[SciPy-dev] Absoft & Scipy

Stephen Walton stephen.walton at csun.edu
Wed Feb 2 18:06:26 CST 2005

Serendipity:  I thought Nils's bsp5.py would make a nice little test of 
my Scipy installation, built with Fernando's tools.  I immediately found 
that I'd built LAPACK with Absoft Fortran but Scipy with g77.  In 
attempting to fix that, I found that the Absoft compiler support is 
slightly broken, and also that INSTALL.txt is incorrect. I know I should 
do a formal diff, but the changes are so slight:

INSTALL.txt:  the information under COMPILER NOTES is out of date and 
should be replaced with the corresponding information on the 
installation Web page.

absoftcompiler.py:  line 66 should change to "opt.extend(['g2c', 
'f90math', 'fio', 'f77math'])"

spblas.f.src: line 379 needs a declaration of 'ra' as integer, as it is 
used as an array index.  g77 will let you get away with the code as is, 
but Absoft won't.

I also found, unfortunately, that bdist_rpm doesn't respect the compiler 
setting.  The following command

python setup.py config_fc --fcompiler=absoft bdist_rpm --release=P4SSE2 

fails to use the Absoft compiler;  it still uses g77.  Replacing the 
"bdist_rpm..." with "build" works fine.

Finally, even after a successful scipy build against Absoft, I'm still 
getting an error and would appreciate a clue where to look in tracking 
it down:

In [1]: linalg.eigvals(rand(5,5))
/usr/lib/python2.3/site-packages/scipy/linalg/_flinalg.so: undefined 
symbol: ddet_c_
Out[1]: NumPy array, format: long
[ 2.90519028+0.j         -0.42248621+0.j         -0.01764924+0.50381647j
      -0.01764924-0.50381647j  0.40434219+0.j        ]

More information about the Scipy-dev mailing list