[maleadt]

Test PR to see whether the Aurora LTS stack (NEO 25.18, IGC 2.11) accepts SPIR-V from the LLVM back-end. #574 switched the LTS stack to the Khronos translator because its IGC reportedly didn't accept the back-end's output, but there's no recorded reproducer.

The translator is now the limiting factor: it doesn't support llvm.ssub.with.overflow or llvm.smul.with.overflow, so signed checked_sub/checked_mul in a kernel fail to compile with ONEAPI_LTS=1 (this is what broke the LTS job on #707). If the ALCF job passes here, the translator could be dropped altogether.

The other LTS workarounds (BFloat16 off, strided reductions, queue draining) are unchanged.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 82.43%. Comparing base (f644589) to head (7aa2afe).
⚠️ Report is 9 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #708      +/-   ##
==========================================
+ Coverage   82.40%   82.43%   +0.03%     
==========================================
  Files          52       51       -1     
  Lines        4415     4241     -174     
==========================================
- Hits         3638     3496     -142     
+ Misses        777      745      -32     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[maleadt]

@michel2323 Could you ask Intel to include a SPIRV-LLVM-Translator bump in the LTS IGC update? IGC 2.11 bundles llvm_release_150 @ 2d4f2e7, which turns SPIR-V OpIAddCarry/OpISubBorrow into calls to __spirv_IAddCarry/__spirv_ISubBorrow that IGC doesn't implement ("undefined reference to _Z17__spirv_IAddCarryll", module build failure). It's fixed by KhronosGroup/SPIRV-LLVM-Translator#2877, backported to llvm_release_150 as KhronosGroup/SPIRV-LLVM-Translator#3761 (998b2cc6); upstream IGC has had it since 2.38.2. It breaks unsigned checked_add/checked_sub in kernels today, and it's what keeps oneAPI.jl from using the LLVM SPIR-V back-end on Aurora.

[michel2323]

I ran into this issue too, and it seems to be the only remaining blocker to bring LTS to the LLVM translator. I've created a SYCL reproducer, and I'll confirm it on Aurora tomorrow and forward it to our bug tracker.