int Colls::ibarrier(MPI_Comm comm, MPI_Request* request)
{
- int i;
int size = comm->size();
int rank = comm->rank();
- MPI_Request* requests;
(*request) = new Request( nullptr, 0, MPI_BYTE,
rank,rank, COLL_TAG_BARRIER, comm, MPI_REQ_PERSISTENT);
if (rank > 0) {
- requests = new MPI_Request[2];
+ MPI_Request* requests = new MPI_Request[2];
requests[0] = Request::isend (nullptr, 0, MPI_BYTE, 0,
COLL_TAG_BARRIER,
comm);
(*request)->set_nbc_requests(requests, 2);
}
else {
- requests = new MPI_Request[(size-1)*2];
- for (i = 1; i < 2*size-1; i+=2) {
- requests[i-1] = Request::irecv(nullptr, 0, MPI_BYTE, MPI_ANY_SOURCE,
- COLL_TAG_BARRIER, comm
- );
- requests[i] = Request::isend(nullptr, 0, MPI_BYTE, (i+1)/2,
- COLL_TAG_BARRIER,
- comm
- );
+ MPI_Request* requests = new MPI_Request[(size - 1) * 2];
+ for (int i = 1; i < 2 * size - 1; i += 2) {
+ requests[i - 1] = Request::irecv(nullptr, 0, MPI_BYTE, MPI_ANY_SOURCE, COLL_TAG_BARRIER, comm);
+ requests[i] = Request::isend(nullptr, 0, MPI_BYTE, (i + 1) / 2, COLL_TAG_BARRIER, comm);
}
(*request)->set_nbc_requests(requests, 2*(size-1));
}
int Colls::ibcast(void *buf, int count, MPI_Datatype datatype, int root, MPI_Comm comm, MPI_Request* request)
{
- int i;
int size = comm->size();
int rank = comm->rank();
- MPI_Request* requests;
(*request) = new Request( nullptr, 0, MPI_BYTE,
- rank,rank, COLL_TAG_BARRIER, comm, MPI_REQ_PERSISTENT);
+ rank,rank, COLL_TAG_BCAST, comm, MPI_REQ_PERSISTENT);
if (rank != root) {
- requests = new MPI_Request[1];
+ MPI_Request* requests = new MPI_Request[1];
requests[0] = Request::irecv (buf, count, datatype, root,
COLL_TAG_BCAST,
comm);
(*request)->set_nbc_requests(requests, 1);
}
else {
- requests = new MPI_Request[size-1];
+ MPI_Request* requests = new MPI_Request[size - 1];
int n = 0;
- for (i = 0; i < size; i++) {
+ for (int i = 0; i < size; i++) {
if(i!=root){
requests[n] = Request::isend(buf, count, datatype, i,
COLL_TAG_BCAST,
const int system_tag = COLL_TAG_ALLGATHER;
MPI_Aint lb = 0;
MPI_Aint recvext = 0;
- MPI_Request *requests;
int rank = comm->rank();
int size = comm->size();
(*request) = new Request( nullptr, 0, MPI_BYTE,
- rank,rank, COLL_TAG_BARRIER, comm, MPI_REQ_PERSISTENT);
+ rank,rank, system_tag, comm, MPI_REQ_PERSISTENT);
// FIXME: check for errors
recvtype->extent(&lb, &recvext);
// Local copy from self
Datatype::copy(sendbuf, sendcount, sendtype, static_cast<char *>(recvbuf) + rank * recvcount * recvext, recvcount,
recvtype);
// Send/Recv buffers to/from others;
- requests = new MPI_Request[2 * (size - 1)];
+ MPI_Request* requests = new MPI_Request[2 * (size - 1)];
int index = 0;
for (int other = 0; other < size; other++) {
if(other != rank) {
const int system_tag = COLL_TAG_SCATTER;
MPI_Aint lb = 0;
MPI_Aint sendext = 0;
- MPI_Request *requests;
int rank = comm->rank();
int size = comm->size();
(*request) = new Request( nullptr, 0, MPI_BYTE,
- rank,rank, COLL_TAG_BARRIER, comm, MPI_REQ_PERSISTENT);
+ rank,rank, system_tag, comm, MPI_REQ_PERSISTENT);
if(rank != root) {
- requests = new MPI_Request[1];
+ MPI_Request* requests = new MPI_Request[1];
// Recv buffer from root
requests[0] = Request::irecv(recvbuf, recvcount, recvtype, root, system_tag, comm);
(*request)->set_nbc_requests(requests, 1);
sendcount, sendtype, recvbuf, recvcount, recvtype);
}
// Send buffers to receivers
- requests = new MPI_Request[size - 1];
+ MPI_Request* requests = new MPI_Request[size - 1];
