@llvm/pr-subscribers-offload

@llvm/pr-subscribers-backend-amdgpu

Author: None (Kewen12)

diff --git a/offload/cmake/caches/AMDGPULibcBot.cmake b/offload/cmake/caches/AMDGPULibcBot.cmake
index 728dfe3f0a3f1..a772043c79669 100644
--- a/offload/cmake/caches/AMDGPULibcBot.cmake
+++ b/offload/cmake/caches/AMDGPULibcBot.cmake
@@ -18,3 +18,4 @@ set(CLANG_DEFAULT_RTLIB "compiler-rt" STRING "")
 
 set(LLVM_RUNTIME_TARGETS default;amdgcn-amd-amdhsa CACHE STRING "")
 set(RUNTIMES_amdgcn-amd-amdhsa_LLVM_ENABLE_RUNTIMES "compiler-rt;libc" CACHE STRING "")
+set(RUNTIMES_amdgcn-amd-amdhsa_LIBC_GPU_TEST_JOBS 4 CACHE STRING "")

No write access. would appreciate if you could help review @jplehr @jhuber6 TIA!

Thanks for the help @jhuber6! I might not have full context here, you mean enabling this flag may not be safe?

Thanks for the help @jhuber6! I might not have full context here, you mean enabling this flag may not be safe?

Yes, the HSA runtime would routinely crash when many of these tests were run in parallel. I poked at it through https://github.com/jhuber6/hsa_test awhile back, pretty much just found that loading binaries in parallel would crash depending on the machine.

"routinely crash" love it :-D

pretty much just found that loading binaries in parallel would crash depending on the machine

This sounds like a loader issue. CC @kzhuravl

pretty much just found that loading binaries in parallel would crash depending on the machine

This sounds like a loader issue. CC @kzhuravl

Who knows, maybe they fixed it, haven't checked in awhile.

We've been running this config on the current libc bot for about 6 months now or so (ROCm 6.2 and ROCm 6.3) and did not see spurious fails in that time.
So I guess, something has improved.