Correct way to point to ATLAS/BLAS/LAPACK libraries for numpy build?

Question

I'm building numpy from source on CentOS 6.5 with no root access (python -V=2.7.6). I have the latest numpy source from git. I cannot for the life of me get numpy to acknowledge atlas libs. I have:

ls -1 /usr/lib64/atlas

libatlas.so.3
libatlas.so.3.0
libcblas.so.3
libcblas.so.3.0
libclapack.so.3
libclapack.so.3.0
libf77blas.so.3
libf77blas.so.3.0
liblapack.so.3
liblapack.so.3.0
libptcblas.so.3
libptcblas.so.3.0
libptf77blas.so.3
libptf77blas.so.3.0

I don't know anything about how these libs came about, but I can only assume that the atlas builds would be faster than any standard BLAS/LAPACK builds I could make.

What is the correct way to point numpy to these libraries? Do I export ATLAS, BLAS, LAPACK=... setting each to its corresponding path? or do I edit a site.cfg file to contain something like:

[default]
library_dirs = /usr/lib64/atlas

[atlas]
library_dirs = /usr/lib64/atlas
atlas_libs = lapack, cblas, f77blas, atlas

I've tried just about every variation of the above, and each time I run python setup.py config it tells me each library cannot be found in the paths I specify as well as a bunch of other default search paths. I've pasted the results of running python setup.py config with the site.cfg as above and no environment variables set here: http://pastebin.com/EL9CfaR7. Any help is appreciated.

Solution

Ok this was pretty simple and essentially follows the install guidelines exactly. I suppose this is more of a question on the numberings after shared libs than about numpy or atlas. Anyway, I just had to create some symlinks:

ln -s /usr/lib64/atlas/___.so.3.0 $HOME/local/lib/___.so

Then removed all configs in the site.cfg and updated my .bashrc:

export ATLAS=$HOME/local/lib/libatlas.so
export BLAS=$HOME/local/lib/libptf77blas.so
export LAPACK=$HOME/local/lib/liblapack.so

After running python setup.py install, I'm good:

>>> import numpy.distutils.system_info as si
>>> si.get_info('atlas')
    ATLAS version 3.8.4 built by mockbuild on Wed Mar 21 01:43:44 GMT 2012:
   UNAME    : Linux c6b6.bsys.dev.centos.org 2.6.32-44.2.el6.x86_64 #1 SMP Wed Jul 21 12:48:32 EDT 2010 x86_64 x86_64 x86_64 GNU/Linux
   INSTFLG  : -1 0 -a 1
   ARCHDEFS : -DATL_OS_Linux -DATL_ARCH_PII -DATL_CPUMHZ=2261 -DATL_SSE2 -DATL_SSE1 -DATL_USE64BITS -DATL_GAS_x8664
   F2CDEFS  : -DAdd_ -DF77_INTEGER=int -DStringSunStyle
   CACHEEDGE: 163840
   F77      : gfortran, version GNU Fortran (GCC) 4.4.6 20110731 (Red Hat 4.4.6-3)
   F77FLAGS : -fomit-frame-pointer -mfpmath=387 -O2 -falign-loops=4 -g -Wa,--noexecstack -fPIC -m64
   SMC      : gcc, version gcc (GCC) 4.4.6 20110731 (Red Hat 4.4.6-3)
   SMCFLAGS : -fomit-frame-pointer -mfpmath=387 -O2 -falign-loops=4 -g -Wa,--noexecstack -fPIC -m64
   SKC      : gcc, version gcc (GCC) 4.4.6 20110731 (Red Hat 4.4.6-3)
   SKCFLAGS : -fomit-frame-pointer -mfpmath=387 -O2 -falign-loops=4 -g -Wa,--noexecstack -fPIC -m64
{'libraries': ['lapack', 'f77blas', 'cblas', 'atlas'], 'library_dirs': ['~/local/lib'], 'define_macros': [('ATLAS_INFO', '"\\"3.8.4\\""')], 'language': 'f77'}

Comments

[Davidmh]

This seems awfully complicated. Have you tried a virtualenv instead? You can have your full Python stack in your home folder, so no need for root access for anything. And then you can use pip for everything.

[Matt Hancock]

I'm not familiar with virtualenv, but from glancing at it, I'm not sure if it'd be useful. I already have (well HAD--I'm running into some other issues) python/numpy/scipy installed locally. I'm not trying to install multiple versions of numpy - just telling the one version to use a particular atlas library that was (I'm assuming) optimized for the architecture I'm running on.

[Rich Homolka]

(yes, i know this is old) You need the -devel package. since you don't have root, you can't install it, but you should ask your admin to do so. The -devel package has headers and the .so symlink