int index = 0;
for(int dst = 0; dst < size; dst++) {
if(dst != root) {
int rank = comm->rank();
int size = comm->size();
(*request) = new Request( nullptr, 0, MPI_BYTE,
- rank,rank, COLL_TAG_BARRIER, comm, MPI_REQ_PERSISTENT);
+ rank,rank, system_tag, comm, MPI_REQ_PERSISTENT);
recvtype->extent(&lb, &recvext);
// Local copy from self
Datatype::copy(sendbuf, sendcount, sendtype,
}
int Colls::ialltoall( void *sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, int recvcount, MPI_Datatype recvtype, MPI_Comm comm, MPI_Request* request){
-int system_tag = COLL_TAG_ALLTOALL;
- int i;
- int count;
- MPI_Aint lb = 0, sendext = 0, recvext = 0;
- MPI_Request *requests;
+ int system_tag = COLL_TAG_ALLTOALL;
+ MPI_Aint lb = 0;
+ MPI_Aint sendext = 0;
+ MPI_Aint recvext = 0;
/* Initialize. */
int rank = comm->rank();
int size = comm->size();
(*request) = new Request( nullptr, 0, MPI_BYTE,
- rank,rank, COLL_TAG_ALLTOALL, comm, MPI_REQ_PERSISTENT);
+ rank,rank, system_tag, comm, MPI_REQ_PERSISTENT);
sendtype->extent(&lb, &sendext);
recvtype->extent(&lb, &recvext);
/* simple optimization */
static_cast<char *>(recvbuf) + rank * recvcount * recvext, recvcount, recvtype);
if (err == MPI_SUCCESS && size > 1) {
/* Initiate all send/recv to/from others. */
- requests = new MPI_Request[2 * (size - 1)];
+ MPI_Request* requests = new MPI_Request[2 * (size - 1)];
/* Post all receives first -- a simple optimization */
- count = 0;
- for (i = (rank + 1) % size; i != rank; i = (i + 1) % size) {
+ int count = 0;
+ for (int i = (rank + 1) % size; i != rank; i = (i + 1) % size) {
requests[count] = Request::irecv_init(static_cast<char *>(recvbuf) + i * recvcount * recvext, recvcount,
recvtype, i, system_tag, comm);
count++;
* when messages actually arrive in the order in which they were posted.
* TODO: check the previous assertion
*/
- for (i = (rank + size - 1) % size; i != rank; i = (i + size - 1) % size) {
+ for (int i = (rank + size - 1) % size; i != rank; i = (i + size - 1) % size) {
requests[count] = Request::isend_init(static_cast<char *>(sendbuf) + i * sendcount * sendext, sendcount,
sendtype, i, system_tag, comm);
count++;
MPI_Aint lb = 0;
MPI_Aint sendext = 0;
MPI_Aint recvext = 0;
- MPI_Request *requests;
/* Initialize. */
int rank = comm->rank();
int size = comm->size();
(*request) = new Request( nullptr, 0, MPI_BYTE,
- rank,rank, COLL_TAG_ALLTOALLV, comm, MPI_REQ_PERSISTENT);
+ rank,rank, system_tag, comm, MPI_REQ_PERSISTENT);
sendtype->extent(&lb, &sendext);
recvtype->extent(&lb, &recvext);
/* Local copy from self */
static_cast<char *>(recvbuf) + recvdisps[rank] * recvext, recvcounts[rank], recvtype);
if (err == MPI_SUCCESS && size > 1) {
/* Initiate all send/recv to/from others. */
- requests = new MPI_Request[2 * (size - 1)];
+ MPI_Request* requests = new MPI_Request[2 * (size - 1)];
int count = 0;
/* Create all receives that will be posted first */
for (int i = 0; i < size; ++i) {
int Colls::ialltoallw(void *sendbuf, int *sendcounts, int *senddisps, MPI_Datatype* sendtypes,
void *recvbuf, int *recvcounts, int *recvdisps, MPI_Datatype* recvtypes, MPI_Comm comm, MPI_Request *request){
const int system_tag = COLL_TAG_ALLTOALLV;
- MPI_Request *requests;
/* Initialize. */
int rank = comm->rank();
int size = comm->size();
(*request) = new Request( nullptr, 0, MPI_BYTE,
- rank,rank, COLL_TAG_ALLTOALLV, comm, MPI_REQ_PERSISTENT);
