[SciPy-dev] Intel MKL without Intel Fortran??!!

[Norbert Nemec]

Hi there,

on my system, I have the Intel Fortran compiler installed along with
Intel MKL, as well as the g77 and gfortran 4.0 and Atlas.

The default configuration in this setup seems to be somewhat strange:
the gnu compiler is preferred over the intel compiler, but still, the
intel MKL is preferred over atlas. This means, that the intel MKL is
linked in with GNU compiled libraries. If I'm not mistaken, such a setup
is doomed to fail, because the intel MKL only works with Intel compiled
programs and libraries. (Correct me if I'm wrong.)

What I see is:

-------------------------------------------
In [1]: import numpy

In [2]: numpy.__version__
Out[2]: '0.9.2.1828'

In [2]: numpy.test()
  Found 3 tests for numpy.distutils.misc_util
  Found 2 tests for numpy.core.umath
  Found 42 tests for numpy.lib.type_check
  Found 9 tests for numpy.lib.twodim_base
  Found 3 tests for numpy.dft.helper
  Found 20 tests for numpy.core.ma
  Found 3 tests for numpy.lib.getlimits
  Found 6 tests for numpy.core.defmatrix
  Found 33 tests for numpy.lib.function_base
  Found 6 tests for numpy.core.records
  Found 4 tests for numpy.lib.index_tricks
  Found 44 tests for numpy.lib.shape_base
  Found 0 tests for __main__
..................................................................................MKL
FATAL ERROR: /usr/local/lib/libmkl_def.so: undefined symbol:
__kmpc_global_thread_num
-------------------------------------------

After that, the python interpreter itself crashes. I'm not sure whether
my analysis above is correct or whether there is some other problem. In
any case: the test breaks and it seems to be related to the MKL...

Thanks for help,
Greetings,
Norbert