Commit e9158f02 authored by Sanjoy Das's avatar Sanjoy Das Committed by TensorFlower Gardener
Browse files

[XLA:GPU] (Mostly) Disable fast-math optimizations 

We remain performance neutral on the set of tracked benchmarks by allowing
fast-math like optimizations in two places:

 1. We allow LLVM to generate div.approx when necessary by setting
    -nvptx-prec-divf32=0

 2. By default we emit llvm.min/llvm.max that has non IEEE behavior with NaNs.

PiperOrigin-RevId: 223610526
parent 6be20545
Show whitespace changes
Inline Side-by-side
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Please to comment