void
sc_allgather_alltoall (sc_MPI_Comm mpicomm, char *data, int datasize,
int groupsize, int myoffset, int myrank)
{
int j, peer;
int mpiret;
sc_MPI_Request *request;
SC_ASSERT (myoffset >= 0 && myoffset < groupsize);
request = SC_ALLOC (sc_MPI_Request, 2 * groupsize);
for (j = 0; j < groupsize; ++j) {
if (j == myoffset) {
request[j] = request[groupsize + j] = sc_MPI_REQUEST_NULL;
continue;
}
peer = myrank - (myoffset - j);
mpiret = sc_MPI_Irecv (data + j * datasize, datasize, sc_MPI_BYTE,
peer, SC_TAG_AG_ALLTOALL, mpicomm, request + j);
SC_CHECK_MPI (mpiret);
mpiret = sc_MPI_Isend (data + myoffset * datasize, datasize, sc_MPI_BYTE,
peer, SC_TAG_AG_ALLTOALL,
mpicomm, request + groupsize + j);
SC_CHECK_MPI (mpiret);
}
mpiret = sc_MPI_Waitall (2 * groupsize, request, sc_MPI_STATUSES_IGNORE);
SC_CHECK_MPI (mpiret);
SC_FREE (request);
}
int
p6est_profile_sync (p6est_profile_t * profile)
{
p4est_lnodes_t *lnodes = profile->lnodes;
p4est_locidx_t nln = lnodes->num_local_nodes;
sc_array_t lrview;
p4est_lnodes_buffer_t *countbuf;
sc_array_t *sharers;
size_t zz, nsharers;
int nleft;
int8_t *recv, *send;
int *array_of_indices;
p4est_locidx_t recv_total;
p4est_locidx_t *recv_offsets, recv_offset;
p4est_locidx_t send_total;
p4est_locidx_t *send_offsets, send_offset;
p4est_locidx_t (*lr)[2];
sc_array_t *lc = profile->lnode_columns;
sc_MPI_Request *recv_request, *send_request;
sc_array_t *work;
int any_change = 0;
int any_global_change;
int mpiret, mpirank;
int evenodd = profile->evenodd;
lr = (p4est_locidx_t (*)[2]) profile->lnode_ranges;
sharers = lnodes->sharers;
nsharers = sharers->elem_count;
mpiret = sc_MPI_Comm_rank (lnodes->mpicomm, &mpirank);
SC_CHECK_MPI (mpiret);
sc_array_init_data (&lrview, lr, 2 * sizeof (p4est_locidx_t), nln);
countbuf = p4est_lnodes_share_all_begin (&lrview, lnodes);
send_offsets = P4EST_ALLOC (p4est_locidx_t, nsharers + 1);
send_offset = 0;
for (zz = 0; zz < nsharers; zz++) {
p4est_lnodes_rank_t *sharer;
sc_array_t *send_buf;
size_t zy, nnodes;
send_offsets[zz] = send_offset;
sharer = p4est_lnodes_rank_array_index (sharers, zz);
if (sharer->rank == mpirank) {
continue;
}
send_buf = (sc_array_t *) sc_array_index (countbuf->send_buffers, zz);
nnodes = sharer->shared_nodes.elem_count;
P4EST_ASSERT (nnodes == send_buf->elem_count);
P4EST_ASSERT (send_buf->elem_size == 2 * sizeof (p4est_locidx_t));
for (zy = 0; zy < nnodes; zy++) {
p4est_locidx_t *lp =
(p4est_locidx_t *) sc_array_index (send_buf, zy);
P4EST_ASSERT (lp[0] >= 0);
P4EST_ASSERT (lp[1] >= 0);
send_offset += lp[1];
}
}
send_total = send_offsets[nsharers] = send_offset;
p4est_lnodes_share_all_end (countbuf);
recv_offsets = P4EST_ALLOC (p4est_locidx_t, nsharers + 1);
recv_offset = 0;
for (zz = 0; zz < nsharers; zz++) {
p4est_lnodes_rank_t *sharer;
sc_array_t *recv_buf;
size_t zy, nnodes;
recv_offsets[zz] = recv_offset;
sharer = p4est_lnodes_rank_array_index (sharers, zz);
if (sharer->rank == mpirank) {
continue;
}
recv_buf = (sc_array_t *) sc_array_index (countbuf->recv_buffers, zz);
nnodes = sharer->shared_nodes.elem_count;
P4EST_ASSERT (nnodes == recv_buf->elem_count);
