Have replay always use shared buffers instead of allocating new ones, even inside...

[simgrid.git] / src / smpi / colls / allreduce-rab1.c

diff --git a/src/smpi/colls/allreduce-rab1.c b/src/smpi/colls/allreduce-rab1.c
index ae72f74..0a04e54 100644 (file)
--- a/src/smpi/colls/allreduce-rab1.c
+++ b/src/smpi/colls/allreduce-rab1.c
@@ -1,3 +1,9 @@
+/* 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"
 //#include <star-reduction.c>
 
@@ -8,7 +14,7 @@ int smpi_coll_tuned_allreduce_rab1(void *sbuff, void *rbuff,
 {
   MPI_Status status;
   MPI_Aint extent;
-  int tag = 4321, rank, nprocs, send_size, newcnt, share;
+  int tag = COLL_TAG_ALLREDUCE, rank, nprocs, send_size, newcnt, share;
   int pof2 = 1, mask, send_idx, recv_idx, dst, send_cnt, recv_cnt;
 
   void *recv, *tmp_buf;
@@ -16,6 +22,9 @@ int smpi_coll_tuned_allreduce_rab1(void *sbuff, void *rbuff,
   rank = smpi_comm_rank(comm);
   nprocs = smpi_comm_size(comm);
 
+  if((nprocs&(nprocs-1)))
+    THROWF(arg_error,0, "allreduce rab1 algorithm can't be used with non power of two number of processes ! ");
+
   extent = smpi_datatype_get_extent(dtype);
 
   pof2 = 1;
@@ -23,7 +32,6 @@ int smpi_coll_tuned_allreduce_rab1(void *sbuff, void *rbuff,
     pof2 <<= 1;
   pof2 >>= 1;
 
-  mask = 1;
   send_idx = recv_idx = 0;
 
   // uneven count
@@ -31,8 +39,8 @@ int smpi_coll_tuned_allreduce_rab1(void *sbuff, void *rbuff,
     send_size = (count + nprocs) / nprocs;
     newcnt = send_size * nprocs;
 
-    recv = (void *) xbt_malloc(extent * newcnt);
-    tmp_buf = (void *) xbt_malloc(extent * newcnt);
+    recv = (void *) smpi_get_tmp_recvbuffer(extent * newcnt);
+    tmp_buf = (void *) smpi_get_tmp_sendbuffer(extent * newcnt);
     memcpy(recv, sbuff, extent * count);
 
 
@@ -62,13 +70,13 @@ int smpi_coll_tuned_allreduce_rab1(void *sbuff, void *rbuff,
     mpi_coll_allgather_fun(tmp_buf, recv_cnt, dtype, recv, recv_cnt, dtype, comm);
 
     memcpy(rbuff, recv, count * extent);
-    free(recv);
-    free(tmp_buf);
+    smpi_free_tmp_buffer(recv);
+    smpi_free_tmp_buffer(tmp_buf);
 
   }
 
   else {
-    tmp_buf = (void *) xbt_malloc(extent * count);
+    tmp_buf = (void *) smpi_get_tmp_sendbuffer(extent * count);
     memcpy(rbuff, sbuff, count * extent);
     mask = pof2 / 2;
     share = count / pof2;
@@ -94,7 +102,7 @@ int smpi_coll_tuned_allreduce_rab1(void *sbuff, void *rbuff,
 
     memcpy(tmp_buf, (char *) rbuff + recv_idx * extent, recv_cnt * extent);
     mpi_coll_allgather_fun(tmp_buf, recv_cnt, dtype, rbuff, recv_cnt, dtype, comm);
-    free(tmp_buf);
+    smpi_free_tmp_buffer(tmp_buf);
   }
 
   return MPI_SUCCESS;