, 1)); +} + +TEST(PowOpModel, Float) { + PowOpModel model({TensorType_FLOAT32, {1, 2, 2, 1}}, + {TensorType_FLOAT32, {1, 2, 2, 1}}, + {TensorType_FLOAT32, {}}); + model.PopulateTensor(model.input1(), {0.3, 0.4, 0.7, 5.8}); + model.PopulateTensor(model.input2(), {0.5, 2.7, 3.1, 3.2}); + model.Invoke(); + EXPECT_THAT(model.GetOutputShape(), ElementsAre(1, 2, 2, 1)); + EXPECT_THAT(model.GetOutput(), + ElementsAreArray(ArrayFloatNear( + {0.5477226, 0.08424846, 0.33098164, 277.313}, 1e-3))); +} + +TEST(PowOpModel, NegativeFloatTest) { + PowOpModel model({TensorType_FLOAT32, {1, 2, 2, 1}}, + {TensorType_FLOAT32, {1, 2, 2, 1}}, + {TensorType_FLOAT32, {}}); + model.PopulateTensor(model.input1(), {0.3, 0.4, 0.7, 5.8}); + model.PopulateTensor(model.input2(), {0.5, -2.7, 3.1, -3.2}); + model.Invoke(); + EXPECT_THAT(model.GetOutputShape(), ElementsAre(1, 2, 2, 1)); + EXPECT_THAT(model.GetOutput(), + ElementsAreArray(ArrayFloatNear( + {0.5477226, 11.869653, 0.33098164, 0.003606}, 1e-3))); +} + +TEST(PowOpModel, BroadcastTest) { + PowOpModel model({TensorType_INT32, {1, 2, 2, 1}}, + {TensorType_INT32, {1}}, {TensorType_INT32, {}}); + model.PopulateTensor(model.input1(), {12, 2, 7, 8}); + model.PopulateTensor(model.input2(), {4}); + model.Invoke(); + EXPECT_THAT(model.GetOutputShape(), ElementsAre(1, 2, 2, 1)); + EXPECT_THAT(model.GetOutput(), ElementsAre(20736, 16, 2401, 4096)); +} + +} // namespace +} // namespace tflite + +int main(int argc, char** argv) { + ::tflite::LogToStderr(); + ::testing::InitGoogleTest(&argc, argv); + return RUN_ALL_TESTS(); +} diff --git a/tensorflow/contrib/lite/kernels/register.cc b/tensorflow/contrib/lite/kernels/register.cc index 67f6caea678f840076f839a2203d047d1e63329d..22a507e6a43e74182dce29fe79a777ab2335db05 100644 --- a/tensorflow/contrib/lite/kernels/register.cc +++ b/tensorflow/contrib/lite/kernels/register.cc @@ -82,6 +82,7 @@ TfLiteRegistration* Register_PRELU(); TfLiteRegistration* Register_MAXIMUM(); TfLiteRegistration* Register_MINIMUM(); TfLiteRegistration* Register_ARG_MAX(); +TfLiteRegistration* Register_ARG_MIN(); TfLiteRegistration* Register_GREATER(); TfLiteRegistration* Register_GREATER_EQUAL(); TfLiteRegistration* Register_LESS(); @@ -101,6 +102,8 @@ TfLiteRegistration* Register_NOT_EQUAL(); TfLiteRegistration* Register_SQRT(); TfLiteRegistration* Register_RSQRT(); TfLiteRegistration* Register_SHAPE(); +TfLiteRegistration* Register_POW(); +TfLiteRegistration* Register_FAKE_QUANT(); BuiltinOpResolver::BuiltinOpResolver() { AddBuiltin(BuiltinOperator_RELU, Register_RELU()); @@ -122,7 +125,9 @@ BuiltinOpResolver::BuiltinOpResolver() { AddBuiltin(BuiltinOperator_EMBEDDING_LOOKUP, Register_EMBEDDING_LOOKUP()); AddBuiltin(BuiltinOperator_EMBEDDING_LOOKUP_SPARSE, Register_EMBEDDING_LOOKUP_SPARSE()); - AddBuiltin(BuiltinOperator_FULLY_CONNECTED, Register_FULLY_CONNECTED()); + AddBuiltin(BuiltinOperator_FULLY_CONNECTED, Register_FULLY_CONNECTED(), + /* min_version */ 1, + /* max_version */ 2); AddBuiltin(BuiltinOperator_LSH_PROJECTION, Register_LSH_PROJECTION()); AddBuiltin(BuiltinOperator_HASHTABLE_LOOKUP, Register_HASHTABLE_LOOKUP()); AddBuiltin(BuiltinOperator_SOFTMAX, Register_SOFTMAX()); @@ -164,6 +169,7 @@ BuiltinOpResolver::BuiltinOpResolver() { AddBuiltin(BuiltinOperator_MAXIMUM, Register_MAXIMUM()); AddBuiltin(BuiltinOperator_MINIMUM, Register_MINIMUM()); AddBuiltin(BuiltinOperator_ARG_MAX, Register_ARG_MAX()); + AddBuiltin(BuiltinOperator_ARG_MIN, Register_ARG_MIN()); AddBuiltin(BuiltinOperator_GREATER, Register_GREATER()); AddBuiltin(BuiltinOperator_GREATER_EQUAL, Register_GREATER_EQUAL()); AddBuiltin(BuiltinOperator_LESS, Register_LESS()); @@ -183,6 +189,8 @@ BuiltinOpResolver::BuiltinOpResolver() { AddBuiltin(BuiltinOperator_SQRT, Register_SQRT()); AddBuiltin(BuiltinOperator_RSQRT, Register_RSQRT()); AddBuiltin(BuiltinOperator_SHAPE, Register_SHAPE()); + AddBuiltin(BuiltinOperator_POW, Register_POW()); + AddBuiltin(BuiltinOperator_FAKE_QUANT, Register_FAKE_QUANT(), 1, 2); // TODO(andrewharp, ahentz): Move these somewhere more appropriate so that // custom ops aren't always included by default. diff --git a/tensorflow/contrib/lite/kernels/select.cc b/tensorflow/contrib/lite/kernels/select.cc index 9b6cee3cb55bf93b987fa8e59bdf9c591f5c0372..3cdb5db2090a3cb3eeb43c6e20a4fec09fe8a069 100644 --- a/tensorflow/contrib/lite/kernels/select.cc +++ b/tensorflow/contrib/lite/kernels/select.cc @@ -89,6 +89,9 @@ TfLiteStatus SelectEval(TfLiteContext* context, TfLiteNode* node) { case kTfLiteUInt8: \ TF_LITE_SELECT(uint8_t, op); \ break; \ + case kTfLiteInt16: \ + TF_LITE_SELECT(int16_t, op); \ + break; \ case kTfLiteInt32: \ TF_LITE_SELECT(int32_t, op); \ break; \ diff --git a/tensorflow/contrib/lite/kernels/select_test.cc b/tensorflow/contrib/lite/kernels/select_test.cc index 4664b9acb444747167f991944ddc120e9941ccd6..5b2e61cd29a7fd7c699fd81cb81e5f9a12c4b18f 100644 --- a/tensorflow/contrib/lite/kernels/select_test.cc +++ b/tensorflow/contrib/lite/kernels/select_test.cc @@ -96,6 +96,19 @@ TEST(SelectOpTest, SelectUInt8) { EXPECT_THAT(model.GetOutputShape(), ElementsAreArray({1, 1, 1, 4})); } +TEST(SelectOpTest, SelectInt16) { + SelectOpModel model({1, 1, 1, 4}, {1, 1, 1, 4}, {1, 1, 1, 4}, + TensorType_INT16); + + model.PopulateTensor(model.input1(), {false, true, false, false}); + model.PopulateTensor(model.input2(), {1, 2, 3, 4}); + model.PopulateTensor(model.input3(), {5, 6, 7, 8}); + model.Invoke(); + + EXPECT_THAT(model.GetOutput(), ElementsAreArray({5, 2, 7, 8})); + EXPECT_THAT(model.GetOutputShape(), ElementsAreArray({1, 1, 1, 4})); +} + TEST(SelectOpTest, SelectInt32) { SelectOpModel model({1, 1, 1, 4}, {1, 1, 1, 4}, {1, 1, 1, 4}, TensorType_INT32); diff --git a/tensorflow/contrib/lite/kernels/strided_slice.cc b/tensorflow/contrib/lite/kernels/strided_slice.cc index 725dd8105ab9506d5203ed38a11f8e06abdab603..bed2117f9ae3a64e963478eb03b46f0547f4c05f 100644 --- a/tensorflow/contrib/lite/kernels/strided_slice.cc +++ b/tensorflow/contrib/lite/kernels/strided_slice.cc @@ -121,10 +121,19 @@ TfLiteStatus ResizeOutputTensor(TfLiteContext* context, int32_t begin = GetBeginValueAtIndex(op_context, idx); int32_t end = GetEndValueAtIndex(op_context, idx); + // When shrinking an axis, the end position does not matter (and can be + // incorrect when negative indexing is used, see Issue #19260). Always use + // begin + 1 to generate a length 1 slice, since begin has + // already been adjusted for negative indices by GetBeginValueAtIndex. + const bool shrink_axis = op_context->params->shrink_axis_mask & (1 << idx); + if (shrink_axis) { + end = begin + 1; + } + // This is valid for both positive and negative strides int32_t dim_shape = ceil((end - begin) / static_cast(stride)); dim_shape = dim_shape < 0 ? 0 : dim_shape; - if (!(op_context->params->shrink_axis_mask & (1 << idx))) { + if (!shrink_axis) { output_shape_vector.push_back(dim_shape); } } @@ -204,13 +213,15 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { int begin_mask = ReverseMaskBits(op_context.params->begin_mask, op_context.dims); int end_mask = ReverseMaskBits(op_context.params->end_mask, op_context.dims); - -#define TF_LITE_STRIDED_SLICE(kernel_type, data_type) \ - kernel_type::StridedSlice(GetTensorData(op_context.input), \ - GetTensorDims(op_context.input), begin_mask, \ - end_mask, starts, stops, strides, \ - GetTensorData(op_context.output), \ - GetTensorDims(op_context.output)) + int shrink_axis_mask = + ReverseMaskBits(op_context.params->shrink_axis_mask, op_context.dims); + +#define TF_LITE_STRIDED_SLICE(kernel_type, data_type) \ + kernel_type::StridedSlice( \ + GetTensorData(op_context.input), \ + GetTensorDims(op_context.input), begin_mask, end_mask, shrink_axis_mask, \ + starts, stops, strides, GetTensorData(op_context.output), \ + GetTensorDims(op_context.output)) switch (op_context.input->type) { case kTfLiteFloat32: diff --git a/tensorflow/contrib/lite/kernels/strided_slice_test.cc b/tensorflow/contrib/lite/kernels/strided_slice_test.cc index e2be41d9588762869f639e8a7caff90f8c494561..c5d4f9affb46c82b4dec15bc0653d7315d132335 100644 --- a/tensorflow/contrib/lite/kernels/strided_slice_test.cc +++ b/tensorflow/contrib/lite/kernels/strided_slice_test.cc @@ -383,6 +383,45 @@ TEST(StridedSliceOpTest, In1D_ShrinkAxisMask1) { EXPECT_THAT(m.GetOutput(), ElementsAreArray({2})); } +TEST(StridedSliceOpTest, In1D_ShrinkAxisMask1_NegativeSlice) { + // This is equivalent to tf.range(4)[-1]. + StridedSliceOpModel<> m({4}, {1}, {1}, {1}, 0, 0, 0, 0, 1); + m.SetInput({0, 1, 2, 3}); + m.SetBegin({-1}); + m.SetEnd({0}); + m.SetStrides({1}); + + m.Invoke(); + EXPECT_TRUE(m.GetOutputShape().empty()); + EXPECT_THAT(m.GetOutput(), ElementsAreArray({3})); +} + +TEST(StridedSliceOpTest, In2D_ShrinkAxis3_NegativeSlice) { + // This is equivalent to tf.range(4)[:, tf.newaxis][-2, -1]. + StridedSliceOpModel<> m({4, 1}, {2}, {2}, {2}, 0, 0, 0, 0, 3); + m.SetInput({0, 1, 2, 3}); + m.SetBegin({-2, -1}); + m.SetEnd({-1, 0}); + m.SetStrides({1, 1}); + + m.Invoke(); + EXPECT_TRUE(m.GetOutputShape().empty()); + EXPECT_THAT(m.GetOutput(), ElementsAreArray({2})); +} + +TEST(StridedSliceOpTest, In2D_ShrinkAxis2_BeginEndAxis1_NegativeSlice) { + // This is equivalent to tf.range(4)[:, tf.newaxis][:, -1]. + StridedSliceOpModel<> m({4, 1}, {2}, {2}, {2}, 1, 1, 0, 0, 2); + m.SetInput({0, 1, 2, 3}); + m.SetBegin({0, -1}); + m.SetEnd({0, 0}); + m.SetStrides({1, 1}); + + m.Invoke(); + EXPECT_THAT(m.GetOutputShape(), ElementsAreArray({4})); + EXPECT_THAT(m.GetOutput(), ElementsAreArray({0, 1, 2, 3})); +} + TEST(StridedSliceOpTest, In1D_BeginMaskShrinkAxisMask1) { StridedSliceOpModel<> m({4}, {1}, {1}, {1}, 1, 0, 0, 0, 1); m.SetInput({1, 2, 3, 4}); @@ -394,17 +433,6 @@ TEST(StridedSliceOpTest, In1D_BeginMaskShrinkAxisMask1) { EXPECT_THAT(m.GetOutput(), ElementsAreArray({1})); } -TEST(StridedSliceOpTest, In1D_NegativeBeginNegativeStrideShrinkAxisMask1) { - StridedSliceOpModel<> m({4}, {1}, {1}, {1}, 0, 0, 0, 0, 1); - m.SetInput({1, 2, 3, 4}); - m.SetBegin({-2}); - m.SetEnd({-3}); - m.SetStrides({-1}); - m.Invoke(); - EXPECT_TRUE(m.GetOutputShape().empty()); - EXPECT_THAT(m.GetOutput(), ElementsAreArray({3})); -} - TEST(StridedSliceOpTest, In2D_ShrinkAxisMask1) { StridedSliceOpModel<> m({2, 3}, {2}, {2}, {2}, 0, 0, 0, 0, 1); m.SetInput({1, 2, 3, 4, 5, 6}); diff --git a/tensorflow/contrib/lite/kernels/sub.cc b/tensorflow/contrib/lite/kernels/sub.cc index a8b803589962032db3ed579d31e8b736c3afada0..1247525d416e8166a9e2e1d67c7907c00b0f6723 100644 --- a/tensorflow/contrib/lite/kernels/sub.cc +++ b/tensorflow/contrib/lite/kernels/sub.cc @@ -83,8 +83,8 @@ void EvalFloat(TfLiteContext* context, TfLiteNode* node, const TfLiteTensor* input1, const TfLiteTensor* input2, TfLiteTensor* output) { float output_activation_min, output_activation_max; - CalculateActivationRangeFloat(params->activation, &output_activation_min, - &output_activation_max); + CalculateActivationRange(params->activation, &output_activation_min, + &output_activation_max); #define TF_LITE_SUB(type, opname) \ type::opname(GetTensorData(input1), GetTensorDims(input1), \ GetTensorData(input2), GetTensorDims(input2), \ diff --git a/tensorflow/contrib/lite/kernels/svdf.cc b/tensorflow/contrib/lite/kernels/svdf.cc index 308860c299e9d74729d35b760e0f605437872c92..22eebdd4ceb16aeabc5e799c708f7236b3e2be37 100644 --- a/tensorflow/contrib/lite/kernels/svdf.cc +++ b/tensorflow/contrib/lite/kernels/svdf.cc @@ -12,6 +12,10 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License. ==============================================================================*/ + +// SVDF op that compresses a fully connected op via low-rank matrix +// factorization. See https://research.google.com/pubs/archive/43813.pdf for +// details. #include #include #include @@ -32,6 +36,67 @@ namespace ops { namespace builtin { namespace svdf { +namespace { + +struct OpData { + int scratch_tensor_index; + bool float_weights_time_initialized; +}; + +static inline void ApplyTimeWeightsBiasAndActivation( + int batch_size, int memory_size, int num_filters, int num_units, int rank, + const TfLiteTensor* weights_time, const TfLiteTensor* bias, + TfLiteFusedActivation activation, TfLiteTensor* state, + TfLiteTensor* scratch, TfLiteTensor* output) { + // Compute matmul(state, weights_time). + // The right most column is used to save temporary output (with the size of + // num_filters). This is achieved by starting at state->data.f and having the + // stride equal to memory_size. + for (int b = 0; b < batch_size; ++b) { + float* state_ptr_batch = state->data.f + b * memory_size * num_filters; + float* scratch_ptr_batch = scratch->data.f + b * num_filters; + tensor_utils::BatchVectorBatchVectorDotProduct( + weights_time->data.f, state_ptr_batch, memory_size, num_filters, + scratch_ptr_batch, /*result_stride=*/1); + } + + // Initialize output with bias if provided. + if (bias) { + tensor_utils::VectorBatchVectorAssign(bias->data.f, num_units, batch_size, + output->data.f); + } else { + tensor_utils::ZeroVector(output->data.f, batch_size * num_units); + } + + // Reduction sum. + for (int b = 0; b < batch_size; ++b) { + float* output_ptr_batch = output->data.f + b * num_units; + float* scratch_ptr_batch = scratch->data.f + b * num_filters; + tensor_utils::ReductionSumVector(scratch_ptr_batch, output_ptr_batch, + num_units, rank); + } + + // Apply activation. + for (int b = 0; b < batch_size; ++b) { + float* output_ptr_batch = output->data.f + b * num_units; + tensor_utils::ApplyActivationToVector(output_ptr_batch, num_units, + activation, output_ptr_batch); + } + + // Left shift the state to make room for next cycle's activation. + // TODO(alanchiao): explore collapsing this into a single loop. + for (int b = 0; b < batch_size; ++b) { + float* state_ptr_batch = state->data.f + b * memory_size * num_filters; + for (int f = 0; f < num_filters; ++f) { + tensor_utils::VectorShiftLeft(state_ptr_batch, memory_size, + /*shift_value=*/0.0); + state_ptr_batch += memory_size; + } + } +} + +} // namespace + constexpr int kInputTensor = 0; constexpr int kWeightsFeatureTensor = 1; constexpr int kWeightsTimeTensor = 2; @@ -40,29 +105,34 @@ constexpr int kStateTensor = 0; constexpr int kOutputTensor = 1; void* Init(TfLiteContext* context, const char* buffer, size_t length) { - auto* scratch_tensor_index = new int; - context->AddTensors(context, 1, scratch_tensor_index); - return scratch_tensor_index; + auto* op_data = new OpData; + op_data->float_weights_time_initialized = false; + context->AddTensors(context, /*tensors_to_add=*/4, + &op_data->scratch_tensor_index); + return op_data; } void Free(TfLiteContext* context, void* buffer) { - delete reinterpret_cast(buffer); + delete reinterpret_cast(buffer); } TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) { - auto* params = reinterpret_cast(node->builtin_data); - int* scratch_tensor_index = reinterpret_cast(node->user_data); + const auto* params = reinterpret_cast(node->builtin_data); + OpData* op_data = reinterpret_cast(node->user_data); + int scratch_tensor_index = op_data->scratch_tensor_index; // Check we have all the inputs and outputs we need. TF_LITE_ENSURE_EQ(context, node->inputs->size, 4); TF_LITE_ENSURE_EQ(context, node->outputs->size, 2); - TfLiteTensor* input = &context->tensors[node->inputs->data[kInputTensor]]; + const TfLiteTensor* input = GetInput(context, node, kInputTensor); const TfLiteTensor* weights_feature = GetInput(context, node, kWeightsFeatureTensor); const TfLiteTensor* weights_time = GetInput(context, node, kWeightsTimeTensor); + TF_LITE_ENSURE_EQ(context, input->type, kTfLiteFloat32); + // Check all the parameters of tensor match within themselves and match the // input configuration. const int rank = params->rank; @@ -103,10 +173,18 @@ TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) { TF_LITE_ENSURE_OK(context, context->ResizeTensor(context, output, output_size_array)); + // The weights are of consistent type, so it suffices to check one. + const bool is_hybrid_op = + (input->type == kTfLiteFloat32 && weights_feature->type == kTfLiteUInt8); + // Resize scratch. TfLiteIntArrayFree(node->temporaries); - node->temporaries = TfLiteIntArrayCreate(1); - node->temporaries->data[0] = *scratch_tensor_index; + if (is_hybrid_op) { + node->temporaries = TfLiteIntArrayCreate(4); + } else { + node->temporaries = TfLiteIntArrayCreate(1); + } + node->temporaries->data[0] = scratch_tensor_index; TfLiteIntArray* scratch_size_array = TfLiteIntArrayCreate(2); scratch_size_array->data[0] = batch_size; @@ -118,24 +196,56 @@ TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) { TF_LITE_ENSURE_OK(context, context->ResizeTensor(context, scratch_tensor, scratch_size_array)); - return kTfLiteOk; -} - -TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { - auto* params = reinterpret_cast(node->builtin_data); - - const TfLiteTensor* input = GetInput(context, node, kInputTensor); - const TfLiteTensor* weights_feature = - GetInput(context, node, kWeightsFeatureTensor); - const TfLiteTensor* weights_time = - GetInput(context, node, kWeightsTimeTensor); + if (is_hybrid_op) { + // Tell interpreter to allocate temporary tensors to store quantized values + // of input tensors. + node->temporaries->data[1] = scratch_tensor_index + 1; + TfLiteTensor* input_quantized = GetTemporary(context, node, /*index=*/1); + input_quantized->type = kTfLiteUInt8; + input_quantized->allocation_type = kTfLiteArenaRw; + if (!TfLiteIntArrayEqual(input_quantized->dims, input->dims)) { + TfLiteIntArray* input_quantized_size = TfLiteIntArrayCopy(input->dims); + TF_LITE_ENSURE_OK(context, context->ResizeTensor(context, input_quantized, + input_quantized_size)); + } - TfLiteTensor* state = GetOutput(context, node, kStateTensor); - TfLiteTensor* output = GetOutput(context, node, kOutputTensor); - TfLiteTensor* scratch = GetTemporary(context, node, /*index=*/0); + // Tell interpreter to allocate temporary tensors to store scaling factors. + node->temporaries->data[2] = scratch_tensor_index + 2; + TfLiteTensor* scaling_factors = GetTemporary(context, node, /*index=*/2); + scaling_factors->type = kTfLiteFloat32; + scaling_factors->allocation_type = kTfLiteArenaRw; + TfLiteIntArray* scaling_factors_size = TfLiteIntArrayCreate(1); + scaling_factors_size->data[0] = batch_size; + if (!TfLiteIntArrayEqual(scaling_factors->dims, scaling_factors_size)) { + TF_LITE_ENSURE_OK(context, context->ResizeTensor(context, scaling_factors, + scaling_factors_size)); + } - const TfLiteTensor* bias = GetOptionalInputTensor(context, node, kBiasTensor); + // Used to store dequantized weights_time matrix for hybrid computation + // of matmul(state, weights_time), which occurs in floating point. + node->temporaries->data[3] = scratch_tensor_index + 3; + TfLiteTensor* float_weights_time = GetTemporary(context, node, /*index=*/3); + float_weights_time->type = kTfLiteFloat32; + // Persistent so that we can compute the dequantized weights only once. + float_weights_time->allocation_type = kTfLiteArenaRwPersistent; + if (!TfLiteIntArrayEqual(float_weights_time->dims, weights_time->dims)) { + TfLiteIntArray* float_weights_time_size = + TfLiteIntArrayCopy(weights_time->dims); + TF_LITE_ENSURE_OK(context, + context->ResizeTensor(context, float_weights_time, + float_weights_time_size)); + } + } + return kTfLiteOk; +} +TfLiteStatus EvalFloat(TfLiteContext* context, TfLiteNode* node, + const TfLiteTensor* input, + const TfLiteTensor* weights_feature, + const TfLiteTensor* weights_time, + const TfLiteTensor* bias, const TfLiteSVDFParams* params, + TfLiteTensor* scratch, TfLiteTensor* state, + TfLiteTensor* output) { const int rank = params->rank; const int batch_size = input->dims->data[0]; const int input_size = input->dims->data[1]; @@ -146,67 +256,151 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { // Clear the activation (state left most column). // TODO(ghodrat): Add a test which initialize state with invalid values in // left most column and make sure it passes. - for (int b = 0; b < batch_size; b++) { + for (int b = 0; b < batch_size; ++b) { float* state_ptr_batch = state->data.f + b * memory_size * num_filters; - for (int c = 0; c < num_filters; c++) { + for (int c = 0; c < num_filters; ++c) { float* state_ptr = state_ptr_batch + c * memory_size; state_ptr[memory_size - 1] = 0.0; } } // Compute conv1d(inputs, weights_feature). - // The state left most column is used to save current cycle activation. This + // The state right most column is used to save current cycle activation. This // is achieved by starting at state->data.f[memory_size - 1] and having the // stride equal to memory_size. tensor_utils::MatrixBatchVectorMultiplyAccumulate( weights_feature->data.f, num_filters, input_size, input->data.f, batch_size, &state->data.f[memory_size - 1], memory_size); - // Compute matmul(state, weights_time). - // The right most column is used to save temporary output (with the size of - // num_filters). This is achieved by starting at state->data.f and having the - // stride equal to memory_size. - for (int b = 0; b < batch_size; b++) { + ApplyTimeWeightsBiasAndActivation(batch_size, memory_size, num_filters, + num_units, rank, weights_time, bias, + params->activation, state, scratch, output); + return kTfLiteOk; +} + +TfLiteStatus EvalHybrid( + TfLiteContext* context, TfLiteNode* node, const TfLiteTensor* input, + const TfLiteTensor* weights_feature, const TfLiteTensor* weights_time, + const TfLiteTensor* bias, const TfLiteSVDFParams* params, + TfLiteTensor* scratch, TfLiteTensor* scaling_factors, + TfLiteTensor* input_quantized, TfLiteTensor* state, TfLiteTensor* output) { + const int rank = params->rank; + const int batch_size = input->dims->data[0]; + const int input_size = input->dims->data[1]; + const int num_filters = weights_feature->dims->data[0]; + const int num_units = num_filters / rank; + const int memory_size = weights_time->dims->data[1]; + + // Initialize the pointer to input. + const float* input_ptr_batch = input->data.f; + + // Initialize the pointer to storage for quantized values and + // scaling factors. + int8_t* quantized_input_ptr_batch = + reinterpret_cast(input_quantized->data.uint8); + + float* scaling_factors_ptr = scaling_factors->data.f; + + // Other initializations. + const int8_t* weights_feature_ptr = + reinterpret_cast(weights_feature->data.uint8); + const float weights_feature_scale = weights_feature->params.scale; + + // Clear the activation (state left most column). + // TODO(ghodrat): Add a test which initialize state with invalid values in + // left most column and make sure it passes. + for (int b = 0; b < batch_size; ++b) { float* state_ptr_batch = state->data.f + b * memory_size * num_filters; - float* scratch_ptr_batch = scratch->data.f + b * num_filters; - tensor_utils::BatchVectorBatchVectorDotProduct( - weights_time->data.f, state_ptr_batch, memory_size, num_filters, - scratch_ptr_batch, /*result_stride=*/1); + for (int c = 0; c < num_filters; ++c) { + float* state_ptr = state_ptr_batch + c * memory_size; + state_ptr[memory_size - 1] = 0.0; + } } - // Initialize output with bias if provided. - if (bias) { - tensor_utils::VectorBatchVectorAssign(bias->data.f, num_units, batch_size, - output->data.f); - } else { - tensor_utils::ZeroVector(output->data.f, batch_size * num_units); - } + if (!tensor_utils::IsZeroVector(input_ptr_batch, batch_size * input_size)) { + // Quantize input from float to int8. + float unused_min, unused_max; + for (int b = 0; b < batch_size; ++b) { + const int offset = b * input_size; + tensor_utils::SymmetricQuantizeFloats( + input_ptr_batch + offset, input_size, + quantized_input_ptr_batch + offset, &unused_min, &unused_max, + &scaling_factors_ptr[b]); + scaling_factors_ptr[b] *= weights_feature_scale; + } - // Reduction sum - for (int b = 0; b < batch_size; b++) { - float* output_ptr_batch = output->data.f + b * num_units; - float* scratch_ptr_batch = scratch->data.f + b * num_filters; - tensor_utils::ReductionSumVector(scratch_ptr_batch, output_ptr_batch, - num_units, rank); + // Compute conv1d(inputs, weights_feature). + // The state right most column is used to save current cycle activation. + // This is achieved by starting at state->data.f[memory_size - 1] and having + // the stride equal to memory_size. + tensor_utils::MatrixBatchVectorMultiplyAccumulate( + weights_feature_ptr, num_filters, input_size, quantized_input_ptr_batch, + scaling_factors_ptr, batch_size, &state->data.f[memory_size - 1], + memory_size); } - // Apply activation. - for (int b = 0; b < batch_size; b++) { - float* output_ptr_batch = output->data.f + b * num_units; - tensor_utils::ApplyActivationToVector(output_ptr_batch, num_units, - params->activation, output_ptr_batch); - } + // TODO(alanchiao): can optimize hybrid case ~5% by unrolling loop in applying + // time weights so that the inner loop multiplies eight elements at a time. + ApplyTimeWeightsBiasAndActivation(batch_size, memory_size, num_filters, + num_units, rank, weights_time, bias, + params->activation, state, scratch, output); + return kTfLiteOk; +} - // Right shift the state. - for (int b = 0; b < batch_size; b++) { - float* state_ptr_batch = state->data.f + b * memory_size * num_filters; - for (int f = 0; f < num_filters; f++) { - tensor_utils::VectorShiftLeft(state_ptr_batch, memory_size, - /*shift_value=*/0.0); - state_ptr_batch += memory_size; +TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { + auto* params = reinterpret_cast(node->builtin_data); + OpData* op_data = reinterpret_cast(node->user_data); + + const TfLiteTensor* input = GetInput(context, node, kInputTensor); + const TfLiteTensor* weights_feature = + GetInput(context, node, kWeightsFeatureTensor); + const TfLiteTensor* weights_time = + GetInput(context, node, kWeightsTimeTensor); + const TfLiteTensor* bias = GetOptionalInputTensor(context, node, kBiasTensor); + + TfLiteTensor* scratch = GetTemporary(context, node, /*index=*/0); + + TfLiteTensor* state = GetOutput(context, node, kStateTensor); + TfLiteTensor* output = GetOutput(context, node, kOutputTensor); + + switch (weights_feature->type) { + case kTfLiteFloat32: { + return EvalFloat(context, node, input, weights_feature, weights_time, + bias, params, scratch, state, output); + break; } + case kTfLiteUInt8: { + TfLiteTensor* input_quantized = GetTemporary(context, node, /*index=*/1); + TfLiteTensor* scaling_factors = GetTemporary(context, node, /*index=*/2); + TfLiteTensor* float_weights_time = + GetTemporary(context, node, /*index=*/3); + + // Dequantize weights time. + // TODO(alanchiao): this dequantization initialization only needs to + // happen once per model and should theoretically be placed in either Init + // or Prepare. However, TFLite doesn't allocate float_weights_time until + // the Eval function. + // TODO(alanchiao): refactor logic out into dequantize function. + if (!op_data->float_weights_time_initialized) { + const float dequantization_scale = weights_time->params.scale; + const int8_t* weights_time_ptr = + reinterpret_cast(weights_time->data.uint8); + for (int i = 0; i < NumElements(float_weights_time); ++i) { + float_weights_time->data.f[i] = + weights_time_ptr[i] * dequantization_scale; + } + op_data->float_weights_time_initialized = true; + } + return EvalHybrid(context, node, input, weights_feature, + float_weights_time, bias, params, scratch, + scaling_factors, input_quantized, state, output); + break; + } + default: + context->ReportError(context, "Type %d not currently supported.", + weights_feature->type); + return kTfLiteError; } - return kTfLiteOk; } } // namespace svdf diff --git a/tensorflow/contrib/lite/kernels/svdf_test.cc b/tensorflow/contrib/lite/kernels/svdf_test.cc index 0f166dc69b95f3459388135b3a6c4d9b73a31cb4..5af3ff85004ce43c5b75c6f12761f121c0d8deca 100644 --- a/tensorflow/contrib/lite/kernels/svdf_test.cc +++ b/tensorflow/contrib/lite/kernels/svdf_test.cc @@ -126,17 +126,20 @@ static float svdf_golden_output_rank_2[] = { }; // Derived class of SingleOpModel, which is used to test SVDF TFLite op. -class SVDFOpModel : public SingleOpModel { +class BaseSVDFOpModel : public SingleOpModel { public: - SVDFOpModel(int batches, int units, int input_size, int memory_size, int rank) + BaseSVDFOpModel(int batches, int units, int input_size, int memory_size, + int rank, + TensorType weights_feature_type = TensorType_FLOAT32, + TensorType weights_time_type = TensorType_FLOAT32) : batches_(batches), units_(units), input_size_(input_size), memory_size_(memory_size), rank_(rank) { input_ = AddInput(TensorType_FLOAT32); - weights_feature_ = AddInput(TensorType_FLOAT32); - weights_time_ = AddInput(TensorType_FLOAT32); + weights_feature_ = AddInput(weights_feature_type); + weights_time_ = AddInput(weights_time_type); bias_ = AddNullInput(); state_ = AddOutput(TensorType_FLOAT32); output_ = AddOutput(TensorType_FLOAT32); @@ -182,7 +185,7 @@ class SVDFOpModel : public SingleOpModel { int num_units() { return units_; } int num_batches() { return batches_; } - private: + protected: int input_; int weights_feature_; int weights_time_; @@ -197,7 +200,61 @@ class SVDFOpModel : public SingleOpModel { int rank_; }; -TEST(SVDFOpTest, BlackBoxTestRank1) { +class SVDFOpModel : public BaseSVDFOpModel { + public: + using BaseSVDFOpModel::BaseSVDFOpModel; +}; + +class HybridSVDFOpModel : public BaseSVDFOpModel { + public: + HybridSVDFOpModel(int batches, int units, int input_size, int memory_size, + int rank) + : BaseSVDFOpModel(batches, units, input_size, memory_size, rank, + TensorType_UINT8, TensorType_UINT8) {} + + void SetWeightsFeature(std::initializer_list f) { + SymmetricQuantizeAndPopulate(weights_feature_, f); + } + + void SetWeightsTime(std::initializer_list f) { + SymmetricQuantizeAndPopulate(weights_time_, f); + } +}; + +class SVDFOpTest : public ::testing::Test { + protected: + void VerifyGoldens(float golden_input[], float golden_output[], + int golden_size, BaseSVDFOpModel* svdf, + float tolerance = 1e-5) { + const int svdf_num_batches = svdf->num_batches(); + const int svdf_input_size = svdf->input_size(); + const int svdf_num_units = svdf->num_units(); + const int input_sequence_size = + golden_size / sizeof(float) / (svdf_input_size * svdf_num_batches); + // Going over each input batch, setting the input tensor, invoking the SVDF + // op and checking the output with the expected golden values. + for (int i = 0; i < input_sequence_size; i++) { + float* batch_start = + golden_input + i * svdf_input_size * svdf_num_batches; + float* batch_end = batch_start + svdf_input_size * svdf_num_batches; + svdf->SetInput(0, batch_start, batch_end); + + svdf->Invoke(); + + const float* golden_start = + golden_output + i * svdf_num_units * svdf_num_batches; + const float* golden_end = + golden_start + svdf_num_units * svdf_num_batches; + std::vector expected; + expected.insert(expected.end(), golden_start, golden_end); + + EXPECT_THAT(svdf->GetOutput(), + ElementsAreArray(ArrayFloatNear(expected, tolerance))); + } + } +}; + +TEST_F(SVDFOpTest, BlackBoxTestRank1) { SVDFOpModel svdf(/*batches=*/2, /*units=*/4, /*input_size=*/3, /*memory_size=*/10, /*rank=*/1); svdf.SetWeightsFeature({-0.31930989, -0.36118156, 0.0079667, 0.37613347, @@ -218,31 +275,11 @@ TEST(SVDFOpTest, BlackBoxTestRank1) { -0.01580888, -0.14943552, 0.15465137, 0.09784451, -0.0337657}); svdf.ResetState(); - const int svdf_num_batches = svdf.num_batches(); - const int svdf_input_size = svdf.input_size(); - const int svdf_num_units = svdf.num_units(); - const int input_sequence_size = - sizeof(svdf_input) / sizeof(float) / (svdf_input_size * svdf_num_batches); - // Going over each input batch, setting the input tensor, invoking the SVDF op - // and checking the output with the expected golden values. - for (int i = 0; i < input_sequence_size; i++) { - float* batch_start = svdf_input + i * svdf_input_size * svdf_num_batches; - float* batch_end = batch_start + svdf_input_size * svdf_num_batches; - svdf.SetInput(0, batch_start, batch_end); - - svdf.Invoke(); - - float* golden_start = - svdf_golden_output_rank_1 + i * svdf_num_units * svdf_num_batches; - float* golden_end = golden_start + svdf_num_units * svdf_num_batches; - std::vector expected; - expected.insert(expected.end(), golden_start, golden_end); - - EXPECT_THAT(svdf.GetOutput(), ElementsAreArray(ArrayFloatNear(expected))); - } + VerifyGoldens(svdf_input, svdf_golden_output_rank_1, sizeof(svdf_input), + &svdf); } -TEST(SVDFOpTest, BlackBoxTestRank2) { +TEST_F(SVDFOpTest, BlackBoxTestRank2) { SVDFOpModel svdf(/*batches=*/2, /*units=*/4, /*input_size=*/3, /*memory_size=*/10, /*rank=*/2); svdf.SetWeightsFeature({-0.31930989, 0.0079667, 0.39296314, 0.37613347, @@ -278,28 +315,75 @@ TEST(SVDFOpTest, BlackBoxTestRank2) { 0.08682203, 0.1258215, 0.1851041, 0.29228821, 0.12366763}); svdf.ResetState(); - const int svdf_num_batches = svdf.num_batches(); - const int svdf_input_size = svdf.input_size(); - const int svdf_num_units = svdf.num_units(); - const int input_sequence_size = - sizeof(svdf_input) / sizeof(float) / (svdf_input_size * svdf_num_batches); - // Going over each input batch, setting the input tensor, invoking the SVDF op - // and checking the output with the expected golden values. - for (int i = 0; i < input_sequence_size; i++) { - float* batch_start = svdf_input + i * svdf_input_size * svdf_num_batches; - float* batch_end = batch_start + svdf_input_size * svdf_num_batches; - svdf.SetInput(0, batch_start, batch_end); - - svdf.Invoke(); - - float* golden_start = - svdf_golden_output_rank_2 + i * svdf_num_units * svdf_num_batches; - float* golden_end = golden_start + svdf_num_units * svdf_num_batches; - std::vector expected; - expected.insert(expected.end(), golden_start, golden_end); - - EXPECT_THAT(svdf.GetOutput(), ElementsAreArray(ArrayFloatNear(expected))); - } + VerifyGoldens(svdf_input, svdf_golden_output_rank_2, sizeof(svdf_input), + &svdf); +} + +TEST_F(SVDFOpTest, BlackBoxTestHybridRank1) { + HybridSVDFOpModel svdf(/*batches=*/2, /*units=*/4, /*input_size=*/3, + /*memory_size=*/10, /*rank=*/1); + svdf.SetWeightsFeature({-0.31930989, -0.36118156, 0.0079667, 0.37613347, + 0.22197971, 0.12416199, 0.27901134, 0.27557442, + 0.3905206, -0.36137494, -0.06634006, -0.10640851}); + + svdf.SetWeightsTime( + {-0.31930989, 0.37613347, 0.27901134, -0.36137494, -0.36118156, + 0.22197971, 0.27557442, -0.06634006, 0.0079667, 0.12416199, + + 0.3905206, -0.10640851, -0.0976817, 0.15294972, 0.39635518, + -0.02702999, 0.39296314, 0.15785322, 0.21931258, 0.31053296, + + -0.36916667, 0.38031587, -0.21580373, 0.27072677, 0.23622236, + 0.34936687, 0.18174365, 0.35907319, -0.17493086, 0.324846, + + -0.10781813, 0.27201805, 0.14324132, -0.23681851, -0.27115166, + -0.01580888, -0.14943552, 0.15465137, 0.09784451, -0.0337657}); + + svdf.ResetState(); + VerifyGoldens(svdf_input, svdf_golden_output_rank_1, sizeof(svdf_input), + &svdf, + /*tolerance=*/0.002945); +} + +TEST_F(SVDFOpTest, BlackBoxTestHybridRank2) { + HybridSVDFOpModel svdf(/*batches=*/2, /*units=*/4, /*input_size=*/3, + /*memory_size=*/10, /*rank=*/2); + svdf.SetWeightsFeature({-0.31930989, 0.0079667, 0.39296314, 0.37613347, + 0.12416199, 0.15785322, 0.27901134, 0.3905206, + 0.21931258, -0.36137494, -0.10640851, 0.31053296, + -0.36118156, -0.0976817, -0.36916667, 0.22197971, + 0.15294972, 0.38031587, 0.27557442, 0.39635518, + -0.21580373, -0.06634006, -0.02702999, 0.27072677}); + + svdf.SetWeightsTime( + {-0.31930989, 0.37613347, 0.27901134, -0.36137494, -0.36118156, + 0.22197971, 0.27557442, -0.06634006, 0.0079667, 0.12416199, + + 0.3905206, -0.10640851, -0.0976817, 0.15294972, 0.39635518, + -0.02702999, 0.39296314, 0.15785322, 0.21931258, 0.31053296, + + -0.36916667, 0.38031587, -0.21580373, 0.27072677, 0.23622236, + 0.34936687, 0.18174365, 0.35907319, -0.17493086, 0.324846, + + -0.10781813, 0.27201805, 0.14324132, -0.23681851, -0.27115166, + -0.01580888, -0.14943552, 0.15465137, 0.09784451, -0.0337657, + + -0.14884081, 0.19931212, -0.36002168, 0.34663299, -0.11405486, + 0.12672701, 0.39463779, -0.07886535, -0.06384811, 0.08249187, + + -0.26816407, -0.19905911, 0.29211238, 0.31264046, -0.28664589, + 0.05698794, 0.11613581, 0.14078894, 0.02187902, -0.21781836, + + -0.15567942, 0.08693647, -0.38256618, 0.36580828, -0.22922277, + -0.0226903, 0.12878349, -0.28122205, -0.10850525, -0.11955214, + + 0.27179423, -0.04710215, 0.31069002, 0.22672787, 0.09580326, + 0.08682203, 0.1258215, 0.1851041, 0.29228821, 0.12366763}); + + svdf.ResetState(); + VerifyGoldens(svdf_input, svdf_golden_output_rank_2, sizeof(svdf_input), + &svdf, + /*tolerance=*/0.00625109); } } // namespace diff --git a/tensorflow/contrib/lite/kernels/test_util.h b/tensorflow/contrib/lite/kernels/test_util.h index 5094e1343aa7b31537333ef5a770dbbfe3954f55..bedbe93ae65662647f6a0fb0c9c6a6a921e148bb 100644 --- a/tensorflow/contrib/lite/kernels/test_util.h +++ b/tensorflow/contrib/lite/kernels/test_util.h @@ -148,20 +148,18 @@ class SingleOpModel { int AddOutput(const TensorData& t); template - void QuantizeAndPopulate(int index, std::initializer_list data) { + void QuantizeAndPopulate(int index, const std::vector& data) { TfLiteTensor* t = interpreter_->tensor(index); auto q = Quantize(data, t->params.scale, t->params.zero_point); PopulateTensor(index, 0, q.data(), q.data() + q.size()); } - void SymmetricQuantizeAndPopulate(int index, - std::initializer_list data) { + void SymmetricQuantizeAndPopulate(int index, const std::vector& data) { TfLiteTensor* t = interpreter_->tensor(index); - std::vector values(data); - const int length = values.size(); + const int length = data.size(); std::vector q(length); float min, max, scaling_factor; - tensor_utils::SymmetricQuantizeFloats(values.data(), length, q.data(), &min, + tensor_utils::SymmetricQuantizeFloats(data.data(), length, q.data(), &min, &max, &scaling_factor); // Update quantization params. t->params.scale = scaling_factor; @@ -198,8 +196,22 @@ class SingleOpModel { } // Populate the tensor given its index. + // TODO(b/110696148) clean up and merge with vector-taking variant below. template - void PopulateTensor(int index, std::initializer_list data) { + void PopulateTensor(int index, const std::initializer_list& data) { + T* v = interpreter_->typed_tensor(index); + CHECK(v) << "No tensor with index '" << index << "'."; + for (T f : data) { + *v = f; + ++v; + } + } + + // Populate the tensor given its index. + // TODO(b/110696148) clean up and merge with initializer_list-taking variant + // above. + template + void PopulateTensor(int index, const std::vector& data) { T* v = interpreter_->typed_tensor(index); CHECK(v) << "No tensor with index '" << index << "'."; for (T f : data) { diff --git a/tensorflow/contrib/lite/kernels/topk_v2.cc b/tensorflow/contrib/lite/kernels/topk_v2.cc index fb0e49c90c41747f9b7e53570276c8b8045030fd..2dd760bbfebd1faa8b7ff9158bc1a1b1d4647525 100644 --- a/tensorflow/contrib/lite/kernels/topk_v2.cc +++ b/tensorflow/contrib/lite/kernels/topk_v2.cc @@ -56,11 +56,13 @@ TfLiteStatus ResizeOutput(TfLiteContext* context, TfLiteNode* node) { output_values_shape->data[num_dimensions - 1] = k; TfLiteTensor* output_indexes = GetOutput(context, node, kOutputIndexes); TfLiteTensor* output_values = GetOutput(context, node, kOutputValues); + // Force output types. + output_indexes->type = kTfLiteInt32; + output_values->type = input->type; auto resize_tensor = [context](TfLiteTensor* tensor, TfLiteIntArray* new_size, TfLiteIntArray* delete_on_error) { TfLiteStatus status = context->ResizeTensor(context, tensor, new_size); if (status != kTfLiteOk) { - TfLiteIntArrayFree(new_size); if (delete_on_error != nullptr) { TfLiteIntArrayFree(delete_on_error); } diff --git a/tensorflow/contrib/lite/kernels/unidirectional_sequence_lstm.cc b/tensorflow/contrib/lite/kernels/unidirectional_sequence_lstm.cc index 1c28123a24edd9886476bf8e9ea3ba4c692baa2b..c48b470f929da277b09b5c58363ad9081e8966a7 100644 --- a/tensorflow/contrib/lite/kernels/unidirectional_sequence_lstm.cc +++ b/tensorflow/contrib/lite/kernels/unidirectional_sequence_lstm.cc @@ -70,9 +70,21 @@ constexpr int kOutputStateTensor = 0; constexpr int kCellStateTensor = 1; constexpr int kOutputTensor = 2; +// Temporary tensors +enum TemporaryTensor { + kScratchBuffer = 0, + kInputQuantized = 1, + kOutputStateQuantized = 2, + kCellStateQuantized = 3, + kScalingFactors = 4, + kProductScalingFactors = 5, + kRecoveredCellWeights = 6, + kNumTemporaryTensors = 7 +}; + void* Init(TfLiteContext* context, const char* buffer, size_t length) { auto* scratch_tensor_index = new int; - context->AddTensors(context, 1, scratch_tensor_index); + context->AddTensors(context, kNumTemporaryTensors, scratch_tensor_index); return scratch_tensor_index; } @@ -84,7 +96,7 @@ void Free(TfLiteContext* context, void* buffer) { TfLiteStatus CheckInputTensorDimensions(TfLiteContext* context, TfLiteNode* node, int n_input, int n_output, int n_cell) { - auto* params = reinterpret_cast(node->builtin_data); + const auto* params = reinterpret_cast(node->builtin_data); // Making sure clipping parameters have valid values. // == 0 means no clipping @@ -242,6 +254,7 @@ TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) { // Inferring batch size, number of outputs and sequence length and // number of cells from the input tensors. const TfLiteTensor* input = GetInput(context, node, kInputTensor); + TF_LITE_ENSURE_EQ(context, input->type, kTfLiteFloat32); TF_LITE_ENSURE(context, input->dims->size > 1); const int max_time = input->dims->data[0]; const int n_batch = input->dims->data[1]; @@ -261,7 +274,8 @@ TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) { const int n_output = recurrent_to_output_weights->dims->data[1]; // Check that input tensor dimensions matches with each other. - CheckInputTensorDimensions(context, node, n_input, n_output, n_cell); + TF_LITE_ENSURE_OK(context, CheckInputTensorDimensions(context, node, n_input, + n_output, n_cell)); // Get the pointer to output, output_state and cell_state buffer tensors. TfLiteTensor* output = GetOutput(context, node, kOutputTensor); @@ -288,86 +302,156 @@ TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) { TF_LITE_ENSURE_OK(context, context->ResizeTensor(context, cell_state, cell_size)); - // Create a scratch buffer tensor. + // Mark state tensors as persistent tensors. + output_state->allocation_type = kTfLiteArenaRwPersistent; + cell_state->allocation_type = kTfLiteArenaRwPersistent; + + // The weights are of consistent type, so it suffices to check one. + // TODO(mirkov): create a utility/macro for this check, so all Ops can use it. + const bool is_hybrid_op = (input_to_output_weights->type == kTfLiteUInt8 && + input->type == kTfLiteFloat32); + TfLiteIntArrayFree(node->temporaries); - node->temporaries = TfLiteIntArrayCreate(1); + if (is_hybrid_op) { + node->temporaries = TfLiteIntArrayCreate(kNumTemporaryTensors); + } else { + node->temporaries = TfLiteIntArrayCreate(1); + } node->temporaries->data[0] = *scratch_tensor_index; - TfLiteTensor* scratch_buffer = GetTemporary(context, node, /*index=*/0); + + // Create a scratch buffer tensor. + TfLiteTensor* scratch_buffer = GetTemporary(context, node, kScratchBuffer); scratch_buffer->type = input->type; scratch_buffer->allocation_type = kTfLiteArenaRw; - // Mark state tensors as persistent tensors. - output_state->allocation_type = kTfLiteArenaRwPersistent; - cell_state->allocation_type = kTfLiteArenaRwPersistent; - const TfLiteTensor* input_to_input_weights = GetOptionalInputTensor(context, node, kInputToInputWeightsTensor); const bool use_cifg = (input_to_input_weights == nullptr); + TfLiteIntArray* scratch_buffer_size = TfLiteIntArrayCreate(2); + scratch_buffer_size->data[0] = n_batch; if (use_cifg) { - TfLiteIntArray* scratch_buffer_size = TfLiteIntArrayCreate(2); - scratch_buffer_size->data[0] = n_batch; // Reserving space for Cell, Forget, Output gates scratch_buffer_size->data[1] = n_cell * 3; - TF_LITE_ENSURE_OK(context, context->ResizeTensor(context, scratch_buffer, - scratch_buffer_size)); } else { - TfLiteIntArray* scratch_buffer_size = TfLiteIntArrayCreate(2); - scratch_buffer_size->data[0] = n_batch; // Reserving space for Input, Cell, Forget, Output gates scratch_buffer_size->data[1] = n_cell * 4; - TF_LITE_ENSURE_OK(context, context->ResizeTensor(context, scratch_buffer, - scratch_buffer_size)); + } + TF_LITE_ENSURE_OK(context, context->ResizeTensor(context, scratch_buffer, + scratch_buffer_size)); + + if (is_hybrid_op) { + // Allocate temporary tensors to store quantized values of input, + // output_state and cell_state tensors. + node->temporaries->data[kInputQuantized] = + *scratch_tensor_index + kInputQuantized; + TfLiteTensor* input_quantized = + GetTemporary(context, node, kInputQuantized); + input_quantized->type = kTfLiteUInt8; + input_quantized->allocation_type = kTfLiteArenaRw; + if (!TfLiteIntArrayEqual(input_quantized->dims, input->dims)) { + TfLiteIntArray* input_quantized_size = TfLiteIntArrayCopy(input->dims); + TF_LITE_ENSURE_OK(context, context->ResizeTensor(context, input_quantized, + input_quantized_size)); + } + node->temporaries->data[kOutputStateQuantized] = + *scratch_tensor_index + kOutputStateQuantized; + TfLiteTensor* output_state_quantized = + GetTemporary(context, node, kOutputStateQuantized); + output_state_quantized->type = kTfLiteUInt8; + output_state_quantized->allocation_type = kTfLiteArenaRw; + if (!TfLiteIntArrayEqual(output_state_quantized->dims, + output_state->dims)) { + TfLiteIntArray* output_state_quantized_size = + TfLiteIntArrayCopy(output_state->dims); + TF_LITE_ENSURE_OK(context, + context->ResizeTensor(context, output_state_quantized, + output_state_quantized_size)); + } + node->temporaries->data[kCellStateQuantized] = + *scratch_tensor_index + kCellStateQuantized; + TfLiteTensor* cell_state_quantized = + GetTemporary(context, node, kCellStateQuantized); + cell_state_quantized->type = kTfLiteUInt8; + cell_state_quantized->allocation_type = kTfLiteArenaRw; + if (!TfLiteIntArrayEqual(cell_state_quantized->dims, cell_state->dims)) { + TfLiteIntArray* cell_state_quantized_size = + TfLiteIntArrayCopy(cell_state->dims); + TF_LITE_ENSURE_OK(context, + context->ResizeTensor(context, cell_state_quantized, + cell_state_quantized_size)); + } + + // Allocate temporary tensors to store scaling factors and product scaling + // factors. The latter is a convenience storage which allows to quantize + // a vector once (which produces the scaling factors) and multiply it with + // different matrices (which requires multiplying the scaling factors with + // the scaling factor of the matrix). + node->temporaries->data[kScalingFactors] = + *scratch_tensor_index + kScalingFactors; + TfLiteTensor* scaling_factors = + GetTemporary(context, node, kScalingFactors); + scaling_factors->type = kTfLiteFloat32; + scaling_factors->allocation_type = kTfLiteArenaRw; + TfLiteIntArray* scaling_factors_size = TfLiteIntArrayCreate(1); + scaling_factors_size->data[0] = n_batch; + if (!TfLiteIntArrayEqual(scaling_factors->dims, scaling_factors_size)) { + TF_LITE_ENSURE_OK(context, context->ResizeTensor(context, scaling_factors, + scaling_factors_size)); + } + node->temporaries->data[kProductScalingFactors] = + *scratch_tensor_index + kProductScalingFactors; + TfLiteTensor* prod_scaling_factors = + GetTemporary(context, node, kProductScalingFactors); + prod_scaling_factors->type = kTfLiteFloat32; + prod_scaling_factors->allocation_type = kTfLiteArenaRw; + TfLiteIntArray* prod_scaling_factors_size = TfLiteIntArrayCreate(1); + prod_scaling_factors_size->data[0] = n_batch; + if (!TfLiteIntArrayEqual(prod_scaling_factors->dims, + prod_scaling_factors_size)) { + TF_LITE_ENSURE_OK(context, + context->ResizeTensor(context, prod_scaling_factors, + prod_scaling_factors_size)); + } + + // Allocate a temporary tensor to store the recovered cell weights. Since + // this is used for diagonal matrices, only need to store n_cell values. + node->temporaries->data[kRecoveredCellWeights] = + *scratch_tensor_index + kRecoveredCellWeights; + TfLiteTensor* recovered_cell_weights = + GetTemporary(context, node, kRecoveredCellWeights); + recovered_cell_weights->type = kTfLiteFloat32; + recovered_cell_weights->allocation_type = kTfLiteArenaRw; + TfLiteIntArray* recovered_cell_weights_size = TfLiteIntArrayCreate(1); + recovered_cell_weights_size->data[0] = n_cell; + if (!TfLiteIntArrayEqual(recovered_cell_weights->dims, + recovered_cell_weights_size)) { + TF_LITE_ENSURE_OK(context, + context->ResizeTensor(context, recovered_cell_weights, + recovered_cell_weights_size)); + } } return kTfLiteOk; } // The LSTM Op engine. -TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { - auto* params = reinterpret_cast(node->builtin_data); - const TfLiteTensor* input = GetInput(context, node, kInputTensor); - - const TfLiteTensor* input_to_input_weights = - GetOptionalInputTensor(context, node, kInputToInputWeightsTensor); - const TfLiteTensor* input_to_forget_weights = - GetInput(context, node, kInputToForgetWeightsTensor); - const TfLiteTensor* input_to_cell_weights = - GetInput(context, node, kInputToCellWeightsTensor); - const TfLiteTensor* input_to_output_weights = - GetInput(context, node, kInputToOutputWeightsTensor); - - const TfLiteTensor* recurrent_to_input_weights = - GetOptionalInputTensor(context, node, kRecurrentToInputWeightsTensor); - const TfLiteTensor* recurrent_to_forget_weights = - GetInput(context, node, kRecurrentToForgetWeightsTensor); - const TfLiteTensor* recurrent_to_cell_weights = - GetInput(context, node, kRecurrentToCellWeightsTensor); - const TfLiteTensor* recurrent_to_output_weights = - GetInput(context, node, kRecurrentToOutputWeightsTensor); - - const TfLiteTensor* cell_to_input_weights = - GetOptionalInputTensor(context, node, kCellToInputWeightsTensor); - const TfLiteTensor* cell_to_forget_weights = - GetOptionalInputTensor(context, node, kCellToForgetWeightsTensor); - const TfLiteTensor* cell_to_output_weights = - GetOptionalInputTensor(context, node, kCellToOutputWeightsTensor); - - const TfLiteTensor* input_gate_bias = - GetOptionalInputTensor(context, node, kInputGateBiasTensor); - const TfLiteTensor* forget_gate_bias = - GetInput(context, node, kForgetGateBiasTensor); - const TfLiteTensor* cell_bias = GetInput(context, node, kCellGateBiasTensor); - const TfLiteTensor* output_gate_bias = - GetInput(context, node, kOutputGateBiasTensor); - - const TfLiteTensor* projection_weights = - GetOptionalInputTensor(context, node, kProjectionWeightsTensor); - const TfLiteTensor* projection_bias = - GetOptionalInputTensor(context, node, kProjectionBiasTensor); - - TfLiteTensor* output_state = GetOutput(context, node, kOutputStateTensor); - TfLiteTensor* cell_state = GetOutput(context, node, kCellStateTensor); - TfLiteTensor* output = GetOutput(context, node, kOutputTensor); - +TfLiteStatus EvalFloat( + const TfLiteTensor* input, const TfLiteTensor* input_to_input_weights, + const TfLiteTensor* input_to_forget_weights, + const TfLiteTensor* input_to_cell_weights, + const TfLiteTensor* input_to_output_weights, + const TfLiteTensor* recurrent_to_input_weights, + const TfLiteTensor* recurrent_to_forget_weights, + const TfLiteTensor* recurrent_to_cell_weights, + const TfLiteTensor* recurrent_to_output_weights, + const TfLiteTensor* cell_to_input_weights, + const TfLiteTensor* cell_to_forget_weights, + const TfLiteTensor* cell_to_output_weights, + const TfLiteTensor* input_gate_bias, const TfLiteTensor* forget_gate_bias, + const TfLiteTensor* cell_bias, const TfLiteTensor* output_gate_bias, + const TfLiteTensor* projection_weights, const TfLiteTensor* projection_bias, + const TfLiteLSTMParams* params, TfLiteTensor* scratch_buffer, + TfLiteTensor* output_state, TfLiteTensor* cell_state, + TfLiteTensor* output) { const int max_time = input->dims->data[0]; const int n_batch = input->dims->data[1]; const int n_input = input->dims->data[2]; @@ -380,8 +464,6 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { const bool use_cifg = (input_to_input_weights == nullptr); const bool use_peephole = (cell_to_output_weights != nullptr); - // Index the scratch buffers pointers to the global scratch buffer. - TfLiteTensor* scratch_buffer = GetTemporary(context, node, /*index=*/0); float* input_gate_scratch = nullptr; float* cell_scratch = nullptr; float* forget_gate_scratch = nullptr; @@ -432,6 +514,7 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { float* output_state_ptr = output_state->data.f; float* cell_state_ptr = cell_state->data.f; + // Feed the sequence into the LSTM step-by-step. for (int t = 0; t < max_time; t++) { const float* input_ptr_batch = input->data.f + t * n_batch * n_input; float* output_ptr_batch = output->data.f + t * n_batch * n_output; @@ -452,6 +535,262 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { return kTfLiteOk; } +TfLiteStatus EvalHybrid( + const TfLiteTensor* input, const TfLiteTensor* input_to_input_weights, + const TfLiteTensor* input_to_forget_weights, + const TfLiteTensor* input_to_cell_weights, + const TfLiteTensor* input_to_output_weights, + const TfLiteTensor* recurrent_to_input_weights, + const TfLiteTensor* recurrent_to_forget_weights, + const TfLiteTensor* recurrent_to_cell_weights, + const TfLiteTensor* recurrent_to_output_weights, + const TfLiteTensor* cell_to_input_weights, + const TfLiteTensor* cell_to_forget_weights, + const TfLiteTensor* cell_to_output_weights, + const TfLiteTensor* input_gate_bias, const TfLiteTensor* forget_gate_bias, + const TfLiteTensor* cell_bias, const TfLiteTensor* output_gate_bias, + const TfLiteTensor* projection_weights, const TfLiteTensor* projection_bias, + const TfLiteLSTMParams* params, TfLiteTensor* scratch_buffer, + TfLiteTensor* scaling_factors, TfLiteTensor* prod_scaling_factors, + TfLiteTensor* recovered_cell_weights, TfLiteTensor* input_quantized, + TfLiteTensor* output_state_quantized, TfLiteTensor* cell_state_quantized, + TfLiteTensor* output_state, TfLiteTensor* cell_state, + TfLiteTensor* output) { + const int max_time = input->dims->data[0]; + const int n_batch = input->dims->data[1]; + const int n_input = input->dims->data[2]; + // n_cell and n_output will be the same size when there is no projection. + const int n_cell = input_to_output_weights->dims->data[0]; + const int n_output = recurrent_to_output_weights->dims->data[1]; + + // Since we have already checked that weights are all there or none, we can + // check the existence of only one to get the condition. + const bool use_cifg = (input_to_input_weights == nullptr); + const bool use_peephole = (cell_to_output_weights != nullptr); + + float* input_gate_scratch = nullptr; + float* cell_scratch = nullptr; + float* forget_gate_scratch = nullptr; + float* output_gate_scratch = nullptr; + if (use_cifg) { + cell_scratch = scratch_buffer->data.f; + forget_gate_scratch = scratch_buffer->data.f + n_cell * n_batch; + output_gate_scratch = scratch_buffer->data.f + 2 * n_cell * n_batch; + } else { + input_gate_scratch = scratch_buffer->data.f; + cell_scratch = scratch_buffer->data.f + n_cell * n_batch; + forget_gate_scratch = scratch_buffer->data.f + 2 * n_cell * n_batch; + output_gate_scratch = scratch_buffer->data.f + 3 * n_cell * n_batch; + } + + // Check optional tensors, the respective pointers can be null. + int8_t* input_to_input_weights_ptr = nullptr; + float input_to_input_weights_scale = 1.0f; + int8_t* recurrent_to_input_weights_ptr = nullptr; + float recurrent_to_input_weights_scale = 1.0f; + float* input_gate_bias_ptr = nullptr; + if (!use_cifg) { + input_to_input_weights_ptr = + reinterpret_cast(input_to_input_weights->data.uint8); + recurrent_to_input_weights_ptr = + reinterpret_cast(recurrent_to_input_weights->data.uint8); + input_gate_bias_ptr = input_gate_bias->data.f; + input_to_input_weights_scale = input_to_input_weights->params.scale; + recurrent_to_input_weights_scale = recurrent_to_input_weights->params.scale; + } + + int8_t* cell_to_input_weights_ptr = nullptr; + int8_t* cell_to_forget_weights_ptr = nullptr; + int8_t* cell_to_output_weights_ptr = nullptr; + float cell_to_input_weights_scale = 1.0f; + float cell_to_forget_weights_scale = 1.0f; + float cell_to_output_weights_scale = 1.0f; + if (use_peephole) { + if (!use_cifg) { + cell_to_input_weights_ptr = + reinterpret_cast(cell_to_input_weights->data.uint8); + cell_to_input_weights_scale = cell_to_input_weights->params.scale; + } + cell_to_forget_weights_ptr = + reinterpret_cast(cell_to_forget_weights->data.uint8); + cell_to_output_weights_ptr = + reinterpret_cast(cell_to_output_weights->data.uint8); + cell_to_forget_weights_scale = cell_to_forget_weights->params.scale; + cell_to_output_weights_scale = cell_to_output_weights->params.scale; + } + + const int8_t* projection_weights_ptr = + (projection_weights == nullptr) + ? nullptr + : reinterpret_cast(projection_weights->data.uint8); + float projection_weights_scale = + (projection_weights == nullptr) ? 1.0f : projection_weights->params.scale; + const float* projection_bias_ptr = + (projection_bias == nullptr) ? nullptr : projection_bias->data.f; + + // Required tensors, pointers are non-null. + const int8_t* input_to_forget_weights_ptr = + reinterpret_cast(input_to_forget_weights->data.uint8); + const float input_to_forget_weights_scale = + input_to_forget_weights->params.scale; + const int8_t* input_to_cell_weights_ptr = + reinterpret_cast(input_to_cell_weights->data.uint8); + const float input_to_cell_weights_scale = input_to_cell_weights->params.scale; + const int8_t* input_to_output_weights_ptr = + reinterpret_cast(input_to_output_weights->data.uint8); + const float input_to_output_weights_scale = + input_to_output_weights->params.scale; + const int8_t* recurrent_to_forget_weights_ptr = + reinterpret_cast(recurrent_to_forget_weights->data.uint8); + const float recurrent_to_forget_weights_scale = + recurrent_to_forget_weights->params.scale; + const int8_t* recurrent_to_cell_weights_ptr = + reinterpret_cast(recurrent_to_cell_weights->data.uint8); + const float recurrent_to_cell_weights_scale = + recurrent_to_cell_weights->params.scale; + const int8_t* recurrent_to_output_weights_ptr = + reinterpret_cast(recurrent_to_output_weights->data.uint8); + const float recurrent_to_output_weights_scale = + recurrent_to_output_weights->params.scale; + const float* forget_gate_bias_ptr = forget_gate_bias->data.f; + const float* cell_bias_ptr = cell_bias->data.f; + const float* output_gate_bias_ptr = output_gate_bias->data.f; + + float* output_state_ptr = output_state->data.f; + float* cell_state_ptr = cell_state->data.f; + + // Temporary storage for quantized values and scaling factors. + int8_t* quantized_input_ptr = + reinterpret_cast(input_quantized->data.uint8); + int8_t* quantized_output_state_ptr = + reinterpret_cast(output_state_quantized->data.uint8); + int8_t* quantized_cell_state_ptr = + reinterpret_cast(cell_state_quantized->data.uint8); + float* scaling_factors_ptr = scaling_factors->data.f; + float* prod_scaling_factors_ptr = prod_scaling_factors->data.f; + float* recovered_cell_weights_ptr = recovered_cell_weights->data.f; + + // Feed the sequence into the LSTM step-by-step. + for (int t = 0; t < max_time; t++) { + const float* input_ptr_batch = input->data.f + t * n_batch * n_input; + float* output_ptr_batch = output->data.f + t * n_batch * n_output; + + kernel_utils::LstmStep( + input_ptr_batch, input_to_input_weights_ptr, + input_to_input_weights_scale, input_to_forget_weights_ptr, + input_to_forget_weights_scale, input_to_cell_weights_ptr, + input_to_cell_weights_scale, input_to_output_weights_ptr, + input_to_output_weights_scale, recurrent_to_input_weights_ptr, + recurrent_to_input_weights_scale, recurrent_to_forget_weights_ptr, + recurrent_to_forget_weights_scale, recurrent_to_cell_weights_ptr, + recurrent_to_cell_weights_scale, recurrent_to_output_weights_ptr, + recurrent_to_output_weights_scale, cell_to_input_weights_ptr, + cell_to_input_weights_scale, cell_to_forget_weights_ptr, + cell_to_forget_weights_scale, cell_to_output_weights_ptr, + cell_to_output_weights_scale, input_gate_bias_ptr, forget_gate_bias_ptr, + cell_bias_ptr, output_gate_bias_ptr, projection_weights_ptr, + projection_weights_scale, projection_bias_ptr, params, n_batch, n_cell, + n_input, n_output, input_gate_scratch, forget_gate_scratch, + cell_scratch, output_gate_scratch, scaling_factors_ptr, + prod_scaling_factors_ptr, recovered_cell_weights_ptr, + quantized_input_ptr, quantized_output_state_ptr, + quantized_cell_state_ptr, output_state_ptr, cell_state_ptr, + output_ptr_batch); + } + return kTfLiteOk; +} + +TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) { + auto* params = reinterpret_cast(node->builtin_data); + const TfLiteTensor* input = GetInput(context, node, kInputTensor); + + const TfLiteTensor* input_to_input_weights = + GetOptionalInputTensor(context, node, kInputToInputWeightsTensor); + const TfLiteTensor* input_to_forget_weights = + GetInput(context, node, kInputToForgetWeightsTensor); + const TfLiteTensor* input_to_cell_weights = + GetInput(context, node, kInputToCellWeightsTensor); + const TfLiteTensor* input_to_output_weights = + GetInput(context, node, kInputToOutputWeightsTensor); + + const TfLiteTensor* recurrent_to_input_weights = + GetOptionalInputTensor(context, node, kRecurrentToInputWeightsTensor); + const TfLiteTensor* recurrent_to_forget_weights = + GetInput(context, node, kRecurrentToForgetWeightsTensor); + const TfLiteTensor* recurrent_to_cell_weights = + GetInput(context, node, kRecurrentToCellWeightsTensor); + const TfLiteTensor* recurrent_to_output_weights = + GetInput(context, node, kRecurrentToOutputWeightsTensor); + + const TfLiteTensor* cell_to_input_weights = + GetOptionalInputTensor(context, node, kCellToInputWeightsTensor); + const TfLiteTensor* cell_to_forget_weights = + GetOptionalInputTensor(context, node, kCellToForgetWeightsTensor); + const TfLiteTensor* cell_to_output_weights = + GetOptionalInputTensor(context, node, kCellToOutputWeightsTensor); + + const TfLiteTensor* input_gate_bias = + GetOptionalInputTensor(context, node, kInputGateBiasTensor); + const TfLiteTensor* forget_gate_bias = + GetInput(context, node, kForgetGateBiasTensor); + const TfLiteTensor* cell_bias = GetInput(context, node, kCellGateBiasTensor); + const TfLiteTensor* output_gate_bias = + GetInput(context, node, kOutputGateBiasTensor); + + const TfLiteTensor* projection_weights = + GetOptionalInputTensor(context, node, kProjectionWeightsTensor); + const TfLiteTensor* projection_bias = + GetOptionalInputTensor(context, node, kProjectionBiasTensor); + + // Index the scratch buffers pointers to the global scratch buffer. + TfLiteTensor* scratch_buffer = GetTemporary(context, node, /*index=*/0); + + TfLiteTensor* output_state = GetOutput(context, node, kOutputStateTensor); + TfLiteTensor* cell_state = GetOutput(context, node, kCellStateTensor); + TfLiteTensor* output = GetOutput(context, node, kOutputTensor); + + switch (input_to_output_weights->type) { + case kTfLiteFloat32: { + return EvalFloat(input, input_to_input_weights, input_to_forget_weights, + input_to_cell_weights, input_to_output_weights, + recurrent_to_input_weights, recurrent_to_forget_weights, + recurrent_to_cell_weights, recurrent_to_output_weights, + cell_to_input_weights, cell_to_forget_weights, + cell_to_output_weights, input_gate_bias, + forget_gate_bias, cell_bias, output_gate_bias, + projection_weights, projection_bias, params, + scratch_buffer, output_state, cell_state, output); + } + case kTfLiteUInt8: { + TfLiteTensor* input_quantized = GetTemporary(context, node, /*index=*/1); + TfLiteTensor* output_state_quantized = + GetTemporary(context, node, /*index=*/2); + TfLiteTensor* cell_state_quantized = + GetTemporary(context, node, /*index=*/3); + TfLiteTensor* scaling_factors = GetTemporary(context, node, /*index=*/4); + TfLiteTensor* prod_scaling_factors = + GetTemporary(context, node, /*index=*/5); + TfLiteTensor* recovered_cell_weights = + GetTemporary(context, node, /*index=*/6); + return EvalHybrid( + input, input_to_input_weights, input_to_forget_weights, + input_to_cell_weights, input_to_output_weights, + recurrent_to_input_weights, recurrent_to_forget_weights, + recurrent_to_cell_weights, recurrent_to_output_weights, + cell_to_input_weights, cell_to_forget_weights, cell_to_output_weights, + input_gate_bias, forget_gate_bias, cell_bias, output_gate_bias, + projection_weights, projection_bias, params, scratch_buffer, + scaling_factors, prod_scaling_factors, recovered_cell_weights, + input_quantized, output_state_quantized, cell_state_quantized, + output_state, cell_state, output); + } + default: + context->ReportError(context, "Type %d is not currently supported.", + input_to_output_weights->type); + return kTfLiteError; + } + return kTfLiteOk; +} } // namespace unidirectional_sequence_lstm TfLiteRegistration* Register_UNIDIRECTIONAL_SEQUENCE_LSTM() { diff --git a/tensorflow/contrib/lite/kernels/unidirectional_sequence_lstm_test.cc b/tensorflow/contrib/lite/kernels/unidirectional_sequence_lstm_test.cc index 5881ced7c7a616ef2c24db60892cbbf9eec7c42e..de38bdef6fd1b019c7790a664b29cd45d29e5dcc 100644 --- a/tensorflow/contrib/lite/kernels/unidirectional_sequence_lstm_test.cc +++ b/tensorflow/contrib/lite/kernels/unidirectional_sequence_lstm_test.cc @@ -14,7 +14,6 @@ limitations under the License. ==============================================================================*/ // Unit test for TFLite Sequential LSTM op. -#include #include #include @@ -37,7 +36,8 @@ class UnidirectionalLSTMOpModel : public SingleOpModel { bool use_peephole, bool use_projection_weights, bool use_projection_bias, float cell_clip, float proj_clip, - const std::vector>& input_shapes) + const std::vector>& input_shapes, + const TensorType& weights_type = TensorType_FLOAT32) : n_batch_(n_batch), n_input_(n_input), n_cell_(n_cell), @@ -48,31 +48,31 @@ class UnidirectionalLSTMOpModel : public SingleOpModel { if (use_cifg) { input_to_input_weights_ = AddNullInput(); } else { - input_to_input_weights_ = AddInput(TensorType_FLOAT32); + input_to_input_weights_ = AddInput(weights_type); } - input_to_forget_weights_ = AddInput(TensorType_FLOAT32); - input_to_cell_weights_ = AddInput(TensorType_FLOAT32); - input_to_output_weights_ = AddInput(TensorType_FLOAT32); + input_to_forget_weights_ = AddInput(weights_type); + input_to_cell_weights_ = AddInput(weights_type); + input_to_output_weights_ = AddInput(weights_type); if (use_cifg) { recurrent_to_input_weights_ = AddNullInput(); } else { - recurrent_to_input_weights_ = AddInput(TensorType_FLOAT32); + recurrent_to_input_weights_ = AddInput(weights_type); } - recurrent_to_forget_weights_ = AddInput(TensorType_FLOAT32); - recurrent_to_cell_weights_ = AddInput(TensorType_FLOAT32); - recurrent_to_output_weights_ = AddInput(TensorType_FLOAT32); + recurrent_to_forget_weights_ = AddInput(weights_type); + recurrent_to_cell_weights_ = AddInput(weights_type); + recurrent_to_output_weights_ = AddInput(weights_type); if (use_peephole) { if (use_cifg) { cell_to_input_weights_ = AddNullInput(); } else { - cell_to_input_weights_ = AddInput(TensorType_FLOAT32); + cell_to_input_weights_ = AddInput(weights_type); } - cell_to_forget_weights_ = AddInput(TensorType_FLOAT32); - cell_to_output_weights_ = AddInput(TensorType_FLOAT32); + cell_to_forget_weights_ = AddInput(weights_type); + cell_to_output_weights_ = AddInput(weights_type); } else { cell_to_input_weights_ = AddNullInput(); cell_to_forget_weights_ = AddNullInput(); @@ -89,7 +89,7 @@ class UnidirectionalLSTMOpModel : public SingleOpModel { output_gate_bias_ = AddInput(TensorType_FLOAT32); if (use_projection_weights) { - projection_weights_ = AddInput(TensorType_FLOAT32); + projection_weights_ = AddInput(weights_type); if (use_projection_bias) { projection_bias_ = AddInput(TensorType_FLOAT32); } else { @@ -196,8 +196,9 @@ class UnidirectionalLSTMOpModel : public SingleOpModel { zero_buffer.get() + zero_buffer_size); } - void SetInput(int offset, float* begin, float* end) { - PopulateTensor(input_, offset, begin, end); + void SetInput(int offset, const float* begin, const float* end) { + PopulateTensor(input_, offset, const_cast(begin), + const_cast(end)); } std::vector GetOutput() { return ExtractVector(output_); } @@ -208,7 +209,7 @@ class UnidirectionalLSTMOpModel : public SingleOpModel { int num_batches() { return n_batch_; } int sequence_length() { return sequence_length_; } - private: + protected: int input_; int input_to_input_weights_; int input_to_forget_weights_; @@ -243,7 +244,183 @@ class UnidirectionalLSTMOpModel : public SingleOpModel { int sequence_length_; }; -TEST(LSTMOpTest, BlackBoxTestNoCifgNoPeepholeNoProjectionNoClipping) { +// The hybrid model has quantized weights. +class HybridUnidirectionalLSTMOpModel : public UnidirectionalLSTMOpModel { + public: + HybridUnidirectionalLSTMOpModel( + int n_batch, int n_input, int n_cell, int n_output, int sequence_length, + bool use_cifg, bool use_peephole, bool use_projection_weights, + bool use_projection_bias, float cell_clip, float proj_clip, + const std::vector>& input_shapes) + : UnidirectionalLSTMOpModel( + n_batch, n_input, n_cell, n_output, sequence_length, use_cifg, + use_peephole, use_projection_weights, use_projection_bias, + cell_clip, proj_clip, input_shapes, TensorType_UINT8) {} + + void SetInputToInputWeights(std::initializer_list f) { + SymmetricQuantizeAndPopulate(input_to_input_weights_, f); + } + + void SetInputToForgetWeights(std::initializer_list f) { + SymmetricQuantizeAndPopulate(input_to_forget_weights_, f); + } + + void SetInputToCellWeights(std::initializer_list f) { + SymmetricQuantizeAndPopulate(input_to_cell_weights_, f); + } + + void SetInputToOutputWeights(std::initializer_list f) { + SymmetricQuantizeAndPopulate(input_to_output_weights_, f); + } + + void SetRecurrentToInputWeights(std::initializer_list f) { + SymmetricQuantizeAndPopulate(recurrent_to_input_weights_, f); + } + + void SetRecurrentToForgetWeights(std::initializer_list f) { + SymmetricQuantizeAndPopulate(recurrent_to_forget_weights_, f); + } + + void SetRecurrentToCellWeights(std::initializer_list f) { + SymmetricQuantizeAndPopulate(recurrent_to_cell_weights_, f); + } + + void SetRecurrentToOutputWeights(std::initializer_list f) { + SymmetricQuantizeAndPopulate(recurrent_to_output_weights_, f); + } + + void SetCellToInputWeights(std::initializer_list f) { + SymmetricQuantizeAndPopulate(cell_to_input_weights_, f); + } + + void SetCellToForgetWeights(std::initializer_list f) { + SymmetricQuantizeAndPopulate(cell_to_forget_weights_, f); + } + + void SetCellToOutputWeights(std::initializer_list f) { + SymmetricQuantizeAndPopulate(cell_to_output_weights_, f); + } + + void SetProjectionWeights(std::initializer_list f) { + SymmetricQuantizeAndPopulate(projection_weights_, f); + } +}; + +class BaseLstmTest : public ::testing::Test { + protected: + // Weights of the LSTM model. Some are optional. + std::initializer_list input_to_input_weights_; + std::initializer_list input_to_cell_weights_; + std::initializer_list input_to_forget_weights_; + std::initializer_list input_to_output_weights_; + std::initializer_list input_gate_bias_; + std::initializer_list cell_gate_bias_; + std::initializer_list forget_gate_bias_; + std::initializer_list output_gate_bias_; + std::initializer_list recurrent_to_input_weights_; + std::initializer_list recurrent_to_cell_weights_; + std::initializer_list recurrent_to_forget_weights_; + std::initializer_list recurrent_to_output_weights_; + std::initializer_list cell_to_input_weights_; + std::initializer_list cell_to_forget_weights_; + std::initializer_list cell_to_output_weights_; + std::initializer_list projection_weights_; + + // LSTM input is stored as num_batch x num_inputs vector. + std::vector> lstm_input_; + // LSTM output is stored as num_batch x num_outputs vector. + std::vector> lstm_golden_output_; + + // Compares output up to tolerance to the result of the lstm given the input. + void VerifyGoldens(const std::vector>& input, + const std::vector>& output, + UnidirectionalLSTMOpModel* lstm, float tolerance = 1e-5) { + const int num_batches = input.size(); + EXPECT_GT(num_batches, 0); + const int num_inputs = lstm->num_inputs(); + EXPECT_GT(num_inputs, 0); + const int input_sequence_size = input[0].size() / num_inputs; + EXPECT_GT(input_sequence_size, 0); + // Feed the whole sequence as input. + for (int i = 0; i < input_sequence_size; ++i) { + for (int b = 0; b < num_batches; ++b) { + const float* batch_start = input[b].data() + i * num_inputs; + const float* batch_end = batch_start + num_inputs; + + lstm->SetInput(((i * num_batches) + b) * lstm->num_inputs(), + batch_start, batch_end); + } + } + + lstm->Invoke(); + + const int num_outputs = lstm->num_outputs(); + EXPECT_GT(num_outputs, 0); + std::vector expected; + for (int i = 0; i < input_sequence_size; ++i) { + for (int b = 0; b < num_batches; ++b) { + const float* golden_start_batch = output[b].data() + i * num_outputs; + const float* golden_end_batch = golden_start_batch + num_outputs; + + expected.insert(expected.end(), golden_start_batch, golden_end_batch); + } + } + + EXPECT_THAT(lstm->GetOutput(), + ElementsAreArray(ArrayFloatNear(expected, tolerance))); + } +}; + +class NoCifgNoPeepholeNoProjectionNoClippingLstmTest : public BaseLstmTest { + void SetUp() override { + input_to_input_weights_ = {-0.45018822, -0.02338299, -0.0870589, + -0.34550029, 0.04266912, -0.15680569, + -0.34856534, 0.43890524}; + input_to_cell_weights_ = {-0.50013041, 0.1370284, 0.11810488, 0.2013163, + -0.20583314, 0.44344562, 0.22077113, -0.29909778}; + input_to_forget_weights_ = {0.09701663, 0.20334584, -0.50592935, + -0.31343272, -0.40032279, 0.44781327, + 0.01387155, -0.35593212}; + input_to_output_weights_ = {-0.25065863, -0.28290087, 0.04613829, + 0.40525138, 0.44272184, 0.03897077, + -0.1556896, 0.19487578}; + input_gate_bias_ = {0., 0., 0., 0.}; + cell_gate_bias_ = {0., 0., 0., 0.}; + forget_gate_bias_ = {1., 1., 1., 1.}; + output_gate_bias_ = {0., 0., 0., 0.}; + + recurrent_to_input_weights_ = { + -0.0063535, -0.2042388, 0.31454784, -0.35746509, + 0.28902304, 0.08183324, -0.16555229, 0.02286911, + -0.13566875, 0.03034258, 0.48091322, -0.12528998, + 0.24077177, -0.51332325, -0.33502164, 0.10629296}; + + recurrent_to_cell_weights_ = { + -0.3407414, 0.24443203, -0.2078532, 0.26320225, + 0.05695659, -0.00123841, -0.4744786, -0.35869038, + -0.06418842, -0.13502428, -0.501764, 0.22830659, + -0.46367589, 0.26016325, -0.03894562, -0.16368064}; + + recurrent_to_forget_weights_ = { + -0.48684245, -0.06655136, 0.42224967, 0.2112639, + 0.27654213, 0.20864892, -0.07646349, 0.45877004, + 0.00141793, -0.14609534, 0.36447752, 0.09196436, + 0.28053468, 0.01560611, -0.20127171, -0.01140004}; + + recurrent_to_output_weights_ = { + 0.43385774, -0.17194885, 0.2718237, 0.09215671, + 0.24107647, -0.39835793, 0.18212086, 0.01301402, + 0.48572797, -0.50656658, 0.20047462, -0.20607421, + -0.51818722, -0.15390486, 0.0468148, 0.39922136}; + + lstm_input_ = {{2., 3., 3., 4., 1., 1.}}; + lstm_golden_output_ = {{-0.02973187, 0.1229473, 0.20885126, -0.15358765, + -0.03716109, 0.12507336, 0.41193449, -0.20860538, + -0.15053082, 0.09120187, 0.24278517, -0.12222792}}; + } +}; + +TEST_F(NoCifgNoPeepholeNoProjectionNoClippingLstmTest, LstmBlackBoxTest) { const int n_batch = 1; const int n_input = 2; // n_cell and n_output have the same size when there is no projection. @@ -252,9 +429,11 @@ TEST(LSTMOpTest, BlackBoxTestNoCifgNoPeepholeNoProjectionNoClipping) { const int sequence_length = 3; UnidirectionalLSTMOpModel lstm( - n_batch, n_input, n_cell, n_output, sequence_length, /*use_cifg=*/false, - /*use_peephole=*/false, /*use_projection_weights=*/false, - /*use_projection_bias=*/false, /*cell_clip=*/0.0, /*proj_clip=*/0.0, + n_batch, n_input, n_cell, n_output, sequence_length, + /*use_cifg=*/false, /*use_peephole=*/false, + /*use_projection_weights=*/false, + /*use_projection_bias=*/false, + /*cell_clip=*/0.0, /*proj_clip=*/0.0, { {sequence_length, n_batch, n_input}, // input tensor @@ -281,77 +460,138 @@ TEST(LSTMOpTest, BlackBoxTestNoCifgNoPeepholeNoProjectionNoClipping) { {0}, // projection_bias tensor }); - lstm.SetInputToInputWeights({-0.45018822, -0.02338299, -0.0870589, - -0.34550029, 0.04266912, -0.15680569, - -0.34856534, 0.43890524}); + lstm.SetInputToInputWeights(input_to_input_weights_); + lstm.SetInputToCellWeights(input_to_cell_weights_); + lstm.SetInputToForgetWeights(input_to_forget_weights_); + lstm.SetInputToOutputWeights(input_to_output_weights_); - lstm.SetInputToCellWeights({-0.50013041, 0.1370284, 0.11810488, 0.2013163, - -0.20583314, 0.44344562, 0.22077113, - -0.29909778}); + lstm.SetInputGateBias(input_gate_bias_); + lstm.SetCellBias(cell_gate_bias_); + lstm.SetForgetGateBias(forget_gate_bias_); + lstm.SetOutputGateBias(output_gate_bias_); - lstm.SetInputToForgetWeights({0.09701663, 0.20334584, -0.50592935, - -0.31343272, -0.40032279, 0.44781327, - 0.01387155, -0.35593212}); + lstm.SetRecurrentToInputWeights(recurrent_to_input_weights_); + lstm.SetRecurrentToCellWeights(recurrent_to_cell_weights_); + lstm.SetRecurrentToForgetWeights(recurrent_to_forget_weights_); + lstm.SetRecurrentToOutputWeights(recurrent_to_output_weights_); + + // Resetting cell_state and output_state + lstm.ResetCellState(); + lstm.ResetOutputState(); + + VerifyGoldens(lstm_input_, lstm_golden_output_, &lstm); +} - lstm.SetInputToOutputWeights({-0.25065863, -0.28290087, 0.04613829, - 0.40525138, 0.44272184, 0.03897077, -0.1556896, - 0.19487578}); +TEST_F(NoCifgNoPeepholeNoProjectionNoClippingLstmTest, HybridLstmBlackBoxTest) { + const int n_batch = 1; + const int n_input = 2; + // n_cell and n_output have the same size when there is no projection. + const int n_cell = 4; + const int n_output = 4; + const int sequence_length = 3; - lstm.SetInputGateBias({0., 0., 0., 0.}); + HybridUnidirectionalLSTMOpModel lstm( + n_batch, n_input, n_cell, n_output, sequence_length, + /*use_cifg=*/false, /*use_peephole=*/false, + /*use_projection_weights=*/false, + /*use_projection_bias=*/false, /*cell_clip=*/0.0, /*proj_clip=*/0.0, + { + {sequence_length, n_batch, n_input}, // input tensor - lstm.SetCellBias({0., 0., 0., 0.}); + {n_cell, n_input}, // input_to_input_weight tensor + {n_cell, n_input}, // input_to_forget_weight tensor + {n_cell, n_input}, // input_to_cell_weight tensor + {n_cell, n_input}, // input_to_output_weight tensor - lstm.SetForgetGateBias({1., 1., 1., 1.}); + {n_cell, n_output}, // recurrent_to_input_weight tensor + {n_cell, n_output}, // recurrent_to_forget_weight tensor + {n_cell, n_output}, // recurrent_to_cell_weight tensor + {n_cell, n_output}, // recurrent_to_output_weight tensor - lstm.SetOutputGateBias({0., 0., 0., 0.}); + {0}, // cell_to_input_weight tensor + {0}, // cell_to_forget_weight tensor + {0}, // cell_to_output_weight tensor - lstm.SetRecurrentToInputWeights( - {-0.0063535, -0.2042388, 0.31454784, -0.35746509, 0.28902304, 0.08183324, - -0.16555229, 0.02286911, -0.13566875, 0.03034258, 0.48091322, - -0.12528998, 0.24077177, -0.51332325, -0.33502164, 0.10629296}); + {n_cell}, // input_gate_bias tensor + {n_cell}, // forget_gate_bias tensor + {n_cell}, // cell_bias tensor + {n_cell}, // output_gate_bias tensor - lstm.SetRecurrentToCellWeights( - {-0.3407414, 0.24443203, -0.2078532, 0.26320225, 0.05695659, -0.00123841, - -0.4744786, -0.35869038, -0.06418842, -0.13502428, -0.501764, 0.22830659, - -0.46367589, 0.26016325, -0.03894562, -0.16368064}); + {0, 0}, // projection_weight tensor + {0}, // projection_bias tensor + }); - lstm.SetRecurrentToForgetWeights( - {-0.48684245, -0.06655136, 0.42224967, 0.2112639, 0.27654213, 0.20864892, - -0.07646349, 0.45877004, 0.00141793, -0.14609534, 0.36447752, 0.09196436, - 0.28053468, 0.01560611, -0.20127171, -0.01140004}); + lstm.SetInputToInputWeights(input_to_input_weights_); + lstm.SetInputToCellWeights(input_to_cell_weights_); + lstm.SetInputToForgetWeights(input_to_forget_weights_); + lstm.SetInputToOutputWeights(input_to_output_weights_); - lstm.SetRecurrentToOutputWeights( - {0.43385774, -0.17194885, 0.2718237, 0.09215671, 0.24107647, -0.39835793, - 0.18212086, 0.01301402, 0.48572797, -0.50656658, 0.20047462, -0.20607421, - -0.51818722, -0.15390486, 0.0468148, 0.39922136}); + lstm.SetInputGateBias(input_gate_bias_); + lstm.SetCellBias(cell_gate_bias_); + lstm.SetForgetGateBias(forget_gate_bias_); + lstm.SetOutputGateBias(output_gate_bias_); - // Input should have n_input * sequence_length many values. - static float lstm_input[] = {2., 3., 3., 4., 1., 1.}; - static float lstm_golden_output[] = {-0.02973187, 0.1229473, 0.20885126, - -0.15358765, -0.03716109, 0.12507336, - 0.41193449, -0.20860538, -0.15053082, - 0.09120187, 0.24278517, -0.12222792}; + lstm.SetRecurrentToInputWeights(recurrent_to_input_weights_); + lstm.SetRecurrentToCellWeights(recurrent_to_cell_weights_); + lstm.SetRecurrentToForgetWeights(recurrent_to_forget_weights_); + lstm.SetRecurrentToOutputWeights(recurrent_to_output_weights_); // Resetting cell_state and output_state lstm.ResetCellState(); lstm.ResetOutputState(); - float* batch0_start = lstm_input; - float* batch0_end = batch0_start + lstm.num_inputs() * lstm.sequence_length(); + VerifyGoldens(lstm_input_, lstm_golden_output_, &lstm, + /*tolerance=*/0.0157651); +} - lstm.SetInput(0, batch0_start, batch0_end); +class CifgPeepholeNoProjectionNoClippingLstmTest : public BaseLstmTest { + void SetUp() override { + input_to_cell_weights_ = {-0.49770179, -0.27711356, -0.09624726, + 0.05100781, 0.04717243, 0.48944736, + -0.38535351, -0.17212132}; - lstm.Invoke(); + input_to_forget_weights_ = {-0.55291498, -0.42866567, 0.13056988, + -0.3633365, -0.22755712, 0.28253698, + 0.24407166, 0.33826375}; - float* golden_start = lstm_golden_output; - float* golden_end = - golden_start + lstm.num_outputs() * lstm.sequence_length(); - std::vector expected; - expected.insert(expected.end(), golden_start, golden_end); - EXPECT_THAT(lstm.GetOutput(), ElementsAreArray(ArrayFloatNear(expected))); -} + input_to_output_weights_ = {0.10725588, -0.02335852, -0.55932593, + -0.09426838, -0.44257352, 0.54939759, + 0.01533556, 0.42751634}; + cell_gate_bias_ = {0., 0., 0., 0.}; + forget_gate_bias_ = {1., 1., 1., 1.}; + output_gate_bias_ = {0., 0., 0., 0.}; + + recurrent_to_cell_weights_ = { + 0.54066205, -0.32668582, -0.43562764, -0.56094903, + 0.42957711, 0.01841056, -0.32764608, -0.33027974, + -0.10826075, 0.20675004, 0.19069612, -0.03026325, + -0.54532051, 0.33003211, 0.44901288, 0.21193194}; + + recurrent_to_forget_weights_ = { + -0.13832897, -0.0515101, -0.2359007, -0.16661474, + -0.14340827, 0.36986142, 0.23414481, 0.55899, + 0.10798943, -0.41174671, 0.17751795, -0.34484994, + -0.35874045, -0.11352962, 0.27268326, 0.54058349}; + + recurrent_to_output_weights_ = { + 0.41613156, 0.42610586, -0.16495961, -0.5663873, + 0.30579174, -0.05115908, -0.33941799, 0.23364776, + 0.11178309, 0.09481031, -0.26424935, 0.46261835, + 0.50248802, 0.26114327, -0.43736315, 0.33149987}; + + cell_to_forget_weights_ = {0.47485286, -0.51955009, -0.24458408, + 0.31544167}; + cell_to_output_weights_ = {-0.17135078, 0.82760304, 0.85573703, + -0.77109635}; + + lstm_input_ = {{2., 3., 3., 4., 1., 1.}}; + lstm_golden_output_ = {{-0.36444446, -0.00352185, 0.12886585, -0.05163646, + -0.42312205, -0.01218222, 0.24201041, -0.08124574, + -0.358325, -0.04621704, 0.21641694, -0.06471302}}; + } +}; -TEST(LSTMOpTest, BlackBoxTestWithCifgWithPeepholeNoProjectionNoClipping) { +TEST_F(CifgPeepholeNoProjectionNoClippingLstmTest, LstmBlackBoxTest) { const int n_batch = 1; const int n_input = 2; // n_cell and n_output have the same size when there is no projection. @@ -360,9 +600,11 @@ TEST(LSTMOpTest, BlackBoxTestWithCifgWithPeepholeNoProjectionNoClipping) { const int sequence_length = 3; UnidirectionalLSTMOpModel lstm( - n_batch, n_input, n_cell, n_output, sequence_length, /*use_cifg=*/true, - /*use_peephole=*/true, /*use_projection_weights=*/false, - /*use_projection_bias=*/false, /*cell_clip=*/0.0, /*proj_clip=*/0.0, + n_batch, n_input, n_cell, n_output, sequence_length, + /*use_cifg=*/true, /*use_peephole=*/true, + /*use_projection_weights=*/false, + /*use_projection_bias=*/false, + /*cell_clip=*/0.0, /*proj_clip=*/0.0, { {sequence_length, n_batch, n_input}, // input tensor @@ -389,71 +631,690 @@ TEST(LSTMOpTest, BlackBoxTestWithCifgWithPeepholeNoProjectionNoClipping) { {0}, // projection_bias tensor }); - lstm.SetInputToCellWeights({-0.49770179, -0.27711356, -0.09624726, 0.05100781, - 0.04717243, 0.48944736, -0.38535351, - -0.17212132}); + lstm.SetInputToCellWeights(input_to_cell_weights_); + lstm.SetInputToForgetWeights(input_to_forget_weights_); + lstm.SetInputToOutputWeights(input_to_output_weights_); - lstm.SetInputToForgetWeights({-0.55291498, -0.42866567, 0.13056988, - -0.3633365, -0.22755712, 0.28253698, 0.24407166, - 0.33826375}); + lstm.SetCellBias(cell_gate_bias_); + lstm.SetForgetGateBias(forget_gate_bias_); + lstm.SetOutputGateBias(output_gate_bias_); - lstm.SetInputToOutputWeights({0.10725588, -0.02335852, -0.55932593, - -0.09426838, -0.44257352, 0.54939759, - 0.01533556, 0.42751634}); + lstm.SetRecurrentToCellWeights(recurrent_to_cell_weights_); + lstm.SetRecurrentToForgetWeights(recurrent_to_forget_weights_); + lstm.SetRecurrentToOutputWeights(recurrent_to_output_weights_); + + lstm.SetCellToForgetWeights(cell_to_forget_weights_); + lstm.SetCellToOutputWeights(cell_to_output_weights_); + + // Resetting cell_state and output_state + lstm.ResetCellState(); + lstm.ResetOutputState(); + + VerifyGoldens(lstm_input_, lstm_golden_output_, &lstm); +} + +TEST_F(CifgPeepholeNoProjectionNoClippingLstmTest, HybridLstmBlackBoxTest) { + const int n_batch = 1; + const int n_input = 2; + // n_cell and n_output have the same size when there is no projection. + const int n_cell = 4; + const int n_output = 4; + const int sequence_length = 3; + + HybridUnidirectionalLSTMOpModel lstm( + n_batch, n_input, n_cell, n_output, sequence_length, + /*use_cifg=*/true, /*use_peephole=*/true, + /*use_projection_weights=*/false, + /*use_projection_bias=*/false, + /*cell_clip=*/0.0, /*proj_clip=*/0.0, + { + {sequence_length, n_batch, n_input}, // input tensor + + {0, 0}, // input_to_input_weight tensor + {n_cell, n_input}, // input_to_forget_weight tensor + {n_cell, n_input}, // input_to_cell_weight tensor + {n_cell, n_input}, // input_to_output_weight tensor - lstm.SetCellBias({0., 0., 0., 0.}); + {0, 0}, // recurrent_to_input_weight tensor + {n_cell, n_output}, // recurrent_to_forget_weight tensor + {n_cell, n_output}, // recurrent_to_cell_weight tensor + {n_cell, n_output}, // recurrent_to_output_weight tensor - lstm.SetForgetGateBias({1., 1., 1., 1.}); + {0}, // cell_to_input_weight tensor + {n_cell}, // cell_to_forget_weight tensor + {n_cell}, // cell_to_output_weight tensor - lstm.SetOutputGateBias({0., 0., 0., 0.}); + {0}, // input_gate_bias tensor + {n_cell}, // forget_gate_bias tensor + {n_cell}, // cell_bias tensor + {n_cell}, // output_gate_bias tensor - lstm.SetRecurrentToCellWeights( - {0.54066205, -0.32668582, -0.43562764, -0.56094903, 0.42957711, - 0.01841056, -0.32764608, -0.33027974, -0.10826075, 0.20675004, - 0.19069612, -0.03026325, -0.54532051, 0.33003211, 0.44901288, - 0.21193194}); + {0, 0}, // projection_weight tensor + {0}, // projection_bias tensor + }); - lstm.SetRecurrentToForgetWeights( - {-0.13832897, -0.0515101, -0.2359007, -0.16661474, -0.14340827, - 0.36986142, 0.23414481, 0.55899, 0.10798943, -0.41174671, 0.17751795, - -0.34484994, -0.35874045, -0.11352962, 0.27268326, 0.54058349}); + lstm.SetInputToCellWeights(input_to_cell_weights_); + lstm.SetInputToForgetWeights(input_to_forget_weights_); + lstm.SetInputToOutputWeights(input_to_output_weights_); - lstm.SetRecurrentToOutputWeights( - {0.41613156, 0.42610586, -0.16495961, -0.5663873, 0.30579174, -0.05115908, - -0.33941799, 0.23364776, 0.11178309, 0.09481031, -0.26424935, 0.46261835, - 0.50248802, 0.26114327, -0.43736315, 0.33149987}); + lstm.SetCellBias(cell_gate_bias_); + lstm.SetForgetGateBias(forget_gate_bias_); + lstm.SetOutputGateBias(output_gate_bias_); - lstm.SetCellToForgetWeights( - {0.47485286, -0.51955009, -0.24458408, 0.31544167}); - lstm.SetCellToOutputWeights( - {-0.17135078, 0.82760304, 0.85573703, -0.77109635}); + lstm.SetRecurrentToCellWeights(recurrent_to_cell_weights_); + lstm.SetRecurrentToForgetWeights(recurrent_to_forget_weights_); + lstm.SetRecurrentToOutputWeights(recurrent_to_output_weights_); - static float lstm_input[] = {2., 3., 3., 4., 1., 1.}; - static float lstm_golden_output[] = {-0.36444446, -0.00352185, 0.12886585, - -0.05163646, -0.42312205, -0.01218222, - 0.24201041, -0.08124574, -0.358325, - -0.04621704, 0.21641694, -0.06471302}; + lstm.SetCellToForgetWeights(cell_to_forget_weights_); + lstm.SetCellToOutputWeights(cell_to_output_weights_); // Resetting cell_state and output_state lstm.ResetCellState(); lstm.ResetOutputState(); - float* batch0_start = lstm_input; - float* batch0_end = batch0_start + lstm.num_inputs() * lstm.sequence_length(); - - lstm.SetInput(0, batch0_start, batch0_end); - - lstm.Invoke(); - - float* golden_start = lstm_golden_output; - float* golden_end = - golden_start + lstm.num_outputs() * lstm.sequence_length(); - std::vector expected; - expected.insert(expected.end(), golden_start, golden_end); - EXPECT_THAT(lstm.GetOutput(), ElementsAreArray(ArrayFloatNear(expected))); + VerifyGoldens(lstm_input_, lstm_golden_output_, &lstm, /*tolerance=*/0.03573); } -TEST(LSTMOpTest, BlackBoxTestWithPeepholeWithProjectionNoClipping) { +class NoCifgPeepholeProjectionClippingLstmTest : public BaseLstmTest { + void SetUp() override { + input_to_input_weights_ = { + 0.021393683, 0.06124551, 0.046905167, -0.014657677, -0.03149463, + 0.09171803, 0.14647801, 0.10797193, -0.0057968358, 0.0019193048, + -0.2726754, 0.10154029, -0.018539885, 0.080349885, -0.10262385, + -0.022599787, -0.09121155, -0.008675967, -0.045206103, -0.0821282, + -0.008045952, 0.015478081, 0.055217247, 0.038719587, 0.044153627, + -0.06453243, 0.05031825, -0.046935108, -0.008164439, 0.014574226, + -0.1671009, -0.15519552, -0.16819797, -0.13971269, -0.11953059, + 0.25005487, -0.22790983, 0.009855087, -0.028140958, -0.11200698, + 0.11295408, -0.0035217577, 0.054485075, 0.05184695, 0.064711206, + 0.10989193, 0.11674786, 0.03490607, 0.07727357, 0.11390585, + -0.1863375, -0.1034451, -0.13945189, -0.049401227, -0.18767063, + 0.042483903, 0.14233552, 0.13832581, 0.18350165, 0.14545603, + -0.028545704, 0.024939531, 0.050929718, 0.0076203286, -0.0029723682, + -0.042484224, -0.11827596, -0.09171104, -0.10808628, -0.16327988, + -0.2273378, -0.0993647, -0.017155107, 0.0023917493, 0.049272764, + 0.0038534778, 0.054764505, 0.089753784, 0.06947234, 0.08014476, + -0.04544234, -0.0497073, -0.07135631, -0.048929106, -0.004042012, + -0.009284026, 0.018042054, 0.0036860977, -0.07427302, -0.11434604, + -0.018995456, 0.031487543, 0.012834908, 0.019977754, 0.044256654, + -0.39292613, -0.18519334, -0.11651281, -0.06809892, 0.011373677}; + + input_to_forget_weights_ = { + -0.0018401089, -0.004852237, 0.03698424, 0.014181704, + 0.028273236, -0.016726194, -0.05249759, -0.10204261, + 0.00861066, -0.040979505, -0.009899187, 0.01923892, + -0.028177269, -0.08535103, -0.14585495, 0.10662567, + -0.01909731, -0.017883534, -0.0047269356, -0.045103323, + 0.0030784295, 0.076784775, 0.07463696, 0.094531395, + 0.0814421, -0.12257899, -0.033945758, -0.031303465, + 0.045630626, 0.06843887, -0.13492945, -0.012480007, + -0.0811829, -0.07224499, -0.09628791, 0.045100946, + 0.0012300825, 0.013964662, 0.099372394, 0.02543059, + 0.06958324, 0.034257296, 0.0482646, 0.06267997, + 0.052625068, 0.12784666, 0.07077897, 0.025725935, + 0.04165009, 0.07241905, 0.018668644, -0.037377294, + -0.06277783, -0.08833636, -0.040120605, -0.011405586, + -0.007808335, -0.010301386, -0.005102167, 0.027717464, + 0.05483423, 0.11449111, 0.11289652, 0.10939839, + 0.13396506, -0.08402166, -0.01901462, -0.044678304, + -0.07720565, 0.014350063, -0.11757958, -0.0652038, + -0.08185733, -0.076754324, -0.092614375, 0.10405491, + 0.052960336, 0.035755895, 0.035839386, -0.012540553, + 0.036881298, 0.02913376, 0.03420159, 0.05448447, + -0.054523353, 0.02582715, 0.02327355, -0.011857179, + -0.0011980024, -0.034641717, -0.026125094, -0.17582615, + -0.15923657, -0.27486774, -0.0006143371, 0.0001771948, + -8.470171e-05, 0.02651807, 0.045790765, 0.06956496}; + + input_to_cell_weights_ = { + -0.04580283, -0.09549462, -0.032418985, -0.06454633, + -0.043528453, 0.043018587, -0.049152344, -0.12418144, + -0.078985475, -0.07596889, 0.019484362, -0.11434962, + -0.0074034138, -0.06314844, -0.092981495, 0.0062155537, + -0.025034338, -0.0028890965, 0.048929527, 0.06235075, + 0.10665918, -0.032036792, -0.08505916, -0.10843358, + -0.13002433, -0.036816437, -0.02130134, -0.016518239, + 0.0047691227, -0.0025825808, 0.066017866, 0.029991534, + -0.10652836, -0.1037554, -0.13056071, -0.03266643, + -0.033702414, -0.006473424, -0.04611692, 0.014419339, + -0.025174323, 0.0396852, 0.081777506, 0.06157468, + 0.10210095, -0.009658194, 0.046511717, 0.03603906, + 0.0069369148, 0.015960095, -0.06507666, 0.09551598, + 0.053568836, 0.06408714, 0.12835667, -0.008714329, + -0.20211966, -0.12093674, 0.029450472, 0.2849013, + -0.029227901, 0.1164364, -0.08560263, 0.09941786, + -0.036999565, -0.028842626, -0.0033637602, -0.017012902, + -0.09720865, -0.11193351, -0.029155117, -0.017936034, + -0.009768936, -0.04223324, -0.036159635, 0.06505112, + -0.021742892, -0.023377212, -0.07221364, -0.06430552, + 0.05453865, 0.091149814, 0.06387331, 0.007518393, + 0.055960953, 0.069779344, 0.046411168, 0.10509911, + 0.07463894, 0.0075130584, 0.012850982, 0.04555431, + 0.056955688, 0.06555285, 0.050801456, -0.009862683, + 0.00826772, -0.026555609, -0.0073611983, -0.0014897042}; + + input_to_output_weights_ = { + -0.0998932, -0.07201956, -0.052803773, -0.15629593, -0.15001918, + -0.07650751, 0.02359855, -0.075155355, -0.08037709, -0.15093534, + 0.029517552, -0.04751393, 0.010350531, -0.02664851, -0.016839722, + -0.023121163, 0.0077019283, 0.012851257, -0.05040649, -0.0129761, + -0.021737747, -0.038305793, -0.06870586, -0.01481247, -0.001285394, + 0.10124236, 0.083122835, 0.053313006, -0.062235646, -0.075637154, + -0.027833903, 0.029774971, 0.1130802, 0.09218906, 0.09506135, + -0.086665764, -0.037162706, -0.038880914, -0.035832845, -0.014481564, + -0.09825003, -0.12048569, -0.097665586, -0.05287633, -0.0964047, + -0.11366429, 0.035777505, 0.13568819, 0.052451383, 0.050649304, + 0.05798951, -0.021852335, -0.099848844, 0.014740475, -0.078897946, + 0.04974699, 0.014160473, 0.06973932, 0.04964942, 0.033364646, + 0.08190124, 0.025535367, 0.050893165, 0.048514254, 0.06945813, + -0.078907564, -0.06707616, -0.11844508, -0.09986688, -0.07509403, + 0.06263226, 0.14925587, 0.20188436, 0.12098451, 0.14639415, + 0.0015017595, -0.014267382, -0.03417257, 0.012711468, 0.0028300495, + -0.024758482, -0.05098548, -0.0821182, 0.014225672, 0.021544158, + 0.08949725, 0.07505268, -0.0020780868, 0.04908258, 0.06476295, + -0.022907063, 0.027562456, 0.040185735, 0.019567577, -0.015598739, + -0.049097303, -0.017121866, -0.083368234, -0.02332002, -0.0840956}; + + input_gate_bias_ = {0.02234832, 0.14757581, 0.18176508, 0.10380666, + 0.053110216, -0.06928846, -0.13942584, -0.11816189, + 0.19483899, 0.03652339, -0.10250295, 0.036714908, + -0.18426876, 0.036065217, 0.21810818, 0.02383196, + -0.043370757, 0.08690144, -0.04444982, 0.00030581196}; + + forget_gate_bias_ = {0.035185695, -0.042891346, -0.03032477, 0.23027696, + 0.11098921, 0.15378423, 0.09263801, 0.09790885, + 0.09508917, 0.061199076, 0.07665568, -0.015443159, + -0.03499149, 0.046190713, 0.08895977, 0.10899629, + 0.40694186, 0.06030037, 0.012413437, -0.06108739}; + + cell_gate_bias_ = {-0.024379363, 0.0055531194, 0.23377132, 0.033463873, + -0.1483596, -0.10639995, -0.091433935, 0.058573797, + -0.06809782, -0.07889636, -0.043246906, -0.09829136, + -0.4279842, 0.034901652, 0.18797937, 0.0075234566, + 0.016178843, 0.1749513, 0.13975595, 0.92058027}; + + output_gate_bias_ = {0.046159424, -0.0012809046, 0.03563469, 0.12648113, + 0.027195795, 0.35373217, -0.018957434, 0.008907322, + -0.0762701, 0.12018895, 0.04216877, 0.0022856654, + 0.040952638, 0.3147856, 0.08225149, -0.057416286, + -0.14995944, -0.008040261, 0.13208859, 0.029760877}; + + recurrent_to_input_weights_ = { + -0.001374326, -0.078856036, 0.10672688, 0.029162422, + -0.11585556, 0.02557986, -0.13446963, -0.035785314, + -0.01244275, 0.025961924, -0.02337298, -0.044228926, + -0.055839065, -0.046598054, -0.010546039, -0.06900766, + 0.027239809, 0.022582639, -0.013296484, -0.05459212, + 0.08981, -0.045407712, 0.08682226, -0.06867011, + -0.14390695, -0.02916037, 0.000996957, 0.091420636, + 0.14283475, -0.07390571, -0.06402044, 0.062524505, + -0.093129106, 0.04860203, -0.08364217, -0.08119002, + 0.009352075, 0.22920375, 0.0016303885, 0.11583097, + -0.13732095, 0.012405723, -0.07551853, 0.06343048, + 0.12162708, -0.031923793, -0.014335606, 0.01790974, + -0.10650317, -0.0724401, 0.08554849, -0.05727212, + 0.06556731, -0.042729504, -0.043227166, 0.011683251, + -0.013082158, -0.029302018, -0.010899579, -0.062036745, + -0.022509435, -0.00964907, -0.01567329, 0.04260106, + -0.07787477, -0.11576462, 0.017356863, 0.048673786, + -0.017577527, -0.05527947, -0.082487635, -0.040137455, + -0.10820036, -0.04666372, 0.022746278, -0.07851417, + 0.01068115, 0.032956902, 0.022433773, 0.0026891115, + 0.08944216, -0.0685835, 0.010513544, 0.07228705, + 0.02032331, -0.059686817, -0.0005566496, -0.086984694, + 0.040414046, -0.1380399, 0.094208956, -0.05722982, + 0.012092817, -0.04989123, -0.086576, -0.003399834, + -0.04696032, -0.045747425, 0.10091314, 0.048676282, + -0.029037097, 0.031399418, -0.0040285117, 0.047237843, + 0.09504992, 0.041799378, -0.049185462, -0.031518843, + -0.10516937, 0.026374253, 0.10058866, -0.0033195973, + -0.041975245, 0.0073591834, 0.0033782164, -0.004325073, + -0.10167381, 0.042500053, -0.01447153, 0.06464186, + -0.017142897, 0.03312627, 0.009205989, 0.024138335, + -0.011337001, 0.035530265, -0.010912711, 0.0706555, + -0.005894094, 0.051841937, -0.1401738, -0.02351249, + 0.0365468, 0.07590991, 0.08838724, 0.021681072, + -0.10086113, 0.019608743, -0.06195883, 0.077335775, + 0.023646897, -0.095322326, 0.02233014, 0.09756986, + -0.048691444, -0.009579111, 0.07595467, 0.11480546, + -0.09801813, 0.019894179, 0.08502348, 0.004032281, + 0.037211012, 0.068537936, -0.048005626, -0.091520436, + -0.028379958, -0.01556313, 0.06554592, -0.045599163, + -0.01672207, -0.020169014, -0.011877351, -0.20212261, + 0.010889619, 0.0047078193, 0.038385306, 0.08540671, + -0.017140968, -0.0035865551, 0.016678626, 0.005633034, + 0.015963363, 0.00871737, 0.060130805, 0.028611384, + 0.10109069, -0.015060172, -0.07894427, 0.06401885, + 0.011584063, -0.024466386, 0.0047652307, -0.09041358, + 0.030737216, -0.0046374933, 0.14215417, -0.11823516, + 0.019899689, 0.006106124, -0.027092824, 0.0786356, + 0.05052217, -0.058925, -0.011402121, -0.024987547, + -0.0013661642, -0.06832946, -0.015667673, -0.1083353, + -0.00096863037, -0.06988685, -0.053350925, -0.027275559, + -0.033664223, -0.07978348, -0.025200296, -0.017207067, + -0.058403496, -0.055697463, 0.005798788, 0.12965427, + -0.062582195, 0.0013350133, -0.10482091, 0.0379771, + 0.072521195, -0.0029455067, -0.13797039, -0.03628521, + 0.013806405, -0.017858358, -0.01008298, -0.07700066, + -0.017081132, 0.019358726, 0.0027079724, 0.004635139, + 0.062634714, -0.02338735, -0.039547626, -0.02050681, + 0.03385117, -0.083611414, 0.002862572, -0.09421313, + 0.058618143, -0.08598433, 0.00972939, 0.023867095, + -0.053934585, -0.023203006, 0.07452513, -0.048767887, + -0.07314807, -0.056307215, -0.10433547, -0.06440842, + 0.04328182, 0.04389765, -0.020006588, -0.09076438, + -0.11652589, -0.021705797, 0.03345259, -0.010329105, + -0.025767034, 0.013057034, -0.07316461, -0.10145612, + 0.06358255, 0.18531723, 0.07759293, 0.12006465, + 0.1305557, 0.058638252, -0.03393652, 0.09622831, + -0.16253184, -2.4580743e-06, 0.079869635, -0.070196845, + -0.005644518, 0.06857898, -0.12598175, -0.035084512, + 0.03156317, -0.12794146, -0.031963028, 0.04692781, + 0.030070418, 0.0071660685, -0.095516115, -0.004643372, + 0.040170413, -0.062104587, -0.0037324072, 0.0554317, + 0.08184801, -0.019164372, 0.06791302, 0.034257166, + -0.10307039, 0.021943003, 0.046745934, 0.0790918, + -0.0265588, -0.007824208, 0.042546265, -0.00977924, + -0.0002440307, -0.017384544, -0.017990116, 0.12252321, + -0.014512694, -0.08251313, 0.08861942, 0.13589665, + 0.026351685, 0.012641483, 0.07466548, 0.044301085, + -0.045414884, -0.051112458, 0.03444247, -0.08502782, + -0.04106223, -0.028126027, 0.028473156, 0.10467447}; + + recurrent_to_cell_weights_ = { + -0.037322544, 0.018592842, 0.0056175636, -0.06253426, + 0.055647098, -0.05713207, -0.05626563, 0.005559383, + 0.03375411, -0.025757805, -0.088049285, 0.06017052, + -0.06570978, 0.007384076, 0.035123326, -0.07920549, + 0.053676967, 0.044480428, -0.07663568, 0.0071805613, + 0.08089997, 0.05143358, 0.038261272, 0.03339287, + -0.027673481, 0.044746667, 0.028349208, 0.020090483, + -0.019443132, -0.030755889, -0.0040000007, 0.04465846, + -0.021585021, 0.0031670958, 0.0053199246, -0.056117613, + -0.10893326, 0.076739706, -0.08509834, -0.027997585, + 0.037871376, 0.01449768, -0.09002357, -0.06111149, + -0.046195522, 0.0422062, -0.005683705, -0.1253618, + -0.012925729, -0.04890792, 0.06985068, 0.037654128, + 0.03398274, -0.004781977, 0.007032333, -0.031787455, + 0.010868644, -0.031489216, 0.09525667, 0.013939797, + 0.0058680447, 0.0167067, 0.02668468, -0.04797466, + -0.048885044, -0.12722108, 0.035304096, 0.06554885, + 0.00972396, -0.039238118, -0.05159735, -0.11329045, + 0.1613692, -0.03750952, 0.06529313, -0.071974665, + -0.11769596, 0.015524369, -0.0013754242, -0.12446318, + 0.02786344, -0.014179351, 0.005264273, 0.14376344, + 0.015983658, 0.03406988, -0.06939408, 0.040699873, + 0.02111075, 0.09669095, 0.041345075, -0.08316494, + -0.07684199, -0.045768797, 0.032298047, -0.041805092, + 0.0119405, 0.0061010392, 0.12652606, 0.0064572375, + -0.024950314, 0.11574242, 0.04508852, -0.04335324, + 0.06760663, -0.027437469, 0.07216407, 0.06977076, + -0.05438599, 0.034033038, -0.028602652, 0.05346137, + 0.043184172, -0.037189785, 0.10420091, 0.00882477, + -0.054019816, -0.074273005, -0.030617684, -0.0028467078, + 0.024302477, -0.0038869337, 0.005332455, 0.0013399826, + 0.04361412, -0.007001822, 0.09631092, -0.06702025, + -0.042049985, -0.035070654, -0.04103342, -0.10273396, + 0.0544271, 0.037184782, -0.13150354, -0.0058036847, + -0.008264958, 0.042035464, 0.05891794, 0.029673764, + 0.0063542654, 0.044788733, 0.054816857, 0.062257513, + -0.00093483756, 0.048938446, -0.004952862, -0.007730018, + -0.04043371, -0.017094059, 0.07229206, -0.023670016, + -0.052195564, -0.025616996, -0.01520939, 0.045104615, + -0.007376126, 0.003533447, 0.006570588, 0.056037236, + 0.12436656, 0.051817212, 0.028532185, -0.08686856, + 0.11868599, 0.07663395, -0.07323171, 0.03463402, + -0.050708205, -0.04458982, -0.11590894, 0.021273347, + 0.1251325, -0.15313013, -0.12224372, 0.17228661, + 0.023029093, 0.086124025, 0.006445803, -0.03496501, + 0.028332196, 0.04449512, -0.042436164, -0.026587414, + -0.006041347, -0.09292539, -0.05678812, 0.03897832, + 0.09465633, 0.008115513, -0.02171956, 0.08304309, + 0.071401566, 0.019622514, 0.032163795, -0.004167056, + 0.02295182, 0.030739572, 0.056506045, 0.004612461, + 0.06524936, 0.059999723, 0.046395954, -0.0045512207, + -0.1335546, -0.030136576, 0.11584653, -0.014678886, + 0.0020118146, -0.09688814, -0.0790206, 0.039770417, + -0.0329582, 0.07922767, 0.029322514, 0.026405897, + 0.04207835, -0.07073373, 0.063781224, 0.0859677, + -0.10925287, -0.07011058, 0.048005477, 0.03438226, + -0.09606514, -0.006669445, -0.043381985, 0.04240257, + -0.06955775, -0.06769346, 0.043903265, -0.026784198, + -0.017840602, 0.024307009, -0.040079936, -0.019946516, + 0.045318738, -0.12233574, 0.026170589, 0.0074471775, + 0.15978073, 0.10185836, 0.10298046, -0.015476589, + -0.039390966, -0.072174534, 0.0739445, -0.1211869, + -0.0347889, -0.07943156, 0.014809798, -0.12412325, + -0.0030663363, 0.039695457, 0.0647603, -0.08291318, + -0.018529687, -0.004423833, 0.0037507233, 0.084633216, + -0.01514876, -0.056505352, -0.012800942, -0.06994386, + 0.012962922, -0.031234352, 0.07029052, 0.016418684, + 0.03618972, 0.055686004, -0.08663945, -0.017404709, + -0.054761406, 0.029065743, 0.052404847, 0.020238016, + 0.0048197987, -0.0214882, 0.07078733, 0.013016777, + 0.06262858, 0.009184685, 0.020785125, -0.043904778, + -0.0270329, -0.03299152, -0.060088247, -0.015162964, + -0.001828936, 0.12642565, -0.056757294, 0.013586685, + 0.09232601, -0.035886683, 0.06000002, 0.05229691, + -0.052580316, -0.082029596, -0.010794592, 0.012947712, + -0.036429964, -0.085508935, -0.13127148, -0.017744139, + 0.031502828, 0.036232427, -0.031581745, 0.023051167, + -0.05325106, -0.03421577, 0.028793324, -0.034633752, + -0.009881397, -0.043551125, -0.018609839, 0.0019097115, + -0.008799762, 0.056595087, 0.0022273948, 0.055752404}; + + recurrent_to_forget_weights_ = { + -0.057784554, -0.026057621, -0.068447545, -0.022581743, + 0.14811787, 0.10826372, 0.09471067, 0.03987225, + -0.0039523416, 0.00030638507, 0.053185795, 0.10572994, + 0.08414449, -0.022036452, -0.00066928595, -0.09203576, + 0.032950465, -0.10985798, -0.023809856, 0.0021431844, + -0.02196096, -0.00326074, 0.00058621005, -0.074678116, + -0.06193199, 0.055729095, 0.03736828, 0.020123724, + 0.061878487, -0.04729229, 0.034919553, -0.07585433, + -0.04421272, -0.044019096, 0.085488975, 0.04058006, + -0.06890133, -0.030951202, -0.024628663, -0.07672815, + 0.034293607, 0.08556707, -0.05293577, -0.033561368, + -0.04899627, 0.0241671, 0.015736353, -0.095442444, + -0.029564252, 0.016493602, -0.035026584, 0.022337519, + -0.026871363, 0.004780428, 0.0077918363, -0.03601621, + 0.016435321, -0.03263031, -0.09543275, -0.047392778, + 0.013454138, 0.028934088, 0.01685226, -0.086110644, + -0.046250615, -0.01847454, 0.047608484, 0.07339695, + 0.034546845, -0.04881143, 0.009128804, -0.08802852, + 0.03761666, 0.008096139, -0.014454086, 0.014361001, + -0.023502491, -0.0011840804, -0.07607001, 0.001856849, + -0.06509276, -0.006021153, -0.08570962, -0.1451793, + 0.060212336, 0.055259194, 0.06974018, 0.049454916, + -0.027794661, -0.08077226, -0.016179763, 0.1169753, + 0.17213494, -0.0056326236, -0.053934924, -0.0124349, + -0.11520337, 0.05409887, 0.088759385, 0.0019655675, + 0.0042065294, 0.03881498, 0.019844765, 0.041858196, + -0.05695512, 0.047233116, 0.038937137, -0.06542224, + 0.014429736, -0.09719407, 0.13908425, -0.05379757, + 0.012321099, 0.082840554, -0.029899208, 0.044217527, + 0.059855383, 0.07711018, -0.045319796, 0.0948846, + -0.011724666, -0.0033288454, -0.033542685, -0.04764985, + -0.13873616, 0.040668588, 0.034832682, -0.015319203, + -0.018715994, 0.046002675, 0.0599172, -0.043107376, + 0.0294216, -0.002314414, -0.022424703, 0.0030315618, + 0.0014641669, 0.0029166266, -0.11878115, 0.013738511, + 0.12375372, -0.0006038222, 0.029104086, 0.087442465, + 0.052958444, 0.07558703, 0.04817258, 0.044462286, + -0.015213451, -0.08783778, -0.0561384, -0.003008196, + 0.047060397, -0.002058388, 0.03429439, -0.018839769, + 0.024734668, 0.024614193, -0.042046934, 0.09597743, + -0.0043254104, 0.04320769, 0.0064070094, -0.0019131786, + -0.02558259, -0.022822596, -0.023273505, -0.02464396, + -0.10991725, -0.006240552, 0.0074488563, 0.024044557, + 0.04383914, -0.046476185, 0.028658995, 0.060410924, + 0.050786525, 0.009452605, -0.0073054377, -0.024810238, + 0.0052906186, 0.0066939713, -0.0020913032, 0.014515517, + 0.015898481, 0.021362653, -0.030262267, 0.016587038, + -0.011442813, 0.041154444, -0.007631438, -0.03423484, + -0.010977775, 0.036152758, 0.0066366293, 0.11915515, + 0.02318443, -0.041350313, 0.021485701, -0.10906167, + -0.028218046, -0.00954771, 0.020531068, -0.11995105, + -0.03672871, 0.024019798, 0.014255957, -0.05221243, + -0.00661567, -0.04630967, 0.033188973, 0.10107534, + -0.014027541, 0.030796422, -0.10270911, -0.035999842, + 0.15443139, 0.07684145, 0.036571592, -0.035900835, + -0.0034699554, 0.06209149, 0.015920248, -0.031122351, + -0.03858649, 0.01849943, 0.13872518, 0.01503974, + 0.069941424, -0.06948533, -0.0088794185, 0.061282158, + -0.047401894, 0.03100163, -0.041533746, -0.10430945, + 0.044574402, -0.01425562, -0.024290353, 0.034563623, + 0.05866852, 0.023947537, -0.09445152, 0.035450947, + 0.02247216, -0.0042998926, 0.061146557, -0.10250651, + 0.020881841, -0.06747029, 0.10062043, -0.0023941975, + 0.03532124, -0.016341697, 0.09685456, -0.016764693, + 0.051808182, 0.05875331, -0.04536488, 0.001626336, + -0.028892258, -0.01048663, -0.009793449, -0.017093895, + 0.010987891, 0.02357273, -0.00010856845, 0.0099760275, + -0.001845119, -0.03551521, 0.0018358806, 0.05763657, + -0.01769146, 0.040995963, 0.02235177, -0.060430344, + 0.11475477, -0.023854522, 0.10071741, 0.0686208, + -0.014250481, 0.034261297, 0.047418304, 0.08562733, + -0.030519066, 0.0060542435, 0.014653856, -0.038836084, + 0.04096551, 0.032249358, -0.08355519, -0.026823482, + 0.056386515, -0.010401743, -0.028396193, 0.08507674, + 0.014410365, 0.020995233, 0.17040324, 0.11511526, + 0.02459721, 0.0066619175, 0.025853224, -0.023133837, + -0.081302024, 0.017264642, -0.009585969, 0.09491168, + -0.051313367, 0.054532815, -0.014298593, 0.10657464, + 0.007076659, 0.10964551, 0.0409152, 0.008275321, + -0.07283536, 0.07937492, 0.04192024, -0.1075027}; + + recurrent_to_output_weights_ = { + 0.025825322, -0.05813119, 0.09495884, -0.045984812, + -0.01255415, -0.0026479573, -0.08196161, -0.054914974, + -0.0046604523, -0.029587349, -0.044576716, -0.07480124, + -0.082868785, 0.023254942, 0.027502948, -0.0039728214, + -0.08683098, -0.08116779, -0.014675607, -0.037924774, + -0.023314456, -0.007401714, -0.09255757, 0.029460307, + -0.08829125, -0.005139627, -0.08989442, -0.0555066, + 0.13596267, -0.025062224, -0.048351806, -0.03850004, + 0.07266485, -0.022414139, 0.05940088, 0.075114764, + 0.09597592, -0.010211725, -0.0049794707, -0.011523867, + -0.025980417, 0.072999895, 0.11091378, -0.081685916, + 0.014416728, 0.043229222, 0.034178585, -0.07530371, + 0.035837382, -0.085607, -0.007721233, -0.03287832, + -0.043848954, -0.06404588, -0.06632928, -0.073643476, + 0.008214239, -0.045984086, 0.039764922, 0.03474462, + 0.060612556, -0.080590084, 0.049127717, 0.04151091, + -0.030063879, 0.008801774, -0.023021035, -0.019558564, + 0.05158114, -0.010947698, -0.011825728, 0.0075720972, + 0.0699727, -0.0039981045, 0.069350146, 0.08799282, + 0.016156472, 0.035502106, 0.11695009, 0.006217345, + 0.13392477, -0.037875112, 0.025745004, 0.08940699, + -0.00924166, 0.0046702605, -0.036598757, -0.08811812, + 0.10522024, -0.032441203, 0.008176899, -0.04454919, + 0.07058152, 0.0067963637, 0.039206743, 0.03259838, + 0.03725492, -0.09515802, 0.013326398, -0.052055415, + -0.025676316, 0.03198509, -0.015951829, -0.058556724, + 0.036879618, 0.043357447, 0.028362012, -0.05908629, + 0.0059240665, -0.04995891, -0.019187413, 0.0276265, + -0.01628143, 0.0025863599, 0.08800015, 0.035250366, + -0.022165963, -0.07328642, -0.009415526, -0.07455109, + 0.11690406, 0.0363299, 0.07411125, 0.042103454, + -0.009660886, 0.019076364, 0.018299393, -0.046004917, + 0.08891175, 0.0431396, -0.026327137, -0.051502608, + 0.08979574, -0.051670972, 0.04940282, -0.07491107, + -0.021240504, 0.022596184, -0.034280192, 0.060163025, + -0.058211457, -0.051837247, -0.01349775, -0.04639988, + -0.035936575, -0.011681591, 0.064818054, 0.0073146066, + -0.021745546, -0.043124277, -0.06471268, -0.07053354, + -0.029321948, -0.05330136, 0.016933719, -0.053782392, + 0.13747959, -0.1361751, -0.11569455, 0.0033329215, + 0.05693899, -0.053219706, 0.063698, 0.07977434, + -0.07924483, 0.06936997, 0.0034815092, -0.007305279, + -0.037325785, -0.07251102, -0.033633437, -0.08677009, + 0.091591336, -0.14165086, 0.021752775, 0.019683983, + 0.0011612234, -0.058154266, 0.049996935, 0.0288841, + -0.0024567875, -0.14345716, 0.010955264, -0.10234828, + 0.1183656, -0.0010731248, -0.023590032, -0.072285876, + -0.0724771, -0.026382286, -0.0014920527, 0.042667855, + 0.0018776858, 0.02986552, 0.009814309, 0.0733756, + 0.12289186, 0.018043943, -0.0458958, 0.049412545, + 0.033632483, 0.05495232, 0.036686596, -0.013781798, + -0.010036754, 0.02576849, -0.08307328, 0.010112348, + 0.042521734, -0.05869831, -0.071689695, 0.03876447, + -0.13275425, -0.0352966, -0.023077697, 0.10285965, + 0.084736146, 0.15568255, -0.00040734606, 0.027835453, + -0.10292561, -0.032401145, 0.10053256, -0.026142767, + -0.08271222, -0.0030240538, -0.016368777, 0.1070414, + 0.042672627, 0.013456989, -0.0437609, -0.022309763, + 0.11576483, 0.04108048, 0.061026827, -0.0190714, + -0.0869359, 0.037901703, 0.0610107, 0.07202949, + 0.01675338, 0.086139716, -0.08795751, -0.014898893, + -0.023771819, -0.01965048, 0.007955471, -0.043740474, + 0.03346837, -0.10549954, 0.090567775, 0.042013682, + -0.03176985, 0.12569028, -0.02421228, -0.029526481, + 0.023851605, 0.031539805, 0.05292009, -0.02344001, + -0.07811758, -0.08834428, 0.10094801, 0.16594367, + -0.06861939, -0.021256343, -0.041093912, -0.06669611, + 0.035498552, 0.021757556, -0.09302526, -0.015403468, + -0.06614931, -0.051798206, -0.013874718, 0.03630673, + 0.010412845, -0.08077351, 0.046185967, 0.0035662893, + 0.03541868, -0.094149634, -0.034814864, 0.003128424, + -0.020674974, -0.03944324, -0.008110165, -0.11113267, + 0.08484226, 0.043586485, 0.040582247, 0.0968012, + -0.065249965, -0.028036479, 0.0050708856, 0.0017462453, + 0.0326779, 0.041296225, 0.09164146, -0.047743853, + -0.015952192, -0.034451712, 0.084197424, -0.05347844, + -0.11768019, 0.085926116, -0.08251791, -0.045081906, + 0.0948852, 0.068401024, 0.024856757, 0.06978981, + -0.057309967, -0.012775832, -0.0032452994, 0.01977615, + -0.041040014, -0.024264973, 0.063464895, 0.05431621, + }; + + cell_to_input_weights_ = { + 0.040369894, 0.030746894, 0.24704495, 0.018586371, -0.037586458, + -0.15312155, -0.11812848, -0.11465643, 0.20259799, 0.11418174, + -0.10116027, -0.011334949, 0.12411352, -0.076769054, -0.052169047, + 0.21198851, -0.38871562, -0.09061183, -0.09683246, -0.21929175}; + + cell_to_forget_weights_ = { + -0.01998659, -0.15568835, -0.24248174, -0.012770197, 0.041331276, + -0.072311886, -0.052123554, -0.0066330447, -0.043891653, 0.036225766, + -0.047248036, 0.021479502, 0.033189066, 0.11952997, -0.020432774, + 0.64658105, -0.06650122, -0.03467612, 0.095340036, 0.23647355}; + + cell_to_output_weights_ = { + 0.08286371, -0.08261836, -0.51210177, 0.002913762, 0.17764764, + -0.5495371, -0.08460716, -0.24552552, 0.030037103, 0.04123544, + -0.11940523, 0.007358328, 0.1890978, 0.4833202, -0.34441817, + 0.36312827, -0.26375428, 0.1457655, -0.19724406, 0.15548733}; + + projection_weights_ = { + -0.009802181, 0.09401916, 0.0717386, -0.13895074, + 0.09641832, 0.060420845, 0.08539281, 0.054285463, + 0.061395317, 0.034448683, -0.042991187, 0.019801661, + -0.16840284, -0.015726732, -0.23041931, -0.024478018, + -0.10959692, -0.013875541, 0.18600968, -0.061274476, + 0.0138165, -0.08160894, -0.07661644, 0.032372914, + 0.16169067, 0.22465782, -0.03993472, -0.004017731, + 0.08633481, -0.28869787, 0.08682067, 0.17240396, + 0.014975425, 0.056431185, 0.031037588, 0.16702051, + 0.0077946745, 0.15140012, 0.29405436, 0.120285, + -0.188994, -0.027265169, 0.043389652, -0.022061434, + 0.014777949, -0.20203483, 0.094781205, 0.19100232, + 0.13987629, -0.036132768, -0.06426278, -0.05108664, + 0.13221376, 0.009441198, -0.16715929, 0.15859416, + -0.040437475, 0.050779544, -0.022187516, 0.012166504, + 0.027685808, -0.07675938, -0.0055694645, -0.09444123, + 0.0046453946, 0.050794356, 0.10770313, -0.20790008, + -0.07149004, -0.11425117, 0.008225835, -0.035802525, + 0.14374903, 0.15262283, 0.048710253, 0.1847461, + -0.007487823, 0.11000021, -0.09542012, 0.22619456, + -0.029149994, 0.08527916, 0.009043713, 0.0042746216, + 0.016261552, 0.022461696, 0.12689082, -0.043589946, + -0.12035478, -0.08361797, -0.050666027, -0.1248618, + -0.1275799, -0.071875185, 0.07377272, 0.09944291, + -0.18897448, -0.1593054, -0.06526116, -0.040107165, + -0.004618631, -0.067624845, -0.007576253, 0.10727444, + 0.041546922, -0.20424393, 0.06907816, 0.050412357, + 0.00724631, 0.039827548, 0.12449835, 0.10747581, + 0.13708383, 0.09134148, -0.12617786, -0.06428341, + 0.09956831, 0.1208086, -0.14676677, -0.0727722, + 0.1126304, 0.010139365, 0.015571211, -0.038128063, + 0.022913318, -0.042050496, 0.16842307, -0.060597885, + 0.10531834, -0.06411776, -0.07451711, -0.03410368, + -0.13393489, 0.06534304, 0.003620307, 0.04490757, + 0.05970546, 0.05197996, 0.02839995, 0.10434969, + -0.013699693, -0.028353551, -0.07260381, 0.047201227, + -0.024575593, -0.036445823, 0.07155557, 0.009672501, + -0.02328883, 0.009533515, -0.03606021, -0.07421458, + -0.028082801, -0.2678904, -0.13221288, 0.18419984, + -0.13012612, -0.014588381, -0.035059117, -0.04824723, + 0.07830115, -0.056184657, 0.03277091, 0.025466874, + 0.14494097, -0.12522776, -0.098633975, -0.10766018, + -0.08317623, 0.08594209, 0.07749552, 0.039474737, + 0.1776665, -0.07409566, -0.0477268, 0.29323658, + 0.10801441, 0.1154011, 0.013952499, 0.10739139, + 0.10708251, -0.051456142, 0.0074137426, -0.10430189, + 0.10034707, 0.045594677, 0.0635285, -0.0715442, + -0.089667566, -0.10811871, 0.00026344223, 0.08298446, + -0.009525053, 0.006585689, -0.24567553, -0.09450807, + 0.09648481, 0.026996298, -0.06419476, -0.04752702, + -0.11063944, -0.23441927, -0.17608605, -0.052156363, + 0.067035615, 0.19271925, -0.0032889997, -0.043264326, + 0.09663576, -0.057112187, -0.10100678, 0.0628376, + 0.04447668, 0.017961001, -0.10094388, -0.10190601, + 0.18335468, 0.10494553, -0.052095775, -0.0026118709, + 0.10539724, -0.04383912, -0.042349473, 0.08438151, + -0.1947263, 0.02251204, 0.11216432, -0.10307853, + 0.17351969, -0.039091777, 0.08066188, -0.00561982, + 0.12633002, 0.11335965, -0.0088127935, -0.019777594, + 0.06864014, -0.059751723, 0.016233567, -0.06894641, + -0.28651384, -0.004228674, 0.019708522, -0.16305895, + -0.07468996, -0.0855457, 0.099339016, -0.07580735, + -0.13775392, 0.08434318, 0.08330512, -0.12131499, + 0.031935584, 0.09180414, -0.08876437, -0.08049874, + 0.008753825, 0.03498998, 0.030215185, 0.03907079, + 0.089751154, 0.029194152, -0.03337423, -0.019092513, + 0.04331237, 0.04299654, -0.036394123, -0.12915532, + 0.09793732, 0.07512415, -0.11319543, -0.032502122, + 0.15661901, 0.07671967, -0.005491124, -0.19379048, + -0.218606, 0.21448623, 0.017840758, 0.1416943, + -0.07051762, 0.19488361, 0.02664691, -0.18104725, + -0.09334311, 0.15026465, -0.15493552, -0.057762887, + -0.11604192, -0.262013, -0.01391798, 0.012185008, + 0.11156489, -0.07483202, 0.06693364, -0.26151478, + 0.046425626, 0.036540434, -0.16435726, 0.17338543, + -0.21401681, -0.11385144, -0.08283257, -0.069031075, + 0.030635102, 0.010969227, 0.11109743, 0.010919218, + 0.027526086, 0.13519906, 0.01891392, -0.046839405, + -0.040167913, 0.017953383, -0.09700955, 0.0061885654, + -0.07000971, 0.026893595, -0.038844477, 0.14543656}; + + lstm_input_ = { + {// Batch0: 4 (input_sequence_size) * 5 (n_input) + 0.787926, 0.151646, 0.071352, 0.118426, 0.458058, // step 0 + 0.596268, 0.998386, 0.568695, 0.864524, 0.571277, // step 1 + 0.073204, 0.296072, 0.743333, 0.069199, 0.045348, // step 2 + 0.867394, 0.291279, 0.013714, 0.482521, 0.626339}, // step 3 + + {// Batch1: 4 (input_sequence_size) * 5 (n_input) + 0.295743, 0.544053, 0.690064, 0.858138, 0.497181, // step 0 + 0.642421, 0.524260, 0.134799, 0.003639, 0.162482, // step 1 + 0.640394, 0.930399, 0.050782, 0.432485, 0.988078, // step 2 + 0.082922, 0.563329, 0.865614, 0.333232, 0.259916} // step 3 + }; + + lstm_golden_output_ = { + {// Batch0: 4 (input_sequence_size) * 16 (n_output) + -0.00396806, 0.029352, -0.00279226, 0.0159977, -0.00835576, + -0.0211779, 0.0283512, -0.0114597, 0.00907307, -0.0244004, + -0.0152191, -0.0259063, 0.00914318, 0.00415118, 0.017147, + 0.0134203, -0.0166936, 0.0381209, 0.000889694, 0.0143363, + -0.0328911, -0.0234288, 0.0333051, -0.012229, 0.0110322, + -0.0457725, -0.000832209, -0.0202817, 0.0327257, 0.0121308, + 0.0155969, 0.0312091, -0.0213783, 0.0350169, 0.000324794, + 0.0276012, -0.0263374, -0.0371449, 0.0446149, -0.0205474, + 0.0103729, -0.0576349, -0.0150052, -0.0292043, 0.0376827, + 0.0136115, 0.0243435, 0.0354492, -0.0189322, 0.0464512, + -0.00251373, 0.0225745, -0.0308346, -0.0317124, 0.0460407, + -0.0189395, 0.0149363, -0.0530162, -0.0150767, -0.0340193, + 0.0286833, 0.00824207, 0.0264887, 0.0305169}, + {// Batch1: 4 (input_sequence_size) * 16 (n_output) + -0.013869, 0.0287268, -0.00334693, 0.00733398, -0.0287926, + -0.0186926, 0.0193662, -0.0115437, 0.00422612, -0.0345232, + 0.00223253, -0.00957321, 0.0210624, 0.013331, 0.0150954, + 0.02168, -0.0141913, 0.0322082, 0.00227024, 0.0260507, + -0.0188721, -0.0296489, 0.0399134, -0.0160509, 0.0116039, + -0.0447318, -0.0150515, -0.0277406, 0.0316596, 0.0118233, + 0.0214762, 0.0293641, -0.0204549, 0.0450315, -0.00117378, + 0.0167673, -0.0375007, -0.0238314, 0.038784, -0.0174034, + 0.0131743, -0.0506589, -0.0048447, -0.0240239, 0.0325789, + 0.00790065, 0.0220157, 0.0333314, -0.0264787, 0.0387855, + -0.000764675, 0.0217599, -0.037537, -0.0335206, 0.0431679, + -0.0211424, 0.010203, -0.062785, -0.00832363, -0.025181, + 0.0412031, 0.0118723, 0.0239643, 0.0394009}}; + } +}; + +TEST_F(NoCifgPeepholeProjectionClippingLstmTest, LstmBlackBoxTest) { const int n_batch = 2; const int n_input = 5; const int n_cell = 20; @@ -461,8 +1322,9 @@ TEST(LSTMOpTest, BlackBoxTestWithPeepholeWithProjectionNoClipping) { const int sequence_length = 4; UnidirectionalLSTMOpModel lstm( - n_batch, n_input, n_cell, n_output, sequence_length, /*use_cifg=*/false, - /*use_peephole=*/true, /*use_projection_weights=*/true, + n_batch, n_input, n_cell, n_output, sequence_length, + /*use_cifg=*/false, /*use_peephole=*/true, + /*use_projection_weights=*/true, /*use_projection_bias=*/false, /*cell_clip=*/0.0, /*proj_clip=*/0.0, { @@ -491,588 +1353,99 @@ TEST(LSTMOpTest, BlackBoxTestWithPeepholeWithProjectionNoClipping) { {0}, // projection_bias tensor }); - lstm.SetInputToInputWeights( - {0.021393683, 0.06124551, 0.046905167, -0.014657677, -0.03149463, - 0.09171803, 0.14647801, 0.10797193, -0.0057968358, 0.0019193048, - -0.2726754, 0.10154029, -0.018539885, 0.080349885, -0.10262385, - -0.022599787, -0.09121155, -0.008675967, -0.045206103, -0.0821282, - -0.008045952, 0.015478081, 0.055217247, 0.038719587, 0.044153627, - -0.06453243, 0.05031825, -0.046935108, -0.008164439, 0.014574226, - -0.1671009, -0.15519552, -0.16819797, -0.13971269, -0.11953059, - 0.25005487, -0.22790983, 0.009855087, -0.028140958, -0.11200698, - 0.11295408, -0.0035217577, 0.054485075, 0.05184695, 0.064711206, - 0.10989193, 0.11674786, 0.03490607, 0.07727357, 0.11390585, - -0.1863375, -0.1034451, -0.13945189, -0.049401227, -0.18767063, - 0.042483903, 0.14233552, 0.13832581, 0.18350165, 0.14545603, - -0.028545704, 0.024939531, 0.050929718, 0.0076203286, -0.0029723682, - -0.042484224, -0.11827596, -0.09171104, -0.10808628, -0.16327988, - -0.2273378, -0.0993647, -0.017155107, 0.0023917493, 0.049272764, - 0.0038534778, 0.054764505, 0.089753784, 0.06947234, 0.08014476, - -0.04544234, -0.0497073, -0.07135631, -0.048929106, -0.004042012, - -0.009284026, 0.018042054, 0.0036860977, -0.07427302, -0.11434604, - -0.018995456, 0.031487543, 0.012834908, 0.019977754, 0.044256654, - -0.39292613, -0.18519334, -0.11651281, -0.06809892, 0.011373677}); - - lstm.SetInputToForgetWeights( - {-0.0018401089, -0.004852237, 0.03698424, 0.014181704, 0.028273236, - -0.016726194, -0.05249759, -0.10204261, 0.00861066, -0.040979505, - -0.009899187, 0.01923892, -0.028177269, -0.08535103, -0.14585495, - 0.10662567, -0.01909731, -0.017883534, -0.0047269356, -0.045103323, - 0.0030784295, 0.076784775, 0.07463696, 0.094531395, 0.0814421, - -0.12257899, -0.033945758, -0.031303465, 0.045630626, 0.06843887, - -0.13492945, -0.012480007, -0.0811829, -0.07224499, -0.09628791, - 0.045100946, 0.0012300825, 0.013964662, 0.099372394, 0.02543059, - 0.06958324, 0.034257296, 0.0482646, 0.06267997, 0.052625068, - 0.12784666, 0.07077897, 0.025725935, 0.04165009, 0.07241905, - 0.018668644, -0.037377294, -0.06277783, -0.08833636, -0.040120605, - -0.011405586, -0.007808335, -0.010301386, -0.005102167, 0.027717464, - 0.05483423, 0.11449111, 0.11289652, 0.10939839, 0.13396506, - -0.08402166, -0.01901462, -0.044678304, -0.07720565, 0.014350063, - -0.11757958, -0.0652038, -0.08185733, -0.076754324, -0.092614375, - 0.10405491, 0.052960336, 0.035755895, 0.035839386, -0.012540553, - 0.036881298, 0.02913376, 0.03420159, 0.05448447, -0.054523353, - 0.02582715, 0.02327355, -0.011857179, -0.0011980024, -0.034641717, - -0.026125094, -0.17582615, -0.15923657, -0.27486774, -0.0006143371, - 0.0001771948, -8.470171e-05, 0.02651807, 0.045790765, 0.06956496}); - - lstm.SetInputToCellWeights( - {-0.04580283, -0.09549462, -0.032418985, -0.06454633, - -0.043528453, 0.043018587, -0.049152344, -0.12418144, - -0.078985475, -0.07596889, 0.019484362, -0.11434962, - -0.0074034138, -0.06314844, -0.092981495, 0.0062155537, - -0.025034338, -0.0028890965, 0.048929527, 0.06235075, - 0.10665918, -0.032036792, -0.08505916, -0.10843358, - -0.13002433, -0.036816437, -0.02130134, -0.016518239, - 0.0047691227, -0.0025825808, 0.066017866, 0.029991534, - -0.10652836, -0.1037554, -0.13056071, -0.03266643, - -0.033702414, -0.006473424, -0.04611692, 0.014419339, - -0.025174323, 0.0396852, 0.081777506, 0.06157468, - 0.10210095, -0.009658194, 0.046511717, 0.03603906, - 0.0069369148, 0.015960095, -0.06507666, 0.09551598, - 0.053568836, 0.06408714, 0.12835667, -0.008714329, - -0.20211966, -0.12093674, 0.029450472, 0.2849013, - -0.029227901, 0.1164364, -0.08560263, 0.09941786, - -0.036999565, -0.028842626, -0.0033637602, -0.017012902, - -0.09720865, -0.11193351, -0.029155117, -0.017936034, - -0.009768936, -0.04223324, -0.036159635, 0.06505112, - -0.021742892, -0.023377212, -0.07221364, -0.06430552, - 0.05453865, 0.091149814, 0.06387331, 0.007518393, - 0.055960953, 0.069779344, 0.046411168, 0.10509911, - 0.07463894, 0.0075130584, 0.012850982, 0.04555431, - 0.056955688, 0.06555285, 0.050801456, -0.009862683, - 0.00826772, -0.026555609, -0.0073611983, -0.0014897042}); - - lstm.SetInputToOutputWeights( - {-0.0998932, -0.07201956, -0.052803773, -0.15629593, -0.15001918, - -0.07650751, 0.02359855, -0.075155355, -0.08037709, -0.15093534, - 0.029517552, -0.04751393, 0.010350531, -0.02664851, -0.016839722, - -0.023121163, 0.0077019283, 0.012851257, -0.05040649, -0.0129761, - -0.021737747, -0.038305793, -0.06870586, -0.01481247, -0.001285394, - 0.10124236, 0.083122835, 0.053313006, -0.062235646, -0.075637154, - -0.027833903, 0.029774971, 0.1130802, 0.09218906, 0.09506135, - -0.086665764, -0.037162706, -0.038880914, -0.035832845, -0.014481564, - -0.09825003, -0.12048569, -0.097665586, -0.05287633, -0.0964047, - -0.11366429, 0.035777505, 0.13568819, 0.052451383, 0.050649304, - 0.05798951, -0.021852335, -0.099848844, 0.014740475, -0.078897946, - 0.04974699, 0.014160473, 0.06973932, 0.04964942, 0.033364646, - 0.08190124, 0.025535367, 0.050893165, 0.048514254, 0.06945813, - -0.078907564, -0.06707616, -0.11844508, -0.09986688, -0.07509403, - 0.06263226, 0.14925587, 0.20188436, 0.12098451, 0.14639415, - 0.0015017595, -0.014267382, -0.03417257, 0.012711468, 0.0028300495, - -0.024758482, -0.05098548, -0.0821182, 0.014225672, 0.021544158, - 0.08949725, 0.07505268, -0.0020780868, 0.04908258, 0.06476295, - -0.022907063, 0.027562456, 0.040185735, 0.019567577, -0.015598739, - -0.049097303, -0.017121866, -0.083368234, -0.02332002, -0.0840956}); - - lstm.SetInputGateBias( - {0.02234832, 0.14757581, 0.18176508, 0.10380666, 0.053110216, - -0.06928846, -0.13942584, -0.11816189, 0.19483899, 0.03652339, - -0.10250295, 0.036714908, -0.18426876, 0.036065217, 0.21810818, - 0.02383196, -0.043370757, 0.08690144, -0.04444982, 0.00030581196}); - - lstm.SetForgetGateBias({0.035185695, -0.042891346, -0.03032477, 0.23027696, - 0.11098921, 0.15378423, 0.09263801, 0.09790885, - 0.09508917, 0.061199076, 0.07665568, -0.015443159, - -0.03499149, 0.046190713, 0.08895977, 0.10899629, - 0.40694186, 0.06030037, 0.012413437, -0.06108739}); - - lstm.SetCellBias({-0.024379363, 0.0055531194, 0.23377132, 0.033463873, - -0.1483596, -0.10639995, -0.091433935, 0.058573797, - -0.06809782, -0.07889636, -0.043246906, -0.09829136, - -0.4279842, 0.034901652, 0.18797937, 0.0075234566, - 0.016178843, 0.1749513, 0.13975595, 0.92058027}); - - lstm.SetOutputGateBias( - {0.046159424, -0.0012809046, 0.03563469, 0.12648113, 0.027195795, - 0.35373217, -0.018957434, 0.008907322, -0.0762701, 0.12018895, - 0.04216877, 0.0022856654, 0.040952638, 0.3147856, 0.08225149, - -0.057416286, -0.14995944, -0.008040261, 0.13208859, 0.029760877}); - - lstm.SetRecurrentToInputWeights( - {-0.001374326, -0.078856036, 0.10672688, 0.029162422, - -0.11585556, 0.02557986, -0.13446963, -0.035785314, - -0.01244275, 0.025961924, -0.02337298, -0.044228926, - -0.055839065, -0.046598054, -0.010546039, -0.06900766, - 0.027239809, 0.022582639, -0.013296484, -0.05459212, - 0.08981, -0.045407712, 0.08682226, -0.06867011, - -0.14390695, -0.02916037, 0.000996957, 0.091420636, - 0.14283475, -0.07390571, -0.06402044, 0.062524505, - -0.093129106, 0.04860203, -0.08364217, -0.08119002, - 0.009352075, 0.22920375, 0.0016303885, 0.11583097, - -0.13732095, 0.012405723, -0.07551853, 0.06343048, - 0.12162708, -0.031923793, -0.014335606, 0.01790974, - -0.10650317, -0.0724401, 0.08554849, -0.05727212, - 0.06556731, -0.042729504, -0.043227166, 0.011683251, - -0.013082158, -0.029302018, -0.010899579, -0.062036745, - -0.022509435, -0.00964907, -0.01567329, 0.04260106, - -0.07787477, -0.11576462, 0.017356863, 0.048673786, - -0.017577527, -0.05527947, -0.082487635, -0.040137455, - -0.10820036, -0.04666372, 0.022746278, -0.07851417, - 0.01068115, 0.032956902, 0.022433773, 0.0026891115, - 0.08944216, -0.0685835, 0.010513544, 0.07228705, - 0.02032331, -0.059686817, -0.0005566496, -0.086984694, - 0.040414046, -0.1380399, 0.094208956, -0.05722982, - 0.012092817, -0.04989123, -0.086576, -0.003399834, - -0.04696032, -0.045747425, 0.10091314, 0.048676282, - -0.029037097, 0.031399418, -0.0040285117, 0.047237843, - 0.09504992, 0.041799378, -0.049185462, -0.031518843, - -0.10516937, 0.026374253, 0.10058866, -0.0033195973, - -0.041975245, 0.0073591834, 0.0033782164, -0.004325073, - -0.10167381, 0.042500053, -0.01447153, 0.06464186, - -0.017142897, 0.03312627, 0.009205989, 0.024138335, - -0.011337001, 0.035530265, -0.010912711, 0.0706555, - -0.005894094, 0.051841937, -0.1401738, -0.02351249, - 0.0365468, 0.07590991, 0.08838724, 0.021681072, - -0.10086113, 0.019608743, -0.06195883, 0.077335775, - 0.023646897, -0.095322326, 0.02233014, 0.09756986, - -0.048691444, -0.009579111, 0.07595467, 0.11480546, - -0.09801813, 0.019894179, 0.08502348, 0.004032281, - 0.037211012, 0.068537936, -0.048005626, -0.091520436, - -0.028379958, -0.01556313, 0.06554592, -0.045599163, - -0.01672207, -0.020169014, -0.011877351, -0.20212261, - 0.010889619, 0.0047078193, 0.038385306, 0.08540671, - -0.017140968, -0.0035865551, 0.016678626, 0.005633034, - 0.015963363, 0.00871737, 0.060130805, 0.028611384, - 0.10109069, -0.015060172, -0.07894427, 0.06401885, - 0.011584063, -0.024466386, 0.0047652307, -0.09041358, - 0.030737216, -0.0046374933, 0.14215417, -0.11823516, - 0.019899689, 0.006106124, -0.027092824, 0.0786356, - 0.05052217, -0.058925, -0.011402121, -0.024987547, - -0.0013661642, -0.06832946, -0.015667673, -0.1083353, - -0.00096863037, -0.06988685, -0.053350925, -0.027275559, - -0.033664223, -0.07978348, -0.025200296, -0.017207067, - -0.058403496, -0.055697463, 0.005798788, 0.12965427, - -0.062582195, 0.0013350133, -0.10482091, 0.0379771, - 0.072521195, -0.0029455067, -0.13797039, -0.03628521, - 0.013806405, -0.017858358, -0.01008298, -0.07700066, - -0.017081132, 0.019358726, 0.0027079724, 0.004635139, - 0.062634714, -0.02338735, -0.039547626, -0.02050681, - 0.03385117, -0.083611414, 0.002862572, -0.09421313, - 0.058618143, -0.08598433, 0.00972939, 0.023867095, - -0.053934585, -0.023203006, 0.07452513, -0.048767887, - -0.07314807, -0.056307215, -0.10433547, -0.06440842, - 0.04328182, 0.04389765, -0.020006588, -0.09076438, - -0.11652589, -0.021705797, 0.03345259, -0.010329105, - -0.025767034, 0.013057034, -0.07316461, -0.10145612, - 0.06358255, 0.18531723, 0.07759293, 0.12006465, - 0.1305557, 0.058638252, -0.03393652, 0.09622831, - -0.16253184, -2.4580743e-06, 0.079869635, -0.070196845, - -0.005644518, 0.06857898, -0.12598175, -0.035084512, - 0.03156317, -0.12794146, -0.031963028, 0.04692781, - 0.030070418, 0.0071660685, -0.095516115, -0.004643372, - 0.040170413, -0.062104587, -0.0037324072, 0.0554317, - 0.08184801, -0.019164372, 0.06791302, 0.034257166, - -0.10307039, 0.021943003, 0.046745934, 0.0790918, - -0.0265588, -0.007824208, 0.042546265, -0.00977924, - -0.0002440307, -0.017384544, -0.017990116, 0.12252321, - -0.014512694, -0.08251313, 0.08861942, 0.13589665, - 0.026351685, 0.012641483, 0.07466548, 0.044301085, - -0.045414884, -0.051112458, 0.03444247, -0.08502782, - -0.04106223, -0.028126027, 0.028473156, 0.10467447}); - - lstm.SetRecurrentToForgetWeights( - {-0.057784554, -0.026057621, -0.068447545, -0.022581743, - 0.14811787, 0.10826372, 0.09471067, 0.03987225, - -0.0039523416, 0.00030638507, 0.053185795, 0.10572994, - 0.08414449, -0.022036452, -0.00066928595, -0.09203576, - 0.032950465, -0.10985798, -0.023809856, 0.0021431844, - -0.02196096, -0.00326074, 0.00058621005, -0.074678116, - -0.06193199, 0.055729095, 0.03736828, 0.020123724, - 0.061878487, -0.04729229, 0.034919553, -0.07585433, - -0.04421272, -0.044019096, 0.085488975, 0.04058006, - -0.06890133, -0.030951202, -0.024628663, -0.07672815, - 0.034293607, 0.08556707, -0.05293577, -0.033561368, - -0.04899627, 0.0241671, 0.015736353, -0.095442444, - -0.029564252, 0.016493602, -0.035026584, 0.022337519, - -0.026871363, 0.004780428, 0.0077918363, -0.03601621, - 0.016435321, -0.03263031, -0.09543275, -0.047392778, - 0.013454138, 0.028934088, 0.01685226, -0.086110644, - -0.046250615, -0.01847454, 0.047608484, 0.07339695, - 0.034546845, -0.04881143, 0.009128804, -0.08802852, - 0.03761666, 0.008096139, -0.014454086, 0.014361001, - -0.023502491, -0.0011840804, -0.07607001, 0.001856849, - -0.06509276, -0.006021153, -0.08570962, -0.1451793, - 0.060212336, 0.055259194, 0.06974018, 0.049454916, - -0.027794661, -0.08077226, -0.016179763, 0.1169753, - 0.17213494, -0.0056326236, -0.053934924, -0.0124349, - -0.11520337, 0.05409887, 0.088759385, 0.0019655675, - 0.0042065294, 0.03881498, 0.019844765, 0.041858196, - -0.05695512, 0.047233116, 0.038937137, -0.06542224, - 0.014429736, -0.09719407, 0.13908425, -0.05379757, - 0.012321099, 0.082840554, -0.029899208, 0.044217527, - 0.059855383, 0.07711018, -0.045319796, 0.0948846, - -0.011724666, -0.0033288454, -0.033542685, -0.04764985, - -0.13873616, 0.040668588, 0.034832682, -0.015319203, - -0.018715994, 0.046002675, 0.0599172, -0.043107376, - 0.0294216, -0.002314414, -0.022424703, 0.0030315618, - 0.0014641669, 0.0029166266, -0.11878115, 0.013738511, - 0.12375372, -0.0006038222, 0.029104086, 0.087442465, - 0.052958444, 0.07558703, 0.04817258, 0.044462286, - -0.015213451, -0.08783778, -0.0561384, -0.003008196, - 0.047060397, -0.002058388, 0.03429439, -0.018839769, - 0.024734668, 0.024614193, -0.042046934, 0.09597743, - -0.0043254104, 0.04320769, 0.0064070094, -0.0019131786, - -0.02558259, -0.022822596, -0.023273505, -0.02464396, - -0.10991725, -0.006240552, 0.0074488563, 0.024044557, - 0.04383914, -0.046476185, 0.028658995, 0.060410924, - 0.050786525, 0.009452605, -0.0073054377, -0.024810238, - 0.0052906186, 0.0066939713, -0.0020913032, 0.014515517, - 0.015898481, 0.021362653, -0.030262267, 0.016587038, - -0.011442813, 0.041154444, -0.007631438, -0.03423484, - -0.010977775, 0.036152758, 0.0066366293, 0.11915515, - 0.02318443, -0.041350313, 0.021485701, -0.10906167, - -0.028218046, -0.00954771, 0.020531068, -0.11995105, - -0.03672871, 0.024019798, 0.014255957, -0.05221243, - -0.00661567, -0.04630967, 0.033188973, 0.10107534, - -0.014027541, 0.030796422, -0.10270911, -0.035999842, - 0.15443139, 0.07684145, 0.036571592, -0.035900835, - -0.0034699554, 0.06209149, 0.015920248, -0.031122351, - -0.03858649, 0.01849943, 0.13872518, 0.01503974, - 0.069941424, -0.06948533, -0.0088794185, 0.061282158, - -0.047401894, 0.03100163, -0.041533746, -0.10430945, - 0.044574402, -0.01425562, -0.024290353, 0.034563623, - 0.05866852, 0.023947537, -0.09445152, 0.035450947, - 0.02247216, -0.0042998926, 0.061146557, -0.10250651, - 0.020881841, -0.06747029, 0.10062043, -0.0023941975, - 0.03532124, -0.016341697, 0.09685456, -0.016764693, - 0.051808182, 0.05875331, -0.04536488, 0.001626336, - -0.028892258, -0.01048663, -0.009793449, -0.017093895, - 0.010987891, 0.02357273, -0.00010856845, 0.0099760275, - -0.001845119, -0.03551521, 0.0018358806, 0.05763657, - -0.01769146, 0.040995963, 0.02235177, -0.060430344, - 0.11475477, -0.023854522, 0.10071741, 0.0686208, - -0.014250481, 0.034261297, 0.047418304, 0.08562733, - -0.030519066, 0.0060542435, 0.014653856, -0.038836084, - 0.04096551, 0.032249358, -0.08355519, -0.026823482, - 0.056386515, -0.010401743, -0.028396193, 0.08507674, - 0.014410365, 0.020995233, 0.17040324, 0.11511526, - 0.02459721, 0.0066619175, 0.025853224, -0.023133837, - -0.081302024, 0.017264642, -0.009585969, 0.09491168, - -0.051313367, 0.054532815, -0.014298593, 0.10657464, - 0.007076659, 0.10964551, 0.0409152, 0.008275321, - -0.07283536, 0.07937492, 0.04192024, -0.1075027}); - - lstm.SetRecurrentToCellWeights( - {-0.037322544, 0.018592842, 0.0056175636, -0.06253426, - 0.055647098, -0.05713207, -0.05626563, 0.005559383, - 0.03375411, -0.025757805, -0.088049285, 0.06017052, - -0.06570978, 0.007384076, 0.035123326, -0.07920549, - 0.053676967, 0.044480428, -0.07663568, 0.0071805613, - 0.08089997, 0.05143358, 0.038261272, 0.03339287, - -0.027673481, 0.044746667, 0.028349208, 0.020090483, - -0.019443132, -0.030755889, -0.0040000007, 0.04465846, - -0.021585021, 0.0031670958, 0.0053199246, -0.056117613, - -0.10893326, 0.076739706, -0.08509834, -0.027997585, - 0.037871376, 0.01449768, -0.09002357, -0.06111149, - -0.046195522, 0.0422062, -0.005683705, -0.1253618, - -0.012925729, -0.04890792, 0.06985068, 0.037654128, - 0.03398274, -0.004781977, 0.007032333, -0.031787455, - 0.010868644, -0.031489216, 0.09525667, 0.013939797, - 0.0058680447, 0.0167067, 0.02668468, -0.04797466, - -0.048885044, -0.12722108, 0.035304096, 0.06554885, - 0.00972396, -0.039238118, -0.05159735, -0.11329045, - 0.1613692, -0.03750952, 0.06529313, -0.071974665, - -0.11769596, 0.015524369, -0.0013754242, -0.12446318, - 0.02786344, -0.014179351, 0.005264273, 0.14376344, - 0.015983658, 0.03406988, -0.06939408, 0.040699873, - 0.02111075, 0.09669095, 0.041345075, -0.08316494, - -0.07684199, -0.045768797, 0.032298047, -0.041805092, - 0.0119405, 0.0061010392, 0.12652606, 0.0064572375, - -0.024950314, 0.11574242, 0.04508852, -0.04335324, - 0.06760663, -0.027437469, 0.07216407, 0.06977076, - -0.05438599, 0.034033038, -0.028602652, 0.05346137, - 0.043184172, -0.037189785, 0.10420091, 0.00882477, - -0.054019816, -0.074273005, -0.030617684, -0.0028467078, - 0.024302477, -0.0038869337, 0.005332455, 0.0013399826, - 0.04361412, -0.007001822, 0.09631092, -0.06702025, - -0.042049985, -0.035070654, -0.04103342, -0.10273396, - 0.0544271, 0.037184782, -0.13150354, -0.0058036847, - -0.008264958, 0.042035464, 0.05891794, 0.029673764, - 0.0063542654, 0.044788733, 0.054816857, 0.062257513, - -0.00093483756, 0.048938446, -0.004952862, -0.007730018, - -0.04043371, -0.017094059, 0.07229206, -0.023670016, - -0.052195564, -0.025616996, -0.01520939, 0.045104615, - -0.007376126, 0.003533447, 0.006570588, 0.056037236, - 0.12436656, 0.051817212, 0.028532185, -0.08686856, - 0.11868599, 0.07663395, -0.07323171, 0.03463402, - -0.050708205, -0.04458982, -0.11590894, 0.021273347, - 0.1251325, -0.15313013, -0.12224372, 0.17228661, - 0.023029093, 0.086124025, 0.006445803, -0.03496501, - 0.028332196, 0.04449512, -0.042436164, -0.026587414, - -0.006041347, -0.09292539, -0.05678812, 0.03897832, - 0.09465633, 0.008115513, -0.02171956, 0.08304309, - 0.071401566, 0.019622514, 0.032163795, -0.004167056, - 0.02295182, 0.030739572, 0.056506045, 0.004612461, - 0.06524936, 0.059999723, 0.046395954, -0.0045512207, - -0.1335546, -0.030136576, 0.11584653, -0.014678886, - 0.0020118146, -0.09688814, -0.0790206, 0.039770417, - -0.0329582, 0.07922767, 0.029322514, 0.026405897, - 0.04207835, -0.07073373, 0.063781224, 0.0859677, - -0.10925287, -0.07011058, 0.048005477, 0.03438226, - -0.09606514, -0.006669445, -0.043381985, 0.04240257, - -0.06955775, -0.06769346, 0.043903265, -0.026784198, - -0.017840602, 0.024307009, -0.040079936, -0.019946516, - 0.045318738, -0.12233574, 0.026170589, 0.0074471775, - 0.15978073, 0.10185836, 0.10298046, -0.015476589, - -0.039390966, -0.072174534, 0.0739445, -0.1211869, - -0.0347889, -0.07943156, 0.014809798, -0.12412325, - -0.0030663363, 0.039695457, 0.0647603, -0.08291318, - -0.018529687, -0.004423833, 0.0037507233, 0.084633216, - -0.01514876, -0.056505352, -0.012800942, -0.06994386, - 0.012962922, -0.031234352, 0.07029052, 0.016418684, - 0.03618972, 0.055686004, -0.08663945, -0.017404709, - -0.054761406, 0.029065743, 0.052404847, 0.020238016, - 0.0048197987, -0.0214882, 0.07078733, 0.013016777, - 0.06262858, 0.009184685, 0.020785125, -0.043904778, - -0.0270329, -0.03299152, -0.060088247, -0.015162964, - -0.001828936, 0.12642565, -0.056757294, 0.013586685, - 0.09232601, -0.035886683, 0.06000002, 0.05229691, - -0.052580316, -0.082029596, -0.010794592, 0.012947712, - -0.036429964, -0.085508935, -0.13127148, -0.017744139, - 0.031502828, 0.036232427, -0.031581745, 0.023051167, - -0.05325106, -0.03421577, 0.028793324, -0.034633752, - -0.009881397, -0.043551125, -0.018609839, 0.0019097115, - -0.008799762, 0.056595087, 0.0022273948, 0.055752404}); - - lstm.SetRecurrentToOutputWeights({ - 0.025825322, -0.05813119, 0.09495884, -0.045984812, -0.01255415, - -0.0026479573, -0.08196161, -0.054914974, -0.0046604523, -0.029587349, - -0.044576716, -0.07480124, -0.082868785, 0.023254942, 0.027502948, - -0.0039728214, -0.08683098, -0.08116779, -0.014675607, -0.037924774, - -0.023314456, -0.007401714, -0.09255757, 0.029460307, -0.08829125, - -0.005139627, -0.08989442, -0.0555066, 0.13596267, -0.025062224, - -0.048351806, -0.03850004, 0.07266485, -0.022414139, 0.05940088, - 0.075114764, 0.09597592, -0.010211725, -0.0049794707, -0.011523867, - -0.025980417, 0.072999895, 0.11091378, -0.081685916, 0.014416728, - 0.043229222, 0.034178585, -0.07530371, 0.035837382, -0.085607, - -0.007721233, -0.03287832, -0.043848954, -0.06404588, -0.06632928, - -0.073643476, 0.008214239, -0.045984086, 0.039764922, 0.03474462, - 0.060612556, -0.080590084, 0.049127717, 0.04151091, -0.030063879, - 0.008801774, -0.023021035, -0.019558564, 0.05158114, -0.010947698, - -0.011825728, 0.0075720972, 0.0699727, -0.0039981045, 0.069350146, - 0.08799282, 0.016156472, 0.035502106, 0.11695009, 0.006217345, - 0.13392477, -0.037875112, 0.025745004, 0.08940699, -0.00924166, - 0.0046702605, -0.036598757, -0.08811812, 0.10522024, -0.032441203, - 0.008176899, -0.04454919, 0.07058152, 0.0067963637, 0.039206743, - 0.03259838, 0.03725492, -0.09515802, 0.013326398, -0.052055415, - -0.025676316, 0.03198509, -0.015951829, -0.058556724, 0.036879618, - 0.043357447, 0.028362012, -0.05908629, 0.0059240665, -0.04995891, - -0.019187413, 0.0276265, -0.01628143, 0.0025863599, 0.08800015, - 0.035250366, -0.022165963, -0.07328642, -0.009415526, -0.07455109, - 0.11690406, 0.0363299, 0.07411125, 0.042103454, -0.009660886, - 0.019076364, 0.018299393, -0.046004917, 0.08891175, 0.0431396, - -0.026327137, -0.051502608, 0.08979574, -0.051670972, 0.04940282, - -0.07491107, -0.021240504, 0.022596184, -0.034280192, 0.060163025, - -0.058211457, -0.051837247, -0.01349775, -0.04639988, -0.035936575, - -0.011681591, 0.064818054, 0.0073146066, -0.021745546, -0.043124277, - -0.06471268, -0.07053354, -0.029321948, -0.05330136, 0.016933719, - -0.053782392, 0.13747959, -0.1361751, -0.11569455, 0.0033329215, - 0.05693899, -0.053219706, 0.063698, 0.07977434, -0.07924483, - 0.06936997, 0.0034815092, -0.007305279, -0.037325785, -0.07251102, - -0.033633437, -0.08677009, 0.091591336, -0.14165086, 0.021752775, - 0.019683983, 0.0011612234, -0.058154266, 0.049996935, 0.0288841, - -0.0024567875, -0.14345716, 0.010955264, -0.10234828, 0.1183656, - -0.0010731248, -0.023590032, -0.072285876, -0.0724771, -0.026382286, - -0.0014920527, 0.042667855, 0.0018776858, 0.02986552, 0.009814309, - 0.0733756, 0.12289186, 0.018043943, -0.0458958, 0.049412545, - 0.033632483, 0.05495232, 0.036686596, -0.013781798, -0.010036754, - 0.02576849, -0.08307328, 0.010112348, 0.042521734, -0.05869831, - -0.071689695, 0.03876447, -0.13275425, -0.0352966, -0.023077697, - 0.10285965, 0.084736146, 0.15568255, -0.00040734606, 0.027835453, - -0.10292561, -0.032401145, 0.10053256, -0.026142767, -0.08271222, - -0.0030240538, -0.016368777, 0.1070414, 0.042672627, 0.013456989, - -0.0437609, -0.022309763, 0.11576483, 0.04108048, 0.061026827, - -0.0190714, -0.0869359, 0.037901703, 0.0610107, 0.07202949, - 0.01675338, 0.086139716, -0.08795751, -0.014898893, -0.023771819, - -0.01965048, 0.007955471, -0.043740474, 0.03346837, -0.10549954, - 0.090567775, 0.042013682, -0.03176985, 0.12569028, -0.02421228, - -0.029526481, 0.023851605, 0.031539805, 0.05292009, -0.02344001, - -0.07811758, -0.08834428, 0.10094801, 0.16594367, -0.06861939, - -0.021256343, -0.041093912, -0.06669611, 0.035498552, 0.021757556, - -0.09302526, -0.015403468, -0.06614931, -0.051798206, -0.013874718, - 0.03630673, 0.010412845, -0.08077351, 0.046185967, 0.0035662893, - 0.03541868, -0.094149634, -0.034814864, 0.003128424, -0.020674974, - -0.03944324, -0.008110165, -0.11113267, 0.08484226, 0.043586485, - 0.040582247, 0.0968012, -0.065249965, -0.028036479, 0.0050708856, - 0.0017462453, 0.0326779, 0.041296225, 0.09164146, -0.047743853, - -0.015952192, -0.034451712, 0.084197424, -0.05347844, -0.11768019, - 0.085926116, -0.08251791, -0.045081906, 0.0948852, 0.068401024, - 0.024856757, 0.06978981, -0.057309967, -0.012775832, -0.0032452994, - 0.01977615, -0.041040014, -0.024264973, 0.063464895, 0.05431621, - }); - - lstm.SetCellToInputWeights( - {0.040369894, 0.030746894, 0.24704495, 0.018586371, -0.037586458, - -0.15312155, -0.11812848, -0.11465643, 0.20259799, 0.11418174, - -0.10116027, -0.011334949, 0.12411352, -0.076769054, -0.052169047, - 0.21198851, -0.38871562, -0.09061183, -0.09683246, -0.21929175}); - - lstm.SetCellToForgetWeights( - {-0.01998659, -0.15568835, -0.24248174, -0.012770197, 0.041331276, - -0.072311886, -0.052123554, -0.0066330447, -0.043891653, 0.036225766, - -0.047248036, 0.021479502, 0.033189066, 0.11952997, -0.020432774, - 0.64658105, -0.06650122, -0.03467612, 0.095340036, 0.23647355}); - - lstm.SetCellToOutputWeights( - {0.08286371, -0.08261836, -0.51210177, 0.002913762, 0.17764764, - -0.5495371, -0.08460716, -0.24552552, 0.030037103, 0.04123544, - -0.11940523, 0.007358328, 0.1890978, 0.4833202, -0.34441817, - 0.36312827, -0.26375428, 0.1457655, -0.19724406, 0.15548733}); - - lstm.SetProjectionWeights( - {-0.009802181, 0.09401916, 0.0717386, -0.13895074, 0.09641832, - 0.060420845, 0.08539281, 0.054285463, 0.061395317, 0.034448683, - -0.042991187, 0.019801661, -0.16840284, -0.015726732, -0.23041931, - -0.024478018, -0.10959692, -0.013875541, 0.18600968, -0.061274476, - 0.0138165, -0.08160894, -0.07661644, 0.032372914, 0.16169067, - 0.22465782, -0.03993472, -0.004017731, 0.08633481, -0.28869787, - 0.08682067, 0.17240396, 0.014975425, 0.056431185, 0.031037588, - 0.16702051, 0.0077946745, 0.15140012, 0.29405436, 0.120285, - -0.188994, -0.027265169, 0.043389652, -0.022061434, 0.014777949, - -0.20203483, 0.094781205, 0.19100232, 0.13987629, -0.036132768, - -0.06426278, -0.05108664, 0.13221376, 0.009441198, -0.16715929, - 0.15859416, -0.040437475, 0.050779544, -0.022187516, 0.012166504, - 0.027685808, -0.07675938, -0.0055694645, -0.09444123, 0.0046453946, - 0.050794356, 0.10770313, -0.20790008, -0.07149004, -0.11425117, - 0.008225835, -0.035802525, 0.14374903, 0.15262283, 0.048710253, - 0.1847461, -0.007487823, 0.11000021, -0.09542012, 0.22619456, - -0.029149994, 0.08527916, 0.009043713, 0.0042746216, 0.016261552, - 0.022461696, 0.12689082, -0.043589946, -0.12035478, -0.08361797, - -0.050666027, -0.1248618, -0.1275799, -0.071875185, 0.07377272, - 0.09944291, -0.18897448, -0.1593054, -0.06526116, -0.040107165, - -0.004618631, -0.067624845, -0.007576253, 0.10727444, 0.041546922, - -0.20424393, 0.06907816, 0.050412357, 0.00724631, 0.039827548, - 0.12449835, 0.10747581, 0.13708383, 0.09134148, -0.12617786, - -0.06428341, 0.09956831, 0.1208086, -0.14676677, -0.0727722, - 0.1126304, 0.010139365, 0.015571211, -0.038128063, 0.022913318, - -0.042050496, 0.16842307, -0.060597885, 0.10531834, -0.06411776, - -0.07451711, -0.03410368, -0.13393489, 0.06534304, 0.003620307, - 0.04490757, 0.05970546, 0.05197996, 0.02839995, 0.10434969, - -0.013699693, -0.028353551, -0.07260381, 0.047201227, -0.024575593, - -0.036445823, 0.07155557, 0.009672501, -0.02328883, 0.009533515, - -0.03606021, -0.07421458, -0.028082801, -0.2678904, -0.13221288, - 0.18419984, -0.13012612, -0.014588381, -0.035059117, -0.04824723, - 0.07830115, -0.056184657, 0.03277091, 0.025466874, 0.14494097, - -0.12522776, -0.098633975, -0.10766018, -0.08317623, 0.08594209, - 0.07749552, 0.039474737, 0.1776665, -0.07409566, -0.0477268, - 0.29323658, 0.10801441, 0.1154011, 0.013952499, 0.10739139, - 0.10708251, -0.051456142, 0.0074137426, -0.10430189, 0.10034707, - 0.045594677, 0.0635285, -0.0715442, -0.089667566, -0.10811871, - 0.00026344223, 0.08298446, -0.009525053, 0.006585689, -0.24567553, - -0.09450807, 0.09648481, 0.026996298, -0.06419476, -0.04752702, - -0.11063944, -0.23441927, -0.17608605, -0.052156363, 0.067035615, - 0.19271925, -0.0032889997, -0.043264326, 0.09663576, -0.057112187, - -0.10100678, 0.0628376, 0.04447668, 0.017961001, -0.10094388, - -0.10190601, 0.18335468, 0.10494553, -0.052095775, -0.0026118709, - 0.10539724, -0.04383912, -0.042349473, 0.08438151, -0.1947263, - 0.02251204, 0.11216432, -0.10307853, 0.17351969, -0.039091777, - 0.08066188, -0.00561982, 0.12633002, 0.11335965, -0.0088127935, - -0.019777594, 0.06864014, -0.059751723, 0.016233567, -0.06894641, - -0.28651384, -0.004228674, 0.019708522, -0.16305895, -0.07468996, - -0.0855457, 0.099339016, -0.07580735, -0.13775392, 0.08434318, - 0.08330512, -0.12131499, 0.031935584, 0.09180414, -0.08876437, - -0.08049874, 0.008753825, 0.03498998, 0.030215185, 0.03907079, - 0.089751154, 0.029194152, -0.03337423, -0.019092513, 0.04331237, - 0.04299654, -0.036394123, -0.12915532, 0.09793732, 0.07512415, - -0.11319543, -0.032502122, 0.15661901, 0.07671967, -0.005491124, - -0.19379048, -0.218606, 0.21448623, 0.017840758, 0.1416943, - -0.07051762, 0.19488361, 0.02664691, -0.18104725, -0.09334311, - 0.15026465, -0.15493552, -0.057762887, -0.11604192, -0.262013, - -0.01391798, 0.012185008, 0.11156489, -0.07483202, 0.06693364, - -0.26151478, 0.046425626, 0.036540434, -0.16435726, 0.17338543, - -0.21401681, -0.11385144, -0.08283257, -0.069031075, 0.030635102, - 0.010969227, 0.11109743, 0.010919218, 0.027526086, 0.13519906, - 0.01891392, -0.046839405, -0.040167913, 0.017953383, -0.09700955, - 0.0061885654, -0.07000971, 0.026893595, -0.038844477, 0.14543656}); - - static float lstm_input[][20] = { - {// Batch0: 4 (input_sequence_size) * 5 (n_input) - 0.787926, 0.151646, 0.071352, 0.118426, 0.458058, 0.596268, 0.998386, - 0.568695, 0.864524, 0.571277, 0.073204, 0.296072, 0.743333, 0.069199, - 0.045348, 0.867394, 0.291279, 0.013714, 0.482521, 0.626339}, - - {// Batch1: 4 (input_sequence_size) * 5 (n_input) - 0.295743, 0.544053, 0.690064, 0.858138, 0.497181, 0.642421, 0.524260, - 0.134799, 0.003639, 0.162482, 0.640394, 0.930399, 0.050782, 0.432485, - 0.988078, 0.082922, 0.563329, 0.865614, 0.333232, 0.259916}}; - - static float lstm_golden_output[][64] = { - {// Batch0: 4 (input_sequence_size) * 16 (n_output) - -0.00396806, 0.029352, -0.00279226, 0.0159977, -0.00835576, - -0.0211779, 0.0283512, -0.0114597, 0.00907307, -0.0244004, - -0.0152191, -0.0259063, 0.00914318, 0.00415118, 0.017147, - 0.0134203, -0.0166936, 0.0381209, 0.000889694, 0.0143363, - -0.0328911, -0.0234288, 0.0333051, -0.012229, 0.0110322, - -0.0457725, -0.000832209, -0.0202817, 0.0327257, 0.0121308, - 0.0155969, 0.0312091, -0.0213783, 0.0350169, 0.000324794, - 0.0276012, -0.0263374, -0.0371449, 0.0446149, -0.0205474, - 0.0103729, -0.0576349, -0.0150052, -0.0292043, 0.0376827, - 0.0136115, 0.0243435, 0.0354492, -0.0189322, 0.0464512, - -0.00251373, 0.0225745, -0.0308346, -0.0317124, 0.0460407, - -0.0189395, 0.0149363, -0.0530162, -0.0150767, -0.0340193, - 0.0286833, 0.00824207, 0.0264887, 0.0305169}, - {// Batch1: 4 (input_sequence_size) * 16 (n_output) - -0.013869, 0.0287268, -0.00334693, 0.00733398, -0.0287926, - -0.0186926, 0.0193662, -0.0115437, 0.00422612, -0.0345232, - 0.00223253, -0.00957321, 0.0210624, 0.013331, 0.0150954, - 0.02168, -0.0141913, 0.0322082, 0.00227024, 0.0260507, - -0.0188721, -0.0296489, 0.0399134, -0.0160509, 0.0116039, - -0.0447318, -0.0150515, -0.0277406, 0.0316596, 0.0118233, - 0.0214762, 0.0293641, -0.0204549, 0.0450315, -0.00117378, - 0.0167673, -0.0375007, -0.0238314, 0.038784, -0.0174034, - 0.0131743, -0.0506589, -0.0048447, -0.0240239, 0.0325789, - 0.00790065, 0.0220157, 0.0333314, -0.0264787, 0.0387855, - -0.000764675, 0.0217599, -0.037537, -0.0335206, 0.0431679, - -0.0211424, 0.010203, -0.062785, -0.00832363, -0.025181, - 0.0412031, 0.0118723, 0.0239643, 0.0394009}}; + lstm.SetInputToInputWeights(input_to_input_weights_); + lstm.SetInputToCellWeights(input_to_cell_weights_); + lstm.SetInputToForgetWeights(input_to_forget_weights_); + lstm.SetInputToOutputWeights(input_to_output_weights_); + + lstm.SetInputGateBias(input_gate_bias_); + lstm.SetCellBias(cell_gate_bias_); + lstm.SetForgetGateBias(forget_gate_bias_); + lstm.SetOutputGateBias(output_gate_bias_); + + lstm.SetRecurrentToInputWeights(recurrent_to_input_weights_); + lstm.SetRecurrentToCellWeights(recurrent_to_cell_weights_); + lstm.SetRecurrentToForgetWeights(recurrent_to_forget_weights_); + lstm.SetRecurrentToOutputWeights(recurrent_to_output_weights_); + + lstm.SetCellToInputWeights(cell_to_input_weights_); + lstm.SetCellToForgetWeights(cell_to_forget_weights_); + lstm.SetCellToOutputWeights(cell_to_output_weights_); + + lstm.SetProjectionWeights(projection_weights_); // Resetting cell_state and output_state lstm.ResetCellState(); lstm.ResetOutputState(); - for (int i = 0; i < lstm.sequence_length(); i++) { - float* batch0_start = lstm_input[0] + i * lstm.num_inputs(); - float* batch0_end = batch0_start + lstm.num_inputs(); + VerifyGoldens(lstm_input_, lstm_golden_output_, &lstm); +} - lstm.SetInput(2 * i * lstm.num_inputs(), batch0_start, batch0_end); +TEST_F(NoCifgPeepholeProjectionClippingLstmTest, HybridLstmBlackBoxTest) { + const int n_batch = 2; + const int n_input = 5; + const int n_cell = 20; + const int n_output = 16; + const int sequence_length = 4; - float* batch1_start = lstm_input[1] + i * lstm.num_inputs(); - float* batch1_end = batch1_start + lstm.num_inputs(); - lstm.SetInput((2 * i + 1) * lstm.num_inputs(), batch1_start, batch1_end); - } + HybridUnidirectionalLSTMOpModel lstm( + n_batch, n_input, n_cell, n_output, sequence_length, + /*use_cifg=*/false, /*use_peephole=*/true, + /*use_projection_weights=*/true, + /*use_projection_bias=*/false, + /*cell_clip=*/0.0, /*proj_clip=*/0.0, + { + {sequence_length, n_batch, n_input}, // input tensor - lstm.Invoke(); + {n_cell, n_input}, // input_to_input_weight tensor + {n_cell, n_input}, // input_to_forget_weight tensor + {n_cell, n_input}, // input_to_cell_weight tensor + {n_cell, n_input}, // input_to_output_weight tensor - std::vector expected; - for (int i = 0; i < lstm.sequence_length(); i++) { - float* golden_start_batch0 = lstm_golden_output[0] + i * lstm.num_outputs(); - float* golden_end_batch0 = golden_start_batch0 + lstm.num_outputs(); - float* golden_start_batch1 = lstm_golden_output[1] + i * lstm.num_outputs(); - float* golden_end_batch1 = golden_start_batch1 + lstm.num_outputs(); - expected.insert(expected.end(), golden_start_batch0, golden_end_batch0); - expected.insert(expected.end(), golden_start_batch1, golden_end_batch1); - } - EXPECT_THAT(lstm.GetOutput(), ElementsAreArray(ArrayFloatNear(expected))); + {n_cell, n_output}, // recurrent_to_input_weight tensor + {n_cell, n_output}, // recurrent_to_forget_weight tensor + {n_cell, n_output}, // recurrent_to_cell_weight tensor + {n_cell, n_output}, // recurrent_to_output_weight tensor + + {n_cell}, // cell_to_input_weight tensor + {n_cell}, // cell_to_forget_weight tensor + {n_cell}, // cell_to_output_weight tensor + + {n_cell}, // input_gate_bias tensor + {n_cell}, // forget_gate_bias tensor + {n_cell}, // cell_bias tensor + {n_cell}, // output_gate_bias tensor + + {n_output, n_cell}, // projection_weight tensor + {0}, // projection_bias tensor + }); + + lstm.SetInputToInputWeights(input_to_input_weights_); + lstm.SetInputToCellWeights(input_to_cell_weights_); + lstm.SetInputToForgetWeights(input_to_forget_weights_); + lstm.SetInputToOutputWeights(input_to_output_weights_); + + lstm.SetInputGateBias(input_gate_bias_); + lstm.SetCellBias(cell_gate_bias_); + lstm.SetForgetGateBias(forget_gate_bias_); + lstm.SetOutputGateBias(output_gate_bias_); + + lstm.SetRecurrentToInputWeights(recurrent_to_input_weights_); + lstm.SetRecurrentToCellWeights(recurrent_to_cell_weights_); + lstm.SetRecurrentToForgetWeights(recurrent_to_forget_weights_); + lstm.SetRecurrentToOutputWeights(recurrent_to_output_weights_); + + lstm.SetCellToInputWeights(cell_to_input_weights_); + lstm.SetCellToForgetWeights(cell_to_forget_weights_); + lstm.SetCellToOutputWeights(cell_to_output_weights_); + + lstm.SetProjectionWeights(projection_weights_); + + // Resetting cell_state and output_state + lstm.ResetCellState(); + lstm.ResetOutputState(); + + VerifyGoldens(lstm_input_, lstm_golden_output_, &lstm, /*tolerance=*/0.00467); } } // namespace diff --git a/tensorflow/contrib/lite/model.cc b/tensorflow/contrib/lite/model.cc index e1ec2d6d5789da1f5d31981501a503d7b610b336..71e38c3f134412383935323041268870c1a9338a 100644 --- a/tensorflow/contrib/lite/model.cc +++ b/tensorflow/contrib/lite/model.cc @@ -63,6 +63,9 @@ TfLiteStatus ConvertTensorType(TensorType tensor_type, TfLiteType* type, case TensorType_BOOL: *type = kTfLiteBool; break; + case TensorType_COMPLEX64: + *type = kTfLiteComplex64; + break; default: error_reporter->Report("Unimplemented data type %s (%d) in tensor\n", EnumNameTensorType(tensor_type), tensor_type); @@ -183,6 +186,8 @@ InterpreterBuilder::InterpreterBuilder(const ::tflite::Model* model, op_resolver_(op_resolver), error_reporter_(ValidateErrorReporter(error_reporter)) {} +InterpreterBuilder::~InterpreterBuilder() {} + TfLiteStatus InterpreterBuilder::BuildLocalIndexToRegistrationMapping() { TfLiteStatus status = kTfLiteOk; auto opcodes = model_->operator_codes(); @@ -201,8 +206,9 @@ TfLiteStatus InterpreterBuilder::BuildLocalIndexToRegistrationMapping() { } else if (builtin_code != BuiltinOperator_CUSTOM) { registration = op_resolver_.FindOp(builtin_code, version); if (registration == nullptr) { - error_reporter_->Report("Didn't find op for builtin opcode '%s'\n", - EnumNameBuiltinOperator(builtin_code)); + error_reporter_->Report( + "Didn't find op for builtin opcode '%s' version '%d'\n", + EnumNameBuiltinOperator(builtin_code), version); status = kTfLiteError; } } else if (!opcode->custom_code()) { @@ -444,6 +450,18 @@ TfLiteStatus ParseOpData(const Operator* op, BuiltinOperator op_type, op->builtin_options_as_FullyConnectedOptions()) { params->activation = parse_activation( fully_connected_params->fused_activation_function()); + switch (fully_connected_params->weights_format()) { + case FullyConnectedOptionsWeightsFormat_DEFAULT: + params->weights_format = kTfLiteFullyConnectedWeightsFormatDefault; + break; + case FullyConnectedOptionsWeightsFormat_SHUFFLED4x16INT8: + params->weights_format = + kTfLiteFullyConnectedWeightsFormatShuffled4x16Int8; + break; + default: + error_reporter->Report("Unhandled fully-connected weights format."); + return kTfLiteError; + } } *builtin_data = reinterpret_cast(params); break; @@ -646,6 +664,15 @@ TfLiteStatus ParseOpData(const Operator* op, BuiltinOperator op_type, *builtin_data = reinterpret_cast(params); break; } + case BuiltinOperator_ARG_MIN: { + auto* params = MallocPOD(); + if (const auto* schema_params = op->builtin_options_as_ArgMinOptions()) { + ConvertTensorType(schema_params->output_type(), ¶ms->output_type, + error_reporter); + } + *builtin_data = reinterpret_cast(params); + break; + } case BuiltinOperator_TRANSPOSE_CONV: { TfLiteTransposeConvParams* params = MallocPOD(); @@ -682,6 +709,17 @@ TfLiteStatus ParseOpData(const Operator* op, BuiltinOperator op_type, error_reporter->Report("DELEGATE op shouldn't exist in model."); return kTfLiteError; } + case BuiltinOperator_FAKE_QUANT: { + auto* params = MallocPOD(); + if (auto* schema_params = op->builtin_options_as_FakeQuantOptions()) { + params->min = schema_params->min(); + params->max = schema_params->max(); + params->num_bits = schema_params->num_bits(); + params->narrow_range = schema_params->narrow_range(); + } + *builtin_data = static_cast(params); + break; + } // Below are the ops with no builtin_data strcture. case BuiltinOperator_BATCH_TO_SPACE_ND: @@ -723,6 +761,7 @@ TfLiteStatus ParseOpData(const Operator* op, BuiltinOperator op_type, case BuiltinOperator_TILE: case BuiltinOperator_TOPK_V2: case BuiltinOperator_TRANSPOSE: + case BuiltinOperator_POW: break; } return kTfLiteOk; @@ -745,7 +784,7 @@ TfLiteStatus InterpreterBuilder::ParseNodes( } const TfLiteRegistration* registration = - flatbuffer_op_index_to_registration_[op->opcode_index()]; + flatbuffer_op_index_to_registration_[index]; if (registration == nullptr) { error_reporter_->Report("Skipping op for opcode_index %d\n", index); status = kTfLiteError; @@ -975,7 +1014,7 @@ TfLiteStatus InterpreterBuilder::operator()( variables.push_back(i); } } - (**interpreter).SetVariables(variables); + (**interpreter).SetVariables(std::move(variables)); return kTfLiteOk; } diff --git a/tensorflow/contrib/lite/model.h b/tensorflow/contrib/lite/model.h index 3946b490417104f620ecb55bb22d4ef99fd33bb7..8bc9ecd7ce9725c3d88985ccd92d48aed169fe31 100644 --- a/tensorflow/contrib/lite/model.h +++ b/tensorflow/contrib/lite/model.h @@ -156,6 +156,7 @@ class InterpreterBuilder { InterpreterBuilder(const ::tflite::Model* model, const OpResolver& op_resolver, ErrorReporter* error_reporter = DefaultErrorReporter()); + ~InterpreterBuilder(); InterpreterBuilder(const InterpreterBuilder&) = delete; InterpreterBuilder& operator=(const InterpreterBuilder&) = delete; TfLiteStatus operator()(std::unique_ptr* interpreter); diff --git a/tensorflow/contrib/lite/nnapi_delegate.cc b/tensorflow/contrib/lite/nnapi_delegate.cc index ab007993afc35a30179814df23d6c0175f1d955e..cc668485a4fc0cc9e104d2046651092293911912 100644 --- a/tensorflow/contrib/lite/nnapi_delegate.cc +++ b/tensorflow/contrib/lite/nnapi_delegate.cc @@ -29,27 +29,46 @@ limitations under the License. namespace tflite { -// TODO(aselle): FATAL leaves resources hanging. -void FATAL(const char* format, ...) { +void logError(const char* format, ...) { + // TODO(mikie): use android logging, stderr is not captured for Java + // applications va_list args; va_start(args, format); vfprintf(stderr, format, args); va_end(args); + fprintf(stderr, "\n"); fflush(stderr); - exit(1); } +#define FATAL(...) \ + logError(__VA_ARGS__); \ + exit(1); + // TODO(aselle): Change the error model to use status codes. -#define CHECK_TFLITE_SUCCESS(x) \ - if (x != kTfLiteOk) { \ - FATAL("Aborting since tflite returned failure."); \ +#define CHECK_TFLITE_SUCCESS(x) \ + if (x != kTfLiteOk) { \ + FATAL("Aborting since tflite returned failure nnapi_delegate.cc:%d.", \ + __LINE__); \ } -#define CHECK_NN(x) \ - if (x != ANEURALNETWORKS_NO_ERROR) { \ - FATAL("Aborting since tflite returned failure."); \ +#define CHECK_NN(x) \ + if (x != ANEURALNETWORKS_NO_ERROR) { \ + FATAL("Aborting since NNAPI returned failure nnapi_delegate.cc:%d", \ + __LINE__); \ } +#define RETURN_ERROR_IF_NN_FAILED(x) \ + if (x != ANEURALNETWORKS_NO_ERROR) { \ + logError( \ + "Returning error since NNAPI returned failure nnapi_delegate.cc:%d.", \ + __LINE__); \ + return kTfLiteError; \ + } + +// Tracking of NNAPI operand ids +static const int64_t kOperandIdNotSet = -1; +static const int64_t kOperandNotNeeded = -2; + namespace { int32_t GetAndroidSdkVersion() { @@ -104,21 +123,16 @@ NNAPIDelegate::~NNAPIDelegate() { } // Adds the tensors of the interpreter to the NN API model. -// Returns the number of operands added. -uint32_t addTensorOperands(tflite::Interpreter* interpreter, - ANeuralNetworksModel* nn_model, - const std::vector& skip_list) { +TfLiteStatus addTensorOperands(tflite::Interpreter* interpreter, + ANeuralNetworksModel* nn_model, + uint32_t* no_of_operands_added, + std::vector* nnapi_ids) { uint32_t next_id = 0; for (size_t i = 0; i < interpreter->tensors_size(); i++) { - // skip temporaries tensors. - bool shouldSkip = false; - for (auto skip_idx : skip_list) { - if (i == skip_idx) { - shouldSkip = true; - break; - } - } - if (shouldSkip) continue; + // Skip temporaries and RNN back-edges. + if ((*nnapi_ids)[i] == kOperandNotNeeded) continue; + + (*nnapi_ids)[i] = int64_t(next_id); int32_t nn_type = 0; // NNAPI requires 32-bit float scale to be zero, tflite doesn't care @@ -144,7 +158,18 @@ uint32_t addTensorOperands(tflite::Interpreter* interpreter, zeroPoint = tensor->params.zero_point; break; default: - FATAL("Unsupported type."); + logError("Unsupported tensor type %d", tensor->type); + return kTfLiteError; + } + if (tensor->dims->size == 0) { + logError("NNAPI doesn't support tensors with rank 0 (index %d name %s)", + i, tensor->name); + return kTfLiteError; + } + if (tensor->dims->size > 4) { + logError("NNAPI doesn't support tensors with rank > 4 (index %d name %s)", + i, tensor->name); + return kTfLiteError; } // TODO(aselle): Note, many of these are intermediate results. Do I need // to ever specify these sizes. I am currently below doing setValue @@ -154,36 +179,53 @@ uint32_t addTensorOperands(tflite::Interpreter* interpreter, ANeuralNetworksOperandType operand_type{ nn_type, static_cast(tensor->dims->size), reinterpret_cast(tensor->dims->data), scale, zeroPoint}; - CHECK_NN(ANeuralNetworksModel_addOperand(nn_model, &operand_type)); + RETURN_ERROR_IF_NN_FAILED( + ANeuralNetworksModel_addOperand(nn_model, &operand_type)); // TODO(aselle): Based on Michael's suggestion, limiting this to read // only memory if (tensor->allocation_type == kTfLiteMmapRo) { if (const NNAPIAllocation* alloc = dynamic_cast( static_cast(tensor->allocation))) { - CHECK_NN(ANeuralNetworksModel_setOperandValueFromMemory( - nn_model, next_id, alloc->memory(), alloc->offset(tensor->data.raw), - tensor->bytes)); + RETURN_ERROR_IF_NN_FAILED( + ANeuralNetworksModel_setOperandValueFromMemory( + nn_model, next_id, alloc->memory(), + alloc->offset(tensor->data.raw), tensor->bytes)); } else { - CHECK_NN(ANeuralNetworksModel_setOperandValue( + RETURN_ERROR_IF_NN_FAILED(ANeuralNetworksModel_setOperandValue( nn_model, next_id, tensor->data.raw, tensor->bytes)); } } else if (tensor->bytes == 0) { // These size 0 tensors are optional tensors reserved. - CHECK_NN( + RETURN_ERROR_IF_NN_FAILED( ANeuralNetworksModel_setOperandValue(nn_model, next_id, nullptr, 0)); } ++next_id; } - return next_id; + *no_of_operands_added = next_id; + return kTfLiteOk; +} + +void MapAndAddTensorIds(const int* from_ids_buf, size_t from_ids_count, + std::vector* into, + const std::vector& map) { + for (size_t i = 0; i < from_ids_count; i++) { + int from_id = from_ids_buf[i]; + if (from_id == kOptionalTensor) { + into->push_back(from_id); + } else { + into->push_back(map[from_id]); + } + } } // Adds the operations and their parameters to the NN API model. // 'next-id' is the operand ID of the next operand of the model. -void AddOpsAndParams(tflite::Interpreter* interpreter, - ANeuralNetworksModel* nn_model, uint32_t next_id, - std::vector* model_state_inputs, - std::vector* model_state_outputs) { +TfLiteStatus AddOpsAndParams( + tflite::Interpreter* interpreter, ANeuralNetworksModel* nn_model, + uint32_t next_id, std::vector* model_state_inputs, + std::vector* model_state_outputs, + const std::vector& tensor_id_to_nnapi_id) { for (size_t i = 0; i < interpreter->nodes_size(); i++) { const auto* node_and_registration = interpreter->node_and_registration(i); const TfLiteNode& node = node_and_registration->first; @@ -192,10 +234,11 @@ void AddOpsAndParams(tflite::Interpreter* interpreter, static_cast(registration.builtin_code); // Add the parameters. - std::vector augmented_inputs( - node.inputs->data, node.inputs->data + node.inputs->size); - std::vector augmented_outputs( - node.outputs->data, node.outputs->data + node.outputs->size); + std::vector augmented_inputs, augmented_outputs; + MapAndAddTensorIds(node.inputs->data, node.inputs->size, &augmented_inputs, + tensor_id_to_nnapi_id); + MapAndAddTensorIds(node.outputs->data, node.outputs->size, + &augmented_outputs, tensor_id_to_nnapi_id); auto add_scalar_int32 = [&nn_model, &augmented_inputs, &next_id](int value) { @@ -215,6 +258,17 @@ void AddOpsAndParams(tflite::Interpreter* interpreter, augmented_inputs.push_back(next_id++); }; + auto add_vector_int32 = [&](const int* values, uint32_t num_values) { + ANeuralNetworksOperandType operand_type{ + .type = ANEURALNETWORKS_TENSOR_INT32, + .dimensionCount = 1, + .dimensions = &num_values}; + CHECK_NN(ANeuralNetworksModel_addOperand(nn_model, &operand_type)) + CHECK_NN(ANeuralNetworksModel_setOperandValue( + nn_model, next_id, values, sizeof(int32_t) * num_values)); + augmented_inputs.push_back(next_id++); + }; + // Handle state tensors of RNN, LSTM, SVDF. // For each state_out tensor, a corresponding state_in operand needs to be // created for NNAPI. @@ -233,42 +287,54 @@ void AddOpsAndParams(tflite::Interpreter* interpreter, model_state_outputs->push_back(tensor_id); next_id++; }; + auto check_and_add_activation = [&add_scalar_int32](int activation) { + if (activation > kTfLiteActRelu6) { + FATAL("NNAPI only supports RELU, RELU1 and RELU6 activations"); + } + add_scalar_int32(activation); + }; auto add_add_params = [&add_scalar_int32](void* data) { auto* builtin = reinterpret_cast(data); + if (builtin->activation > kTfLiteActRelu6) { + FATAL("NNAPI only supports RELU, RELU1 and RELU6 activations"); + } add_scalar_int32(builtin->activation); }; - auto add_pooling_params = [&add_scalar_int32](void* data) { + auto add_pooling_params = [&add_scalar_int32, + &check_and_add_activation](void* data) { auto builtin = reinterpret_cast(data); add_scalar_int32(builtin->padding); add_scalar_int32(builtin->stride_width); add_scalar_int32(builtin->stride_height); add_scalar_int32(builtin->filter_width); add_scalar_int32(builtin->filter_height); - add_scalar_int32(builtin->activation); + check_and_add_activation(builtin->activation); }; - auto add_convolution_params = [&add_scalar_int32](void* data) { + auto add_convolution_params = [&add_scalar_int32, + &check_and_add_activation](void* data) { auto builtin = reinterpret_cast(data); add_scalar_int32(builtin->padding); add_scalar_int32(builtin->stride_width); add_scalar_int32(builtin->stride_height); - add_scalar_int32(builtin->activation); + check_and_add_activation(builtin->activation); }; - auto add_depthwise_conv_params = [&add_scalar_int32](void* data) { + auto add_depthwise_conv_params = [&add_scalar_int32, + &check_and_add_activation](void* data) { auto builtin = reinterpret_cast(data); add_scalar_int32(builtin->padding); add_scalar_int32(builtin->stride_width); add_scalar_int32(builtin->stride_height); add_scalar_int32(builtin->depth_multiplier); - add_scalar_int32(builtin->activation); + check_and_add_activation(builtin->activation); }; - auto add_fully_connected_params = [&add_scalar_int32](void* data) { + auto add_fully_connected_params = [&check_and_add_activation](void* data) { auto builtin = reinterpret_cast(data); - add_scalar_int32(builtin->activation); + check_and_add_activation(builtin->activation); }; auto add_concatenation_params = [&add_scalar_int32](void* data) { @@ -300,6 +366,7 @@ void AddOpsAndParams(tflite::Interpreter* interpreter, // LSTM in NNAPI requires scratch tensor as an output operand. auto add_lstm_scratch_tensor_float32 = [interpreter, &node, &nn_model, &next_id, &augmented_outputs]() { + if (node.temporaries->size == 0) return; int scratch_buffer_index = node.temporaries->data[0]; const TfLiteTensor* tensor = interpreter->tensor(scratch_buffer_index); ANeuralNetworksOperandType operand_type{ @@ -327,6 +394,14 @@ void AddOpsAndParams(tflite::Interpreter* interpreter, add_scalar_int32(builtin->activation); }; + auto add_squeeze_params = [&](void* data) { + const auto* builtin = reinterpret_cast(data); + // Note that we add the squeeze dimensions even if the dimensions were + // unspecified (empty), as NNAPI requires the operand. + add_vector_int32(builtin->squeeze_dims, + static_cast(builtin->num_squeeze_dims)); + }; + // Handle optional input tensors. auto add_optional_tensors = [&nn_model, &augmented_inputs, &next_id](int nn_type) { @@ -366,7 +441,14 @@ void AddOpsAndParams(tflite::Interpreter* interpreter, add_pooling_params(node.builtin_data); nn_op_type = ANEURALNETWORKS_L2_POOL_2D; break; - case tflite::BuiltinOperator_CONV_2D: + case tflite::BuiltinOperator_CONV_2D: { + auto builtin = reinterpret_cast(node.builtin_data); + if (builtin->dilation_width_factor != 1 || + builtin->dilation_height_factor != 1 || node.inputs->size != 3) { + logError("NNAPI does not support dilated Conv2D."); + return kTfLiteError; + } + } add_convolution_params(node.builtin_data); nn_op_type = ANEURALNETWORKS_CONV_2D; break; @@ -410,6 +492,10 @@ void AddOpsAndParams(tflite::Interpreter* interpreter, nn_op_type = ANEURALNETWORKS_SPACE_TO_DEPTH; break; case tflite::BuiltinOperator_LSTM: { + if (node.inputs->size + /* no of params */ 3 != 21) { + logError("NNAPI only supports 21-input LSTMs"); + return kTfLiteError; + } duplicate_state_tensor_float32( node.outputs->data[/*kOutputStateTensor*/ 0]); duplicate_state_tensor_float32( @@ -448,10 +534,31 @@ void AddOpsAndParams(tflite::Interpreter* interpreter, case tflite::BuiltinOperator_DIV: nnapi_version = 11; // require NNAPI 1.1 nn_op_type = ANEURALNETWORKS_DIV; + check_and_add_activation( + reinterpret_cast(node.builtin_data)->activation); break; case tflite::BuiltinOperator_SUB: nnapi_version = 11; // require NNAPI 1.1 nn_op_type = ANEURALNETWORKS_SUB; + check_and_add_activation( + reinterpret_cast(node.builtin_data)->activation); + break; + case tflite::BuiltinOperator_SQUEEZE: + nnapi_version = 11; // requires NNAPI 1.1 + add_squeeze_params(node.builtin_data); + nn_op_type = ANEURALNETWORKS_SQUEEZE; + break; + case tflite::BuiltinOperator_TRANSPOSE: + // The permutation input tensor value dictates the output dimensions. + // TODO(b/110888333): Support dynamically-sized tensors in delegates. + if ((node.inputs->size > 1) && + (interpreter->tensor(node.inputs->data[1])->allocation_type != + kTfLiteMmapRo)) { + logError("NNAPI does not yet support dynamic tensors."); + return kTfLiteError; + } + nnapi_version = 11; // require NNAPI 1.1 + nn_op_type = ANEURALNETWORKS_TRANSPOSE; break; case tflite::BuiltinOperator_CONCAT_EMBEDDINGS: case tflite::BuiltinOperator_LSH_PROJECTION: @@ -472,9 +579,7 @@ void AddOpsAndParams(tflite::Interpreter* interpreter, case tflite::BuiltinOperator_SPACE_TO_BATCH_ND: case tflite::BuiltinOperator_BATCH_TO_SPACE_ND: case tflite::BuiltinOperator_TOPK_V2: - case tflite::BuiltinOperator_TRANSPOSE: case tflite::BuiltinOperator_SPLIT: - case tflite::BuiltinOperator_SQUEEZE: case tflite::BuiltinOperator_STRIDED_SLICE: case tflite::BuiltinOperator_EXP: case tflite::BuiltinOperator_LOG_SOFTMAX: @@ -485,6 +590,7 @@ void AddOpsAndParams(tflite::Interpreter* interpreter, case tflite::BuiltinOperator_MAXIMUM: case tflite::BuiltinOperator_MINIMUM: case tflite::BuiltinOperator_ARG_MAX: + case tflite::BuiltinOperator_ARG_MIN: case tflite::BuiltinOperator_GREATER: case tflite::BuiltinOperator_GREATER_EQUAL: case tflite::BuiltinOperator_LESS: @@ -504,12 +610,14 @@ void AddOpsAndParams(tflite::Interpreter* interpreter, case tflite::BuiltinOperator_SQRT: case tflite::BuiltinOperator_RSQRT: case tflite::BuiltinOperator_SHAPE: - FATAL("Op code %d is currently not delegated to NNAPI", builtin); - nn_op_type = -1; // set to invalid + case tflite::BuiltinOperator_POW: + case tflite::BuiltinOperator_FAKE_QUANT: + logError("Op code %d is currently not delegated to NNAPI", builtin); + return kTfLiteError; break; case tflite::BuiltinOperator_CUSTOM: - FATAL("Custom operations are not supported when using NNAPI."); - nn_op_type = -1; // set to invalid + logError("Custom operations are not supported when using NNAPI."); + return kTfLiteError; break; } @@ -518,47 +626,70 @@ void AddOpsAndParams(tflite::Interpreter* interpreter, } // Add the operation. - CHECK_NN(ANeuralNetworksModel_addOperation( + RETURN_ERROR_IF_NN_FAILED(ANeuralNetworksModel_addOperation( nn_model, nn_op_type, static_cast(augmented_inputs.size()), augmented_inputs.data(), static_cast(augmented_outputs.size()), reinterpret_cast(augmented_outputs.data()))); } + return kTfLiteOk; } TfLiteStatus NNAPIDelegate::BuildGraph(Interpreter* interpreter) { - // TODO(aselle): This is not correct. need to handle resize invalidation. - if (nn_model_ && nn_compiled_model_) return kTfLiteOk; + if (nn_model_ && nn_compiled_model_) return model_status_; + // TODO(aselle): This is not correct. need to handle resize invalidation. if (!nn_model_) { CHECK_NN(ANeuralNetworksModel_create(&nn_model_)); - // Find all the temporary tensors and put them in a skip_list. - std::vector skip_list; + // Find which tensors should be added to NNAPI. TFLite has temporaries + // and RNN back-edges which are are not valid for NNAPI. We look through all + // inputs and outputs and mark the mapping in tensor_id_to_nnapi_id with + // kOperandIdNotSet. addTensorOperands will replace those with the + // corresponding NNAPI operand ids and skip kOperandNotNeeded entries. + std::vector tensor_id_to_nnapi_id(interpreter->tensors_size(), + kOperandNotNeeded); + auto set_ids_to_not_set = [&tensor_id_to_nnapi_id](const int* buf, + size_t count) { + for (int j = 0; j < count; j++) { + auto tensor_id = buf[j]; + if (tensor_id != kOptionalTensor) { + tensor_id_to_nnapi_id[tensor_id] = kOperandIdNotSet; + } + } + }; for (size_t i = 0; i < interpreter->nodes_size(); i++) { const auto* node_and_registration = interpreter->node_and_registration(i); const TfLiteNode& node = node_and_registration->first; - if (node.temporaries != nullptr) { - for (int j = 0; j < node.temporaries->size; j++) { - skip_list.push_back(static_cast(node.temporaries->data[j])); - } - } + set_ids_to_not_set(node.inputs->data, node.inputs->size); + set_ids_to_not_set(node.outputs->data, node.outputs->size); } - - uint32_t next_id = addTensorOperands(interpreter, nn_model_, skip_list); - AddOpsAndParams(interpreter, nn_model_, next_id, &model_states_inputs_, - &model_states_outputs_); - - std::vector augmented_inputs = interpreter->inputs(); - std::vector augmented_outputs = interpreter->outputs(); - - // All state tensors input/output need to be treated as model input/output. + set_ids_to_not_set(interpreter->inputs().data(), + interpreter->inputs().size()); + set_ids_to_not_set(interpreter->outputs().data(), + interpreter->outputs().size()); + + uint32_t next_id = 0; + RETURN_ERROR_IF_NN_FAILED(addTensorOperands( + interpreter, nn_model_, &next_id, &tensor_id_to_nnapi_id)); + RETURN_ERROR_IF_NN_FAILED( + AddOpsAndParams(interpreter, nn_model_, next_id, &model_states_inputs_, + &model_states_outputs_, tensor_id_to_nnapi_id)); + + std::vector augmented_inputs; + MapAndAddTensorIds(interpreter->inputs().data(), + interpreter->inputs().size(), &augmented_inputs, + tensor_id_to_nnapi_id); augmented_inputs.insert(augmented_inputs.end(), model_states_inputs_.begin(), model_states_inputs_.end()); - augmented_outputs.insert(augmented_outputs.end(), - model_states_outputs_.begin(), - model_states_outputs_.end()); + std::vector augmented_outputs; + MapAndAddTensorIds(interpreter->outputs().data(), + interpreter->outputs().size(), &augmented_outputs, + tensor_id_to_nnapi_id); + MapAndAddTensorIds(model_states_outputs_.data(), + model_states_outputs_.size(), &augmented_outputs, + tensor_id_to_nnapi_id); CHECK_NN(ANeuralNetworksModel_identifyInputsAndOutputs( nn_model_, static_cast(augmented_inputs.size()), @@ -576,7 +707,13 @@ TfLiteStatus NNAPIDelegate::BuildGraph(Interpreter* interpreter) { TfLiteStatus NNAPIDelegate::Invoke(Interpreter* interpreter) { if (!nn_model_) { - TF_LITE_ENSURE_STATUS(BuildGraph(interpreter)); + model_status_ = BuildGraph(interpreter); + if (model_status_ != kTfLiteOk) { + logError("Failed to build graph for NNAPI"); + } + } + if (model_status_ != kTfLiteOk) { + return model_status_; } ANeuralNetworksExecution* execution = nullptr; diff --git a/tensorflow/contrib/lite/nnapi_delegate.h b/tensorflow/contrib/lite/nnapi_delegate.h index 94dea4f9b23f208fddbacd3c77d889ea753a8a1d..8dc7d38a303f51b7ccefefd8c9d2990b443e6827 100644 --- a/tensorflow/contrib/lite/nnapi_delegate.h +++ b/tensorflow/contrib/lite/nnapi_delegate.h @@ -59,14 +59,16 @@ class NNAPIDelegate { ANeuralNetworksModel* nn_model_ = nullptr; // The NN API compilation handle ANeuralNetworksCompilation* nn_compiled_model_ = nullptr; + // Model status + TfLiteStatus model_status_ = kTfLiteOk; // List of state tensors for LSTM, RNN, SVDF. // NN API does not allow ops to maintain states across multiple // invocations. We need to manually create state input tensors from // corresponding state output tensors of TFLite operations, and map them // correctly. - std::vector model_states_inputs_; - std::vector model_states_outputs_; + std::vector model_states_inputs_; // holds NNAPI operand ids + std::vector model_states_outputs_; // holds TFLite tensor ids }; } // namespace tflite diff --git a/tensorflow/contrib/lite/optional_debug_tools.cc b/tensorflow/contrib/lite/optional_debug_tools.cc index 99c35b9cafd82c7dd7ffface33f9c6c59b404c58..f1f025f777c987c5ee47bdea457a973896b9bb82 100644 --- a/tensorflow/contrib/lite/optional_debug_tools.cc +++ b/tensorflow/contrib/lite/optional_debug_tools.cc @@ -52,6 +52,8 @@ const char* TensorTypeName(TfLiteType type) { return "kTfLiteBool"; case kTfLiteInt16: return "kTfLiteInt16"; + case kTfLiteComplex64: + return "kTfLiteComplex64"; } return "(invalid)"; } diff --git a/tensorflow/contrib/lite/python/BUILD b/tensorflow/contrib/lite/python/BUILD index 27909a9458f6b09f96cb556a5254f01e54f46e05..8c9608db049576515bd50be474c529c75f72f139 100644 --- a/tensorflow/contrib/lite/python/BUILD +++ b/tensorflow/contrib/lite/python/BUILD @@ -19,6 +19,7 @@ py_library( visibility = ["//visibility:public"], deps = [ "//tensorflow/contrib/lite/python/interpreter_wrapper:tensorflow_wrap_interpreter_wrapper", + "//tensorflow/python:util", ], ) @@ -30,9 +31,10 @@ py_test( tags = ["no_oss"], deps = [ ":interpreter", - "//tensorflow/python:array_ops", "//tensorflow/python:client_testlib", - "//tensorflow/python:platform_test", + "//tensorflow/python:framework_test_lib", + "//tensorflow/python:platform", + "//third_party/py/numpy", ], ) diff --git a/tensorflow/contrib/lite/python/interpreter.py b/tensorflow/contrib/lite/python/interpreter.py index fd908234254185e0a0639618e936ca8ff58631da..e1981ceae2a14fe42f4725c2fcd9f7f460770e21 100644 --- a/tensorflow/contrib/lite/python/interpreter.py +++ b/tensorflow/contrib/lite/python/interpreter.py @@ -56,9 +56,6 @@ class Interpreter(object): self._interpreter = ( _interpreter_wrapper.InterpreterWrapper_CreateWrapperCPPFromBuffer( model_content)) - if not self._interpreter: - raise ValueError( - 'Failed to create model from {} bytes'.format(len(model_content))) elif not model_path and not model_path: raise ValueError('`model_path` or `model_content` must be specified.') else: @@ -66,8 +63,7 @@ class Interpreter(object): def allocate_tensors(self): self._ensure_safe() - if not self._interpreter.AllocateTensors(): - raise ValueError('Failed to allocate tensors') + return self._interpreter.AllocateTensors() def _safe_to_run(self): """Returns true if there exist no numpy array buffers. @@ -152,8 +148,7 @@ class Interpreter(object): Raises: ValueError: If the interpreter could not set the tensor. """ - if not self._interpreter.SetTensor(tensor_index, value): - raise ValueError('Failed to set tensor') + self._interpreter.SetTensor(tensor_index, value) def resize_tensor_input(self, input_index, tensor_size): """Resizes an input tensor. @@ -167,8 +162,7 @@ class Interpreter(object): ValueError: If the interpreter could not resize the input tensor. """ self._ensure_safe() - if not self._interpreter.ResizeInputTensor(input_index, tensor_size): - raise ValueError('Failed to resize input') + self._interpreter.ResizeInputTensor(input_index, tensor_size) def get_output_details(self): """Gets model output details. @@ -181,7 +175,9 @@ class Interpreter(object): ] def get_tensor(self, tensor_index): - """Gets the value of the input tensor. Note this makes a copy so prefer `tensor()`. + """Gets the value of the input tensor (get a copy). + + If you wish to avoid the copy, use `tensor()`. Args: tensor_index: Tensor index of tensor to get. This value can be gotten from @@ -247,5 +243,7 @@ class Interpreter(object): ValueError: When the underlying interpreter fails raise ValueError. """ self._ensure_safe() - if not self._interpreter.Invoke(): - raise ValueError('Failed to invoke TFLite model') + self._interpreter.Invoke() + + def reset_all_variables_to_zero(self): + return self._interpreter.ResetVariableTensorsToZero() diff --git a/tensorflow/contrib/lite/python/interpreter_test.py b/tensorflow/contrib/lite/python/interpreter_test.py index 5f1fa26c3b7f76309a6f1f80aa3c1e4889781528..95fa4b8584567771a7c603e035d2335d590d3f78 100644 --- a/tensorflow/contrib/lite/python/interpreter_test.py +++ b/tensorflow/contrib/lite/python/interpreter_test.py @@ -19,6 +19,7 @@ from __future__ import print_function import io import numpy as np +import six from tensorflow.contrib.lite.python import interpreter as interpreter_wrapper from tensorflow.python.framework import test_util @@ -91,6 +92,28 @@ class InterpreterTest(test_util.TensorFlowTestCase): self.assertTrue((expected_output == output_data).all()) +class InterpreterTestErrorPropagation(test_util.TensorFlowTestCase): + + def testInvalidModelContent(self): + with self.assertRaisesRegexp(ValueError, + 'Model provided has model identifier \''): + interpreter_wrapper.Interpreter(model_content=six.b('garbage')) + + def testInvalidModelFile(self): + with self.assertRaisesRegexp( + ValueError, 'Could not open \'totally_invalid_file_name\''): + interpreter_wrapper.Interpreter( + model_path='totally_invalid_file_name') + + def testInvokeBeforeReady(self): + interpreter = interpreter_wrapper.Interpreter( + model_path=resource_loader.get_path_to_datafile( + 'testdata/permute_float.tflite')) + with self.assertRaisesRegexp(RuntimeError, + 'Invoke called on model that is not ready'): + interpreter.invoke() + + class InterpreterTensorAccessorTest(test_util.TensorFlowTestCase): def setUp(self): diff --git a/tensorflow/contrib/lite/python/interpreter_wrapper/BUILD b/tensorflow/contrib/lite/python/interpreter_wrapper/BUILD index 634c2a1e1f5005208b4eea5c853a43cccf4d244c..69ee95c320b72b68052c6f76f32c1493707f34b1 100644 --- a/tensorflow/contrib/lite/python/interpreter_wrapper/BUILD +++ b/tensorflow/contrib/lite/python/interpreter_wrapper/BUILD @@ -13,7 +13,6 @@ cc_library( deps = [ "//tensorflow/contrib/lite:framework", "//tensorflow/contrib/lite/kernels:builtin_ops", - "//tensorflow/core:lib", "//third_party/py/numpy:headers", "//third_party/python_runtime:headers", "@com_google_absl//absl/memory", diff --git a/tensorflow/contrib/lite/python/interpreter_wrapper/interpreter_wrapper.cc b/tensorflow/contrib/lite/python/interpreter_wrapper/interpreter_wrapper.cc index b283551c45d3d75aecb50043f1c7486b3345118d..c38b692dcd3969a982954a81716d8fb5d914b621 100644 --- a/tensorflow/contrib/lite/python/interpreter_wrapper/interpreter_wrapper.cc +++ b/tensorflow/contrib/lite/python/interpreter_wrapper/interpreter_wrapper.cc @@ -14,13 +14,13 @@ limitations under the License. ==============================================================================*/ #include "tensorflow/contrib/lite/python/interpreter_wrapper/interpreter_wrapper.h" +#include #include #include "absl/memory/memory.h" #include "tensorflow/contrib/lite/interpreter.h" #include "tensorflow/contrib/lite/kernels/register.h" #include "tensorflow/contrib/lite/model.h" -#include "tensorflow/core/platform/logging.h" // Disallow Numpy 1.7 deprecated symbols. #define NPY_NO_DEPRECATED_API NPY_1_7_API_VERSION @@ -38,9 +38,58 @@ limitations under the License. #define CPP_TO_PYSTRING PyString_FromStringAndSize #endif +#define TFLITE_PY_CHECK(x) \ + if ((x) != kTfLiteOk) { \ + return error_reporter_->exception(); \ + } + +#define TFLITE_PY_TENSOR_BOUNDS_CHECK(i) \ + if (i >= interpreter_->tensors_size() || i < 0) { \ + PyErr_Format(PyExc_ValueError, \ + "Invalid tensor index %d exceeds max tensor index %lu", i, \ + interpreter_->tensors_size()); \ + return nullptr; \ + } + +#define TFLITE_PY_ENSURE_VALID_INTERPRETER() \ + if (!interpreter_) { \ + PyErr_SetString(PyExc_ValueError, "Interpreter was not initialized."); \ + return nullptr; \ + } + namespace tflite { namespace interpreter_wrapper { +class PythonErrorReporter : public tflite::ErrorReporter { + public: + PythonErrorReporter() {} + + // Report an error message + int Report(const char* format, va_list args) override { + char buf[1024]; + int formatted = vsnprintf(buf, sizeof(buf), format, args); + buffer_ << buf; + return formatted; + } + + // Set's a Python runtime exception with the last error. + PyObject* exception() { + std::string last_message = message(); + PyErr_SetString(PyExc_RuntimeError, last_message.c_str()); + return nullptr; + } + + // Gets the last error message and clears the buffer. + std::string message() { + std::string value = buffer_.str(); + buffer_.clear(); + return value; + } + + private: + std::stringstream buffer_; +}; + namespace { // Calls PyArray's initialization to initialize all the API pointers. Note that @@ -60,19 +109,6 @@ std::unique_ptr CreateInterpreter( std::unique_ptr interpreter; tflite::InterpreterBuilder(*model, resolver)(&interpreter); - if (interpreter) { - for (const int input_index : interpreter->inputs()) { - const TfLiteTensor* tensor = interpreter->tensor(input_index); - CHECK(tensor); - const TfLiteIntArray* dims = tensor->dims; - if (!dims) { - continue; - } - - std::vector input_dims(dims->data, dims->data + dims->size); - interpreter->ResizeInputTensor(input_index, input_dims); - } - } return interpreter; } @@ -92,11 +128,13 @@ int TfLiteTypeToPyArrayType(TfLiteType tf_lite_type) { return NPY_OBJECT; case kTfLiteBool: return NPY_BOOL; + case kTfLiteComplex64: + return NPY_COMPLEX64; case kTfLiteNoType: - return -1; + return NPY_NOTYPE; + // Avoid default so compiler errors created when new types are made. } - LOG(ERROR) << "Unknown TfLiteType " << tf_lite_type; - return -1; + return NPY_NOTYPE; } TfLiteType TfLiteTypeFromPyArray(PyArrayObject* array) { @@ -118,8 +156,10 @@ TfLiteType TfLiteTypeFromPyArray(PyArrayObject* array) { case NPY_STRING: case NPY_UNICODE: return kTfLiteString; + case NPY_COMPLEX64: + return kTfLiteComplex64; + // Avoid default so compiler errors created when new types are made. } - LOG(ERROR) << "Unknown PyArray dtype " << pyarray_type; return kTfLiteNoType; } @@ -143,32 +183,29 @@ PyObject* PyTupleFromQuantizationParam(const TfLiteQuantizationParams& param) { } // namespace InterpreterWrapper::InterpreterWrapper( - std::unique_ptr model) + std::unique_ptr model, + std::unique_ptr error_reporter) : model_(std::move(model)), + error_reporter_(std::move(error_reporter)), resolver_(absl::make_unique()), interpreter_(CreateInterpreter(model_.get(), *resolver_)) {} InterpreterWrapper::~InterpreterWrapper() {} -bool InterpreterWrapper::AllocateTensors() { - if (!interpreter_) { - LOG(ERROR) << "Cannot allocate tensors: invalid interpreter."; - return false; - } - - if (interpreter_->AllocateTensors() != kTfLiteOk) { - LOG(ERROR) << "Unable to allocate tensors."; - return false; - } - - return true; +PyObject* InterpreterWrapper::AllocateTensors() { + TFLITE_PY_ENSURE_VALID_INTERPRETER(); + TFLITE_PY_CHECK(interpreter_->AllocateTensors()); + Py_RETURN_NONE; } -bool InterpreterWrapper::Invoke() { - return interpreter_ ? (interpreter_->Invoke() == kTfLiteOk) : false; +PyObject* InterpreterWrapper::Invoke() { + TFLITE_PY_ENSURE_VALID_INTERPRETER(); + TFLITE_PY_CHECK(interpreter_->Invoke()); + Py_RETURN_NONE; } PyObject* InterpreterWrapper::InputIndices() const { + TFLITE_PY_ENSURE_VALID_INTERPRETER(); PyObject* np_array = PyArrayFromIntVector(interpreter_->inputs().data(), interpreter_->inputs().size()); @@ -182,35 +219,36 @@ PyObject* InterpreterWrapper::OutputIndices() const { return PyArray_Return(reinterpret_cast(np_array)); } -bool InterpreterWrapper::ResizeInputTensor(int i, PyObject* value) { - if (!interpreter_) { - LOG(ERROR) << "Invalid interpreter."; - return false; - } +PyObject* InterpreterWrapper::ResizeInputTensor(int i, PyObject* value) { + TFLITE_PY_ENSURE_VALID_INTERPRETER(); std::unique_ptr array_safe( PyArray_FromAny(value, nullptr, 0, 0, NPY_ARRAY_CARRAY, nullptr)); if (!array_safe) { - LOG(ERROR) << "Failed to convert value into readable tensor."; - return false; + PyErr_SetString(PyExc_ValueError, + "Failed to convert numpy value into readable tensor."); + return nullptr; } PyArrayObject* array = reinterpret_cast(array_safe.get()); if (PyArray_NDIM(array) != 1) { - LOG(ERROR) << "Expected 1-D defining input shape."; - return false; + PyErr_Format(PyExc_ValueError, "Shape should be 1D instead of %d.", + PyArray_NDIM(array)); + return nullptr; } if (PyArray_TYPE(array) != NPY_INT32) { - LOG(ERROR) << "Shape must be an int32 array"; - return false; + PyErr_Format(PyExc_ValueError, "Shape must be type int32 (was %d).", + PyArray_TYPE(array)); + return nullptr; } std::vector dims(PyArray_SHAPE(array)[0]); memcpy(dims.data(), PyArray_BYTES(array), dims.size() * sizeof(int)); - return (interpreter_->ResizeInputTensor(i, dims) == kTfLiteOk); + TFLITE_PY_CHECK(interpreter_->ResizeInputTensor(i, dims)); + Py_RETURN_NONE; } std::string InterpreterWrapper::TensorName(int i) const { @@ -223,21 +261,21 @@ std::string InterpreterWrapper::TensorName(int i) const { } PyObject* InterpreterWrapper::TensorType(int i) const { - if (!interpreter_ || i >= interpreter_->tensors_size() || i < 0) { - return nullptr; - } + TFLITE_PY_ENSURE_VALID_INTERPRETER(); + TFLITE_PY_TENSOR_BOUNDS_CHECK(i); const TfLiteTensor* tensor = interpreter_->tensor(i); - int typenum = TfLiteTypeToPyArrayType(tensor->type); - return PyArray_TypeObjectFromType(typenum); + int code = TfLiteTypeToPyArrayType(tensor->type); + if (code == -1) { + PyErr_Format(PyExc_ValueError, "Invalid tflite type code %d", code); + return nullptr; + } + return PyArray_TypeObjectFromType(code); } PyObject* InterpreterWrapper::TensorSize(int i) const { - if (!interpreter_ || i >= interpreter_->tensors_size() || i < 0) { - Py_INCREF(Py_None); - return Py_None; - } - + TFLITE_PY_ENSURE_VALID_INTERPRETER(); + TFLITE_PY_TENSOR_BOUNDS_CHECK(i); const TfLiteTensor* tensor = interpreter_->tensor(i); PyObject* np_array = PyArrayFromIntVector(tensor->dims->data, tensor->dims->size); @@ -246,97 +284,82 @@ PyObject* InterpreterWrapper::TensorSize(int i) const { } PyObject* InterpreterWrapper::TensorQuantization(int i) const { - if (!interpreter_ || i >= interpreter_->tensors_size() || i < 0) { - Py_INCREF(Py_None); - return Py_None; - } - + TFLITE_PY_ENSURE_VALID_INTERPRETER(); + TFLITE_PY_TENSOR_BOUNDS_CHECK(i); const TfLiteTensor* tensor = interpreter_->tensor(i); return PyTupleFromQuantizationParam(tensor->params); } -bool InterpreterWrapper::SetTensor(int i, PyObject* value) { - if (!interpreter_) { - LOG(ERROR) << "Invalid interpreter."; - return false; - } - - if (i >= interpreter_->tensors_size()) { - LOG(ERROR) << "Invalid tensor index: " << i << " exceeds max tensor index " - << interpreter_->tensors_size(); - return false; - } +PyObject* InterpreterWrapper::SetTensor(int i, PyObject* value) { + TFLITE_PY_ENSURE_VALID_INTERPRETER(); + TFLITE_PY_TENSOR_BOUNDS_CHECK(i); std::unique_ptr array_safe( PyArray_FromAny(value, nullptr, 0, 0, NPY_ARRAY_CARRAY, nullptr)); if (!array_safe) { - LOG(ERROR) << "Failed to convert value into readable tensor."; - return false; + PyErr_SetString(PyExc_ValueError, + "Failed to convert value into readable tensor."); + return nullptr; } PyArrayObject* array = reinterpret_cast(array_safe.get()); const TfLiteTensor* tensor = interpreter_->tensor(i); if (TfLiteTypeFromPyArray(array) != tensor->type) { - LOG(ERROR) << "Cannot set tensor:" - << " Got tensor of type " << TfLiteTypeFromPyArray(array) - << " but expected type " << tensor->type << " for input " << i; - return false; + PyErr_Format(PyExc_ValueError, + "Cannot set tensor:" + " Got tensor of type %d" + " but expected type %d for input %d ", + TfLiteTypeFromPyArray(array), tensor->type, i); + return nullptr; } if (PyArray_NDIM(array) != tensor->dims->size) { - LOG(ERROR) << "Cannot set tensor: Dimension mismatch"; - return false; + PyErr_SetString(PyExc_ValueError, "Cannot set tensor: Dimension mismatch"); + return nullptr; } for (int j = 0; j < PyArray_NDIM(array); j++) { if (tensor->dims->data[j] != PyArray_SHAPE(array)[j]) { - LOG(ERROR) << "Cannot set tensor: Dimension mismatch"; - return false; + PyErr_SetString(PyExc_ValueError, + "Cannot set tensor: Dimension mismatch"); + return nullptr; } } size_t size = PyArray_NBYTES(array); - DCHECK_EQ(size, tensor->bytes); + if (size != tensor->bytes) { + PyErr_Format(PyExc_ValueError, + "numpy array had %zu bytes but expected %zu bytes.", size, + tensor->bytes); + return nullptr; + } memcpy(tensor->data.raw, PyArray_DATA(array), size); - return true; + Py_RETURN_NONE; } namespace { -PyObject* CheckGetTensorArgs(Interpreter* interpreter, int tensor_index, +PyObject* CheckGetTensorArgs(Interpreter* interpreter_, int tensor_index, TfLiteTensor** tensor, int* type_num) { - if (!interpreter) { - LOG(ERROR) << "Invalid interpreter."; - Py_INCREF(Py_None); - return Py_None; - } - - if (tensor_index >= interpreter->tensors_size() || tensor_index < 0) { - LOG(ERROR) << "Invalid tensor index: " << tensor_index - << " exceeds max tensor index " << interpreter->inputs().size(); - Py_INCREF(Py_None); - return Py_None; - } + TFLITE_PY_ENSURE_VALID_INTERPRETER(); + TFLITE_PY_TENSOR_BOUNDS_CHECK(tensor_index); - *tensor = interpreter->tensor(tensor_index); + *tensor = interpreter_->tensor(tensor_index); if ((*tensor)->bytes == 0) { - LOG(ERROR) << "Invalid tensor size"; - Py_INCREF(Py_None); - return Py_None; + PyErr_SetString(PyExc_ValueError, "Invalid tensor size."); + return nullptr; } *type_num = TfLiteTypeToPyArrayType((*tensor)->type); if (*type_num == -1) { - LOG(ERROR) << "Unknown tensor type " << (*tensor)->type; - Py_INCREF(Py_None); - return Py_None; + PyErr_SetString(PyExc_ValueError, "Unknown tensor type."); + return nullptr; } if (!(*tensor)->data.raw) { - LOG(ERROR) << "Tensor data is null."; - Py_INCREF(Py_None); - return Py_None; + PyErr_SetString(PyExc_ValueError, "Tensor data is null."); + return nullptr; } return nullptr; @@ -358,9 +381,8 @@ PyObject* InterpreterWrapper::GetTensor(int i) const { // it will leak. void* data = malloc(tensor->bytes); if (!data) { - LOG(ERROR) << "Malloc to copy tensor failed."; - Py_INCREF(Py_None); - return Py_None; + PyErr_SetString(PyExc_ValueError, "Malloc to copy tensor failed."); + return nullptr; } memcpy(data, tensor->data.raw, tensor->bytes); PyObject* np_array = @@ -390,22 +412,39 @@ PyObject* InterpreterWrapper::tensor(PyObject* base_object, int i) { } InterpreterWrapper* InterpreterWrapper::CreateWrapperCPPFromFile( - const char* model_path) { + const char* model_path, std::string* error_msg) { + std::unique_ptr error_reporter(new PythonErrorReporter); std::unique_ptr model = - tflite::FlatBufferModel::BuildFromFile(model_path); - return model ? new InterpreterWrapper(std::move(model)) : nullptr; + tflite::FlatBufferModel::BuildFromFile(model_path, error_reporter.get()); + if (!model) { + *error_msg = error_reporter->message(); + return nullptr; + } + return new InterpreterWrapper(std::move(model), std::move(error_reporter)); } InterpreterWrapper* InterpreterWrapper::CreateWrapperCPPFromBuffer( - PyObject* data) { + PyObject* data, std::string* error_msg) { char * buf = nullptr; Py_ssize_t length; + std::unique_ptr error_reporter(new PythonErrorReporter); if (PY_TO_CPPSTRING(data, &buf, &length) == -1) { return nullptr; } std::unique_ptr model = - tflite::FlatBufferModel::BuildFromBuffer(buf, length); - return model ? new InterpreterWrapper(std::move(model)) : nullptr; + tflite::FlatBufferModel::BuildFromBuffer(buf, length, + error_reporter.get()); + if (!model) { + *error_msg = error_reporter->message(); + return nullptr; + } + return new InterpreterWrapper(std::move(model), std::move(error_reporter)); +} + +PyObject* InterpreterWrapper::ResetVariableTensorsToZero() { + TFLITE_PY_ENSURE_VALID_INTERPRETER(); + TFLITE_PY_CHECK(interpreter_->ResetVariableTensorsToZero()); + Py_RETURN_NONE; } } // namespace interpreter_wrapper diff --git a/tensorflow/contrib/lite/python/interpreter_wrapper/interpreter_wrapper.h b/tensorflow/contrib/lite/python/interpreter_wrapper/interpreter_wrapper.h index e7343cb388d657e472464f69fa8cd0c6ddc60923..febfd2dc56d0368e04ab2ccb8234ffb9133eb193 100644 --- a/tensorflow/contrib/lite/python/interpreter_wrapper/interpreter_wrapper.h +++ b/tensorflow/contrib/lite/python/interpreter_wrapper/interpreter_wrapper.h @@ -20,8 +20,8 @@ limitations under the License. #include // Place `` before to avoid build failures in macOS. -#include #include +#include // We forward declare TFLite classes here to avoid exposing them to SWIG. namespace tflite { @@ -36,34 +36,41 @@ class Interpreter; namespace interpreter_wrapper { +class PythonErrorReporter; + class InterpreterWrapper { public: // SWIG caller takes ownership of pointer. - static InterpreterWrapper* CreateWrapperCPPFromFile(const char* model_path); + static InterpreterWrapper* CreateWrapperCPPFromFile(const char* model_path, + std::string* error_msg); // SWIG caller takes ownership of pointer. - static InterpreterWrapper* CreateWrapperCPPFromBuffer(PyObject* data); + static InterpreterWrapper* CreateWrapperCPPFromBuffer(PyObject* data, + std::string* error_msg); ~InterpreterWrapper(); - bool AllocateTensors(); - bool Invoke(); + PyObject* AllocateTensors(); + PyObject* Invoke(); PyObject* InputIndices() const; PyObject* OutputIndices() const; - bool ResizeInputTensor(int i, PyObject* value); + PyObject* ResizeInputTensor(int i, PyObject* value); std::string TensorName(int i) const; PyObject* TensorType(int i) const; PyObject* TensorSize(int i) const; PyObject* TensorQuantization(int i) const; - bool SetTensor(int i, PyObject* value); + PyObject* SetTensor(int i, PyObject* value); PyObject* GetTensor(int i) const; + PyObject* ResetVariableTensorsToZero(); + // Returns a reference to tensor index i as a numpy array. The base_object // should be the interpreter object providing the memory. PyObject* tensor(PyObject* base_object, int i); private: - InterpreterWrapper(std::unique_ptr model); + InterpreterWrapper(std::unique_ptr model, + std::unique_ptr error_reporter); // InterpreterWrapper is not copyable or assignable. We avoid the use of // InterpreterWrapper() = delete here for SWIG compatibility. @@ -71,6 +78,7 @@ class InterpreterWrapper { InterpreterWrapper(const InterpreterWrapper& rhs); const std::unique_ptr model_; + const std::unique_ptr error_reporter_; const std::unique_ptr resolver_; const std::unique_ptr interpreter_; }; diff --git a/tensorflow/contrib/lite/python/interpreter_wrapper/interpreter_wrapper.i b/tensorflow/contrib/lite/python/interpreter_wrapper/interpreter_wrapper.i index 7f51f9f00d1b2fe057052f7b7bd52bcb65231164..afb2092eacab1d8dcccf8c75cee1d8d5c34d7e75 100644 --- a/tensorflow/contrib/lite/python/interpreter_wrapper/interpreter_wrapper.i +++ b/tensorflow/contrib/lite/python/interpreter_wrapper/interpreter_wrapper.i @@ -18,8 +18,51 @@ limitations under the License. %{ #define SWIG_FILE_WITH_INIT +#include "tensorflow/contrib/lite/interpreter.h" +#include "tensorflow/contrib/lite/model.h" #include "tensorflow/contrib/lite/python/interpreter_wrapper/interpreter_wrapper.h" %} %include "tensorflow/contrib/lite/python/interpreter_wrapper/interpreter_wrapper.h" + +namespace tflite { +namespace interpreter_wrapper { +%extend InterpreterWrapper { + + // Version of the constructor that handles producing Python exceptions + // that propagate strings. + static PyObject* CreateWrapperCPPFromFile(const char* model_path) { + std::string error; + if(tflite::interpreter_wrapper::InterpreterWrapper* ptr = + tflite::interpreter_wrapper::InterpreterWrapper + ::CreateWrapperCPPFromFile( + model_path, &error)) { + return SWIG_NewPointerObj( + ptr, SWIGTYPE_p_tflite__interpreter_wrapper__InterpreterWrapper, 1); + } else { + PyErr_SetString(PyExc_ValueError, error.c_str()); + return nullptr; + } + } + + // Version of the constructor that handles producing Python exceptions + // that propagate strings. + static PyObject* CreateWrapperCPPFromBuffer( + PyObject* data) { + std::string error; + if(tflite::interpreter_wrapper::InterpreterWrapper* ptr = + tflite::interpreter_wrapper::InterpreterWrapper + ::CreateWrapperCPPFromBuffer( + data, &error)) { + return SWIG_NewPointerObj( + ptr, SWIGTYPE_p_tflite__interpreter_wrapper__InterpreterWrapper, 1); + } else { + PyErr_SetString(PyExc_ValueError, error.c_str()); + return nullptr; + } + } +} + +} // namespace interpreter_wrapper +} // namespace tflite diff --git a/tensorflow/contrib/lite/python/lite.py b/tensorflow/contrib/lite/python/lite.py index 69a2f638af2a7926001be4c517e8dac9d2b6fe3d..29a1487c1f468055dde85ef6c2657a50f3d2f32b 100644 --- a/tensorflow/contrib/lite/python/lite.py +++ b/tensorflow/contrib/lite/python/lite.py @@ -50,6 +50,7 @@ from tensorflow.contrib.lite.python.interpreter import Interpreter # pylint: di from tensorflow.contrib.lite.python.op_hint import convert_op_hints_to_stubs # pylint: disable=unused-import from tensorflow.contrib.lite.python.op_hint import OpHint # pylint: disable=unused-import from tensorflow.core.framework import graph_pb2 as _graph_pb2 +from tensorflow.python import keras as _keras from tensorflow.python.client import session as _session from tensorflow.python.framework import graph_util as tf_graph_util from tensorflow.python.framework.importer import import_graph_def @@ -131,7 +132,7 @@ class TocoConverter(object): Args: - graph_def: TensorFlow GraphDef. + graph_def: Frozen TensorFlow GraphDef. input_tensors: List of input tensors. Type and shape are computed using `foo.get_shape()` and `foo.dtype`. output_tensors: List of output tensors (only .name is used from this). @@ -177,7 +178,7 @@ class TocoConverter(object): """Creates a TocoConverter class from a file containing a frozen GraphDef. Args: - graph_def_file: Full filepath of file containing TensorFlow GraphDef. + graph_def_file: Full filepath of file containing frozen GraphDef. input_arrays: List of input tensors to freeze graph with. output_arrays: List of output tensors to freeze graph with. input_shapes: Dict of strings representing input tensor names to list of @@ -269,6 +270,48 @@ class TocoConverter(object): return cls( graph_def=result[0], input_tensors=result[1], output_tensors=result[2]) + @classmethod + def from_keras_model_file(cls, + model_file, + input_arrays=None, + input_shapes=None, + output_arrays=None): + """Creates a TocoConverter class from a tf.keras model file. + + Args: + model_file: Full filepath of HDF5 file containing the tf.keras model. + input_arrays: List of input tensors to freeze graph with. Uses input + arrays from SignatureDef when none are provided. (default None) + input_shapes: Dict of strings representing input tensor names to list of + integers representing input shapes (e.g., {"foo" : [1, 16, 16, 3]}). + Automatically determined when input shapes is None (e.g., {"foo" : + None}). (default None) + output_arrays: List of output tensors to freeze graph with. Uses output + arrays from SignatureDef when none are provided. (default None) + + Returns: + TocoConverter class. + """ + _keras.backend.clear_session() + _keras.backend.set_learning_phase(False) + keras_model = _keras.models.load_model(model_file) + sess = _keras.backend.get_session() + + # Get input and output tensors. + if input_arrays: + input_tensors = get_tensors_from_tensor_names(sess.graph, input_arrays) + else: + input_tensors = keras_model.inputs + + if output_arrays: + output_tensors = get_tensors_from_tensor_names(sess.graph, output_arrays) + else: + output_tensors = keras_model.outputs + set_tensor_shapes(input_tensors, input_shapes) + + graph_def = _freeze_graph(sess, output_tensors) + return cls(graph_def, input_tensors, output_tensors) + def convert(self): """Converts a TensorFlow GraphDef based on instance variables. @@ -366,7 +409,7 @@ def _is_frozen_graph(sess): Bool. """ for op in sess.graph.get_operations(): - if op.type.startswith("Variable"): + if op.type.startswith("Variable") or op.type.endswith("VariableOp"): return False return True diff --git a/tensorflow/contrib/lite/python/lite_test.py b/tensorflow/contrib/lite/python/lite_test.py index a9475de47408d7d451663cdf021d40eaa85c7c63..ca2af5aaed3ee4f4fce5f0d31eaa61df0e11f364 100644 --- a/tensorflow/contrib/lite/python/lite_test.py +++ b/tensorflow/contrib/lite/python/lite_test.py @@ -19,11 +19,13 @@ from __future__ import division from __future__ import print_function import os +import tempfile import numpy as np from tensorflow.contrib.lite.python import lite from tensorflow.contrib.lite.python import lite_constants from tensorflow.contrib.lite.python.interpreter import Interpreter +from tensorflow.python import keras from tensorflow.python.client import session from tensorflow.python.framework import constant_op from tensorflow.python.framework import dtypes @@ -618,5 +620,279 @@ class FromSavedModelTest(test_util.TensorFlowTestCase): self.assertTrue(tflite_model) +class FromKerasFile(test_util.TensorFlowTestCase): + + def setUp(self): + keras.backend.clear_session() + + def _getSequentialModel(self): + model = keras.models.Sequential() + model.add(keras.layers.Dense(2, input_shape=(3,))) + model.add(keras.layers.RepeatVector(3)) + model.add(keras.layers.TimeDistributed(keras.layers.Dense(3))) + model.compile( + loss=keras.losses.MSE, + optimizer=keras.optimizers.RMSprop(), + metrics=[keras.metrics.categorical_accuracy], + sample_weight_mode='temporal') + x = np.random.random((1, 3)) + y = np.random.random((1, 3, 3)) + model.train_on_batch(x, y) + model.predict(x) + + try: + fd, keras_file = tempfile.mkstemp('.h5') + keras.models.save_model(model, keras_file) + finally: + os.close(fd) + return keras_file + + def testSequentialModel(self): + """Test a Sequential tf.keras model with default inputs.""" + keras_file = self._getSequentialModel() + + converter = lite.TocoConverter.from_keras_model_file(keras_file) + tflite_model = converter.convert() + self.assertTrue(tflite_model) + + os.remove(keras_file) + + # Check values from converted model. + interpreter = Interpreter(model_content=tflite_model) + interpreter.allocate_tensors() + + input_details = interpreter.get_input_details() + self.assertEqual(1, len(input_details)) + self.assertEqual('dense_input', input_details[0]['name']) + self.assertEqual(np.float32, input_details[0]['dtype']) + self.assertTrue(([1, 3] == input_details[0]['shape']).all()) + self.assertEqual((0., 0.), input_details[0]['quantization']) + + output_details = interpreter.get_output_details() + self.assertEqual(1, len(output_details)) + self.assertEqual('time_distributed/Reshape_1', output_details[0]['name']) + self.assertEqual(np.float32, output_details[0]['dtype']) + self.assertTrue(([1, 3, 3] == output_details[0]['shape']).all()) + self.assertEqual((0., 0.), output_details[0]['quantization']) + + def testSequentialModelInputArray(self): + """Test a Sequential tf.keras model testing input arrays argument.""" + keras_file = self._getSequentialModel() + + # Invalid input array raises error. + with self.assertRaises(ValueError) as error: + lite.TocoConverter.from_keras_model_file( + keras_file, input_arrays=['invalid-input']) + self.assertEqual("Invalid tensors 'invalid-input' were found.", + str(error.exception)) + + # Valid input array. + converter = lite.TocoConverter.from_keras_model_file( + keras_file, input_arrays=['dense_input']) + tflite_model = converter.convert() + os.remove(keras_file) + self.assertTrue(tflite_model) + + def testSequentialModelInputShape(self): + """Test a Sequential tf.keras model testing input shapes argument.""" + keras_file = self._getSequentialModel() + + # Passing in shape of invalid input array has no impact as long as all input + # arrays have a shape. + converter = lite.TocoConverter.from_keras_model_file( + keras_file, input_shapes={'invalid-input': [2, 3]}) + tflite_model = converter.convert() + self.assertTrue(tflite_model) + + # Passing in shape of valid input array. + converter = lite.TocoConverter.from_keras_model_file( + keras_file, input_shapes={'dense_input': [2, 3]}) + tflite_model = converter.convert() + os.remove(keras_file) + self.assertTrue(tflite_model) + + # Check input shape from converted model. + interpreter = Interpreter(model_content=tflite_model) + interpreter.allocate_tensors() + + input_details = interpreter.get_input_details() + self.assertEqual(1, len(input_details)) + self.assertEqual('dense_input', input_details[0]['name']) + self.assertTrue(([2, 3] == input_details[0]['shape']).all()) + + def testSequentialModelOutputArray(self): + """Test a Sequential tf.keras model testing output arrays argument.""" + keras_file = self._getSequentialModel() + + # Invalid output array raises error. + with self.assertRaises(ValueError) as error: + lite.TocoConverter.from_keras_model_file( + keras_file, output_arrays=['invalid-output']) + self.assertEqual("Invalid tensors 'invalid-output' were found.", + str(error.exception)) + + # Valid output array. + converter = lite.TocoConverter.from_keras_model_file( + keras_file, output_arrays=['time_distributed/Reshape_1']) + tflite_model = converter.convert() + os.remove(keras_file) + self.assertTrue(tflite_model) + + def testFunctionalModel(self): + """Test a Functional tf.keras model with default inputs.""" + inputs = keras.layers.Input(shape=(3,), name='input') + x = keras.layers.Dense(2)(inputs) + output = keras.layers.Dense(3)(x) + + model = keras.models.Model(inputs, output) + model.compile( + loss=keras.losses.MSE, + optimizer=keras.optimizers.RMSprop(), + metrics=[keras.metrics.categorical_accuracy]) + x = np.random.random((1, 3)) + y = np.random.random((1, 3)) + model.train_on_batch(x, y) + + model.predict(x) + fd, keras_file = tempfile.mkstemp('.h5') + keras.models.save_model(model, keras_file) + + # Convert to TFLite model. + converter = lite.TocoConverter.from_keras_model_file(keras_file) + tflite_model = converter.convert() + self.assertTrue(tflite_model) + + os.close(fd) + os.remove(keras_file) + + # Check values from converted model. + interpreter = Interpreter(model_content=tflite_model) + interpreter.allocate_tensors() + + input_details = interpreter.get_input_details() + self.assertEqual(1, len(input_details)) + self.assertEqual('input', input_details[0]['name']) + self.assertEqual(np.float32, input_details[0]['dtype']) + self.assertTrue(([1, 3] == input_details[0]['shape']).all()) + self.assertEqual((0., 0.), input_details[0]['quantization']) + + output_details = interpreter.get_output_details() + self.assertEqual(1, len(output_details)) + self.assertEqual('dense_1/BiasAdd', output_details[0]['name']) + self.assertEqual(np.float32, output_details[0]['dtype']) + self.assertTrue(([1, 3] == output_details[0]['shape']).all()) + self.assertEqual((0., 0.), output_details[0]['quantization']) + + def testFunctionalModelMultipleInputs(self): + """Test a Functional tf.keras model with multiple inputs and outputs.""" + a = keras.layers.Input(shape=(3,), name='input_a') + b = keras.layers.Input(shape=(3,), name='input_b') + dense = keras.layers.Dense(4, name='dense') + c = dense(a) + d = dense(b) + e = keras.layers.Dropout(0.5, name='dropout')(c) + + model = keras.models.Model([a, b], [d, e]) + model.compile( + loss=keras.losses.MSE, + optimizer=keras.optimizers.RMSprop(), + metrics=[keras.metrics.mae], + loss_weights=[1., 0.5]) + + input_a_np = np.random.random((10, 3)) + input_b_np = np.random.random((10, 3)) + output_d_np = np.random.random((10, 4)) + output_e_np = np.random.random((10, 4)) + model.train_on_batch([input_a_np, input_b_np], [output_d_np, output_e_np]) + + model.predict([input_a_np, input_b_np], batch_size=5) + fd, keras_file = tempfile.mkstemp('.h5') + keras.models.save_model(model, keras_file) + + # Convert to TFLite model. + converter = lite.TocoConverter.from_keras_model_file(keras_file) + tflite_model = converter.convert() + self.assertTrue(tflite_model) + + os.close(fd) + os.remove(keras_file) + + # Check values from converted model. + interpreter = Interpreter(model_content=tflite_model) + interpreter.allocate_tensors() + + input_details = interpreter.get_input_details() + self.assertEqual(2, len(input_details)) + self.assertEqual('input_a', input_details[0]['name']) + self.assertEqual(np.float32, input_details[0]['dtype']) + self.assertTrue(([1, 3] == input_details[0]['shape']).all()) + self.assertEqual((0., 0.), input_details[0]['quantization']) + + self.assertEqual('input_b', input_details[1]['name']) + self.assertEqual(np.float32, input_details[1]['dtype']) + self.assertTrue(([1, 3] == input_details[1]['shape']).all()) + self.assertEqual((0., 0.), input_details[1]['quantization']) + + output_details = interpreter.get_output_details() + self.assertEqual(2, len(output_details)) + self.assertEqual('dense_1/BiasAdd', output_details[0]['name']) + self.assertEqual(np.float32, output_details[0]['dtype']) + self.assertTrue(([1, 4] == output_details[0]['shape']).all()) + self.assertEqual((0., 0.), output_details[0]['quantization']) + + self.assertEqual('dropout/Identity', output_details[1]['name']) + self.assertEqual(np.float32, output_details[1]['dtype']) + self.assertTrue(([1, 4] == output_details[1]['shape']).all()) + self.assertEqual((0., 0.), output_details[1]['quantization']) + + def testFunctionalSequentialModel(self): + """Test a Functional tf.keras model containing a Sequential model.""" + model = keras.models.Sequential() + model.add(keras.layers.Dense(2, input_shape=(3,))) + model.add(keras.layers.RepeatVector(3)) + model.add(keras.layers.TimeDistributed(keras.layers.Dense(3))) + model = keras.models.Model(model.input, model.output) + + model.compile( + loss=keras.losses.MSE, + optimizer=keras.optimizers.RMSprop(), + metrics=[keras.metrics.categorical_accuracy], + sample_weight_mode='temporal') + x = np.random.random((1, 3)) + y = np.random.random((1, 3, 3)) + model.train_on_batch(x, y) + model.predict(x) + + model.predict(x) + fd, keras_file = tempfile.mkstemp('.h5') + keras.models.save_model(model, keras_file) + + # Convert to TFLite model. + converter = lite.TocoConverter.from_keras_model_file(keras_file) + tflite_model = converter.convert() + self.assertTrue(tflite_model) + + os.close(fd) + os.remove(keras_file) + + # Check values from converted model. + interpreter = Interpreter(model_content=tflite_model) + interpreter.allocate_tensors() + + input_details = interpreter.get_input_details() + self.assertEqual(1, len(input_details)) + self.assertEqual('dense_input', input_details[0]['name']) + self.assertEqual(np.float32, input_details[0]['dtype']) + self.assertTrue(([1, 3] == input_details[0]['shape']).all()) + self.assertEqual((0., 0.), input_details[0]['quantization']) + + output_details = interpreter.get_output_details() + self.assertEqual(1, len(output_details)) + self.assertEqual('time_distributed/Reshape_1', output_details[0]['name']) + self.assertEqual(np.float32, output_details[0]['dtype']) + self.assertTrue(([1, 3, 3] == output_details[0]['shape']).all()) + self.assertEqual((0., 0.), output_details[0]['quantization']) + + if __name__ == '__main__': test.main() diff --git a/tensorflow/contrib/lite/python/tflite_convert.py b/tensorflow/contrib/lite/python/tflite_convert.py index d18a29834b6341938277ceea8233e535b55ddc9d..9bd1f4f76ee693414a8515a5bd2567001b53e2ea 100644 --- a/tensorflow/contrib/lite/python/tflite_convert.py +++ b/tensorflow/contrib/lite/python/tflite_convert.py @@ -74,6 +74,9 @@ def _get_toco_converter(flags): converter_kwargs["saved_model_dir"] = flags.saved_model_dir converter_kwargs["tag_set"] = _parse_set(flags.saved_model_tag_set) converter_kwargs["signature_key"] = flags.saved_model_signature_key + elif flags.keras_model_file: + converter_fn = lite.TocoConverter.from_keras_model_file + converter_kwargs["model_file"] = flags.keras_model_file return converter_fn(**converter_kwargs) @@ -102,7 +105,7 @@ def _convert_model(flags): input_arrays = converter.get_input_arrays() std_dev_values = _parse_array(flags.std_dev_values, type_fn=int) mean_values = _parse_array(flags.mean_values, type_fn=int) - quant_stats = zip(mean_values, std_dev_values) + quant_stats = list(zip(mean_values, std_dev_values)) if ((not flags.input_arrays and len(input_arrays) > 1) or (len(input_arrays) != len(quant_stats))): raise ValueError("Mismatching --input_arrays, --std_dev_values, and " @@ -222,11 +225,15 @@ def run_main(_): input_file_group.add_argument( "--graph_def_file", type=str, - help="Full filepath of file containing TensorFlow GraphDef.") + help="Full filepath of file containing frozen TensorFlow GraphDef.") input_file_group.add_argument( "--saved_model_dir", type=str, help="Full filepath of directory containing the SavedModel.") + input_file_group.add_argument( + "--keras_model_file", + type=str, + help="Full filepath of HDF5 file containing tf.Keras model.") # Model format flags. parser.add_argument( diff --git a/tensorflow/contrib/lite/schema/BUILD b/tensorflow/contrib/lite/schema/BUILD index 9717a4a1a496b888348514584888e62c4e3703b4..f095151cae835aa202ff4c9f43e175246f54f1cf 100644 --- a/tensorflow/contrib/lite/schema/BUILD +++ b/tensorflow/contrib/lite/schema/BUILD @@ -65,6 +65,7 @@ cc_test( ], tags = [ "tflite_not_portable_android", + "tflite_not_portable_ios", ], deps = [ "//tensorflow/core:lib_platform", diff --git a/tensorflow/contrib/lite/schema/schema.fbs b/tensorflow/contrib/lite/schema/schema.fbs index df43f1e5abf921410b14912e33562bf1a7067795..64830b1dc31ed9ee4711fe30badfc6b4e72593e5 100644 --- a/tensorflow/contrib/lite/schema/schema.fbs +++ b/tensorflow/contrib/lite/schema/schema.fbs @@ -35,6 +35,7 @@ enum TensorType : byte { STRING = 5, BOOL = 6, INT16 = 7, + COMPLEX64 = 8, } // Parameters for converting a quantized tensor back to float. Given a @@ -43,7 +44,7 @@ enum TensorType : byte { table QuantizationParameters { min:[float]; // For importing back into tensorflow. max:[float]; // For importing back into tensorflow. - scale:[float]; + scale:[float]; // For dequantizing the tensor's values. zero_point:[long]; } @@ -158,6 +159,9 @@ enum BuiltinOperator : byte { SQRT = 75, RSQRT = 76, SHAPE = 77, + POW = 78, + ARG_MIN = 79, + FAKE_QUANT = 80, } // Options for the builtin operators. @@ -217,6 +221,9 @@ union BuiltinOptions { EqualOptions, NotEqualOptions, ShapeOptions, + PowOptions, + ArgMinOptions, + FakeQuantOptions, } enum Padding : byte { SAME, VALID } @@ -294,9 +301,18 @@ table BidirectionalSequenceRNNOptions { fused_activation_function:ActivationFunctionType; } +enum FullyConnectedOptionsWeightsFormat: byte { + DEFAULT = 0, + SHUFFLED4x16INT8 = 1, +} + // An implementation of TensorFlow fully_connected (a.k.a Dense) layer. table FullyConnectedOptions { + // Parameters for FullyConnected version 1 or above. fused_activation_function:ActivationFunctionType; + + // Parameters for FullyConnected version 2 or above. + weights_format:FullyConnectedOptionsWeightsFormat = DEFAULT; } table SoftmaxOptions { @@ -457,6 +473,10 @@ table ArgMaxOptions { output_type : TensorType; } +table ArgMinOptions { + output_type : TensorType; +} + table GreaterOptions { } @@ -502,6 +522,19 @@ table ShapeOptions { out_type : TensorType; } +table PowOptions { +} + +table FakeQuantOptions { + // Parameters supported by version 1: + min:float; + max:float; + num_bits:int; + + // Parameters supported by version 2: + narrow_range:bool; +} + // An OperatorCode can be an enum value (BuiltinOperator) if the operator is a // builtin, or a string if the operator is custom. table OperatorCode { diff --git a/tensorflow/contrib/lite/schema/schema_generated.h b/tensorflow/contrib/lite/schema/schema_generated.h index 8c0660dfe201ac7ad0f45b6fd234c213a06416b6..c0b57039cbb90b468742bd39d0c6fdd4bf6e1d3c 100755 --- a/tensorflow/contrib/lite/schema/schema_generated.h +++ b/tensorflow/contrib/lite/schema/schema_generated.h @@ -157,6 +157,9 @@ struct TileOptionsT; struct ArgMaxOptions; struct ArgMaxOptionsT; +struct ArgMinOptions; +struct ArgMinOptionsT; + struct GreaterOptions; struct GreaterOptionsT; @@ -196,6 +199,12 @@ struct NotEqualOptionsT; struct ShapeOptions; struct ShapeOptionsT; +struct PowOptions; +struct PowOptionsT; + +struct FakeQuantOptions; +struct FakeQuantOptionsT; + struct OperatorCode; struct OperatorCodeT; @@ -220,11 +229,12 @@ enum TensorType { TensorType_STRING = 5, TensorType_BOOL = 6, TensorType_INT16 = 7, + TensorType_COMPLEX64 = 8, TensorType_MIN = TensorType_FLOAT32, - TensorType_MAX = TensorType_INT16 + TensorType_MAX = TensorType_COMPLEX64 }; -inline TensorType (&EnumValuesTensorType())[8] { +inline TensorType (&EnumValuesTensorType())[9] { static TensorType values[] = { TensorType_FLOAT32, TensorType_FLOAT16, @@ -233,7 +243,8 @@ inline TensorType (&EnumValuesTensorType())[8] { TensorType_INT64, TensorType_STRING, TensorType_BOOL, - TensorType_INT16 + TensorType_INT16, + TensorType_COMPLEX64 }; return values; } @@ -248,6 +259,7 @@ inline const char **EnumNamesTensorType() { "STRING", "BOOL", "INT16", + "COMPLEX64", nullptr }; return names; @@ -336,11 +348,14 @@ enum BuiltinOperator { BuiltinOperator_SQRT = 75, BuiltinOperator_RSQRT = 76, BuiltinOperator_SHAPE = 77, + BuiltinOperator_POW = 78, + BuiltinOperator_ARG_MIN = 79, + BuiltinOperator_FAKE_QUANT = 80, BuiltinOperator_MIN = BuiltinOperator_ADD, - BuiltinOperator_MAX = BuiltinOperator_SHAPE + BuiltinOperator_MAX = BuiltinOperator_FAKE_QUANT }; -inline BuiltinOperator (&EnumValuesBuiltinOperator())[77] { +inline BuiltinOperator (&EnumValuesBuiltinOperator())[80] { static BuiltinOperator values[] = { BuiltinOperator_ADD, BuiltinOperator_AVERAGE_POOL_2D, @@ -418,7 +433,10 @@ inline BuiltinOperator (&EnumValuesBuiltinOperator())[77] { BuiltinOperator_SUM, BuiltinOperator_SQRT, BuiltinOperator_RSQRT, - BuiltinOperator_SHAPE + BuiltinOperator_SHAPE, + BuiltinOperator_POW, + BuiltinOperator_ARG_MIN, + BuiltinOperator_FAKE_QUANT }; return values; } @@ -503,6 +521,9 @@ inline const char **EnumNamesBuiltinOperator() { "SQRT", "RSQRT", "SHAPE", + "POW", + "ARG_MIN", + "FAKE_QUANT", nullptr }; return names; @@ -570,11 +591,14 @@ enum BuiltinOptions { BuiltinOptions_EqualOptions = 53, BuiltinOptions_NotEqualOptions = 54, BuiltinOptions_ShapeOptions = 55, + BuiltinOptions_PowOptions = 56, + BuiltinOptions_ArgMinOptions = 57, + BuiltinOptions_FakeQuantOptions = 58, BuiltinOptions_MIN = BuiltinOptions_NONE, - BuiltinOptions_MAX = BuiltinOptions_ShapeOptions + BuiltinOptions_MAX = BuiltinOptions_FakeQuantOptions }; -inline BuiltinOptions (&EnumValuesBuiltinOptions())[56] { +inline BuiltinOptions (&EnumValuesBuiltinOptions())[59] { static BuiltinOptions values[] = { BuiltinOptions_NONE, BuiltinOptions_Conv2DOptions, @@ -631,7 +655,10 @@ inline BuiltinOptions (&EnumValuesBuiltinOptions())[56] { BuiltinOptions_ExpandDimsOptions, BuiltinOptions_EqualOptions, BuiltinOptions_NotEqualOptions, - BuiltinOptions_ShapeOptions + BuiltinOptions_ShapeOptions, + BuiltinOptions_PowOptions, + BuiltinOptions_ArgMinOptions, + BuiltinOptions_FakeQuantOptions }; return values; } @@ -694,6 +721,9 @@ inline const char **EnumNamesBuiltinOptions() { "EqualOptions", "NotEqualOptions", "ShapeOptions", + "PowOptions", + "ArgMinOptions", + "FakeQuantOptions", nullptr }; return names; @@ -928,6 +958,18 @@ template<> struct BuiltinOptionsTraits { static const BuiltinOptions enum_value = BuiltinOptions_ShapeOptions; }; +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_PowOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_ArgMinOptions; +}; + +template<> struct BuiltinOptionsTraits { + static const BuiltinOptions enum_value = BuiltinOptions_FakeQuantOptions; +}; + struct BuiltinOptionsUnion { BuiltinOptions type; void *value; @@ -1399,6 +1441,30 @@ struct BuiltinOptionsUnion { return type == BuiltinOptions_ShapeOptions ? reinterpret_cast(value) : nullptr; } + PowOptionsT *AsPowOptions() { + return type == BuiltinOptions_PowOptions ? + reinterpret_cast(value) : nullptr; + } + const PowOptionsT *AsPowOptions() const { + return type == BuiltinOptions_PowOptions ? + reinterpret_cast(value) : nullptr; + } + ArgMinOptionsT *AsArgMinOptions() { + return type == BuiltinOptions_ArgMinOptions ? + reinterpret_cast(value) : nullptr; + } + const ArgMinOptionsT *AsArgMinOptions() const { + return type == BuiltinOptions_ArgMinOptions ? + reinterpret_cast(value) : nullptr; + } + FakeQuantOptionsT *AsFakeQuantOptions() { + return type == BuiltinOptions_FakeQuantOptions ? + reinterpret_cast(value) : nullptr; + } + const FakeQuantOptionsT *AsFakeQuantOptions() const { + return type == BuiltinOptions_FakeQuantOptions ? + reinterpret_cast(value) : nullptr; + } }; bool VerifyBuiltinOptions(flatbuffers::Verifier &verifier, const void *obj, BuiltinOptions type); @@ -1506,6 +1572,35 @@ inline const char *EnumNameLSHProjectionType(LSHProjectionType e) { return EnumNamesLSHProjectionType()[index]; } +enum FullyConnectedOptionsWeightsFormat { + FullyConnectedOptionsWeightsFormat_DEFAULT = 0, + FullyConnectedOptionsWeightsFormat_SHUFFLED4x16INT8 = 1, + FullyConnectedOptionsWeightsFormat_MIN = FullyConnectedOptionsWeightsFormat_DEFAULT, + FullyConnectedOptionsWeightsFormat_MAX = FullyConnectedOptionsWeightsFormat_SHUFFLED4x16INT8 +}; + +inline FullyConnectedOptionsWeightsFormat (&EnumValuesFullyConnectedOptionsWeightsFormat())[2] { + static FullyConnectedOptionsWeightsFormat values[] = { + FullyConnectedOptionsWeightsFormat_DEFAULT, + FullyConnectedOptionsWeightsFormat_SHUFFLED4x16INT8 + }; + return values; +} + +inline const char **EnumNamesFullyConnectedOptionsWeightsFormat() { + static const char *names[] = { + "DEFAULT", + "SHUFFLED4x16INT8", + nullptr + }; + return names; +} + +inline const char *EnumNameFullyConnectedOptionsWeightsFormat(FullyConnectedOptionsWeightsFormat e) { + const size_t index = static_cast(e); + return EnumNamesFullyConnectedOptionsWeightsFormat()[index]; +} + enum LSTMKernelType { LSTMKernelType_FULL = 0, LSTMKernelType_BASIC = 1, @@ -2558,22 +2653,29 @@ flatbuffers::Offset CreateBidirectionalSequence struct FullyConnectedOptionsT : public flatbuffers::NativeTable { typedef FullyConnectedOptions TableType; ActivationFunctionType fused_activation_function; + FullyConnectedOptionsWeightsFormat weights_format; FullyConnectedOptionsT() - : fused_activation_function(ActivationFunctionType_NONE) { + : fused_activation_function(ActivationFunctionType_NONE), + weights_format(FullyConnectedOptionsWeightsFormat_DEFAULT) { } }; struct FullyConnectedOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { typedef FullyConnectedOptionsT NativeTableType; enum { - VT_FUSED_ACTIVATION_FUNCTION = 4 + VT_FUSED_ACTIVATION_FUNCTION = 4, + VT_WEIGHTS_FORMAT = 6 }; ActivationFunctionType fused_activation_function() const { return static_cast(GetField(VT_FUSED_ACTIVATION_FUNCTION, 0)); } + FullyConnectedOptionsWeightsFormat weights_format() const { + return static_cast(GetField(VT_WEIGHTS_FORMAT, 0)); + } bool Verify(flatbuffers::Verifier &verifier) const { return VerifyTableStart(verifier) && VerifyField(verifier, VT_FUSED_ACTIVATION_FUNCTION) && + VerifyField(verifier, VT_WEIGHTS_FORMAT) && verifier.EndTable(); } FullyConnectedOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; @@ -2587,6 +2689,9 @@ struct FullyConnectedOptionsBuilder { void add_fused_activation_function(ActivationFunctionType fused_activation_function) { fbb_.AddElement(FullyConnectedOptions::VT_FUSED_ACTIVATION_FUNCTION, static_cast(fused_activation_function), 0); } + void add_weights_format(FullyConnectedOptionsWeightsFormat weights_format) { + fbb_.AddElement(FullyConnectedOptions::VT_WEIGHTS_FORMAT, static_cast(weights_format), 0); + } explicit FullyConnectedOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) : fbb_(_fbb) { start_ = fbb_.StartTable(); @@ -2601,8 +2706,10 @@ struct FullyConnectedOptionsBuilder { inline flatbuffers::Offset CreateFullyConnectedOptions( flatbuffers::FlatBufferBuilder &_fbb, - ActivationFunctionType fused_activation_function = ActivationFunctionType_NONE) { + ActivationFunctionType fused_activation_function = ActivationFunctionType_NONE, + FullyConnectedOptionsWeightsFormat weights_format = FullyConnectedOptionsWeightsFormat_DEFAULT) { FullyConnectedOptionsBuilder builder_(_fbb); + builder_.add_weights_format(weights_format); builder_.add_fused_activation_function(fused_activation_function); return builder_.Finish(); } @@ -4421,6 +4528,60 @@ inline flatbuffers::Offset CreateArgMaxOptions( flatbuffers::Offset CreateArgMaxOptions(flatbuffers::FlatBufferBuilder &_fbb, const ArgMaxOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +struct ArgMinOptionsT : public flatbuffers::NativeTable { + typedef ArgMinOptions TableType; + TensorType output_type; + ArgMinOptionsT() + : output_type(TensorType_FLOAT32) { + } +}; + +struct ArgMinOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef ArgMinOptionsT NativeTableType; + enum { + VT_OUTPUT_TYPE = 4 + }; + TensorType output_type() const { + return static_cast(GetField(VT_OUTPUT_TYPE, 0)); + } + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyField(verifier, VT_OUTPUT_TYPE) && + verifier.EndTable(); + } + ArgMinOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(ArgMinOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const ArgMinOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct ArgMinOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + void add_output_type(TensorType output_type) { + fbb_.AddElement(ArgMinOptions::VT_OUTPUT_TYPE, static_cast(output_type), 0); + } + explicit ArgMinOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + ArgMinOptionsBuilder &operator=(const ArgMinOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateArgMinOptions( + flatbuffers::FlatBufferBuilder &_fbb, + TensorType output_type = TensorType_FLOAT32) { + ArgMinOptionsBuilder builder_(_fbb); + builder_.add_output_type(output_type); + return builder_.Finish(); +} + +flatbuffers::Offset CreateArgMinOptions(flatbuffers::FlatBufferBuilder &_fbb, const ArgMinOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + struct GreaterOptionsT : public flatbuffers::NativeTable { typedef GreaterOptions TableType; GreaterOptionsT() { @@ -5007,6 +5168,136 @@ inline flatbuffers::Offset CreateShapeOptions( flatbuffers::Offset CreateShapeOptions(flatbuffers::FlatBufferBuilder &_fbb, const ShapeOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +struct PowOptionsT : public flatbuffers::NativeTable { + typedef PowOptions TableType; + PowOptionsT() { + } +}; + +struct PowOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef PowOptionsT NativeTableType; + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + verifier.EndTable(); + } + PowOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(PowOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const PowOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct PowOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + explicit PowOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + PowOptionsBuilder &operator=(const PowOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreatePowOptions( + flatbuffers::FlatBufferBuilder &_fbb) { + PowOptionsBuilder builder_(_fbb); + return builder_.Finish(); +} + +flatbuffers::Offset CreatePowOptions(flatbuffers::FlatBufferBuilder &_fbb, const PowOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + +struct FakeQuantOptionsT : public flatbuffers::NativeTable { + typedef FakeQuantOptions TableType; + float min; + float max; + int32_t num_bits; + bool narrow_range; + FakeQuantOptionsT() + : min(0.0f), + max(0.0f), + num_bits(0), + narrow_range(false) { + } +}; + +struct FakeQuantOptions FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { + typedef FakeQuantOptionsT NativeTableType; + enum { + VT_MIN = 4, + VT_MAX = 6, + VT_NUM_BITS = 8, + VT_NARROW_RANGE = 10 + }; + float min() const { + return GetField(VT_MIN, 0.0f); + } + float max() const { + return GetField(VT_MAX, 0.0f); + } + int32_t num_bits() const { + return GetField(VT_NUM_BITS, 0); + } + bool narrow_range() const { + return GetField(VT_NARROW_RANGE, 0) != 0; + } + bool Verify(flatbuffers::Verifier &verifier) const { + return VerifyTableStart(verifier) && + VerifyField(verifier, VT_MIN) && + VerifyField(verifier, VT_MAX) && + VerifyField(verifier, VT_NUM_BITS) && + VerifyField(verifier, VT_NARROW_RANGE) && + verifier.EndTable(); + } + FakeQuantOptionsT *UnPack(const flatbuffers::resolver_function_t *_resolver = nullptr) const; + void UnPackTo(FakeQuantOptionsT *_o, const flatbuffers::resolver_function_t *_resolver = nullptr) const; + static flatbuffers::Offset Pack(flatbuffers::FlatBufferBuilder &_fbb, const FakeQuantOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); +}; + +struct FakeQuantOptionsBuilder { + flatbuffers::FlatBufferBuilder &fbb_; + flatbuffers::uoffset_t start_; + void add_min(float min) { + fbb_.AddElement(FakeQuantOptions::VT_MIN, min, 0.0f); + } + void add_max(float max) { + fbb_.AddElement(FakeQuantOptions::VT_MAX, max, 0.0f); + } + void add_num_bits(int32_t num_bits) { + fbb_.AddElement(FakeQuantOptions::VT_NUM_BITS, num_bits, 0); + } + void add_narrow_range(bool narrow_range) { + fbb_.AddElement(FakeQuantOptions::VT_NARROW_RANGE, static_cast(narrow_range), 0); + } + explicit FakeQuantOptionsBuilder(flatbuffers::FlatBufferBuilder &_fbb) + : fbb_(_fbb) { + start_ = fbb_.StartTable(); + } + FakeQuantOptionsBuilder &operator=(const FakeQuantOptionsBuilder &); + flatbuffers::Offset Finish() { + const auto end = fbb_.EndTable(start_); + auto o = flatbuffers::Offset(end); + return o; + } +}; + +inline flatbuffers::Offset CreateFakeQuantOptions( + flatbuffers::FlatBufferBuilder &_fbb, + float min = 0.0f, + float max = 0.0f, + int32_t num_bits = 0, + bool narrow_range = false) { + FakeQuantOptionsBuilder builder_(_fbb); + builder_.add_num_bits(num_bits); + builder_.add_max(max); + builder_.add_min(min); + builder_.add_narrow_range(narrow_range); + return builder_.Finish(); +} + +flatbuffers::Offset CreateFakeQuantOptions(flatbuffers::FlatBufferBuilder &_fbb, const FakeQuantOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher = nullptr); + struct OperatorCodeT : public flatbuffers::NativeTable { typedef OperatorCode TableType; BuiltinOperator builtin_code; @@ -5305,6 +5596,15 @@ struct Operator FLATBUFFERS_FINAL_CLASS : private flatbuffers::Table { const ShapeOptions *builtin_options_as_ShapeOptions() const { return builtin_options_type() == BuiltinOptions_ShapeOptions ? static_cast(builtin_options()) : nullptr; } + const PowOptions *builtin_options_as_PowOptions() const { + return builtin_options_type() == BuiltinOptions_PowOptions ? static_cast(builtin_options()) : nullptr; + } + const ArgMinOptions *builtin_options_as_ArgMinOptions() const { + return builtin_options_type() == BuiltinOptions_ArgMinOptions ? static_cast(builtin_options()) : nullptr; + } + const FakeQuantOptions *builtin_options_as_FakeQuantOptions() const { + return builtin_options_type() == BuiltinOptions_FakeQuantOptions ? static_cast(builtin_options()) : nullptr; + } const flatbuffers::Vector *custom_options() const { return GetPointer *>(VT_CUSTOM_OPTIONS); } @@ -5556,6 +5856,18 @@ template<> inline const ShapeOptions *Operator::builtin_options_as return builtin_options_as_ShapeOptions(); } +template<> inline const PowOptions *Operator::builtin_options_as() const { + return builtin_options_as_PowOptions(); +} + +template<> inline const ArgMinOptions *Operator::builtin_options_as() const { + return builtin_options_as_ArgMinOptions(); +} + +template<> inline const FakeQuantOptions *Operator::builtin_options_as() const { + return builtin_options_as_FakeQuantOptions(); +} + struct OperatorBuilder { flatbuffers::FlatBufferBuilder &fbb_; flatbuffers::uoffset_t start_; @@ -6335,6 +6647,7 @@ inline void FullyConnectedOptions::UnPackTo(FullyConnectedOptionsT *_o, const fl (void)_o; (void)_resolver; { auto _e = fused_activation_function(); _o->fused_activation_function = _e; }; + { auto _e = weights_format(); _o->weights_format = _e; }; } inline flatbuffers::Offset FullyConnectedOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const FullyConnectedOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { @@ -6346,9 +6659,11 @@ inline flatbuffers::Offset CreateFullyConnectedOptions(fl (void)_o; struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const FullyConnectedOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; auto _fused_activation_function = _o->fused_activation_function; + auto _weights_format = _o->weights_format; return tflite::CreateFullyConnectedOptions( _fbb, - _fused_activation_function); + _fused_activation_function, + _weights_format); } inline SoftmaxOptionsT *SoftmaxOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { @@ -7218,6 +7533,32 @@ inline flatbuffers::Offset CreateArgMaxOptions(flatbuffers::FlatB _output_type); } +inline ArgMinOptionsT *ArgMinOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new ArgMinOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void ArgMinOptions::UnPackTo(ArgMinOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; + { auto _e = output_type(); _o->output_type = _e; }; +} + +inline flatbuffers::Offset ArgMinOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const ArgMinOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateArgMinOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateArgMinOptions(flatbuffers::FlatBufferBuilder &_fbb, const ArgMinOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const ArgMinOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + auto _output_type = _o->output_type; + return tflite::CreateArgMinOptions( + _fbb, + _output_type); +} + inline GreaterOptionsT *GreaterOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { auto _o = new GreaterOptionsT(); UnPackTo(_o, _resolver); @@ -7532,6 +7873,64 @@ inline flatbuffers::Offset CreateShapeOptions(flatbuffers::FlatBuf _out_type); } +inline PowOptionsT *PowOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new PowOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void PowOptions::UnPackTo(PowOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; +} + +inline flatbuffers::Offset PowOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const PowOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreatePowOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreatePowOptions(flatbuffers::FlatBufferBuilder &_fbb, const PowOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const PowOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + return tflite::CreatePowOptions( + _fbb); +} + +inline FakeQuantOptionsT *FakeQuantOptions::UnPack(const flatbuffers::resolver_function_t *_resolver) const { + auto _o = new FakeQuantOptionsT(); + UnPackTo(_o, _resolver); + return _o; +} + +inline void FakeQuantOptions::UnPackTo(FakeQuantOptionsT *_o, const flatbuffers::resolver_function_t *_resolver) const { + (void)_o; + (void)_resolver; + { auto _e = min(); _o->min = _e; }; + { auto _e = max(); _o->max = _e; }; + { auto _e = num_bits(); _o->num_bits = _e; }; + { auto _e = narrow_range(); _o->narrow_range = _e; }; +} + +inline flatbuffers::Offset FakeQuantOptions::Pack(flatbuffers::FlatBufferBuilder &_fbb, const FakeQuantOptionsT* _o, const flatbuffers::rehasher_function_t *_rehasher) { + return CreateFakeQuantOptions(_fbb, _o, _rehasher); +} + +inline flatbuffers::Offset CreateFakeQuantOptions(flatbuffers::FlatBufferBuilder &_fbb, const FakeQuantOptionsT *_o, const flatbuffers::rehasher_function_t *_rehasher) { + (void)_rehasher; + (void)_o; + struct _VectorArgs { flatbuffers::FlatBufferBuilder *__fbb; const FakeQuantOptionsT* __o; const flatbuffers::rehasher_function_t *__rehasher; } _va = { &_fbb, _o, _rehasher}; (void)_va; + auto _min = _o->min; + auto _max = _o->max; + auto _num_bits = _o->num_bits; + auto _narrow_range = _o->narrow_range; + return tflite::CreateFakeQuantOptions( + _fbb, + _min, + _max, + _num_bits, + _narrow_range); +} + inline OperatorCodeT *OperatorCode::UnPack(const flatbuffers::resolver_function_t *_resolver) const { auto _o = new OperatorCodeT(); UnPackTo(_o, _resolver); @@ -7941,6 +8340,18 @@ inline bool VerifyBuiltinOptions(flatbuffers::Verifier &verifier, const void *ob auto ptr = reinterpret_cast(obj); return verifier.VerifyTable(ptr); } + case BuiltinOptions_PowOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_ArgMinOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } + case BuiltinOptions_FakeQuantOptions: { + auto ptr = reinterpret_cast(obj); + return verifier.VerifyTable(ptr); + } default: return false; } } @@ -8179,6 +8590,18 @@ inline void *BuiltinOptionsUnion::UnPack(const void *obj, BuiltinOptions type, c auto ptr = reinterpret_cast(obj); return ptr->UnPack(resolver); } + case BuiltinOptions_PowOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_ArgMinOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } + case BuiltinOptions_FakeQuantOptions: { + auto ptr = reinterpret_cast(obj); + return ptr->UnPack(resolver); + } default: return nullptr; } } @@ -8405,6 +8828,18 @@ inline flatbuffers::Offset BuiltinOptionsUnion::Pack(flatbuffers::FlatBuff auto ptr = reinterpret_cast(value); return CreateShapeOptions(_fbb, ptr, _rehasher).Union(); } + case BuiltinOptions_PowOptions: { + auto ptr = reinterpret_cast(value); + return CreatePowOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_ArgMinOptions: { + auto ptr = reinterpret_cast(value); + return CreateArgMinOptions(_fbb, ptr, _rehasher).Union(); + } + case BuiltinOptions_FakeQuantOptions: { + auto ptr = reinterpret_cast(value); + return CreateFakeQuantOptions(_fbb, ptr, _rehasher).Union(); + } default: return 0; } } @@ -8631,6 +9066,18 @@ inline BuiltinOptionsUnion::BuiltinOptionsUnion(const BuiltinOptionsUnion &u) FL value = new ShapeOptionsT(*reinterpret_cast(u.value)); break; } + case BuiltinOptions_PowOptions: { + value = new PowOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_ArgMinOptions: { + value = new ArgMinOptionsT(*reinterpret_cast(u.value)); + break; + } + case BuiltinOptions_FakeQuantOptions: { + value = new FakeQuantOptionsT(*reinterpret_cast(u.value)); + break; + } default: break; } @@ -8913,6 +9360,21 @@ inline void BuiltinOptionsUnion::Reset() { delete ptr; break; } + case BuiltinOptions_PowOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_ArgMinOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } + case BuiltinOptions_FakeQuantOptions: { + auto ptr = reinterpret_cast(value); + delete ptr; + break; + } default: break; } value = nullptr; diff --git a/tensorflow/contrib/lite/testing/BUILD b/tensorflow/contrib/lite/testing/BUILD index b823c97f38e7660652aa0ce3538b11de59dc9aea..789bc695f8e9f8721edeb3b3a3f2af59b36adeed 100644 --- a/tensorflow/contrib/lite/testing/BUILD +++ b/tensorflow/contrib/lite/testing/BUILD @@ -172,6 +172,7 @@ cc_test( data = ["//tensorflow/contrib/lite:testdata/multi_add.bin"], tags = [ "tflite_not_portable_android", + "tflite_not_portable_ios", ], deps = [ ":tflite_driver", diff --git a/tensorflow/contrib/lite/testing/generate_examples.py b/tensorflow/contrib/lite/testing/generate_examples.py index c4d2d7ca52ad9b3652682e3d5127d11246b14005..1093bd2cbe5096e7043e6757a98d5dc909705420 100644 --- a/tensorflow/contrib/lite/testing/generate_examples.py +++ b/tensorflow/contrib/lite/testing/generate_examples.py @@ -94,8 +94,8 @@ KNOWN_BUGS = { r"sigmoid.*input_shape=\[\]": "67645668", # Concat doesn't work with a single input tensor r"concat.*num_tensors=1": "67378344", - # Transposition in MatMul is not supported. - r"fully_connected.*transpose_.=True": "67586970", + # Transposition in MatMul is not fully supported. + "fully_connected.*transpose_a=True": "67586970", # Softmax graphs are too complex. r"softmax.*dim=0": "67749831", # BatchToSpaceND only supports 4D tensors. @@ -678,6 +678,55 @@ def make_relu6_tests(zip_path): make_zip_of_tests(zip_path, test_parameters, build_graph, build_inputs) +def make_prelu_tests(zip_path): + """Make a set of tests to do PReLU.""" + + test_parameters = [{ + # The canonical case for image processing is having a 4D `input` (NHWC) + # and `shared_axes`=[1, 2], so the alpha parameter is per channel. + "input_shape": [[1, 10, 10, 3], [3, 3, 3, 3]], + "shared_axes": [[1, 2], [1]], + }] + + def build_graph(parameters): + """Build the graph for the test case.""" + + input_tensor = tf.placeholder( + dtype=tf.float32, name="input", shape=parameters["input_shape"]) + prelu = tf.keras.layers.PReLU(shared_axes=parameters["shared_axes"]) + out = prelu(input_tensor) + return [input_tensor], [out] + + def build_inputs(parameters, sess, inputs, outputs): + """Build the inputs for the test case.""" + + input_shape = parameters["input_shape"] + input_values = create_tensor_data( + np.float32, input_shape, min_value=-10, max_value=10) + shared_axes = parameters["shared_axes"] + + alpha_shape = [] + for dim in range(1, len(input_shape)): + alpha_shape.append(1 if dim in shared_axes else input_shape[dim]) + + alpha_values = create_tensor_data(np.float32, alpha_shape) + + # There should be only 1 trainable variable tensor. + variables = tf.all_variables() + assert len(variables) == 1 + sess.run(variables[0].assign(alpha_values)) + + return [input_values], sess.run( + outputs, feed_dict=dict(zip(inputs, [input_values]))) + + make_zip_of_tests( + zip_path, + test_parameters, + build_graph, + build_inputs, + use_frozen_graph=True) + + # This function tests various TensorFLow functions that generates Const op, # including `tf.ones`, `tf.zeros` and random functions. def make_constant_tests(zip_path): @@ -705,7 +754,7 @@ def make_constant_tests(zip_path): def make_binary_op_tests(zip_path, binary_operator): - """Make a set of tests to do add with and without broadcast.""" + """Make a set of tests to do binary ops with and without broadcast.""" # These parameters are split because we don't support broadcasting. test_parameters = [{ @@ -990,6 +1039,10 @@ def make_mul_tests(zip_path): make_binary_op_tests(zip_path, tf.multiply) +def make_pow_tests(zip_path): + make_binary_op_tests(zip_path, tf.pow) + + def make_gather_tests(zip_path): """Make a set of tests to do gather.""" @@ -1321,6 +1374,12 @@ def make_fully_connected_tests(zip_path): "transpose_a": [False], "transpose_b": [False], "constant_filter": [True, False], + }, { + "shape1": [[40, 37]], + "shape2": [[40, 37]], + "transpose_a": [False], + "transpose_b": [True], + "constant_filter": [True, False], }] def build_graph(parameters): @@ -2165,7 +2224,7 @@ def make_topk_tests(zip_path): make_zip_of_tests(zip_path, test_parameters, build_graph, build_inputs) -def make_arg_max_tests(zip_path): +def make_arg_min_max_tests(zip_path): """Make a set of tests to do arg_max.""" test_parameters = [{ @@ -2173,6 +2232,7 @@ def make_arg_max_tests(zip_path): "input_shape": [[1, 1, 1, 3], [2, 3, 4, 5], [2, 3, 3], [5, 5], [10]], "output_type": [tf.int32, tf.int64], "axis_is_last_dim": [True, False], + "is_arg_max": [True], }] def build_graph(parameters): @@ -2185,7 +2245,10 @@ def make_arg_max_tests(zip_path): axis = len(parameters["input_shape"]) - 1 else: axis = random.randint(0, max(len(parameters["input_shape"]) - 2, 0)) - out = tf.arg_max(input_value, axis, output_type=parameters["output_type"]) + if parameters["is_arg_max"]: + out = tf.arg_max(input_value, axis, output_type=parameters["output_type"]) + else: + out = tf.arg_min(input_value, axis, output_type=parameters["output_type"]) return [input_value], [out] def build_inputs(parameters, sess, inputs, outputs): diff --git a/tensorflow/contrib/lite/testing/generate_testspec.cc b/tensorflow/contrib/lite/testing/generate_testspec.cc index c0c861ff6da2fc144b9303dfdd48f19794cebeca..c1092e4d25567f0374e3cd5a27bde32419d3db19 100644 --- a/tensorflow/contrib/lite/testing/generate_testspec.cc +++ b/tensorflow/contrib/lite/testing/generate_testspec.cc @@ -25,7 +25,7 @@ namespace testing { template void GenerateCsv(const std::vector& shape, float min, float max, string* out) { - auto random_float = [](int min, int max) { + auto random_float = [](float min, float max) { static unsigned int seed; return min + (max - min) * static_cast(rand_r(&seed)) / RAND_MAX; }; @@ -37,16 +37,10 @@ void GenerateCsv(const std::vector& shape, float min, float max, *out = Join(data.data(), data.size(), ","); } -bool GenerateTestSpecFromTensorflowModel( - std::iostream& stream, const string& tensorflow_model_path, - const string& tflite_model_path, const std::vector& input_layer, +std::vector GenerateInputValues( + const std::vector& input_layer, const std::vector& input_layer_type, - const std::vector& input_layer_shape, - const std::vector& output_layer) { - CHECK_EQ(input_layer.size(), input_layer_type.size()); - CHECK_EQ(input_layer.size(), input_layer_shape.size()); - - // Generate inputs. + const std::vector& input_layer_shape) { std::vector input_values; input_values.resize(input_layer.size()); for (int i = 0; i < input_layer.size(); i++) { @@ -73,9 +67,22 @@ bool GenerateTestSpecFromTensorflowModel( default: fprintf(stderr, "Unsupported type %d (%s) when generating testspec.\n", type, input_layer_type[i].c_str()); - return false; + input_values.clear(); + return input_values; } } + return input_values; +} + +bool GenerateTestSpecFromTensorflowModel( + std::iostream& stream, const string& tensorflow_model_path, + const string& tflite_model_path, int num_invocations, + const std::vector& input_layer, + const std::vector& input_layer_type, + const std::vector& input_layer_shape, + const std::vector& output_layer) { + CHECK_EQ(input_layer.size(), input_layer_type.size()); + CHECK_EQ(input_layer.size(), input_layer_shape.size()); // Invoke tensorflow model. TfDriver runner(input_layer, input_layer_type, input_layer_shape, @@ -91,39 +98,51 @@ bool GenerateTestSpecFromTensorflowModel( return false; } - for (int i = 0; i < input_values.size(); i++) { - runner.SetInput(i, input_values[i]); - if (!runner.IsValid()) { - cerr << runner.GetErrorMessage() << endl; - return false; - } - } - - runner.Invoke(); - if (!runner.IsValid()) { - cerr << runner.GetErrorMessage() << endl; - return false; - } - - // Write test spec. + // Write first part of test spec, defining model and input shapes. stream << "load_model: " << tflite_model_path << "\n"; stream << "reshape {\n"; for (const auto& shape : input_layer_shape) { stream << " input: \"" << shape << "\"\n"; } stream << "}\n"; - stream << "invoke {\n"; - for (const auto& value : input_values) { - stream << " input: \"" << value << "\"\n"; - } - for (int i = 0; i < output_layer.size(); i++) { - stream << " output: \"" << runner.ReadOutput(i) << "\"\n"; + + // Generate inputs. + for (int i = 0; i < num_invocations; ++i) { + // Note that the input values are random, so each invocation will have a + // different set. + std::vector input_values = + GenerateInputValues(input_layer, input_layer_type, input_layer_shape); + if (input_values.empty()) return false; + + // Run TensorFlow. + for (int j = 0; j < input_values.size(); j++) { + runner.SetInput(j, input_values[j]); + if (!runner.IsValid()) { + cerr << runner.GetErrorMessage() << endl; + return false; + } + } + + runner.Invoke(); if (!runner.IsValid()) { cerr << runner.GetErrorMessage() << endl; return false; } + + // Write second part of test spec, with inputs and outputs. + stream << "invoke {\n"; + for (const auto& value : input_values) { + stream << " input: \"" << value << "\"\n"; + } + for (int j = 0; j < output_layer.size(); j++) { + stream << " output: \"" << runner.ReadOutput(j) << "\"\n"; + if (!runner.IsValid()) { + cerr << runner.GetErrorMessage() << endl; + return false; + } + } + stream << "}\n"; } - stream << "}\n"; return true; } diff --git a/tensorflow/contrib/lite/testing/generate_testspec.h b/tensorflow/contrib/lite/testing/generate_testspec.h index 6e31a853c3f7f82a89126ff83af784ffd418741a..bfaf5e7ec89bbdd85b68a7dc45d7686e143e5d3d 100644 --- a/tensorflow/contrib/lite/testing/generate_testspec.h +++ b/tensorflow/contrib/lite/testing/generate_testspec.h @@ -30,13 +30,15 @@ namespace testing { // stream: mutable iostream that contains the contents of test spec. // tensorflow_model_path: path to TensorFlow model. // tflite_model_path: path to tflite_model_path that the test spec runs +// num_invocations: how many pairs of inputs and outputs will be generated. // against. input_layer: names of input tensors. Example: input1 // input_layer_type: datatypes of input tensors. Example: float // input_layer_shape: shapes of input tensors, separated by comma. example: // 1,3,4 output_layer: names of output tensors. Example: output bool GenerateTestSpecFromTensorflowModel( std::iostream& stream, const string& tensorflow_model_path, - const string& tflite_model_path, const std::vector& input_layer, + const string& tflite_model_path, int num_invocations, + const std::vector& input_layer, const std::vector& input_layer_type, const std::vector& input_layer_shape, const std::vector& output_layer); diff --git a/tensorflow/contrib/lite/testing/generated_examples_zip_test.cc b/tensorflow/contrib/lite/testing/generated_examples_zip_test.cc index 8a59d756f8dbbcefc930b5285c1ced8ce6b08845..58f6bb538268515d2f1b1bc8173b366faf11157b 100644 --- a/tensorflow/contrib/lite/testing/generated_examples_zip_test.cc +++ b/tensorflow/contrib/lite/testing/generated_examples_zip_test.cc @@ -42,6 +42,7 @@ string* FLAGS_unzip_binary_path = new string("/usr/bin/unzip"); string* FLAGS_unzip_binary_path = new string("/system/bin/unzip"); #endif bool FLAGS_use_nnapi = false; +bool FLAGS_ignore_unsupported_nnapi = false; } // namespace // TensorFlow system environment for file system called. @@ -52,10 +53,6 @@ tensorflow::Env* env = tensorflow::Env::Default(); // Key is a substring of the test name and value is a bug number. // TODO(ahentz): make sure we clean this list up frequently. std::map kBrokenTests = { - // Add only supports float32. (and "constant" tests use Add) - {R"(^\/add_a.*int32)", "68808744"}, - {R"(^\/constant.*int32)", "68808744"}, - {R"(^\/mul.*int32)", "68808744"}, {R"(^\/div.*int32)", "68808744"}, {R"(^\/sub.*int32)", "68808744"}, @@ -99,11 +96,12 @@ std::map kBrokenTests = { {R"(^\/gather.*axis=1)", "76910444"}, // No support for arbitrary dimensions in ArgMax. - {R"(^\/arg_max.*axis_is_last_dim=False.*input_shape=\[.,.,.,.\])", + {R"(^\/arg_min_max.*axis_is_last_dim=False.*input_shape=\[.,.,.,.\])", "77546240"}, - {R"(^\/arg_max.*axis_is_last_dim=False.*input_shape=\[.,.,.\])", + {R"(^\/arg_min_max.*axis_is_last_dim=False.*input_shape=\[.,.,.\])", + "77546240"}, + {R"(^\/arg_min_max.*axis_is_last_dim=False.*input_shape=\[.,.\])", "77546240"}, - {R"(^\/arg_max.*axis_is_last_dim=False.*input_shape=\[.,.\])", "77546240"}, }; // Allows test data to be unzipped into a temporary directory and makes @@ -228,16 +226,21 @@ TEST_P(OpsTest, RunZipTests) { } bool result = tflite::testing::ParseAndRunTests(&tflite_stream, &test_driver); + string message = test_driver.GetErrorMessage(); if (bug_number.empty()) { - EXPECT_TRUE(result) << test_driver.GetErrorMessage(); + if (FLAGS_use_nnapi && FLAGS_ignore_unsupported_nnapi && !result) { + EXPECT_EQ(message, string("Failed to invoke interpreter")) << message; + } else { + EXPECT_TRUE(result) << message; + } } else { if (FLAGS_ignore_known_bugs) { EXPECT_FALSE(result) << "Test was expected to fail but is now passing; " "you can mark http://b/" << bug_number << " as fixed! Yay!"; } else { - EXPECT_TRUE(result) << test_driver.GetErrorMessage() - << ": Possibly due to http://b/" << bug_number; + EXPECT_TRUE(result) << message << ": Possibly due to http://b/" + << bug_number; } } } @@ -280,8 +283,11 @@ int main(int argc, char** argv) { tflite::testing::FLAGS_unzip_binary_path, "Required: Location of a suitable unzip binary."), tensorflow::Flag("use_nnapi", &tflite::testing::FLAGS_use_nnapi, - "Whether to enable the NNAPI delegate")}; - + "Whether to enable the NNAPI delegate"), + tensorflow::Flag("ignore_unsupported_nnapi", + &tflite::testing::FLAGS_ignore_unsupported_nnapi, + "Don't fail tests just because delegation to NNAPI " + "is not possible")}; bool success = tensorflow::Flags::Parse(&argc, argv, flags); if (!success || (argc == 2 && !strcmp(argv[1], "--helpfull"))) { fprintf(stderr, "%s", tensorflow::Flags::Usage(argv[0], flags).c_str()); diff --git a/tensorflow/contrib/lite/testing/tflite_diff_example_test.cc b/tensorflow/contrib/lite/testing/tflite_diff_example_test.cc index 5afa0f800cdaa8bf70a11cb6e2ac64ace8138e79..f2c49fe389763110279b3dd1e4f13b1522de0460 100644 --- a/tensorflow/contrib/lite/testing/tflite_diff_example_test.cc +++ b/tensorflow/contrib/lite/testing/tflite_diff_example_test.cc @@ -20,12 +20,29 @@ int main(int argc, char** argv) { ::tflite::testing::DiffOptions options = ::tflite::testing::ParseTfliteDiffFlags(&argc, argv); if (options.tensorflow_model.empty()) return 1; + int failure_count = 0; - for (int i = 0; i < 100; i++) { - if (!tflite::testing::RunDiffTest(options)) { + for (int i = 0; i < options.num_runs_per_pass; i++) { + if (!tflite::testing::RunDiffTest(options, /*num_invocations=*/1)) { ++failure_count; } } - fprintf(stderr, "Num errors: %d\n", failure_count); + int failures_in_first_pass = failure_count; + + if (failure_count == 0) { + // Let's try again with num_invocations > 1 to make sure we can do multiple + // invocations without resetting the interpreter. + for (int i = 0; i < options.num_runs_per_pass; i++) { + if (!tflite::testing::RunDiffTest(options, /*num_invocations=*/2)) { + ++failure_count; + } + } + } + + fprintf(stderr, "Num errors in single-inference pass: %d\n", + failures_in_first_pass); + fprintf(stderr, "Num errors in multi-inference pass : %d\n", + failure_count - failures_in_first_pass); + return failure_count != 0 ? 1 : 0; } diff --git a/tensorflow/contrib/lite/testing/tflite_diff_flags.h b/tensorflow/contrib/lite/testing/tflite_diff_flags.h index 706108ed73bb3fd9bd784cffffe322d6981433e6..7a57e8d3fba29cd106eb038992bb5ed12bb457ae 100644 --- a/tensorflow/contrib/lite/testing/tflite_diff_flags.h +++ b/tensorflow/contrib/lite/testing/tflite_diff_flags.h @@ -30,6 +30,7 @@ DiffOptions ParseTfliteDiffFlags(int* argc, char** argv) { string input_layer_type; string input_layer_shape; string output_layer; + int32_t num_runs_per_pass = 100; } values; std::vector flags = { @@ -49,6 +50,8 @@ DiffOptions ParseTfliteDiffFlags(int* argc, char** argv) { tensorflow::Flag("output_layer", &values.output_layer, "Names of output tensors, separated by comma. Example " "output_1,output_2"), + tensorflow::Flag("num_runs_per_pass", &values.num_runs_per_pass, + "Number of full runs in each pass."), }; bool no_inputs = *argc == 1; @@ -63,7 +66,8 @@ DiffOptions ParseTfliteDiffFlags(int* argc, char** argv) { Split(values.input_layer, ","), Split(values.input_layer_type, ","), Split(values.input_layer_shape, ":"), - Split(values.output_layer, ",")}; + Split(values.output_layer, ","), + values.num_runs_per_pass}; } } // namespace testing diff --git a/tensorflow/contrib/lite/testing/tflite_diff_util.cc b/tensorflow/contrib/lite/testing/tflite_diff_util.cc index f601d3752ddb5df9f2b5ac73d9bc303efaade4a5..19f34c0a51e442804bf2824adc3a1d8bde1eb4b0 100644 --- a/tensorflow/contrib/lite/testing/tflite_diff_util.cc +++ b/tensorflow/contrib/lite/testing/tflite_diff_util.cc @@ -25,13 +25,14 @@ limitations under the License. namespace tflite { namespace testing { -bool RunDiffTest(const DiffOptions& options) { +bool RunDiffTest(const DiffOptions& options, int num_invocations) { std::stringstream tflite_stream; if (!GenerateTestSpecFromTensorflowModel( tflite_stream, options.tensorflow_model, options.tflite_model, - options.input_layer, options.input_layer_type, - options.input_layer_shape, options.output_layer)) + num_invocations, options.input_layer, options.input_layer_type, + options.input_layer_shape, options.output_layer)) { return false; + } TfLiteDriver tflite_driver(/*use_nnapi=*/true); tflite_driver.LoadModel(options.tflite_model); return tflite::testing::ParseAndRunTests(&tflite_stream, &tflite_driver); diff --git a/tensorflow/contrib/lite/testing/tflite_diff_util.h b/tensorflow/contrib/lite/testing/tflite_diff_util.h index 326fa6c3e28000dee9b6eb9cc5b3a6c5c87e28d0..4ab2f230fdcdfe4616ab1706aa41f0e806665f66 100644 --- a/tensorflow/contrib/lite/testing/tflite_diff_util.h +++ b/tensorflow/contrib/lite/testing/tflite_diff_util.h @@ -40,10 +40,14 @@ struct DiffOptions { // Names of output tensors. // Example output_1,output_2 std::vector output_layer; + // Number of full runs (from building interpreter to checking outputs) in + // each of the passes. The first pass has a single inference, while the + // second pass does multiple inferences back to back. + int num_runs_per_pass; }; // Run a single TensorFLow Lite diff test with a given options. -bool RunDiffTest(const DiffOptions& options); +bool RunDiffTest(const DiffOptions& options, int num_invocations); } // namespace testing } // namespace tflite diff --git a/tensorflow/contrib/lite/toco/BUILD b/tensorflow/contrib/lite/toco/BUILD index be102faa4c6e95668b68a92efda19a4f938ae178..2c469c0e75d4a7e637b947336f5aaf9efdd53103 100644 --- a/tensorflow/contrib/lite/toco/BUILD +++ b/tensorflow/contrib/lite/toco/BUILD @@ -143,7 +143,6 @@ cc_library( ":toco_graphviz_dump_options", ":toco_port", ":types_proto_cc", - "//tensorflow/cc/saved_model:tag_constants", "//tensorflow/core:framework_internal", "//tensorflow/core:lib", "@com_google_absl//absl/strings", @@ -169,41 +168,6 @@ cc_library( ], ) -cc_library( - name = "toco_saved_model", - srcs = [ - "toco_saved_model.cc", - ], - hdrs = [ - "toco_saved_model.h", - ], - visibility = ["//visibility:public"], - deps = [ - ":model_cmdline_flags", - ":model_flags_proto_cc", - ":toco_flags_proto_cc", - ":types_proto_cc", - "//tensorflow/cc/tools:freeze_saved_model", - "//tensorflow/core:protos_all_cc", - "@com_google_absl//absl/strings", - ], -) - -tf_cc_test( - name = "toco_saved_model_test", - srcs = ["toco_saved_model_test.cc"], - deps = [ - ":model_cmdline_flags", - ":toco_cmdline_flags", - ":toco_saved_model", - "//tensorflow/cc:cc_ops", - "//tensorflow/cc:scope", - "//tensorflow/core:test", - "@com_google_absl//absl/strings", - "@com_google_googletest//:gtest_main", - ], -) - cc_library( name = "graph_transformations", srcs = [ @@ -238,6 +202,7 @@ cc_library( "graph_transformations/lstm_utils.cc", "graph_transformations/make_initial_dequantize_operator.cc", "graph_transformations/merge_reshape_into_preceding_transpose.cc", + "graph_transformations/move_binary_operator_before_reshape.cc", "graph_transformations/propagate_activation_function_into_constants.cc", "graph_transformations/propagate_array_data_types.cc", "graph_transformations/propagate_default_min_max.cc", @@ -247,7 +212,7 @@ cc_library( "graph_transformations/quantization_util.h", "graph_transformations/quantize.cc", "graph_transformations/quantize_weights.cc", - "graph_transformations/read_fake_quant_min_max.cc", + "graph_transformations/read_array_minmax_and_narrow_range_from_fake_quant.cc", "graph_transformations/remove_final_dequantize_op.cc", "graph_transformations/remove_tensorflow_assert.cc", "graph_transformations/remove_tensorflow_identity.cc", @@ -280,6 +245,7 @@ cc_library( "graph_transformations/resolve_constant_strided_slice.cc", "graph_transformations/resolve_constant_transpose.cc", "graph_transformations/resolve_constant_unary.cc", + "graph_transformations/resolve_fake_quant_args_from_vars.cc", "graph_transformations/resolve_mean_attributes.cc", "graph_transformations/resolve_multiply_by_zero.cc", "graph_transformations/resolve_pad_attributes.cc", @@ -431,7 +397,6 @@ tf_cc_binary( ":toco_cmdline_flags", ":toco_flags_proto_cc", ":toco_port", - ":toco_saved_model", ":toco_tooling", ":types_proto_cc", "//tensorflow/core:lib", diff --git a/tensorflow/contrib/lite/toco/README.md b/tensorflow/contrib/lite/toco/README.md index ee83c7a6e3253d02fb1a2c791fc22428473c1832..2db6a627ab59604a99cafe3b38df08b70092d989 100644 --- a/tensorflow/contrib/lite/toco/README.md +++ b/tensorflow/contrib/lite/toco/README.md @@ -17,11 +17,12 @@ Usage information is given in these documents: Once an application developer has a trained TensorFlow model, TOCO will accept that model and generate a TensorFlow Lite [FlatBuffer](https://google.github.io/flatbuffers/) file. TOCO currently supports -[SavedModels](https://www.tensorflow.org/guide/saved_model#using_savedmodel_with_estimators) -and frozen graphs (models generated via -[freeze_graph.py](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/python/tools/freeze_graph.py)). -The TensorFlow Lite FlatBuffer file can be shipped to client devices, generally -mobile devices, where the TensorFlow Lite interpreter handles them on-device. -This flow is represented in the diagram below. +[SavedModels](https://www.tensorflow.org/guide/saved_model#using_savedmodel_with_estimators), +frozen graphs (models generated via +[freeze_graph.py](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/python/tools/freeze_graph.py)), +and `tf.Keras` model files. The TensorFlow Lite FlatBuffer file can be shipped +to client devices, generally mobile devices, where the TensorFlow Lite +interpreter handles them on-device. This flow is represented in the diagram +below. ![drawing](g3doc/toco_landscape.svg) diff --git a/tensorflow/contrib/lite/toco/args.h b/tensorflow/contrib/lite/toco/args.h index 9f5ca66d050f0ead9b8856c77dba8d9bbd182d10..aef35ad490656c09a7d7314aa033bc985b3af661 100644 --- a/tensorflow/contrib/lite/toco/args.h +++ b/tensorflow/contrib/lite/toco/args.h @@ -21,13 +21,13 @@ limitations under the License. #include #include #include +#include "tensorflow/contrib/lite/toco/toco_port.h" #if defined(PLATFORM_GOOGLE) #include "strings/split.h" +#include "strings/strip.h" #endif #include "absl/strings/numbers.h" #include "absl/strings/str_split.h" -#include "tensorflow/cc/saved_model/tag_constants.h" -#include "tensorflow/contrib/lite/toco/toco_port.h" #include "tensorflow/contrib/lite/toco/toco_types.h" namespace toco { @@ -145,8 +145,10 @@ class Arg final { } string outer_member_copy = outer_member; absl::StripAsciiWhitespace(&outer_member); - if (!TryStripPrefixString(outer_member, "{", &outer_member)) return false; - if (!TryStripSuffixString(outer_member, "}", &outer_member)) return false; + if (!strings::TryStripPrefixString(outer_member, "{", &outer_member)) + return false; + if (!strings::TryStripSuffixString(outer_member, "}", &outer_member)) + return false; const std::vector inner_fields_vector = absl::StrSplit(outer_member, ','); @@ -223,7 +225,7 @@ struct ParsedTocoFlags { Arg output_file; Arg input_format = Arg("TENSORFLOW_GRAPHDEF"); Arg output_format = Arg("TFLITE"); - Arg savedmodel_tagset = Arg(tensorflow::kSavedModelTagServe); + Arg savedmodel_tagset; // TODO(aselle): command_line_flags doesn't support doubles Arg default_ranges_min = Arg(0.); Arg default_ranges_max = Arg(0.); diff --git a/tensorflow/contrib/lite/toco/export_tensorflow.cc b/tensorflow/contrib/lite/toco/export_tensorflow.cc index 6b78f1c05ee777e0d456cd70d07b58ff51271fec..bf9a51a52553f8cc41b61a0c6c660108036cd632 100644 --- a/tensorflow/contrib/lite/toco/export_tensorflow.cc +++ b/tensorflow/contrib/lite/toco/export_tensorflow.cc @@ -145,7 +145,7 @@ void ConvertFloatTensorConst(const string& name, const Shape& input_shape, if (HasAlreadyExportedConst(name, *tensorflow_graph)) { return; } - auto* const_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* const_op = tensorflow_graph->add_node(); const_op->set_op("Const"); const_op->set_name(name); (*const_op->mutable_attr())["dtype"].set_type(DT_FLOAT); @@ -162,7 +162,7 @@ void ConvertFloatTensorConst(const string& name, const Shape& input_shape, if (HasAlreadyExportedConst(name, *tensorflow_graph)) { return; } - auto* const_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* const_op = tensorflow_graph->add_node(); const_op->set_op("Const"); const_op->set_name(name); (*const_op->mutable_attr())["dtype"].set_type(DT_FLOAT); @@ -178,7 +178,7 @@ void ConvertFloatTensorConst(const Model& model, const string& name, if (HasAlreadyExportedConst(name, *tensorflow_graph)) { return; } - auto* const_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* const_op = tensorflow_graph->add_node(); const_op->set_op("Const"); const_op->set_name(name); (*const_op->mutable_attr())["dtype"].set_type(DT_FLOAT); @@ -199,7 +199,7 @@ void ConvertFloatTensorConst(const Model& model, const string& name, if (HasAlreadyExportedConst(name, *tensorflow_graph)) { return; } - auto* const_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* const_op = tensorflow_graph->add_node(); const_op->set_op("Const"); const_op->set_name(name); (*const_op->mutable_attr())["dtype"].set_type(DT_FLOAT); @@ -222,7 +222,7 @@ void ConvertIntTensorConst(const Model& model, const string& name, } CHECK(model.HasArray(name)); const auto& array = model.GetArray(name); - auto* const_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* const_op = tensorflow_graph->add_node(); const_op->set_op("Const"); const_op->set_name(name); (*const_op->mutable_attr())["dtype"].set_type(DT_INT32); @@ -245,7 +245,7 @@ void CreateIntTensorConst(const string& name, const std::vector& data, if (HasAlreadyExportedConst(name, *tensorflow_graph)) { return; } - auto* const_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* const_op = tensorflow_graph->add_node(); const_op->set_op("Const"); const_op->set_name(name); (*const_op->mutable_attr())["dtype"].set_type(DT_INT32); @@ -268,7 +268,7 @@ void CreateMatrixShapeTensorConst(const string& name, int rows, int cols, if (HasAlreadyExportedConst(name, *tensorflow_graph)) { return; } - auto* const_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* const_op = tensorflow_graph->add_node(); const_op->set_op("Const"); const_op->set_name(name); (*const_op->mutable_attr())["dtype"].set_type(DT_INT32); @@ -286,7 +286,7 @@ void CreateDummyConcatDimTensorConst(const string& name, int dim, if (HasAlreadyExportedConst(name, *tensorflow_graph)) { return; } - auto* const_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* const_op = tensorflow_graph->add_node(); const_op->set_op("Const"); const_op->set_name(name); (*const_op->mutable_attr())["dtype"].set_type(DT_INT32); @@ -301,7 +301,7 @@ void CreateReshapeShapeTensorConst(const string& name, if (HasAlreadyExportedConst(name, *tensorflow_graph)) { return; } - auto* const_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* const_op = tensorflow_graph->add_node(); const_op->set_op("Const"); const_op->set_name(name); (*const_op->mutable_attr())["dtype"].set_type(DT_INT32); @@ -341,7 +341,7 @@ void ConvertConvOperator(const Model& model, const ConvOperator& src_op, conv_output += "/conv"; } - auto* conv2d_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* conv2d_op = tensorflow_graph->add_node(); conv2d_op->set_op("Conv2D"); conv2d_op->set_name(conv_output); *conv2d_op->add_input() = src_op.inputs[0]; @@ -377,7 +377,7 @@ void ConvertConvOperator(const Model& model, const ConvOperator& src_op, (*conv2d_op->mutable_attr())["padding"].set_s(padding); if (has_bias) { - auto* biasadd_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* biasadd_op = tensorflow_graph->add_node(); biasadd_op->set_op("BiasAdd"); biasadd_op->set_name(src_op.outputs[0]); biasadd_op->add_input(conv_output); @@ -409,7 +409,7 @@ void ConvertDepthwiseConvOperator(const Model& model, conv_output += "/conv"; } - auto* dc2d_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* dc2d_op = tensorflow_graph->add_node(); dc2d_op->set_op("DepthwiseConv2dNative"); dc2d_op->set_name(conv_output); *dc2d_op->add_input() = src_op.inputs[0]; @@ -457,7 +457,7 @@ void ConvertDepthwiseConvOperator(const Model& model, (*dc2d_op->mutable_attr())["padding"].set_s(padding); if (has_bias) { - auto* biasadd_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* biasadd_op = tensorflow_graph->add_node(); biasadd_op->set_op("BiasAdd"); biasadd_op->set_name(src_op.outputs[0]); biasadd_op->add_input(conv_output); @@ -482,7 +482,7 @@ void ConvertDepthwiseConvOperator(const Model& model, void ConvertTransposeConvOperator(const Model& model, const TransposeConvOperator& src_op, GraphDef* tensorflow_graph) { - auto* conv2d_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* conv2d_op = tensorflow_graph->add_node(); conv2d_op->set_op("Conv2DBackpropInput"); conv2d_op->set_name(src_op.outputs[0]); *conv2d_op->add_input() = src_op.inputs[0]; @@ -514,7 +514,7 @@ void ConvertTransposeConvOperator(const Model& model, void ConvertDepthToSpaceOperator(const Model& model, const DepthToSpaceOperator& src_op, GraphDef* tensorflow_graph) { - auto* op = tensorflow_graph->add_node(); + tensorflow::NodeDef* op = tensorflow_graph->add_node(); op->set_op("DepthToSpace"); op->set_name(src_op.outputs[0]); *op->add_input() = src_op.inputs[0]; @@ -525,7 +525,7 @@ void ConvertDepthToSpaceOperator(const Model& model, void ConvertSpaceToDepthOperator(const Model& model, const SpaceToDepthOperator& src_op, GraphDef* tensorflow_graph) { - auto* op = tensorflow_graph->add_node(); + tensorflow::NodeDef* op = tensorflow_graph->add_node(); op->set_op("SpaceToDepth"); op->set_name(src_op.outputs[0]); *op->add_input() = src_op.inputs[0]; @@ -546,7 +546,7 @@ void ConvertFullyConnectedOperator(const Model& model, CHECK_EQ(fc_weights_shape.dimensions_count(), 2); CreateMatrixShapeTensorConst(reshape_shape, fc_weights_shape.dims(1), -1, tensorflow_graph); - auto* reshape_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* reshape_op = tensorflow_graph->add_node(); reshape_op->set_op("Reshape"); reshape_op->set_name(reshape_output); reshape_op->add_input(src_op.inputs[0]); @@ -568,7 +568,7 @@ void ConvertFullyConnectedOperator(const Model& model, const string transpose_perm = AvailableArrayName(model, transpose_output + "/perm"); CreateIntTensorConst(transpose_perm, {1, 0}, {2}, tensorflow_graph); - auto transpose_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* transpose_op = tensorflow_graph->add_node(); transpose_op->set_op("Transpose"); transpose_op->set_name(transpose_output); *transpose_op->add_input() = src_op.inputs[1]; @@ -577,7 +577,7 @@ void ConvertFullyConnectedOperator(const Model& model, GetTensorFlowDataType(model, src_op.inputs[1])); (*transpose_op->mutable_attr())["Tperm"].set_type(DT_INT32); - auto* matmul_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* matmul_op = tensorflow_graph->add_node(); matmul_op->set_op("MatMul"); matmul_op->set_name(matmul_output); *matmul_op->add_input() = reshape_output; @@ -590,7 +590,7 @@ void ConvertFullyConnectedOperator(const Model& model, // Add the bias, if it exists. if (has_bias) { - auto* biasadd_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* biasadd_op = tensorflow_graph->add_node(); biasadd_op->set_op("BiasAdd"); biasadd_op->set_name(src_op.outputs[0]); biasadd_op->add_input(matmul_output); @@ -615,7 +615,7 @@ void ConvertFullyConnectedOperator(const Model& model, void ConvertAddOperator(const Model& model, const AddOperator& src_op, GraphDef* tensorflow_graph) { - auto* add_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* add_op = tensorflow_graph->add_node(); add_op->set_op("Add"); add_op->set_name(src_op.outputs[0]); CHECK_EQ(src_op.inputs.size(), 2); @@ -626,7 +626,7 @@ void ConvertAddOperator(const Model& model, const AddOperator& src_op, void ConvertAddNOperator(const Model& model, const AddNOperator& src_op, GraphDef* tensorflow_graph) { - auto* add_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* add_op = tensorflow_graph->add_node(); add_op->set_op("AddN"); add_op->set_name(src_op.outputs[0]); for (const auto& input : src_op.inputs) { @@ -638,7 +638,7 @@ void ConvertAddNOperator(const Model& model, const AddNOperator& src_op, void ConvertMulOperator(const Model& model, const MulOperator& src_op, GraphDef* tensorflow_graph) { - auto* add_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* add_op = tensorflow_graph->add_node(); add_op->set_op("Mul"); add_op->set_name(src_op.outputs[0]); CHECK_EQ(src_op.inputs.size(), 2); @@ -649,7 +649,7 @@ void ConvertMulOperator(const Model& model, const MulOperator& src_op, void ConvertReluOperator(const ReluOperator& src_op, GraphDef* tensorflow_graph) { - auto* relu_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* relu_op = tensorflow_graph->add_node(); relu_op->set_op("Relu"); relu_op->set_name(src_op.outputs[0]); *relu_op->add_input() = src_op.inputs[0]; @@ -662,7 +662,7 @@ void ConvertRelu1Operator(const Relu1Operator& src_op, const string min_bounds = src_op.outputs[0] + "/min_bounds"; const string max_output = src_op.outputs[0] + "/max_output"; - auto* max_bounds_const_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* max_bounds_const_op = tensorflow_graph->add_node(); max_bounds_const_op->set_op("Const"); max_bounds_const_op->set_name(max_bounds); (*max_bounds_const_op->mutable_attr())["dtype"].set_type(DT_FLOAT); @@ -671,7 +671,7 @@ void ConvertRelu1Operator(const Relu1Operator& src_op, max_bounds_const_op_tensor->set_dtype(DT_FLOAT); max_bounds_const_op_tensor->add_float_val(-1.0f); - auto* min_bounds_const_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* min_bounds_const_op = tensorflow_graph->add_node(); min_bounds_const_op->set_op("Const"); min_bounds_const_op->set_name(min_bounds); (*min_bounds_const_op->mutable_attr())["dtype"].set_type(DT_FLOAT); @@ -680,14 +680,14 @@ void ConvertRelu1Operator(const Relu1Operator& src_op, min_bounds_const_op_tensor->set_dtype(DT_FLOAT); min_bounds_const_op_tensor->add_float_val(1.0f); - auto* max_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* max_op = tensorflow_graph->add_node(); max_op->set_op("Maximum"); max_op->set_name(max_output); *max_op->add_input() = src_op.inputs[0]; *max_op->add_input() = max_bounds; (*max_op->mutable_attr())["T"].set_type(DT_FLOAT); - auto* min_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* min_op = tensorflow_graph->add_node(); min_op->set_op("Minimum"); min_op->set_name(src_op.outputs[0]); *min_op->add_input() = max_output; @@ -697,7 +697,7 @@ void ConvertRelu1Operator(const Relu1Operator& src_op, void ConvertRelu6Operator(const Relu6Operator& src_op, GraphDef* tensorflow_graph) { - auto* relu_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* relu_op = tensorflow_graph->add_node(); relu_op->set_op("Relu6"); relu_op->set_name(src_op.outputs[0]); *relu_op->add_input() = src_op.inputs[0]; @@ -705,7 +705,7 @@ void ConvertRelu6Operator(const Relu6Operator& src_op, } void ConvertLogOperator(const LogOperator& src_op, GraphDef* tensorflow_graph) { - auto* op = tensorflow_graph->add_node(); + tensorflow::NodeDef* op = tensorflow_graph->add_node(); op->set_op("Log"); op->set_name(src_op.outputs[0]); CHECK_EQ(src_op.inputs.size(), 1); @@ -715,7 +715,7 @@ void ConvertLogOperator(const LogOperator& src_op, GraphDef* tensorflow_graph) { void ConvertLogisticOperator(const LogisticOperator& src_op, GraphDef* tensorflow_graph) { - auto* relu_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* relu_op = tensorflow_graph->add_node(); relu_op->set_op("Sigmoid"); relu_op->set_name(src_op.outputs[0]); *relu_op->add_input() = src_op.inputs[0]; @@ -724,7 +724,7 @@ void ConvertLogisticOperator(const LogisticOperator& src_op, void ConvertTanhOperator(const TanhOperator& src_op, GraphDef* tensorflow_graph) { - auto* tanh_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* tanh_op = tensorflow_graph->add_node(); tanh_op->set_op("Tanh"); tanh_op->set_name(src_op.outputs[0]); *tanh_op->add_input() = src_op.inputs[0]; @@ -744,7 +744,7 @@ void ConvertSoftmaxOperator(const Model& model, const SoftmaxOperator& src_op, const string softmax_size = src_op.outputs[0] + "/softmax_insert_size"; softmax_input = reshape_output; - auto* reshape_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* reshape_op = tensorflow_graph->add_node(); reshape_op->set_op("Reshape"); reshape_op->set_name(reshape_output); *reshape_op->add_input() = src_op.inputs[0]; @@ -761,7 +761,7 @@ void ConvertSoftmaxOperator(const Model& model, const SoftmaxOperator& src_op, CreateReshapeShapeTensorConst(softmax_size, shape_data, tensorflow_graph); } - auto* softmax_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* softmax_op = tensorflow_graph->add_node(); softmax_op->set_op("Softmax"); softmax_op->set_name(src_op.outputs[0]); *softmax_op->add_input() = softmax_input; @@ -785,7 +785,7 @@ void ConvertLogSoftmaxOperator(const Model& model, const string softmax_size = src_op.outputs[0] + "/log_softmax_insert_size"; softmax_input = reshape_output; - auto* reshape_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* reshape_op = tensorflow_graph->add_node(); reshape_op->set_op("Reshape"); reshape_op->set_name(reshape_output); *reshape_op->add_input() = src_op.inputs[0]; @@ -802,7 +802,7 @@ void ConvertLogSoftmaxOperator(const Model& model, CreateReshapeShapeTensorConst(softmax_size, shape_data, tensorflow_graph); } - auto* log_softmax_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* log_softmax_op = tensorflow_graph->add_node(); log_softmax_op->set_op("LogSoftmax"); log_softmax_op->set_name(src_op.outputs[0]); *log_softmax_op->add_input() = softmax_input; @@ -817,7 +817,7 @@ void ConvertL2NormalizationOperator(const L2NormalizationOperator& src_op, const string rsqrt_output = src_op.outputs[0] + "/rsqrt"; const string rsqrt_tiled_output = src_op.outputs[0] + "/rsqrt_tiled"; - auto* sum_reduction_indices_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* sum_reduction_indices_op = tensorflow_graph->add_node(); sum_reduction_indices_op->set_op("Const"); sum_reduction_indices_op->set_name(sum_reduction_indices); (*sum_reduction_indices_op->mutable_attr())["dtype"].set_type(DT_INT32); @@ -831,26 +831,26 @@ void ConvertL2NormalizationOperator(const L2NormalizationOperator& src_op, sum_reduction_indices_tensor->add_int_val(0); sum_reduction_indices_tensor->add_int_val(1); - auto* square_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* square_op = tensorflow_graph->add_node(); square_op->set_op("Square"); square_op->set_name(square_output); *square_op->add_input() = src_op.inputs[0]; (*square_op->mutable_attr())["T"].set_type(DT_FLOAT); - auto* sum_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* sum_op = tensorflow_graph->add_node(); sum_op->set_op("Sum"); sum_op->set_name(sum_output); *sum_op->add_input() = square_output; *sum_op->add_input() = sum_reduction_indices; (*sum_op->mutable_attr())["T"].set_type(DT_FLOAT); - auto* rsqrt_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* rsqrt_op = tensorflow_graph->add_node(); rsqrt_op->set_op("Rsqrt"); rsqrt_op->set_name(rsqrt_output); *rsqrt_op->add_input() = sum_output; (*rsqrt_op->mutable_attr())["T"].set_type(DT_FLOAT); - auto* mul_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* mul_op = tensorflow_graph->add_node(); mul_op->set_op("Mul"); mul_op->set_name(src_op.outputs[0]); *mul_op->add_input() = src_op.inputs[0]; @@ -861,7 +861,7 @@ void ConvertL2NormalizationOperator(const L2NormalizationOperator& src_op, void ConvertLocalResponseNormalizationOperator( const LocalResponseNormalizationOperator& src_op, GraphDef* tensorflow_graph) { - auto* lrn_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* lrn_op = tensorflow_graph->add_node(); lrn_op->set_op("LRN"); lrn_op->set_name(src_op.outputs[0]); *lrn_op->add_input() = src_op.inputs[0]; @@ -873,7 +873,7 @@ void ConvertLocalResponseNormalizationOperator( void ConvertFakeQuantOperator(const FakeQuantOperator& src_op, GraphDef* tensorflow_graph) { - auto* fakequant_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* fakequant_op = tensorflow_graph->add_node(); fakequant_op->set_op("FakeQuantWithMinMaxArgs"); fakequant_op->set_name(src_op.outputs[0]); CHECK_EQ(src_op.inputs.size(), 1); @@ -884,11 +884,14 @@ void ConvertFakeQuantOperator(const FakeQuantOperator& src_op, if (src_op.num_bits) { (*fakequant_op->mutable_attr())["num_bits"].set_i(src_op.num_bits); } + if (src_op.narrow_range) { + (*fakequant_op->mutable_attr())["narrow_range"].set_b(src_op.narrow_range); + } } void ConvertMaxPoolOperator(const MaxPoolOperator& src_op, GraphDef* tensorflow_graph) { - auto* maxpool_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* maxpool_op = tensorflow_graph->add_node(); maxpool_op->set_op("MaxPool"); maxpool_op->set_name(src_op.outputs[0]); *maxpool_op->add_input() = src_op.inputs[0]; @@ -916,7 +919,7 @@ void ConvertMaxPoolOperator(const MaxPoolOperator& src_op, void ConvertAveragePoolOperator(const AveragePoolOperator& src_op, GraphDef* tensorflow_graph) { - auto* avgpool_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* avgpool_op = tensorflow_graph->add_node(); avgpool_op->set_op("AvgPool"); avgpool_op->set_name(src_op.outputs[0]); *avgpool_op->add_input() = src_op.inputs[0]; @@ -945,7 +948,7 @@ void ConvertAveragePoolOperator(const AveragePoolOperator& src_op, void ConvertConcatenationOperator(const Model& model, const ConcatenationOperator& src_op, GraphDef* tensorflow_graph) { - auto* dc_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* dc_op = tensorflow_graph->add_node(); dc_op->set_op("ConcatV2"); dc_op->set_name(src_op.outputs[0]); const string dummy_axis = src_op.outputs[0] + "/axis"; @@ -963,7 +966,7 @@ void ConvertConcatenationOperator(const Model& model, void ConvertTensorFlowReshapeOperator(const Model& model, const TensorFlowReshapeOperator& src_op, GraphDef* tensorflow_graph) { - auto* reshape_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* reshape_op = tensorflow_graph->add_node(); reshape_op->set_op("Reshape"); reshape_op->set_name(src_op.outputs[0]); CHECK_EQ(src_op.inputs.size(), 2); @@ -985,7 +988,7 @@ void ConvertL2PoolOperator(const L2PoolOperator& src_op, const string square_output = src_op.outputs[0] + "/square"; const string avgpool_output = src_op.outputs[0] + "/avgpool"; - auto* square_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* square_op = tensorflow_graph->add_node(); square_op->set_op("Square"); square_op->set_name(square_output); *square_op->add_input() = src_op.inputs[0]; @@ -1000,7 +1003,7 @@ void ConvertL2PoolOperator(const L2PoolOperator& src_op, LOG(FATAL) << "Bad padding (only SAME and VALID are supported)"; } - auto* avgpool_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* avgpool_op = tensorflow_graph->add_node(); avgpool_op->set_op("AvgPool"); avgpool_op->set_name(avgpool_output); *avgpool_op->add_input() = square_output; @@ -1018,7 +1021,7 @@ void ConvertL2PoolOperator(const L2PoolOperator& src_op, ksize.mutable_list()->add_i(src_op.kwidth); ksize.mutable_list()->add_i(1); - auto* sqrt_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* sqrt_op = tensorflow_graph->add_node(); sqrt_op->set_op("Sqrt"); sqrt_op->set_name(src_op.outputs[0]); *sqrt_op->add_input() = avgpool_output; @@ -1027,7 +1030,7 @@ void ConvertL2PoolOperator(const L2PoolOperator& src_op, void ConvertSquareOperator(const TensorFlowSquareOperator& src_op, GraphDef* tensorflow_graph) { - auto* square_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* square_op = tensorflow_graph->add_node(); square_op->set_op("Square"); square_op->set_name(src_op.outputs[0]); CHECK_EQ(src_op.inputs.size(), 1); @@ -1037,7 +1040,7 @@ void ConvertSquareOperator(const TensorFlowSquareOperator& src_op, void ConvertSqrtOperator(const TensorFlowSqrtOperator& src_op, GraphDef* tensorflow_graph) { - auto* sqrt_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* sqrt_op = tensorflow_graph->add_node(); sqrt_op->set_op("Sqrt"); sqrt_op->set_name(src_op.outputs[0]); CHECK_EQ(src_op.inputs.size(), 1); @@ -1048,19 +1051,20 @@ void ConvertSqrtOperator(const TensorFlowSqrtOperator& src_op, void ConvertRsqrtOperator(const Model& model, const TensorFlowRsqrtOperator& src_op, GraphDef* tensorflow_graph) { - auto* rsqrt_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* rsqrt_op = tensorflow_graph->add_node(); rsqrt_op->set_op("Rsqrt"); rsqrt_op->set_name(src_op.outputs[0]); CHECK_EQ(src_op.inputs.size(), 1); *rsqrt_op->add_input() = src_op.inputs[0]; - const auto data_type = GetTensorFlowDataType(model, src_op.inputs[0]); + const tensorflow::DataType data_type = + GetTensorFlowDataType(model, src_op.inputs[0]); (*rsqrt_op->mutable_attr())["T"].set_type(data_type); } void ConvertSplitOperator(const Model& model, const TensorFlowSplitOperator& src_op, GraphDef* tensorflow_graph) { - auto* split_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* split_op = tensorflow_graph->add_node(); split_op->set_op("Split"); split_op->set_name(src_op.outputs[0]); for (const auto& input : src_op.inputs) { @@ -1081,7 +1085,7 @@ void ConvertSplitOperator(const Model& model, void ConvertCastOperator(const Model& model, const CastOperator& src_op, GraphDef* tensorflow_graph) { - auto* cast_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* cast_op = tensorflow_graph->add_node(); cast_op->set_op("Cast"); cast_op->set_name(src_op.outputs[0]); CHECK_EQ(src_op.inputs.size(), 1); @@ -1095,7 +1099,7 @@ void ConvertCastOperator(const Model& model, const CastOperator& src_op, void ConvertFloorOperator(const Model& model, const FloorOperator& src_op, GraphDef* tensorflow_graph) { - auto* floor_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* floor_op = tensorflow_graph->add_node(); floor_op->set_op("Floor"); floor_op->set_name(src_op.outputs[0]); CHECK_EQ(src_op.inputs.size(), 1); @@ -1105,7 +1109,7 @@ void ConvertFloorOperator(const Model& model, const FloorOperator& src_op, void ConvertGatherOperator(const Model& model, const GatherOperator& src_op, GraphDef* tensorflow_graph) { - auto* gather_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* gather_op = tensorflow_graph->add_node(); gather_op->set_op("Gather"); gather_op->set_name(src_op.outputs[0]); CHECK_EQ(src_op.inputs.size(), 2); @@ -1113,13 +1117,14 @@ void ConvertGatherOperator(const Model& model, const GatherOperator& src_op, *gather_op->add_input() = src_op.inputs[1]; (*gather_op->mutable_attr())["Tindices"].set_type(DT_INT32); - const auto params_type = GetTensorFlowDataType(model, src_op.inputs[0]); + const tensorflow::DataType params_type = + GetTensorFlowDataType(model, src_op.inputs[0]); (*gather_op->mutable_attr())["Tparams"].set_type(params_type); } void ConvertArgMaxOperator(const Model& model, const ArgMaxOperator& src_op, GraphDef* tensorflow_graph) { - auto* argmax_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* argmax_op = tensorflow_graph->add_node(); argmax_op->set_op("ArgMax"); argmax_op->set_name(src_op.outputs[0]); CHECK_EQ(src_op.inputs.size(), 2); @@ -1133,10 +1138,26 @@ void ConvertArgMaxOperator(const Model& model, const ArgMaxOperator& src_op, GetTensorFlowDataType(model, src_op.outputs[0])); } +void ConvertArgMinOperator(const Model& model, const ArgMinOperator& src_op, + GraphDef* tensorflow_graph) { + tensorflow::NodeDef* argmin_op = tensorflow_graph->add_node(); + argmin_op->set_op("ArgMin"); + argmin_op->set_name(src_op.outputs[0]); + CHECK_EQ(src_op.inputs.size(), 2); + *argmin_op->add_input() = src_op.inputs[0]; + *argmin_op->add_input() = src_op.inputs[1]; + (*argmin_op->mutable_attr())["T"].set_type( + GetTensorFlowDataType(model, src_op.inputs[0])); + (*argmin_op->mutable_attr())["Tidx"].set_type( + GetTensorFlowDataType(model, src_op.inputs[1])); + (*argmin_op->mutable_attr())["output_type"].set_type( + GetTensorFlowDataType(model, src_op.outputs[0])); +} + void ConvertTransposeOperator(const Model& model, const TransposeOperator& src_op, GraphDef* tensorflow_graph) { - auto* transpose_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* transpose_op = tensorflow_graph->add_node(); transpose_op->set_op("Transpose"); transpose_op->set_name(src_op.outputs[0]); CHECK_EQ(src_op.inputs.size(), 2); @@ -1151,7 +1172,7 @@ void ConvertTransposeOperator(const Model& model, void ConvertTensorFlowShapeOperator(const Model& model, const TensorFlowShapeOperator& src_op, GraphDef* tensorflow_graph) { - auto* shape_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* shape_op = tensorflow_graph->add_node(); shape_op->set_op("Shape"); shape_op->set_name(src_op.outputs[0]); CHECK_EQ(src_op.inputs.size(), 1); @@ -1164,7 +1185,7 @@ void ConvertTensorFlowShapeOperator(const Model& model, void ConvertRankOperator(const Model& model, const RankOperator& src_op, GraphDef* tensorflow_graph) { - auto* rank_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* rank_op = tensorflow_graph->add_node(); rank_op->set_op("Rank"); rank_op->set_name(src_op.outputs[0]); CHECK_EQ(src_op.inputs.size(), 1); @@ -1175,7 +1196,7 @@ void ConvertRankOperator(const Model& model, const RankOperator& src_op, void ConvertRangeOperator(const Model& model, const RangeOperator& src_op, GraphDef* tensorflow_graph) { - auto* range_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* range_op = tensorflow_graph->add_node(); range_op->set_op("Range"); range_op->set_name(src_op.outputs[0]); CHECK_EQ(src_op.inputs.size(), 3); @@ -1188,7 +1209,7 @@ void ConvertRangeOperator(const Model& model, const RangeOperator& src_op, void ConvertStackOperator(const Model& model, const StackOperator& src_op, GraphDef* tensorflow_graph) { - auto* stack_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* stack_op = tensorflow_graph->add_node(); stack_op->set_op("Stack"); stack_op->set_name(src_op.outputs[0]); for (const auto& input : src_op.inputs) { @@ -1201,7 +1222,7 @@ void ConvertStackOperator(const Model& model, const StackOperator& src_op, void ConvertFillOperator(const Model& model, const FillOperator& src_op, GraphDef* tensorflow_graph) { - auto* fill_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* fill_op = tensorflow_graph->add_node(); fill_op->set_op("Fill"); fill_op->set_name(src_op.outputs[0]); CHECK_EQ(src_op.inputs.size(), 2); @@ -1215,7 +1236,7 @@ void ConvertFillOperator(const Model& model, const FillOperator& src_op, void ConvertFloorDivOperator(const Model& model, const FloorDivOperator& src_op, GraphDef* tensorflow_graph) { - auto* floor_div_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* floor_div_op = tensorflow_graph->add_node(); floor_div_op->set_op("FloorDiv"); floor_div_op->set_name(src_op.outputs[0]); CHECK_EQ(src_op.inputs.size(), 2); @@ -1228,7 +1249,7 @@ void ConvertFloorDivOperator(const Model& model, const FloorDivOperator& src_op, void ConvertExpandDimsOperator(const Model& model, const ExpandDimsOperator& src_op, GraphDef* tensorflow_graph) { - auto* expand_dims_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* expand_dims_op = tensorflow_graph->add_node(); expand_dims_op->set_op("ExpandDims"); expand_dims_op->set_name(src_op.outputs[0]); CHECK_EQ(src_op.inputs.size(), 2); @@ -1243,7 +1264,7 @@ void ConvertExpandDimsOperator(const Model& model, void ConvertResizeBilinearOperator(const Model& model, const ResizeBilinearOperator& src_op, GraphDef* tensorflow_graph) { - auto* resize_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* resize_op = tensorflow_graph->add_node(); resize_op->set_op("ResizeBilinear"); resize_op->set_name(src_op.outputs[0]); CHECK_EQ(src_op.inputs.size(), 2); @@ -1293,7 +1314,7 @@ void ConvertLstmCellOperator(const Model& model, const LstmCellOperator& src_op, // works the same since the tensor has the same underlying data layout. const string axis_output = concat_output + "/axis"; CreateDummyConcatDimTensorConst(axis_output, axis, tensorflow_graph); - auto* concat_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* concat_op = tensorflow_graph->add_node(); concat_op->set_op("ConcatV2"); concat_op->set_name(concat_output); *concat_op->add_input() = src_op.inputs[LstmCellOperator::DATA_INPUT]; @@ -1321,7 +1342,7 @@ void ConvertLstmCellOperator(const Model& model, const LstmCellOperator& src_op, // Fully connected matrix multiply const string matmul_output = base + "MatMul"; - auto* matmul_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* matmul_op = tensorflow_graph->add_node(); matmul_op->set_op("MatMul"); matmul_op->set_name(matmul_output); *matmul_op->add_input() = concat_output; @@ -1350,7 +1371,7 @@ void ConvertLstmCellOperator(const Model& model, const LstmCellOperator& src_op, // Add biases string biasadd_output = base + "BiasAdd"; - auto* biasadd_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* biasadd_op = tensorflow_graph->add_node(); biasadd_op->set_op("BiasAdd"); biasadd_op->set_name(biasadd_output); biasadd_op->add_input(matmul_output); @@ -1363,7 +1384,7 @@ void ConvertLstmCellOperator(const Model& model, const LstmCellOperator& src_op, // The dimension is the same as the concatenation dimension CreateDummyConcatDimTensorConst(split_dim_output, axis, tensorflow_graph); string split_output = base + "split"; - auto* split_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* split_op = tensorflow_graph->add_node(); split_op->set_op("Split"); split_op->set_name(split_output); *split_op->add_input() = split_dim_output; @@ -1373,21 +1394,21 @@ void ConvertLstmCellOperator(const Model& model, const LstmCellOperator& src_op, // Activation functions and memory computations const string tanh_0_output = base + "Tanh"; - auto* tanh_0_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* tanh_0_op = tensorflow_graph->add_node(); tanh_0_op->set_op("Tanh"); tanh_0_op->set_name(tanh_0_output); *tanh_0_op->add_input() = split_output + ":1"; (*tanh_0_op->mutable_attr())["T"].set_type(DT_FLOAT); const string sigmoid_1_output = base + "Sigmoid_1"; - auto* logistic_1_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* logistic_1_op = tensorflow_graph->add_node(); logistic_1_op->set_op("Sigmoid"); logistic_1_op->set_name(sigmoid_1_output); *logistic_1_op->add_input() = split_output; (*logistic_1_op->mutable_attr())["T"].set_type(DT_FLOAT); const string mul_1_output = base + "mul_1"; - auto* mul_1_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* mul_1_op = tensorflow_graph->add_node(); mul_1_op->set_op("Mul"); mul_1_op->set_name(mul_1_output); *mul_1_op->add_input() = sigmoid_1_output; @@ -1395,21 +1416,21 @@ void ConvertLstmCellOperator(const Model& model, const LstmCellOperator& src_op, (*mul_1_op->mutable_attr())["T"].set_type(DT_FLOAT); const string sigmoid_0_output = base + "Sigmoid"; - auto* logistic_2_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* logistic_2_op = tensorflow_graph->add_node(); logistic_2_op->set_op("Sigmoid"); logistic_2_op->set_name(sigmoid_0_output); *logistic_2_op->add_input() = split_output + ":2"; (*logistic_2_op->mutable_attr())["T"].set_type(DT_FLOAT); const string sigmoid_2_output = base + "Sigmoid_2"; - auto* logistic_3_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* logistic_3_op = tensorflow_graph->add_node(); logistic_3_op->set_op("Sigmoid"); logistic_3_op->set_name(sigmoid_2_output); *logistic_3_op->add_input() = split_output + ":3"; (*logistic_3_op->mutable_attr())["T"].set_type(DT_FLOAT); const string mul_0_output = base + "mul"; - auto* mul_0_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* mul_0_op = tensorflow_graph->add_node(); mul_0_op->set_op("Mul"); mul_0_op->set_name(mul_0_output); *mul_0_op->add_input() = src_op.inputs[LstmCellOperator::PREV_STATE_INPUT]; @@ -1417,7 +1438,7 @@ void ConvertLstmCellOperator(const Model& model, const LstmCellOperator& src_op, (*mul_0_op->mutable_attr())["T"].set_type(DT_FLOAT); const string add_1_output = src_op.outputs[LstmCellOperator::STATE_OUTPUT]; - auto* add_1_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* add_1_op = tensorflow_graph->add_node(); add_1_op->set_op("Add"); add_1_op->set_name(add_1_output); *add_1_op->add_input() = mul_0_output; @@ -1425,14 +1446,14 @@ void ConvertLstmCellOperator(const Model& model, const LstmCellOperator& src_op, (*add_1_op->mutable_attr())["T"].set_type(DT_FLOAT); const string tanh_1_output = base + "Tanh_1"; - auto* tanh_1_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* tanh_1_op = tensorflow_graph->add_node(); tanh_1_op->set_op("Tanh"); tanh_1_op->set_name(tanh_1_output); *tanh_1_op->add_input() = add_1_output; (*tanh_1_op->mutable_attr())["T"].set_type(DT_FLOAT); const string mul_2_output = src_op.outputs[LstmCellOperator::ACTIV_OUTPUT]; - auto* mul_2_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* mul_2_op = tensorflow_graph->add_node(); mul_2_op->set_op("Mul"); mul_2_op->set_name(mul_2_output); *mul_2_op->add_input() = tanh_1_output; @@ -1443,14 +1464,15 @@ void ConvertLstmCellOperator(const Model& model, const LstmCellOperator& src_op, void ConvertSpaceToBatchNDOperator(const Model& model, const SpaceToBatchNDOperator& src_op, GraphDef* tensorflow_graph) { - auto* new_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* new_op = tensorflow_graph->add_node(); new_op->set_op("SpaceToBatchND"); new_op->set_name(src_op.outputs[0]); CHECK_EQ(src_op.inputs.size(), 3); *new_op->add_input() = src_op.inputs[0]; *new_op->add_input() = src_op.inputs[1]; *new_op->add_input() = src_op.inputs[2]; - const auto params_type = GetTensorFlowDataType(model, src_op.inputs[0]); + const tensorflow::DataType params_type = + GetTensorFlowDataType(model, src_op.inputs[0]); (*new_op->mutable_attr())["T"].set_type(params_type); (*new_op->mutable_attr())["Tblock_shape"].set_type(DT_INT32); (*new_op->mutable_attr())["Tpaddings"].set_type(DT_INT32); @@ -1459,14 +1481,15 @@ void ConvertSpaceToBatchNDOperator(const Model& model, void ConvertBatchToSpaceNDOperator(const Model& model, const BatchToSpaceNDOperator& src_op, GraphDef* tensorflow_graph) { - auto* new_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* new_op = tensorflow_graph->add_node(); new_op->set_op("BatchToSpaceND"); new_op->set_name(src_op.outputs[0]); CHECK_EQ(src_op.inputs.size(), 3); *new_op->add_input() = src_op.inputs[0]; *new_op->add_input() = src_op.inputs[1]; *new_op->add_input() = src_op.inputs[2]; - const auto params_type = GetTensorFlowDataType(model, src_op.inputs[0]); + const tensorflow::DataType params_type = + GetTensorFlowDataType(model, src_op.inputs[0]); (*new_op->mutable_attr())["T"].set_type(params_type); (*new_op->mutable_attr())["Tblock_shape"].set_type(DT_INT32); (*new_op->mutable_attr())["Tcrops"].set_type(DT_INT32); @@ -1474,18 +1497,19 @@ void ConvertBatchToSpaceNDOperator(const Model& model, void ConvertPadOperator(const Model& model, const PadOperator& src_op, GraphDef* tensorflow_graph) { - auto* new_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* new_op = tensorflow_graph->add_node(); new_op->set_op("Pad"); new_op->set_name(src_op.outputs[0]); CHECK_EQ(src_op.inputs.size(), 2); *new_op->add_input() = src_op.inputs[0]; *new_op->add_input() = src_op.inputs[1]; - const auto params_type = GetTensorFlowDataType(model, src_op.inputs[0]); + const tensorflow::DataType params_type = + GetTensorFlowDataType(model, src_op.inputs[0]); (*new_op->mutable_attr())["T"].set_type(params_type); // Create the params tensor. - auto* params_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* params_op = tensorflow_graph->add_node(); params_op->set_op("Const"); params_op->set_name(src_op.inputs[1]); (*params_op->mutable_attr())["dtype"].set_type(DT_INT32); @@ -1504,7 +1528,7 @@ void ConvertPadOperator(const Model& model, const PadOperator& src_op, void ConvertPadV2Operator(const Model& model, const PadV2Operator& src_op, GraphDef* tensorflow_graph) { - auto* new_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* new_op = tensorflow_graph->add_node(); new_op->set_op("PadV2"); new_op->set_name(src_op.outputs[0]); CHECK_EQ(src_op.inputs.size(), 2); @@ -1512,11 +1536,12 @@ void ConvertPadV2Operator(const Model& model, const PadV2Operator& src_op, *new_op->add_input() = src_op.inputs[1]; *new_op->add_input() = src_op.inputs[2]; - const auto params_type = GetTensorFlowDataType(model, src_op.inputs[0]); + const tensorflow::DataType params_type = + GetTensorFlowDataType(model, src_op.inputs[0]); (*new_op->mutable_attr())["T"].set_type(params_type); // Create the params tensor. - auto* params_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* params_op = tensorflow_graph->add_node(); params_op->set_op("Const"); params_op->set_name(src_op.inputs[1]); (*params_op->mutable_attr())["dtype"].set_type(DT_INT32); @@ -1535,7 +1560,7 @@ void ConvertPadV2Operator(const Model& model, const PadV2Operator& src_op, void CreateSliceInput(const string& input_name, const std::vector& values, GraphDef* tensorflow_graph) { - auto* params_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* params_op = tensorflow_graph->add_node(); params_op->set_op("Const"); params_op->set_name(input_name); (*params_op->mutable_attr())["dtype"].set_type(DT_INT32); @@ -1552,7 +1577,7 @@ void CreateSliceInput(const string& input_name, const std::vector& values, void ConvertStridedSliceOperator(const Model& model, const StridedSliceOperator& src_op, GraphDef* tensorflow_graph) { - auto* new_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* new_op = tensorflow_graph->add_node(); new_op->set_op("StridedSlice"); new_op->set_name(src_op.outputs[0]); CHECK_EQ(src_op.inputs.size(), 4); @@ -1561,7 +1586,8 @@ void ConvertStridedSliceOperator(const Model& model, *new_op->add_input() = src_op.inputs[2]; *new_op->add_input() = src_op.inputs[3]; - const auto params_type = GetTensorFlowDataType(model, src_op.inputs[0]); + const tensorflow::DataType params_type = + GetTensorFlowDataType(model, src_op.inputs[0]); (*new_op->mutable_attr())["T"].set_type(params_type); (*new_op->mutable_attr())["Index"].set_type(DT_INT32); @@ -1579,7 +1605,7 @@ void ConvertStridedSliceOperator(const Model& model, void ConvertSliceOperator(const Model& model, const SliceOperator& src_op, GraphDef* tensorflow_graph) { - auto* new_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* new_op = tensorflow_graph->add_node(); new_op->set_op("Slice"); new_op->set_name(src_op.outputs[0]); CHECK_EQ(src_op.inputs.size(), 3); @@ -1587,7 +1613,8 @@ void ConvertSliceOperator(const Model& model, const SliceOperator& src_op, *new_op->add_input() = src_op.inputs[1]; *new_op->add_input() = src_op.inputs[2]; - const auto params_type = GetTensorFlowDataType(model, src_op.inputs[0]); + const tensorflow::DataType params_type = + GetTensorFlowDataType(model, src_op.inputs[0]); (*new_op->mutable_attr())["T"].set_type(params_type); (*new_op->mutable_attr())["Index"].set_type(DT_INT32); @@ -1598,14 +1625,15 @@ void ConvertSliceOperator(const Model& model, const SliceOperator& src_op, void ConvertMeanOperator(const Model& model, const MeanOperator& src_op, GraphDef* tensorflow_graph) { - auto* new_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* new_op = tensorflow_graph->add_node(); new_op->set_op("Mean"); new_op->set_name(src_op.outputs[0]); CHECK_EQ(src_op.inputs.size(), 2); *new_op->add_input() = src_op.inputs[0]; *new_op->add_input() = src_op.inputs[1]; - const auto params_type = GetTensorFlowDataType(model, src_op.inputs[0]); + const tensorflow::DataType params_type = + GetTensorFlowDataType(model, src_op.inputs[0]); (*new_op->mutable_attr())["T"].set_type(params_type); if (src_op.keep_dims) { @@ -1613,7 +1641,7 @@ void ConvertMeanOperator(const Model& model, const MeanOperator& src_op, } // Create the params tensor. - auto* params_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* params_op = tensorflow_graph->add_node(); params_op->set_op("Const"); params_op->set_name(src_op.inputs[1]); (*params_op->mutable_attr())["dtype"].set_type(DT_INT32); @@ -1629,13 +1657,14 @@ void ConvertMeanOperator(const Model& model, const MeanOperator& src_op, void ConvertSqueezeOperator(const Model& model, const SqueezeOperator& src_op, GraphDef* tensorflow_graph) { - auto* new_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* new_op = tensorflow_graph->add_node(); new_op->set_op("Squeeze"); new_op->set_name(src_op.outputs[0]); CHECK_EQ(src_op.inputs.size(), 1); *new_op->add_input() = src_op.inputs[0]; - const auto params_type = GetTensorFlowDataType(model, src_op.inputs[0]); + const tensorflow::DataType params_type = + GetTensorFlowDataType(model, src_op.inputs[0]); (*new_op->mutable_attr())["T"].set_type(params_type); if (!src_op.squeeze_dims.empty()) { @@ -1648,74 +1677,79 @@ void ConvertSqueezeOperator(const Model& model, const SqueezeOperator& src_op, void ConvertSubOperator(const Model& model, const SubOperator& src_op, GraphDef* tensorflow_graph) { - auto* sub_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* sub_op = tensorflow_graph->add_node(); sub_op->set_op("Sub"); sub_op->set_name(src_op.outputs[0]); CHECK_EQ(src_op.inputs.size(), 2); *sub_op->add_input() = src_op.inputs[0]; *sub_op->add_input() = src_op.inputs[1]; - const auto data_type = GetTensorFlowDataType(model, src_op.inputs[0]); + const tensorflow::DataType data_type = + GetTensorFlowDataType(model, src_op.inputs[0]); (*sub_op->mutable_attr())["T"].set_type(data_type); } void ConvertTensorFlowMinimumOperator(const Model& model, const TensorFlowMinimumOperator& src_op, GraphDef* tensorflow_graph) { - auto* sub_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* sub_op = tensorflow_graph->add_node(); sub_op->set_op("Minimum"); sub_op->set_name(src_op.outputs[0]); CHECK_EQ(src_op.inputs.size(), 2); *sub_op->add_input() = src_op.inputs[0]; *sub_op->add_input() = src_op.inputs[1]; - const auto data_type = GetTensorFlowDataType(model, src_op.inputs[0]); + const tensorflow::DataType data_type = + GetTensorFlowDataType(model, src_op.inputs[0]); (*sub_op->mutable_attr())["T"].set_type(data_type); } void ConvertTensorFlowMaximumOperator(const Model& model, const TensorFlowMaximumOperator& src_op, GraphDef* tensorflow_graph) { - auto* sub_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* sub_op = tensorflow_graph->add_node(); sub_op->set_op("Maximum"); sub_op->set_name(src_op.outputs[0]); CHECK_EQ(src_op.inputs.size(), 2); *sub_op->add_input() = src_op.inputs[0]; *sub_op->add_input() = src_op.inputs[1]; - const auto data_type = GetTensorFlowDataType(model, src_op.inputs[0]); + const tensorflow::DataType data_type = + GetTensorFlowDataType(model, src_op.inputs[0]); (*sub_op->mutable_attr())["T"].set_type(data_type); } void ConvertSelectOperator(const Model& model, const SelectOperator& src_op, GraphDef* tensorflow_graph) { - auto* sub_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* sub_op = tensorflow_graph->add_node(); sub_op->set_op("Select"); sub_op->set_name(src_op.outputs[0]); CHECK_EQ(src_op.inputs.size(), 3); *sub_op->add_input() = src_op.inputs[0]; *sub_op->add_input() = src_op.inputs[1]; *sub_op->add_input() = src_op.inputs[2]; - const auto data_type = GetTensorFlowDataType(model, src_op.inputs[1]); + const tensorflow::DataType data_type = + GetTensorFlowDataType(model, src_op.inputs[1]); (*sub_op->mutable_attr())["T"].set_type(data_type); } void ConvertTileOperator(const Model& model, const TensorFlowTileOperator& src_op, GraphDef* tensorflow_graph) { - auto* tile_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* tile_op = tensorflow_graph->add_node(); tile_op->set_op("Tile"); tile_op->set_name(src_op.outputs[0]); CHECK_EQ(src_op.inputs.size(), 2); *tile_op->add_input() = src_op.inputs[0]; *tile_op->add_input() = src_op.inputs[1]; - const auto data_type = GetTensorFlowDataType(model, src_op.inputs[0]); + const tensorflow::DataType data_type = + GetTensorFlowDataType(model, src_op.inputs[0]); (*tile_op->mutable_attr())["T"].set_type(data_type); - const auto multiples_data_type = + const tensorflow::DataType multiples_data_type = GetTensorFlowDataType(model, src_op.inputs[1]); (*tile_op->mutable_attr())["Tmultiples"].set_type(multiples_data_type); } void ConvertTopKV2Operator(const Model& model, const TopKV2Operator& src_op, GraphDef* tensorflow_graph) { - auto* topk_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* topk_op = tensorflow_graph->add_node(); topk_op->set_op("TOPKV2"); topk_op->set_name(src_op.outputs[0]); CHECK_EQ(src_op.inputs.size(), 2); @@ -1728,12 +1762,13 @@ void ConvertRandomUniformOperator(const Model& model, const RandomUniformOperator& src_op, GraphDef* tensorflow_graph) { CHECK(tensorflow_graph != nullptr); - auto* new_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* new_op = tensorflow_graph->add_node(); new_op->set_op("RandomUniform"); CHECK_EQ(src_op.inputs.size(), 1); new_op->set_name(src_op.outputs[0]); *new_op->add_input() = src_op.inputs[0]; - const auto shape_type = GetTensorFlowDataType(model, src_op.inputs[0]); + const tensorflow::DataType shape_type = + GetTensorFlowDataType(model, src_op.inputs[0]); (*new_op->mutable_attr())["T"].set_type(shape_type); (*new_op->mutable_attr())["dtype"].set_type( GetTensorFlowDataType(src_op.dtype)); @@ -1744,13 +1779,14 @@ void ConvertRandomUniformOperator(const Model& model, void ConvertComparisonOperator(const Model& model, const Operator& src_op, const char* op_name, GraphDef* tensorflow_graph) { - auto* comparison_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* comparison_op = tensorflow_graph->add_node(); comparison_op->set_op(op_name); comparison_op->set_name(src_op.outputs[0]); CHECK_EQ(src_op.inputs.size(), 2); *comparison_op->add_input() = src_op.inputs[0]; *comparison_op->add_input() = src_op.inputs[1]; - const auto data_type = GetTensorFlowDataType(model, src_op.inputs[0]); + const tensorflow::DataType data_type = + GetTensorFlowDataType(model, src_op.inputs[0]); (*comparison_op->mutable_attr())["T"].set_type(data_type); } @@ -1758,21 +1794,37 @@ void ConvertSparseToDenseOperator(const Model& model, const SparseToDenseOperator& src_op, const char* op_name, GraphDef* tensorflow_graph) { - auto* sparse_to_dense_op = tensorflow_graph->add_node(); + tensorflow::NodeDef* sparse_to_dense_op = tensorflow_graph->add_node(); sparse_to_dense_op->set_op(op_name); sparse_to_dense_op->set_name(src_op.outputs[0]); CHECK_EQ(src_op.inputs.size(), 4); for (int i = 0; i < 4; ++i) { *sparse_to_dense_op->add_input() = src_op.inputs[i]; } - const auto data_type = GetTensorFlowDataType(model, src_op.inputs[3]); + const tensorflow::DataType data_type = + GetTensorFlowDataType(model, src_op.inputs[3]); (*sparse_to_dense_op->mutable_attr())["T"].set_type(data_type); - const auto index_type = GetTensorFlowDataType(model, src_op.inputs[0]); + const tensorflow::DataType index_type = + GetTensorFlowDataType(model, src_op.inputs[0]); (*sparse_to_dense_op->mutable_attr())["Tindices"].set_type(index_type); (*sparse_to_dense_op->mutable_attr())["Tindices"].set_b( src_op.validate_indices); } +void ConvertPowOperator(const Model& model, const PowOperator& src_op, + const char* op_name, GraphDef* tensorflow_graph) { + tensorflow::NodeDef* pow_op = tensorflow_graph->add_node(); + pow_op->set_op(op_name); + pow_op->set_name(src_op.outputs[0]); + CHECK_EQ(src_op.inputs.size(), 2); + for (int i = 0; i < 2; ++i) { + *pow_op->add_input() = src_op.inputs[i]; + } + const tensorflow::DataType data_type = + GetTensorFlowDataType(model, src_op.inputs[0]); + (*pow_op->mutable_attr())["T"].set_type(data_type); +} + void ConvertOperator(const Model& model, const Operator& src_op, GraphDef* tensorflow_graph) { if (src_op.fused_activation_function != FusedActivationFunctionType::kNone) { @@ -1931,6 +1983,9 @@ void ConvertOperator(const Model& model, const Operator& src_op, } else if (src_op.type == OperatorType::kArgMax) { ConvertArgMaxOperator(model, static_cast(src_op), tensorflow_graph); + } else if (src_op.type == OperatorType::kArgMin) { + ConvertArgMinOperator(model, static_cast(src_op), + tensorflow_graph); } else if (src_op.type == OperatorType::kTopK_V2) { ConvertTopKV2Operator(model, static_cast(src_op), tensorflow_graph); @@ -1987,6 +2042,9 @@ void ConvertOperator(const Model& model, const Operator& src_op, ConvertTileOperator(model, static_cast(src_op), tensorflow_graph); + } else if (src_op.type == OperatorType::kPow) { + ConvertPowOperator(model, static_cast(src_op), "Pow", + tensorflow_graph); } else { LOG(FATAL) << "Unhandled operator type " << OperatorTypeName(src_op.type); } @@ -1994,7 +2052,7 @@ void ConvertOperator(const Model& model, const Operator& src_op, void AddPlaceholder(const string& name, ArrayDataType type, GraphDef* tensorflow_graph) { - auto* placeholder = tensorflow_graph->add_node(); + tensorflow::NodeDef* placeholder = tensorflow_graph->add_node(); placeholder->set_op("Placeholder"); switch (type) { case ArrayDataType::kBool: @@ -2023,7 +2081,7 @@ void AddPlaceholder(const string& name, ArrayDataType type, void AddPlaceholderForRNNState(const Model& model, const string& name, int size, GraphDef* tensorflow_graph) { - auto* placeholder = tensorflow_graph->add_node(); + tensorflow::NodeDef* placeholder = tensorflow_graph->add_node(); placeholder->set_op("Placeholder"); placeholder->set_name(name); (*placeholder->mutable_attr())["dtype"].set_type(DT_FLOAT); diff --git a/tensorflow/contrib/lite/toco/g3doc/cmdline_examples.md b/tensorflow/contrib/lite/toco/g3doc/cmdline_examples.md index 0ab024c6187ffed2acf860505812f16ab12a32f5..18b7848db86e553ec645fa87298420012b5f753f 100644 --- a/tensorflow/contrib/lite/toco/g3doc/cmdline_examples.md +++ b/tensorflow/contrib/lite/toco/g3doc/cmdline_examples.md @@ -11,8 +11,10 @@ Table of contents: * [Command-line tools](#tools) * [Converting models prior to TensorFlow 1.9.](#pre-tensorflow-1.9) -* [Convert a TensorFlow GraphDef](#graphdef) -* [Convert a TensorFlow SavedModel](#savedmodel) +* [Basic examples](#basic) + * [Convert a TensorFlow GraphDef](#graphdef) + * [Convert a TensorFlow SavedModel](#savedmodel) + * [Convert a tf.keras model](#keras) * [Quantization](#quantization) * [Convert a TensorFlow GraphDef for quantized inference](#graphdef-quant) * [Use "dummy-quantization" to try out quantized inference on a float @@ -51,7 +53,12 @@ API](python_api.md#pre-tensorflow-1.9). If a command line tool is desired, the Terminal for additional details on the command-line flags available. There were no command line tools in TensorFlow 1.8. -## Convert a TensorFlow GraphDef +## Basic examples + +The following section shows examples of how to convert a basic float-point model +from each of the supported data formats into a TensorFlow Lite FlatBuffers. + +### Convert a TensorFlow GraphDef The follow example converts a basic TensorFlow GraphDef (frozen by [freeze_graph.py](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/python/tools/freeze_graph.py)) @@ -70,7 +77,7 @@ tflite_convert \ The value for `input_shapes` is automatically determined whenever possible. -## Convert a TensorFlow SavedModel +### Convert a TensorFlow SavedModel The follow example converts a basic TensorFlow SavedModel into a Tensorflow Lite FlatBuffer to perform floating-point inference. @@ -95,6 +102,17 @@ There is currently no support for MetaGraphDefs without a SignatureDef or for MetaGraphDefs that use the [`assets/` directory](https://www.tensorflow.org/guide/saved_model#structure_of_a_savedmodel_directory). +### Convert a tf.Keras model + +The following example converts a `tf.keras` model into a TensorFlow Lite +Flatbuffer. The `tf.keras` file must contain both the model and the weights. + +``` +tflite_convert \ + --output_file=/tmp/foo.tflite \ + --keras_model_file=/tmp/keras_model.h5 +``` + ## Quantization ### Convert a TensorFlow GraphDef for quantized inference diff --git a/tensorflow/contrib/lite/toco/g3doc/cmdline_reference.md b/tensorflow/contrib/lite/toco/g3doc/cmdline_reference.md index 2d44b871c6616dbb415bdec550eed867ab49657b..decc8a45a40ffba2a27320ce8391b1916391d744 100644 --- a/tensorflow/contrib/lite/toco/g3doc/cmdline_reference.md +++ b/tensorflow/contrib/lite/toco/g3doc/cmdline_reference.md @@ -19,7 +19,7 @@ Table of contents: The following high level flags specify the details of the input and output files. The flag `--output_file` is always required. Additionally, either -`--graph_def_file` or `--saved_model_dir` is required. +`--graph_def_file`, `--saved_model_dir` or `--keras_model_file` is required. * `--output_file`. Type: string. Specifies the full path of the output file. * `--graph_def_file`. Type: string. Specifies the full path of the input @@ -27,6 +27,8 @@ files. The flag `--output_file` is always required. Additionally, either [freeze_graph.py](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/python/tools/freeze_graph.py). * `--saved_model_dir`. Type: string. Specifies the full path to the directory containing the SavedModel. +* `--keras_model_file`. Type: string. Specifies the full path of the HDF5 file + containing the tf.keras model. * `--output_format`. Type: string. Default: `TFLITE`. Specifies the format of the output file. Allowed values: * `TFLITE`: TensorFlow Lite FlatBuffer format. diff --git a/tensorflow/contrib/lite/toco/g3doc/python_api.md b/tensorflow/contrib/lite/toco/g3doc/python_api.md index afa6fd69572695d37a7193d5007375b11b942ce1..3799eac0a1181afe3b63d2f8651745c2ec61f5e0 100644 --- a/tensorflow/contrib/lite/toco/g3doc/python_api.md +++ b/tensorflow/contrib/lite/toco/g3doc/python_api.md @@ -15,6 +15,7 @@ Table of contents: * [Exporting a GraphDef from tf.Session](#basic-graphdef-sess) * [Exporting a GraphDef from file](#basic-graphdef-file) * [Exporting a SavedModel](#basic-savedmodel) + * [Exporting a tf.keras File](#basic-keras-file) * [Complex examples](#complex) * [Exporting a quantized GraphDef](#complex-quant) * [TensorFlow Lite Python interpreter](#interpreter) @@ -40,9 +41,11 @@ is `tf.contrib.lite.TocoConverter`. The API for calling the Python intepreter is `TocoConverter` provides class methods based on the original format of the model. `TocoConverter.from_session()` is available for GraphDefs. -`TocoConverter.from_saved_model()` is available for SavedModels. Example usages -for simple float-point models are shown in [Basic Examples](#basic). Examples -usages for more complex models is shown in [Complex Examples](#complex). +`TocoConverter.from_saved_model()` is available for SavedModels. +`TocoConverter.from_keras_model_file()` is available for `tf.Keras` files. +Example usages for simple float-point models are shown in [Basic +Examples](#basic). Examples usages for more complex models is shown in [Complex +Examples](#complex). **NOTE**: Currently, `TocoConverter` will cause a fatal error to the Python interpreter when the conversion fails. This will be remedied as soon as @@ -114,6 +117,51 @@ For more complex SavedModels, the optional parameters that can be passed into `output_arrays`, `tag_set` and `signature_key`. Details of each parameter are available by running `help(tf.contrib.lite.TocoConverter)`. +### Exporting a tf.keras File + +The following example shows how to convert a `tf.keras` model into a TensorFlow +Lite FlatBuffer. + +```python +import tensorflow as tf + +converter = tf.contrib.lite.TocoConverter.from_keras_model_file("keras_model.h5") +tflite_model = converter.convert() +open("converted_model.tflite", "wb").write(tflite_model) +``` + +The `tf.keras` file must contain both the model and the weights. A comprehensive +example including model construction can be seen below. + +```python +import numpy as np +import tensorflow as tf + +# Generate tf.keras model. +model = tf.keras.models.Sequential() +model.add(tf.keras.layers.Dense(2, input_shape=(3,))) +model.add(tf.keras.layers.RepeatVector(3)) +model.add(tf.keras.layers.TimeDistributed(tf.keras.layers.Dense(3))) +model.compile(loss=tf.keras.losses.MSE, + optimizer=tf.keras.optimizers.RMSprop(lr=0.0001), + metrics=[tf.keras.metrics.categorical_accuracy], + sample_weight_mode='temporal') + +x = np.random.random((1, 3)) +y = np.random.random((1, 3, 3)) +model.train_on_batch(x, y) +model.predict(x) + +# Save tf.keras model in HDF5 format. +keras_file = "keras_model.h5" +tf.keras.models.save_model(model, keras_file) + +# Convert to TensorFlow Lite model. +converter = tf.contrib.lite.TocoConverter.from_keras_model_file(keras_file) +tflite_model = converter.convert() +open("converted_model.tflite", "wb").write(tflite_model) +``` + ## Complex examples For models where the default value of the attributes is not sufficient, the diff --git a/tensorflow/contrib/lite/toco/graph_transformations/dequantize.cc b/tensorflow/contrib/lite/toco/graph_transformations/dequantize.cc index 2c7ffe488477ef1a544dfe6f36a6e0d1ac40aa96..1688586733b0434c7fc98686a19f0ceb8092f33b 100644 --- a/tensorflow/contrib/lite/toco/graph_transformations/dequantize.cc +++ b/tensorflow/contrib/lite/toco/graph_transformations/dequantize.cc @@ -159,6 +159,7 @@ bool DequantizeArray(const string& array_name, new_array.GetOrCreateMinMax() = array->GetMinMax(); fakequant_op->minmax.reset(new MinMax); *fakequant_op->minmax = array->GetMinMax(); + fakequant_op->narrow_range = array->narrow_range; if (must_insert_fakequant_before) { for (const auto& op : model->operators) { for (string& output : op->outputs) { diff --git a/tensorflow/contrib/lite/toco/graph_transformations/ensure_bias_vectors.cc b/tensorflow/contrib/lite/toco/graph_transformations/ensure_bias_vectors.cc index 708ecf6e0a96811ab274fbb25f748f562cd3afad..e80ed036b311cfc586c40ece410ef6a6432a0cd9 100644 --- a/tensorflow/contrib/lite/toco/graph_transformations/ensure_bias_vectors.cc +++ b/tensorflow/contrib/lite/toco/graph_transformations/ensure_bias_vectors.cc @@ -26,17 +26,38 @@ namespace toco { namespace { +int GetOutputDepthFromWeights(const Model& model, const Operator& op) { + const string& weights_name = op.inputs[1]; + const auto& weights_shape = model.GetArray(weights_name).shape(); + if (op.type == OperatorType::kConv || + op.type == OperatorType::kFullyConnected) { + return weights_shape.dims(0); + } + if (op.type == OperatorType::kDepthwiseConv) { + return weights_shape.dims(3); + } + LOG(FATAL) << "Unhandled operator type"; + return 0; +} + bool ProcessLinearOperator(Model* model, Operator* op) { if (op->inputs.size() >= 3) { return false; } const string& output_name = op->outputs[0]; + const string& weights_name = op->inputs[1]; + if (!model->GetArray(weights_name).has_shape()) { + return false; + } + const int depth = GetOutputDepthFromWeights(*model, *op); const string& bias_name = AvailableArrayName(*model, output_name + "_bias"); op->inputs.push_back(bias_name); DCHECK_EQ(op->inputs.size(), 3); auto& bias_array = model->GetOrCreateArray(bias_name); bias_array.data_type = ArrayDataType::kFloat; - + bias_array.mutable_shape()->mutable_dims()->push_back(depth); + auto& bias_buffer = bias_array.GetMutableBuffer(); + bias_buffer.data.resize(depth, 0.f); return true; } } // namespace diff --git a/tensorflow/contrib/lite/toco/graph_transformations/graph_transformations.h b/tensorflow/contrib/lite/toco/graph_transformations/graph_transformations.h index 4025fede6f160d7ad0fb09be99c246adb93b43a6..7cc9bb75d7debbe28ec9b345fc639f9fc91b6844 100644 --- a/tensorflow/contrib/lite/toco/graph_transformations/graph_transformations.h +++ b/tensorflow/contrib/lite/toco/graph_transformations/graph_transformations.h @@ -135,6 +135,7 @@ DECLARE_GRAPH_TRANSFORMATION(IdentifyRelu1) DECLARE_GRAPH_TRANSFORMATION(IdentifyPRelu) DECLARE_GRAPH_TRANSFORMATION(IdentifyDilatedConv) DECLARE_GRAPH_TRANSFORMATION(MakeInitialDequantizeOperator) +DECLARE_GRAPH_TRANSFORMATION(MoveBinaryOperatorBeforeReshape) DECLARE_GRAPH_TRANSFORMATION(PropagateActivationFunctionIntoConstants) DECLARE_GRAPH_TRANSFORMATION(PropagateArrayDataTypes) DECLARE_GRAPH_TRANSFORMATION(PropagateFakeQuantNumBits); @@ -158,7 +159,7 @@ DECLARE_GRAPH_TRANSFORMATION(ResolveConstantBinaryOperator) DECLARE_GRAPH_TRANSFORMATION(ResolveConstantUnaryOperator) DECLARE_GRAPH_TRANSFORMATION(CreateIm2colArrays) DECLARE_GRAPH_TRANSFORMATION(DropIm2colArrays) -DECLARE_GRAPH_TRANSFORMATION(ReadFakeQuantMinMax) +DECLARE_GRAPH_TRANSFORMATION(ReadArrayMinmaxAndNarrowRangeFromFakeQuant) DECLARE_GRAPH_TRANSFORMATION(ReorderElementwiseUnary) DECLARE_GRAPH_TRANSFORMATION(ReorderReshapeTranspose) DECLARE_GRAPH_TRANSFORMATION(ResolveReorderAxes) @@ -193,6 +194,7 @@ DECLARE_GRAPH_TRANSFORMATION(ResolveMultiplyByZero) DECLARE_GRAPH_TRANSFORMATION(Dequantize) DECLARE_GRAPH_TRANSFORMATION(UnpartitionEmbeddingLookup) DECLARE_GRAPH_TRANSFORMATION(ShuffleFCWeights) +DECLARE_GRAPH_TRANSFORMATION(ResolveFakeQuantArgsFromVars) class PropagateDefaultMinMax : public GraphTransformation { public: diff --git a/tensorflow/contrib/lite/toco/graph_transformations/hardcode_min_max.cc b/tensorflow/contrib/lite/toco/graph_transformations/hardcode_min_max.cc index 82a4308ecb134d28c37f4519ae783b50bf35477a..2f1bb8f0ad6374243e5a094701eef54cd086151a 100644 --- a/tensorflow/contrib/lite/toco/graph_transformations/hardcode_min_max.cc +++ b/tensorflow/contrib/lite/toco/graph_transformations/hardcode_min_max.cc @@ -133,24 +133,20 @@ bool HardcodeMinMaxForConcatenation(Model* model, Operator* op) { } bool HardcodeMinMaxForSplit(Model* model, Operator* op) { - for (const auto& output : op->outputs) { - if (model->GetArray(output).minmax) { - LOG(WARNING) << "Skipping min-max setting for " << LogName(*op) - << " because output " << output << " already has min-max."; - return false; - } - } // Data is in second input. auto& input_array = model->GetArray(op->inputs[1]); if (!input_array.minmax) { return false; - } else { - for (const auto& output : op->outputs) { - auto& array = model->GetArray(output); + } + bool changed = false; + for (const auto& output : op->outputs) { + auto& array = model->GetArray(output); + if (!array.minmax || !(array.GetMinMax() == input_array.GetMinMax())) { + changed = true; array.GetOrCreateMinMax() = *input_array.minmax; } - return true; } + return changed; } // The output of average or max pooling is within the same range as its input. @@ -232,6 +228,14 @@ bool HardcodeMinMaxForOutput(Model* model, Operator* op, double min, return true; } +bool MinMaxApproximatelyEqual(const MinMax& minmax1, const MinMax& minmax2) { + const double magnitude = + std::min(minmax1.max - minmax1.min, minmax2.max - minmax2.min); + const double tolerated = 1e-6 * magnitude; + return std::abs(minmax1.min - minmax2.min) < tolerated && + std::abs(minmax1.max - minmax2.max) < tolerated; +} + // Propagates MinMax from any of the listed arrays, to all others. // If multiple of these arrays have MinMax, then these are required // to agree with each other. @@ -254,7 +258,7 @@ bool PropagateMinMaxAmongArrays(Model* model, for (const string& array_name : array_names) { auto& array = model->GetArray(array_name); if (array.minmax) { - CHECK(*array.minmax == *reference_minmax) + CHECK(MinMaxApproximatelyEqual(*array.minmax, *reference_minmax)) << "Both the following arrays have minmax, and they disagree: " << reference_array_name << " (" << reference_minmax->min << "," << reference_minmax->max << ") and " << array_name << " (" diff --git a/tensorflow/contrib/lite/toco/graph_transformations/identify_lstm.cc b/tensorflow/contrib/lite/toco/graph_transformations/identify_lstm.cc index 685353a846f706571dc8872310683d2eb24998fa..c0b014b45eb1df25173ce3ca3fa488b0655c3c76 100644 --- a/tensorflow/contrib/lite/toco/graph_transformations/identify_lstm.cc +++ b/tensorflow/contrib/lite/toco/graph_transformations/identify_lstm.cc @@ -35,19 +35,24 @@ std::vector>::iterator FindOperator( return it; } -bool GetStateArrayForBackEdge(const Model& model, - const string& back_edge_source_array, - string* state_array = nullptr) { - for (const auto& rnn_state : model.flags.rnn_states()) { - if (back_edge_source_array == rnn_state.back_edge_source_array()) { - // Found LSTM cell output - if (state_array) { - *state_array = rnn_state.state_array(); - } - return true; +bool ValidateSourceOp(const Model& model, const string& array_name, + OperatorType op_type, Operator** source_op) { + if (op_type == OperatorType::kNone) { + CHECK(!source_op); + } else { + CHECK(source_op); + *source_op = GetOpWithOutput(model, array_name); + if (*source_op == nullptr) { + return false; + } + + // Check that first operator, if connected, is of correct type + if ((*source_op)->type != op_type) { + return false; } } - return false; + + return true; } // Returns true if the given operator has exactly 1 input, and is connected to @@ -62,24 +67,10 @@ bool MatchOperatorInputs(const Operator& op, const Model& model, } // Check if first input is disconnected/connected to an operator - Operator* x = GetOpWithOutput(model, op.inputs[0]); - if ((op_type == OperatorType::kNone) && (x != nullptr)) { - return false; - } - if ((op_type != OperatorType::kNone) && (x == nullptr)) { + if (!ValidateSourceOp(model, op.inputs[0], op_type, connected_op)) { return false; } - // Check that first operator, if connected, is of correct type - if ((x != nullptr) && (x->type != op_type)) { - return false; - } - - // Successfully matched. Optionally return matching input operators. - if (connected_op) { - *connected_op = x; - } - return true; } @@ -96,40 +87,15 @@ bool MatchOperatorInputs(const Operator& op, const Model& model, } // Check if first input is disconnected/connected to an operator - Operator* x = GetOpWithOutput(model, op.inputs[0]); - if ((a_op_type == OperatorType::kNone) && (x != nullptr)) { - return false; - } - if ((a_op_type != OperatorType::kNone) && (x == nullptr)) { - return false; - } - - // Check that first operator, if connected, is of correct type - if ((x != nullptr) && (x->type != a_op_type)) { + if (!ValidateSourceOp(model, op.inputs[0], a_op_type, a_op)) { return false; } // Check if second input is disconnected/connected to an operator - Operator* y = GetOpWithOutput(model, op.inputs[1]); - if ((b_op_type == OperatorType::kNone) && (y != nullptr)) { - return false; - } - if ((b_op_type != OperatorType::kNone) && (y == nullptr)) { + if (!ValidateSourceOp(model, op.inputs[1], b_op_type, b_op)) { return false; } - // Check that second operator, if connected, is of correct type - if ((y != nullptr) && (y->type != b_op_type)) { - return false; - } - - // Successfully matched. Optionally return matching input operators. - if (a_op != nullptr) { - *a_op = x; - } - if (b_op != nullptr) { - *b_op = y; - } return true; } @@ -147,57 +113,20 @@ bool MatchOperatorInputs(const Operator& op, const Model& model, } // Check if first input is disconnected/connected to an operator - Operator* x = GetOpWithOutput(model, op.inputs[0]); - if ((a_op_type == OperatorType::kNone) && (x != nullptr)) { - return false; - } - if ((a_op_type != OperatorType::kNone) && (x == nullptr)) { - return false; - } - - // Check that first operator, if connected, is of correct type - if ((x != nullptr) && (x->type != a_op_type)) { + if (!ValidateSourceOp(model, op.inputs[0], a_op_type, a_op)) { return false; } // Check if second input is disconnected/connected to an operator - Operator* y = GetOpWithOutput(model, op.inputs[1]); - if ((b_op_type == OperatorType::kNone) && (y != nullptr)) { - return false; - } - if ((b_op_type != OperatorType::kNone) && (y == nullptr)) { - return false; - } - - // Check that second operator, if connected, is of correct type - if ((y != nullptr) && (y->type != b_op_type)) { + if (!ValidateSourceOp(model, op.inputs[1], b_op_type, b_op)) { return false; } // Check if third input is disconnected/connected to an operator - Operator* z = GetOpWithOutput(model, op.inputs[2]); - if ((c_op_type == OperatorType::kNone) && (z != nullptr)) { - return false; - } - if ((c_op_type != OperatorType::kNone) && (z == nullptr)) { - return false; - } - - // Check that third operator, if connected, is of correct type - if ((z != nullptr) && (z->type != c_op_type)) { + if (!ValidateSourceOp(model, op.inputs[2], c_op_type, c_op)) { return false; } - // Successfully matched. Optionally return matching input operators. - if (a_op != nullptr) { - *a_op = x; - } - if (b_op != nullptr) { - *b_op = y; - } - if (c_op != nullptr) { - *c_op = z; - } return true; } @@ -231,11 +160,6 @@ bool IdentifyLstmCell::Run(Model* model, std::size_t op_index) { &state_combine_add)) { return false; } - string prev_state; - if (!GetStateArrayForBackEdge(*model, state_output_tanh->inputs[0], - &prev_state)) { - return false; - } // State forget & remember addition Operator *state_forget_mul, *state_remember_mul; @@ -244,9 +168,7 @@ bool IdentifyLstmCell::Run(Model* model, std::size_t op_index) { &state_remember_mul)) { return false; } - if (state_forget_mul->inputs[0] != prev_state) { - return false; - } + const string prev_state = state_forget_mul->inputs[0]; // State forget gate Operator* state_forget_sig; diff --git a/tensorflow/contrib/lite/toco/graph_transformations/identify_prelu.cc b/tensorflow/contrib/lite/toco/graph_transformations/identify_prelu.cc index 30be4ac0aa5e9f639bbf0630e142c2806faa3260..b90a156a0dcfcd77c3e2b47bb0d77e246f2fc625 100644 --- a/tensorflow/contrib/lite/toco/graph_transformations/identify_prelu.cc +++ b/tensorflow/contrib/lite/toco/graph_transformations/identify_prelu.cc @@ -74,14 +74,30 @@ bool IdentifyPRelu::Run(Model* model, std::size_t op_index) { const auto* relu_neg_input_op = GetOpWithOutput(*model, mul_op->inputs[1]); if (relu_neg_input_op == nullptr || - relu_neg_input_op->type != OperatorType::kNeg || - relu_neg_input_op->fused_activation_function != - FusedActivationFunctionType::kRelu || relu_neg_input_op->inputs.size() != 1) { return false; } - if (relu_input_op->inputs[0] != relu_neg_input_op->inputs[0]) { + const Operator* final_input_op; + if (relu_neg_input_op->type == OperatorType::kNeg && + relu_neg_input_op->fused_activation_function == + FusedActivationFunctionType::kRelu) { + // This detects a Neg op with fused Relu activation function. + final_input_op = relu_neg_input_op; + } else { + // This detects a Neg op followed by a separated Relu op. + const auto* neg_input_op = + GetOpWithOutput(*model, relu_neg_input_op->inputs[0]); + if (neg_input_op == nullptr || neg_input_op->inputs.size() != 1 || + relu_neg_input_op->type != OperatorType::kRelu || + relu_neg_input_op->fused_activation_function != + FusedActivationFunctionType::kNone) { + return false; + } + final_input_op = neg_input_op; + } + + if (relu_input_op->inputs[0] != final_input_op->inputs[0]) { return false; } @@ -112,7 +128,6 @@ bool IdentifyPRelu::Run(Model* model, std::size_t op_index) { // intermediate tensors aren't used by other ops, those will be removed by // other graph transformation rules. model->operators.erase(FindOp(*model, add_op)); - return true; } diff --git a/tensorflow/contrib/lite/toco/graph_transformations/make_initial_dequantize_operator.cc b/tensorflow/contrib/lite/toco/graph_transformations/make_initial_dequantize_operator.cc index 45d9f73a1e6416b8f3fe3936c740da637961b7fc..f684de08abf72d05d4408bf6341fa5a3c2ed11cd 100644 --- a/tensorflow/contrib/lite/toco/graph_transformations/make_initial_dequantize_operator.cc +++ b/tensorflow/contrib/lite/toco/graph_transformations/make_initial_dequantize_operator.cc @@ -85,15 +85,8 @@ bool AddDequantizeOperatorToInput(const string& input_name, const Operator* op, dequantized_input_minmax = input_minmax; auto& input_qparams = input_array.GetOrCreateQuantizationParams(); input_array.data_type = input_array.final_data_type; - if (input_array.data_type == ArrayDataType::kUint8) { - GetQuantizationParamsFromMinMax(input_minmax, - &input_qparams); - } else if (input_array.data_type == ArrayDataType::kInt16) { - GetQuantizationParamsFromMinMax(input_minmax, - &input_qparams); - } else { - LOG(FATAL) << "unhandled data type"; - } + ChooseQuantizationParamsForArrayAndQuantizedDataType( + input_array, input_array.data_type, &input_qparams); transformation->AddMessageF( "Created %s" diff --git a/tensorflow/contrib/lite/toco/graph_transformations/move_binary_operator_before_reshape.cc b/tensorflow/contrib/lite/toco/graph_transformations/move_binary_operator_before_reshape.cc new file mode 100644 index 0000000000000000000000000000000000000000..7f44c65285bdef6ba314b16122fdd550bfa47e6a --- /dev/null +++ b/tensorflow/contrib/lite/toco/graph_transformations/move_binary_operator_before_reshape.cc @@ -0,0 +1,178 @@ +/* Copyright 2017 The TensorFlow Authors. All Rights Reserved. + + Licensed under the Apache License, Version 2.0 (the "License"); + you may not use this file except in compliance with the License. + You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + + Unless required by applicable law or agreed to in writing, software + distributed under the License is distributed on an "AS IS" BASIS, + WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + See the License for the specific language governing permissions and + limitations under the License. + ==============================================================================*/ +#include + +#include "tensorflow/contrib/lite/toco/graph_transformations/graph_transformations.h" +#include "tensorflow/contrib/lite/toco/model.h" +#include "tensorflow/contrib/lite/toco/tooling_util.h" + +namespace toco { + +namespace { + +bool IsTailOfShape(const Shape& tail, const Shape& shape) { + // Return true if 'tail' dimensions are the same as the ending dimensions of + // 'shape'. + + int shape_end = shape.dimensions_count() - 1; + int tail_end = tail.dimensions_count() - 1; + + if (tail_end > shape_end) { + // tail cannot be longer than shape. + return false; + } + + // Walk dimensions back to front and compare + for (int i = 0; i <= tail_end; i++) { + if (shape.dims(shape_end - i) != tail.dims(tail_end - i)) { + return false; + } + } + return true; +} + +} // namespace + +// If a binary operator is doing a broadcast operation from a constant array, +// and the constant array shape is the tail of both the other input shape, and a +// subsequent reshape op's output shape, we can swap their order. Since we +// prefer to have reshape ops after mathematic ops, this can allow for the +// collapsing of some reshapes. The WaveNet model in particular benefits from +// this transformation. +// +// Note we are testing for one particular case of a broader set of possible +// binary-reshape op transformations. This transformation could be generalized. +bool MoveBinaryOperatorBeforeReshape::Run(Model* model, std::size_t op_index) { + const auto binary_it = model->operators.begin() + op_index; + Operator* binary_op = binary_it->get(); + if (binary_op->type != OperatorType::kAdd && + binary_op->type != OperatorType::kMul && + binary_op->type != OperatorType::kSub && + binary_op->type != OperatorType::kDiv && + binary_op->type != OperatorType::kFloorDiv && + binary_op->type != OperatorType::kFloorMod && + binary_op->type != OperatorType::kMinimum && + binary_op->type != OperatorType::kMaximum && + binary_op->type != OperatorType::kLess && + binary_op->type != OperatorType::kLessEqual && + binary_op->type != OperatorType::kGreater && + binary_op->type != OperatorType::kGreaterEqual) { + return false; + } + + // BINARY OP INPUT CHECKS + CHECK_EQ(binary_op->inputs.size(), 2); + const bool input_is_const[2] = { + IsConstantParameterArray(*model, binary_op->inputs[0]), + IsConstantParameterArray(*model, binary_op->inputs[1]), + }; + if (!input_is_const[0] && !input_is_const[1]) { + // To limit our scope, we require one constant input. Though there's no + // reason this transformation wouldn't work with all variable inputs. + return false; + } + if (input_is_const[0] && input_is_const[1]) { + // Both inputs are constants. Leave this for constants propagation. + return false; + } + const int constant_input_idx = input_is_const[0] ? 0 : 1; + const int variable_input_idx = input_is_const[0] ? 1 : 0; + CHECK(input_is_const[constant_input_idx]); + CHECK(!input_is_const[variable_input_idx]); + + const auto& variable_input_array = + model->GetArray(binary_op->inputs[variable_input_idx]); + if (!variable_input_array.has_shape()) { + AddMessageF( + "Not moving %s because it's non-constant input shape is not resolved.", + LogName(*binary_op)); + return false; + } + if (!IsTailOfShape( + model->GetArray(binary_op->inputs[constant_input_idx]).shape(), + model->GetArray(binary_op->inputs[variable_input_idx]).shape())) { + // Constant array shape must be the latter part of the variable shape. + return false; + } + + // RESHAPE OP CHECKS + auto reshape_it = + FindOpWithOutput(*model, binary_op->inputs[variable_input_idx]); + if (reshape_it == model->operators.end()) { + AddMessageF("Not moving %s because it's variable input is not connected.", + LogName(*binary_op)); + return false; + } + Operator* reshape_op = reshape_it->get(); + if (reshape_op->type != OperatorType::kReshape) { + AddMessageF("Not moving %s because the preceding %s is not a reshape op", + LogName(*binary_op), LogName(*reshape_op)); + return false; + } + const auto& reshape_input_array = model->GetArray(reshape_op->inputs[0]); + if (!reshape_input_array.has_shape()) { + AddMessageF( + "Not moving %s because it's non-constant input shape is not resolved " + "yet", + LogName(*binary_op)); + return false; + } + if (!IsTailOfShape( + model->GetArray(binary_op->inputs[constant_input_idx]).shape(), + model->GetArray(reshape_op->outputs[0]).shape())) { + // Constant array shape must be the latter part of the binary op output + // shape. + return false; + } + + // EXTRA CHECKS ON CONNECTING ARRAY + for (const string& output_array : model->flags.output_arrays()) { + if (binary_op->inputs[variable_input_idx] == output_array) { + AddMessageF( + "Not moving %s because the output of reshape op %s is an output op.", + LogName(*binary_op), LogName(*reshape_op)); + return false; + } + } + int count_ops_consuming_output = + CountOpsWithInput(*model, binary_op->inputs[variable_input_idx]); + DCHECK_GE(count_ops_consuming_output, 1); + if (count_ops_consuming_output > 1) { + AddMessageF( + "Not moving %s because the output of reshape op %s is consumed by " + "another op", + LogName(*binary_op), LogName(*reshape_op)); + return false; + } + + // SWAP ORDER OF BINARY AND RESHAPE OPS + AddMessageF("Moving op %s before reshape op %s", LogName(*binary_op), + LogName(*reshape_op)); + + // Swap op input and outputs + std::iter_swap(reshape_op->inputs.begin(), + binary_op->inputs.begin() + variable_input_idx); + std::iter_swap(reshape_op->outputs.begin(), binary_op->outputs.begin()); + + // Swap operator ordering + std::iter_swap(binary_it, reshape_it); + + // Clear binary output shape so it will be re-propagated + model->GetArray(binary_op->outputs[0]).clear_shape(); + + return true; +} + +} // namespace toco diff --git a/tensorflow/contrib/lite/toco/graph_transformations/propagate_array_data_types.cc b/tensorflow/contrib/lite/toco/graph_transformations/propagate_array_data_types.cc index 27a1049eaf830e2c690dbc68f80d37107eb76772..670bcf64e7aaaf761bceba41a2f17bb05aa4f2b6 100644 --- a/tensorflow/contrib/lite/toco/graph_transformations/propagate_array_data_types.cc +++ b/tensorflow/contrib/lite/toco/graph_transformations/propagate_array_data_types.cc @@ -100,6 +100,13 @@ bool PropagateArrayDataTypes::Run(Model* model, std::size_t op_index) { model->GetArray(op->outputs[0]).data_type = argmax_op->output_data_type; break; } + case OperatorType::kArgMin: { + // Data type of the ArgMin op is specified. + CHECK_EQ(op->outputs.size(), 1); + auto* argmin_op = static_cast(op); + model->GetArray(op->outputs[0]).data_type = argmin_op->output_data_type; + break; + } case OperatorType::kRange: { auto* range_op = static_cast(op); // Output type of the Range op can be set via an attribute @@ -175,6 +182,14 @@ bool PropagateArrayDataTypes::Run(Model* model, std::size_t op_index) { SetDataTypeForAllOutputs(model, op, data_type); break; } + case OperatorType::kPow: { + CHECK_EQ(op->inputs.size(), 2); + CHECK(model->GetArray(op->inputs[0]).data_type == + model->GetArray(op->inputs[1]).data_type); + const ArrayDataType data_type = model->GetArray(op->inputs[0]).data_type; + SetDataTypeForAllOutputs(model, op, data_type); + break; + } default: { // These operators produce outputs with the same type as their 1st input CHECK_GT(op->inputs.size(), 0); diff --git a/tensorflow/contrib/lite/toco/graph_transformations/propagate_default_min_max.cc b/tensorflow/contrib/lite/toco/graph_transformations/propagate_default_min_max.cc index 50b90e7c2bfddb0382a4d44ad6c90fc7f7701273..cd078ef189e922682098a0ec8dc4743060181aac 100644 --- a/tensorflow/contrib/lite/toco/graph_transformations/propagate_default_min_max.cc +++ b/tensorflow/contrib/lite/toco/graph_transformations/propagate_default_min_max.cc @@ -25,6 +25,14 @@ limitations under the License. namespace toco { +namespace { + +bool SupportsMinMax(const Array& array) { + return array.data_type == ArrayDataType::kFloat; +} + +} // namespace + // Propagates default min/max values to any operator input/output array that // is missing them. // @@ -39,14 +47,16 @@ bool PropagateDefaultMinMax::Run(Model* model, std::size_t op_index) { for (const auto& input : op->inputs) { auto& input_array = model->GetArray(input); - if (!input_array.minmax && !input_array.buffer) { + if (!input_array.minmax && !input_array.buffer && + SupportsMinMax(input_array)) { did_change |= SetArrayMinMax(input, &input_array); } } for (const auto& output : op->outputs) { auto& output_array = model->GetArray(output); - if (!output_array.minmax && !output_array.buffer) { + if (!output_array.minmax && !output_array.buffer && + SupportsMinMax(output_array)) { did_change |= SetArrayMinMax(output, &output_array); } } diff --git a/tensorflow/contrib/lite/toco/graph_transformations/propagate_fake_quant_num_bits.cc b/tensorflow/contrib/lite/toco/graph_transformations/propagate_fake_quant_num_bits.cc index e25125b429a7e33fa83c603eb85b931ab45ecb50..3ad6b0ec6f7a3c4a9a0ab3964c1198ee757ea4b5 100644 --- a/tensorflow/contrib/lite/toco/graph_transformations/propagate_fake_quant_num_bits.cc +++ b/tensorflow/contrib/lite/toco/graph_transformations/propagate_fake_quant_num_bits.cc @@ -27,11 +27,15 @@ namespace toco { namespace { -void ChangeArrayDataType(GraphTransformation* transformation, Array* array, +bool ChangeArrayDataType(GraphTransformation* transformation, Array* array, ArrayDataType new_data_type, const MinMax* new_minmax) { // Ensure the array ends up in the new type (if it hasn't yet been quantized). - array->final_data_type = new_data_type; + bool changed = false; + if (array->final_data_type != new_data_type) { + array->final_data_type = new_data_type; + changed = true; + } if (array->minmax && array->quantization_params) { // The array is already quantized and has min/max info. @@ -62,18 +66,16 @@ void ChangeArrayDataType(GraphTransformation* transformation, Array* array, "Rescaling min/max from %g,%g (%s) to %g,%g (%s)", array_minmax.min, array_minmax.max, ArrayDataTypeName(array->data_type), min, max, ArrayDataTypeName(new_data_type)); - array_minmax.min = min; array_minmax.max = max; - GetQuantizationParamsFromMinMax( - array_minmax, array->quantization_params.get()); - + ChooseQuantizationParamsForArrayAndQuantizedDataType( + *array, new_data_type, array->quantization_params.get()); // Directly change the type as the array was already quantized. array->data_type = new_data_type; - } else { + changed = true; + } else if (!array->quantization_params) { // Array has not yet been quantized so we can just set the final data type // and assign the new min/max value (if provided). - CHECK(!array->quantization_params); if (!array->minmax && new_minmax) { transformation->AddMessageF("Forcing new minmax to %g,%g (%s)", @@ -82,8 +84,11 @@ void ChangeArrayDataType(GraphTransformation* transformation, Array* array, auto& array_minmax = array->GetOrCreateMinMax(); array_minmax.min = new_minmax->min; array_minmax.max = new_minmax->max; + changed = true; } } + + return changed; } // Returns true if the op blocks our backward recursive data type propagation. @@ -159,9 +164,8 @@ bool RecursivelyBackwardPropagateDataType(GraphTransformation* transformation, "Adjusting input final data type of array %s from %s to %s", input, ArrayDataTypeName(input_array.final_data_type), ArrayDataTypeName(new_data_type)); - did_change = true; - ChangeArrayDataType(transformation, &input_array, new_data_type, - &new_minmax); + did_change |= ChangeArrayDataType(transformation, &input_array, + new_data_type, &new_minmax); // Walk up into all ops producing the inputs to this op. for (auto& producing_op : model->operators) { @@ -212,9 +216,8 @@ bool RecursivelyForwardPropagateDataType(GraphTransformation* transformation, "Adjusting output final data type of array %s from %s to %s", output, ArrayDataTypeName(output_array.final_data_type), ArrayDataTypeName(new_data_type)); - did_change = true; - ChangeArrayDataType(transformation, &output_array, new_data_type, - nullptr); + did_change |= ChangeArrayDataType(transformation, &output_array, + new_data_type, nullptr); // Walk down into all ops consuming the output of this op. for (auto& consuming_op : model->operators) { diff --git a/tensorflow/contrib/lite/toco/graph_transformations/propagate_fixed_sizes.cc b/tensorflow/contrib/lite/toco/graph_transformations/propagate_fixed_sizes.cc index c61da203c63ae0b449c7d4b0cb63945bb3551f3a..4f95c57451d618c59e8caa49b6fba5044a94681e 100644 --- a/tensorflow/contrib/lite/toco/graph_transformations/propagate_fixed_sizes.cc +++ b/tensorflow/contrib/lite/toco/graph_transformations/propagate_fixed_sizes.cc @@ -120,49 +120,7 @@ void ComputeBinaryOperatorOutputSize(const Shape& input_shape_x, CHECK(output_array->has_shape()); } -int GetOutputDepthFromWeights(const Model& model, const Operator& op) { - const string& weights_name = op.inputs[1]; - const auto& weights_shape = model.GetArray(weights_name).shape(); - if (op.type == OperatorType::kConv || - op.type == OperatorType::kFullyConnected) { - return weights_shape.dims(0); - } else if (op.type == OperatorType::kDepthwiseConv) { - return weights_shape.dims(3); - } else { - LOG(FATAL) << "Unhandled operator type"; - } -} - -bool EnsureBiasVectorShape(Model* model, Operator* op) { - const string& weights_name = op->inputs[1]; - const auto& weights_array = model->GetArray(weights_name); - // Yield until weights shape has been resolved. - if (!weights_array.has_shape()) { - return false; - } - - if (op->inputs.size() < 3) { - return false; - } - auto& bias_array = model->GetArray(op->inputs[2]); - if (bias_array.has_shape()) { - return true; - } - - const int output_depth = GetOutputDepthFromWeights(*model, *op); - bias_array.copy_shape(Shape({output_depth})); - - auto& float_buffer = bias_array.GetMutableBuffer(); - float_buffer.data.resize(output_depth, 0); - - return true; -} - void ProcessConvOperator(Model* model, ConvOperator* op) { - if (!EnsureBiasVectorShape(model, op)) { - return; - } - const auto& input_array = model->GetArray(op->inputs[0]); // Yield until input dims have been resolved. if (!input_array.has_shape()) { @@ -292,10 +250,6 @@ void ProcessTransposeConvOperator(Model* model, TransposeConvOperator* op) { } void ProcessDepthwiseConvOperator(Model* model, DepthwiseConvOperator* op) { - if (!EnsureBiasVectorShape(model, op)) { - return; - } - const auto& input_array = model->GetArray(op->inputs[0]); // Yield until input dims have been resolved. if (!input_array.has_shape()) { @@ -410,10 +364,6 @@ void ProcessOpWithShapeInput(Model* model, Operator* op) { } void ProcessFullyConnectedOperator(Model* model, FullyConnectedOperator* op) { - if (!EnsureBiasVectorShape(model, op)) { - return; - } - const auto& input_array = model->GetArray(op->inputs[0]); // Yield until input dims have been resolved. if (!input_array.has_shape()) { @@ -1089,9 +1039,6 @@ void ProcessGatherOperator(Model* model, GatherOperator* op) { QCHECK_GE(input_shape.dimensions_count(), 1); op->input_rank = input_shape.dimensions_count(); - // We only support 1-D indices. - QCHECK_EQ(indices_shape.dimensions_count(), 1); - // Copy the input dimensions to the output except for dimension 0, // where the dimension of indices_shape is used. // TODO(mgubin): if axis != 0 this is not true, change when it's supported. @@ -1341,8 +1288,8 @@ void ProcessStridedSliceOperator(Model* model, StridedSliceOperator* op) { op->begin_mask, op->start_indices, op->strides, input_array.shape().dims().data(), axis); int stop_index = tflite::strided_slice::StopForAxis( - op->end_mask, op->stop_indices, op->strides, - input_array.shape().dims().data(), axis); + op->end_mask, op->shrink_axis_mask, op->stop_indices, op->strides, + input_array.shape().dims().data(), axis, start_index); int dim_size = ceil(static_cast(stop_index - start_index) / op->strides[axis]); @@ -1457,7 +1404,8 @@ void ProcessTransposeOperator(Model* model, TransposeOperator* op) { } } -void ProcessArgMaxOperator(Model* model, ArgMaxOperator* op) { +template +void ProcessArgMinMaxOperator(Model* model, Op* op) { CHECK_EQ(op->inputs.size(), 2); const auto& input_array = model->GetArray(op->inputs[0]); // Yield until input dims have been resolved. @@ -1611,6 +1559,7 @@ bool PropagateFixedSizes::Run(Model* model, std::size_t op_index) { case OperatorType::kGreaterEqual: case OperatorType::kEqual: case OperatorType::kNotEqual: + case OperatorType::kPow: ProcessSimpleBinaryOperator(model, op); break; case OperatorType::kAddN: @@ -1748,7 +1697,12 @@ bool PropagateFixedSizes::Run(Model* model, std::size_t op_index) { static_cast(op)); break; case OperatorType::kArgMax: - ProcessArgMaxOperator(model, static_cast(op)); + ProcessArgMinMaxOperator( + model, static_cast(op)); + break; + case OperatorType::kArgMin: + ProcessArgMinMaxOperator( + model, static_cast(op)); break; case OperatorType::kUnsupported: break; diff --git a/tensorflow/contrib/lite/toco/graph_transformations/quantization_util.cc b/tensorflow/contrib/lite/toco/graph_transformations/quantization_util.cc index d74cad9a626b3a472e2740d6bdaaaf7aab5bd484..44733391f5a1d9ebf9a24f4f31b425a35354e1fc 100644 --- a/tensorflow/contrib/lite/toco/graph_transformations/quantization_util.cc +++ b/tensorflow/contrib/lite/toco/graph_transformations/quantization_util.cc @@ -74,46 +74,54 @@ ArrayDataType GetQuantizedDataType(const Array& array, } } -void GetQuantizationParams(ArrayDataType data_type, const MinMax& minmax, - QuantizationParams* quantization_params) { - switch (data_type) { +template +void ChooseQuantizationParamsForArrayAndQuantizedDataType( + const Array& array, QuantizationParams* quantization_params) { + *quantization_params = ::tflite::ChooseQuantizationParams>( + array.minmax->min, array.minmax->max, array.narrow_range); +} + +void ChooseQuantizationParamsForArrayAndQuantizedDataType( + const Array& array, ArrayDataType quantized_data_type, + QuantizationParams* quantization_params) { + switch (quantized_data_type) { case ArrayDataType::kInt8: - GetQuantizationParamsFromMinMax( - minmax, quantization_params); + ChooseQuantizationParamsForArrayAndQuantizedDataType< + ArrayDataType::kInt8>(array, quantization_params); break; case ArrayDataType::kUint8: - GetQuantizationParamsFromMinMax( - minmax, quantization_params); + ChooseQuantizationParamsForArrayAndQuantizedDataType< + ArrayDataType::kUint8>(array, quantization_params); break; case ArrayDataType::kInt16: - GetQuantizationParamsFromMinMax( - minmax, quantization_params); + ChooseQuantizationParamsForArrayAndQuantizedDataType< + ArrayDataType::kInt16>(array, quantization_params); break; case ArrayDataType::kUint16: - GetQuantizationParamsFromMinMax( - minmax, quantization_params); + ChooseQuantizationParamsForArrayAndQuantizedDataType< + ArrayDataType::kUint16>(array, quantization_params); break; case ArrayDataType::kInt32: - GetQuantizationParamsFromMinMax( - minmax, quantization_params); + ChooseQuantizationParamsForArrayAndQuantizedDataType< + ArrayDataType::kInt32>(array, quantization_params); break; case ArrayDataType::kUint32: - GetQuantizationParamsFromMinMax( - minmax, quantization_params); + ChooseQuantizationParamsForArrayAndQuantizedDataType< + ArrayDataType::kUint32>(array, quantization_params); break; case ArrayDataType::kInt64: - GetQuantizationParamsFromMinMax( - minmax, quantization_params); + ChooseQuantizationParamsForArrayAndQuantizedDataType< + ArrayDataType::kInt64>(array, quantization_params); break; case ArrayDataType::kUint64: - GetQuantizationParamsFromMinMax( - minmax, quantization_params); + ChooseQuantizationParamsForArrayAndQuantizedDataType< + ArrayDataType::kUint64>(array, quantization_params); break; case ArrayDataType::kFloat: case ArrayDataType::kNone: default: LOG(FATAL) << "Unhandled final quantization type " - << static_cast(data_type); + << static_cast(quantized_data_type); } } @@ -121,8 +129,8 @@ namespace { template std::unique_ptr QuantizeBuffer( - const GenericBuffer& buffer, - const QuantizationParams& quantization_params) { + const Array& array, const QuantizationParams& quantization_params) { + const GenericBuffer& buffer = *array.buffer; const auto inverse_scale = 1. / quantization_params.scale; CHECK(buffer.type == ArrayDataType::kFloat); const auto& float_buffer = @@ -140,8 +148,15 @@ std::unique_ptr QuantizeBuffer( } else { scaled_val = quantization_params.zero_point + inverse_scale * src_val; } - quantized_buffer->data[i] = - tflite::SafeCast>(std::round(scaled_val)); + auto integer_val = tflite::SafeCast>(std::round(scaled_val)); + // In addition to its effect on the choice of quantization params upstream + // of here, narrow_range also means nudge the min quantized value by +1, + // so e.g. uint8 values get constrained to [1, 255]. + if (integer_val == std::numeric_limits>::min() && + array.narrow_range) { + integer_val++; + } + quantized_buffer->data[i] = integer_val; } return std::unique_ptr(quantized_buffer); } @@ -155,7 +170,7 @@ void QuantizeArray(GraphTransformation* transformation, Model* model, CHECK(!array.quantization_params); array.GetOrCreateQuantizationParams() = quantization_params; if (array.buffer) { - array.buffer = QuantizeBuffer(*array.buffer, quantization_params); + array.buffer = QuantizeBuffer(array, quantization_params); } array.data_type = A; array.final_data_type = A; @@ -210,8 +225,8 @@ bool IsArrayQuantizedRangeSubset(GraphTransformation* transformation, } else { // Work around cases where we are asking for this prior to the Quantize // transformation having added the quantization_params. - GetQuantizationParams(quantized_data_type, *array.minmax, - &quantization_params); + ChooseQuantizationParamsForArrayAndQuantizedDataType( + array, quantized_data_type, &quantization_params); transformation->AddMessageF( "No quantization params - infering from data type %s with minmax " "%g,%g as zero_point=%g, scale=%g", diff --git a/tensorflow/contrib/lite/toco/graph_transformations/quantization_util.h b/tensorflow/contrib/lite/toco/graph_transformations/quantization_util.h index 79a2ce7e50887b4608b278471da0e5e63b5673e3..cf093c6f17b45839156dae0d06ca2fc7e5e2f3c6 100644 --- a/tensorflow/contrib/lite/toco/graph_transformations/quantization_util.h +++ b/tensorflow/contrib/lite/toco/graph_transformations/quantization_util.h @@ -38,21 +38,11 @@ bool GetQuantizedDataTypeNumericalRange(ArrayDataType data_type, ArrayDataType GetQuantizedDataType(const Array& array, ArrayDataType default_type); -// Returns the quantization params for the array with the given data type and -// minmax. -void GetQuantizationParams(ArrayDataType data_type, const MinMax& minmax, - QuantizationParams* quantization_params); - -// Returns the quantization params for the data type and minmax values. -template -void GetQuantizationParamsFromMinMax(const MinMax& minmax, - QuantizationParams* quantization_params) { - using Integer = DataType; - const double rmin = minmax.min; - const double rmax = minmax.max; - *quantization_params = - ::tflite::ChooseQuantizationParams(rmin, rmax); -} +// Chooses the quantization params for a given array and a given target +// quantized data type (which may not be the array's current data type). +void ChooseQuantizationParamsForArrayAndQuantizedDataType( + const Array& array, ArrayDataType quantized_data_type, + QuantizationParams* quantization_params); // Quantizes an array by setting its data type and (if constant) quantizing // all values in the array. diff --git a/tensorflow/contrib/lite/toco/graph_transformations/quantize.cc b/tensorflow/contrib/lite/toco/graph_transformations/quantize.cc index 1c61b8cb36ef9968c55f64c023fca8361162beb1..5be275747906fc8b30194f2e3d062db45fbb088f 100644 --- a/tensorflow/contrib/lite/toco/graph_transformations/quantize.cc +++ b/tensorflow/contrib/lite/toco/graph_transformations/quantize.cc @@ -59,7 +59,8 @@ bool SupportsQuantization(const Operator& op) { type == OperatorType::kGreater || type == OperatorType::kGreaterEqual || type == OperatorType::kLess || type == OperatorType::kLessEqual || type == OperatorType::kSelect || - type == OperatorType::kArgMax; + type == OperatorType::kArgMax || type == OperatorType::kRelu || + type == OperatorType::kRelu1 || type == OperatorType::kRelu6; } const MinMax& GetOrComputeMinMax(Model* model, const string& array_name) { @@ -211,13 +212,15 @@ bool ChooseQuantizationForOperatorInput( if (op.type == OperatorType::kLstmCell) { if (input_index == LstmCellOperator::PREV_STATE_INPUT) { *quantized_data_type = ArrayDataType::kInt16; - GetQuantizationParams(*quantized_data_type, minmax, quantization_params); + ChooseQuantizationParamsForArrayAndQuantizedDataType( + array, *quantized_data_type, quantization_params); return true; } } *quantized_data_type = GetQuantizedDataType(array, ArrayDataType::kUint8); - GetQuantizationParams(*quantized_data_type, minmax, quantization_params); + ChooseQuantizationParamsForArrayAndQuantizedDataType( + array, *quantized_data_type, quantization_params); transformation->AddMessageF( "For input array %s with min=%g, max=%g, chose to quantize as %s (f=%s) " "with zero_point=%d, scale=%g", @@ -325,12 +328,13 @@ bool ChooseQuantizationForOperatorOutput( output, OperatorTypeName(op.type)); return true; } - if ((op.type == OperatorType::kDepthToSpace) || - (op.type == OperatorType::kSpaceToDepth) || - (op.type == OperatorType::kReshape) || - (op.type == OperatorType::kSplit) || - (op.type == OperatorType::kConcatenation && - model->flags.change_concat_input_ranges())) { + if ((op.type == OperatorType::kConcatenation && + model->flags.change_concat_input_ranges()) || + op.type == OperatorType::kDepthToSpace || + op.type == OperatorType::kSpaceToDepth || + op.type == OperatorType::kReshape || op.type == OperatorType::kSplit || + op.type == OperatorType::kRelu || op.type == OperatorType::kRelu1 || + op.type == OperatorType::kRelu6) { int data_input_index = 0; if (op.type == OperatorType::kSplit) { data_input_index = 1; @@ -356,12 +360,14 @@ bool ChooseQuantizationForOperatorOutput( if (output_index == LstmCellOperator::STATE_OUTPUT || output_index == LstmCellOperator::ACTIV_TEMP) { *quantized_data_type = ArrayDataType::kInt16; - GetQuantizationParams(*quantized_data_type, minmax, quantization_params); + ChooseQuantizationParamsForArrayAndQuantizedDataType( + array, *quantized_data_type, quantization_params); return true; } } *quantized_data_type = GetQuantizedDataType(array, ArrayDataType::kUint8); - GetQuantizationParams(*quantized_data_type, minmax, quantization_params); + ChooseQuantizationParamsForArrayAndQuantizedDataType( + array, *quantized_data_type, quantization_params); transformation->AddMessageF( "For output array %s with min=%g, max=%g" ", chose to quantize as %s with zero_point=%d" @@ -505,36 +511,47 @@ bool Quantize::Run(Model* model, std::size_t op_index) { // Check if the output of that Dequantize op was not used by any // other operator. We will then erase that Dequantize op. if (!CountOpsWithInput(*model, dequantize_op->outputs[0])) { - // If any of the model's output_arrays was pointing to the - // Dequantize op's output, let it point to the Dequantize op's - // input instead. - for (int i = 0; i < model->flags.output_arrays_size(); i++) { - if (model->flags.output_arrays(i) == dequantize_op->outputs[0]) { - // TODO(b/78013785): never rename output arrays. - if (IsInputArray(*model, dequantize_op->inputs[0])) { - // The op input is an input array and the output is an output - // array and we can't have an array be both. Insert a copy - // op to ensure the two arrays stay separate. - AddMessageF( - "Tried to rename output array %d while removing dequant " - "op %s but array is also an input; inserting copy %s " - "-> %s", - i, LogName(*dequantize_op), model->flags.output_arrays(i), - dequantize_op->inputs[0]); - InsertCopyOperator(model, dequantize_op->inputs[0], - dequantize_op->outputs[0]); - } else { - // Op output is strictly used as an output array, so we can - // just rename the array and directly bypass the op. - AddMessageF( - "Renaming output array %d after removing dequant op %s: " - "%s -> %s", - i, LogName(*dequantize_op), model->flags.output_arrays(i), - dequantize_op->inputs[0]); - model->flags.set_output_arrays(i, dequantize_op->inputs[0]); - model->EraseArray(dequantize_op->outputs[0]); + if (IsDiscardableArray(*model, dequantize_op->outputs[0])) { + // Usual case: we can just discard the dequantize output. + model->EraseArray(dequantize_op->outputs[0]); + } else { + // The dequantize output is not discardable. Special care needed. + // If any of the model's output_arrays was pointing to the + // Dequantize op's output, let it point to the Dequantize op's + // input instead. + for (int i = 0; i < model->flags.output_arrays_size(); i++) { + if (model->flags.output_arrays(i) == + dequantize_op->outputs[0]) { + // TODO(b/78013785): never rename output arrays. + if (IsInputArray(*model, dequantize_op->inputs[0])) { + // The op input is an input array and the output is an + // output array and we can't have an array be both. Insert a + // copy op to ensure the two arrays stay separate. + AddMessageF( + "Tried to rename output array %d while removing " + "dequant " + "op %s but array is also an input; inserting copy %s " + "-> %s", + i, LogName(*dequantize_op), + model->flags.output_arrays(i), + dequantize_op->inputs[0]); + InsertCopyOperator(model, dequantize_op->inputs[0], + dequantize_op->outputs[0]); + } else { + // Op output is strictly used as an output array, so we can + // just rename the array and directly bypass the op. + AddMessageF( + "Renaming output array %d after removing dequant op " + "%s: " + "%s -> %s", + i, LogName(*dequantize_op), + model->flags.output_arrays(i), + dequantize_op->inputs[0]); + model->flags.set_output_arrays(i, dequantize_op->inputs[0]); + model->EraseArray(dequantize_op->outputs[0]); + } + break; } - break; } } model->operators.erase(dequantize_it); diff --git a/tensorflow/contrib/lite/toco/graph_transformations/quantize_weights.cc b/tensorflow/contrib/lite/toco/graph_transformations/quantize_weights.cc index 88ea0945e7dd15ba325d34ea3fdbf34ff7d91381..7a8515f6d12f96d464ea0764907f9cc2a487d3e7 100644 --- a/tensorflow/contrib/lite/toco/graph_transformations/quantize_weights.cc +++ b/tensorflow/contrib/lite/toco/graph_transformations/quantize_weights.cc @@ -36,10 +36,8 @@ void GetQuantizationParamsFromArray(const Array& array, const std::vector& float_vals = array.GetBuffer().data; auto minmax = std::minmax_element(float_vals.begin(), float_vals.end()); - MinMax toco_minmax; - toco_minmax.min = *minmax.first; - toco_minmax.max = *minmax.second; - GetQuantizationParams(ArrayDataType::kUint8, toco_minmax, params); + *params = tflite::ChooseQuantizationParams( + *minmax.first, *minmax.second, array.narrow_range); } } // namespace diff --git a/tensorflow/contrib/lite/toco/graph_transformations/read_array_minmax_and_narrow_range_from_fake_quant.cc b/tensorflow/contrib/lite/toco/graph_transformations/read_array_minmax_and_narrow_range_from_fake_quant.cc new file mode 100644 index 0000000000000000000000000000000000000000..5b41c49bfaff245d599d26989e4ed3f9b0d582cf --- /dev/null +++ b/tensorflow/contrib/lite/toco/graph_transformations/read_array_minmax_and_narrow_range_from_fake_quant.cc @@ -0,0 +1,78 @@ +/* Copyright 2017 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#include +#include +#include +#include +#include + +#include "tensorflow/contrib/lite/toco/graph_transformations/graph_transformations.h" +#include "tensorflow/contrib/lite/toco/model.h" +#include "tensorflow/contrib/lite/toco/tooling_util.h" +#include "tensorflow/core/platform/logging.h" + +namespace toco { + +namespace { + +bool ApplyAttrsToArray(GraphTransformation* transformation, Model* model, + const FakeQuantOperator& fq_op, + const string& array_name) { + bool changed = false; + auto& annotated_array = model->GetArray(array_name); + if (!annotated_array.minmax) { + const MinMax& minmax = *fq_op.minmax; + annotated_array.GetOrCreateMinMax() = minmax; + transformation->AddMessageF( + "Read min/max annotation for array %s: min=%g, max=%g", array_name, + minmax.min, minmax.max); + changed = true; + } + if (fq_op.narrow_range && !annotated_array.narrow_range) { + annotated_array.narrow_range = true; + transformation->AddMessageF("Read narrow_range annotation for array %s", + array_name); + changed = true; + } + return changed; +} + +} // end namespace + +bool ReadArrayMinmaxAndNarrowRangeFromFakeQuant::Run(Model* model, + std::size_t op_index) { + const auto fakequant_it = model->operators.begin() + op_index; + auto* fakequant_base_op = fakequant_it->get(); + if (fakequant_base_op->type != OperatorType::kFakeQuant) { + return false; + } + auto* fq_op = static_cast(fakequant_base_op); + + if (!fq_op->minmax) { + // Need to be resolved first by ResolveFakeQuantArgsFromVars. + return false; + } + + // At this point, this FakeQuantOperator should have a MinMax + // attached to it, and should only have 1 input (it should not have + // 2nd and 3rd input arrays giving min and max anymore). + CHECK(fq_op->minmax); + CHECK_EQ(1, fq_op->inputs.size()); + + return ApplyAttrsToArray(this, model, *fq_op, fq_op->inputs[0]) || + ApplyAttrsToArray(this, model, *fq_op, fq_op->outputs[0]); +} + +} // namespace toco diff --git a/tensorflow/contrib/lite/toco/graph_transformations/read_fake_quant_min_max.cc b/tensorflow/contrib/lite/toco/graph_transformations/read_fake_quant_min_max.cc deleted file mode 100644 index bdcca5b7caf61a62203debaa32c4d7a9b2eb43fa..0000000000000000000000000000000000000000 --- a/tensorflow/contrib/lite/toco/graph_transformations/read_fake_quant_min_max.cc +++ /dev/null @@ -1,112 +0,0 @@ -/* Copyright 2017 The TensorFlow Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. -==============================================================================*/ -#include -#include -#include -#include -#include - -#include "tensorflow/contrib/lite/toco/graph_transformations/graph_transformations.h" -#include "tensorflow/contrib/lite/toco/model.h" -#include "tensorflow/contrib/lite/toco/tooling_util.h" -#include "tensorflow/core/platform/logging.h" - -namespace toco { - -namespace { - -bool ApplyMinMaxToArray(GraphTransformation* transformation, Model* model, - const MinMax& minmax, const string& array_name) { - auto& annotated_array = model->GetArray(array_name); - if (annotated_array.minmax) { - return false; - } - annotated_array.GetOrCreateMinMax() = minmax; - transformation->AddMessageF( - "Read min/max annotation for array %s: min=%g, max=%g", array_name, - minmax.min, minmax.max); - return true; -} - -} // end namespace - -bool ReadFakeQuantMinMax::Run(Model* model, std::size_t op_index) { - const auto fakequant_it = model->operators.begin() + op_index; - auto* fakequant_base_op = fakequant_it->get(); - if (fakequant_base_op->type != OperatorType::kFakeQuant) { - return false; - } - auto* fakequant_op = static_cast(fakequant_base_op); - - bool changed = false; - - if (!fakequant_op->minmax) { - CHECK_EQ(fakequant_op->inputs.size(), 3); - // We need to yield until the min and max parameters have been - // resolved to constant arrays. - for (int i = 1; i <= 2; i++) { - if (!IsConstantParameterArray(*model, fakequant_op->inputs[1])) { - return false; - } - } - - // Obtain the final min/max values - const auto& min_array = model->GetArray(fakequant_op->inputs[1]); - const auto& max_array = model->GetArray(fakequant_op->inputs[2]); - CHECK_EQ(RequiredBufferSizeForShape(min_array.shape()), 1); - CHECK_EQ(RequiredBufferSizeForShape(max_array.shape()), 1); - fakequant_op->minmax.reset(new MinMax); - MinMax& minmax = *fakequant_op->minmax; - minmax.min = min_array.GetBuffer().data[0]; - minmax.max = max_array.GetBuffer().data[0]; - // We always want [min, max] to contain 0. - if (minmax.min > 0 || minmax.max < 0) { - LOG(ERROR) << "For " << LogName(*fakequant_op) << " the MinMax range " - << "[" << minmax.min << ", " << minmax.max - << "] does not contain 0. " - << "Proceeding by tweaking it to contain 0, which will result " - "in poor accuracy."; - } - minmax.min = std::min(minmax.min, 0.); - minmax.max = std::max(minmax.max, 0.); - - // We won't use the input arrays that provided these min and max - // values, anymore. Delete them unless they are used by something - // else. - for (int i = 1; i <= 2; i++) { - if (CountOpsWithInput(*model, fakequant_op->inputs[i]) == 1) { - model->EraseArray(fakequant_op->inputs[i]); - } - } - fakequant_op->inputs.resize(1); - changed = true; - } - - // At this point, this FakeQuantOperator should have a MinMax - // attached to it, and should only have 1 input (it should not have - // 2nd and 3rd input arrays giving min and max anymore). - CHECK(fakequant_op->minmax); - CHECK_EQ(1, fakequant_op->inputs.size()); - - const MinMax& minmax = *fakequant_op->minmax; - - // Record the MinMax info on the input and output arrays - changed |= ApplyMinMaxToArray(this, model, minmax, fakequant_op->inputs[0]); - changed |= ApplyMinMaxToArray(this, model, minmax, fakequant_op->outputs[0]); - - return changed; -} - -} // namespace toco diff --git a/tensorflow/contrib/lite/toco/graph_transformations/resolve_batch_to_space_nd_attributes.cc b/tensorflow/contrib/lite/toco/graph_transformations/resolve_batch_to_space_nd_attributes.cc index a06919e228dc2084f8943a714a0ca111d013c159..b8b35161d77e5b6dd8c30e03959dba3c60d1d56c 100644 --- a/tensorflow/contrib/lite/toco/graph_transformations/resolve_batch_to_space_nd_attributes.cc +++ b/tensorflow/contrib/lite/toco/graph_transformations/resolve_batch_to_space_nd_attributes.cc @@ -50,7 +50,7 @@ bool ResolveBatchToSpaceNDAttributes::Run(Model* model, std::size_t op_index) { // will delete this op. return false; } - std::vector crops_buffer = + const std::vector& crops_buffer = crops_array.GetBuffer().data; for (int i = 0; i < crops_dims[0]; ++i) { op->before_crops.push_back(crops_buffer[i * 2]); @@ -62,7 +62,7 @@ bool ResolveBatchToSpaceNDAttributes::Run(Model* model, std::size_t op_index) { if (!block_shape_array.has_shape()) return false; const std::vector& block_shape_dims = block_shape_array.shape().dims(); CHECK_EQ(block_shape_dims.size(), 1); - std::vector block_shape_buffer = + const std::vector& block_shape_buffer = block_shape_array.GetBuffer().data; for (int i = 0; i < block_shape_dims[0]; ++i) { op->block_shape.push_back(block_shape_buffer[i]); diff --git a/tensorflow/contrib/lite/toco/graph_transformations/resolve_constant_fake_quant.cc b/tensorflow/contrib/lite/toco/graph_transformations/resolve_constant_fake_quant.cc index efb7bb218421dd045e3e8e2a38b9c70989f222e1..058f314b338aeeab94cb11fb8c1163427b559d3e 100644 --- a/tensorflow/contrib/lite/toco/graph_transformations/resolve_constant_fake_quant.cc +++ b/tensorflow/contrib/lite/toco/graph_transformations/resolve_constant_fake_quant.cc @@ -25,6 +25,37 @@ limitations under the License. namespace toco { +template +void GetBoundsForQuantizedDataType(double* min, double* max) { + using limits = std::numeric_limits>; + *min = limits::min(); + *max = limits::max(); +} + +void GetBoundsForQuantizedDataType(ArrayDataType quantized_data_type, + double* min, double* max) { + switch (quantized_data_type) { + case ArrayDataType::kUint8: + return GetBoundsForQuantizedDataType(min, max); + case ArrayDataType::kInt8: + return GetBoundsForQuantizedDataType(min, max); + case ArrayDataType::kUint16: + return GetBoundsForQuantizedDataType(min, max); + case ArrayDataType::kInt16: + return GetBoundsForQuantizedDataType(min, max); + case ArrayDataType::kUint32: + return GetBoundsForQuantizedDataType(min, max); + case ArrayDataType::kInt32: + return GetBoundsForQuantizedDataType(min, max); + case ArrayDataType::kUint64: + return GetBoundsForQuantizedDataType(min, max); + case ArrayDataType::kInt64: + return GetBoundsForQuantizedDataType(min, max); + default: + LOG(FATAL) << "unhandled quantized data type"; + } +} + bool ResolveConstantFakeQuant::Run(Model* model, std::size_t op_index) { const auto fakequant_it = model->operators.begin() + op_index; const auto* fakequant_base_op = fakequant_it->get(); @@ -76,14 +107,21 @@ bool ResolveConstantFakeQuant::Run(Model* model, std::size_t op_index) { const int size = input_buffer.data.size(); output_buffer.data.resize(size); QuantizationParams qparams; - GetQuantizationParamsFromMinMax(*fakequant_op->minmax, - &qparams); + ChooseQuantizationParamsForArrayAndQuantizedDataType( + output_array, quantized_data_type, &qparams); + double quantized_min, quantized_max; + GetBoundsForQuantizedDataType(quantized_data_type, &quantized_min, + &quantized_max); + if (fakequant_op->narrow_range) { + quantized_min++; + } + for (int i = 0; i < size; i++) { const double src_val = input_buffer.data[i]; const double unclamped_quantized_val = std::round(qparams.zero_point + src_val / qparams.scale); - const double quantized_val = - std::min(255., std::max(0., unclamped_quantized_val)); + const double quantized_val = std::min( + quantized_max, std::max(quantized_min, unclamped_quantized_val)); const double dst_val = qparams.scale * (quantized_val - qparams.zero_point); output_buffer.data[i] = dst_val; } diff --git a/tensorflow/contrib/lite/toco/graph_transformations/resolve_constant_strided_slice.cc b/tensorflow/contrib/lite/toco/graph_transformations/resolve_constant_strided_slice.cc index 6ee231465fae5127e3769bd6b9060ea60d59eb2c..9d8bd4fc39344a4ea1fa4942a2a99ec535b5bee8 100644 --- a/tensorflow/contrib/lite/toco/graph_transformations/resolve_constant_strided_slice.cc +++ b/tensorflow/contrib/lite/toco/graph_transformations/resolve_constant_strided_slice.cc @@ -38,6 +38,7 @@ void StridedSlice(StridedSliceOperator const& op, Array const& input_array, CHECK_EQ(op.new_axis_mask, 0); int num_input_axes = op.start_indices.size(); + CHECK_EQ(num_input_axes, op.start_indices.size()); CHECK_EQ(num_input_axes, op.stop_indices.size()); CHECK_EQ(num_input_axes, op.strides.size()); @@ -49,11 +50,16 @@ void StridedSlice(StridedSliceOperator const& op, Array const& input_array, // Initialize source coordinate Shape const& input_shape = input_array.shape(); Buffer const& input_buffer = input_array.GetBuffer(); - std::vector src_coord(op.start_indices.size()); + std::vector src_coord(num_input_axes); + std::vector stop_for_axis(num_input_axes); for (int axis = 0; axis < num_input_axes; axis++) { - src_coord[axis] = tflite::strided_slice::StartForAxis( + int start = tflite::strided_slice::StartForAxis( op.begin_mask, op.start_indices, op.strides, input_shape.dims().data(), axis); + src_coord[axis] = start; + stop_for_axis[axis] = tflite::strided_slice::StopForAxis( + op.end_mask, op.shrink_axis_mask, op.stop_indices, op.strides, + input_shape.dims().data(), axis, start); } // In order to handle any number (N) of dimensions, we copy elements one by @@ -76,9 +82,7 @@ void StridedSlice(StridedSliceOperator const& op, Array const& input_array, } // Check if we've overflowed. - int stop = tflite::strided_slice::StopForAxis( - op.end_mask, op.stop_indices, op.strides, input_shape.dims().data(), - axis); + int stop = stop_for_axis[axis]; if (tflite::strided_slice::LoopCondition(src_coord[axis], stop, stride)) { // Reset axis and set carry src_coord[axis] = tflite::strided_slice::StartForAxis( diff --git a/tensorflow/contrib/lite/toco/graph_transformations/resolve_fake_quant_args_from_vars.cc b/tensorflow/contrib/lite/toco/graph_transformations/resolve_fake_quant_args_from_vars.cc new file mode 100644 index 0000000000000000000000000000000000000000..0dda1fd0b35fb0cdc3c605360df5126c52c05403 --- /dev/null +++ b/tensorflow/contrib/lite/toco/graph_transformations/resolve_fake_quant_args_from_vars.cc @@ -0,0 +1,80 @@ +/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ +#include +#include +#include +#include +#include + +#include "tensorflow/contrib/lite/toco/graph_transformations/graph_transformations.h" +#include "tensorflow/contrib/lite/toco/model.h" +#include "tensorflow/contrib/lite/toco/tooling_util.h" +#include "tensorflow/core/platform/logging.h" + +namespace toco { + +bool ResolveFakeQuantArgsFromVars::Run(Model* model, std::size_t op_index) { + const auto fakequant_it = model->operators.begin() + op_index; + auto* fakequant_base_op = fakequant_it->get(); + if (fakequant_base_op->type != OperatorType::kFakeQuant) { + return false; + } + auto* fakequant_op = static_cast(fakequant_base_op); + + if (fakequant_op->minmax) { + // Already resolved. + return false; + } + + CHECK_EQ(fakequant_op->inputs.size(), 3); + // We need to yield until the min and max parameters have been + // resolved to constant arrays. + for (int i = 1; i <= 2; i++) { + if (!IsConstantParameterArray(*model, fakequant_op->inputs[i])) { + return false; + } + } + + // Obtain the final min/max values + const auto& min_array = model->GetArray(fakequant_op->inputs[1]); + const auto& max_array = model->GetArray(fakequant_op->inputs[2]); + CHECK_EQ(RequiredBufferSizeForShape(min_array.shape()), 1); + CHECK_EQ(RequiredBufferSizeForShape(max_array.shape()), 1); + fakequant_op->minmax.reset(new MinMax); + MinMax& minmax = *fakequant_op->minmax; + minmax.min = min_array.GetBuffer().data[0]; + minmax.max = max_array.GetBuffer().data[0]; + // We always want [min, max] to contain 0. + if (minmax.min > 0 || minmax.max < 0) { + LOG(ERROR) << "For " << LogName(*fakequant_op) << " the MinMax range " + << "[" << minmax.min << ", " << minmax.max + << "] does not contain 0. " + << "Proceeding by tweaking it to contain 0, which will result " + "in poor accuracy."; + } + minmax.min = std::min(minmax.min, 0.); + minmax.max = std::max(minmax.max, 0.); + + // We won't use the input arrays that provided these min and max + // values, anymore. Delete them unless they are used by something + // else. + for (int i = 1; i <= 2; i++) { + DeleteArrayIfUsedOnce(fakequant_op->inputs[i], model); + } + fakequant_op->inputs.resize(1); + return true; +} + +} // namespace toco diff --git a/tensorflow/contrib/lite/toco/graph_transformations/resolve_reorder_axes.cc b/tensorflow/contrib/lite/toco/graph_transformations/resolve_reorder_axes.cc index bc70db0bd8c26319fa140616de96452260a01058..8266e2c205b65e9d8a969643f102bb852be9125b 100644 --- a/tensorflow/contrib/lite/toco/graph_transformations/resolve_reorder_axes.cc +++ b/tensorflow/contrib/lite/toco/graph_transformations/resolve_reorder_axes.cc @@ -51,11 +51,12 @@ void ReorderAxes(AxesOrder input_axes_order, AxesOrder output_axes_order, } bool ResolveReorderAxes::Run(Model* model, std::size_t op_index) { - auto reorder_it = model->operators.begin() + op_index; - auto* reorder_op = static_cast(reorder_it->get()); - if (reorder_op->type != OperatorType::kReorderAxes) { + auto it = model->operators.begin() + op_index; + auto* op = it->get(); + if (op->type != OperatorType::kReorderAxes) { return false; } + auto* reorder_op = static_cast(op); const auto& input_array_name = reorder_op->inputs[0]; const auto& output_array_name = reorder_op->outputs[0]; auto& input_array = model->GetArray(input_array_name); @@ -95,7 +96,7 @@ bool ResolveReorderAxes::Run(Model* model, std::size_t op_index) { // Remove the op and output array. model->EraseArray(output_array_name); - model->operators.erase(reorder_it); + model->operators.erase(it); return true; } diff --git a/tensorflow/contrib/lite/toco/graph_transformations/resolve_space_to_batch_nd_attributes.cc b/tensorflow/contrib/lite/toco/graph_transformations/resolve_space_to_batch_nd_attributes.cc index dad6aceccfd201b3db07c29c99a8c6ef75bb89a1..fab50bec1fc5ec50cecba53845457931ed59c0b8 100644 --- a/tensorflow/contrib/lite/toco/graph_transformations/resolve_space_to_batch_nd_attributes.cc +++ b/tensorflow/contrib/lite/toco/graph_transformations/resolve_space_to_batch_nd_attributes.cc @@ -53,7 +53,7 @@ bool ResolveSpaceToBatchNDAttributes::Run(Model* model, std::size_t op_index) { // will delete this op. return false; } - std::vector paddings_buffer = + const std::vector& paddings_buffer = paddings_array.GetBuffer().data; for (int i = 0; i < paddings_dims[0]; ++i) { op->before_paddings.push_back(paddings_buffer[i * 2]); @@ -66,7 +66,7 @@ bool ResolveSpaceToBatchNDAttributes::Run(Model* model, std::size_t op_index) { if (!block_shape_array.has_shape()) return false; const std::vector& block_shape_dims = block_shape_array.shape().dims(); CHECK_EQ(block_shape_dims.size(), 1); - std::vector block_shape_buffer = + const std::vector& block_shape_buffer = block_shape_array.GetBuffer().data; for (int i = 0; i < block_shape_dims[0]; ++i) { op->block_shape.push_back(block_shape_buffer[i]); diff --git a/tensorflow/contrib/lite/toco/graph_transformations/resolve_tensorflow_matmul.cc b/tensorflow/contrib/lite/toco/graph_transformations/resolve_tensorflow_matmul.cc index d496f5ae5eeeca5063e23b25498b0ac450e9f946..fcf30bd34725fc59bb819e75deda0dadf330f372 100644 --- a/tensorflow/contrib/lite/toco/graph_transformations/resolve_tensorflow_matmul.cc +++ b/tensorflow/contrib/lite/toco/graph_transformations/resolve_tensorflow_matmul.cc @@ -32,21 +32,34 @@ bool ResolveTensorFlowMatMul::Run(Model* model, std::size_t op_index) { const auto* matmul_op = static_cast(matmul_it->get()); + // Handling transposition of the first input here isn't very simple because + // we need to know the actual shape in order to produce a proper + // TransposeOperator. However, the second input is supposed to be 2D, so we + // can actually handle transposition of that matrix, which happens to be more + // common anyway. + CHECK(!matmul_op->transpose_a); + // Reorder the axes on the second input. TensorFlow uses row-major ordering // on both inputs, however this is inefficient for the FullyConnected // operator. We'll transpose the second input to be in column-major order now // and let constant propagation optimize things (if possible). - auto* transpose_op = new TransposeOperator; - transpose_op->inputs = { - matmul_op->inputs[1], - CreateInt32Array( - model, - AvailableArrayName(*model, matmul_op->inputs[1] + "/transpose/perm"), - {1, 0})}; - transpose_op->outputs = { - AvailableArrayName(*model, matmul_op->inputs[1] + "/transpose")}; - model->GetOrCreateArray(transpose_op->outputs[0]); - model->operators.emplace(matmul_it, transpose_op); + string input_lhs = matmul_op->inputs[0]; + string input_rhs = matmul_op->inputs[1]; + if (!matmul_op->transpose_b) { + auto* transpose_op = new TransposeOperator; + transpose_op->inputs = { + matmul_op->inputs[1], + CreateInt32Array(model, + AvailableArrayName( + *model, matmul_op->inputs[1] + "/transpose/perm"), + {1, 0})}; + transpose_op->outputs = { + AvailableArrayName(*model, matmul_op->inputs[1] + "/transpose")}; + model->GetOrCreateArray(transpose_op->outputs[0]); + model->operators.emplace(matmul_it, transpose_op); + + input_rhs = transpose_op->outputs[0]; + } // Refresh iterator. matmul_it = model->operators.begin(); @@ -57,9 +70,6 @@ bool ResolveTensorFlowMatMul::Run(Model* model, std::size_t op_index) { } DCHECK_EQ(matmul_it->get(), matmul_op); - string input_lhs = matmul_op->inputs[0]; - string input_rhs = transpose_op->outputs[0]; - // Construct the new FullyConnectedOperator. auto* fc_op = new FullyConnectedOperator; fc_op->outputs = matmul_op->outputs; diff --git a/tensorflow/contrib/lite/toco/import_tensorflow.cc b/tensorflow/contrib/lite/toco/import_tensorflow.cc index da7e5add7e15456448d87a5fcaa1ba0529eb24ec..ab3762e7eaa8b2ca9bcf657882f760211a735c39 100644 --- a/tensorflow/contrib/lite/toco/import_tensorflow.cc +++ b/tensorflow/contrib/lite/toco/import_tensorflow.cc @@ -378,7 +378,7 @@ tensorflow::Status ImportBoolArray(const TensorProto& input_tensor, for (int i = 0; i < input_flat_size; i++) { output_bool_data[i] = input_tensor.bool_val(0); } - } else if (input_tensor.int_val_size() == input_flat_size) { + } else if (input_tensor.bool_val_size() == input_flat_size) { for (int i = 0; i < input_tensor.bool_val_size(); i++) { output_bool_data[i] = input_tensor.bool_val(i); } @@ -755,6 +755,9 @@ tensorflow::Status ConvertFakeQuantWithMinMaxArgs( op->outputs.push_back(node.name()); // tf.fake_quant_with_min_max_args num_bits defaults to 8. op->num_bits = HasAttr(node, "num_bits") ? GetIntAttr(node, "num_bits") : 8; + if (HasAttr(node, "narrow_range")) { + op->narrow_range = GetBoolAttr(node, "narrow_range"); + } model->operators.emplace_back(op); return tensorflow::Status::OK(); } @@ -774,6 +777,9 @@ tensorflow::Status ConvertFakeQuantWithMinMaxVars( } op->outputs.push_back(node.name()); op->num_bits = HasAttr(node, "num_bits") ? GetIntAttr(node, "num_bits") : 8; + if (HasAttr(node, "narrow_range")) { + op->narrow_range = GetBoolAttr(node, "narrow_range"); + } model->operators.emplace_back(op); return tensorflow::Status::OK(); } @@ -984,18 +990,19 @@ tensorflow::Status ConvertMatMulOperator( Model* model) { TF_QCHECK_OK(CheckInputsCount(node, tf_import_flags, 2)); - // Transpose flags should be easy to support, but we don't have a - // GraphDef with them to test on at the moment. - CHECK_EQ(HasAttr(node, "transpose_a") && GetBoolAttr(node, "transpose_a"), - false); - CHECK_EQ(HasAttr(node, "transpose_b") && GetBoolAttr(node, "transpose_b"), - false); CHECK(!HasAttr(node, "adjoint_a") || (GetBoolAttr(node, "adjoint_a") == false)); CHECK(!HasAttr(node, "adjoint_b") || (GetBoolAttr(node, "adjoint_b") == false)); auto* matmul = new TensorFlowMatMulOperator; + if (HasAttr(node, "transpose_a")) { + matmul->transpose_a = GetBoolAttr(node, "transpose_a"); + } + if (HasAttr(node, "transpose_b")) { + matmul->transpose_b = GetBoolAttr(node, "transpose_b"); + } + matmul->inputs = {node.input(0), node.input(1)}; matmul->outputs = {node.name()}; model->operators.emplace_back(matmul); @@ -1229,10 +1236,11 @@ tensorflow::Status ConvertGatherOperator( return tensorflow::Status::OK(); } -tensorflow::Status ConvertArgMaxOperator( +template +tensorflow::Status ConvertArgMinMaxOperator( const NodeDef& node, const TensorFlowImportFlags& tf_import_flags, Model* model) { - CHECK_EQ(node.op(), "ArgMax"); + CHECK_EQ(node.op(), op_name); TF_QCHECK_OK(CheckInputsCount(node, tf_import_flags, 2)); const auto axis_data_type = HasAttr(node, "Tidx") ? GetDataTypeAttr(node, "Tidx") : DT_INT32; @@ -1241,7 +1249,7 @@ tensorflow::Status ConvertArgMaxOperator( : DT_INT64; CHECK(axis_data_type == DT_INT64 || axis_data_type == DT_INT32); CHECK(output_type == DT_INT64 || output_type == DT_INT32); - auto* op = new ArgMaxOperator; + auto* op = new Op; op->output_data_type = ConvertDataType(output_type); op->inputs.push_back(node.input(0)); op->inputs.push_back(node.input(1)); @@ -1832,12 +1840,16 @@ using ConverterType = tensorflow::Status (*)( Model* model); using ConverterMapType = std::unordered_map; +constexpr char kArgMax[] = "ArgMax"; +constexpr char kArgMin[] = "ArgMin"; + ConverterMapType GetTensorFlowNodeConverterMap() { return std::unordered_map({ {"Add", ConvertSimpleOperator}, {"AddN", ConvertSimpleOperator}, {"All", ConvertSimpleOperator}, - {"ArgMax", ConvertArgMaxOperator}, + {"ArgMax", ConvertArgMinMaxOperator}, + {"ArgMin", ConvertArgMinMaxOperator}, {"Assert", ConvertSimpleOperator}, {"AvgPool", ConvertAvgPoolOperator}, {"BatchMatMul", ConvertBatchMatMulOperator}, @@ -1899,6 +1911,7 @@ ConverterMapType GetTensorFlowNodeConverterMap() { {"ParallelDynamicStitch", ConvertDynamicStitchOperator}, {"Placeholder", ConvertPlaceholderOperator}, {"PlaceholderWithDefault", ConvertIdentityOperator}, + {"Pow", ConvertSimpleOperator}, {"RandomUniform", ConvertRandomUniform}, {"Range", ConvertRangeOperator}, {"Rank", ConvertSimpleOperator}, diff --git a/tensorflow/contrib/lite/toco/model.h b/tensorflow/contrib/lite/toco/model.h index 89cb061499f865a9cbb9e9dcb27580465c8a29fd..d06a30b6389fb7375ebc3c01b5b0cf85726fecf9 100644 --- a/tensorflow/contrib/lite/toco/model.h +++ b/tensorflow/contrib/lite/toco/model.h @@ -15,6 +15,7 @@ limitations under the License. #ifndef TENSORFLOW_CONTRIB_LITE_TOCO_MODEL_H_ #define TENSORFLOW_CONTRIB_LITE_TOCO_MODEL_H_ +#include #include #include #include @@ -138,6 +139,8 @@ enum class OperatorType : uint8 { kSparseToDense, kEqual, kNotEqual, + kPow, + kArgMin, }; // Helper to deal with TensorFlow arrays using a different ordering of @@ -160,15 +163,16 @@ enum class AxesOrder { // The type of the scalars in an array. // Note that the type does not by itself tell whether the values in the array -// are real (are literally interpreted as real numbers) or quantized (only -// acquire a meaning as real numbers in conjunction with QuantizationParams). +// are non-quantized (can be accessed directly) or quantized (must be +// interpreted in conjunction with QuantizationParams). // // In practice though: -// float values are always real +// float values are never quantized // uint8 values are always quantized -// int32 values are either real or quantized (depending on whether +// int32 values are sometimes quantized (depending on whether // QuantizationParams are present). -// other types are unused at the moment. +// complex values are never quantized +// other types are never quantized at the moment. // // kNone means that we don't know the data type yet, or that we don't care // because we'll be dropping the array anyway (e.g. some exotic array types @@ -186,7 +190,8 @@ enum class ArrayDataType : uint8 { kUint32, kInt64, kUint64, // 10 - kString + kString, + kComplex64, }; // Compile-time logic to map ArrayDataType to the corresponding C++ scalar type @@ -240,6 +245,10 @@ template <> struct DataTypeImpl { typedef string Type; }; +template <> +struct DataTypeImpl { + typedef std::complex Type; +}; template using DataType = typename DataTypeImpl::Type; @@ -782,6 +791,7 @@ struct FakeQuantOperator : Operator { FakeQuantOperator() : Operator(OperatorType::kFakeQuant) {} std::unique_ptr minmax; int num_bits = 8; + bool narrow_range = false; }; // Element-wise division operator. @@ -829,6 +839,8 @@ struct BatchMatMulOperator : Operator { // TensorFlow equivalent: MatMul struct TensorFlowMatMulOperator : Operator { TensorFlowMatMulOperator() : Operator(OperatorType::kMatMul) {} + bool transpose_a = false; + bool transpose_b = false; }; // Padding operator. Pads a tensor with zeros. @@ -1518,6 +1530,17 @@ struct ArgMaxOperator : Operator { ArrayDataType output_data_type = ArrayDataType::kInt64; }; +// ArgMin operator. It returns the index of the minimum value along axis. +// +// Inputs: +// inputs[0]: required: the input tensor +// +// TensorFlow equivalent: ArgMin +struct ArgMinOperator : Operator { + ArgMinOperator() : Operator(OperatorType::kArgMin) {} + ArrayDataType output_data_type = ArrayDataType::kInt64; +}; + // ResizeBilinear operator. It resizes input images with bilinear interpolation. // It does not support align_corners at the moment. // @@ -1637,6 +1660,17 @@ struct SparseToDenseOperator : Operator { bool validate_indices; }; +// Pow operator: +// +// Inputs: +// Inputs[0]: required: A tensor. +// Inputs[1]: required: A tensor. +// +// TensorFlow equivalent: Pow. +struct PowOperator : Operator { + PowOperator() : Operator(OperatorType::kPow) {} +}; + // Alloc's are used for transient arrays only. An Alloc specifies which interval // of the "transient_data" workspace buffer passed to inference functions, is to // be used for the transient array at hand. The 'start' and 'end' values are @@ -1821,6 +1855,40 @@ struct Array { // If this is non-null, then these quantization parameters are to be used // to assign a meaning as real numbers to the elements of this array. std::unique_ptr quantization_params; + // narrow_range is a detail of how toco handles FakeQuant operators with + // narrow_range, see + // https://www.tensorflow.org/api_docs/python/tf/fake_quant_with_min_max_vars + // + // For more context about what that is useful for, see the big comment in + // graph_transformations/ensure_uint8_weights_safe_for_fast_int8_kernels.cc + // + // The narrow_range flag applies only to quantized arrays, and changes + // their quantization in the following way when it is set to 'true': + // 1. The computation of {zero_point, scale} from {min, max} needs to be + // amended so that the real min value will get quantized to + // (min_quantized_value + 1) instead of just (min_quantized_value). + // E.g. for uint8 quantization, the real min value should get quantized to + // the uint8 value 1, not 0. + // 2. Quantized values should get clamped to the interval + // [min_quantized_value + 1, max_value]. Equivalently, the + // min_quantized_value should get nudged to (min_quantized_value + 1). + // The reason why 1. does not imply 2. is that real values may not belong to + // the stated [min, max] interval. Concretely, weights recorded at the last + // learning step may not fall in the [min, max] interval recorded over + // previous learning steps, as the values evolve across learning steps. + // + // Rationale why this is directly a field on Array: + // - This can't be just a field on FakeQuantOperator, because + // FakeQuantOperators are gone (DropFakeQuant) before we get to using that + // information (Quantize). We need a place to store that bit in the interim. + // - This can't be in QuantizationParams because we need to record this + // ahead of quantization, and QuantizationParams are only created during + // quantization. + // - This could be in MinMax, but that would be an abuse of what MinMax is + // about, and would break existing code that assumes that a MinMax is just + // a min and a max. Unlike MinMax which is agnostic as to the quantized + // data type, narrow_range refers to values in the quantized data type. + bool narrow_range = false; private: std::unique_ptr array_shape; diff --git a/tensorflow/contrib/lite/toco/model_cmdline_flags.cc b/tensorflow/contrib/lite/toco/model_cmdline_flags.cc index 4c9f1aa4b0274b5123bb3baa9b9fca1463bda4c3..06072d1fcb0612ed8193b3a0be1317923fe95bcc 100644 --- a/tensorflow/contrib/lite/toco/model_cmdline_flags.cc +++ b/tensorflow/contrib/lite/toco/model_cmdline_flags.cc @@ -74,10 +74,10 @@ bool ParseModelFlagsFromCommandLineFlags( "height, input array width, input array depth."), Flag("batch_size", parsed_flags.batch_size.bind(), parsed_flags.batch_size.default_value(), - "Batch size for the model. Replaces the first dimension of an " - "input size array if undefined. Use only with SavedModels when " - "--input_shapes flag is not specified. Always use --input_shapes " - "flag with frozen graphs."), + "Deprecated. Batch size for the model. Replaces the first dimension " + "of an input size array if undefined. Use only with SavedModels " + "when --input_shapes flag is not specified. Always use " + "--input_shapes flag with frozen graphs."), Flag("input_data_type", parsed_flags.input_data_type.bind(), parsed_flags.input_data_type.default_value(), "Deprecated: use --input_data_types instead. Input array type, if " diff --git a/tensorflow/contrib/lite/toco/tflite/export.cc b/tensorflow/contrib/lite/toco/tflite/export.cc index 19722468079a32b76f6952db6ca818da470a03ac..5ad307af14a0613188482ae17aed491dea06f984 100644 --- a/tensorflow/contrib/lite/toco/tflite/export.cc +++ b/tensorflow/contrib/lite/toco/tflite/export.cc @@ -336,17 +336,13 @@ void Export( auto op_codes = ExportOperatorCodes(model, ops_by_type, operators_map, &builder, &error_summary); - const string fake_quant_operation_name = "FAKE_QUANT"; - - if (error_summary.count(fake_quant_operation_name) != 0) { - LOG(ERROR) - << fake_quant_operation_name - << " operation was not converted. If running quantized make sure you " - "are passing --inference_type=QUANTIZED_UINT8 and values for " - "--std_values and --mean_values."; - // Remove the fake quant operation from the errors, since it shouldn't - // be provided a custom implementation. - error_summary.erase(fake_quant_operation_name); + for (const auto& op : model.operators) { + if (op->type == OperatorType::kFakeQuant) { + LOG(WARNING) << "FAKE_QUANT operation " << LogName(*op) + << " was not converted. If running quantized make sure you " + "are passing --inference_type=QUANTIZED_UINT8 and values " + "for --std_values and --mean_values."; + } } if (!allow_custom_ops && !error_summary.empty()) { // Remove ExpandDims and ReorderAxes from unimplemented list unless they diff --git a/tensorflow/contrib/lite/toco/tflite/import.cc b/tensorflow/contrib/lite/toco/tflite/import.cc index d1867bd4fa46a8a9dcd4c6abd4ef20b82c3854b4..1dd4915b31413e5afb04b45ee7c4893a2eded66d 100644 --- a/tensorflow/contrib/lite/toco/tflite/import.cc +++ b/tensorflow/contrib/lite/toco/tflite/import.cc @@ -221,6 +221,8 @@ std::unique_ptr Import(const ModelFlags& model_flags, model.get()); ImportIOTensors(*input_model, tensors_table, model.get()); + UndoWeightsShuffling(model.get()); + return model; } diff --git a/tensorflow/contrib/lite/toco/tflite/operator.cc b/tensorflow/contrib/lite/toco/tflite/operator.cc index 290a925c1ef68315473fcd06006114836cd08a4f..a791e60f91d9bca4be593115cd7ff34596500b4b 100644 --- a/tensorflow/contrib/lite/toco/tflite/operator.cc +++ b/tensorflow/contrib/lite/toco/tflite/operator.cc @@ -282,25 +282,31 @@ class DepthToSpace : public CustomOperator { int GetVersion(const Operator& op) const override { return 1; } }; -class FakeQuant : public CustomOperator { +class FakeQuant + : public BuiltinOperator { public: - using CustomOperator::CustomOperator; - void WriteOptions(const TocoOperator& op, - flexbuffers::Builder* fbb) const override { - fbb->Float("min", op.minmax->min); - fbb->Float("max", op.minmax->max); - fbb->Int("num_bits", op.num_bits); + using BuiltinOperator::BuiltinOperator; + flatbuffers::Offset WriteOptions( + const TocoOperator& op, + flatbuffers::FlatBufferBuilder* builder) const override { + return ::tflite::CreateFakeQuantOptions( + *builder, op.minmax->min, op.minmax->max, op.num_bits, op.narrow_range); } - void ReadOptions(const flexbuffers::Map& m, TocoOperator* op) const override { + void ReadOptions(const TfLiteOptions& options, + TocoOperator* op) const override { auto* minmax = new MinMax; - minmax->min = m["min"].AsFloat(); - minmax->max = m["max"].AsFloat(); + minmax->min = options.min(); + minmax->max = options.max(); op->minmax.reset(minmax); - const auto& num_bits = m["num_bits"]; - op->num_bits = num_bits.IsInt() ? num_bits.AsInt32() : 8; + op->num_bits = options.num_bits(); + op->narrow_range = options.narrow_range(); } - int GetVersion(const Operator& op) const override { return 1; } + int GetVersion(const Operator& op) const override { + const auto& fq_op = static_cast(op); + return fq_op.narrow_range ? 2 : 1; + } }; class FullyConnected @@ -314,16 +320,47 @@ class FullyConnected flatbuffers::FlatBufferBuilder* builder) const override { auto activation_function = ActivationFunction::Serialize(op.fused_activation_function); - return ::tflite::CreateFullyConnectedOptions(*builder, activation_function); + ::tflite::FullyConnectedOptionsWeightsFormat tflite_weights_format; + switch (op.weights_format) { + case FullyConnectedWeightsFormat::kDefault: + tflite_weights_format = + ::tflite::FullyConnectedOptionsWeightsFormat_DEFAULT; + break; + case FullyConnectedWeightsFormat::kShuffled4x16Int8: + tflite_weights_format = + ::tflite::FullyConnectedOptionsWeightsFormat_SHUFFLED4x16INT8; + break; + default: + LOG(ERROR) << "Unhandled FC weights format"; + tflite_weights_format = + ::tflite::FullyConnectedOptionsWeightsFormat_DEFAULT; + } + return ::tflite::CreateFullyConnectedOptions(*builder, activation_function, + tflite_weights_format); } void ReadOptions(const TfLiteOptions& options, TocoOperator* op) const override { op->fused_activation_function = ActivationFunction::Deserialize(options.fused_activation_function()); + switch (options.weights_format()) { + case ::tflite::FullyConnectedOptionsWeightsFormat_DEFAULT: + op->weights_format = FullyConnectedWeightsFormat::kDefault; + break; + case ::tflite::FullyConnectedOptionsWeightsFormat_SHUFFLED4x16INT8: + op->weights_format = FullyConnectedWeightsFormat::kShuffled4x16Int8; + break; + default: + LOG(ERROR) << "Unhandled FC weights format"; + op->weights_format = FullyConnectedWeightsFormat::kDefault; + } } - int GetVersion(const Operator& op) const override { return 1; } + int GetVersion(const Operator& op) const override { + const auto& fc_op = static_cast(op); + return fc_op.weights_format == FullyConnectedWeightsFormat::kDefault ? 1 + : 2; + } }; class Gather : public BuiltinOperator { + public: + using BuiltinOperator::BuiltinOperator; + flatbuffers::Offset WriteOptions( + const TocoOperator& op, + flatbuffers::FlatBufferBuilder* builder) const override { + return ::tflite::CreateArgMinOptions( + *builder, DataType::Serialize(op.output_data_type)); + } + + void ReadOptions(const TfLiteOptions& options, + TocoOperator* op) const override { + op->output_data_type = DataType::Deserialize(options.output_type()); + } + + int GetVersion(const Operator& op) const override { return 1; } +}; + class TransposeConv : public BuiltinOperator> BuildOperatorList() { new Cast(::tflite::BuiltinOperator_CAST, OperatorType::kCast)); ops.emplace_back( new ArgMax(::tflite::BuiltinOperator_ARG_MAX, OperatorType::kArgMax)); + ops.emplace_back( + new ArgMin(::tflite::BuiltinOperator_ARG_MIN, OperatorType::kArgMin)); ops.emplace_back( new Tile(::tflite::BuiltinOperator_TILE, OperatorType::kTile)); ops.emplace_back(new ExpandDims(::tflite::BuiltinOperator_EXPAND_DIMS, @@ -1153,11 +1211,12 @@ std::vector> BuildOperatorList() { OperatorType::kSparseToDense)); ops.emplace_back( new Shape(::tflite::BuiltinOperator_SHAPE, OperatorType::kShape)); + ops.emplace_back(new FakeQuant(::tflite::BuiltinOperator_FAKE_QUANT, + OperatorType::kFakeQuant)); // Custom Operators. ops.emplace_back( new DepthToSpace("DEPTH_TO_SPACE", OperatorType::kDepthToSpace)); - ops.emplace_back(new FakeQuant("FAKE_QUANT", OperatorType::kFakeQuant)); ops.emplace_back(new TensorFlowUnsupported("TENSORFLOW_UNSUPPORTED", OperatorType::kUnsupported)); @@ -1206,6 +1265,7 @@ std::vector> BuildOperatorList() { new SimpleOperator("SELECT", OperatorType::kSelect)); ops.emplace_back( new SimpleOperator("SLICE", OperatorType::kSlice)); + ops.emplace_back(new SimpleOperator("POW", OperatorType::kPow)); // Element-wise operator ops.emplace_back(new SimpleOperator("SIN", OperatorType::kSin)); ops.emplace_back(new SimpleOperator("LOG", OperatorType::kLog)); diff --git a/tensorflow/contrib/lite/toco/tflite/operator_test.cc b/tensorflow/contrib/lite/toco/tflite/operator_test.cc index 79c8e5d738ab7da12a279c86df8b03d39a924fa1..ff2d35b1f5c6613e01e665cb616b85f281ba800e 100644 --- a/tensorflow/contrib/lite/toco/tflite/operator_test.cc +++ b/tensorflow/contrib/lite/toco/tflite/operator_test.cc @@ -126,6 +126,7 @@ TEST_F(OperatorTest, SimpleOperators) { CheckSimpleOperator("LOG", OperatorType::kLog); CheckSimpleOperator("SQRT", OperatorType::kSqrt); CheckSimpleOperator("RSQRT", OperatorType::kRsqrt); + CheckSimpleOperator("POW", OperatorType::kPow); } TEST_F(OperatorTest, BuiltinAdd) { @@ -415,6 +416,13 @@ TEST_F(OperatorTest, BuiltinArgMax) { EXPECT_EQ(op.output_data_type, output_toco_op->output_data_type); } +TEST_F(OperatorTest, BuiltinArgMin) { + ArgMinOperator op; + auto output_toco_op = SerializeAndDeserialize( + GetOperator("ARG_MIN", OperatorType::kArgMin), op); + EXPECT_EQ(op.output_data_type, output_toco_op->output_data_type); +} + TEST_F(OperatorTest, BuiltinTransposeConv) { TransposeConvOperator op; op.stride_width = 123; diff --git a/tensorflow/contrib/lite/toco/tflite/types.cc b/tensorflow/contrib/lite/toco/tflite/types.cc index 42c5d7e8ebc3a7b90963a92843af616d9e6532d6..754f0b4b8c661355c99d9e5a86f2d7844414a303 100644 --- a/tensorflow/contrib/lite/toco/tflite/types.cc +++ b/tensorflow/contrib/lite/toco/tflite/types.cc @@ -100,6 +100,8 @@ void CopyBuffer(const ::tflite::Buffer& buffer, Array* array) { return ::tflite::TensorType_STRING; case ArrayDataType::kBool: return ::tflite::TensorType_BOOL; + case ArrayDataType::kComplex64: + return ::tflite::TensorType_COMPLEX64; default: // FLOAT32 is filled for unknown data types. // TODO(ycling): Implement type inference in TF Lite interpreter. @@ -123,6 +125,8 @@ ArrayDataType DataType::Deserialize(int tensor_type) { return ArrayDataType::kUint8; case ::tflite::TensorType_BOOL: return ArrayDataType::kBool; + case ::tflite::TensorType_COMPLEX64: + return ArrayDataType::kComplex64; default: LOG(FATAL) << "Unhandled tensor type '" << tensor_type << "'."; } @@ -147,6 +151,8 @@ flatbuffers::Offset> DataBuffer::Serialize( return CopyBuffer(array, builder); case ArrayDataType::kBool: return CopyBoolToBuffer(array, builder); + case ArrayDataType::kComplex64: + return CopyBuffer(array, builder); default: LOG(FATAL) << "Unhandled array data type."; } @@ -172,6 +178,8 @@ void DataBuffer::Deserialize(const ::tflite::Tensor& tensor, return CopyBuffer(buffer, array); case ::tflite::TensorType_BOOL: return CopyBuffer(buffer, array); + case ::tflite::TensorType_COMPLEX64: + return CopyBuffer(buffer, array); default: LOG(FATAL) << "Unhandled tensor type."; } diff --git a/tensorflow/contrib/lite/toco/tflite/types_test.cc b/tensorflow/contrib/lite/toco/tflite/types_test.cc index 8c6ef95bfab0a5e9b410748eabf9570eec52c2e0..8e9f30ba3a6e6b98fa9c4237567b0797a5a797aa 100644 --- a/tensorflow/contrib/lite/toco/tflite/types_test.cc +++ b/tensorflow/contrib/lite/toco/tflite/types_test.cc @@ -14,6 +14,8 @@ limitations under the License. ==============================================================================*/ #include "tensorflow/contrib/lite/toco/tflite/types.h" +#include + #include #include @@ -71,7 +73,8 @@ TEST(DataType, SupportedTypes) { {ArrayDataType::kInt32, ::tflite::TensorType_INT32}, {ArrayDataType::kInt64, ::tflite::TensorType_INT64}, {ArrayDataType::kFloat, ::tflite::TensorType_FLOAT32}, - {ArrayDataType::kBool, ::tflite::TensorType_BOOL}}; + {ArrayDataType::kBool, ::tflite::TensorType_BOOL}, + {ArrayDataType::kComplex64, ::tflite::TensorType_COMPLEX64}}; for (auto x : testdata) { EXPECT_EQ(x.second, DataType::Serialize(x.first)); EXPECT_EQ(x.first, DataType::Deserialize(x.second)); @@ -171,6 +174,14 @@ TEST(DataBuffer, Bool) { ::testing::ElementsAre(true, false, true)); } +TEST(DataBuffer, Complex64) { + Array recovered = ToFlatBufferAndBack( + {std::complex(1.0f, 2.0f), std::complex(3.0f, 4.0f)}); + EXPECT_THAT(recovered.GetBuffer().data, + ::testing::ElementsAre(std::complex(1.0f, 2.0f), + std::complex(3.0f, 4.0f))); +} + TEST(Padding, All) { EXPECT_EQ(::tflite::Padding_SAME, Padding::Serialize(PaddingType::kSame)); EXPECT_EQ(PaddingType::kSame, Padding::Deserialize(::tflite::Padding_SAME)); diff --git a/tensorflow/contrib/lite/toco/toco.cc b/tensorflow/contrib/lite/toco/toco.cc index 8041aa9e7fbfdaf44134395fee4b2bb01633893a..0b460bd178a49cafefd3438b7ae1c38a07b2ab7c 100644 --- a/tensorflow/contrib/lite/toco/toco.cc +++ b/tensorflow/contrib/lite/toco/toco.cc @@ -23,7 +23,6 @@ limitations under the License. #include "tensorflow/contrib/lite/toco/toco_cmdline_flags.h" #include "tensorflow/contrib/lite/toco/toco_flags.pb.h" #include "tensorflow/contrib/lite/toco/toco_port.h" -#include "tensorflow/contrib/lite/toco/toco_saved_model.h" #include "tensorflow/contrib/lite/toco/toco_tooling.h" #include "tensorflow/contrib/lite/toco/toco_types.h" #include "tensorflow/core/platform/logging.h" @@ -49,17 +48,6 @@ void CheckFrozenModelPermissions(const Arg& input_file) { << input_file.value() << ".\n"; } -// Checks the permissions of the SavedModel directory. -void CheckSavedModelPermissions(const Arg& savedmodel_directory) { - QCHECK(savedmodel_directory.specified()) - << "Missing required flag --savedmodel_directory.\n"; - QCHECK( - port::file::Exists(savedmodel_directory.value(), port::file::Defaults()) - .ok()) - << "Specified savedmodel_directory does not exist: " - << savedmodel_directory.value() << ".\n"; -} - // Reads the contents of the GraphDef from either the frozen graph file or the // SavedModel directory. If it reads the SavedModel directory, it updates the // ModelFlags and TocoFlags accordingly. @@ -69,24 +57,16 @@ void ReadInputData(const ParsedTocoFlags& parsed_toco_flags, string* graph_def_contents) { port::CheckInitGoogleIsDone("InitGoogle is not done yet.\n"); - bool has_input_file = parsed_toco_flags.input_file.specified(); - bool has_savedmodel_dir = parsed_toco_flags.savedmodel_directory.specified(); - - // Ensure either input_file or savedmodel_directory flag has been set. - QCHECK_NE(has_input_file, has_savedmodel_dir) - << "Specify either input_file or savedmodel_directory flag.\n"; + // Ensure savedmodel_directory is not set. + QCHECK(!parsed_toco_flags.savedmodel_directory.specified()) + << "Use `tensorflow/contrib/lite/python/tflite_convert` script with " + << "SavedModel directories.\n"; // Checks the input file permissions and reads the contents. - if (has_input_file) { - CheckFrozenModelPermissions(parsed_toco_flags.input_file); - CHECK(port::file::GetContents(parsed_toco_flags.input_file.value(), - graph_def_contents, port::file::Defaults()) - .ok()); - } else { - CheckSavedModelPermissions(parsed_toco_flags.savedmodel_directory); - GetSavedModelContents(parsed_toco_flags, parsed_model_flags, toco_flags, - model_flags, graph_def_contents); - } + CheckFrozenModelPermissions(parsed_toco_flags.input_file); + CHECK(port::file::GetContents(parsed_toco_flags.input_file.value(), + graph_def_contents, port::file::Defaults()) + .ok()); } void ToolMain(const ParsedTocoFlags& parsed_toco_flags, diff --git a/tensorflow/contrib/lite/toco/toco_cmdline_flags.cc b/tensorflow/contrib/lite/toco/toco_cmdline_flags.cc index 87a1e429b928bf59cb14597980602953732a7659..c6d0a03452f7477841d7e68665baf32dff45f41c 100644 --- a/tensorflow/contrib/lite/toco/toco_cmdline_flags.cc +++ b/tensorflow/contrib/lite/toco/toco_cmdline_flags.cc @@ -41,7 +41,7 @@ bool ParseTocoFlagsFromCommandLineFlags( "extension."), Flag("savedmodel_directory", parsed_flags.savedmodel_directory.bind(), parsed_flags.savedmodel_directory.default_value(), - "Full path to the directory containing the SavedModel."), + "Deprecated. Full path to the directory containing the SavedModel."), Flag("output_file", parsed_flags.output_file.bind(), parsed_flags.output_file.default_value(), "Output file. " @@ -55,9 +55,9 @@ bool ParseTocoFlagsFromCommandLineFlags( "One of TENSORFLOW_GRAPHDEF, TFLITE, GRAPHVIZ_DOT."), Flag("savedmodel_tagset", parsed_flags.savedmodel_tagset.bind(), parsed_flags.savedmodel_tagset.default_value(), - "Comma-separated set of tags identifying the MetaGraphDef within " - "the SavedModel to analyze. All tags in the tag set must be " - "specified."), + "Deprecated. Comma-separated set of tags identifying the " + "MetaGraphDef within the SavedModel to analyze. All tags in the tag " + "set must be specified."), Flag("default_ranges_min", parsed_flags.default_ranges_min.bind(), parsed_flags.default_ranges_min.default_value(), "If defined, will be used as the default value for the min bound " diff --git a/tensorflow/contrib/lite/toco/toco_flags.proto b/tensorflow/contrib/lite/toco/toco_flags.proto index ad4e94ded9f9730842a257e065d9aec2b1cbfac8..b4a9870d5834d1d5689d15ebc131ac0ead3e9850 100644 --- a/tensorflow/contrib/lite/toco/toco_flags.proto +++ b/tensorflow/contrib/lite/toco/toco_flags.proto @@ -37,7 +37,7 @@ enum FileFormat { // of as properties of models, instead describing how models are to be // processed in the context of the present tooling job. // -// Next ID to use: 21. +// Next ID to use: 26. message TocoFlags { // Input file format optional FileFormat input_format = 1; diff --git a/tensorflow/contrib/lite/toco/toco_saved_model.cc b/tensorflow/contrib/lite/toco/toco_saved_model.cc deleted file mode 100644 index 26f55a66c729894a990258080e397bb42ea98a13..0000000000000000000000000000000000000000 --- a/tensorflow/contrib/lite/toco/toco_saved_model.cc +++ /dev/null @@ -1,189 +0,0 @@ -/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. -==============================================================================*/ - -#include -#include - -#include "absl/strings/numbers.h" -#include "tensorflow/contrib/lite/toco/model_cmdline_flags.h" -#include "tensorflow/contrib/lite/toco/toco_saved_model.h" -#include "tensorflow/core/framework/attr_value.pb.h" -#include "tensorflow/core/framework/node_def.pb.h" -#include "tensorflow/core/framework/tensor_shape.pb.h" - -namespace toco { -namespace { - -// Loads a SavedModel from the directory specified in parsed_toco_flags. -// Returns a SavedModelBundle with the requested MetaGraphDef. -const tensorflow::SavedModelBundle* LoadSavedModel( - const ParsedTocoFlags& parsed_toco_flags) { - const string model_path = parsed_toco_flags.savedmodel_directory.value(); - QCHECK(tensorflow::MaybeSavedModelDirectory(model_path)) - << "Model is not saved in the supported SavedModel format.\n"; - - // Gets the tags identifying the MetaGraphDef from the command line arguments. - string tags_str; - if (parsed_toco_flags.savedmodel_tagset.specified()) { - tags_str = parsed_toco_flags.savedmodel_tagset.value(); - } else { - tags_str = parsed_toco_flags.savedmodel_tagset.default_value(); - } - auto tags = absl::StrSplit(tags_str, ','); - - // Loads MetaGraphDef. - auto* bundle = new tensorflow::SavedModelBundle; - TF_CHECK_OK(tensorflow::LoadSavedModel(tensorflow::SessionOptions(), - tensorflow::RunOptions(), model_path, - tags, bundle)) - << "Failed to load exported model from " << model_path - << ". Ensure the model contains the required tags '" << tags_str - << "'.\n"; - return bundle; -} - -// Returns the array name without the postfix. -// -// e.g. reduces "input:0" to "input". -string GetArrayName(const string& name) { - const std::vector& names = absl::StrSplit(name, ':'); - return names[0]; -} - -// Returns the list of array names without the postfix sorted alphabetically. -std::set GetSortedNames(const std::unordered_set& names) { - std::vector final_names; - final_names.reserve(names.size()); - for (const auto& name : names) { - final_names.push_back(GetArrayName(name)); - } - return std::set(final_names.begin(), final_names.end()); -} - -// Gets the final shape after replacing the first dimension with batch size, if -// it is undefined (containing the value -1). Returns whether the shape is -// valid. -bool ReplaceShapeBatchSize(const tensorflow::TensorShapeProto& shape, - int batch_size, - tensorflow::TensorShapeProto* final_shape) { - for (int idx = 0; idx < shape.dim().size(); ++idx) { - int64 final_dim = shape.dim()[idx].size(); - if (final_dim == -1) { - if (idx > 0) return false; - final_dim = batch_size; - } - final_shape->add_dim()->set_size(final_dim); - } - return true; -} - -// Updates the input arrays in ModelFlags to contain the shape of the array. -void ProcessInputShapes(const tensorflow::GraphDef& graph_def, int batch_size, - ModelFlags* model_flags) { - // Build map of input array names to input arrays. - std::unordered_map input_data_map; - for (auto& input : *model_flags->mutable_input_arrays()) { - input_data_map[input.name()] = &input; - } - - // Adds shapes to the input arrays if the shape is valid. - for (const tensorflow::NodeDef& node_def : graph_def.node()) { - if (input_data_map.find(node_def.name()) != input_data_map.end()) { - const auto shape_it = node_def.attr().find("shape"); - if (shape_it != node_def.attr().end()) { - tensorflow::TensorShapeProto final_shape; - bool is_valid = ReplaceShapeBatchSize(shape_it->second.shape(), - batch_size, &final_shape); - - if (is_valid) { - auto* shape = input_data_map.at(node_def.name())->mutable_shape(); - QCHECK_EQ(shape->dims_size(), 0) - << "The shape for the input '" << node_def.name() - << "' was previously defined. For clarity please define inputs " - << "via --input_arrays and input_shapes flags.\n"; - for (const auto& dim : final_shape.dim()) { - shape->add_dims(dim.size()); - } - } - } - } - } - - // Checks all input arrays have a shape. - for (auto const& input : model_flags->input_arrays()) { - QCHECK(input.shape().dims_size() > 0) - << "A valid input shape was not found for input '" << input.name() - << "'. Please define via --input_arrays and --input_shapes flags.\n"; - } -} - -} // namespace - -void ParseMetaData(const tensorflow::GraphDef& graph_def, - const std::unordered_set& inputs, - const std::unordered_set& outputs, - const ParsedTocoFlags& parsed_toco_flags, - const ParsedModelFlags& parsed_model_flags, - TocoFlags* toco_flags, ModelFlags* model_flags) { - if (!parsed_model_flags.input_arrays.specified()) { - const std::set sorted_inputs = GetSortedNames(inputs); - for (const auto& input_name : sorted_inputs) { - model_flags->add_input_arrays()->set_name(input_name); - } - } - - if (!parsed_model_flags.output_arrays.specified()) { - const std::set sorted_outputs = GetSortedNames(outputs); - for (const auto& output_name : sorted_outputs) { - model_flags->add_output_arrays(GetArrayName(output_name)); - } - } - - if (!parsed_model_flags.input_shapes.specified()) { - int batch_size = parsed_model_flags.batch_size.value(); - ProcessInputShapes(graph_def, batch_size, model_flags); - } - - if (!parsed_toco_flags.inference_type.specified()) { - toco_flags->set_inference_type(IODataType::FLOAT); - } -} - -// TODO(nupurgarg): Add top level tests. -void GetSavedModelContents(const ParsedTocoFlags& parsed_toco_flags, - const ParsedModelFlags& parsed_model_flags, - TocoFlags* toco_flags, ModelFlags* model_flags, - string* graph_def_contents) { - // Loads the MetaGraphDef within a SavedModelBundle. - auto bundle = LoadSavedModel(parsed_toco_flags); - - // Converts the MetaGraphDef to frozen GraphDef. - tensorflow::GraphDef frozen_graph_def; - std::unordered_set inputs; - std::unordered_set outputs; - TF_CHECK_OK(tensorflow::FreezeSavedModel(*bundle, &frozen_graph_def, &inputs, - &outputs)); - - // Reads the frozen GraphDef into a string. - QCHECK(frozen_graph_def.SerializeToString(graph_def_contents)) - << "Unable to generate serialized GraphDef.\n"; - - // Process inputs and outputs and metadata within GraphDef. - const tensorflow::GraphDef graph_def = bundle->meta_graph_def.graph_def(); - ParseMetaData(graph_def, inputs, outputs, parsed_toco_flags, - parsed_model_flags, toco_flags, model_flags); -} - -} // namespace toco diff --git a/tensorflow/contrib/lite/toco/toco_saved_model.h b/tensorflow/contrib/lite/toco/toco_saved_model.h deleted file mode 100644 index 7a0fabd82d90131a3b2d28c757c08dcb0f9e3988..0000000000000000000000000000000000000000 --- a/tensorflow/contrib/lite/toco/toco_saved_model.h +++ /dev/null @@ -1,53 +0,0 @@ -/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. -==============================================================================*/ - -#ifndef TENSORFLOW_CONTRIB_LITE_TOCO_TOCO_SAVED_MODEL_H_ -#define TENSORFLOW_CONTRIB_LITE_TOCO_TOCO_SAVED_MODEL_H_ - -#include -#include - -#include "tensorflow/cc/tools/freeze_saved_model.h" -#include "tensorflow/contrib/lite/toco/args.h" -#include "tensorflow/contrib/lite/toco/model_flags.pb.h" -#include "tensorflow/contrib/lite/toco/toco_flags.pb.h" -#include "tensorflow/contrib/lite/toco/types.pb.h" - -namespace toco { - -// Parses metadata into `toco_flags` and `model_flags`. -// -// Stores `inputs` as input_arrays and `outputs` as output_arrays in -// `model_flags`. Infers input_shapes from the GraphDef and stores it in -// `model_flags` as part of the input_arrays. Assumes inference_type is FLOAT -// and stores it in `toco_flags`. -void ParseMetaData(const tensorflow::GraphDef& graph_def, - const std::unordered_set& inputs, - const std::unordered_set& outputs, - const ParsedTocoFlags& parsed_toco_flags, - const ParsedModelFlags& parsed_model_flags, - TocoFlags* toco_flags, ModelFlags* model_flags); - -// Generates a frozen graph from the SavedModel in the directory specified in -// `toco_flags`. Reads frozen graph contents into `graph_def_contents`. Parses -// metadata relating to the GraphDef into `toco_flags` and `model_flags`. -void GetSavedModelContents(const ParsedTocoFlags& parsed_toco_flags, - const ParsedModelFlags& parsed_model_flags, - TocoFlags* toco_flags, ModelFlags* model_flags, - string* graph_def_contents); - -} // namespace toco - -#endif // TENSORFLOW_CONTRIB_LITE_TOCO_TOCO_SAVED_MODEL_H_ diff --git a/tensorflow/contrib/lite/toco/toco_saved_model_test.cc b/tensorflow/contrib/lite/toco/toco_saved_model_test.cc deleted file mode 100644 index 5e122afe65dc29abc85f142f4019aae5058ace51..0000000000000000000000000000000000000000 --- a/tensorflow/contrib/lite/toco/toco_saved_model_test.cc +++ /dev/null @@ -1,274 +0,0 @@ -/* Copyright 2018 The TensorFlow Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. -==============================================================================*/ - -#include "tensorflow/contrib/lite/toco/toco_saved_model.h" -#include "absl/strings/str_join.h" -#include "tensorflow/cc/framework/scope.h" -#include "tensorflow/cc/ops/standard_ops.h" -#include "tensorflow/contrib/lite/toco/model_cmdline_flags.h" -#include "tensorflow/contrib/lite/toco/toco_cmdline_flags.h" -#include "tensorflow/core/lib/core/status_test_util.h" - -#include -#include - -namespace toco { -namespace { - -using tensorflow::ops::Add; -using tensorflow::ops::Const; -using tensorflow::ops::FakeQuantWithMinMaxArgs; -using tensorflow::ops::Placeholder; - -class TocoSavedModelTest : public ::testing::Test { - protected: - // Calls functions to process cmdline arguments and calls ParseMetaData. - // ParseMetaData parses input_arrays, output_arrays, and gets metadata from - // SavedModel it is not defined in the cmdline arguments. - void ProcessGraphDefMetadata(const std::unordered_set& inputs, - const std::unordered_set& outputs, - const tensorflow::GraphDef& graph_def) { - ReadTocoFlagsFromCommandLineFlags(parsed_toco_flags_, &toco_flags_); - ReadModelFlagsFromCommandLineFlags(parsed_model_flags_, &model_flags_); - ParseMetaData(graph_def, inputs, outputs, parsed_toco_flags_, - parsed_model_flags_, &toco_flags_, &model_flags_); - } - - // Gets the GraphDef from the SavedModelBundle and processes metadata. - void ProcessSavedModelMetadata(const std::unordered_set& inputs, - const std::unordered_set& outputs) { - const tensorflow::GraphDef graph_def = bundle_.meta_graph_def.graph_def(); - ProcessGraphDefMetadata(inputs, outputs, graph_def); - } - - // Returns a GraphDef representing a simple float model with a single input. - tensorflow::GraphDef GetFloatGraphDef(const std::vector& shape) { - tensorflow::GraphDef graph_def; - tensorflow::Scope scope = tensorflow::Scope::NewRootScope(); - - tensorflow::Output input = - Placeholder(scope.WithOpName("input"), tensorflow::DT_FLOAT, - Placeholder::Shape(tensorflow::PartialTensorShape(shape))); - tensorflow::Output zero = Const(scope.WithOpName("zero"), 0.0f, {}); - tensorflow::Output add = Add(scope.WithOpName("add"), input, zero); - - TF_EXPECT_OK(scope.ToGraphDef(&graph_def)); - return graph_def; - } - - // Returns a GraphDef representing a simple float model with two inputs. - tensorflow::GraphDef GetComplexFloatGraphDef() { - tensorflow::GraphDef graph_def; - tensorflow::Scope scope = tensorflow::Scope::NewRootScope(); - - tensorflow::Output inputA = - Placeholder(scope.WithOpName("inputA"), tensorflow::DT_FLOAT, - Placeholder::Shape(tensorflow::TensorShape({1, 3, 3, 1}))); - tensorflow::Output inputB = - Placeholder(scope.WithOpName("inputB"), tensorflow::DT_FLOAT, - Placeholder::Shape(tensorflow::TensorShape({1, 3, 3, 1}))); - tensorflow::Output add = Add(scope.WithOpName("add"), inputB, inputA); - - TF_EXPECT_OK(scope.ToGraphDef(&graph_def)); - return graph_def; - } - - // Returns a GraphDef representing a simple quantized model. - tensorflow::GraphDef GetQuantizedGraphDef() { - tensorflow::GraphDef graph_def; - tensorflow::Scope scope = tensorflow::Scope::NewRootScope(); - - tensorflow::Output input = - Placeholder(scope.WithOpName("input"), tensorflow::DT_FLOAT, - Placeholder::Shape(tensorflow::TensorShape({1, 3, 3, 1}))); - tensorflow::Output zero = Const(scope.WithOpName("zero"), 0.0f, {}); - tensorflow::Output fake_quant = - FakeQuantWithMinMaxArgs(scope.WithOpName("quant"), zero); - tensorflow::Output add = Add(scope.WithOpName("add"), input, fake_quant); - - TF_EXPECT_OK(scope.ToGraphDef(&graph_def)); - return graph_def; - } - - // Gets the values in the input_arrays flag. - std::vector GetInputArrays() { - std::vector actual; - for (const auto& input : model_flags_.input_arrays()) { - actual.push_back(input.name()); - } - return actual; - } - - // Gets the values in the output_arrays flag. - std::vector GetOutputArrays() { - std::vector actual(model_flags_.output_arrays().begin(), - model_flags_.output_arrays().end()); - return actual; - } - - // Gets the shape of the given input array. - string GetInputShape(const string& input_array) { - for (const auto& input : model_flags_.input_arrays()) { - if (input.name() == input_array) { - std::vector dims; - for (int idx = 0; idx < input.shape().dims_size(); ++idx) { - dims.push_back(std::to_string(input.shape().dims(idx))); - } - return absl::StrJoin(dims, ","); - } - } - return ""; - } - - tensorflow::SavedModelBundle bundle_; - ParsedTocoFlags parsed_toco_flags_; - ParsedModelFlags parsed_model_flags_; - TocoFlags toco_flags_; - ModelFlags model_flags_; -}; - -// Tests if input_arrays, output_arrays, inference_type, and output_arrays are -// added to ModelFlags if they are not specified in cmdline arguments. -// Tests if the default batch size replaces a -1 in the first dimension. -TEST_F(TocoSavedModelTest, NoCmdLine) { - tensorflow::GraphDef graph_def = GetFloatGraphDef({-1, 3, 3, 1}); - - ProcessGraphDefMetadata({"input"}, {"add"}, graph_def); - EXPECT_EQ(GetInputArrays(), std::vector({"input"})); - EXPECT_EQ(GetOutputArrays(), std::vector({"add"})); - EXPECT_EQ(GetInputShape("input"), "1,3,3,1"); - EXPECT_EQ(toco_flags_.inference_type(), IODataType::FLOAT); -} - -// Tests if the order of input_arrays and output_arrays is deterministic when -// they are taken from the SavedModel. -TEST_F(TocoSavedModelTest, NoCmdLineMultipleArrays) { - tensorflow::GraphDef graph_def = GetComplexFloatGraphDef(); - - // Note: The model does not have two outputs. However, the function does not - // need an accurate output_array list. This is only meant to test order. - ProcessGraphDefMetadata({"inputB", "inputA"}, {"add", "invalid"}, graph_def); - EXPECT_EQ(GetInputArrays(), std::vector({"inputA", "inputB"})); - EXPECT_EQ(GetOutputArrays(), std::vector({"add", "invalid"})); - EXPECT_EQ(GetInputShape("inputA"), "1,3,3,1"); - EXPECT_EQ(GetInputShape("inputB"), "1,3,3,1"); - EXPECT_EQ(toco_flags_.inference_type(), IODataType::FLOAT); -} - -// Tests if input_shapes is inferred when input_arrays is passed in via cmdline -// arguments. -TEST_F(TocoSavedModelTest, InputNameWithoutInputShape) { - parsed_model_flags_.input_arrays.bind()("input"); - tensorflow::GraphDef graph_def = GetFloatGraphDef({2, 3, 3, 1}); - - ProcessGraphDefMetadata({"not_used_input"}, {"add"}, graph_def); - EXPECT_EQ(GetInputArrays(), std::vector({"input"})); - EXPECT_EQ(GetOutputArrays(), std::vector({"add"})); - EXPECT_EQ(GetInputShape("input"), "2,3,3,1"); - EXPECT_EQ(toco_flags_.inference_type(), IODataType::FLOAT); -} - -// Ensures a failure occurs when input_shapes is defined without input_arrays. -TEST_F(TocoSavedModelTest, InputShapeWithoutInputName) { - parsed_model_flags_.input_shapes.bind()("1,224,224,1:9,12"); - tensorflow::GraphDef graph_def = GetFloatGraphDef({1, 3, 3, 1}); - - EXPECT_DEATH(ProcessGraphDefMetadata({"input"}, {"add"}, graph_def), - "failed: input_shapes.size\\(\\) == " - "model_flags->input_arrays_size\\(\\)"); -} - -// Tests if the cmdline values of input_arrays, input_shapes are used when -// specified with an empty GraphDef. -TEST_F(TocoSavedModelTest, InputArraysCmdLine) { - parsed_model_flags_.input_arrays.bind()("inputA,inputB"); - parsed_model_flags_.input_shapes.bind()("1,224,224,1:9,12"); - - ProcessSavedModelMetadata({"input0", "input1"}, {"output0", "output1"}); - EXPECT_EQ(GetInputArrays(), std::vector({"inputA", "inputB"})); - EXPECT_EQ(GetOutputArrays(), std::vector({"output0", "output1"})); - EXPECT_EQ(GetInputShape("inputA"), "1,224,224,1"); - EXPECT_EQ(GetInputShape("inputB"), "9,12"); - EXPECT_EQ(toco_flags_.inference_type(), IODataType::FLOAT); -} - -// Tests if the cmdline values of input_arrays, input_shapes are used when -// specified even if values exist within the GraphDef. -TEST_F(TocoSavedModelTest, InputArraysCmdLineWithGraphDef) { - parsed_model_flags_.input_arrays.bind()("inputA"); - parsed_model_flags_.input_shapes.bind()("1,224,224,1"); - tensorflow::GraphDef graph_def = GetFloatGraphDef({1, 3, 3, 1}); - - ProcessGraphDefMetadata({"inputA"}, {"add"}, graph_def); - EXPECT_EQ(GetInputArrays(), std::vector({"inputA"})); - EXPECT_EQ(GetOutputArrays(), std::vector({"add"})); - EXPECT_EQ(GetInputShape("inputA"), "1,224,224,1"); - EXPECT_EQ(toco_flags_.inference_type(), IODataType::FLOAT); -} - -// Tests if the cmdline values of input_arrays, input_shapes, inference_type, -// and output_arrays are used when specified with an empty GraphDef. -TEST_F(TocoSavedModelTest, AllParamsCmdLine) { - parsed_model_flags_.input_arrays.bind()("inputA,inputB"); - parsed_model_flags_.output_arrays.bind()("outputA,outputB"); - parsed_model_flags_.input_shapes.bind()("1,224,224,1:9,12"); - parsed_toco_flags_.inference_type.bind()("FLOAT"); - - ProcessSavedModelMetadata({"input0", "input1"}, {"output0", "output1"}); - EXPECT_EQ(GetInputArrays(), std::vector({"inputA", "inputB"})); - EXPECT_EQ(GetOutputArrays(), std::vector({"outputA", "outputB"})); - EXPECT_EQ(GetInputShape("inputA"), "1,224,224,1"); - EXPECT_EQ(GetInputShape("inputB"), "9,12"); - EXPECT_EQ(toco_flags_.inference_type(), IODataType::FLOAT); -} - -// Tests if a quantized graph gives the correct values assuming type is passed -// in via command line. -TEST_F(TocoSavedModelTest, QuantizedNoCmdLine) { - parsed_toco_flags_.inference_type.bind()("QUANTIZED_UINT8"); - tensorflow::GraphDef graph_def = GetQuantizedGraphDef(); - - ProcessGraphDefMetadata({"input"}, {"add"}, graph_def); - EXPECT_EQ(GetInputArrays(), std::vector({"input"})); - EXPECT_EQ(GetOutputArrays(), std::vector({"add"})); - EXPECT_EQ(GetInputShape("input"), "1,3,3,1"); - EXPECT_EQ(toco_flags_.inference_type(), IODataType::QUANTIZED_UINT8); -} - -// Tests if the provided batch size replaces a -1 in the first dimension of -// input shape. -TEST_F(TocoSavedModelTest, MissingShapeParameterValid) { - parsed_model_flags_.batch_size.bind()(3); - tensorflow::GraphDef graph_def = GetFloatGraphDef({-1, 3, 3, 1}); - - ProcessGraphDefMetadata({"input"}, {"add"}, graph_def); - EXPECT_EQ(GetInputArrays(), std::vector({"input"})); - EXPECT_EQ(GetOutputArrays(), std::vector({"add"})); - EXPECT_EQ(GetInputShape("input"), "3,3,3,1"); - EXPECT_EQ(toco_flags_.inference_type(), IODataType::FLOAT); -} - -// Ensures a failure occurs if there is a -1 in a dimension aside from the first -// position of input shape. -TEST_F(TocoSavedModelTest, MissingShapeParameterInvalid) { - parsed_model_flags_.batch_size.bind()(3); - tensorflow::GraphDef graph_def = GetFloatGraphDef({1, -1, 3, 1}); - - EXPECT_DEATH(ProcessGraphDefMetadata({"input"}, {"add"}, graph_def), - "A valid input shape was not found for input 'input'."); -} - -} // namespace -} // namespace toco diff --git a/tensorflow/contrib/lite/toco/toco_tooling.cc b/tensorflow/contrib/lite/toco/toco_tooling.cc index 2534d1ef2ad3409a42836ad9470d8ac53d62894a..a4dc1bbe93fb4e65b4531c7dc10b79bb96cb339a 100644 --- a/tensorflow/contrib/lite/toco/toco_tooling.cc +++ b/tensorflow/contrib/lite/toco/toco_tooling.cc @@ -79,6 +79,7 @@ void MakeGeneralGraphTransformationsSet( transformations->Add(new FuseBinaryIntoFollowingAffine); transformations->Add(new FuseBroadcastIntoFollowingBinary); transformations->Add(new MergeReshapeIntoPrecedingTranspose); + transformations->Add(new MoveBinaryOperatorBeforeReshape); transformations->Add(new ReorderElementwiseUnary); transformations->Add(new ReorderReshapeTranspose); transformations->Add(new ResolveBatchNormalization); @@ -104,7 +105,8 @@ void MakeGeneralGraphTransformationsSet( transformations->Add(new IdentifyRelu1); transformations->Add(new IdentifyPRelu); transformations->Add(new RemoveTrivialBinaryOperator); - transformations->Add(new ReadFakeQuantMinMax); + transformations->Add(new ResolveFakeQuantArgsFromVars); + transformations->Add(new ReadArrayMinmaxAndNarrowRangeFromFakeQuant); transformations->Add(new ResolveSpaceToBatchNDAttributes); transformations->Add(new ResolveBatchToSpaceNDAttributes); transformations->Add(new ResolvePadAttributes); @@ -134,6 +136,8 @@ bool SupportsPreallocatedWorkspace(FileFormat format) { return (format == TFLITE); } +bool SupportsShuffledFCWeights(FileFormat format) { return format == TFLITE; } + bool IsRealValued(toco::ArrayDataType type) { // TODO(benoitjacob) - this is hardcoding that uint8 and int16 are only used // for quantized real-number values, and no other integer type is ever used @@ -270,13 +274,16 @@ void Transform(const TocoFlags& toco_flags, Model* model) { transformations.Add(new toco::MergeLstmCellInputs); } } - if (toco_flags.quantize_weights()) { - transformations.Add(new QuantizeWeights); - } transformations.Add(new ResolveConstantConcatenation); RunGraphTransformations(model, "general graph transformations", transformations); + if (toco_flags.quantize_weights()) { + // Run the quantize weights transformation after batchnorms have been + // folded into the weights. + RunGraphTransformations(model, "quantize weights transformation", + {new QuantizeWeights}); + } if (quantize_output) { if (toco_flags.propagate_fake_quant_num_bits()) { RunGraphTransformations(model, @@ -335,6 +342,10 @@ void Transform(const TocoFlags& toco_flags, Model* model) { new RemoveFinalDequantizeOp, ensure_safe_for_int8_kernels, }); + if (SupportsShuffledFCWeights(output_format)) { + RunGraphTransformations(model, "shuffling of FC weights", + {new ShuffleFCWeights}); + } } else { GraphTransformationsSet dequantization_transformations{new Dequantize}; // Dequantize creates FakeQuant nodes. We may want to discard diff --git a/tensorflow/contrib/lite/toco/tooling_util.cc b/tensorflow/contrib/lite/toco/tooling_util.cc index a52c812ef45a3f82c6ca7812067e5a4b9bda3a67..4ec74e351f059df20233cf46d2edc46fd239c176 100644 --- a/tensorflow/contrib/lite/toco/tooling_util.cc +++ b/tensorflow/contrib/lite/toco/tooling_util.cc @@ -387,6 +387,7 @@ const char* OperatorTypeName(OperatorType type) { HANDLE_OPERATORTYPENAME_CASE(Mean) HANDLE_OPERATORTYPENAME_CASE(Svdf) HANDLE_OPERATORTYPENAME_CASE(ArgMax) + HANDLE_OPERATORTYPENAME_CASE(ArgMin) HANDLE_OPERATORTYPENAME_CASE(TopK_V2) HANDLE_OPERATORTYPENAME_CASE(Unsupported) HANDLE_OPERATORTYPENAME_CASE(Exp) @@ -396,6 +397,7 @@ const char* OperatorTypeName(OperatorType type) { HANDLE_OPERATORTYPENAME_CASE(SparseToDense) HANDLE_OPERATORTYPENAME_CASE(Equal) HANDLE_OPERATORTYPENAME_CASE(NotEqual) + HANDLE_OPERATORTYPENAME_CASE(Pow) default: LOG(FATAL) << "Unhandled op type"; #undef HANDLE_OPERATORTYPENAME_CASE @@ -446,8 +448,12 @@ void LogSummary(int log_level, const Model& model) { } void LogArray(int log_level, const Model& model, const string& name) { - const auto& array = model.GetArray(name); VLOG(log_level) << "Array: " << name; + if (!model.HasArray(name)) { + VLOG(log_level) << " DOES NOT EXIST"; + return; + } + const auto& array = model.GetArray(name); VLOG(log_level) << " Data type: " << ArrayDataTypeName(array.data_type); VLOG(log_level) << " Final type: " << ArrayDataTypeName(array.final_data_type); @@ -1260,8 +1266,13 @@ void InsertCopyOperator(Model* model, const string& source_array_name, auto* copy_op = new TensorFlowReshapeOperator; copy_op->inputs = { source_array_name, - CreateInt32Array(model, target_array_name + "_copy_shape", shape)}; + CreateInt32Array( + model, AvailableArrayName(*model, target_array_name + "_copy_shape"), + shape)}; copy_op->outputs = {target_array_name}; + if (target_array.has_shape()) { + copy_op->shape = target_array.shape().dims(); + } model->operators.emplace_back(copy_op); } @@ -2200,4 +2211,51 @@ void UseArraysExtraInfo(Model* model, bool quantize_output) { } } +void UndoWeightsShuffling(Model* model) { + for (const auto& op : model->operators) { + if (op->type != toco::OperatorType::kFullyConnected) { + continue; + } + const auto& fc_op = static_cast(*op); + if (fc_op.weights_format == FullyConnectedWeightsFormat::kDefault) { + continue; + } + const string& weights_name = fc_op.inputs[1]; + QCHECK_EQ(CountOpsWithInput(*model, weights_name), 1); + auto& weights_array = model->GetArray(weights_name); + QCHECK(weights_array.data_type == ArrayDataType::kUint8); + auto& weights_data = + weights_array.GetMutableBuffer().data; + const auto& weights_shape = weights_array.shape(); + QCHECK_EQ(weights_shape.dimensions_count(), 2); + const int rows = weights_shape.dims(0); + const int cols = weights_shape.dims(1); + QCHECK_EQ(rows % 4, 0); + QCHECK_EQ(cols % 16, 0); + CHECK_EQ(rows * cols, weights_data.size()); + // Compute the de-shuffled weights + std::vector deshuffled_data(weights_data.size()); + uint8* shuffled_data_ptr = weights_data.data(); + for (int r = 0; r < rows; r += 4) { + for (int c = 0; c < cols; c += 16) { + for (int i = 0; i < 4; i++) { + uint8* deshuffled_data_ptr = + deshuffled_data.data() + (r + i) * cols + c; + for (int j = 0; j < 16; j++) { + uint8 shuffled_val = *shuffled_data_ptr++; + // Deshuffling isn't only about deshuffling the storage layout, + // it's also about undoing the flipping of the sign bit, which is + // performed on the shuffled weights. + uint8 deshuffled_val = shuffled_val ^ 0x80; + *deshuffled_data_ptr++ = deshuffled_val; + } + } + } + } + CHECK_EQ(shuffled_data_ptr, weights_data.data() + rows * cols); + // Switch this FC op to using the deshuffled weights. + weights_data = std::move(deshuffled_data); + } +} + } // namespace toco diff --git a/tensorflow/contrib/lite/toco/tooling_util.h b/tensorflow/contrib/lite/toco/tooling_util.h index 791ced8d012209867f0ce7ce417d4d11b59b2ead..5dbfa54fa0369676dce638aec171b409a468da9f 100644 --- a/tensorflow/contrib/lite/toco/tooling_util.h +++ b/tensorflow/contrib/lite/toco/tooling_util.h @@ -344,6 +344,11 @@ tensorflow::Status NumElements(const std::vector& shape, U* num_elements) { return tensorflow::Status::OK(); } +// A model file may have shuffled FC weights. +// When that happens, we want to de-shuffle them immediately on import, +// so that the rest of toco doesn't need to know about shuffled weights. +void UndoWeightsShuffling(Model* model); + } // namespace toco #endif // TENSORFLOW_CONTRIB_LITE_TOCO_TOOLING_UTIL_H_ diff --git a/tensorflow/contrib/lite/tools/BUILD b/tensorflow/contrib/lite/tools/BUILD index 5913847329eeae7373d0d21834dd37327e4068c4..d070018e833f642f47d8e495b71ef7adebbb1562 100644 --- a/tensorflow/contrib/lite/tools/BUILD +++ b/tensorflow/contrib/lite/tools/BUILD @@ -14,6 +14,7 @@ py_binary( srcs = ["visualize.py"], data = [ "//tensorflow/contrib/lite/schema:schema.fbs", + "//tensorflow/python:platform", "@flatbuffers//:flatc", ], srcs_version = "PY2AND3", @@ -53,6 +54,7 @@ cc_test( ], tags = [ "tflite_not_portable_android", + "tflite_not_portable_ios", ], deps = [ ":gen_op_registration", diff --git a/tensorflow/contrib/lite/tools/benchmark/README.md b/tensorflow/contrib/lite/tools/benchmark/README.md index c10826afff6d5569545d4b7df73c88d24d9dcd1a..f1e257ad104885a23cd7f17b9c21317c0881ccc0 100644 --- a/tensorflow/contrib/lite/tools/benchmark/README.md +++ b/tensorflow/contrib/lite/tools/benchmark/README.md @@ -3,7 +3,38 @@ ## Description A simple C++ binary to benchmark a TFLite model and its individual operators, -both on desktop machines and on Android. +both on desktop machines and on Android. The binary takes a TFLite model, +generates random inputs and then repeatedly runs the model for specified number +of runs. Aggregrate latency statistics are reported after running the benchmark. + +The instructions below are for running the binary on Desktop and Android, +for iOS please use the +[iOS benchmark app] (https://github.com/tensorflow/tensorflow/tree/master/tensorflow/contrib/lite/tools/benchmark/ios). + +## Parameters + +The binary takes the following required parameters: + +* `graph`: `string` \ + The path to the TFLite model file. +* `input_layer`: `string` \ + The name of the input layer, this is typically the first layer of the model. +* `input_layer_shape`: `string` \ + The shape of the input layer. This is a comma separated string of the shape + of tensor of input layer. + +and the following optional parameters: + +* `num_threads`: `int` (default=1) \ + The number of threads to use for running TFLite interpreter. +* `warmup_runs`: `int` (default=1) \ + The number of warmup runs to do before starting the benchmark. +* `run_delay`: `float` (default=-1.0) \ + The delay in seconds between subsequent benchmark runs. Non-positive values + mean use no delay. +* `use_nnapi`: `bool` (default=false) \ + Whether to use [Android NNAPI] (https://developer.android.com/ndk/guides/neuralnetworks/). + This API is available on recent Android devices. ## To build/install/run @@ -44,7 +75,7 @@ adb push mobilenet_quant_v1_224.tflite /data/local/tmp ``` adb shell /data/local/tmp/benchmark_model \ --graph=/data/local/tmp/mobilenet_quant_v1_224.tflite \ - --input_layer="Placeholder" \ + --input_layer="input" \ --input_layer_shape="1,224,224,3" \ --num_threads=4 ``` @@ -70,6 +101,30 @@ bazel-bin/tensorflow/contrib/lite/tools/benchmark/benchmark_model \ The MobileNet graph used as an example here may be downloaded from https://storage.googleapis.com/download.tensorflow.org/models/tflite/mobilenet_v1_224_android_quant_2017_11_08.zip + +## Reducing variance between runs on Android. + +Most modern Android phones use [ARM big.LITTLE](https://en.wikipedia.org/wiki/ARM_big.LITTLE) +architecture where some cores are more power hungry but faster than other cores. +When running benchmarks on these phones there can be significant variance +between different runs of the benchmark. One way to reduce variance between runs +is to set the [CPU affinity](https://en.wikipedia.org/wiki/Processor_affinity) +before running the benchmark. On Android this can be done using the `taskset` +command. +E.g. for running the benchmark on big cores on Pixel 2 with a single thread one +can use the following command: + +``` +adb shell taskset f0 /data/local/tmp/benchmark_model \ + --graph=/data/local/tmp/mobilenet_quant_v1_224.tflite \ + --input_layer="input" \ + --input_layer_shape="1,224,224,3" \ + --num_threads=1 +``` + +where `f0` is the affinity mask for big cores on Pixel 2. +Note: The affinity mask varies with the device. + ## Profiling model operators The benchmark model binary also allows you to profile operators and give execution times of each operator. To do this, compile the binary with a compiler flag that enables profiling to be compiled in. Pass **--copt=-DTFLITE_PROFILING_ENABLED** diff --git a/tensorflow/contrib/lite/tools/benchmark/benchmark_model.cc b/tensorflow/contrib/lite/tools/benchmark/benchmark_model.cc index 08648bcfe26365d180d984fde8f8e04b22eb45dd..19b9a9c7ba40ad8241632dda77db77a5e1ce8e63 100644 --- a/tensorflow/contrib/lite/tools/benchmark/benchmark_model.cc +++ b/tensorflow/contrib/lite/tools/benchmark/benchmark_model.cc @@ -98,10 +98,13 @@ void BenchmarkModel::LogFlags() { << "]"; } +void BenchmarkModel::PrepareInputsAndOutputs() {} + Stat BenchmarkModel::Run(int num_times, RunType run_type) { Stat run_stats; TFLITE_LOG(INFO) << "Running benchmark for " << num_times << " iterations "; for (int run = 0; run < num_times; run++) { + PrepareInputsAndOutputs(); listeners_.OnSingleRunStart(run_type); int64_t start_us = profiling::time::NowMicros(); RunImpl(); diff --git a/tensorflow/contrib/lite/tools/benchmark/benchmark_model.h b/tensorflow/contrib/lite/tools/benchmark/benchmark_model.h index 942e21f67a7f864f16b7b1b85b2599d5c872b5c7..3c7063b2d49b4d0de22a74490aea0d62383da6a8 100644 --- a/tensorflow/contrib/lite/tools/benchmark/benchmark_model.h +++ b/tensorflow/contrib/lite/tools/benchmark/benchmark_model.h @@ -150,6 +150,7 @@ class BenchmarkModel { virtual std::vector GetFlags(); virtual uint64_t ComputeInputBytes() = 0; virtual tensorflow::Stat Run(int num_times, RunType run_type); + virtual void PrepareInputsAndOutputs(); virtual void RunImpl() = 0; BenchmarkParams params_; BenchmarkListeners listeners_; diff --git a/tensorflow/contrib/lite/tools/visualize.py b/tensorflow/contrib/lite/tools/visualize.py index f571dd59da0a3f4aff264b48fba3e41f75b50404..e07f899e4d8c249cb03d4251a722df0614007fed 100644 --- a/tensorflow/contrib/lite/tools/visualize.py +++ b/tensorflow/contrib/lite/tools/visualize.py @@ -28,11 +28,24 @@ import json import os import sys +from tensorflow.python.platform import resource_loader + # Schema to use for flatbuffers _SCHEMA = "third_party/tensorflow/contrib/lite/schema/schema.fbs" -# Where the binary will be once built in for the flatc converter -_BINARY = "third_party/flatbuffers/flatc" +# TODO(angerson): fix later when rules are simplified.. +_SCHEMA = resource_loader.get_path_to_datafile("../schema/schema.fbs") +_BINARY = resource_loader.get_path_to_datafile("../../../../flatbuffers/flatc") +# Account for different package positioning internal vs. external. +if not os.path.exists(_BINARY): + _BINARY = resource_loader.get_path_to_datafile( + "../../../../../flatbuffers/flatc") + +if not os.path.exists(_SCHEMA): + raise RuntimeError("Sorry, schema file cannot be found at %r" % _SCHEMA) +if not os.path.exists(_BINARY): + raise RuntimeError("Sorry, flatc is not available at %r" % _BINARY) + # A CSS description for making the visualizer _CSS = """ diff --git a/tensorflow/contrib/makefile/tf_op_files.txt b/tensorflow/contrib/makefile/tf_op_files.txt index 89db9ee2794ddf0a99951dca327e74c5d9694d23..6e7423f85e3b66e2f40b25c0b83d0fcaa54817a9 100644 --- a/tensorflow/contrib/makefile/tf_op_files.txt +++ b/tensorflow/contrib/makefile/tf_op_files.txt @@ -92,6 +92,7 @@ tensorflow/core/kernels/reduction_ops_common.cc tensorflow/core/kernels/reduction_ops_any.cc tensorflow/core/kernels/reduction_ops_all.cc tensorflow/core/kernels/roll_op.cc +tensorflow/core/kernels/queue_op.cc tensorflow/core/kernels/queue_ops.cc tensorflow/core/kernels/queue_base.cc tensorflow/core/kernels/pooling_ops_common.cc diff --git a/tensorflow/contrib/metrics/BUILD b/tensorflow/contrib/metrics/BUILD index 66cb493e5c5bb9b8645e87dc7f5b274d916f64fc..21cd34f73ffbbf615a81c18b9d365bffa61397f4 100644 --- a/tensorflow/contrib/metrics/BUILD +++ b/tensorflow/contrib/metrics/BUILD @@ -31,6 +31,7 @@ py_library( "//tensorflow/python:check_ops", "//tensorflow/python:confusion_matrix", "//tensorflow/python:control_flow_ops", + "//tensorflow/python:distribute", "//tensorflow/python:framework_for_generated_wrappers", "//tensorflow/python:histogram_ops", "//tensorflow/python:init_ops", diff --git a/tensorflow/contrib/metrics/__init__.py b/tensorflow/contrib/metrics/__init__.py index 5effea3596bb83a08e0a8627e411684262aef5f7..88798d61b71388de63e492ba69284a72303d32ab 100644 --- a/tensorflow/contrib/metrics/__init__.py +++ b/tensorflow/contrib/metrics/__init__.py @@ -63,6 +63,7 @@ See the @{$python/contrib.metrics} guide. @@aggregate_metrics @@aggregate_metric_map @@confusion_matrix +@@f1_score @@set_difference @@set_intersection @@set_size diff --git a/tensorflow/contrib/metrics/python/metrics/classification.py b/tensorflow/contrib/metrics/python/metrics/classification.py index 26aba1cc51446e589856013d69526007fbe9d921..e5536122698a50852c4cb96f12ce52ab5d5f6e39 100644 --- a/tensorflow/contrib/metrics/python/metrics/classification.py +++ b/tensorflow/contrib/metrics/python/metrics/classification.py @@ -22,6 +22,9 @@ from tensorflow.python.framework import dtypes from tensorflow.python.framework import ops from tensorflow.python.ops import array_ops from tensorflow.python.ops import math_ops +from tensorflow.python.ops import metrics_impl +from tensorflow.python.ops import variable_scope +from tensorflow.python.training import distribute as distribute_lib # TODO(nsilberman): move into metrics/python/ops/ @@ -62,3 +65,121 @@ def accuracy(predictions, labels, weights=None, name=None): return math_ops.div(math_ops.reduce_sum(is_correct), math_ops.reduce_sum(num_values)) return math_ops.reduce_mean(is_correct) + + +def f1_score(labels, predictions, weights=None, num_thresholds=200, + metrics_collections=None, updates_collections=None, name=None): + """Computes the approximately best F1-score across different thresholds. + + The f1_score function applies a range of thresholds to the predictions to + convert them from [0, 1] to bool. Precision and recall are computed by + comparing them to the labels. The F1-Score is then defined as + 2 * precision * recall / (precision + recall). The best one across the + thresholds is returned. + + Disclaimer: In practice it may be desirable to choose the best threshold on + the validation set and evaluate the F1 score with this threshold on a + separate test set. Or it may be desirable to use a fixed threshold (e.g. 0.5). + + This function internally creates four local variables, `true_positives`, + `true_negatives`, `false_positives` and `false_negatives` that are used to + compute the pairs of recall and precision values for a linearly spaced set of + thresholds from which the best f1-score is derived. + + This value is ultimately returned as `f1-score`, an idempotent operation that + computes the F1-score (computed using the aforementioned variables). The + `num_thresholds` variable controls the degree of discretization with larger + numbers of thresholds more closely approximating the true best F1-score. + + For estimation of the metric over a stream of data, the function creates an + `update_op` operation that updates these variables and returns the F1-score. + + Example usage with a custom estimator: + def model_fn(features, labels, mode): + predictions = make_predictions(features) + loss = make_loss(predictions, labels) + train_op = tf.contrib.training.create_train_op( + total_loss=loss, + optimizer='Adam') + eval_metric_ops = {'f1': f1_score(labels, predictions)} + return tf.estimator.EstimatorSpec( + mode=mode, + predictions=predictions, + loss=loss, + train_op=train_op, + eval_metric_ops=eval_metric_ops, + export_outputs=export_outputs) + estimator = tf.estimator.Estimator(model_fn=model_fn) + + If `weights` is `None`, weights default to 1. Use weights of 0 to mask values. + + Args: + labels: A `Tensor` whose shape matches `predictions`. Will be cast to + `bool`. + predictions: A floating point `Tensor` of arbitrary shape and whose values + are in the range `[0, 1]`. + weights: Optional `Tensor` whose rank is either 0, or the same rank as + `labels`, and must be broadcastable to `labels` (i.e., all dimensions must + be either `1`, or the same as the corresponding `labels` dimension). + num_thresholds: The number of thresholds to use when discretizing the roc + curve. + metrics_collections: An optional list of collections that `f1_score` should + be added to. + updates_collections: An optional list of collections that `update_op` should + be added to. + name: An optional variable_scope name. + + Returns: + f1_score: A scalar `Tensor` representing the current best f1-score across + different thresholds. + update_op: An operation that increments the `true_positives`, + `true_negatives`, `false_positives` and `false_negatives` variables + appropriately and whose value matches the `f1_score`. + + Raises: + ValueError: If `predictions` and `labels` have mismatched shapes, or if + `weights` is not `None` and its shape doesn't match `predictions`, or if + either `metrics_collections` or `updates_collections` are not a list or + tuple. + """ + with variable_scope.variable_scope( + name, 'f1', (labels, predictions, weights)): + predictions, labels, weights = metrics_impl._remove_squeezable_dimensions( # pylint: disable=protected-access + predictions=predictions, labels=labels, weights=weights) + # To account for floating point imprecisions / avoid division by zero. + epsilon = 1e-7 + thresholds = [(i + 1) * 1.0 / (num_thresholds - 1) + for i in range(num_thresholds - 2)] + thresholds = [0.0 - epsilon] + thresholds + [1.0 + epsilon] + + # Confusion matrix. + values, update_ops = metrics_impl._confusion_matrix_at_thresholds( # pylint: disable=protected-access + labels, predictions, thresholds, weights, includes=('tp', 'fp', 'fn')) + + # Compute precision and recall at various thresholds. + def compute_best_f1_score(tp, fp, fn, name): + precision_at_t = math_ops.div(tp, epsilon + tp + fp, + name='precision_' + name) + recall_at_t = math_ops.div(tp, epsilon + tp + fn, name='recall_' + name) + # Compute F1 score. + f1_at_thresholds = ( + 2.0 * precision_at_t * recall_at_t / + (precision_at_t + recall_at_t + epsilon)) + return math_ops.reduce_max(f1_at_thresholds) + + def f1_across_towers(_, values): + best_f1 = compute_best_f1_score(tp=values['tp'], fp=values['fp'], + fn=values['fn'], name='value') + if metrics_collections: + ops.add_to_collections(metrics_collections, best_f1) + return best_f1 + + best_f1 = distribute_lib.get_tower_context().merge_call( + f1_across_towers, values) + + update_op = compute_best_f1_score(tp=update_ops['tp'], fp=update_ops['fp'], + fn=update_ops['fn'], name='update') + if updates_collections: + ops.add_to_collections(updates_collections, update_op) + + return best_f1, update_op diff --git a/tensorflow/contrib/metrics/python/metrics/classification_test.py b/tensorflow/contrib/metrics/python/metrics/classification_test.py index fa0f12d029620ad6427f715f035ff69f15c133e7..3d0b81c1bed02dae013141367fb052e16d31fe08 100644 --- a/tensorflow/contrib/metrics/python/metrics/classification_test.py +++ b/tensorflow/contrib/metrics/python/metrics/classification_test.py @@ -18,9 +18,16 @@ from __future__ import absolute_import from __future__ import division from __future__ import print_function +import numpy as np + from tensorflow.contrib.metrics.python.metrics import classification +from tensorflow.python.data.ops import dataset_ops +from tensorflow.python.framework import constant_op from tensorflow.python.framework import dtypes +from tensorflow.python.framework import ops from tensorflow.python.ops import array_ops +from tensorflow.python.ops import random_ops +from tensorflow.python.ops import variables from tensorflow.python.platform import test @@ -108,5 +115,200 @@ class ClassificationTest(test.TestCase): self.assertEqual(result, 0.5) +class F1ScoreTest(test.TestCase): + + def setUp(self): + super(F1ScoreTest, self).setUp() + np.random.seed(1) + + def testVars(self): + classification.f1_score( + predictions=array_ops.ones((10, 1)), + labels=array_ops.ones((10, 1)), + num_thresholds=3) + expected = {'f1/true_positives:0', 'f1/false_positives:0', + 'f1/false_negatives:0'} + self.assertEquals( + expected, set(v.name for v in variables.local_variables())) + self.assertEquals( + set(expected), set(v.name for v in variables.local_variables())) + self.assertEquals( + set(expected), + set(v.name for v in ops.get_collection(ops.GraphKeys.METRIC_VARIABLES))) + + def testMetricsCollection(self): + my_collection_name = '__metrics__' + f1, _ = classification.f1_score( + predictions=array_ops.ones((10, 1)), + labels=array_ops.ones((10, 1)), + num_thresholds=3, + metrics_collections=[my_collection_name]) + self.assertListEqual(ops.get_collection(my_collection_name), [f1]) + + def testUpdatesCollection(self): + my_collection_name = '__updates__' + _, f1_op = classification.f1_score( + predictions=array_ops.ones((10, 1)), + labels=array_ops.ones((10, 1)), + num_thresholds=3, + updates_collections=[my_collection_name]) + self.assertListEqual(ops.get_collection(my_collection_name), [f1_op]) + + def testValueTensorIsIdempotent(self): + predictions = random_ops.random_uniform( + (10, 3), maxval=1, dtype=dtypes.float32, seed=1) + labels = random_ops.random_uniform( + (10, 3), maxval=2, dtype=dtypes.int64, seed=2) + f1, f1_op = classification.f1_score(predictions, labels, num_thresholds=3) + + with self.test_session() as sess: + sess.run(variables.local_variables_initializer()) + + # Run several updates. + for _ in range(10): + sess.run([f1_op]) + + # Then verify idempotency. + initial_f1 = f1.eval() + for _ in range(10): + self.assertAllClose(initial_f1, f1.eval()) + + def testAllCorrect(self): + inputs = np.random.randint(0, 2, size=(100, 1)) + + with self.test_session() as sess: + predictions = constant_op.constant(inputs, dtype=dtypes.float32) + labels = constant_op.constant(inputs) + f1, f1_op = classification.f1_score(predictions, labels, num_thresholds=3) + + sess.run(variables.local_variables_initializer()) + sess.run([f1_op]) + + self.assertEqual(1, f1.eval()) + + def testSomeCorrect(self): + predictions = constant_op.constant( + [1, 0, 1, 0], shape=(1, 4), dtype=dtypes.float32) + labels = constant_op.constant([0, 1, 1, 0], shape=(1, 4)) + f1, f1_op = classification.f1_score(predictions, labels, num_thresholds=1) + with self.test_session() as sess: + sess.run(variables.local_variables_initializer()) + sess.run([f1_op]) + # Threshold 0 will have around 0.5 precision and 1 recall yielding an F1 + # score of 2 * 0.5 * 1 / (1 + 0.5). + self.assertAlmostEqual(2 * 0.5 * 1 / (1 + 0.5), f1.eval()) + + def testAllIncorrect(self): + inputs = np.random.randint(0, 2, size=(10000, 1)) + + with self.test_session() as sess: + predictions = constant_op.constant(inputs, dtype=dtypes.float32) + labels = constant_op.constant(1 - inputs, dtype=dtypes.float32) + f1, f1_op = classification.f1_score(predictions, labels, num_thresholds=3) + + sess.run(variables.local_variables_initializer()) + sess.run([f1_op]) + + # Threshold 0 will have around 0.5 precision and 1 recall yielding an F1 + # score of 2 * 0.5 * 1 / (1 + 0.5). + self.assertAlmostEqual(2 * 0.5 * 1 / (1 + 0.5), f1.eval(), places=2) + + def testWeights1d(self): + with self.test_session() as sess: + predictions = constant_op.constant( + [[1, 0], [1, 0]], shape=(2, 2), dtype=dtypes.float32) + labels = constant_op.constant([[0, 1], [1, 0]], shape=(2, 2)) + weights = constant_op.constant( + [[0], [1]], shape=(2, 1), dtype=dtypes.float32) + f1, f1_op = classification.f1_score(predictions, labels, weights, + num_thresholds=3) + sess.run(variables.local_variables_initializer()) + sess.run([f1_op]) + + self.assertAlmostEqual(1.0, f1.eval(), places=5) + + def testWeights2d(self): + with self.test_session() as sess: + predictions = constant_op.constant( + [[1, 0], [1, 0]], shape=(2, 2), dtype=dtypes.float32) + labels = constant_op.constant([[0, 1], [1, 0]], shape=(2, 2)) + weights = constant_op.constant( + [[0, 0], [1, 1]], shape=(2, 2), dtype=dtypes.float32) + f1, f1_op = classification.f1_score(predictions, labels, weights, + num_thresholds=3) + sess.run(variables.local_variables_initializer()) + sess.run([f1_op]) + + self.assertAlmostEqual(1.0, f1.eval(), places=5) + + def testZeroLabelsPredictions(self): + with self.test_session() as sess: + predictions = array_ops.zeros([4], dtype=dtypes.float32) + labels = array_ops.zeros([4]) + f1, f1_op = classification.f1_score(predictions, labels, num_thresholds=3) + sess.run(variables.local_variables_initializer()) + sess.run([f1_op]) + + self.assertAlmostEqual(0.0, f1.eval(), places=5) + + def testWithMultipleUpdates(self): + num_samples = 1000 + batch_size = 10 + num_batches = int(num_samples / batch_size) + + # Create the labels and data. + labels = np.random.randint(0, 2, size=(num_samples, 1)) + noise = np.random.normal(0.0, scale=0.2, size=(num_samples, 1)) + predictions = 0.4 + 0.2 * labels + noise + predictions[predictions > 1] = 1 + predictions[predictions < 0] = 0 + thresholds = [-0.01, 0.5, 1.01] + + expected_max_f1 = -1.0 + for threshold in thresholds: + tp = 0 + fp = 0 + fn = 0 + tn = 0 + for i in range(num_samples): + if predictions[i] >= threshold: + if labels[i] == 1: + tp += 1 + else: + fp += 1 + else: + if labels[i] == 1: + fn += 1 + else: + tn += 1 + epsilon = 1e-7 + expected_prec = tp / (epsilon + tp + fp) + expected_rec = tp / (epsilon + tp + fn) + expected_f1 = (2 * expected_prec * expected_rec / + (epsilon + expected_prec + expected_rec)) + if expected_f1 > expected_max_f1: + expected_max_f1 = expected_f1 + + labels = labels.astype(np.float32) + predictions = predictions.astype(np.float32) + tf_predictions, tf_labels = (dataset_ops.Dataset + .from_tensor_slices((predictions, labels)) + .repeat() + .batch(batch_size) + .make_one_shot_iterator() + .get_next()) + f1, f1_op = classification.f1_score(tf_labels, tf_predictions, + num_thresholds=3) + + with self.test_session() as sess: + sess.run(variables.local_variables_initializer()) + for _ in range(num_batches): + sess.run([f1_op]) + # Since this is only approximate, we can't expect a 6 digits match. + # Although with higher number of samples/thresholds we should see the + # accuracy improving + self.assertAlmostEqual(expected_max_f1, f1.eval(), 2) + + if __name__ == '__main__': test.main() diff --git a/tensorflow/contrib/mixed_precision/python/loss_scale_optimizer.py b/tensorflow/contrib/mixed_precision/python/loss_scale_optimizer.py index ef34f7bf7bf3eba047b50ce8abf883b0ed741a63..93050a3ae373603c516c7eb72c22f327f4a60a00 100644 --- a/tensorflow/contrib/mixed_precision/python/loss_scale_optimizer.py +++ b/tensorflow/contrib/mixed_precision/python/loss_scale_optimizer.py @@ -77,7 +77,7 @@ class LossScaleOptimizer(optimizer.Optimizer): If gradients clipping is applied, one can call `optimizer.compute_gradients()` and `optimizer.apply_gradients()` - seperately. + separately. Notice the following way of using LossScaleOptimizer is not intended. Always use `loss_scale_optimizer.compute_gradients()` to compute gradients instead of diff --git a/tensorflow/contrib/mpi_collectives/BUILD b/tensorflow/contrib/mpi_collectives/BUILD index a7be92a35e0d62a61f7923ac61bb2c1267d039c6..ecac06354d2ce796f2a6021cdf2370d7c30ccab7 100644 --- a/tensorflow/contrib/mpi_collectives/BUILD +++ b/tensorflow/contrib/mpi_collectives/BUILD @@ -52,6 +52,7 @@ tf_custom_op_library( deps = [ ":mpi_defines", ":mpi_message_proto_cc", + "//tensorflow/stream_executor:stream_executor_headers_lib", "//third_party/mpi", ], ) diff --git a/tensorflow/contrib/mpi_collectives/kernels/mpi_ops.cc b/tensorflow/contrib/mpi_collectives/kernels/mpi_ops.cc index ed22ee667f1d73b3f86f77e09bad9bfec7e46391..e4b0c2c6541836243347d2950686c60ef06d2bfc 100644 --- a/tensorflow/contrib/mpi_collectives/kernels/mpi_ops.cc +++ b/tensorflow/contrib/mpi_collectives/kernels/mpi_ops.cc @@ -73,7 +73,7 @@ limitations under the License. */ template -using StatusOr = se::port::StatusOr; +using StatusOr = stream_executor::port::StatusOr; using CPUDevice = Eigen::ThreadPoolDevice; using GPUDevice = Eigen::GpuDevice; diff --git a/tensorflow/contrib/mpi_collectives/mpi_ops.py b/tensorflow/contrib/mpi_collectives/mpi_ops.py new file mode 100644 index 0000000000000000000000000000000000000000..bd7096d9cee2d32bde5227a95038ae65cd8a6e18 --- /dev/null +++ b/tensorflow/contrib/mpi_collectives/mpi_ops.py @@ -0,0 +1,163 @@ +# Copyright 2017 The TensorFlow Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================= +"""Inter-process communication using MPI.""" + +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + +import tensorflow as tf + +from tensorflow.python.framework import errors +from tensorflow.python.framework import load_library +from tensorflow.python.framework import ops +from tensorflow.python.platform import resource_loader +from tensorflow.python.platform import tf_logging as logging + + +def _load_library(name, op_list=None): + """Loads a .so file containing the specified operators. + + Args: + name: The name of the .so file to load. + op_list: A list of names of operators that the library should have. If None + then the .so file's contents will not be verified. + + Raises: + NameError if one of the required ops is missing. + """ + try: + filename = resource_loader.get_path_to_datafile(name) + library = load_library.load_op_library(filename) + for expected_op in (op_list or []): + for lib_op in library.OP_LIST.op: + if lib_op.name == expected_op: + break + else: + raise NameError('Could not find operator %s in dynamic library %s' % + (expected_op, name)) + return library + except errors.NotFoundError: + logging.warning('%s file could not be loaded.', name) + + +MPI_LIB = _load_library( + 'mpi_collectives.so', + ['MPISize', 'MPIRank', 'MPILocalRank', 'MPIAllgather', 'MPIAllreduce']) + + +def size(name=None): + """An op which returns the number of MPI processes. + + This is equivalent to running `MPI_Comm_size(MPI_COMM_WORLD, ...)` to get the + size of the global communicator. + + Returns: + An integer scalar containing the number of MPI processes. + """ + return MPI_LIB.mpi_size(name=name) + + +ops.NotDifferentiable('MPISize') + + +def rank(name=None): + """An op which returns the MPI rank of the calling process. + + This is equivalent to running `MPI_Comm_rank(MPI_COMM_WORLD, ...)` to get the + rank of the current process in the global communicator. + + Returns: + An integer scalar with the MPI rank of the calling process. + """ + return MPI_LIB.mpi_rank(name=name) + + +ops.NotDifferentiable('MPIRank') + + +def init(name=None): + """An op which initializes MPI on the device on which it is run. + + All future MPI ops must be run on the same device that the `init` op was run + on. + """ + return MPI_LIB.mpi_init(name=name) + + +ops.NotDifferentiable('MPIInit') + + +def local_rank(name=None): + """An op which returns the local MPI rank of the calling process, within the + node that it is running on. For example, if there are seven processes running + on a node, their local ranks will be zero through six, inclusive. + + This is equivalent to running `MPI_Comm_rank(...)` on a new communicator + which only includes processes on the same node. + + Returns: + An integer scalar with the local MPI rank of the calling process. + """ + return MPI_LIB.mpi_local_rank(name=name) + + +ops.NotDifferentiable('MPILocalRank') + + +def _allreduce(tensor, name=None): + """An op which sums an input tensor over all the MPI processes. + + The reduction operation is keyed by the name of the op. The tensor type and + shape must be the same on all MPI processes for a given name. The reduction + will not start until all processes are ready to send and receive the tensor. + + Returns: + A tensor of the same shape and type as `tensor`, summed across all + processes. + """ + return MPI_LIB.mpi_allreduce(tensor, name=name) + + +ops.NotDifferentiable('MPIAllreduce') + + +def allgather(tensor, name=None): + """An op which concatenates the input tensor with the same input tensor on + all other MPI processes. + + The concatenation is done on the first dimension, so the input tensors on the + different processes must have the same rank and shape, except for the first + dimension, which is allowed to be different. + + Returns: + A tensor of the same type as `tensor`, concatenated on dimension zero + across all processes. The shape is identical to the input shape, except for + the first dimension, which may be greater and is the sum of all first + dimensions of the tensors in different MPI processes. + """ + # Specify that first allgather is to collect the tensor gather sizes, + # indicated by passing in a scalar (0-D tensor) of value 0 + sizes_flag = tf.constant(0, dtype=tf.int64, name='size_flag_const') + my_size = tf.slice( + tf.shape(tensor, out_type=tf.int64), [0], [1], name='size_slice') + if name is None: + name = 'allgather' + sizing_name = '{}_sizing'.format(name) + sizes = MPI_LIB.mpi_allgather(my_size, sizes_flag, name=sizing_name) + return MPI_LIB.mpi_allgather(tensor, sizes, name=name) + + +ops.NotDifferentiable('MPIAllgather') diff --git a/tensorflow/contrib/mpi_collectives/ring.cc b/tensorflow/contrib/mpi_collectives/ring.cc new file mode 100644 index 0000000000000000000000000000000000000000..d93233eb210b80df10fd9c2c7975ce77112d18a2 --- /dev/null +++ b/tensorflow/contrib/mpi_collectives/ring.cc @@ -0,0 +1,80 @@ +/* Copyright 2017 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#ifdef TENSORFLOW_USE_MPI + +#define EIGEN_USE_THREADS + +#include "tensorflow/contrib/mpi_collectives/ring.h" + +namespace tensorflow { +namespace contrib { +namespace mpi { + +using CPUDevice = Eigen::ThreadPoolDevice; + +extern template MPI_Datatype MPIType(); +extern template MPI_Datatype MPIType(); +extern template MPI_Datatype MPIType(); +extern template DataType TensorFlowDataType(); +extern template DataType TensorFlowDataType(); +extern template DataType TensorFlowDataType(); + +// Generate all necessary specializations for RingAllreduce. +template Status RingAllreduce(OpKernelContext*, const Tensor*, + Tensor*, Tensor*); +template Status RingAllreduce(OpKernelContext*, + const Tensor*, Tensor*, + Tensor*); +template Status RingAllreduce(OpKernelContext*, const Tensor*, + Tensor*, Tensor*); + +// Generate all necessary specializations for RingAllgather. +template Status RingAllgather(OpKernelContext*, const Tensor*, + const std::vector&, + Tensor*); +template Status RingAllgather(OpKernelContext*, + const Tensor*, + const std::vector&, + Tensor*); +template Status RingAllgather(OpKernelContext*, const Tensor*, + const std::vector&, + Tensor*); + +// Copy data on a CPU using a straight-forward memcpy. +template <> +void CopyTensorData(void* dst, void* src, size_t size) { + std::memcpy(dst, src, size); +}; + +// Accumulate values on a CPU. +#define GENERATE_ACCUMULATE(type) \ + template <> \ + void AccumulateTensorData(type * dst, type * src, \ + size_t size) { \ + for (unsigned int i = 0; i < size; i++) { \ + dst[i] += src[i]; \ + } \ + }; +GENERATE_ACCUMULATE(int); +GENERATE_ACCUMULATE(long long); +GENERATE_ACCUMULATE(float); +#undef GENERATE_ACCUMULATE + +} // namespace mpi +} // namespace contrib +} // namespace tensorflow + +#endif // TENSORFLOW_USE_MPI diff --git a/tensorflow/contrib/mpi_collectives/ring.cu.cc b/tensorflow/contrib/mpi_collectives/ring.cu.cc new file mode 100644 index 0000000000000000000000000000000000000000..2f3eef366a9a3c10e59cd5298fc1626e1094dff8 --- /dev/null +++ b/tensorflow/contrib/mpi_collectives/ring.cu.cc @@ -0,0 +1,117 @@ +/* Copyright 2016 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#ifdef TENSORFLOW_USE_MPI + +#if GOOGLE_CUDA + +#define EIGEN_USE_GPU + +#include "tensorflow/contrib/mpi_collectives/ring.h" + +namespace tensorflow { +namespace contrib { +namespace mpi { + +using CPUDevice = Eigen::ThreadPoolDevice; + +template <> +MPI_Datatype MPIType() { + return MPI_FLOAT; +}; +template <> +MPI_Datatype MPIType() { + return MPI_INT; +}; +template <> +MPI_Datatype MPIType() { + return MPI_LONG_LONG; +}; + +template <> +DataType TensorFlowDataType() { + return DT_FLOAT; +}; +template <> +DataType TensorFlowDataType() { + return DT_INT32; +}; +template <> +DataType TensorFlowDataType() { + return DT_INT64; +}; + +// Generate all necessary specializations for RingAllreduce. +template Status RingAllreduce(OpKernelContext*, const Tensor*, + Tensor*, Tensor*); +template Status RingAllreduce(OpKernelContext*, + const Tensor*, Tensor*, + Tensor*); +template Status RingAllreduce(OpKernelContext*, const Tensor*, + Tensor*, Tensor*); + +// Generate all necessary specializations for RingAllgather. +template Status RingAllgather(OpKernelContext*, const Tensor*, + const std::vector&, + Tensor*); +template Status RingAllgather(OpKernelContext*, + const Tensor*, + const std::vector&, + Tensor*); +template Status RingAllgather(OpKernelContext*, const Tensor*, + const std::vector&, + Tensor*); + +// Synchronously copy data on the GPU, using a different stream than the default +// and than TensorFlow to avoid synchronizing on operations unrelated to the +// allreduce. +template <> +void CopyTensorData(void* dst, void* src, size_t size) { + auto stream = CudaStreamForMPI(); + cudaMemcpyAsync(dst, src, size, cudaMemcpyDeviceToDevice, stream); + cudaStreamSynchronize(stream); +}; + +// Elementwise accumulation kernel for GPU. +template +__global__ void elemwise_accum(T* out, const T* in, const size_t N) { + for (size_t i = blockIdx.x * blockDim.x + threadIdx.x; i < N; + i += blockDim.x * gridDim.x) { + out[i] += in[i]; + } +} + +// Synchronously accumulate tensors on the GPU, using a different stream than +// the default and than TensorFlow to avoid synchronizing on operations +// unrelated to the allreduce. +#define GENERATE_ACCUMULATE(type) \ + template <> \ + void AccumulateTensorData(type * dst, type * src, \ + size_t size) { \ + auto stream = CudaStreamForMPI(); \ + elemwise_accum<<<32, 256, 0, stream>>>(dst, src, size); \ + cudaStreamSynchronize(stream); \ + }; +GENERATE_ACCUMULATE(int); +GENERATE_ACCUMULATE(long long); +GENERATE_ACCUMULATE(float); +#undef GENERATE_ACCUMULATE + +} // namespace mpi +} // namespace contrib +} // namespace tensorflow +#endif // GOOGLE_CUDA + +#endif // TENSORFLOW_USE_MPI diff --git a/tensorflow/contrib/mpi_collectives/ring.h b/tensorflow/contrib/mpi_collectives/ring.h new file mode 100644 index 0000000000000000000000000000000000000000..cae57ce60eb09509af69f8ccab9eacedea361548 --- /dev/null +++ b/tensorflow/contrib/mpi_collectives/ring.h @@ -0,0 +1,327 @@ +/* Copyright 2016 The TensorFlow Authors. All Rights Reserved. + +Licensed under the Apache License, Version 2.0 (the "License"); +you may not use this file except in compliance with the License. +You may obtain a copy of the License at + + http://www.apache.org/licenses/LICENSE-2.0 + +Unless required by applicable law or agreed to in writing, software +distributed under the License is distributed on an "AS IS" BASIS, +WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +See the License for the specific language governing permissions and +limitations under the License. +==============================================================================*/ + +#ifndef TENSORFLOW_CONTRIB_MPI_H_ +#define TENSORFLOW_CONTRIB_MPI_H_ + +#ifdef TENSORFLOW_USE_MPI + +#include "tensorflow/core/framework/op.h" +#include "tensorflow/core/framework/op_kernel.h" +#include "tensorflow/core/framework/shape_inference.h" + +#include "third_party/eigen3/unsupported/Eigen/CXX11/Tensor" +#include "tensorflow/core/framework/tensor_types.h" + +#if GOOGLE_CUDA +#include "cuda_runtime.h" +#endif + +// Needed to avoid header issues with C++-supporting MPI implementations +#define OMPI_SKIP_MPICXX +#include "third_party/mpi/mpi.h" + +#define TAG_TENSOR 12 + +namespace tensorflow { +namespace contrib { +namespace mpi { + +using CPUDevice = Eigen::ThreadPoolDevice; +using GPUDevice = Eigen::GpuDevice; + +// Convert from templated types to values we can pass to MPI. +template +MPI_Datatype MPIType(); + +// Convert from templated types to TensorFlow data types. +template +DataType TensorFlowDataType(); + +#define MPI_REQUIRES_OK(MPI_STATUS) \ + if ((MPI_STATUS) != MPI_SUCCESS) { \ + return errors::Unknown("MPI operation failed unexpectedly."); \ + } + +// Copy data from one tensor to another tensor. +// This uses a custom CUDA stream on GPU, which is necessary to overlay the +// backpropagation computations with the allreduce. +template +void CopyTensorData(void* destination, void* source, size_t size); + +// Add a tensor into another tensor, accumulating in place. +// This uses a custom CUDA stream on GPU, which is necessary to overlay the +// backpropagation computations with the allreduce. +template +void AccumulateTensorData(T* destination, T* source, size_t size); + +// We need to get the right stream for doing CUDA memory transfers and +// operations, which is possibly different from the standard TensorFlow stream. +#if GOOGLE_CUDA +cudaStream_t CudaStreamForMPI(); +#endif + +/* Perform a ring allreduce on the data. Allocate the necessary output tensor + * and store it in the output parameter. + * + * Assumes that all MPI processes are doing an allreduce of the same tensor, + * with the same dimensions. + * + * A ring allreduce is a bandwidth-optimal way to do an allreduce. To do the + * allreduce, the nodes involved are arranged in a ring: + * + * .--0--. + * / \ + * 3 1 + * \ / + * *--2--* + * + * Each node always sends to the next clockwise node in the ring, and receives + * from the previous one. + * + * The allreduce is done in two parts: a scatter-reduce and an allgather. In + * the scatter reduce, a reduction is done, so that each node ends up with a + * chunk of the final output tensor which has contributions from all other + * nodes. In the allgather, those chunks are distributed among all the nodes, + * so that all nodes have the entire output tensor. + * + * Both of these operations are done by dividing the input tensor into N + * evenly sized chunks (where N is the number of nodes in the ring). + * + * The scatter-reduce is done in N-1 steps. In the ith step, node j will send + * the (j - i)th chunk and receive the (j - i - 1)th chunk, adding it in to + * its existing data for that chunk. For example, in the first iteration with + * the ring depicted above, you will have the following transfers: + * + * Segment 0: Node 0 --> Node 1 + * Segment 1: Node 1 --> Node 2 + * Segment 2: Node 2 --> Node 3 + * Segment 3: Node 3 --> Node 0 + * + * In the second iteration, you'll have the following transfers: + * + * Segment 0: Node 1 --> Node 2 + * Segment 1: Node 2 --> Node 3 + * Segment 2: Node 3 --> Node 0 + * Segment 3: Node 0 --> Node 1 + * + * After this iteration, Node 2 has 3 of the four contributions to Segment 0. + * The last iteration has the following transfers: + * + * Segment 0: Node 2 --> Node 3 + * Segment 1: Node 3 --> Node 0 + * Segment 2: Node 0 --> Node 1 + * Segment 3: Node 1 --> Node 2 + * + * After this iteration, Node 3 has the fully accumulated Segment 0; Node 0 + * has the fully accumulated Segment 1; and so on. The scatter-reduce is + * complete. + * + * Next, the allgather distributes these fully accumululated chunks across all + * nodes. Communication proceeds in the same ring, once again in N-1 steps. At + * the ith step, node j will send chunk (j - i + 1) and receive chunk (j - i). + * For example, at the first iteration, the following transfers will occur: + * + * Segment 0: Node 3 --> Node 0 + * Segment 1: Node 0 --> Node 1 + * Segment 2: Node 1 --> Node 2 + * Segment 3: Node 2 --> Node 3 + * + * After the first iteration, Node 0 will have a fully accumulated Segment 0 + * (from Node 3) and Segment 1. In the next iteration, Node 0 will send its + * just-received Segment 0 onward to Node 1, and receive Segment 3 from Node 3. + * After this has continued for N - 1 iterations, all nodes will have a the + * fully accumulated tensor. + * + * Each node will do (N-1) sends for the scatter-reduce and (N-1) sends for the + * allgather. Each send will contain K / N bytes, if there are K bytes in the + * original tensor on every node. Thus, each node sends and receives 2K(N - 1)/N + * bytes of data, and the performance of the allreduce (assuming no latency in + * connections) is constrained by the slowest interconnect between the nodes. + * + */ +template +Status RingAllreduce(OpKernelContext* context, const Tensor* input, + Tensor* temp, Tensor* output) { + // Acquire MPI size and rank + int n, r; + MPI_REQUIRES_OK(MPI_Comm_size(MPI_COMM_WORLD, &n)); + MPI_REQUIRES_OK(MPI_Comm_rank(MPI_COMM_WORLD, &r)); + + T* buffer = (T*)output->tensor_data().data(); + + CopyTensorData((void*)buffer, (void*)input->tensor_data().data(), + output->tensor_data().size()); + + // Calculate segment sizes and segment ends + const size_t elements_to_reduce = input->NumElements(); + const size_t segment_size = elements_to_reduce / n; + std::vector segment_sizes(n, segment_size); + + const size_t residual = elements_to_reduce % n; + for (size_t i = 0; i < residual; ++i) { + segment_sizes[i]++; + } + + std::vector segment_starts(n); + segment_starts[0] = 0; + for (size_t i = 1; i < segment_starts.size(); ++i) { + segment_starts[i] = segment_starts[i - 1] + segment_sizes[i - 1]; + } + + assert(segment_starts[n - 1] + segment_sizes[n - 1] == elements_to_reduce); + + T* segment_recv = (T*)temp->tensor_data().data(); + + // Receive from your left neighbor with wrap-around + const size_t recv_from = ((r - 1) + n) % n; + + // Send to your right neighbor with wrap-around + const size_t send_to = (r + 1) % n; + + MPI_Status recv_status; + MPI_Request recv_req; + + // Now start ring. At every step, for every rank, we iterate through + // segments with wraparound and send and recv from our neighbors and reduce + // locally. At the i'th iteration, rank r, sends segment (r-i) and receives + // segment (r-i-1). + for (int i = 0; i < n - 1; i++) { + const size_t send_seg_id = ((r - i) + n) % n; + const size_t recv_seg_id = ((r - i - 1) + n) % n; + + T* segment_send = &(buffer[segment_starts[send_seg_id]]); + + MPI_REQUIRES_OK(MPI_Irecv(segment_recv, segment_sizes[recv_seg_id], + MPIType(), recv_from, TAG_TENSOR, + MPI_COMM_WORLD, &recv_req)); + + MPI_REQUIRES_OK(MPI_Send(segment_send, segment_sizes[send_seg_id], + MPIType(), send_to, TAG_TENSOR, + MPI_COMM_WORLD)); + + T* segment_update = &(buffer[segment_starts[recv_seg_id]]); + + // Wait for recv to complete before reduction + MPI_REQUIRES_OK(MPI_Wait(&recv_req, &recv_status)); + + const size_t recv_seg_size = segment_sizes[recv_seg_id]; + AccumulateTensorData(segment_update, segment_recv, + recv_seg_size); + } + + // Now start pipelined ring allgather. At every step, for every rank, we + // iterate through segments with wraparound and send and recv from our + // neighbors. At the i'th iteration, rank r, sends segment (r-i+1) and + // receives segment (r-i). + for (size_t i = 0; i < n - 1; ++i) { + const size_t send_seg_id = ((r - i + 1) + n) % n; + const size_t recv_seg_id = ((r - i) + n) % n; + + // Segment to send - at every iteration we send segment (r-i+1) + T* segment_send = &(buffer[segment_starts[send_seg_id]]); + + // Segment to recv - at every iteration we receive segment (r-i) + T* segment_recv = &(buffer[segment_starts[recv_seg_id]]); + + MPI_REQUIRES_OK(MPI_Sendrecv( + segment_send, segment_sizes[send_seg_id], MPIType(), send_to, + TAG_TENSOR, segment_recv, segment_sizes[recv_seg_id], MPIType(), + recv_from, TAG_TENSOR, MPI_COMM_WORLD, &recv_status)); + } + + return Status::OK(); +} + +// Perform a ring allgather on a Tensor. Other ranks may allgather with a +// tensor which differs in the first dimension only; all other dimensions must +// be the same. +// +// For more information on the ring allgather, read the documentation for the +// ring allreduce, which includes a ring allgather. +template +Status RingAllgather(OpKernelContext* context, const Tensor* input, + const std::vector& sizes, Tensor* output) { + // Acquire MPI size and rank + int n, r; + MPI_REQUIRES_OK(MPI_Comm_size(MPI_COMM_WORLD, &n)); + MPI_REQUIRES_OK(MPI_Comm_rank(MPI_COMM_WORLD, &r)); + + assert(sizes.size() == n); + assert(input->dim_size(0) == sizes[r]); + + // Compute number of elements in every "row". We can't compute number of + // elements in every chunks, because those chunks are variable length. + size_t elements_per_row = 1; + for (int i = 1; i < input->shape().dims(); i++) { + elements_per_row *= input->dim_size(i); + } + + // Copy data from input tensor to correct place in output tensor. + std::vector segment_starts(n); + segment_starts[0] = 0; + for (int i = 1; i < n; i++) { + segment_starts[i] = segment_starts[i - 1] + elements_per_row * sizes[i - 1]; + } + size_t offset = segment_starts[r]; + + // Copy data to the right offset for this rank. + T* buffer = (T*)output->tensor_data().data(); + CopyTensorData((void*)(buffer + offset), + (void*)input->tensor_data().data(), + elements_per_row * sizes[r] * sizeof(T)); + + // Receive from your left neighbor with wrap-around + const size_t recv_from = ((r - 1) + n) % n; + + // Send to your right neighbor with wrap-around + const size_t send_to = (r + 1) % n; + + // Perform a ring allgather. At every step, for every rank, we iterate + // through segments with wraparound and send and recv from our neighbors. + // At the i'th iteration, rank r, sends segment (r-i) and receives segment + // (r-1-i). + MPI_Status recv_status; + for (size_t i = 0; i < n - 1; ++i) { + const size_t send_seg_id = ((r - i) + n) % n; + const size_t recv_seg_id = ((r - i - 1) + n) % n; + + // Segment to send - at every iteration we send segment (r-i) + size_t offset_send = segment_starts[send_seg_id]; + size_t rows_send = sizes[send_seg_id]; + T* segment_send = &(buffer[offset_send]); + + // Segment to recv - at every iteration we receive segment (r-1-i) + size_t offset_recv = segment_starts[recv_seg_id]; + size_t rows_recv = sizes[recv_seg_id]; + T* segment_recv = &(buffer[offset_recv]); + + MPI_REQUIRES_OK(MPI_Sendrecv( + segment_send, elements_per_row * rows_send, MPIType(), send_to, + TAG_TENSOR, segment_recv, elements_per_row * rows_recv, MPIType(), + recv_from, TAG_TENSOR, MPI_COMM_WORLD, &recv_status)); + } + + return Status::OK(); +} + +} // namespace mpi +} // namespace contrib +} // namespace tensorflow + +#endif // TENSORFLOW_USE_MPI + +#undef TENSORFLOW_CONTRIB_MPI_H_ +#endif // TENSORFLOW_CONTRIB_MPI_H_ diff --git a/tensorflow/contrib/nccl/BUILD b/tensorflow/contrib/nccl/BUILD index 7cfdf0f607033479f03827ca20f4ad609e51cdfe..62996d1fd83f46145e9a1b773b1be57e27903127 100644 --- a/tensorflow/contrib/nccl/BUILD +++ b/tensorflow/contrib/nccl/BUILD @@ -19,17 +19,18 @@ load("//tensorflow:tensorflow.bzl", "cuda_py_test") load("@local_config_cuda//cuda:build_defs.bzl", "if_cuda") load("//tensorflow:tensorflow.bzl", "tf_kernel_library") load("//tensorflow:tensorflow.bzl", "tf_custom_op_py_library") +load("//tensorflow:tensorflow.bzl", "if_not_windows_cuda") tf_custom_op_library( name = "python/ops/_nccl_ops.so", srcs = [ "ops/nccl_ops.cc", ], - gpu_srcs = [ + gpu_srcs = if_not_windows_cuda([ "kernels/nccl_manager.cc", "kernels/nccl_manager.h", "kernels/nccl_ops.cc", - ], + ]), deps = if_cuda([ "@local_config_nccl//:nccl", "//tensorflow/core:gpu_headers_lib", diff --git a/tensorflow/contrib/nccl/python/ops/nccl_ops.py b/tensorflow/contrib/nccl/python/ops/nccl_ops.py index 029b01412d96ca03d4ecf7bf4d7d9872864e3ddc..fa597cf3efcf915311047f3a483772c45cc314fd 100644 --- a/tensorflow/contrib/nccl/python/ops/nccl_ops.py +++ b/tensorflow/contrib/nccl/python/ops/nccl_ops.py @@ -63,12 +63,12 @@ def _all_sum_grad(op, grad): Raises: LookupError: If `reduction` is not `sum`. """ - if op.get_attr('reduction') != 'sum': + if op.get_attr('reduction') != b'sum': raise LookupError('No gradient defined for NcclAllReduce except sum.') _check_device(grad, expected=op.device) num_devices = op.get_attr('num_devices') - shared_name = op.get_attr('shared_name') + '_grad' + shared_name = op.get_attr('shared_name') + b'_grad' with ops.device(op.device): return gen_nccl_ops.nccl_all_reduce( @@ -162,7 +162,7 @@ def _reduce_sum_grad(op, grad): Raises: LookupError: If the reduction attribute of op is not `sum`. """ - if op.get_attr('reduction') != 'sum': + if op.get_attr('reduction') != b'sum': raise LookupError('No gradient defined for NcclReduce except sum.') _check_device(grad, expected=op.device) diff --git a/tensorflow/contrib/opt/__init__.py b/tensorflow/contrib/opt/__init__.py index 157ed6a278bb699724d3854426d780a3a58823db..3e63e99030c46c254625ca8fdccce614cd60e8b0 100644 --- a/tensorflow/contrib/opt/__init__.py +++ b/tensorflow/contrib/opt/__init__.py @@ -22,17 +22,18 @@ from __future__ import print_function from tensorflow.contrib.opt.python.training.adamax import * from tensorflow.contrib.opt.python.training.addsign import * from tensorflow.contrib.opt.python.training.drop_stale_gradient_optimizer import * +from tensorflow.contrib.opt.python.training.elastic_average_optimizer import * from tensorflow.contrib.opt.python.training.external_optimizer import * +from tensorflow.contrib.opt.python.training.ggt import * from tensorflow.contrib.opt.python.training.lazy_adam_optimizer import * +from tensorflow.contrib.opt.python.training.model_average_optimizer import * from tensorflow.contrib.opt.python.training.moving_average_optimizer import * from tensorflow.contrib.opt.python.training.multitask_optimizer_wrapper import * from tensorflow.contrib.opt.python.training.nadam_optimizer import * from tensorflow.contrib.opt.python.training.weight_decay_optimizers import * from tensorflow.contrib.opt.python.training.powersign import * from tensorflow.contrib.opt.python.training.variable_clipping_optimizer import * -from tensorflow.contrib.opt.python.training.elastic_average_optimizer import * -from tensorflow.contrib.opt.python.training.model_average_optimizer import * -from tensorflow.contrib.opt.python.training.ggt import * +from tensorflow.contrib.opt.python.training.weight_decay_optimizers import * # pylint: enable=wildcard-import from tensorflow.python.util.all_util import remove_undocumented diff --git a/tensorflow/contrib/opt/python/training/addsign_test.py b/tensorflow/contrib/opt/python/training/addsign_test.py index 08d45ed73f3ae4b580d7078272e79fef22ef67c5..628a735e721d2f0c594dd59b5193499dfd7da02e 100644 --- a/tensorflow/contrib/opt/python/training/addsign_test.py +++ b/tensorflow/contrib/opt/python/training/addsign_test.py @@ -214,7 +214,7 @@ class AddSignTest(test.TestCase): # Run 7 steps of AddSign # first 4 steps with positive gradient # last 3 steps with negative gradient (sign(gm) should be -1) - for t in range(1, 4): + for t in range(1, 8): if t < 5: update.run() else: @@ -222,7 +222,7 @@ class AddSignTest(test.TestCase): var0_np, m0 = addsign_update_numpy( var0_np, - grads0_np, + grads0_np if t < 5 else -grads0_np, m0, learning_rate, alpha=alpha, @@ -232,7 +232,7 @@ class AddSignTest(test.TestCase): ) var1_np, m1 = addsign_update_numpy( var1_np, - grads1_np, + grads1_np if t < 5 else -grads1_np, m1, learning_rate, alpha=alpha, diff --git a/tensorflow/contrib/opt/python/training/ggt.py b/tensorflow/contrib/opt/python/training/ggt.py index 928c453517f825ed2d305ec498d07ac29c065f1a..cae952d8f50acbc3a176697fb3989db6c9ac3e9b 100644 --- a/tensorflow/contrib/opt/python/training/ggt.py +++ b/tensorflow/contrib/opt/python/training/ggt.py @@ -33,7 +33,7 @@ class GGTOptimizer(optimizer_v2.OptimizerV2): GGT has an advantage over sgd and adam on large models with poor conditioning, for example language models and CNNs, - see [ABCHSZZ 2018]([pdf](https://arxiv.org/pdf/1806.02958.pdf)). + see [[ABCHSZZ 2018]](https://arxiv.org/pdf/1806.02958.pdf). """ def __init__(self, diff --git a/tensorflow/contrib/opt/python/training/powersign_test.py b/tensorflow/contrib/opt/python/training/powersign_test.py index 5214082dd66f00eadadad71d50f7e00b178b8c10..0bcf5d230a8b7b5b778d233a79922dc34449f8dd 100644 --- a/tensorflow/contrib/opt/python/training/powersign_test.py +++ b/tensorflow/contrib/opt/python/training/powersign_test.py @@ -216,7 +216,7 @@ class PowerSignTest(test.TestCase): self.assertAllClose([1.0, 2.0], var0.eval()) self.assertAllClose([3.0, 4.0], var1.eval()) - # Run 3 steps of powersign + # Run 7 steps of powersign # first 4 steps with positive gradient # last 3 steps with negative gradient (sign(gm) should be -1) for t in range(1, 8): diff --git a/tensorflow/contrib/opt/python/training/weight_decay_optimizers.py b/tensorflow/contrib/opt/python/training/weight_decay_optimizers.py index 8aa40aeb45d4ec15140bdfc5ebd824e8aa08d8d9..b9cf40eb7b2d11c98b93c51213145ca4e2670318 100644 --- a/tensorflow/contrib/opt/python/training/weight_decay_optimizers.py +++ b/tensorflow/contrib/opt/python/training/weight_decay_optimizers.py @@ -19,13 +19,13 @@ from __future__ import division from __future__ import print_function from tensorflow.python.framework import ops -from tensorflow.python.training import optimizer from tensorflow.python.ops import control_flow_ops +from tensorflow.python.ops import resource_variable_ops +from tensorflow.python.ops import state_ops from tensorflow.python.training import adam from tensorflow.python.training import momentum as momentum_opt +from tensorflow.python.training import optimizer from tensorflow.python.util.tf_export import tf_export -from tensorflow.python.ops import state_ops -from tensorflow.python.ops import resource_variable_ops class DecoupledWeightDecayExtension(object): @@ -65,7 +65,7 @@ class DecoupledWeightDecayExtension(object): Args: weight_decay: A `Tensor` or a floating point value, the factor by which a variable is decayed in the update step. - decay_var_list: Optional list or tuple or set of `Variable` objects to + **kwargs: Optional list or tuple or set of `Variable` objects to decay. """ self._decay_var_list = None # is set in minimize or apply_gradients @@ -85,6 +85,28 @@ class DecoupledWeightDecayExtension(object): If decay_var_list is None, all variables in var_list are decayed. For more information see the documentation of Optimizer.minimize. + + Args: + loss: A `Tensor` containing the value to minimize. + global_step: Optional `Variable` to increment by one after the + variables have been updated. + var_list: Optional list or tuple of `Variable` objects to update to + minimize `loss`. Defaults to the list of variables collected in + the graph under the key `GraphKeys.TRAINABLE_VARIABLES`. + gate_gradients: How to gate the computation of gradients. Can be + `GATE_NONE`, `GATE_OP`, or `GATE_GRAPH`. + aggregation_method: Specifies the method used to combine gradient terms. + Valid values are defined in the class `AggregationMethod`. + colocate_gradients_with_ops: If True, try colocating gradients with + the corresponding op. + name: Optional name for the returned operation. + grad_loss: Optional. A `Tensor` holding the gradient computed for `loss`. + decay_var_list: Optional list of decay variables. + + Returns: + An Operation that updates the variables in `var_list`. If `global_step` + was not `None`, that operation also increments `global_step`. + """ self._decay_var_list = set(decay_var_list) if decay_var_list else False return super(DecoupledWeightDecayExtension, self).minimize( @@ -103,6 +125,19 @@ class DecoupledWeightDecayExtension(object): are decayed. For more information see the documentation of Optimizer.apply_gradients. + + Args: + grads_and_vars: List of (gradient, variable) pairs as returned by + `compute_gradients()`. + global_step: Optional `Variable` to increment by one after the + variables have been updated. + name: Optional name for the returned operation. Default to the + name passed to the `Optimizer` constructor. + decay_var_list: Optional list of decay variables. + + Returns: + An `Operation` that applies the specified gradients. If `global_step` + was not None, that operation also increments `global_step`. """ self._decay_var_list = set(decay_var_list) if decay_var_list else False return super(DecoupledWeightDecayExtension, self).apply_gradients( @@ -197,6 +232,7 @@ def extend_with_decoupled_weight_decay(base_optimizer): A new optimizer class that inherits from DecoupledWeightDecayExtension and base_optimizer. """ + class OptimizerWithDecoupledWeightDecay(DecoupledWeightDecayExtension, base_optimizer): """Base_optimizer with decoupled weight decay. diff --git a/tensorflow/contrib/opt/python/training/weight_decay_optimizers_test.py b/tensorflow/contrib/opt/python/training/weight_decay_optimizers_test.py index 74d1cdbbdac8724518937d141a976abf9fec6ce3..76d8a5697acb79e7748175c4a81dfdd85807dd49 100644 --- a/tensorflow/contrib/opt/python/training/weight_decay_optimizers_test.py +++ b/tensorflow/contrib/opt/python/training/weight_decay_optimizers_test.py @@ -20,6 +20,7 @@ from __future__ import print_function import numpy as np +from tensorflow.contrib.opt.python.training import weight_decay_optimizers from tensorflow.python.eager import context from tensorflow.python.framework import constant_op from tensorflow.python.framework import dtypes @@ -29,7 +30,6 @@ from tensorflow.python.ops import resource_variable_ops from tensorflow.python.ops import variables from tensorflow.python.platform import test from tensorflow.python.training import adam -from tensorflow.contrib.opt.python.training import weight_decay_optimizers WEIGHT_DECAY = 0.01 @@ -91,7 +91,6 @@ class WeightDecayOptimizerTest(test.TestCase): opt = optimizer() update = opt.apply_gradients(zip([grads0, grads1], [var0, var1])) - if not context.executing_eagerly(): with ops.Graph().as_default(): # Shouldn't return non-slot variables from other graphs. @@ -171,9 +170,9 @@ class ExtendWithWeightDecayTest(WeightDecayOptimizerTest): @staticmethod def get_optimizer(): - AdamW = weight_decay_optimizers.extend_with_decoupled_weight_decay( + adamw = weight_decay_optimizers.extend_with_decoupled_weight_decay( adam.AdamOptimizer) - return AdamW(WEIGHT_DECAY) + return adamw(WEIGHT_DECAY) def testBasic(self): self.doTest(self.get_optimizer, adamw_update_numpy, "Adam", "m", @@ -185,6 +184,5 @@ class ExtendWithWeightDecayTest(WeightDecayOptimizerTest): use_resource=True) - if __name__ == "__main__": test.main() diff --git a/tensorflow/contrib/optimizer_v2/optimizer_v2.py b/tensorflow/contrib/optimizer_v2/optimizer_v2.py index c6f3bd6ee18fa353944e2fc303573894933f5b27..8c11d8bcfdf76bc12e13ffb58f917978e966476e 100644 --- a/tensorflow/contrib/optimizer_v2/optimizer_v2.py +++ b/tensorflow/contrib/optimizer_v2/optimizer_v2.py @@ -766,7 +766,8 @@ class OptimizerV2(optimizer_v1.Optimizer): # *after* loss() is evaluated, so we know what loss reduction it uses. if scale_loss_by_num_towers is None: scale_loss_by_num_towers = ( - distribute_lib.get_loss_reduction() == "mean") + distribute_lib.get_loss_reduction() == + variable_scope.VariableAggregation.MEAN) if scale_loss_by_num_towers: num_towers = distribute_lib.get_distribution_strategy().num_towers if num_towers > 1: @@ -784,7 +785,8 @@ class OptimizerV2(optimizer_v1.Optimizer): # Scale loss for number of towers (non-callable-loss case). if scale_loss_by_num_towers is None: scale_loss_by_num_towers = ( - distribute_lib.get_loss_reduction() == "mean") + distribute_lib.get_loss_reduction() == + variable_scope.VariableAggregation.MEAN) if scale_loss_by_num_towers: num_towers = distribute_lib.get_distribution_strategy().num_towers if num_towers > 1: @@ -896,7 +898,8 @@ class OptimizerV2(optimizer_v1.Optimizer): def _distributed_apply(self, distribution, grads_and_vars, global_step, name): """`apply_gradients` for use with a `DistributionStrategy`.""" - reduced_grads = distribution.batch_reduce("sum", grads_and_vars) + reduced_grads = distribution.batch_reduce( + variable_scope.VariableAggregation.SUM, grads_and_vars) var_list = [v for _, v in grads_and_vars] grads_and_vars = zip(reduced_grads, var_list) diff --git a/tensorflow/contrib/proto/BUILD b/tensorflow/contrib/proto/BUILD index 3e9b1a0b8d8ec7c3c5fe5d1f2cf896dbb6c3de72..d45622174f21d9d104321cd56e47a3d120bcc03d 100644 --- a/tensorflow/contrib/proto/BUILD +++ b/tensorflow/contrib/proto/BUILD @@ -19,9 +19,7 @@ py_library( py_library( name = "proto_pip", - data = [ - "//tensorflow/contrib/proto/python/kernel_tests:test_messages", - ] + if_static( + data = if_static( [], otherwise = ["//tensorflow/contrib/proto/python/kernel_tests:libtestexample.so"], ), diff --git a/tensorflow/contrib/proto/python/kernel_tests/BUILD b/tensorflow/contrib/proto/python/kernel_tests/BUILD index a380a131f86abc8dd921a123afdb964bf6c2466c..3c6fde23d2a321d5b9ba24b610f08a2646f5b2d1 100644 --- a/tensorflow/contrib/proto/python/kernel_tests/BUILD +++ b/tensorflow/contrib/proto/python/kernel_tests/BUILD @@ -4,45 +4,18 @@ licenses(["notice"]) # Apache 2.0 exports_files(["LICENSE"]) -# Much of the work in this BUILD file actually happens in the corresponding -# build_defs.bzl, which creates an individual testcase for each example .pbtxt -# file in this directory. -# -load(":build_defs.bzl", "decode_proto_test_suite") -load(":build_defs.bzl", "encode_proto_test_suite") - -# This expands to a tf_py_test for each test file. -# It defines the test_suite :decode_proto_op_tests. -decode_proto_test_suite( - name = "decode_proto_tests", - examples = glob(["*.pbtxt"]), -) - -# This expands to a tf_py_test for each test file. -# It defines the test_suite :encode_proto_op_tests. -encode_proto_test_suite( - name = "encode_proto_tests", - examples = glob(["*.pbtxt"]), -) - -# Below here are tests that are not tied to an example text proto. -filegroup( - name = "test_messages", - srcs = glob(["*.pbtxt"]), -) - load("//tensorflow:tensorflow.bzl", "tf_py_test") load("//tensorflow:tensorflow.bzl", "tf_cc_shared_object") load("//tensorflow/core:platform/default/build_config_root.bzl", "if_static") load("//tensorflow/core:platform/default/build_config.bzl", "tf_proto_library") tf_py_test( - name = "decode_proto_fail_test", + name = "decode_proto_op_test", size = "small", - srcs = ["decode_proto_fail_test.py"], + srcs = ["decode_proto_op_test.py"], additional_deps = [ + ":decode_proto_op_test_base", ":py_test_deps", - "//third_party/py/numpy", "//tensorflow/contrib/proto:proto", "//tensorflow/contrib/proto/python/ops:decode_proto_op_py", ], @@ -56,20 +29,63 @@ tf_py_test( ], ) +tf_py_test( + name = "encode_proto_op_test", + size = "small", + srcs = ["encode_proto_op_test.py"], + additional_deps = [ + ":encode_proto_op_test_base", + ":py_test_deps", + "//tensorflow/contrib/proto:proto", + "//tensorflow/contrib/proto/python/ops:decode_proto_op_py", + "//tensorflow/contrib/proto/python/ops:encode_proto_op_py", + ], + data = if_static( + [], + otherwise = [":libtestexample.so"], + ), + tags = [ + "no_pip", # TODO(b/78026780) + "no_windows", # TODO(b/78028010) + ], +) + +py_library( + name = "proto_op_test_base", + testonly = 1, + srcs = ["proto_op_test_base.py"], + deps = [ + ":test_example_proto_py", + "//tensorflow/python:client_testlib", + ], +) + py_library( - name = "test_case", - srcs = ["test_case.py"], - deps = ["//tensorflow/python:client_testlib"], + name = "decode_proto_op_test_base", + testonly = 1, + srcs = ["decode_proto_op_test_base.py"], + deps = [ + ":proto_op_test_base", + ":test_example_proto_py", + "//third_party/py/numpy", + "@absl_py//absl/testing:parameterized", + ], ) py_library( - name = "py_test_deps", + name = "encode_proto_op_test_base", + testonly = 1, + srcs = ["encode_proto_op_test_base.py"], deps = [ - ":test_case", + ":proto_op_test_base", ":test_example_proto_py", + "//third_party/py/numpy", + "@absl_py//absl/testing:parameterized", ], ) +py_library(name = "py_test_deps") + tf_proto_library( name = "test_example_proto", srcs = ["test_example.proto"], diff --git a/tensorflow/contrib/proto/python/kernel_tests/build_defs.bzl b/tensorflow/contrib/proto/python/kernel_tests/build_defs.bzl deleted file mode 100644 index f425601691e21b36914f340d53ccadf9b4e3641f..0000000000000000000000000000000000000000 --- a/tensorflow/contrib/proto/python/kernel_tests/build_defs.bzl +++ /dev/null @@ -1,89 +0,0 @@ -"""BUILD rules for generating file-driven proto test cases. - -The decode_proto_test_suite() and encode_proto_test_suite() rules take a list -of text protos and generates a tf_py_test() for each one. -""" - -load("//tensorflow:tensorflow.bzl", "tf_py_test") -load("//tensorflow:tensorflow.bzl", "register_extension_info") -load("//tensorflow/core:platform/default/build_config_root.bzl", "if_static") - -def _test_name(test, path): - return "%s_%s_test" % (test, path.split("/")[-1].split(".")[0]) - -def decode_proto_test_suite(name, examples): - """Build the decode_proto py_test for each test filename.""" - for test_filename in examples: - tf_py_test( - name = _test_name("decode_proto", test_filename), - srcs = ["decode_proto_op_test.py"], - size = "small", - data = [test_filename] + if_static( - [], - otherwise = [":libtestexample.so"], - ), - main = "decode_proto_op_test.py", - args = [ - "--message_text_file=\"%s/%s\"" % (native.package_name(), test_filename), - ], - additional_deps = [ - ":py_test_deps", - "//third_party/py/numpy", - "//tensorflow/contrib/proto:proto", - "//tensorflow/contrib/proto/python/ops:decode_proto_op_py", - ], - tags = [ - "no_pip", # TODO(b/78026780) - "no_windows", # TODO(b/78028010) - ], - ) - native.test_suite( - name = name, - tests = [":" + _test_name("decode_proto", test_filename) - for test_filename in examples], - ) - -def encode_proto_test_suite(name, examples): - """Build the encode_proto py_test for each test filename.""" - for test_filename in examples: - tf_py_test( - name = _test_name("encode_proto", test_filename), - srcs = ["encode_proto_op_test.py"], - size = "small", - data = [test_filename] + if_static( - [], - otherwise = [":libtestexample.so"], - ), - main = "encode_proto_op_test.py", - args = [ - "--message_text_file=\"%s/%s\"" % (native.package_name(), test_filename), - ], - additional_deps = [ - ":py_test_deps", - "//third_party/py/numpy", - "//tensorflow/contrib/proto:proto", - "//tensorflow/contrib/proto/python/ops:decode_proto_op_py", - "//tensorflow/contrib/proto/python/ops:encode_proto_op_py", - ], - tags = [ - "no_pip", # TODO(b/78026780) - "no_windows", # TODO(b/78028010) - ], - ) - native.test_suite( - name = name, - tests = [":" + _test_name("encode_proto", test_filename) - for test_filename in examples], - ) - -register_extension_info( - extension_name = "decode_proto_test_suite", - label_regex_map = { - "deps": "deps:decode_example_.*", - }) - -register_extension_info( - extension_name = "encode_proto_test_suite", - label_regex_map = { - "deps": "deps:encode_example_.*", - }) diff --git a/tensorflow/contrib/proto/python/kernel_tests/decode_proto_fail_test.py b/tensorflow/contrib/proto/python/kernel_tests/decode_proto_fail_test.py deleted file mode 100644 index 5298342ee79b08a50b13ce8715e891a332efb3bc..0000000000000000000000000000000000000000 --- a/tensorflow/contrib/proto/python/kernel_tests/decode_proto_fail_test.py +++ /dev/null @@ -1,68 +0,0 @@ -# ============================================================================= -# Copyright 2018 The TensorFlow Authors. All Rights Reserved. -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. -# ============================================================================= - -# Python3 preparedness imports. -from __future__ import absolute_import -from __future__ import division -from __future__ import print_function - -import numpy as np - -from tensorflow.contrib.proto.python.kernel_tests import test_case -from tensorflow.contrib.proto.python.ops import decode_proto_op -from tensorflow.python.framework import dtypes -from tensorflow.python.framework import errors -from tensorflow.python.platform import test - - -class DecodeProtoFailTest(test_case.ProtoOpTestCase): - """Test failure cases for DecodeToProto.""" - - def _TestCorruptProtobuf(self, sanitize): - """Test failure cases for DecodeToProto.""" - - # The goal here is to check the error reporting. - # Testing against a variety of corrupt protobufs is - # done by fuzzing. - corrupt_proto = 'This is not a binary protobuf' - - # Numpy silently truncates the strings if you don't specify dtype=object. - batch = np.array(corrupt_proto, dtype=object) - msg_type = 'tensorflow.contrib.proto.TestCase' - field_names = ['sizes'] - field_types = [dtypes.int32] - - with self.test_session() as sess: - ctensor, vtensor = decode_proto_op.decode_proto( - batch, - message_type=msg_type, - field_names=field_names, - output_types=field_types, - sanitize=sanitize) - with self.assertRaisesRegexp(errors.DataLossError, - 'Unable to parse binary protobuf' - '|Failed to consume entire buffer'): - _ = sess.run([ctensor] + vtensor) - - def testCorrupt(self): - self._TestCorruptProtobuf(sanitize=False) - - def testSanitizerCorrupt(self): - self._TestCorruptProtobuf(sanitize=True) - - -if __name__ == '__main__': - test.main() diff --git a/tensorflow/contrib/proto/python/kernel_tests/decode_proto_op_test.py b/tensorflow/contrib/proto/python/kernel_tests/decode_proto_op_test.py index d1c13c82bc264bc8bcc721eb68ee3916f32ef7a8..934035ec4c97e04846f493817d4b4ed65db94f14 100644 --- a/tensorflow/contrib/proto/python/kernel_tests/decode_proto_op_test.py +++ b/tensorflow/contrib/proto/python/kernel_tests/decode_proto_op_test.py @@ -13,287 +13,22 @@ # See the License for the specific language governing permissions and # limitations under the License. # ============================================================================= -"""Table-driven test for decode_proto op. +"""Tests for decode_proto op.""" -This test is run once with each of the *.TestCase.pbtxt files -in the test directory. -""" # Python3 preparedness imports. from __future__ import absolute_import from __future__ import division from __future__ import print_function -import numpy as np - -from google.protobuf import text_format - -from tensorflow.contrib.proto.python.kernel_tests import test_case -from tensorflow.contrib.proto.python.kernel_tests import test_example_pb2 +from tensorflow.contrib.proto.python.kernel_tests import decode_proto_op_test_base as test_base from tensorflow.contrib.proto.python.ops import decode_proto_op -from tensorflow.python.framework import dtypes -from tensorflow.python.platform import flags from tensorflow.python.platform import test -FLAGS = flags.FLAGS - -flags.DEFINE_string('message_text_file', None, - 'A file containing a text serialized TestCase protobuf.') - - -class DecodeProtoOpTest(test_case.ProtoOpTestCase): - - def _compareValues(self, fd, vs, evs): - """Compare lists/arrays of field values.""" - - if len(vs) != len(evs): - self.fail('Field %s decoded %d outputs, expected %d' % - (fd.name, len(vs), len(evs))) - for i, ev in enumerate(evs): - # Special case fuzzy match for float32. TensorFlow seems to mess with - # MAX_FLT slightly and the test doesn't work otherwise. - # TODO(nix): ask on TF list about why MAX_FLT doesn't pass through. - if fd.cpp_type == fd.CPPTYPE_FLOAT: - # Numpy isclose() is better than assertIsClose() which uses an absolute - # value comparison. - self.assertTrue( - np.isclose(vs[i], ev), 'expected %r, actual %r' % (ev, vs[i])) - elif fd.cpp_type == fd.CPPTYPE_STRING: - # In Python3 string tensor values will be represented as bytes, so we - # reencode the proto values to match that. - self.assertEqual(vs[i], ev.encode('ascii')) - else: - # Doubles and other types pass through unscathed. - self.assertEqual(vs[i], ev) - - def _compareRepeatedPrimitiveValue(self, batch_shape, sizes, fields, - field_dict): - """Compare protos of type RepeatedPrimitiveValue. - - Args: - batch_shape: the shape of the input tensor of serialized messages. - sizes: int matrix of repeat counts returned by decode_proto - fields: list of test_example_pb2.FieldSpec (types and expected values) - field_dict: map from field names to decoded numpy tensors of values - """ - - # Check that expected values match. - for field in fields: - values = field_dict[field.name] - self.assertEqual(dtypes.as_dtype(values.dtype), field.dtype) - - fd = field.expected.DESCRIPTOR.fields_by_name[field.name] - - # Values has the same shape as the input plus an extra - # dimension for repeats. - self.assertEqual(list(values.shape)[:-1], batch_shape) - - # Nested messages are represented as TF strings, requiring - # some special handling. - if field.name == 'message_value': - vs = [] - for buf in values.flat: - msg = test_example_pb2.PrimitiveValue() - msg.ParseFromString(buf) - vs.append(msg) - evs = getattr(field.expected, field.name) - if len(vs) != len(evs): - self.fail('Field %s decoded %d outputs, expected %d' % - (fd.name, len(vs), len(evs))) - for v, ev in zip(vs, evs): - self.assertEqual(v, ev) - continue - - # This can be a little confusing. For testing we are using - # RepeatedPrimitiveValue in two ways: it's the proto that we - # decode for testing, and it's used in the expected value as a - # union type. The two cases are slightly different: this is the - # second case. - # We may be fetching the uint64_value from the test proto, but - # in the expected proto we store it in the int64_value field - # because TensorFlow doesn't support unsigned int64. - tf_type_to_primitive_value_field = { - dtypes.float32: - 'float_value', - dtypes.float64: - 'double_value', - dtypes.int32: - 'int32_value', - dtypes.uint8: - 'uint8_value', - dtypes.int8: - 'int8_value', - dtypes.string: - 'string_value', - dtypes.int64: - 'int64_value', - dtypes.bool: - 'bool_value', - # Unhandled TensorFlow types: - # DT_INT16 DT_COMPLEX64 DT_QINT8 DT_QUINT8 DT_QINT32 - # DT_BFLOAT16 DT_QINT16 DT_QUINT16 DT_UINT16 - } - tf_field_name = tf_type_to_primitive_value_field.get(field.dtype) - if tf_field_name is None: - self.fail('Unhandled tensorflow type %d' % field.dtype) - - self._compareValues(fd, values.flat, - getattr(field.expected, tf_field_name)) - - def _runDecodeProtoTests(self, fields, case_sizes, batch_shape, batch, - message_type, message_format, sanitize, - force_disordered=False): - """Run decode tests on a batch of messages. - - Args: - fields: list of test_example_pb2.FieldSpec (types and expected values) - case_sizes: expected sizes array - batch_shape: the shape of the input tensor of serialized messages - batch: list of serialized messages - message_type: descriptor name for messages - message_format: format of messages, 'text' or 'binary' - sanitize: whether to sanitize binary protobuf inputs - force_disordered: whether to force fields encoded out of order. - """ - - if force_disordered: - # Exercise code path that handles out-of-order fields by prepending extra - # fields with tag numbers higher than any real field. Note that this won't - # work with sanitization because that forces reserialization using a - # trusted decoder and encoder. - assert not sanitize - extra_fields = test_example_pb2.ExtraFields() - extra_fields.string_value = 'IGNORE ME' - extra_fields.bool_value = False - extra_msg = extra_fields.SerializeToString() - batch = [extra_msg + msg for msg in batch] - - # Numpy silently truncates the strings if you don't specify dtype=object. - batch = np.array(batch, dtype=object) - batch = np.reshape(batch, batch_shape) - - field_names = [f.name for f in fields] - output_types = [f.dtype for f in fields] - - with self.test_session() as sess: - sizes, vtensor = decode_proto_op.decode_proto( - batch, - message_type=message_type, - field_names=field_names, - output_types=output_types, - message_format=message_format, - sanitize=sanitize) - - vlist = sess.run([sizes] + vtensor) - sizes = vlist[0] - # Values is a list of tensors, one for each field. - value_tensors = vlist[1:] - - # Check that the repeat sizes are correct. - self.assertTrue( - np.all(np.array(sizes.shape) == batch_shape + [len(field_names)])) - - # Check that the decoded sizes match the expected sizes. - self.assertEqual(len(sizes.flat), len(case_sizes)) - self.assertTrue( - np.all(sizes.flat == np.array( - case_sizes, dtype=np.int32))) - - field_dict = dict(zip(field_names, value_tensors)) - - self._compareRepeatedPrimitiveValue(batch_shape, sizes, fields, - field_dict) - - def testBinary(self): - with open(FLAGS.message_text_file, 'r') as fp: - case = text_format.Parse(fp.read(), test_example_pb2.TestCase()) - - batch = [primitive.SerializeToString() for primitive in case.primitive] - self._runDecodeProtoTests( - case.field, - case.sizes, - list(case.shape), - batch, - 'tensorflow.contrib.proto.RepeatedPrimitiveValue', - 'binary', - sanitize=False) - - def testBinaryDisordered(self): - with open(FLAGS.message_text_file, 'r') as fp: - case = text_format.Parse(fp.read(), test_example_pb2.TestCase()) - - batch = [primitive.SerializeToString() for primitive in case.primitive] - self._runDecodeProtoTests( - case.field, - case.sizes, - list(case.shape), - batch, - 'tensorflow.contrib.proto.RepeatedPrimitiveValue', - 'binary', - sanitize=False, - force_disordered=True) - - def testPacked(self): - with open(FLAGS.message_text_file, 'r') as fp: - case = text_format.Parse(fp.read(), test_example_pb2.TestCase()) - - # Now try with the packed serialization. - # We test the packed representations by loading the same test cases - # using PackedPrimitiveValue instead of RepeatedPrimitiveValue. - # To do this we rely on the text format being the same for packed and - # unpacked fields, and reparse the test message using the packed version - # of the proto. - packed_batch = [ - # Note: float_format='.17g' is necessary to ensure preservation of - # doubles and floats in text format. - text_format.Parse( - text_format.MessageToString( - primitive, float_format='.17g'), - test_example_pb2.PackedPrimitiveValue()).SerializeToString() - for primitive in case.primitive - ] - - self._runDecodeProtoTests( - case.field, - case.sizes, - list(case.shape), - packed_batch, - 'tensorflow.contrib.proto.PackedPrimitiveValue', - 'binary', - sanitize=False) - - def testText(self): - with open(FLAGS.message_text_file, 'r') as fp: - case = text_format.Parse(fp.read(), test_example_pb2.TestCase()) - - # Note: float_format='.17g' is necessary to ensure preservation of - # doubles and floats in text format. - text_batch = [ - text_format.MessageToString( - primitive, float_format='.17g') for primitive in case.primitive - ] - - self._runDecodeProtoTests( - case.field, - case.sizes, - list(case.shape), - text_batch, - 'tensorflow.contrib.proto.RepeatedPrimitiveValue', - 'text', - sanitize=False) - def testSanitizerGood(self): - with open(FLAGS.message_text_file, 'r') as fp: - case = text_format.Parse(fp.read(), test_example_pb2.TestCase()) +class DecodeProtoOpTest(test_base.DecodeProtoOpTestBase): - batch = [primitive.SerializeToString() for primitive in case.primitive] - self._runDecodeProtoTests( - case.field, - case.sizes, - list(case.shape), - batch, - 'tensorflow.contrib.proto.RepeatedPrimitiveValue', - 'binary', - sanitize=True) + def __init__(self, methodName='runTest'): # pylint: disable=invalid-name + super(DecodeProtoOpTest, self).__init__(decode_proto_op, methodName) if __name__ == '__main__': diff --git a/tensorflow/contrib/proto/python/kernel_tests/decode_proto_op_test_base.py b/tensorflow/contrib/proto/python/kernel_tests/decode_proto_op_test_base.py new file mode 100644 index 0000000000000000000000000000000000000000..5f7f510352d23c756c226d8826611ba6d2c8de31 --- /dev/null +++ b/tensorflow/contrib/proto/python/kernel_tests/decode_proto_op_test_base.py @@ -0,0 +1,310 @@ +# ============================================================================= +# Copyright 2018 The TensorFlow Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================= +"""Tests for decode_proto op.""" + +# Python3 preparedness imports. +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + +from absl.testing import parameterized +import numpy as np + + +from google.protobuf import text_format + +from tensorflow.contrib.proto.python.kernel_tests import proto_op_test_base as test_base +from tensorflow.contrib.proto.python.kernel_tests import test_example_pb2 +from tensorflow.python.framework import dtypes +from tensorflow.python.framework import errors + + +class DecodeProtoOpTestBase(test_base.ProtoOpTestBase, parameterized.TestCase): + """Base class for testing proto decoding ops.""" + + def __init__(self, decode_module, methodName='runTest'): # pylint: disable=invalid-name + """DecodeProtoOpTestBase initializer. + + Args: + decode_module: a module containing the `decode_proto_op` method + methodName: the name of the test method (same as for test.TestCase) + """ + + super(DecodeProtoOpTestBase, self).__init__(methodName) + self._decode_module = decode_module + + def _compareValues(self, fd, vs, evs): + """Compare lists/arrays of field values.""" + + if len(vs) != len(evs): + self.fail('Field %s decoded %d outputs, expected %d' % + (fd.name, len(vs), len(evs))) + for i, ev in enumerate(evs): + # Special case fuzzy match for float32. TensorFlow seems to mess with + # MAX_FLT slightly and the test doesn't work otherwise. + # TODO(nix): ask on TF list about why MAX_FLT doesn't pass through. + if fd.cpp_type == fd.CPPTYPE_FLOAT: + # Numpy isclose() is better than assertIsClose() which uses an absolute + # value comparison. + self.assertTrue( + np.isclose(vs[i], ev), 'expected %r, actual %r' % (ev, vs[i])) + elif fd.cpp_type == fd.CPPTYPE_STRING: + # In Python3 string tensor values will be represented as bytes, so we + # reencode the proto values to match that. + self.assertEqual(vs[i], ev.encode('ascii')) + else: + # Doubles and other types pass through unscathed. + self.assertEqual(vs[i], ev) + + def _compareProtos(self, batch_shape, sizes, fields, field_dict): + """Compare protos of type TestValue. + + Args: + batch_shape: the shape of the input tensor of serialized messages. + sizes: int matrix of repeat counts returned by decode_proto + fields: list of test_example_pb2.FieldSpec (types and expected values) + field_dict: map from field names to decoded numpy tensors of values + """ + + # Check that expected values match. + for field in fields: + values = field_dict[field.name] + self.assertEqual(dtypes.as_dtype(values.dtype), field.dtype) + + fd = field.value.DESCRIPTOR.fields_by_name[field.name] + + # Values has the same shape as the input plus an extra + # dimension for repeats. + self.assertEqual(list(values.shape)[:-1], batch_shape) + + # Nested messages are represented as TF strings, requiring + # some special handling. + if field.name == 'message_value': + vs = [] + for buf in values.flat: + msg = test_example_pb2.PrimitiveValue() + msg.ParseFromString(buf) + vs.append(msg) + evs = getattr(field.value, field.name) + if len(vs) != len(evs): + self.fail('Field %s decoded %d outputs, expected %d' % + (fd.name, len(vs), len(evs))) + for v, ev in zip(vs, evs): + self.assertEqual(v, ev) + continue + + # This can be a little confusing. For testing we are using TestValue in + # two ways: it's the proto that we decode for testing, and it's used in + # the expected value as a union type. + # + # The two cases are slightly different: this is the second case. We may be + # fetching the uint64_value from the test proto, but in the expected proto + # we store it in the int64_value field because TensorFlow doesn't support + # unsigned int64. + tf_type_to_primitive_value_field = { + dtypes.float32: + 'float_value', + dtypes.float64: + 'double_value', + dtypes.int32: + 'int32_value', + dtypes.uint8: + 'uint8_value', + dtypes.int8: + 'int8_value', + dtypes.string: + 'string_value', + dtypes.int64: + 'int64_value', + dtypes.bool: + 'bool_value', + # Unhandled TensorFlow types: + # DT_INT16 DT_COMPLEX64 DT_QINT8 DT_QUINT8 DT_QINT32 + # DT_BFLOAT16 DT_QINT16 DT_QUINT16 DT_UINT16 + } + tf_field_name = tf_type_to_primitive_value_field.get(field.dtype) + if tf_field_name is None: + self.fail('Unhandled tensorflow type %d' % field.dtype) + + self._compareValues(fd, values.flat, + getattr(field.value, tf_field_name)) + + def _runDecodeProtoTests(self, fields, case_sizes, batch_shape, batch, + message_type, message_format, sanitize, + force_disordered=False): + """Run decode tests on a batch of messages. + + Args: + fields: list of test_example_pb2.FieldSpec (types and expected values) + case_sizes: expected sizes array + batch_shape: the shape of the input tensor of serialized messages + batch: list of serialized messages + message_type: descriptor name for messages + message_format: format of messages, 'text' or 'binary' + sanitize: whether to sanitize binary protobuf inputs + force_disordered: whether to force fields encoded out of order. + """ + + if force_disordered: + # Exercise code path that handles out-of-order fields by prepending extra + # fields with tag numbers higher than any real field. Note that this won't + # work with sanitization because that forces reserialization using a + # trusted decoder and encoder. + assert not sanitize + extra_fields = test_example_pb2.ExtraFields() + extra_fields.string_value = 'IGNORE ME' + extra_fields.bool_value = False + extra_msg = extra_fields.SerializeToString() + batch = [extra_msg + msg for msg in batch] + + # Numpy silently truncates the strings if you don't specify dtype=object. + batch = np.array(batch, dtype=object) + batch = np.reshape(batch, batch_shape) + + field_names = [f.name for f in fields] + output_types = [f.dtype for f in fields] + + with self.test_session() as sess: + sizes, vtensor = self._decode_module.decode_proto( + batch, + message_type=message_type, + field_names=field_names, + output_types=output_types, + message_format=message_format, + sanitize=sanitize) + + vlist = sess.run([sizes] + vtensor) + sizes = vlist[0] + # Values is a list of tensors, one for each field. + value_tensors = vlist[1:] + + # Check that the repeat sizes are correct. + self.assertTrue( + np.all(np.array(sizes.shape) == batch_shape + [len(field_names)])) + + # Check that the decoded sizes match the expected sizes. + self.assertEqual(len(sizes.flat), len(case_sizes)) + self.assertTrue( + np.all(sizes.flat == np.array( + case_sizes, dtype=np.int32))) + + field_dict = dict(zip(field_names, value_tensors)) + + self._compareProtos(batch_shape, sizes, fields, field_dict) + + @parameterized.named_parameters(*test_base.ProtoOpTestBase.named_parameters()) + def testBinary(self, case): + batch = [value.SerializeToString() for value in case.values] + self._runDecodeProtoTests( + case.fields, + case.sizes, + list(case.shapes), + batch, + 'tensorflow.contrib.proto.TestValue', + 'binary', + sanitize=False) + + @parameterized.named_parameters(*test_base.ProtoOpTestBase.named_parameters()) + def testBinaryDisordered(self, case): + batch = [value.SerializeToString() for value in case.values] + self._runDecodeProtoTests( + case.fields, + case.sizes, + list(case.shapes), + batch, + 'tensorflow.contrib.proto.TestValue', + 'binary', + sanitize=False, + force_disordered=True) + + @parameterized.named_parameters(*test_base.ProtoOpTestBase.named_parameters()) + def testPacked(self, case): + # Now try with the packed serialization. + # + # We test the packed representations by loading the same test case using + # PackedTestValue instead of TestValue. To do this we rely on the text + # format being the same for packed and unpacked fields, and reparse the + # test message using the packed version of the proto. + packed_batch = [ + # Note: float_format='.17g' is necessary to ensure preservation of + # doubles and floats in text format. + text_format.Parse( + text_format.MessageToString( + value, float_format='.17g'), + test_example_pb2.PackedTestValue()).SerializeToString() + for value in case.values + ] + + self._runDecodeProtoTests( + case.fields, + case.sizes, + list(case.shapes), + packed_batch, + 'tensorflow.contrib.proto.PackedTestValue', + 'binary', + sanitize=False) + + @parameterized.named_parameters(*test_base.ProtoOpTestBase.named_parameters()) + def testText(self, case): + # Note: float_format='.17g' is necessary to ensure preservation of + # doubles and floats in text format. + text_batch = [ + text_format.MessageToString( + value, float_format='.17g') for value in case.values + ] + + self._runDecodeProtoTests( + case.fields, + case.sizes, + list(case.shapes), + text_batch, + 'tensorflow.contrib.proto.TestValue', + 'text', + sanitize=False) + + @parameterized.named_parameters(*test_base.ProtoOpTestBase.named_parameters()) + def testSanitizerGood(self, case): + batch = [value.SerializeToString() for value in case.values] + self._runDecodeProtoTests( + case.fields, + case.sizes, + list(case.shapes), + batch, + 'tensorflow.contrib.proto.TestValue', + 'binary', + sanitize=True) + + @parameterized.parameters((False), (True)) + def testCorruptProtobuf(self, sanitize): + corrupt_proto = 'This is not a binary protobuf' + + # Numpy silently truncates the strings if you don't specify dtype=object. + batch = np.array(corrupt_proto, dtype=object) + msg_type = 'tensorflow.contrib.proto.TestCase' + field_names = ['sizes'] + field_types = [dtypes.int32] + + with self.test_session() as sess: + ctensor, vtensor = self._decode_module.decode_proto( + batch, + message_type=msg_type, + field_names=field_names, + output_types=field_types, + sanitize=sanitize) + with self.assertRaisesRegexp(errors.DataLossError, + 'Unable to parse binary protobuf' + '|Failed to consume entire buffer'): + _ = sess.run([ctensor] + vtensor) diff --git a/tensorflow/contrib/proto/python/kernel_tests/defaut_values.TestCase.pbtxt b/tensorflow/contrib/proto/python/kernel_tests/defaut_values.TestCase.pbtxt deleted file mode 100644 index 4e316819077c7dbb28beefd4dc260568f26da680..0000000000000000000000000000000000000000 --- a/tensorflow/contrib/proto/python/kernel_tests/defaut_values.TestCase.pbtxt +++ /dev/null @@ -1,94 +0,0 @@ -primitive { - # No fields specified, so we get all defaults -} -shape: 1 -sizes: 0 -field { - name: "double_default" - dtype: DT_DOUBLE - expected { double_value: 1.0 } -} -sizes: 0 -field { - name: "float_default" - dtype: DT_DOUBLE # Try casting the float field to double. - expected { double_value: 2.0 } -} -sizes: 0 -field { - name: "int64_default" - dtype: DT_INT64 - expected { int64_value: 3 } -} -sizes: 0 -field { - name: "uint64_default" - dtype: DT_INT64 - expected { int64_value: 4 } -} -sizes: 0 -field { - name: "int32_default" - dtype: DT_INT32 - expected { int32_value: 5 } -} -sizes: 0 -field { - name: "fixed64_default" - dtype: DT_INT64 - expected { int64_value: 6 } -} -sizes: 0 -field { - name: "fixed32_default" - dtype: DT_INT32 - expected { int32_value: 7 } -} -sizes: 0 -field { - name: "bool_default" - dtype: DT_BOOL - expected { bool_value: true } -} -sizes: 0 -field { - name: "string_default" - dtype: DT_STRING - expected { string_value: "a" } -} -sizes: 0 -field { - name: "bytes_default" - dtype: DT_STRING - expected { string_value: "a longer default string" } -} -sizes: 0 -field { - name: "uint32_default" - dtype: DT_INT32 - expected { int32_value: -1 } -} -sizes: 0 -field { - name: "sfixed32_default" - dtype: DT_INT32 - expected { int32_value: 10 } -} -sizes: 0 -field { - name: "sfixed64_default" - dtype: DT_INT64 - expected { int64_value: 11 } -} -sizes: 0 -field { - name: "sint32_default" - dtype: DT_INT32 - expected { int32_value: 12 } -} -sizes: 0 -field { - name: "sint64_default" - dtype: DT_INT64 - expected { int64_value: 13 } -} diff --git a/tensorflow/contrib/proto/python/kernel_tests/encode_proto_op_test.py b/tensorflow/contrib/proto/python/kernel_tests/encode_proto_op_test.py index 30e58e6336dc66830418c7cd2b3111a851d691b6..fc5cd25d43be1df2480630396c39f7a83e0eb57a 100644 --- a/tensorflow/contrib/proto/python/kernel_tests/encode_proto_op_test.py +++ b/tensorflow/contrib/proto/python/kernel_tests/encode_proto_op_test.py @@ -13,167 +13,24 @@ # See the License for the specific language governing permissions and # limitations under the License. # ============================================================================= -"""Table-driven test for encode_proto op. +"""Tests for encode_proto op.""" -This test is run once with each of the *.TestCase.pbtxt files -in the test directory. - -It tests that encode_proto is a lossless inverse of decode_proto -(for the specified fields). -""" # Python3 readiness boilerplate from __future__ import absolute_import from __future__ import division from __future__ import print_function -import numpy as np - -from google.protobuf import text_format - -from tensorflow.contrib.proto.python.kernel_tests import test_case -from tensorflow.contrib.proto.python.kernel_tests import test_example_pb2 +from tensorflow.contrib.proto.python.kernel_tests import encode_proto_op_test_base as test_base from tensorflow.contrib.proto.python.ops import decode_proto_op from tensorflow.contrib.proto.python.ops import encode_proto_op -from tensorflow.python.framework import dtypes -from tensorflow.python.ops import array_ops -from tensorflow.python.platform import flags from tensorflow.python.platform import test -FLAGS = flags.FLAGS - -flags.DEFINE_string('message_text_file', None, - 'A file containing a text serialized TestCase protobuf.') - - -class EncodeProtoOpTest(test_case.ProtoOpTestCase): - - def testBadInputs(self): - # Invalid field name - with self.test_session(): - with self.assertRaisesOpError('Unknown field: non_existent_field'): - encode_proto_op.encode_proto( - sizes=[[1]], - values=[np.array([[0.0]], dtype=np.int32)], - message_type='tensorflow.contrib.proto.RepeatedPrimitiveValue', - field_names=['non_existent_field']).eval() - - # Incorrect types. - with self.test_session(): - with self.assertRaisesOpError( - 'Incompatible type for field double_value.'): - encode_proto_op.encode_proto( - sizes=[[1]], - values=[np.array([[0.0]], dtype=np.int32)], - message_type='tensorflow.contrib.proto.RepeatedPrimitiveValue', - field_names=['double_value']).eval() - - # Incorrect shapes of sizes. - with self.test_session(): - with self.assertRaisesOpError( - r'sizes should be batch_size \+ \[len\(field_names\)\]'): - sizes = array_ops.placeholder(dtypes.int32) - values = array_ops.placeholder(dtypes.float64) - encode_proto_op.encode_proto( - sizes=sizes, - values=[values], - message_type='tensorflow.contrib.proto.RepeatedPrimitiveValue', - field_names=['double_value']).eval(feed_dict={ - sizes: [[[0, 0]]], - values: [[0.0]] - }) - - # Inconsistent shapes of values. - with self.test_session(): - with self.assertRaisesOpError( - 'Values must match up to the last dimension'): - sizes = array_ops.placeholder(dtypes.int32) - values1 = array_ops.placeholder(dtypes.float64) - values2 = array_ops.placeholder(dtypes.int32) - (encode_proto_op.encode_proto( - sizes=[[1, 1]], - values=[values1, values2], - message_type='tensorflow.contrib.proto.RepeatedPrimitiveValue', - field_names=['double_value', 'int32_value']).eval(feed_dict={ - values1: [[0.0]], - values2: [[0], [0]] - })) - - def _testRoundtrip(self, in_bufs, message_type, fields): - - field_names = [f.name for f in fields] - out_types = [f.dtype for f in fields] - - with self.test_session() as sess: - sizes, field_tensors = decode_proto_op.decode_proto( - in_bufs, - message_type=message_type, - field_names=field_names, - output_types=out_types) - - out_tensors = encode_proto_op.encode_proto( - sizes, - field_tensors, - message_type=message_type, - field_names=field_names) - - out_bufs, = sess.run([out_tensors]) - - # Check that the re-encoded tensor has the same shape. - self.assertEqual(in_bufs.shape, out_bufs.shape) - - # Compare the input and output. - for in_buf, out_buf in zip(in_bufs.flat, out_bufs.flat): - in_obj = test_example_pb2.RepeatedPrimitiveValue() - in_obj.ParseFromString(in_buf) - - out_obj = test_example_pb2.RepeatedPrimitiveValue() - out_obj.ParseFromString(out_buf) - - # Check that the deserialized objects are identical. - self.assertEqual(in_obj, out_obj) - - # Check that the input and output serialized messages are identical. - # If we fail here, there is a difference in the serialized - # representation but the new serialization still parses. This could - # be harmless (a change in map ordering?) or it could be bad (e.g. - # loss of packing in the encoding). - self.assertEqual(in_buf, out_buf) - - def testRoundtrip(self): - with open(FLAGS.message_text_file, 'r') as fp: - case = text_format.Parse(fp.read(), test_example_pb2.TestCase()) - - in_bufs = [primitive.SerializeToString() for primitive in case.primitive] - - # np.array silently truncates strings if you don't specify dtype=object. - in_bufs = np.reshape(np.array(in_bufs, dtype=object), list(case.shape)) - return self._testRoundtrip( - in_bufs, 'tensorflow.contrib.proto.RepeatedPrimitiveValue', case.field) - - def testRoundtripPacked(self): - with open(FLAGS.message_text_file, 'r') as fp: - case = text_format.Parse(fp.read(), test_example_pb2.TestCase()) - # Now try with the packed serialization. - # We test the packed representations by loading the same test cases - # using PackedPrimitiveValue instead of RepeatedPrimitiveValue. - # To do this we rely on the text format being the same for packed and - # unpacked fields, and reparse the test message using the packed version - # of the proto. - in_bufs = [ - # Note: float_format='.17g' is necessary to ensure preservation of - # doubles and floats in text format. - text_format.Parse( - text_format.MessageToString( - primitive, float_format='.17g'), - test_example_pb2.PackedPrimitiveValue()).SerializeToString() - for primitive in case.primitive - ] +class EncodeProtoOpTest(test_base.EncodeProtoOpTestBase): - # np.array silently truncates strings if you don't specify dtype=object. - in_bufs = np.reshape(np.array(in_bufs, dtype=object), list(case.shape)) - return self._testRoundtrip( - in_bufs, 'tensorflow.contrib.proto.PackedPrimitiveValue', case.field) + def __init__(self, methodName='runTest'): # pylint: disable=invalid-name + super(EncodeProtoOpTest, self).__init__(decode_proto_op, encode_proto_op, + methodName) if __name__ == '__main__': diff --git a/tensorflow/contrib/proto/python/kernel_tests/encode_proto_op_test_base.py b/tensorflow/contrib/proto/python/kernel_tests/encode_proto_op_test_base.py new file mode 100644 index 0000000000000000000000000000000000000000..07dfb924d3ede5bdb9b848c5eb0d3382ec053121 --- /dev/null +++ b/tensorflow/contrib/proto/python/kernel_tests/encode_proto_op_test_base.py @@ -0,0 +1,177 @@ +# ============================================================================= +# Copyright 2018 The TensorFlow Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================= +"""Table-driven test for encode_proto op. + +This test is run once with each of the *.TestCase.pbtxt files +in the test directory. + +It tests that encode_proto is a lossless inverse of decode_proto +(for the specified fields). +""" +# Python3 readiness boilerplate +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + +from absl.testing import parameterized +import numpy as np + +from google.protobuf import text_format + +from tensorflow.contrib.proto.python.kernel_tests import proto_op_test_base as test_base +from tensorflow.contrib.proto.python.kernel_tests import test_example_pb2 +from tensorflow.python.framework import dtypes +from tensorflow.python.ops import array_ops + + +class EncodeProtoOpTestBase(test_base.ProtoOpTestBase, parameterized.TestCase): + """Base class for testing proto encoding ops.""" + + def __init__(self, decode_module, encode_module, methodName='runTest'): # pylint: disable=invalid-name + """EncodeProtoOpTestBase initializer. + + Args: + decode_module: a module containing the `decode_proto_op` method + encode_module: a module containing the `encode_proto_op` method + methodName: the name of the test method (same as for test.TestCase) + """ + + super(EncodeProtoOpTestBase, self).__init__(methodName) + self._decode_module = decode_module + self._encode_module = encode_module + + def testBadInputs(self): + # Invalid field name + with self.test_session(): + with self.assertRaisesOpError('Unknown field: non_existent_field'): + self._encode_module.encode_proto( + sizes=[[1]], + values=[np.array([[0.0]], dtype=np.int32)], + message_type='tensorflow.contrib.proto.TestValue', + field_names=['non_existent_field']).eval() + + # Incorrect types. + with self.test_session(): + with self.assertRaisesOpError( + 'Incompatible type for field double_value.'): + self._encode_module.encode_proto( + sizes=[[1]], + values=[np.array([[0.0]], dtype=np.int32)], + message_type='tensorflow.contrib.proto.TestValue', + field_names=['double_value']).eval() + + # Incorrect shapes of sizes. + with self.test_session(): + with self.assertRaisesOpError( + r'sizes should be batch_size \+ \[len\(field_names\)\]'): + sizes = array_ops.placeholder(dtypes.int32) + values = array_ops.placeholder(dtypes.float64) + self._encode_module.encode_proto( + sizes=sizes, + values=[values], + message_type='tensorflow.contrib.proto.TestValue', + field_names=['double_value']).eval(feed_dict={ + sizes: [[[0, 0]]], + values: [[0.0]] + }) + + # Inconsistent shapes of values. + with self.test_session(): + with self.assertRaisesOpError( + 'Values must match up to the last dimension'): + sizes = array_ops.placeholder(dtypes.int32) + values1 = array_ops.placeholder(dtypes.float64) + values2 = array_ops.placeholder(dtypes.int32) + (self._encode_module.encode_proto( + sizes=[[1, 1]], + values=[values1, values2], + message_type='tensorflow.contrib.proto.TestValue', + field_names=['double_value', 'int32_value']).eval(feed_dict={ + values1: [[0.0]], + values2: [[0], [0]] + })) + + def _testRoundtrip(self, in_bufs, message_type, fields): + + field_names = [f.name for f in fields] + out_types = [f.dtype for f in fields] + + with self.test_session() as sess: + sizes, field_tensors = self._decode_module.decode_proto( + in_bufs, + message_type=message_type, + field_names=field_names, + output_types=out_types) + + out_tensors = self._encode_module.encode_proto( + sizes, + field_tensors, + message_type=message_type, + field_names=field_names) + + out_bufs, = sess.run([out_tensors]) + + # Check that the re-encoded tensor has the same shape. + self.assertEqual(in_bufs.shape, out_bufs.shape) + + # Compare the input and output. + for in_buf, out_buf in zip(in_bufs.flat, out_bufs.flat): + in_obj = test_example_pb2.TestValue() + in_obj.ParseFromString(in_buf) + + out_obj = test_example_pb2.TestValue() + out_obj.ParseFromString(out_buf) + + # Check that the deserialized objects are identical. + self.assertEqual(in_obj, out_obj) + + # Check that the input and output serialized messages are identical. + # If we fail here, there is a difference in the serialized + # representation but the new serialization still parses. This could + # be harmless (a change in map ordering?) or it could be bad (e.g. + # loss of packing in the encoding). + self.assertEqual(in_buf, out_buf) + + @parameterized.named_parameters(*test_base.ProtoOpTestBase.named_parameters()) + def testRoundtrip(self, case): + in_bufs = [value.SerializeToString() for value in case.values] + + # np.array silently truncates strings if you don't specify dtype=object. + in_bufs = np.reshape(np.array(in_bufs, dtype=object), list(case.shapes)) + return self._testRoundtrip( + in_bufs, 'tensorflow.contrib.proto.TestValue', case.fields) + + @parameterized.named_parameters(*test_base.ProtoOpTestBase.named_parameters()) + def testRoundtripPacked(self, case): + # Now try with the packed serialization. + # We test the packed representations by loading the same test cases using + # PackedTestValue instead of TestValue. To do this we rely on the text + # format being the same for packed and unpacked fields, and reparse the test + # message using the packed version of the proto. + in_bufs = [ + # Note: float_format='.17g' is necessary to ensure preservation of + # doubles and floats in text format. + text_format.Parse( + text_format.MessageToString( + value, float_format='.17g'), + test_example_pb2.PackedTestValue()).SerializeToString() + for value in case.values + ] + + # np.array silently truncates strings if you don't specify dtype=object. + in_bufs = np.reshape(np.array(in_bufs, dtype=object), list(case.shapes)) + return self._testRoundtrip( + in_bufs, 'tensorflow.contrib.proto.PackedTestValue', case.fields) diff --git a/tensorflow/contrib/proto/python/kernel_tests/minmax.TestCase.pbtxt b/tensorflow/contrib/proto/python/kernel_tests/minmax.TestCase.pbtxt deleted file mode 100644 index b170f89c0f00dd9dffd5785197bb3bfd1ca2cfee..0000000000000000000000000000000000000000 --- a/tensorflow/contrib/proto/python/kernel_tests/minmax.TestCase.pbtxt +++ /dev/null @@ -1,161 +0,0 @@ -primitive { - double_value: -1.7976931348623158e+308 - double_value: 2.2250738585072014e-308 - double_value: 1.7976931348623158e+308 - float_value: -3.402823466e+38 - float_value: 1.175494351e-38 - float_value: 3.402823466e+38 - int64_value: -9223372036854775808 - int64_value: 9223372036854775807 - uint64_value: 0 - uint64_value: 18446744073709551615 - int32_value: -2147483648 - int32_value: 2147483647 - fixed64_value: 0 - fixed64_value: 18446744073709551615 - fixed32_value: 0 - fixed32_value: 4294967295 - bool_value: false - bool_value: true - string_value: "" - string_value: "I refer to the infinite." - uint32_value: 0 - uint32_value: 4294967295 - sfixed32_value: -2147483648 - sfixed32_value: 2147483647 - sfixed64_value: -9223372036854775808 - sfixed64_value: 9223372036854775807 - sint32_value: -2147483648 - sint32_value: 2147483647 - sint64_value: -9223372036854775808 - sint64_value: 9223372036854775807 -} -shape: 1 -sizes: 3 -sizes: 3 -sizes: 2 -sizes: 2 -sizes: 2 -sizes: 2 -sizes: 2 -sizes: 2 -sizes: 2 -sizes: 2 -sizes: 2 -sizes: 2 -sizes: 2 -sizes: 2 -field { - name: "double_value" - dtype: DT_DOUBLE - expected { - double_value: -1.7976931348623158e+308 - double_value: 2.2250738585072014e-308 - double_value: 1.7976931348623158e+308 - } -} -field { - name: "float_value" - dtype: DT_FLOAT - expected { - float_value: -3.402823466e+38 - float_value: 1.175494351e-38 - float_value: 3.402823466e+38 - } -} -field { - name: "int64_value" - dtype: DT_INT64 - expected { - int64_value: -9223372036854775808 - int64_value: 9223372036854775807 - } -} -field { - name: "uint64_value" - dtype: DT_INT64 - expected { - int64_value: 0 - int64_value: -1 - } -} -field { - name: "int32_value" - dtype: DT_INT32 - expected { - int32_value: -2147483648 - int32_value: 2147483647 - } -} -field { - name: "fixed64_value" - dtype: DT_INT64 - expected { - int64_value: 0 - int64_value: -1 # unsigned is 18446744073709551615 - } -} -field { - name: "fixed32_value" - dtype: DT_INT32 - expected { - int32_value: 0 - int32_value: -1 # unsigned is 4294967295 - } -} -field { - name: "bool_value" - dtype: DT_BOOL - expected { - bool_value: false - bool_value: true - } -} -field { - name: "string_value" - dtype: DT_STRING - expected { - string_value: "" - string_value: "I refer to the infinite." - } -} -field { - name: "uint32_value" - dtype: DT_INT32 - expected { - int32_value: 0 - int32_value: -1 # unsigned is 4294967295 - } -} -field { - name: "sfixed32_value" - dtype: DT_INT32 - expected { - int32_value: -2147483648 - int32_value: 2147483647 - } -} -field { - name: "sfixed64_value" - dtype: DT_INT64 - expected { - int64_value: -9223372036854775808 - int64_value: 9223372036854775807 - } -} -field { - name: "sint32_value" - dtype: DT_INT32 - expected { - int32_value: -2147483648 - int32_value: 2147483647 - } -} -field { - name: "sint64_value" - dtype: DT_INT64 - expected { - int64_value: -9223372036854775808 - int64_value: 9223372036854775807 - } -} diff --git a/tensorflow/contrib/proto/python/kernel_tests/nested.TestCase.pbtxt b/tensorflow/contrib/proto/python/kernel_tests/nested.TestCase.pbtxt deleted file mode 100644 index c664e52851b5bb3c439544537ce6402fc7cf3362..0000000000000000000000000000000000000000 --- a/tensorflow/contrib/proto/python/kernel_tests/nested.TestCase.pbtxt +++ /dev/null @@ -1,16 +0,0 @@ -primitive { - message_value { - double_value: 23.5 - } -} -shape: 1 -sizes: 1 -field { - name: "message_value" - dtype: DT_STRING - expected { - message_value { - double_value: 23.5 - } - } -} diff --git a/tensorflow/contrib/proto/python/kernel_tests/optional.TestCase.pbtxt b/tensorflow/contrib/proto/python/kernel_tests/optional.TestCase.pbtxt deleted file mode 100644 index 125651d7eaa1901e4804712bb807322b02ed5bc6..0000000000000000000000000000000000000000 --- a/tensorflow/contrib/proto/python/kernel_tests/optional.TestCase.pbtxt +++ /dev/null @@ -1,20 +0,0 @@ -primitive { - bool_value: true -} -shape: 1 -sizes: 1 -sizes: 0 -field { - name: "bool_value" - dtype: DT_BOOL - expected { - bool_value: true - } -} -field { - name: "double_value" - dtype: DT_DOUBLE - expected { - double_value: 0.0 - } -} diff --git a/tensorflow/contrib/proto/python/kernel_tests/promote_unsigned.TestCase.pbtxt b/tensorflow/contrib/proto/python/kernel_tests/promote_unsigned.TestCase.pbtxt deleted file mode 100644 index bc07efc8f3038c6c540855c97b2254575e517ef3..0000000000000000000000000000000000000000 --- a/tensorflow/contrib/proto/python/kernel_tests/promote_unsigned.TestCase.pbtxt +++ /dev/null @@ -1,29 +0,0 @@ -primitive { - fixed32_value: 4294967295 - uint32_value: 4294967295 -} -shape: 1 -sizes: 1 -field { - name: "fixed32_value" - dtype: DT_INT64 - expected { - int64_value: 4294967295 - } -} -sizes: 1 -field { - name: "uint32_value" - dtype: DT_INT64 - expected { - int64_value: 4294967295 - } -} -sizes: 0 -field { - name: "uint32_default" - dtype: DT_INT64 - expected { - int64_value: 4294967295 # Comes from an explicitly-specified default - } -} diff --git a/tensorflow/contrib/proto/python/kernel_tests/proto_op_test_base.py b/tensorflow/contrib/proto/python/kernel_tests/proto_op_test_base.py new file mode 100644 index 0000000000000000000000000000000000000000..cbc7b3d3f81bae42d525b7049c67f1337933b2de --- /dev/null +++ b/tensorflow/contrib/proto/python/kernel_tests/proto_op_test_base.py @@ -0,0 +1,407 @@ +# ============================================================================= +# Copyright 2018 The TensorFlow Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================= +"""Test case base for testing proto operations.""" + +# Python3 preparedness imports. +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + +import ctypes as ct +import os + +from tensorflow.contrib.proto.python.kernel_tests import test_example_pb2 +from tensorflow.core.framework import types_pb2 +from tensorflow.python.platform import test + + +class ProtoOpTestBase(test.TestCase): + """Base class for testing proto decoding and encoding ops.""" + + def __init__(self, methodName="runTest"): # pylint: disable=invalid-name + super(ProtoOpTestBase, self).__init__(methodName) + lib = os.path.join(os.path.dirname(__file__), "libtestexample.so") + if os.path.isfile(lib): + ct.cdll.LoadLibrary(lib) + + @staticmethod + def named_parameters(): + return ( + ("defaults", ProtoOpTestBase.defaults_test_case()), + ("minmax", ProtoOpTestBase.minmax_test_case()), + ("nested", ProtoOpTestBase.nested_test_case()), + ("optional", ProtoOpTestBase.optional_test_case()), + ("promote_unsigned", ProtoOpTestBase.promote_unsigned_test_case()), + ("ragged", ProtoOpTestBase.ragged_test_case()), + ("shaped_batch", ProtoOpTestBase.shaped_batch_test_case()), + ("simple", ProtoOpTestBase.simple_test_case()), + ) + + @staticmethod + def defaults_test_case(): + test_case = test_example_pb2.TestCase() + test_case.values.add() # No fields specified, so we get all defaults. + test_case.shapes.append(1) + test_case.sizes.append(0) + field = test_case.fields.add() + field.name = "double_value_with_default" + field.dtype = types_pb2.DT_DOUBLE + field.value.double_value.append(1.0) + test_case.sizes.append(0) + field = test_case.fields.add() + field.name = "float_value_with_default" + field.dtype = types_pb2.DT_FLOAT + field.value.float_value.append(2.0) + test_case.sizes.append(0) + field = test_case.fields.add() + field.name = "int64_value_with_default" + field.dtype = types_pb2.DT_INT64 + field.value.int64_value.append(3) + test_case.sizes.append(0) + field = test_case.fields.add() + field.name = "sfixed64_value_with_default" + field.dtype = types_pb2.DT_INT64 + field.value.int64_value.append(11) + test_case.sizes.append(0) + field = test_case.fields.add() + field.name = "sint64_value_with_default" + field.dtype = types_pb2.DT_INT64 + field.value.int64_value.append(13) + test_case.sizes.append(0) + field = test_case.fields.add() + field.name = "uint64_value_with_default" + field.dtype = types_pb2.DT_INT64 + field.value.int64_value.append(4) + test_case.sizes.append(0) + field = test_case.fields.add() + field.name = "fixed64_value_with_default" + field.dtype = types_pb2.DT_INT64 + field.value.int64_value.append(6) + test_case.sizes.append(0) + field = test_case.fields.add() + field.name = "int32_value_with_default" + field.dtype = types_pb2.DT_INT32 + field.value.int32_value.append(5) + test_case.sizes.append(0) + field = test_case.fields.add() + field.name = "sfixed32_value_with_default" + field.dtype = types_pb2.DT_INT32 + field.value.int32_value.append(10) + test_case.sizes.append(0) + field = test_case.fields.add() + field.name = "sint32_value_with_default" + field.dtype = types_pb2.DT_INT32 + field.value.int32_value.append(12) + test_case.sizes.append(0) + field = test_case.fields.add() + field.name = "uint32_value_with_default" + field.dtype = types_pb2.DT_INT32 + field.value.int32_value.append(9) + test_case.sizes.append(0) + field = test_case.fields.add() + field.name = "fixed32_value_with_default" + field.dtype = types_pb2.DT_INT32 + field.value.int32_value.append(7) + test_case.sizes.append(0) + field = test_case.fields.add() + field.name = "bool_value_with_default" + field.dtype = types_pb2.DT_BOOL + field.value.bool_value.append(True) + test_case.sizes.append(0) + field = test_case.fields.add() + field.name = "string_value_with_default" + field.dtype = types_pb2.DT_STRING + field.value.string_value.append("a") + test_case.sizes.append(0) + field = test_case.fields.add() + field.name = "bytes_value_with_default" + field.dtype = types_pb2.DT_STRING + field.value.string_value.append("a longer default string") + return test_case + + @staticmethod + def minmax_test_case(): + test_case = test_example_pb2.TestCase() + value = test_case.values.add() + value.double_value.append(-1.7976931348623158e+308) + value.double_value.append(2.2250738585072014e-308) + value.double_value.append(1.7976931348623158e+308) + value.float_value.append(-3.402823466e+38) + value.float_value.append(1.175494351e-38) + value.float_value.append(3.402823466e+38) + value.int64_value.append(-9223372036854775808) + value.int64_value.append(9223372036854775807) + value.sfixed64_value.append(-9223372036854775808) + value.sfixed64_value.append(9223372036854775807) + value.sint64_value.append(-9223372036854775808) + value.sint64_value.append(9223372036854775807) + value.uint64_value.append(0) + value.uint64_value.append(18446744073709551615) + value.fixed64_value.append(0) + value.fixed64_value.append(18446744073709551615) + value.int32_value.append(-2147483648) + value.int32_value.append(2147483647) + value.sfixed32_value.append(-2147483648) + value.sfixed32_value.append(2147483647) + value.sint32_value.append(-2147483648) + value.sint32_value.append(2147483647) + value.uint32_value.append(0) + value.uint32_value.append(4294967295) + value.fixed32_value.append(0) + value.fixed32_value.append(4294967295) + value.bool_value.append(False) + value.bool_value.append(True) + value.string_value.append("") + value.string_value.append("I refer to the infinite.") + test_case.shapes.append(1) + test_case.sizes.append(3) + field = test_case.fields.add() + field.name = "double_value" + field.dtype = types_pb2.DT_DOUBLE + field.value.double_value.append(-1.7976931348623158e+308) + field.value.double_value.append(2.2250738585072014e-308) + field.value.double_value.append(1.7976931348623158e+308) + test_case.sizes.append(3) + field = test_case.fields.add() + field.name = "float_value" + field.dtype = types_pb2.DT_FLOAT + field.value.float_value.append(-3.402823466e+38) + field.value.float_value.append(1.175494351e-38) + field.value.float_value.append(3.402823466e+38) + test_case.sizes.append(2) + field = test_case.fields.add() + field.name = "int64_value" + field.dtype = types_pb2.DT_INT64 + field.value.int64_value.append(-9223372036854775808) + field.value.int64_value.append(9223372036854775807) + test_case.sizes.append(2) + field = test_case.fields.add() + field.name = "sfixed64_value" + field.dtype = types_pb2.DT_INT64 + field.value.int64_value.append(-9223372036854775808) + field.value.int64_value.append(9223372036854775807) + test_case.sizes.append(2) + field = test_case.fields.add() + field.name = "sint64_value" + field.dtype = types_pb2.DT_INT64 + field.value.int64_value.append(-9223372036854775808) + field.value.int64_value.append(9223372036854775807) + test_case.sizes.append(2) + field = test_case.fields.add() + field.name = "uint64_value" + field.dtype = types_pb2.DT_INT64 + field.value.int64_value.append(0) + field.value.int64_value.append(-1) + test_case.sizes.append(2) + field = test_case.fields.add() + field.name = "fixed64_value" + field.dtype = types_pb2.DT_INT64 + field.value.int64_value.append(0) + field.value.int64_value.append(-1) + test_case.sizes.append(2) + field = test_case.fields.add() + field.name = "int32_value" + field.dtype = types_pb2.DT_INT32 + field.value.int32_value.append(-2147483648) + field.value.int32_value.append(2147483647) + test_case.sizes.append(2) + field = test_case.fields.add() + field.name = "sfixed32_value" + field.dtype = types_pb2.DT_INT32 + field.value.int32_value.append(-2147483648) + field.value.int32_value.append(2147483647) + test_case.sizes.append(2) + field = test_case.fields.add() + field.name = "sint32_value" + field.dtype = types_pb2.DT_INT32 + field.value.int32_value.append(-2147483648) + field.value.int32_value.append(2147483647) + test_case.sizes.append(2) + field = test_case.fields.add() + field.name = "uint32_value" + field.dtype = types_pb2.DT_INT32 + field.value.int32_value.append(0) + field.value.int32_value.append(-1) + test_case.sizes.append(2) + field = test_case.fields.add() + field.name = "fixed32_value" + field.dtype = types_pb2.DT_INT32 + field.value.int32_value.append(0) + field.value.int32_value.append(-1) + test_case.sizes.append(2) + field = test_case.fields.add() + field.name = "bool_value" + field.dtype = types_pb2.DT_BOOL + field.value.bool_value.append(False) + field.value.bool_value.append(True) + test_case.sizes.append(2) + field = test_case.fields.add() + field.name = "string_value" + field.dtype = types_pb2.DT_STRING + field.value.string_value.append("") + field.value.string_value.append("I refer to the infinite.") + return test_case + + @staticmethod + def nested_test_case(): + test_case = test_example_pb2.TestCase() + value = test_case.values.add() + message_value = value.message_value.add() + message_value.double_value = 23.5 + test_case.shapes.append(1) + test_case.sizes.append(1) + field = test_case.fields.add() + field.name = "message_value" + field.dtype = types_pb2.DT_STRING + message_value = field.value.message_value.add() + message_value.double_value = 23.5 + return test_case + + @staticmethod + def optional_test_case(): + test_case = test_example_pb2.TestCase() + value = test_case.values.add() + value.bool_value.append(True) + test_case.shapes.append(1) + test_case.sizes.append(1) + field = test_case.fields.add() + field.name = "bool_value" + field.dtype = types_pb2.DT_BOOL + field.value.bool_value.append(True) + test_case.sizes.append(0) + field = test_case.fields.add() + field.name = "double_value" + field.dtype = types_pb2.DT_DOUBLE + field.value.double_value.append(0.0) + return test_case + + @staticmethod + def promote_unsigned_test_case(): + test_case = test_example_pb2.TestCase() + value = test_case.values.add() + value.fixed32_value.append(4294967295) + value.uint32_value.append(4294967295) + test_case.shapes.append(1) + test_case.sizes.append(1) + field = test_case.fields.add() + field.name = "fixed32_value" + field.dtype = types_pb2.DT_INT64 + field.value.int64_value.append(4294967295) + test_case.sizes.append(1) + field = test_case.fields.add() + field.name = "uint32_value" + field.dtype = types_pb2.DT_INT64 + field.value.int64_value.append(4294967295) + # Comes from an explicitly-specified default + test_case.sizes.append(0) + field = test_case.fields.add() + field.name = "uint32_value_with_default" + field.dtype = types_pb2.DT_INT64 + field.value.int64_value.append(9) + return test_case + + @staticmethod + def ragged_test_case(): + test_case = test_example_pb2.TestCase() + value = test_case.values.add() + value.double_value.append(23.5) + value.double_value.append(123.0) + value.bool_value.append(True) + value = test_case.values.add() + value.double_value.append(3.1) + value.bool_value.append(False) + test_case.shapes.append(2) + test_case.sizes.append(2) + test_case.sizes.append(1) + test_case.sizes.append(1) + test_case.sizes.append(1) + field = test_case.fields.add() + field.name = "double_value" + field.dtype = types_pb2.DT_DOUBLE + field.value.double_value.append(23.5) + field.value.double_value.append(123.0) + field.value.double_value.append(3.1) + field.value.double_value.append(0.0) + field = test_case.fields.add() + field.name = "bool_value" + field.dtype = types_pb2.DT_BOOL + field.value.bool_value.append(True) + field.value.bool_value.append(False) + return test_case + + @staticmethod + def shaped_batch_test_case(): + test_case = test_example_pb2.TestCase() + value = test_case.values.add() + value.double_value.append(23.5) + value.bool_value.append(True) + value = test_case.values.add() + value.double_value.append(44.0) + value.bool_value.append(False) + value = test_case.values.add() + value.double_value.append(3.14159) + value.bool_value.append(True) + value = test_case.values.add() + value.double_value.append(1.414) + value.bool_value.append(True) + value = test_case.values.add() + value.double_value.append(-32.2) + value.bool_value.append(False) + value = test_case.values.add() + value.double_value.append(0.0001) + value.bool_value.append(True) + test_case.shapes.append(3) + test_case.shapes.append(2) + for _ in range(12): + test_case.sizes.append(1) + field = test_case.fields.add() + field.name = "double_value" + field.dtype = types_pb2.DT_DOUBLE + field.value.double_value.append(23.5) + field.value.double_value.append(44.0) + field.value.double_value.append(3.14159) + field.value.double_value.append(1.414) + field.value.double_value.append(-32.2) + field.value.double_value.append(0.0001) + field = test_case.fields.add() + field.name = "bool_value" + field.dtype = types_pb2.DT_BOOL + field.value.bool_value.append(True) + field.value.bool_value.append(False) + field.value.bool_value.append(True) + field.value.bool_value.append(True) + field.value.bool_value.append(False) + field.value.bool_value.append(True) + return test_case + + @staticmethod + def simple_test_case(): + test_case = test_example_pb2.TestCase() + value = test_case.values.add() + value.double_value.append(23.5) + value.bool_value.append(True) + test_case.shapes.append(1) + test_case.sizes.append(1) + field = test_case.fields.add() + field.name = "double_value" + field.dtype = types_pb2.DT_DOUBLE + field.value.double_value.append(23.5) + test_case.sizes.append(1) + field = test_case.fields.add() + field.name = "bool_value" + field.dtype = types_pb2.DT_BOOL + field.value.bool_value.append(True) + return test_case diff --git a/tensorflow/contrib/proto/python/kernel_tests/ragged.TestCase.pbtxt b/tensorflow/contrib/proto/python/kernel_tests/ragged.TestCase.pbtxt deleted file mode 100644 index 61c7ac53f72b0764a0d57241cbdcdd93fcbd9279..0000000000000000000000000000000000000000 --- a/tensorflow/contrib/proto/python/kernel_tests/ragged.TestCase.pbtxt +++ /dev/null @@ -1,32 +0,0 @@ -primitive { - double_value: 23.5 - double_value: 123.0 - bool_value: true -} -primitive { - double_value: 3.1 - bool_value: false -} -shape: 2 -sizes: 2 -sizes: 1 -sizes: 1 -sizes: 1 -field { - name: "double_value" - dtype: DT_DOUBLE - expected { - double_value: 23.5 - double_value: 123.0 - double_value: 3.1 - double_value: 0.0 - } -} -field { - name: "bool_value" - dtype: DT_BOOL - expected { - bool_value: true - bool_value: false - } -} diff --git a/tensorflow/contrib/proto/python/kernel_tests/shaped_batch.TestCase.pbtxt b/tensorflow/contrib/proto/python/kernel_tests/shaped_batch.TestCase.pbtxt deleted file mode 100644 index f4828076d52dc5d03a887c4a445dbcf52414c361..0000000000000000000000000000000000000000 --- a/tensorflow/contrib/proto/python/kernel_tests/shaped_batch.TestCase.pbtxt +++ /dev/null @@ -1,62 +0,0 @@ -primitive { - double_value: 23.5 - bool_value: true -} -primitive { - double_value: 44.0 - bool_value: false -} -primitive { - double_value: 3.14159 - bool_value: true -} -primitive { - double_value: 1.414 - bool_value: true -} -primitive { - double_value: -32.2 - bool_value: false -} -primitive { - double_value: 0.0001 - bool_value: true -} -shape: 3 -shape: 2 -sizes: 1 -sizes: 1 -sizes: 1 -sizes: 1 -sizes: 1 -sizes: 1 -sizes: 1 -sizes: 1 -sizes: 1 -sizes: 1 -sizes: 1 -sizes: 1 -field { - name: "double_value" - dtype: DT_DOUBLE - expected { - double_value: 23.5 - double_value: 44.0 - double_value: 3.14159 - double_value: 1.414 - double_value: -32.2 - double_value: 0.0001 - } -} -field { - name: "bool_value" - dtype: DT_BOOL - expected { - bool_value: true - bool_value: false - bool_value: true - bool_value: true - bool_value: false - bool_value: true - } -} diff --git a/tensorflow/contrib/proto/python/kernel_tests/simple.TestCase.pbtxt b/tensorflow/contrib/proto/python/kernel_tests/simple.TestCase.pbtxt deleted file mode 100644 index dc20ac147b0e772f05b4fc614f9f56513aceb1d5..0000000000000000000000000000000000000000 --- a/tensorflow/contrib/proto/python/kernel_tests/simple.TestCase.pbtxt +++ /dev/null @@ -1,21 +0,0 @@ -primitive { - double_value: 23.5 - bool_value: true -} -shape: 1 -sizes: 1 -sizes: 1 -field { - name: "double_value" - dtype: DT_DOUBLE - expected { - double_value: 23.5 - } -} -field { - name: "bool_value" - dtype: DT_BOOL - expected { - bool_value: true - } -} diff --git a/tensorflow/contrib/proto/python/kernel_tests/test_example.proto b/tensorflow/contrib/proto/python/kernel_tests/test_example.proto index a2c88e372bf7c6b7f14c5bb55776b66c4c06bcd4..674d881220a1113631def47c5111e3ef401b99f3 100644 --- a/tensorflow/contrib/proto/python/kernel_tests/test_example.proto +++ b/tensorflow/contrib/proto/python/kernel_tests/test_example.proto @@ -1,6 +1,4 @@ // Test description and protos to work with it. -// -// Many of the protos in this file are for unit tests that haven't been written yet. syntax = "proto2"; @@ -8,54 +6,27 @@ import "tensorflow/core/framework/types.proto"; package tensorflow.contrib.proto; -// A TestCase holds a proto and a bunch of assertions -// about how it should decode. +// A TestCase holds a proto and assertions about how it should decode. message TestCase { - // A batch of primitives to be serialized and decoded. - repeated RepeatedPrimitiveValue primitive = 1; - // The shape of the batch. - repeated int32 shape = 2; + // Batches of primitive values. + repeated TestValue values = 1; + // The batch shapes. + repeated int32 shapes = 2; // Expected sizes for each field. repeated int32 sizes = 3; // Expected values for each field. - repeated FieldSpec field = 4; + repeated FieldSpec fields = 4; }; // FieldSpec describes the expected output for a single field. message FieldSpec { optional string name = 1; optional tensorflow.DataType dtype = 2; - optional RepeatedPrimitiveValue expected = 3; + optional TestValue value = 3; }; +// NOTE: This definition must be kept in sync with PackedTestValue. message TestValue { - optional PrimitiveValue primitive_value = 1; - optional EnumValue enum_value = 2; - optional MessageValue message_value = 3; - optional RepeatedMessageValue repeated_message_value = 4; - optional RepeatedPrimitiveValue repeated_primitive_value = 6; -} - -message PrimitiveValue { - optional double double_value = 1; - optional float float_value = 2; - optional int64 int64_value = 3; - optional uint64 uint64_value = 4; - optional int32 int32_value = 5; - optional fixed64 fixed64_value = 6; - optional fixed32 fixed32_value = 7; - optional bool bool_value = 8; - optional string string_value = 9; - optional bytes bytes_value = 12; - optional uint32 uint32_value = 13; - optional sfixed32 sfixed32_value = 15; - optional sfixed64 sfixed64_value = 16; - optional sint32 sint32_value = 17; - optional sint64 sint64_value = 18; -} - -// NOTE: This definition must be kept in sync with PackedPrimitiveValue. -message RepeatedPrimitiveValue { repeated double double_value = 1; repeated float float_value = 2; repeated int64 int64_value = 3; @@ -74,30 +45,31 @@ message RepeatedPrimitiveValue { repeated PrimitiveValue message_value = 19; // Optional fields with explicitly-specified defaults. - optional double double_default = 20 [default = 1.0]; - optional float float_default = 21 [default = 2.0]; - optional int64 int64_default = 22 [default = 3]; - optional uint64 uint64_default = 23 [default = 4]; - optional int32 int32_default = 24 [default = 5]; - optional fixed64 fixed64_default = 25 [default = 6]; - optional fixed32 fixed32_default = 26 [default = 7]; - optional bool bool_default = 27 [default = true]; - optional string string_default = 28 [default = "a"]; - optional bytes bytes_default = 29 [default = "a longer default string"]; - optional uint32 uint32_default = 30 [default = 4294967295]; - optional sfixed32 sfixed32_default = 31 [default = 10]; - optional sfixed64 sfixed64_default = 32 [default = 11]; - optional sint32 sint32_default = 33 [default = 12]; - optional sint64 sint64_default = 34 [default = 13]; + optional double double_value_with_default = 20 [default = 1.0]; + optional float float_value_with_default = 21 [default = 2.0]; + optional int64 int64_value_with_default = 22 [default = 3]; + optional uint64 uint64_value_with_default = 23 [default = 4]; + optional int32 int32_value_with_default = 24 [default = 5]; + optional fixed64 fixed64_value_with_default = 25 [default = 6]; + optional fixed32 fixed32_value_with_default = 26 [default = 7]; + optional bool bool_value_with_default = 27 [default = true]; + optional string string_value_with_default = 28 [default = "a"]; + optional bytes bytes_value_with_default = 29 + [default = "a longer default string"]; + optional uint32 uint32_value_with_default = 30 [default = 9]; + optional sfixed32 sfixed32_value_with_default = 31 [default = 10]; + optional sfixed64 sfixed64_value_with_default = 32 [default = 11]; + optional sint32 sint32_value_with_default = 33 [default = 12]; + optional sint64 sint64_value_with_default = 34 [default = 13]; } -// A PackedPrimitiveValue looks exactly the same as a RepeatedPrimitiveValue -// in the text format, but the binary serializion is different. -// We test the packed representations by loading the same test cases -// using this definition instead of RepeatedPrimitiveValue. -// NOTE: This definition must be kept in sync with RepeatedPrimitiveValue -// in every way except the packed=true declaration. -message PackedPrimitiveValue { +// A PackedTestValue looks exactly the same as a TestValue in the text format, +// but the binary serializion is different. We test the packed representations +// by loading the same test cases using this definition instead of TestValue. +// +// NOTE: This definition must be kept in sync with TestValue in every way except +// the packed=true declaration. +message PackedTestValue { repeated double double_value = 1 [packed = true]; repeated float float_value = 2 [packed = true]; repeated int64 int64_value = 3 [packed = true]; @@ -115,23 +87,53 @@ message PackedPrimitiveValue { repeated sint64 sint64_value = 18 [packed = true]; repeated PrimitiveValue message_value = 19; - optional double double_default = 20 [default = 1.0]; - optional float float_default = 21 [default = 2.0]; - optional int64 int64_default = 22 [default = 3]; - optional uint64 uint64_default = 23 [default = 4]; - optional int32 int32_default = 24 [default = 5]; - optional fixed64 fixed64_default = 25 [default = 6]; - optional fixed32 fixed32_default = 26 [default = 7]; - optional bool bool_default = 27 [default = true]; - optional string string_default = 28 [default = "a"]; - optional bytes bytes_default = 29 [default = "a longer default string"]; - optional uint32 uint32_default = 30 [default = 4294967295]; - optional sfixed32 sfixed32_default = 31 [default = 10]; - optional sfixed64 sfixed64_default = 32 [default = 11]; - optional sint32 sint32_default = 33 [default = 12]; - optional sint64 sint64_default = 34 [default = 13]; + optional double double_value_with_default = 20 [default = 1.0]; + optional float float_value_with_default = 21 [default = 2.0]; + optional int64 int64_value_with_default = 22 [default = 3]; + optional uint64 uint64_value_with_default = 23 [default = 4]; + optional int32 int32_value_with_default = 24 [default = 5]; + optional fixed64 fixed64_value_with_default = 25 [default = 6]; + optional fixed32 fixed32_value_with_default = 26 [default = 7]; + optional bool bool_value_with_default = 27 [default = true]; + optional string string_value_with_default = 28 [default = "a"]; + optional bytes bytes_value_with_default = 29 + [default = "a longer default string"]; + optional uint32 uint32_value_with_default = 30 [default = 9]; + optional sfixed32 sfixed32_value_with_default = 31 [default = 10]; + optional sfixed64 sfixed64_value_with_default = 32 [default = 11]; + optional sint32 sint32_value_with_default = 33 [default = 12]; + optional sint64 sint64_value_with_default = 34 [default = 13]; } +message PrimitiveValue { + optional double double_value = 1; + optional float float_value = 2; + optional int64 int64_value = 3; + optional uint64 uint64_value = 4; + optional int32 int32_value = 5; + optional fixed64 fixed64_value = 6; + optional fixed32 fixed32_value = 7; + optional bool bool_value = 8; + optional string string_value = 9; + optional bytes bytes_value = 12; + optional uint32 uint32_value = 13; + optional sfixed32 sfixed32_value = 15; + optional sfixed64 sfixed64_value = 16; + optional sint32 sint32_value = 17; + optional sint64 sint64_value = 18; +} + +// Message containing fields with field numbers higher than any field above. +// An instance of this message is prepended to each binary message in the test +// to exercise the code path that handles fields encoded out of order of field +// number. +message ExtraFields { + optional string string_value = 1776; + optional bool bool_value = 1777; +} + +// The messages below are for yet-to-be created tests. + message EnumValue { enum Color { RED = 0; @@ -171,12 +173,3 @@ message RepeatedMessageValue { repeated NestedMessageValue message_values = 11; } - -// Message containing fields with field numbers higher than any field above. An -// instance of this message is prepended to each binary message in the test to -// exercise the code path that handles fields encoded out of order of field -// number. -message ExtraFields { - optional string string_value = 1776; - optional bool bool_value = 1777; -} diff --git a/tensorflow/contrib/quantize/python/fold_batch_norms.py b/tensorflow/contrib/quantize/python/fold_batch_norms.py index 55479bf5f74299bf09f131a6127f9f11d6192d90..e3c48998305e9d9b6c185fd4c0f324fa0449c691 100644 --- a/tensorflow/contrib/quantize/python/fold_batch_norms.py +++ b/tensorflow/contrib/quantize/python/fold_batch_norms.py @@ -121,7 +121,8 @@ def _FoldFusedBatchNorms(graph, is_training, freeze_batch_norm_delay): scaled_weight_tensor = math_ops.multiply( weights, multiplier_tensor, name='mul_fold') new_layer_tensor = _CloneWithNewOperands( - match.layer_op, match.input_tensor, scaled_weight_tensor) + match.layer_op, match.input_tensor, scaled_weight_tensor, + match.batch_to_space_op) if correction_recip is not None: new_layer_tensor = math_ops.multiply( @@ -149,6 +150,8 @@ def _FindFusedBatchNorms(graph): _FusedBatchNormMatches. """ input_pattern = graph_matcher.OpTypePattern('*') + # In practice, the weight pattern can match a Variable or a SpaceToBatchND + # operation that follows a variable for atrous convolutions. weight_pattern = graph_matcher.OpTypePattern('*') gamma_pattern = graph_matcher.OpTypePattern('*') beta_pattern = graph_matcher.OpTypePattern('*') @@ -160,16 +163,27 @@ def _FindFusedBatchNorms(graph): layer_pattern = graph_matcher.OpTypePattern( 'Conv2D|DepthwiseConv2dNative|MatMul', inputs=[input_pattern, weight_pattern]) + batch_to_space_pattern = graph_matcher.OpTypePattern( + 'BatchToSpaceND', + inputs=[ + layer_pattern, + graph_matcher.OpTypePattern('*'), + graph_matcher.OpTypePattern('*') + ]) + layer_output_pattern = graph_matcher.OneofPattern( + [layer_pattern, batch_to_space_pattern]) # MatMul has a Reshape between it and FusedBatchNorm. matmul_reshape_pattern = graph_matcher.OpTypePattern( - 'Reshape', inputs=[layer_pattern, - graph_matcher.OpTypePattern('*')]) + 'Reshape', + inputs=[layer_output_pattern, + graph_matcher.OpTypePattern('*')]) batch_norm_pattern = graph_matcher.OpTypePattern( 'FusedBatchNorm', inputs=[ - graph_matcher.OneofPattern([matmul_reshape_pattern, layer_pattern]), - gamma_pattern, beta_pattern, mean_pattern, variance_pattern + graph_matcher.OneofPattern( + [matmul_reshape_pattern, layer_output_pattern]), gamma_pattern, + beta_pattern, mean_pattern, variance_pattern ]) matmul_bn_output_reshape_pattern = graph_matcher.OpTypePattern( 'Reshape', inputs=[batch_norm_pattern, @@ -192,6 +206,7 @@ def _FindFusedBatchNorms(graph): moving_variance_tensor = None bn_decay_mean_tensor = None bn_decay_var_tensor = None + batch_to_space_op = None layer_op = match_result.get_op(layer_pattern) layer_tensor = match_result.get_tensor(layer_pattern) bn_op = match_result.get_op(batch_norm_pattern) @@ -213,6 +228,7 @@ def _FindFusedBatchNorms(graph): if not output_tensor.consumers(): continue + batch_to_space_op = match_result.get_op(batch_to_space_pattern) input_tensor = match_result.get_tensor(input_pattern) weight_tensor = match_result.get_tensor(weight_pattern) gamma_tensor = match_result.get_tensor(gamma_pattern) @@ -276,7 +292,8 @@ def _FindFusedBatchNorms(graph): moving_variance_tensor=moving_variance_tensor, bn_decay_mean_tensor=bn_decay_mean_tensor, bn_decay_var_tensor=bn_decay_var_tensor, - batch_epsilon=batch_epsilon) + batch_epsilon=batch_epsilon, + batch_to_space_op=batch_to_space_op) def _ComputeBatchNormCorrections(context, match, freeze_batch_norm_delay, @@ -380,7 +397,8 @@ def _ComputeBatchNormCorrections(context, match, freeze_batch_norm_delay, return correction_scale, correction_recip, correction_offset -def _CloneWithNewOperands(layer_op, input_tensor, weight_tensor): +def _CloneWithNewOperands(layer_op, input_tensor, weight_tensor, + batch_to_space_op): """Clones layer_op with input_tensor and weight_tensor as new inputs.""" new_layer_name = layer_op.name.split('/')[-1] + '_Fold' if layer_op.type == 'Conv2D': @@ -400,12 +418,25 @@ def _CloneWithNewOperands(layer_op, input_tensor, weight_tensor): transpose_b=layer_op.get_attr('transpose_b'), name=new_layer_name) elif layer_op.type == 'DepthwiseConv2dNative': - return nn.depthwise_conv2d( + conv = nn.depthwise_conv2d( input_tensor, weight_tensor, + rate=layer_op.get_attr('dilations'), strides=layer_op.get_attr('strides'), padding=layer_op.get_attr('padding'), name=new_layer_name) + # Copy the batch to space operation if we have a atrous convolution. + if batch_to_space_op: + batch_to_space_op = layer_op.outputs[0].consumers()[0] + # TODO(suharshs): It's hard to make this name match with the unfused name. + # Restructure this code to not rely on scope at all. + new_batch_to_space_name = batch_to_space_op.name.split('/')[-1] + '_Fold' + conv = array_ops.batch_to_space_nd( + conv, + batch_to_space_op.inputs[1], + batch_to_space_op.inputs[2], + name=new_batch_to_space_name) + return conv else: raise ValueError('Cannot handle operation of type: %s' % layer_op.type) @@ -617,7 +648,8 @@ def _GetBatchNormParams(graph, context, has_scaling): moving_variance_tensor=moving_variance_tensor, bn_decay_mean_tensor=bn_decay_mean_tensor, bn_decay_var_tensor=bn_decay_var_tensor, - batch_epsilon=batch_epsilon) + batch_epsilon=batch_epsilon, + batch_to_space_op=None) def _CreateFoldedOp(graph, context, has_scaling, freeze_batch_norm_delay, @@ -651,6 +683,11 @@ def _CreateFoldedOp(graph, context, has_scaling, freeze_batch_norm_delay, '/BatchNorm/batchnorm_1/' + mul_scale_name) op_below = mul_scale.inputs[0].op + # Skip over the BatchToSpace operation in the case of atrous convolutions. + batch_to_space_op = None + if op_below.type == 'BatchToSpaceND': + batch_to_space_op = op_below + op_below = op_below.inputs[0].op weights = op_below.inputs[1] match = _GetBatchNormParams( graph=graph, context=context, has_scaling=has_scaling) @@ -691,7 +728,7 @@ def _CreateFoldedOp(graph, context, has_scaling, freeze_batch_norm_delay, context + '/correction_mult') mul_fold = _CloneOp(mul_scale, context + '/mul_fold', [(0, weights)]) else: - raise ValueError('Cannot handle operation of type: %s' % op_below.op) + raise ValueError('Cannot handle operation of type: %s' % op_below.type) _AssertShapesMatch('mul_fold', mul_fold.inputs[0], mul_fold.outputs[0]) conv_or_fc_folded = _CloneOp(op_below, op_below.name + '_Fold', @@ -701,6 +738,13 @@ def _CreateFoldedOp(graph, context, has_scaling, freeze_batch_norm_delay, context + '/BatchNorm/batchnorm_1/add_1') corrected_output = conv_or_fc_folded.outputs[0] + # Copy the batch to space operation if we have a atrous convolution. + if batch_to_space_op: + corrected_output = array_ops.batch_to_space_nd( + corrected_output, + batch_to_space_op.inputs[1], + batch_to_space_op.inputs[2], + name=batch_to_space_op.name + '_Fold') if correction_offset is not None: with ops.device(conv_or_fc_folded.device): corrected_output = math_ops.multiply(correction_recip, corrected_output, @@ -898,7 +942,8 @@ class _BatchNormMatch(object): def __init__(self, layer_op, bn_op, output_tensor, input_tensor, weight_tensor, gamma_tensor, beta_tensor, mean_tensor, variance_tensor, moving_mean_tensor, moving_variance_tensor, - bn_decay_mean_tensor, bn_decay_var_tensor, batch_epsilon): + bn_decay_mean_tensor, bn_decay_var_tensor, batch_epsilon, + batch_to_space_op): self._layer_op = layer_op self._bn_op = bn_op self._output_tensor = output_tensor @@ -913,6 +958,7 @@ class _BatchNormMatch(object): self._bn_decay_mean_tensor = bn_decay_mean_tensor self._bn_decay_var_tensor = bn_decay_var_tensor self._batch_epsilon = batch_epsilon + self._batch_to_space_op = batch_to_space_op @property def layer_op(self): @@ -969,3 +1015,7 @@ class _BatchNormMatch(object): @property def bn_decay_var_tensor(self): return self._bn_decay_var_tensor + + @property + def batch_to_space_op(self): + return self._batch_to_space_op diff --git a/tensorflow/contrib/quantize/python/fold_batch_norms_test.py b/tensorflow/contrib/quantize/python/fold_batch_norms_test.py index bfa9d3bf705e327091098a8e416b7902f852605a..7c907ffd92c1ae0c762e41cc429b0e6ce053f6b9 100644 --- a/tensorflow/contrib/quantize/python/fold_batch_norms_test.py +++ b/tensorflow/contrib/quantize/python/fold_batch_norms_test.py @@ -438,6 +438,90 @@ class FoldBatchNormsTest(test_util.TensorFlowTestCase): def testFoldDepthwiseConv2d(self): self._RunTestOverParameters(self._TestFoldDepthwiseConv2d) + def _TestFoldAtrousConv2d(self, relu, relu_op_name, with_bypass, has_scaling, + fused_batch_norm, freeze_batch_norm_delay): + """Tests folding: inputs -> AtrousConv2d with batch norm -> Relu*. + + Args: + relu: Callable that returns an Operation, a factory method for the Relu*. + relu_op_name: String, name of the Relu* operation. + with_bypass: Bool, when true there is an extra connection added from + inputs to just before Relu*. + has_scaling: Bool, when true the batch norm has scaling. + fused_batch_norm: Bool, when true the batch norm is fused. + freeze_batch_norm_delay: None or the number of steps after which training + switches to using frozen mean and variance + """ + g = ops.Graph() + with g.as_default(): + batch_size, height, width = 5, 128, 128 + inputs = array_ops.zeros((batch_size, height, width, 3)) + dilation_rate = 2 + activation_fn = None if with_bypass else relu + scope = 'test/test2' if with_bypass else 'test' + node = separable_conv2d( + inputs, + None, [3, 3], + rate=dilation_rate, + depth_multiplier=1.0, + padding='SAME', + weights_initializer=self._WeightInit(0.09), + activation_fn=activation_fn, + normalizer_fn=batch_norm, + normalizer_params=self._BatchNormParams( + scale=has_scaling, fused=fused_batch_norm), + scope=scope) + if with_bypass: + node = math_ops.add(inputs, node, name='test/Add') + relu(node, name='test/' + relu_op_name) + + fold_batch_norms.FoldBatchNorms( + g, is_training=True, freeze_batch_norm_delay=freeze_batch_norm_delay) + + folded_mul = g.get_operation_by_name(scope + '/mul_fold') + self.assertEqual(folded_mul.type, 'Mul') + if fused_batch_norm: + scale_reshape_op_name = scope + '/BatchNorm_Fold/scale_reshape' + else: + scale_reshape_op_name = scope + '/scale_reshape' + self._AssertInputOpsAre(folded_mul, + [scope + '/correction_mult', scale_reshape_op_name]) + self._AssertOutputGoesToOps(folded_mul, g, [scope + '/depthwise_Fold']) + + scale_reshape = g.get_operation_by_name(scale_reshape_op_name) + self.assertEqual(scale_reshape.type, 'Reshape') + self._AssertInputOpsAre(scale_reshape, [ + self._BatchNormMultiplierName(scope, has_scaling, fused_batch_norm), + scale_reshape_op_name + '/shape' + ]) + self._AssertOutputGoesToOps(scale_reshape, g, [scope + '/mul_fold']) + + folded_conv = g.get_operation_by_name(scope + '/depthwise_Fold') + self.assertEqual(folded_conv.type, 'DepthwiseConv2dNative') + self._AssertInputOpsAre( + folded_conv, [scope + '/mul_fold', scope + '/depthwise/SpaceToBatchND']) + if fused_batch_norm: + self._AssertOutputGoesToOps(folded_conv, g, + [scope + '/BatchToSpaceND_Fold']) + else: + self._AssertOutputGoesToOps(folded_conv, g, + [scope + '/depthwise/BatchToSpaceND_Fold']) + + folded_add = g.get_operation_by_name(scope + '/add_fold') + self.assertEqual(folded_add.type, 'Add') + self._AssertInputOpsAre(folded_add, [ + scope + '/correction_add', + self._BathNormBiasName(scope, fused_batch_norm) + ]) + output_op_names = ['test/Add' if with_bypass else 'test/' + relu_op_name] + self._AssertOutputGoesToOps(folded_add, g, output_op_names) + + for op in g.get_operations(): + self.assertFalse('//' in op.name, 'Double slash in op %s' % op.name) + + def testFoldAtrousConv2d(self): + self._RunTestOverParameters(self._TestFoldAtrousConv2d) + def _TestCompareFoldAndUnfolded(self, relu, relu_op_name, with_bypass, has_scaling, fused_batch_norm, freeze_batch_norm_delay): diff --git a/tensorflow/contrib/quantize/python/quantize.py b/tensorflow/contrib/quantize/python/quantize.py index cbba72643f7f166c473b6181edc292f695c4cbc2..4fc315d901a86ac235513aad6eb34d7f90f61801 100644 --- a/tensorflow/contrib/quantize/python/quantize.py +++ b/tensorflow/contrib/quantize/python/quantize.py @@ -194,6 +194,8 @@ def _FindLayersToQuantize(graph): / conv|fc | + [batch_to_space_nd] + | [post_conv_correction] | biasadd|folded_bias @@ -247,9 +249,21 @@ def _FindLayersToQuantize(graph): ], ordered_inputs=False) + # For atrous convolutions a BatchToSpaceND will occur after the depthwise + # convolution. + batch_to_space_pattern = graph_matcher.OpTypePattern( + 'BatchToSpaceND', + inputs=[ + layer_pattern, + graph_matcher.OpTypePattern('*'), + graph_matcher.OpTypePattern('*') + ]) + + layer_output_pattern = graph_matcher.OneofPattern( + [batch_to_space_pattern, layer_pattern]) folded_bias_mul_pattern = graph_matcher.OpTypePattern( 'Mul', - inputs=[graph_matcher.OpTypePattern('*'), layer_pattern], + inputs=[graph_matcher.OpTypePattern('*'), layer_output_pattern], ordered_inputs=False) post_layer_op_correction_pattern = graph_matcher.OpTypePattern( 'Add', @@ -264,28 +278,37 @@ def _FindLayersToQuantize(graph): ], ordered_inputs=False) + # batch_norms with forced updates have an Identity operation at the end. + # TODO(suharshs): Find a way to easily skip extra Identity operations. The + # current issue is that doing so can often match patterns across many layers + # incorrectly. + batch_norm_identity = graph_matcher.OpTypePattern( + 'Identity', inputs=[folded_bias_add_pattern]) + bias_add_pattern = graph_matcher.OpTypePattern( - 'Add|BiasAdd', inputs=[layer_pattern, '*'], ordered_inputs=False) + 'Add|BiasAdd', inputs=[layer_output_pattern, '*'], ordered_inputs=False) # The bias can come from the bias add or the folded bias add. bypass_pattern = graph_matcher.OpTypePattern( 'Add', inputs=[ graph_matcher.OneofPattern( - [bias_add_pattern, folded_bias_add_pattern]), '*' + [bias_add_pattern, folded_bias_add_pattern, batch_norm_identity]), + '*' ], ordered_inputs=False) # The input to the activation can come from bias add, fold bias add, the # bypasses. # TODO(suharshs): We should ideally skip Identity operations instead of - # treating them as an activation. + # treating them as activations. activation_pattern = graph_matcher.OpTypePattern( '|'.join(_ACTIVATION_TYPES) + '|Identity', inputs=[ graph_matcher.OneofPattern([ bias_add_pattern, folded_bias_add_pattern, + batch_norm_identity, bypass_pattern, ]) ]) @@ -373,14 +396,6 @@ def _FindLayersToQuantize(graph): return layer_matches -def _HasPostActivationBypass(activation_op): - for activation_tensor in activation_op.outputs: - for output_op in activation_tensor.consumers(): - if output_op.type == 'Add': - return True - return False - - class _LayerMatch(object): """Contains all information related to a matched Layer.""" diff --git a/tensorflow/contrib/quantize/python/quantize_graph.py b/tensorflow/contrib/quantize/python/quantize_graph.py index 11d052d7f491dc029d1bda9b47364d6e9c880a67..2944f964c7078814111c96890f18abe1607b68fc 100644 --- a/tensorflow/contrib/quantize/python/quantize_graph.py +++ b/tensorflow/contrib/quantize/python/quantize_graph.py @@ -191,6 +191,7 @@ def experimental_create_training_graph(input_graph=None, def experimental_create_eval_graph(input_graph=None, weight_bits=8, activation_bits=8, + quant_delay=None, scope=None): """Rewrites an eval input_graph in place for simulated quantization. @@ -209,6 +210,8 @@ def experimental_create_eval_graph(input_graph=None, default graph. weight_bits: Number of bits to use for quantizing weights. activation_bits: Number of bits to use for quantizing activations. + quant_delay: Number of steps after which weights and activations are + quantized during eval. scope: The scope to be transformed. If it's not None, only the ops which are in this scope will be transformed. @@ -221,4 +224,5 @@ def experimental_create_eval_graph(input_graph=None, is_training=False, weight_bits=weight_bits, activation_bits=activation_bits, + quant_delay=quant_delay, scope=scope) diff --git a/tensorflow/contrib/quantize/python/quantize_parameterized_test.py b/tensorflow/contrib/quantize/python/quantize_parameterized_test.py index db745aa56212af6a9c20e06ee9e4e5d6e27cf3c3..31a2955ddb3b32f2b07c6125c8f83ffba335cc5f 100644 --- a/tensorflow/contrib/quantize/python/quantize_parameterized_test.py +++ b/tensorflow/contrib/quantize/python/quantize_parameterized_test.py @@ -276,6 +276,52 @@ class QuantizeTest(test_util.TensorFlowTestCase): graph, scope, 'DepthwiseConv2dNative', activation_op_name, with_bypass, delay, use_resource) + def testQuantize_AtrousConvWithoutBatchNorm(self): + self._RunWithoutBatchNormTestOverParameters( + self._TestQuantize_AtrousConvWithoutBatchNorm) + + def _TestQuantize_AtrousConvWithoutBatchNorm( + self, activation, activation_op_name, with_bypass, delay, use_resource): + """Tests quantization: inputs -> atrous conv no batch norm -> Activation. + + Args: + activation: Callable that returns an Operation, a factory method for the + Activation. + activation_op_name: String, name of the Activation operation. + with_bypass: Bool, when true there is an extra connection added from + inputs to just before Activation. + delay: Int (optional), delay in number of steps until quantization starts. + use_resource: Bool, when true uses resource variables. + """ + graph = ops.Graph() + with graph.as_default(): + variable_scope.get_variable_scope().set_use_resource(use_resource) + batch_size, height, width, depth = 5, 128, 128, 3 + inputs = array_ops.zeros((batch_size, height, width, depth)) + dilation_rate = 2 + activation_fn = None if with_bypass else activation + scope = 'test/test2' if with_bypass else 'test' + node = separable_conv2d( + inputs, + None, [3, 3], + rate=dilation_rate, + depth_multiplier=1.0, + padding='SAME', + weights_initializer=self._WeightInit(0.09), + activation_fn=activation_fn, + scope=scope) + if with_bypass: + node = math_ops.add(inputs, node, name='test/Add') + node = activation(node, name='test/' + activation_op_name) + update_barrier = control_flow_ops.no_op(name='update_barrier') + with ops.control_dependencies([update_barrier]): + array_ops.identity(node, name='control_dependency') + quantize.Quantize(graph, True, quant_delay=delay) + + self._AssertCorrectQuantizedGraphWithoutBatchNorm( + graph, scope, 'DepthwiseConv2dNative', activation_op_name, with_bypass, + delay, use_resource) + def _RunBatchNormTestOverParameters(self, test_fn): # TODO(suharshs): Use parameterized test once OSS TF supports it. parameters_list = [ @@ -543,6 +589,61 @@ class QuantizeTest(test_util.TensorFlowTestCase): graph, scope, 'DepthwiseConv2dNative', activation_op_name, with_bypass, delay, use_resource) + def testQuantize_AtrousConvWithBatchNorm(self): + self._RunBatchNormTestOverParameters( + self._TestQuantize_AtrousConvWithBatchNorm) + + def _TestQuantize_AtrousConvWithBatchNorm( + self, activation, activation_op_name, with_bypass, delay, + fused_batch_norm, use_resource): + """Tests quantization: inputs -> atrous conv with batch norm -> Activation. + + Args: + activation: Callable that returns an Operation, a factory method for the + Activation. + activation_op_name: String, name of the Activation operation. + with_bypass: Bool, when true there is an extra connection added from + inputs to just before Activation. + delay: Int (optional), delay in number of steps until quantization starts. + fused_batch_norm: Bool, when true use FusedBatchNorm. + use_resource: Bool, when true uses resource variables. + """ + graph = ops.Graph() + with graph.as_default(): + variable_scope.get_variable_scope().set_use_resource(use_resource) + batch_size, height, width, depth = 5, 128, 128, 3 + inputs = array_ops.zeros((batch_size, height, width, depth)) + dilation_rate = 2 + scope = 'test/test2' if with_bypass else 'test' + node = separable_conv2d( + inputs, + None, [3, 3], + rate=dilation_rate, + depth_multiplier=1.0, + padding='SAME', + weights_initializer=self._WeightInit(0.09), + activation_fn=None, + normalizer_fn=batch_norm, + normalizer_params=self._BatchNormParams(fused_batch_norm), + scope=scope) + + # Manually add a bypass (optional) and an activation. + if with_bypass: + node = math_ops.add(inputs, node, name='test/Add') + + node = activation(node, name='test/' + activation_op_name) + + update_barrier = control_flow_ops.no_op(name='update_barrier') + with ops.control_dependencies([update_barrier]): + array_ops.identity(node, name='control_dependency') + + fold_batch_norms.FoldBatchNorms(graph, is_training=True) + quantize.Quantize(graph, True, quant_delay=delay) + + self._AssertCorrectQuantizedGraphWithBatchNorm( + graph, scope, 'DepthwiseConv2dNative', activation_op_name, + with_bypass, delay, use_resource) + def _AssertIdempotent(self, graph): # Ensure that calling the rewrite again doesn't change the graph. graph_def_before = str(graph.as_graph_def()) @@ -553,8 +654,80 @@ class QuantizeTest(test_util.TensorFlowTestCase): graph_def_after = str(graph.as_graph_def()) self.assertEqual(graph_def_before, graph_def_after) - def _BatchNormParams(self, fused=False): - return {'center': True, 'scale': True, 'decay': 1.0 - 0.003, 'fused': fused} + def testBatchNormForcedUpdates(self): + parameter_list = [ + # (activation, activation_op_name, fused_batch_norm) + (nn_ops.relu6, 'Relu6', False), + (nn_ops.relu, 'Relu', False), + (array_ops.identity, 'Identity', False), + (nn_ops.relu6, 'Relu6', True), + (nn_ops.relu, 'Relu', True), + (array_ops.identity, 'Identity', True), + ] + for params in parameter_list: + self._TestBatchNormForcedUpdates(params[0], params[1], params[2], False) + self._TestBatchNormForcedUpdates(params[0], params[1], params[2], True) + + def _TestBatchNormForcedUpdates(self, activation, activation_op_name, + fused_batch_norm, use_resource): + """post_activation bypass quantization should happen with forced updates.""" + graph = ops.Graph() + with graph.as_default(): + variable_scope.get_variable_scope().set_use_resource(use_resource) + batch_size, height, width, depth = 5, 128, 128, 3 + input1 = array_ops.zeros((batch_size, height, width, depth)) + input2 = array_ops.zeros((batch_size, height / 2, width / 2, 32)) + # Setting updates_collections to None forces updates adding an extra + # identity operation following batch norms. + bn_params = self._BatchNormParams( + fused=fused_batch_norm, force_updates=True) + conv = conv2d( + input1, + 32, [5, 5], + stride=2, + padding='SAME', + weights_initializer=self._WeightInit(0.09), + activation_fn=activation, + normalizer_fn=batch_norm, + normalizer_params=bn_params, + scope='test/test') + bypass_tensor = math_ops.add(conv, input2, name='test/add') + # The output of the post_activation bypass will be another layer. + _ = conv2d( + bypass_tensor, + 32, [5, 5], + stride=2, + padding='SAME', + weights_initializer=self._WeightInit(0.09), + normalizer_fn=batch_norm, + normalizer_params=bn_params, + activation_fn=activation, + scope='test/unused') + + fold_batch_norms.FoldBatchNorms(graph, is_training=True) + quantize.Quantize(graph, is_training=True) + + # Ensure that the bypass node is preceded by and followed by a + # FakeQuantWithMinMaxVar operation, since the output of the Add isn't an + # activation. + self.assertTrue('FakeQuantWithMinMaxVars' in + [c.type for c in bypass_tensor.consumers()]) + self.assertTrue('FakeQuantWithMinMaxVars' in + [i.op.type for i in bypass_tensor.op.inputs]) + + with open('/tmp/bn_quant_test.pbtxt', 'w') as f: + f.write(str(graph.as_graph_def())) + + def _BatchNormParams(self, fused=False, force_updates=False): + params = { + 'center': True, + 'scale': True, + 'decay': 1.0 - 0.003, + 'fused': fused + } + if force_updates: + params['updates_collections'] = None + return params def _WeightInit(self, stddev): """Returns truncated normal variable initializer. diff --git a/tensorflow/contrib/rnn/BUILD b/tensorflow/contrib/rnn/BUILD index 4eb5c920b3517a8968ff730003e786ae2a9c9e26..2a84629080d20e38807a4be87e51646c3046ebf3 100644 --- a/tensorflow/contrib/rnn/BUILD +++ b/tensorflow/contrib/rnn/BUILD @@ -118,7 +118,6 @@ cuda_py_tests( "//tensorflow/python:framework_for_generated_wrappers", "//tensorflow/python:init_ops", "//tensorflow/python:math_ops", - "//tensorflow/python:random_ops", "//tensorflow/python:rnn", "//tensorflow/python:rnn_cell", "//tensorflow/python:variable_scope", diff --git a/tensorflow/contrib/rnn/__init__.py b/tensorflow/contrib/rnn/__init__.py index 67f31785b57fddef67733c18c3b744322532c28c..cb437f2a2f252fcb0763587b07fed19be5887282 100644 --- a/tensorflow/contrib/rnn/__init__.py +++ b/tensorflow/contrib/rnn/__init__.py @@ -58,6 +58,10 @@ See @{$python/contrib.rnn} guide. @@Conv3DLSTMCell @@HighwayWrapper @@GLSTMCell +@@SRUCell +@@IndRNNCell +@@IndyGRUCell +@@IndyLSTMCell @@AttentionCellWrapper diff --git a/tensorflow/contrib/rnn/python/kernel_tests/core_rnn_cell_test.py b/tensorflow/contrib/rnn/python/kernel_tests/core_rnn_cell_test.py index 86f1e27abd53d011f37f06851dd6d0977853c8f4..85f0f8ced91e15cd0f9b3bc51f3a9e3aee12c978 100644 --- a/tensorflow/contrib/rnn/python/kernel_tests/core_rnn_cell_test.py +++ b/tensorflow/contrib/rnn/python/kernel_tests/core_rnn_cell_test.py @@ -18,7 +18,6 @@ from __future__ import absolute_import from __future__ import division from __future__ import print_function -import functools import os import numpy as np @@ -35,7 +34,6 @@ from tensorflow.python.framework import test_util from tensorflow.python.ops import array_ops from tensorflow.python.ops import init_ops from tensorflow.python.ops import math_ops -from tensorflow.python.ops import random_ops from tensorflow.python.ops import rnn from tensorflow.python.ops import rnn_cell_impl from tensorflow.python.ops import variable_scope @@ -117,6 +115,27 @@ class RNNCellTest(test.TestCase): }) self.assertEqual(res[0].shape, (1, 2)) + def testIndRNNCell(self): + with self.test_session() as sess: + with variable_scope.variable_scope( + "root", initializer=init_ops.constant_initializer(0.5)): + x = array_ops.zeros([1, 2]) + m = array_ops.zeros([1, 2]) + cell = contrib_rnn_cell.IndRNNCell(2) + g, _ = cell(x, m) + self.assertEqual([ + "root/ind_rnn_cell/%s_w:0" % rnn_cell_impl._WEIGHTS_VARIABLE_NAME, + "root/ind_rnn_cell/%s_u:0" % rnn_cell_impl._WEIGHTS_VARIABLE_NAME, + "root/ind_rnn_cell/%s:0" % rnn_cell_impl._BIAS_VARIABLE_NAME + ], [v.name for v in cell.trainable_variables]) + self.assertFalse(cell.non_trainable_variables) + sess.run([variables_lib.global_variables_initializer()]) + res = sess.run([g], { + x.name: np.array([[1., 1.]]), + m.name: np.array([[0.1, 0.1]]) + }) + self.assertEqual(res[0].shape, (1, 2)) + def testGRUCell(self): with self.test_session() as sess: with variable_scope.variable_scope( @@ -145,6 +164,34 @@ class RNNCellTest(test.TestCase): # Smoke test self.assertAllClose(res[0], [[0.156736, 0.156736]]) + def testIndyGRUCell(self): + with self.test_session() as sess: + with variable_scope.variable_scope( + "root", initializer=init_ops.constant_initializer(0.5)): + x = array_ops.zeros([1, 2]) + m = array_ops.zeros([1, 2]) + g, _ = contrib_rnn_cell.IndyGRUCell(2)(x, m) + sess.run([variables_lib.global_variables_initializer()]) + res = sess.run([g], { + x.name: np.array([[1., 1.]]), + m.name: np.array([[0.1, 0.1]]) + }) + # Smoke test + self.assertAllClose(res[0], [[0.185265, 0.17704]]) + with variable_scope.variable_scope( + "other", initializer=init_ops.constant_initializer(0.5)): + # Test IndyGRUCell with input_size != num_units. + x = array_ops.zeros([1, 3]) + m = array_ops.zeros([1, 2]) + g, _ = contrib_rnn_cell.IndyGRUCell(2)(x, m) + sess.run([variables_lib.global_variables_initializer()]) + res = sess.run([g], { + x.name: np.array([[1., 1., 1.]]), + m.name: np.array([[0.1, 0.1]]) + }) + # Smoke test + self.assertAllClose(res[0], [[0.155127, 0.157328]]) + def testSRUCell(self): with self.test_session() as sess: with variable_scope.variable_scope( @@ -345,6 +392,72 @@ class RNNCellTest(test.TestCase): self.assertAllClose(res[1], expected_mem0) self.assertAllClose(res[2], expected_mem1) + def testIndyLSTMCell(self): + for dtype in [dtypes.float16, dtypes.float32]: + np_dtype = dtype.as_numpy_dtype + with self.test_session(graph=ops.Graph()) as sess: + with variable_scope.variable_scope( + "root", initializer=init_ops.constant_initializer(0.5)): + x = array_ops.zeros([1, 2], dtype=dtype) + state_0 = (array_ops.zeros([1, 2], dtype=dtype),) * 2 + state_1 = (array_ops.zeros([1, 2], dtype=dtype),) * 2 + cell = rnn_cell_impl.MultiRNNCell( + [contrib_rnn_cell.IndyLSTMCell(2) for _ in range(2)]) + self.assertEqual(cell.dtype, None) + self.assertEqual("cell-0", cell._checkpoint_dependencies[0].name) + self.assertEqual("cell-1", cell._checkpoint_dependencies[1].name) + cell.get_config() # Should not throw an error + g, (out_state_0, out_state_1) = cell(x, (state_0, state_1)) + # Layer infers the input type. + self.assertEqual(cell.dtype, dtype.name) + expected_variable_names = [ + "root/multi_rnn_cell/cell_0/indy_lstm_cell/%s_w:0" % + rnn_cell_impl._WEIGHTS_VARIABLE_NAME, + "root/multi_rnn_cell/cell_0/indy_lstm_cell/%s_u:0" % + rnn_cell_impl._WEIGHTS_VARIABLE_NAME, + "root/multi_rnn_cell/cell_0/indy_lstm_cell/%s:0" % + rnn_cell_impl._BIAS_VARIABLE_NAME, + "root/multi_rnn_cell/cell_1/indy_lstm_cell/%s_w:0" % + rnn_cell_impl._WEIGHTS_VARIABLE_NAME, + "root/multi_rnn_cell/cell_1/indy_lstm_cell/%s_u:0" % + rnn_cell_impl._WEIGHTS_VARIABLE_NAME, + "root/multi_rnn_cell/cell_1/indy_lstm_cell/%s:0" % + rnn_cell_impl._BIAS_VARIABLE_NAME + ] + self.assertEqual(expected_variable_names, + [v.name for v in cell.trainable_variables]) + self.assertFalse(cell.non_trainable_variables) + sess.run([variables_lib.global_variables_initializer()]) + res = sess.run( + [g, out_state_0, out_state_1], { + x.name: np.array([[1., 1.]]), + state_0[0].name: 0.1 * np.ones([1, 2]), + state_0[1].name: 0.1 * np.ones([1, 2]), + state_1[0].name: 0.1 * np.ones([1, 2]), + state_1[1].name: 0.1 * np.ones([1, 2]), + }) + self.assertEqual(len(res), 3) + variables = variables_lib.global_variables() + self.assertEqual(expected_variable_names, [v.name for v in variables]) + # Only check the range of outputs as this is just a smoke test. + self.assertAllInRange(res[0], -1.0, 1.0) + self.assertAllInRange(res[1], -1.0, 1.0) + self.assertAllInRange(res[2], -1.0, 1.0) + with variable_scope.variable_scope( + "other", initializer=init_ops.constant_initializer(0.5)): + # Test IndyLSTMCell with input_size != num_units. + x = array_ops.zeros([1, 3], dtype=dtype) + state = (array_ops.zeros([1, 2], dtype=dtype),) * 2 + g, out_state = contrib_rnn_cell.IndyLSTMCell(2)(x, state) + sess.run([variables_lib.global_variables_initializer()]) + res = sess.run( + [g, out_state], { + x.name: np.array([[1., 1., 1.]], dtype=np_dtype), + state[0].name: 0.1 * np.ones([1, 2], dtype=np_dtype), + state[1].name: 0.1 * np.ones([1, 2], dtype=np_dtype), + }) + self.assertEqual(len(res), 2) + def testLSTMCell(self): with self.test_session() as sess: num_units = 8 @@ -935,50 +1048,6 @@ class DropoutWrapperTest(test.TestCase): self.assertAllClose(res0[1].h, res1[1].h) -class SlimRNNCellTest(test.TestCase): - - def testBasicRNNCell(self): - with self.test_session() as sess: - with variable_scope.variable_scope( - "root", initializer=init_ops.constant_initializer(0.5)): - x = array_ops.zeros([1, 2]) - m = array_ops.zeros([1, 2]) - my_cell = functools.partial(basic_rnn_cell, num_units=2) - # pylint: disable=protected-access - g, _ = rnn_cell_impl._SlimRNNCell(my_cell)(x, m) - # pylint: enable=protected-access - sess.run([variables_lib.global_variables_initializer()]) - res = sess.run([g], { - x.name: np.array([[1., 1.]]), - m.name: np.array([[0.1, 0.1]]) - }) - self.assertEqual(res[0].shape, (1, 2)) - - def testBasicRNNCellMatch(self): - batch_size = 32 - input_size = 100 - num_units = 10 - with self.test_session() as sess: - with variable_scope.variable_scope( - "root", initializer=init_ops.constant_initializer(0.5)): - inputs = random_ops.random_uniform((batch_size, input_size)) - _, initial_state = basic_rnn_cell(inputs, None, num_units) - rnn_cell = rnn_cell_impl.BasicRNNCell(num_units) - outputs, state = rnn_cell(inputs, initial_state) - variable_scope.get_variable_scope().reuse_variables() - my_cell = functools.partial(basic_rnn_cell, num_units=num_units) - # pylint: disable=protected-access - slim_cell = rnn_cell_impl._SlimRNNCell(my_cell) - # pylint: enable=protected-access - slim_outputs, slim_state = slim_cell(inputs, initial_state) - self.assertEqual(slim_outputs.get_shape(), outputs.get_shape()) - self.assertEqual(slim_state.get_shape(), state.get_shape()) - sess.run([variables_lib.global_variables_initializer()]) - res = sess.run([slim_outputs, slim_state, outputs, state]) - self.assertAllClose(res[0], res[2]) - self.assertAllClose(res[1], res[3]) - - def basic_rnn_cell(inputs, state, num_units, scope=None): if state is None: if inputs is not None: diff --git a/tensorflow/contrib/rnn/python/ops/rnn_cell.py b/tensorflow/contrib/rnn/python/ops/rnn_cell.py index b12e2cd5eddc3f8abdba62781692673a40e41d9b..1816b469ee5bf338453a82d18663f97f6565dc0c 100644 --- a/tensorflow/contrib/rnn/python/ops/rnn_cell.py +++ b/tensorflow/contrib/rnn/python/ops/rnn_cell.py @@ -23,6 +23,7 @@ import math from tensorflow.contrib.compiler import jit from tensorflow.contrib.layers.python.layers import layers from tensorflow.contrib.rnn.python.ops import core_rnn_cell +from tensorflow.python.framework import constant_op from tensorflow.python.framework import dtypes from tensorflow.python.framework import op_def_registry from tensorflow.python.framework import ops @@ -30,6 +31,7 @@ from tensorflow.python.framework import tensor_shape from tensorflow.python.layers import base as base_layer from tensorflow.python.ops import array_ops from tensorflow.python.ops import clip_ops +from tensorflow.python.ops import gen_array_ops from tensorflow.python.ops import init_ops from tensorflow.python.ops import math_ops from tensorflow.python.ops import nn_impl # pylint: disable=unused-import @@ -3050,3 +3052,343 @@ class WeightNormLSTMCell(rnn_cell_impl.RNNCell): new_state = rnn_cell_impl.LSTMStateTuple(new_c, new_h) return new_h, new_state + + +class IndRNNCell(rnn_cell_impl.LayerRNNCell): + """Independently Recurrent Neural Network (IndRNN) cell + (cf. https://arxiv.org/abs/1803.04831). + + Args: + num_units: int, The number of units in the RNN cell. + activation: Nonlinearity to use. Default: `tanh`. + reuse: (optional) Python boolean describing whether to reuse variables + in an existing scope. If not `True`, and the existing scope already has + the given variables, an error is raised. + name: String, the name of the layer. Layers with the same name will + share weights, but to avoid mistakes we require reuse=True in such + cases. + dtype: Default dtype of the layer (default of `None` means use the type + of the first input). Required when `build` is called before `call`. + """ + + def __init__(self, + num_units, + activation=None, + reuse=None, + name=None, + dtype=None): + super(IndRNNCell, self).__init__(_reuse=reuse, name=name, dtype=dtype) + + # Inputs must be 2-dimensional. + self.input_spec = base_layer.InputSpec(ndim=2) + + self._num_units = num_units + self._activation = activation or math_ops.tanh + + @property + def state_size(self): + return self._num_units + + @property + def output_size(self): + return self._num_units + + def build(self, inputs_shape): + if inputs_shape[1].value is None: + raise ValueError( + "Expected inputs.shape[-1] to be known, saw shape: %s" % inputs_shape) + + input_depth = inputs_shape[1].value + # pylint: disable=protected-access + self._kernel_w = self.add_variable( + "%s_w" % rnn_cell_impl._WEIGHTS_VARIABLE_NAME, + shape=[input_depth, self._num_units]) + self._kernel_u = self.add_variable( + "%s_u" % rnn_cell_impl._WEIGHTS_VARIABLE_NAME, + shape=[1, self._num_units], + initializer=init_ops.random_uniform_initializer( + minval=-1, maxval=1, dtype=self.dtype)) + self._bias = self.add_variable( + rnn_cell_impl._BIAS_VARIABLE_NAME, + shape=[self._num_units], + initializer=init_ops.zeros_initializer(dtype=self.dtype)) + # pylint: enable=protected-access + + self.built = True + + def call(self, inputs, state): + """IndRNN: output = new_state = act(W * input + u * state + B).""" + + gate_inputs = math_ops.matmul(inputs, self._kernel_w) + ( + state * self._kernel_u) + gate_inputs = nn_ops.bias_add(gate_inputs, self._bias) + output = self._activation(gate_inputs) + return output, output + + +class IndyGRUCell(rnn_cell_impl.LayerRNNCell): + r"""Independently Gated Recurrent Unit cell. + + Based on IndRNNs (https://arxiv.org/abs/1803.04831) and similar to GRUCell, + yet with the \(U_r\), \(U_z\), and \(U\) matrices in equations 5, 6, and + 8 of http://arxiv.org/abs/1406.1078 respectively replaced by diagonal + matrices, i.e. a Hadamard product with a single vector: + + $$r_j = \sigma\left([\mathbf W_r\mathbf x]_j + + [\mathbf u_r\circ \mathbf h_{(t-1)}]_j\right)$$ + $$z_j = \sigma\left([\mathbf W_z\mathbf x]_j + + [\mathbf u_z\circ \mathbf h_{(t-1)}]_j\right)$$ + $$\tilde{h}^{(t)}_j = \phi\left([\mathbf W \mathbf x]_j + + [\mathbf u \circ \mathbf r \circ \mathbf h_{(t-1)}]_j\right)$$ + + where \(\circ\) denotes the Hadamard operator. This means that each IndyGRU + node sees only its own state, as opposed to seeing all states in the same + layer. + + TODO(gonnet): Write a paper describing this and add a reference here. + + Args: + num_units: int, The number of units in the GRU cell. + activation: Nonlinearity to use. Default: `tanh`. + reuse: (optional) Python boolean describing whether to reuse variables + in an existing scope. If not `True`, and the existing scope already has + the given variables, an error is raised. + kernel_initializer: (optional) The initializer to use for the weight + matrices applied to the input. + bias_initializer: (optional) The initializer to use for the bias. + name: String, the name of the layer. Layers with the same name will + share weights, but to avoid mistakes we require reuse=True in such + cases. + dtype: Default dtype of the layer (default of `None` means use the type + of the first input). Required when `build` is called before `call`. + """ + + def __init__(self, + num_units, + activation=None, + reuse=None, + kernel_initializer=None, + bias_initializer=None, + name=None, + dtype=None): + super(IndyGRUCell, self).__init__(_reuse=reuse, name=name, dtype=dtype) + + # Inputs must be 2-dimensional. + self.input_spec = base_layer.InputSpec(ndim=2) + + self._num_units = num_units + self._activation = activation or math_ops.tanh + self._kernel_initializer = kernel_initializer + self._bias_initializer = bias_initializer + + @property + def state_size(self): + return self._num_units + + @property + def output_size(self): + return self._num_units + + def build(self, inputs_shape): + if inputs_shape[1].value is None: + raise ValueError( + "Expected inputs.shape[-1] to be known, saw shape: %s" % inputs_shape) + + input_depth = inputs_shape[1].value + # pylint: disable=protected-access + self._gate_kernel_w = self.add_variable( + "gates/%s_w" % rnn_cell_impl._WEIGHTS_VARIABLE_NAME, + shape=[input_depth, 2 * self._num_units], + initializer=self._kernel_initializer) + self._gate_kernel_u = self.add_variable( + "gates/%s_u" % rnn_cell_impl._WEIGHTS_VARIABLE_NAME, + shape=[1, 2 * self._num_units], + initializer=init_ops.random_uniform_initializer( + minval=-1, maxval=1, dtype=self.dtype)) + self._gate_bias = self.add_variable( + "gates/%s" % rnn_cell_impl._BIAS_VARIABLE_NAME, + shape=[2 * self._num_units], + initializer=(self._bias_initializer + if self._bias_initializer is not None else + init_ops.constant_initializer(1.0, dtype=self.dtype))) + self._candidate_kernel_w = self.add_variable( + "candidate/%s" % rnn_cell_impl._WEIGHTS_VARIABLE_NAME, + shape=[input_depth, self._num_units], + initializer=self._kernel_initializer) + self._candidate_kernel_u = self.add_variable( + "candidate/%s_u" % rnn_cell_impl._WEIGHTS_VARIABLE_NAME, + shape=[1, self._num_units], + initializer=init_ops.random_uniform_initializer( + minval=-1, maxval=1, dtype=self.dtype)) + self._candidate_bias = self.add_variable( + "candidate/%s" % rnn_cell_impl._BIAS_VARIABLE_NAME, + shape=[self._num_units], + initializer=(self._bias_initializer + if self._bias_initializer is not None else + init_ops.zeros_initializer(dtype=self.dtype))) + # pylint: enable=protected-access + + self.built = True + + def call(self, inputs, state): + """Gated recurrent unit (GRU) with nunits cells.""" + + gate_inputs = math_ops.matmul(inputs, self._gate_kernel_w) + ( + gen_array_ops.tile(state, [1, 2]) * self._gate_kernel_u) + gate_inputs = nn_ops.bias_add(gate_inputs, self._gate_bias) + + value = math_ops.sigmoid(gate_inputs) + r, u = array_ops.split(value=value, num_or_size_splits=2, axis=1) + + r_state = r * state + + candidate = math_ops.matmul(inputs, self._candidate_kernel_w) + ( + r_state * self._candidate_kernel_u) + candidate = nn_ops.bias_add(candidate, self._candidate_bias) + + c = self._activation(candidate) + new_h = u * state + (1 - u) * c + return new_h, new_h + + +class IndyLSTMCell(rnn_cell_impl.LayerRNNCell): + r"""Basic IndyLSTM recurrent network cell. + + Based on IndRNNs (https://arxiv.org/abs/1803.04831) and similar to + BasicLSTMCell, yet with the \(U_f\), \(U_i\), \(U_o\) and \(U_c\) + matrices in + https://en.wikipedia.org/wiki/Long_short-term_memory#LSTM_with_a_forget_gate + replaced by diagonal matrices, i.e. a Hadamard product with a single vector: + + $$f_t = \sigma_g\left(W_f x_t + u_f \circ h_{t-1} + b_f\right)$$ + $$i_t = \sigma_g\left(W_i x_t + u_i \circ h_{t-1} + b_i\right)$$ + $$o_t = \sigma_g\left(W_o x_t + u_o \circ h_{t-1} + b_o\right)$$ + $$c_t = f_t \circ c_{t-1} + + i_t \circ \sigma_c\left(W_c x_t + u_c \circ h_{t-1} + b_c\right)$$ + + where \(\circ\) denotes the Hadamard operator. This means that each IndyLSTM + node sees only its own state \(h\) and \(c\), as opposed to seeing all + states in the same layer. + + We add forget_bias (default: 1) to the biases of the forget gate in order to + reduce the scale of forgetting in the beginning of the training. + + It does not allow cell clipping, a projection layer, and does not + use peep-hole connections: it is the basic baseline. + + For advanced models, please use the full @{tf.nn.rnn_cell.LSTMCell} + that follows. + + TODO(gonnet): Write a paper describing this and add a reference here. + """ + + def __init__(self, + num_units, + forget_bias=1.0, + activation=None, + reuse=None, + kernel_initializer=None, + bias_initializer=None, + name=None, + dtype=None): + """Initialize the IndyLSTM cell. + + Args: + num_units: int, The number of units in the LSTM cell. + forget_bias: float, The bias added to forget gates (see above). + Must set to `0.0` manually when restoring from CudnnLSTM-trained + checkpoints. + activation: Activation function of the inner states. Default: `tanh`. + reuse: (optional) Python boolean describing whether to reuse variables + in an existing scope. If not `True`, and the existing scope already has + the given variables, an error is raised. + kernel_initializer: (optional) The initializer to use for the weight + matrix applied to the inputs. + bias_initializer: (optional) The initializer to use for the bias. + name: String, the name of the layer. Layers with the same name will + share weights, but to avoid mistakes we require reuse=True in such + cases. + dtype: Default dtype of the layer (default of `None` means use the type + of the first input). Required when `build` is called before `call`. + """ + super(IndyLSTMCell, self).__init__(_reuse=reuse, name=name, dtype=dtype) + + # Inputs must be 2-dimensional. + self.input_spec = base_layer.InputSpec(ndim=2) + + self._num_units = num_units + self._forget_bias = forget_bias + self._activation = activation or math_ops.tanh + self._kernel_initializer = kernel_initializer + self._bias_initializer = bias_initializer + + @property + def state_size(self): + return rnn_cell_impl.LSTMStateTuple(self._num_units, self._num_units) + + @property + def output_size(self): + return self._num_units + + def build(self, inputs_shape): + if inputs_shape[1].value is None: + raise ValueError( + "Expected inputs.shape[-1] to be known, saw shape: %s" % inputs_shape) + + input_depth = inputs_shape[1].value + # pylint: disable=protected-access + self._kernel_w = self.add_variable( + "%s_w" % rnn_cell_impl._WEIGHTS_VARIABLE_NAME, + shape=[input_depth, 4 * self._num_units], + initializer=self._kernel_initializer) + self._kernel_u = self.add_variable( + "%s_u" % rnn_cell_impl._WEIGHTS_VARIABLE_NAME, + shape=[1, 4 * self._num_units], + initializer=init_ops.random_uniform_initializer( + minval=-1, maxval=1, dtype=self.dtype)) + self._bias = self.add_variable( + rnn_cell_impl._BIAS_VARIABLE_NAME, + shape=[4 * self._num_units], + initializer=(self._bias_initializer + if self._bias_initializer is not None else + init_ops.zeros_initializer(dtype=self.dtype))) + # pylint: enable=protected-access + + self.built = True + + def call(self, inputs, state): + """Independent Long short-term memory cell (IndyLSTM). + + Args: + inputs: `2-D` tensor with shape `[batch_size, input_size]`. + state: An `LSTMStateTuple` of state tensors, each shaped + `[batch_size, num_units]`. + + Returns: + A pair containing the new hidden state, and the new state (a + `LSTMStateTuple`). + """ + sigmoid = math_ops.sigmoid + one = constant_op.constant(1, dtype=dtypes.int32) + c, h = state + + gate_inputs = math_ops.matmul(inputs, self._kernel_w) + gate_inputs += gen_array_ops.tile(h, [1, 4]) * self._kernel_u + gate_inputs = nn_ops.bias_add(gate_inputs, self._bias) + + # i = input_gate, j = new_input, f = forget_gate, o = output_gate + i, j, f, o = array_ops.split( + value=gate_inputs, num_or_size_splits=4, axis=one) + + forget_bias_tensor = constant_op.constant(self._forget_bias, dtype=f.dtype) + # Note that using `add` and `multiply` instead of `+` and `*` gives a + # performance improvement. So using those at the cost of readability. + add = math_ops.add + multiply = math_ops.multiply + new_c = add( + multiply(c, sigmoid(add(f, forget_bias_tensor))), + multiply(sigmoid(i), self._activation(j))) + new_h = multiply(self._activation(new_c), sigmoid(o)) + + new_state = rnn_cell_impl.LSTMStateTuple(new_c, new_h) + return new_h, new_state diff --git a/tensorflow/contrib/rpc/python/kernel_tests/BUILD b/tensorflow/contrib/rpc/python/kernel_tests/BUILD index 2311c15a68c46090cec0f97bd950296506b0817e..cb0b89ae55b96361428c7845d4d6aab72543feb7 100644 --- a/tensorflow/contrib/rpc/python/kernel_tests/BUILD +++ b/tensorflow/contrib/rpc/python/kernel_tests/BUILD @@ -1,5 +1,3 @@ -# TODO(b/76425722): Port everything in here to OS (currently excluded). - package(default_visibility = ["//visibility:public"]) licenses(["notice"]) # Apache 2.0 @@ -17,7 +15,6 @@ tf_proto_library( srcs = ["test_example.proto"], has_services = 1, cc_api_version = 2, - protodeps = ["//tensorflow/core:protos_all"], ) py_library( diff --git a/tensorflow/contrib/rpc/python/kernel_tests/rpc_op_test_base.py b/tensorflow/contrib/rpc/python/kernel_tests/rpc_op_test_base.py index 27273d16b1c09eba60e124e632b353b09ea2d063..1c23c28860dac6203ea4ec8e808f63d3e9e467e2 100644 --- a/tensorflow/contrib/rpc/python/kernel_tests/rpc_op_test_base.py +++ b/tensorflow/contrib/rpc/python/kernel_tests/rpc_op_test_base.py @@ -51,23 +51,23 @@ class RpcOpTestBase(object): def testScalarHostPortRpc(self): with self.test_session() as sess: request_tensors = ( - test_example_pb2.TestCase(shape=[1, 2, 3]).SerializeToString()) + test_example_pb2.TestCase(values=[1, 2, 3]).SerializeToString()) response_tensors = self.rpc( - method=self.get_method_name('IncrementTestShapes'), + method=self.get_method_name('Increment'), address=self._address, request=request_tensors) self.assertEqual(response_tensors.shape, ()) response_values = sess.run(response_tensors) response_message = test_example_pb2.TestCase() self.assertTrue(response_message.ParseFromString(response_values)) - self.assertAllEqual([2, 3, 4], response_message.shape) + self.assertAllEqual([2, 3, 4], response_message.values) def testScalarHostPortTryRpc(self): with self.test_session() as sess: request_tensors = ( - test_example_pb2.TestCase(shape=[1, 2, 3]).SerializeToString()) + test_example_pb2.TestCase(values=[1, 2, 3]).SerializeToString()) response_tensors, status_code, status_message = self.try_rpc( - method=self.get_method_name('IncrementTestShapes'), + method=self.get_method_name('Increment'), address=self._address, request=request_tensors) self.assertEqual(status_code.shape, ()) @@ -77,7 +77,7 @@ class RpcOpTestBase(object): sess.run((response_tensors, status_code, status_message))) response_message = test_example_pb2.TestCase() self.assertTrue(response_message.ParseFromString(response_values)) - self.assertAllEqual([2, 3, 4], response_message.shape) + self.assertAllEqual([2, 3, 4], response_message.values) # For the base Rpc op, don't expect to get error status back. self.assertEqual(errors.OK, status_code_values) self.assertEqual(b'', status_message_values) @@ -86,7 +86,7 @@ class RpcOpTestBase(object): with self.test_session() as sess: request_tensors = [] response_tensors = self.rpc( - method=self.get_method_name('IncrementTestShapes'), + method=self.get_method_name('Increment'), address=self._address, request=request_tensors) self.assertAllEqual(response_tensors.shape, [0]) @@ -95,7 +95,7 @@ class RpcOpTestBase(object): def testInvalidMethod(self): for method in [ - '/InvalidService.IncrementTestShapes', + '/InvalidService.Increment', self.get_method_name('InvalidMethodName') ]: with self.test_session() as sess: @@ -115,12 +115,12 @@ class RpcOpTestBase(object): with self.assertRaises(errors.UnavailableError): sess.run( self.rpc( - method=self.get_method_name('IncrementTestShapes'), + method=self.get_method_name('Increment'), address=address, request='')) _, status_code_value, status_message_value = sess.run( self.try_rpc( - method=self.get_method_name('IncrementTestShapes'), + method=self.get_method_name('Increment'), address=address, request='')) self.assertEqual(errors.UNAVAILABLE, status_code_value) @@ -182,10 +182,10 @@ class RpcOpTestBase(object): with self.test_session() as sess: request_tensors = [ test_example_pb2.TestCase( - shape=[i, i + 1, i + 2]).SerializeToString() for i in range(20) + values=[i, i + 1, i + 2]).SerializeToString() for i in range(20) ] response_tensors = self.rpc( - method=self.get_method_name('IncrementTestShapes'), + method=self.get_method_name('Increment'), address=self._address, request=request_tensors) self.assertEqual(response_tensors.shape, (20,)) @@ -194,17 +194,17 @@ class RpcOpTestBase(object): for i in range(20): response_message = test_example_pb2.TestCase() self.assertTrue(response_message.ParseFromString(response_values[i])) - self.assertAllEqual([i + 1, i + 2, i + 3], response_message.shape) + self.assertAllEqual([i + 1, i + 2, i + 3], response_message.values) def testVecHostPortManyParallelRpcs(self): with self.test_session() as sess: request_tensors = [ test_example_pb2.TestCase( - shape=[i, i + 1, i + 2]).SerializeToString() for i in range(20) + values=[i, i + 1, i + 2]).SerializeToString() for i in range(20) ] many_response_tensors = [ self.rpc( - method=self.get_method_name('IncrementTestShapes'), + method=self.get_method_name('Increment'), address=self._address, request=request_tensors) for _ in range(10) ] @@ -216,25 +216,25 @@ class RpcOpTestBase(object): for i in range(20): response_message = test_example_pb2.TestCase() self.assertTrue(response_message.ParseFromString(response_values[i])) - self.assertAllEqual([i + 1, i + 2, i + 3], response_message.shape) + self.assertAllEqual([i + 1, i + 2, i + 3], response_message.values) def testVecHostPortRpcUsingEncodeAndDecodeProto(self): with self.test_session() as sess: request_tensors = encode_proto_op.encode_proto( message_type='tensorflow.contrib.rpc.TestCase', - field_names=['shape'], + field_names=['values'], sizes=[[3]] * 20, values=[ [[i, i + 1, i + 2] for i in range(20)], ]) response_tensor_strings = self.rpc( - method=self.get_method_name('IncrementTestShapes'), + method=self.get_method_name('Increment'), address=self._address, request=request_tensors) _, (response_shape,) = decode_proto_op.decode_proto( bytes=response_tensor_strings, message_type='tensorflow.contrib.rpc.TestCase', - field_names=['shape'], + field_names=['values'], output_types=[dtypes.int32]) response_shape_values = sess.run(response_shape) self.assertAllEqual([[i + 1, i + 2, i + 3] @@ -285,9 +285,9 @@ class RpcOpTestBase(object): addresses = flatten([[ self._address, 'unix:/tmp/this_unix_socket_doesnt_exist_97820348!!@' ] for _ in range(10)]) - request = test_example_pb2.TestCase(shape=[0, 1, 2]).SerializeToString() + request = test_example_pb2.TestCase(values=[0, 1, 2]).SerializeToString() response_tensors, status_code, _ = self.try_rpc( - method=self.get_method_name('IncrementTestShapes'), + method=self.get_method_name('Increment'), address=addresses, request=request) response_tensors_values, status_code_values = sess.run((response_tensors, @@ -303,9 +303,9 @@ class RpcOpTestBase(object): flatten = lambda x: list(itertools.chain.from_iterable(x)) with self.test_session() as sess: methods = flatten( - [[self.get_method_name('IncrementTestShapes'), 'InvalidMethodName'] + [[self.get_method_name('Increment'), 'InvalidMethodName'] for _ in range(10)]) - request = test_example_pb2.TestCase(shape=[0, 1, 2]).SerializeToString() + request = test_example_pb2.TestCase(values=[0, 1, 2]).SerializeToString() response_tensors, status_code, _ = self.try_rpc( method=methods, address=self._address, request=request) response_tensors_values, status_code_values = sess.run((response_tensors, @@ -325,10 +325,10 @@ class RpcOpTestBase(object): ] for _ in range(10)]) requests = [ test_example_pb2.TestCase( - shape=[i, i + 1, i + 2]).SerializeToString() for i in range(20) + values=[i, i + 1, i + 2]).SerializeToString() for i in range(20) ] response_tensors, status_code, _ = self.try_rpc( - method=self.get_method_name('IncrementTestShapes'), + method=self.get_method_name('Increment'), address=addresses, request=requests) response_tensors_values, status_code_values = sess.run((response_tensors, @@ -343,4 +343,4 @@ class RpcOpTestBase(object): response_message = test_example_pb2.TestCase() self.assertTrue( response_message.ParseFromString(response_tensors_values[i])) - self.assertAllEqual([i + 1, i + 2, i + 3], response_message.shape) + self.assertAllEqual([i + 1, i + 2, i + 3], response_message.values) diff --git a/tensorflow/contrib/rpc/python/kernel_tests/rpc_op_test_servicer.py b/tensorflow/contrib/rpc/python/kernel_tests/rpc_op_test_servicer.py index 7cbd636cb16e3befc9ae27cb231696634e859a22..265254aa51c64ff5a76ad3a9f7e081c56dd639e7 100644 --- a/tensorflow/contrib/rpc/python/kernel_tests/rpc_op_test_servicer.py +++ b/tensorflow/contrib/rpc/python/kernel_tests/rpc_op_test_servicer.py @@ -30,8 +30,8 @@ from tensorflow.contrib.rpc.python.kernel_tests import test_example_pb2_grpc class RpcOpTestServicer(test_example_pb2_grpc.TestCaseServiceServicer): """Test servicer for RpcOp tests.""" - def IncrementTestShapes(self, request, context): - """Increment the entries in the shape attribute of request. + def Increment(self, request, context): + """Increment the entries in the `values` attribute of request. Args: request: input TestCase. @@ -40,8 +40,8 @@ class RpcOpTestServicer(test_example_pb2_grpc.TestCaseServiceServicer): Returns: output TestCase. """ - for i in range(len(request.shape)): - request.shape[i] += 1 + for i in range(len(request.values)): + request.values[i] += 1 return request def AlwaysFailWithInvalidArgument(self, request, context): diff --git a/tensorflow/contrib/rpc/python/kernel_tests/test_example.proto b/tensorflow/contrib/rpc/python/kernel_tests/test_example.proto index 96f4550f62bc17e713abe1f3843ec0964f57b046..8141466349afcebcd104153a9f28c8f382458098 100644 --- a/tensorflow/contrib/rpc/python/kernel_tests/test_example.proto +++ b/tensorflow/contrib/rpc/python/kernel_tests/test_example.proto @@ -1,29 +1,17 @@ // Test description and protos to work with it. -// -// Many of the protos in this file are for unit tests that haven't been written yet. syntax = "proto2"; -import "tensorflow/core/framework/types.proto"; - package tensorflow.contrib.rpc; -// A TestCase holds a proto and a bunch of assertions -// about how it should decode. +// A TestCase holds a sequence of values. message TestCase { - // A batch of primitives to be serialized and decoded. - repeated RepeatedPrimitiveValue primitive = 1; - // The shape of the batch. - repeated int32 shape = 2; - // Expected sizes for each field. - repeated int32 sizes = 3; - // Expected values for each field. - repeated FieldSpec field = 4; + repeated int32 values = 1; }; service TestCaseService { - // Copy input, and increment each entry in 'shape' by 1. - rpc IncrementTestShapes(TestCase) returns (TestCase) { + // Copy input, and increment each entry in 'values' by 1. + rpc Increment(TestCase) returns (TestCase) { } // Sleep forever. @@ -42,130 +30,3 @@ service TestCaseService { rpc SometimesFailWithInvalidArgument(TestCase) returns (TestCase) { } }; - -// FieldSpec describes the expected output for a single field. -message FieldSpec { - optional string name = 1; - optional tensorflow.DataType dtype = 2; - optional RepeatedPrimitiveValue expected = 3; -}; - -message TestValue { - optional PrimitiveValue primitive_value = 1; - optional EnumValue enum_value = 2; - optional MessageValue message_value = 3; - optional RepeatedMessageValue repeated_message_value = 4; - optional RepeatedPrimitiveValue repeated_primitive_value = 6; -} - -message PrimitiveValue { - optional double double_value = 1; - optional float float_value = 2; - optional int64 int64_value = 3; - optional uint64 uint64_value = 4; - optional int32 int32_value = 5; - optional fixed64 fixed64_value = 6; - optional fixed32 fixed32_value = 7; - optional bool bool_value = 8; - optional string string_value = 9; - optional bytes bytes_value = 12; - optional uint32 uint32_value = 13; - optional sfixed32 sfixed32_value = 15; - optional sfixed64 sfixed64_value = 16; - optional sint32 sint32_value = 17; - optional sint64 sint64_value = 18; -} - -// NOTE: This definition must be kept in sync with PackedPrimitiveValue. -message RepeatedPrimitiveValue { - repeated double double_value = 1; - repeated float float_value = 2; - repeated int64 int64_value = 3; - repeated uint64 uint64_value = 4; - repeated int32 int32_value = 5; - repeated fixed64 fixed64_value = 6; - repeated fixed32 fixed32_value = 7; - repeated bool bool_value = 8; - repeated string string_value = 9; - repeated bytes bytes_value = 12; - repeated uint32 uint32_value = 13; - repeated sfixed32 sfixed32_value = 15; - repeated sfixed64 sfixed64_value = 16; - repeated sint32 sint32_value = 17; - repeated sint64 sint64_value = 18; - repeated PrimitiveValue message_value = 19; -} - -// A PackedPrimitiveValue looks exactly the same as a RepeatedPrimitiveValue -// in the text format, but the binary serializion is different. -// We test the packed representations by loading the same test cases -// using this definition instead of RepeatedPrimitiveValue. -// NOTE: This definition must be kept in sync with RepeatedPrimitiveValue -// in every way except the packed=true declaration. -message PackedPrimitiveValue { - repeated double double_value = 1 [packed = true]; - repeated float float_value = 2 [packed = true]; - repeated int64 int64_value = 3 [packed = true]; - repeated uint64 uint64_value = 4 [packed = true]; - repeated int32 int32_value = 5 [packed = true]; - repeated fixed64 fixed64_value = 6 [packed = true]; - repeated fixed32 fixed32_value = 7 [packed = true]; - repeated bool bool_value = 8 [packed = true]; - repeated string string_value = 9; - repeated bytes bytes_value = 12; - repeated uint32 uint32_value = 13 [packed = true]; - repeated sfixed32 sfixed32_value = 15 [packed = true]; - repeated sfixed64 sfixed64_value = 16 [packed = true]; - repeated sint32 sint32_value = 17 [packed = true]; - repeated sint64 sint64_value = 18 [packed = true]; - repeated PrimitiveValue message_value = 19; -} - -message EnumValue { - enum Color { - RED = 0; - ORANGE = 1; - YELLOW = 2; - GREEN = 3; - BLUE = 4; - INDIGO = 5; - VIOLET = 6; - }; - optional Color enum_value = 14; - repeated Color repeated_enum_value = 15; -} - - -message InnerMessageValue { - optional float float_value = 2; - repeated bytes bytes_values = 8; -} - -message MiddleMessageValue { - repeated int32 int32_values = 5; - optional InnerMessageValue message_value = 11; - optional uint32 uint32_value = 13; -} - -message MessageValue { - optional double double_value = 1; - optional MiddleMessageValue message_value = 11; -} - -message RepeatedMessageValue { - message NestedMessageValue { - optional float float_value = 2; - repeated bytes bytes_values = 8; - } - - repeated NestedMessageValue message_values = 11; -} - -// Message containing fields with field numbers higher than any field above. An -// instance of this message is prepended to each binary message in the test to -// exercise the code path that handles fields encoded out of order of field -// number. -message ExtraFields { - optional string string_value = 1776; - optional bool bool_value = 1777; -} diff --git a/tensorflow/contrib/seq2seq/python/ops/beam_search_decoder.py b/tensorflow/contrib/seq2seq/python/ops/beam_search_decoder.py index 184144f64a56358206014a0f75473b4a9b16617a..c7fbeea3105ae4c9c9ec2fd131f3468018990028 100644 --- a/tensorflow/contrib/seq2seq/python/ops/beam_search_decoder.py +++ b/tensorflow/contrib/seq2seq/python/ops/beam_search_decoder.py @@ -250,7 +250,7 @@ class BeamSearchDecoder(decoder.Decoder): ``` tiled_encoder_outputs = tf.contrib.seq2seq.tile_batch( encoder_outputs, multiplier=beam_width) - tiled_encoder_final_state = tf.conrib.seq2seq.tile_batch( + tiled_encoder_final_state = tf.contrib.seq2seq.tile_batch( encoder_final_state, multiplier=beam_width) tiled_sequence_length = tf.contrib.seq2seq.tile_batch( sequence_length, multiplier=beam_width) diff --git a/tensorflow/contrib/seq2seq/python/ops/decoder.py b/tensorflow/contrib/seq2seq/python/ops/decoder.py index e69725ff8ab1ba4de880c914a6f5fdad5e54566d..f58268eff525a4b592c79acb32207e1a3f62bdc7 100644 --- a/tensorflow/contrib/seq2seq/python/ops/decoder.py +++ b/tensorflow/contrib/seq2seq/python/ops/decoder.py @@ -21,6 +21,7 @@ from __future__ import print_function import abc import six +from tensorflow.python.eager import context from tensorflow.python.framework import constant_op from tensorflow.python.framework import dtypes from tensorflow.python.framework import ops @@ -182,19 +183,20 @@ def dynamic_decode(decoder, raise TypeError("Expected decoder to be type Decoder, but saw: %s" % type(decoder)) - def _is_xla_tensor(tensor): - try: - op = tensor.op - except AttributeError: - return False - if control_flow_util.IsInXLAContext(op): - return True - return False - with variable_scope.variable_scope(scope, "decoder") as varscope: - # Properly cache variable values inside the while_loop - if varscope.caching_device is None: - varscope.set_caching_device(lambda op: op.device) + # Determine context types. + ctxt = ops.get_default_graph()._get_control_flow_context() # pylint: disable=protected-access + is_xla = control_flow_util.GetContainingXLAContext(ctxt) is not None + in_while_loop = ( + control_flow_util.GetContainingWhileContext(ctxt) is not None) + # Properly cache variable values inside the while_loop. + # Don't set a caching device when running in a loop, since it is possible + # that train steps could be wrapped in a tf.while_loop. In that scenario + # caching prevents forward computations in loop iterations from re-reading + # the updated weights. + if not context.executing_eagerly() and not in_while_loop: + if varscope.caching_device is None: + varscope.set_caching_device(lambda op: op.device) if maximum_iterations is not None: maximum_iterations = ops.convert_to_tensor( @@ -208,9 +210,6 @@ def dynamic_decode(decoder, decoder.output_dtype, decoder.batch_size) - is_xla = False - if any([_is_xla_tensor(i) for i in nest.flatten(initial_inputs)]): - is_xla = True if is_xla and maximum_iterations is None: raise ValueError("maximum_iterations is required for XLA compilation.") if maximum_iterations is not None: diff --git a/tensorflow/contrib/slim/python/slim/evaluation_test.py b/tensorflow/contrib/slim/python/slim/evaluation_test.py index 3d0308aaf3da3b5b16fd22a2905db36917e8c97b..2c97834523424d0fab56330b4d9355a75427e0ef 100644 --- a/tensorflow/contrib/slim/python/slim/evaluation_test.py +++ b/tensorflow/contrib/slim/python/slim/evaluation_test.py @@ -33,7 +33,6 @@ from tensorflow.python.debug.lib import debug_data from tensorflow.python.debug.wrappers import hooks from tensorflow.python.framework import constant_op from tensorflow.python.framework import dtypes -from tensorflow.python.framework import errors from tensorflow.python.ops import control_flow_ops from tensorflow.python.ops import math_ops from tensorflow.python.ops import metrics @@ -242,7 +241,7 @@ class SingleEvaluationTest(test.TestCase): checkpoint_path = os.path.join(self.get_temp_dir(), 'this_file_doesnt_exist') log_dir = os.path.join(self.get_temp_dir(), 'error_raised') - with self.assertRaises(errors.NotFoundError): + with self.assertRaises(ValueError): evaluation.evaluate_once('', checkpoint_path, log_dir) def _prepareCheckpoint(self, checkpoint_path): diff --git a/tensorflow/contrib/summary/summary_ops_test.py b/tensorflow/contrib/summary/summary_ops_test.py index f1ef218e74bbd225071324a8269fdfeb5de0e038..3e41e3d0b48ea06f9cb8c1862e27eacb5ebc4417 100644 --- a/tensorflow/contrib/summary/summary_ops_test.py +++ b/tensorflow/contrib/summary/summary_ops_test.py @@ -81,6 +81,19 @@ class EagerFileTest(test_util.TensorFlowTestCase): # test here that we're calling them correctly. self.assertTrue(gfile.Exists(logdir)) + @test_util.assert_no_new_pyobjects_executing_eagerly + def testEagerMemory(self): + training_util.get_or_create_global_step() + logdir = self.get_temp_dir() + with summary_ops.create_file_writer( + logdir, max_queue=0, + name='t0').as_default(), summary_ops.always_record_summaries(): + summary_ops.generic('tensor', 1, '') + summary_ops.scalar('scalar', 2.0) + summary_ops.histogram('histogram', [1.0]) + summary_ops.image('image', [[[[1.0]]]]) + summary_ops.audio('audio', [[1.0]], 1.0, 1) + def testDefunSummarys(self): training_util.get_or_create_global_step() logdir = tempfile.mkdtemp() diff --git a/tensorflow/contrib/tensorboard/db/BUILD b/tensorflow/contrib/tensorboard/db/BUILD index 3f6b4cdc9ad10f5089f28af35a8be408918c7f90..6507546ee9f81108add181a9c83064c9860005e2 100644 --- a/tensorflow/contrib/tensorboard/db/BUILD +++ b/tensorflow/contrib/tensorboard/db/BUILD @@ -106,6 +106,7 @@ cc_library( "//tensorflow/core:framework", "//tensorflow/core:lib", "//tensorflow/core:lib_internal", + "//tensorflow/core:png_internal", "//tensorflow/core:protos_all_cc", ], ) diff --git a/tensorflow/contrib/tensorrt/convert/convert_graph.cc b/tensorflow/contrib/tensorrt/convert/convert_graph.cc index e7ce0821460a3e4fac0b9e0c896fc397dc7a4370..8578ca07f193c6da619ac03740575d29830c037d 100644 --- a/tensorflow/contrib/tensorrt/convert/convert_graph.cc +++ b/tensorflow/contrib/tensorrt/convert/convert_graph.cc @@ -149,7 +149,7 @@ tensorflow::Status ConvertCalibGraphToInferGraph( "Need to run graph with calibration data first!"); } if (cres->calibrator_) { - cres->calibrator_->setDone(); + cres->calibrator_->waitAndSetDone(); cres->thr_->join(); const auto& calibration_table = cres->calibrator_->getCalibrationTableAsString(); @@ -165,7 +165,7 @@ tensorflow::Status ConvertCalibGraphToInferGraph( "Can't get TRTCalibrator from resource manager!"); } cres->Unref(); - calib_rm->Cleanup(container_name); + TF_RETURN_IF_ERROR(calib_rm->Cleanup(container_name)); } } return tensorflow::Status::OK(); @@ -719,8 +719,8 @@ tensorflow::Status ConvertAfterShapes(ConversionParams& params) { } else { // Graph is not modified. LOG(WARNING) << "Engine creation for segment " << i << ", composed of " - << converted_segments.at(i).first.size() << " nodes failed: " - << status << ". Skipping..."; + << converted_segments.at(i).first.size() + << " nodes failed: " << status << ". Skipping..."; } } cudaSetDevice(old_cuda_device); diff --git a/tensorflow/contrib/tensorrt/convert/convert_nodes.h b/tensorflow/contrib/tensorrt/convert/convert_nodes.h index 7684d8d4a23ae22c855d82fc54931151a976eb2f..1a4c0e755d1cd1e88ac26c39996eb3a750421a0a 100644 --- a/tensorflow/contrib/tensorrt/convert/convert_nodes.h +++ b/tensorflow/contrib/tensorrt/convert/convert_nodes.h @@ -46,8 +46,8 @@ const int INT8MODE = 2; struct EngineConnection { EngineConnection(const string& outside, int out_id, int out_port, - const string& inside, int in_id, int in_port, - bool input_edge, int port) + const string& inside, int in_id, int in_port, + bool input_edge, int port) : outside_node_name(outside), outside_id(out_id), outside_port(out_port), diff --git a/tensorflow/contrib/tensorrt/kernels/trt_engine_op.cc b/tensorflow/contrib/tensorrt/kernels/trt_engine_op.cc index 75e32559bb055a49ccef2100c208c6277c0c4b60..8a17eb02f1af7c8f148c9cd4e14cc3876b6e13e3 100644 --- a/tensorflow/contrib/tensorrt/kernels/trt_engine_op.cc +++ b/tensorflow/contrib/tensorrt/kernels/trt_engine_op.cc @@ -319,7 +319,7 @@ void TRTEngineOp::ComputeAsync(tensorflow::OpKernelContext* ctx, default: LOG(ERROR) << "Unknown TRT data type: " << int(dtype); ctx->SetStatus(tensorflow::errors::InvalidArgument( - "Unknown ouput TRT data type! ", static_cast(dtype))); + "Unknown output TRT data type! ", static_cast(dtype))); return; } } @@ -327,8 +327,8 @@ void TRTEngineOp::ComputeAsync(tensorflow::OpKernelContext* ctx, for (int i = 0; i < ctx->num_outputs(); i++) { // Create an output tensor const string output_name = StrCat(kOutputPHName, i); - const size_t binding_index = trt_engine_ptr->getBindingIndex( - output_name.c_str()); + const size_t binding_index = + trt_engine_ptr->getBindingIndex(output_name.c_str()); Tensor* output_tensor = nullptr; TensorShape output_shape; @@ -371,7 +371,7 @@ void TRTEngineOp::ComputeAsync(tensorflow::OpKernelContext* ctx, default: LOG(ERROR) << "Unknown TRT data type: " << static_cast(dtype); ctx->SetStatus(tensorflow::errors::InvalidArgument( - "Unsupported output data type! ", int(dtype))); + "Unsupported output data type! ", static_cast(dtype))); return; } } @@ -420,10 +420,10 @@ nvinfer1::IGpuAllocator* TRTEngineOp::GetAllocator(OpKernelContext* ctx) { } TRTEngineOp::EngineCtxPair& TRTEngineOp::GetEngine(int batch_size, - OpKernelContext* ctx) { + OpKernelContext* ctx) { static EngineCtxPair null_pair = { - TrtUniquePtrType(nullptr), - TrtUniquePtrType(nullptr)}; + TrtUniquePtrType(nullptr), + TrtUniquePtrType(nullptr)}; // TODO(sami): This method needs to be re-written to use resource manager and // with LRU mechanism option. tensorflow::mutex_lock lock(engine_mutex_); @@ -450,9 +450,9 @@ TRTEngineOp::EngineCtxPair& TRTEngineOp::GetEngine(int batch_size, auto raw_static_engine = static_engine.get(); const auto max_batch_size = raw_static_engine->getMaxBatchSize(); engine_map_[max_batch_size] = { - std::move(static_engine), - TrtUniquePtrType( - raw_static_engine->createExecutionContext())}; + std::move(static_engine), + TrtUniquePtrType( + raw_static_engine->createExecutionContext())}; // Runtime is safe to delete after engine creation serialized_segment_.clear(); if (max_batch_size < batch_size) return null_pair; diff --git a/tensorflow/contrib/tensorrt/resources/trt_int8_calibrator.cc b/tensorflow/contrib/tensorrt/resources/trt_int8_calibrator.cc index 32e81858b95d76a2baebb4804a1326fbbb6144c7..dab1dd9343be7d5b033a3e04bf0b49fbbf37e9e5 100644 --- a/tensorflow/contrib/tensorrt/resources/trt_int8_calibrator.cc +++ b/tensorflow/contrib/tensorrt/resources/trt_int8_calibrator.cc @@ -36,13 +36,14 @@ TRTInt8Calibrator::TRTInt8Calibrator( : batch_size_(batch_size), done_(false), dev_buffers_(dev_buffers), + // Make sure setBatch() waits until getBatch() is called (the first time). calib_running_(true), batch_is_set_(false), engine_name_(engine_name) {} TRTInt8Calibrator::TRTInt8Calibrator(const string& calib_data) : batch_size_(0), - done_(false), + done_(true), calib_running_(false), batch_is_set_(false), calibration_table_(calib_data) {} @@ -50,13 +51,14 @@ TRTInt8Calibrator::TRTInt8Calibrator(const string& calib_data) bool TRTInt8Calibrator::setBatch(const std::unordered_map& data, const cudaStream_t stream) { tensorflow::mutex_lock lock(cond_mtx_); - // wait while calibration is running. - while ((calib_running_ || batch_is_set_) && !done_) { - cond_.wait(lock); - } + + // Wait while the queue is full or calibration is running. + while ((calib_running_ || batch_is_set_) && !done_) cond_.wait(lock); if (done_) return false; CHECK(!calib_running_ && !batch_is_set_); VLOG(1) << "Set Batch Waiting finished"; + + // Sets the batch. for (const auto it : data) { auto devptr = dev_buffers_.find(it.first); if (devptr == dev_buffers_.end()) { @@ -76,8 +78,8 @@ bool TRTInt8Calibrator::setBatch(const std::unordered_map& data, } // TODO(Sami, aaorey): Find an alternative way! - cudaStreamSynchronize( - stream); // we have to wait for the stream before returning! + // we have to wait for the stream before returning! + cudaStreamSynchronize(stream); batch_is_set_ = true; cond_.notify_all(); return true; @@ -86,21 +88,21 @@ bool TRTInt8Calibrator::setBatch(const std::unordered_map& data, bool TRTInt8Calibrator::getBatch(void** bindings, const char** names, int num_bindings) { tensorflow::mutex_lock lock(cond_mtx_); + // Notify finish of last round of calibration. calib_running_ = false; cond_.notify_all(); - // wait until new batch arrives - while ((!batch_is_set_ && !done_)) { - cond_.wait(lock); - } + + // Wait until new batch arrives + while ((!batch_is_set_ && !done_)) cond_.wait(lock); if (done_) return false; + // Gets the batch for (int i = 0; i < num_bindings; i++) { auto it = dev_buffers_.find(names[i]); if (it == dev_buffers_.end()) { LOG(FATAL) << "Calibration engine asked for unknown tensor name '" << names[i] << "' at position " << i; } - bindings[i] = it->second.first; } batch_is_set_ = false; @@ -108,6 +110,17 @@ bool TRTInt8Calibrator::getBatch(void** bindings, const char** names, return true; } +void TRTInt8Calibrator::waitAndSetDone() { + tensorflow::mutex_lock lock(cond_mtx_); + // Wait while the queue is full or calibration is running, so we don't miss + // the last batch. + while ((calib_running_ || batch_is_set_) && !done_) cond_.wait(lock); + if (!done_) { + done_ = true; + cond_.notify_all(); + } +} + const void* TRTInt8Calibrator::readCalibrationCache(std::size_t& length) { if (calibration_table_.empty()) return nullptr; length = calibration_table_.size(); diff --git a/tensorflow/contrib/tensorrt/resources/trt_int8_calibrator.h b/tensorflow/contrib/tensorrt/resources/trt_int8_calibrator.h index 994312d7c3c93ba04394b7d9542d261c57c5609b..65466c9741989fda5f82fc27d813d026f35fe386 100644 --- a/tensorflow/contrib/tensorrt/resources/trt_int8_calibrator.h +++ b/tensorflow/contrib/tensorrt/resources/trt_int8_calibrator.h @@ -36,10 +36,13 @@ namespace tensorrt { struct TRTInt8Calibrator : public nvinfer1::IInt8EntropyCalibrator { public: + // Construct a calibrator for future calibration. TRTInt8Calibrator( const std::unordered_map>& dev_buffers, int batch_size, string engine_name); + // Construct a finalized calibrator where we don't need to run calibration any + // more, as the calibration data is provided. TRTInt8Calibrator(const string& calibration_data); ~TRTInt8Calibrator(); @@ -52,6 +55,11 @@ struct TRTInt8Calibrator : public nvinfer1::IInt8EntropyCalibrator { bool setBatch(const std::unordered_map& data, const cudaStream_t stream); + // Wait until the last batch is consumed by the calibrator and set done. + void waitAndSetDone(); + + // Notify that calibration is done and future batches provided by setBatch() + // will be ignored. void setDone(); // If not null, calibration is skipped. diff --git a/tensorflow/contrib/timeseries/python/timeseries/BUILD b/tensorflow/contrib/timeseries/python/timeseries/BUILD index e4963596d38dbe8aea98fddbc67dbbf761c215c8..7020989d6895fd6322db45cda6f7dd99d417d937 100644 --- a/tensorflow/contrib/timeseries/python/timeseries/BUILD +++ b/tensorflow/contrib/timeseries/python/timeseries/BUILD @@ -157,6 +157,7 @@ py_library( py_test( name = "head_test", + size = "large", srcs = [ "head_test.py", ], @@ -184,6 +185,7 @@ py_test( "//tensorflow/python/saved_model:loader", "//tensorflow/python/saved_model:tag_constants", "//third_party/py/numpy", + "@absl_py//absl/testing:parameterized", "@six_archive//:six", ], ) diff --git a/tensorflow/contrib/timeseries/python/timeseries/estimators.py b/tensorflow/contrib/timeseries/python/timeseries/estimators.py index 4ec8d26116159fee3ac00581010d1603ac9e19f3..769183f40ad269954dac70db393207c266052144 100644 --- a/tensorflow/contrib/timeseries/python/timeseries/estimators.py +++ b/tensorflow/contrib/timeseries/python/timeseries/estimators.py @@ -288,7 +288,7 @@ class StateSpaceRegressor(TimeSeriesRegressor): """An Estimator for general state space models.""" def __init__(self, model, state_manager=None, optimizer=None, model_dir=None, - config=None): + config=None, head_type=ts_head_lib.TimeSeriesRegressionHead): """See TimeSeriesRegressor. Uses the ChainingStateManager by default.""" if not isinstance(model, state_space_model.StateSpaceModel): raise ValueError( @@ -301,7 +301,8 @@ class StateSpaceRegressor(TimeSeriesRegressor): state_manager=state_manager, optimizer=optimizer, model_dir=model_dir, - config=config) + config=config, + head_type=head_type) class StructuralEnsembleRegressor(StateSpaceRegressor): @@ -344,7 +345,8 @@ class StructuralEnsembleRegressor(StateSpaceRegressor): anomaly_prior_probability=None, optimizer=None, model_dir=None, - config=None): + config=None, + head_type=ts_head_lib.TimeSeriesRegressionHead): """Initialize the Estimator. Args: @@ -401,6 +403,8 @@ class StructuralEnsembleRegressor(StateSpaceRegressor): from tf.train.Optimizer. Defaults to Adam with step size 0.02. model_dir: See `Estimator`. config: See `Estimator`. + head_type: The kind of head to use for the model (inheriting from + `TimeSeriesRegressionHead`). """ if anomaly_prior_probability is not None: filtering_postprocessor = StateInterpolatingAnomalyDetector( @@ -424,4 +428,5 @@ class StructuralEnsembleRegressor(StateSpaceRegressor): model=model, optimizer=optimizer, model_dir=model_dir, - config=config) + config=config, + head_type=head_type) diff --git a/tensorflow/contrib/timeseries/python/timeseries/head.py b/tensorflow/contrib/timeseries/python/timeseries/head.py index f236329fdb038ba5ab432c6b97f44bda7ccfe815..8686a803e5bb023bbddb7df3203080fee0e13fea 100644 --- a/tensorflow/contrib/timeseries/python/timeseries/head.py +++ b/tensorflow/contrib/timeseries/python/timeseries/head.py @@ -19,11 +19,7 @@ from __future__ import print_function import re -from tensorflow.python.training import training_util -from tensorflow.contrib.layers.python.layers import optimizers - from tensorflow.contrib.timeseries.python.timeseries import feature_keys - from tensorflow.python.estimator import estimator_lib from tensorflow.python.estimator.canned import head as head_lib from tensorflow.python.estimator.canned import metric_keys @@ -35,8 +31,9 @@ from tensorflow.python.ops import control_flow_ops from tensorflow.python.ops import math_ops from tensorflow.python.ops import state_ops from tensorflow.python.ops import variable_scope -from tensorflow.python.util import nest from tensorflow.python.summary import summary +from tensorflow.python.training import training_util +from tensorflow.python.util import nest class _NoStatePredictOutput(export_lib.PredictOutput): @@ -102,12 +99,9 @@ class TimeSeriesRegressionHead(head_lib._Head): # pylint:disable=protected-acce use_resource=True): model_outputs = self.create_loss(features, mode) - train_op = optimizers.optimize_loss( + train_op = self.optimizer.minimize( model_outputs.loss, - global_step=training_util.get_global_step(), - optimizer=self.optimizer, - # Learning rate is set in the Optimizer object - learning_rate=None) + global_step=training_util.get_global_step()) return estimator_lib.EstimatorSpec( loss=model_outputs.loss, mode=mode, diff --git a/tensorflow/contrib/timeseries/python/timeseries/head_test.py b/tensorflow/contrib/timeseries/python/timeseries/head_test.py index ed8f29c321719e552c25f4d2183fdf4eb282e4b7..78c2cec21cf4b6ccf6c314e54de41f3e95466adf 100644 --- a/tensorflow/contrib/timeseries/python/timeseries/head_test.py +++ b/tensorflow/contrib/timeseries/python/timeseries/head_test.py @@ -18,6 +18,9 @@ from __future__ import absolute_import from __future__ import division from __future__ import print_function +import os + +from absl.testing import parameterized import numpy import six @@ -317,10 +320,38 @@ class PredictFeatureCheckingTests(test.TestCase): mode=estimator_lib.ModeKeys.PREDICT) -class OneShotTests(test.TestCase): - - def test_one_shot_prediction_head_export(self): - model_dir = self.get_temp_dir() +def _custom_time_series_regressor( + model_dir, head_type, exogenous_feature_columns): + return ts_estimators.TimeSeriesRegressor( + model=lstm_example._LSTMModel( + num_features=5, num_units=128, + exogenous_feature_columns=exogenous_feature_columns), + optimizer=adam.AdamOptimizer(0.001), + config=estimator_lib.RunConfig(tf_random_seed=4), + state_manager=state_management.ChainingStateManager(), + head_type=head_type, + model_dir=model_dir) + + +def _structural_ensemble_regressor( + model_dir, head_type, exogenous_feature_columns): + return ts_estimators.StructuralEnsembleRegressor( + periodicities=None, + num_features=5, + exogenous_feature_columns=exogenous_feature_columns, + head_type=head_type, + model_dir=model_dir) + + +class OneShotTests(parameterized.TestCase): + + @parameterized.named_parameters( + {"testcase_name": "custom_time_series_regressor", + "estimator_factory": _custom_time_series_regressor}, + {"testcase_name": "structural_ensemble_regressor", + "estimator_factory": _structural_ensemble_regressor}) + def test_one_shot_prediction_head_export(self, estimator_factory): + model_dir = os.path.join(test.get_temp_dir(), str(ops.uid())) categorical_column = feature_column.categorical_column_with_hash_bucket( key="categorical_exogenous_feature", hash_bucket_size=16) exogenous_feature_columns = [ @@ -328,15 +359,10 @@ class OneShotTests(test.TestCase): "2d_exogenous_feature", shape=(2,)), feature_column.embedding_column( categorical_column=categorical_column, dimension=10)] - estimator = ts_estimators.TimeSeriesRegressor( - model=lstm_example._LSTMModel( - num_features=5, num_units=128, - exogenous_feature_columns=exogenous_feature_columns), - optimizer=adam.AdamOptimizer(0.001), - config=estimator_lib.RunConfig(tf_random_seed=4), - state_manager=state_management.ChainingStateManager(), - head_type=ts_head_lib.OneShotPredictionHead, - model_dir=model_dir) + estimator = estimator_factory( + model_dir=model_dir, + exogenous_feature_columns=exogenous_feature_columns, + head_type=ts_head_lib.OneShotPredictionHead) train_features = { feature_keys.TrainEvalFeatures.TIMES: numpy.arange( 20, dtype=numpy.int64), @@ -351,7 +377,7 @@ class OneShotTests(test.TestCase): num_threads=1, batch_size=16, window_size=16) estimator.train(input_fn=train_input_fn, steps=5) input_receiver_fn = estimator.build_raw_serving_input_receiver_fn() - export_location = estimator.export_savedmodel(self.get_temp_dir(), + export_location = estimator.export_savedmodel(test.get_temp_dir(), input_receiver_fn) graph = ops.Graph() with graph.as_default(): @@ -385,7 +411,7 @@ class OneShotTests(test.TestCase): for output_key, output_value in predict_signature.outputs.items()} output = session.run(fetches, feed_dict=feeds) - self.assertAllEqual((2, 15, 5), output["mean"].shape) + self.assertEqual((2, 15, 5), output["mean"].shape) if __name__ == "__main__": diff --git a/tensorflow/contrib/tpu/BUILD b/tensorflow/contrib/tpu/BUILD index 16696793bc2dab977a3dbbfa338e33e5771d0699..0044fde9d020d79eafa5af4dd8fcb345ef102e1a 100644 --- a/tensorflow/contrib/tpu/BUILD +++ b/tensorflow/contrib/tpu/BUILD @@ -160,13 +160,44 @@ py_library( ], ) +py_library( + name = "keras_support", + srcs = [ + "python/tpu/keras_support.py", + ], + srcs_version = "PY2AND3", + deps = [ + ":tpu_lib", + ":tpu_py", + "//tensorflow/contrib/cluster_resolver:tpu_cluster_resolver_py", + "//tensorflow/contrib/distribute/python:tpu_strategy", + "//tensorflow/contrib/framework:framework_py", + "//tensorflow/contrib/tpu/proto:compilation_result_proto_py", + "//tensorflow/core:protos_all_py", + "//tensorflow/python:array_ops", + "//tensorflow/python:dtypes", + "//tensorflow/python:framework_ops", + "//tensorflow/python:linalg_ops", + "//tensorflow/python:math_ops", + "//tensorflow/python:platform", + "//tensorflow/python:random_ops", + "//tensorflow/python:session", + "//tensorflow/python:tensor_spec", + "//tensorflow/python:variable_scope", + "//tensorflow/python/estimator:model_fn", + "//tensorflow/python/keras:backend", + "//tensorflow/python/keras:engine", + "//tensorflow/python/keras:layers", + "//third_party/py/numpy", + ], +) + py_library( name = "tpu_lib", srcs = [ "python/tpu/__init__.py", "python/tpu/bfloat16.py", "python/tpu/device_assignment.py", - "python/tpu/keras_support.py", "python/tpu/session_support.py", "python/tpu/topology.py", "python/tpu/tpu.py", @@ -307,3 +338,13 @@ tf_py_test( "//tensorflow/python:framework_test_lib", ], ) + +tf_py_test( + name = "topology_test", + size = "small", + srcs = ["python/tpu/topology_test.py"], + additional_deps = [ + ":tpu", + "//tensorflow/python:framework_test_lib", + ], +) diff --git a/tensorflow/contrib/tpu/__init__.py b/tensorflow/contrib/tpu/__init__.py index dc9066855990f372c28dc481959117daa4c2da97..d62338680e9d8445b1a2405e92763f4f70d07d96 100644 --- a/tensorflow/contrib/tpu/__init__.py +++ b/tensorflow/contrib/tpu/__init__.py @@ -45,6 +45,8 @@ @@RunConfig @@InputPipelineConfig @@TPUConfig + +@@bfloat16_scope """ from __future__ import absolute_import diff --git a/tensorflow/contrib/tpu/profiler/pip_package/cloud_tpu_profiler/main.py b/tensorflow/contrib/tpu/profiler/pip_package/cloud_tpu_profiler/main.py index 7f1d25732e21b5dea4e605f6caa141ca9d3d02c6..7a5d01cca42351f6d4d8b41d43756560ce7874d3 100644 --- a/tensorflow/contrib/tpu/profiler/pip_package/cloud_tpu_profiler/main.py +++ b/tensorflow/contrib/tpu/profiler/pip_package/cloud_tpu_profiler/main.py @@ -17,12 +17,11 @@ from __future__ import absolute_import from __future__ import division from __future__ import print_function -from absl import flags - import os import subprocess import sys - +from absl import flags +from distutils.version import LooseVersion import tensorflow as tf # Cloud TPU Cluster Resolvers @@ -35,9 +34,9 @@ flags.DEFINE_string( None, help='GCE zone where the Cloud TPU is located in. If not specified, we ' 'will attempt to automatically detect the GCE project from metadata.') -flags.DEFINE_string('tpu', None, - 'Name of the Cloud TPU for Cluster Resolvers. You must ' - 'specify either this flag or --service_addr.') +flags.DEFINE_string( + 'tpu', None, 'Name of the Cloud TPU for Cluster Resolvers. You must ' + 'specify either this flag or --service_addr.') # Tool specific parameters flags.DEFINE_string( @@ -48,13 +47,13 @@ flags.DEFINE_string( ' e.g. 10.0.1.2, 10.0.1.3. You can specify this flag with --tpu or ' '--service_addr to profile a subset of tpu nodes. You can also use only' '--tpu and leave this flag unspecified to profile all the tpus.') -flags.DEFINE_string('logdir', None, - 'Path of TensorBoard log directory e.g. /tmp/tb_log, ' - 'gs://tb_bucket') +flags.DEFINE_string( + 'logdir', None, 'Path of TensorBoard log directory e.g. /tmp/tb_log, ' + 'gs://tb_bucket') flags.DEFINE_integer('duration_ms', 2000, 'Duration of tracing in ms.') -flags.DEFINE_integer('num_tracing_attempts', 3, - 'Automatically retry N times when no trace ' - 'event is collected.') +flags.DEFINE_integer( + 'num_tracing_attempts', 3, 'Automatically retry N times when no trace ' + 'event is collected.') flags.DEFINE_boolean('include_dataset_ops', True, 'Set to false to profile longer TPU ' 'device traces.') @@ -63,18 +62,24 @@ FLAGS = flags.FLAGS EXECUTABLE = 'data/capture_tpu_profile' JOB_NAME = 'worker' + def get_workers_list(cluster_resolver): cluster_spec = cluster_resolver.cluster_spec() task_indices = cluster_spec.task_indices(JOB_NAME) - workers_list = [cluster_spec.task_address(JOB_NAME, i).split(':')[0] - for i in task_indices] + workers_list = [ + cluster_spec.task_address(JOB_NAME, i).split(':')[0] for i in task_indices + ] return ','.join(workers_list) + def run_main(): tf.app.run(main) + def main(unused_argv=None): tf.logging.set_verbosity(tf.logging.INFO) + tf_version = tf.__version__ + print('TensorFlow version %s detected' % tf_version) if FLAGS.service_addr is None and FLAGS.tpu is None: sys.exit('You must specify either --service_addr or --tpu.') @@ -88,17 +93,19 @@ def main(unused_argv=None): else: tpu_cluster_resolver = ( tf.contrib.cluster_resolver.TPUClusterResolver( - [FLAGS.tpu], - zone=FLAGS.tpu_zone, - project=FLAGS.gcp_project)) + [FLAGS.tpu], zone=FLAGS.tpu_zone, project=FLAGS.gcp_project)) service_addr = tpu_cluster_resolver.get_master() service_addr = service_addr.replace('grpc://', '').replace(':8470', ':8466') - workers_list = "" - if FLAGS.workers_list is not None: - workers_list = FLAGS.workers_list - elif tpu_cluster_resolver is not None: - workers_list = get_workers_list(tpu_cluster_resolver) + workers_list = '' + if LooseVersion(tf_version) < LooseVersion('1.9'): + tf.logging.warn('Attempt to profile with legacy support under TensorFlow ' + 'version %s' % tf_version) + else: + if FLAGS.workers_list is not None: + workers_list = FLAGS.workers_list + elif tpu_cluster_resolver is not None: + workers_list = get_workers_list(tpu_cluster_resolver) if not FLAGS.logdir: sys.exit('logdir must be provided.') diff --git a/tensorflow/contrib/tpu/profiler/pip_package/setup.py b/tensorflow/contrib/tpu/profiler/pip_package/setup.py index f97a972f01a3ba5582df3675439aa962886f796e..19f088f8b862ce7b114490151f2b6a8c260b8580 100644 --- a/tensorflow/contrib/tpu/profiler/pip_package/setup.py +++ b/tensorflow/contrib/tpu/profiler/pip_package/setup.py @@ -20,7 +20,7 @@ from __future__ import print_function from setuptools import setup -_VERSION = '1.7.0' +_VERSION = '1.9.0' CONSOLE_SCRIPTS = [ 'capture_tpu_profile=cloud_tpu_profiler.main:run_main', diff --git a/tensorflow/contrib/tpu/profiler/version.h b/tensorflow/contrib/tpu/profiler/version.h index bd9ba6697edd9ef14dd3af0d2c9b77df9ec6917a..1bf49966d12db83f1e6904f8c00453bba278847c 100644 --- a/tensorflow/contrib/tpu/profiler/version.h +++ b/tensorflow/contrib/tpu/profiler/version.h @@ -16,6 +16,6 @@ limitations under the License. #ifndef TENSORFLOW_CONTRIB_TPU_PROFILER_VERSION_H_ #define TENSORFLOW_CONTRIB_TPU_PROFILER_VERSION_H_ -#define TPU_PROFILER_VERSION "1.7.0" +#define TPU_PROFILER_VERSION "1.9.0" #endif // TENSORFLOW_CONTRIB_TPU_PROFILER_VERSION_H_ diff --git a/tensorflow/contrib/tpu/proto/BUILD b/tensorflow/contrib/tpu/proto/BUILD index 7ecb36852c53bb74d70ed0f8c70ca1ce860a037a..26016f47dfb36990fd73267c70619878ac3450e5 100644 --- a/tensorflow/contrib/tpu/proto/BUILD +++ b/tensorflow/contrib/tpu/proto/BUILD @@ -2,7 +2,12 @@ licenses(["notice"]) # Apache 2.0 exports_files(["LICENSE"]) -load("//tensorflow/core:platform/default/build_config.bzl", "tf_proto_library") +load( + "//tensorflow/core:platform/default/build_config.bzl", + "tf_additional_all_protos", + "tf_proto_library", + "tf_proto_library_py", +) tf_proto_library( name = "tpu_embedding_config_proto", @@ -22,12 +27,14 @@ tf_proto_library( visibility = ["//visibility:public"], ) -tf_proto_library( +tf_proto_library_py( name = "compilation_result_proto", srcs = [ "compilation_result.proto", ], - cc_api_version = 2, - protodeps = ["//tensorflow/core:protos_all"], + protodeps = tf_additional_all_protos() + [ + "//tensorflow/compiler/xla:xla_data_proto", + "//tensorflow/compiler/xla/service:hlo_proto", + ], visibility = ["//visibility:public"], ) diff --git a/tensorflow/contrib/tpu/proto/compilation_result.proto b/tensorflow/contrib/tpu/proto/compilation_result.proto index cf52897de3d0fefa55e68a6b889ae9af7b45864a..88585a5bd10fc28aa34bb0de72de970e21b2adb2 100644 --- a/tensorflow/contrib/tpu/proto/compilation_result.proto +++ b/tensorflow/contrib/tpu/proto/compilation_result.proto @@ -3,6 +3,7 @@ syntax = "proto3"; option cc_enable_arenas = true; package tensorflow.tpu; +import "tensorflow/compiler/xla/service/hlo.proto"; import "tensorflow/core/lib/core/error_codes.proto"; // Describes the result of a TPU compilation. @@ -10,4 +11,7 @@ message CompilationResultProto { // The error message, if any, returned during compilation. error.Code status_code = 1; string status_error_message = 2; + + // HLO proto. + repeated xla.HloProto hlo_protos = 3; } diff --git a/tensorflow/contrib/tpu/python/tpu/keras_support.py b/tensorflow/contrib/tpu/python/tpu/keras_support.py index 293e162059205cad572f0ca78217985b6932a239..722e31abb27deaf49d45b526dd894124c9d93b25 100644 --- a/tensorflow/contrib/tpu/python/tpu/keras_support.py +++ b/tensorflow/contrib/tpu/python/tpu/keras_support.py @@ -19,15 +19,16 @@ To use, wrap your model with the `keras_support.tpu_model` function. Example usage: ``` -# Must activate before building TPU models -keras_support.setup_tpu_session(master_address) - image = tf.keras.layers.Input(shape=(28, 28, 3), name='image') c1 = tf.keras.layers.Conv2D(filters=16, kernel_size=(3, 3))( image) flattened = tf.keras.layers.Flatten()(c1) logits = tf.keras.layers.Dense(10, activation='softmax')(flattened) model = tf.keras.Model(inputs=[image], outputs=[logits]) -model = keras_support.tpu_model(model) + +strategy = keras_support.TPUDistributionStrategy(num_cores_per_host=8) +model = keras_support.tpu_model(model, + strategy=strategy, + tpu_name_or_address=tpu_name) # Only TF optimizers are currently supported. model.compile(optimizer=tf.train.AdamOptimizer(), ...) @@ -35,9 +36,6 @@ model.compile(optimizer=tf.train.AdamOptimizer(), ...) # `images` and `labels` should be Numpy arrays. Support for tensor input # (e.g. datasets) is planned. model.fit(images, labels) - -# Invoke before shutting down -keras_support.shutdown_tpu_session() ``` """ @@ -48,29 +46,41 @@ from __future__ import division from __future__ import print_function import collections +import contextlib import re +import sys import time +import numpy as np + from tensorflow.contrib.cluster_resolver.python.training import tpu_cluster_resolver +from tensorflow.contrib.distribute.python import tpu_strategy from tensorflow.contrib.framework.python.framework import experimental from tensorflow.contrib.tpu.proto import compilation_result_pb2 as tpu_compilation_result from tensorflow.contrib.tpu.python.ops import tpu_ops from tensorflow.contrib.tpu.python.tpu import tpu +from tensorflow.contrib.tpu.python.tpu import tpu_function from tensorflow.contrib.tpu.python.tpu import tpu_optimizer from tensorflow.core.protobuf import config_pb2 from tensorflow.python.client import session as tf_session from tensorflow.python.estimator import model_fn as model_fn_lib +from tensorflow.python.framework import dtypes from tensorflow.python.framework import ops from tensorflow.python.framework import tensor_spec from tensorflow.python.keras import backend as K -from tensorflow.python.keras import layers from tensorflow.python.keras import models from tensorflow.python.keras import optimizers as keras_optimizers +from tensorflow.python.keras.engine import base_layer from tensorflow.python.keras.layers import embeddings from tensorflow.python.ops import array_ops +from tensorflow.python.ops import gen_linalg_ops from tensorflow.python.ops import math_ops +from tensorflow.python.ops import random_ops +from tensorflow.python.ops import variable_scope from tensorflow.python.platform import tf_logging as logging +TPUDistributionStrategy = tpu_strategy.TPUStrategy # pylint: disable=invalid-name + class TPUEmbedding(embeddings.Embedding): """TPU compatible embedding layer. @@ -93,11 +103,49 @@ class TPUEmbedding(embeddings.Embedding): return math_ops.tensordot(inputs, self.embeddings, 1) +class KerasCrossShardOptimizer(keras_optimizers.Optimizer): + """An optimizer that averages gradients across TPU shards.""" + + def __init__(self, opt, name='KerasCrossShardOptimizer'): + """Construct a new cross-shard optimizer. + + Args: + opt: An existing `Optimizer` to encapsulate. + name: Optional name prefix for the operations created when applying + gradients. Defaults to "KerasCrossShardOptimizer". + + Raises: + ValueError: If reduction is not a valid cross-shard reduction. + """ + super(KerasCrossShardOptimizer, self).__init__() + self._name = name + self._opt = opt + + def get_updates(self, loss, params): + logging.info('Get updates: %s', loss) + self._opt.get_gradients = self.get_gradients + return self._opt.get_updates(loss, params) + + def get_gradients(self, loss, params): + num_shards = tpu_function.get_tpu_context().number_of_shards + grads = super(KerasCrossShardOptimizer, self).get_gradients(loss, params) + return [tpu_ops.cross_replica_sum(grad) / num_shards for grad in grads] + + def set_weights(self, weights): + self._opt.set_weights() + + def get_weights(self): + return self._opt.get_weights() + + @property + def lr(self): + return self._opt.lr + + class TPUModelOp( - collections.namedtuple( - 'TPUModelOp', - ['compile_op', 'execute_op', 'infeed_tensors', 'infeed_op', - 'outfeed_op'])): + collections.namedtuple('TPUModelOp', [ + 'compile_op', 'execute_op', 'infeed_tensors', 'infeed_op', 'outfeed_op' + ])): pass @@ -106,13 +154,119 @@ def _valid_name(tensor_name): return re.sub('[^a-zA-Z0-9_-]+', '', tensor_name) -def _replicated_optimizer(opt, num_replicas): +def _replicated_optimizer(opt): """Wrap the optimizer `opt` with CrossShardOptimizer if applicable.""" - if num_replicas == 1: + if tpu_function.get_tpu_context().number_of_shards == 1: return opt - return keras_optimizers.TFOptimizer( - optimizer=tpu_optimizer.CrossShardOptimizer(opt.optimizer) - ) + + if isinstance(opt, keras_optimizers.TFOptimizer): + return tpu_optimizer.CrossShardOptimizer(opt.optimizer) + else: + return KerasCrossShardOptimizer(opt) + + +class TPURewriteContext(object): + """Prepare the environment for a Keras model during `tpu.rewrite`. + + This overrides the default placeholder behaviour to instead refer to a preset + input mapping. Placeholders are unsupported in TPU compiled code, and must + be replaced with explicit inputs or values from the infeed queue. + + Instead of explicitly threading inputs all the way through the Keras codebase, + we override the behavior of the placeholder while compiling and inject the + Tensors from the infeed in place of the placeholder. + + Similarly, as we compile a new sub-graph for each unique shape and execution + mode, we need to override the behavior of an embedded `name_scope` call in + the base Keras layer code. This allows us to re-use the same weights across + many compiles and share a single session/graph. + """ + + def __init__(self, input_map): + self._input_map = input_map + self._default_placeholder = None + self._default_name_scope = None + + def __enter__(self): + + def _placeholder(dtype, shape=None, name=None): # pylint: disable=unused-argument + logging.info('Remapping placeholder for %s', name) + if name in self._input_map: + return self._input_map[name] + else: + logging.info('Default: %s', name) + return self._default_placeholder(dtype, shape, name) + + def _name_scope(name, default_name=None, values=None): + caller_frame = sys._getframe().f_back + caller_obj = caller_frame.f_locals.get('self') + if (caller_obj is not None and + isinstance(caller_obj, base_layer.Layer) and name is not None): + logging.info('Intercepted name_scope: %s', caller_obj) + return variable_scope.variable_scope( + name, default_name, values, reuse=variable_scope.AUTO_REUSE) + + return self._default_name_scope(name, default_name, values) + + self._default_placeholder = array_ops.placeholder + self._default_name_scope = ops.name_scope + self._default_make_variable = base_layer.make_variable + self._default_random_normal = random_ops.random_normal + self._default_qr = gen_linalg_ops.qr + + array_ops.placeholder = _placeholder + + # Replace random_ops.random_normal with a dummy function because + # `random_normal` isn't yet implemented on the TPU. Because these + # initialized values are overwritten by the CPU values, this is okay. + def random_normal(shape, + mean=0.0, + stddev=1.0, + dtype=dtypes.float32, + seed=None, + name=None): + del mean + del stddev + del seed + return array_ops.zeros(shape, dtype=dtype, name=name) + + random_ops.random_normal = random_normal + + # Replace gen_linalg_ops.qr because QR decomposition is not yet implemented. + # TODO(saeta): Remove qr override once we confirm the qr implementation is + # ok. + # pylint: disable=redefined-builtin + def qr(input, full_matrices=False, name=None): + """Dummy implementation of qr decomposition.""" + del full_matrices # TODO(saeta): Properly handle the full matrix case. + input_shape = input.shape + if len(input_shape) < 2: + raise ValueError('Invalid shape passed to qr: %s' % input_shape) + p = min(input_shape[-1], input_shape[-2]) + if len(input_shape) == 2: + q = array_ops.zeros((p, p), name=name) + r = array_ops.zeros(input_shape, name=name) + return (r, q) + elif len(input_shape) == 3: + n = input_shape[0] + q = array_ops.zeros((n, p, p), name=name) + r = array_ops.zeros(input_shape, name=name) + return (r, q) + else: + raise ValueError('Invalid shape passed to qr: %s' % input_shape) + gen_linalg_ops.qr = qr + + ops.name_scope = _name_scope + base_layer.make_variable = variable_scope.get_variable + logging.info('Overriding default placeholder.') + return + + def __exit__(self, exc_type, exc_val, exc_tb): + array_ops.placeholder = self._default_placeholder + ops.name_scope = self._default_name_scope + base_layer.make_variable = self._default_make_variable + random_ops.random_normal = self._default_random_normal + gen_linalg_ops.qr = self._default_qr class TPUFunction(object): @@ -127,19 +281,24 @@ class TPUFunction(object): instead of being injected as `feed_dict` items or fetches. """ - def __init__(self, model, execution_mode, num_replicas=1): + def __init__(self, model, execution_mode, strategy): self.model = model self.execution_mode = execution_mode + self._strategy = strategy self._compilation_cache = {} - self.num_replicas = num_replicas + self._cloned_model = None + + # Copy optimizer configuration. This is done prior to `_specialize_model` + # as the configuration may require evaluating variables in the CPU session. + self._optimizer_config = None + if not isinstance(self.model.optimizer, keras_optimizers.TFOptimizer): + self._optimizer_config = self.model.optimizer.get_config() def _specialize_model(self, input_specs): """Specialize `self.model` (a Keras model) for the given input shapes.""" # Re-create our input and output layers inside our subgraph. They will be # attached to the true computation when we clone our model in `tpu_fn`. - K.set_learning_phase( - self.execution_mode == model_fn_lib.ModeKeys.TRAIN - ) + K.set_learning_phase(self.execution_mode == model_fn_lib.ModeKeys.TRAIN) # functools.partial and callable objects are not supported by tpu.rewrite def _model_fn(): @@ -165,23 +324,34 @@ class TPUFunction(object): infeed_tensors)) tpu_targets = [] - tpu_inputs = [] + tpu_input_map = {} # Sort infeed outputs into inputs and labels for calling our Keras model. for tensor, layer in zip(infeed_tensors, infeed_layers): if layer in self.model._input_layers: - tpu_inputs.append(layers.Input(name=layer.name, tensor=tensor)) + tpu_input_map[layer.name] = tensor if layer in self.model._output_layers: tpu_targets.append(tensor) - # Call our model with our infeed inputs (re-using the weights). - model_outputs = self.model(tpu_inputs) - child_model = models.Model(inputs=tpu_inputs, outputs=model_outputs) + # Clone our CPU model, running within the TPU device context. + with TPURewriteContext(tpu_input_map): + # TODO(power): Replicate variables. + with ops.device('/device:TPU:0'): + self._cloned_model = models.clone_model(self.model) + + # Create a copy of the optimizer for this graph. + if isinstance(self.model.optimizer, keras_optimizers.TFOptimizer): + cloned_optimizer = keras_optimizers.TFOptimizer( + self.model.optimizer.optimizer) + else: + logging.info('Cloning %s %s', self.model.optimizer.__class__.__name__, + self._optimizer_config) + cloned_optimizer = self.model.optimizer.__class__.from_config( + self._optimizer_config) if is_training or is_test: - child_model.compile( - optimizer=_replicated_optimizer(self.model.optimizer, - self.num_replicas), + self._cloned_model.compile( + optimizer=_replicated_optimizer(cloned_optimizer), loss=self.model.loss, loss_weights=self.model.loss_weights, metrics=self.model.metrics, @@ -191,37 +361,37 @@ class TPUFunction(object): # Compute our outfeed depending on the execution mode if is_training: - child_model._make_train_function() + self._cloned_model._make_train_function() self._outfeed_spec = [ tensor_spec.TensorSpec(tensor.shape, tensor.dtype, tensor.name) - for tensor in child_model.train_function.outputs + for tensor in self._cloned_model.train_function.outputs ] return [ - child_model.train_function.updates_op, + self._cloned_model.train_function.updates_op, tpu_ops.outfeed_enqueue_tuple( - child_model.train_function.outputs, + self._cloned_model.train_function.outputs, name='outfeed-enqueue-train') ] elif is_test: - child_model._make_test_function() + self._cloned_model._make_test_function() self._outfeed_spec = [ tensor_spec.TensorSpec(tensor.shape, tensor.dtype, tensor.name) - for tensor in child_model.test_function.outputs + for tensor in self._cloned_model.test_function.outputs ] return [ tpu_ops.outfeed_enqueue_tuple( - child_model.test_function.outputs, + self._cloned_model.test_function.outputs, name='outfeed-enqueue-test') ] elif is_predict: - child_model._make_predict_function() + self._cloned_model._make_predict_function() self._outfeed_spec = [ tensor_spec.TensorSpec(tensor.shape, tensor.dtype, tensor.name) - for tensor in child_model.predict_function.outputs + for tensor in self._cloned_model.predict_function.outputs ] return [ tpu_ops.outfeed_enqueue_tuple( - child_model.predict_function.outputs, + self._cloned_model.predict_function.outputs, name='outfeed-enqueue-predict', ) ] @@ -236,7 +406,7 @@ class TPUFunction(object): # `execute op` replicates `_model_fn` `num_replicas` times, with each shard # running on a different logical core. compile_op, execute_op = tpu.split_compile_and_replicate( - _model_fn, inputs=[[]] * self.num_replicas) + _model_fn, inputs=[[]] * self._strategy.num_towers) # Generate CPU side operations to enqueue features/labels and dequeue # outputs from the model call. @@ -244,7 +414,7 @@ class TPUFunction(object): outfeed_op = [] shard_infeed_tensors = [] - for shard_id in range(self.num_replicas): + for shard_id in range(self._strategy.num_towers): with ops.device('/device:TPU:%d' % shard_id): infeed_tensors = [] for spec in input_specs: @@ -255,32 +425,35 @@ class TPUFunction(object): name='infeed-enqueue-%s-%d' % (spec.name, shard_id))) shard_infeed_tensors.append(infeed_tensors) - infeed_op.append(tpu_ops.infeed_enqueue_tuple( - infeed_tensors, [spec.shape for spec in input_specs], - name='infeed-enqueue-%s-%d' % (self.execution_mode, shard_id))) + infeed_op.append( + tpu_ops.infeed_enqueue_tuple( + infeed_tensors, [spec.shape for spec in input_specs], + name='infeed-enqueue-%s-%d' % (self.execution_mode, shard_id))) - outfeed_op.extend(tpu_ops.outfeed_dequeue_tuple( - dtypes=[spec.dtype for spec in self._outfeed_spec], - shapes=[spec.shape for spec in self._outfeed_spec], - name='outfeed-dequeue-%s-%d' % (self.execution_mode, shard_id))) + outfeed_op.extend( + tpu_ops.outfeed_dequeue_tuple( + dtypes=[spec.dtype for spec in self._outfeed_spec], + shapes=[spec.shape for spec in self._outfeed_spec], + name='outfeed-dequeue-%s-%d' % (self.execution_mode, shard_id))) return TPUModelOp( - compile_op, execute_op, infeed_tensors=shard_infeed_tensors, - infeed_op=infeed_op, outfeed_op=outfeed_op) + compile_op, + execute_op, + infeed_tensors=shard_infeed_tensors, + infeed_op=infeed_op, + outfeed_op=outfeed_op) def _test_model_compiles(self, tpu_model_ops): """Verifies that the given TPUModelOp can be compiled via XLA.""" - session = K.get_session() - logging.info('Started compiling') start_time = time.clock() - result = session.run(tpu_model_ops.compile_op) + result = K.get_session().run(tpu_model_ops.compile_op) proto = tpu_compilation_result.CompilationResultProto() proto.ParseFromString(result) if proto.status_error_message: - raise RuntimeError( - 'Compilation failed: {}'.format(proto.status_error_message)) + raise RuntimeError('Compilation failed: {}'.format( + proto.status_error_message)) end_time = time.clock() logging.info('Finished compiling. Time elapsed: %s secs', @@ -297,17 +470,19 @@ class TPUFunction(object): Returns: List of lists containing the input to feed to each TPU shard. """ - if self.num_replicas == 1: + if self._strategy.num_towers == 1: return [inputs] batch_size = inputs[0].shape[0] - assert batch_size % self.num_replicas == 0, ( - 'batch_size must be divisible by num_replicas') - shard_size = batch_size // self.num_replicas + assert batch_size % self._strategy.num_towers == 0, ( + 'batch_size must be divisible by strategy.num_towers (%s vs %s)' % + (batch_size, self._strategy.num_towers)) + shard_size = batch_size // self._strategy.num_towers input_list = [] - for index in range(self.num_replicas): - shard_inputs = [x[index * shard_size:(index + 1) * shard_size] - for x in inputs] + for index in range(self._strategy.num_towers): + shard_inputs = [ + x[index * shard_size:(index + 1) * shard_size] for x in inputs + ] input_list.append(shard_inputs) return input_list @@ -344,12 +519,15 @@ class TPUFunction(object): shape_key = tuple([tuple(spec.shape.as_list()) for spec in input_specs]) if shape_key not in self._compilation_cache: - logging.info('New input shapes; (re-)compiling: mode=%s, %s', - self.execution_mode, input_specs) - new_tpu_model_ops = self._specialize_model(input_specs) - self._compilation_cache[shape_key] = new_tpu_model_ops - self._test_model_compiles(new_tpu_model_ops) - + with self.model.tpu_session(): + logging.info('New input shapes; (re-)compiling: mode=%s, %s', + self.execution_mode, input_specs) + new_tpu_model_ops = self._specialize_model(input_specs) + self._compilation_cache[shape_key] = new_tpu_model_ops + self._test_model_compiles(new_tpu_model_ops) + + # Initialize our TPU weights on the first compile. + self.model._initialize_weights(self._cloned_model) tpu_model_ops = self._compilation_cache[shape_key] infeed_dict = {} @@ -358,75 +536,84 @@ class TPUFunction(object): for tensor, value in zip(infeed_tensors, inputs): infeed_dict[tensor] = value - session = K.get_session() - _, _, outfeed_outputs = session.run([ - tpu_model_ops.infeed_op, tpu_model_ops.execute_op, - tpu_model_ops.outfeed_op - ], infeed_dict) + with self.model.tpu_session() as session: + _, _, outfeed_outputs = session.run([ + tpu_model_ops.infeed_op, tpu_model_ops.execute_op, + tpu_model_ops.outfeed_op + ], infeed_dict) # TODO(xiejw): Decide how to reduce outputs, or just discard all but first. - return outfeed_outputs[:len(outfeed_outputs) // self.num_replicas] - + if self.execution_mode == model_fn_lib.ModeKeys.PREDICT: + outputs = [[]] * len(self._outfeed_spec) + outputs_per_replica = len(self._outfeed_spec) -@experimental -def setup_tpu_session(tpu_name_or_address): - """Initializes and returns a Keras/TF session connected the TPU `master`. + for i in range(self._strategy.num_towers): + output_group = outfeed_outputs[i * outputs_per_replica:(i + 1) * + outputs_per_replica] + for j in range(outputs_per_replica): + outputs[j].append(output_group[j]) - Args: - tpu_name_or_address: A string that is either the name of the Cloud TPU, - the grpc address of the Cloud TPU, or (Googlers only) the BNS name of the - Cloud TPU. If tpu_name_or_address is None, the TPUClusterResolver will - examine the environment to determine a potential Cloud TPU to use. - - Returns: - A `tf.Session`. - """ - cluster_resolver = tpu_cluster_resolver.TPUClusterResolver( - tpu_name_or_address) - cluster_spec = cluster_resolver.cluster_spec() - session = tf_session.Session( - target=cluster_resolver.master(), - config=config_pb2.ConfigProto( - isolate_session_state=True)) - if cluster_spec: - session.cluster_def.CopyFrom(cluster_spec.as_cluster_def()) - K.set_session(session) - K.get_session().run(tpu.initialize_system()) - return session - - -@experimental -def shutdown_tpu_session(session=None): - """Shutdown the TPU attached to session. - - This should be called to cleanly shut down the TPU system before the client - exits. - - Args: - session: Session to shutdown, or None to use the default session. - - Returns: - - """ - if session is None: - session = K.get_session() - - session.run(tpu.shutdown_system()) + return [np.concatenate(group) for group in outputs] + else: + return outfeed_outputs[:len(outfeed_outputs) // self._strategy.num_towers] class KerasTPUModel(models.Model): """TPU compatible Keras model wrapper.""" - def __init__(self, inputs, outputs, name, replicas=1): + def __init__(self, cpu_model, tpu_name_or_address, strategy): super(models.Model, self).__init__( # pylint: disable=bad-super-call - inputs=inputs, - outputs=outputs, - name=name, + inputs=cpu_model.inputs, + outputs=cpu_model.outputs, + name=cpu_model.name, ) + self.predict_function = None self.test_function = None self.train_function = None - self.replicas = replicas + self._strategy = strategy + + self._tpu_name_or_address = tpu_name_or_address + self._cpu_model = cpu_model + self._tpu_model = None + self._tpu_weights_initialized = False + self._graph = ops.Graph() + + self._cluster_resolver = tpu_cluster_resolver.TPUClusterResolver( + tpu_name_or_address) + master = self._cluster_resolver.master() + cluster_spec = self._cluster_resolver.cluster_spec() + self._session = tf_session.Session( + graph=self._graph, + target=master, + config=config_pb2.ConfigProto(isolate_session_state=True)) + + # TODO(saeta): Confirm the lines below work in ClusterSpec propagation env. + if cluster_spec: + self._session.cluster_def.CopyFrom(cluster_spec.as_cluster_def()) + + with self._graph.as_default(): + self._session.run(tpu.initialize_system()) + + # If the input CPU model has already been compiled, compile our TPU model + # immediately. + if self._cpu_model.optimizer: + self.compile( + self._cpu_model.optimizer, + self._cpu_model.loss, + self._cpu_model.metrics, + self._cpu_model.loss_weights, + self._cpu_model.sample_weight_mode, + self._cpu_model.weighted_metrics, + self._cpu_model.target_tensors, + ) + + def get_config(self): + return { + 'cpu_model': self._cpu_model, + 'tpu_name_or_address': self._tpu_name_or_address, + 'strategy': self._strategy, + } def compile(self, optimizer, @@ -448,44 +635,97 @@ class KerasTPUModel(models.Model): sample_weight_mode, weighted_metrics, target_tensors, **kwargs) - # Keras optimizers are not compatible with TPU rewrite - if not isinstance(self.optimizer, keras_optimizers.TFOptimizer): - raise ValueError( - 'Optimizer must be a TFOptimizer, got: %s' % self.optimizer) + if not self._cpu_model.optimizer: + self._cpu_model.compile(optimizer, loss, metrics, loss_weights, + sample_weight_mode, weighted_metrics, + target_tensors, **kwargs) def _make_train_function(self): if not self.train_function: - self.train_function = TPUFunction(self, model_fn_lib.ModeKeys.TRAIN, - num_replicas=self.replicas) + self.train_function = TPUFunction( + self, model_fn_lib.ModeKeys.TRAIN, strategy=self._strategy) return self.train_function def _make_test_function(self): if not self.test_function: - self.test_function = TPUFunction(self, model_fn_lib.ModeKeys.EVAL) + self.test_function = TPUFunction( + self, model_fn_lib.ModeKeys.EVAL, strategy=self._strategy) return self.test_function def _make_predict_function(self): if not self.predict_function: - self.predict_function = TPUFunction(self, model_fn_lib.ModeKeys.PREDICT) + self.predict_function = TPUFunction( + self, model_fn_lib.ModeKeys.PREDICT, strategy=self._strategy) return self.predict_function - def cpu_model(self): - cpu_model = models.Model( - inputs=self.inputs, - outputs=self.outputs, - name=self.name, - ) + def _initialize_weights(self, cloned_model): + """Initialize TPU weights. - if self.optimizer: - cpu_model.compile( - optimizer=self.optimizer, - loss=self.loss, - metrics=self.metrics, - loss_weights=self.loss_weights, - ) + This is called on the first compile of the TPU model (first call to + fit/predict/evaluate). - return cpu_model + Args: + cloned_model: `keras.Model`, TPU model to initialize. + """ + if self._tpu_weights_initialized: + return + + self._tpu_model = cloned_model + self._tpu_weights_initialized = True + + weights = self._cpu_model.get_weights() + with self.tpu_session(): + logging.info('Setting weights on TPU model.') + cloned_model.set_weights(weights) + + def sync_to_cpu(self): + """Copy weights from the CPU, returning a synchronized CPU model.""" + if self._tpu_weights_initialized: + with self.tpu_session(): + logging.info('Copying TPU weights to the CPU') + tpu_weights = self._tpu_model.get_weights() + + self._cpu_model.set_weights(tpu_weights) + + return self._cpu_model + + def get_weights(self): + return self.sync_to_cpu().get_weights() + + def save_weights(self, *args, **kw): + return self.sync_to_cpu().save_weights(*args, **kw) + + def save(self, *args, **kw): + return self.sync_to_cpu().save(*args, **kw) + + def set_weights(self, weights): + # We may not have a TPU model available if we haven't run fit/predict, so + # we can't directly set the TPU weights here. + # Instead, reset CPU model weights and force TPU re-initialization at the + # next call. + self._cpu_model.set_weights(weights) + self._tpu_weights_initialized = False + + @contextlib.contextmanager + def tpu_session(self): + """Yields a TPU session and sets it as the default Keras session.""" + with self._graph.as_default(): + default_session = K.get_session() + # N.B. We have to call `K.set_session()` AND set our session as the + # TF default. `K.get_session()` surprisingly does not return the value + # supplied by K.set_session otherwise. + K.set_session(self._session) + with self._session.as_default(): + yield self._session + K.set_session(default_session) + + def shutdown(self): + # TODO(b/111364423): Actually shut down the system. + logging.info('Skipping shutting down TPU system.') + # with self.tpu_session() as session: + # session.run(tpu.shutdown_system()) + self._session.close() def _validate_shapes(model): @@ -522,8 +762,8 @@ Output shape: %(output_shape)s @experimental -def tpu_model(model, replicas=None): - """Runs a model on TPU(s). +def tpu_model(model, tpu_name_or_address=None, strategy=None): + """Copy `model` along with weights to the TPU. Returns a TPU model. Usage: ``` @@ -531,44 +771,39 @@ def tpu_model(model, replicas=None): b = Dense(32)(a) model = Model(inputs=a, outputs=b) - model = keras_support.tpu_model(model) - model.compile( - optimizer=tf.train.GradientDescentOptimizer(learning_rate=1.0), - ...) - ``` - - If `replicas` is set, replicates the model computation on all TPU cores. The - model computation is replicated `num_replicas` times; each shard will run on a - different TPU core. - - Limitation: Currently, replication is only supported for training. - - Usage: - ``` - a = Input(shape=(32,)) - b = Dense(32)(a) - model = Model(inputs=a, outputs=b) - - model = keras_support.tpu_model(model, replicas=2) + # If `num_cores_per_host` is greater than one, batch parallelism will be used + # to run on multiple TPU cores. + strategy = keras_support.TPUDistributionStrategy(num_cores_per_host=8) + model = keras_support.tpu_model(model, strategy) model.compile( optimizer=tf.train.GradientDescentOptimizer(learning_rate=1.0), ...) + model.shutdown() ``` Args: model: A `KerasTPUModel`. - replicas: (Optional) Int, number of TPU cores which to create model - replicas. If `None`, the model runs on single core only, i.e., no - replication. + tpu_name_or_address: A string that is either the name of the Cloud TPU, + the grpc address of the Cloud TPU, or (Googlers only) the BNS name of the + Cloud TPU. If tpu_name_or_address is None, the TPUClusterResolver will + examine the environment to determine a potential Cloud TPU to use. + strategy: `TPUDistributionStrategy`. The strategy to use for replicating + model across multiple TPU cores. Returns: A new `KerasTPUModel` instance. """ + # Force initialization of the CPU model. + model.get_weights() + model.reset_states() + _validate_shapes(model) # TODO(xiejw): Validate TPU model. TPUModel only? # TODO(xiejw): Validate replicas. Full or 1. Shall we allow subset? # TODO(xiejw): Adds reduction option. - replicas = 1 if replicas is None else replicas + if strategy is None: + strategy = TPUDistributionStrategy(num_cores_per_host=1) return KerasTPUModel( - inputs=model.inputs, outputs=model.outputs, name=model.name, - replicas=replicas) + cpu_model=model, + tpu_name_or_address=tpu_name_or_address, + strategy=strategy) diff --git a/tensorflow/contrib/tpu/python/tpu/topology.py b/tensorflow/contrib/tpu/python/tpu/topology.py index cda9a63f204ed686b527c95dd5b4fd7786ac60cf..1fb26e701a392d5ef3bc40d5772d4541fa38f773 100644 --- a/tensorflow/contrib/tpu/python/tpu/topology.py +++ b/tensorflow/contrib/tpu/python/tpu/topology.py @@ -55,8 +55,9 @@ class Topology(object): rank 3 numpy int32 array that describes a valid coordinate mapping. """ + self._serialized = serialized + if serialized: - self._serialized = serialized self._parse_topology(serialized) else: self._mesh_shape = np.asarray(mesh_shape, dtype=np.int32) @@ -131,7 +132,7 @@ class Topology(object): proto.mesh_shape[:] = list(self._mesh_shape) proto.num_tasks = self._device_coordinates.shape[0] proto.num_tpu_devices_per_task = self._device_coordinates.shape[1] - proto.device_coordinates = list(self._device_coordinates.flatten()) + proto.device_coordinates.extend(list(self._device_coordinates.flatten())) self._serialized = proto.SerializeToString() return self._serialized diff --git a/tensorflow/contrib/tpu/python/tpu/topology_test.py b/tensorflow/contrib/tpu/python/tpu/topology_test.py new file mode 100644 index 0000000000000000000000000000000000000000..e67fdb263aa48a37f65c3623365ebcf8f98bebd4 --- /dev/null +++ b/tensorflow/contrib/tpu/python/tpu/topology_test.py @@ -0,0 +1,46 @@ +# Copyright 2018 The TensorFlow Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================= + +"""Tests for topology.py.""" + +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + +from tensorflow.contrib.tpu.python.tpu import topology + +from tensorflow.python.platform import test + + +class TopologyTest(test.TestCase): + + def testSerialization(self): + """Test if the class is able to generate serialzied string.""" + original_topology = topology.Topology( + mesh_shape=[1, 1, 2], + device_coordinates=[[[0, 0, 0], [0, 0, 1]]], + ) + serialized_str = original_topology.serialized() + new_topology = topology.Topology(serialized=serialized_str) + + # Make sure the topology recovered from serialized str is same as the + # original topology. + self.assertAllEqual( + original_topology.mesh_shape, new_topology.mesh_shape) + self.assertAllEqual( + original_topology.device_coordinates, new_topology.device_coordinates) + +if __name__ == "__main__": + test.main() diff --git a/tensorflow/contrib/tpu/python/tpu/tpu.py b/tensorflow/contrib/tpu/python/tpu/tpu.py index dc473c5846aafc5a92756dfb8259f7f8dc14b98d..6a64893d9abcd64360554ab00502cdf360b820b6 100644 --- a/tensorflow/contrib/tpu/python/tpu/tpu.py +++ b/tensorflow/contrib/tpu/python/tpu/tpu.py @@ -227,19 +227,26 @@ class TPUReplicateContext(control_flow_ops.XLAControlFlowContext): class FakeOp(object): """A helper class to determine the current device. - Supports only the device set/get methods needed to run the + Supports only the type and device set/get methods needed to run the graph's _apply_device_function method. """ def __init__(self): self._device = "" + @property + def type(self): + return "FakeOp" + @property def device(self): return self._device def _set_device(self, device): - self._device = device.to_string() + if isinstance(device, pydev.DeviceSpec): + self._device = device.to_string() + else: + self._device = device if self._outside_compilation_cluster: raise NotImplementedError("Cannot nest outside_compilation clusters") diff --git a/tensorflow/contrib/tpu/python/tpu/tpu_config.py b/tensorflow/contrib/tpu/python/tpu/tpu_config.py index 6d7331e3c79ade9c12c15de79f550cf3973c4e6c..9e010922dcf565e78944bd77d49f7d3fa07f2cc4 100644 --- a/tensorflow/contrib/tpu/python/tpu/tpu_config.py +++ b/tensorflow/contrib/tpu/python/tpu/tpu_config.py @@ -23,8 +23,6 @@ import collections import json import os -import numpy as np - from tensorflow.contrib.tpu.python.tpu import util as util_lib from tensorflow.core.protobuf import config_pb2 from tensorflow.python.estimator import run_config as run_config_lib @@ -43,6 +41,7 @@ class InputPipelineConfig(object): PER_SHARD_V1 = 1 PER_HOST_V1 = 2 PER_HOST_V2 = 3 + BROADCAST = 4 # TODO(b/72511246) Provide a simplified api to configure model parallelism. @@ -50,7 +49,7 @@ class TPUConfig( collections.namedtuple('TPUConfig', [ 'iterations_per_loop', 'num_shards', - 'computation_shape', + 'num_cores_per_replica', 'per_host_input_for_training', 'tpu_job_name', 'initial_infeed_sleep_secs', @@ -67,22 +66,22 @@ class TPUConfig( case, this number equals the total number of TPU cores. For model-parallelism, the total number of TPU cores equals product(computation_shape) * num_shards. - computation_shape: Defaults to `None`, which disables model parallelism. A - list of size 3 which describes the shape of a model replica's block of - cores. This is required by model-parallelism which enables partitioning - the model to multiple cores. For example, [2, 2, 1] means the model is - partitioned across 4 cores which span two cores in both x and y - coordinates. Please refer to @{tf.contrib.tpu.Topology} for the - geometry of a TPU mesh. + num_cores_per_replica: Defaults to `None`, which disables model parallelism. + An integer which describes the number of TPU cores per model replica. This + is required by model-parallelism which enables partitioning + the model to multiple cores. Currently num_cores_per_replica must be + 1, 2, 4, or 8. per_host_input_for_training: If `True`, `PER_HOST_V1`, or `PER_HOST_V2`, - `input_fn` is invoked per-host rather than per-core. With per-host input - pipeline configuration, `input_fn` is invoked once on each host. With the - per-core input pipeline configuration, it is invoked once for each core. + `input_fn` is invoked once on each host. With the per-core input pipeline + configuration, it is invoked once for each core. With a global batch size `train_batch_size` in `TPUEstimator` constructor, the batch size for each shard is `train_batch_size` // #hosts in the `True` or `PER_HOST_V1` mode. In `PER_HOST_V2` mode, it is - `train_batch_size` // #cores. With the per-core input pipeline - configuration, the shard batch size is also `train_batch_size` // #cores. + `train_batch_size` // #cores. In `BROADCAST` mode, `input_fn` is only + invoked once on host 0 and the tensors are broadcasted to all other + replicas. The batch size equals to train_batch_size`. With the per-core + input pipeline configuration, the shard batch size is also + `train_batch_size` // #cores. Note: per_host_input_for_training==PER_SHARD_V1 only supports mode.TRAIN. tpu_job_name: The name of the TPU job. Typically, this name is auto-inferred within TPUEstimator, however when using ClusterSpec propagation in more @@ -99,7 +98,7 @@ class TPUConfig( def __new__(cls, iterations_per_loop=2, num_shards=None, - computation_shape=None, + num_cores_per_replica=None, per_host_input_for_training=True, tpu_job_name=None, initial_infeed_sleep_secs=None): @@ -112,19 +111,12 @@ class TPUConfig( if num_shards is not None: util_lib.check_positive_integer(num_shards, 'TPUConfig num_shards') - # Check computation_shape - if computation_shape is not None and len(computation_shape) != 3: - raise ValueError( - 'computation_shape must be a list with length 3 or None; got {}'. - format(str(computation_shape))) - - if computation_shape is not None: - computation_shape_array = np.asarray(computation_shape, dtype=np.int32) - # This prevents any computation being replicated across multiple hosts, so - # that each host feeds the same number of computations. - if any(computation_shape_array < 1) or any(computation_shape_array > 2): - raise ValueError('computation_shape elements can only be 1 or 2; got ' - 'computation_shape={}'.format(computation_shape)) + # Parse computation_shape + if num_cores_per_replica is not None: + if num_cores_per_replica not in [1, 2, 4, 8]: + raise ValueError( + 'num_cores_per_replica must be 1, 2, 4, or 8; got {}'.format( + str(num_cores_per_replica))) # per_host_input_for_training may be True, False, or integer in [1..3]. # Map legacy values (True, False) to numeric values. @@ -144,7 +136,7 @@ class TPUConfig( cls, iterations_per_loop=iterations_per_loop, num_shards=num_shards, - computation_shape=computation_shape, + num_cores_per_replica=num_cores_per_replica, per_host_input_for_training=per_host_input_for_training, tpu_job_name=tpu_job_name, initial_infeed_sleep_secs=initial_infeed_sleep_secs) @@ -214,6 +206,12 @@ class RunConfig(run_config_lib.RunConfig): self._session_config.cluster_def.CopyFrom( self._cluster_spec.as_cluster_def()) + def _maybe_overwrite_session_config_for_distributed_training(self): + # Overrides the parent class session_config overwrite for between-graph. TPU + # runs with in-graph, which should not have device filter. Doing nothing + # ("pass") basically disables it. + pass + @property def evaluation_master(self): return self._evaluation_master diff --git a/tensorflow/contrib/tpu/python/tpu/tpu_config_test.py b/tensorflow/contrib/tpu/python/tpu/tpu_config_test.py index 37ef3dbe1e66efe18b13ab9153ee346c08b9774a..2326fe97a807e6708a9cdc24fea889b998025a45 100644 --- a/tensorflow/contrib/tpu/python/tpu/tpu_config_test.py +++ b/tensorflow/contrib/tpu/python/tpu/tpu_config_test.py @@ -21,6 +21,7 @@ from __future__ import print_function import json from tensorflow.contrib.tpu.python.tpu import tpu_config as tpu_config_lib +from tensorflow.core.protobuf import config_pb2 from tensorflow.python.estimator import run_config as run_config_lib from tensorflow.python.platform import test @@ -33,6 +34,46 @@ def _set_tf_config_env_variable(tf_config): class TPURunConfigTest(test.TestCase): + def test_no_session_config_set_in_local_case(self): + run_config = tpu_config_lib.RunConfig() + self.assertIsNone(run_config.session_config) + + def test_no_session_config_overwrite_in_local_case(self): + session_config = config_pb2.ConfigProto(allow_soft_placement=True) + run_config = tpu_config_lib.RunConfig(session_config=session_config) + self.assertEqual(session_config, run_config.session_config) + + def test_no_session_config_set_with_cluster_spec(self): + tf_config = { + 'cluster': { + run_config_lib.TaskType.CHIEF: ['host3:3'], + run_config_lib.TaskType.WORKER: ['host3:4'] + }, + 'task': { + 'type': run_config_lib.TaskType.CHIEF, + 'index': 0 + } + } + with _set_tf_config_env_variable(tf_config): + run_config = tpu_config_lib.RunConfig() + self.assertIsNone(run_config.session_config) + + def test_no_session_config_overwrite_with_cluster_spec(self): + tf_config = { + 'cluster': { + run_config_lib.TaskType.CHIEF: ['host3:3'], + run_config_lib.TaskType.WORKER: ['host3:4'] + }, + 'task': { + 'type': run_config_lib.TaskType.CHIEF, + 'index': 0 + } + } + with _set_tf_config_env_variable(tf_config): + session_config = config_pb2.ConfigProto(allow_soft_placement=True) + run_config = tpu_config_lib.RunConfig(session_config=session_config) + self.assertEqual(session_config, run_config.session_config) + def test_fail_with_invalid_num_shards(self): with self.assertRaisesRegexp(ValueError, 'must be positive'): tpu_config_lib.RunConfig( @@ -43,15 +84,11 @@ class TPURunConfigTest(test.TestCase): tpu_config_lib.RunConfig( tpu_config=tpu_config_lib.TPUConfig(iterations_per_loop=0)) - def test_fail_with_invalid_computation_shape(self): - with self.assertRaisesRegexp(ValueError, - 'computation_shape must be a list with length' - ' 3 or None'): - tpu_config_lib.TPUConfig(computation_shape=[2, 1]) - - with self.assertRaisesRegexp(ValueError, - 'computation_shape elements can only be'): - tpu_config_lib.TPUConfig(computation_shape=[1, 3, 1]) + def test_fail_with_invalid_num_cores_per_replica(self): + with self.assertRaisesRegexp( + ValueError, 'num_cores_per_replica must be 1, 2, 4, or 8;' + ' got 7'): + tpu_config_lib.TPUConfig(num_cores_per_replica=7) class TPURunConfigMasterTest(test.TestCase): diff --git a/tensorflow/contrib/tpu/python/tpu/tpu_context.py b/tensorflow/contrib/tpu/python/tpu/tpu_context.py index c4c69902f95e73e90832b3fd6538d73e474e330a..211c59cb90c78e3bd6cfcdccbf3bfe697bacfe24 100644 --- a/tensorflow/contrib/tpu/python/tpu/tpu_context.py +++ b/tensorflow/contrib/tpu/python/tpu/tpu_context.py @@ -21,8 +21,6 @@ from __future__ import print_function from contextlib import contextmanager import copy -import numpy as np - from tensorflow.contrib.tpu.python.tpu import device_assignment as tpu_device_assignment from tensorflow.contrib.tpu.python.tpu import tpu_config from tensorflow.contrib.tpu.python.tpu import tpu_system_metadata as tpu_system_metadata_lib @@ -33,15 +31,26 @@ from tensorflow.python.platform import tf_logging as logging _DEFAULT_JOB_NAME = 'tpu_worker' _DEFAULT_COORDINATOR_JOB_NAME = 'coordinator' _LOCAL_MASTERS = ('', 'local') +_NUM_CORES_TO_COMPUTATION_SHAPE = { + 1: [1, 1, 1], + 2: [1, 1, 2], + 4: [1, 2, 2], + 8: [2, 2, 2] +} class TPUContext(object): """The context of current input_fn invocation.""" - def __init__(self, internal_ctx, input_device=None, invocation_index=None): + def __init__(self, + internal_ctx, + input_device=None, + invocation_index=None, + call_from_input_fn=True): self._internal_ctx = internal_ctx self._input_device = input_device self._invocation_index = invocation_index + self._call_from_input_fn = call_from_input_fn def current_input_fn_deployment(self): """The configuration of the current input_fn invocation. @@ -69,11 +78,21 @@ class TPUContext(object): total invocation count is equal to the number of hosts in the system and num replicas consumed by current invocation is equal to number of cores per host. + + Raises: + RuntimeError: If this method must not be called from input_fn. """ + if not self._call_from_input_fn: + raise RuntimeError('This TPUContext instance must not be called from' + ' model_fn.') + if self._internal_ctx.is_input_sharded_per_core(): total_invocation_count = (self._internal_ctx.num_hosts * self._internal_ctx.num_of_replicas_per_host) replicas_consumed = 1 + elif self._internal_ctx.is_input_broadcast_with_iterators(): + total_invocation_count = 1 + replicas_consumed = self._internal_ctx.num_replicas else: total_invocation_count = self._internal_ctx.num_hosts replicas_consumed = self._internal_ctx.num_of_replicas_per_host @@ -92,6 +111,27 @@ class TPUContext(object): """ return self._internal_ctx.num_replicas + @property + def num_hosts(self): + """The number of hosts for the TPU system.""" + return self._internal_ctx.num_hosts + + @property + def num_of_replicas_per_host(self): + """The number of replicas for each host.""" + if self._internal_ctx.model_parallelism_enabled: + raise ValueError( + 'num_of_replicas_per_host is not supported for model_parallelism') + return self._internal_ctx.num_of_replicas_per_host + + @property + def device_assignment(self): + """Returns device_assignment object.""" + if self._call_from_input_fn: + raise RuntimeError('This TPUContext instance must not be called from' + ' input_fn.') + return self._internal_ctx.device_assignment + def device_for_replica(self, replica_id): """Returns the tuple of (CPU device and device ordinal) for replica. @@ -108,8 +148,8 @@ class TPUContext(object): # as far as model is replicated to all cores in the system. # If the precise replica_id to device mapping is required, please - # set the computation_shape as [1,1,1] in TPUConfig to enable - # the model parallelism. + # set the num_cores_per_replica to 1 in TPUConfig to enable the + # model parallelism. if self._internal_ctx.model_parallelism_enabled: return RuntimeError( 'device_for_replica is not yet implemented for model parallelism. ' @@ -162,9 +202,14 @@ class _InternalTPUContext(object): self._eval_on_tpu = eval_on_tpu self._model_parallelism_enabled = ( - use_tpu and config.tpu_config.computation_shape) + use_tpu and config.tpu_config.num_cores_per_replica) self._mode = None - + num_cores_per_replica = config.tpu_config.num_cores_per_replica + if num_cores_per_replica: + self._computation_shape = _NUM_CORES_TO_COMPUTATION_SHAPE[ + num_cores_per_replica] + else: + self._computation_shape = None self._lazy_tpu_system_metadata_dict = {} # key by master address self._lazy_device_assignment_dict = {} # key by master address self._lazy_validation_dict = {} # key by ModeKeys @@ -225,11 +270,12 @@ class _InternalTPUContext(object): device_assignment = tpu_device_assignment.device_assignment( tpu_system_metadata.topology, - computation_shape=self._config.tpu_config.computation_shape, + computation_shape=self._computation_shape, num_replicas=self.num_replicas) - logging.info('computation_shape: %s', - str(self._config.tpu_config.computation_shape)) + logging.info('num_cores_per_replica: %s', + str(self._config.tpu_config.num_cores_per_replica)) + logging.info('computation_shape: %s', str(self._computation_shape)) logging.info('num_replicas: %d', self.num_replicas) logging.info('device_assignment.topology.device_coordinates: %s', str(device_assignment.topology.device_coordinates)) @@ -270,23 +316,20 @@ class _InternalTPUContext(object): num_cores_in_system = self.num_cores if self.model_parallelism_enabled: - computation_shape_array = np.asarray( - self._config.tpu_config.computation_shape, dtype=np.int32) - num_cores_per_replica = np.prod(computation_shape_array) + num_cores_per_replica = self._config.tpu_config.num_cores_per_replica if num_cores_per_replica > num_cores_in_system: raise ValueError( 'The num of cores required by the model parallelism, specified by ' - 'TPUConfig.computation_shape, is larger than the total num of ' - 'TPU cores in the system. computation_shape: {}, num cores ' - 'in the system: {}'.format( - self._config.tpu_config.computation_shape, - num_cores_in_system)) + 'TPUConfig.num_cores_per_replica, is larger than the total num of ' + 'TPU cores in the system. num_cores_per_replica: {}, num cores ' + 'in the system: {}'.format(num_cores_per_replica, + num_cores_in_system)) if num_cores_in_system % num_cores_per_replica != 0: raise RuntimeError( 'The num of cores in the system ({}) is not divisible by the num ' 'of cores ({}) required by the model parallelism, specified by ' - 'TPUConfig.computation_shape. This should never happen!'.format( + 'TPUConfig.num_cores_per_replica. This should never happen!'.format( num_cores_in_system, num_cores_per_replica)) return num_cores_in_system // num_cores_per_replica @@ -314,6 +357,11 @@ class _InternalTPUContext(object): return (self._config.tpu_config.per_host_input_for_training is tpu_config.InputPipelineConfig.PER_HOST_V2) + def is_input_broadcast_with_iterators(self): + """Return true if input_fn should be run in the full_replicae config.""" + return (self._config.tpu_config.per_host_input_for_training is + tpu_config.InputPipelineConfig.BROADCAST) + def is_running_on_cpu(self, is_export_mode=False): """Determines whether the input_fn and model_fn should be invoked on CPU. @@ -378,7 +426,7 @@ class _InternalTPUContext(object): """Returns the shard batch size for `input_fn`.""" global_batch_size = self.global_batch_size - if self.is_running_on_cpu(): + if (self.is_running_on_cpu() or self.is_input_broadcast_with_iterators()): return global_batch_size # On TPU @@ -393,7 +441,7 @@ class _InternalTPUContext(object): """Returns the shard batch size for `model_fn`.""" global_batch_size = self.global_batch_size - if self.is_running_on_cpu(): + if (self.is_running_on_cpu() or self.is_input_broadcast_with_iterators()): return global_batch_size # On TPU. always sharded per shard. @@ -450,17 +498,23 @@ class _InternalTPUContext(object): master = self.master_job - def _placement_function(_sentinal=None, core_id=None, host_id=None): # pylint: disable=invalid-name + def _placement_function(_sentinal=None, replica_id=None, host_id=None): # pylint: disable=invalid-name + """Return the host device given replica_id or host_id.""" assert _sentinal is None - if core_id is not None and host_id is not None: + if replica_id is not None and host_id is not None: raise RuntimeError( - 'core_id and host_id can have only one non-None value.') + 'replica_id and host_id can have only one non-None value.') if master is None: return '/replica:0/task:0/device:CPU:0' else: - if core_id is not None: - host_id = core_id / self.num_of_cores_per_host + if replica_id is not None: + if self.model_parallelism_enabled: + return self.device_assignment.host_device( + replica=replica_id, job=master) + else: + host_id = replica_id / self.num_of_cores_per_host + return '/job:%s/task:%d/device:CPU:0' % (master, host_id) return _placement_function @@ -533,7 +587,7 @@ class _InternalTPUContext(object): 'be ({}), got ({}). For non-model-parallelism, num_replicas should ' 'be the total num of TPU cores in the system. For ' 'model-parallelism, the total number of TPU cores should be ' - 'product(computation_shape) * num_replicas. Please set it ' + 'num_cores_per_replica * num_replicas. Please set it ' 'accordingly or leave it as `None`'.format( self._get_master_address(), num_replicas, user_provided_num_replicas)) @@ -612,7 +666,7 @@ def _get_tpu_context(config, train_batch_size, eval_batch_size, """Returns an instance of `_InternalTPUContext`.""" if (config.tpu_config.num_shards == 1 and - config.tpu_config.computation_shape is None): + config.tpu_config.num_cores_per_replica is None): logging.warning( 'Setting TPUConfig.num_shards==1 is an unsupported behavior. ' 'Please fix as soon as possible (leaving num_shards as None.') diff --git a/tensorflow/contrib/tpu/python/tpu/tpu_estimator.py b/tensorflow/contrib/tpu/python/tpu/tpu_estimator.py index 2b1cb4245eb7f69801bb28d2bad5ab5b45ff3514..74157a6193f1eca852f22c26f1007a334865fcb5 100644 --- a/tensorflow/contrib/tpu/python/tpu/tpu_estimator.py +++ b/tensorflow/contrib/tpu/python/tpu/tpu_estimator.py @@ -81,12 +81,17 @@ _TPU_ESTIMATOR = 'tpu_estimator' _ITERATIONS_PER_LOOP_VAR = 'iterations_per_loop' _BATCH_SIZE_KEY = 'batch_size' _CTX_KEY = 'context' +_USE_TPU_KEY = 'use_tpu' _CROSS_REPLICA_SUM_OP = 'CrossReplicaSum' _ONE_GIGABYTE = 1024 * 1024 * 1024 _TPU_ENQUEUE_OPS = '_tpu_enqueue_ops' _TPU_TRAIN_OP = '_tpu_train_op' _REWRITE_FOR_INFERENCE_MODE = '_rewrite_for_inference' +# Ideally _USE_TPU_KEY should be reserved as well. However there are already +# models that make use of this key, thus it can not be reserved now to prevent +# breakage. In the long run, we would like to mitigate this by migrating models +# off of using _USE_TPU_KEY. _RESERVED_PARAMS_KEYS = [_BATCH_SIZE_KEY, _CTX_KEY] @@ -211,8 +216,8 @@ class _SIGNAL(object): class TPUEstimatorSpec(model_fn_lib._TPUEstimatorSpec): # pylint: disable=protected-access """Ops and objects returned from a `model_fn` and passed to `TPUEstimator`. - See `EstimatorSpec` for `mode`, 'predictions, 'loss', 'train_op', and - 'export_outputs`. + See `EstimatorSpec` for `mode`, `predictions`, `loss`, `train_op`, and + `export_outputs`. For evaluation, `eval_metrics `is a tuple of `metric_fn` and `tensors`, where `metric_fn` runs on CPU to generate metrics and `tensors` represents the @@ -226,7 +231,7 @@ class TPUEstimatorSpec(model_fn_lib._TPUEstimatorSpec): # pylint: disable=prote size is the first dimension. Once all tensors are available at CPU host from all shards, they are concatenated (on CPU) and passed as positional arguments to the `metric_fn` if `tensors` is list or keyword arguments if `tensors` is - dict. `metric_fn` takes the `tensors` and returns a dict from metric string + a dict. `metric_fn` takes the `tensors` and returns a dict from metric string name to the result of calling a metric function, namely a `(metric_tensor, update_op)` tuple. See `TPUEstimator` for MNIST example how to specify the `eval_metrics`. @@ -842,6 +847,65 @@ def generate_per_host_v2_enqueue_ops_fn_for_host( return enqueue_ops_fn, captured_infeed_queue, hooks, is_dataset +def generate_broadcast_enqueue_ops_fn(ctx, input_fn, inputs_structure_recorder, + num_hosts): + """Generates infeed enqueue ops for one input_fn on all the hosts.""" + captured_infeed_queue = _CapturedObject() + hooks = [] + device_0 = ctx.tpu_host_placement_function(host_id=0) + with ops.device(device_0): + user_context = tpu_context.TPUContext( + internal_ctx=ctx, input_device=device_0, invocation_index=0) + inputs = _Inputs.from_input_fn(input_fn(user_context)) + + is_dataset = inputs.is_dataset + if ctx.mode == model_fn_lib.ModeKeys.PREDICT: + raise TypeError('Mode PREDICT not yet supported in BROADCAST mode.') + + hooks.append(inputs.dataset_initializer_hook()) + num_replicas_per_host = ctx.num_of_replicas_per_host + + def tpu_ordinal_function_impl(replica_id): + if ctx.device_assignment: + return ctx.device_assignment.tpu_ordinal(replica_id=replica_id) + else: + return replica_id % num_replicas_per_host + + def device_function_impl(replica_id): + return ctx.tpu_host_placement_function(replica_id=replica_id) + + def enqueue_ops_fn(): + """Generates enqueue ops for all the hosts.""" + broadcasted_inputs = [] + flattened_inputs = None # Cache result from input_fn. + for host_id in xrange(num_hosts): + with ops.device(ctx.tpu_host_placement_function(host_id=host_id)): + for _ in xrange(ctx.num_of_replicas_per_host): + # Note: input_fn is only called once at host 0 for the first replica. + # The features and labels returned from that invocation are + # broadcasted to other replicas(including the replicas on other + # hosts). + if flattened_inputs is None: + features, labels = inputs.features_and_labels() # Calls get_next() + inputs_structure_recorder.validate_and_record_structure( + features, labels) + flattened_inputs = ( + inputs_structure_recorder.flatten_features_and_labels( + features, labels)) + broadcasted_inputs.append(flattened_inputs) + + infeed_queue = tpu_feed.InfeedQueue( + number_of_tuple_elements=len(broadcasted_inputs[0])) + captured_infeed_queue.capture(infeed_queue) + enqueue_ops = infeed_queue.generate_enqueue_ops( + broadcasted_inputs, + tpu_ordinal_function=tpu_ordinal_function_impl, + placement_function=device_function_impl) + return enqueue_ops + + return enqueue_ops_fn, captured_infeed_queue, hooks, is_dataset + + class _InputPipeline(object): """`_InputPipeline` handles invoking `input_fn` and piping to infeed queue. @@ -1074,6 +1138,22 @@ class _InputPipeline(object): # Infeed_queue_getter must be called after enqueue_ops_fn is called. infeed_queues.append(captured_infeed_queue.get()) + elif self._ctx.is_input_broadcast_with_iterators(): + # Only calls input_fn in host 0. + host_device = tpu_host_placement_fn(host_id=0) + enqueue_ops_fn, captured_infeed_queue, hooks, is_dataset = ( + generate_broadcast_enqueue_ops_fn(self._ctx, self._input_fn, + self._inputs_structure_recorder, + num_hosts)) + all_hooks.extend(hooks) + if is_dataset: + run_infeed_loop_on_coordinator = False + enqueue_ops.append( + _wrap_computation_in_while_loop( + device=host_device, op_fn=enqueue_ops_fn)) + else: + enqueue_ops.append(enqueue_ops_fn()) + infeed_queues.append(captured_infeed_queue.get()) else: for host_id in range(num_hosts): host_device = tpu_host_placement_fn(host_id=host_id) @@ -1414,8 +1494,16 @@ class _ModelFnWrapper(object): if batch_size_for_model_fn is not None: _add_item_to_params(params, _BATCH_SIZE_KEY, batch_size_for_model_fn) + running_on_cpu = self._ctx.is_running_on_cpu(is_export_mode) + _add_item_to_params(params, _USE_TPU_KEY, not running_on_cpu) + + if not running_on_cpu: + user_context = tpu_context.TPUContext( + internal_ctx=self._ctx, call_from_input_fn=False) + _add_item_to_params(params, _CTX_KEY, user_context) + estimator_spec = self._model_fn(features=features, **kwargs) - if (self._ctx.is_running_on_cpu(is_export_mode) and + if (running_on_cpu and isinstance(estimator_spec, model_fn_lib._TPUEstimatorSpec)): # pylint: disable=protected-access # The estimator_spec will be passed to `Estimator` directly, which expects # type `EstimatorSpec`. @@ -1593,7 +1681,7 @@ class _OutfeedHostCall(object): # place all ops on tpu host if possible. # # TODO(jhseu): Evaluate whether this is right for summaries. - with ops.device(self._ctx.tpu_host_placement_function(core_id=0)): + with ops.device(self._ctx.tpu_host_placement_function(replica_id=0)): for name in self._names: dequeue_ops = dequeue_ops_by_name[name] for i, item in enumerate(dequeue_ops): @@ -1978,7 +2066,7 @@ class TPUEstimator(estimator_lib.Estimator): if (config.tpu_config.per_host_input_for_training is tpu_config.InputPipelineConfig.PER_SHARD_V1 and - config.tpu_config.computation_shape): + config.tpu_config.num_cores_per_replica): raise ValueError( 'Model parallelism only supports per host input for training. ' 'Please adjust TPURunconfig.per_host_input_for_training.') @@ -2033,24 +2121,29 @@ class TPUEstimator(estimator_lib.Estimator): strip_default_attrs, save_variables=True, mode=model_fn_lib.ModeKeys.PREDICT, - export_tags=None): + export_tags=None, + check_variables=True): if mode != model_fn_lib.ModeKeys.PREDICT: raise NotImplementedError( 'TPUEstimator only handles mode PREDICT for export_savedmodel(); ' 'got {}.'.format(mode)) - super(TPUEstimator, self)._add_meta_graph_for_mode(builder, - input_receiver_fn_map, - checkpoint_path, - strip_default_attrs, - save_variables, - mode=mode) + (super(TPUEstimator, self). + _add_meta_graph_for_mode(builder, + input_receiver_fn_map, + checkpoint_path, + strip_default_attrs, + save_variables, + mode=mode, + export_tags=export_tags, + check_variables=check_variables)) if self._export_to_tpu: input_receiver_fn_map = {_REWRITE_FOR_INFERENCE_MODE: input_receiver_fn_map[mode]} export_tags = [tag_constants.SERVING, tag_constants.TPU] mode = _REWRITE_FOR_INFERENCE_MODE + # See b/110052256 for why `check_variables` is `False`. (super(TPUEstimator, self). _add_meta_graph_for_mode(builder, input_receiver_fn_map, @@ -2058,7 +2151,8 @@ class TPUEstimator(estimator_lib.Estimator): strip_default_attrs, save_variables=False, mode=mode, - export_tags=export_tags)) + export_tags=export_tags, + check_variables=False)) def _call_model_fn(self, features, labels, mode, config): if mode == _REWRITE_FOR_INFERENCE_MODE: @@ -2284,10 +2378,20 @@ class TPUEstimator(estimator_lib.Estimator): # Clear the bit. self._is_input_fn_invoked = None + # examples_hook is added to training_hooks for both CPU and TPU + # execution. + examples_hook = ExamplesPerSecondHook( + ctx.global_batch_size, + output_dir=self.model_dir, + every_n_steps=self._log_every_n_steps) + if ctx.is_running_on_cpu(is_export_mode=is_export_mode): logging.info('Running %s on CPU', mode) - return model_fn_wrapper.call_without_tpu( + estimator_spec = model_fn_wrapper.call_without_tpu( features, labels, is_export_mode=is_export_mode) + estimator_spec = estimator_spec._replace( + training_hooks=estimator_spec.training_hooks + (examples_hook,)) + return estimator_spec assert labels is None, '`labels` passed to `model_fn` must be `None`.' # TPUEstimator._call_input_fn passes `input_fn` as features to here. @@ -2355,10 +2459,6 @@ class TPUEstimator(estimator_lib.Estimator): }, every_n_iter=logging_hook_frequency) ]) - examples_hook = ExamplesPerSecondHook( - ctx.global_batch_size, - output_dir=self.model_dir, - every_n_steps=self._log_every_n_steps) examples_hook._set_steps_per_run( # pylint: disable=protected-access self._config.tpu_config.iterations_per_loop) hooks.append(examples_hook) @@ -3143,7 +3243,7 @@ def _add_item_to_params(params, key, value): if isinstance(params, hparam.HParams): # For HParams, we need to use special API. if key in params: - params.key = value + params.set_hparam(key, value) else: params.add_hparam(key, value) else: diff --git a/tensorflow/contrib/tpu/python/tpu/tpu_feed.py b/tensorflow/contrib/tpu/python/tpu/tpu_feed.py index 604e6600c81a4136a1f10e79a725a887a96f4d86..a44b4f4622afabced9cb1b801acedb0e7b1e5d12 100644 --- a/tensorflow/contrib/tpu/python/tpu/tpu_feed.py +++ b/tensorflow/contrib/tpu/python/tpu/tpu_feed.py @@ -461,7 +461,10 @@ class InfeedQueue(object): name=full_name, device_ordinal=tpu_ordinal) - def generate_enqueue_ops(self, sharded_inputs, tpu_ordinal_function=None): + def generate_enqueue_ops(self, + sharded_inputs, + tpu_ordinal_function=None, + placement_function=None): """Generates the host-side Ops to enqueue the shards of a tuple. sharded_inputs is a list, one for each shard, of lists of @@ -483,6 +486,9 @@ class InfeedQueue(object): shard index as input and returns the ordinal of the TPU device the shard's infeed should be placed on. tpu_ordinal_function must be set if the inputs are placed on CPU devices. + placement_function: if not None, a function that takes the shard index as + input and returns the host device where the enqueue op should be placed + on. Returns: A list of host-side Ops, one for each shard, that when executed together @@ -508,8 +514,12 @@ class InfeedQueue(object): tpu_ordinal_function = lambda index: -1 name_prefix = "%s/enqueue" % self._name return [ - self._generate_enqueue_op(shard, name_prefix, index, - tpu_ordinal=tpu_ordinal_function(index)) + self._generate_enqueue_op( + shard, + name_prefix, + index, + tpu_ordinal=tpu_ordinal_function(index), + device=placement_function(index) if placement_function else None) for (shard, index) in zip(sharded_inputs, xrange(self.number_of_shards)) ] diff --git a/tensorflow/contrib/tpu/python/tpu/tpu_optimizer.py b/tensorflow/contrib/tpu/python/tpu/tpu_optimizer.py index 15f99d7eebddd46f9f6902b68f01e42359a72cbe..53d33f40777a1c6d93f19c30b2ef5902d63ad2fd 100644 --- a/tensorflow/contrib/tpu/python/tpu/tpu_optimizer.py +++ b/tensorflow/contrib/tpu/python/tpu/tpu_optimizer.py @@ -23,6 +23,7 @@ import collections from tensorflow.contrib.tpu.python.ops import tpu_ops from tensorflow.contrib.tpu.python.tpu import tpu_function +from tensorflow.python.framework import ops from tensorflow.python.ops.losses import losses from tensorflow.python.platform import tf_logging as logging from tensorflow.python.training import optimizer @@ -153,8 +154,9 @@ class CrossShardOptimizer(optimizer.Optimizer): if grad is None: summed_grads_and_vars.append((grad, var)) else: - summed_grads_and_vars.append((tpu_ops.cross_replica_sum( - grad, self._group_assignment), var)) + with ops.colocate_with(grad): + summed_grads_and_vars.append((tpu_ops.cross_replica_sum( + grad, self._group_assignment), var)) return self._opt.apply_gradients(summed_grads_and_vars, global_step, name) def get_slot(self, *args, **kwargs): diff --git a/tensorflow/contrib/training/python/training/sgdr_learning_rate_decay.py b/tensorflow/contrib/training/python/training/sgdr_learning_rate_decay.py new file mode 100644 index 0000000000000000000000000000000000000000..ed0f398e30a7f3c0b1b9378f8fc5d5bfbea1536a --- /dev/null +++ b/tensorflow/contrib/training/python/training/sgdr_learning_rate_decay.py @@ -0,0 +1,187 @@ +# Copyright 2015 The TensorFlow Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================== + +"""SGDR learning rate decay function.""" +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + +import math + +from tensorflow.python.framework import constant_op +from tensorflow.python.framework import ops +from tensorflow.python.ops import math_ops, control_flow_ops + + +def sgdr_decay(learning_rate, global_step, initial_period_steps, + t_mul=2.0, m_mul=1.0, name=None): + """Implements Stochastic Gradient Descent with Warm Restarts (SGDR). + + As described in "SGDR: Stochastic Gradient Descent + with Warm Restarts" by Ilya Loshchilov & Frank Hutter, Proceedings of + ICLR'2017, available at https://arxiv.org/pdf/1608.03983.pdf + + The learning rate decreases according to cosine annealing: + + ```python + learning_rate * 0.5 * (1 + cos(x_val * pi)) # for x_val defined in [0, 1] + ``` + + Thus, at the beginning (when the restart index i = 0), + the learning rate decreases for `initial_period_steps` steps from the initial + learning rate `learning_rate` (when `x_val=0`, we get `cos(0)=1`) to + 0 (when `x_val=1`, we get `cos(pi)=-1`). + + The decrease within the i-th period takes `t_i` steps, + where `t_0` = `initial_period_steps` is the user-defined number of batch + iterations (not epochs as in the paper) to be performed before the first + restart is launched. + + Then, we perform the first restart (i=1) by setting the learning rate to + `learning_rate*(m_mul^i)`, where `m_mul in [0,1]` (set to 1 by default). + The i-th restart runs for `t_i=t_0*(t_mul^i)` steps, i.e., every new + restart runs `t_mul` times longer than the previous one. + + Importantly, when one has no access to a validation set, SGDR suggests + to report the best expected / recommended solution in the following way: + When we are within our initial run (i=0), every new solution represents + SGDR's recommended solution. Instead, when i>0, the recommended solution is + the one obtained at the end of each restart. + + Note that the minimum learning rate is set to 0 for simplicity, + you can adjust the code to deal with any positive minimum learning rate + as defined in the paper. + + `initial_period_steps` is the duration of the first period measured in terms + of number of minibatch updates. If one wants to use epochs, one should compute + the number of updates required for an epoch. + + For example, assume the following parameters and intention: + Minibatch size: 100 + Training dataset size: 10000 + If the user wants the first decay period to span across 5 epochs, then + `initial_period_steps` = 5 * 10000/100 = 500 + + Train for 10000 batch iterations with the initial learning rate set to + 0.1, then restart to run 2 times longer, i.e, for 20000 batch iterations + and with the initial learning rate 0.05, then restart again and again, + doubling the runtime of each new period and with two times smaller + initial learning rate. + + To accomplish the above, one would write: + + ```python + ... + global_step = tf.Variable(0, trainable=False) + starter_learning_rate = 0.1 + learning_rate = sgdr_decay(starter_learning_rate, global_step, + initial_period_steps=10000, t_mul=2, m_mul=0.5) + # Passing global_step to minimize() will increment it at each step. + learning_step = ( + tf.train.GradientDescentOptimizer(learning_rate) + .minimize(...my loss..., global_step=global_step) + ) + + # Step | 0 | 1000 | 5000 | 9000 | 9999 | 10000 | 11000 | + # LR | 0.1 | 0.097 | 0.05 | 0.002 | 0.00 | 0.05 | 0.0496 | + + # Step | 20000 | 29000 | 29999 | 30000 | + # LR | 0.025 | 0.0003 | 0.00 | 0.025 | + ``` + + Args: + learning_rate: A scalar `float32` or `float64` `Tensor` or a + Python number. The initial learning rate. + global_step: A scalar `int32` or `int64` `Tensor` or a Python number. + Global step to use for the decay computation. Must not be negative. + initial_period_steps: Duration of the first period measured as the number + of minibatch updates, if one wants to use epochs, one should compute + the number of updates required for an epoch. + t_mul: A scalar `float32` or `float64` `Tensor` or a Python number. + Must be positive. + Used to derive the number of iterations in the i-th period: + `initial_period_steps * (t_mul^i)`. Defaults to 2.0. + m_mul: A scalar `float32` or `float64` `Tensor` or a Python number. + Must be positive. + Used to derive the initial learning rate of the i-th period: + `learning_rate * (m_mul^i)`. Defaults to 1.0 + + Returns: + A scalar `Tensor` of the same type as `learning_rate`. + The learning rate for a provided global_step. + Raises: + ValueError: if `global_step` is not supplied. + """ + + if global_step is None: + raise ValueError("global_step is required for sgdr_decay.") + with ops.name_scope(name, "SGDRDecay", + [learning_rate, global_step, + initial_period_steps, t_mul, m_mul]) as name: + learning_rate = ops.convert_to_tensor(learning_rate, + name="initial_learning_rate") + dtype = learning_rate.dtype + global_step = math_ops.cast(global_step, dtype) + t_0 = math_ops.cast(initial_period_steps, dtype) + t_mul = math_ops.cast(t_mul, dtype) + m_mul = math_ops.cast(m_mul, dtype) + + c_one = math_ops.cast(constant_op.constant(1.0), dtype) + c_half = math_ops.cast(constant_op.constant(0.5), dtype) + c_pi = math_ops.cast(constant_op.constant(math.pi), dtype) + + # Find normalized value of the current step + x_val = math_ops.div(global_step, t_0) + + def compute_step(x_val, geometric=False): + if geometric: + # Consider geometric series where t_mul != 1 + # 1 + t_mul + t_mul^2 ... = (1 - t_mul^i_restart) / (1 - t_mul) + + # First find how many restarts were performed for a given x_val + # Find maximal integer i_restart value for which this equation holds + # x_val >= (1 - t_mul^i_restart) / (1 - t_mul) + # x_val * (1 - t_mul) <= (1 - t_mul^i_restart) + # t_mul^i_restart <= (1 - x_val * (1 - t_mul)) + + # tensorflow allows only log with base e + # i_restart <= log(1 - x_val * (1 - t_mul) / log(t_mul) + # Find how many restarts were performed + + i_restart = math_ops.floor( + math_ops.log(c_one - x_val * (c_one - t_mul)) / math_ops.log(t_mul)) + # Compute the sum of all restarts before the current one + sum_r = (c_one - t_mul ** i_restart) / (c_one - t_mul) + # Compute our position within the current restart + x_val = (x_val - sum_r) / t_mul ** i_restart + + else: + # Find how many restarts were performed + i_restart = math_ops.floor(x_val) + # Compute our position within the current restart + x_val = x_val - i_restart + return i_restart, x_val + + i_restart, x_val = control_flow_ops.cond( + math_ops.equal(t_mul, c_one), + lambda: compute_step(x_val, geometric=False), + lambda: compute_step(x_val, geometric=True)) + + # If m_mul < 1, then the initial learning rate of every new restart will be + # smaller, i.e., by a factor of m_mul ** i_restart at i_restart-th restart + m_fac = learning_rate * (m_mul ** i_restart) + + return math_ops.multiply(c_half * m_fac, + (math_ops.cos(x_val * c_pi) + c_one), name=name) diff --git a/tensorflow/contrib/training/python/training/sgdr_learning_rate_decay_test.py b/tensorflow/contrib/training/python/training/sgdr_learning_rate_decay_test.py new file mode 100644 index 0000000000000000000000000000000000000000..4a46e9a49ef203384e36698f81d6cbe3a3881ef8 --- /dev/null +++ b/tensorflow/contrib/training/python/training/sgdr_learning_rate_decay_test.py @@ -0,0 +1,145 @@ +# Copyright 2015 The TensorFlow Authors. All Rights Reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +# ============================================================================== + +"""Functional test for sgdr learning rate decay.""" +from __future__ import absolute_import +from __future__ import division +from __future__ import print_function + +import math + +from sgdr_learning_rate_decay import sgdr_decay +from tensorflow.python.platform import googletest +from tensorflow.python.framework import test_util +from tensorflow.python.framework import dtypes +from tensorflow import placeholder + + +class SGDRDecayTest(test_util.TensorFlowTestCase): + """Unit tests for SGDR learning rate decay.""" + + def get_original_values(self, lr, t_e, mult_factor, iter_per_epoch, epochs): + """Get an array with learning rate values from the consecutive steps using + the original implementation + (https://github.com/loshchil/SGDR/blob/master/SGDR_WRNs.py).""" + t0 = math.pi / 2.0 + tt = 0 + te_next = t_e + + lr_values = [] + sh_lr = lr + for epoch in range(epochs): + for _ in range(iter_per_epoch): + # In the original approach training function is executed here + lr_values.append(sh_lr) + dt = 2.0 * math.pi / float(2.0 * t_e) + tt = tt + float(dt) / iter_per_epoch + if tt >= math.pi: + tt = tt - math.pi + cur_t = t0 + tt + new_lr = lr * (1.0 + math.sin(cur_t)) / 2.0 # lr_min = 0, lr_max = lr + sh_lr = new_lr + if (epoch + 1) == te_next: # time to restart + sh_lr = lr + tt = 0 # by setting to 0 we set lr to lr_max, see above + t_e = t_e * mult_factor # change the period of restarts + te_next = te_next + t_e # note the next restart's epoch + + return lr_values + + def get_sgdr_values(self, lr, initial_period_steps, t_mul, iters): + """Get an array with learning rate values from the consecutive steps + using current tensorflow implementation.""" + with self.test_session(): + step = placeholder(dtypes.int32) + + decay = sgdr_decay(lr, step, initial_period_steps, t_mul) + lr_values = [] + for i in range(iters): + lr_values.append(decay.eval(feed_dict={step: i})) + + return lr_values + + def testCompareToOriginal(self): + """Compare values generated by tensorflow implementation to the values + generated by the original implementation + (https://github.com/loshchil/SGDR/blob/master/SGDR_WRNs.py).""" + with self.test_session(): + lr = 10.0 + init_steps = 2 + t_mul = 3 + iters = 10 + epochs = 50 + + org_lr = self.get_original_values(lr, init_steps, t_mul, iters, epochs) + sgdr_lr = self.get_sgdr_values(lr, init_steps*iters, t_mul, iters*epochs) + + for org, sgdr in zip(org_lr, sgdr_lr): + self.assertAllClose(org, sgdr) + + def testMDecay(self): + """Test m_mul argument. Check values for learning rate at the beginning + of the first, second, third and fourth period. """ + with self.test_session(): + step = placeholder(dtypes.int32) + + lr = 0.1 + t_e = 10 + t_mul = 3 + m_mul = 0.9 + + decay = sgdr_decay(lr, step, t_e, t_mul, m_mul) + + test_step = 0 + self.assertAllClose(decay.eval(feed_dict={step: test_step}), + lr) + + test_step = t_e + self.assertAllClose(decay.eval(feed_dict={step: test_step}), + lr * m_mul) + + test_step = t_e + t_e*t_mul + self.assertAllClose(decay.eval(feed_dict={step: test_step}), + lr * m_mul**2) + + test_step = t_e + t_e*t_mul + t_e * (t_mul**2) + self.assertAllClose(decay.eval(feed_dict={step: test_step}), + lr * (m_mul**3)) + + def testCos(self): + """Check learning rate values at the beginning, in the middle + and at the end of the period.""" + with self.test_session(): + step = placeholder(dtypes.int32) + lr = 0.2 + t_e = 1000 + t_mul = 1 + + decay = sgdr_decay(lr, step, t_e, t_mul) + + test_step = 0 + self.assertAllClose(decay.eval(feed_dict={step: test_step}), lr) + + test_step = t_e//2 + self.assertAllClose(decay.eval(feed_dict={step: test_step}), lr/2) + + test_step = t_e + self.assertAllClose(decay.eval(feed_dict={step: test_step}), lr) + + test_step = t_e*3//2 + self.assertAllClose(decay.eval(feed_dict={step: test_step}), lr/2) + +if __name__ == "__main__": + googletest.main() diff --git a/tensorflow/contrib/verbs/rdma.cc b/tensorflow/contrib/verbs/rdma.cc index 86350a08e57e5050f18d019fe80d70f6381c1f7d..f7c979e86320d59ad033e2b8d7fcdff89ce0d133 100644 --- a/tensorflow/contrib/verbs/rdma.cc +++ b/tensorflow/contrib/verbs/rdma.cc @@ -24,8 +24,8 @@ limitations under the License. #include "tensorflow/core/common_runtime/dma_helper.h" #include "tensorflow/core/common_runtime/process_util.h" #if GOOGLE_CUDA +#include "tensorflow/core/common_runtime/gpu/gpu_process_state.h" #include "tensorflow/core/common_runtime/gpu/gpu_util.h" -#include "tensorflow/core/common_runtime/gpu/process_state.h" #endif #include "tensorflow/core/distributed_runtime/rendezvous_mgr_interface.h" #include "tensorflow/core/distributed_runtime/rpc/grpc_util.h" @@ -1084,7 +1084,7 @@ void RdmaTensorResponse::RecvHandler(Rendezvous::ParsedKey parsed, // The tensor must be copied from GPU to CPU, because either: // 1. The tensor is located on a non GDR compatible GPU. // 2. The tensor's meta-data has changed. - Allocator* alloc = ProcessState::singleton()->GetCUDAHostAllocator(0); + Allocator* alloc = GPUProcessState::singleton()->GetCUDAHostAllocator(0); copy = Tensor(alloc, in.dtype(), in.shape()); CountCopies(rm_.name_, (void*)DMAHelper::base(&in), (void*)DMAHelper::base(©), in.TotalBytes(), true); @@ -1541,7 +1541,7 @@ bool RdmaTensorRequest::AllocateTensors() { if (mr_ == nullptr) { // Can't RDMA directly to result. Use a proxy. proxy_tensor_ = - new Tensor(ProcessState::singleton()->GetCUDAHostAllocator(0), + new Tensor(GPUProcessState::singleton()->GetCUDAHostAllocator(0), result_tensor_->dtype(), result_tensor_->shape()); rdma_addr_ = DMAHelper::base(proxy_tensor_); mr_ = diff --git a/tensorflow/contrib/verbs/rdma_mgr.cc b/tensorflow/contrib/verbs/rdma_mgr.cc index 369bd986df5313955bc22d6e5c6d38815908ada3..9cb3d1fbbfdbc6d85a7a9799bd82438f0bf70c4f 100644 --- a/tensorflow/contrib/verbs/rdma_mgr.cc +++ b/tensorflow/contrib/verbs/rdma_mgr.cc @@ -21,8 +21,9 @@ limitations under the License. #include "tensorflow/contrib/verbs/grpc_verbs_client.h" #include "tensorflow/contrib/verbs/verbs_service.pb.h" #include "tensorflow/core/common_runtime/bfc_allocator.h" +#include "tensorflow/core/common_runtime/gpu/gpu_process_state.h" #include "tensorflow/core/common_runtime/gpu/gpu_util.h" -#include "tensorflow/core/common_runtime/gpu/process_state.h" +#include "tensorflow/core/common_runtime/process_state.h" #include "tensorflow/core/distributed_runtime/rpc/grpc_worker_cache.h" #include "tensorflow/core/distributed_runtime/session_mgr.h" #include "tensorflow/core/framework/allocator_registry.h" @@ -282,7 +283,7 @@ void RdmaMgr::InitAllocators() { Allocator* allocators[] = { #if GOOGLE_CUDA - ProcessState::singleton()->GetCUDAHostAllocator(0), + GPUProcessState::singleton()->GetCUDAHostAllocator(0), ProcessState::singleton()->GetCPUAllocator(0), #endif // GOOGLE_CUDA cpu_allocator(), @@ -323,7 +324,8 @@ void RdmaMgr::InitAllocators() { std::bind(&RdmaMemoryMgr::InsertMemoryRegion, &RdmaMemoryMgr::Singleton(), _1, _2, std::string(buf)); - ProcessState::singleton()->AddGPUAllocVisitor(bus_id, cuda_alloc_visitor); + GPUProcessState::singleton()->AddGPUAllocVisitor(bus_id, + cuda_alloc_visitor); LOG(INFO) << "Instrumenting GPU allocator with bus_id " << bus_id; } #endif // GOOGLE_CUDA diff --git a/tensorflow/core/BUILD b/tensorflow/core/BUILD index 59e76cb575ffd3036b6f1c574e13bd0f9fe520b5..dbe87a6dbbb91416eb01e300ebc2c45162587218 100644 --- a/tensorflow/core/BUILD +++ b/tensorflow/core/BUILD @@ -150,7 +150,6 @@ load( "//third_party/mkl:build_defs.bzl", "if_mkl", ) -load("@io_bazel_rules_closure//closure:defs.bzl", "closure_proto_library") exports_files(["ops/ops.pbtxt"]) @@ -234,7 +233,6 @@ tf_proto_library( srcs = [], cc_api_version = 2, default_header = True, - j2objc_api_version = 1, java_api_version = 2, js_api_version = 2, protodeps = [ @@ -335,6 +333,7 @@ filegroup( "platform/init_main.h", "platform/mem.h", "platform/mutex.h", + "platform/numa.h", "platform/thread_annotations.h", ], visibility = ["//visibility:private"], @@ -793,6 +792,7 @@ tf_cuda_library( "framework/graph_def_util.h", "framework/graph_to_functiondef.h", "framework/kernel_def_builder.h", + "framework/kernel_def_util.h", "framework/log_memory.h", "framework/lookup_interface.h", "framework/memory_types.h", @@ -902,6 +902,15 @@ cc_library( hdrs = ["util/ptr_util.h"], ) +cc_library( + name = "status_util", + hdrs = ["util/status_util.h"], + deps = [ + ":graph", + ":lib", + ], +) + cc_library( name = "reader_base", srcs = ["framework/reader_base.cc"], @@ -1198,6 +1207,7 @@ tf_cuda_library( hdrs = [ "common_runtime/device.h", "common_runtime/device_factory.h", + "common_runtime/function.h", "common_runtime/optimization_registry.h", "common_runtime/shape_refiner.h", "graph/algorithm.h", @@ -1252,6 +1262,7 @@ cc_library( "//tensorflow/core/kernels:fake_quant_ops", "//tensorflow/core/kernels:function_ops", "//tensorflow/core/kernels:functional_ops", + "//tensorflow/core/kernels:grappler", "//tensorflow/core/kernels:histogram_op", "//tensorflow/core/kernels:image", "//tensorflow/core/kernels:io", @@ -1941,8 +1952,10 @@ LIB_INTERNAL_PRIVATE_HEADERS = ["framework/resource_handle.h"] + glob( "**/*test*", "lib/gif/**/*", "lib/jpeg/**/*", + "lib/png/**/*", "platform/gif.h", "platform/jpeg.h", + "platform/png.h", "platform/**/cuda.h", "platform/**/stream_executor.h", ], @@ -2037,6 +2050,7 @@ cc_library( "lib/hash/crc32c_accelerate.cc", "lib/gif/**/*", "lib/jpeg/**/*", + "lib/png/**/*", "platform/**/env_time.cc", "platform/**/cuda_libdevice_path.cc", "platform/**/device_tracer.cc", @@ -2132,6 +2146,39 @@ cc_library( ], ) +cc_library( + name = "png_internal", + srcs = ["lib/png/png_io.cc"], + hdrs = [ + "lib/bfloat16/bfloat16.h", + "lib/core/casts.h", + "lib/core/stringpiece.h", + "lib/png/png_io.h", + "platform/byte_order.h", + "platform/cpu_info.h", + "platform/default/integral_types.h", + "platform/default/logging.h", + "platform/logging.h", + "platform/macros.h", + "platform/platform.h", + "platform/png.h", + "platform/types.h", + ], + copts = tf_copts(), + linkopts = select({ + "//tensorflow:freebsd": [], + "//tensorflow:windows": [], + "//tensorflow:windows_msvc": [], + "//conditions:default": ["-ldl"], + }), + deps = [ + ":lib", + ":lib_internal", + "//tensorflow/core/platform/default/build_config:png", + "@zlib_archive//:zlib", + ], +) + cc_library( name = "tflite_portable_logging", srcs = [], @@ -2240,7 +2287,6 @@ tf_proto_library( srcs = ERROR_CODES_PROTO_SRCS, cc_api_version = 2, default_header = True, - j2objc_api_version = 1, java_api_version = 2, js_api_version = 2, provide_cc_alias = True, @@ -2262,7 +2308,6 @@ tf_proto_library( srcs = COMMON_PROTO_SRCS + ADDITIONAL_CORE_PROTO_SRCS, cc_api_version = 2, default_header = True, - j2objc_api_version = 1, java_api_version = 2, js_api_version = 2, protodeps = [ @@ -2660,6 +2705,8 @@ CORE_CPU_LIB_HEADERS = CORE_CPU_BASE_HDRS + [ "common_runtime/step_stats_collector.h", "common_runtime/threadpool_device.h", "common_runtime/visitable_allocator.h", + "common_runtime/process_state.h", + "common_runtime/pool_allocator.h", "graph/gradients.h", "graph/quantize_training.h", ] + if_mkl(["graph/mkl_graph_util.h"]) @@ -2698,7 +2745,9 @@ tf_cuda_library( "common_runtime/optimization_registry.cc", "common_runtime/parallel_concat_optimizer.cc", "common_runtime/placer.cc", + "common_runtime/pool_allocator.cc", "common_runtime/process_function_library_runtime.cc", + "common_runtime/process_state.cc", "common_runtime/process_util.cc", "common_runtime/renamed_device.cc", "common_runtime/rendezvous_mgr.cc", @@ -2885,6 +2934,7 @@ cc_library( ) GPU_RUNTIME_HEADERS = [ + "common_runtime/gpu/cuda_host_allocator.h", "common_runtime/gpu/gpu_bfc_allocator.h", "common_runtime/gpu/gpu_cudamalloc_allocator.h", "common_runtime/gpu/gpu_debug_allocator.h", @@ -2894,10 +2944,9 @@ GPU_RUNTIME_HEADERS = [ "common_runtime/gpu/gpu_id_utils.h", "common_runtime/gpu/gpu_init.h", "common_runtime/gpu/gpu_managed_allocator.h", + "common_runtime/gpu/gpu_process_state.h", "common_runtime/gpu/gpu_stream_util.h", "common_runtime/gpu/gpu_util.h", - "common_runtime/gpu/pool_allocator.h", - "common_runtime/gpu/process_state.h", "common_runtime/gpu_device_context.h", ] @@ -2910,11 +2959,10 @@ tf_cuda_library( "common_runtime/gpu/gpu_device.cc", "common_runtime/gpu/gpu_device_factory.cc", "common_runtime/gpu/gpu_managed_allocator.cc", + "common_runtime/gpu/gpu_process_state.cc", "common_runtime/gpu/gpu_stream_util.cc", "common_runtime/gpu/gpu_util.cc", "common_runtime/gpu/gpu_util_platform_specific.cc", - "common_runtime/gpu/pool_allocator.cc", - "common_runtime/gpu/process_state.cc", ], hdrs = GPU_RUNTIME_HEADERS, copts = tf_copts(), @@ -3224,6 +3272,28 @@ tf_cc_test( ], ) +tf_cc_test( + name = "platform_numa_test", + size = "small", + srcs = ["platform/numa_test.cc"], + tags = [ + # This test will not pass unless it has access to all NUMA nodes + # on the executing machine. + "manual", + "notap", + ], + deps = [ + ":framework", + ":lib", + ":lib_internal", + ":lib_test_internal", + ":protos_all_cc", + ":test", + ":test_main", + "//third_party/eigen3", + ], +) + tf_cc_test( name = "platform_setround_test", size = "small", @@ -3377,6 +3447,7 @@ tf_cc_tests( "framework/graph_def_util_test.cc", "framework/graph_to_functiondef_test.cc", "framework/kernel_def_builder_test.cc", + "framework/kernel_def_util_test.cc", "framework/memory_types_test.cc", "framework/node_def_builder_test.cc", "framework/node_def_util_test.cc", @@ -3426,6 +3497,7 @@ tf_cc_tests( "util/semver_test.cc", "util/sparse/sparse_tensor_test.cc", "util/stat_summarizer_test.cc", + "util/status_util_test.cc", "util/tensor_format_test.cc", "util/tensor_slice_reader_test.cc", "util/tensor_slice_set_test.cc", @@ -3450,6 +3522,7 @@ tf_cc_tests( ":ops", ":protos_all_cc", ":protos_test_cc", + ":status_util", ":test", ":test_main", ":testlib", @@ -3585,6 +3658,7 @@ tf_cc_test_mkl( deps = [ ":core", ":core_cpu", + ":core_cpu_internal", ":framework", ":framework_internal", ":test", @@ -3908,13 +3982,13 @@ tf_cc_test( ], ) -tf_cc_test( +tf_cuda_cc_test( name = "common_runtime_direct_session_test", size = "small", srcs = ["common_runtime/direct_session_test.cc"], + args = [] + if_cuda(["--heap_check=local"]), # The GPU tracer leaks memory linkstatic = tf_kernel_tests_linkstatic(), deps = [ - ":core", ":core_cpu", ":core_cpu_internal", ":direct_session_internal", @@ -3927,6 +4001,7 @@ tf_cc_test( ":test", ":test_main", ":testlib", + "//third_party/eigen3", "//tensorflow/cc:cc_ops", "//tensorflow/core/kernels:control_flow_ops", "//tensorflow/core/kernels:cwise_op", @@ -3940,8 +4015,7 @@ tf_cc_test( "//tensorflow/core/kernels:queue_ops", "//tensorflow/core/kernels:session_ops", "//tensorflow/core/kernels:variable_ops", - "//third_party/eigen3", - ], + ] + if_cuda([":cuda"]), ) # This is identical to :common_runtime_direct_session_test with the addition of diff --git a/tensorflow/core/api_def/api_test.cc b/tensorflow/core/api_def/api_test.cc index 477a0b670e49f8aa4ee8c250d4957886eb865ed5..ae03a61ae66ec8d0119d91eefe8c64e61348e9b4 100644 --- a/tensorflow/core/api_def/api_test.cc +++ b/tensorflow/core/api_def/api_test.cc @@ -149,6 +149,33 @@ void TestAllApiDefAttributeNamesAreValid( } } } + +void TestDeprecatedAttributesSetCorrectly( + const std::unordered_map& api_defs_map) { + for (const auto& name_and_api_def : api_defs_map) { + int num_deprecated_endpoints = 0; + const auto& api_def = name_and_api_def.second; + for (const auto& endpoint : api_def.endpoint()) { + if (endpoint.deprecated()) { + ++num_deprecated_endpoints; + } + } + + const auto& name = name_and_api_def.first; + ASSERT_TRUE(api_def.deprecation_message().empty() || + num_deprecated_endpoints == 0) + << "Endpoints are set to 'deprecated' for deprecated op " << name + << ". If an op is deprecated (i.e. deprecation_message is set), " + << "all the endpoints are deprecated implicitly and 'deprecated' " + << "field should not be set."; + if (num_deprecated_endpoints > 0) { + ASSERT_NE(num_deprecated_endpoints, api_def.endpoint_size()) + << "All " << name << " endpoints are deprecated. Please, set " + << "deprecation_message in api_def_" << name << ".pbtxt instead. " + << "to indicate that the op is deprecated."; + } + } +} } // namespace class BaseApiTest : public ::testing::Test { @@ -171,7 +198,7 @@ TEST_F(BaseApiTest, AllOpsAreInApiDef) { if (excluded_ops->find(op.name()) != excluded_ops->end()) { continue; } - ASSERT_TRUE(api_defs_map_.find(op.name()) != api_defs_map_.end()) + EXPECT_TRUE(api_defs_map_.find(op.name()) != api_defs_map_.end()) << op.name() << " op does not have api_def_*.pbtxt file. " << "Please add api_def_" << op.name() << ".pbtxt file " << "under tensorflow/core/api_def/base_api/ directory."; @@ -236,6 +263,11 @@ TEST_F(BaseApiTest, AllApiDefAttributeNamesAreValid) { TestAllApiDefAttributeNamesAreValid(ops_, api_defs_map_); } +// Checks that deprecation is set correctly. +TEST_F(BaseApiTest, DeprecationSetCorrectly) { + TestDeprecatedAttributesSetCorrectly(api_defs_map_); +} + class PythonApiTest : public ::testing::Test { protected: PythonApiTest() { @@ -272,4 +304,9 @@ TEST_F(PythonApiTest, AllApiDefAttributeNamesAreValid) { TestAllApiDefAttributeNamesAreValid(ops_, api_defs_map_); } +// Checks that deprecation is set correctly. +TEST_F(PythonApiTest, DeprecationSetCorrectly) { + TestDeprecatedAttributesSetCorrectly(api_defs_map_); +} + } // namespace tensorflow diff --git a/tensorflow/core/api_def/base_api/api_def_BoostedTreesCenterBias.pbtxt b/tensorflow/core/api_def/base_api/api_def_BoostedTreesCenterBias.pbtxt new file mode 100644 index 0000000000000000000000000000000000000000..b58b974eb4e43b49d6630449de1a0a6c37a15859 --- /dev/null +++ b/tensorflow/core/api_def/base_api/api_def_BoostedTreesCenterBias.pbtxt @@ -0,0 +1,41 @@ +op { + graph_op_name: "BoostedTreesCenterBias" + visibility: HIDDEN + in_arg { + name: "tree_ensemble_handle" + description: < ## TensorFlow GPU support @@ -511,6 +511,8 @@ on your system: list of supported GPU cards. * [GPU drivers](http://nvidia.com/drivers) that support your version of the CUDA Toolkit. +* NCCL 2.2 to use TensorFlow with multiple GPUs. For details, see [NVIDIA's + documentation](https://developer.nvidia.com/nccl). * The `libcupti-dev` library is the NVIDIA CUDA Profile Tools Interface. This library provides advanced profiling support. To install this library, use the following command for CUDA Toolkit >= 8.0: diff --git a/tensorflow/docs_src/install/install_mac.md b/tensorflow/docs_src/install/install_mac.md index 584f1e2e35caff32a4f8aea5ab5fe94114470219..c6f0c17924c95e11d22b08c8976d9044c365dce2 100644 --- a/tensorflow/docs_src/install/install_mac.md +++ b/tensorflow/docs_src/install/install_mac.md @@ -403,8 +403,7 @@ writing TensorFlow programs: If the system outputs an error message instead of a greeting, see [Common installation problems](#common_installation_problems). -To learn more, see [Get Started with TensorFlow](https://www.tensorflow.org/get_started). - +To learn more, see the [TensorFlow tutorials](../tutorials/). ## Common installation problems diff --git a/tensorflow/docs_src/install/install_raspbian.md b/tensorflow/docs_src/install/install_raspbian.md index 0caab6d335544bfc291894a79f9ed0441eb03561..46c4944ca7448df2c993ee44d5099494b759dea8 100644 --- a/tensorflow/docs_src/install/install_raspbian.md +++ b/tensorflow/docs_src/install/install_raspbian.md @@ -230,7 +230,7 @@ problems, despite the log message. If the system outputs an error message instead of a greeting, see [Common installation problems](#common_installation_problems). -To learn more, see [Get Started with TensorFlow](https://www.tensorflow.org/get_started). +To learn more, see the [TensorFlow tutorials](../tutorials/). ## Common installation problems diff --git a/tensorflow/docs_src/install/install_sources.md b/tensorflow/docs_src/install/install_sources.md index e55520ceaa63129072c684a02e7d9ae21f80073c..fc1f6d05bdc26785090e1fc2c6f47826660090ac 100644 --- a/tensorflow/docs_src/install/install_sources.md +++ b/tensorflow/docs_src/install/install_sources.md @@ -289,17 +289,27 @@ Note: If you're only interested in building the libraries for the TensorFlow C or Java APIs, see [Build the C or Java libraries](#BuildCorJava), you do not need to build the pip package in that case. -To build a pip package for TensorFlow with CPU-only support, -you would typically invoke the following command: +### CPU-only support + +To build a pip package for TensorFlow with CPU-only support: + +

