[Bug 273219] math/openblas: update to 0.3.24

From: <bugzilla-noreply_at_freebsd.org>
Date: Sat, 30 Sep 2023 00:15:38 UTC

--- Comment #7 from Eijiro Shibusawa <phd_kimberlite@yahoo.co.jp> ---
(In reply to Daniel Engberg from comment #5)

I understood a packing problem, at least, partially. 
I'll fix distfile issue when I have a time, because I'll be busy for a while.
However I am really disappointed in you because you don't provide any solution
for that you pointed.
You should understand the background of openblas software and also consider
it's usability. 
I think that the most part of users are not interested in developer's issue but
use of fast BLAS.

> https://gitweb.gentoo.org/repo/gentoo.git/tree/sci-libs/openblas
With one glance, I did not find the patch that contributs to abstraction of CPU
Could please you specify the patch that may contribute the solution, otherwise
that issue is also not related the problem that you pointed.

> Both AVX and AVX2 are SMID instructions?
YES, Both are SIMD instructions.

I think your provided log may be the result of respecting your CFLAGS.
The following is corresponding part of poudriere log of my freebsd box.
cc -DCBLAS -c -O2 -pipe  -fstack-protector-strong -fno-strict-aliasing  -O2
-DASMFNAME=cblas_strmv_ -DNAME=cblas_strmv_ -DCNAME=cblas_strmv
-DCHAR_NAME=\"cblas_strmv_\" -DCHAR_CNAME=\"cblas_strmv\" -DNO_AFFINITY -I..
-I. -UDOUBLE  -UCOMPLEX trmv.c -o cblas_strmv.o
I didn't understand why -march=tigerlake is generated, the port may generate
-march=skylake-avx512 however tigerlake is not.
There is no doubt that somewhat strange things occurred.
The module cblas_* is only I/F and not the optimized kernel function, so it is
hard to think that they contain SIMD instructions.
The openblas is highly HW dependent so the discussion should be based on the
poudriere log with default settings.
Even if the default settings some of issue may not be reproduced.

You are receiving this mail because:
You are the assignee for the bug.