This document describes a procedure for setting up Python so that it is able to call LAPACK routines, on a Linux distribution with broken existing BLAS and LAPACK dynamic libraries, all from the point of view of a non-administrative user.

Why might you care about this document? In a nutshell: You might care about this document if you wish to dynamically load the BLAS or LAPACK libraries in Linux, and you have one of the many distributions of Linux appear to ship with broken shared BLAS and LAPACK libraries.

I use Python a lot, and I had desires I'd thought would be easy to satisfy: calling LAPACK routines from Python. On a properly configured machine, this would normally be a straightforward matter. For example, consider the following call to the BLAS function ISAMAX from Python, which will work out-of-the-box on Mac OS X 10.5 if you just paste this code into your Python interpreter:

from ctypes import *
blaslib = cdll.LoadLibrary('libblas.dylib')
array = (c_float*4)()
array[:] = [4.0, 3.14159, 4.1, 2.0]
print blaslib.isamax_(byref(c_int(len(array))), array, byref(c_int(1)))

This code will dutifully print out that 3 is the index of the maximum element. (Recall that Fortran indexes arrays from 1, not 0, unlike Python and C.) (If you want to try this on Linux, replace the dylib in the string with so.) LAPACK routines are, naturally, more complicated, but the calling convention is quite similar, except there are usually a lot more arguments.

Unfortunately, on our department Linux machines, this did not work. I've observed that many common distributions of Linux ship with broken version of BLAS and LAPACK dynamic libraries unsuitable for dynamic loading, since whomever decided to build the libraries neglected to compile with an -fPIC flag. Somehow, this mistake was repeated or else copied in many common Linux distributions. Also unfortunately, being a non-admin user of department machines means I have to build my own since the admins, for their own perfectly valid reasons, are unwilling to do changes to basic libraries outside of their general maintenance schedule.

Getting to the point of having completely broken libraries to having the ability to call LAPACK routines from Python was sufficiently non-trivial that I felt motivated to document my efforts, especially after the second and third times I had to perform the procedure.

Instructions

As an overview, these instructions follow this general plan: you will create an installation heirarchy in your user directory, and then into it you shall install Python, F2C, BLAS and LAPACK, then NumPy. For those with very peculiar needs, the installation of SVM^python is also included.

User Local

Set up your own special "user local" directory structure. This is so that you can have your own libraries, headers, etc., without administrative priviledges on the machine. Just make the directory local in your home directory.

mkdir ~/local

Then you'll want to add the ~/local/bin and ~/local/lib directories to your paths. In Bash this might look something like the following appended to ~/.bash_profile.

export PATH=$HOME/local/bin:$PATH
export LIBRARY_PATH=$HOME/local/lib:$LIBRARY_PATH
export LD_LIBRARY_PATH=$HOME/local/lib:$LD_LIBRARY_PATH
export MANPATH=$HOME/local/man:$MANPATH

Then your can source the file you changed, or just relogin, or what have you. Note that the directories do not exist yet, but should once you install Python.

Python

Download the Python source. You will probably want the most recent version rather than 2.5.1 assuming you're reading this significantly after these instructions have been written, but these instructions just install Python 2.5.1. These directions assume you are using at least Python 2.5 for the ctypes module.

wget http://www.python.org/ftp/python/2.5.1/Python-2.5.1.tgz
tar -zxf Python-2.5.1.tgz
cd Python-2.5.1

Next you want to configure, being careful to specify that you want your ~/local directory as the installation prefix.

./configure --prefix=$HOME/local

Then, just make and install Python. This should create the ~/local/[lib,include,man,bin] directories, along with some associated content, including the Python executable, libraries, includes for C-language extension modules, and other such things.

make install

Under csh or tcsh, once it is installed, rehash to gain access to that executable. Then confirm that which python points to the python executable in your binary directory (e.g., /home/yourname/local/bin/python).

F2C

First see if there is a libf2c.so in your library search path, e.g., a file exists called /usr/lib/libf2c.so in a typical configuration. If it is there, you can probably skip this step. If this is not there, then you should install it in your ~/local structure! First, download and unpack the source.

wget http://www.netlib.org/f2c/libf2c.zip
unzip libf2c.zip
cp libf2c

Then, we have to configure the package, which unfortunately must be done manually. The start of this process is to copy over the Unix Makefile template to the appropriate location.

cp makefile.u Makefile

Next, we have to do a little tweeking to some of the files so that it builds properly.

1. Edit Makefile and add -fPIC to the CFLAGS (so that CFLAGS = -O becomes CFLAGS = -O -fPIC).

   Original	Changed
   CFLAGS = -O	CFLAGS = -O -fPIC

2. We also need to provide a dummy MAIN__ method, or else links to the dynamic library will fail. So, edit main.c, and right after line 54, add the following dummy function definition: int MAIN__(void) { return 0; }. This should provide a change like this:

   Original	Changed
   #ifdef KR_headers extern VOID f_init(), sig_die(); extern int MAIN__(); #define Int /* int */ #else extern void f_init(void), sig_die(char*, int); extern int MAIN__(void); #define Int int #endif	#ifdef KR_headers extern VOID f_init(), sig_die(); extern int MAIN__(); #define Int /* int */ #else extern void f_init(void), sig_die(char*, int); extern int MAIN__(void); #define Int int #endif int MAIN__(void) { return 0; }

Finally, we can make the package. We first make and copy over the f2c.h header to the include directory, then we make and copy over the libf2c.so to the library directory.

make hadd
cp f2c.h ~/local/include
make
make libf2c.so
cp libf2c.so ~/local/lib

from ctypes import *
f2clib = cdll.LoadLibrary('libf2c.so')
f2clib.r_cos.restype, f2clib.r_sin.restype = c_double, c_double
val = c_float(0.8)
print f2clib.r_cos(byref(val))**2 + f2clib.r_sin(byref(val))**2

BLAS and LAPACK

It is now time to build the BLAS and LAPACK shared libraries. Right out of the box (more or less), these libraries shall (probably) build statically, but the shared dynamic libraries are another story. We need the shared libraries. (I am not going to use ATLAS, but I have no good reason not to, except that that's just what I got working.) Invaluable to this effort was this thread, which, while insufficient by itself, at least got me on the right track.

