+ int rank;
+ int retval = MPI_SUCCESS;
+ smpi_mpi_request_t request;
+
+ rank = smpi_mpi_comm_rank(comm);
+ if (rank == root) {
+ retval = smpi_create_request(buf, count, datatype, root,
+ (root + 1) % comm->size, 0, comm, &request);
+ request->forward = comm->size - 1;
+ smpi_mpi_isend(request);
+ } else {
+ retval = smpi_create_request(buf, count, datatype, MPI_ANY_SOURCE, rank,
+ 0, comm, &request);
+ smpi_mpi_irecv(request);
+ }
+
+ smpi_mpi_wait(request, MPI_STATUS_IGNORE);
+ xbt_mallocator_release(smpi_global->request_mallocator, request);
+
+ return(retval);
+
+}
+/**
+ * Bcast internal level
+ **/
+int smpi_mpi_bcast(void *buf, int count, MPI_Datatype datatype, int root,
+ MPI_Comm comm)
+{
+ int retval = MPI_SUCCESS;
+ //retval = flat_tree_bcast(buf, count, datatype, root, comm);
+ retval = nary_tree_bcast(buf, count, datatype, root, comm, 2 );
+ return retval;
+}
+
+/**
+ * Bcast user entry point
+ **/
+int SMPI_MPI_Bcast(void *buf, int count, MPI_Datatype datatype, int root,
+ MPI_Comm comm)
+{
+ int retval = MPI_SUCCESS;
+
+ smpi_bench_end();
+ smpi_mpi_bcast(buf,count,datatype,root,comm);
+ smpi_bench_begin();
+
+ return retval;
+}
+
+
+
+#ifdef DEBUG_REDUCE
+/**
+ * debugging helper function
+ **/
+static void print_buffer_int(void *buf, int len, char *msg, int rank)
+{
+ int tmp, *v;
+ printf("**[%d] %s: ", rank, msg);
+ for (tmp = 0; tmp < len; tmp++) {
+ v = buf;
+ printf("[%d]", v[tmp]);
+ }
+ printf("\n");
+ free(msg);
+}
+static void print_buffer_double(void *buf, int len, char *msg, int rank)
+{
+ int tmp;
+ double *v;
+ printf("**[%d] %s: ", rank, msg);
+ for (tmp = 0; tmp < len; tmp++) {
+ v = buf;
+ printf("[%lf]", v[tmp]);
+ }
+ printf("\n");
+ free(msg);
+}
+
+
+#endif
+/**
+ * MPI_Reduce internal level
+ **/
+int smpi_mpi_reduce(void *sendbuf, void *recvbuf, int count,
+ MPI_Datatype datatype, MPI_Op op, int root, MPI_Comm comm)
+{
+ int retval = MPI_SUCCESS;
+ int rank;
+ int size;
+ int i;
+ int tag = 0;
+ smpi_mpi_request_t *requests;
+ smpi_mpi_request_t request;
+
+ smpi_bench_end();
+
+ rank = smpi_mpi_comm_rank(comm);
+ size = comm->size;
+
+ if (rank != root) { // if i am not ROOT, simply send my buffer to root
+
+#ifdef DEBUG_REDUCE
+ print_buffer_int(sendbuf, count, xbt_strdup("sndbuf"), rank);
+#endif
+ retval = smpi_create_request(sendbuf, count, datatype, rank, root, tag, comm,
+ &request);
+ smpi_mpi_isend(request);
+ smpi_mpi_wait(request, MPI_STATUS_IGNORE);
+ xbt_mallocator_release(smpi_global->request_mallocator, request);
+
+ } else {
+ // i am the ROOT: wait for all buffers by creating one request by sender
+ int src;
+ requests = xbt_malloc((size-1) * sizeof(smpi_mpi_request_t));
+
+ void **tmpbufs = xbt_malloc((size-1) * sizeof(void *));
+ for (i = 0; i < size-1; i++) {
+ // we need 1 buffer per request to store intermediate receptions
+ tmpbufs[i] = xbt_malloc(count * datatype->size);
+ }
+ // root: initiliaze recv buf with my own snd buf
+ memcpy(recvbuf, sendbuf, count * datatype->size * sizeof(char));
+
+ // i can not use: 'request->forward = size-1;' (which would progagate size-1 receive reqs)
+ // since we should op values as soon as one receiving request matches.
+ for (i = 0; i < size-1; i++) {
+ // reminder: for smpi_create_request() the src is always the process sending.
