[SciPy-dev] scipy distutils questions

[Arnd Baecker]

On Wed, 14 Dec 2005, Pearu Peterson wrote:

> On Wed, 14 Dec 2005, Arnd Baecker wrote:
>
> > I got some further info on  this:
> >
> > The lapack stuff is provided in
> >  libmkl_lapack64.so  and  libmkl.so
> > which are in /opt/intel/mkl72/lib/64
> > To link correctly, they need
> >  libifcore.so   libguide.so
> > which are in /opt/intel/fc_90/lib/
> > and
> >  libpthread.so
> > which is in /usr/lib/
> >
> > We tried the following, but without success
> >
> > [atlas]
> > library_dirs = /opt/intel/mkl72/lib/64:/opt/intel/fc_90/lib/
> > include_dirs = /opt/intel/mkl72/include/
> > atlas_libs = mkl_lapack64,mkl,ifcore
>
> First, mkl is NOT atlas.
>
> mkl: http://www.intel.com/cd/software/products/asmo-na/eng/perflib/mkl/index.htm
> atlas: http://math-atlas.sourceforge.net/
>
> system_info assumes that atlas libraries are used
> according to official instructions given in
> http://math-atlas.sourceforge.net/errata.html#LINK
>
> mkl and atlas are different libraries of optimized blas/lapack/..
> libraries. So the problem boils down to adding mkl support to system_info.
> Btw, then also fftpack, ufunc's, random could take advantage of mkl some
> day..

This would indeed be fantastic!
(Similarly, this also applies to the AMD Core Math Library ACML,
which also provides BLAS/LAPACK, FFT and
"Fast scalar, vector, and array math transcendental library routines
optimized for high performance on AMD Opteron processors.",
http://developer.amd.com/acml.aspx
)
If one can believe the descriptions there, then this can
make quite a difference ...

Best - and thanks for all your effort -

Arnd