+$ bazel build --config=opt //tensorflow/tools/pip_package:build_pip_package
+
+ +To build a pip package for TensorFlow with CPU-only support for the IntelĀ® MKL-DNN: 
-$ bazel build --config=opt //tensorflow/tools/pip_package:build_pip_package
+$ bazel build --config=mkl --config=opt //tensorflow/tools/pip_package:build_pip_package
-To build a pip package for TensorFlow with GPU support, -invoke the following command: +### GPU support + +To build a pip package for TensorFlow with GPU support: -
$ bazel build --config=opt --config=cuda //tensorflow/tools/pip_package:build_pip_package
+
+$ bazel build --config=opt --config=cuda //tensorflow/tools/pip_package:build_pip_package
+
**NOTE on gcc 5 or later:** the binary pip packages available on the TensorFlow website are built with gcc 4, which uses the older ABI. To @@ -362,7 +372,7 @@ TensorFlow programs: 
Hello, TensorFlow!
-To learn more, see [Get Started with TensorFlow](https://www.tensorflow.org/get_started). +To learn more, see the [TensorFlow tutorials](../tutorials/). If the system outputs an error message instead of a greeting, see [Common installation problems](#common_installation_problems). diff --git a/tensorflow/docs_src/install/install_windows.md b/tensorflow/docs_src/install/install_windows.md index 7fe94f0bc3850b7210e83f746f8f8fd5b343cbd3..7b7b17ce81407bbbff837a00bb43162b4b2d44f3 100644 --- a/tensorflow/docs_src/install/install_windows.md +++ b/tensorflow/docs_src/install/install_windows.md @@ -157,7 +157,7 @@ TensorFlow programs: If the system outputs an error message instead of a greeting, see [Common installation problems](#common_installation_problems). -To learn more, see [Get Started with TensorFlow](https://www.tensorflow.org/get_started). +To learn more, see the [TensorFlow tutorials](../tutorials/). ## Common installation problems diff --git a/tensorflow/docs_src/javascript/index.md b/tensorflow/docs_src/javascript/index.md deleted file mode 100644 index ad63eeb255d870064567a0de8a28815ce2ae0172..0000000000000000000000000000000000000000 --- a/tensorflow/docs_src/javascript/index.md +++ /dev/null @@ -1,5 +0,0 @@ -# JavaScript - -You may develop TensorFlow programs in JavaScript, training and deploying -models right in your browser. For details, see -[js.tensorflow.org](https://js.tensorflow.org). diff --git a/tensorflow/docs_src/javascript/leftnav_files b/tensorflow/docs_src/javascript/leftnav_files deleted file mode 100644 index fc0ab8a5435943f6442969ec5787305b98c7908b..0000000000000000000000000000000000000000 --- a/tensorflow/docs_src/javascript/leftnav_files +++ /dev/null @@ -1 +0,0 @@ -index.md diff --git a/tensorflow/docs_src/mobile/leftnav_files b/tensorflow/docs_src/mobile/leftnav_files index 585470d5f0847716863ba6129bf75c26631fecbd..97340ef7e1af64634f8590b5d21a344b5181cb73 100644 --- a/tensorflow/docs_src/mobile/leftnav_files +++ b/tensorflow/docs_src/mobile/leftnav_files @@ -4,6 +4,7 @@ tflite/index.md tflite/devguide.md tflite/demo_android.md tflite/demo_ios.md +tflite/performance.md >>> ### TensorFlow Mobile mobile_intro.md diff --git a/tensorflow/docs_src/mobile/mobile_intro.md b/tensorflow/docs_src/mobile/mobile_intro.md index 241f01d460ae35e818a61be4c4914b3bd8dae00a..baad4433083d18a19ea3dd5ec0c1bae498ac2da9 100644 --- a/tensorflow/docs_src/mobile/mobile_intro.md +++ b/tensorflow/docs_src/mobile/mobile_intro.md @@ -38,7 +38,8 @@ speech-driven interface, and many of these require on-device processing. Most of the time a user isnā€™t giving commands, and so streaming audio continuously to a remote server would be a waste of bandwidth, since it would mostly be silence or background noises. To solve this problem itā€™s common to have a small neural -network running on-device @{$tutorials/audio_recognition$listening out for a particular keyword}. +network running on-device +[listening out for a particular keyword](../tutorials/sequences/audio_recognition). Once that keyword has been spotted, the rest of the conversation can be transmitted over to the server for further processing if more computing power is needed. diff --git a/tensorflow/docs_src/mobile/tflite/demo_android.md b/tensorflow/docs_src/mobile/tflite/demo_android.md index 6f9893f8f18b4d94dee887ce797f4a9440ed1a8a..fdf0bcf3c1135f0e702c7dda4d1d608a26169470 100644 --- a/tensorflow/docs_src/mobile/tflite/demo_android.md +++ b/tensorflow/docs_src/mobile/tflite/demo_android.md @@ -1,7 +1,7 @@ # Android Demo App An example Android application using TensorFLow Lite is available -[on GitHub](https://github.com/tensorflow/tensorflow/tree/master/tensorflow/contrib/lite/java/demo/app). +[on GitHub](https://github.com/tensorflow/tensorflow/tree/master/tensorflow/contrib/lite/java/demo). The demo is a sample camera app that classifies images continuously using either a quantized Mobilenet model or a floating point Inception-v3 model. To run the demo, a device running Android 5.0 ( API 21) or higher is required. diff --git a/tensorflow/docs_src/mobile/tflite/devguide.md b/tensorflow/docs_src/mobile/tflite/devguide.md index 4133bc172a1924f0ce8bb515d66fc03d716923c8..b168d6c18366708ebaa7216481d262b02051168d 100644 --- a/tensorflow/docs_src/mobile/tflite/devguide.md +++ b/tensorflow/docs_src/mobile/tflite/devguide.md @@ -54,10 +54,11 @@ both floating point and quantized inference. ### Train a custom model A developer may choose to train a custom model using Tensorflow (see the -@{$tutorials} for examples of building and training models). If you have already -written a model, the first step is to export this to a @{tf.GraphDef} file. This -is required because some formats do not store the model structure outside the -code, and we must communicate with other parts of the framework. See +[TensorFlow tutorials](../../tutorials/) for examples of building and training +models). If you have already written a model, the first step is to export this +to a @{tf.GraphDef} file. This is required because some formats do not store the +model structure outside the code, and we must communicate with other parts of the +framework. See [Exporting the Inference Graph](https://github.com/tensorflow/models/blob/master/research/slim/README.md) to create .pb file for the custom model. diff --git a/tensorflow/docs_src/mobile/tflite/performance.md b/tensorflow/docs_src/mobile/tflite/performance.md new file mode 100644 index 0000000000000000000000000000000000000000..79bacaaa1b889a8711e5c09c7fd4e4912e70d3bd --- /dev/null +++ b/tensorflow/docs_src/mobile/tflite/performance.md @@ -0,0 +1,174 @@ +# Performance + +This document lists TensorFlow Lite performance benchmarks when running well +known models on some Android and iOS devices. + +These performance benchmark numbers were generated with the +[Android TFLite benchmark binary](https://github.com/tensorflow/tensorflow/tree/master/tensorflow/contrib/lite/tools/benchmark) +and the [iOS benchmark app](https://github.com/tensorflow/tensorflow/tree/master/tensorflow/contrib/lite/tools/benchmark/ios). + +# Android performance benchmarks + +For Android benchmarks, the CPU affinity is set to use big cores on the device to +reduce variance (see [details](https://github.com/tensorflow/tensorflow/tree/master/tensorflow/contrib/lite/tools/benchmark#reducing-variance-between-runs-on-android)). + +It assumes that models were download and unzipped to the +`/data/local/tmp/tflite_models` directory. The benchmark binary is built +using [these instructions](https://github.com/tensorflow/tensorflow/tree/master/tensorflow/contrib/lite/tools/benchmark#on-android) +and assumed in the `/data/local/tmp` directory. + +To run the benchmark: + +``` +adb shell taskset ${CPU_MASK} /data/local/tmp/benchmark_model \ + --num_threads=1 \ + --graph=/data/local/tmp/tflite_models/${GRAPH} \ + --warmup_runs=1 \ + --num_runs=50 \ + --use_nnapi=false +``` + +Here, `${GRAPH}` is the name of model and `${CPU_MASK}` is the CPU affinity +chosen according to the following table: + +Device | CPU_MASK | +-------| ---------- +Pixel 2 | f0 | +Pixel xl | 0c | + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +
Model NameDevice Mean inference time (std dev)
+ Mobilenet_1.0_224(float) + Pixel 2 166.5 ms (2.6 ms)
Pixel xl 122.9 ms (1.8 ms)
+ Mobilenet_1.0_224 (quant) + Pixel 2 69.5 ms (0.9 ms)
Pixel xl 78.9 ms (2.2 ms)
+ NASNet mobile + Pixel 2 273.8 ms (3.5 ms)
Pixel xl 210.8 ms (4.2 ms)
+ SqueezeNet + Pixel 2 234.0 ms (2.1 ms)
Pixel xl 158.0 ms (2.1 ms)
+ Inception_ResNet_V2 + Pixel 2 2846.0 ms (15.0 ms)
Pixel xl 1973.0 ms (15.0 ms)
+ Inception_V4 + Pixel 2 3180.0 ms (11.7 ms)
Pixel xl 2262.0 ms (21.0 ms)
+ +# iOS benchmarks + +To run iOS benchmarks, the [benchmark +app](https://github.com/tensorflow/tensorflow/tree/master/tensorflow/contrib/lite/tools/benchmark/ios) +was modified to include the appropriate model and `benchmark_params.json` was +modified to set `num_threads` to 1. + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +
Model NameDevice Mean inference time (std dev)
+ Mobilenet_1.0_224(float) + iPhone 8 32.2 ms (0.8 ms)
+ Mobilenet_1.0_224 (quant) + iPhone 8 24.4 ms (0.8 ms)
+ NASNet mobile + iPhone 8 60.3 ms (0.6 ms)
+ SqueezeNet + iPhone 8 44.3 (0.7 ms)
+ Inception_ResNet_V2 + iPhone 8562.4 ms (18.2 ms)
+ Inception_V4 + iPhone 8 661.0 ms (29.2 ms)
diff --git a/tensorflow/docs_src/performance/xla/operation_semantics.md b/tensorflow/docs_src/performance/xla/operation_semantics.md index f7e116bf0f85fe94b2167eca8b623207432b38e9..68c427a31661e26adf2d0d599a5d9eb4a18f57be 100644 --- a/tensorflow/docs_src/performance/xla/operation_semantics.md +++ b/tensorflow/docs_src/performance/xla/operation_semantics.md @@ -1308,12 +1308,10 @@ See also : : : parameters of type T and M of : : : : arbitrary type : | `dimensions` | `int64` array | array of map dimensions | -| `static_operands` | sequence of M `XlaOp`s | M arrays of arbitrary type | Applies a scalar function over the given `operands` arrays, producing an array of the same dimensions where each element is the result of the mapped function -applied to the corresponding elements in the input arrays with `static_operands` -given as additional input to `computation`. +applied to the corresponding elements in the input arrays. The mapped function is an arbitrary computation with the restriction that it has N inputs of scalar type `T` and a single output with type `S`. The output has @@ -2012,13 +2010,42 @@ Slice(b, {2, 1}, {4, 3}) produces: See also [`XlaBuilder::Sort`](https://www.tensorflow.org/code/tensorflow/compiler/xla/client/xla_client/xla_builder.h). -Sorts the elements in the operand. +There are two versions of the Sort instruction: a single-operand and a +two-operand version. `Sort(operand)` -Arguments | Type | Semantics ---------- | ------- | ------------------- -`operand` | `XlaOp` | The operand to sort +Arguments | Type | Semantics +----------- | ------- | -------------------- +`operand` | `XlaOp` | The operand to sort. +`dimension` | `int64` | The dimension along which to sort. + +Sorts the elements in the operand in ascending order along the provided +dimension. For example, for a rank-2 (matrix) operand, a `dimension` value of 0 +will sort each column independently, and a `dimension` value of 1 will sort each +row independently. If the operand's elements have floating point type, and the +operand contains NaN elements, the order of elements in the output is +implementation-defined. + +`Sort(key, value)` + +Sorts both the key and the value operands. The keys are sorted as in the +single-operand version. The values are sorted according to the order of their +corresponding keys. For example, if the inputs are `keys = [3, 1]` and +`values = [42, 50]`, then the output of the sort is the tuple +`{[1, 3], [50, 42]}`. + +The sort is not guaranteed to be stable, that is, if the keys array contains +duplicates, the order of their corresponding values may not be preserved. + +Arguments | Type | Semantics +----------- | ------- | ------------------- +`keys` | `XlaOp` | The sort keys. +`values` | `XlaOp` | The values to sort. +`dimension` | `int64` | The dimension along which to sort. + +The `keys` and `values` must have the same dimensions, but may have different +element types. ## Transpose diff --git a/tensorflow/docs_src/get_started/_index.yaml b/tensorflow/docs_src/tutorials/_index.yaml similarity index 75% rename from tensorflow/docs_src/get_started/_index.yaml rename to tensorflow/docs_src/tutorials/_index.yaml index 277fc852fb5377aecb9c82b75555a74afd074ac2..07d561b8a2de364b737f2acc3b6745ea80b1be3d 100644 --- a/tensorflow/docs_src/get_started/_index.yaml +++ b/tensorflow/docs_src/tutorials/_index.yaml @@ -66,9 +66,7 @@ landing_page: }
- -

