Skip site navigation (1)Skip section navigation (2) 
Date:      Sat, 17 Dec 2022 01:54:23 +0000
From:      bugzilla-noreply@freebsd.org
To:        ports-bugs@FreeBSD.org
Subject:   [Bug 268422] math/atlas alapack is not a full lapack after update to 3.10.3
Message-ID:  <bug-268422-7788@https.bugs.freebsd.org/bugzilla/>
next in thread | raw e-mail | index | archive | help 
https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=3D268422

            Bug ID: 268422
           Summary: math/atlas alapack is not a full lapack after update
                    to 3.10.3
           Product: Ports & Packages
           Version: Latest
          Hardware: Any
                OS: Any
            Status: New
          Severity: Affects Some People
          Priority: ---
         Component: Individual Port(s)
          Assignee: ports-bugs@FreeBSD.org
          Reporter: alt2600@icloud.com

Created attachment 238852
  --> https://bugs.freebsd.org/bugzilla/attachment.cgi?id=3D238852&action=
=3Dedit
git-Makefile_full_lapack.diff

So ran into some issues after Atlas update with some errors indicating miss=
ing
symbols notably from py-numpy when configuring opencv, or py-netCDF4

ImportError:
/usr/local/lib/python3.9/site-packages/numpy/linalg/_umath_linalg.cpython-3=
9.so:
Undefined symbol "sgesdd_"

so after poking around the backup package of Atlas and the current ones I
confirmed the old one did provide that symbol, and I recall that becoming
default a while ago to build full lapack.=20

So I starting digging and saw what upstream did with eliminating specifying
just the static library reference to use. I basically manually did in the
do-build section things that configure would have done if if had built its =
own
lapack.a and set it up to use it. This included passing some options to
xconfig, stripping some objects from our reference lapack and copying it in=
to
${opt}/lib where the build expects it. it appears passing flapack to xconfig
doesn't really do anything, all the business setting things up now seems to=
 be
in configure, but can be worked around.

I confirmed this now provides the library, opencv doesn't fail to import the
symbol loading numpy, all the things that use atlas build. I even re-built
py-scipy with atlas using this:
https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=3D257400 and the scikit
packages I have that use it.

I've attached my patch

--=20
You are receiving this mail because:
You are the assignee for the bug.=

Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?bug-268422-7788>