P4EST_ASSERT (recv_buf->elem_size == 2 * sizeof (p4est_locidx_t));
for (zy = 0; zy < nnodes; zy++) {
p4est_locidx_t *lp =
(p4est_locidx_t *) sc_array_index (recv_buf, zy);
P4EST_ASSERT (lp[0] >= 0);
P4EST_ASSERT (lp[1] >= 0);
recv_offset += lp[1];
}
}
recv_total = recv_offsets[nsharers] = recv_offset;
recv = P4EST_ALLOC (int8_t, recv_total);
recv_request = P4EST_ALLOC (sc_MPI_Request, nsharers);
send = P4EST_ALLOC (int8_t, send_total);
send_request = P4EST_ALLOC (sc_MPI_Request, nsharers);
/* post receives */
nleft = 0;
for (zz = 0; zz < nsharers; zz++) {
p4est_lnodes_rank_t *sharer;
int icount = recv_offsets[zz + 1] - recv_offsets[zz];
sharer = p4est_lnodes_rank_array_index (sharers, zz);
if (sharer->rank == mpirank) {
recv_request[zz] = sc_MPI_REQUEST_NULL;
continue;
}
if (icount) {
mpiret =
sc_MPI_Irecv (recv + recv_offsets[zz], icount * sizeof (int8_t),
sc_MPI_BYTE, sharer->rank, P6EST_COMM_BALANCE,
lnodes->mpicomm, recv_request + zz);
SC_CHECK_MPI (mpiret);
nleft++;
}
else {
recv_request[zz] = sc_MPI_REQUEST_NULL;
}
}
/* post sends */
for (zz = 0; zz < nsharers; zz++) {
p4est_lnodes_rank_t *sharer;
size_t zy, nnodes;
int icount;
sc_array_t *shared_nodes;
sharer = p4est_lnodes_rank_array_index (sharers, zz);
if (sharer->rank == mpirank) {
send_request[zz] = sc_MPI_REQUEST_NULL;
continue;
}
shared_nodes = &sharer->shared_nodes;
nnodes = shared_nodes->elem_count;
icount = 0;
for (zy = 0; zy < nnodes; zy++) {
p4est_locidx_t nidx;
int8_t *c;
nidx = *((p4est_locidx_t *) sc_array_index (shared_nodes, zy));
if (lr[nidx][1]) {
c = (int8_t *) sc_array_index (lc, lr[nidx][0]);
memcpy (send + send_offsets[zz] + icount, c,
lr[nidx][1] * sizeof (int8_t));
icount += lr[nidx][1];
}
else {
P4EST_ASSERT (!lr[nidx][0]);
}
}
P4EST_ASSERT (icount == send_offsets[zz + 1] - send_offsets[zz]);
if (icount) {
mpiret =
sc_MPI_Isend (send + send_offsets[zz], icount * sizeof (int8_t),
sc_MPI_BYTE, sharer->rank, P6EST_COMM_BALANCE,
lnodes->mpicomm, send_request + zz);
SC_CHECK_MPI (mpiret);
}
else {
send_request[zz] = sc_MPI_REQUEST_NULL;
}
}
work = sc_array_new (sizeof (int8_t));
array_of_indices = P4EST_ALLOC (int, nsharers);
while (nleft) {
int outcount;
int i;
mpiret = sc_MPI_Waitsome (nsharers, recv_request, &outcount,
array_of_indices, sc_MPI_STATUSES_IGNORE);
SC_CHECK_MPI (mpiret);
for (i = 0; i < outcount; i++) {
p4est_lnodes_rank_t *sharer;
size_t zy, nnode;
sc_array_t *shared_nodes;
sc_array_t *recv_buf;
zz = array_of_indices[i];
sharer = p4est_lnodes_rank_array_index (sharers, zz);
shared_nodes = &sharer->shared_nodes;
recv_buf = (sc_array_t *) sc_array_index (countbuf->recv_buffers, zz);
nnode = shared_nodes->elem_count;
P4EST_ASSERT (nnode == recv_buf->elem_count);
recv_offset = recv_offsets[zz];
for (zy = 0; zy < nnode; zy++) {
p4est_locidx_t *lp;
p4est_locidx_t nidx;
sc_array_t oldview, newview;
nidx = *((p4est_locidx_t *) sc_array_index (shared_nodes, zy));
lp = (p4est_locidx_t *) sc_array_index (recv_buf, zy);
sc_array_init_view (&oldview, lc, lr[nidx][0], lr[nidx][1]);
sc_array_init_data (&newview, recv + recv_offset, sizeof (int8_t),
lp[1]);
if (profile->ptype == P6EST_PROFILE_UNION) {
