[SciPy-dev] Scipy Compile and Testing Problem from SVN

Skipper Seabold jsseabold@gmail....
Sat Apr 11 13:17:06 CDT 2009

On Sat, Apr 11, 2009 at 12:37 AM, Charles R Harris
<charlesr.harris@gmail.com> wrote:
> On Fri, Apr 10, 2009 at 9:21 PM, <josef.pktd@gmail.com> wrote:
>> On Fri, Apr 10, 2009 at 9:42 PM, Skipper Seabold <jsseabold@gmail.com>
>> wrote:
>> > Hello all,
>> >
>> > I am having some trouble compiling the development versions of Scipy
>> > and Numpy from SVN and then getting the tests to run without errors.
>> > >From the errors I've seen and having combed the archives, I wonder if
>> > it's a compiler issue (gnu95 4.3.2).  Or most likely some other user
>> > error...  So far I have tried to avoid installing ATLAS and LAPACK
>> > from source.
>> >
>> > I am using openSUSE 11.1 (32-bit), Python 2.6, and the most recent
>> > version of scipy and numpy from SVN.  I have installed the following
>> > all from the openSUSE repo openSUSE_Education
>> >
>> > libatlas3
>> > libatlas3-devel
>> > blas
>> > libblas3
>> > libumfpack5_2_0
>> > umfpack-devel
>> > lapack
>> > liblapack3
>> > fftw
>> > fftw-devel
>> > fftw3
>> > fftw3-devel
>> > fftw3-threads
>> > amd
>> > amd-devel
>> > libamd2_2_0
>> >
>> > For both Numpy and Scipy, I edited the site.cfg as I discovered was
>> > necessary after a first go around.
>> >
>> > [DEFAULT]
>> > library_dirs = /usr/local/lib:/usr/lib
>> > include_dirs = /usr/local/include:/usr/include:/usr/include/suitesparse
>> >
>> > And uncommented blas, lapack (not the pthreads), amd, umfpack, fftw
>> >
> <snip>
>> If I interpret Davids usually explanations and your logs correctly
>> then you are mixing fortran compilers, lapack is using gfortran and
>> you are using g95, which showed up pretty often on the mailing lists.
>> But this is just pattern recognition, I don't know anything about
>> opensuse.
> IIRC, there was some problem with opensuse reported on the mailing list.
> <searches> hmm, there was a 32/64 bit issue about two weeks ago. And there
> are opensuse repositories with current numpy and scipy here.
> Chuck

Thanks for the pointers.  I've seen discussion of similar problems in
the archives and the build instructions on the site, so I thought it
was a compiler issue.  I still had the same errors when building scipy
with "~>python setup.py build=gfortran" (gfortran as part of gcc 4.3
on opensuse at least), so I wondered if I could get it to work after
compiling BLAS, LAPACK, and ATLAS from source explicitly using
gfortran then installing numpy and scipy explicitly with gfortran.
Still getting the same error messages.

First before I get into the last few things that I can think of, sorry
for all the questions.  If the next few steps don't work, I am open to
switching Linux distributions and seeing if I have more luck because I
am about at the end of my abilities (and time constraint).  I have
seen that Debian-based distributions are usually better for Scipy and
the extensions.  Is there a particular "flavor" that is popular for
Scipy developers?

Ok, last few thoughts.

Is it still true that gcc 3.x compilers are recommended?  Wondering
how old that information is.

Could someone explain this note in the installation instructions found
here: <http://www.scipy.org/Installing_SciPy/BuildingGeneral#head-afafb64f8d0edbe09e7b85c43c08cd5bd5904350>

"NOTE: scipy may not find the libf* names.  You may have to make a
symbolic link from these files to libblas.a and liblapack.a  Numpy
does not seem to have this problem."

I do not have a plain libblas.a after building BLAS, LAPACK, and
ATLAS, only libfblas.a from the BLAS build and libf77blas.a and
libcblas.a from the ATLAS build.

Should I create a symbolic link from libflapack.a to the liblapack.a
that I have after I installed ATLAS following these instructions:

Do I just run
~> ln -s /usr/bin/libflapack.a /usr/bin/liblapack.a

I also wonder if I should have included the compiler flag
"-fno-second-underscore" to make when building LAPACK and ATLAS as is
instructed to do for BLAS.

Thanks for any help,


More information about the Scipy-dev mailing list