[Numpy-discussion] Installing numpy with MKL

Søren Gammelmark gammelmark@phys.au...
Tue Aug 3 09:55:12 CDT 2010


Hi everyone

I realize that this e-mail has been written a lot of times before. I 
have, however, not been able to make it work even though I have looked 
through a lot of e-mails on the mailinglist. I apologize if the solution 
is written neatly somewhere.

As many others I would like to install numpy with MKL support on a 
cluster. The cluster contains dual core AMD Opteron machines. The reason 
for doing this is to enable me to perform multi-threaded 
matrix-operations (matrix-multiplications and singular value 
decompositions primarily). Additionally I need to integrate some Fortran 
code also using Intel MKL into numpy using f2py. I think I need to use 
MKL for this, but correct me if I am wrong.

I have tried to follow the guide at 
http://www.scipy.org/Installing_SciPy/Linux#head-7ce43956a69ec51c6f2cedd894a4715d5bfff974. 
I have compiled my own python and installed it locally using the 
LD_RUN_PATH suggested there using gcc (I only have access to the intel 
9.1 compilers which seems to have a bug preventing me from compiling 
python with it (see http://bugs.python.org/issue1689617)). This works 
without any problems.

I have access to Intel MKL 10.2.5.035 and have the following site.cfg 
for numpy 1.4.1:

[DEFAULT]
library_dirs = /home/sgamark/usr/lib        # for python is installed 
with prefix $HOME/usr
include_dirs = /home/sgamark/usr/include

[mkl]
library_dirs = /com/intel/mkl/10.2.5.035/lib/em64t
include_dirs = /com/intel/mkl/10.2.5.035/include
lapack_libs = mkl_lapack
mkl_libs = mkl_intel_lp64, mkl_intel_thread, mkl_core, iomp5, pthread, 
mkl_def

and have changes numpy/distutils/intelcompiler.py according to 
http://www.scipy.org/Installing_SciPy/Linux#head-7ce43956a69ec51c6f2cedd894a4715d5bfff974 
but with cc_exe = 'icc -fPIC -openmp -fomit-frame-pointer -O2 -g'. I 
have include openmp since I believe it is necessary for using the 
threading features of MKL (correct me if I am wrong).

I have also tried modifiying numpy/distutils/system_info.py as described 
in 
http://article.gmane.org/gmane.comp.python.numeric.general/18877/match=intel+mkl 
but without any effect.

However, executing "python setup.py config" shows that the setup-script 
is unable to find any of the libraries I have (see attached file). If I 
try to compile anyway MKL is not used and setting OMP_NUM_THREADS have 
no effect.

At this point I am a bit frustrated, since it seems that it should not 
be too difficult to get to work since it appears that other people have 
no problems. Indeed, several people I work with have no problem doing 
this with intel 10.1 compilers, MKL 11.1 on Intel Itanium machines. I am 
not particularly experienced with this sort of thing but I usually have 
few problems compiling a stand-alone Fortran or C/C++ program that uses MKL.

I would greatly appreciate it if anyone could help with this.
Regards
Søren Gammelmark

P.s: The operating system is Linux/CentOS 4.4, x86_64


-------------- next part --------------
An embedded and charset-unspecified text was scrubbed...
Name: setup_conf
Url: http://mail.scipy.org/pipermail/numpy-discussion/attachments/20100803/77c4d10a/attachment.pl 


More information about the NumPy-Discussion mailing list