X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/bd9bee9dcbceb25fdb5a293b91abbd5476db7ad2..54ce50db3eb883dfc77d00b9e885e786d7a6d381:/src/smpi/smpi_mpi.c

diff --git a/src/smpi/smpi_mpi.c b/src/smpi/smpi_mpi.c
index 6713ba5e7f..ca94c57c66 100644
--- a/src/smpi/smpi_mpi.c
+++ b/src/smpi/smpi_mpi.c
@@ -181,11 +181,20 @@ int SMPI_MPI_Send(void *buf, int count, MPI_Datatype datatype, int dst,
   return retval;
 }
 
+/**
+ * MPI_Wait and friends
+ **/
 int SMPI_MPI_Wait(MPI_Request * request, MPI_Status * status)
 {
   return smpi_mpi_wait(*request, status);
 }
 
+int SMPI_MPI_Waitall(int count, MPI_Request requests[], MPI_Status status[]) {
+	return smpi_mpi_waitall(count, requests,status);
+}
+int SMPI_MPI_Waitany(int count, MPI_Request requests[], int *index, MPI_Status status[]) {
+	return smpi_mpi_waitany(count, requests, index,status);
+}
 /**
  * MPI_Bcast
  **/
@@ -220,10 +229,27 @@ int SMPI_MPI_Bcast(void *buf, int count, MPI_Datatype datatype, int root,
   return retval;
 }
 
+
+
+#ifdef DEBUG_REDUCE
 /**
- * MPI_Reduce
+ * debugging helper function
  **/
+static void print_buffer_int( void *buf, int len, const 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);
+}
+#endif
 
+/**
+ * MPI_Reduce
+ **/
 int SMPI_MPI_Reduce( void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype,
    		         MPI_Op op, int root, MPI_Comm comm )
 {
@@ -231,52 +257,73 @@ int SMPI_MPI_Reduce( void *sendbuf, void *recvbuf, int count, MPI_Datatype datat
   int rank;
   int size;
   int i;
+  int tag=0;
   smpi_mpi_request_t *tabrequest;
+  smpi_mpi_request_t request;
 
   smpi_bench_end();
 
   rank = smpi_mpi_comm_rank(comm);
   size = comm->size;
 
-  tabrequest = xbt_malloc((size)*sizeof(smpi_mpi_request_t));
+  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);
 
-  if (rank != root) { // if i am not root, simply send my buffer to root
-	    retval = smpi_create_request(sendbuf, count, datatype,
-				  rank, root, 0, comm, &(tabrequest[rank]));
-	    smpi_mpi_isend(tabrequest[rank]);
-  	    smpi_mpi_wait(tabrequest[rank], MPI_STATUS_IGNORE);
-	    //printf("DEBUG: rank %d sent my sendbuf to root (rank %d)\n",rank,root);
   } else {
-	    // i am the root: wait for all buffers by creating requests
+	    // i am the ROOT: wait for all buffers by creating one request by sender
+	    int src;
+	    tabrequest = 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);
+	    }
+	    memcpy(recvbuf,sendbuf,count*datatype->size*sizeof(char)); // initiliaze recv buf with my own snd buf
+
 	    // 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<comm->size; i++) {
-			if ( rank != i ) { // except for me
-				  // reminder: for smpi_create_request() the src is always the process sending.
-				  retval = smpi_create_request(recvbuf, count, datatype, MPI_ANY_SOURCE, root,
-							0, comm, &(tabrequest[i]));
-				  if (NULL != tabrequest[i] && MPI_SUCCESS == retval) {
-					    if (MPI_SUCCESS == retval) {
-							smpi_mpi_irecv(tabrequest[i]);
-					    }
+	    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, &(tabrequest[i]));
+			if (NULL != tabrequest[i] && MPI_SUCCESS == retval) {
+				  if (MPI_SUCCESS == retval) {
+					    smpi_mpi_irecv(tabrequest[i]);
 				  }
 			}
 	    }
 	    // now, wait for completion of all irecv's.
-	    // FIXME: we should implement smpi_wait_all for a more asynchronous behavior
-	    for (i=0; i<comm->size; i++) {
-			if ( rank != i ) { // except for me
-				  smpi_mpi_wait(tabrequest[i], MPI_STATUS_IGNORE);
-
-				  // FIXME: the core part is here. To be written ...
-
-				  fprintf(stderr,"[smpi] %s:%d : MPI_Reduce *Not yet implemented*.\n",__FILE__,__LINE__);
-			}
+	    for (i=0; i<size-1; i++) {
+			int index = MPI_UNDEFINED;
+			smpi_mpi_waitany(size-1, tabrequest, &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_mallocator_release(smpi_global->request_mallocator, tabrequest[i]);
+			xbt_free( tmpbufs[index]);
 	    }
-
+	    xbt_free(tabrequest);
+	    xbt_free(tmpbufs);
   }
-  for (i=0; i<comm->size; i++)
-	    xbt_mallocator_release(smpi_global->request_mallocator, tabrequest[i]);
 
   smpi_bench_begin();
 
@@ -287,8 +334,8 @@ int SMPI_MPI_Reduce( void *sendbuf, void *recvbuf, int count, MPI_Datatype datat
 int smpi_compare_rankkeys(const void *a, const void *b);
 int smpi_compare_rankkeys(const void *a, const void *b)
 {
-  int *x = (int *) a;
-  int *y = (int *) b;
+	  int *x = (int *) a;
+	  int *y = (int *) b;
 
   if (x[1] < y[1])
     return -1;