change some behaviors with MPI_IN_PLACE

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

diff --git a/src/smpi/smpi_pmpi.c b/src/smpi/smpi_pmpi.c
index 26e809b..7859738 100644 (file)
--- a/src/smpi/smpi_pmpi.c
+++ b/src/smpi/smpi_pmpi.c
@@ -303,7 +303,6 @@ int PMPI_Group_free(MPI_Group * group)
   if (group == NULL) {
     retval = MPI_ERR_ARG;
   } else {
-    if(*group!= smpi_comm_group(MPI_COMM_WORLD))// do not free the group of the comm_world
     smpi_group_destroy(*group);
     *group = MPI_GROUP_NULL;
     retval = MPI_SUCCESS;
@@ -419,7 +418,6 @@ int PMPI_Group_union(MPI_Group group1, MPI_Group group2,
         smpi_group_set_mapping(*newgroup, proc2, i);
       }
     }
-    smpi_group_use(*newgroup);
     retval = MPI_SUCCESS;
   }
   smpi_bench_begin();
@@ -459,7 +457,6 @@ int PMPI_Group_intersection(MPI_Group group1, MPI_Group group2,
         }
       }
     }
-    smpi_group_use(*newgroup);
     retval = MPI_SUCCESS;
   }
   smpi_bench_begin();
@@ -496,7 +493,6 @@ int PMPI_Group_difference(MPI_Group group1, MPI_Group group2, MPI_Group * newgro
         }
       }
     }
-    smpi_group_use(*newgroup);
     retval = MPI_SUCCESS;
   }
   smpi_bench_begin();
@@ -517,6 +513,11 @@ int PMPI_Group_incl(MPI_Group group, int n, int *ranks, MPI_Group * newgroup)
       *newgroup = MPI_GROUP_EMPTY;
     } else if (n == smpi_group_size(group)) {
       *newgroup = group;
+      if(group!= smpi_comm_group(MPI_COMM_WORLD)
+                && group != MPI_GROUP_NULL
+                && group != smpi_comm_group(MPI_COMM_SELF)
+                && group != MPI_GROUP_EMPTY)
+      smpi_group_use(group);
     } else {
       *newgroup = smpi_group_new(n);
       for (i = 0; i < n; i++) {
@@ -524,7 +525,6 @@ int PMPI_Group_incl(MPI_Group group, int n, int *ranks, MPI_Group * newgroup)
         smpi_group_set_mapping(*newgroup, index, i);
       }
     }
-    smpi_group_use(*newgroup);
     retval = MPI_SUCCESS;
   }
   smpi_bench_begin();
@@ -543,6 +543,11 @@ int PMPI_Group_excl(MPI_Group group, int n, int *ranks, MPI_Group * newgroup)
   } else {
     if (n == 0) {
       *newgroup = group;
+      if(group!= smpi_comm_group(MPI_COMM_WORLD)
+                && group != MPI_GROUP_NULL
+                && group != smpi_comm_group(MPI_COMM_SELF)
+                && group != MPI_GROUP_EMPTY)
+      smpi_group_use(group);
     } else if (n == smpi_group_size(group)) {
       *newgroup = MPI_GROUP_EMPTY;
     } else {
@@ -567,7 +572,6 @@ int PMPI_Group_excl(MPI_Group group, int n, int *ranks, MPI_Group * newgroup)
 
       xbt_free(to_exclude);
     }
-    smpi_group_use(*newgroup);
     retval = MPI_SUCCESS;
   }
   smpi_bench_begin();
@@ -626,7 +630,6 @@ int PMPI_Group_range_incl(MPI_Group group, int n, int ranges[][3],
         }
       }
     }
-    smpi_group_use(*newgroup);
     retval = MPI_SUCCESS;
   }
   smpi_bench_begin();
@@ -636,7 +639,7 @@ int PMPI_Group_range_incl(MPI_Group group, int n, int ranges[][3],
 int PMPI_Group_range_excl(MPI_Group group, int n, int ranges[][3],
                          MPI_Group * newgroup)
 {
-  int retval, i, newrank, rank, size, index, add;
+  int retval, i, rank, newrank,oldrank, size, index, add;
 
