[gpaw-users] Test failures

[Jens Jørgen Mortensen]

On Mon, 2010-11-15 at 20:40 +0200, Janne Blomqvist wrote:
> On Mon, 15 Nov 2010 12:32:46 +0100, Marcin Dulak <Marcin.Dulak at fysik.dtu.dk> wrote:
> > Hi,
> > 
> > Ondrej Marsalek wrote:
> > > Hi,
> > >
> > > On Mon, Nov 15, 2010 at 10:17, Marcin Dulak <Marcin.Dulak at fysik.dtu.dk> wrote:
> > >   
> > >> Hi,
> > >>
> > >> first thing - avoid icc, i usually  have problems with it.
> > >>     
> > >
> > > Thanks for the warning. I've had trouble with Intel compilers with
> > > other software, but thought that with GPAW it was OK. Is anyone using
> > > icc with GPAW? If so, what version?
> 
> I've tried with 9.1, 10.something and 11.0, never got it to work
> properly. 
> 
> > >> With gcc: the failure of numpy_zdotc_graphite.py is usually a sign
> > >> that there is a problem of gpaw-python/numpy linked blas:
> > >> if your numpy built also with mkl - try to build it with default blas (and
> > >> lapack) coming from numpy source.
> > >> It will less inefficient, but probably work correctly.
> > >>     
> > >
> > > Interesting. I am using the distribution-provided numpy that comes
> > > with Ubuntu 10.10:
> > > Python 2.6.6
> > > numpy 1.3.0
> > >   
> > i would still build numpy with it's internal blas/lapack.
> > The one from ubuntu uses probably atlas.
> 
> Yes, the ubuntu numpy uses atlas. However, at least with Ubuntu 10.04
> AMD64 it seems to work for me with the packaged gcc, atlas and
> openmpi. My customize.py is empty so everything is picked up
> automatically when building.

10.10 also works for me out of the box with atlas-sse2.  Have you tried
to compile GPAW with atlas instead of MKL?  If yes, what atlas library
did you use - maybe try the generic one.

JJ