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

Reorder kernel gen passes.

Browse Source 
This is due to moving patterns between the two  passes to allow proper unranked handling of both binary and special unary functions.

PiperOrigin-RevId: 337840967
Change-Id: Ia4f560526cbe65d13ab455dd83230e05f62c9421
This commit is contained in:
Tres Popp 2020-10-19 06:15:12 -07:00 committed by TensorFlower Gardener
parent 635faa44b2
commit 756a3d71ed
1 changed files with 1 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
tensorflow/compiler/mlir/tools/kernel_gen/kernel_creator.cc
View File

@ -87,8 +87,8 @@ Status LowerTFtoGPU(mlir::ModuleOp module, bool gpu_binary_only,
} else { } else {
pm.addPass(mlir::mhlo::createLegalizeTFPass( pm.addPass(mlir::mhlo::createLegalizeTFPass(
/*allow_partial_conversion=*/false, /*legalize_chlo=*/false)); /*allow_partial_conversion=*/false, /*legalize_chlo=*/false));
pm.addPass(mlir::mhlo::createChloLegalizeToHloPass());
pm.addPass(mlir::createTransformUnrankedHloPass()); pm.addPass(mlir::createTransformUnrankedHloPass());
pm.addPass(mlir::mhlo::createChloLegalizeToHloPass());
pm.addPass(mlir::kernel_gen::transforms::CreateShapeToDescriptorsPass()); pm.addPass(mlir::kernel_gen::transforms::CreateShapeToDescriptorsPass());
// Clean up the IR created above. In particular, operations on descriptors // Clean up the IR created above. In particular, operations on descriptors
// are simplified here. // are simplified here.