PGI compilation of LAPACK (ATLAS, ACML, etc...)
Posted: Fri Oct 29, 2004 7:55 pm
This is a note to other users trying to compile VASP with the Portland Group compiler and linking against ACML or ATLAS or the generic LAPACK library (as well as a request for info on how to optimize the compilation of the code).
We have had numerical instability problems on Opteron/Linux when compiling with PGF90 and the "-fastsse" flag. It turns out that parts of those Lapack libraries we were linking against did not behave well when we tried to compile VASP with the "-fastsse" flag. This caused problems when calling the ZHEGV LAPACK routine for optimizing the wave-function. The problem occured both for IALGO = 38 or 48.
We got rid of the "-fastsse" optimization we compiled using only a regular "-O3" optimization and things now work fine.
However the code is slower too.
It turns out that Portland Group is reporting that their fastsse flag causes some of the sanity checks of ATLAS to fail. This could be related, perhaps:
http://www.pgroup.com/resources/atlas/a ... _pgi52.htm
Does anyone know of a good (fast and reliable) set of PGF90 compiler options for Opterons?
Thanks,
Olivier.
We have had numerical instability problems on Opteron/Linux when compiling with PGF90 and the "-fastsse" flag. It turns out that parts of those Lapack libraries we were linking against did not behave well when we tried to compile VASP with the "-fastsse" flag. This caused problems when calling the ZHEGV LAPACK routine for optimizing the wave-function. The problem occured both for IALGO = 38 or 48.
We got rid of the "-fastsse" optimization we compiled using only a regular "-O3" optimization and things now work fine.
However the code is slower too.
It turns out that Portland Group is reporting that their fastsse flag causes some of the sanity checks of ATLAS to fail. This could be related, perhaps:
http://www.pgroup.com/resources/atlas/a ... _pgi52.htm
Does anyone know of a good (fast and reliable) set of PGF90 compiler options for Opterons?
Thanks,
Olivier.