Merge branch 'master' of git+ssh://scm.gforge.inria.fr//gitroot/simgrid/simgrid

[simgrid.git] / src / smpi / colls / bcast-SMP-binary.c

diff --git a/src/smpi/colls/bcast-SMP-binary.c b/src/smpi/colls/bcast-SMP-binary.c
index ec717e2..5264ca3 100644 (file)
--- a/src/smpi/colls/bcast-SMP-binary.c
+++ b/src/smpi/colls/bcast-SMP-binary.c
@@ -1,7 +1,11 @@
+/* Copyright (c) 2013-2014. The SimGrid Team.
+ * All rights reserved.                                                     */
+
+/* This program is free software; you can redistribute it and/or modify it
+ * under the terms of the license (GNU LGPL) which comes with this package. */
+
 #include "colls_private.h"
-#ifndef NUM_CORE
-#define NUM_CORE 8
-#endif
+
 
 int bcast_SMP_binary_segment_byte = 8192;
 
@@ -9,7 +13,7 @@ int smpi_coll_tuned_bcast_SMP_binary(void *buf, int count,
                                      MPI_Datatype datatype, int root,
                                      MPI_Comm comm)
 {
-  int tag = 5000;
+  int tag = COLL_TAG_BCAST;
   MPI_Status status;
   MPI_Request request;
   MPI_Request *request_array;
@@ -21,23 +25,34 @@ int smpi_coll_tuned_bcast_SMP_binary(void *buf, int count,
 
   rank = smpi_comm_rank(comm);
   size = smpi_comm_size(comm);
+  if(smpi_comm_get_leaders_comm(comm)==MPI_COMM_NULL){
+    smpi_comm_init_smp(comm);
+  }
+  int host_num_core=1;
+  if (smpi_comm_is_uniform(comm)){
+    host_num_core = smpi_comm_size(smpi_comm_get_intra_comm(comm));
+  }else{
+    //implementation buggy in this case
+    return smpi_coll_tuned_bcast_mpich( buf , count, datatype,
+              root, comm);
+  }
 
   int segment = bcast_SMP_binary_segment_byte / extent;
   int pipe_length = count / segment;
   int remainder = count % segment;
 
-  int to_intra_left = (rank / NUM_CORE) * NUM_CORE + (rank % NUM_CORE) * 2 + 1;
-  int to_intra_right = (rank / NUM_CORE) * NUM_CORE + (rank % NUM_CORE) * 2 + 2;
-  int to_inter_left = ((rank / NUM_CORE) * 2 + 1) * NUM_CORE;
-  int to_inter_right = ((rank / NUM_CORE) * 2 + 2) * NUM_CORE;
-  int from_inter = (((rank / NUM_CORE) - 1) / 2) * NUM_CORE;
-  int from_intra = (rank / NUM_CORE) * NUM_CORE + ((rank % NUM_CORE) - 1) / 2;
+  int to_intra_left = (rank / host_num_core) * host_num_core + (rank % host_num_core) * 2 + 1;
+  int to_intra_right = (rank / host_num_core) * host_num_core + (rank % host_num_core) * 2 + 2;
+  int to_inter_left = ((rank / host_num_core) * 2 + 1) * host_num_core;
+  int to_inter_right = ((rank / host_num_core) * 2 + 2) * host_num_core;
+  int from_inter = (((rank / host_num_core) - 1) / 2) * host_num_core;
+  int from_intra = (rank / host_num_core) * host_num_core + ((rank % host_num_core) - 1) / 2;
   int increment = segment * extent;
 
-  int base = (rank / NUM_CORE) * NUM_CORE;
-  int num_core = NUM_CORE;
-  if (((rank / NUM_CORE) * NUM_CORE) == ((size / NUM_CORE) * NUM_CORE))
-    num_core = size - (rank / NUM_CORE) * NUM_CORE;
+  int base = (rank / host_num_core) * host_num_core;
+  int num_core = host_num_core;
+  if (((rank / host_num_core) * host_num_core) == ((size / host_num_core) * host_num_core))
+    num_core = size - (rank / host_num_core) * host_num_core;
 
   // if root is not zero send to rank zero first
   if (root != 0) {
@@ -49,7 +64,7 @@ int smpi_coll_tuned_bcast_SMP_binary(void *buf, int count,
   // when a message is smaller than a block size => no pipeline 
   if (count <= segment) {
     // case ROOT-of-each-SMP
-    if (rank % NUM_CORE == 0) {
+    if (rank % host_num_core == 0) {
       // case ROOT
       if (rank == 0) {
         //printf("node %d left %d right %d\n",rank,to_inter_left,to_inter_right);
@@ -114,7 +129,7 @@ int smpi_coll_tuned_bcast_SMP_binary(void *buf, int count,
         (MPI_Status *) xbt_malloc((size + pipe_length) * sizeof(MPI_Status));
 
     // case ROOT-of-each-SMP
-    if (rank % NUM_CORE == 0) {
+    if (rank % host_num_core == 0) {
       // case ROOT
       if (rank == 0) {
         for (i = 0; i < pipe_length; i++) {