Learn and use ML

- +

Learn and use ML

The high-level Keras API provides building blocks to create and @@ -77,11 +75,11 @@ landing_page: TensorFlow Keras guide.

    -
  1. Basic classification
    2. -
  2. Text classification
    3. -
  3. Regression
    4. -
  4. Overfitting and underfitting
    5. -
  5. Save and load
    6. +
  6. Basic classification
    7. +
  7. Text classification
    8. +
  8. Regression
    9. +
  9. Overfitting and underfitting
    10. +
  10. Save and load
@@ -111,15 +109,13 @@ landing_page: model.evaluate(x_test, y_test) {% dynamic if request.tld != 'cn' %} - Run in a Notebook + Run in a Notebook {% dynamic endif %} - items: - custom_html: >
- -

Research and experimentation

- +

Research and experimentation

Eager execution provides an imperative, define-by-run interface for advanced operations. Write custom layers, forward passes, and training loops with autoā€‘differentiation. Start with @@ -128,38 +124,38 @@ landing_page:

  1. {% dynamic if request.tld == 'cn' %} - Eager execution basics + Eager execution basics {% dynamic else %} - Eager execution basics + Eager execution basics {% dynamic endif %}
  2. {% dynamic if request.tld == 'cn' %} - Automatic differentiation and gradient tapes + Automatic differentiation and gradient tape {% dynamic else %} - Automatic differentiation and gradient tapes + Automatic differentiation and gradient tape {% dynamic endif %}
  3. {% dynamic if request.tld == 'cn' %} - Variables, models, and training + Custom training: basics {% dynamic else %} - Variables, models, and training + Custom training: basics {% dynamic endif %}
  4. {% dynamic if request.tld == 'cn' %} - Custom layers + Custom layers {% dynamic else %} - Custom layers + Custom layers {% dynamic endif %}
    5. -
  5. Custom training walkthrough
    6. +
  6. Custom training: walkthrough
  7. {% dynamic if request.tld == 'cn' %} Example: Neural machine translation w/ attention {% dynamic else %} - Example: Neural machine translation w/ attention + Example: Neural machine translation w/ attention {% dynamic endif %}
@@ -170,19 +166,20 @@ landing_page:
- custom_html: >
- -

ML at production scale

- +

ML at production scale

Estimators can train large models on multiple machines in a - production environment. Try the examples below and read the + production environment. TensorFlow provides a collection of + pre-made Estimators to implement common ML algorithms. See the Estimators guide.

    -
  1. How to build a simple text classifier with TF-Hub
    2. -
  2. Classifying Higgs boson processes
    3. -
  3. Wide and deep learning using estimators
    4. +
  4. Premade Estimators guide
    5. +
  5. Wide and deep learning with Estimators
    6. +
  6. Boosted trees
    7. +
  7. How to build a simple text classifier with TF-Hub
    8. +
  8. Build a Convolutional Neural Network using Estimators
@@ -193,7 +190,7 @@ landing_page: - description: >

Google Colab: An easy way to learn and use TensorFlow

- Colaboratory + Colaboratory is a Google research project created to help disseminate machine learning education and research. It's a Jupyter notebook environment that requires no setup to use and runs entirely in the cloud. diff --git a/tensorflow/docs_src/tutorials/_toc.yaml b/tensorflow/docs_src/tutorials/_toc.yaml new file mode 100644 index 0000000000000000000000000000000000000000..4db97e35fc5cfe8575200d26692b7fee0b1eef49 --- /dev/null +++ b/tensorflow/docs_src/tutorials/_toc.yaml @@ -0,0 +1,101 @@ +toc: +- title: Get started with TensorFlow + path: /tutorials/ + +- title: Learn and use ML + style: accordion + section: + - title: Overview + path: /tutorials/keras/ + - title: Basic classification + path: /tutorials/keras/basic_classification + - title: Text classification + path: /tutorials/keras/basic_text_classification + - title: Regression + path: /tutorials/keras/basic_regression + - title: Overfitting and underfitting + path: /tutorials/keras/overfit_and_underfit + - title: Save and restore models + path: /tutorials/keras/save_and_restore_models + +- title: Research and experimentation + style: accordion + section: + - title: Overview + path: /tutorials/eager/ + - title: Eager execution + path: https://github.com/tensorflow/tensorflow/blob/master/tensorflow/contrib/eager/python/examples/notebooks/eager_basics.ipynb + status: external + - title: Automatic differentiation + path: https://github.com/tensorflow/tensorflow/blob/master/tensorflow/contrib/eager/python/examples/notebooks/automatic_differentiation.ipynb + status: external + - title: "Custom training: basics" + path: https://github.com/tensorflow/tensorflow/blob/master/tensorflow/contrib/eager/python/examples/notebooks/custom_training.ipynb + status: external + - title: Custom layers + path: https://github.com/tensorflow/tensorflow/blob/master/tensorflow/contrib/eager/python/examples/notebooks/custom_layers.ipynb + status: external + - title: "Custom training: walkthrough" + path: /tutorials/eager/custom_training_walkthrough + - title: Translation with attention + path: https://github.com/tensorflow/tensorflow/blob/master/tensorflow/contrib/eager/python/examples/nmt_with_attention/nmt_with_attention.ipynb + status: external + +- title: ML at production scale + style: accordion + section: + - title: Wide and deep learning + path: https://github.com/tensorflow/models/tree/master/official/wide_deep + status: external + - title: Boosted trees + path: https://github.com/tensorflow/models/tree/master/official/boosted_trees + status: external + - title: Text classifier with TF-Hub + path: /hub/tutorials/text_classification_with_tf_hub + - title: Build a CNN using Estimators + path: /tutorials/estimators/cnn + +- title: Images + style: accordion + section: + - title: Image recognition + path: /tutorials/images/image_recognition + - title: Image retraining + path: /hub/tutorials/image_retraining + - title: Advanced CNN + path: /tutorials/images/deep_cnn + +- title: Sequences + style: accordion + section: + - title: Recurrent neural network + path: /tutorials/sequences/recurrent + - title: Drawing classification + path: /tutorials/sequences/recurrent_quickdraw + - title: Simple audio recognition + path: /tutorials/sequences/audio_recognition + - title: Neural machine translation + path: https://github.com/tensorflow/nmt + status: external + +- title: Data representation + style: accordion + section: + - title: Vector representations of words + path: /tutorials/representation/word2vec + - title: Kernel methods + path: /tutorials/representation/kernel_methods + - title: Large-scale linear models + path: /tutorials/representation/linear + +- title: Non-ML + style: accordion + section: + - title: Mandelbrot set + path: /tutorials/non-ml/mandelbrot + - title: Partial differential equations + path: /tutorials/non-ml/pdes + +- break: True +- title: Next steps + path: /tutorials/next_steps diff --git a/tensorflow/docs_src/tutorials/eager/custom_training_walkthrough.md b/tensorflow/docs_src/tutorials/eager/custom_training_walkthrough.md new file mode 100644 index 0000000000000000000000000000000000000000..b45fbefac01c575515798af4692318ea1e905607 --- /dev/null +++ b/tensorflow/docs_src/tutorials/eager/custom_training_walkthrough.md @@ -0,0 +1,3 @@ +# Custom training: walkthrough + +[Colab notebook](https://colab.research.google.com/github/tensorflow/models/blob/master/samples/core/get_started/eager.ipynb) diff --git a/tensorflow/docs_src/tutorials/eager/index.md b/tensorflow/docs_src/tutorials/eager/index.md new file mode 100644 index 0000000000000000000000000000000000000000..5445e0c3439392d4eeb8a6b3e9d229407b5b014e --- /dev/null +++ b/tensorflow/docs_src/tutorials/eager/index.md @@ -0,0 +1,13 @@ +# Research and experimentation + +Eager execution provides an imperative, define-by-run interface for advanced +operations. Write custom layers, forward passes, and training loops with +auto differentiation. Start with these notebooks, then read the +[eager execution guide](../../guide/eager). + +1. [Eager execution](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/contrib/eager/python/examples/notebooks/eager_intro.ipynb){:.external} +2. [Automatic differentiation and gradient tape](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/contrib/eager/python/examples/notebooks/automatic_differentiation.ipynb){:.external} +3. [Custom training: basics](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/contrib/eager/python/examples/notebooks/custom_training.ipynb){:.external} +4. [Custom layers](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/contrib/eager/python/examples/notebooks/custom_layers.ipynb){:.external} +5. [Custom training: walkthrough](/tutorials/eager/custom_training_walkthrough) +6. [Advanced example: Neural machine translation with attention](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/contrib/eager/python/examples/nmt_with_attention/nmt_with_attention.ipynb){:.external} diff --git a/tensorflow/docs_src/tutorials/layers.md b/tensorflow/docs_src/tutorials/estimators/cnn.md similarity index 94% rename from tensorflow/docs_src/tutorials/layers.md rename to tensorflow/docs_src/tutorials/estimators/cnn.md index 212e33763779d2a712c0f83a72fc346ea5ac801a..12a215b50c54f276f3c084885810c7a496769681 100644 --- a/tensorflow/docs_src/tutorials/layers.md +++ b/tensorflow/docs_src/tutorials/estimators/cnn.md @@ -1,4 +1,4 @@ -# A Guide to TF Layers: Building a Convolutional Neural Network +# Build a Convolutional Neural Network using Estimators The TensorFlow @{tf.layers$`layers` module} provides a high-level API that makes it easy to construct a neural network. It provides methods that facilitate the @@ -470,51 +470,18 @@ as the loss metric. The following code calculates cross entropy when the model runs in either `TRAIN` or `EVAL` mode: ```python -onehot_labels = tf.one_hot(indices=tf.cast(labels, tf.int32), depth=10) -loss = tf.losses.softmax_cross_entropy( - onehot_labels=onehot_labels, logits=logits) +loss = tf.losses.sparse_softmax_cross_entropy(labels=labels, logits=logits) ``` Let's take a closer look at what's happening above. -Our `labels` tensor contains a list of predictions for our examples, e.g. `[1, -9, ...]`. In order to calculate cross-entropy, first we need to convert `labels` -to the corresponding -[one-hot encoding](https://www.quora.com/What-is-one-hot-encoding-and-when-is-it-used-in-data-science): +Our `labels` tensor contains a list of prediction indices for our examples, e.g. `[1, +9, ...]`. `logits` contains the linear outputs of our last layer. -```none -[[0, 1, 0, 0, 0, 0, 0, 0, 0, 0], - [0, 0, 0, 0, 0, 0, 0, 0, 0, 1], - ...] -``` - -We use the @{tf.one_hot} function -to perform this conversion. `tf.one_hot()` has two required arguments: - -* `indices`. The locations in the one-hot tensor that will have "on - values"ā€”i.e., the locations of `1` values in the tensor shown above. -* `depth`. The depth of the one-hot tensorā€”i.e., the number of target classes. - Here, the depth is `10`. +`tf.losses.sparse_softmax_cross_entropy`, calculates the softmax crossentropy +(aka: categorical crossentropy, negative log-likelihood) from these two inputs +in an efficient, numerically stable way. -The following code creates the one-hot tensor for our labels, `onehot_labels`: - -```python -onehot_labels = tf.one_hot(indices=tf.cast(labels, tf.int32), depth=10) -``` - -Because `labels` contains a series of values from 0ā€“9, `indices` is just our -`labels` tensor, with values cast to integers. The `depth` is `10` because we -have 10 possible target classes, one for each digit. - -Next, we compute cross-entropy of `onehot_labels` and the softmax of the -predictions from our logits layer. `tf.losses.softmax_cross_entropy()` takes -`onehot_labels` and `logits` as arguments, performs softmax activation on -`logits`, calculates cross-entropy, and returns our `loss` as a scalar `Tensor`: - -```python -loss = tf.losses.softmax_cross_entropy( - onehot_labels=onehot_labels, logits=logits) -``` ### Configure the Training Op diff --git a/tensorflow/docs_src/tutorials/image_retraining.md b/tensorflow/docs_src/tutorials/image_retraining.md deleted file mode 100644 index 27784eef9cdb5c6f8b9af44b3fc3f876cda39d13..0000000000000000000000000000000000000000 --- a/tensorflow/docs_src/tutorials/image_retraining.md +++ /dev/null @@ -1,4 +0,0 @@ -# How to Retrain Inception's Final Layer for New Categories - -**NOTE: This tutorial has moved to** -https://github.com/tensorflow/hub/tree/master/docs/tutorials/image_retraining.md diff --git a/tensorflow/docs_src/tutorials/deep_cnn.md b/tensorflow/docs_src/tutorials/images/deep_cnn.md similarity index 93% rename from tensorflow/docs_src/tutorials/deep_cnn.md rename to tensorflow/docs_src/tutorials/images/deep_cnn.md index 44a32d9d1dcbd7d4be7a2063e9c5ae4affffe487..27963575f5a02eb8a91b490fdfcc33d35749963c 100644 --- a/tensorflow/docs_src/tutorials/deep_cnn.md +++ b/tensorflow/docs_src/tutorials/images/deep_cnn.md @@ -1,7 +1,4 @@ -# Convolutional Neural Networks - -> **NOTE:** This tutorial is intended for *advanced* users of TensorFlow -and assumes expertise and experience in machine learning. +# Advanced Convolutional Neural Networks ## Overview @@ -83,21 +80,21 @@ for details. It consists of 1,068,298 learnable parameters and requires about ## Code Organization The code for this tutorial resides in -[`models/tutorials/image/cifar10/`](https://www.tensorflow.org/code/tensorflow_models/tutorials/image/cifar10/). +[`models/tutorials/image/cifar10/`](https://github.com/tensorflow/models/tree/master/tutorials/image/cifar10/). File | Purpose --- | --- -[`cifar10_input.py`](https://www.tensorflow.org/code/tensorflow_models/tutorials/image/cifar10/cifar10_input.py) | Reads the native CIFAR-10 binary file format. -[`cifar10.py`](https://www.tensorflow.org/code/tensorflow_models/tutorials/image/cifar10/cifar10.py) | Builds the CIFAR-10 model. -[`cifar10_train.py`](https://www.tensorflow.org/code/tensorflow_models/tutorials/image/cifar10/cifar10_train.py) | Trains a CIFAR-10 model on a CPU or GPU. -[`cifar10_multi_gpu_train.py`](https://www.tensorflow.org/code/tensorflow_models/tutorials/image/cifar10/cifar10_multi_gpu_train.py) | Trains a CIFAR-10 model on multiple GPUs. -[`cifar10_eval.py`](https://www.tensorflow.org/code/tensorflow_models/tutorials/image/cifar10/cifar10_eval.py) | Evaluates the predictive performance of a CIFAR-10 model. +[`cifar10_input.py`](https://github.com/tensorflow/models/tree/master/tutorials/image/cifar10/cifar10_input.py) | Reads the native CIFAR-10 binary file format. +[`cifar10.py`](https://github.com/tensorflow/models/tree/master/tutorials/image/cifar10/cifar10.py) | Builds the CIFAR-10 model. +[`cifar10_train.py`](https://github.com/tensorflow/models/tree/master/tutorials/image/cifar10/cifar10_train.py) | Trains a CIFAR-10 model on a CPU or GPU. +[`cifar10_multi_gpu_train.py`](https://github.com/tensorflow/models/tree/master/tutorials/image/cifar10/cifar10_multi_gpu_train.py) | Trains a CIFAR-10 model on multiple GPUs. +[`cifar10_eval.py`](https://github.com/tensorflow/models/tree/master/tutorials/image/cifar10/cifar10_eval.py) | Evaluates the predictive performance of a CIFAR-10 model. ## CIFAR-10 Model The CIFAR-10 network is largely contained in -[`cifar10.py`](https://www.tensorflow.org/code/tensorflow_models/tutorials/image/cifar10/cifar10.py). +[`cifar10.py`](https://github.com/tensorflow/models/tree/master/tutorials/image/cifar10/cifar10.py). The complete training graph contains roughly 765 operations. We find that we can make the code most reusable by constructing the graph with the following modules: @@ -438,9 +435,6 @@ with a batch size of 64 and compare the training speed. ## Next Steps -[Congratulations!](https://www.youtube.com/watch?v=9bZkp7q19f0) You have -completed the CIFAR-10 tutorial. - If you are now interested in developing and training your own image classification system, we recommend forking this tutorial and replacing components to address your image classification problem. diff --git a/tensorflow/docs_src/tutorials/image_recognition.md b/tensorflow/docs_src/tutorials/images/image_recognition.md similarity index 99% rename from tensorflow/docs_src/tutorials/image_recognition.md rename to tensorflow/docs_src/tutorials/images/image_recognition.md index 332bcf54f02e6e3c7d805746011dfab642943cfe..d545de73df57a7bc775a83cc1fc41ffa185874c5 100644 --- a/tensorflow/docs_src/tutorials/image_recognition.md +++ b/tensorflow/docs_src/tutorials/images/image_recognition.md @@ -434,7 +434,6 @@ should be able to transfer some of that understanding to solving related problems. One way to perform transfer learning is to remove the final classification layer of the network and extract the [next-to-last layer of the CNN](https://arxiv.org/abs/1310.1531), in this case a 2048 dimensional vector. -There's a guide to doing this @{$image_retraining$in the how-to section}. ## Resources for Learning More @@ -450,7 +449,7 @@ covering them. To find out more about implementing convolutional neural networks, you can jump to the TensorFlow @{$deep_cnn$deep convolutional networks tutorial}, -or start a bit more gently with our @{$layers$MNIST starter tutorial}. +or start a bit more gently with our [Estimator MNIST tutorial](../estimators/cnn.md). Finally, if you want to get up to speed on research in this area, you can read the recent work of all the papers referenced in this tutorial. diff --git a/tensorflow/docs_src/tutorials/index.md b/tensorflow/docs_src/tutorials/index.md deleted file mode 100644 index 6bd3a3a897d9cc11e9172e4ccde6fcad4f075ad1..0000000000000000000000000000000000000000 --- a/tensorflow/docs_src/tutorials/index.md +++ /dev/null @@ -1,59 +0,0 @@ -# Tutorials - - -This section contains tutorials demonstrating how to do specific tasks -in TensorFlow. If you are new to TensorFlow, we recommend reading -[Get Started with TensorFlow](/get_started/). - -## Images - -These tutorials cover different aspects of image recognition: - - * @{$layers$MNIST}, which introduces convolutional neural networks (CNNs) and - demonstrates how to build a CNN in TensorFlow. - * @{$image_recognition}, which introduces the field of image recognition and - uses a pre-trained model (Inception) for recognizing images. - * @{$image_retraining}, which has a wonderfully self-explanatory title. - * @{$deep_cnn}, which demonstrates how to build a small CNN for recognizing - images. This tutorial is aimed at advanced TensorFlow users. - - -## Sequences - -These tutorials focus on machine learning problems dealing with sequence data. - - * @{$recurrent}, which demonstrates how to use a - recurrent neural network to predict the next word in a sentence. - * @{$seq2seq}, which demonstrates how to use a - sequence-to-sequence model to translate text from English to French. - * @{$recurrent_quickdraw} - builds a classification model for drawings, directly from the sequence of - pen strokes. - * @{$audio_recognition}, which shows how to - build a basic speech recognition network. - -## Data representation - -These tutorials demonstrate various data representations that can be used in -TensorFlow. - - * @{$wide}, uses - @{tf.feature_column$feature columns} to feed a variety of data types - to linear model, to solve a classification problem. - * @{$wide_and_deep}, builds on the - above linear model tutorial, adding a deep feed-forward neural network - component and a DNN-compatible data representation. - * @{$word2vec}, which demonstrates how to - create an embedding for words. - * @{$kernel_methods}, - which shows how to improve the quality of a linear model by using explicit - kernel mappings. - -## Non Machine Learning - -Although TensorFlow specializes in machine learning, the core of TensorFlow is -a powerful numeric computation system which you can also use to solve other -kinds of math problems. For example: - - * @{$mandelbrot} - * @{$pdes} diff --git a/tensorflow/docs_src/get_started/basic_classification.md b/tensorflow/docs_src/tutorials/keras/basic_classification.md similarity index 100% rename from tensorflow/docs_src/get_started/basic_classification.md rename to tensorflow/docs_src/tutorials/keras/basic_classification.md diff --git a/tensorflow/docs_src/get_started/basic_regression.md b/tensorflow/docs_src/tutorials/keras/basic_regression.md similarity index 100% rename from tensorflow/docs_src/get_started/basic_regression.md rename to tensorflow/docs_src/tutorials/keras/basic_regression.md diff --git a/tensorflow/docs_src/get_started/basic_text_classification.md b/tensorflow/docs_src/tutorials/keras/basic_text_classification.md similarity index 100% rename from tensorflow/docs_src/get_started/basic_text_classification.md rename to tensorflow/docs_src/tutorials/keras/basic_text_classification.md diff --git a/tensorflow/docs_src/tutorials/keras/index.md b/tensorflow/docs_src/tutorials/keras/index.md new file mode 100644 index 0000000000000000000000000000000000000000..9d42281c8f97fd8930770c0bc30c9bcf1e50fde6 --- /dev/null +++ b/tensorflow/docs_src/tutorials/keras/index.md @@ -0,0 +1,22 @@ +# Learn and use machine learning + +This notebook collection is inspired by the book +*[Deep Learning with Python](https://books.google.com/books?id=Yo3CAQAACAAJ)*. +These tutorials use `tf.keras`, TensorFlow's high-level Python API for building +and training deep learning models. To learn more about using Keras with +TensorFlow, see the [TensorFlow Keras Guide](../../guide/keras). + +Publisher's note: *Deep Learning with Python* introduces the field of deep +learning using the Python language and the powerful Keras library. Written by +Keras creator and Google AI researcher FranƧois Chollet, this book builds your +understanding through intuitive explanations and practical examples. + +To learn about machine learning fundamentals and concepts, consider taking the +[Machine Learning Crash Course](https://developers.google.com/machine-learning/crash-course/). +Additional TensorFlow and machine learning resources are listed in [next steps](../next_steps). + +1. [Basic classification](./basic_classification) +2. [Text classification](./basic_text_classification) +3. [Regression](./basic_regression) +4. [Overfitting and underfitting](./overfit_and_underfit) +5. [Save and restore models](./save_and_restore_models) diff --git a/tensorflow/docs_src/get_started/overfit_and_underfit.md b/tensorflow/docs_src/tutorials/keras/overfit_and_underfit.md similarity index 100% rename from tensorflow/docs_src/get_started/overfit_and_underfit.md rename to tensorflow/docs_src/tutorials/keras/overfit_and_underfit.md diff --git a/tensorflow/docs_src/get_started/save_and_restore_models.md b/tensorflow/docs_src/tutorials/keras/save_and_restore_models.md similarity index 100% rename from tensorflow/docs_src/get_started/save_and_restore_models.md rename to tensorflow/docs_src/tutorials/keras/save_and_restore_models.md diff --git a/tensorflow/docs_src/tutorials/leftnav_files b/tensorflow/docs_src/tutorials/leftnav_files deleted file mode 100644 index 888052428f951fa1a7cbd9c6d35497a056387097..0000000000000000000000000000000000000000 --- a/tensorflow/docs_src/tutorials/leftnav_files +++ /dev/null @@ -1,23 +0,0 @@ -index.md - -### Images -layers.md: MNIST -image_recognition.md: Image Recognition -image_retraining.md: Image Retraining -deep_cnn.md - -### Sequences -recurrent.md -seq2seq.md: Neural Machine Translation -recurrent_quickdraw.md: Drawing Classification -audio_recognition.md - -### Data Representation -wide.md: Linear Models -wide_and_deep.md: Wide & Deep Learning -word2vec.md -kernel_methods.md: Kernel Methods - -### Non-ML -mandelbrot.md -pdes.md diff --git a/tensorflow/docs_src/get_started/next_steps.md b/tensorflow/docs_src/tutorials/next_steps.md similarity index 99% rename from tensorflow/docs_src/get_started/next_steps.md rename to tensorflow/docs_src/tutorials/next_steps.md index 6318a39c6cda86f2cc1ec95232e886181fa38fd8..01c9f7204a7ddae16bcbd9eb5702516a39f8ce4c 100644 --- a/tensorflow/docs_src/get_started/next_steps.md +++ b/tensorflow/docs_src/tutorials/next_steps.md @@ -1,4 +1,4 @@ -# Next Steps +# Next steps ## Learn more about TensorFlow diff --git a/tensorflow/docs_src/tutorials/mandelbrot.md b/tensorflow/docs_src/tutorials/non-ml/mandelbrot.md old mode 100755 new mode 100644 similarity index 100% rename from tensorflow/docs_src/tutorials/mandelbrot.md rename to tensorflow/docs_src/tutorials/non-ml/mandelbrot.md diff --git a/tensorflow/docs_src/tutorials/pdes.md b/tensorflow/docs_src/tutorials/non-ml/pdes.md old mode 100755 new mode 100644 similarity index 98% rename from tensorflow/docs_src/tutorials/pdes.md rename to tensorflow/docs_src/tutorials/non-ml/pdes.md index 425e8d7084e7f2505b7a3013b431345b72b38cf0..b5a0fa834a8a0a51421657180f8c7817c0e3d140 --- a/tensorflow/docs_src/tutorials/pdes.md +++ b/tensorflow/docs_src/tutorials/non-ml/pdes.md @@ -135,7 +135,6 @@ for i in range(1000): DisplayArray(U.eval(), rng=[-0.1, 0.1]) ``` -![jpeg](../images/pde_output_2.jpg) +![jpeg](../../images/pde_output_2.jpg) Look! Ripples! - diff --git a/tensorflow/docs_src/tutorials/kernel_methods.md b/tensorflow/docs_src/tutorials/representation/kernel_methods.md similarity index 99% rename from tensorflow/docs_src/tutorials/kernel_methods.md rename to tensorflow/docs_src/tutorials/representation/kernel_methods.md index 205e2a2d2c1d1008e62ca4c2caf9f1b0895dff1a..f3c232c51155927a4b8e5abdd6e1e04403f8caa4 100644 --- a/tensorflow/docs_src/tutorials/kernel_methods.md +++ b/tensorflow/docs_src/tutorials/representation/kernel_methods.md @@ -27,7 +27,7 @@ TensorFlow will provide support for sparse features at a later release. This tutorial uses [tf.contrib.learn](https://www.tensorflow.org/code/tensorflow/contrib/learn/python/learn) (TensorFlow's high-level Machine Learning API) Estimators for our ML models. -If you are not familiar with this API, [tf.estimator Quickstart](https://www.tensorflow.org/get_started/estimator) +If you are not familiar with this API, The [Estimator guide](../../guide/estimators.md) is a good place to start. We will use the MNIST dataset. The tutorial consists of the following steps: diff --git a/tensorflow/docs_src/tutorials/linear.md b/tensorflow/docs_src/tutorials/representation/linear.md similarity index 95% rename from tensorflow/docs_src/tutorials/linear.md rename to tensorflow/docs_src/tutorials/representation/linear.md index 3f247ade266d2675eac4d0f59a4744daa61f27ea..1b418cf065a141dc46833bb0d3c2048658efc388 100644 --- a/tensorflow/docs_src/tutorials/linear.md +++ b/tensorflow/docs_src/tutorials/representation/linear.md @@ -11,8 +11,9 @@ those tools. It explains: deep learning to get the advantages of both. Read this overview to decide whether the Estimator's linear model tools might -be useful to you. Then do the @{$wide$Linear Models tutorial} to -give it a try. This overview uses code samples from the tutorial, but the +be useful to you. Then work through the +[Estimator wide and deep learning tutorial](https://github.com/tensorflow/models/tree/master/official/wide_deep) +to give it a try. This overview uses code samples from the tutorial, but the tutorial walks through the code in greater detail. To understand this overview it will help to have some familiarity @@ -176,7 +177,7 @@ the name of a `FeatureColumn`. Each key's value is a tensor containing the values of that feature for all data instances. See @{$premade_estimators#input_fn} for a more comprehensive look at input functions, and `input_fn` in the -[linear models tutorial code](https://github.com/tensorflow/models/tree/master/official/wide_deep/wide_deep.py) +[wide and deep learning tutorial](https://github.com/tensorflow/models/tree/master/official/wide_deep) for an example implementation of an input function. The input function is passed to the `train()` and `evaluate()` calls that @@ -234,4 +235,5 @@ e = tf.estimator.DNNLinearCombinedClassifier( dnn_feature_columns=deep_columns, dnn_hidden_units=[100, 50]) ``` -For more information, see the @{$wide_and_deep$Wide and Deep Learning tutorial}. +For more information, see the +[wide and deep learning tutorial](https://github.com/tensorflow/models/tree/master/official/wide_deep). diff --git a/tensorflow/docs_src/tutorials/word2vec.md b/tensorflow/docs_src/tutorials/representation/word2vec.md similarity index 96% rename from tensorflow/docs_src/tutorials/word2vec.md rename to tensorflow/docs_src/tutorials/representation/word2vec.md index 3fe7352bd2383177ca200a0265dee41dba430144..0a1c41c84a3971cb6237e37ccaaa884e53de2aae 100644 --- a/tensorflow/docs_src/tutorials/word2vec.md +++ b/tensorflow/docs_src/tutorials/representation/word2vec.md @@ -23,7 +23,7 @@ straight in, feel free to look at the minimalistic implementation in This basic example contains the code needed to download some data, train on it a bit and visualize the result. Once you get comfortable with reading and running the basic version, you can graduate to -[models/tutorials/embedding/word2vec.py](https://www.tensorflow.org/code/tensorflow_models/tutorials/embedding/word2vec.py) +[models/tutorials/embedding/word2vec.py](https://github.com/tensorflow/models/tree/master/tutorials/embedding/word2vec.py) which is a more serious implementation that showcases some more advanced TensorFlow principles about how to efficiently use threads to move data into a text model, how to checkpoint during training, etc. @@ -341,7 +341,7 @@ t-SNE. Et voila! As expected, words that are similar end up clustering nearby each other. For a more heavyweight implementation of word2vec that showcases more of the advanced features of TensorFlow, see the implementation in -[models/tutorials/embedding/word2vec.py](https://www.tensorflow.org/code/tensorflow_models/tutorials/embedding/word2vec.py). +[models/tutorials/embedding/word2vec.py](https://github.com/tensorflow/models/tree/master/tutorials/embedding/word2vec.py). ## Evaluating Embeddings: Analogical Reasoning @@ -357,7 +357,7 @@ Download the dataset for this task from To see how we do this evaluation, have a look at the `build_eval_graph()` and `eval()` functions in -[models/tutorials/embedding/word2vec.py](https://www.tensorflow.org/code/tensorflow_models/tutorials/embedding/word2vec.py). +[models/tutorials/embedding/word2vec.py](https://github.com/tensorflow/models/tree/master/tutorials/embedding/word2vec.py). The choice of hyperparameters can strongly influence the accuracy on this task. To achieve state-of-the-art performance on this task requires training over a @@ -385,13 +385,13 @@ your model is seriously bottlenecked on input data, you may want to implement a custom data reader for your problem, as described in @{$new_data_formats$New Data Formats}. For the case of Skip-Gram modeling, we've actually already done this for you as an example in -[models/tutorials/embedding/word2vec.py](https://www.tensorflow.org/code/tensorflow_models/tutorials/embedding/word2vec.py). +[models/tutorials/embedding/word2vec.py](https://github.com/tensorflow/models/tree/master/tutorials/embedding/word2vec.py). If your model is no longer I/O bound but you want still more performance, you can take things further by writing your own TensorFlow Ops, as described in @{$adding_an_op$Adding a New Op}. Again we've provided an example of this for the Skip-Gram case -[models/tutorials/embedding/word2vec_optimized.py](https://www.tensorflow.org/code/tensorflow_models/tutorials/embedding/word2vec_optimized.py). +[models/tutorials/embedding/word2vec_optimized.py](https://github.com/tensorflow/models/tree/master/tutorials/embedding/word2vec_optimized.py). Feel free to benchmark these against each other to measure performance improvements at each stage. diff --git a/tensorflow/docs_src/tutorials/seq2seq.md b/tensorflow/docs_src/tutorials/seq2seq.md deleted file mode 100644 index 8928ba4f7da26ae2e8e9351e2c7c03f0e657f613..0000000000000000000000000000000000000000 --- a/tensorflow/docs_src/tutorials/seq2seq.md +++ /dev/null @@ -1,5 +0,0 @@ -# Sequence-to-Sequence Models - -Please check out the -[tensorflow neural machine translation tutorial](https://github.com/tensorflow/nmt) -for building sequence-to-sequence models with the latest Tensorflow API. diff --git a/tensorflow/docs_src/tutorials/audio_recognition.md b/tensorflow/docs_src/tutorials/sequences/audio_recognition.md similarity index 100% rename from tensorflow/docs_src/tutorials/audio_recognition.md rename to tensorflow/docs_src/tutorials/sequences/audio_recognition.md diff --git a/tensorflow/docs_src/tutorials/recurrent.md b/tensorflow/docs_src/tutorials/sequences/recurrent.md similarity index 98% rename from tensorflow/docs_src/tutorials/recurrent.md rename to tensorflow/docs_src/tutorials/sequences/recurrent.md index 14da2c8785276abb34d6959d738f5b39e6c6a2e8..715cc7856af1d6a3422b65a796a3d48b6c1c3e0f 100644 --- a/tensorflow/docs_src/tutorials/recurrent.md +++ b/tensorflow/docs_src/tutorials/sequences/recurrent.md @@ -2,8 +2,8 @@ ## Introduction -Take a look at [this great article](https://colah.github.io/posts/2015-08-Understanding-LSTMs/) -for an introduction to recurrent neural networks and LSTMs in particular. +See [Understanding LSTM Networks](https://colah.github.io/posts/2015-08-Understanding-LSTMs/){:.external} +for an introduction to recurrent neural networks and LSTMs. ## Language Modeling diff --git a/tensorflow/docs_src/tutorials/recurrent_quickdraw.md b/tensorflow/docs_src/tutorials/sequences/recurrent_quickdraw.md similarity index 98% rename from tensorflow/docs_src/tutorials/recurrent_quickdraw.md rename to tensorflow/docs_src/tutorials/sequences/recurrent_quickdraw.md index 1afd861738512f20de5171548d539d256f5f5225..37bce5b76d46741dfe04cbf3612f71863adb02c6 100644 --- a/tensorflow/docs_src/tutorials/recurrent_quickdraw.md +++ b/tensorflow/docs_src/tutorials/sequences/recurrent_quickdraw.md @@ -13,7 +13,7 @@ In this tutorial we'll show how to build an RNN-based recognizer for this problem. The model will use a combination of convolutional layers, LSTM layers, and a softmax output layer to classify the drawings: -

![RNN model structure](../images/quickdraw_model.png)
+
![RNN model structure](../../images/quickdraw_model.png)
The figure above shows the structure of the model that we will build in this tutorial. The input is a drawing that is encoded as a sequence of strokes of @@ -208,7 +208,7 @@ This data is then reformatted into a tensor of shape `[num_training_samples, max_length, 3]`. Then we determine the bounding box of the original drawing in screen coordinates and normalize the size such that the drawing has unit height. -
![Size normalization](../images/quickdraw_sizenormalization.png)
+
![Size normalization](../../images/quickdraw_sizenormalization.png)
Finally, we compute the differences between consecutive points and store these as a `VarLenFeature` in a diff --git a/tensorflow/docs_src/tutorials/wide.md b/tensorflow/docs_src/tutorials/wide.md deleted file mode 100644 index 27ce75a30dd2acd5925702611042270e767b0c73..0000000000000000000000000000000000000000 --- a/tensorflow/docs_src/tutorials/wide.md +++ /dev/null @@ -1,461 +0,0 @@ -# TensorFlow Linear Model Tutorial - -In this tutorial, we will use the tf.estimator API in TensorFlow to solve a -binary classification problem: Given census data about a person such as age, -education, marital status, and occupation (the features), we will try to predict -whether or not the person earns more than 50,000 dollars a year (the target -label). We will train a **logistic regression** model, and given an individual's -information our model will output a number between 0 and 1, which can be -interpreted as the probability that the individual has an annual income of over -50,000 dollars. - -## Setup - -To try the code for this tutorial: - -1. @{$install$Install TensorFlow} if you haven't already. - -2. Download [the tutorial code](https://github.com/tensorflow/models/tree/master/official/wide_deep/). - -3. Execute the data download script we provide to you: - - $ python data_download.py - -4. Execute the tutorial code with the following command to train the linear -model described in this tutorial: - - $ python wide_deep.py --model_type=wide - -Read on to find out how this code builds its linear model. - -## Reading The Census Data - -The dataset we'll be using is the -[Census Income Dataset](https://archive.ics.uci.edu/ml/datasets/Census+Income). -We have provided -[data_download.py](https://github.com/tensorflow/models/tree/master/official/wide_deep/data_download.py) -which downloads the code and performs some additional cleanup. - -Since the task is a binary classification problem, we'll construct a label -column named "label" whose value is 1 if the income is over 50K, and 0 -otherwise. For reference, see `input_fn` in -[wide_deep.py](https://github.com/tensorflow/models/tree/master/official/wide_deep/wide_deep.py). - -Next, let's take a look at the dataframe and see which columns we can use to -predict the target label. The columns can be grouped into two typesā€”categorical -and continuous columns: - -* A column is called **categorical** if its value can only be one of the - categories in a finite set. For example, the relationship status of a person - (wife, husband, unmarried, etc.) or the education level (high school, - college, etc.) are categorical columns. -* A column is called **continuous** if its value can be any numerical value in - a continuous range. For example, the capital gain of a person (e.g. $14,084) - is a continuous column. - -Here's a list of columns available in the Census Income dataset: - -| Column Name | Type | Description | -| -------------- | ----------- | --------------------------------- | -| age | Continuous | The age of the individual | -| workclass | Categorical | The type of employer the | -: : : individual has (government, : -: : : military, private, etc.). : -| fnlwgt | Continuous | The number of people the census | -: : : takers believe that observation : -: : : represents (sample weight). Final : -: : : weight will not be used. : -| education | Categorical | The highest level of education | -: : : achieved for that individual. : -| education_num | Continuous | The highest level of education in | -: : : numerical form. : -| marital_status | Categorical | Marital status of the individual. | -| occupation | Categorical | The occupation of the individual. | -| relationship | Categorical | Wife, Own-child, Husband, | -: : : Not-in-family, Other-relative, : -: : : Unmarried. : -| race | Categorical | Amer-Indian-Eskimo, Asian-Pac- | -: : : Islander, Black, White, Other. : -| gender | Categorical | Female, Male. | -| capital_gain | Continuous | Capital gains recorded. | -| capital_loss | Continuous | Capital Losses recorded. | -| hours_per_week | Continuous | Hours worked per week. | -| native_country | Categorical | Country of origin of the | -: : : individual. : -| income_bracket | Categorical | ">50K" or "<=50K", meaning | -: : : whether the person makes more : -: : : than $50,000 annually. : - -## Converting Data into Tensors - -When building a tf.estimator model, the input data is specified by means of an -Input Builder function. This builder function will not be called until it is -later passed to tf.estimator.Estimator methods such as `train` and `evaluate`. -The purpose of this function is to construct the input data, which is -represented in the form of @{tf.Tensor}s or @{tf.SparseTensor}s. -In more detail, the input builder function returns the following as a pair: - -1. `features`: A dict from feature column names to `Tensors` or - `SparseTensors`. -2. `labels`: A `Tensor` containing the label column. - -The keys of the `features` will be used to construct columns in the next -section. Because we want to call the `train` and `evaluate` methods with -different data, we define a method that returns an input function based on the -given data. Note that the returned input function will be called while -constructing the TensorFlow graph, not while running the graph. What it is -returning is a representation of the input data as the fundamental unit of -TensorFlow computations, a `Tensor` (or `SparseTensor`). - -Each continuous column in the train or test data will be converted into a -`Tensor`, which in general is a good format to represent dense data. For -categorical data, we must represent the data as a `SparseTensor`. This data -format is good for representing sparse data. Our `input_fn` uses the `tf.data` -API, which makes it easy to apply transformations to our dataset: - -```python -def input_fn(data_file, num_epochs, shuffle, batch_size): - """Generate an input function for the Estimator.""" - assert tf.gfile.Exists(data_file), ( - '%s not found. Please make sure you have either run data_download.py or ' - 'set both arguments --train_data and --test_data.' % data_file) - - def parse_csv(value): - print('Parsing', data_file) - columns = tf.decode_csv(value, record_defaults=_CSV_COLUMN_DEFAULTS) - features = dict(zip(_CSV_COLUMNS, columns)) - labels = features.pop('income_bracket') - return features, tf.equal(labels, '>50K') - - # Extract lines from input files using the Dataset API. - dataset = tf.data.TextLineDataset(data_file) - - if shuffle: - dataset = dataset.shuffle(buffer_size=_SHUFFLE_BUFFER) - - dataset = dataset.map(parse_csv, num_parallel_calls=5) - - # We call repeat after shuffling, rather than before, to prevent separate - # epochs from blending together. - dataset = dataset.repeat(num_epochs) - dataset = dataset.batch(batch_size) - - iterator = dataset.make_one_shot_iterator() - features, labels = iterator.get_next() - return features, labels -``` - -## Selecting and Engineering Features for the Model - -Selecting and crafting the right set of feature columns is key to learning an -effective model. A **feature column** can be either one of the raw columns in -the original dataframe (let's call them **base feature columns**), or any new -columns created based on some transformations defined over one or multiple base -columns (let's call them **derived feature columns**). Basically, "feature -column" is an abstract concept of any raw or derived variable that can be used -to predict the target label. - -### Base Categorical Feature Columns - -To define a feature column for a categorical feature, we can create a -`CategoricalColumn` using the tf.feature_column API. If you know the set of all -possible feature values of a column and there are only a few of them, you can -use `categorical_column_with_vocabulary_list`. Each key in the list will get -assigned an auto-incremental ID starting from 0. For example, for the -`relationship` column we can assign the feature string "Husband" to an integer -ID of 0 and "Not-in-family" to 1, etc., by doing: - -```python -relationship = tf.feature_column.categorical_column_with_vocabulary_list( - 'relationship', [ - 'Husband', 'Not-in-family', 'Wife', 'Own-child', 'Unmarried', - 'Other-relative']) -``` - -What if we don't know the set of possible values in advance? Not a problem. We -can use `categorical_column_with_hash_bucket` instead: - -```python -occupation = tf.feature_column.categorical_column_with_hash_bucket( - 'occupation', hash_bucket_size=1000) -``` - -What will happen is that each possible value in the feature column `occupation` -will be hashed to an integer ID as we encounter them in training. See an example -illustration below: - -ID | Feature ---- | ------------- -... | -9 | `"Machine-op-inspct"` -... | -103 | `"Farming-fishing"` -... | -375 | `"Protective-serv"` -... | - -No matter which way we choose to define a `SparseColumn`, each feature string -will be mapped into an integer ID by looking up a fixed mapping or by hashing. -Note that hashing collisions are possible, but may not significantly impact the -model quality. Under the hood, the `LinearModel` class is responsible for -managing the mapping and creating `tf.Variable` to store the model parameters -(also known as model weights) for each feature ID. The model parameters will be -learned through the model training process we'll go through later. - -We'll do the similar trick to define the other categorical features: - -```python -education = tf.feature_column.categorical_column_with_vocabulary_list( - 'education', [ - 'Bachelors', 'HS-grad', '11th', 'Masters', '9th', 'Some-college', - 'Assoc-acdm', 'Assoc-voc', '7th-8th', 'Doctorate', 'Prof-school', - '5th-6th', '10th', '1st-4th', 'Preschool', '12th']) - -marital_status = tf.feature_column.categorical_column_with_vocabulary_list( - 'marital_status', [ - 'Married-civ-spouse', 'Divorced', 'Married-spouse-absent', - 'Never-married', 'Separated', 'Married-AF-spouse', 'Widowed']) - -relationship = tf.feature_column.categorical_column_with_vocabulary_list( - 'relationship', [ - 'Husband', 'Not-in-family', 'Wife', 'Own-child', 'Unmarried', - 'Other-relative']) - -workclass = tf.feature_column.categorical_column_with_vocabulary_list( - 'workclass', [ - 'Self-emp-not-inc', 'Private', 'State-gov', 'Federal-gov', - 'Local-gov', '?', 'Self-emp-inc', 'Without-pay', 'Never-worked']) - -# To show an example of hashing: -occupation = tf.feature_column.categorical_column_with_hash_bucket( - 'occupation', hash_bucket_size=1000) -``` - -### Base Continuous Feature Columns - -Similarly, we can define a `NumericColumn` for each continuous feature column -that we want to use in the model: - -```python -age = tf.feature_column.numeric_column('age') -education_num = tf.feature_column.numeric_column('education_num') -capital_gain = tf.feature_column.numeric_column('capital_gain') -capital_loss = tf.feature_column.numeric_column('capital_loss') -hours_per_week = tf.feature_column.numeric_column('hours_per_week') -``` - -### Making Continuous Features Categorical through Bucketization - -Sometimes the relationship between a continuous feature and the label is not -linear. As a hypothetical example, a person's income may grow with age in the -early stage of one's career, then the growth may slow at some point, and finally -the income decreases after retirement. In this scenario, using the raw `age` as -a real-valued feature column might not be a good choice because the model can -only learn one of the three cases: - -1. Income always increases at some rate as age grows (positive correlation), -1. Income always decreases at some rate as age grows (negative correlation), or -1. Income stays the same no matter at what age (no correlation) - -If we want to learn the fine-grained correlation between income and each age -group separately, we can leverage **bucketization**. Bucketization is a process -of dividing the entire range of a continuous feature into a set of consecutive -bins/buckets, and then converting the original numerical feature into a bucket -ID (as a categorical feature) depending on which bucket that value falls into. -So, we can define a `bucketized_column` over `age` as: - -```python -age_buckets = tf.feature_column.bucketized_column( - age, boundaries=[18, 25, 30, 35, 40, 45, 50, 55, 60, 65]) -``` - -where the `boundaries` is a list of bucket boundaries. In this case, there are -10 boundaries, resulting in 11 age group buckets (from age 17 and below, 18-24, -25-29, ..., to 65 and over). - -### Intersecting Multiple Columns with CrossedColumn - -Using each base feature column separately may not be enough to explain the data. -For example, the correlation between education and the label (earning > 50,000 -dollars) may be different for different occupations. Therefore, if we only learn -a single model weight for `education="Bachelors"` and `education="Masters"`, we -won't be able to capture every single education-occupation combination (e.g. -distinguishing between `education="Bachelors" AND occupation="Exec-managerial"` -and `education="Bachelors" AND occupation="Craft-repair"`). To learn the -differences between different feature combinations, we can add **crossed feature -columns** to the model. - -```python -education_x_occupation = tf.feature_column.crossed_column( - ['education', 'occupation'], hash_bucket_size=1000) -``` - -We can also create a `CrossedColumn` over more than two columns. Each -constituent column can be either a base feature column that is categorical -(`SparseColumn`), a bucketized real-valued feature column (`BucketizedColumn`), -or even another `CrossColumn`. Here's an example: - -```python -age_buckets_x_education_x_occupation = tf.feature_column.crossed_column( - [age_buckets, 'education', 'occupation'], hash_bucket_size=1000) -``` - -## Defining The Logistic Regression Model - -After processing the input data and defining all the feature columns, we're now -ready to put them all together and build a Logistic Regression model. In the -previous section we've seen several types of base and derived feature columns, -including: - -* `CategoricalColumn` -* `NumericColumn` -* `BucketizedColumn` -* `CrossedColumn` - -All of these are subclasses of the abstract `FeatureColumn` class, and can be -added to the `feature_columns` field of a model: - -```python -base_columns = [ - education, marital_status, relationship, workclass, occupation, - age_buckets, -] -crossed_columns = [ - tf.feature_column.crossed_column( - ['education', 'occupation'], hash_bucket_size=1000), - tf.feature_column.crossed_column( - [age_buckets, 'education', 'occupation'], hash_bucket_size=1000), -] - -model_dir = tempfile.mkdtemp() -model = tf.estimator.LinearClassifier( - model_dir=model_dir, feature_columns=base_columns + crossed_columns) -``` - -The model also automatically learns a bias term, which controls the prediction -one would make without observing any features (see the section "How Logistic -Regression Works" for more explanations). The learned model files will be stored -in `model_dir`. - -## Training and Evaluating Our Model - -After adding all the features to the model, now let's look at how to actually -train the model. Training a model is just a single command using the -tf.estimator API: - -```python -model.train(input_fn=lambda: input_fn(train_data, num_epochs, True, batch_size)) -``` - -After the model is trained, we can evaluate how good our model is at predicting -the labels of the holdout data: - -```python -results = model.evaluate(input_fn=lambda: input_fn( - test_data, 1, False, batch_size)) -for key in sorted(results): - print('%s: %s' % (key, results[key])) -``` - -The first line of the final output should be something like -`accuracy: 0.83557522`, which means the accuracy is 83.6%. Feel free to try more -features and transformations and see if you can do even better! - -After the model is evaluated, we can use the model to predict whether an individual has an annual income of over -50,000 dollars given an individual's information input. -```python - pred_iter = model.predict(input_fn=lambda: input_fn(FLAGS.test_data, 1, False, 1)) - for pred in pred_iter: - print(pred['classes']) -``` - -The model prediction output would be like `[b'1']` or `[b'0']` which means whether corresponding individual has an annual income of over 50,000 dollars or not. - -If you'd like to see a working end-to-end example, you can download our -[example code](https://github.com/tensorflow/models/tree/master/official/wide_deep/wide_deep.py) -and set the `model_type` flag to `wide`. - -## Adding Regularization to Prevent Overfitting - -Regularization is a technique used to avoid **overfitting**. Overfitting happens -when your model does well on the data it is trained on, but worse on test data -that the model has not seen before, such as live traffic. Overfitting generally -occurs when a model is excessively complex, such as having too many parameters -relative to the number of observed training data. Regularization allows for you -to control your model's complexity and makes the model more generalizable to -unseen data. - -In the Linear Model library, you can add L1 and L2 regularizations to the model -as: - -``` -model = tf.estimator.LinearClassifier( - model_dir=model_dir, feature_columns=base_columns + crossed_columns, - optimizer=tf.train.FtrlOptimizer( - learning_rate=0.1, - l1_regularization_strength=1.0, - l2_regularization_strength=1.0)) -``` - -One important difference between L1 and L2 regularization is that L1 -regularization tends to make model weights stay at zero, creating sparser -models, whereas L2 regularization also tries to make the model weights closer to -zero but not necessarily zero. Therefore, if you increase the strength of L1 -regularization, you will have a smaller model size because many of the model -weights will be zero. This is often desirable when the feature space is very -large but sparse, and when there are resource constraints that prevent you from -serving a model that is too large. - -In practice, you should try various combinations of L1, L2 regularization -strengths and find the best parameters that best control overfitting and give -you a desirable model size. - -## How Logistic Regression Works - -Finally, let's take a minute to talk about what the Logistic Regression model -actually looks like in case you're not already familiar with it. We'll denote -the label as \\(Y\\), and the set of observed features as a feature vector -\\(\mathbf{x}=[x_1, x_2, ..., x_d]\\). We define \\(Y=1\\) if an individual -earned > 50,000 dollars and \\(Y=0\\) otherwise. In Logistic Regression, the -probability of the label being positive (\\(Y=1\\)) given the features -\\(\mathbf{x}\\) is given as: - -$$ P(Y=1|\mathbf{x}) = \frac{1}{1+\exp(-(\mathbf{w}^T\mathbf{x}+b))}$$ - -where \\(\mathbf{w}=[w_1, w_2, ..., w_d]\\) are the model weights for the -features \\(\mathbf{x}=[x_1, x_2, ..., x_d]\\). \\(b\\) is a constant that is -often called the **bias** of the model. The equation consists of two partsā€”A -linear model and a logistic function: - -* **Linear Model**: First, we can see that \\(\mathbf{w}^T\mathbf{x}+b = b + - w_1x_1 + ... +w_dx_d\\) is a linear model where the output is a linear - function of the input features \\(\mathbf{x}\\). The bias \\(b\\) is the - prediction one would make without observing any features. The model weight - \\(w_i\\) reflects how the feature \\(x_i\\) is correlated with the positive - label. If \\(x_i\\) is positively correlated with the positive label, the - weight \\(w_i\\) increases, and the probability \\(P(Y=1|\mathbf{x})\\) will - be closer to 1. On the other hand, if \\(x_i\\) is negatively correlated - with the positive label, then the weight \\(w_i\\) decreases and the - probability \\(P(Y=1|\mathbf{x})\\) will be closer to 0. - -* **Logistic Function**: Second, we can see that there's a logistic function - (also known as the sigmoid function) \\(S(t) = 1/(1+\exp(-t))\\) being - applied to the linear model. The logistic function is used to convert the - output of the linear model \\(\mathbf{w}^T\mathbf{x}+b\\) from any real - number into the range of \\([0, 1]\\), which can be interpreted as a - probability. - -Model training is an optimization problem: The goal is to find a set of model -weights (i.e. model parameters) to minimize a **loss function** defined over the -training data, such as logistic loss for Logistic Regression models. The loss -function measures the discrepancy between the ground-truth label and the model's -prediction. If the prediction is very close to the ground-truth label, the loss -value will be low; if the prediction is very far from the label, then the loss -value would be high. - -## Learn Deeper - -If you're interested in learning more, check out our -@{$wide_and_deep$Wide & Deep Learning Tutorial} where we'll show you how to -combine the strengths of linear models and deep neural networks by jointly -training them using the tf.estimator API. diff --git a/tensorflow/docs_src/tutorials/wide_and_deep.md b/tensorflow/docs_src/tutorials/wide_and_deep.md deleted file mode 100644 index 44677a810bc5c253c198d81fae2be723c4f8ae4e..0000000000000000000000000000000000000000 --- a/tensorflow/docs_src/tutorials/wide_and_deep.md +++ /dev/null @@ -1,243 +0,0 @@ -# TensorFlow Wide & Deep Learning Tutorial - -In the previous @{$wide$TensorFlow Linear Model Tutorial}, we trained a logistic -regression model to predict the probability that the individual has an annual -income of over 50,000 dollars using the -[Census Income Dataset](https://archive.ics.uci.edu/ml/datasets/Census+Income). -TensorFlow is great for training deep neural networks too, and you might be -thinking which one you should chooseā€”well, why not both? Would it be possible to -combine the strengths of both in one model? - -In this tutorial, we'll introduce how to use the tf.estimator API to jointly -train a wide linear model and a deep feed-forward neural network. This approach -combines the strengths of memorization and generalization. It's useful for -generic large-scale regression and classification problems with sparse input -features (e.g., categorical features with a large number of possible feature -values). If you're interested in learning more about how Wide & Deep Learning -works, please check out our [research paper](https://arxiv.org/abs/1606.07792). - -![Wide & Deep Spectrum of Models](https://www.tensorflow.org/images/wide_n_deep.svg "Wide & Deep") - -The figure above shows a comparison of a wide model (logistic regression with -sparse features and transformations), a deep model (feed-forward neural network -with an embedding layer and several hidden layers), and a Wide & Deep model -(joint training of both). At a high level, there are only 3 steps to configure a -wide, deep, or Wide & Deep model using the tf.estimator API: - -1. Select features for the wide part: Choose the sparse base columns and - crossed columns you want to use. -1. Select features for the deep part: Choose the continuous columns, the - embedding dimension for each categorical column, and the hidden layer sizes. -1. Put them all together in a Wide & Deep model - (`DNNLinearCombinedClassifier`). - -And that's it! Let's go through a simple example. - -## Setup - -To try the code for this tutorial: - -1. @{$install$Install TensorFlow} if you haven't already. - -2. Download [the tutorial code](https://github.com/tensorflow/models/tree/master/official/wide_deep/). - -3. Execute the data download script we provide to you: - - $ python data_download.py - -4. Execute the tutorial code with the following command to train the wide and -deep model described in this tutorial: - - $ python wide_deep.py - -Read on to find out how this code builds its model. - - -## Define Base Feature Columns - -First, let's define the base categorical and continuous feature columns that -we'll use. These base columns will be the building blocks used by both the wide -part and the deep part of the model. - -```python -import tensorflow as tf - -# Continuous columns -age = tf.feature_column.numeric_column('age') -education_num = tf.feature_column.numeric_column('education_num') -capital_gain = tf.feature_column.numeric_column('capital_gain') -capital_loss = tf.feature_column.numeric_column('capital_loss') -hours_per_week = tf.feature_column.numeric_column('hours_per_week') - -education = tf.feature_column.categorical_column_with_vocabulary_list( - 'education', [ - 'Bachelors', 'HS-grad', '11th', 'Masters', '9th', 'Some-college', - 'Assoc-acdm', 'Assoc-voc', '7th-8th', 'Doctorate', 'Prof-school', - '5th-6th', '10th', '1st-4th', 'Preschool', '12th']) - -marital_status = tf.feature_column.categorical_column_with_vocabulary_list( - 'marital_status', [ - 'Married-civ-spouse', 'Divorced', 'Married-spouse-absent', - 'Never-married', 'Separated', 'Married-AF-spouse', 'Widowed']) - -relationship = tf.feature_column.categorical_column_with_vocabulary_list( - 'relationship', [ - 'Husband', 'Not-in-family', 'Wife', 'Own-child', 'Unmarried', - 'Other-relative']) - -workclass = tf.feature_column.categorical_column_with_vocabulary_list( - 'workclass', [ - 'Self-emp-not-inc', 'Private', 'State-gov', 'Federal-gov', - 'Local-gov', '?', 'Self-emp-inc', 'Without-pay', 'Never-worked']) - -# To show an example of hashing: -occupation = tf.feature_column.categorical_column_with_hash_bucket( - 'occupation', hash_bucket_size=1000) - -# Transformations. -age_buckets = tf.feature_column.bucketized_column( - age, boundaries=[18, 25, 30, 35, 40, 45, 50, 55, 60, 65]) -``` - -## The Wide Model: Linear Model with Crossed Feature Columns - -The wide model is a linear model with a wide set of sparse and crossed feature -columns: - -```python -base_columns = [ - education, marital_status, relationship, workclass, occupation, - age_buckets, -] - -crossed_columns = [ - tf.feature_column.crossed_column( - ['education', 'occupation'], hash_bucket_size=1000), - tf.feature_column.crossed_column( - [age_buckets, 'education', 'occupation'], hash_bucket_size=1000), -] -``` - -You can also see the @{$wide$TensorFlow Linear Model Tutorial} for more details. - -Wide models with crossed feature columns can memorize sparse interactions -between features effectively. That being said, one limitation of crossed feature -columns is that they do not generalize to feature combinations that have not -appeared in the training data. Let's add a deep model with embeddings to fix -that. - -## The Deep Model: Neural Network with Embeddings - -The deep model is a feed-forward neural network, as shown in the previous -figure. Each of the sparse, high-dimensional categorical features are first -converted into a low-dimensional and dense real-valued vector, often referred to -as an embedding vector. These low-dimensional dense embedding vectors are -concatenated with the continuous features, and then fed into the hidden layers -of a neural network in the forward pass. The embedding values are initialized -randomly, and are trained along with all other model parameters to minimize the -training loss. If you're interested in learning more about embeddings, check out -the TensorFlow tutorial on @{$word2vec$Vector Representations of Words} or -[Word embedding](https://en.wikipedia.org/wiki/Word_embedding) on Wikipedia. - -Another way to represent categorical columns to feed into a neural network is -via a one-hot or multi-hot representation. This is often appropriate for -categorical columns with only a few possible values. As an example of a one-hot -representation, for the relationship column, `"Husband"` can be represented as -[1, 0, 0, 0, 0, 0], and `"Not-in-family"` as [0, 1, 0, 0, 0, 0], etc. This is a -fixed representation, whereas embeddings are more flexible and calculated at -training time. - -We'll configure the embeddings for the categorical columns using -`embedding_column`, and concatenate them with the continuous columns. -We also use `indicator_column` to create multi-hot representations of some -categorical columns. - -```python -deep_columns = [ - age, - education_num, - capital_gain, - capital_loss, - hours_per_week, - tf.feature_column.indicator_column(workclass), - tf.feature_column.indicator_column(education), - tf.feature_column.indicator_column(marital_status), - tf.feature_column.indicator_column(relationship), - # To show an example of embedding - tf.feature_column.embedding_column(occupation, dimension=8), -] -``` - -The higher the `dimension` of the embedding is, the more degrees of freedom the -model will have to learn the representations of the features. For simplicity, we -set the dimension to 8 for all feature columns here. Empirically, a more -informed decision for the number of dimensions is to start with a value on the -order of \\(\log_2(n)\\) or \\(k\sqrt[4]n\\), where \\(n\\) is the number of -unique features in a feature column and \\(k\\) is a small constant (usually -smaller than 10). - -Through dense embeddings, deep models can generalize better and make predictions -on feature pairs that were previously unseen in the training data. However, it -is difficult to learn effective low-dimensional representations for feature -columns when the underlying interaction matrix between two feature columns is -sparse and high-rank. In such cases, the interaction between most feature pairs -should be zero except a few, but dense embeddings will lead to nonzero -predictions for all feature pairs, and thus can over-generalize. On the other -hand, linear models with crossed features can memorize these ā€œexception rulesā€ -effectively with fewer model parameters. - -Now, let's see how to jointly train wide and deep models and allow them to -complement each otherā€™s strengths and weaknesses. - -## Combining Wide and Deep Models into One - -The wide models and deep models are combined by summing up their final output -log odds as the prediction, then feeding the prediction to a logistic loss -function. All the graph definition and variable allocations have already been -handled for you under the hood, so you simply need to create a -`DNNLinearCombinedClassifier`: - -```python -model = tf.estimator.DNNLinearCombinedClassifier( - model_dir='/tmp/census_model', - linear_feature_columns=base_columns + crossed_columns, - dnn_feature_columns=deep_columns, - dnn_hidden_units=[100, 50]) -``` - -## Training and Evaluating The Model - -Before we train the model, let's read in the Census dataset as we did in the -@{$wide$TensorFlow Linear Model tutorial}. See `data_download.py` as well as -`input_fn` within -[`wide_deep.py`](https://github.com/tensorflow/models/tree/master/official/wide_deep/wide_deep.py). - -After reading in the data, you can train and evaluate the model: - -```python -# Train and evaluate the model every `FLAGS.epochs_per_eval` epochs. -for n in range(FLAGS.train_epochs // FLAGS.epochs_per_eval): - model.train(input_fn=lambda: input_fn( - FLAGS.train_data, FLAGS.epochs_per_eval, True, FLAGS.batch_size)) - - results = model.evaluate(input_fn=lambda: input_fn( - FLAGS.test_data, 1, False, FLAGS.batch_size)) - - # Display evaluation metrics - print('Results at epoch', (n + 1) * FLAGS.epochs_per_eval) - print('-' * 30) - - for key in sorted(results): - print('%s: %s' % (key, results[key])) -``` - -The final output accuracy should be somewhere around 85.5%. If you'd like to -see a working end-to-end example, you can download our -[example code](https://github.com/tensorflow/models/tree/master/official/wide_deep/wide_deep.py). - -Note that this tutorial is just a quick example on a small dataset to get you -familiar with the API. Wide & Deep Learning will be even more powerful if you -try it on a large dataset with many sparse feature columns that have a large -number of possible feature values. Again, feel free to take a look at our -[research paper](https://arxiv.org/abs/1606.07792) for more ideas about how to -apply Wide & Deep Learning in real-world large-scale machine learning problems. diff --git a/tensorflow/examples/speech_commands/BUILD b/tensorflow/examples/speech_commands/BUILD index 13bca34a86b0c2fba7e5e8e3527d13587feacaae..7a44e2ee4fdf690ce576f720bb371785f88779b4 100644 --- a/tensorflow/examples/speech_commands/BUILD +++ b/tensorflow/examples/speech_commands/BUILD @@ -56,6 +56,7 @@ tf_py_test( srcs = ["input_data_test.py"], additional_deps = [ ":input_data", + ":models", "//tensorflow/python:client_testlib", ], ) diff --git a/tensorflow/examples/speech_commands/freeze.py b/tensorflow/examples/speech_commands/freeze.py index c8671d9c41169c07ce3134a49bf81a4ac29a8c60..7657b23c600035af2d230142f3ffc51a6836adc9 100644 --- a/tensorflow/examples/speech_commands/freeze.py +++ b/tensorflow/examples/speech_commands/freeze.py @@ -54,7 +54,7 @@ FLAGS = None def create_inference_graph(wanted_words, sample_rate, clip_duration_ms, clip_stride_ms, window_size_ms, window_stride_ms, - dct_coefficient_count, model_architecture): + feature_bin_count, model_architecture, preprocess): """Creates an audio model with the nodes needed for inference. Uses the supplied arguments to create a model, and inserts the input and @@ -67,14 +67,19 @@ def create_inference_graph(wanted_words, sample_rate, clip_duration_ms, clip_stride_ms: How often to run recognition. Useful for models with cache. window_size_ms: Time slice duration to estimate frequencies from. window_stride_ms: How far apart time slices should be. - dct_coefficient_count: Number of frequency bands to analyze. + feature_bin_count: Number of frequency bands to analyze. model_architecture: Name of the kind of model to generate. + preprocess: How the spectrogram is processed to produce features, for + example 'mfcc' or 'average'. + + Raises: + Exception: If the preprocessing mode isn't recognized. """ words_list = input_data.prepare_words_list(wanted_words.split(',')) model_settings = models.prepare_model_settings( len(words_list), sample_rate, clip_duration_ms, window_size_ms, - window_stride_ms, dct_coefficient_count) + window_stride_ms, feature_bin_count, preprocess) runtime_settings = {'clip_stride_ms': clip_stride_ms} wav_data_placeholder = tf.placeholder(tf.string, [], name='wav_data') @@ -88,15 +93,25 @@ def create_inference_graph(wanted_words, sample_rate, clip_duration_ms, window_size=model_settings['window_size_samples'], stride=model_settings['window_stride_samples'], magnitude_squared=True) - fingerprint_input = contrib_audio.mfcc( - spectrogram, - decoded_sample_data.sample_rate, - dct_coefficient_count=dct_coefficient_count) - fingerprint_frequency_size = model_settings['dct_coefficient_count'] - fingerprint_time_size = model_settings['spectrogram_length'] - reshaped_input = tf.reshape(fingerprint_input, [ - -1, fingerprint_time_size * fingerprint_frequency_size - ]) + + if preprocess == 'average': + fingerprint_input = tf.nn.pool( + tf.expand_dims(spectrogram, -1), + window_shape=[1, model_settings['average_window_width']], + strides=[1, model_settings['average_window_width']], + pooling_type='AVG', + padding='SAME') + elif preprocess == 'mfcc': + fingerprint_input = contrib_audio.mfcc( + spectrogram, + sample_rate, + dct_coefficient_count=model_settings['fingerprint_width']) + else: + raise Exception('Unknown preprocess mode "%s" (should be "mfcc" or' + ' "average")' % (preprocess)) + + fingerprint_size = model_settings['fingerprint_size'] + reshaped_input = tf.reshape(fingerprint_input, [-1, fingerprint_size]) logits = models.create_model( reshaped_input, model_settings, model_architecture, is_training=False, @@ -110,10 +125,12 @@ def main(_): # Create the model and load its weights. sess = tf.InteractiveSession() - create_inference_graph(FLAGS.wanted_words, FLAGS.sample_rate, - FLAGS.clip_duration_ms, FLAGS.clip_stride_ms, - FLAGS.window_size_ms, FLAGS.window_stride_ms, - FLAGS.dct_coefficient_count, FLAGS.model_architecture) + create_inference_graph( + FLAGS.wanted_words, FLAGS.sample_rate, FLAGS.clip_duration_ms, + FLAGS.clip_stride_ms, FLAGS.window_size_ms, FLAGS.window_stride_ms, + FLAGS.feature_bin_count, FLAGS.model_architecture, FLAGS.preprocess) + if FLAGS.quantize: + tf.contrib.quantize.create_training_graph(quant_delay=0) models.load_variables_from_checkpoint(sess, FLAGS.start_checkpoint) # Turn all the variables into inline constants inside the graph and save it. @@ -155,10 +172,11 @@ if __name__ == '__main__': default=10.0, help='How long the stride is between spectrogram timeslices',) parser.add_argument( - '--dct_coefficient_count', + '--feature_bin_count', type=int, default=40, - help='How many bins to use for the MFCC fingerprint',) + help='How many bins to use for the MFCC fingerprint', + ) parser.add_argument( '--start_checkpoint', type=str, @@ -176,5 +194,15 @@ if __name__ == '__main__': help='Words to use (others will be added to an unknown label)',) parser.add_argument( '--output_file', type=str, help='Where to save the frozen graph.') + parser.add_argument( + '--quantize', + type=bool, + default=False, + help='Whether to train the model for eight-bit deployment') + parser.add_argument( + '--preprocess', + type=str, + default='mfcc', + help='Spectrogram processing mode. Can be "mfcc" or "average"') FLAGS, unparsed = parser.parse_known_args() tf.app.run(main=main, argv=[sys.argv[0]] + unparsed) diff --git a/tensorflow/examples/speech_commands/freeze_test.py b/tensorflow/examples/speech_commands/freeze_test.py index 97c6eac675f696d89d069258edf6eec901cfad0b..c8de6c2152909cd6dfca9acc895c25b0ae8e09ca 100644 --- a/tensorflow/examples/speech_commands/freeze_test.py +++ b/tensorflow/examples/speech_commands/freeze_test.py @@ -24,14 +24,62 @@ from tensorflow.python.platform import test class FreezeTest(test.TestCase): - def testCreateInferenceGraph(self): + def testCreateInferenceGraphWithMfcc(self): with self.test_session() as sess: - freeze.create_inference_graph('a,b,c,d', 16000, 1000.0, 30.0, 30.0, 10.0, - 40, 'conv') + freeze.create_inference_graph( + wanted_words='a,b,c,d', + sample_rate=16000, + clip_duration_ms=1000.0, + clip_stride_ms=30.0, + window_size_ms=30.0, + window_stride_ms=10.0, + feature_bin_count=40, + model_architecture='conv', + preprocess='mfcc') self.assertIsNotNone(sess.graph.get_tensor_by_name('wav_data:0')) self.assertIsNotNone( sess.graph.get_tensor_by_name('decoded_sample_data:0')) self.assertIsNotNone(sess.graph.get_tensor_by_name('labels_softmax:0')) + ops = [node.op for node in sess.graph_def.node] + self.assertEqual(1, ops.count('Mfcc')) + + def testCreateInferenceGraphWithoutMfcc(self): + with self.test_session() as sess: + freeze.create_inference_graph( + wanted_words='a,b,c,d', + sample_rate=16000, + clip_duration_ms=1000.0, + clip_stride_ms=30.0, + window_size_ms=30.0, + window_stride_ms=10.0, + feature_bin_count=40, + model_architecture='conv', + preprocess='average') + self.assertIsNotNone(sess.graph.get_tensor_by_name('wav_data:0')) + self.assertIsNotNone( + sess.graph.get_tensor_by_name('decoded_sample_data:0')) + self.assertIsNotNone(sess.graph.get_tensor_by_name('labels_softmax:0')) + ops = [node.op for node in sess.graph_def.node] + self.assertEqual(0, ops.count('Mfcc')) + + def testFeatureBinCount(self): + with self.test_session() as sess: + freeze.create_inference_graph( + wanted_words='a,b,c,d', + sample_rate=16000, + clip_duration_ms=1000.0, + clip_stride_ms=30.0, + window_size_ms=30.0, + window_stride_ms=10.0, + feature_bin_count=80, + model_architecture='conv', + preprocess='average') + self.assertIsNotNone(sess.graph.get_tensor_by_name('wav_data:0')) + self.assertIsNotNone( + sess.graph.get_tensor_by_name('decoded_sample_data:0')) + self.assertIsNotNone(sess.graph.get_tensor_by_name('labels_softmax:0')) + ops = [node.op for node in sess.graph_def.node] + self.assertEqual(0, ops.count('Mfcc')) if __name__ == '__main__': diff --git a/tensorflow/examples/speech_commands/generate_streaming_test_wav.py b/tensorflow/examples/speech_commands/generate_streaming_test_wav.py index 053206ae2f144ce05efa7eb490626aef01a6bc49..9858906927737cd520a9fd02f04437d01e0f6d31 100644 --- a/tensorflow/examples/speech_commands/generate_streaming_test_wav.py +++ b/tensorflow/examples/speech_commands/generate_streaming_test_wav.py @@ -87,11 +87,12 @@ def main(_): words_list = input_data.prepare_words_list(FLAGS.wanted_words.split(',')) model_settings = models.prepare_model_settings( len(words_list), FLAGS.sample_rate, FLAGS.clip_duration_ms, - FLAGS.window_size_ms, FLAGS.window_stride_ms, FLAGS.dct_coefficient_count) + FLAGS.window_size_ms, FLAGS.window_stride_ms, FLAGS.feature_bin_count, + 'mfcc') audio_processor = input_data.AudioProcessor( '', FLAGS.data_dir, FLAGS.silence_percentage, 10, FLAGS.wanted_words.split(','), FLAGS.validation_percentage, - FLAGS.testing_percentage, model_settings) + FLAGS.testing_percentage, model_settings, FLAGS.data_dir) output_audio_sample_count = FLAGS.sample_rate * FLAGS.test_duration_seconds output_audio = np.zeros((output_audio_sample_count,), dtype=np.float32) @@ -242,10 +243,11 @@ if __name__ == '__main__': default=10.0, help='How long the stride is between spectrogram timeslices',) parser.add_argument( - '--dct_coefficient_count', + '--feature_bin_count', type=int, default=40, - help='How many bins to use for the MFCC fingerprint',) + help='How many bins to use for the MFCC fingerprint', + ) parser.add_argument( '--wanted_words', type=str, diff --git a/tensorflow/examples/speech_commands/input_data.py b/tensorflow/examples/speech_commands/input_data.py index 63dd18457fea42acb09058b9ddd4623d72d1fd04..30f2cfa9fef7d0b5800c7e557bde4702dbafaf26 100644 --- a/tensorflow/examples/speech_commands/input_data.py +++ b/tensorflow/examples/speech_commands/input_data.py @@ -153,14 +153,14 @@ class AudioProcessor(object): def __init__(self, data_url, data_dir, silence_percentage, unknown_percentage, wanted_words, validation_percentage, testing_percentage, - model_settings): + model_settings, summaries_dir): self.data_dir = data_dir self.maybe_download_and_extract_dataset(data_url, data_dir) self.prepare_data_index(silence_percentage, unknown_percentage, wanted_words, validation_percentage, testing_percentage) self.prepare_background_data() - self.prepare_processing_graph(model_settings) + self.prepare_processing_graph(model_settings, summaries_dir) def maybe_download_and_extract_dataset(self, data_url, dest_directory): """Download and extract data set tar file. @@ -325,7 +325,7 @@ class AudioProcessor(object): if not self.background_data: raise Exception('No background wav files were found in ' + search_path) - def prepare_processing_graph(self, model_settings): + def prepare_processing_graph(self, model_settings, summaries_dir): """Builds a TensorFlow graph to apply the input distortions. Creates a graph that loads a WAVE file, decodes it, scales the volume, @@ -341,48 +341,88 @@ class AudioProcessor(object): - time_shift_offset_placeholder_: How much to move the clip in time. - background_data_placeholder_: PCM sample data for background noise. - background_volume_placeholder_: Loudness of mixed-in background. - - mfcc_: Output 2D fingerprint of processed audio. + - output_: Output 2D fingerprint of processed audio. Args: model_settings: Information about the current model being trained. + summaries_dir: Path to save training summary information to. + + Raises: + ValueError: If the preprocessing mode isn't recognized. """ - desired_samples = model_settings['desired_samples'] - self.wav_filename_placeholder_ = tf.placeholder(tf.string, []) - wav_loader = io_ops.read_file(self.wav_filename_placeholder_) - wav_decoder = contrib_audio.decode_wav( - wav_loader, desired_channels=1, desired_samples=desired_samples) - # Allow the audio sample's volume to be adjusted. - self.foreground_volume_placeholder_ = tf.placeholder(tf.float32, []) - scaled_foreground = tf.multiply(wav_decoder.audio, - self.foreground_volume_placeholder_) - # Shift the sample's start position, and pad any gaps with zeros. - self.time_shift_padding_placeholder_ = tf.placeholder(tf.int32, [2, 2]) - self.time_shift_offset_placeholder_ = tf.placeholder(tf.int32, [2]) - padded_foreground = tf.pad( - scaled_foreground, - self.time_shift_padding_placeholder_, - mode='CONSTANT') - sliced_foreground = tf.slice(padded_foreground, - self.time_shift_offset_placeholder_, - [desired_samples, -1]) - # Mix in background noise. - self.background_data_placeholder_ = tf.placeholder(tf.float32, - [desired_samples, 1]) - self.background_volume_placeholder_ = tf.placeholder(tf.float32, []) - background_mul = tf.multiply(self.background_data_placeholder_, - self.background_volume_placeholder_) - background_add = tf.add(background_mul, sliced_foreground) - background_clamp = tf.clip_by_value(background_add, -1.0, 1.0) - # Run the spectrogram and MFCC ops to get a 2D 'fingerprint' of the audio. - spectrogram = contrib_audio.audio_spectrogram( - background_clamp, - window_size=model_settings['window_size_samples'], - stride=model_settings['window_stride_samples'], - magnitude_squared=True) - self.mfcc_ = contrib_audio.mfcc( - spectrogram, - wav_decoder.sample_rate, - dct_coefficient_count=model_settings['dct_coefficient_count']) + with tf.get_default_graph().name_scope('data'): + desired_samples = model_settings['desired_samples'] + self.wav_filename_placeholder_ = tf.placeholder( + tf.string, [], name='wav_filename') + wav_loader = io_ops.read_file(self.wav_filename_placeholder_) + wav_decoder = contrib_audio.decode_wav( + wav_loader, desired_channels=1, desired_samples=desired_samples) + # Allow the audio sample's volume to be adjusted. + self.foreground_volume_placeholder_ = tf.placeholder( + tf.float32, [], name='foreground_volume') + scaled_foreground = tf.multiply(wav_decoder.audio, + self.foreground_volume_placeholder_) + # Shift the sample's start position, and pad any gaps with zeros. + self.time_shift_padding_placeholder_ = tf.placeholder( + tf.int32, [2, 2], name='time_shift_padding') + self.time_shift_offset_placeholder_ = tf.placeholder( + tf.int32, [2], name='time_shift_offset') + padded_foreground = tf.pad( + scaled_foreground, + self.time_shift_padding_placeholder_, + mode='CONSTANT') + sliced_foreground = tf.slice(padded_foreground, + self.time_shift_offset_placeholder_, + [desired_samples, -1]) + # Mix in background noise. + self.background_data_placeholder_ = tf.placeholder( + tf.float32, [desired_samples, 1], name='background_data') + self.background_volume_placeholder_ = tf.placeholder( + tf.float32, [], name='background_volume') + background_mul = tf.multiply(self.background_data_placeholder_, + self.background_volume_placeholder_) + background_add = tf.add(background_mul, sliced_foreground) + background_clamp = tf.clip_by_value(background_add, -1.0, 1.0) + # Run the spectrogram and MFCC ops to get a 2D 'fingerprint' of the audio. + spectrogram = contrib_audio.audio_spectrogram( + background_clamp, + window_size=model_settings['window_size_samples'], + stride=model_settings['window_stride_samples'], + magnitude_squared=True) + tf.summary.image( + 'spectrogram', tf.expand_dims(spectrogram, -1), max_outputs=1) + # The number of buckets in each FFT row in the spectrogram will depend on + # how many input samples there are in each window. This can be quite + # large, with a 160 sample window producing 127 buckets for example. We + # don't need this level of detail for classification, so we often want to + # shrink them down to produce a smaller result. That's what this section + # implements. One method is to use average pooling to merge adjacent + # buckets, but a more sophisticated approach is to apply the MFCC + # algorithm to shrink the representation. + if model_settings['preprocess'] == 'average': + self.output_ = tf.nn.pool( + tf.expand_dims(spectrogram, -1), + window_shape=[1, model_settings['average_window_width']], + strides=[1, model_settings['average_window_width']], + pooling_type='AVG', + padding='SAME') + tf.summary.image('shrunk_spectrogram', self.output_, max_outputs=1) + elif model_settings['preprocess'] == 'mfcc': + self.output_ = contrib_audio.mfcc( + spectrogram, + wav_decoder.sample_rate, + dct_coefficient_count=model_settings['fingerprint_width']) + tf.summary.image( + 'mfcc', tf.expand_dims(self.output_, -1), max_outputs=1) + else: + raise ValueError('Unknown preprocess mode "%s" (should be "mfcc" or' + ' "average")' % (model_settings['preprocess'])) + + # Merge all the summaries and write them out to /tmp/retrain_logs (by + # default) + self.merged_summaries_ = tf.summary.merge_all(scope='data') + self.summary_writer_ = tf.summary.FileWriter(summaries_dir + '/data', + tf.get_default_graph()) def set_size(self, mode): """Calculates the number of samples in the dataset partition. @@ -418,6 +458,9 @@ class AudioProcessor(object): Returns: List of sample data for the transformed samples, and list of label indexes + + Raises: + ValueError: If background samples are too short. """ # Pick one of the partitions to choose samples from. candidates = self.data_index[mode] @@ -460,6 +503,11 @@ class AudioProcessor(object): if use_background or sample['label'] == SILENCE_LABEL: background_index = np.random.randint(len(self.background_data)) background_samples = self.background_data[background_index] + if len(background_samples) <= model_settings['desired_samples']: + raise ValueError( + 'Background sample is too short! Need more than %d' + ' samples but only %d were found' % + (model_settings['desired_samples'], len(background_samples))) background_offset = np.random.randint( 0, len(background_samples) - model_settings['desired_samples']) background_clipped = background_samples[background_offset:( @@ -482,7 +530,10 @@ class AudioProcessor(object): else: input_dict[self.foreground_volume_placeholder_] = 1 # Run the graph to produce the output audio. - data[i - offset, :] = sess.run(self.mfcc_, feed_dict=input_dict).flatten() + summary, data_tensor = sess.run( + [self.merged_summaries_, self.output_], feed_dict=input_dict) + self.summary_writer_.add_summary(summary) + data[i - offset, :] = data_tensor.flatten() label_index = self.word_to_index[sample['label']] labels[i - offset] = label_index return data, labels diff --git a/tensorflow/examples/speech_commands/input_data_test.py b/tensorflow/examples/speech_commands/input_data_test.py index 13f294d39dbf89367496d2a16f466f8e2195d900..2e551be9a208221dc8b788e4d795e68bde21c9e5 100644 --- a/tensorflow/examples/speech_commands/input_data_test.py +++ b/tensorflow/examples/speech_commands/input_data_test.py @@ -25,6 +25,7 @@ import tensorflow as tf from tensorflow.contrib.framework.python.ops import audio_ops as contrib_audio from tensorflow.examples.speech_commands import input_data +from tensorflow.examples.speech_commands import models from tensorflow.python.platform import test @@ -32,7 +33,7 @@ class InputDataTest(test.TestCase): def _getWavData(self): with self.test_session() as sess: - sample_data = tf.zeros([1000, 2]) + sample_data = tf.zeros([32000, 2]) wav_encoder = contrib_audio.encode_wav(sample_data, 16000) wav_data = sess.run(wav_encoder) return wav_data @@ -57,9 +58,31 @@ class InputDataTest(test.TestCase): "label_count": 4, "window_size_samples": 100, "window_stride_samples": 100, - "dct_coefficient_count": 40, + "fingerprint_width": 40, + "preprocess": "mfcc", } + def _runGetDataTest(self, preprocess, window_length_ms): + tmp_dir = self.get_temp_dir() + wav_dir = os.path.join(tmp_dir, "wavs") + os.mkdir(wav_dir) + self._saveWavFolders(wav_dir, ["a", "b", "c"], 100) + background_dir = os.path.join(wav_dir, "_background_noise_") + os.mkdir(background_dir) + wav_data = self._getWavData() + for i in range(10): + file_path = os.path.join(background_dir, "background_audio_%d.wav" % i) + self._saveTestWavFile(file_path, wav_data) + model_settings = models.prepare_model_settings( + 4, 16000, 1000, window_length_ms, 20, 40, preprocess) + with self.test_session() as sess: + audio_processor = input_data.AudioProcessor( + "", wav_dir, 10, 10, ["a", "b"], 10, 10, model_settings, tmp_dir) + result_data, result_labels = audio_processor.get_data( + 10, 0, model_settings, 0.3, 0.1, 100, "training", sess) + self.assertEqual(10, len(result_data)) + self.assertEqual(10, len(result_labels)) + def testPrepareWordsList(self): words_list = ["a", "b"] self.assertGreater( @@ -76,8 +99,9 @@ class InputDataTest(test.TestCase): def testPrepareDataIndex(self): tmp_dir = self.get_temp_dir() self._saveWavFolders(tmp_dir, ["a", "b", "c"], 100) - audio_processor = input_data.AudioProcessor("", tmp_dir, 10, 10, ["a", "b"], - 10, 10, self._model_settings()) + audio_processor = input_data.AudioProcessor("", tmp_dir, 10, 10, + ["a", "b"], 10, 10, + self._model_settings(), tmp_dir) self.assertLess(0, audio_processor.set_size("training")) self.assertTrue("training" in audio_processor.data_index) self.assertTrue("validation" in audio_processor.data_index) @@ -90,7 +114,7 @@ class InputDataTest(test.TestCase): self._saveWavFolders(tmp_dir, ["a", "b", "c"], 0) with self.assertRaises(Exception) as e: _ = input_data.AudioProcessor("", tmp_dir, 10, 10, ["a", "b"], 10, 10, - self._model_settings()) + self._model_settings(), tmp_dir) self.assertTrue("No .wavs found" in str(e.exception)) def testPrepareDataIndexMissing(self): @@ -98,7 +122,7 @@ class InputDataTest(test.TestCase): self._saveWavFolders(tmp_dir, ["a", "b", "c"], 100) with self.assertRaises(Exception) as e: _ = input_data.AudioProcessor("", tmp_dir, 10, 10, ["a", "b", "d"], 10, - 10, self._model_settings()) + 10, self._model_settings(), tmp_dir) self.assertTrue("Expected to find" in str(e.exception)) def testPrepareBackgroundData(self): @@ -110,8 +134,9 @@ class InputDataTest(test.TestCase): file_path = os.path.join(background_dir, "background_audio_%d.wav" % i) self._saveTestWavFile(file_path, wav_data) self._saveWavFolders(tmp_dir, ["a", "b", "c"], 100) - audio_processor = input_data.AudioProcessor("", tmp_dir, 10, 10, ["a", "b"], - 10, 10, self._model_settings()) + audio_processor = input_data.AudioProcessor("", tmp_dir, 10, 10, + ["a", "b"], 10, 10, + self._model_settings(), tmp_dir) self.assertEqual(10, len(audio_processor.background_data)) def testLoadWavFile(self): @@ -148,44 +173,27 @@ class InputDataTest(test.TestCase): "label_count": 4, "window_size_samples": 100, "window_stride_samples": 100, - "dct_coefficient_count": 40, + "fingerprint_width": 40, + "preprocess": "mfcc", } audio_processor = input_data.AudioProcessor("", wav_dir, 10, 10, ["a", "b"], - 10, 10, model_settings) + 10, 10, model_settings, tmp_dir) self.assertIsNotNone(audio_processor.wav_filename_placeholder_) self.assertIsNotNone(audio_processor.foreground_volume_placeholder_) self.assertIsNotNone(audio_processor.time_shift_padding_placeholder_) self.assertIsNotNone(audio_processor.time_shift_offset_placeholder_) self.assertIsNotNone(audio_processor.background_data_placeholder_) self.assertIsNotNone(audio_processor.background_volume_placeholder_) - self.assertIsNotNone(audio_processor.mfcc_) + self.assertIsNotNone(audio_processor.output_) - def testGetData(self): - tmp_dir = self.get_temp_dir() - wav_dir = os.path.join(tmp_dir, "wavs") - os.mkdir(wav_dir) - self._saveWavFolders(wav_dir, ["a", "b", "c"], 100) - background_dir = os.path.join(wav_dir, "_background_noise_") - os.mkdir(background_dir) - wav_data = self._getWavData() - for i in range(10): - file_path = os.path.join(background_dir, "background_audio_%d.wav" % i) - self._saveTestWavFile(file_path, wav_data) - model_settings = { - "desired_samples": 160, - "fingerprint_size": 40, - "label_count": 4, - "window_size_samples": 100, - "window_stride_samples": 100, - "dct_coefficient_count": 40, - } - audio_processor = input_data.AudioProcessor("", wav_dir, 10, 10, ["a", "b"], - 10, 10, model_settings) - with self.test_session() as sess: - result_data, result_labels = audio_processor.get_data( - 10, 0, model_settings, 0.3, 0.1, 100, "training", sess) - self.assertEqual(10, len(result_data)) - self.assertEqual(10, len(result_labels)) + def testGetDataAverage(self): + self._runGetDataTest("average", 10) + + def testGetDataAverageLongWindow(self): + self._runGetDataTest("average", 30) + + def testGetDataMfcc(self): + self._runGetDataTest("mfcc", 30) def testGetUnprocessedData(self): tmp_dir = self.get_temp_dir() @@ -198,10 +206,11 @@ class InputDataTest(test.TestCase): "label_count": 4, "window_size_samples": 100, "window_stride_samples": 100, - "dct_coefficient_count": 40, + "fingerprint_width": 40, + "preprocess": "mfcc", } audio_processor = input_data.AudioProcessor("", wav_dir, 10, 10, ["a", "b"], - 10, 10, model_settings) + 10, 10, model_settings, tmp_dir) result_data, result_labels = audio_processor.get_unprocessed_data( 10, model_settings, "training") self.assertEqual(10, len(result_data)) diff --git a/tensorflow/examples/speech_commands/models.py b/tensorflow/examples/speech_commands/models.py index ab611f414a8afa1f08b955918071b04ae0ef88db..65ae3b1511135e6aae0adb5f2536797966a107a1 100644 --- a/tensorflow/examples/speech_commands/models.py +++ b/tensorflow/examples/speech_commands/models.py @@ -24,9 +24,21 @@ import math import tensorflow as tf +def _next_power_of_two(x): + """Calculates the smallest enclosing power of two for an input. + + Args: + x: Positive float or integer number. + + Returns: + Next largest power of two integer. + """ + return 1 if x == 0 else 2**(int(x) - 1).bit_length() + + def prepare_model_settings(label_count, sample_rate, clip_duration_ms, - window_size_ms, window_stride_ms, - dct_coefficient_count): + window_size_ms, window_stride_ms, feature_bin_count, + preprocess): """Calculates common settings needed for all models. Args: @@ -35,10 +47,14 @@ def prepare_model_settings(label_count, sample_rate, clip_duration_ms, clip_duration_ms: Length of each audio clip to be analyzed. window_size_ms: Duration of frequency analysis window. window_stride_ms: How far to move in time between frequency windows. - dct_coefficient_count: Number of frequency bins to use for analysis. + feature_bin_count: Number of frequency bins to use for analysis. + preprocess: How the spectrogram is processed to produce features. Returns: Dictionary containing common settings. + + Raises: + ValueError: If the preprocessing mode isn't recognized. """ desired_samples = int(sample_rate * clip_duration_ms / 1000) window_size_samples = int(sample_rate * window_size_ms / 1000) @@ -48,16 +64,28 @@ def prepare_model_settings(label_count, sample_rate, clip_duration_ms, spectrogram_length = 0 else: spectrogram_length = 1 + int(length_minus_window / window_stride_samples) - fingerprint_size = dct_coefficient_count * spectrogram_length + if preprocess == 'average': + fft_bin_count = 1 + (_next_power_of_two(window_size_samples) / 2) + average_window_width = int(math.floor(fft_bin_count / feature_bin_count)) + fingerprint_width = int(math.ceil(fft_bin_count / average_window_width)) + elif preprocess == 'mfcc': + average_window_width = -1 + fingerprint_width = feature_bin_count + else: + raise ValueError('Unknown preprocess mode "%s" (should be "mfcc" or' + ' "average")' % (preprocess)) + fingerprint_size = fingerprint_width * spectrogram_length return { 'desired_samples': desired_samples, 'window_size_samples': window_size_samples, 'window_stride_samples': window_stride_samples, 'spectrogram_length': spectrogram_length, - 'dct_coefficient_count': dct_coefficient_count, + 'fingerprint_width': fingerprint_width, 'fingerprint_size': fingerprint_size, 'label_count': label_count, 'sample_rate': sample_rate, + 'preprocess': preprocess, + 'average_window_width': average_window_width, } @@ -106,10 +134,14 @@ def create_model(fingerprint_input, model_settings, model_architecture, elif model_architecture == 'low_latency_svdf': return create_low_latency_svdf_model(fingerprint_input, model_settings, is_training, runtime_settings) + elif model_architecture == 'tiny_conv': + return create_tiny_conv_model(fingerprint_input, model_settings, + is_training) else: raise Exception('model_architecture argument "' + model_architecture + '" not recognized, should be one of "single_fc", "conv",' + - ' "low_latency_conv, or "low_latency_svdf"') + ' "low_latency_conv, "low_latency_svdf",' + + ' or "tiny_conv"') def load_variables_from_checkpoint(sess, start_checkpoint): @@ -152,9 +184,12 @@ def create_single_fc_model(fingerprint_input, model_settings, is_training): dropout_prob = tf.placeholder(tf.float32, name='dropout_prob') fingerprint_size = model_settings['fingerprint_size'] label_count = model_settings['label_count'] - weights = tf.Variable( - tf.truncated_normal([fingerprint_size, label_count], stddev=0.001)) - bias = tf.Variable(tf.zeros([label_count])) + weights = tf.get_variable( + name='weights', + initializer=tf.truncated_normal_initializer(stddev=0.001), + shape=[fingerprint_size, label_count]) + bias = tf.get_variable( + name='bias', initializer=tf.zeros_initializer, shape=[label_count]) logits = tf.matmul(fingerprint_input, weights) + bias if is_training: return logits, dropout_prob @@ -212,18 +247,21 @@ def create_conv_model(fingerprint_input, model_settings, is_training): """ if is_training: dropout_prob = tf.placeholder(tf.float32, name='dropout_prob') - input_frequency_size = model_settings['dct_coefficient_count'] + input_frequency_size = model_settings['fingerprint_width'] input_time_size = model_settings['spectrogram_length'] fingerprint_4d = tf.reshape(fingerprint_input, [-1, input_time_size, input_frequency_size, 1]) first_filter_width = 8 first_filter_height = 20 first_filter_count = 64 - first_weights = tf.Variable( - tf.truncated_normal( - [first_filter_height, first_filter_width, 1, first_filter_count], - stddev=0.01)) - first_bias = tf.Variable(tf.zeros([first_filter_count])) + first_weights = tf.get_variable( + name='first_weights', + initializer=tf.truncated_normal_initializer(stddev=0.01), + shape=[first_filter_height, first_filter_width, 1, first_filter_count]) + first_bias = tf.get_variable( + name='first_bias', + initializer=tf.zeros_initializer, + shape=[first_filter_count]) first_conv = tf.nn.conv2d(fingerprint_4d, first_weights, [1, 1, 1, 1], 'SAME') + first_bias first_relu = tf.nn.relu(first_conv) @@ -235,14 +273,17 @@ def create_conv_model(fingerprint_input, model_settings, is_training): second_filter_width = 4 second_filter_height = 10 second_filter_count = 64 - second_weights = tf.Variable( - tf.truncated_normal( - [ - second_filter_height, second_filter_width, first_filter_count, - second_filter_count - ], - stddev=0.01)) - second_bias = tf.Variable(tf.zeros([second_filter_count])) + second_weights = tf.get_variable( + name='second_weights', + initializer=tf.truncated_normal_initializer(stddev=0.01), + shape=[ + second_filter_height, second_filter_width, first_filter_count, + second_filter_count + ]) + second_bias = tf.get_variable( + name='second_bias', + initializer=tf.zeros_initializer, + shape=[second_filter_count]) second_conv = tf.nn.conv2d(max_pool, second_weights, [1, 1, 1, 1], 'SAME') + second_bias second_relu = tf.nn.relu(second_conv) @@ -259,10 +300,14 @@ def create_conv_model(fingerprint_input, model_settings, is_training): flattened_second_conv = tf.reshape(second_dropout, [-1, second_conv_element_count]) label_count = model_settings['label_count'] - final_fc_weights = tf.Variable( - tf.truncated_normal( - [second_conv_element_count, label_count], stddev=0.01)) - final_fc_bias = tf.Variable(tf.zeros([label_count])) + final_fc_weights = tf.get_variable( + name='final_fc_weights', + initializer=tf.truncated_normal_initializer, + shape=[second_conv_element_count, label_count]) + final_fc_bias = tf.get_variable( + name='final_fc_bias', + initializer=tf.zeros_initializer, + shape=[label_count]) final_fc = tf.matmul(flattened_second_conv, final_fc_weights) + final_fc_bias if is_training: return final_fc, dropout_prob @@ -318,7 +363,7 @@ def create_low_latency_conv_model(fingerprint_input, model_settings, """ if is_training: dropout_prob = tf.placeholder(tf.float32, name='dropout_prob') - input_frequency_size = model_settings['dct_coefficient_count'] + input_frequency_size = model_settings['fingerprint_width'] input_time_size = model_settings['spectrogram_length'] fingerprint_4d = tf.reshape(fingerprint_input, [-1, input_time_size, input_frequency_size, 1]) @@ -327,11 +372,14 @@ def create_low_latency_conv_model(fingerprint_input, model_settings, first_filter_count = 186 first_filter_stride_x = 1 first_filter_stride_y = 1 - first_weights = tf.Variable( - tf.truncated_normal( - [first_filter_height, first_filter_width, 1, first_filter_count], - stddev=0.01)) - first_bias = tf.Variable(tf.zeros([first_filter_count])) + first_weights = tf.get_variable( + name='first_weights', + initializer=tf.truncated_normal_initializer(stddev=0.01), + shape=[first_filter_height, first_filter_width, 1, first_filter_count]) + first_bias = tf.get_variable( + name='first_bias', + initializer=tf.zeros_initializer, + shape=[first_filter_count]) first_conv = tf.nn.conv2d(fingerprint_4d, first_weights, [ 1, first_filter_stride_y, first_filter_stride_x, 1 ], 'VALID') + first_bias @@ -351,30 +399,42 @@ def create_low_latency_conv_model(fingerprint_input, model_settings, flattened_first_conv = tf.reshape(first_dropout, [-1, first_conv_element_count]) first_fc_output_channels = 128 - first_fc_weights = tf.Variable( - tf.truncated_normal( - [first_conv_element_count, first_fc_output_channels], stddev=0.01)) - first_fc_bias = tf.Variable(tf.zeros([first_fc_output_channels])) + first_fc_weights = tf.get_variable( + name='first_fc_weights', + initializer=tf.truncated_normal_initializer(stddev=0.01), + shape=[first_conv_element_count, first_fc_output_channels]) + first_fc_bias = tf.get_variable( + name='first_fc_bias', + initializer=tf.zeros_initializer, + shape=[first_fc_output_channels]) first_fc = tf.matmul(flattened_first_conv, first_fc_weights) + first_fc_bias if is_training: second_fc_input = tf.nn.dropout(first_fc, dropout_prob) else: second_fc_input = first_fc second_fc_output_channels = 128 - second_fc_weights = tf.Variable( - tf.truncated_normal( - [first_fc_output_channels, second_fc_output_channels], stddev=0.01)) - second_fc_bias = tf.Variable(tf.zeros([second_fc_output_channels])) + second_fc_weights = tf.get_variable( + name='second_fc_weights', + initializer=tf.truncated_normal_initializer(stddev=0.01), + shape=[first_fc_output_channels, second_fc_output_channels]) + second_fc_bias = tf.get_variable( + name='second_fc_bias', + initializer=tf.zeros_initializer, + shape=[second_fc_output_channels]) second_fc = tf.matmul(second_fc_input, second_fc_weights) + second_fc_bias if is_training: final_fc_input = tf.nn.dropout(second_fc, dropout_prob) else: final_fc_input = second_fc label_count = model_settings['label_count'] - final_fc_weights = tf.Variable( - tf.truncated_normal( - [second_fc_output_channels, label_count], stddev=0.01)) - final_fc_bias = tf.Variable(tf.zeros([label_count])) + final_fc_weights = tf.get_variable( + name='final_fc_weights', + initializer=tf.truncated_normal_initializer(stddev=0.01), + shape=[second_fc_output_channels, label_count]) + final_fc_bias = tf.get_variable( + name='final_fc_bias', + initializer=tf.zeros_initializer, + shape=[label_count]) final_fc = tf.matmul(final_fc_input, final_fc_weights) + final_fc_bias if is_training: return final_fc, dropout_prob @@ -422,7 +482,7 @@ def create_low_latency_svdf_model(fingerprint_input, model_settings, Args: fingerprint_input: TensorFlow node that will output audio feature vectors. The node is expected to produce a 2D Tensor of shape: - [batch, model_settings['dct_coefficient_count'] * + [batch, model_settings['fingerprint_width'] * model_settings['spectrogram_length']] with the features corresponding to the same time slot arranged contiguously, and the oldest slot at index [:, 0], and newest at [:, -1]. @@ -440,7 +500,7 @@ def create_low_latency_svdf_model(fingerprint_input, model_settings, if is_training: dropout_prob = tf.placeholder(tf.float32, name='dropout_prob') - input_frequency_size = model_settings['dct_coefficient_count'] + input_frequency_size = model_settings['fingerprint_width'] input_time_size = model_settings['spectrogram_length'] # Validation. @@ -462,8 +522,11 @@ def create_low_latency_svdf_model(fingerprint_input, model_settings, num_filters = rank * num_units # Create the runtime memory: [num_filters, batch, input_time_size] batch = 1 - memory = tf.Variable(tf.zeros([num_filters, batch, input_time_size]), - trainable=False, name='runtime-memory') + memory = tf.get_variable( + initializer=tf.zeros_initializer, + shape=[num_filters, batch, input_time_size], + trainable=False, + name='runtime-memory') # Determine the number of new frames in the input, such that we only operate # on those. For training we do not use the memory, and thus use all frames # provided in the input. @@ -483,8 +546,10 @@ def create_low_latency_svdf_model(fingerprint_input, model_settings, new_fingerprint_input = tf.expand_dims(new_fingerprint_input, 2) # Create the frequency filters. - weights_frequency = tf.Variable( - tf.truncated_normal([input_frequency_size, num_filters], stddev=0.01)) + weights_frequency = tf.get_variable( + name='weights_frequency', + initializer=tf.truncated_normal_initializer(stddev=0.01), + shape=[input_frequency_size, num_filters]) # Expand to add input channels dimensions. # weights_frequency: [input_frequency_size, 1, num_filters] weights_frequency = tf.expand_dims(weights_frequency, 1) @@ -506,8 +571,10 @@ def create_low_latency_svdf_model(fingerprint_input, model_settings, activations_time = new_memory # Create the time filters. - weights_time = tf.Variable( - tf.truncated_normal([num_filters, input_time_size], stddev=0.01)) + weights_time = tf.get_variable( + name='weights_time', + initializer=tf.truncated_normal_initializer(stddev=0.01), + shape=[num_filters, input_time_size]) # Apply the time filter on the outputs of the feature filters. # weights_time: [num_filters, input_time_size, 1] # outputs: [num_filters, batch, 1] @@ -524,7 +591,8 @@ def create_low_latency_svdf_model(fingerprint_input, model_settings, units_output = tf.transpose(units_output) # Appy bias. - bias = tf.Variable(tf.zeros([num_units])) + bias = tf.get_variable( + name='bias', initializer=tf.zeros_initializer, shape=[num_units]) first_bias = tf.nn.bias_add(units_output, bias) # Relu. @@ -536,31 +604,135 @@ def create_low_latency_svdf_model(fingerprint_input, model_settings, first_dropout = first_relu first_fc_output_channels = 256 - first_fc_weights = tf.Variable( - tf.truncated_normal([num_units, first_fc_output_channels], stddev=0.01)) - first_fc_bias = tf.Variable(tf.zeros([first_fc_output_channels])) + first_fc_weights = tf.get_variable( + name='first_fc_weights', + initializer=tf.truncated_normal_initializer(stddev=0.01), + shape=[num_units, first_fc_output_channels]) + first_fc_bias = tf.get_variable( + name='first_fc_bias', + initializer=tf.zeros_initializer, + shape=[first_fc_output_channels]) first_fc = tf.matmul(first_dropout, first_fc_weights) + first_fc_bias if is_training: second_fc_input = tf.nn.dropout(first_fc, dropout_prob) else: second_fc_input = first_fc second_fc_output_channels = 256 - second_fc_weights = tf.Variable( - tf.truncated_normal( - [first_fc_output_channels, second_fc_output_channels], stddev=0.01)) - second_fc_bias = tf.Variable(tf.zeros([second_fc_output_channels])) + second_fc_weights = tf.get_variable( + name='second_fc_weights', + initializer=tf.truncated_normal_initializer(stddev=0.01), + shape=[first_fc_output_channels, second_fc_output_channels]) + second_fc_bias = tf.get_variable( + name='second_fc_bias', + initializer=tf.zeros_initializer, + shape=[second_fc_output_channels]) second_fc = tf.matmul(second_fc_input, second_fc_weights) + second_fc_bias if is_training: final_fc_input = tf.nn.dropout(second_fc, dropout_prob) else: final_fc_input = second_fc label_count = model_settings['label_count'] - final_fc_weights = tf.Variable( - tf.truncated_normal( - [second_fc_output_channels, label_count], stddev=0.01)) - final_fc_bias = tf.Variable(tf.zeros([label_count])) + final_fc_weights = tf.get_variable( + name='final_fc_weights', + initializer=tf.truncated_normal(stddev=0.01), + shape=[second_fc_output_channels, label_count]) + final_fc_bias = tf.get_variable( + name='final_fc_bias', + initializer=tf.zeros_initializer, + shape=[label_count]) final_fc = tf.matmul(final_fc_input, final_fc_weights) + final_fc_bias if is_training: return final_fc, dropout_prob else: return final_fc + + +def create_tiny_conv_model(fingerprint_input, model_settings, is_training): + """Builds a convolutional model aimed at microcontrollers. + + Devices like DSPs and microcontrollers can have very small amounts of + memory and limited processing power. This model is designed to use less + than 20KB of working RAM, and fit within 32KB of read-only (flash) memory. + + Here's the layout of the graph: + + (fingerprint_input) + v + [Conv2D]<-(weights) + v + [BiasAdd]<-(bias) + v + [Relu] + v + [MatMul]<-(weights) + v + [BiasAdd]<-(bias) + v + + This doesn't produce particularly accurate results, but it's designed to be + used as the first stage of a pipeline, running on a low-energy piece of + hardware that can always be on, and then wake higher-power chips when a + possible utterance has been found, so that more accurate analysis can be done. + + During training, a dropout node is introduced after the relu, controlled by a + placeholder. + + Args: + fingerprint_input: TensorFlow node that will output audio feature vectors. + model_settings: Dictionary of information about the model. + is_training: Whether the model is going to be used for training. + + Returns: + TensorFlow node outputting logits results, and optionally a dropout + placeholder. + """ + if is_training: + dropout_prob = tf.placeholder(tf.float32, name='dropout_prob') + input_frequency_size = model_settings['fingerprint_width'] + input_time_size = model_settings['spectrogram_length'] + fingerprint_4d = tf.reshape(fingerprint_input, + [-1, input_time_size, input_frequency_size, 1]) + first_filter_width = 8 + first_filter_height = 10 + first_filter_count = 8 + first_weights = tf.get_variable( + name='first_weights', + initializer=tf.truncated_normal_initializer(stddev=0.01), + shape=[first_filter_height, first_filter_width, 1, first_filter_count]) + first_bias = tf.get_variable( + name='first_bias', + initializer=tf.zeros_initializer, + shape=[first_filter_count]) + first_conv_stride_x = 2 + first_conv_stride_y = 2 + first_conv = tf.nn.conv2d(fingerprint_4d, first_weights, + [1, first_conv_stride_y, first_conv_stride_x, 1], + 'SAME') + first_bias + first_relu = tf.nn.relu(first_conv) + if is_training: + first_dropout = tf.nn.dropout(first_relu, dropout_prob) + else: + first_dropout = first_relu + first_dropout_shape = first_dropout.get_shape() + first_dropout_output_width = first_dropout_shape[2] + first_dropout_output_height = first_dropout_shape[1] + first_dropout_element_count = int( + first_dropout_output_width * first_dropout_output_height * + first_filter_count) + flattened_first_dropout = tf.reshape(first_dropout, + [-1, first_dropout_element_count]) + label_count = model_settings['label_count'] + final_fc_weights = tf.get_variable( + name='final_fc_weights', + initializer=tf.truncated_normal_initializer(stddev=0.01), + shape=[first_dropout_element_count, label_count]) + final_fc_bias = tf.get_variable( + name='final_fc_bias', + initializer=tf.zeros_initializer, + shape=[label_count]) + final_fc = ( + tf.matmul(flattened_first_dropout, final_fc_weights) + final_fc_bias) + if is_training: + return final_fc, dropout_prob + else: + return final_fc diff --git a/tensorflow/examples/speech_commands/models_test.py b/tensorflow/examples/speech_commands/models_test.py index 80c795367fa01f214d78d3fa7df7864b6b243b97..0c373967ed8fb9cddcc82972e0fc8bba186add2e 100644 --- a/tensorflow/examples/speech_commands/models_test.py +++ b/tensorflow/examples/speech_commands/models_test.py @@ -26,12 +26,29 @@ from tensorflow.python.platform import test class ModelsTest(test.TestCase): + def _modelSettings(self): + return models.prepare_model_settings( + label_count=10, + sample_rate=16000, + clip_duration_ms=1000, + window_size_ms=20, + window_stride_ms=10, + feature_bin_count=40, + preprocess="mfcc") + def testPrepareModelSettings(self): self.assertIsNotNone( - models.prepare_model_settings(10, 16000, 1000, 20, 10, 40)) + models.prepare_model_settings( + label_count=10, + sample_rate=16000, + clip_duration_ms=1000, + window_size_ms=20, + window_stride_ms=10, + feature_bin_count=40, + preprocess="mfcc")) def testCreateModelConvTraining(self): - model_settings = models.prepare_model_settings(10, 16000, 1000, 20, 10, 40) + model_settings = self._modelSettings() with self.test_session() as sess: fingerprint_input = tf.zeros([1, model_settings["fingerprint_size"]]) logits, dropout_prob = models.create_model(fingerprint_input, @@ -42,7 +59,7 @@ class ModelsTest(test.TestCase): self.assertIsNotNone(sess.graph.get_tensor_by_name(dropout_prob.name)) def testCreateModelConvInference(self): - model_settings = models.prepare_model_settings(10, 16000, 1000, 20, 10, 40) + model_settings = self._modelSettings() with self.test_session() as sess: fingerprint_input = tf.zeros([1, model_settings["fingerprint_size"]]) logits = models.create_model(fingerprint_input, model_settings, "conv", @@ -51,7 +68,7 @@ class ModelsTest(test.TestCase): self.assertIsNotNone(sess.graph.get_tensor_by_name(logits.name)) def testCreateModelLowLatencyConvTraining(self): - model_settings = models.prepare_model_settings(10, 16000, 1000, 20, 10, 40) + model_settings = self._modelSettings() with self.test_session() as sess: fingerprint_input = tf.zeros([1, model_settings["fingerprint_size"]]) logits, dropout_prob = models.create_model( @@ -62,7 +79,7 @@ class ModelsTest(test.TestCase): self.assertIsNotNone(sess.graph.get_tensor_by_name(dropout_prob.name)) def testCreateModelFullyConnectedTraining(self): - model_settings = models.prepare_model_settings(10, 16000, 1000, 20, 10, 40) + model_settings = self._modelSettings() with self.test_session() as sess: fingerprint_input = tf.zeros([1, model_settings["fingerprint_size"]]) logits, dropout_prob = models.create_model( @@ -73,7 +90,7 @@ class ModelsTest(test.TestCase): self.assertIsNotNone(sess.graph.get_tensor_by_name(dropout_prob.name)) def testCreateModelBadArchitecture(self): - model_settings = models.prepare_model_settings(10, 16000, 1000, 20, 10, 40) + model_settings = self._modelSettings() with self.test_session(): fingerprint_input = tf.zeros([1, model_settings["fingerprint_size"]]) with self.assertRaises(Exception) as e: @@ -81,6 +98,17 @@ class ModelsTest(test.TestCase): "bad_architecture", True) self.assertTrue("not recognized" in str(e.exception)) + def testCreateModelTinyConvTraining(self): + model_settings = self._modelSettings() + with self.test_session() as sess: + fingerprint_input = tf.zeros([1, model_settings["fingerprint_size"]]) + logits, dropout_prob = models.create_model( + fingerprint_input, model_settings, "tiny_conv", True) + self.assertIsNotNone(logits) + self.assertIsNotNone(dropout_prob) + self.assertIsNotNone(sess.graph.get_tensor_by_name(logits.name)) + self.assertIsNotNone(sess.graph.get_tensor_by_name(dropout_prob.name)) + if __name__ == "__main__": test.main() diff --git a/tensorflow/examples/speech_commands/train.py b/tensorflow/examples/speech_commands/train.py index fc28eb0631dc5e1947c2a31a6acdb02ed8d28f3a..eca34f8812b76a60168c97a745f5890bf3ee0269 100644 --- a/tensorflow/examples/speech_commands/train.py +++ b/tensorflow/examples/speech_commands/train.py @@ -98,12 +98,12 @@ def main(_): model_settings = models.prepare_model_settings( len(input_data.prepare_words_list(FLAGS.wanted_words.split(','))), FLAGS.sample_rate, FLAGS.clip_duration_ms, FLAGS.window_size_ms, - FLAGS.window_stride_ms, FLAGS.dct_coefficient_count) + FLAGS.window_stride_ms, FLAGS.feature_bin_count, FLAGS.preprocess) audio_processor = input_data.AudioProcessor( - FLAGS.data_url, FLAGS.data_dir, FLAGS.silence_percentage, - FLAGS.unknown_percentage, + FLAGS.data_url, FLAGS.data_dir, + FLAGS.silence_percentage, FLAGS.unknown_percentage, FLAGS.wanted_words.split(','), FLAGS.validation_percentage, - FLAGS.testing_percentage, model_settings) + FLAGS.testing_percentage, model_settings, FLAGS.summaries_dir) fingerprint_size = model_settings['fingerprint_size'] label_count = model_settings['label_count'] time_shift_samples = int((FLAGS.time_shift_ms * FLAGS.sample_rate) / 1000) @@ -122,8 +122,25 @@ def main(_): 'lists, but are %d and %d long instead' % (len(training_steps_list), len(learning_rates_list))) - fingerprint_input = tf.placeholder( + input_placeholder = tf.placeholder( tf.float32, [None, fingerprint_size], name='fingerprint_input') + if FLAGS.quantize: + # TODO(petewarden): These values have been derived from the observed ranges + # of spectrogram and MFCC inputs. If the preprocessing pipeline changes, + # they may need to be updated. + if FLAGS.preprocess == 'average': + fingerprint_min = 0.0 + fingerprint_max = 2048.0 + elif FLAGS.preprocess == 'mfcc': + fingerprint_min = -247.0 + fingerprint_max = 30.0 + else: + raise Exception('Unknown preprocess mode "%s" (should be "mfcc" or' + ' "average")' % (FLAGS.preprocess)) + fingerprint_input = tf.fake_quant_with_min_max_args( + input_placeholder, fingerprint_min, fingerprint_max) + else: + fingerprint_input = input_placeholder logits, dropout_prob = models.create_model( fingerprint_input, @@ -146,7 +163,8 @@ def main(_): with tf.name_scope('cross_entropy'): cross_entropy_mean = tf.losses.sparse_softmax_cross_entropy( labels=ground_truth_input, logits=logits) - tf.summary.scalar('cross_entropy', cross_entropy_mean) + if FLAGS.quantize: + tf.contrib.quantize.create_training_graph(quant_delay=0) with tf.name_scope('train'), tf.control_dependencies(control_dependencies): learning_rate_input = tf.placeholder( tf.float32, [], name='learning_rate_input') @@ -157,7 +175,9 @@ def main(_): confusion_matrix = tf.confusion_matrix( ground_truth_input, predicted_indices, num_classes=label_count) evaluation_step = tf.reduce_mean(tf.cast(correct_prediction, tf.float32)) - tf.summary.scalar('accuracy', evaluation_step) + with tf.get_default_graph().name_scope('eval'): + tf.summary.scalar('cross_entropy', cross_entropy_mean) + tf.summary.scalar('accuracy', evaluation_step) global_step = tf.train.get_or_create_global_step() increment_global_step = tf.assign(global_step, global_step + 1) @@ -165,7 +185,7 @@ def main(_): saver = tf.train.Saver(tf.global_variables()) # Merge all the summaries and write them out to /tmp/retrain_logs (by default) - merged_summaries = tf.summary.merge_all() + merged_summaries = tf.summary.merge_all(scope='eval') train_writer = tf.summary.FileWriter(FLAGS.summaries_dir + '/train', sess.graph) validation_writer = tf.summary.FileWriter(FLAGS.summaries_dir + '/validation') @@ -207,8 +227,11 @@ def main(_): # Run the graph with this batch of training data. train_summary, train_accuracy, cross_entropy_value, _, _ = sess.run( [ - merged_summaries, evaluation_step, cross_entropy_mean, train_step, - increment_global_step + merged_summaries, + evaluation_step, + cross_entropy_mean, + train_step, + increment_global_step, ], feed_dict={ fingerprint_input: train_fingerprints, @@ -364,10 +387,11 @@ if __name__ == '__main__': default=10.0, help='How far to move in time between spectogram timeslices.',) parser.add_argument( - '--dct_coefficient_count', + '--feature_bin_count', type=int, default=40, - help='How many bins to use for the MFCC fingerprint',) + help='How many bins to use for the MFCC fingerprint', + ) parser.add_argument( '--how_many_training_steps', type=str, @@ -423,6 +447,16 @@ if __name__ == '__main__': type=bool, default=False, help='Whether to check for invalid numbers during processing') + parser.add_argument( + '--quantize', + type=bool, + default=False, + help='Whether to train the model for eight-bit deployment') + parser.add_argument( + '--preprocess', + type=str, + default='mfcc', + help='Spectrogram processing mode. Can be "mfcc" or "average"') FLAGS, unparsed = parser.parse_known_args() tf.app.run(main=main, argv=[sys.argv[0]] + unparsed) diff --git a/tensorflow/go/attrs_test.go b/tensorflow/go/attrs_test.go index 35b0cb352e7a5c1ca2e465720cd4dc125f166675..ea8af221aeef3bf1d2edeab4372ae00f0cc7e92d 100644 --- a/tensorflow/go/attrs_test.go +++ b/tensorflow/go/attrs_test.go @@ -28,7 +28,7 @@ func TestOperationAttrs(t *testing.T) { i := 0 makeConst := func(v interface{}) Output { op, err := Const(g, fmt.Sprintf("const/%d/%+v", i, v), v) - i += 1 + i++ if err != nil { t.Fatal(err) } @@ -71,6 +71,7 @@ func TestOperationAttrs(t *testing.T) { "boundaries": []float32(nil), }, }, + /* TODO(ashankar): debug this issue and add it back later. { Name: "list(type),list(shape)", Type: "InfeedEnqueueTuple", @@ -111,6 +112,7 @@ func TestOperationAttrs(t *testing.T) { "device_ordinal": int64(0), }, }, + */ { Name: "list(int),int", Type: "StringToHashBucketStrong", diff --git a/tensorflow/go/graph.go b/tensorflow/go/graph.go index 08943a527cbdc072b12b066240c213be45ffd54c..32a77550ee2fa5606b402600aa6429950d8e72a5 100644 --- a/tensorflow/go/graph.go +++ b/tensorflow/go/graph.go @@ -177,7 +177,14 @@ type OpSpec struct { // being added. ControlDependencies []*Operation - // Other possible fields: Device, ColocateWith. + // The device on which the operation should be executed. + // If omitted, an appropriate device will automatically be selected. + // + // For example, if set of "/device:GPU:0", then the operation will + // execute on GPU #0. + Device string + + // Other possible fields: ColocateWith. } // AddOperation adds an operation to g. @@ -225,6 +232,11 @@ func (g *Graph) AddOperation(args OpSpec) (*Operation, error) { return nil, fmt.Errorf("%v (memory will be leaked)", err) } } + if len(args.Device) > 0 { + cdevice := C.CString(args.Device) + C.TF_SetDevice(cdesc, cdevice) + C.free(unsafe.Pointer(cdevice)) + } c := C.TF_FinishOperation(cdesc, status.c) if err := status.Err(); err != nil { return nil, err diff --git a/tensorflow/go/op/scope.go b/tensorflow/go/op/scope.go index 13de4294dc2ebdfff9bb68d277c09239d0bc8593..ac39808d838f4737b81b170d3f540d10ed38fe42 100644 --- a/tensorflow/go/op/scope.go +++ b/tensorflow/go/op/scope.go @@ -37,6 +37,7 @@ type Scope struct { namemap map[string]int namespace string controlDependencies []*tf.Operation + device string err *scopeErr } @@ -82,6 +83,7 @@ func (s *Scope) AddOperation(args tf.OpSpec) *tf.Operation { args.Name = s.namespace + "/" + args.Name } args.ControlDependencies = append(args.ControlDependencies, s.controlDependencies...) + args.Device = s.device op, err := s.graph.AddOperation(args) if err != nil { s.UpdateErr(args.Type, err) @@ -98,10 +100,12 @@ func (s *Scope) SubScope(namespace string) *Scope { namespace = s.namespace + "/" + namespace } return &Scope{ - graph: s.graph, - namemap: make(map[string]int), - namespace: namespace, - err: s.err, + graph: s.graph, + namemap: make(map[string]int), + namespace: namespace, + controlDependencies: s.controlDependencies, + device: s.device, + err: s.err, } } @@ -123,6 +127,25 @@ func (s *Scope) WithControlDependencies(ops ...*tf.Operation) *Scope { namemap: s.namemap, namespace: s.namespace, controlDependencies: deps, + device: s.device, + err: s.err, + } +} + +// WithDevice returns a new Scope which will cause all operations added to the +// graph to execute on devices that match the provided device specification. +// +// For example, WithDevice("/device:GPU:0") will cause operations added to +// the graph to execute on GPU #0. +// +// An empty string removes any device restrictions. +func (s *Scope) WithDevice(device string) *Scope { + return &Scope{ + graph: s.graph, + namemap: s.namemap, + namespace: s.namespace, + controlDependencies: s.controlDependencies, + device: device, err: s.err, } } diff --git a/tensorflow/go/op/scope_test.go b/tensorflow/go/op/scope_test.go index b58a61de98b0f5b04959e1eca35c6b6c4d77e42b..be7b0ad8926aadac47218b7625036d7e12b9554b 100644 --- a/tensorflow/go/op/scope_test.go +++ b/tensorflow/go/op/scope_test.go @@ -112,6 +112,21 @@ func TestControlDependencies(t *testing.T) { } } +func TestDevice(t *testing.T) { + s := NewScope() + matrix := Const(s, [][]float32{{3.0}}) + s = s.WithDevice("/device:GPU:0") + square := MatMul(s.SubScope("square"), matrix, matrix) + s = s.WithDevice("") + cube := MatMul(s.SubScope("cube"), square, matrix) + if got, want := square.Op.Device(), "/device:GPU:0"; got != want { + t.Errorf("Got %q, want %q", got, want) + } + if got, want := cube.Op.Device(), ""; got != want { + t.Errorf("Got %q, want %q", got, want) + } +} + func TestScopeFinalize(t *testing.T) { var ( root = NewScope() diff --git a/tensorflow/go/op/wrappers.go b/tensorflow/go/op/wrappers.go index 7f1f0970a6fd697419b4158f3a6517bca5bbe10e..f49e1cecafe14760d9a883e1a91e72d6c76d2649 100644 --- a/tensorflow/go/op/wrappers.go +++ b/tensorflow/go/op/wrappers.go @@ -3045,25 +3045,171 @@ func UnravelIndex(scope *Scope, indices tf.Output, dims tf.Output) (output tf.Ou return op.Output(0) } -// Computes gradients for SparseSegmentSqrtN. +// Subtracts `v` into specified rows of `x`. // -// Returns tensor "output" with same shape as grad, except for dimension 0 whose -// value is output_dim0. +// Computes y = x; y[i, :] -= v; return y. // // Arguments: -// grad: gradient propagated to the SparseSegmentSqrtN op. -// indices: indices passed to the corresponding SparseSegmentSqrtN op. -// segment_ids: segment_ids passed to the corresponding SparseSegmentSqrtN op. -// output_dim0: dimension 0 of "data" passed to SparseSegmentSqrtN op. -func SparseSegmentSqrtNGrad(scope *Scope, grad tf.Output, indices tf.Output, segment_ids tf.Output, output_dim0 tf.Output) (output tf.Output) { +// x: A `Tensor` of type T. +// i: A vector. Indices into the left-most dimension of `x`. +// v: A `Tensor` of type T. Same dimension sizes as x except the first dimension, which must be the same as i's size. +// +// Returns A `Tensor` of type T. An alias of `x`. The content of `y` is undefined if there are duplicates in `i`. +func InplaceSub(scope *Scope, x tf.Output, i tf.Output, v tf.Output) (y tf.Output) { if scope.Err() != nil { return } opspec := tf.OpSpec{ - Type: "SparseSegmentSqrtNGrad", + Type: "InplaceSub", Input: []tf.Input{ - grad, indices, segment_ids, output_dim0, + x, i, v, + }, + } + op := scope.AddOperation(opspec) + return op.Output(0) +} + +// Updates specified rows with values in `v`. +// +// Computes `x[i, :] = v; return x`. +// +// Arguments: +// x: A tensor of type `T`. +// i: A vector. Indices into the left-most dimension of `x`. +// v: A `Tensor` of type T. Same dimension sizes as x except the first dimension, which must be the same as i's size. +// +// Returns A `Tensor` of type T. An alias of `x`. The content of `y` is undefined if there are duplicates in `i`. +func InplaceUpdate(scope *Scope, x tf.Output, i tf.Output, v tf.Output) (y tf.Output) { + if scope.Err() != nil { + return + } + opspec := tf.OpSpec{ + Type: "InplaceUpdate", + Input: []tf.Input{ + x, i, v, + }, + } + op := scope.AddOperation(opspec) + return op.Output(0) +} + +// Makes a copy of `x`. +// +// Arguments: +// x: The source tensor of type `T`. +// +// Returns y: A `Tensor` of type `T`. A copy of `x`. Guaranteed that `y` +// is not an alias of `x`. +func DeepCopy(scope *Scope, x tf.Output) (y tf.Output) { + if scope.Err() != nil { + return + } + opspec := tf.OpSpec{ + Type: "DeepCopy", + Input: []tf.Input{ + x, + }, + } + op := scope.AddOperation(opspec) + return op.Output(0) +} + +// PackAttr is an optional argument to Pack. +type PackAttr func(optionalAttr) + +// PackAxis sets the optional axis attribute to value. +// +// value: Dimension along which to pack. Negative values wrap around, so the +// valid range is `[-(R+1), R+1)`. +// If not specified, defaults to 0 +func PackAxis(value int64) PackAttr { + return func(m optionalAttr) { + m["axis"] = value + } +} + +// Packs a list of `N` rank-`R` tensors into one rank-`(R+1)` tensor. +// +// Packs the `N` tensors in `values` into a tensor with rank one higher than each +// tensor in `values`, by packing them along the `axis` dimension. +// Given a list of tensors of shape `(A, B, C)`; +// +// if `axis == 0` then the `output` tensor will have the shape `(N, A, B, C)`. +// if `axis == 1` then the `output` tensor will have the shape `(A, N, B, C)`. +// Etc. +// +// For example: +// +// ``` +// # 'x' is [1, 4] +// # 'y' is [2, 5] +// # 'z' is [3, 6] +// pack([x, y, z]) => [[1, 4], [2, 5], [3, 6]] # Pack along first dim. +// pack([x, y, z], axis=1) => [[1, 2, 3], [4, 5, 6]] +// ``` +// +// This is the opposite of `unpack`. +// +// Arguments: +// values: Must be of same shape and type. +// +// Returns The packed tensor. +func Pack(scope *Scope, values []tf.Output, optional ...PackAttr) (output tf.Output) { + if scope.Err() != nil { + return + } + attrs := map[string]interface{}{} + for _, a := range optional { + a(attrs) + } + opspec := tf.OpSpec{ + Type: "Pack", + Input: []tf.Input{ + tf.OutputList(values), + }, + Attrs: attrs, + } + op := scope.AddOperation(opspec) + return op.Output(0) +} + +// Concatenates a list of `N` tensors along the first dimension. +// +// The input tensors are all required to have size 1 in the first dimension. +// +// For example: +// +// ``` +// # 'x' is [[1, 4]] +// # 'y' is [[2, 5]] +// # 'z' is [[3, 6]] +// parallel_concat([x, y, z]) => [[1, 4], [2, 5], [3, 6]] # Pack along first dim. +// ``` +// +// The difference between concat and parallel_concat is that concat requires all +// of the inputs be computed before the operation will begin but doesn't require +// that the input shapes be known during graph construction. Parallel concat +// will copy pieces of the input into the output as they become available, in +// some situations this can provide a performance benefit. +// +// Arguments: +// values: Tensors to be concatenated. All must have size 1 in the first dimension +// and same shape. +// shape: the final shape of the result; should be equal to the shapes of any input +// but with the number of input values in the first dimension. +// +// Returns The concatenated tensor. +func ParallelConcat(scope *Scope, values []tf.Output, shape tf.Shape) (output tf.Output) { + if scope.Err() != nil { + return + } + attrs := map[string]interface{}{"shape": shape} + opspec := tf.OpSpec{ + Type: "ParallelConcat", + Input: []tf.Input{ + tf.OutputList(values), }, + Attrs: attrs, } op := scope.AddOperation(opspec) return op.Output(0) @@ -3121,6 +3267,57 @@ func StackPopV2(scope *Scope, handle tf.Output, elem_type tf.DataType) (elem tf. return op.Output(0) } +// Computes the sum along sparse segments of a tensor. +// +// Like `SparseSegmentSum`, but allows missing ids in `segment_ids`. If an id is +// misisng, the `output` tensor at that position will be zeroed. +// +// Read @{$math_ops#Segmentation$the section on segmentation} for an explanation of +// segments. +// +// For example: +// +// ```python +// c = tf.constant([[1,2,3,4], [-1,-2,-3,-4], [5,6,7,8]]) +// +// tf.sparse_segment_sum_with_num_segments( +// c, tf.constant([0, 1]), tf.constant([0, 0]), num_segments=3) +// # => [[0 0 0 0] +// # [0 0 0 0] +// # [0 0 0 0]] +// +// tf.sparse_segment_sum_with_num_segments(c, +// tf.constant([0, 1]), +// tf.constant([0, 2], +// num_segments=4)) +// # => [[ 1 2 3 4] +// # [ 0 0 0 0] +// # [-1 -2 -3 -4] +// # [ 0 0 0 0]] +// ``` +// +// Arguments: +// +// indices: A 1-D tensor. Has same rank as `segment_ids`. +// segment_ids: A 1-D tensor. Values should be sorted and can be repeated. +// num_segments: Should equal the number of distinct segment IDs. +// +// Returns Has same shape as data, except for dimension 0 which +// has size `num_segments`. +func SparseSegmentSumWithNumSegments(scope *Scope, data tf.Output, indices tf.Output, segment_ids tf.Output, num_segments tf.Output) (output tf.Output) { + if scope.Err() != nil { + return + } + opspec := tf.OpSpec{ + Type: "SparseSegmentSumWithNumSegments", + Input: []tf.Input{ + data, indices, segment_ids, num_segments, + }, + } + op := scope.AddOperation(opspec) + return op.Output(0) +} + // PreventGradientAttr is an optional argument to PreventGradient. type PreventGradientAttr func(optionalAttr) @@ -6071,66 +6268,19 @@ func MutexV2(scope *Scope, optional ...MutexV2Attr) (resource tf.Output) { return op.Output(0) } -// AvgPool3DAttr is an optional argument to AvgPool3D. -type AvgPool3DAttr func(optionalAttr) - -// AvgPool3DDataFormat sets the optional data_format attribute to value. -// -// value: The data format of the input and output data. With the -// default format "NDHWC", the data is stored in the order of: -// [batch, in_depth, in_height, in_width, in_channels]. -// Alternatively, the format could be "NCDHW", the data storage order is: -// [batch, in_channels, in_depth, in_height, in_width]. -// If not specified, defaults to "NDHWC" -func AvgPool3DDataFormat(value string) AvgPool3DAttr { - return func(m optionalAttr) { - m["data_format"] = value - } -} - -// Performs 3D average pooling on the input. +// Returns element-wise remainder of division. This emulates C semantics in that // -// Arguments: -// input: Shape `[batch, depth, rows, cols, channels]` tensor to pool over. -// ksize: 1-D tensor of length 5. The size of the window for each dimension of -// the input tensor. Must have `ksize[0] = ksize[4] = 1`. -// strides: 1-D tensor of length 5. The stride of the sliding window for each -// dimension of `input`. Must have `strides[0] = strides[4] = 1`. -// padding: The type of padding algorithm to use. +// the result here is consistent with a truncating divide. E.g. +// `tf.truncatediv(x, y) * y + truncate_mod(x, y) = x`. // -// Returns The average pooled output tensor. -func AvgPool3D(scope *Scope, input tf.Output, ksize []int64, strides []int64, padding string, optional ...AvgPool3DAttr) (output tf.Output) { +// *NOTE*: `Mod` supports broadcasting. More about broadcasting +// [here](http://docs.scipy.org/doc/numpy/user/basics.broadcasting.html) +func Mod(scope *Scope, x tf.Output, y tf.Output) (z tf.Output) { if scope.Err() != nil { return } - attrs := map[string]interface{}{"ksize": ksize, "strides": strides, "padding": padding} - for _, a := range optional { - a(attrs) - } opspec := tf.OpSpec{ - Type: "AvgPool3D", - Input: []tf.Input{ - input, - }, - Attrs: attrs, - } - op := scope.AddOperation(opspec) - return op.Output(0) -} - -// Returns element-wise remainder of division. This emulates C semantics in that -// -// the result here is consistent with a truncating divide. E.g. -// `tf.truncatediv(x, y) * y + truncate_mod(x, y) = x`. -// -// *NOTE*: `Mod` supports broadcasting. More about broadcasting -// [here](http://docs.scipy.org/doc/numpy/user/basics.broadcasting.html) -func Mod(scope *Scope, x tf.Output, y tf.Output) (z tf.Output) { - if scope.Err() != nil { - return - } - opspec := tf.OpSpec{ - Type: "Mod", + Type: "Mod", Input: []tf.Input{ x, y, }, @@ -7677,6 +7827,124 @@ func AccumulateNV2(scope *Scope, inputs []tf.Output, shape tf.Shape) (sum tf.Out return op.Output(0) } +// RandomShuffleAttr is an optional argument to RandomShuffle. +type RandomShuffleAttr func(optionalAttr) + +// RandomShuffleSeed sets the optional seed attribute to value. +// +// value: If either `seed` or `seed2` are set to be non-zero, the random number +// generator is seeded by the given seed. Otherwise, it is seeded by a +// random seed. +// If not specified, defaults to 0 +func RandomShuffleSeed(value int64) RandomShuffleAttr { + return func(m optionalAttr) { + m["seed"] = value + } +} + +// RandomShuffleSeed2 sets the optional seed2 attribute to value. +// +// value: A second seed to avoid seed collision. +// If not specified, defaults to 0 +func RandomShuffleSeed2(value int64) RandomShuffleAttr { + return func(m optionalAttr) { + m["seed2"] = value + } +} + +// Randomly shuffles a tensor along its first dimension. +// +// The tensor is shuffled along dimension 0, such that each `value[j]` is mapped +// to one and only one `output[i]`. For example, a mapping that might occur for a +// 3x2 tensor is: +// +// ``` +// [[1, 2], [[5, 6], +// [3, 4], ==> [1, 2], +// [5, 6]] [3, 4]] +// ``` +// +// Arguments: +// value: The tensor to be shuffled. +// +// Returns A tensor of same shape and type as `value`, shuffled along its first +// dimension. +func RandomShuffle(scope *Scope, value tf.Output, optional ...RandomShuffleAttr) (output tf.Output) { + if scope.Err() != nil { + return + } + attrs := map[string]interface{}{} + for _, a := range optional { + a(attrs) + } + opspec := tf.OpSpec{ + Type: "RandomShuffle", + Input: []tf.Input{ + value, + }, + Attrs: attrs, + } + op := scope.AddOperation(opspec) + return op.Output(0) +} + +// OrderedMapIncompleteSizeAttr is an optional argument to OrderedMapIncompleteSize. +type OrderedMapIncompleteSizeAttr func(optionalAttr) + +// OrderedMapIncompleteSizeCapacity sets the optional capacity attribute to value. +// If not specified, defaults to 0 +// +// REQUIRES: value >= 0 +func OrderedMapIncompleteSizeCapacity(value int64) OrderedMapIncompleteSizeAttr { + return func(m optionalAttr) { + m["capacity"] = value + } +} + +// OrderedMapIncompleteSizeMemoryLimit sets the optional memory_limit attribute to value. +// If not specified, defaults to 0 +// +// REQUIRES: value >= 0 +func OrderedMapIncompleteSizeMemoryLimit(value int64) OrderedMapIncompleteSizeAttr { + return func(m optionalAttr) { + m["memory_limit"] = value + } +} + +// OrderedMapIncompleteSizeContainer sets the optional container attribute to value. +// If not specified, defaults to "" +func OrderedMapIncompleteSizeContainer(value string) OrderedMapIncompleteSizeAttr { + return func(m optionalAttr) { + m["container"] = value + } +} + +// OrderedMapIncompleteSizeSharedName sets the optional shared_name attribute to value. +// If not specified, defaults to "" +func OrderedMapIncompleteSizeSharedName(value string) OrderedMapIncompleteSizeAttr { + return func(m optionalAttr) { + m["shared_name"] = value + } +} + +// Op returns the number of incomplete elements in the underlying container. +func OrderedMapIncompleteSize(scope *Scope, dtypes []tf.DataType, optional ...OrderedMapIncompleteSizeAttr) (size tf.Output) { + if scope.Err() != nil { + return + } + attrs := map[string]interface{}{"dtypes": dtypes} + for _, a := range optional { + a(attrs) + } + opspec := tf.OpSpec{ + Type: "OrderedMapIncompleteSize", + + Attrs: attrs, + } + op := scope.AddOperation(opspec) + return op.Output(0) +} + // DepthwiseConv2dNativeBackpropFilterAttr is an optional argument to DepthwiseConv2dNativeBackpropFilter. type DepthwiseConv2dNativeBackpropFilterAttr func(optionalAttr) @@ -7996,27 +8264,6 @@ func CollectiveBcastSend(scope *Scope, input tf.Output, group_size int64, group_ return op.Output(0) } -// Makes a copy of `x`. -// -// Arguments: -// x: The source tensor of type `T`. -// -// Returns y: A `Tensor` of type `T`. A copy of `x`. Guaranteed that `y` -// is not an alias of `x`. -func DeepCopy(scope *Scope, x tf.Output) (y tf.Output) { - if scope.Err() != nil { - return - } - opspec := tf.OpSpec{ - Type: "DeepCopy", - Input: []tf.Input{ - x, - }, - } - op := scope.AddOperation(opspec) - return op.Output(0) -} - // Split a `SparseTensor` into `num_split` tensors along one dimension. // // If the `shape[split_dim]` is not an integer multiple of `num_split`. Slices @@ -11210,7 +11457,7 @@ func SampleDistortedBoundingBoxAspectRatioRange(value []float32) SampleDistorted // SampleDistortedBoundingBoxAreaRange sets the optional area_range attribute to value. // // value: The cropped area of the image must contain a fraction of the -// supplied image within this range. +// supplied image within in this range. // If not specified, defaults to func SampleDistortedBoundingBoxAreaRange(value []float32) SampleDistortedBoundingBoxAttr { return func(m optionalAttr) { @@ -13013,36 +13260,75 @@ func Conv3DBackpropInput(scope *Scope, input tf.Output, filter tf.Output, out_ba return op.Output(0) } -// ResourceApplyProximalAdagradAttr is an optional argument to ResourceApplyProximalAdagrad. -type ResourceApplyProximalAdagradAttr func(optionalAttr) +// Subtracts sparse updates from the variable referenced by `resource`. +// +// This operation computes +// +// # Scalar indices +// ref[indices, ...] -= updates[...] +// +// # Vector indices (for each i) +// ref[indices[i], ...] -= updates[i, ...] +// +// # High rank indices (for each i, ..., j) +// ref[indices[i, ..., j], ...] -= updates[i, ..., j, ...] +// +// Duplicate entries are handled correctly: if multiple `indices` reference +// the same location, their contributions add. +// +// Requires `updates.shape = indices.shape + ref.shape[1:]` or `updates.shape = []`. +// +//
+// +//
+// +// Arguments: +// resource: Should be from a `Variable` node. +// indices: A tensor of indices into the first dimension of `ref`. +// updates: A tensor of updated values to add to `ref`. +// +// Returns the created operation. +func ResourceScatterSub(scope *Scope, resource tf.Output, indices tf.Output, updates tf.Output) (o *tf.Operation) { + if scope.Err() != nil { + return + } + opspec := tf.OpSpec{ + Type: "ResourceScatterSub", + Input: []tf.Input{ + resource, indices, updates, + }, + } + return scope.AddOperation(opspec) +} -// ResourceApplyProximalAdagradUseLocking sets the optional use_locking attribute to value. +// ResourceApplyProximalGradientDescentAttr is an optional argument to ResourceApplyProximalGradientDescent. +type ResourceApplyProximalGradientDescentAttr func(optionalAttr) + +// ResourceApplyProximalGradientDescentUseLocking sets the optional use_locking attribute to value. // -// value: If True, updating of the var and accum tensors will be protected by -// a lock; otherwise the behavior is undefined, but may exhibit less contention. +// value: If True, the subtraction will be protected by a lock; +// otherwise the behavior is undefined, but may exhibit less contention. // If not specified, defaults to false -func ResourceApplyProximalAdagradUseLocking(value bool) ResourceApplyProximalAdagradAttr { +func ResourceApplyProximalGradientDescentUseLocking(value bool) ResourceApplyProximalGradientDescentAttr { return func(m optionalAttr) { m["use_locking"] = value } } -// Update '*var' and '*accum' according to FOBOS with Adagrad learning rate. +// Update '*var' as FOBOS algorithm with fixed learning rate. // -// accum += grad * grad -// prox_v = var - lr * grad * (1 / sqrt(accum)) -// var = sign(prox_v)/(1+lr*l2) * max{|prox_v|-lr*l1,0} +// prox_v = var - alpha * delta +// var = sign(prox_v)/(1+alpha*l2) * max{|prox_v|-alpha*l1,0} // // Arguments: // var_: Should be from a Variable(). -// accum: Should be from a Variable(). -// lr: Scaling factor. Must be a scalar. +// alpha: Scaling factor. Must be a scalar. // l1: L1 regularization. Must be a scalar. // l2: L2 regularization. Must be a scalar. -// grad: The gradient. +// delta: The change. // // Returns the created operation. -func ResourceApplyProximalAdagrad(scope *Scope, var_ tf.Output, accum tf.Output, lr tf.Output, l1 tf.Output, l2 tf.Output, grad tf.Output, optional ...ResourceApplyProximalAdagradAttr) (o *tf.Operation) { +func ResourceApplyProximalGradientDescent(scope *Scope, var_ tf.Output, alpha tf.Output, l1 tf.Output, l2 tf.Output, delta tf.Output, optional ...ResourceApplyProximalGradientDescentAttr) (o *tf.Operation) { if scope.Err() != nil { return } @@ -13051,186 +13337,31 @@ func ResourceApplyProximalAdagrad(scope *Scope, var_ tf.Output, accum tf.Output, a(attrs) } opspec := tf.OpSpec{ - Type: "ResourceApplyProximalAdagrad", + Type: "ResourceApplyProximalGradientDescent", Input: []tf.Input{ - var_, accum, lr, l1, l2, grad, + var_, alpha, l1, l2, delta, }, Attrs: attrs, } return scope.AddOperation(opspec) } -// MutableHashTableOfTensorsV2Attr is an optional argument to MutableHashTableOfTensorsV2. -type MutableHashTableOfTensorsV2Attr func(optionalAttr) - -// MutableHashTableOfTensorsV2Container sets the optional container attribute to value. +// Computes the gradient for the sqrt of `x` wrt its input. // -// value: If non-empty, this table is placed in the given container. -// Otherwise, a default container is used. -// If not specified, defaults to "" -func MutableHashTableOfTensorsV2Container(value string) MutableHashTableOfTensorsV2Attr { - return func(m optionalAttr) { - m["container"] = value +// Specifically, `grad = dy * 0.5 / y`, where `y = sqrt(x)`, and `dy` +// is the corresponding input gradient. +func SqrtGrad(scope *Scope, y tf.Output, dy tf.Output) (z tf.Output) { + if scope.Err() != nil { + return } -} - -// MutableHashTableOfTensorsV2SharedName sets the optional shared_name attribute to value. -// -// value: If non-empty, this table is shared under the given name across -// multiple sessions. -// If not specified, defaults to "" -func MutableHashTableOfTensorsV2SharedName(value string) MutableHashTableOfTensorsV2Attr { - return func(m optionalAttr) { - m["shared_name"] = value - } -} - -// MutableHashTableOfTensorsV2UseNodeNameSharing sets the optional use_node_name_sharing attribute to value. -// If not specified, defaults to false -func MutableHashTableOfTensorsV2UseNodeNameSharing(value bool) MutableHashTableOfTensorsV2Attr { - return func(m optionalAttr) { - m["use_node_name_sharing"] = value - } -} - -// MutableHashTableOfTensorsV2ValueShape sets the optional value_shape attribute to value. -// If not specified, defaults to <> -func MutableHashTableOfTensorsV2ValueShape(value tf.Shape) MutableHashTableOfTensorsV2Attr { - return func(m optionalAttr) { - m["value_shape"] = value - } -} - -// Creates an empty hash table. -// -// This op creates a mutable hash table, specifying the type of its keys and -// values. Each value must be a vector. Data can be inserted into the table using -// the insert operations. It does not support the initialization operation. -// -// Arguments: -// key_dtype: Type of the table keys. -// value_dtype: Type of the table values. -// -// Returns Handle to a table. -func MutableHashTableOfTensorsV2(scope *Scope, key_dtype tf.DataType, value_dtype tf.DataType, optional ...MutableHashTableOfTensorsV2Attr) (table_handle tf.Output) { - if scope.Err() != nil { - return - } - attrs := map[string]interface{}{"key_dtype": key_dtype, "value_dtype": value_dtype} - for _, a := range optional { - a(attrs) - } - opspec := tf.OpSpec{ - Type: "MutableHashTableOfTensorsV2", - - Attrs: attrs, - } - op := scope.AddOperation(opspec) - return op.Output(0) -} - -// Subtracts sparse updates from the variable referenced by `resource`. -// -// This operation computes -// -// # Scalar indices -// ref[indices, ...] -= updates[...] -// -// # Vector indices (for each i) -// ref[indices[i], ...] -= updates[i, ...] -// -// # High rank indices (for each i, ..., j) -// ref[indices[i, ..., j], ...] -= updates[i, ..., j, ...] -// -// Duplicate entries are handled correctly: if multiple `indices` reference -// the same location, their contributions add. -// -// Requires `updates.shape = indices.shape + ref.shape[1:]` or `updates.shape = []`. -// -//
-// -//
-// -// Arguments: -// resource: Should be from a `Variable` node. -// indices: A tensor of indices into the first dimension of `ref`. -// updates: A tensor of updated values to add to `ref`. -// -// Returns the created operation. -func ResourceScatterSub(scope *Scope, resource tf.Output, indices tf.Output, updates tf.Output) (o *tf.Operation) { - if scope.Err() != nil { - return - } - opspec := tf.OpSpec{ - Type: "ResourceScatterSub", - Input: []tf.Input{ - resource, indices, updates, - }, - } - return scope.AddOperation(opspec) -} - -// ResourceApplyProximalGradientDescentAttr is an optional argument to ResourceApplyProximalGradientDescent. -type ResourceApplyProximalGradientDescentAttr func(optionalAttr) - -// ResourceApplyProximalGradientDescentUseLocking sets the optional use_locking attribute to value. -// -// value: If True, the subtraction will be protected by a lock; -// otherwise the behavior is undefined, but may exhibit less contention. -// If not specified, defaults to false -func ResourceApplyProximalGradientDescentUseLocking(value bool) ResourceApplyProximalGradientDescentAttr { - return func(m optionalAttr) { - m["use_locking"] = value - } -} - -// Update '*var' as FOBOS algorithm with fixed learning rate. -// -// prox_v = var - alpha * delta -// var = sign(prox_v)/(1+alpha*l2) * max{|prox_v|-alpha*l1,0} -// -// Arguments: -// var_: Should be from a Variable(). -// alpha: Scaling factor. Must be a scalar. -// l1: L1 regularization. Must be a scalar. -// l2: L2 regularization. Must be a scalar. -// delta: The change. -// -// Returns the created operation. -func ResourceApplyProximalGradientDescent(scope *Scope, var_ tf.Output, alpha tf.Output, l1 tf.Output, l2 tf.Output, delta tf.Output, optional ...ResourceApplyProximalGradientDescentAttr) (o *tf.Operation) { - if scope.Err() != nil { - return - } - attrs := map[string]interface{}{} - for _, a := range optional { - a(attrs) - } - opspec := tf.OpSpec{ - Type: "ResourceApplyProximalGradientDescent", - Input: []tf.Input{ - var_, alpha, l1, l2, delta, - }, - Attrs: attrs, - } - return scope.AddOperation(opspec) -} - -// Computes the gradient for the sqrt of `x` wrt its input. -// -// Specifically, `grad = dy * 0.5 / y`, where `y = sqrt(x)`, and `dy` -// is the corresponding input gradient. -func SqrtGrad(scope *Scope, y tf.Output, dy tf.Output) (z tf.Output) { - if scope.Err() != nil { - return - } - opspec := tf.OpSpec{ - Type: "SqrtGrad", - Input: []tf.Input{ - y, dy, - }, - } - op := scope.AddOperation(opspec) - return op.Output(0) + opspec := tf.OpSpec{ + Type: "SqrtGrad", + Input: []tf.Input{ + y, dy, + }, + } + op := scope.AddOperation(opspec) + return op.Output(0) } // Get the value of the tensor specified by its handle. @@ -13874,6 +14005,83 @@ func Minimum(scope *Scope, x tf.Output, y tf.Output) (z tf.Output) { return op.Output(0) } +// MfccAttr is an optional argument to Mfcc. +type MfccAttr func(optionalAttr) + +// MfccUpperFrequencyLimit sets the optional upper_frequency_limit attribute to value. +// +// value: The highest frequency to use when calculating the +// ceptstrum. +// If not specified, defaults to 4000 +func MfccUpperFrequencyLimit(value float32) MfccAttr { + return func(m optionalAttr) { + m["upper_frequency_limit"] = value + } +} + +// MfccLowerFrequencyLimit sets the optional lower_frequency_limit attribute to value. +// +// value: The lowest frequency to use when calculating the +// ceptstrum. +// If not specified, defaults to 20 +func MfccLowerFrequencyLimit(value float32) MfccAttr { + return func(m optionalAttr) { + m["lower_frequency_limit"] = value + } +} + +// MfccFilterbankChannelCount sets the optional filterbank_channel_count attribute to value. +// +// value: Resolution of the Mel bank used internally. +// If not specified, defaults to 40 +func MfccFilterbankChannelCount(value int64) MfccAttr { + return func(m optionalAttr) { + m["filterbank_channel_count"] = value + } +} + +// MfccDctCoefficientCount sets the optional dct_coefficient_count attribute to value. +// +// value: How many output channels to produce per time slice. +// If not specified, defaults to 13 +func MfccDctCoefficientCount(value int64) MfccAttr { + return func(m optionalAttr) { + m["dct_coefficient_count"] = value + } +} + +// Transforms a spectrogram into a form that's useful for speech recognition. +// +// Mel Frequency Cepstral Coefficients are a way of representing audio data that's +// been effective as an input feature for machine learning. They are created by +// taking the spectrum of a spectrogram (a 'cepstrum'), and discarding some of the +// higher frequencies that are less significant to the human ear. They have a long +// history in the speech recognition world, and https://en.wikipedia.org/wiki/Mel-frequency_cepstrum +// is a good resource to learn more. +// +// Arguments: +// spectrogram: Typically produced by the Spectrogram op, with magnitude_squared +// set to true. +// sample_rate: How many samples per second the source audio used. +func Mfcc(scope *Scope, spectrogram tf.Output, sample_rate tf.Output, optional ...MfccAttr) (output tf.Output) { + if scope.Err() != nil { + return + } + attrs := map[string]interface{}{} + for _, a := range optional { + a(attrs) + } + opspec := tf.OpSpec{ + Type: "Mfcc", + Input: []tf.Input{ + spectrogram, sample_rate, + }, + Attrs: attrs, + } + op := scope.AddOperation(opspec) + return op.Output(0) +} + // AudioSummaryAttr is an optional argument to AudioSummary. type AudioSummaryAttr func(optionalAttr) @@ -14292,65 +14500,6 @@ func TensorArraySplitV2(scope *Scope, handle tf.Output, value tf.Output, lengths return op.Output(0) } -// PackAttr is an optional argument to Pack. -type PackAttr func(optionalAttr) - -// PackAxis sets the optional axis attribute to value. -// -// value: Dimension along which to pack. Negative values wrap around, so the -// valid range is `[-(R+1), R+1)`. -// If not specified, defaults to 0 -func PackAxis(value int64) PackAttr { - return func(m optionalAttr) { - m["axis"] = value - } -} - -// Packs a list of `N` rank-`R` tensors into one rank-`(R+1)` tensor. -// -// Packs the `N` tensors in `values` into a tensor with rank one higher than each -// tensor in `values`, by packing them along the `axis` dimension. -// Given a list of tensors of shape `(A, B, C)`; -// -// if `axis == 0` then the `output` tensor will have the shape `(N, A, B, C)`. -// if `axis == 1` then the `output` tensor will have the shape `(A, N, B, C)`. -// Etc. -// -// For example: -// -// ``` -// # 'x' is [1, 4] -// # 'y' is [2, 5] -// # 'z' is [3, 6] -// pack([x, y, z]) => [[1, 4], [2, 5], [3, 6]] # Pack along first dim. -// pack([x, y, z], axis=1) => [[1, 2, 3], [4, 5, 6]] -// ``` -// -// This is the opposite of `unpack`. -// -// Arguments: -// values: Must be of same shape and type. -// -// Returns The packed tensor. -func Pack(scope *Scope, values []tf.Output, optional ...PackAttr) (output tf.Output) { - if scope.Err() != nil { - return - } - attrs := map[string]interface{}{} - for _, a := range optional { - a(attrs) - } - opspec := tf.OpSpec{ - Type: "Pack", - Input: []tf.Input{ - tf.OutputList(values), - }, - Attrs: attrs, - } - op := scope.AddOperation(opspec) - return op.Output(0) -} - // Reorders a SparseTensor into the canonical, row-major ordering. // // Note that by convention, all sparse ops preserve the canonical ordering along @@ -15010,30 +15159,6 @@ func Sigmoid(scope *Scope, x tf.Output) (y tf.Output) { return op.Output(0) } -// Updates specified rows with values in `v`. -// -// Computes `x[i, :] = v; return x`. -// -// Arguments: -// x: A tensor of type `T`. -// i: A vector. Indices into the left-most dimension of `x`. -// v: A `Tensor` of type T. Same dimension sizes as x except the first dimension, which must be the same as i's size. -// -// Returns A `Tensor` of type T. An alias of `x`. The content of `y` is undefined if there are duplicates in `i`. -func InplaceUpdate(scope *Scope, x tf.Output, i tf.Output, v tf.Output) (y tf.Output) { - if scope.Err() != nil { - return - } - opspec := tf.OpSpec{ - Type: "InplaceUpdate", - Input: []tf.Input{ - x, i, v, - }, - } - op := scope.AddOperation(opspec) - return op.Output(0) -} - // FusedBatchNormAttr is an optional argument to FusedBatchNorm. type FusedBatchNormAttr func(optionalAttr) @@ -17734,35 +17859,216 @@ func SparseConcat(scope *Scope, indices []tf.Output, values []tf.Output, shapes // Fingerprint64("e"), Fingerprint64("c"))) // // Arguments: -// indices: 2-D. Indices of each input `SparseTensor`. -// values: 1-D. values of each `SparseTensor`. -// shapes: 1-D. Shapes of each `SparseTensor`. -// dense_inputs: 2-D. Columns represented by dense `Tensor`. -// hashed_output: If true, returns the hash of the cross instead of the string. -// This will allow us avoiding string manipulations. -// num_buckets: It is used if hashed_output is true. -// output = hashed_value%num_buckets if num_buckets > 0 else hashed_value. -// hash_key: Specify the hash_key that will be used by the `FingerprintCat64` -// function to combine the crosses fingerprints. +// indices: 2-D. Indices of each input `SparseTensor`. +// values: 1-D. values of each `SparseTensor`. +// shapes: 1-D. Shapes of each `SparseTensor`. +// dense_inputs: 2-D. Columns represented by dense `Tensor`. +// hashed_output: If true, returns the hash of the cross instead of the string. +// This will allow us avoiding string manipulations. +// num_buckets: It is used if hashed_output is true. +// output = hashed_value%num_buckets if num_buckets > 0 else hashed_value. +// hash_key: Specify the hash_key that will be used by the `FingerprintCat64` +// function to combine the crosses fingerprints. +// +// +// +// Returns 2-D. Indices of the concatenated `SparseTensor`.1-D. Non-empty values of the concatenated or hashed +// `SparseTensor`.1-D. Shape of the concatenated `SparseTensor`. +func SparseCross(scope *Scope, indices []tf.Output, values []tf.Output, shapes []tf.Output, dense_inputs []tf.Output, hashed_output bool, num_buckets int64, hash_key int64, out_type tf.DataType, internal_type tf.DataType) (output_indices tf.Output, output_values tf.Output, output_shape tf.Output) { + if scope.Err() != nil { + return + } + attrs := map[string]interface{}{"hashed_output": hashed_output, "num_buckets": num_buckets, "hash_key": hash_key, "out_type": out_type, "internal_type": internal_type} + opspec := tf.OpSpec{ + Type: "SparseCross", + Input: []tf.Input{ + tf.OutputList(indices), tf.OutputList(values), tf.OutputList(shapes), tf.OutputList(dense_inputs), + }, + Attrs: attrs, + } + op := scope.AddOperation(opspec) + return op.Output(0), op.Output(1), op.Output(2) +} + +// ResourceApplyProximalAdagradAttr is an optional argument to ResourceApplyProximalAdagrad. +type ResourceApplyProximalAdagradAttr func(optionalAttr) + +// ResourceApplyProximalAdagradUseLocking sets the optional use_locking attribute to value. +// +// value: If True, updating of the var and accum tensors will be protected by +// a lock; otherwise the behavior is undefined, but may exhibit less contention. +// If not specified, defaults to false +func ResourceApplyProximalAdagradUseLocking(value bool) ResourceApplyProximalAdagradAttr { + return func(m optionalAttr) { + m["use_locking"] = value + } +} + +// Update '*var' and '*accum' according to FOBOS with Adagrad learning rate. +// +// accum += grad * grad +// prox_v = var - lr * grad * (1 / sqrt(accum)) +// var = sign(prox_v)/(1+lr*l2) * max{|prox_v|-lr*l1,0} +// +// Arguments: +// var_: Should be from a Variable(). +// accum: Should be from a Variable(). +// lr: Scaling factor. Must be a scalar. +// l1: L1 regularization. Must be a scalar. +// l2: L2 regularization. Must be a scalar. +// grad: The gradient. +// +// Returns the created operation. +func ResourceApplyProximalAdagrad(scope *Scope, var_ tf.Output, accum tf.Output, lr tf.Output, l1 tf.Output, l2 tf.Output, grad tf.Output, optional ...ResourceApplyProximalAdagradAttr) (o *tf.Operation) { + if scope.Err() != nil { + return + } + attrs := map[string]interface{}{} + for _, a := range optional { + a(attrs) + } + opspec := tf.OpSpec{ + Type: "ResourceApplyProximalAdagrad", + Input: []tf.Input{ + var_, accum, lr, l1, l2, grad, + }, + Attrs: attrs, + } + return scope.AddOperation(opspec) +} + +// MutableHashTableOfTensorsV2Attr is an optional argument to MutableHashTableOfTensorsV2. +type MutableHashTableOfTensorsV2Attr func(optionalAttr) + +// MutableHashTableOfTensorsV2Container sets the optional container attribute to value. +// +// value: If non-empty, this table is placed in the given container. +// Otherwise, a default container is used. +// If not specified, defaults to "" +func MutableHashTableOfTensorsV2Container(value string) MutableHashTableOfTensorsV2Attr { + return func(m optionalAttr) { + m["container"] = value + } +} + +// MutableHashTableOfTensorsV2SharedName sets the optional shared_name attribute to value. +// +// value: If non-empty, this table is shared under the given name across +// multiple sessions. +// If not specified, defaults to "" +func MutableHashTableOfTensorsV2SharedName(value string) MutableHashTableOfTensorsV2Attr { + return func(m optionalAttr) { + m["shared_name"] = value + } +} + +// MutableHashTableOfTensorsV2UseNodeNameSharing sets the optional use_node_name_sharing attribute to value. +// If not specified, defaults to false +func MutableHashTableOfTensorsV2UseNodeNameSharing(value bool) MutableHashTableOfTensorsV2Attr { + return func(m optionalAttr) { + m["use_node_name_sharing"] = value + } +} + +// MutableHashTableOfTensorsV2ValueShape sets the optional value_shape attribute to value. +// If not specified, defaults to <> +func MutableHashTableOfTensorsV2ValueShape(value tf.Shape) MutableHashTableOfTensorsV2Attr { + return func(m optionalAttr) { + m["value_shape"] = value + } +} + +// Creates an empty hash table. +// +// This op creates a mutable hash table, specifying the type of its keys and +// values. Each value must be a vector. Data can be inserted into the table using +// the insert operations. It does not support the initialization operation. +// +// Arguments: +// key_dtype: Type of the table keys. +// value_dtype: Type of the table values. +// +// Returns Handle to a table. +func MutableHashTableOfTensorsV2(scope *Scope, key_dtype tf.DataType, value_dtype tf.DataType, optional ...MutableHashTableOfTensorsV2Attr) (table_handle tf.Output) { + if scope.Err() != nil { + return + } + attrs := map[string]interface{}{"key_dtype": key_dtype, "value_dtype": value_dtype} + for _, a := range optional { + a(attrs) + } + opspec := tf.OpSpec{ + Type: "MutableHashTableOfTensorsV2", + + Attrs: attrs, + } + op := scope.AddOperation(opspec) + return op.Output(0) +} + +// Computes the gradient of the sigmoid of `x` wrt its input. +// +// Specifically, `grad = dy * y * (1 - y)`, where `y = sigmoid(x)`, and +// `dy` is the corresponding input gradient. +func SigmoidGrad(scope *Scope, y tf.Output, dy tf.Output) (z tf.Output) { + if scope.Err() != nil { + return + } + opspec := tf.OpSpec{ + Type: "SigmoidGrad", + Input: []tf.Input{ + y, dy, + }, + } + op := scope.AddOperation(opspec) + return op.Output(0) +} + +// Convert one or more images from HSV to RGB. +// +// Outputs a tensor of the same shape as the `images` tensor, containing the RGB +// value of the pixels. The output is only well defined if the value in `images` +// are in `[0,1]`. +// +// See `rgb_to_hsv` for a description of the HSV encoding. +// +// Arguments: +// images: 1-D or higher rank. HSV data to convert. Last dimension must be size 3. // +// Returns `images` converted to RGB. +func HSVToRGB(scope *Scope, images tf.Output) (output tf.Output) { + if scope.Err() != nil { + return + } + opspec := tf.OpSpec{ + Type: "HSVToRGB", + Input: []tf.Input{ + images, + }, + } + op := scope.AddOperation(opspec) + return op.Output(0) +} + +// Retrieves the tree ensemble resource stamp token, number of trees and growing statistics. // +// Arguments: +// tree_ensemble_handle: Handle to the tree ensemble. // -// Returns 2-D. Indices of the concatenated `SparseTensor`.1-D. Non-empty values of the concatenated or hashed -// `SparseTensor`.1-D. Shape of the concatenated `SparseTensor`. -func SparseCross(scope *Scope, indices []tf.Output, values []tf.Output, shapes []tf.Output, dense_inputs []tf.Output, hashed_output bool, num_buckets int64, hash_key int64, out_type tf.DataType, internal_type tf.DataType) (output_indices tf.Output, output_values tf.Output, output_shape tf.Output) { +// Returns Stamp token of the tree ensemble resource.The number of trees in the tree ensemble resource.The number of trees that were finished successfully.The number of layers we attempted to build (but not necessarily succeeded).Rank size 2 tensor that contains start and end ids of the nodes in the latest +// layer. +func BoostedTreesGetEnsembleStates(scope *Scope, tree_ensemble_handle tf.Output) (stamp_token tf.Output, num_trees tf.Output, num_finalized_trees tf.Output, num_attempted_layers tf.Output, last_layer_nodes_range tf.Output) { if scope.Err() != nil { return } - attrs := map[string]interface{}{"hashed_output": hashed_output, "num_buckets": num_buckets, "hash_key": hash_key, "out_type": out_type, "internal_type": internal_type} opspec := tf.OpSpec{ - Type: "SparseCross", + Type: "BoostedTreesGetEnsembleStates", Input: []tf.Input{ - tf.OutputList(indices), tf.OutputList(values), tf.OutputList(shapes), tf.OutputList(dense_inputs), + tree_ensemble_handle, }, - Attrs: attrs, } op := scope.AddOperation(opspec) - return op.Output(0), op.Output(1), op.Output(2) + return op.Output(0), op.Output(1), op.Output(2), op.Output(3), op.Output(4) } // Returns the element-wise min of two SparseTensors. @@ -17969,9 +18275,8 @@ func SparseFillEmptyRowsGrad(scope *Scope, reverse_index_map tf.Output, grad_val } // Computes scaled exponential linear: `scale * alpha * (exp(features) - 1)` -// if < 0, `scale * features` otherwise. // -// Assumes weights to have zero mean and variance 1.0 / fan_in. +// if < 0, `scale * features` otherwise. // // See [Self-Normalizing Neural Networks](https://arxiv.org/abs/1706.02515) func Selu(scope *Scope, features tf.Output) (activations tf.Output) { @@ -19204,119 +19509,25 @@ func RandomUniformInt(scope *Scope, shape tf.Output, minval tf.Output, maxval tf return op.Output(0) } -// RandomShuffleAttr is an optional argument to RandomShuffle. -type RandomShuffleAttr func(optionalAttr) - -// RandomShuffleSeed sets the optional seed attribute to value. -// -// value: If either `seed` or `seed2` are set to be non-zero, the random number -// generator is seeded by the given seed. Otherwise, it is seeded by a -// random seed. -// If not specified, defaults to 0 -func RandomShuffleSeed(value int64) RandomShuffleAttr { - return func(m optionalAttr) { - m["seed"] = value - } -} - -// RandomShuffleSeed2 sets the optional seed2 attribute to value. -// -// value: A second seed to avoid seed collision. -// If not specified, defaults to 0 -func RandomShuffleSeed2(value int64) RandomShuffleAttr { - return func(m optionalAttr) { - m["seed2"] = value - } -} - -// Randomly shuffles a tensor along its first dimension. -// -// The tensor is shuffled along dimension 0, such that each `value[j]` is mapped -// to one and only one `output[i]`. For example, a mapping that might occur for a -// 3x2 tensor is: +// Computes gradients for SparseSegmentSqrtN. // -// ``` -// [[1, 2], [[5, 6], -// [3, 4], ==> [1, 2], -// [5, 6]] [3, 4]] -// ``` +// Returns tensor "output" with same shape as grad, except for dimension 0 whose +// value is output_dim0. // // Arguments: -// value: The tensor to be shuffled. -// -// Returns A tensor of same shape and type as `value`, shuffled along its first -// dimension. -func RandomShuffle(scope *Scope, value tf.Output, optional ...RandomShuffleAttr) (output tf.Output) { +// grad: gradient propagated to the SparseSegmentSqrtN op. +// indices: indices passed to the corresponding SparseSegmentSqrtN op. +// segment_ids: segment_ids passed to the corresponding SparseSegmentSqrtN op. +// output_dim0: dimension 0 of "data" passed to SparseSegmentSqrtN op. +func SparseSegmentSqrtNGrad(scope *Scope, grad tf.Output, indices tf.Output, segment_ids tf.Output, output_dim0 tf.Output) (output tf.Output) { if scope.Err() != nil { return } - attrs := map[string]interface{}{} - for _, a := range optional { - a(attrs) - } opspec := tf.OpSpec{ - Type: "RandomShuffle", + Type: "SparseSegmentSqrtNGrad", Input: []tf.Input{ - value, + grad, indices, segment_ids, output_dim0, }, - Attrs: attrs, - } - op := scope.AddOperation(opspec) - return op.Output(0) -} - -// OrderedMapIncompleteSizeAttr is an optional argument to OrderedMapIncompleteSize. -type OrderedMapIncompleteSizeAttr func(optionalAttr) - -// OrderedMapIncompleteSizeCapacity sets the optional capacity attribute to value. -// If not specified, defaults to 0 -// -// REQUIRES: value >= 0 -func OrderedMapIncompleteSizeCapacity(value int64) OrderedMapIncompleteSizeAttr { - return func(m optionalAttr) { - m["capacity"] = value - } -} - -// OrderedMapIncompleteSizeMemoryLimit sets the optional memory_limit attribute to value. -// If not specified, defaults to 0 -// -// REQUIRES: value >= 0 -func OrderedMapIncompleteSizeMemoryLimit(value int64) OrderedMapIncompleteSizeAttr { - return func(m optionalAttr) { - m["memory_limit"] = value - } -} - -// OrderedMapIncompleteSizeContainer sets the optional container attribute to value. -// If not specified, defaults to "" -func OrderedMapIncompleteSizeContainer(value string) OrderedMapIncompleteSizeAttr { - return func(m optionalAttr) { - m["container"] = value - } -} - -// OrderedMapIncompleteSizeSharedName sets the optional shared_name attribute to value. -// If not specified, defaults to "" -func OrderedMapIncompleteSizeSharedName(value string) OrderedMapIncompleteSizeAttr { - return func(m optionalAttr) { - m["shared_name"] = value - } -} - -// Op returns the number of incomplete elements in the underlying container. -func OrderedMapIncompleteSize(scope *Scope, dtypes []tf.DataType, optional ...OrderedMapIncompleteSizeAttr) (size tf.Output) { - if scope.Err() != nil { - return - } - attrs := map[string]interface{}{"dtypes": dtypes} - for _, a := range optional { - a(attrs) - } - opspec := tf.OpSpec{ - Type: "OrderedMapIncompleteSize", - - Attrs: attrs, } op := scope.AddOperation(opspec) return op.Output(0) @@ -20322,100 +20533,23 @@ func QuantizedAddToutput(value tf.DataType) QuantizedAddAttr { return func(m optionalAttr) { m["Toutput"] = value } -} - -// Returns x + y element-wise, working on quantized buffers. -// -// Arguments: -// -// -// min_x: The float value that the lowest quantized `x` value represents. -// max_x: The float value that the highest quantized `x` value represents. -// min_y: The float value that the lowest quantized `y` value represents. -// max_y: The float value that the highest quantized `y` value represents. -// -// Returns The float value that the lowest quantized output value represents.The float value that the highest quantized output value represents. -// -// *NOTE*: `QuantizedAdd` supports limited forms of broadcasting. More about -// broadcasting [here](http://docs.scipy.org/doc/numpy/user/basics.broadcasting.html) -func QuantizedAdd(scope *Scope, x tf.Output, y tf.Output, min_x tf.Output, max_x tf.Output, min_y tf.Output, max_y tf.Output, optional ...QuantizedAddAttr) (z tf.Output, min_z tf.Output, max_z tf.Output) { - if scope.Err() != nil { - return - } - attrs := map[string]interface{}{} - for _, a := range optional { - a(attrs) - } - opspec := tf.OpSpec{ - Type: "QuantizedAdd", - Input: []tf.Input{ - x, y, min_x, max_x, min_y, max_y, - }, - Attrs: attrs, - } - op := scope.AddOperation(opspec) - return op.Output(0), op.Output(1), op.Output(2) -} - -// MfccAttr is an optional argument to Mfcc. -type MfccAttr func(optionalAttr) - -// MfccUpperFrequencyLimit sets the optional upper_frequency_limit attribute to value. -// -// value: The highest frequency to use when calculating the -// ceptstrum. -// If not specified, defaults to 4000 -func MfccUpperFrequencyLimit(value float32) MfccAttr { - return func(m optionalAttr) { - m["upper_frequency_limit"] = value - } -} - -// MfccLowerFrequencyLimit sets the optional lower_frequency_limit attribute to value. -// -// value: The lowest frequency to use when calculating the -// ceptstrum. -// If not specified, defaults to 20 -func MfccLowerFrequencyLimit(value float32) MfccAttr { - return func(m optionalAttr) { - m["lower_frequency_limit"] = value - } -} - -// MfccFilterbankChannelCount sets the optional filterbank_channel_count attribute to value. -// -// value: Resolution of the Mel bank used internally. -// If not specified, defaults to 40 -func MfccFilterbankChannelCount(value int64) MfccAttr { - return func(m optionalAttr) { - m["filterbank_channel_count"] = value - } -} - -// MfccDctCoefficientCount sets the optional dct_coefficient_count attribute to value. -// -// value: How many output channels to produce per time slice. -// If not specified, defaults to 13 -func MfccDctCoefficientCount(value int64) MfccAttr { - return func(m optionalAttr) { - m["dct_coefficient_count"] = value - } -} - -// Transforms a spectrogram into a form that's useful for speech recognition. -// -// Mel Frequency Cepstral Coefficients are a way of representing audio data that's -// been effective as an input feature for machine learning. They are created by -// taking the spectrum of a spectrogram (a 'cepstrum'), and discarding some of the -// higher frequencies that are less significant to the human ear. They have a long -// history in the speech recognition world, and https://en.wikipedia.org/wiki/Mel-frequency_cepstrum -// is a good resource to learn more. +} + +// Returns x + y element-wise, working on quantized buffers. // // Arguments: -// spectrogram: Typically produced by the Spectrogram op, with magnitude_squared -// set to true. -// sample_rate: How many samples per second the source audio used. -func Mfcc(scope *Scope, spectrogram tf.Output, sample_rate tf.Output, optional ...MfccAttr) (output tf.Output) { +// +// +// min_x: The float value that the lowest quantized `x` value represents. +// max_x: The float value that the highest quantized `x` value represents. +// min_y: The float value that the lowest quantized `y` value represents. +// max_y: The float value that the highest quantized `y` value represents. +// +// Returns The float value that the lowest quantized output value represents.The float value that the highest quantized output value represents. +// +// *NOTE*: `QuantizedAdd` supports limited forms of broadcasting. More about +// broadcasting [here](http://docs.scipy.org/doc/numpy/user/basics.broadcasting.html) +func QuantizedAdd(scope *Scope, x tf.Output, y tf.Output, min_x tf.Output, max_x tf.Output, min_y tf.Output, max_y tf.Output, optional ...QuantizedAddAttr) (z tf.Output, min_z tf.Output, max_z tf.Output) { if scope.Err() != nil { return } @@ -20424,14 +20558,14 @@ func Mfcc(scope *Scope, spectrogram tf.Output, sample_rate tf.Output, optional . a(attrs) } opspec := tf.OpSpec{ - Type: "Mfcc", + Type: "QuantizedAdd", Input: []tf.Input{ - spectrogram, sample_rate, + x, y, min_x, max_x, min_y, max_y, }, Attrs: attrs, } op := scope.AddOperation(opspec) - return op.Output(0) + return op.Output(0), op.Output(1), op.Output(2) } // Given a quantized tensor described by (input, input_min, input_max), outputs a @@ -21656,7 +21790,7 @@ func ImageSummaryBadColor(value tf.Tensor) ImageSummaryAttr { // generated sequentially as '*tag*/image/0', '*tag*/image/1', etc. // // The `bad_color` argument is the color to use in the generated images for -// non-finite input values. It is a `uint8` 1-D tensor of length `channels`. +// non-finite input values. It is a `unit8` 1-D tensor of length `channels`. // Each element must be in the range `[0, 255]` (It represents the value of a // pixel in the output image). Non-finite values in the input tensor are // replaced by this tensor in the output image. The default value is the color @@ -23914,71 +24048,6 @@ func DecodeGif(scope *Scope, contents tf.Output) (image tf.Output) { return op.Output(0) } -// Computes the gradient of the sigmoid of `x` wrt its input. -// -// Specifically, `grad = dy * y * (1 - y)`, where `y = sigmoid(x)`, and -// `dy` is the corresponding input gradient. -func SigmoidGrad(scope *Scope, y tf.Output, dy tf.Output) (z tf.Output) { - if scope.Err() != nil { - return - } - opspec := tf.OpSpec{ - Type: "SigmoidGrad", - Input: []tf.Input{ - y, dy, - }, - } - op := scope.AddOperation(opspec) - return op.Output(0) -} - -// Convert one or more images from HSV to RGB. -// -// Outputs a tensor of the same shape as the `images` tensor, containing the RGB -// value of the pixels. The output is only well defined if the value in `images` -// are in `[0,1]`. -// -// See `rgb_to_hsv` for a description of the HSV encoding. -// -// Arguments: -// images: 1-D or higher rank. HSV data to convert. Last dimension must be size 3. -// -// Returns `images` converted to RGB. -func HSVToRGB(scope *Scope, images tf.Output) (output tf.Output) { - if scope.Err() != nil { - return - } - opspec := tf.OpSpec{ - Type: "HSVToRGB", - Input: []tf.Input{ - images, - }, - } - op := scope.AddOperation(opspec) - return op.Output(0) -} - -// Retrieves the tree ensemble resource stamp token, number of trees and growing statistics. -// -// Arguments: -// tree_ensemble_handle: Handle to the tree ensemble. -// -// Returns Stamp token of the tree ensemble resource.The number of trees in the tree ensemble resource.The number of trees that were finished successfully.The number of layers we attempted to build (but not necessarily succeeded).Rank size 2 tensor that contains start and end ids of the nodes in the latest -// layer. -func BoostedTreesGetEnsembleStates(scope *Scope, tree_ensemble_handle tf.Output) (stamp_token tf.Output, num_trees tf.Output, num_finalized_trees tf.Output, num_attempted_layers tf.Output, last_layer_nodes_range tf.Output) { - if scope.Err() != nil { - return - } - opspec := tf.OpSpec{ - Type: "BoostedTreesGetEnsembleStates", - Input: []tf.Input{ - tree_ensemble_handle, - }, - } - op := scope.AddOperation(opspec) - return op.Output(0), op.Output(1), op.Output(2), op.Output(3), op.Output(4) -} - // Gets the next output from the given iterator. // // This operation is a synchronous version IteratorGetNext. It should only be used @@ -24049,7 +24118,7 @@ func SampleDistortedBoundingBoxV2AspectRatioRange(value []float32) SampleDistort // SampleDistortedBoundingBoxV2AreaRange sets the optional area_range attribute to value. // // value: The cropped area of the image must contain a fraction of the -// supplied image within this range. +// supplied image within in this range. // If not specified, defaults to func SampleDistortedBoundingBoxV2AreaRange(value []float32) SampleDistortedBoundingBoxV2Attr { return func(m optionalAttr) { @@ -24745,7 +24814,8 @@ type DecodeProtoV2Attr func(optionalAttr) // If not specified, defaults to "local://" func DecodeProtoV2DescriptorSource(value string) DecodeProtoV2Attr { return func(m optionalAttr) { - m["descriptor_source"] = value } + m["descriptor_source"] = value + } } // DecodeProtoV2MessageFormat sets the optional message_format attribute to value. @@ -25778,57 +25848,6 @@ func CacheDataset(scope *Scope, input_dataset tf.Output, filename tf.Output, out return op.Output(0) } -// Computes the sum along sparse segments of a tensor. -// -// Like `SparseSegmentSum`, but allows missing ids in `segment_ids`. If an id is -// misisng, the `output` tensor at that position will be zeroed. -// -// Read @{$math_ops#Segmentation$the section on segmentation} for an explanation of -// segments. -// -// For example: -// -// ```python -// c = tf.constant([[1,2,3,4], [-1,-2,-3,-4], [5,6,7,8]]) -// -// tf.sparse_segment_sum_with_num_segments( -// c, tf.constant([0, 1]), tf.constant([0, 0]), num_segments=3) -// # => [[0 0 0 0] -// # [0 0 0 0] -// # [0 0 0 0]] -// -// tf.sparse_segment_sum_with_num_segments(c, -// tf.constant([0, 1]), -// tf.constant([0, 2], -// num_segments=4)) -// # => [[ 1 2 3 4] -// # [ 0 0 0 0] -// # [-1 -2 -3 -4] -// # [ 0 0 0 0]] -// ``` -// -// Arguments: -// -// indices: A 1-D tensor. Has same rank as `segment_ids`. -// segment_ids: A 1-D tensor. Values should be sorted and can be repeated. -// num_segments: Should equal the number of distinct segment IDs. -// -// Returns Has same shape as data, except for dimension 0 which -// has size `num_segments`. -func SparseSegmentSumWithNumSegments(scope *Scope, data tf.Output, indices tf.Output, segment_ids tf.Output, num_segments tf.Output) (output tf.Output) { - if scope.Err() != nil { - return - } - opspec := tf.OpSpec{ - Type: "SparseSegmentSumWithNumSegments", - Input: []tf.Input{ - data, indices, segment_ids, num_segments, - }, - } - op := scope.AddOperation(opspec) - return op.Output(0) -} - // Creates a dataset that executes a SQL query and emits rows of the result set. // // Arguments: @@ -26443,6 +26462,53 @@ func Cross(scope *Scope, a tf.Output, b tf.Output) (product tf.Output) { return op.Output(0) } +// AvgPool3DAttr is an optional argument to AvgPool3D. +type AvgPool3DAttr func(optionalAttr) + +// AvgPool3DDataFormat sets the optional data_format attribute to value. +// +// value: The data format of the input and output data. With the +// default format "NDHWC", the data is stored in the order of: +// [batch, in_depth, in_height, in_width, in_channels]. +// Alternatively, the format could be "NCDHW", the data storage order is: +// [batch, in_channels, in_depth, in_height, in_width]. +// If not specified, defaults to "NDHWC" +func AvgPool3DDataFormat(value string) AvgPool3DAttr { + return func(m optionalAttr) { + m["data_format"] = value + } +} + +// Performs 3D average pooling on the input. +// +// Arguments: +// input: Shape `[batch, depth, rows, cols, channels]` tensor to pool over. +// ksize: 1-D tensor of length 5. The size of the window for each dimension of +// the input tensor. Must have `ksize[0] = ksize[4] = 1`. +// strides: 1-D tensor of length 5. The stride of the sliding window for each +// dimension of `input`. Must have `strides[0] = strides[4] = 1`. +// padding: The type of padding algorithm to use. +// +// Returns The average pooled output tensor. +func AvgPool3D(scope *Scope, input tf.Output, ksize []int64, strides []int64, padding string, optional ...AvgPool3DAttr) (output tf.Output) { + if scope.Err() != nil { + return + } + attrs := map[string]interface{}{"ksize": ksize, "strides": strides, "padding": padding} + for _, a := range optional { + a(attrs) + } + opspec := tf.OpSpec{ + Type: "AvgPool3D", + Input: []tf.Input{ + input, + }, + Attrs: attrs, + } + op := scope.AddOperation(opspec) + return op.Output(0) +} + // Performs a padding as a preprocess during a convolution. // // Similar to FusedResizeAndPadConv2d, this op allows for an optimized @@ -30644,69 +30710,3 @@ func DecodeWav(scope *Scope, contents tf.Output, optional ...DecodeWavAttr) (aud op := scope.AddOperation(opspec) return op.Output(0), op.Output(1) } - -// Concatenates a list of `N` tensors along the first dimension. -// -// The input tensors are all required to have size 1 in the first dimension. -// -// For example: -// -// ``` -// # 'x' is [[1, 4]] -// # 'y' is [[2, 5]] -// # 'z' is [[3, 6]] -// parallel_concat([x, y, z]) => [[1, 4], [2, 5], [3, 6]] # Pack along first dim. -// ``` -// -// The difference between concat and parallel_concat is that concat requires all -// of the inputs be computed before the operation will begin but doesn't require -// that the input shapes be known during graph construction. Parallel concat -// will copy pieces of the input into the output as they become available, in -// some situations this can provide a performance benefit. -// -// Arguments: -// values: Tensors to be concatenated. All must have size 1 in the first dimension -// and same shape. -// shape: the final shape of the result; should be equal to the shapes of any input -// but with the number of input values in the first dimension. -// -// Returns The concatenated tensor. -func ParallelConcat(scope *Scope, values []tf.Output, shape tf.Shape) (output tf.Output) { - if scope.Err() != nil { - return - } - attrs := map[string]interface{}{"shape": shape} - opspec := tf.OpSpec{ - Type: "ParallelConcat", - Input: []tf.Input{ - tf.OutputList(values), - }, - Attrs: attrs, - } - op := scope.AddOperation(opspec) - return op.Output(0) -} - -// Subtracts `v` into specified rows of `x`. -// -// Computes y = x; y[i, :] -= v; return y. -// -// Arguments: -// x: A `Tensor` of type T. -// i: A vector. Indices into the left-most dimension of `x`. -// v: A `Tensor` of type T. Same dimension sizes as x except the first dimension, which must be the same as i's size. -// -// Returns A `Tensor` of type T. An alias of `x`. The content of `y` is undefined if there are duplicates in `i`. -func InplaceSub(scope *Scope, x tf.Output, i tf.Output, v tf.Output) (y tf.Output) { - if scope.Err() != nil { - return - } - opspec := tf.OpSpec{ - Type: "InplaceSub", - Input: []tf.Input{ - x, i, v, - }, - } - op := scope.AddOperation(opspec) - return op.Output(0) -} diff --git a/tensorflow/go/operation.go b/tensorflow/go/operation.go index 25ec71870315917351d68db6a16d25fe037d543b..d6a37e0a8633f936fda7ec9612c6c097c9029c31 100644 --- a/tensorflow/go/operation.go +++ b/tensorflow/go/operation.go @@ -45,6 +45,12 @@ func (op *Operation) NumOutputs() int { return int(C.TF_OperationNumOutputs(op.c)) } +// Device returns a specification of the device on which this operation +// will be executed, or the empty string if there is no such specification. +func (op *Operation) Device() string { + return C.GoString(C.TF_OperationDevice(op.c)) +} + // OutputListSize returns the size of the list of Outputs that is produced by a // named output of op. // diff --git a/tensorflow/go/operation_test.go b/tensorflow/go/operation_test.go index 06b65bdfb7eb814a2bead191374029cc0fdf025e..4af9e33ad0aea5d269d876f154f96cbc99243cad 100644 --- a/tensorflow/go/operation_test.go +++ b/tensorflow/go/operation_test.go @@ -228,6 +228,29 @@ func TestOperationConsumers(t *testing.T) { } } +func TestOperationDevice(t *testing.T) { + graph := NewGraph() + v, err := NewTensor(float32(1.0)) + if err != nil { + t.Fatal(err) + } + op, err := graph.AddOperation(OpSpec{ + Type: "Const", + Name: "Const", + Attrs: map[string]interface{}{ + "dtype": v.DataType(), + "value": v, + }, + Device: "/device:GPU:0", + }) + if err != nil { + t.Fatal(err) + } + if got, want := op.Device(), "/device:GPU:0"; got != want { + t.Errorf("Got %q, want %q", got, want) + } +} + func forceGC() { var mem runtime.MemStats runtime.ReadMemStats(&mem) diff --git a/tensorflow/java/maven/hadoop/pom.xml b/tensorflow/java/maven/hadoop/pom.xml index 0642be06fa148933902ab450c5cf2f771e268828..7391dfb965c7af4fbd54a6fad69dde435a40812d 100644 --- a/tensorflow/java/maven/hadoop/pom.xml +++ b/tensorflow/java/maven/hadoop/pom.xml @@ -1,12 +1,30 @@ - - + 4.0.0 - TensorFlow TFRecord InputFormat/OutputFormat for Apache Hadoop + org.tensorflow hadoop jar + 1.9.0 + tensorflow-hadoop + https://www.tensorflow.org + TensorFlow TFRecord InputFormat/OutputFormat for Apache Hadoop + + + UTF-8 + 1.6 + 1.6 + 2.6.0 + 3.3.1 + 4.11 + + + + + Apache License Version 2.0 + http://www.apache.org/licenses/LICENSE-2.0.txt + + https://github.com/tensorflow/ecosystem.git @@ -14,11 +32,161 @@ scm:git:https://github.com/tensorflow/ecosystem.git - https://github.com/tensorflow/ecosystem/ - - org.tensorflow - parentpom - 1.9.0-rc0 - ../ - - \ No newline at end of file + + + + + org.apache.maven.plugins + maven-gpg-plugin + 1.5 + + + sign-artifacts + verify + + sign + + + + + + + + + org.apache.maven.plugins + maven-source-plugin + 2.2.1 + + + attach-sources + + jar-no-fork + + + + + + org.apache.maven.plugins + maven-javadoc-plugin + 2.9.1 + + + attach-javadocs + + jar + + + + + + + + + + org.tensorflow + proto + ${project.version} + + + org.apache.hadoop + hadoop-common + ${hadoop.version} + + + com.google.protobuf + protobuf-java + + + + + org.apache.hadoop + hadoop-mapreduce-client-core + ${hadoop.version} + + + com.google.protobuf + protobuf-java + + + + + com.google.protobuf + protobuf-java + ${protobuf.version} + + + junit + junit + ${junit.version} + test + + + org.apache.hadoop + hadoop-mapreduce-client-jobclient + ${hadoop.version} + test-jar + true + test + + + com.google.protobuf + protobuf-java + + + + + + + + + ossrh + + + + ossrh + https://oss.sonatype.org/content/repositories/snapshots + + + ossrh + https://oss.sonatype.org/service/local/staging/deploy/maven2/ + + + + + + org.apache.maven.plugins + maven-gpg-plugin + + + + + + bintray + + + + bintray + https://api.bintray.com/maven/google/tensorflow/tensorflow/;publish=0 + + + + + + org.apache.maven.plugins + maven-gpg-plugin + + + + + + + + + TensorFlowers + TensorFlow + http://www.tensorflow.org + + + diff --git a/tensorflow/java/maven/libtensorflow/pom.xml b/tensorflow/java/maven/libtensorflow/pom.xml index a7fa9ea5cc78f9d83cfb105f09837e958c60d5b4..d44bdf8f81f83c0d61a1a499c83481f2a74d0998 100644 --- a/tensorflow/java/maven/libtensorflow/pom.xml +++ b/tensorflow/java/maven/libtensorflow/pom.xml @@ -6,7 +6,7 @@ org.tensorflow parentpom - 1.9.0-rc1 + 1.9.0 ../ libtensorflow diff --git a/tensorflow/java/maven/libtensorflow_jni/pom.xml b/tensorflow/java/maven/libtensorflow_jni/pom.xml index 83aae29f1ea0f893c40597a1be6f77668d8206e9..e8925c6fb18c19ca65cff8a4b02239d0e2edd915 100644 --- a/tensorflow/java/maven/libtensorflow_jni/pom.xml +++ b/tensorflow/java/maven/libtensorflow_jni/pom.xml @@ -6,7 +6,7 @@ org.tensorflow parentpom - 1.9.0-rc1 + 1.9.0 ../ libtensorflow_jni diff --git a/tensorflow/java/maven/libtensorflow_jni_gpu/pom.xml b/tensorflow/java/maven/libtensorflow_jni_gpu/pom.xml index 50bd8ee5f9e6d268976540ca8180380447bc8f18..3bf4a2590cbe365bb3fd6972e00ddc5189758f01 100644 --- a/tensorflow/java/maven/libtensorflow_jni_gpu/pom.xml +++ b/tensorflow/java/maven/libtensorflow_jni_gpu/pom.xml @@ -6,7 +6,7 @@ org.tensorflow parentpom - 1.9.0-rc1 + 1.9.0 ../ libtensorflow_jni_gpu diff --git a/tensorflow/java/maven/pom.xml b/tensorflow/java/maven/pom.xml index b4746794ea9e417bb0bb9253ca356976a48eb1e8..b96dcf2888e4f4e587b27d246099d7d93b416645 100644 --- a/tensorflow/java/maven/pom.xml +++ b/tensorflow/java/maven/pom.xml @@ -6,7 +6,7 @@ 4.0.0 org.tensorflow parentpom - 1.9.0-rc1 + 1.9.0 pom https://www.tensorflow.org diff --git a/tensorflow/java/maven/proto/pom.xml b/tensorflow/java/maven/proto/pom.xml index 618a2a124c77240b0a2b65f33577a6330929ae83..5581d864d70e63d7121f6db912bae85f89cfdc9a 100644 --- a/tensorflow/java/maven/proto/pom.xml +++ b/tensorflow/java/maven/proto/pom.xml @@ -6,7 +6,7 @@ org.tensorflow parentpom - 1.9.0-rc1 + 1.9.0 ../ proto diff --git a/tensorflow/java/maven/run_inside_container.sh b/tensorflow/java/maven/run_inside_container.sh index 2e771064e4a9a9ca4c50f5ecf8cae91cf8b5ce6c..2240d6b7b9957b480cf2053ecb65487fa64bbf08 100644 --- a/tensorflow/java/maven/run_inside_container.sh +++ b/tensorflow/java/maven/run_inside_container.sh @@ -203,7 +203,10 @@ download_tf_ecosystem() { cd "${ECOSYSTEM_DIR}" git clone "${TF_ECOSYSTEM_URL}" cd ecosystem - git checkout r${TF_VERSION} + # TF_VERSION is a semver string (..[-suffix]) + # but the branch is just (r.). + RELEASE_BRANCH=$(echo "${TF_VERSION}" | sed -e 's/\([0-9]\+\.[0-9]\+\)\.[0-9]\+.*/\1/') + git checkout r${RELEASE_BRANCH} # Copy the TensorFlow Hadoop source cp -r "${ECOSYSTEM_DIR}/ecosystem/hadoop/src" "${HADOOP_DIR}" diff --git a/tensorflow/java/maven/spark-connector/pom.xml b/tensorflow/java/maven/spark-connector/pom.xml index 19c752d08be1deec40042bc1aa8fd1159b2f2be9..64956be02c18464b10bfee5fd8c3bc83007c30c9 100644 --- a/tensorflow/java/maven/spark-connector/pom.xml +++ b/tensorflow/java/maven/spark-connector/pom.xml @@ -1,12 +1,23 @@ - - + + 4.0.0 - TensorFlow TFRecord connector for Apache Spark DataFrames - spark-connector + org.tensorflow + spark-connector_2.11 jar + 1.9.0 + spark-tensorflow-connector + https://www.tensorflow.org + TensorFlow TFRecord connector for Apache Spark DataFrames + + + + The Apache Software License, Version 2.0 + http://www.apache.org/licenses/LICENSE-2.0.txt + repo + + https://github.com/tensorflow/ecosystem.git @@ -14,11 +25,325 @@ scm:git:https://github.com/tensorflow/ecosystem.git - https://github.com/tensorflow/ecosystem/ - - org.tensorflow - parentpom - 1.9.0-rc0 - ../ - - \ No newline at end of file + + UTF-8 + 3.2.2 + 2.11 + 1.0 + 2.2.6 + 3.0 + 1.8 + 2.3.0 + 2.7.3 + 4.11 + + + + + + + true + net.alchim31.maven + scala-maven-plugin + ${scala.maven.version} + + + compile + + add-source + compile + + + + -Xms256m + -Xmx512m + + + -g:vars + -deprecation + -feature + -unchecked + -Xfatal-warnings + -language:implicitConversions + -language:existentials + + + + + test + + add-source + testCompile + + + + attach-javadocs + + doc-jar + + + + + incremental + true + ${scala.binary.version} + false + + + + true + org.scalatest + scalatest-maven-plugin + ${scalatest.maven.version} + + + scalaTest + test + + test + + + + + + + maven-shade-plugin + 3.1.0 + + + package + + shade + + + true + + + com.google.protobuf:protobuf-java + org.tensorflow:hadoop + org.tensorflow:proto + + + + + + com.google.protobuf:protobuf-java + + **/*.java + + + + + + com.google.protobuf + + org.tensorflow.spark.shaded.com.google.protobuf + + + + + + + + + + org.apache.maven.plugins + maven-gpg-plugin + 1.5 + + + sign-artifacts + verify + + sign + + + + + + + + + net.alchim31.maven + scala-maven-plugin + + + org.apache.maven.plugins + maven-shade-plugin + + + org.scalatest + scalatest-maven-plugin + + + org.apache.maven.plugins + maven-compiler-plugin + ${maven.compiler.version} + + ${java.version} + ${java.version} + + + + org.apache.maven.plugins + maven-source-plugin + 2.2.1 + + + attach-sources + + jar-no-fork + + + + + + org.apache.maven.plugins + maven-javadoc-plugin + 2.9.1 + + + attach-javadocs + + jar + + + + + + + + + + test + + true + + !NEVERSETME + + + + + + net.alchim31.maven + scala-maven-plugin + + + + + + + org.scalatest + scalatest_${scala.binary.version} + ${scala.test.version} + test + + + + + + org.scalatest + scalatest_${scala.binary.version} + test + + + + + + + ossrh + + + + ossrh + https://oss.sonatype.org/content/repositories/snapshots + + + ossrh + https://oss.sonatype.org/service/local/staging/deploy/maven2/ + + + + + + org.apache.maven.plugins + maven-gpg-plugin + + + + + + bintray + + + + bintray + https://api.bintray.com/maven/google/tensorflow/tensorflow/;publish=0 + + + + + + org.apache.maven.plugins + maven-gpg-plugin + + + + + + + + + TensorFlowers + TensorFlow + http://www.tensorflow.org + + + + + + org.tensorflow + hadoop + ${project.version} + + + org.apache.spark + spark-core_${scala.binary.version} + ${spark.version} + provided + + + org.apache.spark + spark-sql_${scala.binary.version} + ${spark.version} + provided + + + org.apache.spark + spark-mllib_${scala.binary.version} + ${spark.version} + provided + + + org.apache.hadoop + hadoop-yarn-api + ${yarn.api.version} + provided + + + org.apache.spark + spark-mllib_${scala.binary.version} + ${spark.version} + test-jar + test + + + junit + junit + ${junit.version} + test + + + diff --git a/tensorflow/java/maven/tensorflow/pom.xml b/tensorflow/java/maven/tensorflow/pom.xml index 157c4b8e82d6b8062ce8c9c98432cfe97a20d190..92e15aa2c7082f03fb542eba8b2bf222174e8ea2 100644 --- a/tensorflow/java/maven/tensorflow/pom.xml +++ b/tensorflow/java/maven/tensorflow/pom.xml @@ -6,7 +6,7 @@ org.tensorflow parentpom - 1.9.0-rc1 + 1.9.0 ../ tensorflow diff --git a/tensorflow/java/src/gen/cc/java_defs.h b/tensorflow/java/src/gen/cc/java_defs.h index f5f54bf4d31af159624c668f1abb106f68944737..d9d6f8adc8ac9e58dbfe3609171803b55e76e42d 100644 --- a/tensorflow/java/src/gen/cc/java_defs.h +++ b/tensorflow/java/src/gen/cc/java_defs.h @@ -16,9 +16,9 @@ limitations under the License. #ifndef TENSORFLOW_JAVA_SRC_GEN_CC_JAVA_DEFS_H_ #define TENSORFLOW_JAVA_SRC_GEN_CC_JAVA_DEFS_H_ -#include #include #include +#include #include namespace tensorflow { diff --git a/tensorflow/java/src/gen/cc/op_generator.cc b/tensorflow/java/src/gen/cc/op_generator.cc index 2df69ee29996304569320c1dbbcaa46f214d4ea0..d5bd99bdd9d71f73288661380ec45e76c797fa75 100644 --- a/tensorflow/java/src/gen/cc/op_generator.cc +++ b/tensorflow/java/src/gen/cc/op_generator.cc @@ -36,20 +36,21 @@ namespace java { namespace { constexpr const char kLicense[] = - "/* Copyright 2018 The TensorFlow Authors. All Rights Reserved.\n" - "\n" - "Licensed under the Apache License, Version 2.0 (the \"License\");\n" - "you may not use this file except in compliance with the License.\n" - "You may obtain a copy of the License at\n" - "\n" - " http://www.apache.org/licenses/LICENSE-2.0\n" - "\n" - "Unless required by applicable law or agreed to in writing, software\n" - "distributed under the License is distributed on an \"AS IS\" BASIS,\n" - "WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n" - "See the License for the specific language governing permissions and\n" - "limitations under the License.\n" - "=======================================================================*/\n"; + "/* Copyright 2018 The TensorFlow Authors. All Rights Reserved.\n" + "\n" + "Licensed under the Apache License, Version 2.0 (the \"License\");\n" + "you may not use this file except in compliance with the License.\n" + "You may obtain a copy of the License at\n" + "\n" + " http://www.apache.org/licenses/LICENSE-2.0\n" + "\n" + "Unless required by applicable law or agreed to in writing, software\n" + "distributed under the License is distributed on an \"AS IS\" BASIS,\n" + "WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n" + "See the License for the specific language governing permissions and\n" + "limitations under the License.\n" + "=======================================================================*/" + "\n"; // There is three different modes to render an op class, depending on the // number and type of outputs it has: diff --git a/tensorflow/java/src/gen/cc/op_generator.h b/tensorflow/java/src/gen/cc/op_generator.h index 759d800ecfb5bec10b7bf8454baf5fc4c389e990..05decd6b54944f18205cce4d2341d7009ce7d806 100644 --- a/tensorflow/java/src/gen/cc/op_generator.h +++ b/tensorflow/java/src/gen/cc/op_generator.h @@ -19,10 +19,10 @@ limitations under the License. #include #include -#include "tensorflow/core/framework/op_def.pb.h" #include "tensorflow/core/framework/api_def.pb.h" #include "tensorflow/core/framework/op_def.pb.h" #include "tensorflow/core/lib/core/status.h" +#include "tensorflow/core/platform/env.h" #include "tensorflow/java/src/gen/cc/op_specs.h" namespace tensorflow { diff --git a/tensorflow/java/src/gen/cc/op_specs.cc b/tensorflow/java/src/gen/cc/op_specs.cc index 63e99fbb04fd6ba34f2bbd2bc3fe7644a31ddf7f..941ab2699cb887375987f14200664b9bfaf6815a 100644 --- a/tensorflow/java/src/gen/cc/op_specs.cc +++ b/tensorflow/java/src/gen/cc/op_specs.cc @@ -14,9 +14,9 @@ limitations under the License. ==============================================================================*/ #include -#include #include #include +#include #include "re2/re2.h" #include "tensorflow/core/framework/op.h" @@ -50,7 +50,7 @@ class TypeResolver { // For example, if the argument's datatype is DT_STRING, this method will // return "java.lang.String", so the argument can become "Operand" // in the Ops API - Type TypeOf(const OpDef_ArgDef& arg_def, bool *iterable_out); + Type TypeOf(const OpDef_ArgDef& arg_def, bool* iterable_out); // Returns types of an input attribute // @@ -62,7 +62,7 @@ class TypeResolver { // , so the attribute can be used as a "Float" object // in the Ops API and casted to a "float" when passing through the JNI layer. std::pair TypesOf(const OpDef_AttrDef& attr_def, - bool *iterable_out); + bool* iterable_out); // Returns true if the type of this attribute has already been resolved bool IsAttributeVisited(const string& attr_name) { @@ -89,8 +89,7 @@ class TypeResolver { } }; -Type TypeResolver::TypeOf(const OpDef_ArgDef& arg_def, - bool* iterable_out) { +Type TypeResolver::TypeOf(const OpDef_ArgDef& arg_def, bool* iterable_out) { *iterable_out = false; if (!arg_def.number_attr().empty()) { // when number_attr is set, argument has to be a list of tensors @@ -154,13 +153,13 @@ Type TypeResolver::TypeOf(const OpDef_ArgDef& arg_def, } else { LOG(FATAL) << "Cannot resolve data type of argument \"" << arg_def.name() - << "\" in operation \"" << op_def_.name() << "\""; + << "\" in operation \"" << op_def_.name() << "\""; } return type; } std::pair TypeResolver::TypesOf(const OpDef_AttrDef& attr_def, - bool* iterable_out) { + bool* iterable_out) { std::pair types = MakeTypePair(Type::Wildcard()); *iterable_out = false; StringPiece attr_type = attr_def.type(); @@ -185,7 +184,7 @@ std::pair TypeResolver::TypesOf(const OpDef_AttrDef& attr_def, } else if (attr_type == "tensor") { types = MakeTypePair(Type::Class("Tensor", "org.tensorflow") - .add_parameter(Type::Wildcard())); + .add_parameter(Type::Wildcard())); } else if (attr_type == "type") { Type type = *iterable_out ? Type::Wildcard() : NextGeneric(); @@ -196,7 +195,7 @@ std::pair TypeResolver::TypesOf(const OpDef_AttrDef& attr_def, } else { LOG(FATAL) << "Cannot resolve data type for attribute \"" << attr_type - << "\" in operation \"" << op_def_.name() << "\""; + << "\" in operation \"" << op_def_.name() << "\""; } visited_attrs_.insert(std::make_pair(attr_def.name(), types.first)); return types; @@ -219,47 +218,43 @@ string SnakeToCamelCase(const string& str, bool upper = false) { return result; } -bool FindAndCut(re2::StringPiece* input, const RE2& expr, - re2::StringPiece* before_match, re2::StringPiece* ret_match = nullptr) { - re2::StringPiece match; - if (!expr.Match(*input, 0, input->size(), RE2::UNANCHORED, &match, 1)) { - return false; - } - before_match->set(input->data(), match.begin() - input->begin()); - input->remove_prefix(match.end() - before_match->begin()); - if (ret_match != nullptr) { - *ret_match = match; - } +bool FindAndCut(string* input, const RE2& expr, string* before_match, + string* ret_match = nullptr) { + string match; + if (!RE2::PartialMatch(*input, expr, &match)) return false; + *before_match = input->substr(0, input->find(match)); + *input = input->substr(before_match->size() + match.size()); + if (ret_match != nullptr) *ret_match = match; return true; } -string ParseDocumentation(re2::StringPiece input) { +string ParseDocumentation(const string& inp) { std::stringstream javadoc_text; // TODO(karllessard) This is a very minimalist utility method for converting // markdown syntax, as found in ops descriptions, to Javadoc/html tags. Check // for alternatives to increase the level of support for markups. std::vector markups_subexpr; - markups_subexpr.push_back("\n+\\*\\s+"); // lists - markups_subexpr.push_back("\n{2,}"); // paragraphs + markups_subexpr.push_back("\n+\\*\\s+"); // lists + markups_subexpr.push_back("\n{2,}"); // paragraphs markups_subexpr.push_back("`{3,}\\s*[^\\s\n]*\\s*\n"); // code blocks - markups_subexpr.push_back("`+"); // inlined code and code blocks + markups_subexpr.push_back("`+"); // inlined code and code blocks markups_subexpr.push_back("\\*{1,2}\\b"); // text emphasis - markups_subexpr.push_back("\\["); // hyperlinks - const RE2 markup_expr(str_util::Join(markups_subexpr, "|")); + markups_subexpr.push_back("\\["); // hyperlinks + const RE2 markup_expr("(" + str_util::Join(markups_subexpr, "|") + ")"); bool in_list = false; + string input = inp; while (true) { - re2::StringPiece text; - re2::StringPiece markup; + string text, markup; if (!FindAndCut(&input, markup_expr, &text, &markup)) { javadoc_text << input; break; // end of loop } javadoc_text << text; - if (markup.starts_with("\n")) { + if (str_util::StartsWith(markup, "\n")) { javadoc_text << "\n"; - if (markup.contains("*")) { + if (str_util::StrContains(markup, "*")) { // new list item javadoc_text << (in_list ? "\n" : "
    \n") << "
  • \n"; in_list = true; @@ -267,18 +262,18 @@ string ParseDocumentation(re2::StringPiece input) { // end of list javadoc_text << "
    • \n
\n"; in_list = false; - } else if (!input.starts_with("```")) { + } else if (!str_util::StartsWith(input, "```")) { // new paragraph (not required if a 
 block follows)
         javadoc_text << "
\n";
       }
-    } else if (markup.starts_with("```")) {
+    } else if (str_util::StartsWith(markup, "```")) {
       // code blocks
-      if (FindAndCut(&input, "```\\s*\n*", &text)) {
+      if (FindAndCut(&input, "(```\\s*\n*)", &text)) {
         javadoc_text << "
{@code\n" << text << "}
\n";
       } else {
         javadoc_text << markup;
       }
-    } else if (markup.starts_with("`")) {
+    } else if (str_util::StartsWith("(" + markup + ")", "`")) {
       // inlined code
       if (FindAndCut(&input, markup, &text)) {
         javadoc_text << "{@code " << text << "}";
@@ -287,26 +282,28 @@ string ParseDocumentation(re2::StringPiece input) {
       }
     } else if (markup == "**") {
       // text emphasis (strong)
-      if (FindAndCut(&input, "\\b\\*{2}", &text)) {
+      if (FindAndCut(&input, "(\\b\\*{2})", &text)) {
         javadoc_text << "" << ParseDocumentation(text) << "";
       } else {
         javadoc_text << markup;
       }
     } else if (markup == "*") {
       // text emphasis (normal)
-      if (FindAndCut(&input, "\\b\\*{1}", &text)) {
+      if (FindAndCut(&input, "(\\b\\*{1})", &text)) {
         javadoc_text << "" << ParseDocumentation(text) << "";
       } else {
         javadoc_text << markup;
       }
-    } else if (markup.starts_with("[")) {
+    } else if (str_util::StartsWith(markup, "[")) {
       // hyperlinks
       string label;
       string link;
-      if (RE2::Consume(&input, "([^\\[]+)\\]\\((http.+)\\)", &label, &link)) {
+      if (RE2::PartialMatch(input, "([^\\[]+)\\]\\((http.+)\\)", &label,
+                            &link) &&
+          str_util::StartsWith(input, label + link)) {
+        input = input.substr(label.size() + link.size());
         javadoc_text << ""
-            << ParseDocumentation(label)
-            << "";
+                     << ParseDocumentation(label) << "";
       } else {
         javadoc_text << markup;
       }
@@ -319,57 +316,56 @@ string ParseDocumentation(re2::StringPiece input) {
 }
 
 ArgumentSpec CreateInput(const OpDef_ArgDef& input_def,
-    const ApiDef::Arg& input_api_def, TypeResolver* type_resolver) {
+                         const ApiDef::Arg& input_api_def,
+                         TypeResolver* type_resolver) {
   bool iterable = false;
   Type type = type_resolver->TypeOf(input_def, &iterable);
-  Type var_type = Type::Interface("Operand", "org.tensorflow")
-    .add_parameter(type);
+  Type var_type =
+      Type::Interface("Operand", "org.tensorflow").add_parameter(type);
   if (iterable) {
     var_type = Type::IterableOf(var_type);
   }
-  return ArgumentSpec(input_api_def.name(),
+  return ArgumentSpec(
+      input_api_def.name(),
       Variable::Create(SnakeToCamelCase(input_api_def.rename_to()), var_type),
-      type,
-      ParseDocumentation(input_api_def.description()),
-      iterable);
+      type, ParseDocumentation(input_api_def.description()), iterable);
 }
 
 AttributeSpec CreateAttribute(const OpDef_AttrDef& attr_def,
-    const ApiDef::Attr& attr_api_def, TypeResolver* type_resolver) {
+                              const ApiDef::Attr& attr_api_def,
+                              TypeResolver* type_resolver) {
   bool iterable = false;
   std::pair types = type_resolver->TypesOf(attr_def, &iterable);
-  Type var_type = types.first.kind() == Type::GENERIC ?
-      Type::Class("Class").add_parameter(types.first) : types.first;
+  Type var_type = types.first.kind() == Type::GENERIC
+                      ? Type::Class("Class").add_parameter(types.first)
+                      : types.first;
   if (iterable) {
     var_type = Type::ListOf(var_type);
   }
-  return AttributeSpec(attr_api_def.name(),
+  return AttributeSpec(
+      attr_api_def.name(),
       Variable::Create(SnakeToCamelCase(attr_api_def.rename_to()), var_type),
-      types.first,
-      types.second,
-      ParseDocumentation(attr_api_def.description()),
-      iterable,
-      attr_api_def.has_default_value());
+      types.first, types.second, ParseDocumentation(attr_api_def.description()),
+      iterable, attr_api_def.has_default_value());
 }
 
 ArgumentSpec CreateOutput(const OpDef_ArgDef& output_def,
-    const ApiDef::Arg& output_api, TypeResolver* type_resolver) {
+                          const ApiDef::Arg& output_api,
+                          TypeResolver* type_resolver) {
   bool iterable = false;
   Type type = type_resolver->TypeOf(output_def, &iterable);
-  Type var_type = Type::Class("Output", "org.tensorflow")
-    .add_parameter(type);
+  Type var_type = Type::Class("Output", "org.tensorflow").add_parameter(type);
   if (iterable) {
     var_type = Type::ListOf(var_type);
   }
-  return ArgumentSpec(output_api.name(),
+  return ArgumentSpec(
+      output_api.name(),
       Variable::Create(SnakeToCamelCase(output_api.rename_to()), var_type),
-      type,
-      ParseDocumentation(output_api.description()),
-      iterable);
+      type, ParseDocumentation(output_api.description()), iterable);
 }
 
 EndpointSpec CreateEndpoint(const OpDef& op_def, const ApiDef& api_def,
-    const ApiDef_Endpoint& endpoint_def) {
+                            const ApiDef_Endpoint& endpoint_def) {
   std::vector name_tokens = str_util::Split(endpoint_def.name(), ".");
   string package;
   string name;
@@ -377,27 +373,25 @@ EndpointSpec CreateEndpoint(const OpDef& op_def, const ApiDef& api_def,
     package = name_tokens.at(0);
     name = name_tokens.at(1);
   } else {
-    package = kDefaultEndpointPackage;
+    package = "core";  // generate unclassified ops in the 'core' package
     name = name_tokens.at(0);
   }
-  return EndpointSpec(package,
-      name,
-      Javadoc::Create(ParseDocumentation(api_def.summary()))
-          .details(ParseDocumentation(api_def.description())));
+  return EndpointSpec(package, name,
+                      Javadoc::Create(ParseDocumentation(api_def.summary()))
+                          .details(ParseDocumentation(api_def.description())));
 }
 
 }  // namespace
 
 OpSpec OpSpec::Create(const OpDef& op_def, const ApiDef& api_def) {
-  OpSpec op(api_def.graph_op_name(),
-      api_def.visibility() == ApiDef::HIDDEN,
-      op_def.deprecation().explanation());
+  OpSpec op(api_def.graph_op_name(), api_def.visibility() == ApiDef::HIDDEN,
+            op_def.deprecation().explanation());
   TypeResolver type_resolver(op_def);
   for (const string& next_input_name : api_def.arg_order()) {
     for (int i = 0; i < op_def.input_arg().size(); ++i) {
       if (op_def.input_arg(i).name() == next_input_name) {
         op.inputs_.push_back(CreateInput(op_def.input_arg(i), api_def.in_arg(i),
-            &type_resolver));
+                                         &type_resolver));
         break;
       }
     }
@@ -406,8 +400,8 @@ OpSpec OpSpec::Create(const OpDef& op_def, const ApiDef& api_def) {
     // do not parse attributes already visited, they have probably been inferred
     // before as an input argument type
     if (!type_resolver.IsAttributeVisited(op_def.attr(i).name())) {
-      AttributeSpec attr = CreateAttribute(op_def.attr(i), api_def.attr(i),
-          &type_resolver);
+      AttributeSpec attr =
+          CreateAttribute(op_def.attr(i), api_def.attr(i), &type_resolver);
       // attributes with a default value are optional
       if (attr.has_default_value() && attr.type().kind() != Type::GENERIC) {
         op.optional_attributes_.push_back(attr);
@@ -417,8 +411,8 @@ OpSpec OpSpec::Create(const OpDef& op_def, const ApiDef& api_def) {
     }
   }
   for (int i = 0; i < op_def.output_arg().size(); ++i) {
-    op.outputs_.push_back(CreateOutput(op_def.output_arg(i), api_def.out_arg(i),
-        &type_resolver));
+    op.outputs_.push_back(
+        CreateOutput(op_def.output_arg(i), api_def.out_arg(i), &type_resolver));
   }
   for (const auto& endpoint_def : api_def.endpoint()) {
     op.endpoints_.push_back(CreateEndpoint(op_def, api_def, endpoint_def));
diff --git a/tensorflow/java/src/gen/cc/op_specs.h b/tensorflow/java/src/gen/cc/op_specs.h
index 3b53c730df23c6f81f968f09b9d145a8efa1030a..30ecb8ce53d15372606981639183d3ba0e4466a4 100644
--- a/tensorflow/java/src/gen/cc/op_specs.h
+++ b/tensorflow/java/src/gen/cc/op_specs.h
@@ -19,9 +19,9 @@ limitations under the License.
 #include 
 #include 
 
-#include "tensorflow/core/framework/op_def.pb.h"
 #include "tensorflow/core/framework/api_def.pb.h"
 #include "tensorflow/core/framework/attr_value.pb.h"
+#include "tensorflow/core/framework/op_def.pb.h"
 #include "tensorflow/java/src/gen/cc/java_defs.h"
 
 namespace tensorflow {
@@ -38,9 +38,8 @@ class EndpointSpec {
   // javadoc: the endpoint class documentation
   // TODO(annarev): hardcode depcreated to false until deprecated is possible
   EndpointSpec(const string& package, const string& name,
-      const Javadoc& javadoc)
-    : package_(package), name_(name), javadoc_(javadoc),
-      deprecated_(false) {}
+               const Javadoc& javadoc)
+      : package_(package), name_(name), javadoc_(javadoc), deprecated_(false) {}
 
   const string& package() const { return package_; }
   const string& name() const { return name_; }
@@ -63,10 +62,13 @@ class ArgumentSpec {
   // type: the tensor type of this argument
   // description: a description of this argument, in javadoc
   // iterable: true if this argument is a list
-  ArgumentSpec(const string& op_def_name, const Variable& var,
-      const Type& type, const string& description, bool iterable)
-    : op_def_name_(op_def_name), var_(var), type_(type),
-      description_(description), iterable_(iterable) {}
+  ArgumentSpec(const string& op_def_name, const Variable& var, const Type& type,
+               const string& description, bool iterable)
+      : op_def_name_(op_def_name),
+        var_(var),
+        type_(type),
+        description_(description),
+        iterable_(iterable) {}
 
   const string& op_def_name() const { return op_def_name_; }
   const Variable& var() const { return var_; }
@@ -94,11 +96,16 @@ class AttributeSpec {
   // iterable: true if this attribute is a list
   // has_default_value: true if this attribute has a default value if not set
   AttributeSpec(const string& op_def_name, const Variable& var,
-      const Type& type, const Type& jni_type, const string& description,
-      bool iterable, bool has_default_value)
-    : op_def_name_(op_def_name), var_(var), type_(type),
-      description_(description), iterable_(iterable),
-      jni_type_(jni_type), has_default_value_(has_default_value) {}
+                const Type& type, const Type& jni_type,
+                const string& description, bool iterable,
+                bool has_default_value)
+      : op_def_name_(op_def_name),
+        var_(var),
+        type_(type),
+        description_(description),
+        iterable_(iterable),
+        jni_type_(jni_type),
+        has_default_value_(has_default_value) {}
 
   const string& op_def_name() const { return op_def_name_; }
   const Variable& var() const { return var_; }
@@ -147,9 +154,10 @@ class OpSpec {
   // hidden: true if this op should not be visible through the Graph Ops API
   // deprecation_explanation: message to show if all endpoints are deprecated
   explicit OpSpec(const string& graph_op_name, bool hidden,
-      const string& deprecation_explanation)
-    : graph_op_name_(graph_op_name), hidden_(hidden),
-      deprecation_explanation_(deprecation_explanation) {}
+                  const string& deprecation_explanation)
+      : graph_op_name_(graph_op_name),
+        hidden_(hidden),
+        deprecation_explanation_(deprecation_explanation) {}
 
   const string graph_op_name_;
   const bool hidden_;
diff --git a/tensorflow/java/src/gen/java/org/tensorflow/processor/OperatorProcessor.java b/tensorflow/java/src/gen/java/org/tensorflow/processor/OperatorProcessor.java
index 3524160d876ac89306203891357f27946d9e368f..796d6a62dcf8551d8d68d9ff62077e7f09db4401 100644
--- a/tensorflow/java/src/gen/java/org/tensorflow/processor/OperatorProcessor.java
+++ b/tensorflow/java/src/gen/java/org/tensorflow/processor/OperatorProcessor.java
@@ -15,6 +15,18 @@ limitations under the License.
 
 package org.tensorflow.processor;
 
+import com.google.common.base.CaseFormat;
+import com.google.common.base.Strings;
+import com.google.common.collect.HashMultimap;
+import com.google.common.collect.Multimap;
+import com.squareup.javapoet.ClassName;
+import com.squareup.javapoet.FieldSpec;
+import com.squareup.javapoet.JavaFile;
+import com.squareup.javapoet.MethodSpec;
+import com.squareup.javapoet.ParameterSpec;
+import com.squareup.javapoet.TypeName;
+import com.squareup.javapoet.TypeSpec;
+import com.squareup.javapoet.TypeVariableName;
 import java.io.IOException;
 import java.util.Collection;
 import java.util.Collections;
@@ -23,7 +35,6 @@ import java.util.Map;
 import java.util.Set;
 import java.util.regex.Matcher;
 import java.util.regex.Pattern;
-
 import javax.annotation.processing.AbstractProcessor;
 import javax.annotation.processing.Filer;
 import javax.annotation.processing.Messager;
@@ -44,19 +55,6 @@ import javax.lang.model.util.ElementFilter;
 import javax.lang.model.util.Elements;
 import javax.tools.Diagnostic.Kind;
 
-import com.google.common.base.CaseFormat;
-import com.google.common.base.Strings;
-import com.google.common.collect.HashMultimap;
-import com.google.common.collect.Multimap;
-import com.squareup.javapoet.ClassName;
-import com.squareup.javapoet.FieldSpec;
-import com.squareup.javapoet.JavaFile;
-import com.squareup.javapoet.MethodSpec;
-import com.squareup.javapoet.ParameterSpec;
-import com.squareup.javapoet.TypeName;
-import com.squareup.javapoet.TypeSpec;
-import com.squareup.javapoet.TypeVariableName;
-
 /**
  * A compile-time Processor that aggregates classes annotated with {@link
  * org.tensorflow.op.annotation.Operator} and generates the {@code Ops} convenience API. Please
@@ -115,10 +113,12 @@ public final class OperatorProcessor extends AbstractProcessor {
     // generated our code, flag the location of each such class.
     if (hasRun) {
       for (Element e : annotated) {
-        error(e, "The Operator processor has already processed @Operator annotated sources\n" +
-            "and written out an Ops API. It cannot process additional @Operator sources.\n" +
-            "One reason this can happen is if other annotation processors generate\n" +
-            "new @Operator source files.");
+        error(
+            e,
+            "The Operator processor has already processed @Operator annotated sources\n"
+                + "and written out an Ops API. It cannot process additional @Operator sources.\n"
+                + "One reason this can happen is if other annotation processors generate\n"
+                + "new @Operator source files.");
       }
       return true;
     }
@@ -146,9 +146,11 @@ public final class OperatorProcessor extends AbstractProcessor {
     return Collections.singleton("org.tensorflow.op.annotation.Operator");
   }
 
-  private static final Pattern JAVADOC_TAG_PATTERN = Pattern.compile("@(?:param|return|throws|exception|see)\\s+.*");
+  private static final Pattern JAVADOC_TAG_PATTERN =
+      Pattern.compile("@(?:param|return|throws|exception|see)\\s+.*");
   private static final TypeName T_OPS = ClassName.get("org.tensorflow.op", "Ops");
-  private static final TypeName T_OPERATOR = ClassName.get("org.tensorflow.op.annotation", "Operator");
+  private static final TypeName T_OPERATOR =
+      ClassName.get("org.tensorflow.op.annotation", "Operator");
   private static final TypeName T_SCOPE = ClassName.get("org.tensorflow.op", "Scope");
   private static final TypeName T_GRAPH = ClassName.get("org.tensorflow", "Graph");
   private static final TypeName T_STRING = ClassName.get(String.class);
@@ -167,20 +169,17 @@ public final class OperatorProcessor extends AbstractProcessor {
 
   private void write(TypeSpec spec) {
     try {
-      JavaFile.builder("org.tensorflow.op", spec)
-          .skipJavaLangImports(true)
-          .build()
-          .writeTo(filer);
+      JavaFile.builder("org.tensorflow.op", spec).skipJavaLangImports(true).build().writeTo(filer);
     } catch (IOException e) {
       throw new AssertionError(e);
     }
   }
 
   private void writeApi(Multimap groupedMethods) {
-    Map groups = new HashMap();
-    
+    Map groups = new HashMap<>();
+
     // Generate a API class for each group collected other than the default one (= empty string)
-    for (Map.Entry> entry: groupedMethods.asMap().entrySet()) {
+    for (Map.Entry> entry : groupedMethods.asMap().entrySet()) {
       if (!entry.getKey().isEmpty()) {
         TypeSpec groupClass = buildGroupClass(entry.getKey(), entry.getValue());
         write(groupClass);
@@ -193,12 +192,17 @@ public final class OperatorProcessor extends AbstractProcessor {
   }
 
   private boolean collectOpsMethods(
-      RoundEnvironment roundEnv, Multimap groupedMethods, TypeElement annotation) {
+      RoundEnvironment roundEnv,
+      Multimap groupedMethods,
+      TypeElement annotation) {
     boolean result = true;
     for (Element e : roundEnv.getElementsAnnotatedWith(annotation)) {
       // @Operator can only apply to types, so e must be a TypeElement.
       if (!(e instanceof TypeElement)) {
-        error(e, "@Operator can only be applied to classes, but this is a %s", e.getKind().toString());
+        error(
+            e,
+            "@Operator can only be applied to classes, but this is a %s",
+            e.getKind().toString());
         result = false;
         continue;
       }
@@ -210,38 +214,42 @@ public final class OperatorProcessor extends AbstractProcessor {
     }
     return result;
   }
-  
-  private void collectOpMethods(Multimap groupedMethods, TypeElement opClass, TypeElement annotation) {
+
+  private void collectOpMethods(
+      Multimap groupedMethods, TypeElement opClass, TypeElement annotation) {
     AnnotationMirror am = getAnnotationMirror(opClass, annotation);
     String groupName = getAnnotationElementValueAsString("group", am);
     String methodName = getAnnotationElementValueAsString("name", am);
     ClassName opClassName = ClassName.get(opClass);
     if (Strings.isNullOrEmpty(methodName)) {
-      methodName = CaseFormat.UPPER_CAMEL.to(CaseFormat.LOWER_CAMEL, opClassName.simpleName()); 
+      methodName = CaseFormat.UPPER_CAMEL.to(CaseFormat.LOWER_CAMEL, opClassName.simpleName());
     }
-    // Build a method for each @Operator found in the class path. There should be one method per operation factory called
+    // Build a method for each @Operator found in the class path. There should be one method per
+    // operation factory called
     // "create", which takes in parameter a scope and, optionally, a list of arguments
     for (ExecutableElement opMethod : ElementFilter.methodsIn(opClass.getEnclosedElements())) {
-      if (opMethod.getModifiers().contains(Modifier.STATIC) && opMethod.getSimpleName().contentEquals("create")) {
+      if (opMethod.getModifiers().contains(Modifier.STATIC)
+          && opMethod.getSimpleName().contentEquals("create")) {
         MethodSpec method = buildOpMethod(methodName, opClassName, opMethod);
         groupedMethods.put(groupName, method);
       }
     }
   }
 
-  private MethodSpec buildOpMethod(String methodName, ClassName opClassName, ExecutableElement factoryMethod) {
+  private MethodSpec buildOpMethod(
+      String methodName, ClassName opClassName, ExecutableElement factoryMethod) {
     MethodSpec.Builder builder =
         MethodSpec.methodBuilder(methodName)
-        .addModifiers(Modifier.PUBLIC)
-        .returns(TypeName.get(factoryMethod.getReturnType()))
-        .varargs(factoryMethod.isVarArgs())
-        .addJavadoc("$L", buildOpMethodJavadoc(opClassName, factoryMethod));
+            .addModifiers(Modifier.PUBLIC)
+            .returns(TypeName.get(factoryMethod.getReturnType()))
+            .varargs(factoryMethod.isVarArgs())
+            .addJavadoc("$L", buildOpMethodJavadoc(opClassName, factoryMethod));
 
-    for (TypeParameterElement tp: factoryMethod.getTypeParameters()) {
+    for (TypeParameterElement tp : factoryMethod.getTypeParameters()) {
       TypeVariableName tvn = TypeVariableName.get((TypeVariable) tp.asType());
       builder.addTypeVariable(tvn);
     }
-    for (TypeMirror thrownType: factoryMethod.getThrownTypes()) {
+    for (TypeMirror thrownType : factoryMethod.getThrownTypes()) {
       builder.addException(TypeName.get(thrownType));
     }
     StringBuilder call = new StringBuilder("return $T.create(scope");
@@ -259,13 +267,17 @@ public final class OperatorProcessor extends AbstractProcessor {
     call.append(")");
     builder.addStatement(call.toString(), opClassName);
     return builder.build();
-  }    
-  
+  }
+
   private String buildOpMethodJavadoc(ClassName opClassName, ExecutableElement factoryMethod) {
     StringBuilder javadoc = new StringBuilder();
-    javadoc.append("Adds an {@link ").append(opClassName.simpleName()).append("} operation to the graph\n\n");
+    javadoc
+        .append("Adds an {@link ")
+        .append(opClassName.simpleName())
+        .append("} operation to the graph\n\n");
 
-    // Add all javadoc tags found in the operator factory method but the first one, which should be in all cases the
+    // Add all javadoc tags found in the operator factory method but the first one, which should be
+    // in all cases the
     // 'scope' parameter that is implicitly passed by this API
     Matcher tagMatcher = JAVADOC_TAG_PATTERN.matcher(elements.getDocComment(factoryMethod));
     boolean firstParam = true;
@@ -277,136 +289,144 @@ public final class OperatorProcessor extends AbstractProcessor {
       } else {
         javadoc.append(tag).append('\n');
       }
-    }    
+    }
     javadoc.append("@see {@link ").append(opClassName).append("}\n");
 
     return javadoc.toString();
   }
- 
+
   private static TypeSpec buildGroupClass(String group, Collection methods) {
     MethodSpec.Builder ctorBuilder =
         MethodSpec.constructorBuilder()
-        .addParameter(T_SCOPE, "scope")
-        .addStatement("this.scope = scope");
-    
+            .addParameter(T_SCOPE, "scope")
+            .addStatement("this.scope = scope");
+
     TypeSpec.Builder builder =
         TypeSpec.classBuilder(CaseFormat.LOWER_CAMEL.to(CaseFormat.UPPER_CAMEL, group) + "Ops")
-        .addModifiers(Modifier.PUBLIC, Modifier.FINAL)
-        .addJavadoc("An API for adding {@code $L} operations to a {@link $T Graph}\n\n" +
-            "@see {@link $T}\n", group, T_GRAPH, T_OPS)
-        .addMethods(methods)
-        .addMethod(ctorBuilder.build());
+            .addModifiers(Modifier.PUBLIC, Modifier.FINAL)
+            .addJavadoc(
+                "An API for adding {@code $L} operations to a {@link $T Graph}\n\n"
+                    + "@see {@link $T}\n",
+                group,
+                T_GRAPH,
+                T_OPS)
+            .addMethods(methods)
+            .addMethod(ctorBuilder.build());
 
     builder.addField(
-        FieldSpec.builder(T_SCOPE, "scope")
-        .addModifiers(Modifier.PRIVATE, Modifier.FINAL)
-        .build());
+        FieldSpec.builder(T_SCOPE, "scope").addModifiers(Modifier.PRIVATE, Modifier.FINAL).build());
 
     return builder.build();
   }
 
-  private static TypeSpec buildTopClass(Map groupToClass, Collection methods) {
+  private static TypeSpec buildTopClass(
+      Map groupToClass, Collection methods) {
     MethodSpec.Builder ctorBuilder =
         MethodSpec.constructorBuilder()
-        .addModifiers(Modifier.PRIVATE)
-        .addParameter(T_SCOPE, "scope")
-        .addStatement("this.scope = scope", T_SCOPE);
+            .addModifiers(Modifier.PRIVATE)
+            .addParameter(T_SCOPE, "scope")
+            .addStatement("this.scope = scope", T_SCOPE);
 
-    for (Map.Entry entry: groupToClass.entrySet()) {
+    for (Map.Entry entry : groupToClass.entrySet()) {
       ctorBuilder.addStatement("$L = new $T(scope)", entry.getKey(), entry.getValue());
     }
 
     TypeSpec.Builder opsBuilder =
         TypeSpec.classBuilder("Ops")
-        .addModifiers(Modifier.PUBLIC, Modifier.FINAL)
-        .addJavadoc("An API for building a {@link $T} with operation wrappers\n
\n" +
-            "Any operation wrapper found in the classpath properly annotated as an {@link $T @Operator} is exposed\n" +
-            "by this API or one of its subgroup.\n
Example usage:\n
{@code\n" +
-            "try (Graph g = new Graph()) {\n" +
-            "  Ops ops = new Ops(g);\n" +
-            "  // Operations are typed classes with convenience\n" +
-            "  // builders in Ops.\n" +
-            "  Constant three = ops.constant(3);\n" +
-            "  // Single-result operations implement the Operand\n" +
-            "  // interface, so this works too.\n" +
-            "  Operand four = ops.constant(4);\n" +
-            "  // Most builders are found within a group, and accept\n" +
-            "  // Operand types as operands\n" +
-            "  Operand nine = ops.math().add(four, ops.constant(5));\n" +
-            "  // Multi-result operations however offer methods to\n" +
-            "  // select a particular result for use.\n" +
-            "  Operand result = \n" +
-            "      ops.math().add(ops.array().unique(s, a).y(), b);\n" +
-            "  // Optional attributes\n" +
-            "  ops.math().matMul(a, b, MatMul.transposeA(true));\n" +
-            "  // Naming operators\n" +
-            "  ops.withName(ā€œfooā€).constant(5); // name ā€œfooā€\n" +
-            "  // Names can exist in a hierarchy\n" +
-            "  Ops sub = ops.withSubScope(ā€œsubā€);\n" +
-            "  sub.withName(ā€œbarā€).constant(4); // ā€œsub/barā€\n" +
-            "}\n" +
-            "}
\n", T_GRAPH, T_OPERATOR)
-        .addMethods(methods)
-        .addMethod(ctorBuilder.build());
+            .addModifiers(Modifier.PUBLIC, Modifier.FINAL)
+            .addJavadoc(
+                "An API for building a {@link $T} with operation wrappers\n
\n"
+                    + "Any operation wrapper found in the classpath properly annotated as an"
+                    + "{@link $T @Operator} is exposed\n"
+                    + "by this API or one of its subgroup.\n
Example usage:\n
{@code\n"
+                    + "try (Graph g = new Graph()) {\n"
+                    + "  Ops ops = new Ops(g);\n"
+                    + "  // Operations are typed classes with convenience\n"
+                    + "  // builders in Ops.\n"
+                    + "  Constant three = ops.constant(3);\n"
+                    + "  // Single-result operations implement the Operand\n"
+                    + "  // interface, so this works too.\n"
+                    + "  Operand four = ops.constant(4);\n"
+                    + "  // Most builders are found within a group, and accept\n"
+                    + "  // Operand types as operands\n"
+                    + "  Operand nine = ops.math().add(four, ops.constant(5));\n"
+                    + "  // Multi-result operations however offer methods to\n"
+                    + "  // select a particular result for use.\n"
+                    + "  Operand result = \n"
+                    + "      ops.math().add(ops.array().unique(s, a).y(), b);\n"
+                    + "  // Optional attributes\n"
+                    + "  ops.math().matMul(a, b, MatMul.transposeA(true));\n"
+                    + "  // Naming operators\n"
+                    + "  ops.withName(ā€œfooā€).constant(5); // name ā€œfooā€\n"
+                    + "  // Names can exist in a hierarchy\n"
+                    + "  Ops sub = ops.withSubScope(ā€œsubā€);\n"
+                    + "  sub.withName(ā€œbarā€).constant(4); // ā€œsub/barā€\n"
+                    + "}\n"
+                    + "}
\n",
+                T_GRAPH,
+                T_OPERATOR)
+            .addMethods(methods)
+            .addMethod(ctorBuilder.build());
 
     opsBuilder.addMethod(
         MethodSpec.methodBuilder("withSubScope")
-        .addModifiers(Modifier.PUBLIC)
-        .addParameter(T_STRING, "childScopeName")
-        .returns(T_OPS)
-        .addStatement("return new $T(scope.withSubScope(childScopeName))", T_OPS)
-        .addJavadoc(
-            "Returns an API that adds operations to the graph with the provided name prefix.\n\n" +
-            "@see {@link $T#withSubScope(String)}\n", T_SCOPE)
-        .build());
+            .addModifiers(Modifier.PUBLIC)
+            .addParameter(T_STRING, "childScopeName")
+            .returns(T_OPS)
+            .addStatement("return new $T(scope.withSubScope(childScopeName))", T_OPS)
+            .addJavadoc(
+                "Returns an API that adds operations to the graph with the provided name prefix.\n"
+                    + "\n@see {@link $T#withSubScope(String)}\n",
+                T_SCOPE)
+            .build());
 
     opsBuilder.addMethod(
         MethodSpec.methodBuilder("withName")
-        .addModifiers(Modifier.PUBLIC)
-        .addParameter(T_STRING, "opName")
-        .returns(T_OPS)
-        .addStatement("return new Ops(scope.withName(opName))")
-        .addJavadoc(
-            "Returns an API that uses the provided name for an op.\n\n" +
-            "@see {@link $T#withName(String)}\n", T_SCOPE)
-        .build());
+            .addModifiers(Modifier.PUBLIC)
+            .addParameter(T_STRING, "opName")
+            .returns(T_OPS)
+            .addStatement("return new Ops(scope.withName(opName))")
+            .addJavadoc(
+                "Returns an API that uses the provided name for an op.\n\n"
+                    + "@see {@link $T#withName(String)}\n",
+                T_SCOPE)
+            .build());
 
     opsBuilder.addField(
-        FieldSpec.builder(T_SCOPE, "scope")
-        .addModifiers(Modifier.PRIVATE, Modifier.FINAL)
-        .build());
+        FieldSpec.builder(T_SCOPE, "scope").addModifiers(Modifier.PRIVATE, Modifier.FINAL).build());
 
     opsBuilder.addMethod(
         MethodSpec.methodBuilder("scope")
-        .addModifiers(Modifier.PUBLIC, Modifier.FINAL)
-        .returns(T_SCOPE)
-        .addStatement("return scope")
-        .addJavadoc("Returns the current {@link $T scope} of this API\n", T_SCOPE)
-        .build());
+            .addModifiers(Modifier.PUBLIC, Modifier.FINAL)
+            .returns(T_SCOPE)
+            .addStatement("return scope")
+            .addJavadoc("Returns the current {@link $T scope} of this API\n", T_SCOPE)
+            .build());
 
-    for (Map.Entry entry: groupToClass.entrySet()) {
+    for (Map.Entry entry : groupToClass.entrySet()) {
       opsBuilder.addField(
           FieldSpec.builder(entry.getValue(), entry.getKey())
-          .addModifiers(Modifier.PUBLIC, Modifier.FINAL)
-          .build());
-      
+              .addModifiers(Modifier.PUBLIC, Modifier.FINAL)
+              .build());
+
       opsBuilder.addMethod(
           MethodSpec.methodBuilder(entry.getKey())
-          .addModifiers(Modifier.PUBLIC, Modifier.FINAL)
-          .returns(entry.getValue())
-          .addStatement("return $L", entry.getKey())
-          .addJavadoc("Returns an API for adding {@code $L} operations to the graph\n", entry.getKey())
-          .build());
+              .addModifiers(Modifier.PUBLIC, Modifier.FINAL)
+              .returns(entry.getValue())
+              .addStatement("return $L", entry.getKey())
+              .addJavadoc(
+                  "Returns an API for adding {@code $L} operations to the graph\n", entry.getKey())
+              .build());
     }
 
     opsBuilder.addMethod(
         MethodSpec.methodBuilder("create")
-        .addModifiers(Modifier.PUBLIC, Modifier.STATIC)
-        .addParameter(T_GRAPH, "graph")
-        .returns(T_OPS)
-        .addStatement("return new Ops(new $T(graph))", T_SCOPE)
-        .addJavadoc("Creates an API for adding operations to the provided {@code graph}\n")
-        .build());
+            .addModifiers(Modifier.PUBLIC, Modifier.STATIC)
+            .addParameter(T_GRAPH, "graph")
+            .returns(T_OPS)
+            .addStatement("return new Ops(new $T(graph))", T_SCOPE)
+            .addJavadoc("Creates an API for adding operations to the provided {@code graph}\n")
+            .build());
 
     return opsBuilder.build();
   }
@@ -417,12 +437,16 @@ public final class OperatorProcessor extends AbstractProcessor {
         return am;
       }
     }
-    throw new IllegalArgumentException("Annotation " + annotation.getSimpleName() + " not present on element "
-        + element.getSimpleName());
+    throw new IllegalArgumentException(
+        "Annotation "
+            + annotation.getSimpleName()
+            + " not present on element "
+            + element.getSimpleName());
   }
-  
+
   private static String getAnnotationElementValueAsString(String elementName, AnnotationMirror am) {
-    for (Map.Entry entry :  am.getElementValues().entrySet()) {
+    for (Map.Entry entry :
+        am.getElementValues().entrySet()) {
       if (entry.getKey().getSimpleName().contentEquals(elementName)) {
         return entry.getValue().getValue().toString();
       }
diff --git a/tensorflow/java/src/main/java/org/tensorflow/Graph.java b/tensorflow/java/src/main/java/org/tensorflow/Graph.java
index d4fd3db5f7325ae891832ff7b658f5d3ea0789a6..7d19696749bbbb944e591daf596562f13f6dc103 100644
--- a/tensorflow/java/src/main/java/org/tensorflow/Graph.java
+++ b/tensorflow/java/src/main/java/org/tensorflow/Graph.java
@@ -143,6 +143,82 @@ public final class Graph implements AutoCloseable {
     }
   }
 
+  /**
+   * Adds operations to compute the partial derivatives of sum of {@code y}s w.r.t {@code x}s,
+   * i.e., {@code d(y_1 + y_2 + ...)/dx_1, d(y_1 + y_2 + ...)/dx_2...}
+   * 
 
+   * {@code dx} are used as initial gradients (which represent the symbolic partial derivatives of some loss function 
+   * {@code L} w.r.t. {@code y}). {@code dx} must be null or have size of {@code y}.
+   * 

+   * If {@code dx} is null, the implementation will use dx of {@link org.tensorflow.op.core.OnesLike OnesLike} for all
+   * shapes in {@code y}.
+   * 
+   * @param y output of the function to derive
+   * @param x inputs of the function for which partial derivatives are computed
+   * @param dx if not null, the partial derivatives of some loss function {@code L} w.r.t. {@code y}
+   * @return the partial derivatives {@code dy} with the size of {@code x}
+   */
+  public Output[] addGradients(Output[] y, Output[] x, Output[] dx) {
+    Output[] dy = new Output[x.length];
+    final long[] yHandles = new long[y.length];
+    final int[] yIndices = new int[y.length];
+    final long[] xHandles = new long[x.length];
+    final int[] xIndices = new int[x.length];
+    long[] dxHandles = null;
+    int[] dxIndices = null;
+
+    try (Reference ref = ref()) {
+      for (int i = 0; i < y.length; ++i) {
+        yHandles[i] = y[i].op().getUnsafeNativeHandle();
+        yIndices[i] = y[i].index();
+      }
+      for (int i = 0; i < x.length; ++i) {
+        xHandles[i] = x[i].op().getUnsafeNativeHandle();
+        xIndices[i] = x[i].index();
+      }
+      if (dx != null && dx.length > 0) {
+        dxHandles = new long[dx.length];
+        dxIndices = new int[dx.length];
+
+        for (int i = 0; i < dx.length; ++i) {
+          dxHandles[i] = dx[i].op().getUnsafeNativeHandle();
+          dxIndices[i] = dx[i].index();
+        }
+      }
+      // Gradient outputs are returned in two continuous arrays concatenated into one. The first holds the native handles 
+      // of the gradient operations while the second holds the index of their output
+      // e.g. given xHandles = [x0Handle, x1Handle, ...] and xIndices = [x0Index, x1Index, ..], we obtain 
+      // dy = [dy0Handle, dy1Handle, ..., dy0Index, dy1Index, ...]
+      long[] dyHandlesAndIndices =
+            addGradients(ref.nativeHandle(), yHandles, yIndices, xHandles, xIndices, dxHandles, dxIndices);
+      int ndy = dyHandlesAndIndices.length >> 1;
+      if (ndy != dy.length) {
+        throw new IllegalStateException(String.valueOf(ndy) + " gradients were added to the graph when " + dy.length
+            + " were expected");
+      }
+      for (int i = 0, j = ndy; i < ndy; ++i, ++j) {
+        Operation op = new Operation(this, dyHandlesAndIndices[i]);
+        dy[i] = new Output<>(op, (int) dyHandlesAndIndices[j]);
+      }
+    }
+    return dy;
+  }
+
+  /**
+   * Adds operations to compute the partial derivatives of sum of {@code y}s w.r.t {@code x}s,
+   * i.e., {@code dy/dx_1, dy/dx_2...}
+   * 
 
+   * This is a simplified version of {@link #addGradients(Output[], Output[], Output[]) where {@code y} is
+   * a single output and {@code dx} is null.
+   * 
+   * @param y output of the function to derive
+   * @param x inputs of the function for which partial derivatives are computed
+   * @return the partial derivatives {@code dy} with the size of {@code x}
+   */
+  public Output[] addGradients(Output y, Output[] x) {
+    return addGradients(new Output[]{y}, x, null);
+  }
+  
   private final Object nativeHandleLock = new Object();
   private long nativeHandle;
   private int refcount = 0;
@@ -254,6 +330,9 @@ public final class Graph implements AutoCloseable {
 
   private static native byte[] toGraphDef(long handle);
 
