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
handle MPI_IN_PLACE for this collective
[simgrid.git]
/
src
/
smpi
/
colls
/
allreduce-smp-rsag-rab.c
diff --git
a/src/smpi/colls/allreduce-smp-rsag-rab.c
b/src/smpi/colls/allreduce-smp-rsag-rab.c
index
8ccb479
..
1ac827e
100644
(file)
--- a/
src/smpi/colls/allreduce-smp-rsag-rab.c
+++ b/
src/smpi/colls/allreduce-smp-rsag-rab.c
@@
-1,14
+1,15
@@
-#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. */
+
/*
* implemented by Pitch Patarasuk, 07/01/2007
*/
/*
* implemented by Pitch Patarasuk, 07/01/2007
*/
+#include "colls_private.h"
//#include <star-reduction.c>
//#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.
/*
This fucntion performs all-reduce operation as follow.
@@
-23,16
+24,26
@@
int smpi_coll_tuned_allreduce_smp_rsag_rab(void *sbuf, void *rbuf, int count,
{
int comm_size, rank;
void *tmp_buf;
{
int comm_size, rank;
void *tmp_buf;
- int tag =
50
;
+ int tag =
COLL_TAG_ALLREDUCE
;
int mask, src, dst;
MPI_Status status;
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));
+ }
- MPI_Comm_size(comm, &comm_size);
- MPI_Comm_rank(comm, &rank);
+ comm_size = smpi_comm_size(comm);
+
+ if((comm_size&(comm_size-1)))
+ THROWF(arg_error,0, "allreduce smp rsag rab algorithm can't be used with non power of two number of processes ! ");
+
+ rank = smpi_comm_rank(comm);
MPI_Aint extent;
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;
int intra_rank, inter_rank;
intra_rank = rank % num_core;
@@
-42,7
+53,7
@@
int smpi_coll_tuned_allreduce_smp_rsag_rab(void *sbuf, void *rbuf, int count,
int inter_comm_size = (comm_size + num_core - 1) / num_core;
int inter_comm_size = (comm_size + num_core - 1) / num_core;
-
MPI_S
endrecv(sbuf, count, dtype, rank, tag,
+
smpi_mpi_s
endrecv(sbuf, count, dtype, rank, tag,
rbuf, count, dtype, rank, tag, comm, &status);
// SMP_binomial_reduce
rbuf, count, dtype, rank, tag, comm, &status);
// SMP_binomial_reduce
@@
-52,14
+63,14
@@
int smpi_coll_tuned_allreduce_smp_rsag_rab(void *sbuf, void *rbuf, int count,
src = (inter_rank * num_core) + (intra_rank | mask);
// if (src < ((inter_rank + 1) * num_core)) {
if (src < comm_size) {
src = (inter_rank * num_core) + (intra_rank | mask);
// if (src < ((inter_rank + 1) * num_core)) {
if (src < comm_size) {
-
MPI_R
ecv(tmp_buf, count, dtype, src, tag, comm, &status);
- s
tar_reduction
(op, tmp_buf, rbuf, &count, &dtype);
+
smpi_mpi_r
ecv(tmp_buf, count, dtype, src, tag, comm, &status);
+ s
mpi_op_apply
(op, tmp_buf, rbuf, &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));
//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_S
end(rbuf, count, dtype, dst, tag, comm);
+
smpi_mpi_s
end(rbuf, count, dtype, dst, tag, comm);
//printf("Node %d send to node %d when mask is %d\n", rank, dst, mask);
break;
}
//printf("Node %d send to node %d when mask is %d\n", rank, dst, mask);
break;
}
@@
-78,19
+89,14
@@
int smpi_coll_tuned_allreduce_smp_rsag_rab(void *sbuf, void *rbuf, int count,
recv_chunk = extent * count / (comm_size / num_core);
mask = 1;
recv_chunk = extent * count / (comm_size / num_core);
