This is an archive of the discontinued LLVM Phabricator instance.

[AMDGPU][OpenMP] Disable tests when amdgpu-arch fails
ClosedPublic

Authored by pdhaliwal on May 7 2021, 1:19 AM.

Details

Summary

This patch prevents runtime tests running on systems without amdgpu.

Diff Detail

Event Timeline

pdhaliwal created this revision.May 7 2021, 1:19 AM
pdhaliwal requested review of this revision.May 7 2021, 1:19 AM
Herald added a project: Restricted Project. · View Herald Transcript
pdhaliwal updated this revision to Diff 343608.May 7 2021, 1:22 AM

Remove useless libomptarget_say

Looks reasonable to me. Let's wait for a second opinion though

tianshilei1992 accepted this revision.May 7 2021, 7:51 AM

LGTM. Actually I'm wondering if it is possible to do something like: check the eligibility in CMake and set something, then in lit if corresponding platform is not ready (no matter what reason), simply ignore corresponding execution line.

This revision is now accepted and ready to land.May 7 2021, 7:51 AM

Similarly we could do the test in lit.

I'd like something similar for cuda, where a system may have cuda installed but no nvidia card present. I wonder if there's an executable in the cuda toolkit that could serve the same purpose.

protze.joachim accepted this revision.May 7 2021, 9:45 AM

lgtm as well.

I'd like something similar for cuda, where a system may have cuda installed but no nvidia card present. I wonder if there's an executable in the cuda toolkit that could serve the same purpose.

$ nvidia-smi -L
GPU 0: Tesla V100-SXM2-16GB (UUID: GPU-xyz...)
GPU 1: Tesla V100-SXM2-16GB (UUID: GPU-xyz...)
SUMMARY OPTIONS
-L, --list-gpus
    List each of the NVIDIA GPUs in the system, along with their UUIDs.

I cannot test, what the tool does, if no NVIDIA GPU is available in a system. We only have the tool installed on nodes with GPUs. It might return an error code or just have an empty output.

This revision was automatically updated to reflect the committed changes.