change some behaviors with MPI_IN_PLACE

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

diff --git a/src/smpi/smpi_base.c b/src/smpi/smpi_base.c
index 8afbcae..88e9307 100644 (file)
--- a/src/smpi/smpi_base.c
+++ b/src/smpi/smpi_base.c
@@ -421,7 +421,6 @@ void smpi_mpi_startall(int count, MPI_Request * requests)
 
 void smpi_mpi_request_free(MPI_Request * request)
 {
-
   if((*request) != MPI_REQUEST_NULL){
     (*request)->refcount--;
     if((*request)->refcount<0) xbt_die("wrong refcount");
@@ -1110,8 +1109,10 @@ void smpi_mpi_scatter(void *sendbuf, int sendcount, MPI_Datatype sendtype,
     // FIXME: check for errors
     smpi_datatype_extent(sendtype, &lb, &sendext);
     // Local copy from root
-    smpi_datatype_copy((char *)sendbuf + root * sendcount * sendext,
-                       sendcount, sendtype, recvbuf, recvcount, recvtype);
+    if(recvbuf!=MPI_IN_PLACE){
+        smpi_datatype_copy((char *)sendbuf + root * sendcount * sendext,
+                           sendcount, sendtype, recvbuf, recvcount, recvtype);
+    }
     // Send buffers to receivers
     requests = xbt_new(MPI_Request, size - 1);
     index = 0;
@@ -1149,8 +1150,10 @@ void smpi_mpi_scatterv(void *sendbuf, int *sendcounts, int *displs,
     // FIXME: check for errors
     smpi_datatype_extent(sendtype, &lb, &sendext);
     // Local copy from root
-    smpi_datatype_copy((char *)sendbuf + displs[root] * sendext, sendcounts[root],
+    if(recvbuf!=MPI_IN_PLACE){
+      smpi_datatype_copy((char *)sendbuf + displs[root] * sendext, sendcounts[root],
                        sendtype, recvbuf, recvcount, recvtype);
+    }
     // Send buffers to receivers
     requests = xbt_new(MPI_Request, size - 1);
     index = 0;
@@ -1179,9 +1182,11 @@ void smpi_mpi_reduce(void *sendbuf, void *recvbuf, int count,
   MPI_Request *requests;
   void **tmpbufs;
 
+
   char* sendtmpbuf = (char*) sendbuf;
   if( sendbuf == MPI_IN_PLACE ) {
-      sendtmpbuf = (char *)recvbuf;
+    sendtmpbuf = (char *)xbt_malloc(count*smpi_datatype_get_extent(datatype));
+    smpi_datatype_copy(recvbuf, count, datatype,sendtmpbuf, count, datatype);
   }
 
   rank = smpi_comm_rank(comm);
@@ -1234,6 +1239,10 @@ void smpi_mpi_reduce(void *sendbuf, void *recvbuf, int count,
     }
     xbt_free(tmpbufs);
     xbt_free(requests);
+
+    if( sendbuf == MPI_IN_PLACE ) {
+      xbt_free(sendtmpbuf);
+    }
   }
 }
 
@@ -1247,7 +1256,7 @@ void smpi_mpi_allreduce(void *sendbuf, void *recvbuf, int count,
 void smpi_mpi_scan(void *sendbuf, void *recvbuf, int count,
                    MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
 {
-  int system_tag = 888;
+  int system_tag = -888;
   int rank, size, other, index;
   MPI_Aint lb = 0, dataext = 0;
   MPI_Request *requests;
@@ -1282,14 +1291,25 @@ void smpi_mpi_scan(void *sendbuf, void *recvbuf, int count,
   }
   // Wait for completion of all comms.
   smpi_mpi_startall(size - 1, requests);
-  for(other = 0; other < size - 1; other++) {
-    index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
-    if(index == MPI_UNDEFINED) {
-      break;
+
+  if(smpi_op_is_commute(op)){
+    for(other = 0; other < size - 1; other++) {
+      index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
+      if(index == MPI_UNDEFINED) {
+        break;
+      }
+      if(index < rank) {
+        // #Request is below rank: it's a irecv
+        smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
+      }
     }
-    if(index < rank) {
-      // #Request is below rank: it's a irecv
-      smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
+  }else{
+    //non commutative case, wait in order
+    for(other = 0; other < size - 1; other++) {
+      smpi_mpi_wait(&(requests[other]), MPI_STATUS_IGNORE);
+      if(index < rank) {
+        smpi_op_apply(op, tmpbufs[other], recvbuf, &count, &datatype);
+      }
     }
   }
   for(index = 0; index < rank; index++) {
@@ -1302,7 +1322,7 @@ void smpi_mpi_scan(void *sendbuf, void *recvbuf, int count,
 void smpi_mpi_exscan(void *sendbuf, void *recvbuf, int count,
                    MPI_Datatype datatype, MPI_Op op, MPI_Comm comm)
 {
-  int system_tag = 888;
+  int system_tag = -888;
   int rank, size, other, index;
   MPI_Aint lb = 0, dataext = 0;
   MPI_Request *requests;
@@ -1334,15 +1354,31 @@ void smpi_mpi_exscan(void *sendbuf, void *recvbuf, int count,
   }
   // Wait for completion of all comms.
   smpi_mpi_startall(size - 1, requests);
-  for(other = 0; other < size - 1; other++) {
-    index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
-    if(index == MPI_UNDEFINED) {
-      break;
+  if(smpi_op_is_commute(op)){
+    for(other = 0; other < size - 1; other++) {
+      index = smpi_mpi_waitany(size - 1, requests, MPI_STATUS_IGNORE);
+      if(index == MPI_UNDEFINED) {
+        break;
+      }
+      if(index < rank) {
+        if(recvbuf_is_empty){
+          smpi_datatype_copy(tmpbufs[index], count, datatype, recvbuf, count, datatype);
+          recvbuf_is_empty=0;
+        }else
+        // #Request is below rank: it's a irecv
+        smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
+      }
     }
-    if(index < rank) {
-      if(recvbuf_is_empty)  smpi_datatype_copy(tmpbufs[index], count, datatype, recvbuf, count, datatype);
-      // #Request is below rank: it's a irecv
-      else smpi_op_apply(op, tmpbufs[index], recvbuf, &count, &datatype);
+  }else{
+    //non commutative case, wait in order
+    for(other = 0; other < size - 1; other++) {
+      smpi_mpi_wait(&(requests[other]), MPI_STATUS_IGNORE);
+      if(index < rank) {
+          if(recvbuf_is_empty){
+            smpi_datatype_copy(tmpbufs[other], count, datatype, recvbuf, count, datatype);
+            recvbuf_is_empty=0;
+          }else smpi_op_apply(op, tmpbufs[other], recvbuf, &count, &datatype);
+      }
     }
   }
   for(index = 0; index < rank; index++) {