experiments/STT-tensorflow
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

[XLA] Add TODO comment explaining why the IsScalar check exists.

Browse Source 
PiperOrigin-RevId: 163292777
This commit is contained in:
A. Unique TensorFlower 2017-07-26 19:56:14 -07:00 committed by TensorFlower Gardener
parent 43036ac16c
commit 4404aa7cbc
1 changed files with 5 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

5
tensorflow/compiler/xla/service/reduce_precision_insertion.cc
View File

@ -41,6 +41,11 @@ StatusOr<bool> ReducePrecisionInsertion::Run(HloModule* module) {
// equivalent behavior can be obtained by adding ReducePrecision // equivalent behavior can be obtained by adding ReducePrecision
// instructions after the instructions that pull the F32 arrays out of // instructions after the instructions that pull the F32 arrays out of
// the tuples. // the tuples.
//
// TODO(b/64093391): Remove the IsScalar check once this won't cause
// failures on the GPU backend if the ReducePrecision instruction ends up
// inserted between a scalar constant and the init_value argument of a
// Reduce operation.
if (instruction->shape().element_type() == PrimitiveType::F32 && if (instruction->shape().element_type() == PrimitiveType::F32 &&
!ShapeUtil::IsScalar(instruction->shape()) && !ShapeUtil::IsScalar(instruction->shape()) &&
should_reduce_output_precision_(instruction->opcode())) { should_reduce_output_precision_(instruction->opcode())) {