Numpy multiarray.so: undefined symbol: cblas_sgemm

Question

I'm trying to install numpy==1.10.2, scipy and matplotlib on server based on opensuse. I've installed numpy in virtualenv from source (I've tried by pip also - same result of course). Now when I'm trying to import numpy in python console I'm receiving following error:

ImportError: /home/user/.virtualenvs/project/lib/python2.7/site-packages/numpy/core/multiarray.so: undefined symbol: cblas_sgemm

Note: I'm not superuser on this server.

Edit:

ldd /home/user/.virtualenvs/project/lib/python2.7/site-packages/numpy/core/multiarray.so`
linux-vdso.so.1 (0x00007fffa0d69000)
libtatlas.so.3 => /home/user/.local/usr/lib64/atlas/libtatlas.so.3 (0x00007fe366d66000)
libm.so.6 => /lib64/libm.so.6 (0x00007fe366a50000)
libpython2.7.so.1.0 => /usr/lib64/libpython2.7.so.1.0 (0x00007fe3666b2000)
libpthread.so.0 => /lib64/libpthread.so.0 (0x00007fe366496000)
libc.so.6 => /lib64/libc.so.6 (0x00007fe3660f0000)
/lib64/ld-linux-x86-64.so.2 (0x00007fe367a15000)
libdl.so.2 => /lib64/libdl.so.2 (0x00007fe365eec000)
libutil.so.1 => /lib64/libutil.so.1 (0x00007fe365ce9000)

libtatlas.so.3 => /home/user/.local/usr/lib64/atlas/libtatlas.so.3 - I linked this because of other problem with libatlas: ImportError: /usr/lib64/atlas/libtatlas.so.3: undefined symbol: clapack_ilaenv

SOLLUTION

The problem was with ATLAS lib. According to @ali_m advice I installed numpy from source with OpenBLAS instead of ATLAS. Here link to instruction how to install numpy with BLAS.

Answer 1

As you mentioned in the comments above, you extracted libtatlas.so.3 from an RPM package and copied it into a local directory, rather than either building it from source or installing it via the package manager. I'm sure that this is the cause of the problem.

In particular, the clapack_* and cblas_* symbols don't seem to be defined within libtatlas.so.3. On my Ubuntu machine, if I sudo apt-get install libatlas-base-dev I get two different shared libraries:

~$ ldconfig -p | grep atlas
        liblapack_atlas.so.3 (libc6,x86-64) => /usr/lib/liblapack_atlas.so.3
        liblapack_atlas.so (libc6,x86-64) => /usr/lib/liblapack_atlas.so
        libatlas.so.3 (libc6,x86-64) => /usr/lib/libatlas.so.3
        libatlas.so (libc6,x86-64) => /usr/lib/libatlas.so

The clapack_* and cgemm_* symbols are defined in liblapack_atlas.so.3 rather than libatlas_so.3:

~$ nm -D /usr/lib/libatlas.so.3 | grep clapack_ilaenv

~$ nm -D /usr/lib/libatlas.so.3 | grep cblas_sgemm

~$ nm -D /usr/lib/liblapack_atlas.so.3 | grep clapack_ilaenv
0000000000041d90 T clapack_ilaenv

~$ nm -D /usr/lib/liblapack_atlas.so.3 | grep cblas_sgemm
                 U cblas_sgemm

Given that you have no admin rights, and therefore can't install ATLAS via the normal system package manager, here are three potential solutions:

• It might be possible to extract a binary of liblapack_atlas.so.3 (or whatever the OpenSUSE equivalent is) from a different RPM package, and get numpy to link against it by modifying the site.cfg file in the numpy source directory.
• Another option would be to build ATLAS from source, but in my experience this is a long and painful process.
• My personal recommendation would be to build OpenBLAS instead, as described in my previous answer here. It's much easier to compile, and has better performance in every benchmark that I've come across so far.