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
die blank lines and dummy comments, die
[simgrid.git]
/
src
/
smpi
/
colls
/
allreduce-smp-rdb.c
diff --git
a/src/smpi/colls/allreduce-smp-rdb.c
b/src/smpi/colls/allreduce-smp-rdb.c
index
2bdfacd
..
9770d29
100644
(file)
--- a/
src/smpi/colls/allreduce-smp-rdb.c
+++ b/
src/smpi/colls/allreduce-smp-rdb.c
@@
-1,15
+1,16
@@
-#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"
/* IMPLEMENTED BY PITCH PATARASUK
Non-topoloty-specific (however, number of cores/node need to be changed)
all-reduce operation designed for smp clusters
It uses 2-layer communication: binomial for intra-communication
and rdb for inter-communication*/
/* IMPLEMENTED BY PITCH PATARASUK
Non-topoloty-specific (however, number of cores/node need to be changed)
all-reduce operation designed for smp clusters
It uses 2-layer communication: binomial for intra-communication
and rdb for inter-communication*/
-/* 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
/* ** NOTE **
Use -DMPICH2 if this code does not compile.
/* ** NOTE **
Use -DMPICH2 if this code does not compile.
@@
-31,10
+32,16
@@
int smpi_coll_tuned_allreduce_smp_rdb(void *send_buf, void *recv_buf, 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));
+ }
/*
#ifdef MPICH2_REDUCTION
MPI_User_function * uop = MPIR_Op_table[op % 16 - 1];
/*
#ifdef MPICH2_REDUCTION
MPI_User_function * uop = MPIR_Op_table[op % 16 - 1];
@@
-45,11
+52,11
@@
int smpi_coll_tuned_allreduce_smp_rdb(void *send_buf, void *recv_buf, int count,
uop = op_ptr->op;
#endif
*/
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_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);
/* compute intra and inter ranking */
int intra_rank, inter_rank;
/* compute intra and inter ranking */
int intra_rank, inter_rank;
@@
-61,7
+68,7
@@
int smpi_coll_tuned_allreduce_smp_rdb(void *send_buf, void *recv_buf, int count,
int inter_comm_size = (comm_size + num_core - 1) / num_core;
/* copy input buffer to output buffer */
int inter_comm_size = (comm_size + num_core - 1) / num_core;
/* copy input buffer to output buffer */
-
MPI_S
endrecv(send_buf, count, dtype, rank, tag,
+
smpi_mpi_s
endrecv(send_buf, count, dtype, rank, tag,
recv_buf, count, dtype, rank, tag, comm, &status);
/* start binomial reduce intra communication inside each SMP node */
recv_buf, count, dtype, rank, tag, comm, &status);
/* start binomial reduce intra communication inside each SMP node */
@@
-70,12
+77,12
@@
int smpi_coll_tuned_allreduce_smp_rdb(void *send_buf, void *recv_buf, int count,
if ((mask & intra_rank) == 0) {
src = (inter_rank * num_core) + (intra_rank | mask);
if (src < comm_size) {
if ((mask & intra_rank) == 0) {
src = (inter_rank * num_core) + (intra_rank | mask);
if (src < comm_size) {
-
MPI_R
ecv(tmp_buf, count, dtype, src, tag, comm, &status);
- s
tar_reduction
(op, tmp_buf, recv_buf, &count, &dtype);
+
smpi_mpi_r
ecv(tmp_buf, count, dtype, src, tag, comm, &status);
+ s
mpi_op_apply
(op, tmp_buf, recv_buf, &count, &dtype);
}
} else {
dst = (inter_rank * num_core) + (intra_rank & (~mask));
}
} else {
