X-Git-Url: http://info.iut-bm.univ-fcomte.fr/pub/gitweb/simgrid.git/blobdiff_plain/4fa2931658e07433d8bd09d002697114b1c67d9b..3cfd4823046b64b2c8645a655dfb47b795e4f412:/src/smpi/smpi_mpi.c

diff --git a/src/smpi/smpi_mpi.c b/src/smpi/smpi_mpi.c
index 6b767a95c3..53799a0e94 100644
--- a/src/smpi/smpi_mpi.c
+++ b/src/smpi/smpi_mpi.c
@@ -1,7 +1,19 @@
+/* $Id: $tag */
+
+/* smpi_mpi.c -- 
+ *
+ * Eventually will contain the user level MPI primitives and its corresponding 
+ * internal wrapper. The implementations of these primitives should go to specific
+ * files. For example, a SMPI_MPI_Bcast() in this file, should call the wrapper 
+ * smpi_mpi_bcast(), which decides which implementation to call. Currently, it
+ * calls nary_tree_bcast() in smpi_coll.c.  (Stéphane Genaud).
+ * */
+
 
 
 #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)");
@@ -66,25 +78,11 @@ int SMPI_MPI_Comm_rank(MPI_Comm comm, int *rank)
   return retval;
 }
 
-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;
-}
 
+/**
+ * Barrier
+ **/
 int SMPI_MPI_Barrier(MPI_Comm comm)
 {
   int retval = MPI_SUCCESS;
@@ -109,6 +107,8 @@ 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)
 {
@@ -168,6 +168,9 @@ 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)
 {
@@ -191,6 +194,60 @@ int SMPI_MPI_Send(void *buf, int count, MPI_Datatype datatype, int dst,
   return retval;
 }
 
+
+/**
+ * MPI_Sendrecv internal level 
+ **/
+int smpi_mpi_sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype, int dest, int sendtag, 
+		    void *recvbuf, int recvcount, MPI_Datatype recvtype, int source, int recvtag,
+		    MPI_Comm comm, MPI_Status *status)
+{
+int rank;
+int retval = MPI_SUCCESS;
+smpi_mpi_request_t srequest;
+smpi_mpi_request_t rrequest;
+
+	  rank = smpi_mpi_comm_rank(comm);
+
+	  /* send */
+	  retval = smpi_create_request(sendbuf, sendcount, sendtype, 
+				rank,dest,sendtag, 
+				comm, &srequest);
+	  smpi_mpi_isend(srequest);
+        
+	  /* recv */
+	  retval = smpi_create_request(recvbuf, recvcount, recvtype, 
+				source, rank,recvtag, 
+				comm, &rrequest);
+	  smpi_mpi_irecv(rrequest);
+
+	  smpi_mpi_wait(srequest, MPI_STATUS_IGNORE);
+	  //printf("[%d] isend request src=%d dst=%d tag=%d COMPLETED (retval=%d) \n",rank,rank,dest,sendtag,retval);
+	  smpi_mpi_wait(rrequest, MPI_STATUS_IGNORE);
+	  //printf("[%d] irecv request src=%d -> dst=%d tag=%d COMPLETED (retval=%d)\n",rank,source,rank,recvtag,retval);
+
+	  return(retval);
+}
+/**
+ * MPI_Sendrecv user entry point
+ **/
+int SMPI_MPI_Sendrecv(void *sendbuf, int sendcount, MPI_Datatype sendtype, int dest, int sendtag, 
+		    void *recvbuf, int recvcount, MPI_Datatype recvtype, int source, int recvtag,
+		    MPI_Comm comm, MPI_Status *status)
+{
+int retval = MPI_SUCCESS;
+
+  smpi_bench_end();
+  smpi_mpi_sendrecv( sendbuf, sendcount, sendtype, dest, sendtag, 
+		         recvbuf, recvcount, recvtype, source, recvtag,
+		         comm, status);
+  smpi_bench_begin();
+
+  return retval;
+
+	
+}
+
 /**
  * MPI_Wait and friends
  **/
@@ -243,6 +300,20 @@ int flat_tree_bcast(void *buf, int count, MPI_Datatype datatype, int root,
         return(retval);
 
 }
+/**
+ * Bcast internal level 
+ **/
+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;
+}
 
 /**
  * Bcast user entry point
@@ -253,10 +324,7 @@ int SMPI_MPI_Bcast(void *buf, int count, MPI_Datatype datatype, int root,
   int retval = MPI_SUCCESS;
 
   smpi_bench_end();
-
-  //retval = flat_tree_bcast(buf, count, datatype, root, comm);
-  retval = nary_tree_bcast(buf, count, datatype, root, comm, 2 );
-
+  smpi_mpi_bcast(buf,count,datatype,root,comm);
   smpi_bench_begin();
 
   return retval;
@@ -264,7 +332,7 @@ int SMPI_MPI_Bcast(void *buf, int count, MPI_Datatype datatype, int root,
 
 
 
-//#ifdef DEBUG_REDUCE
+#ifdef DEBUG_REDUCE
 /**
  * debugging helper function
  **/
@@ -293,18 +361,18 @@ static void print_buffer_double(void *buf, int len, char *msg, int rank)
 }
 
 
-//#endif
+#endif
 /**
- * MPI_Reduce
+ * MPI_Reduce internal level 
  **/
-int SMPI_MPI_Reduce(void *sendbuf, void *recvbuf, int count,
+int smpi_mpi_reduce(void *sendbuf, void *recvbuf, int count,
                 MPI_Datatype datatype, MPI_Op op, int root, MPI_Comm comm)
 {
         int retval = MPI_SUCCESS;
         int rank;
         int size;
         int i;
-        int tag = 0;
+        int system_tag = 666;
         smpi_mpi_request_t *requests;
         smpi_mpi_request_t request;
 
