LAPACK failures

[Thomas Koenig]

On Tue, Jan 04, 2005 at 08:16:15PM -0800, Jerry DeLisle wrote:

> LAPACK is an excellent torture test for gfortran.  Unfortunately it is
> so complex that it is not useful for identifying the actual bugs.  It's
> useful only for identifying the existence of bugs.
 
I concur, especially for complex calculations.

(Apart from that, I think some Lapack bugs are due to PR 18398).

> I think we need to avoid optimizations for comparison purposes so we
> have an even playing field.

It depends.  Optimization bugs are bad, too :-)

> I can set up my machine to run the tests regularly.  What I would really
> like to do is come up with a way to isolate the test cases that fail and
> be able to run those repeatedly and singularly so we have a chance to
> nail the bug(s).

The format of the output files is pretty regular.  Using a short
perl script, it shouldn't be too hard to parse the individual
failures and find regressions.

I am under the suspicion that many Lapack bugs are actually
middle or back end bugs, which are exposed by the complexity of
the Lapack code.