int
sc_allgather (void *sendbuf, int sendcount, sc_MPI_Datatype sendtype,
void *recvbuf, int recvcount, sc_MPI_Datatype recvtype,
sc_MPI_Comm mpicomm)
{
int mpiret;
int mpisize;
int mpirank;
size_t datasize;
#ifdef SC_DEBUG
size_t datasize2;
#endif
SC_ASSERT (sendcount >= 0 && recvcount >= 0);
/* *INDENT-OFF* HORRIBLE indent bug */
datasize = (size_t) sendcount * sc_mpi_sizeof (sendtype);
#ifdef SC_DEBUG
datasize2 = (size_t) recvcount * sc_mpi_sizeof (recvtype);
#endif
/* *INDENT-ON* */
SC_ASSERT (datasize == datasize2);
mpiret = sc_MPI_Comm_size (mpicomm, &mpisize);
SC_CHECK_MPI (mpiret);
mpiret = sc_MPI_Comm_rank (mpicomm, &mpirank);
SC_CHECK_MPI (mpiret);
memcpy (((char *) recvbuf) + mpirank * datasize, sendbuf, datasize);
sc_allgather_recursive (mpicomm, (char *) recvbuf, (int) datasize,
mpisize, mpirank, mpirank);
return sc_MPI_SUCCESS;
}
static void
sc_reduce_alltoall (MPI_Comm mpicomm,
void *data, int count, MPI_Datatype datatype,
int groupsize, int target,
int maxlevel, int level, int branch,
sc_reduce_t reduce_fn)
{
int i, l;
int mpiret;
int doall, allcount;
int myrank, peer, peer2;
int shift;
char *alldata;
size_t datasize;
MPI_Request *request, *rrequest, *srequest;
doall = 0;
if (target == -1) {
doall = 1;
target = 0;
}
SC_ASSERT (0 <= target && target < groupsize);
myrank = sc_search_bias (maxlevel, level, branch, target);
SC_ASSERT (0 <= myrank && myrank < groupsize);
SC_ASSERT (reduce_fn != NULL);
/* *INDENT-OFF* HORRIBLE indent bug */
datasize = (size_t) count * sc_mpi_sizeof (datatype);
/* *INDENT-ON* */
if (doall || target == myrank) {
allcount = 1 << level;
alldata = SC_ALLOC (char, allcount * datasize);
request = SC_ALLOC (MPI_Request, 2 * allcount);
rrequest = request;
srequest = request + allcount;
for (i = 0; i < allcount; ++i) {
peer = sc_search_bias (maxlevel, level, i, target);
/* communicate with existing peers */
if (peer == myrank) {
memcpy (alldata + i * datasize, data, datasize);
rrequest[i] = srequest[i] = MPI_REQUEST_NULL;
}
else {
if (peer < groupsize) {
mpiret = MPI_Irecv (alldata + i * datasize, datasize, MPI_BYTE,
peer, SC_TAG_REDUCE, mpicomm, rrequest + i);
SC_CHECK_MPI (mpiret);
if (doall) {
mpiret = MPI_Isend (data, datasize, MPI_BYTE,
peer, SC_TAG_REDUCE, mpicomm, srequest + i);
SC_CHECK_MPI (mpiret);
}
else {
srequest[i] = MPI_REQUEST_NULL; /* unused */
}
}
else {
/* ignore non-existing ranks greater or equal mpisize */
rrequest[i] = srequest[i] = MPI_REQUEST_NULL;
}
}
}
/* complete receive operations */
mpiret = MPI_Waitall (allcount, rrequest, MPI_STATUSES_IGNORE);
SC_CHECK_MPI (mpiret);
/* process received data in the same order as sc_reduce_recursive */
for (shift = 0, l = level - 1; l >= 0; ++shift, --l) {
for (i = 0; i < 1 << l; ++i) {
#ifdef SC_DEBUG
peer = sc_search_bias (maxlevel, l + 1, 2 * i, target);
#endif
peer2 = sc_search_bias (maxlevel, l + 1, 2 * i + 1, target);
SC_ASSERT (peer < peer2);
if (peer2 < groupsize) {
reduce_fn (alldata + ((2 * i + 1) << shift) * datasize,
alldata + ((2 * i) << shift) * datasize,
count, datatype);
}
}
}
memcpy (data, alldata, datasize);
SC_FREE (alldata); /* alldata is not used in send buffers */
/* wait for sends only after computation is done */
if (doall) {
mpiret = MPI_Waitall (allcount, srequest, MPI_STATUSES_IGNORE);
SC_CHECK_MPI (mpiret);
}
SC_FREE (request);
}
