cvs commit: ports/math/atlas Makefile

[Brooks Davis]

On Wed, Sep 07, 2005 at 11:35:36AM +0900, NAKATA Maho wrote:
> In Message-ID: <200509070152.j871qL11077549 at repoman.freebsd.org> 
> Kris Kennaway <kris at FreeBSD.org> wrote:
> 
> > kris        2005-09-07 01:52:21 UTC
> > 
> >   FreeBSD ports repository
> > 
> >   Modified files:
> >     math/atlas           Makefile 
> >   Log:
> >   Set NO_PACKAGE: atlas optimizes for the local machine, so a package is not
> >   generally useful.
> >   
> >   Reviewed by:    maho (maintainer)
> >   
> >   Revision  Changes    Path
> >   1.35      +2 -0      ports/math/atlas/Makefile
> 
> As of 5.4-RELEASE, build of atlas became extremely
> fragile, namely; type make 20 times or so to build.
> apparently something has been changed between 5.4 and 5.3.
> 
> I suspect this was due to:
> o timing are not correctly calculated, errors become larger?
> o I don't use SCHD_ULE.
> 
> do you know something about it?

Weird.  I sucessfully built it in one go on a 2x 2GHz Xeon with a
7-CURRRENT kernel in a ~6.0-BETA2 chroot.

I'm on our cluster, I'm looking in to having per hardware configuration
libmap.conf files so users will link against a version optimized for the
head node and then pick up the right version on each node.  If that
works out it might be useful to have packages some day (even ones that
deliberatly disable much of autotuning) just to let me install something
to link against.

-- Brooks

-- 
Any statement of the form "X is the one, true Y" is FALSE.
PGP fingerprint 655D 519C 26A7 82E7 2529  9BF0 5D8E 8BE9 F238 1AD4
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 189 bytes
Desc: not available
Url : http://lists.freebsd.org/pipermail/cvs-all/attachments/20050906/fd15b8e4/attachment.bin