The first step is to acquire and unpack the source, and move into the main directory for the package.

wget http://www.netlib.org/lapack/lapack.tgz
tar -zxf lapack.tgz
cd lapack-3.1.1

The second step is to create the make.inc file. This is done by copying over the included make.inc.example, file. The make.inc file is included in many of the makefiles of the code, and is used in this source as the dumping ground for declarations common to all the build processes.

cp make.inc.example make.inc

We are going to change it so that we are using the -fPIC (position-independent code) option, so that the code may be linked into an already running process (like, hypothetically, a Python interpreter running a script that decides to load a module that links to this library we shall build ;) ). So edit make.inc, add -fPIC to both the OPTS and NOOPT variables, and just for kicks add a new MYLIBDIR = $(HOME)/local declaration. The original versus new relevant section of the file should lok like this:

FORTRAN  = g77 
OPTS     = -funroll-all-loops -O3
DRVOPTS  = $(OPTS)
NOOPT    =
LOADER   = g77
LOADOPTS =

FORTRAN  = g77 
OPTS     = -funroll-all-loops -O3 -fPIC
DRVOPTS  = $(OPTS)
NOOPT    = -fPIC
LOADER   = g77
LOADOPTS =
MYLIBDIR = $(HOME)/local

Making BLAS

Now we are going to edit BLAS/SRC/Makefile to include a target for the shared library file libblas.so. This requires changing the file in two places: adding libblas.so to the all rule, and adding a rule for the libblas.so target after the $(BLASLIB) target as seen here.

all: $(BLASLIB)

all: $(BLASLIB) libblas.so

$(BLASLIB): $(ALLOBJ)
	$(ARCH) $(ARCHFLAGS) $@ $(ALLOBJ)
	$(RANLIB) $@

$(BLASLIB): $(ALLOBJ)
	$(ARCH) $(ARCHFLAGS) $@ $(ALLOBJ)
	$(RANLIB) $@
  
libblas.so: $(ALLOBJ)
	cc -shared -Wl,-soname,$@ -lf2c -o $@ *.o
	cp $@ $(MYLIBDIR)/$@

Now, make the library.

make blaslib

If all goes well, you should have a happy, healthy libblas.so in your ~/local/lib library directory.

from ctypes import *
blaslib = cdll.LoadLibrary('libblas.so')
array = (c_float*4)()
array[:] = [4.0, 3.14159, 4.1, 2.0]
print blaslib.isamax_(byref(c_int(len(array))), array, byref(c_int(1)))

Making LAPACK

Next we shall make similar changes to the SRC/Makefile, by again changing the file in two places: add liblapack.so to the all rule, and add a rule for the liblapack.so target after the ../$(LAPACKLIB) target as seen here. Note the declaration of the SO_FILES variable.

all: ../$(LAPACKLIB)

all: ../$(LAPACKLIB) liblapack.so

../$(LAPACKLIB): $(ALLOBJ)
	$(ARCH) $(ARCHFLAGS) $@ $(ALLOBJ)
	$(RANLIB) $@

../$(LAPACKLIB): $(ALLOBJ)
	$(ARCH) $(ARCHFLAGS) $@ $(ALLOBJ)
	$(RANLIB) $@
  
SO_FILES := dlamch slamch
SO_FILES := $(patsubst %,../INSTALL/%.o,$(SO_FILES))
liblapack.so: $(ALLOBJ) 
	cc -shared -Wl,-soname,$@ -lblas -o $@ *.o $(SO_FILES)
	cp $@ $(MYLIBDIR)/$@

Now, make this library.

make lapacklib

If all goes well, liblapack.so should exist in your ~/local/lib library directory.

from ctypes import *
lapacklib = cdll.LoadLibrary('liblapack.so')
n, info = 5, c_int()
diag, offdiag = (c_double*n)(), (c_double*(n-1))()
diag[:], offdiag[:] = [2.0]*n, [-1.0]*(n-1)
r=lapacklib.dsterf_(byref(c_int(n)), diag, offdiag, byref(info))
print '[', ', '.join('%g'%v for v in diag[:]), ']'

NumPy

Next we want to get NumPy installed and running. (This is not really required. However, since installing NumPy is something I often do, I decided to include it.) However, compared to what we have done so far, this should be quite easy. First, download and unpack the most recent version of NumPy. Then run

python setup.py install

Note, if you want to just install NumPy without doing anything else in this document, and you have a broken dynamic BLAS and LAPACK library that prevents it from working, then including the following environment variable declarations should stop NumPy from attempting to link in the bad libraries, and instead build its own suboptimal-but-still-OK mini-LAPACK:

export ATLAS=None
export BLAS=None
export LAPACK=None
python setup.py install

(In csh or tcsh instead of export VARNAME=None you'd use setenv VARNAME None.)

Note that in many situations this will suffice, but NumPy uses only the most general subset of LAPACK, rather insufficient in some calculations where you know your operands have special structure that can be exploited for speed. Nonetheless, it is often useful to know how to build NumPy quickly, even under the typical broken Linux distribution.

SVM^python

A final completely optional step is to install SVM^python. Almost no one but those in my research group and I will want to do this, but of course as this document is primarily for my benefit, I'm going to describe it anyway. So there.

make
cp svm_python_{learn,classify} ~/bin

cd pyobjs
make install