This is the mail archive of the
gsl-discuss@sourceware.org
mailing list for the GSL project.
Re: BLAS documentation
On Wed, Mar 15, 2006 at 10:46:50AM +0100, Alxneit-Kamber Ivo wrote:
> On Tue, 2006-03-14 at 23:34 -0500, Arthur K. Barnes wrote:
> > Hi all,
> >
> > Thanks for a great library. I notice that the GSL documentation for
> > the BLAS wrappers include function prototypes, but not a description
> > of what each function does. Granted, it is possible to infer
> > functionality from the function name, and there is a link to the BLAS
> > documentation, but it still is a bit of a pain. I'm willing to produce
> > a patch for a completed version, but I'm curious if there is a reason
> > for referring to the BLAS documentation.
> >
> > Thanks again,
> > -Art
> >
> i assume you refer to appendix d of the manual. you are right, only the
> prototypes are listed. for the complete documentation refer to chapter
> 12. the naming convention is gsl_blas_function <-> cblas_function. the
> only part that is not obvious are the dimensions of the various
> matrices/vectors. they are contained in the gsl_matrix or gsl_vevtor but
> have to be supplied for the wrapper.
> --
> Dr. Ivo Alxneit
> Laboratory for Solar Technology phone: +41 56 310 4092
> Paul Scherrer Institute fax: +41 56 310 2688
> CH-5232 Villigen http://solar.web.psi.ch
> Switzerland gnupg key: 0x515E30C7
>
There are actually more differences than that. Lots of useful blas functions do
not appear as gsl_blas_xxxx (gbmv, tpmv, others?). There are two tricky things
about such functions: how to store the elements of matrices in memory, and where
to put the various parameters in the call. The original blas documentation is
pretty terse, so I can understand why it might be better to use GSL's
documentation to make it better.
On the other hand, there is a newer specification of the BLAS that doesn't seem
to be very popular with vendors, but it looks to me like a good extension!
(http://www.netlib.org/blas/blast-forum/)
In constrast to the original blas standard (the new document refers to it as
legacy blas) the new standard is very *well* documented. Specifically, it
carefully documents the different storage techniques and parameter semantics for
the complicated Level2 and Level 3 functions. The catch is that it documents
the new versions of the functions. (The new ones are similar to the old ones
for the most part.)
What would list readers think of:
- linking to the blast-forum documentation as a means of explaining how to use
the legacy interface (which is implemented by GSL and others).
- actually implementing a reference library of the new standard as a companion
(extension?) to the GSL.
PS. The newer standard is not totally unsupported by tuned implementations.
ATLAS already implements [ds]set and [ds]axpby, but exports them as catlas_xxx
instead of cblas_xxxxx.
--
james bergstra
http://www-etud.iro.umontreal.ca/~bergstrj