+ rank,rank, system_tag, comm, MPI_REQ_PERSISTENT);
/* Local copy from self */
int err = (sendcounts[rank]>0 && recvcounts[rank]) ? Datatype::copy(static_cast<char *>(sendbuf) + senddisps[rank], sendcounts[rank], sendtypes[rank],
static_cast<char *>(recvbuf) + recvdisps[rank], recvcounts[rank], recvtypes[rank]): MPI_SUCCESS;
if (err == MPI_SUCCESS && size > 1) {
/* Initiate all send/recv to/from others. */
- requests = new MPI_Request[2 * (size - 1)];
+ MPI_Request* requests = new MPI_Request[2 * (size - 1)];
int count = 0;
/* Create all receives that will be posted first */
for (int i = 0; i < size; ++i) {
const int system_tag = COLL_TAG_GATHER;
MPI_Aint lb = 0;
MPI_Aint recvext = 0;
- MPI_Request *requests;
int rank = comm->rank();
int size = comm->size();
(*request) = new Request( nullptr, 0, MPI_BYTE,
- rank,rank, COLL_TAG_GATHER, comm, MPI_REQ_PERSISTENT);
+ rank,rank, system_tag, comm, MPI_REQ_PERSISTENT);
if(rank != root) {
// Send buffer to root
- requests = new MPI_Request[1];
+ MPI_Request* requests = new MPI_Request[1];
requests[0]=Request::isend(sendbuf, sendcount, sendtype, root, system_tag, comm);
(*request)->set_nbc_requests(requests, 1);
} else {
Datatype::copy(sendbuf, sendcount, sendtype, static_cast<char*>(recvbuf) + root * recvcount * recvext,
recvcount, recvtype);
// Receive buffers from senders
- requests = new MPI_Request[size - 1];
+ MPI_Request* requests = new MPI_Request[size - 1];
int index = 0;
for (int src = 0; src < size; src++) {
if(src != root) {
int system_tag = COLL_TAG_GATHERV;
MPI_Aint lb = 0;
MPI_Aint recvext = 0;
- MPI_Request *requests;
int rank = comm->rank();
int size = comm->size();
(*request) = new Request( nullptr, 0, MPI_BYTE,
- rank,rank, COLL_TAG_GATHERV, comm, MPI_REQ_PERSISTENT);
+ rank,rank, system_tag, comm, MPI_REQ_PERSISTENT);
if (rank != root) {
// Send buffer to root
- requests = new MPI_Request[1];
+ MPI_Request* requests = new MPI_Request[1];
requests[0]=Request::isend(sendbuf, sendcount, sendtype, root, system_tag, comm);
(*request)->set_nbc_requests(requests, 1);
} else {
Datatype::copy(sendbuf, sendcount, sendtype, static_cast<char*>(recvbuf) + displs[root] * recvext,
recvcounts[root], recvtype);
// Receive buffers from senders
- requests = new MPI_Request[size - 1];
+ MPI_Request* requests = new MPI_Request[size - 1];
int index = 0;
for (int src = 0; src < size; src++) {
if(src != root) {
int system_tag = COLL_TAG_SCATTERV;
MPI_Aint lb = 0;
MPI_Aint sendext = 0;
- MPI_Request* requests;
int rank = comm->rank();
int size = comm->size();
(*request) = new Request( nullptr, 0, MPI_BYTE,
- rank,rank, COLL_TAG_SCATTERV, comm, MPI_REQ_PERSISTENT);
+ rank,rank, system_tag, comm, MPI_REQ_PERSISTENT);
if(rank != root) {
// Recv buffer from root
- requests = new MPI_Request[1];
+ MPI_Request* requests = new MPI_Request[1];
requests[0]=Request::irecv(recvbuf, recvcount, recvtype, root, system_tag, comm);
(*request)->set_nbc_requests(requests, 1);
} else {
}
return MPI_SUCCESS;
}
+
+int Colls::ireduce(void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, int root,
+ MPI_Comm comm, MPI_Request* request)
+{
+ const int system_tag = COLL_TAG_REDUCE;
+ MPI_Aint lb = 0;
+ MPI_Aint dataext = 0;
+
+ char* sendtmpbuf = static_cast<char *>(sendbuf);
+
+ int rank = comm->rank();
+ int size = comm->size();
+
+ if (size <= 0)
+ return MPI_ERR_COMM;
+
+ if( sendbuf == MPI_IN_PLACE ) {
+ sendtmpbuf = static_cast<char *>(smpi_get_tmp_sendbuffer(count*datatype->get_extent()));
