This is an archive of the discontinued LLVM Phabricator instance.

Revert "Reorder linalg.conv indexing_maps loop order"
ClosedPublic

Authored by antiagainst on Nov 19 2020, 6:39 AM.

Details

Summary

This reverts commit 9b47525824df9be5ae23c39e7ce0d220d12f85e5
and falls back to the original parallel-iterators-as-leading-
dimensions convention. We can control the loop order by first
converting the named op into linalg.generic and then performing
interchange.

Diff Detail

Event Timeline

antiagainst created this revision.Nov 19 2020, 6:39 AM
antiagainst requested review of this revision.Nov 19 2020, 6:39 AM
nicolasvasilache accepted this revision.Nov 19 2020, 7:33 AM

I'm good with this but please wait for @asaadaldien to confirm he can recoup the perf on the things he is looking at before pushing.

This revision is now accepted and ready to land.Nov 19 2020, 7:33 AM
asaadaldien accepted this revision.Nov 19 2020, 9:32 AM

Thanks @antiagainst! A pass added to in IREE to convert linalg.conv to linalg.generic then apply loop interchange.

This revision was automatically updated to reflect the committed changes.