Building Scipy/Numpy on Windows with Optimized Numerical Libraries¶
This involves compiling several libraries (ATLAS, LAPACK, FFTW and UMFPACK) and then building numpy and scipy from SVN source. But as with most things Windows, this turns out to be a slightly tricky affair.
The following has been tested on Windows Vista Enterprise 32bit only, but should theoretically work on other Windows platforms. It also used Python 2.5.
Ideally, a big chunk of this page should move to the scipy/numpy site. And also ideally should become a single script. But it’s also good to know exactly how you got there.
Prerequisites¶
You need Windows Vista enterprise/ultimate with SUA enabled and installed or Windows (others, including other Vista variants) with Cygwin installed. You cannot install the SUA package on a non enterprise or ultimate Vista edition.
MinGW (installer) with gcc 3.4.5 (choose the candidate option when installing) and the msys environment installed. You will need to download the following packages for msys:
bzip2-1.0.3-MSYS-1.0.11-snapshot.tar.bz2
coreutils-5.97-MSYS-1.0.11-snapshot.tar.bz2
diffutils-2.8.7-MSYS-1.0.11-snapshot.tar.bz2
gawk-3.1.5-MSYS-1.0.11-snapshot.tar.bz2
make-3.81-MSYS-1.0.11-snapshot.tar.bz2
msysCORE-1.0.11-2007.01.19-1.tar.bz2
binutils-2.17.50-20070129-1.tar.gz
Just unpack all the package contents in a single directory and copy them over to the MinGW installation directory. You may want to add the following to the system path:
set PATH=[PATH TO]\MinGW;[PATH TO]\MinGW\libexec\gcc\mingw32\3.4.5;%PATH%
- Numerical Libraries
Installation¶
Create a directory called BUILDS, BUILDS/lib, BUILDS/include
Unpack all the numerical library files in BUILDS
Create subversion check out directories for scipy and numpy in BUILDS
Start SUA c-shell or cygwin shell
Start msys.bat:
PATH=/mingw/libexec/gcc/mingw32/3.4.5:$PATH; export PATH
Change directory to location of BUILDS. (/dev/fs/driveletter/… in SUA, /cygdrive/driveletter/… in cygwin, /driveletter/… in msys)
Compiling ATLAS¶
This is done in the SUA/Cygwin shell. In Cygwin you probably want to follow the instructions at Installing Scipy on Windows
cd ATLAS; mkdir build; cd build
Run ../configure (This will probably fail but will leave you with xconfig)
Run ./xconfig –help (to see all options)
Run ../configure -O 8 -A 16 -m 3189 -b 32 (replacing the values with your machine configuration)
Edit Make.inc to provide correct L2SIZE
Run make (leave your computer and go do something else for about an hour)
Compiling LAPACK¶
This is done in the msys shell
cd lapack_XX
Copy make.inc.example to make.inc
Edit the following lines in make.inc:
PLAT = _NT OPTS = -funroll-all-loops -O3 -malign-double -msse2 BLASLIB = -L/driveletter/[PATH TO]/BUILDS/ATLAS/build/lib -lf77blas -latlas
Run make lib
Combining LAPACK and ATLAS¶
Stay in the msys shell after compiling LAPACK
Go to the ATLAS/build/lib directory
Execute the following commands:
mkdir tmp; cd tmp cp ../liblapack.a ../liblapack_ATLAS.a ar -x ../liblapack.a cp [PATH TO]/lapack_NT.a ../liblapack.a ar -r ../liblapack.a *.o rm *.o ar -x ../liblapack.a xerbla.o ar -r ../libf77blas.a xerbla.o
Copy liblapack.a, libf77blas.a, libcblas.a, libatlas.a to BUILDS/lib
Copy the ATLAS/include to BUILDS/include/ATLAS
Compiling UMFPACK¶
Stay in msys shell
Goto UFconfig
Edit UFConfig/UFconfig.mk:
BLAS = -L/driveletter/[PATH TO]/BUILDS/lib -llapack -lf77blas -lcblas -latlas -lg2c LAPACK = -L/driveletter/[PATH TO]/BUILDS/lib -llapack -lf77blas -lcblas -latlas -lg2c XERBLA =
Run the following commands:
cd ..\AMD make cd ..\UMFPACK make
Copy libamd.a (from AMD), libumfpack.a (from UMFPACK) to BUILDS/lib
Copy UMFPACK/include to BUILDS/include/UMFPACK
Copy UFconfig/ufconfig.h to BUILDS/include
Copy AMD/include/amd.h to BUILDS/include
Compiling fftw¶
Note
The latest versions of scipy do not link to FFTW, so this step is no longer useful for scipy
Stay in msys shell
Goto fftw_XX
mkdir build; cd build
Run the following command:
../configure --prefix=/c/DOWNLOADS/BUILDS/ --enable-sse2 --disable-dependency-tracking --enable-threads --with-our-malloc16 --with-windows-f77-mangling --with-combined-threads
Run make OR make -j 4 if you have multiple processors (it’ll make things go faster. This build on msys in vista takes a while)
Copy .libs/libfftw3.a to BUILDS/lib
Copy fftw_XX/api/fftw3.h to BUILDS/include
Compiling numpy/scipy¶
Note
As above, note that the FFTW linking here is no longer useful for the scipy install
Open a Windows cmd window and make sure you can execute python.
Make a copy of each of the libs in BUILDS/lib and rename them from libname.a to name.lib
Rename lapack.lib to flapack.lib
rename site.cfg.example to site.cfg
Edit site.cfg in the numpy directory. Replace the blas_opt and lapack_opt section with:
[atlas] libraries = f77blas, cblas, atlas, g2c library_dirs = driveletter:\[PATH TO]\MinGW\lib;driveletter:\[PATH TO]\BUILDS\lib include_dirs = driveletter:\[PATH TO]\BUILDS\include\ATLAS [lapack] libraries = flapack, f77blas, cblas, atlas library_dirs = driveletter:\[PATH TO]\MinGW\lib;driveletter:\[PATH TO]\BUILDS\lib [amd] library_dirs = driveletter:\[PATH TO]\MinGW\lib;driveletter:\[PATH TO]\BUILDS\lib include_dirs = driveletter:\[PATH TO]\BUILDS\include libraries = amd [umfpack] library_dirs = driveletter:\[PATH TO]\MinGW\lib;driveletter:\[PATH TO]\BUILDS\lib include_dirs = driveletter:\[PATH TO]\BUILDS\include\UMFPACK libraries = umfpack [fftw3] library_dirs = driveletter:\[PATH TO]\MinGW\lib;driveletter:\[PATH TO]\BUILDS\lib include_dirs = driveletter:\[PATH TO]\BUILDS\include libraries = fftw3
Edit numpy/distutils/fcompiler/gnu.py. Find the line that says opt.append(‘gcc’) and comment it # opt.append(‘gcc’). This is probably a Vista SUA thing and perhaps won’t be required when using Cygwin to compile ATLAS.
Copy site.cfg to ../scipy/site.cfg
Compile numpy:
cd numpy python setup.py config --compiler=mingw32 build --compiler=mingw32 bdist_wininst
Install numpy from the numpy/dist folder
Compile scipy:
cd scipy python setup.py config --compiler=mingw32 build --compiler=mingw32 bdist_wininst
Install scipy from the scipy/dist folder
Test installations. In python run:
import numpy import scipy numpy.test() scipy.test() numpy.show_config() scipy.show_config()