[Bug 276789] math/openblas: Update to 0.3.26

https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=276789

--- Comment #5 from Artyom Davidov <ard_1@mail.ru> ---
(In reply to Daniel Engberg from comment #4)
Hello Daniel!
I'm sorry, but I'm totally against a proposed approach.
It doesn't really simplify port it and makes it harder to debug and maintain.

First of all, combining several AVX options into a single global one makes the
port less flexible. There are several reasons why there are three AVX options
provided by OpenBLAS developers and I don't see any reason why we should
combine them into a single one by ourselves.

The second one is that there is no reason why we should follow an Arch Linux
approach, since the current math/openblas port structure is almost optimal and
provides clean and easy way to a build process.

In fact what the current FreeBSD port do is it's setting up an OpenBLAS
Makefile.rules standard variables according to PORTOPTIONS and running a
generic OpenBLAS build pipeline after that.
It doesn't directly set up any custom compiler options nor it interfere in any
other ways with the OpenBLAS build process - and it's the way the port should
be done 'cause it allows to easily update it to a newer version and report bugs
to upstream if we'll encounter some of them.

On the other hand I'm not sure why the port fetching several archives with the
source code files and uses a GH tag instead of release archive but may be
Eijiro Shibusawa can shed some light on it.

IMO what this port is really missing is an option to enable FreeBSD CPUTYPE to
OpenBLAS TARGET variable mapping. It's presence will allow a more granular
control of the resulting optimizations especially during DYNAMIC_ARCH builds.
But it seems that it'll be easy to implement it.

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