[SYCL][NFC] Fix check-work-group-attributes-match.cpp for AMDGPU#22446
Conversation
b3ca5fb introduced amdgpu-flat-work-group-size attributes with per-kernel values, causing each kernel to get a distinct attribute group (#0, intel#2, intel#4) on amdgcn targets. Relax the attribute group checks from the hardcoded #0 to #{{[0-9]+}}. Fixes CodeGenSYCL/check-work-group-attributes-match.cpp CMPLRLLVM-76303 Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
) b3ca5fb introduced amdgpu-flat-work-group-size attributes with per-kernel values, causing each kernel to get a distinct attribute group. Fixes CodeGenSYCL/check-work-group-attributes-match.cpp CMPLRLLVM-76303 --------- Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
|
@wenju-he, SYCL doesn't support AMDGPU anymore. Do we need to maintain both b3ca5fb and c53885f changes in this repository? @intel/dpcpp-cfe-reviewers, @tahonermann, what do you think? |
|
Since intel/llvm/dpclang is now being productized and beginning to be included in Linux distributions, I think caution is warranted with regard to discarding support for other vendor hardware. We might begin to see contributions outside of Intel as the SYCL ecosystem grows; which could be a considerable benefit to Intel. For that reason, I think it is worth putting minimal effort towards keeping what currently works working as a gesture of good will to the SYCL community. I don't think doing so conflicts with Intel's goals to prioritize support for its own products. |
|
will try to upstream #22452 |
sorry, I mean it is not the task from pulldown coordinator. |
b3ca5fb introduced amdgpu-flat-work-group-size attributes with per-kernel values, causing each kernel to get a distinct attribute group (#0, #2, #4) on amdgcn targets. Relax the attribute group checks from the hardcoded #0 to #{{[0-9]+}}.
Fixes CodeGenSYCL/check-work-group-attributes-match.cpp CMPLRLLVM-76303