This is an archive of the discontinued LLVM Phabricator instance.

[mlir] Add sparse linalg.dot to LLVMAVX512 translation test.
AbandonedPublic

Authored by springerm on Mar 15 2021, 3:22 AM.

Details

Summary

Add a transform that rewrites a sparse linalg.dot into two nested for loops.

Diff Detail

Event Timeline

springerm created this revision.Mar 15 2021, 3:22 AM
springerm requested review of this revision.Mar 15 2021, 3:22 AM

LLVMAVX512 no longer exists, please just use AVX512 in names

LLVMAVX512 no longer exists, please just use AVX512 in names

Ah yes, bad naming. This should be AVX512.

springerm updated this revision to Diff 330594.Mar 15 2021, 3:32 AM

Rename LLVMAVX512 to AVX512.

nicolasvasilache accepted this revision.Mar 17 2021, 1:31 AM
nicolasvasilache added inline comments.
mlir/test/lib/Transforms/TestSparseLinalgDotToAVX512.cpp
54

Can you please add a TODO that we want to decompose this in smaller composable pieces later?

This revision is now accepted and ready to land.Mar 17 2021, 1:31 AM
springerm updated this revision to Diff 331437.Mar 17 2021, 6:34 PM

Add comment.

mehdi_amini requested changes to this revision.Mar 17 2021, 8:47 PM
mehdi_amini added inline comments.
mlir/test/lib/Transforms/TestSparseLinalgDotToAVX512.cpp
248

That looks like a lot of code for the "test" directory. It does not seems adequate to me: this directory should not add significant "business" logic but only glue to test the compiler libraries.
Does this belong to some sort of lowering / conversion?

This revision now requires changes to proceed.Mar 17 2021, 8:47 PM

This is a very old revision. Still relevant?

springerm abandoned this revision.Jun 24 2022, 1:03 AM
springerm marked an inline comment as done.