-#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
- **/
-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;
+int MPI_Send(void *buf, int count, MPI_Datatype datatype, int dst, int tag,
+ MPI_Comm comm)
+{
+ int retval;
+ int rank = comm != MPI_COMM_NULL ? smpi_comm_rank(comm) : -1;
+
+ smpi_bench_end(rank, "Send");
+#ifdef HAVE_TRACING
+ int dst_traced = smpi_group_rank(smpi_comm_group(comm), dst);
+ TRACE_smpi_ptp_in(rank, rank, dst_traced, __FUNCTION__);
+ TRACE_smpi_send(rank, rank, dst_traced);
+#endif
+ if (comm == MPI_COMM_NULL) {
+ retval = MPI_ERR_COMM;
+ } else {
+ smpi_mpi_send(buf, count, datatype, dst, tag, comm);
+ retval = MPI_SUCCESS;
+ }
+#ifdef HAVE_TRACING
+ TRACE_smpi_ptp_out(rank, rank, dst_traced, __FUNCTION__);
+#endif
+ smpi_bench_begin(rank, "Send");
+ return retval;