   smpi_bench_end();
   if (group == MPI_GROUP_NULL) {
@@ -646,41 +649,65 @@ int PMPI_Group_range_excl(MPI_Group group, int n, int ranges[][3],
   } else {
     if (n == 0) {
       *newgroup = group;
+      if(group!= smpi_comm_group(MPI_COMM_WORLD)
+                && group != MPI_GROUP_NULL
+                && group != smpi_comm_group(MPI_COMM_SELF)
+                && group != MPI_GROUP_EMPTY)
+      smpi_group_use(group);
     } else {
       size = smpi_group_size(group);
       for (i = 0; i < n; i++) {
         for (rank = ranges[i][0];       /* First */
-             rank >= 0 && rank <= ranges[i][1]; /* Last */
-             rank += ranges[i][2] /* Stride */ ) {
+             rank >= 0; /* Last */
+              ) {
           size--;
+
+          rank += ranges[i][2]; /* Stride */
+         if (ranges[i][0]<ranges[i][1]){
+             if(rank > ranges[i][1])
+               break;
+         }else{
+             if(rank < ranges[i][1])
+               break;
+         }
         }
       }
       if (size == 0) {
         *newgroup = MPI_GROUP_EMPTY;
       } else {
         *newgroup = smpi_group_new(size);
-        newrank = 0;
+        newrank=0;
+        oldrank=0;
         while (newrank < size) {
+          add=1;
           for (i = 0; i < n; i++) {
-            add = 1;
-            for (rank = ranges[i][0];   /* First */
-                 rank >= 0 && rank <= ranges[i][1];     /* Last */
-                 rank += ranges[i][2] /* Stride */ ) {
-              if (rank == newrank) {
-                add = 0;
-                break;
+            for (rank = ranges[i][0];rank >= 0;){
+              if(rank==oldrank){
+                  add=0;
+                  break;
+              }
+
+              rank += ranges[i][2]; /* Stride */
+
+              if (ranges[i][0]<ranges[i][1]){
+                  if(rank > ranges[i][1])
+                    break;
+              }else{
+                  if(rank < ranges[i][1])
+                    break;
               }
-            }
-            if (add == 1) {
-              index = smpi_group_index(group, newrank);
-              smpi_group_set_mapping(*newgroup, index, newrank);
             }
           }
-          newrank++; //added to avoid looping, need to be checked ..
+          if(add==1){
+            index = smpi_group_index(group, oldrank);
+            smpi_group_set_mapping(*newgroup, index, newrank);
+            newrank++;
+          }
+          oldrank++;
         }
       }
     }
-    smpi_group_use(*newgroup);
+
     retval = MPI_SUCCESS;
   }
   smpi_bench_begin();
@@ -749,6 +776,11 @@ int PMPI_Comm_group(MPI_Comm comm, MPI_Group * group)
     retval = MPI_ERR_ARG;
   } else {
     *group = smpi_comm_group(comm);
+    if(*group!= smpi_comm_group(MPI_COMM_WORLD)
+              && *group != MPI_GROUP_NULL
+              && *group != smpi_comm_group(MPI_COMM_SELF)
+              && *group != MPI_GROUP_EMPTY)
+    smpi_group_use(*group);
     retval = MPI_SUCCESS;
   }
   smpi_bench_begin();
@@ -1414,6 +1446,7 @@ int PMPI_Iprobe(int source, int tag, MPI_Comm comm, int* flag, MPI_Status* statu
   } else if (comm == MPI_COMM_NULL) {
     retval = MPI_ERR_COMM;
   } else if (source == MPI_PROC_NULL) {
+    *flag=TRUE;
     smpi_empty_status(status);
     status->MPI_SOURCE = MPI_PROC_NULL;
     retval = MPI_SUCCESS;
@@ -1832,9 +1865,9 @@ int PMPI_Scatter(void *sendbuf, int sendcount, MPI_Datatype sendtype,
              || ((recvbuf !=MPI_IN_PLACE) && (recvtype == MPI_DATATYPE_NULL))) {
     retval = MPI_ERR_TYPE;
   } else {
-
-    if(recvbuf==MPI_IN_PLACE){
-       recvcount=0;
+    if (recvbuf == MPI_IN_PLACE) {
+        recvtype=sendtype;
+        recvcount=sendcount;
     }
     mpi_coll_scatter_fun(sendbuf, sendcount, sendtype, recvbuf, recvcount,
                      recvtype, root, comm);
@@ -1869,11 +1902,10 @@ int PMPI_Scatterv(void *sendbuf, int *sendcounts, int *displs,
              || ((recvbuf !=MPI_IN_PLACE) && (recvtype == MPI_DATATYPE_NULL))) {
     retval = MPI_ERR_TYPE;
   } else {
-
-    if(recvbuf==MPI_IN_PLACE){
-       recvcount=0;
+    if (recvbuf == MPI_IN_PLACE) {
+        recvtype=sendtype;
+        recvcount=sendcounts[smpi_comm_rank(comm)];
     }
-
     smpi_mpi_scatterv(sendbuf, sendcounts, displs, sendtype, recvbuf,
                       recvcount, recvtype, root, comm);
     retval = MPI_SUCCESS;
@@ -1904,17 +1936,7 @@ int PMPI_Reduce(void *sendbuf, void *recvbuf, int count,
     retval = MPI_ERR_ARG;
   } else {
 
-    char* sendtmpbuf = (char*) sendbuf;
-    if( sendbuf == MPI_IN_PLACE ) {
-      sendtmpbuf = (char *)xbt_malloc(count*smpi_datatype_get_extent(datatype));
-      smpi_datatype_copy(recvbuf, count, datatype,sendtmpbuf, count, datatype);
-    }
-
-    mpi_coll_reduce_fun(sendtmpbuf, recvbuf, count, datatype, op, root, comm);
-
-    if( sendbuf == MPI_IN_PLACE ) {
-      xbt_free(sendtmpbuf);
-    }
+    mpi_coll_reduce_fun(sendbuf, recvbuf, count, datatype, op, root, comm);
 
     retval = MPI_SUCCESS;
   }