Preparation for LIBXSMM 1.8 (#13)

* Fixed libxsmm_config_arguments: Fixed the incorrect value supposed to trigger auto-prefetch. Fixed the 0-threshold, which is now accounted for in LIBXSMM (by just populating the default threshold). The problem arised from the assumption "threshold: fallback to BLAS if n*m*k above this", which is wrong (the threshold populates an upper bound until which JIT code is generated). The previous configuration perhaps caused all sorts of issues due to other values derived from the 0-threshold. Note, explicitly JIT'ting code is/was never subject to a threshold.

* Upgraded to libxsmm 1.6.5

* Enable the use of libxsmm for matrix multiplications

* Enable the use of libxsmm to speedup 1x1 convolutions (which are
computed using matrix multiplications)

* Fixed libxsmm_config_arguments in libxsmm.BUILD (#7)

* Fixed libxsmm_config_arguments: Fixed the incorrect value supposed to trigger auto-prefetch. Fixed the 0-threshold, which is now accounted for in LIBXSMM (by just populating the default threshold). The problem arised from the assumption "threshold: fallback to BLAS if n*m*k above this", which is wrong (the threshold populates an upper bound until which JIT code is generated). The previous configuration perhaps caused all sorts of issues due to other values derived from the 0-threshold. Note, explicitly JIT'ting code is/was never subject to a threshold.

* Make use of TensorFlow's allocation infrastructure even when using LIBXSMM allocation functions. In particular, the (cached) libxsmm_spmdm_init now relies on TF's cpu_allocator().

For C++ code, one can use a libxsmm_scoped_allocator<kind> in order to (temporarily) setup a different allocation mechanism. For instance, using libxsmm_tf_allocator<libxsmm_scratch_allocator> changes LIBXSMM's scratch allocator to rely on TensorFlow. The libxsmm_tf_allocator provides two kinds of c'tors: (1) the no-argument variant adopts TF's cpu_allocator(), whereas the one-argument form (2) adopts the allocator from the given OpKernelContext. Changing the allocator in LIBXSMM with pending buffers (from different allocators) is valid, and all other services in LIBXSMM's "malloc domain" work regardless of the allocation mechanism (e.g., libxsmm_malloc_size).

* Simply renamed API items in order to follow changes in LIBXSMM 1.7. This is incomplete as more changes/adjustments are needed.

* Account for removed non-check API.

* Include libxsmm_malloc.h now that libxsmm_tf_allocator is used.

* Renamed libxsmm_dnn_create_conv_handle to libxsmm_dnn_create_conv_layer.

* Renamed LIBXSMM_DNN_CONV_FORMAT_* to LIBXSMM_DNN_TENSOR_FORMAT_*.

* Renamed libxsmm_dnn_destroy_conv_handle to libxsmm_dnn_destroy_conv_layer.

* Include missing header file (libxsmm_malloc.h).

* Renamed LIBXSMM_DNN_CONV_KIND_* to LIBXSMM_DNN_COMPUTE_KIND_*.

* Account for the fact that datatype_in/out is now only datatype (libxsmm_dnn_conv_desc structure).

* Updated to new libxsmm_dnn_link_* functions.

* Updated to use new libxsmm_dnn_bind_* functions.

* Fixed calling libxsmm_dnn_transpose_filter.

* Updates in preparation of LIBXSMM 1.7 (#8)

* Fixed libxsmm_config_arguments: Fixed the incorrect value supposed to trigger auto-prefetch. Fixed the 0-threshold, which is now accounted for in LIBXSMM (by just populating the default threshold). The problem arised from the assumption "threshold: fallback to BLAS if n*m*k above this", which is wrong (the threshold populates an upper bound until which JIT code is generated). The previous configuration perhaps caused all sorts of issues due to other values derived from the 0-threshold. Note, explicitly JIT'ting code is/was never subject to a threshold.

* Upgraded to libxsmm 1.6.5

* Enable the use of libxsmm for matrix multiplications

* Enable the use of libxsmm to speedup 1x1 convolutions (which are
computed using matrix multiplications)

* Make use of TensorFlow's allocation infrastructure even when using LIBXSMM allocation functions. In particular, the (cached) libxsmm_spmdm_init now relies on TF's cpu_allocator().

For C++ code, one can use a libxsmm_scoped_allocator<kind> in order to (temporarily) setup a different allocation mechanism. For instance, using libxsmm_tf_allocator<libxsmm_scratch_allocator> changes LIBXSMM's scratch allocator to rely on TensorFlow. The libxsmm_tf_allocator provides two kinds of c'tors: (1) the no-argument variant adopts TF's cpu_allocator(), whereas the one-argument form (2) adopts the allocator from the given OpKernelContext. Changing the allocator in LIBXSMM with pending buffers (from different allocators) is valid, and all other services in LIBXSMM's "malloc domain" work regardless of the allocation mechanism (e.g., libxsmm_malloc_size).

* Simply renamed API items in order to follow changes in LIBXSMM 1.7. This is incomplete as more changes/adjustments are needed.

* Account for removed non-check API.

* Include libxsmm_malloc.h now that libxsmm_tf_allocator is used.

* Renamed libxsmm_dnn_create_conv_handle to libxsmm_dnn_create_conv_layer.

* Renamed LIBXSMM_DNN_CONV_FORMAT_* to LIBXSMM_DNN_TENSOR_FORMAT_*.

* Renamed libxsmm_dnn_destroy_conv_handle to libxsmm_dnn_destroy_conv_layer.

* Include missing header file (libxsmm_malloc.h).

* Renamed LIBXSMM_DNN_CONV_KIND_* to LIBXSMM_DNN_COMPUTE_KIND_*.

* Account for the fact that datatype_in/out is now only datatype (libxsmm_dnn_conv_desc structure).

* Updated to new libxsmm_dnn_link_* functions.

* Updated to use new libxsmm_dnn_bind_* functions.

* Fixed calling libxsmm_dnn_transpose_filter.

* integrated LIBXSMM 1.7

* support for LIBXSMM 1.7 (#9)

* Upgraded to libxsmm 1.6.5

* Enable the use of libxsmm for matrix multiplications

* Enable the use of libxsmm to speedup 1x1 convolutions (which are
computed using matrix multiplications)

* integrated LIBXSMM 1.7

* updated LIBXSMM to 1.7.1

* updated to LIBXSMM 1.7.1 (#10)

* Upgraded to libxsmm 1.6.5

* Enable the use of libxsmm for matrix multiplications

* Enable the use of libxsmm to speedup 1x1 convolutions (which are
computed using matrix multiplications)

* integrated LIBXSMM 1.7

* updated LIBXSMM to 1.7.1

* merge alheinecke master (#11)

* Upgraded to libxsmm 1.6.5

* Enable the use of libxsmm for matrix multiplications

* Enable the use of libxsmm to speedup 1x1 convolutions (which are
computed using matrix multiplications)

* integrated LIBXSMM 1.7

* updated LIBXSMM to 1.7.1

* Take new translation units into account (LIBXSMM 1.8).

* Account for adjusted header dependency in LIBXSMM (TODO: API to avoid incl. header from LIBXSMM's src).

* Trigger rebuild if template changed (LIBXSMM).