fix behavior of waitany

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

diff --git a/src/smpi/smpi_pmpi.c b/src/smpi/smpi_pmpi.c
index df6074b..59083e6 100644 (file)
--- a/src/smpi/smpi_pmpi.c
+++ b/src/smpi/smpi_pmpi.c
@@ -27,6 +27,7 @@ void TRACE_smpi_set_category(const char *category)
 int PMPI_Init(int *argc, char ***argv)
 {
   smpi_process_init(argc, argv);
+  smpi_process_mark_as_initialized();
 #ifdef HAVE_TRACING
   int rank = smpi_process_index();
   TRACE_smpi_init(rank);
@@ -303,7 +304,6 @@ 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 +419,6 @@ 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 +458,6 @@ int PMPI_Group_intersection(MPI_Group group1, MPI_Group group2,
         }
       }
     }
-    smpi_group_use(*newgroup);
     retval = MPI_SUCCESS;
   }
   smpi_bench_begin();
@@ -496,7 +494,6 @@ 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 +514,11 @@ 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 +526,6 @@ 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 +544,11 @@ 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 +573,6 @@ 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();
@@ -606,28 +611,26 @@ int PMPI_Group_range_incl(MPI_Group group, int n, int ranges[][3],
         }
       }
 
-        *newgroup = smpi_group_new(size);
-        j = 0;
-        for (i = 0; i < n; i++) {
-          for (rank = ranges[i][0];     /* First */
-               rank >= 0; /* Last */
-               ) {
-            index = smpi_group_index(group, rank);
-            smpi_group_set_mapping(*newgroup, index, j);
-            j++;
-            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;
-           }
-          }
+      *newgroup = smpi_group_new(size);
+      j = 0;
+      for (i = 0; i < n; i++) {
+        for (rank = ranges[i][0];     /* First */
+             rank >= 0; /* Last */
+             ) {
+          index = smpi_group_index(group, rank);
+          smpi_group_set_mapping(*newgroup, index, j);
+          j++;
+          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;
+         }
         }
-      //}
+      }
     }
-    smpi_group_use(*newgroup);
     retval = MPI_SUCCESS;
   }
   smpi_bench_begin();
@@ -637,7 +640,7 @@ 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) {
@@ -647,41 +650,65 @@ 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();
@@ -750,6 +777,11 @@ 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();
@@ -1415,6 +1447,7 @@ 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;
@@ -1446,6 +1479,7 @@ int PMPI_Wait(MPI_Request * request, MPI_Status * status)
 
   int src_traced = (*request)->src;
   int dst_traced = (*request)->dst;
+  MPI_Comm comm = (*request)->comm;
   int is_wait_for_receive = (*request)->recv;
   TRACE_smpi_ptp_in(rank, src_traced, dst_traced, __FUNCTION__);
 #endif
@@ -1454,8 +1488,13 @@ int PMPI_Wait(MPI_Request * request, MPI_Status * status)
     retval = MPI_SUCCESS;
 
