This is an archive of the discontinued LLVM Phabricator instance.

[mlir][VectorToGPU] Support transposed broadcasted MMA load
ClosedPublic

Authored by antiagainst on Dec 8 2022, 11:30 AM.

Details

Summary

This is now possible with transpose semantics on subgroup MMA
load ops.

Diff Detail

Event Timeline

antiagainst created this revision.Dec 8 2022, 11:30 AM
Herald added a project: Restricted Project. · View Herald Transcript
antiagainst requested review of this revision.Dec 8 2022, 11:30 AM
ThomasRaoux added inline comments.Dec 8 2022, 12:11 PM
mlir/lib/Conversion/VectorToGPU/VectorToGPU.cpp
469–474

for a better check I would check getResult(isTranspose ? 1 : 0)

mlir/test/Conversion/VectorToGPU/vector-to-mma-ops.mlir
198

it should be 0 here

Address comments

antiagainst marked 2 inline comments as done.Dec 14 2022, 10:52 AM
antiagainst added inline comments.
mlir/test/Conversion/VectorToGPU/vector-to-mma-ops.mlir
198

Chatted offline; this is irrelevant anymore.

ThomasRaoux accepted this revision.Dec 14 2022, 10:55 AM
This revision is now accepted and ready to land.Dec 14 2022, 10:55 AM
This revision was automatically updated to reflect the committed changes.
antiagainst marked an inline comment as done.