p6est_profile_union (&oldview, &newview, work);
if (work->elem_count > oldview.elem_count) {
int8_t *c;
any_change = 1;
lr[nidx][0] = lc->elem_count;
lr[nidx][1] = work->elem_count;
profile->lnode_changed[evenodd][nidx] = 1;
c = (int8_t *) sc_array_push_count (lc, work->elem_count);
memcpy (c, work->array, work->elem_count * work->elem_size);
}
}
else {
p6est_profile_intersection (&oldview, &newview, work);
P4EST_ASSERT (work->elem_count <= oldview.elem_count);
if (work->elem_count < oldview.elem_count) {
lr[nidx][1] = work->elem_count;
memcpy (oldview.array, work->array,
work->elem_count * work->elem_size);
}
}
recv_offset += lp[1];
}
P4EST_ASSERT (recv_offset == recv_offsets[zz + 1]);
}
nleft -= outcount;
P4EST_ASSERT (nleft >= 0);
}
P4EST_FREE (array_of_indices);
sc_array_destroy (work);
p6est_profile_compress (profile);
p4est_lnodes_buffer_destroy (countbuf);
P4EST_FREE (recv_request);
P4EST_FREE (recv_offsets);
P4EST_FREE (recv);
{
mpiret = sc_MPI_Waitall (nsharers, send_request, sc_MPI_STATUSES_IGNORE);
SC_CHECK_MPI (mpiret);
P4EST_FREE (send_request);
P4EST_FREE (send_offsets);
P4EST_FREE (send);
any_global_change = any_change;
mpiret = sc_MPI_Allreduce (&any_change, &any_global_change, 1, sc_MPI_INT,
sc_MPI_LOR, lnodes->mpicomm);
SC_CHECK_MPI (mpiret);
}
return any_global_change;
}
void
sc_allgather_recursive (sc_MPI_Comm mpicomm, char *data, int datasize,
int groupsize, int myoffset, int myrank)
{
const int g2 = groupsize / 2;
const int g2B = groupsize - g2;
int mpiret;
sc_MPI_Request request[3];
SC_ASSERT (myoffset >= 0 && myoffset < groupsize);
if (groupsize > SC_AG_ALLTOALL_MAX) {
if (myoffset < g2) {
sc_allgather_recursive (mpicomm, data, datasize, g2, myoffset, myrank);
mpiret = sc_MPI_Irecv (data + g2 * datasize, g2B * datasize,
sc_MPI_BYTE, myrank + g2, SC_TAG_AG_RECURSIVE_B,
mpicomm, request + 0);
SC_CHECK_MPI (mpiret);
mpiret = sc_MPI_Isend (data, g2 * datasize, sc_MPI_BYTE,
myrank + g2, SC_TAG_AG_RECURSIVE_A,
mpicomm, request + 1);
SC_CHECK_MPI (mpiret);
if (myoffset == g2 - 1 && g2 != g2B) {
mpiret = sc_MPI_Isend (data, g2 * datasize, sc_MPI_BYTE,
myrank + g2B, SC_TAG_AG_RECURSIVE_C,
mpicomm, request + 2);
SC_CHECK_MPI (mpiret);
}
else {
request[2] = sc_MPI_REQUEST_NULL;
}
}
else {
sc_allgather_recursive (mpicomm, data + g2 * datasize, datasize, g2B,
myoffset - g2, myrank);
if (myoffset == groupsize - 1 && g2 != g2B) {
request[0] = sc_MPI_REQUEST_NULL;
request[1] = sc_MPI_REQUEST_NULL;
mpiret = sc_MPI_Irecv (data, g2 * datasize, sc_MPI_BYTE,
myrank - g2B, SC_TAG_AG_RECURSIVE_C,
mpicomm, request + 2);
SC_CHECK_MPI (mpiret);
}
else {
mpiret = sc_MPI_Irecv (data, g2 * datasize, sc_MPI_BYTE,
myrank - g2, SC_TAG_AG_RECURSIVE_A,
mpicomm, request + 0);
SC_CHECK_MPI (mpiret);
mpiret = sc_MPI_Isend (data + g2 * datasize, g2B * datasize,
sc_MPI_BYTE, myrank - g2,
SC_TAG_AG_RECURSIVE_B, mpicomm, request + 1);
SC_CHECK_MPI (mpiret);
request[2] = sc_MPI_REQUEST_NULL;
}
}
mpiret = sc_MPI_Waitall (3, request, sc_MPI_STATUSES_IGNORE);
SC_CHECK_MPI (mpiret);
}
else {
sc_allgather_alltoall (mpicomm, data, datasize, groupsize, myoffset,
myrank);
}
}