oops

[simgrid.git] / src / smpi / colls / allreduce-smp-binomial-pipeline.c

diff --git a/src/smpi/colls/allreduce-smp-binomial-pipeline.c b/src/smpi/colls/allreduce-smp-binomial-pipeline.c
index 05ecfd6..06a80a7 100644 (file)
--- a/src/smpi/colls/allreduce-smp-binomial-pipeline.c
+++ b/src/smpi/colls/allreduce-smp-binomial-pipeline.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"
 /* IMPLEMENTED BY PITCH PATARASUK 
    Non-topoloty-specific (however, number of cores/node need to be changed) 
@@ -6,11 +12,7 @@
    inter-communication
    The communication are done in a pipeline fashion */
 
-/* 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 is a default segment size for pipelining, 
    but it is typically passed as a command line argument */
@@ -29,18 +31,6 @@ int allreduce_smp_binomial_pipeline_segment_size = 4096;
    This code assume commutative and associative reduce operator (MPI_SUM, MPI_MAX, etc).
 */
 
-#ifndef MPICH2
-extern void *MPIR_ToPointer();
-
-struct MPIR_OP {
-  MPI_User_function *op;
-  int commute, permanent;
-};
-
-#else
-extern MPI_User_function *MPIR_Op_table[];
-#endif
-
 /*
 This fucntion performs all-reduce operation as follow. ** in a pipeline fashion **
 1) binomial_tree reduce inside each SMP node
@@ -55,24 +45,22 @@ int smpi_coll_tuned_allreduce_smp_binomial_pipeline(void *send_buf,
 {
   int comm_size, rank;
   void *tmp_buf;
-  int tag = 50;
+  int tag = COLL_TAG_ALLREDUCE;
   int mask, src, dst;
   MPI_Status status;
-  int num_core = NUM_CORE;
-
-  MPI_User_function *uop;
-#ifndef MPICH2
-  struct MPIR_OP *op_ptr = MPIR_ToPointer(op);
-  uop = (MPI_User_function *) op_ptr->op;
-#else
-  uop = MPIR_Op_table[op % 16 - 1];
-#endif
+  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);
   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;
@@ -111,7 +99,7 @@ int smpi_coll_tuned_allreduce_smp_binomial_pipeline(void *send_buf,
           if (src < comm_size) {
             recv_offset = phase * pcount * extent;
             smpi_mpi_recv(tmp_buf, pcount, dtype, src, tag, comm, &status);
-            (*uop) (tmp_buf, (char *)recv_buf + recv_offset, &pcount, &dtype);
+            smpi_op_apply(op, tmp_buf, (char *)recv_buf + recv_offset, &pcount, &dtype);
           }
         } else {
           send_offset = phase * pcount * extent;
@@ -135,7 +123,7 @@ int smpi_coll_tuned_allreduce_smp_binomial_pipeline(void *send_buf,
             if (src < comm_size) {
               recv_offset = (phase - 1) * pcount * extent;
               smpi_mpi_recv(tmp_buf, pcount, dtype, src, tag, comm, &status);
-              (*uop) (tmp_buf, (char *)recv_buf + recv_offset, &pcount, &dtype);
+              smpi_op_apply(op, tmp_buf, (char *)recv_buf + recv_offset, &pcount, &dtype);
             }
           } else {
             dst = (inter_rank & (~mask)) * num_core;
@@ -209,6 +197,6 @@ int smpi_coll_tuned_allreduce_smp_binomial_pipeline(void *send_buf,
     }
   }                             // for phase
 
-  free(tmp_buf);
+  smpi_free_tmp_buffer(tmp_buf);
   return MPI_SUCCESS;
 }