Page MenuHomePhabricator

[libomptarget] Run GPU offloading tests on both new and old runtime
ClosedPublic

Authored by JonChesterfield on Oct 21 2021, 8:10 AM.

Details

Summary

Implemented by patching python config instead of modifying all
the tests so that -generic and XFAIL work as usual. Expectation is for
this to be reverted once the old runtime is deleted.

Diff Detail

Event Timeline

JonChesterfield requested review of this revision.Oct 21 2021, 8:10 AM
Herald added a project: Restricted Project. · View Herald TranscriptOct 21 2021, 8:10 AM

As it stands, this would double the number of offloading tests run on the cuda buildbot

Do you know whether all current regression tests are passing with the newRTL?

openmp/libomptarget/plugins/cuda/CMakeLists.txt
68–69
JonChesterfield added a comment.EditedOct 21 2021, 9:20 AM

They definitely don't on amdgpu. Cuda I assume they are, my local nvidia box is still building clang to confirm. I think they were passing on nvptx a couple of days ago.

I get a failure on sm_50:

$HOME/llvm-project/openmp/libomptarget/test/unified_shared_memory/api.c:18:22: error: Target architecture sm_50 does not support unified addressing
#pragma omp requires unified_shared_memory

which presumably means I've missed whatever should have disabled that test on my machine

edit: patch was fine, test in question doesn't work on the old devicertl either. Copied the xfail.

  • add xfail for new cuda target
  • review comment
openmp/libomptarget/plugins/cuda/CMakeLists.txt
68–69

better, thanks!

This revision is now accepted and ready to land.Oct 22 2021, 10:40 AM
openmp/libomptarget/plugins/cuda/CMakeLists.txt
68–69

this ^ turned out to be a functional change. As originally written, it would only run the newRTL cases. With this form, both run