[SciPy-user] installation on Suse8.1 (blas problem)
Pearu Peterson
pearu at cens.ioc.ee
Fri Mar 14 02:41:29 EST 2003
On Fri, 14 Mar 2003 baecker at physik.tu-dresden.de wrote:
> I am trying to install scipy (CVS version) under
> SUSE linux.
> Unfortunately, it seems the "delivered" blas is one of those
> incomplete ones which don't have srotmg_ .
>
>
> As the whole installation is supposed to work
> for a set of Intel PIII machines and a set of more recent Athlons
> I don't think that building ATLAS would be an optimal solution.
> Do you have any recommendation in this situation ?
> (Should I build lapack and blas from scratch ?)
Using ATLAS instead of blas/lapack gives certainly a huge execution speed
advantage. However, I you wish not to use ATLAS, then rebuilding blas (no
need to rebuild lapack) would solve the problem. Another solution is
to look at linalg/setup_linalg.py and append missing symbols like srotmg
to
skip_names['cblas']
skip_names['fblas']
lists in appropiate places (should be clear from the code).
Whenever making changes to skip_names, you should manually remove
fblas.pyf and cblas.pyf files before building.
> (Another thing is, that python scipy/scipy_core/system_info.py
> does not detect the fftw routines (I hope they are the
> right ones, but I could not test this yet) which
> are found under the names:
> /usr/lib/libdfftw.so.2
> /usr/lib/libdfftw.so.2.0.5
> /usr/lib/libdrfftw.so.2
> /usr/lib/libdrfftw.so.2.0.5
> /usr/lib/libsfftw.so.2
> /usr/lib/libsfftw.so.2.0.5
> /usr/lib/libsrfftw.so.2
> /usr/lib/libsrfftw.so.2.0.5
>
> Can one just change the names in site.cfg accordingly
> to get them detected ?
Not recommended.
You should install libfftw-dev package (on suse libfftw-dev may
have different name, though). Note that building fftw wrapper requires
also fftw header files.
> ((One more thing: it might be nice if system_info.py
> reports which of the possible alternatives
> (atlas, lapack/blas etc.) will be used in the end.))
Actually, system_info does not decide what will be used at the end,
setup_linalg.py makes that choice. So, you can see which resources will be
used from the output of building scipy.
The following rule applies:
If atlas libraries are available then atlas is used;
else if lapack/blas libraries are available then they are used;
else if lapack/blas sources are available then they are used;
else not found error exception is raised.
Pearu
More information about the SciPy-User
mailing list