+ src = i < root ? i : i + 1;
+ retval = smpi_create_request(tmpbufs[i], count, datatype,
+ src, root, tag, comm, &(requests[i]));
+ if (NULL != requests[i] && MPI_SUCCESS == retval) {
+ if (MPI_SUCCESS == retval) {
+ smpi_mpi_irecv(requests[i]);
+ }
+ }
+ }
+ // now, wait for completion of all irecv's.
+ for (i = 0; i < size-1; i++) {
+ int index = MPI_UNDEFINED;
+ smpi_mpi_waitany( size-1, requests, &index, MPI_STATUS_IGNORE);
+#ifdef DEBUG_REDUCE
+ printf ("MPI_Waitany() unblocked: root received (completes req[index=%d])\n",index);
+ print_buffer_int(tmpbufs[index], count, bprintf("tmpbufs[index=%d] (value received)", index),
+ rank);
+#endif
+
+ // arg 2 is modified
+ op->func(tmpbufs[index], recvbuf, &count, &datatype);
+#ifdef DEBUG_REDUCE
+ print_buffer_int(recvbuf, count, xbt_strdup("rcvbuf"), rank);
+#endif
+ xbt_free(tmpbufs[index]);
+ /* FIXME: with the following line, it generates an
+ * [xbt_ex/CRITICAL] Conditional list not empty 162518800.
+ */
+ // xbt_mallocator_release(smpi_global->request_mallocator, requests[index]);
+ }
+ xbt_free(requests);
+ xbt_free(tmpbufs);
+ }
+ return retval;
+}
+
+/**
+ * MPI_Reduce user entry point
+ **/
+int SMPI_MPI_Reduce(void *sendbuf, void *recvbuf, int count,
+ MPI_Datatype datatype, MPI_Op op, int root, MPI_Comm comm)
+{
+int retval = MPI_SUCCESS;
+
+ smpi_bench_end();
+
+ retval = smpi_mpi_reduce(sendbuf, recvbuf, count, datatype, op, root, comm);
+
+ smpi_bench_begin();
+ return retval;
+}
+
+
+
+/**
+ * MPI_Allreduce
+ *
+ * Same as MPI_REDUCE except that the result appears in the receive buffer of all the group members.
+ **/
+int SMPI_MPI_Allreduce( void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype,
+ MPI_Op op, MPI_Comm comm )
+{
+int retval = MPI_SUCCESS;
+int root=0; // arbitrary choice
+
+ smpi_bench_end();
+
+ retval = smpi_mpi_reduce( sendbuf, recvbuf, count, datatype, op, root, comm);
+ if (MPI_SUCCESS != retval)
+ return(retval);
+
+ retval = smpi_mpi_bcast( sendbuf, count, datatype, root, comm);
+
+ smpi_bench_end();
+ return( retval );
+}
+
+
+/**
+ * MPI_Scatter user entry point
+ **/
+int SMPI_MPI_Scatter(void *sendbuf, int sendcount, MPI_Datatype datatype,
+ void *recvbuf, int recvcount, MPI_Datatype recvtype,
+ int root, MPI_Comm comm)
+{
+ int retval = MPI_SUCCESS;
+ int i;
+ int cnt=0;
+ int rank;
+ int tag=0;
+ char *cptr; // to manipulate the void * buffers
+ smpi_mpi_request_t *requests;
+ smpi_mpi_request_t request;
+ smpi_mpi_status_t status;
+
+
+ smpi_bench_end();
+
+ rank = smpi_mpi_comm_rank(comm);
+
+ requests = xbt_malloc((comm->size-1) * sizeof(smpi_mpi_request_t));
+ if (rank == root) {
+ // i am the root: distribute my sendbuf
+ //print_buffer_int(sendbuf, comm->size, xbt_strdup("rcvbuf"), rank);
+ cptr = sendbuf;
+ for (i=0; i < comm->size; i++) {
+ if ( i!=root ) { // send to processes ...
+
+ retval = smpi_create_request((void *)cptr, sendcount,
+ datatype, root, i, tag, comm, &(requests[cnt]));
+ if (NULL != requests[cnt] && MPI_SUCCESS == retval) {
+ if (MPI_SUCCESS == retval) {
+ smpi_mpi_isend(requests[cnt]);
+ }
+ }
+ cnt++;
+ }
+ else { // ... except if it's me.
+ memcpy(recvbuf, (void *)cptr, recvcount*recvtype->size*sizeof(char));