On Dec 5, 2008, at 6:58 PM, Anthony Chan wrote:
AFAIK, all known/popular MPI implemention's fortran binding
layer is implemented with C MPI functions, including
MPICH2 and OpenMPI. If MPICH2's fortran layer was implemented
the way you said, typical profiling tools including MPE will
fail to work with fortran applications.
FWIW, NEC's MPI has their Fortran functions directly call back-end
functionality (vs. calling the C MPI API). We've considered doing
this for at least some key Fortran MPI API functions in Open MPI as
well. Specifically, have MPI_SEND [Fortran] directly call the back-
end sending functionality rather than calling MPI_Send [C] or
PMPI_Send [C].
Hence, I don't think it's a good assumption to rely on (that the MPI
Fortran API always calls the [P]MPI C API).
I think one of George's middle points is the most relevant here: those
who are interested in this stuff should definitely participate in the
MPI Forum's Tools group:
https://svn.mpi-forum.org/trac/mpi-forum-web/wiki/MPI3Tools
--
Jeff Squyres
Cisco Systems
