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

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

diff --git a/src/smpi/colls/allreduce-smp-rsag.c b/src/smpi/colls/allreduce-smp-rsag.c
index a40f189..b4a6885 100644 (file)
--- a/src/smpi/colls/allreduce-smp-rsag.c
+++ b/src/smpi/colls/allreduce-smp-rsag.c
@@ -1,10 +1,10 @@
-#include "colls.h"
+/* Copyright (c) 2013-2014. The SimGrid Team.
+ * All rights reserved.                                                     */
 
-/* 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 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"
 
 /*
 This fucntion performs all-reduce operation as follow.
@@ -19,10 +19,16 @@ int smpi_coll_tuned_allreduce_smp_rsag(void *send_buf, void *recv_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;
+  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];
@@ -33,11 +39,11 @@ int smpi_coll_tuned_allreduce_smp_rsag(void *send_buf, void *recv_buf,
      uop  = op_ptr->op;
      #endif
    */
-  MPI_Comm_size(comm, &comm_size);
-  MPI_Comm_rank(comm, &rank);
+  comm_size = smpi_comm_size(comm);
+  rank = smpi_comm_rank(comm);
   MPI_Aint extent;
-  MPI_Type_extent(dtype, &extent);
-  tmp_buf = (void *) malloc(count * extent);
+  extent = smpi_datatype_get_extent(dtype);
+  tmp_buf = (void *) smpi_get_tmp_sendbuffer(count * extent);
 
   int intra_rank, inter_rank;
   intra_rank = rank % num_core;
@@ -53,7 +59,7 @@ int smpi_coll_tuned_allreduce_smp_rsag(void *send_buf, void *recv_buf,
   }
 
 
-  MPI_Sendrecv(send_buf, count, dtype, rank, tag,
+  smpi_mpi_sendrecv(send_buf, count, dtype, rank, tag,
                recv_buf, count, dtype, rank, tag, comm, &status);
 
 
@@ -64,14 +70,14 @@ int smpi_coll_tuned_allreduce_smp_rsag(void *send_buf, void *recv_buf,
       src = (inter_rank * num_core) + (intra_rank | mask);
       //      if (src < ((inter_rank + 1) * num_core)) {
       if (src < comm_size) {
-        MPI_Recv(tmp_buf, count, dtype, src, tag, comm, &status);
-        star_reduction(op, tmp_buf, recv_buf, &count, &dtype);
+        smpi_mpi_recv(tmp_buf, count, dtype, src, tag, comm, &status);
+        smpi_op_apply(op, tmp_buf, recv_buf, &count, &dtype);
         //printf("Node %d recv from node %d when mask is %d\n", rank, src, mask);
       }
     } else {
 
       dst = (inter_rank * num_core) + (intra_rank & (~mask));
-      MPI_Send(recv_buf, count, dtype, dst, tag, comm);
+      smpi_mpi_send(recv_buf, count, dtype, dst, tag, comm);
       //printf("Node %d send to node %d when mask is %d\n", rank, dst, mask);
       break;
     }
@@ -100,12 +106,12 @@ int smpi_coll_tuned_allreduce_smp_rsag(void *send_buf, void *recv_buf,
           ((inter_rank - 2 - i +
             inter_comm_size) % inter_comm_size) * seg_count * extent;
 
-      MPI_Sendrecv((char *) recv_buf + send_offset, seg_count, dtype, to,
+      smpi_mpi_sendrecv((char *) recv_buf + send_offset, seg_count, dtype, to,
                    tag + i, tmp_buf, seg_count, dtype, from, tag + i, comm,
                    &status);
 
       // result is in rbuf
-      star_reduction(op, tmp_buf, (char *) recv_buf + recv_offset, &seg_count,
+      smpi_op_apply(op, tmp_buf, (char *) recv_buf + recv_offset, &seg_count,
                      &dtype);
     }
 
