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

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

diff --git a/src/smpi/colls/allreduce-smp-rsag-lr.c b/src/smpi/colls/allreduce-smp-rsag-lr.c
index bc72027..fa9296f 100644 (file)
--- a/src/smpi/colls/allreduce-smp-rsag-lr.c
+++ b/src/smpi/colls/allreduce-smp-rsag-lr.c
@@ -1,12 +1,12 @@
+/* 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>
 
-/* 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.
 1) binomial_tree reduce inside each SMP node
@@ -23,10 +23,13 @@ int smpi_coll_tuned_allreduce_smp_rsag_lr(void *send_buf, void *recv_buf,
   int tag = COLL_TAG_ALLREDUCE;
   int mask, src, dst;
   MPI_Status status;
-  int num_core = simcall_host_get_core(SIMIX_host_self());
-  // do we use the default one or the number of cores in the platform ?
-  // if the number of cores is one, the platform may be simulated with 1 node = 1 core
-  if (num_core == 1) 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));
+  }
   /*
      #ifdef MPICH2_REDUCTION
      MPI_User_function * uop = MPIR_Op_table[op % 16 - 1];
@@ -41,7 +44,7 @@ int smpi_coll_tuned_allreduce_smp_rsag_lr(void *send_buf, void *recv_buf,
   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;
@@ -161,10 +164,10 @@ int smpi_coll_tuned_allreduce_smp_rsag_lr(void *send_buf, void *recv_buf,
 
 
 
-  /*
      // INTER_binomial_reduce
 
      // only root node for each SMP
+  /*
      if (intra_rank == 0) {
 
      mask = 1;
@@ -188,38 +191,35 @@ int smpi_coll_tuned_allreduce_smp_rsag_lr(void *send_buf, void *recv_buf,
      }
    */
 
-  /*
      // INTER_binomial_bcast
 
 
-     if (intra_rank == 0) {
-     mask = 1;
-     while (mask < inter_comm_size) {
-     if (inter_rank & mask) {
-     src = (inter_rank - mask) * num_core;
+//     if (intra_rank == 0) {
+//     mask = 1;
+//     while (mask < inter_comm_size) {
+//     if (inter_rank & mask) {
+//     src = (inter_rank - mask) * num_core;
      //printf("Node %d recv from node %d when mask is %d\n", rank, src, mask);
-     smpi_mpi_recv(recv_buf, count, dtype, src, tag, comm, &status);
-     break;
-     }
-     mask <<= 1;
-     }
-
-     mask >>= 1;
+//     smpi_mpi_recv(recv_buf, count, dtype, src, tag, comm, &status);
+//     break;
+//     }
+//     mask <<= 1;
+//     }
+//
+//     mask >>= 1;
      //printf("My rank = %d my mask = %d\n", rank,mask);
 
-     while (mask > 0) {
-     if (inter_rank < inter_comm_size) {
-     dst = (inter_rank + mask) * num_core;
-     if (dst < comm_size) {
-     //printf("Node %d send to node %d when mask is %d\n", rank, dst, mask);
-     smpi_mpi_send(recv_buf, count, dtype, dst, tag, comm);
-     }
-     }
-     mask >>= 1;
-     }
-     }
-   */
-
+//     while (mask > 0) {
+//     if (inter_rank < inter_comm_size) {
+//     dst = (inter_rank + mask) * num_core;
+//     if (dst < comm_size) {
+//     //printf("Node %d send to node %d when mask is %d\n", rank, dst, mask);
+//     smpi_mpi_send(recv_buf, count, dtype, dst, tag, comm);
+//     }
+//     }
+//     mask >>= 1;
+//     }
+//     }
 
   // INTRA_binomial_bcast
 
@@ -252,6 +252,6 @@ int smpi_coll_tuned_allreduce_smp_rsag_lr(void *send_buf, void *recv_buf,
   }
 
 
-  free(tmp_buf);
+  smpi_free_tmp_buffer(tmp_buf);
   return MPI_SUCCESS;
 }