Pmpi fortran #4651
Pmpi fortran #4651
Conversation
The Fortran wrappers like send_f.c have been defining mpi_send() and
calling C PMPI_Send(). This checkin defines a set of function pointers like
ompi_fptr_MPI_Send
to be used instead. And these can dynamically point to either MPI_Send or
PMPI_Send.
There's a corresponding checkin that involves a bunch of scripted changes
to make the calls happen as described above. This checkin has all the
by-hand changes.
Currently the arguments needed to define function pointers for all the MPI C
routines are coming from parsing mpi.h in a script normalize_mpih.pl which
outputs a file "flist" in a very uniform layout. Then a second script mkcode.pl
generates the declarations and settings for all those function pointers.
A small number of fortran wrappers contain an ompi_fptr_init() call that
initializes the function pointers with values. Then at the end of
ompi_mpi_init() there's another initialization. The second initialization
is there since the MCA system is available then so the setting for
--mca mpi_fortcall MPI
--mca mpi_fortcall PMPI
is expected to be available at that point. The earlier initialization is
based on less complete information.
Signed-off-by: Mark Allen <[email protected]>
This is to enable the fortran wrappers to have a runtime ability to call either MPI_Foo() or PMPI_Foo() for the calls that are language wrapped. The code change actually uses a macro so we could put it back to vanilla PMPI_Foo() if desired, possibly with configure options. Signed-off-by: Mark Allen <[email protected]>
|
@markalle what happens if a |
|
Oh no, I think you're right. The way this is done there's only one send_f.o and it uses pragma weak to create both mpi_send and pmpi_send which aren't distinguished from each other. So yeah, a fortran app calling pmpi_send() would go back out into MPI_Send() and through any wrapper that might exist there. I don't see how to fix that without a more significant change to separate the behavior of mpi_send() from pmpi_send(). Hmm, the current PR is within my threshold of "triviality" for adding convenience features like this, but I'm not so sure about the next step of doubling up on the mpi_send()/pmpi_send() definitions... I may have to retract this PR then and go back to our layered entrypoint feature which is a different solution to the same problem. |
|
@markalle Given your last reply, I marked this PR as "WIP DNM" (work in progress / do not merge). Feel free to update this PR or close it. |
|
@markalle did you still want this PR or are you planning to pick it up again? |
|
Can one of the admins verify this patch? |
|
closing no intent to merge |
7 participants
This is a convenience feature for PMPI wrapper library writers, that allows them to have F77 language wrapping for free (something allowed but not required by the standard).
Eg if a wrapper library redefines
int MPI_Send(...args...) { ... record stuff... ; rv = PMPI_Send; return(rv); }
and a Fortran app calls mpi_send() it's possible to have the mpi_send() definition in fortran/mpif-h/send_f.c call MPI_Send which would resolve back out into the wrapper library.
A longer discussion about this feature is here:
#3954
The main argument against this is it's not required by the standard, and hence there's no good way for wrapper library writers to use this functionality: If they want to support all MPI vendors they can't rely on having access to the above. If they want to be complete, they have to write their own mpi_send() as well, and call either pmpi_send or PMPI_Send from it.
The argument for this feature is that pragmatically these C-only wrapper tools already exist, and customers try to use them with their Fortran apps, so it's a nice convenience feature to offer (and it's not random, it's an optional part of the standard).
This PR changes the fortran/mpif-h/send_f.c (etc) code from a direct PMPI_Send() call to a function pointer where the function pointer defaults to PMPI_Send(), and gives runtime control:
-mca mpi_fortcall PMPI : the default behavior (mpi_send calls PMPI_Send)
-mca mpi_fortcall MPI : make mpi_send call MPI_Send
The code's got an extra
#if OMPI_FORTRAN_USE_FPTR
that isn't being used right now, but is a stub for the idea of making this configurable. Eg if someone wanted no trace of the above changes and just wanted the original direct PMPI_Send() calls.