Re: 2.9.6 build fails on an Opteron RHEL4

[Javier Fernández Baldomero]

> > > | I use K. Goto's BLAS (and a LAPACK compiled against  the GOTO BLAS lib).
> > >
> > > So, is it compiled in a way that ensures that the integer quantities
> > > are 8 byte signed integers?
> > >
> > >      

I am supervising a class assignment comparing Octave performance with 
and without ATLAS and Goto's BLAS. There were/are lots of things I was 
not aware of. For instance, ATLAS includes a BLAS _and_ a LAPACK. The 
student found a nice benchmark from sciviews.org, more detailed (and 
adaptable to Octave) than the one available from Mathworks' File Exchange.

There is some weird result that GotoBlas + LAPACK is worse than GotoBLAS 
alone. A previous weirdness that GotoBLAS wasn't better than ATLAS BLAS 
was due to not using LD_PRELOAD before running Octave. I was suggesting 
LD_LIBRARY_PATH to the student, without any results. Thanks to David for 
his e-mail from 2003.
http://velveeta.che.wisc.edu/octave/lists/help-octave/2003/754
I would have suggested that same benchmark to the student if I had 
remembered that e-mail earlier.

I'm not sure why the LD_LIBRARY_PATH didn't make the trick. In fact, 
right now I'm a bit messed up with shared libraries concepts. The 
symbols defined in the GotoBLAS library are not exactly the same as in 
ATLAS BLAS (there are underscore_letter variants). Anyways, from reading 
this thread I deduce I should instruct the student to link his ATLAS 
LAPACK against the improved GotoBLAS, isn't it? How would he do it? Does 
it really depend on link-time flags? Or does it depend on which library 
is loaded at run-time?

In other words: as long as you build your libraries as shared objects, 
it really doesn't matter which one you compile Octave against. It only 
matters which one is found at load/runtime. Is that true?
If so, it doesn't matter either, which BLAS do you link LAPACK against. 
It only matters which BLAS is preloaded when you load LAPACK (ok, when 
Octave loads it).

If LAPACK depends on BLAS, precisely on ATLAS BLAS, that could be the 
explanation for the weird student result that GotoBLAS+LAPACK is worse 
than GotoBLAS alone (if loading ATLAS LAPACK overrides the GotoBLAS 
preload). Is there any good reference on these issues? Is there any 
other likely explanation?

Thanks for any help, sincerely confused :-)

-javier