You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
A user reported ocrs crashing with an "Illegal instruction" error on an M1 Mac, though I was not able to reproduce testing on an M1 myself.
When they ran the tests for RTen, the test_gemm_with_fma_kernel test ran and crashed but everything else passed. This test shouldn't even be run on an Intel Mac. So it looks like the binary was compiled for x86 and run under Rosetta.
AFAIK Rosetta doesn't implement AVX intructions, so that explains the issue. What should happen in this environment is that the fallback SSE kernel gets used instead, so there is likely something in FmaKernel::supported that needs changing.
The text was updated successfully, but these errors were encountered:
This might be fixed now. There was an issue where FmaKernel::supported checked only for fma, but FmaKernel::kernel used avx2 + fma target features. If it is the case that is_x86_feature_detected!("fma") returned true under Apple Silicon but is_x86_feature_detected!("avx2") return false, this could have explained the issue. This inconsistency has now been resolved.
A user reported ocrs crashing with an "Illegal instruction" error on an M1 Mac, though I was not able to reproduce testing on an M1 myself.
When they ran the tests for RTen, the
test_gemm_with_fma_kernel
test ran and crashed but everything else passed. This test shouldn't even be run on an Intel Mac. So it looks like the binary was compiled for x86 and run under Rosetta.AFAIK Rosetta doesn't implement AVX intructions, so that explains the issue. What should happen in this environment is that the fallback SSE kernel gets used instead, so there is likely something in
FmaKernel::supported
that needs changing.The text was updated successfully, but these errors were encountered: