+void smpi_mpi_wait(MPI_Request* request, MPI_Status* status) {
+ print_request("wait", *request);
+ SIMIX_network_wait((*request)->pair, -1.0);
+ finish_wait(request, status);
+}
+
+int smpi_mpi_waitany(int count, MPI_Request requests[], MPI_Status* status) {
+ xbt_dynar_t comms;
+ int i, size, index;
+ int* map;
+
+ index = MPI_UNDEFINED;
+ if(count > 0) {
+ // First check for already completed requests
+ for(i = 0; i < count; i++) {
+ if(requests[i] != MPI_REQUEST_NULL && requests[i]->complete) {
+ index = i;
+ smpi_mpi_wait(&requests[index], status);
+ break;
+ }
+ }
+ if(index == MPI_UNDEFINED) {
+ // Otherwise, wait for a request to complete
+ comms = xbt_dynar_new(sizeof(smx_comm_t), NULL);
+ map = xbt_new(int, count);
+ size = 0;
+ DEBUG0("Wait for one of");
+ for(i = 0; i < count; i++) {
+ if(requests[i] != MPI_REQUEST_NULL && requests[i]->complete == 0) {
+ print_request(" ", requests[i]);
+ xbt_dynar_push(comms, &requests[i]->pair);
+ map[size] = i;
+ size++;
+ }
+ }
+ if(size > 0) {
+ index = SIMIX_network_waitany(comms);
+ index = map[index];
+ finish_wait(&requests[index], status);
+ }
+ xbt_free(map);
+ xbt_dynar_free(&comms);
+ }
+ }
+ return index;
+}
+
+void smpi_mpi_waitall(int count, MPI_Request requests[], MPI_Status status[]) {
+ int index;
+ MPI_Status stat;
+
+ while(count > 0) {
+ index = smpi_mpi_waitany(count, requests, &stat);
+ if(index == MPI_UNDEFINED) {
+ break;
+ }
+ if(status != MPI_STATUS_IGNORE) {
+ memcpy(&status[index], &stat, sizeof(stat));
+ }
+ // FIXME: check this -v
+ // Move the last request to the found position
+ requests[index] = requests[count - 1];
+ requests[count - 1] = MPI_REQUEST_NULL;
+ count--;
+ }
+}
+
+int smpi_mpi_waitsome(int incount, MPI_Request requests[], int* indices, MPI_Status status[]) {
+ int i, count;
+
+ count = 0;
+ for(i = 0; i < incount; i++) {
+ if(requests[i] != MPI_REQUEST_NULL && requests[i]->complete) {
+ smpi_mpi_wait(&requests[i], status != MPI_STATUS_IGNORE ? &status[i] : MPI_STATUS_IGNORE);
+ indices[count] = i;
+ count++;
+ }
+ }
+ return count;
+}
+
+void smpi_mpi_bcast(void* buf, int count, MPI_Datatype datatype, int root, MPI_Comm comm) {
+ // arity=2: a binary tree, arity=4 seem to be a good setting (see P2P-MPI))
+ nary_tree_bcast(buf, count, datatype, root, comm, 4);
+}
+
+void smpi_mpi_barrier(MPI_Comm comm) {
+ // arity=2: a binary tree, arity=4 seem to be a good setting (see P2P-MPI))
+ nary_tree_barrier(comm, 4);
+}
+
+void smpi_mpi_gather(void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, int recvcount, MPI_Datatype recvtype, int root, MPI_Comm comm) {
+ int system_tag = 666;
+ int rank, size, src, index, sendsize, recvsize;
+ MPI_Request* requests;
+
+ rank = smpi_comm_rank(comm);
+ size = smpi_comm_size(comm);
+ if(rank != root) {
+ // Send buffer to root
+ smpi_mpi_send(sendbuf, sendcount, sendtype, root, system_tag, comm);
+ } else {
+ sendsize = smpi_datatype_size(sendtype);
+ recvsize = smpi_datatype_size(recvtype);
+ // Local copy from root
+ memcpy(&((char*)recvbuf)[root * recvcount * recvsize], sendbuf, sendcount * sendsize * sizeof(char));
+ // Receive buffers from senders
+ requests = xbt_new(MPI_Request, size - 1);
+ index = 0;
+ for(src = 0; src < size; src++) {
+ if(src != root) {
+ requests[index] = smpi_irecv_init(&((char*)recvbuf)[src * recvcount * recvsize], recvcount, recvtype, src, system_tag, comm);
+ index++;
+ }
+ }
+ // Wait for completion of irecv's.
+ smpi_mpi_startall(size - 1, requests);
+ smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
+ xbt_free(requests);
+ }
+}
+
+void smpi_mpi_gatherv(void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, int* recvcounts, int* displs, MPI_Datatype recvtype, int root, MPI_Comm comm) {
+ int system_tag = 666;
+ int rank, size, src, index, sendsize;
+ MPI_Request* requests;
+
+ rank = smpi_comm_rank(comm);
+ size = smpi_comm_size(comm);
+ if(rank != root) {
+ // Send buffer to root
+ smpi_mpi_send(sendbuf, sendcount, sendtype, root, system_tag, comm);
+ } else {
+ sendsize = smpi_datatype_size(sendtype);
+ // Local copy from root
+ memcpy(&((char*)recvbuf)[displs[root]], sendbuf, sendcount * sendsize * sizeof(char));
+ // Receive buffers from senders
+ requests = xbt_new(MPI_Request, size - 1);
+ index = 0;
+ for(src = 0; src < size; src++) {
+ if(src != root) {
+ requests[index] = smpi_irecv_init(&((char*)recvbuf)[displs[src]], recvcounts[src], recvtype, src, system_tag, comm);
+ index++;
+ }
+ }
+ // Wait for completion of irecv's.
+ smpi_mpi_startall(size - 1, requests);
+ smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
+ xbt_free(requests);
+ }
+}
+
+void smpi_mpi_allgather(void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, int recvcount, MPI_Datatype recvtype, MPI_Comm comm) {
+ int system_tag = 666;
+ int rank, size, other, index, sendsize, recvsize;
+ MPI_Request* requests;
+
+ rank = smpi_comm_rank(comm);
+ size = smpi_comm_size(comm);
+ sendsize = smpi_datatype_size(sendtype);
+ recvsize = smpi_datatype_size(recvtype);
+ // Local copy from self
+ memcpy(&((char*)recvbuf)[rank * recvcount * recvsize], sendbuf, sendcount * sendsize * sizeof(char));
+ // Send/Recv buffers to/from others;
+ requests = xbt_new(MPI_Request, 2 * (size - 1));
+ index = 0;
+ for(other = 0; other < size; other++) {
+ if(other != rank) {
+ requests[index] = smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag, comm);
+ index++;
+ requests[index] = smpi_irecv_init(&((char*)recvbuf)[other * recvcount * recvsize], recvcount, recvtype, other, system_tag, comm);
+ index++;
+ }
+ }
+ // Wait for completion of all comms.
+ smpi_mpi_startall(2 * (size - 1), requests);
+ smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE);
+ xbt_free(requests);
+}
+
+void smpi_mpi_allgatherv(void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, int* recvcounts, int* displs, MPI_Datatype recvtype, MPI_Comm comm) {
+ int system_tag = 666;
+ int rank, size, other, index, sendsize, recvsize;
+ MPI_Request* requests;
+
+ rank = smpi_comm_rank(comm);
+ size = smpi_comm_size(comm);
+ sendsize = smpi_datatype_size(sendtype);
+ recvsize = smpi_datatype_size(recvtype);
+ // Local copy from self
+ memcpy(&((char*)recvbuf)[displs[rank]], sendbuf, sendcount * sendsize * sizeof(char));
+ // Send buffers to others;
+ requests = xbt_new(MPI_Request, 2 * (size - 1));
+ index = 0;
+ for(other = 0; other < size; other++) {
+ if(other != rank) {
+ requests[index] = smpi_isend_init(sendbuf, sendcount, sendtype, other, system_tag, comm);
+ index++;
+ requests[index] = smpi_irecv_init(&((char*)recvbuf)[displs[other]], recvcounts[other], recvtype, other, system_tag, comm);
+ index++;
+ }
+ }
+ // Wait for completion of all comms.
+ smpi_mpi_startall(2 * (size - 1), requests);
+ smpi_mpi_waitall(2 * (size - 1), requests, MPI_STATUS_IGNORE);
+ xbt_free(requests);
+}
+
+void smpi_mpi_scatter(void* sendbuf, int sendcount, MPI_Datatype sendtype, void* recvbuf, int recvcount, MPI_Datatype recvtype, int root, MPI_Comm comm) {
+ int system_tag = 666;
+ int rank, size, dst, index, sendsize, recvsize;
+ MPI_Request* requests;
+
+ rank = smpi_comm_rank(comm);
+ size = smpi_comm_size(comm);
+ if(rank != root) {
+ // Recv buffer from root
+ smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm, MPI_STATUS_IGNORE);
+ } else {
+ sendsize = smpi_datatype_size(sendtype);
+ recvsize = smpi_datatype_size(recvtype);
+ // Local copy from root
+ memcpy(recvbuf, &((char*)sendbuf)[root * sendcount * sendsize], recvcount * recvsize * sizeof(char));
+ // Send buffers to receivers
+ requests = xbt_new(MPI_Request, size - 1);
+ index = 0;
+ for(dst = 0; dst < size; dst++) {
+ if(dst != root) {
+ requests[index] = smpi_isend_init(&((char*)sendbuf)[dst * sendcount * sendsize], sendcount, sendtype, dst, system_tag, comm);
+ index++;
+ }
+ }
+ // Wait for completion of isend's.
+ smpi_mpi_startall(size - 1, requests);
+ smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
+ xbt_free(requests);
+ }
+}
+
+void smpi_mpi_scatterv(void* sendbuf, int* sendcounts, int* displs, MPI_Datatype sendtype, void* recvbuf, int recvcount, MPI_Datatype recvtype, int root, MPI_Comm comm) {
+ int system_tag = 666;
+ int rank, size, dst, index, sendsize, recvsize;
+ MPI_Request* requests;
+
+ rank = smpi_comm_rank(comm);
+ size = smpi_comm_size(comm);
+ if(rank != root) {
+ // Recv buffer from root
+ smpi_mpi_recv(recvbuf, recvcount, recvtype, root, system_tag, comm, MPI_STATUS_IGNORE);
+ } else {
+ sendsize = smpi_datatype_size(sendtype);
+ recvsize = smpi_datatype_size(recvtype);
+ // Local copy from root
+ memcpy(recvbuf, &((char*)sendbuf)[displs[root]], recvcount * recvsize * sizeof(char));
+ // Send buffers to receivers
+ requests = xbt_new(MPI_Request, size - 1);
+ index = 0;
+ for(dst = 0; dst < size; dst++) {
+ if(dst != root) {
+ requests[index] = smpi_isend_init(&((char*)sendbuf)[displs[dst]], sendcounts[dst], sendtype, dst, system_tag, comm);
+ index++;
+ }
+ }
+ // Wait for completion of isend's.
+ smpi_mpi_startall(size - 1, requests);
+ smpi_mpi_waitall(size - 1, requests, MPI_STATUS_IGNORE);
+ xbt_free(requests);
+ }
+}
+
+void smpi_mpi_reduce(void* sendbuf, void* recvbuf, int count, MPI_Datatype datatype, MPI_Op op, int root, MPI_Comm comm) {
+ int system_tag = 666;
+ int rank, size, src, index, datasize;
+ MPI_Request* requests;
+ void** tmpbufs;
+
+ rank = smpi_comm_rank(comm);
+ size = smpi_comm_size(comm);
+ if(rank != root) {
+ // Send buffer to root
+ smpi_mpi_send(sendbuf, count, datatype, root, system_tag, comm);
+ } else {
+ datasize = smpi_datatype_size(datatype);
+ // Local copy from root
+ memcpy(recvbuf, sendbuf, count * datasize * sizeof(char));
+ // Receive buffers from senders
+ //TODO: make a MPI_barrier here ?
+ requests = xbt_new(MPI_Request, size - 1);
+ tmpbufs = xbt_new(void*, size - 1);
+ index = 0;
+ for(src = 0; src < size; src++) {
+ if(src != root) {
+ tmpbufs[index] = xbt_malloc(count * datasize);
+ requests[index] = smpi_irecv_init(tmpbufs[index], count, datatype, src, system_tag, comm);
+ index++;
+ }
+ }
+ // Wait for completion of irecv's.
+ smpi_mpi_startall(size - 1, requests);
+ for(src = 0; src < size - 1; src++) {
+ index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
+ if(index == MPI_UNDEFINED) {
+ break;
+ }
+ smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
+ }
+ for(index = 0; index < size - 1; index++) {
+ xbt_free(tmpbufs[index]);
+ }
+ xbt_free(tmpbufs);
+ xbt_free(requests);
+ }
+}
+
+void smpi_mpi_allreduce(void* sendbuf, void* recvbuf, int count, MPI_Datatype datatype, MPI_Op op, MPI_Comm comm) {
+ smpi_mpi_reduce(sendbuf, recvbuf, count, datatype, op, 0, comm);
+ smpi_mpi_bcast(recvbuf, count, datatype, 0, comm);
+
+/*
+FIXME: buggy implementation
+
+ int system_tag = 666;
+ int rank, size, other, index, datasize;
+ MPI_Request* requests;
+ void** tmpbufs;
+
+ rank = smpi_comm_rank(comm);
+ size = smpi_comm_size(comm);
+ datasize = smpi_datatype_size(datatype);
+ // Local copy from self
+ memcpy(recvbuf, sendbuf, count * datasize * sizeof(char));
+ // Send/Recv buffers to/from others;
+ //TODO: make a MPI_barrier here ?
+ requests = xbt_new(MPI_Request, 2 * (size - 1));
+ tmpbufs = xbt_new(void*, size - 1);
+ index = 0;
+ for(other = 0; other < size; other++) {
+ if(other != rank) {
+ tmpbufs[index / 2] = xbt_malloc(count * datasize);
+ requests[index] = smpi_mpi_isend(sendbuf, count, datatype, other, system_tag, comm);
+ requests[index + 1] = smpi_mpi_irecv(tmpbufs[index / 2], count, datatype, other, system_tag, comm);
+ index += 2;
+ }
+ }
+ // Wait for completion of all comms.
+ for(other = 0; other < 2 * (size - 1); other++) {
+ index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
+ if(index == MPI_UNDEFINED) {
+ break;
+ }
+ if((index & 1) == 1) {
+ // Request is odd: it's a irecv
+ smpi_op_apply(op, tmpbufs[index / 2], recvbuf, &count, &datatype);
+ }
+ }
+ for(index = 0; index < size - 1; index++) {
+ xbt_free(tmpbufs[index]);
+ }
+ xbt_free(tmpbufs);
+ xbt_free(requests);
+*/
+}