dst = (inter_rank * num_core) + (intra_rank & (~mask));
-
MPI_S
end(recv_buf, count, dtype, dst, tag, comm);
+
smpi_mpi_s
end(recv_buf, count, dtype, dst, tag, comm);
break;
}
mask <<= 1;
break;
}
mask <<= 1;
@@
-103,12
+110,12
@@
int smpi_coll_tuned_allreduce_smp_rdb(void *send_buf, void *recv_buf, int count,
if (inter_rank < 2 * rem) {
if (inter_rank % 2 == 0) {
dst = rank + num_core;
if (inter_rank < 2 * rem) {
if (inter_rank % 2 == 0) {
dst = rank + num_core;
-
MPI_S
end(recv_buf, count, dtype, dst, tag, comm);
+
smpi_mpi_s
end(recv_buf, count, dtype, dst, tag, comm);
newrank = -1;
} else {
src = rank - num_core;
newrank = -1;
} else {
src = rank - num_core;
-
MPI_R
ecv(tmp_buf, count, dtype, src, tag, comm, &status);
- s
tar_reduction
(op, tmp_buf, recv_buf, &count, &dtype);
+
smpi_mpi_r
ecv(tmp_buf, count, dtype, src, tag, comm, &status);
+ s
mpi_op_apply
(op, tmp_buf, recv_buf, &count, &dtype);
newrank = inter_rank / 2;
}
} else {
newrank = inter_rank / 2;
}
} else {
@@
-132,9
+139,9
@@
int smpi_coll_tuned_allreduce_smp_rdb(void *send_buf, void *recv_buf, int count,
dst *= num_core;
/* exchange data in rdb manner */
dst *= num_core;
/* exchange data in rdb manner */
-
MPI_S
endrecv(recv_buf, count, dtype, dst, tag, tmp_buf, count, dtype,
+
smpi_mpi_s
endrecv(recv_buf, count, dtype, dst, tag, tmp_buf, count, dtype,
dst, tag, comm, &status);
dst, tag, comm, &status);
- s
tar_reduction
(op, tmp_buf, recv_buf, &count, &dtype);
+ s
mpi_op_apply
(op, tmp_buf, recv_buf, &count, &dtype);
mask <<= 1;
}
}
mask <<= 1;
}
}
@@
-144,9
+151,9
@@
int smpi_coll_tuned_allreduce_smp_rdb(void *send_buf, void *recv_buf, int count,
*/
if (inter_rank < 2 * rem) {
if (inter_rank % 2) {
*/
if (inter_rank < 2 * rem) {
if (inter_rank % 2) {
-
MPI_S
end(recv_buf, count, dtype, rank - num_core, tag, comm);
+
smpi_mpi_s
end(recv_buf, count, dtype, rank - num_core, tag, comm);
} else {
} else {
-
MPI_R
ecv(recv_buf, count, dtype, rank + num_core, tag, comm, &status);
+
smpi_mpi_r
ecv(recv_buf, count, dtype, rank + num_core, tag, comm, &status);
}
}
}
}
}
}
@@
-160,7
+167,7
@@
int smpi_coll_tuned_allreduce_smp_rdb(void *send_buf, void *recv_buf, int count,
while (mask < num_core_in_current_smp) {
if (intra_rank & mask) {
src = (inter_rank * num_core) + (intra_rank - mask);
while (mask < num_core_in_current_smp) {
if (intra_rank & mask) {
src = (inter_rank * num_core) + (intra_rank - mask);
-
MPI_R
ecv(recv_buf, count, dtype, src, tag, comm, &status);
+
smpi_mpi_r
ecv(recv_buf, count, dtype, src, tag, comm, &status);
break;
}
mask <<= 1;
break;
}
mask <<= 1;
@@
-170,11
+177,11
@@
int smpi_coll_tuned_allreduce_smp_rdb(void *send_buf, void *recv_buf, int count,
while (mask > 0) {
dst = (inter_rank * num_core) + (intra_rank + mask);
if (dst < comm_size) {
while (mask > 0) {
dst = (inter_rank * num_core) + (intra_rank + mask);
if (dst < comm_size) {
-
MPI_S
end(recv_buf, count, dtype, dst, tag, comm);
+
smpi_mpi_s
end(recv_buf, count, dtype, dst, tag, comm);
}
mask >>= 1;
}
}
mask >>= 1;
}
-
free
(tmp_buf);
+
smpi_free_tmp_buffer
(tmp_buf);
return MPI_SUCCESS;
}
return MPI_SUCCESS;
}