[SciPy-dev] Compiling scipy with Intel ifort & MKL

rex rex@nosyntax....
Tue Jun 12 14:02:07 CDT 2007


Robert Kern <robert.kern@gmail.com> [2007-06-12 11:24]:
> Pearu Peterson wrote:
> 
> > Are you using numpy 1.0.3 tar-ball? There are at least two error reports
> > of the given kind (building scipy fails with `dfftpack/*.f` not existing 
> > error) that was caused by this numpy.distutils bug.
> 
> Ah, I am using the first 1.0.3 tarball, not the one with r3845 in it. Yes, I
> think we should release a 1.0.3.1 that fixes that issue and the problem that
> Travis had that caused him to do that in the first place. We should branch it
> from the tag rather than the trunk because of David Cooke's merge.

In the current numpy svn there are at least two distutil problems:
In  system_info.py, the line:

lapack_libs = self.get_libs('lapack_libs',['mkl_lapack32','mkl_lapack64'])

will not work with MKL 9.1 because the '32' and '64' have been
removed. The line below works for me (but of course it won't work for
the old MKL 8.1, so there probably should be a version check).

lapack_libs = self.get_libs('lapack_libs',['mkl_lapack'])


In fcompiler/intel.py, the line(s) 

'version_cmd'  : ['<F77>', None],

causes the scipy build to fail when ifort 10.0 (and, ifort 9.1, I think)
is used. Following George Nurser's lead, I changed all instances to:

'version_cmd'  : ['<F77>', '-V'],

and now scipy svn builds. ('-V' may not be correct for the visual
compilers)

The problem of scipy ignoring 

cc_exe = 'icc ...

in intelccompiler.py and using gcc remains. 

-rex


More information about the Scipy-dev mailing list