+ Datatype::copy(recvbuf, count, datatype,sendtmpbuf, count, datatype);
+ }
+
+ if(rank == root){
+ (*request) = new Request( recvbuf, count, datatype,
+ rank,rank, system_tag, comm, MPI_REQ_PERSISTENT, op);
+ }
+ else
+ (*request) = new Request( nullptr, count, datatype,
+ rank,rank, system_tag, comm, MPI_REQ_PERSISTENT);
+
+ if(rank != root) {
+ // Send buffer to root
+ MPI_Request* requests = new MPI_Request[1];
+ requests[0]=Request::isend(sendtmpbuf, count, datatype, root, system_tag, comm);
+ (*request)->set_nbc_requests(requests, 1);
+ } else {
+ datatype->extent(&lb, &dataext);
+ // Local copy from root
+ if (sendtmpbuf != nullptr && recvbuf != nullptr)
+ Datatype::copy(sendtmpbuf, count, datatype, recvbuf, count, datatype);
+ // Receive buffers from senders
+ MPI_Request *requests = new MPI_Request[size - 1];
+ int index = 0;
+ for (int src = 0; src < size; src++) {
+ if (src != root) {
+ requests[index] =
+ Request::irecv_init(smpi_get_tmp_sendbuffer(count * dataext), count, datatype, src, system_tag, comm);
+ index++;
+ }
+ }
+ // Wait for completion of irecv's.
+ Request::startall(size - 1, requests);
+ (*request)->set_nbc_requests(requests, size - 1);
+ }
+ if( sendbuf == MPI_IN_PLACE ) {
+ smpi_free_tmp_buffer(sendtmpbuf);
+ }
+ return MPI_SUCCESS;
+}
+
+int Colls::iallreduce(void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype,
+ MPI_Op op, MPI_Comm comm, MPI_Request* request)
+{
+
+ const int system_tag = COLL_TAG_ALLREDUCE;
+ MPI_Aint lb = 0;
+ MPI_Aint dataext = 0;
+
+ int rank = comm->rank();
+ int size = comm->size();
+ (*request) = new Request( recvbuf, count, datatype,
+ rank,rank, system_tag, comm, MPI_REQ_PERSISTENT, op);
+ // FIXME: check for errors
+ datatype->extent(&lb, &dataext);
+ // Local copy from self
+ Datatype::copy(sendbuf, count, datatype, recvbuf, count, datatype);
+ // Send/Recv buffers to/from others;
+ MPI_Request* requests = new MPI_Request[2 * (size - 1)];
+ int index = 0;
+ for (int other = 0; other < size; other++) {
+ if(other != rank) {
+ requests[index] = Request::isend_init(sendbuf, count, datatype, other, system_tag,comm);
+ index++;
+ requests[index] = Request::irecv_init(smpi_get_tmp_sendbuffer(count * dataext), count, datatype,
+ other, system_tag, comm);
+ index++;
+ }
+ }
+ Request::startall(2 * (size - 1), requests);
+ (*request)->set_nbc_requests(requests, 2 * (size - 1));
+ return MPI_SUCCESS;
+}
+
+int Colls::iscan(void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm, MPI_Request* request)
+{
+ int system_tag = -888;
+ MPI_Aint lb = 0;
+ MPI_Aint dataext = 0;
+
+ int rank = comm->rank();
+ int size = comm->size();
+ (*request) = new Request( recvbuf, count, datatype,
+ rank,rank, system_tag, comm, MPI_REQ_PERSISTENT, op);
+ datatype->extent(&lb, &dataext);
+
+ // Local copy from self
+ Datatype::copy(sendbuf, count, datatype, recvbuf, count, datatype);
+
+ // Send/Recv buffers to/from others
+ MPI_Request *requests = new MPI_Request[size - 1];
+ int index = 0;
+ for (int other = 0; other < rank; other++) {
+ requests[index] = Request::irecv_init(smpi_get_tmp_sendbuffer(count * dataext), count, datatype, other, system_tag, comm);
+ index++;
+ }
+ for (int other = rank + 1; other < size; other++) {
+ requests[index] = Request::isend_init(sendbuf, count, datatype, other, system_tag, comm);
+ index++;
+ }
+ // Wait for completion of all comms.
