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
avoid warning by checking index in the right order
[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
2c7b49b
..
ce690e2
100644
(file)
--- a/
src/smpi/colls/allreduce-rab-rdb.c
+++ b/
src/smpi/colls/allreduce-rab-rdb.c
@@
-1,34
+1,29
@@
-#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"
int smpi_coll_tuned_allreduce_rab_rdb(void *sbuff, void *rbuff, int count,
MPI_Datatype dtype, MPI_Op op,
MPI_Comm comm)
{
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, t
ype_size, tag = 543
;
+ int nprocs, rank, t
ag = COLL_TAG_ALLREDUCE
;
int mask, dst, pof2, newrank, rem, newdst, i,
send_idx, recv_idx, last_idx, send_cnt, recv_cnt, *cnts, *disps;
int mask, dst, pof2, newrank, rem, newdst, i,
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);
- MPI_Comm_size(comm, &nprocs);
- MPI_Comm_rank(comm, &rank);
-
- MPI_Type_extent(dtype, &extent);
+ extent = smpi_datatype_get_extent(dtype);
tmp_buf = (void *) xbt_malloc(count * extent);
tmp_buf = (void *) xbt_malloc(count * extent);
- MPIR_Localcopy(sbuff, count, dtype, rbuff, count, dtype);
-
- MPI_Type_size(dtype, &type_size);
+ 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;
@@
-48,7
+43,7
@@
int smpi_coll_tuned_allreduce_rab_rdb(void *sbuff, void *rbuff, int count,
// even
if (rank % 2 == 0) {
// even
if (rank % 2 == 0) {
-
MPI_S
end(rbuff, count, dtype, rank + 1, tag, comm);
+
smpi_mpi_s
end(rbuff, count, dtype, rank + 1, tag, comm);
// temporarily set the rank to -1 so that this
// process does not pariticipate in recursive
// temporarily set the rank to -1 so that this
// process does not pariticipate in recursive
@@
-56,11
+51,11
@@
int smpi_coll_tuned_allreduce_rab_rdb(void *sbuff, void *rbuff, int count,
newrank = -1;
} else // odd
{
newrank = -1;
} else // odd
{
-
MPI_R
ecv(tmp_buf, count, dtype, rank - 1, tag, comm, &status);
+
smpi_mpi_r
ecv(tmp_buf, count, dtype, rank - 1, tag, comm, &status);
// 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;
@@
-84,8
+79,8
@@
int smpi_coll_tuned_allreduce_rab_rdb(void *sbuff, void *rbuff, int count,
// reduce-scatter, calculate the count that each process receives
// and the displacement within the buffer
// reduce-scatter, calculate the count that each process receives
// and the displacement within the buffer
- cnts = (int *) malloc(pof2 * sizeof(int));
- disps = (int *) malloc(pof2 * sizeof(int));
+ cnts = (int *)
xbt_
malloc(pof2 * sizeof(int));
+ disps = (int *)
xbt_
malloc(pof2 * sizeof(int));
for (i = 0; i < (pof2 - 1); i++)
cnts[i] = count / pof2;
for (i = 0; i < (pof2 - 1); i++)
cnts[i] = count / pof2;
@@
-119,7
+114,7
@@
int smpi_coll_tuned_allreduce_rab_rdb(void *sbuff, void *rbuff, int count,
}
// Send data from recvbuf. Recv into tmp_buf
}
// Send data from recvbuf. Recv into tmp_buf
-
MPI_S
endrecv((char *) rbuff + disps[send_idx] * extent, send_cnt,
+
smpi_mpi_s
endrecv((char *) rbuff + disps[send_idx] * extent, send_cnt,
dtype, dst, tag,
(char *) tmp_buf + disps[recv_idx] * extent, recv_cnt,
dtype, dst, tag, comm, &status);
dtype, dst, tag,
(char *) tmp_buf + disps[recv_idx] * extent, recv_cnt,
dtype, dst, tag, comm, &status);
@@
-129,8
+124,8
@@
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;
@@
-169,7
+164,7
@@
int smpi_coll_tuned_allreduce_rab_rdb(void *sbuff, void *rbuff, int count,
recv_cnt += cnts[i];
}
recv_cnt += cnts[i];
}
-
MPI_S
endrecv((char *) rbuff + disps[send_idx] * extent, send_cnt,
+
smpi_mpi_s
endrecv((char *) rbuff + disps[send_idx] * extent, send_cnt,
dtype, dst, tag,
(char *) rbuff + disps[recv_idx] * extent, recv_cnt,
dtype, dst, tag, comm, &status);
dtype, dst, tag,
(char *) rbuff + disps[recv_idx] * extent, recv_cnt,
dtype, dst, tag, comm, &status);
@@
-190,9
+185,9
@@
int smpi_coll_tuned_allreduce_rab_rdb(void *sbuff, void *rbuff, int count,
if (rank < 2 * rem) {
if (rank % 2) // odd
if (rank < 2 * rem) {
if (rank % 2) // odd
-
MPI_S
end(rbuff, count, dtype, rank - 1, tag, comm);
+
smpi_mpi_s
end(rbuff, count, dtype, rank - 1, tag, comm);
else // even
else // even
-
MPI_R
ecv(rbuff, count, dtype, rank + 1, tag, comm, &status);
+
smpi_mpi_r
ecv(rbuff, count, dtype, rank + 1, tag, comm, &status);
}
free(tmp_buf);
}
free(tmp_buf);