[SciPy-Dev] arpack failures 64-bit

[Ralf Gommers]

On Sat, Jun 19, 2010 at 10:03 PM, Aric Hagberg <aric.hagberg at gmail.com>wrote:

> On Fri, Jun 18, 2010 at 11:00 PM, Ralf Gommers
> <ralf.gommers at googlemail.com> wrote:
> > Arpack test failures have been reported by several people, all on 64 bit.
> On
> > 32-bit it's fine.
> >
> > ======================================================================
> > ERROR: test_nonsymmetric_modes (test_arpack.TestEigenNonSymmetric)
> > ----------------------------------------------------------------------
> > Traceback (most recent call last):
> >  File
> >
> "/sw/lib/python2.6/site-packages/scipy/sparse/linalg/eigen/arpack/tests/test_arpack.py",
> > line 204, in test_nonsymmetric_modes
> >    self.eval_evec(m,typ,k,which)
> >  File
> >
> "/sw/lib/python2.6/site-packages/scipy/sparse/linalg/eigen/arpack/tests/test_arpack.py",
> > line 186, in eval_evec
> >    eval,evec=eigen(a,k,which=which,**kwds)
> >  File
> >
> "/sw/lib/python2.6/site-packages/scipy/sparse/linalg/eigen/arpack/arpack.py",
> > line 397,
> > in eigen
> >    params.iterate()
> >  File
> >
> "/sw/lib/python2.6/site-packages/scipy/sparse/linalg/eigen/arpack/arpack.py",
> > line 222,
> > in iterate
> >    raise RuntimeError("Error info=%d in arpack" % self.info)
> > RuntimeError: Error info=-8 in arpack
> >
> > For the full list of errors see
> >
> http://article.gmane.org/gmane.comp.python.scientific.devel/13706/match=arpack
> > (on OS X).
> > Similar report on Linux:
> > http://thread.gmane.org/gmane.comp.python.scientific.devel/8551
> > This seems also related: http://projects.scipy.org/scipy/ticket/725.
> >
> > Does anyone have time to look at this soon? If not, can I mark them as
> > knownfail for 0.8.0?
>
> I took a quick look at this but I don't have a fix and I also can't
> reproduce the problem on 64 bit Linux.
>
> The error is coming from the LAPACK routine zlahqr
> which is what the -8 error code refers to.  See error codes at
>
> http://projects.scipy.org/scipy/browser/trunk/scipy/sparse/linalg/eigen/arpack/ARPACK/SRC/zneupd.f#L149
> :
>
> c          = -8: Error return from LAPACK eigenvalue calculation.
> c                This should never happened.
>
> (and around line 557 where the error is generated after calling zlahqr)
> and the section the
>
> http://projects.scipy.org/scipy/browser/trunk/scipy/sparse/linalg/eigen/arpack/ARPACK/SRC/zneupd.f#L557
>
> The LAPACK subroutine zlahqr in currently included in scipy since in
> earlier
> testing (on Linux) we found that this function as included in the
> packaged libraries was producing errors.
>
> Further, the interface to complex number functions had to be hacked
> to make it work on OSX since the OSX distributed LAPACK had a different
> (C, rather than Fortran) interface:
> http://projects.scipy.org/scipy/ticket/725
>
> Is this an OSX issue only? Those tests pass for me on 64 bit Linux.
> Perhaps we need to revisit the code David inserted (see ticket 725 above)?
>

Yes, looks like it is OS X (10.6) only. This report:
http://article.gmane.org/gmane.comp.python.scientific.devel/13705/match=test_arpack+testeigennonsymmetricseems
to claim that these tests did not fail for 0.7.2.

Ralf


> $ uname -a
> Linux ll 2.6.31-22-generic #60-Ubuntu SMP Thu May 27 02:41:03 UTC 2010
> x86_64 GNU/Linux
> $ dpkg -l |grep atlas
> ii  libatlas-base-dev                    3.6.0-22ubuntu2
>             Automatically Tuned Linear Algebra Software,
> ii  libatlas-headers                     3.6.0-22ubuntu2
>             Automatically Tuned Linear Algebra Software,
> ii  libatlas3gf-base                     3.6.0-22ubuntu2
>             Automatically Tuned Linear Algebra Software,
> $ dpkg -l |grep -i lapack
> ii  liblapack-dev                        3.2.1-1
>             library of linear algebra routines 3 - stati
> ii  liblapack3gf                         3.2.1-1
>             library of linear algebra routines 3 - share
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.python.org/pipermail/scipy-dev/attachments/20100619/25c04c14/attachment.html>