This is an archive of the discontinued LLVM Phabricator instance.

AMDGPU: Leave WQM earlier before VMEM loads
AbandonedPublic

Authored by nhaehnle on Jul 10 2016, 7:36 AM.

Details

Summary

When the initial propagation determines that a VMEM load does not have to run
in WQM, make it run in Exact mode instead if the remainder of the shader
doesn't care. This can reduce the bandwidth required for loads.

This typically introduces two additional SALU instructions and uses two
additional SGPRs in pixel shaders for the livemask.

Diff Detail

Event Timeline

nhaehnle updated this revision to Diff 63427.Jul 10 2016, 7:36 AM
nhaehnle retitled this revision from to AMDGPU: Leave WQM earlier before VMEM loads.
nhaehnle updated this object.
nhaehnle added a subscriber: llvm-commits.
nhaehnle updated this revision to Diff 64253.Jul 17 2016, 6:19 AM

Rebase and add test case changes that I missed.

arsenm added inline comments.Jul 21 2016, 3:34 PM
lib/Target/AMDGPU/SIWholeQuadMode.cpp
372

TII->usesVM_CNT(MI)

nhaehnle updated this revision to Diff 65683.Jul 27 2016, 2:05 AM

Use usesVM_CNT

tstellarAMD accepted this revision.Sep 8 2016, 6:50 AM
tstellarAMD edited edge metadata.
This revision is now accepted and ready to land.Sep 8 2016, 6:50 AM

Is this still needed?

nhaehnle abandoned this revision.Jul 27 2020, 2:12 AM