@@ -119,7 +125,7 @@ int smpi_coll_tuned_allreduce_smp_rsag(void *send_buf, void *recv_buf,
           ((inter_rank - 1 - i +
             inter_comm_size) % inter_comm_size) * seg_count * extent;
 
-      MPI_Sendrecv((char *) recv_buf + send_offset, seg_count, dtype, to,
+      smpi_mpi_sendrecv((char *) recv_buf + send_offset, seg_count, dtype, to,
                    tag + i, (char *) recv_buf + recv_offset, seg_count, dtype,
                    from, tag + i, comm, &status);
 
@@ -128,64 +134,60 @@ int smpi_coll_tuned_allreduce_smp_rsag(void *send_buf, void *recv_buf,
 
 
 
-  /*
      // INTER_binomial_reduce
 
      // only root node for each SMP
-     if (intra_rank == 0) {
-
-     mask = 1;
-     while (mask < inter_comm_size) {
-     if ((mask & inter_rank) == 0) {
-     src = (inter_rank | mask) * num_core;
-     if (src < comm_size) {
-     MPI_Recv(tmp_buf, count, dtype, src, tag, comm, &status);
-     (* uop) (tmp_buf, recv_buf, &count, &dtype);
+//     if (intra_rank == 0) {
+//
+//     mask = 1;
+//     while (mask < inter_comm_size) {
+//     if ((mask & inter_rank) == 0) {
+//     src = (inter_rank | mask) * num_core;
+//     if (src < comm_size) {
+//     smpi_mpi_recv(tmp_buf, count, dtype, src, tag, comm, &status);
+//     (* uop) (tmp_buf, recv_buf, &count, &dtype);
      //printf("Node %d recv from node %d when mask is %d\n", rank, src, mask);
-     }
-     }
-     else {
-     dst = (inter_rank & (~mask)) * num_core;
-     MPI_Send(recv_buf, count, dtype, dst, tag, comm);
+//     }
+//     }
+//     else {
+//     dst = (inter_rank & (~mask)) * num_core;
+//     smpi_mpi_send(recv_buf, count, dtype, dst, tag, comm);
      //printf("Node %d send to node %d when mask is %d\n", rank, dst, mask);
-     break;
-     }
-     mask <<=1;
-     }
-     }
-   */
+//     break;
+//     }
+//     mask <<=1;
+//     }
+//     }
 
-  /*
      // 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);
-     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) {
+//     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);
-     MPI_Send(recv_buf, count, dtype, dst, tag, comm);
-     }
-     }
-     mask >>= 1;
-     }
-     }
-   */
+//     smpi_mpi_send(recv_buf, count, dtype, dst, tag, comm);
+//     }
+//     }
+//     mask >>= 1;
+//     }
+//     }
 
 
   // INTRA_binomial_bcast
@@ -200,7 +202,7 @@ int smpi_coll_tuned_allreduce_smp_rsag(void *send_buf, void *recv_buf,
     if (intra_rank & mask) {
       src = (inter_rank * num_core) + (intra_rank - mask);
       //printf("Node %d recv from node %d when mask is %d\n", rank, src, mask);
-      MPI_Recv(recv_buf, count, dtype, src, tag, comm, &status);
+      smpi_mpi_recv(recv_buf, count, dtype, src, tag, comm, &status);
       break;
     }
     mask <<= 1;
@@ -213,12 +215,11 @@ int smpi_coll_tuned_allreduce_smp_rsag(void *send_buf, void *recv_buf,
     dst = (inter_rank * num_core) + (intra_rank + mask);
     if (dst < comm_size) {
       //printf("Node %d send to node %d when mask is %d\n", rank, dst, mask);
-      MPI_Send(recv_buf, count, dtype, dst, tag, comm);
+      smpi_mpi_send(recv_buf, count, dtype, dst, tag, comm);
     }
     mask >>= 1;
   }
 
-
-  free(tmp_buf);
+  smpi_free_tmp_buffer(tmp_buf);
   return MPI_SUCCESS;
 }