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

[simgrid.git] / src / smpi / colls / allreduce-rab2.c

diff --git a/src/smpi/colls/allreduce-rab2.c b/src/smpi/colls/allreduce-rab2.c
index d09cd11..372d5e8 100644 (file)
--- a/src/smpi/colls/allreduce-rab2.c
+++ b/src/smpi/colls/allreduce-rab2.c
@@ -1,4 +1,10 @@
-#include "colls.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. */
+
+#include "colls_private.h"
 //#include <star-reduction.c>
 
 // this requires that count >= NP
@@ -20,11 +26,11 @@ int smpi_coll_tuned_allreduce_rab2(void *sbuff, void *rbuff,
      uop  = op_ptr->op;
      #endif
    */
-  MPI_Comm_rank(comm, &rank);
-  MPI_Comm_size(comm, &nprocs);
+  rank = smpi_comm_rank(comm);
+  nprocs = smpi_comm_size(comm);
 
 
-  MPI_Type_extent(dtype, &s_extent);
+  s_extent = smpi_datatype_get_extent(dtype);
 
   // uneven count
   if (count % nprocs) {
@@ -34,45 +40,45 @@ int smpi_coll_tuned_allreduce_rab2(void *sbuff, void *rbuff,
       send_size = (count + nprocs) / nprocs;
     nbytes = send_size * s_extent;
 
-    send = (void *) malloc(s_extent * send_size * nprocs);
-    recv = (void *) malloc(s_extent * send_size * nprocs);
-    tmp = (void *) malloc(nbytes);
+    send = (void *) smpi_get_tmp_sendbuffer(s_extent * send_size * nprocs);
+    recv = (void *) smpi_get_tmp_recvbuffer(s_extent * send_size * nprocs);
+    tmp = (void *) smpi_get_tmp_sendbuffer(nbytes);
 
     memcpy(send, sbuff, s_extent * count);
 
-    MPI_Alltoall(send, send_size, dtype, recv, send_size, dtype, comm);
+    mpi_coll_alltoall_fun(send, send_size, dtype, recv, send_size, dtype, comm);
 
     memcpy(tmp, recv, nbytes);
 
     for (i = 1, s_offset = nbytes; i < nprocs; i++, s_offset = i * nbytes)
-      star_reduction(op, (char *) recv + s_offset, tmp, &send_size, &dtype);
+      smpi_op_apply(op, (char *) recv + s_offset, tmp, &send_size, &dtype);
 
-    MPI_Allgather(tmp, send_size, dtype, recv, send_size, dtype, comm);
+    mpi_coll_allgather_fun(tmp, send_size, dtype, recv, send_size, dtype, comm);
     memcpy(rbuff, recv, count * s_extent);
 
-    free(recv);
-    free(tmp);
-    free(send);
+    smpi_free_tmp_buffer(recv);
+    smpi_free_tmp_buffer(tmp);
+    smpi_free_tmp_buffer(send);
   } else {
     send = sbuff;
     send_size = count / nprocs;
     nbytes = send_size * s_extent;
     r_offset = rank * nbytes;
 
-    recv = (void *) malloc(s_extent * send_size * nprocs);
+    recv = (void *) smpi_get_tmp_recvbuffer(s_extent * send_size * nprocs);
 
-    MPI_Alltoall(send, send_size, dtype, recv, send_size, dtype, comm);
+    mpi_coll_alltoall_fun(send, send_size, dtype, recv, send_size, dtype, comm);
 
     memcpy((char *) rbuff + r_offset, recv, nbytes);
 
     for (i = 1, s_offset = nbytes; i < nprocs; i++, s_offset = i * nbytes)
-      star_reduction(op, (char *) recv + s_offset, (char *) rbuff + r_offset,
+      smpi_op_apply(op, (char *) recv + s_offset, (char *) rbuff + r_offset,
                      &send_size, &dtype);
 
-    MPI_Allgather((char *) rbuff + r_offset, send_size, dtype, rbuff, send_size,
+    mpi_coll_allgather_fun((char *) rbuff + r_offset, send_size, dtype, rbuff, send_size,
                   dtype, comm);
-    free(recv);
+    smpi_free_tmp_buffer(recv);
   }
 
-  return 0;
+  return MPI_SUCCESS;
 }