void MultiIsendIrecvImpl::send(const Address &dest, void *msgPtr, int32 msgSize)
{
Action *action = new Action;
action->addr = dest;
action->data = msgPtr;
action->size = msgSize;
Group *g = (Group *)dest.group;
MPI_CALL(Isend(action->data,action->size,MPI_BYTE,
dest.rank,g->tag,g->comm,&action->request));
g->sendQueue.put(action);
}
/*
**************************************************************************
* Parallel prefix sum for ints.
*
* This method implements the parallel prefix sum algorithm.
* The distance loop is expected to execute (ln d_size) times,
* doing up to 1 send and 1 receive each time.
*
* Note: I'm not sure we have to use all non-blocking calls to get
* good performance, but it probably can't hurt. --BTNG
**************************************************************************
*/
int
SAMRAI_MPI::parallelPrefixSum(
int* x,
int count,
int tag) const
{
#ifndef HAVE_MPI
NULL_USE(x);
NULL_USE(count);
NULL_USE(tag);
#endif
// Scratch data.
std::vector<int> send_scr(count), recv_scr(count);
Request send_req, recv_req;
Status send_stat, recv_stat;
int mpi_err = MPI_SUCCESS;
for (int distance = 1; distance < d_size; distance *= 2) {
const int recv_from = d_rank - distance;
const int send_to = d_rank + distance;
if (recv_from >= 0) {
mpi_err = Irecv(&recv_scr[0], count, MPI_INT, recv_from, tag, &recv_req);
if (mpi_err != MPI_SUCCESS) {
return mpi_err;
}
}
if (send_to < d_size) {
send_scr.clear();
send_scr.insert(send_scr.end(), x, x + count);
mpi_err = Isend(&send_scr[0], count, MPI_INT, send_to, tag, &send_req);
if (mpi_err != MPI_SUCCESS) {
return mpi_err;
}
}
if (recv_from >= 0) {
mpi_err = Wait(&recv_req, &recv_stat);
if (mpi_err != MPI_SUCCESS) {
return mpi_err;
}
for (int i = 0; i < count; ++i) {
x[i] += recv_scr[i];
}
}
if (send_to < d_size) {
mpi_err = Wait(&send_req, &send_stat);
if (mpi_err != MPI_SUCCESS) {
return mpi_err;
}
}
}
return MPI_SUCCESS;
}