mask = 1;
- i = 0;
curr_count = count / 2;
int phase = 0;
base_offset = 0;
curr_count = count / 2;
int phase = 0;
base_offset = 0;
- send_base_offset = 0;
- recv_base_offset = 0;
while (mask < (comm_size / num_core)) {
dst = inter_rank ^ mask;
// compute offsets
while (mask < (comm_size / num_core)) {
dst = inter_rank ^ mask;
// compute offsets
- send_base_offset = base_offset;
-
// right-handside
if (inter_rank & mask) {
recv_base_offset = base_offset + curr_count;
// right-handside
if (inter_rank & mask) {
recv_base_offset = base_offset + curr_count;
@@
-108,11
+114,11
@@
int smpi_coll_tuned_allreduce_smp_rsag_rab(void *sbuf, void *rbuf, int count,
// if (rank==7)
// printf("node %d send to %d in phase %d s_offset = %d r_offset = %d count = %d\n",rank,dst,phase, send_offset, recv_offset, curr_count);
// if (rank==7)
// printf("node %d send to %d in phase %d s_offset = %d r_offset = %d count = %d\n",rank,dst,phase, send_offset, recv_offset, curr_count);
-
MPI_S
endrecv((char *)rbuf + send_offset, curr_count, dtype, (dst * num_core), tag,
+
smpi_mpi_s
endrecv((char *)rbuf + send_offset, curr_count, dtype, (dst * num_core), tag,
tmp_buf, curr_count, dtype, (dst * num_core), tag,
comm, &status);
tmp_buf, curr_count, dtype, (dst * num_core), tag,
comm, &status);
- s
tar_reduction
(op, tmp_buf, (char *)rbuf + recv_offset, &curr_count, &dtype);
+ s
mpi_op_apply
(op, tmp_buf, (char *)rbuf + recv_offset, &curr_count, &dtype);
mask *= 2;
curr_count /= 2;
mask *= 2;
curr_count /= 2;
@@
-155,7
+161,7
@@
int smpi_coll_tuned_allreduce_smp_rsag_rab(void *sbuf, void *rbuf, int count,
// if (rank==7)
//printf("node %d send to %d in phase %d s_offset = %d r_offset = %d count = %d\n",rank,dst,phase, send_offset, recv_offset, curr_count);
// if (rank==7)
//printf("node %d send to %d in phase %d s_offset = %d r_offset = %d count = %d\n",rank,dst,phase, send_offset, recv_offset, curr_count);
-
MPI_S
endrecv((char *)rbuf + send_offset, curr_count, dtype, (dst * num_core), tag,
+
smpi_mpi_s
endrecv((char *)rbuf + send_offset, curr_count, dtype, (dst * num_core), tag,
(char *)rbuf + recv_offset, curr_count, dtype, (dst * num_core), tag,
comm, &status);
(char *)rbuf + recv_offset, curr_count, dtype, (dst * num_core), tag,
comm, &status);
@@
-181,7
+187,7
@@
int smpi_coll_tuned_allreduce_smp_rsag_rab(void *sbuf, void *rbuf, int count,
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);
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_R
ecv(rbuf, count, dtype, src, tag, comm, &status);
+
smpi_mpi_r
ecv(rbuf, count, dtype, src, tag, comm, &status);
break;
}
mask <<= 1;
break;
}
mask <<= 1;
@@
-194,12
+200,12
@@
int smpi_coll_tuned_allreduce_smp_rsag_rab(void *sbuf, void *rbuf, int count,
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);
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_S
end(rbuf, count, dtype, dst, tag, comm);
+
smpi_mpi_s
end(rbuf, count, dtype, dst, tag, comm);
}
mask >>= 1;
}
}
mask >>= 1;
}
-
free
(tmp_buf);
+
smpi_free_tmp_buffer
(tmp_buf);
return MPI_SUCCESS;
}
return MPI_SUCCESS;
}