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

[simgrid.git] / src / smpi / colls / allreduce-smp-rsag-rab.c

diff --git a/src/smpi/colls/allreduce-smp-rsag-rab.c b/src/smpi/colls/allreduce-smp-rsag-rab.c
index cec7521..1ac827e 100644 (file)
--- a/src/smpi/colls/allreduce-smp-rsag-rab.c
+++ b/src/smpi/colls/allreduce-smp-rsag-rab.c
@@ -1,14 +1,15 @@
-#include "colls_private.h"
+/* 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. */
+
 /* 
  * implemented by Pitch Patarasuk, 07/01/2007
  */
+#include "colls_private.h"
 //#include <star-reduction.c>
 
-/* change number of core per smp-node
-   we assume that number of core per process will be the same for all implementations */
-#ifndef NUM_CORE
-#define NUM_CORE 8
-#endif
 
 /*
 This fucntion performs all-reduce operation as follow.
@@ -26,13 +27,23 @@ int smpi_coll_tuned_allreduce_smp_rsag_rab(void *sbuf, void *rbuf, int count,
   int tag = COLL_TAG_ALLREDUCE;
   int mask, src, dst;
   MPI_Status status;
-  int num_core = NUM_CORE;
+  if(smpi_comm_get_leaders_comm(comm)==MPI_COMM_NULL){
+    smpi_comm_init_smp(comm);
+  }
+  int num_core=1;
+  if (smpi_comm_is_uniform(comm)){
+    num_core = smpi_comm_size(smpi_comm_get_intra_comm(comm));
+  }
 
   comm_size = smpi_comm_size(comm);
+
+  if((comm_size&(comm_size-1)))
+    THROWF(arg_error,0, "allreduce smp rsag rab algorithm can't be used with non power of two number of processes ! ");
+
   rank = smpi_comm_rank(comm);
   MPI_Aint extent;
   extent = smpi_datatype_get_extent(dtype);
-  tmp_buf = (void *) xbt_malloc(count * extent);
+  tmp_buf = (void *) smpi_get_tmp_sendbuffer(count * extent);
 
   int intra_rank, inter_rank;
   intra_rank = rank % num_core;
@@ -78,19 +89,14 @@ int smpi_coll_tuned_allreduce_smp_rsag_rab(void *sbuf, void *rbuf, int count,
     recv_chunk = extent * count / (comm_size / num_core);
 
     mask = 1;
-    i = 0;
     curr_count = count / 2;
     int phase = 0;
     base_offset = 0;
-    send_base_offset = 0;
-    recv_base_offset = 0;
 
     while (mask < (comm_size / num_core)) {
       dst = inter_rank ^ mask;
 
       // compute offsets
-      send_base_offset = base_offset;
-
       // right-handside
       if (inter_rank & mask) {
         recv_base_offset = base_offset + curr_count;
@@ -200,6 +206,6 @@ int smpi_coll_tuned_allreduce_smp_rsag_rab(void *sbuf, void *rbuf, int count,
   }
 
 
-  free(tmp_buf);
+  smpi_free_tmp_buffer(tmp_buf);
   return MPI_SUCCESS;
 }