[SciPy-Dev] Many errors/failures running scipy.test() using latest numpy and scipy

Benjamin Root ben.root@ou....
Tue Jun 21 12:23:54 CDT 2011


On Tue, Jun 21, 2011 at 11:15 AM, Pauli Virtanen <pav@iki.fi> wrote:

> Tue, 21 Jun 2011 10:51:18 -0500, Benjamin Root wrote:
> [clip]
> > The first error has to do with the netcdf file (I thought that was fixed
> > by now?). Then, some of the errors are related to the gmres module (I
> > see that is already known).  Then, there are many errors for
> > "test_arpack.test_real_nonsymmetric_modes",
> > "test_arpack.test_complex_nonsymmetric_modes", and
> > "test_arpack.test_hermitian_modes". Lastly, there is a single failure
> > for "test_morestats.TestAnderson".
> >
> > I hope this is useful.
>
> All of the ARPACK failures are GMRES-related.
>
> These appear to be very strongly platform-dependent for some reason.
> I don't see them, and it is somewhat difficult to believe that the
> calculation could be this unstable because of rounding error.
>
> Can you check if your GMRES works at all, i.e., try a simple 6 x 6
> complex-valued matrix? The current test cases don't actually cover
> it properly, and at least Gfortran 4.5.2 does miscompile the GMRES
> routines in some cases. Not sure about 4.6.0 that you use, but
> compiler issues remain a possibility (esp. since it is a .0 release).
>
>
Using the example given in the other mail thread:

import numpy as np
import scipy.sparse.linalg
A = np.random.rand(6, 6) + 1j*np.random.rand(6, 6)
b = np.random.rand(6) + 1j*np.random.rand(6)
x, info = scipy.sparse.linalg.gmres(A, b)
y = np.linalg.solve(A, b)
print abs(x - y).max()

produces:

0.427293469893



> Maybe using a different gfortran version makes these go away?
>
> F90=gfortran-4.4 python setup.py build
>
>
I don't see another gfortran available in the repos.  I will look further.

Ben Root
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mail.scipy.org/pipermail/scipy-dev/attachments/20110621/83cdab41/attachment.html 


More information about the SciPy-Dev mailing list