+  private static native long[] addGradients(long handle, long[] inputHandles, int[] inputIndices,
+      long[] outputHandles, int[] outputIndices, long[] gradInputHandles, int[] gradInputIndices);
+
   static {
     TensorFlow.init();
   }
diff --git a/tensorflow/java/src/main/java/org/tensorflow/Input.java b/tensorflow/java/src/main/java/org/tensorflow/Input.java
new file mode 100644
index 0000000000000000000000000000000000000000..13bc463e7d6a991858332a353681b24fff417547
--- /dev/null
+++ b/tensorflow/java/src/main/java/org/tensorflow/Input.java
@@ -0,0 +1,48 @@
+/* Copyright 2017 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+package org.tensorflow;
+
+/**
+ * Interface implemented by operands of a TensorFlow operation.
+ *
+ * 
Example usage:
+ *
+ * 
{@code
+ * // The "decodeJpeg" operation can be used as input to the "cast" operation
+ * Input decodeJpeg = ops.image().decodeJpeg(...);
+ * ops.math().cast(decodeJpeg, DataType.FLOAT);
+ *
+ * // The output "y" of the "unique" operation can be used as input to the "cast" operation
+ * Output y = ops.array().unique(...).y();
+ * ops.math().cast(y, DataType.FLOAT);
+ *
+ * // The "split" operation can be used as input list to the "concat" operation
+ * Iterable split = ops.array().split(...);
+ * ops.array().concat(0, split);
+ * }

+ */
+public interface Input {
+
+  /**
+   * Returns the symbolic handle of a tensor.
+   *
+   * 
Inputs to TensorFlow operations are outputs of another TensorFlow operation. This method is
+   * used to obtain a symbolic handle that represents the computation of the input.
+   *
+   * @see OperationBuilder#addInput(Output)
+   */
+  Output asOutput();
+}
diff --git a/tensorflow/java/src/main/java/org/tensorflow/op/core/Gradients.java b/tensorflow/java/src/main/java/org/tensorflow/op/core/Gradients.java
new file mode 100644
index 0000000000000000000000000000000000000000..f4671c8af941dd732859080238fa48e0a22672b6
--- /dev/null
+++ b/tensorflow/java/src/main/java/org/tensorflow/op/core/Gradients.java
@@ -0,0 +1,153 @@
+/* Copyright 2018 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+package org.tensorflow.op.core;
+
+import java.util.Arrays;
+import java.util.Iterator;
+import java.util.List;
+
+import org.tensorflow.Operand;
+import org.tensorflow.Output;
+import org.tensorflow.op.Op;
+import org.tensorflow.op.Operands;
+import org.tensorflow.op.Scope;
+import org.tensorflow.op.annotation.Operator;
+
+/**
+ * Adds operations to compute the partial derivatives of sum of {@code y}s w.r.t {@code x}s,
+ * i.e., {@code d(y_1 + y_2 + ...)/dx_1, d(y_1 + y_2 + ...)/dx_2...}
+ * 
 
+ * If {@code Options.dx()} values are set, they are as the initial symbolic partial derivatives of some loss 
+ * function {@code L} w.r.t. {@code y}. {@code Options.dx()} must have the size of {@code y}.
+ * 

+ * If {@code Options.dx()} is not set, the implementation will use dx of {@code OnesLike} for all
+ * shapes in {@code y}.
+ * 

+ * The partial derivatives are returned in output {@code dy}, with the size of {@code x}.
+ * 

+ * Example of usage:
+ * 
{@code
+ * Gradients gradients = Gradients.create(scope, Arrays.asList(loss), Arrays.asList(w, b));
+ * 
+ * Constant alpha = ops.constant(1.0f, Float.class);
+ * ApplyGradientDescent.create(scope, w, alpha, gradients.dy(0));
+ * ApplyGradientDescent.create(scope, b, alpha, gradients.dy(1));
+ * }

+ */
+@Operator
+public class Gradients implements Op, Iterable> {
+
+  /**
+   * Optional attributes for {@link Gradients}
+   */
+  public static class Options {
+    
+    /**
+     * @param dx partial derivatives of some loss function {@code L} w.r.t. {@code y}
+     * @return this option builder
+     */
+    public Options dx(Iterable> dx) {
+      this.dx = dx;
+      return this;
+    }
+    
+    private Iterable> dx;
+    
+    private Options() {
+    }
+  }
+
+  /**
+   * Adds gradients computation ops to the graph according to scope.
+   * 
+   * @param scope current graph scope
+   * @param y outputs of the function to derive
+   * @param x inputs of the function for which partial derivatives are computed
+   * @param options carries optional attributes values
+   * @return a new instance of {@code Gradients}
+   */
+  public static Gradients create(Scope scope, Iterable> y, Iterable> x, Options... options) {
+    Output[] dx = null;
+    if (options != null) {
+      for (Options opts : options) {
+        if (opts.dx != null) {
+          dx = Operands.asOutputs(opts.dx);
+        }
+      }
+    }
+    Output[] gradOutputs = scope.graph().addGradients(Operands.asOutputs(y), Operands.asOutputs(x), dx);
+    return new Gradients(Arrays.asList(gradOutputs));
+  }
+
+  /**
+   * Adds gradients computation ops to the graph according to scope.
+   * 
+   * This is a simplified version of {@link #create(Scope, Iterable, Iterable, Options...)} where {@code y} is
+   * a single output.
+   * 
+   * @param scope current graph scope
+   * @param y output of the function to derive
+   * @param x inputs of the function for which partial derivatives are computed
+   * @param options carries optional attributes values
+   * @return a new instance of {@code Gradients}
+   */
+  @SuppressWarnings({"unchecked", "rawtypes"})
+  public static Gradients create(Scope scope, Operand y, Iterable> x, Options... options) {
+    return create(scope, (Iterable) Arrays.asList(y), x, options);
+  }
+
+  /**
+   * @param dx partial derivatives of some loss function {@code L} w.r.t. {@code y}
+   * @return builder to add more options to this operation
+   */
+  public Options dx(Iterable> dx) {
+    return new Options().dx(dx);
+  }
+
+  @Override
+  @SuppressWarnings({"rawtypes", "unchecked"})
+  public Iterator> iterator() {
+    return (Iterator) dy.iterator();
+  }
+  
+  /**
+   * Partial derivatives of {@code y}s w.r.t. {@code x}s, with the size of {@code x}
+   */
+  public List> dy() {
+    return dy;
+  }
+  
+  /**
+   * Returns a symbolic handle to one of the gradient operation output
+   * 

+   * Warning: Does not check that the type of the tensor matches T. It is recommended to call
+   * this method with an explicit type parameter rather than letting it be inferred, e.g. {@code
+   * gradients.dy(0)}
+   *
+   * @param  The expected element type of the tensors produced by this output.
+   * @param index The index of the output among the gradients added by this operation
+   */
+  @SuppressWarnings("unchecked")
+  public  Output dy(int index) {
+    return (Output) dy.get(index);
+  }
+
+  private List> dy;
+  
+  private Gradients(List> dy) {
+    this.dy = dy;
+  }
+}
diff --git a/tensorflow/java/src/main/java/org/tensorflow/types/TFBool.java b/tensorflow/java/src/main/java/org/tensorflow/types/TFBool.java
new file mode 100644
index 0000000000000000000000000000000000000000..ab34f6aa125eded4f7acafea1439559d084c9780
--- /dev/null
+++ b/tensorflow/java/src/main/java/org/tensorflow/types/TFBool.java
@@ -0,0 +1,30 @@
+/* Copyright 2017 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+// GENERATED FILE. To update, edit tftypes.pl instead.
+
+package org.tensorflow.types;
+
+import org.tensorflow.DataType;
+
+/** Represents a boolean. */
+public class TFBool implements TFType {
+  private TFBool() {}
+  static {
+    Types.typeCodes.put(TFBool.class, DataType.BOOL);
+  }
+  static {
+    Types.scalars.put(TFBool.class, false);
+  }
+}
diff --git a/tensorflow/java/src/main/java/org/tensorflow/types/TFDouble.java b/tensorflow/java/src/main/java/org/tensorflow/types/TFDouble.java
new file mode 100644
index 0000000000000000000000000000000000000000..49e5d9f2f3a6627201dd9af67b5698f095a9c0f0
--- /dev/null
+++ b/tensorflow/java/src/main/java/org/tensorflow/types/TFDouble.java
@@ -0,0 +1,30 @@
+/* Copyright 2017 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+// GENERATED FILE. To update, edit tftypes.pl instead.
+
+package org.tensorflow.types;
+
+import org.tensorflow.DataType;
+
+/** Represents a 64-bit double precision floating point number. */
+public class TFDouble implements TFType {
+  private TFDouble() {}
+  static {
+    Types.typeCodes.put(TFDouble.class, DataType.DOUBLE);
+  }
+  static {
+    Types.scalars.put(TFDouble.class, 0.0);
+  }
+}
diff --git a/tensorflow/java/src/main/java/org/tensorflow/types/TFFloat.java b/tensorflow/java/src/main/java/org/tensorflow/types/TFFloat.java
new file mode 100644
index 0000000000000000000000000000000000000000..8426ee41f01efa71cac4dc2dd8aabcafe500e1cc
--- /dev/null
+++ b/tensorflow/java/src/main/java/org/tensorflow/types/TFFloat.java
@@ -0,0 +1,30 @@
+/* Copyright 2017 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+// GENERATED FILE. To update, edit tftypes.pl instead.
+
+package org.tensorflow.types;
+
+import org.tensorflow.DataType;
+
+/** Represents a 32-bit single precision floating point number. */
+public class TFFloat implements TFType {
+  private TFFloat() {}
+  static {
+    Types.typeCodes.put(TFFloat.class, DataType.FLOAT);
+  }
+  static {
+    Types.scalars.put(TFFloat.class, 0f);
+  }
+}
diff --git a/tensorflow/java/src/main/java/org/tensorflow/types/TFInt32.java b/tensorflow/java/src/main/java/org/tensorflow/types/TFInt32.java
new file mode 100644
index 0000000000000000000000000000000000000000..3947b6ad095b5a4e8bc8b55561961fc91bc73966
--- /dev/null
+++ b/tensorflow/java/src/main/java/org/tensorflow/types/TFInt32.java
@@ -0,0 +1,30 @@
+/* Copyright 2017 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+// GENERATED FILE. To update, edit tftypes.pl instead.
+
+package org.tensorflow.types;
+
+import org.tensorflow.DataType;
+
+/** Represents a 32-bit signed integer. */
+public class TFInt32 implements TFType {
+  private TFInt32() {}
+  static {
+    Types.typeCodes.put(TFInt32.class, DataType.INT32);
+  }
+  static {
+    Types.scalars.put(TFInt32.class, 0);
+  }
+}
diff --git a/tensorflow/java/src/main/java/org/tensorflow/types/TFInt64.java b/tensorflow/java/src/main/java/org/tensorflow/types/TFInt64.java
new file mode 100644
index 0000000000000000000000000000000000000000..ccdded86939c0bbe1265145d76b8470a9099fb94
--- /dev/null
+++ b/tensorflow/java/src/main/java/org/tensorflow/types/TFInt64.java
@@ -0,0 +1,30 @@
+/* Copyright 2017 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+// GENERATED FILE. To update, edit tftypes.pl instead.
+
+package org.tensorflow.types;
+
+import org.tensorflow.DataType;
+
+/** Represents a 64-bit signed integer. */
+public class TFInt64 implements TFType {
+  private TFInt64() {}
+  static {
+    Types.typeCodes.put(TFInt64.class, DataType.INT64);
+  }
+  static {
+    Types.scalars.put(TFInt64.class, 0L);
+  }
+}
diff --git a/tensorflow/java/src/main/java/org/tensorflow/types/TFString.java b/tensorflow/java/src/main/java/org/tensorflow/types/TFString.java
new file mode 100644
index 0000000000000000000000000000000000000000..e7327e8c57fd41e1c4441f7f79eb44614afdc325
--- /dev/null
+++ b/tensorflow/java/src/main/java/org/tensorflow/types/TFString.java
@@ -0,0 +1,27 @@
+/* Copyright 2017 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+// GENERATED FILE. To update, edit tftypes.pl instead.
+
+package org.tensorflow.types;
+
+import org.tensorflow.DataType;
+
+/** Represents an arbitrary sequence of bytes. */
+public class TFString implements TFType {
+  private TFString() {}
+  static {
+    Types.typeCodes.put(TFString.class, DataType.STRING);
+  }
+}
diff --git a/tensorflow/java/src/main/java/org/tensorflow/types/TFType.java b/tensorflow/java/src/main/java/org/tensorflow/types/TFType.java
new file mode 100644
index 0000000000000000000000000000000000000000..562953ac9dc0abf8cac172338025bac9e1dae81c
--- /dev/null
+++ b/tensorflow/java/src/main/java/org/tensorflow/types/TFType.java
@@ -0,0 +1,20 @@
+/* Copyright 2017 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+package org.tensorflow.types;
+
+/**
+ * A marker interface for classes representing TensorFlow types.
+ */
+public interface TFType {}
diff --git a/tensorflow/java/src/main/java/org/tensorflow/types/TFUInt8.java b/tensorflow/java/src/main/java/org/tensorflow/types/TFUInt8.java
new file mode 100644
index 0000000000000000000000000000000000000000..d7305ca5a80311f520e446353d31d376c210d6a3
--- /dev/null
+++ b/tensorflow/java/src/main/java/org/tensorflow/types/TFUInt8.java
@@ -0,0 +1,30 @@
+/* Copyright 2017 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+// GENERATED FILE. To update, edit tftypes.pl instead.
+
+package org.tensorflow.types;
+
+import org.tensorflow.DataType;
+
+/** Represents an 8-bit unsigned integer. */
+public class TFUInt8 implements TFType {
+  private TFUInt8() {}
+  static {
+    Types.typeCodes.put(TFUInt8.class, DataType.UINT8);
+  }
+  static {
+    Types.scalars.put(TFUInt8.class, (byte)0);
+  }
+}
diff --git a/tensorflow/java/src/main/java/org/tensorflow/types/Types.java b/tensorflow/java/src/main/java/org/tensorflow/types/Types.java
new file mode 100644
index 0000000000000000000000000000000000000000..976cd9fd347fb48020f186ece8d23dae89626b1d
--- /dev/null
+++ b/tensorflow/java/src/main/java/org/tensorflow/types/Types.java
@@ -0,0 +1,52 @@
+/* Copyright 2017 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+package org.tensorflow.types;
+
+import java.util.HashMap;
+import java.util.Map;
+import org.tensorflow.DataType;
+
+/**
+ * Utility class for managing the representation of TensorFlow types as Java
+ * types. For each TensorFlow type (e.g., int32), there is a corresponding Java
+ * type (e.g., TFInt32) that represents it at compile time and a corresponding
+ * class object (e.g., TFInt32.class) that represents it at run time. There is
+ * also an enumeration value in DataType that can be used to represent the
+ * type, though that should rarely be required.
+ */
+public class Types {
+
+  private Types() {} // not instantiable
+
+  static final Map, DataType> typeCodes = new HashMap<>();
+
+  /** Returns the DataType value corresponding to a TensorFlow type class. */
+  public static DataType dataType(Class c) {
+    DataType dtype = typeCodes.get(c);
+    if (dtype == null) {
+      throw new IllegalArgumentException("" + c + " is not a TensorFlow type.");
+    }
+    return dtype;
+  }
+
+  static final Map, Object> scalars = new HashMap<>();
+
+  /** Returns the zero value of type described by {@code c}, or null if
+   *  the type (e.g., string) is not numeric and therefore has no zero value.
+   */
+  public static Object zeroValue(Class c) {
+    return scalars.get(c);
+  }
+}
diff --git a/tensorflow/java/src/main/native/graph_jni.cc b/tensorflow/java/src/main/native/graph_jni.cc
index 0fef15527586555e7d3fc2c76403c6e5888fb236..dac6a345e917b618f7f1234c27959069650b51b7 100644
--- a/tensorflow/java/src/main/native/graph_jni.cc
+++ b/tensorflow/java/src/main/native/graph_jni.cc
@@ -16,7 +16,9 @@ limitations under the License.
 #include "tensorflow/java/src/main/native/graph_jni.h"
 
 #include 
+#include 
 #include "tensorflow/c/c_api.h"
+#include "tensorflow/java/src/main/native/utils_jni.h"
 #include "tensorflow/java/src/main/native/exception_jni.h"
 
 namespace {
@@ -130,3 +132,55 @@ Java_org_tensorflow_Graph_toGraphDef(JNIEnv* env, jclass clazz, jlong handle) {
   TF_DeleteBuffer(buf);
   return ret;
 }
+
+JNIEXPORT jlongArray JNICALL
+Java_org_tensorflow_Graph_addGradients(JNIEnv* env, jclass clazz, jlong handle,
+    jlongArray y_handles, jintArray y_indices,
+    jlongArray x_handles, jintArray x_indices,
+    jlongArray dx_handles, jintArray dx_indices) {
+
+  TF_Graph* g = requireHandle(env, handle);
+  if (g == nullptr) return nullptr;
+
+  const jint ny = env->GetArrayLength(y_handles);
+  const jint nx = env->GetArrayLength(x_handles);
+
+  std::unique_ptr y(new TF_Output[ny]);
+  std::unique_ptr x(new TF_Output[nx]);
+  std::unique_ptr dx(nullptr);
+  std::unique_ptr dy(new TF_Output[nx]);
+
+  resolveOutputs(env, "y", y_handles, y_indices, y.get(), ny);
+  resolveOutputs(env, "x", x_handles, x_indices, x.get(), nx);
+  if (dx_handles != nullptr) {
+    if (env->GetArrayLength(dx_handles) != ny) {
+      throwException(env, kIllegalArgumentException,
+                     "expected %d, got %d dx handles", ny,
+                     env->GetArrayLength(dx_handles));
+    }
+    dx.reset(new TF_Output[ny]);
+    resolveOutputs(env, "dx", dx_handles, dx_indices, dx.get(), ny);
+  }
+  if (env->ExceptionCheck()) return nullptr;
+
+  TF_Status* status = TF_NewStatus();
+  TF_AddGradients(g, y.get(), ny, x.get(), nx, dx.get(), status, dy.get());
+
+  if (!throwExceptionIfNotOK(env, status)) {
+    TF_DeleteStatus(status);
+    return nullptr;
+  }
+  TF_DeleteStatus(status);
+
+  // returned array contains both op handles and output indices, in pair
+  jlongArray dy_handles_and_indices = env->NewLongArray(nx << 1);
+  jlong* dy_elems = env->GetLongArrayElements(dy_handles_and_indices, nullptr);
+  for (int i = 0, j = nx; i < nx; ++i, ++j) {
+    TF_Output dy_output = dy.get()[i];
+    dy_elems[i] = reinterpret_cast(dy_output.oper);
+    dy_elems[j] = static_cast(dy_output.index);
+  }
+  env->ReleaseLongArrayElements(dy_handles_and_indices, dy_elems, 0);
+
+  return dy_handles_and_indices;
+}
diff --git a/tensorflow/java/src/main/native/graph_jni.h b/tensorflow/java/src/main/native/graph_jni.h
index dd2e038332f7d39e6460d6cfef40a9df7e348758..4f87e8d5a79d3ac46f7813ba4344bbfda069b557 100644
--- a/tensorflow/java/src/main/native/graph_jni.h
+++ b/tensorflow/java/src/main/native/graph_jni.h
@@ -73,6 +73,15 @@ JNIEXPORT jbyteArray JNICALL Java_org_tensorflow_Graph_toGraphDef(JNIEnv *,
                                                                   jclass,
                                                                   jlong);
 
+/*
+ * Class:     org_tensorflow_Graph
+ * Method:    name
+ * Signature: (J[J[I[J[I[J[I)[J
+ */
+JNIEXPORT jlongArray JNICALL Java_org_tensorflow_Graph_addGradients(JNIEnv *,
+    jclass, jlong, jlongArray, jintArray, jlongArray, jintArray, jlongArray,
+    jintArray);
+
 #ifdef __cplusplus
 }  // extern "C"
 #endif  // __cplusplus
diff --git a/tensorflow/java/src/main/native/session_jni.cc b/tensorflow/java/src/main/native/session_jni.cc
index 2cd542d3c9be536a42037e9ef533ed629dd3ac9f..8b1152578555c0d9b5b4b383460116050c89c3d5 100644
--- a/tensorflow/java/src/main/native/session_jni.cc
+++ b/tensorflow/java/src/main/native/session_jni.cc
@@ -17,6 +17,7 @@ limitations under the License.
 #include 
 
 #include "tensorflow/c/c_api.h"
+#include "tensorflow/java/src/main/native/utils_jni.h"
 #include "tensorflow/java/src/main/native/exception_jni.h"
 #include "tensorflow/java/src/main/native/session_jni.h"
 
@@ -55,37 +56,6 @@ void resolveHandles(JNIEnv* env, const char* type, jlongArray src_array,
   env->ReleaseLongArrayElements(src_array, src_start, JNI_ABORT);
 }
 
-void resolveOutputs(JNIEnv* env, const char* type, jlongArray src_op,
-                    jintArray src_index, TF_Output* dst, jint n) {
-  if (env->ExceptionCheck()) return;
-  jint len = env->GetArrayLength(src_op);
-  if (len != n) {
-    throwException(env, kIllegalArgumentException,
-                   "expected %d, got %d %s Operations", n, len, type);
-    return;
-  }
-  len = env->GetArrayLength(src_index);
-  if (len != n) {
-    throwException(env, kIllegalArgumentException,
-                   "expected %d, got %d %s Operation output indices", n, len,
-                   type);
-    return;
-  }
-  jlong* op_handles = env->GetLongArrayElements(src_op, nullptr);
-  jint* indices = env->GetIntArrayElements(src_index, nullptr);
-  for (int i = 0; i < n; ++i) {
-    if (op_handles[i] == 0) {
-      throwException(env, kNullPointerException, "invalid %s (#%d of %d)", type,
-                     i, n);
-      break;
-    }
-    dst[i] = TF_Output{reinterpret_cast(op_handles[i]),
-                       static_cast(indices[i])};
-  }
-  env->ReleaseIntArrayElements(src_index, indices, JNI_ABORT);
-  env->ReleaseLongArrayElements(src_op, op_handles, JNI_ABORT);
-}
-
 void TF_MaybeDeleteBuffer(TF_Buffer* buf) {
   if (buf == nullptr) return;
   TF_DeleteBuffer(buf);
@@ -116,20 +86,22 @@ JNIEXPORT jlong JNICALL Java_org_tensorflow_Session_allocate2(
   TF_Graph* graph = reinterpret_cast(graph_handle);
   TF_Status* status = TF_NewStatus();
   TF_SessionOptions* opts = TF_NewSessionOptions();
-  const char* ctarget = nullptr;
   jbyte* cconfig = nullptr;
-  if (target != nullptr) {
-    ctarget = env->GetStringUTFChars(target, nullptr);
-  }
   if (config != nullptr) {
     cconfig = env->GetByteArrayElements(config, nullptr);
     TF_SetConfig(opts, cconfig,
                  static_cast(env->GetArrayLength(config)), status);
     if (!throwExceptionIfNotOK(env, status)) {
       env->ReleaseByteArrayElements(config, cconfig, JNI_ABORT);
+      TF_DeleteSessionOptions(opts);
+      TF_DeleteStatus(status);
       return 0;
     }
   }
+  const char* ctarget = nullptr;
+  if (target != nullptr) {
+    ctarget = env->GetStringUTFChars(target, nullptr);
+  }
   TF_Session* session = TF_NewSession(graph, opts, status);
   if (config != nullptr) {
     env->ReleaseByteArrayElements(config, cconfig, JNI_ABORT);
diff --git a/tensorflow/java/src/main/native/utils_jni.cc b/tensorflow/java/src/main/native/utils_jni.cc
new file mode 100644
index 0000000000000000000000000000000000000000..069ac05a1c39408dc02f5bbf9a7fc50fd095cc96
--- /dev/null
+++ b/tensorflow/java/src/main/native/utils_jni.cc
@@ -0,0 +1,53 @@
+/* Copyright 2018 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#include "tensorflow/java/src/main/native/utils_jni.h"
+
+#include "tensorflow/java/src/main/native/exception_jni.h"
+
+void resolveOutputs(JNIEnv* env, const char* type, jlongArray src_op,
+                    jintArray src_index, TF_Output* dst, jint n) {
+  if (env->ExceptionCheck()) return;
+  jint len = env->GetArrayLength(src_op);
+  if (len != n) {
+    throwException(env, kIllegalArgumentException,
+                   "expected %d, got %d %s Operations", n, len, type);
+    return;
+  }
+  len = env->GetArrayLength(src_index);
+  if (len != n) {
+    throwException(env, kIllegalArgumentException,
+                   "expected %d, got %d %s Operation output indices", n, len,
+                   type);
+    return;
+  }
+  jlong* op_handles = env->GetLongArrayElements(src_op, nullptr);
+  jint* indices = env->GetIntArrayElements(src_index, nullptr);
+  for (int i = 0; i < n; ++i) {
+    if (op_handles[i] == 0) {
+      throwException(env, kNullPointerException, "invalid %s (#%d of %d)", type,
+                     i, n);
+      break;
+    }
+    dst[i] = TF_Output{reinterpret_cast(op_handles[i]),
+                       static_cast(indices[i])};
+  }
+  env->ReleaseIntArrayElements(src_index, indices, JNI_ABORT);
+  env->ReleaseLongArrayElements(src_op, op_handles, JNI_ABORT);
+}
+
+
+
+
diff --git a/tensorflow/java/src/main/native/utils_jni.h b/tensorflow/java/src/main/native/utils_jni.h
new file mode 100644
index 0000000000000000000000000000000000000000..352298e7de1d07cebc1a287774c9bef85c9a6ae4
--- /dev/null
+++ b/tensorflow/java/src/main/native/utils_jni.h
@@ -0,0 +1,33 @@
+/* Copyright 2018 The TensorFlow Authors. All Rights Reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+==============================================================================*/
+
+#ifndef TENSORFLOW_JAVA_UTILS_JNI_H_
+#define TENSORFLOW_JAVA_UTILS_JNI_H_
+
+#include 
+
+#include "tensorflow/c/c_api.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif  // __cplusplus
+
+void resolveOutputs(JNIEnv* env, const char* type, jlongArray src_op,
+                    jintArray src_index, TF_Output* dst, jint n);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif  // __cplusplus
+#endif /* TENSORFLOW_JAVA_UTILS_JNI_H_ */
diff --git a/tensorflow/java/src/test/java/org/tensorflow/GraphTest.java b/tensorflow/java/src/test/java/org/tensorflow/GraphTest.java
index c540299bdcfcd7bc5969caf82b29144bad24201f..c2e52c22c6dc58a3002b536e64c4607b675804f7 100644
--- a/tensorflow/java/src/test/java/org/tensorflow/GraphTest.java
+++ b/tensorflow/java/src/test/java/org/tensorflow/GraphTest.java
@@ -22,6 +22,7 @@ import static org.junit.Assert.assertTrue;
 
 import java.util.HashSet;
 import java.util.Iterator;
+
 import org.junit.Test;
 import org.junit.runner.RunWith;
 import org.junit.runners.JUnit4;
@@ -129,4 +130,106 @@ public class GraphTest {
       // expected exception.
     }
   }
+
+  @Test
+  public void addGradientsToGraph() {
+    try (Graph g = new Graph();
+        Session s = new Session(g)) {
+
+      Output x1 = TestUtil.placeholder(g, "x1", Float.class);
+      Output x2 = TestUtil.placeholder(g, "x2", Float.class);
+      Output y0 = TestUtil.square(g, "y0", x1);
+      Output y1 = TestUtil.square(g, "y1", y0);
+      Output y2 = TestUtil.addN(g, y0, x2);
+      
+      Output[] grads0 = g.addGradients(y1, toArray(x1));
+      assertNotNull(grads0);
+      assertEquals(1, grads0.length);
+      assertEquals(DataType.FLOAT, grads0[0].dataType());
+
+      Output[] grads1 = g.addGradients(y2, toArray(x1, x2));
+      assertNotNull(grads1);
+      assertEquals(2, grads1.length);
+      assertEquals(DataType.FLOAT, grads1[0].dataType());
+      assertEquals(DataType.FLOAT, grads1[1].dataType());
+      
+      try (Tensor c1 = Tensors.create(3.0f);
+          Tensor c2 = Tensors.create(2.0f);
+          TestUtil.AutoCloseableList> outputs = new TestUtil.AutoCloseableList<>(
+              s.runner()
+                  .feed(x1, c1)
+                  .feed(x2, c2)
+                  .fetch(grads0[0])
+                  .fetch(grads1[0])
+                  .fetch(grads1[1])
+                  .run())) {
+     
+        assertEquals(3, outputs.size());
+        assertEquals(108.0f, outputs.get(0).floatValue(), 0.0f);
+        assertEquals(6.0f, outputs.get(1).floatValue(), 0.0f);
+        assertEquals(1.0f, outputs.get(2).floatValue(), 0.0f);
+      }
+    }
+  }
+
+  @Test
+  public void addGradientSumsToGraph() {
+    try (Graph g = new Graph();
+        Session s = new Session(g)) {
+
+      Output x = TestUtil.placeholder(g, "x", Float.class);
+      Output y0 = TestUtil.square(g, "y0", x);
+      Output y1 = TestUtil.square(g, "y1", y0);
+      
+      Output[] grad = g.addGradients(toArray(y0, y1), toArray(x), null);
+      assertNotNull(grad);
+      assertEquals(1, grad.length);
+      assertEquals(DataType.FLOAT, grad[0].dataType());
+
+      try (Tensor c = Tensors.create(3.0f);
+          Tensor output = s.runner()
+              .feed(x, c)
+              .fetch(grad[0])
+              .run()
+              .get(0)) {
+     
+        assertEquals(114.0f, output.floatValue(), 0.0f);
+      }
+    }
+  }
+
+  @Test
+  public void addGradientsWithInitialValuesToGraph() {
+    try (Graph g = new Graph();
+        Session s = new Session(g)) {
+
+      Output x = TestUtil.placeholder(g, "x", Float.class);
+      Output y0 = TestUtil.square(g, "y0", x);
+      Output y1 = TestUtil.square(g, "y1", y0);
+      
+      Output[] grad0 = g.addGradients(y1, toArray(y0));
+      assertNotNull(grad0);
+      assertEquals(1, grad0.length);
+      assertEquals(DataType.FLOAT, grad0[0].dataType());
+
+      Output[] grad1 = g.addGradients(toArray(y0), toArray(x), toArray(grad0[0]));
+      assertNotNull(grad1);
+      assertEquals(1, grad1.length);
+      assertEquals(DataType.FLOAT, grad1[0].dataType());
+
+      try (Tensor c = Tensors.create(3.0f);
+          Tensor output = s.runner()
+              .feed(x, c)
+              .fetch(grad1[0])
+              .run()
+              .get(0)) {
+     
+        assertEquals(108.0f, output.floatValue(), 0.0f);
+      }
+    }
+  }
+  
+  private static Output[] toArray(Output... outputs) {
+    return outputs;
+  }
 }
diff --git a/tensorflow/java/src/test/java/org/tensorflow/SessionTest.java b/tensorflow/java/src/test/java/org/tensorflow/SessionTest.java
index e8cc76c2a6458193161a98e17483fe73de107b77..7d5980bcdedebedcd2fa4722e85abc1d598fb4fd 100644
--- a/tensorflow/java/src/test/java/org/tensorflow/SessionTest.java
+++ b/tensorflow/java/src/test/java/org/tensorflow/SessionTest.java
@@ -20,8 +20,6 @@ import static org.junit.Assert.assertEquals;
 import static org.junit.Assert.assertTrue;
 import static org.junit.Assert.fail;
 
-import java.util.ArrayList;
-import java.util.Collection;
 import org.junit.Test;
 import org.junit.runner.RunWith;
 import org.junit.runners.JUnit4;
@@ -36,8 +34,8 @@ public class SessionTest {
         Session s = new Session(g)) {
       TestUtil.transpose_A_times_X(g, new int[][] {{2}, {3}});
       try (Tensor x = Tensors.create(new int[][] {{5}, {7}});
-          AutoCloseableList> outputs =
-              new AutoCloseableList>(s.runner().feed("X", x).fetch("Y").run())) {
+          TestUtil.AutoCloseableList> outputs =
+              new TestUtil.AutoCloseableList>(s.runner().feed("X", x).fetch("Y").run())) {
         assertEquals(1, outputs.size());
         final int[][] expected = {{31}};
         assertArrayEquals(expected, outputs.get(0).copyTo(new int[1][1]));
@@ -53,8 +51,8 @@ public class SessionTest {
       Output feed = g.operation("X").output(0);
       Output fetch = g.operation("Y").output(0);
       try (Tensor x = Tensors.create(new int[][] {{5}, {7}});
-          AutoCloseableList> outputs =
-              new AutoCloseableList>(s.runner().feed(feed, x).fetch(fetch).run())) {
+          TestUtil.AutoCloseableList> outputs =
+              new TestUtil.AutoCloseableList>(s.runner().feed(feed, x).fetch(fetch).run())) {
         assertEquals(1, outputs.size());
         final int[][] expected = {{31}};
         assertArrayEquals(expected, outputs.get(0).copyTo(new int[1][1]));
@@ -112,7 +110,7 @@ public class SessionTest {
                 .setOptions(fullTraceRunOptions())
                 .runAndFetchMetadata();
         // Sanity check on outputs.
-        AutoCloseableList> outputs = new AutoCloseableList>(result.outputs);
+        TestUtil.AutoCloseableList> outputs = new TestUtil.AutoCloseableList>(result.outputs);
         assertEquals(1, outputs.size());
         final int[][] expected = {{31}};
         assertArrayEquals(expected, outputs.get(0).copyTo(new int[1][1]));
@@ -135,8 +133,8 @@ public class SessionTest {
         Session s = new Session(g)) {
       TestUtil.constant(g, "c1", 2718);
       TestUtil.constant(g, "c2", 31415);
-      AutoCloseableList> outputs =
-          new AutoCloseableList>(s.runner().fetch("c2").fetch("c1").run());
+      TestUtil.AutoCloseableList> outputs =
+          new TestUtil.AutoCloseableList>(s.runner().fetch("c2").fetch("c1").run());
       assertEquals(2, outputs.size());
       assertEquals(31415, outputs.get(0).intValue());
       assertEquals(2718, outputs.get(1).intValue());
@@ -164,28 +162,6 @@ public class SessionTest {
         Session s = new Session(g, singleThreadConfigProto())) {}
   }
 
-  private static final class AutoCloseableList extends ArrayList
-      implements AutoCloseable {
-    AutoCloseableList(Collection c) {
-      super(c);
-    }
-
-    @Override
-    public void close() {
-      Exception toThrow = null;
-      for (AutoCloseable c : this) {
-        try {
-          c.close();
-        } catch (Exception e) {
-          toThrow = e;
-        }
-      }
-      if (toThrow != null) {
-        throw new RuntimeException(toThrow);
-      }
-    }
-  }
-
   private static byte[] fullTraceRunOptions() {
     // Ideally this would use the generated Java sources for protocol buffers
     // and end up with something like the snippet below. However, generating
diff --git a/tensorflow/java/src/test/java/org/tensorflow/TestUtil.java b/tensorflow/java/src/test/java/org/tensorflow/TestUtil.java
index c973b5a3d8b2be8ee21710d65732bc1e5c3b520a..4e848864167982c750b390a77a1ab7f5d0d40fe9 100644
--- a/tensorflow/java/src/test/java/org/tensorflow/TestUtil.java
+++ b/tensorflow/java/src/test/java/org/tensorflow/TestUtil.java
@@ -16,9 +16,34 @@ limitations under the License.
 package org.tensorflow;
 
 import java.lang.reflect.Array;
+import java.util.ArrayList;
+import java.util.Collection;
 
 /** Static utility functions. */
 public class TestUtil {
+
+  public static final class AutoCloseableList extends ArrayList
+      implements AutoCloseable {
+    AutoCloseableList(Collection c) {
+      super(c);
+    }
+
+    @Override
+    public void close() {
+      Exception toThrow = null;
+      for (AutoCloseable c : this) {
+        try {
+          c.close();
+        } catch (Exception e) {
+          toThrow = e;
+        }
+      }
+      if (toThrow != null) {
+        throw new RuntimeException(toThrow);
+      }
+    }
+  }
+
   public static  Output constant(Graph g, String name, Object value) {
     try (Tensor t = Tensor.create(value)) {
       return g.opBuilder("Const", name)
@@ -36,7 +61,7 @@ public class TestUtil {
         .output(0);
   }
 
-  public static Output addN(Graph g, Output... inputs) {
+  public static  Output addN(Graph g, Output... inputs) {
     return g.opBuilder("AddN", "AddN").addInputList(inputs).build().output(0);
   }
 
@@ -58,6 +83,13 @@ public class TestUtil {
         .setAttr("num_split", numSplit)
         .build();
   }
+  
+  public static  Output square(Graph g, String name, Output value) {
+    return g.opBuilder("Square", name)
+        .addInput(value)
+        .build()
+        .output(0);
+  }
 
   public static void transpose_A_times_X(Graph g, int[][] a) {
     Output aa = constant(g, "A", a);
diff --git a/tensorflow/python/BUILD b/tensorflow/python/BUILD
index 5d9a5130a08904edb77907e5ce19ffd4a379017d..924db54cbcf287a5387d297ebe7896614fc56b1c 100644
--- a/tensorflow/python/BUILD
+++ b/tensorflow/python/BUILD
@@ -73,7 +73,7 @@ py_library(
     visibility = [
         "//tensorflow:__pkg__",
         "//tensorflow/python/tools:__pkg__",
-        "//tensorflow/tools/api/generator:__pkg__",
+        "//tensorflow/python/tools/api/generator:__pkg__",
     ],
     deps = [
         ":array_ops",
@@ -127,12 +127,14 @@ py_library(
         ":util",
         ":weights_broadcast_ops",
         "//tensorflow/core:protos_all_py",
+        "//tensorflow/python/compat",
         "//tensorflow/python/data",
         "//tensorflow/python/feature_column:feature_column_py",
         "//tensorflow/python/keras",
         "//tensorflow/python/ops/distributions",
         "//tensorflow/python/ops/linalg",
         "//tensorflow/python/ops/losses",
+        "//tensorflow/python/ops/parallel_for",
         "//tensorflow/python/profiler",
         "//tensorflow/python/saved_model",
         "//third_party/py/numpy",
@@ -279,6 +281,9 @@ cc_library(
     name = "ndarray_tensor_bridge",
     srcs = ["lib/core/ndarray_tensor_bridge.cc"],
     hdrs = ["lib/core/ndarray_tensor_bridge.h"],
+    visibility = visibility + [
+        "//learning/deepmind/courier:__subpackages__",
+    ],
     deps = [
         ":bfloat16_lib",
         ":numpy_lib",
@@ -694,6 +699,15 @@ py_library(
     ],
 )
 
+py_library(
+    name = "error_interpolation",
+    srcs = [
+        "framework/error_interpolation.py",
+    ],
+    srcs_version = "PY2AND3",
+    deps = [],
+)
+
 py_library(
     name = "function",
     srcs = ["framework/function.py"],
@@ -808,6 +822,7 @@ py_library(
         ":platform",
         ":registry",
         ":tensor_shape",
+        ":traceable_stack",
         ":util",
         ":versions",
         "//tensorflow/core:protos_all_py",
@@ -873,6 +888,17 @@ py_library(
     ],
 )
 
+# This target is maintained separately from :util to provide separate visibility
+# for legacy users who were granted visibility when the functions were private
+# members of ops.Graph.
+py_library(
+    name = "tf_stack",
+    srcs = ["util/tf_stack.py"],
+    srcs_version = "PY2AND3",
+    visibility = ["//visibility:public"],
+    deps = [],
+)
+
 py_library(
     name = "tensor_shape",
     srcs = ["framework/tensor_shape.py"],
@@ -907,6 +933,16 @@ py_library(
     ],
 )
 
+py_library(
+    name = "traceable_stack",
+    srcs = ["framework/traceable_stack.py"],
+    srcs_version = "PY2AND3",
+    visibility = ["//visibility:public"],
+    deps = [
+        ":util",
+    ],
+)
+
 py_library(
     name = "versions",
     srcs = ["framework/versions.py"],
@@ -996,6 +1032,18 @@ py_test(
     ],
 )
 
+py_test(
+    name = "framework_error_interpolation_test",
+    size = "small",
+    srcs = ["framework/error_interpolation_test.py"],
+    main = "framework/error_interpolation_test.py",
+    srcs_version = "PY2AND3",
+    deps = [
+        ":client_testlib",
+        ":error_interpolation",
+    ],
+)
+
 py_test(
     name = "framework_subscribe_test",
     size = "small",
@@ -1181,6 +1229,21 @@ py_test(
     ],
 )
 
+py_test(
+    name = "framework_traceable_stack_test",
+    size = "small",
+    srcs = ["framework/traceable_stack_test.py"],
+    main = "framework/traceable_stack_test.py",
+    srcs_version = "PY2AND3",
+    deps = [
+        ":framework_test_lib",
+        ":platform_test",
+        ":test_ops",
+        ":traceable_stack",
+        ":util",
+    ],
+)
+
 tf_gen_op_wrapper_py(
     name = "test_ops",
     out = "framework/test_ops.py",
@@ -1961,6 +2024,8 @@ py_library(
         ":math_ops",
         ":platform",
         ":resource_variable_ops",
+        ":sparse_ops",
+        ":tensor_shape",
         ":variables",
     ],
 )
@@ -3725,6 +3790,7 @@ py_library(
     srcs_version = "PY2AND3",
     deps = [
         ":c_api_util",
+        ":error_interpolation",
         ":errors",
         ":framework",
         ":framework_for_generated_wrappers",
@@ -3925,7 +3991,7 @@ tf_cuda_library(
 
 tf_py_test(
     name = "session_test",
-    size = "small",
+    size = "medium",
     srcs = ["client/session_test.py"],
     additional_deps = [
         ":array_ops",
@@ -4050,6 +4116,7 @@ cuda_py_test(
         ":math_ops",
         "//tensorflow/core:protos_all_py",
     ],
+    tags = ["no_windows_gpu"],
 )
 
 py_test(
diff --git a/tensorflow/python/client/session.py b/tensorflow/python/client/session.py
index 35aa37ac6dd721750cd72b54d1b8ef6a70402038..e037925961f2bfc8b8906fa81c2d7908ea590a62 100644
--- a/tensorflow/python/client/session.py
+++ b/tensorflow/python/client/session.py
@@ -361,7 +361,7 @@ class _ListFetchMapper(_FetchMapper):
     for m, vi in zip(self._mappers, self._value_indices):
       results.append(m.build_results([values[j] for j in vi]))
     # Return a value of the original type of the fetches.
-    if self._fetch_type == list:
+    if issubclass(self._fetch_type, list):
       return results
     elif self._fetch_type == tuple:
       return tuple(results)
@@ -1291,7 +1291,7 @@ class BaseSession(SessionInterface):
       raise type(e)(node_def, op, message)
 
   def _extend_graph(self):
-    with self._graph._lock:  # pylint: disable=protected-access
+    with self._graph._session_run_lock():  # pylint: disable=protected-access
       tf_session.ExtendSession(self._session)
 
   # The threshold to run garbage collection to delete dead tensors.
diff --git a/tensorflow/python/client/session_test.py b/tensorflow/python/client/session_test.py
index e49d0671050f557842ad1d3305331d61cd8c9672..b72e029d1ccb688f5992f6cc8695969be5e5e2e3 100644
--- a/tensorflow/python/client/session_test.py
+++ b/tensorflow/python/client/session_test.py
@@ -18,6 +18,7 @@ from __future__ import division
 from __future__ import print_function
 
 import collections
+import random
 import os
 import sys
 import threading
@@ -1040,40 +1041,72 @@ class SessionTest(test_util.TensorFlowTestCase):
       for t in threads:
         t.join()
 
-  def testParallelRunAndBuild(self):
+  @staticmethod
+  def _build_graph():
+    time.sleep(random.random() * 0.1)
+    # Do some graph construction. Try to exercise non-trivial paths.
+    graph = ops.get_default_graph()
+    gdef = None
+    for _ in range(10):
+      x = array_ops.placeholder(dtype=dtypes.float32)
+      with ops.colocate_with(x):
+        y = array_ops.placeholder(dtype=dtypes.float32)
+      with ops.device('/cpu:0'):
+        z = control_flow_ops.while_loop(
+            lambda x, y: x < 10, lambda x, y: (x + 1, x * y), [x, y])
+      with graph._attr_scope({'_a': attr_value_pb2.AttrValue(b=False)}):
+        gradients_impl.gradients(z, [x, y])
+        if gdef is None:
+          gdef = graph.as_graph_def()
+        else:
+          importer.import_graph_def(gdef, name='import')
+
+  def testParallelRunAndSingleBuild(self):
     with session.Session() as sess:
       c = constant_op.constant(5.0)
       stop = threading.Event()
 
       def run_loop():
         while not stop.is_set():
+          time.sleep(random.random() * 0.1)
           self.assertEqual(sess.run(c), 5.0)
 
-      threads = [self.checkedThread(target=run_loop) for _ in range(100)]
+      threads = [self.checkedThread(target=run_loop) for _ in range(10)]
       for t in threads:
         t.start()
 
-      # Do some graph construction. Try to exercise non-trivial paths.
-      graph = ops.get_default_graph()
-      gdef = None
-      for _ in range(10):
-        x = array_ops.placeholder(dtype=dtypes.float32)
-        with ops.colocate_with(x):
-          y = array_ops.placeholder(dtype=dtypes.float32)
-        with ops.device('/cpu:0'):
-          z = control_flow_ops.while_loop(
-              lambda x, y: x < 10, lambda x, y: (x + 1, x * y), [x, y])
-        with graph._attr_scope({'_a': attr_value_pb2.AttrValue(b=False)}):
-          gradients_impl.gradients(z, [x, y])
-          if gdef is None:
-            gdef = graph.as_graph_def()
-          else:
-            importer.import_graph_def(gdef, name='import')
+      SessionTest._build_graph()
 
       stop.set()
       for t in threads:
         t.join()
 
+  def testParallelRunAndParallelBuild(self):
+    with session.Session() as sess:
+      c = constant_op.constant(5.0)
+      stop = threading.Event()
+
+      def run_loop():
+        while not stop.is_set():
+          time.sleep(random.random() * 0.1)
+          self.assertEqual(sess.run(c), 5.0)
+
+      run_threads = [self.checkedThread(target=run_loop) for _ in range(10)]
+      for t in run_threads:
+        t.start()
+
+      build_threads = [self.checkedThread(target=SessionTest._build_graph)
+                       for _ in range(10)]
+      for t in build_threads:
+        t.start()
+      for t in build_threads:
+        t.join()
+
+      # Let the run_threads run until the build threads are finished.
+      stop.set()
+      for t in run_threads:
+        t.join()
+
   def testRunFeedDict(self):
     with session.Session() as s:
       x = array_ops.zeros([2])
diff --git a/tensorflow/python/compat/BUILD b/tensorflow/python/compat/BUILD
new file mode 100644
index 0000000000000000000000000000000000000000..58ceafca0638a90c2e66ddea0e4bbb1547455f48
--- /dev/null
+++ b/tensorflow/python/compat/BUILD
@@ -0,0 +1,22 @@
+licenses(["notice"])  # Apache 2.0
+
+exports_files(["LICENSE"])
+
+load("//tensorflow:tensorflow.bzl", "tf_py_test")
+
+py_library(
+    name = "compat",
+    srcs = ["compat.py"],
+    srcs_version = "PY2AND3",
+    visibility = ["//tensorflow:internal"],
+)
+
+tf_py_test(
+    name = "compat_test",
+    size = "small",
+    srcs = ["compat_test.py"],
+    additional_deps = [
+        ":compat",
+        "//tensorflow/python:client_testlib",
+    ],
+)
diff --git a/tensorflow/python/compat/compat.py b/tensorflow/python/compat/compat.py
new file mode 100644
index 0000000000000000000000000000000000000000..68a6421c2c56c9f007cbd8aee3111c4abfde691c
--- /dev/null
+++ b/tensorflow/python/compat/compat.py
@@ -0,0 +1,125 @@
+# Copyright 2018 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""Utilities for API compatibility between TensorFlow release versions.
+
+See
+@{$guide/version_compat#backward_and_partial_forward_compatibility}
+"""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import datetime
+from tensorflow.python.util import tf_contextlib
+
+_FORWARD_COMPATIBILITY_HORIZON = datetime.date(2018, 8, 1)
+
+
+def forward_compatible(year, month, day):
+  """Return true if the forward compatibility window has expired.
+
+  Forward-compatibility refers to scenarios where the producer of a TensorFlow
+  model (a GraphDef or SavedModel) is compiled against a version of the
+  TensorFlow library newer than what the consumer was compiled against. The
+  "producer" is typically a Python program that constructs and trains a model
+  while the "consumer" is typically another program that loads and serves the
+  model.
+
+  TensorFlow has been supporting a 3 week forward-compatibility window for
+  programs compiled from source at HEAD.
+
+  For example, consider the case where a new operation `MyNewAwesomeAdd` is
+  created with the intent of replacing the implementation of an existing Python
+  wrapper - `tf.add`.  The Python wrapper implementation should change from
+  something like:
+
+  ```python
+  def add(inputs, name=None):
+    return gen_math_ops.add(inputs, name)
+  ```
+
+  to:
+
+  ```python
+  from tensorflow.python.compat import compat
+
+  def add(inputs, name=None):
+    if compat.forward_compatible(year, month, day):
+      # Can use the awesome new implementation.
+      return gen_math_ops.my_new_awesome_add(inputs, name)
+    # To maintain forward compatibiltiy, use the old implementation.
+    return gen_math_ops.add(inputs, name)
+  ```
+
+  Where `year`, `month`, and `day` specify the date beyond which binaries
+  that consume a model are expected to have been updated to include the
+  new operations. This date is typically at least 3 weeks beyond the date
+  the code that adds the new operation is committed.
+
+  Args:
+    year:  A year (e.g., 2018).
+    month: A month (1 <= month <= 12) in year.
+    day:   A day (1 <= day <= 31, or 30, or 29, or 28) in month.
+
+  Returns:
+    True if the caller can expect that serialized TensorFlow graphs produced
+    can be consumed by programs that are compiled with the TensorFlow library
+    source code after (year, month, day).
+  """
+  return _FORWARD_COMPATIBILITY_HORIZON > datetime.date(year, month, day)
+
+
+@tf_contextlib.contextmanager
+def forward_compatibility_horizon(year, month, day):
+  """Context manager for testing forward compatibility of generated graphs.
+
+  To ensure forward compatibility of generated graphs (see `forward_compatible`)
+  with older binaries, new features can be gated with:
+
+  ```python
+  if compat.forward_compatible(year=2018, month=08, date=01):
+    generate_graph_with_new_features()
+  else:
+    generate_graph_so_older_binaries_can_consume_it()
+  ```
+
+  However, when adding new features, one may want to unittest it before
+  the forward compatibility window expires. This context manager enables
+  such tests. For example:
+
+  ```python
+  from tensorflow.python.compat import compat
+
+  def testMyNewFeature(self):
+    with compat.forward_compatibility_horizon(2018, 08, 02):
+       # Test that generate_graph_with_new_features() has an effect
+  ```
+
+  Args :
+    year:  A year (e.g. 2018).
+    month: A month (1 <= month <= 12) in year.
+    day:   A day (1 <= day <= 31, or 30, or 29, or 28) in month.
+
+  Yields:
+    Nothing.
+  """
+  global _FORWARD_COMPATIBILITY_HORIZON
+  try:
+    old_compat_date = _FORWARD_COMPATIBILITY_HORIZON
+    _FORWARD_COMPATIBILITY_HORIZON = datetime.date(year, month, day)
+    yield
+  finally:
+    _FORWARD_COMPATIBILITY_HORIZON = old_compat_date
diff --git a/tensorflow/python/compat/compat_test.py b/tensorflow/python/compat/compat_test.py
new file mode 100644
index 0000000000000000000000000000000000000000..946abbb300d66e7be5ea317e365bc75cbcf6941c
--- /dev/null
+++ b/tensorflow/python/compat/compat_test.py
@@ -0,0 +1,70 @@
+# Copyright 2018 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""Tests for forward and backwards compatibility utilties."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import datetime
+from tensorflow.python.compat import compat
+from tensorflow.python.platform import test
+
+
+class CompatTest(test.TestCase):
+
+  def _compatibility_date(self):
+    date = compat._FORWARD_COMPATIBILITY_HORIZON  # pylint: disable=protected-access
+    return (date.year, date.month, date.day)
+
+  def _n_days_after(self, n):
+    date = compat._FORWARD_COMPATIBILITY_HORIZON + datetime.timedelta(days=n)  # pylint: disable=protected-access
+    return (date.year, date.month, date.day)
+
+  def test_basic(self):
+    compatibility_date = self._compatibility_date()
+    one_day_before = self._n_days_after(-1)
+    self.assertTrue(compat.forward_compatible(*one_day_before))
+    self.assertFalse(compat.forward_compatible(*compatibility_date))
+
+  def test_decorator(self):
+    compatibility_date = self._compatibility_date()
+    one_day_after = self._n_days_after(1)
+    with compat.forward_compatibility_horizon(*one_day_after):
+      self.assertTrue(compat.forward_compatible(*compatibility_date))
+      self.assertFalse(compat.forward_compatible(*one_day_after))
+
+    # After exiting context manager, value should be reset.
+    self.assertFalse(compat.forward_compatible(*compatibility_date))
+
+  def test_decorator_with_failure(self):
+    compatibility_date = self._compatibility_date()
+    one_day_after = self._n_days_after(1)
+
+    class DummyError(Exception):
+      pass
+
+    try:
+      with compat.forward_compatibility_horizon(*one_day_after):
+        raise DummyError()
+    except DummyError:
+      pass  # silence DummyError
+
+    # After exiting context manager, value should be reset.
+    self.assertFalse(compat.forward_compatible(*compatibility_date))
+
+
+if __name__ == '__main__':
+  test.main()
diff --git a/tensorflow/python/data/kernel_tests/BUILD b/tensorflow/python/data/kernel_tests/BUILD
index 3bde62fa1d8a71c0d6f2bbfbff29bb842a9248f0..38505c0a01133509e682e8750ddd62192bcceb82 100644
--- a/tensorflow/python/data/kernel_tests/BUILD
+++ b/tensorflow/python/data/kernel_tests/BUILD
@@ -349,6 +349,7 @@ tf_py_test(
         "//tensorflow/python:sparse_tensor",
         "//tensorflow/python:tensor_shape",
         "//tensorflow/python:training",
+        "//tensorflow/python/compat:compat",
     ],
     grpc_enabled = True,
 )
diff --git a/tensorflow/python/data/kernel_tests/batch_dataset_op_test.py b/tensorflow/python/data/kernel_tests/batch_dataset_op_test.py
index 50bb0837b7052d67ced4fdf5c9c7e96212bdb415..89de55dd4f9fdc612663c839b926684d27d48c54 100644
--- a/tensorflow/python/data/kernel_tests/batch_dataset_op_test.py
+++ b/tensorflow/python/data/kernel_tests/batch_dataset_op_test.py
@@ -18,9 +18,12 @@ from __future__ import absolute_import
 from __future__ import division
 from __future__ import print_function
 
+import time
+
 from absl.testing import parameterized
 import numpy as np
 
+from tensorflow.python.client import session
 from tensorflow.python.data.ops import dataset_ops
 from tensorflow.python.framework import constant_op
 from tensorflow.python.framework import dtypes
@@ -275,7 +278,7 @@ class PaddedBatchDatasetTest(test.TestCase, parameterized.TestCase):
         result = sess.run(get_next)
         padded_len = padded_shapes[0]
         if padded_len is None or padded_len == -1:
-          padded_len = np.max(result)
+          padded_len = np.max(result) if result.size > 0 else 0
         self.assertEqual((batch_size, padded_len), result.shape)
         for j in range(batch_size):
           seq_len = seq_lens[(i * batch_size) + j]
@@ -285,7 +288,7 @@ class PaddedBatchDatasetTest(test.TestCase, parameterized.TestCase):
 
       if not drop_remainder and len(seq_lens) % batch_size > 0:
         result = sess.run(get_next)
-        padded_len = np.max(result)
+        padded_len = np.max(result) if result.size > 0 else 0
         self.assertEqual((len(seq_lens) % batch_size, padded_len),
                          result.shape)
         for j in range(len(seq_lens) % batch_size):
@@ -461,5 +464,55 @@ class PaddedBatchDatasetTest(test.TestCase, parameterized.TestCase):
           5, padded_shapes=shape_as_tensor)
 
 
+class BatchDatasetBenchmark(test.Benchmark):
+
+  def benchmarkBatchSparse(self):
+    non_zeros_per_row_values = [0, 1, 5, 10, 100]
+    batch_size_values = [1, 32, 64, 128, 1024]
+
+    sparse_placeholder = array_ops.sparse_placeholder(dtype=dtypes.int64)
+    batch_size_placeholder = array_ops.placeholder(dtype=dtypes.int64, shape=[])
+
+    dataset = dataset_ops.Dataset.from_tensors(sparse_placeholder).repeat(
+        ).batch(batch_size_placeholder)
+    iterator = dataset.make_initializable_iterator()
+    next_element = iterator.get_next()
+
+    for non_zeros_per_row in non_zeros_per_row_values:
+
+      sparse_value = sparse_tensor.SparseTensorValue(
+          indices=np.arange(non_zeros_per_row, dtype=np.int64)[:, np.newaxis],
+          values=np.arange(non_zeros_per_row, dtype=np.int64),
+          dense_shape=[1000])
+
+      for batch_size in batch_size_values:
+
+        with session.Session() as sess:
+          sess.run(iterator.initializer, feed_dict={
+              sparse_placeholder: sparse_value,
+              batch_size_placeholder: batch_size})
+          # Run five steps to warm up the session caches before taking the
+          # first measurement.
+          for _ in range(5):
+            sess.run(next_element.indices.op)
+          deltas = []
+          for _ in range(100):
+            start = time.time()
+            for _ in range(100):
+              sess.run(next_element.indices.op)
+            end = time.time()
+            deltas.append(end - start)
+
+        median_wall_time = np.median(deltas) / 100.0
+
+        print('Batch sparse dataset non-zeros per row: %d batch_size: %d '
+              'wall time: %f'
+              % (non_zeros_per_row, batch_size, median_wall_time))
+        self.report_benchmark(
+            iters=10000, wall_time=median_wall_time,
+            name='benchmark_batch_sparse_dataset_nnz_%d_batch_size_%d' % (
+                non_zeros_per_row, batch_size))
+
+
 if __name__ == '__main__':
   test.main()
diff --git a/tensorflow/python/data/kernel_tests/iterator_ops_test.py b/tensorflow/python/data/kernel_tests/iterator_ops_test.py
index 820c167b6bb9dc3b1c25d9c6156cef17ad20eb1b..b434fa7334398674a442f2ee5aa21de41b290cc4 100644
--- a/tensorflow/python/data/kernel_tests/iterator_ops_test.py
+++ b/tensorflow/python/data/kernel_tests/iterator_ops_test.py
@@ -25,6 +25,7 @@ import numpy as np
 from tensorflow.core.protobuf import cluster_pb2
 from tensorflow.core.protobuf import config_pb2
 from tensorflow.python.client import session
+from tensorflow.python.compat import compat as forward_compat
 from tensorflow.python.data.ops import dataset_ops
 from tensorflow.python.data.ops import iterator_ops
 from tensorflow.python.data.ops import readers
@@ -415,6 +416,69 @@ class IteratorTest(test.TestCase):
         sess.run(
             next_element, feed_dict={handle_placeholder: iterator_4_handle})
 
+  def testIteratorStringHandleFuture(self):
+    with forward_compat.forward_compatibility_horizon(2018, 8, 4):
+      dataset_3 = dataset_ops.Dataset.from_tensor_slices([1, 2, 3])
+      dataset_4 = dataset_ops.Dataset.from_tensor_slices([10, 20, 30, 40])
+
+      iterator_3 = dataset_3.make_one_shot_iterator()
+      iterator_4 = dataset_4.make_one_shot_iterator()
+
+      handle_placeholder = array_ops.placeholder(dtypes.string, shape=[])
+      feedable_iterator = iterator_ops.Iterator.from_string_handle(
+          handle_placeholder, dataset_3.output_types, dataset_3.output_shapes)
+      next_element = feedable_iterator.get_next()
+
+      self.assertEqual(dataset_3.output_types, feedable_iterator.output_types)
+      self.assertEqual(dataset_4.output_types, feedable_iterator.output_types)
+      self.assertEqual([], feedable_iterator.output_shapes)
+
+      with self.test_session() as sess:
+        iterator_3_handle = sess.run(iterator_3.string_handle())
+        iterator_4_handle = sess.run(iterator_4.string_handle())
+
+        self.assertEqual(
+            10,
+            sess.run(
+                next_element,
+                feed_dict={handle_placeholder: iterator_4_handle}))
+        self.assertEqual(
+            1,
+            sess.run(
+                next_element,
+                feed_dict={handle_placeholder: iterator_3_handle}))
+        self.assertEqual(
+            20,
+            sess.run(
+                next_element,
+                feed_dict={handle_placeholder: iterator_4_handle}))
+        self.assertEqual(
+            2,
+            sess.run(
+                next_element,
+                feed_dict={handle_placeholder: iterator_3_handle}))
+        self.assertEqual(
+            30,
+            sess.run(
+                next_element,
+                feed_dict={handle_placeholder: iterator_4_handle}))
+        self.assertEqual(
+            3,
+            sess.run(
+                next_element,
+                feed_dict={handle_placeholder: iterator_3_handle}))
+        self.assertEqual(
+            40,
+            sess.run(
+                next_element,
+                feed_dict={handle_placeholder: iterator_4_handle}))
+        with self.assertRaises(errors.OutOfRangeError):
+          sess.run(
+              next_element, feed_dict={handle_placeholder: iterator_3_handle})
+        with self.assertRaises(errors.OutOfRangeError):
+          sess.run(
+              next_element, feed_dict={handle_placeholder: iterator_4_handle})
+
   def testIteratorStringHandleReuseTensorObject(self):
     dataset = dataset_ops.Dataset.from_tensor_slices([1, 2, 3])
     one_shot_iterator = dataset.make_one_shot_iterator()
diff --git a/tensorflow/python/data/kernel_tests/map_dataset_op_test.py b/tensorflow/python/data/kernel_tests/map_dataset_op_test.py
index 0ecd821e9e473522b0cf4bd7bbceb071ecf5bb9e..637bde9ae4eb839e2b983ceec082f868f3ed2728 100644
--- a/tensorflow/python/data/kernel_tests/map_dataset_op_test.py
+++ b/tensorflow/python/data/kernel_tests/map_dataset_op_test.py
@@ -666,6 +666,13 @@ class MapDatasetTest(test.TestCase):
         "currently support nested datasets as outputs."):
       _ = dataset.map(dataset_ops.Dataset.from_tensor_slices)
 
+  def testReturnValueError(self):
+    dataset = dataset_ops.Dataset.from_tensors([1.0, 2.0, 3.0])
+    with self.assertRaisesRegexp(
+        TypeError, r"Unsupported return value from function passed to "
+        r"Dataset.map\(\): None."):
+      _ = dataset.map(lambda x: None)
+
 
 class MapDatasetBenchmark(test.Benchmark):
 
diff --git a/tensorflow/python/data/ops/BUILD b/tensorflow/python/data/ops/BUILD
index fa2e86eab18b0b97ea01a96e309b0ea82d91b267..f15eb6310f6176338155c4c0b370f59db7cfa210 100644
--- a/tensorflow/python/data/ops/BUILD
+++ b/tensorflow/python/data/ops/BUILD
@@ -40,6 +40,7 @@ py_library(
         "//tensorflow/python:dtypes",
         "//tensorflow/python:framework_ops",
         "//tensorflow/python:tensor_shape",
+        "//tensorflow/python/compat",
         "//tensorflow/python/data/util:convert",
     ],
 )
@@ -54,6 +55,7 @@ py_library(
         "//tensorflow/python:framework_ops",
         "//tensorflow/python:resource_variable_ops",
         "//tensorflow/python:tensor_shape",
+        "//tensorflow/python/compat",
         "//tensorflow/python/data/util:nest",
         "//tensorflow/python/data/util:sparse",
         "//tensorflow/python/eager:context",
diff --git a/tensorflow/python/data/ops/dataset_ops.py b/tensorflow/python/data/ops/dataset_ops.py
index 7cb6627615461efec074d9ae02ce7dd4c57f86b9..88de4b588cc3369e9d67a03c600e68186bb267ad 100644
--- a/tensorflow/python/data/ops/dataset_ops.py
+++ b/tensorflow/python/data/ops/dataset_ops.py
@@ -24,6 +24,7 @@ import warnings
 import numpy as np
 import six
 
+from tensorflow.python.compat import compat
 from tensorflow.python.data.ops import iterator_ops
 from tensorflow.python.data.util import nest
 from tensorflow.python.data.util import random_seed
@@ -107,8 +108,12 @@ class Dataset(object):
           "execution is enabled.")
     if shared_name is None:
       shared_name = ""
-    iterator_resource = gen_dataset_ops.iterator(
-        container="", shared_name=shared_name, **flat_structure(self))
+    if compat.forward_compatible(2018, 8, 3):
+      iterator_resource = gen_dataset_ops.iterator_v2(
+          container="", shared_name=shared_name, **flat_structure(self))
+    else:
+      iterator_resource = gen_dataset_ops.iterator(
+          container="", shared_name=shared_name, **flat_structure(self))
     with ops.colocate_with(iterator_resource):
       initializer = gen_dataset_ops.make_iterator(self._as_variant_tensor(),
                                                   iterator_resource)
@@ -888,7 +893,83 @@ class Dataset(object):
                               drop_remainder)
 
   def map(self, map_func, num_parallel_calls=None):
-    """Maps `map_func` across this dataset.
+    """Maps `map_func` across the elements of this dataset.
+
+    This transformation applies `map_func` to each element of this dataset, and
+    returns a new dataset containing the transformed elements, in the same
+    order as they appeared in the input.
+
+    For example:
+
+    ```python
+    # NOTE: The following examples use `{ ... }` to represent the
+    # contents of a dataset.
+    a = { 1, 2, 3, 4, 5 }
+
+    a.map(lambda x: x + 1) = { 2, 3, 4, 5, 6 }
+    ```
+
+    The input signature of `map_func` is determined by the structure of each
+    element in this dataset. For example:
+
+    ```python
+    # Each element is a `tf.Tensor` object.
+    a = { 1, 2, 3, 4, 5 }
+    # `map_func` takes a single argument of type `tf.Tensor` with the same
+    # shape and dtype.
+    result = a.map(lambda x: ...)
+
+    # Each element is a tuple containing two `tf.Tensor` objects.
+    b = { (1, "foo"), (2, "bar"), (3, "baz") }
+    # `map_func` takes two arguments of type `tf.Tensor`.
+    result = b.map(lambda x_int, y_str: ...)
+
+    # Each element is a dictionary mapping strings to `tf.Tensor` objects.
+    c = { {"a": 1, "b": "foo"}, {"a": 2, "b": "bar"}, {"a": 3, "b": "baz"} }
+    # `map_func` takes a single argument of type `dict` with the same keys as
+    # the elements.
+    result = c.map(lambda d: ...)
+    ```
+
+    The value or values returned by `map_func` determine the structure of each
+    element in the returned dataset.
+
+    ```python
+    # `map_func` returns a scalar `tf.Tensor` of type `tf.float32`.
+    def f(...):
+      return tf.constant(37.0)
+    result = dataset.map(f)
+    result.output_classes == tf.Tensor
+    result.output_types == tf.float32
+    result.output_shapes == []  # scalar
+
+    # `map_func` returns two `tf.Tensor` objects.
+    def g(...):
+      return tf.constant(37.0), tf.constant(["Foo", "Bar", "Baz"])
+    result = dataset.map(g)
+    result.output_classes == (tf.Tensor, tf.Tensor)
+    result.output_types == (tf.float32, tf.string)
+    result.output_shapes == ([], [3])
+
+    # Python primitives, lists, and NumPy arrays are implicitly converted to
+    # `tf.Tensor`.
+    def h(...):
+      return 37.0, ["Foo", "Bar", "Baz"], np.array([1.0, 2.0] dtype=np.float64)
+    result = dataset.map(h)
+    result.output_classes == (tf.Tensor, tf.Tensor, tf.Tensor)
+    result.output_types == (tf.float32, tf.string, tf.float64)
+    result.output_shapes == ([], [3], [2])
+
+    # `map_func` can return nested structures.
+    def i(...):
+      return {"a": 37.0, "b": [42, 16]}, "foo"
+    result.output_classes == ({"a": tf.Tensor, "b": tf.Tensor}, tf.Tensor)
+    result.output_types == ({"a": tf.float32, "b": tf.int32}, tf.string)
+    result.output_shapes == ({"a": [], "b": [2]}, [])
+    ```
+
+    In addition to `tf.Tensor` objects, `map_func` can accept as arguments and
+    return `tf.SparseTensor` objects.
 
     Args:
       map_func: A function mapping a nested structure of tensors (having
@@ -1168,10 +1249,29 @@ class _NestedDatasetComponent(object):
   custom component types.
   """
 
-  def __init__(self, dataset):
-    self._output_classes = dataset.output_classes
-    self._output_shapes = dataset.output_shapes
-    self._output_types = dataset.output_types
+  def __init__(self,
+               dataset=None,
+               output_shapes=None,
+               output_types=None,
+               output_classes=None):
+    if dataset is None:
+      if (output_classes is None or output_shapes is None or
+          output_types is None):
+        raise ValueError(
+            "Either `dataset`, or all of `output_classes`, "
+            "`output_shapes`, and `output_types` must be specified.")
+      self._output_classes = output_classes
+      self._output_shapes = output_shapes
+      self._output_types = output_types
+    else:
+      if not (output_classes is None and output_shapes is None and
+              output_types is None):
+        raise ValueError(
+            "Either `dataset`, or all of `output_classes`, "
+            "`output_shapes`, and `output_types` must be specified.")
+      self._output_classes = dataset.output_classes
+      self._output_shapes = dataset.output_shapes
+      self._output_types = dataset.output_types
 
   @property
   def output_classes(self):
@@ -1330,7 +1430,11 @@ class StructuredFunctionWrapper(object):
           flat_shapes.append(component)
           flat_types.append(component)
         else:
-          t = ops.convert_to_tensor(t)
+          try:
+            t = ops.convert_to_tensor(t)
+          except (ValueError, TypeError):
+            raise TypeError("Unsupported return value from function passed to "
+                            "%s: %s." % (transformation_name, t))
           flat_ret.append(t)
           flat_classes.append(ops.Tensor)
           flat_shapes.append(t.get_shape())
@@ -1406,11 +1510,30 @@ def flat_structure(dataset):
     A dictionary of keyword arguments that can be passed to many Dataset op
     constructors.
   """
+  output_classes = []
+  output_shapes = []
+  output_types = []
+  for output_class, output_shape, output_type in zip(
+      nest.flatten(dataset.output_classes), nest.flatten(dataset.output_shapes),
+      nest.flatten(dataset.output_types)):
+    if isinstance(output_class, _NestedDatasetComponent):
+      output_classes.append(output_class.output_classes)
+      output_shapes.append(output_shape.output_shapes)
+      output_types.append(output_type.output_types)
+    else:
+      output_classes.append(output_class)
+      output_shapes.append(output_shape)
+      output_types.append(output_type)
+
+  output_classes = nest.pack_sequence_as(dataset.output_classes, output_classes)
+  output_shapes = nest.pack_sequence_as(dataset.output_shapes, output_shapes)
+  output_types = nest.pack_sequence_as(dataset.output_types, output_types)
+
   return {
-      "output_shapes": nest.flatten(sparse.as_dense_shapes(
-          dataset.output_shapes, dataset.output_classes)),
-      "output_types": nest.flatten(sparse.as_dense_types(
-          dataset.output_types, dataset.output_classes)),
+      "output_shapes":
+          nest.flatten(sparse.as_dense_shapes(output_shapes, output_classes)),
+      "output_types":
+          nest.flatten(sparse.as_dense_types(output_types, output_classes)),
   }
 
 
diff --git a/tensorflow/python/data/ops/iterator_ops.py b/tensorflow/python/data/ops/iterator_ops.py
index b6dba4e3ca3874b8e9bc3b7ea92fb91fe41759d8..35de2f2841604d95fa3363b0b8e194ec1723f554 100644
--- a/tensorflow/python/data/ops/iterator_ops.py
+++ b/tensorflow/python/data/ops/iterator_ops.py
@@ -20,6 +20,7 @@ from __future__ import print_function
 import threading
 import warnings
 
+from tensorflow.python.compat import compat
 from tensorflow.python.data.util import nest
 from tensorflow.python.data.util import sparse
 from tensorflow.python.eager import context
@@ -172,13 +173,32 @@ class Iterator(object):
     nest.assert_same_structure(output_types, output_shapes)
     if shared_name is None:
       shared_name = ""
-    iterator_resource = gen_dataset_ops.iterator(
-        container="",
-        shared_name=shared_name,
-        output_types=nest.flatten(
-            sparse.as_dense_types(output_types, output_classes)),
-        output_shapes=nest.flatten(
-            sparse.as_dense_shapes(output_shapes, output_classes)))
+    if compat.forward_compatible(2018, 8, 3):
+      if not ops.get_default_graph()._graph_device_function_stack:  # pylint: disable=protected-access
+        with ops.device("/cpu:0"):
+          iterator_resource = gen_dataset_ops.iterator_v2(
+              container="",
+              shared_name=shared_name,
+              output_types=nest.flatten(
+                  sparse.as_dense_types(output_types, output_classes)),
+              output_shapes=nest.flatten(
+                  sparse.as_dense_shapes(output_shapes, output_classes)))
+      else:
+        iterator_resource = gen_dataset_ops.iterator_v2(
+            container="",
+            shared_name=shared_name,
+            output_types=nest.flatten(
+                sparse.as_dense_types(output_types, output_classes)),
+            output_shapes=nest.flatten(
+                sparse.as_dense_shapes(output_shapes, output_classes)))
+    else:
+      iterator_resource = gen_dataset_ops.iterator(
+          container="",
+          shared_name=shared_name,
+          output_types=nest.flatten(
+              sparse.as_dense_types(output_types, output_classes)),
+          output_shapes=nest.flatten(
+              sparse.as_dense_shapes(output_shapes, output_classes)))
     return Iterator(iterator_resource, None, output_types, output_shapes,
                     output_classes)
 
@@ -242,12 +262,29 @@ class Iterator(object):
       output_classes = nest.map_structure(lambda _: ops.Tensor, output_types)
     nest.assert_same_structure(output_types, output_shapes)
     string_handle = ops.convert_to_tensor(string_handle, dtype=dtypes.string)
-    iterator_resource = gen_dataset_ops.iterator_from_string_handle(
-        string_handle,
-        output_types=nest.flatten(
-            sparse.as_dense_types(output_types, output_classes)),
-        output_shapes=nest.flatten(
-            sparse.as_dense_shapes(output_shapes, output_classes)))
+    if compat.forward_compatible(2018, 8, 3):
+      if not ops.get_default_graph()._graph_device_function_stack:  # pylint: disable=protected-access
+        with ops.device("/cpu:0"):
+          iterator_resource = gen_dataset_ops.iterator_from_string_handle_v2(
+              string_handle,
+              output_types=nest.flatten(
+                  sparse.as_dense_types(output_types, output_classes)),
+              output_shapes=nest.flatten(
+                  sparse.as_dense_shapes(output_shapes, output_classes)))
+      else:
+        iterator_resource = gen_dataset_ops.iterator_from_string_handle_v2(
+            string_handle,
+            output_types=nest.flatten(
+                sparse.as_dense_types(output_types, output_classes)),
+            output_shapes=nest.flatten(
+                sparse.as_dense_shapes(output_shapes, output_classes)))
+    else:
+      iterator_resource = gen_dataset_ops.iterator_from_string_handle(
+          string_handle,
+          output_types=nest.flatten(
+              sparse.as_dense_types(output_types, output_classes)),
+          output_shapes=nest.flatten(
+              sparse.as_dense_shapes(output_shapes, output_classes)))
     return Iterator(iterator_resource, None, output_types, output_shapes,
                     output_classes)
 
diff --git a/tensorflow/python/debug/BUILD b/tensorflow/python/debug/BUILD
index 6941cacf23e225cb90086a54070a05c5a9dea09a..27b8ebd362eea4468d20c65ee39e1b55e8dcd17d 100644
--- a/tensorflow/python/debug/BUILD
+++ b/tensorflow/python/debug/BUILD
@@ -404,6 +404,7 @@ py_library(
     deps = [
         ":debug_errors",
         ":debug_fibonacci",
+        ":debug_keras",
         ":debug_mnist",
         ":debug_tflearn_iris",
     ],
@@ -454,6 +455,17 @@ py_binary(
     ],
 )
 
+py_binary(
+    name = "debug_keras",
+    srcs = ["examples/debug_keras.py"],
+    srcs_version = "PY2AND3",
+    deps = [
+        ":debug_py",
+        "//tensorflow:tensorflow_py",
+        "//third_party/py/numpy",
+    ],
+)
+
 py_test(
     name = "common_test",
     size = "small",
@@ -791,6 +803,7 @@ cuda_py_test(
         "//tensorflow/python:platform_test",
         "//tensorflow/python:variables",
     ],
+    tags = ["no_windows_gpu"],
 )
 
 py_test(
@@ -1086,6 +1099,7 @@ py_test(
         "//tensorflow/python:state_ops",
         "//tensorflow/python:training",
         "//tensorflow/python:variables",
+        "//third_party/py/numpy",
     ],
 )
 
@@ -1096,6 +1110,7 @@ sh_test(
     data = [
         ":debug_errors",
         ":debug_fibonacci",
+        ":debug_keras",
         ":debug_mnist",
         ":debug_tflearn_iris",
         ":offline_analyzer",
diff --git a/tensorflow/python/debug/examples/debug_keras.py b/tensorflow/python/debug/examples/debug_keras.py
new file mode 100644
index 0000000000000000000000000000000000000000..3272d85ade957b254b2c1a0977156179cd71bb9d
--- /dev/null
+++ b/tensorflow/python/debug/examples/debug_keras.py
@@ -0,0 +1,89 @@
+# Copyright 2016 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""tfdbg example: debugging tf.keras models training on tf.data.Dataset."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import argparse
+import sys
+
+import numpy as np
+import tensorflow as tf
+
+from tensorflow.python import debug as tf_debug
+
+
+def main(_):
+  # Create a dummy dataset.
+  num_examples = 8
+  steps_per_epoch = 2
+  input_dims = 3
+  output_dims = 1
+  xs = np.zeros([num_examples, input_dims])
+  ys = np.zeros([num_examples, output_dims])
+  dataset = tf.data.Dataset.from_tensor_slices(
+      (xs, ys)).repeat(num_examples).batch(int(num_examples / steps_per_epoch))
+
+  sess = tf.Session()
+  if FLAGS.debug:
+    # Use the command-line interface (CLI) of tfdbg.
+    sess = tf_debug.LocalCLIDebugWrapperSession(sess, ui_type=FLAGS.ui_type)
+  elif FLAGS.tensorboard_debug_address:
+    # Use the TensorBoard Debugger Plugin (GUI of tfdbg).
+    sess = tf_debug.TensorBoardDebugWrapperSession(
+        sess, FLAGS.tensorboard_debug_address)
+  tf.keras.backend.set_session(sess)
+
+  # Create a dummy model.
+  model = tf.keras.Sequential([
+      tf.keras.layers.Dense(1, input_shape=[input_dims])])
+  model.compile(loss="mse", optimizer="sgd")
+
+  # Train the model using the dummy dataset created above.
+  model.fit(dataset, epochs=FLAGS.epochs, steps_per_epoch=steps_per_epoch)
+
+
+if __name__ == "__main__":
+  parser = argparse.ArgumentParser()
+  parser.register("type", "bool", lambda v: v.lower() == "true")
+  parser.add_argument(
+      "--debug",
+      type="bool",
+      nargs="?",
+      const=True,
+      default=False,
+      help="Use debugger to track down bad values during training. "
+      "Mutually exclusive with the --tensorboard_debug_address flag.")
+  parser.add_argument(
+      "--ui_type",
+      type=str,
+      default="curses",
+      help="Command-line user interface type (curses | readline).")
+  parser.add_argument(
+      "--tensorboard_debug_address",
+      type=str,
+      default=None,
+      help="Connect to the TensorBoard Debugger Plugin backend specified by "
+      "the gRPC address (e.g., localhost:1234). Mutually exclusive with the "
+      "--debug flag.")
+  parser.add_argument(
+      "--epochs",
+      type=int,
+      default=2,
+      help="Number of epochs to train the model for.")
+  FLAGS, unparsed = parser.parse_known_args()
+  tf.app.run(main=main, argv=[sys.argv[0]] + unparsed)
diff --git a/tensorflow/python/debug/examples/examples_test.sh b/tensorflow/python/debug/examples/examples_test.sh
index e9c45a7e6e92d069f51648647620f7a7c3a5aadc..2d35b2d8bb10d17decfa404afd5004d3409c06e5 100755
--- a/tensorflow/python/debug/examples/examples_test.sh
+++ b/tensorflow/python/debug/examples/examples_test.sh
@@ -48,12 +48,14 @@ if [[ -z "${PYTHON_BIN_PATH}" ]]; then
   DEBUG_ERRORS_BIN="$TEST_SRCDIR/org_tensorflow/tensorflow/python/debug/debug_errors"
   DEBUG_MNIST_BIN="$TEST_SRCDIR/org_tensorflow/tensorflow/python/debug/debug_mnist"
   DEBUG_TFLEARN_IRIS_BIN="$TEST_SRCDIR/org_tensorflow/tensorflow/python/debug/debug_tflearn_iris"
+  DEBUG_KERAS_BIN="$TEST_SRCDIR/org_tensorflow/tensorflow/python/debug/debug_keras"
   OFFLINE_ANALYZER_BIN="$TEST_SRCDIR/org_tensorflow/tensorflow/python/debug/offline_analyzer"
 else
   DEBUG_FIBONACCI_BIN="${PYTHON_BIN_PATH} -m tensorflow.python.debug.examples.debug_fibonacci"
   DEBUG_ERRORS_BIN="${PYTHON_BIN_PATH} -m tensorflow.python.debug.examples.debug_errors"
   DEBUG_MNIST_BIN="${PYTHON_BIN_PATH} -m tensorflow.python.debug.examples.debug_mnist"
   DEBUG_TFLEARN_IRIS_BIN="${PYTHON_BIN_PATH} -m tensorflow.python.debug.examples.debug_tflearn_iris"
+  DEBUG_KERAS_BIN="${PYTHON_BIN_PATH} -m tensorflow.python.debug.examples.debug_keras"
   OFFLINE_ANALYZER_BIN="${PYTHON_BIN_PATH} -m tensorflow.python.debug.cli.offline_analyzer"
 fi
 
@@ -96,6 +98,11 @@ if [[ -d "${CUSTOM_DUMP_ROOT}" ]]; then
   exit 1
 fi
 
+# Test debugging of tf.keras.
+cat << EOF | "${DEBUG_KERAS_BIN}" --debug --ui_type=readline
+run -f has_inf_or_nan
+EOF
+
 # Test offline_analyzer.
 echo
 echo "Testing offline_analyzer"
diff --git a/tensorflow/python/debug/wrappers/framework.py b/tensorflow/python/debug/wrappers/framework.py
index c530204bbf6959f56a72c6e67add91f1e575f067..b9524ce649c7d6d888affacc22cfadd41dbe2e40 100644
--- a/tensorflow/python/debug/wrappers/framework.py
+++ b/tensorflow/python/debug/wrappers/framework.py
@@ -392,6 +392,9 @@ class BaseDebugWrapperSession(session.SessionInterface):
 
     self._default_session_context_manager = None
 
+    # A cache for callables created from CallableOptions.
+    self._cached_callables_from_options = dict()
+
   @property
   def graph(self):
     return self._sess.graph
@@ -414,7 +417,8 @@ class BaseDebugWrapperSession(session.SessionInterface):
           options=None,
           run_metadata=None,
           callable_runner=None,
-          callable_runner_args=None):
+          callable_runner_args=None,
+          callable_options=None):
     """Wrapper around Session.run() that inserts tensor watch options.
 
     Args:
@@ -424,7 +428,12 @@ class BaseDebugWrapperSession(session.SessionInterface):
       run_metadata: Same as the `run_metadata` arg to regular `Session.run()`.
       callable_runner: A `callable` returned by `Session.make_callable()`.
         If not `None`, `fetches` and `feed_dict` must both be `None`.
-      callable_runner_args: An optional list of arguments to `callable_runner`.
+        Mutually exclusive with `callable_options`.
+      callable_runner_args: An optional list of arguments to `callable_runner`
+        or for `callable_options`.
+      callable_options: An instance of `config_pb2.CallableOptions`, to be
+        used with `Session._make_callable_from_options()`. Mutually exclusive
+        with `callable_runner`.
 
     Returns:
       Simply forwards the output of the wrapped `Session.run()` call.
@@ -433,13 +442,17 @@ class BaseDebugWrapperSession(session.SessionInterface):
       ValueError: On invalid `OnRunStartAction` value. Or if `callable_runner`
         is not `None` and either or both of `fetches` and `feed_dict` is `None`.
     """
-    if not callable_runner:
+    if callable_runner and callable_options:
+      raise ValueError(
+          "callable_runner and callable_options are mutually exclusive, but "
+          "are both specified in this call to BaseDebugWrapperSession.run().")
+
+    if not (callable_runner or callable_options):
       self.increment_run_call_count()
-    else:
-      if fetches or feed_dict:
-        raise ValueError(
-            "callable_runner and fetches/feed_dict are mutually exclusive, but "
-            "are used simultaneously.")
+    elif callable_runner and (fetches or feed_dict):
+      raise ValueError(
+          "callable_runner and fetches/feed_dict are mutually exclusive, "
+          "but are used simultaneously.")
 
     empty_fetches = not nest.flatten(fetches)
     if empty_fetches:
@@ -449,6 +462,11 @@ class BaseDebugWrapperSession(session.SessionInterface):
     if self._is_disabled_thread() or empty_fetches:
       if callable_runner:
         return callable_runner(*callable_runner_args)
+      elif callable_options:
+        # pylint:disable=protected-access
+        return self._sess._make_callable_from_options(
+            callable_options)(*callable_runner_args)
+        # pylint:enable=protected-access
       else:
         return self._sess.run(fetches,
                               feed_dict=feed_dict,
@@ -464,19 +482,30 @@ class BaseDebugWrapperSession(session.SessionInterface):
 
     if run_start_resp.action == OnRunStartAction.DEBUG_RUN:
       # Decorate RunOption to fill in debugger tensor watch specifications.
-      decorated_run_options = options or config_pb2.RunOptions()
+      decorated_run_options = None
+      if callable_options:
+        callable_options_id = id(callable_options)
+        if callable_options_id not in self._cached_callables_from_options:
+          # Make a copy of callable_options to avoid mutating it.
+          new_callable_options = config_pb2.CallableOptions()
+          new_callable_options.CopyFrom(callable_options)
+          decorated_run_options = new_callable_options.run_options
+      else:
+        decorated_run_options = options or config_pb2.RunOptions()
+
       run_metadata = run_metadata or config_pb2.RunMetadata()
 
-      self._decorate_run_options_for_debug(
-          decorated_run_options,
-          run_start_resp.debug_urls,
-          debug_ops=run_start_resp.debug_ops,
-          node_name_regex_whitelist=run_start_resp.node_name_regex_whitelist,
-          op_type_regex_whitelist=run_start_resp.op_type_regex_whitelist,
-          tensor_dtype_regex_whitelist=(
-              run_start_resp.tensor_dtype_regex_whitelist),
-          tolerate_debug_op_creation_failures=(
-              run_start_resp.tolerate_debug_op_creation_failures))
+      if decorated_run_options:
+        self._decorate_run_options_for_debug(
+            decorated_run_options,
+            run_start_resp.debug_urls,
+            debug_ops=run_start_resp.debug_ops,
+            node_name_regex_whitelist=run_start_resp.node_name_regex_whitelist,
+            op_type_regex_whitelist=run_start_resp.op_type_regex_whitelist,
+            tensor_dtype_regex_whitelist=(
+                run_start_resp.tensor_dtype_regex_whitelist),
+            tolerate_debug_op_creation_failures=(
+                run_start_resp.tolerate_debug_op_creation_failures))
 
       # Invoke the run() method of the wrapped Session. Catch any TensorFlow
       # runtime errors.
@@ -486,6 +515,19 @@ class BaseDebugWrapperSession(session.SessionInterface):
           retvals = callable_runner(*callable_runner_args,
                                     options=decorated_run_options,
                                     run_metadata=run_metadata)
+        elif callable_options:
+          # pylint:disable=protected-access
+          if callable_options_id in self._cached_callables_from_options:
+            callable_object = self._cached_callables_from_options[
+                callable_options_id]
+          else:
+            callable_object = self._sess._make_callable_from_options(
+                new_callable_options)
+            self._cached_callables_from_options[
+                callable_options_id] = callable_object
+          # pylint:enable=protected-access
+          retvals = callable_object(
+              *callable_runner_args, run_metadata=run_metadata)
         else:
           retvals = self._sess.run(fetches,
                                    feed_dict=feed_dict,
@@ -590,7 +632,14 @@ class BaseDebugWrapperSession(session.SessionInterface):
                       run_metadata=kwargs.get("run_metadata", None),
                       callable_runner=runner,
                       callable_runner_args=runner_args)
+    return wrapped_runner
 
+  def _make_callable_from_options(self, callable_options):
+    def wrapped_runner(*feed_values, **kwargs):
+      return self.run(None,
+                      run_metadata=kwargs.get("run_metadata", None),
+                      callable_options=callable_options,
+                      callable_runner_args=feed_values)
     return wrapped_runner
 
   @property
diff --git a/tensorflow/python/debug/wrappers/grpc_wrapper.py b/tensorflow/python/debug/wrappers/grpc_wrapper.py
index 1f9c8fa5a96b4d6826fae0870608e0e737c7cd88..85944fa61118114cc73f9288f3f974f0a5a8a839 100644
--- a/tensorflow/python/debug/wrappers/grpc_wrapper.py
+++ b/tensorflow/python/debug/wrappers/grpc_wrapper.py
@@ -215,7 +215,8 @@ class TensorBoardDebugWrapperSession(GrpcDebugWrapperSession):
           options=None,
           run_metadata=None,
           callable_runner=None,
-          callable_runner_args=None):
+          callable_runner_args=None,
+          callable_options=None):
     if self._send_traceback_and_source_code:
       self._sent_graph_version = publish_traceback(
           self._grpc_debug_server_urls, self.graph, feed_dict, fetches,
@@ -226,4 +227,5 @@ class TensorBoardDebugWrapperSession(GrpcDebugWrapperSession):
         options=options,
         run_metadata=run_metadata,
         callable_runner=callable_runner,
-        callable_runner_args=callable_runner_args)
+        callable_runner_args=callable_runner_args,
+        callable_options=callable_options)
diff --git a/tensorflow/python/debug/wrappers/local_cli_wrapper.py b/tensorflow/python/debug/wrappers/local_cli_wrapper.py
index 4e551ab9955b79360afc544cff245efedc9c6b7c..668ffb57f10a69ce7e11e889fe613afbd618e823 100644
--- a/tensorflow/python/debug/wrappers/local_cli_wrapper.py
+++ b/tensorflow/python/debug/wrappers/local_cli_wrapper.py
@@ -596,7 +596,7 @@ class LocalCLIDebugWrapperSession(framework.BaseDebugWrapperSession):
       # Register tab completion for the filter names.
       curses_cli.register_tab_comp_context(["run", "r"],
                                            list(self._tensor_filters.keys()))
-    if self._feed_dict:
+    if self._feed_dict and hasattr(self._feed_dict, "keys"):
       # Register tab completion for feed_dict keys.
       feed_keys = [common.get_graph_element_name(key)
                    for key in self._feed_dict.keys()]
diff --git a/tensorflow/python/debug/wrappers/local_cli_wrapper_test.py b/tensorflow/python/debug/wrappers/local_cli_wrapper_test.py
index b06fa26a935b42709575f8e400e0bda951ffbbc7..05c9eaa4d27319ecf5e12fdeb0a973246c61704a 100644
--- a/tensorflow/python/debug/wrappers/local_cli_wrapper_test.py
+++ b/tensorflow/python/debug/wrappers/local_cli_wrapper_test.py
@@ -21,7 +21,10 @@ import os
 import shutil
 import tempfile
 
+import numpy as np
+
 from tensorflow.core.protobuf import config_pb2
+from tensorflow.core.protobuf import rewriter_config_pb2
 from tensorflow.python.client import session
 from tensorflow.python.debug.cli import cli_shared
 from tensorflow.python.debug.cli import debugger_cli_common
@@ -149,7 +152,13 @@ class LocalCLIDebugWrapperSessionTest(test_util.TensorFlowTestCase):
         dtypes.float32, shape=([5, 5]), name="sparse_placeholder")
     self.sparse_add = sparse_ops.sparse_add(self.sparse_ph, self.sparse_ph)
 
-    self.sess = session.Session()
+    rewriter_config = rewriter_config_pb2.RewriterConfig(
+        disable_model_pruning=True,
+        arithmetic_optimization=rewriter_config_pb2.RewriterConfig.OFF,
+        dependency_optimization=rewriter_config_pb2.RewriterConfig.OFF)
+    graph_options = config_pb2.GraphOptions(rewrite_options=rewriter_config)
+    config_proto = config_pb2.ConfigProto(graph_options=graph_options)
+    self.sess = session.Session(config=config_proto)
 
     # Initialize variable.
     self.sess.run(variables.global_variables_initializer())
@@ -393,6 +402,113 @@ class LocalCLIDebugWrapperSessionTest(test_util.TensorFlowTestCase):
     self.assertAllClose(42.0, tensor_runner(41.0, 1.0))
     self.assertEqual(1, len(wrapped_sess.observers["debug_dumps"]))
 
+  def testDebuggingMakeCallableFromOptionsWithZeroFeedWorks(self):
+    variable_1 = variables.Variable(
+        10.5, dtype=dtypes.float32, name="variable_1")
+    a = math_ops.add(variable_1, variable_1, "callable_a")
+    math_ops.add(a, a, "callable_b")
+    self.sess.run(variable_1.initializer)
+
+    wrapped_sess = LocalCLIDebuggerWrapperSessionForTest(
+        [["run"]] * 3, self.sess, dump_root=self._tmp_dir)
+    callable_options = config_pb2.CallableOptions()
+    callable_options.fetch.append("callable_b")
+    sess_callable = wrapped_sess._make_callable_from_options(callable_options)
+
+    for _ in range(2):
+      callable_output = sess_callable()
+      self.assertAllClose(np.array(42.0, dtype=np.float32), callable_output[0])
+
+    debug_dumps = wrapped_sess.observers["debug_dumps"]
+    self.assertEqual(2, len(debug_dumps))
+    for debug_dump in debug_dumps:
+      node_names = [datum.node_name for datum in debug_dump.dumped_tensor_data]
+      self.assertItemsEqual(
+          ["callable_a", "callable_b", "variable_1", "variable_1/read"],
+          node_names)
+
+  def testDebuggingMakeCallableFromOptionsWithOneFeedWorks(self):
+    ph1 = array_ops.placeholder(dtypes.float32, name="callable_ph1")
+    a = math_ops.add(ph1, ph1, "callable_a")
+    math_ops.add(a, a, "callable_b")
+
+    wrapped_sess = LocalCLIDebuggerWrapperSessionForTest(
+        [["run"]] * 3, self.sess, dump_root=self._tmp_dir)
+    callable_options = config_pb2.CallableOptions()
+    callable_options.feed.append("callable_ph1")
+    callable_options.fetch.append("callable_b")
+    sess_callable = wrapped_sess._make_callable_from_options(callable_options)
+
+    ph1_value = np.array([10.5, -10.5], dtype=np.float32)
+
+    for _ in range(2):
+      callable_output = sess_callable(ph1_value)
+      self.assertAllClose(
+          np.array([42.0, -42.0], dtype=np.float32), callable_output[0])
+
+    debug_dumps = wrapped_sess.observers["debug_dumps"]
+    self.assertEqual(2, len(debug_dumps))
+    for debug_dump in debug_dumps:
+      node_names = [datum.node_name for datum in debug_dump.dumped_tensor_data]
+      self.assertItemsEqual(["callable_a", "callable_b"], node_names)
+
+  def testDebuggingMakeCallableFromOptionsWithTwoFeedsWorks(self):
+    ph1 = array_ops.placeholder(dtypes.float32, name="callable_ph1")
+    ph2 = array_ops.placeholder(dtypes.float32, name="callable_ph2")
+    a = math_ops.add(ph1, ph2, "callable_a")
+    math_ops.add(a, a, "callable_b")
+
+    wrapped_sess = LocalCLIDebuggerWrapperSessionForTest(
+        [["run"]] * 3, self.sess, dump_root=self._tmp_dir)
+    callable_options = config_pb2.CallableOptions()
+    callable_options.feed.append("callable_ph1")
+    callable_options.feed.append("callable_ph2")
+    callable_options.fetch.append("callable_b")
+    sess_callable = wrapped_sess._make_callable_from_options(callable_options)
+
+    ph1_value = np.array(5.0, dtype=np.float32)
+    ph2_value = np.array(16.0, dtype=np.float32)
+
+    for _ in range(2):
+      callable_output = sess_callable(ph1_value, ph2_value)
+      self.assertAllClose(np.array(42.0, dtype=np.float32), callable_output[0])
+
+    debug_dumps = wrapped_sess.observers["debug_dumps"]
+    self.assertEqual(2, len(debug_dumps))
+    for debug_dump in debug_dumps:
+      node_names = [datum.node_name for datum in debug_dump.dumped_tensor_data]
+      self.assertItemsEqual(["callable_a", "callable_b"], node_names)
+
+  def testDebugMakeCallableFromOptionsWithCustomOptionsAndMetadataWorks(self):
+    variable_1 = variables.Variable(
+        10.5, dtype=dtypes.float32, name="variable_1")
+    a = math_ops.add(variable_1, variable_1, "callable_a")
+    math_ops.add(a, a, "callable_b")
+    self.sess.run(variable_1.initializer)
+
+    wrapped_sess = LocalCLIDebuggerWrapperSessionForTest(
+        [["run"], ["run"]], self.sess, dump_root=self._tmp_dir)
+    callable_options = config_pb2.CallableOptions()
+    callable_options.fetch.append("callable_b")
+    callable_options.run_options.trace_level = config_pb2.RunOptions.FULL_TRACE
+
+    sess_callable = wrapped_sess._make_callable_from_options(callable_options)
+
+    run_metadata = config_pb2.RunMetadata()
+    # Call the callable with a custom run_metadata.
+    callable_output = sess_callable(run_metadata=run_metadata)
+    # Verify that step_stats is populated in the custom run_metadata.
+    self.assertTrue(run_metadata.step_stats)
+    self.assertAllClose(np.array(42.0, dtype=np.float32), callable_output[0])
+
+    debug_dumps = wrapped_sess.observers["debug_dumps"]
+    self.assertEqual(1, len(debug_dumps))
+    debug_dump = debug_dumps[0]
+    node_names = [datum.node_name for datum in debug_dump.dumped_tensor_data]
+    self.assertItemsEqual(
+        ["callable_a", "callable_b", "variable_1", "variable_1/read"],
+        node_names)
+
   def testRuntimeErrorShouldBeCaught(self):
     wrapped_sess = LocalCLIDebuggerWrapperSessionForTest(
         [["run"], ["run"]], self.sess, dump_root=self._tmp_dir)
diff --git a/tensorflow/python/eager/backprop.py b/tensorflow/python/eager/backprop.py
index bd97b181ff7fa5a38ea8ab16e55b3ade7b599261..9e0bbce4a15ddf8b9aff848a94ae6e7773d18373 100644
--- a/tensorflow/python/eager/backprop.py
+++ b/tensorflow/python/eager/backprop.py
@@ -605,7 +605,9 @@ def _zeros(shape, dtype):
     # TODO(apassos): need to save enough information about variant tensors to do
     # a zeros
     return None
-  cache_key = shape, dtype, device
+  # pylint: disable=protected-access
+  cache_key = shape, dtype, device, context.context()._eager_context.mode
+  # pylint: enable=protected-access
   cached = _zeros_cache.get(cache_key)
   if cached is None:
     cached = _fast_fill(0, shape, dtype)
@@ -711,10 +713,15 @@ class GradientTape(object):
     if self._recording:
       self._pop_tape()
 
-  def _push_tape(self):
+  def _push_tape(self, existing_tape=False):
     if self._recording:
       raise ValueError("Tape is already recording.")
-    self._tape = tape.push_new_tape(persistent=self._persistent)
+    if existing_tape:
+      if self._tape is None:
+        raise ValueError("There is no existing tape.")
+      tape.push_tape(self._tape)
+    else:
+      self._tape = tape.push_new_tape(persistent=self._persistent)
     self._recording = True
 
   def _pop_tape(self):
@@ -762,7 +769,7 @@ class GradientTape(object):
     try:
       yield
     finally:
-      self._push_tape()
+      self._push_tape(existing_tape=True)
 
   def reset(self):
     """Clears all information stored in this tape.
diff --git a/tensorflow/python/eager/backprop_test.py b/tensorflow/python/eager/backprop_test.py
index e129c2756af1e30667c3e25e86d9530c6ea50481..bdda200ff641d03bae40924cdc738bbf7cd60c4e 100644
--- a/tensorflow/python/eager/backprop_test.py
+++ b/tensorflow/python/eager/backprop_test.py
@@ -223,11 +223,23 @@ class BackpropTest(test.TestCase):
 
   def testTapeStopRecording(self):
     with backprop.GradientTape() as t:
-      x = constant_op.constant(1.0)
+      x = resource_variable_ops.ResourceVariable(1.0)
       with t.stop_recording():
         y = x * x
     self.assertEqual(t.gradient(y, x), None)
 
+  def testTapeStopStartRecording(self):
+    with backprop.GradientTape(persistent=True) as t:
+      x = resource_variable_ops.ResourceVariable(1.0)
+      x2 = x * 2  # This should be differentiated through.
+      with t.stop_recording():
+        y = x2 * x2
+      z = x2 * x2
+    self.assertEqual(t.gradient(y, x2), None)
+
+    # If the x*2 was not differentiated through, this would be 2.0, not 4.0
+    self.assertEqual(t.gradient(z, x2).numpy(), 4.0)
+
   def testTapeReset(self):
     with backprop.GradientTape() as t:
       v = resource_variable_ops.ResourceVariable(1.0)
@@ -900,6 +912,33 @@ class BackpropTest(test.TestCase):
         'did you forget to return a value from fn?'):
       val_and_grads_fn(x, y)
 
+  def testZerosCacheDoesntLeakAcrossModes(self):
+    with ops.Graph().as_default():
+      t = random_ops.random_normal(shape=[100, 2])
+      x = random_ops.random_normal(shape=[100, 4])
+      dy = random_ops.random_normal(shape=[100, 4])
+      with backprop.GradientTape() as gradient_tape:
+        gradient_tape.watch(x)
+        x1, _ = array_ops.split(x, num_or_size_splits=2, axis=1)
+        y1 = x1 ** 2.
+        y = array_ops.concat([y1, t], axis=1)
+
+      dx = gradient_tape.gradient(y, x, output_gradients=dy)
+      with self.test_session() as sess:
+        sess.run(variables.global_variables_initializer())
+        sess.run(dx)
+
+    t = random_ops.random_normal(shape=[100, 2])
+    x = random_ops.random_normal(shape=[100, 4])
+    dy = random_ops.random_normal(shape=[100, 4])
+    with backprop.GradientTape() as gradient_tape:
+      gradient_tape.watch(x)
+      x1, _ = array_ops.split(x, num_or_size_splits=2, axis=1)
+      y1 = x1 ** 2.
+      y = array_ops.concat([y1, t], axis=1)
+
+    dx = gradient_tape.gradient(y, x, output_gradients=dy)
+
 
 if __name__ == '__main__':
   test.main()
diff --git a/tensorflow/python/eager/function.py b/tensorflow/python/eager/function.py
index fc68e945c0c3177dfdebeed843583492112e9b27..a6906f9efdb698205f2ef28245c9cdce5cd01537 100644
--- a/tensorflow/python/eager/function.py
+++ b/tensorflow/python/eager/function.py
@@ -21,6 +21,7 @@ from __future__ import print_function
 
 import collections
 import functools
+import threading
 
 import numpy as np
 
@@ -36,6 +37,7 @@ from tensorflow.python.framework import dtypes as dtypes_module
 from tensorflow.python.framework import ops
 from tensorflow.python.ops import array_ops
 from tensorflow.python.ops import control_flow_ops
+from tensorflow.python.ops import functional_ops
 from tensorflow.python.ops import gradients_impl
 from tensorflow.python.ops import resource_variable_ops
 from tensorflow.python.util import compat
@@ -47,8 +49,11 @@ def capture_value(tensor_map, value, dtype, name):
   """Capture a value from outside the function, to pass in as an extra arg."""
   captured_value = tensor_map.get(ops.tensor_id(value), None)
   if captured_value is None:
-    captured_value = graph_placeholder(
-        dtype=dtype or value.dtype, shape=value.shape, name=name)
+    # Note: setting ops.control_dependencies(None) ensures we always put
+    # capturing placeholders outside of any control flow context.
+    with ops.control_dependencies(None):
+      captured_value = graph_placeholder(
+          dtype=dtype or value.dtype, shape=value.shape, name=name)
     if captured_value.dtype == dtypes_module.resource:
       if ops._USE_C_SHAPES:  # pylint: disable=protected-access
         if isinstance(value, ops.EagerTensor):
@@ -133,7 +138,7 @@ class CapturingGraph(ops.Graph):
       inputs[i] = self.capture(inp)
     return super(CapturingGraph, self).create_op(
         op_type, inputs, dtypes, input_types, name, attrs, op_def,
-        compute_shapes, compute_device)
+        compute_device=compute_device)
 
 
 # pylint: disable=invalid-name
@@ -228,11 +233,20 @@ def _register(fn):
   context.context().add_function(fn)
 
 
+_xla_compile_attr = "_XlaCompile"
+
+
 # TODO(apassos) get rid of this by splitting framework.function._DefinedFunction
 # so it doesn't have the definition-generating logic and is just a container for
 # an already-defined function.
 class _EagerDefinedFunction(object):
-  """Function object with the interface of tf _DefinedFunction."""
+  """Callable with the interface of `framework.function._DefinedFunction.`
+
+  `_EagerDefinedFunction` encapsulates a function definition and its properties,
+  and it provides a method for calling the encapsulated function. Some Ops
+  take functions as attributes, which have type `func`; an instance of this
+  class may be provided as the value of these `func` attributes.
+  """
 
   def __init__(self, name, graph, operations, inputs, outputs, attrs):
     """Initializes an eager defined function.
@@ -263,6 +277,7 @@ class _EagerDefinedFunction(object):
       # It might be worth creating a convenient way to re-use status.
       pywrap_tensorflow.TF_FunctionSetAttrValueProto(
           fn, compat.as_str(name), serialized)
+    self._xla_compile = _xla_compile_attr in attrs
 
     # TODO(apassos) avoid creating a FunctionDef (specially to grab the
     # signature, but also in general it's nice not to depend on it.
@@ -274,12 +289,92 @@ class _EagerDefinedFunction(object):
     if context.executing_eagerly():
       _register(fn)
     self.definition = function_def
-    self.name = function_def.signature.name
+    self.name = compat.as_bytes(function_def.signature.name)
     self.signature = function_def.signature
+    self._num_outputs = len(self.signature.output_arg)
+    self._output_types = [o.type for o in self.signature.output_arg]
     self.grad_func_name = None
     self.python_grad_func = None
     self._c_func = c_api_util.ScopedTFFunction(fn)
     self._grad_func = None
+    self._graph = graph
+    self._stateful_ops = tuple(op for op in operations if op.op_def.is_stateful)
+
+  def add_to_graph(self, g):
+    # pylint: disable=protected-access
+    if self.name not in g._functions:
+      g._add_function(self)
+    for f in self._graph._functions.values():
+      if f.name not in g._functions:
+        g._add_function(f)
+    # pylint: enable=protected-access
+
+  @property
+  def stateful_ops(self):
+    return self._stateful_ops
+
+  def call(self, ctx, args, output_shapes):
+    """Calls this function with `args` as inputs.
+
+    Function execution respects device annotations only if the function won't
+    be compiled with xla.
+
+    Args:
+      ctx: a Context object
+      args: a list of arguments to supply this function with.
+      output_shapes: shapes to which outputs should be set; ignored when
+        executing eagerly.
+
+    Returns:
+      The outputs of the function call.
+    """
+
+    executing_eagerly = ctx.executing_eagerly()
+
+    xla_compile = self._xla_compile or (executing_eagerly and
+                                        ctx.device_spec.device_type == "TPU")
+
+    if xla_compile:
+      # XLA compilation relies upon a custom kernel creator to run functions.
+      signature = self.signature
+      if executing_eagerly:
+        outputs = execute.execute(
+            str(signature.name),
+            num_outputs=self._num_outputs,
+            inputs=args,
+            attrs=None,
+            ctx=ctx)
+      else:
+        g = ops.get_default_graph()
+        self.add_to_graph(g)
+        op = g.create_op(
+            signature.name,
+            [ops.internal_convert_to_tensor(x, ctx=ctx) for x in args],
+            tuple(dtypes_module.DType(x.type) for x in signature.output_arg),
+            op_def=signature,
+            name="FunctionCall",
+            compute_shapes=False)
+        outputs = op.outputs
+        if not outputs:
+          return op
+        outputs = [outputs] if isinstance(
+            outputs, (ops.Tensor, type(None))) else list(outputs)
+    else:
+      # TODO(akshayka): Either remove this if the FunctionLibraryRuntime
+      # creates `PartitionedCallOp` kernels by default, or remove the previous
+      # branch if a TPU kernel is registered for `PartitionedCall`.
+      outputs = functional_ops.partitioned_call(
+          args=args,
+          f=self,
+          tout=self._output_types,
+          executing_eagerly=executing_eagerly)
+
+    if executing_eagerly:
+      return outputs
+    else:
+      for i, shape in enumerate(output_shapes):
+        outputs[i].set_shape(shape)
+      return outputs
 
 
 def _map_sequence_obj_to_idx(sequence):
@@ -303,8 +398,12 @@ def _flatten(sequence):
   return outputs
 
 
+# TODO(akshayka): Perhaps rename to something more appropriate.
 class GraphModeFunction(object):
-  """Callable object representing a graph-mode function.
+  """Callable object encapsulating a function definition and its gradient.
+
+  `GraphModeFunction` is a callable that encapsulates a function definition and
+  is differentiable under `tf.GradientTape` objects.
   """
 
   def __init__(self,
@@ -371,7 +470,7 @@ class GraphModeFunction(object):
 
   def _construct_backprop_function(self):
     """Constructs the backprop function object for this function."""
-    with self._graph.as_default(), context.graph_mode():
+    with self._graph.as_default():
       c_known_ops = set()
       c_captured_tensors = set()
 
@@ -385,7 +484,7 @@ class GraphModeFunction(object):
           grad_ys=self._out_grad_placeholders)
       for op in self._graph.get_operations()[existing_op_len:]:
         if op.type in ["Variable", "VariableV2", "VarHandleOp"]:
-          raise ValueError("tfe.defun cannot capture variables created without "
+          raise ValueError("defun cannot capture variables created without "
                            "using tf.get_variable. Op: %s" % op)
         c_known_ops.add(op)
         for i in op.inputs:
@@ -427,35 +526,10 @@ class GraphModeFunction(object):
       The call output.
     """
     all_args = args + self._extra_inputs
-    signature = self._forward_fdef.signature
     ctx = context.context()
-    if ctx.executing_eagerly():
-      outputs = execute.execute(
-          str(signature.name),
-          num_outputs=len(signature.output_arg),
-          inputs=all_args,
-          attrs=None,
-          ctx=ctx)
-      if not outputs:
-        return None
-    else:
-      g = ops.get_default_graph()
-      g._add_function(self._forward_fdef)  # pylint: disable=protected-access
-      op = g.create_op(
-          signature.name,
-          [ops.internal_convert_to_tensor(x, ctx=ctx) for x in all_args],
-          tuple(dtypes_module.DType(x.type) for x in signature.output_arg),
-          op_def=signature,
-          name="FunctionCall",
-          compute_shapes=False)
-      outputs = op.outputs
-      if not outputs:
-        return op
-      outputs = [outputs] if isinstance(outputs, ops.Tensor) else list(outputs)
-
-      shapes = [shape for shape in self._output_shapes if shape is not None]
-      for i, shape in enumerate(shapes):
-        outputs[i].set_shape(shape)
+    outputs = self._forward_fdef.call(ctx, all_args, self._output_shapes)
+    if isinstance(outputs, ops.Operation) or outputs is None:
+      return outputs
 
     # `real_outputs` are the actual outputs of the inference graph function;
     # `side_outputs` are the intermediate Tensors that were added as outputs to
@@ -467,7 +541,7 @@ class GraphModeFunction(object):
       return self._backward_function(*(list(args) + side_outputs))  # pylint: disable=not-callable
 
     tape.record_operation(
-        signature.name,
+        self._forward_fdef.signature.name,
         real_outputs,
         (args + self._extra_inputs),
         backward_function)
@@ -509,13 +583,6 @@ class GraphModeFunction(object):
     """Returns the name of the function in Eager-compatible format."""
     return self._function_def.name.encode("utf-8")
 
-  def add_to_graph(self, g):
-    if self._function_def.name not in g._functions:  # pylint: disable=protected-access
-      g._add_function(self._function_def)  # pylint: disable=protected-access
-    for f in self._graph._functions.values():  # pylint: disable=protected-access
-      if f.name not in g._functions:  # pylint: disable=protected-access
-        g._add_function(f)  # pylint: disable=protected-access
-
   def __call__(self, *args):
     """Executes the passed function in eager mode."""
     for v in self._variables:
@@ -530,34 +597,9 @@ class GraphModeFunction(object):
       return self._backprop_call(tensor_inputs)
 
     ctx = context.context()
-    if ctx.executing_eagerly():
-      result = execute.execute(
-          str(self._func_name),
-          num_outputs=self._num_outputs,
-          inputs=tensor_inputs + self._extra_inputs,
-          attrs=None,
-          ctx=ctx)
-    else:
-      g = ops.get_default_graph()
-      self.add_to_graph(g)
-      signature = self._function_def.definition.signature
-      args = list(tensor_inputs) + self._extra_inputs
-      op = g.create_op(
-          signature.name,
-          [ops.internal_convert_to_tensor(x, ctx=ctx) for x in args],
-          tuple(dtypes_module.DType(x.type) for x in signature.output_arg),
-          op_def=signature,
-          name="FunctionCall",
-          compute_shapes=False)
-      result = op.outputs
-      if not result:
-        return op
-
-      shapes = [shape for shape in self._output_shapes if shape is not None]
-      for i, shape in enumerate(shapes):
-        result[i].set_shape(shape)
-
-    return self._build_call_outputs(result)
+    args = tensor_inputs + self._extra_inputs
+    outputs = self._function_def.call(ctx, args, self._output_shapes)
+    return self._build_call_outputs(outputs)
 
   def _build_call_outputs(self, result):
     """Maps the fdef output list to actual output structure.
@@ -568,7 +610,8 @@ class GraphModeFunction(object):
       The actual call output.
     """
     if self._python_func_outputs is None:
-      return None
+      return result
+
     # Use `nest.flatten` instead of `_flatten` in order to preserve any
     # IndexedSlices in `self._python_func_outputs`.
     outputs_list = nest.flatten(self._python_func_outputs)
@@ -614,55 +657,58 @@ def _deterministic_dict_values(kwds):
 def _trace_and_define_function(name, func, compiled, args, kwds):
   """Defines and returns graph-mode version of func."""
   graph_key = ops.get_default_graph()._graph_key  # pylint: disable=protected-access
-  with context.graph_mode():
-    captures = {}
-    tmp_graph = CapturingGraph(captures)
-    # Inherit the graph key, since this is used for matching variables in
-    # optimizers.
-    tmp_graph._graph_key = graph_key  # pylint: disable=protected-access
-    # Copy the graph collections to ensure summaries and other things work. This
-    # lets the function access (but not mutate) collections of the containing
-    # graph, such as the global step and the summary writer collections.
-    curr_graph = ops.get_default_graph()
-    for collection in curr_graph.collections:
-      tmp_graph.get_collection_ref(collection)[:] = curr_graph.get_collection(
-          collection)
-    with tmp_graph.as_default(), AutomaticControlDependencies() as a:
-      func_args = _get_defun_inputs(args)
-      func_kwds = _get_defun_inputs(kwds)
-
-      def convert(x):
-        if x is None:
-          return None
-        x = ops.convert_to_tensor_or_indexed_slices(x)
-        x = a.mark_as_return(x)
-        return x
+  captures = {}
+  tmp_graph = CapturingGraph(captures)
+  # Inherit the graph key, since this is used for matching variables in
+  # optimizers.
+  tmp_graph._graph_key = graph_key  # pylint: disable=protected-access
+  # Copy the graph collections to ensure summaries and other things work. This
+  # lets the function access (but not mutate) collections of the containing
+  # graph, such as the global step and the summary writer collections.
+  curr_graph = ops.get_default_graph()
+  for collection in curr_graph.collections:
+    tmp_graph.get_collection_ref(collection)[:] = curr_graph.get_collection(
+        collection)
+  if context.executing_eagerly():
+    tmp_graph.seed = context.global_seed()
+  else:
+    tmp_graph.seed = curr_graph.seed
+  with tmp_graph.as_default(), AutomaticControlDependencies() as a:
+    func_args = _get_defun_inputs(args)
+    func_kwds = _get_defun_inputs(kwds)
 
-      this_tape = tape.push_new_tape()
-      try:
-        func_outputs = func(*func_args, **func_kwds)
-        func_outputs = nest.map_structure(convert, func_outputs)
-      finally:
-        tape.pop_tape(this_tape)
-      variables = this_tape.watched_variables()
-
-      # Returning a closed-over tensor as an output does not trigger a
-      # call to convert_to_tensor, so we manually capture all such tensors.
-      outputs_list = _flatten(func_outputs)
-      func_def_outputs = [
-          tmp_graph.capture(x) for x in outputs_list
-          if x is not None
-      ]
-
-      ids = list(sorted(captures.keys()))
-      if ids:
-        extra_inputs, extra_placeholders = zip(* [captures[x] for x in ids])
-      else:
-        extra_inputs = []
-        extra_placeholders = []
-      output_shapes = tuple(
-          x.shape if isinstance(x, ops.Tensor) else None
-          for x in outputs_list)
+    def convert(x):
+      if x is None:
+        return None
+      x = ops.convert_to_tensor_or_indexed_slices(x)
+      x = a.mark_as_return(x)
+      return x
+
+    this_tape = tape.push_new_tape()
+    try:
+      func_outputs = func(*func_args, **func_kwds)
+      func_outputs = nest.map_structure(convert, func_outputs)
+    finally:
+      tape.pop_tape(this_tape)
+    variables = this_tape.watched_variables()
+
+    # Returning a closed-over tensor as an output does not trigger a
+    # call to convert_to_tensor, so we manually capture all such tensors.
+    outputs_list = _flatten(func_outputs)
+    func_def_outputs = [
+        tmp_graph.capture(x) for x in outputs_list
+        if x is not None
+    ]
+
+    ids = list(sorted(captures.keys()))
+    if ids:
+      extra_inputs, extra_placeholders = zip(* [captures[x] for x in ids])
+    else:
+      extra_inputs = []
+      extra_placeholders = []
+    output_shapes = tuple(
+        x.shape if isinstance(x, ops.Tensor) else None
+        for x in func_def_outputs)
 
   func_kwds_values = _deterministic_dict_values(func_kwds)
   flat_inputs = [
@@ -683,7 +729,7 @@ def _trace_and_define_function(name, func, compiled, args, kwds):
 
   attrs = {}
   if compiled:
-    attrs["_XlaCompile"] = attr_value_pb2.AttrValue(b=True)
+    attrs[_xla_compile_attr] = attr_value_pb2.AttrValue(b=True)
 
   return GraphModeFunction(
       fname, all_inputs, extra_inputs, tmp_graph, operations, func_def_outputs,
@@ -728,6 +774,11 @@ class _PolymorphicFunction(object):
 
   See the documentation for `defun` for more information on the semantics of
   defined functions.
+
+  _PolymorphicFunction class is thread-compatible meaning that minimal
+  usage of defuns (defining and calling) is thread-safe, but if users call other
+  methods or invoke the base `python_function` themselves, external
+  synchronization is necessary.
   """
 
   def __init__(self, python_function, name, compiled=False):
@@ -745,6 +796,8 @@ class _PolymorphicFunction(object):
     self._arguments_to_functions = {}
     self._variables = []
 
+    self._lock = threading.Lock()
+
   def __get__(self, instance, owner):
     """Makes it possible to defun instance methods."""
     del owner
@@ -779,22 +832,30 @@ class _PolymorphicFunction(object):
     kwd_values = _deterministic_dict_values(kwds)
     inputs = args + kwd_values
     signature = tuple(_cache_key(x) for x in inputs)
-
-    if signature not in self._arguments_to_functions:
-      graph_function = _trace_and_define_function(
-          self._name, self._python_function, self._compiled, args, kwds)
-      self._arguments_to_functions[signature] = graph_function
-      self._variables.extend(
-          [v for v in graph_function.variables if v not in self._variables])
-      return graph_function, inputs
-    else:
-      return self._arguments_to_functions[signature], inputs
+    # The graph, or whether we're executing eagerly, should be a part of the
+    # signature so we don't improperly capture tensors such as variables.
+    signature += tuple([context.executing_eagerly() or ops.get_default_graph()])
+
+    with self._lock:
+      if signature not in self._arguments_to_functions:
+        graph_function = _trace_and_define_function(
+            self._name, self._python_function, self._compiled, args, kwds)
+        self._arguments_to_functions[signature] = graph_function
+        self._variables.extend(
+            [v for v in graph_function.variables if v not in self._variables])
+        return graph_function, inputs
+      else:
+        return self._arguments_to_functions[signature], inputs
 
   def __call__(self, *args, **kwds):
     """Calls a graph function specialized for this input signature."""
     graph_function, inputs = self._maybe_define_function(*args, **kwds)
     return graph_function(*inputs)
 
+  def call_python_function(self, *args, **kwargs):
+    """Directly calls the wrapped python function."""
+    return self._python_function(*args, **kwargs)
+
   @property
   def variables(self):
     """Returns a list of variables used in any of the defined functions."""
@@ -832,6 +893,11 @@ def defun(func=None, compiled=False):
   be hashable Python objects or lists thereof. Additionally, it must return zero
   or more @{tf.Tensor} objects.
 
+  Executing a graph generated by `defun` respects device annotations (i.e.,
+  all `with tf.device` directives present in a Python function will also be
+  present in its corresponding graph), but it is not yet possible to execute the
+  generated graphs across multiple machines.
+
   _Example Usage_
 
