[SciPy-user] Building 64bit scipy with ATLAS/LAPACK/UMFPACK in Fedora core 8

David Cournapeau david@ar.media.kyoto-u.ac...
Fri Jul 24 04:15:28 CDT 2009

Antonino Ingargiola wrote:
> Hi to the list,
> I have fedora core 8 64bit in an 8 core server. I've installed the
> latest numpy/scipy from sources but using Sfepy[1] I found that the
> tests run very slowly.
> Here it is the procedure I used to install numpy and scipy. Firstly I installed:
>     lapack.x86_64
>     lapack-devel.x86_64
>     atlas.x86_64
>     atlas-devel.x86_64
>     suitesparse.x86_64
>     suitesparse-devel.x86_64
> then I downloaded numpy 1.3.0, edited site.cfg adding the following lines:
>     [umfpack]
>     library_dirs=/usr/lib64
>     include_dirs = /usr/include/suitesparse
> and built it with
>     python setup.py build
>     python setup.py install --prefix=/my/dir
> After that I built scipy in exactly the same way (including site.cfg editing).
> For umfpack I installed the scikits using "python setup.py install
> --prefix=/my/dir"
> I can I check if this scipy is correctly using the fast
> atlas/lapack/umfpack libraries?

You can check as follows:
    >>> import scipy
    >>> scipy.show_config()
    >>> import scipy.linalg.atlas_version

To check the speed, you can just use numpy.dot. Basic blas vs. optimized
ATLAS can be one order of magnitude different for large matrices (say
1000x1000, for example).

> Are there any pitfalls on fedora core 8 regarding the numerical
> libraries? I've read that some libraries are broken...

Generally, problems are with the packaged versions - ATLAS in particular
is very hard to package correctly. If you build them by yourself
correctly, there should be no problem. Just make sure you use the last
released ATLAS (3.8.3 I believe at this time), and avoid LAPACK 3.2.0
(i.e. use 3.1.1).



More information about the SciPy-User mailing list