[SciPy-dev] scipy distutils questions

Arnd Baecker arnd.baecker at web.de
Thu Dec 15 03:40:39 CST 2005


Good morning alltogether ;-)

thanks to Pearu things look very good now:

- scipy core (0.8.6.1668) installs fine on the Itanium 2
  and mkl72. All tests pass.
- full scipy: installs fine.

  scipy.test(10,10):

  ****************************************************************
  WARNING: clapack module is empty
  -----------
  See scipy/INSTALL.txt for troubleshooting.
  Notes:
  * If atlas library is not found by scipy/system_info.py,
    then scipy uses flapack instead of clapack.
  ****************************************************************

  ****************************************************************
  WARNING: cblas module is empty
  -----------
  See scipy/INSTALL.txt for troubleshooting.
  Notes:
  * If atlas library is not found by scipy/system_info.py,
    then scipy uses fblas instead of cblas.
  ****************************************************************

======================================================================
FAIL: check_tandg (scipy.special.basic.test_basic.test_cephes)
----------------------------------------------------------------------
Traceback (most recent call last):
  File
"/home/baecker/python2/scipy2/lib/python2.4/site-packages/scipy/special/tests/test_basic.py",
line 440, in check_tandg
    assert_equal(cephes.tandg(45),1.0)
  File
"/home/baecker/python2//scipy2/lib/python2.4/site-packages/scipy/test/testing.py",
line 661, in assert_equal
    assert desired == actual, msg
AssertionError:
Items are not equal:
DESIRED: 1.0
ACTUAL: 1.0000000000000002

----------------------------------------------------------------------
Ran 1384 tests in 89.783s

  Compared to .... tests on a machine with ATLAS
  The number of tests (apart from 3 tests from distuils)
  is the one which I also get on a different machine,
  so things look very good on this side!

While we are at it:
- I would also like to test the FFT provided by mkl,
  as it claims to be faster than fftw3 by a factor 1.5 and more
  in a couple of cases:
  http://www.intel.com/cd/software/products/asmo-na/eng/perflib/mkl/219662.htm

  For MKL 8.0, there is a fftw interface which might make
  this possible without too much problems.

- ACML: On an Opteron the following worked with site.cfg

  [atlas]
  library_dirs = /opt/cluster/acml/2.7.0/lib
  atlas_libs = acml

  [blas]
  library_dirs = /opt/cluster/acml/2.7.0/lib
  blas_libs = acml

  [lapack]
  library_dirs = /opt/cluster/acml/2.7.0/lib

  Do you also plan a `site.cfg`-less configuration for ACML?
  (or does this already work?)

- For the SGI ALtix we have one more candidate:
  scsl -  Scientific Computing Software Library from SGI
  (http://www.sgi.com/products/software/scsl.html)
  which provides BLAS/LAPACK, FFT and more...

  I would be surprised, if you had scsl, so I think that one
  will be our job, right?
  Some questions: Similarly to mkl we would have to provide
          'scsl':scsll_info,
          'lapack_scsl':lapack_scsl_info,
          'blas_scls':blas_scsl_info,
  add the corresponding routines accordingly, add a
      default_scls_dirs
  and add the corresponding stuff in
    class lapack_opt_info(system_info):
  ?

One more general question:
In the end it might well happen, that ATLAS is faster for
one set of problems and MKL is faster for another bunch of problems.
This would mean to have two different installations of scipy around.
Adding ACML and scsl to the mess leads to 4 installations
and allowing for different types of FFTs would lead
to a nice tree of installations ...

Do you think it is feasible to make it possible to
choose between  different
(for example) flapack.so from within one scipy installation?

((I am just running a second install for fftw3 and
see if it is also slow on this machine, but that's
for a separate thread...))

Many thanks,

Arnd




More information about the Scipy-dev mailing list