   ```python
@@ -1242,7 +1308,7 @@ class AutomaticControlDependencies(object):
     # Ensures the merge always runs
     ops_which_must_run.add(new_merge[0].op)
     if inp in last_op_using_resource_tensor:
-      # Ensures the switch exectutes after the previous op using the resource.
+      # Ensures the switch executes after the previous op using the resource.
       switch_op._add_control_input(last_op_using_resource_tensor[inp])  # pylint: disable=protected-access
     # Ensure the next op outside the cond happens after the merge.
     last_op_using_resource_tensor[inp] = new_merge[0].op
diff --git a/tensorflow/python/eager/function_test.py b/tensorflow/python/eager/function_test.py
index a5df3ef530d1c2ac31ee44545d0d8ada2b108984..13c4ee7f15514b347bbead9c0ef9f2e19dd289ba 100644
--- a/tensorflow/python/eager/function_test.py
+++ b/tensorflow/python/eager/function_test.py
@@ -19,15 +19,18 @@ from __future__ import print_function
 
 import collections
 
+from tensorflow.core.protobuf import config_pb2
+from tensorflow.python.data.ops import iterator_ops
 from tensorflow.python.eager import backprop
 from tensorflow.python.eager import context
 from tensorflow.python.eager import function
 from tensorflow.python.eager import tape
-from tensorflow.python.eager import test
 from tensorflow.python.framework import constant_op
 from tensorflow.python.framework import dtypes
+from tensorflow.python.framework import errors
 from tensorflow.python.framework import function as tf_function
 from tensorflow.python.framework import ops
+from tensorflow.python.framework import random_seed
 from tensorflow.python.framework import tensor_shape
 from tensorflow.python.framework import test_util
 from tensorflow.python.layers import convolutional
@@ -37,10 +40,14 @@ from tensorflow.python.ops import control_flow_ops
 from tensorflow.python.ops import gradients_impl
 from tensorflow.python.ops import init_ops
 from tensorflow.python.ops import math_ops
+from tensorflow.python.ops import random_ops
 from tensorflow.python.ops import resource_variable_ops
 from tensorflow.python.ops import variable_scope
 from tensorflow.python.ops import variables
-from tensorflow.python.training import gradient_descent
+from tensorflow.python.platform import test
+from tensorflow.python.training import momentum
+from tensorflow.python.training import training_ops
+from tensorflow.python.util import compat
 
 
 @test_util.with_c_shapes
@@ -103,6 +110,19 @@ class FunctionTest(test.TestCase):
       grads, = gradients_impl.gradients(node, v)
       v.initializer.run()
       self.assertAllEqual(grads.eval(), 2.0)
+      self.assertEqual(grads.shape, v.shape)
+
+  def testGraphEagerIsolation(self):
+
+    @function.defun
+    def f():
+      v = resource_variable_ops.ResourceVariable(1.0)
+      return v.read_value()
+
+    self.assertAllEqual(f(), 1.0)
+
+    with ops.Graph().as_default():
+      self.assertEqual(f().shape, ())
 
   def testBasicDefunOpGraphMode(self):
     matmul = function.defun(math_ops.matmul)
@@ -118,6 +138,18 @@ class FunctionTest(test.TestCase):
     out = sq_op(t)
     self.assertAllEqual(out, math_ops.matmul(t, t).numpy())
 
+  def testRandomSeed(self):
+
+    @function.defun
+    def f():
+      return random_ops.random_normal(())
+
+    random_seed.set_random_seed(1)
+    x = f()
+    self.assertNotEqual(x, f())
+    random_seed.set_random_seed(1)
+    self.assertAllEqual(f(), x)
+
   def testNestedInputsDefunOpGraphMode(self):
     matmul = function.defun(math_ops.matmul)
 
@@ -180,6 +212,15 @@ class FunctionTest(test.TestCase):
     self.assertEqual(fn_op.output_shapes, None)
     self.assertAllEqual(fn_op(x, x), None)
 
+  def testDefunCapturedInt32(self):
+    x = constant_op.constant(1, dtype=dtypes.int32)
+
+    @function.defun
+    def add_int32s():
+      return x + x
+
+    self.assertEqual(2, int(add_int32s()))
+
   def testDefunReadVariable(self):
     v = resource_variable_ops.ResourceVariable(1.0)
 
@@ -191,13 +232,14 @@ class FunctionTest(test.TestCase):
 
   def testDefunAssignAddVariable(self):
     v = resource_variable_ops.ResourceVariable(1.0)
+    x = constant_op.constant(2.0)
 
     @function.defun
-    def f():
-      v.assign_add(2.0)
+    def test_assign_add():
+      v.assign_add(x)
       return v.read_value()
 
-    self.assertEqual(3.0, float(f()))
+    self.assertEqual(3.0, float(test_assign_add()))
 
   def testDefunShapeInferenceWithCapturedResourceVariable(self):
     v = resource_variable_ops.ResourceVariable([[1, 2], [3, 4]])
@@ -210,6 +252,21 @@ class FunctionTest(test.TestCase):
     compiled = function.defun(f)
     compiled()
 
+  def testVariableInLoopInFunction(self):
+
+    @function.defun
+    def test_function():
+
+      def loop_test(_):
+        return False
+
+      def loop_body(_):
+        return variable_scope.get_variable('a', shape=())
+
+      return control_flow_ops.while_loop(loop_test, loop_body, [0.0])
+
+    self.assertEqual(test_function().shape, [])
+
   def testDefunShapeInferenceWithCapturedResourceVariableInGraphMode(self):
     with context.graph_mode():
       v = resource_variable_ops.ResourceVariable([[1, 2], [3, 4]])
@@ -412,24 +469,33 @@ class FunctionTest(test.TestCase):
 
     self.assertAllEqual(f(constant_op.constant(1.0)), 2.0)
 
-  def testGradientOfGatherWithDefun(self):
+  def testGatherResourceWithDefun(self):
     with ops.device('cpu:0'):
       v = resource_variable_ops.ResourceVariable([0.0, 1.0, 2.0])
 
-      def sum_gather():
-        return math_ops.reduce_sum(array_ops.gather(v, [1, 2]))
+    def sum_gather():
+      return math_ops.reduce_sum(array_ops.gather(v, [1, 2]))
+
+    defined = function.defun(sum_gather)
+    self.assertAllEqual(sum_gather(), defined())
+
+  def testGradientOfGatherWithDefun(self):
+    v = resource_variable_ops.ResourceVariable([0.0, 1.0, 2.0])
+
+    def sum_gather():
+      return math_ops.reduce_sum(array_ops.gather(v, [1, 2]))
 
-      grad_fn = backprop.implicit_grad(sum_gather)
-      gradient = grad_fn()
-      defun_grad_fn = backprop.implicit_grad(function.defun(sum_gather))
-      defun_gradient = defun_grad_fn()
-      self.assertEqual(len(gradient), len(defun_gradient))
+    grad_fn = backprop.implicit_grad(sum_gather)
+    gradient = grad_fn()
+    defun_grad_fn = backprop.implicit_grad(function.defun(sum_gather))
+    defun_gradient = defun_grad_fn()
+    self.assertEqual(len(gradient), len(defun_gradient))
 
-      gradient = gradient[0][0]
-      defun_gradient = defun_gradient[0][0]
-      self.assertAllEqual(gradient.values, defun_gradient.values)
-      self.assertAllEqual(gradient.indices, defun_gradient.indices)
-      self.assertAllEqual(gradient.dense_shape, defun_gradient.dense_shape)
+    gradient = gradient[0][0]
+    defun_gradient = defun_gradient[0][0]
+    self.assertAllEqual(gradient.values, defun_gradient.values)
+    self.assertAllEqual(gradient.indices, defun_gradient.indices)
+    self.assertAllEqual(gradient.dense_shape, defun_gradient.dense_shape)
 
   def testReturningIndexedSlicesWithDefun(self):
 
@@ -493,6 +559,66 @@ class FunctionTest(test.TestCase):
     y = f(x, x).cpu()
     self.assertAllEqual(y, [2.])
 
+  @test_util.run_in_graph_and_eager_modes
+  def testFunctionWithResourcesOnDifferentDevices(self):
+    if not context.context().num_gpus():
+      self.skipTest('No GPUs found.')
+
+    with ops.device('/cpu:0'):
+      v_cpu = resource_variable_ops.ResourceVariable([0.0, 1.0, 2.0])
+
+    with ops.device('/gpu:0'):
+      v_gpu = resource_variable_ops.ResourceVariable([0.0, 1.0, 2.0])
+
+    def sum_gather():
+      cpu_result = math_ops.reduce_sum(array_ops.gather(v_cpu, [1, 2]))
+      gpu_result = math_ops.reduce_sum(array_ops.gather(v_gpu, [1, 2]))
+      return cpu_result, gpu_result
+
+    defined = function.defun(sum_gather)
+    if not context.executing_eagerly():
+      self.evaluate(variables.global_variables_initializer())
+    expected = self.evaluate(sum_gather())
+    self.assertAllEqual(expected, self.evaluate(defined()))
+
+  @test_util.run_in_graph_and_eager_modes
+  def testOpInFunctionWithConflictingResourceInputs(self):
+    if not context.context().num_gpus():
+      self.skipTest('No GPUs found.')
+
+    with ops.device('/cpu:0'):
+      v_cpu = resource_variable_ops.ResourceVariable(
+          [0.0, 1.0, 2.0], name='cpu')
+      v_also_cpu = resource_variable_ops.ResourceVariable(
+          [0.0, 1.0, 2.0], name='also_cpu')
+
+    with ops.device('/gpu:0'):
+      v_gpu = resource_variable_ops.ResourceVariable(
+          [0.0, 1.0, 2.0], name='gpu')
+
+    @function.defun
+    def resource_apply_adam():
+      training_ops.resource_apply_adam(
+          v_cpu.handle,
+          v_gpu.handle,
+          v_also_cpu.handle,
+          1.0,  # beta1_power
+          1.0,  # beta2_power
+          1.0,  # learning_rate
+          1.0,  # beta1
+          1.0,  # beta2
+          1.0,  # epsilon,
+          [1.0, 1.0, 1.0],  # grad
+          False)  # use_locking
+      return None
+
+    with self.assertRaisesRegexp(
+        errors.InvalidArgumentError, 'Could not colocate node with its '
+        'resource and reference inputs.*'):
+      if not context.executing_eagerly():
+        self.evaluate(variables.global_variables_initializer())
+      self.evaluate(resource_apply_adam())
+
   def testFunctionHandlesInputsOnDifferentDevices(self):
     if not context.context().num_gpus():
       self.skipTest('No GPUs found')
@@ -542,17 +668,17 @@ class FunctionTest(test.TestCase):
 
   def testNestedDifferentiableFunction(self):
     @function.defun
-    def foo(a, b):
+    def inner_fn(a, b):
       return a * math_ops.add(a, b)
 
     @function.defun
-    def bar(x):
-      return foo(x, 1.0)
+    def outer_fn(x):
+      return inner_fn(x, 1.0)
 
     x = constant_op.constant(5.0)
     with backprop.GradientTape() as tp:
       tp.watch(x)
-      result = bar(x)
+      result = outer_fn(x)
     grad = tp.gradient(result, x)
 
     self.assertAllEqual(grad, 2 * 5.0 + 1.0)
@@ -602,15 +728,15 @@ class FunctionTest(test.TestCase):
     self.assertAllEqual(3, add_one(constant_op.constant(2)))
 
   def testVariableCaptureInNestedFunctions(self):
-    v = resource_variable_ops.ResourceVariable(1)
+    v = resource_variable_ops.ResourceVariable(1, dtype=dtypes.int32)
 
     @function.defun
-    def read():
+    def inner_read():
       return v.read_value()
 
     @function.defun
     def outer():
-      return read()
+      return inner_read()
 
     self.assertEqual(1, int(outer()))
 
@@ -701,6 +827,27 @@ class FunctionTest(test.TestCase):
     y = model(x)
     self.assertAllEqual([[[[4.0]]]], y.numpy())
 
+  @test_util.run_in_graph_and_eager_modes(
+      config=config_pb2.ConfigProto(device_count={'CPU': 3}))
+  def testDeviceAnnotationsRespected(self):
+    @function.defun
+    def multi_device_fn():
+      with ops.device('/cpu:0'):
+        s1 = iterator_ops.Iterator.from_structure(
+            (dtypes.float32,)).string_handle()
+      with ops.device('/cpu:1'):
+        s2 = iterator_ops.Iterator.from_structure(
+            (dtypes.float32,)).string_handle()
+      with ops.device('/cpu:2'):
+        s3 = iterator_ops.Iterator.from_structure(
+            (dtypes.float32,)).string_handle()
+      return s1, s2, s3
+
+    outputs = multi_device_fn()
+    self.assertTrue(compat.as_bytes('CPU:0') in self.evaluate(outputs[0]))
+    self.assertTrue(compat.as_bytes('CPU:1') in self.evaluate(outputs[1]))
+    self.assertTrue(compat.as_bytes('CPU:2') in self.evaluate(outputs[2]))
+
   def testVariablesAreTracked(self):
     v = resource_variable_ops.ResourceVariable(1.0)
 
@@ -801,6 +948,25 @@ class FunctionTest(test.TestCase):
     out = foo.two(t)
     self.assertEqual(float(out), 1.0)
 
+  def testPythonCallWithSideEffects(self):
+    state = []
+
+    @function.defun
+    def side_effecting_function():
+      state.append(0)
+
+    side_effecting_function()
+    self.assertAllEqual(state, [0])
+
+    # The second invocation should call the graph function, which shouldn't
+    # trigger the list append.
+    side_effecting_function()
+    self.assertAllEqual(state, [0])
+
+    # Whereas calling the python function directly should create a side-effect.
+    side_effecting_function.call_python_function()
+    self.assertAllEqual(state, [0, 0])
+
 
 @test_util.with_c_shapes
 class AutomaticControlDependenciesTest(test.TestCase):
@@ -988,7 +1154,7 @@ class AutomaticControlDependenciesTest(test.TestCase):
     def loss(v):
       return v**2
 
-    optimizer = gradient_descent.GradientDescentOptimizer(learning_rate=1.0)
+    optimizer = momentum.MomentumOptimizer(learning_rate=1.0, momentum=1.0)
 
     @function.defun
     def train():
@@ -1005,7 +1171,7 @@ class AutomaticControlDependenciesTest(test.TestCase):
     def loss():
       return v**2
 
-    optimizer = gradient_descent.GradientDescentOptimizer(learning_rate=1.0)
+    optimizer = momentum.MomentumOptimizer(learning_rate=1.0, momentum=1.0)
 
     @function.defun
     def train():
@@ -1017,4 +1183,6 @@ class AutomaticControlDependenciesTest(test.TestCase):
 
 
 if __name__ == '__main__':
+  ops.enable_eager_execution(
+      config=config_pb2.ConfigProto(device_count={'CPU': 3}))
   test.main()
diff --git a/tensorflow/python/eager/graph_callable.py b/tensorflow/python/eager/graph_callable.py
index 760a1485523798c6587e95804488a14b42a69bc0..2c6f04d8ad3cd121e52e56a388d5ff7951da5e33 100644
--- a/tensorflow/python/eager/graph_callable.py
+++ b/tensorflow/python/eager/graph_callable.py
@@ -110,13 +110,25 @@ class _VariableCapturingScope(object):
     """
     # TODO(apassos) ignoring the regularizer and partitioner here; figure out
     # how to deal with these.
-    def _custom_getter(getter=None, name=None, shape=None, dtype=dtypes.float32,  # pylint: disable=missing-docstring
-                       initializer=None, regularizer=None, reuse=None,
-                       trainable=True, collections=None, caching_device=None,  # pylint: disable=redefined-outer-name
-                       partitioner=None, validate_shape=True,
-                       use_resource=None):
+    def _custom_getter(  # pylint: disable=missing-docstring
+        getter=None,
+        name=None,
+        shape=None,
+        dtype=dtypes.float32,
+        initializer=None,
+        regularizer=None,
+        reuse=None,
+        trainable=None,
+        collections=None,
+        caching_device=None,  # pylint: disable=redefined-outer-name
+        partitioner=None,
+        validate_shape=True,
+        use_resource=None,
+        aggregation=variable_scope.VariableAggregation.NONE,
+        synchronization=variable_scope.VariableSynchronization.AUTO):
       del getter, regularizer, partitioner, validate_shape, use_resource, dtype
-      del collections, initializer, trainable, reuse, caching_device, shape,
+      del collections, initializer, trainable, reuse, caching_device, shape
+      del aggregation, synchronization
       assert name in self.variables
       v = self.variables[name]
       return v.variable
@@ -136,13 +148,24 @@ class _VariableCapturingScope(object):
     """
     # TODO(apassos) ignoring the regularizer and partitioner here; figure out
     # how to deal with these.
-    def _custom_getter(getter=None, name=None, shape=None, dtype=dtypes.float32,  # pylint: disable=missing-docstring
-                       initializer=None, regularizer=None, reuse=None,
-                       trainable=True, collections=None, caching_device=None,  # pylint: disable=redefined-outer-name
-                       partitioner=None, validate_shape=True,
-                       use_resource=None):
+    def _custom_getter(  # pylint: disable=missing-docstring
+        getter=None,
+        name=None,
+        shape=None,
+        dtype=dtypes.float32,
+        initializer=None,
+        regularizer=None,
+        reuse=None,
+        trainable=None,
+        collections=None,
+        caching_device=None,  # pylint: disable=redefined-outer-name
+        partitioner=None,
+        validate_shape=True,
+        use_resource=None,
+        aggregation=variable_scope.VariableAggregation.NONE,
+        synchronization=variable_scope.VariableSynchronization.AUTO):
       del getter, regularizer, collections, caching_device, partitioner
-      del use_resource, validate_shape
+      del use_resource, validate_shape, aggregation, synchronization
       if name in self.tf_variables:
         if reuse:
           return self.tf_variables[name].initialized_value()
diff --git a/tensorflow/python/eager/pywrap_tfe_src.cc b/tensorflow/python/eager/pywrap_tfe_src.cc
index b797a3f82d14fd856388c7368c96b9cd5bdd8c20..ec7e2371e96dc5e1877c510f72c5c894f0f8a43a 100644
--- a/tensorflow/python/eager/pywrap_tfe_src.cc
+++ b/tensorflow/python/eager/pywrap_tfe_src.cc
@@ -948,7 +948,7 @@ class GradientTape
         : id(id), variable(variable) {}
   };
   struct CompareById {
-    bool operator()(const IdAndVariable& lhs, const IdAndVariable& rhs) {
+    bool operator()(const IdAndVariable& lhs, const IdAndVariable& rhs) const {
       return lhs.id < rhs.id;
     }
   };
@@ -1898,14 +1898,39 @@ PyObject* RecordGradient(PyObject* op_name, PyObject* inputs, PyObject* attrs,
 
 void MaybeWatchVariable(PyObject* input) {
   DCHECK(CheckResourceVariable(input));
-  DCHECK(PyObject_HasAttrString(input, "trainable"));
+  DCHECK(PyObject_HasAttrString(input, "_trainable"));
 
   tensorflow::Safe_PyObjectPtr trainable(
-      PyObject_GetAttrString(input, "trainable"));
+      PyObject_GetAttrString(input, "_trainable"));
   if (trainable.get() == Py_False) return;
   TFE_Py_TapeSetWatchVariable(input);
 }
 
+bool CastTensor(const FastPathOpExecInfo& op_exec_info,
+                const TF_DataType& desired_dtype,
+                tensorflow::Safe_TFE_TensorHandlePtr* handle,
+                TF_Status* status) {
+  TF_DataType input_dtype = TFE_TensorHandleDataType(handle->get());
+  TF_DataType output_dtype = input_dtype;
+
+  if (desired_dtype >= 0 && desired_dtype != input_dtype) {
+    *handle = tensorflow::make_safe(
+        tensorflow::EagerCast(op_exec_info.ctx, handle->get(), input_dtype,
+                              static_cast(desired_dtype), status));
+    if (!status->status.ok()) return false;
+    output_dtype = desired_dtype;
+  }
+
+  if (output_dtype != TF_INT32) {
+    // Note that this is a shallow copy and will share the underlying buffer
+    // if copying to the same device.
+    *handle = tensorflow::make_safe(TFE_TensorHandleCopyToDevice(
+        handle->get(), op_exec_info.ctx, op_exec_info.device_name, status));
+    if (!status->status.ok()) return false;
+  }
+  return true;
+}
+
 bool ReadVariableOp(const FastPathOpExecInfo& parent_op_exec_info,
                     PyObject* input, tensorflow::Safe_PyObjectPtr* output,
                     TF_Status* status) {
@@ -1938,9 +1963,31 @@ bool ReadVariableOp(const FastPathOpExecInfo& parent_op_exec_info,
   TFE_Execute(op, &output_handle, &num_retvals, status);
   if (MaybeRaiseExceptionFromTFStatus(status, nullptr)) return false;
 
-  // Always create the py object (and correctly DECREF it) from the returned
-  // value, else the data will leak.
-  output->reset(EagerTensorFromHandle(output_handle));
+  if (!PyObject_HasAttrString(input, "_read_dtype")) {
+    // Always create the py object (and correctly DECREF it) from the returned
+    // value, else the data will leak.
+    output->reset(EagerTensorFromHandle(output_handle));
+  } else {
+    // This is a _MixedPrecisionVariable which potentially does casting when
+    // being read.
+    tensorflow::Safe_PyObjectPtr read_dtype(
+        PyObject_GetAttrString(input, "_read_dtype"));
+    int desired_dtype = -1;
+    if (!ParseTypeValue("_read_dtype", read_dtype.get(), status,
+                        &desired_dtype)) {
+      return false;
+    }
+
+    auto safe_output_handle = tensorflow::make_safe(output_handle);
+    // Retires output_handle in the future.
+    output_handle = nullptr;
+    if (!CastTensor(parent_op_exec_info,
+                    static_cast(desired_dtype),
+                    &safe_output_handle, status)) {
+      return false;
+    }
+    output->reset(EagerTensorFromHandle(safe_output_handle.release()));
+  }
 
   // TODO(nareshmodi): Should we run post exec callbacks here?
   if (parent_op_exec_info.run_gradient_callback) {
@@ -2010,27 +2057,13 @@ bool ConvertToTensor(
     }
   }
 
-  TF_DataType handle_dtype = TFE_TensorHandleDataType(handle.get());
-  if (desired_dtype >= 0 && desired_dtype != handle_dtype) {
-    handle = tensorflow::make_safe(
-        tensorflow::EagerCast(op_exec_info.ctx, handle.get(), handle_dtype,
-                              static_cast(desired_dtype), status));
-    if (!status->status.ok()) return false;
-
-    handle_dtype = TFE_TensorHandleDataType(handle.get());
-  }
-
-  if (handle_dtype != TF_INT32) {
-    // Note that this is a shallow copy and will share the underlying buffer
-    // if copying to the same device.
-    handle = tensorflow::make_safe(TFE_TensorHandleCopyToDevice(
-        handle.get(), op_exec_info.ctx, op_exec_info.device_name, status));
-    if (!status->status.ok()) return false;
+  if (!CastTensor(op_exec_info, static_cast(desired_dtype),
+                  &handle, status)) {
+    return false;
   }
-
+  TF_DataType output_dtype = TFE_TensorHandleDataType(handle.get());
   output_handle->reset(EagerTensorFromHandle(handle.release()));
-
-  dtype_setter(handle_dtype);
+  dtype_setter(output_dtype);
 
   return true;
 }
diff --git a/tensorflow/python/eager/pywrap_tfe_test.py b/tensorflow/python/eager/pywrap_tfe_test.py
index faaae40b3f1ef02984a7a75c23ae4acae65ac335..fd8ab695b8fbb732bb853cd4affadf98d4861cc2 100644
--- a/tensorflow/python/eager/pywrap_tfe_test.py
+++ b/tensorflow/python/eager/pywrap_tfe_test.py
@@ -23,6 +23,7 @@ from tensorflow.python.eager import backprop
 from tensorflow.python.eager import context
 from tensorflow.python.eager import test
 from tensorflow.python.framework import constant_op
+from tensorflow.python.framework import dtypes
 from tensorflow.python.framework import test_util
 from tensorflow.python.ops import array_ops
 from tensorflow.python.ops import math_ops
@@ -69,6 +70,25 @@ class Tests(test.TestCase):
 
     self.assertAllEqual(x, y)
 
+  @test_util.assert_no_new_tensors
+  @test_util.assert_no_garbage_created
+  def testFastpathExecute_MixedPrecisionVariableMatMulCorrectResponse(self):
+    ctx = context.context()
+    a_2_by_2 = constant_op.constant(1.0, shape=[2, 2])
+    a_2_by_2_fp16 = math_ops.cast(a_2_by_2, dtype=dtypes.float16)
+    m = resource_variable_ops.ResourceVariable(a_2_by_2)
+    m = resource_variable_ops._MixedPrecisionVariable(
+        m, read_dtype=dtypes.float16)
+    x = pywrap_tensorflow.TFE_Py_FastPathExecute(
+        ctx._handle, ctx.device_name, "MatMul", None, None, m, m, "transpose_a",
+        False, "transpose_b", False)
+    y = pywrap_tensorflow.TFE_Py_FastPathExecute(
+        ctx._handle, ctx.device_name, "MatMul", None, None, a_2_by_2_fp16,
+        a_2_by_2_fp16, "transpose_a", False, "transpose_b", False)
+
+    self.assertEqual(x.dtype, dtypes.float16)
+    self.assertAllEqual(x, y)
+
   @test_util.assert_no_new_tensors
   @test_util.assert_no_garbage_created
   def testFastpathExecute_TapeWrite(self):
@@ -98,6 +118,29 @@ class Tests(test.TestCase):
     self.assertAllEqual(dz_dy.numpy(),
                         constant_op.constant(4.0, shape=[2, 2]).numpy())
 
+  @test_util.assert_no_new_tensors
+  @test_util.assert_no_garbage_created
+  def testFastpathExecute_MixedPrecisionVariableTapeWrite(self):
+    ctx = context.context()
+    with backprop.GradientTape(persistent=True) as tape:
+      a_2_by_2 = constant_op.constant(
+          [[1.0, 2.0], [3.0, 4.0]], dtype=dtypes.float32)
+      a_2_by_2_fp16 = math_ops.cast(a_2_by_2, dtype=dtypes.float16)
+      m1 = resource_variable_ops.ResourceVariable(a_2_by_2)
+      m2 = resource_variable_ops._MixedPrecisionVariable(
+          m1, read_dtype=dtypes.float16)
+      tape.watch(m2)
+      z = pywrap_tensorflow.TFE_Py_FastPathExecute(
+          ctx._handle, ctx.device_name, "MatMul", None, None, a_2_by_2_fp16, m2,
+          "transpose_a", False, "transpose_b", False)
+    dz_dy = tape.gradient(z, [m2])[0]
+    self.assertEqual(dz_dy.dtype, dtypes.float16)
+
+    expected_grads = math_ops.matmul(
+        array_ops.transpose(a_2_by_2_fp16),
+        constant_op.constant(1., shape=[2, 2], dtype=dtypes.float16)).numpy()
+    self.assertAllEqual(dz_dy.numpy(), expected_grads)
+
   # Tests homogeneous list op
   @test_util.assert_no_new_tensors
   @test_util.assert_no_garbage_created
diff --git a/tensorflow/python/eager/test.py b/tensorflow/python/eager/test.py
index f6a46e7eb3d03982f07bf4162d94c6038217bf61..33ee797678ed73c52ebb17723f688cec4feca402 100644
--- a/tensorflow/python/eager/test.py
+++ b/tensorflow/python/eager/test.py
@@ -23,6 +23,7 @@ from tensorflow.python.platform import test as _test
 from tensorflow.python.platform.test import *  # pylint: disable=wildcard-import
 
 
+# TODO(akshayka): Do away with this file.
 def main(argv=None):
   _ops.enable_eager_execution()
   _test.main(argv)
diff --git a/tensorflow/python/estimator/api/BUILD b/tensorflow/python/estimator/api/BUILD
index aa5a29e6dd148c39ebb098cb99cb1907d9c5a9d9..a75fa7d0aee56c4fd4faccfaf2fa07c399cedcc9 100644
--- a/tensorflow/python/estimator/api/BUILD
+++ b/tensorflow/python/estimator/api/BUILD
@@ -6,13 +6,14 @@ package(
 
 licenses(["notice"])  # Apache 2.0
 
-load("//tensorflow/tools/api/generator:api_gen.bzl", "gen_api_init_files")
-load("//tensorflow/tools/api/generator:api_gen.bzl", "ESTIMATOR_API_INIT_FILES")
+load("//tensorflow/python/tools/api/generator:api_gen.bzl", "gen_api_init_files")
+load("//tensorflow/python/tools/api/generator:api_gen.bzl", "ESTIMATOR_API_INIT_FILES")
 
 gen_api_init_files(
     name = "estimator_python_api_gen",
     api_name = "estimator",
     output_files = ESTIMATOR_API_INIT_FILES,
+    output_package = "tensorflow.python.estimator.api",
     package = "tensorflow.python.estimator",
     package_dep = "//tensorflow/python/estimator:estimator_py",
 )
diff --git a/tensorflow/python/estimator/canned/baseline_test.py b/tensorflow/python/estimator/canned/baseline_test.py
index 7bf2e62da9c4598c28ad38825aac2031c9d51905..e46a3a156dfd546b733067299906857fbd705736 100644
--- a/tensorflow/python/estimator/canned/baseline_test.py
+++ b/tensorflow/python/estimator/canned/baseline_test.py
@@ -154,6 +154,8 @@ class BaselineRegressorEvaluationTest(test.TestCase):
     self.assertDictEqual({
         metric_keys.MetricKeys.LOSS: 9.,
         metric_keys.MetricKeys.LOSS_MEAN: 9.,
+        metric_keys.MetricKeys.PREDICTION_MEAN: 13.,
+        metric_keys.MetricKeys.LABEL_MEAN: 10.,
         ops.GraphKeys.GLOBAL_STEP: 100
     }, eval_metrics)
 
@@ -176,6 +178,8 @@ class BaselineRegressorEvaluationTest(test.TestCase):
     self.assertDictEqual({
         metric_keys.MetricKeys.LOSS: 18.,
         metric_keys.MetricKeys.LOSS_MEAN: 9.,
+        metric_keys.MetricKeys.PREDICTION_MEAN: 13.,
+        metric_keys.MetricKeys.LABEL_MEAN: 10.,
         ops.GraphKeys.GLOBAL_STEP: 100
     }, eval_metrics)
 
@@ -204,6 +208,8 @@ class BaselineRegressorEvaluationTest(test.TestCase):
     self.assertDictEqual({
         metric_keys.MetricKeys.LOSS: 27.,
         metric_keys.MetricKeys.LOSS_MEAN: 9.,
+        metric_keys.MetricKeys.PREDICTION_MEAN: 13.,
+        metric_keys.MetricKeys.LABEL_MEAN: 10.,
         ops.GraphKeys.GLOBAL_STEP: 100
     }, eval_metrics)
 
@@ -229,7 +235,9 @@ class BaselineRegressorEvaluationTest(test.TestCase):
 
     self.assertItemsEqual(
         (metric_keys.MetricKeys.LOSS, metric_keys.MetricKeys.LOSS_MEAN,
-         ops.GraphKeys.GLOBAL_STEP), eval_metrics.keys())
+         metric_keys.MetricKeys.PREDICTION_MEAN,
+         metric_keys.MetricKeys.LABEL_MEAN, ops.GraphKeys.GLOBAL_STEP),
+        eval_metrics.keys())
 
     # Logit is bias which is [46, 58]
     self.assertAlmostEqual(0, eval_metrics[metric_keys.MetricKeys.LOSS])
diff --git a/tensorflow/python/estimator/canned/boosted_trees.py b/tensorflow/python/estimator/canned/boosted_trees.py
index 8afef1b65a8d57e2b7ce3e4e512c622ca107ab83..3c832c7569f4936cfc11dd3742b00caed7a3c539 100644
--- a/tensorflow/python/estimator/canned/boosted_trees.py
+++ b/tensorflow/python/estimator/canned/boosted_trees.py
@@ -18,6 +18,7 @@ from __future__ import division
 from __future__ import print_function
 
 import collections
+import functools
 
 from tensorflow.python.estimator import estimator
 from tensorflow.python.estimator import model_fn
@@ -44,12 +45,13 @@ from tensorflow.python.util.tf_export import estimator_export
 # TODO(nponomareva): Reveal pruning params here.
 _TreeHParams = collections.namedtuple('TreeHParams', [
     'n_trees', 'max_depth', 'learning_rate', 'l1', 'l2', 'tree_complexity',
-    'min_node_weight'
+    'min_node_weight', 'center_bias'
 ])
 
 _HOLD_FOR_MULTI_CLASS_SUPPORT = object()
 _HOLD_FOR_MULTI_DIM_SUPPORT = object()
 _DUMMY_NUM_BUCKETS = -1
+_DUMMY_NODE_ID = -1
 
 
 def _get_transformed_features(features, sorted_feature_columns):
@@ -279,7 +281,9 @@ class _CacheTrainingStatesUsingHashTable(object):
     """Returns cached_tree_ids, cached_node_ids, cached_logits."""
     cached_tree_ids, cached_node_ids, cached_logits = array_ops.split(
         lookup_ops.lookup_table_find_v2(
-            self._table_ref, self._example_ids, default_value=[0.0, 0.0, 0.0]),
+            self._table_ref,
+            self._example_ids,
+            default_value=[0.0, _DUMMY_NODE_ID, 0.0]),
         [1, 1, self._logits_dimension],
         axis=1)
     cached_tree_ids = array_ops.squeeze(
@@ -330,7 +334,7 @@ class _CacheTrainingStatesUsingVariables(object):
         array_ops.zeros([batch_size], dtype=dtypes.int32),
         name='tree_ids_cache')
     self._node_ids = _local_variable(
-        array_ops.zeros([batch_size], dtype=dtypes.int32),
+        _DUMMY_NODE_ID*array_ops.ones([batch_size], dtype=dtypes.int32),
         name='node_ids_cache')
     self._logits = _local_variable(
         array_ops.zeros([batch_size, logits_dimension], dtype=dtypes.float32),
@@ -425,8 +429,8 @@ def _bt_model_fn(
     ValueError: mode or params are invalid, or features has the wrong type.
   """
   is_single_machine = (config.num_worker_replicas <= 1)
-
   sorted_feature_columns = sorted(feature_columns, key=lambda tc: tc.name)
+  center_bias = tree_hparams.center_bias
   if train_in_memory:
     assert n_batches_per_layer == 1, (
         'When train_in_memory is enabled, input_fn should return the entire '
@@ -469,6 +473,9 @@ def _bt_model_fn(
 
     # Create Ensemble resources.
     tree_ensemble = boosted_trees_ops.TreeEnsemble(name=name)
+    # Variable that determines whether bias centering is needed.
+    center_bias_var = variable_scope.variable(
+        initial_value=center_bias, name='center_bias_needed', trainable=False)
     # Create logits.
     if mode != model_fn.ModeKeys.TRAIN:
       logits = boosted_trees_ops.predict(
@@ -489,6 +496,7 @@ def _bt_model_fn(
         # TODO(soroush): Do partial updates if this becomes a bottleneck.
         ensemble_reload = local_tree_ensemble.deserialize(
             *tree_ensemble.serialize())
+
       if training_state_cache:
         cached_tree_ids, cached_node_ids, cached_logits = (
             training_state_cache.lookup())
@@ -497,9 +505,10 @@ def _bt_model_fn(
         batch_size = array_ops.shape(labels)[0]
         cached_tree_ids, cached_node_ids, cached_logits = (
             array_ops.zeros([batch_size], dtype=dtypes.int32),
-            array_ops.zeros([batch_size], dtype=dtypes.int32),
+            _DUMMY_NODE_ID * array_ops.ones([batch_size], dtype=dtypes.int32),
             array_ops.zeros(
                 [batch_size, head.logits_dimension], dtype=dtypes.float32))
+
       with ops.control_dependencies([ensemble_reload]):
         (stamp_token, num_trees, num_finalized_trees, num_attempted_layers,
          last_layer_nodes_range) = local_tree_ensemble.get_states()
@@ -513,13 +522,20 @@ def _bt_model_fn(
             cached_node_ids=cached_node_ids,
             bucketized_features=input_feature_list,
             logits_dimension=head.logits_dimension)
+
       logits = cached_logits + partial_logits
 
     # Create training graph.
     def _train_op_fn(loss):
       """Run one training iteration."""
       if training_state_cache:
-        train_op.append(training_state_cache.insert(tree_ids, node_ids, logits))
+        # Cache logits only after center_bias is complete, if it's in progress.
+        train_op.append(
+            control_flow_ops.cond(
+                center_bias_var, control_flow_ops.no_op,
+                lambda: training_state_cache.insert(tree_ids, node_ids, logits))
+        )
+
       if closed_form_grad_and_hess_fn:
         gradients, hessians = closed_form_grad_and_hess_fn(logits, labels)
       else:
@@ -543,8 +559,7 @@ def _bt_model_fn(
         ]
         stats_summaries_list.append(summaries)
 
-      accumulators = []
-
+      # ========= Helper methods for both in and not in memory. ==============
       def grow_tree_from_stats_summaries(stats_summaries_list,
                                          feature_ids_list):
         """Updates ensemble based on the best gains from stats summaries."""
@@ -591,55 +606,126 @@ def _bt_model_fn(
             pruning_mode=boosted_trees_ops.PruningMode.NO_PRUNING)
         return grow_op
 
+      def _center_bias_fn(mean_gradients, mean_hessians):
+        """Updates the ensembles and cache (if needed) with logits prior."""
+        continue_centering = boosted_trees_ops.center_bias(
+            tree_ensemble.resource_handle,
+            mean_gradients=mean_gradients,
+            mean_hessians=mean_hessians,
+            l1=tree_hparams.l1,
+            l2=tree_hparams.l2
+        )
+        return center_bias_var.assign(continue_centering)
+
+      # ========= End of helper methods. ==============
+
       if train_in_memory and is_single_machine:
         train_op.append(distribute_lib.increment_var(global_step))
+
+        mean_gradients = array_ops.expand_dims(
+            math_ops.reduce_mean(gradients, 0), 0)
+        mean_heassians = array_ops.expand_dims(
+            math_ops.reduce_mean(hessians, 0), 0)
+
         train_op.append(
-            grow_tree_from_stats_summaries(stats_summaries_list,
-                                           feature_ids_list))
+            control_flow_ops.cond(
+                center_bias_var,
+                lambda: _center_bias_fn(mean_gradients, mean_heassians),
+                functools.partial(grow_tree_from_stats_summaries,
+                                  stats_summaries_list, feature_ids_list)))
       else:
-        dependencies = []
 
-        for i, feature_ids in enumerate(feature_ids_list):
-          stats_summaries = stats_summaries_list[i]
-          accumulator = data_flow_ops.ConditionalAccumulator(
+        def center_bias_not_in_mem():
+          """Accumulates the data and updates the logits bias, when ready."""
+          bias_dependencies = []
+
+          bias_accumulator = data_flow_ops.ConditionalAccumulator(
               dtype=dtypes.float32,
-              # The stats consist of grads and hessians (the last dimension).
-              shape=[len(feature_ids), max_splits, bucket_size_list[i], 2],
-              shared_name='numeric_stats_summary_accumulator_' + str(i))
-          accumulators.append(accumulator)
-
-          apply_grad = accumulator.apply_grad(
-              array_ops.stack(stats_summaries, axis=0), stamp_token)
-          dependencies.append(apply_grad)
-
-        def grow_tree_from_accumulated_summaries_fn():
-          """Updates the tree with the best layer from accumulated summaries."""
-          # Take out the accumulated summaries from the accumulator and grow.
-          stats_summaries_list = []
-
-          stats_summaries_list = [
-              array_ops.unstack(accumulator.take_grad(1), axis=0)
-              for accumulator in accumulators
-          ]
-
-          grow_op = grow_tree_from_stats_summaries(stats_summaries_list,
-                                                   feature_ids_list)
-          return grow_op
-
-        with ops.control_dependencies(dependencies):
-          train_op.append(distribute_lib.increment_var(global_step))
-          if config.is_chief:
-            min_accumulated = math_ops.reduce_min(
-                array_ops.stack(
-                    [acc.num_accumulated() for acc in accumulators]))
-
-            train_op.append(
-                control_flow_ops.cond(
-                    math_ops.greater_equal(min_accumulated,
-                                           n_batches_per_layer),
-                    grow_tree_from_accumulated_summaries_fn,
-                    control_flow_ops.no_op,
-                    name='wait_until_n_batches_accumulated'))
+              # The stats consist of grads and hessians means only.
+              # TODO(nponomareva): this will change for a multiclass
+              shape=[2, 1],
+              shared_name='bias_accumulator')
+
+          grads_and_hess = array_ops.stack([gradients, hessians], axis=0)
+          grads_and_hess = math_ops.reduce_mean(grads_and_hess, axis=1)
+
+          apply_grad = bias_accumulator.apply_grad(grads_and_hess, stamp_token)
+          bias_dependencies.append(apply_grad)
+
+          def center_bias_from_accumulator():
+            accumulated = array_ops.unstack(
+                bias_accumulator.take_grad(1), axis=0)
+            return _center_bias_fn(
+                array_ops.expand_dims(accumulated[0], 0),
+                array_ops.expand_dims(accumulated[1], 0))
+
+          with ops.control_dependencies(bias_dependencies):
+            if config.is_chief:
+              center_bias_op = control_flow_ops.cond(
+                  math_ops.greater_equal(bias_accumulator.num_accumulated(),
+                                         n_batches_per_layer),
+                  center_bias_from_accumulator,
+                  control_flow_ops.no_op,
+                  name='wait_until_n_batches_for_bias_accumulated')
+
+              return center_bias_op
+            else:
+              return control_flow_ops.no_op()
+
+        def grow_not_in_mem():
+          """Accumulates the data and grows a layer when ready."""
+
+          accumulators = []
+          dependencies = []
+          for i, feature_ids in enumerate(feature_ids_list):
+            stats_summaries = stats_summaries_list[i]
+            accumulator = data_flow_ops.ConditionalAccumulator(
+                dtype=dtypes.float32,
+                # The stats consist of grads and hessians (the last dimension).
+                shape=[len(feature_ids), max_splits, bucket_size_list[i], 2],
+                shared_name='numeric_stats_summary_accumulator_' + str(i))
+            accumulators.append(accumulator)
+
+            apply_grad = accumulator.apply_grad(
+                array_ops.stack(stats_summaries, axis=0), stamp_token)
+            dependencies.append(apply_grad)
+
+          def grow_tree_from_accumulated_summaries_fn():
+            """Updates tree with the best layer from accumulated summaries."""
+            # Take out the accumulated summaries from the accumulator and grow.
+            stats_summaries_list = []
+
+            stats_summaries_list = [
+                array_ops.unstack(accumulator.take_grad(1), axis=0)
+                for accumulator in accumulators
+            ]
+
+            grow_op = grow_tree_from_stats_summaries(stats_summaries_list,
+                                                     feature_ids_list)
+            return grow_op
+
+          with ops.control_dependencies(dependencies):
+            if config.is_chief:
+              min_accumulated = math_ops.reduce_min(
+                  array_ops.stack(
+                      [acc.num_accumulated() for acc in accumulators]))
+
+              grow_model = control_flow_ops.cond(
+                  math_ops.greater_equal(min_accumulated, n_batches_per_layer),
+                  grow_tree_from_accumulated_summaries_fn,
+                  control_flow_ops.no_op,
+                  name='wait_until_n_batches_accumulated')
+
+              return grow_model
+            else:
+              return control_flow_ops.no_op()
+
+        update_model = control_flow_ops.cond(
+            center_bias_var, center_bias_not_in_mem, grow_not_in_mem)
+        train_op.append(update_model)
+        with ops.control_dependencies([update_model]):
+          increment_global = distribute_lib.increment_var(global_step)
+          train_op.append(increment_global)
 
       return control_flow_ops.group(train_op, name='train_op')
 
@@ -739,7 +825,8 @@ class BoostedTreesClassifier(estimator.Estimator):
                l2_regularization=0.,
                tree_complexity=0.,
                min_node_weight=0.,
-               config=None):
+               config=None,
+               center_bias=False):
     """Initializes a `BoostedTreesClassifier` instance.
 
     Example:
@@ -807,6 +894,13 @@ class BoostedTreesClassifier(estimator.Estimator):
         split to be considered. The value will be compared with
         sum(leaf_hessian)/(batch_size * n_batches_per_layer).
       config: `RunConfig` object to configure the runtime settings.
+      center_bias: Whether bias centering needs to occur. Bias centering refers
+        to the first node in the very first tree returning the prediction that
+        is aligned with the original labels distribution. For example, for
+        regression problems, the first node will return the mean of the labels.
+        For binary classification problems, it will return a logit for a prior
+        probability of label 1.
+
 
     Raises:
       ValueError: when wrong arguments are given or unsupported functionalities
@@ -821,7 +915,7 @@ class BoostedTreesClassifier(estimator.Estimator):
     # HParams for the model.
     tree_hparams = _TreeHParams(n_trees, max_depth, learning_rate,
                                 l1_regularization, l2_regularization,
-                                tree_complexity, min_node_weight)
+                                tree_complexity, min_node_weight, center_bias)
 
     def _model_fn(features, labels, mode, config):
       return _bt_model_fn(  # pylint: disable=protected-access
@@ -864,7 +958,8 @@ class BoostedTreesRegressor(estimator.Estimator):
                l2_regularization=0.,
                tree_complexity=0.,
                min_node_weight=0.,
-               config=None):
+               config=None,
+               center_bias=False):
     """Initializes a `BoostedTreesRegressor` instance.
 
     Example:
@@ -925,6 +1020,12 @@ class BoostedTreesRegressor(estimator.Estimator):
         split to be considered. The value will be compared with
         sum(leaf_hessian)/(batch_size * n_batches_per_layer).
       config: `RunConfig` object to configure the runtime settings.
+      center_bias: Whether bias centering needs to occur. Bias centering refers
+        to the first node in the very first tree returning the prediction that
+        is aligned with the original labels distribution. For example, for
+        regression problems, the first node will return the mean of the labels.
+        For binary classification problems, it will return a logit for a prior
+        probability of label 1.
 
     Raises:
       ValueError: when wrong arguments are given or unsupported functionalities
@@ -938,7 +1039,7 @@ class BoostedTreesRegressor(estimator.Estimator):
     # HParams for the model.
     tree_hparams = _TreeHParams(n_trees, max_depth, learning_rate,
                                 l1_regularization, l2_regularization,
-                                tree_complexity, min_node_weight)
+                                tree_complexity, min_node_weight, center_bias)
 
     def _model_fn(features, labels, mode, config):
       return _bt_model_fn(  # pylint: disable=protected-access
diff --git a/tensorflow/python/estimator/canned/boosted_trees_test.py b/tensorflow/python/estimator/canned/boosted_trees_test.py
index 33e9e69b041a7d250c9d86bdf8912bf0585f7d81..f807641057990971407f69ff0ba4d3513302e452 100644
--- a/tensorflow/python/estimator/canned/boosted_trees_test.py
+++ b/tensorflow/python/estimator/canned/boosted_trees_test.py
@@ -554,37 +554,495 @@ class ModelFnTests(test_util.TensorFlowTestCase):
             feature_column.numeric_column('f_%d' % i, dtype=dtypes.float32),
             BUCKET_BOUNDARIES) for i in range(NUM_FEATURES)
     }
-    self._tree_hparams = boosted_trees._TreeHParams(  # pylint:disable=protected-access
-        n_trees=2,
-        max_depth=2,
-        learning_rate=0.1,
-        l1=0.,
-        l2=0.01,
-        tree_complexity=0.,
-        min_node_weight=0.)
 
-  def _get_expected_ensembles_for_classification(self):
+  def _get_expected_ensembles_for_classification(self):
+    first_round = """
+        trees {
+          nodes {
+            bucketized_split {
+              feature_id: 2
+              threshold: 2
+              left_id: 1
+              right_id: 2
+            }
+            metadata {
+              gain: 0.387675
+            }
+          }
+          nodes {
+            leaf {
+              scalar: -0.181818
+            }
+          }
+          nodes {
+            leaf {
+              scalar: 0.0625
+            }
+          }
+        }
+        tree_weights: 1.0
+        tree_metadata {
+          num_layers_grown: 1
+          is_finalized: false
+        }
+        growing_metadata {
+          num_trees_attempted: 1
+          num_layers_attempted: 1
+          last_layer_node_start: 1
+          last_layer_node_end: 3
+        }
+        """
+    second_round = """
+        trees {
+          nodes {
+            bucketized_split {
+              feature_id: 2
+              threshold: 2
+              left_id: 1
+              right_id: 2
+            }
+            metadata {
+              gain: 0.387675
+            }
+          }
+          nodes {
+            bucketized_split {
+              feature_id: 0
+              threshold: 3
+              left_id: 3
+              right_id: 4
+            }
+            metadata {
+              gain: 0.0
+              original_leaf {
+                scalar: -0.181818
+              }
+            }
+          }
+          nodes {
+            bucketized_split {
+              feature_id: 0
+              threshold: 0
+              left_id: 5
+              right_id: 6
+            }
+            metadata {
+              gain: 0.105518
+              original_leaf {
+                scalar: 0.0625
+              }
+            }
+          }
+          nodes {
+            leaf {
+              scalar: -0.348397
+            }
+          }
+          nodes {
+            leaf {
+              scalar: -0.181818
+            }
+          }
+          nodes {
+            leaf {
+              scalar: 0.224091
+            }
+          }
+          nodes {
+            leaf {
+              scalar: 0.056815
+            }
+          }
+        }
+        trees {
+          nodes {
+            leaf {
+              scalar: 0.0
+            }
+          }
+        }
+        tree_weights: 1.0
+        tree_weights: 1.0
+        tree_metadata {
+          num_layers_grown: 2
+          is_finalized: true
+        }
+        tree_metadata {
+          num_layers_grown: 0
+          is_finalized: false
+        }
+        growing_metadata {
+          num_trees_attempted: 1
+          num_layers_attempted: 2
+          last_layer_node_start: 0
+          last_layer_node_end: 1
+        }
+        """
+    third_round = """
+        trees {
+          nodes {
+            bucketized_split {
+              feature_id: 2
+              threshold: 2
+              left_id: 1
+              right_id: 2
+            }
+            metadata {
+              gain: 0.387675
+            }
+          }
+          nodes {
+            bucketized_split {
+              feature_id: 0
+              threshold: 3
+              left_id: 3
+              right_id: 4
+            }
+            metadata {
+              gain: 0.0
+              original_leaf {
+                scalar: -0.181818
+              }
+            }
+          }
+          nodes {
+            bucketized_split {
+              feature_id: 0
+              threshold: 0
+              left_id: 5
+              right_id: 6
+            }
+            metadata {
+              gain: 0.105518
+              original_leaf {
+                scalar: 0.0625
+              }
+            }
+          }
+          nodes {
+            leaf {
+              scalar: -0.348397
+            }
+          }
+          nodes {
+            leaf {
+              scalar: -0.181818
+            }
+          }
+          nodes {
+            leaf {
+              scalar: 0.224091
+            }
+          }
+          nodes {
+            leaf {
+              scalar: 0.056815
+            }
+          }
+        }
+        trees {
+          nodes {
+            bucketized_split {
+              feature_id: 1
+              threshold: 0
+              left_id: 1
+              right_id: 2
+            }
+            metadata {
+              gain: 0.287131
+            }
+          }
+          nodes {
+            leaf {
+              scalar: 0.162042
+            }
+          }
+          nodes {
+            leaf {
+              scalar: -0.086986
+            }
+          }
+        }
+        tree_weights: 1.0
+        tree_weights: 1.0
+        tree_metadata {
+          num_layers_grown: 2
+          is_finalized: true
+        }
+        tree_metadata {
+          num_layers_grown: 1
+          is_finalized: false
+        }
+        growing_metadata {
+          num_trees_attempted: 2
+          num_layers_attempted: 3
+          last_layer_node_start: 1
+          last_layer_node_end: 3
+        }
+        """
+    return (first_round, second_round, third_round)
+
+  def _get_expected_ensembles_for_classification_with_bias(self):
+    first_round = """
+        trees {
+          nodes {
+            leaf {
+              scalar: -0.405086
+            }
+          }
+        }
+        tree_weights: 1.0
+        tree_metadata {
+        }
+        """
+    second_round = """
+        trees {
+          nodes {
+            bucketized_split {
+              feature_id: 2
+              threshold: 2
+              left_id: 1
+              right_id: 2
+            }
+            metadata {
+              gain: 0.407711
+              original_leaf {
+                scalar: -0.405086
+              }
+            }
+          }
+          nodes {
+            leaf {
+              scalar: -0.556054
+            }
+          }
+          nodes {
+            leaf {
+              scalar: -0.301233
+            }
+          }
+        }
+        tree_weights: 1.0
+        tree_metadata {
+          num_layers_grown: 1
+          is_finalized: false
+        }
+        growing_metadata {
+          num_trees_attempted: 1
+          num_layers_attempted: 1
+          last_layer_node_start: 1
+          last_layer_node_end: 3
+        }
+        """
+    third_round = """
+        trees {
+          nodes {
+            bucketized_split {
+              feature_id: 2
+              threshold: 2
+              left_id: 1
+              right_id: 2
+            }
+            metadata {
+              gain: 0.407711
+              original_leaf {
+                scalar: -0.405086
+              }
+            }
+          }
+          nodes {
+            bucketized_split {
+              feature_id: 0
+              threshold: 3
+              left_id: 3
+              right_id: 4
+            }
+            metadata {
+              original_leaf {
+                scalar: -0.556054
+              }
+            }
+          }
+          nodes {
+            bucketized_split {
+              feature_id: 0
+              threshold: 0
+              left_id: 5
+              right_id: 6
+            }
+            metadata {
+              gain: 0.09876
+              original_leaf {
+                scalar: -0.301233
+              }
+            }
+          }
+          nodes {
+            leaf {
+              scalar: -0.698072
+            }
+          }
+          nodes {
+            leaf {
+              scalar: -0.556054
+            }
+          }
+          nodes {
+            leaf {
+              scalar: -0.106016
+            }
+          }
+          nodes {
+            leaf {
+              scalar: -0.27349
+            }
+          }
+        }
+        trees {
+          nodes {
+            leaf {
+            }
+          }
+        }
+        tree_weights: 1.0
+        tree_weights: 1.0
+        tree_metadata {
+          num_layers_grown: 2
+          is_finalized: true
+        }
+        tree_metadata {
+        }
+        growing_metadata {
+          num_trees_attempted: 1
+          num_layers_attempted: 2
+          last_layer_node_end: 1
+        }
+        """
+    forth_round = """
+        trees {
+          nodes {
+            bucketized_split {
+              feature_id: 2
+              threshold: 2
+              left_id: 1
+              right_id: 2
+            }
+            metadata {
+              gain: 0.4077113
+              original_leaf {
+                scalar: -0.405086
+              }
+            }
+          }
+          nodes {
+            bucketized_split {
+              threshold: 3
+              left_id: 3
+              right_id: 4
+            }
+            metadata {
+              original_leaf {
+                scalar: -0.556054
+              }
+            }
+          }
+          nodes {
+            bucketized_split {
+              threshold: 0
+              left_id: 5
+              right_id: 6
+            }
+            metadata {
+              gain: 0.09876
+              original_leaf {
+                scalar: -0.301233
+              }
+            }
+          }
+          nodes {
+            leaf {
+              scalar: -0.698072
+            }
+          }
+          nodes {
+            leaf {
+              scalar: -0.556054
+            }
+          }
+          nodes {
+            leaf {
+              scalar: -0.106016
+            }
+          }
+          nodes {
+            leaf {
+              scalar: -0.27349
+            }
+          }
+        }
+        trees {
+          nodes {
+            bucketized_split {
+              feature_id: 2
+              threshold: 2
+              left_id: 1
+              right_id: 2
+            }
+            metadata {
+              gain: 0.289927
+            }
+          }
+          nodes {
+            leaf {
+              scalar: -0.134588
+            }
+          }
+          nodes {
+            leaf {
+              scalar: 0.083838            
+            }
+          }
+        }
+        tree_weights: 1.0
+        tree_weights: 1.0
+        tree_metadata {
+          num_layers_grown: 2
+          is_finalized: true
+        }
+        tree_metadata {
+          num_layers_grown: 1
+        }
+        growing_metadata {
+          num_trees_attempted: 2
+          num_layers_attempted: 3
+          last_layer_node_start: 1
+          last_layer_node_end: 3
+        }
+        """
+    return (first_round, second_round, third_round, forth_round)
+
+  def _get_expected_ensembles_for_regression(self):
     first_round = """
         trees {
           nodes {
             bucketized_split {
-              feature_id: 2
-              threshold: 2
+              feature_id: 1
+              threshold: 1
               left_id: 1
               right_id: 2
             }
             metadata {
-              gain: 0.387675
+              gain: 1.169714
             }
           }
           nodes {
             leaf {
-              scalar: -0.181818
+              scalar: 0.241322
             }
           }
           nodes {
             leaf {
-              scalar: 0.0625
+              scalar: 0.083951
             }
           }
         }
@@ -604,26 +1062,26 @@ class ModelFnTests(test_util.TensorFlowTestCase):
         trees {
           nodes {
             bucketized_split {
-              feature_id: 2
-              threshold: 2
+              feature_id: 1
+              threshold: 1
               left_id: 1
               right_id: 2
             }
             metadata {
-              gain: 0.387675
+              gain: 1.169714
             }
           }
           nodes {
             bucketized_split {
               feature_id: 0
-              threshold: 3
+              threshold: 1
               left_id: 3
               right_id: 4
             }
             metadata {
-              gain: 0.0
+              gain: 2.673407
               original_leaf {
-                scalar: -0.181818
+                scalar: 0.241322
               }
             }
           }
@@ -635,30 +1093,30 @@ class ModelFnTests(test_util.TensorFlowTestCase):
               right_id: 6
             }
             metadata {
-              gain: 0.105518
+              gain: 0.324102
               original_leaf {
-                scalar: 0.0625
+                scalar: 0.083951
               }
             }
           }
           nodes {
             leaf {
-              scalar: -0.348397
+              scalar: 0.563167
             }
           }
           nodes {
             leaf {
-              scalar: -0.181818
+              scalar: 0.247047
             }
           }
           nodes {
             leaf {
-              scalar: 0.224091
+              scalar: 0.095273
             }
           }
           nodes {
             leaf {
-              scalar: 0.056815
+              scalar: 0.222102
             }
           }
         }
@@ -690,26 +1148,26 @@ class ModelFnTests(test_util.TensorFlowTestCase):
         trees {
           nodes {
             bucketized_split {
-              feature_id: 2
-              threshold: 2
+              feature_id: 1
+              threshold: 1
               left_id: 1
               right_id: 2
             }
             metadata {
-              gain: 0.387675
+              gain: 1.169714
             }
           }
           nodes {
             bucketized_split {
               feature_id: 0
-              threshold: 3
+              threshold: 1
               left_id: 3
               right_id: 4
             }
             metadata {
-              gain: 0.0
+              gain: 2.673407
               original_leaf {
-                scalar: -0.181818
+                scalar: 0.241322
               }
             }
           }
@@ -721,30 +1179,30 @@ class ModelFnTests(test_util.TensorFlowTestCase):
               right_id: 6
             }
             metadata {
-              gain: 0.105518
+              gain: 0.324102
               original_leaf {
-                scalar: 0.0625
+                scalar: 0.083951
               }
             }
           }
           nodes {
             leaf {
-              scalar: -0.348397
+              scalar: 0.563167
             }
           }
           nodes {
             leaf {
-              scalar: -0.181818
+              scalar: 0.247047
             }
           }
           nodes {
             leaf {
-              scalar: 0.224091
+              scalar: 0.095273
             }
           }
           nodes {
             leaf {
-              scalar: 0.056815
+              scalar: 0.222102
             }
           }
         }
@@ -757,17 +1215,17 @@ class ModelFnTests(test_util.TensorFlowTestCase):
               right_id: 2
             }
             metadata {
-              gain: 0.287131
+              gain: 0.981026
             }
           }
           nodes {
             leaf {
-              scalar: 0.162042
+              scalar: 0.005166
             }
           }
           nodes {
             leaf {
-              scalar: -0.086986
+              scalar: 0.180281
             }
           }
         }
@@ -790,8 +1248,20 @@ class ModelFnTests(test_util.TensorFlowTestCase):
         """
     return (first_round, second_round, third_round)
 
-  def _get_expected_ensembles_for_regression(self):
+  def _get_expected_ensembles_for_regression_with_bias(self):
     first_round = """
+        trees {
+          nodes {
+            leaf {
+              scalar: 1.799974
+            }
+          }
+        }
+        tree_weights: 1.0
+        tree_metadata {
+        }
+        """
+    second_round = """
         trees {
           nodes {
             bucketized_split {
@@ -801,17 +1271,20 @@ class ModelFnTests(test_util.TensorFlowTestCase):
               right_id: 2
             }
             metadata {
-              gain: 1.169714
+              gain: 1.190442
+              original_leaf {
+                scalar: 1.799974
+              }
             }
           }
           nodes {
             leaf {
-              scalar: 0.241322
+              scalar: 1.862786
             }
           }
           nodes {
             leaf {
-              scalar: 0.083951
+              scalar: 1.706149
             }
           }
         }
@@ -827,7 +1300,7 @@ class ModelFnTests(test_util.TensorFlowTestCase):
           last_layer_node_end: 3
         }
         """
-    second_round = """
+    third_round = """
         trees {
           nodes {
             bucketized_split {
@@ -837,7 +1310,10 @@ class ModelFnTests(test_util.TensorFlowTestCase):
               right_id: 2
             }
             metadata {
-              gain: 1.169714
+              gain: 1.190442
+              original_leaf {
+                scalar: 1.799974
+              }
             }
           }
           nodes {
@@ -848,9 +1324,9 @@ class ModelFnTests(test_util.TensorFlowTestCase):
               right_id: 4
             }
             metadata {
-              gain: 2.673407
+              gain: 2.683594
               original_leaf {
-                scalar: 0.241322
+                scalar: 1.862786
               }
             }
           }
@@ -862,30 +1338,30 @@ class ModelFnTests(test_util.TensorFlowTestCase):
               right_id: 6
             }
             metadata {
-              gain: 0.324102
+              gain: 0.322693
               original_leaf {
-                scalar: 0.083951
+                scalar: 1.706149
               }
             }
           }
           nodes {
             leaf {
-              scalar: 0.563167
+              scalar: 2.024487
             }
           }
           nodes {
             leaf {
-              scalar: 0.247047
+              scalar: 1.710319
             }
           }
           nodes {
             leaf {
-              scalar: 0.095273
+              scalar: 1.559208
             }
           }
           nodes {
             leaf {
-              scalar: 0.222102
+              scalar: 1.686037
             }
           }
         }
@@ -913,7 +1389,7 @@ class ModelFnTests(test_util.TensorFlowTestCase):
           last_layer_node_end: 1
         }
         """
-    third_round = """
+    forth_round = """
         trees {
           nodes {
             bucketized_split {
@@ -923,55 +1399,55 @@ class ModelFnTests(test_util.TensorFlowTestCase):
               right_id: 2
             }
             metadata {
-              gain: 1.169714
+              gain: 1.190442
+              original_leaf {
+                scalar:  1.799974
+              }
             }
           }
           nodes {
             bucketized_split {
-              feature_id: 0
               threshold: 1
               left_id: 3
               right_id: 4
             }
             metadata {
-              gain: 2.673407
+              gain: 2.683594
               original_leaf {
-                scalar: 0.241322
+                scalar: 1.8627863
               }
             }
           }
           nodes {
             bucketized_split {
-              feature_id: 0
-              threshold: 0
               left_id: 5
               right_id: 6
             }
             metadata {
-              gain: 0.324102
+              gain: 0.322693
               original_leaf {
-                scalar: 0.083951
+                scalar: 1.706149
               }
             }
           }
           nodes {
             leaf {
-              scalar: 0.563167
+              scalar: 2.024487
             }
           }
           nodes {
             leaf {
-              scalar: 0.247047
+              scalar: 1.710319
             }
           }
           nodes {
             leaf {
-              scalar: 0.095273
+              scalar: 1.5592078
             }
           }
           nodes {
             leaf {
-              scalar: 0.222102
+              scalar: 1.686037
             }
           }
         }
@@ -979,22 +1455,21 @@ class ModelFnTests(test_util.TensorFlowTestCase):
           nodes {
             bucketized_split {
               feature_id: 1
-              threshold: 0
               left_id: 1
               right_id: 2
             }
             metadata {
-              gain: 0.981026
+              gain: 0.972589
             }
           }
           nodes {
             leaf {
-              scalar: 0.005166
+              scalar: -0.137592
             }
           }
           nodes {
             leaf {
-              scalar: 0.180281
+              scalar: 0.034926
             }
           }
         }
@@ -1006,7 +1481,6 @@ class ModelFnTests(test_util.TensorFlowTestCase):
         }
         tree_metadata {
           num_layers_grown: 1
-          is_finalized: false
         }
         growing_metadata {
           num_trees_attempted: 2
@@ -1015,19 +1489,34 @@ class ModelFnTests(test_util.TensorFlowTestCase):
           last_layer_node_end: 3
         }
         """
-    return (first_round, second_round, third_round)
-
-  def _get_train_op_and_ensemble(self, head, config, is_classification,
-                                 train_in_memory):
+    return (first_round, second_round, third_round, forth_round)
+
+  def _get_train_op_and_ensemble(self,
+                                 head,
+                                 config,
+                                 is_classification,
+                                 train_in_memory,
+                                 center_bias=False):
     """Calls bt_model_fn() and returns the train_op and ensemble_serialzed."""
     features, labels = _make_train_input_fn(is_classification)()
+
+    tree_hparams = boosted_trees._TreeHParams(  # pylint:disable=protected-access
+        n_trees=2,
+        max_depth=2,
+        learning_rate=0.1,
+        l1=0.,
+        l2=0.01,
+        tree_complexity=0.,
+        min_node_weight=0.,
+        center_bias=center_bias)
+
     estimator_spec = boosted_trees._bt_model_fn(  # pylint:disable=protected-access
         features=features,
         labels=labels,
         mode=model_fn.ModeKeys.TRAIN,
         head=head,
         feature_columns=self._feature_columns,
-        tree_hparams=self._tree_hparams,
+        tree_hparams=tree_hparams,
         example_id_column_name=EXAMPLE_ID_COLUMN,
         n_batches_per_layer=1,
         config=config,
@@ -1076,6 +1565,49 @@ class ModelFnTests(test_util.TensorFlowTestCase):
       ensemble_proto.ParseFromString(serialized)
       self.assertProtoEquals(expected_third, ensemble_proto)
 
+  def testTrainClassifierWithCenterBiasInMemory(self):
+    ops.reset_default_graph()
+
+    # When bias centering is on, we expect the very first node to have the
+    expected_first, expected_second, expected_third, expected_forth = (
+        self._get_expected_ensembles_for_classification_with_bias())
+
+    with self.test_session() as sess:
+      with sess.graph.as_default():
+        train_op, ensemble_serialized = self._get_train_op_and_ensemble(
+            boosted_trees._create_classification_head(n_classes=2),
+            run_config.RunConfig(),
+            is_classification=True,
+            train_in_memory=True,
+            center_bias=True)
+
+      # 4 iterations to center bias.
+      for _ in range(4):
+        _, serialized = sess.run([train_op, ensemble_serialized])
+
+      # Validate the trained ensemble.
+      ensemble_proto = boosted_trees_pb2.TreeEnsemble()
+      ensemble_proto.ParseFromString(serialized)
+      self.assertProtoEquals(expected_first, ensemble_proto)
+
+      _, serialized = sess.run([train_op, ensemble_serialized])
+      ensemble_proto = boosted_trees_pb2.TreeEnsemble()
+      ensemble_proto.ParseFromString(serialized)
+      self.assertProtoEquals(expected_second, ensemble_proto)
+
+      # Third round training and validation.
+      _, serialized = sess.run([train_op, ensemble_serialized])
+      ensemble_proto = boosted_trees_pb2.TreeEnsemble()
+      ensemble_proto.ParseFromString(serialized)
+      self.assertProtoEquals(expected_third, ensemble_proto)
+
+      # Forth round training and validation.
+      _, serialized = sess.run([train_op, ensemble_serialized])
+      ensemble_proto = boosted_trees_pb2.TreeEnsemble()
+      ensemble_proto.ParseFromString(serialized)
+
+      self.assertProtoEquals(expected_forth, ensemble_proto)
+
   def testTrainClassifierNonInMemory(self):
     ops.reset_default_graph()
     expected_first, expected_second, expected_third = (
@@ -1106,6 +1638,47 @@ class ModelFnTests(test_util.TensorFlowTestCase):
       ensemble_proto.ParseFromString(serialized)
       self.assertProtoEquals(expected_third, ensemble_proto)
 
+  def testTrainClassifierWithCenterBiasNonInMemory(self):
+    ops.reset_default_graph()
+
+    # When bias centering is on, we expect the very first node to have the
+    expected_first, expected_second, expected_third, expected_forth = (
+        self._get_expected_ensembles_for_classification_with_bias())
+
+    with self.test_session() as sess:
+      with sess.graph.as_default():
+        train_op, ensemble_serialized = self._get_train_op_and_ensemble(
+            boosted_trees._create_classification_head(n_classes=2),
+            run_config.RunConfig(),
+            is_classification=True,
+            train_in_memory=False,
+            center_bias=True)
+      # 4 iterations to center bias.
+      for _ in range(4):
+        _, serialized = sess.run([train_op, ensemble_serialized])
+      # Validate the trained ensemble.
+      ensemble_proto = boosted_trees_pb2.TreeEnsemble()
+      ensemble_proto.ParseFromString(serialized)
+      self.assertProtoEquals(expected_first, ensemble_proto)
+
+      # Run one more time and validate the trained ensemble.
+      _, serialized = sess.run([train_op, ensemble_serialized])
+      ensemble_proto = boosted_trees_pb2.TreeEnsemble()
+      ensemble_proto.ParseFromString(serialized)
+      self.assertProtoEquals(expected_second, ensemble_proto)
+
+      # Third round training and validation.
+      _, serialized = sess.run([train_op, ensemble_serialized])
+      ensemble_proto = boosted_trees_pb2.TreeEnsemble()
+      ensemble_proto.ParseFromString(serialized)
+      self.assertProtoEquals(expected_third, ensemble_proto)
+
+      # Forth round training and validation.
+      _, serialized = sess.run([train_op, ensemble_serialized])
+      ensemble_proto = boosted_trees_pb2.TreeEnsemble()
+      ensemble_proto.ParseFromString(serialized)
+      self.assertProtoEquals(expected_forth, ensemble_proto)
+
   def testTrainRegressorInMemory(self):
     ops.reset_default_graph()
     expected_first, expected_second, expected_third = (
@@ -1136,6 +1709,46 @@ class ModelFnTests(test_util.TensorFlowTestCase):
       ensemble_proto.ParseFromString(serialized)
       self.assertProtoEquals(expected_third, ensemble_proto)
 
+  def testTrainRegressorInMemoryWithCenterBias(self):
+    ops.reset_default_graph()
+    expected_first, expected_second, expected_third, expected_forth = (
+        self._get_expected_ensembles_for_regression_with_bias())
+    with self.test_session() as sess:
+      # Train with train_in_memory mode.
+      with sess.graph.as_default():
+        train_op, ensemble_serialized = self._get_train_op_and_ensemble(
+            boosted_trees._create_regression_head(label_dimension=1),
+            run_config.RunConfig(),
+            is_classification=False,
+            train_in_memory=True,
+            center_bias=True)
+      # 3 iterations to center bias.
+      for _ in range(3):
+        _, serialized = sess.run([train_op, ensemble_serialized])
+      # Validate the trained ensemble.
+      ensemble_proto = boosted_trees_pb2.TreeEnsemble()
+      ensemble_proto.ParseFromString(serialized)
+
+      self.assertProtoEquals(expected_first, ensemble_proto)
+
+      # Run one more time and validate the trained ensemble.
+      _, serialized = sess.run([train_op, ensemble_serialized])
+      ensemble_proto = boosted_trees_pb2.TreeEnsemble()
+      ensemble_proto.ParseFromString(serialized)
+      self.assertProtoEquals(expected_second, ensemble_proto)
+
+      # Third round training and validation.
+      _, serialized = sess.run([train_op, ensemble_serialized])
+      ensemble_proto = boosted_trees_pb2.TreeEnsemble()
+      ensemble_proto.ParseFromString(serialized)
+      self.assertProtoEquals(expected_third, ensemble_proto)
+
+      # Forth round training and validation.
+      _, serialized = sess.run([train_op, ensemble_serialized])
+      ensemble_proto = boosted_trees_pb2.TreeEnsemble()
+      ensemble_proto.ParseFromString(serialized)
+      self.assertProtoEquals(expected_forth, ensemble_proto)
+
   def testTrainRegressorNonInMemory(self):
     ops.reset_default_graph()
     expected_first, expected_second, expected_third = (
@@ -1166,6 +1779,46 @@ class ModelFnTests(test_util.TensorFlowTestCase):
       ensemble_proto.ParseFromString(serialized)
       self.assertProtoEquals(expected_third, ensemble_proto)
 
+  def testTrainRegressorNotInMemoryWithCenterBias(self):
+    ops.reset_default_graph()
+    expected_first, expected_second, expected_third, expected_forth = (
+        self._get_expected_ensembles_for_regression_with_bias())
+    with self.test_session() as sess:
+      # Train with train_in_memory mode.
+      with sess.graph.as_default():
+        train_op, ensemble_serialized = self._get_train_op_and_ensemble(
+            boosted_trees._create_regression_head(label_dimension=1),
+            run_config.RunConfig(),
+            is_classification=False,
+            train_in_memory=False,
+            center_bias=True)
+      # 3 iterations to center the bias (because we are using regularization).
+      for _ in range(3):
+        _, serialized = sess.run([train_op, ensemble_serialized])
+
+      # Validate the trained ensemble.
+      ensemble_proto = boosted_trees_pb2.TreeEnsemble()
+      ensemble_proto.ParseFromString(serialized)
+      self.assertProtoEquals(expected_first, ensemble_proto)
+
+      # Run one more time and validate the trained ensemble.
+      _, serialized = sess.run([train_op, ensemble_serialized])
+      ensemble_proto = boosted_trees_pb2.TreeEnsemble()
+      ensemble_proto.ParseFromString(serialized)
+      self.assertProtoEquals(expected_second, ensemble_proto)
+
+      # Third round training and validation.
+      _, serialized = sess.run([train_op, ensemble_serialized])
+      ensemble_proto = boosted_trees_pb2.TreeEnsemble()
+      ensemble_proto.ParseFromString(serialized)
+      self.assertProtoEquals(expected_third, ensemble_proto)
+
+      # Forth round training and validation.
+      _, serialized = sess.run([train_op, ensemble_serialized])
+      ensemble_proto = boosted_trees_pb2.TreeEnsemble()
+      ensemble_proto.ParseFromString(serialized)
+      self.assertProtoEquals(expected_forth, ensemble_proto)
+
 
 if __name__ == '__main__':
   googletest.main()
diff --git a/tensorflow/python/estimator/canned/dnn.py b/tensorflow/python/estimator/canned/dnn.py
index 90889e3e5d9f022f53c1f9f754bb01ae0a292f9c..c08cf61220716730fa495c6e327b91e8f3c69cd5 100644
--- a/tensorflow/python/estimator/canned/dnn.py
+++ b/tensorflow/python/estimator/canned/dnn.py
@@ -26,6 +26,7 @@ from tensorflow.python.estimator.canned import head as head_lib
 from tensorflow.python.estimator.canned import optimizers
 from tensorflow.python.feature_column import feature_column as feature_column_lib
 from tensorflow.python.layers import core as core_layers
+from tensorflow.python.layers import normalization
 from tensorflow.python.ops import init_ops
 from tensorflow.python.ops import nn
 from tensorflow.python.ops import partitioned_variables
@@ -45,7 +46,7 @@ def _add_hidden_layer_summary(value, tag):
 
 
 def _dnn_logit_fn_builder(units, hidden_units, feature_columns, activation_fn,
-                          dropout, input_layer_partitioner):
+                          dropout, input_layer_partitioner, batch_norm):
   """Function builder for a dnn logit_fn.
 
   Args:
@@ -58,6 +59,7 @@ def _dnn_logit_fn_builder(units, hidden_units, feature_columns, activation_fn,
     dropout: When not `None`, the probability we will drop out a given
       coordinate.
     input_layer_partitioner: Partitioner for input layer.
+    batch_norm: Whether to use batch normalization after each hidden layer.
 
   Returns:
     A logit_fn (see below).
@@ -83,6 +85,7 @@ def _dnn_logit_fn_builder(units, hidden_units, feature_columns, activation_fn,
       A `Tensor` representing the logits, or a list of `Tensor`'s representing
       multiple logits in the MultiHead case.
     """
+    is_training = mode == model_fn.ModeKeys.TRAIN
     with variable_scope.variable_scope(
         'input_from_feature_columns',
         values=tuple(six.itervalues(features)),
@@ -98,8 +101,20 @@ def _dnn_logit_fn_builder(units, hidden_units, feature_columns, activation_fn,
             activation=activation_fn,
             kernel_initializer=init_ops.glorot_uniform_initializer(),
             name=hidden_layer_scope)
-        if dropout is not None and mode == model_fn.ModeKeys.TRAIN:
+        if dropout is not None and is_training:
           net = core_layers.dropout(net, rate=dropout, training=True)
+        if batch_norm:
+          # TODO(hjm): In future, if this becomes popular, we can enable
+          # customization of the batch normalization params by accepting a
+          # list of `BatchNormalization` instances as `batch_norm`.
+          net = normalization.batch_normalization(
+              net,
+              # The default momentum 0.99 actually crashes on certain
+              # problem, so here we use 0.999, which is the default of
+              # tf.contrib.layers.batch_norm.
+              momentum=0.999,
+              training=is_training,
+              name='batchnorm_%d' % layer_id)
       _add_hidden_layer_summary(net, hidden_layer_scope.name)
 
     with variable_scope.variable_scope('logits', values=(net,)) as logits_scope:
@@ -127,7 +142,8 @@ def _dnn_model_fn(features,
                   dropout=None,
                   input_layer_partitioner=None,
                   config=None,
-                  tpu_estimator_spec=False):
+                  tpu_estimator_spec=False,
+                  batch_norm=False):
   """Deep Neural Net model_fn.
 
   Args:
@@ -150,6 +166,7 @@ def _dnn_model_fn(features,
     config: `RunConfig` object to configure the runtime settings.
     tpu_estimator_spec: Whether to return a `_TPUEstimatorSpec` or
       or `model_fn.EstimatorSpec` instance.
+    batch_norm: Whether to use batch normalization after each hidden layer.
 
   Returns:
     An `EstimatorSpec` instance.
@@ -182,7 +199,8 @@ def _dnn_model_fn(features,
         feature_columns=feature_columns,
         activation_fn=activation_fn,
         dropout=dropout,
-        input_layer_partitioner=input_layer_partitioner)
+        input_layer_partitioner=input_layer_partitioner,
+        batch_norm=batch_norm)
     logits = logit_fn(features=features, mode=mode)
 
     if tpu_estimator_spec:
@@ -230,6 +248,17 @@ class DNNClassifier(estimator.Estimator):
         l1_regularization_strength=0.001
       ))
 
+  # Or estimator using an optimizer with a learning rate decay.
+  estimator = DNNClassifier(
+      feature_columns=[categorical_feature_a_emb, categorical_feature_b_emb],
+      hidden_units=[1024, 512, 256],
+      optimizer=lambda: tf.AdamOptimizer(
+          learning_rate=tf.exponential_decay(
+              learning_rate=0.1,
+              global_step=tf.get_global_step(),
+              decay_steps=10000,
+              decay_rate=0.96))
+
   # Or estimator with warm-starting from a previous checkpoint.
   estimator = DNNClassifier(
       feature_columns=[categorical_feature_a_emb, categorical_feature_b_emb],
@@ -288,6 +317,7 @@ class DNNClassifier(estimator.Estimator):
       config=None,
       warm_start_from=None,
       loss_reduction=losses.Reduction.SUM,
+      batch_norm=False,
   ):
     """Initializes a `DNNClassifier` instance.
 
@@ -317,8 +347,9 @@ class DNNClassifier(estimator.Estimator):
         encoded as integer values in {0, 1,..., n_classes-1} for `n_classes`>2 .
         Also there will be errors if vocabulary is not provided and labels are
         string.
-      optimizer: An instance of `tf.Optimizer` used to train the model. Defaults
-        to Adagrad optimizer.
+      optimizer: An instance of `tf.Optimizer` used to train the model. Can also
+        be a string (one of 'Adagrad', 'Adam', 'Ftrl', 'RMSProp', 'SGD'), or
+        callable. Defaults to Adagrad optimizer.
       activation_fn: Activation function applied to each layer. If `None`, will
         use `tf.nn.relu`.
       dropout: When not `None`, the probability we will drop out a given
@@ -333,6 +364,7 @@ class DNNClassifier(estimator.Estimator):
         names are unchanged.
       loss_reduction: One of `tf.losses.Reduction` except `NONE`. Describes how
         to reduce training loss over batch. Defaults to `SUM`.
+      batch_norm: Whether to use batch normalization after each hidden layer.
     """
     head = head_lib._binary_logistic_or_multi_class_head(  # pylint: disable=protected-access
         n_classes, weight_column, label_vocabulary, loss_reduction)
@@ -349,7 +381,8 @@ class DNNClassifier(estimator.Estimator):
           activation_fn=activation_fn,
           dropout=dropout,
           input_layer_partitioner=input_layer_partitioner,
-          config=config)
+          config=config,
+          batch_norm=batch_norm)
 
     super(DNNClassifier, self).__init__(
         model_fn=_model_fn, model_dir=model_dir, config=config,
@@ -385,6 +418,17 @@ class DNNRegressor(estimator.Estimator):
         l1_regularization_strength=0.001
       ))
 
+  # Or estimator using an optimizer with a learning rate decay.
+  estimator = DNNRegressor(
+      feature_columns=[categorical_feature_a_emb, categorical_feature_b_emb],
+      hidden_units=[1024, 512, 256],
+      optimizer=lambda: tf.AdamOptimizer(
+          learning_rate=tf.exponential_decay(
+              learning_rate=0.1,
+              global_step=tf.get_global_step(),
+              decay_steps=10000,
+              decay_rate=0.96))
+
   # Or estimator with warm-starting from a previous checkpoint.
   estimator = DNNRegressor(
       feature_columns=[categorical_feature_a_emb, categorical_feature_b_emb],
@@ -442,6 +486,7 @@ class DNNRegressor(estimator.Estimator):
       config=None,
       warm_start_from=None,
       loss_reduction=losses.Reduction.SUM,
+      batch_norm=False,
   ):
     """Initializes a `DNNRegressor` instance.
 
@@ -465,8 +510,9 @@ class DNNRegressor(estimator.Estimator):
         used as a key to fetch weight tensor from the `features`. If it is a
         `_NumericColumn`, raw tensor is fetched by key `weight_column.key`,
         then weight_column.normalizer_fn is applied on it to get weight tensor.
-      optimizer: An instance of `tf.Optimizer` used to train the model. Defaults
-        to Adagrad optimizer.
+      optimizer: An instance of `tf.Optimizer` used to train the model. Can also
+        be a string (one of 'Adagrad', 'Adam', 'Ftrl', 'RMSProp', 'SGD'), or
+        callable. Defaults to Adagrad optimizer.
       activation_fn: Activation function applied to each layer. If `None`, will
         use `tf.nn.relu`.
       dropout: When not `None`, the probability we will drop out a given
@@ -481,6 +527,7 @@ class DNNRegressor(estimator.Estimator):
         names are unchanged.
       loss_reduction: One of `tf.losses.Reduction` except `NONE`. Describes how
         to reduce training loss over batch. Defaults to `SUM`.
+      batch_norm: Whether to use batch normalization after each hidden layer.
     """
 
     def _model_fn(features, labels, mode, config):
@@ -498,7 +545,8 @@ class DNNRegressor(estimator.Estimator):
           activation_fn=activation_fn,
           dropout=dropout,
           input_layer_partitioner=input_layer_partitioner,
-          config=config)
+          config=config,
+          batch_norm=batch_norm)
 
     super(DNNRegressor, self).__init__(
         model_fn=_model_fn, model_dir=model_dir, config=config,
diff --git a/tensorflow/python/estimator/canned/dnn_linear_combined.py b/tensorflow/python/estimator/canned/dnn_linear_combined.py
index 3d1ad1365bc66b3ef8b973257dd8b86ded0ea847..efa7812452427a6cdd7854b50b7d95a9a003abbb 100644
--- a/tensorflow/python/estimator/canned/dnn_linear_combined.py
+++ b/tensorflow/python/estimator/canned/dnn_linear_combined.py
@@ -88,7 +88,9 @@ def _dnn_linear_combined_model_fn(features,
                                   dnn_activation_fn=nn.relu,
                                   dnn_dropout=None,
                                   input_layer_partitioner=None,
-                                  config=None):
+                                  config=None,
+                                  batch_norm=False,
+                                  linear_sparse_combiner='sum'):
   """Deep Neural Net and Linear combined model_fn.
 
   Args:
@@ -115,7 +117,10 @@ def _dnn_linear_combined_model_fn(features,
       coordinate.
     input_layer_partitioner: Partitioner for input layer.
     config: `RunConfig` object to configure the runtime settings.
-
+    batch_norm: Whether to use batch normalization after each hidden layer.
+    linear_sparse_combiner: A string specifying how to reduce the linear model
+      if a categorical column is multivalent.  One of "mean", "sqrtn", and
+      "sum".
   Returns:
     An `EstimatorSpec` instance.
 
@@ -164,7 +169,8 @@ def _dnn_linear_combined_model_fn(features,
           feature_columns=dnn_feature_columns,
           activation_fn=dnn_activation_fn,
           dropout=dnn_dropout,
-          input_layer_partitioner=input_layer_partitioner)
+          input_layer_partitioner=input_layer_partitioner,
+          batch_norm=batch_norm)
       dnn_logits = dnn_logit_fn(features=features, mode=mode)
 
   linear_parent_scope = 'linear'
@@ -182,7 +188,8 @@ def _dnn_linear_combined_model_fn(features,
         partitioner=input_layer_partitioner) as scope:
       logit_fn = linear._linear_logit_fn_builder(  # pylint: disable=protected-access
           units=head.logits_dimension,
-          feature_columns=linear_feature_columns)
+          feature_columns=linear_feature_columns,
+          sparse_combiner=linear_sparse_combiner)
       linear_logits = logit_fn(features=features)
       _add_layer_summary(linear_logits, scope.name)
 
@@ -257,12 +264,19 @@ class DNNLinearCombinedClassifier(estimator.Estimator):
       # warm-start settings
       warm_start_from="/path/to/checkpoint/dir")
 
-  # To apply L1 and L2 regularization, you can set optimizers as follows:
+  # To apply L1 and L2 regularization, you can set dnn_optimizer to:
   tf.train.ProximalAdagradOptimizer(
       learning_rate=0.1,
       l1_regularization_strength=0.001,
       l2_regularization_strength=0.001)
-  # It is same for FtrlOptimizer.
+  # To apply learning rate decay, you can set dnn_optimizer to a callable:
+  lambda: tf.AdamOptimizer(
+      learning_rate=tf.exponential_decay(
+          learning_rate=0.1,
+          global_step=tf.get_global_step(),
+          decay_steps=10000,
+          decay_rate=0.96)
+  # It is the same for linear_optimizer.
 
   # Input builders
   def input_fn_train: # returns x, y
@@ -314,7 +328,9 @@ class DNNLinearCombinedClassifier(estimator.Estimator):
                input_layer_partitioner=None,
                config=None,
                warm_start_from=None,
-               loss_reduction=losses.Reduction.SUM):
+               loss_reduction=losses.Reduction.SUM,
+               batch_norm=False,
+               linear_sparse_combiner='sum'):
     """Initializes a DNNLinearCombinedClassifier instance.
 
     Args:
@@ -325,12 +341,16 @@ class DNNLinearCombinedClassifier(estimator.Estimator):
         used by linear part of the model. All items in the set must be
         instances of classes derived from `FeatureColumn`.
       linear_optimizer: An instance of `tf.Optimizer` used to apply gradients to
-        the linear part of the model. Defaults to FTRL optimizer.
+        the linear part of the model. Can also be a string (one of 'Adagrad',
+        'Adam', 'Ftrl', 'RMSProp', 'SGD'), or callable. Defaults to FTRL
+        optimizer.
       dnn_feature_columns: An iterable containing all the feature columns used
         by deep part of the model. All items in the set must be instances of
         classes derived from `FeatureColumn`.
       dnn_optimizer: An instance of `tf.Optimizer` used to apply gradients to
-        the deep part of the model. Defaults to Adagrad optimizer.
+        the deep part of the model. Can also be a string (one of 'Adagrad',
+        'Adam', 'Ftrl', 'RMSProp', 'SGD'), or callable. Defaults to Adagrad
+        optimizer.
       dnn_hidden_units: List of hidden units per layer. All layers are fully
         connected.
       dnn_activation_fn: Activation function applied to each layer. If None,
@@ -363,6 +383,12 @@ class DNNLinearCombinedClassifier(estimator.Estimator):
         names are unchanged.
       loss_reduction: One of `tf.losses.Reduction` except `NONE`. Describes how
         to reduce training loss over batch. Defaults to `SUM`.
+      batch_norm: Whether to use batch normalization after each hidden layer.
+      linear_sparse_combiner: A string specifying how to reduce the linear model
+        if a categorical column is multivalent.  One of "mean", "sqrtn", and
+        "sum" -- these are effectively different ways to do example-level
+        normalization, which can be useful for bag-of-words features.  For more
+        details, see @{tf.feature_column.linear_model$linear_model}.
 
     Raises:
       ValueError: If both linear_feature_columns and dnn_features_columns are
@@ -402,7 +428,9 @@ class DNNLinearCombinedClassifier(estimator.Estimator):
           dnn_activation_fn=dnn_activation_fn,
           dnn_dropout=dnn_dropout,
           input_layer_partitioner=input_layer_partitioner,
-          config=config)
+          config=config,
+          batch_norm=batch_norm,
+          linear_sparse_combiner=linear_sparse_combiner)
 
     super(DNNLinearCombinedClassifier, self).__init__(
         model_fn=_model_fn, model_dir=model_dir, config=config,
@@ -441,12 +469,19 @@ class DNNLinearCombinedRegressor(estimator.Estimator):
       # warm-start settings
       warm_start_from="/path/to/checkpoint/dir")
 
-  # To apply L1 and L2 regularization, you can set optimizers as follows:
+  # To apply L1 and L2 regularization, you can set dnn_optimizer to:
   tf.train.ProximalAdagradOptimizer(
       learning_rate=0.1,
       l1_regularization_strength=0.001,
       l2_regularization_strength=0.001)
-  # It is same for FtrlOptimizer.
+  # To apply learning rate decay, you can set dnn_optimizer to a callable:
+  lambda: tf.AdamOptimizer(
+      learning_rate=tf.exponential_decay(
+          learning_rate=0.1,
+          global_step=tf.get_global_step(),
+          decay_steps=10000,
+          decay_rate=0.96)
+  # It is the same for linear_optimizer.
 
   # Input builders
   def input_fn_train: # returns x, y
@@ -497,7 +532,9 @@ class DNNLinearCombinedRegressor(estimator.Estimator):
                input_layer_partitioner=None,
                config=None,
                warm_start_from=None,
-               loss_reduction=losses.Reduction.SUM):
+               loss_reduction=losses.Reduction.SUM,
+               batch_norm=False,
+               linear_sparse_combiner='sum'):
     """Initializes a DNNLinearCombinedRegressor instance.
 
     Args:
@@ -508,12 +545,16 @@ class DNNLinearCombinedRegressor(estimator.Estimator):
         used by linear part of the model. All items in the set must be
         instances of classes derived from `FeatureColumn`.
       linear_optimizer: An instance of `tf.Optimizer` used to apply gradients to
-        the linear part of the model. Defaults to FTRL optimizer.
+        the linear part of the model. Can also be a string (one of 'Adagrad',
+        'Adam', 'Ftrl', 'RMSProp', 'SGD'), or callable. Defaults to FTRL
+        optimizer.
       dnn_feature_columns: An iterable containing all the feature columns used
         by deep part of the model. All items in the set must be instances of
         classes derived from `FeatureColumn`.
       dnn_optimizer: An instance of `tf.Optimizer` used to apply gradients to
-        the deep part of the model. Defaults to Adagrad optimizer.
+        the deep part of the model. Can also be a string (one of 'Adagrad',
+        'Adam', 'Ftrl', 'RMSProp', 'SGD'), or callable. Defaults to Adagrad
+        optimizer.
       dnn_hidden_units: List of hidden units per layer. All layers are fully
         connected.
       dnn_activation_fn: Activation function applied to each layer. If None,
@@ -540,6 +581,12 @@ class DNNLinearCombinedRegressor(estimator.Estimator):
         names are unchanged.
       loss_reduction: One of `tf.losses.Reduction` except `NONE`. Describes how
         to reduce training loss over batch. Defaults to `SUM`.
+      batch_norm: Whether to use batch normalization after each hidden layer.
+      linear_sparse_combiner: A string specifying how to reduce the linear model
+        if a categorical column is multivalent.  One of "mean", "sqrtn", and
+        "sum" -- these are effectively different ways to do example-level
+        normalization, which can be useful for bag-of-words features.  For more
+        details, see @{tf.feature_column.linear_model$linear_model}.
 
     Raises:
       ValueError: If both linear_feature_columns and dnn_features_columns are
@@ -570,7 +617,9 @@ class DNNLinearCombinedRegressor(estimator.Estimator):
           dnn_activation_fn=dnn_activation_fn,
           dnn_dropout=dnn_dropout,
           input_layer_partitioner=input_layer_partitioner,
-          config=config)
+          config=config,
+          batch_norm=batch_norm,
+          linear_sparse_combiner=linear_sparse_combiner)
 
     super(DNNLinearCombinedRegressor, self).__init__(
         model_fn=_model_fn, model_dir=model_dir, config=config,
diff --git a/tensorflow/python/estimator/canned/dnn_linear_combined_test.py b/tensorflow/python/estimator/canned/dnn_linear_combined_test.py
index d275695eb319117cf94aefd7038ab5ee685e05a9..d16318659ba8fac70486e88fff07d71e060eac9b 100644
--- a/tensorflow/python/estimator/canned/dnn_linear_combined_test.py
+++ b/tensorflow/python/estimator/canned/dnn_linear_combined_test.py
@@ -100,7 +100,8 @@ def _linear_regressor_fn(feature_columns,
                          weight_column=None,
                          optimizer='Ftrl',
                          config=None,
-                         partitioner=None):
+                         partitioner=None,
+                         sparse_combiner='sum'):
   return dnn_linear_combined.DNNLinearCombinedRegressor(
       model_dir=model_dir,
       linear_feature_columns=feature_columns,
@@ -108,7 +109,8 @@ def _linear_regressor_fn(feature_columns,
       label_dimension=label_dimension,
       weight_column=weight_column,
       input_layer_partitioner=partitioner,
-      config=config)
+      config=config,
+      linear_sparse_combiner=sparse_combiner)
 
 
 class LinearOnlyRegressorPartitionerTest(
@@ -163,7 +165,8 @@ def _linear_classifier_fn(feature_columns,
                           label_vocabulary=None,
                           optimizer='Ftrl',
                           config=None,
-                          partitioner=None):
+                          partitioner=None,
+                          sparse_combiner='sum'):
   return dnn_linear_combined.DNNLinearCombinedClassifier(
       model_dir=model_dir,
       linear_feature_columns=feature_columns,
@@ -172,7 +175,8 @@ def _linear_classifier_fn(feature_columns,
       weight_column=weight_column,
       label_vocabulary=label_vocabulary,
       input_layer_partitioner=partitioner,
-      config=config)
+      config=config,
+      linear_sparse_combiner=sparse_combiner)
 
 
 class LinearOnlyClassifierTrainingTest(
diff --git a/tensorflow/python/estimator/canned/dnn_testing_utils.py b/tensorflow/python/estimator/canned/dnn_testing_utils.py
index 06a648777f8f730b4c739a69528090c5821f2681..de226ed0ef28e6a026e5df6ce128e178254a8c93 100644
--- a/tensorflow/python/estimator/canned/dnn_testing_utils.py
+++ b/tensorflow/python/estimator/canned/dnn_testing_utils.py
@@ -65,6 +65,11 @@ from tensorflow.python.training import training_util
 LEARNING_RATE_NAME = 'dnn/regression_head/dnn/learning_rate'
 HIDDEN_WEIGHTS_NAME_PATTERN = 'dnn/hiddenlayer_%d/kernel'
 HIDDEN_BIASES_NAME_PATTERN = 'dnn/hiddenlayer_%d/bias'
+BATCH_NORM_BETA_NAME_PATTERN = 'dnn/hiddenlayer_%d/batchnorm_%d/beta'
+BATCH_NORM_GAMMA_NAME_PATTERN = 'dnn/hiddenlayer_%d/batchnorm_%d/gamma'
+BATCH_NORM_MEAN_NAME_PATTERN = 'dnn/hiddenlayer_%d/batchnorm_%d/moving_mean'
+BATCH_NORM_VARIANCE_NAME_PATTERN = (
+    'dnn/hiddenlayer_%d/batchnorm_%d/moving_variance')
 LOGITS_WEIGHTS_NAME = 'dnn/logits/kernel'
 LOGITS_BIASES_NAME = 'dnn/logits/bias'
 OCCUPATION_EMBEDDING_NAME = ('dnn/input_from_feature_columns/input_layer/'
@@ -89,7 +94,10 @@ def assert_close(expected, actual, rtol=1e-04, message='', name='assert_close'):
         name=scope)
 
 
-def create_checkpoint(weights_and_biases, global_step, model_dir):
+def create_checkpoint(weights_and_biases,
+                      global_step,
+                      model_dir,
+                      batch_norm_vars=None):
   """Create checkpoint file with provided model weights.
 
   Args:
@@ -98,12 +106,20 @@ def create_checkpoint(weights_and_biases, global_step, model_dir):
     model_dir: Directory into which checkpoint is saved.
   """
   weights, biases = zip(*weights_and_biases)
+  if batch_norm_vars:
+    assert len(batch_norm_vars) == len(weights_and_biases) - 1
+    (bn_betas, bn_gammas, bn_means, bn_variances) = zip(*batch_norm_vars)
   model_weights = {}
 
   # Hidden layer weights.
   for i in range(0, len(weights) - 1):
     model_weights[HIDDEN_WEIGHTS_NAME_PATTERN % i] = weights[i]
     model_weights[HIDDEN_BIASES_NAME_PATTERN % i] = biases[i]
+    if batch_norm_vars:
+      model_weights[BATCH_NORM_BETA_NAME_PATTERN % (i, i)] = bn_betas[i]
+      model_weights[BATCH_NORM_GAMMA_NAME_PATTERN % (i, i)] = bn_gammas[i]
+      model_weights[BATCH_NORM_MEAN_NAME_PATTERN % (i, i)] = bn_means[i]
+      model_weights[BATCH_NORM_VARIANCE_NAME_PATTERN % (i, i)] = bn_variances[i]
 
   # Output layer weights.
   model_weights[LOGITS_WEIGHTS_NAME] = weights[-1]
@@ -503,8 +519,13 @@ class BaseDNNLogitFnTest(object):
       writer_cache.FileWriterCache.clear()
       shutil.rmtree(self._model_dir)
 
-  def _test_logits(self, mode, hidden_units, logits_dimension, inputs,
-                   expected_logits):
+  def _test_logits(self,
+                   mode,
+                   hidden_units,
+                   logits_dimension,
+                   inputs,
+                   expected_logits,
+                   batch_norm=False):
     """Tests that the expected logits are calculated."""
     with ops.Graph().as_default():
       # Global step needed for MonitoredSession, which is in turn used to
@@ -525,7 +546,8 @@ class BaseDNNLogitFnTest(object):
             ],
             activation_fn=nn.relu,
             dropout=None,
-            input_layer_partitioner=input_layer_partitioner)
+            input_layer_partitioner=input_layer_partitioner,
+            batch_norm=batch_norm)
         logits = logit_fn(
             features={'age': constant_op.constant(inputs)}, mode=mode)
         with monitored_session.MonitoredTrainingSession(
@@ -556,6 +578,69 @@ class BaseDNNLogitFnTest(object):
           inputs=[[10.]],
           expected_logits=[[-2.08]])
 
+  def test_one_dim_logits_with_batch_norm(self):
+    """Tests one-dimensional logits.
+
+    input_layer = [[10]]
+    hidden_layer_0 = [[relu(0.6*10 +1), relu(0.5*10 -1)]] = [[7, 4]]
+    hidden_layer_0 = [[relu(0.6*20 +1), relu(0.5*20 -1)]] = [[13, 9]]
+
+    batch_norm_0, training (epsilon = 0.001):
+      mean1 = 1/2*(7+13) = 10,
+      variance1 = 1/2*(3^2+3^2) = 9
+      x11 = (7-10)/sqrt(9+0.001) = -0.999944449,
+      x21 = (13-10)/sqrt(9+0.001) = 0.999944449,
+
+      mean2 = 1/2*(4+9) = 6.5,
+      variance2 = 1/2*(2.5^2+.2.5^2) = 6.25
+      x12 = (4-6.5)/sqrt(6.25+0.001) = -0.99992001,
+      x22 = (9-6.5)/sqrt(6.25+0.001) = 0.99992001,
+
+    logits = [[-1*(-0.999944449) + 2*(-0.99992001) + 0.3],
+              [-1*0.999944449 + 2*0.99992001 + 0.3]]
+           = [[-0.699895571],[1.299895571]]
+
+    batch_norm_0, not training (epsilon = 0.001):
+      moving_mean1 = 0, moving_variance1 = 1
+      x11 = (7-0)/sqrt(1+0.001) = 6.996502623,
+      x21 = (13-0)/sqrt(1+0.001) = 12.993504871,
+      moving_mean2 = 0, moving_variance2 = 1
+      x12 = (4-0)/sqrt(1+0.001) = 3.998001499,
+      x22 = (9-0)/sqrt(1+0.001) = 8.995503372,
+
+    logits = [[-1*6.996502623 + 2*3.998001499 + 0.3],
+              [-1*12.993504871 + 2*8.995503372 + 0.3]]
+           = [[1.299500375],[5.297501873]]
+    """
+    base_global_step = 100
+    create_checkpoint(
+        (
+            ([[.6, .5]], [1., -1.]),
+            ([[-1.], [2.]], [.3]),
+        ),
+        base_global_step,
+        self._model_dir,
+        batch_norm_vars=([[0, 0],  # beta.
+                          [1, 1],  # gamma.
+                          [0, 0],  # moving mean.
+                          [1, 1],  # moving variance.
+                         ],))
+    self._test_logits(
+        model_fn.ModeKeys.TRAIN,
+        hidden_units=[2],
+        logits_dimension=1,
+        inputs=[[10.], [20.]],
+        expected_logits=[[-0.699895571], [1.299895571]],
+        batch_norm=True)
+    for mode in [model_fn.ModeKeys.EVAL, model_fn.ModeKeys.PREDICT]:
+      self._test_logits(
+          mode,
+          hidden_units=[2],
+          logits_dimension=1,
+          inputs=[[10.], [20.]],
+          expected_logits=[[1.299500375], [5.297501873]],
+          batch_norm=True)
+
   def test_multi_dim_logits(self):
     """Tests multi-dimensional logits.
 
@@ -706,7 +791,8 @@ class BaseDNNLogitFnTest(object):
               ],
               activation_fn=nn.relu,
               dropout=None,
-              input_layer_partitioner=input_layer_partitioner)
+              input_layer_partitioner=input_layer_partitioner,
+              batch_norm=False)
           logits = logit_fn(
               features={
                   'age': constant_op.constant(inputs[0]),
@@ -1185,6 +1271,8 @@ class BaseDNNRegressorEvaluateTest(object):
     self.assertAllClose({
         metric_keys.MetricKeys.LOSS: expected_loss,
         metric_keys.MetricKeys.LOSS_MEAN: expected_loss,
+        metric_keys.MetricKeys.PREDICTION_MEAN: -2.08,
+        metric_keys.MetricKeys.LABEL_MEAN: 1.0,
         ops.GraphKeys.GLOBAL_STEP: global_step
     }, dnn_regressor.evaluate(input_fn=_input_fn, steps=1))
 
@@ -1215,6 +1303,8 @@ class BaseDNNRegressorEvaluateTest(object):
     self.assertAllClose({
         metric_keys.MetricKeys.LOSS: expected_loss,
         metric_keys.MetricKeys.LOSS_MEAN: expected_loss / label_dimension,
+        metric_keys.MetricKeys.PREDICTION_MEAN: 0.39 / 3.0,
+        metric_keys.MetricKeys.LABEL_MEAN: 0.5 / 3.0,
         ops.GraphKeys.GLOBAL_STEP: global_step
     }, dnn_regressor.evaluate(input_fn=_input_fn, steps=1))
 
diff --git a/tensorflow/python/estimator/canned/head.py b/tensorflow/python/estimator/canned/head.py
index b74ef1015cc564c20370e17e94e3a09d460c4f85..da9a64c2bc9f6b6797ef6cc115f36a73616b2e1e 100644
--- a/tensorflow/python/estimator/canned/head.py
+++ b/tensorflow/python/estimator/canned/head.py
@@ -1398,15 +1398,21 @@ class _RegressionHeadWithMeanSquaredErrorLoss(_Head):
         weights=weights,
         processed_labels=labels)
 
-  def _eval_metric_ops(self, weights, unreduced_loss, regularization_loss):
+  def _eval_metric_ops(self, predicted_value, labels, weights, unreduced_loss,
+                       regularization_loss):
     """Returns the Eval metric ops."""
     keys = metric_keys.MetricKeys
     # Estimator already adds a metric for loss.
     eval_metric_ops = {
         _summary_key(self._name, keys.LOSS_MEAN):
-            metrics_lib.mean(
-                values=unreduced_loss,
-                weights=weights)
+            metrics_lib.mean(values=unreduced_loss, weights=weights),
+        _summary_key(self._name, keys.PREDICTION_MEAN):
+            _predictions_mean(
+                predictions=predicted_value,
+                weights=weights,
+                name=keys.PREDICTION_MEAN),
+        _summary_key(self._name, keys.LABEL_MEAN):
+            metrics_lib.mean(values=labels, weights=weights)
     }
     if regularization_loss is not None:
       regularization_loss_key = _summary_key(
@@ -1489,13 +1495,13 @@ class _RegressionHeadWithMeanSquaredErrorLoss(_Head):
             predictions=predictions,
             loss=regularized_training_loss,
             eval_metrics=_create_eval_metrics_tuple(
-                self._eval_metric_ops,
-                {
+                self._eval_metric_ops, {
+                    'predicted_value': predicted_value,
+                    'labels': labels,
                     'weights': weights,
                     'unreduced_loss': unreduced_loss,
                     'regularization_loss': regularization_loss,
-                }
-            ))
+                }))
 
       # Train.
       if optimizer is not None:
diff --git a/tensorflow/python/estimator/canned/head_test.py b/tensorflow/python/estimator/canned/head_test.py
index 08ce5ca8e833fdd88f9c45b668f0914fcc70acd0..bd2e0ae943fb4da2acc09b120db59cf08e4ed9e6 100644
--- a/tensorflow/python/estimator/canned/head_test.py
+++ b/tensorflow/python/estimator/canned/head_test.py
@@ -3103,8 +3103,10 @@ class RegressionHead(test.TestCase):
     self.assertItemsEqual((prediction_key,), spec.predictions.keys())
     self.assertEqual(dtypes.float32, spec.predictions[prediction_key].dtype)
     self.assertEqual(dtypes.float32, spec.loss.dtype)
-    self.assertItemsEqual(
-        (metric_keys.MetricKeys.LOSS_MEAN,), spec.eval_metric_ops.keys())
+    self.assertItemsEqual((metric_keys.MetricKeys.LOSS_MEAN,
+                           metric_keys.MetricKeys.PREDICTION_MEAN,
+                           metric_keys.MetricKeys.LABEL_MEAN),
+                          spec.eval_metric_ops.keys())
     self.assertIsNone(spec.train_op)
     self.assertIsNone(spec.export_outputs)
     _assert_no_hooks(self, spec)
@@ -3140,6 +3142,9 @@ class RegressionHead(test.TestCase):
 
     expected_metric_keys = [
         '{}/some_regression_head'.format(metric_keys.MetricKeys.LOSS_MEAN),
+        '{}/some_regression_head'.format(
+            metric_keys.MetricKeys.PREDICTION_MEAN),
+        '{}/some_regression_head'.format(metric_keys.MetricKeys.LABEL_MEAN),
     ]
     self.assertItemsEqual(expected_metric_keys, spec.eval_metric_ops.keys())
 
@@ -3170,6 +3175,8 @@ class RegressionHead(test.TestCase):
     expected_metrics = {
         keys.LOSS_MEAN: expected_unregularized_loss,
         keys.LOSS_REGULARIZATION: expected_regularization_loss,
+        keys.PREDICTION_MEAN: (45 + 41) / 2.0,
+        keys.LABEL_MEAN: (43 + 44) / 2.0,
     }
 
     # Assert predictions, loss, and metrics.
@@ -3471,8 +3478,10 @@ class RegressionHead(test.TestCase):
     self.assertItemsEqual((prediction_key,), spec.predictions.keys())
     self.assertEqual(dtypes.float32, spec.predictions[prediction_key].dtype)
     self.assertEqual(dtypes.float32, spec.loss.dtype)
-    self.assertItemsEqual(
-        (metric_keys.MetricKeys.LOSS_MEAN,), spec.eval_metric_ops.keys())
+    self.assertItemsEqual((metric_keys.MetricKeys.LOSS_MEAN,
+                           metric_keys.MetricKeys.PREDICTION_MEAN,
+                           metric_keys.MetricKeys.LABEL_MEAN),
+                          spec.eval_metric_ops.keys())
     self.assertIsNone(spec.train_op)
     self.assertIsNone(spec.export_outputs)
     _assert_no_hooks(self, spec)
@@ -3700,8 +3709,10 @@ class RegressionHead(test.TestCase):
     self.assertItemsEqual((prediction_key,), spec.predictions.keys())
     self.assertEqual(dtypes.float32, spec.predictions[prediction_key].dtype)
     self.assertEqual(dtypes.float32, spec.loss.dtype)
-    self.assertItemsEqual(
-        (metric_keys.MetricKeys.LOSS_MEAN,), spec.eval_metric_ops.keys())
+    self.assertItemsEqual((metric_keys.MetricKeys.LOSS_MEAN,
+                           metric_keys.MetricKeys.PREDICTION_MEAN,
+                           metric_keys.MetricKeys.LABEL_MEAN),
+                          spec.eval_metric_ops.keys())
     self.assertIsNone(spec.train_op)
     self.assertIsNone(spec.export_outputs)
     _assert_no_hooks(self, spec)
@@ -3832,7 +3843,13 @@ class RegressionHead(test.TestCase):
     # losses = [1*(35-45)^2, .1*(42-41)^2, 1.5*(45-44)^2] = [100, .1, 1.5]
     # loss = sum(losses) = 100+.1+1.5 = 101.6
     # loss_mean = loss/(1+.1+1.5) = 101.6/2.6 = 39.076923
-    expected_metrics = {metric_keys.MetricKeys.LOSS_MEAN: 39.076923}
+    expected_metrics = {
+        metric_keys.MetricKeys.LOSS_MEAN:
+            39.076923,
+        metric_keys.MetricKeys.PREDICTION_MEAN:
+            (45 + 41 * 0.1 + 44 * 1.5) / 2.6,
+        metric_keys.MetricKeys.LABEL_MEAN: (35 + 42 * 0.1 + 45 * 1.5) / 2.6,
+    }
 
     # Assert spec contains expected tensors.
     self.assertEqual(dtypes.float32, spec.loss.dtype)
diff --git a/tensorflow/python/estimator/canned/linear.py b/tensorflow/python/estimator/canned/linear.py
index ac59e786c414f2093f8ab2c6eeb26101acdb2600..58a71603488198373bc4d1fd716538c2cee4d86f 100644
--- a/tensorflow/python/estimator/canned/linear.py
+++ b/tensorflow/python/estimator/canned/linear.py
@@ -66,13 +66,15 @@ def _compute_fraction_of_zero(cols_to_vars):
   return nn.zero_fraction(array_ops.concat(all_weight_vars, axis=0))
 
 
-def _linear_logit_fn_builder(units, feature_columns):
+def _linear_logit_fn_builder(units, feature_columns, sparse_combiner='sum'):
   """Function builder for a linear logit_fn.
 
   Args:
     units: An int indicating the dimension of the logit layer.
     feature_columns: An iterable containing all the feature columns used by
       the model.
+    sparse_combiner: A string specifying how to reduce if a categorical column
+      is multivalent.  One of "mean", "sqrtn", and "sum".
 
   Returns:
     A logit_fn (see below).
@@ -95,6 +97,7 @@ def _linear_logit_fn_builder(units, feature_columns):
         features=features,
         feature_columns=feature_columns,
         units=units,
+        sparse_combiner=sparse_combiner,
         cols_to_vars=cols_to_vars)
     bias = cols_to_vars.pop('bias')
     if units > 1:
@@ -111,7 +114,7 @@ def _linear_logit_fn_builder(units, feature_columns):
 
 
 def _linear_model_fn(features, labels, mode, head, feature_columns, optimizer,
-                     partitioner, config):
+                     partitioner, config, sparse_combiner='sum'):
   """A model_fn for linear models that use a gradient-based optimizer.
 
   Args:
@@ -126,6 +129,8 @@ def _linear_model_fn(features, labels, mode, head, feature_columns, optimizer,
       optimizer to use for training. If `None`, will use a FTRL optimizer.
     partitioner: Partitioner for variables.
     config: `RunConfig` object to configure the runtime settings.
+    sparse_combiner: A string specifying how to reduce if a categorical column
+      is multivalent.  One of "mean", "sqrtn", and "sum".
 
   Returns:
     An `EstimatorSpec` instance.
@@ -153,7 +158,8 @@ def _linear_model_fn(features, labels, mode, head, feature_columns, optimizer,
       partitioner=partitioner):
 
     logit_fn = _linear_logit_fn_builder(
-        units=head.logits_dimension, feature_columns=feature_columns)
+        units=head.logits_dimension, feature_columns=feature_columns,
+        sparse_combiner=sparse_combiner)
     logits = logit_fn(features=features)
 
     return head.create_estimator_spec(
@@ -193,6 +199,17 @@ class LinearClassifier(estimator.Estimator):
         l1_regularization_strength=0.001
       ))
 
+  # Or estimator using an optimizer with a learning rate decay.
+  estimator = LinearClassifier(
+      feature_columns=[categorical_column_a,
+                       categorical_feature_a_x_categorical_feature_b],
+      optimizer=lambda: tf.train.FtrlOptimizer(
+          learning_rate=tf.exponential_decay(
+              learning_rate=0.1,
+              global_step=tf.get_global_step(),
+              decay_steps=10000,
+              decay_rate=0.96))
+
   # Or estimator with warm-starting from a previous checkpoint.
   estimator = LinearClassifier(
       feature_columns=[categorical_column_a,
@@ -244,7 +261,8 @@ class LinearClassifier(estimator.Estimator):
                config=None,
                partitioner=None,
                warm_start_from=None,
-               loss_reduction=losses.Reduction.SUM):
+               loss_reduction=losses.Reduction.SUM,
+               sparse_combiner='sum'):
     """Construct a `LinearClassifier` estimator object.
 
     Args:
@@ -272,8 +290,9 @@ class LinearClassifier(estimator.Estimator):
         encoded as integer values in {0, 1,..., n_classes-1} for `n_classes`>2 .
         Also there will be errors if vocabulary is not provided and labels are
         string.
-      optimizer: An instance of `tf.Optimizer` used to train the model. Defaults
-        to FTRL optimizer.
+      optimizer: An instance of `tf.Optimizer` used to train the model. Can also
+        be a string (one of 'Adagrad', 'Adam', 'Ftrl', 'RMSProp', 'SGD'), or
+        callable. Defaults to FTRL optimizer.
       config: `RunConfig` object to configure the runtime settings.
       partitioner: Optional. Partitioner for input layer.
       warm_start_from: A string filepath to a checkpoint to warm-start from, or
@@ -283,6 +302,11 @@ class LinearClassifier(estimator.Estimator):
         and Tensor names are unchanged.
       loss_reduction: One of `tf.losses.Reduction` except `NONE`. Describes how
         to reduce training loss over batch. Defaults to `SUM`.
+      sparse_combiner: A string specifying how to reduce if a categorical column
+        is multivalent.  One of "mean", "sqrtn", and "sum" -- these are
+        effectively different ways to do example-level normalization, which can
+        be useful for bag-of-words features. for more details, see
+        @{tf.feature_column.linear_model$linear_model}.
 
     Returns:
       A `LinearClassifier` estimator.
@@ -311,7 +335,8 @@ class LinearClassifier(estimator.Estimator):
           feature_columns=tuple(feature_columns or []),
           optimizer=optimizer,
           partitioner=partitioner,
-          config=config)
+          config=config,
+          sparse_combiner=sparse_combiner)
 
     super(LinearClassifier, self).__init__(
         model_fn=_model_fn,
@@ -335,10 +360,31 @@ class LinearRegressor(estimator.Estimator):
 
   categorical_feature_a_x_categorical_feature_b = crossed_column(...)
 
+  # Estimator using the default optimizer.
   estimator = LinearRegressor(
       feature_columns=[categorical_column_a,
                        categorical_feature_a_x_categorical_feature_b])
 
+  # Or estimator using the FTRL optimizer with regularization.
+  estimator = LinearRegressor(
+      feature_columns=[categorical_column_a,
+                       categorical_feature_a_x_categorical_feature_b],
+      optimizer=tf.train.FtrlOptimizer(
+        learning_rate=0.1,
+        l1_regularization_strength=0.001
+      ))
+
+  # Or estimator using an optimizer with a learning rate decay.
+  estimator = LinearRegressor(
+      feature_columns=[categorical_column_a,
+                       categorical_feature_a_x_categorical_feature_b],
+      optimizer=lambda: tf.train.FtrlOptimizer(
+          learning_rate=tf.exponential_decay(
+              learning_rate=0.1,
+              global_step=tf.get_global_step(),
+              decay_steps=10000,
+              decay_rate=0.96))
+
   # Or estimator with warm-starting from a previous checkpoint.
   estimator = LinearRegressor(
       feature_columns=[categorical_column_a,
@@ -389,7 +435,8 @@ class LinearRegressor(estimator.Estimator):
                config=None,
                partitioner=None,
                warm_start_from=None,
-               loss_reduction=losses.Reduction.SUM):
+               loss_reduction=losses.Reduction.SUM,
+               sparse_combiner='sum'):
     """Initializes a `LinearRegressor` instance.
 
     Args:
@@ -409,8 +456,9 @@ class LinearRegressor(estimator.Estimator):
         used as a key to fetch weight tensor from the `features`. If it is a
         `_NumericColumn`, raw tensor is fetched by key `weight_column.key`,
         then weight_column.normalizer_fn is applied on it to get weight tensor.
-      optimizer: An instance of `tf.Optimizer` used to train the model. Defaults
-        to FTRL optimizer.
+      optimizer: An instance of `tf.Optimizer` used to train the model. Can also
+        be a string (one of 'Adagrad', 'Adam', 'Ftrl', 'RMSProp', 'SGD'), or
+        callable. Defaults to FTRL optimizer.
       config: `RunConfig` object to configure the runtime settings.
       partitioner: Optional. Partitioner for input layer.
       warm_start_from: A string filepath to a checkpoint to warm-start from, or
@@ -420,6 +468,11 @@ class LinearRegressor(estimator.Estimator):
         and Tensor names are unchanged.
       loss_reduction: One of `tf.losses.Reduction` except `NONE`. Describes how
         to reduce training loss over batch. Defaults to `SUM`.
+      sparse_combiner: A string specifying how to reduce if a categorical column
+        is multivalent.  One of "mean", "sqrtn", and "sum" -- these are
+        effectively different ways to do example-level normalization, which can
+        be useful for bag-of-words features. for more details, see
+        @{tf.feature_column.linear_model$linear_model}.
     """
     head = head_lib._regression_head(  # pylint: disable=protected-access
         label_dimension=label_dimension, weight_column=weight_column,
@@ -435,7 +488,8 @@ class LinearRegressor(estimator.Estimator):
           feature_columns=tuple(feature_columns or []),
           optimizer=optimizer,
           partitioner=partitioner,
-          config=config)
+          config=config,
+          sparse_combiner=sparse_combiner)
 
     super(LinearRegressor, self).__init__(
         model_fn=_model_fn,
diff --git a/tensorflow/python/estimator/canned/linear_testing_utils.py b/tensorflow/python/estimator/canned/linear_testing_utils.py
index 0e6436b42143f4b136165d47c41e143dacb4d476..c3934c7a801033d587465f0926301f30d4257fc7 100644
--- a/tensorflow/python/estimator/canned/linear_testing_utils.py
+++ b/tensorflow/python/estimator/canned/linear_testing_utils.py
@@ -29,6 +29,7 @@ import six
 from tensorflow.core.example import example_pb2
 from tensorflow.core.example import feature_pb2
 from tensorflow.python.client import session as tf_session
+from tensorflow.python.data.ops import dataset_ops
 from tensorflow.python.estimator import estimator
 from tensorflow.python.estimator import run_config
 from tensorflow.python.estimator.canned import linear
@@ -260,6 +261,8 @@ class BaseLinearRegressorEvaluationTest(object):
     self.assertDictEqual({
         metric_keys.MetricKeys.LOSS: 9.,
         metric_keys.MetricKeys.LOSS_MEAN: 9.,
+        metric_keys.MetricKeys.PREDICTION_MEAN: 13.,
+        metric_keys.MetricKeys.LABEL_MEAN: 10.,
         ops.GraphKeys.GLOBAL_STEP: 100
     }, eval_metrics)
 
@@ -285,6 +288,8 @@ class BaseLinearRegressorEvaluationTest(object):
     self.assertDictEqual({
         metric_keys.MetricKeys.LOSS: 18.,
         metric_keys.MetricKeys.LOSS_MEAN: 9.,
+        metric_keys.MetricKeys.PREDICTION_MEAN: 13.,
+        metric_keys.MetricKeys.LABEL_MEAN: 10.,
         ops.GraphKeys.GLOBAL_STEP: 100
     }, eval_metrics)
 
@@ -315,6 +320,8 @@ class BaseLinearRegressorEvaluationTest(object):
     self.assertDictEqual({
         metric_keys.MetricKeys.LOSS: 27.,
         metric_keys.MetricKeys.LOSS_MEAN: 9.,
+        metric_keys.MetricKeys.PREDICTION_MEAN: 13.,
+        metric_keys.MetricKeys.LABEL_MEAN: 10.,
         ops.GraphKeys.GLOBAL_STEP: 100
     }, eval_metrics)
 
@@ -345,7 +352,9 @@ class BaseLinearRegressorEvaluationTest(object):
 
     self.assertItemsEqual(
         (metric_keys.MetricKeys.LOSS, metric_keys.MetricKeys.LOSS_MEAN,
-         ops.GraphKeys.GLOBAL_STEP), eval_metrics.keys())
+         metric_keys.MetricKeys.PREDICTION_MEAN,
+         metric_keys.MetricKeys.LABEL_MEAN, ops.GraphKeys.GLOBAL_STEP),
+        eval_metrics.keys())
 
     # Logit is
     #   [2., 4., 5.] * [1.0, 2.0] + [7.0, 8.0] = [39, 50] + [7.0, 8.0]
@@ -382,7 +391,9 @@ class BaseLinearRegressorEvaluationTest(object):
     eval_metrics = est.evaluate(input_fn=input_fn, steps=1)
     self.assertItemsEqual(
         (metric_keys.MetricKeys.LOSS, metric_keys.MetricKeys.LOSS_MEAN,
-         ops.GraphKeys.GLOBAL_STEP), eval_metrics.keys())
+         metric_keys.MetricKeys.PREDICTION_MEAN,
+         metric_keys.MetricKeys.LABEL_MEAN, ops.GraphKeys.GLOBAL_STEP),
+        eval_metrics.keys())
 
     # Logit is [(20. * 10.0 + 4 * 2.0 + 5.0), (40. * 10.0 + 8 * 2.0 + 5.0)] =
     # [213.0, 421.0], while label is [213., 421.]. Loss = 0.
@@ -484,6 +495,69 @@ class BaseLinearRegressorPredictTest(object):
     # x0 * weight0 + x1 * weight1 + bias = 2. * 10. + 3. * 20 + .2 = 80.2
     self.assertAllClose([[80.2]], predicted_scores)
 
+  def testSparseCombiner(self):
+    w_a = 2.0
+    w_b = 3.0
+    w_c = 5.0
+    bias = 5.0
+    with ops.Graph().as_default():
+      variables_lib.Variable([[w_a], [w_b], [w_c]], name=LANGUAGE_WEIGHT_NAME)
+      variables_lib.Variable([bias], name=BIAS_NAME)
+      variables_lib.Variable(1, name=ops.GraphKeys.GLOBAL_STEP,
+                             dtype=dtypes.int64)
+      save_variables_to_ckpt(self._model_dir)
+
+    def _input_fn():
+      return dataset_ops.Dataset.from_tensors({
+          'language': sparse_tensor.SparseTensor(
+              values=['a', 'c', 'b', 'c'],
+              indices=[[0, 0], [0, 1], [1, 0], [1, 1]],
+              dense_shape=[2, 2]),
+      })
+
+    feature_columns = (
+        feature_column_lib.categorical_column_with_vocabulary_list(
+            'language', vocabulary_list=['a', 'b', 'c']),)
+
+    # Check prediction for each sparse_combiner.
+    # With sparse_combiner = 'sum', we have
+    # logits_1 = w_a + w_c + bias
+    #          = 2.0 + 5.0 + 5.0 = 12.0
+    # logits_2 = w_b + w_c + bias
+    #          = 3.0 + 5.0 + 5.0 = 13.0
+    linear_regressor = self._linear_regressor_fn(
+        feature_columns=feature_columns,
+        model_dir=self._model_dir)
+    predictions = linear_regressor.predict(input_fn=_input_fn)
+    predicted_scores = list([x['predictions'] for x in predictions])
+    self.assertAllClose([[12.0], [13.0]], predicted_scores)
+
+    # With sparse_combiner = 'mean', we have
+    # logits_1 = 1/2 * (w_a + w_c) + bias
+    #          = 1/2 * (2.0 + 5.0) + 5.0 = 8.5
+    # logits_2 = 1/2 * (w_b + w_c) + bias
+    #          = 1/2 * (3.0 + 5.0) + 5.0 = 9.0
+    linear_regressor = self._linear_regressor_fn(
+        feature_columns=feature_columns,
+        model_dir=self._model_dir,
+        sparse_combiner='mean')
+    predictions = linear_regressor.predict(input_fn=_input_fn)
+    predicted_scores = list([x['predictions'] for x in predictions])
+    self.assertAllClose([[8.5], [9.0]], predicted_scores)
+
+    # With sparse_combiner = 'sqrtn', we have
+    # logits_1 = sqrt(2)/2 * (w_a + w_c) + bias
+    #          = sqrt(2)/2 * (2.0 + 5.0) + 5.0 = 9.94974
+    # logits_2 = sqrt(2)/2 * (w_b + w_c) + bias
+    #          = sqrt(2)/2 * (3.0 + 5.0) + 5.0 = 10.65685
+    linear_regressor = self._linear_regressor_fn(
+        feature_columns=feature_columns,
+        model_dir=self._model_dir,
+        sparse_combiner='sqrtn')
+    predictions = linear_regressor.predict(input_fn=_input_fn)
+    predicted_scores = list([x['predictions'] for x in predictions])
+    self.assertAllClose([[9.94974], [10.65685]], predicted_scores)
+
 
 class BaseLinearRegressorIntegrationTest(object):
 
@@ -1636,6 +1710,69 @@ class BaseLinearClassifierPredictTest(object):
                           for i in range(n_classes)],
         label_output_fn=lambda x: ('class_vocab_%s' % x).encode())
 
+  def testSparseCombiner(self):
+    w_a = 2.0
+    w_b = 3.0
+    w_c = 5.0
+    bias = 5.0
+    with ops.Graph().as_default():
+      variables_lib.Variable([[w_a], [w_b], [w_c]], name=LANGUAGE_WEIGHT_NAME)
+      variables_lib.Variable([bias], name=BIAS_NAME)
+      variables_lib.Variable(1, name=ops.GraphKeys.GLOBAL_STEP,
+                             dtype=dtypes.int64)
+      save_variables_to_ckpt(self._model_dir)
+
+    def _input_fn():
+      return dataset_ops.Dataset.from_tensors({
+          'language': sparse_tensor.SparseTensor(
+              values=['a', 'c', 'b', 'c'],
+              indices=[[0, 0], [0, 1], [1, 0], [1, 1]],
+              dense_shape=[2, 2]),
+      })
+
+    feature_columns = (
+        feature_column_lib.categorical_column_with_vocabulary_list(
+            'language', vocabulary_list=['a', 'b', 'c']),)
+
+    # Check prediction for each sparse_combiner.
+    # With sparse_combiner = 'sum', we have
+    # logits_1 = w_a + w_c + bias
+    #          = 2.0 + 5.0 + 5.0 = 12.0
+    # logits_2 = w_b + w_c + bias
+    #          = 3.0 + 5.0 + 5.0 = 13.0
+    linear_classifier = self._linear_classifier_fn(
+        feature_columns=feature_columns,
+        model_dir=self._model_dir)
+    predictions = linear_classifier.predict(input_fn=_input_fn)
+    predicted_scores = list([x['logits'] for x in predictions])
+    self.assertAllClose([[12.0], [13.0]], predicted_scores)
+
+    # With sparse_combiner = 'mean', we have
+    # logits_1 = 1/2 * (w_a + w_c) + bias
+    #          = 1/2 * (2.0 + 5.0) + 5.0 = 8.5
+    # logits_2 = 1/2 * (w_b + w_c) + bias
+    #          = 1/2 * (3.0 + 5.0) + 5.0 = 9.0
+    linear_classifier = self._linear_classifier_fn(
+        feature_columns=feature_columns,
+        model_dir=self._model_dir,
+        sparse_combiner='mean')
+    predictions = linear_classifier.predict(input_fn=_input_fn)
+    predicted_scores = list([x['logits'] for x in predictions])
+    self.assertAllClose([[8.5], [9.0]], predicted_scores)
+
+    # With sparse_combiner = 'sqrtn', we have
+    # logits_1 = sqrt(2)/2 * (w_a + w_c) + bias
+    #          = sqrt(2)/2 * (2.0 + 5.0) + 5.0 = 9.94974
+    # logits_2 = sqrt(2)/2 * (w_b + w_c) + bias
+    #          = sqrt(2)/2 * (3.0 + 5.0) + 5.0 = 10.65685
+    linear_classifier = self._linear_classifier_fn(
+        feature_columns=feature_columns,
+        model_dir=self._model_dir,
+        sparse_combiner='sqrtn')
+    predictions = linear_classifier.predict(input_fn=_input_fn)
+    predicted_scores = list([x['logits'] for x in predictions])
+    self.assertAllClose([[9.94974], [10.65685]], predicted_scores)
+
 
 class BaseLinearClassifierIntegrationTest(object):
 
diff --git a/tensorflow/python/estimator/canned/optimizers.py b/tensorflow/python/estimator/canned/optimizers.py
index f72c5ca5cbb2721d967ad9ef9dfa896f7ccce240..8f51cc3a80dd9b91eb24a83577b7d0614615e008 100644
--- a/tensorflow/python/estimator/canned/optimizers.py
+++ b/tensorflow/python/estimator/canned/optimizers.py
@@ -72,6 +72,8 @@ def get_optimizer_instance(opt, learning_rate=None):
     raise ValueError(
         'Unsupported optimizer name: {}. Supported names are: {}'.format(
             opt, tuple(sorted(six.iterkeys(_OPTIMIZER_CLS_NAMES)))))
+  if callable(opt):
+    opt = opt()
   if not isinstance(opt, optimizer_lib.Optimizer):
     raise ValueError(
         'The given object is not an Optimizer instance. Given: {}'.format(opt))
diff --git a/tensorflow/python/estimator/canned/optimizers_test.py b/tensorflow/python/estimator/canned/optimizers_test.py
index ee28756155afd5ae3421475c3d41542db9411345..eadabdbc496334270cd792f5b8d5ff39a446bcf7 100644
--- a/tensorflow/python/estimator/canned/optimizers_test.py
+++ b/tensorflow/python/estimator/canned/optimizers_test.py
@@ -28,6 +28,13 @@ from tensorflow.python.training import optimizer as optimizer_lib
 from tensorflow.python.training import rmsprop
 
 
+class _TestOptimizer(optimizer_lib.Optimizer):
+
+  def __init__(self):
+    super(_TestOptimizer, self).__init__(
+        use_locking=False, name='TestOptimizer')
+
+
 class GetOptimizerInstance(test.TestCase):
 
   def test_unsupported_name(self):
@@ -66,12 +73,6 @@ class GetOptimizerInstance(test.TestCase):
     self.assertAlmostEqual(0.1, opt._learning_rate)
 
   def test_object(self):
-    class _TestOptimizer(optimizer_lib.Optimizer):
-
-      def __init__(self):
-        super(_TestOptimizer, self).__init__(
-            use_locking=False, name='TestOptimizer')
-
     opt = optimizers.get_optimizer_instance(_TestOptimizer())
     self.assertIsInstance(opt, _TestOptimizer)
 
@@ -80,6 +81,23 @@ class GetOptimizerInstance(test.TestCase):
         ValueError, 'The given object is not an Optimizer instance'):
       optimizers.get_optimizer_instance((1, 2, 3))
 
+  def test_callable(self):
+    def _optimizer_fn():
+      return _TestOptimizer()
+    opt = optimizers.get_optimizer_instance(_optimizer_fn)
+    self.assertIsInstance(opt, _TestOptimizer)
+
+  def test_lambda(self):
+    opt = optimizers.get_optimizer_instance(lambda: _TestOptimizer())  # pylint: disable=unnecessary-lambda
+    self.assertIsInstance(opt, _TestOptimizer)
+
+  def test_callable_returns_invalid(self):
+    def _optimizer_fn():
+      return (1, 2, 3)
+    with self.assertRaisesRegexp(
+        ValueError, 'The given object is not an Optimizer instance'):
+      optimizers.get_optimizer_instance(_optimizer_fn)
+
 
 if __name__ == '__main__':
   test.main()
diff --git a/tensorflow/python/estimator/estimator.py b/tensorflow/python/estimator/estimator.py
index 8df75d9eeeb6538c4b826fd27abc5a7e7208fcfa..253716b43ed8d12512c7bdc6fe100aa868e90bd1 100644
--- a/tensorflow/python/estimator/estimator.py
+++ b/tensorflow/python/estimator/estimator.py
@@ -38,6 +38,7 @@ from tensorflow.python.estimator import run_config
 from tensorflow.python.estimator import util as estimator_util
 from tensorflow.python.estimator.export import export as export_helpers
 from tensorflow.python.estimator.export import export_output
+from tensorflow.python.framework import constant_op
 from tensorflow.python.framework import dtypes
 from tensorflow.python.framework import errors
 from tensorflow.python.framework import ops
@@ -575,7 +576,9 @@ class Estimator(object):
     allowed_overrides = set([
         '_call_input_fn', '_create_global_step',
         '_convert_train_steps_to_hooks', '_convert_eval_steps_to_hooks',
-        '_tf_api_names', '_estimator_api_names', '_estimator_api_constants',
+        '_tf_api_names', '_tf_api_names_v1', '_estimator_api_names',
+        '_estimator_api_names_v1', '_estimator_api_constants',
+        '_estimator_api_constants_v1',
         '_validate_features_in_predict_input',
         '_call_model_fn', '_add_meta_graph_for_mode'
     ])
@@ -848,7 +851,8 @@ class Estimator(object):
                                strip_default_attrs,
                                save_variables=True,
                                mode=model_fn_lib.ModeKeys.PREDICT,
-                               export_tags=None):
+                               export_tags=None,
+                               check_variables=True):
     # pylint: disable=line-too-long
     """Loads variables and adds them along with a MetaGraphDef for saving.
 
@@ -869,6 +873,10 @@ class Estimator(object):
       mode: tf.estimator.ModeKeys value indicating which mode will be exported.
       export_tags: The set of tags with which to save `MetaGraphDef`. If None,
         a default set will be selected to matched the passed mode.
+      check_variables: bool, whether to check the checkpoint has all variables.
+
+    Raises:
+      ValueError: if `save_variables` is `True` and `check_variable` is `False`.
     """
     # pylint: enable=line-too-long
     if export_tags is None:
@@ -909,16 +917,20 @@ class Estimator(object):
         # SavedModel for restore later.
         graph_saver = estimator_spec.scaffold.saver or saver.Saver(sharded=True)
 
-        try:
-          graph_saver.restore(session, checkpoint_path)
-        except errors.NotFoundError as e:
-          msg = ('Could not load all requested variables from the checkpoint. '
-                 'Please make sure your model_fn does not expect variables '
-                 'that were not saved in the checkpoint.\n\n'
-                 'Encountered error with mode `{}` while restoring checkpoint '
-                 'from: `{}`. Full Traceback:\n\n{}').format(
-                     mode, checkpoint_path, e)
-          raise ValueError(msg)
+        if save_variables and not check_variables:
+          raise ValueError('If `save_variables` is `True, `check_variables`'
+                           'must not be `False`.')
+        if check_variables:
+          try:
+            graph_saver.restore(session, checkpoint_path)
+          except errors.NotFoundError as e:
+            msg = ('Could not load all requested variables from checkpoint. '
+                   'Please make sure your model_fn does not expect variables '
+                   'that were not saved in the checkpoint.\n\n'
+                   'Encountered error with mode `{}` while restoring '
+                   'checkpoint from: `{}`. Full Traceback:\n\n{}').format(
+                       mode, checkpoint_path, e)
+            raise ValueError(msg)
 
         # We add the train op explicitly for now, so that we don't have to
         # change the Builder public interface. Note that this is a no-op
@@ -1174,25 +1186,73 @@ class Estimator(object):
       Loss from training
     """
     self._distribution.configure(self._session_config)
+
+    # TODO(sourabhbajaj): Remove this hack once we migrate the other strategies
+    # to use the new API
+    is_tpu_strategy = self._distribution.__class__.__name__ == 'TPUStrategy'
+
     worker_hooks = []
     with ops.Graph().as_default() as g:
       with self._distribution.scope():
         random_seed.set_random_seed(self._config.tf_random_seed)
-        features, labels, input_hooks = (
-            self._get_features_and_labels_from_input_fn(
-                input_fn, model_fn_lib.ModeKeys.TRAIN))
-        worker_hooks.extend(input_hooks)
-        global_step_tensor = self._create_and_assert_global_step(g)
-        # we want to add to the global collection in the main thread not the
-        # tower threads.
-        ops.add_to_collection(training_util.GLOBAL_STEP_READ_KEY,
-                              self._distribution.read_var(global_step_tensor))
-        grouped_estimator_spec = self._distribution.call_for_each_tower(
-            self._call_model_fn,
-            features,
-            labels,  # although this will be None it seems
-            model_fn_lib.ModeKeys.TRAIN,
-            self.config)
+
+        if is_tpu_strategy:
+          # Create the iterator for run_on_dataset function
+          # TODO(sourabhbajaj): refactor this out to call a function on the
+          # strategy
+          dataset = self._distribution.distribute_dataset(
+              lambda: self._call_input_fn(input_fn,  # pylint: disable=g-long-lambda
+                                          model_fn_lib.ModeKeys.TRAIN))
+          iterator = dataset.make_initializable_iterator()
+          worker_hooks.append(
+              estimator_util._DatasetInitializerHook(iterator))  # pylint: disable=protected-access
+
+          global_step_tensor = self._create_and_assert_global_step(g)
+          # we want to add to the global collection in the main thread not the
+          # tower threads.
+          ops.add_to_collection(training_util.GLOBAL_STEP_READ_KEY,
+                                self._distribution.read_var(global_step_tensor))
+
+          # Create a step_fn from the train_op of grouped_estimator_spec
+          def step_fn(ctx, inputs):
+            """A single step that is passed to run_on_dataset."""
+            features, labels = inputs
+            estimator_spec = self._distribution.call_for_each_tower(
+                self._call_model_fn,
+                features,
+                labels,
+                model_fn_lib.ModeKeys.TRAIN,
+                self.config)
+            ctx.last_step_outputs = estimator_spec.loss
+            ctx.non_tensor_outputs = {'estimator_spec': estimator_spec}
+            with ops.control_dependencies([estimator_spec.train_op]):
+              return array_ops.identity(estimator_spec.loss)
+
+          # Create new train_op post graph rewrites
+          # TODO(sourabhbajaj): Make sure train_steps and tpu_iterations
+          # work correctly. Currently hardcoded at 2
+          initial_training_loss = constant_op.constant(1e7)
+          distributed_train_op, tpu_result, ctx = \
+              self._distribution._run_steps_on_dataset(  # pylint: disable=protected-access
+                  step_fn, iterator, iterations=2,
+                  initial_loop_values=initial_training_loss)
+          grouped_estimator_spec = ctx.non_tensor_outputs['estimator_spec']
+        else:
+          features, labels, input_hooks = (
+              self._get_features_and_labels_from_input_fn(
+                  input_fn, model_fn_lib.ModeKeys.TRAIN))
+          worker_hooks.extend(input_hooks)
+          global_step_tensor = self._create_and_assert_global_step(g)
+          # we want to add to the global collection in the main thread not the
+          # tower threads.
+          ops.add_to_collection(training_util.GLOBAL_STEP_READ_KEY,
+                                self._distribution.read_var(global_step_tensor))
+          grouped_estimator_spec = self._distribution.call_for_each_tower(
+              self._call_model_fn,
+              features,
+              labels,  # although this will be None it seems
+              model_fn_lib.ModeKeys.TRAIN,
+              self.config)
 
         # TODO(anjalisridhar): Figure out how to resolve the following scaffold
         # parameters: init_feed_dict, init_fn.
@@ -1278,13 +1338,28 @@ class Estimator(object):
         training_chief_hooks = get_hooks_from_the_first_device(
             grouped_estimator_spec.training_chief_hooks)
 
+        # TODO(sourabhbajaj): Merge the two code paths once we can
+        # handle per device variables correctly in reduce and can output
+        # the loss scaler.
+        if is_tpu_strategy:
+          loss = self._distribution.unwrap(
+              self._distribution.reduce(distribute_lib.get_loss_reduction(),
+                                        tpu_result)[0])[0]
+          worker_hooks.append(
+              estimator_util.StrategyInitFinalizeHook(
+                  self._distribution.get_initialization_ops,
+                  self._distribution.get_finalize_ops))
+        else:
+          loss = self._distribution.unwrap(
+              self._distribution.reduce(distribute_lib.get_loss_reduction(),
+                                        grouped_estimator_spec.loss,
+                                        destinations='/device:CPU:0'))[0]
+          distributed_train_op = grouped_estimator_spec.train_op
+
         estimator_spec = model_fn_lib.EstimatorSpec(
             mode=grouped_estimator_spec.mode,
-            loss=self._distribution.unwrap(
-                self._distribution.reduce(distribute_lib.get_loss_reduction(),
-                                          grouped_estimator_spec.loss,
-                                          destinations='/device:CPU:0'))[0],
-            train_op=self._distribution.group(grouped_estimator_spec.train_op),
+            loss=loss,
+            train_op=self._distribution.group(distributed_train_op),
             training_hooks=training_hooks,
             training_chief_hooks=training_chief_hooks,
             scaffold=scaffold)
diff --git a/tensorflow/python/estimator/estimator_test.py b/tensorflow/python/estimator/estimator_test.py
index 733c7fb95dd2035a5cb63bcf37c06905ebdd24fb..2a0e4e761755e272a316ce2d326b0c0a51ecbaba 100644
--- a/tensorflow/python/estimator/estimator_test.py
+++ b/tensorflow/python/estimator/estimator_test.py
@@ -38,6 +38,7 @@ from tensorflow.python.estimator.export import export_output
 from tensorflow.python.estimator.inputs import numpy_io
 from tensorflow.python.framework import constant_op
 from tensorflow.python.framework import dtypes
+from tensorflow.python.framework import errors
 from tensorflow.python.framework import ops
 from tensorflow.python.framework import tensor_util
 from tensorflow.python.framework import test_util
@@ -1296,6 +1297,31 @@ class EstimatorEvaluateTest(test.TestCase):
         dummy_input_fn, steps=1, checkpoint_path=est1.latest_checkpoint())
     self.assertEqual(5, scores['global_step'])
 
+  def test_wrong_shape_throws_reasonable_error(self):
+    """Make sure we are helpful when model_fns change. See b/110263146."""
+    def _get_model_fn(val=1):
+      def _model_fn(features, labels, mode):
+        del features, labels  # unused
+        variables.Variable(val, name='weight')
+        return model_fn_lib.EstimatorSpec(
+            mode=mode,
+            predictions=constant_op.constant([[1.]]),
+            loss=constant_op.constant(0.),
+            train_op=state_ops.assign_add(training.get_global_step(), 1))
+      return _model_fn
+
+    model_fn_1 = _get_model_fn()
+    model_fn_2 = _get_model_fn(val=[1])
+
+    est1 = estimator.Estimator(model_fn=model_fn_1)
+    est1.train(dummy_input_fn, steps=5)
+    est2 = estimator.Estimator(
+        model_fn=model_fn_2, model_dir=est1.model_dir)
+
+    expected_msg = 'Restoring from checkpoint failed.*a mismatch between'
+    with self.assertRaisesRegexp(errors.InvalidArgumentError, expected_msg):
+      est2.train(dummy_input_fn, steps=1,)
+
   def test_scaffold_is_used(self):
 
     def _model_fn_scaffold(features, labels, mode):
diff --git a/tensorflow/python/estimator/inputs/pandas_io.py b/tensorflow/python/estimator/inputs/pandas_io.py
index 57f8e5fd6aff366ad1f574d5dd40a6c457966b52..616bcb410f8119e170e991f8320c5b6448ee85c9 100644
--- a/tensorflow/python/estimator/inputs/pandas_io.py
+++ b/tensorflow/python/estimator/inputs/pandas_io.py
@@ -18,6 +18,8 @@
 from __future__ import absolute_import
 from __future__ import division
 from __future__ import print_function
+import six
+import uuid
 
 import numpy as np
 from tensorflow.python.estimator.inputs.queues import feeding_functions
@@ -35,6 +37,22 @@ except ImportError:
   HAS_PANDAS = False
 
 
+def _get_unique_target_key(features, target_column_name):
+  """Returns a key that does not exist in the input DataFrame `features`.
+
+  Args:
+    features: DataFrame
+    target_column_name: Name of the target column as a `str`
+
+  Returns:
+    A unique key that can be used to insert the target into
+      features.
+  """
+  if target_column_name in features:
+    target_column_name += '_' + str(uuid.uuid4())
+  return target_column_name
+
+
 @estimator_export('estimator.inputs.pandas_input_fn')
 def pandas_input_fn(x,
                     y=None,
@@ -50,7 +68,7 @@ def pandas_input_fn(x,
 
   Args:
     x: pandas `DataFrame` object.
-    y: pandas `Series` object. `None` if absent.
+    y: pandas `Series` object or `DataFrame`. `None` if absent.
     batch_size: int, size of batches to return.
     num_epochs: int, number of epochs to iterate over data. If not `None`,
       read attempts that would exceed this value will raise `OutOfRangeError`.
@@ -60,7 +78,8 @@ def pandas_input_fn(x,
     num_threads: Integer, number of threads used for reading and enqueueing. In
       order to have predicted and repeatable order of reading and enqueueing,
       such as in prediction and evaluation mode, `num_threads` should be 1.
-    target_column: str, name to give the target column `y`.
+    target_column: str, name to give the target column `y`. This parameter
+      is not used when `y` is a `DataFrame`.
 
   Returns:
     Function, that has signature of ()->(dict of `features`, `target`)
@@ -79,6 +98,9 @@ def pandas_input_fn(x,
                      '(it is recommended to set it as True for training); '
                      'got {}'.format(shuffle))
 
+  if not isinstance(target_column, six.string_types):
+    raise TypeError('target_column must be a string type')
+
   x = x.copy()
   if y is not None:
     if target_column in x:
@@ -88,7 +110,13 @@ def pandas_input_fn(x,
     if not np.array_equal(x.index, y.index):
       raise ValueError('Index for x and y are mismatched.\nIndex for x: %s\n'
                        'Index for y: %s\n' % (x.index, y.index))
-    x[target_column] = y
+    if isinstance(y, pd.DataFrame):
+      y_columns = [(column, _get_unique_target_key(x, column))
+                   for column in list(y)]
+      target_column = [v for _, v in y_columns]
+      x[target_column] = y
+    else:
+      x[target_column] = y
 
   # TODO(mdan): These are memory copies. We probably don't need 4x slack space.
   # The sizes below are consistent with what I've seen elsewhere.
@@ -118,7 +146,12 @@ def pandas_input_fn(x,
     features = features[1:]
     features = dict(zip(list(x.columns), features))
     if y is not None:
-      target = features.pop(target_column)
+      if isinstance(target_column, list):
+        keys = [k for k, _ in y_columns]
+        values = [features.pop(column) for column in target_column]
+        target = {k: v for k, v in zip(keys, values)}
+      else:
+        target = features.pop(target_column)
       return features, target
     return features
   return input_fn
diff --git a/tensorflow/python/estimator/inputs/pandas_io_test.py b/tensorflow/python/estimator/inputs/pandas_io_test.py
index dcecf6dd61c4d24a36b2be8f054c066050d088fc..6f13bc95d2d315ad1aabfd89d5d479d65fe08502 100644
--- a/tensorflow/python/estimator/inputs/pandas_io_test.py
+++ b/tensorflow/python/estimator/inputs/pandas_io_test.py
@@ -47,6 +47,16 @@ class PandasIoTest(test.TestCase):
     y = pd.Series(np.arange(-32, -28), index=index)
     return x, y
 
+  def makeTestDataFrameWithYAsDataFrame(self):
+    index = np.arange(100, 104)
+    a = np.arange(4)
+    b = np.arange(32, 36)
+    a_label = np.arange(10, 14)
+    b_label = np.arange(50, 54)
+    x = pd.DataFrame({'a': a, 'b': b}, index=index)
+    y = pd.DataFrame({'a_target': a_label, 'b_target': b_label}, index=index)
+    return x, y
+
   def callInputFnOnce(self, input_fn, session):
     results = input_fn()
     coord = coordinator.Coordinator()
@@ -65,6 +75,19 @@ class PandasIoTest(test.TestCase):
       pandas_io.pandas_input_fn(
           x, y_noindex, batch_size=2, shuffle=False, num_epochs=1)
 
+  def testPandasInputFn_RaisesWhenTargetColumnIsAList(self):
+    if not HAS_PANDAS:
+      return
+
+    x, y = self.makeTestDataFrame()
+
+    with self.assertRaisesRegexp(TypeError,
+                                 'target_column must be a string type'):
+      pandas_io.pandas_input_fn(x, y, batch_size=2,
+                                shuffle=False,
+                                num_epochs=1,
+                                target_column=['one', 'two'])
+
   def testPandasInputFn_NonBoolShuffle(self):
     if not HAS_PANDAS:
       return
@@ -90,6 +113,53 @@ class PandasIoTest(test.TestCase):
       self.assertAllEqual(features['b'], [32, 33])
       self.assertAllEqual(target, [-32, -31])
 
+  def testPandasInputFnWhenYIsDataFrame_ProducesExpectedOutput(self):
+    if not HAS_PANDAS:
+      return
+    with self.test_session() as session:
+      x, y = self.makeTestDataFrameWithYAsDataFrame()
+      input_fn = pandas_io.pandas_input_fn(
+          x, y, batch_size=2, shuffle=False, num_epochs=1)
+
+      features, targets = self.callInputFnOnce(input_fn, session)
+
+      self.assertAllEqual(features['a'], [0, 1])
+      self.assertAllEqual(features['b'], [32, 33])
+      self.assertAllEqual(targets['a_target'], [10, 11])
+      self.assertAllEqual(targets['b_target'], [50, 51])
+
+  def testPandasInputFnYIsDataFrame_HandlesOverlappingColumns(self):
+    if not HAS_PANDAS:
+      return
+    with self.test_session() as session:
+      x, y = self.makeTestDataFrameWithYAsDataFrame()
+      y = y.rename(columns={'a_target': 'a', 'b_target': 'b'})
+      input_fn = pandas_io.pandas_input_fn(
+          x, y, batch_size=2, shuffle=False, num_epochs=1)
+
+      features, targets = self.callInputFnOnce(input_fn, session)
+
+      self.assertAllEqual(features['a'], [0, 1])
+      self.assertAllEqual(features['b'], [32, 33])
+      self.assertAllEqual(targets['a'], [10, 11])
+      self.assertAllEqual(targets['b'], [50, 51])
+
+  def testPandasInputFnYIsDataFrame_HandlesOverlappingColumnsInTargets(self):
+    if not HAS_PANDAS:
+      return
+    with self.test_session() as session:
+      x, y = self.makeTestDataFrameWithYAsDataFrame()
+      y = y.rename(columns={'a_target': 'a', 'b_target': 'a_n'})
+      input_fn = pandas_io.pandas_input_fn(
+          x, y, batch_size=2, shuffle=False, num_epochs=1)
+
+      features, targets = self.callInputFnOnce(input_fn, session)
+
+      self.assertAllEqual(features['a'], [0, 1])
+      self.assertAllEqual(features['b'], [32, 33])
+      self.assertAllEqual(targets['a'], [10, 11])
+      self.assertAllEqual(targets['a_n'], [50, 51])
+
   def testPandasInputFn_ProducesOutputsForLargeBatchAndMultipleEpochs(self):
     if not HAS_PANDAS:
       return
diff --git a/tensorflow/python/estimator/keras.py b/tensorflow/python/estimator/keras.py
index 408752d3601b7c0679de3ba03fc172a50d7cbd13..076359b503b37b0088282c941199257432ca1230 100644
--- a/tensorflow/python/estimator/keras.py
+++ b/tensorflow/python/estimator/keras.py
@@ -39,12 +39,13 @@ from tensorflow.python.keras.utils.generic_utils import CustomObjectScope
 from tensorflow.python.ops import check_ops
 from tensorflow.python.ops import math_ops
 from tensorflow.python.ops import metrics as metrics_module
-from tensorflow.python.ops import variables as variables_module
 from tensorflow.python.platform import tf_logging as logging
 from tensorflow.python.saved_model import signature_constants
 from tensorflow.python.training import distribute as distribute_lib
 from tensorflow.python.training import saver as saver_lib
 from tensorflow.python.training import training_util
+from tensorflow.python.training.checkpointable import base as checkpointable
+from tensorflow.python.training.checkpointable import data_structures
 
 
 _DEFAULT_SERVING_KEY = signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY
@@ -69,16 +70,22 @@ def _convert_tensor(x):
   return x
 
 
-def _any_variable_initialized():
-  """Check if any variable has been initialized in the Keras model.
+def _any_weight_initialized(keras_model):
+  """Check if any weights has been initialized in the Keras model.
+
+  Args:
+    keras_model: An instance of compiled keras model.
 
   Returns:
-    boolean, True if at least one variable has been initialized, else False.
+    boolean, True if at least one weight has been initialized, else False.
+    Currently keras initialize all weights at get_session().
   """
-  variables = variables_module.global_variables()
-  for v in variables:
-    if getattr(v, '_keras_initialized', False):
-      return True
+  if keras_model is None:
+    return False
+  for layer in keras_model.layers:
+    for weight in layer.weights:
+      if hasattr(weight, '_keras_initialized'):
+        return True
   return False
 
 
@@ -122,8 +129,8 @@ def _create_ordered_io(keras_model, estimator_io, is_input=True):
             'It needs to match one '
             'of the following: %s' % ('input' if is_input else 'output', key,
                                       ', '.join(keras_io_names)))
-      tensors = [_convert_tensor(estimator_io[io_name])
-                 for io_name in keras_io_names]
+    tensors = [_convert_tensor(estimator_io[io_name])
+               for io_name in keras_io_names]
     return tensors
   else:
     # Plain array.
@@ -241,8 +248,17 @@ def _in_place_subclassed_model_state_restoration(model):
   # Restore layers and build attributes
   if (hasattr(model, '_original_attributes_cache') and
       model._original_attributes_cache is not None):
-    model._layers = []
+    # Models have sticky attribute assignment, so we want to be careful to add
+    # back the previous attributes and track Layers by their original names
+    # without adding dependencies on "utility" attributes which Models exempt
+    # when they're constructed.
+    model._layers = data_structures.NoDependency([])
     for name, value in model._original_attributes_cache.items():
+      if not isinstance(value, checkpointable.CheckpointableBase):
+        # If this value is not already checkpointable, it's probably that way
+        # for a reason; we don't want to start tracking data structures that the
+        # original Model didn't.
+        value = data_structures.NoDependency(value)
       setattr(model, name, value)
     model._original_attributes_cache = None
   else:
@@ -509,7 +525,7 @@ def model_to_estimator(keras_model=None,
       keras_model_fn, model_dir=model_dir, config=config)
 
   # Check if we need to call get_weights:
-  if _any_variable_initialized():
+  if _any_weight_initialized(keras_model):
     keras_weights = keras_model.get_weights()
     # Warn if config passed to estimator tries to update GPUOptions. If a
     # session has already been created, the GPUOptions passed to the first
diff --git a/tensorflow/python/estimator/keras_test.py b/tensorflow/python/estimator/keras_test.py
index 5e094ae92bcf88a48d7afe3fb88bbced4971b587..7a4457f5a491357ee9ee5176c80416ab9eb2a73b 100644
--- a/tensorflow/python/estimator/keras_test.py
+++ b/tensorflow/python/estimator/keras_test.py
@@ -204,6 +204,7 @@ class TestKerasEstimator(test_util.TensorFlowTestCase):
       writer_cache.FileWriterCache.clear()
       gfile.DeleteRecursively(self._config.model_dir)
 
+  @test_util.run_in_graph_and_eager_modes
   def test_train_with_tf_optimizer(self):
     for model_type in ['sequential', 'functional']:
       keras_model, (_, _), (
@@ -231,6 +232,7 @@ class TestKerasEstimator(test_util.TensorFlowTestCase):
       writer_cache.FileWriterCache.clear()
       gfile.DeleteRecursively(self._config.model_dir)
 
+  @test_util.run_in_graph_and_eager_modes
   def test_train_with_subclassed_model(self):
     keras_model, (_, _), (
         _, _), train_input_fn, eval_input_fn = get_resource_for_simple_model(
diff --git a/tensorflow/python/estimator/run_config.py b/tensorflow/python/estimator/run_config.py
index 3d60c63b68968c98a00364948bd3de0581daadd4..aa594af2e4261cbd4356496376c9a74207b72ddd 100644
--- a/tensorflow/python/estimator/run_config.py
+++ b/tensorflow/python/estimator/run_config.py
@@ -485,7 +485,16 @@ class RunConfig(object):
 
     self._init_distributed_setting_from_environment_var(tf_config)
 
-    # Get session_config only for distributed mode (cluster_spec is present).
+    self._maybe_overwrite_session_config_for_distributed_training()
+
+  def _maybe_overwrite_session_config_for_distributed_training(self):
+    """Overwrites the session_config for distributed training.
+
+    The default overwrite is optimized for between-graph training. Subclass
+    should override this method if necessary.
+    """
+    # Get session_config only for between-graph distributed mode (cluster_spec
+    # is present).
     if not self._session_config and self._cluster_spec:
       RunConfig._replace(
           self,
diff --git a/tensorflow/python/estimator/training.py b/tensorflow/python/estimator/training.py
index 37b123217a97a64db65a956d9cefb1948a03f9ef..f5ac79ced2c276305e4cef038f6ca48bdcc00af6 100644
--- a/tensorflow/python/estimator/training.py
+++ b/tensorflow/python/estimator/training.py
@@ -278,10 +278,7 @@ def train_and_evaluate(estimator, train_spec, eval_spec):
   supported distributed training configuration is between-graph replication.
 
   Overfitting: In order to avoid overfitting, it is recommended to set up the
-  training `input_fn` to shuffle the training data properly. It is also
-  recommended to train the model a little longer, say multiple epochs, before
-  performing evaluation, as the input pipeline starts from scratch for each
-  training. It is particularly important for local training and evaluation.
+  training `input_fn` to shuffle the training data properly.
 
   Stop condition: In order to support both distributed and non-distributed
   configuration reliably, the only supported stop condition for model
@@ -315,10 +312,10 @@ def train_and_evaluate(estimator, train_spec, eval_spec):
   #       hidden_units=[1024, 512, 256])
 
   # Input pipeline for train and evaluate.
-  def train_input_fn: # returns x, y
+  def train_input_fn(): # returns x, y
     # please shuffle the data.
     pass
-  def eval_input_fn_eval: # returns x, y
+  def eval_input_fn(): # returns x, y
     pass
 
   train_spec = tf.estimator.TrainSpec(input_fn=train_input_fn, max_steps=1000)
diff --git a/tensorflow/python/estimator/util.py b/tensorflow/python/estimator/util.py
index 924ca309ff0455d3bb06be61ce65bb0a61e84fb0..d4a75478d53f5b3dc8e66df98a78b51a6d25aab8 100644
--- a/tensorflow/python/estimator/util.py
+++ b/tensorflow/python/estimator/util.py
@@ -22,6 +22,7 @@ from __future__ import print_function
 import os
 import time
 
+from tensorflow.core.protobuf import config_pb2
 from tensorflow.python.platform import gfile
 from tensorflow.python.platform import tf_logging as logging
 from tensorflow.python.training import training
@@ -129,3 +130,24 @@ class _DatasetInitializerHook(training.SessionRunHook):
   def after_create_session(self, session, coord):
     del coord
     session.run(self._initializer)
+
+
+class StrategyInitFinalizeHook(training.SessionRunHook):
+  """Creates a SessionRunHook that initializes and shutsdown devices."""
+
+  def __init__(self, initialization_fn, finalize_fn):
+    self._initialization_fn = initialization_fn
+    self._finalize_fn = finalize_fn
+
+  def begin(self):
+    self._init_ops = self._initialization_fn()
+    self._finalize_ops = self._finalize_fn()
+
+  def after_create_session(self, session, coord):
+    logging.info('Initialize system')
+    session.run(self._init_ops,
+                options=config_pb2.RunOptions(timeout_in_ms=5 * 60 * 1000))
+
+  def end(self, session):
+    logging.info('Finalize system.')
+    session.run(self._finalize_ops)
diff --git a/tensorflow/python/feature_column/BUILD b/tensorflow/python/feature_column/BUILD
index 295d4ca094cc8cb85c0f1f7fd47c20b910c270df..80707030e6eb3c423a1b8ae38624ddad3e87fb04 100644
--- a/tensorflow/python/feature_column/BUILD
+++ b/tensorflow/python/feature_column/BUILD
@@ -48,6 +48,39 @@ py_library(
     ],
 )
 
+py_library(
+    name = "feature_column_v2",
+    srcs = ["feature_column_v2.py"],
+    srcs_version = "PY2AND3",
+    deps = [
+        "//tensorflow/python:array_ops",
+        "//tensorflow/python:check_ops",
+        "//tensorflow/python:control_flow_ops",
+        "//tensorflow/python:dtypes",
+        "//tensorflow/python:embedding_ops",
+        "//tensorflow/python:framework_ops",
+        "//tensorflow/python:init_ops",
+        "//tensorflow/python:lookup_ops",
+        "//tensorflow/python:math_ops",
+        "//tensorflow/python:nn_ops",
+        "//tensorflow/python:parsing_ops",
+        "//tensorflow/python:platform",
+        "//tensorflow/python:resource_variable_ops",
+        "//tensorflow/python:sparse_ops",
+        "//tensorflow/python:sparse_tensor",
+        "//tensorflow/python:string_ops",
+        "//tensorflow/python:template",
+        "//tensorflow/python:tensor_shape",
+        "//tensorflow/python:training",
+        "//tensorflow/python:util",
+        "//tensorflow/python:variable_scope",
+        "//tensorflow/python:variables",
+        "//tensorflow/python/keras",
+        "//third_party/py/numpy",
+        "@six_archive//:six",
+    ],
+)
+
 filegroup(
     name = "vocabulary_testdata",
     srcs = [
@@ -92,3 +125,38 @@ py_test(
         "//tensorflow/python/estimator:numpy_io",
     ],
 )
+
+py_test(
+    name = "feature_column_v2_test",
+    srcs = ["feature_column_v2_test.py"],
+    data = [":vocabulary_testdata"],
+    srcs_version = "PY2AND3",
+    tags = [
+        "no_cuda_on_cpu_tap",
+        "no_pip",
+    ],
+    deps = [
+        ":feature_column_py",
+        ":feature_column_v2",
+        "//tensorflow/core:protos_all_py",
+        "//tensorflow/python:array_ops",
+        "//tensorflow/python:client_testlib",
+        "//tensorflow/python:constant_op",
+        "//tensorflow/python:dtypes",
+        "//tensorflow/python:errors",
+        "//tensorflow/python:framework_ops",
+        "//tensorflow/python:framework_test_lib",
+        "//tensorflow/python:lookup_ops",
+        "//tensorflow/python:parsing_ops",
+        "//tensorflow/python:partitioned_variables",
+        "//tensorflow/python:session",
+        "//tensorflow/python:sparse_tensor",
+        "//tensorflow/python:training",
+        "//tensorflow/python:variable_scope",
+        "//tensorflow/python:variables",
+        "//tensorflow/python/eager:backprop",
+        "//tensorflow/python/eager:context",
+        "//tensorflow/python/estimator:numpy_io",
+        "//third_party/py/numpy",
+    ],
+)
diff --git a/tensorflow/python/feature_column/feature_column.py b/tensorflow/python/feature_column/feature_column.py
index 40219e4b342de8e69f0b45f32a1f7b3eccfa3b80..d091d2fe0ac688773b27d80f37fbf3083b8ffa1f 100644
--- a/tensorflow/python/feature_column/feature_column.py
+++ b/tensorflow/python/feature_column/feature_column.py
@@ -2158,7 +2158,7 @@ def _create_categorical_column_weighted_sum(column,
         initializer=init_ops.zeros_initializer(),
         trainable=trainable,
         collections=weight_collections)
-  return _safe_embedding_lookup_sparse(
+  return embedding_ops.safe_embedding_lookup_sparse(
       weight,
       id_tensor,
       sparse_weights=weight_tensor,
@@ -2594,7 +2594,7 @@ class _EmbeddingColumn(
       })
 
     # Return embedding lookup result.
-    return _safe_embedding_lookup_sparse(
+    return embedding_ops.safe_embedding_lookup_sparse(
         embedding_weights=embedding_weights,
         sparse_ids=sparse_ids,
         sparse_weights=sparse_weights,
@@ -2736,7 +2736,7 @@ class _SharedEmbeddingColumn(
         })
 
       # Return embedding lookup result.
-      return _safe_embedding_lookup_sparse(
+      return embedding_ops.safe_embedding_lookup_sparse(
           embedding_weights=embedding_weights,
           sparse_ids=sparse_ids,
           sparse_weights=sparse_weights,
@@ -3228,161 +3228,6 @@ def _collect_leaf_level_keys(cross):
   return leaf_level_keys
 
 
-# TODO(zakaria): Move this to embedding_ops and make it public.
-def _safe_embedding_lookup_sparse(embedding_weights,
-                                  sparse_ids,
-                                  sparse_weights=None,
-                                  combiner='mean',
-                                  default_id=None,
-                                  name=None,
-                                  partition_strategy='div',
-                                  max_norm=None):
-  """Lookup embedding results, accounting for invalid IDs and empty features.
-
-  The partitioned embedding in `embedding_weights` must all be the same shape
-  except for the first dimension. The first dimension is allowed to vary as the
-  vocabulary size is not necessarily a multiple of `P`.  `embedding_weights`
-  may be a `PartitionedVariable` as returned by using `tf.get_variable()` with a
-  partitioner.
-
-  Invalid IDs (< 0) are pruned from input IDs and weights, as well as any IDs
-  with non-positive weight. For an entry with no features, the embedding vector
-  for `default_id` is returned, or the 0-vector if `default_id` is not supplied.
-
-  The ids and weights may be multi-dimensional. Embeddings are always aggregated
-  along the last dimension.
-
-  Args:
-    embedding_weights:  A list of `P` float `Tensor`s or values representing
-        partitioned embedding `Tensor`s.  Alternatively, a `PartitionedVariable`
-        created by partitioning along dimension 0.  The total unpartitioned
-        shape should be `[e_0, e_1, ..., e_m]`, where `e_0` represents the
-        vocab size and `e_1, ..., e_m` are the embedding dimensions.
-    sparse_ids: `SparseTensor` of shape `[d_0, d_1, ..., d_n]` containing the
-        ids. `d_0` is typically batch size.
-    sparse_weights: `SparseTensor` of same shape as `sparse_ids`, containing
-        float weights corresponding to `sparse_ids`, or `None` if all weights
-        are be assumed to be 1.0.
-    combiner: A string specifying how to combine embedding results for each
-        entry. Currently "mean", "sqrtn" and "sum" are supported, with "mean"
-        the default.
-    default_id: The id to use for an entry with no features.
-    name: A name for this operation (optional).
-    partition_strategy: A string specifying the partitioning strategy.
-        Currently `"div"` and `"mod"` are supported. Default is `"div"`.
-    max_norm: If not `None`, all embeddings are l2-normalized to max_norm before
-        combining.
-
-
-  Returns:
-    Dense `Tensor` of shape `[d_0, d_1, ..., d_{n-1}, e_1, ..., e_m]`.
-
-  Raises:
-    ValueError: if `embedding_weights` is empty.
-  """
-  if embedding_weights is None:
-    raise ValueError('Missing embedding_weights %s.' % embedding_weights)
-  if isinstance(embedding_weights, variables.PartitionedVariable):
-    embedding_weights = list(embedding_weights)  # get underlying Variables.
-  if not isinstance(embedding_weights, list):
-    embedding_weights = [embedding_weights]
-  if len(embedding_weights) < 1:
-    raise ValueError('Missing embedding_weights %s.' % embedding_weights)
-
-  dtype = sparse_weights.dtype if sparse_weights is not None else None
-  embedding_weights = [
-      ops.convert_to_tensor(w, dtype=dtype) for w in embedding_weights
-  ]
-
-  with ops.name_scope(name, 'embedding_lookup',
-                      embedding_weights + [sparse_ids,
-                                           sparse_weights]) as scope:
-    # Reshape higher-rank sparse ids and weights to linear segment ids.
-    original_shape = sparse_ids.dense_shape
-    original_rank_dim = sparse_ids.dense_shape.get_shape()[0]
-    original_rank = (
-        array_ops.size(original_shape)
-        if original_rank_dim.value is None
-        else original_rank_dim.value)
-    sparse_ids = sparse_ops.sparse_reshape(sparse_ids, [
-        math_ops.reduce_prod(
-            array_ops.slice(original_shape, [0], [original_rank - 1])),
-        array_ops.gather(original_shape, original_rank - 1)])
-    if sparse_weights is not None:
-      sparse_weights = sparse_tensor_lib.SparseTensor(
-          sparse_ids.indices,
-          sparse_weights.values, sparse_ids.dense_shape)
-
-    # Prune invalid ids and weights.
-    sparse_ids, sparse_weights = _prune_invalid_ids(sparse_ids, sparse_weights)
-    if combiner != 'sum':
-      sparse_ids, sparse_weights = _prune_invalid_weights(
-          sparse_ids, sparse_weights)
-
-    # Fill in dummy values for empty features, if necessary.
-    sparse_ids, is_row_empty = sparse_ops.sparse_fill_empty_rows(sparse_ids,
-                                                                 default_id or
-                                                                 0)
-    if sparse_weights is not None:
-      sparse_weights, _ = sparse_ops.sparse_fill_empty_rows(sparse_weights, 1.0)
-
-    result = embedding_ops.embedding_lookup_sparse(
-        embedding_weights,
-        sparse_ids,
-        sparse_weights,
-        combiner=combiner,
-        partition_strategy=partition_strategy,
-        name=None if default_id is None else scope,
-        max_norm=max_norm)
-
-    if default_id is None:
-      # Broadcast is_row_empty to the same shape as embedding_lookup_result,
-      # for use in Select.
-      is_row_empty = array_ops.tile(
-          array_ops.reshape(is_row_empty, [-1, 1]),
-          array_ops.stack([1, array_ops.shape(result)[1]]))
-
-      result = array_ops.where(is_row_empty,
-                               array_ops.zeros_like(result),
-                               result,
-                               name=scope)
-
-    # Reshape back from linear ids back into higher-dimensional dense result.
-    final_result = array_ops.reshape(
-        result,
-        array_ops.concat([
-            array_ops.slice(
-                math_ops.cast(original_shape, dtypes.int32), [0],
-                [original_rank - 1]),
-            array_ops.slice(array_ops.shape(result), [1], [-1])
-        ], 0))
-    final_result.set_shape(tensor_shape.unknown_shape(
-        (original_rank_dim - 1).value).concatenate(result.get_shape()[1:]))
-    return final_result
-
-
-def _prune_invalid_ids(sparse_ids, sparse_weights):
-  """Prune invalid IDs (< 0) from the input ids and weights."""
-  is_id_valid = math_ops.greater_equal(sparse_ids.values, 0)
-  if sparse_weights is not None:
-    is_id_valid = math_ops.logical_and(
-        is_id_valid,
-        array_ops.ones_like(sparse_weights.values, dtype=dtypes.bool))
-  sparse_ids = sparse_ops.sparse_retain(sparse_ids, is_id_valid)
-  if sparse_weights is not None:
-    sparse_weights = sparse_ops.sparse_retain(sparse_weights, is_id_valid)
-  return sparse_ids, sparse_weights
-
-
-def _prune_invalid_weights(sparse_ids, sparse_weights):
-  """Prune invalid weights (< 0) from the input ids and weights."""
-  if sparse_weights is not None:
-    is_weights_valid = math_ops.greater(sparse_weights.values, 0)
-    sparse_ids = sparse_ops.sparse_retain(sparse_ids, is_weights_valid)
-    sparse_weights = sparse_ops.sparse_retain(sparse_weights, is_weights_valid)
-  return sparse_ids, sparse_weights
-
-
 class _IndicatorColumn(_DenseColumn, _SequenceDenseColumn,
                        collections.namedtuple('_IndicatorColumn',
                                               ['categorical_column'])):
@@ -3419,10 +3264,14 @@ class _IndicatorColumn(_DenseColumn, _SequenceDenseColumn,
           sp_ids=id_tensor,
           sp_values=weight_tensor,
           vocab_size=int(self._variable_shape[-1]))
-      # Remove (?, -1) index
+      # Remove (?, -1) index.
       weighted_column = sparse_ops.sparse_slice(weighted_column, [0, 0],
                                                 weighted_column.dense_shape)
-      return sparse_ops.sparse_tensor_to_dense(weighted_column)
+      # Use scatter_nd to merge duplicated indices if existed,
+      # instead of sparse_tensor_to_dense.
+      return array_ops.scatter_nd(weighted_column.indices,
+                                  weighted_column.values,
+                                  weighted_column.dense_shape)
 
     dense_id_tensor = sparse_ops.sparse_tensor_to_dense(
         id_tensor, default_value=-1)
diff --git a/tensorflow/python/feature_column/feature_column_test.py b/tensorflow/python/feature_column/feature_column_test.py
index 511205451cdee707d80993bd37eaad395625e773..5bb47bfa47cf8fe0311d63f325198bcb7ecd5f9c 100644
--- a/tensorflow/python/feature_column/feature_column_test.py
+++ b/tensorflow/python/feature_column/feature_column_test.py
@@ -4580,12 +4580,12 @@ class IndicatorColumnTest(test.TestCase):
     weights = fc.weighted_categorical_column(ids, 'weights')
     indicator = fc.indicator_column(weights)
     features = {
-        'ids': constant_op.constant([['c', 'b', 'a']]),
-        'weights': constant_op.constant([[2., 4., 6.]])
+        'ids': constant_op.constant([['c', 'b', 'a', 'c']]),
+        'weights': constant_op.constant([[2., 4., 6., 1.]])
     }
     indicator_tensor = _transform_features(features, [indicator])[indicator]
     with _initialized_session():
-      self.assertAllEqual([[6., 4., 2.]], indicator_tensor.eval())
+      self.assertAllEqual([[6., 4., 3.]], indicator_tensor.eval())
 
   def test_transform_with_missing_value_in_weighted_column(self):
     # Github issue 12583
diff --git a/tensorflow/python/feature_column/feature_column_v2.py b/tensorflow/python/feature_column/feature_column_v2.py
new file mode 100644
index 0000000000000000000000000000000000000000..b4dd23f58de60bacae68f9b67ed30c5d4ae49b15
--- /dev/null
+++ b/tensorflow/python/feature_column/feature_column_v2.py
@@ -0,0 +1,3600 @@
+# Copyright 2017 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""This API defines FeatureColumn abstraction.
+
+FeatureColumns provide a high level abstraction for ingesting and representing
+features. FeatureColumns are also the primary way of encoding features for
+canned @{tf.estimator.Estimator}s.
+
+When using FeatureColumns with `Estimators`, the type of feature column you
+should choose depends on (1) the feature type and (2) the model type.
+
+1. Feature type:
+
+  * Continuous features can be represented by `numeric_column`.
+  * Categorical features can be represented by any `categorical_column_with_*`
+  column:
+    - `categorical_column_with_vocabulary_list`
+    - `categorical_column_with_vocabulary_file`
+    - `categorical_column_with_hash_bucket`
+    - `categorical_column_with_identity`
+    - `weighted_categorical_column`
+
+2. Model type:
+
+  * Deep neural network models (`DNNClassifier`, `DNNRegressor`).
+
+    Continuous features can be directly fed into deep neural network models.
+
+      age_column = numeric_column("age")
+
+    To feed sparse features into DNN models, wrap the column with
+    `embedding_column` or `indicator_column`. `indicator_column` is recommended
+    for features with only a few possible values. For features with many
+    possible values, to reduce the size of your model, `embedding_column` is
+    recommended.
+
+      embedded_dept_column = embedding_column(
+          categorical_column_with_vocabulary_list(
+              "department", ["math", "philosophy", ...]), dimension=10)
+
+  * Wide (aka linear) models (`LinearClassifier`, `LinearRegressor`).
+
+    Sparse features can be fed directly into linear models. They behave like an
+    indicator column but with an efficient implementation.
+
+      dept_column = categorical_column_with_vocabulary_list("department",
+          ["math", "philosophy", "english"])
+
+    It is recommended that continuous features be bucketized before being
+    fed into linear models.
+
+      bucketized_age_column = bucketized_column(
+          source_column=age_column,
+          boundaries=[18, 25, 30, 35, 40, 45, 50, 55, 60, 65])
+
+    Sparse features can be crossed (also known as conjuncted or combined) in
+    order to form non-linearities, and then fed into linear models.
+
+      cross_dept_age_column = crossed_column(
+          columns=["department", bucketized_age_column],
+          hash_bucket_size=1000)
+
+Example of building canned `Estimator`s using FeatureColumns:
+
+  ```python
+  # Define features and transformations
+  deep_feature_columns = [age_column, embedded_dept_column]
+  wide_feature_columns = [dept_column, bucketized_age_column,
+      cross_dept_age_column]
+
+  # Build deep model
+  estimator = DNNClassifier(
+      feature_columns=deep_feature_columns,
+      hidden_units=[500, 250, 50])
+  estimator.train(...)
+
+  # Or build a wide model
+  estimator = LinearClassifier(
+      feature_columns=wide_feature_columns)
+  estimator.train(...)
+
+  # Or build a wide and deep model!
+  estimator = DNNLinearCombinedClassifier(
+      linear_feature_columns=wide_feature_columns,
+      dnn_feature_columns=deep_feature_columns,
+      dnn_hidden_units=[500, 250, 50])
+  estimator.train(...)
+  ```
+
+
+FeatureColumns can also be transformed into a generic input layer for
+custom models using `input_layer`.
+
+Example of building model using FeatureColumns, this can be used in a
+`model_fn` which is given to the {tf.estimator.Estimator}:
+
+  ```python
+  # Building model via layers
+
+  deep_feature_columns = [age_column, embedded_dept_column]
+  columns_to_tensor = parse_feature_columns_from_examples(
+      serialized=my_data,
+      feature_columns=deep_feature_columns)
+  first_layer = input_layer(
+      features=columns_to_tensor,
+      feature_columns=deep_feature_columns)
+  second_layer = fully_connected(first_layer, ...)
+  ```
+
+NOTE: Functions prefixed with "_" indicate experimental or private parts of
+the API subject to change, and should not be relied upon!
+"""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import abc
+import collections
+import math
+
+import numpy as np
+import six
+
+
+from tensorflow.python.eager import context
+from tensorflow.python.feature_column import feature_column as fc_old
+from tensorflow.python.framework import dtypes
+from tensorflow.python.framework import ops
+from tensorflow.python.framework import sparse_tensor as sparse_tensor_lib
+from tensorflow.python.framework import tensor_shape
+from tensorflow.python.keras.engine import training
+from tensorflow.python.layers import base
+from tensorflow.python.ops import array_ops
+from tensorflow.python.ops import check_ops
+from tensorflow.python.ops import control_flow_ops
+from tensorflow.python.ops import embedding_ops
+from tensorflow.python.ops import init_ops
+from tensorflow.python.ops import lookup_ops
+from tensorflow.python.ops import math_ops
+from tensorflow.python.ops import nn_ops
+from tensorflow.python.ops import parsing_ops
+from tensorflow.python.ops import resource_variable_ops
+from tensorflow.python.ops import sparse_ops
+from tensorflow.python.ops import string_ops
+from tensorflow.python.ops import template
+from tensorflow.python.ops import variable_scope
+from tensorflow.python.ops import variables
+from tensorflow.python.platform import gfile
+from tensorflow.python.platform import tf_logging as logging
+from tensorflow.python.training import checkpoint_utils
+from tensorflow.python.util import nest
+
+
+def _internal_input_layer(features,
+                          feature_columns,
+                          weight_collections=None,
+                          trainable=True,
+                          cols_to_vars=None,
+                          scope=None):
+  """See input_layer. `scope` is a name or variable scope to use."""
+
+  feature_columns = fc_old._normalize_feature_columns(feature_columns)  # pylint: disable=protected-access
+  for column in feature_columns:
+    if not isinstance(column, fc_old._DenseColumn):  # pylint: disable=protected-access
+      raise ValueError(
+          'Items of feature_columns must be a _DenseColumn. '
+          'You can wrap a categorical column with an '
+          'embedding_column or indicator_column. Given: {}'.format(column))
+  weight_collections = list(weight_collections or [])
+  if ops.GraphKeys.GLOBAL_VARIABLES not in weight_collections:
+    weight_collections.append(ops.GraphKeys.GLOBAL_VARIABLES)
+  if ops.GraphKeys.MODEL_VARIABLES not in weight_collections:
+    weight_collections.append(ops.GraphKeys.MODEL_VARIABLES)
+
+  # a non-None `scope` can allow for variable reuse, when, e.g., this function
+  # is wrapped by a `make_template`.
+  with variable_scope.variable_scope(
+      scope, default_name='input_layer', values=features.values()):
+    builder = fc_old._LazyBuilder(features)  # pylint: disable=protected-access
+    output_tensors = []
+    ordered_columns = []
+    for column in sorted(feature_columns, key=lambda x: x.name):
+      ordered_columns.append(column)
+      with variable_scope.variable_scope(
+          None, default_name=column._var_scope_name):  # pylint: disable=protected-access
+        tensor = column._get_dense_tensor(  # pylint: disable=protected-access
+            builder,
+            weight_collections=weight_collections,
+            trainable=trainable)
+        num_elements = column._variable_shape.num_elements()  # pylint: disable=protected-access
+        batch_size = array_ops.shape(tensor)[0]
+        output_tensors.append(
+            array_ops.reshape(tensor, shape=(batch_size, num_elements)))
+        if cols_to_vars is not None:
+          # Retrieve any variables created (some _DenseColumn's don't create
+          # variables, in which case an empty list is returned).
+          cols_to_vars[column] = ops.get_collection(
+              ops.GraphKeys.GLOBAL_VARIABLES,
+              scope=variable_scope.get_variable_scope().name)
+    _verify_static_batch_size_equality(output_tensors, ordered_columns)
+    return array_ops.concat(output_tensors, 1)
+
+
+def input_layer(features,
+                feature_columns,
+                weight_collections=None,
+                trainable=True,
+                cols_to_vars=None):
+  """Returns a dense `Tensor` as input layer based on given `feature_columns`.
+
+  Generally a single example in training data is described with FeatureColumns.
+  At the first layer of the model, this column oriented data should be converted
+  to a single `Tensor`.
+
+  Example:
+
+  ```python
+  price = numeric_column('price')
+  keywords_embedded = embedding_column(
+      categorical_column_with_hash_bucket("keywords", 10K), dimensions=16)
+  columns = [price, keywords_embedded, ...]
+  features = tf.parse_example(..., features=make_parse_example_spec(columns))
+  dense_tensor = input_layer(features, columns)
+  for units in [128, 64, 32]:
+    dense_tensor = tf.layers.dense(dense_tensor, units, tf.nn.relu)
+  prediction = tf.layers.dense(dense_tensor, 1)
+  ```
+
+  Args:
+    features: A mapping from key to tensors. `_FeatureColumn`s look up via these
+      keys. For example `numeric_column('price')` will look at 'price' key in
+      this dict. Values can be a `SparseTensor` or a `Tensor` depends on
+      corresponding `_FeatureColumn`.
+    feature_columns: An iterable containing the FeatureColumns to use as inputs
+      to your model. All items should be instances of classes derived from
+      `_DenseColumn` such as `numeric_column`, `embedding_column`,
+      `bucketized_column`, `indicator_column`. If you have categorical features,
+      you can wrap them with an `embedding_column` or `indicator_column`.
+    weight_collections: A list of collection names to which the Variable will be
+      added. Note that variables will also be added to collections
+      `tf.GraphKeys.GLOBAL_VARIABLES` and `ops.GraphKeys.MODEL_VARIABLES`.
+    trainable: If `True` also add the variable to the graph collection
+      `GraphKeys.TRAINABLE_VARIABLES` (see `tf.Variable`).
+    cols_to_vars: If not `None`, must be a dictionary that will be filled with a
+      mapping from `_FeatureColumn` to list of `Variable`s.  For example, after
+      the call, we might have cols_to_vars =
+      {_EmbeddingColumn(
+        categorical_column=_HashedCategoricalColumn(
+          key='sparse_feature', hash_bucket_size=5, dtype=tf.string),
+        dimension=10): [],
+        'bias': [],
+        _NumericColumn(
+          key='numeric_feature2', shape=(2,)):
+        []}
+      If a column creates no variables, its value will be an empty list. Note
+      that cols_to_vars will also contain a string key 'bias' that maps to a
+      list of Variables.
+
+  Returns:
+    A `Tensor` which represents predictions/logits of a linear model. Its shape
+    is (batch_size, units) and its dtype is `float32`.
+
+  Raises:
+    ValueError: if an item in `feature_columns` is neither a `_DenseColumn`
+      nor `_CategoricalColumn`.
+  """
+  with variable_scope.variable_scope(None, 'linear_model') as vs:
+    model_name = _strip_leading_slashes(vs.name)
+  linear_model_layer = _LinearModel(
+      feature_columns=feature_columns,
+      units=units,
+      sparse_combiner=sparse_combiner,
+      weight_collections=weight_collections,
+      trainable=trainable,
+      name=model_name)
+  retval = linear_model_layer(features)  # pylint: disable=not-callable
+  if cols_to_vars is not None:
+    cols_to_vars.update(linear_model_layer.cols_to_vars())
+  return retval
+
+
+def _add_to_collections(var, weight_collections):
+  """Adds a var to the list of weight_collections provided.
+
+  Handles the case for partitioned and non-partitioned variables.
+
+  Args:
+    var: A variable or Partitioned Variable.
+    weight_collections: List of collections to add variable to.
+  """
+  for weight_collection in weight_collections:
+    # The layer self.add_variable call already adds it to GLOBAL_VARIABLES.
+    if weight_collection == ops.GraphKeys.GLOBAL_VARIABLES:
+      continue
+    # TODO(rohanj): Explore adding a _get_variable_list method on `Variable`
+    # so that we don't have to do this check.
+    if isinstance(var, variables.PartitionedVariable):
+      for constituent_var in list(var):
+        ops.add_to_collection(weight_collection, constituent_var)
+    else:
+      ops.add_to_collection(weight_collection, var)
+
+
+class _FCLinearWrapper(base.Layer):
+  """Wraps a _FeatureColumn in a layer for use in a linear model.
+
+  See `linear_model` above.
+  """
+
+  def __init__(self,
+               feature_column,
+               units=1,
+               sparse_combiner='sum',
+               weight_collections=None,
+               trainable=True,
+               name=None,
+               **kwargs):
+    super(_FCLinearWrapper, self).__init__(
+        trainable=trainable, name=name, **kwargs)
+    self._feature_column = feature_column
+    self._units = units
+    self._sparse_combiner = sparse_combiner
+    self._weight_collections = weight_collections
+
+  def build(self, _):
+    if isinstance(self._feature_column, fc_old._CategoricalColumn):  # pylint: disable=protected-access
+      weight = self.add_variable(
+          name='weights',
+          shape=(self._feature_column._num_buckets, self._units),  # pylint: disable=protected-access
+          initializer=init_ops.zeros_initializer(),
+          trainable=self.trainable)
+    else:
+      num_elements = self._feature_column._variable_shape.num_elements()  # pylint: disable=protected-access
+      weight = self.add_variable(
+          name='weights',
+          shape=[num_elements, self._units],
+          initializer=init_ops.zeros_initializer(),
+          trainable=self.trainable)
+    _add_to_collections(weight, self._weight_collections)
+    self._weight_var = weight
+    self.built = True
+
+  def call(self, builder):
+    weighted_sum = fc_old._create_weighted_sum(  # pylint: disable=protected-access
+        column=self._feature_column,
+        builder=builder,
+        units=self._units,
+        sparse_combiner=self._sparse_combiner,
+        weight_collections=self._weight_collections,
+        trainable=self.trainable,
+        weight_var=self._weight_var)
+    return weighted_sum
+
+
+class _BiasLayer(base.Layer):
+  """A layer for the bias term.
+  """
+
+  def __init__(self,
+               units=1,
+               trainable=True,
+               weight_collections=None,
+               name=None,
+               **kwargs):
+    super(_BiasLayer, self).__init__(trainable=trainable, name=name, **kwargs)
+    self._units = units
+    self._weight_collections = weight_collections
+
+  def build(self, _):
+    self._bias_variable = self.add_variable(
+        'bias_weights',
+        shape=[self._units],
+        initializer=init_ops.zeros_initializer(),
+        trainable=self.trainable)
+    _add_to_collections(self._bias_variable, self._weight_collections)
+    self.built = True
+
+  def call(self, _):
+    return self._bias_variable
+
+
+def _get_expanded_variable_list(variable):
+  if (isinstance(variable, variables.Variable) or
+      resource_variable_ops.is_resource_variable(variable)):
+    return [variable]  # Single variable case.
+  else:  # Must be a PartitionedVariable, so convert into a list.
+    return list(variable)
+
+
+def _strip_leading_slashes(name):
+  return name.rsplit('/', 1)[-1]
+
+
+class _LinearModel(training.Model):
+  """Creates a linear model using feature columns.
+
+  See `linear_model` for details.
+  """
+
+  def __init__(self,
+               feature_columns,
+               units=1,
+               sparse_combiner='sum',
+               weight_collections=None,
+               trainable=True,
+               name=None,
+               **kwargs):
+    super(_LinearModel, self).__init__(name=name, **kwargs)
+    self._feature_columns = fc_old._normalize_feature_columns(  # pylint: disable=protected-access
+        feature_columns)
+    self._weight_collections = list(weight_collections or [])
+    if ops.GraphKeys.GLOBAL_VARIABLES not in self._weight_collections:
+      self._weight_collections.append(ops.GraphKeys.GLOBAL_VARIABLES)
+    if ops.GraphKeys.MODEL_VARIABLES not in self._weight_collections:
+      self._weight_collections.append(ops.GraphKeys.MODEL_VARIABLES)
+
+    column_layers = {}
+    for column in sorted(self._feature_columns, key=lambda x: x.name):
+      with variable_scope.variable_scope(
+          None, default_name=column._var_scope_name) as vs:  # pylint: disable=protected-access
+        # Having the fully expressed variable scope name ends up doubly
+        # expressing the outer scope (scope with which this method was called)
+        # in the name of the variable that would get created.
+        column_name = _strip_leading_slashes(vs.name)
+      column_layer = _FCLinearWrapper(column, units, sparse_combiner,
+                                      self._weight_collections, trainable,
+                                      column_name, **kwargs)
+      column_layers[column_name] = column_layer
+    self._column_layers = self._add_layers(column_layers)
+    self._bias_layer = _BiasLayer(
+        units=units,
+        trainable=trainable,
+        weight_collections=self._weight_collections,
+        name='bias_layer',
+        **kwargs)
+    self._cols_to_vars = {}
+
+  def cols_to_vars(self):
+    """Returns a dict mapping _FeatureColumns to variables.
+
+    See `linear_model` for more information.
+    This is not populated till `call` is called i.e. layer is built.
+    """
+    return self._cols_to_vars
+
+  def call(self, features):
+    with variable_scope.variable_scope(self.name):
+      for column in self._feature_columns:
+        if not isinstance(
+            column,
+            (
+                fc_old._DenseColumn,  # pylint: disable=protected-access
+                fc_old._CategoricalColumn)):  # pylint: disable=protected-access
+          raise ValueError(
+              'Items of feature_columns must be either a '
+              '_DenseColumn or _CategoricalColumn. Given: {}'.format(column))
+      weighted_sums = []
+      ordered_columns = []
+      builder = fc_old._LazyBuilder(features)  # pylint: disable=protected-access
+      for layer in sorted(self._column_layers.values(), key=lambda x: x.name):
+        column = layer._feature_column  # pylint: disable=protected-access
+        ordered_columns.append(column)
+        weighted_sum = layer(builder)
+        weighted_sums.append(weighted_sum)
+        self._cols_to_vars[column] = ops.get_collection(
+            ops.GraphKeys.GLOBAL_VARIABLES, scope=layer.scope_name)
+
+      _verify_static_batch_size_equality(weighted_sums, ordered_columns)
+      predictions_no_bias = math_ops.add_n(
+          weighted_sums, name='weighted_sum_no_bias')
+      predictions = nn_ops.bias_add(
+          predictions_no_bias,
+          self._bias_layer(  # pylint: disable=not-callable
+              builder,
+              scope=variable_scope.get_variable_scope()),  # pylint: disable=not-callable
+          name='weighted_sum')
+      bias = self._bias_layer.variables[0]
+      self._cols_to_vars['bias'] = _get_expanded_variable_list(bias)
+    return predictions
+
+  def _add_layers(self, layers):
+    # "Magic" required for keras.Model classes to track all the variables in
+    # a list of layers.Layer objects.
+    # TODO(ashankar): Figure out API so user code doesn't have to do this.
+    for name, layer in layers.items():
+      setattr(self, 'layer-%s' % name, layer)
+    return layers
+
+
+def _transform_features(features, feature_columns, state_manager):
+  """Returns transformed features based on features columns passed in.
+
+  Please note that most probably you would not need to use this function. Please
+  check `input_layer` and `linear_model` to see whether they will
+  satisfy your use case or not.
+
+  Example:
+
+  ```python
+  # Define features and transformations
+  crosses_a_x_b = crossed_column(
+      columns=["sparse_feature_a", "sparse_feature_b"], hash_bucket_size=10000)
+  price_buckets = bucketized_column(
+      source_column=numeric_column("price"), boundaries=[...])
+
+  columns = [crosses_a_x_b, price_buckets]
+  features = tf.parse_example(..., features=make_parse_example_spec(columns))
+  transformed = transform_features(features=features, feature_columns=columns)
+
+  assertCountEqual(columns, transformed.keys())
+  ```
+
+  Args:
+    features: A mapping from key to tensors. `FeatureColumn`s look up via these
+      keys. For example `numeric_column('price')` will look at 'price' key in
+      this dict. Values can be a `SparseTensor` or a `Tensor` depends on
+      corresponding `FeatureColumn`.
+    feature_columns: An iterable containing all the `FeatureColumn`s.
+    state_manager: A StateManager object that holds the FeatureColumn state.
+
+  Returns:
+    A `dict` mapping `FeatureColumn` to `Tensor` and `SparseTensor` values.
+  """
+  feature_columns = _normalize_feature_columns(feature_columns)
+  outputs = {}
+  with ops.name_scope(
+      None, default_name='transform_features', values=features.values()):
+    transformation_cache = FeatureTransformationCache(features)
+    for column in sorted(feature_columns, key=lambda x: x.name):
+      with ops.name_scope(None, default_name=column.name):
+        outputs[column] = transformation_cache.get(column, state_manager)
+  return outputs
+
+
+def make_parse_example_spec(feature_columns):
+  """Creates parsing spec dictionary from input feature_columns.
+
+  The returned dictionary can be used as arg 'features' in `tf.parse_example`.
+
+  Typical usage example:
+
+  ```python
+  # Define features and transformations
+  feature_a = categorical_column_with_vocabulary_file(...)
+  feature_b = numeric_column(...)
+  feature_c_bucketized = bucketized_column(numeric_column("feature_c"), ...)
+  feature_a_x_feature_c = crossed_column(
+      columns=["feature_a", feature_c_bucketized], ...)
+
+  feature_columns = set(
+      [feature_b, feature_c_bucketized, feature_a_x_feature_c])
+  features = tf.parse_example(
+      serialized=serialized_examples,
+      features=make_parse_example_spec(feature_columns))
+  ```
+
+  For the above example, make_parse_example_spec would return the dict:
+
+  ```python
+  {
+      "feature_a": parsing_ops.VarLenFeature(tf.string),
+      "feature_b": parsing_ops.FixedLenFeature([1], dtype=tf.float32),
+      "feature_c": parsing_ops.FixedLenFeature([1], dtype=tf.float32)
+  }
+  ```
+
+  Args:
+    feature_columns: An iterable containing all feature columns. All items
+      should be instances of classes derived from `FeatureColumn`.
+
+  Returns:
+    A dict mapping each feature key to a `FixedLenFeature` or `VarLenFeature`
+    value.
+
+  Raises:
+    ValueError: If any of the given `feature_columns` is not a `FeatureColumn`
+      instance.
+  """
+  result = {}
+  for column in feature_columns:
+    if not isinstance(column, FeatureColumn):
+      raise ValueError('All feature_columns must be FeatureColumn instances. '
+                       'Given: {}'.format(column))
+    config = column.parse_example_spec
+    for key, value in six.iteritems(config):
+      if key in result and value != result[key]:
+        raise ValueError(
+            'feature_columns contain different parse_spec for key '
+            '{}. Given {} and {}'.format(key, value, result[key]))
+    result.update(config)
+  return result
+
+
+def embedding_column(
+    categorical_column, dimension, combiner='mean', initializer=None,
+    ckpt_to_load_from=None, tensor_name_in_ckpt=None, max_norm=None,
+    trainable=True):
+  """`_DenseColumn` that converts from sparse, categorical input.
+
+  Use this when your inputs are sparse, but you want to convert them to a dense
+  representation (e.g., to feed to a DNN).
+
+  Inputs must be a `_CategoricalColumn` created by any of the
+  `categorical_column_*` function. Here is an example of using
+  `embedding_column` with `DNNClassifier`:
+
+  ```python
+  video_id = categorical_column_with_identity(
+      key='video_id', num_buckets=1000000, default_value=0)
+  columns = [embedding_column(video_id, 9),...]
+
+  estimator = tf.estimator.DNNClassifier(feature_columns=columns, ...)
+
+  label_column = ...
+  def input_fn():
+    features = tf.parse_example(
+        ..., features=make_parse_example_spec(columns + [label_column]))
+    labels = features.pop(label_column.name)
+    return features, labels
+
+  estimator.train(input_fn=input_fn, steps=100)
+  ```
+
+  Here is an example using `embedding_column` with model_fn:
+
+  ```python
+  def model_fn(features, ...):
+    video_id = categorical_column_with_identity(
+        key='video_id', num_buckets=1000000, default_value=0)
+    columns = [embedding_column(video_id, 9),...]
+    dense_tensor = input_layer(features, columns)
+    # Form DNN layers, calculate loss, and return EstimatorSpec.
+    ...
+  ```
+
+  Args:
+    categorical_column: A `_CategoricalColumn` created by a
+      `categorical_column_with_*` function. This column produces the sparse IDs
+      that are inputs to the embedding lookup.
+    dimension: An integer specifying dimension of the embedding, must be > 0.
+    combiner: A string specifying how to reduce if there are multiple entries
+      in a single row. Currently 'mean', 'sqrtn' and 'sum' are supported, with
+      'mean' the default. 'sqrtn' often achieves good accuracy, in particular
+      with bag-of-words columns. Each of this can be thought as example level
+      normalizations on the column. For more information, see
+      `tf.embedding_lookup_sparse`.
+    initializer: A variable initializer function to be used in embedding
+      variable initialization. If not specified, defaults to
+      `tf.truncated_normal_initializer` with mean `0.0` and standard deviation
+      `1/sqrt(dimension)`.
+    ckpt_to_load_from: String representing checkpoint name/pattern from which to
+      restore column weights. Required if `tensor_name_in_ckpt` is not `None`.
+    tensor_name_in_ckpt: Name of the `Tensor` in `ckpt_to_load_from` from
+      which to restore the column weights. Required if `ckpt_to_load_from` is
+      not `None`.
+    max_norm: If not `None`, embedding values are l2-normalized to this value.
+    trainable: Whether or not the embedding is trainable. Default is True.
+
+  Returns:
+    `_DenseColumn` that converts from sparse input.
+
+  Raises:
+    ValueError: if `dimension` not > 0.
+    ValueError: if exactly one of `ckpt_to_load_from` and `tensor_name_in_ckpt`
+      is specified.
+    ValueError: if `initializer` is specified and is not callable.
+    RuntimeError: If eager execution is enabled.
+  """
+  if (dimension is None) or (dimension < 1):
+    raise ValueError('Invalid dimension {}.'.format(dimension))
+  if (ckpt_to_load_from is None) != (tensor_name_in_ckpt is None):
+    raise ValueError('Must specify both `ckpt_to_load_from` and '
+                     '`tensor_name_in_ckpt` or none of them.')
+
+  if (initializer is not None) and (not callable(initializer)):
+    raise ValueError('initializer must be callable if specified. '
+                     'Embedding of column_name: {}'.format(
+                         categorical_column.name))
+  if initializer is None:
+    initializer = init_ops.truncated_normal_initializer(
+        mean=0.0, stddev=1 / math.sqrt(dimension))
+
+  return EmbeddingColumn(
+      categorical_column=categorical_column,
+      dimension=dimension,
+      combiner=combiner,
+      initializer=initializer,
+      ckpt_to_load_from=ckpt_to_load_from,
+      tensor_name_in_ckpt=tensor_name_in_ckpt,
+      max_norm=max_norm,
+      trainable=trainable)
+
+
+def shared_embedding_columns(
+    categorical_columns, dimension, combiner='mean', initializer=None,
+    shared_embedding_collection_name=None, ckpt_to_load_from=None,
+    tensor_name_in_ckpt=None, max_norm=None, trainable=True):
+  """List of dense columns that convert from sparse, categorical input.
+
+  This is similar to `embedding_column`, except that it produces a list of
+  embedding columns that share the same embedding weights.
+
+  Use this when your inputs are sparse and of the same type (e.g. watched and
+  impression video IDs that share the same vocabulary), and you want to convert
+  them to a dense representation (e.g., to feed to a DNN).
+
+  Inputs must be a list of categorical columns created by any of the
+  `categorical_column_*` function. They must all be of the same type and have
+  the same arguments except `key`. E.g. they can be
+  categorical_column_with_vocabulary_file with the same vocabulary_file. Some or
+  all columns could also be weighted_categorical_column.
+
+  Here is an example embedding of two features for a DNNClassifier model:
+
+  ```python
+  watched_video_id = categorical_column_with_vocabulary_file(
+      'watched_video_id', video_vocabulary_file, video_vocabulary_size)
+  impression_video_id = categorical_column_with_vocabulary_file(
+      'impression_video_id', video_vocabulary_file, video_vocabulary_size)
+  columns = shared_embedding_columns(
+      [watched_video_id, impression_video_id], dimension=10)
+
+  estimator = tf.estimator.DNNClassifier(feature_columns=columns, ...)
+
+  label_column = ...
+  def input_fn():
+    features = tf.parse_example(
+        ..., features=make_parse_example_spec(columns + [label_column]))
+    labels = features.pop(label_column.name)
+    return features, labels
+
+  estimator.train(input_fn=input_fn, steps=100)
+  ```
+
+  Here is an example using `shared_embedding_columns` with model_fn:
+
+  ```python
+  def model_fn(features, ...):
+    watched_video_id = categorical_column_with_vocabulary_file(
+        'watched_video_id', video_vocabulary_file, video_vocabulary_size)
+    impression_video_id = categorical_column_with_vocabulary_file(
+        'impression_video_id', video_vocabulary_file, video_vocabulary_size)
+    columns = shared_embedding_columns(
+        [watched_video_id, impression_video_id], dimension=10)
+    dense_tensor = input_layer(features, columns)
+    # Form DNN layers, calculate loss, and return EstimatorSpec.
+    ...
+  ```
+
+  Args:
+    categorical_columns: List of categorical columns created by a
+      `categorical_column_with_*` function. These columns produce the sparse IDs
+      that are inputs to the embedding lookup. All columns must be of the same
+      type and have the same arguments except `key`. E.g. they can be
+      categorical_column_with_vocabulary_file with the same vocabulary_file.
+      Some or all columns could also be weighted_categorical_column.
+    dimension: An integer specifying dimension of the embedding, must be > 0.
+    combiner: A string specifying how to reduce if there are multiple entries
+      in a single row. Currently 'mean', 'sqrtn' and 'sum' are supported, with
+      'mean' the default. 'sqrtn' often achieves good accuracy, in particular
+      with bag-of-words columns. Each of this can be thought as example level
+      normalizations on the column. For more information, see
+      `tf.embedding_lookup_sparse`.
+    initializer: A variable initializer function to be used in embedding
+      variable initialization. If not specified, defaults to
+      `tf.truncated_normal_initializer` with mean `0.0` and standard deviation
+      `1/sqrt(dimension)`.
+    shared_embedding_collection_name: Optional collective name of these columns.
+      If not given, a reasonable name will be chosen based on the names of
+      `categorical_columns`.
+    ckpt_to_load_from: String representing checkpoint name/pattern from which to
+      restore column weights. Required if `tensor_name_in_ckpt` is not `None`.
+    tensor_name_in_ckpt: Name of the `Tensor` in `ckpt_to_load_from` from
+      which to restore the column weights. Required if `ckpt_to_load_from` is
+      not `None`.
+    max_norm: If not `None`, each embedding is clipped if its l2-norm is
+      larger than this value, before combining.
+    trainable: Whether or not the embedding is trainable. Default is True.
+
+  Returns:
+    A list of dense columns that converts from sparse input. The order of
+    results follows the ordering of `categorical_columns`.
+
+  Raises:
+    ValueError: if `dimension` not > 0.
+    ValueError: if any of the given `categorical_columns` is of different type
+      or has different arguments than the others.
+    ValueError: if exactly one of `ckpt_to_load_from` and `tensor_name_in_ckpt`
+      is specified.
+    ValueError: if `initializer` is specified and is not callable.
+    RuntimeError: if eager execution is enabled.
+  """
+  if context.executing_eagerly():
+    raise RuntimeError('shared_embedding_columns are not supported when eager '
+                       'execution is enabled.')
+
+  if (dimension is None) or (dimension < 1):
+    raise ValueError('Invalid dimension {}.'.format(dimension))
+  if (ckpt_to_load_from is None) != (tensor_name_in_ckpt is None):
+    raise ValueError('Must specify both `ckpt_to_load_from` and '
+                     '`tensor_name_in_ckpt` or none of them.')
+
+  if (initializer is not None) and (not callable(initializer)):
+    raise ValueError('initializer must be callable if specified.')
+  if initializer is None:
+    initializer = init_ops.truncated_normal_initializer(
+        mean=0.0, stddev=1. / math.sqrt(dimension))
+
+  # Sort the columns so the default collection name is deterministic even if the
+  # user passes columns from an unsorted collection, such as dict.values().
+  sorted_columns = sorted(categorical_columns, key=lambda x: x.name)
+
+  c0 = sorted_columns[0]
+  num_buckets = c0.num_buckets
+  if not isinstance(c0, CategoricalColumn):
+    raise ValueError(
+        'All categorical_columns must be subclasses of CategoricalColumn. '
+        'Given: {}, of type: {}'.format(c0, type(c0)))
+  if isinstance(c0, WeightedCategoricalColumn):
+    c0 = c0.categorical_column
+  for c in sorted_columns[1:]:
+    if isinstance(c, WeightedCategoricalColumn):
+      c = c.categorical_column
+    if not isinstance(c, type(c0)):
+      raise ValueError(
+          'To use shared_embedding_column, all categorical_columns must have '
+          'the same type, or be weighted_categorical_column of the same type. '
+          'Given column: {} of type: {} does not match given column: {} of '
+          'type: {}'.format(c0, type(c0), c, type(c)))
+    if num_buckets != c.num_buckets:
+      raise ValueError(
+          'To use shared_embedding_column, all categorical_columns must have '
+          'the same number of buckets. Given column: {} with buckets: {} does  '
+          'not match column: {} with buckets: {}'.format(
+              c0, num_buckets, c, c.num_buckets))
+
+  if not shared_embedding_collection_name:
+    shared_embedding_collection_name = '_'.join(c.name for c in sorted_columns)
+    shared_embedding_collection_name += '_shared_embedding'
+
+  result = []
+  for column in categorical_columns:
+    result.append(
+        SharedEmbeddingColumn(
+            categorical_column=column,
+            initializer=initializer,
+            dimension=dimension,
+            combiner=combiner,
+            shared_embedding_collection_name=shared_embedding_collection_name,
+            ckpt_to_load_from=ckpt_to_load_from,
+            tensor_name_in_ckpt=tensor_name_in_ckpt,
+            max_norm=max_norm,
+            trainable=trainable))
+
+  return result
+
+
+def numeric_column(key,
+                   shape=(1,),
+                   default_value=None,
+                   dtype=dtypes.float32,
+                   normalizer_fn=None):
+  """Represents real valued or numerical features.
+
+  Example:
+
+  ```python
+  price = numeric_column('price')
+  columns = [price, ...]
+  features = tf.parse_example(..., features=make_parse_example_spec(columns))
+  dense_tensor = input_layer(features, columns)
+
+  # or
+  bucketized_price = bucketized_column(price, boundaries=[...])
+  columns = [bucketized_price, ...]
+  features = tf.parse_example(..., features=make_parse_example_spec(columns))
+  linear_prediction = linear_model(features, columns)
+  ```
+
+  Args:
+    key: A unique string identifying the input feature. It is used as the
+      column name and the dictionary key for feature parsing configs, feature
+      `Tensor` objects, and feature columns.
+    shape: An iterable of integers specifies the shape of the `Tensor`. An
+      integer can be given which means a single dimension `Tensor` with given
+      width. The `Tensor` representing the column will have the shape of
+      [batch_size] + `shape`.
+    default_value: A single value compatible with `dtype` or an iterable of
+      values compatible with `dtype` which the column takes on during
+      `tf.Example` parsing if data is missing. A default value of `None` will
+      cause `tf.parse_example` to fail if an example does not contain this
+      column. If a single value is provided, the same value will be applied as
+      the default value for every item. If an iterable of values is provided,
+      the shape of the `default_value` should be equal to the given `shape`.
+    dtype: defines the type of values. Default value is `tf.float32`. Must be a
+      non-quantized, real integer or floating point type.
+    normalizer_fn: If not `None`, a function that can be used to normalize the
+      value of the tensor after `default_value` is applied for parsing.
+      Normalizer function takes the input `Tensor` as its argument, and returns
+      the output `Tensor`. (e.g. lambda x: (x - 3.0) / 4.2). Please note that
+      even though the most common use case of this function is normalization, it
+      can be used for any kind of Tensorflow transformations.
+
+  Returns:
+    A `NumericColumn`.
+
+  Raises:
+    TypeError: if any dimension in shape is not an int
+    ValueError: if any dimension in shape is not a positive integer
+    TypeError: if `default_value` is an iterable but not compatible with `shape`
+    TypeError: if `default_value` is not compatible with `dtype`.
+    ValueError: if `dtype` is not convertible to `tf.float32`.
+  """
+  shape = _check_shape(shape, key)
+  if not (dtype.is_integer or dtype.is_floating):
+    raise ValueError('dtype must be convertible to float. '
+                     'dtype: {}, key: {}'.format(dtype, key))
+  default_value = _check_default_value(shape, default_value, dtype, key)
+
+  if normalizer_fn is not None and not callable(normalizer_fn):
+    raise TypeError(
+        'normalizer_fn must be a callable. Given: {}'.format(normalizer_fn))
+
+  _assert_key_is_string(key)
+  return NumericColumn(
+      key,
+      shape=shape,
+      default_value=default_value,
+      dtype=dtype,
+      normalizer_fn=normalizer_fn)
+
+
+def bucketized_column(source_column, boundaries):
+  """Represents discretized dense input.
+
+  Buckets include the left boundary, and exclude the right boundary. Namely,
+  `boundaries=[0., 1., 2.]` generates buckets `(-inf, 0.)`, `[0., 1.)`,
+  `[1., 2.)`, and `[2., +inf)`.
+
+  For example, if the inputs are
+
+  ```python
+  boundaries = [0, 10, 100]
+  input tensor = [[-5, 10000]
+                  [150,   10]
+                  [5,    100]]
+  ```
+
+  then the output will be
+
+  ```python
+  output = [[0, 3]
+            [3, 2]
+            [1, 3]]
+  ```
+
+  Example:
+
+  ```python
+  price = numeric_column('price')
+  bucketized_price = bucketized_column(price, boundaries=[...])
+  columns = [bucketized_price, ...]
+  features = tf.parse_example(..., features=make_parse_example_spec(columns))
+  linear_prediction = linear_model(features, columns)
+
+  # or
+  columns = [bucketized_price, ...]
+  features = tf.parse_example(..., features=make_parse_example_spec(columns))
+  dense_tensor = input_layer(features, columns)
+  ```
+
+  `bucketized_column` can also be crossed with another categorical column using
+  `crossed_column`:
+
+  ```python
+  price = numeric_column('price')
+  # bucketized_column converts numerical feature to a categorical one.
+  bucketized_price = bucketized_column(price, boundaries=[...])
+  # 'keywords' is a string feature.
+  price_x_keywords = crossed_column([bucketized_price, 'keywords'], 50K)
+  columns = [price_x_keywords, ...]
+  features = tf.parse_example(..., features=make_parse_example_spec(columns))
+  linear_prediction = linear_model(features, columns)
+  ```
+
+  Args:
+    source_column: A one-dimensional dense column which is generated with
+      `numeric_column`.
+    boundaries: A sorted list or tuple of floats specifying the boundaries.
+
+  Returns:
+    A `BucketizedColumn`.
+
+  Raises:
+    ValueError: If `source_column` is not a numeric column, or if it is not
+      one-dimensional.
+    ValueError: If `boundaries` is not a sorted list or tuple.
+  """
+  if not isinstance(source_column, NumericColumn):
+    raise ValueError(
+        'source_column must be a column generated with numeric_column(). '
+        'Given: {}'.format(source_column))
+  if len(source_column.shape) > 1:
+    raise ValueError(
+        'source_column must be one-dimensional column. '
+        'Given: {}'.format(source_column))
+  if (not boundaries or
+      not (isinstance(boundaries, list) or isinstance(boundaries, tuple))):
+    raise ValueError('boundaries must be a sorted list.')
+  for i in range(len(boundaries) - 1):
+    if boundaries[i] >= boundaries[i + 1]:
+      raise ValueError('boundaries must be a sorted list.')
+  return BucketizedColumn(source_column, tuple(boundaries))
+
+
+def _assert_string_or_int(dtype, prefix):
+  if (dtype != dtypes.string) and (not dtype.is_integer):
+    raise ValueError(
+        '{} dtype must be string or integer. dtype: {}.'.format(prefix, dtype))
+
+
+def _assert_key_is_string(key):
+  if not isinstance(key, six.string_types):
+    raise ValueError(
+        'key must be a string. Got: type {}. Given key: {}.'.format(
+            type(key), key))
+
+
+def categorical_column_with_hash_bucket(key,
+                                        hash_bucket_size,
+                                        dtype=dtypes.string):
+  """Represents sparse feature where ids are set by hashing.
+
+  Use this when your sparse features are in string or integer format, and you
+  want to distribute your inputs into a finite number of buckets by hashing.
+  output_id = Hash(input_feature_string) % bucket_size for string type input.
+  For int type input, the value is converted to its string representation first
+  and then hashed by the same formula.
+
+  For input dictionary `features`, `features[key]` is either `Tensor` or
+  `SparseTensor`. If `Tensor`, missing values can be represented by `-1` for int
+  and `''` for string, which will be dropped by this feature column.
+
+  Example:
+
+  ```python
+  keywords = categorical_column_with_hash_bucket("keywords", 10K)
+  columns = [keywords, ...]
+  features = tf.parse_example(..., features=make_parse_example_spec(columns))
+  linear_prediction = linear_model(features, columns)
+
+  # or
+  keywords_embedded = embedding_column(keywords, 16)
+  columns = [keywords_embedded, ...]
+  features = tf.parse_example(..., features=make_parse_example_spec(columns))
+  dense_tensor = input_layer(features, columns)
+  ```
+
+  Args:
+    key: A unique string identifying the input feature. It is used as the
+      column name and the dictionary key for feature parsing configs, feature
+      `Tensor` objects, and feature columns.
+    hash_bucket_size: An int > 1. The number of buckets.
+    dtype: The type of features. Only string and integer types are supported.
+
+  Returns:
+    A `HashedCategoricalColumn`.
+
+  Raises:
+    ValueError: `hash_bucket_size` is not greater than 1.
+    ValueError: `dtype` is neither string nor integer.
+  """
+  if hash_bucket_size is None:
+    raise ValueError('hash_bucket_size must be set. ' 'key: {}'.format(key))
+
+  if hash_bucket_size < 1:
+    raise ValueError('hash_bucket_size must be at least 1. '
+                     'hash_bucket_size: {}, key: {}'.format(
+                         hash_bucket_size, key))
+
+  _assert_key_is_string(key)
+  _assert_string_or_int(dtype, prefix='column_name: {}'.format(key))
+
+  return HashedCategoricalColumn(key, hash_bucket_size, dtype)
+
+
+def categorical_column_with_vocabulary_file(key,
+                                            vocabulary_file,
+                                            vocabulary_size=None,
+                                            num_oov_buckets=0,
+                                            default_value=None,
+                                            dtype=dtypes.string):
+  """A `CategoricalColumn` with a vocabulary file.
+
+  Use this when your inputs are in string or integer format, and you have a
+  vocabulary file that maps each value to an integer ID. By default,
+  out-of-vocabulary values are ignored. Use either (but not both) of
+  `num_oov_buckets` and `default_value` to specify how to include
+  out-of-vocabulary values.
+
+  For input dictionary `features`, `features[key]` is either `Tensor` or
+  `SparseTensor`. If `Tensor`, missing values can be represented by `-1` for int
+  and `''` for string, which will be dropped by this feature column.
+
+  Example with `num_oov_buckets`:
+  File '/us/states.txt' contains 50 lines, each with a 2-character U.S. state
+  abbreviation. All inputs with values in that file are assigned an ID 0-49,
+  corresponding to its line number. All other values are hashed and assigned an
+  ID 50-54.
+
+  ```python
+  states = categorical_column_with_vocabulary_file(
+      key='states', vocabulary_file='/us/states.txt', vocabulary_size=50,
+      num_oov_buckets=5)
+  columns = [states, ...]
+  features = tf.parse_example(..., features=make_parse_example_spec(columns))
+  linear_prediction = linear_model(features, columns)
+  ```
+
+  Example with `default_value`:
+  File '/us/states.txt' contains 51 lines - the first line is 'XX', and the
+  other 50 each have a 2-character U.S. state abbreviation. Both a literal 'XX'
+  in input, and other values missing from the file, will be assigned ID 0. All
+  others are assigned the corresponding line number 1-50.
+
+  ```python
+  states = categorical_column_with_vocabulary_file(
+      key='states', vocabulary_file='/us/states.txt', vocabulary_size=51,
+      default_value=0)
+  columns = [states, ...]
+  features = tf.parse_example(..., features=make_parse_example_spec(columns))
+  linear_prediction, _, _ = linear_model(features, columns)
+  ```
+
+  And to make an embedding with either:
+
+  ```python
+  columns = [embedding_column(states, 3),...]
+  features = tf.parse_example(..., features=make_parse_example_spec(columns))
+  dense_tensor = input_layer(features, columns)
+  ```
+
+  Args:
+    key: A unique string identifying the input feature. It is used as the
+      column name and the dictionary key for feature parsing configs, feature
+      `Tensor` objects, and feature columns.
+    vocabulary_file: The vocabulary file name.
+    vocabulary_size: Number of the elements in the vocabulary. This must be no
+      greater than length of `vocabulary_file`, if less than length, later
+      values are ignored. If None, it is set to the length of `vocabulary_file`.
+    num_oov_buckets: Non-negative integer, the number of out-of-vocabulary
+      buckets. All out-of-vocabulary inputs will be assigned IDs in the range
+      `[vocabulary_size, vocabulary_size+num_oov_buckets)` based on a hash of
+      the input value. A positive `num_oov_buckets` can not be specified with
+      `default_value`.
+    default_value: The integer ID value to return for out-of-vocabulary feature
+      values, defaults to `-1`. This can not be specified with a positive
+      `num_oov_buckets`.
+    dtype: The type of features. Only string and integer types are supported.
+
+  Returns:
+    A `CategoricalColumn` with a vocabulary file.
+
+  Raises:
+    ValueError: `vocabulary_file` is missing or cannot be opened.
+    ValueError: `vocabulary_size` is missing or < 1.
+    ValueError: `num_oov_buckets` is a negative integer.
+    ValueError: `num_oov_buckets` and `default_value` are both specified.
+    ValueError: `dtype` is neither string nor integer.
+  """
+  if not vocabulary_file:
+    raise ValueError('Missing vocabulary_file in {}.'.format(key))
+
+  if vocabulary_size is None:
+    if not gfile.Exists(vocabulary_file):
+      raise ValueError('vocabulary_file in {} does not exist.'.format(key))
+
+    with gfile.GFile(vocabulary_file) as f:
+      vocabulary_size = sum(1 for _ in f)
+    logging.info(
+        'vocabulary_size = %d in %s is inferred from the number of elements '
+        'in the vocabulary_file %s.', vocabulary_size, key, vocabulary_file)
+
+  # `vocabulary_size` isn't required for lookup, but it is for `_num_buckets`.
+  if vocabulary_size < 1:
+    raise ValueError('Invalid vocabulary_size in {}.'.format(key))
+  if num_oov_buckets:
+    if default_value is not None:
+      raise ValueError(
+          'Can\'t specify both num_oov_buckets and default_value in {}.'.format(
+              key))
+    if num_oov_buckets < 0:
+      raise ValueError('Invalid num_oov_buckets {} in {}.'.format(
+          num_oov_buckets, key))
+  _assert_string_or_int(dtype, prefix='column_name: {}'.format(key))
+  _assert_key_is_string(key)
+  return VocabularyFileCategoricalColumn(
+      key=key,
+      vocabulary_file=vocabulary_file,
+      vocabulary_size=vocabulary_size,
+      num_oov_buckets=0 if num_oov_buckets is None else num_oov_buckets,
+      default_value=-1 if default_value is None else default_value,
+      dtype=dtype)
+
+
+def categorical_column_with_vocabulary_list(
+    key, vocabulary_list, dtype=None, default_value=-1, num_oov_buckets=0):
+  """A `_CategoricalColumn` with in-memory vocabulary.
+
+  Use this when your inputs are in string or integer format, and you have an
+  in-memory vocabulary mapping each value to an integer ID. By default,
+  out-of-vocabulary values are ignored. Use either (but not both) of
+  `num_oov_buckets` and `default_value` to specify how to include
+  out-of-vocabulary values.
+
+  For input dictionary `features`, `features[key]` is either `Tensor` or
+  `SparseTensor`. If `Tensor`, missing values can be represented by `-1` for int
+  and `''` for string, which will be dropped by this feature column.
+
+  Example with `num_oov_buckets`:
+  In the following example, each input in `vocabulary_list` is assigned an ID
+  0-3 corresponding to its index (e.g., input 'B' produces output 2). All other
+  inputs are hashed and assigned an ID 4-5.
+
+  ```python
+  colors = categorical_column_with_vocabulary_list(
+      key='colors', vocabulary_list=('R', 'G', 'B', 'Y'),
+      num_oov_buckets=2)
+  columns = [colors, ...]
+  features = tf.parse_example(..., features=make_parse_example_spec(columns))
+  linear_prediction, _, _ = linear_model(features, columns)
+  ```
+
+  Example with `default_value`:
+  In the following example, each input in `vocabulary_list` is assigned an ID
+  0-4 corresponding to its index (e.g., input 'B' produces output 3). All other
+  inputs are assigned `default_value` 0.
+
+
+  ```python
+  colors = categorical_column_with_vocabulary_list(
+      key='colors', vocabulary_list=('X', 'R', 'G', 'B', 'Y'), default_value=0)
+  columns = [colors, ...]
+  features = tf.parse_example(..., features=make_parse_example_spec(columns))
+  linear_prediction, _, _ = linear_model(features, columns)
+  ```
+
+  And to make an embedding with either:
+
+  ```python
+  columns = [embedding_column(colors, 3),...]
+  features = tf.parse_example(..., features=make_parse_example_spec(columns))
+  dense_tensor = input_layer(features, columns)
+  ```
+
+  Args:
+    key: A unique string identifying the input feature. It is used as the
+      column name and the dictionary key for feature parsing configs, feature
+      `Tensor` objects, and feature columns.
+    vocabulary_list: An ordered iterable defining the vocabulary. Each feature
+      is mapped to the index of its value (if present) in `vocabulary_list`.
+      Must be castable to `dtype`.
+    dtype: The type of features. Only string and integer types are supported.
+      If `None`, it will be inferred from `vocabulary_list`.
+    default_value: The integer ID value to return for out-of-vocabulary feature
+      values, defaults to `-1`. This can not be specified with a positive
+      `num_oov_buckets`.
+    num_oov_buckets: Non-negative integer, the number of out-of-vocabulary
+      buckets. All out-of-vocabulary inputs will be assigned IDs in the range
+      `[len(vocabulary_list), len(vocabulary_list)+num_oov_buckets)` based on a
+      hash of the input value. A positive `num_oov_buckets` can not be specified
+      with `default_value`.
+
+  Returns:
+    A `CategoricalColumn` with in-memory vocabulary.
+
+  Raises:
+    ValueError: if `vocabulary_list` is empty, or contains duplicate keys.
+    ValueError: `num_oov_buckets` is a negative integer.
+    ValueError: `num_oov_buckets` and `default_value` are both specified.
+    ValueError: if `dtype` is not integer or string.
+  """
+  if (vocabulary_list is None) or (len(vocabulary_list) < 1):
+    raise ValueError(
+        'vocabulary_list {} must be non-empty, column_name: {}'.format(
+            vocabulary_list, key))
+  if len(set(vocabulary_list)) != len(vocabulary_list):
+    raise ValueError(
+        'Duplicate keys in vocabulary_list {}, column_name: {}'.format(
+            vocabulary_list, key))
+  vocabulary_dtype = dtypes.as_dtype(np.array(vocabulary_list).dtype)
+  if num_oov_buckets:
+    if default_value != -1:
+      raise ValueError(
+          'Can\'t specify both num_oov_buckets and default_value in {}.'.format(
+              key))
+    if num_oov_buckets < 0:
+      raise ValueError('Invalid num_oov_buckets {} in {}.'.format(
+          num_oov_buckets, key))
+  _assert_string_or_int(
+      vocabulary_dtype, prefix='column_name: {} vocabulary'.format(key))
+  if dtype is None:
+    dtype = vocabulary_dtype
+  elif dtype.is_integer != vocabulary_dtype.is_integer:
+    raise ValueError(
+        'dtype {} and vocabulary dtype {} do not match, column_name: {}'.format(
+            dtype, vocabulary_dtype, key))
+  _assert_string_or_int(dtype, prefix='column_name: {}'.format(key))
+  _assert_key_is_string(key)
+
+  return VocabularyListCategoricalColumn(
+      key=key,
+      vocabulary_list=tuple(vocabulary_list),
+      dtype=dtype,
+      default_value=default_value,
+      num_oov_buckets=num_oov_buckets)
+
+
+def categorical_column_with_identity(key, num_buckets, default_value=None):
+  """A `CategoricalColumn` that returns identity values.
+
+  Use this when your inputs are integers in the range `[0, num_buckets)`, and
+  you want to use the input value itself as the categorical ID. Values outside
+  this range will result in `default_value` if specified, otherwise it will
+  fail.
+
+  Typically, this is used for contiguous ranges of integer indexes, but
+  it doesn't have to be. This might be inefficient, however, if many of IDs
+  are unused. Consider `categorical_column_with_hash_bucket` in that case.
+
+  For input dictionary `features`, `features[key]` is either `Tensor` or
+  `SparseTensor`. If `Tensor`, missing values can be represented by `-1` for int
+  and `''` for string, which will be dropped by this feature column.
+
+  In the following examples, each input in the range `[0, 1000000)` is assigned
+  the same value. All other inputs are assigned `default_value` 0. Note that a
+  literal 0 in inputs will result in the same default ID.
+
+  Linear model:
+
+  ```python
+  video_id = categorical_column_with_identity(
+      key='video_id', num_buckets=1000000, default_value=0)
+  columns = [video_id, ...]
+  features = tf.parse_example(..., features=make_parse_example_spec(columns))
+  linear_prediction, _, _ = linear_model(features, columns)
+  ```
+
+  Embedding for a DNN model:
+
+  ```python
+  columns = [embedding_column(video_id, 9),...]
+  features = tf.parse_example(..., features=make_parse_example_spec(columns))
+  dense_tensor = input_layer(features, columns)
+  ```
+
+  Args:
+    key: A unique string identifying the input feature. It is used as the
+      column name and the dictionary key for feature parsing configs, feature
+      `Tensor` objects, and feature columns.
+    num_buckets: Range of inputs and outputs is `[0, num_buckets)`.
+    default_value: If `None`, this column's graph operations will fail for
+      out-of-range inputs. Otherwise, this value must be in the range
+      `[0, num_buckets)`, and will replace inputs in that range.
+
+  Returns:
+    A `CategoricalColumn` that returns identity values.
+
+  Raises:
+    ValueError: if `num_buckets` is less than one.
+    ValueError: if `default_value` is not in range `[0, num_buckets)`.
+  """
+  if num_buckets < 1:
+    raise ValueError(
+        'num_buckets {} < 1, column_name {}'.format(num_buckets, key))
+  if (default_value is not None) and (
+      (default_value < 0) or (default_value >= num_buckets)):
+    raise ValueError(
+        'default_value {} not in range [0, {}), column_name {}'.format(
+            default_value, num_buckets, key))
+  _assert_key_is_string(key)
+  return IdentityCategoricalColumn(
+      key=key, number_buckets=num_buckets, default_value=default_value)
+
+
+def indicator_column(categorical_column):
+  """Represents multi-hot representation of given categorical column.
+
+  - For DNN model, `indicator_column` can be used to wrap any
+    `categorical_column_*` (e.g., to feed to DNN). Consider to Use
+    `embedding_column` if the number of buckets/unique(values) are large.
+
+  - For Wide (aka linear) model, `indicator_column` is the internal
+    representation for categorical column when passing categorical column
+    directly (as any element in feature_columns) to `linear_model`. See
+    `linear_model` for details.
+
+  ```python
+  name = indicator_column(categorical_column_with_vocabulary_list(
+      'name', ['bob', 'george', 'wanda'])
+  columns = [name, ...]
+  features = tf.parse_example(..., features=make_parse_example_spec(columns))
+  dense_tensor = input_layer(features, columns)
+
+  dense_tensor == [[1, 0, 0]]  # If "name" bytes_list is ["bob"]
+  dense_tensor == [[1, 0, 1]]  # If "name" bytes_list is ["bob", "wanda"]
+  dense_tensor == [[2, 0, 0]]  # If "name" bytes_list is ["bob", "bob"]
+  ```
+
+  Args:
+    categorical_column: A `CategoricalColumn` which is created by
+      `categorical_column_with_*` or `crossed_column` functions.
+
+  Returns:
+    An `IndicatorColumn`.
+  """
+  return IndicatorColumn(categorical_column)
+
+
+def weighted_categorical_column(
+    categorical_column, weight_feature_key, dtype=dtypes.float32):
+  """Applies weight values to a `_CategoricalColumn`.
+
+  Use this when each of your sparse inputs has both an ID and a value. For
+  example, if you're representing text documents as a collection of word
+  frequencies, you can provide 2 parallel sparse input features ('terms' and
+  'frequencies' below).
+
+  Example:
+
+  Input `tf.Example` objects:
+
+  ```proto
+  [
+    features {
+      feature {
+        key: "terms"
+        value {bytes_list {value: "very" value: "model"}}
+      }
+      feature {
+        key: "frequencies"
+        value {float_list {value: 0.3 value: 0.1}}
+      }
+    },
+    features {
+      feature {
+        key: "terms"
+        value {bytes_list {value: "when" value: "course" value: "human"}}
+      }
+      feature {
+        key: "frequencies"
+        value {float_list {value: 0.4 value: 0.1 value: 0.2}}
+      }
+    }
+  ]
+  ```
+
+  ```python
+  categorical_column = categorical_column_with_hash_bucket(
+      column_name='terms', hash_bucket_size=1000)
+  weighted_column = weighted_categorical_column(
+      categorical_column=categorical_column, weight_feature_key='frequencies')
+  columns = [weighted_column, ...]
+  features = tf.parse_example(..., features=make_parse_example_spec(columns))
+  linear_prediction, _, _ = linear_model(features, columns)
+  ```
+
+  This assumes the input dictionary contains a `SparseTensor` for key
+  'terms', and a `SparseTensor` for key 'frequencies'. These 2 tensors must have
+  the same indices and dense shape.
+
+  Args:
+    categorical_column: A `_CategoricalColumn` created by
+      `categorical_column_with_*` functions.
+    weight_feature_key: String key for weight values.
+    dtype: Type of weights, such as `tf.float32`. Only float and integer weights
+      are supported.
+
+  Returns:
+    A `CategoricalColumn` composed of two sparse features: one represents id,
+    the other represents weight (value) of the id feature in that example.
+
+  Raises:
+    ValueError: if `dtype` is not convertible to float.
+  """
+  if (dtype is None) or not (dtype.is_integer or dtype.is_floating):
+    raise ValueError('dtype {} is not convertible to float.'.format(dtype))
+  return WeightedCategoricalColumn(
+      categorical_column=categorical_column,
+      weight_feature_key=weight_feature_key,
+      dtype=dtype)
+
+
+def crossed_column(keys, hash_bucket_size, hash_key=None):
+  """Returns a column for performing crosses of categorical features.
+
+  Crossed features will be hashed according to `hash_bucket_size`. Conceptually,
+  the transformation can be thought of as:
+    Hash(cartesian product of features) % `hash_bucket_size`
+
+  For example, if the input features are:
+
+  * SparseTensor referred by first key:
+
+    ```python
+    shape = [2, 2]
+    {
+        [0, 0]: "a"
+        [1, 0]: "b"
+        [1, 1]: "c"
+    }
+    ```
+
+  * SparseTensor referred by second key:
+
+    ```python
+    shape = [2, 1]
+    {
+        [0, 0]: "d"
+        [1, 0]: "e"
+    }
+    ```
+
+  then crossed feature will look like:
+
+  ```python
+   shape = [2, 2]
+  {
+      [0, 0]: Hash64("d", Hash64("a")) % hash_bucket_size
+      [1, 0]: Hash64("e", Hash64("b")) % hash_bucket_size
+      [1, 1]: Hash64("e", Hash64("c")) % hash_bucket_size
+  }
+  ```
+
+  Here is an example to create a linear model with crosses of string features:
+
+  ```python
+  keywords_x_doc_terms = crossed_column(['keywords', 'doc_terms'], 50K)
+  columns = [keywords_x_doc_terms, ...]
+  features = tf.parse_example(..., features=make_parse_example_spec(columns))
+  linear_prediction = linear_model(features, columns)
+  ```
+
+  You could also use vocabulary lookup before crossing:
+
+  ```python
+  keywords = categorical_column_with_vocabulary_file(
+      'keywords', '/path/to/vocabulary/file', vocabulary_size=1K)
+  keywords_x_doc_terms = crossed_column([keywords, 'doc_terms'], 50K)
+  columns = [keywords_x_doc_terms, ...]
+  features = tf.parse_example(..., features=make_parse_example_spec(columns))
+  linear_prediction = linear_model(features, columns)
+  ```
+
+  If an input feature is of numeric type, you can use
+  `categorical_column_with_identity`, or `bucketized_column`, as in the example:
+
+  ```python
+  # vertical_id is an integer categorical feature.
+  vertical_id = categorical_column_with_identity('vertical_id', 10K)
+  price = numeric_column('price')
+  # bucketized_column converts numerical feature to a categorical one.
+  bucketized_price = bucketized_column(price, boundaries=[...])
+  vertical_id_x_price = crossed_column([vertical_id, bucketized_price], 50K)
+  columns = [vertical_id_x_price, ...]
+  features = tf.parse_example(..., features=make_parse_example_spec(columns))
+  linear_prediction = linear_model(features, columns)
+  ```
+
+  To use crossed column in DNN model, you need to add it in an embedding column
+  as in this example:
+
+  ```python
+  vertical_id_x_price = crossed_column([vertical_id, bucketized_price], 50K)
+  vertical_id_x_price_embedded = embedding_column(vertical_id_x_price, 10)
+  dense_tensor = input_layer(features, [vertical_id_x_price_embedded, ...])
+  ```
+
+  Args:
+    keys: An iterable identifying the features to be crossed. Each element can
+      be either:
+      * string: Will use the corresponding feature which must be of string type.
+      * `CategoricalColumn`: Will use the transformed tensor produced by this
+        column. Does not support hashed categorical column.
+    hash_bucket_size: An int > 1. The number of buckets.
+    hash_key: Specify the hash_key that will be used by the `FingerprintCat64`
+      function to combine the crosses fingerprints on SparseCrossOp (optional).
+
+  Returns:
+    A `CrossedColumn`.
+
+  Raises:
+    ValueError: If `len(keys) < 2`.
+    ValueError: If any of the keys is neither a string nor `CategoricalColumn`.
+    ValueError: If any of the keys is `HashedCategoricalColumn`.
+    ValueError: If `hash_bucket_size < 1`.
+  """
+  if not hash_bucket_size or hash_bucket_size < 1:
+    raise ValueError('hash_bucket_size must be > 1. '
+                     'hash_bucket_size: {}'.format(hash_bucket_size))
+  if not keys or len(keys) < 2:
+    raise ValueError(
+        'keys must be a list with length > 1. Given: {}'.format(keys))
+  for key in keys:
+    if (not isinstance(key, six.string_types) and
+        not isinstance(key, CategoricalColumn)):
+      raise ValueError(
+          'Unsupported key type. All keys must be either string, or '
+          'categorical column except HashedCategoricalColumn. '
+          'Given: {}'.format(key))
+    if isinstance(key, HashedCategoricalColumn):
+      raise ValueError(
+          'categorical_column_with_hash_bucket is not supported for crossing. '
+          'Hashing before crossing will increase probability of collision. '
+          'Instead, use the feature name as a string. Given: {}'.format(key))
+  return CrossedColumn(
+      keys=tuple(keys), hash_bucket_size=hash_bucket_size, hash_key=hash_key)
+
+
+class StateManager(object):
+  """Manages the state associated with FeatureColumns.
+
+  Some `FeatureColumn`s create variables or resources to assist their
+  computation. The `StateManager` is responsible for creating and storing these
+  objects since `FeatureColumn`s are supposed to be stateless configuration
+  only.
+  """
+
+  def get_variable(self,
+                   feature_column,
+                   name,
+                   shape,
+                   dtype=None,
+                   initializer=None):
+    """Creates a new variable or returns an existing one.
+
+    Args:
+      feature_column: A `FeatureColumn` object this variable corresponds to.
+      name: variable name.
+      shape: variable shape.
+      dtype: The type of the variable. Defaults to `self.dtype` or `float32`.
+      initializer: initializer instance (callable).
+
+    Returns:
+      The variable.
+    """
+    raise NotImplementedError('StateManager.get_variable')
+
+  def get_resource(self, feature_column, name, resource_creator):
+    """Creates a new resource or returns an existing one.
+
+    Resources can be things such as tables etc.
+
+    Args:
+      feature_column: A `FeatureColumn` object this variable corresponds to.
+      name: Name of the resource.
+      resource_creator: A callable that can create the resource.
+
+    Returns:
+      The resource.
+    """
+    raise NotImplementedError('StateManager.get_resource')
+
+
+class FeatureColumn(object):
+  """Represents a feature column abstraction.
+
+  WARNING: Do not subclass this layer unless you know what you are doing:
+  the API is subject to future changes.
+
+  To distinguish between the concept of a feature family and a specific binary
+  feature within a family, we refer to a feature family like "country" as a
+  feature column. For example, we can have a feature in a `tf.Example` format:
+    {key: "country",  value: [ "US" ]}
+  In this example the value of feature is "US" and "country" refers to the
+  column of the feature.
+
+  This class is an abstract class. Users should not create instances of this.
+  """
+  __metaclass__ = abc.ABCMeta
+
+  @abc.abstractproperty
+  def name(self):
+    """Returns string. Used for naming."""
+    pass
+
+  @abc.abstractmethod
+  def transform_feature(self, transformation_cache, state_manager):
+    """Returns intermediate representation (usually a `Tensor`).
+
+    Uses `transformation_cache` to create an intermediate representation
+    (usually a `Tensor`) that other feature columns can use.
+
+    Example usage of `transformation_cache`:
+    Let's say a Feature column depends on raw feature ('raw') and another
+    `FeatureColumn` (input_fc). To access corresponding `Tensor`s,
+    transformation_cache will be used as follows:
+
+    ```python
+    raw_tensor = transformation_cache.get('raw', state_manager)
+    fc_tensor = transformation_cache.get(input_fc, state_manager)
+    ```
+
+    Args:
+      transformation_cache: A `FeatureTransformationCache` object to access
+        features.
+      state_manager: A `StateManager` to create / access resources such as
+        lookup tables.
+
+    Returns:
+      Transformed feature `Tensor`.
+    """
+    pass
+
+  @abc.abstractproperty
+  def parse_example_spec(self):
+    """Returns a `tf.Example` parsing spec as dict.
+
+    It is used for get_parsing_spec for `tf.parse_example`. Returned spec is a
+    dict from keys ('string') to `VarLenFeature`, `FixedLenFeature`, and other
+    supported objects. Please check documentation of @{tf.parse_example} for all
+    supported spec objects.
+
+    Let's say a Feature column depends on raw feature ('raw') and another
+    `FeatureColumn` (input_fc). One possible implementation of
+    parse_example_spec is as follows:
+
+    ```python
+    spec = {'raw': tf.FixedLenFeature(...)}
+    spec.update(input_fc.parse_example_spec)
+    return spec
+    ```
+    """
+    pass
+
+  def create_state(self, state_manager):
+    """Uses the `state_manager` to create state for the FeatureColumn.
+
+    Args:
+      state_manager: A `StateManager` to create / access resources such as
+        lookup tables and variables.
+    """
+    pass
+
+
+class DenseColumn(FeatureColumn):
+  """Represents a column which can be represented as `Tensor`.
+
+  Some examples of this type are: numeric_column, embedding_column,
+  indicator_column.
+  """
+
+  __metaclass__ = abc.ABCMeta
+
+  @abc.abstractproperty
+  def variable_shape(self):
+    """`TensorShape` of `get_dense_tensor`, without batch dimension."""
+    pass
+
+  @abc.abstractmethod
+  def get_dense_tensor(self, transformation_cache, state_manager):
+    """Returns a `Tensor`.
+
+    The output of this function will be used by model-builder-functions. For
+    example the pseudo code of `input_layer` will be like:
+
+    ```python
+    def input_layer(features, feature_columns, ...):
+      outputs = [fc.get_dense_tensor(...) for fc in feature_columns]
+      return tf.concat(outputs)
+    ```
+
+    Args:
+      transformation_cache: A `FeatureTransformationCache` object to access
+        features.
+      state_manager: A `StateManager` to create / access resources such as
+        lookup tables.
+
+    Returns:
+      `Tensor` of shape [batch_size] + `variable_shape`.
+    """
+    pass
+
+
+def _create_weighted_sum(column,
+                         transformation_cache,
+                         state_manager,
+                         units,
+                         sparse_combiner,
+                         weight_collections,
+                         trainable,
+                         weight_var=None):
+  """Creates a weighted sum for a dense/categorical column for linear_model."""
+  if isinstance(column, CategoricalColumn):
+    return _create_categorical_column_weighted_sum(
+        column=column,
+        transformation_cache=transformation_cache,
+        state_manager=state_manager,
+        units=units,
+        sparse_combiner=sparse_combiner,
+        weight_collections=weight_collections,
+        trainable=trainable,
+        weight_var=weight_var)
+  else:
+    return _create_dense_column_weighted_sum(
+        column=column,
+        transformation_cache=transformation_cache,
+        state_manager=state_manager,
+        units=units,
+        weight_collections=weight_collections,
+        trainable=trainable,
+        weight_var=weight_var)
+
+
+def _create_dense_column_weighted_sum(column,
+                                      transformation_cache,
+                                      state_manager,
+                                      units,
+                                      weight_collections,
+                                      trainable,
+                                      weight_var=None):
+  """Create a weighted sum of a dense column for linear_model."""
+  tensor = column.get_dense_tensor(transformation_cache, state_manager)
+  num_elements = column.variable_shape.num_elements()
+  batch_size = array_ops.shape(tensor)[0]
+  tensor = array_ops.reshape(tensor, shape=(batch_size, num_elements))
+  if weight_var is not None:
+    weight = weight_var
+  else:
+    weight = variable_scope.get_variable(
+        name='weights',
+        shape=[num_elements, units],
+        initializer=init_ops.zeros_initializer(),
+        trainable=trainable,
+        collections=weight_collections)
+  return math_ops.matmul(tensor, weight, name='weighted_sum')
+
+
+class CategoricalColumn(FeatureColumn):
+  """Represents a categorical feature.
+
+  A categorical feature typically handled with a @{tf.SparseTensor} of IDs.
+  """
+  __metaclass__ = abc.ABCMeta
+
+  IdWeightPair = collections.namedtuple(  # pylint: disable=invalid-name
+      'IdWeightPair', ('id_tensor', 'weight_tensor'))
+
+  @abc.abstractproperty
+  def num_buckets(self):
+    """Returns number of buckets in this sparse feature."""
+    pass
+
+  @abc.abstractmethod
+  def get_sparse_tensors(self, transformation_cache, state_manager):
+    """Returns an IdWeightPair.
+
+    `IdWeightPair` is a pair of `SparseTensor`s which represents ids and
+    weights.
+
+    `IdWeightPair.id_tensor` is typically a `batch_size` x `num_buckets`
+    `SparseTensor` of `int64`. `IdWeightPair.weight_tensor` is either a
+    `SparseTensor` of `float` or `None` to indicate all weights should be
+    taken to be 1. If specified, `weight_tensor` must have exactly the same
+    shape and indices as `sp_ids`. Expected `SparseTensor` is same as parsing
+    output of a `VarLenFeature` which is a ragged matrix.
+
+    Args:
+      transformation_cache: A `FeatureTransformationCache` object to access
+        features.
+      state_manager: A `StateManager` to create / access resources such as
+        lookup tables.
+    """
+    pass
+
+
+def _create_categorical_column_weighted_sum(column,
+                                            transformation_cache,
+                                            state_manager,
+                                            units,
+                                            sparse_combiner,
+                                            weight_collections,
+                                            trainable,
+                                            weight_var=None):
+  # pylint: disable=g-doc-return-or-yield,g-doc-args
+  """Create a weighted sum of a categorical column for linear_model.
+
+  Note to maintainer: As implementation details, the weighted sum is
+  implemented via embedding_lookup_sparse toward efficiency. Mathematically,
+  they are the same.
+
+  To be specific, conceptually, categorical column can be treated as multi-hot
+  vector. Say:
+
+  ```python
+    x = [0 0 1]  # categorical column input
+    w = [a b c]  # weights
+  ```
+  The weighted sum is `c` in this case, which is same as `w[2]`.
+
+  Another example is
+
+  ```python
+    x = [0 1 1]  # categorical column input
+    w = [a b c]  # weights
+  ```
+  The weighted sum is `b + c` in this case, which is same as `w[2] + w[3]`.
+
+  For both cases, we can implement weighted sum via embedding_lookup with
+  sparse_combiner = "sum".
+  """
+
+  sparse_tensors = column.get_sparse_tensors(transformation_cache,
+                                             state_manager)
+  id_tensor = sparse_ops.sparse_reshape(sparse_tensors.id_tensor, [
+      array_ops.shape(sparse_tensors.id_tensor)[0], -1
+  ])
+  weight_tensor = sparse_tensors.weight_tensor
+  if weight_tensor is not None:
+    weight_tensor = sparse_ops.sparse_reshape(
+        weight_tensor, [array_ops.shape(weight_tensor)[0], -1])
+
+  if weight_var is not None:
+    weight = weight_var
+  else:
+    weight = variable_scope.get_variable(
+        name='weights',
+        shape=(column.num_buckets, units),
+        initializer=init_ops.zeros_initializer(),
+        trainable=trainable,
+        collections=weight_collections)
+  return _safe_embedding_lookup_sparse(
+      weight,
+      id_tensor,
+      sparse_weights=weight_tensor,
+      combiner=sparse_combiner,
+      name='weighted_sum')
+
+
+class SequenceDenseColumn(FeatureColumn):
+  """Represents dense sequence data."""
+
+  __metaclass__ = abc.ABCMeta
+
+  TensorSequenceLengthPair = collections.namedtuple(  # pylint: disable=invalid-name
+      'TensorSequenceLengthPair', ('dense_tensor', 'sequence_length'))
+
+  @abc.abstractmethod
+  def get_sequence_dense_tensor(self, transformation_cache, state_manager):
+    """Returns a `TensorSequenceLengthPair`.
+
+    Args:
+      transformation_cache: A `FeatureTransformationCache` object to access
+        features.
+      state_manager: A `StateManager` to create / access resources such as
+        lookup tables.
+    """
+    pass
+
+
+class FeatureTransformationCache(object):
+  """Handles caching of transformations while building the model.
+
+  `FeatureColumn` specifies how to digest an input column to the network. Some
+  feature columns require data transformations. This class caches those
+  transformations.
+
+  Some features may be used in more than one place. For example, one can use a
+  bucketized feature by itself and a cross with it. In that case we
+  should create only one bucketization op instead of creating ops for each
+  feature column separately. To handle re-use of transformed columns,
+  `FeatureTransformationCache` caches all previously transformed columns.
+
+  Example:
+  We're trying to use the following `FeatureColumn`s:
+
+  ```python
+  bucketized_age = fc.bucketized_column(fc.numeric_column("age"), ...)
+  keywords = fc.categorical_column_with_hash_buckets("keywords", ...)
+  age_X_keywords = fc.crossed_column([bucketized_age, "keywords"])
+  ... = linear_model(features,
+                          [bucketized_age, keywords, age_X_keywords]
+  ```
+
+  If we transform each column independently, then we'll get duplication of
+  bucketization (one for cross, one for bucketization itself).
+  The `FeatureTransformationCache` eliminates this duplication.
+  """
+
+  def __init__(self, features):
+    """Creates a `FeatureTransformationCache`.
+
+    Args:
+      features: A mapping from feature column to objects that are `Tensor` or
+        `SparseTensor`, or can be converted to same via
+        `sparse_tensor.convert_to_tensor_or_sparse_tensor`. A `string` key
+        signifies a base feature (not-transformed). A `FeatureColumn` key
+        means that this `Tensor` is the output of an existing `FeatureColumn`
+        which can be reused.
+    """
+    self._features = features.copy()
+    self._feature_tensors = {}
+
+  def get(self, key, state_manager):
+    """Returns a `Tensor` for the given key.
+
+    A `str` key is used to access a base feature (not-transformed). When a
+    `FeatureColumn` is passed, the transformed feature is returned if it
+    already exists, otherwise the given `FeatureColumn` is asked to provide its
+    transformed output, which is then cached.
+
+    Args:
+      key: a `str` or a `FeatureColumn`.
+      state_manager: A StateManager object that holds the FeatureColumn state.
+
+    Returns:
+      The transformed `Tensor` corresponding to the `key`.
+
+    Raises:
+      ValueError: if key is not found or a transformed `Tensor` cannot be
+        computed.
+    """
+    if key in self._feature_tensors:
+      # FeatureColumn is already transformed or converted.
+      return self._feature_tensors[key]
+
+    if key in self._features:
+      feature_tensor = self._get_raw_feature_as_tensor(key)
+      self._feature_tensors[key] = feature_tensor
+      return feature_tensor
+
+    if isinstance(key, six.string_types):
+      raise ValueError('Feature {} is not in features dictionary.'.format(key))
+
+    if not isinstance(key, FeatureColumn):
+      raise TypeError('"key" must be either a "str" or "FeatureColumn". '
+                      'Provided: {}'.format(key))
+
+    column = key
+    logging.debug('Transforming feature_column %s.', column)
+    transformed = column.transform_feature(self, state_manager)
+    if transformed is None:
+      raise ValueError('Column {} is not supported.'.format(column.name))
+    self._feature_tensors[column] = transformed
+    return transformed
+
+  def _get_raw_feature_as_tensor(self, key):
+    """Gets the raw_feature (keyed by `key`) as `tensor`.
+
+    The raw feature is converted to (sparse) tensor and maybe expand dim.
+
+    For both `Tensor` and `SparseTensor`, the rank will be expanded (to 2) if
+    the rank is 1. This supports dynamic rank also. For rank 0 raw feature, will
+    error out as it is not supported.
+
+    Args:
+      key: A `str` key to access the raw feature.
+
+    Returns:
+      A `Tensor` or `SparseTensor`.
+
+    Raises:
+      ValueError: if the raw feature has rank 0.
+    """
+    raw_feature = self._features[key]
+    feature_tensor = sparse_tensor_lib.convert_to_tensor_or_sparse_tensor(
+        raw_feature)
+
+    def expand_dims(input_tensor):
+      # Input_tensor must have rank 1.
+      if isinstance(input_tensor, sparse_tensor_lib.SparseTensor):
+        return sparse_ops.sparse_reshape(
+            input_tensor, [array_ops.shape(input_tensor)[0], -1])
+      else:
+        return array_ops.expand_dims(input_tensor, -1)
+
+    rank = feature_tensor.get_shape().ndims
+    if rank is not None:
+      if rank == 0:
+        raise ValueError(
+            'Feature (key: {}) cannot have rank 0. Give: {}'.format(
+                key, feature_tensor))
+      return feature_tensor if rank != 1 else expand_dims(feature_tensor)
+
+    # Handle dynamic rank.
+    with ops.control_dependencies([
+        check_ops.assert_positive(
+            array_ops.rank(feature_tensor),
+            message='Feature (key: {}) cannot have rank 0. Given: {}'.format(
+                key, feature_tensor))]):
+      return control_flow_ops.cond(
+          math_ops.equal(1, array_ops.rank(feature_tensor)),
+          lambda: expand_dims(feature_tensor),
+          lambda: feature_tensor)
+
+
+# TODO(ptucker): Move to third_party/tensorflow/python/ops/sparse_ops.py
+def _shape_offsets(shape):
+  """Returns moving offset for each dimension given shape."""
+  offsets = []
+  for dim in reversed(shape):
+    if offsets:
+      offsets.append(dim * offsets[-1])
+    else:
+      offsets.append(dim)
+  offsets.reverse()
+  return offsets
+
+
+# TODO(ptucker): Move to third_party/tensorflow/python/ops/sparse_ops.py
+def _to_sparse_input_and_drop_ignore_values(input_tensor, ignore_value=None):
+  """Converts a `Tensor` to a `SparseTensor`, dropping ignore_value cells.
+
+  If `input_tensor` is already a `SparseTensor`, just return it.
+
+  Args:
+    input_tensor: A string or integer `Tensor`.
+    ignore_value: Entries in `dense_tensor` equal to this value will be
+      absent from the resulting `SparseTensor`. If `None`, default value of
+      `dense_tensor`'s dtype will be used ('' for `str`, -1 for `int`).
+
+  Returns:
+    A `SparseTensor` with the same shape as `input_tensor`.
+
+  Raises:
+    ValueError: when `input_tensor`'s rank is `None`.
+  """
+  input_tensor = sparse_tensor_lib.convert_to_tensor_or_sparse_tensor(
+      input_tensor)
+  if isinstance(input_tensor, sparse_tensor_lib.SparseTensor):
+    return input_tensor
+  with ops.name_scope(None, 'to_sparse_input', (input_tensor, ignore_value,)):
+    if ignore_value is None:
+      if input_tensor.dtype == dtypes.string:
+        # Exception due to TF strings are converted to numpy objects by default.
+        ignore_value = ''
+      elif input_tensor.dtype.is_integer:
+        ignore_value = -1  # -1 has a special meaning of missing feature
+      else:
+        # NOTE: `as_numpy_dtype` is a property, so with the parentheses this is
+        # constructing a new numpy object of the given type, which yields the
+        # default value for that type.
+        ignore_value = input_tensor.dtype.as_numpy_dtype()
+    ignore_value = math_ops.cast(
+        ignore_value, input_tensor.dtype, name='ignore_value')
+    indices = array_ops.where(
+        math_ops.not_equal(input_tensor, ignore_value), name='indices')
+    return sparse_tensor_lib.SparseTensor(
+        indices=indices,
+        values=array_ops.gather_nd(input_tensor, indices, name='values'),
+        dense_shape=array_ops.shape(
+            input_tensor, out_type=dtypes.int64, name='dense_shape'))
+
+
+def _normalize_feature_columns(feature_columns):
+  """Normalizes the `feature_columns` input.
+
+  This method converts the `feature_columns` to list type as best as it can. In
+  addition, verifies the type and other parts of feature_columns, required by
+  downstream library.
+
+  Args:
+    feature_columns: The raw feature columns, usually passed by users.
+
+  Returns:
+    The normalized feature column list.
+
+  Raises:
+    ValueError: for any invalid inputs, such as empty, duplicated names, etc.
+  """
+  if isinstance(feature_columns, FeatureColumn):
+    feature_columns = [feature_columns]
+
+  if isinstance(feature_columns, collections.Iterator):
+    feature_columns = list(feature_columns)
+
+  if isinstance(feature_columns, dict):
+    raise ValueError('Expected feature_columns to be iterable, found dict.')
+
+  for column in feature_columns:
+    if not isinstance(column, FeatureColumn):
+      raise ValueError('Items of feature_columns must be a FeatureColumn. '
+                       'Given (type {}): {}.'.format(type(column), column))
+  if not feature_columns:
+    raise ValueError('feature_columns must not be empty.')
+  name_to_column = dict()
+  for column in feature_columns:
+    if column.name in name_to_column:
+      raise ValueError('Duplicate feature column name found for columns: {} '
+                       'and {}. This usually means that these columns refer to '
+                       'same base feature. Either one must be discarded or a '
+                       'duplicated but renamed item must be inserted in '
+                       'features dict.'.format(column,
+                                               name_to_column[column.name]))
+    name_to_column[column.name] = column
+
+  return feature_columns
+
+
+class NumericColumn(
+    DenseColumn,
+    collections.namedtuple(
+        'NumericColumn',
+        ('key', 'shape', 'default_value', 'dtype', 'normalizer_fn'))):
+  """see `numeric_column`."""
+
+  @property
+  def name(self):
+    """See `FeatureColumn` base class."""
+    return self.key
+
+  @property
+  def parse_example_spec(self):
+    """See `FeatureColumn` base class."""
+    return {
+        self.key:
+            parsing_ops.FixedLenFeature(self.shape, self.dtype,
+                                        self.default_value)
+    }
+
+  def transform_feature(self, transformation_cache, state_manager):
+    """See `FeatureColumn` base class.
+
+    In this case, we apply the `normalizer_fn` to the input tensor.
+
+    Args:
+      transformation_cache: A `FeatureTransformationCache` object to access
+        features.
+      state_manager: A `StateManager` to create / access resources such as
+        lookup tables.
+
+    Returns:
+      Normalized input tensor.
+    Raises:
+      ValueError: If a SparseTensor is passed in.
+    """
+    input_tensor = transformation_cache.get(self.key, state_manager)
+    if isinstance(input_tensor, sparse_tensor_lib.SparseTensor):
+      raise ValueError(
+          'The corresponding Tensor of numerical column must be a Tensor. '
+          'SparseTensor is not supported. key: {}'.format(self.key))
+    if self.normalizer_fn is not None:
+      input_tensor = self.normalizer_fn(input_tensor)
+    return math_ops.to_float(input_tensor)
+
+  @property
+  def variable_shape(self):
+    """See `DenseColumn` base class."""
+    return tensor_shape.TensorShape(self.shape)
+
+  def get_dense_tensor(self, transformation_cache, state_manager):
+    """Returns dense `Tensor` representing numeric feature.
+
+    Args:
+      transformation_cache: A `FeatureTransformationCache` object to access
+        features.
+      state_manager: A `StateManager` to create / access resources such as
+        lookup tables.
+
+    Returns:
+      Dense `Tensor` created within `transform_feature`.
+    """
+    # Feature has been already transformed. Return the intermediate
+    # representation created by _transform_feature.
+    return transformation_cache.get(self, state_manager)
+
+
+class BucketizedColumn(DenseColumn, CategoricalColumn,
+                       collections.namedtuple('BucketizedColumn',
+                                              ('source_column', 'boundaries'))):
+  """See `bucketized_column`."""
+
+  @property
+  def name(self):
+    """See `FeatureColumn` base class."""
+    return '{}_bucketized'.format(self.source_column.name)
+
+  @property
+  def parse_example_spec(self):
+    """See `FeatureColumn` base class."""
+    return self.source_column.parse_example_spec
+
+  def transform_feature(self, transformation_cache, state_manager):
+    """Returns bucketized categorical `source_column` tensor."""
+    source_tensor = transformation_cache.get(self.source_column, state_manager)
+    return math_ops._bucketize(  # pylint: disable=protected-access
+        source_tensor,
+        boundaries=self.boundaries)
+
+  @property
+  def variable_shape(self):
+    """See `DenseColumn` base class."""
+    return tensor_shape.TensorShape(
+        tuple(self.source_column.shape) + (len(self.boundaries) + 1,))
+
+  def get_dense_tensor(self, transformation_cache, state_manager):
+    """Returns one hot encoded dense `Tensor`."""
+    input_tensor = transformation_cache.get(self, state_manager)
+    return array_ops.one_hot(
+        indices=math_ops.to_int64(input_tensor),
+        depth=len(self.boundaries) + 1,
+        on_value=1.,
+        off_value=0.)
+
+  @property
+  def num_buckets(self):
+    """See `CategoricalColumn` base class."""
+    # By construction, source_column is always one-dimensional.
+    return (len(self.boundaries) + 1) * self.source_column.shape[0]
+
+  def get_sparse_tensors(self, transformation_cache, state_manager):
+    """Converts dense inputs to SparseTensor so downstream code can use it."""
+    input_tensor = transformation_cache.get(self, state_manager)
+    batch_size = array_ops.shape(input_tensor)[0]
+    # By construction, source_column is always one-dimensional.
+    source_dimension = self.source_column.shape[0]
+
+    i1 = array_ops.reshape(
+        array_ops.tile(
+            array_ops.expand_dims(math_ops.range(0, batch_size), 1),
+            [1, source_dimension]),
+        (-1,))
+    i2 = array_ops.tile(math_ops.range(0, source_dimension), [batch_size])
+    # Flatten the bucket indices and unique them across dimensions
+    # E.g. 2nd dimension indices will range from k to 2*k-1 with k buckets
+    bucket_indices = (
+        array_ops.reshape(input_tensor, (-1,)) +
+        (len(self.boundaries) + 1) * i2)
+
+    indices = math_ops.to_int64(array_ops.transpose(array_ops.stack((i1, i2))))
+    dense_shape = math_ops.to_int64(array_ops.stack(
+        [batch_size, source_dimension]))
+    sparse_tensor = sparse_tensor_lib.SparseTensor(
+        indices=indices,
+        values=bucket_indices,
+        dense_shape=dense_shape)
+    return CategoricalColumn.IdWeightPair(sparse_tensor, None)
+
+
+class EmbeddingColumn(
+    DenseColumn, SequenceDenseColumn,
+    collections.namedtuple(
+        'EmbeddingColumn',
+        ('categorical_column', 'dimension', 'combiner', 'initializer',
+         'ckpt_to_load_from', 'tensor_name_in_ckpt', 'max_norm', 'trainable'))):
+  """See `embedding_column`."""
+
+  @property
+  def name(self):
+    """See `FeatureColumn` base class."""
+    return '{}_embedding'.format(self.categorical_column.name)
+
+  @property
+  def parse_example_spec(self):
+    """See `FeatureColumn` base class."""
+    return self.categorical_column.parse_example_spec
+
+  def transform_feature(self, transformation_cache, state_manager):
+    """Transforms underlying `categorical_column`."""
+    return transformation_cache.get(self.categorical_column, state_manager)
+
+  @property
+  def variable_shape(self):
+    """See `DenseColumn` base class."""
+    return tensor_shape.vector(self.dimension)
+
+  def _get_dense_tensor_internal(self, transformation_cache, state_manager):
+    """Private method that follows the signature of _get_dense_tensor."""
+    # Get sparse IDs and weights.
+    sparse_tensors = self.categorical_column.get_sparse_tensors(
+        transformation_cache, state_manager)
+    sparse_ids = sparse_tensors.id_tensor
+    sparse_weights = sparse_tensors.weight_tensor
+
+    embedding_shape = (self.categorical_column.num_buckets, self.dimension)
+    embedding_weights = state_manager.get_variable(
+        self,
+        name='embedding_weights',
+        shape=embedding_shape,
+        dtype=dtypes.float32,
+        initializer=self.initializer)
+
+    if self.ckpt_to_load_from is not None:
+      to_restore = embedding_weights
+      if isinstance(to_restore, variables.PartitionedVariable):
+        to_restore = to_restore._get_variable_list()  # pylint: disable=protected-access
+      checkpoint_utils.init_from_checkpoint(self.ckpt_to_load_from, {
+          self.tensor_name_in_ckpt: to_restore
+      })
+
+    # Return embedding lookup result.
+    return _safe_embedding_lookup_sparse(
+        embedding_weights=embedding_weights,
+        sparse_ids=sparse_ids,
+        sparse_weights=sparse_weights,
+        combiner=self.combiner,
+        name='%s_weights' % self.name,
+        max_norm=self.max_norm)
+
+  def get_dense_tensor(self, transformation_cache, state_manager):
+    """Returns tensor after doing the embedding lookup.
+
+    Args:
+      transformation_cache: A `FeatureTransformationCache` object to access
+        features.
+      state_manager: A `StateManager` to create / access resources such as
+        lookup tables.
+
+    Returns:
+      Embedding lookup tensor.
+
+    Raises:
+      ValueError: `categorical_column` is SequenceCategoricalColumn.
+    """
+    if isinstance(self.categorical_column, SequenceCategoricalColumn):
+      raise ValueError(
+          'In embedding_column: {}. '
+          'categorical_column must not be of type SequenceCategoricalColumn. '
+          'Suggested fix A: If you wish to use input_layer, use a '
+          'non-sequence categorical_column_with_*. '
+          'Suggested fix B: If you wish to create sequence input, use '
+          'sequence_input_layer instead of input_layer. '
+          'Given (type {}): {}'.format(self.name, type(self.categorical_column),
+                                       self.categorical_column))
+    return self._get_dense_tensor_internal(transformation_cache, state_manager)
+
+  def get_sequence_dense_tensor(self, transformation_cache, state_manager):
+    """See `SequenceDenseColumn` base class."""
+    if not isinstance(self.categorical_column, SequenceCategoricalColumn):
+      raise ValueError(
+          'In embedding_column: {}. '
+          'categorical_column must be of type SequenceCategoricalColumn '
+          'to use sequence_input_layer. '
+          'Suggested fix: Use one of sequence_categorical_column_with_*. '
+          'Given (type {}): {}'.format(self.name, type(self.categorical_column),
+                                       self.categorical_column))
+    dense_tensor = self._get_dense_tensor_internal(  # pylint: disable=protected-access
+        transformation_cache, state_manager)
+    sparse_tensors = self.categorical_column.get_sparse_tensors(
+        transformation_cache, state_manager)
+    sequence_length = _sequence_length_from_sparse_tensor(
+        sparse_tensors.id_tensor)
+    return SequenceDenseColumn.TensorSequenceLengthPair(
+        dense_tensor=dense_tensor, sequence_length=sequence_length)
+
+
+def _get_graph_for_variable(var):
+  if isinstance(var, variables.PartitionedVariable):
+    return list(var)[0].graph
+  else:
+    return var.graph
+
+
+class SharedEmbeddingColumn(
+    DenseColumn, SequenceDenseColumn,
+    collections.namedtuple(
+        'SharedEmbeddingColumn',
+        ('categorical_column', 'dimension', 'combiner', 'initializer',
+         'shared_embedding_collection_name', 'ckpt_to_load_from',
+         'tensor_name_in_ckpt', 'max_norm', 'trainable'))):
+  """See `embedding_column`."""
+
+  @property
+  def name(self):
+    """See `FeatureColumn` base class."""
+    return '{}_shared_embedding'.format(self.categorical_column.name)
+
+  @property
+  def shared_collection_name(self):
+    """Returns the shared name of this column.
+
+    A group of columns share an embedding. Each one of those columns would have
+    the same `shared_collection_name` by which they could be collectively
+    referred to.
+    """
+    return self.shared_embedding_collection_name
+
+  @property
+  def parse_example_spec(self):
+    """See `FeatureColumn` base class."""
+    return self.categorical_column.parse_example_spec
+
+  def transform_feature(self, transformation_cache, state_manager):
+    """See `FeatureColumn` base class."""
+    return transformation_cache.get(self.categorical_column, state_manager)
+
+  @property
+  def variable_shape(self):
+    """See `DenseColumn` base class."""
+    return tensor_shape.vector(self.dimension)
+
+  def _get_dense_tensor_internal(self, transformation_cache, state_manager):
+    """Private method that follows the signature of _get_dense_tensor."""
+    # This method is called from a variable_scope with name _var_scope_name,
+    # which is shared among all shared embeddings. Open a name_scope here, so
+    # that the ops for different columns have distinct names.
+    with ops.name_scope(None, default_name=self.name):
+      # Get sparse IDs and weights.
+      sparse_tensors = self.categorical_column.get_sparse_tensors(
+          transformation_cache, state_manager)
+      sparse_ids = sparse_tensors.id_tensor
+      sparse_weights = sparse_tensors.weight_tensor
+
+      embedding_shape = (self.categorical_column.num_buckets, self.dimension)
+      embedding_weights = state_manager.get_variable(
+          self,
+          name='embedding_weights',
+          shape=embedding_shape,
+          dtype=dtypes.float32,
+          initializer=self.initializer)
+
+      if self.ckpt_to_load_from is not None:
+        to_restore = embedding_weights
+        if isinstance(to_restore, variables.PartitionedVariable):
+          to_restore = to_restore._get_variable_list()  # pylint: disable=protected-access
+        checkpoint_utils.init_from_checkpoint(self.ckpt_to_load_from, {
+            self.tensor_name_in_ckpt: to_restore
+        })
+
+      # Return embedding lookup result.
+      return _safe_embedding_lookup_sparse(
+          embedding_weights=embedding_weights,
+          sparse_ids=sparse_ids,
+          sparse_weights=sparse_weights,
+          combiner=self.combiner,
+          name='%s_weights' % self.name,
+          max_norm=self.max_norm)
+
+  def get_dense_tensor(self, transformation_cache, state_manager):
+    """Returns the embedding lookup result."""
+    if isinstance(self.categorical_column, SequenceCategoricalColumn):
+      raise ValueError(
+          'In embedding_column: {}. '
+          'categorical_column must not be of type SequenceCategoricalColumn. '
+          'Suggested fix A: If you wish to use input_layer, use a '
+          'non-sequence categorical_column_with_*. '
+          'Suggested fix B: If you wish to create sequence input, use '
+          'sequence_input_layer instead of input_layer. '
+          'Given (type {}): {}'.format(self.name, type(self.categorical_column),
+                                       self.categorical_column))
+    return self._get_dense_tensor_internal(transformation_cache, state_manager)
+
+  def get_sequence_dense_tensor(self, transformation_cache, state_manager):
+    """See `SequenceDenseColumn` base class."""
+    if not isinstance(self.categorical_column, SequenceCategoricalColumn):
+      raise ValueError(
+          'In embedding_column: {}. '
+          'categorical_column must be of type SequenceCategoricalColumn '
+          'to use sequence_input_layer. '
+          'Suggested fix: Use one of sequence_categorical_column_with_*. '
+          'Given (type {}): {}'.format(self.name, type(self.categorical_column),
+                                       self.categorical_column))
+    dense_tensor = self.get_dense_tensor_internal(transformation_cache,
+                                                  state_manager)
+    sparse_tensors = self.categorical_column.get_sparse_tensors(
+        transformation_cache, state_manager)
+    sequence_length = _sequence_length_from_sparse_tensor(
+        sparse_tensors.id_tensor)
+    return SequenceDenseColumn.TensorSequenceLengthPair(
+        dense_tensor=dense_tensor, sequence_length=sequence_length)
+
+
+def _create_tuple(shape, value):
+  """Returns a tuple with given shape and filled with value."""
+  if shape:
+    return tuple([_create_tuple(shape[1:], value) for _ in range(shape[0])])
+  return value
+
+
+def _as_tuple(value):
+  if not nest.is_sequence(value):
+    return value
+  return tuple([_as_tuple(v) for v in value])
+
+
+def _check_shape(shape, key):
+  """Returns shape if it's valid, raises error otherwise."""
+  assert shape is not None
+  if not nest.is_sequence(shape):
+    shape = [shape]
+  shape = tuple(shape)
+  for dimension in shape:
+    if not isinstance(dimension, int):
+      raise TypeError('shape dimensions must be integer. '
+                      'shape: {}, key: {}'.format(shape, key))
+    if dimension < 1:
+      raise ValueError('shape dimensions must be greater than 0. '
+                       'shape: {}, key: {}'.format(shape, key))
+  return shape
+
+
+def _is_shape_and_default_value_compatible(default_value, shape):
+  """Verifies compatibility of shape and default_value."""
+  # Invalid condition:
+  #  * if default_value is not a scalar and shape is empty
+  #  * or if default_value is an iterable and shape is not empty
+  if nest.is_sequence(default_value) != bool(shape):
+    return False
+  if not shape:
+    return True
+  if len(default_value) != shape[0]:
+    return False
+  for i in range(shape[0]):
+    if not _is_shape_and_default_value_compatible(default_value[i], shape[1:]):
+      return False
+  return True
+
+
+def _check_default_value(shape, default_value, dtype, key):
+  """Returns default value as tuple if it's valid, otherwise raises errors.
+
+  This function verifies that `default_value` is compatible with both `shape`
+  and `dtype`. If it is not compatible, it raises an error. If it is compatible,
+  it casts default_value to a tuple and returns it. `key` is used only
+  for error message.
+
+  Args:
+    shape: An iterable of integers specifies the shape of the `Tensor`.
+    default_value: If a single value is provided, the same value will be applied
+      as the default value for every item. If an iterable of values is
+      provided, the shape of the `default_value` should be equal to the given
+      `shape`.
+    dtype: defines the type of values. Default value is `tf.float32`. Must be a
+      non-quantized, real integer or floating point type.
+    key: Column name, used only for error messages.
+
+  Returns:
+    A tuple which will be used as default value.
+
+  Raises:
+    TypeError: if `default_value` is an iterable but not compatible with `shape`
+    TypeError: if `default_value` is not compatible with `dtype`.
+    ValueError: if `dtype` is not convertible to `tf.float32`.
+  """
+  if default_value is None:
+    return None
+
+  if isinstance(default_value, int):
+    return _create_tuple(shape, default_value)
+
+  if isinstance(default_value, float) and dtype.is_floating:
+    return _create_tuple(shape, default_value)
+
+  if callable(getattr(default_value, 'tolist', None)):  # Handles numpy arrays
+    default_value = default_value.tolist()
+
+  if nest.is_sequence(default_value):
+    if not _is_shape_and_default_value_compatible(default_value, shape):
+      raise ValueError(
+          'The shape of default_value must be equal to given shape. '
+          'default_value: {}, shape: {}, key: {}'.format(
+              default_value, shape, key))
+    # Check if the values in the list are all integers or are convertible to
+    # floats.
+    is_list_all_int = all(
+        isinstance(v, int) for v in nest.flatten(default_value))
+    is_list_has_float = any(
+        isinstance(v, float) for v in nest.flatten(default_value))
+    if is_list_all_int:
+      return _as_tuple(default_value)
+    if is_list_has_float and dtype.is_floating:
+      return _as_tuple(default_value)
+  raise TypeError('default_value must be compatible with dtype. '
+                  'default_value: {}, dtype: {}, key: {}'.format(
+                      default_value, dtype, key))
+
+
+class HashedCategoricalColumn(
+    CategoricalColumn,
+    collections.namedtuple('HashedCategoricalColumn',
+                           ('key', 'hash_bucket_size', 'dtype'))):
+  """see `categorical_column_with_hash_bucket`."""
+
+  @property
+  def name(self):
+    """See `FeatureColumn` base class."""
+    return self.key
+
+  @property
+  def parse_example_spec(self):
+    """See `FeatureColumn` base class."""
+    return {self.key: parsing_ops.VarLenFeature(self.dtype)}
+
+  def transform_feature(self, transformation_cache, state_manager):
+    """Hashes the values in the feature_column."""
+    input_tensor = _to_sparse_input_and_drop_ignore_values(
+        transformation_cache.get(self.key, state_manager))
+    if not isinstance(input_tensor, sparse_tensor_lib.SparseTensor):
+      raise ValueError('SparseColumn input must be a SparseTensor.')
+
+    _assert_string_or_int(
+        input_tensor.dtype,
+        prefix='column_name: {} input_tensor'.format(self.key))
+
+    if self.dtype.is_integer != input_tensor.dtype.is_integer:
+      raise ValueError(
+          'Column dtype and SparseTensors dtype must be compatible. '
+          'key: {}, column dtype: {}, tensor dtype: {}'.format(
+              self.key, self.dtype, input_tensor.dtype))
+
+    if self.dtype == dtypes.string:
+      sparse_values = input_tensor.values
+    else:
+      sparse_values = string_ops.as_string(input_tensor.values)
+
+    sparse_id_values = string_ops.string_to_hash_bucket_fast(
+        sparse_values, self.hash_bucket_size, name='lookup')
+    return sparse_tensor_lib.SparseTensor(
+        input_tensor.indices, sparse_id_values, input_tensor.dense_shape)
+
+  @property
+  def num_buckets(self):
+    """Returns number of buckets in this sparse feature."""
+    return self.hash_bucket_size
+
+  def get_sparse_tensors(self, transformation_cache, state_manager):
+    """See `CategoricalColumn` base class."""
+    return CategoricalColumn.IdWeightPair(
+        transformation_cache.get(self, state_manager), None)
+
+
+class VocabularyFileCategoricalColumn(
+    CategoricalColumn,
+    collections.namedtuple('VocabularyFileCategoricalColumn',
+                           ('key', 'vocabulary_file', 'vocabulary_size',
+                            'num_oov_buckets', 'dtype', 'default_value'))):
+  """See `categorical_column_with_vocabulary_file`."""
+
+  @property
+  def name(self):
+    """See `FeatureColumn` base class."""
+    return self.key
+
+  @property
+  def parse_example_spec(self):
+    """See `FeatureColumn` base class."""
+    return {self.key: parsing_ops.VarLenFeature(self.dtype)}
+
+  def transform_feature(self, transformation_cache, state_manager):
+    """Creates a lookup table for the vocabulary."""
+    input_tensor = _to_sparse_input_and_drop_ignore_values(
+        transformation_cache.get(self.key, state_manager))
+
+    if self.dtype.is_integer != input_tensor.dtype.is_integer:
+      raise ValueError(
+          'Column dtype and SparseTensors dtype must be compatible. '
+          'key: {}, column dtype: {}, tensor dtype: {}'.format(
+              self.key, self.dtype, input_tensor.dtype))
+
+    _assert_string_or_int(
+        input_tensor.dtype,
+        prefix='column_name: {} input_tensor'.format(self.key))
+
+    key_dtype = self.dtype
+    if input_tensor.dtype.is_integer:
+      # `index_table_from_file` requires 64-bit integer keys.
+      key_dtype = dtypes.int64
+      input_tensor = math_ops.to_int64(input_tensor)
+
+    # TODO(rohanj): Use state manager to manage the index table creation.
+    return lookup_ops.index_table_from_file(
+        vocabulary_file=self.vocabulary_file,
+        num_oov_buckets=self.num_oov_buckets,
+        vocab_size=self.vocabulary_size,
+        default_value=self.default_value,
+        key_dtype=key_dtype,
+        name='{}_lookup'.format(self.key)).lookup(input_tensor)
+
+  @property
+  def num_buckets(self):
+    """Returns number of buckets in this sparse feature."""
+    return self.vocabulary_size + self.num_oov_buckets
+
+  def get_sparse_tensors(self, transformation_cache, state_manager):
+    """See `CategoricalColumn` base class."""
+    return CategoricalColumn.IdWeightPair(
+        transformation_cache.get(self, state_manager), None)
+
+
+class VocabularyListCategoricalColumn(
+    CategoricalColumn,
+    collections.namedtuple(
+        'VocabularyListCategoricalColumn',
+        ('key', 'vocabulary_list', 'dtype', 'default_value', 'num_oov_buckets'))
+):
+  """See `categorical_column_with_vocabulary_list`."""
+
+  @property
+  def name(self):
+    """See `FeatureColumn` base class."""
+    return self.key
+
+  @property
+  def parse_example_spec(self):
+    """See `FeatureColumn` base class."""
+    return {self.key: parsing_ops.VarLenFeature(self.dtype)}
+
+  def transform_feature(self, transformation_cache, state_manager):
+    """Creates a lookup table for the vocabulary list."""
+    input_tensor = _to_sparse_input_and_drop_ignore_values(
+        transformation_cache.get(self.key, state_manager))
+
+    if self.dtype.is_integer != input_tensor.dtype.is_integer:
+      raise ValueError(
+          'Column dtype and SparseTensors dtype must be compatible. '
+          'key: {}, column dtype: {}, tensor dtype: {}'.format(
+              self.key, self.dtype, input_tensor.dtype))
+
+    _assert_string_or_int(
+        input_tensor.dtype,
+        prefix='column_name: {} input_tensor'.format(self.key))
+
+    key_dtype = self.dtype
+    if input_tensor.dtype.is_integer:
+      # `index_table_from_tensor` requires 64-bit integer keys.
+      key_dtype = dtypes.int64
+      input_tensor = math_ops.to_int64(input_tensor)
+
+    # TODO(rohanj): Use state manager to manage the index table creation.
+    return lookup_ops.index_table_from_tensor(
+        vocabulary_list=tuple(self.vocabulary_list),
+        default_value=self.default_value,
+        num_oov_buckets=self.num_oov_buckets,
+        dtype=key_dtype,
+        name='{}_lookup'.format(self.key)).lookup(input_tensor)
+
+  @property
+  def num_buckets(self):
+    """Returns number of buckets in this sparse feature."""
+    return len(self.vocabulary_list) + self.num_oov_buckets
+
+  def get_sparse_tensors(self, transformation_cache, state_manager):
+    """See `CategoricalColumn` base class."""
+    return CategoricalColumn.IdWeightPair(
+        transformation_cache.get(self, state_manager), None)
+
+
+class IdentityCategoricalColumn(
+    CategoricalColumn,
+    collections.namedtuple('IdentityCategoricalColumn',
+                           ('key', 'number_buckets', 'default_value'))):
+
+  """See `categorical_column_with_identity`."""
+
+  @property
+  def name(self):
+    """See `FeatureColumn` base class."""
+    return self.key
+
+  @property
+  def parse_example_spec(self):
+    """See `FeatureColumn` base class."""
+    return {self.key: parsing_ops.VarLenFeature(dtypes.int64)}
+
+  def transform_feature(self, transformation_cache, state_manager):
+    """Returns a SparseTensor with identity values."""
+    input_tensor = _to_sparse_input_and_drop_ignore_values(
+        transformation_cache.get(self.key, state_manager))
+
+    if not input_tensor.dtype.is_integer:
+      raise ValueError(
+          'Invalid input, not integer. key: {} dtype: {}'.format(
+              self.key, input_tensor.dtype))
+
+    values = math_ops.to_int64(input_tensor.values, name='values')
+    num_buckets = math_ops.to_int64(self.num_buckets, name='num_buckets')
+    zero = math_ops.to_int64(0, name='zero')
+    if self.default_value is None:
+      # Fail if values are out-of-range.
+      assert_less = check_ops.assert_less(
+          values, num_buckets, data=(values, num_buckets),
+          name='assert_less_than_num_buckets')
+      assert_greater = check_ops.assert_greater_equal(
+          values, zero, data=(values,),
+          name='assert_greater_or_equal_0')
+      with ops.control_dependencies((assert_less, assert_greater)):
+        values = array_ops.identity(values)
+    else:
+      # Assign default for out-of-range values.
+      values = array_ops.where(
+          math_ops.logical_or(
+              values < zero, values >= num_buckets, name='out_of_range'),
+          array_ops.fill(
+              dims=array_ops.shape(values),
+              value=math_ops.to_int64(self.default_value),
+              name='default_values'),
+          values)
+
+    return sparse_tensor_lib.SparseTensor(
+        indices=input_tensor.indices,
+        values=values,
+        dense_shape=input_tensor.dense_shape)
+
+  @property
+  def num_buckets(self):
+    """Returns number of buckets in this sparse feature."""
+    return self.number_buckets
+
+  def get_sparse_tensors(self, transformation_cache, state_manager):
+    """See `CategoricalColumn` base class."""
+    return CategoricalColumn.IdWeightPair(
+        transformation_cache.get(self, state_manager), None)
+
+
+class WeightedCategoricalColumn(
+    CategoricalColumn,
+    collections.namedtuple(
+        'WeightedCategoricalColumn',
+        ('categorical_column', 'weight_feature_key', 'dtype'))):
+  """See `weighted_categorical_column`."""
+
+  @property
+  def name(self):
+    """See `FeatureColumn` base class."""
+    return '{}_weighted_by_{}'.format(
+        self.categorical_column.name, self.weight_feature_key)
+
+  @property
+  def parse_example_spec(self):
+    """See `FeatureColumn` base class."""
+    config = self.categorical_column.parse_example_spec
+    if self.weight_feature_key in config:
+      raise ValueError('Parse config {} already exists for {}.'.format(
+          config[self.weight_feature_key], self.weight_feature_key))
+    config[self.weight_feature_key] = parsing_ops.VarLenFeature(self.dtype)
+    return config
+
+  @property
+  def num_buckets(self):
+    """See `DenseColumn` base class."""
+    return self.categorical_column.num_buckets
+
+  def transform_feature(self, transformation_cache, state_manager):
+    """Applies weights to tensor generated from `categorical_column`'."""
+    weight_tensor = transformation_cache.get(self.weight_feature_key,
+                                             state_manager)
+    if weight_tensor is None:
+      raise ValueError('Missing weights {}.'.format(self.weight_feature_key))
+    weight_tensor = sparse_tensor_lib.convert_to_tensor_or_sparse_tensor(
+        weight_tensor)
+    if self.dtype != weight_tensor.dtype.base_dtype:
+      raise ValueError('Bad dtype, expected {}, but got {}.'.format(
+          self.dtype, weight_tensor.dtype))
+    if not isinstance(weight_tensor, sparse_tensor_lib.SparseTensor):
+      # The weight tensor can be a regular Tensor. In this case, sparsify it.
+      weight_tensor = _to_sparse_input_and_drop_ignore_values(
+          weight_tensor, ignore_value=0.0)
+    if not weight_tensor.dtype.is_floating:
+      weight_tensor = math_ops.to_float(weight_tensor)
+    return (transformation_cache.get(self.categorical_column, state_manager),
+            weight_tensor)
+
+  def get_sparse_tensors(self, transformation_cache, state_manager):
+    """See `CategoricalColumn` base class."""
+    tensors = transformation_cache.get(self, state_manager)
+    return CategoricalColumn.IdWeightPair(tensors[0], tensors[1])
+
+
+class CrossedColumn(
+    CategoricalColumn,
+    collections.namedtuple('CrossedColumn',
+                           ('keys', 'hash_bucket_size', 'hash_key'))):
+  """See `crossed_column`."""
+
+  @property
+  def name(self):
+    """See `FeatureColumn` base class."""
+    feature_names = []
+    for key in _collect_leaf_level_keys(self):
+      if isinstance(key, FeatureColumn):
+        feature_names.append(key.name)
+      else:  # key must be a string
+        feature_names.append(key)
+    return '_X_'.join(sorted(feature_names))
+
+  @property
+  def parse_example_spec(self):
+    """See `FeatureColumn` base class."""
+    config = {}
+    for key in self.keys:
+      if isinstance(key, FeatureColumn):
+        config.update(key.parse_example_spec)
+      else:  # key must be a string
+        config.update({key: parsing_ops.VarLenFeature(dtypes.string)})
+    return config
+
+  def transform_feature(self, transformation_cache, state_manager):
+    """Generates a hashed sparse cross from the input tensors."""
+    feature_tensors = []
+    for key in _collect_leaf_level_keys(self):
+      if isinstance(key, six.string_types):
+        feature_tensors.append(transformation_cache.get(key, state_manager))
+      elif isinstance(key, CategoricalColumn):
+        ids_and_weights = key.get_sparse_tensors(transformation_cache,
+                                                 state_manager)
+        if ids_and_weights.weight_tensor is not None:
+          raise ValueError(
+              'crossed_column does not support weight_tensor, but the given '
+              'column populates weight_tensor. '
+              'Given column: {}'.format(key.name))
+        feature_tensors.append(ids_and_weights.id_tensor)
+      else:
+        raise ValueError('Unsupported column type. Given: {}'.format(key))
+    return sparse_ops.sparse_cross_hashed(
+        inputs=feature_tensors,
+        num_buckets=self.hash_bucket_size,
+        hash_key=self.hash_key)
+
+  @property
+  def num_buckets(self):
+    """Returns number of buckets in this sparse feature."""
+    return self.hash_bucket_size
+
+  def get_sparse_tensors(self, transformation_cache, state_manager):
+    """See `CategoricalColumn` base class."""
+    return CategoricalColumn.IdWeightPair(
+        transformation_cache.get(self, state_manager), None)
+
+
+def _collect_leaf_level_keys(cross):
+  """Collects base keys by expanding all nested crosses.
+
+  Args:
+    cross: A `CrossedColumn`.
+
+  Returns:
+    A list of strings or `CategoricalColumn` instances.
+  """
+  leaf_level_keys = []
+  for k in cross.keys:
+    if isinstance(k, CrossedColumn):
+      leaf_level_keys.extend(_collect_leaf_level_keys(k))
+    else:
+      leaf_level_keys.append(k)
+  return leaf_level_keys
+
+
+# TODO(zakaria): Move this to embedding_ops and make it public.
+def _safe_embedding_lookup_sparse(embedding_weights,
+                                  sparse_ids,
+                                  sparse_weights=None,
+                                  combiner='mean',
+                                  default_id=None,
+                                  name=None,
+                                  partition_strategy='div',
+                                  max_norm=None):
+  """Lookup embedding results, accounting for invalid IDs and empty features.
+
+  The partitioned embedding in `embedding_weights` must all be the same shape
+  except for the first dimension. The first dimension is allowed to vary as the
+  vocabulary size is not necessarily a multiple of `P`.  `embedding_weights`
+  may be a `PartitionedVariable` as returned by using `tf.get_variable()` with a
+  partitioner.
+
+  Invalid IDs (< 0) are pruned from input IDs and weights, as well as any IDs
+  with non-positive weight. For an entry with no features, the embedding vector
+  for `default_id` is returned, or the 0-vector if `default_id` is not supplied.
+
+  The ids and weights may be multi-dimensional. Embeddings are always aggregated
+  along the last dimension.
+
+  Args:
+    embedding_weights:  A list of `P` float `Tensor`s or values representing
+        partitioned embedding `Tensor`s.  Alternatively, a `PartitionedVariable`
+        created by partitioning along dimension 0.  The total unpartitioned
+        shape should be `[e_0, e_1, ..., e_m]`, where `e_0` represents the
+        vocab size and `e_1, ..., e_m` are the embedding dimensions.
+    sparse_ids: `SparseTensor` of shape `[d_0, d_1, ..., d_n]` containing the
+        ids. `d_0` is typically batch size.
+    sparse_weights: `SparseTensor` of same shape as `sparse_ids`, containing
+        float weights corresponding to `sparse_ids`, or `None` if all weights
+        are be assumed to be 1.0.
+    combiner: A string specifying how to combine embedding results for each
+        entry. Currently "mean", "sqrtn" and "sum" are supported, with "mean"
+        the default.
+    default_id: The id to use for an entry with no features.
+    name: A name for this operation (optional).
+    partition_strategy: A string specifying the partitioning strategy.
+        Currently `"div"` and `"mod"` are supported. Default is `"div"`.
+    max_norm: If not `None`, all embeddings are l2-normalized to max_norm before
+        combining.
+
+
+  Returns:
+    Dense `Tensor` of shape `[d_0, d_1, ..., d_{n-1}, e_1, ..., e_m]`.
+
+  Raises:
+    ValueError: if `embedding_weights` is empty.
+  """
+  if embedding_weights is None:
+    raise ValueError('Missing embedding_weights %s.' % embedding_weights)
+  if isinstance(embedding_weights, variables.PartitionedVariable):
+    embedding_weights = list(embedding_weights)  # get underlying Variables.
+  if not isinstance(embedding_weights, list):
+    embedding_weights = [embedding_weights]
+  if len(embedding_weights) < 1:
+    raise ValueError('Missing embedding_weights %s.' % embedding_weights)
+
+  dtype = sparse_weights.dtype if sparse_weights is not None else None
+  embedding_weights = [
+      ops.convert_to_tensor(w, dtype=dtype) for w in embedding_weights
+  ]
+
+  with ops.name_scope(name, 'embedding_lookup',
+                      embedding_weights + [sparse_ids,
+                                           sparse_weights]) as scope:
+    # Reshape higher-rank sparse ids and weights to linear segment ids.
+    original_shape = sparse_ids.dense_shape
+    original_rank_dim = sparse_ids.dense_shape.get_shape()[0]
+    original_rank = (
+        array_ops.size(original_shape)
+        if original_rank_dim.value is None
+        else original_rank_dim.value)
+    sparse_ids = sparse_ops.sparse_reshape(sparse_ids, [
+        math_ops.reduce_prod(
+            array_ops.slice(original_shape, [0], [original_rank - 1])),
+        array_ops.gather(original_shape, original_rank - 1)])
+    if sparse_weights is not None:
+      sparse_weights = sparse_tensor_lib.SparseTensor(
+          sparse_ids.indices,
+          sparse_weights.values, sparse_ids.dense_shape)
+
+    # Prune invalid ids and weights.
+    sparse_ids, sparse_weights = _prune_invalid_ids(sparse_ids, sparse_weights)
+    if combiner != 'sum':
+      sparse_ids, sparse_weights = _prune_invalid_weights(
+          sparse_ids, sparse_weights)
+
+    # Fill in dummy values for empty features, if necessary.
+    sparse_ids, is_row_empty = sparse_ops.sparse_fill_empty_rows(sparse_ids,
+                                                                 default_id or
+                                                                 0)
+    if sparse_weights is not None:
+      sparse_weights, _ = sparse_ops.sparse_fill_empty_rows(sparse_weights, 1.0)
+
+    result = embedding_ops.embedding_lookup_sparse(
+        embedding_weights,
+        sparse_ids,
+        sparse_weights,
+        combiner=combiner,
+        partition_strategy=partition_strategy,
+        name=None if default_id is None else scope,
+        max_norm=max_norm)
+
+    if default_id is None:
+      # Broadcast is_row_empty to the same shape as embedding_lookup_result,
+      # for use in Select.
+      is_row_empty = array_ops.tile(
+          array_ops.reshape(is_row_empty, [-1, 1]),
+          array_ops.stack([1, array_ops.shape(result)[1]]))
+
+      result = array_ops.where(is_row_empty,
+                               array_ops.zeros_like(result),
+                               result,
+                               name=scope)
+
+    # Reshape back from linear ids back into higher-dimensional dense result.
+    final_result = array_ops.reshape(
+        result,
+        array_ops.concat([
+            array_ops.slice(
+                math_ops.cast(original_shape, dtypes.int32), [0],
+                [original_rank - 1]),
+            array_ops.slice(array_ops.shape(result), [1], [-1])
+        ], 0))
+    final_result.set_shape(tensor_shape.unknown_shape(
+        (original_rank_dim - 1).value).concatenate(result.get_shape()[1:]))
+    return final_result
+
+
+def _prune_invalid_ids(sparse_ids, sparse_weights):
+  """Prune invalid IDs (< 0) from the input ids and weights."""
+  is_id_valid = math_ops.greater_equal(sparse_ids.values, 0)
+  if sparse_weights is not None:
+    is_id_valid = math_ops.logical_and(
+        is_id_valid,
+        array_ops.ones_like(sparse_weights.values, dtype=dtypes.bool))
+  sparse_ids = sparse_ops.sparse_retain(sparse_ids, is_id_valid)
+  if sparse_weights is not None:
+    sparse_weights = sparse_ops.sparse_retain(sparse_weights, is_id_valid)
+  return sparse_ids, sparse_weights
+
+
+def _prune_invalid_weights(sparse_ids, sparse_weights):
+  """Prune invalid weights (< 0) from the input ids and weights."""
+  if sparse_weights is not None:
+    is_weights_valid = math_ops.greater(sparse_weights.values, 0)
+    sparse_ids = sparse_ops.sparse_retain(sparse_ids, is_weights_valid)
+    sparse_weights = sparse_ops.sparse_retain(sparse_weights, is_weights_valid)
+  return sparse_ids, sparse_weights
+
+
+class IndicatorColumn(DenseColumn, SequenceDenseColumn,
+                      collections.namedtuple('IndicatorColumn',
+                                             ('categorical_column'))):
+  """Represents a one-hot column for use in deep networks.
+
+  Args:
+    categorical_column: A `CategoricalColumn` which is created by
+      `categorical_column_with_*` function.
+  """
+
+  @property
+  def name(self):
+    """See `FeatureColumn` base class."""
+    return '{}_indicator'.format(self.categorical_column.name)
+
+  def transform_feature(self, transformation_cache, state_manager):
+    """Returns dense `Tensor` representing feature.
+
+    Args:
+      transformation_cache: A `FeatureTransformationCache` object to access
+        features.
+      state_manager: A `StateManager` to create / access resources such as
+        lookup tables.
+
+    Returns:
+      Transformed feature `Tensor`.
+
+    Raises:
+      ValueError: if input rank is not known at graph building time.
+    """
+    id_weight_pair = self.categorical_column.get_sparse_tensors(
+        transformation_cache, state_manager)
+    id_tensor = id_weight_pair.id_tensor
+    weight_tensor = id_weight_pair.weight_tensor
+
+    # If the underlying column is weighted, return the input as a dense tensor.
+    if weight_tensor is not None:
+      weighted_column = sparse_ops.sparse_merge(
+          sp_ids=id_tensor,
+          sp_values=weight_tensor,
+          vocab_size=int(self.variable_shape[-1]))
+      # Remove (?, -1) index
+      weighted_column = sparse_ops.sparse_slice(weighted_column, [0, 0],
+                                                weighted_column.dense_shape)
+      return sparse_ops.sparse_tensor_to_dense(weighted_column)
+
+    dense_id_tensor = sparse_ops.sparse_tensor_to_dense(
+        id_tensor, default_value=-1)
+
+    # One hot must be float for tf.concat reasons since all other inputs to
+    # input_layer are float32.
+    one_hot_id_tensor = array_ops.one_hot(
+        dense_id_tensor,
+        depth=self.variable_shape[-1],
+        on_value=1.0,
+        off_value=0.0)
+
+    # Reduce to get a multi-hot per example.
+    return math_ops.reduce_sum(one_hot_id_tensor, axis=[-2])
+
+  @property
+  def parse_example_spec(self):
+    """See `FeatureColumn` base class."""
+    return self.categorical_column.parse_example_spec
+
+  @property
+  def variable_shape(self):
+    """Returns a `TensorShape` representing the shape of the dense `Tensor`."""
+    return tensor_shape.TensorShape([1, self.categorical_column.num_buckets])
+
+  def get_dense_tensor(self, transformation_cache, state_manager):
+    """Returns dense `Tensor` representing feature.
+
+    Args:
+      transformation_cache: A `FeatureTransformationCache` object to access
+        features.
+      state_manager: A `StateManager` to create / access resources such as
+        lookup tables.
+
+    Returns:
+      Dense `Tensor` created within `transform_feature`.
+
+    Raises:
+      ValueError: If `categorical_column` is a `SequenceCategoricalColumn`.
+    """
+    if isinstance(self.categorical_column, SequenceCategoricalColumn):
+      raise ValueError(
+          'In indicator_column: {}. '
+          'categorical_column must not be of type SequenceCategoricalColumn. '
+          'Suggested fix A: If you wish to use input_layer, use a '
+          'non-sequence categorical_column_with_*. '
+          'Suggested fix B: If you wish to create sequence input, use '
+          'sequence_input_layer instead of input_layer. '
+          'Given (type {}): {}'.format(self.name, type(self.categorical_column),
+                                       self.categorical_column))
+    # Feature has been already transformed. Return the intermediate
+    # representation created by transform_feature.
+    return transformation_cache.get(self, state_manager)
+
+  def get_sequence_dense_tensor(self, transformation_cache, state_manager):
+    """See `SequenceDenseColumn` base class."""
+    if not isinstance(self.categorical_column, SequenceCategoricalColumn):
+      raise ValueError(
+          'In indicator_column: {}. '
+          'categorical_column must be of type SequenceCategoricalColumn '
+          'to use sequence_input_layer. '
+          'Suggested fix: Use one of sequence_categorical_column_with_*. '
+          'Given (type {}): {}'.format(self.name, type(self.categorical_column),
+                                       self.categorical_column))
+    # Feature has been already transformed. Return the intermediate
+    # representation created by transform_feature.
+    dense_tensor = transformation_cache.get(self, state_manager)
+    sparse_tensors = self.categorical_column.get_sparse_tensors(
+        transformation_cache, state_manager)
+    sequence_length = _sequence_length_from_sparse_tensor(
+        sparse_tensors.id_tensor)
+    return SequenceDenseColumn.TensorSequenceLengthPair(
+        dense_tensor=dense_tensor, sequence_length=sequence_length)
+
+
+def _verify_static_batch_size_equality(tensors, columns):
+  # bath_size is a tf.Dimension object.
+  expected_batch_size = None
+  for i in range(0, len(tensors)):
+    if tensors[i].shape[0].value is not None:
+      if expected_batch_size is None:
+        bath_size_column_index = i
+        expected_batch_size = tensors[i].shape[0]
+      elif not expected_batch_size.is_compatible_with(tensors[i].shape[0]):
+        raise ValueError(
+            'Batch size (first dimension) of each feature must be same. '
+            'Batch size of columns ({}, {}): ({}, {})'.format(
+                columns[bath_size_column_index].name, columns[i].name,
+                expected_batch_size, tensors[i].shape[0]))
+
+
+def _sequence_length_from_sparse_tensor(sp_tensor, num_elements=1):
+  """Returns a [batch_size] Tensor with per-example sequence length."""
+  with ops.name_scope(None, 'sequence_length') as name_scope:
+    row_ids = sp_tensor.indices[:, 0]
+    column_ids = sp_tensor.indices[:, 1]
+    column_ids += array_ops.ones_like(column_ids)
+    seq_length = math_ops.to_int64(
+        math_ops.segment_max(column_ids, segment_ids=row_ids) / num_elements)
+    # If the last n rows do not have ids, seq_length will have shape
+    # [batch_size - n]. Pad the remaining values with zeros.
+    n_pad = array_ops.shape(sp_tensor)[:1] - array_ops.shape(seq_length)[:1]
+    padding = array_ops.zeros(n_pad, dtype=seq_length.dtype)
+    return array_ops.concat([seq_length, padding], axis=0, name=name_scope)
+
+
+class SequenceCategoricalColumn(FeatureColumn,
+                                collections.namedtuple(
+                                    'SequenceCategoricalColumn',
+                                    ('categorical_column'))):
+  """Represents sequences of categorical data."""
+
+  @property
+  def name(self):
+    """See `FeatureColumn` base class."""
+    return self.categorical_column.name
+
+  @property
+  def parse_example_spec(self):
+    """See `FeatureColumn` base class."""
+    return self.categorical_column.parse_example_spec
+
+  def transform_feature(self, transformation_cache, state_manager):
+    """See `FeatureColumn` base class."""
+    return self.categorical_column.transform_feature(transformation_cache,
+                                                     state_manager)
+
+  @property
+  def num_buckets(self):
+    """Returns number of buckets in this sparse feature."""
+    return self.categorical_column.num_buckets
+
+  def get_sequence_sparse_tensors(self, transformation_cache, state_manager):
+    """Returns an IdWeightPair.
+
+    `IdWeightPair` is a pair of `SparseTensor`s which represents ids and
+    weights.
+
+    `IdWeightPair.id_tensor` is typically a `batch_size` x `num_buckets`
+    `SparseTensor` of `int64`. `IdWeightPair.weight_tensor` is either a
+    `SparseTensor` of `float` or `None` to indicate all weights should be
+    taken to be 1. If specified, `weight_tensor` must have exactly the same
+    shape and indices as `sp_ids`. Expected `SparseTensor` is same as parsing
+    output of a `VarLenFeature` which is a ragged matrix.
+
+    Args:
+      transformation_cache: A `FeatureTransformationCache` object to access
+        features.
+      state_manager: A `StateManager` to create / access resources such as
+        lookup tables.
+    """
+    sparse_tensors = self.categorical_column.get_sparse_tensors(
+        transformation_cache, state_manager)
+    id_tensor = sparse_tensors.id_tensor
+    weight_tensor = sparse_tensors.weight_tensor
+    # Expands final dimension, so that embeddings are not combined during
+    # embedding lookup.
+    check_id_rank = check_ops.assert_equal(
+        array_ops.rank(id_tensor), 2,
+        data=[
+            'Column {} expected ID tensor of rank 2. '.format(self.name),
+            'id_tensor shape: ', array_ops.shape(id_tensor)])
+    with ops.control_dependencies([check_id_rank]):
+      id_tensor = sparse_ops.sparse_reshape(
+          id_tensor,
+          shape=array_ops.concat([id_tensor.dense_shape, [1]], axis=0))
+    if weight_tensor is not None:
+      check_weight_rank = check_ops.assert_equal(
+          array_ops.rank(weight_tensor), 2,
+          data=[
+              'Column {} expected weight tensor of rank 2.'.format(self.name),
+              'weight_tensor shape:', array_ops.shape(weight_tensor)])
+      with ops.control_dependencies([check_weight_rank]):
+        weight_tensor = sparse_ops.sparse_reshape(
+            weight_tensor,
+            shape=array_ops.concat([weight_tensor.dense_shape, [1]], axis=0))
+    return CategoricalColumn.IdWeightPair(id_tensor, weight_tensor)
diff --git a/tensorflow/python/feature_column/feature_column_v2_test.py b/tensorflow/python/feature_column/feature_column_v2_test.py
new file mode 100644
index 0000000000000000000000000000000000000000..80a9d5d40e275fce664ef52e5d5413930432d683
--- /dev/null
+++ b/tensorflow/python/feature_column/feature_column_v2_test.py
@@ -0,0 +1,6583 @@
+# Copyright 2017 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""Tests for feature_column."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import collections
+import copy
+
+import numpy as np
+
+from tensorflow.core.example import example_pb2
+from tensorflow.core.example import feature_pb2
+from tensorflow.core.protobuf import config_pb2
+from tensorflow.core.protobuf import rewriter_config_pb2
+from tensorflow.python.client import session
+from tensorflow.python.eager import backprop
+from tensorflow.python.eager import context
+from tensorflow.python.estimator.inputs import numpy_io
+from tensorflow.python.feature_column import feature_column as fc_old
+from tensorflow.python.feature_column import feature_column_v2 as fc
+from tensorflow.python.feature_column.feature_column_v2 import FeatureColumn
+from tensorflow.python.feature_column.feature_column_v2 import FeatureTransformationCache
+from tensorflow.python.feature_column.feature_column_v2 import InputLayer
+from tensorflow.python.feature_column.feature_column_v2 import StateManager
+from tensorflow.python.feature_column.feature_column_v2 import _LinearModel
+from tensorflow.python.feature_column.feature_column_v2 import _transform_features
+from tensorflow.python.framework import constant_op
+from tensorflow.python.framework import dtypes
+from tensorflow.python.framework import errors
+from tensorflow.python.framework import ops
+from tensorflow.python.framework import sparse_tensor
+from tensorflow.python.framework import test_util
+from tensorflow.python.ops import array_ops
+from tensorflow.python.ops import lookup_ops
+from tensorflow.python.ops import parsing_ops
+from tensorflow.python.ops import partitioned_variables
+from tensorflow.python.ops import variable_scope
+from tensorflow.python.ops import variables as variables_lib
+from tensorflow.python.platform import test
+from tensorflow.python.training import coordinator
+from tensorflow.python.training import queue_runner_impl
+
+
+def _initialized_session(config=None):
+  sess = session.Session(config=config)
+  sess.run(variables_lib.global_variables_initializer())
+  sess.run(lookup_ops.tables_initializer())
+  return sess
+
+
+class LazyColumnTest(test.TestCase):
+
+  def test_transformations_called_once(self):
+
+    class TransformCounter(FeatureColumn):
+
+      def __init__(self):
+        self.num_transform = 0
+
+      @property
+      def name(self):
+        return 'TransformCounter'
+
+      def transform_feature(self, transformation_cache, state_manager):
+        self.num_transform += 1  # Count transform calls.
+        return transformation_cache.get('a', state_manager)
+
+      @property
+      def parse_example_spec(self):
+        pass
+
+    transformation_cache = FeatureTransformationCache(
+        features={'a': [[2], [3.]]})
+    column = TransformCounter()
+    self.assertEqual(0, column.num_transform)
+    transformation_cache.get(column, None)
+    self.assertEqual(1, column.num_transform)
+    transformation_cache.get(column, None)
+    self.assertEqual(1, column.num_transform)
+
+  def test_returns_transform_output(self):
+
+    class Transformer(FeatureColumn):
+
+      @property
+      def name(self):
+        return 'Transformer'
+
+      def transform_feature(self, transformation_cache, state_manager):
+        return 'Output'
+
+      @property
+      def parse_example_spec(self):
+        pass
+
+    transformation_cache = FeatureTransformationCache(
+        features={'a': [[2], [3.]]})
+    column = Transformer()
+    self.assertEqual('Output', transformation_cache.get(column, None))
+    self.assertEqual('Output', transformation_cache.get(column, None))
+
+  def test_does_not_pollute_given_features_dict(self):
+
+    class Transformer(FeatureColumn):
+
+      @property
+      def name(self):
+        return 'Transformer'
+
+      def transform_feature(self, transformation_cache, state_manager):
+        return 'Output'
+
+      @property
+      def parse_example_spec(self):
+        pass
+
+    features = {'a': [[2], [3.]]}
+    transformation_cache = FeatureTransformationCache(features=features)
+    transformation_cache.get(Transformer(), None)
+    self.assertEqual(['a'], list(features.keys()))
+
+  def test_error_if_feature_is_not_found(self):
+    transformation_cache = FeatureTransformationCache(
+        features={'a': [[2], [3.]]})
+    with self.assertRaisesRegexp(ValueError,
+                                 'bbb is not in features dictionary'):
+      transformation_cache.get('bbb', None)
+    with self.assertRaisesRegexp(ValueError,
+                                 'bbb is not in features dictionary'):
+      transformation_cache.get(u'bbb', None)
+
+  def test_not_supported_feature_column(self):
+
+    class NotAProperColumn(FeatureColumn):
+
+      @property
+      def name(self):
+        return 'NotAProperColumn'
+
+      def transform_feature(self, transformation_cache, state_manager):
+        # It should return not None.
+        pass
+
+      @property
+      def parse_example_spec(self):
+        pass
+
+    transformation_cache = FeatureTransformationCache(
+        features={'a': [[2], [3.]]})
+    with self.assertRaisesRegexp(ValueError,
+                                 'NotAProperColumn is not supported'):
+      transformation_cache.get(NotAProperColumn(), None)
+
+  def test_key_should_be_string_or_feature_colum(self):
+
+    class NotAFeatureColumn(object):
+      pass
+
+    transformation_cache = FeatureTransformationCache(
+        features={'a': [[2], [3.]]})
+    with self.assertRaisesRegexp(
+        TypeError, '"key" must be either a "str" or "FeatureColumn".'):
+      transformation_cache.get(NotAFeatureColumn(), None)
+
+
+class NumericColumnTest(test.TestCase):
+
+  def test_defaults(self):
+    a = fc.numeric_column('aaa')
+    self.assertEqual('aaa', a.key)
+    self.assertEqual('aaa', a.name)
+    self.assertEqual((1,), a.shape)
+    self.assertIsNone(a.default_value)
+    self.assertEqual(dtypes.float32, a.dtype)
+    self.assertIsNone(a.normalizer_fn)
+
+  def test_key_should_be_string(self):
+    with self.assertRaisesRegexp(ValueError, 'key must be a string.'):
+      fc.numeric_column(key=('aaa',))
+
+  def test_shape_saved_as_tuple(self):
+    a = fc.numeric_column('aaa', shape=[1, 2], default_value=[[3, 2.]])
+    self.assertEqual((1, 2), a.shape)
+
+  def test_default_value_saved_as_tuple(self):
+    a = fc.numeric_column('aaa', default_value=4.)
+    self.assertEqual((4.,), a.default_value)
+    a = fc.numeric_column('aaa', shape=[1, 2], default_value=[[3, 2.]])
+    self.assertEqual(((3., 2.),), a.default_value)
+
+  def test_shape_and_default_value_compatibility(self):
+    fc.numeric_column('aaa', shape=[2], default_value=[1, 2.])
+    with self.assertRaisesRegexp(ValueError, 'The shape of default_value'):
+      fc.numeric_column('aaa', shape=[2], default_value=[1, 2, 3.])
+    fc.numeric_column(
+        'aaa', shape=[3, 2], default_value=[[2, 3], [1, 2], [2, 3.]])
+    with self.assertRaisesRegexp(ValueError, 'The shape of default_value'):
+      fc.numeric_column(
+          'aaa', shape=[3, 1], default_value=[[2, 3], [1, 2], [2, 3.]])
+    with self.assertRaisesRegexp(ValueError, 'The shape of default_value'):
+      fc.numeric_column(
+          'aaa', shape=[3, 3], default_value=[[2, 3], [1, 2], [2, 3.]])
+
+  def test_default_value_type_check(self):
+    fc.numeric_column(
+        'aaa', shape=[2], default_value=[1, 2.], dtype=dtypes.float32)
+    fc.numeric_column(
+        'aaa', shape=[2], default_value=[1, 2], dtype=dtypes.int32)
+    with self.assertRaisesRegexp(TypeError, 'must be compatible with dtype'):
+      fc.numeric_column(
+          'aaa', shape=[2], default_value=[1, 2.], dtype=dtypes.int32)
+    with self.assertRaisesRegexp(TypeError,
+                                 'default_value must be compatible with dtype'):
+      fc.numeric_column('aaa', default_value=['string'])
+
+  def test_shape_must_be_positive_integer(self):
+    with self.assertRaisesRegexp(TypeError, 'shape dimensions must be integer'):
+      fc.numeric_column(
+          'aaa', shape=[
+              1.0,
+          ])
+
+    with self.assertRaisesRegexp(ValueError,
+                                 'shape dimensions must be greater than 0'):
+      fc.numeric_column(
+          'aaa', shape=[
+              0,
+          ])
+
+  def test_dtype_is_convertible_to_float(self):
+    with self.assertRaisesRegexp(ValueError,
+                                 'dtype must be convertible to float'):
+      fc.numeric_column('aaa', dtype=dtypes.string)
+
+  def test_scalar_default_value_fills_the_shape(self):
+    a = fc.numeric_column('aaa', shape=[2, 3], default_value=2.)
+    self.assertEqual(((2., 2., 2.), (2., 2., 2.)), a.default_value)
+
+  def test_parse_spec(self):
+    a = fc.numeric_column('aaa', shape=[2, 3], dtype=dtypes.int32)
+    self.assertEqual({
+        'aaa': parsing_ops.FixedLenFeature((2, 3), dtype=dtypes.int32)
+    }, a.parse_example_spec)
+
+  def test_parse_example_no_default_value(self):
+    price = fc.numeric_column('price', shape=[2])
+    data = example_pb2.Example(features=feature_pb2.Features(
+        feature={
+            'price':
+                feature_pb2.Feature(float_list=feature_pb2.FloatList(
+                    value=[20., 110.]))
+        }))
+    features = parsing_ops.parse_example(
+        serialized=[data.SerializeToString()],
+        features=fc.make_parse_example_spec([price]))
+    self.assertIn('price', features)
+    with self.test_session():
+      self.assertAllEqual([[20., 110.]], features['price'].eval())
+
+  def test_parse_example_with_default_value(self):
+    price = fc.numeric_column('price', shape=[2], default_value=11.)
+    data = example_pb2.Example(features=feature_pb2.Features(
+        feature={
+            'price':
+                feature_pb2.Feature(float_list=feature_pb2.FloatList(
+                    value=[20., 110.]))
+        }))
+    no_data = example_pb2.Example(features=feature_pb2.Features(
+        feature={
+            'something_else':
+                feature_pb2.Feature(float_list=feature_pb2.FloatList(
+                    value=[20., 110.]))
+        }))
+    features = parsing_ops.parse_example(
+        serialized=[data.SerializeToString(),
+                    no_data.SerializeToString()],
+        features=fc.make_parse_example_spec([price]))
+    self.assertIn('price', features)
+    with self.test_session():
+      self.assertAllEqual([[20., 110.], [11., 11.]], features['price'].eval())
+
+  def test_normalizer_fn_must_be_callable(self):
+    with self.assertRaisesRegexp(TypeError, 'must be a callable'):
+      fc.numeric_column('price', normalizer_fn='NotACallable')
+
+  def test_normalizer_fn_transform_feature(self):
+
+    def _increment_two(input_tensor):
+      return input_tensor + 2.
+
+    price = fc.numeric_column('price', shape=[2], normalizer_fn=_increment_two)
+    output = _transform_features({'price': [[1., 2.], [5., 6.]]}, [price], None)
+    with self.test_session():
+      self.assertAllEqual([[3., 4.], [7., 8.]], output[price].eval())
+
+  def test_get_dense_tensor(self):
+
+    def _increment_two(input_tensor):
+      return input_tensor + 2.
+
+    price = fc.numeric_column('price', shape=[2], normalizer_fn=_increment_two)
+    transformation_cache = FeatureTransformationCache({
+        'price': [[1., 2.], [5., 6.]]
+    })
+    self.assertEqual(
+        transformation_cache.get(price, None),
+        price.get_dense_tensor(transformation_cache, None))
+
+  def test_sparse_tensor_not_supported(self):
+    price = fc.numeric_column('price')
+    transformation_cache = FeatureTransformationCache({
+        'price':
+            sparse_tensor.SparseTensor(
+                indices=[[0, 0]], values=[0.3], dense_shape=[1, 1])
+    })
+    with self.assertRaisesRegexp(ValueError, 'must be a Tensor'):
+      price.transform_feature(transformation_cache, None)
+
+  def test_deep_copy(self):
+    a = fc.numeric_column('aaa', shape=[1, 2], default_value=[[3., 2.]])
+    a_copy = copy.deepcopy(a)
+    self.assertEqual(a_copy.name, 'aaa')
+    self.assertEqual(a_copy.shape, (1, 2))
+    self.assertEqual(a_copy.default_value, ((3., 2.),))
+
+  def test_numpy_default_value(self):
+    a = fc.numeric_column(
+        'aaa', shape=[1, 2], default_value=np.array([[3., 2.]]))
+    self.assertEqual(a.default_value, ((3., 2.),))
+
+  def test_linear_model(self):
+    price = fc_old.numeric_column('price')
+    with ops.Graph().as_default():
+      features = {'price': [[1.], [5.]]}
+      predictions = fc.linear_model(features, [price])
+      bias = get_linear_model_bias()
+      price_var = get_linear_model_column_var(price)
+      with _initialized_session() as sess:
+        self.assertAllClose([0.], bias.eval())
+        self.assertAllClose([[0.]], price_var.eval())
+        self.assertAllClose([[0.], [0.]], predictions.eval())
+        sess.run(price_var.assign([[10.]]))
+        self.assertAllClose([[10.], [50.]], predictions.eval())
+
+  def test_keras_linear_model(self):
+    price = fc_old.numeric_column('price')
+    with ops.Graph().as_default():
+      features = {'price': [[1.], [5.]]}
+      predictions = get_keras_linear_model_predictions(features, [price])
+      bias = get_linear_model_bias()
+      price_var = get_linear_model_column_var(price)
+      with _initialized_session() as sess:
+        self.assertAllClose([0.], bias.eval())
+        self.assertAllClose([[0.]], price_var.eval())
+        self.assertAllClose([[0.], [0.]], predictions.eval())
+        sess.run(price_var.assign([[10.]]))
+        self.assertAllClose([[10.], [50.]], predictions.eval())
+
+
+class BucketizedColumnTest(test.TestCase):
+
+  def test_invalid_source_column_type(self):
+    a = fc.categorical_column_with_hash_bucket('aaa', hash_bucket_size=10)
+    with self.assertRaisesRegexp(
+        ValueError,
+        'source_column must be a column generated with numeric_column'):
+      fc.bucketized_column(a, boundaries=[0, 1])
+
+  def test_invalid_source_column_shape(self):
+    a = fc.numeric_column('aaa', shape=[2, 3])
+    with self.assertRaisesRegexp(
+        ValueError, 'source_column must be one-dimensional column'):
+      fc.bucketized_column(a, boundaries=[0, 1])
+
+  def test_invalid_boundaries(self):
+    a = fc.numeric_column('aaa')
+    with self.assertRaisesRegexp(
+        ValueError, 'boundaries must be a sorted list'):
+      fc.bucketized_column(a, boundaries=None)
+    with self.assertRaisesRegexp(
+        ValueError, 'boundaries must be a sorted list'):
+      fc.bucketized_column(a, boundaries=1.)
+    with self.assertRaisesRegexp(
+        ValueError, 'boundaries must be a sorted list'):
+      fc.bucketized_column(a, boundaries=[1, 0])
+    with self.assertRaisesRegexp(
+        ValueError, 'boundaries must be a sorted list'):
+      fc.bucketized_column(a, boundaries=[1, 1])
+
+  def test_name(self):
+    a = fc.numeric_column('aaa', dtype=dtypes.int32)
+    b = fc.bucketized_column(a, boundaries=[0, 1])
+    self.assertEqual('aaa_bucketized', b.name)
+
+  def test_parse_spec(self):
+    a = fc.numeric_column('aaa', shape=[2], dtype=dtypes.int32)
+    b = fc.bucketized_column(a, boundaries=[0, 1])
+    self.assertEqual({
+        'aaa': parsing_ops.FixedLenFeature((2,), dtype=dtypes.int32)
+    }, b.parse_example_spec)
+
+  def test_variable_shape(self):
+    a = fc.numeric_column('aaa', shape=[2], dtype=dtypes.int32)
+    b = fc.bucketized_column(a, boundaries=[0, 1])
+    # Column 'aaa` has shape [2] times three buckets -> variable_shape=[2, 3].
+    self.assertAllEqual((2, 3), b.variable_shape)
+
+  def test_num_buckets(self):
+    a = fc.numeric_column('aaa', shape=[2], dtype=dtypes.int32)
+    b = fc.bucketized_column(a, boundaries=[0, 1])
+    # Column 'aaa` has shape [2] times three buckets -> num_buckets=6.
+    self.assertEqual(6, b.num_buckets)
+
+  def test_parse_example(self):
+    price = fc.numeric_column('price', shape=[2])
+    bucketized_price = fc.bucketized_column(price, boundaries=[0, 50])
+    data = example_pb2.Example(features=feature_pb2.Features(
+        feature={
+            'price':
+                feature_pb2.Feature(float_list=feature_pb2.FloatList(
+                    value=[20., 110.]))
+        }))
+    features = parsing_ops.parse_example(
+        serialized=[data.SerializeToString()],
+        features=fc.make_parse_example_spec([bucketized_price]))
+    self.assertIn('price', features)
+    with self.test_session():
+      self.assertAllEqual([[20., 110.]], features['price'].eval())
+
+  def test_transform_feature(self):
+    price = fc.numeric_column('price', shape=[2])
+    bucketized_price = fc.bucketized_column(price, boundaries=[0, 2, 4, 6])
+    with ops.Graph().as_default():
+      transformed_tensor = _transform_features({
+          'price': [[-1., 1.], [5., 6.]]
+      }, [bucketized_price], None)
+      with _initialized_session():
+        self.assertAllEqual([[0, 1], [3, 4]],
+                            transformed_tensor[bucketized_price].eval())
+
+  def test_get_dense_tensor_one_input_value(self):
+    """Tests _get_dense_tensor() for input with shape=[1]."""
+    price = fc.numeric_column('price', shape=[1])
+    bucketized_price = fc.bucketized_column(price, boundaries=[0, 2, 4, 6])
+    with ops.Graph().as_default():
+      transformation_cache = FeatureTransformationCache({
+          'price': [[-1.], [1.], [5.], [6.]]
+      })
+      with _initialized_session():
+        bucketized_price_tensor = bucketized_price.get_dense_tensor(
+            transformation_cache, None)
+        self.assertAllClose(
+            # One-hot tensor.
+            [[[1., 0., 0., 0., 0.]],
+             [[0., 1., 0., 0., 0.]],
+             [[0., 0., 0., 1., 0.]],
+             [[0., 0., 0., 0., 1.]]],
+            bucketized_price_tensor.eval())
+
+  def test_get_dense_tensor_two_input_values(self):
+    """Tests _get_dense_tensor() for input with shape=[2]."""
+    price = fc.numeric_column('price', shape=[2])
+    bucketized_price = fc.bucketized_column(price, boundaries=[0, 2, 4, 6])
+    with ops.Graph().as_default():
+      transformation_cache = FeatureTransformationCache({
+          'price': [[-1., 1.], [5., 6.]]
+      })
+      with _initialized_session():
+        bucketized_price_tensor = bucketized_price.get_dense_tensor(
+            transformation_cache, None)
+        self.assertAllClose(
+            # One-hot tensor.
+            [[[1., 0., 0., 0., 0.], [0., 1., 0., 0., 0.]],
+             [[0., 0., 0., 1., 0.], [0., 0., 0., 0., 1.]]],
+            bucketized_price_tensor.eval())
+
+  def test_get_sparse_tensors_one_input_value(self):
+    """Tests _get_sparse_tensors() for input with shape=[1]."""
+    price = fc.numeric_column('price', shape=[1])
+    bucketized_price = fc.bucketized_column(price, boundaries=[0, 2, 4, 6])
+    with ops.Graph().as_default():
+      transformation_cache = FeatureTransformationCache({
+          'price': [[-1.], [1.], [5.], [6.]]
+      })
+      with _initialized_session() as sess:
+        id_weight_pair = bucketized_price.get_sparse_tensors(
+            transformation_cache, None)
+        self.assertIsNone(id_weight_pair.weight_tensor)
+        id_tensor_value = sess.run(id_weight_pair.id_tensor)
+        self.assertAllEqual(
+            [[0, 0], [1, 0], [2, 0], [3, 0]], id_tensor_value.indices)
+        self.assertAllEqual([0, 1, 3, 4], id_tensor_value.values)
+        self.assertAllEqual([4, 1], id_tensor_value.dense_shape)
+
+  def test_get_sparse_tensors_two_input_values(self):
+    """Tests _get_sparse_tensors() for input with shape=[2]."""
+    price = fc.numeric_column('price', shape=[2])
+    bucketized_price = fc.bucketized_column(price, boundaries=[0, 2, 4, 6])
+    with ops.Graph().as_default():
+      transformation_cache = FeatureTransformationCache({
+          'price': [[-1., 1.], [5., 6.]]
+      })
+      with _initialized_session() as sess:
+        id_weight_pair = bucketized_price.get_sparse_tensors(
+            transformation_cache, None)
+        self.assertIsNone(id_weight_pair.weight_tensor)
+        id_tensor_value = sess.run(id_weight_pair.id_tensor)
+        self.assertAllEqual(
+            [[0, 0], [0, 1], [1, 0], [1, 1]], id_tensor_value.indices)
+        # Values 0-4 correspond to the first column of the input price.
+        # Values 5-9 correspond to the second column of the input price.
+        self.assertAllEqual([0, 6, 3, 9], id_tensor_value.values)
+        self.assertAllEqual([2, 2], id_tensor_value.dense_shape)
+
+  def test_sparse_tensor_input_not_supported(self):
+    price = fc.numeric_column('price')
+    bucketized_price = fc.bucketized_column(price, boundaries=[0, 1])
+    transformation_cache = FeatureTransformationCache({
+        'price':
+            sparse_tensor.SparseTensor(
+                indices=[[0, 0]], values=[0.3], dense_shape=[1, 1])
+    })
+    with self.assertRaisesRegexp(ValueError, 'must be a Tensor'):
+      bucketized_price.transform_feature(transformation_cache, None)
+
+  def test_deep_copy(self):
+    a = fc.numeric_column('aaa', shape=[2])
+    a_bucketized = fc.bucketized_column(a, boundaries=[0, 1])
+    a_bucketized_copy = copy.deepcopy(a_bucketized)
+    self.assertEqual(a_bucketized_copy.name, 'aaa_bucketized')
+    self.assertAllEqual(a_bucketized_copy.variable_shape, (2, 3))
+    self.assertEqual(a_bucketized_copy.boundaries, (0, 1))
+
+  def test_linear_model_one_input_value(self):
+    """Tests linear_model() for input with shape=[1]."""
+    price = fc_old.numeric_column('price', shape=[1])
+    bucketized_price = fc_old.bucketized_column(price, boundaries=[0, 2, 4, 6])
+    with ops.Graph().as_default():
+      features = {'price': [[-1.], [1.], [5.], [6.]]}
+      predictions = fc.linear_model(features, [bucketized_price])
+      bias = get_linear_model_bias()
+      bucketized_price_var = get_linear_model_column_var(bucketized_price)
+      with _initialized_session() as sess:
+        self.assertAllClose([0.], bias.eval())
+        # One weight variable per bucket, all initialized to zero.
+        self.assertAllClose(
+            [[0.], [0.], [0.], [0.], [0.]], bucketized_price_var.eval())
+        self.assertAllClose([[0.], [0.], [0.], [0.]], predictions.eval())
+        sess.run(bucketized_price_var.assign(
+            [[10.], [20.], [30.], [40.], [50.]]))
+        # price -1. is in the 0th bucket, whose weight is 10.
+        # price 1. is in the 1st bucket, whose weight is 20.
+        # price 5. is in the 3rd bucket, whose weight is 40.
+        # price 6. is in the 4th bucket, whose weight is 50.
+        self.assertAllClose([[10.], [20.], [40.], [50.]], predictions.eval())
+        sess.run(bias.assign([1.]))
+        self.assertAllClose([[11.], [21.], [41.], [51.]], predictions.eval())
+
+  def test_linear_model_two_input_values(self):
+    """Tests linear_model() for input with shape=[2]."""
+    price = fc_old.numeric_column('price', shape=[2])
+    bucketized_price = fc_old.bucketized_column(price, boundaries=[0, 2, 4, 6])
+    with ops.Graph().as_default():
+      features = {'price': [[-1., 1.], [5., 6.]]}
+      predictions = fc.linear_model(features, [bucketized_price])
+      bias = get_linear_model_bias()
+      bucketized_price_var = get_linear_model_column_var(bucketized_price)
+      with _initialized_session() as sess:
+        self.assertAllClose([0.], bias.eval())
+        # One weight per bucket per input column, all initialized to zero.
+        self.assertAllClose(
+            [[0.], [0.], [0.], [0.], [0.], [0.], [0.], [0.], [0.], [0.]],
+            bucketized_price_var.eval())
+        self.assertAllClose([[0.], [0.]], predictions.eval())
+        sess.run(bucketized_price_var.assign(
+            [[10.], [20.], [30.], [40.], [50.],
+             [60.], [70.], [80.], [90.], [100.]]))
+        # 1st example:
+        #   price -1. is in the 0th bucket, whose weight is 10.
+        #   price 1. is in the 6th bucket, whose weight is 70.
+        # 2nd example:
+        #   price 5. is in the 3rd bucket, whose weight is 40.
+        #   price 6. is in the 9th bucket, whose weight is 100.
+        self.assertAllClose([[80.], [140.]], predictions.eval())
+        sess.run(bias.assign([1.]))
+        self.assertAllClose([[81.], [141.]], predictions.eval())
+
+  def test_keras_linear_model_one_input_value(self):
+    """Tests _LinearModel for input with shape=[1]."""
+    price = fc_old.numeric_column('price', shape=[1])
+    bucketized_price = fc_old.bucketized_column(price, boundaries=[0, 2, 4, 6])
+    with ops.Graph().as_default():
+      features = {'price': [[-1.], [1.], [5.], [6.]]}
+      predictions = get_keras_linear_model_predictions(features,
+                                                       [bucketized_price])
+      bias = get_linear_model_bias()
+      bucketized_price_var = get_linear_model_column_var(bucketized_price)
+      with _initialized_session() as sess:
+        self.assertAllClose([0.], bias.eval())
+        # One weight variable per bucket, all initialized to zero.
+        self.assertAllClose([[0.], [0.], [0.], [0.], [0.]],
+                            bucketized_price_var.eval())
+        self.assertAllClose([[0.], [0.], [0.], [0.]], predictions.eval())
+        sess.run(
+            bucketized_price_var.assign([[10.], [20.], [30.], [40.], [50.]]))
+        # price -1. is in the 0th bucket, whose weight is 10.
+        # price 1. is in the 1st bucket, whose weight is 20.
+        # price 5. is in the 3rd bucket, whose weight is 40.
+        # price 6. is in the 4th bucket, whose weight is 50.
+        self.assertAllClose([[10.], [20.], [40.], [50.]], predictions.eval())
+        sess.run(bias.assign([1.]))
+        self.assertAllClose([[11.], [21.], [41.], [51.]], predictions.eval())
+
+  def test_keras_linear_model_two_input_values(self):
+    """Tests _LinearModel for input with shape=[2]."""
+    price = fc_old.numeric_column('price', shape=[2])
+    bucketized_price = fc_old.bucketized_column(price, boundaries=[0, 2, 4, 6])
+    with ops.Graph().as_default():
+      features = {'price': [[-1., 1.], [5., 6.]]}
+      predictions = get_keras_linear_model_predictions(features,
+                                                       [bucketized_price])
+      bias = get_linear_model_bias()
+      bucketized_price_var = get_linear_model_column_var(bucketized_price)
+      with _initialized_session() as sess:
+        self.assertAllClose([0.], bias.eval())
+        # One weight per bucket per input column, all initialized to zero.
+        self.assertAllClose(
+            [[0.], [0.], [0.], [0.], [0.], [0.], [0.], [0.], [0.], [0.]],
+            bucketized_price_var.eval())
+        self.assertAllClose([[0.], [0.]], predictions.eval())
+        sess.run(
+            bucketized_price_var.assign([[10.], [20.], [30.], [40.], [50.],
+                                         [60.], [70.], [80.], [90.], [100.]]))
+        # 1st example:
+        #   price -1. is in the 0th bucket, whose weight is 10.
+        #   price 1. is in the 6th bucket, whose weight is 70.
+        # 2nd example:
+        #   price 5. is in the 3rd bucket, whose weight is 40.
+        #   price 6. is in the 9th bucket, whose weight is 100.
+        self.assertAllClose([[80.], [140.]], predictions.eval())
+        sess.run(bias.assign([1.]))
+        self.assertAllClose([[81.], [141.]], predictions.eval())
+
+
+class HashedCategoricalColumnTest(test.TestCase):
+
+  def test_defaults(self):
+    a = fc.categorical_column_with_hash_bucket('aaa', 10)
+    self.assertEqual('aaa', a.name)
+    self.assertEqual('aaa', a.key)
+    self.assertEqual(10, a.hash_bucket_size)
+    self.assertEqual(dtypes.string, a.dtype)
+
+  def test_key_should_be_string(self):
+    with self.assertRaisesRegexp(ValueError, 'key must be a string.'):
+      fc.categorical_column_with_hash_bucket(('key',), 10)
+
+  def test_bucket_size_should_be_given(self):
+    with self.assertRaisesRegexp(ValueError, 'hash_bucket_size must be set.'):
+      fc.categorical_column_with_hash_bucket('aaa', None)
+
+  def test_bucket_size_should_be_positive(self):
+    with self.assertRaisesRegexp(ValueError,
+                                 'hash_bucket_size must be at least 1'):
+      fc.categorical_column_with_hash_bucket('aaa', 0)
+
+  def test_dtype_should_be_string_or_integer(self):
+    fc.categorical_column_with_hash_bucket('aaa', 10, dtype=dtypes.string)
+    fc.categorical_column_with_hash_bucket('aaa', 10, dtype=dtypes.int32)
+    with self.assertRaisesRegexp(ValueError, 'dtype must be string or integer'):
+      fc.categorical_column_with_hash_bucket('aaa', 10, dtype=dtypes.float32)
+
+  def test_deep_copy(self):
+    original = fc.categorical_column_with_hash_bucket('aaa', 10)
+    for column in (original, copy.deepcopy(original)):
+      self.assertEqual('aaa', column.name)
+      self.assertEqual(10, column.hash_bucket_size)
+      self.assertEqual(10, column.num_buckets)
+      self.assertEqual(dtypes.string, column.dtype)
+
+  def test_parse_spec_string(self):
+    a = fc.categorical_column_with_hash_bucket('aaa', 10)
+    self.assertEqual({
+        'aaa': parsing_ops.VarLenFeature(dtypes.string)
+    }, a.parse_example_spec)
+
+  def test_parse_spec_int(self):
+    a = fc.categorical_column_with_hash_bucket('aaa', 10, dtype=dtypes.int32)
+    self.assertEqual({
+        'aaa': parsing_ops.VarLenFeature(dtypes.int32)
+    }, a.parse_example_spec)
+
+  def test_parse_example(self):
+    a = fc.categorical_column_with_hash_bucket('aaa', 10)
+    data = example_pb2.Example(features=feature_pb2.Features(
+        feature={
+            'aaa':
+                feature_pb2.Feature(bytes_list=feature_pb2.BytesList(
+                    value=[b'omar', b'stringer']))
+        }))
+    features = parsing_ops.parse_example(
+        serialized=[data.SerializeToString()],
+        features=fc.make_parse_example_spec([a]))
+    self.assertIn('aaa', features)
+    with self.test_session():
+      _assert_sparse_tensor_value(
+          self,
+          sparse_tensor.SparseTensorValue(
+              indices=[[0, 0], [0, 1]],
+              values=np.array([b'omar', b'stringer'], dtype=np.object_),
+              dense_shape=[1, 2]),
+          features['aaa'].eval())
+
+  def test_strings_should_be_hashed(self):
+    hashed_sparse = fc.categorical_column_with_hash_bucket('wire', 10)
+    wire_tensor = sparse_tensor.SparseTensor(
+        values=['omar', 'stringer', 'marlo'],
+        indices=[[0, 0], [1, 0], [1, 1]],
+        dense_shape=[2, 2])
+    outputs = _transform_features({'wire': wire_tensor}, [hashed_sparse], None)
+    output = outputs[hashed_sparse]
+    # Check exact hashed output. If hashing changes this test will break.
+    expected_values = [6, 4, 1]
+    with self.test_session():
+      self.assertEqual(dtypes.int64, output.values.dtype)
+      self.assertAllEqual(expected_values, output.values.eval())
+      self.assertAllEqual(wire_tensor.indices.eval(), output.indices.eval())
+      self.assertAllEqual(wire_tensor.dense_shape.eval(),
+                          output.dense_shape.eval())
+
+  def test_tensor_dtype_should_be_string_or_integer(self):
+    string_fc = fc.categorical_column_with_hash_bucket(
+        'a_string', 10, dtype=dtypes.string)
+    int_fc = fc.categorical_column_with_hash_bucket(
+        'a_int', 10, dtype=dtypes.int32)
+    float_fc = fc.categorical_column_with_hash_bucket(
+        'a_float', 10, dtype=dtypes.string)
+    int_tensor = sparse_tensor.SparseTensor(
+        values=[101],
+        indices=[[0, 0]],
+        dense_shape=[1, 1])
+    string_tensor = sparse_tensor.SparseTensor(
+        values=['101'],
+        indices=[[0, 0]],
+        dense_shape=[1, 1])
+    float_tensor = sparse_tensor.SparseTensor(
+        values=[101.],
+        indices=[[0, 0]],
+        dense_shape=[1, 1])
+    transformation_cache = FeatureTransformationCache({
+        'a_int': int_tensor,
+        'a_string': string_tensor,
+        'a_float': float_tensor
+    })
+    transformation_cache.get(string_fc, None)
+    transformation_cache.get(int_fc, None)
+    with self.assertRaisesRegexp(ValueError, 'dtype must be string or integer'):
+      transformation_cache.get(float_fc, None)
+
+  def test_dtype_should_match_with_tensor(self):
+    hashed_sparse = fc.categorical_column_with_hash_bucket(
+        'wire', 10, dtype=dtypes.int64)
+    wire_tensor = sparse_tensor.SparseTensor(
+        values=['omar'], indices=[[0, 0]], dense_shape=[1, 1])
+    transformation_cache = FeatureTransformationCache({'wire': wire_tensor})
+    with self.assertRaisesRegexp(ValueError, 'dtype must be compatible'):
+      transformation_cache.get(hashed_sparse, None)
+
+  def test_ints_should_be_hashed(self):
+    hashed_sparse = fc.categorical_column_with_hash_bucket(
+        'wire', 10, dtype=dtypes.int64)
+    wire_tensor = sparse_tensor.SparseTensor(
+        values=[101, 201, 301],
+        indices=[[0, 0], [1, 0], [1, 1]],
+        dense_shape=[2, 2])
+    transformation_cache = FeatureTransformationCache({'wire': wire_tensor})
+    output = transformation_cache.get(hashed_sparse, None)
+    # Check exact hashed output. If hashing changes this test will break.
+    expected_values = [3, 7, 5]
+    with self.test_session():
+      self.assertAllEqual(expected_values, output.values.eval())
+
+  def test_int32_64_is_compatible(self):
+    hashed_sparse = fc.categorical_column_with_hash_bucket(
+        'wire', 10, dtype=dtypes.int64)
+    wire_tensor = sparse_tensor.SparseTensor(
+        values=constant_op.constant([101, 201, 301], dtype=dtypes.int32),
+        indices=[[0, 0], [1, 0], [1, 1]],
+        dense_shape=[2, 2])
+    transformation_cache = FeatureTransformationCache({'wire': wire_tensor})
+    output = transformation_cache.get(hashed_sparse, None)
+    # Check exact hashed output. If hashing changes this test will break.
+    expected_values = [3, 7, 5]
+    with self.test_session():
+      self.assertAllEqual(expected_values, output.values.eval())
+
+  def test_get_sparse_tensors(self):
+    hashed_sparse = fc.categorical_column_with_hash_bucket('wire', 10)
+    transformation_cache = FeatureTransformationCache({
+        'wire':
+            sparse_tensor.SparseTensor(
+                values=['omar', 'stringer', 'marlo'],
+                indices=[[0, 0], [1, 0], [1, 1]],
+                dense_shape=[2, 2])
+    })
+    id_weight_pair = hashed_sparse.get_sparse_tensors(transformation_cache,
+                                                      None)
+    self.assertIsNone(id_weight_pair.weight_tensor)
+    self.assertEqual(
+        transformation_cache.get(hashed_sparse, None), id_weight_pair.id_tensor)
+
+  def DISABLED_test_get_sparse_tensors_weight_collections(self):
+    column = fc.categorical_column_with_hash_bucket('aaa', 10)
+    inputs = sparse_tensor.SparseTensor(
+        values=['omar', 'stringer', 'marlo'],
+        indices=[[0, 0], [1, 0], [1, 1]],
+        dense_shape=[2, 2])
+    column._get_sparse_tensors(
+        FeatureTransformationCache({
+            'aaa': inputs
+        }),
+        weight_collections=('my_weights',))
+
+    self.assertItemsEqual(
+        [], ops.get_collection(ops.GraphKeys.GLOBAL_VARIABLES))
+    self.assertItemsEqual([], ops.get_collection('my_weights'))
+
+  def test_get_sparse_tensors_dense_input(self):
+    hashed_sparse = fc.categorical_column_with_hash_bucket('wire', 10)
+    transformation_cache = FeatureTransformationCache({
+        'wire': (('omar', ''), ('stringer', 'marlo'))
+    })
+    id_weight_pair = hashed_sparse.get_sparse_tensors(transformation_cache,
+                                                      None)
+    self.assertIsNone(id_weight_pair.weight_tensor)
+    self.assertEqual(
+        transformation_cache.get(hashed_sparse, None), id_weight_pair.id_tensor)
+
+  def test_linear_model(self):
+    wire_column = fc_old.categorical_column_with_hash_bucket('wire', 4)
+    self.assertEqual(4, wire_column._num_buckets)
+    with ops.Graph().as_default():
+      predictions = fc.linear_model({
+          wire_column.name: sparse_tensor.SparseTensorValue(
+              indices=((0, 0), (1, 0), (1, 1)),
+              values=('marlo', 'skywalker', 'omar'),
+              dense_shape=(2, 2))
+      }, (wire_column,))
+      bias = get_linear_model_bias()
+      wire_var = get_linear_model_column_var(wire_column)
+      with _initialized_session():
+        self.assertAllClose((0.,), bias.eval())
+        self.assertAllClose(((0.,), (0.,), (0.,), (0.,)), wire_var.eval())
+        self.assertAllClose(((0.,), (0.,)), predictions.eval())
+        wire_var.assign(((1.,), (2.,), (3.,), (4.,))).eval()
+        # 'marlo' -> 3: wire_var[3] = 4
+        # 'skywalker' -> 2, 'omar' -> 2: wire_var[2] + wire_var[2] = 3+3 = 6
+        self.assertAllClose(((4.,), (6.,)), predictions.eval())
+
+  def test_keras_linear_model(self):
+    wire_column = fc_old.categorical_column_with_hash_bucket('wire', 4)
+    self.assertEqual(4, wire_column._num_buckets)
+    with ops.Graph().as_default():
+      predictions = get_keras_linear_model_predictions({
+          wire_column.name:
+              sparse_tensor.SparseTensorValue(
+                  indices=((0, 0), (1, 0), (1, 1)),
+                  values=('marlo', 'skywalker', 'omar'),
+                  dense_shape=(2, 2))
+      }, (wire_column,))
+      bias = get_linear_model_bias()
+      wire_var = get_linear_model_column_var(wire_column)
+      with _initialized_session():
+        self.assertAllClose((0.,), bias.eval())
+        self.assertAllClose(((0.,), (0.,), (0.,), (0.,)), wire_var.eval())
+        self.assertAllClose(((0.,), (0.,)), predictions.eval())
+        wire_var.assign(((1.,), (2.,), (3.,), (4.,))).eval()
+        # 'marlo' -> 3: wire_var[3] = 4
+        # 'skywalker' -> 2, 'omar' -> 2: wire_var[2] + wire_var[2] = 3+3 = 6
+        self.assertAllClose(((4.,), (6.,)), predictions.eval())
+
+
+class CrossedColumnTest(test.TestCase):
+
+  def test_keys_empty(self):
+    with self.assertRaisesRegexp(
+        ValueError, 'keys must be a list with length > 1'):
+      fc.crossed_column([], 10)
+
+  def test_keys_length_one(self):
+    with self.assertRaisesRegexp(
+        ValueError, 'keys must be a list with length > 1'):
+      fc.crossed_column(['a'], 10)
+
+  def test_key_type_unsupported(self):
+    with self.assertRaisesRegexp(ValueError, 'Unsupported key type'):
+      fc.crossed_column(['a', fc.numeric_column('c')], 10)
+
+    with self.assertRaisesRegexp(
+        ValueError, 'categorical_column_with_hash_bucket is not supported'):
+      fc.crossed_column(
+          ['a', fc.categorical_column_with_hash_bucket('c', 10)], 10)
+
+  def test_hash_bucket_size_negative(self):
+    with self.assertRaisesRegexp(
+        ValueError, 'hash_bucket_size must be > 1'):
+      fc.crossed_column(['a', 'c'], -1)
+
+  def test_hash_bucket_size_zero(self):
+    with self.assertRaisesRegexp(
+        ValueError, 'hash_bucket_size must be > 1'):
+      fc.crossed_column(['a', 'c'], 0)
+
+  def test_hash_bucket_size_none(self):
+    with self.assertRaisesRegexp(
+        ValueError, 'hash_bucket_size must be > 1'):
+      fc.crossed_column(['a', 'c'], None)
+
+  def test_name(self):
+    a = fc.numeric_column('a', dtype=dtypes.int32)
+    b = fc.bucketized_column(a, boundaries=[0, 1])
+    crossed1 = fc.crossed_column(['d1', 'd2'], 10)
+
+    crossed2 = fc.crossed_column([b, 'c', crossed1], 10)
+    self.assertEqual('a_bucketized_X_c_X_d1_X_d2', crossed2.name)
+
+  def test_name_ordered_alphabetically(self):
+    """Tests that the name does not depend on the order of given columns."""
+    a = fc.numeric_column('a', dtype=dtypes.int32)
+    b = fc.bucketized_column(a, boundaries=[0, 1])
+    crossed1 = fc.crossed_column(['d1', 'd2'], 10)
+
+    crossed2 = fc.crossed_column([crossed1, 'c', b], 10)
+    self.assertEqual('a_bucketized_X_c_X_d1_X_d2', crossed2.name)
+
+  def test_name_leaf_keys_ordered_alphabetically(self):
+    """Tests that the name does not depend on the order of given columns."""
+    a = fc.numeric_column('a', dtype=dtypes.int32)
+    b = fc.bucketized_column(a, boundaries=[0, 1])
+    crossed1 = fc.crossed_column(['d2', 'c'], 10)
+
+    crossed2 = fc.crossed_column([crossed1, 'd1', b], 10)
+    self.assertEqual('a_bucketized_X_c_X_d1_X_d2', crossed2.name)
+
+  def test_parse_spec(self):
+    a = fc.numeric_column('a', shape=[2], dtype=dtypes.int32)
+    b = fc.bucketized_column(a, boundaries=[0, 1])
+    crossed = fc.crossed_column([b, 'c'], 10)
+    self.assertEqual({
+        'a': parsing_ops.FixedLenFeature((2,), dtype=dtypes.int32),
+        'c': parsing_ops.VarLenFeature(dtypes.string),
+    }, crossed.parse_example_spec)
+
+  def test_num_buckets(self):
+    a = fc.numeric_column('a', shape=[2], dtype=dtypes.int32)
+    b = fc.bucketized_column(a, boundaries=[0, 1])
+    crossed = fc.crossed_column([b, 'c'], 15)
+    self.assertEqual(15, crossed.num_buckets)
+
+  def test_deep_copy(self):
+    a = fc.numeric_column('a', dtype=dtypes.int32)
+    b = fc.bucketized_column(a, boundaries=[0, 1])
+    crossed1 = fc.crossed_column(['d1', 'd2'], 10)
+    crossed2 = fc.crossed_column([b, 'c', crossed1], 15, hash_key=5)
+    crossed2_copy = copy.deepcopy(crossed2)
+    self.assertEqual('a_bucketized_X_c_X_d1_X_d2', crossed2_copy.name,)
+    self.assertEqual(15, crossed2_copy.hash_bucket_size)
+    self.assertEqual(5, crossed2_copy.hash_key)
+
+  def test_parse_example(self):
+    price = fc.numeric_column('price', shape=[2])
+    bucketized_price = fc.bucketized_column(price, boundaries=[0, 50])
+    price_cross_wire = fc.crossed_column([bucketized_price, 'wire'], 10)
+    data = example_pb2.Example(features=feature_pb2.Features(
+        feature={
+            'price':
+                feature_pb2.Feature(float_list=feature_pb2.FloatList(
+                    value=[20., 110.])),
+            'wire':
+                feature_pb2.Feature(bytes_list=feature_pb2.BytesList(
+                    value=[b'omar', b'stringer'])),
+        }))
+    features = parsing_ops.parse_example(
+        serialized=[data.SerializeToString()],
+        features=fc.make_parse_example_spec([price_cross_wire]))
+    self.assertIn('price', features)
+    self.assertIn('wire', features)
+    with self.test_session():
+      self.assertAllEqual([[20., 110.]], features['price'].eval())
+      wire_sparse = features['wire']
+      self.assertAllEqual([[0, 0], [0, 1]], wire_sparse.indices.eval())
+      # Use byte constants to pass the open-source test.
+      self.assertAllEqual([b'omar', b'stringer'], wire_sparse.values.eval())
+      self.assertAllEqual([1, 2], wire_sparse.dense_shape.eval())
+
+  def test_transform_feature(self):
+    price = fc.numeric_column('price', shape=[2])
+    bucketized_price = fc.bucketized_column(price, boundaries=[0, 50])
+    hash_bucket_size = 10
+    price_cross_wire = fc.crossed_column(
+        [bucketized_price, 'wire'], hash_bucket_size)
+    features = {
+        'price': constant_op.constant([[1., 2.], [5., 6.]]),
+        'wire': sparse_tensor.SparseTensor(
+            values=['omar', 'stringer', 'marlo'],
+            indices=[[0, 0], [1, 0], [1, 1]],
+            dense_shape=[2, 2]),
+    }
+    outputs = _transform_features(features, [price_cross_wire], None)
+    output = outputs[price_cross_wire]
+    with self.test_session() as sess:
+      output_val = sess.run(output)
+      self.assertAllEqual(
+          [[0, 0], [0, 1], [1, 0], [1, 1], [1, 2], [1, 3]], output_val.indices)
+      for val in output_val.values:
+        self.assertIn(val, list(range(hash_bucket_size)))
+      self.assertAllEqual([2, 4], output_val.dense_shape)
+
+  def test_get_sparse_tensors(self):
+    a = fc.numeric_column('a', dtype=dtypes.int32, shape=(2,))
+    b = fc.bucketized_column(a, boundaries=(0, 1))
+    crossed1 = fc.crossed_column(['d1', 'd2'], 10)
+    crossed2 = fc.crossed_column([b, 'c', crossed1], 15, hash_key=5)
+    with ops.Graph().as_default():
+      transformation_cache = FeatureTransformationCache({
+          'a':
+              constant_op.constant(((-1., .5), (.5, 1.))),
+          'c':
+              sparse_tensor.SparseTensor(
+                  indices=((0, 0), (1, 0), (1, 1)),
+                  values=['cA', 'cB', 'cC'],
+                  dense_shape=(2, 2)),
+          'd1':
+              sparse_tensor.SparseTensor(
+                  indices=((0, 0), (1, 0), (1, 1)),
+                  values=['d1A', 'd1B', 'd1C'],
+                  dense_shape=(2, 2)),
+          'd2':
+              sparse_tensor.SparseTensor(
+                  indices=((0, 0), (1, 0), (1, 1)),
+                  values=['d2A', 'd2B', 'd2C'],
+                  dense_shape=(2, 2)),
+      })
+      id_weight_pair = crossed2.get_sparse_tensors(transformation_cache, None)
+      with _initialized_session():
+        id_tensor_eval = id_weight_pair.id_tensor.eval()
+        self.assertAllEqual(
+            ((0, 0), (0, 1), (1, 0), (1, 1), (1, 2), (1, 3), (1, 4), (1, 5),
+             (1, 6), (1, 7), (1, 8), (1, 9), (1, 10), (1, 11), (1, 12), (1, 13),
+             (1, 14), (1, 15)),
+            id_tensor_eval.indices)
+        # Check exact hashed output. If hashing changes this test will break.
+        # All values are within [0, hash_bucket_size).
+        expected_values = (
+            6, 14, 0, 13, 8, 8, 10, 12, 2, 0, 1, 9, 8, 12, 2, 0, 10, 11)
+        self.assertAllEqual(expected_values, id_tensor_eval.values)
+        self.assertAllEqual((2, 16), id_tensor_eval.dense_shape)
+
+  def test_get_sparse_tensors_simple(self):
+    """Same as test_get_sparse_tensors, but with simpler values."""
+    a = fc.numeric_column('a', dtype=dtypes.int32, shape=(2,))
+    b = fc.bucketized_column(a, boundaries=(0, 1))
+    crossed = fc.crossed_column([b, 'c'], hash_bucket_size=5, hash_key=5)
+    with ops.Graph().as_default():
+      transformation_cache = FeatureTransformationCache({
+          'a':
+              constant_op.constant(((-1., .5), (.5, 1.))),
+          'c':
+              sparse_tensor.SparseTensor(
+                  indices=((0, 0), (1, 0), (1, 1)),
+                  values=['cA', 'cB', 'cC'],
+                  dense_shape=(2, 2)),
+      })
+      id_weight_pair = crossed.get_sparse_tensors(transformation_cache, None)
+      with _initialized_session():
+        id_tensor_eval = id_weight_pair.id_tensor.eval()
+        self.assertAllEqual(
+            ((0, 0), (0, 1), (1, 0), (1, 1), (1, 2), (1, 3)),
+            id_tensor_eval.indices)
+        # Check exact hashed output. If hashing changes this test will break.
+        # All values are within [0, hash_bucket_size).
+        expected_values = (1, 0, 1, 3, 4, 2)
+        self.assertAllEqual(expected_values, id_tensor_eval.values)
+        self.assertAllEqual((2, 4), id_tensor_eval.dense_shape)
+
+  def test_linear_model(self):
+    """Tests linear_model.
+
+    Uses data from test_get_sparse_tesnsors_simple.
+    """
+    a = fc_old.numeric_column('a', dtype=dtypes.int32, shape=(2,))
+    b = fc_old.bucketized_column(a, boundaries=(0, 1))
+    crossed = fc_old.crossed_column([b, 'c'], hash_bucket_size=5, hash_key=5)
+    with ops.Graph().as_default():
+      predictions = fc.linear_model({
+          'a': constant_op.constant(((-1., .5), (.5, 1.))),
+          'c': sparse_tensor.SparseTensor(
+              indices=((0, 0), (1, 0), (1, 1)),
+              values=['cA', 'cB', 'cC'],
+              dense_shape=(2, 2)),
+      }, (crossed,))
+      bias = get_linear_model_bias()
+      crossed_var = get_linear_model_column_var(crossed)
+      with _initialized_session() as sess:
+        self.assertAllClose((0.,), bias.eval())
+        self.assertAllClose(
+            ((0.,), (0.,), (0.,), (0.,), (0.,)), crossed_var.eval())
+        self.assertAllClose(((0.,), (0.,)), predictions.eval())
+        sess.run(crossed_var.assign(((1.,), (2.,), (3.,), (4.,), (5.,))))
+        # Expected ids after cross = (1, 0, 1, 3, 4, 2)
+        self.assertAllClose(((3.,), (14.,)), predictions.eval())
+        sess.run(bias.assign((.1,)))
+        self.assertAllClose(((3.1,), (14.1,)), predictions.eval())
+
+  def test_linear_model_with_weights(self):
+
+    class _TestColumnWithWeights(fc_old._CategoricalColumn):
+      """Produces sparse IDs and sparse weights."""
+
+      @property
+      def name(self):
+        return 'test_column'
+
+      @property
+      def _parse_example_spec(self):
+        return {
+            self.name: parsing_ops.VarLenFeature(dtypes.int32),
+            '{}_weights'.format(self.name): parsing_ops.VarLenFeature(
+                dtypes.float32),
+            }
+
+      @property
+      def _num_buckets(self):
+        return 5
+
+      def _transform_feature(self, inputs):
+        return (inputs.get(self.name),
+                inputs.get('{}_weights'.format(self.name)))
+
+      def _get_sparse_tensors(self, inputs, weight_collections=None,
+                              trainable=None):
+        """Populates both id_tensor and weight_tensor."""
+        ids_and_weights = inputs.get(self)
+        return fc_old._CategoricalColumn.IdWeightPair(
+            id_tensor=ids_and_weights[0], weight_tensor=ids_and_weights[1])
+
+    t = _TestColumnWithWeights()
+    crossed = fc_old.crossed_column([t, 'c'], hash_bucket_size=5, hash_key=5)
+    with ops.Graph().as_default():
+      with self.assertRaisesRegexp(
+          ValueError,
+          'crossed_column does not support weight_tensor.*{}'.format(t.name)):
+        fc.linear_model({
+            t.name: sparse_tensor.SparseTensor(
+                indices=((0, 0), (1, 0), (1, 1)),
+                values=[0, 1, 2],
+                dense_shape=(2, 2)),
+            '{}_weights'.format(t.name): sparse_tensor.SparseTensor(
+                indices=((0, 0), (1, 0), (1, 1)),
+                values=[1., 10., 2.],
+                dense_shape=(2, 2)),
+            'c': sparse_tensor.SparseTensor(
+                indices=((0, 0), (1, 0), (1, 1)),
+                values=['cA', 'cB', 'cC'],
+                dense_shape=(2, 2)),
+        }, (crossed,))
+
+  def test_keras_linear_model(self):
+    """Tests _LinearModel.
+
+    Uses data from test_get_sparse_tesnsors_simple.
+    """
+    a = fc_old.numeric_column('a', dtype=dtypes.int32, shape=(2,))
+    b = fc_old.bucketized_column(a, boundaries=(0, 1))
+    crossed = fc_old.crossed_column([b, 'c'], hash_bucket_size=5, hash_key=5)
+    with ops.Graph().as_default():
+      predictions = get_keras_linear_model_predictions({
+          'a':
+              constant_op.constant(((-1., .5), (.5, 1.))),
+          'c':
+              sparse_tensor.SparseTensor(
+                  indices=((0, 0), (1, 0), (1, 1)),
+                  values=['cA', 'cB', 'cC'],
+                  dense_shape=(2, 2)),
+      }, (crossed,))
+      bias = get_linear_model_bias()
+      crossed_var = get_linear_model_column_var(crossed)
+      with _initialized_session() as sess:
+        self.assertAllClose((0.,), bias.eval())
+        self.assertAllClose(((0.,), (0.,), (0.,), (0.,), (0.,)),
+                            crossed_var.eval())
+        self.assertAllClose(((0.,), (0.,)), predictions.eval())
+        sess.run(crossed_var.assign(((1.,), (2.,), (3.,), (4.,), (5.,))))
+        # Expected ids after cross = (1, 0, 1, 3, 4, 2)
+        self.assertAllClose(((3.,), (14.,)), predictions.eval())
+        sess.run(bias.assign((.1,)))
+        self.assertAllClose(((3.1,), (14.1,)), predictions.eval())
+
+  def test_keras_linear_model_with_weights(self):
+
+    class _TestColumnWithWeights(fc_old._CategoricalColumn):
+      """Produces sparse IDs and sparse weights."""
+
+      @property
+      def name(self):
+        return 'test_column'
+
+      @property
+      def _parse_example_spec(self):
+        return {
+            self.name:
+                parsing_ops.VarLenFeature(dtypes.int32),
+            '{}_weights'.format(self.name):
+                parsing_ops.VarLenFeature(dtypes.float32),
+        }
+
+      @property
+      def _num_buckets(self):
+        return 5
+
+      def _transform_feature(self, inputs):
+        return (inputs.get(self.name),
+                inputs.get('{}_weights'.format(self.name)))
+
+      def _get_sparse_tensors(self,
+                              inputs,
+                              weight_collections=None,
+                              trainable=None):
+        """Populates both id_tensor and weight_tensor."""
+        ids_and_weights = inputs.get(self)
+        return fc_old._CategoricalColumn.IdWeightPair(
+            id_tensor=ids_and_weights[0], weight_tensor=ids_and_weights[1])
+
+    t = _TestColumnWithWeights()
+    crossed = fc_old.crossed_column([t, 'c'], hash_bucket_size=5, hash_key=5)
+    with ops.Graph().as_default():
+      with self.assertRaisesRegexp(
+          ValueError,
+          'crossed_column does not support weight_tensor.*{}'.format(t.name)):
+        get_keras_linear_model_predictions({
+            t.name:
+                sparse_tensor.SparseTensor(
+                    indices=((0, 0), (1, 0), (1, 1)),
+                    values=[0, 1, 2],
+                    dense_shape=(2, 2)),
+            '{}_weights'.format(t.name):
+                sparse_tensor.SparseTensor(
+                    indices=((0, 0), (1, 0), (1, 1)),
+                    values=[1., 10., 2.],
+                    dense_shape=(2, 2)),
+            'c':
+                sparse_tensor.SparseTensor(
+                    indices=((0, 0), (1, 0), (1, 1)),
+                    values=['cA', 'cB', 'cC'],
+                    dense_shape=(2, 2)),
+        }, (crossed,))
+
+
+def get_linear_model_bias(name='linear_model'):
+  with variable_scope.variable_scope(name, reuse=True):
+    return variable_scope.get_variable('bias_weights')
+
+
+def get_linear_model_column_var(column, name='linear_model'):
+  return ops.get_collection(ops.GraphKeys.GLOBAL_VARIABLES,
+                            name + '/' + column.name)[0]
+
+
+def get_keras_linear_model_predictions(features,
+                                       feature_columns,
+                                       units=1,
+                                       sparse_combiner='sum',
+                                       weight_collections=None,
+                                       trainable=True,
+                                       cols_to_vars=None):
+  keras_linear_model = _LinearModel(
+      feature_columns,
+      units,
+      sparse_combiner,
+      weight_collections,
+      trainable,
+      name='linear_model')
+  retval = keras_linear_model(features)  # pylint: disable=not-callable
+  if cols_to_vars is not None:
+    cols_to_vars.update(keras_linear_model.cols_to_vars())
+  return retval
+
+
+class LinearModelTest(test.TestCase):
+
+  def test_raises_if_empty_feature_columns(self):
+    with self.assertRaisesRegexp(ValueError,
+                                 'feature_columns must not be empty'):
+      fc.linear_model(features={}, feature_columns=[])
+
+  def test_should_be_feature_column(self):
+    with self.assertRaisesRegexp(ValueError, 'must be a _FeatureColumn'):
+      fc.linear_model(features={'a': [[0]]}, feature_columns='NotSupported')
+
+  def test_should_be_dense_or_categorical_column(self):
+
+    class NotSupportedColumn(fc_old._FeatureColumn):
+
+      @property
+      def name(self):
+        return 'NotSupportedColumn'
+
+      def _transform_feature(self, cache):
+        pass
+
+      @property
+      def _parse_example_spec(self):
+        pass
+
+    with self.assertRaisesRegexp(
+        ValueError, 'must be either a _DenseColumn or _CategoricalColumn'):
+      fc.linear_model(
+          features={'a': [[0]]}, feature_columns=[NotSupportedColumn()])
+
+  def test_does_not_support_dict_columns(self):
+    with self.assertRaisesRegexp(
+        ValueError, 'Expected feature_columns to be iterable, found dict.'):
+      fc.linear_model(
+          features={'a': [[0]]},
+          feature_columns={'a': fc_old.numeric_column('a')})
+
+  def test_raises_if_duplicate_name(self):
+    with self.assertRaisesRegexp(
+        ValueError, 'Duplicate feature column name found for columns'):
+      fc.linear_model(
+          features={'a': [[0]]},
+          feature_columns=[
+              fc_old.numeric_column('a'),
+              fc_old.numeric_column('a')
+          ])
+
+  def test_dense_bias(self):
+    price = fc_old.numeric_column('price')
+    with ops.Graph().as_default():
+      features = {'price': [[1.], [5.]]}
+      predictions = fc.linear_model(features, [price])
+      bias = get_linear_model_bias()
+      price_var = get_linear_model_column_var(price)
+      with _initialized_session() as sess:
+        self.assertAllClose([0.], bias.eval())
+        sess.run(price_var.assign([[10.]]))
+        sess.run(bias.assign([5.]))
+        self.assertAllClose([[15.], [55.]], predictions.eval())
+
+  def test_sparse_bias(self):
+    wire_cast = fc_old.categorical_column_with_hash_bucket('wire_cast', 4)
+    with ops.Graph().as_default():
+      wire_tensor = sparse_tensor.SparseTensor(
+          values=['omar', 'stringer', 'marlo'],  # hashed to = [2, 0, 3]
+          indices=[[0, 0], [1, 0], [1, 1]],
+          dense_shape=[2, 2])
+      features = {'wire_cast': wire_tensor}
+      predictions = fc.linear_model(features, [wire_cast])
+      bias = get_linear_model_bias()
+      wire_cast_var = get_linear_model_column_var(wire_cast)
+      with _initialized_session() as sess:
+        self.assertAllClose([0.], bias.eval())
+        self.assertAllClose([[0.], [0.], [0.], [0.]], wire_cast_var.eval())
+        sess.run(wire_cast_var.assign([[10.], [100.], [1000.], [10000.]]))
+        sess.run(bias.assign([5.]))
+        self.assertAllClose([[1005.], [10015.]], predictions.eval())
+
+  def test_dense_and_sparse_bias(self):
+    wire_cast = fc_old.categorical_column_with_hash_bucket('wire_cast', 4)
+    price = fc_old.numeric_column('price')
+    with ops.Graph().as_default():
+      wire_tensor = sparse_tensor.SparseTensor(
+          values=['omar', 'stringer', 'marlo'],  # hashed to = [2, 0, 3]
+          indices=[[0, 0], [1, 0], [1, 1]],
+          dense_shape=[2, 2])
+      features = {'wire_cast': wire_tensor, 'price': [[1.], [5.]]}
+      predictions = fc.linear_model(features, [wire_cast, price])
+      bias = get_linear_model_bias()
+      wire_cast_var = get_linear_model_column_var(wire_cast)
+      price_var = get_linear_model_column_var(price)
+      with _initialized_session() as sess:
+        sess.run(wire_cast_var.assign([[10.], [100.], [1000.], [10000.]]))
+        sess.run(bias.assign([5.]))
+        sess.run(price_var.assign([[10.]]))
+        self.assertAllClose([[1015.], [10065.]], predictions.eval())
+
+  def test_dense_and_sparse_column(self):
+    """When the column is both dense and sparse, uses sparse tensors."""
+
+    class _DenseAndSparseColumn(fc_old._DenseColumn, fc_old._CategoricalColumn):
+
+      @property
+      def name(self):
+        return 'dense_and_sparse_column'
+
+      @property
+      def _parse_example_spec(self):
+        return {self.name: parsing_ops.VarLenFeature(self.dtype)}
+
+      def _transform_feature(self, inputs):
+        return inputs.get(self.name)
+
+      @property
+      def _variable_shape(self):
+        raise ValueError('Should not use this method.')
+
+      def _get_dense_tensor(self, inputs, weight_collections=None,
+                            trainable=None):
+        raise ValueError('Should not use this method.')
+
+      @property
+      def _num_buckets(self):
+        return 4
+
+      def _get_sparse_tensors(self, inputs, weight_collections=None,
+                              trainable=None):
+        sp_tensor = sparse_tensor.SparseTensor(
+            indices=[[0, 0], [1, 0], [1, 1]],
+            values=[2, 0, 3],
+            dense_shape=[2, 2])
+        return fc_old._CategoricalColumn.IdWeightPair(sp_tensor, None)
+
+    dense_and_sparse_column = _DenseAndSparseColumn()
+    with ops.Graph().as_default():
+      sp_tensor = sparse_tensor.SparseTensor(
+          values=['omar', 'stringer', 'marlo'],
+          indices=[[0, 0], [1, 0], [1, 1]],
+          dense_shape=[2, 2])
+      features = {dense_and_sparse_column.name: sp_tensor}
+      predictions = fc.linear_model(features, [dense_and_sparse_column])
+      bias = get_linear_model_bias()
+      dense_and_sparse_column_var = get_linear_model_column_var(
+          dense_and_sparse_column)
+      with _initialized_session() as sess:
+        sess.run(dense_and_sparse_column_var.assign(
+            [[10.], [100.], [1000.], [10000.]]))
+        sess.run(bias.assign([5.]))
+        self.assertAllClose([[1005.], [10015.]], predictions.eval())
+
+  def test_dense_multi_output(self):
+    price = fc_old.numeric_column('price')
+    with ops.Graph().as_default():
+      features = {'price': [[1.], [5.]]}
+      predictions = fc.linear_model(features, [price], units=3)
+      bias = get_linear_model_bias()
+      price_var = get_linear_model_column_var(price)
+      with _initialized_session() as sess:
+        self.assertAllClose(np.zeros((3,)), bias.eval())
+        self.assertAllClose(np.zeros((1, 3)), price_var.eval())
+        sess.run(price_var.assign([[10., 100., 1000.]]))
+        sess.run(bias.assign([5., 6., 7.]))
+        self.assertAllClose([[15., 106., 1007.], [55., 506., 5007.]],
+                            predictions.eval())
+
+  def test_sparse_multi_output(self):
+    wire_cast = fc_old.categorical_column_with_hash_bucket('wire_cast', 4)
+    with ops.Graph().as_default():
+      wire_tensor = sparse_tensor.SparseTensor(
+          values=['omar', 'stringer', 'marlo'],  # hashed to = [2, 0, 3]
+          indices=[[0, 0], [1, 0], [1, 1]],
+          dense_shape=[2, 2])
+      features = {'wire_cast': wire_tensor}
+      predictions = fc.linear_model(features, [wire_cast], units=3)
+      bias = get_linear_model_bias()
+      wire_cast_var = get_linear_model_column_var(wire_cast)
+      with _initialized_session() as sess:
+        self.assertAllClose(np.zeros((3,)), bias.eval())
+        self.assertAllClose(np.zeros((4, 3)), wire_cast_var.eval())
+        sess.run(
+            wire_cast_var.assign([[10., 11., 12.], [100., 110., 120.], [
+                1000., 1100., 1200.
+            ], [10000., 11000., 12000.]]))
+        sess.run(bias.assign([5., 6., 7.]))
+        self.assertAllClose([[1005., 1106., 1207.], [10015., 11017., 12019.]],
+                            predictions.eval())
+
+  def test_dense_multi_dimension(self):
+    price = fc_old.numeric_column('price', shape=2)
+    with ops.Graph().as_default():
+      features = {'price': [[1., 2.], [5., 6.]]}
+      predictions = fc.linear_model(features, [price])
+      price_var = get_linear_model_column_var(price)
+      with _initialized_session() as sess:
+        self.assertAllClose([[0.], [0.]], price_var.eval())
+        sess.run(price_var.assign([[10.], [100.]]))
+        self.assertAllClose([[210.], [650.]], predictions.eval())
+
+  def test_sparse_multi_rank(self):
+    wire_cast = fc_old.categorical_column_with_hash_bucket('wire_cast', 4)
+    with ops.Graph().as_default():
+      wire_tensor = array_ops.sparse_placeholder(dtypes.string)
+      wire_value = sparse_tensor.SparseTensorValue(
+          values=['omar', 'stringer', 'marlo', 'omar'],  # hashed = [2, 0, 3, 2]
+          indices=[[0, 0, 0], [0, 1, 0], [1, 0, 0], [1, 0, 1]],
+          dense_shape=[2, 2, 2])
+      features = {'wire_cast': wire_tensor}
+      predictions = fc.linear_model(features, [wire_cast])
+      wire_cast_var = get_linear_model_column_var(wire_cast)
+      with _initialized_session() as sess:
+        self.assertAllClose(np.zeros((4, 1)), wire_cast_var.eval())
+        self.assertAllClose(
+            np.zeros((2, 1)),
+            predictions.eval(feed_dict={wire_tensor: wire_value}))
+        sess.run(wire_cast_var.assign([[10.], [100.], [1000.], [10000.]]))
+        self.assertAllClose(
+            [[1010.], [11000.]],
+            predictions.eval(feed_dict={wire_tensor: wire_value}))
+
+  def test_sparse_combiner(self):
+    wire_cast = fc_old.categorical_column_with_hash_bucket('wire_cast', 4)
+    with ops.Graph().as_default():
+      wire_tensor = sparse_tensor.SparseTensor(
+          values=['omar', 'stringer', 'marlo'],  # hashed to = [2, 0, 3]
+          indices=[[0, 0], [1, 0], [1, 1]],
+          dense_shape=[2, 2])
+      features = {'wire_cast': wire_tensor}
+      predictions = fc.linear_model(
+          features, [wire_cast], sparse_combiner='mean')
+      bias = get_linear_model_bias()
+      wire_cast_var = get_linear_model_column_var(wire_cast)
+      with _initialized_session() as sess:
+        sess.run(wire_cast_var.assign([[10.], [100.], [1000.], [10000.]]))
+        sess.run(bias.assign([5.]))
+        self.assertAllClose([[1005.], [5010.]], predictions.eval())
+
+  def test_sparse_combiner_with_negative_weights(self):
+    wire_cast = fc_old.categorical_column_with_hash_bucket('wire_cast', 4)
+    wire_cast_weights = fc_old.weighted_categorical_column(wire_cast, 'weights')
+
+    with ops.Graph().as_default():
+      wire_tensor = sparse_tensor.SparseTensor(
+          values=['omar', 'stringer', 'marlo'],  # hashed to = [2, 0, 3]
+          indices=[[0, 0], [1, 0], [1, 1]],
+          dense_shape=[2, 2])
+      features = {
+          'wire_cast': wire_tensor,
+          'weights': constant_op.constant([[1., 1., -1.0]])
+      }
+      predictions = fc.linear_model(
+          features, [wire_cast_weights], sparse_combiner='sum')
+      bias = get_linear_model_bias()
+      wire_cast_var = get_linear_model_column_var(wire_cast)
+      with _initialized_session() as sess:
+        sess.run(wire_cast_var.assign([[10.], [100.], [1000.], [10000.]]))
+        sess.run(bias.assign([5.]))
+        self.assertAllClose([[1005.], [-9985.]], predictions.eval())
+
+  def test_dense_multi_dimension_multi_output(self):
+    price = fc_old.numeric_column('price', shape=2)
+    with ops.Graph().as_default():
+      features = {'price': [[1., 2.], [5., 6.]]}
+      predictions = fc.linear_model(features, [price], units=3)
+      bias = get_linear_model_bias()
+      price_var = get_linear_model_column_var(price)
+      with _initialized_session() as sess:
+        self.assertAllClose(np.zeros((3,)), bias.eval())
+        self.assertAllClose(np.zeros((2, 3)), price_var.eval())
+        sess.run(price_var.assign([[1., 2., 3.], [10., 100., 1000.]]))
+        sess.run(bias.assign([2., 3., 4.]))
+        self.assertAllClose([[23., 205., 2007.], [67., 613., 6019.]],
+                            predictions.eval())
+
+  def test_raises_if_shape_mismatch(self):
+    price = fc_old.numeric_column('price', shape=2)
+    with ops.Graph().as_default():
+      features = {'price': [[1.], [5.]]}
+      with self.assertRaisesRegexp(
+          Exception,
+          r'Cannot reshape a tensor with 2 elements to shape \[2,2\]'):
+        fc.linear_model(features, [price])
+
+  def test_dense_reshaping(self):
+    price = fc_old.numeric_column('price', shape=[1, 2])
+    with ops.Graph().as_default():
+      features = {'price': [[[1., 2.]], [[5., 6.]]]}
+      predictions = fc.linear_model(features, [price])
+      bias = get_linear_model_bias()
+      price_var = get_linear_model_column_var(price)
+      with _initialized_session() as sess:
+        self.assertAllClose([0.], bias.eval())
+        self.assertAllClose([[0.], [0.]], price_var.eval())
+        self.assertAllClose([[0.], [0.]], predictions.eval())
+        sess.run(price_var.assign([[10.], [100.]]))
+        self.assertAllClose([[210.], [650.]], predictions.eval())
+
+  def test_dense_multi_column(self):
+    price1 = fc_old.numeric_column('price1', shape=2)
+    price2 = fc_old.numeric_column('price2')
+    with ops.Graph().as_default():
+      features = {
+          'price1': [[1., 2.], [5., 6.]],
+          'price2': [[3.], [4.]]
+      }
+      predictions = fc.linear_model(features, [price1, price2])
+      bias = get_linear_model_bias()
+      price1_var = get_linear_model_column_var(price1)
+      price2_var = get_linear_model_column_var(price2)
+      with _initialized_session() as sess:
+        self.assertAllClose([0.], bias.eval())
+        self.assertAllClose([[0.], [0.]], price1_var.eval())
+        self.assertAllClose([[0.]], price2_var.eval())
+        self.assertAllClose([[0.], [0.]], predictions.eval())
+        sess.run(price1_var.assign([[10.], [100.]]))
+        sess.run(price2_var.assign([[1000.]]))
+        sess.run(bias.assign([7.]))
+        self.assertAllClose([[3217.], [4657.]], predictions.eval())
+
+  def test_fills_cols_to_vars(self):
+    price1 = fc_old.numeric_column('price1', shape=2)
+    price2 = fc_old.numeric_column('price2')
+    with ops.Graph().as_default():
+      features = {'price1': [[1., 2.], [5., 6.]], 'price2': [[3.], [4.]]}
+      cols_to_vars = {}
+      fc.linear_model(features, [price1, price2], cols_to_vars=cols_to_vars)
+      bias = get_linear_model_bias()
+      price1_var = get_linear_model_column_var(price1)
+      price2_var = get_linear_model_column_var(price2)
+      self.assertAllEqual(cols_to_vars['bias'], [bias])
+      self.assertAllEqual(cols_to_vars[price1], [price1_var])
+      self.assertAllEqual(cols_to_vars[price2], [price2_var])
+
+  def test_fills_cols_to_vars_partitioned_variables(self):
+    price1 = fc_old.numeric_column('price1', shape=2)
+    price2 = fc_old.numeric_column('price2', shape=3)
+    with ops.Graph().as_default():
+      features = {
+          'price1': [[1., 2.], [6., 7.]],
+          'price2': [[3., 4., 5.], [8., 9., 10.]]
+      }
+      cols_to_vars = {}
+      with variable_scope.variable_scope(
+          'linear',
+          partitioner=partitioned_variables.fixed_size_partitioner(2, axis=0)):
+        fc.linear_model(features, [price1, price2], cols_to_vars=cols_to_vars)
+      with _initialized_session():
+        self.assertEqual([0.], cols_to_vars['bias'][0].eval())
+        # Partitioning shards the [2, 1] price1 var into 2 [1, 1] Variables.
+        self.assertAllEqual([[0.]], cols_to_vars[price1][0].eval())
+        self.assertAllEqual([[0.]], cols_to_vars[price1][1].eval())
+        # Partitioning shards the [3, 1] price2 var into a [2, 1] Variable and
+        # a [1, 1] Variable.
+        self.assertAllEqual([[0.], [0.]], cols_to_vars[price2][0].eval())
+        self.assertAllEqual([[0.]], cols_to_vars[price2][1].eval())
+
+  def test_dense_collection(self):
+    price = fc_old.numeric_column('price')
+    with ops.Graph().as_default() as g:
+      features = {'price': [[1.], [5.]]}
+      fc.linear_model(features, [price], weight_collections=['my-vars'])
+      my_vars = g.get_collection('my-vars')
+      bias = get_linear_model_bias()
+      price_var = get_linear_model_column_var(price)
+      self.assertIn(bias, my_vars)
+      self.assertIn(price_var, my_vars)
+
+  def test_sparse_collection(self):
+    wire_cast = fc_old.categorical_column_with_hash_bucket('wire_cast', 4)
+    with ops.Graph().as_default() as g:
+      wire_tensor = sparse_tensor.SparseTensor(
+          values=['omar'], indices=[[0, 0]], dense_shape=[1, 1])
+      features = {'wire_cast': wire_tensor}
+      fc.linear_model(
+          features, [wire_cast], weight_collections=['my-vars'])
+      my_vars = g.get_collection('my-vars')
+      bias = get_linear_model_bias()
+      wire_cast_var = get_linear_model_column_var(wire_cast)
+      self.assertIn(bias, my_vars)
+      self.assertIn(wire_cast_var, my_vars)
+
+  def test_dense_trainable_default(self):
+    price = fc_old.numeric_column('price')
+    with ops.Graph().as_default() as g:
+      features = {'price': [[1.], [5.]]}
+      fc.linear_model(features, [price])
+      bias = get_linear_model_bias()
+      price_var = get_linear_model_column_var(price)
+      trainable_vars = g.get_collection(ops.GraphKeys.TRAINABLE_VARIABLES)
+      self.assertIn(bias, trainable_vars)
+      self.assertIn(price_var, trainable_vars)
+
+  def test_sparse_trainable_default(self):
+    wire_cast = fc_old.categorical_column_with_hash_bucket('wire_cast', 4)
+    with ops.Graph().as_default() as g:
+      wire_tensor = sparse_tensor.SparseTensor(
+          values=['omar'], indices=[[0, 0]], dense_shape=[1, 1])
+      features = {'wire_cast': wire_tensor}
+      fc.linear_model(features, [wire_cast])
+      trainable_vars = g.get_collection(ops.GraphKeys.TRAINABLE_VARIABLES)
+      bias = get_linear_model_bias()
+      wire_cast_var = get_linear_model_column_var(wire_cast)
+      self.assertIn(bias, trainable_vars)
+      self.assertIn(wire_cast_var, trainable_vars)
+
+  def test_dense_trainable_false(self):
+    price = fc_old.numeric_column('price')
+    with ops.Graph().as_default() as g:
+      features = {'price': [[1.], [5.]]}
+      fc.linear_model(features, [price], trainable=False)
+      trainable_vars = g.get_collection(ops.GraphKeys.TRAINABLE_VARIABLES)
+      self.assertEqual([], trainable_vars)
+
+  def test_sparse_trainable_false(self):
+    wire_cast = fc_old.categorical_column_with_hash_bucket('wire_cast', 4)
+    with ops.Graph().as_default() as g:
+      wire_tensor = sparse_tensor.SparseTensor(
+          values=['omar'], indices=[[0, 0]], dense_shape=[1, 1])
+      features = {'wire_cast': wire_tensor}
+      fc.linear_model(features, [wire_cast], trainable=False)
+      trainable_vars = g.get_collection(ops.GraphKeys.TRAINABLE_VARIABLES)
+      self.assertEqual([], trainable_vars)
+
+  def test_column_order(self):
+    price_a = fc_old.numeric_column('price_a')
+    price_b = fc_old.numeric_column('price_b')
+    wire_cast = fc_old.categorical_column_with_hash_bucket('wire_cast', 4)
+    with ops.Graph().as_default() as g:
+      features = {
+          'price_a': [[1.]],
+          'price_b': [[3.]],
+          'wire_cast':
+              sparse_tensor.SparseTensor(
+                  values=['omar'], indices=[[0, 0]], dense_shape=[1, 1])
+      }
+      fc.linear_model(
+          features, [price_a, wire_cast, price_b],
+          weight_collections=['my-vars'])
+      my_vars = g.get_collection('my-vars')
+      self.assertIn('price_a', my_vars[0].name)
+      self.assertIn('price_b', my_vars[1].name)
+      self.assertIn('wire_cast', my_vars[2].name)
+
+    with ops.Graph().as_default() as g:
+      features = {
+          'price_a': [[1.]],
+          'price_b': [[3.]],
+          'wire_cast':
+              sparse_tensor.SparseTensor(
+                  values=['omar'], indices=[[0, 0]], dense_shape=[1, 1])
+      }
+      fc.linear_model(
+          features, [wire_cast, price_b, price_a],
+          weight_collections=['my-vars'])
+      my_vars = g.get_collection('my-vars')
+      self.assertIn('price_a', my_vars[0].name)
+      self.assertIn('price_b', my_vars[1].name)
+      self.assertIn('wire_cast', my_vars[2].name)
+
+  def test_static_batch_size_mismatch(self):
+    price1 = fc_old.numeric_column('price1')
+    price2 = fc_old.numeric_column('price2')
+    with ops.Graph().as_default():
+      features = {
+          'price1': [[1.], [5.], [7.]],  # batchsize = 3
+          'price2': [[3.], [4.]]  # batchsize = 2
+      }
+    with self.assertRaisesRegexp(
+        ValueError,
+        'Batch size \(first dimension\) of each feature must be same.'):  # pylint: disable=anomalous-backslash-in-string
+      fc.linear_model(features, [price1, price2])
+
+  def test_subset_of_static_batch_size_mismatch(self):
+    price1 = fc_old.numeric_column('price1')
+    price2 = fc_old.numeric_column('price2')
+    price3 = fc_old.numeric_column('price3')
+    with ops.Graph().as_default():
+      features = {
+          'price1': array_ops.placeholder(dtype=dtypes.int64),  # batchsize = 3
+          'price2': [[3.], [4.]],  # batchsize = 2
+          'price3': [[3.], [4.], [5.]]  # batchsize = 3
+      }
+      with self.assertRaisesRegexp(
+          ValueError,
+          'Batch size \(first dimension\) of each feature must be same.'):  # pylint: disable=anomalous-backslash-in-string
+        fc.linear_model(features, [price1, price2, price3])
+
+  def test_runtime_batch_size_mismatch(self):
+    price1 = fc_old.numeric_column('price1')
+    price2 = fc_old.numeric_column('price2')
+    with ops.Graph().as_default():
+      features = {
+          'price1': array_ops.placeholder(dtype=dtypes.int64),  # batchsize = 3
+          'price2': [[3.], [4.]]  # batchsize = 2
+      }
+      predictions = fc.linear_model(features, [price1, price2])
+      with _initialized_session() as sess:
+        with self.assertRaisesRegexp(errors.OpError,
+                                     'must have the same size and shape'):
+          sess.run(
+              predictions, feed_dict={features['price1']: [[1.], [5.], [7.]]})
+
+  def test_runtime_batch_size_matches(self):
+    price1 = fc_old.numeric_column('price1')
+    price2 = fc_old.numeric_column('price2')
+    with ops.Graph().as_default():
+      features = {
+          'price1': array_ops.placeholder(dtype=dtypes.int64),  # batchsize = 2
+          'price2': array_ops.placeholder(dtype=dtypes.int64),  # batchsize = 2
+      }
+      predictions = fc.linear_model(features, [price1, price2])
+      with _initialized_session() as sess:
+        sess.run(
+            predictions,
+            feed_dict={
+                features['price1']: [[1.], [5.]],
+                features['price2']: [[1.], [5.]],
+            })
+
+  def test_with_numpy_input_fn(self):
+    price = fc_old.numeric_column('price')
+    price_buckets = fc_old.bucketized_column(
+        price, boundaries=[
+            0.,
+            10.,
+            100.,
+        ])
+    body_style = fc_old.categorical_column_with_vocabulary_list(
+        'body-style', vocabulary_list=['hardtop', 'wagon', 'sedan'])
+
+    input_fn = numpy_io.numpy_input_fn(
+        x={
+            'price': np.array([-1., 2., 13., 104.]),
+            'body-style': np.array(['sedan', 'hardtop', 'wagon', 'sedan']),
+        },
+        batch_size=2,
+        shuffle=False)
+    features = input_fn()
+    net = fc.linear_model(features, [price_buckets, body_style])
+    # self.assertEqual(1 + 3 + 5, net.shape[1])
+    with _initialized_session() as sess:
+      coord = coordinator.Coordinator()
+      threads = queue_runner_impl.start_queue_runners(sess, coord=coord)
+
+      bias = get_linear_model_bias()
+      price_buckets_var = get_linear_model_column_var(price_buckets)
+      body_style_var = get_linear_model_column_var(body_style)
+
+      sess.run(price_buckets_var.assign([[10.], [100.], [1000.], [10000.]]))
+      sess.run(body_style_var.assign([[-10.], [-100.], [-1000.]]))
+      sess.run(bias.assign([5.]))
+
+      self.assertAllClose([[10 - 1000 + 5.], [100 - 10 + 5.]], sess.run(net))
+
+      coord.request_stop()
+      coord.join(threads)
+
+  def test_with_1d_sparse_tensor(self):
+    price = fc_old.numeric_column('price')
+    price_buckets = fc_old.bucketized_column(
+        price, boundaries=[
+            0.,
+            10.,
+            100.,
+        ])
+    body_style = fc_old.categorical_column_with_vocabulary_list(
+        'body-style', vocabulary_list=['hardtop', 'wagon', 'sedan'])
+
+    # Provides 1-dim tensor and dense tensor.
+    features = {
+        'price': constant_op.constant([-1., 12.,]),
+        'body-style': sparse_tensor.SparseTensor(
+            indices=((0,), (1,)),
+            values=('sedan', 'hardtop'),
+            dense_shape=(2,)),
+    }
+    self.assertEqual(1, features['price'].shape.ndims)
+    self.assertEqual(1, features['body-style'].dense_shape.get_shape()[0])
+
+    net = fc.linear_model(features, [price_buckets, body_style])
+    with _initialized_session() as sess:
+      bias = get_linear_model_bias()
+      price_buckets_var = get_linear_model_column_var(price_buckets)
+      body_style_var = get_linear_model_column_var(body_style)
+
+      sess.run(price_buckets_var.assign([[10.], [100.], [1000.], [10000.]]))
+      sess.run(body_style_var.assign([[-10.], [-100.], [-1000.]]))
+      sess.run(bias.assign([5.]))
+
+      self.assertAllClose([[10 - 1000 + 5.], [1000 - 10 + 5.]], sess.run(net))
+
+  def test_with_1d_unknown_shape_sparse_tensor(self):
+    price = fc_old.numeric_column('price')
+    price_buckets = fc_old.bucketized_column(
+        price, boundaries=[
+            0.,
+            10.,
+            100.,
+        ])
+    body_style = fc_old.categorical_column_with_vocabulary_list(
+        'body-style', vocabulary_list=['hardtop', 'wagon', 'sedan'])
+    country = fc_old.categorical_column_with_vocabulary_list(
+        'country', vocabulary_list=['US', 'JP', 'CA'])
+
+    # Provides 1-dim tensor and dense tensor.
+    features = {
+        'price': array_ops.placeholder(dtypes.float32),
+        'body-style': array_ops.sparse_placeholder(dtypes.string),
+        'country': array_ops.placeholder(dtypes.string),
+    }
+    self.assertIsNone(features['price'].shape.ndims)
+    self.assertIsNone(features['body-style'].get_shape().ndims)
+
+    price_data = np.array([-1., 12.])
+    body_style_data = sparse_tensor.SparseTensorValue(
+        indices=((0,), (1,)),
+        values=('sedan', 'hardtop'),
+        dense_shape=(2,))
+    country_data = np.array(['US', 'CA'])
+
+    net = fc.linear_model(features, [price_buckets, body_style, country])
+    bias = get_linear_model_bias()
+    price_buckets_var = get_linear_model_column_var(price_buckets)
+    body_style_var = get_linear_model_column_var(body_style)
+    with _initialized_session() as sess:
+      sess.run(price_buckets_var.assign([[10.], [100.], [1000.], [10000.]]))
+      sess.run(body_style_var.assign([[-10.], [-100.], [-1000.]]))
+      sess.run(bias.assign([5.]))
+
+      self.assertAllClose([[10 - 1000 + 5.], [1000 - 10 + 5.]],
+                          sess.run(
+                              net,
+                              feed_dict={
+                                  features['price']: price_data,
+                                  features['body-style']: body_style_data,
+                                  features['country']: country_data
+                              }))
+
+  def test_with_rank_0_feature(self):
+    price = fc_old.numeric_column('price')
+    features = {
+        'price': constant_op.constant(0),
+    }
+    self.assertEqual(0, features['price'].shape.ndims)
+
+    # Static rank 0 should fail
+    with self.assertRaisesRegexp(ValueError, 'Feature .* cannot have rank 0'):
+      fc.linear_model(features, [price])
+
+    # Dynamic rank 0 should fail
+    features = {
+        'price': array_ops.placeholder(dtypes.float32),
+    }
+    net = fc.linear_model(features, [price])
+    self.assertEqual(1, net.shape[1])
+    with _initialized_session() as sess:
+      with self.assertRaisesOpError('Feature .* cannot have rank 0'):
+        sess.run(net, feed_dict={features['price']: np.array(1)})
+
+  def test_multiple_linear_models(self):
+    price = fc_old.numeric_column('price')
+    with ops.Graph().as_default():
+      features1 = {'price': [[1.], [5.]]}
+      features2 = {'price': [[2.], [10.]]}
+      predictions1 = fc.linear_model(features1, [price])
+      predictions2 = fc.linear_model(features2, [price])
+      bias1 = get_linear_model_bias(name='linear_model')
+      bias2 = get_linear_model_bias(name='linear_model_1')
+      price_var1 = get_linear_model_column_var(price, name='linear_model')
+      price_var2 = get_linear_model_column_var(price, name='linear_model_1')
+      with _initialized_session() as sess:
+        self.assertAllClose([0.], bias1.eval())
+        sess.run(price_var1.assign([[10.]]))
+        sess.run(bias1.assign([5.]))
+        self.assertAllClose([[15.], [55.]], predictions1.eval())
+        self.assertAllClose([0.], bias2.eval())
+        sess.run(price_var2.assign([[10.]]))
+        sess.run(bias2.assign([5.]))
+        self.assertAllClose([[25.], [105.]], predictions2.eval())
+
+
+class _LinearModelTest(test.TestCase):
+
+  def test_raises_if_empty_feature_columns(self):
+    with self.assertRaisesRegexp(ValueError,
+                                 'feature_columns must not be empty'):
+      get_keras_linear_model_predictions(features={}, feature_columns=[])
+
+  def test_should_be_feature_column(self):
+    with self.assertRaisesRegexp(ValueError, 'must be a _FeatureColumn'):
+      get_keras_linear_model_predictions(
+          features={'a': [[0]]}, feature_columns='NotSupported')
+
+  def test_should_be_dense_or_categorical_column(self):
+
+    class NotSupportedColumn(fc_old._FeatureColumn):
+
+      @property
+      def name(self):
+        return 'NotSupportedColumn'
+
+      def _transform_feature(self, cache):
+        pass
+
+      @property
+      def _parse_example_spec(self):
+        pass
+
+    with self.assertRaisesRegexp(
+        ValueError, 'must be either a _DenseColumn or _CategoricalColumn'):
+      get_keras_linear_model_predictions(
+          features={'a': [[0]]}, feature_columns=[NotSupportedColumn()])
+
+  def test_does_not_support_dict_columns(self):
+    with self.assertRaisesRegexp(
+        ValueError, 'Expected feature_columns to be iterable, found dict.'):
+      fc.linear_model(
+          features={'a': [[0]]},
+          feature_columns={'a': fc_old.numeric_column('a')})
+
+  def test_raises_if_duplicate_name(self):
+    with self.assertRaisesRegexp(
+        ValueError, 'Duplicate feature column name found for columns'):
+      get_keras_linear_model_predictions(
+          features={'a': [[0]]},
+          feature_columns=[
+              fc_old.numeric_column('a'),
+              fc_old.numeric_column('a')
+          ])
+
+  def test_dense_bias(self):
+    price = fc_old.numeric_column('price')
+    with ops.Graph().as_default():
+      features = {'price': [[1.], [5.]]}
+      predictions = get_keras_linear_model_predictions(features, [price])
+      bias = get_linear_model_bias()
+      price_var = get_linear_model_column_var(price)
+      with _initialized_session() as sess:
+        self.assertAllClose([0.], bias.eval())
+        sess.run(price_var.assign([[10.]]))
+        sess.run(bias.assign([5.]))
+        self.assertAllClose([[15.], [55.]], predictions.eval())
+
+  def test_sparse_bias(self):
+    wire_cast = fc_old.categorical_column_with_hash_bucket('wire_cast', 4)
+    with ops.Graph().as_default():
+      wire_tensor = sparse_tensor.SparseTensor(
+          values=['omar', 'stringer', 'marlo'],  # hashed to = [2, 0, 3]
+          indices=[[0, 0], [1, 0], [1, 1]],
+          dense_shape=[2, 2])
+      features = {'wire_cast': wire_tensor}
+      predictions = get_keras_linear_model_predictions(features, [wire_cast])
+      bias = get_linear_model_bias()
+      wire_cast_var = get_linear_model_column_var(wire_cast)
+      with _initialized_session() as sess:
+        self.assertAllClose([0.], bias.eval())
+        self.assertAllClose([[0.], [0.], [0.], [0.]], wire_cast_var.eval())
+        sess.run(wire_cast_var.assign([[10.], [100.], [1000.], [10000.]]))
+        sess.run(bias.assign([5.]))
+        self.assertAllClose([[1005.], [10015.]], predictions.eval())
+
+  def test_dense_and_sparse_bias(self):
+    wire_cast = fc_old.categorical_column_with_hash_bucket('wire_cast', 4)
+    price = fc_old.numeric_column('price')
+    with ops.Graph().as_default():
+      wire_tensor = sparse_tensor.SparseTensor(
+          values=['omar', 'stringer', 'marlo'],  # hashed to = [2, 0, 3]
+          indices=[[0, 0], [1, 0], [1, 1]],
+          dense_shape=[2, 2])
+      features = {'wire_cast': wire_tensor, 'price': [[1.], [5.]]}
+      predictions = get_keras_linear_model_predictions(features,
+                                                       [wire_cast, price])
+      bias = get_linear_model_bias()
+      wire_cast_var = get_linear_model_column_var(wire_cast)
+      price_var = get_linear_model_column_var(price)
+      with _initialized_session() as sess:
+        sess.run(wire_cast_var.assign([[10.], [100.], [1000.], [10000.]]))
+        sess.run(bias.assign([5.]))
+        sess.run(price_var.assign([[10.]]))
+        self.assertAllClose([[1015.], [10065.]], predictions.eval())
+
+  def test_dense_and_sparse_column(self):
+    """When the column is both dense and sparse, uses sparse tensors."""
+
+    class _DenseAndSparseColumn(fc_old._DenseColumn, fc_old._CategoricalColumn):
+
+      @property
+      def name(self):
+        return 'dense_and_sparse_column'
+
+      @property
+      def _parse_example_spec(self):
+        return {self.name: parsing_ops.VarLenFeature(self.dtype)}
+
+      def _transform_feature(self, inputs):
+        return inputs.get(self.name)
+
+      @property
+      def _variable_shape(self):
+        raise ValueError('Should not use this method.')
+
+      def _get_dense_tensor(self,
+                            inputs,
+                            weight_collections=None,
+                            trainable=None):
+        raise ValueError('Should not use this method.')
+
+      @property
+      def _num_buckets(self):
+        return 4
+
+      def _get_sparse_tensors(self,
+                              inputs,
+                              weight_collections=None,
+                              trainable=None):
+        sp_tensor = sparse_tensor.SparseTensor(
+            indices=[[0, 0], [1, 0], [1, 1]],
+            values=[2, 0, 3],
+            dense_shape=[2, 2])
+        return fc_old._CategoricalColumn.IdWeightPair(sp_tensor, None)
+
+    dense_and_sparse_column = _DenseAndSparseColumn()
+    with ops.Graph().as_default():
+      sp_tensor = sparse_tensor.SparseTensor(
+          values=['omar', 'stringer', 'marlo'],
+          indices=[[0, 0], [1, 0], [1, 1]],
+          dense_shape=[2, 2])
+      features = {dense_and_sparse_column.name: sp_tensor}
+      predictions = get_keras_linear_model_predictions(
+          features, [dense_and_sparse_column])
+      bias = get_linear_model_bias()
+      dense_and_sparse_column_var = get_linear_model_column_var(
+          dense_and_sparse_column)
+      with _initialized_session() as sess:
+        sess.run(
+            dense_and_sparse_column_var.assign([[10.], [100.], [1000.],
+                                                [10000.]]))
+        sess.run(bias.assign([5.]))
+        self.assertAllClose([[1005.], [10015.]], predictions.eval())
+
+  def test_dense_multi_output(self):
+    price = fc_old.numeric_column('price')
+    with ops.Graph().as_default():
+      features = {'price': [[1.], [5.]]}
+      predictions = get_keras_linear_model_predictions(
+          features, [price], units=3)
+      bias = get_linear_model_bias()
+      price_var = get_linear_model_column_var(price)
+      with _initialized_session() as sess:
+        self.assertAllClose(np.zeros((3,)), bias.eval())
+        self.assertAllClose(np.zeros((1, 3)), price_var.eval())
+        sess.run(price_var.assign([[10., 100., 1000.]]))
+        sess.run(bias.assign([5., 6., 7.]))
+        self.assertAllClose([[15., 106., 1007.], [55., 506., 5007.]],
+                            predictions.eval())
+
+  def test_sparse_multi_output(self):
+    wire_cast = fc_old.categorical_column_with_hash_bucket('wire_cast', 4)
+    with ops.Graph().as_default():
+      wire_tensor = sparse_tensor.SparseTensor(
+          values=['omar', 'stringer', 'marlo'],  # hashed to = [2, 0, 3]
+          indices=[[0, 0], [1, 0], [1, 1]],
+          dense_shape=[2, 2])
+      features = {'wire_cast': wire_tensor}
+      predictions = get_keras_linear_model_predictions(
+          features, [wire_cast], units=3)
+      bias = get_linear_model_bias()
+      wire_cast_var = get_linear_model_column_var(wire_cast)
+      with _initialized_session() as sess:
+        self.assertAllClose(np.zeros((3,)), bias.eval())
+        self.assertAllClose(np.zeros((4, 3)), wire_cast_var.eval())
+        sess.run(
+            wire_cast_var.assign([[10., 11., 12.], [100., 110., 120.],
+                                  [1000., 1100.,
+                                   1200.], [10000., 11000., 12000.]]))
+        sess.run(bias.assign([5., 6., 7.]))
+        self.assertAllClose([[1005., 1106., 1207.], [10015., 11017., 12019.]],
+                            predictions.eval())
+
+  def test_dense_multi_dimension(self):
+    price = fc_old.numeric_column('price', shape=2)
+    with ops.Graph().as_default():
+      features = {'price': [[1., 2.], [5., 6.]]}
+      predictions = get_keras_linear_model_predictions(features, [price])
+      price_var = get_linear_model_column_var(price)
+      with _initialized_session() as sess:
+        self.assertAllClose([[0.], [0.]], price_var.eval())
+        sess.run(price_var.assign([[10.], [100.]]))
+        self.assertAllClose([[210.], [650.]], predictions.eval())
+
+  def test_sparse_multi_rank(self):
+    wire_cast = fc_old.categorical_column_with_hash_bucket('wire_cast', 4)
+    with ops.Graph().as_default():
+      wire_tensor = array_ops.sparse_placeholder(dtypes.string)
+      wire_value = sparse_tensor.SparseTensorValue(
+          values=['omar', 'stringer', 'marlo', 'omar'],  # hashed = [2, 0, 3, 2]
+          indices=[[0, 0, 0], [0, 1, 0], [1, 0, 0], [1, 0, 1]],
+          dense_shape=[2, 2, 2])
+      features = {'wire_cast': wire_tensor}
+      predictions = get_keras_linear_model_predictions(features, [wire_cast])
+      wire_cast_var = get_linear_model_column_var(wire_cast)
+      with _initialized_session() as sess:
+        self.assertAllClose(np.zeros((4, 1)), wire_cast_var.eval())
+        self.assertAllClose(
+            np.zeros((2, 1)),
+            predictions.eval(feed_dict={wire_tensor: wire_value}))
+        sess.run(wire_cast_var.assign([[10.], [100.], [1000.], [10000.]]))
+        self.assertAllClose(
+            [[1010.], [11000.]],
+            predictions.eval(feed_dict={wire_tensor: wire_value}))
+
+  def test_sparse_combiner(self):
+    wire_cast = fc_old.categorical_column_with_hash_bucket('wire_cast', 4)
+    with ops.Graph().as_default():
+      wire_tensor = sparse_tensor.SparseTensor(
+          values=['omar', 'stringer', 'marlo'],  # hashed to = [2, 0, 3]
+          indices=[[0, 0], [1, 0], [1, 1]],
+          dense_shape=[2, 2])
+      features = {'wire_cast': wire_tensor}
+      predictions = get_keras_linear_model_predictions(
+          features, [wire_cast], sparse_combiner='mean')
+      bias = get_linear_model_bias()
+      wire_cast_var = get_linear_model_column_var(wire_cast)
+      with _initialized_session() as sess:
+        sess.run(wire_cast_var.assign([[10.], [100.], [1000.], [10000.]]))
+        sess.run(bias.assign([5.]))
+        self.assertAllClose([[1005.], [5010.]], predictions.eval())
+
+  def test_dense_multi_dimension_multi_output(self):
+    price = fc_old.numeric_column('price', shape=2)
+    with ops.Graph().as_default():
+      features = {'price': [[1., 2.], [5., 6.]]}
+      predictions = get_keras_linear_model_predictions(
+          features, [price], units=3)
+      bias = get_linear_model_bias()
+      price_var = get_linear_model_column_var(price)
+      with _initialized_session() as sess:
+        self.assertAllClose(np.zeros((3,)), bias.eval())
+        self.assertAllClose(np.zeros((2, 3)), price_var.eval())
+        sess.run(price_var.assign([[1., 2., 3.], [10., 100., 1000.]]))
+        sess.run(bias.assign([2., 3., 4.]))
+        self.assertAllClose([[23., 205., 2007.], [67., 613., 6019.]],
+                            predictions.eval())
+
+  def test_raises_if_shape_mismatch(self):
+    price = fc_old.numeric_column('price', shape=2)
+    with ops.Graph().as_default():
+      features = {'price': [[1.], [5.]]}
+      with self.assertRaisesRegexp(
+          Exception,
+          r'Cannot reshape a tensor with 2 elements to shape \[2,2\]'):
+        get_keras_linear_model_predictions(features, [price])
+
+  def test_dense_reshaping(self):
+    price = fc_old.numeric_column('price', shape=[1, 2])
+    with ops.Graph().as_default():
+      features = {'price': [[[1., 2.]], [[5., 6.]]]}
+      predictions = get_keras_linear_model_predictions(features, [price])
+      bias = get_linear_model_bias()
+      price_var = get_linear_model_column_var(price)
+      with _initialized_session() as sess:
+        self.assertAllClose([0.], bias.eval())
+        self.assertAllClose([[0.], [0.]], price_var.eval())
+        self.assertAllClose([[0.], [0.]], predictions.eval())
+        sess.run(price_var.assign([[10.], [100.]]))
+        self.assertAllClose([[210.], [650.]], predictions.eval())
+
+  def test_dense_multi_column(self):
+    price1 = fc_old.numeric_column('price1', shape=2)
+    price2 = fc_old.numeric_column('price2')
+    with ops.Graph().as_default():
+      features = {'price1': [[1., 2.], [5., 6.]], 'price2': [[3.], [4.]]}
+      predictions = get_keras_linear_model_predictions(features,
+                                                       [price1, price2])
+      bias = get_linear_model_bias()
+      price1_var = get_linear_model_column_var(price1)
+      price2_var = get_linear_model_column_var(price2)
+      with _initialized_session() as sess:
+        self.assertAllClose([0.], bias.eval())
+        self.assertAllClose([[0.], [0.]], price1_var.eval())
+        self.assertAllClose([[0.]], price2_var.eval())
+        self.assertAllClose([[0.], [0.]], predictions.eval())
+        sess.run(price1_var.assign([[10.], [100.]]))
+        sess.run(price2_var.assign([[1000.]]))
+        sess.run(bias.assign([7.]))
+        self.assertAllClose([[3217.], [4657.]], predictions.eval())
+
+  def test_fills_cols_to_vars(self):
+    price1 = fc_old.numeric_column('price1', shape=2)
+    price2 = fc_old.numeric_column('price2')
+    with ops.Graph().as_default():
+      features = {'price1': [[1., 2.], [5., 6.]], 'price2': [[3.], [4.]]}
+      cols_to_vars = {}
+      get_keras_linear_model_predictions(
+          features, [price1, price2], cols_to_vars=cols_to_vars)
+      bias = get_linear_model_bias()
+      price1_var = get_linear_model_column_var(price1)
+      price2_var = get_linear_model_column_var(price2)
+      self.assertAllEqual(cols_to_vars['bias'], [bias])
+      self.assertAllEqual(cols_to_vars[price1], [price1_var])
+      self.assertAllEqual(cols_to_vars[price2], [price2_var])
+
+  def test_fills_cols_to_vars_partitioned_variables(self):
+    price1 = fc_old.numeric_column('price1', shape=2)
+    price2 = fc_old.numeric_column('price2', shape=3)
+    with ops.Graph().as_default():
+      features = {
+          'price1': [[1., 2.], [6., 7.]],
+          'price2': [[3., 4., 5.], [8., 9., 10.]]
+      }
+      cols_to_vars = {}
+      with variable_scope.variable_scope(
+          'linear',
+          partitioner=partitioned_variables.fixed_size_partitioner(2, axis=0)):
+        get_keras_linear_model_predictions(
+            features, [price1, price2], cols_to_vars=cols_to_vars)
+      with _initialized_session():
+        self.assertEqual([0.], cols_to_vars['bias'][0].eval())
+        # Partitioning shards the [2, 1] price1 var into 2 [1, 1] Variables.
+        self.assertAllEqual([[0.]], cols_to_vars[price1][0].eval())
+        self.assertAllEqual([[0.]], cols_to_vars[price1][1].eval())
+        # Partitioning shards the [3, 1] price2 var into a [2, 1] Variable and
+        # a [1, 1] Variable.
+        self.assertAllEqual([[0.], [0.]], cols_to_vars[price2][0].eval())
+        self.assertAllEqual([[0.]], cols_to_vars[price2][1].eval())
+
+  def test_dense_collection(self):
+    price = fc_old.numeric_column('price')
+    with ops.Graph().as_default() as g:
+      features = {'price': [[1.], [5.]]}
+      get_keras_linear_model_predictions(
+          features, [price], weight_collections=['my-vars'])
+      my_vars = g.get_collection('my-vars')
+      bias = get_linear_model_bias()
+      price_var = get_linear_model_column_var(price)
+      self.assertIn(bias, my_vars)
+      self.assertIn(price_var, my_vars)
+
+  def test_sparse_collection(self):
+    wire_cast = fc_old.categorical_column_with_hash_bucket('wire_cast', 4)
+    with ops.Graph().as_default() as g:
+      wire_tensor = sparse_tensor.SparseTensor(
+          values=['omar'], indices=[[0, 0]], dense_shape=[1, 1])
+      features = {'wire_cast': wire_tensor}
+      get_keras_linear_model_predictions(
+          features, [wire_cast], weight_collections=['my-vars'])
+      my_vars = g.get_collection('my-vars')
+      bias = get_linear_model_bias()
+      wire_cast_var = get_linear_model_column_var(wire_cast)
+      self.assertIn(bias, my_vars)
+      self.assertIn(wire_cast_var, my_vars)
+
+  def test_dense_trainable_default(self):
+    price = fc_old.numeric_column('price')
+    with ops.Graph().as_default() as g:
+      features = {'price': [[1.], [5.]]}
+      get_keras_linear_model_predictions(features, [price])
+      bias = get_linear_model_bias()
+      price_var = get_linear_model_column_var(price)
+      trainable_vars = g.get_collection(ops.GraphKeys.TRAINABLE_VARIABLES)
+      self.assertIn(bias, trainable_vars)
+      self.assertIn(price_var, trainable_vars)
+
+  def test_sparse_trainable_default(self):
+    wire_cast = fc_old.categorical_column_with_hash_bucket('wire_cast', 4)
+    with ops.Graph().as_default() as g:
+      wire_tensor = sparse_tensor.SparseTensor(
+          values=['omar'], indices=[[0, 0]], dense_shape=[1, 1])
+      features = {'wire_cast': wire_tensor}
+      get_keras_linear_model_predictions(features, [wire_cast])
+      trainable_vars = g.get_collection(ops.GraphKeys.TRAINABLE_VARIABLES)
+      bias = get_linear_model_bias()
+      wire_cast_var = get_linear_model_column_var(wire_cast)
+      self.assertIn(bias, trainable_vars)
+      self.assertIn(wire_cast_var, trainable_vars)
+
+  def test_dense_trainable_false(self):
+    price = fc_old.numeric_column('price')
+    with ops.Graph().as_default() as g:
+      features = {'price': [[1.], [5.]]}
+      get_keras_linear_model_predictions(features, [price], trainable=False)
+      trainable_vars = g.get_collection(ops.GraphKeys.TRAINABLE_VARIABLES)
+      self.assertEqual([], trainable_vars)
+
+  def test_sparse_trainable_false(self):
+    wire_cast = fc_old.categorical_column_with_hash_bucket('wire_cast', 4)
+    with ops.Graph().as_default() as g:
+      wire_tensor = sparse_tensor.SparseTensor(
+          values=['omar'], indices=[[0, 0]], dense_shape=[1, 1])
+      features = {'wire_cast': wire_tensor}
+      get_keras_linear_model_predictions(features, [wire_cast], trainable=False)
+      trainable_vars = g.get_collection(ops.GraphKeys.TRAINABLE_VARIABLES)
+      self.assertEqual([], trainable_vars)
+
+  def test_column_order(self):
+    price_a = fc_old.numeric_column('price_a')
+    price_b = fc_old.numeric_column('price_b')
+    wire_cast = fc_old.categorical_column_with_hash_bucket('wire_cast', 4)
+    with ops.Graph().as_default() as g:
+      features = {
+          'price_a': [[1.]],
+          'price_b': [[3.]],
+          'wire_cast':
+              sparse_tensor.SparseTensor(
+                  values=['omar'], indices=[[0, 0]], dense_shape=[1, 1])
+      }
+      get_keras_linear_model_predictions(
+          features, [price_a, wire_cast, price_b],
+          weight_collections=['my-vars'])
+      my_vars = g.get_collection('my-vars')
+      self.assertIn('price_a', my_vars[0].name)
+      self.assertIn('price_b', my_vars[1].name)
+      self.assertIn('wire_cast', my_vars[2].name)
+
+    with ops.Graph().as_default() as g:
+      features = {
+          'price_a': [[1.]],
+          'price_b': [[3.]],
+          'wire_cast':
+              sparse_tensor.SparseTensor(
+                  values=['omar'], indices=[[0, 0]], dense_shape=[1, 1])
+      }
+      get_keras_linear_model_predictions(
+          features, [wire_cast, price_b, price_a],
+          weight_collections=['my-vars'])
+      my_vars = g.get_collection('my-vars')
+      self.assertIn('price_a', my_vars[0].name)
+      self.assertIn('price_b', my_vars[1].name)
+      self.assertIn('wire_cast', my_vars[2].name)
+
+  def test_static_batch_size_mismatch(self):
+    price1 = fc_old.numeric_column('price1')
+    price2 = fc_old.numeric_column('price2')
+    with ops.Graph().as_default():
+      features = {
+          'price1': [[1.], [5.], [7.]],  # batchsize = 3
+          'price2': [[3.], [4.]]  # batchsize = 2
+      }
+    with self.assertRaisesRegexp(
+        ValueError,
+        'Batch size \(first dimension\) of each feature must be same.'):  # pylint: disable=anomalous-backslash-in-string
+      get_keras_linear_model_predictions(features, [price1, price2])
+
+  def test_subset_of_static_batch_size_mismatch(self):
+    price1 = fc_old.numeric_column('price1')
+    price2 = fc_old.numeric_column('price2')
+    price3 = fc_old.numeric_column('price3')
+    with ops.Graph().as_default():
+      features = {
+          'price1': array_ops.placeholder(dtype=dtypes.int64),  # batchsize = 3
+          'price2': [[3.], [4.]],  # batchsize = 2
+          'price3': [[3.], [4.], [5.]]  # batchsize = 3
+      }
+      with self.assertRaisesRegexp(
+          ValueError,
+          'Batch size \(first dimension\) of each feature must be same.'):  # pylint: disable=anomalous-backslash-in-string
+        get_keras_linear_model_predictions(features, [price1, price2, price3])
+
+  def test_runtime_batch_size_mismatch(self):
+    price1 = fc_old.numeric_column('price1')
+    price2 = fc_old.numeric_column('price2')
+    with ops.Graph().as_default():
+      features = {
+          'price1': array_ops.placeholder(dtype=dtypes.int64),  # batchsize = 3
+          'price2': [[3.], [4.]]  # batchsize = 2
+      }
+      predictions = get_keras_linear_model_predictions(features,
+                                                       [price1, price2])
+      with _initialized_session() as sess:
+        with self.assertRaisesRegexp(errors.OpError,
+                                     'must have the same size and shape'):
+          sess.run(
+              predictions, feed_dict={features['price1']: [[1.], [5.], [7.]]})
+
+  def test_runtime_batch_size_matches(self):
+    price1 = fc_old.numeric_column('price1')
+    price2 = fc_old.numeric_column('price2')
+    with ops.Graph().as_default():
+      features = {
+          'price1': array_ops.placeholder(dtype=dtypes.int64),  # batchsize = 2
+          'price2': array_ops.placeholder(dtype=dtypes.int64),  # batchsize = 2
+      }
+      predictions = get_keras_linear_model_predictions(features,
+                                                       [price1, price2])
+      with _initialized_session() as sess:
+        sess.run(
+            predictions,
+            feed_dict={
+                features['price1']: [[1.], [5.]],
+                features['price2']: [[1.], [5.]],
+            })
+
+  def test_with_numpy_input_fn(self):
+    price = fc_old.numeric_column('price')
+    price_buckets = fc_old.bucketized_column(
+        price, boundaries=[
+            0.,
+            10.,
+            100.,
+        ])
+    body_style = fc_old.categorical_column_with_vocabulary_list(
+        'body-style', vocabulary_list=['hardtop', 'wagon', 'sedan'])
+
+    input_fn = numpy_io.numpy_input_fn(
+        x={
+            'price': np.array([-1., 2., 13., 104.]),
+            'body-style': np.array(['sedan', 'hardtop', 'wagon', 'sedan']),
+        },
+        batch_size=2,
+        shuffle=False)
+    features = input_fn()
+    net = get_keras_linear_model_predictions(features,
+                                             [price_buckets, body_style])
+    # self.assertEqual(1 + 3 + 5, net.shape[1])
+    with _initialized_session() as sess:
+      coord = coordinator.Coordinator()
+      threads = queue_runner_impl.start_queue_runners(sess, coord=coord)
+
+      bias = get_linear_model_bias()
+      price_buckets_var = get_linear_model_column_var(price_buckets)
+      body_style_var = get_linear_model_column_var(body_style)
+
+      sess.run(price_buckets_var.assign([[10.], [100.], [1000.], [10000.]]))
+      sess.run(body_style_var.assign([[-10.], [-100.], [-1000.]]))
+      sess.run(bias.assign([5.]))
+
+      self.assertAllClose([[10 - 1000 + 5.], [100 - 10 + 5.]], sess.run(net))
+
+      coord.request_stop()
+      coord.join(threads)
+
+  def test_with_1d_sparse_tensor(self):
+    price = fc_old.numeric_column('price')
+    price_buckets = fc_old.bucketized_column(
+        price, boundaries=[
+            0.,
+            10.,
+            100.,
+        ])
+    body_style = fc_old.categorical_column_with_vocabulary_list(
+        'body-style', vocabulary_list=['hardtop', 'wagon', 'sedan'])
+
+    # Provides 1-dim tensor and dense tensor.
+    features = {
+        'price':
+            constant_op.constant([
+                -1.,
+                12.,
+            ]),
+        'body-style':
+            sparse_tensor.SparseTensor(
+                indices=((0,), (1,)),
+                values=('sedan', 'hardtop'),
+                dense_shape=(2,)),
+    }
+    self.assertEqual(1, features['price'].shape.ndims)
+    self.assertEqual(1, features['body-style'].dense_shape.get_shape()[0])
+
+    net = get_keras_linear_model_predictions(features,
+                                             [price_buckets, body_style])
+    with _initialized_session() as sess:
+      bias = get_linear_model_bias()
+      price_buckets_var = get_linear_model_column_var(price_buckets)
+      body_style_var = get_linear_model_column_var(body_style)
+
+      sess.run(price_buckets_var.assign([[10.], [100.], [1000.], [10000.]]))
+      sess.run(body_style_var.assign([[-10.], [-100.], [-1000.]]))
+      sess.run(bias.assign([5.]))
+
+      self.assertAllClose([[10 - 1000 + 5.], [1000 - 10 + 5.]], sess.run(net))
+
+  def test_with_1d_unknown_shape_sparse_tensor(self):
+    price = fc_old.numeric_column('price')
+    price_buckets = fc_old.bucketized_column(
+        price, boundaries=[
+            0.,
+            10.,
+            100.,
+        ])
+    body_style = fc_old.categorical_column_with_vocabulary_list(
+        'body-style', vocabulary_list=['hardtop', 'wagon', 'sedan'])
+    country = fc_old.categorical_column_with_vocabulary_list(
+        'country', vocabulary_list=['US', 'JP', 'CA'])
+
+    # Provides 1-dim tensor and dense tensor.
+    features = {
+        'price': array_ops.placeholder(dtypes.float32),
+        'body-style': array_ops.sparse_placeholder(dtypes.string),
+        'country': array_ops.placeholder(dtypes.string),
+    }
+    self.assertIsNone(features['price'].shape.ndims)
+    self.assertIsNone(features['body-style'].get_shape().ndims)
+
+    price_data = np.array([-1., 12.])
+    body_style_data = sparse_tensor.SparseTensorValue(
+        indices=((0,), (1,)), values=('sedan', 'hardtop'), dense_shape=(2,))
+    country_data = np.array(['US', 'CA'])
+
+    net = get_keras_linear_model_predictions(
+        features, [price_buckets, body_style, country])
+    bias = get_linear_model_bias()
+    price_buckets_var = get_linear_model_column_var(price_buckets)
+    body_style_var = get_linear_model_column_var(body_style)
+    with _initialized_session() as sess:
+      sess.run(price_buckets_var.assign([[10.], [100.], [1000.], [10000.]]))
+      sess.run(body_style_var.assign([[-10.], [-100.], [-1000.]]))
+      sess.run(bias.assign([5.]))
+
+      self.assertAllClose([[10 - 1000 + 5.], [1000 - 10 + 5.]],
+                          sess.run(
+                              net,
+                              feed_dict={
+                                  features['price']: price_data,
+                                  features['body-style']: body_style_data,
+                                  features['country']: country_data
+                              }))
+
+  def test_with_rank_0_feature(self):
+    price = fc_old.numeric_column('price')
+    features = {
+        'price': constant_op.constant(0),
+    }
+    self.assertEqual(0, features['price'].shape.ndims)
+
+    # Static rank 0 should fail
+    with self.assertRaisesRegexp(ValueError, 'Feature .* cannot have rank 0'):
+      get_keras_linear_model_predictions(features, [price])
+
+    # Dynamic rank 0 should fail
+    features = {
+        'price': array_ops.placeholder(dtypes.float32),
+    }
+    net = get_keras_linear_model_predictions(features, [price])
+    self.assertEqual(1, net.shape[1])
+    with _initialized_session() as sess:
+      with self.assertRaisesOpError('Feature .* cannot have rank 0'):
+        sess.run(net, feed_dict={features['price']: np.array(1)})
+
+
+class InputLayerTest(test.TestCase):
+
+  @test_util.run_in_graph_and_eager_modes()
+  def test_retrieving_input(self):
+    features = {'a': [0.]}
+    input_layer = InputLayer(fc_old.numeric_column('a'))
+    inputs = self.evaluate(input_layer(features))
+    self.assertAllClose([[0.]], inputs)
+
+  def test_reuses_variables(self):
+    with context.eager_mode():
+      sparse_input = sparse_tensor.SparseTensor(
+          indices=((0, 0), (1, 0), (2, 0)),
+          values=(0, 1, 2),
+          dense_shape=(3, 3))
+
+      # Create feature columns (categorical and embedding).
+      categorical_column = fc_old.categorical_column_with_identity(
+          key='a', num_buckets=3)
+      embedding_dimension = 2
+      def _embedding_column_initializer(shape, dtype, partition_info):
+        del shape  # unused
+        del dtype  # unused
+        del partition_info  # unused
+        embedding_values = (
+            (1, 0),  # id 0
+            (0, 1),  # id 1
+            (1, 1))  # id 2
+        return embedding_values
+
+      embedding_column = fc_old.embedding_column(
+          categorical_column,
+          dimension=embedding_dimension,
+          initializer=_embedding_column_initializer)
+
+      input_layer = InputLayer([embedding_column])
+      features = {'a': sparse_input}
+
+      inputs = input_layer(features)
+      variables = input_layer.variables
+
+      # Sanity check: test that the inputs are correct.
+      self.assertAllEqual([[1, 0], [0, 1], [1, 1]], inputs)
+
+      # Check that only one variable was created.
+      self.assertEqual(1, len(variables))
+
+      # Check that invoking input_layer on the same features does not create
+      # additional variables
+      _ = input_layer(features)
+      self.assertEqual(1, len(variables))
+      self.assertEqual(variables[0], input_layer.variables[0])
+
+  def test_feature_column_input_layer_gradient(self):
+    with context.eager_mode():
+      sparse_input = sparse_tensor.SparseTensor(
+          indices=((0, 0), (1, 0), (2, 0)),
+          values=(0, 1, 2),
+          dense_shape=(3, 3))
+
+      # Create feature columns (categorical and embedding).
+      categorical_column = fc_old.categorical_column_with_identity(
+          key='a', num_buckets=3)
+      embedding_dimension = 2
+
+      def _embedding_column_initializer(shape, dtype, partition_info):
+        del shape  # unused
+        del dtype  # unused
+        del partition_info  # unused
+        embedding_values = (
+            (1, 0),  # id 0
+            (0, 1),  # id 1
+            (1, 1))  # id 2
+        return embedding_values
+
+      embedding_column = fc_old.embedding_column(
+          categorical_column,
+          dimension=embedding_dimension,
+          initializer=_embedding_column_initializer)
+
+      input_layer = InputLayer([embedding_column])
+      features = {'a': sparse_input}
+
+      def scale_matrix():
+        matrix = input_layer(features)
+        return 2 * matrix
+
+      # Sanity check: Verify that scale_matrix returns the correct output.
+      self.assertAllEqual([[2, 0], [0, 2], [2, 2]], scale_matrix())
+
+      # Check that the returned gradient is correct.
+      grad_function = backprop.implicit_grad(scale_matrix)
+      grads_and_vars = grad_function()
+      indexed_slice = grads_and_vars[0][0]
+      gradient = grads_and_vars[0][0].values
+
+      self.assertAllEqual([0, 1, 2], indexed_slice.indices)
+      self.assertAllEqual([[2, 2], [2, 2], [2, 2]], gradient)
+
+
+class FunctionalInputLayerTest(test.TestCase):
+
+  def test_raises_if_empty_feature_columns(self):
+    with self.assertRaisesRegexp(ValueError,
+                                 'feature_columns must not be empty'):
+      fc.input_layer(features={}, feature_columns=[])
+
+  def test_should_be_dense_column(self):
+    with self.assertRaisesRegexp(ValueError, 'must be a _DenseColumn'):
+      fc.input_layer(
+          features={'a': [[0]]},
+          feature_columns=[
+              fc_old.categorical_column_with_hash_bucket('wire_cast', 4)
+          ])
+
+  def test_does_not_support_dict_columns(self):
+    with self.assertRaisesRegexp(
+        ValueError, 'Expected feature_columns to be iterable, found dict.'):
+      fc.input_layer(
+          features={'a': [[0]]},
+          feature_columns={'a': fc_old.numeric_column('a')})
+
+  def test_bare_column(self):
+    with ops.Graph().as_default():
+      features = features = {'a': [0.]}
+      net = fc.input_layer(features, fc_old.numeric_column('a'))
+      with _initialized_session():
+        self.assertAllClose([[0.]], net.eval())
+
+  def test_column_generator(self):
+    with ops.Graph().as_default():
+      features = features = {'a': [0.], 'b': [1.]}
+      columns = (fc_old.numeric_column(key) for key in features)
+      net = fc.input_layer(features, columns)
+      with _initialized_session():
+        self.assertAllClose([[0., 1.]], net.eval())
+
+  def test_raises_if_duplicate_name(self):
+    with self.assertRaisesRegexp(
+        ValueError, 'Duplicate feature column name found for columns'):
+      fc.input_layer(
+          features={'a': [[0]]},
+          feature_columns=[
+              fc_old.numeric_column('a'),
+              fc_old.numeric_column('a')
+          ])
+
+  def test_one_column(self):
+    price = fc_old.numeric_column('price')
+    with ops.Graph().as_default():
+      features = {'price': [[1.], [5.]]}
+      net = fc.input_layer(features, [price])
+      with _initialized_session():
+        self.assertAllClose([[1.], [5.]], net.eval())
+
+  def test_multi_dimension(self):
+    price = fc_old.numeric_column('price', shape=2)
+    with ops.Graph().as_default():
+      features = {'price': [[1., 2.], [5., 6.]]}
+      net = fc.input_layer(features, [price])
+      with _initialized_session():
+        self.assertAllClose([[1., 2.], [5., 6.]], net.eval())
+
+  def test_raises_if_shape_mismatch(self):
+    price = fc_old.numeric_column('price', shape=2)
+    with ops.Graph().as_default():
+      features = {'price': [[1.], [5.]]}
+      with self.assertRaisesRegexp(
+          Exception,
+          r'Cannot reshape a tensor with 2 elements to shape \[2,2\]'):
+        fc.input_layer(features, [price])
+
+  def test_reshaping(self):
+    price = fc_old.numeric_column('price', shape=[1, 2])
+    with ops.Graph().as_default():
+      features = {'price': [[[1., 2.]], [[5., 6.]]]}
+      net = fc.input_layer(features, [price])
+      with _initialized_session():
+        self.assertAllClose([[1., 2.], [5., 6.]], net.eval())
+
+  def test_multi_column(self):
+    price1 = fc_old.numeric_column('price1', shape=2)
+    price2 = fc_old.numeric_column('price2')
+    with ops.Graph().as_default():
+      features = {
+          'price1': [[1., 2.], [5., 6.]],
+          'price2': [[3.], [4.]]
+      }
+      net = fc.input_layer(features, [price1, price2])
+      with _initialized_session():
+        self.assertAllClose([[1., 2., 3.], [5., 6., 4.]], net.eval())
+
+  def test_fills_cols_to_vars(self):
+    # Provide three _DenseColumn's to input_layer: a _NumericColumn, a
+    # _BucketizedColumn, and an _EmbeddingColumn.  Only the _EmbeddingColumn
+    # creates a Variable.
+    price1 = fc_old.numeric_column('price1')
+    dense_feature = fc_old.numeric_column('dense_feature')
+    dense_feature_bucketized = fc_old.bucketized_column(
+        dense_feature, boundaries=[0.])
+    some_sparse_column = fc_old.categorical_column_with_hash_bucket(
+        'sparse_feature', hash_bucket_size=5)
+    some_embedding_column = fc_old.embedding_column(
+        some_sparse_column, dimension=10)
+    with ops.Graph().as_default():
+      features = {
+          'price1': [[3.], [4.]],
+          'dense_feature': [[-1.], [4.]],
+          'sparse_feature': [['a'], ['x']],
+      }
+      cols_to_vars = {}
+      all_cols = [price1, dense_feature_bucketized, some_embedding_column]
+      fc.input_layer(features, all_cols, cols_to_vars=cols_to_vars)
+      self.assertItemsEqual(list(cols_to_vars.keys()), all_cols)
+      self.assertEqual(0, len(cols_to_vars[price1]))
+      self.assertEqual(0, len(cols_to_vars[dense_feature_bucketized]))
+      self.assertEqual(1, len(cols_to_vars[some_embedding_column]))
+      self.assertIsInstance(cols_to_vars[some_embedding_column][0],
+                            variables_lib.Variable)
+      self.assertAllEqual(cols_to_vars[some_embedding_column][0].shape, [5, 10])
+
+  def test_fills_cols_to_vars_partitioned_variables(self):
+    price1 = fc_old.numeric_column('price1')
+    dense_feature = fc_old.numeric_column('dense_feature')
+    dense_feature_bucketized = fc_old.bucketized_column(
+        dense_feature, boundaries=[0.])
+    some_sparse_column = fc_old.categorical_column_with_hash_bucket(
+        'sparse_feature', hash_bucket_size=5)
+    some_embedding_column = fc_old.embedding_column(
+        some_sparse_column, dimension=10)
+    with ops.Graph().as_default():
+      features = {
+          'price1': [[3.], [4.]],
+          'dense_feature': [[-1.], [4.]],
+          'sparse_feature': [['a'], ['x']],
+      }
+      cols_to_vars = {}
+      all_cols = [price1, dense_feature_bucketized, some_embedding_column]
+      with variable_scope.variable_scope(
+          'input_from_feature_columns',
+          partitioner=partitioned_variables.fixed_size_partitioner(3, axis=0)):
+        fc.input_layer(features, all_cols, cols_to_vars=cols_to_vars)
+      self.assertItemsEqual(list(cols_to_vars.keys()), all_cols)
+      self.assertEqual(0, len(cols_to_vars[price1]))
+      self.assertEqual(0, len(cols_to_vars[dense_feature_bucketized]))
+      self.assertEqual(3, len(cols_to_vars[some_embedding_column]))
+      self.assertAllEqual(cols_to_vars[some_embedding_column][0].shape, [2, 10])
+      self.assertAllEqual(cols_to_vars[some_embedding_column][1].shape, [2, 10])
+      self.assertAllEqual(cols_to_vars[some_embedding_column][2].shape, [1, 10])
+
+  def test_column_order(self):
+    price_a = fc_old.numeric_column('price_a')
+    price_b = fc_old.numeric_column('price_b')
+    with ops.Graph().as_default():
+      features = {
+          'price_a': [[1.]],
+          'price_b': [[3.]],
+      }
+      net1 = fc.input_layer(features, [price_a, price_b])
+      net2 = fc.input_layer(features, [price_b, price_a])
+      with _initialized_session():
+        self.assertAllClose([[1., 3.]], net1.eval())
+        self.assertAllClose([[1., 3.]], net2.eval())
+
+  def test_fails_for_categorical_column(self):
+    animal = fc_old.categorical_column_with_identity('animal', num_buckets=4)
+    with ops.Graph().as_default():
+      features = {
+          'animal':
+              sparse_tensor.SparseTensor(
+                  indices=[[0, 0], [0, 1]], values=[1, 2], dense_shape=[1, 2])
+      }
+      with self.assertRaisesRegexp(Exception, 'must be a _DenseColumn'):
+        fc.input_layer(features, [animal])
+
+  def test_static_batch_size_mismatch(self):
+    price1 = fc_old.numeric_column('price1')
+    price2 = fc_old.numeric_column('price2')
+    with ops.Graph().as_default():
+      features = {
+          'price1': [[1.], [5.], [7.]],  # batchsize = 3
+          'price2': [[3.], [4.]]  # batchsize = 2
+      }
+      with self.assertRaisesRegexp(
+          ValueError,
+          'Batch size \(first dimension\) of each feature must be same.'):  # pylint: disable=anomalous-backslash-in-string
+        fc.input_layer(features, [price1, price2])
+
+  def test_subset_of_static_batch_size_mismatch(self):
+    price1 = fc_old.numeric_column('price1')
+    price2 = fc_old.numeric_column('price2')
+    price3 = fc_old.numeric_column('price3')
+    with ops.Graph().as_default():
+      features = {
+          'price1': array_ops.placeholder(dtype=dtypes.int64),  # batchsize = 3
+          'price2': [[3.], [4.]],  # batchsize = 2
+          'price3': [[3.], [4.], [5.]]  # batchsize = 3
+      }
+      with self.assertRaisesRegexp(
+          ValueError,
+          'Batch size \(first dimension\) of each feature must be same.'):  # pylint: disable=anomalous-backslash-in-string
+        fc.input_layer(features, [price1, price2, price3])
+
+  def test_runtime_batch_size_mismatch(self):
+    price1 = fc_old.numeric_column('price1')
+    price2 = fc_old.numeric_column('price2')
+    with ops.Graph().as_default():
+      features = {
+          'price1': array_ops.placeholder(dtype=dtypes.int64),  # batchsize = 3
+          'price2': [[3.], [4.]]  # batchsize = 2
+      }
+      net = fc.input_layer(features, [price1, price2])
+      with _initialized_session() as sess:
+        with self.assertRaisesRegexp(errors.OpError,
+                                     'Dimensions of inputs should match'):
+          sess.run(net, feed_dict={features['price1']: [[1.], [5.], [7.]]})
+
+  def test_runtime_batch_size_matches(self):
+    price1 = fc_old.numeric_column('price1')
+    price2 = fc_old.numeric_column('price2')
+    with ops.Graph().as_default():
+      features = {
+          'price1': array_ops.placeholder(dtype=dtypes.int64),  # batchsize = 2
+          'price2': array_ops.placeholder(dtype=dtypes.int64),  # batchsize = 2
+      }
+      net = fc.input_layer(features, [price1, price2])
+      with _initialized_session() as sess:
+        sess.run(
+            net,
+            feed_dict={
+                features['price1']: [[1.], [5.]],
+                features['price2']: [[1.], [5.]],
+            })
+
+  def test_multiple_layers_with_same_embedding_column(self):
+    some_sparse_column = fc_old.categorical_column_with_hash_bucket(
+        'sparse_feature', hash_bucket_size=5)
+    some_embedding_column = fc_old.embedding_column(
+        some_sparse_column, dimension=10)
+
+    with ops.Graph().as_default():
+      features = {
+          'sparse_feature': [['a'], ['x']],
+      }
+      all_cols = [some_embedding_column]
+      fc.input_layer(features, all_cols)
+      fc.input_layer(features, all_cols)
+      # Make sure that 2 variables get created in this case.
+      self.assertEqual(2, len(
+          ops.get_collection(ops.GraphKeys.GLOBAL_VARIABLES)))
+      expected_var_names = [
+          'input_layer/sparse_feature_embedding/embedding_weights:0',
+          'input_layer_1/sparse_feature_embedding/embedding_weights:0'
+      ]
+      self.assertItemsEqual(
+          expected_var_names,
+          [v.name for v in ops.get_collection(ops.GraphKeys.GLOBAL_VARIABLES)])
+
+  def test_multiple_layers_with_same_shared_embedding_column(self):
+    categorical_column_a = fc_old.categorical_column_with_identity(
+        key='aaa', num_buckets=3)
+    categorical_column_b = fc_old.categorical_column_with_identity(
+        key='bbb', num_buckets=3)
+    embedding_dimension = 2
+    embedding_column_b, embedding_column_a = fc_old.shared_embedding_columns(
+        [categorical_column_b, categorical_column_a],
+        dimension=embedding_dimension)
+
+    with ops.Graph().as_default():
+      features = {
+          'aaa':
+              sparse_tensor.SparseTensor(
+                  indices=((0, 0), (1, 0), (1, 1)),
+                  values=(0, 1, 0),
+                  dense_shape=(2, 2)),
+          'bbb':
+              sparse_tensor.SparseTensor(
+                  indices=((0, 0), (1, 0), (1, 1)),
+                  values=(1, 2, 1),
+                  dense_shape=(2, 2)),
+      }
+      all_cols = [embedding_column_a, embedding_column_b]
+      fc.input_layer(features, all_cols)
+      fc.input_layer(features, all_cols)
+      # Make sure that only 1 variable gets created in this case.
+      self.assertEqual(1, len(
+          ops.get_collection(ops.GraphKeys.GLOBAL_VARIABLES)))
+      self.assertItemsEqual(
+          ['input_layer/aaa_bbb_shared_embedding/embedding_weights:0'],
+          [v.name for v in ops.get_collection(ops.GraphKeys.GLOBAL_VARIABLES)])
+
+  def test_multiple_layers_with_same_shared_embedding_column_diff_graphs(self):
+    categorical_column_a = fc_old.categorical_column_with_identity(
+        key='aaa', num_buckets=3)
+    categorical_column_b = fc_old.categorical_column_with_identity(
+        key='bbb', num_buckets=3)
+    embedding_dimension = 2
+    embedding_column_b, embedding_column_a = fc_old.shared_embedding_columns(
+        [categorical_column_b, categorical_column_a],
+        dimension=embedding_dimension)
+    all_cols = [embedding_column_a, embedding_column_b]
+
+    with ops.Graph().as_default():
+      features = {
+          'aaa':
+              sparse_tensor.SparseTensor(
+                  indices=((0, 0), (1, 0), (1, 1)),
+                  values=(0, 1, 0),
+                  dense_shape=(2, 2)),
+          'bbb':
+              sparse_tensor.SparseTensor(
+                  indices=((0, 0), (1, 0), (1, 1)),
+                  values=(1, 2, 1),
+                  dense_shape=(2, 2)),
+      }
+      fc.input_layer(features, all_cols)
+      # Make sure that only 1 variable gets created in this case.
+      self.assertEqual(1, len(
+          ops.get_collection(ops.GraphKeys.GLOBAL_VARIABLES)))
+
+    with ops.Graph().as_default():
+      features1 = {
+          'aaa':
+              sparse_tensor.SparseTensor(
+                  indices=((0, 0), (1, 0), (1, 1)),
+                  values=(0, 1, 0),
+                  dense_shape=(2, 2)),
+          'bbb':
+              sparse_tensor.SparseTensor(
+                  indices=((0, 0), (1, 0), (1, 1)),
+                  values=(1, 2, 1),
+                  dense_shape=(2, 2)),
+      }
+
+      fc.input_layer(features1, all_cols)
+      # Make sure that only 1 variable gets created in this case.
+      self.assertEqual(1, len(
+          ops.get_collection(ops.GraphKeys.GLOBAL_VARIABLES)))
+      self.assertItemsEqual(
+          ['input_layer/aaa_bbb_shared_embedding/embedding_weights:0'],
+          [v.name for v in ops.get_collection(ops.GraphKeys.GLOBAL_VARIABLES)])
+
+  def test_with_numpy_input_fn(self):
+    embedding_values = (
+        (1., 2., 3., 4., 5.),  # id 0
+        (6., 7., 8., 9., 10.),  # id 1
+        (11., 12., 13., 14., 15.)  # id 2
+    )
+    def _initializer(shape, dtype, partition_info):
+      del shape, dtype, partition_info
+      return embedding_values
+
+    # price has 1 dimension in input_layer
+    price = fc_old.numeric_column('price')
+    body_style = fc_old.categorical_column_with_vocabulary_list(
+        'body-style', vocabulary_list=['hardtop', 'wagon', 'sedan'])
+    # one_hot_body_style has 3 dims in input_layer.
+    one_hot_body_style = fc_old.indicator_column(body_style)
+    # embedded_body_style has 5 dims in input_layer.
+    embedded_body_style = fc_old.embedding_column(
+        body_style, dimension=5, initializer=_initializer)
+
+    input_fn = numpy_io.numpy_input_fn(
+        x={
+            'price': np.array([11., 12., 13., 14.]),
+            'body-style': np.array(['sedan', 'hardtop', 'wagon', 'sedan']),
+        },
+        batch_size=2,
+        shuffle=False)
+    features = input_fn()
+    net = fc.input_layer(features,
+                         [price, one_hot_body_style, embedded_body_style])
+    self.assertEqual(1 + 3 + 5, net.shape[1])
+    with _initialized_session() as sess:
+      coord = coordinator.Coordinator()
+      threads = queue_runner_impl.start_queue_runners(sess, coord=coord)
+
+      # Each row is formed by concatenating `embedded_body_style`,
+      # `one_hot_body_style`, and `price` in order.
+      self.assertAllEqual(
+          [[11., 12., 13., 14., 15., 0., 0., 1., 11.],
+           [1., 2., 3., 4., 5., 1., 0., 0., 12]],
+          sess.run(net))
+
+      coord.request_stop()
+      coord.join(threads)
+
+  def test_with_1d_sparse_tensor(self):
+    embedding_values = (
+        (1., 2., 3., 4., 5.),  # id 0
+        (6., 7., 8., 9., 10.),  # id 1
+        (11., 12., 13., 14., 15.)  # id 2
+    )
+    def _initializer(shape, dtype, partition_info):
+      del shape, dtype, partition_info
+      return embedding_values
+
+    # price has 1 dimension in input_layer
+    price = fc_old.numeric_column('price')
+
+    # one_hot_body_style has 3 dims in input_layer.
+    body_style = fc_old.categorical_column_with_vocabulary_list(
+        'body-style', vocabulary_list=['hardtop', 'wagon', 'sedan'])
+    one_hot_body_style = fc_old.indicator_column(body_style)
+
+    # embedded_body_style has 5 dims in input_layer.
+    country = fc_old.categorical_column_with_vocabulary_list(
+        'country', vocabulary_list=['US', 'JP', 'CA'])
+    embedded_country = fc_old.embedding_column(
+        country, dimension=5, initializer=_initializer)
+
+    # Provides 1-dim tensor and dense tensor.
+    features = {
+        'price': constant_op.constant([11., 12.,]),
+        'body-style': sparse_tensor.SparseTensor(
+            indices=((0,), (1,)),
+            values=('sedan', 'hardtop'),
+            dense_shape=(2,)),
+        # This is dense tensor for the categorical_column.
+        'country': constant_op.constant(['CA', 'US']),
+    }
+    self.assertEqual(1, features['price'].shape.ndims)
+    self.assertEqual(1, features['body-style'].dense_shape.get_shape()[0])
+    self.assertEqual(1, features['country'].shape.ndims)
+
+    net = fc.input_layer(features,
+                         [price, one_hot_body_style, embedded_country])
+    self.assertEqual(1 + 3 + 5, net.shape[1])
+    with _initialized_session() as sess:
+
+      # Each row is formed by concatenating `embedded_body_style`,
+      # `one_hot_body_style`, and `price` in order.
+      self.assertAllEqual(
+          [[0., 0., 1., 11., 12., 13., 14., 15., 11.],
+           [1., 0., 0., 1., 2., 3., 4., 5., 12.]],
+          sess.run(net))
+
+  def test_with_1d_unknown_shape_sparse_tensor(self):
+    embedding_values = (
+        (1., 2.),  # id 0
+        (6., 7.),  # id 1
+        (11., 12.)  # id 2
+    )
+    def _initializer(shape, dtype, partition_info):
+      del shape, dtype, partition_info
+      return embedding_values
+
+    # price has 1 dimension in input_layer
+    price = fc_old.numeric_column('price')
+
+    # one_hot_body_style has 3 dims in input_layer.
+    body_style = fc_old.categorical_column_with_vocabulary_list(
+        'body-style', vocabulary_list=['hardtop', 'wagon', 'sedan'])
+    one_hot_body_style = fc_old.indicator_column(body_style)
+
+    # embedded_body_style has 5 dims in input_layer.
+    country = fc_old.categorical_column_with_vocabulary_list(
+        'country', vocabulary_list=['US', 'JP', 'CA'])
+    embedded_country = fc_old.embedding_column(
+        country, dimension=2, initializer=_initializer)
+
+    # Provides 1-dim tensor and dense tensor.
+    features = {
+        'price': array_ops.placeholder(dtypes.float32),
+        'body-style': array_ops.sparse_placeholder(dtypes.string),
+        # This is dense tensor for the categorical_column.
+        'country': array_ops.placeholder(dtypes.string),
+    }
+    self.assertIsNone(features['price'].shape.ndims)
+    self.assertIsNone(features['body-style'].get_shape().ndims)
+    self.assertIsNone(features['country'].shape.ndims)
+
+    price_data = np.array([11., 12.])
+    body_style_data = sparse_tensor.SparseTensorValue(
+        indices=((0,), (1,)),
+        values=('sedan', 'hardtop'),
+        dense_shape=(2,))
+    country_data = np.array([['US'], ['CA']])
+
+    net = fc.input_layer(features,
+                         [price, one_hot_body_style, embedded_country])
+    self.assertEqual(1 + 3 + 2, net.shape[1])
+    with _initialized_session() as sess:
+
+      # Each row is formed by concatenating `embedded_body_style`,
+      # `one_hot_body_style`, and `price` in order.
+      self.assertAllEqual(
+          [[0., 0., 1., 1., 2., 11.], [1., 0., 0., 11., 12., 12.]],
+          sess.run(
+              net,
+              feed_dict={
+                  features['price']: price_data,
+                  features['body-style']: body_style_data,
+                  features['country']: country_data
+              }))
+
+  def test_with_rank_0_feature(self):
+    # price has 1 dimension in input_layer
+    price = fc_old.numeric_column('price')
+    features = {
+        'price': constant_op.constant(0),
+    }
+    self.assertEqual(0, features['price'].shape.ndims)
+
+    # Static rank 0 should fail
+    with self.assertRaisesRegexp(ValueError, 'Feature .* cannot have rank 0'):
+      fc.input_layer(features, [price])
+
+    # Dynamic rank 0 should fail
+    features = {
+        'price': array_ops.placeholder(dtypes.float32),
+    }
+    net = fc.input_layer(features, [price])
+    self.assertEqual(1, net.shape[1])
+    with _initialized_session() as sess:
+      with self.assertRaisesOpError('Feature .* cannot have rank 0'):
+        sess.run(net, feed_dict={features['price']: np.array(1)})
+
+
+class MakeParseExampleSpecTest(test.TestCase):
+
+  class _TestFeatureColumn(FeatureColumn,
+                           collections.namedtuple('_TestFeatureColumn',
+                                                  ('parse_spec'))):
+
+    @property
+    def name(self):
+      return "_TestFeatureColumn"
+
+    def transform_feature(self, transformation_cache, state_manager):
+      pass
+
+    @property
+    def parse_example_spec(self):
+      return self.parse_spec
+
+  def test_no_feature_columns(self):
+    actual = fc.make_parse_example_spec([])
+    self.assertDictEqual({}, actual)
+
+  def test_invalid_type(self):
+    key1 = 'key1'
+    parse_spec1 = parsing_ops.FixedLenFeature(
+        shape=(2,), dtype=dtypes.float32, default_value=0.)
+    with self.assertRaisesRegexp(
+        ValueError,
+        'All feature_columns must be FeatureColumn instances.*invalid_column'):
+      fc.make_parse_example_spec(
+          (self._TestFeatureColumn({key1: parse_spec1}), 'invalid_column'))
+
+  def test_one_feature_column(self):
+    key1 = 'key1'
+    parse_spec1 = parsing_ops.FixedLenFeature(
+        shape=(2,), dtype=dtypes.float32, default_value=0.)
+    actual = fc.make_parse_example_spec(
+        (self._TestFeatureColumn({key1: parse_spec1}),))
+    self.assertDictEqual({key1: parse_spec1}, actual)
+
+  def test_two_feature_columns(self):
+    key1 = 'key1'
+    parse_spec1 = parsing_ops.FixedLenFeature(
+        shape=(2,), dtype=dtypes.float32, default_value=0.)
+    key2 = 'key2'
+    parse_spec2 = parsing_ops.VarLenFeature(dtype=dtypes.string)
+    actual = fc.make_parse_example_spec(
+        (self._TestFeatureColumn({key1: parse_spec1}),
+         self._TestFeatureColumn({key2: parse_spec2})))
+    self.assertDictEqual({key1: parse_spec1, key2: parse_spec2}, actual)
+
+  def test_equal_keys_different_parse_spec(self):
+    key1 = 'key1'
+    parse_spec1 = parsing_ops.FixedLenFeature(
+        shape=(2,), dtype=dtypes.float32, default_value=0.)
+    parse_spec2 = parsing_ops.VarLenFeature(dtype=dtypes.string)
+    with self.assertRaisesRegexp(
+        ValueError,
+        'feature_columns contain different parse_spec for key key1'):
+      fc.make_parse_example_spec(
+          (self._TestFeatureColumn({key1: parse_spec1}),
+           self._TestFeatureColumn({key1: parse_spec2})))
+
+  def test_equal_keys_equal_parse_spec(self):
+    key1 = 'key1'
+    parse_spec1 = parsing_ops.FixedLenFeature(
+        shape=(2,), dtype=dtypes.float32, default_value=0.)
+    actual = fc.make_parse_example_spec(
+        (self._TestFeatureColumn({key1: parse_spec1}),
+         self._TestFeatureColumn({key1: parse_spec1})))
+    self.assertDictEqual({key1: parse_spec1}, actual)
+
+  def test_multiple_features_dict(self):
+    """parse_spc for one column is a dict with length > 1."""
+    key1 = 'key1'
+    parse_spec1 = parsing_ops.FixedLenFeature(
+        shape=(2,), dtype=dtypes.float32, default_value=0.)
+    key2 = 'key2'
+    parse_spec2 = parsing_ops.VarLenFeature(dtype=dtypes.string)
+    key3 = 'key3'
+    parse_spec3 = parsing_ops.VarLenFeature(dtype=dtypes.int32)
+    actual = fc.make_parse_example_spec(
+        (self._TestFeatureColumn({key1: parse_spec1}),
+         self._TestFeatureColumn({key2: parse_spec2, key3: parse_spec3})))
+    self.assertDictEqual(
+        {key1: parse_spec1, key2: parse_spec2, key3: parse_spec3}, actual)
+
+
+def _assert_sparse_tensor_value(test_case, expected, actual):
+  test_case.assertEqual(np.int64, np.array(actual.indices).dtype)
+  test_case.assertAllEqual(expected.indices, actual.indices)
+
+  test_case.assertEqual(
+      np.array(expected.values).dtype, np.array(actual.values).dtype)
+  test_case.assertAllEqual(expected.values, actual.values)
+
+  test_case.assertEqual(np.int64, np.array(actual.dense_shape).dtype)
+  test_case.assertAllEqual(expected.dense_shape, actual.dense_shape)
+
+
+class VocabularyFileCategoricalColumnTest(test.TestCase):
+
+  def setUp(self):
+    super(VocabularyFileCategoricalColumnTest, self).setUp()
+
+    # Contains ints, Golden State Warriors jersey numbers: 30, 35, 11, 23, 22
+    self._warriors_vocabulary_file_name = test.test_src_dir_path(
+        'python/feature_column/testdata/warriors_vocabulary.txt')
+    self._warriors_vocabulary_size = 5
+
+    # Contains strings, character names from 'The Wire': omar, stringer, marlo
+    self._wire_vocabulary_file_name = test.test_src_dir_path(
+        'python/feature_column/testdata/wire_vocabulary.txt')
+    self._wire_vocabulary_size = 3
+
+  def test_defaults(self):
+    column = fc.categorical_column_with_vocabulary_file(
+        key='aaa', vocabulary_file='path_to_file', vocabulary_size=3)
+    self.assertEqual('aaa', column.name)
+    self.assertEqual('aaa', column.key)
+    self.assertEqual(3, column.num_buckets)
+    self.assertEqual({
+        'aaa': parsing_ops.VarLenFeature(dtypes.string)
+    }, column.parse_example_spec)
+
+  def test_key_should_be_string(self):
+    with self.assertRaisesRegexp(ValueError, 'key must be a string.'):
+      fc.categorical_column_with_vocabulary_file(
+          key=('aaa',), vocabulary_file='path_to_file', vocabulary_size=3)
+
+  def test_all_constructor_args(self):
+    column = fc.categorical_column_with_vocabulary_file(
+        key='aaa', vocabulary_file='path_to_file', vocabulary_size=3,
+        num_oov_buckets=4, dtype=dtypes.int32)
+    self.assertEqual(7, column.num_buckets)
+    self.assertEqual({
+        'aaa': parsing_ops.VarLenFeature(dtypes.int32)
+    }, column.parse_example_spec)
+
+  def test_deep_copy(self):
+    original = fc.categorical_column_with_vocabulary_file(
+        key='aaa', vocabulary_file='path_to_file', vocabulary_size=3,
+        num_oov_buckets=4, dtype=dtypes.int32)
+    for column in (original, copy.deepcopy(original)):
+      self.assertEqual('aaa', column.name)
+      self.assertEqual(7, column.num_buckets)
+      self.assertEqual({
+          'aaa': parsing_ops.VarLenFeature(dtypes.int32)
+      }, column.parse_example_spec)
+
+  def test_vocabulary_file_none(self):
+    with self.assertRaisesRegexp(ValueError, 'Missing vocabulary_file'):
+      fc.categorical_column_with_vocabulary_file(
+          key='aaa', vocabulary_file=None, vocabulary_size=3)
+
+  def test_vocabulary_file_empty_string(self):
+    with self.assertRaisesRegexp(ValueError, 'Missing vocabulary_file'):
+      fc.categorical_column_with_vocabulary_file(
+          key='aaa', vocabulary_file='', vocabulary_size=3)
+
+  def test_invalid_vocabulary_file(self):
+    column = fc.categorical_column_with_vocabulary_file(
+        key='aaa', vocabulary_file='file_does_not_exist', vocabulary_size=10)
+    inputs = sparse_tensor.SparseTensorValue(
+        indices=((0, 0), (1, 0), (1, 1)),
+        values=('marlo', 'skywalker', 'omar'),
+        dense_shape=(2, 2))
+    column.get_sparse_tensors(FeatureTransformationCache({'aaa': inputs}), None)
+    with self.assertRaisesRegexp(errors.OpError, 'file_does_not_exist'):
+      with self.test_session():
+        lookup_ops.tables_initializer().run()
+
+  def test_invalid_vocabulary_size(self):
+    with self.assertRaisesRegexp(ValueError, 'Invalid vocabulary_size'):
+      fc.categorical_column_with_vocabulary_file(
+          key='aaa', vocabulary_file=self._wire_vocabulary_file_name,
+          vocabulary_size=-1)
+    with self.assertRaisesRegexp(ValueError, 'Invalid vocabulary_size'):
+      fc.categorical_column_with_vocabulary_file(
+          key='aaa', vocabulary_file=self._wire_vocabulary_file_name,
+          vocabulary_size=0)
+
+  def test_too_large_vocabulary_size(self):
+    column = fc.categorical_column_with_vocabulary_file(
+        key='aaa',
+        vocabulary_file=self._wire_vocabulary_file_name,
+        vocabulary_size=self._wire_vocabulary_size + 1)
+    inputs = sparse_tensor.SparseTensorValue(
+        indices=((0, 0), (1, 0), (1, 1)),
+        values=('marlo', 'skywalker', 'omar'),
+        dense_shape=(2, 2))
+    column.get_sparse_tensors(FeatureTransformationCache({'aaa': inputs}), None)
+    with self.assertRaisesRegexp(errors.OpError, 'Invalid vocab_size'):
+      with self.test_session():
+        lookup_ops.tables_initializer().run()
+
+  def test_invalid_num_oov_buckets(self):
+    with self.assertRaisesRegexp(ValueError, 'Invalid num_oov_buckets'):
+      fc.categorical_column_with_vocabulary_file(
+          key='aaa', vocabulary_file='path', vocabulary_size=3,
+          num_oov_buckets=-1)
+
+  def test_invalid_dtype(self):
+    with self.assertRaisesRegexp(ValueError, 'dtype must be string or integer'):
+      fc.categorical_column_with_vocabulary_file(
+          key='aaa', vocabulary_file='path', vocabulary_size=3,
+          dtype=dtypes.float64)
+
+  def test_invalid_buckets_and_default_value(self):
+    with self.assertRaisesRegexp(
+        ValueError, 'both num_oov_buckets and default_value'):
+      fc.categorical_column_with_vocabulary_file(
+          key='aaa',
+          vocabulary_file=self._wire_vocabulary_file_name,
+          vocabulary_size=self._wire_vocabulary_size,
+          num_oov_buckets=100,
+          default_value=2)
+
+  def test_invalid_input_dtype_int32(self):
+    column = fc.categorical_column_with_vocabulary_file(
+        key='aaa',
+        vocabulary_file=self._wire_vocabulary_file_name,
+        vocabulary_size=self._wire_vocabulary_size,
+        dtype=dtypes.string)
+    inputs = sparse_tensor.SparseTensorValue(
+        indices=((0, 0), (1, 0), (1, 1)),
+        values=(12, 24, 36),
+        dense_shape=(2, 2))
+    with self.assertRaisesRegexp(ValueError, 'dtype must be compatible'):
+      column.get_sparse_tensors(
+          FeatureTransformationCache({
+              'aaa': inputs
+          }), None)
+
+  def test_invalid_input_dtype_string(self):
+    column = fc.categorical_column_with_vocabulary_file(
+        key='aaa',
+        vocabulary_file=self._warriors_vocabulary_file_name,
+        vocabulary_size=self._warriors_vocabulary_size,
+        dtype=dtypes.int32)
+    inputs = sparse_tensor.SparseTensorValue(
+        indices=((0, 0), (1, 0), (1, 1)),
+        values=('omar', 'stringer', 'marlo'),
+        dense_shape=(2, 2))
+    with self.assertRaisesRegexp(ValueError, 'dtype must be compatible'):
+      column.get_sparse_tensors(
+          FeatureTransformationCache({
+              'aaa': inputs
+          }), None)
+
+  def test_parse_example(self):
+    a = fc.categorical_column_with_vocabulary_file(
+        key='aaa', vocabulary_file='path_to_file', vocabulary_size=3)
+    data = example_pb2.Example(features=feature_pb2.Features(
+        feature={
+            'aaa':
+                feature_pb2.Feature(bytes_list=feature_pb2.BytesList(
+                    value=[b'omar', b'stringer']))
+        }))
+    features = parsing_ops.parse_example(
+        serialized=[data.SerializeToString()],
+        features=fc.make_parse_example_spec([a]))
+    self.assertIn('aaa', features)
+    with self.test_session():
+      _assert_sparse_tensor_value(
+          self,
+          sparse_tensor.SparseTensorValue(
+              indices=[[0, 0], [0, 1]],
+              values=np.array([b'omar', b'stringer'], dtype=np.object_),
+              dense_shape=[1, 2]),
+          features['aaa'].eval())
+
+  def test_get_sparse_tensors(self):
+    column = fc.categorical_column_with_vocabulary_file(
+        key='aaa',
+        vocabulary_file=self._wire_vocabulary_file_name,
+        vocabulary_size=self._wire_vocabulary_size)
+    inputs = sparse_tensor.SparseTensorValue(
+        indices=((0, 0), (1, 0), (1, 1)),
+        values=('marlo', 'skywalker', 'omar'),
+        dense_shape=(2, 2))
+    id_weight_pair = column.get_sparse_tensors(
+        FeatureTransformationCache({
+            'aaa': inputs
+        }), None)
+    self.assertIsNone(id_weight_pair.weight_tensor)
+    with _initialized_session():
+      _assert_sparse_tensor_value(
+          self,
+          sparse_tensor.SparseTensorValue(
+              indices=inputs.indices,
+              values=np.array((2, -1, 0), dtype=np.int64),
+              dense_shape=inputs.dense_shape),
+          id_weight_pair.id_tensor.eval())
+
+  def test_get_sparse_tensors_none_vocabulary_size(self):
+    column = fc.categorical_column_with_vocabulary_file(
+        key='aaa', vocabulary_file=self._wire_vocabulary_file_name)
+    inputs = sparse_tensor.SparseTensorValue(
+        indices=((0, 0), (1, 0), (1, 1)),
+        values=('marlo', 'skywalker', 'omar'),
+        dense_shape=(2, 2))
+    id_weight_pair = column.get_sparse_tensors(
+        FeatureTransformationCache({
+            'aaa': inputs
+        }), None)
+    self.assertIsNone(id_weight_pair.weight_tensor)
+    with _initialized_session():
+      _assert_sparse_tensor_value(self,
+                                  sparse_tensor.SparseTensorValue(
+                                      indices=inputs.indices,
+                                      values=np.array(
+                                          (2, -1, 0), dtype=np.int64),
+                                      dense_shape=inputs.dense_shape),
+                                  id_weight_pair.id_tensor.eval())
+
+  def test_transform_feature(self):
+    column = fc.categorical_column_with_vocabulary_file(
+        key='aaa',
+        vocabulary_file=self._wire_vocabulary_file_name,
+        vocabulary_size=self._wire_vocabulary_size)
+    inputs = sparse_tensor.SparseTensorValue(
+        indices=((0, 0), (1, 0), (1, 1)),
+        values=('marlo', 'skywalker', 'omar'),
+        dense_shape=(2, 2))
+    id_tensor = _transform_features({'aaa': inputs}, [column], None)[column]
+    with _initialized_session():
+      _assert_sparse_tensor_value(self,
+                                  sparse_tensor.SparseTensorValue(
+                                      indices=inputs.indices,
+                                      values=np.array(
+                                          (2, -1, 0), dtype=np.int64),
+                                      dense_shape=inputs.dense_shape),
+                                  id_tensor.eval())
+
+  def DISABLED_test_get_sparse_tensors_weight_collections(self):
+    column = fc.categorical_column_with_vocabulary_file(
+        key='aaa',
+        vocabulary_file=self._wire_vocabulary_file_name,
+        vocabulary_size=self._wire_vocabulary_size)
+    inputs = sparse_tensor.SparseTensor(
+        values=['omar', 'stringer', 'marlo'],
+        indices=[[0, 0], [1, 0], [1, 1]],
+        dense_shape=[2, 2])
+    column.get_sparse_tensors(
+        FeatureTransformationCache({
+            'aaa': inputs
+        }),
+        weight_collections=('my_weights',))
+
+    self.assertItemsEqual(
+        [], ops.get_collection(ops.GraphKeys.GLOBAL_VARIABLES))
+    self.assertItemsEqual([], ops.get_collection('my_weights'))
+
+  def test_get_sparse_tensors_dense_input(self):
+    column = fc.categorical_column_with_vocabulary_file(
+        key='aaa',
+        vocabulary_file=self._wire_vocabulary_file_name,
+        vocabulary_size=self._wire_vocabulary_size)
+    id_weight_pair = column.get_sparse_tensors(
+        FeatureTransformationCache({
+            'aaa': (('marlo', ''), ('skywalker', 'omar'))
+        }), None)
+    self.assertIsNone(id_weight_pair.weight_tensor)
+    with _initialized_session():
+      _assert_sparse_tensor_value(
+          self,
+          sparse_tensor.SparseTensorValue(
+              indices=((0, 0), (1, 0), (1, 1)),
+              values=np.array((2, -1, 0), dtype=np.int64),
+              dense_shape=(2, 2)),
+          id_weight_pair.id_tensor.eval())
+
+  def test_get_sparse_tensors_default_value_in_vocabulary(self):
+    column = fc.categorical_column_with_vocabulary_file(
+        key='aaa',
+        vocabulary_file=self._wire_vocabulary_file_name,
+        vocabulary_size=self._wire_vocabulary_size,
+        default_value=2)
+    inputs = sparse_tensor.SparseTensorValue(
+        indices=((0, 0), (1, 0), (1, 1)),
+        values=('marlo', 'skywalker', 'omar'),
+        dense_shape=(2, 2))
+    id_weight_pair = column.get_sparse_tensors(
+        FeatureTransformationCache({
+            'aaa': inputs
+        }), None)
+    self.assertIsNone(id_weight_pair.weight_tensor)
+    with _initialized_session():
+      _assert_sparse_tensor_value(
+          self,
+          sparse_tensor.SparseTensorValue(
+              indices=inputs.indices,
+              values=np.array((2, 2, 0), dtype=np.int64),
+              dense_shape=inputs.dense_shape),
+          id_weight_pair.id_tensor.eval())
+
+  def test_get_sparse_tensors_with_oov_buckets(self):
+    column = fc.categorical_column_with_vocabulary_file(
+        key='aaa',
+        vocabulary_file=self._wire_vocabulary_file_name,
+        vocabulary_size=self._wire_vocabulary_size,
+        num_oov_buckets=100)
+    inputs = sparse_tensor.SparseTensorValue(
+        indices=((0, 0), (1, 0), (1, 1), (1, 2)),
+        values=('marlo', 'skywalker', 'omar', 'heisenberg'),
+        dense_shape=(2, 3))
+    id_weight_pair = column.get_sparse_tensors(
+        FeatureTransformationCache({
+            'aaa': inputs
+        }), None)
+    self.assertIsNone(id_weight_pair.weight_tensor)
+    with _initialized_session():
+      _assert_sparse_tensor_value(
+          self,
+          sparse_tensor.SparseTensorValue(
+              indices=inputs.indices,
+              values=np.array((2, 33, 0, 62), dtype=np.int64),
+              dense_shape=inputs.dense_shape),
+          id_weight_pair.id_tensor.eval())
+
+  def test_get_sparse_tensors_small_vocabulary_size(self):
+    # 'marlo' is the last entry in our vocabulary file, so be setting
+    # `vocabulary_size` to 1 less than number of entries in file, we take
+    # 'marlo' out of the vocabulary.
+    column = fc.categorical_column_with_vocabulary_file(
+        key='aaa',
+        vocabulary_file=self._wire_vocabulary_file_name,
+        vocabulary_size=self._wire_vocabulary_size - 1)
+    inputs = sparse_tensor.SparseTensorValue(
+        indices=((0, 0), (1, 0), (1, 1)),
+        values=('marlo', 'skywalker', 'omar'),
+        dense_shape=(2, 2))
+    id_weight_pair = column.get_sparse_tensors(
+        FeatureTransformationCache({
+            'aaa': inputs
+        }), None)
+    self.assertIsNone(id_weight_pair.weight_tensor)
+    with _initialized_session():
+      _assert_sparse_tensor_value(
+          self,
+          sparse_tensor.SparseTensorValue(
+              indices=inputs.indices,
+              values=np.array((-1, -1, 0), dtype=np.int64),
+              dense_shape=inputs.dense_shape),
+          id_weight_pair.id_tensor.eval())
+
+  def test_get_sparse_tensors_int32(self):
+    column = fc.categorical_column_with_vocabulary_file(
+        key='aaa',
+        vocabulary_file=self._warriors_vocabulary_file_name,
+        vocabulary_size=self._warriors_vocabulary_size,
+        dtype=dtypes.int32)
+    inputs = sparse_tensor.SparseTensorValue(
+        indices=((0, 0), (1, 0), (1, 1), (2, 2)),
+        values=(11, 100, 30, 22),
+        dense_shape=(3, 3))
+    id_weight_pair = column.get_sparse_tensors(
+        FeatureTransformationCache({
+            'aaa': inputs
+        }), None)
+    self.assertIsNone(id_weight_pair.weight_tensor)
+    with _initialized_session():
+      _assert_sparse_tensor_value(
+          self,
+          sparse_tensor.SparseTensorValue(
+              indices=inputs.indices,
+              values=np.array((2, -1, 0, 4), dtype=np.int64),
+              dense_shape=inputs.dense_shape),
+          id_weight_pair.id_tensor.eval())
+
+  def test_get_sparse_tensors_int32_dense_input(self):
+    default_value = -100
+    column = fc.categorical_column_with_vocabulary_file(
+        key='aaa',
+        vocabulary_file=self._warriors_vocabulary_file_name,
+        vocabulary_size=self._warriors_vocabulary_size,
+        dtype=dtypes.int32,
+        default_value=default_value)
+    id_weight_pair = column.get_sparse_tensors(
+        FeatureTransformationCache({
+            'aaa': ((11, -1, -1), (100, 30, -1), (-1, -1, 22))
+        }), None)
+    self.assertIsNone(id_weight_pair.weight_tensor)
+    with _initialized_session():
+      _assert_sparse_tensor_value(
+          self,
+          sparse_tensor.SparseTensorValue(
+              indices=((0, 0), (1, 0), (1, 1), (2, 2)),
+              values=np.array((2, default_value, 0, 4), dtype=np.int64),
+              dense_shape=(3, 3)),
+          id_weight_pair.id_tensor.eval())
+
+  def test_get_sparse_tensors_int32_with_oov_buckets(self):
+    column = fc.categorical_column_with_vocabulary_file(
+        key='aaa',
+        vocabulary_file=self._warriors_vocabulary_file_name,
+        vocabulary_size=self._warriors_vocabulary_size,
+        dtype=dtypes.int32,
+        num_oov_buckets=100)
+    inputs = sparse_tensor.SparseTensorValue(
+        indices=((0, 0), (1, 0), (1, 1), (2, 2)),
+        values=(11, 100, 30, 22),
+        dense_shape=(3, 3))
+    id_weight_pair = column.get_sparse_tensors(
+        FeatureTransformationCache({
+            'aaa': inputs
+        }), None)
+    self.assertIsNone(id_weight_pair.weight_tensor)
+    with _initialized_session():
+      _assert_sparse_tensor_value(
+          self,
+          sparse_tensor.SparseTensorValue(
+              indices=inputs.indices,
+              values=np.array((2, 60, 0, 4), dtype=np.int64),
+              dense_shape=inputs.dense_shape),
+          id_weight_pair.id_tensor.eval())
+
+  def test_linear_model(self):
+    wire_column = fc_old.categorical_column_with_vocabulary_file(
+        key='wire',
+        vocabulary_file=self._wire_vocabulary_file_name,
+        vocabulary_size=self._wire_vocabulary_size,
+        num_oov_buckets=1)
+    self.assertEqual(4, wire_column._num_buckets)
+    with ops.Graph().as_default():
+      predictions = fc.linear_model({
+          wire_column.name: sparse_tensor.SparseTensorValue(
+              indices=((0, 0), (1, 0), (1, 1)),
+              values=('marlo', 'skywalker', 'omar'),
+              dense_shape=(2, 2))
+      }, (wire_column,))
+      bias = get_linear_model_bias()
+      wire_var = get_linear_model_column_var(wire_column)
+      with _initialized_session():
+        self.assertAllClose((0.,), bias.eval())
+        self.assertAllClose(((0.,), (0.,), (0.,), (0.,)), wire_var.eval())
+        self.assertAllClose(((0.,), (0.,)), predictions.eval())
+        wire_var.assign(((1.,), (2.,), (3.,), (4.,))).eval()
+        # 'marlo' -> 2: wire_var[2] = 3
+        # 'skywalker' -> 3, 'omar' -> 0: wire_var[3] + wire_var[0] = 4+1 = 5
+        self.assertAllClose(((3.,), (5.,)), predictions.eval())
+
+  def test_keras_linear_model(self):
+    wire_column = fc_old.categorical_column_with_vocabulary_file(
+        key='wire',
+        vocabulary_file=self._wire_vocabulary_file_name,
+        vocabulary_size=self._wire_vocabulary_size,
+        num_oov_buckets=1)
+    self.assertEqual(4, wire_column._num_buckets)
+    with ops.Graph().as_default():
+      predictions = get_keras_linear_model_predictions({
+          wire_column.name:
+              sparse_tensor.SparseTensorValue(
+                  indices=((0, 0), (1, 0), (1, 1)),
+                  values=('marlo', 'skywalker', 'omar'),
+                  dense_shape=(2, 2))
+      }, (wire_column,))
+      bias = get_linear_model_bias()
+      wire_var = get_linear_model_column_var(wire_column)
+      with _initialized_session():
+        self.assertAllClose((0.,), bias.eval())
+        self.assertAllClose(((0.,), (0.,), (0.,), (0.,)), wire_var.eval())
+        self.assertAllClose(((0.,), (0.,)), predictions.eval())
+        wire_var.assign(((1.,), (2.,), (3.,), (4.,))).eval()
+        # 'marlo' -> 2: wire_var[2] = 3
+        # 'skywalker' -> 3, 'omar' -> 0: wire_var[3] + wire_var[0] = 4+1 = 5
+        self.assertAllClose(((3.,), (5.,)), predictions.eval())
+
+
+class VocabularyListCategoricalColumnTest(test.TestCase):
+
+  def test_defaults_string(self):
+    column = fc.categorical_column_with_vocabulary_list(
+        key='aaa', vocabulary_list=('omar', 'stringer', 'marlo'))
+    self.assertEqual('aaa', column.name)
+    self.assertEqual('aaa', column.key)
+    self.assertEqual(3, column.num_buckets)
+    self.assertEqual({
+        'aaa': parsing_ops.VarLenFeature(dtypes.string)
+    }, column.parse_example_spec)
+
+  def test_key_should_be_string(self):
+    with self.assertRaisesRegexp(ValueError, 'key must be a string.'):
+      fc.categorical_column_with_vocabulary_list(
+          key=('aaa',), vocabulary_list=('omar', 'stringer', 'marlo'))
+
+  def test_defaults_int(self):
+    column = fc.categorical_column_with_vocabulary_list(
+        key='aaa', vocabulary_list=(12, 24, 36))
+    self.assertEqual('aaa', column.name)
+    self.assertEqual('aaa', column.key)
+    self.assertEqual(3, column.num_buckets)
+    self.assertEqual({
+        'aaa': parsing_ops.VarLenFeature(dtypes.int64)
+    }, column.parse_example_spec)
+
+  def test_all_constructor_args(self):
+    column = fc.categorical_column_with_vocabulary_list(
+        key='aaa', vocabulary_list=(12, 24, 36), dtype=dtypes.int32,
+        default_value=-99)
+    self.assertEqual(3, column.num_buckets)
+    self.assertEqual({
+        'aaa': parsing_ops.VarLenFeature(dtypes.int32)
+    }, column.parse_example_spec)
+
+  def test_deep_copy(self):
+    original = fc.categorical_column_with_vocabulary_list(
+        key='aaa', vocabulary_list=(12, 24, 36), dtype=dtypes.int32)
+    for column in (original, copy.deepcopy(original)):
+      self.assertEqual('aaa', column.name)
+      self.assertEqual(3, column.num_buckets)
+      self.assertEqual({
+          'aaa': parsing_ops.VarLenFeature(dtypes.int32)
+      }, column.parse_example_spec)
+
+  def test_invalid_dtype(self):
+    with self.assertRaisesRegexp(ValueError, 'dtype must be string or integer'):
+      fc.categorical_column_with_vocabulary_list(
+          key='aaa', vocabulary_list=('omar', 'stringer', 'marlo'),
+          dtype=dtypes.float32)
+
+  def test_invalid_mapping_dtype(self):
+    with self.assertRaisesRegexp(
+        ValueError, r'vocabulary dtype must be string or integer'):
+      fc.categorical_column_with_vocabulary_list(
+          key='aaa', vocabulary_list=(12., 24., 36.))
+
+  def test_mismatched_int_dtype(self):
+    with self.assertRaisesRegexp(
+        ValueError, r'dtype.*and vocabulary dtype.*do not match'):
+      fc.categorical_column_with_vocabulary_list(
+          key='aaa', vocabulary_list=('omar', 'stringer', 'marlo'),
+          dtype=dtypes.int32)
+
+  def test_mismatched_string_dtype(self):
+    with self.assertRaisesRegexp(
+        ValueError, r'dtype.*and vocabulary dtype.*do not match'):
+      fc.categorical_column_with_vocabulary_list(
+          key='aaa', vocabulary_list=(12, 24, 36), dtype=dtypes.string)
+
+  def test_none_mapping(self):
+    with self.assertRaisesRegexp(
+        ValueError, r'vocabulary_list.*must be non-empty'):
+      fc.categorical_column_with_vocabulary_list(
+          key='aaa', vocabulary_list=None)
+
+  def test_empty_mapping(self):
+    with self.assertRaisesRegexp(
+        ValueError, r'vocabulary_list.*must be non-empty'):
+      fc.categorical_column_with_vocabulary_list(
+          key='aaa', vocabulary_list=tuple([]))
+
+  def test_duplicate_mapping(self):
+    with self.assertRaisesRegexp(ValueError, 'Duplicate keys'):
+      fc.categorical_column_with_vocabulary_list(
+          key='aaa', vocabulary_list=(12, 24, 12))
+
+  def test_invalid_num_oov_buckets(self):
+    with self.assertRaisesRegexp(ValueError, 'Invalid num_oov_buckets'):
+      fc.categorical_column_with_vocabulary_list(
+          key='aaa', vocabulary_list=(12, 24, 36),
+          num_oov_buckets=-1)
+
+  def test_invalid_buckets_and_default_value(self):
+    with self.assertRaisesRegexp(
+        ValueError, 'both num_oov_buckets and default_value'):
+      fc.categorical_column_with_vocabulary_list(
+          key='aaa',
+          vocabulary_list=(12, 24, 36),
+          num_oov_buckets=100,
+          default_value=2)
+
+  def test_invalid_input_dtype_int32(self):
+    column = fc.categorical_column_with_vocabulary_list(
+        key='aaa',
+        vocabulary_list=('omar', 'stringer', 'marlo'))
+    inputs = sparse_tensor.SparseTensorValue(
+        indices=((0, 0), (1, 0), (1, 1)),
+        values=(12, 24, 36),
+        dense_shape=(2, 2))
+    with self.assertRaisesRegexp(ValueError, 'dtype must be compatible'):
+      column.get_sparse_tensors(
+          FeatureTransformationCache({
+              'aaa': inputs
+          }), None)
+
+  def test_invalid_input_dtype_string(self):
+    column = fc.categorical_column_with_vocabulary_list(
+        key='aaa',
+        vocabulary_list=(12, 24, 36))
+    inputs = sparse_tensor.SparseTensorValue(
+        indices=((0, 0), (1, 0), (1, 1)),
+        values=('omar', 'stringer', 'marlo'),
+        dense_shape=(2, 2))
+    with self.assertRaisesRegexp(ValueError, 'dtype must be compatible'):
+      column.get_sparse_tensors(
+          FeatureTransformationCache({
+              'aaa': inputs
+          }), None)
+
+  def test_parse_example_string(self):
+    a = fc.categorical_column_with_vocabulary_list(
+        key='aaa', vocabulary_list=('omar', 'stringer', 'marlo'))
+    data = example_pb2.Example(features=feature_pb2.Features(
+        feature={
+            'aaa':
+                feature_pb2.Feature(bytes_list=feature_pb2.BytesList(
+                    value=[b'omar', b'stringer']))
+        }))
+    features = parsing_ops.parse_example(
+        serialized=[data.SerializeToString()],
+        features=fc.make_parse_example_spec([a]))
+    self.assertIn('aaa', features)
+    with self.test_session():
+      _assert_sparse_tensor_value(
+          self,
+          sparse_tensor.SparseTensorValue(
+              indices=[[0, 0], [0, 1]],
+              values=np.array([b'omar', b'stringer'], dtype=np.object_),
+              dense_shape=[1, 2]),
+          features['aaa'].eval())
+
+  def test_parse_example_int(self):
+    a = fc.categorical_column_with_vocabulary_list(
+        key='aaa', vocabulary_list=(11, 21, 31))
+    data = example_pb2.Example(features=feature_pb2.Features(
+        feature={
+            'aaa':
+                feature_pb2.Feature(int64_list=feature_pb2.Int64List(
+                    value=[11, 21]))
+        }))
+    features = parsing_ops.parse_example(
+        serialized=[data.SerializeToString()],
+        features=fc.make_parse_example_spec([a]))
+    self.assertIn('aaa', features)
+    with self.test_session():
+      _assert_sparse_tensor_value(
+          self,
+          sparse_tensor.SparseTensorValue(
+              indices=[[0, 0], [0, 1]],
+              values=[11, 21],
+              dense_shape=[1, 2]),
+          features['aaa'].eval())
+
+  def test_get_sparse_tensors(self):
+    column = fc.categorical_column_with_vocabulary_list(
+        key='aaa',
+        vocabulary_list=('omar', 'stringer', 'marlo'))
+    inputs = sparse_tensor.SparseTensorValue(
+        indices=((0, 0), (1, 0), (1, 1)),
+        values=('marlo', 'skywalker', 'omar'),
+        dense_shape=(2, 2))
+    id_weight_pair = column.get_sparse_tensors(
+        FeatureTransformationCache({
+            'aaa': inputs
+        }), None)
+    self.assertIsNone(id_weight_pair.weight_tensor)
+    with _initialized_session():
+      _assert_sparse_tensor_value(
+          self,
+          sparse_tensor.SparseTensorValue(
+              indices=inputs.indices,
+              values=np.array((2, -1, 0), dtype=np.int64),
+              dense_shape=inputs.dense_shape),
+          id_weight_pair.id_tensor.eval())
+
+  def test_transform_feature(self):
+    column = fc.categorical_column_with_vocabulary_list(
+        key='aaa',
+        vocabulary_list=('omar', 'stringer', 'marlo'))
+    inputs = sparse_tensor.SparseTensorValue(
+        indices=((0, 0), (1, 0), (1, 1)),
+        values=('marlo', 'skywalker', 'omar'),
+        dense_shape=(2, 2))
+    id_tensor = _transform_features({'aaa': inputs}, [column], None)[column]
+    with _initialized_session():
+      _assert_sparse_tensor_value(
+          self,
+          sparse_tensor.SparseTensorValue(
+              indices=inputs.indices,
+              values=np.array((2, -1, 0), dtype=np.int64),
+              dense_shape=inputs.dense_shape),
+          id_tensor.eval())
+
+  def DISABLED_test_get_sparse_tensors_weight_collections(self):
+    column = fc.categorical_column_with_vocabulary_list(
+        key='aaa',
+        vocabulary_list=('omar', 'stringer', 'marlo'))
+    inputs = sparse_tensor.SparseTensor(
+        values=['omar', 'stringer', 'marlo'],
+        indices=[[0, 0], [1, 0], [1, 1]],
+        dense_shape=[2, 2])
+    column.get_sparse_tensors(
+        FeatureTransformationCache({
+            'aaa': inputs
+        }),
+        weight_collections=('my_weights',))
+
+    self.assertItemsEqual(
+        [], ops.get_collection(ops.GraphKeys.GLOBAL_VARIABLES))
+    self.assertItemsEqual([], ops.get_collection('my_weights'))
+
+  def test_get_sparse_tensors_dense_input(self):
+    column = fc.categorical_column_with_vocabulary_list(
+        key='aaa',
+        vocabulary_list=('omar', 'stringer', 'marlo'))
+    id_weight_pair = column.get_sparse_tensors(
+        FeatureTransformationCache({
+            'aaa': (('marlo', ''), ('skywalker', 'omar'))
+        }), None)
+    self.assertIsNone(id_weight_pair.weight_tensor)
+    with _initialized_session():
+      _assert_sparse_tensor_value(
+          self,
+          sparse_tensor.SparseTensorValue(
+              indices=((0, 0), (1, 0), (1, 1)),
+              values=np.array((2, -1, 0), dtype=np.int64),
+              dense_shape=(2, 2)),
+          id_weight_pair.id_tensor.eval())
+
+  def test_get_sparse_tensors_default_value_in_vocabulary(self):
+    column = fc.categorical_column_with_vocabulary_list(
+        key='aaa',
+        vocabulary_list=('omar', 'stringer', 'marlo'),
+        default_value=2)
+    inputs = sparse_tensor.SparseTensorValue(
+        indices=((0, 0), (1, 0), (1, 1)),
+        values=('marlo', 'skywalker', 'omar'),
+        dense_shape=(2, 2))
+    id_weight_pair = column.get_sparse_tensors(
+        FeatureTransformationCache({
+            'aaa': inputs
+        }), None)
+    self.assertIsNone(id_weight_pair.weight_tensor)
+    with _initialized_session():
+      _assert_sparse_tensor_value(
+          self,
+          sparse_tensor.SparseTensorValue(
+              indices=inputs.indices,
+              values=np.array((2, 2, 0), dtype=np.int64),
+              dense_shape=inputs.dense_shape),
+          id_weight_pair.id_tensor.eval())
+
+  def test_get_sparse_tensors_with_oov_buckets(self):
+    column = fc.categorical_column_with_vocabulary_list(
+        key='aaa',
+        vocabulary_list=('omar', 'stringer', 'marlo'),
+        num_oov_buckets=100)
+    inputs = sparse_tensor.SparseTensorValue(
+        indices=((0, 0), (1, 0), (1, 1), (1, 2)),
+        values=('marlo', 'skywalker', 'omar', 'heisenberg'),
+        dense_shape=(2, 3))
+    id_weight_pair = column.get_sparse_tensors(
+        FeatureTransformationCache({
+            'aaa': inputs
+        }), None)
+    self.assertIsNone(id_weight_pair.weight_tensor)
+    with _initialized_session():
+      _assert_sparse_tensor_value(
+          self,
+          sparse_tensor.SparseTensorValue(
+              indices=inputs.indices,
+              values=np.array((2, 33, 0, 62), dtype=np.int64),
+              dense_shape=inputs.dense_shape),
+          id_weight_pair.id_tensor.eval())
+
+  def test_get_sparse_tensors_int32(self):
+    column = fc.categorical_column_with_vocabulary_list(
+        key='aaa',
+        vocabulary_list=np.array((30, 35, 11, 23, 22), dtype=np.int32),
+        dtype=dtypes.int32)
+    inputs = sparse_tensor.SparseTensorValue(
+        indices=((0, 0), (1, 0), (1, 1), (2, 2)),
+        values=np.array((11, 100, 30, 22), dtype=np.int32),
+        dense_shape=(3, 3))
+    id_weight_pair = column.get_sparse_tensors(
+        FeatureTransformationCache({
+            'aaa': inputs
+        }), None)
+    self.assertIsNone(id_weight_pair.weight_tensor)
+    with _initialized_session():
+      _assert_sparse_tensor_value(
+          self,
+          sparse_tensor.SparseTensorValue(
+              indices=inputs.indices,
+              values=np.array((2, -1, 0, 4), dtype=np.int64),
+              dense_shape=inputs.dense_shape),
+          id_weight_pair.id_tensor.eval())
+
+  def test_get_sparse_tensors_int32_dense_input(self):
+    default_value = -100
+    column = fc.categorical_column_with_vocabulary_list(
+        key='aaa',
+        vocabulary_list=np.array((30, 35, 11, 23, 22), dtype=np.int32),
+        dtype=dtypes.int32,
+        default_value=default_value)
+    id_weight_pair = column.get_sparse_tensors(
+        FeatureTransformationCache({
+            'aaa':
+                np.array(
+                    ((11, -1, -1), (100, 30, -1), (-1, -1, 22)), dtype=np.int32)
+        }), None)
+    self.assertIsNone(id_weight_pair.weight_tensor)
+    with _initialized_session():
+      _assert_sparse_tensor_value(
+          self,
+          sparse_tensor.SparseTensorValue(
+              indices=((0, 0), (1, 0), (1, 1), (2, 2)),
+              values=np.array((2, default_value, 0, 4), dtype=np.int64),
+              dense_shape=(3, 3)),
+          id_weight_pair.id_tensor.eval())
+
+  def test_get_sparse_tensors_int32_with_oov_buckets(self):
+    column = fc.categorical_column_with_vocabulary_list(
+        key='aaa',
+        vocabulary_list=np.array((30, 35, 11, 23, 22), dtype=np.int32),
+        dtype=dtypes.int32,
+        num_oov_buckets=100)
+    inputs = sparse_tensor.SparseTensorValue(
+        indices=((0, 0), (1, 0), (1, 1), (2, 2)),
+        values=(11, 100, 30, 22),
+        dense_shape=(3, 3))
+    id_weight_pair = column.get_sparse_tensors(
+        FeatureTransformationCache({
+            'aaa': inputs
+        }), None)
+    self.assertIsNone(id_weight_pair.weight_tensor)
+    with _initialized_session():
+      _assert_sparse_tensor_value(
+          self,
+          sparse_tensor.SparseTensorValue(
+              indices=inputs.indices,
+              values=np.array((2, 60, 0, 4), dtype=np.int64),
+              dense_shape=inputs.dense_shape),
+          id_weight_pair.id_tensor.eval())
+
+  def test_linear_model(self):
+    wire_column = fc_old.categorical_column_with_vocabulary_list(
+        key='aaa',
+        vocabulary_list=('omar', 'stringer', 'marlo'),
+        num_oov_buckets=1)
+    self.assertEqual(4, wire_column._num_buckets)
+    with ops.Graph().as_default():
+      predictions = fc.linear_model({
+          wire_column.name: sparse_tensor.SparseTensorValue(
+              indices=((0, 0), (1, 0), (1, 1)),
+              values=('marlo', 'skywalker', 'omar'),
+              dense_shape=(2, 2))
+      }, (wire_column,))
+      bias = get_linear_model_bias()
+      wire_var = get_linear_model_column_var(wire_column)
+      with _initialized_session():
+        self.assertAllClose((0.,), bias.eval())
+        self.assertAllClose(((0.,), (0.,), (0.,), (0.,)), wire_var.eval())
+        self.assertAllClose(((0.,), (0.,)), predictions.eval())
+        wire_var.assign(((1.,), (2.,), (3.,), (4.,))).eval()
+        # 'marlo' -> 2: wire_var[2] = 3
+        # 'skywalker' -> 3, 'omar' -> 0: wire_var[3] + wire_var[0] = 4+1 = 5
+        self.assertAllClose(((3.,), (5.,)), predictions.eval())
+
+  def test_keras_linear_model(self):
+    wire_column = fc_old.categorical_column_with_vocabulary_list(
+        key='aaa',
+        vocabulary_list=('omar', 'stringer', 'marlo'),
+        num_oov_buckets=1)
+    self.assertEqual(4, wire_column._num_buckets)
+    with ops.Graph().as_default():
+      predictions = get_keras_linear_model_predictions({
+          wire_column.name:
+              sparse_tensor.SparseTensorValue(
+                  indices=((0, 0), (1, 0), (1, 1)),
+                  values=('marlo', 'skywalker', 'omar'),
+                  dense_shape=(2, 2))
+      }, (wire_column,))
+      bias = get_linear_model_bias()
+      wire_var = get_linear_model_column_var(wire_column)
+      with _initialized_session():
+        self.assertAllClose((0.,), bias.eval())
+        self.assertAllClose(((0.,), (0.,), (0.,), (0.,)), wire_var.eval())
+        self.assertAllClose(((0.,), (0.,)), predictions.eval())
+        wire_var.assign(((1.,), (2.,), (3.,), (4.,))).eval()
+        # 'marlo' -> 2: wire_var[2] = 3
+        # 'skywalker' -> 3, 'omar' -> 0: wire_var[3] + wire_var[0] = 4+1 = 5
+        self.assertAllClose(((3.,), (5.,)), predictions.eval())
+
+
+class IdentityCategoricalColumnTest(test.TestCase):
+
+  def test_constructor(self):
+    column = fc.categorical_column_with_identity(key='aaa', num_buckets=3)
+    self.assertEqual('aaa', column.name)
+    self.assertEqual('aaa', column.key)
+    self.assertEqual(3, column.num_buckets)
+    self.assertEqual({
+        'aaa': parsing_ops.VarLenFeature(dtypes.int64)
+    }, column.parse_example_spec)
+
+  def test_key_should_be_string(self):
+    with self.assertRaisesRegexp(ValueError, 'key must be a string.'):
+      fc.categorical_column_with_identity(key=('aaa',), num_buckets=3)
+
+  def test_deep_copy(self):
+    original = fc.categorical_column_with_identity(key='aaa', num_buckets=3)
+    for column in (original, copy.deepcopy(original)):
+      self.assertEqual('aaa', column.name)
+      self.assertEqual(3, column.num_buckets)
+      self.assertEqual({
+          'aaa': parsing_ops.VarLenFeature(dtypes.int64)
+      }, column.parse_example_spec)
+
+  def test_invalid_num_buckets_zero(self):
+    with self.assertRaisesRegexp(ValueError, 'num_buckets 0 < 1'):
+      fc.categorical_column_with_identity(key='aaa', num_buckets=0)
+
+  def test_invalid_num_buckets_negative(self):
+    with self.assertRaisesRegexp(ValueError, 'num_buckets -1 < 1'):
+      fc.categorical_column_with_identity(key='aaa', num_buckets=-1)
+
+  def test_invalid_default_value_too_small(self):
+    with self.assertRaisesRegexp(ValueError, 'default_value -1 not in range'):
+      fc.categorical_column_with_identity(
+          key='aaa', num_buckets=3, default_value=-1)
+
+  def test_invalid_default_value_too_big(self):
+    with self.assertRaisesRegexp(ValueError, 'default_value 3 not in range'):
+      fc.categorical_column_with_identity(
+          key='aaa', num_buckets=3, default_value=3)
+
+  def test_invalid_input_dtype(self):
+    column = fc.categorical_column_with_identity(key='aaa', num_buckets=3)
+    inputs = sparse_tensor.SparseTensorValue(
+        indices=((0, 0), (1, 0), (1, 1)),
+        values=('omar', 'stringer', 'marlo'),
+        dense_shape=(2, 2))
+    with self.assertRaisesRegexp(ValueError, 'Invalid input, not integer'):
+      column.get_sparse_tensors(
+          FeatureTransformationCache({
+              'aaa': inputs
+          }), None)
+
+  def test_parse_example(self):
+    a = fc.categorical_column_with_identity(key='aaa', num_buckets=30)
+    data = example_pb2.Example(features=feature_pb2.Features(
+        feature={
+            'aaa':
+                feature_pb2.Feature(int64_list=feature_pb2.Int64List(
+                    value=[11, 21]))
+        }))
+    features = parsing_ops.parse_example(
+        serialized=[data.SerializeToString()],
+        features=fc.make_parse_example_spec([a]))
+    self.assertIn('aaa', features)
+    with self.test_session():
+      _assert_sparse_tensor_value(
+          self,
+          sparse_tensor.SparseTensorValue(
+              indices=[[0, 0], [0, 1]],
+              values=np.array([11, 21], dtype=np.int64),
+              dense_shape=[1, 2]),
+          features['aaa'].eval())
+
+  def test_get_sparse_tensors(self):
+    column = fc.categorical_column_with_identity(key='aaa', num_buckets=3)
+    inputs = sparse_tensor.SparseTensorValue(
+        indices=((0, 0), (1, 0), (1, 1)),
+        values=(0, 1, 0),
+        dense_shape=(2, 2))
+    id_weight_pair = column.get_sparse_tensors(
+        FeatureTransformationCache({
+            'aaa': inputs
+        }), None)
+    self.assertIsNone(id_weight_pair.weight_tensor)
+    with _initialized_session():
+      _assert_sparse_tensor_value(
+          self,
+          sparse_tensor.SparseTensorValue(
+              indices=inputs.indices,
+              values=np.array((0, 1, 0), dtype=np.int64),
+              dense_shape=inputs.dense_shape),
+          id_weight_pair.id_tensor.eval())
+
+  def test_transform_feature(self):
+    column = fc.categorical_column_with_identity(key='aaa', num_buckets=3)
+    inputs = sparse_tensor.SparseTensorValue(
+        indices=((0, 0), (1, 0), (1, 1)),
+        values=(0, 1, 0),
+        dense_shape=(2, 2))
+    id_tensor = _transform_features({'aaa': inputs}, [column], None)[column]
+    with _initialized_session():
+      _assert_sparse_tensor_value(
+          self,
+          sparse_tensor.SparseTensorValue(
+              indices=inputs.indices,
+              values=np.array((0, 1, 0), dtype=np.int64),
+              dense_shape=inputs.dense_shape),
+          id_tensor.eval())
+
+  def DISABLED_test_get_sparse_tensors_weight_collections(self):
+    column = fc.categorical_column_with_identity(key='aaa', num_buckets=3)
+    inputs = sparse_tensor.SparseTensorValue(
+        indices=((0, 0), (1, 0), (1, 1)),
+        values=(0, 1, 0),
+        dense_shape=(2, 2))
+    column.get_sparse_tensors(
+        FeatureTransformationCache({
+            'aaa': inputs
+        }),
+        weight_collections=('my_weights',))
+
+    self.assertItemsEqual(
+        [], ops.get_collection(ops.GraphKeys.GLOBAL_VARIABLES))
+    self.assertItemsEqual([], ops.get_collection('my_weights'))
+
+  def test_get_sparse_tensors_dense_input(self):
+    column = fc.categorical_column_with_identity(key='aaa', num_buckets=3)
+    id_weight_pair = column.get_sparse_tensors(
+        FeatureTransformationCache({
+            'aaa': ((0, -1), (1, 0))
+        }), None)
+    self.assertIsNone(id_weight_pair.weight_tensor)
+    with _initialized_session():
+      _assert_sparse_tensor_value(
+          self,
+          sparse_tensor.SparseTensorValue(
+              indices=((0, 0), (1, 0), (1, 1)),
+              values=np.array((0, 1, 0), dtype=np.int64),
+              dense_shape=(2, 2)),
+          id_weight_pair.id_tensor.eval())
+
+  def test_get_sparse_tensors_with_inputs_too_small(self):
+    column = fc.categorical_column_with_identity(key='aaa', num_buckets=3)
+    inputs = sparse_tensor.SparseTensorValue(
+        indices=((0, 0), (1, 0), (1, 1)),
+        values=(1, -1, 0),
+        dense_shape=(2, 2))
+    id_weight_pair = column.get_sparse_tensors(
+        FeatureTransformationCache({
+            'aaa': inputs
+        }), None)
+    self.assertIsNone(id_weight_pair.weight_tensor)
+    with _initialized_session():
+      with self.assertRaisesRegexp(
+          errors.OpError, 'assert_greater_or_equal_0'):
+        id_weight_pair.id_tensor.eval()
+
+  def test_get_sparse_tensors_with_inputs_too_big(self):
+    column = fc.categorical_column_with_identity(key='aaa', num_buckets=3)
+    inputs = sparse_tensor.SparseTensorValue(
+        indices=((0, 0), (1, 0), (1, 1)),
+        values=(1, 99, 0),
+        dense_shape=(2, 2))
+    id_weight_pair = column.get_sparse_tensors(
+        FeatureTransformationCache({
+            'aaa': inputs
+        }), None)
+    self.assertIsNone(id_weight_pair.weight_tensor)
+    with _initialized_session():
+      with self.assertRaisesRegexp(
+          errors.OpError, 'assert_less_than_num_buckets'):
+        id_weight_pair.id_tensor.eval()
+
+  def test_get_sparse_tensors_with_default_value(self):
+    column = fc.categorical_column_with_identity(
+        key='aaa', num_buckets=4, default_value=3)
+    inputs = sparse_tensor.SparseTensorValue(
+        indices=((0, 0), (1, 0), (1, 1)),
+        values=(1, -1, 99),
+        dense_shape=(2, 2))
+    id_weight_pair = column.get_sparse_tensors(
+        FeatureTransformationCache({
+            'aaa': inputs
+        }), None)
+    self.assertIsNone(id_weight_pair.weight_tensor)
+    with _initialized_session():
+      _assert_sparse_tensor_value(
+          self,
+          sparse_tensor.SparseTensorValue(
+              indices=inputs.indices,
+              values=np.array((1, 3, 3), dtype=np.int64),
+              dense_shape=inputs.dense_shape),
+          id_weight_pair.id_tensor.eval())
+
+  def test_get_sparse_tensors_with_default_value_and_placeholder_inputs(self):
+    column = fc.categorical_column_with_identity(
+        key='aaa', num_buckets=4, default_value=3)
+    input_indices = array_ops.placeholder(dtype=dtypes.int64)
+    input_values = array_ops.placeholder(dtype=dtypes.int32)
+    input_shape = array_ops.placeholder(dtype=dtypes.int64)
+    inputs = sparse_tensor.SparseTensorValue(
+        indices=input_indices,
+        values=input_values,
+        dense_shape=input_shape)
+    id_weight_pair = column.get_sparse_tensors(
+        FeatureTransformationCache({
+            'aaa': inputs
+        }), None)
+    self.assertIsNone(id_weight_pair.weight_tensor)
+    with _initialized_session():
+      _assert_sparse_tensor_value(
+          self,
+          sparse_tensor.SparseTensorValue(
+              indices=np.array(((0, 0), (1, 0), (1, 1)), dtype=np.int64),
+              values=np.array((1, 3, 3), dtype=np.int64),
+              dense_shape=np.array((2, 2), dtype=np.int64)),
+          id_weight_pair.id_tensor.eval(feed_dict={
+              input_indices: ((0, 0), (1, 0), (1, 1)),
+              input_values: (1, -1, 99),
+              input_shape: (2, 2),
+          }))
+
+  def test_linear_model(self):
+    column = fc_old.categorical_column_with_identity(key='aaa', num_buckets=3)
+    self.assertEqual(3, column.num_buckets)
+    with ops.Graph().as_default():
+      predictions = fc.linear_model({
+          column.name: sparse_tensor.SparseTensorValue(
+              indices=((0, 0), (1, 0), (1, 1)),
+              values=(0, 2, 1),
+              dense_shape=(2, 2))
+      }, (column,))
+      bias = get_linear_model_bias()
+      weight_var = get_linear_model_column_var(column)
+      with _initialized_session():
+        self.assertAllClose((0.,), bias.eval())
+        self.assertAllClose(((0.,), (0.,), (0.,)), weight_var.eval())
+        self.assertAllClose(((0.,), (0.,)), predictions.eval())
+        weight_var.assign(((1.,), (2.,), (3.,))).eval()
+        # weight_var[0] = 1
+        # weight_var[2] + weight_var[1] = 3+2 = 5
+        self.assertAllClose(((1.,), (5.,)), predictions.eval())
+
+  def test_keras_linear_model(self):
+    column = fc_old.categorical_column_with_identity(key='aaa', num_buckets=3)
+    self.assertEqual(3, column.num_buckets)
+    with ops.Graph().as_default():
+      predictions = get_keras_linear_model_predictions({
+          column.name:
+              sparse_tensor.SparseTensorValue(
+                  indices=((0, 0), (1, 0), (1, 1)),
+                  values=(0, 2, 1),
+                  dense_shape=(2, 2))
+      }, (column,))
+      bias = get_linear_model_bias()
+      weight_var = get_linear_model_column_var(column)
+      with _initialized_session():
+        self.assertAllClose((0.,), bias.eval())
+        self.assertAllClose(((0.,), (0.,), (0.,)), weight_var.eval())
+        self.assertAllClose(((0.,), (0.,)), predictions.eval())
+        weight_var.assign(((1.,), (2.,), (3.,))).eval()
+        # weight_var[0] = 1
+        # weight_var[2] + weight_var[1] = 3+2 = 5
+        self.assertAllClose(((1.,), (5.,)), predictions.eval())
+
+
+class TransformFeaturesTest(test.TestCase):
+
+  # All transform tests are distributed in column test.
+  # Here we only test multi column case and naming
+  def transform_multi_column(self):
+    bucketized_price = fc.bucketized_column(
+        fc.numeric_column('price'), boundaries=[0, 2, 4, 6])
+    hashed_sparse = fc.categorical_column_with_hash_bucket('wire', 10)
+    with ops.Graph().as_default():
+      features = {
+          'price': [[-1.], [5.]],
+          'wire':
+              sparse_tensor.SparseTensor(
+                  values=['omar', 'stringer', 'marlo'],
+                  indices=[[0, 0], [1, 0], [1, 1]],
+                  dense_shape=[2, 2])
+      }
+      transformed = _transform_features(features,
+                                        [bucketized_price, hashed_sparse], None)
+      with _initialized_session():
+        self.assertIn(bucketized_price.name, transformed[bucketized_price].name)
+        self.assertAllEqual([[0], [3]], transformed[bucketized_price].eval())
+        self.assertIn(hashed_sparse.name, transformed[hashed_sparse].name)
+        self.assertAllEqual([6, 4, 1], transformed[hashed_sparse].values.eval())
+
+  def test_column_order(self):
+    """When the column is both dense and sparse, uses sparse tensors."""
+
+    class _LoggerColumn(FeatureColumn):
+
+      def __init__(self, name):
+        self._name = name
+
+      @property
+      def name(self):
+        return self._name
+
+      def transform_feature(self, transformation_cache, state_manager):
+        self.call_order = call_logger['count']
+        call_logger['count'] += 1
+        return 'Anything'
+
+      @property
+      def parse_example_spec(self):
+        pass
+
+    with ops.Graph().as_default():
+      column1 = _LoggerColumn('1')
+      column2 = _LoggerColumn('2')
+      call_logger = {'count': 0}
+      _transform_features({}, [column1, column2], None)
+      self.assertEqual(0, column1.call_order)
+      self.assertEqual(1, column2.call_order)
+
+      call_logger = {'count': 0}
+      _transform_features({}, [column2, column1], None)
+      self.assertEqual(0, column1.call_order)
+      self.assertEqual(1, column2.call_order)
+
+
+class IndicatorColumnTest(test.TestCase):
+
+  def test_indicator_column(self):
+    a = fc.categorical_column_with_hash_bucket('a', 4)
+    indicator_a = fc.indicator_column(a)
+    self.assertEqual(indicator_a.categorical_column.name, 'a')
+    self.assertEqual(indicator_a.name, 'a_indicator')
+    self.assertEqual(indicator_a.variable_shape, [1, 4])
+
+    b = fc.categorical_column_with_hash_bucket('b', hash_bucket_size=100)
+    indicator_b = fc.indicator_column(b)
+    self.assertEqual(indicator_b.categorical_column.name, 'b')
+    self.assertEqual(indicator_b.name, 'b_indicator')
+    self.assertEqual(indicator_b.variable_shape, [1, 100])
+
+  def test_1D_shape_succeeds(self):
+    animal = fc.indicator_column(
+        fc.categorical_column_with_hash_bucket('animal', 4))
+    transformation_cache = FeatureTransformationCache({
+        'animal': ['fox', 'fox']
+    })
+    output = transformation_cache.get(animal, None)
+    with self.test_session():
+      self.assertAllEqual([[0., 0., 1., 0.], [0., 0., 1., 0.]], output.eval())
+
+  def test_2D_shape_succeeds(self):
+    # TODO(ispir/cassandrax): Swith to categorical_column_with_keys when ready.
+    animal = fc.indicator_column(
+        fc.categorical_column_with_hash_bucket('animal', 4))
+    transformation_cache = FeatureTransformationCache({
+        'animal':
+            sparse_tensor.SparseTensor(
+                indices=[[0, 0], [1, 0]],
+                values=['fox', 'fox'],
+                dense_shape=[2, 1])
+    })
+    output = transformation_cache.get(animal, None)
+    with self.test_session():
+      self.assertAllEqual([[0., 0., 1., 0.], [0., 0., 1., 0.]], output.eval())
+
+  def test_multi_hot(self):
+    animal = fc.indicator_column(
+        fc.categorical_column_with_identity('animal', num_buckets=4))
+
+    transformation_cache = FeatureTransformationCache({
+        'animal':
+            sparse_tensor.SparseTensor(
+                indices=[[0, 0], [0, 1]], values=[1, 1], dense_shape=[1, 2])
+    })
+    output = transformation_cache.get(animal, None)
+    with self.test_session():
+      self.assertAllEqual([[0., 2., 0., 0.]], output.eval())
+
+  def test_multi_hot2(self):
+    animal = fc.indicator_column(
+        fc.categorical_column_with_identity('animal', num_buckets=4))
+    transformation_cache = FeatureTransformationCache({
+        'animal':
+            sparse_tensor.SparseTensor(
+                indices=[[0, 0], [0, 1]], values=[1, 2], dense_shape=[1, 2])
+    })
+    output = transformation_cache.get(animal, None)
+    with self.test_session():
+      self.assertAllEqual([[0., 1., 1., 0.]], output.eval())
+
+  def test_deep_copy(self):
+    a = fc.categorical_column_with_hash_bucket('a', 4)
+    column = fc.indicator_column(a)
+    column_copy = copy.deepcopy(column)
+    self.assertEqual(column_copy.categorical_column.name, 'a')
+    self.assertEqual(column.name, 'a_indicator')
+    self.assertEqual(column.variable_shape, [1, 4])
+
+  def test_parse_example(self):
+    a = fc.categorical_column_with_vocabulary_list(
+        key='aaa', vocabulary_list=('omar', 'stringer', 'marlo'))
+    a_indicator = fc.indicator_column(a)
+    data = example_pb2.Example(features=feature_pb2.Features(
+        feature={
+            'aaa':
+                feature_pb2.Feature(bytes_list=feature_pb2.BytesList(
+                    value=[b'omar', b'stringer']))
+        }))
+    features = parsing_ops.parse_example(
+        serialized=[data.SerializeToString()],
+        features=fc.make_parse_example_spec([a_indicator]))
+    self.assertIn('aaa', features)
+    with self.test_session():
+      _assert_sparse_tensor_value(
+          self,
+          sparse_tensor.SparseTensorValue(
+              indices=[[0, 0], [0, 1]],
+              values=np.array([b'omar', b'stringer'], dtype=np.object_),
+              dense_shape=[1, 2]),
+          features['aaa'].eval())
+
+  def test_transform(self):
+    a = fc.categorical_column_with_vocabulary_list(
+        key='aaa', vocabulary_list=('omar', 'stringer', 'marlo'))
+    a_indicator = fc.indicator_column(a)
+    features = {
+        'aaa': sparse_tensor.SparseTensorValue(
+            indices=((0, 0), (1, 0), (1, 1)),
+            values=('marlo', 'skywalker', 'omar'),
+            dense_shape=(2, 2))
+    }
+    indicator_tensor = _transform_features(features, [a_indicator],
+                                           None)[a_indicator]
+    with _initialized_session():
+      self.assertAllEqual([[0, 0, 1], [1, 0, 0]], indicator_tensor.eval())
+
+  def test_transform_with_weighted_column(self):
+    # Github issue 12557
+    ids = fc.categorical_column_with_vocabulary_list(
+        key='ids', vocabulary_list=('a', 'b', 'c'))
+    weights = fc.weighted_categorical_column(ids, 'weights')
+    indicator = fc.indicator_column(weights)
+    features = {
+        'ids': constant_op.constant([['c', 'b', 'a']]),
+        'weights': constant_op.constant([[2., 4., 6.]])
+    }
+    indicator_tensor = _transform_features(features, [indicator],
+                                           None)[indicator]
+    with _initialized_session():
+      self.assertAllEqual([[6., 4., 2.]], indicator_tensor.eval())
+
+  def test_transform_with_missing_value_in_weighted_column(self):
+    # Github issue 12583
+    ids = fc.categorical_column_with_vocabulary_list(
+        key='ids', vocabulary_list=('a', 'b', 'c'))
+    weights = fc.weighted_categorical_column(ids, 'weights')
+    indicator = fc.indicator_column(weights)
+    features = {
+        'ids': constant_op.constant([['c', 'b', 'unknown']]),
+        'weights': constant_op.constant([[2., 4., 6.]])
+    }
+    indicator_tensor = _transform_features(features, [indicator],
+                                           None)[indicator]
+    with _initialized_session():
+      self.assertAllEqual([[0., 4., 2.]], indicator_tensor.eval())
+
+  def test_transform_with_missing_value_in_categorical_column(self):
+    # Github issue 12583
+    ids = fc.categorical_column_with_vocabulary_list(
+        key='ids', vocabulary_list=('a', 'b', 'c'))
+    indicator = fc.indicator_column(ids)
+    features = {
+        'ids': constant_op.constant([['c', 'b', 'unknown']]),
+    }
+    indicator_tensor = _transform_features(features, [indicator],
+                                           None)[indicator]
+    with _initialized_session():
+      self.assertAllEqual([[0., 1., 1.]], indicator_tensor.eval())
+
+  def test_linear_model(self):
+    animal = fc_old.indicator_column(
+        fc_old.categorical_column_with_identity('animal', num_buckets=4))
+    with ops.Graph().as_default():
+      features = {
+          'animal':
+              sparse_tensor.SparseTensor(
+                  indices=[[0, 0], [0, 1]], values=[1, 2], dense_shape=[1, 2])
+      }
+
+      predictions = fc.linear_model(features, [animal])
+      weight_var = get_linear_model_column_var(animal)
+      with _initialized_session():
+        # All should be zero-initialized.
+        self.assertAllClose([[0.], [0.], [0.], [0.]], weight_var.eval())
+        self.assertAllClose([[0.]], predictions.eval())
+        weight_var.assign([[1.], [2.], [3.], [4.]]).eval()
+        self.assertAllClose([[2. + 3.]], predictions.eval())
+
+  def test_keras_linear_model(self):
+    animal = fc_old.indicator_column(
+        fc_old.categorical_column_with_identity('animal', num_buckets=4))
+    with ops.Graph().as_default():
+      features = {
+          'animal':
+              sparse_tensor.SparseTensor(
+                  indices=[[0, 0], [0, 1]], values=[1, 2], dense_shape=[1, 2])
+      }
+
+      predictions = get_keras_linear_model_predictions(features, [animal])
+      weight_var = get_linear_model_column_var(animal)
+      with _initialized_session():
+        # All should be zero-initialized.
+        self.assertAllClose([[0.], [0.], [0.], [0.]], weight_var.eval())
+        self.assertAllClose([[0.]], predictions.eval())
+        weight_var.assign([[1.], [2.], [3.], [4.]]).eval()
+        self.assertAllClose([[2. + 3.]], predictions.eval())
+
+  def test_input_layer(self):
+    animal = fc_old.indicator_column(
+        fc_old.categorical_column_with_identity('animal', num_buckets=4))
+    with ops.Graph().as_default():
+      features = {
+          'animal':
+              sparse_tensor.SparseTensor(
+                  indices=[[0, 0], [0, 1]], values=[1, 2], dense_shape=[1, 2])
+      }
+      net = fc.input_layer(features, [animal])
+      with _initialized_session():
+        self.assertAllClose([[0., 1., 1., 0.]], net.eval())
+
+
+class _TestStateManager(StateManager):
+
+  def __init__(self, trainable=True):
+    # Dict of feature_column to a dict of variables.
+    self._all_variables = {}
+    self._trainable = trainable
+
+  def get_variable(self,
+                   feature_column,
+                   name,
+                   shape,
+                   dtype=None,
+                   initializer=None):
+    if feature_column not in self._all_variables:
+      self._all_variables[feature_column] = {}
+    var_dict = self._all_variables[feature_column]
+    if name in var_dict:
+      return var_dict[name]
+    else:
+      var = variable_scope.get_variable(
+          name=name,
+          shape=shape,
+          initializer=initializer,
+          trainable=self._trainable)
+      var_dict[name] = var
+      return var
+
+
+class EmbeddingColumnTest(test.TestCase):
+
+  def test_defaults(self):
+    categorical_column = fc.categorical_column_with_identity(
+        key='aaa', num_buckets=3)
+    embedding_dimension = 2
+    embedding_column = fc.embedding_column(
+        categorical_column, dimension=embedding_dimension)
+    self.assertIs(categorical_column, embedding_column.categorical_column)
+    self.assertEqual(embedding_dimension, embedding_column.dimension)
+    self.assertEqual('mean', embedding_column.combiner)
+    self.assertIsNone(embedding_column.ckpt_to_load_from)
+    self.assertIsNone(embedding_column.tensor_name_in_ckpt)
+    self.assertIsNone(embedding_column.max_norm)
+    self.assertTrue(embedding_column.trainable)
+    self.assertEqual('aaa_embedding', embedding_column.name)
+    self.assertEqual((embedding_dimension,), embedding_column.variable_shape)
+    self.assertEqual({
+        'aaa': parsing_ops.VarLenFeature(dtypes.int64)
+    }, embedding_column.parse_example_spec)
+
+  def test_all_constructor_args(self):
+    categorical_column = fc.categorical_column_with_identity(
+        key='aaa', num_buckets=3)
+    embedding_dimension = 2
+    embedding_column = fc.embedding_column(
+        categorical_column, dimension=embedding_dimension,
+        combiner='my_combiner', initializer=lambda: 'my_initializer',
+        ckpt_to_load_from='my_ckpt', tensor_name_in_ckpt='my_ckpt_tensor',
+        max_norm=42., trainable=False)
+    self.assertIs(categorical_column, embedding_column.categorical_column)
+    self.assertEqual(embedding_dimension, embedding_column.dimension)
+    self.assertEqual('my_combiner', embedding_column.combiner)
+    self.assertEqual('my_ckpt', embedding_column.ckpt_to_load_from)
+    self.assertEqual('my_ckpt_tensor', embedding_column.tensor_name_in_ckpt)
+    self.assertEqual(42., embedding_column.max_norm)
+    self.assertFalse(embedding_column.trainable)
+    self.assertEqual('aaa_embedding', embedding_column.name)
+    self.assertEqual((embedding_dimension,), embedding_column.variable_shape)
+    self.assertEqual({
+        'aaa': parsing_ops.VarLenFeature(dtypes.int64)
+    }, embedding_column.parse_example_spec)
+
+  def test_deep_copy(self):
+    categorical_column = fc.categorical_column_with_identity(
+        key='aaa', num_buckets=3)
+    embedding_dimension = 2
+    original = fc.embedding_column(
+        categorical_column, dimension=embedding_dimension,
+        combiner='my_combiner', initializer=lambda: 'my_initializer',
+        ckpt_to_load_from='my_ckpt', tensor_name_in_ckpt='my_ckpt_tensor',
+        max_norm=42., trainable=False)
+    for embedding_column in (original, copy.deepcopy(original)):
+      self.assertEqual('aaa', embedding_column.categorical_column.name)
+      self.assertEqual(3, embedding_column.categorical_column.num_buckets)
+      self.assertEqual({
+          'aaa': parsing_ops.VarLenFeature(dtypes.int64)
+      }, embedding_column.categorical_column.parse_example_spec)
+
+      self.assertEqual(embedding_dimension, embedding_column.dimension)
+      self.assertEqual('my_combiner', embedding_column.combiner)
+      self.assertEqual('my_ckpt', embedding_column.ckpt_to_load_from)
+      self.assertEqual('my_ckpt_tensor', embedding_column.tensor_name_in_ckpt)
+      self.assertEqual(42., embedding_column.max_norm)
+      self.assertFalse(embedding_column.trainable)
+      self.assertEqual('aaa_embedding', embedding_column.name)
+      self.assertEqual((embedding_dimension,), embedding_column.variable_shape)
+      self.assertEqual({
+          'aaa': parsing_ops.VarLenFeature(dtypes.int64)
+      }, embedding_column.parse_example_spec)
+
+  def test_invalid_initializer(self):
+    categorical_column = fc.categorical_column_with_identity(
+        key='aaa', num_buckets=3)
+    with self.assertRaisesRegexp(ValueError, 'initializer must be callable'):
+      fc.embedding_column(categorical_column, dimension=2, initializer='not_fn')
+
+  def test_parse_example(self):
+    a = fc.categorical_column_with_vocabulary_list(
+        key='aaa', vocabulary_list=('omar', 'stringer', 'marlo'))
+    a_embedded = fc.embedding_column(a, dimension=2)
+    data = example_pb2.Example(features=feature_pb2.Features(
+        feature={
+            'aaa':
+                feature_pb2.Feature(bytes_list=feature_pb2.BytesList(
+                    value=[b'omar', b'stringer']))
+        }))
+    features = parsing_ops.parse_example(
+        serialized=[data.SerializeToString()],
+        features=fc.make_parse_example_spec([a_embedded]))
+    self.assertIn('aaa', features)
+    with self.test_session():
+      _assert_sparse_tensor_value(
+          self,
+          sparse_tensor.SparseTensorValue(
+              indices=[[0, 0], [0, 1]],
+              values=np.array([b'omar', b'stringer'], dtype=np.object_),
+              dense_shape=[1, 2]),
+          features['aaa'].eval())
+
+  def test_transform_feature(self):
+    a = fc.categorical_column_with_identity(key='aaa', num_buckets=3)
+    a_embedded = fc.embedding_column(a, dimension=2)
+    features = {
+        'aaa': sparse_tensor.SparseTensor(
+            indices=((0, 0), (1, 0), (1, 1)),
+            values=(0, 1, 0),
+            dense_shape=(2, 2))
+    }
+    outputs = _transform_features(features, [a, a_embedded], None)
+    output_a = outputs[a]
+    output_embedded = outputs[a_embedded]
+    with _initialized_session():
+      _assert_sparse_tensor_value(
+          self, output_a.eval(), output_embedded.eval())
+
+  def test_get_dense_tensor(self):
+    # Inputs.
+    vocabulary_size = 3
+    sparse_input = sparse_tensor.SparseTensorValue(
+        # example 0, ids [2]
+        # example 1, ids [0, 1]
+        # example 2, ids []
+        # example 3, ids [1]
+        indices=((0, 0), (1, 0), (1, 4), (3, 0)),
+        values=(2, 0, 1, 1),
+        dense_shape=(4, 5))
+
+    # Embedding variable.
+    embedding_dimension = 2
+    embedding_values = (
+        (1., 2.),  # id 0
+        (3., 5.),  # id 1
+        (7., 11.)  # id 2
+    )
+    def _initializer(shape, dtype, partition_info):
+      self.assertAllEqual((vocabulary_size, embedding_dimension), shape)
+      self.assertEqual(dtypes.float32, dtype)
+      self.assertIsNone(partition_info)
+      return embedding_values
+
+    # Expected lookup result, using combiner='mean'.
+    expected_lookups = (
+        # example 0, ids [2], embedding = [7, 11]
+        (7., 11.),
+        # example 1, ids [0, 1], embedding = mean([1, 2] + [3, 5]) = [2, 3.5]
+        (2., 3.5),
+        # example 2, ids [], embedding = [0, 0]
+        (0., 0.),
+        # example 3, ids [1], embedding = [3, 5]
+        (3., 5.),
+    )
+
+    # Build columns.
+    categorical_column = fc.categorical_column_with_identity(
+        key='aaa', num_buckets=vocabulary_size)
+    embedding_column = fc.embedding_column(
+        categorical_column, dimension=embedding_dimension,
+        initializer=_initializer)
+    state_manager = _TestStateManager()
+
+    # Provide sparse input and get dense result.
+    embedding_lookup = embedding_column.get_dense_tensor(
+        FeatureTransformationCache({
+            'aaa': sparse_input
+        }), state_manager)
+
+    # Assert expected embedding variable and lookups.
+    global_vars = ops.get_collection(ops.GraphKeys.GLOBAL_VARIABLES)
+    self.assertItemsEqual(('embedding_weights:0',),
+                          tuple([v.name for v in global_vars]))
+    with _initialized_session():
+      self.assertAllEqual(embedding_values, global_vars[0].eval())
+      self.assertAllEqual(expected_lookups, embedding_lookup.eval())
+
+  def test_get_dense_tensor_3d(self):
+    # Inputs.
+    vocabulary_size = 4
+    sparse_input = sparse_tensor.SparseTensorValue(
+        # example 0, ids [2]
+        # example 1, ids [0, 1]
+        # example 2, ids []
+        # example 3, ids [1]
+        indices=((0, 0, 0), (1, 1, 0), (1, 1, 4), (3, 0, 0), (3, 1, 2)),
+        values=(2, 0, 1, 1, 2),
+        dense_shape=(4, 2, 5))
+
+    # Embedding variable.
+    embedding_dimension = 3
+    embedding_values = (
+        (1., 2., 4.),   # id 0
+        (3., 5., 1.),   # id 1
+        (7., 11., 2.),  # id 2
+        (2., 7., 12.)   # id 3
+    )
+    def _initializer(shape, dtype, partition_info):
+      self.assertAllEqual((vocabulary_size, embedding_dimension), shape)
+      self.assertEqual(dtypes.float32, dtype)
+      self.assertIsNone(partition_info)
+      return embedding_values
+
+    # Expected lookup result, using combiner='mean'.
+    expected_lookups = (
+        # example 0, ids [[2], []], embedding = [[7, 11, 2], [0, 0, 0]]
+        ((7., 11., 2.), (0., 0., 0.)),
+        # example 1, ids [[], [0, 1]], embedding
+        # = mean([[], [1, 2, 4] + [3, 5, 1]]) = [[0, 0, 0], [2, 3.5, 2.5]]
+        ((0., 0., 0.), (2., 3.5, 2.5)),
+        # example 2, ids [[], []], embedding = [[0, 0, 0], [0, 0, 0]]
+        ((0., 0., 0.), (0., 0., 0.)),
+        # example 3, ids [[1], [2]], embedding = [[3, 5, 1], [7, 11, 2]]
+        ((3., 5., 1.), (7., 11., 2.)),
+    )
+
+    # Build columns.
+    categorical_column = fc.categorical_column_with_identity(
+        key='aaa', num_buckets=vocabulary_size)
+    embedding_column = fc.embedding_column(
+        categorical_column, dimension=embedding_dimension,
+        initializer=_initializer)
+    state_manager = _TestStateManager()
+
+    # Provide sparse input and get dense result.
+    embedding_lookup = embedding_column.get_dense_tensor(
+        FeatureTransformationCache({
+            'aaa': sparse_input
+        }), state_manager)
+
+    # Assert expected embedding variable and lookups.
+    global_vars = ops.get_collection(ops.GraphKeys.GLOBAL_VARIABLES)
+    self.assertItemsEqual(('embedding_weights:0',),
+                          tuple([v.name for v in global_vars]))
+    with _initialized_session():
+      self.assertAllEqual(embedding_values, global_vars[0].eval())
+      self.assertAllEqual(expected_lookups, embedding_lookup.eval())
+
+  def DISABLED_test_get_dense_tensor_weight_collections(self):
+    sparse_input = sparse_tensor.SparseTensorValue(
+        # example 0, ids [2]
+        # example 1, ids [0, 1]
+        # example 2, ids []
+        # example 3, ids [1]
+        indices=((0, 0), (1, 0), (1, 4), (3, 0)),
+        values=(2, 0, 1, 1),
+        dense_shape=(4, 5))
+
+    # Build columns.
+    categorical_column = fc.categorical_column_with_identity(
+        key='aaa', num_buckets=3)
+    embedding_column = fc.embedding_column(categorical_column, dimension=2)
+
+    # Provide sparse input and get dense result.
+    embedding_column.get_dense_tensor(
+        FeatureTransformationCache({
+            'aaa': sparse_input
+        }),
+        weight_collections=('my_vars',))
+
+    # Assert expected embedding variable and lookups.
+    global_vars = ops.get_collection(ops.GraphKeys.GLOBAL_VARIABLES)
+    self.assertItemsEqual(('embedding_weights:0',),
+                          tuple([v.name for v in global_vars]))
+    my_vars = ops.get_collection('my_vars')
+    self.assertItemsEqual(
+        ('embedding_weights:0',), tuple([v.name for v in my_vars]))
+
+  def test_get_dense_tensor_placeholder_inputs(self):
+    # Inputs.
+    vocabulary_size = 3
+    sparse_input = sparse_tensor.SparseTensorValue(
+        # example 0, ids [2]
+        # example 1, ids [0, 1]
+        # example 2, ids []
+        # example 3, ids [1]
+        indices=((0, 0), (1, 0), (1, 4), (3, 0)),
+        values=(2, 0, 1, 1),
+        dense_shape=(4, 5))
+
+    # Embedding variable.
+    embedding_dimension = 2
+    embedding_values = (
+        (1., 2.),  # id 0
+        (3., 5.),  # id 1
+        (7., 11.)  # id 2
+    )
+    def _initializer(shape, dtype, partition_info):
+      self.assertAllEqual((vocabulary_size, embedding_dimension), shape)
+      self.assertEqual(dtypes.float32, dtype)
+      self.assertIsNone(partition_info)
+      return embedding_values
+
+    # Expected lookup result, using combiner='mean'.
+    expected_lookups = (
+        # example 0, ids [2], embedding = [7, 11]
+        (7., 11.),
+        # example 1, ids [0, 1], embedding = mean([1, 2] + [3, 5]) = [2, 3.5]
+        (2., 3.5),
+        # example 2, ids [], embedding = [0, 0]
+        (0., 0.),
+        # example 3, ids [1], embedding = [3, 5]
+        (3., 5.),
+    )
+
+    # Build columns.
+    categorical_column = fc.categorical_column_with_identity(
+        key='aaa', num_buckets=vocabulary_size)
+    embedding_column = fc.embedding_column(
+        categorical_column, dimension=embedding_dimension,
+        initializer=_initializer)
+    state_manager = _TestStateManager()
+
+    # Provide sparse input and get dense result.
+    input_indices = array_ops.placeholder(dtype=dtypes.int64)
+    input_values = array_ops.placeholder(dtype=dtypes.int64)
+    input_shape = array_ops.placeholder(dtype=dtypes.int64)
+    embedding_lookup = embedding_column.get_dense_tensor(
+        FeatureTransformationCache({
+            'aaa':
+                sparse_tensor.SparseTensorValue(
+                    indices=input_indices,
+                    values=input_values,
+                    dense_shape=input_shape)
+        }), state_manager)
+
+    # Assert expected embedding variable and lookups.
+    global_vars = ops.get_collection(ops.GraphKeys.GLOBAL_VARIABLES)
+    self.assertItemsEqual(
+        ('embedding_weights:0',), tuple([v.name for v in global_vars]))
+    with _initialized_session():
+      self.assertAllEqual(embedding_values, global_vars[0].eval())
+      self.assertAllEqual(expected_lookups, embedding_lookup.eval(
+          feed_dict={
+              input_indices: sparse_input.indices,
+              input_values: sparse_input.values,
+              input_shape: sparse_input.dense_shape,
+          }))
+
+  def test_get_dense_tensor_restore_from_ckpt(self):
+    # Inputs.
+    vocabulary_size = 3
+    sparse_input = sparse_tensor.SparseTensorValue(
+        # example 0, ids [2]
+        # example 1, ids [0, 1]
+        # example 2, ids []
+        # example 3, ids [1]
+        indices=((0, 0), (1, 0), (1, 4), (3, 0)),
+        values=(2, 0, 1, 1),
+        dense_shape=(4, 5))
+
+    # Embedding variable. The checkpoint file contains _embedding_values.
+    embedding_dimension = 2
+    embedding_values = (
+        (1., 2.),  # id 0
+        (3., 5.),  # id 1
+        (7., 11.)  # id 2
+    )
+    ckpt_path = test.test_src_dir_path(
+        'python/feature_column/testdata/embedding.ckpt')
+    ckpt_tensor = 'my_embedding'
+
+    # Expected lookup result, using combiner='mean'.
+    expected_lookups = (
+        # example 0, ids [2], embedding = [7, 11]
+        (7., 11.),
+        # example 1, ids [0, 1], embedding = mean([1, 2] + [3, 5]) = [2, 3.5]
+        (2., 3.5),
+        # example 2, ids [], embedding = [0, 0]
+        (0., 0.),
+        # example 3, ids [1], embedding = [3, 5]
+        (3., 5.),
+    )
+
+    # Build columns.
+    categorical_column = fc.categorical_column_with_identity(
+        key='aaa', num_buckets=vocabulary_size)
+    embedding_column = fc.embedding_column(
+        categorical_column, dimension=embedding_dimension,
+        ckpt_to_load_from=ckpt_path,
+        tensor_name_in_ckpt=ckpt_tensor)
+    state_manager = _TestStateManager()
+
+    # Provide sparse input and get dense result.
+    embedding_lookup = embedding_column.get_dense_tensor(
+        FeatureTransformationCache({
+            'aaa': sparse_input
+        }), state_manager)
+
+    # Assert expected embedding variable and lookups.
+    global_vars = ops.get_collection(ops.GraphKeys.GLOBAL_VARIABLES)
+    self.assertItemsEqual(
+        ('embedding_weights:0',), tuple([v.name for v in global_vars]))
+    with _initialized_session():
+      self.assertAllEqual(embedding_values, global_vars[0].eval())
+      self.assertAllEqual(expected_lookups, embedding_lookup.eval())
+
+  def test_linear_model(self):
+    # Inputs.
+    batch_size = 4
+    vocabulary_size = 3
+    sparse_input = sparse_tensor.SparseTensorValue(
+        # example 0, ids [2]
+        # example 1, ids [0, 1]
+        # example 2, ids []
+        # example 3, ids [1]
+        indices=((0, 0), (1, 0), (1, 4), (3, 0)),
+        values=(2, 0, 1, 1),
+        dense_shape=(batch_size, 5))
+
+    # Embedding variable.
+    embedding_dimension = 2
+    embedding_shape = (vocabulary_size, embedding_dimension)
+    zeros_embedding_values = np.zeros(embedding_shape)
+    def _initializer(shape, dtype, partition_info):
+      self.assertAllEqual(embedding_shape, shape)
+      self.assertEqual(dtypes.float32, dtype)
+      self.assertIsNone(partition_info)
+      return zeros_embedding_values
+
+    # Build columns.
+    categorical_column = fc_old.categorical_column_with_identity(
+        key='aaa', num_buckets=vocabulary_size)
+    embedding_column = fc_old.embedding_column(
+        categorical_column,
+        dimension=embedding_dimension,
+        initializer=_initializer)
+
+    with ops.Graph().as_default():
+      predictions = fc.linear_model({
+          categorical_column.name: sparse_input
+      }, (embedding_column,))
+      expected_var_names = (
+          'linear_model/bias_weights:0',
+          'linear_model/aaa_embedding/weights:0',
+          'linear_model/aaa_embedding/embedding_weights:0',
+      )
+      self.assertItemsEqual(
+          expected_var_names,
+          [v.name for v in ops.get_collection(ops.GraphKeys.GLOBAL_VARIABLES)])
+      trainable_vars = {
+          v.name: v for v in ops.get_collection(
+              ops.GraphKeys.TRAINABLE_VARIABLES)
+      }
+      self.assertItemsEqual(expected_var_names, trainable_vars.keys())
+      bias = trainable_vars['linear_model/bias_weights:0']
+      embedding_weights = trainable_vars[
+          'linear_model/aaa_embedding/embedding_weights:0']
+      linear_weights = trainable_vars[
+          'linear_model/aaa_embedding/weights:0']
+      with _initialized_session():
+        # Predictions with all zero weights.
+        self.assertAllClose(np.zeros((1,)), bias.eval())
+        self.assertAllClose(zeros_embedding_values, embedding_weights.eval())
+        self.assertAllClose(
+            np.zeros((embedding_dimension, 1)), linear_weights.eval())
+        self.assertAllClose(np.zeros((batch_size, 1)), predictions.eval())
+
+        # Predictions with all non-zero weights.
+        embedding_weights.assign((
+            (1., 2.),  # id 0
+            (3., 5.),  # id 1
+            (7., 11.)  # id 2
+        )).eval()
+        linear_weights.assign(((4.,), (6.,))).eval()
+        # example 0, ids [2], embedding[0] = [7, 11]
+        # example 1, ids [0, 1], embedding[1] = mean([1, 2] + [3, 5]) = [2, 3.5]
+        # example 2, ids [], embedding[2] = [0, 0]
+        # example 3, ids [1], embedding[3] = [3, 5]
+        # sum(embeddings * linear_weights)
+        # = [4*7 + 6*11, 4*2 + 6*3.5, 4*0 + 6*0, 4*3 + 6*5] = [94, 29, 0, 42]
+        self.assertAllClose(((94.,), (29.,), (0.,), (42.,)), predictions.eval())
+
+  def test_keras_linear_model(self):
+    # Inputs.
+    batch_size = 4
+    vocabulary_size = 3
+    sparse_input = sparse_tensor.SparseTensorValue(
+        # example 0, ids [2]
+        # example 1, ids [0, 1]
+        # example 2, ids []
+        # example 3, ids [1]
+        indices=((0, 0), (1, 0), (1, 4), (3, 0)),
+        values=(2, 0, 1, 1),
+        dense_shape=(batch_size, 5))
+
+    # Embedding variable.
+    embedding_dimension = 2
+    embedding_shape = (vocabulary_size, embedding_dimension)
+    zeros_embedding_values = np.zeros(embedding_shape)
+
+    def _initializer(shape, dtype, partition_info):
+      self.assertAllEqual(embedding_shape, shape)
+      self.assertEqual(dtypes.float32, dtype)
+      self.assertIsNone(partition_info)
+      return zeros_embedding_values
+
+    # Build columns.
+    categorical_column = fc_old.categorical_column_with_identity(
+        key='aaa', num_buckets=vocabulary_size)
+    embedding_column = fc_old.embedding_column(
+        categorical_column,
+        dimension=embedding_dimension,
+        initializer=_initializer)
+
+    with ops.Graph().as_default():
+      predictions = get_keras_linear_model_predictions({
+          categorical_column.name: sparse_input
+      }, (embedding_column,))
+      expected_var_names = (
+          'linear_model/bias_weights:0',
+          'linear_model/aaa_embedding/weights:0',
+          'linear_model/aaa_embedding/embedding_weights:0',
+      )
+      self.assertItemsEqual(
+          expected_var_names,
+          [v.name for v in ops.get_collection(ops.GraphKeys.GLOBAL_VARIABLES)])
+      trainable_vars = {
+          v.name: v
+          for v in ops.get_collection(ops.GraphKeys.TRAINABLE_VARIABLES)
+      }
+      self.assertItemsEqual(expected_var_names, trainable_vars.keys())
+      bias = trainable_vars['linear_model/bias_weights:0']
+      embedding_weights = trainable_vars[
+          'linear_model/aaa_embedding/embedding_weights:0']
+      linear_weights = trainable_vars['linear_model/aaa_embedding/weights:0']
+      with _initialized_session():
+        # Predictions with all zero weights.
+        self.assertAllClose(np.zeros((1,)), bias.eval())
+        self.assertAllClose(zeros_embedding_values, embedding_weights.eval())
+        self.assertAllClose(
+            np.zeros((embedding_dimension, 1)), linear_weights.eval())
+        self.assertAllClose(np.zeros((batch_size, 1)), predictions.eval())
+
+        # Predictions with all non-zero weights.
+        embedding_weights.assign((
+            (1., 2.),  # id 0
+            (3., 5.),  # id 1
+            (7., 11.)  # id 2
+        )).eval()
+        linear_weights.assign(((4.,), (6.,))).eval()
+        # example 0, ids [2], embedding[0] = [7, 11]
+        # example 1, ids [0, 1], embedding[1] = mean([1, 2] + [3, 5]) = [2, 3.5]
+        # example 2, ids [], embedding[2] = [0, 0]
+        # example 3, ids [1], embedding[3] = [3, 5]
+        # sum(embeddings * linear_weights)
+        # = [4*7 + 6*11, 4*2 + 6*3.5, 4*0 + 6*0, 4*3 + 6*5] = [94, 29, 0, 42]
+        self.assertAllClose(((94.,), (29.,), (0.,), (42.,)), predictions.eval())
+
+  def test_input_layer(self):
+    # Inputs.
+    vocabulary_size = 3
+    sparse_input = sparse_tensor.SparseTensorValue(
+        # example 0, ids [2]
+        # example 1, ids [0, 1]
+        # example 2, ids []
+        # example 3, ids [1]
+        indices=((0, 0), (1, 0), (1, 4), (3, 0)),
+        values=(2, 0, 1, 1),
+        dense_shape=(4, 5))
+
+    # Embedding variable.
+    embedding_dimension = 2
+    embedding_values = (
+        (1., 2.),  # id 0
+        (3., 5.),  # id 1
+        (7., 11.)  # id 2
+    )
+    def _initializer(shape, dtype, partition_info):
+      self.assertAllEqual((vocabulary_size, embedding_dimension), shape)
+      self.assertEqual(dtypes.float32, dtype)
+      self.assertIsNone(partition_info)
+      return embedding_values
+
+    # Expected lookup result, using combiner='mean'.
+    expected_lookups = (
+        # example 0, ids [2], embedding = [7, 11]
+        (7., 11.),
+        # example 1, ids [0, 1], embedding = mean([1, 2] + [3, 5]) = [2, 3.5]
+        (2., 3.5),
+        # example 2, ids [], embedding = [0, 0]
+        (0., 0.),
+        # example 3, ids [1], embedding = [3, 5]
+        (3., 5.),
+    )
+
+    # Build columns.
+    categorical_column = fc_old.categorical_column_with_identity(
+        key='aaa', num_buckets=vocabulary_size)
+    embedding_column = fc_old.embedding_column(
+        categorical_column,
+        dimension=embedding_dimension,
+        initializer=_initializer)
+
+    # Provide sparse input and get dense result.
+    input_layer = fc.input_layer({'aaa': sparse_input}, (embedding_column,))
+
+    # Assert expected embedding variable and lookups.
+    global_vars = ops.get_collection(ops.GraphKeys.GLOBAL_VARIABLES)
+    self.assertItemsEqual(
+        ('input_layer/aaa_embedding/embedding_weights:0',),
+        tuple([v.name for v in global_vars]))
+    trainable_vars = ops.get_collection(ops.GraphKeys.TRAINABLE_VARIABLES)
+    self.assertItemsEqual(
+        ('input_layer/aaa_embedding/embedding_weights:0',),
+        tuple([v.name for v in trainable_vars]))
+    with _initialized_session():
+      self.assertAllEqual(embedding_values, trainable_vars[0].eval())
+      self.assertAllEqual(expected_lookups, input_layer.eval())
+
+  def test_input_layer_not_trainable(self):
+    # Inputs.
+    vocabulary_size = 3
+    sparse_input = sparse_tensor.SparseTensorValue(
+        # example 0, ids [2]
+        # example 1, ids [0, 1]
+        # example 2, ids []
+        # example 3, ids [1]
+        indices=((0, 0), (1, 0), (1, 4), (3, 0)),
+        values=(2, 0, 1, 1),
+        dense_shape=(4, 5))
+
+    # Embedding variable.
+    embedding_dimension = 2
+    embedding_values = (
+        (1., 2.),  # id 0
+        (3., 5.),  # id 1
+        (7., 11.)  # id 2
+    )
+    def _initializer(shape, dtype, partition_info):
+      self.assertAllEqual((vocabulary_size, embedding_dimension), shape)
+      self.assertEqual(dtypes.float32, dtype)
+      self.assertIsNone(partition_info)
+      return embedding_values
+
+    # Expected lookup result, using combiner='mean'.
+    expected_lookups = (
+        # example 0, ids [2], embedding = [7, 11]
+        (7., 11.),
+        # example 1, ids [0, 1], embedding = mean([1, 2] + [3, 5]) = [2, 3.5]
+        (2., 3.5),
+        # example 2, ids [], embedding = [0, 0]
+        (0., 0.),
+        # example 3, ids [1], embedding = [3, 5]
+        (3., 5.),
+    )
+
+    # Build columns.
+    categorical_column = fc_old.categorical_column_with_identity(
+        key='aaa', num_buckets=vocabulary_size)
+    embedding_column = fc_old.embedding_column(
+        categorical_column,
+        dimension=embedding_dimension,
+        initializer=_initializer,
+        trainable=False)
+
+    # Provide sparse input and get dense result.
+    input_layer = fc.input_layer({'aaa': sparse_input}, (embedding_column,))
+
+    # Assert expected embedding variable and lookups.
+    global_vars = ops.get_collection(ops.GraphKeys.GLOBAL_VARIABLES)
+    self.assertItemsEqual(
+        ('input_layer/aaa_embedding/embedding_weights:0',),
+        tuple([v.name for v in global_vars]))
+    self.assertItemsEqual(
+        [], ops.get_collection(ops.GraphKeys.TRAINABLE_VARIABLES))
+    with _initialized_session():
+      self.assertAllEqual(embedding_values, global_vars[0].eval())
+      self.assertAllEqual(expected_lookups, input_layer.eval())
+
+
+class _TestSharedEmbeddingStateManager(StateManager):
+  """Manages the state for shared embedding columns.
+
+  This can handle multiple groups of shared embedding columns.
+  """
+
+  def __init__(self, trainable=True):
+    # Dict of shared_embedding_collection_name to a dict of variables.
+    self._all_variables = {}
+    self._trainable = trainable
+
+  def get_variable(self,
+                   feature_column,
+                   name,
+                   shape,
+                   dtype=None,
+                   initializer=None):
+    if not isinstance(feature_column, fc.SharedEmbeddingColumn):
+      raise ValueError(
+          'SharedEmbeddingStateManager can only handle SharedEmbeddingColumns. '
+          'Given type: {} '.format(type(feature_column)))
+
+    collection_name = feature_column.shared_collection_name
+    if collection_name not in self._all_variables:
+      self._all_variables[collection_name] = {}
+    var_dict = self._all_variables[collection_name]
+    if name in var_dict:
+      return var_dict[name]
+    else:
+      var = variable_scope.get_variable(
+          name=name,
+          shape=shape,
+          initializer=initializer,
+          trainable=self._trainable)
+      var_dict[name] = var
+      return var
+
+
+class SharedEmbeddingColumnTest(test.TestCase):
+
+  def test_defaults(self):
+    categorical_column_a = fc.categorical_column_with_identity(
+        key='aaa', num_buckets=3)
+    categorical_column_b = fc.categorical_column_with_identity(
+        key='bbb', num_buckets=3)
+    embedding_dimension = 2
+    embedding_column_b, embedding_column_a = fc.shared_embedding_columns(
+        [categorical_column_b, categorical_column_a],
+        dimension=embedding_dimension)
+    self.assertIs(categorical_column_a, embedding_column_a.categorical_column)
+    self.assertIs(categorical_column_b, embedding_column_b.categorical_column)
+    self.assertEqual(embedding_dimension, embedding_column_a.dimension)
+    self.assertEqual(embedding_dimension, embedding_column_b.dimension)
+    self.assertEqual('mean', embedding_column_a.combiner)
+    self.assertEqual('mean', embedding_column_b.combiner)
+    self.assertIsNone(embedding_column_a.ckpt_to_load_from)
+    self.assertIsNone(embedding_column_b.ckpt_to_load_from)
+    self.assertEqual('aaa_bbb_shared_embedding',
+                     embedding_column_a.shared_collection_name)
+    self.assertEqual('aaa_bbb_shared_embedding',
+                     embedding_column_b.shared_collection_name)
+    self.assertIsNone(embedding_column_a.tensor_name_in_ckpt)
+    self.assertIsNone(embedding_column_b.tensor_name_in_ckpt)
+    self.assertIsNone(embedding_column_a.max_norm)
+    self.assertIsNone(embedding_column_b.max_norm)
+    self.assertTrue(embedding_column_a.trainable)
+    self.assertTrue(embedding_column_b.trainable)
+    self.assertEqual('aaa_shared_embedding', embedding_column_a.name)
+    self.assertEqual('bbb_shared_embedding', embedding_column_b.name)
+    self.assertEqual((embedding_dimension,), embedding_column_a.variable_shape)
+    self.assertEqual((embedding_dimension,), embedding_column_b.variable_shape)
+    self.assertEqual({
+        'aaa': parsing_ops.VarLenFeature(dtypes.int64)
+    }, embedding_column_a.parse_example_spec)
+    self.assertEqual({
+        'bbb': parsing_ops.VarLenFeature(dtypes.int64)
+    }, embedding_column_b.parse_example_spec)
+
+  def test_all_constructor_args(self):
+    categorical_column_a = fc.categorical_column_with_identity(
+        key='aaa', num_buckets=3)
+    categorical_column_b = fc.categorical_column_with_identity(
+        key='bbb', num_buckets=3)
+    embedding_dimension = 2
+    embedding_column_a, embedding_column_b = fc.shared_embedding_columns(
+        [categorical_column_a, categorical_column_b],
+        dimension=embedding_dimension,
+        combiner='my_combiner',
+        initializer=lambda: 'my_initializer',
+        shared_embedding_collection_name='shared_embedding_collection_name',
+        ckpt_to_load_from='my_ckpt',
+        tensor_name_in_ckpt='my_ckpt_tensor',
+        max_norm=42.,
+        trainable=False)
+    self.assertIs(categorical_column_a, embedding_column_a.categorical_column)
+    self.assertIs(categorical_column_b, embedding_column_b.categorical_column)
+    self.assertEqual(embedding_dimension, embedding_column_a.dimension)
+    self.assertEqual(embedding_dimension, embedding_column_b.dimension)
+    self.assertEqual('my_combiner', embedding_column_a.combiner)
+    self.assertEqual('my_combiner', embedding_column_b.combiner)
+    self.assertEqual('shared_embedding_collection_name',
+                     embedding_column_a.shared_collection_name)
+    self.assertEqual('shared_embedding_collection_name',
+                     embedding_column_b.shared_collection_name)
+    self.assertEqual('my_ckpt', embedding_column_a.ckpt_to_load_from)
+    self.assertEqual('my_ckpt', embedding_column_b.ckpt_to_load_from)
+    self.assertEqual('my_ckpt_tensor', embedding_column_a.tensor_name_in_ckpt)
+    self.assertEqual('my_ckpt_tensor', embedding_column_b.tensor_name_in_ckpt)
+    self.assertEqual(42., embedding_column_a.max_norm)
+    self.assertEqual(42., embedding_column_b.max_norm)
+    self.assertFalse(embedding_column_a.trainable)
+    self.assertFalse(embedding_column_b.trainable)
+    self.assertEqual('aaa_shared_embedding', embedding_column_a.name)
+    self.assertEqual('bbb_shared_embedding', embedding_column_b.name)
+    self.assertEqual((embedding_dimension,), embedding_column_a.variable_shape)
+    self.assertEqual((embedding_dimension,), embedding_column_b.variable_shape)
+    self.assertEqual({
+        'aaa': parsing_ops.VarLenFeature(dtypes.int64)
+    }, embedding_column_a.parse_example_spec)
+    self.assertEqual({
+        'bbb': parsing_ops.VarLenFeature(dtypes.int64)
+    }, embedding_column_b.parse_example_spec)
+
+  def test_deep_copy(self):
+    categorical_column_a = fc.categorical_column_with_identity(
+        key='aaa', num_buckets=3)
+    categorical_column_b = fc.categorical_column_with_identity(
+        key='bbb', num_buckets=3)
+    embedding_dimension = 2
+    original_a, _ = fc.shared_embedding_columns(
+        [categorical_column_a, categorical_column_b],
+        dimension=embedding_dimension,
+        combiner='my_combiner',
+        initializer=lambda: 'my_initializer',
+        shared_embedding_collection_name='shared_embedding_collection_name',
+        ckpt_to_load_from='my_ckpt',
+        tensor_name_in_ckpt='my_ckpt_tensor',
+        max_norm=42., trainable=False)
+    for embedding_column_a in (original_a, copy.deepcopy(original_a)):
+      self.assertEqual('aaa', embedding_column_a.categorical_column.name)
+      self.assertEqual(3, embedding_column_a.categorical_column.num_buckets)
+      self.assertEqual({
+          'aaa': parsing_ops.VarLenFeature(dtypes.int64)
+      }, embedding_column_a.categorical_column.parse_example_spec)
+
+      self.assertEqual(embedding_dimension, embedding_column_a.dimension)
+      self.assertEqual('my_combiner', embedding_column_a.combiner)
+      self.assertEqual('shared_embedding_collection_name',
+                       embedding_column_a.shared_collection_name)
+      self.assertEqual('my_ckpt', embedding_column_a.ckpt_to_load_from)
+      self.assertEqual('my_ckpt_tensor', embedding_column_a.tensor_name_in_ckpt)
+      self.assertEqual(42., embedding_column_a.max_norm)
+      self.assertFalse(embedding_column_a.trainable)
+      self.assertEqual('aaa_shared_embedding', embedding_column_a.name)
+      self.assertEqual((embedding_dimension,),
+                       embedding_column_a.variable_shape)
+      self.assertEqual({
+          'aaa': parsing_ops.VarLenFeature(dtypes.int64)
+      }, embedding_column_a.parse_example_spec)
+
+  def test_invalid_initializer(self):
+    categorical_column_a = fc.categorical_column_with_identity(
+        key='aaa', num_buckets=3)
+    categorical_column_b = fc.categorical_column_with_identity(
+        key='bbb', num_buckets=3)
+    with self.assertRaisesRegexp(ValueError, 'initializer must be callable'):
+      fc.shared_embedding_columns(
+          [categorical_column_a, categorical_column_b], dimension=2,
+          initializer='not_fn')
+
+  def test_incompatible_column_type(self):
+    categorical_column_a = fc.categorical_column_with_identity(
+        key='aaa', num_buckets=3)
+    categorical_column_b = fc.categorical_column_with_identity(
+        key='bbb', num_buckets=3)
+    categorical_column_c = fc.categorical_column_with_hash_bucket(
+        key='ccc', hash_bucket_size=3)
+    with self.assertRaisesRegexp(
+        ValueError, 'all categorical_columns must have the same type.*'
+        'IdentityCategoricalColumn.*HashedCategoricalColumn'):
+      fc.shared_embedding_columns(
+          [categorical_column_a, categorical_column_b, categorical_column_c],
+          dimension=2)
+
+  def test_weighted_categorical_column_ok(self):
+    categorical_column_a = fc.categorical_column_with_identity(
+        key='aaa', num_buckets=3)
+    weighted_categorical_column_a = fc.weighted_categorical_column(
+        categorical_column_a, weight_feature_key='aaa_weights')
+    categorical_column_b = fc.categorical_column_with_identity(
+        key='bbb', num_buckets=3)
+    weighted_categorical_column_b = fc.weighted_categorical_column(
+        categorical_column_b, weight_feature_key='bbb_weights')
+    fc.shared_embedding_columns(
+        [weighted_categorical_column_a, categorical_column_b], dimension=2)
+    fc.shared_embedding_columns(
+        [categorical_column_a, weighted_categorical_column_b], dimension=2)
+    fc.shared_embedding_columns(
+        [weighted_categorical_column_a, weighted_categorical_column_b],
+        dimension=2)
+
+  def test_parse_example(self):
+    a = fc.categorical_column_with_vocabulary_list(
+        key='aaa', vocabulary_list=('omar', 'stringer', 'marlo'))
+    b = fc.categorical_column_with_vocabulary_list(
+        key='bbb', vocabulary_list=('omar', 'stringer', 'marlo'))
+    a_embedded, b_embedded = fc.shared_embedding_columns(
+        [a, b], dimension=2)
+    data = example_pb2.Example(features=feature_pb2.Features(
+        feature={
+            'aaa':
+                feature_pb2.Feature(bytes_list=feature_pb2.BytesList(
+                    value=[b'omar', b'stringer'])),
+            'bbb':
+                feature_pb2.Feature(bytes_list=feature_pb2.BytesList(
+                    value=[b'stringer', b'marlo'])),
+        }))
+    features = parsing_ops.parse_example(
+        serialized=[data.SerializeToString()],
+        features=fc.make_parse_example_spec([a_embedded, b_embedded]))
+    self.assertIn('aaa', features)
+    self.assertIn('bbb', features)
+    with self.test_session():
+      _assert_sparse_tensor_value(
+          self,
+          sparse_tensor.SparseTensorValue(
+              indices=[[0, 0], [0, 1]],
+              values=np.array([b'omar', b'stringer'], dtype=np.object_),
+              dense_shape=[1, 2]),
+          features['aaa'].eval())
+      _assert_sparse_tensor_value(
+          self,
+          sparse_tensor.SparseTensorValue(
+              indices=[[0, 0], [0, 1]],
+              values=np.array([b'stringer', b'marlo'], dtype=np.object_),
+              dense_shape=[1, 2]),
+          features['bbb'].eval())
+
+  def test_transform_feature(self):
+    a = fc.categorical_column_with_identity(key='aaa', num_buckets=3)
+    b = fc.categorical_column_with_identity(key='bbb', num_buckets=3)
+    a_embedded, b_embedded = fc.shared_embedding_columns(
+        [a, b], dimension=2)
+    features = {
+        'aaa': sparse_tensor.SparseTensor(
+            indices=((0, 0), (1, 0), (1, 1)),
+            values=(0, 1, 0),
+            dense_shape=(2, 2)),
+        'bbb': sparse_tensor.SparseTensor(
+            indices=((0, 0), (1, 0), (1, 1)),
+            values=(1, 2, 1),
+            dense_shape=(2, 2)),
+    }
+    outputs = _transform_features(features, [a, a_embedded, b, b_embedded],
+                                  None)
+    output_a = outputs[a]
+    output_a_embedded = outputs[a_embedded]
+    output_b = outputs[b]
+    output_b_embedded = outputs[b_embedded]
+    with _initialized_session():
+      _assert_sparse_tensor_value(
+          self, output_a.eval(), output_a_embedded.eval())
+      _assert_sparse_tensor_value(
+          self, output_b.eval(), output_b_embedded.eval())
+
+  def test_get_dense_tensor(self):
+    # Inputs.
+    vocabulary_size = 3
+    # -1 values are ignored.
+    input_a = np.array(
+        [[2, -1, -1],  # example 0, ids [2]
+         [0, 1, -1]])  # example 1, ids [0, 1]
+    input_b = np.array(
+        [[0, -1, -1],  # example 0, ids [0]
+         [-1, -1, -1]])  # example 1, ids []
+    input_features = {
+        'aaa': input_a,
+        'bbb': input_b
+    }
+
+    # Embedding variable.
+    embedding_dimension = 2
+    embedding_values = (
+        (1., 2.),  # id 0
+        (3., 5.),  # id 1
+        (7., 11.)  # id 2
+    )
+    def _initializer(shape, dtype, partition_info):
+      self.assertAllEqual((vocabulary_size, embedding_dimension), shape)
+      self.assertEqual(dtypes.float32, dtype)
+      self.assertIsNone(partition_info)
+      return embedding_values
+
+    # Expected lookup result, using combiner='mean'.
+    expected_lookups_a = (
+        # example 0:
+        (7., 11.),  # ids [2], embedding = [7, 11]
+        # example 1:
+        (2., 3.5),  # ids [0, 1], embedding = mean([1, 2] + [3, 5]) = [2, 3.5]
+    )
+    expected_lookups_b = (
+        # example 0:
+        (1., 2.),  # ids [0], embedding = [1, 2]
+        # example 1:
+        (0., 0.),  # ids [], embedding = [0, 0]
+    )
+
+    # Build columns.
+    categorical_column_a = fc.categorical_column_with_identity(
+        key='aaa', num_buckets=vocabulary_size)
+    categorical_column_b = fc.categorical_column_with_identity(
+        key='bbb', num_buckets=vocabulary_size)
+    embedding_column_a, embedding_column_b = fc.shared_embedding_columns(
+        [categorical_column_a, categorical_column_b],
+        dimension=embedding_dimension, initializer=_initializer)
+    state_manager = _TestSharedEmbeddingStateManager()
+
+    # Provide sparse input and get dense result.
+    embedding_lookup_a = embedding_column_a.get_dense_tensor(
+        FeatureTransformationCache(input_features), state_manager)
+    embedding_lookup_b = embedding_column_b.get_dense_tensor(
+        FeatureTransformationCache(input_features), state_manager)
+
+    # Assert expected embedding variable and lookups.
+    global_vars = ops.get_collection(ops.GraphKeys.GLOBAL_VARIABLES)
+    self.assertItemsEqual(('embedding_weights:0',),
+                          tuple([v.name for v in global_vars]))
+    embedding_var = global_vars[0]
+    with _initialized_session():
+      self.assertAllEqual(embedding_values, embedding_var.eval())
+      self.assertAllEqual(expected_lookups_a, embedding_lookup_a.eval())
+      self.assertAllEqual(expected_lookups_b, embedding_lookup_b.eval())
+
+  def DISABLED_test_get_dense_tensor_weight_collections(self):
+    # Inputs.
+    vocabulary_size = 3
+    # -1 values are ignored.
+    input_a = np.array([
+        [2, -1, -1],  # example 0, ids [2]
+        [0, 1, -1]
+    ])  # example 1, ids [0, 1]
+    input_b = np.array([
+        [0, -1, -1],  # example 0, ids [0]
+        [-1, -1, -1]
+    ])  # example 1, ids []
+    input_features = {'aaa': input_a, 'bbb': input_b}
+
+    # Embedding variable.
+    embedding_dimension = 2
+    embedding_values = (
+        (1., 2.),  # id 0
+        (3., 5.),  # id 1
+        (7., 11.)  # id 2
+    )
+
+    def _initializer(shape, dtype, partition_info):
+      self.assertAllEqual((vocabulary_size, embedding_dimension), shape)
+      self.assertEqual(dtypes.float32, dtype)
+      self.assertIsNone(partition_info)
+      return embedding_values
+
+    # Build columns.
+    categorical_column_a = fc.categorical_column_with_identity(
+        key='aaa', num_buckets=vocabulary_size)
+    categorical_column_b = fc.categorical_column_with_identity(
+        key='bbb', num_buckets=vocabulary_size)
+    embedding_column_a, embedding_column_b = fc.shared_embedding_columns(
+        [categorical_column_a, categorical_column_b],
+        dimension=embedding_dimension,
+        initializer=_initializer)
+
+    fc.input_layer(
+        input_features, [embedding_column_a, embedding_column_b],
+        weight_collections=('my_vars',))
+
+    # Assert expected embedding variable and lookups.
+    global_vars = ops.get_collection(ops.GraphKeys.GLOBAL_VARIABLES)
+    self.assertItemsEqual(
+        ('input_layer/aaa_bbb_shared_embedding/embedding_weights:0',),
+        tuple(v.name for v in global_vars))
+    my_vars = ops.get_collection('my_vars')
+    self.assertItemsEqual(
+        ('input_layer/aaa_bbb_shared_embedding/embedding_weights:0',),
+        tuple(v.name for v in my_vars))
+
+  def test_get_dense_tensor_placeholder_inputs(self):
+    # Inputs.
+    vocabulary_size = 3
+    # -1 values are ignored.
+    input_a = np.array(
+        [[2, -1, -1],  # example 0, ids [2]
+         [0, 1, -1]])  # example 1, ids [0, 1]
+    input_b = np.array(
+        [[0, -1, -1],  # example 0, ids [0]
+         [-1, -1, -1]])  # example 1, ids []
+    # Specify shape, because dense input must have rank specified.
+    input_a_placeholder = array_ops.placeholder(
+        dtype=dtypes.int64, shape=[None, 3])
+    input_b_placeholder = array_ops.placeholder(
+        dtype=dtypes.int64, shape=[None, 3])
+    input_features = {
+        'aaa': input_a_placeholder,
+        'bbb': input_b_placeholder,
+    }
+    feed_dict = {
+        input_a_placeholder: input_a,
+        input_b_placeholder: input_b,
+    }
+
+    # Embedding variable.
+    embedding_dimension = 2
+    embedding_values = (
+        (1., 2.),  # id 0
+        (3., 5.),  # id 1
+        (7., 11.)  # id 2
+    )
+    def _initializer(shape, dtype, partition_info):
+      self.assertAllEqual((vocabulary_size, embedding_dimension), shape)
+      self.assertEqual(dtypes.float32, dtype)
+      self.assertIsNone(partition_info)
+      return embedding_values
+
+    # Build columns.
+    categorical_column_a = fc.categorical_column_with_identity(
+        key='aaa', num_buckets=vocabulary_size)
+    categorical_column_b = fc.categorical_column_with_identity(
+        key='bbb', num_buckets=vocabulary_size)
+    embedding_column_a, embedding_column_b = fc.shared_embedding_columns(
+        [categorical_column_a, categorical_column_b],
+        dimension=embedding_dimension, initializer=_initializer)
+    state_manager = _TestSharedEmbeddingStateManager()
+
+    # Provide sparse input and get dense result.
+    embedding_lookup_a = embedding_column_a.get_dense_tensor(
+        FeatureTransformationCache(input_features), state_manager)
+    embedding_lookup_b = embedding_column_b.get_dense_tensor(
+        FeatureTransformationCache(input_features), state_manager)
+
+    with _initialized_session() as sess:
+      sess.run([embedding_lookup_a, embedding_lookup_b], feed_dict=feed_dict)
+
+  def test_linear_model(self):
+    # Inputs.
+    batch_size = 2
+    vocabulary_size = 3
+    # -1 values are ignored.
+    input_a = np.array(
+        [[2, -1, -1],  # example 0, ids [2]
+         [0, 1, -1]])  # example 1, ids [0, 1]
+    input_b = np.array(
+        [[0, -1, -1],  # example 0, ids [0]
+         [-1, -1, -1]])  # example 1, ids []
+
+    # Embedding variable.
+    embedding_dimension = 2
+    embedding_shape = (vocabulary_size, embedding_dimension)
+    zeros_embedding_values = np.zeros(embedding_shape)
+    def _initializer(shape, dtype, partition_info):
+      self.assertAllEqual(embedding_shape, shape)
+      self.assertEqual(dtypes.float32, dtype)
+      self.assertIsNone(partition_info)
+      return zeros_embedding_values
+
+    # Build columns.
+    categorical_column_a = fc_old.categorical_column_with_identity(
+        key='aaa', num_buckets=vocabulary_size)
+    categorical_column_b = fc_old.categorical_column_with_identity(
+        key='bbb', num_buckets=vocabulary_size)
+    embedding_column_a, embedding_column_b = fc_old.shared_embedding_columns(
+        [categorical_column_a, categorical_column_b],
+        dimension=embedding_dimension,
+        initializer=_initializer)
+
+    with ops.Graph().as_default():
+      predictions = fc.linear_model({
+          categorical_column_a.name: input_a,
+          categorical_column_b.name: input_b,
+      }, (embedding_column_a, embedding_column_b))
+      # Linear weights do not follow the column name. But this is a rare use
+      # case, and fixing it would add too much complexity to the code.
+      expected_var_names = (
+          'linear_model/bias_weights:0',
+          'linear_model/aaa_bbb_shared_embedding/weights:0',
+          'linear_model/aaa_bbb_shared_embedding/embedding_weights:0',
+          'linear_model/aaa_bbb_shared_embedding_1/weights:0',
+      )
+      self.assertItemsEqual(
+          expected_var_names,
+          [v.name for v in ops.get_collection(ops.GraphKeys.GLOBAL_VARIABLES)])
+      trainable_vars = {
+          v.name: v for v in ops.get_collection(
+              ops.GraphKeys.TRAINABLE_VARIABLES)
+      }
+      self.assertItemsEqual(expected_var_names, trainable_vars.keys())
+      bias = trainable_vars['linear_model/bias_weights:0']
+      embedding_weights = trainable_vars[
+          'linear_model/aaa_bbb_shared_embedding/embedding_weights:0']
+      linear_weights_a = trainable_vars[
+          'linear_model/aaa_bbb_shared_embedding/weights:0']
+      linear_weights_b = trainable_vars[
+          'linear_model/aaa_bbb_shared_embedding_1/weights:0']
+      with _initialized_session():
+        # Predictions with all zero weights.
+        self.assertAllClose(np.zeros((1,)), bias.eval())
+        self.assertAllClose(zeros_embedding_values, embedding_weights.eval())
+        self.assertAllClose(
+            np.zeros((embedding_dimension, 1)), linear_weights_a.eval())
+        self.assertAllClose(
+            np.zeros((embedding_dimension, 1)), linear_weights_b.eval())
+        self.assertAllClose(np.zeros((batch_size, 1)), predictions.eval())
+
+        # Predictions with all non-zero weights.
+        embedding_weights.assign((
+            (1., 2.),  # id 0
+            (3., 5.),  # id 1
+            (7., 11.)  # id 2
+        )).eval()
+        linear_weights_a.assign(((4.,), (6.,))).eval()
+        # example 0, ids [2], embedding[0] = [7, 11]
+        # example 1, ids [0, 1], embedding[1] = mean([1, 2] + [3, 5]) = [2, 3.5]
+        # sum(embeddings * linear_weights)
+        # = [4*7 + 6*11, 4*2 + 6*3.5] = [94, 29]
+        linear_weights_b.assign(((3.,), (5.,))).eval()
+        # example 0, ids [0], embedding[0] = [1, 2]
+        # example 1, ids [], embedding[1] = 0, 0]
+        # sum(embeddings * linear_weights)
+        # = [3*1 + 5*2, 3*0 +5*0] = [13, 0]
+        self.assertAllClose([[94. + 13.], [29.]], predictions.eval())
+
+  def test_keras_linear_model(self):
+    # Inputs.
+    batch_size = 2
+    vocabulary_size = 3
+    # -1 values are ignored.
+    input_a = np.array([
+        [2, -1, -1],  # example 0, ids [2]
+        [0, 1, -1]
+    ])  # example 1, ids [0, 1]
+    input_b = np.array([
+        [0, -1, -1],  # example 0, ids [0]
+        [-1, -1, -1]
+    ])  # example 1, ids []
+
+    # Embedding variable.
+    embedding_dimension = 2
+    embedding_shape = (vocabulary_size, embedding_dimension)
+    zeros_embedding_values = np.zeros(embedding_shape)
+
+    def _initializer(shape, dtype, partition_info):
+      self.assertAllEqual(embedding_shape, shape)
+      self.assertEqual(dtypes.float32, dtype)
+      self.assertIsNone(partition_info)
+      return zeros_embedding_values
+
+    # Build columns.
+    categorical_column_a = fc_old.categorical_column_with_identity(
+        key='aaa', num_buckets=vocabulary_size)
+    categorical_column_b = fc_old.categorical_column_with_identity(
+        key='bbb', num_buckets=vocabulary_size)
+    embedding_column_a, embedding_column_b = fc_old.shared_embedding_columns(
+        [categorical_column_a, categorical_column_b],
+        dimension=embedding_dimension,
+        initializer=_initializer)
+
+    with ops.Graph().as_default():
+      predictions = get_keras_linear_model_predictions({
+          categorical_column_a.name: input_a,
+          categorical_column_b.name: input_b,
+      }, (embedding_column_a, embedding_column_b))
+      # Linear weights do not follow the column name. But this is a rare use
+      # case, and fixing it would add too much complexity to the code.
+      expected_var_names = (
+          'linear_model/bias_weights:0',
+          'linear_model/aaa_bbb_shared_embedding/weights:0',
+          'linear_model/aaa_bbb_shared_embedding/embedding_weights:0',
+          'linear_model/aaa_bbb_shared_embedding_1/weights:0',
+      )
+      self.assertItemsEqual(
+          expected_var_names,
+          [v.name for v in ops.get_collection(ops.GraphKeys.GLOBAL_VARIABLES)])
+      trainable_vars = {
+          v.name: v
+          for v in ops.get_collection(ops.GraphKeys.TRAINABLE_VARIABLES)
+      }
+      self.assertItemsEqual(expected_var_names, trainable_vars.keys())
+      bias = trainable_vars['linear_model/bias_weights:0']
+      embedding_weights = trainable_vars[
+          'linear_model/aaa_bbb_shared_embedding/embedding_weights:0']
+      linear_weights_a = trainable_vars[
+          'linear_model/aaa_bbb_shared_embedding/weights:0']
+      linear_weights_b = trainable_vars[
+          'linear_model/aaa_bbb_shared_embedding_1/weights:0']
+      with _initialized_session():
+        # Predictions with all zero weights.
+        self.assertAllClose(np.zeros((1,)), bias.eval())
+        self.assertAllClose(zeros_embedding_values, embedding_weights.eval())
+        self.assertAllClose(
+            np.zeros((embedding_dimension, 1)), linear_weights_a.eval())
+        self.assertAllClose(
+            np.zeros((embedding_dimension, 1)), linear_weights_b.eval())
+        self.assertAllClose(np.zeros((batch_size, 1)), predictions.eval())
+
+        # Predictions with all non-zero weights.
+        embedding_weights.assign((
+            (1., 2.),  # id 0
+            (3., 5.),  # id 1
+            (7., 11.)  # id 2
+        )).eval()
+        linear_weights_a.assign(((4.,), (6.,))).eval()
+        # example 0, ids [2], embedding[0] = [7, 11]
+        # example 1, ids [0, 1], embedding[1] = mean([1, 2] + [3, 5]) = [2, 3.5]
+        # sum(embeddings * linear_weights)
+        # = [4*7 + 6*11, 4*2 + 6*3.5] = [94, 29]
+        linear_weights_b.assign(((3.,), (5.,))).eval()
+        # example 0, ids [0], embedding[0] = [1, 2]
+        # example 1, ids [], embedding[1] = 0, 0]
+        # sum(embeddings * linear_weights)
+        # = [3*1 + 5*2, 3*0 +5*0] = [13, 0]
+        self.assertAllClose([[94. + 13.], [29.]], predictions.eval())
+
+  def _test_input_layer(self, trainable=True):
+    # Inputs.
+    vocabulary_size = 3
+    sparse_input_a = sparse_tensor.SparseTensorValue(
+        # example 0, ids [2]
+        # example 1, ids [0, 1]
+        indices=((0, 0), (1, 0), (1, 4)),
+        values=(2, 0, 1),
+        dense_shape=(2, 5))
+    sparse_input_b = sparse_tensor.SparseTensorValue(
+        # example 0, ids [0]
+        # example 1, ids []
+        indices=((0, 0),),
+        values=(0,),
+        dense_shape=(2, 5))
+
+    # Embedding variable.
+    embedding_dimension = 2
+    embedding_values = (
+        (1., 2.),  # id 0
+        (3., 5.),  # id 1
+        (7., 11.)  # id 2
+    )
+    def _initializer(shape, dtype, partition_info):
+      self.assertAllEqual((vocabulary_size, embedding_dimension), shape)
+      self.assertEqual(dtypes.float32, dtype)
+      self.assertIsNone(partition_info)
+      return embedding_values
+
+    # Expected lookup result, using combiner='mean'.
+    expected_lookups = (
+        # example 0:
+        # A ids [2], embedding = [7, 11]
+        # B ids [0], embedding = [1, 2]
+        (7., 11., 1., 2.),
+        # example 1:
+        # A ids [0, 1], embedding = mean([1, 2] + [3, 5]) = [2, 3.5]
+        # B ids [], embedding = [0, 0]
+        (2., 3.5, 0., 0.),
+    )
+
+    # Build columns.
+    categorical_column_a = fc_old.categorical_column_with_identity(
+        key='aaa', num_buckets=vocabulary_size)
+    categorical_column_b = fc_old.categorical_column_with_identity(
+        key='bbb', num_buckets=vocabulary_size)
+    embedding_column_a, embedding_column_b = fc_old.shared_embedding_columns(
+        [categorical_column_a, categorical_column_b],
+        dimension=embedding_dimension,
+        initializer=_initializer,
+        trainable=trainable)
+
+    # Provide sparse input and get dense result.
+    input_layer = fc.input_layer(
+        features={'aaa': sparse_input_a, 'bbb': sparse_input_b},
+        feature_columns=(embedding_column_b, embedding_column_a))
+
+    # Assert expected embedding variable and lookups.
+    global_vars = ops.get_collection(ops.GraphKeys.GLOBAL_VARIABLES)
+    self.assertItemsEqual(
+        ['input_layer/aaa_bbb_shared_embedding/embedding_weights:0'],
+        tuple([v.name for v in global_vars]))
+    trainable_vars = ops.get_collection(ops.GraphKeys.TRAINABLE_VARIABLES)
+    if trainable:
+      self.assertItemsEqual(
+          ['input_layer/aaa_bbb_shared_embedding/embedding_weights:0'],
+          tuple([v.name for v in trainable_vars]))
+    else:
+      self.assertItemsEqual([], tuple([v.name for v in trainable_vars]))
+    shared_embedding_vars = global_vars
+    with _initialized_session():
+      self.assertAllEqual(embedding_values, shared_embedding_vars[0].eval())
+      self.assertAllEqual(expected_lookups, input_layer.eval())
+
+  def test_input_layer(self):
+    self._test_input_layer()
+
+  def test_input_layer_no_trainable(self):
+    self._test_input_layer(trainable=False)
+
+
+class WeightedCategoricalColumnTest(test.TestCase):
+
+  def test_defaults(self):
+    column = fc.weighted_categorical_column(
+        categorical_column=fc.categorical_column_with_identity(
+            key='ids', num_buckets=3),
+        weight_feature_key='values')
+    self.assertEqual('ids_weighted_by_values', column.name)
+    self.assertEqual(3, column.num_buckets)
+    self.assertEqual({
+        'ids': parsing_ops.VarLenFeature(dtypes.int64),
+        'values': parsing_ops.VarLenFeature(dtypes.float32)
+    }, column.parse_example_spec)
+
+  def test_deep_copy(self):
+    """Tests deepcopy of categorical_column_with_hash_bucket."""
+    original = fc.weighted_categorical_column(
+        categorical_column=fc.categorical_column_with_identity(
+            key='ids', num_buckets=3),
+        weight_feature_key='values')
+    for column in (original, copy.deepcopy(original)):
+      self.assertEqual('ids_weighted_by_values', column.name)
+      self.assertEqual(3, column.num_buckets)
+      self.assertEqual({
+          'ids': parsing_ops.VarLenFeature(dtypes.int64),
+          'values': parsing_ops.VarLenFeature(dtypes.float32)
+      }, column.parse_example_spec)
+
+  def test_invalid_dtype_none(self):
+    with self.assertRaisesRegexp(ValueError, 'is not convertible to float'):
+      fc.weighted_categorical_column(
+          categorical_column=fc.categorical_column_with_identity(
+              key='ids', num_buckets=3),
+          weight_feature_key='values',
+          dtype=None)
+
+  def test_invalid_dtype_string(self):
+    with self.assertRaisesRegexp(ValueError, 'is not convertible to float'):
+      fc.weighted_categorical_column(
+          categorical_column=fc.categorical_column_with_identity(
+              key='ids', num_buckets=3),
+          weight_feature_key='values',
+          dtype=dtypes.string)
+
+  def test_invalid_input_dtype(self):
+    column = fc.weighted_categorical_column(
+        categorical_column=fc.categorical_column_with_identity(
+            key='ids', num_buckets=3),
+        weight_feature_key='values')
+    strings = sparse_tensor.SparseTensorValue(
+        indices=((0, 0), (1, 0), (1, 1)),
+        values=('omar', 'stringer', 'marlo'),
+        dense_shape=(2, 2))
+    with self.assertRaisesRegexp(ValueError, 'Bad dtype'):
+      _transform_features({'ids': strings, 'values': strings}, (column,), None)
+
+  def test_column_name_collision(self):
+    with self.assertRaisesRegexp(ValueError, r'Parse config.*already exists'):
+      fc.weighted_categorical_column(
+          categorical_column=fc.categorical_column_with_identity(
+              key='aaa', num_buckets=3),
+          weight_feature_key='aaa').parse_example_spec()
+
+  def test_missing_weights(self):
+    column = fc.weighted_categorical_column(
+        categorical_column=fc.categorical_column_with_identity(
+            key='ids', num_buckets=3),
+        weight_feature_key='values')
+    inputs = sparse_tensor.SparseTensorValue(
+        indices=((0, 0), (1, 0), (1, 1)),
+        values=('omar', 'stringer', 'marlo'),
+        dense_shape=(2, 2))
+    with self.assertRaisesRegexp(
+        ValueError, 'values is not in features dictionary'):
+      _transform_features({'ids': inputs}, (column,), None)
+
+  def test_parse_example(self):
+    a = fc.categorical_column_with_vocabulary_list(
+        key='aaa', vocabulary_list=('omar', 'stringer', 'marlo'))
+    a_weighted = fc.weighted_categorical_column(a, weight_feature_key='weights')
+    data = example_pb2.Example(features=feature_pb2.Features(
+        feature={
+            'aaa':
+                feature_pb2.Feature(bytes_list=feature_pb2.BytesList(
+                    value=[b'omar', b'stringer'])),
+            'weights':
+                feature_pb2.Feature(float_list=feature_pb2.FloatList(
+                    value=[1., 10.]))
+        }))
+    features = parsing_ops.parse_example(
+        serialized=[data.SerializeToString()],
+        features=fc.make_parse_example_spec([a_weighted]))
+    self.assertIn('aaa', features)
+    self.assertIn('weights', features)
+    with self.test_session():
+      _assert_sparse_tensor_value(
+          self,
+          sparse_tensor.SparseTensorValue(
+              indices=[[0, 0], [0, 1]],
+              values=np.array([b'omar', b'stringer'], dtype=np.object_),
+              dense_shape=[1, 2]),
+          features['aaa'].eval())
+      _assert_sparse_tensor_value(
+          self,
+          sparse_tensor.SparseTensorValue(
+              indices=[[0, 0], [0, 1]],
+              values=np.array([1., 10.], dtype=np.float32),
+              dense_shape=[1, 2]),
+          features['weights'].eval())
+
+  def test_transform_features(self):
+    column = fc.weighted_categorical_column(
+        categorical_column=fc.categorical_column_with_identity(
+            key='ids', num_buckets=3),
+        weight_feature_key='values')
+    inputs = sparse_tensor.SparseTensorValue(
+        indices=((0, 0), (1, 0), (1, 1)),
+        values=(0, 1, 0),
+        dense_shape=(2, 2))
+    weights = sparse_tensor.SparseTensorValue(
+        indices=((0, 0), (1, 0), (1, 1)),
+        values=(0.5, 1.0, 0.1),
+        dense_shape=(2, 2))
+    id_tensor, weight_tensor = _transform_features({
+        'ids': inputs,
+        'values': weights,
+    }, (column,), None)[column]
+    with _initialized_session():
+      _assert_sparse_tensor_value(
+          self,
+          sparse_tensor.SparseTensorValue(
+              indices=inputs.indices,
+              values=np.array(inputs.values, dtype=np.int64),
+              dense_shape=inputs.dense_shape),
+          id_tensor.eval())
+      _assert_sparse_tensor_value(
+          self,
+          sparse_tensor.SparseTensorValue(
+              indices=weights.indices,
+              values=np.array(weights.values, dtype=np.float32),
+              dense_shape=weights.dense_shape),
+          weight_tensor.eval())
+
+  def test_transform_features_dense_input(self):
+    column = fc.weighted_categorical_column(
+        categorical_column=fc.categorical_column_with_identity(
+            key='ids', num_buckets=3),
+        weight_feature_key='values')
+    weights = sparse_tensor.SparseTensorValue(
+        indices=((0, 0), (1, 0), (1, 1)),
+        values=(0.5, 1.0, 0.1),
+        dense_shape=(2, 2))
+    id_tensor, weight_tensor = _transform_features({
+        'ids': ((0, -1), (1, 0)),
+        'values': weights,
+    }, (column,), None)[column]
+    with _initialized_session():
+      _assert_sparse_tensor_value(
+          self,
+          sparse_tensor.SparseTensorValue(
+              indices=((0, 0), (1, 0), (1, 1)),
+              values=np.array((0, 1, 0), dtype=np.int64),
+              dense_shape=(2, 2)),
+          id_tensor.eval())
+      _assert_sparse_tensor_value(
+          self,
+          sparse_tensor.SparseTensorValue(
+              indices=weights.indices,
+              values=np.array(weights.values, dtype=np.float32),
+              dense_shape=weights.dense_shape),
+          weight_tensor.eval())
+
+  def test_transform_features_dense_weights(self):
+    column = fc.weighted_categorical_column(
+        categorical_column=fc.categorical_column_with_identity(
+            key='ids', num_buckets=3),
+        weight_feature_key='values')
+    inputs = sparse_tensor.SparseTensorValue(
+        indices=((0, 0), (1, 0), (1, 1)),
+        values=(2, 1, 0),
+        dense_shape=(2, 2))
+    id_tensor, weight_tensor = _transform_features({
+        'ids': inputs,
+        'values': ((.5, 0.), (1., .1)),
+    }, (column,), None)[column]
+    with _initialized_session():
+      _assert_sparse_tensor_value(
+          self,
+          sparse_tensor.SparseTensorValue(
+              indices=inputs.indices,
+              values=np.array(inputs.values, dtype=np.int64),
+              dense_shape=inputs.dense_shape),
+          id_tensor.eval())
+      _assert_sparse_tensor_value(
+          self,
+          sparse_tensor.SparseTensorValue(
+              indices=((0, 0), (1, 0), (1, 1)),
+              values=np.array((.5, 1., .1), dtype=np.float32),
+              dense_shape=(2, 2)),
+          weight_tensor.eval())
+
+  def test_keras_linear_model(self):
+    column = fc_old.weighted_categorical_column(
+        categorical_column=fc_old.categorical_column_with_identity(
+            key='ids', num_buckets=3),
+        weight_feature_key='values')
+    with ops.Graph().as_default():
+      predictions = get_keras_linear_model_predictions({
+          'ids':
+              sparse_tensor.SparseTensorValue(
+                  indices=((0, 0), (1, 0), (1, 1)),
+                  values=(0, 2, 1),
+                  dense_shape=(2, 2)),
+          'values':
+              sparse_tensor.SparseTensorValue(
+                  indices=((0, 0), (1, 0), (1, 1)),
+                  values=(.5, 1., .1),
+                  dense_shape=(2, 2))
+      }, (column,))
+      bias = get_linear_model_bias()
+      weight_var = get_linear_model_column_var(column)
+      with _initialized_session():
+        self.assertAllClose((0.,), bias.eval())
+        self.assertAllClose(((0.,), (0.,), (0.,)), weight_var.eval())
+        self.assertAllClose(((0.,), (0.,)), predictions.eval())
+        weight_var.assign(((1.,), (2.,), (3.,))).eval()
+        # weight_var[0] * weights[0, 0] = 1 * .5 = .5
+        # weight_var[2] * weights[1, 0] + weight_var[1] * weights[1, 1]
+        # = 3*1 + 2*.1 = 3+.2 = 3.2
+        self.assertAllClose(((.5,), (3.2,)), predictions.eval())
+
+  def test_keras_linear_model_mismatched_shape(self):
+    column = fc_old.weighted_categorical_column(
+        categorical_column=fc_old.categorical_column_with_identity(
+            key='ids', num_buckets=3),
+        weight_feature_key='values')
+    with ops.Graph().as_default():
+      with self.assertRaisesRegexp(ValueError,
+                                   r'Dimensions.*are not compatible'):
+        get_keras_linear_model_predictions({
+            'ids':
+                sparse_tensor.SparseTensorValue(
+                    indices=((0, 0), (1, 0), (1, 1)),
+                    values=(0, 2, 1),
+                    dense_shape=(2, 2)),
+            'values':
+                sparse_tensor.SparseTensorValue(
+                    indices=((0, 0), (0, 1), (1, 0), (1, 1)),
+                    values=(.5, 11., 1., .1),
+                    dense_shape=(2, 2))
+        }, (column,))
+
+  def test_keras_linear_model_mismatched_dense_values(self):
+    column = fc_old.weighted_categorical_column(
+        categorical_column=fc_old.categorical_column_with_identity(
+            key='ids', num_buckets=3),
+        weight_feature_key='values')
+    with ops.Graph().as_default():
+      predictions = get_keras_linear_model_predictions(
+          {
+              'ids':
+                  sparse_tensor.SparseTensorValue(
+                      indices=((0, 0), (1, 0), (1, 1)),
+                      values=(0, 2, 1),
+                      dense_shape=(2, 2)),
+              'values': ((.5,), (1.,))
+          }, (column,),
+          sparse_combiner='mean')
+      # Disabling the constant folding optimizer here since it changes the
+      # error message differently on CPU and GPU.
+      config = config_pb2.ConfigProto()
+      config.graph_options.rewrite_options.constant_folding = (
+          rewriter_config_pb2.RewriterConfig.OFF)
+      with _initialized_session(config):
+        with self.assertRaisesRegexp(errors.OpError, 'Incompatible shapes'):
+          predictions.eval()
+
+  def test_keras_linear_model_mismatched_dense_shape(self):
+    column = fc_old.weighted_categorical_column(
+        categorical_column=fc_old.categorical_column_with_identity(
+            key='ids', num_buckets=3),
+        weight_feature_key='values')
+    with ops.Graph().as_default():
+      predictions = get_keras_linear_model_predictions({
+          'ids':
+              sparse_tensor.SparseTensorValue(
+                  indices=((0, 0), (1, 0), (1, 1)),
+                  values=(0, 2, 1),
+                  dense_shape=(2, 2)),
+          'values': ((.5,), (1.,), (.1,))
+      }, (column,))
+      bias = get_linear_model_bias()
+      weight_var = get_linear_model_column_var(column)
+      with _initialized_session():
+        self.assertAllClose((0.,), bias.eval())
+        self.assertAllClose(((0.,), (0.,), (0.,)), weight_var.eval())
+        self.assertAllClose(((0.,), (0.,)), predictions.eval())
+        weight_var.assign(((1.,), (2.,), (3.,))).eval()
+        # weight_var[0] * weights[0, 0] = 1 * .5 = .5
+        # weight_var[2] * weights[1, 0] + weight_var[1] * weights[1, 1]
+        # = 3*1 + 2*.1 = 3+.2 = 3.2
+        self.assertAllClose(((.5,), (3.2,)), predictions.eval())
+
+  def test_linear_model(self):
+    column = fc_old.weighted_categorical_column(
+        categorical_column=fc_old.categorical_column_with_identity(
+            key='ids', num_buckets=3),
+        weight_feature_key='values')
+    with ops.Graph().as_default():
+      predictions = fc.linear_model({
+          'ids': sparse_tensor.SparseTensorValue(
+              indices=((0, 0), (1, 0), (1, 1)),
+              values=(0, 2, 1),
+              dense_shape=(2, 2)),
+          'values': sparse_tensor.SparseTensorValue(
+              indices=((0, 0), (1, 0), (1, 1)),
+              values=(.5, 1., .1),
+              dense_shape=(2, 2))
+      }, (column,))
+      bias = get_linear_model_bias()
+      weight_var = get_linear_model_column_var(column)
+      with _initialized_session():
+        self.assertAllClose((0.,), bias.eval())
+        self.assertAllClose(((0.,), (0.,), (0.,)), weight_var.eval())
+        self.assertAllClose(((0.,), (0.,)), predictions.eval())
+        weight_var.assign(((1.,), (2.,), (3.,))).eval()
+        # weight_var[0] * weights[0, 0] = 1 * .5 = .5
+        # weight_var[2] * weights[1, 0] + weight_var[1] * weights[1, 1]
+        # = 3*1 + 2*.1 = 3+.2 = 3.2
+        self.assertAllClose(((.5,), (3.2,)), predictions.eval())
+
+  def test_linear_model_mismatched_shape(self):
+    column = fc_old.weighted_categorical_column(
+        categorical_column=fc_old.categorical_column_with_identity(
+            key='ids', num_buckets=3),
+        weight_feature_key='values')
+    with ops.Graph().as_default():
+      with self.assertRaisesRegexp(
+          ValueError, r'Dimensions.*are not compatible'):
+        fc.linear_model({
+            'ids': sparse_tensor.SparseTensorValue(
+                indices=((0, 0), (1, 0), (1, 1)),
+                values=(0, 2, 1),
+                dense_shape=(2, 2)),
+            'values': sparse_tensor.SparseTensorValue(
+                indices=((0, 0), (0, 1), (1, 0), (1, 1)),
+                values=(.5, 11., 1., .1),
+                dense_shape=(2, 2))
+        }, (column,))
+
+  def test_linear_model_mismatched_dense_values(self):
+    column = fc_old.weighted_categorical_column(
+        categorical_column=fc_old.categorical_column_with_identity(
+            key='ids', num_buckets=3),
+        weight_feature_key='values')
+    with ops.Graph().as_default():
+      predictions = fc.linear_model(
+          {
+              'ids':
+                  sparse_tensor.SparseTensorValue(
+                      indices=((0, 0), (1, 0), (1, 1)),
+                      values=(0, 2, 1),
+                      dense_shape=(2, 2)),
+              'values': ((.5,), (1.,))
+          }, (column,),
+          sparse_combiner='mean')
+      # Disabling the constant folding optimizer here since it changes the
+      # error message differently on CPU and GPU.
+      config = config_pb2.ConfigProto()
+      config.graph_options.rewrite_options.constant_folding = (
+          rewriter_config_pb2.RewriterConfig.OFF)
+      with _initialized_session(config):
+        with self.assertRaisesRegexp(errors.OpError, 'Incompatible shapes'):
+          predictions.eval()
+
+  def test_linear_model_mismatched_dense_shape(self):
+    column = fc_old.weighted_categorical_column(
+        categorical_column=fc_old.categorical_column_with_identity(
+            key='ids', num_buckets=3),
+        weight_feature_key='values')
+    with ops.Graph().as_default():
+      predictions = fc.linear_model({
+          'ids': sparse_tensor.SparseTensorValue(
+              indices=((0, 0), (1, 0), (1, 1)),
+              values=(0, 2, 1),
+              dense_shape=(2, 2)),
+          'values': ((.5,), (1.,), (.1,))
+      }, (column,))
+      bias = get_linear_model_bias()
+      weight_var = get_linear_model_column_var(column)
+      with _initialized_session():
+        self.assertAllClose((0.,), bias.eval())
+        self.assertAllClose(((0.,), (0.,), (0.,)), weight_var.eval())
+        self.assertAllClose(((0.,), (0.,)), predictions.eval())
+        weight_var.assign(((1.,), (2.,), (3.,))).eval()
+        # weight_var[0] * weights[0, 0] = 1 * .5 = .5
+        # weight_var[2] * weights[1, 0] + weight_var[1] * weights[1, 1]
+        # = 3*1 + 2*.1 = 3+.2 = 3.2
+        self.assertAllClose(((.5,), (3.2,)), predictions.eval())
+
+  # TODO(ptucker): Add test with embedding of weighted categorical.
+
+if __name__ == '__main__':
+  test.main()
diff --git a/tensorflow/python/framework/common_shapes.py b/tensorflow/python/framework/common_shapes.py
index 3c5aebbce8af117aa1e216f1ef07ded181c997ea..40788e24c486c4357042672e3697063a4c7fb381 100644
--- a/tensorflow/python/framework/common_shapes.py
+++ b/tensorflow/python/framework/common_shapes.py
@@ -28,6 +28,18 @@ from tensorflow.python.framework import tensor_shape
 from tensorflow.python.framework import tensor_util
 
 
+def has_fully_defined_shape(tensor):
+  """Returns true if tensor has a fully defined shape."""
+  return isinstance(tensor, ops.EagerTensor) or tensor.shape.is_fully_defined()
+
+
+def rank(tensor):
+  """Return a rank if it is a tensor, else return None."""
+  if isinstance(tensor, ops.Tensor):
+    return tensor._rank()  # pylint: disable=protected-access
+  return None
+
+
 def scalar_shape(unused_op):
   """Shape function for ops that output a scalar value."""
   return [tensor_shape.scalar()]
diff --git a/tensorflow/python/framework/error_interpolation.py b/tensorflow/python/framework/error_interpolation.py
new file mode 100644
index 0000000000000000000000000000000000000000..9ccae761471e24ddb1d4d6acd89ebcc9650d1320
--- /dev/null
+++ b/tensorflow/python/framework/error_interpolation.py
@@ -0,0 +1,92 @@
+# Copyright 2018 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""Function for interpolating formatted errors from the TensorFlow runtime.
+
+Exposes the function `interpolate` to interpolate messages with tags of the form
+^^type:name:format^^.
+"""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import collections
+import itertools
+import re
+import string
+
+import six
+
+_NAME_REGEX = r"[A-Za-z0-9.][A-Za-z0-9_.\-/]*?"
+_FORMAT_REGEX = r"[A-Za-z0-9_.\-/${}:]+"
+_TAG_REGEX = r"\^\^({name}):({name}):({fmt})\^\^".format(
+    name=_NAME_REGEX, fmt=_FORMAT_REGEX)
+_INTERPOLATION_REGEX = r"^(.*?)({tag})".format(tag=_TAG_REGEX)
+_INTERPOLATION_PATTERN = re.compile(_INTERPOLATION_REGEX)
+
+_ParseTag = collections.namedtuple("_ParseTag", ["type", "name", "format"])
+
+
+def _parse_message(message):
+  """Parses the message.
+
+  Splits the message into separators and tags. Tags are named tuples
+  representing the string ^^type:name:format^^ and they are separated by
+  separators. For example, in
+  "123^^node:Foo:${file}^^456^^node:Bar:${line}^^789", there are two tags and
+  three separators. The separators are the numeric characters.
+
+  Args:
+    message: String to parse
+
+  Returns:
+    (list of separator strings, list of _ParseTags).
+
+    For example, if message is "123^^node:Foo:${file}^^456" then this function
+    returns (["123", "456"], [_ParseTag("node", "Foo", "${file}")])
+  """
+  seps = []
+  tags = []
+  pos = 0
+  while pos < len(message):
+    match = re.match(_INTERPOLATION_PATTERN, message[pos:])
+    if match:
+      seps.append(match.group(1))
+      tags.append(_ParseTag(match.group(3), match.group(4), match.group(5)))
+      pos += match.end()
+    else:
+      break
+  seps.append(message[pos:])
+  return seps, tags
+
+
+# TODO(jtkeeling): Modify to actually interpolate format strings rather than
+# echoing them.
+def interpolate(error_message):
+  """Interpolates an error message.
+
+  The error message can contain tags of the form ^^type:name:format^^ which will
+  be replaced.
+
+  Args:
+    error_message: A string to interpolate.
+
+  Returns:
+    The string with tags of the form ^^type:name:format^^ interpolated.
+  """
+  seps, tags = _parse_message(error_message)
+  subs = [string.Template(tag.format).safe_substitute({}) for tag in tags]
+  return "".join(
+      itertools.chain(*six.moves.zip_longest(seps, subs, fillvalue="")))
diff --git a/tensorflow/python/framework/error_interpolation_test.py b/tensorflow/python/framework/error_interpolation_test.py
new file mode 100644
index 0000000000000000000000000000000000000000..ad448deb622cb6a3d24e502d7238d3f614d5af4d
--- /dev/null
+++ b/tensorflow/python/framework/error_interpolation_test.py
@@ -0,0 +1,49 @@
+# Copyright 2018 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""Tests for tensorflow.python.framework.errors."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+from tensorflow.python.framework import error_interpolation
+from tensorflow.python.platform import test
+
+
+class InterpolateTest(test.TestCase):
+
+  def testNothingToDo(self):
+    normal_string = "This is just a normal string"
+    interpolated_string = error_interpolation.interpolate(normal_string)
+    self.assertEqual(interpolated_string, normal_string)
+
+  def testOneTag(self):
+    one_tag_string = "^^node:Foo:${file}^^"
+    interpolated_string = error_interpolation.interpolate(one_tag_string)
+    self.assertEqual(interpolated_string, "${file}")
+
+  def testTwoTagsNoSeps(self):
+    two_tags_no_seps = "^^node:Foo:${file}^^^^node:Bar:${line}^^"
+    interpolated_string = error_interpolation.interpolate(two_tags_no_seps)
+    self.assertEqual(interpolated_string, "${file}${line}")
+
+  def testTwoTagsWithSeps(self):
+    two_tags_with_seps = "123^^node:Foo:${file}^^456^^node:Bar:${line}^^789"
+    interpolated_string = error_interpolation.interpolate(two_tags_with_seps)
+    self.assertEqual(interpolated_string, "123${file}456${line}789")
+
+
+if __name__ == "__main__":
+  test.main()
diff --git a/tensorflow/python/framework/function_test.py b/tensorflow/python/framework/function_test.py
index 15e41ba91f9ae121d3d4ea48e3e71eace7cd9a3e..1707f929b89203e1890ee96fd153ace2063b449c 100644
--- a/tensorflow/python/framework/function_test.py
+++ b/tensorflow/python/framework/function_test.py
@@ -537,19 +537,25 @@ class FunctionTest(test.TestCase):
   def testResourceVarAsImplicitInput(self):
     g = ops.Graph()
     with g.as_default(), ops.device("cpu:0"):
+      expected_type = dtypes.float32
+      expected_shape = tensor_shape.TensorShape((4, 4))
       v = variable_scope.get_variable(
-          "var", (4, 4), dtypes.float32, use_resource=True)
+          "var", expected_shape, expected_type, use_resource=True)
 
       @function.Defun()
       def Foo():
-        return array_ops.identity(v)
+        captured = array_ops.identity(v)
+        self.assertEqual(expected_type, captured.dtype)
+        self.assertEqual(expected_shape, captured.shape)
+        return captured, array_ops.shape(captured)
 
-      y = v.value()
-      z = Foo()
+      expected_val = v.value()
+      actual_val, actual_shape = Foo()
 
     with self.test_session(graph=g):
       v.initializer.run()
-      self.assertAllEqual(y.eval(), z.eval())
+      self.assertAllEqual(expected_val.eval(), actual_val.eval())
+      self.assertAllEqual(expected_shape, actual_shape.eval())
 
   def testDefineErrors(self):
     with ops.Graph().as_default():
diff --git a/tensorflow/python/framework/importer.py b/tensorflow/python/framework/importer.py
index 72eb7e0eeb73fb1f8725ab2cbd4182e543c79b9f..699d2b70d176db7718a6e480f9f7b08a65ae6a8e 100644
--- a/tensorflow/python/framework/importer.py
+++ b/tensorflow/python/framework/importer.py
@@ -407,11 +407,11 @@ def import_graph_def(graph_def,
   _PopulateTFImportGraphDefOptions(options, prefix, input_map,
                                    return_elements)
 
-  # _ProcessNewOps mutates the new operations. _lock ensures a Session.run
-  # call cannot occur between creating the TF_Operations in the
+  # _ProcessNewOps mutates the new operations. _mutation_lock ensures a
+  # Session.run call cannot occur between creating the TF_Operations in the
   # TF_GraphImportGraphDefWithResults call and mutating the them in
   # _ProcessNewOps.
-  with graph._lock:  # pylint: disable=protected-access
+  with graph._mutation_lock():  # pylint: disable=protected-access
     with c_api_util.tf_buffer(graph_def.SerializeToString()) as serialized:
       try:
         results = c_api.TF_GraphImportGraphDefWithResults(
diff --git a/tensorflow/python/framework/ops.py b/tensorflow/python/framework/ops.py
index 05f9ae21b174d31f909f15c00df2952244050ffa..c4f58f08479656935ddcec578328f9441c2a4360 100644
--- a/tensorflow/python/framework/ops.py
+++ b/tensorflow/python/framework/ops.py
@@ -20,7 +20,6 @@ from __future__ import print_function
 
 import collections
 import copy
-import linecache
 import os
 import re
 import sys
@@ -48,13 +47,16 @@ from tensorflow.python.framework import dtypes
 from tensorflow.python.framework import errors
 from tensorflow.python.framework import op_def_registry
 from tensorflow.python.framework import registry
+from tensorflow.python.util import tf_stack
 from tensorflow.python.framework import tensor_shape
+from tensorflow.python.framework import traceable_stack
 from tensorflow.python.framework import versions
 from tensorflow.python.ops import control_flow_util
 from tensorflow.python.platform import app
 from tensorflow.python.platform import tf_logging as logging
 from tensorflow.python.util import compat
 from tensorflow.python.util import decorator_utils
+from tensorflow.python.util import lock_util
 from tensorflow.python.util import tf_contextlib
 from tensorflow.python.util.deprecation import deprecated_args
 from tensorflow.python.util.tf_export import tf_export
@@ -705,7 +707,7 @@ class _EagerTensorBase(Tensor):
     """
     if self.dtype == dtypes.resource:
       raise ValueError("Resource handles are not convertible to numpy.")
-    return self.cpu()._numpy()  # pylint: disable=protected-access
+    return self._cpu_nograd()._numpy()  # pylint: disable=protected-access
 
   # __int__ and  __float__ may copy the tensor to CPU and
   # only work for scalars; values are cast as per numpy.
@@ -779,8 +781,8 @@ class _EagerTensorBase(Tensor):
   def _override_operator(name, func):
     setattr(_EagerTensorBase, name, func)
 
-  def _copy(self, ctx=None, device_name=None):
-    """Copies tensor to dest device."""
+  def _copy_nograd(self, ctx=None, device_name=None):
+    """Copies tensor to dest device, but doesn't record the operation."""
     # pylint: disable=protected-access
     # Creates a new tensor on the dest device.
     if ctx is None:
@@ -792,7 +794,11 @@ class _EagerTensorBase(Tensor):
       new_tensor = self._copy_to_device(context=ctx._handle, device=device_name)
     except core._NotOkStatusException as e:
       six.raise_from(core._status_to_exception(e.code, e.message), None)
+    return new_tensor
 
+  def _copy(self, ctx=None, device_name=None):
+    """Copies tensor to dest device."""
+    new_tensor = self._copy_nograd(ctx, device_name)
     # Record the copy on tape and define backprop copy as well.
     if context.executing_eagerly():
       self_device = self.device
@@ -823,6 +829,16 @@ class _EagerTensorBase(Tensor):
     """Returns the number of Tensor dimensions."""
     return self.shape.ndims
 
+  def _cpu_nograd(self):
+    """A copy of this Tensor with contents backed by host memory.
+
+    The copy cannot be differentiated through.
+
+    Returns:
+      A CPU-memory backed Tensor object with the same contents as this Tensor.
+    """
+    return self._copy_nograd(context.context(), "CPU:0")
+
   def cpu(self):
     """A copy of this Tensor with contents backed by host memory."""
     return self._copy(context.context(), "CPU:0")
@@ -1698,7 +1714,7 @@ class Operation(object):
 
     self._id_value = self._graph._next_id()  # pylint: disable=protected-access
     self._original_op = original_op
-    self._traceback = self._graph._extract_stack()  # pylint: disable=protected-access
+    self._traceback = tf_stack.extract_stack()
     self._control_flow_context = self.graph._get_control_flow_context()  # pylint: disable=protected-access
 
     # Initialize self._c_op.
@@ -2139,7 +2155,7 @@ class Operation(object):
   @property
   def traceback(self):
     """Returns the call stack from when this operation was constructed."""
-    return self._graph._convert_stack(self._traceback)  # pylint: disable=protected-access
+    return tf_stack.convert_stack(self._traceback)
 
   @property
   def traceback_with_start_lines(self):
@@ -2148,9 +2164,8 @@ class Operation(object):
     Returns:
       A list of 5-tuples (filename, lineno, name, code, func_start_lineno).
     """
-    return self._graph._convert_stack(  # pylint: disable=protected-access
-        self._traceback,
-        include_func_start_lineno=True)
+    return tf_stack.convert_stack(self._traceback,
+                                  include_func_start_lineno=True)
 
   def _set_attr(self, attr_name, attr_value):
     """Private method used to set an attribute in the node_def."""
@@ -2599,6 +2614,9 @@ def _name_from_scope_name(name):
   return name[:-1] if (name and name[-1] == "/") else name
 
 
+_MUTATION_LOCK_GROUP = 0
+_SESSION_RUN_LOCK_GROUP = 1
+
 @tf_export("Graph")
 class Graph(object):
   """A TensorFlow computation, represented as a dataflow graph.
@@ -2648,20 +2666,21 @@ class Graph(object):
 
   def __init__(self):
     """Creates a new, empty Graph."""
-    # Protects core state that can be returned via public accessors, as well as
-    # synchronizes Session.run calls with methods that create and mutate ops
-    # (e.g. Graph.create_op()). This synchronization is necessary because it's
-    # illegal to modify an operation after it's been run. Thread-safety is
-    # provided on a best-effort basis to support buggy programs, and is not
-    # guaranteed by the public `tf.Graph` API.
-    #
-    # The lock must be reentrant because create_op can be called recursively due
-    # to control flow. Without a reentrant lock, many methods would also need a
-    # "locked" version or parameter (including generated code).
+    # Protects core state that can be returned via public accessors.
+    # Thread-safety is provided on a best-effort basis to support buggy
+    # programs, and is not guaranteed by the public `tf.Graph` API.
     #
     # NOTE(mrry): This does not protect the various stacks. A warning will
     # be reported if these are used from multiple threads
     self._lock = threading.RLock()
+    # The group lock synchronizes Session.run calls with methods that create
+    # and mutate ops (e.g. Graph.create_op()). This synchronization is
+    # necessary because it's illegal to modify an operation after it's been run.
+    # The group lock allows any number of threads to mutate ops at the same time
+    # but if any modification is going on, all Session.run calls have to wait.
+    # Similarly, if one or more Session.run calls are going on, all mutate ops
+    # have to wait until all Session.run calls have finished.
+    self._group_lock = lock_util.GroupLock(num_groups=2)
     self._nodes_by_id = dict()  # GUARDED_BY(self._lock)
     self._next_id_counter = 0  # GUARDED_BY(self._lock)
     self._nodes_by_name = dict()  # GUARDED_BY(self._lock)
@@ -2706,7 +2725,7 @@ class Graph(object):
     self._building_function = False
     # Stack of colocate_with ops. After switch_to_thread_local(),
     # self._thread_local._colocation_stack is used instead.
-    self._graph_colocation_stack = []
+    self._graph_colocation_stack = traceable_stack.TraceableStack()
     # Set of tensors that are dangerous to feed!
     self._unfeedable_tensors = set()
     # Set of operations that are dangerous to fetch!
@@ -2746,36 +2765,6 @@ class Graph(object):
     """Temporary hack; can be overridden to force C API usage."""
     return _USE_C_API
 
-  def _convert_stack(self, stack, include_func_start_lineno=False):
-    """Converts a stack extracted using _extract_stack() to a traceback stack.
-
-    Args:
-      stack: A list of n 5-tuples,
-        (filename, lineno, name, frame_globals, func_start_lineno).
-      include_func_start_lineno: True if function start line number should be
-        included as the 5th entry in return tuples.
-
-    Returns:
-      A list of n 4-tuples or 5-tuples
-      (filename, lineno, name, code, [optional: func_start_lineno]), where the
-      code tuple element is calculated from the corresponding elements of the
-      input tuple.
-    """
-    ret = []
-    for (filename, lineno, name, frame_globals, func_start_lineno,
-         unused_frame_info) in stack:
-      linecache.checkcache(filename)
-      line = linecache.getline(filename, lineno, frame_globals)
-      if line:
-        line = line.strip()
-      else:
-        line = None
-      if include_func_start_lineno:
-        ret.append((filename, lineno, name, line, func_start_lineno))
-      else:
-        ret.append((filename, lineno, name, line))
-    return ret
-
   # Note: this method is private because the API of tf.Graph() is public and
   # frozen, and this functionality is still not ready for public visibility.
   @tf_contextlib.contextmanager
@@ -2801,46 +2790,6 @@ class Graph(object):
   def _variable_creator_stack(self, variable_creator_stack):
     self._thread_local._variable_creator_stack = variable_creator_stack
 
-  def _extract_stack(self):
-    """A lightweight, extensible re-implementation of traceback.extract_stack.
-
-    NOTE(mrry): traceback.extract_stack eagerly retrieves the line of code for
-      each stack frame using linecache, which results in an abundance of stat()
-      calls. This implementation does not retrieve the code, and any consumer
-      should apply _convert_stack to the result to obtain a traceback that can
-      be formatted etc. using traceback methods.
-
-    Derived classes can implement _extract_frame_info() to add extra information
-    to the traceback.
-
-    Returns:
-      A list of 6-tuples
-      (filename, lineno, name, frame_globals, func_start_lineno, custom_info)
-      corresponding to the call stack of the current thread.
-    """
-    try:
-      raise ZeroDivisionError
-    except ZeroDivisionError:
-      f = sys.exc_info()[2].tb_frame.f_back
-    ret = []
-    while f is not None:
-      lineno = f.f_lineno
-      co = f.f_code
-      filename = co.co_filename
-      name = co.co_name
-      frame_globals = f.f_globals
-      func_start_lineno = co.co_firstlineno
-      frame_info = self._extract_frame_info(f)
-      ret.append((filename, lineno, name, frame_globals, func_start_lineno,
-                  frame_info))
-      f = f.f_back
-    ret.reverse()
-    return ret
-
-  def _extract_frame_info(self, frame):  # pylint: disable=unused-argument
-    """Extracts custom information from a frame in an op traceback."""
-    return None
-
   def _check_not_finalized(self):
     """Check if the graph is finalized.
 
@@ -3192,9 +3141,9 @@ class Graph(object):
 
     input_ops = set([t.op for t in inputs])
     control_inputs = self._control_dependencies_for_inputs(input_ops)
-    # _create_op_helper mutates the new Operation. _lock ensures a Session.run
-    # call cannot occur between creating and mutating the op.
-    with self._lock:
+    # _create_op_helper mutates the new Operation. `_mutation_lock` ensures a
+    # Session.run call cannot occur between creating and mutating the op.
+    with self._mutation_lock():
       ret = Operation(
           node_def,
           self,
@@ -3281,7 +3230,7 @@ class Graph(object):
 
     if self._colocation_stack:
       all_colocation_groups = []
-      for colocation_op in self._colocation_stack:
+      for colocation_op in self._colocation_stack.peek_objs():
         all_colocation_groups.extend(colocation_op.colocation_groups())
         if colocation_op.device:
           # Make this device match the device of the colocated op, to provide
@@ -4054,10 +4003,10 @@ class Graph(object):
 
     if ignore_existing:
       current_stack = self._colocation_stack
-      self._colocation_stack = []
+      self._colocation_stack = traceable_stack.TraceableStack()
 
     if op is not None:
-      self._colocation_stack.append(op)
+      self._colocation_stack.push_obj(op, name=op.name, offset=1)
 
     try:
       yield
@@ -4065,7 +4014,7 @@ class Graph(object):
       # Restore device function stack
       self._device_function_stack = device_fn_tmp
       if op is not None:
-        self._colocation_stack.pop()
+        self._colocation_stack.pop_obj()
 
       # Reset the colocation stack if requested.
       if ignore_existing:
@@ -4692,11 +4641,15 @@ class Graph(object):
 
   @property
   def _colocation_stack(self):
+    """Return thread-local copy of colocation stack."""
     if self._stack_state_is_thread_local:
       # This may be called from a thread where colocation_stack doesn't yet
       # exist.
       if not hasattr(self._thread_local, "_colocation_stack"):
-        self._thread_local._colocation_stack = self._graph_colocation_stack[:]
+        stack_copy_for_this_thread = self._graph_colocation_stack.copy()
+        # pylint: disable=protected-access
+        self._thread_local._colocation_stack = stack_copy_for_this_thread
+        # pylint: enable=protected-access
       return self._thread_local._colocation_stack
     else:
       return self._graph_colocation_stack
@@ -4727,6 +4680,20 @@ class Graph(object):
     else:
       self._graph_control_dependencies_stack = control_dependencies
 
+  def _mutation_lock(self):
+    """Returns a lock to guard code that creates & mutates ops.
+
+    See the comment for self._group_lock for more info.
+    """
+    return self._group_lock.group(_MUTATION_LOCK_GROUP)
+
+  def _session_run_lock(self):
+    """Returns a lock to guard code for Session.run.
+
+    See the comment for self._group_lock for more info.
+    """
+    return self._group_lock.group(_SESSION_RUN_LOCK_GROUP)
+
 
 # TODO(agarwal): currently device directives in an outer eager scope will not
 # apply to inner graph mode code. Fix that.
diff --git a/tensorflow/python/framework/tensor_util_test.py b/tensorflow/python/framework/tensor_util_test.py
index d6edc1364369e1b4d06093879571cdb4e9ffe409..395cf43b3f189e7ed61ab4bcf479d24de801f3ef 100644
--- a/tensorflow/python/framework/tensor_util_test.py
+++ b/tensorflow/python/framework/tensor_util_test.py
@@ -50,13 +50,13 @@ class TensorUtilTest(test.TestCase):
   def testFloatN(self):
     t = tensor_util.make_tensor_proto([10.0, 20.0, 30.0])
     if sys.byteorder == "big":
-      self.assertProtoEquals("""
+      self.assertProtoEquals(r"""
         dtype: DT_FLOAT
         tensor_shape { dim { size: 3 } }
         tensor_content: "A \000\000A\240\000\000A\360\000\000"
         """, t)
     else:
-      self.assertProtoEquals("""
+      self.assertProtoEquals(r"""
         dtype: DT_FLOAT
         tensor_shape { dim { size: 3 } }
         tensor_content: "\000\000 A\000\000\240A\000\000\360A"
@@ -68,13 +68,13 @@ class TensorUtilTest(test.TestCase):
   def testFloatTyped(self):
     t = tensor_util.make_tensor_proto([10.0, 20.0, 30.0], dtype=dtypes.float32)
     if sys.byteorder == "big":
-      self.assertProtoEquals("""
+      self.assertProtoEquals(r"""
         dtype: DT_FLOAT
         tensor_shape { dim { size: 3 } }
         tensor_content: "A \000\000A\240\000\000A\360\000\000"
         """, t)
     else:
-      self.assertProtoEquals("""
+      self.assertProtoEquals(r"""
         dtype: DT_FLOAT
         tensor_shape { dim { size: 3 } }
         tensor_content: "\000\000 A\000\000\240A\000\000\360A"
@@ -86,13 +86,13 @@ class TensorUtilTest(test.TestCase):
   def testFloatTypeCoerce(self):
     t = tensor_util.make_tensor_proto([10, 20, 30], dtype=dtypes.float32)
     if sys.byteorder == "big":
-      self.assertProtoEquals("""
+      self.assertProtoEquals(r"""
         dtype: DT_FLOAT
         tensor_shape { dim { size: 3 } }
         tensor_content: "A \000\000A\240\000\000A\360\000\000"
         """, t)
     else:
-      self.assertProtoEquals("""
+      self.assertProtoEquals(r"""
         dtype: DT_FLOAT
         tensor_shape { dim { size: 3 } }
         tensor_content: "\000\000 A\000\000\240A\000\000\360A"
@@ -105,13 +105,13 @@ class TensorUtilTest(test.TestCase):
     arr = np.asarray([10, 20, 30], dtype="int")
     t = tensor_util.make_tensor_proto(arr, dtype=dtypes.float32)
     if sys.byteorder == "big":
-      self.assertProtoEquals("""
+      self.assertProtoEquals(r"""
         dtype: DT_FLOAT
         tensor_shape { dim { size: 3 } }
         tensor_content: "A \000\000A\240\000\000A\360\000\000"
         """, t)
     else:
-      self.assertProtoEquals("""
+      self.assertProtoEquals(r"""
         dtype: DT_FLOAT
         tensor_shape { dim { size: 3 } }
         tensor_content: "\000\000 A\000\000\240A\000\000\360A"
@@ -123,13 +123,13 @@ class TensorUtilTest(test.TestCase):
   def testFloatSizes(self):
     t = tensor_util.make_tensor_proto([10.0, 20.0, 30.0], shape=[1, 3])
     if sys.byteorder == "big":
-      self.assertProtoEquals("""
+      self.assertProtoEquals(r"""
         dtype: DT_FLOAT
         tensor_shape { dim { size: 1 } dim { size: 3 } }
         tensor_content: "A \000\000A\240\000\000A\360\000\000"
         """, t)
     else:
-      self.assertProtoEquals("""
+      self.assertProtoEquals(r"""
         dtype: DT_FLOAT
         tensor_shape { dim { size: 1 } dim { size: 3 } }
         tensor_content: "\000\000 A\000\000\240A\000\000\360A"
@@ -141,13 +141,13 @@ class TensorUtilTest(test.TestCase):
   def testFloatSizes2(self):
     t = tensor_util.make_tensor_proto([10.0, 20.0, 30.0], shape=[3, 1])
     if sys.byteorder == "big":
-      self.assertProtoEquals("""
+      self.assertProtoEquals(r"""
         dtype: DT_FLOAT
         tensor_shape { dim { size: 3 } dim { size: 1 } }
         tensor_content: "A \000\000A\240\000\000A\360\000\000"
         """, t)
     else:
-      self.assertProtoEquals("""
+      self.assertProtoEquals(r"""
         dtype: DT_FLOAT
         tensor_shape { dim { size: 3 } dim { size: 1 } }
         tensor_content: "\000\000 A\000\000\240A\000\000\360A"
@@ -169,13 +169,13 @@ class TensorUtilTest(test.TestCase):
     t = tensor_util.make_tensor_proto(
         np.array([[10.0, 20.0, 30.0]], dtype=np.float64))
     if sys.byteorder == "big":
-      self.assertProtoEquals("""
+      self.assertProtoEquals(r"""
         dtype: DT_DOUBLE
         tensor_shape { dim { size: 1 } dim { size: 3 } }
         tensor_content: "@$\000\000\000\000\000\000@4\000\000\000\000\000\000@>\000\000\000\000\000\000"
         """, t)
     else:
-      self.assertProtoEquals("""
+      self.assertProtoEquals(r"""
         dtype: DT_DOUBLE
         tensor_shape { dim { size: 1 } dim { size: 3 } }
         tensor_content: "\000\000\000\000\000\000$@\000\000\000\000\000\0004@\000\000\000\000\000\000>@"
@@ -206,13 +206,13 @@ class TensorUtilTest(test.TestCase):
     self.assertEquals(np.float32, a.dtype)
     self.assertAllClose(np.array([5.0, 20.0, 30.0], dtype=np.float32), a)
     if sys.byteorder == "big":
-      self.assertProtoEquals("""
+      self.assertProtoEquals(r"""
         dtype: DT_FLOAT
         tensor_shape { dim { size: 3 } }
         tensor_content: "A \000\000A\240\000\000A\360\000\000"
         """, t)
     else:
-      self.assertProtoEquals("""
+      self.assertProtoEquals(r"""
         dtype: DT_FLOAT
         tensor_shape { dim { size: 3 } }
         tensor_content: "\000\000 A\000\000\240A\000\000\360A"
@@ -299,16 +299,16 @@ class TensorUtilTest(test.TestCase):
   def testIntNDefaultType(self):
     t = tensor_util.make_tensor_proto([10, 20, 30, 40], shape=[2, 2])
     if sys.byteorder == "big":
-      self.assertProtoEquals("""
+      self.assertProtoEquals(r"""
         dtype: DT_INT32
         tensor_shape { dim { size: 2 } dim { size: 2 } }
-        tensor_content: "\000\000\000\\n\000\000\000\024\000\000\000\036\000\000\000("
+        tensor_content: "\000\000\000\n\000\000\000\024\000\000\000\036\000\000\000("
         """, t)
     else:
-      self.assertProtoEquals("""
+      self.assertProtoEquals(r"""
         dtype: DT_INT32
         tensor_shape { dim { size: 2 } dim { size: 2 } }
-        tensor_content: "\\n\000\000\000\024\000\000\000\036\000\000\000(\000\000\000"
+        tensor_content: "\n\000\000\000\024\000\000\000\036\000\000\000(\000\000\000"
         """, t)
     a = tensor_util.MakeNdarray(t)
     self.assertEquals(np.int32, a.dtype)
@@ -380,16 +380,16 @@ class TensorUtilTest(test.TestCase):
     t = tensor_util.make_tensor_proto(
         [10, 20, 30], shape=[1, 3], dtype=dtypes.int64)
     if sys.byteorder == "big":
-      self.assertProtoEquals("""
+      self.assertProtoEquals(r"""
         dtype: DT_INT64
         tensor_shape { dim { size: 1 } dim { size: 3 } }
-        tensor_content: "\000\000\000\000\000\000\000\\n\000\000\000\000\000\000\000\024\000\000\000\000\000\000\000\036"
+        tensor_content: "\000\000\000\000\000\000\000\n\000\000\000\000\000\000\000\024\000\000\000\000\000\000\000\036"
         """, t)
     else:
-      self.assertProtoEquals("""
+      self.assertProtoEquals(r"""
         dtype: DT_INT64
         tensor_shape { dim { size: 1 } dim { size: 3 } }
-        tensor_content: "\\n\000\000\000\000\000\000\000\024\000\000\000\000\000\000\000\036\000\000\000\000\000\000\000"
+        tensor_content: "\n\000\000\000\000\000\000\000\024\000\000\000\000\000\000\000\036\000\000\000\000\000\000\000"
         """, t)
     a = tensor_util.MakeNdarray(t)
     self.assertEquals(np.int64, a.dtype)
@@ -398,16 +398,16 @@ class TensorUtilTest(test.TestCase):
   def testLongNpArray(self):
     t = tensor_util.make_tensor_proto(np.array([10, 20, 30]))
     if sys.byteorder == "big":
-      self.assertProtoEquals("""
+      self.assertProtoEquals(r"""
         dtype: DT_INT64
         tensor_shape { dim { size: 3 } }
-        tensor_content: "\000\000\000\000\000\000\000\\n\000\000\000\000\000\000\000\024\000\000\000\000\000\000\000\036"
+        tensor_content: "\000\000\000\000\000\000\000\n\000\000\000\000\000\000\000\024\000\000\000\000\000\000\000\036"
         """, t)
     else:
-      self.assertProtoEquals("""
+      self.assertProtoEquals(r"""
         dtype: DT_INT64
         tensor_shape { dim { size: 3 } }
-        tensor_content: "\\n\000\000\000\000\000\000\000\024\000\000\000\000\000\000\000\036\000\000\000\000\000\000\000"
+        tensor_content: "\n\000\000\000\000\000\000\000\024\000\000\000\000\000\000\000\036\000\000\000\000\000\000\000"
         """, t)
     a = tensor_util.MakeNdarray(t)
     self.assertEquals(np.int64, a.dtype)
@@ -419,13 +419,13 @@ class TensorUtilTest(test.TestCase):
 
     t = tensor_util.make_tensor_proto(data, dtype=dtypes.qint32)
     if sys.byteorder == "big":
-      self.assertProtoEquals("""
+      self.assertProtoEquals(r"""
         dtype: DT_QINT32
         tensor_shape { dim { size: 3 } }
         tensor_content: "\000\000\000\025\000\000\000\026\000\000\000\027"
         """, t)
     else:
-      self.assertProtoEquals("""
+      self.assertProtoEquals(r"""
         dtype: DT_QINT32
         tensor_shape { dim { size: 3 } }
         tensor_content: "\025\000\000\000\026\000\000\000\027\000\000\000"
@@ -435,7 +435,7 @@ class TensorUtilTest(test.TestCase):
     self.assertAllEqual(np.array(data, dtype=a.dtype), a)
 
     t = tensor_util.make_tensor_proto(data, dtype=dtypes.quint8)
-    self.assertProtoEquals("""
+    self.assertProtoEquals(r"""
       dtype: DT_QUINT8
       tensor_shape { dim { size: 3 } }
       tensor_content: "\025\026\027"
@@ -445,7 +445,7 @@ class TensorUtilTest(test.TestCase):
     self.assertAllEqual(np.array(data, dtype=a.dtype), a)
 
     t = tensor_util.make_tensor_proto(data, dtype=dtypes.qint8)
-    self.assertProtoEquals("""
+    self.assertProtoEquals(r"""
       dtype: DT_QINT8
       tensor_shape { dim { size: 3 } }
       tensor_content: "\025\026\027"
@@ -456,13 +456,13 @@ class TensorUtilTest(test.TestCase):
 
     t = tensor_util.make_tensor_proto(data, dtype=dtypes.quint16)
     if sys.byteorder == "big":
-      self.assertProtoEquals("""
+      self.assertProtoEquals(r"""
         dtype: DT_QUINT16
         tensor_shape { dim { size: 3 } }
         tensor_content: "\000\025\000\026\000\027"
         """, t)
     else:
-      self.assertProtoEquals("""
+      self.assertProtoEquals(r"""
         dtype: DT_QUINT16
         tensor_shape { dim { size: 3 } }
         tensor_content: "\025\000\026\000\027\000"
@@ -473,13 +473,13 @@ class TensorUtilTest(test.TestCase):
 
     t = tensor_util.make_tensor_proto(data, dtype=dtypes.qint16)
     if sys.byteorder == "big":
-      self.assertProtoEquals("""
+      self.assertProtoEquals(r"""
         dtype: DT_QINT16
         tensor_shape { dim { size: 3 } }
         tensor_content: "\000\025\000\026\000\027"
         """, t)
     else:
-      self.assertProtoEquals("""
+      self.assertProtoEquals(r"""
         dtype: DT_QINT16
         tensor_shape { dim { size: 3 } }
         tensor_content: "\025\000\026\000\027\000"
diff --git a/tensorflow/python/framework/test_util.py b/tensorflow/python/framework/test_util.py
index 3988238609754d84892ed9284f346e462a0d721e..2bc2a189fa8e825613ca834e2c06ea916074d455 100644
--- a/tensorflow/python/framework/test_util.py
+++ b/tensorflow/python/framework/test_util.py
@@ -27,6 +27,7 @@ import random
 import re
 import tempfile
 import threading
+import unittest
 
 import numpy as np
 import six
@@ -414,8 +415,28 @@ def assert_no_new_pyobjects_executing_eagerly(f):
       f(self, **kwargs)
       gc.collect()
       previous_count = len(gc.get_objects())
+      collection_sizes_before = {
+          collection: len(ops.get_collection(collection))
+          for collection in ops.get_default_graph().collections}
       for _ in range(3):
         f(self, **kwargs)
+      # Note that gc.get_objects misses anything that isn't subject to garbage
+      # collection (C types). Collections are a common source of leaks, so we
+      # test for collection sizes explicitly.
+      for collection_key in ops.get_default_graph().collections:
+        collection = ops.get_collection(collection_key)
+        size_before = collection_sizes_before.get(collection_key, 0)
+        if len(collection) > size_before:
+          raise AssertionError(
+              ("Collection %s increased in size from "
+               "%d to %d (current items %s).")
+              % (collection_key, size_before, len(collection), collection))
+        # Make sure our collection checks don't show up as leaked memory by
+        # removing references to temporary variables.
+        del collection
+        del collection_key
+        del size_before
+      del collection_sizes_before
       gc.collect()
       # There should be no new Python objects hanging around.
       new_count = len(gc.get_objects())
@@ -625,16 +646,12 @@ def run_in_graph_and_eager_modes(func=None,
           "Did you mean to use `run_all_tests_in_graph_and_eager_modes`?")
 
     def decorated(self, **kwargs):
-      with context.graph_mode():
-        with self.test_session(use_gpu=use_gpu, config=config):
-          f(self, **kwargs)
-
-      if reset_test:
-        # This decorator runs the wrapped test twice.
-        # Reset the test environment between runs.
-        self.tearDown()
-        self._tempdir = None
-        self.setUp()
+      try:
+        with context.graph_mode():
+          with self.test_session(use_gpu=use_gpu, config=config):
+            f(self, **kwargs)
+      except unittest.case.SkipTest:
+        pass
 
       def run_eagerly(self, **kwargs):
         if not use_gpu:
@@ -649,6 +666,13 @@ def run_in_graph_and_eager_modes(func=None,
             assert_no_garbage_created(run_eagerly))
 
       with context.eager_mode():
+        if reset_test:
+          # This decorator runs the wrapped test twice.
+          # Reset the test environment between runs.
+          self.tearDown()
+          self._tempdir = None
+          self.setUp()
+
         run_eagerly(self, **kwargs)
 
     return decorated
diff --git a/tensorflow/python/framework/test_util_test.py b/tensorflow/python/framework/test_util_test.py
index 54983761812a624318a8675d5ec4e704ad6d8034..122c14c8473f133f6a3bed1e6297394eaa1b845c 100644
--- a/tensorflow/python/framework/test_util_test.py
+++ b/tensorflow/python/framework/test_util_test.py
@@ -616,7 +616,7 @@ class TestUtilTest(test_util.TensorFlowTestCase):
     self.assertIs(test_util.get_node_def_from_graph("foo", graph_def), node_foo)
     self.assertIsNone(test_util.get_node_def_from_graph("bar", graph_def))
 
-  def testRunInGraphAndEagerModesOnTestCase(self):
+  def test_run_in_eager_and_graph_modes_test_class(self):
     msg = "`run_test_in_graph_and_eager_modes` only supports test methods.*"
     with self.assertRaisesRegexp(ValueError, msg):
       @test_util.run_in_graph_and_eager_modes()
@@ -624,6 +624,47 @@ class TestUtilTest(test_util.TensorFlowTestCase):
         pass
       del Foo  # Make pylint unused happy.
 
+  def test_run_in_eager_and_graph_modes_skip_graph_runs_eager(self):
+    modes = []
+    def _test(self):
+      if not context.executing_eagerly():
+        self.skipTest("Skipping in graph mode")
+      modes.append("eager" if context.executing_eagerly() else "graph")
+    test_util.run_in_graph_and_eager_modes(_test)(self)
+    self.assertEqual(modes, ["eager"])
+
+  def test_run_in_eager_and_graph_modes_skip_eager_runs_graph(self):
+    modes = []
+    def _test(self):
+      if context.executing_eagerly():
+        self.skipTest("Skipping in eager mode")
+      modes.append("eager" if context.executing_eagerly() else "graph")
+    test_util.run_in_graph_and_eager_modes(_test)(self)
+    self.assertEqual(modes, ["graph"])
+
+  def test_run_in_graph_and_eager_modes_setup_in_same_mode(self):
+    modes = []
+    mode_name = lambda: "eager" if context.executing_eagerly() else "graph"
+
+    class ExampleTest(test_util.TensorFlowTestCase):
+
+      def runTest(self):
+        pass
+
+      def setUp(self):
+        modes.append("setup_" + mode_name())
+
+      @test_util.run_in_graph_and_eager_modes
+      def testBody(self):
+        modes.append("run_" + mode_name())
+
+    e = ExampleTest()
+    e.setUp()
+    e.testBody()
+
+    self.assertEqual(modes[0:2], ["setup_graph", "run_graph"])
+    self.assertEqual(modes[2:], ["setup_eager", "run_eager"])
+
 
 class GarbageCollectionTest(test_util.TensorFlowTestCase):
 
diff --git a/tensorflow/python/framework/traceable_stack.py b/tensorflow/python/framework/traceable_stack.py
new file mode 100644
index 0000000000000000000000000000000000000000..1b7c6bd7c56e40cfa632c6d73c04e958ce24363e
--- /dev/null
+++ b/tensorflow/python/framework/traceable_stack.py
@@ -0,0 +1,135 @@
+# Copyright 2015 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""A simple stack that associates filename and line numbers with each object."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+from tensorflow.python.util import tf_stack
+
+
+class TraceableObject(object):
+  """Wrap an object together with its the code definition location."""
+
+  # Return codes for the set_filename_and_line_from_caller() method.
+  SUCCESS, HEURISTIC_USED, FAILURE = (0, 1, 2)
+
+  def __init__(self, obj, name=None, filename=None, lineno=None):
+    self.obj = obj
+    self.name = name
+    self.filename = filename
+    self.lineno = lineno
+
+  def set_filename_and_line_from_caller(self, offset=0):
+    """Set filename and line using the caller's stack frame.
+
+    If the requested stack information is not available, a heuristic may
+    be applied and self.HEURISTIC USED will be returned.  If the heuristic
+    fails then no change will be made to the filename and lineno members
+    (None by default) and self.FAILURE will be returned.
+
+    Args:
+      offset: Integer.  If 0, the caller's stack frame is used.  If 1,
+          the caller's caller's stack frame is used.  Larger values are
+          permissible but if out-of-range (larger than the number of stack
+          frames available) the outermost stack frame will be used.
+
+    Returns:
+      TraceableObject.SUCCESS if appropriate stack information was found,
+      TraceableObject.HEURISTIC_USED if the offset was larger than the stack,
+      and TraceableObject.FAILURE if the stack was empty.
+    """
+    # Offset is defined in "Args" as relative to the caller.  We are one frame
+    # beyond the caller.
+    local_offset = offset + 1
+
+    frame_records = tf_stack.extract_stack()
+    if not frame_records:
+      return self.FAILURE
+    if len(frame_records) >= local_offset:
+      # Negative indexing is one-indexed instead of zero-indexed.
+      negative_offset = -(local_offset + 1)
+      self.filename, self.lineno = frame_records[negative_offset][:2]
+      return self.SUCCESS
+    else:
+      # If the offset is too large then we use the largest offset possible,
+      # meaning we use the outermost stack frame at index 0.
+      self.filename, self.lineno = frame_records[0][:2]
+      return self.HEURISTIC_USED
+
+  def copy_metadata(self):
+    """Return a TraceableObject like this one, but without the object."""
+    return self.__class__(None, name=self.name, filename=self.filename,
+                          lineno=self.lineno)
+
+
+class TraceableStack(object):
+  """A stack of TraceableObjects."""
+
+  def __init__(self, existing_stack=None):
+    """Constructor.
+
+    Args:
+      existing_stack: [TraceableObject, ...] If provided, this object will
+        set its new stack to a SHALLOW COPY of existing_stack.
+    """
+    self._stack = existing_stack[:] if existing_stack else []
+
+  def push_obj(self, obj, name=None, offset=0):
+    """Add object to the stack and record its filename and line information.
+
+    Args:
+      obj: An object to store on the stack.
+      name: A name for the object, used for dict keys in get_item_metadata_dict.
+      offset: Integer.  If 0, the caller's stack frame is used.  If 1,
+          the caller's caller's stack frame is used.
+
+    Returns:
+      TraceableObject.SUCCESS if appropriate stack information was found,
+      TraceableObject.HEURISTIC_USED if the stack was smaller than expected,
+      and TraceableObject.FAILURE if the stack was empty.
+    """
+    traceable_obj = TraceableObject(obj, name=name)
+    self._stack.append(traceable_obj)
+    # Offset is defined in "Args" as relative to the caller.  We are 1 frame
+    # beyond the caller and need to compensate.
+    return traceable_obj.set_filename_and_line_from_caller(offset + 1)
+
+  def pop_obj(self):
+    """Remove last-inserted object and return it, without filename/line info."""
+    return self._stack.pop().obj
+
+  def peek_objs(self):
+    """Return list of stored objects ordered newest to oldest."""
+    return [t_obj.obj for t_obj in reversed(self._stack)]
+
+  def peek_traceable_objs(self):
+    """Return list of stored TraceableObjects ordered newest to oldest."""
+    return list(reversed(self._stack))
+
+  def __len__(self):
+    """Return number of items on the stack, and used for truth-value testing."""
+    return len(self._stack)
+
+  def copy(self):
+    """Return a copy of self referencing the same objects but in a new list.
+
+    This method is implemented to support thread-local stacks.
+
+    Returns:
+      TraceableStack with a new list that holds existing objects.
+    """
+    return TraceableStack(self._stack)
diff --git a/tensorflow/python/framework/traceable_stack_test.py b/tensorflow/python/framework/traceable_stack_test.py
new file mode 100644
index 0000000000000000000000000000000000000000..3e7876f6318da368a373ca554e674a21b0d869c3
--- /dev/null
+++ b/tensorflow/python/framework/traceable_stack_test.py
@@ -0,0 +1,133 @@
+# Copyright 2015 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""Tests for tensorflow.python.framework.traceable_stack."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+from tensorflow.python.framework import test_util
+from tensorflow.python.framework import traceable_stack
+from tensorflow.python.platform import googletest
+from tensorflow.python.util import tf_inspect as inspect
+
+_LOCAL_OBJECT = lambda x: x
+_THIS_FILENAME = inspect.getsourcefile(_LOCAL_OBJECT)
+
+
+class TraceableObjectTest(test_util.TensorFlowTestCase):
+
+  def testSetFilenameAndLineFromCallerUsesCallersStack(self):
+    t_obj = traceable_stack.TraceableObject(17)
+
+    # Do not separate placeholder from the set_filename_and_line_from_caller()
+    # call one line below it as it is used to calculate the latter's line
+    # number.
+    placeholder = lambda x: x
+    result = t_obj.set_filename_and_line_from_caller()
+
+    expected_lineno = inspect.getsourcelines(placeholder)[1] + 1
+    self.assertEqual(expected_lineno, t_obj.lineno)
+    self.assertEqual(_THIS_FILENAME, t_obj.filename)
+    self.assertEqual(t_obj.SUCCESS, result)
+
+  def testSetFilenameAndLineFromCallerRespectsOffset(self):
+
+    def call_set_filename_and_line_from_caller(t_obj):
+      # We expect to retrieve the line number from _our_ caller.
+      return t_obj.set_filename_and_line_from_caller(offset=1)
+
+    t_obj = traceable_stack.TraceableObject(None)
+    # Do not separate placeholder from the
+    # call_set_filename_and_line_from_caller() call one line below it as it is
+    # used to calculate the latter's line number.
+    placeholder = lambda x: x
+    result = call_set_filename_and_line_from_caller(t_obj)
+
+    expected_lineno = inspect.getsourcelines(placeholder)[1] + 1
+    self.assertEqual(expected_lineno, t_obj.lineno)
+    self.assertEqual(t_obj.SUCCESS, result)
+
+  def testSetFilenameAndLineFromCallerHandlesRidiculousOffset(self):
+    t_obj = traceable_stack.TraceableObject('The quick brown fox.')
+    # This line shouldn't die.
+    result = t_obj.set_filename_and_line_from_caller(offset=300)
+
+    # We expect a heuristic to be used because we are not currently 300 frames
+    # down on the stack.  The filename and lineno of the outermost frame are not
+    # predictable -- in some environments the filename is this test file, but in
+    # other environments it is not (e.g. due to a test runner calling this
+    # file).  Therefore we only test that the called function knows it applied a
+    # heuristic for the ridiculous stack offset.
+    self.assertEqual(t_obj.HEURISTIC_USED, result)
+
+
+class TraceableStackTest(test_util.TensorFlowTestCase):
+
+  def testPushPeekPopObj(self):
+    t_stack = traceable_stack.TraceableStack()
+    t_stack.push_obj(42.0)
+    t_stack.push_obj('hope')
+
+    expected_lifo_peek = ['hope', 42.0]
+    self.assertEqual(expected_lifo_peek, t_stack.peek_objs())
+
+    self.assertEqual('hope', t_stack.pop_obj())
+    self.assertEqual(42.0, t_stack.pop_obj())
+
+  def testPushPopPreserveLifoOrdering(self):
+    t_stack = traceable_stack.TraceableStack()
+    t_stack.push_obj(0)
+    t_stack.push_obj(1)
+    t_stack.push_obj(2)
+    t_stack.push_obj(3)
+
+    obj_3 = t_stack.pop_obj()
+    obj_2 = t_stack.pop_obj()
+    obj_1 = t_stack.pop_obj()
+    obj_0 = t_stack.pop_obj()
+
+    self.assertEqual(3, obj_3)
+    self.assertEqual(2, obj_2)
+    self.assertEqual(1, obj_1)
+    self.assertEqual(0, obj_0)
+
+  def testPushObjSetsFilenameAndLineInfoForCaller(self):
+    t_stack = traceable_stack.TraceableStack()
+
+    # We expect that the line number recorded for the 1-object will come from
+    # the call to t_stack.push_obj(1).  Do not separate the next two lines!
+    placeholder_1 = lambda x: x
+    t_stack.push_obj(1)
+
+    # We expect that the line number recorded for the 2-object will come from
+    # the call to call_push_obj() and _not_ the call to t_stack.push_obj().
+    def call_push_obj(obj):
+      t_stack.push_obj(obj, offset=1)
+
+    # Do not separate the next two lines!
+    placeholder_2 = lambda x: x
+    call_push_obj(2)
+
+    expected_lineno_1 = inspect.getsourcelines(placeholder_1)[1] + 1
+    expected_lineno_2 = inspect.getsourcelines(placeholder_2)[1] + 1
+
+    t_obj_2, t_obj_1 = t_stack.peek_traceable_objs()
+    self.assertEqual(expected_lineno_2, t_obj_2.lineno)
+    self.assertEqual(expected_lineno_1, t_obj_1.lineno)
+
+
+if __name__ == '__main__':
+  googletest.main()
diff --git a/tensorflow/python/keras/BUILD b/tensorflow/python/keras/BUILD
index 8b6b28bc776fa500a93d0a3fb3bf91081ba86967..4056818a951907283a82d924650aca5de74383de 100755
--- a/tensorflow/python/keras/BUILD
+++ b/tensorflow/python/keras/BUILD
@@ -451,6 +451,7 @@ cuda_py_test(
         "//tensorflow/python:client_testlib",
     ],
     shard_count = 2,
+    tags = ["no_windows_gpu"],
 )
 
 py_test(
@@ -720,6 +721,7 @@ py_test(
     size = "medium",
     srcs = ["preprocessing/image_test.py"],
     srcs_version = "PY2AND3",
+    tags = ["nomsan"],  # TODO(b/110990716) reenable
     deps = [
         ":keras",
         "//tensorflow/python:client_testlib",
diff --git a/tensorflow/python/keras/backend.py b/tensorflow/python/keras/backend.py
index fed779650ebd580cc9c3830014102133c0c901ab..cb3423598b4740c20cc3cbd3f0e4c997f0eaf858 100644
--- a/tensorflow/python/keras/backend.py
+++ b/tensorflow/python/keras/backend.py
@@ -963,13 +963,14 @@ def zeros(shape, dtype=None, name=None):
              [ 0.,  0.,  0.,  0.]], dtype=float32)
   ```
   """
-  if dtype is None:
-    dtype = floatx()
-  tf_dtype = dtypes_module.as_dtype(dtype)
-  v = array_ops.zeros(shape=shape, dtype=tf_dtype, name=name)
-  if py_all(v.get_shape().as_list()):
-    return variable(v, dtype=dtype, name=name)
-  return v
+  with ops.init_scope():
+    if dtype is None:
+      dtype = floatx()
+    tf_dtype = dtypes_module.as_dtype(dtype)
+    v = array_ops.zeros(shape=shape, dtype=tf_dtype, name=name)
+    if py_all(v.get_shape().as_list()):
+      return variable(v, dtype=dtype, name=name)
+    return v
 
