Segfaults with scipy/openblas since 4989e039a5

[carlkl]

Since 'Define SBGEMM_ALIGN_K for DYNAMIC_ARCH build' I get segfaults and failures with scipy.test() in the following setting:

OS: Windows 64-bit
Compiler: msys2 GCC-12
OpenBLAS: 4989e039a5
Python: vanilla CPython-3.10
Scipy: local scipy-meson build

No further examinations were carried out right now other than narrowing down the problem to the mentioned commit.

[martin-frbg]

Can you be more specific about your build please ? It seems surprising to me that this particular commit would cause segfaults (where?) on an unrelated platform - do the scipy tests even use bfloat16 ?

[carlkl]

In kernel/setparam-ref.c the parameter align_k was removed, and this is related to a Windows DYNAMIC_ARCH build, maybe by accident, in this commit.

[martin-frbg]

uh-oh, there is a stray "1" in kernel/setparam-ref.c as a leftover from the initial placement of the "align_k"....

[carlkl]

Maybe (wild guess) related to: #3801?

[carlkl]

I will try to test this PR with scipy later this evening.

[martin-frbg]

unlikely to be related to 3801 as I only merged the "offending" PR yesterday while the (original) Julia ticket is a week old. (plus I don't think they use bleeding-edge builds in Julia, so more likely something to do with "recent" LAPACK 3.10.1 or a general (Windows-specific?) threading issue that only happened to be "felt" now

[carlkl]

I can confirm, that with Remove excess initializer (leftover from rework of PR 3793) scipy.test() is running without failures and segfaults again.