[SYCL][CUDA] Bug fix: non-uniform work-groups error on CUDA #2857
Conversation
There was a problem hiding this comment.
LGTM, but I'd like @v-klochkov to review. Also a test would be great.
|
@huanghua1994, could add a regression test following this guide, please? |
|
@bader Sure, will do |
|
Tested on vs-sycl-2.hd.intel.com. Test results: |
|
Sorry for confusion. What I meant is that we should add a new regression test to validate that issue #2821 is fixed. |
|
Hi, I am in process of debugging a bit bigger and more complete patch for exactly same issue. The intention was to upload it for review yesterday, but the new LIT test added by me there showed some new unexpected issues with work-group size for Aux kernels with reductions. I'll take some more time to analyze it. |
|
I uploaded a fix that includes similar code as this patch, but is a bit bigger patch and includes a LIT test: #2859 |
|
@v-klochkov Thanks! That's a better fix and should be merged. @Pennycook please review Vyecheslav's PR first, and close my PR after passing his PR. |
|
The bug is fixed by #2859. |
Ensure that we do correct translation of matrices (SPV_KHR_cooperative_matrix and SPV_INTEL_joint_matrix extensions) when untyped pointers are enabled. This mainly fixes mangling issues in reverse translation for the untyped pointer. Also added handling for typed and untyped SPIR-V friendly access chain instructions in forward translation (a point to review and discuss). Original commit: KhronosGroup/SPIRV-LLVM-Translator@a1f6742cd6a2ce2
Bug fix for issue #2821