[Mpi-forum] Missing user data in MPI_User_function?

Phil Ruffwind rf at rufflewind.com
Mon Apr 10 04:23:57 CDT 2017


I am experimenting with an interface for doing generic vector
operations, which would have MPI as a potential backend.  To implement a
reduce over a user-provided function, I would wrap their function inside
a wrapper compatible with the MPI_User_function signature, and then call
MPI_Op_create.  But I can only make a single wrapper as standard C does
not provide a way to create function pointers out of thin air, so the
pointer to user's original function needs to be somehow passed in as
“context” into the wrapper.

Normally, C APIs that accept callbacks also accept an opaque void
pointer which is then passed into the callback function – this would
serve as the context.  But I do not see that in the declaration of
MPI_User_function:

    typedef void MPI_User_function(void *invec,
                                   void *inoutvec,
                                   int *len,
                                   MPI_Datatype *datatype);

This leaves me with several undesirable options:

- Use global data, which, even ignoring the thread-safety implications,
means that I can only wrap a single user function at any given time.

- Use thread-local data, which I'm not sure is sensible either since
there's no guarantee MPI is going to call my wrapper function on the
same thread (and it still has the one-instance limitation).

- Use invec or inoutvec to pass in the contextual data, which is awkward
because it means now the data is complicated by the addition of
unrelated contextual data, and it also slightly bloats the data with
stuff that needn't be transmitted.

- Create a wrapper function dynamically at run time, which means
throwing away portability.

Is there another more sensible option for this dilemma?

Thanks,
Phil


More information about the mpi-forum mailing list