[SciPy-dev] Reverting eigh code ?

[David Cournapeau]

Hi Tiziano,

Tiziano Zito wrote:
> I have added the code for eigh. The corresponding fortran wrappers
> have been running without any problems since 6 years in symeig. symeig
> has been dowloaded more than 1000 times since its first appearence and
> the MDP package (more than 11000 dowloads) depends on it. No bug has
> been reported regarding symeig in 6 years.

I have no doubt the fortran code itself is OK. But there is a vast
variety of Fortran/Blas/Lapack configurations, and sometimes, the code
has to be modified to run on some platforms. That's why I am a bit
worried about last-minute change with fortran-related code; experience
says that it has always been a source of problems.

> I feel responsible for the eigh code, so if someone cares to submit a
> proper bug report, or a link to a scipy ticket, I would be glad to
> also have a look at it. 

http://projects.scipy.org/scipy/scipy/ticket/795

I could reproduce a similar problem (indeed related to work size) on
Ubuntu Hardy 32 bits with custom-build ATLAS, and RHEL 5 64 bits - the
tests fail, but do not crash, though.

> Is there a buildbot that we can use for such cases?

Not for scipy, unfortunately. Here is the exact configurations which
fail for me:
    - Ubuntu 8.04 - 32 bits
    - packaged gcc and g77
    - custom-made ATLAS (3.8.2) against LAPACK 3.1.1

And
    - RHEL 5 - 64 bits
    - packaged gcc and gfortran
    - custom made ATLAS (3.8.2) against LAPACK 3.1.1

In the later case, I cannot reproduce the problem if I use LAPACK
without ATLAS (I have not taken the time to test this with Ubuntu, but I
can do it if that's useful).

David