return retval;
}
+
+
+#ifdef DEBUG_REDUCE
/**
* debugging helper function
**/
-void print_buffer_int( void *buf, int len, const char *msg) ;
-void print_buffer_int( void *buf, int len, const char *msg) {
+static void print_buffer_int( void *buf, int len, const char *msg, int rank) {
int tmp, *v;
- printf("**%s: ",msg);
+ 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 )
{
smpi_mpi_request_t *tabrequest;
smpi_mpi_request_t request;
-
smpi_bench_end();
rank = smpi_mpi_comm_rank(comm);
size = comm->size;
- //printf("-->rank %d. Entering ....\n",rank);
- //print_buffer_int( sendbuf, count, "sendbuf");
-
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);
} else {
// 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);
+ tmpbufs[i] = xbt_malloc(count*datatype->size);
}
- memcpy(recvbuf,sendbuf,count*datatype->size*sizeof(char)); // initiliaze recv buf with my own snd buf
+ 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-1; 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,
- MPI_ANY_SOURCE, root,
+ src , root,
tag, comm, &(tabrequest[i]));
if (NULL != tabrequest[i] && MPI_SUCCESS == retval) {
if (MPI_SUCCESS == retval) {
}
}
// now, wait for completion of all irecv's.
- for (i=0; i<comm->size-1; i++) {
+ for (i=0; i<size-1; i++) {
int index = MPI_UNDEFINED;
- smpi_mpi_waitany(comm->size-1, tabrequest, &index, MPI_STATUS_IGNORE);
+ smpi_mpi_waitany(size-1, tabrequest, &index, MPI_STATUS_IGNORE);
- //print_buffer_int( recvbuf, count, "rcvbuf");
- //printf("MPI_Waitany() unblocked: root received (completes req[index=%d]): ",index);
- //print_buffer_int( tmpbufs[index], count, "tmpbufs[index]");
+#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
+ // arg 2 is modified
op->func (tmpbufs[index],recvbuf,&count,&datatype);
+#ifdef DEBUG_REDUCE
+ print_buffer_int( recvbuf, count, xbt_strdup("rcvbuf"),rank);
- // print_buffer_int( recvbuf, count, "recvbuf after func");
+#endif
//xbt_mallocator_release(smpi_global->request_mallocator, tabrequest[i]);
- //xbt_free( tmpbufs[index]);
+ xbt_free( tmpbufs[index]);
}
xbt_free(tabrequest);
xbt_free(tmpbufs);