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

Move layout assignment before instruction fusion for the CPU backend.

Browse Source 
This means we will also have layouts inside fusion nodes, which in turn can
potentially enable us to calculate linear index values for array accesses
which can be reused.

PiperOrigin-RevId: 236815199
This commit is contained in:
Adrian Kuegel 2019-03-05 02:48:02 -08:00 committed by TensorFlower Gardener
parent 55cd3abd02
commit b0315f0960
1 changed files with 5 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

9
tensorflow/compiler/xla/service/cpu/cpu_compiler.cc
View File

@ -318,6 +318,11 @@ Status CpuCompiler::RunHloPassesThroughLayoutAssn(
},
TransposeFolding::NeverFoldTranspose);
pipeline.AddPass<HloCSE>(/*is_layout_sensitive=*/false);
pipeline.AddPass<CpuLayoutAssignment>(
module->mutable_entry_computation_layout(),
LayoutAssignment::InstructionCanChangeLayout, target_machine_features);
pipeline.AddPass<CpuInstructionFusion>();
pipeline.AddPass<ScatterExpander>();
@ -325,10 +330,6 @@ Status CpuCompiler::RunHloPassesThroughLayoutAssn(
ReducePrecisionInsertion::AddPasses(
&pipeline, module->config().debug_options(),
ReducePrecisionInsertion::PassTiming::AFTER_FUSION);
pipeline.AddPass<CpuLayoutAssignment>(
module->mutable_entry_computation_layout(),
LayoutAssignment::InstructionCanChangeLayout, target_machine_features);
return pipeline.Run(module).status();
}