Bernard,
I've added you to the dev list, and CCed it since you seem have some
questions you wanted input on . . .
> In principle I can (and do) build ATLAS first, build LAPACK using
>the options stolen from Make.inc as generated by ATLAS, patch in
>the ATLAS partial LAPACK library, and re-link. Still, ATLAS
>is documented as doing the job going the other way (though I dare
>say it will only generate a static library - I can do better!), and I was
>annoyed to find it not working.
>
> When I add --with-netlib-lapack=<xyz> to the ATLAS
>configuration arguments, the build breaks. I haven't yet followed the
>whole process down in detail, but 'make' tells me that libf77blas.a is
>required, but there is no rule to build it. Remove --with-netlib-lapack
>and it all starts working again. I'ts all the odder because ( unless I've
>misunderstood something) the LAPACK installation isn't relevant to
>the build until *after* the ATLAS libraries (or, at least, their
>contributory object modules ) have been built.
I believe all the fancy lapack options that are available in 3.9.x work fine
with lapack3.1.1 (I use them every day). Unfortunately, they changed a lot
of things with LAPACK 3.2, and so the fancy options don't work there yet. I
will be rewriting lapack handling completely when I get the chance to fix this,
but that may be a while. Essentially, I plan on rewriting it so that
ATLAS just spews out a make.inc, and then uses LAPACK's own Makefiles for
the build of lapack, which ATLAS then takes as a starting point to its own
build; I hope this approach will insulate me from further changes in LAPACK
(except, of course, if they change the make.inc!). I'm not sure if this
will break assumptions I've made in the present build or not, so this
will require some time for me to fix. At that time, I will probably
eliminate my current favorite way of building, -Ss lasrc.
So, when building LAPACK with ATLAS 3.9.x, you have two choices:
(1) Use the 3.8 manual way of getting LAPACK 3.2 working
(2) Use one of the convenient method with LAPACK 3.1.1
Let me know if these two options don't work for you.
> Human life is short, and I've not finished tracing the build process to
>identify the difference, Still, does someone out there have
>mileage that differs? I might even find some spare time, if necessary,
>to make it all work... though building shared libraries is still likely
>to require surgery to the LAPACK build process.
The basic trick is to build LAPACK with -fPIC *first*, and then I think
ATLAS *might* handle the shared building for you (though others who use
shared objects might know better) if you use the old --with-netlib-lapack=.
Cheers,
Clint
**************************************************************************
** R. Clint Whaley, PhD ** Assist Prof, UTSA ** www.cs.utsa.edu/~whaley **
**************************************************************************
------------------------------------------------------------------------------
Open Source Business Conference (OSBC), March 24-25, 2009, San Francisco, CA
-OSBC tackles the biggest issue in open source: Open Sourcing the Enterprise
-Strategies to boost innovation and cut costs with open source participation
-Receive a $600 discount off the registration fee with the source code: SFAD
http://p.sf.net/sfu/XcvMzF8H_______________________________________________
Math-atlas-devel mailing list
Math-atlas-devel@...
https://lists.sourceforge.net/lists/listinfo/math-atlas-devel
