-#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
*/
+#include "colls_private.h"
//#include <star-reduction.c>
/* change number of core per smp-node
{
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;
+ int num_core = simcall_host_get_core(SIMIX_host_self());
+ // do we use the default one or the number of cores in the platform ?
+ // if the number of cores is one, the platform may be simulated with 1 node = 1 core
+ if (num_core == 1) num_core = NUM_CORE;
+
+ comm_size = smpi_comm_size(comm);
- MPI_Comm_size(comm, &comm_size);
- MPI_Comm_rank(comm, &rank);
+ 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_Type_extent(dtype, &extent);
- tmp_buf = (void *) malloc(count * extent);
+ extent = smpi_datatype_get_extent(dtype);
+ tmp_buf = (void *) xbt_malloc(count * extent);
int intra_rank, inter_rank;
intra_rank = rank % num_core;
int inter_comm_size = (comm_size + num_core - 1) / num_core;
- MPI_Sendrecv(sbuf, count, dtype, rank, tag,
+ smpi_mpi_sendrecv(sbuf, count, dtype, rank, tag,
rbuf, count, dtype, rank, tag, comm, &status);
// SMP_binomial_reduce
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, rbuf, &count, &dtype);
+ smpi_mpi_recv(tmp_buf, count, dtype, src, tag, comm, &status);
+ smpi_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));
- MPI_Send(rbuf, count, dtype, dst, tag, comm);
+ smpi_mpi_send(rbuf, count, dtype, dst, tag, comm);
//printf("Node %d send to node %d when mask is %d\n", rank, dst, mask);
break;
}
recv_chunk = extent * count / (comm_size / num_core);
mask = 1;
- i = 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
- send_base_offset = base_offset;
-
// right-handside
if (inter_rank & mask) {
recv_base_offset = base_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_Sendrecv((char *)rbuf + send_offset, curr_count, dtype, (dst * num_core), tag,
+ smpi_mpi_sendrecv((char *)rbuf + send_offset, curr_count, dtype, (dst * num_core), tag,
tmp_buf, curr_count, dtype, (dst * num_core), tag,
comm, &status);
- star_reduction(op, tmp_buf, (char *)rbuf + recv_offset, &curr_count, &dtype);
+ smpi_op_apply(op, tmp_buf, (char *)rbuf + recv_offset, &curr_count, &dtype);
mask *= 2;
curr_count /= 2;
// 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_Sendrecv((char *)rbuf + send_offset, curr_count, dtype, (dst * num_core), tag,
+ smpi_mpi_sendrecv((char *)rbuf + send_offset, curr_count, dtype, (dst * num_core), tag,
(char *)rbuf + recv_offset, curr_count, dtype, (dst * num_core), tag,
comm, &status);
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(rbuf, count, dtype, src, tag, comm, &status);
+ smpi_mpi_recv(rbuf, count, dtype, src, tag, comm, &status);
break;
}
mask <<= 1;
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(rbuf, count, dtype, dst, tag, comm);
+ smpi_mpi_send(rbuf, count, dtype, dst, tag, comm);
}
mask >>= 1;
}