- int retval = MPI_SUCCESS;
-
- int index, rank;
- smpi_mpi_request_t request;
- int colorkey[2];
- smpi_mpi_status_t status;
-
- smpi_bench_end();
-
- // FIXME: need to test parameters
-
- index = smpi_host_index();
- rank = comm->index_to_rank_map[index];
-
- if (0 == rank) {
- int *colors = xbt_new(int, comm->size);
- int *keys = xbt_new(int, comm->size);
- int i, j, k;
- smpi_mpi_communicator_t tempcomm = NULL;
- int colortmp;
- int keycount;
- int *keystmp = xbt_new(int, comm->size);
- int *rankstmp = xbt_new(int, comm->size);
- int tmpval;
- int indextmp;
-
- colors[0] = color;
- keys[0] = key;
-
- // FIXME: not efficient
- for (i = 1; i < comm->size; i++) {
- retval = smpi_create_request(colorkey, 2, MPI_INT, MPI_ANY_SOURCE, rank, MPI_ANY_TAG, comm, &request);
- smpi_mpi_irecv(request);
- smpi_mpi_wait(request, &status);
- colors[status.MPI_SOURCE] = colorkey[0];
- keys[status.MPI_SOURCE] = colorkey[1];
- xbt_mallocator_release(smpi_global->request_mallocator, request);
- }
-
- for (i = 0; i < comm->size; i++) {
- if (-1 == colors[i]) {
- continue;
- }
- colortmp = colors[i];
- keycount = 0;
- for (j = i; j < comm->size; j++) {
- if(colortmp == colors[j]) {
- colors[j] = -1;
- keystmp[keycount] = keys[j];
- rankstmp[keycount] = j;
- keycount++;
- }
- }
- // FIXME: yes, mock me, bubble sort...
- for (j = 0; j < keycount; j++) {
- for (k = j; k < keycount - 1; k++) {
- if (keystmp[k] > keystmp[k + 1]) {
- tmpval = keystmp[k];
- keystmp[k] = keystmp[k + 1];
- keystmp[k + 1] = tmpval;
-
- tmpval = rankstmp[k];
- rankstmp[k] = rankstmp[k + 1];
- rankstmp[k + 1] = tmpval;
- }
- }
- }
- tempcomm = xbt_new(s_smpi_mpi_communicator_t, 1);
- tempcomm->barrier_count = 0;
- tempcomm->size = keycount;
- tempcomm->barrier_mutex = SIMIX_mutex_init();
- tempcomm->barrier_cond = SIMIX_cond_init();
- tempcomm->rank_to_index_map = xbt_new(int, keycount);
- tempcomm->index_to_rank_map = xbt_new(int, smpi_global->host_count);
- for (j = 0; j < smpi_global->host_count; j++) {
- tempcomm->index_to_rank_map[j] = -1;
- }
- for (j = 0; j < keycount; j++) {
- indextmp = comm->rank_to_index_map[rankstmp[j]];
- tempcomm->rank_to_index_map[j] = indextmp;
- tempcomm->index_to_rank_map[indextmp] = j;
+ int tmp, *v;
+ printf("**[%d] %s: ", rank, msg);
+ for (tmp = 0; tmp < len; tmp++) {
+ v = buf;
+ printf("[%d]", 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);
+ }
+ 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 SMPI_MPI_Allreduce( void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype,
+ MPI_Op op, MPI_Comm comm )
+{
+ int retval = MPI_SUCCESS;
+ int root=1; // 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_begin();
+ 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 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 *cbuf; // 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
+ for (i=0; i < comm->size; i++) {
+ cbuf = sendbuf;
+ cbuf += i*sendcount*datatype->size;
+ if ( i!=root ) { // send to processes ...
+
+ retval = smpi_create_request((void *)cbuf, 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 *)cbuf, recvcount*recvtype->size*sizeof(char));