Re: [ublas] MUMPS and latest traits

Subject: Re: [ublas] MUMPS and latest traits
From: Thomas Klimpel (Thomas.Klimpel_at_[hidden])
Date: 2009-02-22 12:49:18

Thomas Klimpel wrote:
> Karl Meerbergen wrote:
> > I have no experience with bindings + mumps on windows. Someone else?
>
> It's on my to do list, as a part of what I call "backlog".
> Give me time till next Monday, and I will report whether the regression tests passes.

The regression tests for atlas (build under cygwin with gcc 3.4.4) pass under cygwin with gcc (3.4.4), with gcc-4 (4.3.2) and under VC9.
The linking under VC9 was done by renaming/copying libatlas.a to atlas.lib, libcblas.a to cblas.lib and liblapack.a to clapack.lib.

I was more or less forced to use gcc-4/gfortran-4 (4.3.2) to build mumps, because mumps uses fortran90+, and gfortran is not available for gcc 3.4.4.
The regression test for blas, lapack and mumps (build under cygwin with gcc-4-gfortran-4 4.3.2) passes under cygwin with gcc-4. (I had link problems with when using gcc 3.4.4).

I was able to link the regression tests for blas, lapack and mumps (build under cygwin with gcc-4-gfortran-4 4.3.2) under VC9, by copying /lib/gcc/4.3.2/libgfortran.a to gfortran.lib, /lib/gcc/4.3.2/libgcc.a to gcc.lib and lib/libcygwin.a to cygwin.lib (as well as renaming/copying the blas, lapack and mumps libraries to the appropriate names).

However, most of the tests produce crashes when executed, including mumps. I guess the problem is that the executables are unable to find the cygwin1.dll (libcygwin.a is probably not a static library), and crash as soon as a call requiring the cygwin1.dll happens. I copied cygwin1.dll into the same place as the executable, but the crash still happens. Perhaps the crash is related to the fact that "gcc-4 --version" still says "gcc-4 (GCC) 4.3.2 20080827 (alpha-testing) 1". Or the the crash may be related to Clint Whaley's complaint (http://math-atlas.sourceforge.net/errata.html#gccCrazy). I could try building atlas with gcc-4 and see whether it will then stop working with VC9. I could also try to apply Clint Whaley suggested workarounds. I could also try to link against blas and lapack compiled with g77 (3.4.4) and see whether it will still crash. (But this wouldn't help for mumps, because mumps more or less requires gfortran-4).

However, I only promised to try whether it works and report how it went. I have done this now, and I also tried to avoid creating FUD by giving possible reasons why it didn't work. So it is OK from my point of view to give the investigations of the problems a low priority on my to do list.

Regards,
Thomas