[SciPy-user] compiler & lapack issues

[Pearu Peterson]

On Sun, 26 May 2002 gvermeul at polycnrs-gre.fr wrote:

> Pearu wrote:
> > 
> > > test_svd fails (I commented out almost everything except for this test, but
> > > impossible the get rid of the zswap:n=3" message). No failure if I also
> > > comment out test_svd.
> > 
> > I think zswap:n=3 comes from test_fblas.py.
> > 
> > > 
> > > I will integrate the latest patches in my lapack. Maybe a patch has been
> > > superseded by another.
> > >
> test_svd fails, with newest patches applied.
> 
> So, I tried to mask dgesdd, as you suggested. Problem is that dgessd
> is linked into _flinalg.so and flapack.so. On my system,
> the one in flapack.so is called during the test.
> 
> Attached (hackish) patch replaces dgesdd.o in both *.so's.
> So, in principle, it is possible to repair "broken" LAPACK
> installations.

Thanks for the patch. I am now able to produce this error and already
fixed it (see below). Also I have implemented hooks to use LAPACK and BLAS
libriaries if ATLAS is not available. I need to clean up things and run
the tests and then I'll commit my changes to CVS.

Fix: replace the following line at the end of subroutine gesdd in file
scipy/linalg/src/calc_lwork.f

      max_lwork = MAXWRK

with

      max_lwork = MAX(MINWRK,MAXWRK)

With this fix both official and patched LAPACK should work so that no
additional special hooks will be needed. I hope that you can confirm this.

Regards,
	Pearu