(corrected) Simplified version of generalized requests

Nathan E. Doss (doss@ERC.MsState.Edu)
Fri, 28 Jul 1995 08:12:43 -0500 (CDT)

The example is more correct than in the one I'd sent out earlier.

-- 
Nathan Doss                  doss@ERC.MsState.Edu




This is a simplified version of the current generalized request
proposal.  It consists of the following four functions:


MPI_REQUEST_TYPE_CREATE ( init, start, free, request_type )
   IN    init          Callback invoked when request is init'ed
   IN    start         Callback invoked when request is started
   IN    free          Callback invoked when request is freed
   OUT   request_type  MPI created request type


MPI_REQUEST_TYPE_FREE ( request_type )
   INOUT request_type  Type of request to be freed


MPI_REQUEST_INIT ( request_type, extra_state, comm, request )
   IN    request_type  Type of request
   IN    extra_state   Extra request_type specific state 
   IN    comm          MPI communicator
   OUT   request       MPI request


MPI_REQUEST_MARK_COMPLETED ( request ) 
   OUT   request       Request to be marked complete


One thing the current version of the proposal attempts to manage
is handlers that fire as a result of the completion of an associated
request.  IF we are going to have hsend/hrecv, this functionality doesn't
necessarily have to be in the generalized request mechanism.


Below is a version of the allgather example from section 3.2.4 of the July
24 draft using these functions and a persistant version of the
hrecv/hsend proposed by Snir.


Main routine that calls the non-blocking operation:


MPI_Request_type_create( init_fn, start_fn, free_fn, &type_req );
Fill_in_extra_state( &extra_state );
MPI_Request_init( type_req, extra_state, comm, &request );
MPI_Start( &request );
Do_other_stuff();
MPI_Wait( &request );
MPI_Type_request_free( &type_req );




The init_fn, start_fn, & free_fn:


init_fn( extra, comm, request )
{    
    MPI_Comm_size ( comm, &size );
    MPI_Comm_rank ( comm, &rank );
    extra->comm    = comm;
    extra->request = request;
    extra->req_completed = 0;
    
    from = (i-1) % p;
    to   = (i+1) % p;


    MPI_Hrecv_init(extra->rbuf,1,MPI_DOUBLE,from,13,comm,handler_fn,extra,
                   &extra->rreq);
    MPI_Hsend_init(extra->sbuf,1,MPI_DOUBLE,to,  13,comm,handler_fn,extra,
                   &extra->sreq);


    return (MPI_SUCCESS);
}


start_fn( extra, comm, request )
{
    extra->sbuf = extra->buf[rank];
    extra->num_done = 0;


    MPI_Start( extra->rreq );
    MPI_Start( extra->sreq );


    return (MPI_SUCCESS);
}


free_fn( extra, comm, request )
{
    MPI_Request_free( &extra->sreq );
    MPI_Request_free( &extra->rreq );


    return (MPI_SUCCESS);
}






The handler_fn used by the hrecv/hsend:


handler_fn( extra, status )
{
    extra->num_completed++;
    if (extra->num_completed != 2)
        return (MPI_SUCCESS);
    extra->num_completed = 0; /* reset */
    
    /* We don't get here unless both the send & recv completed */ 
    MPI_Comm_size ( extra->comm, &size );
    extra->num_done++;


    if (extra->num_done == size - 1) {
        MPI_Request_mark_completed( extra->request );
        return (MPI_SUCCESS);
    }


    /* store away received data/set data to be sent */
    extra->buf[?number determined by num_done?] = extra->rbuf;
    extra->sbuf    = extra->buf[?number determined by num_done?];


    MPI_Start( extra->sreq );
    MPI_Start( extra->rreq );


    return (MPI_SUCCESS);
}