change some behaviors with MPI_IN_PLACE

[simgrid.git] / src / smpi / colls / smpi_mpich_selector.c

diff --git a/src/smpi/colls/smpi_mpich_selector.c b/src/smpi/colls/smpi_mpich_selector.c
index c956135..a2a44ae 100644 (file)
--- a/src/smpi/colls/smpi_mpich_selector.c
+++ b/src/smpi/colls/smpi_mpich_selector.c
@@ -176,7 +176,7 @@ int smpi_coll_tuned_alltoall_mpich( void *sbuf, int scount,
                                                            comm);
     }
 
-    return smpi_coll_tuned_alltoall_pair (sbuf, scount, sdtype, 
+    return smpi_coll_tuned_alltoall_ompi_pairwise (sbuf, scount, sdtype,
                                                     rbuf, rcount, rdtype,
                                                     comm);
 }
@@ -424,20 +424,17 @@ int smpi_coll_tuned_reduce_scatter_mpich( void *sbuf, void *rbuf,
                                                     )
 {
     int comm_size, i;
-    size_t total_message_size, dsize;
-    int zerocounts = 0;
+    size_t total_message_size;
+
+    if(sbuf==rbuf)sbuf=MPI_IN_PLACE; //restore MPI_IN_PLACE as these algos handle it
 
     XBT_DEBUG("smpi_coll_tuned_reduce_scatter_mpich");
     
     comm_size = smpi_comm_size(comm);
     // We need data size for decision function 
-    dsize=smpi_datatype_size(dtype);
     total_message_size = 0;
     for (i = 0; i < comm_size; i++) { 
         total_message_size += rcounts[i];
-        if (0 == rcounts[i]) {
-            zerocounts = 1;
-        }
     }
 
     if( smpi_op_is_commute(op) &&  total_message_size > 524288) { 
@@ -600,7 +597,7 @@ int smpi_coll_tuned_allgatherv_mpich(void *sbuf, int scount,
                                                MPI_Comm  comm
                                                )
 {
-    int communicator_size, pow2_size;
+    int communicator_size, pow2_size,i;
     size_t dsize, total_dsize;
 
     communicator_size = smpi_comm_size(comm);
@@ -608,7 +605,12 @@ int smpi_coll_tuned_allgatherv_mpich(void *sbuf, int scount,
     /* Determine complete data size */
     dsize=smpi_datatype_size(sdtype);
     total_dsize = dsize * scount * communicator_size;   
-   
+
+    total_dsize = 0;
+    for (i=0; i<communicator_size; i++)
+        total_dsize += rcounts[i];
+    if (total_dsize == 0) return MPI_SUCCESS;
+    
     for (pow2_size  = 1; pow2_size < communicator_size; pow2_size <<=1); 
 
     if ((pow2_size == communicator_size) && (total_dsize < 524288)) {
@@ -689,6 +691,11 @@ int smpi_coll_tuned_scatter_mpich(void *sbuf, int scount,
                                             int root, MPI_Comm  comm
                                             )
 {
+  if(smpi_comm_rank(comm)!=root){
+      sbuf=xbt_malloc(rcount*smpi_datatype_get_extent(rdtype));
+      scount=rcount;
+      sdtype=rdtype;
+  }
         return smpi_coll_tuned_scatter_ompi_binomial (sbuf, scount, sdtype, 
                                                        rbuf, rcount, rdtype, 
                                                        root, comm);