@@ -312,14 +380,14 @@ 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);
@@ -344,7 +412,7 @@ 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]);
@@ -355,8 +423,9 @@ 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
@@ -375,41 +444,54 @@ int SMPI_MPI_Reduce(void *sendbuf, void *recvbuf, int count,
                 xbt_free(requests);
                 xbt_free(tmpbufs);
         }
-        smpi_bench_begin();
         return retval;
 }
 
+/**
+ * MPI_Reduce user entry point
+ **/
+int SMPI_MPI_Reduce(void *sendbuf, void *recvbuf, int count,
+		    MPI_Datatype datatype, MPI_Op op, int root, MPI_Comm comm)
+{
+int retval = MPI_SUCCESS;
+
+	  smpi_bench_end();
+
+	  retval = smpi_mpi_reduce(sendbuf, recvbuf, count, datatype, op, root, comm);
+
+	  smpi_bench_begin();
+	  return retval;
+}
+
+
+
 /**
  * MPI_Allreduce
  *
- * Same as MPI_REDUCE except that the result appears in the receive buffer of all the group members.
+ * Same as MPI_Reduce except that the result appears in the receive buffer of all the group members.
  **/
 int SMPI_MPI_Allreduce( void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype,
-   		         MPI_Op op, MPI_Comm comm );
-int SMPI_MPI_Allreduce( void *sendbuf, void *recvbuf, int count, MPI_Datatype datatype,
-   		         MPI_Op op, MPI_Comm comm )
+		    MPI_Op op, MPI_Comm comm )
 {
-  int retval = MPI_SUCCESS;
-  int root=1;  // arbitrary choice
+int retval = MPI_SUCCESS;
+int root=0;  // arbitrary choice
 
-  smpi_bench_end();
+	  smpi_bench_end();
 
-  retval = SMPI_MPI_Reduce( sendbuf, recvbuf, count, datatype, op, root, comm);
-  if (MPI_SUCCESS != retval)
-	    return(retval);
+	  retval = smpi_mpi_reduce( sendbuf, recvbuf, count, datatype, op, root, comm);
+	  if (MPI_SUCCESS != retval)
+		    return(retval);
 
-  retval = SMPI_MPI_Bcast( sendbuf, count, datatype, root, comm);
-  smpi_bench_begin();
-  return( retval );
+	  retval = smpi_mpi_bcast( sendbuf, count, datatype, root, comm);
+
+	  smpi_bench_end();
+	  return( retval );
 }
 
 
 /**
  * MPI_Scatter user entry point
  **/
-//int SMPI_MPI_Scatter(void *sendbuf, int sendcount, MPI_Datatype datatype, 
-//		         void *recvbuf, int recvcount, MPI_Datatype recvtype,int root,
-//		        MPI_Comm comm);
 int SMPI_MPI_Scatter(void *sendbuf, int sendcount, MPI_Datatype datatype, 
 		         void *recvbuf, int recvcount, MPI_Datatype recvtype,
 			   int root, MPI_Comm comm)
@@ -477,6 +559,66 @@ int SMPI_MPI_Scatter(void *sendbuf, int sendcount, MPI_Datatype datatype,
 }
 
 
+/**
+ * MPI_Alltoall user entry point
+ * 
+ * Uses the logic of OpenMPI (upto 1.2.7 or greater) for the optimizations
+ * ompi/mca/coll/tuned/coll_tuned_module.c
+ **/
+int SMPI_MPI_Alltoall(void *sendbuf, int sendcount, MPI_Datatype datatype, 
+		          void *recvbuf, int recvcount, MPI_Datatype recvtype,
+			    MPI_Comm comm)
+{
+  int retval = MPI_SUCCESS;
+  int block_dsize;
+  int rank;
+
+  smpi_bench_end();
+
+  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,
+				  recvbuf, recvcount, recvtype, comm);
+
+  } else if (block_dsize < 3000) {
+	    retval = smpi_coll_tuned_alltoall_basic_linear(sendbuf, sendcount, datatype,
+				  recvbuf, recvcount, recvtype, comm);
+  } else {
+
+  retval = smpi_coll_tuned_alltoall_pairwise(sendbuf, sendcount, datatype,
+				  recvbuf, recvcount, recvtype, comm);
+  }
+
+  smpi_bench_begin();
+
+  return retval;
+}
+
+/**
+ * MPI_Alltoallv user entry point
+ * 
+ * As in OpenMPI, alltoallv is not optimized
+ * ompi/mca/coll/basic/coll_basic_alltoallv.c 
+ **/
+int SMPI_MPI_Alltoallv(void *sendbuf, int *scounts, int *sdisps, MPI_Datatype datatype, 
+		           void *recvbuf, int *rcounts, int *rdisps, MPI_Datatype recvtype,
+			     MPI_Comm comm)
+{
+  int retval = MPI_SUCCESS;
+  int rank;
+
+  rank = smpi_mpi_comm_rank(comm);
+  DEBUG1("<%d> basic alltoallv() called.",rank);
+
+  retval = smpi_coll_basic_alltoallv(sendbuf, scounts, sdisps, datatype, 
+                                     recvbuf, rcounts, rdisps, recvtype,
+                                     comm); 
+
+  return retval;
+}