alltoall implemented (almost opmpi algorithms)

[simgrid.git] / src / smpi / smpi_mpi.c

diff --git a/src/smpi/smpi_mpi.c b/src/smpi/smpi_mpi.c
index fb2e86b..fdd8a4f 100644 (file)
--- a/src/smpi/smpi_mpi.c
+++ b/src/smpi/smpi_mpi.c
@@ -2,6 +2,7 @@
 
 #include "private.h"
 #include "smpi_coll_private.h"
+#include "smpi_mpi_dt_private.h"
 
 XBT_LOG_NEW_DEFAULT_SUBCATEGORY(smpi_mpi, smpi,
                                 "Logging specific to SMPI (mpi)");
@@ -67,41 +68,10 @@ int SMPI_MPI_Comm_rank(MPI_Comm comm, int *rank)
 }
 
 
-//------------------------------- Datatypes ---------------------------------------
-/**
- * query the size of the type
- **/
-int SMPI_MPI_Type_size(MPI_Datatype datatype, size_t * size)
-{
-  int retval = MPI_SUCCESS;
-
-  smpi_bench_end();
-
-  if (NULL == datatype) {
-    retval = MPI_ERR_TYPE;
-  } else if (NULL == size) {
-    retval = MPI_ERR_ARG;
-  } else {
-    *size = datatype->size;
-  }
-
-  smpi_bench_begin();
-
-  return retval;
-}
-
 
 /**
- * query extent and lower bound of the type 
+ * Barrier
  **/
-int SMPI_MPI_Type_get_extent( MPI_Datatype datatype, int *lb, int *extent) 
-{
-        return( smpi_mpi_type_get_extent( datatype, lb, extent));
-}
-
-
-
-
 int SMPI_MPI_Barrier(MPI_Comm comm)
 {
   int retval = MPI_SUCCESS;
@@ -126,6 +96,8 @@ int SMPI_MPI_Barrier(MPI_Comm comm)
   return retval;
 }
 
+
+
 int SMPI_MPI_Irecv(void *buf, int count, MPI_Datatype datatype, int src,
                    int tag, MPI_Comm comm, MPI_Request * request)
 {
@@ -185,6 +157,9 @@ int SMPI_MPI_Isend(void *buf, int count, MPI_Datatype datatype, int dst,
   return retval;
 }
 
+/**
+ * MPI_Send user level
+ **/
 int SMPI_MPI_Send(void *buf, int count, MPI_Datatype datatype, int dst,
                   int tag, MPI_Comm comm)
 {
@@ -321,6 +296,9 @@ int smpi_mpi_bcast(void *buf, int count, MPI_Datatype datatype, int root,
                    MPI_Comm comm)
 {
   int retval = MPI_SUCCESS;
+  int rank = smpi_mpi_comm_rank(comm);
+
+  DEBUG1("<%d> entered smpi_mpi_bcast(). Calls nary_tree_bcast()",rank);
   //retval = flat_tree_bcast(buf, count, datatype, root, comm);
   retval = nary_tree_bcast(buf, count, datatype, root, comm, 2 );
   return retval;
@@ -383,7 +361,7 @@ int smpi_mpi_reduce(void *sendbuf, void *recvbuf, int count,
         int rank;
         int size;
         int i;
-        int tag = 0;
+        int system_tag = 666;
         smpi_mpi_request_t *requests;
         smpi_mpi_request_t request;
 
@@ -391,13 +369,14 @@ int smpi_mpi_reduce(void *sendbuf, void *recvbuf, int count,
 
         rank = smpi_mpi_comm_rank(comm);
         size = comm->size;
+        DEBUG1("<%d> entered smpi_mpi_reduce()",rank);
 
         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,
+                retval = smpi_create_request(sendbuf, count, datatype, rank, root, system_tag, comm,
                                         &request);
                 smpi_mpi_isend(request);
                 smpi_mpi_wait(request, MPI_STATUS_IGNORE);
@@ -422,7 +401,7 @@ int smpi_mpi_reduce(void *sendbuf, void *recvbuf, int count,
                         // 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]));
+                                        src, root, system_tag, comm, &(requests[i]));
                         if (NULL != requests[i] && MPI_SUCCESS == retval) {
                                 if (MPI_SUCCESS == retval) {
                                         smpi_mpi_irecv(requests[i]);
@@ -433,8 +412,9 @@ int smpi_mpi_reduce(void *sendbuf, void *recvbuf, int count,
                 for (i = 0; i < size-1; i++) {
                         int index = MPI_UNDEFINED;
                         smpi_mpi_waitany( size-1, requests, &index, MPI_STATUS_IGNORE);
+                        DEBUG3("<%d> waitany() unblocked by reception (completes request[%d]) (%d reqs remaining)",
+                                        rank,index,size-i-2);
 #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
@@ -586,6 +566,7 @@ int SMPI_MPI_Alltoall(void *sendbuf, int sendcount, MPI_Datatype datatype,
 
   rank = smpi_mpi_comm_rank(comm);
   block_dsize = datatype->size * sendcount;
+  DEBUG2("<%d> optimized alltoall() called. Block size sent to each rank: %d bytes.",rank,block_dsize);
 
   if ((block_dsize < 200) && (comm->size > 12)) {
            retval = smpi_coll_tuned_alltoall_bruck(sendbuf, sendcount, datatype,