Page MenuHomePhabricator

[mlir] Fix gpu MMA integrations tests
ClosedPublic

Authored by ThomasRaoux on May 25 2021, 9:34 AM.

Details

Summary

Don't outline the kernel in the test file as this prevent some debug info from being stripped out. Cuda driver doesn't support PTX with debug info causing conversion to cubin to fail.

Diff Detail

Event Timeline

ThomasRaoux created this revision.May 25 2021, 9:34 AM
ThomasRaoux requested review of this revision.May 25 2021, 9:34 AM

The test had started failing with https://github.com/llvm/llvm-project/commit/81467f500f6ad106a69088bc276024c5e1938571. I'll also enable those tests in google build bots that have Tesla T4 GPUs once this is fixed.

navdeepkk accepted this revision.May 27 2021, 9:12 AM

Thanks! this is something I wasn't aware of. BTW I tested these on a Turing with CUDA10.2, and they passed, but maybe they fail on some other devices.

mlir/test/Integration/GPU/CUDA/TensorCore/wmma-matmul-f16.mlir
46

This is I think present since this file was added but not required anymore. Can you please drop this? Or should I remove this in a subsequent patch?

This revision is now accepted and ready to land.May 27 2021, 9:12 AM

Thanks! this is something I wasn't aware of. BTW I tested these on a Turing with CUDA10.2, and they passed, but maybe they fail on some other devices.

Yes I ended up doing a more generic fix for the problem as it causing more general problems: https://reviews.llvm.org/D103187

This still feels like a small improvement so I'll move forward with this patch unless you have any concerns.

Thanks! this is something I wasn't aware of. BTW I tested these on a Turing with CUDA10.2, and they passed, but maybe they fail on some other devices.

Yes I ended up doing a more generic fix for the problem as it causing more general problems: https://reviews.llvm.org/D103187

This still feels like a small improvement so I'll move forward with this patch unless you have any concerns.

Yes. We can go ahead with this patch.

ThomasRaoux marked an inline comment as done.
This revision was landed with ongoing or failed builds.May 27 2021, 9:45 AM
This revision was automatically updated to reflect the committed changes.