 #ifdef HAVE_TRACING
+  //the src may not have been known at the beginning of the recv (MPI_ANY_SOURCE)
   TRACE_smpi_ptp_out(rank, src_traced, dst_traced, __FUNCTION__);
   if (is_wait_for_receive) {
+    if(src_traced==MPI_ANY_SOURCE)
+    src_traced = (status!=MPI_STATUS_IGNORE) ?
+                                smpi_group_rank(smpi_comm_group(comm), status->MPI_SOURCE) :
+                                src_traced;
     TRACE_smpi_recv(rank, src_traced, dst_traced);
   }
   TRACE_smpi_computing_in(rank);
@@ -1478,12 +1517,15 @@ int PMPI_Waitany(int count, MPI_Request requests[], int *index, MPI_Status * sta
   int *srcs = xbt_new(int, count);
   int *dsts = xbt_new(int, count);
   int *recvs = xbt_new(int, count);
+  MPI_Comm *comms = xbt_new(MPI_Comm, count);
+
   for (i = 0; i < count; i++) {
     MPI_Request req = requests[i];      //already received requests are no longer valid
     if (req) {
       srcs[i] = req->src;
       dsts[i] = req->dst;
       recvs[i] = req->recv;
+      comms[i] = req->comm;
     }
   }
   int rank_traced = smpi_process_index();
@@ -1501,15 +1543,22 @@ int PMPI_Waitany(int count, MPI_Request requests[], int *index, MPI_Status * sta
 #ifdef HAVE_TRACING
   if(*index!=MPI_UNDEFINED){
     int src_traced = srcs[*index];
+    //the src may not have been known at the beginning of the recv (MPI_ANY_SOURCE)
     int dst_traced = dsts[*index];
     int is_wait_for_receive = recvs[*index];
     if (is_wait_for_receive) {
+      if(srcs[*index]==MPI_ANY_SOURCE)
+        src_traced = (status!=MPI_STATUSES_IGNORE) ?
+                      smpi_group_rank(smpi_comm_group(comms[*index]), status->MPI_SOURCE) :
+                      srcs[*index];
       TRACE_smpi_recv(rank_traced, src_traced, dst_traced);
     }
     TRACE_smpi_ptp_out(rank_traced, src_traced, dst_traced, __FUNCTION__);
     xbt_free(srcs);
     xbt_free(dsts);
     xbt_free(recvs);
+    xbt_free(comms);
+
   }
   TRACE_smpi_computing_in(rank_traced);
 #endif
@@ -1527,14 +1576,20 @@ int PMPI_Waitall(int count, MPI_Request requests[], MPI_Status status[])
   int *srcs = xbt_new(int, count);
   int *dsts = xbt_new(int, count);
   int *recvs = xbt_new(int, count);
-  int valid_count = 0;
+  int *valid = xbt_new(int, count);
+  MPI_Comm *comms = xbt_new(MPI_Comm, count);
+
+  //int valid_count = 0;
   for (i = 0; i < count; i++) {
     MPI_Request req = requests[i];
     if(req!=MPI_REQUEST_NULL){
-      srcs[valid_count] = req->src;
-      dsts[valid_count] = req->dst;
-      recvs[valid_count] = req->recv;
-      valid_count++;
+      srcs[i] = req->src;
+      dsts[i] = req->dst;
+      recvs[i] = req->recv;
+      comms[i] = req->comm;
+      valid[i]=1;;
+    }else{
+      valid[i]=0;
     }
   }
   int rank_traced = smpi_process_index();
@@ -1544,18 +1599,29 @@ int PMPI_Waitall(int count, MPI_Request requests[], MPI_Status status[])
 #endif
   int retval = smpi_mpi_waitall(count, requests, status);
 #ifdef HAVE_TRACING
-  for (i = 0; i < valid_count; i++) {
-    int src_traced = srcs[i];
-    int dst_traced = dsts[i];
-    int is_wait_for_receive = recvs[i];
-    if (is_wait_for_receive) {
-      TRACE_smpi_recv(rank_traced, src_traced, dst_traced);
+  for (i = 0; i < count; i++) {
+    if(valid[i]){
+    //int src_traced = srcs[*index];
+    //the src may not have been known at the beginning of the recv (MPI_ANY_SOURCE)
+      int src_traced = srcs[i];
+      int dst_traced = dsts[i];
+      int is_wait_for_receive = recvs[i];
+      if (is_wait_for_receive) {
+        if(src_traced==MPI_ANY_SOURCE)
+        src_traced = (status!=MPI_STATUSES_IGNORE) ?
+                          smpi_group_rank(smpi_comm_group(comms[i]), status[i].MPI_SOURCE) :
+                          srcs[i];
+        TRACE_smpi_recv(rank_traced, src_traced, dst_traced);
+      }
     }
   }
   TRACE_smpi_ptp_out(rank_traced, -1, -1, __FUNCTION__);
   xbt_free(srcs);
   xbt_free(dsts);
   xbt_free(recvs);
+  xbt_free(valid);
+  xbt_free(comms);
+
   TRACE_smpi_computing_in(rank_traced);
 #endif
   smpi_bench_begin();
@@ -1833,9 +1899,9 @@ 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);
@@ -1870,11 +1936,10 @@ 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;
@@ -1905,17 +1970,7 @@ 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;
   }
@@ -2295,6 +2350,21 @@ int PMPI_Type_indexed(int count, int* blocklens, int* indices, MPI_Datatype old_
   return retval;
 }
 
+int PMPI_Type_create_indexed(int count, int* blocklens, int* indices, MPI_Datatype old_type, MPI_Datatype* new_type) {
+  int retval;
+
+  smpi_bench_end();
+  if (old_type == MPI_DATATYPE_NULL) {
+    retval = MPI_ERR_TYPE;
+  } else if (count<0){
+    retval = MPI_ERR_COUNT;
+  } else {
+    retval = smpi_datatype_indexed(count, blocklens, indices, old_type, new_type);
+  }
+  smpi_bench_begin();
+  return retval;
+}
+
 int PMPI_Type_create_indexed_block(int count, int blocklength, int* indices, MPI_Datatype old_type, MPI_Datatype* new_type) {
   int retval,i;
 
@@ -2329,6 +2399,10 @@ int PMPI_Type_hindexed(int count, int* blocklens, MPI_Aint* indices, MPI_Datatyp
   return retval;
 }
 
+int PMPI_Type_create_hindexed(int count, int* blocklens, MPI_Aint* indices, MPI_Datatype old_type, MPI_Datatype* new_type) {
+  return PMPI_Type_hindexed(count, blocklens,indices,old_type,new_type);
+}
+
 int PMPI_Type_create_hindexed_block(int count, int blocklength, MPI_Aint* indices, MPI_Datatype old_type, MPI_Datatype* new_type) {
   int retval,i;
 
@@ -2374,7 +2448,7 @@ int PMPI_Error_class(int errorcode, int* errorclass) {
 
 
 int PMPI_Initialized(int* flag) {
-   *flag=(smpi_process_data()!=NULL);
+   *flag=smpi_process_initialized();
    return MPI_SUCCESS;
 }