 
 @tf_export('keras.backend.ones')
@@ -996,13 +997,14 @@ def ones(shape, dtype=None, name=None):
              [ 1.,  1.,  1.,  1.]], dtype=float32)
   ```
   """
-  if dtype is None:
-    dtype = floatx()
-  tf_dtype = dtypes_module.as_dtype(dtype)
-  v = array_ops.ones(shape=shape, dtype=tf_dtype, name=name)
-  if py_all(v.get_shape().as_list()):
-    return variable(v, dtype=dtype, name=name)
-  return v
+  with ops.init_scope():
+    if dtype is None:
+      dtype = floatx()
+    tf_dtype = dtypes_module.as_dtype(dtype)
+    v = array_ops.ones(shape=shape, dtype=tf_dtype, name=name)
+    if py_all(v.get_shape().as_list()):
+      return variable(v, dtype=dtype, name=name)
+    return v
 
 
 @tf_export('keras.backend.eye')
@@ -2795,10 +2797,15 @@ class Function(object):
     if not isinstance(self.fetches, list):
       self.fetches = [self.fetches]
     # The main use case of `fetches` being passed to a model is the ability
-    # to run custom updates (since the outputs of fetches are never returned).
+    # to run custom updates
     # This requires us to wrap fetches in `identity` ops.
     self.fetches = [array_ops.identity(x) for x in self.fetches]
     self.session_kwargs = session_kwargs
+    # This mapping keeps track of the function that should receive the
+    # output from a fetch in `fetches`: { fetch: function(fetch_output) }
+    # A Callback can use this to register a function with access to the
+    # output values for a fetch it added.
+    self.fetch_callbacks = dict()
 
     if session_kwargs:
       raise ValueError('Some keys in session_kwargs are not supported at this '
@@ -2808,6 +2815,7 @@ class Function(object):
     self._feed_arrays = None
     self._feed_symbols = None
     self._symbol_vals = None
+    self._fetches = None
     self._session = None
 
   def _make_callable(self, feed_arrays, feed_symbols, symbol_vals, session):
@@ -2853,8 +2861,14 @@ class Function(object):
     self._feed_arrays = feed_arrays
     self._feed_symbols = feed_symbols
     self._symbol_vals = symbol_vals
+    self._fetches = list(self.fetches)
     self._session = session
 
+  def _call_fetch_callbacks(self, fetches_output):
+    for fetch, output in zip(self._fetches, fetches_output):
+      if fetch in self.fetch_callbacks:
+        self.fetch_callbacks[fetch](output)
+
   def __call__(self, inputs):
     if not isinstance(inputs, (list, tuple)):
       raise TypeError('`inputs` should be a list or tuple.')
@@ -2891,14 +2905,14 @@ class Function(object):
             np.asarray(self.feed_dict[key], dtype=key.dtype.base_dtype.name))
 
     # Refresh callable if anything has changed.
-    if (self._callable_fn is None or
-        feed_arrays != self._feed_arrays or
+    if (self._callable_fn is None or feed_arrays != self._feed_arrays or
         symbol_vals != self._symbol_vals or
-        feed_symbols != self._feed_symbols or
+        feed_symbols != self._feed_symbols or self.fetches != self._fetches or
         session != self._session):
       self._make_callable(feed_arrays, feed_symbols, symbol_vals, session)
 
     fetched = self._callable_fn(*array_vals)
+    self._call_fetch_callbacks(fetched[-len(self._fetches):])
     return fetched[:len(self.outputs)]
 
 
@@ -3161,10 +3175,16 @@ def rnn(step_function,
                                       array_ops.stack(
                                           [1, array_ops.shape(output)[1]]))
         output = array_ops.where(tiled_mask_t, output, states[0])
-        new_states = [
-            array_ops.where(tiled_mask_t, new_states[i], states[i])
-            for i in range(len(states))
-        ]
+
+        masked_states = []
+        for i in range(len(states)):
+          states_dim = array_ops.shape(new_states[i])[1]
+          stacked_states_dim = array_ops.stack([1, states_dim])
+          tiled_mask = array_ops.tile(mask_t, stacked_states_dim)
+          masked_state = array_ops.where(tiled_mask, new_states[i], states[i])
+          masked_states.append(masked_state)
+        new_states = masked_states
+
         output_ta_t = output_ta_t.write(time, output)
         return (time + 1, output_ta_t) + tuple(new_states)
     else:
diff --git a/tensorflow/python/keras/backend_test.py b/tensorflow/python/keras/backend_test.py
index 2ba6c8ef15ecadb8cf62eb70617623001435c2fb..36478ea089a871667908d70e33422aef8444a3e4 100644
--- a/tensorflow/python/keras/backend_test.py
+++ b/tensorflow/python/keras/backend_test.py
@@ -276,6 +276,36 @@ class BackendUtilsTest(test.TestCase):
       self.assertEqual(
           keras.backend.get_session().run(fetches=[x, y]), [30., 40.])
 
+  def test_function_fetch_callbacks(self):
+
+    class CallbackStub(object):
+
+      def __init__(self):
+        self.times_called = 0
+        self.callback_result = 0
+
+      def _fetch_callback(self, result):
+        self.times_called += 1
+        self.callback_result = result
+
+    with self.test_session():
+      callback = CallbackStub()
+      x_placeholder = keras.backend.placeholder(shape=())
+      y_placeholder = keras.backend.placeholder(shape=())
+
+      callback_op = x_placeholder * y_placeholder
+
+      f = keras.backend.function(
+          inputs=[x_placeholder, y_placeholder],
+          outputs=[x_placeholder + y_placeholder])
+      f.fetches.append(callback_op)
+      f.fetch_callbacks[callback_op] = callback._fetch_callback
+
+      _ = f([10., 20.])
+
+      self.assertEqual(callback.times_called, 1)
+      self.assertEqual(callback.callback_result, 200)
+
 
 class BackendVariableTest(test.TestCase):
 
@@ -1077,7 +1107,7 @@ class BackendNNOpsTest(test.TestCase, parameterized.TestCase):
         {'go_backwards': False, 'mask': mask, 'unroll': True},
     ]
     with self.test_session():
-      for (i, kwargs) in enumerate(kwargs_list):
+      for i, kwargs in enumerate(kwargs_list):
         last_output, outputs, new_states = keras.backend.rnn(rnn_fn, inputs,
                                                              initial_states,
                                                              **kwargs)
@@ -1124,6 +1154,115 @@ class BackendNNOpsTest(test.TestCase, parameterized.TestCase):
       for b_s, b_u_s in zip(state_list[2], state_list[3]):
         self.assertAllClose(b_s, b_u_s, atol=1e-04)
 
+  def test_rnn_additional_states(self):
+    # implement a simple RNN
+    num_samples = 4
+    input_dim = 5
+    output_dim = 3
+    timesteps = 6
+
+    input_val = np.random.random(
+        (num_samples, timesteps, input_dim)).astype(np.float32)
+    init_state_val = np.random.random(
+        (num_samples, output_dim)).astype(np.float32)
+    w_i_val = np.random.random((input_dim, output_dim)).astype(np.float32)
+    w_o_val = np.random.random((output_dim, output_dim)).astype(np.float32)
+    np_mask = np.random.randint(2, size=(num_samples, timesteps))
+
+    def rnn_step_fn():
+      w_i = keras.backend.variable(w_i_val)
+      w_o = keras.backend.variable(w_o_val)
+
+      def step_function(x, states):
+        assert len(states) == 2
+        prev_output = states[0]
+        output = keras.backend.dot(x, w_i) + keras.backend.dot(prev_output, w_o)
+        return output, [output,
+                        keras.backend.concatenate([output, output], axis=-1)]
+
+      return step_function
+
+    # test default setup
+    last_output_list = [[], [], [], [], [], []]
+    outputs_list = [[], [], [], [], [], []]
+    state_list = [[], [], [], [], [], []]
+    additional_state_list = [[], [], [], [], [], []]
+
+    rnn_fn = rnn_step_fn()
+    inputs = keras.backend.variable(input_val)
+    initial_states = [keras.backend.variable(init_state_val),
+                      np.concatenate([init_state_val, init_state_val], axis=-1)]
+    mask = keras.backend.variable(np_mask)
+
+    kwargs_list = [
+        {'go_backwards': False, 'mask': None},
+        {'go_backwards': False, 'mask': None, 'unroll': True},
+        {'go_backwards': True, 'mask': None},
+        {'go_backwards': True, 'mask': None, 'unroll': True},
+        {'go_backwards': False, 'mask': mask},
+        {'go_backwards': False, 'mask': mask, 'unroll': True},
+    ]
+    with self.test_session():
+      for i, kwargs in enumerate(kwargs_list):
+        last_output, outputs, new_states = keras.backend.rnn(rnn_fn, inputs,
+                                                             initial_states,
+                                                             **kwargs)
+        # check static shape inference
+        self.assertEqual(last_output.get_shape().as_list(),
+                         [num_samples, output_dim])
+        self.assertEqual(outputs.get_shape().as_list(),
+                         [num_samples, timesteps, output_dim])
+        # for state in new_states:
+        #   self.assertEquals(state.get_shape().as_list(),
+        #                     [num_samples, output_dim])
+        self.assertEqual(new_states[0].get_shape().as_list(),
+                         [num_samples, output_dim])
+        self.assertEqual(new_states[1].get_shape().as_list(),
+                         [num_samples, 2 * output_dim])
+
+        last_output_list[i].append(keras.backend.eval(last_output))
+        outputs_list[i].append(keras.backend.eval(outputs))
+        self.assertEqual(len(new_states), 2)
+        state_list[i].append(keras.backend.eval(new_states[0]))
+        additional_state_list[i].append(keras.backend.eval(new_states[1]))
+
+      def assert_list_pairwise(z_list, atol=1e-05):
+        for (z1, z2) in zip(z_list[1:], z_list[:-1]):
+          self.assertAllClose(z1, z2, atol=atol)
+
+      assert_list_pairwise(last_output_list[0], atol=1e-04)
+      assert_list_pairwise(outputs_list[0], atol=1e-04)
+      assert_list_pairwise(state_list[0], atol=1e-04)
+      assert_list_pairwise(additional_state_list[0], atol=1e-04)
+      assert_list_pairwise(last_output_list[2], atol=1e-04)
+      assert_list_pairwise(outputs_list[2], atol=1e-04)
+      assert_list_pairwise(state_list[2], atol=1e-04)
+      assert_list_pairwise(additional_state_list[2], atol=1e-04)
+
+      for l, u_l in zip(last_output_list[0], last_output_list[1]):
+        self.assertAllClose(l, u_l, atol=1e-04)
+
+      for o, u_o in zip(outputs_list[0], outputs_list[1]):
+        self.assertAllClose(o, u_o, atol=1e-04)
+
+      for s, u_s in zip(state_list[0], state_list[1]):
+        self.assertAllClose(s, u_s, atol=1e-04)
+
+      for s, u_s in zip(additional_state_list[0], additional_state_list[1]):
+        self.assertAllClose(s, u_s, atol=1e-04)
+
+      for b_l, b_u_l in zip(last_output_list[2], last_output_list[3]):
+        self.assertAllClose(b_l, b_u_l, atol=1e-04)
+
+      for b_o, b_u_o in zip(outputs_list[2], outputs_list[3]):
+        self.assertAllClose(b_o, b_u_o, atol=1e-04)
+
+      for b_s, b_u_s in zip(state_list[2], state_list[3]):
+        self.assertAllClose(b_s, b_u_s, atol=1e-04)
+
+      for s, u_s in zip(additional_state_list[2], additional_state_list[3]):
+        self.assertAllClose(s, u_s, atol=1e-04)
+
   def test_normalize_batch_in_training(self):
     val = np.random.random((10, 3, 10, 10))
     x = keras.backend.variable(val)
diff --git a/tensorflow/python/keras/callbacks.py b/tensorflow/python/keras/callbacks.py
index 9f91368e5bd772b47ac951a600f458126c1e12a6..5d66db232afb35df81707c106780370783eadf31 100644
--- a/tensorflow/python/keras/callbacks.py
+++ b/tensorflow/python/keras/callbacks.py
@@ -24,6 +24,7 @@ from collections import Iterable
 from collections import OrderedDict
 import csv
 import json
+import math
 import os
 import time
 
@@ -31,8 +32,10 @@ import numpy as np
 import six
 
 from tensorflow.python.keras import backend as K
+from tensorflow.python.keras import optimizers
 from tensorflow.python.keras.utils.generic_utils import Progbar
 from tensorflow.python.ops import array_ops
+from tensorflow.python.ops.resource_variable_ops import ResourceVariable as Variable
 from tensorflow.python.platform import tf_logging as logging
 from tensorflow.python.summary import summary as tf_summary
 from tensorflow.python.util.tf_export import tf_export
@@ -496,6 +499,9 @@ class EarlyStopping(Callback):
           monitored has stopped increasing; in `auto`
           mode, the direction is automatically inferred
           from the name of the monitored quantity.
+      baseline: baseline value for the monitored quantity.
+          Training will stop if the model doesn't show improvement over the
+          baseline.
   """
 
   def __init__(self,
@@ -503,13 +509,15 @@ class EarlyStopping(Callback):
                min_delta=0,
                patience=0,
                verbose=0,
-               mode='auto'):
+               mode='auto',
+               baseline=None):
     super(EarlyStopping, self).__init__()
 
     self.monitor = monitor
     self.patience = patience
     self.verbose = verbose
-    self.min_delta = min_delta
+    self.baseline = baseline
+    self.min_delta = abs(min_delta)
     self.wait = 0
     self.stopped_epoch = 0
 
@@ -537,7 +545,10 @@ class EarlyStopping(Callback):
     # Allow instances to be re-used
     self.wait = 0
     self.stopped_epoch = 0
-    self.best = np.Inf if self.monitor_op == np.less else -np.Inf
+    if self.baseline is not None:
+      self.best = self.baseline
+    else:
+      self.best = np.Inf if self.monitor_op == np.less else -np.Inf
 
   def on_epoch_end(self, epoch, logs=None):
     current = logs.get(self.monitor)
@@ -633,17 +644,35 @@ class LearningRateScheduler(Callback):
     self.verbose = verbose
 
   def on_epoch_begin(self, epoch, logs=None):
-    if not hasattr(self.model.optimizer, 'lr'):
-      raise ValueError('Optimizer must have a "lr" attribute.')
+    # TODO(yashkatariya): Change the property checking when the learning
+    # rate attribute is unified across all TF Optimizers.
+    if isinstance(self.model.optimizer, optimizers.TFOptimizer):
+      if not hasattr(self.model.optimizer.optimizer, '_lr') and not hasattr(
+          self.model.optimizer.optimizer, '_learning_rate'):
+        raise ValueError(
+            'TF Optimizer must have a "_lr" or "_learning_rate" attribute.')
+      else:
+        opt = self.model.optimizer.optimizer
+        if hasattr(opt, '_lr'):
+          opt_lr = Variable(opt._lr)  # pylint: disable=protected-access
+        elif hasattr(opt, '_learning_rate'):
+          opt_lr = Variable(opt._learning_rate)  # pylint: disable=protected-access
+    else:
+      if not hasattr(self.model.optimizer, 'lr'):
+        raise ValueError('Optimizer must have a "lr" attribute.')
+      else:
+        opt = self.model.optimizer
+        opt_lr = opt.lr
+
     try:  # new API
-      lr = float(K.get_value(self.model.optimizer.lr))
+      lr = float(K.get_value(opt_lr))
       lr = self.schedule(epoch, lr)
     except TypeError:  # Support for old API for backward compatibility
       lr = self.schedule(epoch)
     if not isinstance(lr, (float, np.float32, np.float64)):
       raise ValueError('The output of the "schedule" function '
                        'should be float.')
-    K.set_value(self.model.optimizer.lr, lr)
+    K.set_value(opt_lr, lr)
     if self.verbose > 0:
       print('\nEpoch %05d: LearningRateScheduler reducing learning '
             'rate to %s.' % (epoch + 1, lr))
@@ -715,10 +744,16 @@ class TensorBoard(Callback):
     self.write_grads = write_grads
     self.write_images = write_images
     self.batch_size = batch_size
+    self._current_batch = 0
+    # abstracted writer class to be able to stub for testing
+    self._writer_class = tf_summary.FileWriter
 
   def set_model(self, model):
+    """Sets Keras model and creates summary ops."""
+
     self.model = model
     self.sess = K.get_session()
+    # only make histogram summary op if it hasn't already been made
     if self.histogram_freq and self.merged is None:
       for layer in self.model.layers:
         for weight in layer.weights:
@@ -767,54 +802,56 @@ class TensorBoard(Callback):
     self.merged = tf_summary.merge_all()
 
     if self.write_graph:
-      self.writer = tf_summary.FileWriter(self.log_dir, self.sess.graph)
+      self.writer = self._writer_class(self.log_dir, self.sess.graph)
     else:
-      self.writer = tf_summary.FileWriter(self.log_dir)
+      self.writer = self._writer_class(self.log_dir)
+
+  def _fetch_callback(self, summary):
+    self.writer.add_summary(
+        summary,
+        self._epoch + self._current_val_batch / self._validation_batches)
+    self._current_val_batch += 1
+
+  def on_train_begin(self, logs=None):
+    """Checks if histogram summaries can be run."""
+
+    if self.histogram_freq:
+      if 'validation_steps' in self.params:
+        self._validation_batches = self.params['validation_steps']
+      elif self.validation_data:
+        self._validation_batches = math.ceil(
+            self.validation_data[0].shape[0] / self.batch_size)
+      else:
+        raise ValueError('If printing histograms, validation data must be '
+                         'provided.')
+      if self._validation_batches == 0:
+        raise ValueError(
+            'If printing histograms, validation data must have length > 0.')
+
+  def on_epoch_begin(self, epoch, logs=None):
+    """Add histogram op to Model test_function callbacks, reset batch count."""
+
+    # check if histogram summary should be run for this epoch
+    if self.histogram_freq and epoch % self.histogram_freq == 0:
+      self._epoch = epoch
+      self._current_val_batch = 0
+      # add the histogram summary op if it should run this epoch
+      if self.merged not in self.model.test_function.fetches:
+        self.model.test_function.fetches.append(self.merged)
+        self.model.test_function.fetch_callbacks[
+            self.merged] = self._fetch_callback
 
   def on_epoch_end(self, epoch, logs=None):
+    """Checks if summary ops should run next epoch, logs scalar summaries."""
+
     logs = logs or {}
 
-    if not self.validation_data and self.histogram_freq:
-      raise ValueError('If printing histograms, validation_data must be '
-                       'provided, and cannot be a generator.')
-    if self.validation_data and self.histogram_freq:
-      if epoch % self.histogram_freq == 0:
-
-        val_data = self.validation_data
-        tensors = (
-            self.model.inputs + self.model.targets + self.model.sample_weights)
-
-        if self.model.uses_learning_phase:
-          tensors += [K.learning_phase()]
-
-        assert len(val_data) == len(tensors)
-        val_size = val_data[0].shape[0]
-        i = 0
-        while i < val_size:
-          step = min(self.batch_size, val_size - i)
-          batch_val = []
-          batch_val.append(val_data[0][i:i + step]
-                           if val_data[0] is not None else None)
-          batch_val.append(val_data[1][i:i + step]
-                           if val_data[1] is not None else None)
-          batch_val.append(val_data[2][i:i + step]
-                           if val_data[2] is not None else None)
-          if self.model.uses_learning_phase:
-            # do not slice the learning phase
-            batch_val = [x[i:i + step] if x is not None else None
-                         for x in val_data[:-1]]
-            batch_val.append(val_data[-1])
-          else:
-            batch_val = [x[i:i + step] if x is not None else None
-                         for x in val_data]
-          feed_dict = {}
-          for key, val in zip(tensors, batch_val):
-            if val is not None:
-              feed_dict[key] = val
-          result = self.sess.run([self.merged], feed_dict=feed_dict)
-          summary_str = result[0]
-          self.writer.add_summary(summary_str, epoch)
-          i += self.batch_size
+    # pop the histogram summary op after each epoch
+    if self.histogram_freq:
+      if self.merged in self.model.test_function.fetches:
+        self.model.test_function.fetches.remove(self.merged)
+      if self.merged in self.model.test_function.fetch_callbacks:
+        self.model.test_function.fetch_callbacks.pop(self.merged)
 
     for name, value in logs.items():
       if name in ['batch', 'size']:
diff --git a/tensorflow/python/keras/callbacks_test.py b/tensorflow/python/keras/callbacks_test.py
index 5062a26580ddb10011fd04f9a6e75ee6d2adbc68..244d48591ca462ad52dc73e9b4d85c96baabd5a0 100644
--- a/tensorflow/python/keras/callbacks_test.py
+++ b/tensorflow/python/keras/callbacks_test.py
@@ -27,11 +27,18 @@ import unittest
 
 import numpy as np
 
+from tensorflow.core.framework import summary_pb2
 from tensorflow.python import keras
+from tensorflow.python.eager import context
+from tensorflow.python.framework import test_util
 from tensorflow.python.keras import testing_utils
+from tensorflow.python.ops.resource_variable_ops import ResourceVariable as Variable
 from tensorflow.python.platform import test
 from tensorflow.python.platform import tf_logging as logging
 from tensorflow.python.summary.writer import writer_cache
+from tensorflow.python.training.adam import AdamOptimizer
+from tensorflow.python.training.gradient_descent import GradientDescentOptimizer
+
 
 try:
   import h5py  # pylint:disable=g-import-not-at-top
@@ -273,16 +280,43 @@ class KerasCallbacksTest(test.TestCase):
               1, activation='sigmoid'),))
       model.compile(
           optimizer='sgd', loss='binary_crossentropy', metrics=['accuracy'])
-      stopper = keras.callbacks.EarlyStopping(monitor='acc', patience=patience)
       weights = model.get_weights()
 
+      stopper = keras.callbacks.EarlyStopping(monitor='acc', patience=patience)
       hist = model.fit(data, labels, callbacks=[stopper], verbose=0, epochs=20)
       assert len(hist.epoch) >= patience
 
       # This should allow training to go for at least `patience` epochs
       model.set_weights(weights)
       hist = model.fit(data, labels, callbacks=[stopper], verbose=0, epochs=20)
-    assert len(hist.epoch) >= patience
+      assert len(hist.epoch) >= patience
+
+  def test_EarlyStopping_with_baseline(self):
+    with self.test_session():
+      np.random.seed(1337)
+      baseline = 0.5
+      (data, labels), _ = testing_utils.get_test_data(
+          train_samples=100,
+          test_samples=50,
+          input_shape=(1,),
+          num_classes=NUM_CLASSES)
+      model = keras.models.Sequential((keras.layers.Dense(
+          1, input_dim=1, activation='relu'), keras.layers.Dense(
+              1, activation='sigmoid'),))
+      model.compile(
+          optimizer='sgd', loss='binary_crossentropy', metrics=['accuracy'])
+
+      stopper = keras.callbacks.EarlyStopping(monitor='acc',
+                                              baseline=baseline)
+      hist = model.fit(data, labels, callbacks=[stopper], verbose=0, epochs=20)
+      assert len(hist.epoch) == 1
+
+      patience = 3
+      stopper = keras.callbacks.EarlyStopping(monitor='acc',
+                                              patience=patience,
+                                              baseline=baseline)
+      hist = model.fit(data, labels, callbacks=[stopper], verbose=0, epochs=20)
+      assert len(hist.epoch) >= patience
 
   def test_RemoteMonitor(self):
     if requests is None:
@@ -342,6 +376,76 @@ class KerasCallbacksTest(test.TestCase):
           float(keras.backend.get_value(
               model.optimizer.lr)) - 0.01 / 4) < keras.backend.epsilon()
 
+  @test_util.run_in_graph_and_eager_modes
+  def test_TF_LearningRateScheduler_Adam(self):
+    with self.test_session():
+      with context.eager_mode():
+        np.random.seed(1337)
+        (x_train, y_train), (x_test, y_test) = testing_utils.get_test_data(
+            train_samples=TRAIN_SAMPLES,
+            test_samples=TEST_SAMPLES,
+            input_shape=(INPUT_DIM,),
+            num_classes=NUM_CLASSES)
+        y_test = keras.utils.to_categorical(y_test)
+        y_train = keras.utils.to_categorical(y_train)
+        model = keras.models.Sequential()
+        model.add(
+            keras.layers.Dense(
+                NUM_HIDDEN, input_dim=INPUT_DIM, activation='relu'))
+        model.add(keras.layers.Dense(NUM_CLASSES, activation='softmax'))
+        model.compile(
+            loss='categorical_crossentropy',
+            optimizer=AdamOptimizer(),
+            metrics=['accuracy'])
+        cbks = [keras.callbacks.LearningRateScheduler(lambda x: 1. / (1. + x))]
+        model.fit(
+            x_train,
+            y_train,
+            batch_size=BATCH_SIZE,
+            validation_data=(x_test, y_test),
+            callbacks=cbks,
+            epochs=5,
+            verbose=0)
+        opt_lr = model.optimizer.optimizer._lr
+        self.assertLess(
+            float(keras.backend.get_value(
+                Variable(opt_lr))) - 0.2, keras.backend.epsilon())
+
+  @test_util.run_in_graph_and_eager_modes
+  def test_TF_LearningRateScheduler_GradientDescent(self):
+    with self.test_session():
+      with context.eager_mode():
+        np.random.seed(1337)
+        (x_train, y_train), (x_test, y_test) = testing_utils.get_test_data(
+            train_samples=TRAIN_SAMPLES,
+            test_samples=TEST_SAMPLES,
+            input_shape=(INPUT_DIM,),
+            num_classes=NUM_CLASSES)
+        y_test = keras.utils.to_categorical(y_test)
+        y_train = keras.utils.to_categorical(y_train)
+        model = keras.models.Sequential()
+        model.add(
+            keras.layers.Dense(
+                NUM_HIDDEN, input_dim=INPUT_DIM, activation='relu'))
+        model.add(keras.layers.Dense(NUM_CLASSES, activation='softmax'))
+        model.compile(
+            loss='categorical_crossentropy',
+            optimizer=GradientDescentOptimizer(1e-3),
+            metrics=['accuracy'])
+        cbks = [keras.callbacks.LearningRateScheduler(lambda x: 1. / (1. + x))]
+        model.fit(
+            x_train,
+            y_train,
+            batch_size=BATCH_SIZE,
+            validation_data=(x_test, y_test),
+            callbacks=cbks,
+            epochs=5,
+            verbose=0)
+        opt_lr = model.optimizer.optimizer._learning_rate
+        self.assertLess(
+            float(keras.backend.get_value(
+                Variable(opt_lr))) - 0.2, keras.backend.epsilon())
+
   def test_ReduceLROnPlateau(self):
     with self.test_session():
       np.random.seed(1337)
@@ -785,21 +889,6 @@ class KerasCallbacksTest(test.TestCase):
       for cb in cbs:
         cb.on_train_end()
 
-      # fit generator with validation data generator should raise ValueError if
-      # histogram_freq > 0
-      cbs = callbacks_factory(histogram_freq=1)
-      with self.assertRaises(ValueError):
-        model.fit_generator(
-            data_generator(True),
-            len(x_train),
-            epochs=2,
-            validation_data=data_generator(False),
-            validation_steps=1,
-            callbacks=cbs)
-
-      for cb in cbs:
-        cb.on_train_end()
-
       # Make sure file writer cache is clear to avoid failures during cleanup.
       writer_cache.FileWriterCache.clear()
 
@@ -874,6 +963,130 @@ class KerasCallbacksTest(test.TestCase):
                           callbacks=callbacks_factory(histogram_freq=1))
       assert os.path.isdir(filepath)
 
+  def test_Tensorboard_histogram_summaries_in_test_function(self):
+
+    class FileWriterStub(object):
+
+      def __init__(self, logdir, graph=None):
+        self.logdir = logdir
+        self.graph = graph
+        self.steps_seen = []
+
+      def add_summary(self, summary, global_step):
+        summary_obj = summary_pb2.Summary()
+
+        # ensure a valid Summary proto is being sent
+        if isinstance(summary, bytes):
+          summary_obj.ParseFromString(summary)
+        else:
+          assert isinstance(summary, summary_pb2.Summary)
+          summary_obj = summary
+
+        # keep track of steps seen for the merged_summary op,
+        # which contains the histogram summaries
+        if len(summary_obj.value) > 1:
+          self.steps_seen.append(global_step)
+
+      def flush(self):
+        pass
+
+      def close(self):
+        pass
+
+    np.random.seed(1337)
+    tmpdir = self.get_temp_dir()
+    self.addCleanup(shutil.rmtree, tmpdir)
+    (x_train, y_train), (x_test, y_test) = testing_utils.get_test_data(
+        train_samples=TRAIN_SAMPLES,
+        test_samples=TEST_SAMPLES,
+        input_shape=(INPUT_DIM,),
+        num_classes=NUM_CLASSES)
+    y_test = keras.utils.to_categorical(y_test)
+    y_train = keras.utils.to_categorical(y_train)
+
+    with self.test_session():
+      model = keras.models.Sequential()
+      model.add(
+          keras.layers.Dense(
+              NUM_HIDDEN, input_dim=INPUT_DIM, activation='relu'))
+      # non_trainable_weights: moving_variance, moving_mean
+      model.add(keras.layers.BatchNormalization())
+      model.add(keras.layers.Dense(NUM_CLASSES, activation='softmax'))
+      model.compile(
+          loss='categorical_crossentropy',
+          optimizer='sgd',
+          metrics=['accuracy'])
+      tsb = keras.callbacks.TensorBoard(
+          log_dir=tmpdir,
+          histogram_freq=1,
+          write_images=True,
+          write_grads=True,
+          batch_size=5)
+      tsb._writer_class = FileWriterStub
+      cbks = [tsb]
+
+      # fit with validation data
+      model.fit(
+          x_train,
+          y_train,
+          batch_size=BATCH_SIZE,
+          validation_data=(x_test, y_test),
+          callbacks=cbks,
+          epochs=3,
+          verbose=0)
+
+      self.assertAllEqual(tsb.writer.steps_seen, [0, 0.5, 1, 1.5, 2, 2.5])
+
+  def test_Tensorboard_histogram_summaries_with_generator(self):
+    np.random.seed(1337)
+    tmpdir = self.get_temp_dir()
+    self.addCleanup(shutil.rmtree, tmpdir)
+
+    def generator():
+      x = np.random.randn(10, 100).astype(np.float32)
+      y = np.random.randn(10, 10).astype(np.float32)
+      while True:
+        yield x, y
+
+    with self.test_session():
+      model = keras.models.Sequential()
+      model.add(keras.layers.Dense(10, input_dim=100, activation='relu'))
+      model.add(keras.layers.Dense(10, activation='softmax'))
+      model.compile(
+          loss='categorical_crossentropy',
+          optimizer='sgd',
+          metrics=['accuracy'])
+      tsb = keras.callbacks.TensorBoard(
+          log_dir=tmpdir,
+          histogram_freq=1,
+          write_images=True,
+          write_grads=True,
+          batch_size=5)
+      cbks = [tsb]
+
+      # fit with validation generator
+      model.fit_generator(
+          generator(),
+          steps_per_epoch=2,
+          epochs=2,
+          validation_data=generator(),
+          validation_steps=2,
+          callbacks=cbks,
+          verbose=0)
+
+      with self.assertRaises(ValueError):
+        # fit with validation generator but no
+        # validation_steps
+        model.fit_generator(
+            generator(),
+            steps_per_epoch=2,
+            epochs=2,
+            validation_data=generator(),
+            callbacks=cbks,
+            verbose=0)
+
+      self.assertTrue(os.path.exists(tmpdir))
+
   @unittest.skipIf(
       os.name == 'nt',
       'use_multiprocessing=True does not work on windows properly.')
diff --git a/tensorflow/python/keras/datasets/mnist.py b/tensorflow/python/keras/datasets/mnist.py
index 2a1c8d5f51818ff85617808b9a0779373b878f7b..a96b581960f3d5f60994fe92a1424e793d7e39c7 100644
--- a/tensorflow/python/keras/datasets/mnist.py
+++ b/tensorflow/python/keras/datasets/mnist.py
@@ -50,5 +50,5 @@ def load_data(path='mnist.npz'):
   with np.load(path) as f:
     x_train, y_train = f['x_train'], f['y_train']
     x_test, y_test = f['x_test'], f['y_test']
-    
+
     return (x_train, y_train), (x_test, y_test)
diff --git a/tensorflow/python/keras/engine/base_layer.py b/tensorflow/python/keras/engine/base_layer.py
index 4814275fd5ba53e7845f383b3447a6ef9f47f6c2..e02792208bdedb85601f0eacdf19836d331b1804 100644
--- a/tensorflow/python/keras/engine/base_layer.py
+++ b/tensorflow/python/keras/engine/base_layer.py
@@ -116,6 +116,7 @@ class Layer(checkpointable.CheckpointableBase):
       constraints on inputs that can be accepted by the layer.
   """
 
+  @checkpointable.no_automatic_dependency_tracking
   def __init__(self, trainable=True, name=None, dtype=None, **kwargs):
     # These properties should be set by the user via keyword arguments.
     # note that 'dtype', 'input_shape' and 'batch_input_shape'
@@ -217,7 +218,7 @@ class Layer(checkpointable.CheckpointableBase):
   @activity_regularizer.setter
   def activity_regularizer(self, regularizer):
     """Optional regularizer function for the output of this layer."""
-    self._activity_regularizer = regularizer
+    self._activity_regularizer = self._no_dependency(regularizer)
 
   @property
   def trainable_weights(self):
@@ -459,14 +460,18 @@ class Layer(checkpointable.CheckpointableBase):
     """Alias for `add_weight`."""
     return self.add_weight(*args, **kwargs)
 
-  def add_weight(self, name, shape,
+  def add_weight(self,
+                 name,
+                 shape,
                  dtype=None,
                  initializer=None,
                  regularizer=None,
-                 trainable=True,
+                 trainable=None,
                  constraint=None,
                  partitioner=None,
                  use_resource=None,
+                 synchronization=vs.VariableSynchronization.AUTO,
+                 aggregation=vs.VariableAggregation.NONE,
                  getter=None):
     """Adds a new variable to the layer, or gets an existing one; returns it.
 
@@ -481,10 +486,20 @@ class Layer(checkpointable.CheckpointableBase):
         or "non_trainable_variables" (e.g. BatchNorm mean, stddev).
         Note, if the current variable scope is marked as non-trainable
         then this parameter is ignored and any added variables are also
-        marked as non-trainable.
+        marked as non-trainable. `trainable` defaults to `True` unless
+        `synchronization` is set to `ON_READ`.
       constraint: constraint instance (callable).
       partitioner: Partitioner to be passed to the `Checkpointable` API.
       use_resource: Whether to use `ResourceVariable`.
+      synchronization: Indicates when a distributed a variable will be
+        aggregated. Accepted values are constants defined in the class
+        @{tf.VariableSynchronization}. By default the synchronization is set to
+        `AUTO` and the current `DistributionStrategy` chooses
+        when to synchronize. If `synchronization` is set to `ON_READ`,
+        `trainable` must not be set to `True`.
+      aggregation: Indicates how a distributed variable will be aggregated.
+        Accepted values are constants defined in the class
+        @{tf.VariableAggregation}.
       getter: Variable getter argument to be passed to the `Checkpointable` API.
 
     Returns:
@@ -495,7 +510,8 @@ class Layer(checkpointable.CheckpointableBase):
     Raises:
       RuntimeError: If called with partioned variable regularization and
         eager execution is enabled.
-      ValueError: When giving unsupported dtype and no initializer.
+      ValueError: When giving unsupported dtype and no initializer or when
+        trainable has been set to True with synchronization set as `ON_READ`.
     """
     if dtype is None:
       dtype = self.dtype or backend.floatx()
@@ -504,6 +520,19 @@ class Layer(checkpointable.CheckpointableBase):
     regularizer = regularizers.get(regularizer)
     constraint = constraints.get(constraint)
 
+    if synchronization == vs.VariableSynchronization.ON_READ:
+      if trainable:
+        raise ValueError(
+            'Synchronization value can be set to '
+            'VariableSynchronization.ON_READ only for non-trainable variables. '
+            'You have specified trainable=True and '
+            'synchronization=VariableSynchronization.ON_READ.')
+      else:
+        # Set trainable to be false when variable is to be synced on read.
+        trainable = False
+    elif trainable is None:
+      trainable = True
+
     # Initialize variable when no initializer provided
     if initializer is None:
       # If dtype is DT_FLOAT, provide a uniform unit scaling initializer
@@ -531,7 +560,9 @@ class Layer(checkpointable.CheckpointableBase):
         constraint=constraint,
         trainable=trainable and self.trainable,
         partitioner=partitioner,
-        use_resource=use_resource)
+        use_resource=use_resource,
+        synchronization=synchronization,
+        aggregation=aggregation)
 
     if regularizer is not None:
       # TODO(fchollet): in the future, this should be handled at the
@@ -654,11 +685,12 @@ class Layer(checkpointable.CheckpointableBase):
 
     # Handle Keras mask propagation from previous layer to current layer.
     previous_mask = None
-    if (not hasattr(self, '_compute_previous_mask') or
-        self._compute_previous_mask):
+    if build_graph and (not hasattr(self, '_compute_previous_mask') or
+                        self._compute_previous_mask):
       previous_mask = collect_previous_mask(inputs)
       if not hasattr(self, '_call_fn_args'):
-        self._call_fn_args = function_utils.fn_args(self.call)
+        self._call_fn_args = self._no_dependency(
+            function_utils.fn_args(self.call))
       if ('mask' in self._call_fn_args and 'mask' not in kwargs and
           not generic_utils.is_all_none(previous_mask)):
         # The previous layer generated a mask, and mask was not explicitly pass
@@ -691,9 +723,10 @@ class Layer(checkpointable.CheckpointableBase):
             self._dtype = input_list[0].dtype.base_dtype.name
           except AttributeError:
             pass
-        if all(hasattr(x, 'get_shape') for x in input_list):
-          input_shapes = nest.map_structure(lambda x: x.get_shape(), inputs)
+        if all(hasattr(x, 'shape') for x in input_list):
+          input_shapes = nest.map_structure(lambda x: x.shape, inputs)
         self.build(input_shapes)
+        self.built = True
 
       # Check input assumptions set after layer building, e.g. input shape.
       if build_graph or in_deferred_mode:
@@ -709,7 +742,7 @@ class Layer(checkpointable.CheckpointableBase):
         # Deferred mode behavior: use `compute_output_shape` to
         # infer the number of outputs of the layer and their shapes.
         if input_shapes is None:
-          input_shapes = nest.map_structure(lambda x: x.get_shape(), inputs)
+          input_shapes = nest.map_structure(lambda x: x.shape, inputs)
 
         output_shapes = self.compute_output_shape(input_shapes)
         output_shapes = nest.flatten(output_shapes)
@@ -729,8 +762,6 @@ class Layer(checkpointable.CheckpointableBase):
       if in_deferred_mode or build_graph and have_all_keras_metadata(inputs):
         inputs, outputs = self._set_connectivity_metadata_(
             inputs, outputs, args, kwargs)
-
-      self.built = True
       if context.executing_eagerly():
         return outputs
 
@@ -1293,7 +1324,7 @@ class Layer(checkpointable.CheckpointableBase):
                          ', but the layer isn\'t built. '
                          'You can build it manually via: `' + self.name +
                          '.build(batch_input_shape)`.')
-    weight_shapes = [w.get_shape().as_list() for w in self.weights]
+    weight_shapes = [w.shape.as_list() for w in self.weights]
     return int(sum([np.prod(w) for w in weight_shapes]))
 
   @property
@@ -1376,7 +1407,7 @@ class Layer(checkpointable.CheckpointableBase):
       if (spec.ndim is not None or
           spec.min_ndim is not None or
           spec.max_ndim is not None):
-        if x.get_shape().ndims is None:
+        if x.shape.ndims is None:
           raise ValueError('Input ' + str(input_index) + ' of layer ' +
                            self.name + ' is incompatible with the layer: '
                            'its rank is undefined, but the layer requires a '
@@ -1384,29 +1415,29 @@ class Layer(checkpointable.CheckpointableBase):
 
       # Check ndim.
       if spec.ndim is not None:
-        ndim = x.get_shape().ndims
+        ndim = x.shape.ndims
         if ndim != spec.ndim:
           raise ValueError('Input ' + str(input_index) + ' of layer ' +
                            self.name + ' is incompatible with the layer: '
                            'expected ndim=' + str(spec.ndim) + ', found ndim=' +
                            str(ndim) + '. Full shape received: ' +
-                           str(x.get_shape().as_list()))
+                           str(x.shape.as_list()))
       if spec.max_ndim is not None:
-        ndim = x.get_shape().ndims
+        ndim = x.shape.ndims
         if ndim is not None and ndim > spec.max_ndim:
           raise ValueError('Input ' + str(input_index) + ' of layer ' +
                            self.name + ' is incompatible with the layer: '
                            'expected max_ndim=' + str(spec.max_ndim) +
                            ', found ndim=' + str(ndim))
       if spec.min_ndim is not None:
-        ndim = x.get_shape().ndims
+        ndim = x.shape.ndims
         if ndim is not None and ndim < spec.min_ndim:
           raise ValueError('Input ' + str(input_index) + ' of layer ' +
                            self.name + ' is incompatible with the layer: '
                            ': expected min_ndim=' + str(spec.min_ndim) +
                            ', found ndim=' + str(ndim) +
                            '. Full shape received: ' +
-                           str(x.get_shape().as_list()))
+                           str(x.shape.as_list()))
       # Check dtype.
       if spec.dtype is not None:
         if x.dtype != spec.dtype:
@@ -1416,7 +1447,7 @@ class Layer(checkpointable.CheckpointableBase):
                            ', found dtype=' + str(x.dtype))
       # Check specific shape axes.
       if spec.axes:
-        shape = x.get_shape().as_list()
+        shape = x.shape.as_list()
         if shape is not None:
           for axis, value in spec.axes.items():
             if hasattr(value, 'value'):
@@ -1429,7 +1460,7 @@ class Layer(checkpointable.CheckpointableBase):
                   ' but received input with shape ' + str(shape))
       # Check shape.
       if spec.shape is not None:
-        shape = x.get_shape().as_list()
+        shape = x.shape.as_list()
         if shape is not None:
           for spec_dim, dim in zip(spec.shape, shape):
             if spec_dim is not None and dim is not None:
@@ -1704,12 +1735,12 @@ class DeferredTensor(object):
 
   def __str__(self):
     return "DeferredTensor('%s', shape=%s, dtype=%s)" % (self.name,
-                                                         self.get_shape(),
+                                                         self.shape,
                                                          self.dtype.name)
 
   def __repr__(self):
     return "" % (self.name,
-                                                        self.get_shape(),
+                                                        self.shape,
                                                         self.dtype.name)
 
 
@@ -1804,11 +1835,13 @@ def make_variable(name,
                   dtype=dtypes.float32,
                   initializer=None,
                   partition_info=None,
-                  trainable=True,
+                  trainable=None,
                   caching_device=None,
                   validate_shape=True,
                   constraint=None,
                   use_resource=None,
+                  synchronization=vs.VariableSynchronization.AUTO,
+                  aggregation=vs.VariableAggregation.NONE,
                   partitioner=None):  # pylint: disable=unused-argument
   """Temporary util to create a variable (relies on `variable_scope.variable`).
 
@@ -1834,11 +1867,21 @@ def make_variable(name,
       or "non_trainable_variables" (e.g. BatchNorm mean, stddev).
       Note, if the current variable scope is marked as non-trainable
       then this parameter is ignored and any added variables are also
-      marked as non-trainable.
+      marked as non-trainable. `trainable` defaults to `True` unless
+      `synchronization` is set to `ON_READ`.
     caching_device: Passed to `vs.variable`.
     validate_shape: Passed to `vs.variable`.
     constraint: Constraint instance (callable).
     use_resource: Whether to use a `ResourceVariable`.
+    synchronization: Indicates when a distributed a variable will be
+      aggregated. Accepted values are constants defined in the class
+      @{tf.VariableSynchronization}. By default the synchronization is set to
+      `AUTO` and the current `DistributionStrategy` chooses
+      when to synchronize. If `synchronization` is set to `ON_READ`,
+      `trainable` must not be set to `True`.
+    aggregation: Indicates how a distributed variable will be aggregated.
+      Accepted values are constants defined in the class
+      @{tf.VariableAggregation}.
     partitioner: Not handled at this time.
 
   Returns:
@@ -1870,5 +1913,7 @@ def make_variable(name,
       dtype=variable_dtype,
       validate_shape=validate_shape,
       constraint=constraint,
-      use_resource=use_resource)
+      use_resource=use_resource,
+      synchronization=synchronization,
+      aggregation=aggregation)
   return v
diff --git a/tensorflow/python/keras/engine/network.py b/tensorflow/python/keras/engine/network.py
index 3edb8033ff7fcf47c0f632a05d5dd3c2a4864834..a4d96de74fc90e31d52f9a67e845a84f9ceb5034 100644
--- a/tensorflow/python/keras/engine/network.py
+++ b/tensorflow/python/keras/engine/network.py
@@ -44,6 +44,7 @@ from tensorflow.python.ops import variables
 from tensorflow.python.platform import tf_logging as logging
 from tensorflow.python.training.checkpointable import base as checkpointable
 from tensorflow.python.training.checkpointable import data_structures
+from tensorflow.python.training.checkpointable import layer_utils as checkpointable_layer_utils
 from tensorflow.python.training.checkpointable import util as checkpointable_utils
 from tensorflow.python.util import nest
 from tensorflow.python.util import tf_inspect
@@ -80,6 +81,20 @@ class Network(base_layer.Layer):
       # Subclassed network
       self._init_subclassed_network(**kwargs)
 
+  # Several Network methods have "no_automatic_dependency_tracking"
+  # annotations. Since Network does automatic dependency tracking on attribute
+  # assignment, including for common data structures such as lists, by default
+  # we'd have quite a few empty dependencies which users don't care about (or
+  # would need some way to ignore dependencies automatically, which is confusing
+  # when applied to user code). Some attributes, such as _layers, would cause
+  # structural issues (_layers being the place where Layers assigned to tracked
+  # attributes are stored).
+  #
+  # Aside from these aesthetic and structural issues, useless dependencies on
+  # empty lists shouldn't cause issues; adding or removing them will not break
+  # checkpoints, but may cause "all Python objects matched" assertions to fail
+  # (in which case less strict assertions may be substituted if necessary).
+  @checkpointable.no_automatic_dependency_tracking
   def _base_init(self, name=None):
     # The following are implemented as property functions:
     # self.trainable_weights
@@ -134,6 +149,7 @@ class Network(base_layer.Layer):
     # restore operations when graph building.
     self._in_progress_restore_finalizer = None
 
+  @checkpointable.no_automatic_dependency_tracking
   def _init_graph_network(self, inputs, outputs, name=None):
     self._call_convention = base_layer.CallConvention.EXPLICIT_INPUTS_ARGUMENT
     # Normalize and set self.inputs, self.outputs.
@@ -292,6 +308,7 @@ class Network(base_layer.Layer):
     for layer in self._output_layers:
       self.output_names.append(layer.name)
 
+  @checkpointable.no_automatic_dependency_tracking
   def _init_subclassed_network(self, name=None):
     self._base_init(name=name)
     self._is_graph_network = False
@@ -361,10 +378,31 @@ class Network(base_layer.Layer):
       self._track_checkpointable(
           layer, name='layer-%d' % layer_index, overwrite=True)
 
+  def _no_dependency(self, value):
+    """Override to allow `Layer` to disable dependency tracking.
+
+    `CheckpointableBase` defines this method, whose semantics are "if a subclass
+    does dependency tracking, this method exempts `value`." Layer uses
+    `_no_dependency` to exempt some of its attribute assignments (conditional on
+    attribute assignment causing tracking in the subclass).
+
+    Args:
+      value: An object which will be assigned to an object attribute, whose
+        value should not be tracked.
+
+    Returns:
+      A wrapped object which, when assigned to an attribute, will not be
+      tracked (`value` will be stored in the attribute).
+    """
+    return data_structures.NoDependency(value)
+
   def __setattr__(self, name, value):
-    no_dependency = isinstance(value, checkpointable.NoDependency)
-    if no_dependency:
-      value = value.value
+    if not getattr(self, '_setattr_tracking', True):
+      super(Network, self).__setattr__(name, value)
+      return
+    no_dependency = isinstance(value, data_structures.NoDependency)
+    value = data_structures.sticky_attribute_assignment(
+        checkpointable=self, value=value, name=name)
     if isinstance(value, (
         base_layer.Layer,
         Network,
@@ -376,7 +414,9 @@ class Network(base_layer.Layer):
                            'forgot to call `super(YourClass, self).__init__()`.'
                            ' Always start with this line.')
       if not is_graph_network:
-        if value not in self._layers:
+        # We need to check object identity to avoid de-duplicating empty
+        # container types which compare equal.
+        if not any((layer is value for layer in self._layers)):
           self._layers.append(value)
           if hasattr(value, '_use_resource_variables'):
             # In subclassed models, legacy layers (tf.layers) must always use
@@ -384,12 +424,6 @@ class Network(base_layer.Layer):
             value._use_resource_variables = True
     if (not no_dependency
         and isinstance(value, checkpointable.CheckpointableBase)):
-      # Layer (and therefore Network/Model) inherit from CheckpointableBase
-      # rather than Checkpointable, which means there is no Checkpointable
-      # __setattr__ override (it would be a performance issue for functional
-      # layers). Therefore Model tracks Checkpointable objects itself.
-      self._track_checkpointable(
-          checkpointable=value, name=name, overwrite=True)
       if (  # For subclassed models only, users may add extra weights/variables
             # simply by assigning them to attributes.
           not self._is_graph_network
@@ -492,7 +526,8 @@ class Network(base_layer.Layer):
 
   @property
   def layers(self):
-    return self._layers
+    return checkpointable_layer_utils.filter_empty_layer_containers(
+        self._layers)
 
   def get_layer(self, name=None, index=None):
     """Retrieves a layer based on either its name (unique) or index.
@@ -665,14 +700,14 @@ class Network(base_layer.Layer):
 
   @property
   def trainable_weights(self):
-    return layer_utils.gather_trainable_weights(
+    return checkpointable_layer_utils.gather_trainable_weights(
         trainable=self.trainable,
         sub_layers=self.layers,
         extra_variables=self._extra_variables)
 
   @property
   def non_trainable_weights(self):
-    return layer_utils.gather_non_trainable_weights(
+    return checkpointable_layer_utils.gather_non_trainable_weights(
         trainable=self.trainable,
         sub_layers=self.layers,
         extra_variables=self._extra_variables)
diff --git a/tensorflow/python/keras/engine/saving.py b/tensorflow/python/keras/engine/saving.py
index b9a2e1f25f637dc8017f751bbdd400c1e5c9dd44..d5ccd44604b6b84ea0ceb4fa1c270b2c7dddc147 100644
--- a/tensorflow/python/keras/engine/saving.py
+++ b/tensorflow/python/keras/engine/saving.py
@@ -351,7 +351,10 @@ def preprocess_weights_for_loading(layer,
                                    weights,
                                    original_keras_version=None,
                                    original_backend=None):
-  """Converts layers weights from Keras 1 format to Keras 2.
+  """Preprocess layer weights between different Keras formats.
+
+  Converts layers weights from Keras 1 format to Keras 2 and also weights of
+  CuDNN layers in Keras 2.
 
   Arguments:
       layer: Layer instance.
@@ -363,7 +366,18 @@ def preprocess_weights_for_loading(layer,
   Returns:
       A list of weights values (Numpy arrays).
   """
-  if layer.__class__.__name__ == 'Bidirectional':
+  def convert_nested_bidirectional(weights):
+    """Converts layers nested in `Bidirectional` wrapper.
+
+    This function uses `preprocess_weights_for_loading()` for converting
+    layers.
+
+    Arguments:
+        weights: List of weights values (Numpy arrays).
+
+    Returns:
+        A list of weights values (Numpy arrays).
+    """
     num_weights_per_layer = len(weights) // 2
     forward_weights = preprocess_weights_for_loading(
         layer.forward_layer, weights[:num_weights_per_layer],
@@ -371,7 +385,69 @@ def preprocess_weights_for_loading(layer,
     backward_weights = preprocess_weights_for_loading(
         layer.backward_layer, weights[num_weights_per_layer:],
         original_keras_version, original_backend)
-    weights = forward_weights + backward_weights
+    return forward_weights + backward_weights
+
+  def convert_nested_time_distributed(weights):
+    """Converts layers nested in `TimeDistributed` wrapper.
+
+    This function uses `preprocess_weights_for_loading()` for converting nested
+    layers.
+
+    Arguments:
+        weights: List of weights values (Numpy arrays).
+
+    Returns:
+        A list of weights values (Numpy arrays).
+    """
+    return preprocess_weights_for_loading(
+        layer.layer, weights, original_keras_version, original_backend)
+
+  def convert_nested_model(weights):
+    """Converts layers nested in `Model` or `Sequential`.
+
+    This function uses `preprocess_weights_for_loading()` for converting nested
+    layers.
+
+    Arguments:
+        weights: List of weights values (Numpy arrays).
+
+    Returns:
+        A list of weights values (Numpy arrays).
+    """
+    new_weights = []
+    # trainable weights
+    for sublayer in layer.layers:
+      num_weights = len(sublayer.trainable_weights)
+      if num_weights > 0:
+        new_weights.extend(preprocess_weights_for_loading(
+            layer=sublayer,
+            weights=weights[:num_weights],
+            original_keras_version=original_keras_version,
+            original_backend=original_backend))
+        weights = weights[num_weights:]
+
+    # non-trainable weights
+    for sublayer in layer.layers:
+      num_weights = len([l for l in sublayer.weights
+                         if l not in sublayer.trainable_weights])
+      if num_weights > 0:
+        new_weights.extend(preprocess_weights_for_loading(
+            layer=sublayer,
+            weights=weights[:num_weights],
+            original_keras_version=original_keras_version,
+            original_backend=original_backend))
+        weights = weights[num_weights:]
+    return new_weights
+
+  # Convert layers nested in Bidirectional/Model/Sequential.
+  # Both transformation should be ran for both Keras 1->2 conversion
+  # and for conversion of CuDNN layers.
+  if layer.__class__.__name__ == 'Bidirectional':
+    weights = convert_nested_bidirectional(weights)
+  if layer.__class__.__name__ == 'TimeDistributed':
+    weights = convert_nested_time_distributed(weights)
+  elif layer.__class__.__name__ in ['Model', 'Sequential']:
+    weights = convert_nested_model(weights)
 
   if original_keras_version == '1':
     if layer.__class__.__name__ == 'TimeDistributed':
@@ -446,35 +522,6 @@ def preprocess_weights_for_loading(layer,
           recurrent_kernel = np.transpose(recurrent_kernel, (2, 3, 1, 0))
         weights = [kernel, recurrent_kernel, bias]
 
-    if layer.__class__.__name__ in ['Model', 'Sequential']:
-      new_weights = []
-      # trainable weights
-      for sublayer in layer.layers:
-        num_weights = len(sublayer.trainable_weights)
-        if num_weights > 0:
-          new_weights.extend(
-              preprocess_weights_for_loading(
-                  layer=sublayer,
-                  weights=weights[:num_weights],
-                  original_keras_version=original_keras_version,
-                  original_backend=original_backend))
-          weights = weights[num_weights:]
-
-      # non-trainable weights
-      for sublayer in layer.layers:
-        num_weights = len([
-            l for l in sublayer.weights if l not in sublayer.trainable_weights
-        ])
-        if num_weights > 0:
-          new_weights.extend(
-              preprocess_weights_for_loading(
-                  layer=sublayer,
-                  weights=weights[:num_weights],
-                  original_keras_version=original_keras_version,
-                  original_backend=original_backend))
-          weights = weights[num_weights:]
-      weights = new_weights
-
   conv_layers = ['Conv1D', 'Conv2D', 'Conv3D', 'Conv2DTranspose', 'ConvLSTM2D']
   if layer.__class__.__name__ in conv_layers:
     if original_backend == 'theano':
@@ -486,6 +533,7 @@ def preprocess_weights_for_loading(layer,
       if layer.__class__.__name__ == 'ConvLSTM2D':
         weights[1] = np.transpose(weights[1], (3, 2, 0, 1))
 
+  # convert CuDNN layers
   return _convert_rnn_weights(layer, weights)
 
 
@@ -624,7 +672,7 @@ def _convert_rnn_weights(layer, weights):
       kernels = transform_kernels(weights[0], transpose_input(from_cudnn),
                                   n_gates)
       recurrent_kernels = transform_kernels(weights[1], lambda k: k.T, n_gates)
-      biases = weights[2].reshape((2, -1) if from_cudnn else -1)
+      biases = np.array(weights[2]).reshape((2, -1) if from_cudnn else -1)
       return [kernels, recurrent_kernels, biases]
 
     if bias_shape == (2 * units * n_gates,):
@@ -806,7 +854,16 @@ def load_weights_from_hdf5_group_by_name(f, layers):
                          str(len(weight_values)) + ' element(s).')
       # Set values.
       for i in range(len(weight_values)):
-        weight_value_tuples.append((symbolic_weights[i], weight_values[i]))
+        if K.int_shape(symbolic_weights[i]) != weight_values[i].shape:
+          raise ValueError('Layer #' + str(k) +' (named "' + layer.name +
+                           '"), weight ' + str(symbolic_weights[i]) +
+                           ' has shape {}'.format(K.int_shape(
+                               symbolic_weights[i])) +
+                           ', but the saved weight has shape ' +
+                           str(weight_values[i].shape) + '.')
+
+        else:
+          weight_value_tuples.append((symbolic_weights[i], weight_values[i]))
   K.batch_set_value(weight_value_tuples)
 
 
diff --git a/tensorflow/python/keras/engine/saving_test.py b/tensorflow/python/keras/engine/saving_test.py
index 1a0aa60609216e3d39ec0e7af680f39011e7d6ce..030328f2a66f0ec406ac271aecfbf2dbebf22f5f 100644
--- a/tensorflow/python/keras/engine/saving_test.py
+++ b/tensorflow/python/keras/engine/saving_test.py
@@ -21,7 +21,6 @@ from __future__ import print_function
 import os
 import shutil
 import tempfile
-
 from absl.testing import parameterized
 import numpy as np
 
@@ -31,6 +30,7 @@ from tensorflow.python.framework import constant_op
 from tensorflow.python.framework import dtypes
 from tensorflow.python.framework import ops
 from tensorflow.python.framework import test_util
+from tensorflow.python.keras.engine import saving
 from tensorflow.python.keras.engine import training
 from tensorflow.python.ops import array_ops
 from tensorflow.python.ops import random_ops
@@ -248,6 +248,82 @@ class TestWeightSavingAndLoading(test.TestCase, parameterized.TestCase):
 
       self.assertAllClose(y, ref_y)
 
+  def test_sequential_weight_loading_group_name_with_incorrect_length(self):
+    if h5py is None:
+      return
+
+    temp_dir = self.get_temp_dir()
+    self.addCleanup(shutil.rmtree, temp_dir)
+    h5_path = os.path.join(temp_dir, 'test.h5')
+
+    num_hidden = 5
+    input_dim = 3
+    num_classes = 2
+    with self.test_session():
+      ref_model = keras.models.Sequential()
+      ref_model.add(keras.layers.Dense(num_hidden, input_dim=input_dim,
+                                       name='d1'))
+      ref_model.add(keras.layers.Dense(num_classes, name='d2'))
+      ref_model.compile(loss=keras.losses.MSE,
+                        optimizer=keras.optimizers.RMSprop(lr=0.0001),
+                        metrics=[keras.metrics.categorical_accuracy])
+
+      f_ref_model = h5py.File(h5_path, 'w')
+      saving.save_weights_to_hdf5_group(f_ref_model, ref_model.layers)
+
+      f_model = h5py.File(h5_path, 'r')
+      model = keras.models.Sequential()
+      model.add(keras.layers.Dense(num_hidden, use_bias=False,
+                                   input_dim=input_dim, name='d1'))
+      model.add(keras.layers.Dense(num_classes, name='d2'))
+      model.compile(loss=keras.losses.MSE,
+                    optimizer=keras.optimizers.RMSprop(lr=0.0001),
+                    metrics=[keras.metrics.categorical_accuracy])
+    with self.assertRaisesRegexp(ValueError,
+                                 r'Layer #0 \(named \"d1\"\) expects 1 '
+                                 r'weight\(s\), but the saved weights have 2 '
+                                 r'element\(s\)\.'):
+      saving.load_weights_from_hdf5_group_by_name(f_model, model.layers)
+
+  def test_sequential_weight_loading_group_name_with_incorrect_shape(self):
+    if h5py is None:
+      return
+
+    temp_dir = self.get_temp_dir()
+    self.addCleanup(shutil.rmtree, temp_dir)
+    h5_path = os.path.join(temp_dir, 'test.h5')
+
+    num_hidden = 5
+    input_dim = 3
+    num_classes = 2
+    with self.test_session():
+      ref_model = keras.models.Sequential()
+      ref_model.add(keras.layers.Dense(num_hidden, input_dim=input_dim,
+                                       name='d1'))
+      ref_model.add(keras.layers.Dense(num_classes, name='d2'))
+      ref_model.compile(loss=keras.losses.MSE,
+                        optimizer=keras.optimizers.RMSprop(lr=0.0001),
+                        metrics=[keras.metrics.categorical_accuracy])
+
+      f_ref_model = h5py.File(h5_path, 'w')
+      saving.save_weights_to_hdf5_group(f_ref_model, ref_model.layers)
+
+      f_model = h5py.File(h5_path, 'r')
+      model = keras.models.Sequential()
+      model.add(keras.layers.Dense(num_hidden + 5, input_dim=input_dim,
+                                   name='d1'))
+      model.add(keras.layers.Dense(num_classes, name='d2'))
+      model.compile(loss=keras.losses.MSE,
+                    optimizer=keras.optimizers.RMSprop(lr=0.0001),
+                    metrics=[keras.metrics.categorical_accuracy])
+      with self.assertRaisesRegexp(ValueError,
+                                   r'Layer #0 \(named "d1"\), weight '
+                                   r' has '
+                                   r'shape \(3, 10\), but the saved weight has '
+                                   r'shape \(3, 5\)\.'):
+        saving.load_weights_from_hdf5_group_by_name(f_model, model.layers)
+
 
 class TestWholeModelSaving(test.TestCase):
 
diff --git a/tensorflow/python/keras/engine/sequential.py b/tensorflow/python/keras/engine/sequential.py
index 89b40b5d3869e33254229626854cc768ea4c91f9..371504a503168e7443895bb22a57126b274da226 100644
--- a/tensorflow/python/keras/engine/sequential.py
+++ b/tensorflow/python/keras/engine/sequential.py
@@ -29,6 +29,7 @@ from tensorflow.python.keras.engine.input_layer import InputLayer
 from tensorflow.python.keras.engine.training import Model
 from tensorflow.python.keras.utils import layer_utils
 from tensorflow.python.platform import tf_logging as logging
+from tensorflow.python.training.checkpointable import base as checkpointable
 from tensorflow.python.util.tf_export import tf_export
 
 
@@ -108,6 +109,7 @@ class Sequential(Model):
       return self._layers[1:]
     return self._layers
 
+  @checkpointable.no_automatic_dependency_tracking
   def add(self, layer):
     """Adds a layer instance on top of the layer stack.
 
@@ -146,8 +148,6 @@ class Sequential(Model):
           first_layer = layer.layers[0]
           while isinstance(first_layer, (Model, Sequential)):
             first_layer = first_layer.layers[0]
-          batch_shape = first_layer._batch_input_shape
-          dtype = first_layer.dtype
 
         if hasattr(first_layer, '_batch_input_shape'):
           batch_shape = first_layer._batch_input_shape
@@ -193,6 +193,7 @@ class Sequential(Model):
     else:
       self._layers.append(layer)
 
+  @checkpointable.no_automatic_dependency_tracking
   def pop(self):
     """Removes the last layer in the model.
 
@@ -212,6 +213,7 @@ class Sequential(Model):
       self.outputs = [self.layers[-1].output]
       self.build()
 
+  @checkpointable.no_automatic_dependency_tracking
   def build(self, input_shape=None):
     if input_shape and not self.inputs:
       batch_shape = tuple(input_shape)
diff --git a/tensorflow/python/keras/engine/training.py b/tensorflow/python/keras/engine/training.py
index fce6cbdb7a09d5a4daa70d0483732735e70c28b7..bd03f4871f1abc506858f7c10818c3efb24c2716 100644
--- a/tensorflow/python/keras/engine/training.py
+++ b/tensorflow/python/keras/engine/training.py
@@ -42,6 +42,7 @@ from tensorflow.python.keras.utils.generic_utils import slice_arrays
 from tensorflow.python.ops import array_ops
 from tensorflow.python.platform import tf_logging as logging
 from tensorflow.python.training import optimizer as tf_optimizer_module
+from tensorflow.python.training.checkpointable import base as checkpointable
 from tensorflow.python.util.tf_export import tf_export
 
 
@@ -115,6 +116,7 @@ class Model(Network):
     # Create a cache for dataset - uninitialized iterators
     self._dataset_iterator_cache = weakref.WeakKeyDictionary()
 
+  @checkpointable.no_automatic_dependency_tracking
   def compile(self,
               optimizer,
               loss=None,
@@ -178,6 +180,11 @@ class Model(Network):
       raise ValueError('Only TF native optimizers are supported in Eager mode.')
 
     self.optimizer = optimizers.get(optimizer)
+    # We've disabled automatic dependency tracking for this method, but do want
+    # to add a checkpoint dependency on the optimizer if it's checkpointable.
+    if isinstance(self.optimizer, checkpointable.CheckpointableBase):
+      self._track_checkpointable(
+          self.optimizer, name='optimizer', overwrite=True)
     self.loss = loss
     self.metrics = metrics or []
     self.loss_weights = loss_weights
@@ -592,7 +599,7 @@ class Model(Network):
         # Unconditional updates
         updates += self.get_updates_for(None)
         # Conditional updates relevant to this model
-        updates += self.get_updates_for(self._feed_inputs)
+        updates += self.get_updates_for(self.inputs)
         # Stateful metrics updates
         updates += self.metrics_updates
         # Gets loss and metrics. Updates weights at each call.
@@ -941,6 +948,7 @@ class Model(Network):
                          str(x[0].shape[0]) + ' samples')
     return x, y, sample_weights
 
+  @checkpointable.no_automatic_dependency_tracking
   def _set_inputs(self, inputs, training=None):
     """Set model's input and output specs based on the input data received.
 
@@ -989,6 +997,7 @@ class Model(Network):
     else:
       self._symbolic_set_inputs(inputs, training=training)
 
+  @checkpointable.no_automatic_dependency_tracking
   def _eager_set_inputs(self, inputs):
     """Set model's input and output specs based on the input data received.
 
@@ -1041,6 +1050,7 @@ class Model(Network):
         'output_%d' % (i + 1) for i in range(len(dummy_output_values))]
     self.built = True
 
+  @checkpointable.no_automatic_dependency_tracking
   def _symbolic_set_inputs(self, inputs, outputs=None, training=None):
     """Set model's inputs and output specs based.
 
diff --git a/tensorflow/python/keras/engine/training_arrays.py b/tensorflow/python/keras/engine/training_arrays.py
index 281ad9bd50edf519f520fe4aa664ae05b72528d8..adefffab11093bbe60ee2342706d09d4ff006b5c 100644
--- a/tensorflow/python/keras/engine/training_arrays.py
+++ b/tensorflow/python/keras/engine/training_arrays.py
@@ -124,6 +124,10 @@ def fit_loop(model,
     callback_metrics = copy.copy(out_labels) + [
         'val_' + n for n in out_labels
     ]
+    # need to create the test_function before start of the first epoch
+    # because TensorBoard callback on_epoch_begin adds summary to the
+    # list of fetches of the test_function
+    model._make_test_function()
   else:
     callback_metrics = copy.copy(out_labels)
 
@@ -156,7 +160,7 @@ def fit_loop(model,
 
   callbacks.set_model(callback_model)
 
-  callbacks.set_params({
+  callback_params = {
       'batch_size': batch_size,
       'epochs': epochs,
       'steps': steps_per_epoch,
@@ -164,11 +168,17 @@ def fit_loop(model,
       'verbose': verbose,
       'do_validation': do_validation,
       'metrics': callback_metrics or [],
-  })
-  callbacks.on_train_begin()
-  callback_model.stop_training = False
+  }
+  if validation_steps:
+    callback_params.update({'validation_steps': validation_steps})
+  callbacks.set_params(callback_params)
+
   for cbk in callbacks:
     cbk.validation_data = val_ins
+  # validation_data must be set before on_train_begin() is called
+  # so that TensorboardCallback can validate its input
+  callbacks.on_train_begin()
+  callback_model.stop_training = False
 
   # To prevent a slowdown, we find beforehand the arrays that need conversion.
   feed = model._feed_inputs + model._feed_targets + model._feed_sample_weights
diff --git a/tensorflow/python/keras/engine/training_eager.py b/tensorflow/python/keras/engine/training_eager.py
index e8838cd3bca7b3afba80504f9e705943474423c5..c78684c9f4e6d32b1572def9a855361841fa9af2 100644
--- a/tensorflow/python/keras/engine/training_eager.py
+++ b/tensorflow/python/keras/engine/training_eager.py
@@ -989,7 +989,7 @@ def fit_loop(model,
 
     callbacks.set_model(callback_model)
 
-    callbacks.set_params({
+    callback_params = {
         'batch_size': batch_size,
         'epochs': epochs,
         'steps': steps_per_epoch,
@@ -997,9 +997,11 @@ def fit_loop(model,
         'verbose': verbose,
         'do_validation': do_validation,
         'metrics': callback_metrics or [],
-    })
-    callbacks.on_train_begin()
-    callback_model.stop_training = False
+    }
+    if validation_steps:
+      callback_params.update({'validation_steps': validation_steps})
+    callbacks.set_params(callback_params)
+
     for cbk in callbacks:
       if not val_inputs:
         cbk.validation_data = []
@@ -1009,6 +1011,10 @@ def fit_loop(model,
         cbk.validation_data = val_inputs + val_targets + val_sample_weights
       else:
         cbk.validation_data = val_inputs + val_targets
+    # validation_data must be set before on_train_begin() is called
+    # so that TensorboardCallback can validate its input
+    callbacks.on_train_begin()
+    callback_model.stop_training = False
 
     for epoch in range(initial_epoch, epochs):
       callbacks.on_epoch_begin(epoch)
diff --git a/tensorflow/python/keras/engine/training_generator.py b/tensorflow/python/keras/engine/training_generator.py
index d81b384f0e1810614bd98e3861b4324f0f8a4dca..432cf2bddd052b40dd80dc530c9c6ce23d57d57b 100644
--- a/tensorflow/python/keras/engine/training_generator.py
+++ b/tensorflow/python/keras/engine/training_generator.py
@@ -96,14 +96,25 @@ def fit_generator(model,
   else:
     callback_model = model
   callbacks.set_model(callback_model)
-  callbacks.set_params({
+
+  callback_params = {
       'epochs': epochs,
       'steps': steps_per_epoch,
       'verbose': verbose,
       'do_validation': do_validation,
       'metrics': callback_metrics,
-  })
-  callbacks.on_train_begin()
+  }
+  if do_validation:
+    # need to create the test_function before start of the first epoch
+    # because TensorBoard callback on_epoch_begin adds summary to the
+    # list of fetches of the test_function
+    model._make_test_function()
+    # determine the number of validation batches given a generator
+    if validation_steps:
+      callback_params.update({'validation_steps': validation_steps})
+    elif isinstance(validation_data, Sequence):
+      callback_params.update({'validation_steps': len(validation_data)})
+  callbacks.set_params(callback_params)
 
   enqueuer = None
   val_enqueuer = None
@@ -149,6 +160,9 @@ def fit_generator(model,
         output_generator = generator
 
     callback_model.stop_training = False
+    # validation_data must be set before on_train_begin() is called
+    # so that TensorboardCallback can validate its input
+    callbacks.on_train_begin()
     # Construct epoch logs.
     epoch_logs = {}
     while epoch < epochs:
diff --git a/tensorflow/python/keras/estimator/__init__.py b/tensorflow/python/keras/estimator/__init__.py
index cb86a69990b99881d0f068b4fc94b42241cf396d..b244beb5b58cf339a4687216b87418c88b953c17 100644
--- a/tensorflow/python/keras/estimator/__init__.py
+++ b/tensorflow/python/keras/estimator/__init__.py
@@ -25,7 +25,7 @@ from tensorflow.python.util.tf_export import tf_export
 # everything will work as normal.
 
 try:
-  import tensorflow.python.estimator.keras as keras_lib  # pylint: disable=g-import-not-at-top
+  from tensorflow.python.estimator import keras as keras_lib  # pylint: disable=g-import-not-at-top
   model_to_estimator = tf_export('keras.estimator.model_to_estimator')(
       keras_lib.model_to_estimator)
 except Exception:  # pylint: disable=broad-except
diff --git a/tensorflow/python/keras/initializers.py b/tensorflow/python/keras/initializers.py
index b9b2e9ad598fabe8cbfbbcbd57d4d71ddf630df7..28beb6760d11514fe41d5c6c4b65e5f1772b84f8 100644
--- a/tensorflow/python/keras/initializers.py
+++ b/tensorflow/python/keras/initializers.py
@@ -23,6 +23,9 @@ import six
 from tensorflow.python.keras.utils.generic_utils import deserialize_keras_object
 from tensorflow.python.keras.utils.generic_utils import serialize_keras_object
 from tensorflow.python.ops.init_ops import Constant
+from tensorflow.python.ops.init_ops import glorot_normal_initializer
+from tensorflow.python.ops.init_ops import glorot_uniform_initializer
+
 from tensorflow.python.ops.init_ops import Identity
 from tensorflow.python.ops.init_ops import Initializer  # pylint: disable=unused-import
 from tensorflow.python.ops.init_ops import Ones
@@ -80,52 +83,6 @@ def lecun_uniform(seed=None):
       scale=1., mode='fan_in', distribution='uniform', seed=seed)
 
 
-@tf_export('keras.initializers.glorot_normal')
-def glorot_normal(seed=None):
-  """Glorot normal initializer, also called Xavier normal initializer.
-
-  It draws samples from a truncated normal distribution centered on 0
-  with `stddev = sqrt(2 / (fan_in + fan_out))`
-  where `fan_in` is the number of input units in the weight tensor
-  and `fan_out` is the number of output units in the weight tensor.
-
-  Arguments:
-      seed: A Python integer. Used to seed the random generator.
-
-  Returns:
-      An initializer.
-
-  References:
-      Glorot & Bengio, AISTATS 2010
-      http://jmlr.org/proceedings/papers/v9/glorot10a/glorot10a.pdf
-  """
-  return VarianceScaling(
-      scale=1., mode='fan_avg', distribution='normal', seed=seed)
-
-
-@tf_export('keras.initializers.glorot_uniform')
-def glorot_uniform(seed=None):
-  """Glorot uniform initializer, also called Xavier uniform initializer.
-
-  It draws samples from a uniform distribution within [-limit, limit]
-  where `limit` is `sqrt(6 / (fan_in + fan_out))`
-  where `fan_in` is the number of input units in the weight tensor
-  and `fan_out` is the number of output units in the weight tensor.
-
-  Arguments:
-      seed: A Python integer. Used to seed the random generator.
-
-  Returns:
-      An initializer.
-
-  References:
-      Glorot & Bengio, AISTATS 2010
-      http://jmlr.org/proceedings/papers/v9/glorot10a/glorot10a.pdf
-  """
-  return VarianceScaling(
-      scale=1., mode='fan_avg', distribution='uniform', seed=seed)
-
-
 @tf_export('keras.initializers.he_normal')
 def he_normal(seed=None):
   """He normal initializer.
@@ -179,6 +136,8 @@ normal = random_normal = RandomNormal
 truncated_normal = TruncatedNormal
 identity = Identity
 orthogonal = Orthogonal
+glorot_normal = glorot_normal_initializer
+glorot_uniform = glorot_uniform_initializer
 
 # pylint: enable=invalid-name
 
diff --git a/tensorflow/python/keras/layers/core.py b/tensorflow/python/keras/layers/core.py
index 2bf6229ccba808360e73a333bdec3dac624d81ce..f28cade474e450174f95c9a8e06e26b04e95eb69 100644
--- a/tensorflow/python/keras/layers/core.py
+++ b/tensorflow/python/keras/layers/core.py
@@ -26,6 +26,7 @@ import warnings
 import numpy as np
 
 from tensorflow.python.eager import context
+from tensorflow.python.framework import common_shapes
 from tensorflow.python.framework import ops
 from tensorflow.python.framework import tensor_shape
 from tensorflow.python.keras import activations
@@ -929,13 +930,13 @@ class Dense(Layer):
 
   def call(self, inputs):
     inputs = ops.convert_to_tensor(inputs, dtype=self.dtype)
-    shape = inputs.get_shape().as_list()
-    if len(shape) > 2:
+    rank = common_shapes.rank(inputs)
+    if rank > 2:
       # Broadcasting is required for the inputs.
-      outputs = standard_ops.tensordot(inputs, self.kernel, [[len(shape) - 1],
-                                                             [0]])
+      outputs = standard_ops.tensordot(inputs, self.kernel, [[rank - 1], [0]])
       # Reshape the output back to the original ndim of the input.
       if not context.executing_eagerly():
+        shape = inputs.get_shape().as_list()
         output_shape = shape[:-1] + [self.units]
         outputs.set_shape(output_shape)
     else:
diff --git a/tensorflow/python/keras/layers/cudnn_recurrent_test.py b/tensorflow/python/keras/layers/cudnn_recurrent_test.py
index f1ee441f5f49443e4b149bbd698782c1a50fb26b..8fd970239f205031954c728474abdf10ea80e99e 100644
--- a/tensorflow/python/keras/layers/cudnn_recurrent_test.py
+++ b/tensorflow/python/keras/layers/cudnn_recurrent_test.py
@@ -18,6 +18,8 @@ from __future__ import absolute_import
 from __future__ import division
 from __future__ import print_function
 
+import os
+import tempfile
 from absl.testing import parameterized
 import numpy as np
 
@@ -217,27 +219,14 @@ class CuDNNTest(test.TestCase, parameterized.TestCase):
         out5 = model.predict(np.ones((num_samples, timesteps)))
         self.assertNotEqual(out4.max(), out5.max())
 
-  # TODO(psv): Add generic cross product helper function for parametrized tests.
   @parameterized.named_parameters(
-      ('cudnnlstm_to_lstm_unidirectional_impl_1', 'LSTM', False, False, 1),
-      ('cudnnlstm_to_lstm_bidirectional_impl_1', 'LSTM', False, True, 1),
-      ('lstm_to_cudnnlstm_unidirectional_impl_1', 'LSTM', True, False, 1),
-      ('lstm_to_cudnnlstm_bidirectional_impl_1', 'LSTM', True, True, 1),
-      ('cudnngru_to_gru_unidirectional_impl_1', 'GRU', False, False, 1),
-      ('cudnngru_to_gru_bidirectional_impl_1', 'GRU', False, True, 1),
-      ('gru_to_cudnngru_unidirectional_impl_1', 'GRU', True, False, 1),
-      ('gru_to_cudnngru_bidirectional_impl_1', 'GRU', True, True, 1),
-      ('cudnnlstm_to_lstm_unidirectional_impl_2', 'LSTM', False, False, 2),
-      ('cudnnlstm_to_lstm_bidirectional_impl_2', 'LSTM', False, True, 2),
-      ('lstm_to_cudnnlstm_unidirectional_impl_2', 'LSTM', True, False, 2),
-      ('lstm_to_cudnnlstm_bidirectional_impl_2', 'LSTM', True, True, 2),
-      ('cudnngru_to_gru_unidirectional_impl_2', 'GRU', False, False, 2),
-      ('cudnngru_to_gru_bidirectional_impl_2', 'GRU', False, True, 2),
-      ('gru_to_cudnngru_unidirectional_impl_2', 'GRU', True, False, 2),
-      ('gru_to_cudnngru_bidirectional_impl_2', 'GRU', True, True, 2),
-  )
+      *testing_utils.generate_combinations_with_testcase_name(
+          rnn_type=['LSTM', 'GRU'], to_cudnn=[True, False],
+          bidirectional=[True, False], implementation=[1, 2],
+          model_nest_level=[1, 2], model_type=['seq', 'func']))
   def test_load_weights_between_noncudnn_rnn(self, rnn_type, to_cudnn,
-                                             bidirectional, implementation):
+                                             bidirectional, implementation,
+                                             model_nest_level, model_type):
     if test.is_gpu_available(cuda_only=True):
       with self.test_session(use_gpu=True):
         input_size = 10
@@ -261,14 +250,6 @@ class CuDNNTest(test.TestCase, parameterized.TestCase):
           cudnn_rnn_layer_class = keras.layers.CuDNNGRU
           rnn_layer_kwargs['reset_after'] = True
 
-        def convert_weights(source_layer, target_layer):
-          weights = source_layer.get_weights()
-          weights = keras.engine.saving.preprocess_weights_for_loading(
-              target_layer, weights)
-          target_layer.set_weights(weights)
-
-        input_layer = keras.layers.InputLayer(input_shape)
-
         layer = rnn_layer_class(units, **rnn_layer_kwargs)
         if bidirectional:
           layer = keras.layers.Bidirectional(layer)
@@ -277,16 +258,94 @@ class CuDNNTest(test.TestCase, parameterized.TestCase):
         if bidirectional:
           cudnn_layer = keras.layers.Bidirectional(cudnn_layer)
 
-        model = keras.models.Sequential([input_layer, layer])
-        cudnn_model = keras.models.Sequential([input_layer, cudnn_layer])
+        model = self._make_nested_model(input_shape, layer, model_nest_level,
+                                        model_type)
+        cudnn_model = self._make_nested_model(input_shape, cudnn_layer,
+                                              model_nest_level, model_type)
+
+        if to_cudnn:
+          self._convert_model_weights(model, cudnn_model)
+        else:
+          self._convert_model_weights(cudnn_model, model)
+
+        self.assertAllClose(model.predict(inputs), cudnn_model.predict(inputs),
+                            atol=1e-4)
+
+  def _make_nested_model(self, input_shape, layer, level=1, model_type='func'):
+    # example: make_nested_seq_model((1,), Dense(10), level=2).summary()
+    def make_nested_seq_model(input_shape, layer, level=1):
+      model = layer
+      for i in range(1, level + 1):
+        layers = [keras.layers.InputLayer(input_shape),
+                  model] if (i == 1) else [model]
+        model = keras.models.Sequential(layers)
+      return model
+
+    # example: make_nested_func_model((1,), Dense(10), level=2).summary()
+    def make_nested_func_model(input_shape, layer, level=1):
+      model_input = keras.layers.Input(input_shape)
+      model = layer
+      for _ in range(level):
+        model = keras.models.Model(model_input, model(model_input))
+      return model
+
+    if model_type == 'func':
+      return make_nested_func_model(input_shape, layer, level)
+    elif model_type == 'seq':
+      return make_nested_seq_model(input_shape, layer, level)
+
+  def _convert_model_weights(self, source_model, target_model):
+    _, fname = tempfile.mkstemp('.h5')
+    source_model.save_weights(fname)
+    target_model.load_weights(fname)
+    os.remove(fname)
+
+  @parameterized.named_parameters(
+      *testing_utils.generate_combinations_with_testcase_name(
+          rnn_type=['LSTM', 'GRU'], to_cudnn=[True, False]))
+  def test_load_weights_between_noncudnn_rnn_time_distributed(self, rnn_type,
+                                                              to_cudnn):
+    # Similar test as test_load_weights_between_noncudnn_rnn() but has different
+    # rank of input due to usage of TimeDistributed. Issue: #10356.
+    if test.is_gpu_available(cuda_only=True):
+      with self.test_session(use_gpu=True):
+        input_size = 10
+        steps = 6
+        timesteps = 6
+        input_shape = (timesteps, steps, input_size)
+        units = 2
+        num_samples = 32
+        inputs = np.random.random((num_samples, timesteps, steps, input_size))
+
+        rnn_layer_kwargs = {
+            'recurrent_activation': 'sigmoid',
+            # ensure biases are non-zero and properly converted
+            'bias_initializer': 'random_uniform',
+        }
+        if rnn_type == 'LSTM':
+          rnn_layer_class = keras.layers.LSTM
+          cudnn_rnn_layer_class = keras.layers.CuDNNLSTM
+        else:
+          rnn_layer_class = keras.layers.GRU
+          cudnn_rnn_layer_class = keras.layers.CuDNNGRU
+          rnn_layer_kwargs['reset_after'] = True
+
+        layer = rnn_layer_class(units, **rnn_layer_kwargs)
+        layer = keras.layers.TimeDistributed(layer)
+
+        cudnn_layer = cudnn_rnn_layer_class(units)
+        cudnn_layer = keras.layers.TimeDistributed(cudnn_layer)
+
+        model = self._make_nested_model(input_shape, layer)
+        cudnn_model = self._make_nested_model(input_shape, cudnn_layer)
 
         if to_cudnn:
-          convert_weights(layer, cudnn_layer)
+          self._convert_model_weights(model, cudnn_model)
         else:
-          convert_weights(cudnn_layer, layer)
+          self._convert_model_weights(cudnn_model, model)
 
-        self.assertAllClose(
-            model.predict(inputs), cudnn_model.predict(inputs), atol=1e-4)
+        self.assertAllClose(model.predict(inputs), cudnn_model.predict(inputs),
+                            atol=1e-4)
 
   @test_util.run_in_graph_and_eager_modes
   def test_cudnnrnn_bidirectional(self):
diff --git a/tensorflow/python/keras/layers/embeddings.py b/tensorflow/python/keras/layers/embeddings.py
index 910fff720f6312041a25922cf5c63dfa8f83ec76..629a9ec9a10c8afd4d98174a9183a2e9b08269ea 100644
--- a/tensorflow/python/keras/layers/embeddings.py
+++ b/tensorflow/python/keras/layers/embeddings.py
@@ -112,6 +112,7 @@ class Embedding(Layer):
     self.activity_regularizer = regularizers.get(activity_regularizer)
     self.embeddings_constraint = constraints.get(embeddings_constraint)
     self.mask_zero = mask_zero
+    self.supports_masking = mask_zero
     self.input_length = input_length
 
   @tf_utils.shape_type_conversion
@@ -127,8 +128,8 @@ class Embedding(Layer):
   def compute_mask(self, inputs, mask=None):
     if not self.mask_zero:
       return None
-    else:
-      return math_ops.not_equal(inputs, 0)
+
+    return math_ops.not_equal(inputs, 0)
 
   @tf_utils.shape_type_conversion
   def compute_output_shape(self, input_shape):
diff --git a/tensorflow/python/keras/layers/normalization.py b/tensorflow/python/keras/layers/normalization.py
index d4c213eedd9eb3da0a3644540da29fa22a60f453..a7835bc0a2ad1865c2d98b5f539a6643f2272b81 100644
--- a/tensorflow/python/keras/layers/normalization.py
+++ b/tensorflow/python/keras/layers/normalization.py
@@ -34,6 +34,7 @@ from tensorflow.python.ops import init_ops
 from tensorflow.python.ops import math_ops
 from tensorflow.python.ops import nn
 from tensorflow.python.ops import state_ops
+from tensorflow.python.ops import variable_scope
 from tensorflow.python.platform import tf_logging as logging
 from tensorflow.python.training import distribute as distribute_lib
 from tensorflow.python.util.tf_export import tf_export
@@ -180,11 +181,6 @@ class BatchNormalization(Layer):
       self.renorm_clipping = renorm_clipping
       self.renorm_momentum = renorm_momentum
 
-  def _add_tower_local_variable(self, *args, **kwargs):
-    tower_context = distribute_lib.get_tower_context()
-    with tower_context.tower_local_var_scope('mean'):
-      return self.add_weight(*args, **kwargs)
-
   def build(self, input_shape):
     input_shape = tensor_shape.TensorShape(input_shape)
     if not input_shape.ndims:
@@ -312,19 +308,23 @@ class BatchNormalization(Layer):
         self._scope.set_partitioner(None)
       else:
         partitioner = None
-      self.moving_mean = self._add_tower_local_variable(
+      self.moving_mean = self.add_weight(
           name='moving_mean',
           shape=param_shape,
           dtype=param_dtype,
           initializer=self.moving_mean_initializer,
-          trainable=False)
+          synchronization=variable_scope.VariableSynchronization.ON_READ,
+          trainable=False,
+          aggregation=variable_scope.VariableAggregation.MEAN)
 
-      self.moving_variance = self._add_tower_local_variable(
+      self.moving_variance = self.add_weight(
           name='moving_variance',
           shape=param_shape,
           dtype=param_dtype,
           initializer=self.moving_variance_initializer,
-          trainable=False)
+          synchronization=variable_scope.VariableSynchronization.ON_READ,
+          trainable=False,
+          aggregation=variable_scope.VariableAggregation.MEAN)
 
       if self.renorm:
         # Create variables to maintain the moving mean and standard deviation.
@@ -335,12 +335,14 @@ class BatchNormalization(Layer):
         # stack to be cleared. The nested ones use a `lambda` to set the desired
         # device and ignore any devices that may be set by the custom getter.
         def _renorm_variable(name, shape):
-          var = self._add_tower_local_variable(
+          var = self.add_weight(
               name=name,
               shape=shape,
               dtype=param_dtype,
               initializer=init_ops.zeros_initializer(),
-              trainable=False)
+              synchronization=variable_scope.VariableSynchronization.ON_READ,
+              trainable=False,
+              aggregation=variable_scope.VariableAggregation.MEAN)
           return var
 
         with distribute_lib.get_distribution_strategy().colocate_vars_with(
@@ -368,7 +370,7 @@ class BatchNormalization(Layer):
         decay = ops.convert_to_tensor(1.0 - momentum, name='decay')
         if decay.dtype != variable.dtype.base_dtype:
           decay = math_ops.cast(decay, variable.dtype.base_dtype)
-        update_delta = (variable - value) * decay
+        update_delta = (variable - math_ops.cast(value, variable.dtype)) * decay
         return state_ops.assign_sub(variable, update_delta, name=scope)
 
   def _fused_batch_norm(self, inputs, training):
@@ -617,6 +619,10 @@ class BatchNormalization(Layer):
     else:
       mean, variance = self.moving_mean, self.moving_variance
 
+    mean = math_ops.cast(mean, inputs.dtype)
+    variance = math_ops.cast(variance, inputs.dtype)
+    if offset is not None:
+      offset = math_ops.cast(offset, inputs.dtype)
     outputs = nn.batch_normalization(inputs,
                                      _broadcast(mean),
                                      _broadcast(variance),
diff --git a/tensorflow/python/keras/layers/normalization_test.py b/tensorflow/python/keras/layers/normalization_test.py
index b22f3bd1529812f6b5f63efe5cf6b6133db97f07..a97b4cac469f596112481e1b3b3f93b17ea20074 100644
--- a/tensorflow/python/keras/layers/normalization_test.py
+++ b/tensorflow/python/keras/layers/normalization_test.py
@@ -95,6 +95,24 @@ class NormalizationLayersTest(test.TestCase):
       np.testing.assert_allclose(out.mean(), 0.0, atol=1e-1)
       np.testing.assert_allclose(out.std(), 1.0, atol=1e-1)
 
+  def test_batchnorm_mixed_precision(self):
+    with self.test_session():
+      model = keras.models.Sequential()
+      norm = keras.layers.BatchNormalization(input_shape=(10,), momentum=0.8)
+      model.add(norm)
+      model.compile(loss='mse', optimizer='sgd')
+
+      # centered on 5.0, variance 10.0
+      x = np.random.normal(
+          loc=5.0, scale=10.0, size=(1000, 10)).astype(np.float16)
+      model.fit(x, x, epochs=4, verbose=0)
+      out = model.predict(x)
+      out -= keras.backend.eval(norm.beta)
+      out /= keras.backend.eval(norm.gamma)
+
+      np.testing.assert_allclose(out.mean(), 0.0, atol=1e-1)
+      np.testing.assert_allclose(out.std(), 1.0, atol=1e-1)
+
   def test_batchnorm_convnet(self):
     if test.is_gpu_available(cuda_only=True):
       with self.test_session(use_gpu=True):
diff --git a/tensorflow/python/keras/layers/recurrent.py b/tensorflow/python/keras/layers/recurrent.py
index 32d25c5a650d3b66d944eee945cafa2d6f54d405..61775da47b885be89a89b68fbd2209d6d1f38b4c 100644
--- a/tensorflow/python/keras/layers/recurrent.py
+++ b/tensorflow/python/keras/layers/recurrent.py
@@ -235,7 +235,8 @@ class RNN(Layer):
   """Base class for recurrent layers.
 
   Arguments:
-      cell: A RNN cell instance. A RNN cell is a class that has:
+      cell: A RNN cell instance or a list of RNN cell instances.
+          A RNN cell is a class that has:
           - a `call(input_at_t, states_at_t)` method, returning
               `(output_at_t, states_at_t_plus_1)`. The call method of the
               cell can also take the optional argument `constants`, see
@@ -248,9 +249,9 @@ class RNN(Layer):
               (one size per state). In this case, the first entry
               (`state_size[0]`) should be the same as
               the size of the cell output.
-          It is also possible for `cell` to be a list of RNN cell instances,
-          in which cases the cells get stacked on after the other in the RNN,
-          implementing an efficient stacked RNN.
+          In the case that `cell` is a list of RNN cell instances, the cells
+          will be stacked on after the other in the RNN, implementing an
+          efficient stacked RNN.
       return_sequences: Boolean. Whether to return the last output
           in the output sequence, or the full sequence.
       return_state: Boolean. Whether to return the last state
diff --git a/tensorflow/python/keras/layers/wrappers.py b/tensorflow/python/keras/layers/wrappers.py
index 22e1cf0b366232564398cd4d0d101ff3c5b52126..f651e03874afa5843399dcfa19188c4d8d733397 100644
--- a/tensorflow/python/keras/layers/wrappers.py
+++ b/tensorflow/python/keras/layers/wrappers.py
@@ -169,13 +169,48 @@ class TimeDistributed(Wrapper):
     super(TimeDistributed, self).__init__(layer, **kwargs)
     self.supports_masking = True
 
+  def _get_shape_tuple(self, init_tuple, tensor, start_idx, int_shape=None):
+    """Finds non-specific dimensions in the static shapes.
+
+    The static shapes are replaced with the corresponding dynamic shapes of the
+    tensor.
+
+    Arguments:
+        init_tuple: a tuple, the first part of the output shape
+        tensor: the tensor from which to get the (static and dynamic) shapes
+            as the last part of the output shape
+        start_idx: int, which indicate the first dimension to take from
+            the static shape of the tensor
+        int_shape: an alternative static shape to take as the last part
+            of the output shape
+    Returns:
+        The new int_shape with the first part from init_tuple
+        and the last part from either `int_shape` (if provided)
+        or `tensor.shape`, where every `None` is replaced by
+        the corresponding dimension from `tf.shape(tensor)`.
+    """
+    # replace all None in int_shape by K.shape
+    if int_shape is None:
+      int_shape = K.int_shape(tensor)[start_idx:]
+    if not any(not s for s in int_shape):
+      return init_tuple + tuple(int_shape)
+    shape = K.shape(tensor)
+    int_shape = list(int_shape)
+    for i, s in enumerate(int_shape):
+      if not s:
+        int_shape[i] = shape[start_idx + i]
+    return init_tuple + tuple(int_shape)
+
   def build(self, input_shape):
     input_shape = tensor_shape.TensorShape(input_shape).as_list()
     assert len(input_shape) >= 3
     self.input_spec = InputSpec(shape=input_shape)
     child_input_shape = [input_shape[0]] + input_shape[2:]
     if not self.layer.built:
-      self.layer.build(child_input_shape)
+      # The base layer class calls a conversion function on the input shape to
+      # convert it to a TensorShape. The conversion function requires a
+      # tuple which is why we cast the shape.
+      self.layer.build(tuple(child_input_shape))
       self.layer.built = True
     super(TimeDistributed, self).build()
     self.built = True
@@ -221,18 +256,24 @@ class TimeDistributed(Wrapper):
       input_length = input_shape[1]
       if not input_length:
         input_length = array_ops.shape(inputs)[1]
+      inner_input_shape = self._get_shape_tuple((-1,), inputs, 2)
       # Shape: (num_samples * timesteps, ...). And track the
       # transformation in self._input_map.
       input_uid = generic_utils.object_list_uid(inputs)
-      inputs = array_ops.reshape(inputs, (-1,) + input_shape[2:])
+      inputs = array_ops.reshape(inputs, inner_input_shape)
       self._input_map[input_uid] = inputs
       # (num_samples * timesteps, ...)
+      if generic_utils.has_arg(self.layer.call, 'mask') and mask is not None:
+        inner_mask_shape = self._get_shape_tuple((-1,), mask, 2)
+        kwargs['mask'] = K.reshape(mask, inner_mask_shape)
       y = self.layer.call(inputs, **kwargs)
       if hasattr(y, '_uses_learning_phase'):
         uses_learning_phase = y._uses_learning_phase
       # Shape: (num_samples, timesteps, ...)
       output_shape = self.compute_output_shape(input_shape).as_list()
-      y = array_ops.reshape(y, (-1, input_length) + tuple(output_shape[2:]))
+      output_shape = self._get_shape_tuple(
+          (-1, input_length), y, 1, output_shape[2:])
+      y = array_ops.reshape(y, output_shape)
 
     # Apply activity regularizer if any:
     if (hasattr(self.layer, 'activity_regularizer') and
@@ -244,6 +285,80 @@ class TimeDistributed(Wrapper):
       y._uses_learning_phase = True
     return y
 
+  def compute_mask(self, inputs, mask=None):
+    """Computes an output mask tensor for Embedding layer.
+
+    This is based on the inputs, mask, and the inner layer.
+    If batch size is specified:
+    Simply return the input `mask`. (An rnn-based implementation with
+    more than one rnn inputs is required but not supported in tf.keras yet.)
+    Otherwise we call `compute_mask` of the inner layer at each time step.
+    If the output mask at each time step is not `None`:
+    (E.g., inner layer is Masking or RNN)
+    Concatenate all of them and return the concatenation.
+    If the output mask at each time step is `None` and the input mask is not
+    `None`:(E.g., inner layer is Dense)
+    Reduce the input_mask to 2 dimensions and return it.
+    Otherwise (both the output mask and the input mask are `None`):
+    (E.g., `mask` is not used at all)
+    Return `None`.
+
+    Arguments:
+      inputs: Tensor with shape [batch size, timesteps, ...] indicating the
+          input to TimeDistributed. If static shape information is available for
+          "batch size", `mask` is returned unmodified.
+      mask: Either None (indicating no masking) or a Tensor indicating the
+          input mask for TimeDistributed. The shape can be static or dynamic.
+
+    Returns:
+      Either None (no masking), or a [batch size, timesteps, ...] Tensor with
+      an output mask for the TimeDistributed layer with the shape beyond the
+      second dimension being the value of the input mask shape(if the computed
+      output mask is none), an output mask with the shape beyond the first
+      dimension being the value of the mask shape(if mask is not None) or
+      output mask with the shape beyond the first dimension being the
+      value of the computed output shape.
+
+    """
+    # cases need to call the layer.compute_mask when input_mask is None:
+    # Masking layer and Embedding layer with mask_zero
+    input_shape = K.int_shape(inputs)
+    if input_shape[0]:
+      # batch size matters, we currently do not handle mask explicitly
+      return mask
+    inner_mask = mask
+    if inner_mask is not None:
+      inner_mask_shape = self._get_shape_tuple((-1,), mask, 2)
+      inner_mask = K.reshape(inner_mask, inner_mask_shape)
+    input_uid = generic_utils.object_list_uid(inputs)
+    inner_inputs = self._input_map[input_uid]
+    output_mask = self.layer.compute_mask(inner_inputs, inner_mask)
+    if output_mask is None:
+      if mask is None:
+        return None
+      # input_mask is not None, and output_mask is None:
+      # we should return a not-None mask
+      output_mask = mask
+      for _ in range(2, len(K.int_shape(mask))):
+        output_mask = K.any(output_mask, axis=-1)
+    else:
+      # output_mask is not None. We need to reshape it
+      input_length = input_shape[1]
+      if not input_length:
+        input_length = K.shape(inputs)[1]
+      output_mask_int_shape = K.int_shape(output_mask)
+      if output_mask_int_shape is None:
+        # if the output_mask does not have a static shape,
+        # its shape must be the same as mask's
+        if mask is not None:
+          output_mask_int_shape = K.int_shape(mask)
+        else:
+          output_mask_int_shape = K.compute_output_shape(input_shape)[:-1]
+      output_mask_shape = self._get_shape_tuple(
+          (-1, input_length), output_mask, 1, output_mask_int_shape[1:])
+      output_mask = K.reshape(output_mask, output_mask_shape)
+    return output_mask
+
 
 @tf_export('keras.layers.Bidirectional')
 class Bidirectional(Wrapper):
diff --git a/tensorflow/python/keras/layers/wrappers_test.py b/tensorflow/python/keras/layers/wrappers_test.py
index c8f0d216e6f7a3bb715286bd6e7975a5dc1ac1cc..3f268acf5cab464a1967f9df2c92cc723d03778d 100644
--- a/tensorflow/python/keras/layers/wrappers_test.py
+++ b/tensorflow/python/keras/layers/wrappers_test.py
@@ -190,8 +190,8 @@ class TimeDistributedTest(test.TestCase):
     x = keras.layers.Input(shape=(3, 2))
     layer = keras.layers.TimeDistributed(keras.layers.BatchNormalization())
     _ = layer(x)
-    assert len(layer.updates) == 2
-    assert len(layer.trainable_weights) == 2
+    self.assertEquals(len(layer.updates), 2)
+    self.assertEquals(len(layer.trainable_weights), 2)
     layer.trainable = False
     assert not layer.updates
     assert not layer.trainable_weights
@@ -199,6 +199,62 @@ class TimeDistributedTest(test.TestCase):
     assert len(layer.updates) == 2
     assert len(layer.trainable_weights) == 2
 
+  def test_TimeDistributed_with_masked_embedding_and_unspecified_shape(self):
+    with self.test_session():
+      # test with unspecified shape and Embeddings with mask_zero
+      model = keras.models.Sequential()
+      model.add(keras.layers.TimeDistributed(
+          keras.layers.Embedding(5, 6, mask_zero=True),
+          input_shape=(None, None)))  # N by t_1 by t_2 by 6
+      model.add(keras.layers.TimeDistributed(
+          keras.layers.SimpleRNN(7, return_sequences=True)))
+      model.add(keras.layers.TimeDistributed(
+          keras.layers.SimpleRNN(8, return_sequences=False)))
+      model.add(keras.layers.SimpleRNN(1, return_sequences=False))
+      model.compile(optimizer='rmsprop', loss='mse')
+      model_input = np.random.randint(low=1, high=5, size=(10, 3, 4),
+                                      dtype='int32')
+      for i in range(4):
+        model_input[i, i:, i:] = 0
+      model.fit(model_input,
+                np.random.random((10, 1)), epochs=1, batch_size=10)
+      mask_outputs = [model.layers[0].compute_mask(model.input)]
+      for layer in model.layers[1:]:
+        mask_outputs.append(layer.compute_mask(layer.input, mask_outputs[-1]))
+      func = keras.backend.function([model.input], mask_outputs[:-1])
+      mask_outputs_val = func([model_input])
+      ref_mask_val_0 = model_input > 0         # embedding layer
+      ref_mask_val_1 = ref_mask_val_0          # first RNN layer
+      ref_mask_val_2 = np.any(ref_mask_val_1, axis=-1)     # second RNN layer
+      ref_mask_val = [ref_mask_val_0, ref_mask_val_1, ref_mask_val_2]
+      for i in range(3):
+        self.assertAllEqual(mask_outputs_val[i], ref_mask_val[i])
+      self.assertIs(mask_outputs[-1], None)  # final layer
+
+  def test_TimeDistributed_with_masking_layer(self):
+    with self.test_session():
+      # test with Masking layer
+      model = keras.models.Sequential()
+      model.add(keras.layers.TimeDistributed(keras.layers.Masking(
+          mask_value=0.,), input_shape=(None, 4)))
+      model.add(keras.layers.TimeDistributed(keras.layers.Dense(5)))
+      model.compile(optimizer='rmsprop', loss='mse')
+      model_input = np.random.randint(low=1, high=5, size=(10, 3, 4))
+      for i in range(4):
+        model_input[i, i:, :] = 0.
+      model.compile(optimizer='rmsprop', loss='mse')
+      model.fit(model_input,
+                np.random.random((10, 3, 5)), epochs=1, batch_size=6)
+      mask_outputs = [model.layers[0].compute_mask(model.input)]
+      mask_outputs += [model.layers[1].compute_mask(model.layers[1].input,
+                                                    mask_outputs[-1])]
+      func = keras.backend.function([model.input], mask_outputs)
+      mask_outputs_val = func([model_input])
+      self.assertEqual((mask_outputs_val[0]).all(),
+                       model_input.all())
+      self.assertEqual((mask_outputs_val[1]).all(),
+                       model_input.all())
+
 
 class BidirectionalTest(test.TestCase):
 
diff --git a/tensorflow/python/keras/model_subclassing_test.py b/tensorflow/python/keras/model_subclassing_test.py
index b7e16a41ddaa4fc1f34ffbc0be7150cb10c7a10f..3ac4852eff6910a9861ae959f990978cea33d595 100644
--- a/tensorflow/python/keras/model_subclassing_test.py
+++ b/tensorflow/python/keras/model_subclassing_test.py
@@ -31,7 +31,7 @@ from tensorflow.python.framework import test_util
 from tensorflow.python.ops import array_ops
 from tensorflow.python.ops import resource_variable_ops
 from tensorflow.python.platform import test
-from tensorflow.python.training.checkpointable import base as checkpointable
+from tensorflow.python.training.checkpointable import data_structures
 from tensorflow.python.training.rmsprop import RMSPropOptimizer
 
 try:
@@ -679,8 +679,8 @@ class ModelSubclassingTest(test.TestCase):
       def __init__(self):
         super(Foo, self).__init__()
         self.isdep = keras.layers.Dense(1)
-        self.notdep = checkpointable.NoDependency(keras.layers.Dense(2))
-        self.notdep_var = checkpointable.NoDependency(
+        self.notdep = data_structures.NoDependency(keras.layers.Dense(2))
+        self.notdep_var = data_structures.NoDependency(
             resource_variable_ops.ResourceVariable(1., name='notdep_var'))
 
     m = Foo()
diff --git a/tensorflow/python/keras/models_test.py b/tensorflow/python/keras/models_test.py
index ad3819e6e730b48e294b340d39fddeb6d7f2d6bf..1525104ac92e4c8fc9d52d28a187083d4fc91a4a 100644
--- a/tensorflow/python/keras/models_test.py
+++ b/tensorflow/python/keras/models_test.py
@@ -37,6 +37,7 @@ class TestModelCloning(test.TestCase):
 
       model = keras.models.Sequential()
       model.add(keras.layers.Dense(4, input_shape=(4,)))
+      model.add(keras.layers.BatchNormalization())
       model.add(keras.layers.Dropout(0.5))
       model.add(keras.layers.Dense(4))
 
@@ -46,6 +47,8 @@ class TestModelCloning(test.TestCase):
     with self.test_session():
       # With placeholder creation
       new_model = keras.models.clone_model(model)
+      # update ops from batch norm needs to be included
+      self.assertEquals(len(new_model.get_updates_for(new_model.inputs)), 2)
       new_model.compile('rmsprop', 'mse')
       new_model.train_on_batch(val_a, val_out)
 
@@ -53,6 +56,7 @@ class TestModelCloning(test.TestCase):
       input_a = keras.Input(shape=(4,))
       new_model = keras.models.clone_model(
           model, input_tensors=input_a)
+      self.assertEquals(len(new_model.get_updates_for(new_model.inputs)), 2)
       new_model.compile('rmsprop', 'mse')
       new_model.train_on_batch(val_a, val_out)
 
@@ -60,6 +64,7 @@ class TestModelCloning(test.TestCase):
       input_a = keras.backend.variable(val_a)
       new_model = keras.models.clone_model(
           model, input_tensors=input_a)
+      self.assertEquals(len(new_model.get_updates_for(new_model.inputs)), 2)
       new_model.compile('rmsprop', 'mse')
       new_model.train_on_batch(None, val_out)
 
@@ -76,6 +81,7 @@ class TestModelCloning(test.TestCase):
 
       x_a = dense_1(input_a)
       x_a = keras.layers.Dropout(0.5)(x_a)
+      x_a = keras.layers.BatchNormalization()(x_a)
       x_b = dense_1(input_b)
       x_a = dense_2(x_a)
       outputs = keras.layers.add([x_a, x_b])
@@ -87,6 +93,7 @@ class TestModelCloning(test.TestCase):
     with self.test_session():
       # With placeholder creation
       new_model = keras.models.clone_model(model)
+      self.assertEquals(len(new_model.get_updates_for(new_model.inputs)), 2)
       new_model.compile('rmsprop', 'mse')
       new_model.train_on_batch([val_a, val_b], val_out)
 
@@ -95,6 +102,7 @@ class TestModelCloning(test.TestCase):
       input_b = keras.Input(shape=(4,), name='b')
       new_model = keras.models.clone_model(
           model, input_tensors=[input_a, input_b])
+      self.assertEquals(len(new_model.get_updates_for(new_model.inputs)), 2)
       new_model.compile('rmsprop', 'mse')
       new_model.train_on_batch([val_a, val_b], val_out)
 
@@ -103,6 +111,7 @@ class TestModelCloning(test.TestCase):
       input_b = keras.backend.variable(val_b)
       new_model = keras.models.clone_model(
           model, input_tensors=[input_a, input_b])
+      self.assertEquals(len(new_model.get_updates_for(new_model.inputs)), 2)
       new_model.compile('rmsprop', 'mse')
       new_model.train_on_batch(None, val_out)
 
diff --git a/tensorflow/python/keras/optimizers.py b/tensorflow/python/keras/optimizers.py
index 34951791b50e46533abb3e732441c777d71522bd..0b440185ca7ccfc4fadf5419e6ceb4c64a554e1d 100644
--- a/tensorflow/python/keras/optimizers.py
+++ b/tensorflow/python/keras/optimizers.py
@@ -19,57 +19,22 @@ from __future__ import absolute_import
 from __future__ import division
 from __future__ import print_function
 
-import copy
-
 import six
 from six.moves import zip  # pylint: disable=redefined-builtin
 
-from tensorflow.python.framework import dtypes as dtypes_module
-from tensorflow.python.framework import ops
 from tensorflow.python.keras import backend as K
 from tensorflow.python.keras.utils.generic_utils import deserialize_keras_object
 from tensorflow.python.keras.utils.generic_utils import serialize_keras_object
-from tensorflow.python.ops import control_flow_ops
+from tensorflow.python.ops import clip_ops
 from tensorflow.python.ops import math_ops
 from tensorflow.python.ops import state_ops
 from tensorflow.python.training import distribute as distribute_lib
 from tensorflow.python.training import optimizer as tf_optimizer_module
 from tensorflow.python.training import training_util
-from tensorflow.python.training.checkpointable import tracking as checkpointable
+from tensorflow.python.training.checkpointable import base as checkpointable
 from tensorflow.python.util.tf_export import tf_export
 
 
-def clip_norm(g, c, n):
-  """Clip a tensor by norm.
-
-  Arguments:
-    g: gradient tensor to clip.
-    c: clipping threshold.
-    n: norm of gradient tensor.
-
-  Returns:
-    Clipped gradient tensor.
-  """
-  if c > 0:
-    condition = n >= c
-    then_expression = lambda: math_ops.scalar_mul(c / n, g)
-    else_expression = lambda: g
-
-    # saving the shape to avoid converting sparse tensor to dense
-    if isinstance(g, ops.Tensor):
-      g_shape = copy.copy(g.get_shape())
-    elif isinstance(g, ops.IndexedSlices):
-      g_shape = copy.copy(g.dense_shape)
-    if condition.dtype != dtypes_module.bool:
-      condition = math_ops.cast(condition, 'bool')
-    g = control_flow_ops.cond(condition, then_expression, else_expression)
-    if isinstance(g, ops.Tensor):
-      g.set_shape(g_shape)
-    elif isinstance(g, ops.IndexedSlices):
-      g._dense_shape = g_shape  # pylint: disable=protected-access
-  return g
-
-
 @tf_export('keras.optimizers.Optimizer')
 class Optimizer(object):
   """Abstract optimizer base class.
@@ -91,6 +56,9 @@ class Optimizer(object):
       if k not in allowed_kwargs:
         raise TypeError('Unexpected keyword argument '
                         'passed to optimizer: ' + str(k))
+      # checks that clipnorm >= 0 and clipvalue >= 0
+      if kwargs[k] < 0:
+        raise ValueError('Expected {} >= 0, received: {}'.format(k, kwargs[k]))
     self.__dict__.update(kwargs)
     self.updates = []
     self.weights = []
@@ -119,12 +87,13 @@ class Optimizer(object):
                        'gradient defined (i.e. are differentiable). '
                        'Common ops without gradient: '
                        'K.argmax, K.round, K.eval.')
-    if hasattr(self, 'clipnorm') and self.clipnorm > 0:
-      norm = K.sqrt(
-          sum([math_ops.reduce_sum(math_ops.square(g)) for g in grads]))
-      grads = [clip_norm(g, self.clipnorm, norm) for g in grads]
-    if hasattr(self, 'clipvalue') and self.clipvalue > 0:
-      grads = [K.clip(g, -self.clipvalue, self.clipvalue) for g in grads]
+    if hasattr(self, 'clipnorm'):
+      grads = [clip_ops.clip_by_norm(g, self.clipnorm) for g in grads]
+    if hasattr(self, 'clipvalue'):
+      grads = [
+          clip_ops.clip_by_value(g, -self.clipvalue, self.clipvalue)
+          for g in grads
+      ]
     return grads
 
   def set_weights(self, weights):
@@ -719,12 +688,13 @@ class Nadam(Optimizer):
     return dict(list(base_config.items()) + list(config.items()))
 
 
-class TFOptimizer(Optimizer, checkpointable.Checkpointable):
+class TFOptimizer(Optimizer, checkpointable.CheckpointableBase):
   """Wrapper class for native TensorFlow optimizers.
   """
 
   def __init__(self, optimizer):  # pylint: disable=super-init-not-called
     self.optimizer = optimizer
+    self._track_checkpointable(optimizer, name='optimizer')
     with K.name_scope(self.__class__.__name__):
       self.iterations = K.variable(0, dtype='int64', name='iterations')
 
diff --git a/tensorflow/python/keras/optimizers_test.py b/tensorflow/python/keras/optimizers_test.py
index 92b0cf326158adb1c6124384571a075196dbd3cc..55fc3fdcf47b4e5589e2253fffdc97d33f5b481b 100644
--- a/tensorflow/python/keras/optimizers_test.py
+++ b/tensorflow/python/keras/optimizers_test.py
@@ -145,6 +145,12 @@ class KerasOptimizersTest(test.TestCase):
     with self.assertRaises(NotImplementedError):
       optimizer.from_config(None)
 
+  def test_negative_clipvalue_or_clipnorm(self):
+    with self.assertRaises(ValueError):
+      _ = keras.optimizers.SGD(lr=0.01, clipvalue=-0.5)
+    with self.assertRaises(ValueError):
+      _ = keras.optimizers.Adam(clipnorm=-2.0)
+
 
 if __name__ == '__main__':
   test.main()
diff --git a/tensorflow/python/keras/testing_utils.py b/tensorflow/python/keras/testing_utils.py
index e7cb45d5e110dcb749ae2b1b86dd8dd5b8ded4ef..17aba7d86c236d9bb30d3a3376b3aac40b69e77d 100644
--- a/tensorflow/python/keras/testing_utils.py
+++ b/tensorflow/python/keras/testing_utils.py
@@ -18,6 +18,7 @@ from __future__ import absolute_import
 from __future__ import division
 from __future__ import print_function
 
+from collections import OrderedDict
 import numpy as np
 
 from tensorflow.python import keras
@@ -183,3 +184,76 @@ def layer_test(layer_cls, kwargs=None, input_shape=None, input_dtype=None,
 
   # for further checks in the caller function
   return actual_output
+
+
+def _combine_named_parameters(**kwargs):
+  """Generate combinations based on its keyword arguments.
+
+  Two sets of returned combinations can be concatenated using +.  Their product
+  can be computed using `times()`.
+
+  Args:
+    **kwargs: keyword arguments of form `option=[possibilities, ...]`
+         or `option=the_only_possibility`.
+
+  Returns:
+    a list of dictionaries for each combination. Keys in the dictionaries are
+    the keyword argument names.  Each key has one value - one of the
+    corresponding keyword argument values.
+  """
+  if not kwargs:
+    return [OrderedDict()]
+
+  sort_by_key = lambda k: k[0][0]
+  kwargs = OrderedDict(sorted(kwargs.items(), key=sort_by_key))
+  first = list(kwargs.items())[0]
+
+  rest = dict(list(kwargs.items())[1:])
+  rest_combined = _combine_named_parameters(**rest)
+
+  key = first[0]
+  values = first[1]
+  if not isinstance(values, list):
+    values = [values]
+
+  combinations = [
+      OrderedDict(sorted(list(combined.items()) + [(key, v)], key=sort_by_key))
+      for v in values
+      for combined in rest_combined
+  ]
+  return combinations
+
+
+def generate_combinations_with_testcase_name(**kwargs):
+  """Generate combinations based on its keyword arguments using combine().
+
+  This function calls combine() and appends a testcase name to the list of
+  dictionaries returned. The 'testcase_name' key is a required for named
+  parameterized tests.
+
+  Args:
+    **kwargs: keyword arguments of form `option=[possibilities, ...]`
+         or `option=the_only_possibility`.
+
+  Returns:
+    a list of dictionaries for each combination. Keys in the dictionaries are
+    the keyword argument names.  Each key has one value - one of the
+    corresponding keyword argument values.
+  """
+  combinations = _combine_named_parameters(**kwargs)
+  named_combinations = []
+  for combination in combinations:
+    assert isinstance(combination, OrderedDict)
+    name = ''.join([
+        '_{}_{}'.format(
+            ''.join(filter(str.isalnum, key)),
+            ''.join(filter(str.isalnum, str(value))))
+        for key, value in combination.items()
+    ])
+    named_combinations.append(
+        OrderedDict(
+            list(combination.items()) + [('testcase_name',
+                                          '_test{}'.format(name))]))
+
+  return named_combinations
+
diff --git a/tensorflow/python/kernel_tests/BUILD b/tensorflow/python/kernel_tests/BUILD
index 8a6614c8371744351b352243476ab1877b84b637..838cf836f1784ed97373d87b9af8889aa4d40145 100644
--- a/tensorflow/python/kernel_tests/BUILD
+++ b/tensorflow/python/kernel_tests/BUILD
@@ -1525,6 +1525,7 @@ cuda_py_test(
         "//tensorflow/python:framework_for_generated_wrappers",
         "//tensorflow/python:math_ops",
     ],
+    tags = ["no_windows_gpu"],
 )
 
 cuda_py_test(
@@ -2057,6 +2058,7 @@ cuda_py_test(
         "//tensorflow/python:framework_for_generated_wrappers",
         "//tensorflow/python:math_ops",
     ],
+    tags = ["no_windows_gpu"],
 )
 
 tf_py_test(
@@ -2755,6 +2757,7 @@ cuda_py_test(
         "//tensorflow/python:embedding_ops",
         "//tensorflow/python:framework",
         "//tensorflow/python:framework_for_generated_wrappers",
+        "//tensorflow/python:init_ops",
         "//tensorflow/python:linalg_ops",
         "//tensorflow/python:math_ops",
         "//tensorflow/python:partitioned_variables",
@@ -2842,6 +2845,7 @@ cuda_py_test(
         "//tensorflow/python:math_ops",
     ],
     shard_count = 20,
+    tags = ["nomsan"],  # TODO(b/110990716) reenable
 )
 
 cuda_py_test(
diff --git a/tensorflow/python/kernel_tests/boosted_trees/prediction_ops_test.py b/tensorflow/python/kernel_tests/boosted_trees/prediction_ops_test.py
index 92cd53a031e73d4ff4ac50c2465f32a2c20545a7..4e31b1ea2a796a2e83696d278cf1b4784d177150 100644
--- a/tensorflow/python/kernel_tests/boosted_trees/prediction_ops_test.py
+++ b/tensorflow/python/kernel_tests/boosted_trees/prediction_ops_test.py
@@ -910,7 +910,7 @@ class PredictionOpsTest(test_util.TensorFlowTestCase):
       feature_1_values = [11, 27]
 
       # Example 1: tree 0: 1.14, tree 1: 5.0, tree 2: 5.0 = >
-      #            logit = 0.1*5.0+0.2*5.0+1*5
+      #            logit = 0.1*1.14+0.2*5.0+1*5
       # Example 2: tree 0: 1.14, tree 1: 7.0, tree 2: -7 = >
       #            logit= 0.1*1.14+0.2*7.0-1*7.0
       expected_logits = [[6.114], [-5.486]]
@@ -925,5 +925,147 @@ class PredictionOpsTest(test_util.TensorFlowTestCase):
       self.assertAllClose(expected_logits, logits)
 
 
+class FeatureContribsOpsTest(test_util.TensorFlowTestCase):
+  """Tests feature contribs ops for model understanding."""
+
+  def testContribsMultipleTree(self):
+    """Tests that the contribs work when we have multiple trees."""
+    with self.test_session() as session:
+      tree_ensemble_config = boosted_trees_pb2.TreeEnsemble()
+      text_format.Merge(
+          """
+        trees {
+          nodes {
+            bucketized_split {
+              feature_id: 2
+              threshold: 28
+              left_id: 1
+              right_id: 2
+            }
+            metadata {
+              gain: 7.62
+              original_leaf: {scalar: 2.1}
+            }
+          }
+          nodes {
+            leaf {
+              scalar: 1.14
+            }
+          }
+          nodes {
+            leaf {
+              scalar: 8.79
+            }
+          }
+        }
+        trees {
+          nodes {
+            bucketized_split {
+              feature_id: 2
+              threshold: 26
+              left_id: 1
+              right_id: 2
+            }
+          }
+          nodes {
+            bucketized_split {
+              feature_id: 0
+              threshold: 50
+              left_id: 3
+              right_id: 4
+            }
+            metadata {
+              original_leaf: {scalar: 5.5}
+            }
+          }
+          nodes {
+            leaf {
+              scalar: 7.0
+            }
+          }
+          nodes {
+            leaf {
+              scalar: 5.0
+            }
+          }
+          nodes {
+            leaf {
+              scalar: 6.0
+            }
+          }
+        }
+        trees {
+          nodes {
+            bucketized_split {
+              feature_id: 0
+              threshold: 34
+              left_id: 1
+              right_id: 2
+            }
+          }
+          nodes {
+            leaf {
+              scalar: -7.0
+            }
+          }
+          nodes {
+            leaf {
+              scalar: 5.0
+            }
+          }
+        }
+        tree_weights: 0.1
+        tree_weights: 0.2
+        tree_weights: 1.0
+        tree_metadata: {
+          num_layers_grown: 1}
+        tree_metadata: {
+          num_layers_grown: 2}
+        tree_metadata: {
+          num_layers_grown: 1}
+      """, tree_ensemble_config)
+
+      tree_ensemble = boosted_trees_ops.TreeEnsemble(
+          'ensemble', serialized_proto=tree_ensemble_config.SerializeToString())
+      tree_ensemble_handle = tree_ensemble.resource_handle
+      resources.initialize_resources(resources.shared_resources()).run()
+
+      feature_0_values = [36, 32]
+      feature_1_values = [13, -29]  # Unused. Feature is not in above ensemble.
+      feature_2_values = [11, 27]
+
+      # Expected logits are computed by traversing the logit path and
+      # subtracting child logits from parent logits.
+      bias = 2.1 * 0.1  # Root node of tree_0.
+      expected_feature_ids = ((2, 2, 0, 0), (2, 2, 0))
+      # example_0 :  (bias, 0.1 * 1.14, 0.2 * 5.5 + .114, 0.2 * 5. + .114,
+      # 1.0 * 5.0 + 0.2 * 5. + .114)
+      # example_1 :  (bias, 0.1 * 1.14, 0.2 * 7 + .114,
+      # 1.0 * -7. + 0.2 * 7 + .114)
+      expected_logits_paths = ((bias, 0.114, 1.214, 1.114, 6.114),
+                               (bias, 0.114, 1.514, -5.486))
+
+      bucketized_features = [
+          feature_0_values, feature_1_values, feature_2_values
+      ]
+
+      debug_op = boosted_trees_ops.example_debug_outputs(
+          tree_ensemble_handle,
+          bucketized_features=bucketized_features,
+          logits_dimension=1)
+
+      serialized_examples_debug_outputs = session.run(debug_op)
+      feature_ids = []
+      logits_paths = []
+      for example in serialized_examples_debug_outputs:
+        example_debug_outputs = boosted_trees_pb2.DebugOutput()
+        example_debug_outputs.ParseFromString(example)
+        feature_ids.append(example_debug_outputs.feature_ids)
+        logits_paths.append(example_debug_outputs.logits_path)
+
+      self.assertAllClose(feature_ids, expected_feature_ids)
+      self.assertAllClose(logits_paths, expected_logits_paths)
+
+
 if __name__ == '__main__':
   googletest.main()
diff --git a/tensorflow/python/kernel_tests/boosted_trees/training_ops_test.py b/tensorflow/python/kernel_tests/boosted_trees/training_ops_test.py
index 13b804875e94a9f8acc9c441ba2525876a3ef58f..d55240297a8b972ea926186c2fa38da5da780612 100644
--- a/tensorflow/python/kernel_tests/boosted_trees/training_ops_test.py
+++ b/tensorflow/python/kernel_tests/boosted_trees/training_ops_test.py
@@ -139,6 +139,49 @@ class UpdateTreeEnsembleOpTest(test_util.TensorFlowTestCase):
       self.assertEqual(new_stamp, 1)
       self.assertProtoEquals(expected_result, tree_ensemble)
 
+  def testBiasCenteringOnEmptyEnsemble(self):
+    """Test growing with bias centering on an empty ensemble."""
+    with self.test_session() as session:
+      # Create empty ensemble.
+      tree_ensemble = boosted_trees_ops.TreeEnsemble('ensemble')
+      tree_ensemble_handle = tree_ensemble.resource_handle
+      resources.initialize_resources(resources.shared_resources()).run()
+
+      gradients = np.array([[5.]], dtype=np.float32)
+      hessians = np.array([[24.]], dtype=np.float32)
+
+      # Grow tree ensemble.
+      grow_op = boosted_trees_ops.center_bias(
+          tree_ensemble_handle,
+          mean_gradients=gradients,
+          mean_hessians=hessians,
+          l1=0.0,
+          l2=1.0
+      )
+      session.run(grow_op)
+
+      new_stamp, serialized = session.run(tree_ensemble.serialize())
+
+      tree_ensemble = boosted_trees_pb2.TreeEnsemble()
+      tree_ensemble.ParseFromString(serialized)
+
+      expected_result = """
+        trees {
+         nodes {
+            leaf {
+              scalar: -0.2
+            }
+          }
+        }
+        tree_weights: 1.0
+        tree_metadata {
+          num_layers_grown: 0
+          is_finalized: false
+        }
+      """
+      self.assertEqual(new_stamp, 1)
+      self.assertProtoEquals(expected_result, tree_ensemble)
+
   def testGrowExistingEnsembleTreeNotFinalized(self):
     """Test growing an existing ensemble with the last tree not finalized."""
     with self.test_session() as session:
@@ -666,7 +709,6 @@ class UpdateTreeEnsembleOpTest(test_util.TensorFlowTestCase):
           num_layers_attempted: 1
           last_layer_node_start: 1
           last_layer_node_end: 3
-
         }
       """, tree_ensemble_config)
 
diff --git a/tensorflow/python/kernel_tests/constant_op_eager_test.py b/tensorflow/python/kernel_tests/constant_op_eager_test.py
index 8e9d75667d49bf9e377ccb9290a3a91786b5a1cb..a0d5557b925162b254e34e9fc0971393ec119059 100644
--- a/tensorflow/python/kernel_tests/constant_op_eager_test.py
+++ b/tensorflow/python/kernel_tests/constant_op_eager_test.py
@@ -32,6 +32,9 @@ from tensorflow.python.util import compat
 
 
 # TODO(josh11b): add tests with lists/tuples, Shape.
+# TODO(ashankar): Collapse with tests in constant_op_test.py and use something
+# like the test_util.run_in_graph_and_eager_modes decorator to confirm
+# equivalence between graph and eager execution.
 class ConstantTest(test.TestCase):
 
   def _testCpu(self, x):
@@ -280,6 +283,34 @@ class ConstantTest(test.TestCase):
     with self.assertRaisesRegexp(ValueError, None):
       constant_op.constant([[1, 2], [3], [4, 5]])
 
+  # TODO(ashankar): This test fails with graph construction since
+  # tensor_util.make_tensor_proto (invoked from constant_op.constant)
+  # does not handle iterables (it relies on numpy conversion).
+  # For consistency, should graph construction handle Python objects
+  # that implement the sequence protocol (but not numpy conversion),
+  # or should eager execution fail on such sequences?
+  def testCustomSequence(self):
+
+    # This is inspired by how many objects in pandas are implemented:
+    # - They implement the Python sequence protocol
+    # - But may raise a KeyError on __getitem__(self, 0)
+    # See https://github.com/tensorflow/tensorflow/issues/20347
+    class MySeq(object):
+
+      def __getitem__(self, key):
+        if key != 1 and key != 3:
+          raise KeyError(key)
+        return key
+
+      def __len__(self):
+        return 2
+
+      def __iter__(self):
+        l = list([1, 3])
+        return l.__iter__()
+
+    self.assertAllEqual([1, 3], self.evaluate(constant_op.constant(MySeq())))
+
 
 class AsTensorTest(test.TestCase):
 
diff --git a/tensorflow/python/kernel_tests/dct_ops_test.py b/tensorflow/python/kernel_tests/dct_ops_test.py
index 93b2ff4561bcc8fd13855cde444c4b6237d7949b..97d7e2d8f90a620b693e2c81adc616d399e13bd6 100644
--- a/tensorflow/python/kernel_tests/dct_ops_test.py
+++ b/tensorflow/python/kernel_tests/dct_ops_test.py
@@ -40,50 +40,92 @@ def try_import(name):  # pylint: disable=invalid-name
 fftpack = try_import("scipy.fftpack")
 
 
+def _np_dct2(signals, norm=None):
+  """Computes the DCT-II manually with NumPy."""
+  # X_k = sum_{n=0}^{N-1} x_n * cos(\frac{pi}{N} * (n + 0.5) * k)  k=0,...,N-1
+  dct_size = signals.shape[-1]
+  dct = np.zeros_like(signals)
+  for k in range(dct_size):
+    phi = np.cos(np.pi * (np.arange(dct_size) + 0.5) * k / dct_size)
+    dct[..., k] = np.sum(signals * phi, axis=-1)
+  # SciPy's `dct` has a scaling factor of 2.0 which we follow.
+  # https://github.com/scipy/scipy/blob/v0.15.1/scipy/fftpack/src/dct.c.src
+  if norm == "ortho":
+    # The orthonormal scaling includes a factor of 0.5 which we combine with
+    # the overall scaling of 2.0 to cancel.
+    dct[..., 0] *= np.sqrt(1.0 / dct_size)
+    dct[..., 1:] *= np.sqrt(2.0 / dct_size)
+  else:
+    dct *= 2.0
+  return dct
+
+
+def _np_dct3(signals, norm=None):
+  """Computes the DCT-III manually with NumPy."""
+  # SciPy's `dct` has a scaling factor of 2.0 which we follow.
+  # https://github.com/scipy/scipy/blob/v0.15.1/scipy/fftpack/src/dct.c.src
+  dct_size = signals.shape[-1]
+  signals = np.array(signals)  # make a copy so we can modify
+  if norm == "ortho":
+    signals[..., 0] *= np.sqrt(4.0 / dct_size)
+    signals[..., 1:] *= np.sqrt(2.0 / dct_size)
+  else:
+    signals *= 2.0
+  dct = np.zeros_like(signals)
+  # X_k = 0.5 * x_0 +
+  #       sum_{n=1}^{N-1} x_n * cos(\frac{pi}{N} * n * (k + 0.5))  k=0,...,N-1
+  half_x0 = 0.5 * signals[..., 0]
+  for k in range(dct_size):
+    phi = np.cos(np.pi * np.arange(1, dct_size) * (k + 0.5) / dct_size)
+    dct[..., k] = half_x0 + np.sum(signals[..., 1:] * phi, axis=-1)
+  return dct
+
+
+NP_DCT = {2: _np_dct2, 3: _np_dct3}
+NP_IDCT = {2: _np_dct3, 3: _np_dct2}
+
+
 class DCTOpsTest(test.TestCase):
 
-  def _np_dct2(self, signals, norm=None):
-    """Computes the DCT-II manually with NumPy."""
-    # X_k = sum_{n=0}^{N-1} x_n * cos(\frac{pi}{N} * (n + 0.5) * k)  k=0,...,N-1
-    dct_size = signals.shape[-1]
-    dct = np.zeros_like(signals)
-    for k in range(dct_size):
-      phi = np.cos(np.pi * (np.arange(dct_size) + 0.5) * k / dct_size)
-      dct[..., k] = np.sum(signals * phi, axis=-1)
-    # SciPy's `dct` has a scaling factor of 2.0 which we follow.
-    # https://github.com/scipy/scipy/blob/v0.15.1/scipy/fftpack/src/dct.c.src
-    if norm == "ortho":
-      # The orthonormal scaling includes a factor of 0.5 which we combine with
-      # the overall scaling of 2.0 to cancel.
-      dct[..., 0] *= np.sqrt(1.0 / dct_size)
-      dct[..., 1:] *= np.sqrt(2.0 / dct_size)
-    else:
-      dct *= 2.0
-    return dct
-
-  def _compare(self, signals, norm, atol=5e-4, rtol=5e-4):
-    """Compares the DCT to SciPy (if available) and a NumPy implementation."""
-    np_dct = self._np_dct2(signals, norm)
-    tf_dct = spectral_ops.dct(signals, type=2, norm=norm).eval()
+  def _compare(self, signals, norm, dct_type, atol=5e-4, rtol=5e-4):
+    """Compares (I)DCT to SciPy (if available) and a NumPy implementation."""
+    np_dct = NP_DCT[dct_type](signals, norm)
+    tf_dct = spectral_ops.dct(signals, type=dct_type, norm=norm).eval()
     self.assertAllClose(np_dct, tf_dct, atol=atol, rtol=rtol)
+    np_idct = NP_IDCT[dct_type](signals, norm)
+    tf_idct = spectral_ops.idct(signals, type=dct_type, norm=norm).eval()
+    self.assertAllClose(np_idct, tf_idct, atol=atol, rtol=rtol)
     if fftpack:
-      scipy_dct = fftpack.dct(signals, type=2, norm=norm)
+      scipy_dct = fftpack.dct(signals, type=dct_type, norm=norm)
       self.assertAllClose(scipy_dct, tf_dct, atol=atol, rtol=rtol)
+      scipy_idct = fftpack.idct(signals, type=dct_type, norm=norm)
+      self.assertAllClose(scipy_idct, tf_idct, atol=atol, rtol=rtol)
+    # Verify inverse(forward(s)) == s, up to a normalization factor.
+    tf_idct_dct = spectral_ops.idct(
+        tf_dct, type=dct_type, norm=norm).eval()
+    tf_dct_idct = spectral_ops.dct(
+        tf_idct, type=dct_type, norm=norm).eval()
+    if norm is None:
+      tf_idct_dct *= 0.5 / signals.shape[-1]
+      tf_dct_idct *= 0.5 / signals.shape[-1]
+    self.assertAllClose(signals, tf_idct_dct, atol=atol, rtol=rtol)
+    self.assertAllClose(signals, tf_dct_idct, atol=atol, rtol=rtol)
 
   def test_random(self):
     """Test randomly generated batches of data."""
     with spectral_ops_test_util.fft_kernel_label_map():
       with self.test_session(use_gpu=True):
-        for shape in ([2, 20], [1], [2], [3], [10], [2, 20], [2, 3, 25]):
+        for shape in ([1], [2], [3], [10], [2, 20], [2, 3, 25]):
           signals = np.random.rand(*shape).astype(np.float32)
           for norm in (None, "ortho"):
-            self._compare(signals, norm)
+            self._compare(signals, norm, 2)
+            self._compare(signals, norm, 3)
 
   def test_error(self):
     signals = np.random.rand(10)
     # Unsupported type.
     with self.assertRaises(ValueError):
-      spectral_ops.dct(signals, type=3)
+      spectral_ops.dct(signals, type=1)
     # Unknown normalization.
     with self.assertRaises(ValueError):
       spectral_ops.dct(signals, norm="bad")
diff --git a/tensorflow/python/kernel_tests/embedding_ops_test.py b/tensorflow/python/kernel_tests/embedding_ops_test.py
index e53ca1dcaa520b6937aefa45e2740f1c94188b09..55d75cb4749d6f1a33d6cf7a993a336d1afcf992 100644
--- a/tensorflow/python/kernel_tests/embedding_ops_test.py
+++ b/tensorflow/python/kernel_tests/embedding_ops_test.py
@@ -19,6 +19,7 @@ from __future__ import division
 from __future__ import print_function
 
 import itertools
+import math
 
 import numpy as np
 from six.moves import xrange  # pylint: disable=redefined-builtin
@@ -31,6 +32,7 @@ from tensorflow.python.ops import array_ops
 from tensorflow.python.ops import data_flow_ops
 from tensorflow.python.ops import embedding_ops
 from tensorflow.python.ops import gradient_checker
+from tensorflow.python.ops import init_ops
 from tensorflow.python.ops import linalg_ops
 from tensorflow.python.ops import math_ops
 from tensorflow.python.ops import partitioned_variables
@@ -736,6 +738,222 @@ class EmbeddingLookupSparseTest(test.TestCase):
             x, sp_ids, sp_weights, combiner="mean")
 
 
+class SafeEmbeddingLookupSparseTest(test.TestCase):
+
+  def _random_weights(self, vocab_size=4, embed_dim=4, num_shards=1):
+    assert vocab_size > 0
+    assert embed_dim > 0
+    assert num_shards > 0
+    assert num_shards <= vocab_size
+
+    embedding_weights = partitioned_variables.create_partitioned_variables(
+        shape=[vocab_size, embed_dim],
+        slicing=[num_shards, 1],
+        initializer=init_ops.truncated_normal_initializer(
+            mean=0.0, stddev=1.0 / math.sqrt(vocab_size), dtype=dtypes.float32))
+    for w in embedding_weights:
+      w.initializer.run()
+    embedding_weights = [w.eval() for w in embedding_weights]
+    return embedding_weights
+
+  def _ids_and_weights_2d(self):
+    # Each row demonstrates a test case:
+    #   Row 0: multiple valid ids, 1 invalid id, weighted mean
+    #   Row 1: all ids are invalid (leaving no valid ids after pruning)
+    #   Row 2: no ids to begin with
+    #   Row 3: single id
+    #   Row 4: all ids have <=0 weight
+    indices = [[0, 0], [0, 1], [0, 2], [1, 0], [3, 0], [4, 0], [4, 1]]
+    ids = [0, 1, -1, -1, 2, 0, 1]
+    weights = [1.0, 2.0, 1.0, 1.0, 3.0, 0.0, -0.5]
+    shape = [5, 4]
+
+    sparse_ids = sparse_tensor.SparseTensor(
+        constant_op.constant(indices, dtypes.int64),
+        constant_op.constant(ids, dtypes.int64),
+        constant_op.constant(shape, dtypes.int64))
+
+    sparse_weights = sparse_tensor.SparseTensor(
+        constant_op.constant(indices, dtypes.int64),
+        constant_op.constant(weights, dtypes.float32),
+        constant_op.constant(shape, dtypes.int64))
+
+    return sparse_ids, sparse_weights
+
+  def _ids_and_weights_3d(self):
+    # Each (2-D) index demonstrates a test case:
+    #   Index 0, 0: multiple valid ids, 1 invalid id, weighted mean
+    #   Index 0, 1: all ids are invalid (leaving no valid ids after pruning)
+    #   Index 0, 2: no ids to begin with
+    #   Index 1, 0: single id
+    #   Index 1, 1: all ids have <=0 weight
+    #   Index 1, 2: no ids to begin with
+    indices = [[0, 0, 0], [0, 0, 1], [0, 0, 2], [0, 1, 0], [1, 0, 0], [1, 1, 0],
+               [1, 1, 1]]
+    ids = [0, 1, -1, -1, 2, 0, 1]
+    weights = [1.0, 2.0, 1.0, 1.0, 3.0, 0.0, -0.5]
+    shape = [2, 3, 4]
+
+    sparse_ids = sparse_tensor.SparseTensor(
+        constant_op.constant(indices, dtypes.int64),
+        constant_op.constant(ids, dtypes.int64),
+        constant_op.constant(shape, dtypes.int64))
+
+    sparse_weights = sparse_tensor.SparseTensor(
+        constant_op.constant(indices, dtypes.int64),
+        constant_op.constant(weights, dtypes.float32),
+        constant_op.constant(shape, dtypes.int64))
+
+    return sparse_ids, sparse_weights
+
+  def test_safe_embedding_lookup_sparse_return_zero_vector(self):
+    with self.test_session():
+      embedding_weights = self._random_weights()
+      sparse_ids, sparse_weights = self._ids_and_weights_2d()
+
+      embedding_lookup_result = (embedding_ops.safe_embedding_lookup_sparse(
+          embedding_weights, sparse_ids, sparse_weights).eval())
+
+      self.assertAllClose(
+          embedding_lookup_result,
+          [(1.0 * embedding_weights[0][0] + 2.0 * embedding_weights[0][1]) /
+           3.0, [0] * 4, [0] * 4, embedding_weights[0][2], [0] * 4])
+
+  def test_safe_embedding_lookup_sparse_return_special_vector(self):
+    with self.test_session():
+      embedding_weights = self._random_weights()
+      sparse_ids, sparse_weights = self._ids_and_weights_2d()
+
+      embedding_lookup_result = (embedding_ops.safe_embedding_lookup_sparse(
+          embedding_weights, sparse_ids, sparse_weights, default_id=3).eval())
+
+      self.assertAllClose(
+          embedding_lookup_result,
+          [(1.0 * embedding_weights[0][0] + 2.0 * embedding_weights[0][1]) /
+           3.0, embedding_weights[0][3], embedding_weights[0][3],
+           embedding_weights[0][2], embedding_weights[0][3]])
+
+  def test_safe_embedding_lookup_sparse_no_weights(self):
+    with self.test_session():
+      embedding_weights = self._random_weights()
+      sparse_ids, _ = self._ids_and_weights_2d()
+
+      embedding_lookup_result = (embedding_ops.safe_embedding_lookup_sparse(
+          embedding_weights, sparse_ids, None).eval())
+
+      self.assertAllClose(
+          embedding_lookup_result,
+          [(embedding_weights[0][0] + embedding_weights[0][1]) / 2.0, [0] * 4,
+           [0] * 4, embedding_weights[0][2], (
+               embedding_weights[0][0] + embedding_weights[0][1]) / 2.0])
+
+  def test_safe_embedding_lookup_sparse_partitioned(self):
+    with self.test_session():
+      embedding_weights = self._random_weights(num_shards=3)
+      sparse_ids, _ = self._ids_and_weights_2d()
+
+      embedding_lookup_result = (embedding_ops.safe_embedding_lookup_sparse(
+          embedding_weights, sparse_ids, None).eval())
+
+      embedding_weights = list(itertools.chain(*embedding_weights))
+      self.assertAllClose(embedding_lookup_result,
+                          [(embedding_weights[0] + embedding_weights[1]) / 2.0,
+                           [0] * 4, [0] * 4, embedding_weights[2],
+                           (embedding_weights[0] + embedding_weights[1]) / 2.0])
+
+  def test_safe_embedding_lookup_sparse_partitioned_inconsistent_weights(self):
+    with self.test_session():
+      embedding_weights = self._random_weights(num_shards=3)
+      sparse_ids, sparse_weights = self._ids_and_weights_2d()
+
+      embedding_weights[1] = embedding_weights[1].astype(np.float64)
+      self.assertRaises(TypeError, embedding_ops.safe_embedding_lookup_sparse,
+                        embedding_weights, sparse_ids)
+      embedding_weights = [
+          constant_op.constant(w, dtype=dtypes.float64)
+          for w in embedding_weights
+      ]
+      self.assertRaises(ValueError, embedding_ops.safe_embedding_lookup_sparse,
+                        embedding_weights, sparse_ids, sparse_weights)
+
+  def test_safe_embedding_lookup_sparse_3d_return_zero_vector(self):
+    with self.test_session():
+      embedding_weights = self._random_weights()
+      sparse_ids, sparse_weights = self._ids_and_weights_3d()
+
+      embedding_lookup_result = (embedding_ops.safe_embedding_lookup_sparse(
+          embedding_weights, sparse_ids, sparse_weights).eval())
+
+      self.assertAllClose(embedding_lookup_result, [[
+          (1.0 * embedding_weights[0][0] + 2.0 * embedding_weights[0][1]) / 3.0,
+          [0] * 4, [0] * 4
+      ], [embedding_weights[0][2], [0] * 4, [0] * 4]])
+
+  def test_safe_embedding_lookup_sparse_3d_return_special_vector(self):
+    with self.test_session():
+      embedding_weights = self._random_weights()
+      sparse_ids, sparse_weights = self._ids_and_weights_3d()
+
+      embedding_lookup_result = (embedding_ops.safe_embedding_lookup_sparse(
+          embedding_weights, sparse_ids, sparse_weights, default_id=3).eval())
+
+      self.assertAllClose(
+          embedding_lookup_result,
+          [[(1.0 * embedding_weights[0][0] + 2.0 * embedding_weights[0][1]) /
+            3.0, embedding_weights[0][3], embedding_weights[0][3]], [
+                embedding_weights[0][2], embedding_weights[0][3],
+                embedding_weights[0][3]
+            ]])
+
+  def test_safe_embedding_lookup_sparse_3d_no_weights(self):
+    with self.test_session():
+      embedding_weights = self._random_weights()
+      sparse_ids, _ = self._ids_and_weights_3d()
+
+      embedding_lookup_result = (embedding_ops.safe_embedding_lookup_sparse(
+          embedding_weights, sparse_ids, None).eval())
+
+      self.assertAllClose(embedding_lookup_result, [[(
+          embedding_weights[0][0] + embedding_weights[0][1]) / 2.0, [0] * 4, [
+              0
+          ] * 4], [
+              embedding_weights[0][2],
+              (embedding_weights[0][0] + embedding_weights[0][1]) / 2.0, [0] * 4
+          ]])
+
+  def test_safe_embedding_lookup_sparse_3d_partitioned(self):
+    with self.test_session():
+      embedding_weights = self._random_weights(num_shards=3)
+      sparse_ids, _ = self._ids_and_weights_3d()
+
+      embedding_lookup_result = (embedding_ops.safe_embedding_lookup_sparse(
+          embedding_weights, sparse_ids, None).eval())
+
+      embedding_weights = list(itertools.chain(*embedding_weights))
+      self.assertAllClose(embedding_lookup_result, [[
+          (embedding_weights[0] + embedding_weights[1]) / 2.0, [0] * 4, [0] * 4
+      ], [
+          embedding_weights[2],
+          (embedding_weights[0] + embedding_weights[1]) / 2.0, [0] * 4
+      ]])
+
+  def test_safe_embedding_lookup_sparse_3d_partitioned_inconsistent_weights(
+      self):
+    with self.test_session():
+      embedding_weights = self._random_weights(num_shards=3)
+      sparse_ids, sparse_weights = self._ids_and_weights_3d()
+
+      embedding_weights[1] = embedding_weights[1].astype(np.float64)
+      self.assertRaises(TypeError, embedding_ops.safe_embedding_lookup_sparse,
+                        embedding_weights, sparse_ids)
+      embedding_weights = [
+          constant_op.constant(w, dtype=dtypes.float64)
+          for w in embedding_weights
+      ]
+      self.assertRaises(ValueError, embedding_ops.safe_embedding_lookup_sparse,
+                        embedding_weights, sparse_ids, sparse_weights)
+
+
 class DynamicStitchOpTest(test.TestCase):
 
   def testCint32Cpu(self):
diff --git a/tensorflow/python/kernel_tests/functional_ops_test.py b/tensorflow/python/kernel_tests/functional_ops_test.py
index 1beb0e396e686084a3bb4ee23c35b2ef93cfb01a..24800d2b7a7aec9e43419d65c73a5a7ec3e64e24 100644
--- a/tensorflow/python/kernel_tests/functional_ops_test.py
+++ b/tensorflow/python/kernel_tests/functional_ops_test.py
@@ -35,6 +35,7 @@ from tensorflow.python.ops import functional_ops
 from tensorflow.python.ops import gradients_impl
 from tensorflow.python.ops import init_ops
 from tensorflow.python.ops import math_ops
+from tensorflow.python.ops import resource_variable_ops
 from tensorflow.python.ops import variable_scope
 from tensorflow.python.ops import variables
 import tensorflow.python.ops.tensor_array_grad  # pylint: disable=unused-import
@@ -604,6 +605,25 @@ class FunctionalOpsTest(test.TestCase):
       mul = sess.run(remote_op)
       self.assertEqual(mul, [6])
 
+  def testRemoteFunctionSameDeviceDirectSession(self):
+
+    @function.Defun(dtypes.int32, dtypes.int32)
+    def _remote_fn(a, b):
+      return math_ops.multiply(a, b)
+
+    with ops.device("/cpu:0"):
+      a = variables.Variable(2, dtype=dtypes.int32)
+      b = variables.Variable(3, dtype=dtypes.int32)
+
+    with ops.device("/cpu:0"):
+      remote_op = functional_ops.remote_call(
+          args=[a, b], Tout=[dtypes.int32], f=_remote_fn, target="/cpu:0")
+
+    with self.test_session() as sess:
+      sess.run(variables.global_variables_initializer())
+      mul = sess.run(remote_op)
+      self.assertEqual(mul, [6])
+
   def testRemoteFunctionCPUGPU(self):
     if not test_util.is_gpu_available():
       self.skipTest("No GPU available")
@@ -652,6 +672,24 @@ class FunctionalOpsTest(test.TestCase):
       mul = sess.run(remote_op)
       self.assertEqual(mul, 9.0)
 
+  def testRemoteFunctionGPUCPUStrings(self):
+    if not test_util.is_gpu_available():
+      self.skipTest("No GPU available")
+
+    @function.Defun(dtypes.string)
+    def _remote_fn(inp):
+      return array_ops.identity(inp)
+
+    a = array_ops.constant("a")
+
+    with ops.device("/gpu:0"):
+      remote_op = functional_ops.remote_call(
+          args=[a], Tout=[dtypes.string], f=_remote_fn, target="/cpu:0")
+
+    with self.test_session() as sess:
+      ret = sess.run(remote_op)
+      self.assertAllEqual(ret, [b"a"])
+
   def testRemoteFunctionCrossProcess(self):
     workers, _ = test_util.create_local_cluster(2, 1)
 
@@ -1043,6 +1081,56 @@ class PartitionedCallTest(test.TestCase):
       self.assertTrue(compat.as_bytes("CPU:1") in outputs[1].eval())
       self.assertTrue(compat.as_bytes("CPU:2") in outputs[2].eval())
 
+  def testAssignAddResourceVariable(self):
+
+    v = resource_variable_ops.ResourceVariable(1.0)
+
+    @function.Defun()
+    def AssignAdd():
+      v.assign_add(1.0)
+
+    op = functional_ops.partitioned_call(
+        args=AssignAdd.captured_inputs, f=AssignAdd)
+    _ = self.evaluate(variables.global_variables_initializer())
+    _ = self.evaluate(op)
+    value = self.evaluate(v.read_value())
+    self.assertEqual(value, 2.0)
+
+  def testFunctionWithResourcesOnDifferentDevices(self):
+    if not test_util.is_gpu_available():
+      self.skipTest("No GPUs available.")
+
+    with ops.device("/cpu:0"):
+      v_cpu_zero = resource_variable_ops.ResourceVariable(
+          [0.0, 1.0, 2.0], name="v_cpu_zero")
+
+    with ops.device("/cpu:1"):
+      v_cpu_one = resource_variable_ops.ResourceVariable(
+          [0.0, 1.0, 2.0], name="v_cpu_one")
+
+    with ops.device("/gpu:0"):
+      v_gpu = resource_variable_ops.ResourceVariable(
+          [0.0, 1.0, 2.0], name="v_gpu")
+
+    def sum_gather():
+      cpu_result = math_ops.reduce_sum(array_ops.gather(v_cpu_zero, [1, 2]))
+      also_cpu_result = math_ops.reduce_sum(array_ops.gather(v_cpu_one, [1, 2]))
+      gpu_result = math_ops.reduce_sum(array_ops.gather(v_gpu, [1, 2]))
+      return cpu_result, also_cpu_result, gpu_result
+
+    defined = function.Defun()(sum_gather)
+    with self.test_session(
+        config=config_pb2.ConfigProto(
+            allow_soft_placement=False,
+            log_device_placement=True,
+            device_count={"CPU": 2})) as sess:
+      sess.run(variables.global_variables_initializer())
+      expected = sess.run(sum_gather())
+      result = sess.run(
+          functional_ops.partitioned_call(
+              args=defined.captured_inputs, f=defined))
+      self.assertAllEqual(expected, result)
+
 
 if __name__ == "__main__":
   test.main()
diff --git a/tensorflow/python/kernel_tests/init_ops_test.py b/tensorflow/python/kernel_tests/init_ops_test.py
index 795aa67248f66e72f8f772845c4ca5b2b1b06d3d..f6097ad48984a1bb62708185ebf9782b72036e6a 100644
--- a/tensorflow/python/kernel_tests/init_ops_test.py
+++ b/tensorflow/python/kernel_tests/init_ops_test.py
@@ -364,14 +364,52 @@ class UniformUnitScalingInitializationTest(test.TestCase):
 
 class VarianceScalingInitializationTest(test.TestCase):
 
+  def testTruncatedNormalDistribution(self):
+    shape = [100, 100]
+    expect_mean = 0.
+    expect_var = 1. / shape[0]
+    init = init_ops.variance_scaling_initializer(
+        distribution='truncated_normal')
+
+    with self.test_session(use_gpu=True), \
+      test.mock.patch.object(
+          random_ops, 'truncated_normal', wraps=random_ops.truncated_normal) \
+          as mock_truncated_normal:
+      x = init(shape).eval()
+      self.assertTrue(mock_truncated_normal.called)
+
+    self.assertNear(np.mean(x), expect_mean, err=1e-2)
+    self.assertNear(np.var(x), expect_var, err=1e-2)
+
   def testNormalDistribution(self):
     shape = [100, 100]
     expect_mean = 0.
     expect_var = 1. / shape[0]
     init = init_ops.variance_scaling_initializer(distribution='normal')
 
-    with self.test_session(use_gpu=True):
+    with self.test_session(use_gpu=True), \
+      test.mock.patch.object(
+          random_ops, 'truncated_normal', wraps=random_ops.truncated_normal) \
+          as mock_truncated_normal:
+      x = init(shape).eval()
+      self.assertTrue(mock_truncated_normal.called)
+
+    self.assertNear(np.mean(x), expect_mean, err=1e-2)
+    self.assertNear(np.var(x), expect_var, err=1e-2)
+
+  def testUntruncatedNormalDistribution(self):
+    shape = [100, 100]
+    expect_mean = 0.
+    expect_var = 1. / shape[0]
+    init = init_ops.variance_scaling_initializer(
+        distribution='untruncated_normal')
+
+    with self.test_session(use_gpu=True), \
+      test.mock.patch.object(
+          random_ops, 'random_normal', wraps=random_ops.random_normal) \
+          as mock_random_normal:
       x = init(shape).eval()
+      self.assertTrue(mock_random_normal.called)
 
     self.assertNear(np.mean(x), expect_mean, err=1e-2)
     self.assertNear(np.var(x), expect_var, err=1e-2)
@@ -792,7 +830,7 @@ class ConvolutionOrthogonal1dInitializerTest(test.TestCase):
     tol = 1e-3
     gain = 3.14
     # Check orthogonality/isometry by computing the ratio between
-    # the 2-norms of the inputs and ouputs.
+    # the 2-norms of the inputs and outputs.
     for kernel_size in [[1], [2], [3], [4], [5], [6]]:
       convolution = convolutional.conv1d
       inputs = random_ops.random_normal(shape, dtype=dtype)
@@ -887,7 +925,7 @@ class ConvolutionOrthogonal2dInitializerTest(test.TestCase):
     tol = 1e-3
     gain = 3.14
     # Check orthogonality/isometry by computing the ratio between
-    # the 2-norms of the inputs and ouputs.
+    # the 2-norms of the inputs and outputs.
     for kernel_size in [[1, 1], [2, 2], [3, 3], [4, 4], [5, 5]]:
       convolution = convolutional.conv2d
       inputs = random_ops.random_normal(shape, dtype=dtype)
@@ -1012,7 +1050,7 @@ class ConvolutionOrthogonal3dInitializerTest(test.TestCase):
     tol = 1e-3
     gain = 3.14
     # Check orthogonality/isometry by computing the ratio between
-    # the 2-norms of the inputs and ouputs.
+    # the 2-norms of the inputs and outputs.
     for kernel_size in [[1, 1, 1], [2, 2, 2], [3, 3, 3]]:
       convolution = convolutional.conv3d
       inputs = random_ops.random_normal(shape, dtype=dtype)
diff --git a/tensorflow/python/kernel_tests/linalg/BUILD b/tensorflow/python/kernel_tests/linalg/BUILD
index 0123adc2c3e5c32fd86ef11e7b1f552964232abd..69d3aa401751f56ea338a5ac4b24d65e68dbddeb 100644
--- a/tensorflow/python/kernel_tests/linalg/BUILD
+++ b/tensorflow/python/kernel_tests/linalg/BUILD
@@ -107,6 +107,10 @@ cuda_py_test(
         "//tensorflow/python:random_ops",
     ],
     shard_count = 5,
+    tags = [
+        "noasan",
+        "optonly",
+    ],
 )
 
 cuda_py_test(
@@ -124,7 +128,10 @@ cuda_py_test(
         "//tensorflow/python:random_ops",
     ],
     shard_count = 5,
-    tags = ["optonly"],  # Test is flaky without optimization.
+    tags = [
+        "noasan",
+        "optonly",
+    ],
 )
 
 cuda_py_test(
@@ -141,6 +148,10 @@ cuda_py_test(
         "//tensorflow/python:platform_test",
     ],
     shard_count = 5,
+    tags = [
+        "noasan",
+        "optonly",
+    ],
 )
 
 cuda_py_test(
@@ -178,6 +189,10 @@ cuda_py_test(
         "//tensorflow/python:framework_test_lib",
         "//tensorflow/python:platform_test",
     ],
+    tags = [
+        "noasan",
+        "optonly",
+    ],
 )
 
 cuda_py_test(
@@ -214,4 +229,8 @@ cuda_py_test(
         "//tensorflow/python:platform_test",
     ],
     shard_count = 5,
+    tags = [
+        "noasan",
+        "optonly",
+    ],
 )
diff --git a/tensorflow/python/kernel_tests/linalg/linear_operator_block_diag_test.py b/tensorflow/python/kernel_tests/linalg/linear_operator_block_diag_test.py
index 2b80f01b73441185281a3e2ef4db003b150c1e12..3ede2aceaa51c2795029ba13b763fed3e2ddc441 100644
--- a/tensorflow/python/kernel_tests/linalg/linear_operator_block_diag_test.py
+++ b/tensorflow/python/kernel_tests/linalg/linear_operator_block_diag_test.py
@@ -80,7 +80,7 @@ class SquareLinearOperatorBlockDiagTest(
         build_info((2, 1, 5, 5), blocks=[(2, 1, 2, 2), (1, 3, 3)]),
     ]
 
-  def _operator_and_mat_and_feed_dict(self, build_info, dtype, use_placeholder):
+  def _operator_and_matrix(self, build_info, dtype, use_placeholder):
     shape = list(build_info.shape)
     expected_blocks = (
         build_info.__dict__["blocks"] if "blocks" in build_info.__dict__
@@ -91,26 +91,19 @@ class SquareLinearOperatorBlockDiagTest(
         for block_shape in expected_blocks
     ]
 
+    lin_op_matrices = matrices
+
     if use_placeholder:
-      matrices_ph = [
-          array_ops.placeholder(dtype=dtype) for _ in expected_blocks
-      ]
-      # Evaluate here because (i) you cannot feed a tensor, and (ii)
-      # values are random and we want the same value used for both mat and
-      # feed_dict.
-      matrices = self.evaluate(matrices)
-      operator = block_diag.LinearOperatorBlockDiag(
-          [linalg.LinearOperatorFullMatrix(
-              m_ph, is_square=True) for m_ph in matrices_ph],
-          is_square=True)
-      feed_dict = {m_ph: m for (m_ph, m) in zip(matrices_ph, matrices)}
-    else:
-      operator = block_diag.LinearOperatorBlockDiag(
-          [linalg.LinearOperatorFullMatrix(
-              m, is_square=True) for m in matrices])
-      feed_dict = None
-      # Should be auto-set.
-      self.assertTrue(operator.is_square)
+      lin_op_matrices = [
+          array_ops.placeholder_with_default(
+              matrix, shape=None) for matrix in matrices]
+
+    operator = block_diag.LinearOperatorBlockDiag(
+        [linalg.LinearOperatorFullMatrix(
+            l, is_square=True) for l in lin_op_matrices])
+
+    # Should be auto-set.
+    self.assertTrue(operator.is_square)
 
     # Broadcast the shapes.
     expected_shape = list(build_info.shape)
@@ -123,7 +116,7 @@ class SquareLinearOperatorBlockDiagTest(
       block_diag_dense.set_shape(
           expected_shape[:-2] + [expected_shape[-1], expected_shape[-1]])
 
-    return operator, block_diag_dense, feed_dict
+    return operator, block_diag_dense
 
   def test_is_x_flags(self):
     # Matrix with two positive eigenvalues, 1, and 1.
diff --git a/tensorflow/python/kernel_tests/linalg/linear_operator_circulant_test.py b/tensorflow/python/kernel_tests/linalg/linear_operator_circulant_test.py
index 5713d169696c78e996332b7a515a3ee2eedca839..7261d4bb3bc4aa24f51be21f9ac261549dca58d5 100644
--- a/tensorflow/python/kernel_tests/linalg/linear_operator_circulant_test.py
+++ b/tensorflow/python/kernel_tests/linalg/linear_operator_circulant_test.py
@@ -95,7 +95,7 @@ class LinearOperatorCirculantTestSelfAdjointOperator(
     # real, the matrix will not be real.
     return [dtypes.complex64]
 
-  def _operator_and_mat_and_feed_dict(self, build_info, dtype, use_placeholder):
+  def _operator_and_matrix(self, build_info, dtype, use_placeholder):
     shape = build_info.shape
     # For this test class, we are creating real spectrums.
     # We also want the spectrum to have eigenvalues bounded away from zero.
@@ -107,22 +107,18 @@ class LinearOperatorCirculantTestSelfAdjointOperator(
     # zero, so the operator will still be self-adjoint.
     spectrum = math_ops.cast(spectrum, dtype)
 
+    lin_op_spectrum = spectrum
+
     if use_placeholder:
-      spectrum_ph = array_ops.placeholder(dtypes.complex64)
-      # Evaluate here because (i) you cannot feed a tensor, and (ii)
-      # it is random and we want the same value used for both mat and feed_dict.
-      spectrum = spectrum.eval()
-      operator = linalg.LinearOperatorCirculant(
-          spectrum_ph, is_self_adjoint=True, input_output_dtype=dtype)
-      feed_dict = {spectrum_ph: spectrum}
-    else:
-      operator = linalg.LinearOperatorCirculant(
-          spectrum, is_self_adjoint=True, input_output_dtype=dtype)
-      feed_dict = None
+      lin_op_spectrum = array_ops.placeholder_with_default(
+          spectrum, shape=None)
+
+    operator = linalg.LinearOperatorCirculant(
+        lin_op_spectrum, is_self_adjoint=True, input_output_dtype=dtype)
 
     mat = self._spectrum_to_circulant_1d(spectrum, shape, dtype=dtype)
 
-    return operator, mat, feed_dict
+    return operator, mat
 
   def test_simple_hermitian_spectrum_gives_operator_with_zero_imag_part(self):
     with self.test_session():
@@ -149,7 +145,7 @@ class LinearOperatorCirculantTestHermitianSpectrum(
   def _dtypes_to_test(self):
     return [dtypes.float32, dtypes.complex64]
 
-  def _operator_and_mat_and_feed_dict(self, build_info, dtype, use_placeholder):
+  def _operator_and_matrix(self, build_info, dtype, use_placeholder):
     shape = build_info.shape
     # For this test class, we are creating Hermitian spectrums.
     # We also want the spectrum to have eigenvalues bounded away from zero.
@@ -172,22 +168,18 @@ class LinearOperatorCirculantTestHermitianSpectrum(
 
     spectrum = math_ops.fft(h_c)
 
+    lin_op_spectrum = spectrum
+
     if use_placeholder:
-      spectrum_ph = array_ops.placeholder(dtypes.complex64)
-      # Evaluate here because (i) you cannot feed a tensor, and (ii)
-      # it is random and we want the same value used for both mat and feed_dict.
-      spectrum = spectrum.eval()
-      operator = linalg.LinearOperatorCirculant(
-          spectrum_ph, input_output_dtype=dtype)
-      feed_dict = {spectrum_ph: spectrum}
-    else:
-      operator = linalg.LinearOperatorCirculant(
-          spectrum, input_output_dtype=dtype)
-      feed_dict = None
+      lin_op_spectrum = array_ops.placeholder_with_default(
+          spectrum, shape=None)
+
+    operator = linalg.LinearOperatorCirculant(
+        lin_op_spectrum, input_output_dtype=dtype)
 
     mat = self._spectrum_to_circulant_1d(spectrum, shape, dtype=dtype)
 
-    return operator, mat, feed_dict
+    return operator, mat
 
   def test_simple_hermitian_spectrum_gives_operator_with_zero_imag_part(self):
     with self.test_session():
@@ -213,7 +205,7 @@ class LinearOperatorCirculantTestNonHermitianSpectrum(
   def _dtypes_to_test(self):
     return [dtypes.complex64]
 
-  def _operator_and_mat_and_feed_dict(self, build_info, dtype, use_placeholder):
+  def _operator_and_matrix(self, build_info, dtype, use_placeholder):
     shape = build_info.shape
     # Will be well conditioned enough to get accurate solves.
     spectrum = linear_operator_test_util.random_sign_uniform(
@@ -222,22 +214,18 @@ class LinearOperatorCirculantTestNonHermitianSpectrum(
         minval=1.,
         maxval=2.)
 
+    lin_op_spectrum = spectrum
+
     if use_placeholder:
-      spectrum_ph = array_ops.placeholder(dtypes.complex64)
-      # Evaluate here because (i) you cannot feed a tensor, and (ii)
-      # it is random and we want the same value used for both mat and feed_dict.
-      spectrum = spectrum.eval()
-      operator = linalg.LinearOperatorCirculant(
-          spectrum_ph, input_output_dtype=dtype)
-      feed_dict = {spectrum_ph: spectrum}
-    else:
-      operator = linalg.LinearOperatorCirculant(
-          spectrum, input_output_dtype=dtype)
-      feed_dict = None
+      lin_op_spectrum = array_ops.placeholder_with_default(
+          spectrum, shape=None)
+
+    operator = linalg.LinearOperatorCirculant(
+        lin_op_spectrum, input_output_dtype=dtype)
 
     mat = self._spectrum_to_circulant_1d(spectrum, shape, dtype=dtype)
 
-    return operator, mat, feed_dict
+    return operator, mat
 
   def test_simple_hermitian_spectrum_gives_operator_with_zero_imag_part(self):
     with self.test_session():
@@ -432,7 +420,7 @@ class LinearOperatorCirculant2DTestHermitianSpectrum(
   def _dtypes_to_test(self):
     return [dtypes.float32, dtypes.complex64]
 
-  def _operator_and_mat_and_feed_dict(self, build_info, dtype, use_placeholder):
+  def _operator_and_matrix(self, build_info, dtype, use_placeholder):
     shape = build_info.shape
     # For this test class, we are creating Hermitian spectrums.
     # We also want the spectrum to have eigenvalues bounded away from zero.
@@ -455,22 +443,18 @@ class LinearOperatorCirculant2DTestHermitianSpectrum(
 
     spectrum = math_ops.fft2d(h_c)
 
+    lin_op_spectrum = spectrum
+
     if use_placeholder:
-      spectrum_ph = array_ops.placeholder(dtypes.complex64)
-      # Evaluate here because (i) you cannot feed a tensor, and (ii)
-      # it is random and we want the same value used for both mat and feed_dict.
-      spectrum = spectrum.eval()
-      operator = linalg.LinearOperatorCirculant2D(
-          spectrum_ph, input_output_dtype=dtype)
-      feed_dict = {spectrum_ph: spectrum}
-    else:
-      operator = linalg.LinearOperatorCirculant2D(
-          spectrum, input_output_dtype=dtype)
-      feed_dict = None
+      lin_op_spectrum = array_ops.placeholder_with_default(
+          spectrum, shape=None)
+
+    operator = linalg.LinearOperatorCirculant2D(
+        lin_op_spectrum, input_output_dtype=dtype)
 
     mat = self._spectrum_to_circulant_2d(spectrum, shape, dtype=dtype)
 
-    return operator, mat, feed_dict
+    return operator, mat
 
 
 class LinearOperatorCirculant2DTestNonHermitianSpectrum(
@@ -486,7 +470,7 @@ class LinearOperatorCirculant2DTestNonHermitianSpectrum(
   def _dtypes_to_test(self):
     return [dtypes.complex64]
 
-  def _operator_and_mat_and_feed_dict(self, build_info, dtype, use_placeholder):
+  def _operator_and_matrix(self, build_info, dtype, use_placeholder):
     shape = build_info.shape
     # Will be well conditioned enough to get accurate solves.
     spectrum = linear_operator_test_util.random_sign_uniform(
@@ -495,22 +479,18 @@ class LinearOperatorCirculant2DTestNonHermitianSpectrum(
         minval=1.,
         maxval=2.)
 
+    lin_op_spectrum = spectrum
+
     if use_placeholder:
-      spectrum_ph = array_ops.placeholder(dtypes.complex64)
-      # Evaluate here because (i) you cannot feed a tensor, and (ii)
-      # it is random and we want the same value used for both mat and feed_dict.
-      spectrum = spectrum.eval()
-      operator = linalg.LinearOperatorCirculant2D(
-          spectrum_ph, input_output_dtype=dtype)
-      feed_dict = {spectrum_ph: spectrum}
-    else:
-      operator = linalg.LinearOperatorCirculant2D(
-          spectrum, input_output_dtype=dtype)
-      feed_dict = None
+      lin_op_spectrum = array_ops.placeholder_with_default(
+          spectrum, shape=None)
+
+    operator = linalg.LinearOperatorCirculant2D(
+        lin_op_spectrum, input_output_dtype=dtype)
 
     mat = self._spectrum_to_circulant_2d(spectrum, shape, dtype=dtype)
 
-    return operator, mat, feed_dict
+    return operator, mat
 
   def test_real_hermitian_spectrum_gives_real_symmetric_operator(self):
     with self.test_session() as sess:
diff --git a/tensorflow/python/kernel_tests/linalg/linear_operator_composition_test.py b/tensorflow/python/kernel_tests/linalg/linear_operator_composition_test.py
index f96b9ccdaacae7d8e0552ed3d74ce53808fed963..612a50bcec771f8511d20d19b312a797d531f109 100644
--- a/tensorflow/python/kernel_tests/linalg/linear_operator_composition_test.py
+++ b/tensorflow/python/kernel_tests/linalg/linear_operator_composition_test.py
@@ -44,7 +44,7 @@ class SquareLinearOperatorCompositionTest(
     self._rtol[dtypes.float32] = 1e-4
     self._rtol[dtypes.complex64] = 1e-4
 
-  def _operator_and_mat_and_feed_dict(self, build_info, dtype, use_placeholder):
+  def _operator_and_matrix(self, build_info, dtype, use_placeholder):
     sess = ops.get_default_session()
     shape = list(build_info.shape)
 
@@ -56,33 +56,23 @@ class SquareLinearOperatorCompositionTest(
         for _ in range(num_operators)
     ]
 
+    lin_op_matrices = matrices
+
     if use_placeholder:
-      matrices_ph = [
-          array_ops.placeholder(dtype=dtype) for _ in range(num_operators)
-      ]
-      # Evaluate here because (i) you cannot feed a tensor, and (ii)
-      # values are random and we want the same value used for both mat and
-      # feed_dict.
-      matrices = sess.run(matrices)
-      operator = linalg.LinearOperatorComposition(
-          [linalg.LinearOperatorFullMatrix(m_ph) for m_ph in matrices_ph],
-          is_square=True)
-      feed_dict = {m_ph: m for (m_ph, m) in zip(matrices_ph, matrices)}
-    else:
-      operator = linalg.LinearOperatorComposition(
-          [linalg.LinearOperatorFullMatrix(m) for m in matrices])
-      feed_dict = None
-      # Should be auto-set.
-      self.assertTrue(operator.is_square)
-
-    # Convert back to Tensor.  Needed if use_placeholder, since then we have
-    # already evaluated each matrix to a numpy array.
+      lin_op_matrices = [
+          array_ops.placeholder_with_default(
+              matrix, shape=None) for matrix in matrices]
+
+    operator = linalg.LinearOperatorComposition(
+        [linalg.LinearOperatorFullMatrix(l) for l in lin_op_matrices],
+        is_square=True)
+
     matmul_order_list = list(reversed(matrices))
-    mat = ops.convert_to_tensor(matmul_order_list[0])
+    mat = matmul_order_list[0]
     for other_mat in matmul_order_list[1:]:
       mat = math_ops.matmul(other_mat, mat)
 
-    return operator, mat, feed_dict
+    return operator, mat
 
   def test_is_x_flags(self):
     # Matrix with two positive eigenvalues, 1, and 1.
@@ -148,7 +138,7 @@ class NonSquareLinearOperatorCompositionTest(
     self._rtol[dtypes.float32] = 1e-4
     self._rtol[dtypes.complex64] = 1e-4
 
-  def _operator_and_mat_and_feed_dict(self, build_info, dtype, use_placeholder):
+  def _operator_and_matrix(self, build_info, dtype, use_placeholder):
     sess = ops.get_default_session()
     shape = list(build_info.shape)
 
@@ -170,30 +160,22 @@ class NonSquareLinearOperatorCompositionTest(
                 shape_2, dtype=dtype)
     ]
 
+    lin_op_matrices = matrices
+
     if use_placeholder:
-      matrices_ph = [
-          array_ops.placeholder(dtype=dtype) for _ in range(num_operators)
-      ]
-      # Evaluate here because (i) you cannot feed a tensor, and (ii)
-      # values are random and we want the same value used for both mat and
-      # feed_dict.
-      matrices = sess.run(matrices)
-      operator = linalg.LinearOperatorComposition(
-          [linalg.LinearOperatorFullMatrix(m_ph) for m_ph in matrices_ph])
-      feed_dict = {m_ph: m for (m_ph, m) in zip(matrices_ph, matrices)}
-    else:
-      operator = linalg.LinearOperatorComposition(
-          [linalg.LinearOperatorFullMatrix(m) for m in matrices])
-      feed_dict = None
-
-    # Convert back to Tensor.  Needed if use_placeholder, since then we have
-    # already evaluated each matrix to a numpy array.
+      lin_op_matrices = [
+          array_ops.placeholder_with_default(
+              matrix, shape=None) for matrix in matrices]
+
+    operator = linalg.LinearOperatorComposition(
+        [linalg.LinearOperatorFullMatrix(l) for l in lin_op_matrices])
+
     matmul_order_list = list(reversed(matrices))
-    mat = ops.convert_to_tensor(matmul_order_list[0])
+    mat = matmul_order_list[0]
     for other_mat in matmul_order_list[1:]:
       mat = math_ops.matmul(other_mat, mat)
 
-    return operator, mat, feed_dict
+    return operator, mat
 
   def test_static_shapes(self):
     operators = [
diff --git a/tensorflow/python/kernel_tests/linalg/linear_operator_diag_test.py b/tensorflow/python/kernel_tests/linalg/linear_operator_diag_test.py
index 0a0e31c716ecfa10ed93cff92fa908a240f8495e..83cc8c483f9aec6dd0ddf3f961a8180af7515e40 100644
--- a/tensorflow/python/kernel_tests/linalg/linear_operator_diag_test.py
+++ b/tensorflow/python/kernel_tests/linalg/linear_operator_diag_test.py
@@ -34,25 +34,21 @@ class LinearOperatorDiagTest(
     linear_operator_test_util.SquareLinearOperatorDerivedClassTest):
   """Most tests done in the base class LinearOperatorDerivedClassTest."""
 
-  def _operator_and_mat_and_feed_dict(self, build_info, dtype, use_placeholder):
+  def _operator_and_matrix(self, build_info, dtype, use_placeholder):
     shape = list(build_info.shape)
     diag = linear_operator_test_util.random_sign_uniform(
         shape[:-1], minval=1., maxval=2., dtype=dtype)
+
+    lin_op_diag = diag
+
     if use_placeholder:
-      diag_ph = array_ops.placeholder(dtype=dtype)
-      # Evaluate the diag here because (i) you cannot feed a tensor, and (ii)
-      # diag is random and we want the same value used for both mat and
-      # feed_dict.
-      diag = diag.eval()
-      operator = linalg.LinearOperatorDiag(diag_ph)
-      feed_dict = {diag_ph: diag}
-    else:
-      operator = linalg.LinearOperatorDiag(diag)
-      feed_dict = None
+      lin_op_diag = array_ops.placeholder_with_default(diag, shape=None)
+
+    operator = linalg.LinearOperatorDiag(lin_op_diag)
 
-    mat = array_ops.matrix_diag(diag)
+    matrix = array_ops.matrix_diag(diag)
 
-    return operator, mat, feed_dict
+    return operator, matrix
 
   def test_assert_positive_definite_raises_for_zero_eigenvalue(self):
     # Matrix with one positive eigenvalue and one zero eigenvalue.
diff --git a/tensorflow/python/kernel_tests/linalg/linear_operator_full_matrix_test.py b/tensorflow/python/kernel_tests/linalg/linear_operator_full_matrix_test.py
index b3da623b5e8d8c99c6777e75e2d49f24dab1c96b..1a40a29ec6a040ca3d98e0b27492b1379d30cb4b 100644
--- a/tensorflow/python/kernel_tests/linalg/linear_operator_full_matrix_test.py
+++ b/tensorflow/python/kernel_tests/linalg/linear_operator_full_matrix_test.py
@@ -20,7 +20,6 @@ from __future__ import print_function
 import numpy as np
 
 from tensorflow.python.framework import dtypes
-from tensorflow.python.framework import ops
 from tensorflow.python.framework import random_seed
 from tensorflow.python.ops import array_ops
 from tensorflow.python.ops import math_ops
@@ -36,30 +35,20 @@ class SquareLinearOperatorFullMatrixTest(
     linear_operator_test_util.SquareLinearOperatorDerivedClassTest):
   """Most tests done in the base class LinearOperatorDerivedClassTest."""
 
-  def _operator_and_mat_and_feed_dict(self, build_info, dtype, use_placeholder):
+  def _operator_and_matrix(self, build_info, dtype, use_placeholder):
     shape = list(build_info.shape)
 
     matrix = linear_operator_test_util.random_positive_definite_matrix(
         shape, dtype)
 
+    lin_op_matrix = matrix
+
     if use_placeholder:
-      matrix_ph = array_ops.placeholder(dtype=dtype)
-      # Evaluate here because (i) you cannot feed a tensor, and (ii)
-      # values are random and we want the same value used for both mat and
-      # feed_dict.
-      matrix = matrix.eval()
-      operator = linalg.LinearOperatorFullMatrix(matrix_ph, is_square=True)
-      feed_dict = {matrix_ph: matrix}
-    else:
-      # is_square should be auto-detected here.
-      operator = linalg.LinearOperatorFullMatrix(matrix)
-      feed_dict = None
+      lin_op_matrix = array_ops.placeholder_with_default(matrix, shape=None)
 
-    # Convert back to Tensor.  Needed if use_placeholder, since then we have
-    # already evaluated matrix to a numpy array.
-    mat = ops.convert_to_tensor(matrix)
+    operator = linalg.LinearOperatorFullMatrix(lin_op_matrix, is_square=True)
 
-    return operator, mat, feed_dict
+    return operator, matrix
 
   def test_is_x_flags(self):
     # Matrix with two positive eigenvalues.
@@ -136,32 +125,20 @@ class SquareLinearOperatorFullMatrixSymmetricPositiveDefiniteTest(
   def _dtypes_to_test(self):
     return [dtypes.float32, dtypes.float64]
 
-  def _operator_and_mat_and_feed_dict(self, build_info, dtype, use_placeholder):
+  def _operator_and_matrix(self, build_info, dtype, use_placeholder):
     shape = list(build_info.shape)
 
     matrix = linear_operator_test_util.random_positive_definite_matrix(
         shape, dtype, force_well_conditioned=True)
 
+    lin_op_matrix = matrix
+
     if use_placeholder:
-      matrix_ph = array_ops.placeholder(dtype=dtype)
-      # Evaluate here because (i) you cannot feed a tensor, and (ii)
-      # values are random and we want the same value used for both mat and
-      # feed_dict.
-      matrix = matrix.eval()
-      # is_square is auto-set because of self_adjoint/pd.
-      operator = linalg.LinearOperatorFullMatrix(
-          matrix_ph, is_self_adjoint=True, is_positive_definite=True)
-      feed_dict = {matrix_ph: matrix}
-    else:
-      operator = linalg.LinearOperatorFullMatrix(
-          matrix, is_self_adjoint=True, is_positive_definite=True)
-      feed_dict = None
-
-    # Convert back to Tensor.  Needed if use_placeholder, since then we have
-    # already evaluated matrix to a numpy array.
-    mat = ops.convert_to_tensor(matrix)
-
-    return operator, mat, feed_dict
+      lin_op_matrix = array_ops.placeholder_with_default(matrix, shape=None)
+
+    operator = linalg.LinearOperatorFullMatrix(lin_op_matrix, is_square=True)
+
+    return operator, matrix
 
   def test_is_x_flags(self):
     # Matrix with two positive eigenvalues.
@@ -210,26 +187,18 @@ class NonSquareLinearOperatorFullMatrixTest(
     linear_operator_test_util.NonSquareLinearOperatorDerivedClassTest):
   """Most tests done in the base class LinearOperatorDerivedClassTest."""
 
-  def _operator_and_mat_and_feed_dict(self, build_info, dtype, use_placeholder):
+  def _operator_and_matrix(self, build_info, dtype, use_placeholder):
     shape = list(build_info.shape)
     matrix = linear_operator_test_util.random_normal(shape, dtype=dtype)
+
+    lin_op_matrix = matrix
+
     if use_placeholder:
-      matrix_ph = array_ops.placeholder(dtype=dtype)
-      # Evaluate here because (i) you cannot feed a tensor, and (ii)
-      # values are random and we want the same value used for both mat and
-      # feed_dict.
-      matrix = matrix.eval()
-      operator = linalg.LinearOperatorFullMatrix(matrix_ph)
-      feed_dict = {matrix_ph: matrix}
-    else:
-      operator = linalg.LinearOperatorFullMatrix(matrix)
-      feed_dict = None
+      lin_op_matrix = array_ops.placeholder_with_default(matrix, shape=None)
 
-    # Convert back to Tensor.  Needed if use_placeholder, since then we have
-    # already evaluated matrix to a numpy array.
-    mat = ops.convert_to_tensor(matrix)
+    operator = linalg.LinearOperatorFullMatrix(lin_op_matrix, is_square=True)
 
-    return operator, mat, feed_dict
+    return operator, matrix
 
   def test_is_x_flags(self):
     matrix = [[3., 2., 1.], [1., 1., 1.]]
diff --git a/tensorflow/python/kernel_tests/linalg/linear_operator_identity_test.py b/tensorflow/python/kernel_tests/linalg/linear_operator_identity_test.py
index 59f63f949e96991193412d3574603e58a75cb6e5..35dcf4417c313f5cbc00c8b66b4c5d1f2e157212 100644
--- a/tensorflow/python/kernel_tests/linalg/linear_operator_identity_test.py
+++ b/tensorflow/python/kernel_tests/linalg/linear_operator_identity_test.py
@@ -43,7 +43,7 @@ class LinearOperatorIdentityTest(
     # 16bit.
     return [dtypes.float32, dtypes.float64, dtypes.complex64, dtypes.complex128]
 
-  def _operator_and_mat_and_feed_dict(self, build_info, dtype, use_placeholder):
+  def _operator_and_matrix(self, build_info, dtype, use_placeholder):
     shape = list(build_info.shape)
     assert shape[-1] == shape[-2]
 
@@ -54,13 +54,7 @@ class LinearOperatorIdentityTest(
         num_rows, batch_shape=batch_shape, dtype=dtype)
     mat = linalg_ops.eye(num_rows, batch_shape=batch_shape, dtype=dtype)
 
-    # Nothing to feed since LinearOperatorIdentity takes no Tensor args.
-    if use_placeholder:
-      feed_dict = {}
-    else:
-      feed_dict = None
-
-    return operator, mat, feed_dict
+    return operator, mat
 
   def test_assert_positive_definite(self):
     with self.test_session():
@@ -261,7 +255,7 @@ class LinearOperatorScaledIdentityTest(
     # 16bit.
     return [dtypes.float32, dtypes.float64, dtypes.complex64, dtypes.complex128]
 
-  def _operator_and_mat_and_feed_dict(self, build_info, dtype, use_placeholder):
+  def _operator_and_matrix(self, build_info, dtype, use_placeholder):
     shape = list(build_info.shape)
     assert shape[-1] == shape[-2]
 
@@ -274,24 +268,23 @@ class LinearOperatorScaledIdentityTest(
     multiplier = linear_operator_test_util.random_sign_uniform(
         shape=batch_shape, minval=1., maxval=2., dtype=dtype)
 
-    operator = linalg_lib.LinearOperatorScaledIdentity(num_rows, multiplier)
 
     # Nothing to feed since LinearOperatorScaledIdentity takes no Tensor args.
+    lin_op_multiplier = multiplier
+
     if use_placeholder:
-      multiplier_ph = array_ops.placeholder(dtype=dtype)
-      multiplier = multiplier.eval()
-      operator = linalg_lib.LinearOperatorScaledIdentity(
-          num_rows, multiplier_ph)
-      feed_dict = {multiplier_ph: multiplier}
-    else:
-      feed_dict = None
+      lin_op_multiplier = array_ops.placeholder_with_default(
+          multiplier, shape=None)
+
+    operator = linalg_lib.LinearOperatorScaledIdentity(
+        num_rows, lin_op_multiplier)
 
     multiplier_matrix = array_ops.expand_dims(
         array_ops.expand_dims(multiplier, -1), -1)
-    mat = multiplier_matrix * linalg_ops.eye(
+    matrix = multiplier_matrix * linalg_ops.eye(
         num_rows, batch_shape=batch_shape, dtype=dtype)
 
-    return operator, mat, feed_dict
+    return operator, matrix
 
   def test_assert_positive_definite_does_not_raise_when_positive(self):
     with self.test_session():
diff --git a/tensorflow/python/kernel_tests/linalg/linear_operator_kronecker_test.py b/tensorflow/python/kernel_tests/linalg/linear_operator_kronecker_test.py
index 784c730bbc8179dd1302294b2d558e8a0c532c0c..e26b946151dd8ddb923e34352feb6b483f9752fc 100644
--- a/tensorflow/python/kernel_tests/linalg/linear_operator_kronecker_test.py
+++ b/tensorflow/python/kernel_tests/linalg/linear_operator_kronecker_test.py
@@ -101,7 +101,7 @@ class SquareLinearOperatorKroneckerTest(
   def _tests_to_skip(self):
     return ["det", "solve", "solve_with_broadcast"]
 
-  def _operator_and_mat_and_feed_dict(self, build_info, dtype, use_placeholder):
+  def _operator_and_matrix(self, build_info, dtype, use_placeholder):
     shape = list(build_info.shape)
     expected_factors = build_info.__dict__["factors"]
     matrices = [
@@ -110,26 +110,15 @@ class SquareLinearOperatorKroneckerTest(
         for block_shape in expected_factors
     ]
 
+    lin_op_matrices = matrices
+
     if use_placeholder:
-      matrices_ph = [
-          array_ops.placeholder(dtype=dtype) for _ in expected_factors
-      ]
-      # Evaluate here because (i) you cannot feed a tensor, and (ii)
-      # values are random and we want the same value used for both mat and
-      # feed_dict.
-      matrices = self.evaluate(matrices)
-      operator = kronecker.LinearOperatorKronecker(
-          [linalg.LinearOperatorFullMatrix(
-              m_ph, is_square=True) for m_ph in matrices_ph],
-          is_square=True)
-      feed_dict = {m_ph: m for (m_ph, m) in zip(matrices_ph, matrices)}
-    else:
-      operator = kronecker.LinearOperatorKronecker(
-          [linalg.LinearOperatorFullMatrix(
-              m, is_square=True) for m in matrices])
-      feed_dict = None
-      # Should be auto-set.
-      self.assertTrue(operator.is_square)
+      lin_op_matrices = [
+          array_ops.placeholder_with_default(m, shape=None) for m in matrices]
+
+    operator = kronecker.LinearOperatorKronecker(
+        [linalg.LinearOperatorFullMatrix(
+            l, is_square=True) for l in lin_op_matrices])
 
     matrices = linear_operator_util.broadcast_matrix_batch_dims(matrices)
 
@@ -138,7 +127,7 @@ class SquareLinearOperatorKroneckerTest(
     if not use_placeholder:
       kronecker_dense.set_shape(shape)
 
-    return operator, kronecker_dense, feed_dict
+    return operator, kronecker_dense
 
   def test_is_x_flags(self):
     # Matrix with two positive eigenvalues, 1, and 1.
diff --git a/tensorflow/python/kernel_tests/linalg/linear_operator_low_rank_update_test.py b/tensorflow/python/kernel_tests/linalg/linear_operator_low_rank_update_test.py
index 8095f6419ef0d9543339cf1f4ee9cd4783f852b9..0e38dbd48d2252be4b3f0455ad69994ac5814126 100644
--- a/tensorflow/python/kernel_tests/linalg/linear_operator_low_rank_update_test.py
+++ b/tensorflow/python/kernel_tests/linalg/linear_operator_low_rank_update_test.py
@@ -48,12 +48,6 @@ class BaseLinearOperatorLowRankUpdatetest(object):
   # If False, A = L + UDU^H or A = L + UU^H, depending on _use_diag_update
   _use_v = None
 
-  @property
-  def _dtypes_to_test(self):
-    # TODO(langmore) Test complex types once cholesky works with them.
-    # See comment in LinearOperatorLowRankUpdate.__init__.
-    return [dtypes.float32, dtypes.float64]
-
   @property
   def _operator_build_infos(self):
     build_info = linear_operator_test_util.OperatorBuildInfo
@@ -68,7 +62,16 @@ class BaseLinearOperatorLowRankUpdatetest(object):
         build_info((3, 4, 4)),
         build_info((2, 1, 4, 4))]
 
-  def _operator_and_mat_and_feed_dict(self, build_info, dtype, use_placeholder):
+  def _gen_positive_diag(self, dtype, diag_shape):
+    if dtype.is_complex:
+      diag = linear_operator_test_util.random_uniform(
+          diag_shape, minval=1e-4, maxval=1., dtype=dtypes.float32)
+      return math_ops.cast(diag, dtype=dtype)
+
+    return linear_operator_test_util.random_uniform(
+        diag_shape, minval=1e-4, maxval=1., dtype=dtype)
+
+  def _operator_and_matrix(self, build_info, dtype, use_placeholder):
     # Recall A = L + UDV^H
     shape = list(build_info.shape)
     diag_shape = shape[:-1]
@@ -78,63 +81,46 @@ class BaseLinearOperatorLowRankUpdatetest(object):
 
     # base_operator L will be a symmetric positive definite diagonal linear
     # operator, with condition number as high as 1e4.
-    base_diag = linear_operator_test_util.random_uniform(
-        diag_shape, minval=1e-4, maxval=1., dtype=dtype)
-    base_diag_ph = array_ops.placeholder(dtype=dtype)
+    base_diag = self._gen_positive_diag(dtype, diag_shape)
+    lin_op_base_diag = base_diag
 
     # U
     u = linear_operator_test_util.random_normal_correlated_columns(
         u_perturbation_shape, dtype=dtype)
-    u_ph = array_ops.placeholder(dtype=dtype)
+    lin_op_u = u
 
     # V
     v = linear_operator_test_util.random_normal_correlated_columns(
         u_perturbation_shape, dtype=dtype)
-    v_ph = array_ops.placeholder(dtype=dtype)
+    lin_op_v = v
 
     # D
     if self._is_diag_update_positive:
-      diag_update = linear_operator_test_util.random_uniform(
-          diag_update_shape, minval=1e-4, maxval=1., dtype=dtype)
+      diag_update = self._gen_positive_diag(dtype, diag_update_shape)
     else:
       diag_update = linear_operator_test_util.random_normal(
           diag_update_shape, stddev=1e-4, dtype=dtype)
-    diag_update_ph = array_ops.placeholder(dtype=dtype)
+    lin_op_diag_update = diag_update
 
     if use_placeholder:
-      # Evaluate here because (i) you cannot feed a tensor, and (ii)
-      # values are random and we want the same value used for both mat and
-      # feed_dict.
-      base_diag = base_diag.eval()
-      u = u.eval()
-      v = v.eval()
-      diag_update = diag_update.eval()
-
-      # In all cases, set base_operator to be positive definite.
-      base_operator = linalg.LinearOperatorDiag(
-          base_diag_ph, is_positive_definite=True)
-
-      operator = linalg.LinearOperatorLowRankUpdate(
-          base_operator,
-          u=u_ph,
-          v=v_ph if self._use_v else None,
-          diag_update=diag_update_ph if self._use_diag_update else None,
-          is_diag_update_positive=self._is_diag_update_positive)
-      feed_dict = {
-          base_diag_ph: base_diag,
-          u_ph: u,
-          v_ph: v,
-          diag_update_ph: diag_update}
-    else:
-      base_operator = linalg.LinearOperatorDiag(
-          base_diag, is_positive_definite=True)
-      operator = linalg.LinearOperatorLowRankUpdate(
-          base_operator,
-          u,
-          v=v if self._use_v else None,
-          diag_update=diag_update if self._use_diag_update else None,
-          is_diag_update_positive=self._is_diag_update_positive)
-      feed_dict = None
+      lin_op_base_diag = array_ops.placeholder_with_default(
+          base_diag, shape=None)
+      lin_op_u = array_ops.placeholder_with_default(u, shape=None)
+      lin_op_v = array_ops.placeholder_with_default(v, shape=None)
+      lin_op_diag_update = array_ops.placeholder_with_default(
+          diag_update, shape=None)
+
+    base_operator = linalg.LinearOperatorDiag(
+        lin_op_base_diag,
+        is_positive_definite=True,
+        is_self_adjoint=True)
+
+    operator = linalg.LinearOperatorLowRankUpdate(
+        base_operator,
+        lin_op_u,
+        v=lin_op_v if self._use_v else None,
+        diag_update=lin_op_diag_update if self._use_diag_update else None,
+        is_diag_update_positive=self._is_diag_update_positive)
 
     # The matrix representing L
     base_diag_mat = array_ops.matrix_diag(base_diag)
@@ -146,28 +132,28 @@ class BaseLinearOperatorLowRankUpdatetest(object):
     if self._use_v and self._use_diag_update:
       # In this case, we have L + UDV^H and it isn't symmetric.
       expect_use_cholesky = False
-      mat = base_diag_mat + math_ops.matmul(
+      matrix = base_diag_mat + math_ops.matmul(
           u, math_ops.matmul(diag_update_mat, v, adjoint_b=True))
     elif self._use_v:
       # In this case, we have L + UDV^H and it isn't symmetric.
       expect_use_cholesky = False
-      mat = base_diag_mat + math_ops.matmul(u, v, adjoint_b=True)
+      matrix = base_diag_mat + math_ops.matmul(u, v, adjoint_b=True)
     elif self._use_diag_update:
       # In this case, we have L + UDU^H, which is PD if D > 0, since L > 0.
       expect_use_cholesky = self._is_diag_update_positive
-      mat = base_diag_mat + math_ops.matmul(
+      matrix = base_diag_mat + math_ops.matmul(
           u, math_ops.matmul(diag_update_mat, u, adjoint_b=True))
     else:
       # In this case, we have L + UU^H, which is PD since L > 0.
       expect_use_cholesky = True
-      mat = base_diag_mat + math_ops.matmul(u, u, adjoint_b=True)
+      matrix = base_diag_mat + math_ops.matmul(u, u, adjoint_b=True)
 
     if expect_use_cholesky:
       self.assertTrue(operator._use_cholesky)
     else:
       self.assertFalse(operator._use_cholesky)
 
-    return operator, mat, feed_dict
+    return operator, matrix
 
 
 class LinearOperatorLowRankUpdatetestWithDiagUseCholesky(
@@ -186,6 +172,7 @@ class LinearOperatorLowRankUpdatetestWithDiagUseCholesky(
     self._rtol[dtypes.float32] = 1e-5
     self._atol[dtypes.float64] = 1e-10
     self._rtol[dtypes.float64] = 1e-10
+    self._rtol[dtypes.complex64] = 1e-4
 
 
 class LinearOperatorLowRankUpdatetestWithDiagCannotUseCholesky(
@@ -205,6 +192,7 @@ class LinearOperatorLowRankUpdatetestWithDiagCannotUseCholesky(
     self._rtol[dtypes.float32] = 1e-4
     self._atol[dtypes.float64] = 1e-9
     self._rtol[dtypes.float64] = 1e-9
+    self._rtol[dtypes.complex64] = 1e-4
 
 
 class LinearOperatorLowRankUpdatetestNoDiagUseCholesky(
@@ -223,6 +211,7 @@ class LinearOperatorLowRankUpdatetestNoDiagUseCholesky(
     self._rtol[dtypes.float32] = 1e-5
     self._atol[dtypes.float64] = 1e-10
     self._rtol[dtypes.float64] = 1e-10
+    self._rtol[dtypes.complex64] = 1e-4
 
 
 class LinearOperatorLowRankUpdatetestNoDiagCannotUseCholesky(
@@ -242,6 +231,7 @@ class LinearOperatorLowRankUpdatetestNoDiagCannotUseCholesky(
     self._rtol[dtypes.float32] = 1e-4
     self._atol[dtypes.float64] = 1e-9
     self._rtol[dtypes.float64] = 1e-9
+    self._rtol[dtypes.complex64] = 1e-4
 
 
 class LinearOperatorLowRankUpdatetestWithDiagNotSquare(
diff --git a/tensorflow/python/kernel_tests/linalg/linear_operator_lower_triangular_test.py b/tensorflow/python/kernel_tests/linalg/linear_operator_lower_triangular_test.py
index a57d2f085e089fb913f09fdd9b07cf13aa7f3c35..b389e0cbdf72f2cd43751bd75e5b103b313df4b7 100644
--- a/tensorflow/python/kernel_tests/linalg/linear_operator_lower_triangular_test.py
+++ b/tensorflow/python/kernel_tests/linalg/linear_operator_lower_triangular_test.py
@@ -17,7 +17,6 @@ from __future__ import absolute_import
 from __future__ import division
 from __future__ import print_function
 
-from tensorflow.python.framework import dtypes
 from tensorflow.python.framework import random_seed
 from tensorflow.python.ops import array_ops
 from tensorflow.python.ops.linalg import linalg as linalg_lib
@@ -32,34 +31,23 @@ class LinearOperatorLowerTriangularTest(
     linear_operator_test_util.SquareLinearOperatorDerivedClassTest):
   """Most tests done in the base class LinearOperatorDerivedClassTest."""
 
-  @property
-  def _dtypes_to_test(self):
-    # TODO(langmore) Test complex types once supported by
-    # matrix_triangular_solve.
-    return [dtypes.float32, dtypes.float64]
-
-  def _operator_and_mat_and_feed_dict(self, build_info, dtype, use_placeholder):
+  def _operator_and_matrix(self, build_info, dtype, use_placeholder):
     shape = list(build_info.shape)
     # Upper triangle will be nonzero, but ignored.
     # Use a diagonal that ensures this matrix is well conditioned.
     tril = linear_operator_test_util.random_tril_matrix(
         shape, dtype=dtype, force_well_conditioned=True, remove_upper=False)
 
+    lin_op_tril = tril
+
     if use_placeholder:
-      tril_ph = array_ops.placeholder(dtype=dtype)
-      # Evaluate the tril here because (i) you cannot feed a tensor, and (ii)
-      # tril is random and we want the same value used for both mat and
-      # feed_dict.
-      tril = tril.eval()
-      operator = linalg.LinearOperatorLowerTriangular(tril_ph)
-      feed_dict = {tril_ph: tril}
-    else:
-      operator = linalg.LinearOperatorLowerTriangular(tril)
-      feed_dict = None
+      lin_op_tril = array_ops.placeholder_with_default(lin_op_tril, shape=None)
+
+    operator = linalg.LinearOperatorLowerTriangular(lin_op_tril)
 
-    mat = array_ops.matrix_band_part(tril, -1, 0)
+    matrix = array_ops.matrix_band_part(tril, -1, 0)
 
-    return operator, mat, feed_dict
+    return operator, matrix
 
   def test_assert_non_singular(self):
     # Singlular matrix with one positive eigenvalue and one zero eigenvalue.
diff --git a/tensorflow/python/kernel_tests/resource_variable_ops_test.py b/tensorflow/python/kernel_tests/resource_variable_ops_test.py
index 0fb0b8895cbc847639999ad1bd23e7fb04c86034..e358293a9077a9e8b9b7b0a0947a8db7e4864ded 100644
--- a/tensorflow/python/kernel_tests/resource_variable_ops_test.py
+++ b/tensorflow/python/kernel_tests/resource_variable_ops_test.py
@@ -852,5 +852,62 @@ class ResourceVariableOpsTest(test_util.TensorFlowTestCase):
       state_ops.scatter_update(v, [0, 1], [0, 1, 2])
 
 
+class _MixedPrecisionVariableTest(test_util.TensorFlowTestCase):
+
+  @test_util.run_in_graph_and_eager_modes()
+  def test_dense_var_to_tensor_read_dtype_same_as_var_dtype(self):
+    # read_dtype is same as dtype
+    v = resource_variable_ops.ResourceVariable(1.0, dtype=dtypes.float32)
+    v = resource_variable_ops._MixedPrecisionVariable(v, dtypes.float32)
+    if not context.executing_eagerly():
+      v.initializer.run()
+
+    # dtype is not read_dtype, return NotImplemented
+    self.assertEqual(
+        NotImplemented, v._dense_var_to_tensor(dtype=dtypes.float16))
+    self.assertEqual(NotImplemented,
+                     v._dense_var_to_tensor(dtype=dtypes.float16, as_ref=True))
+
+    # as_ref is False
+    t = v._dense_var_to_tensor(as_ref=False)
+    self.assertTrue(isinstance(t, ops.Tensor))
+    self.assertEqual(t.dtype, dtypes.float32)
+    self.assertEqual(self.evaluate(t), 1.0)
+
+    t = v._dense_var_to_tensor(dtype=dtypes.float32, as_ref=False)
+    self.assertTrue(isinstance(t, ops.Tensor))
+    self.assertEqual(t.dtype, dtypes.float32)
+    self.assertEqual(self.evaluate(t), 1.0)
+
+    # as_ref is True
+    self.assertEqual(NotImplemented, v._dense_var_to_tensor(as_ref=True))
+    self.assertEqual(NotImplemented,
+                     v._dense_var_to_tensor(dtype=dtypes.float32, as_ref=True))
+
+  @test_util.run_in_graph_and_eager_modes()
+  def test_dense_var_to_tensor_read_dtype_different_from_var_dtype(self):
+    # read_dtype is different from dtype
+    v = resource_variable_ops.ResourceVariable(1.0, dtype=dtypes.float32)
+    v = resource_variable_ops._MixedPrecisionVariable(v, dtypes.float16)
+    if not context.executing_eagerly():
+      v.initializer.run()
+
+    # as_ref is False
+    t = v._dense_var_to_tensor(as_ref=False)
+    self.assertTrue(isinstance(t, ops.Tensor))
+    self.assertEqual(t.dtype, dtypes.float16)
+    self.assertEqual(self.evaluate(t), 1.0)
+
+    t = v._dense_var_to_tensor(dtype=dtypes.float16, as_ref=False)
+    self.assertTrue(isinstance(t, ops.Tensor))
+    self.assertEqual(t.dtype, dtypes.float16)
+    self.assertEqual(self.evaluate(t), 1.0)
+
+    # as_ref is True
+    self.assertEqual(NotImplemented, v._dense_var_to_tensor(as_ref=True))
+    self.assertEqual(NotImplemented,
+                     v._dense_var_to_tensor(dtype=dtypes.float16, as_ref=True))
+
+
 if __name__ == "__main__":
   test.main()
diff --git a/tensorflow/python/kernel_tests/rnn_test.py b/tensorflow/python/kernel_tests/rnn_test.py
index 957baf8c6089a6a033f54762fef290399d80cd09..acee180a6c3e55643052b439d95a65b073288ac6 100644
--- a/tensorflow/python/kernel_tests/rnn_test.py
+++ b/tensorflow/python/kernel_tests/rnn_test.py
@@ -268,6 +268,12 @@ class RNNTest(test.TestCase):
     self._assert_cell_builds(rnn_cell_impl.GRUCell, f64, 5, 7, 3)
     self._assert_cell_builds(rnn_cell_impl.LSTMCell, f32, 5, 7, 3)
     self._assert_cell_builds(rnn_cell_impl.LSTMCell, f64, 5, 7, 3)
+    self._assert_cell_builds(contrib_rnn.IndRNNCell, f32, 5, 7, 3)
+    self._assert_cell_builds(contrib_rnn.IndRNNCell, f64, 5, 7, 3)
+    self._assert_cell_builds(contrib_rnn.IndyGRUCell, f32, 5, 7, 3)
+    self._assert_cell_builds(contrib_rnn.IndyGRUCell, f64, 5, 7, 3)
+    self._assert_cell_builds(contrib_rnn.IndyLSTMCell, f32, 5, 7, 3)
+    self._assert_cell_builds(contrib_rnn.IndyLSTMCell, f64, 5, 7, 3)
 
 
 ######### Benchmarking RNN code
diff --git a/tensorflow/python/kernel_tests/sparse_serialization_ops_test.py b/tensorflow/python/kernel_tests/sparse_serialization_ops_test.py
index 27b39a626fcc6b2705bf9e797b5293ed3f1c7820..3847cebc7dcabd66c26a4e4551e5856c6a927a33 100644
--- a/tensorflow/python/kernel_tests/sparse_serialization_ops_test.py
+++ b/tensorflow/python/kernel_tests/sparse_serialization_ops_test.py
@@ -300,6 +300,51 @@ class SerializeSparseTest(test.TestCase):
         sparse_ops.serialize_many_sparse, sparse_ops.deserialize_sparse,
         dtypes.variant)
 
+  def testVariantSerializeDeserializeScalar(self):
+    with self.test_session(use_gpu=False) as sess:
+      indices_value = np.array([[]], dtype=np.int64)
+      values_value = np.array([37], dtype=np.int32)
+      shape_value = np.array([], dtype=np.int64)
+      sparse_tensor = self._SparseTensorPlaceholder()
+      serialized = sparse_ops.serialize_sparse(
+          sparse_tensor, out_type=dtypes.variant)
+      deserialized = sparse_ops.deserialize_sparse(
+          serialized, dtype=dtypes.int32)
+      deserialized_value = sess.run(
+          deserialized,
+          feed_dict={
+              sparse_tensor.indices: indices_value,
+              sparse_tensor.values: values_value,
+              sparse_tensor.dense_shape: shape_value
+          })
+      self.assertAllEqual(deserialized_value.indices, indices_value)
+      self.assertAllEqual(deserialized_value.values, values_value)
+      self.assertAllEqual(deserialized_value.dense_shape, shape_value)
+
+  def testVariantSerializeDeserializeScalarBatch(self):
+    with self.test_session(use_gpu=False) as sess:
+      indices_value = np.array([[]], dtype=np.int64)
+      values_value = np.array([37], dtype=np.int32)
+      shape_value = np.array([], dtype=np.int64)
+      sparse_tensor = self._SparseTensorPlaceholder()
+      serialized = sparse_ops.serialize_sparse(
+          sparse_tensor, out_type=dtypes.variant)
+      stacked = array_ops.stack([serialized, serialized])
+      deserialized = sparse_ops.deserialize_sparse(stacked, dtype=dtypes.int32)
+      deserialized_value = sess.run(
+          deserialized,
+          feed_dict={
+              sparse_tensor.indices: indices_value,
+              sparse_tensor.values: values_value,
+              sparse_tensor.dense_shape: shape_value
+          })
+      self.assertAllEqual(deserialized_value.indices,
+                          np.array([[0], [1]], dtype=np.int64))
+      self.assertAllEqual(deserialized_value.values,
+                          np.array([37, 37], dtype=np.int32))
+      self.assertAllEqual(deserialized_value.dense_shape,
+                          np.array([2], dtype=np.int64))
+
   def _testDeserializeFailsWrongTypeHelper(self,
                                            serialize_fn,
                                            deserialize_fn,
diff --git a/tensorflow/python/kernel_tests/variable_scope_test.py b/tensorflow/python/kernel_tests/variable_scope_test.py
index 1e59a8c9bf58c92c6c8ef5c92ca6340027c985f8..ae2a0ab29abed2902c0095f7b0886c1afa704af4 100644
--- a/tensorflow/python/kernel_tests/variable_scope_test.py
+++ b/tensorflow/python/kernel_tests/variable_scope_test.py
@@ -1054,7 +1054,7 @@ class VariableScopeTest(test.TestCase):
           "testGetCollection_foo/testGetCollection_a:0"
       ])
 
-  def testGetTrainableVariables(self):
+  def testGetTrainableVariablesWithGetVariable(self):
     with self.test_session():
       _ = variable_scope.get_variable("testGetTrainableVariables_a", [])
       with variable_scope.variable_scope(
@@ -1062,10 +1062,72 @@ class VariableScopeTest(test.TestCase):
         _ = variable_scope.get_variable("testGetTrainableVariables_b", [])
         _ = variable_scope.get_variable(
             "testGetTrainableVariables_c", [], trainable=False)
+
+        # sync `ON_READ` sets trainable=False
+        _ = variable_scope.get_variable(
+            "testGetTrainableVariables_d", [],
+            synchronization=variable_scope.VariableSynchronization.ON_READ)
         self.assertEqual(
             [v.name for v in scope.trainable_variables()],
-            ["testGetTrainableVariables_foo/"
-             "testGetTrainableVariables_b:0"])
+            ["testGetTrainableVariables_foo/testGetTrainableVariables_b:0"])
+
+        # All other sync values sets trainable=True
+        _ = variable_scope.get_variable(
+            "testGetTrainableVariables_e", [],
+            synchronization=variable_scope.VariableSynchronization.ON_WRITE)
+        self.assertEqual([v.name for v in scope.trainable_variables()], [
+            "testGetTrainableVariables_foo/testGetTrainableVariables_b:0",
+            "testGetTrainableVariables_foo/testGetTrainableVariables_e:0"
+        ])
+
+      with self.assertRaisesRegexp(
+          ValueError, "Synchronization value can be set to "
+          "VariableSynchronization.ON_READ only for non-trainable variables. "
+          "You have specified trainable=True and "
+          "synchronization=VariableSynchronization.ON_READ."):
+        _ = variable_scope.get_variable(
+            "testGetTrainableVariables_e", [],
+            synchronization=variable_scope.VariableSynchronization.ON_READ,
+            trainable=True)
+
+  def testGetTrainableVariablesWithVariable(self):
+    with self.test_session():
+      _ = variable_scope.variable(1.0, name="testGetTrainableVariables_a")
+      with variable_scope.variable_scope(
+          "testGetTrainableVariables_foo") as scope:
+        _ = variable_scope.variable(1.0, name="testGetTrainableVariables_b")
+        _ = variable_scope.variable(
+            1.0, name="testGetTrainableVariables_c", trainable=False)
+
+        # sync `ON_READ` sets trainable=False
+        _ = variable_scope.variable(
+            1.0,
+            name="testGetTrainableVariables_d",
+            synchronization=variable_scope.VariableSynchronization.ON_READ)
+        self.assertEqual(
+            [v.name for v in scope.trainable_variables()],
+            ["testGetTrainableVariables_foo/testGetTrainableVariables_b:0"])
+
+        # All other sync values sets trainable=True
+        _ = variable_scope.variable(
+            1.0,
+            name="testGetTrainableVariables_e",
+            synchronization=variable_scope.VariableSynchronization.ON_WRITE)
+        self.assertEqual([v.name for v in scope.trainable_variables()], [
+            "testGetTrainableVariables_foo/testGetTrainableVariables_b:0",
+            "testGetTrainableVariables_foo/testGetTrainableVariables_e:0"
+        ])
+
+      with self.assertRaisesRegexp(
+          ValueError, "Synchronization value can be set to "
+          "VariableSynchronization.ON_READ only for non-trainable variables. "
+          "You have specified trainable=True and "
+          "synchronization=VariableSynchronization.ON_READ."):
+        _ = variable_scope.variable(
+            1.0,
+            name="testGetTrainableVariables_e",
+            synchronization=variable_scope.VariableSynchronization.ON_READ,
+            trainable=True)
 
   def testGetGlobalVariables(self):
     with self.test_session():
@@ -1253,6 +1315,31 @@ class VariableScopeWithCustomGetterTest(test.TestCase):
     self.assertEqual(v3, v4)
     self.assertEqual(3, called[0])  # skipped one in the first new_scope
 
+  def testSynchronizationAndAggregationWithCustomGetter(self):
+    called = [0]
+    synchronization = variable_scope.VariableSynchronization.AUTO
+    aggregation = variable_scope.VariableAggregation.NONE
+
+    def custom_getter(getter, *args, **kwargs):
+      called[0] += 1
+
+      # Verify synchronization and aggregation kwargs are as expected.
+      self.assertEqual(kwargs["synchronization"], synchronization)
+      self.assertEqual(kwargs["aggregation"], aggregation)
+      return getter(*args, **kwargs)
+
+    with variable_scope.variable_scope("scope", custom_getter=custom_getter):
+      variable_scope.get_variable("v", [1])
+    self.assertEqual(1, called[0])
+
+    with variable_scope.variable_scope("scope", custom_getter=custom_getter):
+      synchronization = variable_scope.VariableSynchronization.ON_READ
+      aggregation = variable_scope.VariableAggregation.MEAN
+      variable_scope.get_variable(
+          "v1", [1], synchronization=synchronization, aggregation=aggregation)
+
+    self.assertEqual(2, called[0])
+
   def testCustomGetterWithReuse(self):
     # Custom getter can choose to behave differently on reused variables.
     def custom_getter(getter, *args, **kwargs):
@@ -1355,6 +1442,23 @@ class VariableScopeWithCustomGetterTest(test.TestCase):
 
     self.assertAllEqual(variable_names, ["forced_name"])
 
+    called = [False]
+
+    def creater_c(next_creator, **kwargs):
+      called[0] = True
+      self.assertEqual(kwargs["synchronization"],
+                       variable_scope.VariableSynchronization.ON_WRITE)
+      self.assertEqual(kwargs["aggregation"],
+                       variable_scope.VariableAggregation.MEAN)
+      return next_creator(**kwargs)
+
+    with variable_scope.variable_creator_scope(creater_c):
+      variable_scope.get_variable(
+          "v", [],
+          synchronization=variable_scope.VariableSynchronization.ON_WRITE,
+          aggregation=variable_scope.VariableAggregation.MEAN)
+    self.assertTrue(called[0])
+
 
 class PartitionInfoTest(test.TestCase):
 
diff --git a/tensorflow/python/kernel_tests/variables_test.py b/tensorflow/python/kernel_tests/variables_test.py
index 62d596da91682c396c04efbc64cf063c8e29e7cc..2b9c62ad6f15aea65bd8d504b2f5e713ee38fc83 100644
--- a/tensorflow/python/kernel_tests/variables_test.py
+++ b/tensorflow/python/kernel_tests/variables_test.py
@@ -642,6 +642,8 @@ class PartitionedVariableTest(test.TestCase):
       iterated_partitions = list(partitioned_variable)
       self.assertEqual(2, num_partitions)
       self.assertEqual([v0, v1], iterated_partitions)
+      self.assertEqual([2], partitioned_variable.get_shape())
+      self.assertEqual([2], partitioned_variable.shape)
       self.assertEqual([2], concatenated.get_shape())
       self.assertEqual([2], concatenated.shape)
 
diff --git a/tensorflow/python/layers/base.py b/tensorflow/python/layers/base.py
index b8969a41aba1f8ee84233ce7ac398193183d292f..cf13b526175c232d0bc7389bd7c2dc9b23f75353 100644
--- a/tensorflow/python/layers/base.py
+++ b/tensorflow/python/layers/base.py
@@ -152,10 +152,17 @@ class Layer(base_layer.Layer):
             scope, default_name=self._base_name) as captured_scope:
           self._scope = captured_scope
 
-  def add_weight(self, name, shape, dtype=None,
-                 initializer=None, regularizer=None,
-                 trainable=True, constraint=None,
+  def add_weight(self,
+                 name,
+                 shape,
+                 dtype=None,
+                 initializer=None,
+                 regularizer=None,
+                 trainable=None,
+                 constraint=None,
                  use_resource=None,
+                 synchronization=vs.VariableSynchronization.AUTO,
+                 aggregation=vs.VariableAggregation.NONE,
                  partitioner=None):
     """Adds a new variable to the layer, or gets an existing one; returns it.
 
@@ -170,9 +177,19 @@ class Layer(base_layer.Layer):
         or "non_trainable_variables" (e.g. BatchNorm mean, stddev).
         Note, if the current variable scope is marked as non-trainable
         then this parameter is ignored and any added variables are also
-        marked as non-trainable.
+        marked as non-trainable. `trainable` defaults to `True` unless
+        `synchronization` is set to `ON_READ`.
       constraint: constraint instance (callable).
       use_resource: Whether to use `ResourceVariable`.
+      synchronization: Indicates when a distributed a variable will be
+        aggregated. Accepted values are constants defined in the class
+        @{tf.VariableSynchronization}. By default the synchronization is set to
+        `AUTO` and the current `DistributionStrategy` chooses
+        when to synchronize. If `synchronization` is set to `ON_READ`,
+        `trainable` must not be set to `True`.
+      aggregation: Indicates how a distributed variable will be aggregated.
+        Accepted values are constants defined in the class
+        @{tf.VariableAggregation}.
       partitioner: (optional) partitioner instance (callable).  If
         provided, when the requested variable is created it will be split
         into multiple partitions according to `partitioner`.  In this case,
@@ -190,7 +207,21 @@ class Layer(base_layer.Layer):
     Raises:
       RuntimeError: If called with partioned variable regularization and
         eager execution is enabled.
+      ValueError: When trainable has been set to True with synchronization
+        set as `ON_READ`.
     """
+    if synchronization == vs.VariableSynchronization.ON_READ:
+      if trainable:
+        raise ValueError(
+            'Synchronization value can be set to '
+            'VariableSynchronization.ON_READ only for non-trainable variables. '
+            'You have specified trainable=True and '
+            'synchronization=VariableSynchronization.ON_READ.')
+      else:
+        # Set trainable to be false when variable is to be synced on read.
+        trainable = False
+    elif trainable is None:
+      trainable = True
 
     def _should_add_regularizer(variable, existing_variable_set):
       if isinstance(variable, tf_variables.PartitionedVariable):
@@ -240,6 +271,8 @@ class Layer(base_layer.Layer):
             constraint=constraint,
             partitioner=partitioner,
             use_resource=use_resource,
+            synchronization=synchronization,
+            aggregation=aggregation,
             getter=vs.get_variable)
 
         if regularizer:
diff --git a/tensorflow/python/layers/base_test.py b/tensorflow/python/layers/base_test.py
index 298e96e711cbf8a0f625f95d737d1e7a83f4431d..d2443db6651cdab2aaf5fb2b9d678080b48bb254 100644
--- a/tensorflow/python/layers/base_test.py
+++ b/tensorflow/python/layers/base_test.py
@@ -90,12 +90,34 @@ class BaseLayerTest(test.TestCase):
 
       # regularizers only supported in GRAPH mode.
       regularizer = lambda x: math_ops.reduce_sum(x) * 1e-3
-      variable = layer.add_variable(
+      _ = layer.add_variable(
           'reg_var', [2, 2],
           initializer=init_ops.zeros_initializer(),
           regularizer=regularizer)
       self.assertEqual(len(layer.losses), 1)
 
+    # Test that sync `ON_READ` variables are defaulted to be non-trainable.
+    variable_3 = layer.add_variable(
+        'sync_on_read_var', [2, 2],
+        initializer=init_ops.zeros_initializer(),
+        synchronization=variable_scope.VariableSynchronization.ON_READ,
+        aggregation=variable_scope.VariableAggregation.SUM)
+    self.assertEqual(layer.non_trainable_variables, [variable_2, variable_3])
+
+  def testInvalidTrainableSynchronizationCombination(self):
+    layer = base_layers.Layer(name='my_layer')
+
+    with self.assertRaisesRegexp(
+        ValueError, 'Synchronization value can be set to '
+        'VariableSynchronization.ON_READ only for non-trainable variables. '
+        'You have specified trainable=True and '
+        'synchronization=VariableSynchronization.ON_READ.'):
+      _ = layer.add_variable(
+          'v', [2, 2],
+          initializer=init_ops.zeros_initializer(),
+          synchronization=variable_scope.VariableSynchronization.ON_READ,
+          trainable=True)
+
   def testReusePartitionedVaraiblesAndRegularizers(self):
     regularizer = lambda x: math_ops.reduce_sum(x) * 1e-3
     partitioner = partitioned_variables.fixed_size_partitioner(3)
@@ -104,7 +126,7 @@ class BaseLayerTest(test.TestCase):
                                          partitioner=partitioner,
                                          reuse=reuse):
         layer = base_layers.Layer(name='my_layer')
-        variable = layer.add_variable(
+        _ = layer.add_variable(
             'reg_part_var', [4, 4],
             initializer=init_ops.zeros_initializer(),
             regularizer=regularizer)
diff --git a/tensorflow/python/layers/normalization.py b/tensorflow/python/layers/normalization.py
index ece6667981bb48e6d3353fbb526fa83bcbb902f0..f7bc10a6a634d4f821894f1f07106ba340d421af 100644
--- a/tensorflow/python/layers/normalization.py
+++ b/tensorflow/python/layers/normalization.py
@@ -44,7 +44,7 @@ class BatchNormalization(keras_layers.BatchNormalization, base.Layer):
         normalized, typically the features axis/axes. For instance, after a
         `Conv2D` layer with `data_format="channels_first"`, set `axis=1`. If a
         list of axes is provided, each axis in `axis` will be normalized
-        simultaneously. Default is `-1` which takes uses last axis. Note: when
+        simultaneously. Default is `-1` which uses the last axis. Note: when
         using multi-axis batch norm, the `beta`, `gamma`, `moving_mean`, and
         `moving_variance` variables are the same rank as the input Tensor, with
         dimension size 1 in all reduced (non-axis) dimensions).
diff --git a/tensorflow/python/lib/core/numpy.h b/tensorflow/python/lib/core/numpy.h
index 98354083c7e06103166a6fe535b153eaaf201c17..d4621d61ee98b9eb4b19213145059d242c88f40c 100644
--- a/tensorflow/python/lib/core/numpy.h
+++ b/tensorflow/python/lib/core/numpy.h
@@ -30,8 +30,8 @@ limitations under the License.
 #endif
 
 // Place `` before  to avoid build failure in macOS.
-#include 
 #include 
+#include 
 
 #include "numpy/arrayobject.h"
 #include "numpy/ufuncobject.h"
diff --git a/tensorflow/python/lib/core/py_seq_tensor.cc b/tensorflow/python/lib/core/py_seq_tensor.cc
index 386be35ba2ff1fed07d6b6f5ee5d60a0f2039441..3b4f12ae31b9e905ed15e86533e648b4c95736e1 100644
--- a/tensorflow/python/lib/core/py_seq_tensor.cc
+++ b/tensorflow/python/lib/core/py_seq_tensor.cc
@@ -88,6 +88,41 @@ bool IsPyDimension(PyObject* obj) {
   return ret;
 }
 
+// Sets *elem to a NEW reference to an element in seq on success.
+// REQUIRES: PySequence_Check(seq) && PySequence_Length(seq) > 0.
+Status SampleElementFromSequence(PyObject* seq, PyObject** elem) {
+  *elem = PySequence_GetItem(seq, 0);
+  if (*elem != nullptr) return Status::OK();
+  // seq may implement the sequence protocol (i.e., implement __getitem__)
+  // but may legitimately not have a 0-th element (__getitem__(self, 0)
+  // raises a KeyError). For example:
+  // seq = pandas.Series([0, 1, 2], index=[2, 4, 6])
+  //
+  // We don't actually care for the element at key 0, any element will do
+  // for inferring the element types. All elements are expected to
+  // have the same type, and this will be validated when converting
+  // to an EagerTensor.
+  PyErr_Clear();
+  Safe_PyObjectPtr iter(PyObject_GetIter(seq));
+  if (PyErr_Occurred()) {
+    return errors::InvalidArgument("Cannot infer dtype of a ",
+                                   Py_TYPE(seq)->tp_name,
+                                   " object: ", PyExceptionFetch());
+  }
+  *elem = PyIter_Next(iter.get());
+  if (PyErr_Occurred()) {
+    return errors::InvalidArgument(
+        "Cannot infer dtype of a ", Py_TYPE(seq)->tp_name,
+        " object, as iter().next() failed: ", PyExceptionFetch());
+  }
+  if (*elem == nullptr) {
+    return errors::InvalidArgument("Cannot infer dtype of a ",
+                                   Py_TYPE(seq)->tp_name,
+                                   " object since it is an empty sequence");
+  }
+  return Status::OK();
+}
+
 Status InferShapeAndType(PyObject* obj, TensorShape* shape, DataType* dtype) {
   std::vector refs_to_clean;
   while (true) {
@@ -98,7 +133,9 @@ Status InferShapeAndType(PyObject* obj, TensorShape* shape, DataType* dtype) {
       auto length = PySequence_Length(obj);
       if (length > 0) {
         shape->AddDim(length);
-        obj = PySequence_GetItem(obj, 0);
+        PyObject* elem = nullptr;
+        TF_RETURN_IF_ERROR(SampleElementFromSequence(obj, &elem));
+        obj = elem;
         refs_to_clean.push_back(make_safe(obj));
         continue;
       } else if (length == 0) {
diff --git a/tensorflow/python/lib/core/py_util.cc b/tensorflow/python/lib/core/py_util.cc
index 572693b1cfafa04a7716e09464885faa4c92e299..6b6c82015fd2b73e410d64306ecbd613ccf1967c 100644
--- a/tensorflow/python/lib/core/py_util.cc
+++ b/tensorflow/python/lib/core/py_util.cc
@@ -16,8 +16,8 @@ limitations under the License.
 #include "tensorflow/python/lib/core/py_util.h"
 
 // Place `` before  to avoid build failure in macOS.
-#include 
 #include 
+#include 
 
 #include "tensorflow/core/lib/core/errors.h"
 #include "tensorflow/core/lib/strings/strcat.h"
diff --git a/tensorflow/python/ops/boosted_trees_ops.py b/tensorflow/python/ops/boosted_trees_ops.py
index 2a2bcdd9d69b7a0aed1e7f3d3197cf6d7dd98451..868a4f6b84df2c0d1b8b55a254f16f1be5ee1f1d 100644
--- a/tensorflow/python/ops/boosted_trees_ops.py
+++ b/tensorflow/python/ops/boosted_trees_ops.py
@@ -25,6 +25,8 @@ from tensorflow.python.ops import resources
 # Re-exporting ops used by other modules.
 # pylint: disable=unused-import
 from tensorflow.python.ops.gen_boosted_trees_ops import boosted_trees_calculate_best_gains_per_feature as calculate_best_gains_per_feature
+from tensorflow.python.ops.gen_boosted_trees_ops import boosted_trees_center_bias as center_bias
+from tensorflow.python.ops.gen_boosted_trees_ops import boosted_trees_example_debug_outputs as example_debug_outputs
 from tensorflow.python.ops.gen_boosted_trees_ops import boosted_trees_make_stats_summary as make_stats_summary
 from tensorflow.python.ops.gen_boosted_trees_ops import boosted_trees_predict as predict
 from tensorflow.python.ops.gen_boosted_trees_ops import boosted_trees_training_predict as training_predict
diff --git a/tensorflow/python/ops/collective_ops.py b/tensorflow/python/ops/collective_ops.py
index a05fd15eca12a423bf02dfb13044dd1f7630b99c..98668facd5bc56892fa00f258dfebcbe93c063da 100644
--- a/tensorflow/python/ops/collective_ops.py
+++ b/tensorflow/python/ops/collective_ops.py
@@ -22,7 +22,7 @@ from tensorflow.python.ops import gen_collective_ops
 
 
 def all_reduce(t, group_size, group_key, instance_key, merge_op, final_op,
-               subdiv_offsets=(0)):
+               subdiv_offsets=(0,)):
   """Reduces tensors collectively, across devices.
 
   Args:
diff --git a/tensorflow/python/ops/collective_ops_test.py b/tensorflow/python/ops/collective_ops_test.py
index 8e16cffdf4917ba361a3c313047e39af514273bc..9cc64ef9f631faf2f76c3dbb3e70e1f37bbe4b1a 100644
--- a/tensorflow/python/ops/collective_ops_test.py
+++ b/tensorflow/python/ops/collective_ops_test.py
@@ -37,11 +37,11 @@ class CollectiveOpTest(test.TestCase):
       with ops.device('/CPU:0'):
         in0 = constant_op.constant(t0)
         colred0 = collective_ops.all_reduce(in0, 2, group_key, instance_key,
-                                            'Add', 'Div', [0])
+                                            'Add', 'Div')
       with ops.device('/CPU:1'):
         in1 = constant_op.constant(t1)
         colred1 = collective_ops.all_reduce(in1, 2, group_key, instance_key,
-                                            'Add', 'Div', [0])
+                                            'Add', 'Div')
       run_options = config_pb2.RunOptions()
       run_options.experimental.collective_graph_key = 1
       results = sess.run([colred0, colred1], options=run_options)
diff --git a/tensorflow/python/ops/control_flow_ops.py b/tensorflow/python/ops/control_flow_ops.py
index 837c144467703f6a4a853e23f0736f27e2b5fc62..04545cceb7e166d227a46974ba3602e3cfd36512 100644
--- a/tensorflow/python/ops/control_flow_ops.py
+++ b/tensorflow/python/ops/control_flow_ops.py
@@ -2932,7 +2932,8 @@ class WhileContext(ControlFlowContext):
 
     return original_body_result, exit_vars
 
-  def BuildLoop(self, pred, body, loop_vars, shape_invariants):
+  def BuildLoop(self, pred, body, loop_vars, shape_invariants,
+                return_same_structure):
     """Add the loop termination condition and body to the graph."""
 
     # Keep original_loop_vars to identify which are TensorArrays
@@ -2943,9 +2944,10 @@ class WhileContext(ControlFlowContext):
     loop_vars = ops.convert_n_to_tensor_or_indexed_slices(loop_vars)
     try:
       self.Enter()
-      # _BuildLoop calls _update_input in several places. _lock ensures a
-      # Session.run call cannot occur between creating and mutating new ops.
-      with ops.get_default_graph()._lock:  # pylint: disable=protected-access
+      # _BuildLoop calls _update_input in several places. _mutation_lock()
+      # ensures a Session.run call cannot occur between creating and mutating
+      # new ops.
+      with ops.get_default_graph()._mutation_lock():  # pylint: disable=protected-access
         original_body_result, exit_vars = self._BuildLoop(
             pred, body, original_loop_vars, loop_vars, shape_invariants)
     finally:
@@ -2959,7 +2961,11 @@ class WhileContext(ControlFlowContext):
     packed_exit_vars = nest.pack_sequence_as(
         structure=original_body_result,
         flat_sequence=exit_vars_with_tensor_arrays)
-    return packed_exit_vars[0] if len(exit_vars) == 1 else packed_exit_vars
+
+    if return_same_structure:
+      return packed_exit_vars
+    else:
+      return packed_exit_vars[0] if len(exit_vars) == 1 else packed_exit_vars
 
   def _FixControlInputsAndContext(self, enters):
     graph = ops.get_default_graph()
@@ -2999,7 +3005,8 @@ def while_loop(cond,
                back_prop=True,
                swap_memory=False,
                name=None,
-               maximum_iterations=None):
+               maximum_iterations=None,
+               return_same_structure=False):
   """Repeat `body` while the condition `cond` is true.
 
   `cond` is a callable returning a boolean scalar tensor. `body` is a callable
@@ -3075,11 +3082,16 @@ def while_loop(cond,
       to run.  If provided, the `cond` output is AND-ed with an additional
       condition ensuring the number of iterations executed is no greater than
       `maximum_iterations`.
+    return_same_structure: If True, output has same structure as `loop_vars`. If
+      eager execution is enabled, this is ignored (and always treated as True).
 
   Returns:
-    The output tensors for the loop variables after the loop. When the length
-    of `loop_vars` is 1 this is a Tensor, TensorArray or IndexedSlice and when
-    the length of `loop_vars` is greater than 1 it returns a list.
+    The output tensors for the loop variables after the loop.
+     If `return_same_structure` is True, the return value has the same
+     structure as `loop_vars`.
+     If `return_same_structure` is False, the return value is a Tensor,
+     TensorArray or IndexedSlice if the length of `loop_vars` is 1, or a list
+     otherwise.
 
   Raises:
     TypeError: if `cond` or `body` is not callable.
@@ -3134,6 +3146,7 @@ def while_loop(cond,
   happen is that the thread updating `x` can never get ahead of the
   counter thread because the thread incrementing `x` depends on the value
   of the counter.
+  
   ```python
   import tensorflow as tf
 
@@ -3215,7 +3228,8 @@ def while_loop(cond,
     # be encapsulated in the root context.
     if loop_context.outer_context is None:
       ops.add_to_collection(ops.GraphKeys.WHILE_CONTEXT, loop_context)
-    result = loop_context.BuildLoop(cond, body, loop_vars, shape_invariants)
+    result = loop_context.BuildLoop(cond, body, loop_vars, shape_invariants,
+                                    return_same_structure)
     if maximum_iterations is not None:
       return result[1]
     else:
diff --git a/tensorflow/python/ops/control_flow_ops_test.py b/tensorflow/python/ops/control_flow_ops_test.py
index 43fe045bcb10d2fc383381f92f2bc44c5362ac7d..153548ae92cfecfe5c750746b1425abcf3747b1b 100644
--- a/tensorflow/python/ops/control_flow_ops_test.py
+++ b/tensorflow/python/ops/control_flow_ops_test.py
@@ -958,6 +958,28 @@ class WhileLoopTestCase(test_util.TensorFlowTestCase):
     # Expect a tuple since that is what the body returns.
     self.assertEqual(self.evaluate(r), (10,))
 
+  def testWhileLoopSameReturnShape_False(self):
+    i = constant_op.constant(0)
+    c = lambda i, _: math_ops.less(i, 10)
+
+    # Body returns a [tensor, []]
+    b = lambda i, _: [math_ops.add(i, 1), []]
+
+    # Should only return the tensor.
+    r = control_flow_ops.while_loop(c, b, [i, []])
+    self.assertEqual(self.evaluate(r), 10)
+
+  def testWhileLoopSameReturnShape_True(self):
+    i = constant_op.constant(0)
+    c = lambda i, _: math_ops.less(i, 10)
+
+    # Body returns a [tensor, []]
+    b = lambda i, _: [math_ops.add(i, 1), []]
+
+    # Should only return the original structure.
+    r = control_flow_ops.while_loop(c, b, [i, []], return_same_structure=True)
+    self.assertEqual(self.evaluate(r), [10, []])
+
 
 if __name__ == "__main__":
   googletest.main()
diff --git a/tensorflow/python/ops/distributions/distribution.py b/tensorflow/python/ops/distributions/distribution.py
index 41dcd401887a124780a35c3dbd84140553860485..c03ef967e68474b0313de01d48252c8274e37a21 100644
--- a/tensorflow/python/ops/distributions/distribution.py
+++ b/tensorflow/python/ops/distributions/distribution.py
@@ -212,7 +212,7 @@ class ReparameterizationType(object):
     reparameterized, and straight-through gradients are either partially
     unsupported or are not supported at all. In this case, for purposes of
     e.g. RL or variational inference, it is generally safest to wrap the
-    sample results in a `stop_gradients` call and instead use policy
+    sample results in a `stop_gradients` call and use policy
     gradients / surrogate loss instead.
   """
 
diff --git a/tensorflow/python/ops/distributions/exponential.py b/tensorflow/python/ops/distributions/exponential.py
index 24bc3f3d3eb06a01d5173cb6c7fb0f09172a0587..4325a14449dd9a13dabb65a240ede452544c761a 100644
--- a/tensorflow/python/ops/distributions/exponential.py
+++ b/tensorflow/python/ops/distributions/exponential.py
@@ -103,9 +103,6 @@ class Exponential(gamma.Gamma):
         allow_nan_stats=allow_nan_stats,
         validate_args=validate_args,
         name=name)
-    # While the Gamma distribution is not reparameterizable, the exponential
-    # distribution is.
-    self._reparameterization_type = True
     self._parameters = parameters
     self._graph_parents += [self._rate]
 
diff --git a/tensorflow/python/ops/embedding_ops.py b/tensorflow/python/ops/embedding_ops.py
index c7919e4d4c7b7be2b49a501c18861053cc81d798..27c2fa701760f000db2463aaba0b496b3550ddff 100644
--- a/tensorflow/python/ops/embedding_ops.py
+++ b/tensorflow/python/ops/embedding_ops.py
@@ -23,6 +23,7 @@ from tensorflow.python.framework import constant_op
 from tensorflow.python.framework import dtypes
 from tensorflow.python.framework import ops
 from tensorflow.python.framework import sparse_tensor
+from tensorflow.python.framework import tensor_shape
 from tensorflow.python.ops import array_ops
 from tensorflow.python.ops import clip_ops
 # Imports gradient definitions.
@@ -30,6 +31,7 @@ from tensorflow.python.ops import data_flow_grad  # pylint: disable=unused-impor
 from tensorflow.python.ops import data_flow_ops
 from tensorflow.python.ops import math_ops
 from tensorflow.python.ops import resource_variable_ops
+from tensorflow.python.ops import sparse_ops
 from tensorflow.python.ops import variables
 from tensorflow.python.platform import tf_logging as logging
 from tensorflow.python.util.tf_export import tf_export
@@ -479,3 +481,158 @@ def embedding_lookup_sparse(params,
         assert False, "Unrecognized combiner"
 
     return embeddings
+
+
+@tf_export("nn.safe_embedding_lookup_sparse")
+def safe_embedding_lookup_sparse(embedding_weights,
+                                 sparse_ids,
+                                 sparse_weights=None,
+                                 combiner='mean',
+                                 default_id=None,
+                                 name=None,
+                                 partition_strategy='div',
+                                 max_norm=None):
+  """Lookup embedding results, accounting for invalid IDs and empty features.
+
+  The partitioned embedding in `embedding_weights` must all be the same shape
+  except for the first dimension. The first dimension is allowed to vary as the
+  vocabulary size is not necessarily a multiple of `P`.  `embedding_weights`
+  may be a `PartitionedVariable` as returned by using `tf.get_variable()` with a
+  partitioner.
+
+  Invalid IDs (< 0) are pruned from input IDs and weights, as well as any IDs
+  with non-positive weight. For an entry with no features, the embedding vector
+  for `default_id` is returned, or the 0-vector if `default_id` is not supplied.
+
+  The ids and weights may be multi-dimensional. Embeddings are always aggregated
+  along the last dimension.
+
+  Args:
+    embedding_weights:  A list of `P` float `Tensor`s or values representing
+        partitioned embedding `Tensor`s.  Alternatively, a `PartitionedVariable`
+        created by partitioning along dimension 0.  The total unpartitioned
+        shape should be `[e_0, e_1, ..., e_m]`, where `e_0` represents the
+        vocab size and `e_1, ..., e_m` are the embedding dimensions.
+    sparse_ids: `SparseTensor` of shape `[d_0, d_1, ..., d_n]` containing the
+        ids. `d_0` is typically batch size.
+    sparse_weights: `SparseTensor` of same shape as `sparse_ids`, containing
+        float weights corresponding to `sparse_ids`, or `None` if all weights
+        are be assumed to be 1.0.
+    combiner: A string specifying how to combine embedding results for each
+        entry. Currently "mean", "sqrtn" and "sum" are supported, with "mean"
+        the default.
+    default_id: The id to use for an entry with no features.
+    name: A name for this operation (optional).
+    partition_strategy: A string specifying the partitioning strategy.
+        Currently `"div"` and `"mod"` are supported. Default is `"div"`.
+    max_norm: If not `None`, all embeddings are l2-normalized to max_norm before
+        combining.
+
+
+  Returns:
+    Dense `Tensor` of shape `[d_0, d_1, ..., d_{n-1}, e_1, ..., e_m]`.
+
+  Raises:
+    ValueError: if `embedding_weights` is empty.
+  """
+  if embedding_weights is None:
+    raise ValueError('Missing embedding_weights %s.' % embedding_weights)
+  if isinstance(embedding_weights, variables.PartitionedVariable):
+    embedding_weights = list(embedding_weights)  # get underlying Variables.
+  if not isinstance(embedding_weights, list):
+    embedding_weights = [embedding_weights]
+  if len(embedding_weights) < 1:
+    raise ValueError('Missing embedding_weights %s.' % embedding_weights)
+
+  dtype = sparse_weights.dtype if sparse_weights is not None else None
+  embedding_weights = [
+      ops.convert_to_tensor(w, dtype=dtype) for w in embedding_weights
+  ]
+
+  with ops.name_scope(name, 'embedding_lookup',
+                      embedding_weights + [sparse_ids,
+                                           sparse_weights]) as scope:
+    # Reshape higher-rank sparse ids and weights to linear segment ids.
+    original_shape = sparse_ids.dense_shape
+    original_rank_dim = sparse_ids.dense_shape.get_shape()[0]
+    original_rank = (
+        array_ops.size(original_shape)
+        if original_rank_dim.value is None
+        else original_rank_dim.value)
+    sparse_ids = sparse_ops.sparse_reshape(sparse_ids, [
+        math_ops.reduce_prod(
+            array_ops.slice(original_shape, [0], [original_rank - 1])),
+        array_ops.gather(original_shape, original_rank - 1)])
+    if sparse_weights is not None:
+      sparse_weights = sparse_tensor.SparseTensor(
+          sparse_ids.indices,
+          sparse_weights.values, sparse_ids.dense_shape)
+
+    # Prune invalid ids and weights.
+    sparse_ids, sparse_weights = _prune_invalid_ids(sparse_ids, sparse_weights)
+    if combiner != 'sum':
+      sparse_ids, sparse_weights = _prune_invalid_weights(
+          sparse_ids, sparse_weights)
+
+    # Fill in dummy values for empty features, if necessary.
+    sparse_ids, is_row_empty = sparse_ops.sparse_fill_empty_rows(sparse_ids,
+                                                                 default_id or
+                                                                 0)
+    if sparse_weights is not None:
+      sparse_weights, _ = sparse_ops.sparse_fill_empty_rows(sparse_weights, 1.0)
+
+    result = embedding_lookup_sparse(
+        embedding_weights,
+        sparse_ids,
+        sparse_weights,
+        combiner=combiner,
+        partition_strategy=partition_strategy,
+        name=None if default_id is None else scope,
+        max_norm=max_norm)
+
+    if default_id is None:
+      # Broadcast is_row_empty to the same shape as embedding_lookup_result,
+      # for use in Select.
+      is_row_empty = array_ops.tile(
+          array_ops.reshape(is_row_empty, [-1, 1]),
+          array_ops.stack([1, array_ops.shape(result)[1]]))
+
+      result = array_ops.where(is_row_empty,
+                               array_ops.zeros_like(result),
+                               result,
+                               name=scope)
+
+    # Reshape back from linear ids back into higher-dimensional dense result.
+    final_result = array_ops.reshape(
+        result,
+        array_ops.concat([
+            array_ops.slice(
+                math_ops.cast(original_shape, dtypes.int32), [0],
+                [original_rank - 1]),
+            array_ops.slice(array_ops.shape(result), [1], [-1])
+        ], 0))
+    final_result.set_shape(tensor_shape.unknown_shape(
+        (original_rank_dim - 1).value).concatenate(result.get_shape()[1:]))
+    return final_result
+
+
+def _prune_invalid_ids(sparse_ids, sparse_weights):
+  """Prune invalid IDs (< 0) from the input ids and weights."""
+  is_id_valid = math_ops.greater_equal(sparse_ids.values, 0)
+  if sparse_weights is not None:
+    is_id_valid = math_ops.logical_and(
+        is_id_valid,
+        array_ops.ones_like(sparse_weights.values, dtype=dtypes.bool))
+  sparse_ids = sparse_ops.sparse_retain(sparse_ids, is_id_valid)
+  if sparse_weights is not None:
+    sparse_weights = sparse_ops.sparse_retain(sparse_weights, is_id_valid)
+  return sparse_ids, sparse_weights
+
+
+def _prune_invalid_weights(sparse_ids, sparse_weights):
+  """Prune invalid weights (< 0) from the input ids and weights."""
+  if sparse_weights is not None:
+    is_weights_valid = math_ops.greater(sparse_weights.values, 0)
+    sparse_ids = sparse_ops.sparse_retain(sparse_ids, is_weights_valid)
+    sparse_weights = sparse_ops.sparse_retain(sparse_weights, is_weights_valid)
+  return sparse_ids, sparse_weights
diff --git a/tensorflow/python/ops/functional_ops.py b/tensorflow/python/ops/functional_ops.py
index 30413f289a0674db61406153cc05d12c7cc98f9b..4ecc74675ae673bcc30f18dde75a396ff673bfaa 100644
--- a/tensorflow/python/ops/functional_ops.py
+++ b/tensorflow/python/ops/functional_ops.py
@@ -775,7 +775,7 @@ def While(input_, cond, body, name=None, hostmem=None):
       a string, non-empty means True and empty means False. If the
       tensor is not a scalar, non-emptiness means True and False
       otherwise.
-    body: . A funcion takes a list of tensors and returns another
+    body: . A function takes a list of tensors and returns another
       list tensors. Both lists have the same types as specified
       by T.
     name: A name for the operation (optional).
@@ -945,6 +945,61 @@ def For(start,
 # pylint: enable=invalid-name,protected-access
 
 
-def partitioned_call(args, f):
-  return gen_functional_ops.partitioned_call(
-      args=args, Tout=[o.type for o in f.definition.signature.output_arg], f=f)
+def partitioned_call(args, f, tout=None, executing_eagerly=None):
+  """Executes a function while respecting device annotations.
+
+  Currently, only those functions that execute within the same address space
+  can be executed.
+
+  Args:
+    args: The arguments of the function, including captured inputs.
+    f: The function to execute; an instance of `_DefinedFunction` or
+      `_EagerDefinedFunction`.
+    tout: a list containing the output dtypes enums; if `None`, inferred from
+      the signature of `f`.
+    executing_eagerly: (Optional) A boolean indicating whether the context is
+      executing eagerly. If `None`, fetched from the global context.
+
+  Returns:
+    The list of `Tensor`s returned by invoking `f(args)`. If the function does
+    not return anything, then returns `None` if eager execution is enabled, or
+    the `Operation` if not.
+  """
+
+  if tout is None:
+    tout = tuple(x.type for x in f.definition.signature.output_arg)
+
+  if executing_eagerly is None:
+    executing_eagerly = context.executing_eagerly()
+
+  if executing_eagerly or len(tout):
+    if f.stateful_ops:
+      outputs = gen_functional_ops.stateful_partitioned_call(
+          args=args, Tout=tout, f=f)
+    else:
+      outputs = gen_functional_ops.partitioned_call(args=args, Tout=tout, f=f)
+    return outputs if outputs else None
+
+  # The generated binding returns an empty list for functions that don't
+  # return any Tensors, hence the need to use `create_op` directly.
+  args = [ops.internal_convert_to_tensor(x) for x in args]
+  tin_attr = attr_value_pb2.AttrValue(
+      list=attr_value_pb2.AttrValue.ListValue(
+          type=[x.dtype.as_datatype_enum for x in args]))
+  tout_attr = attr_value_pb2.AttrValue(
+      list=attr_value_pb2.AttrValue.ListValue(type=tout))
+  func_attr = attr_value_pb2.AttrValue(
+      func=attr_value_pb2.NameAttrList(name=f.name))
+
+  graph = ops.get_default_graph()
+  f.add_to_graph(graph)
+  op_name = "StatefulPartitionedCall" if f.stateful_ops else "PartitionedCall"
+  op = graph.create_op(
+      op_name,
+      args,
+      tout,
+      compute_shapes=False,
+      name="PartitionedFunctionCall",
+      attrs={"Tin": tin_attr, "Tout": tout_attr, "f": func_attr})
+  outputs = op.outputs
+  return outputs if outputs else op
diff --git a/tensorflow/python/ops/gradients_impl.py b/tensorflow/python/ops/gradients_impl.py
index 99909ac38ee631780bffba2689297b79bb7feb94..b64a66be03ba09e0660b7067420b61f91cf191a3 100644
--- a/tensorflow/python/ops/gradients_impl.py
+++ b/tensorflow/python/ops/gradients_impl.py
@@ -31,6 +31,7 @@ from tensorflow.core.framework import attr_value_pb2
 from tensorflow.python.eager import context
 from tensorflow.python.framework import constant_op
 from tensorflow.python.framework import dtypes
+from tensorflow.python.framework import function
 from tensorflow.python.framework import ops
 from tensorflow.python.framework import tensor_shape
 from tensorflow.python.framework import tensor_util
@@ -54,6 +55,7 @@ from tensorflow.python.ops import resource_variable_ops
 from tensorflow.python.ops import spectral_grad  # pylint: disable=unused-import
 from tensorflow.python.ops import tensor_array_ops
 from tensorflow.python.platform import tf_logging as logging
+from tensorflow.python.util import compat
 from tensorflow.python.util.tf_export import tf_export
 
 # This is to avoid a circular dependency with cond_v2_impl.
@@ -113,12 +115,14 @@ ops.register_tensor_conversion_function(ops.IndexedSlices,
                                         _IndexedSlicesToTensor)
 
 
-def _MarkReachedOps(from_ops, reached_ops):
+def _MarkReachedOps(from_ops, reached_ops, func_graphs):
   """Mark all ops reached from "from_ops".
 
   Args:
     from_ops: list of Operations.
     reached_ops: set of Operations.
+    func_graphs: list of function._FuncGraphs. This method will traverse through
+      these functions if they capture from_ops or any reachable ops.
   """
   queue = collections.deque()
   queue.extend(from_ops)
@@ -128,36 +132,11 @@ def _MarkReachedOps(from_ops, reached_ops):
       reached_ops.add(op)
       for output in op.outputs:
         if _IsBackpropagatable(output):
-          queue.extend(output.consumers())
+          queue.extend(_Consumers(output, func_graphs))
 
 
-def _GatherInputs(to_ops, reached_ops):
-  """List all inputs of to_ops that are in reached_ops.
-
-  Args:
-    to_ops: list of Operations.
-    reached_ops: set of Operations.
-
-  Returns:
-    The list of all inputs of to_ops that are in reached_ops.
-    That list includes all elements of to_ops.
-  """
-  inputs = []
-  queue = collections.deque()
-  queue.extend(to_ops)
-  while queue:
-    op = queue.popleft()
-    # We are interested in this op.
-    if op in reached_ops:
-      inputs.append(op)
-      # Clear the boolean so we won't add the inputs again.
-      reached_ops.remove(op)
-      for inp in op.inputs:
-        queue.append(inp.op)
-  return inputs
-
-
-def _PendingCount(to_ops, from_ops, colocate_gradients_with_ops):
+def _PendingCount(to_ops, from_ops, colocate_gradients_with_ops, func_graphs,
+                  xs):
   """Initialize the pending count for ops between two lists of Operations.
 
   'pending_count[op]' indicates the number of backprop inputs
@@ -167,6 +146,11 @@ def _PendingCount(to_ops, from_ops, colocate_gradients_with_ops):
     to_ops: list of Operations.
     from_ops: list of Operations.
     colocate_gradients_with_ops: Python bool.  See docstring of gradients().
+    func_graphs: list of function._FuncGraphs. This method will traverse through
+      these functions if they capture from_ops or any reachable ops. This is
+      useful if to_ops occur in a function and from_ops are in an outer function
+      or graph.
+    xs: list of Tensors.
 
   Returns:
     A tuple containing: (1) the subset of to_ops reachable from from_ops by a
@@ -177,7 +161,7 @@ def _PendingCount(to_ops, from_ops, colocate_gradients_with_ops):
   """
   # Mark reachable ops from from_ops.
   reached_ops = set()
-  _MarkReachedOps(from_ops, reached_ops)
+  _MarkReachedOps(from_ops, reached_ops, func_graphs)
   # X in reached_ops iff X is reachable from from_ops by a path of zero or more
   # backpropagatable tensors.
 
@@ -196,7 +180,7 @@ def _PendingCount(to_ops, from_ops, colocate_gradients_with_ops):
       between_op_list.append(op)
       # Clear the boolean so we won't add the inputs again.
       reached_ops.remove(op)
-      for inp in op.inputs:
+      for inp in _Inputs(op, xs):
         queue.append(inp.op)
   # X in between_ops iff X is on a path of zero or more backpropagatable tensors
   # between from_ops and to_ops
@@ -208,7 +192,7 @@ def _PendingCount(to_ops, from_ops, colocate_gradients_with_ops):
   # Initialize pending count for between ops.
   pending_count = collections.defaultdict(int)
   for op in between_op_list:
-    for x in op.inputs:
+    for x in _Inputs(op, xs):
       if x.op in between_ops:
         pending_count[x.op] += 1
 
@@ -329,7 +313,7 @@ def _VerifyGeneratedGradients(grads, op):
                      "inputs %d" % (len(grads), op.node_def, len(op.inputs)))
 
 
-def _StopOps(from_ops, stop_gradient_ops, pending_count):
+def _StopOps(from_ops, stop_gradient_ops, pending_count, xs):
   """The set of ops that terminate the gradient computation.
 
   This computes the frontier of the forward graph *before* which backprop
@@ -345,6 +329,7 @@ def _StopOps(from_ops, stop_gradient_ops, pending_count):
     from_ops: list of Operations.
     stop_gradient_ops: list of Operations never to backprop through.
     pending_count: mapping from operation to number of backprop inputs.
+    xs: list of Tensors.
 
   Returns:
     The set of operations.
@@ -352,7 +337,7 @@ def _StopOps(from_ops, stop_gradient_ops, pending_count):
   stop_ops = set()
   for op in from_ops:
     is_stop_op = True
-    for inp in op.inputs:
+    for inp in _Inputs(op, xs):
       if pending_count[inp.op] > 0:
         is_stop_op = False
         break
@@ -372,12 +357,19 @@ def _maybe_colocate_with(op, gradient_uid, colocate_gradients_with_ops):  # pyli
     yield
 
 
-def _SymGrad(op, out_grads):
+def _IsPartitionedCall(op):
+  return op.type == "PartitionedCall" or op.type == "StatefulPartitionedCall"
+
+
+def _SymGrad(op, out_grads, xs):
   """Backprop through a function call node op given its outputs' gradients."""
-  f_in = [x for x in op.inputs] + out_grads
-  f_types = [x.dtype for x in op.inputs]
+  f_in = [x for x in _Inputs(op, xs)] + out_grads
+  f_types = [x.dtype for x in _Inputs(op, xs)]
   f = attr_value_pb2.NameAttrList()
-  f.name = op.type
+  if _IsPartitionedCall(op):
+    f.name = op.get_attr("f").name
+  else:
+    f.name = op.type
   for k in op.node_def.attr:
     f.attr[k].CopyFrom(op.node_def.attr[k])
   # TODO(apassos) use a better dtype here
@@ -425,7 +417,7 @@ def _MaybeCompile(scope, op, func, grad_fn):
     return grad_fn()
 
 
-def _RaiseNoGradWrtInitialLoopValError(op, from_ops):
+def _RaiseNoGradWrtInitialLoopValError(op, from_ops, xs):
   """Raises an error if we backprop through a loop var."""
   # Find the nearest 'to_op' reachable from 'op' to provide a more helpful error
   # message.
@@ -439,7 +431,7 @@ def _RaiseNoGradWrtInitialLoopValError(op, from_ops):
     if curr_op in from_ops:
       target_op = curr_op
       break
-    queue.extend(t.op for t in curr_op.inputs)
+    queue.extend(t.op for t in _Inputs(curr_op, xs))
   assert target_op
   raise ValueError(
       "Cannot compute gradient inside while loop with respect to op '%s'. "
@@ -449,6 +441,68 @@ def _RaiseNoGradWrtInitialLoopValError(op, from_ops):
       % target_op.name)
 
 
+def _MaybeCaptured(t):
+  """If t is a captured value placeholder, returns the original captured value.
+
+  Args:
+    t: Tensor
+
+  Returns:
+    A tensor, potentially from a different Graph/function._FuncGraph.
+  """
+  # pylint: disable=protected-access
+  if isinstance(t.op.graph, function._FuncGraph) and t.op.type == "Placeholder":
+    for input_t, placeholder_t in t.op.graph._captured.items():
+      if t == placeholder_t:
+        return _MaybeCaptured(input_t)
+  # pylint: enable=protected-access
+  return t
+
+
+# TODO(skyewm): plumbing xs through everywhere is ugly, consider making
+# _GradientsHelper a class with xs as a member variable.
+def _Inputs(op, xs):
+  """Returns the inputs of op, crossing closure boundaries where necessary.
+
+  Args:
+    op: Operation
+    xs: list of Tensors we are differentiating w.r.t.
+
+  Returns:
+    A list of tensors. The tensors may be from multiple
+    Graph/function._FuncGraphs if op is in a function._FuncGraph and has
+    captured inputs.
+  """
+  if isinstance(op.graph, function._FuncGraph):  # pylint: disable=protected-access
+    # If we're differentiating w.r.t. `t`, do not attempt to traverse through it
+    # to a captured value. The algorithm needs to "see" `t` in this case, even
+    # if it's a function input for a captured value, whereas usually we'd like
+    # to traverse through these closures as if the captured value was the direct
+    # input to op.
+    return [t if (t in xs) else _MaybeCaptured(t) for t in op.inputs]
+  else:
+    return op.inputs
+
+
+def _Consumers(t, func_graphs):
+  """Returns the consumers of t, crossing closure boundaries where necessary.
+
+  Args:
+    t: Tensor
+    func_graphs: a list of function._FuncGraphs that may have captured t.
+
+  Returns:
+    A list of tensors. The tensors will be from the current graph and/or
+    func_graphs.
+  """
+  consumers = t.consumers()
+  for func in func_graphs:
+    for input_t, placeholder in func._captured.items():  # pylint: disable=protected-access
+      if input_t == t:
+        consumers.extend(_Consumers(placeholder, func_graphs))
+  return consumers
+
+
 @tf_export("gradients")
 def gradients(ys,
               xs,
@@ -534,10 +588,10 @@ def gradients(ys,
     RuntimeError: if called in Eager mode.
 
   """
-  # Creating the gradient graph for control flow mutates Operations. _lock
-  # ensures a Session.run call cannot occur between creating and mutating new
-  # ops.
-  with ops.get_default_graph()._lock:  # pylint: disable=protected-access
+  # Creating the gradient graph for control flow mutates Operations.
+  # _mutation_lock ensures a Session.run call cannot occur between creating and
+  # mutating new ops.
+  with ops.get_default_graph()._mutation_lock():  # pylint: disable=protected-access
     return _GradientsHelper(ys, xs, grad_ys, name, colocate_gradients_with_ops,
                             gate_gradients, aggregation_method, stop_gradients)
 
@@ -558,6 +612,14 @@ def _GradientsHelper(ys,
   if src_graph is None:
     src_graph = ops.get_default_graph()
 
+  # If src_graph is a _FuncGraph (i.e. a function body), gather it and all
+  # ancestor graphs. This is necessary for correctly handling captured values.
+  func_graphs = []
+  curr_graph = src_graph
+  while isinstance(curr_graph, function._FuncGraph):  # pylint: disable=protected-access
+    func_graphs.append(curr_graph)
+    curr_graph = curr_graph._outer_graph  # pylint: disable=protected-access
+
   ys = _AsList(ys)
   xs = _AsList(xs)
   stop_gradients = [] if stop_gradients is None else _AsList(stop_gradients)
@@ -592,12 +654,13 @@ def _GradientsHelper(ys,
     # Initialize the pending count for ops in the connected subgraph from ys
     # to the xs.
     if len(ys) > 1:
-      ys = [array_ops.identity(y) if y.consumers() else y for y in ys]
+      ys = [array_ops.identity(y) if _Consumers(y, func_graphs) else y
+            for y in ys]
     to_ops = [t.op for t in ys]
     from_ops = [t.op for t in xs]
     stop_gradient_ops = [t.op for t in stop_gradients]
     reachable_to_ops, pending_count, loop_state = _PendingCount(
-        to_ops, from_ops, colocate_gradients_with_ops)
+        to_ops, from_ops, colocate_gradients_with_ops, func_graphs, xs)
 
     # Iterate over the collected ops.
     #
@@ -631,7 +694,7 @@ def _GradientsHelper(ys,
           _SetGrad(grads, y, loop_state.ZerosLikeForExit(y))
           queue.append(y.op)
 
-    stop_ops = _StopOps(from_ops, stop_gradient_ops, pending_count)
+    stop_ops = _StopOps(from_ops, stop_gradient_ops, pending_count, xs)
     while queue:
       # generate gradient subgraph for op.
       op = queue.popleft()
@@ -645,13 +708,19 @@ def _GradientsHelper(ys,
 
         grad_fn = None
         func_call = None
+        is_partitioned_call = _IsPartitionedCall(op)
         # pylint: disable=protected-access
-        is_func_call = src_graph._is_function(op.type)
+        is_func_call = (
+            src_graph._is_function(op.type) or is_partitioned_call)
         # pylint: enable=protected-access
         has_out_grads = any(isinstance(g, ops.Tensor) or g for g in out_grads)
         if has_out_grads and (op not in stop_ops):
           if is_func_call:
-            func_call = src_graph._get_function(op.type)  # pylint: disable=protected-access
+            if is_partitioned_call:
+              func_call = src_graph._get_function(  # pylint: disable=protected-access
+                  compat.as_bytes(op.get_attr("f").name))
+            else:
+              func_call = src_graph._get_function(op.type)  # pylint: disable=protected-access
             # Note that __defun is not set if the graph is
             # imported. If it's set, we prefer to access the original
             # defun.
@@ -680,7 +749,7 @@ def _GradientsHelper(ys,
             op._control_flow_context.IsWhileContext() and
             op._control_flow_context ==
             ops.get_default_graph()._get_control_flow_context()):
-          _RaiseNoGradWrtInitialLoopValError(op, from_ops)
+          _RaiseNoGradWrtInitialLoopValError(op, from_ops, xs)
         # pylint: enable=protected-access
 
         if (grad_fn or is_func_call) and has_out_grads:
@@ -712,7 +781,7 @@ def _GradientsHelper(ys,
                 # For function call ops, we add a 'SymbolicGradient'
                 # node to the graph to compute gradients.
                 in_grads = _MaybeCompile(grad_scope, op, func_call,
-                                         lambda: _SymGrad(op, out_grads))
+                                         lambda: _SymGrad(op, out_grads, xs))
               in_grads = _AsList(in_grads)
               _VerifyGeneratedGradients(in_grads, op)
               if gate_gradients and len([x for x in in_grads
@@ -727,8 +796,8 @@ def _GradientsHelper(ys,
         else:
           # If no grad_fn is defined or none of out_grads is available,
           # just propagate a list of None backwards.
-          in_grads = [None] * len(op.inputs)
-        for i, (t_in, in_grad) in enumerate(zip(op.inputs, in_grads)):
+          in_grads = [None] * len(_Inputs(op, xs))
+        for i, (t_in, in_grad) in enumerate(zip(_Inputs(op, xs), in_grads)):
           if in_grad is not None:
             if (isinstance(in_grad, ops.Tensor) and
                 t_in.dtype != dtypes.resource):
@@ -746,7 +815,8 @@ def _GradientsHelper(ys,
           loop_state.ExitGradWhileContext(op, before=False)
 
       # Update pending count for the inputs of op and enqueue ready ops.
-      _UpdatePendingAndEnqueueReady(grads, op, queue, pending_count, loop_state)
+      _UpdatePendingAndEnqueueReady(grads, op, queue, pending_count, loop_state,
+                                    xs)
 
   if loop_state:
     loop_state.PostProcessing()
@@ -765,9 +835,10 @@ def _HasAnyNotNoneGrads(grads, op):
   return False
 
 
-def _UpdatePendingAndEnqueueReady(grads, op, queue, pending_count, loop_state):
+def _UpdatePendingAndEnqueueReady(grads, op, queue, pending_count, loop_state,
+                                  xs):
   """Update pending count for the inputs of op and enqueue ready ops."""
-  for x in op.inputs:
+  for x in _Inputs(op, xs):
     pending_count[x.op] -= 1
     ready = (pending_count[x.op] == 0)
     if loop_state and not ready:
diff --git a/tensorflow/python/ops/gradients_test.py b/tensorflow/python/ops/gradients_test.py
index d81c756f1cbc0a46d094066cda369067f7d3d1f6..d02fcf4ee27c180003e5b026e486a4ec0ad11e7d 100644
--- a/tensorflow/python/ops/gradients_test.py
+++ b/tensorflow/python/ops/gradients_test.py
@@ -57,90 +57,8 @@ from tensorflow.python.ops.nn_ops import bias_add
 from tensorflow.python.platform import googletest
 
 
-def _OpsBetween(to_ops, from_ops):
-  """Build the list of operations between two lists of Operations.
-
-  Args:
-    to_ops: list of Operations.
-    from_ops: list of Operations.
-
-  Returns:
-    The list of operations between "from_ops" and "to_ops", sorted by
-    decreasing operation id. This list contains all elements of to_ops.
-
-    TODO(touts): Think about returning an empty list if from_ops are not
-    reachable from to_ops.  Presently it returns to_ops in that case.
-  """
-  # Ops that are reachable from the output of "input_ops".
-  reached_ops = set()
-  # We only care to reach up to "output_ops" so we mark the
-  # output ops as reached to avoid recursing past them.
-  for op in to_ops:
-    reached_ops.add(op)
-  gradients_impl._MarkReachedOps(from_ops, reached_ops)
-  between_ops = gradients_impl._GatherInputs(to_ops, reached_ops)
-  between_ops.sort(key=lambda x: -x._id)
-  return between_ops
-
-
 class GradientsTest(test_util.TensorFlowTestCase):
 
-  def _OpNames(self, op_list):
-    return ["%s/%d" % (str(op.name), op._id) for op in op_list]
-
-  def _assertOpListEqual(self, ops1, ops2):
-    self.assertEquals(self._OpNames(ops1), self._OpNames(ops2))
-
-  def testOpsBetweenSimple(self):
-    with ops.Graph().as_default():
-      t1 = constant(1.0)
-      t2 = constant(2.0)
-      t3 = array_ops.stack([t1, t2])
-    # Full graph
-    self._assertOpListEqual([t3.op, t2.op, t1.op],
-                            _OpsBetween([t3.op], [t1.op, t2.op]))
-    # Only t1, t3.
-    self._assertOpListEqual([t3.op, t1.op], _OpsBetween([t3.op], [t1.op]))
-
-  def testOpsBetweenUnreachable(self):
-    with ops.Graph().as_default():
-      t1 = constant(1.0)
-      t2 = constant(2.0)
-      _ = array_ops.stack([t1, t2])
-      t4 = constant(1.0)
-      t5 = constant(2.0)
-      t6 = array_ops.stack([t4, t5])
-    # Elements of to_ops are always listed.
-    self._assertOpListEqual([t6.op], _OpsBetween([t6.op], [t1.op]))
-
-  def testOpsBetweenCut(self):
-    with ops.Graph().as_default():
-      t1 = constant(1.0)
-      t2 = constant(2.0)
-      t3 = array_ops.stack([t1, t2])
-      t4 = constant([1.0])
-      t5 = array_ops.concat([t4, t3], 0)
-      t6 = constant([2.0])
-      t7 = array_ops.concat([t5, t6], 0)
-    self._assertOpListEqual([t7.op, t5.op, t4.op],
-                            _OpsBetween([t7.op], [t4.op]))
-
-  def testOpsBetweenCycle(self):
-    with ops.Graph().as_default():
-      t1 = constant(1.0)
-      t2 = constant(2.0)
-      t3 = array_ops.stack([t1, t2])
-      t4 = array_ops.concat([t3, t3, t3], 0)
-      t5 = constant([1.0])
-      t6 = array_ops.concat([t4, t5], 0)
-      t7 = array_ops.concat([t6, t3], 0)
-    self._assertOpListEqual([t6.op, t4.op, t3.op],
-                            _OpsBetween([t6.op], [t3.op]))
-    self._assertOpListEqual([t7.op, t6.op, t5.op, t4.op, t3.op, t1.op],
-                            _OpsBetween([t7.op], [t1.op, t5.op]))
-    self._assertOpListEqual([t6.op, t5.op, t4.op, t3.op, t2.op],
-                            _OpsBetween([t6.op], [t2.op, t5.op]))
-
   def testGradients(self):
     with ops.Graph().as_default():
       inp = constant(1.0, shape=[32, 100], name="in")
@@ -519,6 +437,96 @@ class FunctionGradientsTest(test_util.TensorFlowTestCase):
             grad_func=grad_func, python_grad_func=self._PythonGradient)
         f.add_to_graph(ops.Graph())
 
+  def testGradientWrtCaptured(self):
+    with ops.Graph().as_default():
+      x = constant_op.constant(1.0, name="x")
+
+      @function.Defun()
+      def Foo():
+        y = math_ops.multiply(x, 2.0, name="y")
+        g = gradients_impl.gradients(y, x)
+        return g[0]
+
+      f = Foo()
+      with self.test_session() as sess:
+        self.assertEqual(sess.run(f), 2.0)
+
+  def testGradientOfCaptured(self):
+    with ops.Graph().as_default():
+      x = constant_op.constant(1.0, name="x")
+      y = math_ops.multiply(x, 2.0, name="y")
+
+      @function.Defun()
+      def Foo():
+        g = gradients_impl.gradients(y, x)
+        return g[0]
+
+      f = Foo()
+      with self.test_session() as sess:
+        self.assertEqual(sess.run(f), 2.0)
+
+  def testCapturedResourceVariable(self):
+    with ops.Graph().as_default():
+      var = resource_variable_ops.ResourceVariable(1.0, name="var")
+
+      @function.Defun()
+      def Foo():
+        y = math_ops.multiply(var, 2.0, name="y")
+        g = gradients_impl.gradients(y, var)
+        return g[0]
+
+      f = Foo()
+      with self.test_session() as sess:
+        sess.run(variables.global_variables_initializer())
+        self.assertEqual(sess.run(f), 2.0)
+
+  def testCapturedNested(self):
+    with ops.Graph().as_default():
+      x1 = constant_op.constant(1.0, name="x1")
+      x2 = constant_op.constant(2.0, name="x2")
+      x3 = math_ops.multiply(x1, x2, name="x3")
+
+      @function.Defun()
+      def Outer():
+        outer1 = array_ops.identity(x1, name="outer1")
+
+        @function.Defun()
+        def Inner():
+          inner1 = array_ops.identity(outer1, name="inner1")
+          inner2 = array_ops.identity(x2, name="inner2")
+          inner3 = array_ops.identity(x3, name="inner3")
+          return gradients_impl.gradients([inner1, inner2, inner3, x1],
+                                          [x1, x2])
+
+        return Inner()
+
+      x1_grad, x2_grad = Outer()
+      with self.test_session() as sess:
+        # 1.0 + None + 2.0 + 1.0 = 4.0
+        self.assertEqual(sess.run(x1_grad), 4.0)
+        # None + 1.0 + 1.0 + None = 2.0
+        self.assertEqual(sess.run(x2_grad), 2.0)
+
+  def testCapturedFromFunction(self):
+    with ops.Graph().as_default():
+      x = constant_op.constant(1.0, name="x")
+
+      @function.Defun()
+      def Outer():
+        y = math_ops.multiply(x, 2.0, name="y")
+
+        @function.Defun()
+        def Inner():
+          z = math_ops.multiply(y, 3.0, name="z")
+          g = gradients_impl.gradients(z, y)
+          return g[0]
+
+        return Inner()
+
+      z_grad = Outer()
+      with self.test_session() as sess:
+        self.assertEqual(sess.run(z_grad), 3.0)
+
 
 class StopGradientTest(test_util.TensorFlowTestCase):
 
diff --git a/tensorflow/python/ops/image_ops_impl.py b/tensorflow/python/ops/image_ops_impl.py
index 2c7751f7923dca4d0c4f907a673b06ba86b9f342..9440bab9ee5ee15037d9b879faea265fee608cba 100644
--- a/tensorflow/python/ops/image_ops_impl.py
+++ b/tensorflow/python/ops/image_ops_impl.py
@@ -55,8 +55,10 @@ ops.NotDifferentiable('SampleDistortedBoundingBoxV2')
 ops.NotDifferentiable('ExtractGlimpse')
 ops.NotDifferentiable('NonMaxSuppression')
 ops.NotDifferentiable('NonMaxSuppressionV2')
+ops.NotDifferentiable('NonMaxSuppressionWithOverlaps')
 
 
+# pylint: disable=invalid-name
 def _assert(cond, ex_type, msg):
   """A polymorphic assert, works with tensors and boolean expressions.
 
@@ -1070,15 +1072,16 @@ def resize_images(images,
 
 
 @tf_export('image.resize_image_with_pad')
-def resize_image_with_pad(image, target_height, target_width,
+def resize_image_with_pad(image,
+                          target_height,
+                          target_width,
                           method=ResizeMethod.BILINEAR):
-  """
-  Resizes and pads an image to a target width and height.
+  """Resizes and pads an image to a target width and height.
 
   Resizes an image to a target width and height by keeping
   the aspect ratio the same without distortion. If the target
   dimensions don't match the image dimensions, the image
-  is resized and then padded with zeroes to match requested 
+  is resized and then padded with zeroes to match requested
   dimensions.
 
   Args:
@@ -1139,10 +1142,10 @@ def resize_image_with_pad(image, target_height, target_width,
     ratio = max_(f_width / f_target_width, f_height / f_target_height)
     resized_height_float = f_height / ratio
     resized_width_float = f_width / ratio
-    resized_height = math_ops.cast(math_ops.floor(resized_height_float),
-                                   dtype=dtypes.int32)
-    resized_width = math_ops.cast(math_ops.floor(resized_width_float),
-                                  dtype=dtypes.int32)
+    resized_height = math_ops.cast(
+        math_ops.floor(resized_height_float), dtype=dtypes.int32)
+    resized_width = math_ops.cast(
+        math_ops.floor(resized_width_float), dtype=dtypes.int32)
 
     padding_height = (f_target_height - resized_height_float) / 2
     padding_width = (f_target_width - resized_width_float) / 2
@@ -1154,13 +1157,13 @@ def resize_image_with_pad(image, target_height, target_width,
     # Resize first, then pad to meet requested dimensions
     resized = resize_images(image, [resized_height, resized_width], method)
 
-    padded = pad_to_bounding_box(resized, p_height, p_width,
-                                 target_height, target_width)
+    padded = pad_to_bounding_box(resized, p_height, p_width, target_height,
+                                 target_width)
 
     if padded.get_shape().ndims is None:
       raise ValueError('padded contains no shape.')
 
-    _, padded_height, padded_width, _ = _ImageDimensions(padded, rank=4)
+    _ImageDimensions(padded, rank=4)
 
     if not is_batch:
       padded = array_ops.squeeze(padded, squeeze_dims=[0])
@@ -1750,6 +1753,22 @@ def is_jpeg(contents, name=None):
     return math_ops.equal(substr, b'\xff\xd8\xff', name=name)
 
 
+def _is_png(contents, name=None):
+  r"""Convenience function to check if the 'contents' encodes a PNG image.
+
+  Args:
+    contents: 0-D `string`. The encoded image bytes.
+    name: A name for the operation (optional)
+
+  Returns:
+     A scalar boolean tensor indicating if 'contents' may be a PNG image.
+     is_png is susceptible to false positives.
+  """
+  with ops.name_scope(name, 'is_png'):
+    substr = string_ops.substr(contents, 0, 3)
+    return math_ops.equal(substr, b'\211PN', name=name)
+
+
 @tf_export('image.decode_image')
 def decode_image(contents, channels=None, dtype=dtypes.uint8, name=None):
   """Convenience function for `decode_bmp`, `decode_gif`, `decode_jpeg`,
@@ -1827,8 +1846,8 @@ def decode_image(contents, channels=None, dtype=dtypes.uint8, name=None):
 
     def check_png():
       """Checks if an image is PNG."""
-      is_png = math_ops.equal(substr, b'\211PN', name='is_png')
-      return control_flow_ops.cond(is_png, _png, check_gif, name='cond_png')
+      return control_flow_ops.cond(
+          _is_png(contents), _png, check_gif, name='cond_png')
 
     def _jpeg():
       """Decodes a jpeg image."""
@@ -2091,6 +2110,50 @@ def non_max_suppression(boxes,
                                                 iou_threshold, score_threshold)
 
 
+@tf_export('image.non_max_suppression_overlaps')
+def non_max_suppression_with_overlaps(overlaps,
+                                      scores,
+                                      max_output_size,
+                                      overlap_threshold=0.5,
+                                      score_threshold=float('-inf'),
+                                      name=None):
+  """Greedily selects a subset of bounding boxes in descending order of score.
+
+  Prunes away boxes that have high overlap with previously selected boxes.
+  N-by-n overlap values are supplied as square matrix.
+  The output of this operation is a set of integers indexing into the input
+  collection of bounding boxes representing the selected boxes.  The bounding
+  box coordinates corresponding to the selected indices can then be obtained
+  using the `tf.gather operation`.  For example:
+    selected_indices = tf.image.non_max_suppression_overlaps(
+        overlaps, scores, max_output_size, iou_threshold)
+    selected_boxes = tf.gather(boxes, selected_indices)
+
+  Args:
+    overlaps: A 2-D float `Tensor` of shape `[num_boxes, num_boxes]`.
+    scores: A 1-D float `Tensor` of shape `[num_boxes]` representing a single
+      score corresponding to each box (each row of boxes).
+    max_output_size: A scalar integer `Tensor` representing the maximum number
+      of boxes to be selected by non max suppression.
+    overlap_threshold: A float representing the threshold for deciding whether
+      boxes overlap too much with respect to the provided overlap values.
+    score_threshold: A float representing the threshold for deciding when to
+      remove boxes based on score.
+    name: A name for the operation (optional).
+
+  Returns:
+    selected_indices: A 1-D integer `Tensor` of shape `[M]` representing the
+      selected indices from the overlaps tensor, where `M <= max_output_size`.
+  """
+  with ops.name_scope(name, 'non_max_suppression_overlaps'):
+    overlap_threshold = ops.convert_to_tensor(
+        overlap_threshold, name='overlap_threshold')
+    # pylint: disable=protected-access
+    return gen_image_ops._non_max_suppression_v3(
+        overlaps, scores, max_output_size, overlap_threshold, score_threshold)
+    # pylint: enable=protected-access
+
+
 _rgb_to_yiq_kernel = [[0.299, 0.59590059,
                        0.2115], [0.587, -0.27455667, -0.52273617],
                       [0.114, -0.32134392, 0.31119955]]
diff --git a/tensorflow/python/ops/image_ops_test.py b/tensorflow/python/ops/image_ops_test.py
index 8e40de140df632c9b458c2e2b8a673925ab13634..cf9761803bf9654e21ec12e1f1c7193b3e88c020 100644
--- a/tensorflow/python/ops/image_ops_test.py
+++ b/tensorflow/python/ops/image_ops_test.py
@@ -2731,7 +2731,7 @@ class ResizeImageWithPadTest(test_util.TensorFlowTestCase):
       try:
         self._ResizeImageWithPad(x, target_height, target_width,
                                  use_tensor_inputs)
-      except Exception as e:
+      except Exception as e:  # pylint: disable=broad-except
         if err_msg not in str(e):
           raise
       else:
diff --git a/tensorflow/python/ops/init_ops.py b/tensorflow/python/ops/init_ops.py
index c41e952167d4769972426142e3989d6916df6aed..3132f7467f6cbed4d0f53d6ed4f91019a91614e1 100644
--- a/tensorflow/python/ops/init_ops.py
+++ b/tensorflow/python/ops/init_ops.py
@@ -43,7 +43,8 @@ from tensorflow.python.ops import linalg_ops_impl
 from tensorflow.python.ops import gen_linalg_ops
 from tensorflow.python.ops import math_ops
 from tensorflow.python.ops import random_ops
-from tensorflow.python.util.deprecation import deprecated
+from tensorflow.python.util.deprecation import (
+    deprecated, deprecated_arg_values)
 from tensorflow.python.util.tf_export import tf_export
 
 
@@ -409,8 +410,10 @@ class UniformUnitScaling(Initializer):
 class VarianceScaling(Initializer):
   """Initializer capable of adapting its scale to the shape of weights tensors.
 
-  With `distribution="normal"`, samples are drawn from a truncated normal
-  distribution centered on zero, with `stddev = sqrt(scale / n)`
+  With `distribution="truncated_normal" or "untruncated_normal"`,
+  samples are drawn from a truncated/untruncated normal
+  distribution with a mean of zero and a standard deviation (after truncation,
+  if used) `stddev = sqrt(scale / n)`
   where n is:
     - number of input units in the weight tensor, if mode = "fan_in"
     - number of output units, if mode = "fan_out"
@@ -433,10 +436,14 @@ class VarianceScaling(Initializer):
       "distribution" arguments.
   """
 
+  @deprecated_arg_values(
+      None,
+      "`normal` is a deprecated alias for `truncated_normal`",
+      distribution="normal")
   def __init__(self,
                scale=1.0,
                mode="fan_in",
-               distribution="normal",
+               distribution="truncated_normal",
                seed=None,
                dtype=dtypes.float32):
     if scale <= 0.:
@@ -444,7 +451,8 @@ class VarianceScaling(Initializer):
     if mode not in {"fan_in", "fan_out", "fan_avg"}:
       raise ValueError("Invalid `mode` argument:", mode)
     distribution = distribution.lower()
-    if distribution not in {"normal", "uniform"}:
+    if distribution not in {"normal", "uniform",
+                            "truncated_normal", "untruncated_normal"}:
       raise ValueError("Invalid `distribution` argument:", distribution)
     self.scale = scale
     self.mode = mode
@@ -466,11 +474,15 @@ class VarianceScaling(Initializer):
       scale /= max(1., fan_out)
     else:
       scale /= max(1., (fan_in + fan_out) / 2.)
-    if self.distribution == "normal":
+    if self.distribution == "normal" or self.distribution == "truncated_normal":
       # constant taken from scipy.stats.truncnorm.std(a=-2, b=2, loc=0., scale=1.)
       stddev = math.sqrt(scale) / .87962566103423978
       return random_ops.truncated_normal(
           shape, 0.0, stddev, dtype, seed=self.seed)
+    elif self.distribution == "untruncated_normal":
+      stddev = math.sqrt(scale)
+      return random_ops.random_normal(
+          shape, 0.0, stddev, dtype, seed=self.seed)
     else:
       limit = math.sqrt(3.0 * scale)
       return random_ops.random_uniform(
@@ -1124,7 +1136,7 @@ convolutional_orthogonal_3d = ConvolutionOrthogonal3D
 # pylint: enable=invalid-name
 
 
-@tf_export("glorot_uniform_initializer")
+@tf_export("glorot_uniform_initializer", "keras.initializers.glorot_uniform")
 def glorot_uniform_initializer(seed=None, dtype=dtypes.float32):
   """The Glorot uniform initializer, also called Xavier uniform initializer.
 
@@ -1148,7 +1160,7 @@ def glorot_uniform_initializer(seed=None, dtype=dtypes.float32):
       scale=1.0, mode="fan_avg", distribution="uniform", seed=seed, dtype=dtype)
 
 
-@tf_export("glorot_normal_initializer")
+@tf_export("glorot_normal_initializer", "keras.initializers.glorot_normal")
 def glorot_normal_initializer(seed=None, dtype=dtypes.float32):
   """The Glorot normal initializer, also called Xavier normal initializer.
 
diff --git a/tensorflow/python/ops/linalg/linear_operator.py b/tensorflow/python/ops/linalg/linear_operator.py
index 8cfe964b1c0a572f43a14c66885e74ea105b0916..20c46fbb82b0671c6cc586eafdd7fa346d8b4e6d 100644
--- a/tensorflow/python/ops/linalg/linear_operator.py
+++ b/tensorflow/python/ops/linalg/linear_operator.py
@@ -42,7 +42,7 @@ __all__ = ["LinearOperator"]
 class LinearOperator(object):
   """Base class defining a [batch of] linear operator[s].
 
-  Subclasses of `LinearOperator` provide a access to common methods on a
+  Subclasses of `LinearOperator` provide access to common methods on a
   (batch) matrix, without the need to materialize the matrix.  This allows:
 
   * Matrix free computations
@@ -69,11 +69,11 @@ class LinearOperator(object):
 
   #### Shape compatibility
 
-  `LinearOperator` sub classes should operate on a [batch] matrix with
+  `LinearOperator` subclasses should operate on a [batch] matrix with
   compatible shape.  Class docstrings should define what is meant by compatible
-  shape.  Some sub-classes may not support batching.
+  shape.  Some subclasses may not support batching.
 
-  An example is:
+  Examples:
 
   `x` is a batch matrix with compatible shape for `matmul` if
 
diff --git a/tensorflow/python/ops/linalg/linear_operator_diag.py b/tensorflow/python/ops/linalg/linear_operator_diag.py
index 5beaea65a5171ad7e92042a2afa81c0507e51d0e..ed53decc00dc90df5c6c97d9fd9d5cb124ddf660 100644
--- a/tensorflow/python/ops/linalg/linear_operator_diag.py
+++ b/tensorflow/python/ops/linalg/linear_operator_diag.py
@@ -231,8 +231,11 @@ class LinearOperatorDiag(linear_operator.LinearOperator):
     return math_ops.reduce_prod(self._diag, reduction_indices=[-1])
 
   def _log_abs_determinant(self):
-    return math_ops.reduce_sum(
+    log_det = math_ops.reduce_sum(
         math_ops.log(math_ops.abs(self._diag)), reduction_indices=[-1])
+    if self.dtype.is_complex:
+      log_det = math_ops.cast(log_det, dtype=self.dtype)
+    return log_det
 
   def _solve(self, rhs, adjoint=False, adjoint_arg=False):
     diag_term = math_ops.conj(self._diag) if adjoint else self._diag
diff --git a/tensorflow/python/ops/linalg/linear_operator_low_rank_update.py b/tensorflow/python/ops/linalg/linear_operator_low_rank_update.py
index 08e5896e1034fb1782beacfb18fef16da083bded..2b2bf80f276a62d20aae717ac9fa08f9769f455e 100644
--- a/tensorflow/python/ops/linalg/linear_operator_low_rank_update.py
+++ b/tensorflow/python/ops/linalg/linear_operator_low_rank_update.py
@@ -18,16 +18,15 @@ from __future__ import absolute_import
 from __future__ import division
 from __future__ import print_function
 
-from tensorflow.python.framework import dtypes
 from tensorflow.python.framework import ops
 from tensorflow.python.ops import array_ops
-from tensorflow.python.ops import check_ops
 from tensorflow.python.ops import linalg_ops
 from tensorflow.python.ops import math_ops
 from tensorflow.python.ops.linalg import linear_operator
 from tensorflow.python.ops.linalg import linear_operator_diag
 from tensorflow.python.ops.linalg import linear_operator_identity
 from tensorflow.python.ops.linalg import linear_operator_util
+from tensorflow.python.platform import tf_logging as logging
 from tensorflow.python.util.tf_export import tf_export
 
 __all__ = [
@@ -153,8 +152,7 @@ class LinearOperatorLowRankUpdate(linear_operator.LinearOperator):
     `is_X` matrix property hints, which will trigger the appropriate code path.
 
     Args:
-      base_operator:  Shape `[B1,...,Bb, M, N]` real `float16`, `float32` or
-        `float64` `LinearOperator`.  This is `L` above.
+      base_operator:  Shape `[B1,...,Bb, M, N]`.
       u:  Shape `[B1,...,Bb, M, K]` `Tensor` of same `dtype` as `base_operator`.
         This is `U` above.
       diag_update:  Optional shape `[B1,...,Bb, K]` `Tensor` with same `dtype`
@@ -183,23 +181,12 @@ class LinearOperatorLowRankUpdate(linear_operator.LinearOperator):
     Raises:
       ValueError:  If `is_X` flags are set in an inconsistent way.
     """
-    # TODO(langmore) support complex types.
-    # Complex types are not allowed due to tf.cholesky() requiring float.
-    # If complex dtypes are allowed, we update the following
-    # 1. is_diag_update_positive should still imply that `diag > 0`, but we need
-    #    to remind the user that this implies diag is real.  This is needed
-    #    because if diag has non-zero imaginary part, it will not be
-    #    self-adjoint positive definite.
     dtype = base_operator.dtype
-    allowed_dtypes = [
-        dtypes.float16,
-        dtypes.float32,
-        dtypes.float64,
-    ]
-    if dtype not in allowed_dtypes:
-      raise TypeError(
-          "Argument matrix must have dtype in %s.  Found: %s"
-          % (allowed_dtypes, dtype))
+
+    if diag_update is not None:
+      if is_diag_update_positive and dtype.is_complex:
+        logging.warn("Note: setting is_diag_update_positive with a complex "
+                     "dtype means that diagonal is real and positive.")
 
     if diag_update is None:
       if is_diag_update_positive is False:
@@ -271,8 +258,6 @@ class LinearOperatorLowRankUpdate(linear_operator.LinearOperator):
       self._set_diag_operators(diag_update, is_diag_update_positive)
       self._is_diag_update_positive = is_diag_update_positive
 
-      check_ops.assert_same_float_dtype((base_operator, self.u, self.v,
-                                         self._diag_update))
       self._check_shapes()
 
       # Pre-compute the so-called "capacitance" matrix
@@ -407,6 +392,8 @@ class LinearOperatorLowRankUpdate(linear_operator.LinearOperator):
     else:
       det_c = linalg_ops.matrix_determinant(self._capacitance)
       log_abs_det_c = math_ops.log(math_ops.abs(det_c))
+      if self.dtype.is_complex:
+        log_abs_det_c = math_ops.cast(log_abs_det_c, dtype=self.dtype)
 
     return log_abs_det_c + log_abs_det_d + log_abs_det_l
 
diff --git a/tensorflow/python/ops/linalg/linear_operator_lower_triangular.py b/tensorflow/python/ops/linalg/linear_operator_lower_triangular.py
index fb1eb2fedba5b47ce38f9635527b91e18d894a8f..ca6d3f54051d7bf0ff748804d3cd314b144c2f88 100644
--- a/tensorflow/python/ops/linalg/linear_operator_lower_triangular.py
+++ b/tensorflow/python/ops/linalg/linear_operator_lower_triangular.py
@@ -119,8 +119,7 @@ class LinearOperatorLowerTriangular(linear_operator.LinearOperator):
     Args:
       tril:  Shape `[B1,...,Bb, N, N]` with `b >= 0`, `N >= 0`.
         The lower triangular part of `tril` defines this operator.  The strictly
-        upper triangle is ignored.  Allowed dtypes: `float16`, `float32`,
-        `float64`.
+        upper triangle is ignored.
       is_non_singular:  Expect that this operator is non-singular.
         This operator is non-singular if and only if its diagonal elements are
         all non-zero.
@@ -137,7 +136,6 @@ class LinearOperatorLowerTriangular(linear_operator.LinearOperator):
       name: A name for this `LinearOperator`.
 
     Raises:
-      TypeError:  If `diag.dtype` is not an allowed type.
       ValueError:  If `is_square` is `False`.
     """
 
@@ -163,12 +161,12 @@ class LinearOperatorLowerTriangular(linear_operator.LinearOperator):
 
   def _check_tril(self, tril):
     """Static check of the `tril` argument."""
-    # TODO(langmore) Add complex types once matrix_triangular_solve works for
-    # them.
     allowed_dtypes = [
         dtypes.float16,
         dtypes.float32,
         dtypes.float64,
+        dtypes.complex64,
+        dtypes.complex128,
     ]
     dtype = tril.dtype
     if dtype not in allowed_dtypes:
diff --git a/tensorflow/python/ops/linalg/linear_operator_test_util.py b/tensorflow/python/ops/linalg/linear_operator_test_util.py
index 1b5bb9470c4406ad075f2f6d5c38661311472727..78c85db557047ebcc3dd655deae62acbcef929c7 100644
--- a/tensorflow/python/ops/linalg/linear_operator_test_util.py
+++ b/tensorflow/python/ops/linalg/linear_operator_test_util.py
@@ -102,7 +102,7 @@ class LinearOperatorDerivedClassTest(test.TestCase):
     raise NotImplementedError("operator_build_infos has not been implemented.")
 
   @abc.abstractmethod
-  def _operator_and_mat_and_feed_dict(self, build_info, dtype, use_placeholder):
+  def _operator_and_matrix(self, build_info, dtype, use_placeholder):
     """Build a batch matrix and an Operator that should have similar behavior.
 
     Every operator acts like a (batch) matrix.  This method returns both
@@ -118,9 +118,6 @@ class LinearOperatorDerivedClassTest(test.TestCase):
     Returns:
       operator:  `LinearOperator` subclass instance.
       mat:  `Tensor` representing operator.
-      feed_dict:  Dictionary.
-        If placholder is True, this must contains everything needed to be fed
-          to sess.run calls at runtime to make the operator work.
     """
     # Create a matrix as a numpy array with desired shape/dtype.
     # Create a LinearOperator that should have the same behavior as the matrix.
@@ -189,12 +186,12 @@ class LinearOperatorDerivedClassTest(test.TestCase):
         for dtype in self._dtypes_to_test:
           with self.test_session(graph=ops.Graph()) as sess:
             sess.graph.seed = random_seed.DEFAULT_GRAPH_SEED
-            operator, mat, feed_dict = self._operator_and_mat_and_feed_dict(
+            operator, mat = self._operator_and_matrix(
                 build_info, dtype, use_placeholder=use_placeholder)
             op_dense = operator.to_dense()
             if not use_placeholder:
               self.assertAllEqual(build_info.shape, op_dense.get_shape())
-            op_dense_v, mat_v = sess.run([op_dense, mat], feed_dict=feed_dict)
+            op_dense_v, mat_v = sess.run([op_dense, mat])
             self.assertAC(op_dense_v, mat_v)
 
   def test_det(self):
@@ -204,14 +201,13 @@ class LinearOperatorDerivedClassTest(test.TestCase):
         for dtype in self._dtypes_to_test:
           with self.test_session(graph=ops.Graph()) as sess:
             sess.graph.seed = random_seed.DEFAULT_GRAPH_SEED
-            operator, mat, feed_dict = self._operator_and_mat_and_feed_dict(
+            operator, mat = self._operator_and_matrix(
                 build_info, dtype, use_placeholder=use_placeholder)
             op_det = operator.determinant()
             if not use_placeholder:
               self.assertAllEqual(build_info.shape[:-2], op_det.get_shape())
             op_det_v, mat_det_v = sess.run(
-                [op_det, linalg_ops.matrix_determinant(mat)],
-                feed_dict=feed_dict)
+                [op_det, linalg_ops.matrix_determinant(mat)])
             self.assertAC(op_det_v, mat_det_v)
 
   def test_log_abs_det(self):
@@ -221,7 +217,7 @@ class LinearOperatorDerivedClassTest(test.TestCase):
         for dtype in self._dtypes_to_test:
           with self.test_session(graph=ops.Graph()) as sess:
             sess.graph.seed = random_seed.DEFAULT_GRAPH_SEED
-            operator, mat, feed_dict = self._operator_and_mat_and_feed_dict(
+            operator, mat = self._operator_and_matrix(
                 build_info, dtype, use_placeholder=use_placeholder)
             op_log_abs_det = operator.log_abs_determinant()
             _, mat_log_abs_det = linalg.slogdet(mat)
@@ -229,7 +225,7 @@ class LinearOperatorDerivedClassTest(test.TestCase):
               self.assertAllEqual(
                   build_info.shape[:-2], op_log_abs_det.get_shape())
             op_log_abs_det_v, mat_log_abs_det_v = sess.run(
-                [op_log_abs_det, mat_log_abs_det], feed_dict=feed_dict)
+                [op_log_abs_det, mat_log_abs_det])
             self.assertAC(op_log_abs_det_v, mat_log_abs_det_v)
 
   def _test_matmul(self, with_batch):
@@ -246,7 +242,7 @@ class LinearOperatorDerivedClassTest(test.TestCase):
             for adjoint_arg in self._adjoint_arg_options:
               with self.test_session(graph=ops.Graph()) as sess:
                 sess.graph.seed = random_seed.DEFAULT_GRAPH_SEED
-                operator, mat, feed_dict = self._operator_and_mat_and_feed_dict(
+                operator, mat = self._operator_and_matrix(
                     build_info, dtype, use_placeholder=use_placeholder)
                 x = self._make_x(
                     operator, adjoint=adjoint, with_batch=with_batch)
@@ -264,7 +260,7 @@ class LinearOperatorDerivedClassTest(test.TestCase):
                   self.assertAllEqual(op_matmul.get_shape(),
                                       mat_matmul.get_shape())
                 op_matmul_v, mat_matmul_v = sess.run(
-                    [op_matmul, mat_matmul], feed_dict=feed_dict)
+                    [op_matmul, mat_matmul])
                 self.assertAC(op_matmul_v, mat_matmul_v)
 
   def test_matmul(self):
@@ -289,7 +285,7 @@ class LinearOperatorDerivedClassTest(test.TestCase):
             for adjoint_arg in self._adjoint_arg_options:
               with self.test_session(graph=ops.Graph()) as sess:
                 sess.graph.seed = random_seed.DEFAULT_GRAPH_SEED
-                operator, mat, feed_dict = self._operator_and_mat_and_feed_dict(
+                operator, mat = self._operator_and_matrix(
                     build_info, dtype, use_placeholder=use_placeholder)
                 rhs = self._make_rhs(
                     operator, adjoint=adjoint, with_batch=with_batch)
@@ -307,8 +303,7 @@ class LinearOperatorDerivedClassTest(test.TestCase):
                 if not use_placeholder:
                   self.assertAllEqual(op_solve.get_shape(),
                                       mat_solve.get_shape())
-                op_solve_v, mat_solve_v = sess.run(
-                    [op_solve, mat_solve], feed_dict=feed_dict)
+                op_solve_v, mat_solve_v = sess.run([op_solve, mat_solve])
                 self.assertAC(op_solve_v, mat_solve_v)
 
   def test_solve(self):
@@ -326,14 +321,13 @@ class LinearOperatorDerivedClassTest(test.TestCase):
         for dtype in self._dtypes_to_test:
           with self.test_session(graph=ops.Graph()) as sess:
             sess.graph.seed = random_seed.DEFAULT_GRAPH_SEED
-            operator, mat, feed_dict = self._operator_and_mat_and_feed_dict(
+            operator, mat = self._operator_and_matrix(
                 build_info, dtype, use_placeholder=use_placeholder)
             op_trace = operator.trace()
             mat_trace = math_ops.trace(mat)
             if not use_placeholder:
               self.assertAllEqual(op_trace.get_shape(), mat_trace.get_shape())
-            op_trace_v, mat_trace_v = sess.run(
-                [op_trace, mat_trace], feed_dict=feed_dict)
+            op_trace_v, mat_trace_v = sess.run([op_trace, mat_trace])
             self.assertAC(op_trace_v, mat_trace_v)
 
   def test_add_to_tensor(self):
@@ -343,15 +337,14 @@ class LinearOperatorDerivedClassTest(test.TestCase):
         for dtype in self._dtypes_to_test:
           with self.test_session(graph=ops.Graph()) as sess:
             sess.graph.seed = random_seed.DEFAULT_GRAPH_SEED
-            operator, mat, feed_dict = self._operator_and_mat_and_feed_dict(
+            operator, mat = self._operator_and_matrix(
                 build_info, dtype, use_placeholder=use_placeholder)
             op_plus_2mat = operator.add_to_tensor(2 * mat)
 
             if not use_placeholder:
               self.assertAllEqual(build_info.shape, op_plus_2mat.get_shape())
 
-            op_plus_2mat_v, mat_v = sess.run(
-                [op_plus_2mat, mat], feed_dict=feed_dict)
+            op_plus_2mat_v, mat_v = sess.run([op_plus_2mat, mat])
 
             self.assertAC(op_plus_2mat_v, 3 * mat_v)
 
@@ -362,7 +355,7 @@ class LinearOperatorDerivedClassTest(test.TestCase):
         for dtype in self._dtypes_to_test:
           with self.test_session(graph=ops.Graph()) as sess:
             sess.graph.seed = random_seed.DEFAULT_GRAPH_SEED
-            operator, mat, feed_dict = self._operator_and_mat_and_feed_dict(
+            operator, mat = self._operator_and_matrix(
                 build_info, dtype, use_placeholder=use_placeholder)
             op_diag_part = operator.diag_part()
             mat_diag_part = array_ops.matrix_diag_part(mat)
@@ -372,7 +365,7 @@ class LinearOperatorDerivedClassTest(test.TestCase):
                                   op_diag_part.get_shape())
 
             op_diag_part_, mat_diag_part_ = sess.run(
-                [op_diag_part, mat_diag_part], feed_dict=feed_dict)
+                [op_diag_part, mat_diag_part])
 
             self.assertAC(op_diag_part_, mat_diag_part_)
 
diff --git a/tensorflow/python/ops/linalg_ops.py b/tensorflow/python/ops/linalg_ops.py
index a0dfa543f9b3aee15f11b073dc683b1d2d14388f..f4a93560bee558512f33214148ddec22590b9dd6 100644
--- a/tensorflow/python/ops/linalg_ops.py
+++ b/tensorflow/python/ops/linalg_ops.py
@@ -401,7 +401,7 @@ def svd(tensor, full_matrices=False, compute_uv=True, name=None):
   import tensorflow as tf
   import numpy as np
   s, u, v = tf.linalg.svd(a)
-  tf_a_approx = tf.matmul(u, tf.matmul(tf.linalg.diag(s), v, adjoint_v=True))
+  tf_a_approx = tf.matmul(u, tf.matmul(tf.linalg.diag(s), v, adjoint_b=True))
   u, s, v_adj = np.linalg.svd(a, full_matrices=False)
   np_a_approx = np.dot(u, np.dot(np.diag(s), v_adj))
   # tf_a_approx and np_a_approx should be numerically close.
diff --git a/tensorflow/python/ops/logging_ops.py b/tensorflow/python/ops/logging_ops.py
index 8276047cb678f3d340701718156f8a1cfd6831cb..df41933f8a864be3ada72dbf101420c886dfb36b 100644
--- a/tensorflow/python/ops/logging_ops.py
+++ b/tensorflow/python/ops/logging_ops.py
@@ -35,9 +35,12 @@ from tensorflow.python.util.tf_export import tf_export
 
 
 # Assert and Print are special symbols in python, so we must
-# have an upper-case version of them.  For users with Python 3 or Python 2.7
-# with `from __future__ import print_function`, we also allow lowercase.
-@tf_export("Print", "print")
+# have an upper-case version of them.
+#
+# For users with Python 3 or Python 2.7
+# with `from __future__ import print_function`, we could also allow lowercase.
+# See https://github.com/tensorflow/tensorflow/issues/18053
+@tf_export("Print")
 def Print(input_, data, message=None, first_n=None, summarize=None,
           name=None):
   """Prints a list of tensors.
diff --git a/tensorflow/python/ops/losses/losses_impl.py b/tensorflow/python/ops/losses/losses_impl.py
index 9ba91772f5f8d5466eb5cbde4bd48c79fce2bab6..66633c8b12f60c86760f906aa8e4312c7394e796 100644
--- a/tensorflow/python/ops/losses/losses_impl.py
+++ b/tensorflow/python/ops/losses/losses_impl.py
@@ -878,7 +878,8 @@ def sparse_softmax_cross_entropy(
       exception when this op is run on CPU, and return `NaN` for corresponding
       loss and gradient rows on GPU.
     logits: Unscaled log probabilities of shape
-      `[d_0, d_1, ..., d_{r-1}, num_classes]` and dtype `float32` or `float64`.
+      `[d_0, d_1, ..., d_{r-1}, num_classes]` and dtype `float16`, `float32` or
+      `float64`.
     weights: Coefficients for the loss. This must be scalar or broadcastable to
       `labels` (i.e. same rank and each dimension is either 1 or the same).
     scope: the scope for the operations performed in computing the loss.
diff --git a/tensorflow/python/ops/math_ops.py b/tensorflow/python/ops/math_ops.py
index cdb6dc8f22919420ff44e217578315d17cb93d8c..c28dca5137c40ae1884a2e1407675f82aa4fb407 100644
--- a/tensorflow/python/ops/math_ops.py
+++ b/tensorflow/python/ops/math_ops.py
@@ -37,11 +37,11 @@ from tensorflow.python.ops import gen_math_ops
 from tensorflow.python.ops import gen_nn_ops
 from tensorflow.python.ops import gen_sparse_ops
 from tensorflow.python.ops import gen_spectral_ops
-from tensorflow.python.platform import tf_logging as logging
 # go/tf-wildcard-import
 # pylint: disable=wildcard-import
 from tensorflow.python.ops.gen_math_ops import *
 # pylint: enable=wildcard-import
+from tensorflow.python.platform import tf_logging as logging
 from tensorflow.python.util import compat
 from tensorflow.python.util import deprecation
 from tensorflow.python.util import nest
@@ -651,6 +651,9 @@ def cast(x, dtype, name=None):
     TypeError: If `x` cannot be cast to the `dtype`.
   """
   base_type = dtypes.as_dtype(dtype).base_dtype
+  if isinstance(x,
+                (ops.Tensor, _resource_variable_type)) and base_type == x.dtype:
+    return x
   with ops.name_scope(name, "Cast", [x]) as name:
     if isinstance(x, sparse_tensor.SparseTensor):
       values_cast = cast(x.values, base_type, name=name)
@@ -1222,8 +1225,9 @@ def _ReductionDims(x, axis, reduction_indices):
     return axis
   else:
     # Fast path: avoid creating Rank and Range ops if ndims is known.
-    if isinstance(x, ops.Tensor) and x._rank() is not None:  # pylint: disable=protected-access
-      return constant_op.constant(np.arange(x._rank()), dtype=dtypes.int32)  # pylint: disable=protected-access
+    rank = common_shapes.rank(x)
+    if rank is not None:
+      return constant_op.constant(np.arange(rank), dtype=dtypes.int32)
     if (isinstance(x, sparse_tensor.SparseTensor) and
         x.dense_shape.get_shape().is_fully_defined()):
       rank = x.dense_shape.get_shape()[0].value  # sparse.dense_shape is 1-D.
@@ -1234,8 +1238,8 @@ def _ReductionDims(x, axis, reduction_indices):
 
 
 def _may_reduce_to_scalar(keepdims, axis, reduction_indices, output):
-  """Set a reduction's output's shape to be a scalar if we are certain."""
-  if (not output.shape.is_fully_defined()) and (not keepdims) and (
+  """Set a reduction's output shape to be a scalar if we are certain."""
+  if not common_shapes.has_fully_defined_shape(output) and (not keepdims) and (
       axis is None) and (reduction_indices is None):
     output.set_shape(())
   return output
diff --git a/tensorflow/python/ops/math_ops_test.py b/tensorflow/python/ops/math_ops_test.py
index 45e3bd65d272aaecf6663fd8bc5b12210bb5958a..6b709e5e7faf0a74f966f446ba9d33ee1087908a 100644
--- a/tensorflow/python/ops/math_ops_test.py
+++ b/tensorflow/python/ops/math_ops_test.py
@@ -237,8 +237,8 @@ class ApproximateEqualTest(test_util.TensorFlowTestCase):
 
   def testApproximateEqualShape(self):
     for dtype in [np.float32, np.double]:
-      x = np.array([1, 2], dtype=np.float32)
-      y = np.array([[1, 2]], dtype=np.float32)
+      x = np.array([1, 2], dtype=dtype)
+      y = np.array([[1, 2]], dtype=dtype)
       # The inputs 'x' and 'y' must have the same shape.
       with self.assertRaisesRegexp(
           ValueError, "Shapes must be equal rank, but are 1 and 2"):
diff --git a/tensorflow/python/ops/metrics_impl.py b/tensorflow/python/ops/metrics_impl.py
index 5eab12c41d5f781b496a3a1bcfa9ce35fca4fc54..3aedeb6acd94d1fcef1aa3cff768c5b53cf9fdaf 100644
--- a/tensorflow/python/ops/metrics_impl.py
+++ b/tensorflow/python/ops/metrics_impl.py
@@ -73,15 +73,16 @@ def metric_variable(shape, dtype, validate_shape=True, name=None):
     A (non-trainable) variable initialized to zero, or if inside a
     `DistributionStrategy` scope a tower-local variable container.
   """
-  with distribute_lib.get_tower_context().tower_local_var_scope('sum'):
-    # Note that "tower local" implies trainable=False.
-    return variable_scope.variable(
-        lambda: array_ops.zeros(shape, dtype),
-        collections=[
-            ops.GraphKeys.LOCAL_VARIABLES, ops.GraphKeys.METRIC_VARIABLES
-        ],
-        validate_shape=validate_shape,
-        name=name)
+  # Note that synchronization "ON_READ" implies trainable=False.
+  return variable_scope.variable(
+      lambda: array_ops.zeros(shape, dtype),
+      collections=[
+          ops.GraphKeys.LOCAL_VARIABLES, ops.GraphKeys.METRIC_VARIABLES
+      ],
+      validate_shape=validate_shape,
+      synchronization=variable_scope.VariableSynchronization.ON_READ,
+      aggregation=variable_scope.VariableAggregation.SUM,
+      name=name)
 
 
 def _remove_squeezable_dimensions(predictions, labels, weights):
diff --git a/tensorflow/python/ops/nn_ops.py b/tensorflow/python/ops/nn_ops.py
index 0c2f5b06c497e8ca7db20ac09938c86b425d66a0..41d54a6c2f9d8cd961cea398da679fd81361b848 100644
--- a/tensorflow/python/ops/nn_ops.py
+++ b/tensorflow/python/ops/nn_ops.py
@@ -2009,7 +2009,8 @@ def sparse_softmax_cross_entropy_with_logits(
       exception when this op is run on CPU, and return `NaN` for corresponding
       loss and gradient rows on GPU.
     logits: Unscaled log probabilities of shape
-      `[d_0, d_1, ..., d_{r-1}, num_classes]` and dtype `float32` or `float64`.
+      `[d_0, d_1, ..., d_{r-1}, num_classes]` and dtype `float16`, `float32`, or
+      `float64`.
     name: A name for the operation (optional).
 
