[MPI3 Fortran] Deprecate mpif.h?
crasmussen at newmexicoconsortium.org
Tue Mar 9 12:29:35 CST 2010
On Mar 9, 2010, at 10:52 AM, Supalov, Alexander wrote:
> Thanks. Any MPI implementation probably has a noncontig-to-contig
> copy already; also, there's probably a few trickier functions in the
> MPI than the memcpy; I'd suggest we let the implementors do the
> right thing here.
Good. I didn't expect that the copy would be all that difficult and
could probably make use of existing software.
> What is interesting to learn is what kind of representation will be
> chosen for the array descriptors, and whether it will be more or
> less standard across different compilers.
Yes, the descriptor seen from C will be standard across compilers.
The design is currently open but I can send you a document that will
give you the flavor of where we are heading.
> If it won't, I'll be interested in seeing an estimate how easy it
> will be to support several Fortran compilers at once. Note that
> compilers are often changing their conventions (see GNU, for one),
> which adds an extra dimension to this already complicated matter.
It is likely that GNU will just adopt the interoperability header to
make their life simpler. This means they won't have to copy to/from
different descriptors that cross BIND(C) interfaces.
> Finally, if an array section is passed, and its array descriptor is
> available, what will the MPI datatype argument do there? It looks
> almost superfluous to me in the new Fortran binding, at least in
> this case.
I've been assuming that array sections would only be used for
primitive MPI_Types. The case of array sections and user-defined
types makes my head spin. I'll have to think about it further.
More information about the mpiwg-fortran