+ Request::startall(size - 1, requests);
+ (*request)->set_nbc_requests(requests, size - 1);
+ return MPI_SUCCESS;
+}
+
+int Colls::iexscan(void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm, MPI_Request* request)
+{
+ int system_tag = -888;
+ MPI_Aint lb = 0;
+ MPI_Aint dataext = 0;
+ int rank = comm->rank();
+ int size = comm->size();
+ (*request) = new Request( recvbuf, count, datatype,
+ rank,rank, system_tag, comm, MPI_REQ_PERSISTENT, op);
+ datatype->extent(&lb, &dataext);
+ if(rank != 0)
+ memset(recvbuf, 0, count*dataext);
+
+ // Send/Recv buffers to/from others
+ MPI_Request *requests = new MPI_Request[size - 1];
+ int index = 0;
+ for (int other = 0; other < rank; other++) {
+ requests[index] = Request::irecv_init(smpi_get_tmp_sendbuffer(count * dataext), count, datatype, other, system_tag, comm);
+ index++;
+ }
+ for (int other = rank + 1; other < size; other++) {
+ requests[index] = Request::isend_init(sendbuf, count, datatype, other, system_tag, comm);
+ index++;
+ }
+ // Wait for completion of all comms.
+ Request::startall(size - 1, requests);
+ (*request)->set_nbc_requests(requests, size - 1);
+ return MPI_SUCCESS;
+}
+
+int Colls::ireduce_scatter(void *sendbuf, void *recvbuf, int *recvcounts, MPI_Datatype datatype, MPI_Op op,
+ MPI_Comm comm, MPI_Request* request){
+//Version where each process performs the reduce for its own part. Alltoall pattern for comms.
+ const int system_tag = COLL_TAG_REDUCE_SCATTER;
+ MPI_Aint lb = 0;
+ MPI_Aint dataext = 0;
+
+ int rank = comm->rank();
+ int size = comm->size();
+ int count=recvcounts[rank];
+ (*request) = new Request( recvbuf, count, datatype,
+ rank,rank, system_tag, comm, MPI_REQ_PERSISTENT, op);
+ datatype->extent(&lb, &dataext);
+
+ // Send/Recv buffers to/from others;
+ MPI_Request* requests = new MPI_Request[2 * (size - 1)];
+ int index = 0;
+ int recvdisp=0;
+ for (int other = 0; other < size; other++) {
+ if(other != rank) {
+ requests[index] = Request::isend_init(static_cast<char *>(sendbuf) + recvdisp * dataext, recvcounts[other], datatype, other, system_tag,comm);
+ XBT_VERB("sending with recvdisp %d", recvdisp);
+ index++;
+ requests[index] = Request::irecv_init(smpi_get_tmp_sendbuffer(count * dataext), count, datatype,
+ other, system_tag, comm);
+ index++;
+ }else{
+ Datatype::copy(static_cast<char *>(sendbuf) + recvdisp * dataext, count, datatype, recvbuf, count, datatype);
+ }
+ recvdisp+=recvcounts[other];
+ }
+ Request::startall(2 * (size - 1), requests);
+ (*request)->set_nbc_requests(requests, 2 * (size - 1));
+ return MPI_SUCCESS;
+}
+
}
}