Attention conservation notice: Superseded by an improved text.
ATLAS was my linear algebra backend on Linux machines for years. Installing it usually took a complete day and always failed at the first attempt. Probably because my notes were not as tidy as these.
OpenBLAS is supposed to be much faster than ATLAS and much easier to setup with numpy/scipy, too. At least that's what you can read around the web. Trying both methods didn't work for me though. Luckily enough, Google does Japanese and with some tiny adaptation this works. I'll describe the complete procedure for Ubuntu 12.04.
apt-get and other preparations for 12.04
sudo apt-get install git python-dev gfortran swig
sudo apt-get install rcconf dialog
The first line is clear, you probably have these installed anyway. I use rcconf to disable CPU frequency scaling (search for ONDEMAND on the linked site). 12.04 uses a new linker with some known issues. Therefore
export LIBRARY_PATH=/usr/local/lib
Note that adapting LD_LIBRARY_PATH or updating something in /etc/ld.so.conf.d (even done in the correct way) does not help.
OpenBLAS
git clone git://github.com/xianyi/OpenBLAS
wget http://www.netlib.org/blas/blast-forum/cblas.tgz
Also wget the latest LAPACK from http://www.netlib.org/lapack/. Untar both tgzs into their appropriately named directories.
Change into the cloned OpenBLAS directory and make two tiny modifications to Makefile.rule: Uncomment the lines NO_CBLAS = 1 and NO_LAPACK = 1.
Then
make FC=gfortran
sudo make PREFIX=/usr/local/ install
For building CBLAS, change into the extracted directory. In Makefile.in do
BLLIB = /usr/local/lib/libopenblas.a
CBLIB = ../lib/libcblas.a
...
LOADER = $(FC) -lpthread
...
CFLAGS = -O3 -march=native -m64 -fomit-frame-pointer -fPIC -DADD_
FFLAGS = -O3 -march=native -m64 -fomit-frame-pointer -fPIC
Adapt the last two lines to your needs, -fPIC is the important piece.
Then
make
cd lib
ar -x libcblas.a
gfortran -lopenblas -shared -o libcblas.so *.o
sudo cp libcblas.* /usr/local/lib/
If you get an error about a missing library openblas with the gfortran command then you should first check if your LIBRARY_PATH is set correctly (i.e. pointing to the place where you installed OpenBLAS).
Continuing with LAPACK, change the file make.inc (copied from make.inc.example) in the extracted directory as follows:
OPTS = -O3 -march=native -m64 -fomit-frame-pointer -fPIC
...
NOOPT = -O0 -fPIC
...
LOADOPTS = -lopenblas -lcblas
...
LAPACKLIB = liblapack.3.?.?.a # `?` depends on your version!
Again, -fPIC is the important piece. Then (subsitute the correct version for 3.?.?)
make lapacklib
mkdir tmp
cd tmp
cp ../liblapack.3.?.?.a .
ar -x liblapack.3.?.?.a
gfortran -lopenblas -lcblas -shared -o liblapack.3.?.?.so *.o
sudo cp liblapack.3.?.?.* /usr/local/lib
Finally, add the following symbolic links in /usr/local/lib:
cd /usr/local/lib
sudo ln -sn liblapack.3.?.?.a liblapack.a
sudo ln -sn liblapack.3.?.?.so liblapack.so
Ready for the python part!
Numpy
git clone https://github.com/numpy/numpy
In the numpy folder, create a site.cfg with the following two sections:
[default]
library_dirs = /usr/local/lib
include_dirs = /usr/local/include
[atlas]
atlas_libs = openblas,cblas
Check the detected configuration with python setup.py config. You should see several ATLAS_INFO paired with None. If there is something like 3.8.2 instead of None then you have an ATLAS install somewhere on your system and numpy won't use OpenBLAS. Get rid of this install (e.g. via apt-get purge) and rerun the config command. If ATLAS_INFO is completely missing then numpy couldn't find your libopenblas* libraries — you need to check all previous steps again. Now:
python setup.py build
sudo python setup.py install
Test your installation with this (Before doing a system-wide installation I prefer to do a local installation via python setup.py install --prefix=~/local/to/me. In this case
you need to adopt PYTHONPATH accordingly). On the following computer
cat /proc/cpuinfo
...
model name : Intel(R) Core(TM) i7 CPU 870 @ 2.93GHz
...
I get
FAST BLAS
version: 1.8.0.dev-3f10c36
maxint: 9223372036854775807
dot: 0.162246799469
For reference, an ATLAS-based numpy installation (via apt-get) takes about 0.26 sec, numpy without any optimized BLAS takes about 0.98 sec (on the mentioned computer). You can change the number of cores OpenBLAS utilizes via OPENBLAS_NUM_THREADS, e.g. export OPENBLAS_NUM_THREADS=2 results in
:::bash
FAST BLAS
version: 1.8.0.dev-3f10c36
maxint: 9223372036854775807
dot: 0.0949754238129 sec
Scipy
git clone https://github.com/scipy/scipy
I mentioned conspicuously often some ATLAS installation on your system, but why have one anyway, this is about OpenBLAS, after all?! The funny (or rather not so) fact is that for scipy the following is necessary:
sudo apt-get install libatlas-base-dev
cd /usr/local/lib
sudo ln -sn /usr/lib/atlas-base/liblapack_atlas.a liblapack_atlas.a
sudo ln -sn /usr/lib/atlas-base/liblapack_atlas.so liblapack_atlas.so
But only after you have a working numpy installation! Otherwise it interferes with OpenBLAS. This is a work around for a problem that I don't quite understand and that seems to bother only some people, well, and me. (To the reader: If you know of a solution, please tell me). Your site.cfg looks like that:
[default]
library_dirs = /usr/local/lib
include_dirs = /usr/local/include
[atlas]
atlas_libs = lapack_atlas,openblas,cblas
(There are no sections for umfpack. Adapt accordingly to your needs). Now
python setup.py build
sudo python setup.py install
results in a working scipy installation (i.e. python -c "import scipy.linalg" works without any segfaults.). Timings for the scipy installation can be done via this script (it has some room for improvement, though ...). My machine produces
cholesky: 0.080588388443 sec
svd: 1.13443040848 sec
Finally, if you feel good enough (and do a pip install nose), you can run nose tests via
python -c "import numpy; numpy.test(verbose=2)"
python -c "import scipy; scipy.test(verbose=2)"
(Note: I didn't!). Comments, corrections, pointers, etc. are more than welcome.