the symbol is actually defined in libsci_mpi (blas, lapack, scalapack…this is 
really crazy)
Since libsci_mpi has ScaLAPACK 2.0.2, I guess I will be fine linking with that 
and obtain a working MUMPS.

Thanks
Stefano

On Feb 25, 2015, at 11:10 PM, Barry Smith <[email protected]> wrote:

> 
>   The issue is here in Scalapack.py
> 
> class Configure(config.package.Package):
>  def __init__(self, framework):
>    config.package.Package.__init__(self, framework)
>    self.download         = 
> ['http://www.netlib.org/scalapack/scalapack-2.0.2.tgz',
>                             
> 'http://ftp.mcs.anl.gov/pub/petsc/externalpackages/scalapack-2.0.2.tgz']
>    self.includes         = []
>    self.liblist          = [[],['libscalapack.a']]
>                            ^^^^^^^^
> 
> we first check for scalapack in the "default system libraries" and then only 
> check in libscalapack.a if it is not in the "default system libraries".
> 
> Cray has all kinds of crazy stuff like scalapack in the default system 
> libraries and hence finds the symbol pssytrd before it ever checks in the 
> location where we built it.  
> 
> Now the problem is even if we remove the []   from the 
> [[],['libscalapack.a']] and thus explicitly list the downloaded library 
> location when linking code the linker will find both libraries and we have no 
> control over which one it actually uses so it could end up using the Cray 
> one. 
> 
> If you really want to use the download one instead of the installed one you 
> need to poke around the Cray documentation and figure out if you can remove 
> the automatic linking of the Cray one. Perhaps turning off some module or 
> something?
> 
>  Barry
> 
> 
> 
>> On Feb 25, 2015, at 1:58 PM, Stefano Zampini <[email protected]> 
>> wrote:
>> 
>> I just configured on a Cray XC40 with downloading MUMPS and Scalapack , and 
>> it seems that something is wrong  when creating petscvariables. 
>> SCALAPACK_LIB is empty and -lscalapack does not appear in PETSC_EXTERNAL_LIB.
>> Attached a configure.log
>> 
>> <configure.log>
> 

Reply via email to