AMD Zen 5 architecture is misidentified

[yurivict]

The attached reproducer crashes on AMD Zen 5, but succeeds when run with OPENBLAS_CORETYPE=HASWELL or OPENBLAS_CORETYPE=SKYLAKEX. It also fails with OPENBLAS_CORETYPE=GENERIC which doesn't make sense to me.

The offending instruction is: movaps %xmm0,0x0(%r13)

reproducer.c

openblas-0.3.34
clang version 21.1.8
OS: FreeBSD 15.1

[martin-frbg]

I don't think there's any misidentification involved - your reproducer is creating badly misaligned input on purpose, which happens to hit a "move aligned data" instruction in the CooperLake/Zen5+ kernel where the Haswell and Skylake counterparts have probably been changed to movups already.
The same fault happening with the plain C "generic" kernels suggests that the compiler is defaulting to "movaps" in its code generation as well - I'm not sure how representative your reproducer is of real-world code here ?

[martin-frbg]

Can you tell what your system gets misidentified as (running with OPENBLAS_VERBOSE=2 should show) ? COOPERLAKE is basically SKYLAKEX with additional bfloat16 extensions, so you should be hitting the exact same code paths.

[martin-frbg]

I cannot reproduce your crash on Linux, unless I deliberately force the use of the old AMD Bulldozer kernel (which indeed has a few movaps instructions). It would be trivial to change them to movups, but I'm not convinced that particular hardware is still relevant.

[yurivict]

I'm not sure how representative your reproducer is of real-world code here ?

I didn't mention this in the OP, but I originally hit this problem on this PyTorch example with pytorch-2.14.0 which has exactly same crash and same env vars to prevent it.

---

It shouldn't matter that the data is badly misaligned - it still shouldn't use invalid ASM instructions no matter what.
Are you saying that this is a compiler that creates them?
The compiler is llvm-21.1.8 on FreeBSD 15.1

[martin-frbg]

Yes, I suspect it is the compiler generating them from C code or intrinsics "somewhere" - unless you're actually running Piledriver/Excavator kernels from 10+ years ago (which should 't happen unless you actually request them). Every newer SGEMM kernel should be using (v)movups, and I don't think Linux vs FreeBSD matters in this regard.

[martin-frbg]

Again, what does running the reproducer with OPENBLAS_VERBOSE=2 print on your system ? It should be "Cooperlake" for Zen5, or "Haswell" for earlier generations, but normally nothing like Piledriver (unless you're running BSD under some hypervisor that invents its own cpuid)