New Issue: Unknown issues with ROCm 7 performance testing

28955, "benharsh", "Unknown issues with ROCm 7 performance testing", "2026-06-09T23:14:41Z"

We have recently observed consistent failures in the following tests after the upgrade to ROCm 7.0:

  1. gpu/native/studies/sort.radixSortGpu
internal error: gpu-amd-cub.cc:106: Error calling HIP function: named symbol not found (Code: 500)
  1. gpu/native/studies/chop.queens_GPU_CPU_single_node (empty output)
  2. gpu/native/studies/shoc.shoc-sort
Memory access fault by GPU node-4 (Agent handle: 0xXXXXXXXXXXXX) on address 0xXXXXXXXXXXXX. Reason: Unknown.

I have currently traced these failures down to two related causes:

  • compiling with --fast will always cause these errors, even if we throw --ccflags -O0 to disable LLVM optimizations. This implies that one of our own compiler passes is generating something incorrectly
  • Even without --fast, compiling with --optimize (which throws -O3) will cause the failures. Either the code we're generating is wrong even without optimizations, or we could potentially be dealing with an LLVM bug.

Further work is required to better understand the potential causes.