   Returns:
@@ -2166,7 +2167,7 @@ def _calc_conv_flops(graph, node):
   filter_height = int(filter_shape[0])
   filter_width = int(filter_shape[1])
   filter_in_depth = int(filter_shape[2])
-  output_count = np.prod(output_shape.as_list())
+  output_count = np.prod(output_shape.as_list(), dtype=np.int64)
   return ops.OpStats(
       "flops",
       (output_count * filter_in_depth * filter_height * filter_width * 2))
@@ -2184,7 +2185,7 @@ def _calc_depthwise_conv_flops(graph, node):
   output_shape.assert_is_fully_defined()
   filter_height = int(filter_shape[0])
   filter_width = int(filter_shape[1])
-  output_count = np.prod(output_shape.as_list())
+  output_count = np.prod(output_shape.as_list(), dtype=np.int64)
   return ops.OpStats("flops", (output_count * filter_height * filter_width * 2))
 
 
@@ -2594,7 +2595,7 @@ def _calc_dilation2d_flops(graph, node):
   output_shape.assert_is_fully_defined()
   filter_height = int(filter_shape[0])
   filter_width = int(filter_shape[1])
-  output_count = np.prod(output_shape.as_list())
+  output_count = np.prod(output_shape.as_list(), dtype=np.int64)
   return ops.OpStats("flops", (output_count * filter_height * filter_width * 2))
 
 
diff --git a/tensorflow/python/ops/parallel_for/BUILD b/tensorflow/python/ops/parallel_for/BUILD
new file mode 100644
index 0000000000000000000000000000000000000000..065c2caedc9d334543512941f3513e45360b460f
--- /dev/null
+++ b/tensorflow/python/ops/parallel_for/BUILD
@@ -0,0 +1,129 @@
+package(
+    default_visibility = [
+        "//tensorflow:internal",
+    ],
+)
+
+load("//tensorflow:tensorflow.bzl", "cuda_py_test")
+
+licenses(["notice"])  # Apache 2.0
+
+py_library(
+    name = "parallel_for",
+    srcs = [
+        "__init__.py",
+        "control_flow_ops.py",
+        "gradients.py",
+        "pfor.py",
+    ],
+    srcs_version = "PY2AND3",
+    deps = [
+        ":control_flow_ops",
+        ":gradients",
+        "//tensorflow/python:array_ops",
+        "//tensorflow/python:check_ops",
+        "//tensorflow/python:constant_op",
+        "//tensorflow/python:control_flow_ops",
+        "//tensorflow/python:data_flow_ops",
+        "//tensorflow/python:dtypes",
+        "//tensorflow/python:framework_ops",
+        "//tensorflow/python:functional_ops",
+        "//tensorflow/python:gradients",
+        "//tensorflow/python:math_ops",
+        "//tensorflow/python:nn_ops",
+        "//tensorflow/python:platform",
+        "//tensorflow/python:sparse_ops",
+        "//tensorflow/python:sparse_tensor",
+        "//tensorflow/python:tensor_array_ops",
+        "//tensorflow/python:tensor_shape",
+        "//tensorflow/python:tensor_util",
+        "//tensorflow/python:util",
+        "@absl_py//absl/flags",
+    ],
+)
+
+py_library(
+    name = "pfor_lib",
+    srcs = ["pfor.py"],
+    srcs_version = "PY2AND3",
+    deps = [
+        "//tensorflow/python:array_ops",
+        "//tensorflow/python:check_ops",
+        "//tensorflow/python:constant_op",
+        "//tensorflow/python:control_flow_ops",
+        "//tensorflow/python:data_flow_ops",
+        "//tensorflow/python:dtypes",
+        "//tensorflow/python:framework_ops",
+        "//tensorflow/python:functional_ops",
+        "//tensorflow/python:math_ops",
+        "//tensorflow/python:nn_ops",
+        "//tensorflow/python:platform",
+        "//tensorflow/python:sparse_ops",
+        "//tensorflow/python:sparse_tensor",
+        "//tensorflow/python:tensor_array_ops",
+        "//tensorflow/python:tensor_shape",
+        "//tensorflow/python:tensor_util",
+        "@absl_py//absl/flags",
+    ],
+)
+
+py_library(
+    name = "control_flow_ops",
+    srcs = ["control_flow_ops.py"],
+    srcs_version = "PY2AND3",
+    visibility = ["//visibility:public"],
+    deps = [
+        ":pfor_lib",
+        "//tensorflow/python:array_ops",
+        "//tensorflow/python:control_flow_ops",
+        "//tensorflow/python:dtypes",
+        "//tensorflow/python:framework_ops",
+        "//tensorflow/python:tensor_array_ops",
+        "//tensorflow/python:util",
+    ],
+)
+
+cuda_py_test(
+    name = "control_flow_ops_test",
+    srcs = ["control_flow_ops_test.py"],
+    additional_deps = [
+        ":control_flow_ops",
+        "//tensorflow/core:protos_all_py",
+        "//tensorflow/python:client_testlib",
+        "//tensorflow/python:gradients",
+        "//tensorflow/python:logging_ops",
+        "//tensorflow/python:parsing_ops",
+        "//tensorflow/python:session",
+        "//tensorflow/python:tensor_array_grad",
+        "//tensorflow/python:random_ops",
+        "//tensorflow/python:util",
+    ],
+)
+
+py_library(
+    name = "gradients",
+    srcs = ["gradients.py"],
+    srcs_version = "PY2AND3",
+    deps = [
+        ":control_flow_ops",
+        "//tensorflow/python:array_ops",
+        "//tensorflow/python:gradients",
+        "//tensorflow/python:util",
+    ],
+)
+
+cuda_py_test(
+    name = "gradients_test",
+    size = "large",
+    srcs = ["gradients_test.py"],
+    additional_deps = [
+        ":control_flow_ops",
+        ":gradients",
+        "//third_party/py/numpy",
+        "//tensorflow/python:layers",
+        "//tensorflow/python:client_testlib",
+        "//tensorflow/python:random_ops",
+        "//tensorflow/python/ops/losses",
+    ],
+    tags = ["no_gpu"],  # TODO(b/80127739): test is flaky
+)
diff --git a/tensorflow/python/ops/parallel_for/__init__.py b/tensorflow/python/ops/parallel_for/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..b49d865968b0bab02380cb934431f4933590570e
--- /dev/null
+++ b/tensorflow/python/ops/parallel_for/__init__.py
@@ -0,0 +1,35 @@
+# Copyright 2018 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""Ops for pfor, for_loop, jacobian."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+from tensorflow.python.ops.parallel_for import *  # pylint: disable=wildcard-import
+from tensorflow.python.ops.parallel_for.control_flow_ops import for_loop
+from tensorflow.python.ops.parallel_for.control_flow_ops import pfor
+from tensorflow.python.ops.parallel_for.gradients import batch_jacobian
+from tensorflow.python.ops.parallel_for.gradients import jacobian
+from tensorflow.python.util.all_util import remove_undocumented
+
+_allowed_symbols = [
+    'pfor',
+    'for_loop',
+    'jacobian',
+    'batch_jacobian',
+]
+
+remove_undocumented(__name__, _allowed_symbols)
diff --git a/tensorflow/python/ops/parallel_for/control_flow_ops.py b/tensorflow/python/ops/parallel_for/control_flow_ops.py
new file mode 100644
index 0000000000000000000000000000000000000000..ccf2eb82146969532c84b7d56d40974e94337507
--- /dev/null
+++ b/tensorflow/python/ops/parallel_for/control_flow_ops.py
@@ -0,0 +1,123 @@
+# Copyright 2018 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""for_loop and pfor ops."""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+from tensorflow.python.framework import dtypes
+from tensorflow.python.framework import ops
+from tensorflow.python.ops import array_ops
+from tensorflow.python.ops import control_flow_ops
+from tensorflow.python.ops import tensor_array_ops
+from tensorflow.python.ops.parallel_for.pfor import PFor
+from tensorflow.python.util import nest
+
+
+def for_loop(loop_fn, loop_fn_dtypes, iters):
+  """Runs `loop_fn` `iters` times and stacks the outputs.
+
+
+  Runs `loop_fn` `iters` times, with input values from 0 to `iters - 1`, and
+  stacks corresponding outputs of the different runs.
+
+  Args:
+    loop_fn: A function that takes an int32 scalar tf.Tensor object representing
+      the iteration number, and returns a possibly nested structure of tensor
+      objects. The shape of these outputs should not depend on the input.
+    loop_fn_dtypes: dtypes for the outputs of loop_fn.
+    iters: Number of iterations for which to run loop_fn.
+
+  Returns:
+    Returns a nested structure of stacked output tensor objects with the same
+    nested structure as the output of `loop_fn`.
+  """
+
+  flat_loop_fn_dtypes = nest.flatten(loop_fn_dtypes)
+
+  def while_body(i, *ta_list):
+    """Body of while loop."""
+    fn_output = nest.flatten(loop_fn(i))
+    if len(fn_output) != len(flat_loop_fn_dtypes):
+      raise ValueError(
+          "Number of expected outputs, %d, does not match the number of "
+          "actual outputs, %d, from loop_fn" % (len(flat_loop_fn_dtypes),
+                                                len(fn_output)))
+    outputs = []
+    for out, ta in zip(fn_output, ta_list):
+      # TODO(agarwal): support returning Operation objects from loop_fn.
+      assert isinstance(out, ops.Tensor)
+      outputs.append(ta.write(i, array_ops.expand_dims(out, 0)))
+    return tuple([i + 1] + outputs)
+
+  ta_list = control_flow_ops.while_loop(
+      lambda i, *ta: i < iters, while_body, [0] + [
+          tensor_array_ops.TensorArray(dtype, iters)
+          for dtype in flat_loop_fn_dtypes
+      ])[1:]
+
+  # TODO(rachelim): enable this for sparse tensors
+  return nest.pack_sequence_as(loop_fn_dtypes, [ta.concat() for ta in ta_list])
+
+
+def pfor(loop_fn, iters):
+  """Equivalent to running `loop_fn` `iters` times and stacking the outputs.
+
+  `pfor` has functionality similar to `for_loop`, i.e. running `loop_fn` `iters`
+  times, with input from 0 to `iters - 1`, and stacking corresponding output of
+  each iteration. However the implementation does not use a tf.while_loop.
+  Instead it adds new operations to the graph that collectively compute the same
+  value as what running `loop_fn` in a loop would compute.
+
+
+  This is an experimental feature and currently has a lot of limitations:
+    - There should be no data depenendency between the different iterations. For
+      example, a future iteration should not depend on a value or side-effect of
+      a previous iteration.
+    - Stateful kernels may mostly not be supported since these often imply a
+      data dependency or ordering of the iterations. We do support a limited set
+      of such stateful kernels though (like RandomFoo, Variable operations like
+      reads, etc).
+    - Conversion works only on a limited set of kernels for which a converter
+      has been registered.
+    - loop_fn cannot currently contain control flow operations like
+      tf.while_loop or tf.cond.
+    - `loop_fn` should return nested structure of Tensors or Operations. However
+      if an Operation is returned, it should have zero outputs.
+    - The shape and dtype of `loop_fn` outputs should not depend on the input
+      to loop_fn.
+
+  Args:
+    loop_fn: A function that takes an int32 scalar tf.Tensor object representing
+      the iteration number, and returns a possibly nested structure of Tensor or
+      Operation objects.
+    iters: Number of iterations for which to run loop_fn.
+
+  Returns:
+    Returns a nested structure of stacked tensor objects with the same nested
+    structure as the output of `loop_fn`.
+  """
+  existing_ops = set(ops.get_default_graph().get_operations())
+  with ops.name_scope("loop_body"):
+    loop_var = array_ops.placeholder(dtypes.int32, shape=[])
+    loop_fn_outputs = loop_fn(loop_var)
+  new_ops = set(ops.get_default_graph().get_operations()) - existing_ops
+  iters = ops.convert_to_tensor(iters)
+  with ops.name_scope("pfor"):
+    converter = PFor(loop_var, iters, new_ops)
+    outputs = []
+    for loop_fn_output in nest.flatten(loop_fn_outputs):
+      outputs.append(converter.convert(loop_fn_output))
+    return nest.pack_sequence_as(loop_fn_outputs, outputs)
diff --git a/tensorflow/python/ops/parallel_for/control_flow_ops_test.py b/tensorflow/python/ops/parallel_for/control_flow_ops_test.py
new file mode 100644
index 0000000000000000000000000000000000000000..c0e66cb0b874b183d53cc34dbb3aa3d182e255a4
--- /dev/null
+++ b/tensorflow/python/ops/parallel_for/control_flow_ops_test.py
@@ -0,0 +1,1404 @@
+# Copyright 2018 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""Tests for pfor and for_loop."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import time
+
+from absl import flags
+import numpy as np
+
+from tensorflow.core.example import example_pb2
+from tensorflow.core.example import feature_pb2
+from tensorflow.python.client import session
+from tensorflow.python.framework import constant_op
+from tensorflow.python.framework import dtypes
+from tensorflow.python.framework import ops
+from tensorflow.python.framework import sparse_tensor
+from tensorflow.python.ops import array_ops
+from tensorflow.python.ops import control_flow_ops
+from tensorflow.python.ops import data_flow_ops
+from tensorflow.python.ops import gradients as gradient_ops
+from tensorflow.python.ops import logging_ops
+from tensorflow.python.ops import math_ops
+from tensorflow.python.ops import nn
+from tensorflow.python.ops import parsing_ops
+from tensorflow.python.ops import random_ops
+from tensorflow.python.ops import rnn
+from tensorflow.python.ops import rnn_cell
+from tensorflow.python.ops import tensor_array_grad  # pylint: disable=unused-import
+from tensorflow.python.ops import tensor_array_ops
+from tensorflow.python.ops import variables
+from tensorflow.python.ops.parallel_for import control_flow_ops as pfor_control_flow_ops
+from tensorflow.python.platform import test
+from tensorflow.python.util import nest
+
+
+class PForTest(test.TestCase):
+
+  def _run_targets(self, targets1, targets2=None, run_init=True):
+    targets1 = nest.flatten(targets1)
+    targets2 = ([] if targets2 is None else nest.flatten(targets2))
+    assert len(targets1) == len(targets2) or not targets2
+    if run_init:
+      init = variables.global_variables_initializer()
+      self.evaluate(init)
+    return self.evaluate(targets1 + targets2)
+
+  def run_and_assert_equal(self, targets1, targets2):
+    outputs = self._run_targets(targets1, targets2)
+    outputs = nest.flatten(outputs)  # flatten SparseTensorValues
+    n = len(outputs) // 2
+    for i in range(n):
+      if outputs[i + n].dtype != np.object:
+        self.assertAllClose(outputs[i + n], outputs[i], rtol=1e-4, atol=1e-5)
+      else:
+        self.assertAllEqual(outputs[i + n], outputs[i])
+
+  def _test_loop_fn(self, loop_fn, iters, loop_fn_dtypes=dtypes.float32):
+    t1 = pfor_control_flow_ops.pfor(loop_fn, iters=iters)
+    t2 = pfor_control_flow_ops.for_loop(loop_fn, loop_fn_dtypes, iters=iters)
+    self.run_and_assert_equal(t1, t2)
+
+  def test_op_conversion_fallback_to_while_loop(self):
+    # Note that we used top_k op for this test. If a converter gets defined for
+    # it, we will need to find another op for which a converter has not been
+    # defined.
+    x = random_ops.random_uniform([3, 2, 4])
+
+    def loop_fn(i):
+      x_i = array_ops.gather(x, i)
+      return nn.top_k(x_i)
+
+    with self.assertRaisesRegexp(ValueError, "No converter defined"):
+      self._test_loop_fn(
+          loop_fn, 3, loop_fn_dtypes=[dtypes.float32, dtypes.int32])
+    flags.FLAGS.op_conversion_fallback_to_while_loop = True
+    self._test_loop_fn(
+        loop_fn, 3, loop_fn_dtypes=[dtypes.float32, dtypes.int32])
+    flags.FLAGS.op_conversion_fallback_to_while_loop = False
+
+
+class ArrayTest(PForTest):
+
+  def test_gather(self):
+    x = random_ops.random_uniform([3, 3, 3])
+
+    def loop_fn(i):
+      outputs = []
+      x_i = array_ops.gather(x, i)
+      for y in [x, x_i]:
+        axes = [0, 2, -1] if y == x else [0]
+        for axis in axes:
+          outputs.append(array_ops.gather(y, 2, axis=axis))
+          outputs.append(array_ops.gather(y, i, axis=axis))
+          outputs.append(array_ops.gather(y, [i], axis=axis))
+          outputs.append(array_ops.gather(y, [i, 2], axis=axis))
+          outputs.append(array_ops.gather(y, [[2, i], [i, 1]], axis=axis))
+      return outputs
+
+    self._test_loop_fn(loop_fn, 3, loop_fn_dtypes=[dtypes.float32] * 20)
+
+  def test_shape(self):
+    x = random_ops.random_uniform([3, 2, 3])
+
+    def loop_fn(i):
+      x_i = array_ops.gather(x, i)
+      return array_ops.shape(x_i), array_ops.shape(x_i, out_type=dtypes.int64)
+
+    self._test_loop_fn(loop_fn, 3, loop_fn_dtypes=[dtypes.int32, dtypes.int64])
+
+  def test_size(self):
+    x = random_ops.random_uniform([3, 2, 3])
+
+    def loop_fn(i):
+      x_i = array_ops.gather(x, i)
+      return array_ops.size(x_i), array_ops.size(x_i, out_type=dtypes.int64)
+
+    self._test_loop_fn(loop_fn, 3, loop_fn_dtypes=[dtypes.int32, dtypes.int64])
+
+  def test_rank(self):
+    x = random_ops.random_uniform([3, 2, 3])
+
+    def loop_fn(i):
+      x_i = array_ops.gather(x, i)
+      return array_ops.rank(x_i)
+
+    self._test_loop_fn(loop_fn, 3, loop_fn_dtypes=[dtypes.int32])
+
+  def test_shape_n(self):
+    x = random_ops.random_uniform([3, 2, 3])
+    y = random_ops.random_uniform([3])
+
+    def loop_fn(i):
+      x_i = array_ops.gather(x, i)
+      y_i = array_ops.gather(y, i)
+      return array_ops.shape_n([x_i, x, y, y_i]), array_ops.shape_n(
+          [x_i, x, y, y_i], out_type=dtypes.int64)
+
+    self._test_loop_fn(
+        loop_fn, 3, loop_fn_dtypes=[dtypes.int32] * 4 + [dtypes.int64] * 4)
+
+  def test_reshape(self):
+    x = random_ops.random_uniform([3, 2, 3])
+
+    def loop_fn(i):
+      x1 = array_ops.gather(x, i)
+      return array_ops.reshape(x1, [-1]), array_ops.reshape(x1, [1, 3, 1, -1])
+
+    self._test_loop_fn(loop_fn, 3, loop_fn_dtypes=[dtypes.float32] * 2)
+
+  def test_expand_dims(self):
+    x = random_ops.random_uniform([3, 2, 3])
+
+    def loop_fn(i):
+      x1 = array_ops.gather(x, i)
+      return array_ops.expand_dims(
+          x1, axis=-1), array_ops.expand_dims(
+              x1, axis=1)
+
+    self._test_loop_fn(loop_fn, 3, loop_fn_dtypes=[dtypes.float32] * 2)
+
+  def test_slice(self):
+    x = random_ops.random_uniform([3, 2, 3])
+
+    def loop_fn(i):
+      x1 = array_ops.gather(x, i)
+      return array_ops.slice(x1, begin=(0, 1), size=(2, 1))
+
+    self._test_loop_fn(loop_fn, 3)
+
+  def test_tile(self):
+    x = random_ops.random_uniform([3, 2, 3])
+
+    def loop_fn(i):
+      x1 = array_ops.gather(x, i)
+      return array_ops.tile(x1, [2, 1])
+
+    self._test_loop_fn(loop_fn, 3)
+
+  def test_tile_loop_dependent(self):
+    x = random_ops.random_uniform([3, 2, 3])
+
+    def loop_fn(i):
+      x1 = array_ops.gather(x, i)
+      return array_ops.tile(x1, [i, 1])
+
+    with self.assertRaisesRegexp(ValueError, "expected to be loop invariant"):
+      pfor_control_flow_ops.pfor(loop_fn, 2)
+
+  def test_pack(self):
+    x = random_ops.random_uniform([3, 2, 3])
+    y = random_ops.random_uniform([2, 3])
+
+    def loop_fn(i):
+      x1 = array_ops.gather(x, i)
+      return array_ops.stack([x1, y], axis=-1)
+
+    self._test_loop_fn(loop_fn, 1)
+
+  def test_unpack(self):
+    x = random_ops.random_uniform([3, 2, 3, 4])
+
+    def loop_fn(i):
+      x_i = array_ops.gather(x, i)
+      return array_ops.unstack(
+          x_i, 4, axis=-1), array_ops.unstack(
+              x_i, 3, axis=1)
+
+    self._test_loop_fn(loop_fn, 3, loop_fn_dtypes=[dtypes.float32] * 7)
+
+  def test_pad(self):
+    x = random_ops.random_uniform([3, 2, 3])
+    padding = constant_op.constant([[1, 2], [3, 4]])
+
+    def loop_fn(i):
+      x1 = array_ops.gather(x, i)
+      return array_ops.pad(x1, padding, mode="CONSTANT")
+
+    self._test_loop_fn(loop_fn, 3)
+
+  def test_split(self):
+    x = random_ops.random_uniform([3, 2, 3])
+
+    def loop_fn(i):
+      x1 = array_ops.gather(x, i)
+      return array_ops.split(x1, 2, axis=0), array_ops.split(x1, 3, axis=-1)
+
+    self._test_loop_fn(loop_fn, 3, loop_fn_dtypes=[dtypes.float32] * 5)
+
+  def test_transpose(self):
+    x = random_ops.random_uniform([3, 2, 3, 4])
+
+    def loop_fn(i):
+      x1 = array_ops.gather(x, i)
+      return array_ops.transpose(x1, [2, 1, 0])
+
+    self._test_loop_fn(loop_fn, 3)
+
+  def test_zeros_like(self):
+    x = random_ops.random_uniform([3, 2, 3])
+
+    def loop_fn(i):
+      x1 = array_ops.gather(x, i)
+      z = array_ops.zeros_like(x1),
+      return z, z + x1
+
+    self._test_loop_fn(loop_fn, 3, loop_fn_dtypes=[dtypes.float32] * 2)
+
+  def test_concat_v2(self):
+    x = random_ops.random_uniform([3, 2, 3])
+    y = random_ops.random_uniform([2, 3])
+
+    def loop_fn(i):
+      x1 = array_ops.gather(x, i)
+      return array_ops.concat(
+          [x1, x1, y], axis=0), array_ops.concat(
+              [x1, x1, y], axis=-1)
+
+    self._test_loop_fn(loop_fn, 3, loop_fn_dtypes=[dtypes.float32] * 2)
+
+  def test_unary_cwise_ops(self):
+    for op in [array_ops.identity, array_ops.stop_gradient]:
+      x = random_ops.random_uniform([3, 5])
+
+      # pylint: disable=cell-var-from-loop
+      def loop_fn(i):
+        x1 = array_ops.gather(x, i)
+        y = op(x1) + x1
+        loss = nn.l2_loss(y)
+        return op(x), y, gradient_ops.gradients(loss, x1)
+
+      # pylint: enable=cell-var-from-loop
+
+      self._test_loop_fn(loop_fn, 3, loop_fn_dtypes=[dtypes.float32] * 3)
+
+  def test_strided_slice(self):
+    x = random_ops.random_uniform([3, 3, 4, 4, 2, 2, 2])
+
+    def loop_fn(i):
+      x_i = array_ops.gather(x, i)
+      y = x_i[:2, ::2, 1::3, ..., array_ops.newaxis, 1]
+      loss = nn.l2_loss(y)
+      return y, gradient_ops.gradients(loss, x_i)
+
+    self._test_loop_fn(loop_fn, 3, loop_fn_dtypes=[dtypes.float32] * 2)
+
+
+class MathTest(PForTest):
+
+  def test_unary_cwise_ops(self):
+    for op in [
+        math_ops.tanh, nn.relu, math_ops.sigmoid, math_ops.negative,
+        math_ops.square
+    ]:
+      x = random_ops.random_uniform([3, 5])
+
+      # pylint: disable=cell-var-from-loop
+      def loop_fn(i):
+        x1 = array_ops.gather(x, i)
+        y = op(x1)
+        loss = math_ops.reduce_sum(y * y)
+        return op(x), y, gradient_ops.gradients(loss, x1)
+
+      # pylint: enable=cell-var-from-loop
+
+      self._test_loop_fn(loop_fn, 3, loop_fn_dtypes=[dtypes.float32] * 3)
+
+  def test_unary_cwise_no_grad(self):
+    for op in [math_ops.ceil, math_ops.floor, math_ops.logical_not]:
+      x = random_ops.random_uniform([3, 5])
+      if op == math_ops.logical_not:
+        x = x > 0
+
+      # pylint: disable=cell-var-from-loop
+      def loop_fn(i):
+        return op(array_ops.gather(x, i))
+
+      # pylint: enable=cell-var-from-loop
+
+      self._test_loop_fn(loop_fn, 3, loop_fn_dtypes=x.dtype)
+
+  def test_binary_cwise_ops(self):
+    logical_ops = [
+        math_ops.logical_and, math_ops.logical_or, math_ops.logical_xor
+    ]
+    bool_ops = [
+        math_ops.less, math_ops.less_equal, math_ops.greater,
+        math_ops.greater_equal, math_ops.equal, math_ops.not_equal
+    ]
+    float_ops = [
+        math_ops.add, math_ops.subtract, math_ops.multiply, math_ops.divide,
+        math_ops.maximum, math_ops.minimum
+    ]
+    for op in logical_ops + bool_ops + float_ops:
+      x = random_ops.random_uniform([7, 3, 5])
+      y = random_ops.random_uniform([3, 5])
+      if op in logical_ops:
+        x = x > 0
+        y = y > 0
+
+      # pylint: disable=cell-var-from-loop
+      def loop_fn(i):
+        x1 = array_ops.gather(x, i)
+        y1 = array_ops.gather(y, i)
+        return op(x, y), op(x1, y), op(x, y1), op(x1, y1), op(x1, x1)
+
+      # pylint: enable=cell-var-from-loop
+
+      dtype = dtypes.float32 if op in float_ops else dtypes.bool
+      self._test_loop_fn(loop_fn, 3, loop_fn_dtypes=[dtype] * 5)
+
+  def test_addn(self):
+    x = random_ops.random_uniform([2, 3, 5])
+    y = random_ops.random_uniform([3, 5])
+    z = random_ops.random_uniform([3, 5])
+
+    def loop_fn(i):
+      x1 = array_ops.gather(x, i)
+      return math_ops.add_n([x1, y, z])
+
+    self._test_loop_fn(loop_fn, 2)
+
+  def test_matmul(self):
+    for tr_a in (True, False):
+      for tr_b in (True, False):
+        for stack_a in (True, False):
+          for stack_b in (True, False):
+            shape_a = (5, 3) if tr_a else (3, 5)
+            if stack_a:
+              shape_a = (2,) + shape_a
+            shape_b = (7, 5) if tr_b else (5, 7)
+            if stack_b:
+              shape_b = (2,) + shape_b
+
+            x = random_ops.random_uniform(shape_a)
+            y = random_ops.random_uniform(shape_b)
+
+            # pylint: disable=cell-var-from-loop
+            def loop_fn(i):
+              a = array_ops.gather(x, i) if stack_a else x
+              b = array_ops.gather(y, i) if stack_b else y
+              return math_ops.matmul(a, b, transpose_a=tr_a, transpose_b=tr_b)
+
+            # pylint: enable=cell-var-from-loop
+
+            self._test_loop_fn(loop_fn, 2)
+
+  def test_batch_matmul(self):
+    for tr_a in (True, False):
+      for tr_b in (True, False):
+        for stack_a in (True, False):
+          for stack_b in (True, False):
+            shape_a = (4, 5, 3) if tr_a else (4, 3, 5)
+            if stack_a:
+              shape_a = (2,) + shape_a
+            shape_b = (4, 7, 5) if tr_b else (4, 5, 7)
+            if stack_b:
+              shape_b = (2,) + shape_b
+
+            x = random_ops.random_uniform(shape_a)
+            y = random_ops.random_uniform(shape_b)
+
+            # pylint: disable=cell-var-from-loop
+            def loop_fn(i):
+              a = array_ops.gather(x, i) if stack_a else x
+              b = array_ops.gather(y, i) if stack_b else y
+              return math_ops.matmul(a, b, transpose_a=tr_a, transpose_b=tr_b)
+
+            # pylint: enable=cell-var-from-loop
+
+            self._test_loop_fn(loop_fn, 2)
+
+  def test_reduction(self):
+    x = random_ops.random_uniform([2, 3, 4, 5])
+    for op in [
+        math_ops.reduce_sum, math_ops.reduce_prod, math_ops.reduce_max,
+        math_ops.reduce_min
+    ]:
+      for axis in ([1], None, [0, 2]):
+        for keepdims in (True, False):
+
+          # pylint: disable=cell-var-from-loop
+          def loop_fn(i):
+            a = array_ops.gather(x, i)
+            return op(a, axis=axis, keepdims=keepdims)
+
+          # pylint: enable=cell-var-from-loop
+
+          self._test_loop_fn(loop_fn, 2)
+
+  def test_cum_sum(self):
+    x = random_ops.random_uniform([2, 3, 4, 5])
+    for axis in (1, -2):
+      for exclusive in (True, False):
+        for reverse in (True, False):
+
+          # pylint: disable=cell-var-from-loop
+          def loop_fn(i):
+            a = array_ops.gather(x, i)
+            return math_ops.cumsum(
+                a, axis=axis, exclusive=exclusive, reverse=reverse)
+
+          # pylint: enable=cell-var-from-loop
+
+          self._test_loop_fn(loop_fn, 2)
+
+  def test_cum_prod(self):
+    x = random_ops.random_uniform([2, 3, 4, 5])
+    for axis in (1, -2):
+      for exclusive in (True, False):
+        for reverse in (True, False):
+
+          # pylint: disable=cell-var-from-loop
+          def loop_fn(i):
+            a = array_ops.gather(x, i)
+            return math_ops.cumprod(
+                a, axis=axis, exclusive=exclusive, reverse=reverse)
+
+          # pylint: enable=cell-var-from-loop
+
+          self._test_loop_fn(loop_fn, 2)
+
+  def test_bias_add(self):
+    x_shape = [2, 3, 4, 5, 6]
+    x = random_ops.random_uniform(x_shape)
+    for data_format in ("NCHW", "NHWC"):
+      bias_dim = 2 if data_format == "NCHW" else -1
+      bias_shape = x_shape[bias_dim]
+      bias = random_ops.random_uniform([bias_shape])
+
+      # pylint: disable=cell-var-from-loop
+      def loop_fn(i):
+        a = array_ops.gather(x, i)
+        y = nn.bias_add(a, bias, data_format=data_format)
+        loss = math_ops.reduce_sum(y * y)
+        return y, gradient_ops.gradients(loss, bias)
+
+      # pylint: enable=cell-var-from-loop
+
+      self._test_loop_fn(
+          loop_fn, 2, loop_fn_dtypes=[dtypes.float32, dtypes.float32])
+
+  def test_unsorted_segment_sum(self):
+    t = random_ops.random_uniform([3, 3, 2])
+    segment_ids = constant_op.constant([[0, 0, 2], [0, 1, 2], [2, 2, 2]])
+    num_segments = 3
+
+    def loop_fn(i):
+      data = array_ops.gather(t, i)
+      data_0 = array_ops.gather(t, 0)
+      seg_ids = array_ops.gather(segment_ids, i)
+      return (math_ops.unsorted_segment_sum(data, seg_ids, num_segments),
+              math_ops.unsorted_segment_sum(data_0, seg_ids, num_segments))
+
+    self._test_loop_fn(loop_fn, 3, [dtypes.float32] * 2)
+
+  def test_cast(self):
+    x = constant_op.constant([[1], [2]])
+    y = constant_op.constant([[1.0], [2.0]])
+
+    def loop_fn(i):
+      return (math_ops.cast(array_ops.gather(x, i), dtypes.float32),
+              math_ops.cast(array_ops.gather(y, i), dtypes.int32))
+
+    self._test_loop_fn(
+        loop_fn, 2, loop_fn_dtypes=[dtypes.float32, dtypes.int32])
+
+  def test_tanh_axpy(self):
+    a = constant_op.constant(3.)
+    x = random_ops.random_uniform([4, 5])
+    y = random_ops.random_uniform([6, 5])
+    n = x.shape[0]
+
+    def loop_fn(i):
+      return math_ops.tanh(a * array_ops.gather(x, i) + array_ops.gather(y, i))
+
+    self._test_loop_fn(loop_fn, n)
+
+  def test_select(self):
+    cond = constant_op.constant([True, False])
+    a = random_ops.random_uniform([2, 3, 5])
+    b = random_ops.random_uniform([2, 3, 5])
+    for cond_shape in [2], [2, 3], [2, 3, 5]:
+      cond = random_ops.random_uniform(cond_shape) > 0.5
+
+      # pylint: disable=cell-var-from-loop
+      def loop_fn(i):
+        a_i = array_ops.gather(a, i)
+        b_i = array_ops.gather(b, i)
+        cond_i = array_ops.gather(cond, i)
+        return array_ops.where(cond_i, a_i, b_i)
+
+      # pylint: enable=cell-var-from-loop
+
+      self._test_loop_fn(loop_fn, 2)
+
+
+class NNTest(PForTest):
+
+  def test_conv2d(self):
+    x = random_ops.random_uniform([3, 2, 12, 12, 3])
+    filt = random_ops.random_uniform([3, 3, 3, 7])
+
+    def loop_fn(i):
+      x1 = array_ops.gather(x, i)
+      return nn.conv2d(
+          x1, filt, strides=[1, 2, 2, 1], padding="VALID", data_format="NHWC")
+
+    self._test_loop_fn(loop_fn, 3)
+
+  def test_conv2d_backprop_input(self):
+    x_shape = [2, 12, 12, 3]
+    filt = random_ops.random_uniform([3, 3, 3, 7])
+    grad = random_ops.random_uniform([3, 2, 5, 5, 7])
+
+    def loop_fn(i):
+      grad1 = array_ops.gather(grad, i)
+      return nn.conv2d_backprop_input(
+          x_shape,
+          filt,
+          grad1,
+          strides=[1, 2, 2, 1],
+          padding="VALID",
+          data_format="NHWC")
+
+    self._test_loop_fn(loop_fn, 3)
+
+  def test_conv2d_backprop_filter(self):
+    x = random_ops.random_uniform([3, 2, 12, 12, 3])
+    x_0 = array_ops.gather(x, 0)
+    filter_sizes = [3, 3, 3, 7]
+    grad = random_ops.random_uniform([3, 2, 5, 5, 7])
+
+    def loop_fn(i):
+      x_i = array_ops.gather(x, i)
+      grad_i = array_ops.gather(grad, i)
+      return [
+          nn.conv2d_backprop_filter(
+              inp,
+              filter_sizes,
+              grad_i,
+              strides=[1, 2, 2, 1],
+              padding="VALID",
+              data_format="NHWC") for inp in [x_i, x_0]
+      ]
+
+    self._test_loop_fn(loop_fn, 3, loop_fn_dtypes=[dtypes.float32] * 2)
+
+  def test_avg_pool(self):
+    x = random_ops.random_uniform([3, 2, 12, 12, 3])
+    ksize = [1, 3, 3, 1]
+
+    def loop_fn(i):
+      x1 = array_ops.gather(x, i)
+      output = nn.avg_pool(
+          x1, ksize, strides=[1, 2, 2, 1], padding="VALID", data_format="NHWC")
+      loss = nn.l2_loss(output)
+      return output, gradient_ops.gradients(loss, x1)
+
+    self._test_loop_fn(loop_fn, 3, loop_fn_dtypes=[dtypes.float32] * 2)
+
+  def test_max_pool(self):
+    x = random_ops.random_uniform([3, 2, 12, 12, 3])
+    ksize = [1, 3, 3, 1]
+
+    def loop_fn(i):
+      x1 = array_ops.gather(x, i)
+      output = nn.max_pool(
+          x1, ksize, strides=[1, 2, 2, 1], padding="VALID", data_format="NHWC")
+      loss = nn.l2_loss(output)
+      return output, gradient_ops.gradients(loss, x1)
+
+    self._test_loop_fn(loop_fn, 3, loop_fn_dtypes=[dtypes.float32] * 2)
+
+  def test_fused_batch_norm(self):
+    data_formats = ["NHWC"]
+    if test.is_gpu_available():
+      data_formats.append("NCHW")
+    for is_training in (True, False):
+      for data_format in data_formats:
+        if data_format == "NCHW":
+          x = random_ops.random_uniform([3, 1, 2, 5, 5])
+        else:
+          x = random_ops.random_uniform([3, 1, 5, 5, 2])
+        scale = random_ops.random_uniform([2])
+        offset = random_ops.random_uniform([2])
+        mean = None if is_training else random_ops.random_uniform([2])
+        variance = None if is_training else random_ops.random_uniform([2])
+
+        # pylint: disable=cell-var-from-loop
+        def loop_fn(i):
+          x1 = array_ops.gather(x, i)
+          outputs = nn.fused_batch_norm(
+              x1,
+              scale,
+              offset,
+              mean=mean,
+              variance=variance,
+              epsilon=0.01,
+              data_format=data_format,
+              is_training=is_training)
+          outputs = list(outputs)
+          # We only test the first value of outputs when is_training is False.
+          # It looks like CPU and GPU have different outputs for batch_mean and
+          # batch_variance for this case.
+          if not is_training:
+            outputs[1] = constant_op.constant(0.)
+            outputs[2] = constant_op.constant(0.)
+          loss = nn.l2_loss(outputs[0])
+          gradients = gradient_ops.gradients(loss, [x1, scale, offset])
+          return outputs + gradients
+
+        # pylint: enable=cell-var-from-loop
+
+        self._test_loop_fn(loop_fn, 3, loop_fn_dtypes=[dtypes.float32] * 6)
+
+  def test_softmax_cross_entropy_with_logits(self):
+    logits = random_ops.random_uniform([3, 2, 4])
+    labels = random_ops.random_uniform([3, 2, 4])
+    labels /= math_ops.reduce_sum(labels, axis=[2], keepdims=True)
+
+    def loop_fn(i):
+      logits_i = array_ops.gather(logits, i)
+      labels_i = array_ops.gather(labels, i)
+      loss = nn.softmax_cross_entropy_with_logits(
+          labels=labels_i, logits=logits_i)
+      return loss, gradient_ops.gradients(math_ops.reduce_sum(loss), logits_i)
+
+    self._test_loop_fn(loop_fn, 3, loop_fn_dtypes=[dtypes.float32] * 2)
+
+
+class RandomTest(PForTest):
+
+  # The random values generated in the two implementations are not guaranteed to
+  # match. So we only check the returned shapes.
+  def run_and_assert_equal(self, targets1, targets2):
+    outputs = self._run_targets(targets1, targets2)
+    n = len(outputs) // 2
+    for i in range(n):
+      self.assertAllEqual(outputs[i].shape, outputs[i + n].shape)
+
+  def test_random_uniform(self):
+
+    def loop_fn(_):
+      return random_ops.random_uniform([3])
+
+    self._test_loop_fn(loop_fn, 5)
+
+  def test_random_uniform_int(self):
+
+    def loop_fn(_):
+      return random_ops.random_uniform([3], maxval=1, dtype=dtypes.int32)
+
+    self._test_loop_fn(loop_fn, 5, loop_fn_dtypes=dtypes.int32)
+
+  def test_random_standard_normal(self):
+
+    def loop_fn(_):
+      return random_ops.random_normal([3])
+
+    self._test_loop_fn(loop_fn, 5)
+
+  def test_truncated_normal(self):
+
+    def loop_fn(_):
+      return random_ops.truncated_normal([3])
+
+    self._test_loop_fn(loop_fn, 5)
+
+  def test_random_gamma(self):
+
+    def loop_fn(_):
+      return random_ops.random_gamma([3], alpha=[0.5])
+
+    self._test_loop_fn(loop_fn, 5)
+
+  def test_random_poisson_v2(self):
+
+    def loop_fn(_):
+      return random_ops.random_poisson(lam=[1.3], shape=[3])
+
+    self._test_loop_fn(loop_fn, 5)
+
+
+class LoggingTest(PForTest):
+
+  def test_print(self):
+    x = random_ops.random_uniform([3, 5])
+
+    def loop_fn(i):
+      x1 = array_ops.gather(x, i)
+      return logging_ops.Print(
+          x1, [x1, "x1", array_ops.shape(x1)], summarize=10)
+
+    self._test_loop_fn(loop_fn, 3)
+
+  def test_assert(self):
+
+    def loop_fn(i):
+      return control_flow_ops.Assert(i < 10, [i, [10], [i + 1]])
+
+    # TODO(agarwal): make this work with for_loop.
+    with session.Session() as sess:
+      sess.run(pfor_control_flow_ops.pfor(loop_fn, 3))
+
+
+class TensorArrayTest(PForTest):
+
+  def test_create_outside_and_read(self):
+
+    ta = tensor_array_ops.TensorArray(
+        dtypes.int32, 2, clear_after_read=False).write(0, 0).write(1, 1)
+
+    def loop_fn(i):
+      return ta.read(i), ta.read(0)
+
+    self._test_loop_fn(loop_fn, 2, [dtypes.int32] * 2)
+
+  def test_create_outside_and_gather(self):
+
+    ta = tensor_array_ops.TensorArray(
+        dtypes.int32, 2, clear_after_read=False).write(0, 0).write(1, 1)
+
+    def loop_fn(i):
+      return ta.gather([i]), ta.gather([0, 1])
+
+    self._test_loop_fn(loop_fn, 2, [dtypes.int32] * 2)
+
+  def test_create_outside_and_write_and_scatter(self):
+
+    t = tensor_array_ops.TensorArray(dtypes.int32, 10, clear_after_read=False)
+    handle = t.handle
+
+    def loop_fn(i):
+      ta = t.write(i + 2, 2 * i).write(i, 5)
+      ta = ta.scatter([4 + i], [4]).scatter([6 + i, 8 + i], [6 + i, 8 + i])
+      return ta.flow
+
+    t1 = pfor_control_flow_ops.pfor(loop_fn, iters=2)
+    out1 = tensor_array_ops.TensorArray(
+        dtypes.int32, handle=handle, flow=t1[-1]).stack()
+    output1 = self._run_targets(out1)
+
+    t2 = pfor_control_flow_ops.for_loop(loop_fn, dtypes.float32, iters=2)
+    out2 = tensor_array_ops.TensorArray(
+        dtypes.int32, handle=handle, flow=t2[-1]).stack()
+    output2 = self._run_targets(out2)
+    self.assertAllClose(output2, output1)
+
+  def test_create_inside_and_write(self):
+
+    def loop_fn(i):
+      # TODO(agarwal): switching the order of writes to ta1 does not work.
+      ta1 = tensor_array_ops.TensorArray(dtypes.int32, 2).write(0, i).write(
+          1, 1)
+      ta2 = tensor_array_ops.TensorArray(dtypes.int32, 1).write(0, 1)
+      return ta1.stack(), ta2.stack()
+
+    self._test_loop_fn(loop_fn, 3, [dtypes.int32] * 2)
+
+  def test_create_inside_and_scatter(self):
+
+    def loop_fn(i):
+      # TODO(agarwal): switching the order of scatter to ta1 does not work.
+      ta1 = tensor_array_ops.TensorArray(dtypes.int32, 2).scatter(
+          [0], [[i, 2]]).scatter([1], [[1, 2]])
+      ta2 = tensor_array_ops.TensorArray(dtypes.int32,
+                                         2).scatter([0], [3]).scatter([1], [4])
+      return ta1.stack(), ta2.stack()
+
+    self._test_loop_fn(loop_fn, 3, [dtypes.int32] * 2)
+
+  def test_create_inside_and_read(self):
+
+    def loop_fn(i):
+      ta1 = tensor_array_ops.TensorArray(
+          dtypes.int32, 2, clear_after_read=False).write(0, i).write(1, 1)
+      ta2 = tensor_array_ops.TensorArray(
+          dtypes.int32, 2, clear_after_read=False).write(0, 1).write(1, 2)
+      # TODO(agarwal): ta1.read(i) currently is not supported.
+      return ta1.read(0), ta2.read(0), ta2.read(i)
+
+    self._test_loop_fn(loop_fn, 2, [dtypes.int32] * 3)
+
+  def test_create_inside_and_gather(self):
+
+    def loop_fn(i):
+      ta1 = tensor_array_ops.TensorArray(
+          dtypes.int32, 2, clear_after_read=False).write(0, i).write(1, 1)
+      ta2 = tensor_array_ops.TensorArray(
+          dtypes.int32, 2, clear_after_read=False).write(0, 1).write(1, 2)
+      # TODO(agarwal): ta1.read(i) currently is not supported.
+      return ta1.gather([0, 1]), ta2.gather([0, 1]), ta2.gather([i])
+
+    self._test_loop_fn(loop_fn, 2, [dtypes.int32] * 3)
+
+  def test_grad(self):
+    x = random_ops.random_uniform([3, 2])
+    ta = tensor_array_ops.TensorArray(
+        dtypes.float32, 3, clear_after_read=False).unstack(x)
+    y = math_ops.square(ta.stack())
+
+    def loop_fn(i):
+      y_i = array_ops.gather(y, i)
+      grad = gradient_ops.gradients(y_i, x)[0]
+      return array_ops.gather(grad, i)
+
+    t1 = pfor_control_flow_ops.pfor(loop_fn, iters=3)
+    # y = x * x. Hence dy/dx = 2 * x.
+    actual_grad = 2.0 * x
+    with session.Session() as sess:
+      actual_grad, computed_grad = sess.run([t1, actual_grad])
+      self.assertAllClose(actual_grad, computed_grad)
+
+
+class StackTest(PForTest):
+
+  def test_stack_inside_loop_invariant(self):
+
+    def loop_fn(_):
+      s = data_flow_ops.stack_v2(max_size=4, elem_type=dtypes.int32)
+      op1 = data_flow_ops.stack_push_v2(s, 1)
+      with ops.control_dependencies([op1]):
+        op2 = data_flow_ops.stack_push_v2(s, 2)
+      with ops.control_dependencies([op2]):
+        e2 = data_flow_ops.stack_pop_v2(s, elem_type=dtypes.int32)
+      with ops.control_dependencies([e2]):
+        e1 = data_flow_ops.stack_pop_v2(s, elem_type=dtypes.int32)
+      return e1, e2
+
+    self._test_loop_fn(loop_fn, 2, [dtypes.int32] * 2)
+
+  def test_stack_inside_push_loop_dependent(self):
+
+    def loop_fn(i):
+      s = data_flow_ops.stack_v2(max_size=4, elem_type=dtypes.int32)
+      op1 = data_flow_ops.stack_push_v2(s, i)
+      with ops.control_dependencies([op1]):
+        op2 = data_flow_ops.stack_push_v2(s, 2)
+      with ops.control_dependencies([op2]):
+        e2 = data_flow_ops.stack_pop_v2(s, elem_type=dtypes.int32)
+      with ops.control_dependencies([e2]):
+        e1 = data_flow_ops.stack_pop_v2(s, elem_type=dtypes.int32)
+      return e1, e2
+
+    self._test_loop_fn(loop_fn, 2, [dtypes.int32] * 2)
+
+  def test_stack_outside_pop(self):
+    s = data_flow_ops.stack_v2(max_size=4, elem_type=dtypes.int32)
+    op = data_flow_ops.stack_push_v2(s, 5)
+    with ops.control_dependencies([op]):
+      op = data_flow_ops.stack_push_v2(s, 6)
+    with ops.control_dependencies([op]):
+      op = data_flow_ops.stack_push_v2(s, 7)
+
+    def loop_fn(_):
+      e1 = data_flow_ops.stack_pop_v2(s, elem_type=dtypes.int32)
+      with ops.control_dependencies([e1]):
+        e2 = data_flow_ops.stack_pop_v2(s, elem_type=dtypes.int32)
+      return e1, e2
+
+    with ops.control_dependencies([op]):
+      e1, e2 = pfor_control_flow_ops.pfor(loop_fn, iters=2)
+    with ops.control_dependencies([e1, e2]):
+      e3 = data_flow_ops.stack_pop_v2(s, elem_type=dtypes.int32)
+    v1, v2, v3 = self._run_targets([e1, e2, e3], run_init=False)
+    self.assertAllEqual([7, 7], v1)
+    self.assertAllEqual([6, 6], v2)
+    self.assertAllEqual(5, v3)
+
+  def test_stack_outside_push(self):
+    s = data_flow_ops.stack_v2(max_size=4, elem_type=dtypes.int32)
+
+    def loop_fn(_):
+      return data_flow_ops.stack_push_v2(s, 7)
+
+    with self.assertRaisesRegexp(ValueError, "StackPushV2 not allowed.*"):
+      pfor_control_flow_ops.pfor(loop_fn, iters=2)
+
+
+# TODO(agarwal): test nested while_loops. This currently requires converting a
+# tf.cond.
+class ControlFlowTest(PForTest):
+
+  def test_while_outside_loop(self):
+
+    x = control_flow_ops.while_loop(lambda j: j < 4, lambda j: j + 1, [0])
+
+    def loop_fn(i):
+      return x + i
+
+    self._test_loop_fn(loop_fn, 3, loop_fn_dtypes=[dtypes.int32])
+
+  def test_invariant_while(self):
+
+    def loop_fn(_):
+      return control_flow_ops.while_loop(lambda j: j < 4, lambda j: j + 1, [0])
+
+    self._test_loop_fn(loop_fn, 3, loop_fn_dtypes=[dtypes.int32])
+
+  def test_invariant_while_with_control_dependency(self):
+
+    def loop_fn(i):
+      with ops.control_dependencies([i]):
+        return control_flow_ops.while_loop(lambda j: j < 4, lambda j: j + 1,
+                                           [0])
+
+    self._test_loop_fn(loop_fn, 3, loop_fn_dtypes=[dtypes.int32])
+
+  def test_while_with_stateful_ops(self):
+
+    def loop_fn(_):
+      return control_flow_ops.while_loop(
+          lambda j, x: j < 4,
+          lambda j, x: (j + 1, x + random_ops.random_uniform([])), [0, 0.])[0]
+
+    self._test_loop_fn(loop_fn, 3, loop_fn_dtypes=[dtypes.int32])
+
+  def test_while_unstacked_condition(self):
+
+    def loop_fn(i):
+      return control_flow_ops.while_loop(lambda j, x: j < 4,
+                                         lambda j, x: (j + 1, x + i), [0, 0])
+
+    self._test_loop_fn(loop_fn, 3, loop_fn_dtypes=[dtypes.int32, dtypes.int32])
+
+  def test_while(self):
+    x = random_ops.random_uniform([3, 5])
+    lengths = constant_op.constant([4, 0, 2])
+
+    def loop_fn(i):
+      x_i = array_ops.gather(x, i)
+      lengths_i = array_ops.gather(lengths, i)
+
+      _, total = control_flow_ops.while_loop(
+          lambda j, _: j < lengths_i,
+          lambda j, t: (j + 1, t + array_ops.gather(x_i, j)), [0, 0.])
+      return total
+
+    self._test_loop_fn(loop_fn, 3, loop_fn_dtypes=[dtypes.float32])
+
+  def test_while_jacobian(self):
+    x = random_ops.random_uniform([1, 3])
+    y = random_ops.random_uniform([3, 3])
+
+    # out = x @ y @ y @ y @ y, where @ is matmul operator.
+    _, out = control_flow_ops.while_loop(
+        lambda i, _: i < 4, lambda i, out: (i + 1, math_ops.matmul(out, y)),
+        [0, x])
+
+    def loop_fn(i):
+      out_i = array_ops.gather(out, i, axis=1)
+      return array_ops.reshape(gradient_ops.gradients(out_i, x)[0], [-1])
+
+    out = pfor_control_flow_ops.pfor(loop_fn, iters=3)
+
+    # The above code does not work with tf.while_loop instead of pfor. So we
+    # manually compute the expected output here.
+    # Note that gradient of output w.r.t is (y @ y @ y @ y)^T.
+    expected_output = y
+    for _ in range(3):
+      expected_output = math_ops.matmul(expected_output, y)
+    expected_output = array_ops.transpose(expected_output, [1, 0])
+
+    with session.Session() as sess:
+      out, expected = sess.run([out, expected_output])
+      self.assertAllClose(expected, out)
+
+  def test_tensor_array_as_loop_variable(self):
+
+    def loop_fn(i):
+
+      def body(j, ta):
+        ta = ta.write(j, i + j * j)
+        return j + 1, ta
+
+      _, ta = control_flow_ops.while_loop(
+          lambda j, _: j < 4, body,
+          (0, tensor_array_ops.TensorArray(dtypes.int32, size=4)))
+      return ta.stack()
+
+    self._test_loop_fn(loop_fn, 3, loop_fn_dtypes=[dtypes.int32])
+
+  def test_read_tensor_array_partitioned_indices(self):
+    # Note that tensor array values are pfor loop dependent, and the while loop
+    # termination condition is also dependent on pfor iteration.
+    def loop_fn(i):
+      ta = tensor_array_ops.TensorArray(dtypes.int32, size=6)
+      ta = ta.unstack(i + list(range(5)))
+
+      def body(j, s):
+        return j + 1, s + ta.read(j)
+
+      _, s = control_flow_ops.while_loop(lambda j, _: j < i,
+                                         body,
+                                         (0, 0))
+      return s
+
+    self._test_loop_fn(loop_fn, 3, loop_fn_dtypes=[dtypes.int32])
+
+  def test_external_while_loop_grad(self):
+    # Here we test that external while_loops that are extended from inside pfor
+    # (due to gradient calls) are not actually converted. If the below was
+    # converted all pfor iterations would write to the same tensor array
+    # indices.
+    x = constant_op.constant(1.)
+
+    def body(j, ta):
+      ta = ta.write(j, x)
+      return j + 1, ta
+
+    _, ta = control_flow_ops.while_loop(
+        lambda j, _: j < 4, body,
+        (0, tensor_array_ops.TensorArray(dtypes.float32, size=4)))
+    out = ta.stack()
+
+    def loop_fn(i):
+      out_i = array_ops.gather(out, i)
+      return gradient_ops.gradients(out_i, x)[0]
+
+    with session.Session() as sess:
+      # out is [x, x, x]. Hence the gradients should be [1, 1, 1].
+      self.assertAllEqual([1, 1, 1],
+                          sess.run(pfor_control_flow_ops.pfor(loop_fn, 3)))
+
+  def test_tensor_array_grad(self):
+    inp = constant_op.constant(np.random.rand(3, 4, 2), dtype=dtypes.float32)
+    ta = tensor_array_ops.TensorArray(dtypes.float32, size=3)
+    ta = ta.unstack(inp)
+
+    def loop_fn(i):
+
+      def body(j, x):
+        value = ta.gather([j])
+        value = array_ops.gather(array_ops.reshape(value, [4, 2]), i)
+        return j + 1, x + value
+
+      _, out = control_flow_ops.while_loop(lambda j, _: j < 3, body,
+                                           (0, array_ops.zeros([2])))
+      out = math_ops.reduce_prod(out)
+      return out, gradient_ops.gradients(out, inp)[0]
+
+    pfor_out, pfor_out_grad = pfor_control_flow_ops.pfor(loop_fn, 4)
+    # Note that tf.while_loop does not work in the setup above. So we manually
+    # construct the equivalent computation of the above loops here.
+    real_out = math_ops.reduce_sum(inp, reduction_indices=[0])
+    real_out = math_ops.reduce_prod(real_out, reduction_indices=[1])
+    # Note that gradients of real_out will accumulate the gradients across the
+    # output value. Hence we do the same aggregation on pfor_out_grad.
+    real_out_grad = gradient_ops.gradients(real_out, inp)[0]
+    sum_pfor_out_grad = math_ops.reduce_sum(
+        pfor_out_grad, reduction_indices=[0])
+
+    with session.Session() as sess:
+      v1, v2, v1_grad, v2_grad = sess.run(
+          [pfor_out, real_out, sum_pfor_out_grad, real_out_grad])
+      self.assertAllClose(v1, v2)
+      self.assertAllClose(v1_grad, v2_grad)
+
+
+def dynamic_lstm_input_fn(batch_size, state_size, max_steps):
+  # We make inputs and sequence_length constant so that multiple session.run
+  # calls produce the same result.
+  inputs = constant_op.constant(
+      np.random.rand(batch_size, max_steps, state_size), dtype=dtypes.float32)
+  sequence_length = np.random.randint(0, size=[batch_size], high=max_steps + 1)
+  sequence_length = constant_op.constant(sequence_length, dtype=dtypes.int32)
+  return inputs, sequence_length
+
+
+def create_dynamic_lstm(cell_fn, batch_size, state_size, max_steps):
+  cell = cell_fn(state_size)
+  inputs, sequence_length = dynamic_lstm_input_fn(batch_size,
+                                                  state_size,
+                                                  max_steps)
+  inputs_ta = tensor_array_ops.TensorArray(
+      dtypes.float32, size=max_steps, element_shape=[batch_size, state_size])
+  inputs_time_major = array_ops.transpose(inputs, [1, 0, 2])
+  inputs_ta = inputs_ta.unstack(inputs_time_major)
+  zeros = array_ops.zeros([state_size])
+
+  def loop_fn(i):
+    sequence_length_i = array_ops.gather(sequence_length, i)
+
+    def body_fn(t, state, ta):
+      inputs_t = array_ops.expand_dims(
+          array_ops.gather(inputs_ta.read(t), i), 0)
+      output, new_state = cell(inputs_t, state)
+      output = array_ops.reshape(output, [-1])
+      # TODO(agarwal): one optimization that dynamic_rnn uses is to avoid the
+      # array_ops.where when t < min(sequence_length). Doing that requires
+      # supporting tf.cond pfor conversion.
+      done = t >= sequence_length_i
+      output = array_ops.where(done, zeros, output)
+      ta = ta.write(t, output)
+      new_state = [array_ops.where(done, s, ns) for s, ns in
+                   zip(nest.flatten(state), nest.flatten(new_state))]
+      new_state = nest.pack_sequence_as(state, new_state)
+      return t + 1, new_state, ta
+
+    def condition_fn(t, _, unused):
+      del unused
+      return t < max_steps
+
+    initial_state = cell.zero_state(1, dtypes.float32)
+    _, state, ta = control_flow_ops.while_loop(condition_fn, body_fn, [
+        0, initial_state,
+        tensor_array_ops.TensorArray(dtypes.float32, max_steps)
+    ])
+
+    new_state = [array_ops.reshape(x, [-1]) for x in nest.flatten(state)]
+    new_state = nest.pack_sequence_as(initial_state, new_state)
+    return ta.stack(), new_state
+
+  pfor_output = pfor_control_flow_ops.pfor(loop_fn, batch_size)
+  tf_output = rnn.dynamic_rnn(
+      cell,
+      inputs,
+      sequence_length=sequence_length,
+      initial_state=cell.zero_state(batch_size, dtypes.float32))
+  return pfor_output, tf_output
+
+
+class RNNTest(PForTest):
+
+  def test_dynamic_rnn(self):
+    pfor_outputs, tf_outputs = create_dynamic_lstm(rnn_cell.BasicRNNCell,
+                                                   3, 5, 7)
+    self.run_and_assert_equal(pfor_outputs, tf_outputs)
+
+  def test_dynamic_lstm(self):
+    pfor_outputs, tf_outputs = create_dynamic_lstm(rnn_cell.BasicLSTMCell,
+                                                   3, 5, 7)
+    self.run_and_assert_equal(pfor_outputs, tf_outputs)
+
+
+# TODO(agarwal): benchmark numbers on GPU for graphs based on while_loop
+# conversion don't look good. Some of it seems like lot of copies between host
+# and device. Optimize that.
+class Benchmarks(test.Benchmark):
+
+  def _run(self, targets, iters, name=None):
+
+    def _done(t):
+      # Note that we don't use tf.control_dependencies since that will not make
+      # sure that the computation on GPU has actually finished. So we fetch the
+      # first element of the output, and assume that this will not be called on
+      # empty tensors.
+      return array_ops.gather(array_ops.reshape(t, [-1]), 0)
+
+    targets = [_done(x) for x in nest.flatten(targets)]
+    sess = session.Session()
+    with sess:
+      init = variables.global_variables_initializer()
+      sess.run(init)
+      sess.run(targets)
+      begin = time.time()
+      for _ in range(iters):
+        sess.run(targets)
+      end = time.time()
+    avg_time_ms = 1000 * (end - begin) / iters
+    self.report_benchmark(iters=iters, wall_time=avg_time_ms, name=name)
+    return avg_time_ms
+
+  def benchmark_basic_while(self):
+    with ops.Graph().as_default():
+
+      def loop_fn(i):
+        _, s = control_flow_ops.while_loop(
+            lambda t, x: t < i,
+            lambda t, x: (t + 1, x + i),
+            [0, 0])
+        return s
+
+      iters = 50
+      pfor_output = pfor_control_flow_ops.pfor(loop_fn, iters)
+      for_loop_output = pfor_control_flow_ops.for_loop(loop_fn, dtypes.int32,
+                                                       iters)
+      self._run(pfor_output, 100, name="pfor_basic")
+      self._run(for_loop_output, 100, name="for_loop_basic")
+
+  def benchmark_dynamic_rnn(self):
+    with ops.Graph().as_default():
+      pfor_outputs, tf_outputs = create_dynamic_lstm(rnn_cell.BasicRNNCell,
+                                                     128, 512, 16)
+      self._run(pfor_outputs, 100, name="pfor_rnn")
+      self._run(tf_outputs, 100, name="tf_rnn")
+
+  def benchmark_dynamic_lstm(self):
+    with ops.Graph().as_default():
+      pfor_outputs, tf_outputs = create_dynamic_lstm(rnn_cell.BasicLSTMCell,
+                                                     128, 512, 16)
+      self._run(pfor_outputs, 100, name="pfor_lstm")
+      self._run(tf_outputs, 100, name="tf_lstm")
+
+
+class SparseTest(PForTest):
+
+  def test_var_loop_len(self):
+    num_iters = array_ops.placeholder(dtypes.int32)
+
+    def loop_fn(_):
+      return sparse_tensor.SparseTensor([[0], [1], [2]], [4, 5, 6],
+                                        [3])  # [0, 2, 0]
+
+    pfor = pfor_control_flow_ops.pfor(loop_fn, num_iters)
+    with self.test_session() as sess:
+      sess.run(pfor, feed_dict={num_iters: 3})
+
+  def test_sparse_result_none_stacked(self):
+    num_iters = 10
+
+    def loop_fn(_):
+      return sparse_tensor.SparseTensor([[0], [1], [2]], [4, 5, 6],
+                                        [3])  # [0, 2, 0]
+
+    pfor = pfor_control_flow_ops.pfor(loop_fn, num_iters)
+
+    indices = [[i, j] for i in range(num_iters) for j in range(3)]
+    values = [4, 5, 6] * num_iters
+    dense_shapes = [num_iters, 3]
+    # Expected result: [[4, 5, 6], [4, 5, 6], [4, 5, 6], ...]
+    manual = sparse_tensor.SparseTensor(indices, values, dense_shapes)
+    self.run_and_assert_equal(pfor, manual)
+
+  def test_sparse_result_all_stacked(self):
+    num_iters = 10
+
+    def loop_fn(i):
+      i = array_ops.expand_dims(math_ops.cast(i, dtypes.int64), 0)
+      indices = array_ops.expand_dims(i, 0)
+      return sparse_tensor.SparseTensor(indices, i, i + 1)  # [0, ..., 0, i]
+
+    # Expected result: [[0], [0, 1], [0, 0, 2], [0, 0, 0, 3], ...]
+    pfor = pfor_control_flow_ops.pfor(loop_fn, num_iters)
+    manual = sparse_tensor.SparseTensor([[i, i] for i in range(num_iters)],
+                                        list(range(num_iters)),
+                                        (num_iters, num_iters))
+    self.run_and_assert_equal(pfor, manual)
+
+  def test_sparse_result_indices_stacked(self):
+    num_iters = 10
+
+    def loop_fn(i):
+      i = array_ops.expand_dims(math_ops.cast(i, dtypes.int64), 0)
+      indices = array_ops.expand_dims(i, 0)
+      return sparse_tensor.SparseTensor(indices, [1], [num_iters])
+
+    # Expected result: identity matrix size num_iters * num_iters
+    pfor = pfor_control_flow_ops.pfor(loop_fn, num_iters)
+    manual = sparse_tensor.SparseTensor([[i, i] for i in range(num_iters)],
+                                        [1] * num_iters, (num_iters, num_iters))
+    self.run_and_assert_equal(pfor, manual)
+
+  def test_sparse_result_values_stacked(self):
+    num_iters = 10
+
+    def loop_fn(i):
+      i = array_ops.expand_dims(math_ops.cast(i, dtypes.int64), 0)
+      return sparse_tensor.SparseTensor([[0]], i, [num_iters])  # [i, 0, ..., 0]
+
+    # Expected result: [[1, 0, ...], [2, 0, ...], [3, 0, ...], ...]
+    pfor = pfor_control_flow_ops.pfor(loop_fn, num_iters)
+    manual = sparse_tensor.SparseTensor([[i, 0] for i in range(num_iters)],
+                                        list(range(num_iters)),
+                                        (num_iters, num_iters))
+    self.run_and_assert_equal(pfor, manual)
+
+  def test_sparse_result_shapes_stacked(self):
+    num_iters = 10
+
+    def loop_fn(i):
+      i = array_ops.expand_dims(math_ops.cast(i, dtypes.int64), 0)
+      return sparse_tensor.SparseTensor([[0]], [1], i + 1)  # [1, 0, ..., 0]
+
+    # Expected result: [[1, 0, 0, ...], [1, 0, 0, ...], ...]
+    pfor = pfor_control_flow_ops.pfor(loop_fn, num_iters)
+    manual = sparse_tensor.SparseTensor([[i, 0] for i in range(num_iters)],
+                                        [1] * num_iters, (num_iters, num_iters))
+    self.run_and_assert_equal(pfor, manual)
+
+  def test_sparse_result_shapes_stacked_2D(self):
+    num_iters = 10
+
+    def loop_fn(i):
+      i = array_ops.expand_dims(math_ops.cast(i + 1, dtypes.int64), 0)
+      shape = array_ops.concat([i, i], 0)
+      return sparse_tensor.SparseTensor([[0, 0]], [1], shape)  # [1, 0, ..., 0]
+
+    # Expected result: [[[1, 0, ...], [0, ..., 0], [0, ..., 0], ...], ...]
+    pfor = pfor_control_flow_ops.pfor(loop_fn, num_iters)
+    manual = sparse_tensor.SparseTensor([[i, 0, 0] for i in range(num_iters)],
+                                        [1] * num_iters,
+                                        (num_iters, num_iters, num_iters))
+    self.run_and_assert_equal(pfor, manual)
+
+
+class ParsingTest(PForTest):
+
+  def test_decode_csv(self):
+    csv_tensor = constant_op.constant([["1:2:3"], ["::"], ["7:8:9"]])
+    kwargs = {"record_defaults": [[10], [20], [30]], "field_delim": ":"}
+
+    def loop_fn(i):
+      line = array_ops.gather(csv_tensor, i)
+      return parsing_ops.decode_csv(line, **kwargs)
+
+    self._test_loop_fn(loop_fn, iters=3, loop_fn_dtypes=[dtypes.int32] * 3)
+
+  def test_parse_single_example(self):
+
+    def _int64_feature(*values):
+      return feature_pb2.Feature(int64_list=feature_pb2.Int64List(value=values))
+
+    def _bytes_feature(*values):
+      return feature_pb2.Feature(
+          bytes_list=feature_pb2.BytesList(
+              value=[v.encode("utf-8") for v in values]))
+
+    examples = constant_op.constant([
+        example_pb2.Example(
+            features=feature_pb2.Features(
+                feature={
+                    "dense_int": _int64_feature(i),
+                    "dense_str": _bytes_feature(str(i)),
+                    "sparse_int": _int64_feature(i, i * 2, i * 4, i * 8),
+                    "sparse_str": _bytes_feature(*["abc"] * i)
+                })).SerializeToString() for i in range(10)
+    ])
+
+    features = {
+        "dense_int": parsing_ops.FixedLenFeature((), dtypes.int64, 0),
+        "dense_str": parsing_ops.FixedLenFeature((), dtypes.string, ""),
+        "sparse_int": parsing_ops.VarLenFeature(dtypes.int64),
+        "sparse_str": parsing_ops.VarLenFeature(dtypes.string),
+    }
+
+    def loop_fn(i):
+      example_proto = array_ops.gather(examples, i)
+      f = parsing_ops.parse_single_example(example_proto, features)
+      return f
+
+    pfor = pfor_control_flow_ops.pfor(loop_fn, iters=10)
+    manual = parsing_ops.parse_example(examples, features)
+    self.run_and_assert_equal(pfor, manual)
+
+
+if __name__ == "__main__":
+  test.main()
diff --git a/tensorflow/python/ops/parallel_for/gradients.py b/tensorflow/python/ops/parallel_for/gradients.py
new file mode 100644
index 0000000000000000000000000000000000000000..ee3d5c9b86ed186f76e113351646b3dda153e72b
--- /dev/null
+++ b/tensorflow/python/ops/parallel_for/gradients.py
@@ -0,0 +1,126 @@
+# Copyright 2018 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""Jacobian ops."""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+from tensorflow.python.framework import ops
+from tensorflow.python.ops import array_ops
+from tensorflow.python.ops import check_ops
+from tensorflow.python.ops import gradients as gradient_ops
+from tensorflow.python.ops.parallel_for import control_flow_ops
+from tensorflow.python.util import nest
+
+
+def jacobian(output, inputs, use_pfor=True):
+  """Computes jacobian of `output` w.r.t. `inputs`.
+
+  Args:
+    output: A tensor.
+    inputs: A tensor or a nested structure of tensor objects.
+    use_pfor: If true, uses pfor for computing the jacobian. Else uses
+      tf.while_loop.
+
+  Returns:
+    A tensor or a nested strucutre of tensors with the same structure as
+    `inputs`. Each entry is the jacobian of `output` w.rt. to the corresponding
+    value in `inputs`. If output has shape [y_1, ..., y_n] and inputs_i has
+    shape [x_1, ..., x_m], the corresponding jacobian has shape
+    [y_1, ..., y_n, x_1, ..., x_m].
+  """
+  flat_inputs = nest.flatten(inputs)
+  output_shape = array_ops.shape(output)
+  output = array_ops.reshape(output, [-1])
+
+  def loop_fn(i):
+    y = array_ops.gather(output, i)
+    return gradient_ops.gradients(y, flat_inputs)
+
+  try:
+    output_size = int(output.shape[0])
+  except TypeError:
+    output_size = array_ops.shape(output)[0]
+
+  if use_pfor:
+    pfor_outputs = control_flow_ops.pfor(loop_fn, output_size)
+  else:
+    pfor_outputs = control_flow_ops.for_loop(
+        loop_fn, [output.dtype] * len(flat_inputs), output_size)
+
+  for i, out in enumerate(pfor_outputs):
+    new_shape = array_ops.concat(
+        [output_shape, array_ops.shape(out)[1:]], axis=0)
+    out = array_ops.reshape(out, new_shape)
+    pfor_outputs[i] = out
+
+  return nest.pack_sequence_as(inputs, pfor_outputs)
+
+
+def batch_jacobian(output, inp, use_pfor=True):
+  """Computes and stacks jacobians of `output[i,...]` w.r.t. `input[i,...]`.
+
+  e.g.
+  x = tf.constant([[1, 2], [3, 4]], dtype=tf.float32)
+  y = x * x
+  jacobian = batch_jacobian(y, x)
+  # => [[[2,  0], [0,  4]], [[6,  0], [0,  8]]]
+
+  Args:
+    output: A tensor with shape [b, y1, ..., y_n]. `output[i,...]` should
+      only depend on `inp[i,...]`.
+    inp: A tensor with shape [b, x1, ..., x_m]
+    use_pfor: If true, uses pfor for computing the Jacobian. Else uses a
+      tf.while_loop.
+
+  Returns:
+    A tensor `t` with shape [b, y_1, ..., y_n, x1, ..., x_m] where `t[i, ...]`
+    is the jacobian of `output[i, ...]` w.r.t. `inp[i, ...]`, i.e. stacked
+    per-example jacobians.
+
+  Raises:
+    ValueError: if first dimension of `output` and `inp` do not match.
+  """
+  output_shape = output.shape
+  if not output_shape[0].is_compatible_with(inp.shape[0]):
+    raise ValueError("Need first dimension of output shape (%s) and inp shape "
+                     "(%s) to match." % (output.shape, inp.shape))
+  if output_shape.is_fully_defined():
+    batch_size = int(output_shape[0])
+    output_row_size = output_shape.num_elements() // batch_size
+  else:
+    output_shape = array_ops.shape(output)
+    batch_size = output_shape[0]
+    output_row_size = array_ops.size(output) // batch_size
+  inp_shape = array_ops.shape(inp)
+  # Flatten output to 2-D.
+  with ops.control_dependencies(
+      [check_ops.assert_equal(batch_size, inp_shape[0])]):
+    output = array_ops.reshape(output, [batch_size, output_row_size])
+
+  def loop_fn(i):
+    y = array_ops.gather(output, i, axis=1)
+    return gradient_ops.gradients(y, inp)[0]
+
+  if use_pfor:
+    pfor_output = control_flow_ops.pfor(loop_fn, output_row_size)
+  else:
+    pfor_output = control_flow_ops.for_loop(loop_fn, output.dtype,
+                                            output_row_size)
+  pfor_output = array_ops.reshape(pfor_output,
+                                  [output_row_size, batch_size, -1])
+  output = array_ops.transpose(pfor_output, [1, 0, 2])
+  new_shape = array_ops.concat([output_shape, inp_shape[1:]], axis=0)
+  return array_ops.reshape(output, new_shape)
diff --git a/tensorflow/python/ops/parallel_for/gradients_test.py b/tensorflow/python/ops/parallel_for/gradients_test.py
new file mode 100644
index 0000000000000000000000000000000000000000..310a2154f71c29702de1d43d8fc4af931b3217eb
--- /dev/null
+++ b/tensorflow/python/ops/parallel_for/gradients_test.py
@@ -0,0 +1,568 @@
+# Copyright 2018 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""Tests for jacobian and batch_jacobian ops."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import functools
+import time
+
+import numpy as np
+
+from tensorflow.python.client import session
+from tensorflow.python.framework import constant_op
+from tensorflow.python.framework import dtypes
+from tensorflow.python.framework import errors
+from tensorflow.python.framework import ops
+from tensorflow.python.keras.engine import training as keras_training
+from tensorflow.python.layers import layers as tf_layers
+from tensorflow.python.ops import array_ops
+from tensorflow.python.ops import gradients as gradient_ops
+from tensorflow.python.ops import math_ops
+from tensorflow.python.ops import nn
+from tensorflow.python.ops import random_ops
+from tensorflow.python.ops import rnn
+from tensorflow.python.ops import rnn_cell
+from tensorflow.python.ops import variables
+from tensorflow.python.ops.losses import losses
+from tensorflow.python.ops.parallel_for import control_flow_ops
+from tensorflow.python.ops.parallel_for import gradients
+from tensorflow.python.platform import test
+from tensorflow.python.util import nest
+
+
+class FullyConnectedModel(object):
+
+  def __init__(self, activation_size, num_layers):
+    self._layers = [
+        tf_layers.Dense(activation_size, activation=nn.relu)
+        for _ in range(num_layers)
+    ]
+
+  def __call__(self, inp):
+    activation = inp
+    for layer in self._layers:
+      activation = layer(activation)
+    return activation
+
+
+def fully_connected_model_fn(batch_size, activation_size, num_layers):
+  model = FullyConnectedModel(activation_size, num_layers)
+  inp = random_ops.random_normal([batch_size, activation_size])
+  return inp, model(inp)
+
+
+def lstm_model_fn(batch_size, state_size, steps):
+  inputs = [
+      random_ops.random_normal([batch_size, state_size]) for _ in range(steps)
+  ]
+  cell = rnn_cell.BasicLSTMCell(state_size)
+  init_state = cell.zero_state(batch_size, dtypes.float32)
+  state = init_state
+  for inp in inputs:
+    _, state = cell(inp, state)
+  return init_state.c, state.c
+
+
+def dynamic_lstm_model_fn(batch_size, state_size, max_steps):
+  # We make inputs and sequence_length constant so that multiple session.run
+  # calls produce the same result.
+  inputs = constant_op.constant(
+      np.random.rand(batch_size, max_steps, state_size), dtype=dtypes.float32)
+  sequence_length = constant_op.constant(
+      np.random.randint(0, size=[batch_size], high=max_steps + 1),
+      dtype=dtypes.int32)
+
+  cell = rnn_cell.BasicLSTMCell(state_size)
+  initial_state = cell.zero_state(batch_size, dtypes.float32)
+  return inputs, rnn.dynamic_rnn(
+      cell,
+      inputs,
+      sequence_length=sequence_length,
+      initial_state=initial_state)
+
+
+def create_fc_batch_jacobian(batch_size, activation_size, num_layers):
+  inp, output = fully_connected_model_fn(batch_size, activation_size,
+                                         num_layers)
+  pfor_jacobian = gradients.batch_jacobian(output, inp, use_pfor=True)
+  while_jacobian = gradients.batch_jacobian(output, inp, use_pfor=False)
+  return pfor_jacobian, while_jacobian
+
+
+def create_lstm_batch_jacobian(batch_size, state_size, steps):
+  inp, output = lstm_model_fn(batch_size, state_size, steps)
+  pfor_jacobian = gradients.batch_jacobian(output, inp, use_pfor=True)
+  while_jacobian = gradients.batch_jacobian(output, inp, use_pfor=False)
+  return pfor_jacobian, while_jacobian
+
+
+def create_dynamic_lstm_batch_jacobian(batch_size, state_size, max_steps):
+  inp, (_, final_state) = dynamic_lstm_model_fn(batch_size, state_size,
+                                                max_steps)
+  pfor_jacobian = gradients.batch_jacobian(final_state.c, inp, use_pfor=True)
+  # Note that use_pfor=False does not work above given the current limitations
+  # on implementation of while_loop. So we statically unroll the looping in the
+  # jacobian computation.
+  while_gradients = [
+      gradient_ops.gradients(array_ops.gather(final_state.c, i, axis=1), inp)[0]
+      for i in range(state_size)
+  ]
+  return pfor_jacobian, while_gradients
+
+
+def create_lstm_batch_hessian(batch_size, state_size, steps):
+  inp, output = lstm_model_fn(batch_size, state_size, steps)
+  pfor_jacobian = gradients.batch_jacobian(output, inp, use_pfor=True)
+  pfor_jacobian = array_ops.reshape(pfor_jacobian, [batch_size, -1])
+  pfor_hessian = gradients.batch_jacobian(pfor_jacobian, inp, use_pfor=True)
+  # TODO(agarwal): using two nested while_loop doesn't seem to work here.
+  # Hence we use pfor_jacobian for computing while_hessian.
+  while_jacobian = pfor_jacobian
+  while_hessian = gradients.batch_jacobian(while_jacobian, inp, use_pfor=False)
+  return pfor_hessian, while_hessian
+
+
+def create_lstm_hessian(batch_size, state_size, steps):
+  _, output = lstm_model_fn(batch_size, state_size, steps)
+  weights = variables.trainable_variables()
+  pfor_jacobians = gradients.jacobian(output, weights, use_pfor=True)
+  pfor_hessians = [
+      gradients.jacobian(x, weights, use_pfor=True) for x in pfor_jacobians
+  ]
+  # TODO(agarwal): using two nested while_loop doesn't seem to work here.
+  # Hence we use pfor_jacobians for computing while_hessians.
+  while_jacobians = pfor_jacobians
+  while_hessians = [
+      gradients.jacobian(x, weights, use_pfor=False) for x in while_jacobians
+  ]
+  return pfor_hessians, while_hessians
+
+
+def create_fc_per_eg_grad(batch_size, activation_size, num_layers):
+  inp = random_ops.random_normal([batch_size, activation_size])
+  layers = [
+      tf_layers.Dense(activation_size, activation=nn.relu)
+      for _ in range(num_layers)
+  ]
+  projection = tf_layers.Dense(1)
+
+  def model_fn(activation):
+    for layer in layers:
+      activation = layer(activation)
+    activation = projection(activation)
+    activation = nn.l2_loss(activation)
+    return gradient_ops.gradients(activation, variables.trainable_variables())
+
+  def loop_fn(i):
+    return model_fn(array_ops.expand_dims(array_ops.gather(inp, i), 0))
+
+  pfor_outputs = control_flow_ops.pfor(loop_fn, batch_size)
+  loop_fn_dtypes = [x.dtype for x in variables.trainable_variables()]
+  while_outputs = control_flow_ops.for_loop(loop_fn, loop_fn_dtypes, batch_size)
+  return pfor_outputs, while_outputs
+
+
+def create_lstm_per_eg_grad(batch_size, state_size, steps):
+  inputs = [
+      random_ops.random_normal([batch_size, state_size]) for _ in range(steps)
+  ]
+  cell = rnn_cell.BasicLSTMCell(state_size)
+  init_state = cell.zero_state(batch_size, dtypes.float32)
+
+  def model_fn(inps, init_state):
+    state = init_state
+    for inp in inps:
+      _, state = cell(inp, state)
+    output = nn.l2_loss(state.c)
+    return gradient_ops.gradients(output, variables.trainable_variables())
+
+  def loop_fn(i):
+    loop_inputs = [
+        array_ops.expand_dims(array_ops.gather(x, i), 0) for x in inputs
+    ]
+    loop_init_state = rnn_cell.LSTMStateTuple(
+        *[array_ops.expand_dims(array_ops.gather(x, i), 0) for x in init_state])
+    return model_fn(loop_inputs, loop_init_state)
+
+  pfor_outputs = control_flow_ops.pfor(loop_fn, batch_size)
+  loop_fn_dtypes = [x.dtype for x in variables.trainable_variables()]
+  while_outputs = control_flow_ops.for_loop(loop_fn, loop_fn_dtypes, batch_size)
+  return pfor_outputs, while_outputs
+
+
+# Importing the code from tensorflow_models seems to cause errors. Hence we
+# duplicate the model definition here.
+# TODO(agarwal): Use the version in tensorflow_models/official instead.
+class Mnist(keras_training.Model):
+
+  def __init__(self, data_format):
+    """Creates a model for classifying a hand-written digit.
+
+    Args:
+      data_format: Either 'channels_first' or 'channels_last'.
+    """
+    super(Mnist, self).__init__()
+    if data_format == "channels_first":
+      self._input_shape = [-1, 1, 28, 28]
+    else:
+      assert data_format == "channels_last"
+      self._input_shape = [-1, 28, 28, 1]
+
+    self.conv1 = tf_layers.Conv2D(
+        32, 5, padding="same", data_format=data_format, activation=nn.relu)
+    self.conv2 = tf_layers.Conv2D(
+        64, 5, padding="same", data_format=data_format, activation=nn.relu)
+    self.fc1 = tf_layers.Dense(1024, activation=nn.relu)
+    self.fc2 = tf_layers.Dense(10)
+    self.dropout = tf_layers.Dropout(0.4)
+    self.max_pool2d = tf_layers.MaxPooling2D(
+        (2, 2), (2, 2), padding="same", data_format=data_format)
+
+  def __call__(self, inputs, training):
+    """Add operations to classify a batch of input images.
+
+    Args:
+      inputs: A Tensor representing a batch of input images.
+      training: A boolean. Set to True to add operations required only when
+        training the classifier.
+
+    Returns:
+      A logits Tensor with shape [, 10].
+    """
+    y = array_ops.reshape(inputs, self._input_shape)
+    y = self.conv1(y)
+    y = self.max_pool2d(y)
+    y = self.conv2(y)
+    y = self.max_pool2d(y)
+    y = tf_layers.flatten(y)
+    y = self.fc1(y)
+    y = self.dropout(y, training=training)
+    return self.fc2(y)
+
+
+def create_mnist_per_eg_grad(batch_size, data_format, training):
+  images = random_ops.random_uniform([batch_size, 28, 28])
+  sparse_labels = np.random.randint(
+      low=0, high=10, size=[batch_size]).astype(np.int32)
+  labels = np.zeros((batch_size, 10)).astype(np.float32)
+  labels[np.arange(batch_size), sparse_labels] = 1.
+  model = Mnist(data_format)
+
+  def loop_fn(i):
+    image = array_ops.gather(images, i)
+    label = array_ops.gather(labels, i)
+    logits = array_ops.reshape(model(image, training=training), [-1])
+    loss = losses.softmax_cross_entropy(
+        logits=logits, onehot_labels=label, reduction=losses.Reduction.NONE)
+    return gradient_ops.gradients(loss, variables.trainable_variables())
+
+  pfor_outputs = control_flow_ops.pfor(loop_fn, batch_size)
+  while_outputs = control_flow_ops.for_loop(
+      loop_fn, [dtypes.float32] * len(variables.trainable_variables()),
+      batch_size)
+  return pfor_outputs, while_outputs
+
+
+def create_mnist_per_eg_jacobian(batch_size, data_format, training):
+  images = random_ops.random_uniform([batch_size, 28, 28])
+  model = Mnist(data_format)
+
+  def loop_fn(i, use_pfor):
+    image = array_ops.gather(images, i)
+    logits = array_ops.reshape(model(image, training=training), [-1])
+    return gradients.jacobian(
+        logits, variables.trainable_variables(), use_pfor=use_pfor)
+
+  pfor_outputs = control_flow_ops.pfor(
+      functools.partial(loop_fn, use_pfor=True),
+      batch_size)
+  while_outputs = control_flow_ops.for_loop(
+      functools.partial(loop_fn, use_pfor=False),
+      [dtypes.float32] * len(variables.trainable_variables()), batch_size)
+  return pfor_outputs, while_outputs
+
+
+def create_fc_per_eg_jacobians(batch_size, activation_size, num_layers):
+  model = FullyConnectedModel(activation_size=activation_size,
+                              num_layers=num_layers)
+  inp = random_ops.random_normal([batch_size, activation_size])
+  output = model(inp)
+  jacobians = gradients.jacobian(output, variables.trainable_variables())
+
+  def loop_fn(i, use_pfor):
+    inp_i = array_ops.expand_dims(array_ops.gather(inp, i), 0)
+    output = array_ops.reshape(model(inp_i), [-1])
+    return gradients.jacobian(
+        output, variables.trainable_variables(), use_pfor=use_pfor)
+
+  per_eg_jacobians_pfor = control_flow_ops.pfor(
+      functools.partial(loop_fn, use_pfor=True),
+      batch_size)
+  per_eg_jacobians_while = control_flow_ops.for_loop(
+      functools.partial(loop_fn, use_pfor=False),
+      [dtypes.float32] * len(variables.trainable_variables()), batch_size)
+  return jacobians, per_eg_jacobians_pfor, per_eg_jacobians_while
+
+
+class GradientsTest(test.TestCase):
+
+  def run_and_assert_equal(self, targets1, targets2, atol=1e-4, rtol=1e-4):
+    targets1 = nest.flatten(targets1)
+    targets2 = nest.flatten(targets2)
+    assert len(targets1) == len(targets2)
+    init = variables.global_variables_initializer()
+    self.evaluate(init)
+    outputs = self.evaluate(targets1 + targets2)
+    n = len(outputs) // 2
+    for i in range(n):
+      self.assertAllClose(outputs[i], outputs[i + n], rtol=rtol, atol=atol)
+
+  def test_jacobian_fixed_shape(self):
+    x = random_ops.random_uniform([2, 2])
+    y = math_ops.matmul(x, x, transpose_a=True)
+    jacobian_pfor = gradients.jacobian(y, x, use_pfor=True)
+    jacobian_while = gradients.jacobian(y, x, use_pfor=False)
+    answer = ops.convert_to_tensor([[
+        gradient_ops.gradients(y[0][0], x)[0],
+        gradient_ops.gradients(y[0][1], x)[0]
+    ], [
+        gradient_ops.gradients(y[1][0], x)[0],
+        gradient_ops.gradients(y[1][1], x)[0]
+    ]])
+    self.run_and_assert_equal(answer, jacobian_pfor)
+    self.run_and_assert_equal(answer, jacobian_while)
+
+  def test_jacobian_unknown_shape(self):
+    with self.test_session() as sess:
+      x = array_ops.placeholder(dtypes.float32, shape=[None, None])
+      y = math_ops.matmul(x, x, transpose_a=True)
+      jacobian_pfor = gradients.jacobian(y, x, use_pfor=True)
+      jacobian_while = gradients.jacobian(y, x, use_pfor=False)
+      answer = ops.convert_to_tensor([[
+          gradient_ops.gradients(y[0][0], x)[0],
+          gradient_ops.gradients(y[0][1], x)[0]
+      ], [
+          gradient_ops.gradients(y[1][0], x)[0],
+          gradient_ops.gradients(y[1][1], x)[0]
+      ]])
+      ans, pfor_value, while_value = sess.run(
+          [answer, jacobian_pfor, jacobian_while],
+          feed_dict={x: [[1, 2], [3, 4]]})
+      self.assertAllClose(ans, pfor_value)
+      self.assertAllClose(ans, while_value)
+
+  def test_batch_jacobian_bad_shapes(self):
+    x = random_ops.random_uniform([2, 2])
+    y = random_ops.random_uniform([3, 2])
+    with self.assertRaisesRegexp(ValueError, "Need first dimension of output"):
+      gradients.batch_jacobian(y, x, use_pfor=True)
+
+  def test_batch_jacobian_bad_unknown_shapes(self):
+    with self.test_session() as sess:
+      x = array_ops.placeholder(dtypes.float32)
+      y = array_ops.concat([x, x], axis=0)
+      jacobian = gradients.batch_jacobian(y, x)
+      with self.assertRaisesRegexp(errors.InvalidArgumentError,
+                                   "assertion failed"):
+        sess.run(jacobian, feed_dict={x: [[1, 2], [3, 4]]})
+
+  def test_batch_jacobian_fixed_shape(self):
+    x = random_ops.random_uniform([2, 3, 5])
+    y = x * x
+    batch_jacobian_pfor = gradients.batch_jacobian(y, x, use_pfor=True)
+    batch_jacobian_while = gradients.batch_jacobian(y, x, use_pfor=False)
+    two_x = 2 * x
+    answer = array_ops.stack(
+        [array_ops.diag(two_x[0]),
+         array_ops.diag(two_x[1])])
+    self.run_and_assert_equal(answer, batch_jacobian_pfor)
+    self.run_and_assert_equal(answer, batch_jacobian_while)
+
+  def test_batch_jacobian_unknown_shape(self):
+    with self.test_session() as sess:
+      x = array_ops.placeholder(dtypes.float32)
+      y = x * x
+      batch_jacobian_pfor = gradients.batch_jacobian(y, x, use_pfor=True)
+      batch_jacobian_while = gradients.batch_jacobian(y, x, use_pfor=False)
+      two_x = 2 * x
+      answer = array_ops.stack(
+          [array_ops.diag(two_x[0]),
+           array_ops.diag(two_x[1])])
+      ans, pfor_value, while_value = sess.run(
+          [answer, batch_jacobian_pfor, batch_jacobian_while],
+          feed_dict={x: [[1, 2], [3, 4]]})
+      self.assertAllClose(ans, pfor_value)
+      self.assertAllClose(ans, while_value)
+
+  def test_fc_batch_jacobian(self):
+    pfor_jacobian, while_jacobian = create_fc_batch_jacobian(8, 4, 2)
+    self.run_and_assert_equal(pfor_jacobian, while_jacobian)
+
+  def test_lstm_batch_jacobian(self):
+    pfor_jacobian, while_jacobian = create_lstm_batch_jacobian(8, 4, 2)
+    self.run_and_assert_equal(pfor_jacobian, while_jacobian)
+
+  def test_dynamic_lstm_batch_jacobian(self):
+    pfor_jacobian, while_gradients = create_dynamic_lstm_batch_jacobian(8, 4, 3)
+    with session.Session() as sess:
+      init = variables.global_variables_initializer()
+      sess.run(init)
+      pfor = sess.run(pfor_jacobian)
+      for i in range(4):
+        while_i = sess.run(while_gradients[i])
+        self.assertAllClose(while_i, pfor[:, i, ...])
+
+  def test_lstm_hessian(self):
+    pfor_hessian, while_hessian = create_lstm_hessian(2, 2, 2)
+    self.run_and_assert_equal(pfor_hessian, while_hessian)
+
+  def test_lstm_batch_hessian(self):
+    pfor_hessian, while_hessian = create_lstm_batch_hessian(2, 2, 2)
+    self.run_and_assert_equal(pfor_hessian, while_hessian)
+
+  def test_fc_per_eg_grad(self):
+    pfor_outputs, while_outputs = create_fc_per_eg_grad(8, 4, 2)
+    self.run_and_assert_equal(pfor_outputs, while_outputs)
+
+  def test_lstm_per_eg_grad(self):
+    pfor_outputs, while_outputs = create_lstm_per_eg_grad(8, 4, 2)
+    self.run_and_assert_equal(pfor_outputs, while_outputs)
+
+  def test_mnist_per_eg_grad(self):
+    data_format = ("channels_first"
+                   if test.is_gpu_available() else "channels_last")
+    # Note that we we are setting training=False here so that dropout produces
+    # the same result with pfor and with while_loop.
+    pfor_outputs, while_outputs = create_mnist_per_eg_grad(
+        4, data_format, training=False)
+    self.run_and_assert_equal(pfor_outputs, while_outputs, rtol=1e-3)
+
+  def test_mnist_per_eg_jacobian(self):
+    data_format = ("channels_first"
+                   if test.is_gpu_available() else "channels_last")
+    # Note that we we are setting training=False here so that dropout produces
+    # the same result with pfor and with while_loop.
+    pfor_outputs, while_outputs = create_mnist_per_eg_jacobian(
+        2, data_format, training=False)
+    self.run_and_assert_equal(pfor_outputs, while_outputs, rtol=1e-3)
+
+  def test_fc_jacobian(self):
+    jacobians, per_eg_jacobians_pfor, per_eg_jacobians_while = (
+        create_fc_per_eg_jacobians(batch_size=8,
+                                   activation_size=4,
+                                   num_layers=2))
+    self.run_and_assert_equal(jacobians, per_eg_jacobians_pfor,
+                              rtol=2e-3, atol=1e-3)
+    self.run_and_assert_equal(jacobians, per_eg_jacobians_while,
+                              rtol=2e-3, atol=1e-3)
+
+
+class GradientsBenchmarks(test.Benchmark):
+
+  def _run(self, targets, iters, name=None):
+
+    def _done(t):
+      # Note that we don't use tf.control_dependencies since that will not make
+      # sure that the computation on GPU has actually finished. So we fetch the
+      # first element of the output, and assume that this will not be called on
+      # empty tensors.
+      return array_ops.gather(array_ops.reshape(t, [-1]), 0)
+
+    targets = [_done(x) for x in nest.flatten(targets)]
+    sess = session.Session()
+    with sess:
+      init = variables.global_variables_initializer()
+      sess.run(init)
+      sess.run(targets)
+      begin = time.time()
+      for _ in range(iters):
+        sess.run(targets)
+      end = time.time()
+    avg_time_ms = 1000 * (end - begin) / iters
+    self.report_benchmark(iters=iters, wall_time=avg_time_ms, name=name)
+    return avg_time_ms
+
+  def benchmark_fc_batch_jacobian(self):
+    with ops.Graph().as_default():
+      pfor_jacobian, while_jacobian = create_fc_batch_jacobian(100, 32, 20)
+      self._run(pfor_jacobian, 100, name="fc_batch_jacobian_pfor")
+      self._run(while_jacobian, 20, name="fc_batch_jacobian_while")
+
+  def benchmark_lstm_batch_jacobian(self):
+    with ops.Graph().as_default():
+      pfor_jacobian, while_jacobian = create_lstm_batch_jacobian(100, 32, 8)
+      self._run(pfor_jacobian, 100, name="lstm_batch_jacobian_pfor")
+      self._run(while_jacobian, 20, name="lstm_batch_jacobian_while")
+
+  def benchmark_lstm_hessian(self):
+    with ops.Graph().as_default():
+      pfor_hessian, while_hessian = create_lstm_hessian(2, 2, 10)
+      self._run(pfor_hessian, 20, name="lstm_hessian_pfor")
+      self._run(while_hessian, 3, name="lstm_hessian_while_pfor")
+
+  def benchmark_lstm_batch_hessian(self):
+    with ops.Graph().as_default():
+      pfor_hessian, while_hessian = create_lstm_batch_hessian(4, 4, 10)
+      self._run(pfor_hessian, 100, name="lstm_batch_hessian_pfor")
+      self._run(while_hessian, 20, name="lstm_batch_hessian_while_pfor")
+
+  def benchmark_fc_per_eg_grad(self):
+    with ops.Graph().as_default():
+      pfor_outputs, while_outputs = create_fc_per_eg_grad(100, 32, 3)
+      self._run(pfor_outputs, 100, name="fc_per_eg_grad_pfor")
+      self._run(while_outputs, 20, name="fc_per_eg_grad_while")
+
+  def benchmark_lstm_per_eg_grad(self):
+    with ops.Graph().as_default():
+      pfor_outputs, while_outputs = create_lstm_per_eg_grad(100, 32, 8)
+      self._run(pfor_outputs, 100, name="lstm_per_eg_grad_pfor")
+      self._run(while_outputs, 20, name="lstm_per_eg_grad_while")
+
+  def benchmark_mnist_per_eg_grad(self):
+    with ops.Graph().as_default():
+      data_format = ("channels_first"
+                     if test.is_gpu_available() else "channels_last")
+      pfor_outputs, while_outputs = create_mnist_per_eg_grad(
+          128, data_format, training=True)
+      self._run(pfor_outputs, 20, name="mnist_per_eg_grad_pfor")
+      self._run(while_outputs, 20, name="mnist_per_eg_grad_while")
+
+  def benchmark_mnist_per_eg_jacobian(self):
+    with ops.Graph().as_default():
+      data_format = ("channels_first"
+                     if test.is_gpu_available() else "channels_last")
+      pfor_outputs, while_outputs = create_mnist_per_eg_jacobian(
+          16, data_format, training=True)
+      self._run(pfor_outputs, 20, name="mnist_per_eg_jacobian_pfor")
+      self._run(while_outputs, 20, name="mnist_per_eg_jacobian_while")
+
+  def benchmark_fc_per_eg_jacobian(self):
+    with ops.Graph().as_default():
+      jacobians, per_eg_jacobians_pfor, per_eg_jacobians_while = (
+          create_fc_per_eg_jacobians(batch_size=128,
+                                     activation_size=32,
+                                     num_layers=3))
+      self._run(jacobians, 30, name="fc_jacobians_pfor")
+      self._run(per_eg_jacobians_pfor, 100,
+                name="fc_per_eg_jacobians_pfor")
+      self._run(per_eg_jacobians_while, 10,
+                name="fc_per_eg_jacobians_while")
+
+
+if __name__ == "__main__":
+  test.main()
diff --git a/tensorflow/python/ops/parallel_for/pfor.py b/tensorflow/python/ops/parallel_for/pfor.py
new file mode 100644
index 0000000000000000000000000000000000000000..77ec3bc0d40ecba11c1624af1ad4be0578b5e4f7
--- /dev/null
+++ b/tensorflow/python/ops/parallel_for/pfor.py
@@ -0,0 +1,2552 @@
+# Copyright 2018 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""Compiled parallel-for loop."""
+# pylint: disable=missing-docstring
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import collections
+
+from absl import flags
+
+from tensorflow.python.framework import constant_op
+from tensorflow.python.framework import dtypes
+from tensorflow.python.framework import ops
+from tensorflow.python.framework import sparse_tensor
+from tensorflow.python.framework import tensor_shape
+from tensorflow.python.framework import tensor_util
+from tensorflow.python.ops import array_ops
+from tensorflow.python.ops import check_ops
+from tensorflow.python.ops import control_flow_ops
+from tensorflow.python.ops import data_flow_ops
+from tensorflow.python.ops import functional_ops
+from tensorflow.python.ops import gen_parsing_ops
+from tensorflow.python.ops import gen_sparse_ops
+from tensorflow.python.ops import math_ops
+from tensorflow.python.ops import nn_ops
+from tensorflow.python.ops import parsing_ops
+from tensorflow.python.ops import sparse_ops
+from tensorflow.python.ops import tensor_array_ops
+from tensorflow.python.platform import tf_logging as logging
+from tensorflow.python.util import nest
+
+flags.DEFINE_bool(
+    "op_conversion_fallback_to_while_loop", False,
+    "If true, falls back to using a while loop for ops for "
+    "which a converter is not defined.")
+
+
+def _stack(t, length):
+  """stacks `t` `length` times."""
+  ones = array_ops.ones_like(array_ops.shape(t))
+  multiples = array_ops.concat([length, ones], 0)
+  t = array_ops.tile(array_ops.expand_dims(t, 0), multiples)
+  return wrap(t, True)
+
+
+# The following stateful ops can be safely called once, and with the same
+# signature as the unconverted version, if their inputs are loop invariant.
+# TODO(agarwal): implement a strategy for converting Variable reads/writes. The
+# plan is to map each read/write in the loop_fn to a corresponding merged
+# read/write in the converted graph. Writes need to be mergeable (e.g.
+# AssignAdd) to be used in `pfor`. Given a certain read/write order in the
+# loop_fn, doing a one-to-one conversion will simulate executing such
+# instructions in lock-step across all iterations.
+passthrough_stateful_ops = set([
+    "VariableV2",
+    "VarHandleOp",
+    "ReadVariableOp",
+    "StackV2",
+    "TensorArrayWriteV3",
+    "TensorArrayReadV3",
+    "TensorArraySizeV3",
+])
+
+
+def _is_stateful_pfor_op(op):
+  if isinstance(op, WhileOp):
+    return op.is_stateful
+  if op.type == "Const":
+    # Const didn't have an op_def.
+    return False
+  if op.type in passthrough_stateful_ops:
+    return False
+  assert hasattr(op, "op_def") and op.op_def is not None, op
+  return op.op_def.is_stateful
+
+
+# pylint: disable=protected-access
+class WhileOp(object):
+  """Object for storing state for converting the outputs of a while_loop."""
+
+  def __init__(self, exit_node, pfor_ops):
+    """Initializer.
+
+    Args:
+      exit_node: A tensor output from the while_loop.
+      pfor_ops: list of ops inside the current pfor loop.
+    """
+    self._pfor_ops = set(pfor_ops)
+    self._pfor_op_ids = set([x._id for x in pfor_ops])
+    assert isinstance(exit_node, ops.Tensor)
+    self._while_context = exit_node.op._get_control_flow_context()
+    assert isinstance(self._while_context, control_flow_ops.WhileContext)
+    self._context_name = self._while_context.name
+    self._condition = self._while_context.pivot.op.inputs[0]
+    # Parts of an external while_loop could be created inside a pfor loop.
+    # However for the purpose here, we declare such loops to be external. Also
+    # note that we check if the condition was created inside or outside to
+    # determine if the while_loop was first created inside or outside.
+    # TODO(agarwal): check that the Enter and Exit of this loop are unstacked.
+    self._is_inside_loop = self.op_is_inside_loop(self._condition.op)
+    if self._is_inside_loop:
+      for e in self._while_context.loop_exits:
+        assert self.op_is_inside_loop(e.op)
+
+    # Note the code below tries to reverse engineer an existing while_loop graph
+    # by assuming the following pattern of nodes.
+    #
+    #          NextIteration <---- Body <--- Enter
+    #              |                ^
+    #              V             ___| Y
+    #    Enter -> Merge -> Switch___
+    #                       ^       | N
+    #                       |       V
+    #                  LoopCond    Exit
+
+    # Node that elements in the list below correspond one-to-one with each
+    # other. i.e. these lists are the same size, and the i_th entry corresponds
+    # to different Operations/Tensors of a single cycle as illustrated above.
+    # List of Switch ops (ops.Operation) that feed into an Exit Node.
+    self._exit_switches = []
+    # List of inputs (ops.Tensor) to NextIteration.
+    self._body_outputs = []
+    # List of list of control inputs of the NextIteration nodes.
+    self._next_iter_control_inputs = []
+    # List of Merge ops (ops.Operation).
+    self._enter_merges = []
+    # List of output (ops.Tensor) of Exit nodes.
+    self._outputs = []
+
+    # List of Enter Tensors.
+    # There are two types of Enter nodes:
+    # - The Enter nodes that are used in the `loop_vars` argument to
+    # `while_loop` (see
+    # https://www.tensorflow.org/api_docs/python/tf/while_loop). We collect
+    # these Enter nodes immediately below by tracing backwards from the Exit
+    # nodes via Exit <- Switch <- Merge <- Enter. You can see this chain in the
+    # diagram above. This allows us to have a 1:1 correspondence between the
+    # self._outputs and the first elements in self._enters.
+    # - The Enter nodes that are used only by the body. They don't appear in the
+    # `loop_vars` and are not returned from the `while_loop`. In Python code,
+    # they are usually captured by the body lambda. We collect them below by
+    # iterating over all the ops in the graph. They are appended to the end of
+    # self._enters or self._direct_enters, and don't correspond to any outputs
+    # in self._outputs. Note that we keep the resource/variant Enter nodes in
+    # self._direct_enters and the constructed while_loop's body uses them
+    # directly as opposed to passing them as loop variables. This is done
+    # because the while_body cannot partition the resource/variant Tensors, so
+    # it has to leave them unchanged.
+    self._enters = []
+    self._direct_enters = []
+
+    for e in self._while_context.loop_exits:
+      self._outputs.append(e.op.outputs[0])
+      switch = e.op.inputs[0].op
+      assert switch.type == "Switch", switch
+      self._exit_switches.append(switch)
+      merge = switch.inputs[0].op
+      assert merge.type == "Merge", merge
+      self._enter_merges.append(merge)
+      enter = merge.inputs[0].op
+      assert enter.type == "Enter", enter
+      self._enters.append(enter.outputs[0])
+      next_iter = merge.inputs[1].op
+      assert next_iter.type == "NextIteration", next_iter
+      self._body_outputs.append(next_iter.inputs[0])
+      self._next_iter_control_inputs.append(next_iter.control_inputs)
+
+    # Collect all the Enter nodes that are not part of `loop_vars`, the second
+    # category described above.
+    # Also track whether the loop body has any stateful ops.
+    self._is_stateful = False
+    for op in ops.get_default_graph().get_operations():
+      # TODO(agarwal): make sure this works with nested case.
+      control_flow_context = op._get_control_flow_context()
+      if control_flow_context is None:
+        continue
+      if control_flow_context.name == self._context_name:
+        self._is_stateful |= _is_stateful_pfor_op(op)
+        if op.type == "Enter":
+          output = op.outputs[0]
+          if output not in self._enters:
+            if output.dtype in (dtypes.resource, dtypes.variant):
+              if output not in self._direct_enters:
+                self._direct_enters.append(output)
+            else:
+              self._enters.append(output)
+
+  def __str__(self):
+    """String representation."""
+    return "while_loop(%s)" % self.name
+
+  @property
+  def inputs(self):
+    """Input to all the Enter nodes."""
+    return [x.op.inputs[0] for x in self._enters + self._direct_enters]
+
+  @property
+  def control_inputs(self):
+    """Control input to all the Enter nodes."""
+    control_inputs = []
+    for x in self._enters + self._direct_enters:
+      control_inputs.extend(x.op.control_inputs)
+    return control_inputs
+
+  @property
+  def outputs(self):
+    """Outputs of all the Exit nodes."""
+    return self._outputs
+
+  @property
+  def name(self):
+    """Context name for the while loop."""
+    return self._context_name
+
+  @property
+  def is_inside_loop(self):
+    """Returns true if the while_loop was created inside the pfor."""
+    return self._is_inside_loop
+
+  def op_is_inside_loop(self, op):
+    """True if op was created inside the pfor loop body."""
+    assert isinstance(op, ops.Operation)
+    # Note that we use self._pfor_op_ids for the check and not self._pfor_ops
+    # since it appears there tensorflow API could return different python
+    # objects representing the same Operation node.
+    return op._id in self._pfor_op_ids
+
+  @property
+  def is_stateful(self):
+    return self._is_stateful
+
+  @property
+  def pfor_converter(self):
+    """Return a converter for the while loop."""
+    return self
+
+  def _init_pfor(self, parent_pfor, indices, cond_stacked, inputs,
+                 inputs_stacked):
+    """Create a PFor object for converting parts of the while_loop.
+
+    Args:
+      parent_pfor: PFor object being used for converting the while_loop.
+      indices: int32 Tensor of ids for the iterations that are still active
+        (i.e. did not exit the while_loop).
+      cond_stacked: True if the while_loop condition is stacked.
+      inputs: list of input Tensors corresponding 1-to-1 with self._enters. Note
+        that these Tensors are a subset of the loop variables for the generated
+        while_loop.
+      inputs_stacked: List of booleans corresponding 1-to-1 with `inputs`,
+        indicating if the value is stacked or not.
+
+    Returns:
+      A PFor instance. The instance is initialized by adding conversion mappings
+        of nodes that will be external to the conversion that the returned
+        instance will be used for. e.g. Enter nodes as well as Merge and Switch
+        outputs are mapped to converted values.
+    """
+    num_outputs = len(self._outputs)
+    assert len(inputs) == len(self._enters)
+    assert len(inputs_stacked) == len(self._enters)
+    loop_var = parent_pfor.loop_var
+    loop_len = array_ops.size(indices)
+    pfor = PFor(
+        loop_var,
+        loop_len,
+        pfor_ops=self._pfor_ops,
+        all_indices=indices,
+        all_indices_partitioned=cond_stacked)
+    # Map all inputs of Enter nodes in self._direct_enters to their converted
+    # values.
+    for enter in self._direct_enters:
+      enter_input = enter.op.inputs[0]
+      converted_enter, stacked, is_sparse_stacked = parent_pfor._convert_helper(
+          enter_input)
+      # Since these are resources / variants, they should be unstacked.
+      assert not stacked and not is_sparse_stacked, (enter, converted_enter)
+      pfor._add_conversion(enter, wrap(converted_enter, False))
+
+    # Map all Enter nodes to the inputs.
+    for enter, inp, stacked in zip(self._enters, inputs, inputs_stacked):
+      pfor._add_conversion(enter, wrap(inp, stacked))
+    # Map outputs of Switch and Merge.
+    for i in range(num_outputs):
+      wrapped_inp = wrap(inputs[i], inputs_stacked[i])
+      merge = self._enter_merges[i]
+      pfor._add_conversion(merge.outputs[0], wrapped_inp)
+      # Note that second output of Merge is typically not used, except possibly
+      # as a control dependency. To avoid trying to output the correct value, we
+      # employ a hack here. We output a dummy invalid value with an incorrect
+      # dtype. This will allow control dependency to work but if using it as an
+      # input, it should typically lead to errors during graph construction due
+      # to dtype mismatch.
+      # TODO(agarwal): Check in the original graph to see if there are any
+      # consumers of this Tensor that use it as an input.
+      pfor._add_conversion(merge.outputs[1],
+                           wrap(constant_op.constant(-1.0), False))
+      switch = self._exit_switches[i]
+      # Don't need to worry about switch.output[0] which will feed to Exit node.
+      pfor._add_conversion(switch.outputs[1], wrapped_inp)
+    return pfor
+
+  def _convert_enter(self, parent_pfor, enter):
+    """Converts an Enter node."""
+    inp, stacked, _ = parent_pfor._convert_helper(enter.op.inputs[0])
+    control_inputs = [
+        parent_pfor._convert_helper(x).t for x in enter.op.control_inputs
+    ]
+    if control_inputs:
+      with ops.control_dependencies(control_inputs):
+        inp = array_ops.identity(inp)
+    return inp, stacked
+
+  def _maybe_stacked(self, cache, inp):
+    """Heuristic to figue out if the coverting inp leads to a stacked value.
+
+
+    Args:
+      cache: map from Tensor to boolean indicating stacked/unstacked.
+      inp: input Tensor.
+
+    Returns:
+      True if `inp` could get stacked. If the function returns False, the
+      converted value should be guaranteed to be unstacked. If returning True,
+      it may or may not be stacked.
+    """
+    if inp in cache:
+      return cache[inp]
+    if not self.op_is_inside_loop(inp.op):
+      return False
+    op = inp.op
+    output = False
+    if op.type in [
+        "Shape",
+        "Rank"
+        "ShapeN",
+        "ZerosLike",
+        "TensorArrayV3",
+        "TensorArraySizeV3",
+    ]:
+      output = False
+    elif _is_stateful_pfor_op(op):
+      # This may be fairly aggressive.
+      output = True
+    elif op.type == "Exit":
+      # This may be fairly aggressive.
+      output = True
+    else:
+      for t in op.inputs:
+        if self._maybe_stacked(cache, t):
+          output = True
+          break
+    cache[inp] = output
+    return output
+
+  def _create_init_values(self, pfor_input):
+    """Create arguments passed to converted while_loop."""
+    with ops.name_scope("while_init"):
+      loop_len_vector = pfor_input.pfor.loop_len_vector
+      loop_len = loop_len_vector[0]
+      num_outputs = len(self._outputs)
+
+      inputs = []
+      maybe_stacked_cache = {}
+      # Convert all the Enters. Need to do this before checking for stacking
+      # below.
+      for i, enter in enumerate(self._enters):
+        inp, stacked = self._convert_enter(pfor_input.pfor, enter)
+        inputs.append(inp)
+        maybe_stacked_cache[enter] = stacked
+        # Since this enter node is part of the `loop_vars`, it corresponds to an
+        # output and its preceding switch. We mark this switch's output the same
+        # stackness, to act at the base case for the logic below. Below, we will
+        # be going through the body figuring out which inputs might need to be
+        # stacked and which inputs can safely remain unstacked.
+        if i < num_outputs:
+          maybe_stacked_cache[self._exit_switches[i].outputs[1]] = stacked
+
+      # Shape invariants for init_values corresponding to self._enters.
+      input_shape_invariants = []
+      # TensorArrays for outputs of converted while loop
+      output_tas = []
+      # Shape invariants for output TensorArrays.
+      ta_shape_invariants = []
+      # List of booleans indicating stackness of inputs, i.e. tensors
+      # corresponding to self._enters.
+      inputs_stacked = []
+      for i, inp in enumerate(inputs):
+        enter = self._enters[i]
+        inp_stacked = self._maybe_stacked(maybe_stacked_cache, enter)
+        # Note that even when an input is unstacked, the body could make it
+        # stacked. we use a heuristic below to figure out if body may be making
+        # it stacked.
+        if i < num_outputs:
+          body_output = self._body_outputs[i]
+          if enter.op in self._pfor_ops:
+            body_output_stacked = self._maybe_stacked(maybe_stacked_cache,
+                                                      body_output)
+          else:
+            # If constructed outside of pfor loop, then the output would not be
+            # stacked.
+            body_output_stacked = False
+          if body_output_stacked and not inp_stacked:
+            inp = _stack(inp, loop_len_vector).t
+            inputs[i] = inp
+            inp_stacked = True
+          # TODO(agarwal): other attributes for the TensorArray ?
+          output_tas.append(tensor_array_ops.TensorArray(inp.dtype, loop_len))
+          ta_shape_invariants.append(tensor_shape.TensorShape(None))
+
+        inputs_stacked.append(inp_stacked)
+        input_shape_invariants.append(tensor_shape.TensorShape(None))
+
+      # See documentation for __call__ for the structure of init_values.
+      init_values = [True, pfor_input.pfor.all_indices] + inputs + output_tas
+      # TODO(agarwal): try stricter shape invariants
+      shape_invariants = (
+          [tensor_shape.TensorShape(None),
+           tensor_shape.TensorShape(None)
+          ] + input_shape_invariants + ta_shape_invariants)
+
+      return init_values, inputs_stacked, shape_invariants
+
+  def _process_cond_unstacked(self, conditions, indices, inputs, output_tas):
+    """Handles case when condition is unstacked.
+
+    Note that all iterations end together. So we don't need to partition the
+    inputs. When all iterations are done, we write the inputs to the
+    TensorArrays. Note that we only write to index 0 of output_tas. Since all
+    iterations end together, they can all be output together.
+    """
+    not_all_done = array_ops.reshape(conditions, [])
+    new_output_tas = []
+    # pylint: disable=cell-var-from-loop
+    for i, out_ta in enumerate(output_tas):
+      inp = inputs[i]
+      new_output_tas.append(
+          control_flow_ops.cond(not_all_done,
+                                lambda: out_ta,
+                                lambda: out_ta.write(0, inp)))
+    # pylint: enable=cell-var-from-loop
+    return not_all_done, indices, inputs, new_output_tas
+
+  def _process_cond_stacked(self, conditions, indices, inputs, inputs_stacked,
+                            output_tas):
+    num_outputs = len(self._outputs)
+    # Compute if all iterations are done.
+    not_all_done = math_ops.reduce_any(conditions)
+    conditions_int = math_ops.cast(conditions, dtypes.int32)
+    # Partition the indices.
+    done_indices, new_indices = data_flow_ops.dynamic_partition(
+        indices, conditions_int, 2)
+
+    new_inputs = []
+    new_output_tas = []
+    for i, (inp, stacked) in enumerate(zip(inputs, inputs_stacked)):
+      # Partition the inputs.
+      if stacked:
+        done_inp, new_inp = data_flow_ops.dynamic_partition(
+            inp, conditions_int, 2)
+      else:
+        # TODO(agarwal): avoid this stacking. See TODO earlier in
+        # _process_cond_unstacked.
+        done_inp = _stack(inp, [array_ops.size(done_indices)]).t
+        new_inp = inp
+      new_inputs.append(new_inp)
+      # For iterations that are done, write them to TensorArrays.
+      if i < num_outputs:
+        out_ta = output_tas[i]
+        # Note that done_indices can be empty. done_inp should also be empty in
+        # that case.
+        new_output_tas.append(out_ta.scatter(done_indices, done_inp))
+    return not_all_done, new_indices, new_inputs, new_output_tas
+
+  def _process_body(self, pfor_input, inputs_stacked,
+                    new_indices, cond_stacked, new_inputs,
+                    not_all_done):
+    """Convert the body function."""
+
+    def true_fn(control_inputs, body_pfor, body_output, stacked):
+      """Converts the body function for all but last iteration.
+
+      This essentially converts body_output. Additionally, it needs to handle
+      any control dependencies on the NextIteration node. So it creates another
+      Identity node with the converted dependencies.
+      """
+      converted_control_inp = []
+      for x in control_inputs:
+        for t in x.outputs:
+          converted_control_inp.append(body_pfor._convert_helper(t).t)
+      if stacked:
+        # Note convert always does the stacking.
+        output = body_pfor.convert(body_output)
+      else:
+        output, convert_stacked, _ = body_pfor._convert_helper(body_output)
+        assert convert_stacked == stacked, body_output
+      with ops.control_dependencies(converted_control_inp):
+        return array_ops.identity(output)
+
+    body_pfor = self._init_pfor(pfor_input.pfor, new_indices,
+                                cond_stacked, new_inputs,
+                                inputs_stacked)
+    new_outputs = []
+
+    for i, (body_output, stacked) in enumerate(
+        zip(self._body_outputs, inputs_stacked)):
+      control_inp = self._next_iter_control_inputs[i]
+      out_dtype = body_output.dtype
+      # Note that we want to run the body only if not all pfor iterations are
+      # done. If all are done, we return empty tensors since these values will
+      # not be used. Notice that the value returned by the loop is based on
+      # TensorArrays and not directly on these returned values.
+      # pylint: disable=cell-var-from-loop
+      new_output = control_flow_ops.cond(
+          not_all_done,
+          lambda: true_fn(control_inp, body_pfor, body_output, stacked),
+          lambda: constant_op.constant([], dtype=out_dtype))
+      # pylint: enable=cell-var-from-loop
+      new_outputs.append(new_output)
+    return new_outputs
+
+  def __call__(self, pfor_input):
+    """Converter for the while_loop.
+
+    The conversion of a while_loop is another while_loop.
+
+    The arguments to this converted while_loop are as follows:
+    not_all_done: Boolean scalar Tensor indicating if all the pfor iterations
+      are done.
+    indices: int32 1-D Tensor storing the id of the iterations that are not
+      done.
+    args: Remaining arguments. These can be divided into 3 categories:
+      - First set of arguments are the tensors that correspond to the initial
+        elements of self._enters. The elements that appear in original while
+        loop's `loop_vars`.
+      - The second set of arguments are the tensors that correspond to the
+        remaining elements of self._enters. These are the tensors that directly
+        enter the original while loop body.
+       - Finally, the last set of arguments are TensorArrays. These TensorArrays
+         correspond to the outputs of the original while_loop, i.e. to the
+         elements in self._outputs. Each TensorArray has `PFor.loop_len`
+         elements, i.e. the number of pfor iterations. At the end, the i'th
+         element of each TensorArray will contain the output computed by the
+         i'th iteration of pfor. Note that elements can be written into these
+         tensors arrays in any order, depending on when the corresponding pfor
+         iteration is done.
+      If the original while_loop had `k` tensors in its `loop_vars` and its body
+      directly captured `m` tensors, the `args` will contain `2 * k + m` values.
+
+    In each iteration, the while_loop body recomputes the condition for all
+    active pfor iterations to see which of them are now done. It then partitions
+    all the inputs and passes them along to the converted body. Values for all
+    the iterations that are done are written to TensorArrays indexed by the pfor
+    iteration number. When all iterations are done, the TensorArrays are stacked
+    to get the final value.
+
+    Args:
+      pfor_input: A PForInput object corresponding to the output of any Exit
+        node from this while loop.
+
+    Returns:
+      List of converted outputs.
+    """
+    # Create init_values that will be passed to the while_loop.
+    init_values, inputs_stacked, shape_invariants = self._create_init_values(
+        pfor_input)
+    # Note that we use a list as a hack since we need the nested function body
+    # to set the value of cond_is_stacked. python2.x doesn't support nonlocal
+    # variables.
+    cond_is_stacked = [None]
+
+    def cond(not_all_done, *_):
+      return not_all_done
+
+    def body(not_all_done, indices, *args):
+      # See documentatin for __call__ for the structure of *args.
+      num_enters = len(self._enters)
+      inputs = args[:num_enters]
+      output_tas = args[num_enters:]
+      # TODO(agarwal): see which outputs have consumers and only populate the
+      # TensorArrays corresponding to those. Or do those paths get trimmed out
+      # from inside the while_loop body?
+      assert len(inputs) >= len(output_tas)
+      assert len(inputs) == len(inputs_stacked)
+
+      # Convert condition
+      with ops.name_scope("while_cond"):
+        # Note that we set cond_stacked to True here. At this point we don't
+        # know if it could be loop invariant, hence the conservative value is
+        # to assume stacked.
+        cond_pfor = self._init_pfor(pfor_input.pfor, indices,
+                                    cond_stacked=True,
+                                    inputs=inputs,
+                                    inputs_stacked=inputs_stacked)
+        conditions, cond_stacked, _ = cond_pfor._convert_helper(self._condition)
+        cond_is_stacked[0] = cond_stacked
+
+      # Recompute the new condition, write outputs of done iterations, and
+      # partition the inputs if needed.
+      if not cond_stacked:
+        (not_all_done, new_indices,
+         new_inputs, new_output_tas) = self._process_cond_unstacked(
+             conditions, indices, inputs, output_tas)
+      else:
+        (not_all_done, new_indices,
+         new_inputs, new_output_tas) = self._process_cond_stacked(
+             conditions, indices, inputs, inputs_stacked, output_tas)
+
+      # Convert body
+      with ops.name_scope("while_body"):
+        #  Compute the outputs from the body.
+        new_outputs = self._process_body(pfor_input, inputs_stacked,
+                                         new_indices, cond_stacked, new_inputs,
+                                         not_all_done)
+
+      # Note that the first num_outputs new values of inputs are computed using
+      # the body. Rest of them were direct Enters into the condition/body and
+      # the partitioning done earlier is sufficient to give the new value.
+      num_outputs = len(self._outputs)
+      new_args = ([not_all_done, new_indices] + new_outputs + list(
+          new_inputs[num_outputs:]) + new_output_tas)
+      return tuple(new_args)
+
+    while_outputs = control_flow_ops.while_loop(
+        cond, body, init_values, shape_invariants=shape_invariants)
+    output_tas = while_outputs[-len(self._outputs):]
+    outputs = []
+    assert cond_is_stacked[0] is not None
+    for inp_stacked, ta in zip(inputs_stacked, output_tas):
+      if cond_is_stacked[0]:
+        outputs.append(wrap(ta.stack(), True))
+      else:
+        # Note that if while_loop condition is unstacked, all iterations exit at
+        # the same time and we wrote those outputs in index 0 of the tensor
+        # array.
+        outputs.append(wrap(ta.read(0), inp_stacked))
+    return outputs
+
+
+class _PforInput(object):
+  """Input object passed to registered pfor converters."""
+
+  def __init__(self, pfor, op, inputs):
+    """Creates a _PforInput object.
+
+    Args:
+      pfor: PFor converter object.
+      op: the Operation object that is being converted.
+      inputs: list of WrappedTensor objects representing converted values of the
+        inputs of `op`.
+    """
+    self.pfor = pfor
+    self._op = op
+    self._inputs = inputs
+
+  def stack_inputs(self, stack_indices=None):
+    """Stacks unstacked inputs at `stack_indices`.
+
+    Args:
+      stack_indices: indices of inputs at which stacking is done. If None,
+        stacking is done at all indices.
+    """
+    if stack_indices is None:
+      stack_indices = range(len(self._inputs))
+    length = self.pfor.loop_len_vector
+    for i in stack_indices:
+      inp = self._inputs[i]
+      if not inp.is_stacked:
+        self._inputs[i] = _stack(inp.t, length)
+
+  def expanddim_inputs_for_broadcast(self):
+    """Reshapes stacked inputs to prepare them for broadcast.
+
+    Since stacked inputs have an extra leading dimension, automatic broadcasting
+    rules could incorrectly try to expand dimensions before that leading
+    dimension. To avoid that, we reshape these stacked inputs to the maximum
+    rank they will need to be broadcasted to.
+    """
+    if not self._inputs:
+      return
+
+    # Find max rank
+    def _get_rank(x):
+      rank = array_ops.rank(x.t)
+      if not x.is_stacked:
+        rank += 1
+      return rank
+
+    ranks = [_get_rank(x) for x in self._inputs]
+    max_rank = ranks[0]
+    for rank in ranks[1:]:
+      max_rank = math_ops.maximum(rank, max_rank)
+
+    for i, inp in enumerate(self._inputs):
+      if inp.is_stacked:
+        shape = array_ops.shape(inp.t)
+        rank_diff = array_ops.reshape(max_rank - ranks[i], [1])
+        ones = array_ops.tile([1], rank_diff)
+        new_shape = array_ops.concat([shape[:1], ones, shape[1:]], axis=0)
+        self._inputs[i] = wrap(array_ops.reshape(inp.t, new_shape), True)
+
+  @property
+  def inputs(self):
+    return self._inputs
+
+  @property
+  def num_inputs(self):
+    return len(self._inputs)
+
+  def input(self, index):
+    assert len(self._inputs) > index, (index, self._inputs)
+    return self._inputs[index]
+
+  def stacked_input(self, index):
+    t, is_stacked, _ = self.input(index)
+    if not is_stacked:
+      op_type = self.op_type
+      op_def = getattr(self._op, "op_def", None)
+      if op_def is None:
+        input_name = "at index %d" % index
+      else:
+        input_name = "\"%s\"" % op_def.input_arg[index].name
+      raise ValueError("Input %s of op \"%s\" expected to be not loop invariant"
+                       ".\nError while converting op %s"
+                       "with converted inputs\n%s" % (input_name, op_type,
+                                                      self._op, self.inputs))
+    return t
+
+  def unstacked_input(self, index):
+    t, is_stacked, _ = self.input(index)
+    if is_stacked:
+      op_type = self.op_type
+      op_def = getattr(self._op, "op_def", None)
+      if op_def is None:
+        input_name = "at index %d" % index
+      else:
+        input_name = "\"%s\"" % op_def.input_arg[index].name
+      raise ValueError("Input %s of op \"%s\" expected to be loop invariant"
+                       ".\nError while converting op %s"
+                       "with converted inputs\n%s" % (input_name, op_type,
+                                                      self._op, self.inputs))
+    return t
+
+  @property
+  def op(self):
+    return self._op
+
+  @property
+  def op_type(self):
+    return self._op.type
+
+  def get_attr(self, attr):
+    return self._op.get_attr(attr)
+
+  @property
+  def outputs(self):
+    return self._op.outputs
+
+  def output(self, index):
+    assert index < len(self._op.outputs)
+    return self._op.outputs[index]
+
+
+_pfor_converter_registry = {}
+
+
+class RegisterPFor(object):
+  """Utility to register converters for pfor.
+
+  Usage:
+  @RegisterPFor(foo_op_type)
+  def _foo_converter(pfor_input):
+    ...
+
+  The above will register conversion function `_foo_converter` for handling
+  conversion of `foo_op_type`. During conversion, the registered functin will be
+  called with a single argument of type `PForInput` which will contain state
+  needed for the conversion.  This registered function should output a list of
+  WrappedTensor object with the same length as the number of outputs of op being
+  converted. If the op had zero outputs, then it should return a ops.Operation
+  object.
+  """
+
+  def __init__(self, op_type):
+    """Creates an object to register a converter for op with type `op_type`."""
+    self.op_type = op_type
+
+  def __call__(self, converter):
+    name = self.op_type
+    assert name not in _pfor_converter_registry, "Re-registering %s " % name
+    _pfor_converter_registry[name] = converter
+    return converter
+
+
+class RegisterPForWithArgs(RegisterPFor):
+  """Utility to register converters for pfor.
+
+  Usage:
+  @RegisteRPFor(foo_op_type, foo=value, ....)
+  def _foo_converter(pfor_input, foo=None, ....):
+    ...
+
+  See RegisterPFor for details on the conversion function.
+  `RegisterPForWithArgs` allows binding extra arguments to the
+  conversion function at registration time.
+  """
+
+  def __init__(self, op_type, *args, **kw_args):
+    super(RegisterPForWithArgs, self).__init__(op_type)
+    self._args = args
+    self._kw_args = kw_args
+
+  def __call__(self, converter):
+
+    def _f(pfor_input):
+      return converter(pfor_input, self.op_type, *self._args, **self._kw_args)
+
+    super(RegisterPForWithArgs, self).__call__(_f)
+    return converter
+
+
+def _create_op(op_type, inputs, op_dtypes, attrs=None):
+  """Utility to create an op."""
+  return ops.get_default_graph().create_op(
+      op_type, inputs, op_dtypes, attrs=attrs, compute_device=True)
+
+
+WrappedTensor = collections.namedtuple("WrappedTensor",
+                                       ["t", "is_stacked", "is_sparse_stacked"])
+"""Wrapper around the result of a Tensor conversion.
+
+The additional fields are useful for keeping track of the conversion state as
+data flows through the ops in the loop body. For every op whose output is a
+Tensor, its converter should return either a WrappedTensor or a list of
+WrappedTensors.
+
+Args:
+  t: The converted tensor
+  is_stacked: True if the tensor is stacked, i.e. represents the results of all
+    the iterations of the loop, where each row i of the tensor corresponds to
+    that op's output on iteration i of the loop. False if the tensor is not
+    stacked, i.e. represents the result of the op on of a single iteration of
+    the loop, where the result does not vary between iterations.
+  is_sparse_stacked: True if the tensor corresponds to a component tensor
+    (indices, values, or dense_shape) of a sparse tensor, and has been logically
+    stacked via a sparse conversion.
+"""
+
+
+def wrap(tensor, is_stacked=True, is_sparse_stacked=False):
+  """Helper to create a WrappedTensor object."""
+  assert isinstance(is_stacked, bool)
+  assert isinstance(is_sparse_stacked, bool)
+  assert isinstance(tensor, ops.Tensor)
+  assert not is_sparse_stacked or is_stacked, ("If the wrapped tensor is "
+                                               "stacked via a sparse "
+                                               "conversion, it must also be "
+                                               "stacked.")
+  return WrappedTensor(tensor, is_stacked, is_sparse_stacked)
+
+
+def _fallback_converter(pfor_input):
+  logging.warn("Using a while_loop for converting %s", pfor_input.op_type)
+  output_dtypes = [x.dtype for x in pfor_input.outputs]
+  iters = pfor_input.pfor.loop_len_vector[0]
+
+  def while_body(i, *ta_list):
+    """Body of while loop."""
+    inputs = [
+        x[i, ...] if stacked else x for x, stacked, _ in pfor_input.inputs
+    ]
+    op_outputs = _create_op(
+        pfor_input.op_type,
+        inputs,
+        output_dtypes,
+        attrs=pfor_input.op.node_def.attr).outputs
+
+    outputs = []
+    for out, ta in zip(op_outputs, ta_list):
+      assert isinstance(out, ops.Tensor)
+      outputs.append(ta.write(i, array_ops.expand_dims(out, 0)))
+    return tuple([i + 1] + outputs)
+
+  ta_list = control_flow_ops.while_loop(
+      lambda i, *ta: i < iters, while_body, [0] + [
+          tensor_array_ops.TensorArray(dtype, iters) for dtype in output_dtypes
+      ])[1:]
+  return tuple([wrap(ta.concat(), True) for ta in ta_list])
+
+
+class PFor(object):
+  """Implementation of rewrite of parallel-for loops.
+
+  This class takes a DAG or a set of DAGs representing the body of a
+  parallel-for loop, and adds new operations to the graph that implements
+  functionality equivalent to running that loop body for a specified number of
+  iterations. This new set of nodes may or may not use a tensorflow loop
+  construct.
+
+  The process of conversion does not delete or change any existing operations.
+  It only adds operations that efficiently implement the equivalent
+  functionality. We refer to the added ops as "converted ops".
+
+  The conversion process uses a simple greedy heuristic. It walks the loop body
+  and tries to express the functionality of running each node in a loop with a
+  new set of nodes. When converting an op several cases are possible:
+  - The op is not inside the loop body. Hence it can be used as is.
+  - The op does not depend on the iteration number and is stateless. In this
+    case, it can be used as is.
+  - The op is not stateful, and depends on iteration number only through control
+    dependencies. In this case, we can create a single op with same inputs and
+    attributes, but with "converted" control dependencies.
+  - The op is not stateful, and all its inputs are loop invariant. In this
+    case, similar to above, we can create a single op with same inputs and
+    attributes, but with "converted" control dependencies.
+  - The op is stateful or at least one of the inputs is not loop invariant. In
+    this case, we run the registered converter for that op to create a set of
+    converted ops. All nodes in the set will have converted control dependencies
+    corresponding to control dependencies of the original op. If the op returned
+    multiple outputs, "converted outputs" could be produced by different ops in
+    this set.
+  """
+
+  def __init__(self,
+               loop_var,
+               loop_len,
+               pfor_ops,
+               all_indices=None,
+               all_indices_partitioned=False):
+    """Creates an object to rewrite a parallel-for loop.
+
+    Args:
+      loop_var: ops.Tensor output of a Placeholder operation. The value should
+        be an int32 scalar representing the loop iteration number.
+      loop_len: A scalar or scalar Tensor representing the number of iterations
+        the loop is run for.
+      pfor_ops: List of all ops inside the loop body.
+      all_indices: If not None, an int32 vector with size `loop_len`
+        representing the iteration ids that are still active. These values
+        should be unique and sorted. However they may not be contiguous. This is
+        typically the case when inside a control flow construct which has
+        partitioned the indices of the iterations that are being converted.
+      all_indices_partitioned: If True, this object is being constructed from a
+       control flow construct where not all the pfor iterations are guaranteed
+       to be active.
+    """
+    assert isinstance(loop_var, ops.Tensor)
+    assert loop_var.op.type == "Placeholder"
+    self._loop_var = loop_var
+    loop_len_value = tensor_util.constant_value(loop_len)
+    if loop_len_value is not None:
+      loop_len = loop_len_value
+    self._loop_len_vector = array_ops.reshape(loop_len, [1])
+    self._all_indices_partitioned = all_indices_partitioned
+    if all_indices_partitioned:
+      assert all_indices is not None
+    self.all_indices = (
+        math_ops.range(loop_len) if all_indices is None else all_indices)
+
+    self._conversion_map = {}
+    self._conversion_map[loop_var] = wrap(self.all_indices, True)
+    self._pfor_ops = set(pfor_ops)
+    self._pfor_op_ids = set([x._id for x in pfor_ops])
+
+  def op_is_inside_loop(self, op):
+    """True if op was created inside the pfor loop body."""
+    assert isinstance(op, ops.Operation)
+    # Note that we use self._pfor_op_ids for the check and not self._pfor_ops
+    # since it appears there tensorflow API could return different python
+    # objects representing the same Operation node.
+    return op._id in self._pfor_op_ids
+
+  def _convert_sparse(self, y):
+    """Returns the converted value corresponding to SparseTensor y.
+
+    For SparseTensors, instead of stacking the component tensors separately,
+    resulting in component tensors with shapes (N, m, rank), (N, m), and (N,
+    rank) respectively for indices, values, and dense_shape (where N is the loop
+    length and m is the number of sparse tensor values per loop iter), we want
+    to logically stack the SparseTensors, to create a SparseTensor whose
+    components are size (N * m, rank + 1), (N * m, ), and (rank + 1,)
+    respectively.
+
+    Here, we try to get the conversion of each component tensor.
+    If the tensors are stacked via a sparse conversion, return the resulting
+    SparseTensor composed of the converted components. Otherwise, the component
+    tensors are either unstacked or stacked naively. In the latter case, we
+    unstack the component tensors to reform loop_len SparseTensor elements,
+    then correctly batch them.
+
+    The unstacked tensors must have the same rank. Each dimension of each
+    SparseTensor will expand to be the largest among all SparseTensor elements
+    for that dimension. For example, if there are N SparseTensors of rank 3
+    being stacked, with N dense shapes, where the i_th shape is (x_i, y_i, z_i),
+    the new dense shape will be (N, max_i(x_i), max_i(y_i), max_i(z_i)).
+
+    Args:
+      y: A tf.SparseTensor.
+
+    Returns:
+      A tf.SparseTensor that is the converted value corresponding to y.
+    """
+    outputs = [
+        self._convert_helper(t) for t in (y.indices, y.values, y.dense_shape)
+    ]
+    assert all(isinstance(o, WrappedTensor) for o in outputs)
+
+    if all(w.is_sparse_stacked for w in outputs):
+      return sparse_tensor.SparseTensor(*[w.t for w in outputs])
+
+    assert not any(w.is_sparse_stacked for w in outputs), (
+        "Error converting SparseTensor. All components should be logically "
+        "stacked, or none.")
+
+    # If component tensors were not sparsely stacked, they are either unstacked
+    # or stacked without knowledge that they are components of sparse tensors.
+    # In this case, we have to restack them.
+    return self._restack_sparse_tensor_logically(
+        *[self._unwrap_or_tile(w) for w in outputs])
+
+  def _restack_sparse_tensor_logically(self, indices, values, shape):
+    sparse_tensor_rank = indices.get_shape()[-1].value
+    if sparse_tensor_rank is not None:
+      sparse_tensor_rank += 1
+
+    def map_fn(args):
+      res = gen_sparse_ops.serialize_sparse(
+          args[0], args[1], args[2], out_type=dtypes.variant)
+      return res
+
+    # Applies a map function to the component tensors to serialize each
+    # sparse tensor element and batch them all, then deserializes the batch.
+    # TODO(rachelim): Try to do this without map_fn -- add the right offsets
+    # to shape and indices tensors instead.
+    result = functional_ops.map_fn(
+        map_fn, [indices, values, shape], dtype=dtypes.variant)
+    return sparse_ops.deserialize_sparse(
+        result, dtype=values.dtype, rank=sparse_tensor_rank)
+
+  def _unwrap_or_tile(self, wrapped_tensor):
+    """Given a wrapped tensor, unwrap if stacked. Otherwise, tiles it."""
+    output, is_stacked = wrapped_tensor.t, wrapped_tensor.is_stacked
+    if is_stacked:
+      return output
+    else:
+      return _stack(output, self._loop_len_vector).t
+
+  def convert(self, y):
+    """Returns the converted value corresponding to y.
+
+    Args:
+      y: A ops.Tensor or a ops.Operation object. If latter, y should not have
+        any outputs.
+
+    Returns:
+      If y does not need to be converted, it returns y as is. Else it returns
+      the "converted value" corresponding to y.
+    """
+    if isinstance(y, sparse_tensor.SparseTensor):
+      return self._convert_sparse(y)
+    output = self._convert_helper(y)
+    if isinstance(output, WrappedTensor):
+      assert isinstance(y, ops.Tensor)
+      return self._unwrap_or_tile(output)
+    else:
+      assert isinstance(y, ops.Operation)
+      assert not y.outputs
+      assert isinstance(output, ops.Operation)
+    return output
+
+  def _was_converted(self, t):
+    """True if t is not a conversion of itself."""
+    converted_t = self._conversion_map[t]
+    return converted_t.t is not t
+
+  def _add_conversion(self, old_output, new_output):
+    self._conversion_map[old_output] = new_output
+
+  def _convert_helper(self, op_or_tensor):
+    stack = [op_or_tensor]
+    while stack:
+      y = stack[0]
+      if y in self._conversion_map:
+        assert isinstance(self._conversion_map[y],
+                          (WrappedTensor, ops.Operation))
+        stack.pop(0)
+        continue
+      if isinstance(y, ops.Operation):
+        assert not y.outputs, (
+            "We only support converting Operation objects with no outputs. "
+            "Got %s", y)
+        y_op = y
+      else:
+        assert isinstance(y, ops.Tensor), y
+        y_op = y.op
+
+      is_while_loop = y_op.type == "Exit"
+      if is_while_loop:
+        while_op = WhileOp(y, pfor_ops=self._pfor_ops)
+        is_inside_loop = while_op.is_inside_loop
+        # If all nodes in the while_loop graph were created inside the pfor, we
+        # treat the whole loop subgraph as a single op (y_op) and try to convert
+        # it. For while_loops that are created completely or partially outside,
+        # we treat them as external and should be able to simply return the Exit
+        # node output as is without needing any conversion. Note that for
+        # while_loops that are partially constructed inside, we assume they will
+        # be loop invariant. If that is not the case, it will create runtime
+        # errors since the converted graph would depend on the self._loop_var
+        # placeholder.
+        if is_inside_loop:
+          y_op = while_op
+      else:
+        is_inside_loop = self.op_is_inside_loop(y_op)
+
+      # If this op was not created inside the loop body, we will return as is.
+      # 1. Convert inputs and control inputs.
+
+      def _add_to_stack(x):
+        if x not in self._conversion_map:
+          stack.insert(0, x)
+          return True
+        else:
+          return False
+
+      if is_inside_loop:
+        added_to_stack = False
+        for inp in y_op.inputs:
+          added_to_stack |= _add_to_stack(inp)
+        for cinp in y_op.control_inputs:
+          if cinp.outputs:
+            for t in cinp.outputs:
+              added_to_stack |= _add_to_stack(t)
+          else:
+            added_to_stack |= _add_to_stack(cinp)
+        if added_to_stack:
+          continue
+
+        converted_inputs = [self._conversion_map[inp] for inp in y_op.inputs]
+        some_input_converted = any(
+            [self._was_converted(x) for x in y_op.inputs])
+        some_input_stacked = any([x.is_stacked for x in converted_inputs])
+
+        converted_control_ops = set()
+        some_control_input_converted = False
+        for cinp in y_op.control_inputs:
+          if cinp.outputs:
+            for t in cinp.outputs:
+              converted_t = self._conversion_map[t]
+              if self._was_converted(t):
+                some_control_input_converted = True
+              converted_control_ops.add(converted_t.t.op)
+          else:
+            converted_cinp = self._conversion_map[cinp]
+            assert isinstance(converted_cinp, ops.Operation)
+            if converted_cinp != cinp:
+              some_control_input_converted = True
+            converted_control_ops.add(converted_cinp)
+        converted_control_ops = list(converted_control_ops)
+        is_stateful = _is_stateful_pfor_op(y_op)
+      else:
+        converted_inputs = []
+        converted_control_ops = []
+      logging.vlog(3, "converting op:%s\ninputs:%s\ncontrol_inputs:%s", y_op,
+                   converted_inputs, converted_control_ops)
+
+      # 2. Convert y_op
+      # If converting a while_loop, we let the while_loop convertor deal with
+      # putting the control dependencies appropriately.
+      control_dependencies = [] if is_while_loop else converted_control_ops
+      with ops.control_dependencies(control_dependencies), ops.name_scope(
+          y_op.name + "/pfor/"):
+        # None of the inputs and control inputs were converted.
+        if (not is_inside_loop or
+            (not is_stateful and not some_input_converted and
+             not some_control_input_converted)):
+          if y == y_op:
+            assert not isinstance(y_op, WhileOp)
+            new_outputs = y_op
+          else:
+            new_outputs = [wrap(x, False) for x in y_op.outputs]
+        elif not (is_stateful or is_while_loop or some_input_stacked):
+          # All inputs are unstacked or uncoverted but some control inputs are
+          # converted.
+          # TODO(rachelim): Handle the case where some inputs are sparsely
+          # stacked (i.e. any([x.is_sparse_stacked for x in converted_inputs]))
+          new_op = _create_op(y_op.type, [x.t for x in converted_inputs],
+                              [x.dtype for x in y_op.outputs],
+                              y_op.node_def.attr)
+          if y == y_op:
+            new_outputs = new_op
+          else:
+            new_outputs = [wrap(x, False) for x in new_op.outputs]
+        else:
+          # Either some inputs are not loop invariant or op is stateful.
+          if hasattr(y_op, "pfor_converter"):
+            converter = y_op.pfor_converter
+          else:
+            converter = _pfor_converter_registry.get(y_op.type, None)
+          if converter is None:
+            if flags.FLAGS.op_conversion_fallback_to_while_loop:
+              converter = _fallback_converter
+            else:
+              raise ValueError(
+                  "No converter defined for %s\n%s\ninputs: %s. "
+                  "\nEither add a converter or set "
+                  "--op_conversion_fallback_to_while_loop=True, "
+                  "which may run slower" % (y_op.type, y_op, converted_inputs))
+          # TODO(rachelim): Handle the case where some inputs are sparsely
+          # stacked. We should only call the converter if it supports handling
+          # those inputs.
+          new_outputs = converter(_PforInput(self, y_op, converted_inputs))
+          if isinstance(new_outputs, WrappedTensor):
+            new_outputs = [new_outputs]
+          assert isinstance(new_outputs,
+                            (list, tuple, ops.Operation)), new_outputs
+        logging.vlog(2, "converted %s %s", y_op, new_outputs)
+
+        # Insert into self._conversion_map
+        if y == y_op:
+          assert isinstance(new_outputs, ops.Operation)
+          self._add_conversion(y_op, new_outputs)
+        else:
+          for old_output, new_output in zip(y_op.outputs, new_outputs):
+            assert isinstance(new_output, WrappedTensor), (new_output, y, y_op)
+            self._add_conversion(old_output, new_output)
+        stack.pop(0)
+
+    return self._conversion_map[op_or_tensor]
+
+  @property
+  def loop_len_vector(self):
+    """Returns a single element vector whose value is number of iterations."""
+    return self._loop_len_vector
+
+  @property
+  def loop_var(self):
+    """Returns placeholder loop variable."""
+    return self._loop_var
+
+  @property
+  def pfor_ops(self):
+    return self._pfor_ops
+
+  @property
+  def all_indices_partitioned(self):
+    """all_indices_partitioned property.
+
+    Returns:
+      True if we are inside a control flow construct and not all pfor iterations
+      may be active.
+    """
+    return self._all_indices_partitioned
+
+# nn_ops
+
+
+def _flatten_first_two_dims(x):
+  """Merges first two dimensions."""
+  old_shape = array_ops.shape(x)
+  new_shape = array_ops.concat([[-1], old_shape[2:]], axis=0)
+  return array_ops.reshape(x, new_shape)
+
+
+def _unflatten_first_dim(x, first_dim):
+  """Splits first dimension into [first_dim, -1]."""
+  old_shape = array_ops.shape(x)
+  new_shape = array_ops.concat([first_dim, [-1], old_shape[1:]], axis=0)
+  return array_ops.reshape(x, new_shape)
+
+
+def _inputs_with_flattening(pfor_input, input_indices):
+  """Stacks and flattens first dim of inputs at indices `input_indices`."""
+  if input_indices is None:
+    input_indices = []
+  pfor_input.stack_inputs(stack_indices=input_indices)
+  inputs = []
+  for i in range(pfor_input.num_inputs):
+    if i in input_indices:
+      inp = pfor_input.stacked_input(i)
+      inp = _flatten_first_two_dims(inp)
+    else:
+      inp = pfor_input.unstacked_input(i)
+    inputs.append(inp)
+  return inputs
+
+
+@RegisterPForWithArgs("Conv2D", dims=[0])
+@RegisterPForWithArgs("AvgPool", dims=[0])
+@RegisterPForWithArgs("MaxPool", dims=[0])
+@RegisterPForWithArgs("MaxPoolGrad", dims=[0, 1, 2])
+@RegisterPForWithArgs("SoftmaxCrossEntropyWithLogits", dims=[0, 1])
+def _convert_flatten_batch(pfor_input, op_type, dims):
+  del op_type
+  inputs = _inputs_with_flattening(pfor_input, dims)
+  outputs = _create_op(
+      pfor_input.op_type,
+      inputs, [x.dtype for x in pfor_input.outputs],
+      attrs=pfor_input.op.node_def.attr).outputs
+  n = pfor_input.pfor.loop_len_vector
+  outputs = [_unflatten_first_dim(x, n) for x in outputs]
+  return [wrap(x, True) for x in outputs]
+
+
+_channel_flatten_input_cache = {}
+
+
+def _channel_flatten_input(x, data_format):
+  """Merge the stack dimension with the channel dimension.
+
+  If S is pfor's stacking dimension, then,
+    - for SNCHW, we transpose to NSCHW. If N dimension has size 1, the transpose
+      should be cheap.
+    - for SNHWC, we transpose to NHWCS.
+  We then merge the S and C dimension.
+
+  Args:
+    x: ops.Tensor to transform.
+    data_format: "NCHW" or "NHWC".
+
+  Returns:
+    A 3-element tuple with the transformed value, along with the shape for
+    reshape and order for transpose required to transform back.
+  """
+
+  graph = ops.get_default_graph()
+  cache_key = (graph, x, data_format)
+  if cache_key not in _channel_flatten_input_cache:
+    x_shape = array_ops.shape(x)
+    if data_format == b"NCHW":
+      order = [1, 0, 2, 3, 4]
+      shape = array_ops.concat([x_shape[1:2], [-1], x_shape[3:]], axis=0)
+      reverse_order = order
+    else:
+      order = [1, 2, 3, 0, 4]
+      shape = array_ops.concat([x_shape[1:4], [-1]], axis=0)
+      reverse_order = [3, 0, 1, 2, 4]
+    # Move S dimension next to C dimension.
+    x = array_ops.transpose(x, order)
+    reverse_shape = array_ops.shape(x)
+    # Reshape to merge the S and C dimension.
+    x = array_ops.reshape(x, shape)
+    outputs = x, reverse_order, reverse_shape
+    _channel_flatten_input_cache[cache_key] = outputs
+  else:
+    outputs = _channel_flatten_input_cache[cache_key]
+  return outputs
+
+
+# Note that with training=True, running FusedBatchNorm on individual examples
+# is very different from running FusedBatchNorm on a batch of those examples.
+# This is because, for the latter case, the operation can be considered as first
+# computing the mean and variance over all the examples and then using these
+# to scale all those examples. This creates a data dependency between these
+# different "iterations" since the inputs to the scaling step depends on the
+# statistics coming from all these inputs.
+# As with other kernels, the conversion here effectively runs the kernel
+# independently for each iteration, and returns outputs by stacking outputs from
+# each of those iterations.
+@RegisterPFor("FusedBatchNorm")
+def _convert_fused_batch_norm(pfor_input):
+  is_training = pfor_input.get_attr("is_training")
+  # When BatchNorm is used with training=False, mean and variance are provided
+  # externally and used as is by the op. Thus, we can merge the S and N
+  # dimensions as we do for regular operations.
+  # When BatchNorm is used with training=True, mean and variance are computed
+  # for each channel across the batch dimension (first one). If we merge S and N
+  # dimensions, mean and variances will be computed over a larger set. So, we
+  # merge the S and C dimensions instead.
+  if not is_training:
+    # We return zeros for batch_mean and batch_variance output. Note that CPU
+    # and GPU seem to have different behavior for those two outputs. CPU outputs
+    # zero because these values are not used during inference. GPU outputs
+    # something, probably real means and variances.
+    inputs = _inputs_with_flattening(pfor_input, [0])
+    outputs = _create_op(
+        pfor_input.op_type,
+        inputs, [x.dtype for x in pfor_input.outputs],
+        attrs=pfor_input.op.node_def.attr).outputs
+    y = outputs[0]
+    n = pfor_input.pfor.loop_len_vector
+    y = _unflatten_first_dim(y, n)
+    mean = pfor_input.unstacked_input(3)
+    zeros = array_ops.zeros_like(mean)
+    return [wrap(y, True), wrap(zeros, False), wrap(zeros, False)]
+
+  pfor_input.stack_inputs()
+  data_format = pfor_input.get_attr("data_format")
+  # We merge the first dimension with the "C" dimension, run FusedBatchNorm, and
+  # then transpose back.
+  x = pfor_input.stacked_input(0)
+  x, reverse_order, reverse_shape = _channel_flatten_input(x, data_format)
+  # Note that we stack all the other inputs as well so that they are the same
+  # size as the new size of the channel dimension.
+  inputs = [x] + [
+      array_ops.reshape(pfor_input.stacked_input(i), [-1])
+      for i in range(1, pfor_input.num_inputs)
+  ]
+  outputs = _create_op(
+      pfor_input.op_type,
+      inputs, [x.dtype for x in pfor_input.outputs],
+      attrs=pfor_input.op.node_def.attr).outputs
+  y = outputs[0]
+  y = array_ops.reshape(y, reverse_shape)
+  y = array_ops.transpose(y, reverse_order)
+  n = pfor_input.pfor.loop_len_vector
+  outputs = [_unflatten_first_dim(x, n) for x in outputs[1:]]
+  outputs = [y] + outputs
+  return [wrap(x, True) for x in outputs]
+
+
+@RegisterPFor("FusedBatchNormGrad")
+def _convert_fused_batch_norm_grad(pfor_input):
+  pfor_input.stack_inputs()
+  data_format = pfor_input.get_attr("data_format")
+  y_backprop = pfor_input.stacked_input(0)
+  y_backprop, _, _ = _channel_flatten_input(y_backprop, data_format)
+  x = pfor_input.stacked_input(1)
+  x, x_reverse_order, x_reverse_shape = _channel_flatten_input(x, data_format)
+  inputs = [y_backprop, x] + [
+      array_ops.reshape(pfor_input.stacked_input(i), [-1])
+      for i in range(2, pfor_input.num_inputs)
+  ]
+  outputs = _create_op(
+      pfor_input.op_type,
+      inputs, [x.dtype for x in pfor_input.outputs],
+      attrs=pfor_input.op.node_def.attr).outputs
+  x_backprop = outputs[0]
+  x_backprop = array_ops.reshape(x_backprop, x_reverse_shape)
+  x_backprop = array_ops.transpose(x_backprop, x_reverse_order)
+  n = pfor_input.pfor.loop_len_vector
+  outputs = [_unflatten_first_dim(x, n) for x in outputs[1:]]
+  outputs = [x_backprop] + outputs
+  return [wrap(output, True) for output in outputs]
+
+
+@RegisterPForWithArgs("Conv2DBackpropInput", flatten_dims=[2], shape_dim=0)
+@RegisterPForWithArgs("AvgPoolGrad", flatten_dims=[1], shape_dim=0)
+def _convert_flatten_batch_shape_input(pfor_input, op_type, flatten_dims,
+                                       shape_dim):
+  del op_type
+  inputs = _inputs_with_flattening(pfor_input, flatten_dims)
+  n = pfor_input.pfor.loop_len_vector
+  # Adjust the `input_sizes` input.
+  ones = array_ops.ones(
+      [array_ops.shape(inputs[shape_dim])[0] - 1], dtype=n.dtype)
+  inputs[shape_dim] *= array_ops.concat([n, ones], axis=0)
+  outputs = _create_op(
+      pfor_input.op_type,
+      inputs, [x.dtype for x in pfor_input.outputs],
+      attrs=pfor_input.op.node_def.attr).outputs
+  outputs = [_unflatten_first_dim(x, n) for x in outputs]
+  return [wrap(x, True) for x in outputs]
+
+
+@RegisterPFor("Conv2DBackpropFilter")
+def _convert_conv2d_backprop_filter(pfor_input):
+  pfor_input.stack_inputs(stack_indices=[2])
+  inputs, inputs_stacked, _ = pfor_input.input(0)
+  filter_sizes = pfor_input.unstacked_input(1)
+  grads = pfor_input.stacked_input(2)
+  strides = pfor_input.get_attr("strides")
+  padding = pfor_input.get_attr("padding")
+  use_cudnn_on_gpu = pfor_input.get_attr("use_cudnn_on_gpu")
+  data_format = pfor_input.get_attr("data_format")
+  dilations = pfor_input.get_attr("dilations")
+  if inputs_stacked:
+    # TODO(agarwal): Implement this efficiently.
+    logging.warn("Conv2DBackpropFilter uses a while_loop. Fix that!")
+
+    def while_body(i, ta):
+      inp_i = inputs[i, ...]
+      grad_i = grads[i, ...]
+      output = nn_ops.conv2d_backprop_filter(
+          inp_i,
+          filter_sizes,
+          grad_i,
+          strides=strides,
+          padding=padding,
+          use_cudnn_on_gpu=use_cudnn_on_gpu,
+          data_format=data_format,
+          dilations=dilations)
+      return i + 1, ta.write(i, array_ops.expand_dims(output, 0))
+
+    n = array_ops.reshape(pfor_input.pfor.loop_len_vector, [])
+    _, ta = control_flow_ops.while_loop(
+        lambda i, ta: i < n, while_body,
+        (0, tensor_array_ops.TensorArray(inputs.dtype, n)))
+    output = ta.concat()
+    return wrap(output, True)
+  else:
+    # We merge the stack dimension with the channel dimension of the gradients
+    # and pretend we had a larger filter (see change to filter_sizes below).
+    # Once the filter backprop is computed, we reshape and transpose back
+    # appropriately.
+    grads, _, _ = _channel_flatten_input(grads, data_format)
+    n = pfor_input.pfor.loop_len_vector
+    old_filter_sizes = filter_sizes
+    filter_sizes *= array_ops.concat([[1, 1, 1], n], axis=0)
+    output = nn_ops.conv2d_backprop_filter(
+        inputs,
+        filter_sizes,
+        grads,
+        strides=strides,
+        padding=padding,
+        use_cudnn_on_gpu=use_cudnn_on_gpu,
+        data_format=data_format,
+        dilations=dilations)
+    new_filter_shape = array_ops.concat([old_filter_sizes[:3], n, [-1]], axis=0)
+    output = array_ops.reshape(output, new_filter_shape)
+    output = array_ops.transpose(output, [3, 0, 1, 2, 4])
+    return wrap(output, True)
+
+
+# array_ops
+
+
+@RegisterPForWithArgs("Identity", array_ops.identity)
+@RegisterPForWithArgs("StopGradient", array_ops.stop_gradient)
+def _convert_identity(pfor_input, op_type, op_func):
+  del op_type
+  return wrap(op_func(*[x.t for x in pfor_input.inputs]), True)
+
+
+@RegisterPFor("Reshape")
+def _convert_reshape(pfor_input):
+  t = pfor_input.stacked_input(0)
+  shape = pfor_input.unstacked_input(1)
+  new_dim = array_ops.shape(t)[:1]
+  new_shape = array_ops.concat([new_dim, shape], axis=0)
+  return wrap(array_ops.reshape(t, new_shape), True)
+
+
+@RegisterPFor("ExpandDims")
+def _convert_expanddims(pfor_input):
+  t = pfor_input.stacked_input(0)
+  dim = pfor_input.unstacked_input(1)
+  dim += math_ops.cast(dim >= 0, dtypes.int32)
+  return wrap(array_ops.expand_dims(t, axis=dim), True)
+
+
+@RegisterPFor("Slice")
+def _convert_slice(pfor_input):
+  t = pfor_input.stacked_input(0)
+  begin = pfor_input.unstacked_input(1)
+  size = pfor_input.unstacked_input(2)
+  begin = array_ops.concat([[0], begin], axis=0)
+  size = array_ops.concat([[-1], size], axis=0)
+  return wrap(array_ops.slice(t, begin, size), True)
+
+
+@RegisterPFor("Tile")
+def _convert_tile(pfor_input):
+  t = pfor_input.stacked_input(0)
+  multiples = pfor_input.unstacked_input(1)
+  multiples = array_ops.concat([[1], multiples], 0)
+  return wrap(array_ops.tile(t, multiples), True)
+
+
+@RegisterPFor("Pack")
+def _convert_pack(pfor_input):
+  pfor_input.stack_inputs()
+  axis = pfor_input.get_attr("axis")
+  if axis >= 0:
+    axis += 1
+  return wrap(
+      array_ops.stack([x.t for x in pfor_input.inputs], axis=axis), True)
+
+
+@RegisterPFor("Unpack")
+def _convert_unpack(pfor_input):
+  value = pfor_input.stacked_input(0)
+  axis = pfor_input.get_attr("axis")
+  if axis >= 0:
+    axis += 1
+  num = pfor_input.get_attr("num")
+  return [wrap(x, True) for x in array_ops.unstack(value, axis=axis, num=num)]
+
+
+@RegisterPFor("Pad")
+def _convert_pad(pfor_input):
+  t = pfor_input.stacked_input(0)
+  paddings = pfor_input.unstacked_input(1)
+  paddings = array_ops.concat([[[0, 0]], paddings], 0)
+  return wrap(array_ops.pad(t, paddings, mode="CONSTANT"), True)
+
+
+@RegisterPFor("Split")
+def _convert_split(pfor_input):
+  split_dim = pfor_input.unstacked_input(0)
+  t = pfor_input.stacked_input(1)
+  num_split = pfor_input.get_attr("num_split")
+  split_dim += math_ops.cast(split_dim >= 0, dtypes.int32)
+  return [wrap(x, True) for x in array_ops.split(t, num_split, axis=split_dim)]
+
+
+@RegisterPFor("Transpose")
+def _convert_transpose(pfor_input):
+  t = pfor_input.stacked_input(0)
+  perm = pfor_input.unstacked_input(1)
+  new_perm = array_ops.concat([[0], perm + 1], axis=0)
+  return wrap(array_ops.transpose(t, new_perm), True)
+
+
+@RegisterPFor("ZerosLike")
+def _convert_zeroslike(pfor_input):
+  t = pfor_input.stacked_input(0)
+  shape = array_ops.shape(t)[1:]
+  return wrap(array_ops.zeros(shape, dtype=t.dtype), False)
+
+
+@RegisterPFor("Gather")
+@RegisterPFor("GatherV2")
+def _convert_gather(pfor_input):
+  param, param_stacked, _ = pfor_input.input(0)
+  indices, indices_stacked, _ = pfor_input.input(1)
+  op_type = pfor_input.op_type
+  if op_type == "Gather":
+    validate_indices = pfor_input.get_attr("validate_indices")
+    axis = 0
+  else:
+    validate_indices = None
+    axis = pfor_input.unstacked_input(2)
+    axis_value = tensor_util.constant_value(axis)
+    if axis_value is not None:
+      axis = axis_value
+  if indices_stacked and not param_stacked:
+    if indices == pfor_input.pfor.all_indices and axis == 0:
+      param_shape0 = param.shape[0].value
+      indices_shape0 = indices.shape[0].value
+      if param_shape0 is not None and indices_shape0 == param_shape0:
+        # Note that with loops and conditionals, indices may not be contiguous.
+        # However they will be sorted and unique. So if the shape matches, then
+        # it must be picking up all the rows of param.
+        return wrap(param, True)
+      # TODO(agarwal): use array_ops.slice here.
+    output = array_ops.gather(
+        param, indices, validate_indices=validate_indices, axis=axis)
+    if axis != 0:
+      axis = control_flow_ops.cond(
+          axis < 0, lambda: axis + array_ops.rank(param), lambda: axis)
+      order = array_ops.concat(
+          [[axis],
+           math_ops.range(axis),
+           math_ops.range(axis + 1, array_ops.rank(output))],
+          axis=0)
+      output = control_flow_ops.cond(
+          math_ops.equal(axis, 0), lambda: output,
+          lambda: array_ops.transpose(output, order))
+    return wrap(output, True)
+  if param_stacked:
+    loop_len_vector = pfor_input.pfor.loop_len_vector
+    pfor_input.stack_inputs(stack_indices=[1])
+    indices = pfor_input.stacked_input(1)
+    param_flat = _flatten_first_two_dims(param)
+
+    # Recompute indices to handle stacked param.
+    indices_offset = math_ops.range(
+        loop_len_vector[0]) * array_ops.shape(param)[1]
+    # Reshape indices_offset to allow broadcast addition
+    ones = array_ops.ones([array_ops.rank(indices) - 1], dtype=dtypes.int32)
+    new_shape = array_ops.concat([loop_len_vector, ones], axis=0)
+    indices_offset = array_ops.reshape(indices_offset, new_shape)
+    indices += indices_offset
+
+    # TODO(agarwal): handle axis != 0. May need to transpose param or
+    # array_ops.gather_nd.
+    if isinstance(axis, ops.Tensor):
+      axis_value = tensor_util.constant_value(axis)
+    else:
+      try:
+        axis_value = int(axis)
+      except TypeError:
+        axis_value = None
+    msg = ("Gather, where indices and param are both loop dependent, currently "
+           "requires axis=0")
+    if axis_value is not None and axis_value != 0:
+      raise ValueError("Error while converting %s. %s. Got axis=%d" %
+                       (pfor_input.op, msg, axis))
+    with ops.control_dependencies(
+        [check_ops.assert_equal(axis, 0, message=msg)]):
+      output = array_ops.gather(param_flat, indices)
+    return wrap(output, True)
+
+
+@RegisterPFor("ConcatV2")
+def _convert_concatv2(pfor_input):
+  n = pfor_input.num_inputs
+  pfor_input.stack_inputs(stack_indices=range(n - 1))
+  axis = pfor_input.unstacked_input(n - 1)
+  axis += math_ops.cast(axis >= 0, axis.dtype)
+  return wrap(
+      array_ops.concat([x.t for x in pfor_input.inputs[:n - 1]], axis=axis),
+      True)
+
+
+@RegisterPFor("StridedSlice")
+def _convert_strided_slice(pfor_input):
+  inp = pfor_input.stacked_input(0)
+  begin = pfor_input.unstacked_input(1)
+  end = pfor_input.unstacked_input(2)
+  strides = pfor_input.unstacked_input(3)
+  begin_mask = pfor_input.get_attr("begin_mask")
+  end_mask = pfor_input.get_attr("end_mask")
+  ellipsis_mask = pfor_input.get_attr("ellipsis_mask")
+  new_axis_mask = pfor_input.get_attr("new_axis_mask")
+  shrink_axis_mask = pfor_input.get_attr("shrink_axis_mask")
+
+  begin = array_ops.concat([[0], begin], axis=0)
+  end = array_ops.concat([[0], end], axis=0)
+  strides = array_ops.concat([[1], strides], axis=0)
+  begin_mask = begin_mask << 1 | 1
+  end_mask = end_mask << 1 | 1
+  ellipsis_mask <<= 1
+  new_axis_mask <<= 1
+  shrink_axis_mask <<= 1
+  return wrap(
+      array_ops.strided_slice(
+          inp,
+          begin,
+          end,
+          strides,
+          begin_mask=begin_mask,
+          end_mask=end_mask,
+          ellipsis_mask=ellipsis_mask,
+          new_axis_mask=new_axis_mask,
+          shrink_axis_mask=shrink_axis_mask), True)
+
+
+@RegisterPFor("StridedSliceGrad")
+def _convert_strided_slice_grad(pfor_input):
+  shape = pfor_input.unstacked_input(0)
+  begin = pfor_input.unstacked_input(1)
+  end = pfor_input.unstacked_input(2)
+  strides = pfor_input.unstacked_input(3)
+  dy = pfor_input.stacked_input(4)
+  begin_mask = pfor_input.get_attr("begin_mask")
+  end_mask = pfor_input.get_attr("end_mask")
+  ellipsis_mask = pfor_input.get_attr("ellipsis_mask")
+  new_axis_mask = pfor_input.get_attr("new_axis_mask")
+  shrink_axis_mask = pfor_input.get_attr("shrink_axis_mask")
+
+  shape = array_ops.concat([pfor_input.pfor.loop_len_vector, shape], axis=0)
+  begin = array_ops.concat([[0], begin], axis=0)
+  end = array_ops.concat([[0], end], axis=0)
+  strides = array_ops.concat([[1], strides], axis=0)
+  begin_mask = begin_mask << 1 | 1
+  end_mask = end_mask << 1 | 1
+  ellipsis_mask <<= 1
+  new_axis_mask <<= 1
+  shrink_axis_mask <<= 1
+  return wrap(
+      array_ops.strided_slice_grad(
+          shape,
+          begin,
+          end,
+          strides,
+          dy,
+          begin_mask=begin_mask,
+          end_mask=end_mask,
+          ellipsis_mask=ellipsis_mask,
+          new_axis_mask=new_axis_mask,
+          shrink_axis_mask=shrink_axis_mask), True)
+
+
+# math_ops
+
+
+@RegisterPFor("MatMul")
+def _convert_matmul(pfor_input):
+  # TODO(agarwal): Check if tiling is faster than two transposes.
+  a, a_stacked, _ = pfor_input.input(0)
+  b, b_stacked, _ = pfor_input.input(1)
+  tr_a = pfor_input.get_attr("transpose_a")
+  tr_b = pfor_input.get_attr("transpose_b")
+  if a_stacked and b_stacked:
+    output = wrap(math_ops.matmul(a, b, adjoint_a=tr_a, adjoint_b=tr_b), True)
+    return output
+  elif a_stacked:
+    if tr_a:
+      a = array_ops.transpose(a, [0, 2, 1])
+    if a.shape.is_fully_defined():
+      x, y, z = a.shape
+    else:
+      x, y, z = [
+          array_ops.reshape(i, [])
+          for i in array_ops.split(array_ops.shape(a), 3)
+      ]
+    a = array_ops.reshape(a, [x * y, z])
+    prod = math_ops.matmul(a, b, transpose_b=tr_b)
+    return wrap(array_ops.reshape(prod, [x, y, -1]), True)
+  else:
+    assert b_stacked
+    if tr_b:
+      perm = [2, 0, 1]
+      b = array_ops.transpose(b, perm)
+    else:
+      # As an optimization, if one of the first two dimensions is 1, then we can
+      # reshape instead of transpose.
+      # TODO(agarwal): This check can be done inside Transpose kernel.
+      b_shape = array_ops.shape(b)
+      min_dim = math_ops.minimum(b_shape[0], b_shape[1])
+      perm = control_flow_ops.cond(
+          math_ops.equal(min_dim, 1), lambda: [0, 1, 2], lambda: [1, 0, 2])
+      new_shape = array_ops.stack([b_shape[1], b_shape[0], b_shape[2]])
+      b = array_ops.transpose(b, perm)
+      b = array_ops.reshape(b, new_shape)
+
+    if b.shape.is_fully_defined():
+      x, y, z = b.shape
+    else:
+      x, y, z = [
+          array_ops.reshape(i, [])
+          for i in array_ops.split(array_ops.shape(b), 3)
+      ]
+    b = array_ops.reshape(b, [x, y * z])
+    prod = math_ops.matmul(a, b, transpose_a=tr_a)
+    prod = array_ops.reshape(prod, [-1, y, z])
+    prod = array_ops.transpose(prod, [1, 0, 2])
+    return wrap(prod, True)
+
+
+@RegisterPFor("BatchMatMul")
+def _convert_batch_mat_mul(pfor_input):
+  # TODO(agarwal): There may be a more efficient way to do this instead of
+  # stacking the inputs.
+  pfor_input.stack_inputs()
+  x = pfor_input.stacked_input(0)
+  y = pfor_input.stacked_input(1)
+  adj_x = pfor_input.get_attr("adj_x")
+  adj_y = pfor_input.get_attr("adj_y")
+
+  x = _flatten_first_two_dims(x)
+  y = _flatten_first_two_dims(y)
+  output = math_ops.matmul(x, y, adjoint_a=adj_x, adjoint_b=adj_y)
+  output = _unflatten_first_dim(output, pfor_input.pfor.loop_len_vector)
+  return wrap(output, True)
+
+
+@RegisterPForWithArgs("Sum", math_ops.reduce_sum)
+@RegisterPForWithArgs("Prod", math_ops.reduce_prod)
+@RegisterPForWithArgs("Max", math_ops.reduce_max)
+@RegisterPForWithArgs("Min", math_ops.reduce_min)
+def _convert_reduction(pfor_input, _, op_func):
+  t = pfor_input.stacked_input(0)
+  indices = pfor_input.unstacked_input(1)
+  # Shift positive indices by one to account for the extra dimension.
+  indices += math_ops.cast(indices >= 0, dtypes.int32)
+  keep_dims = pfor_input.get_attr("keep_dims")
+  return wrap(op_func(t, indices, keepdims=keep_dims), True)
+
+
+@RegisterPForWithArgs("Cumsum", math_ops.cumsum)
+@RegisterPForWithArgs("Cumprod", math_ops.cumprod)
+def _convert_cumfoo(pfor_input, _, op_func):
+  t = pfor_input.stacked_input(0)
+  axis = pfor_input.unstacked_input(1)
+  # Shift positive indices by one to account for the extra dimension.
+  axis += math_ops.cast(axis >= 0, dtypes.int32)
+  exclusive = pfor_input.get_attr("exclusive")
+  reverse = pfor_input.get_attr("reverse")
+  return wrap(op_func(t, axis, exclusive=exclusive, reverse=reverse), True)
+
+
+@RegisterPFor("BiasAdd")
+def _convert_biasadd(pfor_input):
+  t = pfor_input.stacked_input(0)
+  bias = pfor_input.unstacked_input(1)
+  data_format = pfor_input.get_attr("data_format")
+  if data_format != b"NCHW":
+    return wrap(nn_ops.bias_add(t, bias, data_format=data_format), True)
+  shape = array_ops.shape(t)
+  flattened_shape = array_ops.concat([[-1], shape[2:]], axis=0)
+  t = array_ops.reshape(t, flattened_shape)
+  t = nn_ops.bias_add(t, bias, data_format=b"NCHW")
+  t = array_ops.reshape(t, shape)
+  return wrap(t, True)
+
+
+@RegisterPFor("UnsortedSegmentSum")
+def _convert_unsortedsegmentsum(pfor_input):
+  data, data_stacked, _ = pfor_input.input(0)
+  # TODO(agarwal): handle unstacked?
+  segment_ids = pfor_input.stacked_input(1)
+  # TODO(agarwal): handle stacked?
+  num_segments = pfor_input.unstacked_input(2)
+  if not data_stacked:
+    data = _stack(data, pfor_input.pfor.loop_len_vector).t
+  segment_shape = array_ops.shape(segment_ids)
+  n = segment_shape[0]
+  ones = array_ops.ones_like(segment_shape)[1:]
+  segment_offset = num_segments * math_ops.range(n)
+  segment_offset = array_ops.reshape(segment_offset,
+                                     array_ops.concat([[n], ones], axis=0))
+  segment_ids += segment_offset
+  num_segments *= n
+  output = math_ops.unsorted_segment_sum(data, segment_ids, num_segments)
+  new_output_shape = array_ops.concat(
+      [[n, -1], array_ops.shape(output)[1:]], axis=0)
+  output = array_ops.reshape(output, new_output_shape)
+  return wrap(output, True)
+
+
+@RegisterPFor("Cast")
+def _convert_cast(pfor_input):
+  inp = pfor_input.stacked_input(0)
+  dtype = pfor_input.get_attr("DstT")
+  return wrap(math_ops.cast(inp, dtype), True)
+
+
+# Note that ops handled here do not have attributes except "T", and hence don't
+# need extra arguments passed to the cwise_op call below.
+@RegisterPForWithArgs("Add", math_ops.add)
+@RegisterPForWithArgs("Ceil", math_ops.ceil)
+@RegisterPForWithArgs("Equal", math_ops.equal)
+@RegisterPForWithArgs("NotEqual", math_ops.not_equal)
+@RegisterPForWithArgs("Floor", math_ops.floor)
+@RegisterPForWithArgs("Greater", math_ops.greater)
+@RegisterPForWithArgs("GreaterEqual", math_ops.greater_equal)
+@RegisterPForWithArgs("Less", math_ops.less)
+@RegisterPForWithArgs("LessEqual", math_ops.less_equal)
+@RegisterPForWithArgs("LogicalOr", math_ops.logical_or)
+@RegisterPForWithArgs("LogicalAnd", math_ops.logical_and)
+@RegisterPForWithArgs("LogicalNot", math_ops.logical_not)
+@RegisterPForWithArgs("LogicalXor", math_ops.logical_xor)
+@RegisterPForWithArgs("Maximum", math_ops.maximum)
+@RegisterPForWithArgs("Minimum", math_ops.minimum)
+@RegisterPForWithArgs("Mul", math_ops.multiply)
+@RegisterPForWithArgs("Neg", math_ops.negative)
+@RegisterPForWithArgs("RealDiv", math_ops.divide)
+@RegisterPForWithArgs("Relu", nn_ops.relu)
+@RegisterPForWithArgs("Sigmoid", math_ops.sigmoid)
+@RegisterPForWithArgs("Square", math_ops.square)
+@RegisterPForWithArgs("Sub", math_ops.subtract)
+@RegisterPForWithArgs("Tanh", math_ops.tanh)
+def _convert_cwise(pfor_input, op_type, op_func):
+  del op_type
+  pfor_input.expanddim_inputs_for_broadcast()
+  return wrap(op_func(*[x.t for x in pfor_input.inputs]), True)
+
+
+@RegisterPFor("Shape")
+def _convert_shape(pfor_input):
+  out_type = pfor_input.get_attr("out_type")
+  return wrap(
+      array_ops.shape(pfor_input.stacked_input(0), out_type=out_type)[1:],
+      False)
+
+
+@RegisterPFor("ShapeN")
+def _convert_shape_n(pfor_input):
+  out_type = pfor_input.get_attr("out_type")
+  shapes = [
+      array_ops.shape(x, out_type=out_type)[1:]
+      if stacked else array_ops.shape(x) for x, stacked, _ in pfor_input.inputs
+  ]
+  return [wrap(x, False) for x in shapes]
+
+
+@RegisterPFor("Size")
+def _convert_size(pfor_input):
+  out_type = pfor_input.get_attr("out_type")
+  n = math_ops.cast(pfor_input.pfor.loop_len_vector[0], out_type)
+  return wrap(
+      array_ops.size(pfor_input.stacked_input(0), out_type=out_type) // n,
+      False)
+
+
+@RegisterPFor("Rank")
+def _convert_rank(pfor_input):
+  return wrap(array_ops.rank(pfor_input.stacked_input(0)) - 1, False)
+
+
+@RegisterPFor("AddN")
+def _convert_addn(pfor_input):
+  # AddN does not support broadcasting.
+  pfor_input.stack_inputs()
+  return wrap(math_ops.add_n([x.t for x in pfor_input.inputs]), True)
+
+
+@RegisterPFor("BiasAddGrad")
+def _convert_biasaddgrad(pfor_input):
+  grad = pfor_input.stacked_input(0)
+  fmt = pfor_input.get_attr("data_format")
+  if fmt == b"NCHW":
+    output = math_ops.reduce_sum(grad, axis=[1, 3, 4], keepdims=False)
+  else:
+    grad_shape = array_ops.shape(grad)
+    last_dim_shape = grad_shape[-1]
+    first_dim_shape = grad_shape[0]
+    output = array_ops.reshape(grad, [first_dim_shape, -1, last_dim_shape])
+    output = math_ops.reduce_sum(output, axis=[1], keepdims=False)
+  return wrap(output, True)
+
+
+# Some required ops are not exposed under the tf namespace. Hence relying on
+# _create_op to create them.
+@RegisterPForWithArgs("ReluGrad")
+@RegisterPForWithArgs("TanhGrad")
+@RegisterPForWithArgs("SigmoidGrad")
+def _convert_grads(pfor_input, op_type, *args, **kw_args):
+  del args
+  del kw_args
+  # TODO(agarwal): Looks like these ops don't support broadcasting. Hence we
+  # have to use tiling here.
+  pfor_input.stack_inputs()
+  outputs = _create_op(
+      op_type, [x.t for x in pfor_input.inputs],
+      [x.dtype for x in pfor_input.outputs],
+      attrs=pfor_input.op.node_def.attr).outputs
+  return [wrap(x, True) for x in outputs]
+
+
+@RegisterPFor("Select")
+def _convert_select(pfor_input):
+  pfor_input.stack_inputs()
+  cond = pfor_input.stacked_input(0)
+  t = pfor_input.stacked_input(1)
+  e = pfor_input.stacked_input(2)
+  cond_rank = array_ops.rank(cond)
+  cond, t, e = control_flow_ops.cond(
+      cond_rank > 1, lambda: _inputs_with_flattening(pfor_input, [0, 1, 2]),
+      lambda: [cond, t, e])
+  outputs = _create_op(
+      pfor_input.op_type, [cond, t, e], [x.dtype for x in pfor_input.outputs],
+      attrs=pfor_input.op.node_def.attr).outputs
+  n = pfor_input.pfor.loop_len_vector
+  out = control_flow_ops.cond(cond_rank > 1,
+                              lambda: _unflatten_first_dim(outputs[0], n),
+                              lambda: outputs[0])
+  return [wrap(out, True) for x in outputs]
+
+
+# random_ops
+
+
+@RegisterPForWithArgs("RandomUniform")
+@RegisterPForWithArgs("RandomUniformInt")
+@RegisterPForWithArgs("RandomStandardNormal")
+@RegisterPForWithArgs("TruncatedNormal")
+@RegisterPForWithArgs("RandomGamma")
+@RegisterPForWithArgs("RandomPoissonV2")
+def _convert_random(pfor_input, op_type, *args, **kw_args):
+  del args
+  del kw_args
+  inputs = [pfor_input.unstacked_input(i) for i in range(pfor_input.num_inputs)]
+  # inputs[0] is "shape"
+  inputs[0] = array_ops.concat(
+      [pfor_input.pfor.loop_len_vector, inputs[0]], axis=0)
+  logging.warning(
+      "Note that %s inside pfor op may not give same output as "
+      "inside a sequential loop.", op_type)
+  outputs = _create_op(
+      op_type,
+      inputs, [x.dtype for x in pfor_input.outputs],
+      attrs=pfor_input.op.node_def.attr).outputs
+  return [wrap(x, True) for x in outputs]
+
+
+# logging_ops
+
+
+@RegisterPFor("Assert")
+def _convert_assert(pfor_input):
+  cond, cond_stacked, _ = pfor_input.input(0)
+  if cond_stacked:
+    cond = math_ops.reduce_all(cond)
+
+  data_list = [x.t for x in pfor_input.inputs][1:]
+  return _create_op("Assert", [cond] + data_list, [],
+                    attrs=pfor_input.op.node_def.attr)
+
+
+@RegisterPFor("Print")
+def _convert_print(pfor_input):
+  # Note that we don't stack all the inputs. Hence unstacked values are printed
+  # once here vs multiple times in a while_loop.
+  pfor_input.stack_inputs([0])
+  outputs = _create_op(
+      "Print", [x.t for x in pfor_input.inputs],
+      [x.dtype for x in pfor_input.outputs],
+      attrs=pfor_input.op.node_def.attr).outputs
+  return [wrap(x, True) for x in outputs]
+
+
+# data_flow_ops
+
+# TensorArray conversion is tricky since we don't support arrays of
+# TensorArrays. For converting them, we consider two distinct cases:
+#
+# 1. The array is constructed outside the pfor call, and read/written inside the
+# loop.
+# This is an easier case since we don't need to make an array of TensorArrays.
+# A correctness requirement is that these parallel iterations shouldn't attempt
+# to write to the same location. Hence at conversion time we disallow indices to
+# be loop-invariant as that would guarantee a collision. Even if the indices are
+# not loop-invariant, they could conflict and that shall trigger runtime errors.
+#
+# 2. The array is constructed and used entirely inside each pfor iteration.
+# For simplicity, here we require that the indices used for write/scatter are
+# "unstacked". Otherwise it becomes hard to merge the TensorArrays created in
+# different pfor iterations. We consider two sub_cases:
+#
+# 2a Elements written to the array are "stacked"
+# To simulate multiple TensorArrays, we may increase the dimension of each
+# element of the array. i.e. the i_th row of the j_th entry of the converted
+# TensorArray corresponds to to the j_th entry of the TensorArray in the i_th
+# pfor iteration.
+#
+# 2b Elements written to the array are "unstacked"
+# In this case we don't increase the dimensions to avoid redundant tiling. Each
+# iteration is trying to write the same value. So we convert that to a single
+# write.
+#
+# Here are some tricks used to implement the above:
+# - TensorArrayV3 constructor encodes the element shape as an attr. Instead of
+# trying to trace whether future writes are stacked or unstacked in order to set
+# this attr, we set it to correspond to unknown shape.
+# - We use the "flow" output of the different ops to track whether the array
+# elements are stacked or unstacked. If a stacked write/scatter is done, we make
+# the flow stacked as well.
+# - We use some heuristic traversal of the graph to track whether the
+# TensorArray handle was created inside or outside the pfor loop.
+
+
+@RegisterPFor("TensorArrayV3")
+def _convert_tensor_array_v3(pfor_input):
+  size = pfor_input.unstacked_input(0)
+  dtype = pfor_input.get_attr("dtype")
+  dynamic_size = pfor_input.get_attr("dynamic_size")
+  clear_after_read = pfor_input.get_attr("clear_after_read")
+  identical_element_shapes = pfor_input.get_attr("identical_element_shapes")
+  tensor_array_name = pfor_input.get_attr("tensor_array_name")
+  handle, flow = data_flow_ops.tensor_array_v3(
+      size,
+      dtype=dtype,
+      # We don't set element shape since we don't know if writes are stacked or
+      # not yet.
+      element_shape=None,
+      dynamic_size=dynamic_size,
+      clear_after_read=clear_after_read,
+      identical_element_shapes=identical_element_shapes,
+      tensor_array_name=tensor_array_name)
+  # Note we keep flow unstacked for now since we don't know if writes will be
+  # stacked or not.
+  return wrap(handle, False), wrap(flow, False)
+
+
+@RegisterPFor("TensorArraySizeV3")
+def _convert_tensor_array_size_v3(pfor_input):
+  handle = pfor_input.unstacked_input(0)
+  flow, flow_stacked, _ = pfor_input.input(1)
+  if flow_stacked:
+    flow = _unstack_flow(flow)
+  size = data_flow_ops.tensor_array_size_v3(handle, flow)
+  return wrap(size, False)
+
+
+def _handle_inside_pfor(pfor_input, handle):
+  """Returns True if handle was created inside the pfor loop."""
+  # We use some heuristic to find the original TensorArray creation op.
+  # The logic should handle the common cases (except cond based subgraphs).
+  # In theory the user could perform different operations on the handle (like
+  # Reshape, stack multiple handles, etc) which could break this logic.
+  # TODO(agarwal): handle Switch/Merge.
+  while handle.op.type in ("Enter", "Identity"):
+    handle = handle.op.inputs[0]
+  if handle.op.type not in [
+      "TensorArrayV3", "TensorArrayGradV3", "TensorArrayGradWithShape"]:
+    raise ValueError("Unable to find source for handle %s" % handle)
+  else:
+    return pfor_input.pfor.op_is_inside_loop(handle.op)
+
+
+def _unstack_flow(value):
+  # TODO(agarwal): consider looking if this is a Tile op then get its input.
+  # This may avoid running the Tile operations.
+  return array_ops.gather(value, 0)
+
+
+@RegisterPFor("TensorArrayReadV3")
+def _convert_tensor_array_read_v3(pfor_input):
+  handle = pfor_input.unstacked_input(0)
+  index, index_stacked, _ = pfor_input.input(1)
+  dtype = pfor_input.get_attr("dtype")
+  flow, flow_stacked, _ = pfor_input.input(2)
+  if flow_stacked:
+    flow = _unstack_flow(flow)
+
+  is_inside_pfor = _handle_inside_pfor(pfor_input, pfor_input.op.inputs[0])
+  if is_inside_pfor:
+    # Note that if we are inside a control flow construct inside the pfor, and
+    # only some of the iterations are doing the read (i.e.
+    # `all_indices_partitioned` is True), then the read operation should only
+    # return values for the currently active pfor iterations (`all_indices`
+    # below). Hence, whenever the returned value is stacked (i.e. `flow` is
+    # stacked), we may need to do an extra gather after reading the values. Also
+    # note that if `is_inside` is false, then values in the tensor array are
+    # unstacked. So the check is only needed in this branch.
+    all_indices = pfor_input.pfor.all_indices
+    all_indices_partitioned = pfor_input.pfor.all_indices_partitioned
+    # Note: flow_stacked indicates if values in the TensorArray are stacked or
+    # not.
+    if index_stacked:
+      if flow_stacked:
+        raise ValueError(
+            "It looks like TensorArrayReadV3 was called on a TensorArray whose"
+            " values are not loop-invariant, and the read indices were also"
+            " not loop invariant. This is currently unsupported.")
+      value = data_flow_ops.tensor_array_gather_v3(
+          handle, index, flow, dtype=dtype)
+      return wrap(value, True)
+    value = data_flow_ops.tensor_array_read_v3(
+        handle, index, flow, dtype=dtype)
+    if flow_stacked and all_indices_partitioned:
+      value = array_ops.gather(value, all_indices)
+    return wrap(value, flow_stacked)
+  # Values in the TensorArray should be unstacked (since different iterations
+  # couldn't write to the same location). So whether output is stacked or not
+  # depends on index_stacked.
+  if index_stacked:
+    value = data_flow_ops.tensor_array_gather_v3(
+        handle, index, flow, dtype=dtype)
+  else:
+    value = data_flow_ops.tensor_array_read_v3(
+        handle, index, flow, dtype=dtype)
+  return wrap(value, index_stacked)
+
+
+@RegisterPFor("TensorArrayWriteV3")
+def _convert_tensor_array_write_v3(pfor_input):
+  handle = pfor_input.unstacked_input(0)
+  index, index_stacked, _ = pfor_input.input(1)
+  value, value_stacked, _ = pfor_input.input(2)
+  flow, flow_stacked, _ = pfor_input.input(3)
+  if value_stacked and pfor_input.pfor.all_indices_partitioned:
+    # Looks like we are in a control flow in a pfor where not all iterations are
+    # active now. We don't allow that since that could lead to different indices
+    # having different shapes which will be hard to merge later.
+    raise ValueError("Writing non loop invariant values to TensorArray from "
+                     "inside a while_loop/cond not supported.")
+  if flow_stacked:
+    flow = _unstack_flow(flow)
+  is_inside = _handle_inside_pfor(pfor_input, pfor_input.op.inputs[0])
+  if is_inside:
+    if index_stacked:
+      raise ValueError("Need indices for %s to be loop invariant" % handle)
+    if not flow_stacked and not value_stacked:
+      flow_out = data_flow_ops.tensor_array_write_v3(handle, index, value, flow)
+      return wrap(flow_out, False)
+    else:
+      if not value_stacked:
+        value = _stack(value, pfor_input.pfor.loop_len_vector).t
+      # TODO(agarwal): Note that if flow is unstacked and value is stacked, then
+      # this may or may not be a safe situation. flow is unstacked both for a
+      # freshly created TensorArray, as well as after unstacked values are
+      # written to it. If it is the latter, then we cannot write a stacked value
+      # now since that may cause runtime errors due to different shapes in the
+      # array. At the moment we are not able to handle this gracefully and
+      # distinguish between the two cases. That would require some heuristic
+      # traversal of the graph to figure out whether all the writes are
+      # unstacked or not.
+      flow_out = data_flow_ops.tensor_array_write_v3(handle, index, value, flow)
+      return _stack(flow_out, pfor_input.pfor.loop_len_vector)
+  else:
+    if not index_stacked:
+      raise ValueError("Need indices for %s to be not loop invariant" % handle)
+    # Note that even when index_stacked is true, actual values in index may
+    # still not be unique. However that will cause runtime error when executing
+    # the scatter operation below.
+    if not value_stacked:
+      value = _stack(value, pfor_input.pfor.loop_len_vector).t
+    flow_out = data_flow_ops.tensor_array_scatter_v3(handle, index, value, flow)
+    return _stack(flow_out, pfor_input.pfor.loop_len_vector)
+
+
+def _transpose_first_two_dims(value):
+  # TODO(agarwal): optimize if one of the dims == 1.
+  value_shape = array_ops.shape(value)
+  v0 = value_shape[0]
+  v1 = value_shape[1]
+  value = array_ops.reshape(value, [v0, v1, -1])
+  value = array_ops.transpose(value, [1, 0, 2])
+  new_shape = array_ops.concat([[v1, v0], value_shape[2:]], axis=0)
+  return array_ops.reshape(value, new_shape)
+
+
+@RegisterPFor("TensorArrayGatherV3")
+def _convert_tensor_array_gather_v3(pfor_input):
+  handle = pfor_input.unstacked_input(0)
+  indices, indices_stacked, _ = pfor_input.input(1)
+  indices = array_ops.reshape(indices, [-1])
+  flow, flow_stacked, _ = pfor_input.input(2)
+  if flow_stacked:
+    flow = _unstack_flow(flow)
+  dtype = pfor_input.get_attr("dtype")
+  # TODO(agarwal): support element_shape attr?
+
+  n = pfor_input.pfor.loop_len_vector
+  value = data_flow_ops.tensor_array_gather_v3(
+      handle, indices, flow, dtype=dtype)
+  is_inside = _handle_inside_pfor(pfor_input, pfor_input.op.inputs[0])
+  if is_inside:
+    # flow_stacked indicates if values in the TensorArray are stacked or not.
+    if indices_stacked:
+      if flow_stacked:
+        raise ValueError(
+            "It looks like TensorArrayGatherV3 was called on a TensorArray "
+            "whose values are not loop-invariant, and the indices were also "
+            "not loop invariant. This is currently unsupported.")
+      else:
+        value = _unflatten_first_dim(value, n)
+        return wrap(value, True)
+    else:
+      if flow_stacked:
+        # Since elements in this array are stacked and `value` was produced by
+        # gather, its first two dims are "gathered elements" and "stack
+        # dimension". Our semantics require these two to be flipped.
+        value = _transpose_first_two_dims(value)
+      return wrap(value, flow_stacked)
+  else:
+    # Values in the TensorArray should be unstacked (since different iterations
+    # couldn't write to the same location). So whether output is stacked or not
+    # depends on indices_stacked.
+    if indices_stacked:
+      value = _unflatten_first_dim(value, n)
+    return wrap(value, indices_stacked)
+
+
+@RegisterPFor("TensorArrayScatterV3")
+def _convert_tensor_array_scatter_v3(pfor_input):
+  handle = pfor_input.unstacked_input(0)
+  indices, indices_stacked, _ = pfor_input.input(1)
+  indices = array_ops.reshape(indices, [-1])
+  value, value_stacked, _ = pfor_input.input(2)
+  flow, flow_stacked, _ = pfor_input.input(3)
+
+  if flow_stacked:
+    flow = _unstack_flow(flow)
+
+  is_inside = _handle_inside_pfor(pfor_input, pfor_input.op.inputs[0])
+  if is_inside:
+    if indices_stacked:
+      raise ValueError("Need indices for %s to be loop invariant" % handle)
+    # Note that flow_stacked indicates if existing values in the array are
+    # stacked or not.
+    if not flow_stacked and not value_stacked:
+      flow_out = data_flow_ops.tensor_array_scatter_v3(handle, indices, value,
+                                                       flow)
+      return wrap(flow_out, False)
+    if not value_stacked:
+      # TODO(agarwal): tile in the second dimension directly instead of
+      # transposing below.
+      value = _stack(value, pfor_input.pfor.loop_len_vector).t
+
+    value = _transpose_first_two_dims(value)
+    # TODO(agarwal): Note that if a previous write was unstacked, flow will be
+    # unstacked, and a stacked value may be written here which may cause
+    # runtime error due to different elements having different shape. We do
+    # not try to prevent that.
+    flow_out = data_flow_ops.tensor_array_scatter_v3(handle, indices, value,
+                                                     flow)
+    return _stack(flow_out, pfor_input.pfor.loop_len_vector)
+  if not indices_stacked:
+    raise ValueError("Need indices for %s to be not loop invariant" % handle)
+  if not value_stacked:
+    value = _stack(value, pfor_input.pfor.loop_len_vector).t
+  value = _flatten_first_two_dims(value)
+  flow_out = data_flow_ops.tensor_array_scatter_v3(handle, indices, value,
+                                                   flow)
+  return _stack(flow_out, pfor_input.pfor.loop_len_vector)
+
+
+@RegisterPFor("TensorArrayGradV3")
+def _convert_tensor_array_grad_v3(pfor_input):
+  handle = pfor_input.unstacked_input(0)
+  flow, flow_stacked, _ = pfor_input.input(1)
+  if flow_stacked:
+    flow = _unstack_flow(flow)
+  source = pfor_input.get_attr("source")
+  # TODO(agarwal): For now, we assume that gradients are stacked if the
+  # TensorArrayGradV3 call is being done inside the pfor. Getting that wrong
+  # will give runtime error due to incorrect shape being written to the
+  # accumulator. It is difficult to know in advance if gradients written will be
+  # stacked or not. Note that flow being stacked is not indicative of the
+  # gradient being stacked or not. Revisit this later.
+  shape_to_prepend = pfor_input.pfor.loop_len_vector
+  grad_handle, flow_out = data_flow_ops.tensor_array_grad_with_shape(
+      handle=handle,
+      flow_in=flow,
+      shape_to_prepend=shape_to_prepend,
+      source=source)
+  flow_out = _stack(flow_out, pfor_input.pfor.loop_len_vector).t
+  return [wrap(grad_handle, False), wrap(flow_out, True)]
+
+
+# StackV2 conversion is tricky since we don't have arrays of StackV2. So similar
+# to TensorArrays, we convert them by changing the dimension of the elements
+# inside the stack.
+#
+# We consider two cases:
+#
+# 1. StackV2 is constructed and used entirely inside the pfor loop.
+# We keep a single Stack and perform the push/pop operations of all the
+# iterations in lock-step. We also assume that all the iterations perform these
+# operations. In case of dynamic control flow, if only some of the iterations
+# try to perform a push/pop, then the conversion may not work correctly and may
+# cause undefined behavior.
+# TODO(agarwal): test StackV2 with dynamic control flow.
+#
+# 2. StackV2 is constructed outside the pfor loop.
+# Performing stack push/pop in a parallel fashion is ill-defined. However given
+# that reading stacks created externally is a common operation when computing
+# jacobians, we provide some special semantics here as follows.
+#  - disallow push operations to the stack
+#  - pop operations are performed in lock step by all iterations, similar to the
+#  case when the stack is created inside. A single value is popped during the
+#  lock-step operation and broadcast to all the iterations. Values in the stack
+#  are assumed to be loop-invariant.
+#
+# Some other implementation details:
+# We use an ugly logic to find whether values in Stack data structure are
+# loop invariant or not. When converting push/pop operations, we keep track of
+# whether the last conversion used a stacked value or not (see _stack_cache
+# below). As a result if an unstacked value is written first, subsequent stacked
+# writes are disallowed when they could have been allowed in theory.
+
+# Map from cache key based on StackV2 handle to a bool indicating whether values
+# are stacked or not.
+# TODO(agarwal): move _stack_cache inside pfor?
+_stack_cache = {}
+
+
+def _stack_cache_key(pfor_input):
+  """Create cache key corresponding to a stack handle."""
+  op_type = pfor_input.op_type
+  assert op_type in ["StackPushV2", "StackPopV2"], op_type
+  orig_handle = pfor_input.op.inputs[0]
+  while orig_handle.op.type in ["Identity", "Enter"]:
+    orig_handle = orig_handle.op.inputs[0]
+  assert orig_handle.op.type == "StackV2", orig_handle.op
+  return ops.get_default_graph(), pfor_input.pfor, orig_handle
+
+
+def _stack_handle_inside_pfor(handle, pfor_input):
+  while handle.op.type in ["Identity", "Enter"]:
+    handle = handle.op.inputs[0]
+  assert handle.op.type == "StackV2", (
+      "Unable to find StackV2 op. Got %s" % handle.op)
+  return pfor_input.pfor.op_is_inside_loop(handle.op)
+
+
+@RegisterPFor("StackPushV2")
+def _convert_stack_push_v2(pfor_input):
+  handle = pfor_input.unstacked_input(0)
+  elem, elem_stacked, _ = pfor_input.input(1)
+  swap_memory = pfor_input.get_attr("swap_memory")
+
+  if not _stack_handle_inside_pfor(pfor_input.op.inputs[0], pfor_input):
+    raise ValueError("StackPushV2 not allowed on stacks created outside pfor")
+  stack_cache_key = _stack_cache_key(pfor_input)
+  stacked = _stack_cache.get(stack_cache_key, None)
+  if stacked is None:
+    stacked = elem_stacked
+    _stack_cache[stack_cache_key] = stacked
+  else:
+    # If we previously made it unstacked then we can't revert to being stacked.
+    if not stacked and elem_stacked:
+      raise ValueError(
+          "It looks like the stack was previously determined to be loop"
+          " invariant, but we are now trying to push a loop dependent value"
+          " to it. This is currently unsupported.")
+    if stacked and not elem_stacked:
+      elem = _stack(elem, pfor_input.pfor.loop_len_vector).t
+  out = data_flow_ops.stack_push_v2(handle, elem, swap_memory=swap_memory)
+  return wrap(out, stacked)
+
+
+# Note that inputs to this convertor will be unstacked. However it should get
+# called since it is a stateful op.
+@RegisterPFor("StackPopV2")
+def _convert_stack_pop_v2(pfor_input):
+  handle = pfor_input.unstacked_input(0)
+  stack_cache_key = _stack_cache_key(pfor_input)
+  stacked = _stack_cache.get(stack_cache_key, None)
+  # If a StackPushV2 has not been converted yet, we default to unstacked since
+  # the push could be outside of pfor, or the covertor may not be called if the
+  # inputs are unconverted.
+  if stacked is None:
+    stacked = False
+    _stack_cache[stack_cache_key] = False
+  elem_type = pfor_input.get_attr("elem_type")
+  out = data_flow_ops.stack_pop_v2(handle, elem_type)
+  return wrap(out, stacked)
+
+
+# parsing_ops
+
+
+@RegisterPFor("DecodeCSV")
+def _convert_decode_csv(pfor_input):
+  lines = pfor_input.stacked_input(0)
+  record_defaults = [
+      pfor_input.unstacked_input(i) for i in range(1, pfor_input.num_inputs)
+  ]
+  field_delim = pfor_input.get_attr("field_delim")
+  use_quote_delim = pfor_input.get_attr("use_quote_delim")
+  select_cols = pfor_input.get_attr("select_cols")
+  if not select_cols:
+    select_cols = None
+  return [
+      wrap(t, True) for t in parsing_ops.decode_csv(
+          lines,
+          record_defaults,
+          field_delim=field_delim,
+          use_quote_delim=use_quote_delim,
+          select_cols=select_cols)
+  ]
+
+
+@RegisterPFor("ParseSingleExample")
+def _convert_parse_single_example(pfor_input):
+  serialized = pfor_input.stacked_input(0)
+  dense_defaults = [
+      pfor_input.unstacked_input(i) for i in range(1, pfor_input.num_inputs)
+  ]
+  sparse_keys = pfor_input.get_attr("sparse_keys")
+  dense_keys = pfor_input.get_attr("dense_keys")
+  sparse_types = pfor_input.get_attr("sparse_types")
+  dense_shapes = pfor_input.get_attr("dense_shapes")
+  output = gen_parsing_ops.parse_example(
+      serialized=serialized,
+      names=[],
+      dense_defaults=dense_defaults,
+      sparse_keys=sparse_keys,
+      dense_keys=dense_keys,
+      sparse_types=sparse_types,
+      dense_shapes=dense_shapes)
+  return [wrap(t, True, True) for t in nest.flatten(output)]
diff --git a/tensorflow/python/ops/resource_variable_ops.py b/tensorflow/python/ops/resource_variable_ops.py
index 2033674a92df32b92c37ceb022a64683ae91b08e..70a89e5ebbb25c74cc5a414d0f2de5d74d070e2d 100644
--- a/tensorflow/python/ops/resource_variable_ops.py
+++ b/tensorflow/python/ops/resource_variable_ops.py
@@ -867,6 +867,19 @@ class ResourceVariable(variables.Variable):
 
   __array_priority__ = 100
 
+  def is_initialized(self, name=None):
+    """Checks whether a resource variable has been initialized.
+
+    Outputs boolean scalar indicating whether the tensor has been initialized.
+
+    Args:
+      name: A name for the operation (optional).
+
+    Returns:
+      A `Tensor` of type `bool`.
+    """
+    return gen_resource_variable_ops.var_is_initialized_op(self.handle, name)
+
   def assign_sub(self, delta, use_locking=None, name=None, read_value=True):
     """Subtracts a value from this variable.
 
@@ -999,32 +1012,28 @@ class ResourceVariable(variables.Variable):
 
   def __imul__(self, unused_other):
     raise RuntimeError("Variable *= value not supported. Use "
-                       "variable.assign_mul(value) to modify the variable "
-                       "value and variable = variable * value to get a new "
-                       "Tensor object.")
+                       "`var.assign(var * value)` to modify the variable or "
+                       "`var = var * value` to get a new Tensor object.")
 
   def __idiv__(self, unused_other):
     raise RuntimeError("Variable /= value not supported. Use "
-                       "variable.assign_div(value) to modify the variable "
-                       "value and variable = variable / value to get a new "
-                       "Tensor object.")
+                       "`var.assign(var / value)` to modify the variable or "
+                       "`var = var / value` to get a new Tensor object.")
 
   def __itruediv__(self, unused_other):
     raise RuntimeError("Variable /= value not supported. Use "
-                       "variable.assign_div(value) to modify the variable "
-                       "value and variable = variable / value to get a new "
-                       "Tensor object.")
+                       "`var.assign(var / value)` to modify the variable or "
+                       "`var = var / value` to get a new Tensor object.")
 
   def __irealdiv__(self, unused_other):
     raise RuntimeError("Variable /= value not supported. Use "
-                       "variable.assign_div(value) to modify the variable "
-                       "value and variable = variable / value to get a new "
-                       "Tensor object.")
+                       "`var.assign(var / value)` to modify the variable or "
+                       "`var = var / value` to get a new Tensor object.")
 
   def __ipow__(self, unused_other):
     raise RuntimeError("Variable **= value not supported. Use "
-                       "value and variable = variable ** value to get a new "
-                       "Tensor object.")
+                       "`var.assign(var ** value)` to modify the variable or "
+                       "`var = var ** value` to get a new Tensor object.")
 
 
 pywrap_tensorflow.TFE_Py_RegisterResourceVariableType(ResourceVariable)
@@ -1095,6 +1104,113 @@ class _UnreadVariable(ResourceVariable):
 ops.register_tensor_conversion_function(_UnreadVariable, _dense_var_to_tensor)
 ops.register_dense_tensor_like_type(_UnreadVariable)
 
+
+class _MixedPrecisionVariable(ResourceVariable):
+  """Represents a variable that can return in desired dtype when read.
+
+  In mixed precision training, it is usually desirable to use different dtypes
+  for variables and computation. This class will be used to wrap created
+  ResourceVariable when mixed precision training is enabled. It allows layers to
+  perform computation in a different dtype than their variable dtypes, in order
+  to achieve higher performance without causing quality loss.
+  """
+
+  def __init__(self, var, read_dtype):
+    """Creates a MixedPrecisionVariable.
+
+    Args:
+      var: A ResourceVariable instance.
+      read_dtype: A tf.DType, the returned dtype when read, default to None.
+        Casting is performed if read_dtype is not None and differs from
+        var.dtype.
+    Returns:
+      An MixedPrecisionVariable instance.
+    Raises:
+      ValueError: if var is not a ResourceVariable instance, or read_dtype is
+        not a tf.DType instance.
+    """
+    # pylint: disable=super-init-not-called
+    # We do not call super init on purpose.
+    if not isinstance(var, ResourceVariable):
+      raise ValueError("InvalidArgument: var must be a ResourceVariable type.")
+    if not isinstance(read_dtype, dtypes.DType):
+      raise ValueError("InvalidArgument: read_dtype must be a tf.DType type.")
+
+    self._var = var
+    self._trainable = var.trainable
+    self._save_slice_info = None
+    self._graph_key = ops.get_default_graph()._graph_key  # pylint: disable=protected-access
+    self._in_graph_mode = var._in_graph_mode  # pylint: disable=protected-access
+    self._handle = var.handle
+    self._shape = var.shape
+    self._initial_value = None
+    if isinstance(self.handle, ops.EagerTensor):
+      self._handle_name = ""
+    else:
+      self._handle_name = self.handle.name
+    self._unique_id = var._unique_id  # pylint: disable=protected-access
+    self._dtype = var.dtype
+    self._constraint = None
+    self._cached_value = None
+    self._is_initialized_op = var._is_initialized_op  # pylint: disable=protected-access
+    self._initializer_op = var._initializer_op  # pylint: disable=protected-access
+    # This needs to be set before read_value() is called.
+    self._read_dtype = read_dtype
+    if context.executing_eagerly():
+      self._graph_element = None
+    else:
+      self._graph_element = self.read_value()
+    self._handle_deleter = (
+        var._handle_deleter if not self._in_graph_mode  # pylint: disable=protected-access
+        else None)
+    # pylint: enable=super-init-not-called
+
+  @property
+  def name(self):
+    return self._var.name
+
+  def value(self):
+    return self._read_variable_op()
+
+  def read_value(self):
+    return self._read_variable_op()
+
+  def _read_variable_op(self):
+    with ops.colocate_with(self._handle):
+      res = gen_resource_variable_ops.read_variable_op(self._handle,
+                                                       self._dtype)
+      if self._read_dtype != self._dtype:
+        return math_ops.cast(res, self._read_dtype)
+      else:
+        return res
+
+  def set_shape(self, shape):
+    self._shape = shape
+    self._cached_shape_as_list = None
+
+  @property
+  def op(self):
+    """The op for this variable."""
+    return self._var.op
+
+  @property
+  def read_dtype(self):
+    """The dtype of the returned tensor when reading the var."""
+    return self._read_dtype
+
+  def _dense_var_to_tensor(self, dtype=None, name=None, as_ref=False):
+    del name
+    dtype = dtype or self.read_dtype
+    if dtype != self.read_dtype or as_ref:
+      return NotImplemented
+    else:
+      res = self.value()
+    return res
+
+  def _should_act_as_resource_variable(self):
+    """To pass resource_variable_ops.is_resource_variable check."""
+    pass
+
 # Register a conversion function which reads the value of the variable,
 # allowing instances of the class to be used as tensors.
 
diff --git a/tensorflow/python/ops/rnn.py b/tensorflow/python/ops/rnn.py
index 215140e9879d8ea1b5cabc24d04c91a68c9abc2a..deba133fb9910f28c7f902f334174734c3c742f7 100644
--- a/tensorflow/python/ops/rnn.py
+++ b/tensorflow/python/ops/rnn.py
@@ -26,6 +26,7 @@ from tensorflow.python.framework import tensor_shape
 from tensorflow.python.framework import tensor_util
 from tensorflow.python.ops import array_ops
 from tensorflow.python.ops import control_flow_ops
+from tensorflow.python.ops import control_flow_util
 from tensorflow.python.ops import math_ops
 from tensorflow.python.ops import rnn_cell_impl
 from tensorflow.python.ops import tensor_array_ops
@@ -131,6 +132,18 @@ def _maybe_tensor_shape_from_tensor(shape):
     return shape
 
 
+def _should_cache():
+  """Returns True if a default caching device should be set, otherwise False."""
+  if context.executing_eagerly():
+    return False
+  # Don't set a caching device when running in a loop, since it is possible that
+  # train steps could be wrapped in a tf.while_loop. In that scenario caching
+  # prevents forward computations in loop iterations from re-reading the
+  # updated weights.
+  ctxt = ops.get_default_graph()._get_control_flow_context()  # pylint: disable=protected-access
+  return control_flow_util.GetContainingWhileContext(ctxt) is None
+
+
 # pylint: disable=unused-argument
 def _rnn_step(
     time, sequence_length, min_sequence_length, max_sequence_length,
@@ -558,7 +571,7 @@ def dynamic_rnn(cell, inputs, sequence_length=None, initial_state=None,
     # Create a new scope in which the caching device is either
     # determined by the parent scope, or is set to place the cached
     # Variable using the same placement as for the rest of the RNN.
-    if not context.executing_eagerly():
+    if _should_cache():
       if varscope.caching_device is None:
         varscope.set_caching_device(lambda op: op.device)
 
@@ -1015,7 +1028,7 @@ def raw_rnn(cell, loop_fn,
   # determined by the parent scope, or is set to place the cached
   # Variable using the same placement as for the rest of the RNN.
   with vs.variable_scope(scope or "rnn") as varscope:
-    if not context.executing_eagerly():
+    if _should_cache():
       if varscope.caching_device is None:
         varscope.set_caching_device(lambda op: op.device)
 
@@ -1228,7 +1241,7 @@ def static_rnn(cell,
   # determined by the parent scope, or is set to place the cached
   # Variable using the same placement as for the rest of the RNN.
   with vs.variable_scope(scope or "rnn") as varscope:
-    if not context.executing_eagerly():
+    if _should_cache():
       if varscope.caching_device is None:
         varscope.set_caching_device(lambda op: op.device)
 
diff --git a/tensorflow/python/ops/rnn_cell_impl.py b/tensorflow/python/ops/rnn_cell_impl.py
index 82a044a0d4c8710f5ade0aa460f4354a0dd35deb..70805fd5725d66e430005f8a43396b99106607c8 100644
--- a/tensorflow/python/ops/rnn_cell_impl.py
+++ b/tensorflow/python/ops/rnn_cell_impl.py
@@ -47,7 +47,6 @@ from tensorflow.python.ops import variable_scope as vs
 from tensorflow.python.ops import variables as tf_variables
 from tensorflow.python.platform import tf_logging as logging
 from tensorflow.python.training.checkpointable import base as checkpointable
-from tensorflow.python.training.checkpointable import tracking as checkpointable_tracking
 from tensorflow.python.util import nest
 from tensorflow.python.util.tf_export import tf_export
 
@@ -55,16 +54,6 @@ from tensorflow.python.util.tf_export import tf_export
 _BIAS_VARIABLE_NAME = "bias"
 _WEIGHTS_VARIABLE_NAME = "kernel"
 
-
-# TODO(jblespiau): Remove this function when we are sure there are no longer
-# any usage (even if protected, it is being used). Prefer assert_like_rnncell.
-def _like_rnncell(cell):
-  """Checks that a given object is an RNNCell by using duck typing."""
-  conditions = [hasattr(cell, "output_size"), hasattr(cell, "state_size"),
-                hasattr(cell, "zero_state"), callable(cell)]
-  return all(conditions)
-
-
 # This can be used with self.assertRaisesRegexp for assert_like_rnncell.
 ASSERT_LIKE_RNNCELL_ERROR_REGEXP = "is not an RNNCell"
 
@@ -1330,48 +1319,3 @@ class MultiRNNCell(RNNCell):
                   array_ops.concat(new_states, 1))
 
     return cur_inp, new_states
-
-
-class _SlimRNNCell(RNNCell, checkpointable_tracking.NotCheckpointable):
-  """A simple wrapper for slim.rnn_cells."""
-
-  def __init__(self, cell_fn):
-    """Create a SlimRNNCell from a cell_fn.
-
-    Args:
-      cell_fn: a function which takes (inputs, state, scope) and produces the
-        outputs and the new_state. Additionally when called with inputs=None and
-        state=None it should return (initial_outputs, initial_state).
-
-    Raises:
-      TypeError: if cell_fn is not callable
-      ValueError: if cell_fn cannot produce a valid initial state.
-    """
-    if not callable(cell_fn):
-      raise TypeError("cell_fn %s needs to be callable", cell_fn)
-    self._cell_fn = cell_fn
-    self._cell_name = cell_fn.func.__name__
-    init_output, init_state = self._cell_fn(None, None)
-    output_shape = init_output.get_shape()
-    state_shape = init_state.get_shape()
-    self._output_size = output_shape.with_rank(2)[1].value
-    self._state_size = state_shape.with_rank(2)[1].value
-    if self._output_size is None:
-      raise ValueError("Initial output created by %s has invalid shape %s" %
-                       (self._cell_name, output_shape))
-    if self._state_size is None:
-      raise ValueError("Initial state created by %s has invalid shape %s" %
-                       (self._cell_name, state_shape))
-
-  @property
-  def state_size(self):
-    return self._state_size
-
-  @property
-  def output_size(self):
-    return self._output_size
-
-  def __call__(self, inputs, state, scope=None):
-    scope = scope or self._cell_name
-    output, state = self._cell_fn(inputs, state, scope=scope)
-    return output, state
diff --git a/tensorflow/python/ops/script_ops.py b/tensorflow/python/ops/script_ops.py
index 1e3f662ff34f67d2b5f226427c8a03d82b9f2a7c..af103d3cc7649128824132c5520b561425819369 100644
--- a/tensorflow/python/ops/script_ops.py
+++ b/tensorflow/python/ops/script_ops.py
@@ -130,7 +130,7 @@ class FuncRegistry(object):
   def __init__(self):
     self._lock = threading.Lock()
     self._unique_id = 0  # GUARDED_BY(self._lock)
-    # Only store weakrefs to the funtions. The strong reference is stored in
+    # Only store weakrefs to the functions. The strong reference is stored in
     # the graph.
     self._funcs = weakref.WeakValueDictionary()
 
diff --git a/tensorflow/python/ops/special_math_ops.py b/tensorflow/python/ops/special_math_ops.py
index 1508873b751c4fa42d3488ff2d18b5795fda9652..9a10abfcf736be783bfcd7907ec6f357912828ab 100644
--- a/tensorflow/python/ops/special_math_ops.py
+++ b/tensorflow/python/ops/special_math_ops.py
@@ -34,7 +34,7 @@ from tensorflow.python.util.tf_export import tf_export
 
 # TODO(b/27419586) Change docstring for required dtype of x once int allowed
 @tf_export('lbeta')
-def lbeta(x, name='lbeta'):
+def lbeta(x, name=None):
   r"""Computes \\(ln(|Beta(x)|)\\), reducing along the last dimension.
 
   Given one-dimensional `z = [z_0,...,z_{K-1}]`, we define
@@ -64,7 +64,7 @@ def lbeta(x, name='lbeta'):
   # This is consistent with a convention that the sum over the empty set 0, and
   # the product is 1.
   # This is standard.  See https://en.wikipedia.org/wiki/Empty_set.
-  with ops.name_scope(name, values=[x]):
+  with ops.name_scope(name, 'lbeta', [x]):
     x = ops.convert_to_tensor(x, name='x')
 
     # Note reduce_sum([]) = 0.
@@ -83,7 +83,7 @@ def lbeta(x, name='lbeta'):
 
 
 @tf_export('math.bessel_i0')
-def bessel_i0(x, name='bessel_i0'):
+def bessel_i0(x, name=None):
   """Computes the Bessel i0 function of `x` element-wise.
 
   Modified Bessel function of order 0.
@@ -102,12 +102,12 @@ def bessel_i0(x, name='bessel_i0'):
   Equivalent to scipy.special.i0
   @end_compatibility
   """
-  with ops.name_scope(name, [x]):
+  with ops.name_scope(name, 'bessel_i0', [x]):
     return math_ops.exp(math_ops.abs(x)) * math_ops.bessel_i0e(x)
 
 
 @tf_export('math.bessel_i1')
-def bessel_i1(x, name='bessel_i1'):
+def bessel_i1(x, name=None):
   """Computes the Bessel i1 function of `x` element-wise.
 
   Modified Bessel function of order 1.
@@ -126,7 +126,7 @@ def bessel_i1(x, name='bessel_i1'):
   Equivalent to scipy.special.i1
   @end_compatibility
   """
-  with ops.name_scope(name, [x]):
+  with ops.name_scope(name, 'bessel_i1', [x]):
     return math_ops.exp(math_ops.abs(x)) * math_ops.bessel_i1e(x)
 
 
@@ -201,8 +201,8 @@ def einsum(equation, *inputs, **kwargs):
         indices in its subscript, or
       - the input shapes are inconsistent along a particular axis.
   """
-  equation = equation.replace(" ", "")
-  
+  equation = equation.replace(' ', '')
+
   name = kwargs.pop('name', None)
   if kwargs:
     raise TypeError('invalid keyword arguments for this function: ' + ', '.join(
diff --git a/tensorflow/python/ops/special_math_ops_test.py b/tensorflow/python/ops/special_math_ops_test.py
index b7e164f149a9cca336fee061ae2cc3a464ca6132..9bc4098d5b63c3e8ee4f9c14332e65b3d2875d8b 100644
--- a/tensorflow/python/ops/special_math_ops_test.py
+++ b/tensorflow/python/ops/special_math_ops_test.py
@@ -25,24 +25,25 @@ from tensorflow.python.framework import constant_op
 from tensorflow.python.framework import dtypes
 from tensorflow.python.framework import ops
 from tensorflow.python.framework import tensor_shape
+from tensorflow.python.framework import test_util
 from tensorflow.python.ops import array_ops
 from tensorflow.python.ops import math_ops
 from tensorflow.python.ops import special_math_ops
 from tensorflow.python.platform import test
 from tensorflow.python.platform import tf_logging
 
-
 class LBetaTest(test.TestCase):
 
+  @test_util.run_in_graph_and_eager_modes
   def test_one_dimensional_arg(self):
     # Should evaluate to 1 and 1/2.
     x_one = [1, 1.]
     x_one_half = [2, 1.]
     with self.test_session(use_gpu=True):
-      self.assertAllClose(1, math_ops.exp(special_math_ops.lbeta(x_one)).eval())
-      self.assertAllClose(0.5,
-                          math_ops.exp(
-                              special_math_ops.lbeta(x_one_half)).eval())
+      self.assertAllClose(
+          1, self.evaluate(math_ops.exp(special_math_ops.lbeta(x_one))))
+      self.assertAllClose(
+          0.5, self.evaluate(math_ops.exp(special_math_ops.lbeta(x_one_half))))
       self.assertEqual([], special_math_ops.lbeta(x_one).get_shape())
 
   def test_one_dimensional_arg_dynamic(self):
@@ -53,7 +54,8 @@ class LBetaTest(test.TestCase):
       ph = array_ops.placeholder(dtypes.float32)
       beta_ph = math_ops.exp(special_math_ops.lbeta(ph))
       self.assertAllClose(1, beta_ph.eval(feed_dict={ph: x_one}))
-      self.assertAllClose(0.5, beta_ph.eval(feed_dict={ph: x_one_half}))
+      self.assertAllClose(0.5,
+                          beta_ph.eval(feed_dict={ph: x_one_half}))
 
   def test_four_dimensional_arg_with_partial_shape_dynamic(self):
     x_ = np.ones((3, 2, 3, 4))
@@ -66,15 +68,17 @@ class LBetaTest(test.TestCase):
     with self.test_session(use_gpu=True):
       x_ph = array_ops.placeholder(dtypes.float32, [3, 2, 3, None])
       beta_ph = math_ops.exp(special_math_ops.lbeta(x_ph))
-      self.assertAllClose(expected_beta_x, beta_ph.eval(feed_dict={x_ph: x_}))
+      self.assertAllClose(expected_beta_x,
+                          beta_ph.eval(feed_dict={x_ph: x_}))
 
+  @test_util.run_in_graph_and_eager_modes
   def test_two_dimensional_arg(self):
     # Should evaluate to 1/2.
     x_one_half = [[2, 1.], [2, 1.]]
     with self.test_session(use_gpu=True):
-      self.assertAllClose([0.5, 0.5],
-                          math_ops.exp(
-                              special_math_ops.lbeta(x_one_half)).eval())
+      self.assertAllClose(
+          [0.5, 0.5],
+          self.evaluate(math_ops.exp(special_math_ops.lbeta(x_one_half))))
       self.assertEqual((2,), special_math_ops.lbeta(x_one_half).get_shape())
 
   def test_two_dimensional_arg_dynamic(self):
@@ -83,50 +87,59 @@ class LBetaTest(test.TestCase):
     with self.test_session(use_gpu=True):
       ph = array_ops.placeholder(dtypes.float32)
       beta_ph = math_ops.exp(special_math_ops.lbeta(ph))
-      self.assertAllClose([0.5, 0.5], beta_ph.eval(feed_dict={ph: x_one_half}))
+      self.assertAllClose([0.5, 0.5],
+                          beta_ph.eval(feed_dict={ph: x_one_half}))
 
+  @test_util.run_in_graph_and_eager_modes
   def test_two_dimensional_proper_shape(self):
     # Should evaluate to 1/2.
     x_one_half = [[2, 1.], [2, 1.]]
     with self.test_session(use_gpu=True):
-      self.assertAllClose([0.5, 0.5],
-                          math_ops.exp(
-                              special_math_ops.lbeta(x_one_half)).eval())
+      self.assertAllClose(
+          [0.5, 0.5],
+          self.evaluate(math_ops.exp(special_math_ops.lbeta(x_one_half))))
       self.assertEqual(
           (2,),
-          array_ops.shape(special_math_ops.lbeta(x_one_half)).eval())
+          self.evaluate(array_ops.shape(special_math_ops.lbeta(x_one_half))))
       self.assertEqual(
           tensor_shape.TensorShape([2]),
           special_math_ops.lbeta(x_one_half).get_shape())
 
+  @test_util.run_in_graph_and_eager_modes
   def test_complicated_shape(self):
     with self.test_session(use_gpu=True):
       x = ops.convert_to_tensor(np.random.rand(3, 2, 2))
-      self.assertAllEqual((3, 2),
-                          array_ops.shape(special_math_ops.lbeta(x)).eval())
+      self.assertAllEqual(
+          (3, 2), self.evaluate(array_ops.shape(special_math_ops.lbeta(x))))
       self.assertEqual(
           tensor_shape.TensorShape([3, 2]),
           special_math_ops.lbeta(x).get_shape())
 
+  @test_util.run_in_graph_and_eager_modes
   def test_length_1_last_dimension_results_in_one(self):
     # If there is only one coefficient, the formula still works, and we get one
     # as the answer, always.
     x_a = [5.5]
     x_b = [0.1]
     with self.test_session(use_gpu=True):
-      self.assertAllClose(1, math_ops.exp(special_math_ops.lbeta(x_a)).eval())
-      self.assertAllClose(1, math_ops.exp(special_math_ops.lbeta(x_b)).eval())
+      self.assertAllClose(
+          1, self.evaluate(math_ops.exp(special_math_ops.lbeta(x_a))))
+      self.assertAllClose(
+          1, self.evaluate(math_ops.exp(special_math_ops.lbeta(x_b))))
       self.assertEqual((), special_math_ops.lbeta(x_a).get_shape())
 
+  @test_util.run_in_graph_and_eager_modes
   def test_empty_rank1_returns_negative_infinity(self):
     with self.test_session(use_gpu=True):
       x = constant_op.constant([], shape=[0])
       lbeta_x = special_math_ops.lbeta(x)
       expected_result = constant_op.constant(-np.inf, shape=())
 
-      self.assertAllEqual(expected_result.eval(), lbeta_x.eval())
+      self.assertAllEqual(self.evaluate(expected_result),
+                          self.evaluate(lbeta_x))
       self.assertEqual(expected_result.get_shape(), lbeta_x.get_shape())
 
+  @test_util.run_in_graph_and_eager_modes
   def test_empty_rank2_with_zero_last_dim_returns_negative_infinity(self):
     with self.test_session(use_gpu=True):
       event_size = 0
@@ -135,9 +148,11 @@ class LBetaTest(test.TestCase):
         lbeta_x = special_math_ops.lbeta(x)
         expected_result = constant_op.constant(-np.inf, shape=[batch_size])
 
-        self.assertAllEqual(expected_result.eval(), lbeta_x.eval())
+        self.assertAllEqual(self.evaluate(expected_result),
+                            self.evaluate(lbeta_x))
         self.assertEqual(expected_result.get_shape(), lbeta_x.get_shape())
 
+  @test_util.run_in_graph_and_eager_modes
   def test_empty_rank2_with_zero_batch_dim_returns_empty(self):
     with self.test_session(use_gpu=True):
       batch_size = 0
@@ -147,12 +162,14 @@ class LBetaTest(test.TestCase):
 
         expected_result = constant_op.constant([], shape=[batch_size])
 
-        self.assertAllEqual(expected_result.eval(), lbeta_x.eval())
+        self.assertAllEqual(self.evaluate(expected_result),
+                            self.evaluate(lbeta_x))
         self.assertEqual(expected_result.get_shape(), lbeta_x.get_shape())
 
 
 class BesselTest(test.TestCase):
 
+  @test_util.run_in_graph_and_eager_modes
   def test_bessel_i0(self):
     x_single = np.arange(-3, 3).reshape(1, 3, 2).astype(np.float32)
     x_double = np.arange(-3, 3).reshape(1, 3, 2).astype(np.float64)
@@ -165,6 +182,7 @@ class BesselTest(test.TestCase):
     except ImportError as e:
       tf_logging.warn('Cannot test special functions: %s' % str(e))
 
+  @test_util.run_in_graph_and_eager_modes
   def test_bessel_i1(self):
     x_single = np.arange(-3, 3).reshape(1, 3, 2).astype(np.float32)
     x_double = np.arange(-3, 3).reshape(1, 3, 2).astype(np.float64)
@@ -316,7 +334,7 @@ class EinsumTest(test.TestCase):
     output_tensor = special_math_ops.einsum(axes, *input_tensors)
 
     with self.test_session(use_gpu=True):
-      output_value = output_tensor.eval()
+      output_value = self.evaluate(output_tensor)
 
     correct_value = np.einsum(axes, *input_vals)
 
diff --git a/tensorflow/python/ops/spectral_ops.py b/tensorflow/python/ops/spectral_ops.py
index 28054f50ef3b1227f12376b4b3700a7618270d65..293aace7282eb0f8dde9da75b0d353a560c0ecb9 100644
--- a/tensorflow/python/ops/spectral_ops.py
+++ b/tensorflow/python/ops/spectral_ops.py
@@ -167,8 +167,8 @@ def _validate_dct_arguments(dct_type, n, axis, norm):
     raise NotImplementedError("The DCT length argument is not implemented.")
   if axis != -1:
     raise NotImplementedError("axis must be -1. Got: %s" % axis)
-  if dct_type != 2:
-    raise ValueError("Only the Type II DCT is supported.")
+  if dct_type not in (2, 3):
+    raise ValueError("Only Types II and III (I)DCT are supported.")
   if norm not in (None, "ortho"):
     raise ValueError(
         "Unknown normalization. Expected None or 'ortho', got: %s" % norm)
@@ -179,18 +179,20 @@ def _validate_dct_arguments(dct_type, n, axis, norm):
 def dct(input, type=2, n=None, axis=-1, norm=None, name=None):  # pylint: disable=redefined-builtin
   """Computes the 1D [Discrete Cosine Transform (DCT)][dct] of `input`.
 
-  Currently only Type II is supported. Implemented using a length `2N` padded
-  @{tf.spectral.rfft}, as described here: https://dsp.stackexchange.com/a/10606
+  Currently only Types II and III are supported. Type II is implemented using a
+  length `2N` padded @{tf.spectral.rfft}, as described here:
+  https://dsp.stackexchange.com/a/10606. Type III is a fairly straightforward
+  inverse of Type II (i.e. using a length `2N` padded @{tf.spectral.irfft}).
 
   @compatibility(scipy)
-  Equivalent to scipy.fftpack.dct for the Type-II DCT.
+  Equivalent to scipy.fftpack.dct for Type-II and Type-III DCT.
   https://docs.scipy.org/doc/scipy-0.14.0/reference/generated/scipy.fftpack.dct.html
   @end_compatibility
 
   Args:
     input: A `[..., samples]` `float32` `Tensor` containing the signals to
       take the DCT of.
-    type: The DCT type to perform. Must be 2.
+    type: The DCT type to perform. Must be 2 or 3.
     n: For future expansion. The length of the transform. Must be `None`.
     axis: For future expansion. The axis to compute the DCT along. Must be `-1`.
     norm: The normalization to apply. `None` for no normalization or `'ortho'`
@@ -201,8 +203,8 @@ def dct(input, type=2, n=None, axis=-1, norm=None, name=None):  # pylint: disabl
     A `[..., samples]` `float32` `Tensor` containing the DCT of `input`.
 
   Raises:
-    ValueError: If `type` is not `2`, `n` is not `None, `axis` is not `-1`, or
-      `norm` is not `None` or `'ortho'`.
+    ValueError: If `type` is not `2` or `3`, `n` is not `None, `axis` is not
+      `-1`, or `norm` is not `None` or `'ortho'`.
 
   [dct]: https://en.wikipedia.org/wiki/Discrete_cosine_transform
   """
@@ -214,22 +216,91 @@ def dct(input, type=2, n=None, axis=-1, norm=None, name=None):  # pylint: disabl
 
     axis_dim = input.shape[-1].value or _array_ops.shape(input)[-1]
     axis_dim_float = _math_ops.to_float(axis_dim)
-    scale = 2.0 * _math_ops.exp(_math_ops.complex(
-        0.0, -_math.pi * _math_ops.range(axis_dim_float) /
-        (2.0 * axis_dim_float)))
-
-    # TODO(rjryan): Benchmark performance and memory usage of the various
-    # approaches to computing a DCT via the RFFT.
-    dct2 = _math_ops.real(
-        rfft(input, fft_length=[2 * axis_dim])[..., :axis_dim] * scale)
-
-    if norm == "ortho":
-      n1 = 0.5 * _math_ops.rsqrt(axis_dim_float)
-      n2 = n1 * _math_ops.sqrt(2.0)
-      # Use tf.pad to make a vector of [n1, n2, n2, n2, ...].
-      weights = _array_ops.pad(
-          _array_ops.expand_dims(n1, 0), [[0, axis_dim - 1]],
-          constant_values=n2)
-      dct2 *= weights
-
-    return dct2
+    if type == 2:
+      scale = 2.0 * _math_ops.exp(
+          _math_ops.complex(
+              0.0, -_math_ops.range(axis_dim_float) * _math.pi * 0.5 /
+              axis_dim_float))
+
+      # TODO(rjryan): Benchmark performance and memory usage of the various
+      # approaches to computing a DCT via the RFFT.
+      dct2 = _math_ops.real(
+          rfft(input, fft_length=[2 * axis_dim])[..., :axis_dim] * scale)
+
+      if norm == "ortho":
+        n1 = 0.5 * _math_ops.rsqrt(axis_dim_float)
+        n2 = n1 * _math_ops.sqrt(2.0)
+        # Use tf.pad to make a vector of [n1, n2, n2, n2, ...].
+        weights = _array_ops.pad(
+            _array_ops.expand_dims(n1, 0), [[0, axis_dim - 1]],
+            constant_values=n2)
+        dct2 *= weights
+
+      return dct2
+
+    elif type == 3:
+      if norm == "ortho":
+        n1 = _math_ops.sqrt(axis_dim_float)
+        n2 = n1 * _math_ops.sqrt(0.5)
+        # Use tf.pad to make a vector of [n1, n2, n2, n2, ...].
+        weights = _array_ops.pad(
+            _array_ops.expand_dims(n1, 0), [[0, axis_dim - 1]],
+            constant_values=n2)
+        input *= weights
+      else:
+        input *= axis_dim_float
+      scale = 2.0 * _math_ops.exp(
+          _math_ops.complex(
+              0.0,
+              _math_ops.range(axis_dim_float) * _math.pi * 0.5 /
+              axis_dim_float))
+      dct3 = _math_ops.real(
+          irfft(
+              scale * _math_ops.complex(input, 0.0),
+              fft_length=[2 * axis_dim]))[..., :axis_dim]
+
+      return dct3
+
+
+# TODO(rjryan): Implement `type`, `n` and `axis` parameters.
+@tf_export("spectral.idct")
+def idct(input, type=2, n=None, axis=-1, norm=None, name=None):  # pylint: disable=redefined-builtin
+  """Computes the 1D [Inverse Discrete Cosine Transform (DCT)][idct] of `input`.
+
+  Currently only Types II and III are supported. Type III is the inverse of
+  Type II, and vice versa.
+
+  Note that you must re-normalize by 1/(2n) to obtain an inverse if `norm` is
+  not `'ortho'`. That is:
+  `signal == idct(dct(signal)) * 0.5 / signal.shape[-1]`.
+  When `norm='ortho'`, we have:
+  `signal == idct(dct(signal, norm='ortho'), norm='ortho')`.
+
+  @compatibility(scipy)
+  Equivalent to scipy.fftpack.idct for Type-II and Type-III DCT.
+  https://docs.scipy.org/doc/scipy-0.14.0/reference/generated/scipy.fftpack.idct.html
+  @end_compatibility
+
+  Args:
+    input: A `[..., samples]` `float32` `Tensor` containing the signals to take
+      the DCT of.
+    type: The IDCT type to perform. Must be 2 or 3.
+    n: For future expansion. The length of the transform. Must be `None`.
+    axis: For future expansion. The axis to compute the DCT along. Must be `-1`.
+    norm: The normalization to apply. `None` for no normalization or `'ortho'`
+      for orthonormal normalization.
+    name: An optional name for the operation.
+
+  Returns:
+    A `[..., samples]` `float32` `Tensor` containing the IDCT of `input`.
+
+  Raises:
+    ValueError: If `type` is not `2` or `3`, `n` is not `None, `axis` is not
+      `-1`, or `norm` is not `None` or `'ortho'`.
+
+  [idct]:
+  https://en.wikipedia.org/wiki/Discrete_cosine_transform#Inverse_transforms
+  """
+  _validate_dct_arguments(type, n, axis, norm)
+  inverse_type = {2: 3, 3: 2}[type]
+  return dct(input, type=inverse_type, n=n, axis=axis, norm=norm, name=name)
diff --git a/tensorflow/python/ops/state_ops.py b/tensorflow/python/ops/state_ops.py
index 8cb6a0537e928effbcf4c475bcc4e974182da2a7..2c93cf72c75ba27145e06abe69bcbef9418b39e0 100644
--- a/tensorflow/python/ops/state_ops.py
+++ b/tensorflow/python/ops/state_ops.py
@@ -19,7 +19,6 @@ from __future__ import absolute_import
 from __future__ import division
 from __future__ import print_function
 
-from tensorflow.python.eager import context
 from tensorflow.python.framework import ops
 from tensorflow.python.framework import tensor_shape
 from tensorflow.python.ops import gen_resource_variable_ops
@@ -124,9 +123,7 @@ def is_variable_initialized(ref, name=None):
   if ref.dtype._is_ref_dtype:
     return gen_state_ops.is_variable_initialized(ref=ref, name=name)
   # Handle resource variables.
-  if context.executing_eagerly() or ref.op.type == "VarHandleOp":
-    return gen_resource_variable_ops.var_is_initialized_op(ref.handle,
-                                                           name=name)
+  return ref.is_initialized(name=name)
 
 
 @tf_export("assign_sub")
diff --git a/tensorflow/python/ops/summary_ops_v2.py b/tensorflow/python/ops/summary_ops_v2.py
index b80f84eb7cde264c5a7c83eafacc344adb50b80a..00150fe68820da711c76f642baced45163a8727c 100644
--- a/tensorflow/python/ops/summary_ops_v2.py
+++ b/tensorflow/python/ops/summary_ops_v2.py
@@ -306,10 +306,11 @@ def create_db_writer(db_uri,
 def _make_summary_writer(name, factory, **kwargs):
   resource = gen_summary_ops.summary_writer(shared_name=name)
   init_op_fn = lambda: factory(resource, **kwargs)
-  # TODO(apassos): Consider doing this instead.
-  # if not context.executing_eagerly():
-  #   ops.get_default_session().run(init_op)
-  ops.add_to_collection(_SUMMARY_WRITER_INIT_COLLECTION_NAME, init_op_fn())
+  init_op = init_op_fn()
+  if not context.executing_eagerly():
+    # TODO(apassos): Consider doing this instead.
+    #   ops.get_default_session().run(init_op)
+    ops.add_to_collection(_SUMMARY_WRITER_INIT_COLLECTION_NAME, init_op)
   return SummaryWriter(resource, init_op_fn)
 
 
@@ -380,7 +381,8 @@ def summary_writer_function(name, tensor, function, family=None):
   with ops.device("cpu:0"):
     op = smart_cond.smart_cond(
         should_record_summaries(), record, _nothing, name="")
-    ops.add_to_collection(ops.GraphKeys._SUMMARY_COLLECTION, op)  # pylint: disable=protected-access
+    if not context.executing_eagerly():
+      ops.add_to_collection(ops.GraphKeys._SUMMARY_COLLECTION, op)  # pylint: disable=protected-access
   return op
 
 
diff --git a/tensorflow/python/ops/variable_scope.py b/tensorflow/python/ops/variable_scope.py
index 47414c28af3d1e53cc0e9f902a0a2b2c6a49cbc9..77f67c18ee871409bab0509fa58ea04d1951434a 100644
--- a/tensorflow/python/ops/variable_scope.py
+++ b/tensorflow/python/ops/variable_scope.py
@@ -1,4 +1,4 @@
- # Copyright 2015 The TensorFlow Authors. All Rights Reserved.
+# Copyright 2015 The TensorFlow Authors. All Rights Reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -44,9 +44,11 @@ from tensorflow.python.util import function_utils
 from tensorflow.python.util import tf_contextlib
 from tensorflow.python.util.tf_export import tf_export
 
-__all__ = ["AUTO_REUSE", "VariableScope", "get_variable_scope",
-           "get_variable", "get_local_variable", "variable_scope",
-           "variable_op_scope", "no_regularizer"]
+__all__ = [
+    "AUTO_REUSE", "VariableScope", "get_variable_scope", "get_variable",
+    "get_local_variable", "variable_scope", "variable_op_scope",
+    "no_regularizer", "VariableSynchronization", "VariableAggregation"
+]
 
 
 class _PartitionInfo(object):
@@ -188,6 +190,38 @@ class _ReuseMode(enum.Enum):
   # REUSE_FALSE = 2
   # REUSE_TRUE = 3
 
+
+@tf_export("VariableSynchronization")
+class VariableSynchronization(enum.Enum):
+  """Indicates when a distributed variable will be synced."""
+
+  # Indicates that the synchronization will be determined by the current
+  # `DistributionStrategy` (eg. With `MirroredStrategy` this would be
+  # `ON_WRITE`).
+  AUTO = 0
+
+  # Indicates that there will only be one copy of the variable, so there is no
+  # need to sync.
+  NONE = 1
+
+  # Indicates that the variable will be aggregated across devices
+  # every time it is updated.
+  ON_WRITE = 2
+
+  # Indicates that the variable will be aggregated across devices
+  # when it is read (eg. when checkpointing or when evaluating an op that uses
+  # the variable).
+  ON_READ = 3
+
+
+@tf_export("VariableAggregation")
+class VariableAggregation(enum.Enum):
+  """Indicates how a distributed variable will be aggregated."""
+  NONE = 0
+  SUM = 1
+  MEAN = 2
+
+
 AUTO_REUSE = _ReuseMode.AUTO_REUSE
 tf_export("AUTO_REUSE").export_constant(__name__, "AUTO_REUSE")
 AUTO_REUSE.__doc__ = """
@@ -214,11 +248,23 @@ class _VariableStore(object):
     self._partitioned_vars = {}  # A dict of the stored PartitionedVariables.
     self._store_eager_variables = False
 
-  def get_variable(self, name, shape=None, dtype=dtypes.float32,
-                   initializer=None, regularizer=None, reuse=None,
-                   trainable=True, collections=None, caching_device=None,
-                   partitioner=None, validate_shape=True, use_resource=None,
-                   custom_getter=None, constraint=None):
+  def get_variable(self,
+                   name,
+                   shape=None,
+                   dtype=dtypes.float32,
+                   initializer=None,
+                   regularizer=None,
+                   reuse=None,
+                   trainable=None,
+                   collections=None,
+                   caching_device=None,
+                   partitioner=None,
+                   validate_shape=True,
+                   use_resource=None,
+                   custom_getter=None,
+                   constraint=None,
+                   synchronization=VariableSynchronization.AUTO,
+                   aggregation=VariableAggregation.NONE):
     """Gets an existing variable with these parameters or create a new one.
 
     If a variable with the given name is already stored, we return the stored
@@ -254,6 +300,8 @@ class _VariableStore(object):
         forced to be False.
       trainable: If `True` also add the variable to the graph collection
         `GraphKeys.TRAINABLE_VARIABLES` (see `tf.Variable`).
+        `trainable` defaults to `True` unless `synchronization` is
+        set to `ON_READ`.
       collections: List of graph collections keys to add the `Variable` to.
         Defaults to `[GraphKeys.GLOBAL_VARIABLES]` (see `tf.Variable`).
       caching_device: Optional device string or function describing where the
@@ -291,6 +339,15 @@ class _VariableStore(object):
         variable and return the Tensor for the projected value
         (which must have the same shape). Constraints are not safe to
         use when doing asynchronous distributed training.
+      synchronization: Indicates when a distributed a variable will be
+        aggregated. Accepted values are constants defined in the class
+        @{tf.VariableSynchronization}. By default the synchronization is set to
+        `AUTO` and the current `DistributionStrategy` chooses
+        when to synchronize. If `synchronization` is set to `ON_READ`,
+        `trainable` must not be set to `True`.
+      aggregation: Indicates how a distributed variable will be aggregated.
+        Accepted values are constants defined in the class
+        @{tf.VariableAggregation}.
 
     Returns:
       The created or existing `Variable` (or `PartitionedVariable`, if a
@@ -343,11 +400,22 @@ class _VariableStore(object):
     # it to custom_getter.
     # Note: the parameters of _true_getter, and their documentation, match
     # *exactly* item-for-item with the docstring of this method.
-    def _true_getter(name, shape=None, dtype=dtypes.float32,  # pylint: disable=missing-docstring
-                     initializer=None, regularizer=None, reuse=None,
-                     trainable=True, collections=None, caching_device=None,
-                     partitioner=None, validate_shape=True, use_resource=None,
-                     constraint=None):
+    def _true_getter(  # pylint: disable=missing-docstring
+        name,
+        shape=None,
+        dtype=dtypes.float32,
+        initializer=None,
+        regularizer=None,
+        reuse=None,
+        trainable=None,
+        collections=None,
+        caching_device=None,
+        partitioner=None,
+        validate_shape=True,
+        use_resource=None,
+        constraint=None,
+        synchronization=VariableSynchronization.AUTO,
+        aggregation=VariableAggregation.NONE):
       is_scalar = (shape is not None
                    and isinstance(shape, collections_lib.Sequence)
                    and not shape)
@@ -397,11 +465,24 @@ class _VariableStore(object):
             "name was already created with partitioning?" % name)
 
       return self._get_single_variable(
-          name=name, shape=shape, dtype=dtype,
-          initializer=initializer, regularizer=regularizer, reuse=reuse,
-          trainable=trainable, collections=collections,
-          caching_device=caching_device, validate_shape=validate_shape,
-          use_resource=use_resource, constraint=constraint)
+          name=name,
+          shape=shape,
+          dtype=dtype,
+          initializer=initializer,
+          regularizer=regularizer,
+          reuse=reuse,
+          trainable=trainable,
+          collections=collections,
+          caching_device=caching_device,
+          validate_shape=validate_shape,
+          use_resource=use_resource,
+          constraint=constraint,
+          synchronization=synchronization,
+          aggregation=aggregation)
+
+    # Set trainable value based on synchronization value.
+    trainable = _get_trainable_value(
+        synchronization=synchronization, trainable=trainable)
 
     if custom_getter is not None:
       # Handle backwards compatibility with getter arguments that were added
@@ -420,6 +501,8 @@ class _VariableStore(object):
           "partitioner": partitioner,
           "validate_shape": validate_shape,
           "use_resource": use_resource,
+          "synchronization": synchronization,
+          "aggregation": aggregation,
       }
       # `fn_args` can handle functions, `functools.partial`, `lambda`.
       if "constraint" in function_utils.fn_args(custom_getter):
@@ -427,18 +510,36 @@ class _VariableStore(object):
       return custom_getter(**custom_getter_kwargs)
     else:
       return _true_getter(
-          name, shape=shape, dtype=dtype,
-          initializer=initializer, regularizer=regularizer,
-          reuse=reuse, trainable=trainable, collections=collections,
-          caching_device=caching_device, partitioner=partitioner,
-          validate_shape=validate_shape, use_resource=use_resource,
-          constraint=constraint)
-
-  def _get_partitioned_variable(
-      self, name, partitioner, shape=None, dtype=dtypes.float32,
-      initializer=None, regularizer=None, reuse=None,
-      trainable=True, collections=None, caching_device=None,
-      validate_shape=True, use_resource=None, constraint=None):
+          name,
+          shape=shape,
+          dtype=dtype,
+          initializer=initializer,
+          regularizer=regularizer,
+          reuse=reuse,
+          trainable=trainable,
+          collections=collections,
+          caching_device=caching_device,
+          partitioner=partitioner,
+          validate_shape=validate_shape,
+          use_resource=use_resource,
+          constraint=constraint,
+          synchronization=synchronization,
+          aggregation=aggregation)
+
+  def _get_partitioned_variable(self,
+                                name,
+                                partitioner,
+                                shape=None,
+                                dtype=dtypes.float32,
+                                initializer=None,
+                                regularizer=None,
+                                reuse=None,
+                                trainable=None,
+                                collections=None,
+                                caching_device=None,
+                                validate_shape=True,
+                                use_resource=None,
+                                constraint=None):
     """Gets or creates a sharded variable list with these parameters.
 
     The `partitioner` must be a callable that accepts a fully defined
@@ -688,12 +789,14 @@ class _VariableStore(object):
                            regularizer=None,
                            partition_info=None,
                            reuse=None,
-                           trainable=True,
+                           trainable=None,
                            collections=None,
                            caching_device=None,
                            validate_shape=True,
                            use_resource=None,
-                           constraint=None):
+                           constraint=None,
+                           synchronization=VariableSynchronization.AUTO,
+                           aggregation=VariableAggregation.NONE):
     """Get or create a single Variable (e.g. a shard or entire variable).
 
     See the documentation of get_variable above (ignore partitioning components)
@@ -713,6 +816,8 @@ class _VariableStore(object):
       validate_shape: see get_variable.
       use_resource: see get_variable.
       constraint: see get_variable.
+      synchronization: see get_variable.
+      aggregation: see get_variable.
 
     Returns:
       A Variable.  See documentation of get_variable above.
@@ -793,7 +898,9 @@ class _VariableStore(object):
         dtype=variable_dtype,
         validate_shape=validate_shape,
         constraint=constraint,
-        use_resource=use_resource)
+        use_resource=use_resource,
+        synchronization=synchronization,
+        aggregation=aggregation)
     if context.executing_eagerly() and self._store_eager_variables:
       if collections:
         ops.add_to_collections(collections, v)
@@ -1045,14 +1152,16 @@ class VariableScope(object):
                    initializer=None,
                    regularizer=None,
                    reuse=None,
-                   trainable=True,
+                   trainable=None,
                    collections=None,
                    caching_device=None,
                    partitioner=None,
                    validate_shape=True,
                    use_resource=None,
                    custom_getter=None,
-                   constraint=None):
+                   constraint=None,
+                   synchronization=VariableSynchronization.AUTO,
+                   aggregation=VariableAggregation.NONE):
     """Gets an existing variable with this name or create a new one."""
     if regularizer is None:
       regularizer = self._regularizer
@@ -1090,12 +1199,22 @@ class VariableScope(object):
       if dtype is None:
         dtype = self._dtype
       return var_store.get_variable(
-          full_name, shape=shape, dtype=dtype, initializer=initializer,
-          regularizer=regularizer, reuse=reuse, trainable=trainable,
-          collections=collections, caching_device=caching_device,
-          partitioner=partitioner, validate_shape=validate_shape,
-          use_resource=use_resource, custom_getter=custom_getter,
-          constraint=constraint)
+          full_name,
+          shape=shape,
+          dtype=dtype,
+          initializer=initializer,
+          regularizer=regularizer,
+          reuse=reuse,
+          trainable=trainable,
+          collections=collections,
+          caching_device=caching_device,
+          partitioner=partitioner,
+          validate_shape=validate_shape,
+          use_resource=use_resource,
+          custom_getter=custom_getter,
+          constraint=constraint,
+          synchronization=synchronization,
+          aggregation=aggregation)
 
   def _get_partitioned_variable(self,
                                 var_store,
@@ -1104,7 +1223,7 @@ class VariableScope(object):
                                 dtype=None,
                                 initializer=None,
                                 regularizer=None,
-                                trainable=True,
+                                trainable=None,
                                 collections=None,
                                 caching_device=None,
                                 partitioner=None,
@@ -1319,21 +1438,35 @@ def get_variable(name,
                  dtype=None,
                  initializer=None,
                  regularizer=None,
-                 trainable=True,
+                 trainable=None,
                  collections=None,
                  caching_device=None,
                  partitioner=None,
                  validate_shape=True,
                  use_resource=None,
                  custom_getter=None,
-                 constraint=None):
+                 constraint=None,
+                 synchronization=VariableSynchronization.AUTO,
+                 aggregation=VariableAggregation.NONE):
   return get_variable_scope().get_variable(
-      _get_default_variable_store(), name, shape=shape, dtype=dtype,
-      initializer=initializer, regularizer=regularizer, trainable=trainable,
-      collections=collections, caching_device=caching_device,
-      partitioner=partitioner, validate_shape=validate_shape,
-      use_resource=use_resource, custom_getter=custom_getter,
-      constraint=constraint)
+      _get_default_variable_store(),
+      name,
+      shape=shape,
+      dtype=dtype,
+      initializer=initializer,
+      regularizer=regularizer,
+      trainable=trainable,
+      collections=collections,
+      caching_device=caching_device,
+      partitioner=partitioner,
+      validate_shape=validate_shape,
+      use_resource=use_resource,
+      custom_getter=custom_getter,
+      constraint=constraint,
+      synchronization=synchronization,
+      aggregation=aggregation)
+
+
 get_variable_or_local_docstring = (
     """%s
 
@@ -1430,29 +1563,44 @@ get_variable.__doc__ = get_variable_or_local_docstring % (
 # The argument list for get_local_variable must match arguments to get_variable.
 # So, if you are updating the arguments, also update arguments to get_variable.
 @tf_export("get_local_variable")
-def get_local_variable(name,
-                       shape=None,
-                       dtype=None,
-                       initializer=None,
-                       regularizer=None,
-                       trainable=False,  # pylint: disable=unused-argument
-                       collections=None,
-                       caching_device=None,
-                       partitioner=None,
-                       validate_shape=True,
-                       use_resource=None,
-                       custom_getter=None,
-                       constraint=None):
+def get_local_variable(  # pylint: disable=missing-docstring
+    name,
+    shape=None,
+    dtype=None,
+    initializer=None,
+    regularizer=None,
+    trainable=False,  # pylint: disable=unused-argument
+    collections=None,
+    caching_device=None,
+    partitioner=None,
+    validate_shape=True,
+    use_resource=None,
+    synchronization=VariableSynchronization.AUTO,
+    aggregation=VariableAggregation.NONE,
+    custom_getter=None,
+    constraint=None):
   if collections:
     collections += [ops.GraphKeys.LOCAL_VARIABLES]
   else:
     collections = [ops.GraphKeys.LOCAL_VARIABLES]
   return get_variable(
-      name, shape=shape, dtype=dtype, initializer=initializer,
-      regularizer=regularizer, trainable=False, collections=collections,
-      caching_device=caching_device, partitioner=partitioner,
-      validate_shape=validate_shape, use_resource=use_resource,
-      custom_getter=custom_getter, constraint=constraint)
+      name,
+      shape=shape,
+      dtype=dtype,
+      initializer=initializer,
+      regularizer=regularizer,
+      trainable=False,
+      collections=collections,
+      caching_device=caching_device,
+      partitioner=partitioner,
+      validate_shape=validate_shape,
+      use_resource=use_resource,
+      synchronization=synchronization,
+      aggregation=aggregation,
+      custom_getter=custom_getter,
+      constraint=constraint)
+
+
 get_local_variable.__doc__ = get_variable_or_local_docstring % (
     "Gets an existing *local* variable or creates a new one.",
     "Behavior is the same as in `get_variable`, except that variables are\n"
@@ -2202,11 +2350,28 @@ def _compute_slice_dim_and_shape(full_shape, slicing):
   return slice_dim, slice_shape
 
 
+def _get_trainable_value(synchronization, trainable):
+  """Computes the trainable value based on the given arguments."""
+  if synchronization == VariableSynchronization.ON_READ:
+    if trainable:
+      raise ValueError(
+          "Synchronization value can be set to "
+          "VariableSynchronization.ON_READ only for non-trainable variables. "
+          "You have specified trainable=True and "
+          "synchronization=VariableSynchronization.ON_READ.")
+    else:
+      # Set trainable to be false when variable is to be synced on read.
+      trainable = False
+  elif trainable is None:
+    trainable = True
+  return trainable
+
+
 def default_variable_creator(next_creator=None, **kwargs):
   """Default variable creator."""
   assert next_creator is None
   initial_value = kwargs.get("initial_value", None)
-  trainable = kwargs.get("trainable", True)
+  trainable = kwargs.get("trainable", None)
   collections = kwargs.get("collections", None)
   validate_shape = kwargs.get("validate_shape", True)
   caching_device = kwargs.get("caching_device", None)
@@ -2214,6 +2379,12 @@ def default_variable_creator(next_creator=None, **kwargs):
   dtype = kwargs.get("dtype", None)
   constraint = kwargs.get("constraint", None)
   use_resource = kwargs.get("use_resource", None)
+
+  # Set trainable value based on synchronization value.
+  synchronization = kwargs.get("synchronization", VariableSynchronization.AUTO)
+  trainable = _get_trainable_value(
+      synchronization=synchronization, trainable=trainable)
+
   if use_resource is None:
     use_resource = get_variable_scope().use_resource
   if use_resource or (use_resource is None and context.executing_eagerly()):
@@ -2241,25 +2412,35 @@ def _make_getter(captured_getter, captured_previous):
 
 
 def variable(initial_value=None,
-             trainable=True,
+             trainable=None,
              collections=None,
              validate_shape=True,
              caching_device=None,
              name=None,
              dtype=None,
              constraint=None,
-             use_resource=None):
+             use_resource=None,
+             synchronization=VariableSynchronization.AUTO,
+             aggregation=VariableAggregation.NONE):
   previous_getter = lambda **kwargs: default_variable_creator(None, **kwargs)
   for getter in ops.get_default_graph()._variable_creator_stack:  # pylint: disable=protected-access
     previous_getter = _make_getter(getter, previous_getter)
-  return previous_getter(initial_value=initial_value,
-                         trainable=trainable,
-                         collections=collections,
-                         validate_shape=validate_shape,
-                         caching_device=caching_device,
-                         name=name, dtype=dtype,
-                         constraint=constraint,
-                         use_resource=use_resource)
+
+  # Reset `aggregation` that is explicitly set as `None` to the enum None value.
+  if aggregation is None:
+    aggregation = VariableAggregation.NONE
+  return previous_getter(
+      initial_value=initial_value,
+      trainable=trainable,
+      collections=collections,
+      validate_shape=validate_shape,
+      caching_device=caching_device,
+      name=name,
+      dtype=dtype,
+      constraint=constraint,
+      use_resource=use_resource,
+      synchronization=synchronization,
+      aggregation=aggregation)
 
 
 @tf_contextlib.contextmanager
@@ -2293,6 +2474,8 @@ def variable_creator_scope(variable_creator):
       trainable: If `True`, the default, also adds the variable to the graph
         collection `GraphKeys.TRAINABLE_VARIABLES`. This collection is used as
         the default list of variables to use by the `Optimizer` classes.
+        `trainable` defaults to `True` unless `synchronization` is
+        set to `ON_READ`.
       collections: List of graph collections keys. The new variable is added to
         these collections. Defaults to `[GraphKeys.GLOBAL_VARIABLES]`.
       validate_shape: If `False`, allows the variable to be initialized with a
@@ -2311,6 +2494,15 @@ def variable_creator_scope(variable_creator):
       constraint: A constraint function to be applied to the variable after
         updates by some algorithms.
       use_resource: if True, a ResourceVariable is always created.
+      synchronization: Indicates when a distributed a variable will be
+        aggregated. Accepted values are constants defined in the class
+        @{tf.VariableSynchronization}. By default the synchronization is set to
+        `AUTO` and the current `DistributionStrategy` chooses
+        when to synchronize. If `synchronization` is set to `ON_READ`,
+        `trainable` must not be set to `True`.
+      aggregation: Indicates how a distributed variable will be aggregated.
+        Accepted values are constants defined in the class
+        @{tf.VariableAggregation}.
 
   This set may grow over time, so it's important the signature of creators is as
   mentioned above.
diff --git a/tensorflow/python/ops/variables.py b/tensorflow/python/ops/variables.py
index 4be9f5eb6864015cd9c3f6f3526285ebbdc180f9..d3b8da6d2ae52bf9d1fd8b6ef02c44a8cf5fe318 100644
--- a/tensorflow/python/ops/variables.py
+++ b/tensorflow/python/ops/variables.py
@@ -1093,39 +1093,40 @@ class Variable(checkpointable.CheckpointableBase):
   def __imul__(self, other):
     logging.log_first_n(
         logging.WARN,
-        "Variable *= will be deprecated. Use variable.assign_mul"
-        " if you want assignment to the variable value or 'x = x * y'"
+        "Variable *= will be deprecated. Use `var.assign(var * other)`"
+        " if you want assignment to the variable value or `x = x * y`"
         " if you want a new python Tensor object.", 1)
     return self * other
 
   def __idiv__(self, other):
     logging.log_first_n(
         logging.WARN,
-        "Variable /= will be deprecated. Use variable.assign_div"
-        " if you want assignment to the variable value or 'x = x / y'"
+        "Variable /= will be deprecated. Use `var.assign(var / other)`"
+        " if you want assignment to the variable value or `x = x / y`"
         " if you want a new python Tensor object.", 1)
     return self / other
 
   def __itruediv__(self, other):
     logging.log_first_n(
         logging.WARN,
-        "Variable /= will be deprecated. Use variable.assign_div"
-        " if you want assignment to the variable value or 'x = x / y'"
+        "Variable /= will be deprecated. Use `var.assign(var / other)`"
+        " if you want assignment to the variable value or `x = x / y`"
         " if you want a new python Tensor object.", 1)
     return self / other
 
   def __irealdiv__(self, other):
     logging.log_first_n(
         logging.WARN,
-        "Variable /= will be deprecated. Use variable.assign_div"
-        " if you want assignment to the variable value or 'x = x / y'"
+        "Variable /= will be deprecated. Use `var.assign(var / other)`"
+        " if you want assignment to the variable value or `x = x / y`"
         " if you want a new python Tensor object.", 1)
     return self / other
 
   def __ipow__(self, other):
     logging.log_first_n(
         logging.WARN,
-        "Variable **= will be deprecated. Use 'x = x ** y'"
+        "Variable **= will be deprecated. Use `var.assign(var ** other)`"
+        " if you want assignment to the variable value or `x = x ** y`"
         " if you want a new python Tensor object.", 1)
     return self ** other
 
@@ -1403,6 +1404,10 @@ class PartitionedVariable(object):
   def dtype(self):
     return self._dtype
 
+  @property
+  def shape(self):
+    return self.get_shape()
+
   def get_shape(self):
     return self._shape
 
@@ -1722,6 +1727,8 @@ def report_uninitialized_variables(var_list=None,
           var_list.append(op.outputs[0])
   with ops.name_scope(name):
     # Run all operations on CPU
+    if var_list:
+      init_vars = [state_ops.is_variable_initialized(v) for v in var_list]
     with ops.device("/cpu:0"):
       if not var_list:
         # Return an empty tensor so we only need to check for returned tensor
@@ -1729,9 +1736,7 @@ def report_uninitialized_variables(var_list=None,
         return array_ops.constant([], dtype=dtypes.string)
       else:
         # Get a 1-D boolean tensor listing whether each variable is initialized.
-        variables_mask = math_ops.logical_not(
-            array_ops.stack(
-                [state_ops.is_variable_initialized(v) for v in var_list]))
+        variables_mask = math_ops.logical_not(array_ops.stack(init_vars))
         # Get a 1-D string tensor containing all the variable names.
         variable_names_tensor = array_ops.constant(
             [s.op.name for s in var_list])
diff --git a/tensorflow/python/platform/benchmark.py b/tensorflow/python/platform/benchmark.py
index eba2baaf6f836c872c8315e558c51733fc013ec2..fa17b17d104221990ed7847b725c4b741cb4aca7 100644
--- a/tensorflow/python/platform/benchmark.py
+++ b/tensorflow/python/platform/benchmark.py
@@ -66,11 +66,11 @@ def _global_report_benchmark(
     if not isinstance(extras, dict):
       raise TypeError("extras must be a dict")
 
-    logging.info("Benchmark [%s] iters: %d, wall_time: %g, cpu_time: %g,"
-                 "throughput: %g %s", name, iters if iters is not None else -1,
-                 wall_time if wall_time is not None else -1, cpu_time if
-                 cpu_time is not None else -1, throughput if
-                 throughput is not None else -1, str(extras) if extras else "")
+  logging.info("Benchmark [%s] iters: %d, wall_time: %g, cpu_time: %g,"
+               "throughput: %g %s", name, iters if iters is not None else -1,
+               wall_time if wall_time is not None else -1, cpu_time if
+               cpu_time is not None else -1, throughput if
+               throughput is not None else -1, str(extras) if extras else "")
 
   entries = test_log_pb2.BenchmarkEntries()
   entry = entries.entry.add()
diff --git a/tensorflow/python/platform/self_check.py b/tensorflow/python/platform/self_check.py
index 966a094e55e09d51c2d5edd36eb3ca29e71935f8..844ae999186f6eed89b113469782840f08502a85 100644
--- a/tensorflow/python/platform/self_check.py
+++ b/tensorflow/python/platform/self_check.py
@@ -78,7 +78,7 @@ def preload_check():
               "Could not find %r. TensorFlow requires that this DLL be "
               "installed in a directory that is named in your %%PATH%% "
               "environment variable. Download and install CUDA %s from "
-              "this URL: https://developer.nvidia.com/cuda-toolkit"
+              "this URL: https://developer.nvidia.com/cuda-90-download-archive"
               % (build_info.cudart_dll_name, build_info.cuda_version_number))
 
       if hasattr(build_info, "cudnn_dll_name") and hasattr(
diff --git a/tensorflow/tools/api/generator/BUILD b/tensorflow/python/tools/api/generator/BUILD
similarity index 71%
rename from tensorflow/tools/api/generator/BUILD
rename to tensorflow/python/tools/api/generator/BUILD
index 8c760e6f52598a5e7399c9250adf99283572d3a4..223d1281ba42afdcb72c84c249471d2dff13722d 100644
--- a/tensorflow/tools/api/generator/BUILD
+++ b/tensorflow/python/tools/api/generator/BUILD
@@ -3,8 +3,9 @@
 
 licenses(["notice"])  # Apache 2.0
 
-load("//tensorflow/tools/api/generator:api_gen.bzl", "ESTIMATOR_API_INIT_FILES")
-load("//tensorflow/tools/api/generator:api_gen.bzl", "TENSORFLOW_API_INIT_FILES")
+load("//tensorflow:tensorflow.bzl", "py_test")
+load("//tensorflow/python/tools/api/generator:api_gen.bzl", "ESTIMATOR_API_INIT_FILES")
+load("//tensorflow/python/tools/api/generator:api_gen.bzl", "TENSORFLOW_API_INIT_FILES")
 
 exports_files(
     [
@@ -13,6 +14,18 @@ exports_files(
     ],
 )
 
+py_binary(
+    name = "create_python_api",
+    srcs = ["//tensorflow/python/tools/api/generator:create_python_api.py"],
+    main = "//tensorflow/python/tools/api/generator:create_python_api.py",
+    srcs_version = "PY2AND3",
+    visibility = ["//visibility:public"],
+    deps = [
+        "//tensorflow/python:no_contrib",
+        "//tensorflow/python/tools/api/generator:doc_srcs",
+    ],
+)
+
 py_library(
     name = "doc_srcs",
     srcs = ["doc_srcs.py"],
diff --git a/tensorflow/tools/api/generator/api_gen.bzl b/tensorflow/python/tools/api/generator/api_gen.bzl
similarity index 70%
rename from tensorflow/tools/api/generator/api_gen.bzl
rename to tensorflow/python/tools/api/generator/api_gen.bzl
index d746b5d3e4f7745d78563eac65ccdf822511a7ef..2a32e8a8933be601b56dc833f583cbb3d807f200 100644
--- a/tensorflow/tools/api/generator/api_gen.bzl
+++ b/tensorflow/python/tools/api/generator/api_gen.bzl
@@ -102,36 +102,41 @@ ESTIMATOR_API_INIT_FILES = [
     # END GENERATED ESTIMATOR FILES
 ]
 
-# Creates a genrule that generates a directory structure with __init__.py
-# files that import all exported modules (i.e. modules with tf_export
-# decorators).
-#
-# Args:
-#   name: name of genrule to create.
-#   output_files: List of __init__.py files that should be generated.
-#     This list should include file name for every module exported using
-#     tf_export. For e.g. if an op is decorated with
-#     @tf_export('module1.module2', 'module3'). Then, output_files should
-#     include module1/module2/__init__.py and module3/__init__.py.
-#   root_init_template: Python init file that should be used as template for
-#     root __init__.py file. "# API IMPORTS PLACEHOLDER" comment inside this
-#     template will be replaced with root imports collected by this genrule.
-#   srcs: genrule sources. If passing root_init_template, the template file
-#     must be included in sources.
-#   api_name: Name of the project that you want to generate API files for
-#     (e.g. "tensorflow" or "estimator").
-#   package: Python package containing the @tf_export decorators you want to
-#     process
-#   package_dep: Python library target containing your package.
-
 def gen_api_init_files(
         name,
         output_files = TENSORFLOW_API_INIT_FILES,
         root_init_template = None,
         srcs = [],
         api_name = "tensorflow",
+        api_version = 2,
         package = "tensorflow.python",
-        package_dep = "//tensorflow/python:no_contrib"):
+        package_dep = "//tensorflow/python:no_contrib",
+        output_package = "tensorflow"):
+    """Creates API directory structure and __init__.py files.
+
+    Creates a genrule that generates a directory structure with __init__.py
+    files that import all exported modules (i.e. modules with tf_export
+    decorators).
+
+    Args:
+      name: name of genrule to create.
+      output_files: List of __init__.py files that should be generated.
+        This list should include file name for every module exported using
+        tf_export. For e.g. if an op is decorated with
+        @tf_export('module1.module2', 'module3'). Then, output_files should
+        include module1/module2/__init__.py and module3/__init__.py.
+      root_init_template: Python init file that should be used as template for
+        root __init__.py file. "# API IMPORTS PLACEHOLDER" comment inside this
+        template will be replaced with root imports collected by this genrule.
+      srcs: genrule sources. If passing root_init_template, the template file
+        must be included in sources.
+      api_name: Name of the project that you want to generate API files for
+        (e.g. "tensorflow" or "estimator").
+      api_version: TensorFlow API version to generate. Must be either 1 or 2.
+      package: Python package containing the @tf_export decorators you want to
+        process
+      package_dep: Python library target containing your package.
+    """
     root_init_template_flag = ""
     if root_init_template:
       root_init_template_flag = "--root_init_template=$(location " + root_init_template + ")"
@@ -139,13 +144,13 @@ def gen_api_init_files(
     api_gen_binary_target = "create_" + package + "_api"
     native.py_binary(
         name = "create_" + package + "_api",
-        srcs = ["//tensorflow/tools/api/generator:create_python_api.py"],
-        main = "//tensorflow/tools/api/generator:create_python_api.py",
+        srcs = ["//tensorflow/python/tools/api/generator:create_python_api.py"],
+        main = "//tensorflow/python/tools/api/generator:create_python_api.py",
         srcs_version = "PY2AND3",
         visibility = ["//visibility:public"],
         deps = [
             package_dep,
-            "//tensorflow/tools/api/generator:doc_srcs",
+            "//tensorflow/python/tools/api/generator:doc_srcs",
         ],
     )
 
@@ -154,7 +159,9 @@ def gen_api_init_files(
         outs = output_files,
         cmd = (
             "$(location :" + api_gen_binary_target + ") " +
-            root_init_template_flag + " --apidir=$(@D) --apiname=" + api_name + " --package=" + package + " $(OUTS)"),
+            root_init_template_flag + " --apidir=$(@D) --apiname=" +
+            api_name + " --apiversion=" + str(api_version) + " --package=" + package +
+            " --output_package=" + output_package + " $(OUTS)"),
         srcs = srcs,
         tools = [":" + api_gen_binary_target ],
         visibility = ["//tensorflow:__pkg__"],
diff --git a/tensorflow/tools/api/generator/create_python_api.py b/tensorflow/python/tools/api/generator/create_python_api.py
similarity index 89%
rename from tensorflow/tools/api/generator/create_python_api.py
rename to tensorflow/python/tools/api/generator/create_python_api.py
index 48d7dcd09eb38f53031afde70fe2e1a9b660ad1a..863c922216fa275fa8a9dda04a212a32a57551c0 100644
--- a/tensorflow/tools/api/generator/create_python_api.py
+++ b/tensorflow/python/tools/api/generator/create_python_api.py
@@ -24,11 +24,12 @@ import importlib
 import os
 import sys
 
+from tensorflow.python.tools.api.generator import doc_srcs
 from tensorflow.python.util import tf_decorator
 from tensorflow.python.util import tf_export
-from tensorflow.tools.api.generator import doc_srcs
 
 API_ATTRS = tf_export.API_ATTRS
+API_ATTRS_V1 = tf_export.API_ATTRS_V1
 
 _DEFAULT_PACKAGE = 'tensorflow.python'
 _GENFILES_DIR_SUFFIX = 'genfiles/'
@@ -38,14 +39,14 @@ _SYMBOLS_TO_SKIP_EXPLICITLY = {
     'tensorflow.python.platform.flags.FLAGS'
 }
 _GENERATED_FILE_HEADER = """# This file is MACHINE GENERATED! Do not edit.
-# Generated by: tensorflow/tools/api/generator/create_python_api.py script.
+# Generated by: tensorflow/python/tools/api/generator/create_python_api.py script.
 \"\"\"%s
 \"\"\"
 
 from __future__ import print_function
 
 """
-_GENERATED_FILE_FOOTER = "\n\ndel print_function\n"
+_GENERATED_FILE_FOOTER = '\n\ndel print_function\n'
 
 
 class SymbolExposedTwiceError(Exception):
@@ -159,13 +160,16 @@ __all__.remove('print_function')
     return module_text_map
 
 
-def get_api_init_text(package, api_name):
+def get_api_init_text(package, output_package, api_name, api_version):
   """Get a map from destination module to __init__.py code for that module.
 
   Args:
     package: Base python package containing python with target tf_export
       decorators.
+    output_package: Base output python package where generated API will
+      be added.
     api_name: API you want to generate (e.g. `tensorflow` or `estimator`).
+    api_version: API version you want to generate (`v1` or `v2`).
 
   Returns:
     A dictionary where
@@ -173,6 +177,12 @@ def get_api_init_text(package, api_name):
       value: (string) text that should be in __init__.py files for
         corresponding modules.
   """
+  if api_version == 1:
+    names_attr = API_ATTRS_V1[api_name].names
+    constants_attr = API_ATTRS_V1[api_name].constants
+  else:
+    names_attr = API_ATTRS[api_name].names
+    constants_attr = API_ATTRS[api_name].constants
   module_code_builder = _ModuleInitCodeBuilder()
 
   # Traverse over everything imported above. Specifically,
@@ -193,7 +203,7 @@ def get_api_init_text(package, api_name):
       attr = getattr(module, module_contents_name)
 
       # If attr is _tf_api_constants attribute, then add the constants.
-      if module_contents_name == API_ATTRS[api_name].constants:
+      if module_contents_name == constants_attr:
         for exports, value in attr:
           for export in exports:
             names = export.split('.')
@@ -205,9 +215,8 @@ def get_api_init_text(package, api_name):
       _, attr = tf_decorator.unwrap(attr)
       # If attr is a symbol with _tf_api_names attribute, then
       # add import for it.
-      if (hasattr(attr, '__dict__') and
-          API_ATTRS[api_name].names in attr.__dict__):
-        for export in getattr(attr, API_ATTRS[api_name].names):  # pylint: disable=protected-access
+      if (hasattr(attr, '__dict__') and names_attr in attr.__dict__):
+        for export in getattr(attr, names_attr):  # pylint: disable=protected-access
           names = export.split('.')
           dest_module = '.'.join(names[:-1])
           module_code_builder.add_import(
@@ -218,7 +227,6 @@ def get_api_init_text(package, api_name):
   # For e.g. if we import 'foo.bar.Value'. Then, we also
   # import 'bar' in 'foo'.
   imported_modules = set(module_code_builder.module_imports.keys())
-  import_from = '.'
   for module in imported_modules:
     if not module:
       continue
@@ -229,6 +237,9 @@ def get_api_init_text(package, api_name):
       if submodule_index > 0:
         parent_module += ('.' + module_split[submodule_index-1] if parent_module
                           else module_split[submodule_index-1])
+      import_from = output_package
+      if submodule_index > 0:
+        import_from += '.' + '.'.join(module_split[:submodule_index])
       module_code_builder.add_import(
           -1, parent_module, import_from,
           module_split[submodule_index], module_split[submodule_index])
@@ -294,7 +305,8 @@ def get_module_docstring(module_name, package, api_name):
 
 
 def create_api_files(
-    output_files, package, root_init_template, output_dir, api_name):
+    output_files, package, root_init_template, output_dir, output_package,
+    api_name, api_version):
   """Creates __init__.py files for the Python API.
 
   Args:
@@ -306,7 +318,9 @@ def create_api_files(
       "#API IMPORTS PLACEHOLDER" comment in the template file will be replaced
       with imports.
     output_dir: output API root directory.
+    output_package: Base output package where generated API will be added.
     api_name: API you want to generate (e.g. `tensorflow` or `estimator`).
+    api_version: API version to generate (`v1` or `v2`).
 
   Raises:
     ValueError: if an output file is not under api/ directory,
@@ -323,7 +337,8 @@ def create_api_files(
       os.makedirs(os.path.dirname(file_path))
     open(file_path, 'a').close()
 
-  module_text_map = get_api_init_text(package, api_name)
+  module_text_map = get_api_init_text(
+      package, output_package, api_name, api_version)
 
   # Add imports to output files.
   missing_output_files = []
@@ -381,6 +396,13 @@ def main():
       '--apiname', required=True, type=str,
       choices=API_ATTRS.keys(),
       help='The API you want to generate.')
+  parser.add_argument(
+      '--apiversion', default=2, type=int,
+      choices=[1, 2],
+      help='The API version you want to generate.')
+  parser.add_argument(
+      '--output_package', default='tensorflow', type=str,
+      help='Root output package.')
 
   args = parser.parse_args()
 
@@ -395,7 +417,8 @@ def main():
   # Populate `sys.modules` with modules containing tf_export().
   importlib.import_module(args.package)
   create_api_files(outputs, args.package, args.root_init_template,
-                   args.apidir, args.apiname)
+                   args.apidir, args.output_package, args.apiname,
+                   args.apiversion)
 
 
 if __name__ == '__main__':
diff --git a/tensorflow/tools/api/generator/create_python_api_test.py b/tensorflow/python/tools/api/generator/create_python_api_test.py
similarity index 90%
rename from tensorflow/tools/api/generator/create_python_api_test.py
rename to tensorflow/python/tools/api/generator/create_python_api_test.py
index 651ec9d040302a4343ae6e0053cf6a4b37a971d4..a565a49d967d3b850058f5370272cfedb43791f4 100644
--- a/tensorflow/tools/api/generator/create_python_api_test.py
+++ b/tensorflow/python/tools/api/generator/create_python_api_test.py
@@ -22,8 +22,8 @@ import imp
 import sys
 
 from tensorflow.python.platform import test
+from tensorflow.python.tools.api.generator import create_python_api
 from tensorflow.python.util.tf_export import tf_export
-from tensorflow.tools.api.generator import create_python_api
 
 
 @tf_export('test_op', 'test_op1')
@@ -58,7 +58,8 @@ class CreatePythonApiTest(test.TestCase):
   def testFunctionImportIsAdded(self):
     imports = create_python_api.get_api_init_text(
         package=create_python_api._DEFAULT_PACKAGE,
-        api_name='tensorflow')
+        output_package='tensorflow',
+        api_name='tensorflow', api_version=1)
     expected_import = (
         'from tensorflow.python.test_module '
         'import test_op as test_op1')
@@ -75,7 +76,8 @@ class CreatePythonApiTest(test.TestCase):
   def testClassImportIsAdded(self):
     imports = create_python_api.get_api_init_text(
         package=create_python_api._DEFAULT_PACKAGE,
-        api_name='tensorflow')
+        output_package='tensorflow',
+        api_name='tensorflow', api_version=2)
     expected_import = ('from tensorflow.python.test_module '
                        'import TestClass')
     self.assertTrue(
@@ -85,7 +87,8 @@ class CreatePythonApiTest(test.TestCase):
   def testConstantIsAdded(self):
     imports = create_python_api.get_api_init_text(
         package=create_python_api._DEFAULT_PACKAGE,
-        api_name='tensorflow')
+        output_package='tensorflow',
+        api_name='tensorflow', api_version=1)
     expected = ('from tensorflow.python.test_module '
                 'import _TEST_CONSTANT')
     self.assertTrue(expected in str(imports),
diff --git a/tensorflow/tools/api/generator/doc_srcs.py b/tensorflow/python/tools/api/generator/doc_srcs.py
similarity index 100%
rename from tensorflow/tools/api/generator/doc_srcs.py
rename to tensorflow/python/tools/api/generator/doc_srcs.py
diff --git a/tensorflow/tools/api/generator/doc_srcs_test.py b/tensorflow/python/tools/api/generator/doc_srcs_test.py
similarity index 75%
rename from tensorflow/tools/api/generator/doc_srcs_test.py
rename to tensorflow/python/tools/api/generator/doc_srcs_test.py
index 7b8f27c1b1cd474462d0eab2ddc8451bb256496b..481d9874a4bcdcdadcdcb16b5b5c1b10b765dc48 100644
--- a/tensorflow/tools/api/generator/doc_srcs_test.py
+++ b/tensorflow/python/tools/api/generator/doc_srcs_test.py
@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 # =============================================================================
-"""Tests for tensorflow.tools.api.generator.doc_srcs."""
+"""Tests for tensorflow.python.tools.api.generator.doc_srcs."""
 
 from __future__ import absolute_import
 from __future__ import division
@@ -23,7 +23,7 @@ import importlib
 import sys
 
 from tensorflow.python.platform import test
-from tensorflow.tools.api.generator import doc_srcs
+from tensorflow.python.tools.api.generator import doc_srcs
 
 
 FLAGS = None
@@ -39,27 +39,27 @@ class DocSrcsTest(test.TestCase):
         file_path += '/'
       file_path += '__init__.py'
 
-      if file_path not in FLAGS.outputs:
-        self.assertFalse('%s is not a valid API module' % module_name)
+      self.assertIn(
+          file_path, FLAGS.outputs,
+          msg='%s is not a valid API module' % module_name)
 
   def testHaveDocstringOrDocstringModule(self):
     for module_name, docsrc in doc_srcs.get_doc_sources(FLAGS.api_name).items():
-      if docsrc.docstring and docsrc.docstring_module_name:
-        self.assertFalse(
-            '%s contains DocSource has both a docstring and a '
-            'docstring_module_name. '
-            'Only one of "docstring" or "docstring_module_name" should be set.'
-            % (module_name))
+      self.assertFalse(
+          docsrc.docstring and docsrc.docstring_module_name,
+          msg=('%s contains DocSource has both a docstring and a '
+               'docstring_module_name. Only one of "docstring" or '
+               '"docstring_module_name" should be set.') % (module_name))
 
   def testDocstringModulesAreValidModules(self):
     for _, docsrc in doc_srcs.get_doc_sources(FLAGS.api_name).items():
       if docsrc.docstring_module_name:
         doc_module_name = '.'.join([
             FLAGS.package, docsrc.docstring_module_name])
-        if doc_module_name not in sys.modules:
-          self.assertFalse(
-              'docsources_module %s is not a valid module under %s.' %
-              (docsrc.docstring_module_name, FLAGS.package))
+        self.assertIn(
+            doc_module_name, sys.modules,
+            msg=('docsources_module %s is not a valid module under %s.' %
+                 (docsrc.docstring_module_name, FLAGS.package)))
 
 
 if __name__ == '__main__':
diff --git a/tensorflow/python/training/checkpointable/BUILD b/tensorflow/python/training/checkpointable/BUILD
index 9232b6089a5775b96c963989f15475f90aaa621b..35007653a09f4b4990be19ef6b14bf6084a7f14c 100644
--- a/tensorflow/python/training/checkpointable/BUILD
+++ b/tensorflow/python/training/checkpointable/BUILD
@@ -47,6 +47,7 @@ py_library(
     srcs_version = "PY2AND3",
     deps = [
         ":base",
+        ":data_structures",
     ],
 )
 
@@ -61,12 +62,19 @@ py_test(
     ],
 )
 
+py_library(
+    name = "layer_utils",
+    srcs = ["layer_utils.py"],
+    srcs_version = "PY2AND3",
+)
+
 py_library(
     name = "data_structures",
     srcs = ["data_structures.py"],
     srcs_version = "PY2AND3",
     deps = [
         ":base",
+        ":layer_utils",
     ],
 )
 
diff --git a/tensorflow/python/training/checkpointable/base.py b/tensorflow/python/training/checkpointable/base.py
index 99c8098eca236549ec5cff10ad6e79badb996a7d..ee35b01328436911fd7926b25b14433377ec4188 100644
--- a/tensorflow/python/training/checkpointable/base.py
+++ b/tensorflow/python/training/checkpointable/base.py
@@ -33,6 +33,7 @@ from tensorflow.python.platform import tf_logging as logging
 from tensorflow.python.training import saveable_object
 from tensorflow.python.util import nest
 from tensorflow.python.util import serialization
+from tensorflow.python.util import tf_decorator
 
 
 # Key where the object graph proto is saved in a TensorBundle
@@ -340,6 +341,34 @@ _SlotVariableRestoration = collections.namedtuple(
     ])
 
 
+def no_automatic_dependency_tracking(method):
+  """Disables automatic dependency tracking on attribute assignment.
+
+  Use to decorate any method of a Checkpointable object. Attribute assignment in
+  that method will not add dependencies (also respected in Model). Harmless if
+  used in a class which does not do automatic dependency tracking (which means
+  it's safe to use in base classes which may have subclasses which also inherit
+  from Checkpointable).
+
+  Args:
+    method: The method to decorate.
+  Returns:
+    A decorated method which sets and un-sets automatic dependency tracking for
+    the object the method is called on (not thread safe).
+  """
+
+  def _method_wrapper(self, *args, **kwargs):
+    previous_value = getattr(self, "_setattr_tracking", True)
+    self._setattr_tracking = False  # pylint: disable=protected-access
+    try:
+      method(self, *args, **kwargs)
+    finally:
+      self._setattr_tracking = previous_value  # pylint: disable=protected-access
+
+  return tf_decorator.make_decorator(
+      target=method, decorator_func=_method_wrapper)
+
+
 class CheckpointableBase(object):
   """Base class for `Checkpointable` objects without automatic dependencies.
 
@@ -349,6 +378,11 @@ class CheckpointableBase(object):
   checks.
   """
 
+  # CheckpointableBase does not do automatic dependency tracking, but uses the
+  # no_automatic_dependency_tracking decorator so it can avoid adding
+  # dependencies if a subclass is Checkpointable / inherits from Model (both of
+  # which have __setattr__ overrides).
+  @no_automatic_dependency_tracking
   def _maybe_initialize_checkpointable(self):
     """Initialize dependency management.
 
@@ -386,6 +420,10 @@ class CheckpointableBase(object):
     # building.
     self._name_based_restores = set()
 
+  def _no_dependency(self, value):
+    """If automatic dependency tracking is enabled, ignores `value`."""
+    return value
+
   def _name_based_attribute_restore(self, checkpoint):
     """Restore the object's attributes from a name-based checkpoint."""
     self._name_based_restores.add(checkpoint)
@@ -463,7 +501,7 @@ class CheckpointableBase(object):
       ValueError: If the variable name is not unique.
     """
     self._maybe_initialize_checkpointable()
-    if not overwrite and self._lookup_dependency(name) is not None:
+    if overwrite and self._lookup_dependency(name) is not None:
       raise ValueError(
           ("A variable named '%s' already exists in this Checkpointable, but "
            "Checkpointable._add_variable called to create another with "
@@ -593,9 +631,9 @@ class CheckpointableBase(object):
           self._unconditional_checkpoint_dependencies[index] = new_reference
     elif current_object is None:
       self._unconditional_checkpoint_dependencies.append(new_reference)
-      self._unconditional_dependency_names[name] = checkpointable
       self._handle_deferred_dependencies(
           name=name, checkpointable=checkpointable)
+    self._unconditional_dependency_names[name] = checkpointable
     return checkpointable
 
   def _handle_deferred_dependencies(self, name, checkpointable):
@@ -733,28 +771,3 @@ class CheckpointableBase(object):
     return {OBJECT_CONFIG_JSON_KEY: functools.partial(
         PythonStringStateSaveable,
         state_callback=_state_callback)}
-
-
-class NoDependency(object):
-  """Allows attribute assignment to `Checkpointable` objects with no dependency.
-
-  Example usage:
-  ```python
-  obj = Checkpointable()
-  obj.has_dependency = tf.Variable(0., name="dep")
-  obj.no_dependency = NoDependency(tf.Variable(1., name="nodep"))
-  assert obj.no_dependency.name == "nodep:0"
-  ```
-
-  `obj` in this example has a dependency on the variable "dep", and both
-  attributes contain un-wrapped `Variable` objects.
-
-  `NoDependency` also works with `tf.keras.Model`, but only for checkpoint
-  dependencies: wrapping a `Layer` in `NoDependency` will assign the (unwrapped)
-  `Layer` to the attribute without a checkpoint dependency, but the `Model` will
-  still track the `Layer` (so it will appear in `Model.layers`, and its
-  variables will appear in `Model.variables`).
-  """
-
-  def __init__(self, value):
-    self.value = value
diff --git a/tensorflow/python/training/checkpointable/data_structures.py b/tensorflow/python/training/checkpointable/data_structures.py
index 680cf3441f8b2095ea0837f92c10978ed2d326ab..019d43f09c10a4975a9b483593af30b5bbe06089 100644
--- a/tensorflow/python/training/checkpointable/data_structures.py
+++ b/tensorflow/python/training/checkpointable/data_structures.py
@@ -21,50 +21,127 @@ import collections
 
 import six
 
-from tensorflow.python.keras.engine import base_layer
-from tensorflow.python.keras.utils import layer_utils
 from tensorflow.python.ops import variables
-from tensorflow.python.training.checkpointable import base as checkpointable_lib
-
-
-# TODO(allenl): We could track regular Python data structures which get assigned
-# to Checkpointable objects. Making this work with restore-on-create would be
-# tricky; we'd need to re-create nested structures with our own wrapped objects
-# on assignment to an attribute, and track the user's original structure to make
-# sure they don't modify it except through the wrappers (since we could save the
-# user's updated structure, but would have no way to support restore-on-create
-# for those modifications).
-# TODO(allenl): A dictionary data structure would be good too.
-class CheckpointableDataStructure(checkpointable_lib.CheckpointableBase):
+from tensorflow.python.training.checkpointable import base
+from tensorflow.python.training.checkpointable import layer_utils
+
+
+class NoDependency(object):
+  """Allows attribute assignment to `Checkpointable` objects with no dependency.
+
+  Example usage:
+  ```python
+  obj = Checkpointable()
+  obj.has_dependency = tf.Variable(0., name="dep")
+  obj.no_dependency = NoDependency(tf.Variable(1., name="nodep"))
+  assert obj.no_dependency.name == "nodep:0"
+  ```
+
+  `obj` in this example has a dependency on the variable "dep", and both
+  attributes contain un-wrapped `Variable` objects.
+
+  `NoDependency` also works with `tf.keras.Model`, but only for checkpoint
+  dependencies: wrapping a `Layer` in `NoDependency` will assign the (unwrapped)
+  `Layer` to the attribute without a checkpoint dependency, but the `Model` will
+  still track the `Layer` (so it will appear in `Model.layers`, and its
+  variables will appear in `Model.variables`).
+  """
+
+  def __init__(self, value):
+    self.value = value
+
+
+def _wrap_or_unwrap(value):
+  """Wraps basic data structures, unwraps NoDependency objects."""
+  if isinstance(value, NoDependency):
+    return value.value
+  if isinstance(value, base.CheckpointableBase):
+    return value  # Skip conversion for already checkpointable objects.
+  elif isinstance(value, list):
+    return _ListWrapper(value)
+  else:
+    return value
+  # TODO(allenl): Handle other common data structures. Tuples will require
+  # special casing (tuple subclasses are not weak referenceable, so replacement
+  # with a wrapper that subclasses tuple on attribute assignment works poorly,
+  # and replacement with a wrapper that isn't a tuple is also problematic),
+  # probably a tree traversal where the leaves are non-tuples(/namedtuples) to
+  # come up with names. Dictionaries should look like lists.
+
+
+def sticky_attribute_assignment(checkpointable, name, value):
+  """Adds dependencies, generally called from __setattr__.
+
+  This behavior is shared between Checkpointable and Model.
+
+  Respects NoDependency indicators, but otherwise makes checkpointable objects
+  out of common data structures and tracks objects by their attribute names.
+
+  Args:
+    checkpointable: The object to add dependencies to (generally the one having
+      an attribute assigned).
+    name: The attribute name being assigned.
+    value: The value being assigned. Not necessarily a checkpointable object.
+
+  Returns:
+    The value which should be stored in the attribute (unwrapped from a
+    NoDependency object if necessary).
+  """
+  if isinstance(value, NoDependency):
+    add_dependency = False
+  else:
+    add_dependency = True
+  value = _wrap_or_unwrap(value)
+  if not add_dependency:
+    return value
+  if isinstance(value, base.CheckpointableBase):
+    checkpointable._track_checkpointable(  # pylint: disable=protected-access
+        value, name=name,
+        # Allow the user to switch the Checkpointable which is tracked by this
+        # name, since assigning a new variable to an attribute has
+        # historically been fine (e.g. Adam did this).
+        overwrite=True)
+  return value
+
+
+class CheckpointableDataStructure(base.CheckpointableBase):
   """Base class for data structures which contain checkpointable objects."""
 
   def __init__(self):
+    # An append-only ordered set
     self._layers = []
+
     self.trainable = True
     self._extra_variables = []
 
   def _track_value(self, value, name):
     """Add a dependency on `value`."""
-    if isinstance(value, checkpointable_lib.CheckpointableBase):
-      self._track_checkpointable(value, name=name)
-      if isinstance(value, variables.Variable):
-        self._extra_variables.append(value)
-    else:
+    value = sticky_attribute_assignment(
+        checkpointable=self, value=value, name=name)
+    if isinstance(value, variables.Variable):
+      self._extra_variables.append(value)
+    if not isinstance(value, base.CheckpointableBase):
       raise ValueError(
           ("Only checkpointable objects (such as Layers or Optimizers) may be "
            "stored in a List object. Got %s, which does not inherit from "
            "CheckpointableBase.") % (value,))
-    if isinstance(value, (base_layer.Layer, CheckpointableDataStructure)):
-      if value not in self._layers:
+    if (isinstance(value, CheckpointableDataStructure)
+        or layer_utils.is_layer(value)):
+      # Check for object-identity rather than with __eq__ to avoid
+      # de-duplicating empty container types. Automatically generated list
+      # wrappers keep things like "[] == []" true, which means "[] in [[]]" is
+      # also true. This becomes not true once one of the lists is mutated.
+      if not any((layer is value for layer in self._layers)):
         self._layers.append(value)
         if hasattr(value, "_use_resource_variables"):
           # In subclassed models, legacy layers (tf.layers) must always use
           # resource variables.
           value._use_resource_variables = True  # pylint: disable=protected-access
+    return value
 
   @property
   def layers(self):
-    return self._layers
+    return layer_utils.filter_empty_layer_containers(self._layers)
 
   @property
   def trainable_weights(self):
@@ -164,24 +241,28 @@ class List(CheckpointableDataStructure, collections.Sequence):
   def __init__(self, *args, **kwargs):
     """Construct a new sequence. Arguments are passed to `list()`."""
     super(List, self).__init__()
-    self._storage = list(*args, **kwargs)
+    self._storage = self._make_storage(*args, **kwargs)
     for index, element in enumerate(self._storage):
-      self._track_value(element, name=self._name_element(index))
+      self._storage[index] = self._track_value(
+          element, name=self._name_element(index))
+
+  def _make_storage(self, *args, **kwargs):
+    """Determines the backing storage (overridden in subclasses)."""
+    return list(*args, **kwargs)
 
   def _name_element(self, index):
     return "%d" % (index,)
 
   def append(self, value):
     """Add a new checkpointable value."""
-    self._track_value(value, self._name_element(len(self._storage)))
+    value = self._track_value(value, self._name_element(len(self._storage)))
     self._storage.append(value)
 
   def extend(self, values):
     """Add a sequence of checkpointable values."""
-    for index_offset, value in enumerate(values):
-      self._track_value(
-          value, name=self._name_element(len(self._storage) + index_offset))
-    self._storage.extend(values)
+    for value in values:
+      self._storage.append(self._track_value(
+          value, name=self._name_element(len(self._storage))))
 
   def __iadd__(self, values):
     self.extend(values)
@@ -189,9 +270,12 @@ class List(CheckpointableDataStructure, collections.Sequence):
 
   def __add__(self, other):
     if isinstance(other, List):
-      return List(self._storage + other._storage)  # pylint: disable=protected-access
+      return self.__class__(self._storage + other._storage)  # pylint: disable=protected-access
     else:
-      return List(self._storage + other)
+      return self.__class__(self._storage + other)
+
+  def __radd__(self, other):
+    return self + other
 
   def __getitem__(self, key):
     return self._storage[key]
@@ -203,6 +287,144 @@ class List(CheckpointableDataStructure, collections.Sequence):
     return "List(%s)" % (repr(self._storage),)
 
 
+class _ListWrapper(List, collections.MutableSequence,
+                   # Shadowed, but there for isinstance checks.
+                   list):
+  """Wraps the built-in `list` to support restore-on-create for variables.
+
+  Unlike `List`, this sequence type is mutable in the same ways built-in lists
+  are. Instead of throwing an error immediately like `List`, it records
+  problematic mutations (e.g. assigning a new element to a position already
+  occupied, meaning both elements get the same names at different times) and
+  refuses to save.
+
+  On assignment to an attribute of a Model or Checkpointable object, Python
+  lists are replaced with _ListWrapper. Wrapping a list in a
+  `tf.contrib.checkpoint.NoDependency` object prevents this.
+  """
+
+  def __init__(self, wrapped_list):
+    """Construct a new list wrapper.
+
+    Args:
+      wrapped_list: The initial value of the data structure. A shallow copy may
+        be maintained for error checking. `wrapped_list` itself should not be
+        modified directly after constructing the `_ListWrapper`, and if changes
+        are detected the `_ListWrapper` will throw an exception on save.
+    """
+    # Monotonic flags which indicate this object would not be restored properly,
+    # and therefore should throw an error on save to avoid giving the impression
+    # that restoring it will work.
+    self._non_append_mutation = False
+    self._external_modification = False
+    super(_ListWrapper, self).__init__(wrapped_list)
+    self._last_wrapped_list_snapshot = list(self._storage)
+
+  def _make_storage(self, wrapped_list):
+    """Use the user's original list for storage."""
+    return wrapped_list
+
+  def _check_external_modification(self):
+    """Checks for any changes to the wrapped list not through the wrapper."""
+    if self._external_modification or self._non_append_mutation:
+      return
+    if self._storage != self._last_wrapped_list_snapshot:
+      self._external_modification = True
+      self._last_wrapped_list_snapshot = None
+
+  def _update_snapshot(self):
+    """Acknowledges tracked changes to the wrapped list."""
+    if self._external_modification or self._non_append_mutation:
+      return
+    self._last_wrapped_list_snapshot = list(self._storage)
+
+  @property
+  def _checkpoint_dependencies(self):
+    self._check_external_modification()
+    if self._non_append_mutation:
+      raise ValueError(
+          ("Unable to save the object %s (a list wrapper constructed to track "
+           "checkpointable TensorFlow objects). A list element was replaced "
+           "(__setitem__), deleted, or inserted. In order to support "
+           "restoration on object creation, tracking is exclusively for "
+           "append-only data structures.\n\nIf you don't need this list "
+           "checkpointed, wrap it in a tf.contrib.checkpoint.NoDependency "
+           "object; it will be automatically un-wrapped and subsequently "
+           "ignored." % (self,)))
+    if self._external_modification:
+      raise ValueError(
+          ("Unable to save the object %s (a list wrapper constructed to track "
+           "checkpointable TensorFlow objects). The wrapped list was modified "
+           "outside the wrapper (its final value was %s, its value when a "
+           "checkpoint dependency was added was %s), which breaks restoration "
+           "on object creation.\n\nIf you don't need this list checkpointed, "
+           "wrap it in a tf.contrib.checkpoint.NoDependency object; it will be "
+           "automatically un-wrapped and subsequently ignored." % (
+               self, self._storage, self._last_wrapped_list_snapshot)))
+    return super(_ListWrapper, self)._checkpoint_dependencies
+
+  def __delitem__(self, key):
+    self._non_append_mutation = True
+    del self._storage[key]
+
+  def __setitem__(self, key, value):
+    self._non_append_mutation = True
+    self._storage[key] = value
+
+  def append(self, value):
+    """Add a new checkpointable value."""
+    self._check_external_modification()
+    super(_ListWrapper, self).append(value)
+    self._update_snapshot()
+
+  def extend(self, values):
+    """Add a sequence of checkpointable values."""
+    self._check_external_modification()
+    super(_ListWrapper, self).extend(values)
+    self._update_snapshot()
+
+  def __eq__(self, other):
+    return self._storage == getattr(other, "_storage", other)
+
+  def __ne__(self, other):
+    return self._storage != getattr(other, "_storage", other)
+
+  def __lt__(self, other):
+    return self._storage < getattr(other, "_storage", other)
+
+  def __le__(self, other):
+    return self._storage <= getattr(other, "_storage", other)
+
+  def __gt__(self, other):
+    return self._storage > getattr(other, "_storage", other)
+
+  def __ge__(self, other):
+    return self._storage >= getattr(other, "_storage", other)
+
+  def __hash__(self):
+    # List wrappers need to compare like regular lists, and so like regular
+    # lists they don't belong in hash tables.
+    raise TypeError("unhashable type: 'ListWrapper'")
+
+  def insert(self, index, obj):
+    self._non_append_mutation = True
+    self._storage.insert(index, obj)
+
+  def _track_value(self, value, name):
+    """Allows storage of non-checkpointable objects."""
+    try:
+      value = super(_ListWrapper, self)._track_value(value=value, name=name)
+    except ValueError:
+      # Even if this value isn't checkpointable, we need to make sure
+      # NoDependency objects get unwrapped.
+      value = sticky_attribute_assignment(
+          checkpointable=self, value=value, name=name)
+    return value
+
+  def __repr__(self):
+    return "ListWrapper(%s)" % (repr(self._storage),)
+
+
 class Mapping(CheckpointableDataStructure, collections.Mapping):
   """An append-only checkpointable mapping data structure with string keys.
 
@@ -217,8 +439,10 @@ class Mapping(CheckpointableDataStructure, collections.Mapping):
     """Construct a new sequence. Arguments are passed to `dict()`."""
     super(Mapping, self).__init__()
     self._storage = dict(*args, **kwargs)
-    for key, value in self._storage.items():
-      self._track_value(value, name=self._name_element(key))
+    self._storage.update(
+        {key: self._track_value(
+            value, name=self._name_element(key))
+         for key, value in self._storage.items()})
 
   def _name_element(self, key):
     if not isinstance(key, six.string_types):
@@ -228,13 +452,14 @@ class Mapping(CheckpointableDataStructure, collections.Mapping):
     return str(key)
 
   def __setitem__(self, key, value):
+    name = self._name_element(key)
+    value = self._track_value(value, name=name)
     current_value = self._storage.setdefault(key, value)
     if current_value is not value:
       raise ValueError(
           ("Mappings are an append-only data structure. Tried to overwrite the "
            "key '%s' with value %s, but it already contains %s")
           % (key, value, current_value))
-    self._track_value(value, name=self._name_element(key))
 
   def update(self, *args, **kwargs):
     for key, value in dict(*args, **kwargs).items():
diff --git a/tensorflow/python/training/checkpointable/data_structures_test.py b/tensorflow/python/training/checkpointable/data_structures_test.py
index ce5852dd6e1acbf36ef58a614148c12b9dbae039..ec8c9da8090c968e8931f96949f5b982dd94f215 100644
--- a/tensorflow/python/training/checkpointable/data_structures_test.py
+++ b/tensorflow/python/training/checkpointable/data_structures_test.py
@@ -31,6 +31,7 @@ from tensorflow.python.ops import array_ops
 from tensorflow.python.ops import math_ops
 from tensorflow.python.ops import resource_variable_ops
 from tensorflow.python.training.checkpointable import data_structures
+from tensorflow.python.training.checkpointable import tracking
 
 
 class HasList(training.Model):
@@ -113,6 +114,19 @@ class ListTests(test.TestCase):
     model(model_input)
     self.assertEqual(2, len(model.losses))
 
+  def testModelContainersCompareEqual(self):
+    class HasEqualContainers(training.Model):
+
+      def __init__(self):
+        super(HasEqualContainers, self).__init__()
+        self.l1 = []
+        self.l2 = []
+
+    model = HasEqualContainers()
+    model.l1.append(HasEqualContainers())
+    model.l2.append(HasEqualContainers())
+    self.assertEqual([model.l1, model.l2], model.layers)
+
   def testNotCheckpointable(self):
     class NotCheckpointable(object):
       pass
@@ -158,11 +172,62 @@ class ListTests(test.TestCase):
     self.assertEqual([v], l.trainable_weights)
     self.assertEqual([v2], l.non_trainable_weights)
 
+  def testListWrapperBasic(self):
+    # _ListWrapper, unlike List, compares like the built-in list type (since it
+    # is used to automatically replace lists).
+    a = tracking.Checkpointable()
+    b = tracking.Checkpointable()
+    self.assertEqual([a, a],
+                     [a, a])
+    self.assertEqual(data_structures._ListWrapper([a, a]),
+                     data_structures._ListWrapper([a, a]))
+    self.assertEqual([a, a],
+                     data_structures._ListWrapper([a, a]))
+    self.assertEqual(data_structures._ListWrapper([a, a]),
+                     [a, a])
+    self.assertNotEqual([a, a],
+                        [b, a])
+    self.assertNotEqual(data_structures._ListWrapper([a, a]),
+                        data_structures._ListWrapper([b, a]))
+    self.assertNotEqual([a, a],
+                        data_structures._ListWrapper([b, a]))
+    self.assertLess([a], [a, b])
+    self.assertLess(data_structures._ListWrapper([a]),
+                    data_structures._ListWrapper([a, b]))
+    self.assertLessEqual([a], [a, b])
+    self.assertLessEqual(data_structures._ListWrapper([a]),
+                         data_structures._ListWrapper([a, b]))
+    self.assertGreater([a, b], [a])
+    self.assertGreater(data_structures._ListWrapper([a, b]),
+                       data_structures._ListWrapper([a]))
+    self.assertGreaterEqual([a, b], [a])
+    self.assertGreaterEqual(data_structures._ListWrapper([a, b]),
+                            data_structures._ListWrapper([a]))
+    self.assertEqual([a], data_structures._ListWrapper([a]))
+    self.assertEqual([a], list(data_structures.List([a])))
+    self.assertEqual([a, a], data_structures._ListWrapper([a]) + [a])
+    self.assertEqual([a, a], [a] + data_structures._ListWrapper([a]))
+    self.assertIsInstance(data_structures._ListWrapper([a]), list)
+
+  def testWrapperChangesList(self):
+    l = []
+    l_wrapper = data_structures._ListWrapper(l)
+    l_wrapper.append(1)
+    self.assertEqual([1], l)
+
+  def testListChangesWrapper(self):
+    l = []
+    l_wrapper = data_structures._ListWrapper(l)
+    l.append(1)
+    self.assertEqual([1], l_wrapper)
+
   def testHashing(self):
     has_sequences = set([data_structures.List(),
                          data_structures.List()])
     self.assertEqual(2, len(has_sequences))
     self.assertNotIn(data_structures.List(), has_sequences)
+    with self.assertRaises(TypeError):
+      has_sequences.add(data_structures._ListWrapper([]))
 
 
 class HasMapping(training.Model):
diff --git a/tensorflow/python/training/checkpointable/layer_utils.py b/tensorflow/python/training/checkpointable/layer_utils.py
new file mode 100644
index 0000000000000000000000000000000000000000..978fcb2252cd4481b8286bdf3afd58b30ce6d665
--- /dev/null
+++ b/tensorflow/python/training/checkpointable/layer_utils.py
@@ -0,0 +1,93 @@
+# Copyright 2018 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""Utilities related to layer/model functionality."""
+
+# TODO(b/110718070): Move these functions back to tensorflow/python/keras/utils
+# once __init__ files no longer require all of tf.keras to be imported together.
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+
+def is_layer(obj):
+  """Implicit check for Layer-like objects."""
+  # TODO(b/110718070): Replace with isinstance(obj, base_layer.Layer).
+  return (hasattr(obj, "call")
+          and hasattr(obj, "build")
+          and hasattr(obj, "variables"))
+
+
+def filter_empty_layer_containers(layer_list):
+  """Filter out empty Layer-like containers."""
+  return [layer for layer in layer_list
+          # Filter out only empty Checkpointable data structures. Empty Networks
+          # will still show up in Model.layers.
+          if is_layer(layer) or getattr(layer, "layers", True)]
+
+
+def gather_trainable_weights(trainable, sub_layers, extra_variables):
+  """Lists the trainable weights for an object with sub-layers.
+
+  Args:
+    trainable: Whether the object collecting the variables is trainable.
+    sub_layers: A flat list of Layer objects owned by this object, to collect
+      variables from.
+    extra_variables: Any extra variables to include. Their `.trainable` property
+      is used to categorize them.
+
+  Returns:
+    A list of collected trainable weights/variables.
+  """
+  if not trainable:
+    return []
+  weights = []
+  for layer in sub_layers:
+    weights += layer.trainable_weights
+  trainable_extra_variables = [
+      v for v in extra_variables if v.trainable]
+  return weights + trainable_extra_variables
+
+
+def gather_non_trainable_weights(trainable, sub_layers, extra_variables):
+  """Lists the non-trainable weights for an object with sub-layers.
+
+  Args:
+    trainable: Whether the object collecting the variables is trainable.
+    sub_layers: A flat list of Layer objects owned by this object, to collect
+      variables from.
+    extra_variables: Any extra variables to include. Their `.trainable` property
+      is used to categorize them.
+
+  Returns:
+    A list of collected non-trainable weights/variables.
+  """
+  trainable_extra_variables = []
+  non_trainable_extra_variables = []
+  for v in extra_variables:
+    if v.trainable:
+      trainable_extra_variables.append(v)
+    else:
+      non_trainable_extra_variables.append(v)
+  weights = []
+  for layer in sub_layers:
+    weights += layer.non_trainable_weights
+  if not trainable:
+    trainable_weights = []
+    for layer in sub_layers:
+      trainable_weights += layer.trainable_weights
+    return (trainable_weights + trainable_extra_variables
+            + weights + non_trainable_extra_variables)
+  return weights + non_trainable_extra_variables
diff --git a/tensorflow/python/training/checkpointable/tracking.py b/tensorflow/python/training/checkpointable/tracking.py
index 00e14ac982358781b379a78d94da05343f88502b..bd0bed9d46f2e75633e3bf1230eded3708ec1c8b 100644
--- a/tensorflow/python/training/checkpointable/tracking.py
+++ b/tensorflow/python/training/checkpointable/tracking.py
@@ -18,31 +18,7 @@ from __future__ import division
 from __future__ import print_function
 
 from tensorflow.python.training.checkpointable import base
-
-
-class NoDependency(object):
-  """Allows attribute assignment to `Checkpointable` objects with no dependency.
-
-  Example usage:
-  ```python
-  obj = Checkpointable()
-  obj.has_dependency = tf.Variable(0., name="dep")
-  obj.no_dependency = NoDependency(tf.Variable(1., name="nodep"))
-  assert obj.no_dependency.name == "nodep:0"
-  ```
-
-  `obj` in this example has a dependency on the variable "dep", and both
-  attributes contain un-wrapped `Variable` objects.
-
-  `NoDependency` also works with `tf.keras.Model`, but only for checkpoint
-  dependencies: wrapping a `Layer` in `NoDependency` will assign the (unwrapped)
-  `Layer` to the attribute without a checkpoint dependency, but the `Model` will
-  still track the `Layer` (so it will appear in `Model.layers`, and its
-  variables will appear in `Model.variables`).
-  """
-
-  def __init__(self, value):
-    self.value = value
+from tensorflow.python.training.checkpointable import data_structures
 
 
 class NotCheckpointable(object):
@@ -86,18 +62,11 @@ class Checkpointable(base.CheckpointableBase):
 
   def __setattr__(self, name, value):
     """Support self.foo = checkpointable syntax."""
-    # Perform the attribute assignment, and potentially call other __setattr__
-    # overrides such as that for tf.keras.Model.
-    no_dependency = isinstance(value, NoDependency)
-    if no_dependency:
-      value = value.value
+    if getattr(self, "_setattr_tracking", True):
+      value = data_structures.sticky_attribute_assignment(
+          checkpointable=self, value=value, name=name)
     super(Checkpointable, self).__setattr__(name, value)
-    if not no_dependency and isinstance(value, base.CheckpointableBase):
-      self._track_checkpointable(
-          value, name=name,
-          # Allow the user to switch the Checkpointable which is tracked by this
-          # name, since assigning a new variable to an attribute has
-          # historically been fine (e.g. Adam did this).
-          # TODO(allenl): Should this be a warning once Checkpointable save/load
-          # is usable?
-          overwrite=True)
+
+  def _no_dependency(self, value):
+    """Override to allow CheckpointableBase to disable dependency tracking."""
+    return data_structures.NoDependency(value)
diff --git a/tensorflow/python/training/checkpointable/tracking_test.py b/tensorflow/python/training/checkpointable/tracking_test.py
index baf6f57efbc5c71ac3cb0d6b0a3d8f8b115fad1e..96da0d6e4720b44815de137c0efdd74645bae0fc 100644
--- a/tensorflow/python/training/checkpointable/tracking_test.py
+++ b/tensorflow/python/training/checkpointable/tracking_test.py
@@ -16,8 +16,19 @@ from __future__ import absolute_import
 from __future__ import division
 from __future__ import print_function
 
+import os
+
+import numpy
+
+from tensorflow.python.framework import test_util
+from tensorflow.python.keras.engine import training
+from tensorflow.python.ops import array_ops
 from tensorflow.python.platform import test
+from tensorflow.python.training.checkpointable import base
+from tensorflow.python.training.checkpointable import data_structures
 from tensorflow.python.training.checkpointable import tracking
+from tensorflow.python.training.checkpointable import util
+from tensorflow.python.util import nest
 
 
 class InterfaceTests(test.TestCase):
@@ -27,23 +38,134 @@ class InterfaceTests(test.TestCase):
     root.leaf = tracking.Checkpointable()
     root.leaf = root.leaf
     duplicate_name_dep = tracking.Checkpointable()
-    with self.assertRaises(ValueError):
+    with self.assertRaisesRegexp(ValueError, "already declared"):
       root._track_checkpointable(duplicate_name_dep, name="leaf")
     # No error; we're overriding __setattr__, so we can't really stop people
     # from doing this while maintaining backward compatibility.
     root.leaf = duplicate_name_dep
     root._track_checkpointable(duplicate_name_dep, name="leaf", overwrite=True)
+    self.assertIs(duplicate_name_dep, root._lookup_dependency("leaf"))
+    (_, dep_object), = root._checkpoint_dependencies
+    self.assertIs(duplicate_name_dep, dep_object)
 
   def testNoDependency(self):
     root = tracking.Checkpointable()
     hasdep = tracking.Checkpointable()
     root.hasdep = hasdep
     nodep = tracking.Checkpointable()
-    root.nodep = tracking.NoDependency(nodep)
+    root.nodep = data_structures.NoDependency(nodep)
     self.assertEqual(1, len(root._checkpoint_dependencies))
     self.assertIs(root._checkpoint_dependencies[0].ref, root.hasdep)
     self.assertIs(root.hasdep, hasdep)
     self.assertIs(root.nodep, nodep)
 
+    class NoDependencyModel(training.Model):
+
+      @base.no_automatic_dependency_tracking
+      def __init__(self):
+        super(NoDependencyModel, self).__init__()
+        self.a = []
+        self.b = tracking.Checkpointable()
+
+    nodeps = NoDependencyModel()
+    self.assertEqual([nodeps], util.list_objects(nodeps))
+
+  def testListBasic(self):
+    a = tracking.Checkpointable()
+    b = tracking.Checkpointable()
+    a.l = [b]
+    c = tracking.Checkpointable()
+    a.l.append(c)
+    a_deps = util.list_objects(a)
+    self.assertIn(b, a_deps)
+    self.assertIn(c, a_deps)
+    direct_a_dep, = a._checkpoint_dependencies
+    self.assertEqual("l", direct_a_dep.name)
+    self.assertIn(b, direct_a_dep.ref)
+    self.assertIn(c, direct_a_dep.ref)
+
+  @test_util.run_in_graph_and_eager_modes
+  def testMutationDirtiesList(self):
+    a = tracking.Checkpointable()
+    b = tracking.Checkpointable()
+    a.l = [b]
+    c = tracking.Checkpointable()
+    a.l.insert(0, c)
+    checkpoint = util.Checkpoint(a=a)
+    with self.assertRaisesRegexp(ValueError, "A list element was replaced"):
+      checkpoint.save(os.path.join(self.get_temp_dir(), "ckpt"))
+
+  @test_util.run_in_graph_and_eager_modes
+  def testOutOfBandEditDirtiesList(self):
+    a = tracking.Checkpointable()
+    b = tracking.Checkpointable()
+    held_reference = [b]
+    a.l = held_reference
+    c = tracking.Checkpointable()
+    held_reference.append(c)
+    checkpoint = util.Checkpoint(a=a)
+    with self.assertRaisesRegexp(ValueError, "The wrapped list was modified"):
+      checkpoint.save(os.path.join(self.get_temp_dir(), "ckpt"))
+
+  @test_util.run_in_graph_and_eager_modes
+  def testNestedLists(self):
+    a = tracking.Checkpointable()
+    a.l = []
+    b = tracking.Checkpointable()
+    a.l.append([b])
+    c = tracking.Checkpointable()
+    a.l[0].append(c)
+    a_deps = util.list_objects(a)
+    self.assertIn(b, a_deps)
+    self.assertIn(c, a_deps)
+    a.l[0].append(1)
+    d = tracking.Checkpointable()
+    a.l[0].append(d)
+    a_deps = util.list_objects(a)
+    self.assertIn(d, a_deps)
+    self.assertIn(b, a_deps)
+    self.assertIn(c, a_deps)
+    self.assertNotIn(1, a_deps)
+    e = tracking.Checkpointable()
+    f = tracking.Checkpointable()
+    a.l1 = [[], [e]]
+    a.l1[0].append(f)
+    a_deps = util.list_objects(a)
+    self.assertIn(e, a_deps)
+    self.assertIn(f, a_deps)
+    checkpoint = util.Checkpoint(a=a)
+    checkpoint.save(os.path.join(self.get_temp_dir(), "ckpt"))
+    a.l[0].append(data_structures.NoDependency([]))
+    a.l[0][-1].append(5)
+    checkpoint.save(os.path.join(self.get_temp_dir(), "ckpt"))
+    # Dirtying the inner list means the root object is unsaveable.
+    a.l[0][1] = 2
+    with self.assertRaisesRegexp(ValueError, "A list element was replaced"):
+      checkpoint.save(os.path.join(self.get_temp_dir(), "ckpt"))
+
+  @test_util.run_in_graph_and_eager_modes
+  def testNoDepList(self):
+    a = training.Model()
+    a.l1 = data_structures.NoDependency([])
+    a.l1.insert(1, 0)
+    self.assertTrue(isinstance(a.l1, list))
+    checkpoint = util.Checkpoint(a=a)
+    checkpoint.save(os.path.join(self.get_temp_dir(), "ckpt"))
+    a.l2 = []
+    a.l2.insert(1, 0)
+    with self.assertRaisesRegexp(ValueError, "A list element was replaced"):
+      checkpoint.save(os.path.join(self.get_temp_dir(), "ckpt"))
+
+  @test_util.run_in_graph_and_eager_modes
+  def testAssertions(self):
+    a = tracking.Checkpointable()
+    a.l = [numpy.zeros([2, 2])]
+    self.assertAllEqual([numpy.zeros([2, 2])], a.l)
+    self.assertAllClose([numpy.zeros([2, 2])], a.l)
+    nest.map_structure(self.assertAllClose, a.l, [numpy.zeros([2, 2])])
+    a.tensors = [array_ops.ones([2, 2]), array_ops.zeros([3, 3])]
+    self.assertAllClose([numpy.ones([2, 2]), numpy.zeros([3, 3])],
+                        self.evaluate(a.tensors))
+
 if __name__ == "__main__":
   test.main()
diff --git a/tensorflow/python/training/checkpointable/util.py b/tensorflow/python/training/checkpointable/util.py
index e0f61137b1026a64a8cc9703ac33997c55f93a4f..6ae5765b133cc72b67f3d9864d0f67abf33f0648 100644
--- a/tensorflow/python/training/checkpointable/util.py
+++ b/tensorflow/python/training/checkpointable/util.py
@@ -40,6 +40,7 @@ from tensorflow.python.training import optimizer as optimizer_lib
 from tensorflow.python.training import saveable_object as saveable_object_lib
 from tensorflow.python.training import saver as saver_lib
 from tensorflow.python.training.checkpointable import base
+from tensorflow.python.training.checkpointable import data_structures
 from tensorflow.python.training.checkpointable import tracking
 from tensorflow.python.util import deprecation
 from tensorflow.python.util import tf_contextlib
@@ -93,7 +94,7 @@ class _CheckpointRestoreCoordinator(object):
     # use them (for example because of inconsistent references when
     # loading). Used to make status assertions fail when loading checkpoints
     # that don't quite match.
-    self.all_python_objects = weakref.WeakSet()
+    self.all_python_objects = _ObjectIdentityWeakSet()
     self.save_path = save_path
     self.dtype_map = dtype_map
     # When graph building, contains a list of ops to run to restore objects from
@@ -272,11 +273,129 @@ def object_metadata(save_path):
   return object_graph_proto
 
 
+class _ObjectIdentityWrapper(object):
+  """Wraps an object, mapping __eq__ on wrapper to "is" on wrapped.
+
+  Since __eq__ is based on object identity, it's safe to also define __hash__
+  based on object ids. This lets us add unhashable types like checkpointable
+  _ListWrapper objects to object-identity collections.
+  """
+
+  def __init__(self, wrapped):
+    self._wrapped = wrapped
+
+  @property
+  def unwrapped(self):
+    return self._wrapped
+
+  def __eq__(self, other):
+    if isinstance(other, _ObjectIdentityWrapper):
+      return self._wrapped is other._wrapped  # pylint: disable=protected-access
+    return self._wrapped is other
+
+  def __hash__(self):
+    # Wrapper id() is also fine for weakrefs. In fact, we rely on
+    # id(weakref.ref(a)) == id(weakref.ref(a)) and weakref.ref(a) is
+    # weakref.ref(a) in _WeakObjectIdentityWrapper.
+    return id(self._wrapped)
+
+
+class _WeakObjectIdentityWrapper(_ObjectIdentityWrapper):
+
+  def __init__(self, wrapped):
+    super(_WeakObjectIdentityWrapper, self).__init__(weakref.ref(wrapped))
+
+  @property
+  def unwrapped(self):
+    return self._wrapped()
+
+
+class _ObjectIdentityDictionary(collections.MutableMapping):
+  """A mutable mapping data structure which compares using "is".
+
+  This is necessary because we have checkpointable objects (_ListWrapper) which
+  have behavior identical to built-in Python lists (including being unhashable
+  and comparing based on the equality of their contents by default).
+  """
+
+  def __init__(self):
+    self._storage = {}
+
+  def _wrap_key(self, key):
+    return _ObjectIdentityWrapper(key)
+
+  def __getitem__(self, key):
+    return self._storage[self._wrap_key(key)]
+
+  def __setitem__(self, key, value):
+    self._storage[self._wrap_key(key)] = value
+
+  def __delitem__(self, key):
+    del self._storage[self._wrap_key(key)]
+
+  def __len__(self):
+    return len(self._storage)
+
+  def __iter__(self):
+    for key in self._storage:
+      yield key.unwrapped
+
+
+class _ObjectIdentityWeakKeyDictionary(_ObjectIdentityDictionary):
+  """Like weakref.WeakKeyDictionary, but compares objects with "is"."""
+
+  def _wrap_key(self, key):
+    return _WeakObjectIdentityWrapper(key)
+
+  def __len__(self):
+    # Iterate, discarding old weak refs
+    return len(list(self._storage))
+
+  def __iter__(self):
+    keys = self._storage.keys()
+    for key in keys:
+      unwrapped = key.unwrapped
+      if unwrapped is None:
+        del self[key]
+      else:
+        yield unwrapped
+
+
+class _ObjectIdentityWeakSet(collections.MutableSet):
+  """Like weakref.WeakSet, but compares objects with "is"."""
+
+  def __init__(self):
+    self._storage = set()
+
+  def __contains__(self, key):
+    return _WeakObjectIdentityWrapper(key) in self._storage
+
+  def discard(self, key):
+    self._storage.discard(_WeakObjectIdentityWrapper(key))
+
+  def add(self, key):
+    self._storage.add(_WeakObjectIdentityWrapper(key))
+
+  def __len__(self):
+    # Iterate, discarding old weak refs
+    return len(list(self))
+
+  def __iter__(self):
+    keys = list(self._storage)
+    for key in keys:
+      unwrapped = key.unwrapped
+      if unwrapped is None:
+        self.discard(key)
+      else:
+        yield unwrapped
+
+
 def _breadth_first_checkpointable_traversal(root_checkpointable):
   """Find shortest paths to all variables owned by dependencies of root."""
   bfs_sorted = []
   to_visit = collections.deque([root_checkpointable])
-  path_to_root = {root_checkpointable: ()}
+  path_to_root = _ObjectIdentityDictionary()
+  path_to_root[root_checkpointable] = ()
   while to_visit:
     current_checkpointable = to_visit.popleft()
     if isinstance(current_checkpointable, tracking.NotCheckpointable):
@@ -337,7 +456,7 @@ def _slot_variable_naming_for_optimizer(optimizer_path):
 def _serialize_slot_variables(checkpointable_objects, node_ids, object_names):
   """Gather and name slot variables."""
   non_slot_objects = list(checkpointable_objects)
-  slot_variables = {}
+  slot_variables = _ObjectIdentityDictionary()
   for checkpointable in non_slot_objects:
     if isinstance(checkpointable, optimizer_lib.Optimizer):
       naming_scheme = _slot_variable_naming_for_optimizer(
@@ -500,11 +619,12 @@ def _serialize_object_graph(root_checkpointable, saveables_cache):
   """
   checkpointable_objects, path_to_root = (
       _breadth_first_checkpointable_traversal(root_checkpointable))
-  object_names = {
-      obj: _object_prefix_from_path(path)
-      for obj, path in path_to_root.items()}
-  node_ids = {node: node_id for node_id, node
-              in enumerate(checkpointable_objects)}
+  object_names = _ObjectIdentityDictionary()
+  for obj, path in path_to_root.items():
+    object_names[obj] = _object_prefix_from_path(path)
+  node_ids = _ObjectIdentityDictionary()
+  for node_id, node in enumerate(checkpointable_objects):
+    node_ids[node] = node_id
   slot_variables = _serialize_slot_variables(
       checkpointable_objects=checkpointable_objects,
       node_ids=node_ids,
@@ -535,11 +655,12 @@ def list_objects(root_checkpointable):
   # to run.
   checkpointable_objects, path_to_root = (
       _breadth_first_checkpointable_traversal(root_checkpointable))
-  object_names = {
-      obj: _object_prefix_from_path(path)
-      for obj, path in path_to_root.items()}
-  node_ids = {node: node_id for node_id, node
-              in enumerate(checkpointable_objects)}
+  object_names = _ObjectIdentityDictionary()
+  for obj, path in path_to_root.items():
+    object_names[obj] = _object_prefix_from_path(path)
+  node_ids = _ObjectIdentityDictionary()
+  for node_id, node in enumerate(checkpointable_objects):
+    node_ids[node] = node_id
   _serialize_slot_variables(
       checkpointable_objects=checkpointable_objects,
       node_ids=node_ids,
@@ -988,7 +1109,7 @@ class CheckpointableSaver(object):
     else:
       # Maps Checkpointable objects -> attribute names -> SaveableObjects, to
       # avoid re-creating SaveableObjects when graph building.
-      self._saveable_object_cache = weakref.WeakKeyDictionary()
+      self._saveable_object_cache = _ObjectIdentityWeakKeyDictionary()
 
   @property
   def _root_checkpointable(self):
@@ -1310,7 +1431,7 @@ class Checkpoint(tracking.Checkpointable):
       with ops.device("/cpu:0"):
         # add_variable creates a dependency named "save_counter"; NoDependency
         # prevents creating a second dependency named "_save_counter".
-        self._save_counter = tracking.NoDependency(
+        self._save_counter = data_structures.NoDependency(
             add_variable(self, name="save_counter", initializer=0,
                          dtype=dtypes.int64))
 
diff --git a/tensorflow/python/training/checkpointable/util_test.py b/tensorflow/python/training/checkpointable/util_test.py
index 896ea47b974a334d34e520e6f3c2ad947dea12a2..3c1a4a6f83c20a74961bf3e1263b2a33d3e36f05 100644
--- a/tensorflow/python/training/checkpointable/util_test.py
+++ b/tensorflow/python/training/checkpointable/util_test.py
@@ -102,7 +102,7 @@ class InterfaceTests(test.TestCase):
         name="duplicate", initial_value=1.)
     duplicate = checkpointable_utils.add_variable(
         obj, name="duplicate", shape=[])
-    with self.assertRaisesRegexp(ValueError, "'duplicate' already exists"):
+    with self.assertRaisesRegexp(ValueError, "'duplicate'.*already declared"):
       checkpointable_utils.add_variable(obj, name="duplicate", shape=[])
 
     self.evaluate(checkpointable_utils.gather_initializers(obj))
diff --git a/tensorflow/python/training/distribute.py b/tensorflow/python/training/distribute.py
index 6a326b65bbe956953bd414c8e89fd9f5cce58f48..c719045c7f8cf3ba7b1a9c0bdb1f610ba8091464 100644
--- a/tensorflow/python/training/distribute.py
+++ b/tensorflow/python/training/distribute.py
@@ -221,11 +221,11 @@ def has_distribution_strategy():
 
 
 def get_loss_reduction():
-  """Reduce `method_string` corresponding to the last loss reduction."""
+  """Reduce `aggregation` corresponding to the last loss reduction."""
   loss_reduction = ops.get_default_graph()._last_loss_reduction  # pylint: disable=protected-access
   if loss_reduction == losses_impl.Reduction.SUM:
-    return "sum"
-  return "mean"
+    return variable_scope.VariableAggregation.SUM
+  return variable_scope.VariableAggregation.MEAN
 
 
 # ------------------------------------------------------------------------------
@@ -539,8 +539,8 @@ class DistributionStrategy(object):
   1. Wrap your input dataset in `d.distribute_dataset()` and create an iterator.
   2. Define each tower `d.call_for_each_tower()` up to the point of
      getting a list of gradient, variable pairs.
-  3. Call `d.reduce("sum", t, v)` or `d.batch_reduce()` to sum the
-     gradients (with locality T) into values with locality V(`v`).
+  3. Call `d.reduce(VariableAggregation.SUM, t, v)` or `d.batch_reduce()` to sum
+     the gradients (with locality T) into values with locality V(`v`).
   4. Call `d.update(v)` for each variable to update its value.
 
   Steps 3 and 4 are done automatically by class `Optimizer` if you call
@@ -614,43 +614,6 @@ class DistributionStrategy(object):
     # Note: should support "colocate_with" argument.
     raise NotImplementedError("must be implemented in descendants")
 
-  def tower_local_var_scope(self, reduce_method):
-    """Inside this scope, new variables will not be mirrored.
-
-    There will still be one component variable per tower, but there is
-    no requirement that they stay in sync. Instead, when saving them
-    or calling `read_var()`, we use the value that results when
-    calling `reduce()` on all the towers' variables.
-
-    Note: tower-local implies not trainable. Instead, it is expected
-    that each tower will directly update (using `assign_add()` or
-    whatever) its local variable instance but only the aggregated
-    value (accessible using `read_var()`) will be exported from the
-    model. When it is acceptable to only aggregate on export, we
-    greatly reduce communication overhead by using tower-local
-    variables.
-
-    Note: All component variables will be initialized to the same
-    value, using the initialization expression from the first tower.
-    The values will match even if the initialization expression uses
-    random numbers.
-
-    Args:
-      reduce_method: String used as a `method_string` to `reduce()`
-        to get the value to save when checkpointing.
-
-    Returns:
-      A context manager.
-    """
-    def create_tower_local_variable(next_creator, *args, **kwargs):
-      _require_distribution_strategy_scope(self)
-      kwargs["use_resource"] = True
-      kwargs["tower_local_reduce_method"] = reduce_method
-      return next_creator(*args, **kwargs)
-
-    _require_distribution_strategy_scope(self)
-    return variable_scope.variable_creator_scope(create_tower_local_variable)
-
   def read_var(self, v):
     """Reads the value of a variable.
 
@@ -816,12 +779,12 @@ class DistributionStrategy(object):
   def _call_for_each_tower(self, fn, *args, **kwargs):
     raise NotImplementedError("must be implemented in descendants")
 
-  def reduce(self, method_string, value, destinations=None):
+  def reduce(self, aggregation, value, destinations=None):
     """Combine (via e.g. sum or mean) values across towers.
 
     Args:
-      method_string: A string indicating how to combine values, either
-        "sum" or "mean".
+      aggregation: Indicates how a variable will be aggregated. Accepted values
+        are @{tf.VariableAggregation.SUM}, @{tf.VariableAggregation.MEAN}.
       value: A per-device value with one value per tower.
       destinations: An optional mirrored variable, a device string,
         list of device strings. The return value will be copied to all
@@ -836,18 +799,21 @@ class DistributionStrategy(object):
     # TODO(josh11b): Return an unwrapped value if colocate_with is a
     # single device.
     _require_cross_tower_context(self)
-    assert method_string in ("sum", "mean")
-    return self._reduce(method_string, value, destinations)
+    assert aggregation in [
+        variable_scope.VariableAggregation.SUM,
+        variable_scope.VariableAggregation.MEAN
+    ]
+    return self._reduce(aggregation, value, destinations)
 
-  def _reduce(self, method_string, value, destinations):
+  def _reduce(self, aggregation, value, destinations):
     raise NotImplementedError("must be implemented in descendants")
 
-  def batch_reduce(self, method_string, value_destination_pairs):
+  def batch_reduce(self, aggregation, value_destination_pairs):
     """Combine multiple `reduce` calls into one for faster execution.
 
     Args:
-      method_string: A string indicating how to combine values, either
-        "sum" or "mean".
+      aggregation: Indicates how a variable will be aggregated. Accepted values
+        are @{tf.VariableAggregation.SUM}, @{tf.VariableAggregation.MEAN}.
       value_destination_pairs: A sequence of (value, destinations)
         pairs. See `reduce()` for a description.
 
@@ -856,12 +822,17 @@ class DistributionStrategy(object):
     """
     # TODO(josh11b): More docstring
     _require_cross_tower_context(self)
-    assert method_string in ("sum", "mean")
-    return self._batch_reduce(method_string, value_destination_pairs)
-
-  def _batch_reduce(self, method_string, value_destination_pairs):
-    return [self.reduce(method_string, t, destinations=v)
-            for t, v in value_destination_pairs]
+    assert aggregation in [
+        variable_scope.VariableAggregation.SUM,
+        variable_scope.VariableAggregation.MEAN
+    ]
+    return self._batch_reduce(aggregation, value_destination_pairs)
+
+  def _batch_reduce(self, aggregation, value_destination_pairs):
+    return [
+        self.reduce(aggregation, t, destinations=v)
+        for t, v in value_destination_pairs
+    ]
 
   def update(self, var, fn, *args, **kwargs):
     """Run `fn` to update `var` using inputs mirrored to the same devices.
@@ -1090,10 +1061,6 @@ class TowerContext(object):
     finally:
       _pop_per_thread_mode()
 
-  def tower_local_var_scope(self, reduce_method):
-    """Alias for distribution_strategy.tower_local_var_scope()."""
-    return self._distribution_strategy.tower_local_var_scope(reduce_method)
-
   @property
   def is_single_tower(self):
     """Returns whether there is a single tower or multiple."""
@@ -1140,22 +1107,11 @@ class _DefaultDistributionStrategy(DistributionStrategy):
 
     def creator(next_creator, *args, **kwargs):
       _require_distribution_strategy_scope(self)
-      kwargs.pop("tower_local_reduce_method", None)
       return next_creator(*args, **kwargs)
 
     return _CurrentDistributionContext(
         self, variable_scope.variable_creator_scope(creator))
 
-  def tower_local_var_scope(self, reduce_method):
-    """Does not set to resource variables."""
-    def create_tower_local_variable(next_creator, *args, **kwargs):
-      _require_distribution_strategy_scope(self)
-      kwargs["trainable"] = False
-      return next_creator(*args, **kwargs)
-
-    _require_distribution_strategy_scope(self)
-    return variable_scope.variable_creator_scope(create_tower_local_variable)
-
   def colocate_vars_with(self, colocate_with_variable):
     """Does not require `self.scope`."""
     _require_distribution_strategy_scope(self)
@@ -1176,9 +1132,9 @@ class _DefaultDistributionStrategy(DistributionStrategy):
     with TowerContext(self, tower_id=0):
       return fn(*args, **kwargs)
 
-  def _reduce(self, method_string, value, destinations):
+  def _reduce(self, aggregation, value, destinations):
     # TODO(josh11b): Use destinations?
-    del method_string, destinations
+    del aggregation, destinations
     return value
 
   def _update(self, var, fn, *args, **kwargs):
diff --git a/tensorflow/python/training/distribute_test.py b/tensorflow/python/training/distribute_test.py
index 0a4f19c31f6714e1211f9deed9703c02192cc2c0..694145ede73c1c9121cbc4c4e2d6f61e93165d09 100644
--- a/tensorflow/python/training/distribute_test.py
+++ b/tensorflow/python/training/distribute_test.py
@@ -29,6 +29,14 @@ class _TestTowerContext(distribute.TowerContext):
     return kwargs["test_arg"]
 
 
+def _get_test_variable(name, synchronization, aggregation):
+  return {
+      "name": name,
+      "synchronization": synchronization,
+      "aggregation": aggregation
+  }
+
+
 class _TestStrategy(distribute.DistributionStrategy):
 
   def _call_for_each_tower(self, fn, *args, **kwargs):
@@ -36,7 +44,8 @@ class _TestStrategy(distribute.DistributionStrategy):
       return fn(*args, **kwargs)
 
   def _create_variable(self, next_creator, *args, **kwargs):
-    return kwargs["name"]
+    return _get_test_variable(kwargs["name"], kwargs["synchronization"],
+                              kwargs["aggregation"])
 
 
 def _assert_in_default_state(t):
@@ -61,7 +70,11 @@ class TestStrategyTest(test.TestCase):
       self.assertTrue(distribute.has_distribution_strategy())
       self.assertIs(dist, distribute.get_distribution_strategy())
       self.assertEqual("foo", tower_context.merge_call(None, test_arg="foo"))
-      self.assertEqual("bar", variable_scope.variable(1.0, name="bar"))
+      expected_value = _get_test_variable(
+          "bar", variable_scope.VariableSynchronization.AUTO,
+          variable_scope.VariableAggregation.NONE)
+      self.assertDictEqual(expected_value,
+                           variable_scope.variable(1.0, name="bar"))
 
     with self.assertRaises(RuntimeError):
       dist.call_for_each_tower(run_fn)
@@ -77,7 +90,27 @@ class TestStrategyTest(test.TestCase):
       self.assertIs(dist, distribute.get_cross_tower_context())
       self.assertTrue(distribute.has_distribution_strategy())
       self.assertIs(dist, distribute.get_distribution_strategy())
-      self.assertEqual("baz", variable_scope.variable(1.0, name="baz"))
+      expected_value = _get_test_variable(
+          "baz", variable_scope.VariableSynchronization.AUTO,
+          variable_scope.VariableAggregation.NONE)
+      self.assertDictEqual(expected_value,
+                           variable_scope.variable(1.0, name="baz"))
+    _assert_in_default_state(self)
+
+  def testSettingSynchronizationAndAggregation(self):
+    _assert_in_default_state(self)
+    dist = _TestStrategy()
+    with dist.scope():
+      expected_value = _get_test_variable(
+          "baz", variable_scope.VariableSynchronization.ON_WRITE,
+          variable_scope.VariableAggregation.MEAN)
+      self.assertDictEqual(
+          expected_value,
+          variable_scope.variable(
+              1.0,
+              name="baz",
+              synchronization=variable_scope.VariableSynchronization.ON_WRITE,
+              aggregation=variable_scope.VariableAggregation.MEAN))
     _assert_in_default_state(self)
 
 
diff --git a/tensorflow/python/training/optimizer.py b/tensorflow/python/training/optimizer.py
index fe9ffde11ce47e1c2ae6c96e59cc2bf0d43d9707..f75db080595c6f348fe7e9302041bf19f72a301f 100644
--- a/tensorflow/python/training/optimizer.py
+++ b/tensorflow/python/training/optimizer.py
@@ -77,9 +77,10 @@ def _deduplicate_indexed_slices(values, indices):
 
 
 def _var_key(var):
-  if context.executing_eagerly():
-    return var._unique_id  # pylint: disable=protected-access
-  return (var.op.graph, var.op.name)
+  # TODO(ashankar): Consolidate handling for eager and graph
+  if hasattr(var, "op"):
+    return (var.op.graph, var.op.name)
+  return var._unique_id  # pylint: disable=protected-access
 
 
 class _OptimizableVariable(object):
@@ -461,7 +462,8 @@ class Optimizer(
         # Have to be careful to call distribute_lib.get_loss_reduction()
         # *after* loss() is evaluated, so we know what loss reduction it uses.
         # TODO(josh11b): Test that we handle weight decay in a reasonable way.
-        if distribute_lib.get_loss_reduction() == "mean":
+        if (distribute_lib.get_loss_reduction() ==
+            variable_scope.VariableAggregation.MEAN):
           num_towers = distribute_lib.get_distribution_strategy().num_towers
           if num_towers > 1:
             loss_value *= (1. / num_towers)
@@ -478,7 +480,8 @@ class Optimizer(
           "be a function when eager execution is enabled.")
 
     # Scale loss if using a "mean" loss reduction and multiple towers.
-    if distribute_lib.get_loss_reduction() == "mean":
+    if (distribute_lib.get_loss_reduction() ==
+        variable_scope.VariableAggregation.MEAN):
       num_towers = distribute_lib.get_distribution_strategy().num_towers
       if num_towers > 1:
         loss *= (1. / num_towers)
@@ -649,7 +652,8 @@ class Optimizer(
       towers. If `global_step` was not None, that operation also
       increments `global_step`.
     """
-    reduced_grads = distribution.batch_reduce("sum", grads_and_vars)
+    reduced_grads = distribution.batch_reduce(
+        variable_scope.VariableAggregation.SUM, grads_and_vars)
     var_list = [v for _, v in grads_and_vars]
     grads_and_vars = zip(reduced_grads, var_list)
     # Note that this is called in a cross-tower context.
diff --git a/tensorflow/python/training/quantize_training.i b/tensorflow/python/training/quantize_training.i
index fb5e47efa0259d02df3ccf2e9b1430e027f8fcfb..54d6789616473382cf87abe4f701092bbd4e272f 100644
--- a/tensorflow/python/training/quantize_training.i
+++ b/tensorflow/python/training/quantize_training.i
@@ -73,6 +73,8 @@ def do_quantize_training_on_graphdef(input_graph, num_bits):
 
 do_quantize_training_on_graphdef._tf_api_names = [
     'train.do_quantize_training_on_graphdef']
+do_quantize_training_on_graphdef._tf_api_names_v1 = [
+    'train.do_quantize_training_on_graphdef']
 %}
 
 %unignoreall
diff --git a/tensorflow/python/training/saver.py b/tensorflow/python/training/saver.py
index 53ed89e4ab8dca876e232209928e61ba9628eb46..1ee975fbe48e8ba724d8f40040b122c5c02aa352 100644
--- a/tensorflow/python/training/saver.py
+++ b/tensorflow/python/training/saver.py
@@ -22,7 +22,6 @@ from __future__ import print_function
 import collections
 import os.path
 import re
-import sys
 import time
 import uuid
 
@@ -1043,8 +1042,8 @@ def get_checkpoint_state(checkpoint_dir, latest_filename=None):
       ckpt = CheckpointState()
       text_format.Merge(file_content, ckpt)
       if not ckpt.model_checkpoint_path:
-        raise ValueError("Invalid checkpoint state loaded from %s",
-                         checkpoint_dir)
+        raise ValueError("Invalid checkpoint state loaded from "
+                         + checkpoint_dir)
       # For relative model_checkpoint_path and all_model_checkpoint_paths,
       # prepend checkpoint_dir.
       if not os.path.isabs(ckpt.model_checkpoint_path):
@@ -1706,12 +1705,17 @@ class Saver(object):
       save_path: Path where parameters were previously saved.
 
     Raises:
-      ValueError: If save_path is None.
+      ValueError: If save_path is None or not a valid checkpoint.
     """
     if self._is_empty:
       return
     if save_path is None:
       raise ValueError("Can't load save_path when it is None.")
+
+    if not checkpoint_exists(compat.as_text(save_path)):
+      raise ValueError("The passed save_path is not a valid checkpoint: "
+                       + compat.as_text(save_path))
+
     logging.info("Restoring parameters from %s", compat.as_text(save_path))
     try:
       if context.executing_eagerly():
@@ -1719,23 +1723,24 @@ class Saver(object):
       else:
         sess.run(self.saver_def.restore_op_name,
                  {self.saver_def.filename_tensor_name: save_path})
-    except errors.NotFoundError:
-      exception_type, exception_value, exception_traceback = sys.exc_info()
-      # The checkpoint would not be loaded successfully as is. Try to parse it
-      # as an object-based checkpoint.
-      should_reraise = False
+    except errors.NotFoundError as err:
+      # There are three common conditions that might cause this error:
+      # 0. The file is missing. We ignore here, as this is checked above.
+      # 1. This is an object-based checkpoint trying name-based loading.
+      # 2. The graph has been altered and a variable or other name is missing.
+
+      # 1. The checkpoint would not be loaded successfully as is. Try to parse
+      # it as an object-based checkpoint.
       try:
         reader = pywrap_tensorflow.NewCheckpointReader(save_path)
         object_graph_string = reader.get_tensor(
             checkpointable.OBJECT_GRAPH_PROTO_KEY)
       except errors.NotFoundError:
-        # This is not an object-based checkpoint, or the checkpoint doesn't
-        # exist. Re-raise the original exception, but do it outside the except
-        # block so the object graph lookup isn't included in the stack trace.
-        should_reraise = True
-      if should_reraise:
-        six.reraise(exception_type, exception_value, exception_traceback)
-      del exception_traceback  # avoid reference cycles
+        # 2. This is not an object-based checkpoint, which likely means there
+        # is a graph mismatch. Re-raise the original error with
+        # a helpful message (b/110263146)
+        raise _wrap_restore_error_with_msg(
+            err, "a Variable name or other graph key that is missing")
 
       # This is an object-based checkpoint. We'll print a warning and then do
       # the restore.
@@ -1747,6 +1752,11 @@ class Saver(object):
       self._restore_from_object_based_checkpoint(
           sess=sess, save_path=save_path,
           object_graph_string=object_graph_string)
+    except errors.InvalidArgumentError as err:
+      # There is a mismatch between the graph and the checkpoint being loaded.
+      # We add a more reasonable error message here to help users (b/110263146)
+      raise _wrap_restore_error_with_msg(
+          err, "a mismatch between the current graph and the graph")
 
   def _restore_from_object_based_checkpoint(self, sess, save_path,
                                             object_graph_string):
@@ -2139,6 +2149,14 @@ def _meta_graph_filename(checkpoint_filename, meta_graph_suffix="meta"):
   return meta_graph_filename
 
 
+def _wrap_restore_error_with_msg(err, extra_verbiage):
+  err_msg = ("Restoring from checkpoint failed. This is most likely "
+             "due to {} from the checkpoint. Please ensure that you "
+             "have not altered the graph expected based on the checkpoint. "
+             "Original error:\n\n{}").format(extra_verbiage, err.message)
+  return err.__class__(err.node_def, err.op, err_msg)
+
+
 ops.register_proto_function(
     ops.GraphKeys.SAVERS,
     proto_type=saver_pb2.SaverDef,
diff --git a/tensorflow/python/training/saver_test.py b/tensorflow/python/training/saver_test.py
index f235300eb5c23cfa7495ab461b512206959a778c..ae9c244aaf372dcbcf365cf3e6a21ae77d9ae7d0 100644
--- a/tensorflow/python/training/saver_test.py
+++ b/tensorflow/python/training/saver_test.py
@@ -24,10 +24,8 @@ import math
 import os
 import random
 import shutil
-import sys
 import tempfile
 import time
-import traceback
 
 import numpy as np
 import six
@@ -369,8 +367,8 @@ class SaverTest(test.TestCase):
     for ver in (saver_pb2.SaverDef.V1, saver_pb2.SaverDef.V2):
       with self.test_session() as sess:
         save = saver_module.Saver({"v0": v0}, write_version=ver)
-        with self.assertRaisesRegexp(errors.NotFoundError,
-                                     "Failed to find any matching files for"):
+        with self.assertRaisesRegexp(
+            ValueError, "The passed save_path is not a valid checkpoint:"):
           save.restore(sess, "invalid path")
 
   def testInt64(self):
@@ -3139,27 +3137,33 @@ class CheckpointableCompatibilityTests(test.TestCase):
           errors.NotFoundError, "Key b not found in checkpoint"):
         b_saver.restore(sess=sess, save_path=save_path)
 
-  def testCheckpointNotFoundErrorRaised(self):
-    # Restore does some tricky exception handling to figure out if it should
-    # load an object-based checkpoint. Tests that the exception handling isn't
-    # too broad.
-    a = resource_variable_ops.ResourceVariable(1., name="a")
-    saver = saver_module.Saver([a])
-    with self.test_session() as sess:
-      with self.assertRaisesRegexp(
-          errors.NotFoundError,
-          "Failed to find any matching files for path_which_does_not_exist"):
-        saver.restore(sess=sess, save_path="path_which_does_not_exist")
-      try:
-        saver.restore(sess=sess, save_path="path_which_does_not_exist")
-      except errors.NotFoundError:
-        # Make sure we don't have a confusing "During handling of the above
-        # exception" block in Python 3.
-        # pylint: disable=no-value-for-parameter
-        exception_string = "\n".join(
-            traceback.format_exception(*sys.exc_info()))
-        # pylint: enable=no-value-for-parameter
-        self.assertNotIn("NewCheckpointReader", exception_string)
+      with self.assertRaises(errors.NotFoundError) as cs:
+        b_saver.restore(sess=sess, save_path=save_path)
+
+      # Make sure we don't have a confusing "During handling of the above
+      # exception" block in Python 3.
+      self.assertNotIn("NewCheckpointReader", cs.exception.message)
+
+  def testGraphChangedForRestoreErrorRaised(self):
+    checkpoint_directory = self.get_temp_dir()
+    checkpoint_prefix = os.path.join(checkpoint_directory, "ckpt")
+
+    with ops_lib.Graph().as_default() as g:
+      a = variables.Variable(1., name="a")
+      a_saver = saver_module.Saver([a])
+
+      with self.test_session(graph=g) as sess:
+        sess.run(a.initializer)
+        save_path = a_saver.save(sess=sess, save_path=checkpoint_prefix)
+
+    with ops_lib.Graph().as_default() as g:
+      a = variables.Variable([1.], name="a")
+      a_saver = saver_module.Saver([a])
+      with self.test_session(graph=g) as sess:
+        with self.assertRaisesRegexp(
+            errors.InvalidArgumentError,
+            "a mismatch between the current graph and the graph"):
+          a_saver.restore(sess=sess, save_path=save_path)
 
   def testLoadFromObjectBasedGraph(self):
     checkpoint_directory = self.get_temp_dir()
diff --git a/tensorflow/python/util/deprecation.py b/tensorflow/python/util/deprecation.py
index 376be39978fb11463ae8a870492a359c89a9f2ce..c8ed2b715d674a4b4cf3fc0acf333221fcaa4352 100644
--- a/tensorflow/python/util/deprecation.py
+++ b/tensorflow/python/util/deprecation.py
@@ -87,6 +87,27 @@ def _call_location(outer=False):
     return '%s:%d' % (entry[1], entry[2])
 
 
+def _wrap_decorator(wrapped_function):
+  """Indicate that one function wraps another.
+
+  This decorator wraps a function using `tf_decorator.make_decorator`
+  so that doc generation scripts can pick up original function
+  signature.
+  It would be better to use @functools.wrap decorator, but it would
+  not update function signature to match wrapped function in Python 2.
+
+  Args:
+    wrapped_function: The function that decorated function wraps.
+
+  Returns:
+    Function that accepts wrapper function as an argument and returns
+    `TFDecorator` instance.
+  """
+  def wrapper(wrapper_func):
+    return tf_decorator.make_decorator(wrapped_function, wrapper_func)
+  return wrapper
+
+
 def deprecated_alias(deprecated_name, name, func_or_class, warn_once=True):
   """Deprecate a symbol in favor of a new name with identical semantics.
 
@@ -144,7 +165,7 @@ def deprecated_alias(deprecated_name, name, func_or_class, warn_once=True):
   if tf_inspect.isclass(func_or_class):
 
     # Make a new class with __init__ wrapped in a warning.
-    class NewClass(func_or_class):  # pylint: disable=missing-docstring
+    class _NewClass(func_or_class):  # pylint: disable=missing-docstring
       __doc__ = decorator_utils.add_notice_to_docstring(
           func_or_class.__doc__, 'Please use %s instead.' % name,
           'DEPRECATED CLASS',
@@ -153,27 +174,28 @@ def deprecated_alias(deprecated_name, name, func_or_class, warn_once=True):
       __name__ = func_or_class.__name__
       __module__ = _call_location(outer=True)
 
+      @_wrap_decorator(func_or_class.__init__)
       def __init__(self, *args, **kwargs):
-        if hasattr(NewClass.__init__, '__func__'):
+        if hasattr(_NewClass.__init__, '__func__'):
           # Python 2
-          NewClass.__init__.__func__.__doc__ = func_or_class.__init__.__doc__
+          _NewClass.__init__.__func__.__doc__ = func_or_class.__init__.__doc__
         else:
           # Python 3
-          NewClass.__init__.__doc__ = func_or_class.__init__.__doc__
+          _NewClass.__init__.__doc__ = func_or_class.__init__.__doc__
 
         if _PRINT_DEPRECATION_WARNINGS:
           # We're making the alias as we speak. The original may have other
           # aliases, so we cannot use it to check for whether it's already been
           # warned about.
-          if NewClass.__init__ not in _PRINTED_WARNING:
+          if _NewClass.__init__ not in _PRINTED_WARNING:
             if warn_once:
-              _PRINTED_WARNING[NewClass.__init__] = True
+              _PRINTED_WARNING[_NewClass.__init__] = True
             logging.warning(
                 'From %s: The name %s is deprecated. Please use %s instead.\n',
                 _call_location(), deprecated_name, name)
-        super(NewClass, self).__init__(*args, **kwargs)
+        super(_NewClass, self).__init__(*args, **kwargs)
 
-    return NewClass
+    return _NewClass
   else:
     decorator_utils.validate_callable(func_or_class, 'deprecated')
 
diff --git a/tensorflow/python/util/deprecation_test.py b/tensorflow/python/util/deprecation_test.py
index bdd0bc48d29319914e184ea4331a5e9d4a1c3328..1ea695e4d68ddc1727b244cc7330815d0db8b7ef 100644
--- a/tensorflow/python/util/deprecation_test.py
+++ b/tensorflow/python/util/deprecation_test.py
@@ -22,6 +22,7 @@ from __future__ import print_function
 from tensorflow.python.platform import test
 from tensorflow.python.platform import tf_logging as logging
 from tensorflow.python.util import deprecation
+from tensorflow.python.util import tf_inspect
 
 
 class DeprecatedAliasTest(test.TestCase):
@@ -73,6 +74,11 @@ class DeprecatedAliasTest(test.TestCase):
     self.assertEqual(["test", "deprecated", "deprecated again"],
                      MyClass.init_args)
 
+    # Check __init__ signature matches for doc generation.
+    self.assertEqual(
+        tf_inspect.getfullargspec(MyClass.__init__),
+        tf_inspect.getfullargspec(deprecated_cls.__init__))
+
 
 class DeprecationTest(test.TestCase):
 
diff --git a/tensorflow/python/util/lock_util_test.py b/tensorflow/python/util/lock_util_test.py
index 2ac640ff990d802986075e2358330fd2536e2f8a..cda8f952259c9e117e0bd7ff3cac35e764856f43 100644
--- a/tensorflow/python/util/lock_util_test.py
+++ b/tensorflow/python/util/lock_util_test.py
@@ -19,7 +19,6 @@ from __future__ import division
 from __future__ import print_function
 
 import random
-import threading
 import time
 
 from absl.testing import parameterized
@@ -48,7 +47,7 @@ class GroupLockTest(test.TestCase, parameterized.TestCase):
         finished.add(thread_id)
 
     threads = [
-        threading.Thread(target=thread_fn, args=(i,))
+        self.checkedThread(target=thread_fn, args=(i,))
         for i in range(num_threads)
     ]
 
diff --git a/tensorflow/python/util/nest.py b/tensorflow/python/util/nest.py
index 1104768ae8f69598f686eb2ffee8b69e43051011..d63f59a8c8e836d3f8ad3686da0b0b3f010a9225 100644
--- a/tensorflow/python/util/nest.py
+++ b/tensorflow/python/util/nest.py
@@ -167,11 +167,14 @@ def assert_same_structure(nest1, nest2, check_types=True):
   Args:
     nest1: an arbitrarily nested structure.
     nest2: an arbitrarily nested structure.
-    check_types: if `True` (default) types of sequences are checked as
-        well, including the keys of dictionaries. If set to `False`, for example
-        a list and a tuple of objects will look the same if they have the same
+    check_types: if `True` (default) types of sequences are checked as well,
+        including the keys of dictionaries. If set to `False`, for example a
+        list and a tuple of objects will look the same if they have the same
         size. Note that namedtuples with identical name and fields are always
-        considered to have the same shallow structure.
+        considered to have the same shallow structure. Two types will also be
+        considered the same if they are both list subtypes (which allows "list"
+        and "_ListWrapper" from checkpointable dependency tracking to compare
+        equal).
 
   Raises:
     ValueError: If the two structures do not have the same number of elements or
diff --git a/tensorflow/python/util/py_checkpoint_reader.i b/tensorflow/python/util/py_checkpoint_reader.i
index 8004898cbcbce7ce593ce35efdc6493e052468bd..1c73f7f06f1937a8db0bd858421c2e884892e25b 100644
--- a/tensorflow/python/util/py_checkpoint_reader.i
+++ b/tensorflow/python/util/py_checkpoint_reader.i
@@ -166,6 +166,7 @@ def NewCheckpointReader(filepattern):
     return CheckpointReader(compat.as_bytes(filepattern), status)
 
 NewCheckpointReader._tf_api_names = ['train.NewCheckpointReader']
+NewCheckpointReader._tf_api_names_v1 = ['train.NewCheckpointReader']
 %}
 
 %include "tensorflow/c/checkpoint_reader.h"
diff --git a/tensorflow/python/util/stat_summarizer.i b/tensorflow/python/util/stat_summarizer.i
index 73fa85494b72d920d00577c826b76c3381d963a4..a5a7984d914f24964c377149f8125ceb3126c009 100644
--- a/tensorflow/python/util/stat_summarizer.i
+++ b/tensorflow/python/util/stat_summarizer.i
@@ -27,8 +27,8 @@ limitations under the License.
 
 %ignoreall
 
-%unignore _NewStatSummarizer;
-%unignore _DeleteStatSummarizer;
+%unignore NewStatSummarizer;
+%unignore DeleteStatSummarizer;
 %unignore tensorflow;
 %unignore tensorflow::StatSummarizer;
 %unignore tensorflow::StatSummarizer::StatSummarizer;
@@ -43,20 +43,20 @@ limitations under the License.
 
 // TODO(ashankar): Remove the unused argument from the API.
 %{
-tensorflow::StatSummarizer* _NewStatSummarizer(
+tensorflow::StatSummarizer* NewStatSummarizer(
       const string& unused) {
   return new tensorflow::StatSummarizer(tensorflow::StatSummarizerOptions());
 }
 %}
 
 %{
-void _DeleteStatSummarizer(tensorflow::StatSummarizer* ss) {
+void DeleteStatSummarizer(tensorflow::StatSummarizer* ss) {
   delete ss;
 }
 %}
 
-tensorflow::StatSummarizer* _NewStatSummarizer(const string& unused);
-void _DeleteStatSummarizer(tensorflow::StatSummarizer* ss);
+tensorflow::StatSummarizer* NewStatSummarizer(const string& unused);
+void DeleteStatSummarizer(tensorflow::StatSummarizer* ss);
 
 %extend tensorflow::StatSummarizer {
   void ProcessStepStatsStr(const string& step_stats_str) {
@@ -76,16 +76,3 @@ void _DeleteStatSummarizer(tensorflow::StatSummarizer* ss);
 %include "tensorflow/core/util/stat_summarizer_options.h"
 %include "tensorflow/core/util/stat_summarizer.h"
 %unignoreall
-
-%insert("python") %{
-
-# Wrapping NewStatSummarizer and DeletStatSummarizer because
-# SWIG-generated functions are built-in functions and do not support
-# setting _tf_api_names attribute.
-
-def NewStatSummarizer(unused):
-  return _NewStatSummarizer(unused)
-
-def DeleteStatSummarizer(stat_summarizer):
-  _DeleteStatSummarizer(stat_summarizer)
-%}
diff --git a/tensorflow/python/util/tf_export.py b/tensorflow/python/util/tf_export.py
index e154ffb68a4f0ccdebf5320cad7d3da056117197..c362d588abfed684932842d757b33a1b1946970c 100644
--- a/tensorflow/python/util/tf_export.py
+++ b/tensorflow/python/util/tf_export.py
@@ -63,6 +63,15 @@ API_ATTRS = {
         '_estimator_api_constants')
 }
 
+API_ATTRS_V1 = {
+    TENSORFLOW_API_NAME: _Attributes(
+        '_tf_api_names_v1',
+        '_tf_api_constants_v1'),
+    ESTIMATOR_API_NAME: _Attributes(
+        '_estimator_api_names_v1',
+        '_estimator_api_constants_v1')
+}
+
 
 class SymbolAlreadyExposedError(Exception):
   """Raised when adding API names to symbol that already has API names."""
@@ -78,13 +87,16 @@ class api_export(object):  # pylint: disable=invalid-name
     Args:
       *args: API names in dot delimited format.
       **kwargs: Optional keyed arguments.
-          overrides: List of symbols that this is overriding
+        v1: Names for the TensorFlow V1 API. If not set, we will use V2 API
+          names both for TensorFlow V1 and V2 APIs.
+        overrides: List of symbols that this is overriding
           (those overrided api exports will be removed). Note: passing overrides
           has no effect on exporting a constant.
-          api_name: Name of the API you want to generate (e.g. `tensorflow` or
+        api_name: Name of the API you want to generate (e.g. `tensorflow` or
           `estimator`). Default is `tensorflow`.
     """
     self._names = args
+    self._names_v1 = kwargs.get('v1', args)
     self._api_name = kwargs.get('api_name', TENSORFLOW_API_NAME)
     self._overrides = kwargs.get('overrides', [])
 
@@ -102,24 +114,27 @@ class api_export(object):  # pylint: disable=invalid-name
         and kwarg `allow_multiple_exports` not set.
     """
     api_names_attr = API_ATTRS[self._api_name].names
-
+    api_names_attr_v1 = API_ATTRS_V1[self._api_name].names
     # Undecorate overridden names
     for f in self._overrides:
       _, undecorated_f = tf_decorator.unwrap(f)
       delattr(undecorated_f, api_names_attr)
+      delattr(undecorated_f, api_names_attr_v1)
 
     _, undecorated_func = tf_decorator.unwrap(func)
+    self.set_attr(undecorated_func, api_names_attr, self._names)
+    self.set_attr(undecorated_func, api_names_attr_v1, self._names_v1)
+    return func
 
+  def set_attr(self, func, api_names_attr, names):
     # Check for an existing api. We check if attribute name is in
     # __dict__ instead of using hasattr to verify that subclasses have
     # their own _tf_api_names as opposed to just inheriting it.
-    if api_names_attr in undecorated_func.__dict__:
+    if api_names_attr in func.__dict__:
       raise SymbolAlreadyExposedError(
           'Symbol %s is already exposed as %s.' %
-          (undecorated_func.__name__, getattr(
-              undecorated_func, api_names_attr)))  # pylint: disable=protected-access
-    setattr(undecorated_func, api_names_attr, self._names)
-    return func
+          (func.__name__, getattr(func, api_names_attr)))  # pylint: disable=protected-access
+    setattr(func, api_names_attr, names)
 
   def export_constant(self, module_name, name):
     """Store export information for constants/string literals.
@@ -140,12 +155,20 @@ class api_export(object):  # pylint: disable=invalid-name
       name: (string) Current constant name.
     """
     module = sys.modules[module_name]
-    if not hasattr(module, API_ATTRS[self._api_name].constants):
-      setattr(module, API_ATTRS[self._api_name].constants, [])
+    api_constants_attr = API_ATTRS[self._api_name].constants
+    api_constants_attr_v1 = API_ATTRS_V1[self._api_name].constants
+
+    if not hasattr(module, api_constants_attr):
+      setattr(module, api_constants_attr, [])
     # pylint: disable=protected-access
-    getattr(module, API_ATTRS[self._api_name].constants).append(
+    getattr(module, api_constants_attr).append(
         (self._names, name))
 
+    if not hasattr(module, api_constants_attr_v1):
+      setattr(module, api_constants_attr_v1, [])
+    getattr(module, api_constants_attr_v1).append(
+        (self._names_v1, name))
+
 
 tf_export = functools.partial(api_export, api_name=TENSORFLOW_API_NAME)
 estimator_export = functools.partial(tf_export, api_name=ESTIMATOR_API_NAME)
diff --git a/tensorflow/python/util/tf_export_test.py b/tensorflow/python/util/tf_export_test.py
index b9e26ecb33383f5aa936a6bc92acea6d91eb996e..4ae1dc55e06b434aeb4a95e2ca9aa68e4eef56de 100644
--- a/tensorflow/python/util/tf_export_test.py
+++ b/tensorflow/python/util/tf_export_test.py
@@ -60,6 +60,8 @@ class ValidateExportTest(test.TestCase):
     for symbol in [_test_function, _test_function, TestClassA, TestClassB]:
       if hasattr(symbol, '_tf_api_names'):
         del symbol._tf_api_names
+      if hasattr(symbol, '_tf_api_names_v1'):
+        del symbol._tf_api_names_v1
 
   def _CreateMockModule(self, name):
     mock_module = self.MockModule(name)
diff --git a/tensorflow/python/util/tf_inspect.py b/tensorflow/python/util/tf_inspect.py
index fbd65617670b15bfc69506bab1e83369081502af..ec20998bdd68444e830d78689465f104177e7fec 100644
--- a/tensorflow/python/util/tf_inspect.py
+++ b/tensorflow/python/util/tf_inspect.py
@@ -300,6 +300,16 @@ def getsource(object):  # pylint: disable=redefined-builtin
   return _inspect.getsource(tf_decorator.unwrap(object)[1])
 
 
+def getsourcefile(object):  # pylint: disable=redefined-builtin
+  """TFDecorator-aware replacement for inspect.getsourcefile."""
+  return _inspect.getsourcefile(tf_decorator.unwrap(object)[1])
+
+
+def getsourcelines(object):  # pylint: disable=redefined-builtin
+  """TFDecorator-aware replacement for inspect.getsourcelines."""
+  return _inspect.getsourcelines(tf_decorator.unwrap(object)[1])
+
+
 def isbuiltin(object):  # pylint: disable=redefined-builtin
   """TFDecorator-aware replacement for inspect.isbuiltin."""
   return _inspect.isbuiltin(tf_decorator.unwrap(object)[1])
diff --git a/tensorflow/python/util/tf_inspect_test.py b/tensorflow/python/util/tf_inspect_test.py
index beaf350de1e469a7675a4b55ff341419262b79b2..2f6021c7d8e64f2474334ff38f203d0f5fc93f81 100644
--- a/tensorflow/python/util/tf_inspect_test.py
+++ b/tensorflow/python/util/tf_inspect_test.py
@@ -326,6 +326,18 @@ def test_decorated_function_with_defaults(a, b=2, c='Hello'):
     self.assertEqual(
         expected, tf_inspect.getsource(test_decorated_function_with_defaults))
 
+  def testGetSourceFile(self):
+    self.assertEqual(
+        __file__,
+        tf_inspect.getsourcefile(test_decorated_function_with_defaults))
+
+  def testGetSourceLines(self):
+    expected = inspect.getsourcelines(
+        test_decorated_function_with_defaults.decorated_target)
+    self.assertEqual(
+        expected,
+        tf_inspect.getsourcelines(test_decorated_function_with_defaults))
+
   def testIsBuiltin(self):
     self.assertEqual(
         tf_inspect.isbuiltin(TestDecoratedClass),
diff --git a/tensorflow/python/util/tf_stack.py b/tensorflow/python/util/tf_stack.py
new file mode 100644
index 0000000000000000000000000000000000000000..dacc1ce83e11a097689de029695f34b2ba2be4e4
--- /dev/null
+++ b/tensorflow/python/util/tf_stack.py
@@ -0,0 +1,97 @@
+# Copyright 2015 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""Functions used to extract and analyze stacks.  Faster than Python libs."""
+# pylint: disable=g-bad-name
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import linecache
+import sys
+
+
+def extract_stack(extract_frame_info_fn=None):
+  """A lightweight, extensible re-implementation of traceback.extract_stack.
+
+  NOTE(mrry): traceback.extract_stack eagerly retrieves the line of code for
+      each stack frame using linecache, which results in an abundance of stat()
+      calls. This implementation does not retrieve the code, and any consumer
+      should apply _convert_stack to the result to obtain a traceback that can
+      be formatted etc. using traceback methods.
+
+  Args:
+    extract_frame_info_fn: Optional callable fn(stack_frame) applied to each
+        stack frame.  This callable's return value is stored as the sixth (last)
+        element of the returned tuples.  If not provided, the returned tuples
+        will have None as their sixth value.
+
+  Returns:
+    A list of 6-tuples
+        (filename, lineno, name, frame_globals, func_start_lineno, custom_info)
+    corresponding to the call stack of the current thread.  The returned tuples
+    have the innermost stack frame at the end, unlike the Python inspect
+    module's stack() function.
+  """
+  default_fn = lambda f: None
+  extract_frame_info_fn = extract_frame_info_fn or default_fn
+  try:
+    raise ZeroDivisionError
+  except ZeroDivisionError:
+    f = sys.exc_info()[2].tb_frame.f_back
+  ret = []
+  while f is not None:
+    lineno = f.f_lineno
+    co = f.f_code
+    filename = co.co_filename
+    name = co.co_name
+    frame_globals = f.f_globals
+    func_start_lineno = co.co_firstlineno
+    frame_info = extract_frame_info_fn(f)
+    ret.append((filename, lineno, name, frame_globals, func_start_lineno,
+                frame_info))
+    f = f.f_back
+  ret.reverse()
+  return ret
+
+
+def convert_stack(stack, include_func_start_lineno=False):
+  """Converts a stack extracted using extract_stack() to a traceback stack.
+
+  Args:
+    stack: A list of n 5-tuples,
+      (filename, lineno, name, frame_globals, func_start_lineno).
+    include_func_start_lineno: True if function start line number should be
+      included as the 5th entry in return tuples.
+
+  Returns:
+    A list of n 4-tuples or 5-tuples
+    (filename, lineno, name, code, [optional: func_start_lineno]), where the
+    code tuple element is calculated from the corresponding elements of the
+    input tuple.
+  """
+  ret = []
+  for (filename, lineno, name, frame_globals, func_start_lineno,
+       unused_frame_info) in stack:
+    linecache.checkcache(filename)
+    line = linecache.getline(filename, lineno, frame_globals)
+    if line:
+      line = line.strip()
+    else:
+      line = None
+    if include_func_start_lineno:
+      ret.append((filename, lineno, name, line, func_start_lineno))
+    else:
+      ret.append((filename, lineno, name, line))
+  return ret
diff --git a/tensorflow/python/util/util.cc b/tensorflow/python/util/util.cc
index c79d8a84458800937e3e51a8dae26605bd834233..366f8a0deb533c3ee258ea618136d44a28160f8f 100644
--- a/tensorflow/python/util/util.cc
+++ b/tensorflow/python/util/util.cc
@@ -394,7 +394,11 @@ bool AssertSameStructureHelper(PyObject* o1, PyObject* o2, bool check_types,
             type2->tp_name);
         return true;
       }
-    } else if (type1 != type2) {
+    } else if (type1 != type2
+               /* If both sequences are list types, don't complain. This allows
+               one to be a list subclass (e.g. _ListWrapper used for automatic
+               dependency tracking.) */
+               && !(PyList_Check(o1) && PyList_Check(o2))) {
       *is_type_error = true;
       *error_msg = tensorflow::strings::StrCat(
           "The two namedtuples don't have the same sequence type. "
diff --git a/tensorflow/stream_executor/BUILD b/tensorflow/stream_executor/BUILD
index c68cda01002b1c5bbc2facb95b1eba214fbad7cb..e742f8e8d51d0217b631ebdc23ee65263c1ce0f0 100644
--- a/tensorflow/stream_executor/BUILD
+++ b/tensorflow/stream_executor/BUILD
@@ -2,6 +2,7 @@ licenses(["restricted"])
 
 load("@local_config_cuda//cuda:build_defs.bzl", "if_cuda_is_configured")
 load("//tensorflow/core:platform/default/build_config_root.bzl", "if_static")
+load("//tensorflow:tensorflow.bzl", "cc_header_only_library")
 
 STREAM_EXECUTOR_HEADERS = glob([
     "*.h",
@@ -33,7 +34,6 @@ cc_library(
     }),
     visibility = ["//visibility:public"],
     deps = [
-        "//tensorflow/compiler/xla:statusor",
         "//tensorflow/core:lib",
         "//tensorflow/core:ptr_util",
         "@local_config_cuda//cuda:cuda_headers",
@@ -48,11 +48,18 @@ cc_library(
     deps = [
         "//tensorflow/core:lib",
         "//tensorflow/core:ptr_util",
-        "//tensorflow/compiler/xla:statusor",
         "@local_config_cuda//cuda:cuda_headers",
     ] + if_static([":stream_executor_impl"]),
 )
 
+cc_header_only_library(
+    name = "stream_executor_headers_lib",
+    visibility = ["//visibility:public"],
+    deps = [
+        ":stream_executor",
+    ],
+)
+
 cc_library(
     name = "cuda_platform",
     srcs = if_cuda_is_configured(
diff --git a/tensorflow/stream_executor/cuda/cuda_dnn.cc b/tensorflow/stream_executor/cuda/cuda_dnn.cc
index d4f2fd262544a6a2771e17a712ae911c7249a7d5..84916385a89b6e2bafb8a3c0a8f435ec9626e816 100644
--- a/tensorflow/stream_executor/cuda/cuda_dnn.cc
+++ b/tensorflow/stream_executor/cuda/cuda_dnn.cc
@@ -3074,6 +3074,22 @@ port::Status CudnnSupport::DoConvolveBackwardDataImpl(
     }
   }
 
+  // Cudnn 7.1.4 has a bug if the workspace of the following convolution is not
+  // zero-initialized.
+  // TODO(timshen): Add an nvbugs/ link.
+  if (CUDNN_VERSION >= 7000 &&
+      algorithm_config.algorithm().algo_id() ==
+          CUDNN_CONVOLUTION_BWD_DATA_ALGO_1 &&
+      cudnn_type == CUDNN_DATA_HALF &&
+      algorithm_config.algorithm().tensor_ops_enabled() &&
+      input_descriptor.layout() == dnn::DataLayout::kBatchYXDepth &&
+      filter_descriptor.layout() == dnn::FilterLayout::kOutputInputYX &&
+      output_descriptor.layout() == dnn::DataLayout::kBatchDepthYX &&
+      (convolution_descriptor.vertical_filter_stride() > 1 ||
+       convolution_descriptor.horizontal_filter_stride() > 1)) {
+    stream->ThenMemZero(&scratch, scratch.size());
+  }
+
   RETURN_IF_CUDNN_ERROR(
       cudnnConvolutionBackwardData(cudnn.handle(),
                                    /*alpha=*/alpha,
diff --git a/tensorflow/stream_executor/event.cc b/tensorflow/stream_executor/event.cc
index 50a6edd80bd39004e32f09bcde36fbc8a8b59ad9..52efe771bc3c43e65b4539f811196e2d8785eb77 100644
--- a/tensorflow/stream_executor/event.cc
+++ b/tensorflow/stream_executor/event.cc
@@ -15,9 +15,9 @@ limitations under the License.
 
 #include "tensorflow/stream_executor/event.h"
 
+#include "tensorflow/stream_executor/stream.h"
 #include "tensorflow/stream_executor/stream_executor_internal.h"
 #include "tensorflow/stream_executor/stream_executor_pimpl.h"
-#include "tensorflow/stream_executor/stream.h"
 
 namespace stream_executor {
 
@@ -27,9 +27,12 @@ Event::Event(StreamExecutor* stream_exec)
           stream_exec_->implementation()->CreateEventImplementation()) {}
 
 Event::~Event() {
-  auto status = stream_exec_->DeallocateEvent(this);
-  if (!status.ok()) {
-    LOG(ERROR) << status.error_message();
+  // Deal with nullptr implementation_, as this event may have been std::moved.
+  if (stream_exec_ && implementation_) {
+    auto status = stream_exec_->DeallocateEvent(this);
+    if (!status.ok()) {
+      LOG(ERROR) << status.error_message();
+    }
   }
 }
 
diff --git a/tensorflow/stream_executor/event.h b/tensorflow/stream_executor/event.h
index 1f37262c78d82f72f8818f35db273e87a47bdc1c..9cc87a7c129962820ed0c84d02faada4ba460d51 100644
--- a/tensorflow/stream_executor/event.h
+++ b/tensorflow/stream_executor/event.h
@@ -61,6 +61,9 @@ class Event {
   // Returns a pointer to the underlying platform-specific implementation.
   internal::EventInterface* implementation() { return implementation_.get(); }
 
+  Event(Event&&) = default;
+  Event& operator=(Event&&) = default;
+
  private:
   friend class Stream;
 
diff --git a/tensorflow/stream_executor/host/host_gpu_executor.cc b/tensorflow/stream_executor/host/host_gpu_executor.cc
index 2c4819651acaa2c6ee99c720b2c3d80e5c2ea1a9..3cd97b3cf165520e236ff6a1ce9280426fe5ed1f 100644
--- a/tensorflow/stream_executor/host/host_gpu_executor.cc
+++ b/tensorflow/stream_executor/host/host_gpu_executor.cc
@@ -26,8 +26,6 @@ limitations under the License.
 #include "tensorflow/stream_executor/lib/statusor.h"
 #include "tensorflow/stream_executor/plugin_registry.h"
 
-bool FLAGS_stream_executor_cpu_real_clock_rate = false;
-
 namespace stream_executor {
 namespace host {
 
@@ -190,11 +188,8 @@ DeviceDescription *HostExecutor::PopulateDeviceDescription() const {
   // doesn't result in thrashing or other badness? 4GiB chosen arbitrarily.
   builder.set_device_memory_size(static_cast(4) * 1024 * 1024 * 1024);
 
-  float cycle_counter_frequency = 1e9;
-  if (FLAGS_stream_executor_cpu_real_clock_rate) {
-    cycle_counter_frequency = static_cast(
-        tensorflow::profile_utils::CpuUtils::GetCycleCounterFrequency());
-  }
+  float cycle_counter_frequency = static_cast(
+      tensorflow::profile_utils::CpuUtils::GetCycleCounterFrequency());
   builder.set_clock_rate_ghz(cycle_counter_frequency / 1e9);
 
   auto built = builder.Build();
diff --git a/tensorflow/compiler/xla/statusor.cc b/tensorflow/stream_executor/lib/statusor.cc
similarity index 89%
rename from tensorflow/compiler/xla/statusor.cc
rename to tensorflow/stream_executor/lib/statusor.cc
index 72ab67ff810e0ec384a22da092363cc7446435bb..e0e851f96ef6fe18ec32ff7d3fd1d1aed18b0343 100644
--- a/tensorflow/compiler/xla/statusor.cc
+++ b/tensorflow/stream_executor/lib/statusor.cc
@@ -13,12 +13,13 @@ See the License for the specific language governing permissions and
 limitations under the License.
 ==============================================================================*/
 
-#include "tensorflow/compiler/xla/statusor.h"
+#include "tensorflow/stream_executor/lib/statusor.h"
 
 #include "tensorflow/core/lib/core/errors.h"
 #include "tensorflow/core/platform/logging.h"
 
-namespace xla {
+namespace stream_executor {
+namespace port {
 namespace internal_statusor {
 
 void Helper::HandleInvalidStatusCtorArg(Status* status) {
@@ -35,4 +36,5 @@ void Helper::Crash(const Status& status) {
 }
 
 }  // namespace internal_statusor
-}  // namespace xla
+}  // namespace port
+}  // namespace stream_executor
diff --git a/tensorflow/stream_executor/lib/statusor.h b/tensorflow/stream_executor/lib/statusor.h
index dab59096740102b94c0ff63c089b83ce052ea264..3c716acb462f1ca25e1d86408386d9eca37265b7 100644
--- a/tensorflow/stream_executor/lib/statusor.h
+++ b/tensorflow/stream_executor/lib/statusor.h
@@ -1,4 +1,4 @@
-/* Copyright 2015 The TensorFlow Authors. All Rights Reserved.
+/* Copyright 2017 The TensorFlow Authors. All Rights Reserved.
 
 Licensed under the Apache License, Version 2.0 (the "License");
 you may not use this file except in compliance with the License.
@@ -13,19 +13,297 @@ See the License for the specific language governing permissions and
 limitations under the License.
 ==============================================================================*/
 
-// IWYU pragma: private, include "third_party/tensorflow/stream_executor/stream_executor.h"
-
+// StatusOr is the union of a Status object and a T object. StatusOr models
+// the concept of an object that is either a value, or an error Status
+// explaining why such a value is not present. To this end, StatusOr does not
+// allow its Status value to be Status::OK.
+//
+// The primary use-case for StatusOr is as the return value of a
+// function which may fail.
+//
+// Example client usage for a StatusOr, where T is not a pointer:
+//
+//  StatusOr result = DoBigCalculationThatCouldFail();
+//  if (result.ok()) {
+//    float answer = result.ValueOrDie();
+//    printf("Big calculation yielded: %f", answer);
+//  } else {
+//    LOG(ERROR) << result.status();
+//  }
+//
+// Example client usage for a StatusOr:
+//
+//  StatusOr result = FooFactory::MakeNewFoo(arg);
+//  if (result.ok()) {
+//    std::unique_ptr foo(result.ValueOrDie());
+//    foo->DoSomethingCool();
+//  } else {
+//    LOG(ERROR) << result.status();
+//  }
+//
+// Example client usage for a StatusOr>:
+//
+//  StatusOr> result = FooFactory::MakeNewFoo(arg);
+//  if (result.ok()) {
+//    std::unique_ptr foo = std::move(result.ValueOrDie());
+//    foo->DoSomethingCool();
+//  } else {
+//    LOG(ERROR) << result.status();
+//  }
+//
+// Example factory implementation returning StatusOr:
+//
+//  StatusOr FooFactory::MakeNewFoo(int arg) {
+//    if (arg <= 0) {
+//      return tensorflow::InvalidArgument("Arg must be positive");
+//    } else {
+//      return new Foo(arg);
+//    }
+//  }
+//
+// Note that the assignment operators require that destroying the currently
+// stored value cannot invalidate the argument; in other words, the argument
+// cannot be an alias for the current value, or anything owned by the current
+// value.
 #ifndef TENSORFLOW_STREAM_EXECUTOR_LIB_STATUSOR_H_
 #define TENSORFLOW_STREAM_EXECUTOR_LIB_STATUSOR_H_
 
-#include "tensorflow/compiler/xla/statusor.h"
+#include "tensorflow/core/platform/macros.h"
+#include "tensorflow/stream_executor/lib/status.h"
+#include "tensorflow/stream_executor/lib/statusor_internals.h"
 
 namespace stream_executor {
 namespace port {
 
-// Use XLA's StatusOr so we don't duplicate code.
+#if defined(__clang__)
+// Only clang supports warn_unused_result as a type annotation.
+template 
+class TF_MUST_USE_RESULT StatusOr;
+#endif
+
+template 
+class StatusOr : private internal_statusor::StatusOrData,
+                 private internal_statusor::TraitsBase<
+                     std::is_copy_constructible::value,
+                     std::is_move_constructible::value> {
+  template 
+  friend class StatusOr;
+
+  typedef internal_statusor::StatusOrData Base;
+
+ public:
+  typedef T element_type;
+
+  // Constructs a new StatusOr with Status::UNKNOWN status.  This is marked
+  // 'explicit' to try to catch cases like 'return {};', where people think
+  // StatusOr> will be initialized with an empty vector,
+  // instead of a Status::UNKNOWN status.
+  explicit StatusOr();
+
+  // StatusOr will be copy constructible/assignable if T is copy
+  // constructible.
+  StatusOr(const StatusOr&) = default;
+  StatusOr& operator=(const StatusOr&) = default;
+
+  // StatusOr will be move constructible/assignable if T is move
+  // constructible.
+  StatusOr(StatusOr&&) = default;
+  StatusOr& operator=(StatusOr&&) = default;
+
+  // Conversion copy/move constructor, T must be convertible from U.
+  template ::value>::type* = nullptr>
+  StatusOr(const StatusOr& other);
+  template ::value>::type* = nullptr>
+  StatusOr(StatusOr&& other);
+
+  // Conversion copy/move assignment operator, T must be convertible from U.
+  template ::value>::type* = nullptr>
+  StatusOr& operator=(const StatusOr& other);
+  template ::value>::type* = nullptr>
+  StatusOr& operator=(StatusOr&& other);
+
+  // Constructs a new StatusOr with the given value. After calling this
+  // constructor, calls to ValueOrDie() will succeed, and calls to status() will
+  // return OK.
+  //
+  // NOTE: Not explicit - we want to use StatusOr as a return type
+  // so it is convenient and sensible to be able to do 'return T()'
+  // when the return type is StatusOr.
+  //
+  // REQUIRES: T is copy constructible.
+  StatusOr(const T& value);
+
+  // Constructs a new StatusOr with the given non-ok status. After calling
+  // this constructor, calls to ValueOrDie() will CHECK-fail.
+  //
+  // NOTE: Not explicit - we want to use StatusOr as a return
+  // value, so it is convenient and sensible to be able to do 'return
+  // Status()' when the return type is StatusOr.
+  //
+  // REQUIRES: !status.ok(). This requirement is DCHECKed.
+  // In optimized builds, passing Status::OK() here will have the effect
+  // of passing tensorflow::error::INTERNAL as a fallback.
+  StatusOr(const Status& status);
+  StatusOr& operator=(const Status& status);
+
+  // TODO(b/62186997): Add operator=(T) overloads.
+
+  // Similar to the `const T&` overload.
+  //
+  // REQUIRES: T is move constructible.
+  StatusOr(T&& value);
+
+  // RValue versions of the operations declared above.
+  StatusOr(Status&& status);
+  StatusOr& operator=(Status&& status);
+
+  // Returns this->status().ok()
+  bool ok() const { return this->status_.ok(); }
+
+  // Returns a reference to our status. If this contains a T, then
+  // returns Status::OK().
+  const Status& status() const &;
+  Status status() &&;
+
+  // Returns a reference to our current value, or CHECK-fails if !this->ok().
+  //
+  // Note: for value types that are cheap to copy, prefer simple code:
+  //
+  //   T value = statusor.ValueOrDie();
+  //
+  // Otherwise, if the value type is expensive to copy, but can be left
+  // in the StatusOr, simply assign to a reference:
+  //
+  //   T& value = statusor.ValueOrDie();  // or `const T&`
+  //
+  // Otherwise, if the value type supports an efficient move, it can be
+  // used as follows:
+  //
+  //   T value = std::move(statusor).ValueOrDie();
+  //
+  // The std::move on statusor instead of on the whole expression enables
+  // warnings about possible uses of the statusor object after the move.
+  // C++ style guide waiver for ref-qualified overloads granted in cl/143176389
+  // See go/ref-qualifiers for more details on such overloads.
+  const T& ValueOrDie() const &;
+  T& ValueOrDie() &;
+  const T&& ValueOrDie() const &&;
+  T&& ValueOrDie() &&;
+
+  T ConsumeValueOrDie() { return std::move(ValueOrDie()); }
+
+  // Ignores any errors. This method does nothing except potentially suppress
+  // complaints from any tools that are checking that errors are not dropped on
+  // the floor.
+  void IgnoreError() const;
+};
+
+////////////////////////////////////////////////////////////////////////////////
+// Implementation details for StatusOr
+
+template 
+StatusOr::StatusOr() : Base(Status(tensorflow::error::UNKNOWN, "")) {}
+
+template 
+StatusOr::StatusOr(const T& value) : Base(value) {}
+
+template 
+StatusOr::StatusOr(const Status& status) : Base(status) {}
+
+template 
+StatusOr& StatusOr::operator=(const Status& status) {
+  this->Assign(status);
+  return *this;
+}
+
+template 
+StatusOr::StatusOr(T&& value) : Base(std::move(value)) {}
+
+template 
+StatusOr::StatusOr(Status&& status) : Base(std::move(status)) {}
+
+template 
+StatusOr& StatusOr::operator=(Status&& status) {
+  this->Assign(std::move(status));
+  return *this;
+}
+
+template 
+template ::value>::type*>
+inline StatusOr::StatusOr(const StatusOr& other)
+    : Base(static_cast::Base&>(other)) {}
+
+template 
+template ::value>::type*>
+inline StatusOr& StatusOr::operator=(const StatusOr& other) {
+  if (other.ok())
+    this->Assign(other.ValueOrDie());
+  else
+    this->Assign(other.status());
+  return *this;
+}
+
+template 
+template ::value>::type*>
+inline StatusOr::StatusOr(StatusOr&& other)
+    : Base(static_cast::Base&&>(other)) {}
+
+template 
+template ::value>::type*>
+inline StatusOr& StatusOr::operator=(StatusOr&& other) {
+  if (other.ok()) {
+    this->Assign(std::move(other).ValueOrDie());
+  } else {
+    this->Assign(std::move(other).status());
+  }
+  return *this;
+}
+
+template 
+const Status& StatusOr::status() const & {
+  return this->status_;
+}
+template 
+Status StatusOr::status() && {
+  return ok() ? Status::OK() : std::move(this->status_);
+}
+
+template 
+const T& StatusOr::ValueOrDie() const & {
+  this->EnsureOk();
+  return this->data_;
+}
+
+template 
+T& StatusOr::ValueOrDie() & {
+  this->EnsureOk();
+  return this->data_;
+}
+
+template 
+const T&& StatusOr::ValueOrDie() const && {
+  this->EnsureOk();
+  return std::move(this->data_);
+}
+
+template 
+T&& StatusOr::ValueOrDie() && {
+  this->EnsureOk();
+  return std::move(this->data_);
+}
+
 template 
-using StatusOr = ::xla::StatusOr;
+void StatusOr::IgnoreError() const {
+  // no-op
+}
 
 }  // namespace port
 }  // namespace stream_executor
diff --git a/tensorflow/compiler/xla/statusor_internals.h b/tensorflow/stream_executor/lib/statusor_internals.h
similarity index 94%
rename from tensorflow/compiler/xla/statusor_internals.h
rename to tensorflow/stream_executor/lib/statusor_internals.h
index 14636bd144bc0a155fc96c5a350c658fd2dadfe6..09f88f5825f57c8e654bd079616a074e84de4f30 100644
--- a/tensorflow/compiler/xla/statusor_internals.h
+++ b/tensorflow/stream_executor/lib/statusor_internals.h
@@ -13,13 +13,15 @@ See the License for the specific language governing permissions and
 limitations under the License.
 ==============================================================================*/
 
-#ifndef TENSORFLOW_COMPILER_XLA_STATUSOR_INTERNALS_H_
-#define TENSORFLOW_COMPILER_XLA_STATUSOR_INTERNALS_H_
+#ifndef TENSORFLOW_STREAM_EXECUTOR_LIB_STATUSOR_INTERNALS_H_
+#define TENSORFLOW_STREAM_EXECUTOR_LIB_STATUSOR_INTERNALS_H_
+
 
-#include "tensorflow/compiler/xla/status.h"
 #include "tensorflow/core/platform/macros.h"
+#include "tensorflow/stream_executor/lib/status.h"
 
-namespace xla {
+namespace stream_executor {
+namespace port {
 namespace internal_statusor {
 
 class Helper {
@@ -240,6 +242,7 @@ struct TraitsBase {
 };
 
 }  // namespace internal_statusor
-}  // namespace xla
+}  // namespace port
+}  // namespace stream_executor
 
-#endif  // TENSORFLOW_COMPILER_XLA_STATUSOR_INTERNALS_H_
+#endif  // TENSORFLOW_STREAM_EXECUTOR_LIB_STATUSOR_INTERNALS_H_
diff --git a/tensorflow/compiler/xla/statusor_test.cc b/tensorflow/stream_executor/lib/statusor_test.cc
similarity index 99%
rename from tensorflow/compiler/xla/statusor_test.cc
rename to tensorflow/stream_executor/lib/statusor_test.cc
index 377a618ffbd99316d409130df8a39f352664dee0..56584e189208b2576f10650fd56bca6d04ecc6c1 100644
--- a/tensorflow/compiler/xla/statusor_test.cc
+++ b/tensorflow/stream_executor/lib/statusor_test.cc
@@ -15,18 +15,18 @@ limitations under the License.
 
 // Unit tests for StatusOr
 
-#include "tensorflow/compiler/xla/statusor.h"
+#include "tensorflow/stream_executor/lib/statusor.h"
 
 #include 
 #include 
 
-#include "tensorflow/compiler/xla/test.h"
-#include "tensorflow/compiler/xla/types.h"
+#include "tensorflow/core/platform/test.h"
 #include "tensorflow/core/lib/core/errors.h"
 #include "tensorflow/core/platform/macros.h"
 #include "tensorflow/core/platform/test_benchmark.h"
 
-namespace xla {
+namespace stream_executor {
+namespace port {
 namespace {
 
 class Base1 {
@@ -672,4 +672,5 @@ void BM_StatusOrFactoryFailLongMsg(int iters) {
 BENCHMARK(BM_StatusOrFactoryFailLongMsg);
 
 }  // namespace
-}  // namespace xla
+}  // namespace port
+}  // namespace stream_executor
diff --git a/tensorflow/stream_executor/stream.cc b/tensorflow/stream_executor/stream.cc
index 0cd0790a72b49bb259b9c72268535b5d74531cf5..9369183133784f614aeb93c4b758a6e6ff8df0f1 100644
--- a/tensorflow/stream_executor/stream.cc
+++ b/tensorflow/stream_executor/stream.cc
@@ -5228,24 +5228,11 @@ port::Status Stream::BlockHostUntilDone() {
     return status;
   }
 
-  port::Status first_error;
-  {
-    // Wait until all active sub-streams have done their tasks.
-    mutex_lock lock(mu_);
-    for (auto &stream : sub_streams_) {
-      if (!stream.second) {
-        first_error.Update(stream.first->BlockHostUntilDone());
-        // Set this sub-stream as available.
-        stream.second = true;
-      }
-    }
-  }
-
   temporary_memory_manager_.DeallocateFinalizedTemporaries();
 
-  first_error.Update(parent_->BlockHostUntilDone(this));
-  CheckError(first_error.ok());
-  return first_error;
+  port::Status error = parent_->BlockHostUntilDone(this);
+  CheckError(error.ok());
+  return error;
 }
 
 }  // namespace stream_executor
diff --git a/tensorflow/stream_executor/stream.h b/tensorflow/stream_executor/stream.h
index a32f4105adea4645a9882df6b264336e8e1a38b4..e8885e1eb682d9ee67c6b7594f96c0911c7c1fa2 100644
--- a/tensorflow/stream_executor/stream.h
+++ b/tensorflow/stream_executor/stream.h
@@ -25,6 +25,7 @@ limitations under the License.
 #include 
 #include 
 
+#include "tensorflow/core/platform/macros.h"
 #include "tensorflow/stream_executor/blas.h"
 #include "tensorflow/stream_executor/device_memory.h"
 #include "tensorflow/stream_executor/dnn.h"
@@ -1349,33 +1350,39 @@ class Stream {
                        DeviceMemory> *x, int incx);
 
   // See BlasSupport::DoBlasGemm.
-  Stream &ThenBlasGemm(blas::Transpose transa, blas::Transpose transb, uint64 m,
-                       uint64 n, uint64 k, float alpha,
-                       const DeviceMemory &a, int lda,
-                       const DeviceMemory &b, int ldb, float beta,
-                       DeviceMemory *c, int ldc);
-  Stream &ThenBlasGemm(blas::Transpose transa, blas::Transpose transb, uint64 m,
-                       uint64 n, uint64 k, float alpha,
-                       const DeviceMemory &a, int lda,
-                       const DeviceMemory &b, int ldb, float beta,
-                       DeviceMemory *c, int ldc);
-  Stream &ThenBlasGemm(blas::Transpose transa, blas::Transpose transb, uint64 m,
-                       uint64 n, uint64 k, double alpha,
-                       const DeviceMemory &a, int lda,
-                       const DeviceMemory &b, int ldb, double beta,
-                       DeviceMemory *c, int ldc);
-  Stream &ThenBlasGemm(blas::Transpose transa, blas::Transpose transb, uint64 m,
-                       uint64 n, uint64 k, std::complex alpha,
-                       const DeviceMemory> &a, int lda,
-                       const DeviceMemory> &b, int ldb,
-                       std::complex beta,
-                       DeviceMemory> *c, int ldc);
-  Stream &ThenBlasGemm(blas::Transpose transa, blas::Transpose transb, uint64 m,
-                       uint64 n, uint64 k, std::complex alpha,
-                       const DeviceMemory> &a, int lda,
-                       const DeviceMemory> &b, int ldb,
-                       std::complex beta,
-                       DeviceMemory> *c, int ldc);
+  TF_EXPORT Stream &ThenBlasGemm(blas::Transpose transa, blas::Transpose transb,
+                                 uint64 m, uint64 n, uint64 k, float alpha,
+                                 const DeviceMemory &a, int lda,
+                                 const DeviceMemory &b, int ldb,
+                                 float beta, DeviceMemory *c,
+                                 int ldc);
+  TF_EXPORT Stream &ThenBlasGemm(blas::Transpose transa, blas::Transpose transb,
+                                 uint64 m, uint64 n, uint64 k, float alpha,
+                                 const DeviceMemory &a, int lda,
+                                 const DeviceMemory &b, int ldb,
+                                 float beta, DeviceMemory *c, int ldc);
+  TF_EXPORT Stream &ThenBlasGemm(blas::Transpose transa, blas::Transpose transb,
+                                 uint64 m, uint64 n, uint64 k, double alpha,
+                                 const DeviceMemory &a, int lda,
+                                 const DeviceMemory &b, int ldb,
+                                 double beta, DeviceMemory *c, int ldc);
+  TF_EXPORT Stream &ThenBlasGemm(blas::Transpose transa, blas::Transpose transb,
+                                 uint64 m, uint64 n, uint64 k,
+                                 std::complex alpha,
+                                 const DeviceMemory> &a,
+                                 int lda,
+                                 const DeviceMemory> &b,
+                                 int ldb, std::complex beta,
+                                 DeviceMemory> *c, int ldc);
+  TF_EXPORT Stream &ThenBlasGemm(blas::Transpose transa, blas::Transpose transb,
+                                 uint64 m, uint64 n, uint64 k,
+                                 std::complex alpha,
+                                 const DeviceMemory> &a,
+                                 int lda,
+                                 const DeviceMemory> &b,
+                                 int ldb, std::complex beta,
+                                 DeviceMemory> *c,
+                                 int ldc);
 
   Stream &ThenBlasGemmWithProfiling(blas::Transpose transa,
                                     blas::Transpose transb, uint64 m, uint64 n,
diff --git a/tensorflow/tensorflow.bzl b/tensorflow/tensorflow.bzl
index 6bb393a3f4a56a0a5f112e0bd91a4a21b818d590..e4241667ad836347f5d3cd543f2f52d22c9885a6 100644
--- a/tensorflow/tensorflow.bzl
+++ b/tensorflow/tensorflow.bzl
@@ -148,6 +148,12 @@ def if_windows(a):
       "//conditions:default": [],
   })
 
+def if_not_windows_cuda(a):
+  return select({
+      clean_dep("//tensorflow:with_cuda_support_windows_override"): [],
+      "//conditions:default": a,
+  })
+
 def if_linux_x86_64(a):
   return select({
       clean_dep("//tensorflow:linux_x86_64"): a,
@@ -819,6 +825,9 @@ def tf_cc_test_mkl(srcs,
                    tags=[],
                    size="medium",
                    args=None):
+  # -fno-exceptions in nocopts breaks compilation if header modules are enabled.
+  disable_header_modules = ["-use_header_modules"]
+
   for src in srcs:
     native.cc_test(
       name=src_to_test_name(src),
@@ -844,6 +853,7 @@ def tf_cc_test_mkl(srcs,
       tags=tags,
       size=size,
       args=args,
+      features=disable_header_modules,
       nocopts="-fno-exceptions")
 
 
@@ -978,16 +988,17 @@ register_extension_info(
     label_regex_for_dep = "{extension_name}",
 )
 
-def tf_kernel_library(name,
-                      prefix=None,
-                      srcs=None,
-                      gpu_srcs=None,
-                      hdrs=None,
-                      deps=None,
-                      alwayslink=1,
-                      copts=None,
-                      is_external=False,
-                      **kwargs):
+def tf_kernel_library(
+        name,
+        prefix = None,
+        srcs = None,
+        gpu_srcs = None,
+        hdrs = None,
+        deps = None,
+        alwayslink = 1,
+        copts = None,
+        is_external = False,
+        **kwargs):
   """A rule to build a TensorFlow OpKernel.
 
