Opinions on packaging ATLAS (for the x86 architecture)

[Christoph Frieben]

2009/9/25 Chris Adams:
>
> Having different packages (that may conflict with each other) means you
> can't easily switch between versions, and if you chose the wrong version
> (or move a hard drive to a different box), your programs crash
> unexpectedly.


That's exactly why

     https://bugzilla.redhat.com/show_bug.cgi?id=510498

had been reported. The default atlas.i586 package on F11 had been built with
SSE support and thus crashed on a generic P6 which was supposed to
(over)-fulfill the minimum requirements for F11 (and probably those of F12,
too)
It could not even be removed from the system although blas/lapack were
installed as well, because packages like octave actually require it! Thus it
is not simply an optional performance optimized package installed by people
doing numerics in order to supersede the standard blas/lapack packages. I
consider it thus intelligible to use a hierarchy

  atlas < atlas-sse < atlas-sse2

the first one being installed by default and and not using SSEx at all. Also
note, that otherwise, one could argue that all x86 packages (including
blas/lapack) should be built using SSEx as well, but this discussion had
already taken place on this list a few months ago .. .

(Likewise, the default x86_64 package is currently called

  atlas [ < atlas-sse3 < ... ]

and is using SSE2 by default as expected for all x86_64 packages. Higher
optimized versions enabling SSE3, SSE4.x which do not exist yet would follow
the same naming scheme.).

~C
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://lists.fedoraproject.org/pipermail/devel/attachments/20090925/301714d6/attachment.html