torch-mlir icon indicating copy to clipboard operation
torch-mlir copied to clipboard

Published 20 hours ago • verified publisher icon llvm

TOSA lowering is very slow

Open silvasean opened this issue 2 years ago • 3 comments

If you run https://github.com/llvm/torch-mlir/blob/main/examples/torchscript_resnet18_all_output_types.py then the TOSA lowering seems to take much longer than the other lowerings.

CC @sjarus

silvasean avatar Apr 20 '22 17:04 silvasean

Very interesting. Thanks for raising this - I forgot to acknowledge seeing this issue posted earlier today.

sjarus avatar Apr 21 '22 02:04 sjarus

So it's a vision model. Maybe this is hitting https://github.com/google/iree/issues/8855 too.

antiagainst avatar Apr 21 '22 02:04 antiagainst

Quickly updating here - I coordinated internally with @GeorgeARM and am helping him work out the refactoring proposed in https://github.com/google/iree/issues/8855 , and in the process increase the number of people internally who are familiar with the LLVM contribution process. Meanwhile I'll test it out with Torch-MLIR separately to see if it indeed addresses this.

sjarus avatar Apr 28 '22 01:04 sjarus