A
lgorithmique
N
umérique
D
istribuée
Public GIT Repository
projects
/
simgrid.git
/ blobdiff
commit
grep
author
committer
pickaxe
?
search:
re
summary
|
shortlog
|
log
|
commit
|
commitdiff
|
tree
raw
|
inline
| side by side
Merge branch 'master' of git+ssh://scm.gforge.inria.fr//gitroot/simgrid/simgrid
[simgrid.git]
/
src
/
smpi
/
colls
/
allreduce-rab-rdb.c
diff --git
a/src/smpi/colls/allreduce-rab-rdb.c
b/src/smpi/colls/allreduce-rab-rdb.c
index
6d49e71
..
7499785
100644
(file)
--- a/
src/smpi/colls/allreduce-rab-rdb.c
+++ b/
src/smpi/colls/allreduce-rab-rdb.c
@@
-1,34
+1,30
@@
+/* 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"
int smpi_coll_tuned_allreduce_rab_rdb(void *sbuff, void *rbuff, int count,
MPI_Datatype dtype, MPI_Op op,
MPI_Comm comm)
{
#include "colls_private.h"
int smpi_coll_tuned_allreduce_rab_rdb(void *sbuff, void *rbuff, int count,
MPI_Datatype dtype, MPI_Op op,
MPI_Comm comm)
{
- int nprocs, rank, type_size, tag = 543;
- int mask, dst, pof2, newrank, rem, newdst, i,
+ int tag = COLL_TAG_ALLREDUCE;
+ unsigned int mask, pof2;
+ int dst, newrank, rem, newdst, i,
send_idx, recv_idx, last_idx, send_cnt, recv_cnt, *cnts, *disps;
send_idx, recv_idx, last_idx, send_cnt, recv_cnt, *cnts, *disps;
- MPI_Aint
lb,
extent;
+ MPI_Aint extent;
MPI_Status status;
void *tmp_buf = NULL;
MPI_Status status;
void *tmp_buf = NULL;
-#ifdef MPICH2_REDUCTION
- MPI_User_function *uop = MPIR_Op_table[op % 16 - 1];
-#else
- MPI_User_function *uop;
- struct MPIR_OP *op_ptr;
- op_ptr = (MPI_User_function *) MPIR_ToPointer(op);
- uop = op_ptr->op;
-#endif
-
- nprocs = smpi_comm_size(comm);
- rank = smpi_comm_rank(comm);
+ unsigned int nprocs = smpi_comm_size(comm);
+ unsigned int rank = smpi_comm_rank(comm);
extent = smpi_datatype_get_extent(dtype);
extent = smpi_datatype_get_extent(dtype);
- tmp_buf = (void *) xbt_malloc(count * extent);
-
- MPIR_Localcopy(sbuff, count, dtype, rbuff, count, dtype);
+ tmp_buf = (void *) smpi_get_tmp_sendbuffer(count * extent);
-
type_size = smpi_datatype_size(
dtype);
+
smpi_datatype_copy(sbuff, count, dtype, rbuff, count,
dtype);
// find nearest power-of-two less than or equal to comm_size
pof2 = 1;
// find nearest power-of-two less than or equal to comm_size
pof2 = 1;
@@
-60,7
+56,7
@@
int smpi_coll_tuned_allreduce_rab_rdb(void *sbuff, void *rbuff, int count,
// do the reduction on received data. since the
// ordering is right, it doesn't matter whether
// the operation is commutative or not.
// do the reduction on received data. since the
// ordering is right, it doesn't matter whether
// the operation is commutative or not.
-
(*uop) (
tmp_buf, rbuff, &count, &dtype);
+
smpi_op_apply(op,
tmp_buf, rbuff, &count, &dtype);
// change the rank
newrank = rank / 2;
// change the rank
newrank = rank / 2;
@@
-129,8
+125,8
@@
int smpi_coll_tuned_allreduce_rab_rdb(void *sbuff, void *rbuff, int count,
// This algorithm is used only for predefined ops
// and predefined ops are always commutative.
// This algorithm is used only for predefined ops
// and predefined ops are always commutative.
-
(*uop) (
(char *) tmp_buf + disps[recv_idx] * extent,
- (char *) rbuff + disps[recv_idx] * extent, &recv_cnt, &dtype);
+
smpi_op_apply(op,
(char *) tmp_buf + disps[recv_idx] * extent,
+
(char *) rbuff + disps[recv_idx] * extent, &recv_cnt, &dtype);
// update send_idx for next iteration
send_idx = recv_idx;
// update send_idx for next iteration
send_idx = recv_idx;
@@
-195,6
+191,6
@@
int smpi_coll_tuned_allreduce_rab_rdb(void *sbuff, void *rbuff, int count,
smpi_mpi_recv(rbuff, count, dtype, rank + 1, tag, comm, &status);
}
smpi_mpi_recv(rbuff, count, dtype, rank + 1, tag, comm, &status);
}
-
free
(tmp_buf);
+
smpi_free_tmp_buffer
(tmp_buf);
return MPI_SUCCESS;
}
return MPI_SUCCESS;
}