   May either specify srcs/hdrs or prefix.  Similar to tf_cuda_library,
@@ -1017,6 +1028,7 @@ def tf_kernel_library(name,
     deps = []
   if not copts:
     copts = []
+  textual_hdrs = []
   copts = copts + tf_copts(is_external=is_external)
   if prefix:
     if native.glob([prefix + "*.cu.cc"], exclude=["*test*"]):
@@ -1027,8 +1039,13 @@ def tf_kernel_library(name,
     srcs = srcs + native.glob(
         [prefix + "*.cc"], exclude=[prefix + "*test*", prefix + "*.cu.cc"])
     hdrs = hdrs + native.glob(
-        [prefix + "*.h"], exclude=[prefix + "*test*", prefix + "*.cu.h"])
-
+            [prefix + "*.h"],
+            exclude = [prefix + "*test*", prefix + "*.cu.h", prefix + "*impl.h"],
+        )
+    textual_hdrs = native.glob(
+            [prefix + "*impl.h"],
+            exclude = [prefix + "*test*", prefix + "*.cu.h"],
+        )
   cuda_deps = [clean_dep("//tensorflow/core:gpu_lib")]
   if gpu_srcs:
     for gpu_src in gpu_srcs:
@@ -1042,6 +1059,7 @@ def tf_kernel_library(name,
       name=name,
       srcs=srcs,
       hdrs=hdrs,
+      textual_hdrs = textual_hdrs,
       copts=copts,
       cuda_deps=cuda_deps,
       linkstatic=1,  # Needed since alwayslink is broken in bazel b/27630669
@@ -1075,6 +1093,9 @@ def tf_mkl_kernel_library(name,
     hdrs = hdrs + native.glob(
         [prefix + "*.h"])
 
+  # -fno-exceptions in nocopts breaks compilation if header modules are enabled.
+  disable_header_modules = ["-use_header_modules"]
+
   native.cc_library(
       name=name,
       srcs=if_mkl(srcs),
@@ -1082,7 +1103,8 @@ def tf_mkl_kernel_library(name,
       deps=deps,
       alwayslink=alwayslink,
       copts=copts,
-      nocopts=nocopts
+      nocopts=nocopts,
+      features = disable_header_modules
   )
 
 register_extension_info(
diff --git a/tensorflow/tools/api/golden/tensorflow.-g-p-u-options.pbtxt b/tensorflow/tools/api/golden/tensorflow.-g-p-u-options.pbtxt
index f819b174c0b701153af4709fade9313efa7f7fb6..353e63127de174a79c209a05327da2de20bf0dd7 100644
--- a/tensorflow/tools/api/golden/tensorflow.-g-p-u-options.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.-g-p-u-options.pbtxt
@@ -72,6 +72,12 @@ tf_proto {
         label: LABEL_OPTIONAL
         type: TYPE_BOOL
       }
+      field {
+        name: "num_dev_to_dev_copy_streams"
+        number: 3
+        label: LABEL_OPTIONAL
+        type: TYPE_INT32
+      }
       nested_type {
         name: "VirtualDevices"
         field {
diff --git a/tensorflow/tools/api/golden/tensorflow.-variable-aggregation.pbtxt b/tensorflow/tools/api/golden/tensorflow.-variable-aggregation.pbtxt
new file mode 100644
index 0000000000000000000000000000000000000000..36b534af360835e3c1cbd1f0fb12a38c42232abf
--- /dev/null
+++ b/tensorflow/tools/api/golden/tensorflow.-variable-aggregation.pbtxt
@@ -0,0 +1,16 @@
+path: "tensorflow.VariableAggregation"
+tf_class {
+  is_instance: ""
+  member {
+    name: "MEAN"
+    mtype: ""
+  }
+  member {
+    name: "NONE"
+    mtype: ""
+  }
+  member {
+    name: "SUM"
+    mtype: ""
+  }
+}
diff --git a/tensorflow/tools/api/golden/tensorflow.-variable-scope.pbtxt b/tensorflow/tools/api/golden/tensorflow.-variable-scope.pbtxt
index 8e539069da05fbb192c383d3f5acff78ab9bfeff..c13eb7b8bb9474f3534582c8af8c3ee4b6c7e076 100644
--- a/tensorflow/tools/api/golden/tensorflow.-variable-scope.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.-variable-scope.pbtxt
@@ -56,7 +56,7 @@ tf_class {
   }
   member_method {
     name: "get_variable"
-    argspec: "args=[\'self\', \'var_store\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'reuse\', \'trainable\', \'collections\', \'caching_device\', \'partitioner\', \'validate_shape\', \'use_resource\', \'custom_getter\', \'constraint\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'var_store\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'reuse\', \'trainable\', \'collections\', \'caching_device\', \'partitioner\', \'validate_shape\', \'use_resource\', \'custom_getter\', \'constraint\', \'synchronization\', \'aggregation\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\'], "
   }
   member_method {
     name: "global_variables"
diff --git a/tensorflow/tools/api/golden/tensorflow.-variable-synchronization.pbtxt b/tensorflow/tools/api/golden/tensorflow.-variable-synchronization.pbtxt
new file mode 100644
index 0000000000000000000000000000000000000000..7589bb28888774839a3011e1e5581f004313f81d
--- /dev/null
+++ b/tensorflow/tools/api/golden/tensorflow.-variable-synchronization.pbtxt
@@ -0,0 +1,20 @@
+path: "tensorflow.VariableSynchronization"
+tf_class {
+  is_instance: ""
+  member {
+    name: "AUTO"
+    mtype: ""
+  }
+  member {
+    name: "NONE"
+    mtype: ""
+  }
+  member {
+    name: "ON_READ"
+    mtype: ""
+  }
+  member {
+    name: "ON_WRITE"
+    mtype: ""
+  }
+}
diff --git a/tensorflow/tools/api/golden/tensorflow.estimator.-boosted-trees-classifier.pbtxt b/tensorflow/tools/api/golden/tensorflow.estimator.-boosted-trees-classifier.pbtxt
index 099838fa65f6a532a594c08e8a44ead8ce008185..9dbb5d16a4e903a755c86bd0a6241180d1999f4d 100644
--- a/tensorflow/tools/api/golden/tensorflow.estimator.-boosted-trees-classifier.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.estimator.-boosted-trees-classifier.pbtxt
@@ -21,7 +21,7 @@ tf_class {
   }
   member_method {
     name: "__init__"
-    argspec: "args=[\'self\', \'feature_columns\', \'n_batches_per_layer\', \'model_dir\', \'n_classes\', \'weight_column\', \'label_vocabulary\', \'n_trees\', \'max_depth\', \'learning_rate\', \'l1_regularization\', \'l2_regularization\', \'tree_complexity\', \'min_node_weight\', \'config\'], varargs=None, keywords=None, defaults=[\'None\', \'\', \'None\', \'None\', \'100\', \'6\', \'0.1\', \'0.0\', \'0.0\', \'0.0\', \'0.0\', \'None\'], "
+    argspec: "args=[\'self\', \'feature_columns\', \'n_batches_per_layer\', \'model_dir\', \'n_classes\', \'weight_column\', \'label_vocabulary\', \'n_trees\', \'max_depth\', \'learning_rate\', \'l1_regularization\', \'l2_regularization\', \'tree_complexity\', \'min_node_weight\', \'config\', \'center_bias\'], varargs=None, keywords=None, defaults=[\'None\', \'\', \'None\', \'None\', \'100\', \'6\', \'0.1\', \'0.0\', \'0.0\', \'0.0\', \'0.0\', \'None\', \'False\'], "
   }
   member_method {
     name: "eval_dir"
diff --git a/tensorflow/tools/api/golden/tensorflow.estimator.-boosted-trees-regressor.pbtxt b/tensorflow/tools/api/golden/tensorflow.estimator.-boosted-trees-regressor.pbtxt
index 87bd19a23a3db727b5c1f13de04e3c11fd91de9b..34a30c2874b90285706c9df6bec8cbbdc3451fe4 100644
--- a/tensorflow/tools/api/golden/tensorflow.estimator.-boosted-trees-regressor.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.estimator.-boosted-trees-regressor.pbtxt
@@ -21,7 +21,7 @@ tf_class {
   }
   member_method {
     name: "__init__"
-    argspec: "args=[\'self\', \'feature_columns\', \'n_batches_per_layer\', \'model_dir\', \'label_dimension\', \'weight_column\', \'n_trees\', \'max_depth\', \'learning_rate\', \'l1_regularization\', \'l2_regularization\', \'tree_complexity\', \'min_node_weight\', \'config\'], varargs=None, keywords=None, defaults=[\'None\', \'\', \'None\', \'100\', \'6\', \'0.1\', \'0.0\', \'0.0\', \'0.0\', \'0.0\', \'None\'], "
+    argspec: "args=[\'self\', \'feature_columns\', \'n_batches_per_layer\', \'model_dir\', \'label_dimension\', \'weight_column\', \'n_trees\', \'max_depth\', \'learning_rate\', \'l1_regularization\', \'l2_regularization\', \'tree_complexity\', \'min_node_weight\', \'config\', \'center_bias\'], varargs=None, keywords=None, defaults=[\'None\', \'\', \'None\', \'100\', \'6\', \'0.1\', \'0.0\', \'0.0\', \'0.0\', \'0.0\', \'None\', \'False\'], "
   }
   member_method {
     name: "eval_dir"
diff --git a/tensorflow/tools/api/golden/tensorflow.estimator.-d-n-n-classifier.pbtxt b/tensorflow/tools/api/golden/tensorflow.estimator.-d-n-n-classifier.pbtxt
index 111914f643a3b192d496c5b0857b4429da12b1d6..0c6b7e4a821ad47c20b6f6074b575bf83c403653 100644
--- a/tensorflow/tools/api/golden/tensorflow.estimator.-d-n-n-classifier.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.estimator.-d-n-n-classifier.pbtxt
@@ -21,7 +21,7 @@ tf_class {
   }
   member_method {
     name: "__init__"
-    argspec: "args=[\'self\', \'hidden_units\', \'feature_columns\', \'model_dir\', \'n_classes\', \'weight_column\', \'label_vocabulary\', \'optimizer\', \'activation_fn\', \'dropout\', \'input_layer_partitioner\', \'config\', \'warm_start_from\', \'loss_reduction\'], varargs=None, keywords=None, defaults=[\'None\', \'2\', \'None\', \'None\', \'Adagrad\', \'\', \'None\', \'None\', \'None\', \'None\', \'weighted_sum\'], "
+    argspec: "args=[\'self\', \'hidden_units\', \'feature_columns\', \'model_dir\', \'n_classes\', \'weight_column\', \'label_vocabulary\', \'optimizer\', \'activation_fn\', \'dropout\', \'input_layer_partitioner\', \'config\', \'warm_start_from\', \'loss_reduction\', \'batch_norm\'], varargs=None, keywords=None, defaults=[\'None\', \'2\', \'None\', \'None\', \'Adagrad\', \'\', \'None\', \'None\', \'None\', \'None\', \'weighted_sum\', \'False\'], "
   }
   member_method {
     name: "eval_dir"
diff --git a/tensorflow/tools/api/golden/tensorflow.estimator.-d-n-n-linear-combined-classifier.pbtxt b/tensorflow/tools/api/golden/tensorflow.estimator.-d-n-n-linear-combined-classifier.pbtxt
index 67e4ee02d0581207e7dd316196aeb782930e7602..9c1c072124083006a1dd8e04526755dd980ba85a 100644
--- a/tensorflow/tools/api/golden/tensorflow.estimator.-d-n-n-linear-combined-classifier.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.estimator.-d-n-n-linear-combined-classifier.pbtxt
@@ -21,7 +21,7 @@ tf_class {
   }
   member_method {
     name: "__init__"
-    argspec: "args=[\'self\', \'model_dir\', \'linear_feature_columns\', \'linear_optimizer\', \'dnn_feature_columns\', \'dnn_optimizer\', \'dnn_hidden_units\', \'dnn_activation_fn\', \'dnn_dropout\', \'n_classes\', \'weight_column\', \'label_vocabulary\', \'input_layer_partitioner\', \'config\', \'warm_start_from\', \'loss_reduction\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'Ftrl\', \'None\', \'Adagrad\', \'None\', \'\', \'None\', \'2\', \'None\', \'None\', \'None\', \'None\', \'None\', \'weighted_sum\'], "
+    argspec: "args=[\'self\', \'model_dir\', \'linear_feature_columns\', \'linear_optimizer\', \'dnn_feature_columns\', \'dnn_optimizer\', \'dnn_hidden_units\', \'dnn_activation_fn\', \'dnn_dropout\', \'n_classes\', \'weight_column\', \'label_vocabulary\', \'input_layer_partitioner\', \'config\', \'warm_start_from\', \'loss_reduction\', \'batch_norm\', \'linear_sparse_combiner\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'Ftrl\', \'None\', \'Adagrad\', \'None\', \'\', \'None\', \'2\', \'None\', \'None\', \'None\', \'None\', \'None\', \'weighted_sum\', \'False\', \'sum\'], "
   }
   member_method {
     name: "eval_dir"
diff --git a/tensorflow/tools/api/golden/tensorflow.estimator.-d-n-n-linear-combined-regressor.pbtxt b/tensorflow/tools/api/golden/tensorflow.estimator.-d-n-n-linear-combined-regressor.pbtxt
index e1289b975e721e94f4a63889f3e0b76b0db23d81..7391d4b07a7e79541091b94fe4a9f38f42d6f68a 100644
--- a/tensorflow/tools/api/golden/tensorflow.estimator.-d-n-n-linear-combined-regressor.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.estimator.-d-n-n-linear-combined-regressor.pbtxt
@@ -21,7 +21,7 @@ tf_class {
   }
   member_method {
     name: "__init__"
-    argspec: "args=[\'self\', \'model_dir\', \'linear_feature_columns\', \'linear_optimizer\', \'dnn_feature_columns\', \'dnn_optimizer\', \'dnn_hidden_units\', \'dnn_activation_fn\', \'dnn_dropout\', \'label_dimension\', \'weight_column\', \'input_layer_partitioner\', \'config\', \'warm_start_from\', \'loss_reduction\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'Ftrl\', \'None\', \'Adagrad\', \'None\', \'\', \'None\', \'1\', \'None\', \'None\', \'None\', \'None\', \'weighted_sum\'], "
+    argspec: "args=[\'self\', \'model_dir\', \'linear_feature_columns\', \'linear_optimizer\', \'dnn_feature_columns\', \'dnn_optimizer\', \'dnn_hidden_units\', \'dnn_activation_fn\', \'dnn_dropout\', \'label_dimension\', \'weight_column\', \'input_layer_partitioner\', \'config\', \'warm_start_from\', \'loss_reduction\', \'batch_norm\', \'linear_sparse_combiner\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'Ftrl\', \'None\', \'Adagrad\', \'None\', \'\', \'None\', \'1\', \'None\', \'None\', \'None\', \'None\', \'weighted_sum\', \'False\', \'sum\'], "
   }
   member_method {
     name: "eval_dir"
diff --git a/tensorflow/tools/api/golden/tensorflow.estimator.-d-n-n-regressor.pbtxt b/tensorflow/tools/api/golden/tensorflow.estimator.-d-n-n-regressor.pbtxt
index d030b2f51f019ecc179a09b76c4484e60ada9dd0..f50e375f7cd392567f5c87536c95eb1f6809bc97 100644
--- a/tensorflow/tools/api/golden/tensorflow.estimator.-d-n-n-regressor.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.estimator.-d-n-n-regressor.pbtxt
@@ -21,7 +21,7 @@ tf_class {
   }
   member_method {
     name: "__init__"
-    argspec: "args=[\'self\', \'hidden_units\', \'feature_columns\', \'model_dir\', \'label_dimension\', \'weight_column\', \'optimizer\', \'activation_fn\', \'dropout\', \'input_layer_partitioner\', \'config\', \'warm_start_from\', \'loss_reduction\'], varargs=None, keywords=None, defaults=[\'None\', \'1\', \'None\', \'Adagrad\', \'\', \'None\', \'None\', \'None\', \'None\', \'weighted_sum\'], "
+    argspec: "args=[\'self\', \'hidden_units\', \'feature_columns\', \'model_dir\', \'label_dimension\', \'weight_column\', \'optimizer\', \'activation_fn\', \'dropout\', \'input_layer_partitioner\', \'config\', \'warm_start_from\', \'loss_reduction\', \'batch_norm\'], varargs=None, keywords=None, defaults=[\'None\', \'1\', \'None\', \'Adagrad\', \'\', \'None\', \'None\', \'None\', \'None\', \'weighted_sum\', \'False\'], "
   }
   member_method {
     name: "eval_dir"
diff --git a/tensorflow/tools/api/golden/tensorflow.estimator.-linear-classifier.pbtxt b/tensorflow/tools/api/golden/tensorflow.estimator.-linear-classifier.pbtxt
index cb578759eee2ed43465195a8c4e8760443a60b71..154f171e89571a43a3f905094a1dbd41cbb000d3 100644
--- a/tensorflow/tools/api/golden/tensorflow.estimator.-linear-classifier.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.estimator.-linear-classifier.pbtxt
@@ -21,7 +21,7 @@ tf_class {
   }
   member_method {
     name: "__init__"
-    argspec: "args=[\'self\', \'feature_columns\', \'model_dir\', \'n_classes\', \'weight_column\', \'label_vocabulary\', \'optimizer\', \'config\', \'partitioner\', \'warm_start_from\', \'loss_reduction\'], varargs=None, keywords=None, defaults=[\'None\', \'2\', \'None\', \'None\', \'Ftrl\', \'None\', \'None\', \'None\', \'weighted_sum\'], "
+    argspec: "args=[\'self\', \'feature_columns\', \'model_dir\', \'n_classes\', \'weight_column\', \'label_vocabulary\', \'optimizer\', \'config\', \'partitioner\', \'warm_start_from\', \'loss_reduction\', \'sparse_combiner\'], varargs=None, keywords=None, defaults=[\'None\', \'2\', \'None\', \'None\', \'Ftrl\', \'None\', \'None\', \'None\', \'weighted_sum\', \'sum\'], "
   }
   member_method {
     name: "eval_dir"
diff --git a/tensorflow/tools/api/golden/tensorflow.estimator.-linear-regressor.pbtxt b/tensorflow/tools/api/golden/tensorflow.estimator.-linear-regressor.pbtxt
index fcd01bb663c7af22791c3855e6da22d93c667f84..4d46d1e6b68758bf634f9b0f82c279fdfa91a0b8 100644
--- a/tensorflow/tools/api/golden/tensorflow.estimator.-linear-regressor.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.estimator.-linear-regressor.pbtxt
@@ -21,7 +21,7 @@ tf_class {
   }
   member_method {
     name: "__init__"
-    argspec: "args=[\'self\', \'feature_columns\', \'model_dir\', \'label_dimension\', \'weight_column\', \'optimizer\', \'config\', \'partitioner\', \'warm_start_from\', \'loss_reduction\'], varargs=None, keywords=None, defaults=[\'None\', \'1\', \'None\', \'Ftrl\', \'None\', \'None\', \'None\', \'weighted_sum\'], "
+    argspec: "args=[\'self\', \'feature_columns\', \'model_dir\', \'label_dimension\', \'weight_column\', \'optimizer\', \'config\', \'partitioner\', \'warm_start_from\', \'loss_reduction\', \'sparse_combiner\'], varargs=None, keywords=None, defaults=[\'None\', \'1\', \'None\', \'Ftrl\', \'None\', \'None\', \'None\', \'weighted_sum\', \'sum\'], "
   }
   member_method {
     name: "eval_dir"
diff --git a/tensorflow/tools/api/golden/tensorflow.image.pbtxt b/tensorflow/tools/api/golden/tensorflow.image.pbtxt
index e89b4dbffdfe85f471fb1dd1b976cc701d526c64..6ec3aba77586a9ffffd1e4375bf58394a118ea82 100644
--- a/tensorflow/tools/api/golden/tensorflow.image.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.image.pbtxt
@@ -120,6 +120,10 @@ tf_module {
     name: "non_max_suppression"
     argspec: "args=[\'boxes\', \'scores\', \'max_output_size\', \'iou_threshold\', \'score_threshold\', \'name\'], varargs=None, keywords=None, defaults=[\'0.5\', \'-inf\', \'None\'], "
   }
+  member_method {
+    name: "non_max_suppression_overlaps"
+    argspec: "args=[\'overlaps\', \'scores\', \'max_output_size\', \'overlap_threshold\', \'score_threshold\', \'name\'], varargs=None, keywords=None, defaults=[\'0.5\', \'-inf\', \'None\'], "
+  }
   member_method {
     name: "pad_to_bounding_box"
     argspec: "args=[\'image\', \'offset_height\', \'offset_width\', \'target_height\', \'target_width\'], varargs=None, keywords=None, defaults=None"
diff --git a/tensorflow/tools/api/golden/tensorflow.initializers.variance_scaling.pbtxt b/tensorflow/tools/api/golden/tensorflow.initializers.variance_scaling.pbtxt
index a6b6e5eceb62654c9ad567a361f7558a2865e57a..86340913e2506c96499aae05a3ed0d5273c93bba 100644
--- a/tensorflow/tools/api/golden/tensorflow.initializers.variance_scaling.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.initializers.variance_scaling.pbtxt
@@ -5,7 +5,7 @@ tf_class {
   is_instance: ""
   member_method {
     name: "__init__"
-    argspec: "args=[\'self\', \'scale\', \'mode\', \'distribution\', \'seed\', \'dtype\'], varargs=None, keywords=None, defaults=[\'1.0\', \'fan_in\', \'normal\', \'None\', \"\"], "
+    argspec: "args=[\'self\', \'scale\', \'mode\', \'distribution\', \'seed\', \'dtype\'], varargs=None, keywords=None, defaults=[\'1.0\', \'fan_in\', \'truncated_normal\', \'None\', \"\"], "
   }
   member_method {
     name: "from_config"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.-model.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.-model.pbtxt
index 11cdd6f0b5e48f5835385fdd4e3e5144fb7d5166..40e82b18b68f9e8353dcb04f76ebb36446d3ab3f 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.-model.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.-model.pbtxt
@@ -119,7 +119,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.-sequential.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.-sequential.pbtxt
index 4afad3e4df308d412a1c18dea3b4e99aa1d2c84f..8295905975f9be76e3608c111b0e7cbe1f152a2b 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.-sequential.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.-sequential.pbtxt
@@ -124,7 +124,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.callbacks.-early-stopping.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.callbacks.-early-stopping.pbtxt
index 7b0ad85eaac5b83835a9e1c4b152e38e7051a2f6..f71292856cd29b2e52194bec8a586686fbfad667 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.callbacks.-early-stopping.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.callbacks.-early-stopping.pbtxt
@@ -5,7 +5,7 @@ tf_class {
   is_instance: ""
   member_method {
     name: "__init__"
-    argspec: "args=[\'self\', \'monitor\', \'min_delta\', \'patience\', \'verbose\', \'mode\'], varargs=None, keywords=None, defaults=[\'val_loss\', \'0\', \'0\', \'0\', \'auto\'], "
+    argspec: "args=[\'self\', \'monitor\', \'min_delta\', \'patience\', \'verbose\', \'mode\', \'baseline\'], varargs=None, keywords=None, defaults=[\'val_loss\', \'0\', \'0\', \'0\', \'auto\', \'None\'], "
   }
   member_method {
     name: "on_batch_begin"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.initializers.-variance-scaling.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.initializers.-variance-scaling.pbtxt
index 32a6f6ee88815b3dc70e9cca855f73099554953b..03f4064b9ef5093044a9cbb897043d643cf7f83e 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.initializers.-variance-scaling.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.initializers.-variance-scaling.pbtxt
@@ -5,7 +5,7 @@ tf_class {
   is_instance: ""
   member_method {
     name: "__init__"
-    argspec: "args=[\'self\', \'scale\', \'mode\', \'distribution\', \'seed\', \'dtype\'], varargs=None, keywords=None, defaults=[\'1.0\', \'fan_in\', \'normal\', \'None\', \"\"], "
+    argspec: "args=[\'self\', \'scale\', \'mode\', \'distribution\', \'seed\', \'dtype\'], varargs=None, keywords=None, defaults=[\'1.0\', \'fan_in\', \'truncated_normal\', \'None\', \"\"], "
   }
   member_method {
     name: "from_config"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.initializers.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.initializers.pbtxt
index 14a667870d3118e48bfac03eee9accb3d48a72ce..8645e5430295dff0a5b7c715b03860fb7734e7f1 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.initializers.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.initializers.pbtxt
@@ -90,11 +90,11 @@ tf_module {
   }
   member_method {
     name: "glorot_normal"
-    argspec: "args=[\'seed\'], varargs=None, keywords=None, defaults=[\'None\'], "
+    argspec: "args=[\'seed\', \'dtype\'], varargs=None, keywords=None, defaults=[\'None\', \"\"], "
   }
   member_method {
     name: "glorot_uniform"
-    argspec: "args=[\'seed\'], varargs=None, keywords=None, defaults=[\'None\'], "
+    argspec: "args=[\'seed\', \'dtype\'], varargs=None, keywords=None, defaults=[\'None\', \"\"], "
   }
   member_method {
     name: "he_normal"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-activation.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-activation.pbtxt
index 2bf973debb175d27bb80e627d7ccbb41b567020d..86e328888e596852caf9ad1020dfdedb71864969 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-activation.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-activation.pbtxt
@@ -98,7 +98,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-activity-regularization.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-activity-regularization.pbtxt
index 03f20e72c2a325cec000cf4a5cfc0f1bbf255c8f..b0ed54578109c6ae8d5bc2c9f5c978b562a9cc84 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-activity-regularization.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-activity-regularization.pbtxt
@@ -98,7 +98,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-add.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-add.pbtxt
index 4b46b8d15afb0a2f636962b762e1808312c2f7c3..42f98ed03d426d60cabeb0b533311d41eb378285 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-add.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-add.pbtxt
@@ -99,7 +99,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-alpha-dropout.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-alpha-dropout.pbtxt
index d8a1c76fd07634ef413152020a397897f2d5b97c..000898a4be928e4e64b4072ef3170b6fbc930bdf 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-alpha-dropout.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-alpha-dropout.pbtxt
@@ -98,7 +98,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-average-pooling1-d.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-average-pooling1-d.pbtxt
index 622926bc4b8b2430ee1ab936665acb5744155e0d..380b49f99ce6e62770a9516ba81db99f194c5b37 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-average-pooling1-d.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-average-pooling1-d.pbtxt
@@ -99,7 +99,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-average-pooling2-d.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-average-pooling2-d.pbtxt
index 82100d8e09c8e95730993527293d2b72ce69f1d4..82db5e6137639e516f6df6f0e130e73be516c9b8 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-average-pooling2-d.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-average-pooling2-d.pbtxt
@@ -99,7 +99,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-average-pooling3-d.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-average-pooling3-d.pbtxt
index 408061077cdeab2f8fd08c7e972744e5ee383f52..b6ff688ec36f8c47b2ac9694fb84350818be25c5 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-average-pooling3-d.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-average-pooling3-d.pbtxt
@@ -99,7 +99,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-average.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-average.pbtxt
index a3c80311043eeb95b06855f662a5e3d344803ba3..b41290f8b067397bf6678d9e98ac53f28a05a3fc 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-average.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-average.pbtxt
@@ -99,7 +99,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-avg-pool1-d.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-avg-pool1-d.pbtxt
index e2dfaca29f86bd9d91d524ec337afad81e7f2da3..88a033e61f42e2fb02b08968ff001ea21195972a 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-avg-pool1-d.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-avg-pool1-d.pbtxt
@@ -99,7 +99,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-avg-pool2-d.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-avg-pool2-d.pbtxt
index 4f068d2066a450bab77becc85a33662b78ad03e2..c1b9b96044ed2e057b8e86dda59ee7f7166cfd43 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-avg-pool2-d.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-avg-pool2-d.pbtxt
@@ -99,7 +99,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-avg-pool3-d.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-avg-pool3-d.pbtxt
index b8c261a74364e9bb6bf8f6c7463993fbff5e9552..f59f7727a3eaeb4fa5631cb1b42901ea6d39b06b 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-avg-pool3-d.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-avg-pool3-d.pbtxt
@@ -99,7 +99,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-batch-normalization.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-batch-normalization.pbtxt
index 4ccd6cace650e2efd1583c75f6639c8598bb8f20..7d3744ed92636a972bae2f9b62a6b2da8f91d106 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-batch-normalization.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-batch-normalization.pbtxt
@@ -98,7 +98,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-bidirectional.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-bidirectional.pbtxt
index 2790e5fd850c24bd3e94cd15a6e079e1c9f79868..3fd4ccdab2573964c2f3192d503e9fb15f442dc5 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-bidirectional.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-bidirectional.pbtxt
@@ -107,7 +107,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-concatenate.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-concatenate.pbtxt
index b1326bd0e6054b2a3fd36e7ad42cd3d4a0cad8dc..ba21b50be41f3adc735b3350bdf9dbeae3c2e358 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-concatenate.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-concatenate.pbtxt
@@ -99,7 +99,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-conv-l-s-t-m2-d.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-conv-l-s-t-m2-d.pbtxt
index e3ac3dbf28da731e14640d5f464547d62391a28f..46f9fa2bbbbe3cfff3aade33c5ebdec92bc70ef0 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-conv-l-s-t-m2-d.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-conv-l-s-t-m2-d.pbtxt
@@ -188,7 +188,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-conv1-d.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-conv1-d.pbtxt
index 1117a695a395f495d988464bbf59d4b8e01877e6..c3ad326589d2822bc5dd381d78216b25f5fb6f95 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-conv1-d.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-conv1-d.pbtxt
@@ -99,7 +99,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-conv2-d-transpose.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-conv2-d-transpose.pbtxt
index b9de1421428dcf61b988df343a22996cfb8fecef..fd9eb43066be580a7df57aeb717b59569c9bba61 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-conv2-d-transpose.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-conv2-d-transpose.pbtxt
@@ -100,7 +100,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-conv2-d.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-conv2-d.pbtxt
index deb535e06e06008a17b80c8e13d8f01ad1535059..40d61688f29a81e873a26c8a5eb823d679320ed6 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-conv2-d.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-conv2-d.pbtxt
@@ -99,7 +99,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-conv3-d-transpose.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-conv3-d-transpose.pbtxt
index 9a9a223fbad11cafd8620110d80b27d5382dd29c..b8c227d7257311578e41abe0a384ed93e6a2866c 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-conv3-d-transpose.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-conv3-d-transpose.pbtxt
@@ -100,7 +100,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-conv3-d.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-conv3-d.pbtxt
index 1c59b0bdf624b09a7454f2d51698951a790f393a..095d35e5749d0113956b04f971e6a8ca1fa277b8 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-conv3-d.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-conv3-d.pbtxt
@@ -99,7 +99,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-convolution1-d.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-convolution1-d.pbtxt
index 30cf5489f4fcd4af3d0bd957fc9c576c57ee2bbd..8f999611982bbfe3c613ef26d93782e299275f19 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-convolution1-d.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-convolution1-d.pbtxt
@@ -99,7 +99,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-convolution2-d-transpose.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-convolution2-d-transpose.pbtxt
index 0ec69508d5a1992b46d1a7c65255cfb5408ab439..96d522a016aedba01032a1c05a69511cb03d19af 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-convolution2-d-transpose.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-convolution2-d-transpose.pbtxt
@@ -100,7 +100,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-convolution2-d.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-convolution2-d.pbtxt
index 4cd8928403c98abad85bc1349a29148c73003c9d..de2824dab4526d90eebf9cef16710cadf82f4850 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-convolution2-d.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-convolution2-d.pbtxt
@@ -99,7 +99,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-convolution3-d-transpose.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-convolution3-d-transpose.pbtxt
index 4b4912496deac2a79a5b0ea3d1ca0f8fa625301a..1d563241d8f0d93bcd19a319eb8383f4bcdf4388 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-convolution3-d-transpose.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-convolution3-d-transpose.pbtxt
@@ -100,7 +100,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-convolution3-d.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-convolution3-d.pbtxt
index d0ad9cf56702e585e31a79de0f93d9efd48ed484..c87e52c53796f0743365a9d8780decf237bba070 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-convolution3-d.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-convolution3-d.pbtxt
@@ -99,7 +99,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-cropping1-d.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-cropping1-d.pbtxt
index 98cff95a7fe9d4e58cf883502df08c58c651cd76..dccf5523e3870b6c1ce0de70c648ab47968a105f 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-cropping1-d.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-cropping1-d.pbtxt
@@ -98,7 +98,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-cropping2-d.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-cropping2-d.pbtxt
index 2357498b46376ef13de102944b69931a9e7d3584..7ac4116d922eea51e5a7e7fe3d02ad919300c459 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-cropping2-d.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-cropping2-d.pbtxt
@@ -98,7 +98,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-cropping3-d.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-cropping3-d.pbtxt
index 3324cbff304c5106360f3f3d3d608a528fa5fc31..024f72705de1e76866a8132246884dffb0c4e72a 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-cropping3-d.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-cropping3-d.pbtxt
@@ -98,7 +98,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-cu-d-n-n-g-r-u.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-cu-d-n-n-g-r-u.pbtxt
index 6c81823654b78a936cded4a1d5a6f54e02dc7fc9..4e0233331bd47e86e8a4df2f84b5392517fbf884 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-cu-d-n-n-g-r-u.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-cu-d-n-n-g-r-u.pbtxt
@@ -108,7 +108,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-cu-d-n-n-l-s-t-m.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-cu-d-n-n-l-s-t-m.pbtxt
index 487e04fd0790cb39ef6aee8d0498b3aae6726084..32d46ce8f3deff6077eaf5a1a8cf7ba64478d9f4 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-cu-d-n-n-l-s-t-m.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-cu-d-n-n-l-s-t-m.pbtxt
@@ -108,7 +108,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-dense.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-dense.pbtxt
index 137e7cced4e8113dd6a54a837e08cfd5af35c94d..858486c725c3be5ecae2a02d0d3134ebeb113ce1 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-dense.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-dense.pbtxt
@@ -98,7 +98,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-depthwise-conv2-d.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-depthwise-conv2-d.pbtxt
index 7161665d2550c1cc3aff1c28f9d7676276b62303..f65d7509262bfeb148588e069c08961058a3fa74 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-depthwise-conv2-d.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-depthwise-conv2-d.pbtxt
@@ -100,7 +100,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-dot.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-dot.pbtxt
index 24affa248121bcb1e1a947417a95ad4f5ba55ab2..2e71ef503d54927edbb3e1ef6c701ac845883e46 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-dot.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-dot.pbtxt
@@ -99,7 +99,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-dropout.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-dropout.pbtxt
index 7ba19a42695da37b4ad43cdde2c0d4978fd0a1eb..42533bcd21b28a0acf183db195a6b5c1848a5d91 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-dropout.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-dropout.pbtxt
@@ -98,7 +98,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-e-l-u.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-e-l-u.pbtxt
index 503aa9162c3a78e9bb42ce16af98451441adbbb7..b5df16941792a29d72f2ee709993b007d342d2d0 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-e-l-u.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-e-l-u.pbtxt
@@ -98,7 +98,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-embedding.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-embedding.pbtxt
index 1737e590a29c5777b5eca2b4cb23081aa8ece738..0ea17919a9bb13ffdedd60ce618bca23dd52712f 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-embedding.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-embedding.pbtxt
@@ -98,7 +98,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-flatten.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-flatten.pbtxt
index 021d024dc2150a75532ea7597d85f36efd2a3cf2..a33248bc005a73d0be679cd62150d6019b475305 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-flatten.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-flatten.pbtxt
@@ -98,7 +98,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-g-r-u-cell.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-g-r-u-cell.pbtxt
index 65387008bf3f78e404d8d8bbd7bb8cd3789bf256..4ba21a25cda83122fbced7fed76d4b1ae28cb4c8 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-g-r-u-cell.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-g-r-u-cell.pbtxt
@@ -98,7 +98,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-g-r-u.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-g-r-u.pbtxt
index 4f791acf0585c95d6c0f1d5ea48e607f9a05188d..a7a570418e0a78873237c1c8cefe36a212e4c9af 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-g-r-u.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-g-r-u.pbtxt
@@ -171,7 +171,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-gaussian-dropout.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-gaussian-dropout.pbtxt
index abc30e54e0630a2d7b4de6074445e155e0ac2782..763bc231136908d469b7f942aec94f6248d2e2d4 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-gaussian-dropout.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-gaussian-dropout.pbtxt
@@ -98,7 +98,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-gaussian-noise.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-gaussian-noise.pbtxt
index 20791bb448d17788ea4aebe4900169a70a9703d6..3c50a3d7f28809b2b810b52951207e48f9f50e34 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-gaussian-noise.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-gaussian-noise.pbtxt
@@ -98,7 +98,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-global-average-pooling1-d.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-global-average-pooling1-d.pbtxt
index 449a91d8735c59f563360307cdb35c5a30344d82..ac78bdafada8c157efd4ab8746be15726eb0bc24 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-global-average-pooling1-d.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-global-average-pooling1-d.pbtxt
@@ -99,7 +99,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-global-average-pooling2-d.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-global-average-pooling2-d.pbtxt
index bb361e129728ddd42c21144937efbc617d98ba30..275282d9d2b1753cf0189b605f921bb039ef5f3c 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-global-average-pooling2-d.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-global-average-pooling2-d.pbtxt
@@ -99,7 +99,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-global-average-pooling3-d.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-global-average-pooling3-d.pbtxt
index e564bf3216104a902fb6cfbe65b1e2b6dafc2524..0e31e6058bd6036a5fb4422335917718f4f82851 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-global-average-pooling3-d.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-global-average-pooling3-d.pbtxt
@@ -99,7 +99,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-global-avg-pool1-d.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-global-avg-pool1-d.pbtxt
index 4cb9cc3ec84d679b78465e43caa5a257466d5676..aacd0b1791dda5babb6eef5d87a1335c8d519b08 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-global-avg-pool1-d.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-global-avg-pool1-d.pbtxt
@@ -99,7 +99,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-global-avg-pool2-d.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-global-avg-pool2-d.pbtxt
index 5ed52b88ae3e2dd25b560206db404952034a04cd..c23654866341818aeb804cfb71dae052049e3f25 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-global-avg-pool2-d.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-global-avg-pool2-d.pbtxt
@@ -99,7 +99,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-global-avg-pool3-d.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-global-avg-pool3-d.pbtxt
index f4559d29d75ef7cd8fcbdeac0a1a2c9e633246bc..6b9c0290aac35d80c7f87acfc44479c57623a645 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-global-avg-pool3-d.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-global-avg-pool3-d.pbtxt
@@ -99,7 +99,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-global-max-pool1-d.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-global-max-pool1-d.pbtxt
index 64e2d061e26997365c461113d3ea15140fef64dd..0d7b2211e6cd35ca331b4a1068f237e7ca07f70c 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-global-max-pool1-d.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-global-max-pool1-d.pbtxt
@@ -99,7 +99,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-global-max-pool2-d.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-global-max-pool2-d.pbtxt
index 3372ad645388beb54f7ed9e3715449facba07f87..d080ad6aedbd5183da890cd63f5f18453d5d476a 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-global-max-pool2-d.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-global-max-pool2-d.pbtxt
@@ -99,7 +99,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-global-max-pool3-d.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-global-max-pool3-d.pbtxt
index 08a6860bcd7d9a260e44af87c51796a9cc2af379..fcb0a109da208ff5bd20447ddced9816a42af311 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-global-max-pool3-d.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-global-max-pool3-d.pbtxt
@@ -99,7 +99,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-global-max-pooling1-d.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-global-max-pooling1-d.pbtxt
index 22c9eab64fde41e1199ecbb1b8b03939653ecd00..1d0e22abd0d8732182881c43ee79400642cef24b 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-global-max-pooling1-d.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-global-max-pooling1-d.pbtxt
@@ -99,7 +99,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-global-max-pooling2-d.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-global-max-pooling2-d.pbtxt
index 74c405ba9b1b465f89c4fef43020181a1a7f3d31..653c9f547bc888a8fec87137f7d495141d4f8599 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-global-max-pooling2-d.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-global-max-pooling2-d.pbtxt
@@ -99,7 +99,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-global-max-pooling3-d.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-global-max-pooling3-d.pbtxt
index 39f6f981931296eb6d31eb6580f93b479ff64ce6..cdbaf82cf6746e878619647439d2256f6e2c4aa3 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-global-max-pooling3-d.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-global-max-pooling3-d.pbtxt
@@ -99,7 +99,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-input-layer.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-input-layer.pbtxt
index 7b25e80b6b7653c5e76bf176b54110b1aabaf5ea..230c5e903438b0a75edf80f0f5c8706987c66a78 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-input-layer.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-input-layer.pbtxt
@@ -98,7 +98,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-l-s-t-m-cell.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-l-s-t-m-cell.pbtxt
index 3619b8bfc44373ba6b8e306b020ac63d4b498573..511456e740837455818ff3f9be270daed03f334f 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-l-s-t-m-cell.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-l-s-t-m-cell.pbtxt
@@ -98,7 +98,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-l-s-t-m.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-l-s-t-m.pbtxt
index 8ef3d71dd82efc79e333770d4a7a7c8aee1a4202..4a3492ebd652e5ab8f0faf8a1583480abc80fba7 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-l-s-t-m.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-l-s-t-m.pbtxt
@@ -171,7 +171,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-lambda.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-lambda.pbtxt
index ecbaa9ce2c76bf3d2964a6c79c96c4d67cc3b80e..5d05cf689fb399d6630f68b09fd123d2d968786b 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-lambda.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-lambda.pbtxt
@@ -98,7 +98,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-layer.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-layer.pbtxt
index 9b90db1e5e56d1e5749669bba8dba1cdbd45bb55..7efa29be77c075a29784d8cd3ebfcd871bc9aa0c 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-layer.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-layer.pbtxt
@@ -97,7 +97,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-leaky-re-l-u.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-leaky-re-l-u.pbtxt
index 3c60eaab7f1df15331004685676d74943d5d538f..0ca8e0b52c4a81c4ff3b756aa6c24b47a664f999 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-leaky-re-l-u.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-leaky-re-l-u.pbtxt
@@ -98,7 +98,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-locally-connected1-d.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-locally-connected1-d.pbtxt
index 3dac1ff342ac1b7f984e9af5a6028ef71da701df..f754fa1da85692c28f31a76bbfa987b3c4c30731 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-locally-connected1-d.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-locally-connected1-d.pbtxt
@@ -98,7 +98,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-locally-connected2-d.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-locally-connected2-d.pbtxt
index 7f1b5db4d34f706f2107ef43ab9c5acf67dac9f6..c9516b8f07d0b6a818bf99d45499d161c2a5cffd 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-locally-connected2-d.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-locally-connected2-d.pbtxt
@@ -98,7 +98,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-masking.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-masking.pbtxt
index b3e31000f3bca0821377d70b1d88a20aa8f8e4ef..850ecff9743b5f5048bb81c5a15b0a4be6b4d0ce 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-masking.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-masking.pbtxt
@@ -98,7 +98,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-max-pool1-d.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-max-pool1-d.pbtxt
index bbd9d1b0dc075bb9241f240b423933db20b38b75..7c69e31f9af9bbd221882d160fa4206997ec3b08 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-max-pool1-d.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-max-pool1-d.pbtxt
@@ -99,7 +99,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-max-pool2-d.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-max-pool2-d.pbtxt
index fe72beea802d12b996948b00436b274ee7e83177..fba42642d7c701688c2bd274cf97e077e7ff571c 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-max-pool2-d.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-max-pool2-d.pbtxt
@@ -99,7 +99,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-max-pool3-d.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-max-pool3-d.pbtxt
index e9bf57b2b0e60376a28c0abfc16fba393df3e73c..9c277411ea5ce26df9c033ada773ad2e45292cb1 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-max-pool3-d.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-max-pool3-d.pbtxt
@@ -99,7 +99,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-max-pooling1-d.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-max-pooling1-d.pbtxt
index 0eecc58a2b6a2846a2c92502cc23bd328f8b5193..7c2f6ccc8a98017aba014ab6a7896e0a4bf40324 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-max-pooling1-d.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-max-pooling1-d.pbtxt
@@ -99,7 +99,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-max-pooling2-d.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-max-pooling2-d.pbtxt
index 96785a7d8559611a19b7f36216dbf0f8a3e39e61..802178dba63d66cca1629bcb7bef0f578c9a6659 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-max-pooling2-d.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-max-pooling2-d.pbtxt
@@ -99,7 +99,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-max-pooling3-d.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-max-pooling3-d.pbtxt
index 42c46cccb37b1ab7ece7760e6858b2180ea833b9..e870dfe9ade75da367f87a4b54d38ba4274bab2e 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-max-pooling3-d.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-max-pooling3-d.pbtxt
@@ -99,7 +99,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-maximum.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-maximum.pbtxt
index ac816f68d492cbfc5503c057a869e3e981de9190..c1337ce0cbac2d1e0e011f5309bfb2722960d3b2 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-maximum.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-maximum.pbtxt
@@ -99,7 +99,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-minimum.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-minimum.pbtxt
index 56e32e9d3690a92c3f6e41bf2b5164c6bf62f443..ed27a62765d5670802d4593b3e648e3f65eaf926 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-minimum.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-minimum.pbtxt
@@ -99,7 +99,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-multiply.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-multiply.pbtxt
index 9ae99563e9a1b3b0700116ed88c13f94fafe1658..b9f05cb3e56f89cb02e1a74c3ec0d362ea27f2bf 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-multiply.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-multiply.pbtxt
@@ -99,7 +99,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-p-re-l-u.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-p-re-l-u.pbtxt
index 815f3bc2d142069adb4e418a4dc6ef82d683373f..336d9f76fb1e6215b763b5064cd6be68d4d0d5a0 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-p-re-l-u.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-p-re-l-u.pbtxt
@@ -98,7 +98,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-permute.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-permute.pbtxt
index e704992b4a18f6bdbd9474af2ee59ea81534d80a..46282217e01e8a137d9fc564f0e3544602d93de4 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-permute.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-permute.pbtxt
@@ -98,7 +98,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-r-n-n.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-r-n-n.pbtxt
index b3a58fa11eda61baa5c932bcc04fdca7459a215f..42cd7e87eebdd969f002d8bcd0dca101168c58e0 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-r-n-n.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-r-n-n.pbtxt
@@ -102,7 +102,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-re-l-u.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-re-l-u.pbtxt
index f3a96ab895dc9dbf8e2362dbcbfdccdf6af749ec..c00fa79adfbe5b986b481f6c9567bafbf3abc1ae 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-re-l-u.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-re-l-u.pbtxt
@@ -98,7 +98,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-repeat-vector.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-repeat-vector.pbtxt
index 78f464583b4e8083f4cdd1a8c6b9f377645cd562..9f094a877a3a47ff89a022db563803f5f391ff2a 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-repeat-vector.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-repeat-vector.pbtxt
@@ -98,7 +98,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-reshape.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-reshape.pbtxt
index 222344fd0497afe9a32d1d05ec37aa160479d88a..2f519a24385ac4e147798ed3e96101cff23e19aa 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-reshape.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-reshape.pbtxt
@@ -98,7 +98,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-separable-conv1-d.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-separable-conv1-d.pbtxt
index 55fddf576cac6afabe984cd51e2ddbf112a55d25..6b93116ba02c2b7e9c5bdf79ddfa1f93050062a4 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-separable-conv1-d.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-separable-conv1-d.pbtxt
@@ -100,7 +100,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-separable-conv2-d.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-separable-conv2-d.pbtxt
index 96314ce49849a50ccc6b968b50c98ddae74c6c70..fd17115e2733d561bff1d53d62d32458b03dc65b 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-separable-conv2-d.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-separable-conv2-d.pbtxt
@@ -100,7 +100,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-separable-convolution1-d.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-separable-convolution1-d.pbtxt
index 88bdf9956603c590940e3ef857765586df7e91d7..4b37a94478857ac8550ea0c4f464058c68770047 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-separable-convolution1-d.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-separable-convolution1-d.pbtxt
@@ -100,7 +100,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-separable-convolution2-d.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-separable-convolution2-d.pbtxt
index 6eeea7a8d1312ada423206378b4c6ee079ffdd73..5bdadca74aeb963adef4999b7e758add1aec4681 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-separable-convolution2-d.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-separable-convolution2-d.pbtxt
@@ -100,7 +100,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-simple-r-n-n-cell.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-simple-r-n-n-cell.pbtxt
index 3050d46249003716eb0778104b729ee9cb52b34f..9dfda96fc81572d70d76ba767b69ee2e41f017ee 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-simple-r-n-n-cell.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-simple-r-n-n-cell.pbtxt
@@ -98,7 +98,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-simple-r-n-n.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-simple-r-n-n.pbtxt
index dda4c9358ba5faa084ad2e6cf75ff83b6a7b2b20..7b7684ccd27a1d4c3fabf56c2669f77095f501ef 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-simple-r-n-n.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-simple-r-n-n.pbtxt
@@ -159,7 +159,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-softmax.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-softmax.pbtxt
index cc6275158b67e94c3c39802cc7c0f9e169c8b144..3b15407fca2cf65f7fa31f29b84db52b5c5d1a7a 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-softmax.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-softmax.pbtxt
@@ -98,7 +98,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-spatial-dropout1-d.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-spatial-dropout1-d.pbtxt
index 5eb7e750477b17571ef861305806894dd2b9ac38..6d04415267c9ce21268b9d86a5b078d8f92db93f 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-spatial-dropout1-d.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-spatial-dropout1-d.pbtxt
@@ -99,7 +99,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-spatial-dropout2-d.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-spatial-dropout2-d.pbtxt
index 500cb8c14ead3eeff28d11b72e2300cc471756d2..04950654d55f30bf095167d176b5b2717e72f2cd 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-spatial-dropout2-d.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-spatial-dropout2-d.pbtxt
@@ -99,7 +99,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-spatial-dropout3-d.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-spatial-dropout3-d.pbtxt
index 1113a7634fa98b499175d90ae7da2d3fb9fb1a13..c424e6dcc869f977100e77fdb543983c3ab7e63c 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-spatial-dropout3-d.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-spatial-dropout3-d.pbtxt
@@ -99,7 +99,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-stacked-r-n-n-cells.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-stacked-r-n-n-cells.pbtxt
index c4b9f93561de6a5d8ecc19bbae17831466b51fe6..1160d2840f5ddd2937db53406af9d4d2132a6515 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-stacked-r-n-n-cells.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-stacked-r-n-n-cells.pbtxt
@@ -102,7 +102,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-subtract.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-subtract.pbtxt
index 35ad87ad5d91f1cc5d413b0adc8e9e5d1403726a..740a03367bd69edf797d3ea8616fdde72f6726b7 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-subtract.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-subtract.pbtxt
@@ -99,7 +99,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-thresholded-re-l-u.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-thresholded-re-l-u.pbtxt
index 282c98d79a6e1da46e4d7ea2e5c7228754792f09..a08c583adb4175ff5ee77869c80c6c0204018166 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-thresholded-re-l-u.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-thresholded-re-l-u.pbtxt
@@ -98,7 +98,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-time-distributed.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-time-distributed.pbtxt
index acab93706b29fedc1bf7b48da2f5b6636dea48e5..c1294fed0fcfca9c8607bf3e5d41efd240fd4d45 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-time-distributed.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-time-distributed.pbtxt
@@ -103,7 +103,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-up-sampling1-d.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-up-sampling1-d.pbtxt
index a5ec228a074721775d4ec0369345b5439d84e186..dc401d3ed0fee5b6fb4bb5563941c3461eb592f4 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-up-sampling1-d.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-up-sampling1-d.pbtxt
@@ -98,7 +98,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-up-sampling2-d.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-up-sampling2-d.pbtxt
index d8d8e0bfe95a6cf2ef61cdb344b963df3f21aabb..4b5165ae9793f900fb474affe52b9abaeb64adbd 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-up-sampling2-d.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-up-sampling2-d.pbtxt
@@ -98,7 +98,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-up-sampling3-d.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-up-sampling3-d.pbtxt
index 97d6dc06fb2e883b20540e4496efa5b39a538263..789af15fea8c0d41dd3f0c00e7be3afd6afafecf 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-up-sampling3-d.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-up-sampling3-d.pbtxt
@@ -98,7 +98,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-wrapper.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-wrapper.pbtxt
index ea9bb41b9979de9049397892372f37aafc719a68..0536a7cee7e6dd5878f532854753cebeaa043c21 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-wrapper.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-wrapper.pbtxt
@@ -102,7 +102,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-zero-padding1-d.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-zero-padding1-d.pbtxt
index e6d1d2e089b01c4eb212d01c456f6fa6b850f7de..8915353ec334f28c4ed058b20a506ff102ca1f61 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-zero-padding1-d.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-zero-padding1-d.pbtxt
@@ -98,7 +98,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-zero-padding2-d.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-zero-padding2-d.pbtxt
index f62017305f26519181b1ef86bdd0946d44d16b88..6efb5ef15a133877666decfd1f2b40fad4463469 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-zero-padding2-d.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-zero-padding2-d.pbtxt
@@ -98,7 +98,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.layers.-zero-padding3-d.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.layers.-zero-padding3-d.pbtxt
index 07a1fde5bdc35535ca5d8443a97cb85adc54b14a..4c33c5d0bf800239e2bff4cc874e594b515a8071 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.layers.-zero-padding3-d.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.layers.-zero-padding3-d.pbtxt
@@ -98,7 +98,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.models.-model.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.models.-model.pbtxt
index 62aa929d32b57518abbe924c036062eb7ccd3acf..85f7c2bfedb936d3b21624448cf8875775de918b 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.models.-model.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.models.-model.pbtxt
@@ -119,7 +119,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.keras.models.-sequential.pbtxt b/tensorflow/tools/api/golden/tensorflow.keras.models.-sequential.pbtxt
index 93ecbbce9b17b9ca6157e65bbabd6c36008c3992..5211657414b24ad1971055446a5d021960941275 100644
--- a/tensorflow/tools/api/golden/tensorflow.keras.models.-sequential.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.keras.models.-sequential.pbtxt
@@ -124,7 +124,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'partitioner\', \'use_resource\', \'synchronization\', \'aggregation\', \'getter\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.layers.-average-pooling1-d.pbtxt b/tensorflow/tools/api/golden/tensorflow.layers.-average-pooling1-d.pbtxt
index 11067058d5852669e1672bf3eb8b7c680d0e5dc9..c82e67526b21696a7d56517dc2cb6998882dc7a5 100644
--- a/tensorflow/tools/api/golden/tensorflow.layers.-average-pooling1-d.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.layers.-average-pooling1-d.pbtxt
@@ -109,7 +109,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'use_resource\', \'partitioner\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'use_resource\', \'synchronization\', \'aggregation\', \'partitioner\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.layers.-average-pooling2-d.pbtxt b/tensorflow/tools/api/golden/tensorflow.layers.-average-pooling2-d.pbtxt
index 3259e706d7f7ea4d0348c1ee586c50f5a2c82b39..1d031cb5f8461145127b0f13d77e6b8774f5a0b3 100644
--- a/tensorflow/tools/api/golden/tensorflow.layers.-average-pooling2-d.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.layers.-average-pooling2-d.pbtxt
@@ -109,7 +109,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'use_resource\', \'partitioner\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'use_resource\', \'synchronization\', \'aggregation\', \'partitioner\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.layers.-average-pooling3-d.pbtxt b/tensorflow/tools/api/golden/tensorflow.layers.-average-pooling3-d.pbtxt
index e561f2f415018840420232a97f0ece3f3c60d0d7..a8dda6655df1d06ca77b74f0a992c8fd7e7a357d 100644
--- a/tensorflow/tools/api/golden/tensorflow.layers.-average-pooling3-d.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.layers.-average-pooling3-d.pbtxt
@@ -109,7 +109,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'use_resource\', \'partitioner\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'use_resource\', \'synchronization\', \'aggregation\', \'partitioner\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.layers.-batch-normalization.pbtxt b/tensorflow/tools/api/golden/tensorflow.layers.-batch-normalization.pbtxt
index 3124a35c7852a97e79a3cfe575017484f2f5731f..97f65ed89436bd0b4027bb0cbeb80b6f1419269c 100644
--- a/tensorflow/tools/api/golden/tensorflow.layers.-batch-normalization.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.layers.-batch-normalization.pbtxt
@@ -108,7 +108,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'use_resource\', \'partitioner\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'use_resource\', \'synchronization\', \'aggregation\', \'partitioner\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.layers.-conv1-d.pbtxt b/tensorflow/tools/api/golden/tensorflow.layers.-conv1-d.pbtxt
index b5ec61255ace78c1fa13370727eb5f5084522f4a..ccd9578f0d62bd70ea252ddeac587d59c926b018 100644
--- a/tensorflow/tools/api/golden/tensorflow.layers.-conv1-d.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.layers.-conv1-d.pbtxt
@@ -109,7 +109,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'use_resource\', \'partitioner\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'use_resource\', \'synchronization\', \'aggregation\', \'partitioner\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.layers.-conv2-d-transpose.pbtxt b/tensorflow/tools/api/golden/tensorflow.layers.-conv2-d-transpose.pbtxt
index b2c89ae66f53299289508eef174b5c44a6be2606..9cbb58d721bb49bde562a57728a9ee46968e611e 100644
--- a/tensorflow/tools/api/golden/tensorflow.layers.-conv2-d-transpose.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.layers.-conv2-d-transpose.pbtxt
@@ -110,7 +110,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'use_resource\', \'partitioner\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'use_resource\', \'synchronization\', \'aggregation\', \'partitioner\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.layers.-conv2-d.pbtxt b/tensorflow/tools/api/golden/tensorflow.layers.-conv2-d.pbtxt
index 9e4f4969dc6e1b6a39cf1d25c5e5e6175fa87c7c..c75ea3911e17bc879d140068ef54521effd2824e 100644
--- a/tensorflow/tools/api/golden/tensorflow.layers.-conv2-d.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.layers.-conv2-d.pbtxt
@@ -109,7 +109,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'use_resource\', \'partitioner\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'use_resource\', \'synchronization\', \'aggregation\', \'partitioner\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.layers.-conv3-d-transpose.pbtxt b/tensorflow/tools/api/golden/tensorflow.layers.-conv3-d-transpose.pbtxt
index 9850e6d7659d311c93dabad73d35f2fcd028dd52..5dc834e5141e58d255357e02d7446a06e6e2aa45 100644
--- a/tensorflow/tools/api/golden/tensorflow.layers.-conv3-d-transpose.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.layers.-conv3-d-transpose.pbtxt
@@ -110,7 +110,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'use_resource\', \'partitioner\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'use_resource\', \'synchronization\', \'aggregation\', \'partitioner\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.layers.-conv3-d.pbtxt b/tensorflow/tools/api/golden/tensorflow.layers.-conv3-d.pbtxt
index be113826cc2b9589e1f8bbde896fbcbe183d4d1b..96ab209874ac14d6acf2e8115e7f04fc35c4b2bd 100644
--- a/tensorflow/tools/api/golden/tensorflow.layers.-conv3-d.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.layers.-conv3-d.pbtxt
@@ -109,7 +109,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'use_resource\', \'partitioner\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'use_resource\', \'synchronization\', \'aggregation\', \'partitioner\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.layers.-dense.pbtxt b/tensorflow/tools/api/golden/tensorflow.layers.-dense.pbtxt
index 0d951bf6336ac7b65be57535c1065e5f87a77a0b..7e9656b3525c1d53940b869607616ff414a466cf 100644
--- a/tensorflow/tools/api/golden/tensorflow.layers.-dense.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.layers.-dense.pbtxt
@@ -108,7 +108,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'use_resource\', \'partitioner\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'use_resource\', \'synchronization\', \'aggregation\', \'partitioner\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.layers.-dropout.pbtxt b/tensorflow/tools/api/golden/tensorflow.layers.-dropout.pbtxt
index f1beeed9ef0cb54318249e42b1279680ea117ba8..e9a2269a6e8de1f9a12f1b54d2e6dced3d4f8902 100644
--- a/tensorflow/tools/api/golden/tensorflow.layers.-dropout.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.layers.-dropout.pbtxt
@@ -108,7 +108,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'use_resource\', \'partitioner\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'use_resource\', \'synchronization\', \'aggregation\', \'partitioner\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.layers.-flatten.pbtxt b/tensorflow/tools/api/golden/tensorflow.layers.-flatten.pbtxt
index b75a012811ff10f055382ea1315eaba506c24ed8..7d2eaaab2a8cb9159214a16ba65473d0b6870ac4 100644
--- a/tensorflow/tools/api/golden/tensorflow.layers.-flatten.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.layers.-flatten.pbtxt
@@ -108,7 +108,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'use_resource\', \'partitioner\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'use_resource\', \'synchronization\', \'aggregation\', \'partitioner\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.layers.-layer.pbtxt b/tensorflow/tools/api/golden/tensorflow.layers.-layer.pbtxt
index 80e0fb228b034727854ab1a4df97e25c6bc2cd97..8bc3eb26e9ca0bf0f129db336b7ca23466fd036f 100644
--- a/tensorflow/tools/api/golden/tensorflow.layers.-layer.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.layers.-layer.pbtxt
@@ -106,7 +106,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'use_resource\', \'partitioner\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'use_resource\', \'synchronization\', \'aggregation\', \'partitioner\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.layers.-max-pooling1-d.pbtxt b/tensorflow/tools/api/golden/tensorflow.layers.-max-pooling1-d.pbtxt
index 50ff484d733633e20e9923dbbf1344af7b51ba9a..6a0dcce56ac0184ffe995662fd62b89e16257a29 100644
--- a/tensorflow/tools/api/golden/tensorflow.layers.-max-pooling1-d.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.layers.-max-pooling1-d.pbtxt
@@ -109,7 +109,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'use_resource\', \'partitioner\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'use_resource\', \'synchronization\', \'aggregation\', \'partitioner\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.layers.-max-pooling2-d.pbtxt b/tensorflow/tools/api/golden/tensorflow.layers.-max-pooling2-d.pbtxt
index cea809744cd07cc6ed0d1655f217cb5821e503e4..b6c84edf2a2f86240369b4053cd7351d0b59442d 100644
--- a/tensorflow/tools/api/golden/tensorflow.layers.-max-pooling2-d.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.layers.-max-pooling2-d.pbtxt
@@ -109,7 +109,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'use_resource\', \'partitioner\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'use_resource\', \'synchronization\', \'aggregation\', \'partitioner\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.layers.-max-pooling3-d.pbtxt b/tensorflow/tools/api/golden/tensorflow.layers.-max-pooling3-d.pbtxt
index ab9e89554c81decf5ee7e42dc963da9ab35e65c7..062a02fa590537b9efbf540a874eeaa6d36697f3 100644
--- a/tensorflow/tools/api/golden/tensorflow.layers.-max-pooling3-d.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.layers.-max-pooling3-d.pbtxt
@@ -109,7 +109,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'use_resource\', \'partitioner\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'use_resource\', \'synchronization\', \'aggregation\', \'partitioner\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.layers.-separable-conv1-d.pbtxt b/tensorflow/tools/api/golden/tensorflow.layers.-separable-conv1-d.pbtxt
index 4362568445e892d6127759c925d47426d49d9927..eaad0fb23ef7501c8c5b7acee6a9677665b7057f 100644
--- a/tensorflow/tools/api/golden/tensorflow.layers.-separable-conv1-d.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.layers.-separable-conv1-d.pbtxt
@@ -110,7 +110,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'use_resource\', \'partitioner\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'use_resource\', \'synchronization\', \'aggregation\', \'partitioner\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.layers.-separable-conv2-d.pbtxt b/tensorflow/tools/api/golden/tensorflow.layers.-separable-conv2-d.pbtxt
index 3cad824cd3b197b91a749347c860ff926610c081..ece28a8ce962d8fafb3f7a397a814b903e915d48 100644
--- a/tensorflow/tools/api/golden/tensorflow.layers.-separable-conv2-d.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.layers.-separable-conv2-d.pbtxt
@@ -110,7 +110,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'use_resource\', \'partitioner\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'use_resource\', \'synchronization\', \'aggregation\', \'partitioner\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.math.pbtxt b/tensorflow/tools/api/golden/tensorflow.math.pbtxt
index 25573cb494d0dbcbb8aaf95f4f7ebc8416ac7066..a308c76ebc08df06c0c360579451ea70e60695d4 100644
--- a/tensorflow/tools/api/golden/tensorflow.math.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.math.pbtxt
@@ -34,7 +34,7 @@ tf_module {
   }
   member_method {
     name: "bessel_i0"
-    argspec: "args=[\'x\', \'name\'], varargs=None, keywords=None, defaults=[\'bessel_i0\'], "
+    argspec: "args=[\'x\', \'name\'], varargs=None, keywords=None, defaults=[\'None\'], "
   }
   member_method {
     name: "bessel_i0e"
@@ -42,7 +42,7 @@ tf_module {
   }
   member_method {
     name: "bessel_i1"
-    argspec: "args=[\'x\', \'name\'], varargs=None, keywords=None, defaults=[\'bessel_i1\'], "
+    argspec: "args=[\'x\', \'name\'], varargs=None, keywords=None, defaults=[\'None\'], "
   }
   member_method {
     name: "bessel_i1e"
diff --git a/tensorflow/tools/api/golden/tensorflow.nn.pbtxt b/tensorflow/tools/api/golden/tensorflow.nn.pbtxt
index 455590d866a4c1ebea65ccff51e34f2e0b0479d7..d9e5b0d0fca8bbcf82feb34304f2a1e4f43f48dd 100644
--- a/tensorflow/tools/api/golden/tensorflow.nn.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.nn.pbtxt
@@ -260,6 +260,10 @@ tf_module {
     name: "relu_layer"
     argspec: "args=[\'x\', \'weights\', \'biases\', \'name\'], varargs=None, keywords=None, defaults=[\'None\'], "
   }
+  member_method {
+    name: "safe_embedding_lookup_sparse"
+    argspec: "args=[\'embedding_weights\', \'sparse_ids\', \'sparse_weights\', \'combiner\', \'default_id\', \'name\', \'partition_strategy\', \'max_norm\'], varargs=None, keywords=None, defaults=[\'None\', \'mean\', \'None\', \'None\', \'div\', \'None\'], "
+  }
   member_method {
     name: "sampled_softmax_loss"
     argspec: "args=[\'weights\', \'biases\', \'labels\', \'inputs\', \'num_sampled\', \'num_classes\', \'num_true\', \'sampled_values\', \'remove_accidental_hits\', \'partition_strategy\', \'name\', \'seed\'], varargs=None, keywords=None, defaults=[\'1\', \'None\', \'True\', \'mod\', \'sampled_softmax_loss\', \'None\'], "
diff --git a/tensorflow/tools/api/golden/tensorflow.nn.rnn_cell.-basic-l-s-t-m-cell.pbtxt b/tensorflow/tools/api/golden/tensorflow.nn.rnn_cell.-basic-l-s-t-m-cell.pbtxt
index a8d9e120cb4aa965c1d85df59de1fbabc196bf54..c74773000aa31b0c51677b49eed6e83cc1f073ae 100644
--- a/tensorflow/tools/api/golden/tensorflow.nn.rnn_cell.-basic-l-s-t-m-cell.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.nn.rnn_cell.-basic-l-s-t-m-cell.pbtxt
@@ -117,7 +117,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'use_resource\', \'partitioner\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'use_resource\', \'synchronization\', \'aggregation\', \'partitioner\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.nn.rnn_cell.-basic-r-n-n-cell.pbtxt b/tensorflow/tools/api/golden/tensorflow.nn.rnn_cell.-basic-r-n-n-cell.pbtxt
index c039890e1f4c1d57e7b795f1f09cff71921f6554..d251f548069b430de0fe9af83b6e9c641ea9237c 100644
--- a/tensorflow/tools/api/golden/tensorflow.nn.rnn_cell.-basic-r-n-n-cell.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.nn.rnn_cell.-basic-r-n-n-cell.pbtxt
@@ -117,7 +117,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'use_resource\', \'partitioner\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'use_resource\', \'synchronization\', \'aggregation\', \'partitioner\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.nn.rnn_cell.-device-wrapper.pbtxt b/tensorflow/tools/api/golden/tensorflow.nn.rnn_cell.-device-wrapper.pbtxt
index 62c393de34475a8806015bed187572f79cf2a196..8a63b4918008674041c9c216a5e5547ed7152fce 100644
--- a/tensorflow/tools/api/golden/tensorflow.nn.rnn_cell.-device-wrapper.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.nn.rnn_cell.-device-wrapper.pbtxt
@@ -116,7 +116,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'use_resource\', \'partitioner\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'use_resource\', \'synchronization\', \'aggregation\', \'partitioner\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.nn.rnn_cell.-dropout-wrapper.pbtxt b/tensorflow/tools/api/golden/tensorflow.nn.rnn_cell.-dropout-wrapper.pbtxt
index f121ba7939acb14681aa6b04b333668dded37aad..db1aae275792dad94c4cf823d0d30f934e397601 100644
--- a/tensorflow/tools/api/golden/tensorflow.nn.rnn_cell.-dropout-wrapper.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.nn.rnn_cell.-dropout-wrapper.pbtxt
@@ -120,7 +120,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'use_resource\', \'partitioner\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'use_resource\', \'synchronization\', \'aggregation\', \'partitioner\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.nn.rnn_cell.-g-r-u-cell.pbtxt b/tensorflow/tools/api/golden/tensorflow.nn.rnn_cell.-g-r-u-cell.pbtxt
index 4583dc32b2e98d4a9912378fe0e3d841882772fd..d76eab7eb874c981ac111cf6f96f28363f5e4375 100644
--- a/tensorflow/tools/api/golden/tensorflow.nn.rnn_cell.-g-r-u-cell.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.nn.rnn_cell.-g-r-u-cell.pbtxt
@@ -117,7 +117,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'use_resource\', \'partitioner\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'use_resource\', \'synchronization\', \'aggregation\', \'partitioner\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.nn.rnn_cell.-l-s-t-m-cell.pbtxt b/tensorflow/tools/api/golden/tensorflow.nn.rnn_cell.-l-s-t-m-cell.pbtxt
index 5016b6ac3010e2e184674db4837173c57c44b97e..944db6ac937acb0d6a134aa2f17dfaa0d3d618ff 100644
--- a/tensorflow/tools/api/golden/tensorflow.nn.rnn_cell.-l-s-t-m-cell.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.nn.rnn_cell.-l-s-t-m-cell.pbtxt
@@ -117,7 +117,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'use_resource\', \'partitioner\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'use_resource\', \'synchronization\', \'aggregation\', \'partitioner\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.nn.rnn_cell.-multi-r-n-n-cell.pbtxt b/tensorflow/tools/api/golden/tensorflow.nn.rnn_cell.-multi-r-n-n-cell.pbtxt
index 59623fc983a63c2966882aa5113423c0a9e23b72..72b40cc9f7a720888a1399a60aa216013e0b9918 100644
--- a/tensorflow/tools/api/golden/tensorflow.nn.rnn_cell.-multi-r-n-n-cell.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.nn.rnn_cell.-multi-r-n-n-cell.pbtxt
@@ -116,7 +116,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'use_resource\', \'partitioner\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'use_resource\', \'synchronization\', \'aggregation\', \'partitioner\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.nn.rnn_cell.-r-n-n-cell.pbtxt b/tensorflow/tools/api/golden/tensorflow.nn.rnn_cell.-r-n-n-cell.pbtxt
index e2ab5aaee9456ffbe42894f2384d7bc9c7ad6a6f..a5c2b4aefd6a1b96cbe63271ca27de06616f1deb 100644
--- a/tensorflow/tools/api/golden/tensorflow.nn.rnn_cell.-r-n-n-cell.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.nn.rnn_cell.-r-n-n-cell.pbtxt
@@ -115,7 +115,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'use_resource\', \'partitioner\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'use_resource\', \'synchronization\', \'aggregation\', \'partitioner\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.nn.rnn_cell.-residual-wrapper.pbtxt b/tensorflow/tools/api/golden/tensorflow.nn.rnn_cell.-residual-wrapper.pbtxt
index bd2a6d61f8578a2a3c8d94d3a8d5eb49679df2f7..61d5f04b22a4b4e3801643958b73a35403b79139 100644
--- a/tensorflow/tools/api/golden/tensorflow.nn.rnn_cell.-residual-wrapper.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.nn.rnn_cell.-residual-wrapper.pbtxt
@@ -116,7 +116,7 @@ tf_class {
   }
   member_method {
     name: "add_weight"
-    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'use_resource\', \'partitioner\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'self\', \'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'constraint\', \'use_resource\', \'synchronization\', \'aggregation\', \'partitioner\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\'], "
   }
   member_method {
     name: "apply"
diff --git a/tensorflow/tools/api/golden/tensorflow.pbtxt b/tensorflow/tools/api/golden/tensorflow.pbtxt
index 20d61aae9daabc430a7f4f10529b41ab28cbdd86..4f90743fec4bf0de42db9d54d801ea9ab620bd75 100644
--- a/tensorflow/tools/api/golden/tensorflow.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.pbtxt
@@ -260,10 +260,18 @@ tf_module {
     name: "Variable"
     mtype: ""
   }
+  member {
+    name: "VariableAggregation"
+    mtype: ""
+  }
   member {
     name: "VariableScope"
     mtype: ""
   }
+  member {
+    name: "VariableSynchronization"
+    mtype: ""
+  }
   member {
     name: "WholeFileReader"
     mtype: ""
@@ -1150,7 +1158,7 @@ tf_module {
   }
   member_method {
     name: "get_local_variable"
-    argspec: "args=[\'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'collections\', \'caching_device\', \'partitioner\', \'validate_shape\', \'use_resource\', \'custom_getter\', \'constraint\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'False\', \'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'collections\', \'caching_device\', \'partitioner\', \'validate_shape\', \'use_resource\', \'synchronization\', \'aggregation\', \'custom_getter\', \'constraint\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'False\', \'None\', \'None\', \'None\', \'True\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\', \'None\', \'None\'], "
   }
   member_method {
     name: "get_seed"
@@ -1166,7 +1174,7 @@ tf_module {
   }
   member_method {
     name: "get_variable"
-    argspec: "args=[\'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'collections\', \'caching_device\', \'partitioner\', \'validate_shape\', \'use_resource\', \'custom_getter\', \'constraint\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\'], "
+    argspec: "args=[\'name\', \'shape\', \'dtype\', \'initializer\', \'regularizer\', \'trainable\', \'collections\', \'caching_device\', \'partitioner\', \'validate_shape\', \'use_resource\', \'custom_getter\', \'constraint\', \'synchronization\', \'aggregation\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'None\', \'True\', \'None\', \'None\', \'None\', \'VariableSynchronization.AUTO\', \'VariableAggregation.NONE\'], "
   }
   member_method {
     name: "get_variable_scope"
@@ -1310,7 +1318,7 @@ tf_module {
   }
   member_method {
     name: "lbeta"
-    argspec: "args=[\'x\', \'name\'], varargs=None, keywords=None, defaults=[\'lbeta\'], "
+    argspec: "args=[\'x\', \'name\'], varargs=None, keywords=None, defaults=[\'None\'], "
   }
   member_method {
     name: "less"
@@ -1552,10 +1560,6 @@ tf_module {
     name: "pow"
     argspec: "args=[\'x\', \'y\', \'name\'], varargs=None, keywords=None, defaults=[\'None\'], "
   }
-  member_method {
-    name: "print"
-    argspec: "args=[\'input_\', \'data\', \'message\', \'first_n\', \'summarize\', \'name\'], varargs=None, keywords=None, defaults=[\'None\', \'None\', \'None\', \'None\'], "
-  }
   member_method {
     name: "py_func"
     argspec: "args=[\'func\', \'inp\', \'Tout\', \'stateful\', \'name\'], varargs=None, keywords=None, defaults=[\'True\', \'None\'], "
@@ -2190,7 +2194,7 @@ tf_module {
   }
   member_method {
     name: "while_loop"
-    argspec: "args=[\'cond\', \'body\', \'loop_vars\', \'shape_invariants\', \'parallel_iterations\', \'back_prop\', \'swap_memory\', \'name\', \'maximum_iterations\'], varargs=None, keywords=None, defaults=[\'None\', \'10\', \'True\', \'False\', \'None\', \'None\'], "
+    argspec: "args=[\'cond\', \'body\', \'loop_vars\', \'shape_invariants\', \'parallel_iterations\', \'back_prop\', \'swap_memory\', \'name\', \'maximum_iterations\', \'return_same_structure\'], varargs=None, keywords=None, defaults=[\'None\', \'10\', \'True\', \'False\', \'None\', \'None\', \'False\'], "
   }
   member_method {
     name: "write_file"
diff --git a/tensorflow/tools/api/golden/tensorflow.spectral.pbtxt b/tensorflow/tools/api/golden/tensorflow.spectral.pbtxt
index 4f306540ccfdeac8ce59a394ec77b24284f13ceb..6a421ef12d58dc047905ec916cbe777b4ce19b9a 100644
--- a/tensorflow/tools/api/golden/tensorflow.spectral.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.spectral.pbtxt
@@ -16,6 +16,10 @@ tf_module {
     name: "fft3d"
     argspec: "args=[\'input\', \'name\'], varargs=None, keywords=None, defaults=[\'None\'], "
   }
+  member_method {
+    name: "idct"
+    argspec: "args=[\'input\', \'type\', \'n\', \'axis\', \'norm\', \'name\'], varargs=None, keywords=None, defaults=[\'2\', \'None\', \'-1\', \'None\', \'None\'], "
+  }
   member_method {
     name: "ifft"
     argspec: "args=[\'input\', \'name\'], varargs=None, keywords=None, defaults=[\'None\'], "
diff --git a/tensorflow/tools/api/golden/tensorflow.variance_scaling_initializer.pbtxt b/tensorflow/tools/api/golden/tensorflow.variance_scaling_initializer.pbtxt
index a58398d645e8397dc8e61a6e0241710c3e34218f..09d7bc03b4f238923db6778ec32ce78ae76eed61 100644
--- a/tensorflow/tools/api/golden/tensorflow.variance_scaling_initializer.pbtxt
+++ b/tensorflow/tools/api/golden/tensorflow.variance_scaling_initializer.pbtxt
@@ -5,7 +5,7 @@ tf_class {
   is_instance: ""
   member_method {
     name: "__init__"
-    argspec: "args=[\'self\', \'scale\', \'mode\', \'distribution\', \'seed\', \'dtype\'], varargs=None, keywords=None, defaults=[\'1.0\', \'fan_in\', \'normal\', \'None\', \"\"], "
+    argspec: "args=[\'self\', \'scale\', \'mode\', \'distribution\', \'seed\', \'dtype\'], varargs=None, keywords=None, defaults=[\'1.0\', \'fan_in\', \'truncated_normal\', \'None\', \"\"], "
   }
   member_method {
     name: "from_config"
diff --git a/tensorflow/tools/api/lib/python_object_to_proto_visitor.py b/tensorflow/tools/api/lib/python_object_to_proto_visitor.py
index 1cf330e70247260cd9e50b18903bdfecad6260e4..3a48cf683c908021a6a87849601227283a8e2034 100644
--- a/tensorflow/tools/api/lib/python_object_to_proto_visitor.py
+++ b/tensorflow/tools/api/lib/python_object_to_proto_visitor.py
@@ -88,6 +88,9 @@ def _SanitizedMRO(obj):
   """
   return_list = []
   for cls in tf_inspect.getmro(obj):
+    if cls.__name__ == '_NewClass':
+      # Ignore class created by @deprecated_alias decorator.
+      continue
     str_repr = str(cls)
     return_list.append(str_repr)
     if 'tensorflow' not in str_repr:
diff --git a/tensorflow/tools/api/tests/api_compatibility_test.py b/tensorflow/tools/api/tests/api_compatibility_test.py
index 90375a794f64a9edd2bab2671f5870ae02e84e3c..d1b34fb242cd6303b61315b64ec60e6fc503aca2 100644
--- a/tensorflow/tools/api/tests/api_compatibility_test.py
+++ b/tensorflow/tools/api/tests/api_compatibility_test.py
@@ -34,6 +34,13 @@ import sys
 import unittest
 
 import tensorflow as tf
+# pylint: disable=g-import-not-at-top
+try:
+  from tensorflow.compat import v1 as tf_v1
+  # We import compat.v1 as tf_v1 instead.
+  del tf.compat.v1
+except ImportError:
+  tf_v1 = None
 
 from google.protobuf import message
 from google.protobuf import text_format
@@ -46,6 +53,7 @@ from tensorflow.tools.api.lib import api_objects_pb2
 from tensorflow.tools.api.lib import python_object_to_proto_visitor
 from tensorflow.tools.common import public_api
 from tensorflow.tools.common import traverse
+# pylint: enable=g-import-not-at-top
 
 
 # FLAGS defined at the bottom:
@@ -215,25 +223,19 @@ class ApiCompatibilityTest(test.TestCase):
     visitor.do_not_descend_map['tf'].append('contrib')
     traverse.traverse(tf, visitor)
 
-  @unittest.skipUnless(
-      sys.version_info.major == 2,
-      'API compabitility test goldens are generated using python2.')
-  def testAPIBackwardsCompatibility(self):
-    # Extract all API stuff.
+  def checkBackwardsCompatibility(self, root, golden_file_pattern):
+     # Extract all API stuff.
     visitor = python_object_to_proto_visitor.PythonObjectToProtoVisitor()
 
     public_api_visitor = public_api.PublicAPIVisitor(visitor)
     public_api_visitor.do_not_descend_map['tf'].append('contrib')
     public_api_visitor.do_not_descend_map['tf.GPUOptions'] = ['Experimental']
-    traverse.traverse(tf, public_api_visitor)
+    traverse.traverse(root, public_api_visitor)
 
     proto_dict = visitor.GetProtos()
 
     # Read all golden files.
-    expression = os.path.join(
-        resource_loader.get_root_dir_with_all_resources(),
-        _KeyToFilePath('*'))
-    golden_file_list = file_io.get_matching_files(expression)
+    golden_file_list = file_io.get_matching_files(golden_file_pattern)
 
     def _ReadFileToProto(filename):
       """Read a filename, create a protobuf from its contents."""
@@ -254,6 +256,26 @@ class ApiCompatibilityTest(test.TestCase):
         verbose=FLAGS.verbose_diffs,
         update_goldens=FLAGS.update_goldens)
 
+  @unittest.skipUnless(
+      sys.version_info.major == 2,
+      'API compabitility test goldens are generated using python2.')
+  def testAPIBackwardsCompatibility(self):
+    golden_file_pattern = os.path.join(
+        resource_loader.get_root_dir_with_all_resources(),
+        _KeyToFilePath('*'))
+    self.checkBackwardsCompatibility(tf, golden_file_pattern)
+
+  @unittest.skipUnless(
+      sys.version_info.major == 2,
+      'API compabitility test goldens are generated using python2.')
+  def testAPIBackwardsCompatibilityV1(self):
+    if not tf_v1:
+      return
+    golden_file_pattern = os.path.join(
+        resource_loader.get_root_dir_with_all_resources(),
+        _KeyToFilePath('*'))
+    self.checkBackwardsCompatibility(tf_v1, golden_file_pattern)
+
 
 if __name__ == '__main__':
   parser = argparse.ArgumentParser()
diff --git a/tensorflow/tools/ci_build/Dockerfile.rbe.cpu b/tensorflow/tools/ci_build/Dockerfile.rbe.cpu
index 3bc52b9ed611a0f0a4a269a2864d5b349ee9232c..7e5860aeec186d908e5d2884bd690b2e5e43cffa 100644
--- a/tensorflow/tools/ci_build/Dockerfile.rbe.cpu
+++ b/tensorflow/tools/ci_build/Dockerfile.rbe.cpu
@@ -1,4 +1,4 @@
-FROM launcher.gcr.io/google/rbe-debian8:r327695
+FROM launcher.gcr.io/google/rbe-ubuntu16-04:r327695
 LABEL maintainer="Yu Yi "
 
 # Copy install scripts
@@ -9,6 +9,6 @@ ENV CC /usr/local/bin/clang
 ENV CXX /usr/local/bin/clang++
 ENV AR /usr/bin/ar
 
-# Run pip install script for RBE Debian8 container.
+# Run pip install script for RBE Ubuntu 16-04 container.
 RUN /install/install_pip_packages_remote.sh
 RUN /install/install_pip_packages.sh
diff --git a/tensorflow/tools/ci_build/ci_parameterized_build.sh b/tensorflow/tools/ci_build/ci_parameterized_build.sh
index d49d4b0c49cf9ed487249a800e3807140a9a03bf..08e2c3edd2d22fbb7b9912c9ce7ec561dc5a7113 100755
--- a/tensorflow/tools/ci_build/ci_parameterized_build.sh
+++ b/tensorflow/tools/ci_build/ci_parameterized_build.sh
@@ -131,7 +131,7 @@ BAZEL_CMD="bazel test"
 BAZEL_BUILD_ONLY_CMD="bazel build"
 BAZEL_CLEAN_CMD="bazel clean"
 
-DEFAULT_BAZEL_CONFIGS="--config=gcp --config=hdfs"
+DEFAULT_BAZEL_CONFIGS=""
 
 PIP_CMD="${CI_BUILD_DIR}/builds/pip.sh"
 PIP_TEST_TUTORIALS_FLAG="--test_tutorials"
diff --git a/tensorflow/tools/ci_build/ci_sanity.sh b/tensorflow/tools/ci_build/ci_sanity.sh
index 05676f9551d4a1e0cb55d0693f99e458381887df..db37edf8097844646236aace5e3517a8080d70cb 100755
--- a/tensorflow/tools/ci_build/ci_sanity.sh
+++ b/tensorflow/tools/ci_build/ci_sanity.sh
@@ -349,12 +349,12 @@ do_external_licenses_check(){
 
   # Blacklist
   echo ${MISSING_LICENSES_FILE}
-  grep -e "@bazel_tools//third_party/" -e "@com_google_absl//absl" -e "@org_tensorflow//" -v ${MISSING_LICENSES_FILE} > temp.txt
+  grep -e "@bazel_tools//third_party/" -e "@com_google_absl//absl" -e "@org_tensorflow//" -e "@com_github_googlecloudplatform_google_cloud_cpp//google" -v ${MISSING_LICENSES_FILE} > temp.txt
   mv temp.txt ${MISSING_LICENSES_FILE}
 
   # Whitelist
   echo ${EXTRA_LICENSE_FILE}
-  grep -e "@bazel_tools//src" -e "@bazel_tools//tools/" -e "@com_google_absl//" -e "//external" -e "@local" -v ${EXTRA_LICENSES_FILE} > temp.txt
+  grep -e "@bazel_tools//src" -e "@bazel_tools//tools/" -e "@com_google_absl//" -e "//external" -e "@local" -e "@com_github_googlecloudplatform_google_cloud_cpp//" -v ${EXTRA_LICENSES_FILE} > temp.txt
   mv temp.txt ${EXTRA_LICENSES_FILE}
 
 
@@ -543,7 +543,7 @@ SANITY_STEPS=("do_pylint PYTHON2" "do_pylint PYTHON3" "do_check_futures_test" "d
 SANITY_STEPS_DESC=("Python 2 pylint" "Python 3 pylint" "Check that python files have certain __future__ imports" "buildifier check" "bazel nobuild" "pip: license check for external dependencies" "C library: license check for external dependencies" "Java Native Library: license check for external dependencies" "Pip Smoke Test: Checking py_test dependencies exist in pip package" "Check load py_test: Check that BUILD files with py_test target properly load py_test" "Code Link Check: Check there are no broken links" "Test entries in /tensorflow/contrib/cmake/python_{modules|protos|protos_cc}.txt for validity and consistency" "Check file names for cases")
 
 INCREMENTAL_FLAG=""
-DEFAULT_BAZEL_CONFIGS="--config=hdfs --config=gcp"
+DEFAULT_BAZEL_CONFIGS=""
 
 # Parse command-line arguments
 BAZEL_FLAGS=${DEFAULT_BAZEL_CONFIGS}
diff --git a/tensorflow/tools/ci_build/linux/mkl/build-dev-container.sh b/tensorflow/tools/ci_build/linux/mkl/build-dev-container.sh
index bd16d580f5bee075c684462221dcc04088d8bca7..ad22ebe4eb304fe6b6f8613f43f2c7c001111503 100755
--- a/tensorflow/tools/ci_build/linux/mkl/build-dev-container.sh
+++ b/tensorflow/tools/ci_build/linux/mkl/build-dev-container.sh
@@ -25,9 +25,14 @@ function upsearch () {
 
 # Set up WORKSPACE.
 WORKSPACE="${WORKSPACE:-$(upsearch WORKSPACE)}"
-TF_DOCKER_BUILD_DEVEL_BRANCH="master"
-TF_DOCKER_BUILD_IMAGE_NAME="intel-mkl/tensorflow"
-TF_DOCKER_BUILD_VERSION="nightly"
+
+TF_DOCKER_BUILD_DEVEL_BRANCH=${TF_DOCKER_BUILD_DEVEL_BRANCH:-master}
+TF_DOCKER_BUILD_IMAGE_NAME=${TF_DOCKER_BUILD_IMAGE_NAME:-intel-mkl/tensorflow}
+TF_DOCKER_BUILD_VERSION=${TF_DOCKER_BUILD_VERSION:-nightly}
+
+echo "TF_DOCKER_BUILD_DEVEL_BRANCH=${TF_DOCKER_BUILD_DEVEL_BRANCH}"
+echo "TF_DOCKER_BUILD_IMAGE_NAME=${TF_DOCKER_BUILD_IMAGE_NAME}"
+echo "TF_DOCKER_BUILD_VERSION=${TF_DOCKER_BUILD_VERSION}"
 
 # build the python 2 container and whl
 TF_DOCKER_BUILD_TYPE="MKL" \
diff --git a/tensorflow/tools/ci_build/pi/build_raspberry_pi.sh b/tensorflow/tools/ci_build/pi/build_raspberry_pi.sh
index b8bce57c87ab39ab2f51288163187f2e87c9135d..3d27e84b81c586729aff21d0859383c24f436a11 100755
--- a/tensorflow/tools/ci_build/pi/build_raspberry_pi.sh
+++ b/tensorflow/tools/ci_build/pi/build_raspberry_pi.sh
@@ -65,6 +65,10 @@ OPENBLAS_SRC_PATH=/tmp/openblas_src/
 sudo rm -rf ${OPENBLAS_SRC_PATH}
 git clone https://github.com/xianyi/OpenBLAS ${OPENBLAS_SRC_PATH}
 cd ${OPENBLAS_SRC_PATH}
+# The commit after this introduced Fortran compile issues. In theory they should
+# be solvable using NOFORTRAN=1 on the make command, but my initial tries didn't
+# work, so pinning to the last know good version.
+git checkout 5a6a2bed9aff0ba8a18651d5514d029c8cae336a
 # If this path is changed, you'll also need to update
 # cxx_builtin_include_directory in third_party/toolchains/cpus/arm/CROSSTOOL.tpl
 OPENBLAS_INSTALL_PATH=/tmp/openblas_install/
diff --git a/tensorflow/tools/ci_build/update_version.py b/tensorflow/tools/ci_build/update_version.py
index 642dde36a7caae35df764d5d7513df972e1e5615..30c318a58fae4c84033ea5e906f3ec88818c4b65 100755
--- a/tensorflow/tools/ci_build/update_version.py
+++ b/tensorflow/tools/ci_build/update_version.py
@@ -248,16 +248,6 @@ def update_md_files(old_version, new_version):
     replace_string_in_line(r"%s<\/version>" % old_version,
                            "%s" % new_version, filepath)
 
-  # Update any links to colab notebooks.
-  def colab_url(version):
-    version_string = "%s.%s.%s" % (version.major, version.minor, version.patch)
-    prefix = "https://colab.research.google.com/github/tensorflow/models/blob/r"
-    return prefix + version_string + "/"
-
-  replace_string_in_line(
-      colab_url(old_version), colab_url(new_version),
-      "%s/docs_src/get_started/eager.md" % TF_SRC_DIR)
-
 
 def major_minor_change(old_version, new_version):
   """Check if a major or minor change occurred."""
diff --git a/tensorflow/tools/ci_build/windows/bazel/bazel_test_lib.sh b/tensorflow/tools/ci_build/windows/bazel/bazel_test_lib.sh
index a3e07737a4fa79de80cf667d058517772db9f103..c03cbd9c66a1d12185920290d4dc3cd52e4a616c 100644
--- a/tensorflow/tools/ci_build/windows/bazel/bazel_test_lib.sh
+++ b/tensorflow/tools/ci_build/windows/bazel/bazel_test_lib.sh
@@ -23,17 +23,20 @@ function run_configure_for_gpu_build {
   # Enable CUDA support
   export TF_NEED_CUDA=1
 
-  # TODO(pcloudy): Remove this after TensorFlow uses its own CRSOOTOOL
-  # for GPU build on Windows
-  export USE_MSVC_WRAPPER=1
-
   yes "" | ./configure
 }
 
-function set_gcs_remote_cache_options {
-  echo "build --experimental_remote_spawn_cache" >> "${TMP_BAZELRC}"
+function set_remote_cache_options {
+  echo "build --remote_instance_name=projects/tensorflow-testing-cpu" >> "${TMP_BAZELRC}"
   echo "build --experimental_remote_platform_override='properties:{name:\"build\" value:\"windows-x64\"}'" >> "${TMP_BAZELRC}"
-  echo "build --remote_http_cache=https://storage.googleapis.com/$GCS_BUCKET_NAME" >> "${TMP_BAZELRC}"
+  echo "build --remote_cache=remotebuildexecution.googleapis.com" >> "${TMP_BAZELRC}"
+  echo "build --tls_enabled=true" >> "${TMP_BAZELRC}"
+  echo "build --remote_timeout=3600" >> "${TMP_BAZELRC}"
+  echo "build --auth_enabled=true" >> "${TMP_BAZELRC}"
+  echo "build --spawn_strategy=remote" >> "${TMP_BAZELRC}"
+  echo "build --strategy=Javac=remote" >> "${TMP_BAZELRC}"
+  echo "build --strategy=Closure=remote" >> "${TMP_BAZELRC}"
+  echo "build --genrule_strategy=remote" >> "${TMP_BAZELRC}"
   echo "build --google_credentials=$GOOGLE_CLOUD_CREDENTIAL" >> "${TMP_BAZELRC}"
 }
 
diff --git a/tensorflow/tools/ci_build/windows/bazel/common_env.sh b/tensorflow/tools/ci_build/windows/bazel/common_env.sh
index eefa8ee2d504945991c91e1574b6a74330ba3a8d..3af132217e3e0a7d514c7e16d16989e7fcdb6c9a 100644
--- a/tensorflow/tools/ci_build/windows/bazel/common_env.sh
+++ b/tensorflow/tools/ci_build/windows/bazel/common_env.sh
@@ -49,3 +49,15 @@ export PATH="/c/Program Files/Git/cmd:$PATH"
 
 # Make sure we have pip in PATH
 export PATH="/c/${PYTHON_BASE_PATH}/Scripts:$PATH"
+
+# Setting default values to CUDA related environment variables
+export TF_CUDA_VERSION=${TF_CUDA_VERSION:-9.0}
+export TF_CUDNN_VERSION=${TF_CUDNN_VERSION:-7.0}
+export TF_CUDA_COMPUTE_CAPABILITIES=${TF_CUDA_COMPUTE_CAPABILITIES:-3.7}
+export CUDA_TOOLKIT_PATH=${CUDA_TOOLKIT_PATH:-"C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v${TF_CUDA_VERSION}"}
+export CUDNN_INSTALL_PATH=${CUDNN_INSTALL_PATH:-"C:/tools/cuda"}
+
+# Add Cuda and Cudnn dll directories into PATH
+export PATH="$(cygpath -u "${CUDA_TOOLKIT_PATH}")/bin:$PATH"
+export PATH="$(cygpath -u "${CUDA_TOOLKIT_PATH}")/extras/CUPTI/libx64:$PATH"
+export PATH="$(cygpath -u "${CUDNN_INSTALL_PATH}")/bin:$PATH"
diff --git a/tensorflow/tools/ci_build/windows/cpu/pip/build_tf_windows.sh b/tensorflow/tools/ci_build/windows/cpu/pip/build_tf_windows.sh
index 5c305f7512852dd6b3e43c4745e7f24c8a4502aa..ed7340146789078bf12fc3bbfba46fb0f740ba54 100644
--- a/tensorflow/tools/ci_build/windows/cpu/pip/build_tf_windows.sh
+++ b/tensorflow/tools/ci_build/windows/cpu/pip/build_tf_windows.sh
@@ -59,8 +59,8 @@ release_build=0
 for ARG in "$@"; do
   if [[ "$ARG" == --skip_test ]]; then
     skip_test=1
-  elif [[ "$ARG" == --enable_gcs_remote_cache ]]; then
-    set_gcs_remote_cache_options
+  elif [[ "$ARG" == --enable_remote_cache ]]; then
+    set_remote_cache_options
   elif [[ "$ARG" == --release_build ]]; then
     release_build=1
   fi
diff --git a/tensorflow/tools/ci_build/windows/gpu/pip/build_tf_windows.sh b/tensorflow/tools/ci_build/windows/gpu/pip/build_tf_windows.sh
index 922bb67bbf6ce34f55acad6d3399bd810032abd0..fe3bce428fb2feb053cb1b8c097f707dd2762a20 100644
--- a/tensorflow/tools/ci_build/windows/gpu/pip/build_tf_windows.sh
+++ b/tensorflow/tools/ci_build/windows/gpu/pip/build_tf_windows.sh
@@ -42,9 +42,58 @@ source "tensorflow/tools/ci_build/windows/bazel/common_env.sh" \
 source "tensorflow/tools/ci_build/windows/bazel/bazel_test_lib.sh" \
   || { echo "Failed to source bazel_test_lib.sh" >&2; exit 1; }
 
+# Recreate an empty bazelrc file under source root
+export TMP_BAZELRC=.tmp.bazelrc
+rm -f "${TMP_BAZELRC}"
+touch "${TMP_BAZELRC}"
+
+function cleanup {
+  # Remove all options in .tmp.bazelrc
+  echo "" > "${TMP_BAZELRC}"
+}
+trap cleanup EXIT
+
+skip_test=0
+release_build=0
+
+for ARG in "$@"; do
+  if [[ "$ARG" == --skip_test ]]; then
+    skip_test=1
+  elif [[ "$ARG" == --enable_remote_cache ]]; then
+    set_remote_cache_options
+  elif [[ "$ARG" == --release_build ]]; then
+    release_build=1
+  fi
+done
+
+if [[ "$release_build" != 1 ]]; then
+  # --define=override_eigen_strong_inline=true speeds up the compiling of conv_grad_ops_3d.cc and conv_ops_3d.cc
+  # by 20 minutes. See https://github.com/tensorflow/tensorflow/issues/10521
+  # Because this hurts the performance of TF, we don't enable it in release build.
+  echo "build --define=override_eigen_strong_inline=true" >> "${TMP_BAZELRC}"
+fi
+
+# The host and target platforms are the same in Windows build. So we don't have
+# to distinct them. This helps avoid building the same targets twice.
+echo "build --distinct_host_configuration=false" >> "${TMP_BAZELRC}"
+
+# Enable short object file path to avoid long path issue on Windows.
+echo "startup --output_user_root=${TMPDIR}" >> "${TMP_BAZELRC}"
+
+# Disable nvcc warnings to reduce log file size.
+echo "build --copt=-nvcc_options=disable-warnings" >> "${TMP_BAZELRC}"
+
+if ! grep -q "import %workspace%/${TMP_BAZELRC}" .bazelrc; then
+  echo "import %workspace%/${TMP_BAZELRC}" >> .bazelrc
+fi
+
 run_configure_for_gpu_build
 
-bazel build -c opt tensorflow/tools/pip_package:build_pip_package || exit $?
+bazel build --announce_rc --config=opt tensorflow/tools/pip_package:build_pip_package || exit $?
+
+if [[ "$skip_test" == 1 ]]; then
+  exit 0
+fi
 
 # Create a python test directory to avoid package name conflict
 PY_TEST_DIR="py_test_dir"
@@ -59,8 +108,11 @@ reinstall_tensorflow_pip ${PIP_NAME}
 # Define no_tensorflow_py_deps=true so that every py_test has no deps anymore,
 # which will result testing system installed tensorflow
 # GPU tests are very flaky when running concurrently, so set local_test_jobs=1
-bazel test -c opt -k --test_output=errors \
+bazel test --announce_rc --config=opt -k --test_output=errors \
   --define=no_tensorflow_py_deps=true --test_lang_filters=py \
-  --test_tag_filters=-no_pip,-no_windows,-no_windows_gpu,-no_gpu,-no_pip_gpu,no_oss \
-  --build_tag_filters=-no_pip,-no_windows,-no_windows_gpu,-no_gpu,-no_pip_gpu,no_oss \
-  --local_test_jobs=1 --build_tests_only //${PY_TEST_DIR}/tensorflow/python/...
+  --test_tag_filters=-no_pip,-no_windows,-no_windows_gpu,-no_gpu,-no_pip_gpu,-no_oss \
+  --build_tag_filters=-no_pip,-no_windows,-no_windows_gpu,-no_gpu,-no_pip_gpu,-no_oss --build_tests_only \
+  --local_test_jobs=1 --test_timeout="300,450,1200,3600" \
+  --flaky_test_attempts=3 \
+  //${PY_TEST_DIR}/tensorflow/python/... \
+  //${PY_TEST_DIR}/tensorflow/contrib/...
diff --git a/tensorflow/tools/compatibility/ast_edits.py b/tensorflow/tools/compatibility/ast_edits.py
new file mode 100644
index 0000000000000000000000000000000000000000..23cc4a21a9e6f81c8dc5016bc2cb6a2f151c7924
--- /dev/null
+++ b/tensorflow/tools/compatibility/ast_edits.py
@@ -0,0 +1,502 @@
+# Copyright 2016 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""Upgrader for Python scripts according to an API change specification."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import ast
+import collections
+import os
+import shutil
+import sys
+import tempfile
+import traceback
+
+
+class APIChangeSpec(object):
+  """This class defines the transformations that need to happen.
+
+  This class must provide the following fields:
+
+  * `function_keyword_renames`: maps function names to a map of old -> new
+    argument names
+  * `function_renames`: maps function names to new function names
+  * `change_to_function`: a set of function names that have changed (for
+    notifications)
+  * `function_reorders`: maps functions whose argument order has changed to the
+    list of arguments in the new order
+  * `function_handle`: maps function names to custom handlers for the function
+
+  For an example, see `TFAPIChangeSpec`.
+  """
+
+
+class _FileEditTuple(
+    collections.namedtuple("_FileEditTuple",
+                           ["comment", "line", "start", "old", "new"])):
+  """Each edit that is recorded by a _FileEditRecorder.
+
+  Fields:
+    comment: A description of the edit and why it was made.
+    line: The line number in the file where the edit occurs (1-indexed).
+    start: The line number in the file where the edit occurs (0-indexed).
+    old: text string to remove (this must match what was in file).
+    new: text string to add in place of `old`.
+  """
+
+  __slots__ = ()
+
+
+class _FileEditRecorder(object):
+  """Record changes that need to be done to the file."""
+
+  def __init__(self, filename):
+    # all edits are lists of chars
+    self._filename = filename
+
+    self._line_to_edit = collections.defaultdict(list)
+    self._errors = []
+
+  def process(self, text):
+    """Process a list of strings, each corresponding to the recorded changes.
+
+    Args:
+      text: A list of lines of text (assumed to contain newlines)
+    Returns:
+      A tuple of the modified text and a textual description of what is done.
+    Raises:
+      ValueError: if substitution source location does not have expected text.
+    """
+
+    change_report = ""
+
+    # Iterate of each line
+    for line, edits in self._line_to_edit.items():
+      offset = 0
+      # sort by column so that edits are processed in order in order to make
+      # indexing adjustments cumulative for changes that change the string
+      # length
+      edits.sort(key=lambda x: x.start)
+
+      # Extract each line to a list of characters, because mutable lists
+      # are editable, unlike immutable strings.
+      char_array = list(text[line - 1])
+
+      # Record a description of the change
+      change_report += "%r Line %d\n" % (self._filename, line)
+      change_report += "-" * 80 + "\n\n"
+      for e in edits:
+        change_report += "%s\n" % e.comment
+      change_report += "\n    Old: %s" % (text[line - 1])
+
+      # Make underscore buffers for underlining where in the line the edit was
+      change_list = [" "] * len(text[line - 1])
+      change_list_new = [" "] * len(text[line - 1])
+
+      # Iterate for each edit
+      for e in edits:
+        # Create effective start, end by accounting for change in length due
+        # to previous edits
+        start_eff = e.start + offset
+        end_eff = start_eff + len(e.old)
+
+        # Make sure the edit is changing what it should be changing
+        old_actual = "".join(char_array[start_eff:end_eff])
+        if old_actual != e.old:
+          raise ValueError("Expected text %r but got %r" %
+                           ("".join(e.old), "".join(old_actual)))
+        # Make the edit
+        char_array[start_eff:end_eff] = list(e.new)
+
+        # Create the underline highlighting of the before and after
+        change_list[e.start:e.start + len(e.old)] = "~" * len(e.old)
+        change_list_new[start_eff:end_eff] = "~" * len(e.new)
+
+        # Keep track of how to generate effective ranges
+        offset += len(e.new) - len(e.old)
+
+      # Finish the report comment
+      change_report += "         %s\n" % "".join(change_list)
+      text[line - 1] = "".join(char_array)
+      change_report += "    New: %s" % (text[line - 1])
+      change_report += "         %s\n\n" % "".join(change_list_new)
+    return "".join(text), change_report, self._errors
+
+  def add(self, comment, line, start, old, new, error=None):
+    """Add a new change that is needed.
+
+    Args:
+      comment: A description of what was changed
+      line: Line number (1 indexed)
+      start: Column offset (0 indexed)
+      old: old text
+      new: new text
+      error: this "edit" is something that cannot be fixed automatically
+    Returns:
+      None
+    """
+
+    self._line_to_edit[line].append(
+        _FileEditTuple(comment, line, start, old, new))
+    if error:
+      self._errors.append("%s:%d: %s" % (self._filename, line, error))
+
+
+class _ASTCallVisitor(ast.NodeVisitor):
+  """AST Visitor that processes function calls.
+
+  Updates function calls from old API version to new API version using a given
+  change spec.
+  """
+
+  def __init__(self, filename, lines, api_change_spec):
+    self._filename = filename
+    self._file_edit = _FileEditRecorder(filename)
+    self._lines = lines
+    self._api_change_spec = api_change_spec
+
+  def process(self, lines):
+    return self._file_edit.process(lines)
+
+  def generic_visit(self, node):
+    ast.NodeVisitor.generic_visit(self, node)
+
+  def _rename_functions(self, node, full_name):
+    function_renames = self._api_change_spec.function_renames
+    try:
+      new_name = function_renames[full_name]
+      self._file_edit.add("Renamed function %r to %r" % (full_name, new_name),
+                          node.lineno, node.col_offset, full_name, new_name)
+    except KeyError:
+      pass
+
+  def _get_attribute_full_path(self, node):
+    """Traverse an attribute to generate a full name e.g. tf.foo.bar.
+
+    Args:
+      node: A Node of type Attribute.
+
+    Returns:
+      a '.'-delimited full-name or None if the tree was not a simple form.
+      i.e. `foo()+b).bar` returns None, while `a.b.c` would return "a.b.c".
+    """
+    curr = node
+    items = []
+    while not isinstance(curr, ast.Name):
+      if not isinstance(curr, ast.Attribute):
+        return None
+      items.append(curr.attr)
+      curr = curr.value
+    items.append(curr.id)
+    return ".".join(reversed(items))
+
+  def _find_true_position(self, node):
+    """Return correct line number and column offset for a given node.
+
+    This is necessary mainly because ListComp's location reporting reports
+    the next token after the list comprehension list opening.
+
+    Args:
+      node: Node for which we wish to know the lineno and col_offset
+    """
+    import re
+    find_open = re.compile("^\s*(\\[).*$")
+    find_string_chars = re.compile("['\"]")
+
+    if isinstance(node, ast.ListComp):
+      # Strangely, ast.ListComp returns the col_offset of the first token
+      # after the '[' token which appears to be a bug. Workaround by
+      # explicitly finding the real start of the list comprehension.
+      line = node.lineno
+      col = node.col_offset
+      # loop over lines
+      while 1:
+        # Reverse the text to and regular expression search for whitespace
+        text = self._lines[line - 1]
+        reversed_preceding_text = text[:col][::-1]
+        # First find if a [ can be found with only whitespace between it and
+        # col.
+        m = find_open.match(reversed_preceding_text)
+        if m:
+          new_col_offset = col - m.start(1) - 1
+          return line, new_col_offset
+        else:
+          if (reversed_preceding_text == "" or
+              reversed_preceding_text.isspace()):
+            line = line - 1
+            prev_line = self._lines[line - 1]
+            # TODO(aselle):
+            # this is poor comment detection, but it is good enough for
+            # cases where the comment does not contain string literal starting/
+            # ending characters. If ast gave us start and end locations of the
+            # ast nodes rather than just start, we could use string literal
+            # node ranges to filter out spurious #'s that appear in string
+            # literals.
+            comment_start = prev_line.find("#")
+            if comment_start == -1:
+              col = len(prev_line) - 1
+            elif find_string_chars.search(prev_line[comment_start:]) is None:
+              col = comment_start
+            else:
+              return None, None
+          else:
+            return None, None
+    # Most other nodes return proper locations (with notably does not), but
+    # it is not possible to use that in an argument.
+    return node.lineno, node.col_offset
+
+  def visit_Call(self, node):  # pylint: disable=invalid-name
+    """Handle visiting a call node in the AST.
+
+    Args:
+      node: Current Node
+    """
+
+    # Find a simple attribute name path e.g. "tf.foo.bar"
+    full_name = self._get_attribute_full_path(node.func)
+
+    # Make sure the func is marked as being part of a call
+    node.func.is_function_for_call = True
+
+    if full_name:
+      # Call special handlers
+      function_handles = self._api_change_spec.function_handle
+      if full_name in function_handles:
+        function_handles[full_name](self._file_edit, node)
+
+      # Examine any non-keyword argument and make it into a keyword argument
+      # if reordering required.
+      function_reorders = self._api_change_spec.function_reorders
+      function_keyword_renames = (
+          self._api_change_spec.function_keyword_renames)
+
+      if full_name in function_reorders:
+        reordered = function_reorders[full_name]
+        for idx, arg in enumerate(node.args):
+          lineno, col_offset = self._find_true_position(arg)
+          if lineno is None or col_offset is None:
+            self._file_edit.add(
+                "Failed to add keyword %r to reordered function %r" %
+                (reordered[idx], full_name),
+                arg.lineno,
+                arg.col_offset,
+                "",
+                "",
+                error="A necessary keyword argument failed to be inserted.")
+          else:
+            keyword_arg = reordered[idx]
+            if (full_name in function_keyword_renames and
+                keyword_arg in function_keyword_renames[full_name]):
+              keyword_arg = function_keyword_renames[full_name][keyword_arg]
+            self._file_edit.add("Added keyword %r to reordered function %r" %
+                                (reordered[idx], full_name), lineno, col_offset,
+                                "", keyword_arg + "=")
+
+      # Examine each keyword argument and convert it to the final renamed form
+      renamed_keywords = ({} if full_name not in function_keyword_renames else
+                          function_keyword_renames[full_name])
+      for keyword in node.keywords:
+        argkey = keyword.arg
+        argval = keyword.value
+
+        if argkey in renamed_keywords:
+          argval_lineno, argval_col_offset = self._find_true_position(argval)
+          if argval_lineno is not None and argval_col_offset is not None:
+            # TODO(aselle): We should scan backward to find the start of the
+            # keyword key. Unfortunately ast does not give you the location of
+            # keyword keys, so we are forced to infer it from the keyword arg
+            # value.
+            key_start = argval_col_offset - len(argkey) - 1
+            key_end = key_start + len(argkey) + 1
+            if (self._lines[argval_lineno - 1][key_start:key_end] == argkey +
+                "="):
+              self._file_edit.add("Renamed keyword argument from %r to %r" %
+                                  (argkey,
+                                   renamed_keywords[argkey]), argval_lineno,
+                                  argval_col_offset - len(argkey) - 1,
+                                  argkey + "=", renamed_keywords[argkey] + "=")
+              continue
+          self._file_edit.add(
+              "Failed to rename keyword argument from %r to %r" %
+              (argkey, renamed_keywords[argkey]),
+              argval.lineno,
+              argval.col_offset - len(argkey) - 1,
+              "",
+              "",
+              error="Failed to find keyword lexographically. Fix manually.")
+
+    ast.NodeVisitor.generic_visit(self, node)
+
+  def visit_Attribute(self, node):  # pylint: disable=invalid-name
+    """Handle bare Attributes i.e. [tf.foo, tf.bar].
+
+    Args:
+      node: Node that is of type ast.Attribute
+    """
+    full_name = self._get_attribute_full_path(node)
+    if full_name:
+      self._rename_functions(node, full_name)
+    if full_name in self._api_change_spec.change_to_function:
+      if not hasattr(node, "is_function_for_call"):
+        new_text = full_name + "()"
+        self._file_edit.add("Changed %r to %r" % (full_name, new_text),
+                            node.lineno, node.col_offset, full_name, new_text)
+
+    ast.NodeVisitor.generic_visit(self, node)
+
+
+class ASTCodeUpgrader(object):
+  """Handles upgrading a set of Python files using a given API change spec."""
+
+  def __init__(self, api_change_spec):
+    if not isinstance(api_change_spec, APIChangeSpec):
+      raise TypeError("Must pass APIChangeSpec to ASTCodeUpgrader, got %s" %
+                      type(api_change_spec))
+    self._api_change_spec = api_change_spec
+
+  def process_file(self, in_filename, out_filename):
+    """Process the given python file for incompatible changes.
+
+    Args:
+      in_filename: filename to parse
+      out_filename: output file to write to
+    Returns:
+      A tuple representing number of files processed, log of actions, errors
+    """
+
+    # Write to a temporary file, just in case we are doing an implace modify.
+    with open(in_filename, "r") as in_file, \
+        tempfile.NamedTemporaryFile("w", delete=False) as temp_file:
+      ret = self.process_opened_file(in_filename, in_file, out_filename,
+                                     temp_file)
+
+    shutil.move(temp_file.name, out_filename)
+    return ret
+
+  # Broad exceptions are required here because ast throws whatever it wants.
+  # pylint: disable=broad-except
+  def process_opened_file(self, in_filename, in_file, out_filename, out_file):
+    """Process the given python file for incompatible changes.
+
+    This function is split out to facilitate StringIO testing from
+    tf_upgrade_test.py.
+
+    Args:
+      in_filename: filename to parse
+      in_file: opened file (or StringIO)
+      out_filename: output file to write to
+      out_file: opened file (or StringIO)
+    Returns:
+      A tuple representing number of files processed, log of actions, errors
+    """
+    process_errors = []
+    text = "-" * 80 + "\n"
+    text += "Processing file %r\n outputting to %r\n" % (in_filename,
+                                                         out_filename)
+    text += "-" * 80 + "\n\n"
+
+    parsed_ast = None
+    lines = in_file.readlines()
+    try:
+      parsed_ast = ast.parse("".join(lines))
+    except Exception:
+      text += "Failed to parse %r\n\n" % in_filename
+      text += traceback.format_exc()
+    if parsed_ast:
+      visitor = _ASTCallVisitor(in_filename, lines, self._api_change_spec)
+      visitor.visit(parsed_ast)
+      out_text, new_text, process_errors = visitor.process(lines)
+      text += new_text
+      if out_file:
+        out_file.write(out_text)
+    text += "\n"
+    return 1, text, process_errors
+
+  # pylint: enable=broad-except
+
+  def process_tree(self, root_directory, output_root_directory,
+                   copy_other_files):
+    """Processes upgrades on an entire tree of python files in place.
+
+    Note that only Python files. If you have custom code in other languages,
+    you will need to manually upgrade those.
+
+    Args:
+      root_directory: Directory to walk and process.
+      output_root_directory: Directory to use as base.
+      copy_other_files: Copy files that are not touched by this converter.
+
+    Returns:
+      A tuple of files processed, the report string ofr all files, and errors
+    """
+
+    # make sure output directory doesn't exist
+    if output_root_directory and os.path.exists(output_root_directory):
+      print("Output directory %r must not already exist." %
+            (output_root_directory))
+      sys.exit(1)
+
+    # make sure output directory does not overlap with root_directory
+    norm_root = os.path.split(os.path.normpath(root_directory))
+    norm_output = os.path.split(os.path.normpath(output_root_directory))
+    if norm_root == norm_output:
+      print("Output directory %r same as input directory %r" %
+            (root_directory, output_root_directory))
+      sys.exit(1)
+
+    # Collect list of files to process (we do this to correctly handle if the
+    # user puts the output directory in some sub directory of the input dir)
+    files_to_process = []
+    files_to_copy = []
+    for dir_name, _, file_list in os.walk(root_directory):
+      py_files = [f for f in file_list if f.endswith(".py")]
+      copy_files = [f for f in file_list if not f.endswith(".py")]
+      for filename in py_files:
+        fullpath = os.path.join(dir_name, filename)
+        fullpath_output = os.path.join(output_root_directory,
+                                       os.path.relpath(fullpath,
+                                                       root_directory))
+        files_to_process.append((fullpath, fullpath_output))
+      if copy_other_files:
+        for filename in copy_files:
+          fullpath = os.path.join(dir_name, filename)
+          fullpath_output = os.path.join(output_root_directory,
+                                         os.path.relpath(
+                                             fullpath, root_directory))
+          files_to_copy.append((fullpath, fullpath_output))
+
+    file_count = 0
+    tree_errors = []
+    report = ""
+    report += ("=" * 80) + "\n"
+    report += "Input tree: %r\n" % root_directory
+    report += ("=" * 80) + "\n"
+
+    for input_path, output_path in files_to_process:
+      output_directory = os.path.dirname(output_path)
+      if not os.path.isdir(output_directory):
+        os.makedirs(output_directory)
+      file_count += 1
+      _, l_report, l_errors = self.process_file(input_path, output_path)
+      tree_errors += l_errors
+      report += l_report
+    for input_path, output_path in files_to_copy:
+      output_directory = os.path.dirname(output_path)
+      if not os.path.isdir(output_directory):
+        os.makedirs(output_directory)
+      shutil.copy(input_path, output_path)
+    return file_count, report, tree_errors
diff --git a/tensorflow/tools/docker/Dockerfile.devel b/tensorflow/tools/docker/Dockerfile.devel
index 57a491255ea968b08e6e9cbaf9dd0178e8d2c3bf..fd94d64268ae69a83f04124e33e5bde3c2f49b75 100644
--- a/tensorflow/tools/docker/Dockerfile.devel
+++ b/tensorflow/tools/docker/Dockerfile.devel
@@ -63,7 +63,7 @@ RUN echo "startup --batch" >>/etc/bazel.bazelrc
 RUN echo "build --spawn_strategy=standalone --genrule_strategy=standalone" \
     >>/etc/bazel.bazelrc
 # Install the most recent bazel release.
-ENV BAZEL_VERSION 0.11.0
+ENV BAZEL_VERSION 0.14.1
 WORKDIR /
 RUN mkdir /bazel && \
     cd /bazel && \
diff --git a/tensorflow/tools/docker/Dockerfile.devel-cpu-mkl b/tensorflow/tools/docker/Dockerfile.devel-cpu-mkl
new file mode 100644
index 0000000000000000000000000000000000000000..6796ad70e5d22ca683343680b142081d8d58a9e4
--- /dev/null
+++ b/tensorflow/tools/docker/Dockerfile.devel-cpu-mkl
@@ -0,0 +1,83 @@
+FROM tensorflow/tensorflow:latest-devel
+
+LABEL maintainer="Clayne Robison"
+
+# These arguments are parameterized. Use --build-args to override.
+ARG TF_BRANCH=r1.9
+ARG WHL_DIR=/whl
+
+RUN apt-get update && apt-get install -y --no-install-recommends \
+        golang \
+        vim \
+        emacs \
+        && \
+    apt-get clean && \
+    rm -rf /var/lib/apt/lists/*
+
+RUN pip --no-cache-dir install --upgrade \
+        pip setuptools
+
+RUN pip --no-cache-dir install wheel 
+
+# Download and build TensorFlow.
+WORKDIR /
+RUN rm -rf tensorflow && \
+    git clone https://github.com/tensorflow/tensorflow.git && \
+    cd tensorflow && \
+    git checkout ${TF_BRANCH}
+WORKDIR /tensorflow
+
+# Configure the build for CPU with MKL by accepting default build options and
+# setting library locations
+ENV CI_BUILD_PYTHON=python \
+   LD_LIBRARY_PATH=${LD_LIBRARY_PATH} \
+    PYTHON_BIN_PATH=/usr/bin/python \
+    PYTHON_LIB_PATH=/usr/local/lib/python2.7/dist-packages \
+    CC_OPT_FLAGS='-march=native' \
+    TF_NEED_JEMALLOC=0 \
+    TF_NEED_GCP=1 \
+    TF_NEED_CUDA=0 \
+    TF_NEED_HDFS=0 \
+    TF_NEED_S3=1 \
+    TF_NEED_OPENCL=0 \
+    TF_NEED_GDR=0 \
+    TF_ENABLE_XLA=0 \
+    TF_NEED_VERBS=0 \
+    TF_NEED_MPI=0
+RUN ./configure
+
+# Build and Install TensorFlow.
+# The 'mkl' option builds with Intel(R) Math Kernel Library (MKL), which detects
+# the platform it is currently running on and takes appropriately optimized 
+# paths. The -march=native option is for code that is not in MKL, and assumes
+# this container will be run on the same architecture on which it is built.
+RUN LD_LIBRARY_PATH=${LD_LIBRARY_PATH} \
+    bazel build --config=mkl \
+                --config="opt" \
+                --copt="-march=broadwell" \
+                --copt="-O3" \
+                //tensorflow/tools/pip_package:build_pip_package && \
+    mkdir ${WHL_DIR} && \
+    bazel-bin/tensorflow/tools/pip_package/build_pip_package ${WHL_DIR}
+
+# Clean up Bazel cache when done, but leave the whl.
+# This will upgrade the default Tensorflow version with the Intel MKL version
+RUN pip --no-cache-dir install --upgrade ${WHL_DIR}/tensorflow-*.whl && \
+    rm -rf /root/.cache
+
+WORKDIR /root
+
+#add welcome message with instructions
+
+RUN echo '[ ! -z "$TERM" -a -r /etc/motd ] && cat /etc/issue && cat /etc/motd' \
+	>> /etc/bash.bashrc \
+	; echo "\
+||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||\n\
+|								\n\
+| Docker container running Ubuntu				\n\
+| with TensorFlow ${TF_BRANCH} optimized for CPU		\n\
+| with Intel(R) MKL						\n\
+|								\n\
+||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||\n\
+\n "\
+	> /etc/motd
diff --git a/tensorflow/tools/docker/Dockerfile.devel-gpu b/tensorflow/tools/docker/Dockerfile.devel-gpu
index 204b5b4dba1b607fb709b7f45d145ceafc33f3e7..5ec43b8cb8dfa96d8c180b56eddf19aa0b8fb925 100644
--- a/tensorflow/tools/docker/Dockerfile.devel-gpu
+++ b/tensorflow/tools/docker/Dockerfile.devel-gpu
@@ -72,7 +72,7 @@ RUN echo "startup --batch" >>/etc/bazel.bazelrc
 RUN echo "build --spawn_strategy=standalone --genrule_strategy=standalone" \
     >>/etc/bazel.bazelrc
 # Install the most recent bazel release.
-ENV BAZEL_VERSION 0.11.0
+ENV BAZEL_VERSION 0.14.1
 WORKDIR /
 RUN mkdir /bazel && \
     cd /bazel && \
diff --git a/tensorflow/tools/docker/Dockerfile.devel-gpu-cuda9-cudnn7 b/tensorflow/tools/docker/Dockerfile.devel-gpu-cuda9-cudnn7
new file mode 100644
index 0000000000000000000000000000000000000000..3bedc8cf3462aabf25f55706b3483907c5d5b467
--- /dev/null
+++ b/tensorflow/tools/docker/Dockerfile.devel-gpu-cuda9-cudnn7
@@ -0,0 +1,115 @@
+FROM nvidia/cuda:9.0-cudnn7-devel-ubuntu16.04
+
+LABEL maintainer="Gunhan Gulsoy "
+
+# It is possible to override these for releases.
+ARG TF_BRANCH=master
+ARG BAZEL_VERSION=0.5.4
+ARG TF_AVAILABLE_CPUS=32
+
+RUN apt-get update && apt-get install -y --no-install-recommends \
+        build-essential \
+        curl \
+        git \
+        golang \
+        libcurl3-dev \
+        libfreetype6-dev \
+        libpng12-dev \
+        libzmq3-dev \
+        pkg-config \
+        python-dev \
+        python-pip \
+        rsync \
+        software-properties-common \
+        unzip \
+        zip \
+        zlib1g-dev \
+        openjdk-8-jdk \
+        openjdk-8-jre-headless \
+        wget \
+        && \
+    apt-get clean && \
+    rm -rf /var/lib/apt/lists/*
+
+RUN pip --no-cache-dir install --upgrade \
+        pip setuptools
+
+RUN pip --no-cache-dir install \
+        ipykernel \
+        jupyter \
+        matplotlib \
+        numpy \
+        scipy \
+        sklearn \
+        pandas \
+        wheel \
+        && \
+    python -m ipykernel.kernelspec
+
+# Set up our notebook config.
+COPY jupyter_notebook_config.py /root/.jupyter/
+
+# Jupyter has issues with being run directly:
+#   https://github.com/ipython/ipython/issues/7062
+# We just add a little wrapper script.
+COPY run_jupyter.sh /
+
+# Set up Bazel.
+
+# Running bazel inside a `docker build` command causes trouble, cf:
+#   https://github.com/bazelbuild/bazel/issues/134
+# The easiest solution is to set up a bazelrc file forcing --batch.
+RUN echo "startup --batch" >>/etc/bazel.bazelrc
+# Similarly, we need to workaround sandboxing issues:
+#   https://github.com/bazelbuild/bazel/issues/418
+RUN echo "build --spawn_strategy=standalone --genrule_strategy=standalone" \
+    >>/etc/bazel.bazelrc
+WORKDIR /
+RUN mkdir /bazel && \
+    cd /bazel && \
+    wget --quiet https://github.com/bazelbuild/bazel/releases/download/$BAZEL_VERSION/bazel-$BAZEL_VERSION-installer-linux-x86_64.sh && \
+    wget --quiet https://raw.githubusercontent.com/bazelbuild/bazel/master/LICENSE && \
+    chmod +x bazel-*.sh && \
+    ./bazel-$BAZEL_VERSION-installer-linux-x86_64.sh && \
+    rm -f /bazel/bazel-$BAZEL_VERSION-installer-linux-x86_64.sh
+
+# Download and build TensorFlow.
+WORKDIR /
+RUN git clone https://github.com/tensorflow/tensorflow.git && \
+    cd tensorflow && \
+    git checkout ${TF_BRANCH}
+WORKDIR /tensorflow
+
+# Configure the build for our CUDA configuration.
+ENV CI_BUILD_PYTHON=python \
+    LD_LIBRARY_PATH=/usr/local/cuda/extras/CUPTI/lib64:${LD_LIBRARY_PATH} \
+    CUDNN_INSTALL_PATH=/usr/lib/x86_64-linux-gnu \
+    PYTHON_BIN_PATH=/usr/bin/python \
+    PYTHON_LIB_PATH=/usr/local/lib/python2.7/dist-packages \
+    TF_NEED_CUDA=1 \
+    TF_CUDA_VERSION=9.0 \
+    TF_CUDA_COMPUTE_CAPABILITIES=3.0,3.5,5.2,6.0,6.1,7.0 \
+    TF_CUDNN_VERSION=7
+RUN ./configure
+
+# Build and Install TensorFlow.
+RUN ln -s /usr/local/cuda/lib64/stubs/libcuda.so /usr/local/cuda/lib64/stubs/libcuda.so.1 && \
+    LD_LIBRARY_PATH=/usr/local/cuda/lib64/stubs:${LD_LIBRARY_PATH} \
+    bazel build -c opt \
+                --config=cuda \
+                --cxxopt="-D_GLIBCXX_USE_CXX11_ABI=0" \
+                --jobs=${TF_AVAILABLE_CPUS} \
+                tensorflow/tools/pip_package:build_pip_package && \
+    mkdir /pip_pkg && \
+    bazel-bin/tensorflow/tools/pip_package/build_pip_package /pip_pkg && \
+    pip --no-cache-dir install --upgrade /pip_pkg/tensorflow-*.whl && \
+    rm -rf /pip_pkg && \
+    rm -rf /root/.cache
+# Clean up pip wheel and Bazel cache when done.
+
+WORKDIR /root
+
+# TensorBoard
+EXPOSE 6006
+# IPython
+EXPOSE 8888
diff --git a/tensorflow/tools/docker/Dockerfile.devel-mkl b/tensorflow/tools/docker/Dockerfile.devel-mkl
index aa6d02766289c9ac3af2c14f87f2d76363f98ead..c85641b38301e90a3dfbc3e67bc0e6deabbd68db 100755
--- a/tensorflow/tools/docker/Dockerfile.devel-mkl
+++ b/tensorflow/tools/docker/Dockerfile.devel-mkl
@@ -73,7 +73,7 @@ RUN echo "startup --batch" >>/etc/bazel.bazelrc
 RUN echo "build --spawn_strategy=standalone --genrule_strategy=standalone" \
     >>/etc/bazel.bazelrc
 # Install the most recent bazel release.
-ENV BAZEL_VERSION 0.11.0
+ENV BAZEL_VERSION 0.14.1
 WORKDIR /
 RUN mkdir /bazel && \
     cd /bazel && \
@@ -86,7 +86,18 @@ RUN mkdir /bazel && \
 
 # Download and build TensorFlow.
 WORKDIR /tensorflow
-RUN git clone --branch=${TF_BUILD_VERSION} --depth=1 https://github.com/tensorflow/tensorflow.git .
+
+# Download and build TensorFlow.
+# Enable checking out both tags and branches
+RUN export TAG_PREFIX="v" && \
+    echo ${TF_BUILD_VERSION} | grep -q ^${TAG_PREFIX}; \
+    if [ $? -eq 0 ]; then \
+        git clone --depth=1 https://github.com/tensorflow/tensorflow.git . && \
+        git fetch --tags && \
+        git checkout ${TF_BUILD_VERSION}; \
+   else \
+        git clone --depth=1 --branch=${TF_BUILD_VERSION} https://github.com/tensorflow/tensorflow.git . ; \
+    fi
 
 RUN yes "" | ${PYTHON} configure.py
 
@@ -103,7 +114,7 @@ COPY .bazelrc /root/.bazelrc
 
 RUN tensorflow/tools/ci_build/builds/configured CPU \
     bazel --bazelrc=/root/.bazelrc build -c opt \
-        tensorflow/tools/pip_package:build_pip_package && \
+    tensorflow/tools/pip_package:build_pip_package && \
     bazel-bin/tensorflow/tools/pip_package/build_pip_package "${WHL_DIR}" && \
     ${PIP} --no-cache-dir install --upgrade "${WHL_DIR}"/tensorflow-*.whl && \
     rm -rf /root/.cache
diff --git a/tensorflow/tools/docs/BUILD b/tensorflow/tools/docs/BUILD
index eea712c279e8685f1b011bf648b4503449a8f0f5..2403e2d966929b86976bf6a31f8144d9b4f58bc6 100644
--- a/tensorflow/tools/docs/BUILD
+++ b/tensorflow/tools/docs/BUILD
@@ -39,6 +39,7 @@ py_library(
     visibility = ["//visibility:public"],
     deps = [
         "//tensorflow/python:platform",
+        "//tensorflow/python:util",
         "@astor_archive//:astor",
     ],
 )
@@ -95,6 +96,7 @@ py_binary(
     deps = [
         ":generate_lib",
         "//tensorflow:tensorflow_py",
+        "//tensorflow/python:util",
         "//tensorflow/python/debug:debug_py",
     ],
 )
diff --git a/tensorflow/tools/docs/generate_lib.py b/tensorflow/tools/docs/generate_lib.py
index 67c413cccb9973aef5a1d722161295fe28876bf0..e7634cd5dcf19d5f21b0bd42b282dfe928659a52 100644
--- a/tensorflow/tools/docs/generate_lib.py
+++ b/tensorflow/tools/docs/generate_lib.py
@@ -388,16 +388,40 @@ def _build_guide_index(guide_src_dir):
 
 
 class _UpdateTags(py_guide_parser.PyGuideParser):
-  """Rewrites a Python guide so that each section has an explicit tag."""
+  """Rewrites a Python guide so that each section has an explicit id tag.
+
+  "section" here refers to blocks delimited by second level headings.
+  """
 
   def process_section(self, line_number, section_title, tag):
     self.replace_line(line_number, '
%s
' % (tag, section_title))
 
 
+def update_id_tags_inplace(src_dir):
+  """Set explicit ids on all second-level headings to ensure back-links work.
+
+  Args:
+    src_dir: The directory of md-files to convert (inplace).
+  """
+  tag_updater = _UpdateTags()
+
+  for dirpath, _, filenames in os.walk(src_dir):
+    for base_name in filenames:
+      if not base_name.endswith('.md'):
+        continue
+      full_path = os.path.join(src_dir, dirpath, base_name)
+
+      # Tag updater loads the file, makes the replacements, and returns the
+      # modified file contents
+      content = tag_updater.process(full_path)
+      with open(full_path, 'w') as f:
+        f.write(content)
+
+
 EXCLUDED = set(['__init__.py', 'OWNERS', 'README.txt'])
 
 
-def _other_docs(src_dir, output_dir, reference_resolver, file_pattern='*.md'):
+def replace_refs(src_dir, output_dir, reference_resolver, file_pattern='*.md'):
   """Fix @{} references in all files under `src_dir` matching `file_pattern`.
 
   A matching directory structure, with the modified files is
@@ -418,7 +442,6 @@ def _other_docs(src_dir, output_dir, reference_resolver, file_pattern='*.md'):
       using fnmatch. Non-matching files are copied unchanged.
   """
   # Iterate through all the source files and process them.
-  tag_updater = _UpdateTags()
   for dirpath, _, filenames in os.walk(src_dir):
     # How to get from `dirpath` to api_docs/python/
     relative_path_to_root = os.path.relpath(
@@ -435,24 +458,25 @@ def _other_docs(src_dir, output_dir, reference_resolver, file_pattern='*.md'):
         continue
       full_in_path = os.path.join(dirpath, base_name)
 
+      # Set the `current_doc_full_name` so bad files can be reported on errors.
       reference_resolver.current_doc_full_name = full_in_path
 
       suffix = os.path.relpath(path=full_in_path, start=src_dir)
       full_out_path = os.path.join(output_dir, suffix)
+      # Copy files that do not match the file_pattern, unmodified.
       if not fnmatch.fnmatch(base_name, file_pattern):
         shutil.copyfile(full_in_path, full_out_path)
         continue
-      if dirpath.endswith('/api_guides/python'):
-        content = tag_updater.process(full_in_path)
-      else:
-        with open(full_in_path, 'rb') as f:
-          content = f.read().decode('utf-8')
+
+      with open(full_in_path, 'rb') as f:
+        content = f.read().decode('utf-8')
 
       content = reference_resolver.replace_references(content,
                                                       relative_path_to_root)
       with open(full_out_path, 'wb') as f:
         f.write(content.encode('utf-8'))
 
+
 class DocGenerator(object):
   """Main entry point for generating docs."""
 
@@ -538,15 +562,43 @@ class DocGenerator(object):
                    self._do_not_descend_map)
 
   def build(self, flags):
-    """Actually build the docs."""
+    """Build all the docs.
+
+    This produces two outputs
+
+    python api docs:
+
+      * generated from modules set with `set_py_modules`.
+      * written to '{FLAGS.output_dir}/api_docs/python/'
+
+    non-api docs:
+
+      * Everything in '{FLAGS.src_dir}' is copied to '{FLAGS.output_dir}'.
+      * '@{}' references in '.md' files are replaced with links.
+      * '.md' files under 'api_guides/python' have explicit ids set for their
+        second level headings.
+
+    Args:
+      flags:
+        * src_dir: Where to fetch the non-api-docs.
+        * base_dir: Base of the docs directory (Used to build correct
+          relative links).
+        * output_dir: Where to write the resulting docs.
+
+    Returns:
+      The number of errors encountered while processing.
+    """
+    # Extract the python api from the _py_modules
     doc_index = build_doc_index(flags.src_dir)
     visitor = self.run_extraction()
     reference_resolver = self.make_reference_resolver(visitor, doc_index)
 
+    # Build the guide_index for the api_docs back links.
     root_title = getattr(flags, 'root_title', 'TensorFlow')
     guide_index = _build_guide_index(
         os.path.join(flags.src_dir, 'api_guides/python'))
 
+    # Write the api docs.
     parser_config = self.make_parser_config(visitor, reference_resolver,
                                             guide_index, flags.base_dir)
     output_dir = os.path.join(flags.output_dir, 'api_docs/python')
@@ -557,8 +609,16 @@ class DocGenerator(object):
         yaml_toc=self.yaml_toc,
         root_title=root_title,
         search_hints=getattr(flags, 'search_hints', True))
-    _other_docs(flags.src_dir, flags.output_dir, reference_resolver)
 
+    # Replace all the @{} references in files under `FLAGS.src_dir`
+    replace_refs(flags.src_dir, flags.output_dir, reference_resolver, '*.md')
+    # Fix the tags in the guide dir.
+    guide_dir = os.path.join(flags.output_dir, 'api_guides/python')
+    if os.path.exists(guide_dir):
+      update_id_tags_inplace(guide_dir)
+
+    # Report all errors found by the reference resolver, and return the error
+    # code.
     parser_config.reference_resolver.log_errors()
 
     return parser_config.reference_resolver.num_errors()
diff --git a/tensorflow/tools/docs/generate_lib_test.py b/tensorflow/tools/docs/generate_lib_test.py
index ea6d28a02b1f3c07fe8783fd59e345dade1fc804..7a6f9fd9f799db5a14015d77e5297955c76a51cd 100644
--- a/tensorflow/tools/docs/generate_lib_test.py
+++ b/tensorflow/tools/docs/generate_lib_test.py
@@ -51,7 +51,9 @@ class DummyVisitor(object):
 
 class GenerateTest(googletest.TestCase):
 
-  def test_write(self):
+  def get_test_objects(self):
+    # These are all mutable objects, so rebuild them for each test.
+    # Don't cache the objects.
     module = sys.modules[__name__]
 
     index = {
@@ -98,6 +100,11 @@ class GenerateTest(googletest.TestCase):
         guide_index={},
         base_dir=base_dir)
 
+    return reference_resolver, parser_config
+
+  def test_write(self):
+    _, parser_config = self.get_test_objects()
+
     output_dir = googletest.GetTempDir()
 
     generate_lib.write_docs(output_dir, parser_config, yaml_toc=True)
@@ -127,6 +134,107 @@ class GenerateTest(googletest.TestCase):
         os.path.exists(
             os.path.join(output_dir, 'tf/TestModule/test_function.md')))
 
+  def test_update_id_tags_inplace(self):
+    test_dir = googletest.GetTempDir()
+    test_sub_dir = os.path.join(test_dir, 'a/b')
+    os.makedirs(test_sub_dir)
+
+    test_path1 = os.path.join(test_dir, 'file1.md')
+    test_path2 = os.path.join(test_sub_dir, 'file2.md')
+    test_path3 = os.path.join(test_sub_dir, 'file3.notmd')
+
+    with open(test_path1, 'w') as f:
+      f.write('## abc&123')
+
+    with open(test_path2, 'w') as f:
+      f.write('# A Level 1 Heading\n')
+      f.write('## A Level 2 Heading')
+
+    with open(test_path3, 'w') as f:
+      f.write("## don\'t change this")
+
+    generate_lib.update_id_tags_inplace(test_dir)
+
+    with open(test_path1) as f:
+      content = f.read()
+
+    self.assertEqual(content, '
abc&123
')
+
+    with open(test_path2) as f:
+      content = f.read()
+
+    self.assertEqual(
+        content, '# A Level 1 Heading\n'
+        '
A Level 2 Heading
')
+
+    with open(test_path3) as f:
+      content = f.read()
+
+    self.assertEqual(content, "## don\'t change this")
+
+  def test_replace_refes(self):
+    test_dir = googletest.GetTempDir()
+    test_in_dir = os.path.join(test_dir, 'in')
+    test_in_dir_a = os.path.join(test_dir, 'in/a')
+    test_in_dir_b = os.path.join(test_dir, 'in/b')
+    os.makedirs(test_in_dir)
+    os.makedirs(test_in_dir_a)
+    os.makedirs(test_in_dir_b)
+
+    test_out_dir = os.path.join(test_dir, 'out')
+    os.makedirs(test_out_dir)
+
+    test_path1 = os.path.join(test_in_dir_a, 'file1.md')
+    test_path2 = os.path.join(test_in_dir_b, 'file2.md')
+    test_path3 = os.path.join(test_in_dir_b, 'file3.notmd')
+    test_path4 = os.path.join(test_in_dir_b, 'OWNERS')
+
+    with open(test_path1, 'w') as f:
+      f.write('Use `tf.test_function` to test things.')
+
+    with open(test_path2, 'w') as f:
+      f.write('Use @{tf.TestModule.TestClass.ChildClass} to test things.\n'
+              "`tf.whatever` doesn't exist")
+
+    with open(test_path3, 'w') as f:
+      file3_content = (
+          'Not a .md file. Should be copied unchanged:'
+          '@{tf.TestModule.TestClass.ChildClass}, `tf.test_function`')
+      f.write(file3_content)
+
+    with open(test_path4, 'w') as f:
+      f.write('')
+
+    reference_resolver, _ = self.get_test_objects()
+    generate_lib.replace_refs(test_in_dir, test_out_dir, reference_resolver,
+                              '*.md')
+
+    with open(os.path.join(test_out_dir, 'a/file1.md')) as f:
+      content = f.read()
+      self.assertEqual(
+          content,
+          'Use '
+          'tf.test_function to test things.')
+
+    with open(os.path.join(test_out_dir, 'b/file2.md')) as f:
+      content = f.read()
+      self.assertEqual(
+          content,
+          'Use '
+          ''
+          'tf.TestModule.TestClass.ChildClass '
+          'to test things.\n'
+          '`tf.whatever` doesn\'t exist')
+
+    with open(os.path.join(test_out_dir, 'b/file3.notmd')) as f:
+      content = f.read()
+      self.assertEqual(content, file3_content)
+
+    with self.assertRaises(IOError):
+      # This should fail. The OWNERS file should not be copied
+      with open(os.path.join(test_out_dir, 'b/OWNERS')) as f:
+        content = f.read()
+
 
 if __name__ == '__main__':
   googletest.main()
diff --git a/tensorflow/tools/lib_package/BUILD b/tensorflow/tools/lib_package/BUILD
index 05c23cd3ee64eca2667ccb12e3e90fab87a5256f..173f418dc8d998bc51d208a04c8671bacf364cdc 100644
--- a/tensorflow/tools/lib_package/BUILD
+++ b/tensorflow/tools/lib_package/BUILD
@@ -115,6 +115,7 @@ genrule(
         "//third_party/fft2d:LICENSE",
         "@aws//:LICENSE",
         "@boringssl//:LICENSE",
+        "@com_github_googlecloudplatform_google_cloud_cpp//:LICENSE",
         "@com_googlesource_code_re2//:LICENSE",
         "@cub_archive//:LICENSE.TXT",
         "@curl//:COPYING",
@@ -156,6 +157,7 @@ genrule(
         "//third_party/fft2d:LICENSE",
         "@aws//:LICENSE",
         "@boringssl//:LICENSE",
+        "@com_github_googlecloudplatform_google_cloud_cpp//:LICENSE",
         "@com_googlesource_code_re2//:LICENSE",
         "@cub_archive//:LICENSE.TXT",
         "@curl//:COPYING",
diff --git a/tensorflow/tools/pip_package/BUILD b/tensorflow/tools/pip_package/BUILD
index a0caf42331288681eee383873b963f865f9555fd..ac252143d7a167db6ed8ae7bebe3cc86fd45143a 100644
--- a/tensorflow/tools/pip_package/BUILD
+++ b/tensorflow/tools/pip_package/BUILD
@@ -104,6 +104,7 @@ COMMON_PIP_DEPS = [
     "//tensorflow/python/kernel_tests/testdata:self_adjoint_eig_op_test_files",
     "//tensorflow/python/saved_model:saved_model",
     "//tensorflow/python/tools:tools_pip",
+    "//tensorflow/python/tools/api/generator:create_python_api",
     "//tensorflow/python:test_ops",
     "//tensorflow/tools/dist_test/server:grpc_tensorflow_server",
 ]
@@ -130,6 +131,8 @@ filegroup(
         "@astor_archive//:LICENSE",
         "@aws//:LICENSE",
         "@boringssl//:LICENSE",
+        "@com_github_googleapis_googleapis//:LICENSE",
+        "@com_github_googlecloudplatform_google_cloud_cpp//:LICENSE",
         "@com_google_absl//:LICENSE",
         "@com_googlesource_code_re2//:LICENSE",
         "@cub_archive//:LICENSE.TXT",
diff --git a/tensorflow/tools/pip_package/build_pip_package.sh b/tensorflow/tools/pip_package/build_pip_package.sh
index 9e41514cfa1a70d649eab6fd23a599db4afae2a8..b0089d33605084b243121cb445d9d0c737899bc9 100755
--- a/tensorflow/tools/pip_package/build_pip_package.sh
+++ b/tensorflow/tools/pip_package/build_pip_package.sh
@@ -27,7 +27,7 @@ function cp_external() {
 
   pushd .
   cd "$src_dir"
-  for f in `find . ! -type d ! -name '*.py' ! -name '*local_config_cuda*' ! -name '*local_config_tensorrt*' ! -name '*org_tensorflow*'`; do
+  for f in `find . ! -type d ! -name '*.py' ! -path '*local_config_cuda*' ! -path '*local_config_tensorrt*' ! -path '*org_tensorflow*'`; do
     mkdir -p "${dest_dir}/$(dirname ${f})"
     cp "${f}" "${dest_dir}/$(dirname ${f})/"
   done
diff --git a/tensorflow/tools/pip_package/setup.py b/tensorflow/tools/pip_package/setup.py
index 55cd4f37c682d95461850d312bb48353efd8194f..c630ca04b885d35da6550d4e5f3e6912b5fd7a00 100644
--- a/tensorflow/tools/pip_package/setup.py
+++ b/tensorflow/tools/pip_package/setup.py
@@ -53,7 +53,7 @@ REQUIRED_PACKAGES = [
     'gast >= 0.2.0',
     'numpy >= 1.13.3',
     'six >= 1.10.0',
-    'protobuf >= 3.4.0',
+    'protobuf >= 3.6.0',
     'setuptools <= 39.1.0',
     'tensorboard >= 1.8.0, < 1.9.0',
     'termcolor >= 1.1.0',
@@ -170,8 +170,9 @@ class InstallHeaders(Command):
     # symlink within the directory hierarchy.
     # NOTE(keveman): Figure out how to customize bdist_wheel package so
     # we can do the symlink.
-    if 'external/eigen_archive/' in install_dir:
-      extra_dir = install_dir.replace('external/eigen_archive', '')
+    if 'tensorflow/include/external/eigen_archive/' in install_dir:
+      extra_dir = install_dir.replace(
+          'tensorflow/include/external/eigen_archive', '')
       if not os.path.exists(extra_dir):
         self.mkpath(extra_dir)
       self.copy_file(header, extra_dir)
@@ -204,13 +205,12 @@ def find_files(pattern, root):
       yield os.path.join(dirpath, filename)
 
 
-matches = ['../' + x for x in find_files('*', 'external') if '.py' not in x]
-
 so_lib_paths = [
     i for i in os.listdir('.')
     if os.path.isdir(i) and fnmatch.fnmatch(i, '_solib_*')
 ]
 
+matches = []
 for path in so_lib_paths:
   matches.extend(
       ['../' + x for x in find_files('*', path) if '.py' not in x]
@@ -225,7 +225,7 @@ headers = (list(find_files('*.h', 'tensorflow/core')) +
            list(find_files('*.h', 'tensorflow/stream_executor')) +
            list(find_files('*.h', 'google/protobuf_archive/src')) +
            list(find_files('*', 'third_party/eigen3')) +
-           list(find_files('*', 'external/eigen_archive')))
+           list(find_files('*', 'tensorflow/include/external/eigen_archive')))
 
 setup(
     name=project_name,
diff --git a/tensorflow/workspace.bzl b/tensorflow/workspace.bzl
index 5cefe377824b58a4692e753dd52182500b4b9790..cd4f17a5ff3856f4f88fdd901b0933baa4278e54 100644
--- a/tensorflow/workspace.bzl
+++ b/tensorflow/workspace.bzl
@@ -163,6 +163,27 @@ def tf_workspace(path_prefix="", tf_repo_name=""):
       strip_prefix = "re2-2018-04-01",
   )
 
+  tf_http_archive(
+      name = "com_github_googlecloudplatform_google_cloud_cpp",
+      urls = [
+          "https://mirror.bazel.build/github.com/GoogleCloudPlatform/google-cloud-cpp/archive/f875700a023bdd706333cde45aee8758b272c357.tar.gz",
+          "https://github.com/GoogleCloudPlatform/google-cloud-cpp/archive/f875700a023bdd706333cde45aee8758b272c357.tar.gz",
+      ],
+      sha256 = "a34f3c50b237686dc870b13baaa6a5836ce3473f2f2a02717299f0ff318372db",
+      strip_prefix = "google-cloud-cpp-f875700a023bdd706333cde45aee8758b272c357",
+  )
+
+  tf_http_archive(
+      name = "com_github_googleapis_googleapis",
+      urls = [
+          "https://mirror.bazel.build/github.com/googleapis/googleapis/archive/f81082ea1e2f85c43649bee26e0d9871d4b41cdb.zip",
+          "https://github.com/googleapis/googleapis/archive/f81082ea1e2f85c43649bee26e0d9871d4b41cdb.zip",
+      ],
+      sha256 = "824870d87a176f26bcef663e92051f532fac756d1a06b404055dc078425f4378",
+      strip_prefix="googleapis-f81082ea1e2f85c43649bee26e0d9871d4b41cdb",
+      build_file = clean_dep("//third_party:googleapis.BUILD"),
+  )
+
   tf_http_archive(
       name = "gemmlowp",
       urls = [
@@ -198,12 +219,12 @@ def tf_workspace(path_prefix="", tf_repo_name=""):
   tf_http_archive(
       name = "nasm",
       urls = [
-          "https://mirror.bazel.build/www.nasm.us/pub/nasm/releasebuilds/2.12.02/nasm-2.12.02.tar.bz2",
-          "http://pkgs.fedoraproject.org/repo/pkgs/nasm/nasm-2.12.02.tar.bz2/d15843c3fb7db39af80571ee27ec6fad/nasm-2.12.02.tar.bz2",
-          "http://www.nasm.us/pub/nasm/releasebuilds/2.12.02/nasm-2.12.02.tar.bz2",
+          "https://mirror.bazel.build/www.nasm.us/pub/nasm/releasebuilds/2.13.03/nasm-2.13.03.tar.bz2",
+          "http://pkgs.fedoraproject.org/repo/pkgs/nasm/nasm-2.13.03.tar.bz2/sha512/d7a6b4cee8dfd603d8d4c976e5287b5cc542fa0b466ff989b743276a6e28114e64289bf02a7819eca63142a5278aa6eed57773007e5f589e15768e6456a8919d/nasm-2.13.03.tar.bz2",
+          "http://www.nasm.us/pub/nasm/releasebuilds/2.13.03/nasm-2.13.03.tar.bz2",
       ],
-      sha256 = "00b0891c678c065446ca59bcee64719d0096d54d6886e6e472aeee2e170ae324",
-      strip_prefix = "nasm-2.12.02",
+      sha256 = "63ec86477ad3f0f6292325fd89e1d93aea2e2fd490070863f17d48f7cd387011",
+      strip_prefix = "nasm-2.13.03",
       build_file = clean_dep("//third_party:nasm.BUILD"),
   )
 
@@ -364,11 +385,11 @@ def tf_workspace(path_prefix="", tf_repo_name=""):
   tf_http_archive(
       name = "nsync",
       urls = [
-          "https://mirror.bazel.build/github.com/google/nsync/archive/5e8b19a81e5729922629dd505daa651f6ffdf107.tar.gz",
-          "https://github.com/google/nsync/archive/5e8b19a81e5729922629dd505daa651f6ffdf107.tar.gz",
+          "https://mirror.bazel.build/github.com/google/nsync/archive/1.20.0.tar.gz",
+          "https://github.com/google/nsync/archive/1.20.0.tar.gz",
       ],
-      sha256 = "2723e6db509779fcf05bd01556e51f2e5179197e2c864cd8010f6b7100a5b1e1",
-      strip_prefix = "nsync-5e8b19a81e5729922629dd505daa651f6ffdf107",
+      sha256 = "0c1b03962b2f8450f21e74a5a46116bf2d6009a807c57eb4207e974a8c4bb7dd",
+      strip_prefix = "nsync-1.20.0",
   )
 
   tf_http_archive(
@@ -384,11 +405,11 @@ def tf_workspace(path_prefix="", tf_repo_name=""):
   tf_http_archive(
       name = "com_github_gflags_gflags",
       urls = [
-          "https://mirror.bazel.build/github.com/gflags/gflags/archive/f8a0efe03aa69b3336d8e228b37d4ccb17324b88.tar.gz",
-          "https://github.com/gflags/gflags/archive/f8a0efe03aa69b3336d8e228b37d4ccb17324b88.tar.gz",
+          "https://mirror.bazel.build/github.com/gflags/gflags/archive/v2.2.1.tar.gz",
+          "https://github.com/gflags/gflags/archive/v2.2.1.tar.gz",
       ],
-      sha256 = "4d222fab8f1ede4709cdff417d15a1336f862d7334a81abf76d09c15ecf9acd1",
-      strip_prefix = "gflags-f8a0efe03aa69b3336d8e228b37d4ccb17324b88",
+      sha256 = "ae27cdbcd6a2f935baa78e4f21f675649271634c092b1be01469440495609d0e",
+      strip_prefix = "gflags-2.2.1",
   )
 
   tf_http_archive(
@@ -428,14 +449,13 @@ def tf_workspace(path_prefix="", tf_repo_name=""):
   tf_http_archive(
       name = "grpc",
       urls = [
-          "https://mirror.bazel.build/github.com/grpc/grpc/archive/v1.12.1.tar.gz",
-          "https://github.com/grpc/grpc/archive/v1.12.1.tar.gz",
+          "https://mirror.bazel.build/github.com/grpc/grpc/archive/v1.13.0.tar.gz",
+          "https://github.com/grpc/grpc/archive/v1.13.0.tar.gz",
       ],
-      sha256 = "f6afbfafa8e7b524727d1ff37ff22fe9c3dcca07bd864e7a9d1efabf1d15d13c",
-      strip_prefix = "grpc-1.12.1",
+      sha256 = "50db9cf2221354485eb7c3bd55a4c27190caef7048a2a1a15fbe60a498f98b44",
+      strip_prefix = "grpc-1.13.0",
   )
 
-
   tf_http_archive(
       name = "linenoise",
       sha256 = "7f51f45887a3d31b4ce4fa5965210a5e64637ceac12720cfce7954d6a2e812f7",
@@ -452,11 +472,11 @@ def tf_workspace(path_prefix="", tf_repo_name=""):
   tf_http_archive(
       name = "llvm",
       urls = [
-          "https://mirror.bazel.build/github.com/llvm-mirror/llvm/archive/7f7cea53068238fca7b7e4299793a0c77bea7219.tar.gz",
-	  "https://github.com/llvm-mirror/llvm/archive/7f7cea53068238fca7b7e4299793a0c77bea7219.tar.gz",
+          "https://mirror.bazel.build/github.com/llvm-mirror/llvm/archive/ae80745b73e435d07e7fb9c12589304ee29e7f59.tar.gz",
+	  "https://github.com/llvm-mirror/llvm/archive/ae80745b73e435d07e7fb9c12589304ee29e7f59.tar.gz",
       ],
-      sha256 = "b645507080e07c845607f212d45e4ee79253c3c9b762531f51fbaeceb6b47391",
-      strip_prefix = "llvm-7f7cea53068238fca7b7e4299793a0c77bea7219",
+      sha256 = "de69b6f92a634b4d12b9e03ebd8eb34c28f997d9480c28358d6efd4c433fe853",
+      strip_prefix = "llvm-ae80745b73e435d07e7fb9c12589304ee29e7f59",
       build_file = clean_dep("//third_party/llvm:llvm.autogenerated.BUILD"),
   )
 
@@ -538,11 +558,11 @@ def tf_workspace(path_prefix="", tf_repo_name=""):
   tf_http_archive(
       name = "kafka",
       urls = [
-          "https://mirror.bazel.build/github.com/edenhill/librdkafka/archive/v0.11.1.tar.gz",
-          "https://github.com/edenhill/librdkafka/archive/v0.11.1.tar.gz",
+          "https://mirror.bazel.build/github.com/edenhill/librdkafka/archive/v0.11.4.tar.gz",
+          "https://github.com/edenhill/librdkafka/archive/v0.11.4.tar.gz",
       ],
-      sha256 = "dd035d57c8f19b0b612dd6eefe6e5eebad76f506e302cccb7c2066f25a83585e",
-      strip_prefix = "librdkafka-0.11.1",
+      sha256 = "9d8f1eb7b0e29e9ab1168347c939cb7ae5dff00a39cef99e7ef033fd8f92737c",
+      strip_prefix = "librdkafka-0.11.4",
       build_file = clean_dep("//third_party:kafka/BUILD"),
       patch_file = clean_dep("//third_party/kafka:config.patch"),
   )
@@ -662,12 +682,12 @@ def tf_workspace(path_prefix="", tf_repo_name=""):
 
   tf_http_archive(
       name = "cython",
-      sha256 = "05e3eb7f06043f5ff2028338370329e71c29f57315e95f4dc6ad7c4971dd4c6f",
+      sha256 = "bccc9aa050ea02595b2440188813b936eaf345e85fb9692790cecfe095cf91aa",
       urls = [
-          "https://mirror.bazel.build/github.com/cython/cython/archive/0.28.3.tar.gz",
-          "https://github.com/cython/cython/archive/0.28.3.tar.gz",
+          "https://mirror.bazel.build/github.com/cython/cython/archive/0.28.4.tar.gz",
+          "https://github.com/cython/cython/archive/0.28.4.tar.gz",
       ],
-      strip_prefix = "cython-0.28.3",
+      strip_prefix = "cython-0.28.4",
       build_file = clean_dep("//third_party:cython.BUILD"),
       delete = ["BUILD.bazel"],
   )
@@ -675,11 +695,11 @@ def tf_workspace(path_prefix="", tf_repo_name=""):
   tf_http_archive(
       name = "bazel_toolchains",
       urls = [
-          "https://mirror.bazel.build/github.com/bazelbuild/bazel-toolchains/archive/44200e0c026d86c53470d107b3697a3e46469c43.tar.gz",
-          "https://github.com/bazelbuild/bazel-toolchains/archive/44200e0c026d86c53470d107b3697a3e46469c43.tar.gz",
+          "https://mirror.bazel.build/github.com/bazelbuild/bazel-toolchains/archive/37acf1841ab1475c98a152cb9e446460c8ae29e1.tar.gz",
+          "https://github.com/bazelbuild/bazel-toolchains/archive/37acf1841ab1475c98a152cb9e446460c8ae29e1.tar.gz",
       ],
-      strip_prefix = "bazel-toolchains-44200e0c026d86c53470d107b3697a3e46469c43",
-      sha256 = "699b55a6916c687f4b7dc092dbbf5f64672cde0dc965f79717735ec4e5416556",
+      strip_prefix = "bazel-toolchains-37acf1841ab1475c98a152cb9e446460c8ae29e1",
+      sha256 = "3b604699685c5c65dd3f6f17425570a4b2f00ddba2f750db15acc72e55bb098b",
   )
 
   tf_http_archive(
@@ -733,6 +753,14 @@ def tf_workspace(path_prefix="", tf_repo_name=""):
       ],
       build_file = str(Label("//third_party:tflite_mobilenet.BUILD")),
   )
+  tf_http_archive(
+      name = "tflite_mobilenet_ssd_quant",
+      sha256 = "a809cd290b4d6a2e8a9d5dad076e0bd695b8091974e0eed1052b480b2f21b6dc",
+      urls = ["https://mirror.bazel.build/storage.googleapis.com/download.tensorflow.org/models/tflite/coco_ssd_mobilenet_v1_0.75_quant_2018_06_29.zip",
+          "https://storage.googleapis.com/download.tensorflow.org/models/tflite/coco_ssd_mobilenet_v1_0.75_quant_2018_06_29.zip",
+      ],
+      build_file = str(Label("//third_party:tflite_mobilenet.BUILD")),
+  )
 
   tf_http_archive(
       name = "tflite_conv_actions_frozen",
diff --git a/third_party/aws.BUILD b/third_party/aws.BUILD
index 2dc921933c310aa9ce2bf21798f1b5143386a12d..5426f79e4650a1ce4dcb4a8408691310c864f06c 100644
--- a/third_party/aws.BUILD
+++ b/third_party/aws.BUILD
@@ -46,6 +46,8 @@ cc_library(
         "aws-cpp-sdk-core/source/utils/xml/**/*.cpp",
         "aws-cpp-sdk-core/source/utils/crypto/*.cpp",
         "aws-cpp-sdk-core/source/utils/crypto/factory/**/*.cpp",
+        "aws-cpp-sdk-kinesis/include/**/*.h",
+        "aws-cpp-sdk-kinesis/source/**/*.cpp",
         "aws-cpp-sdk-s3/include/**/*.h",
         "aws-cpp-sdk-s3/source/**/*.cpp",
     ]),
@@ -72,6 +74,7 @@ cc_library(
     }),
     includes = [
         "aws-cpp-sdk-core/include/",
+        "aws-cpp-sdk-kinesis/include/",
         "aws-cpp-sdk-s3/include/",
     ],
     deps = [
diff --git a/third_party/clang_toolchain/download_clang.bzl b/third_party/clang_toolchain/download_clang.bzl
index b61e901037c3755394452a39b362a9c021673297..ab57b9dfa00094bc2eee727ee98009ce41870379 100644
--- a/third_party/clang_toolchain/download_clang.bzl
+++ b/third_party/clang_toolchain/download_clang.bzl
@@ -35,18 +35,18 @@ def download_clang(repo_ctx, out_folder):
 
   # Latest CLANG_REVISION and CLANG_SUB_REVISION of the Chromiums's release
   # can be found in https://chromium.googlesource.com/chromium/src/tools/clang/+/master/scripts/update.py
-  CLANG_REVISION = '334100'
+  CLANG_REVISION = '336424'
   CLANG_SUB_REVISION = 1
 
   package_version = '%s-%s' % (CLANG_REVISION, CLANG_SUB_REVISION)
 
   checksums = {
       'Linux_x64':
-          '3c57420b591601cd14b5babd74b58fcaefa877112938d70cca6f0a1b0b293ab4',
+          '2ea97e047470da648f5d078af008bce6891287592382cee3d53a1187d996da94',
       'Mac':
-          '97d313996fb97a6138635f963d7ef4efa9f028a8168bb7917cc428b9eab05ebb',
+          'c6e28909cce63ee35e0d51284d9f0f6e8838f7fb8b7a0dc9536c2ea900552df0',
       'Win':
-          '52c1d6d20a0733276597f4ced59d18b545769dbf8beb8c6bdc26a7a862da7fc9',
+          '1299fda7c4378bfb81337f7e5f351c8a1f953f51e0744e2170454b8d722f3db7',
   }
 
   platform_folder = _get_platform_folder(repo_ctx.os.name)
diff --git a/third_party/codegen.BUILD b/third_party/codegen.BUILD
new file mode 100644
index 0000000000000000000000000000000000000000..df436c81635a71421a67fa8d8c84eb8dfcc97d7b
--- /dev/null
+++ b/third_party/codegen.BUILD
@@ -0,0 +1,16 @@
+# -*- mode: python; -*-
+#
+# Description:
+#   Extension to ast that allow ast -> python code generation.
+
+package(default_visibility = ["//visibility:public"])
+
+licenses(["notice"])  # New BSD
+
+exports_files(["LICENSE"])
+
+py_library(
+    name = "com_github_andreif_codegen",
+    srcs = glob(["codegen.py"]),
+    srcs_version = "PY2AND3",
+)
diff --git a/third_party/eigen.BUILD b/third_party/eigen.BUILD
index e54c1a4501d46b6b68a9b8fcc9ce0b1af0535ef4..759f8a9be92e14537d334c3ec37f036d369d8796 100644
--- a/third_party/eigen.BUILD
+++ b/third_party/eigen.BUILD
@@ -69,3 +69,9 @@ cc_library(
     includes = ["."],
     visibility = ["//visibility:public"],
 )
+
+filegroup(
+    name = "eigen_header_files",
+    srcs = EIGEN_MPL2_HEADER_FILES,
+    visibility = ["//visibility:public"],
+)
diff --git a/third_party/eigen3/BUILD b/third_party/eigen3/BUILD
index f661093bc9f68b845f3000b0a931c66773fb3339..203991b50f56086aa76932595f6797ae3bbf58db 100644
--- a/third_party/eigen3/BUILD
+++ b/third_party/eigen3/BUILD
@@ -17,21 +17,23 @@ load("//tensorflow:tensorflow.bzl", "if_mkl")
 # INTEL_MKL end
 load("//tensorflow:tensorflow.bzl", "if_mkl")
 
+EIGEN3_THIRD_PARTY_HEADERS = [
+    "Eigen/Core",
+    "Eigen/LU",
+    "Eigen/Cholesky",
+    "Eigen/Eigenvalues",
+    "Eigen/QR",
+    "Eigen/SVD",
+    "unsupported/Eigen/MatrixFunctions",
+    "unsupported/Eigen/SpecialFunctions",
+    "unsupported/Eigen/CXX11/ThreadPool",
+    "unsupported/Eigen/CXX11/Tensor",
+    "unsupported/Eigen/CXX11/FixedPoint",
+] + glob(["unsupported/Eigen/CXX11/src/FixedPoint/*.h"])
+
 cc_library(
     name = "eigen3",
-    hdrs = glob(["unsupported/Eigen/CXX11/src/FixedPoint/*.h"]) + [
-        "Eigen/Core",
-        "Eigen/LU",
-        "Eigen/Cholesky",
-        "Eigen/Eigenvalues",
-        "Eigen/QR",
-        "Eigen/SVD",
-        "unsupported/Eigen/MatrixFunctions",
-        "unsupported/Eigen/SpecialFunctions",
-        "unsupported/Eigen/CXX11/ThreadPool",
-        "unsupported/Eigen/CXX11/Tensor",
-        "unsupported/Eigen/CXX11/FixedPoint",
-    ],
+    hdrs = EIGEN3_THIRD_PARTY_HEADERS,
     includes = if_mkl(["./mkl_include"]),
     visibility = ["//visibility:public"],
     deps = [
@@ -48,3 +50,35 @@ filegroup(
     ),
     visibility = ["//tensorflow:__subpackages__"],
 )
+
+filegroup(
+    name = "eigen_third_party_header_files",
+    srcs = EIGEN3_THIRD_PARTY_HEADERS,
+    visibility = ["//visibility:public"],
+)
+
+genrule(
+    name = "install_eigen_headers",
+    srcs = [
+        "@eigen_archive//:eigen_header_files",
+        ":eigen_third_party_header_files",
+    ],
+    outs = ["include"],
+    cmd = """
+    mkdir $@
+    for f in $(locations @eigen_archive//:eigen_header_files) ; do
+      d="$${f%/*}"
+      d="$${d#*external/eigen_archive/}"
+
+      mkdir -p "$@/$${d}"
+      cp "$${f}" "$@/$${d}/"
+    done
+
+    for f in $(locations :eigen_third_party_header_files) ; do
+      d="$${f%/*}"
+
+      mkdir -p "$@/$${d}"
+      cp "$${f}" "$@/$${d}/"
+    done
+    """,
+)
diff --git a/third_party/googleapis.BUILD b/third_party/googleapis.BUILD
new file mode 100644
index 0000000000000000000000000000000000000000..95e999af1886576317aa59d133e8d5c88ba368d3
--- /dev/null
+++ b/third_party/googleapis.BUILD
@@ -0,0 +1,45 @@
+# Copyright 2018 Google LLC
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+package(default_visibility = ["//visibility:public"])
+licenses(["notice"])  # Apache 2.0
+exports_files(["LICENSE"])
+
+load("@protobuf_archive//:protobuf.bzl", "cc_proto_library")
+
+cc_proto_library(
+    name = "bigtable_protos",
+    srcs = [
+        "google/bigtable/admin/v2/bigtable_instance_admin.proto",
+        "google/bigtable/admin/v2/bigtable_table_admin.proto",
+        "google/bigtable/admin/v2/common.proto",
+        "google/bigtable/admin/v2/instance.proto",
+        "google/bigtable/admin/v2/table.proto",
+        "google/bigtable/v2/bigtable.proto",
+        "google/bigtable/v2/data.proto",
+        "google/iam/v1/iam_policy.proto",
+        "google/iam/v1/policy.proto",
+        "google/longrunning/operations.proto",
+        "google/rpc/status.proto",
+        "google/rpc/error_details.proto",
+        "google/api/annotations.proto",
+        "google/api/auth.proto",
+        "google/api/http.proto",
+    ],
+    include = ".",
+    protoc = "@protobuf_archive//:protoc",
+    default_runtime = "@protobuf_archive//:protobuf",
+    deps = ["@protobuf_archive//:cc_wkt_protos"],
+    use_grpc_plugin = True,
+)
diff --git a/third_party/gpus/crosstool/BUILD.tpl b/third_party/gpus/crosstool/BUILD.tpl
index 98cb326572e75ac3ea15a656d821c1eade53d313..f638756d2373d3a0d85633be72654091c7982f49 100644
--- a/third_party/gpus/crosstool/BUILD.tpl
+++ b/third_party/gpus/crosstool/BUILD.tpl
@@ -7,6 +7,7 @@ cc_toolchain_suite(
     toolchains = {
         "local|compiler": ":cc-compiler-local",
         "darwin|compiler": ":cc-compiler-darwin",
+        "x64_windows|msvc-cl": ":cc-compiler-windows",
     },
 )
 
@@ -42,6 +43,20 @@ cc_toolchain(
     supports_param_files = 0,
 )
 
+cc_toolchain(
+    name = "cc-compiler-windows",
+    all_files = "%{win_linker_files}",
+    compiler_files = ":empty",
+    cpu = "x64_windows",
+    dwp_files = ":empty",
+    dynamic_runtime_libs = [":empty"],
+    linker_files = "%{win_linker_files}",
+    objcopy_files = ":empty",
+    static_runtime_libs = [":empty"],
+    strip_files = ":empty",
+    supports_param_files = 1,
+)
+
 filegroup(
     name = "empty",
     srcs = [],
@@ -51,3 +66,8 @@ filegroup(
     name = "crosstool_wrapper_driver_is_not_gcc",
     srcs = ["clang/bin/crosstool_wrapper_driver_is_not_gcc"],
 )
+
+filegroup(
+    name = "windows_msvc_wrapper_files",
+    srcs = glob(["windows/msvc_*"]),
+)
diff --git a/third_party/gpus/crosstool/CROSSTOOL.tpl b/third_party/gpus/crosstool/CROSSTOOL.tpl
index 1424ff6511dfe0e7e8eef2843201e825e09a91f1..3972c96a2f726127cd7112265eef4d2a794ed0fc 100644
--- a/third_party/gpus/crosstool/CROSSTOOL.tpl
+++ b/third_party/gpus/crosstool/CROSSTOOL.tpl
@@ -22,6 +22,10 @@ default_toolchain {
   cpu: "ppc"
   toolchain_identifier: "local_linux"
 }
+default_toolchain {
+  cpu: "x64_windows"
+  toolchain_identifier: "local_windows"
+}
 
 toolchain {
   abi_version: "local"
@@ -537,3 +541,868 @@ toolchain {
 
 %{host_compiler_includes}
 }
+
+toolchain {
+  toolchain_identifier: "local_windows"
+  host_system_name: "local"
+  target_system_name: "local"
+
+  abi_version: "local"
+  abi_libc_version: "local"
+  target_cpu: "x64_windows"
+  compiler: "msvc-cl"
+  target_libc: "msvcrt"
+
+%{cxx_builtin_include_directory}
+
+  tool_path {
+    name: "ar"
+    path: "%{msvc_lib_path}"
+  }
+  tool_path {
+    name: "ml"
+    path: "%{msvc_ml_path}"
+  }
+  tool_path {
+    name: "cpp"
+    path: "%{msvc_cl_path}"
+  }
+  tool_path {
+    name: "gcc"
+    path: "%{msvc_cl_path}"
+  }
+  tool_path {
+    name: "gcov"
+    path: "wrapper/bin/msvc_nop.bat"
+  }
+  tool_path {
+    name: "ld"
+    path: "%{msvc_link_path}"
+  }
+  tool_path {
+    name: "nm"
+    path: "wrapper/bin/msvc_nop.bat"
+  }
+  tool_path {
+    name: "objcopy"
+    path: "wrapper/bin/msvc_nop.bat"
+  }
+  tool_path {
+    name: "objdump"
+    path: "wrapper/bin/msvc_nop.bat"
+  }
+  tool_path {
+    name: "strip"
+    path: "wrapper/bin/msvc_nop.bat"
+  }
+  supports_interface_shared_objects: true
+
+  # TODO(pcloudy): Review those flags below, they should be defined by cl.exe
+  compiler_flag: "/DCOMPILER_MSVC"
+
+  # Don't define min/max macros in windows.h.
+  compiler_flag: "/DNOMINMAX"
+
+  # Platform defines.
+  compiler_flag: "/D_WIN32_WINNT=0x0600"
+  # Turn off warning messages.
+  compiler_flag: "/D_CRT_SECURE_NO_DEPRECATE"
+  compiler_flag: "/D_CRT_SECURE_NO_WARNINGS"
+  compiler_flag: "/D_SILENCE_STDEXT_HASH_DEPRECATION_WARNINGS"
+
+  # Useful options to have on for compilation.
+  # Increase the capacity of object files to 2^32 sections.
+  compiler_flag: "/bigobj"
+  # Allocate 500MB for precomputed headers.
+  compiler_flag: "/Zm500"
+  # Use unsigned char by default.
+  compiler_flag: "/J"
+  # Use function level linking.
+  compiler_flag: "/Gy"
+  # Use string pooling.
+  compiler_flag: "/GF"
+  # Catch C++ exceptions only and tell the compiler to assume that functions declared
+  # as extern "C" never throw a C++ exception.
+  compiler_flag: "/EHsc"
+
+  # Globally disabled warnings.
+  # Don't warn about elements of array being be default initialized.
+  compiler_flag: "/wd4351"
+  # Don't warn about no matching delete found.
+  compiler_flag: "/wd4291"
+  # Don't warn about diamond inheritance patterns.
+  compiler_flag: "/wd4250"
+  # Don't warn about insecure functions (e.g. non _s functions).
+  compiler_flag: "/wd4996"
+
+  linker_flag: "/MACHINE:X64"
+
+  feature {
+    name: "no_legacy_features"
+  }
+
+  # Suppress startup banner.
+  feature {
+    name: "nologo"
+    flag_set {
+      action: "c-compile"
+      action: "c++-compile"
+      action: "c++-module-compile"
+      action: "c++-module-codegen"
+      action: "c++-header-parsing"
+      action: "assemble"
+      action: "preprocess-assemble"
+      action: "c++-link-executable"
+      action: "c++-link-dynamic-library"
+      action: "c++-link-nodeps-dynamic-library"
+      action: "c++-link-static-library"
+      flag_group {
+        flag: "/nologo"
+      }
+    }
+  }
+
+  feature {
+    name: 'has_configured_linker_path'
+  }
+
+  # This feature indicates strip is not supported, building stripped binary will just result a copy of orignial binary
+  feature {
+    name: 'no_stripping'
+  }
+
+  # This feature indicates this is a toolchain targeting Windows.
+  feature {
+    name: 'targets_windows'
+    implies: 'copy_dynamic_libraries_to_binary'
+    enabled: true
+  }
+
+  feature {
+    name: 'copy_dynamic_libraries_to_binary'
+  }
+
+  action_config {
+    config_name: 'assemble'
+    action_name: 'assemble'
+    tool {
+      tool_path: '%{msvc_ml_path}'
+    }
+    implies: 'compiler_input_flags'
+    implies: 'compiler_output_flags'
+    implies: 'nologo'
+    implies: 'msvc_env'
+    implies: 'sysroot'
+  }
+
+  action_config {
+    config_name: 'preprocess-assemble'
+    action_name: 'preprocess-assemble'
+    tool {
+      tool_path: '%{msvc_ml_path}'
+    }
+    implies: 'compiler_input_flags'
+    implies: 'compiler_output_flags'
+    implies: 'nologo'
+    implies: 'msvc_env'
+    implies: 'sysroot'
+  }
+
+  action_config {
+    config_name: 'c-compile'
+    action_name: 'c-compile'
+    tool {
+      tool_path: '%{msvc_cl_path}'
+    }
+    implies: 'compiler_input_flags'
+    implies: 'compiler_output_flags'
+    implies: 'legacy_compile_flags'
+    implies: 'nologo'
+    implies: 'msvc_env'
+    implies: 'parse_showincludes'
+    implies: 'user_compile_flags'
+    implies: 'sysroot'
+    implies: 'unfiltered_compile_flags'
+  }
+
+  action_config {
+    config_name: 'c++-compile'
+    action_name: 'c++-compile'
+    tool {
+      tool_path: '%{msvc_cl_path}'
+    }
+    implies: 'compiler_input_flags'
+    implies: 'compiler_output_flags'
+    implies: 'legacy_compile_flags'
+    implies: 'nologo'
+    implies: 'msvc_env'
+    implies: 'parse_showincludes'
+    implies: 'user_compile_flags'
+    implies: 'sysroot'
+    implies: 'unfiltered_compile_flags'
+  }
+
+  action_config {
+    config_name: 'c++-link-executable'
+    action_name: 'c++-link-executable'
+    tool {
+      tool_path: '%{msvc_link_path}'
+    }
+    implies: 'nologo'
+    implies: 'linkstamps'
+    implies: 'output_execpath_flags'
+    implies: 'input_param_flags'
+    implies: 'user_link_flags'
+    implies: 'legacy_link_flags'
+    implies: 'linker_subsystem_flag'
+    implies: 'linker_param_file'
+    implies: 'msvc_env'
+    implies: 'no_stripping'
+  }
+
+  action_config {
+    config_name: 'c++-link-dynamic-library'
+    action_name: 'c++-link-dynamic-library'
+    tool {
+      tool_path: '%{msvc_link_path}'
+    }
+    implies: 'nologo'
+    implies: 'shared_flag'
+    implies: 'linkstamps'
+    implies: 'output_execpath_flags'
+    implies: 'input_param_flags'
+    implies: 'user_link_flags'
+    implies: 'legacy_link_flags'
+    implies: 'linker_subsystem_flag'
+    implies: 'linker_param_file'
+    implies: 'msvc_env'
+    implies: 'no_stripping'
+    implies: 'has_configured_linker_path'
+    implies: 'def_file'
+  }
+
+  action_config {
+      config_name: 'c++-link-nodeps-dynamic-library'
+      action_name: 'c++-link-nodeps-dynamic-library'
+      tool {
+        tool_path: '%{msvc_link_path}'
+      }
+      implies: 'nologo'
+      implies: 'shared_flag'
+      implies: 'linkstamps'
+      implies: 'output_execpath_flags'
+      implies: 'input_param_flags'
+      implies: 'user_link_flags'
+      implies: 'legacy_link_flags'
+      implies: 'linker_subsystem_flag'
+      implies: 'linker_param_file'
+      implies: 'msvc_env'
+      implies: 'no_stripping'
+      implies: 'has_configured_linker_path'
+      implies: 'def_file'
+    }
+
+  action_config {
+    config_name: 'c++-link-static-library'
+    action_name: 'c++-link-static-library'
+    tool {
+      tool_path: '%{msvc_lib_path}'
+    }
+    implies: 'nologo'
+    implies: 'archiver_flags'
+    implies: 'input_param_flags'
+    implies: 'linker_param_file'
+    implies: 'msvc_env'
+  }
+
+  # TODO(b/65151735): Remove legacy_compile_flags feature when legacy fields are
+  # not used in this crosstool
+  feature {
+    name: 'legacy_compile_flags'
+    flag_set {
+      expand_if_all_available: 'legacy_compile_flags'
+      action: 'preprocess-assemble'
+      action: 'c-compile'
+      action: 'c++-compile'
+      action: 'c++-header-parsing'
+      action: 'c++-module-compile'
+      action: 'c++-module-codegen'
+      flag_group {
+        iterate_over: 'legacy_compile_flags'
+        flag: '%{legacy_compile_flags}'
+      }
+    }
+  }
+
+  feature {
+    name: "msvc_env"
+    env_set {
+      action: "c-compile"
+      action: "c++-compile"
+      action: "c++-module-compile"
+      action: "c++-module-codegen"
+      action: "c++-header-parsing"
+      action: "assemble"
+      action: "preprocess-assemble"
+      action: "c++-link-executable"
+      action: "c++-link-dynamic-library"
+      action: "c++-link-nodeps-dynamic-library"
+      action: "c++-link-static-library"
+      env_entry {
+        key: "PATH"
+        value: "%{msvc_env_path}"
+      }
+      env_entry {
+        key: "INCLUDE"
+        value: "%{msvc_env_include}"
+      }
+      env_entry {
+        key: "LIB"
+        value: "%{msvc_env_lib}"
+      }
+      env_entry {
+        key: "TMP"
+        value: "%{msvc_env_tmp}"
+      }
+      env_entry {
+        key: "TEMP"
+        value: "%{msvc_env_tmp}"
+      }
+    }
+  }
+
+  feature {
+    name: 'include_paths'
+    flag_set {
+      action: "assemble"
+      action: 'preprocess-assemble'
+      action: 'c-compile'
+      action: 'c++-compile'
+      action: 'c++-header-parsing'
+      action: 'c++-module-compile'
+      flag_group {
+        iterate_over: 'quote_include_paths'
+        flag: '/I%{quote_include_paths}'
+      }
+      flag_group {
+        iterate_over: 'include_paths'
+        flag: '/I%{include_paths}'
+      }
+      flag_group {
+        iterate_over: 'system_include_paths'
+        flag: '/I%{system_include_paths}'
+      }
+    }
+  }
+
+  feature {
+    name: "preprocessor_defines"
+    flag_set {
+      action: "assemble"
+      action: "preprocess-assemble"
+      action: "c-compile"
+      action: "c++-compile"
+      action: "c++-header-parsing"
+      action: "c++-module-compile"
+      flag_group {
+        flag: "/D%{preprocessor_defines}"
+        iterate_over: "preprocessor_defines"
+      }
+    }
+  }
+
+  # Tell Bazel to parse the output of /showIncludes
+  feature {
+    name: 'parse_showincludes'
+    flag_set {
+      action: 'preprocess-assemble'
+      action: 'c-compile'
+      action: 'c++-compile'
+      action: 'c++-module-compile'
+      action: 'c++-header-parsing'
+      flag_group {
+        flag: "/showIncludes"
+      }
+    }
+  }
+
+
+  feature {
+    name: 'generate_pdb_file'
+    requires: {
+      feature: 'dbg'
+    }
+    requires: {
+      feature: 'fastbuild'
+    }
+  }
+
+  feature {
+    name: 'shared_flag'
+    flag_set {
+      action: 'c++-link-dynamic-library'
+      action: "c++-link-nodeps-dynamic-library"
+      flag_group {
+        flag: '/DLL'
+      }
+    }
+  }
+
+  feature {
+    name: 'linkstamps'
+    flag_set {
+      action: 'c++-link-executable'
+      action: 'c++-link-dynamic-library'
+      action: "c++-link-nodeps-dynamic-library"
+      expand_if_all_available: 'linkstamp_paths'
+      flag_group {
+        iterate_over: 'linkstamp_paths'
+        flag: '%{linkstamp_paths}'
+      }
+    }
+  }
+
+  feature {
+    name: 'output_execpath_flags'
+    flag_set {
+      expand_if_all_available: 'output_execpath'
+      action: 'c++-link-executable'
+      action: 'c++-link-dynamic-library'
+      action: "c++-link-nodeps-dynamic-library"
+      flag_group {
+        flag: '/OUT:%{output_execpath}'
+      }
+    }
+  }
+
+  feature {
+    name: 'archiver_flags'
+    flag_set {
+      expand_if_all_available: 'output_execpath'
+      action: 'c++-link-static-library'
+      flag_group {
+        flag: '/OUT:%{output_execpath}'
+      }
+    }
+  }
+
+  feature {
+    name: 'input_param_flags'
+    flag_set {
+      expand_if_all_available: 'interface_library_output_path'
+      action: 'c++-link-dynamic-library'
+      action: "c++-link-nodeps-dynamic-library"
+      flag_group {
+        flag: "/IMPLIB:%{interface_library_output_path}"
+      }
+    }
+    flag_set {
+      expand_if_all_available: 'libopts'
+      action: 'c++-link-executable'
+      action: 'c++-link-dynamic-library'
+      action: "c++-link-nodeps-dynamic-library"
+      flag_group {
+        iterate_over: 'libopts'
+        flag: '%{libopts}'
+      }
+    }
+    flag_set {
+      expand_if_all_available: 'libraries_to_link'
+      action: 'c++-link-executable'
+      action: 'c++-link-dynamic-library'
+      action: "c++-link-nodeps-dynamic-library"
+      action: 'c++-link-static-library'
+      flag_group {
+        iterate_over: 'libraries_to_link'
+        flag_group {
+          expand_if_equal: {
+            variable: 'libraries_to_link.type'
+            value: 'object_file_group'
+          }
+          iterate_over: 'libraries_to_link.object_files'
+          flag_group {
+            flag: '%{libraries_to_link.object_files}'
+          }
+        }
+        flag_group {
+          expand_if_equal: {
+            variable: 'libraries_to_link.type'
+            value: 'object_file'
+          }
+          flag_group {
+            flag: '%{libraries_to_link.name}'
+          }
+        }
+        flag_group {
+          expand_if_equal: {
+            variable: 'libraries_to_link.type'
+            value: 'interface_library'
+          }
+          flag_group {
+            flag: '%{libraries_to_link.name}'
+          }
+        }
+        flag_group {
+          expand_if_equal: {
+            variable: 'libraries_to_link.type'
+            value: 'static_library'
+          }
+          flag_group {
+            expand_if_false: 'libraries_to_link.is_whole_archive'
+            flag: '%{libraries_to_link.name}'
+          }
+          flag_group {
+            expand_if_true: 'libraries_to_link.is_whole_archive'
+            flag: '/WHOLEARCHIVE:%{libraries_to_link.name}'
+          }
+        }
+      }
+    }
+  }
+
+  # Since this feature is declared earlier in the CROSSTOOL than
+  # "user_link_flags", this feature will be applied prior to it anwyhere they
+  # are both implied. And since "user_link_flags" contains the linkopts from
+  # the build rule, this allows the user to override the /SUBSYSTEM in the BUILD
+  # file.
+  feature {
+    name: 'linker_subsystem_flag'
+    flag_set {
+      action: 'c++-link-executable'
+      action: 'c++-link-dynamic-library'
+      action: "c++-link-nodeps-dynamic-library"
+      flag_group {
+        flag: '/SUBSYSTEM:CONSOLE'
+      }
+    }
+  }
+
+  # The "user_link_flags" contains user-defined linkopts (from build rules)
+  # so it should be defined after features that declare user-overridable flags.
+  # For example the "linker_subsystem_flag" defines a default "/SUBSYSTEM" flag
+  # but we want to let the user override it, therefore "link_flag_subsystem" is
+  # defined earlier in the CROSSTOOL file than "user_link_flags".
+  feature {
+    name: 'user_link_flags'
+    flag_set {
+      expand_if_all_available: 'user_link_flags'
+      action: 'c++-link-executable'
+      action: 'c++-link-dynamic-library'
+      action: "c++-link-nodeps-dynamic-library"
+      flag_group {
+        iterate_over: 'user_link_flags'
+        flag: '%{user_link_flags}'
+      }
+    }
+  }
+  feature {
+    name: 'legacy_link_flags'
+    flag_set {
+      expand_if_all_available: 'legacy_link_flags'
+      action: 'c++-link-executable'
+      action: 'c++-link-dynamic-library'
+      action: "c++-link-nodeps-dynamic-library"
+      flag_group {
+        iterate_over: 'legacy_link_flags'
+        flag: '%{legacy_link_flags}'
+      }
+    }
+  }
+
+  feature {
+    name: 'linker_param_file'
+    flag_set {
+      expand_if_all_available: 'linker_param_file'
+      action: 'c++-link-executable'
+      action: 'c++-link-dynamic-library'
+      action: "c++-link-nodeps-dynamic-library"
+      action: 'c++-link-static-library'
+      flag_group {
+        flag: '@%{linker_param_file}'
+      }
+    }
+  }
+
+  feature {
+    name: 'static_link_msvcrt'
+  }
+
+  feature {
+    name: 'static_link_msvcrt_no_debug'
+    flag_set {
+      action: 'c-compile'
+      action: 'c++-compile'
+      flag_group {
+        flag: "/MT"
+      }
+    }
+    flag_set {
+      action: 'c++-link-executable'
+      action: 'c++-link-dynamic-library'
+      action: "c++-link-nodeps-dynamic-library"
+      flag_group {
+        flag: "/DEFAULTLIB:libcmt.lib"
+      }
+    }
+    requires: { feature: 'fastbuild'}
+    requires: { feature: 'opt'}
+  }
+
+  feature {
+    name: 'dynamic_link_msvcrt_no_debug'
+    flag_set {
+      action: 'c-compile'
+      action: 'c++-compile'
+      flag_group {
+        flag: "/MD"
+      }
+    }
+    flag_set {
+      action: 'c++-link-executable'
+      action: 'c++-link-dynamic-library'
+      action: "c++-link-nodeps-dynamic-library"
+      flag_group {
+        flag: "/DEFAULTLIB:msvcrt.lib"
+      }
+    }
+    requires: { feature: 'fastbuild'}
+    requires: { feature: 'opt'}
+  }
+
+  feature {
+    name: 'static_link_msvcrt_debug'
+    flag_set {
+      action: 'c-compile'
+      action: 'c++-compile'
+      flag_group {
+        flag: "/MTd"
+      }
+    }
+    flag_set {
+      action: 'c++-link-executable'
+      action: 'c++-link-dynamic-library'
+      action: "c++-link-nodeps-dynamic-library"
+      flag_group {
+        flag: "/DEFAULTLIB:libcmtd.lib"
+      }
+    }
+    requires: { feature: 'dbg'}
+  }
+
+  feature {
+    name: 'dynamic_link_msvcrt_debug'
+    flag_set {
+      action: 'c-compile'
+      action: 'c++-compile'
+      flag_group {
+        flag: "/MDd"
+      }
+    }
+    flag_set {
+      action: 'c++-link-executable'
+      action: 'c++-link-dynamic-library'
+      action: "c++-link-nodeps-dynamic-library"
+      flag_group {
+        flag: "/DEFAULTLIB:msvcrtd.lib"
+      }
+    }
+    requires: { feature: 'dbg'}
+  }
+
+  feature {
+    name: 'dbg'
+    flag_set {
+      action: 'c-compile'
+      action: 'c++-compile'
+      flag_group {
+        flag: "/Od"
+        flag: "/Z7"
+        flag: "/DDEBUG"
+      }
+    }
+    flag_set {
+      action: 'c++-link-executable'
+      action: 'c++-link-dynamic-library'
+      action: "c++-link-nodeps-dynamic-library"
+      flag_group {
+        flag: "/DEBUG:FULL"
+        flag: "/INCREMENTAL:NO"
+      }
+    }
+    implies: 'generate_pdb_file'
+  }
+
+  feature {
+    name: 'fastbuild'
+    flag_set {
+      action: 'c-compile'
+      action: 'c++-compile'
+      flag_group {
+        flag: "/Od"
+        flag: "/Z7"
+        flag: "/DDEBUG"
+      }
+    }
+    flag_set {
+      action: 'c++-link-executable'
+      action: 'c++-link-dynamic-library'
+      action: "c++-link-nodeps-dynamic-library"
+      flag_group {
+        flag: "/DEBUG:FASTLINK"
+        flag: "/INCREMENTAL:NO"
+      }
+    }
+    implies: 'generate_pdb_file'
+  }
+
+  feature {
+    name: 'opt'
+    flag_set {
+      action: 'c-compile'
+      action: 'c++-compile'
+      flag_group {
+        flag: "/O2"
+        flag: "/DNDEBUG"
+      }
+    }
+  }
+
+  feature {
+    name: 'user_compile_flags'
+    flag_set {
+      expand_if_all_available: 'user_compile_flags'
+      action: 'preprocess-assemble'
+      action: 'c-compile'
+      action: 'c++-compile'
+      action: 'c++-header-parsing'
+      action: 'c++-module-compile'
+      action: 'c++-module-codegen'
+      flag_group {
+        iterate_over: 'user_compile_flags'
+        flag: '%{user_compile_flags}'
+      }
+    }
+  }
+
+  feature {
+    name: 'sysroot'
+    flag_set {
+      expand_if_all_available: 'sysroot'
+      action: 'assemble'
+      action: 'preprocess-assemble'
+      action: 'c-compile'
+      action: 'c++-compile'
+      action: 'c++-header-parsing'
+      action: 'c++-module-compile'
+      action: 'c++-module-codegen'
+      action: 'c++-link-executable'
+      action: 'c++-link-dynamic-library'
+      action: "c++-link-nodeps-dynamic-library"
+      flag_group {
+        iterate_over: 'sysroot'
+        flag: '--sysroot=%{sysroot}'
+      }
+    }
+  }
+
+  feature {
+    name: 'unfiltered_compile_flags'
+    flag_set {
+      expand_if_all_available: 'unfiltered_compile_flags'
+      action: 'preprocess-assemble'
+      action: 'c-compile'
+      action: 'c++-compile'
+      action: 'c++-header-parsing'
+      action: 'c++-module-compile'
+      action: 'c++-module-codegen'
+      flag_group {
+        iterate_over: 'unfiltered_compile_flags'
+        flag: '%{unfiltered_compile_flags}'
+      }
+    }
+  }
+
+  feature {
+    name: 'compiler_output_flags'
+    flag_set {
+      action: 'assemble'
+      flag_group {
+        expand_if_all_available: 'output_file'
+        expand_if_none_available: 'output_assembly_file'
+        expand_if_none_available: 'output_preprocess_file'
+        flag: '/Fo%{output_file}'
+        flag: '/Zi'
+      }
+    }
+    flag_set {
+      action: 'preprocess-assemble'
+      action: 'c-compile'
+      action: 'c++-compile'
+      action: 'c++-header-parsing'
+      action: 'c++-module-compile'
+      action: 'c++-module-codegen'
+      flag_group {
+        expand_if_all_available: 'output_file'
+        expand_if_none_available: 'output_assembly_file'
+        expand_if_none_available: 'output_preprocess_file'
+        flag: '/Fo%{output_file}'
+      }
+      flag_group {
+        expand_if_all_available: 'output_file'
+        expand_if_all_available: 'output_assembly_file'
+        flag: '/Fa%{output_file}'
+      }
+      flag_group {
+        expand_if_all_available: 'output_file'
+        expand_if_all_available: 'output_preprocess_file'
+        flag: '/P'
+        flag: '/Fi%{output_file}'
+      }
+    }
+  }
+
+  feature {
+    name: 'compiler_input_flags'
+    flag_set {
+      action: 'assemble'
+      action: 'preprocess-assemble'
+      action: 'c-compile'
+      action: 'c++-compile'
+      action: 'c++-header-parsing'
+      action: 'c++-module-compile'
+      action: 'c++-module-codegen'
+      flag_group {
+        expand_if_all_available: 'source_file'
+        flag: '/c'
+        flag: '%{source_file}'
+      }
+    }
+  }
+
+  feature {
+    name : 'def_file',
+    flag_set {
+      expand_if_all_available: 'def_file_path'
+      action: 'c++-link-executable'
+      action: 'c++-link-dynamic-library'
+      action: "c++-link-nodeps-dynamic-library"
+      flag_group {
+        flag: "/DEF:%{def_file_path}"
+        # We can specify a different DLL name in DEF file, /ignore:4070 suppresses
+        # the warning message about DLL name doesn't match the default one.
+        # See https://msdn.microsoft.com/en-us/library/sfkk2fz7.aspx
+        flag: "/ignore:4070"
+      }
+    }
+  }
+
+  feature {
+    name: 'windows_export_all_symbols'
+  }
+
+  feature {
+    name: 'no_windows_export_all_symbols'
+  }
+
+  linking_mode_flags { mode: DYNAMIC }
+}
diff --git a/third_party/gpus/crosstool/clang/bin/crosstool_wrapper_driver_is_not_gcc.tpl b/third_party/gpus/crosstool/clang/bin/crosstool_wrapper_driver_is_not_gcc.tpl
index 2558f46fd55c35b5089cc0119f2654f598e5128a..f4f4d0ee964142b2aa6e010ad5409494438733ea 100755
--- a/third_party/gpus/crosstool/clang/bin/crosstool_wrapper_driver_is_not_gcc.tpl
+++ b/third_party/gpus/crosstool/clang/bin/crosstool_wrapper_driver_is_not_gcc.tpl
@@ -175,6 +175,11 @@ def InvokeNvcc(argv, log=False):
   # any other reliable way to just get the list of source files to be compiled.
   src_files = GetOptionValue(argv, 'c')
 
+  # Pass -w through from host to nvcc, but don't do anything fancier with
+  # warnings-related flags, since they're not necessarily the same across
+  # compilers.
+  warning_options = ' -w' if '-w' in argv else ''
+
   if len(src_files) == 0:
     return 1
   if len(out_file) != 1:
@@ -205,6 +210,7 @@ def InvokeNvcc(argv, log=False):
   nvccopts += defines
   nvccopts += std_options
   nvccopts += m_options
+  nvccopts += warning_options
 
   if depfiles:
     # Generate the dependency file
diff --git a/third_party/gpus/crosstool/windows/msvc_wrapper_for_nvcc.bat.tpl b/third_party/gpus/crosstool/windows/msvc_wrapper_for_nvcc.bat.tpl
new file mode 100644
index 0000000000000000000000000000000000000000..8f8fb3e4231bf1b689cf9b21c53e990d5b9ee354
--- /dev/null
+++ b/third_party/gpus/crosstool/windows/msvc_wrapper_for_nvcc.bat.tpl
@@ -0,0 +1,20 @@
+:: Copyright 2015 The TensorFlow Authors. All Rights Reserved.
+::
+:: Licensed under the Apache License, Version 2.0 (the "License");
+:: you may not use this file except in compliance with the License.
+:: You may obtain a copy of the License at
+::
+::     http://www.apache.org/licenses/LICENSE-2.0
+::
+:: Unless required by applicable law or agreed to in writing, software
+:: distributed under the License is distributed on an "AS IS" BASIS,
+:: WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+:: See the License for the specific language governing permissions and
+:: limitations under the License.
+:: =============================================================================
+
+:: Invoke msvc_wrapper_for_nvcc.py, which is located in the same directory.
+@echo OFF
+set arg0=%~0
+for %%F in ("%arg0%") do set DRIVER_BIN=%%~dpF
+"%{python_binary}" -B "%DRIVER_BIN%\msvc_wrapper_for_nvcc.py" %*
diff --git a/third_party/gpus/crosstool/windows/msvc_wrapper_for_nvcc.py.tpl b/third_party/gpus/crosstool/windows/msvc_wrapper_for_nvcc.py.tpl
new file mode 100644
index 0000000000000000000000000000000000000000..1a09756813e8322b42911dfe7ac80f626e34f98b
--- /dev/null
+++ b/third_party/gpus/crosstool/windows/msvc_wrapper_for_nvcc.py.tpl
@@ -0,0 +1,192 @@
+#!/usr/bin/env python
+# Copyright 2015 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+
+"""Crosstool wrapper for compiling CUDA programs with nvcc on Windows.
+
+DESCRIPTION:
+  This script is the Windows version of //third_party/gpus/crosstool/crosstool_wrapper_is_not_gcc
+"""
+
+from __future__ import print_function
+
+from argparse import ArgumentParser
+import os
+import subprocess
+import re
+import sys
+import pipes
+
+# Template values set by cuda_autoconf.
+CPU_COMPILER = ('%{cpu_compiler}')
+GCC_HOST_COMPILER_PATH = ('%{gcc_host_compiler_path}')
+
+NVCC_PATH = '%{nvcc_path}'
+NVCC_VERSION = '%{cuda_version}'
+NVCC_TEMP_DIR = "%{nvcc_tmp_dir}"
+supported_cuda_compute_capabilities = [ %{cuda_compute_capabilities} ]
+
+def Log(s):
+  print('gpus/crosstool: {0}'.format(s))
+
+
+def GetOptionValue(argv, option):
+  """Extract the list of values for option from options.
+
+  Args:
+    option: The option whose value to extract, without the leading '/'.
+
+  Returns:
+    1. A list of values, either directly following the option,
+    (eg., /opt val1 val2) or values collected from multiple occurrences of
+    the option (eg., /opt val1 /opt val2).
+    2. The leftover options.
+  """
+
+  parser = ArgumentParser(prefix_chars='/')
+  parser.add_argument('/' + option, nargs='*', action='append')
+  args, leftover = parser.parse_known_args(argv)
+  if args and vars(args)[option]:
+    return (sum(vars(args)[option], []), leftover)
+  return ([], leftover)
+
+def _update_options(nvcc_options):
+  if NVCC_VERSION in ("7.0",):
+    return nvcc_options
+
+  update_options = { "relaxed-constexpr" : "expt-relaxed-constexpr" }
+  return [ update_options[opt] if opt in update_options else opt
+                    for opt in nvcc_options ]
+
+def GetNvccOptions(argv):
+  """Collect the -nvcc_options values from argv.
+
+  Args:
+    argv: A list of strings, possibly the argv passed to main().
+
+  Returns:
+    1. The string that can be passed directly to nvcc.
+    2. The leftover options.
+  """
+
+  parser = ArgumentParser()
+  parser.add_argument('-nvcc_options', nargs='*', action='append')
+
+  args, leftover = parser.parse_known_args(argv)
+
+  if args.nvcc_options:
+    options = _update_options(sum(args.nvcc_options, []))
+    return (['--' + a for a in options], leftover)
+  return ([], leftover)
+
+
+def InvokeNvcc(argv, log=False):
+  """Call nvcc with arguments assembled from argv.
+
+  Args:
+    argv: A list of strings, possibly the argv passed to main().
+    log: True if logging is requested.
+
+  Returns:
+    The return value of calling os.system('nvcc ' + args)
+  """
+
+  src_files = [f for f in argv if
+               re.search('\.cpp$|\.cc$|\.c$|\.cxx$|\.C$', f)]
+  if len(src_files) == 0:
+    raise Error('No source files found for cuda compilation.')
+
+  out_file = [ f for f in argv if f.startswith('/Fo') ]
+  if len(out_file) != 1:
+    raise Error('Please sepecify exactly one output file for cuda compilation.')
+  out = ['-o', out_file[0][len('/Fo'):]]
+
+  nvcc_compiler_options, argv = GetNvccOptions(argv)
+
+  opt_option, argv = GetOptionValue(argv, 'O')
+  opt = ['-g', '-G']
+  if (len(opt_option) > 0 and opt_option[0] != 'd'):
+    opt = ['-O2']
+
+  include_options, argv = GetOptionValue(argv, 'I')
+  includes = ["-I " + include for include in include_options]
+
+  defines, argv = GetOptionValue(argv, 'D')
+  defines = ['-D' + define for define in defines]
+
+  undefines, argv = GetOptionValue(argv, 'U')
+  undefines = ['-U' + define for define in undefines]
+
+  # The rest of the unrecongized options should be passed to host compiler
+  host_compiler_options = [option for option in argv if option not in (src_files + out_file)]
+
+  m_options = ["-m64"]
+
+  nvccopts = ['-D_FORCE_INLINES']
+  for capability in supported_cuda_compute_capabilities:
+    capability = capability.replace('.', '')
+    nvccopts += [r'-gencode=arch=compute_%s,"code=sm_%s,compute_%s"' % (
+        capability, capability, capability)]
+  nvccopts += nvcc_compiler_options
+  nvccopts += undefines
+  nvccopts += defines
+  nvccopts += m_options
+  nvccopts += ['--compiler-options="' + " ".join(host_compiler_options) + '"']
+  nvccopts += ['-x', 'cu'] + opt + includes + out + ['-c'] + src_files
+  # If we don't specify --keep-dir, nvcc will generate intermediate files under TEMP
+  # Put them under NVCC_TEMP_DIR instead, then Bazel can ignore files under NVCC_TEMP_DIR during dependency check
+  # http://docs.nvidia.com/cuda/cuda-compiler-driver-nvcc/index.html#options-for-guiding-compiler-driver
+  # Different actions are sharing NVCC_TEMP_DIR, so we cannot remove it if the directory already exists.
+  if os.path.isfile(NVCC_TEMP_DIR):
+    os.remove(NVCC_TEMP_DIR)
+  if not os.path.exists(NVCC_TEMP_DIR):
+    os.makedirs(NVCC_TEMP_DIR)
+  nvccopts += ['--keep', '--keep-dir', NVCC_TEMP_DIR]
+  cmd = [NVCC_PATH] + nvccopts
+  if log:
+    Log(cmd)
+  proc = subprocess.Popen(cmd,
+                          stdout=sys.stdout,
+                          stderr=sys.stderr,
+                          env=os.environ.copy(),
+                          shell=True)
+  proc.wait()
+  return proc.returncode
+
+def main():
+  parser = ArgumentParser()
+  parser.add_argument('-x', nargs=1)
+  parser.add_argument('--cuda_log', action='store_true')
+  args, leftover = parser.parse_known_args(sys.argv[1:])
+
+  if args.x and args.x[0] == 'cuda':
+    if args.cuda_log: Log('-x cuda')
+    leftover = [pipes.quote(s) for s in leftover]
+    if args.cuda_log: Log('using nvcc')
+    return InvokeNvcc(leftover, log=args.cuda_log)
+
+  # Strip our flags before passing through to the CPU compiler for files which
+  # are not -x cuda. We can't just pass 'leftover' because it also strips -x.
+  # We not only want to pass -x to the CPU compiler, but also keep it in its
+  # relative location in the argv list (the compiler is actually sensitive to
+  # this).
+  cpu_compiler_flags = [flag for flag in sys.argv[1:]
+                             if not flag.startswith(('--cuda_log'))
+                             and not flag.startswith(('-nvcc_options'))]
+
+  return subprocess.call([CPU_COMPILER] + cpu_compiler_flags)
+
+if __name__ == '__main__':
+  sys.exit(main())
diff --git a/third_party/gpus/cuda/BUILD.windows.tpl b/third_party/gpus/cuda/BUILD.windows.tpl
new file mode 100644
index 0000000000000000000000000000000000000000..ff6b3cc35144f07c9fba4b42593810ccf50a1b36
--- /dev/null
+++ b/third_party/gpus/cuda/BUILD.windows.tpl
@@ -0,0 +1,163 @@
+licenses(["restricted"])  # MPL2, portions GPL v3, LGPL v3, BSD-like
+
+package(default_visibility = ["//visibility:public"])
+
+config_setting(
+    name = "using_nvcc",
+    values = {
+        "define": "using_cuda_nvcc=true",
+    },
+)
+
+config_setting(
+    name = "using_clang",
+    values = {
+        "define": "using_cuda_clang=true",
+    },
+)
+
+# Equivalent to using_clang && -c opt.
+config_setting(
+    name = "using_clang_opt",
+    values = {
+        "define": "using_cuda_clang=true",
+        "compilation_mode": "opt",
+    },
+)
+
+config_setting(
+    name = "darwin",
+    values = {"cpu": "darwin"},
+    visibility = ["//visibility:public"],
+)
+
+config_setting(
+    name = "freebsd",
+    values = {"cpu": "freebsd"},
+    visibility = ["//visibility:public"],
+)
+
+cc_library(
+    name = "cuda_headers",
+    hdrs = [
+        "cuda/cuda_config.h",
+        %{cuda_headers}
+    ],
+    includes = [
+        ".",
+        "cuda/include",
+        "cuda/include/crt",
+    ],
+    visibility = ["//visibility:public"],
+)
+
+cc_import(
+    name = "cudart_static",
+    # /WHOLEARCHIVE:cudart_static.lib will cause a
+    # "Internal error during CImplib::EmitThunk" error.
+    # Treat this library as interface library to avoid being whole archived when
+    # linking a DLL that depends on this.
+    # TODO(pcloudy): Remove this rule after b/111278841 is resolved.
+    interface_library = "cuda/lib/%{cudart_static_lib}",
+    system_provided = 1,
+    visibility = ["//visibility:public"],
+)
+
+cc_import(
+    name = "cuda_driver",
+    interface_library = "cuda/lib/%{cuda_driver_lib}",
+    system_provided = 1,
+    visibility = ["//visibility:public"],
+)
+
+cc_import(
+    name = "cudart",
+    interface_library = "cuda/lib/%{cudart_lib}",
+    system_provided = 1,
+    visibility = ["//visibility:public"],
+)
+
+cc_import(
+    name = "cublas",
+    interface_library = "cuda/lib/%{cublas_lib}",
+    system_provided = 1,
+    visibility = ["//visibility:public"],
+)
+
+cc_import(
+    name = "cusolver",
+    interface_library = "cuda/lib/%{cusolver_lib}",
+    system_provided = 1,
+    visibility = ["//visibility:public"],
+)
+
+cc_import(
+    name = "cudnn",
+    interface_library = "cuda/lib/%{cudnn_lib}",
+    system_provided = 1,
+    visibility = ["//visibility:public"],
+)
+
+cc_library(
+    name = "cudnn_header",
+    includes = [
+        ".",
+        "cuda/include",
+    ],
+    visibility = ["//visibility:public"],
+)
+
+cc_import(
+    name = "cufft",
+    interface_library = "cuda/lib/%{cufft_lib}",
+    system_provided = 1,
+    visibility = ["//visibility:public"],
+)
+
+cc_import(
+    name = "curand",
+    interface_library = "cuda/lib/%{curand_lib}",
+    system_provided = 1,
+    visibility = ["//visibility:public"],
+)
+
+cc_library(
+    name = "cuda",
+    visibility = ["//visibility:public"],
+    deps = [
+        ":cublas",
+        ":cuda_headers",
+        ":cudart",
+        ":cudnn",
+        ":cufft",
+        ":curand",
+    ],
+)
+
+cc_library(
+    name = "cupti_headers",
+    hdrs = [
+        "cuda/cuda_config.h",
+        ":cuda-extras",
+    ],
+    includes = [
+        ".",
+        "cuda/extras/CUPTI/include/",
+    ],
+    visibility = ["//visibility:public"],
+)
+
+cc_import(
+    name = "cupti_dsos",
+    interface_library = "cuda/lib/%{cupti_lib}",
+    system_provided = 1,
+    visibility = ["//visibility:public"],
+)
+
+cc_library(
+    name = "libdevice_root",
+    data = [":cuda-nvvm"],
+    visibility = ["//visibility:public"],
+)
+
+%{cuda_include_genrules}
diff --git a/third_party/gpus/cuda_configure.bzl b/third_party/gpus/cuda_configure.bzl
index c90c66912d959af109caab51c742d760e0908f30..e848fa175ccb5d39ae9e329837f469b7d5585f05 100644
--- a/third_party/gpus/cuda_configure.bzl
+++ b/third_party/gpus/cuda_configure.bzl
@@ -20,6 +20,7 @@
     `/usr/local/cuda`.
   * `TF_CUDA_COMPUTE_CAPABILITIES`: The CUDA compute capabilities. Default is
     `3.5,5.2`.
+  * `PYTHON_BIN_PATH`: The python binary path
 """
 
 _GCC_HOST_COMPILER_PATH = "GCC_HOST_COMPILER_PATH"
@@ -31,6 +32,7 @@ _CUDNN_INSTALL_PATH = "CUDNN_INSTALL_PATH"
 _TF_CUDA_COMPUTE_CAPABILITIES = "TF_CUDA_COMPUTE_CAPABILITIES"
 _TF_CUDA_CONFIG_REPO = "TF_CUDA_CONFIG_REPO"
 _TF_DOWNLOAD_CLANG = "TF_DOWNLOAD_CLANG"
+_PYTHON_BIN_PATH = "PYTHON_BIN_PATH"
 
 _DEFAULT_CUDA_VERSION = ""
 _DEFAULT_CUDNN_VERSION = ""
@@ -44,12 +46,12 @@ _DEFAULT_CUDA_COMPUTE_CAPABILITIES = ["3.5", "5.2"]
 # will be used. For example, when looking for the cudart libraries, the first
 # attempt will be lib64/cudart inside the CUDA toolkit.
 CUDA_LIB_PATHS = [
-  "lib64/",
-  "lib64/stubs/",
-  "lib/x86_64-linux-gnu/",
-  "lib/x64/",
-  "lib/",
-  "",
+    "lib64/",
+    "lib64/stubs/",
+    "lib/x86_64-linux-gnu/",
+    "lib/x64/",
+    "lib/",
+    "",
 ]
 
 # Lookup paths for cupti.h, relative to the CUDA toolkit directory.
@@ -57,8 +59,8 @@ CUDA_LIB_PATHS = [
 # On most systems, the cupti library is not installed in the same directory as
 # the other CUDA libraries but rather in a special extras/CUPTI directory.
 CUPTI_HEADER_PATHS = [
-  "extras/CUPTI/include/",
-  "include/cuda/CUPTI/",
+    "extras/CUPTI/include/",
+    "include/cuda/CUPTI/",
 ]
 
 # Lookup paths for the cupti library, relative to the
@@ -66,25 +68,25 @@ CUPTI_HEADER_PATHS = [
 # On most systems, the cupti library is not installed in the same directory as
 # the other CUDA libraries but rather in a special extras/CUPTI directory.
 CUPTI_LIB_PATHS = [
-  "extras/CUPTI/lib64/",
-  "lib/x86_64-linux-gnu",
-  "lib64/",
-  "extras/CUPTI/libx64/",
-  "extras/CUPTI/lib/",
-  "lib/",
+    "extras/CUPTI/lib64/",
+    "lib/x86_64-linux-gnu",
+    "lib64/",
+    "extras/CUPTI/libx64/",
+    "extras/CUPTI/lib/",
+    "lib/",
 ]
 
 # Lookup paths for CUDA headers (cuda.h) relative to the CUDA toolkit directory.
 CUDA_INCLUDE_PATHS = [
-  "include/",
-  "include/cuda/"
+    "include/",
+    "include/cuda/",
 ]
 
 # Lookup paths for cudnn.h relative to the CUDNN install directory.
 CUDNN_INCLUDE_PATHS = [
-  "",
-  "include/",
-  "include/cuda/",
+    "",
+    "include/",
+    "include/cuda/",
 ]
 
 # Lookup paths for NVVM libdevice relative to the CUDA directory toolkit.
@@ -92,686 +94,841 @@ CUDNN_INCLUDE_PATHS = [
 # libdevice implements mathematical functions for GPU kernels, and is provided
 # in NVVM bitcode (a subset of LLVM bitcode).
 NVVM_LIBDEVICE_PATHS = [
-  "nvvm/libdevice/",
-  "share/cuda/",
+    "nvvm/libdevice/",
+    "share/cuda/",
+]
+
+# Files used to detect the NVVM libdevice path.
+NVVM_LIBDEVICE_FILES = [
+    # CUDA 9.0 has a single file.
+    "libdevice.10.bc",
+
+    # CUDA 8.0 has separate files for compute versions 2.0, 3.0, 3.5 and 5.0.
+    # Probing for one of them is sufficient.
+    "libdevice.compute_20.10.bc",
 ]
 
 load("//third_party/clang_toolchain:download_clang.bzl", "download_clang")
+load(
+    "@bazel_tools//tools/cpp:lib_cc_configure.bzl",
+    "escape_string",
+    "get_env_var",
+)
+load(
+    "@bazel_tools//tools/cpp:windows_cc_configure.bzl",
+    "find_msvc_tool",
+    "find_vc_path",
+    "setup_vc_env_vars",
+)
+
+def _get_python_bin(repository_ctx):
+    """Gets the python bin path."""
+    python_bin = repository_ctx.os.environ.get(_PYTHON_BIN_PATH)
+    if python_bin != None:
+        return python_bin
+    python_bin_name = "python.exe" if _is_windows(repository_ctx) else "python"
+    python_bin_path = repository_ctx.which(python_bin_name)
+    if python_bin_path != None:
+        return str(python_bin_path)
+    auto_configure_fail("Cannot find python in PATH, please make sure " +
+                        "python is installed and add its directory in PATH, or --define " +
+                        "%s='/something/else'.\nPATH=%s" % (
+                            _PYTHON_BIN_PATH,
+                            repository_ctx.os.environ.get("PATH", ""),
+                        ))
+
+def _get_nvcc_tmp_dir_for_windows(repository_ctx):
+    """Return the tmp directory for nvcc to generate intermediate source files."""
+    escaped_tmp_dir = escape_string(
+        get_env_var(repository_ctx, "TMP", "C:\\Windows\\Temp").replace("\\", "\\\\"),
+    )
+    return escaped_tmp_dir + "\\\\nvcc_inter_files_tmp_dir"
+
+def _get_msvc_compiler(repository_ctx):
+    vc_path = find_vc_path(repository_ctx)
+    return find_msvc_tool(repository_ctx, vc_path, "cl.exe").replace("\\", "/")
+
+def _get_win_cuda_defines(repository_ctx):
+    """Return CROSSTOOL defines for Windows"""
+
+    # If we are not on Windows, return empty vaules for Windows specific fields.
+    # This ensures the CROSSTOOL file parser is happy.
+    if not _is_windows(repository_ctx):
+        return {
+            "%{msvc_env_tmp}": "",
+            "%{msvc_env_path}": "",
+            "%{msvc_env_include}": "",
+            "%{msvc_env_lib}": "",
+            "%{msvc_cl_path}": "",
+            "%{msvc_ml_path}": "",
+            "%{msvc_link_path}": "",
+            "%{msvc_lib_path}": "",
+            "%{cxx_builtin_include_directory}": "",
+        }
+
+    vc_path = find_vc_path(repository_ctx)
+    if not vc_path:
+        auto_configure_fail("Visual C++ build tools not found on your machine." +
+                            "Please check your installation following https://docs.bazel.build/versions/master/windows.html#using")
+        return {}
+
+    env = setup_vc_env_vars(repository_ctx, vc_path)
+    escaped_paths = escape_string(env["PATH"])
+    escaped_include_paths = escape_string(env["INCLUDE"])
+    escaped_lib_paths = escape_string(env["LIB"])
+    escaped_tmp_dir = escape_string(
+        get_env_var(repository_ctx, "TMP", "C:\\Windows\\Temp").replace("\\", "\\\\"),
+    )
+
+    msvc_cl_path = "windows/msvc_wrapper_for_nvcc.bat"
+    msvc_ml_path = find_msvc_tool(repository_ctx, vc_path, "ml64.exe").replace("\\", "/")
+    msvc_link_path = find_msvc_tool(repository_ctx, vc_path, "link.exe").replace("\\", "/")
+    msvc_lib_path = find_msvc_tool(repository_ctx, vc_path, "lib.exe").replace("\\", "/")
+
+    # nvcc will generate some temporary source files under %{nvcc_tmp_dir}
+    # The generated files are guranteed to have unique name, so they can share the same tmp directory
+    escaped_cxx_include_directories = ["cxx_builtin_include_directory: \"%s\"" % _get_nvcc_tmp_dir_for_windows(repository_ctx)]
+    for path in escaped_include_paths.split(";"):
+        if path:
+            escaped_cxx_include_directories.append("cxx_builtin_include_directory: \"%s\"" % path)
+
+    return {
+        "%{msvc_env_tmp}": escaped_tmp_dir,
+        "%{msvc_env_path}": escaped_paths,
+        "%{msvc_env_include}": escaped_include_paths,
+        "%{msvc_env_lib}": escaped_lib_paths,
+        "%{msvc_cl_path}": msvc_cl_path,
+        "%{msvc_ml_path}": msvc_ml_path,
+        "%{msvc_link_path}": msvc_link_path,
+        "%{msvc_lib_path}": msvc_lib_path,
+        "%{cxx_builtin_include_directory}": "\n".join(escaped_cxx_include_directories),
+    }
 
 # TODO(dzc): Once these functions have been factored out of Bazel's
 # cc_configure.bzl, load them from @bazel_tools instead.
 # BEGIN cc_configure common functions.
 def find_cc(repository_ctx):
-  """Find the C++ compiler."""
-  # On Windows, we use Bazel's MSVC CROSSTOOL for GPU build
-  # Return a dummy value for GCC detection here to avoid error
-  if _is_windows(repository_ctx):
-    return "/use/--config=win-cuda --cpu=x64_windows_msvc/instead"
-
-  if _use_cuda_clang(repository_ctx):
-    target_cc_name = "clang"
-    cc_path_envvar = _CLANG_CUDA_COMPILER_PATH
-    if _flag_enabled(repository_ctx, _TF_DOWNLOAD_CLANG):
-      return "extra_tools/bin/clang"
-  else:
-    target_cc_name = "gcc"
-    cc_path_envvar = _GCC_HOST_COMPILER_PATH
-  cc_name = target_cc_name
-
-  if cc_path_envvar in repository_ctx.os.environ:
-    cc_name_from_env = repository_ctx.os.environ[cc_path_envvar].strip()
-    if cc_name_from_env:
-      cc_name = cc_name_from_env
-  if cc_name.startswith("/"):
-    # Absolute path, maybe we should make this supported by our which function.
-    return cc_name
-  cc = repository_ctx.which(cc_name)
-  if cc == None:
-    fail(("Cannot find {}, either correct your path or set the {}" +
-          " environment variable").format(target_cc_name, cc_path_envvar))
-  return cc
-
+    """Find the C++ compiler."""
+    if _is_windows(repository_ctx):
+        return _get_msvc_compiler(repository_ctx)
+
+    if _use_cuda_clang(repository_ctx):
+        target_cc_name = "clang"
+        cc_path_envvar = _CLANG_CUDA_COMPILER_PATH
+        if _flag_enabled(repository_ctx, _TF_DOWNLOAD_CLANG):
+            return "extra_tools/bin/clang"
+    else:
+        target_cc_name = "gcc"
+        cc_path_envvar = _GCC_HOST_COMPILER_PATH
+    cc_name = target_cc_name
+
+    if cc_path_envvar in repository_ctx.os.environ:
+        cc_name_from_env = repository_ctx.os.environ[cc_path_envvar].strip()
+        if cc_name_from_env:
+            cc_name = cc_name_from_env
+    if cc_name.startswith("/"):
+        # Absolute path, maybe we should make this supported by our which function.
+        return cc_name
+    cc = repository_ctx.which(cc_name)
+    if cc == None:
+        fail(("Cannot find {}, either correct your path or set the {}" +
+              " environment variable").format(target_cc_name, cc_path_envvar))
+    return cc
 
 _INC_DIR_MARKER_BEGIN = "#include <...>"
 
-
 # OSX add " (framework directory)" at the end of line, strip it.
 _OSX_FRAMEWORK_SUFFIX = " (framework directory)"
-_OSX_FRAMEWORK_SUFFIX_LEN =  len(_OSX_FRAMEWORK_SUFFIX)
-def _cxx_inc_convert(path):
-  """Convert path returned by cc -E xc++ in a complete path."""
-  path = path.strip()
-  if path.endswith(_OSX_FRAMEWORK_SUFFIX):
-    path = path[:-_OSX_FRAMEWORK_SUFFIX_LEN].strip()
-  return path
+_OSX_FRAMEWORK_SUFFIX_LEN = len(_OSX_FRAMEWORK_SUFFIX)
 
+def _cxx_inc_convert(path):
+    """Convert path returned by cc -E xc++ in a complete path."""
+    path = path.strip()
+    if path.endswith(_OSX_FRAMEWORK_SUFFIX):
+        path = path[:-_OSX_FRAMEWORK_SUFFIX_LEN].strip()
+    return path
 
 def _normalize_include_path(repository_ctx, path):
-  """Normalizes include paths before writing them to the crosstool.
+    """Normalizes include paths before writing them to the crosstool.
 
-  If path points inside the 'crosstool' folder of the repository, a relative
-  path is returned.
-  If path points outside the 'crosstool' folder, an absolute path is returned.
-  """
-  path = str(repository_ctx.path(path))
-  crosstool_folder = str(repository_ctx.path(".").get_child('crosstool'))
-
-  if path.startswith(crosstool_folder):
-    # We drop the path to "$REPO/crosstool" and a trailing path separator.
-    return path[len(crosstool_folder)+1:]
-  return path
+    If path points inside the 'crosstool' folder of the repository, a relative
+    path is returned.
+    If path points outside the 'crosstool' folder, an absolute path is returned.
+    """
+    path = str(repository_ctx.path(path))
+    crosstool_folder = str(repository_ctx.path(".").get_child("crosstool"))
 
+    if path.startswith(crosstool_folder):
+        # We drop the path to "$REPO/crosstool" and a trailing path separator.
+        return path[len(crosstool_folder) + 1:]
+    return path
 
 def _get_cxx_inc_directories_impl(repository_ctx, cc, lang_is_cpp):
-  """Compute the list of default C or C++ include directories."""
-  if lang_is_cpp:
-    lang = "c++"
-  else:
-    lang = "c"
-  result = repository_ctx.execute([cc, "-E", "-x" + lang, "-", "-v"])
-  index1 = result.stderr.find(_INC_DIR_MARKER_BEGIN)
-  if index1 == -1:
-    return []
-  index1 = result.stderr.find("\n", index1)
-  if index1 == -1:
-    return []
-  index2 = result.stderr.rfind("\n ")
-  if index2 == -1 or index2 < index1:
-    return []
-  index2 = result.stderr.find("\n", index2 + 1)
-  if index2 == -1:
-    inc_dirs = result.stderr[index1 + 1:]
-  else:
-    inc_dirs = result.stderr[index1 + 1:index2].strip()
-
-  return [
-      _normalize_include_path(repository_ctx, _cxx_inc_convert(p))
-      for p in inc_dirs.split("\n")
-  ]
+    """Compute the list of default C or C++ include directories."""
+    if lang_is_cpp:
+        lang = "c++"
+    else:
+        lang = "c"
+    result = repository_ctx.execute([cc, "-E", "-x" + lang, "-", "-v"])
+    index1 = result.stderr.find(_INC_DIR_MARKER_BEGIN)
+    if index1 == -1:
+        return []
+    index1 = result.stderr.find("\n", index1)
+    if index1 == -1:
+        return []
+    index2 = result.stderr.rfind("\n ")
+    if index2 == -1 or index2 < index1:
+        return []
+    index2 = result.stderr.find("\n", index2 + 1)
+    if index2 == -1:
+        inc_dirs = result.stderr[index1 + 1:]
+    else:
+        inc_dirs = result.stderr[index1 + 1:index2].strip()
 
+    return [
+        _normalize_include_path(repository_ctx, _cxx_inc_convert(p))
+        for p in inc_dirs.split("\n")
+    ]
 
 def get_cxx_inc_directories(repository_ctx, cc):
-  """Compute the list of default C and C++ include directories."""
-  # For some reason `clang -xc` sometimes returns include paths that are
-  # different from the ones from `clang -xc++`. (Symlink and a dir)
-  # So we run the compiler with both `-xc` and `-xc++` and merge resulting lists
-  includes_cpp = _get_cxx_inc_directories_impl(repository_ctx, cc, True)
-  includes_c = _get_cxx_inc_directories_impl(repository_ctx, cc, False)
+    """Compute the list of default C and C++ include directories."""
 
-  includes_cpp_set = depset(includes_cpp)
-  return includes_cpp + [inc for inc in includes_c
-                         if inc not in includes_cpp_set]
+    # For some reason `clang -xc` sometimes returns include paths that are
+    # different from the ones from `clang -xc++`. (Symlink and a dir)
+    # So we run the compiler with both `-xc` and `-xc++` and merge resulting lists
+    includes_cpp = _get_cxx_inc_directories_impl(repository_ctx, cc, True)
+    includes_c = _get_cxx_inc_directories_impl(repository_ctx, cc, False)
 
+    includes_cpp_set = depset(includes_cpp)
+    return includes_cpp + [
+        inc
+        for inc in includes_c
+        if inc not in includes_cpp_set
+    ]
 
 def auto_configure_fail(msg):
-  """Output failure message when cuda configuration fails."""
-  red = "\033[0;31m"
-  no_color = "\033[0m"
-  fail("\n%sCuda Configuration Error:%s %s\n" % (red, no_color, msg))
-# END cc_configure common functions (see TODO above).
+    """Output failure message when cuda configuration fails."""
+    red = "\033[0;31m"
+    no_color = "\033[0m"
+    fail("\n%sCuda Configuration Error:%s %s\n" % (red, no_color, msg))
 
+# END cc_configure common functions (see TODO above).
 
 def _host_compiler_includes(repository_ctx, cc):
-  """Generates the cxx_builtin_include_directory entries for gcc inc dirs.
-
-  Args:
-    repository_ctx: The repository context.
-    cc: The path to the gcc host compiler.
-
-  Returns:
-    A string containing the cxx_builtin_include_directory for each of the gcc
-    host compiler include directories, which can be added to the CROSSTOOL
-    file.
-  """
-  inc_dirs = get_cxx_inc_directories(repository_ctx, cc)
-  inc_entries = []
-  for inc_dir in inc_dirs:
-    inc_entries.append("  cxx_builtin_include_directory: \"%s\"" % inc_dir)
-  return "\n".join(inc_entries)
+    """Generates the cxx_builtin_include_directory entries for gcc inc dirs.
+
+    Args:
+      repository_ctx: The repository context.
+      cc: The path to the gcc host compiler.
+
+    Returns:
+      A string containing the cxx_builtin_include_directory for each of the gcc
+      host compiler include directories, which can be added to the CROSSTOOL
+      file.
+    """
+    inc_dirs = get_cxx_inc_directories(repository_ctx, cc)
+    inc_entries = []
+    for inc_dir in inc_dirs:
+        inc_entries.append("  cxx_builtin_include_directory: \"%s\"" % inc_dir)
+    return "\n".join(inc_entries)
 
 def _cuda_include_path(repository_ctx, cuda_config):
-  """Generates the cxx_builtin_include_directory entries for cuda inc dirs.
-
-  Args:
-    repository_ctx: The repository context.
-    cc: The path to the gcc host compiler.
-
-  Returns:
-    A string containing the cxx_builtin_include_directory for each of the gcc
-    host compiler include directories, which can be added to the CROSSTOOL
-    file.
-  """
-  nvcc_path = repository_ctx.path("%s/bin/nvcc%s" %
-                                  (cuda_config.cuda_toolkit_path,
-                                   ".exe" if cuda_config.cpu_value == "Windows" else ""))
-  result = repository_ctx.execute([nvcc_path, '-v',
-                                  '/dev/null', '-o', '/dev/null'])
-  target_dir = ""
-  for one_line in result.stderr.splitlines():
-    if one_line.startswith('#$ _TARGET_DIR_='):
-      target_dir = (cuda_config.cuda_toolkit_path + '/' +
-                    one_line.replace('#$ _TARGET_DIR_=', '') + "/include")
-  inc_entries = []
-  if target_dir != "":
-    inc_entries.append("  cxx_builtin_include_directory: \"%s\"" % target_dir)
-  default_include = cuda_config.cuda_toolkit_path + '/include'
-  inc_entries.append("  cxx_builtin_include_directory: \"%s\"" %
-                     default_include)
-  return "\n".join(inc_entries)
+    """Generates the cxx_builtin_include_directory entries for cuda inc dirs.
 
+    Args:
+      repository_ctx: The repository context.
+      cc: The path to the gcc host compiler.
 
-def _enable_cuda(repository_ctx):
-  if "TF_NEED_CUDA" in repository_ctx.os.environ:
-    enable_cuda = repository_ctx.os.environ["TF_NEED_CUDA"].strip()
-    return enable_cuda == "1"
-  return False
+    Returns:
+      A string containing the cxx_builtin_include_directory for each of the gcc
+      host compiler include directories, which can be added to the CROSSTOOL
+      file.
+    """
+    nvcc_path = repository_ctx.path("%s/bin/nvcc%s" %
+                                    (
+                                        cuda_config.cuda_toolkit_path,
+                                        ".exe" if cuda_config.cpu_value == "Windows" else "",
+                                    ))
+    result = repository_ctx.execute([
+        nvcc_path,
+        "-v",
+        "/dev/null",
+        "-o",
+        "/dev/null",
+    ])
+    target_dir = ""
+    for one_line in result.stderr.splitlines():
+        if one_line.startswith("#$ _TARGET_DIR_="):
+            target_dir = (cuda_config.cuda_toolkit_path + "/" +
+                          one_line.replace("#$ _TARGET_DIR_=", "") + "/include")
+    inc_entries = []
+    if target_dir != "":
+        inc_entries.append("  cxx_builtin_include_directory: \"%s\"" % target_dir)
+    default_include = cuda_config.cuda_toolkit_path + "/include"
+    inc_entries.append("  cxx_builtin_include_directory: \"%s\"" %
+                       default_include)
+    return "\n".join(inc_entries)
 
+def _enable_cuda(repository_ctx):
+    if "TF_NEED_CUDA" in repository_ctx.os.environ:
+        enable_cuda = repository_ctx.os.environ["TF_NEED_CUDA"].strip()
+        return enable_cuda == "1"
+    return False
 
 def _cuda_toolkit_path(repository_ctx):
-  """Finds the cuda toolkit directory.
-
-  Args:
-    repository_ctx: The repository context.
+    """Finds the cuda toolkit directory.
 
-  Returns:
-    A speculative real path of the cuda toolkit install directory.
-  """
-  cuda_toolkit_path = _DEFAULT_CUDA_TOOLKIT_PATH
-  if _CUDA_TOOLKIT_PATH in repository_ctx.os.environ:
-    cuda_toolkit_path = repository_ctx.os.environ[_CUDA_TOOLKIT_PATH].strip()
-  if not repository_ctx.path(cuda_toolkit_path).exists:
-    auto_configure_fail("Cannot find cuda toolkit path.")
-  return str(repository_ctx.path(cuda_toolkit_path).realpath)
+    Args:
+      repository_ctx: The repository context.
 
+    Returns:
+      A speculative real path of the cuda toolkit install directory.
+    """
+    cuda_toolkit_path = _DEFAULT_CUDA_TOOLKIT_PATH
+    if _CUDA_TOOLKIT_PATH in repository_ctx.os.environ:
+        cuda_toolkit_path = repository_ctx.os.environ[_CUDA_TOOLKIT_PATH].strip()
+    if not repository_ctx.path(cuda_toolkit_path).exists:
+        auto_configure_fail("Cannot find cuda toolkit path.")
+    return str(repository_ctx.path(cuda_toolkit_path).realpath)
 
 def _cudnn_install_basedir(repository_ctx):
-  """Finds the cudnn install directory."""
-  cudnn_install_path = _DEFAULT_CUDNN_INSTALL_PATH
-  if _CUDNN_INSTALL_PATH in repository_ctx.os.environ:
-    cudnn_install_path = repository_ctx.os.environ[_CUDNN_INSTALL_PATH].strip()
-  if not repository_ctx.path(cudnn_install_path).exists:
-    auto_configure_fail("Cannot find cudnn install path.")
-  return cudnn_install_path
-
+    """Finds the cudnn install directory."""
+    cudnn_install_path = _DEFAULT_CUDNN_INSTALL_PATH
+    if _CUDNN_INSTALL_PATH in repository_ctx.os.environ:
+        cudnn_install_path = repository_ctx.os.environ[_CUDNN_INSTALL_PATH].strip()
+    if not repository_ctx.path(cudnn_install_path).exists:
+        auto_configure_fail("Cannot find cudnn install path.")
+    return cudnn_install_path
 
 def matches_version(environ_version, detected_version):
-  """Checks whether the user-specified version matches the detected version.
-
-  This function performs a weak matching so that if the user specifies only the
-  major or major and minor versions, the versions are still considered matching
-  if the version parts match. To illustrate:
-
-      environ_version  detected_version  result
-      -----------------------------------------
-      5.1.3            5.1.3             True
-      5.1              5.1.3             True
-      5                5.1               True
-      5.1.3            5.1               False
-      5.2.3            5.1.3             False
-
-  Args:
-    environ_version: The version specified by the user via environment
-      variables.
-    detected_version: The version autodetected from the CUDA installation on
-      the system.
-
-  Returns: True if user-specified version matches detected version and False
-    otherwise.
-  """
-  environ_version_parts = environ_version.split(".")
-  detected_version_parts = detected_version.split(".")
-  if len(detected_version_parts) < len(environ_version_parts):
-    return False
-  for i, part in enumerate(detected_version_parts):
-    if i >= len(environ_version_parts):
-      break
-    if part != environ_version_parts[i]:
-      return False
-  return True
-
+    """Checks whether the user-specified version matches the detected version.
+
+    This function performs a weak matching so that if the user specifies only the
+    major or major and minor versions, the versions are still considered matching
+    if the version parts match. To illustrate:
+
+        environ_version  detected_version  result
+        -----------------------------------------
+        5.1.3            5.1.3             True
+        5.1              5.1.3             True
+        5                5.1               True
+        5.1.3            5.1               False
+        5.2.3            5.1.3             False
+
+    Args:
+      environ_version: The version specified by the user via environment
+        variables.
+      detected_version: The version autodetected from the CUDA installation on
+        the system.
+
+    Returns: True if user-specified version matches detected version and False
+      otherwise.
+    """
+    environ_version_parts = environ_version.split(".")
+    detected_version_parts = detected_version.split(".")
+    if len(detected_version_parts) < len(environ_version_parts):
+        return False
+    for i, part in enumerate(detected_version_parts):
+        if i >= len(environ_version_parts):
+            break
+        if part != environ_version_parts[i]:
+            return False
+    return True
 
 _NVCC_VERSION_PREFIX = "Cuda compilation tools, release "
 
-
 def _cuda_version(repository_ctx, cuda_toolkit_path, cpu_value):
-  """Detects the version of CUDA installed on the system.
-
-  Args:
-    repository_ctx: The repository context.
-    cuda_toolkit_path: The CUDA install directory.
-
-  Returns:
-    String containing the version of CUDA.
-  """
-  # Run nvcc --version and find the line containing the CUDA version.
-  nvcc_path = repository_ctx.path("%s/bin/nvcc%s" %
-                                  (cuda_toolkit_path,
-                                   ".exe" if cpu_value == "Windows" else ""))
-  if not nvcc_path.exists:
-    auto_configure_fail("Cannot find nvcc at %s" % str(nvcc_path))
-  result = repository_ctx.execute([str(nvcc_path), '--version'])
-  if result.stderr:
-    auto_configure_fail("Error running nvcc --version: %s" % result.stderr)
-  lines = result.stdout.splitlines()
-  version_line = lines[len(lines) - 1]
-  if version_line.find(_NVCC_VERSION_PREFIX) == -1:
-    auto_configure_fail(
-        "Could not parse CUDA version from nvcc --version. Got: %s" %
-        result.stdout)
-
-  # Parse the CUDA version from the line containing the CUDA version.
-  prefix_removed = version_line.replace(_NVCC_VERSION_PREFIX, '')
-  parts = prefix_removed.split(",")
-  if len(parts) != 2 or len(parts[0]) < 2:
-    auto_configure_fail(
-        "Could not parse CUDA version from nvcc --version. Got: %s" %
-        result.stdout)
-  full_version = parts[1].strip()
-  if full_version.startswith('V'):
-    full_version = full_version[1:]
-
-  # Check whether TF_CUDA_VERSION was set by the user and fail if it does not
-  # match the detected version.
-  environ_version = ""
-  if _TF_CUDA_VERSION in repository_ctx.os.environ:
-    environ_version = repository_ctx.os.environ[_TF_CUDA_VERSION].strip()
-  if environ_version and not matches_version(environ_version, full_version):
-    auto_configure_fail(
-        ("CUDA version detected from nvcc (%s) does not match " +
-         "TF_CUDA_VERSION (%s)") % (full_version, environ_version))
-
-  # We only use the version consisting of the major and minor version numbers.
-  version_parts = full_version.split('.')
-  if len(version_parts) < 2:
-    auto_configure_fail("CUDA version detected from nvcc (%s) is incomplete.")
-  if cpu_value == "Windows":
-    version = "64_%s%s" % (version_parts[0], version_parts[1])
-  else:
-    version = "%s.%s" % (version_parts[0], version_parts[1])
-  return version
+    """Detects the version of CUDA installed on the system.
+
+    Args:
+      repository_ctx: The repository context.
+      cuda_toolkit_path: The CUDA install directory.
+
+    Returns:
+      String containing the version of CUDA.
+    """
+
+    # Run nvcc --version and find the line containing the CUDA version.
+    nvcc_path = repository_ctx.path("%s/bin/nvcc%s" %
+                                    (
+                                        cuda_toolkit_path,
+                                        ".exe" if cpu_value == "Windows" else "",
+                                    ))
+    if not nvcc_path.exists:
+        auto_configure_fail("Cannot find nvcc at %s" % str(nvcc_path))
+    result = repository_ctx.execute([str(nvcc_path), "--version"])
+    if result.stderr:
+        auto_configure_fail("Error running nvcc --version: %s" % result.stderr)
+    lines = result.stdout.splitlines()
+    version_line = lines[len(lines) - 1]
+    if version_line.find(_NVCC_VERSION_PREFIX) == -1:
+        auto_configure_fail(
+            "Could not parse CUDA version from nvcc --version. Got: %s" %
+            result.stdout,
+        )
 
+    # Parse the CUDA version from the line containing the CUDA version.
+    prefix_removed = version_line.replace(_NVCC_VERSION_PREFIX, "")
+    parts = prefix_removed.split(",")
+    if len(parts) != 2 or len(parts[0]) < 2:
+        auto_configure_fail(
+            "Could not parse CUDA version from nvcc --version. Got: %s" %
+            result.stdout,
+        )
+    full_version = parts[1].strip()
+    if full_version.startswith("V"):
+        full_version = full_version[1:]
+
+    # Check whether TF_CUDA_VERSION was set by the user and fail if it does not
+    # match the detected version.
+    environ_version = ""
+    if _TF_CUDA_VERSION in repository_ctx.os.environ:
+        environ_version = repository_ctx.os.environ[_TF_CUDA_VERSION].strip()
+    if environ_version and not matches_version(environ_version, full_version):
+        auto_configure_fail(
+            ("CUDA version detected from nvcc (%s) does not match " +
+             "TF_CUDA_VERSION (%s)") % (full_version, environ_version),
+        )
+
+    # We only use the version consisting of the major and minor version numbers.
+    version_parts = full_version.split(".")
+    if len(version_parts) < 2:
+        auto_configure_fail("CUDA version detected from nvcc (%s) is incomplete.")
+    if cpu_value == "Windows":
+        version = "64_%s%s" % (version_parts[0], version_parts[1])
+    else:
+        version = "%s.%s" % (version_parts[0], version_parts[1])
+    return version
 
 _DEFINE_CUDNN_MAJOR = "#define CUDNN_MAJOR"
 _DEFINE_CUDNN_MINOR = "#define CUDNN_MINOR"
 _DEFINE_CUDNN_PATCHLEVEL = "#define CUDNN_PATCHLEVEL"
 
-
 def find_cuda_define(repository_ctx, header_dir, header_file, define):
-  """Returns the value of a #define in a header file.
-
-  Greps through a header file and returns the value of the specified #define.
-  If the #define is not found, then raise an error.
-
-  Args:
-    repository_ctx: The repository context.
-    header_dir: The directory containing the header file.
-    header_file: The header file name.
-    define: The #define to search for.
-
-  Returns:
-    The value of the #define found in the header.
-  """
-  # Confirm location of the header and grep for the line defining the macro.
-  h_path = repository_ctx.path("%s/%s" % (header_dir, header_file))
-  if not h_path.exists:
-    auto_configure_fail("Cannot find %s at %s" % (header_file, str(h_path)))
-  result = repository_ctx.execute(
-      # Grep one more lines as some #defines are splitted into two lines.
-      ["grep", "--color=never", "-A1", "-E", define, str(h_path)])
-  if result.stderr:
-    auto_configure_fail("Error reading %s: %s" % (str(h_path), result.stderr))
-
-  # Parse the version from the line defining the macro.
-  if result.stdout.find(define) == -1:
-    auto_configure_fail("Cannot find line containing '%s' in %s" %
-                        (define, h_path))
-  # Split results to lines
-  lines = result.stdout.split('\n')
-  num_lines = len(lines)
-  for l in range(num_lines):
-    line = lines[l]
-    if define in line:  # Find the line with define
-      version = line
-      if l != num_lines-1 and line[-1] == '\\':  # Add next line, if multiline
-        version = version[:-1] + lines[l+1]
-      break
-  # Remove any comments
-  version = version.split("//")[0]
-  # Remove define name
-  version = version.replace(define, "").strip()
-  # Remove the code after the version number.
-  version_end = version.find(" ")
-  if version_end != -1:
-    if version_end == 0:
-      auto_configure_fail(
-          "Cannot extract the version from line containing '%s' in %s" %
-          (define, str(h_path)))
-    version = version[:version_end].strip()
-  return version
+    """Returns the value of a #define in a header file.
+
+    Greps through a header file and returns the value of the specified #define.
+    If the #define is not found, then raise an error.
 
+    Args:
+      repository_ctx: The repository context.
+      header_dir: The directory containing the header file.
+      header_file: The header file name.
+      define: The #define to search for.
+
+    Returns:
+      The value of the #define found in the header.
+    """
+
+    # Confirm location of the header and grep for the line defining the macro.
+    h_path = repository_ctx.path("%s/%s" % (header_dir, header_file))
+    if not h_path.exists:
+        auto_configure_fail("Cannot find %s at %s" % (header_file, str(h_path)))
+    result = repository_ctx.execute(
+        # Grep one more lines as some #defines are splitted into two lines.
+        ["grep", "--color=never", "-A1", "-E", define, str(h_path)],
+    )
+    if result.stderr:
+        auto_configure_fail("Error reading %s: %s" % (str(h_path), result.stderr))
+
+    # Parse the version from the line defining the macro.
+    if result.stdout.find(define) == -1:
+        auto_configure_fail("Cannot find line containing '%s' in %s" %
+                            (define, h_path))
+
+    # Split results to lines
+    lines = result.stdout.split("\n")
+    num_lines = len(lines)
+    for l in range(num_lines):
+        line = lines[l]
+        if define in line:  # Find the line with define
+            version = line
+            if l != num_lines - 1 and line[-1] == "\\":  # Add next line, if multiline
+                version = version[:-1] + lines[l + 1]
+            break
+
+    # Remove any comments
+    version = version.split("//")[0]
+
+    # Remove define name
+    version = version.replace(define, "").strip()
+
+    # Remove the code after the version number.
+    version_end = version.find(" ")
+    if version_end != -1:
+        if version_end == 0:
+            auto_configure_fail(
+                "Cannot extract the version from line containing '%s' in %s" %
+                (define, str(h_path)),
+            )
+        version = version[:version_end].strip()
+    return version
 
 def _cudnn_version(repository_ctx, cudnn_install_basedir, cpu_value):
-  """Detects the version of cuDNN installed on the system.
-
-  Args:
-    repository_ctx: The repository context.
-    cpu_value: The name of the host operating system.
-    cudnn_install_basedir: The cuDNN install directory.
-
-  Returns:
-    A string containing the version of cuDNN.
-  """
-  cudnn_header_dir = _find_cudnn_header_dir(repository_ctx,
-                                            cudnn_install_basedir)
-  major_version = find_cuda_define(
-      repository_ctx, cudnn_header_dir, "cudnn.h", _DEFINE_CUDNN_MAJOR)
-  minor_version = find_cuda_define(
-      repository_ctx, cudnn_header_dir, "cudnn.h", _DEFINE_CUDNN_MINOR)
-  patch_version = find_cuda_define(
-      repository_ctx, cudnn_header_dir, "cudnn.h", _DEFINE_CUDNN_PATCHLEVEL)
-  full_version = "%s.%s.%s" % (major_version, minor_version, patch_version)
-
-  # Check whether TF_CUDNN_VERSION was set by the user and fail if it does not
-  # match the detected version.
-  environ_version = ""
-  if _TF_CUDNN_VERSION in repository_ctx.os.environ:
-    environ_version = repository_ctx.os.environ[_TF_CUDNN_VERSION].strip()
-  if environ_version and not matches_version(environ_version, full_version):
-    cudnn_h_path = repository_ctx.path("%s/include/cudnn.h" %
-                                       cudnn_install_basedir)
-    auto_configure_fail(
-        ("cuDNN version detected from %s (%s) does not match " +
-        "TF_CUDNN_VERSION (%s)") %
-        (str(cudnn_h_path), full_version, environ_version))
-
-  # We only use the major version since we use the libcudnn libraries that are
-  # only versioned with the major version (e.g. libcudnn.so.5).
-  version = major_version
-  if cpu_value == "Windows":
-    version = "64_" + version
-  return version
+    """Detects the version of cuDNN installed on the system.
 
+    Args:
+      repository_ctx: The repository context.
+      cpu_value: The name of the host operating system.
+      cudnn_install_basedir: The cuDNN install directory.
 
-def _compute_capabilities(repository_ctx):
-  """Returns a list of strings representing cuda compute capabilities."""
-  if _TF_CUDA_COMPUTE_CAPABILITIES not in repository_ctx.os.environ:
-    return _DEFAULT_CUDA_COMPUTE_CAPABILITIES
-  capabilities_str = repository_ctx.os.environ[_TF_CUDA_COMPUTE_CAPABILITIES]
-  capabilities = capabilities_str.split(",")
-  for capability in capabilities:
-    # Workaround for Skylark's lack of support for regex. This check should
-    # be equivalent to checking:
-    #     if re.match("[0-9]+.[0-9]+", capability) == None:
-    parts = capability.split(".")
-    if len(parts) != 2 or not parts[0].isdigit() or not parts[1].isdigit():
-      auto_configure_fail("Invalid compute capability: %s" % capability)
-  return capabilities
+    Returns:
+      A string containing the version of cuDNN.
+    """
+    cudnn_header_dir = _find_cudnn_header_dir(
+        repository_ctx,
+        cudnn_install_basedir,
+    )
+    major_version = find_cuda_define(
+        repository_ctx,
+        cudnn_header_dir,
+        "cudnn.h",
+        _DEFINE_CUDNN_MAJOR,
+    )
+    minor_version = find_cuda_define(
+        repository_ctx,
+        cudnn_header_dir,
+        "cudnn.h",
+        _DEFINE_CUDNN_MINOR,
+    )
+    patch_version = find_cuda_define(
+        repository_ctx,
+        cudnn_header_dir,
+        "cudnn.h",
+        _DEFINE_CUDNN_PATCHLEVEL,
+    )
+    full_version = "%s.%s.%s" % (major_version, minor_version, patch_version)
+
+    # Check whether TF_CUDNN_VERSION was set by the user and fail if it does not
+    # match the detected version.
+    environ_version = ""
+    if _TF_CUDNN_VERSION in repository_ctx.os.environ:
+        environ_version = repository_ctx.os.environ[_TF_CUDNN_VERSION].strip()
+    if environ_version and not matches_version(environ_version, full_version):
+        cudnn_h_path = repository_ctx.path("%s/include/cudnn.h" %
+                                           cudnn_install_basedir)
+        auto_configure_fail(
+            ("cuDNN version detected from %s (%s) does not match " +
+             "TF_CUDNN_VERSION (%s)") %
+            (str(cudnn_h_path), full_version, environ_version),
+        )
 
+    # We only use the major version since we use the libcudnn libraries that are
+    # only versioned with the major version (e.g. libcudnn.so.5).
+    version = major_version
+    if cpu_value == "Windows":
+        version = "64_" + version
+    return version
 
-def get_cpu_value(repository_ctx):
-  """Returns the name of the host operating system.
+def _compute_capabilities(repository_ctx):
+    """Returns a list of strings representing cuda compute capabilities."""
+    if _TF_CUDA_COMPUTE_CAPABILITIES not in repository_ctx.os.environ:
+        return _DEFAULT_CUDA_COMPUTE_CAPABILITIES
+    capabilities_str = repository_ctx.os.environ[_TF_CUDA_COMPUTE_CAPABILITIES]
+    capabilities = capabilities_str.split(",")
+    for capability in capabilities:
+        # Workaround for Skylark's lack of support for regex. This check should
+        # be equivalent to checking:
+        #     if re.match("[0-9]+.[0-9]+", capability) == None:
+        parts = capability.split(".")
+        if len(parts) != 2 or not parts[0].isdigit() or not parts[1].isdigit():
+            auto_configure_fail("Invalid compute capability: %s" % capability)
+    return capabilities
 
-  Args:
-    repository_ctx: The repository context.
+def get_cpu_value(repository_ctx):
+    """Returns the name of the host operating system.
 
-  Returns:
-    A string containing the name of the host operating system.
-  """
-  os_name = repository_ctx.os.name.lower()
-  if os_name.startswith("mac os"):
-    return "Darwin"
-  if os_name.find("windows") != -1:
-    return "Windows"
-  result = repository_ctx.execute(["uname", "-s"])
-  return result.stdout.strip()
+    Args:
+      repository_ctx: The repository context.
 
+    Returns:
+      A string containing the name of the host operating system.
+    """
+    os_name = repository_ctx.os.name.lower()
+    if os_name.startswith("mac os"):
+        return "Darwin"
+    if os_name.find("windows") != -1:
+        return "Windows"
+    result = repository_ctx.execute(["uname", "-s"])
+    return result.stdout.strip()
 
 def _is_windows(repository_ctx):
-  """Returns true if the host operating system is windows."""
-  return get_cpu_value(repository_ctx) == "Windows"
-
-def _lib_name(lib, cpu_value, version="", static=False):
-  """Constructs the platform-specific name of a library.
-
-  Args:
-    lib: The name of the library, such as "cudart"
-    cpu_value: The name of the host operating system.
-    version: The version of the library.
-    static: True the library is static or False if it is a shared object.
-
-  Returns:
-    The platform-specific name of the library.
-  """
-  if cpu_value in ("Linux", "FreeBSD"):
-    if static:
-      return "lib%s.a" % lib
-    else:
-      if version:
-        version = ".%s" % version
-      return "lib%s.so%s" % (lib, version)
-  elif cpu_value == "Windows":
-    return "%s.lib" % lib
-  elif cpu_value == "Darwin":
-    if static:
-      return "lib%s.a" % lib
-    else:
-      if version:
-        version = ".%s" % version
-    return "lib%s%s.dylib" % (lib, version)
-  else:
-    auto_configure_fail("Invalid cpu_value: %s" % cpu_value)
-
-
-def _find_cuda_lib(lib, repository_ctx, cpu_value, basedir, version="",
-                   static=False):
-  """Finds the given CUDA or cuDNN library on the system.
-
-  Args:
-    lib: The name of the library, such as "cudart"
-    repository_ctx: The repository context.
-    cpu_value: The name of the host operating system.
-    basedir: The install directory of CUDA or cuDNN.
-    version: The version of the library.
-    static: True if static library, False if shared object.
-
-  Returns:
-    Returns a struct with the following fields:
-      file_name: The basename of the library found on the system.
-      path: The full path to the library.
-  """
-  file_name = _lib_name(lib, cpu_value, version, static)
-  for relative_path in CUDA_LIB_PATHS:
-    path = repository_ctx.path("%s/%s%s" % (basedir, relative_path, file_name))
-    if path.exists:
-      return struct(file_name=file_name, path=str(path.realpath))
-  auto_configure_fail("Cannot find cuda library %s" % file_name)
+    """Returns true if the host operating system is windows."""
+    return get_cpu_value(repository_ctx) == "Windows"
 
+def _lib_name(lib, cpu_value, version = "", static = False):
+    """Constructs the platform-specific name of a library.
 
-def _find_cupti_header_dir(repository_ctx, cuda_config):
-  """Returns the path to the directory containing cupti.h
+    Args:
+      lib: The name of the library, such as "cudart"
+      cpu_value: The name of the host operating system.
+      version: The version of the library.
+      static: True the library is static or False if it is a shared object.
+
+    Returns:
+      The platform-specific name of the library.
+    """
+    if cpu_value in ("Linux", "FreeBSD"):
+        if static:
+            return "lib%s.a" % lib
+        else:
+            if version:
+                version = ".%s" % version
+            return "lib%s.so%s" % (lib, version)
+    elif cpu_value == "Windows":
+        return "%s.lib" % lib
+    elif cpu_value == "Darwin":
+        if static:
+            return "lib%s.a" % lib
+        elif version:
+            version = ".%s" % version
+        return "lib%s%s.dylib" % (lib, version)
+    else:
+        auto_configure_fail("Invalid cpu_value: %s" % cpu_value)
+
+def _find_cuda_lib(
+        lib,
+        repository_ctx,
+        cpu_value,
+        basedir,
+        version = "",
+        static = False):
+    """Finds the given CUDA or cuDNN library on the system.
+
+    Args:
+      lib: The name of the library, such as "cudart"
+      repository_ctx: The repository context.
+      cpu_value: The name of the host operating system.
+      basedir: The install directory of CUDA or cuDNN.
+      version: The version of the library.
+      static: True if static library, False if shared object.
+
+    Returns:
+      Returns a struct with the following fields:
+        file_name: The basename of the library found on the system.
+        path: The full path to the library.
+    """
+    file_name = _lib_name(lib, cpu_value, version, static)
+    for relative_path in CUDA_LIB_PATHS:
+        path = repository_ctx.path("%s/%s%s" % (basedir, relative_path, file_name))
+        if path.exists:
+            return struct(file_name = file_name, path = str(path.realpath))
+    auto_configure_fail("Cannot find cuda library %s" % file_name)
 
-  On most systems, the cupti library is not installed in the same directory as
-  the other CUDA libraries but rather in a special extras/CUPTI directory.
+def _find_cupti_header_dir(repository_ctx, cuda_config):
+    """Returns the path to the directory containing cupti.h
 
-  Args:
-    repository_ctx: The repository context.
-    cuda_config: The CUDA config as returned by _get_cuda_config
+    On most systems, the cupti library is not installed in the same directory as
+    the other CUDA libraries but rather in a special extras/CUPTI directory.
 
-  Returns:
-    The path of the directory containing the cupti header.
-  """
-  cuda_toolkit_path = cuda_config.cuda_toolkit_path
-  for relative_path in CUPTI_HEADER_PATHS:
-    if repository_ctx.path("%s/%scupti.h" % (cuda_toolkit_path, relative_path)).exists:
-        return ("%s/%s" % (cuda_toolkit_path, relative_path))[:-1]
-  auto_configure_fail("Cannot find cupti.h under %s" % ", ".join([cuda_toolkit_path + "/" + s for s in CUPTI_HEADER_PATHS]))
+    Args:
+      repository_ctx: The repository context.
+      cuda_config: The CUDA config as returned by _get_cuda_config
 
+    Returns:
+      The path of the directory containing the cupti header.
+    """
+    cuda_toolkit_path = cuda_config.cuda_toolkit_path
+    for relative_path in CUPTI_HEADER_PATHS:
+        if repository_ctx.path("%s/%scupti.h" % (cuda_toolkit_path, relative_path)).exists:
+            return ("%s/%s" % (cuda_toolkit_path, relative_path))[:-1]
+    auto_configure_fail("Cannot find cupti.h under %s" % ", ".join([cuda_toolkit_path + "/" + s for s in CUPTI_HEADER_PATHS]))
 
 def _find_cupti_lib(repository_ctx, cuda_config):
-  """Finds the cupti library on the system.
-
-  On most systems, the cupti library is not installed in the same directory as
-  the other CUDA libraries but rather in a special extras/CUPTI directory.
-
-  Args:
-    repository_ctx: The repository context.
-    cuda_config: The cuda configuration as returned by _get_cuda_config.
-
-  Returns:
-    Returns a struct with the following fields:
-      file_name: The basename of the library found on the system.
-      path: The full path to the library.
-  """
-  file_name = _lib_name("cupti", cuda_config.cpu_value,
-                        cuda_config.cuda_version)
-  cuda_toolkit_path = cuda_config.cuda_toolkit_path
-  for relative_path in CUPTI_LIB_PATHS:
-    path = repository_ctx.path(
-        "%s/%s%s" % (cuda_toolkit_path, relative_path, file_name))
-    if path.exists:
-      return struct(file_name=file_name, path=str(path.realpath))
-
-  auto_configure_fail("Cannot find cupti library %s" % file_name)
+    """Finds the cupti library on the system.
+
+    On most systems, the cupti library is not installed in the same directory as
+    the other CUDA libraries but rather in a special extras/CUPTI directory.
+
+    Args:
+      repository_ctx: The repository context.
+      cuda_config: The cuda configuration as returned by _get_cuda_config.
+
+    Returns:
+      Returns a struct with the following fields:
+        file_name: The basename of the library found on the system.
+        path: The full path to the library.
+    """
+    file_name = _lib_name(
+        "cupti",
+        cuda_config.cpu_value,
+        cuda_config.cuda_version,
+    )
+    cuda_toolkit_path = cuda_config.cuda_toolkit_path
+    for relative_path in CUPTI_LIB_PATHS:
+        path = repository_ctx.path(
+            "%s/%s%s" % (cuda_toolkit_path, relative_path, file_name),
+        )
+        if path.exists:
+            return struct(file_name = file_name, path = str(path.realpath))
+
+    auto_configure_fail("Cannot find cupti library %s" % file_name)
 
 def _find_libs(repository_ctx, cuda_config):
-  """Returns the CUDA and cuDNN libraries on the system.
-
-  Args:
-    repository_ctx: The repository context.
-    cuda_config: The CUDA config as returned by _get_cuda_config
-
-  Returns:
-    Map of library names to structs of filename and path.
-  """
-  cpu_value = cuda_config.cpu_value
-  return {
-      "cuda": _find_cuda_lib("cuda", repository_ctx, cpu_value, cuda_config.cuda_toolkit_path),
-      "cudart": _find_cuda_lib(
-          "cudart", repository_ctx, cpu_value, cuda_config.cuda_toolkit_path,
-          cuda_config.cuda_version),
-      "cudart_static": _find_cuda_lib(
-          "cudart_static", repository_ctx, cpu_value,
-          cuda_config.cuda_toolkit_path, cuda_config.cuda_version, static=True),
-      "cublas": _find_cuda_lib(
-          "cublas", repository_ctx, cpu_value, cuda_config.cuda_toolkit_path,
-          cuda_config.cuda_version),
-      "cusolver": _find_cuda_lib(
-          "cusolver", repository_ctx, cpu_value, cuda_config.cuda_toolkit_path,
-          cuda_config.cuda_version),
-      "curand": _find_cuda_lib(
-          "curand", repository_ctx, cpu_value, cuda_config.cuda_toolkit_path,
-          cuda_config.cuda_version),
-      "cufft": _find_cuda_lib(
-          "cufft", repository_ctx, cpu_value, cuda_config.cuda_toolkit_path,
-          cuda_config.cuda_version),
-      "cudnn": _find_cuda_lib(
-          "cudnn", repository_ctx, cpu_value, cuda_config.cudnn_install_basedir,
-          cuda_config.cudnn_version),
-      "cupti": _find_cupti_lib(repository_ctx, cuda_config)
-  }
+    """Returns the CUDA and cuDNN libraries on the system.
 
+    Args:
+      repository_ctx: The repository context.
+      cuda_config: The CUDA config as returned by _get_cuda_config
 
-def _find_cuda_include_path(repository_ctx, cuda_config):
-  """Returns the path to the directory containing cuda.h
+    Returns:
+      Map of library names to structs of filename and path.
+    """
+    cpu_value = cuda_config.cpu_value
+    return {
+        "cuda": _find_cuda_lib("cuda", repository_ctx, cpu_value, cuda_config.cuda_toolkit_path),
+        "cudart": _find_cuda_lib(
+            "cudart",
+            repository_ctx,
+            cpu_value,
+            cuda_config.cuda_toolkit_path,
+            cuda_config.cuda_version,
+        ),
+        "cudart_static": _find_cuda_lib(
+            "cudart_static",
+            repository_ctx,
+            cpu_value,
+            cuda_config.cuda_toolkit_path,
+            cuda_config.cuda_version,
+            static = True,
+        ),
+        "cublas": _find_cuda_lib(
+            "cublas",
+            repository_ctx,
+            cpu_value,
+            cuda_config.cuda_toolkit_path,
+            cuda_config.cuda_version,
+        ),
+        "cusolver": _find_cuda_lib(
+            "cusolver",
+            repository_ctx,
+            cpu_value,
+            cuda_config.cuda_toolkit_path,
+            cuda_config.cuda_version,
+        ),
+        "curand": _find_cuda_lib(
+            "curand",
+            repository_ctx,
+            cpu_value,
+            cuda_config.cuda_toolkit_path,
+            cuda_config.cuda_version,
+        ),
+        "cufft": _find_cuda_lib(
+            "cufft",
+            repository_ctx,
+            cpu_value,
+            cuda_config.cuda_toolkit_path,
+            cuda_config.cuda_version,
+        ),
+        "cudnn": _find_cuda_lib(
+            "cudnn",
+            repository_ctx,
+            cpu_value,
+            cuda_config.cudnn_install_basedir,
+            cuda_config.cudnn_version,
+        ),
+        "cupti": _find_cupti_lib(repository_ctx, cuda_config),
+    }
 
-  Args:
-    repository_ctx: The repository context.
-    cuda_config: The CUDA config as returned by _get_cuda_config
+def _find_cuda_include_path(repository_ctx, cuda_config):
+    """Returns the path to the directory containing cuda.h
 
-  Returns:
-    The path of the directory containing the CUDA headers.
-  """
-  cuda_toolkit_path = cuda_config.cuda_toolkit_path
-  for relative_path in CUDA_INCLUDE_PATHS:
-    if repository_ctx.path("%s/%scuda.h" % (cuda_toolkit_path, relative_path)).exists:
-        return ("%s/%s" % (cuda_toolkit_path, relative_path))[:-1]
-  auto_configure_fail("Cannot find cuda.h under %s" % cuda_toolkit_path)
+    Args:
+      repository_ctx: The repository context.
+      cuda_config: The CUDA config as returned by _get_cuda_config
 
+    Returns:
+      The path of the directory containing the CUDA headers.
+    """
+    cuda_toolkit_path = cuda_config.cuda_toolkit_path
+    for relative_path in CUDA_INCLUDE_PATHS:
+        if repository_ctx.path("%s/%scuda.h" % (cuda_toolkit_path, relative_path)).exists:
+            return ("%s/%s" % (cuda_toolkit_path, relative_path))[:-1]
+    auto_configure_fail("Cannot find cuda.h under %s" % cuda_toolkit_path)
 
 def _find_cudnn_header_dir(repository_ctx, cudnn_install_basedir):
-  """Returns the path to the directory containing cudnn.h
-
-  Args:
-    repository_ctx: The repository context.
-    cudnn_install_basedir: The cudnn install directory as returned by
-      _cudnn_install_basedir.
+    """Returns the path to the directory containing cudnn.h
 
-  Returns:
-    The path of the directory containing the cudnn header.
-  """
-  for relative_path in CUDA_INCLUDE_PATHS:
-    if repository_ctx.path("%s/%scudnn.h" % (cudnn_install_basedir, relative_path)).exists:
-        return ("%s/%s" % (cudnn_install_basedir, relative_path))[:-1]
-  if repository_ctx.path("/usr/include/cudnn.h").exists:
-    return "/usr/include"
-  auto_configure_fail("Cannot find cudnn.h under %s" % cudnn_install_basedir)
+    Args:
+      repository_ctx: The repository context.
+      cudnn_install_basedir: The cudnn install directory as returned by
+        _cudnn_install_basedir.
 
+    Returns:
+      The path of the directory containing the cudnn header.
+    """
+    for relative_path in CUDA_INCLUDE_PATHS:
+        if repository_ctx.path("%s/%scudnn.h" % (cudnn_install_basedir, relative_path)).exists:
+            return ("%s/%s" % (cudnn_install_basedir, relative_path))[:-1]
+    if repository_ctx.path("/usr/include/cudnn.h").exists:
+        return "/usr/include"
+    auto_configure_fail("Cannot find cudnn.h under %s" % cudnn_install_basedir)
 
 def _find_nvvm_libdevice_dir(repository_ctx, cuda_config):
-  """Returns the path to the directory containing libdevice in bitcode format.
+    """Returns the path to the directory containing libdevice in bitcode format.
 
-  Args:
-    repository_ctx: The repository context.
-    cuda_config: The CUDA config as returned by _get_cuda_config
-
-  Returns:
-    The path of the directory containing the CUDA headers.
-  """
-  cuda_toolkit_path = cuda_config.cuda_toolkit_path
-  for relative_path in NVVM_LIBDEVICE_PATHS:
-    if repository_ctx.path("%s/%slibdevice.10.bc" % (cuda_toolkit_path, relative_path)).exists:
-      return ("%s/%s" % (cuda_toolkit_path, relative_path))[:-1]
-  auto_configure_fail("Cannot find libdevice.10.bc under %s" % cuda_toolkit_path)
+    Args:
+      repository_ctx: The repository context.
+      cuda_config: The CUDA config as returned by _get_cuda_config
 
+    Returns:
+      The path of the directory containing the CUDA headers.
+    """
+    cuda_toolkit_path = cuda_config.cuda_toolkit_path
+    for libdevice_file in NVVM_LIBDEVICE_FILES:
+        for relative_path in NVVM_LIBDEVICE_PATHS:
+            if repository_ctx.path("%s/%s%s" % (cuda_toolkit_path, relative_path, libdevice_file)).exists:
+                return ("%s/%s" % (cuda_toolkit_path, relative_path))[:-1]
+    auto_configure_fail("Cannot find libdevice*.bc files under %s" % cuda_toolkit_path)
 
 def _cudart_static_linkopt(cpu_value):
-  """Returns additional platform-specific linkopts for cudart."""
-  return "" if cpu_value == "Darwin" else "\"-lrt\","
+    """Returns additional platform-specific linkopts for cudart."""
+    return "" if cpu_value == "Darwin" else "\"-lrt\","
 
 def _get_cuda_config(repository_ctx):
-  """Detects and returns information about the CUDA installation on the system.
-
-  Args:
-    repository_ctx: The repository context.
-
-  Returns:
-    A struct containing the following fields:
-      cuda_toolkit_path: The CUDA toolkit installation directory.
-      cudnn_install_basedir: The cuDNN installation directory.
-      cuda_version: The version of CUDA on the system.
-      cudnn_version: The version of cuDNN on the system.
-      compute_capabilities: A list of the system's CUDA compute capabilities.
-      cpu_value: The name of the host operating system.
-  """
-  cpu_value = get_cpu_value(repository_ctx)
-  cuda_toolkit_path = _cuda_toolkit_path(repository_ctx)
-  cuda_version = _cuda_version(repository_ctx, cuda_toolkit_path, cpu_value)
-  cudnn_install_basedir = _cudnn_install_basedir(repository_ctx)
-  cudnn_version = _cudnn_version(repository_ctx, cudnn_install_basedir, cpu_value)
-  return struct(
-      cuda_toolkit_path = cuda_toolkit_path,
-      cudnn_install_basedir = cudnn_install_basedir,
-      cuda_version = cuda_version,
-      cudnn_version = cudnn_version,
-      compute_capabilities = _compute_capabilities(repository_ctx),
-      cpu_value = cpu_value)
-
-
-def _tpl(repository_ctx, tpl, substitutions={}, out=None):
-  if not out:
-    out = tpl.replace(":", "/")
-  repository_ctx.template(
-      out,
-      Label("//third_party/gpus/%s.tpl" % tpl),
-      substitutions)
-
+    """Detects and returns information about the CUDA installation on the system.
+
+    Args:
+      repository_ctx: The repository context.
+
+    Returns:
+      A struct containing the following fields:
+        cuda_toolkit_path: The CUDA toolkit installation directory.
+        cudnn_install_basedir: The cuDNN installation directory.
+        cuda_version: The version of CUDA on the system.
+        cudnn_version: The version of cuDNN on the system.
+        compute_capabilities: A list of the system's CUDA compute capabilities.
+        cpu_value: The name of the host operating system.
+    """
+    cpu_value = get_cpu_value(repository_ctx)
+    cuda_toolkit_path = _cuda_toolkit_path(repository_ctx)
+    cuda_version = _cuda_version(repository_ctx, cuda_toolkit_path, cpu_value)
+    cudnn_install_basedir = _cudnn_install_basedir(repository_ctx)
+    cudnn_version = _cudnn_version(repository_ctx, cudnn_install_basedir, cpu_value)
+    return struct(
+        cuda_toolkit_path = cuda_toolkit_path,
+        cudnn_install_basedir = cudnn_install_basedir,
+        cuda_version = cuda_version,
+        cudnn_version = cudnn_version,
+        compute_capabilities = _compute_capabilities(repository_ctx),
+        cpu_value = cpu_value,
+    )
+
+def _tpl(repository_ctx, tpl, substitutions = {}, out = None):
+    if not out:
+        out = tpl.replace(":", "/")
+    repository_ctx.template(
+        out,
+        Label("//third_party/gpus/%s.tpl" % tpl),
+        substitutions,
+    )
 
 def _file(repository_ctx, label):
-  repository_ctx.template(
-      label.replace(":", "/"),
-      Label("//third_party/gpus/%s.tpl" % label),
-      {})
-
+    repository_ctx.template(
+        label.replace(":", "/"),
+        Label("//third_party/gpus/%s.tpl" % label),
+        {},
+    )
 
 _DUMMY_CROSSTOOL_BZL_FILE = """
 def error_gpu_disabled():
@@ -792,379 +949,498 @@ def error_gpu_disabled():
   )
 """
 
-
 _DUMMY_CROSSTOOL_BUILD_FILE = """
 load("//crosstool:error_gpu_disabled.bzl", "error_gpu_disabled")
 
 error_gpu_disabled()
 """
 
-
 def _create_dummy_repository(repository_ctx):
-  cpu_value = get_cpu_value(repository_ctx)
-
-  # Set up BUILD file for cuda/.
-  _tpl(repository_ctx, "cuda:build_defs.bzl",
-       {
-           "%{cuda_is_configured}": "False",
-           "%{cuda_extra_copts}": "[]",
-       })
-  _tpl(repository_ctx, "cuda:BUILD",
-       {
-           "%{cuda_driver_lib}": _lib_name("cuda", cpu_value),
-           "%{cudart_static_lib}": _lib_name("cudart_static", cpu_value,
-                                             static=True),
-           "%{cudart_static_linkopt}": _cudart_static_linkopt(cpu_value),
-           "%{cudart_lib}": _lib_name("cudart", cpu_value),
-           "%{cublas_lib}": _lib_name("cublas", cpu_value),
-           "%{cusolver_lib}": _lib_name("cusolver", cpu_value),
-           "%{cudnn_lib}": _lib_name("cudnn", cpu_value),
-           "%{cufft_lib}": _lib_name("cufft", cpu_value),
-           "%{curand_lib}": _lib_name("curand", cpu_value),
-           "%{cupti_lib}": _lib_name("cupti", cpu_value),
-           "%{cuda_include_genrules}": '',
-           "%{cuda_headers}": '',
-       })
-
-  # Create dummy files for the CUDA toolkit since they are still required by
-  # tensorflow/core/platform/default/build_config:cuda.
-  repository_ctx.file("cuda/cuda/include/cuda.h", "")
-  repository_ctx.file("cuda/cuda/include/cublas.h", "")
-  repository_ctx.file("cuda/cuda/include/cudnn.h", "")
-  repository_ctx.file("cuda/cuda/extras/CUPTI/include/cupti.h", "")
-  repository_ctx.file("cuda/cuda/lib/%s" % _lib_name("cuda", cpu_value))
-  repository_ctx.file("cuda/cuda/lib/%s" % _lib_name("cudart", cpu_value))
-  repository_ctx.file("cuda/cuda/lib/%s" % _lib_name("cudart_static", cpu_value))
-  repository_ctx.file("cuda/cuda/lib/%s" % _lib_name("cublas", cpu_value))
-  repository_ctx.file("cuda/cuda/lib/%s" % _lib_name("cusolver", cpu_value))
-  repository_ctx.file("cuda/cuda/lib/%s" % _lib_name("cudnn", cpu_value))
-  repository_ctx.file("cuda/cuda/lib/%s" % _lib_name("curand", cpu_value))
-  repository_ctx.file("cuda/cuda/lib/%s" % _lib_name("cufft", cpu_value))
-  repository_ctx.file("cuda/cuda/lib/%s" % _lib_name("cupti", cpu_value))
-
-  # Set up cuda_config.h, which is used by
-  # tensorflow/stream_executor/dso_loader.cc.
-  _tpl(repository_ctx, "cuda:cuda_config.h",
-       {
-           "%{cuda_version}": _DEFAULT_CUDA_VERSION,
-           "%{cudnn_version}": _DEFAULT_CUDNN_VERSION,
-           "%{cuda_compute_capabilities}": ",".join([
-               "CudaVersion(\"%s\")" % c
-               for c in _DEFAULT_CUDA_COMPUTE_CAPABILITIES]),
-           "%{cuda_toolkit_path}": _DEFAULT_CUDA_TOOLKIT_PATH,
-       }, "cuda/cuda/cuda_config.h")
-
-  # If cuda_configure is not configured to build with GPU support, and the user
-  # attempts to build with --config=cuda, add a dummy build rule to intercept
-  # this and fail with an actionable error message.
-  repository_ctx.file("crosstool/error_gpu_disabled.bzl",
-                      _DUMMY_CROSSTOOL_BZL_FILE)
-  repository_ctx.file("crosstool/BUILD", _DUMMY_CROSSTOOL_BUILD_FILE)
-
-
-def _execute(repository_ctx, cmdline, error_msg=None, error_details=None,
-             empty_stdout_fine=False):
-  """Executes an arbitrary shell command.
-
-  Args:
-    repository_ctx: the repository_ctx object
-    cmdline: list of strings, the command to execute
-    error_msg: string, a summary of the error if the command fails
-    error_details: string, details about the error or steps to fix it
-    empty_stdout_fine: bool, if True, an empty stdout result is fine, otherwise
-      it's an error
-  Return:
-    the result of repository_ctx.execute(cmdline)
-  """
-  result = repository_ctx.execute(cmdline)
-  if result.stderr or not (empty_stdout_fine or result.stdout):
-    auto_configure_fail(
-        "\n".join([
-            error_msg.strip() if error_msg else "Repository command failed",
-            result.stderr.strip(),
-            error_details if error_details else ""]))
-  return result
-
+    cpu_value = get_cpu_value(repository_ctx)
+
+    # Set up BUILD file for cuda/.
+    _tpl(
+        repository_ctx,
+        "cuda:build_defs.bzl",
+        {
+            "%{cuda_is_configured}": "False",
+            "%{cuda_extra_copts}": "[]",
+        },
+    )
+    _tpl(
+        repository_ctx,
+        "cuda:BUILD",
+        {
+            "%{cuda_driver_lib}": _lib_name("cuda", cpu_value),
+            "%{cudart_static_lib}": _lib_name(
+                "cudart_static",
+                cpu_value,
+                static = True,
+            ),
+            "%{cudart_static_linkopt}": _cudart_static_linkopt(cpu_value),
+            "%{cudart_lib}": _lib_name("cudart", cpu_value),
+            "%{cublas_lib}": _lib_name("cublas", cpu_value),
+            "%{cusolver_lib}": _lib_name("cusolver", cpu_value),
+            "%{cudnn_lib}": _lib_name("cudnn", cpu_value),
+            "%{cufft_lib}": _lib_name("cufft", cpu_value),
+            "%{curand_lib}": _lib_name("curand", cpu_value),
+            "%{cupti_lib}": _lib_name("cupti", cpu_value),
+            "%{cuda_include_genrules}": "",
+            "%{cuda_headers}": "",
+        },
+    )
+
+    # Create dummy files for the CUDA toolkit since they are still required by
+    # tensorflow/core/platform/default/build_config:cuda.
+    repository_ctx.file("cuda/cuda/include/cuda.h", "")
+    repository_ctx.file("cuda/cuda/include/cublas.h", "")
+    repository_ctx.file("cuda/cuda/include/cudnn.h", "")
+    repository_ctx.file("cuda/cuda/extras/CUPTI/include/cupti.h", "")
+    repository_ctx.file("cuda/cuda/lib/%s" % _lib_name("cuda", cpu_value))
+    repository_ctx.file("cuda/cuda/lib/%s" % _lib_name("cudart", cpu_value))
+    repository_ctx.file("cuda/cuda/lib/%s" % _lib_name("cudart_static", cpu_value))
+    repository_ctx.file("cuda/cuda/lib/%s" % _lib_name("cublas", cpu_value))
+    repository_ctx.file("cuda/cuda/lib/%s" % _lib_name("cusolver", cpu_value))
+    repository_ctx.file("cuda/cuda/lib/%s" % _lib_name("cudnn", cpu_value))
+    repository_ctx.file("cuda/cuda/lib/%s" % _lib_name("curand", cpu_value))
+    repository_ctx.file("cuda/cuda/lib/%s" % _lib_name("cufft", cpu_value))
+    repository_ctx.file("cuda/cuda/lib/%s" % _lib_name("cupti", cpu_value))
+
+    # Set up cuda_config.h, which is used by
+    # tensorflow/stream_executor/dso_loader.cc.
+    _tpl(
+        repository_ctx,
+        "cuda:cuda_config.h",
+        {
+            "%{cuda_version}": _DEFAULT_CUDA_VERSION,
+            "%{cudnn_version}": _DEFAULT_CUDNN_VERSION,
+            "%{cuda_compute_capabilities}": ",".join([
+                "CudaVersion(\"%s\")" % c
+                for c in _DEFAULT_CUDA_COMPUTE_CAPABILITIES
+            ]),
+            "%{cuda_toolkit_path}": _DEFAULT_CUDA_TOOLKIT_PATH,
+        },
+        "cuda/cuda/cuda_config.h",
+    )
+
+    # If cuda_configure is not configured to build with GPU support, and the user
+    # attempts to build with --config=cuda, add a dummy build rule to intercept
+    # this and fail with an actionable error message.
+    repository_ctx.file(
+        "crosstool/error_gpu_disabled.bzl",
+        _DUMMY_CROSSTOOL_BZL_FILE,
+    )
+    repository_ctx.file("crosstool/BUILD", _DUMMY_CROSSTOOL_BUILD_FILE)
+
+def _execute(
+        repository_ctx,
+        cmdline,
+        error_msg = None,
+        error_details = None,
+        empty_stdout_fine = False):
+    """Executes an arbitrary shell command.
+
+    Args:
+      repository_ctx: the repository_ctx object
+      cmdline: list of strings, the command to execute
+      error_msg: string, a summary of the error if the command fails
+      error_details: string, details about the error or steps to fix it
+      empty_stdout_fine: bool, if True, an empty stdout result is fine, otherwise
+        it's an error
+    Return:
+      the result of repository_ctx.execute(cmdline)
+    """
+    result = repository_ctx.execute(cmdline)
+    if result.stderr or not (empty_stdout_fine or result.stdout):
+        auto_configure_fail(
+            "\n".join([
+                error_msg.strip() if error_msg else "Repository command failed",
+                result.stderr.strip(),
+                error_details if error_details else "",
+            ]),
+        )
+    return result
 
 def _norm_path(path):
-  """Returns a path with '/' and remove the trailing slash."""
-  path = path.replace("\\", "/")
-  if path[-1] == "/":
-    path = path[:-1]
-  return path
-
-
-def symlink_genrule_for_dir(repository_ctx, src_dir, dest_dir, genrule_name,
-                            src_files = [], dest_files = []):
-  """Returns a genrule to symlink(or copy if on Windows) a set of files.
-
-  If src_dir is passed, files will be read from the given directory; otherwise
-  we assume files are in src_files and dest_files
-  """
-  if src_dir != None:
-    src_dir = _norm_path(src_dir)
-    dest_dir = _norm_path(dest_dir)
-    files = '\n'.join(sorted(_read_dir(repository_ctx, src_dir).splitlines()))
-    # Create a list with the src_dir stripped to use for outputs.
-    dest_files = files.replace(src_dir, '').splitlines()
-    src_files = files.splitlines()
-  command = []
-  if not _is_windows(repository_ctx):
-    # We clear folders that might have been generated previously to avoid
-    # undesired inclusions
-    command.append('if [ -d "$(@D)/extras" ]; then rm $(@D)/extras -drf; fi')
-    command.append('if [ -d "$(@D)/include" ]; then rm $(@D)/include -drf; fi')
-    command.append('if [ -d "$(@D)/lib" ]; then rm $(@D)/lib -drf; fi')
-    command.append('if [ -d "$(@D)/nvvm" ]; then rm $(@D)/nvvm -drf; fi')
-  outs = []
-  for i in range(len(dest_files)):
-    if dest_files[i] != "":
-      # If we have only one file to link we do not want to use the dest_dir, as
-      # $(@D) will include the full path to the file.
-      dest = '$(@D)/' + dest_dir + dest_files[i] if len(dest_files) != 1 else '$(@D)/' + dest_files[i]
-      # On Windows, symlink is not supported, so we just copy all the files.
-      cmd = 'cp -f' if _is_windows(repository_ctx) else 'ln -s'
-      command.append(cmd + ' "%s" "%s"' % (src_files[i] , dest))
-      outs.append('        "' + dest_dir + dest_files[i] + '",')
-  genrule = _genrule(src_dir, genrule_name, " && ".join(command),
-                     "\n".join(outs))
-  return genrule
-
+    """Returns a path with '/' and remove the trailing slash."""
+    path = path.replace("\\", "/")
+    if path[-1] == "/":
+        path = path[:-1]
+    return path
+
+def symlink_genrule_for_dir(
+        repository_ctx,
+        src_dir,
+        dest_dir,
+        genrule_name,
+        src_files = [],
+        dest_files = []):
+    """Returns a genrule to symlink(or copy if on Windows) a set of files.
+
+    If src_dir is passed, files will be read from the given directory; otherwise
+    we assume files are in src_files and dest_files
+    """
+    if src_dir != None:
+        src_dir = _norm_path(src_dir)
+        dest_dir = _norm_path(dest_dir)
+        files = "\n".join(sorted(_read_dir(repository_ctx, src_dir).splitlines()))
+
+        # Create a list with the src_dir stripped to use for outputs.
+        dest_files = files.replace(src_dir, "").splitlines()
+        src_files = files.splitlines()
+    command = []
+    if not _is_windows(repository_ctx):
+        # We clear folders that might have been generated previously to avoid
+        # undesired inclusions
+        command.append('if [ -d "$(@D)/extras" ]; then rm $(@D)/extras -drf; fi')
+        command.append('if [ -d "$(@D)/include" ]; then rm $(@D)/include -drf; fi')
+        command.append('if [ -d "$(@D)/lib" ]; then rm $(@D)/lib -drf; fi')
+        command.append('if [ -d "$(@D)/nvvm" ]; then rm $(@D)/nvvm -drf; fi')
+    outs = []
+    for i in range(len(dest_files)):
+        if dest_files[i] != "":
+            # If we have only one file to link we do not want to use the dest_dir, as
+            # $(@D) will include the full path to the file.
+            dest = "$(@D)/" + dest_dir + dest_files[i] if len(dest_files) != 1 else "$(@D)/" + dest_files[i]
+
+            # On Windows, symlink is not supported, so we just copy all the files.
+            cmd = "cp -f" if _is_windows(repository_ctx) else "ln -s"
+            command.append(cmd + ' "%s" "%s"' % (src_files[i], dest))
+            outs.append('        "' + dest_dir + dest_files[i] + '",')
+    genrule = _genrule(
+        src_dir,
+        genrule_name,
+        " && ".join(command),
+        "\n".join(outs),
+    )
+    return genrule
 
 def _genrule(src_dir, genrule_name, command, outs):
-  """Returns a string with a genrule.
-
-  Genrule executes the given command and produces the given outputs.
-  """
-  return (
-      'genrule(\n' +
-      '    name = "' +
-      genrule_name + '",\n' +
-      '    outs = [\n' +
-      outs +
-      '\n    ],\n' +
-      '    cmd = """\n' +
-      command +
-      '\n   """,\n' +
-      ')\n'
-  )
+    """Returns a string with a genrule.
 
+    Genrule executes the given command and produces the given outputs.
+    """
+    return (
+        "genrule(\n" +
+        '    name = "' +
+        genrule_name + '",\n' +
+        "    outs = [\n" +
+        outs +
+        "\n    ],\n" +
+        '    cmd = """\n' +
+        command +
+        '\n   """,\n' +
+        ")\n"
+    )
 
 def _read_dir(repository_ctx, src_dir):
-  """Returns a string with all files in a directory.
-
-  Finds all files inside a directory, traversing subfolders and following
-  symlinks. The returned string contains the full path of all files
-  separated by line breaks.
-  """
-  if _is_windows(repository_ctx):
-    src_dir = src_dir.replace("/", "\\")
-    find_result = _execute(
-        repository_ctx, ["cmd.exe", "/c", "dir", src_dir, "/b", "/s", "/a-d"],
-        empty_stdout_fine=True)
-    # src_files will be used in genrule.outs where the paths must
-    # use forward slashes.
-    result = find_result.stdout.replace("\\", "/")
-  else:
-    find_result = _execute(
-        repository_ctx, ["find", src_dir, "-follow", "-type", "f"],
-        empty_stdout_fine=True)
-    result = find_result.stdout
-  return result
+    """Returns a string with all files in a directory.
+
+    Finds all files inside a directory, traversing subfolders and following
+    symlinks. The returned string contains the full path of all files
+    separated by line breaks.
+    """
+    if _is_windows(repository_ctx):
+        src_dir = src_dir.replace("/", "\\")
+        find_result = _execute(
+            repository_ctx,
+            ["cmd.exe", "/c", "dir", src_dir, "/b", "/s", "/a-d"],
+            empty_stdout_fine = True,
+        )
+
+        # src_files will be used in genrule.outs where the paths must
+        # use forward slashes.
+        result = find_result.stdout.replace("\\", "/")
+    else:
+        find_result = _execute(
+            repository_ctx,
+            ["find", src_dir, "-follow", "-type", "f"],
+            empty_stdout_fine = True,
+        )
+        result = find_result.stdout
+    return result
 
 def _flag_enabled(repository_ctx, flag_name):
-  if flag_name in repository_ctx.os.environ:
-    value = repository_ctx.os.environ[flag_name].strip()
-    return value == "1"
-  return False
+    if flag_name in repository_ctx.os.environ:
+        value = repository_ctx.os.environ[flag_name].strip()
+        return value == "1"
+    return False
 
 def _use_cuda_clang(repository_ctx):
-  return _flag_enabled(repository_ctx, "TF_CUDA_CLANG")
+    return _flag_enabled(repository_ctx, "TF_CUDA_CLANG")
 
 def _compute_cuda_extra_copts(repository_ctx, compute_capabilities):
-  if _use_cuda_clang(repository_ctx):
-    capability_flags = ["--cuda-gpu-arch=sm_" +
-        cap.replace(".", "") for cap in compute_capabilities]
-  else:
-    # Capabilities are handled in the "crosstool_wrapper_driver_is_not_gcc" for nvcc
-    capability_flags = []
-  return str(capability_flags)
+    if _use_cuda_clang(repository_ctx):
+        capability_flags = ["--cuda-gpu-arch=sm_" +
+                            cap.replace(".", "") for cap in compute_capabilities]
+    else:
+        # Capabilities are handled in the "crosstool_wrapper_driver_is_not_gcc" for nvcc
+        capability_flags = []
+    return str(capability_flags)
 
 def _create_local_cuda_repository(repository_ctx):
-  """Creates the repository containing files set up to build with CUDA."""
-  cuda_config = _get_cuda_config(repository_ctx)
-
-  cuda_include_path = _find_cuda_include_path(repository_ctx, cuda_config)
-  cudnn_header_dir = _find_cudnn_header_dir(repository_ctx,
-                                            cuda_config.cudnn_install_basedir)
-  cupti_header_dir = _find_cupti_header_dir(repository_ctx, cuda_config)
-  nvvm_libdevice_dir = _find_nvvm_libdevice_dir(repository_ctx, cuda_config)
-
-  # Set up symbolic links for the cuda toolkit by creating genrules to do
-  # symlinking. We create one genrule for each directory we want to track under
-  # cuda_toolkit_path
-  cuda_toolkit_path = cuda_config.cuda_toolkit_path
-  genrules = [symlink_genrule_for_dir(repository_ctx,
-      cuda_include_path, "cuda/include", "cuda-include")]
-  genrules.append(symlink_genrule_for_dir(repository_ctx,
-      nvvm_libdevice_dir, "cuda/nvvm/libdevice", "cuda-nvvm"))
-  genrules.append(symlink_genrule_for_dir(repository_ctx,
-      cupti_header_dir, "cuda/extras/CUPTI/include", "cuda-extras"))
-
-  cuda_libs = _find_libs(repository_ctx, cuda_config)
-  cuda_lib_src = []
-  cuda_lib_dest = []
-  for lib in cuda_libs.values():
-    cuda_lib_src.append(lib.path)
-    cuda_lib_dest.append("cuda/lib/" + lib.file_name)
-  genrules.append(symlink_genrule_for_dir(repository_ctx, None, "", "cuda-lib",
-                                          cuda_lib_src, cuda_lib_dest))
-
-  # Set up the symbolic links for cudnn if cndnn was not installed to
-  # CUDA_TOOLKIT_PATH.
-  included_files = _read_dir(repository_ctx, cuda_include_path).replace(
-      cuda_include_path, '').splitlines()
-  if '/cudnn.h' not in included_files:
-    genrules.append(symlink_genrule_for_dir(repository_ctx, None,
-        "cuda/include/", "cudnn-include", [cudnn_header_dir + "/cudnn.h"],
-        ["cudnn.h"]))
-  else:
-    genrules.append(
-            'filegroup(\n' +
+    """Creates the repository containing files set up to build with CUDA."""
+    cuda_config = _get_cuda_config(repository_ctx)
+
+    cuda_include_path = _find_cuda_include_path(repository_ctx, cuda_config)
+    cudnn_header_dir = _find_cudnn_header_dir(
+        repository_ctx,
+        cuda_config.cudnn_install_basedir,
+    )
+    cupti_header_dir = _find_cupti_header_dir(repository_ctx, cuda_config)
+    nvvm_libdevice_dir = _find_nvvm_libdevice_dir(repository_ctx, cuda_config)
+
+    # Set up symbolic links for the cuda toolkit by creating genrules to do
+    # symlinking. We create one genrule for each directory we want to track under
+    # cuda_toolkit_path
+    cuda_toolkit_path = cuda_config.cuda_toolkit_path
+    genrules = [symlink_genrule_for_dir(
+        repository_ctx,
+        cuda_include_path,
+        "cuda/include",
+        "cuda-include",
+    )]
+    genrules.append(symlink_genrule_for_dir(
+        repository_ctx,
+        nvvm_libdevice_dir,
+        "cuda/nvvm/libdevice",
+        "cuda-nvvm",
+    ))
+    genrules.append(symlink_genrule_for_dir(
+        repository_ctx,
+        cupti_header_dir,
+        "cuda/extras/CUPTI/include",
+        "cuda-extras",
+    ))
+
+    cuda_libs = _find_libs(repository_ctx, cuda_config)
+    cuda_lib_src = []
+    cuda_lib_dest = []
+    for lib in cuda_libs.values():
+        cuda_lib_src.append(lib.path)
+        cuda_lib_dest.append("cuda/lib/" + lib.file_name)
+    genrules.append(symlink_genrule_for_dir(
+        repository_ctx,
+        None,
+        "",
+        "cuda-lib",
+        cuda_lib_src,
+        cuda_lib_dest,
+    ))
+
+    # Set up the symbolic links for cudnn if cndnn was not installed to
+    # CUDA_TOOLKIT_PATH.
+    included_files = _read_dir(repository_ctx, cuda_include_path).replace(
+        cuda_include_path,
+        "",
+    ).splitlines()
+    if "/cudnn.h" not in included_files:
+        genrules.append(symlink_genrule_for_dir(
+            repository_ctx,
+            None,
+            "cuda/include/",
+            "cudnn-include",
+            [cudnn_header_dir + "/cudnn.h"],
+            ["cudnn.h"],
+        ))
+    else:
+        genrules.append(
+            "filegroup(\n" +
             '    name = "cudnn-include",\n' +
-            '    srcs = [],\n' +
-            ')\n'
+            "    srcs = [],\n" +
+            ")\n",
         )
 
-  # Set up BUILD file for cuda/
-  _tpl(repository_ctx, "cuda:build_defs.bzl",
-       {
-           "%{cuda_is_configured}": "True",
-           "%{cuda_extra_copts}": _compute_cuda_extra_copts(
-               repository_ctx, cuda_config.compute_capabilities),
-       })
-  _tpl(repository_ctx, "cuda:BUILD",
-       {
-           "%{cuda_driver_lib}": cuda_libs["cuda"].file_name,
-           "%{cudart_static_lib}": cuda_libs["cudart_static"].file_name,
-           "%{cudart_static_linkopt}": _cudart_static_linkopt(
-               cuda_config.cpu_value),
-           "%{cudart_lib}": cuda_libs["cudart"].file_name,
-           "%{cublas_lib}": cuda_libs["cublas"].file_name,
-           "%{cusolver_lib}": cuda_libs["cusolver"].file_name,
-           "%{cudnn_lib}": cuda_libs["cudnn"].file_name,
-           "%{cufft_lib}": cuda_libs["cufft"].file_name,
-           "%{curand_lib}": cuda_libs["curand"].file_name,
-           "%{cupti_lib}": cuda_libs["cupti"].file_name,
-           "%{cuda_include_genrules}": "\n".join(genrules),
-           "%{cuda_headers}": ('":cuda-include",\n' +
-                               '        ":cudnn-include",')
-       })
-
-  is_cuda_clang = _use_cuda_clang(repository_ctx)
-
-  should_download_clang = is_cuda_clang and _flag_enabled(
-      repository_ctx, _TF_DOWNLOAD_CLANG)
-  if should_download_clang:
-    download_clang(repository_ctx, "crosstool/extra_tools")
-
-  # Set up crosstool/
-  cc = find_cc(repository_ctx)
-  cc_fullpath = cc if not should_download_clang else "crosstool/" + cc
-
-  host_compiler_includes = _host_compiler_includes(repository_ctx, cc_fullpath)
-  cuda_defines = {}
-  if is_cuda_clang:
-    cuda_defines["%{host_compiler_path}"] = str(cc)
-    cuda_defines["%{host_compiler_warnings}"] = """
+    # Set up BUILD file for cuda/
+    _tpl(
+        repository_ctx,
+        "cuda:build_defs.bzl",
+        {
+            "%{cuda_is_configured}": "True",
+            "%{cuda_extra_copts}": _compute_cuda_extra_copts(
+                repository_ctx,
+                cuda_config.compute_capabilities,
+            ),
+        },
+    )
+    _tpl(
+        repository_ctx,
+        "cuda:BUILD.windows" if _is_windows(repository_ctx) else "cuda:BUILD",
+        {
+            "%{cuda_driver_lib}": cuda_libs["cuda"].file_name,
+            "%{cudart_static_lib}": cuda_libs["cudart_static"].file_name,
+            "%{cudart_static_linkopt}": _cudart_static_linkopt(
+                cuda_config.cpu_value,
+            ),
+            "%{cudart_lib}": cuda_libs["cudart"].file_name,
+            "%{cublas_lib}": cuda_libs["cublas"].file_name,
+            "%{cusolver_lib}": cuda_libs["cusolver"].file_name,
+            "%{cudnn_lib}": cuda_libs["cudnn"].file_name,
+            "%{cufft_lib}": cuda_libs["cufft"].file_name,
+            "%{curand_lib}": cuda_libs["curand"].file_name,
+            "%{cupti_lib}": cuda_libs["cupti"].file_name,
+            "%{cuda_include_genrules}": "\n".join(genrules),
+            "%{cuda_headers}": ('":cuda-include",\n' +
+                                '        ":cudnn-include",'),
+        },
+        "cuda/BUILD",
+    )
+
+    is_cuda_clang = _use_cuda_clang(repository_ctx)
+
+    should_download_clang = is_cuda_clang and _flag_enabled(
+        repository_ctx,
+        _TF_DOWNLOAD_CLANG,
+    )
+    if should_download_clang:
+        download_clang(repository_ctx, "crosstool/extra_tools")
+
+    # Set up crosstool/
+    cc = find_cc(repository_ctx)
+    cc_fullpath = cc if not should_download_clang else "crosstool/" + cc
+
+    host_compiler_includes = _host_compiler_includes(repository_ctx, cc_fullpath)
+    cuda_defines = {}
+    if is_cuda_clang:
+        cuda_defines["%{host_compiler_path}"] = str(cc)
+        cuda_defines["%{host_compiler_warnings}"] = """
         # Some parts of the codebase set -Werror and hit this warning, so
         # switch it off for now.
         flag: "-Wno-invalid-partial-specialization"
     """
-    cuda_defines["%{host_compiler_includes}"] = host_compiler_includes
-    _tpl(repository_ctx, "crosstool:BUILD", {"%{linker_files}": ":empty"})
-    repository_ctx.file("crosstool/clang/bin/crosstool_wrapper_driver_is_not_gcc", "")
-  else:
-    cuda_defines["%{host_compiler_path}"] = "clang/bin/crosstool_wrapper_driver_is_not_gcc"
-    cuda_defines["%{host_compiler_warnings}"] = ""
-    # TODO(klimek): We currently need to inject "/" as builtin directory path
-    # to disable bazel's dependency checks.
-    # The problem is that:
-    # - the python rules symlink the python headers into the bazel root
-    # - the rules use 'includes' in the BUILD file to redirect includes of the
-    #   python headers through those paths
-    # - bazel currently uses -isystem for include paths specified via 'includes'
-    # - gcc follows symlinks when resolving files via -isystem paths, and puts
-    #   the resolved paths into the .d file, which makes the dependency check
-    #   fail for bazel
-    # There are multiple possible ways to solve this:
-    # 1. make bazel not use -isystem for paths specified via 'includes'
-    # 2. cp the headers instead of symlinking them
-    #
-    # Once this is fixed, the right builtin directory path is:
-    # (host_compiler_includes +
-    #    "\n  cxx_builtin_include_directory: \"%s\"" % cuda_include_path)
-    # The cuda directory needs to be passed, as there is currently no rule
-    # providing the cuda headers in the same way the python headers are
-    # provided.
-    cuda_defines["%{host_compiler_includes}"] = "\n  cxx_builtin_include_directory: \"/\""
-    nvcc_path = str(repository_ctx.path("%s/bin/nvcc%s" %
-        (cuda_config.cuda_toolkit_path,
-        ".exe" if cuda_config.cpu_value == "Windows" else "")))
-    _tpl(repository_ctx, "crosstool:BUILD",
-         {"%{linker_files}": ":crosstool_wrapper_driver_is_not_gcc"})
-    _tpl(repository_ctx,
-         "crosstool:clang/bin/crosstool_wrapper_driver_is_not_gcc",
-         {
-             "%{cpu_compiler}": str(cc),
-             "%{cuda_version}": cuda_config.cuda_version,
-             "%{nvcc_path}": nvcc_path,
-             "%{gcc_host_compiler_path}": str(cc),
-             "%{cuda_compute_capabilities}": ", ".join(
-                 ["\"%s\"" % c for c in cuda_config.compute_capabilities]),
-         })
-  _tpl(repository_ctx, "crosstool:CROSSTOOL", cuda_defines, out="crosstool/CROSSTOOL")
-
-  # Set up cuda_config.h, which is used by
-  # tensorflow/stream_executor/dso_loader.cc.
-  _tpl(repository_ctx, "cuda:cuda_config.h",
-       {
-           "%{cuda_version}": cuda_config.cuda_version,
-           "%{cudnn_version}": cuda_config.cudnn_version,
-           "%{cuda_compute_capabilities}": ",".join(
-               ["CudaVersion(\"%s\")" % c
-                for c in cuda_config.compute_capabilities]),
-               "%{cuda_toolkit_path}": cuda_config.cuda_toolkit_path,
-       }, "cuda/cuda/cuda_config.h")
+        cuda_defines["%{host_compiler_includes}"] = host_compiler_includes
+        _tpl(repository_ctx, "crosstool:BUILD", {"%{linker_files}": ":empty", "%{win_linker_files}": ":empty"})
+        repository_ctx.file("crosstool/clang/bin/crosstool_wrapper_driver_is_not_gcc", "")
+        repository_ctx.file("crosstool/windows/msvc_wrapper_for_nvcc.py", "")
+        repository_ctx.file("crosstool/windows/msvc_wrapper_for_nvcc.bat", "")
+    else:
+        cuda_defines["%{host_compiler_path}"] = "clang/bin/crosstool_wrapper_driver_is_not_gcc"
+        cuda_defines["%{host_compiler_warnings}"] = ""
+
+        # TODO(klimek): We currently need to inject "/" as builtin directory path
+        # to disable bazel's dependency checks.
+        # The problem is that:
+        # - the python rules symlink the python headers into the bazel root
+        # - the rules use 'includes' in the BUILD file to redirect includes of the
+        #   python headers through those paths
+        # - bazel currently uses -isystem for include paths specified via 'includes'
+        # - gcc follows symlinks when resolving files via -isystem paths, and puts
+        #   the resolved paths into the .d file, which makes the dependency check
+        #   fail for bazel
+        # There are multiple possible ways to solve this:
+        # 1. make bazel not use -isystem for paths specified via 'includes'
+        # 2. cp the headers instead of symlinking them
+        #
+        # Once this is fixed, the right builtin directory path is:
+        # (host_compiler_includes +
+        #    "\n  cxx_builtin_include_directory: \"%s\"" % cuda_include_path)
+        # The cuda directory needs to be passed, as there is currently no rule
+        # providing the cuda headers in the same way the python headers are
+        # provided.
+        cuda_defines["%{host_compiler_includes}"] = "\n  cxx_builtin_include_directory: \"/\""
+        nvcc_path = str(repository_ctx.path("%s/bin/nvcc%s" %
+                                            (
+                                                cuda_config.cuda_toolkit_path,
+                                                ".exe" if _is_windows(repository_ctx) else "",
+                                            )))
+        _tpl(
+            repository_ctx,
+            "crosstool:BUILD",
+            {
+                "%{linker_files}": ":crosstool_wrapper_driver_is_not_gcc",
+                "%{win_linker_files}": ":windows_msvc_wrapper_files",
+            },
+        )
+        wrapper_defines = {
+            "%{cpu_compiler}": str(cc),
+            "%{cuda_version}": cuda_config.cuda_version,
+            "%{nvcc_path}": nvcc_path,
+            "%{gcc_host_compiler_path}": str(cc),
+            "%{cuda_compute_capabilities}": ", ".join(
+                ["\"%s\"" % c for c in cuda_config.compute_capabilities],
+            ),
+            "%{nvcc_tmp_dir}": _get_nvcc_tmp_dir_for_windows(repository_ctx),
+        }
+        _tpl(
+            repository_ctx,
+            "crosstool:clang/bin/crosstool_wrapper_driver_is_not_gcc",
+            wrapper_defines,
+        )
+        _tpl(
+            repository_ctx,
+            "crosstool:windows/msvc_wrapper_for_nvcc.py",
+            wrapper_defines,
+        )
+        _tpl(
+            repository_ctx,
+            "crosstool:windows/msvc_wrapper_for_nvcc.bat",
+            {
+                "%{python_binary}": _get_python_bin(repository_ctx),
+            },
+        )
+
+    _tpl(
+        repository_ctx,
+        "crosstool:CROSSTOOL",
+        cuda_defines + _get_win_cuda_defines(repository_ctx),
+        out = "crosstool/CROSSTOOL",
+    )
+
+    # Set up cuda_config.h, which is used by
+    # tensorflow/stream_executor/dso_loader.cc.
+    _tpl(
+        repository_ctx,
+        "cuda:cuda_config.h",
+        {
+            "%{cuda_version}": cuda_config.cuda_version,
+            "%{cudnn_version}": cuda_config.cudnn_version,
+            "%{cuda_compute_capabilities}": ",".join(
+                [
+                    "CudaVersion(\"%s\")" % c
+                    for c in cuda_config.compute_capabilities
+                ],
+            ),
+            "%{cuda_toolkit_path}": cuda_config.cuda_toolkit_path,
+        },
+        "cuda/cuda/cuda_config.h",
+    )
 
 def _create_remote_cuda_repository(repository_ctx, remote_config_repo):
-  """Creates pointers to a remotely configured repo set up to build with CUDA."""
-  _tpl(repository_ctx, "cuda:build_defs.bzl",
-       {
-           "%{cuda_is_configured}": "True",
-           "%{cuda_extra_copts}": _compute_cuda_extra_copts(
-               repository_ctx, _compute_capabilities(repository_ctx)),
-
-       })
-  _tpl(repository_ctx, "cuda:remote.BUILD",
-       {
-           "%{remote_cuda_repo}": remote_config_repo,
-       }, "cuda/BUILD")
-  _tpl(repository_ctx, "crosstool:remote.BUILD", {
-           "%{remote_cuda_repo}": remote_config_repo,
-       }, "crosstool/BUILD")
+    """Creates pointers to a remotely configured repo set up to build with CUDA."""
+    _tpl(
+        repository_ctx,
+        "cuda:build_defs.bzl",
+        {
+            "%{cuda_is_configured}": "True",
+            "%{cuda_extra_copts}": _compute_cuda_extra_copts(
+                repository_ctx,
+                _compute_capabilities(repository_ctx),
+            ),
+        },
+    )
+    _tpl(
+        repository_ctx,
+        "cuda:remote.BUILD",
+        {
+            "%{remote_cuda_repo}": remote_config_repo,
+        },
+        "cuda/BUILD",
+    )
+    _tpl(repository_ctx, "crosstool:remote.BUILD", {
+        "%{remote_cuda_repo}": remote_config_repo,
+    }, "crosstool/BUILD")
 
 def _cuda_autoconf_impl(repository_ctx):
-  """Implementation of the cuda_autoconf repository rule."""
-  if not _enable_cuda(repository_ctx):
-    _create_dummy_repository(repository_ctx)
-  else:
-    if _TF_CUDA_CONFIG_REPO in repository_ctx.os.environ:
-      _create_remote_cuda_repository(repository_ctx,
-          repository_ctx.os.environ[_TF_CUDA_CONFIG_REPO])
+    """Implementation of the cuda_autoconf repository rule."""
+    if not _enable_cuda(repository_ctx):
+        _create_dummy_repository(repository_ctx)
+    elif _TF_CUDA_CONFIG_REPO in repository_ctx.os.environ:
+        _create_remote_cuda_repository(
+            repository_ctx,
+            repository_ctx.os.environ[_TF_CUDA_CONFIG_REPO],
+        )
     else:
-      _create_local_cuda_repository(repository_ctx)
-
+        _create_local_cuda_repository(repository_ctx)
 
 cuda_configure = repository_rule(
     implementation = _cuda_autoconf_impl,
@@ -1181,6 +1457,7 @@ cuda_configure = repository_rule(
         _TF_CUDA_COMPUTE_CAPABILITIES,
         _TF_CUDA_CONFIG_REPO,
         "NVVMIR_LIBRARY_DIR",
+        _PYTHON_BIN_PATH,
     ],
 )
 
diff --git a/third_party/kafka/BUILD b/third_party/kafka/BUILD
index a839ca717e695f35fac684b510f0a022010e0710..75792b0d87366c304ca29f95f943114ee482dfcd 100644
--- a/third_party/kafka/BUILD
+++ b/third_party/kafka/BUILD
@@ -60,6 +60,8 @@ cc_library(
         "src/rdkafka_event.h",
         "src/rdkafka_feature.c",
         "src/rdkafka_feature.h",
+        "src/rdkafka_header.c",
+        "src/rdkafka_header.h",
         "src/rdkafka_int.h",
         "src/rdkafka_interceptor.c",
         "src/rdkafka_interceptor.h",
@@ -93,7 +95,6 @@ cc_library(
         "src/rdkafka_sasl_int.h",
         "src/rdkafka_sasl_plain.c",
         "src/rdkafka_subscription.c",
-        "src/rdkafka_subscription.h",
         "src/rdkafka_timer.c",
         "src/rdkafka_timer.h",
         "src/rdkafka_topic.c",
@@ -105,6 +106,8 @@ cc_library(
         "src/rdlist.h",
         "src/rdlog.c",
         "src/rdlog.h",
+        "src/rdmurmur2.c",
+        "src/rdmurmur2.h",
         "src/rdports.c",
         "src/rdports.h",
         "src/rdposix.h",
diff --git a/third_party/llvm/llvm.autogenerated.BUILD b/third_party/llvm/llvm.autogenerated.BUILD
index 4f645fa260067584208e7a9681ef1ccf92d74ac0..bf9f9ca9cfeeea99f3d56c692146a51730c0f395 100644
--- a/third_party/llvm/llvm.autogenerated.BUILD
+++ b/third_party/llvm/llvm.autogenerated.BUILD
@@ -11,7 +11,11 @@ load(
     "cmake_var_string",
     "expand_cmake_vars",
     "gentbl",
-    "llvm_target_cmake_vars",
+    "llvm_all_cmake_vars",
+    "llvm_copts",
+    "llvm_defines",
+    "llvm_linkopts",
+    "llvm_support_platform_specific_srcs_glob",
 )
 load(
     "@org_tensorflow//third_party:common.bzl",
@@ -39,147 +43,25 @@ llvm_target_asm_printers = llvm_targets
 
 llvm_target_disassemblers = llvm_targets
 
-# TODO(phawkins): the set of CMake variables was hardcoded for expediency.
-# However, we should really detect many of these via configure-time tests.
-
-# The set of CMake variables common to all targets.
-cmake_vars = {
-    # Headers
-    "HAVE_DIRENT_H": 1,
-    "HAVE_DLFCN_H": 1,
-    "HAVE_ERRNO_H": 1,
-    "HAVE_EXECINFO_H": 1,
-    "HAVE_FCNTL_H": 1,
-    "HAVE_INTTYPES_H": 1,
-    "HAVE_PTHREAD_H": 1,
-    "HAVE_SIGNAL_H": 1,
-    "HAVE_STDINT_H": 1,
-    "HAVE_SYS_IOCTL_H": 1,
-    "HAVE_SYS_MMAN_H": 1,
-    "HAVE_SYS_PARAM_H": 1,
-    "HAVE_SYS_RESOURCE_H": 1,
-    "HAVE_SYS_STAT_H": 1,
-    "HAVE_SYS_TIME_H": 1,
-    "HAVE_SYS_TYPES_H": 1,
-    "HAVE_TERMIOS_H": 1,
-    "HAVE_UNISTD_H": 1,
-    "HAVE_ZLIB_H": 1,
-
-    # Features
-    "HAVE_BACKTRACE": 1,
-    "BACKTRACE_HEADER": "execinfo.h",
-    "HAVE_DLOPEN": 1,
-    "HAVE_FUTIMES": 1,
-    "HAVE_GETCWD": 1,
-    "HAVE_GETPAGESIZE": 1,
-    "HAVE_GETRLIMIT": 1,
-    "HAVE_GETRUSAGE": 1,
-    "HAVE_GETTIMEOFDAY": 1,
-    "HAVE_INT64_T": 1,
-    "HAVE_ISATTY": 1,
-    "HAVE_LIBEDIT": 1,
-    "HAVE_LIBPTHREAD": 1,
-    "HAVE_LIBZ": 1,
-    "HAVE_MKDTEMP": 1,
-    "HAVE_MKSTEMP": 1,
-    "HAVE_MKTEMP": 1,
-    "HAVE_PREAD": 1,
-    "HAVE_PTHREAD_GETSPECIFIC": 1,
-    "HAVE_PTHREAD_MUTEX_LOCK": 1,
-    "HAVE_PTHREAD_RWLOCK_INIT": 1,
-    "HAVE_REALPATH": 1,
-    "HAVE_SBRK": 1,
-    "HAVE_SETENV": 1,
-    "HAVE_SETRLIMIT": 1,
-    "HAVE_SIGALTSTACK": 1,
-    "HAVE_STRERROR": 1,
-    "HAVE_STRERROR_R": 1,
-    "HAVE_STRTOLL": 1,
-    "HAVE_SYSCONF": 1,
-    "HAVE_UINT64_T": 1,
-    "HAVE__UNWIND_BACKTRACE": 1,
-
-    # LLVM features
-    "ENABLE_BACKTRACES": 1,
-    "LLVM_BINDIR": "/dev/null",
-    "LLVM_DISABLE_ABI_BREAKING_CHECKS_ENFORCING": 0,
-    "LLVM_ENABLE_ABI_BREAKING_CHECKS": 0,
-    "LLVM_ENABLE_THREADS": 1,
-    "LLVM_ENABLE_ZLIB": 1,
-    "LLVM_HAS_ATOMICS": 1,
-    "LLVM_INCLUDEDIR": "/dev/null",
-    "LLVM_INFODIR": "/dev/null",
-    "LLVM_MANDIR": "/dev/null",
-    "LLVM_NATIVE_TARGET": 1,
-    "LLVM_NATIVE_TARGETINFO": 1,
-    "LLVM_NATIVE_TARGETMC": 1,
-    "LLVM_NATIVE_ASMPRINTER": 1,
-    "LLVM_NATIVE_ASMPARSER": 1,
-    "LLVM_NATIVE_DISASSEMBLER": 1,
-    "LLVM_ON_UNIX": 1,
-    "LLVM_PREFIX": "/dev/null",
-    "LLVM_VERSION_MAJOR": 0,
-    "LLVM_VERSION_MINOR": 0,
-    "LLVM_VERSION_PATCH": 0,
-    "LTDL_SHLIB_EXT": ".so",
-    "PACKAGE_NAME": "llvm",
-    "PACKAGE_STRING": "llvm tensorflow-trunk",
-    "PACKAGE_VERSION": "tensorflow-trunk",
-    "RETSIGTYPE": "void",
-}
-
-# CMake variables specific to the Linux platform
-linux_cmake_vars = {
-    "HAVE_MALLOC_H": 1,
-    "HAVE_LINK_H": 1,
-    "HAVE_MALLINFO": 1,
-    "HAVE_FUTIMENS": 1,
-}
-
-# CMake variables specific to the Darwin (Mac OS X) platform.
-darwin_cmake_vars = {
-    "HAVE_MALLOC_MALLOC_H": 1,
-}
-
-# Select a set of CMake variables based on the platform.
-# TODO(phawkins): use a better method to select the right host triple, rather
-# than hardcoding x86_64.
-all_cmake_vars = select({
-    "@org_tensorflow//tensorflow:darwin": cmake_var_string(
-        cmake_vars + llvm_target_cmake_vars("X86", "x86_64-apple-darwin") +
-        darwin_cmake_vars,
-    ),
-    "@org_tensorflow//tensorflow:linux_ppc64le": cmake_var_string(
-        cmake_vars +
-        llvm_target_cmake_vars("PowerPC", "powerpc64le-unknown-linux_gnu") +
-        linux_cmake_vars,
-    ),
-    "//conditions:default": cmake_var_string(
-        cmake_vars +
-        llvm_target_cmake_vars("X86", "x86_64-unknown-linux_gnu") +
-        linux_cmake_vars,
-    ),
-})
-
 # Performs CMake variable substitutions on configuration header files.
 expand_cmake_vars(
     name = "config_gen",
     src = "include/llvm/Config/config.h.cmake",
-    cmake_vars = all_cmake_vars,
+    cmake_vars = llvm_all_cmake_vars,
     dst = "include/llvm/Config/config.h",
 )
 
 expand_cmake_vars(
     name = "llvm_config_gen",
     src = "include/llvm/Config/llvm-config.h.cmake",
-    cmake_vars = all_cmake_vars,
+    cmake_vars = llvm_all_cmake_vars,
     dst = "include/llvm/Config/llvm-config.h",
 )
 
 expand_cmake_vars(
     name = "abi_breaking_gen",
     src = "include/llvm/Config/abi-breaking.h.cmake",
-    cmake_vars = all_cmake_vars,
+    cmake_vars = llvm_all_cmake_vars,
     dst = "include/llvm/Config/abi-breaking.h",
 )
 
@@ -240,14 +122,7 @@ cc_library(
         "include/llvm/Config/config.h",
         "include/llvm/Config/llvm-config.h",
     ],
-    defines = [
-        "LLVM_ENABLE_STATS",
-        "__STDC_LIMIT_MACROS",
-        "__STDC_CONSTANT_MACROS",
-        "__STDC_FORMAT_MACROS",
-        "_DEBUG",
-        "LLVM_BUILD_GLOBAL_ISEL",
-    ],
+    defines = llvm_defines,
     includes = ["include"],
 )
 
@@ -262,17 +137,6 @@ genrule(
 )
 
 # Rules that apply the LLVM tblgen tool.
-gentbl(
-    name = "intrinsics_gen",
-    tbl_outs = [("-gen-intrinsic", "include/llvm/IR/Intrinsics.inc")],
-    tblgen = ":llvm-tblgen",
-    td_file = "include/llvm/IR/Intrinsics.td",
-    td_srcs = glob([
-        "include/llvm/CodeGen/*.td",
-        "include/llvm/IR/Intrinsics*.td",
-    ]),
-)
-
 gentbl(
     name = "attributes_gen",
     tbl_outs = [("-gen-attrs", "include/llvm/IR/Attributes.inc")],
@@ -306,6 +170,28 @@ gentbl(
     ]) + ["include/llvm/TableGen/SearchableTable.td"],
 )
 
+gentbl(
+    name = "intrinsic_enums_gen",
+    tbl_outs = [("-gen-intrinsic-enums", "include/llvm/IR/IntrinsicEnums.inc")],
+    tblgen = ":llvm-tblgen",
+    td_file = "include/llvm/IR/Intrinsics.td",
+    td_srcs = glob([
+        "include/llvm/CodeGen/*.td",
+        "include/llvm/IR/Intrinsics*.td",
+    ]),
+)
+
+gentbl(
+    name = "intrinsics_impl_gen",
+    tbl_outs = [("-gen-intrinsic-impl", "include/llvm/IR/IntrinsicImpl.inc")],
+    tblgen = ":llvm-tblgen",
+    td_file = "include/llvm/IR/Intrinsics.td",
+    td_srcs = glob([
+        "include/llvm/CodeGen/*.td",
+        "include/llvm/IR/Intrinsics*.td",
+    ]),
+)
+
 # Binary targets used by Tensorflow.
 cc_binary(
     name = "llvm-tblgen",
@@ -313,11 +199,8 @@ cc_binary(
         "utils/TableGen/*.cpp",
         "utils/TableGen/*.h",
     ]),
-    linkopts = [
-        "-lm",
-        "-ldl",
-        "-lpthread",
-    ],
+    copts = llvm_copts,
+    linkopts = llvm_linkopts,
     stamp = 0,
     deps = [
         ":config",
@@ -333,11 +216,8 @@ cc_binary(
         "utils/FileCheck/*.cpp",
         "utils/FileCheck/*.h",
     ]),
-    linkopts = [
-        "-ldl",
-        "-lm",
-        "-lpthread",
-    ],
+    copts = llvm_copts,
+    linkopts = llvm_linkopts,
     stamp = 0,
     deps = [":support"],
 )
@@ -508,7 +388,7 @@ cc_library(
         "include/llvm/Target/AArch64/AsmParser/*.inc",
         "lib/Target/AArch64/AsmParser/*.h",
     ]),
-    copts = ["-Iexternal/llvm/lib/Target/AArch64"],
+    copts = llvm_copts + ["-Iexternal/llvm/lib/Target/AArch64"],
     deps = [
         ":aarch64_desc",
         ":aarch64_info",
@@ -533,7 +413,7 @@ cc_library(
         "include/llvm/Target/AArch64/InstPrinter/*.inc",
         "lib/Target/AArch64/InstPrinter/*.h",
     ]),
-    copts = ["-Iexternal/llvm/lib/Target/AArch64"],
+    copts = llvm_copts + ["-Iexternal/llvm/lib/Target/AArch64"],
     deps = [
         ":aarch64_target_gen",
         ":aarch64_utils",
@@ -556,7 +436,7 @@ cc_library(
         "include/llvm/Target/AArch64/*.inc",
         "lib/Target/AArch64/*.h",
     ]),
-    copts = ["-Iexternal/llvm/lib/Target/AArch64"],
+    copts = llvm_copts + ["-Iexternal/llvm/lib/Target/AArch64"],
     deps = [
         ":aarch64_asm_printer",
         ":aarch64_desc",
@@ -589,14 +469,15 @@ cc_library(
         "include/llvm/Target/AArch64/MCTargetDesc/*.inc",
         "lib/Target/AArch64/MCTargetDesc/*.h",
     ]),
-    copts = ["-Iexternal/llvm/lib/Target/AArch64"],
+    copts = llvm_copts + ["-Iexternal/llvm/lib/Target/AArch64"],
     deps = [
         ":aarch64_asm_printer",
         ":aarch64_info",
         ":aarch64_target_gen",
         ":attributes_gen",
         ":config",
-        ":intrinsics_gen",
+        ":intrinsic_enums_gen",
+        ":intrinsics_impl_gen",
         ":mc",
         ":support",
     ],
@@ -615,7 +496,7 @@ cc_library(
         "include/llvm/Target/AArch64/Disassembler/*.inc",
         "lib/Target/AArch64/Disassembler/*.h",
     ]),
-    copts = ["-Iexternal/llvm/lib/Target/AArch64"],
+    copts = llvm_copts + ["-Iexternal/llvm/lib/Target/AArch64"],
     deps = [
         ":aarch64_desc",
         ":aarch64_info",
@@ -643,7 +524,7 @@ cc_library(
         "lib/Target/AArch64/AArch64*.h",
         "lib/Target/AArch64/TargetInfo/*.h",
     ]),
-    copts = ["-Iexternal/llvm/lib/Target/AArch64"],
+    copts = llvm_copts + ["-Iexternal/llvm/lib/Target/AArch64"],
     deps = [
         ":code_gen",
         ":config",
@@ -666,7 +547,7 @@ cc_library(
         "include/llvm/Target/AArch64/Utils/*.inc",
         "lib/Target/AArch64/Utils/*.h",
     ]),
-    copts = ["-Iexternal/llvm/lib/Target/AArch64"],
+    copts = llvm_copts + ["-Iexternal/llvm/lib/Target/AArch64"],
     deps = [
         ":aarch64_target_gen",
         ":config",
@@ -688,6 +569,7 @@ cc_library(
         "include/llvm/Transforms/AggressiveInstCombine/*.def",
         "include/llvm/Transforms/AggressiveInstCombine/*.inc",
     ]),
+    copts = llvm_copts,
     deps = [
         ":analysis",
         ":config",
@@ -712,6 +594,7 @@ cc_library(
         "include/llvm/Analysis/*.def",
         "include/llvm/Analysis/*.inc",
     ]),
+    copts = llvm_copts,
     deps = [
         ":binary_format",
         ":config",
@@ -735,7 +618,7 @@ cc_library(
         "include/llvm/Target/AMDGPU/MCTargetDesc/*.inc",
         "lib/Target/AMDGPU/MCTargetDesc/*.h",
     ]),
-    copts = ["-Iexternal/llvm/lib/Target/AMDGPU"],
+    copts = llvm_copts + ["-Iexternal/llvm/lib/Target/AMDGPU"],
     deps = [
         ":amdgpu_asm_printer",
         ":amdgpu_info",
@@ -760,7 +643,7 @@ cc_library(
         "include/llvm/Target/AMDGPU/Disassembler/*.inc",
         "lib/Target/AMDGPU/Disassembler/*.h",
     ]),
-    copts = ["-Iexternal/llvm/lib/Target/AMDGPU"],
+    copts = llvm_copts + ["-Iexternal/llvm/lib/Target/AMDGPU"],
     deps = [
         ":amdgpu_desc",
         ":amdgpu_info",
@@ -785,7 +668,7 @@ cc_library(
         "include/llvm/Target/AMDGPU/TargetInfo/*.inc",
         "lib/Target/AMDGPU/TargetInfo/*.h",
     ]),
-    copts = ["-Iexternal/llvm/lib/Target/AMDGPU"],
+    copts = llvm_copts + ["-Iexternal/llvm/lib/Target/AMDGPU"],
     deps = [
         ":amdgpu_target_gen",
         ":config",
@@ -807,7 +690,7 @@ cc_library(
         "include/llvm/Target/AMDGPU/Utils/*.inc",
         "lib/Target/AMDGPU/Utils/*.h",
     ]),
-    copts = ["-Iexternal/llvm/lib/Target/AMDGPU"],
+    copts = llvm_copts + ["-Iexternal/llvm/lib/Target/AMDGPU"],
     deps = [
         ":amdgpu_target_gen",
         ":config",
@@ -830,7 +713,7 @@ cc_library(
         "include/llvm/Target/AMDGPU/AsmParser/*.inc",
         "lib/Target/AMDGPU/AsmParser/*.h",
     ]),
-    copts = ["-Iexternal/llvm/lib/Target/AMDGPU"],
+    copts = llvm_copts + ["-Iexternal/llvm/lib/Target/AMDGPU"],
     deps = [
         ":amdgpu_desc",
         ":amdgpu_info",
@@ -855,7 +738,7 @@ cc_library(
         "include/llvm/Target/AMDGPU/InstPrinter/*.inc",
         "lib/Target/AMDGPU/InstPrinter/*.h",
     ]),
-    copts = ["-Iexternal/llvm/lib/Target/AMDGPU"],
+    copts = llvm_copts + ["-Iexternal/llvm/lib/Target/AMDGPU"],
     deps = [
         ":amdgpu_utils",
         ":config",
@@ -877,7 +760,7 @@ cc_library(
         "include/llvm/Target/AMDGPU/*.inc",
         "lib/Target/AMDGPU/*.h",
     ]),
-    copts = ["-Iexternal/llvm/lib/Target/AMDGPU"],
+    copts = llvm_copts + ["-Iexternal/llvm/lib/Target/AMDGPU"],
     deps = [
         ":amdgpu_asm_printer",
         ":amdgpu_desc",
@@ -913,7 +796,7 @@ cc_library(
         "include/llvm/Target/ARM/AsmParser/*.inc",
         "lib/Target/ARM/AsmParser/*.h",
     ]),
-    copts = ["-Iexternal/llvm/lib/Target/ARM"],
+    copts = llvm_copts + ["-Iexternal/llvm/lib/Target/ARM"],
     deps = [
         ":arm_desc",
         ":arm_info",
@@ -939,7 +822,7 @@ cc_library(
         "lib/Target/ARM/*.h",
         "lib/Target/ARM/InstPrinter/*.h",
     ]),
-    copts = ["-Iexternal/llvm/lib/Target/ARM"],
+    copts = llvm_copts + ["-Iexternal/llvm/lib/Target/ARM"],
     deps = [
         ":arm_info",
         ":arm_target_gen",
@@ -963,7 +846,7 @@ cc_library(
         "include/llvm/Target/ARM/*.inc",
         "lib/Target/ARM/*.h",
     ]),
-    copts = ["-Iexternal/llvm/lib/Target/ARM"],
+    copts = llvm_copts + ["-Iexternal/llvm/lib/Target/ARM"],
     deps = [
         ":analysis",
         ":arm_asm_printer",
@@ -980,6 +863,7 @@ cc_library(
         ":selection_dag",
         ":support",
         ":target",
+        ":transform_utils",
     ],
 )
 
@@ -998,14 +882,15 @@ cc_library(
         "include/llvm/Target/ARM/MCTargetDesc/*.inc",
         "lib/Target/ARM/MCTargetDesc/*.h",
     ]),
-    copts = ["-Iexternal/llvm/lib/Target/ARM"],
+    copts = llvm_copts + ["-Iexternal/llvm/lib/Target/ARM"],
     deps = [
         ":arm_asm_printer",
         ":arm_info",
         ":arm_target_gen",
         ":attributes_gen",
         ":config",
-        ":intrinsics_gen",
+        ":intrinsic_enums_gen",
+        ":intrinsics_impl_gen",
         ":mc",
         ":mc_disassembler",
         ":support",
@@ -1025,7 +910,7 @@ cc_library(
         "include/llvm/Target/ARM/Disassembler/*.inc",
         "lib/Target/ARM/Disassembler/*.h",
     ]),
-    copts = ["-Iexternal/llvm/lib/Target/ARM"],
+    copts = llvm_copts + ["-Iexternal/llvm/lib/Target/ARM"],
     deps = [
         ":arm_desc",
         ":arm_info",
@@ -1050,7 +935,7 @@ cc_library(
         "include/llvm/Target/ARM/TargetInfo/*.inc",
         "lib/Target/ARM/TargetInfo/*.h",
     ]),
-    copts = ["-Iexternal/llvm/lib/Target/ARM"],
+    copts = llvm_copts + ["-Iexternal/llvm/lib/Target/ARM"],
     deps = [
         ":arm_target_gen",
         ":config",
@@ -1073,7 +958,7 @@ cc_library(
         "include/llvm/Target/ARM/Utils/*.inc",
         "lib/Target/ARM/Utils/*.h",
     ]),
-    copts = ["-Iexternal/llvm/lib/Target/ARM"],
+    copts = llvm_copts + ["-Iexternal/llvm/lib/Target/ARM"],
     deps = [
         ":arm_target_gen",
         ":config",
@@ -1095,6 +980,7 @@ cc_library(
         "include/llvm/AsmParser/*.def",
         "include/llvm/AsmParser/*.inc",
     ]),
+    copts = llvm_copts,
     deps = [
         ":binary_format",
         ":config",
@@ -1117,6 +1003,7 @@ cc_library(
         "include/llvm/CodeGen/AsmPrinter/*.inc",
         "lib/CodeGen/AsmPrinter/*.def",
     ]),
+    copts = llvm_copts,
     deps = [
         ":analysis",
         ":binary_format",
@@ -1147,6 +1034,7 @@ cc_library(
         "include/llvm/BinaryFormat/ELFRelocs/*.def",
         "include/llvm/BinaryFormat/WasmRelocs/*.def",
     ]),
+    copts = llvm_copts,
     deps = [
         ":config",
         ":support",
@@ -1167,6 +1055,7 @@ cc_library(
         "include/llvm/Bitcode/Reader/*.inc",
         "include/llvm/Bitcode/BitstreamReader.h",
     ]),
+    copts = llvm_copts,
     deps = [
         ":config",
         ":core",
@@ -1190,6 +1079,7 @@ cc_library(
         "include/llvm/Bitcode/BitcodeWriterPass.h",
         "include/llvm/Bitcode/BitstreamWriter.h",
     ]),
+    copts = llvm_copts,
     deps = [
         ":analysis",
         ":config",
@@ -1214,6 +1104,7 @@ cc_library(
         "include/llvm/CodeGen/*.inc",
         "include/llvm/CodeGen/**/*.h",
     ]),
+    copts = llvm_copts,
     deps = [
         ":analysis",
         ":bit_reader",
@@ -1251,12 +1142,14 @@ cc_library(
         "include/llvm/*.h",
         "include/llvm/Analysis/*.def",
     ]),
+    copts = llvm_copts,
     deps = [
         ":attributes_compat_gen",
         ":attributes_gen",
         ":binary_format",
         ":config",
-        ":intrinsics_gen",
+        ":intrinsic_enums_gen",
+        ":intrinsics_impl_gen",
         ":support",
     ],
 )
@@ -1274,6 +1167,7 @@ cc_library(
         "include/llvm/DebugInfo/CodeView/*.def",
         "include/llvm/DebugInfo/CodeView/*.inc",
     ]),
+    copts = llvm_copts,
     deps = [
         ":binary_format",
         ":config",
@@ -1295,6 +1189,7 @@ cc_library(
         "include/llvm/DebugInfo/MSF/*.def",
         "include/llvm/DebugInfo/MSF/*.inc",
     ]),
+    copts = llvm_copts,
     deps = [
         ":config",
         ":support",
@@ -1314,6 +1209,7 @@ cc_library(
         "include/llvm/Demangle/*.def",
         "include/llvm/Demangle/*.inc",
     ]),
+    copts = llvm_copts,
     deps = [":config"],
 )
 
@@ -1330,6 +1226,7 @@ cc_library(
         "include/llvm/ExecutionEngine/*.def",
         "include/llvm/ExecutionEngine/*.inc",
     ]),
+    copts = llvm_copts,
     deps = [
         ":config",
         ":core",
@@ -1354,6 +1251,7 @@ cc_library(
         "include/llvm/CodeGen/GlobalISel/*.def",
         "include/llvm/CodeGen/GlobalISel/*.inc",
     ]),
+    copts = llvm_copts,
     deps = [
         ":analysis",
         ":code_gen",
@@ -1383,6 +1281,7 @@ cc_library(
         "include/llvm/Transforms/InstrProfiling.h",
         "include/llvm/Transforms/PGOInstrumentation.h",
     ]),
+    copts = llvm_copts,
     deps = [
         ":analysis",
         ":config",
@@ -1407,6 +1306,7 @@ cc_library(
         "include/llvm/Transforms/InstCombine/*.def",
         "include/llvm/Transforms/InstCombine/*.inc",
     ]),
+    copts = llvm_copts,
     deps = [
         ":analysis",
         ":config",
@@ -1433,6 +1333,7 @@ cc_library(
         "include/llvm/Transforms/IPO/*.def",
         "include/llvm/Transforms/IPO/*.inc",
     ]),
+    copts = llvm_copts,
     deps = [
         ":aggressive_inst_combine",
         ":analysis",
@@ -1466,6 +1367,7 @@ cc_library(
         "include/llvm/IRReader/*.def",
         "include/llvm/IRReader/*.inc",
     ]),
+    copts = llvm_copts,
     deps = [
         ":asm_parser",
         ":bit_reader",
@@ -1488,6 +1390,7 @@ cc_library(
         "include/llvm/Linker/*.def",
         "include/llvm/Linker/*.inc",
     ]),
+    copts = llvm_copts,
     deps = [
         ":config",
         ":core",
@@ -1509,6 +1412,7 @@ cc_library(
         "include/llvm/MC/*.def",
         "include/llvm/MC/*.inc",
     ]),
+    copts = llvm_copts,
     deps = [
         ":binary_format",
         ":config",
@@ -1530,6 +1434,7 @@ cc_library(
         "include/llvm/MC/MCDisassembler/*.def",
         "include/llvm/MC/MCDisassembler/*.inc",
     ]),
+    copts = llvm_copts,
     deps = [
         ":config",
         ":mc",
@@ -1550,6 +1455,7 @@ cc_library(
         "include/llvm/MC/MCParser/*.def",
         "include/llvm/MC/MCParser/*.inc",
     ]),
+    copts = llvm_copts,
     deps = [
         ":config",
         ":mc",
@@ -1570,7 +1476,7 @@ cc_library(
         "include/llvm/Target/NVPTX/InstPrinter/*.inc",
         "lib/Target/NVPTX/InstPrinter/*.h",
     ]),
-    copts = ["-Iexternal/llvm/lib/Target/NVPTX"],
+    copts = llvm_copts + ["-Iexternal/llvm/lib/Target/NVPTX"],
     deps = [
         "nvptx_target_gen",
         ":attributes_gen",
@@ -1594,7 +1500,7 @@ cc_library(
         "include/llvm/Target/NVPTX/*.inc",
         "lib/Target/NVPTX/*.h",
     ]),
-    copts = ["-Iexternal/llvm/lib/Target/NVPTX"],
+    copts = llvm_copts + ["-Iexternal/llvm/lib/Target/NVPTX"],
     deps = [
         ":analysis",
         ":asm_printer",
@@ -1628,7 +1534,7 @@ cc_library(
         "include/llvm/Target/NVPTX/MCTargetDesc/*.inc",
         "lib/Target/NVPTX/MCTargetDesc/*.h",
     ]),
-    copts = ["-Iexternal/llvm/lib/Target/NVPTX"],
+    copts = llvm_copts + ["-Iexternal/llvm/lib/Target/NVPTX"],
     deps = [
         "nvptx_target_gen",
         ":config",
@@ -1654,7 +1560,7 @@ cc_library(
         "lib/Target/NVPTX/NVPTX.h",
         "lib/Target/NVPTX/TargetInfo/*.h",
     ]),
-    copts = ["-Iexternal/llvm/lib/Target/NVPTX"],
+    copts = llvm_copts + ["-Iexternal/llvm/lib/Target/NVPTX"],
     deps = [
         "nvptx_target_gen",
         ":attributes_gen",
@@ -1678,6 +1584,7 @@ cc_library(
         "include/llvm/Object/*.def",
         "include/llvm/Object/*.inc",
     ]),
+    copts = llvm_copts,
     deps = [
         ":binary_format",
         ":bit_reader",
@@ -1703,6 +1610,7 @@ cc_library(
         "include/llvm/Transforms/ObjCARC/*.def",
         "include/llvm/Transforms/ObjCARC/*.inc",
     ]),
+    copts = llvm_copts,
     deps = [
         ":analysis",
         ":config",
@@ -1725,13 +1633,16 @@ cc_library(
         "include/llvm/ExecutionEngine/Orc/*.def",
         "include/llvm/ExecutionEngine/Orc/*.inc",
     ]),
+    copts = llvm_copts,
     deps = [
         ":config",
         ":core",
         ":execution_engine",
+        ":mc",
         ":object",
         ":runtime_dyld",
         ":support",
+        ":target",
         ":transform_utils",
     ],
 )
@@ -1749,7 +1660,7 @@ cc_library(
         "include/llvm/Target/PowerPC/AsmParser/*.inc",
         "lib/Target/PowerPC/AsmParser/*.h",
     ]),
-    copts = ["-Iexternal/llvm/lib/Target/PowerPC"],
+    copts = llvm_copts + ["-Iexternal/llvm/lib/Target/PowerPC"],
     deps = [
         ":config",
         ":mc",
@@ -1773,11 +1684,12 @@ cc_library(
         "include/llvm/Target/PowerPC/InstPrinter/*.inc",
         "lib/Target/PowerPC/InstPrinter/*.h",
     ]),
-    copts = ["-Iexternal/llvm/lib/Target/PowerPC"],
+    copts = llvm_copts + ["-Iexternal/llvm/lib/Target/PowerPC"],
     deps = [
         ":attributes_gen",
         ":config",
-        ":intrinsics_gen",
+        ":intrinsic_enums_gen",
+        ":intrinsics_impl_gen",
         ":mc",
         ":powerpc_info",
         ":powerpc_target_gen",
@@ -1798,7 +1710,7 @@ cc_library(
         "include/llvm/Target/PowerPC/*.inc",
         "lib/Target/PowerPC/*.h",
     ]),
-    copts = ["-Iexternal/llvm/lib/Target/PowerPC"],
+    copts = llvm_copts + ["-Iexternal/llvm/lib/Target/PowerPC"],
     deps = [
         ":analysis",
         ":asm_printer",
@@ -1830,11 +1742,12 @@ cc_library(
         "include/llvm/Target/PowerPC/MCTargetDesc/*.inc",
         "lib/Target/PowerPC/MCTargetDesc/*.h",
     ]),
-    copts = ["-Iexternal/llvm/lib/Target/PowerPC"],
+    copts = llvm_copts + ["-Iexternal/llvm/lib/Target/PowerPC"],
     deps = [
         ":attributes_gen",
         ":config",
-        ":intrinsics_gen",
+        ":intrinsic_enums_gen",
+        ":intrinsics_impl_gen",
         ":mc",
         ":powerpc_asm_printer",
         ":powerpc_info",
@@ -1856,7 +1769,7 @@ cc_library(
         "include/llvm/Target/PowerPC/Disassembler/*.inc",
         "lib/Target/PowerPC/Disassembler/*.h",
     ]),
-    copts = ["-Iexternal/llvm/lib/Target/PowerPC"],
+    copts = llvm_copts + ["-Iexternal/llvm/lib/Target/PowerPC"],
     deps = [
         ":config",
         ":mc_disassembler",
@@ -1880,12 +1793,11 @@ cc_library(
         "lib/Target/PowerPC/PPC*.h",
         "lib/Target/PowerPC/TargetInfo/*.h",
     ]),
-    copts = ["-Iexternal/llvm/lib/Target/PowerPC"],
+    copts = llvm_copts + ["-Iexternal/llvm/lib/Target/PowerPC"],
     deps = [
         ":attributes_gen",
         ":config",
         ":core",
-        ":intrinsics_gen",
         ":powerpc_target_gen",
         ":support",
         ":target",
@@ -1905,6 +1817,7 @@ cc_library(
         "include/llvm/ProfileData/*.def",
         "include/llvm/ProfileData/*.inc",
     ]),
+    copts = llvm_copts,
     deps = [
         ":config",
         ":core",
@@ -1933,6 +1846,7 @@ cc_library(
         "include/llvm/ExecutionEngine/RTDyldMemoryManager.h",
         "include/llvm/ExecutionEngine/RuntimeDyld*.h",
     ]),
+    copts = llvm_copts,
     deps = [
         ":config",
         ":mc",
@@ -1960,6 +1874,7 @@ cc_library(
         "include/llvm/Transforms/IPO.h",
         "include/llvm/Transforms/IPO/SCCP.h",
     ]),
+    copts = llvm_copts,
     deps = [
         ":aggressive_inst_combine",
         ":analysis",
@@ -1985,6 +1900,7 @@ cc_library(
         "include/llvm/CodeGen/SelectionDAG/*.def",
         "include/llvm/CodeGen/SelectionDAG/*.inc",
     ]),
+    copts = llvm_copts,
     deps = [
         ":analysis",
         ":code_gen",
@@ -2003,14 +1919,12 @@ cc_library(
         "lib/Support/*.c",
         "lib/Support/*.cpp",
         "lib/Support/*.inc",
-        "lib/Support/Unix/*.inc",
-        "lib/Support/Unix/*.h",
         "include/llvm-c/*.h",
         "include/llvm/CodeGen/MachineValueType.h",
         "include/llvm/BinaryFormat/COFF.h",
         "include/llvm/BinaryFormat/MachO.h",
         "lib/Support/*.h",
-    ]),
+    ] + llvm_support_platform_specific_srcs_glob),
     hdrs = glob([
         "include/llvm/Support/*.h",
         "include/llvm/Support/*.def",
@@ -2022,6 +1936,7 @@ cc_library(
         "include/llvm/BinaryFormat/MachO.def",
         "include/llvm/Support/VCSRevision.h",
     ],
+    copts = llvm_copts,
     deps = [
         ":config",
         ":demangle",
@@ -2044,6 +1959,7 @@ cc_library(
         "include/llvm/TableGen/*.inc",
         "include/llvm/Target/*.def",
     ]),
+    copts = llvm_copts,
     deps = [
         ":config",
         ":mc",
@@ -2069,6 +1985,7 @@ cc_library(
         "include/llvm/CodeGen/*.def",
         "include/llvm/CodeGen/*.inc",
     ]),
+    copts = llvm_copts,
     deps = [
         ":analysis",
         ":config",
@@ -2093,6 +2010,7 @@ cc_library(
         "include/llvm/Transforms/Utils/*.def",
         "include/llvm/Transforms/Utils/*.inc",
     ]),
+    copts = llvm_copts,
     deps = [
         ":analysis",
         ":config",
@@ -2116,6 +2034,7 @@ cc_library(
         "include/llvm/Transforms/Vectorize/*.inc",
         "include/llvm/Transforms/Vectorize.h",
     ]),
+    copts = llvm_copts,
     deps = [
         ":analysis",
         ":config",
@@ -2139,7 +2058,7 @@ cc_library(
         "include/llvm/Target/X86/AsmParser/*.inc",
         "lib/Target/X86/AsmParser/*.h",
     ]),
-    copts = ["-Iexternal/llvm/lib/Target/X86"],
+    copts = llvm_copts + ["-Iexternal/llvm/lib/Target/X86"],
     deps = [
         ":config",
         ":mc",
@@ -2164,7 +2083,7 @@ cc_library(
         "include/llvm/Target/X86/InstPrinter/*.inc",
         "lib/Target/X86/InstPrinter/*.h",
     ]),
-    copts = ["-Iexternal/llvm/lib/Target/X86"],
+    copts = llvm_copts + ["-Iexternal/llvm/lib/Target/X86"],
     deps = [
         ":config",
         ":mc",
@@ -2188,7 +2107,7 @@ cc_library(
         "include/llvm/Target/X86/*.inc",
         "lib/Target/X86/*.h",
     ]),
-    copts = ["-Iexternal/llvm/lib/Target/X86"],
+    copts = llvm_copts + ["-Iexternal/llvm/lib/Target/X86"],
     deps = [
         ":analysis",
         ":asm_printer",
@@ -2221,7 +2140,7 @@ cc_library(
         "include/llvm/Target/X86/MCTargetDesc/*.inc",
         "lib/Target/X86/MCTargetDesc/*.h",
     ]),
-    copts = ["-Iexternal/llvm/lib/Target/X86"],
+    copts = llvm_copts + ["-Iexternal/llvm/lib/Target/X86"],
     deps = [
         ":config",
         ":mc",
@@ -2246,7 +2165,7 @@ cc_library(
         "include/llvm/Target/X86/Disassembler/*.inc",
         "lib/Target/X86/Disassembler/*.h",
     ]),
-    copts = ["-Iexternal/llvm/lib/Target/X86"],
+    copts = llvm_copts + ["-Iexternal/llvm/lib/Target/X86"],
     deps = [
         ":config",
         ":mc_disassembler",
@@ -2269,7 +2188,7 @@ cc_library(
         "include/llvm/Target/X86/TargetInfo/*.inc",
         "lib/Target/X86/TargetInfo/*.h",
     ]),
-    copts = ["-Iexternal/llvm/lib/Target/X86"],
+    copts = llvm_copts + ["-Iexternal/llvm/lib/Target/X86"],
     deps = [
         ":config",
         ":mc",
@@ -2291,7 +2210,7 @@ cc_library(
         "include/llvm/Target/X86/Utils/*.inc",
         "lib/Target/X86/Utils/*.h",
     ]),
-    copts = ["-Iexternal/llvm/lib/Target/X86"],
+    copts = llvm_copts + ["-Iexternal/llvm/lib/Target/X86"],
     deps = [
         ":code_gen",
         ":config",
diff --git a/third_party/llvm/llvm.bzl b/third_party/llvm/llvm.bzl
index 0efcf319bd99be79263a1b9cd23544523a4c8076..dfdacafceb56440ef7a67fc1352c833b910a7ce5 100644
--- a/third_party/llvm/llvm.bzl
+++ b/third_party/llvm/llvm.bzl
@@ -105,3 +105,143 @@ def expand_cmake_vars(name, src, dst, cmake_vars):
              "< $< > $@")
   )
 
+# TODO(phawkins): the set of CMake variables was hardcoded for expediency.
+# However, we should really detect many of these via configure-time tests.
+
+# The set of CMake variables common to all targets.
+cmake_vars = {
+    # Headers
+    "HAVE_DIRENT_H": 1,
+    "HAVE_DLFCN_H": 1,
+    "HAVE_ERRNO_H": 1,
+    "HAVE_EXECINFO_H": 1,
+    "HAVE_FCNTL_H": 1,
+    "HAVE_INTTYPES_H": 1,
+    "HAVE_PTHREAD_H": 1,
+    "HAVE_SIGNAL_H": 1,
+    "HAVE_STDINT_H": 1,
+    "HAVE_SYS_IOCTL_H": 1,
+    "HAVE_SYS_MMAN_H": 1,
+    "HAVE_SYS_PARAM_H": 1,
+    "HAVE_SYS_RESOURCE_H": 1,
+    "HAVE_SYS_STAT_H": 1,
+    "HAVE_SYS_TIME_H": 1,
+    "HAVE_SYS_TYPES_H": 1,
+    "HAVE_TERMIOS_H": 1,
+    "HAVE_UNISTD_H": 1,
+    "HAVE_ZLIB_H": 1,
+
+    # Features
+    "HAVE_BACKTRACE": 1,
+    "BACKTRACE_HEADER": "execinfo.h",
+    "HAVE_DLOPEN": 1,
+    "HAVE_FUTIMES": 1,
+    "HAVE_GETCWD": 1,
+    "HAVE_GETPAGESIZE": 1,
+    "HAVE_GETRLIMIT": 1,
+    "HAVE_GETRUSAGE": 1,
+    "HAVE_GETTIMEOFDAY": 1,
+    "HAVE_INT64_T": 1,
+    "HAVE_ISATTY": 1,
+    "HAVE_LIBEDIT": 1,
+    "HAVE_LIBPTHREAD": 1,
+    "HAVE_LIBZ": 1,
+    "HAVE_MKDTEMP": 1,
+    "HAVE_MKSTEMP": 1,
+    "HAVE_MKTEMP": 1,
+    "HAVE_PREAD": 1,
+    "HAVE_PTHREAD_GETSPECIFIC": 1,
+    "HAVE_PTHREAD_MUTEX_LOCK": 1,
+    "HAVE_PTHREAD_RWLOCK_INIT": 1,
+    "HAVE_REALPATH": 1,
+    "HAVE_SBRK": 1,
+    "HAVE_SETENV": 1,
+    "HAVE_SETRLIMIT": 1,
+    "HAVE_SIGALTSTACK": 1,
+    "HAVE_STRERROR": 1,
+    "HAVE_STRERROR_R": 1,
+    "HAVE_STRTOLL": 1,
+    "HAVE_SYSCONF": 1,
+    "HAVE_UINT64_T": 1,
+    "HAVE__UNWIND_BACKTRACE": 1,
+
+    # LLVM features
+    "ENABLE_BACKTRACES": 1,
+    "LLVM_BINDIR": "/dev/null",
+    "LLVM_DISABLE_ABI_BREAKING_CHECKS_ENFORCING": 0,
+    "LLVM_ENABLE_ABI_BREAKING_CHECKS": 0,
+    "LLVM_ENABLE_THREADS": 1,
+    "LLVM_ENABLE_ZLIB": 1,
+    "LLVM_HAS_ATOMICS": 1,
+    "LLVM_INCLUDEDIR": "/dev/null",
+    "LLVM_INFODIR": "/dev/null",
+    "LLVM_MANDIR": "/dev/null",
+    "LLVM_NATIVE_TARGET": 1,
+    "LLVM_NATIVE_TARGETINFO": 1,
+    "LLVM_NATIVE_TARGETMC": 1,
+    "LLVM_NATIVE_ASMPRINTER": 1,
+    "LLVM_NATIVE_ASMPARSER": 1,
+    "LLVM_NATIVE_DISASSEMBLER": 1,
+    "LLVM_ON_UNIX": 1,
+    "LLVM_PREFIX": "/dev/null",
+    "LLVM_VERSION_MAJOR": 0,
+    "LLVM_VERSION_MINOR": 0,
+    "LLVM_VERSION_PATCH": 0,
+    "LTDL_SHLIB_EXT": ".so",
+    "PACKAGE_NAME": "llvm",
+    "PACKAGE_STRING": "llvm tensorflow-trunk",
+    "PACKAGE_VERSION": "tensorflow-trunk",
+    "RETSIGTYPE": "void",
+}
+
+# CMake variables specific to the Linux platform
+linux_cmake_vars = {
+    "HAVE_MALLOC_H": 1,
+    "HAVE_LINK_H": 1,
+    "HAVE_MALLINFO": 1,
+    "HAVE_FUTIMENS": 1,
+}
+
+# CMake variables specific to the Darwin (Mac OS X) platform.
+darwin_cmake_vars = {
+    "HAVE_MALLOC_MALLOC_H": 1,
+}
+
+# Select a set of CMake variables based on the platform.
+# TODO(phawkins): use a better method to select the right host triple, rather
+# than hardcoding x86_64.
+llvm_all_cmake_vars = select({
+    "@org_tensorflow//tensorflow:darwin": cmake_var_string(
+        cmake_vars + llvm_target_cmake_vars("X86", "x86_64-apple-darwin") +
+        darwin_cmake_vars),
+    "@org_tensorflow//tensorflow:linux_ppc64le": cmake_var_string(
+        cmake_vars +
+        llvm_target_cmake_vars("PowerPC", "powerpc64le-unknown-linux_gnu") +
+        linux_cmake_vars,
+    ),
+    "//conditions:default": cmake_var_string(
+         cmake_vars +
+         llvm_target_cmake_vars("X86", "x86_64-unknown-linux_gnu") +
+         linux_cmake_vars),
+
+})
+
+llvm_linkopts = ["-ldl", "-lm", "-lpthread"]
+
+llvm_defines = [
+    "LLVM_ENABLE_STATS",
+    "__STDC_LIMIT_MACROS",
+    "__STDC_CONSTANT_MACROS",
+    "__STDC_FORMAT_MACROS",
+    "_DEBUG",
+    "LLVM_BUILD_GLOBAL_ISEL",
+]
+
+llvm_copts = []
+
+# Platform specific sources for libSupport.
+
+llvm_support_platform_specific_srcs_glob = [
+    "lib/Support/Unix/*.inc",
+    "lib/Support/Unix/*.h",
+]
diff --git a/third_party/mkl/LICENSE b/third_party/mkl/LICENSE
new file mode 100644
index 0000000000000000000000000000000000000000..9c8f3ea0871e0bfe81da0fa6e7c1d7d156dc380e
--- /dev/null
+++ b/third_party/mkl/LICENSE
@@ -0,0 +1,201 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "{}"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright {yyyy} {name of copyright owner}
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.
\ No newline at end of file
diff --git a/third_party/nanopb.BUILD b/third_party/nanopb.BUILD
new file mode 100644
index 0000000000000000000000000000000000000000..d21866911b862f0d4adf76c3a07e2732128a6102
--- /dev/null
+++ b/third_party/nanopb.BUILD
@@ -0,0 +1,23 @@
+# Description:
+#   Nanopb, a tiny ANSI C protobuf implementation for use on embedded devices.
+
+licenses(["notice"])  # zlib license
+
+exports_files(["LICENSE.txt"])
+
+cc_library(
+    name = "nanopb",
+    srcs = [
+        "pb_common.c",
+        "pb_decode.c",
+        "pb_encode.c",
+    ],
+    hdrs = [
+        "pb.h",
+        "pb_common.h",
+        "pb_decode.h",
+        "pb_encode.h",
+    ],
+    includes = ["."],
+    visibility = ["//visibility:public"],
+)
diff --git a/third_party/nasm.BUILD b/third_party/nasm.BUILD
index 341d58068be48b1edbbc28718cc104a467efa8d0..89330eac5404934ddded305dfc062017d8abb30c 100644
--- a/third_party/nasm.BUILD
+++ b/third_party/nasm.BUILD
@@ -8,45 +8,93 @@ exports_files(["LICENSE"])
 cc_binary(
     name = "nasm",
     srcs = [
-        "assemble.c",
-        "assemble.h",
-        "compiler.h",
-        "crc64.c",
-        "directiv.c",
-        "directiv.h",
-        "disp8.c",
-        "disp8.h",
-        "eval.c",
-        "eval.h",
-        "exprlib.c",
-        "float.c",
-        "float.h",
-        "hashtbl.c",
-        "hashtbl.h",
-        "iflag.c",
-        "iflag.h",
-        "iflaggen.h",
-        "ilog2.c",
-        "insns.h",
-        "insnsa.c",
-        "insnsb.c",
-        "insnsi.h",
-        "labels.c",
-        "labels.h",
-        "lib/strlcpy.c",
-        "listing.c",
-        "listing.h",
-        "macros.c",
-        "md5.h",
-        "md5c.c",
-        "nasm.c",
-        "nasm.h",
-        "nasmlib.c",
-        "nasmlib.h",
-        "opflags.h",
+        "asm/assemble.c",
+        "asm/assemble.h",
+        "asm/directbl.c",
+        "asm/directiv.c",
+        "asm/directiv.h",
+        "asm/error.c",
+        "asm/eval.c",
+        "asm/eval.h",
+        "asm/exprdump.c",
+        "asm/exprlib.c",
+        "asm/float.c",
+        "asm/float.h",
+        "asm/labels.c",
+        "asm/listing.c",
+        "asm/listing.h",
+        "asm/nasm.c",
+        "asm/parser.c",
+        "asm/parser.h",
+        "asm/pptok.c",
+        "asm/pptok.h",
+        "asm/pragma.c",
+        "asm/preproc.c",
+        "asm/preproc.h",
+        "asm/preproc-nop.c",
+        "asm/quote.c",
+        "asm/quote.h",
+        "asm/rdstrnum.c",
+        "asm/segalloc.c",
+        "asm/stdscan.c",
+        "asm/stdscan.h",
+        "asm/strfunc.c",
+        "asm/tokens.h",
+        "asm/tokhash.c",
+        "common/common.c",
+        "config/unknown.h",
+        "disasm/disasm.c",
+        "disasm/disasm.h",
+        "disasm/sync.c",
+        "disasm/sync.h",
+        "include/compiler.h",
+        "include/disp8.h",
+        "include/error.h",
+        "include/hashtbl.h",
+        "include/iflag.h",
+        "include/insns.h",
+        "include/labels.h",
+        "include/md5.h",
+        "include/nasm.h",
+        "include/nasmint.h",
+        "include/nasmlib.h",
+        "include/opflags.h",
+        "include/perfhash.h",
+        "include/raa.h",
+        "include/rbtree.h",
+        "include/rdoff.h",
+        "include/saa.h",
+        "include/strlist.h",
+        "include/tables.h",
+        "include/ver.h",
+        "macros/macros.c",
+        "nasmlib/badenum.c",
+        "nasmlib/bsi.c",
+        "nasmlib/crc64.c",
+        "nasmlib/file.c",
+        "nasmlib/file.h",
+        "nasmlib/filename.c",
+        "nasmlib/hashtbl.c",
+        "nasmlib/ilog2.c",
+        "nasmlib/malloc.c",
+        "nasmlib/md5c.c",
+        "nasmlib/mmap.c",
+        "nasmlib/path.c",
+        "nasmlib/perfhash.c",
+        "nasmlib/raa.c",
+        "nasmlib/rbtree.c",
+        "nasmlib/readnum.c",
+        "nasmlib/realpath.c",
+        "nasmlib/saa.c",
+        "nasmlib/srcfile.c",
+        "nasmlib/string.c",
+        "nasmlib/strlist.c",
+        "nasmlib/ver.c",
+        "nasmlib/zerobuf.c",
         "output/codeview.c",
         "output/dwarf.h",
         "output/elf.h",
+        "output/legacy.c",
         "output/nulldbg.c",
         "output/nullout.c",
         "output/outaout.c",
@@ -56,9 +104,6 @@ cc_binary(
         "output/outdbg.c",
         "output/outelf.c",
         "output/outelf.h",
-        "output/outelf32.c",
-        "output/outelf64.c",
-        "output/outelfx32.c",
         "output/outform.c",
         "output/outform.h",
         "output/outieee.c",
@@ -69,35 +114,31 @@ cc_binary(
         "output/outrdf2.c",
         "output/pecoff.h",
         "output/stabs.h",
-        "parser.c",
-        "parser.h",
-        "pptok.c",
-        "pptok.h",
-        "preproc.c",
-        "preproc.h",
-        "preproc-nop.c",
-        "quote.c",
-        "quote.h",
-        "raa.c",
-        "raa.h",
-        "rbtree.c",
-        "rbtree.h",
-        "rdoff/rdoff.h",
-        "realpath.c",
-        "regflags.c",
-        "regs.h",
-        "regvals.c",
-        "saa.c",
-        "saa.h",
-        "srcfile.c",
-        "stdscan.c",
-        "stdscan.h",
-        "strfunc.c",
-        "tables.h",
-        "tokens.h",
-        "tokhash.c",
-        "ver.c",
+        "stdlib/snprintf.c",
+        "stdlib/strlcpy.c",
+        "stdlib/strnlen.c",
+        "stdlib/vsnprintf.c",
         "version.h",
+        "x86/disp8.c",
+        "x86/iflag.c",
+        "x86/iflaggen.h",
+        "x86/insnsa.c",
+        "x86/insnsb.c",
+        "x86/insnsd.c",
+        "x86/insnsi.h",
+        "x86/insnsn.c",
+        "x86/regdis.c",
+        "x86/regdis.h",
+        "x86/regflags.c",
+        "x86/regs.c",
+        "x86/regs.h",
+        "x86/regvals.c",
+    ],
+    includes = [
+        "asm",
+        "include",
+        "output",
+        "x86",
     ],
     copts = select({
         ":windows": [],
@@ -110,7 +151,10 @@ cc_binary(
     defines = select({
         ":windows": [],
         ":windows_msvc": [],
-        "//conditions:default": ["HAVE_SNPRINTF"],
+        "//conditions:default": [
+            "HAVE_SNPRINTF",
+            "HAVE_SYS_TYPES_H",
+        ],
     }),
     visibility = ["@jpeg//:__pkg__"],
 )
diff --git a/third_party/repo.bzl b/third_party/repo.bzl
index cb67d3e9617dd1e9374d07cb1536cedf4bc74ae8..9cee1fcc4b5c2b05ecc09b4f372eadeca9e91be8 100644
--- a/third_party/repo.bzl
+++ b/third_party/repo.bzl
@@ -16,7 +16,6 @@
 
 _SINGLE_URL_WHITELIST = depset([
     "arm_compiler",
-    "ortools_archive",
 ])
 
 def _is_windows(ctx):
diff --git a/third_party/toolchains/BUILD b/third_party/toolchains/BUILD
new file mode 100644
index 0000000000000000000000000000000000000000..fc3183a754369fc30dbce40c2bf7b6828ea497c3
--- /dev/null
+++ b/third_party/toolchains/BUILD
@@ -0,0 +1,22 @@
+licenses(["restricted"])
+
+package(default_visibility = ["//visibility:public"])
+
+# Platform for use with remote execution with
+# custom container based off RBE Ubuntu16_04
+# http://gcr.io/cloud-marketplace/google/rbe-ubuntu16-04
+# Built with //tensorflow/tools/ci_build/Dockerfile.rbe.cpu
+platform(
+    name = "rbe_ubuntu16_04-tf",
+    constraint_values = [
+        "@bazel_tools//platforms:x86_64",
+        "@bazel_tools//platforms:linux",
+        "@bazel_tools//tools/cpp:clang",
+        "@bazel_toolchains//constraints:xenial",
+    ],
+    remote_execution_properties = """
+        properties: {
+            name: "container-image"
+            value:"docker://gcr.io/asci-toolchain/nosla-ubuntu16_04-tf@sha256:800a7b68cabef15419695c188ed33ed70adf678c2371b97b236f3ae26c38274d"
+        }""",
+)
diff --git a/third_party/toolchains/clang6/CROSSTOOL.tpl b/third_party/toolchains/clang6/CROSSTOOL.tpl
index 6b7e5a88086f8e5e67fa86a0e9377c3c2afd535d..ffba9850bb80a880d5b95afacbad296ec1f2df54 100644
--- a/third_party/toolchains/clang6/CROSSTOOL.tpl
+++ b/third_party/toolchains/clang6/CROSSTOOL.tpl
@@ -76,9 +76,6 @@ toolchain {
 
   # This adds a little bit more durability to our Clang build.
   #
-  # At the moment, this only only be needed for:
-  # - add_boringssl_s390x.patch: --Wa,--noexecstack
-  #
   # Folks who do maintenance work on TF Bazel Clang should consider
   # commenting out these lines, while doing that work, to gain a better
   # understanding of what the intersection of support looks like between GCC
diff --git a/tools/bazel.rc b/tools/bazel.rc
index 1c1e6afb65ab8da5b689d58ecaec6ac6c8a69bb8..3559375d5cffa70b8a32c0d6d481f8d5aead1431 100644
--- a/tools/bazel.rc
+++ b/tools/bazel.rc
@@ -36,8 +36,6 @@ build:cuda --define=using_cuda=true --define=using_cuda_nvcc=true
 build:cuda_clang --crosstool_top=@local_config_cuda//crosstool:toolchain
 build:cuda_clang --define=using_cuda=true --define=using_cuda_clang=true --define=using_clang=true
 
-build:win-cuda --define=using_cuda=true --define=using_cuda_nvcc=true
-
 build:mkl --define=using_mkl=true
 
 build:sycl --crosstool_top=@local_config_sycl//crosstool:toolchain