This is an archive of the discontinued LLVM Phabricator instance.

[mlir][sparse][gpu] put sparse compiler GPU end-to-end tests back
ClosedPublic

Authored by aartbik on Apr 8 2023, 2:34 PM.

Diff Detail

Event Timeline

aartbik created this revision.Apr 8 2023, 2:34 PM
Herald added a project: Restricted Project. · View Herald TranscriptApr 8 2023, 2:34 PM
aartbik requested review of this revision.Apr 8 2023, 2:34 PM

Do we have tests that require a specific architecture available (things like AVX512 or other fancy things) that we can model against?

aartbik added a comment.EditedApr 10 2023, 10:17 AM

Do we have tests that require a specific architecture available (things like AVX512 or other fancy things) that we can model against?

When I added e.g. the AMX test, I added a new configuration for that

if not config.mlir_run_amx_tests:
    config.unsupported = True

which can be controlled from the setup through DMLIR_RUN_AMX_TESTS.
I can do the same here, of course, but I was really hoping that something like this was already there.

Seems like I am always the first to actually run these things ;-)

Thomas suggests perhaps we should upgrade the bots from turing to ampere ;-)
But until that time, I will add a build flag for these tests.

Any preference for the name?

aartbik updated this revision to Diff 512537.Apr 11 2023, 11:55 AM

add build flag for A100 GPUs

PTAL, I added a MLIR_RUN_CUDA_SM80_TESTS build flag

aartbik updated this revision to Diff 512560.Apr 11 2023, 1:15 PM

rebased with main

aartbik edited the summary of this revision. (Show Details)Apr 11 2023, 1:22 PM
wrengr added inline comments.Apr 11 2023, 2:41 PM
mlir/test/Integration/Dialect/SparseTensor/GPU/CUDA/sparse-mma-2-4-f16.mlir
13

ditto

Bah, phabricator didn't save the comment I was referring to with that "ditto". So here it is again:

mlir/test/Integration/Dialect/SparseTensor/GPU/CUDA/sparse-matvec.mlir
10

Maybe this note should be on the first line, so users see it before all the RUN stuff?

aartbik marked 2 inline comments as done.Apr 11 2023, 2:54 PM
aartbik updated this revision to Diff 512594.Apr 11 2023, 2:56 PM

addressed comment

mehdi_amini accepted this revision.Apr 11 2023, 3:03 PM
mlir/test/Integration/Dialect/SparseTensor/GPU/CUDA/lit.local.cfg
2

Do we plan to not ever have tests in this folder that would run on a <sm80 GPU? (otherwise we should use a REQUIRE annotation in the tests themselves instead)

This revision is now accepted and ready to land.Apr 11 2023, 3:03 PM
aartbik marked an inline comment as done.Apr 11 2023, 3:15 PM
aartbik added inline comments.
mlir/test/Integration/Dialect/SparseTensor/GPU/CUDA/lit.local.cfg
2

Good question, but I am not sure of the answer (yet), since GPU support for the sparse compiler and in particular specifically for CUDA is just ramping up. Perhaps we want to target lower versions too as some point. But I made a note of your suggestion to reconsider when we have more visibility in what targets we will converge on...

This revision was automatically updated to reflect the committed changes.
aartbik marked an inline comment as done.