+/* 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"
-#ifndef NUM_CORE
-#define NUM_CORE 8
-#endif
int bcast_SMP_linear_segment_byte = 8192;
MPI_Datatype datatype, int root,
MPI_Comm comm)
{
- int tag = 5000;
+ int tag = COLL_TAG_BCAST;
MPI_Status status;
MPI_Request request;
MPI_Request *request_array;
rank = smpi_comm_rank(comm);
size = smpi_comm_size(comm);
+ 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));
+ }else{
+ //implementation buggy in this case
+ return smpi_coll_tuned_bcast_mpich( buf , count, datatype,
+ root, comm);
+ }
int segment = bcast_SMP_linear_segment_byte / extent;
+ segment = segment == 0 ? 1 :segment;
int pipe_length = count / segment;
int remainder = count % segment;
int increment = segment * extent;
/* leader of each SMP do inter-communication
and act as a root for intra-communication */
- int to_inter = (rank + NUM_CORE) % size;
+ int to_inter = (rank + num_core) % size;
int to_intra = (rank + 1) % size;
- int from_inter = (rank - NUM_CORE + size) % size;
+ int from_inter = (rank - num_core + size) % size;
int from_intra = (rank + size - 1) % size;
// call native when MPI communication size is too small
- if (size <= NUM_CORE) {
+ if (size <= num_core) {
XBT_WARN("MPI_bcast_SMP_linear use default MPI_bcast.");
smpi_mpi_bcast(buf, count, datatype, root, comm);
return MPI_SUCCESS;
smpi_mpi_send(buf, count, datatype, to_intra, tag, comm);
}
// case last ROOT of each SMP
- else if (rank == (((size - 1) / NUM_CORE) * NUM_CORE)) {
+ else if (rank == (((size - 1) / num_core) * num_core)) {
request = smpi_mpi_irecv(buf, count, datatype, from_inter, tag, comm);
smpi_mpi_wait(&request, &status);
smpi_mpi_send(buf, count, datatype, to_intra, tag, comm);
}
// case intermediate ROOT of each SMP
- else if (rank % NUM_CORE == 0) {
+ else if (rank % num_core == 0) {
request = smpi_mpi_irecv(buf, count, datatype, from_inter, tag, comm);
smpi_mpi_wait(&request, &status);
smpi_mpi_send(buf, count, datatype, to_inter, tag, comm);
smpi_mpi_send(buf, count, datatype, to_intra, tag, comm);
}
// case last non-ROOT of each SMP
- else if (((rank + 1) % NUM_CORE == 0) || (rank == (size - 1))) {
+ else if (((rank + 1) % num_core == 0) || (rank == (size - 1))) {
request = smpi_mpi_irecv(buf, count, datatype, from_intra, tag, comm);
smpi_mpi_wait(&request, &status);
}
(MPI_Status *) xbt_malloc((size + pipe_length) * sizeof(MPI_Status));
// case ROOT of each SMP
- if (rank % NUM_CORE == 0) {
+ if (rank % num_core == 0) {
// case real root
if (rank == 0) {
for (i = 0; i < pipe_length; i++) {
}
}
// case last ROOT of each SMP
- else if (rank == (((size - 1) / NUM_CORE) * NUM_CORE)) {
+ else if (rank == (((size - 1) / num_core) * num_core)) {
for (i = 0; i < pipe_length; i++) {
request_array[i] = smpi_mpi_irecv((char *) buf + (i * increment), segment, datatype,
from_inter, (tag + i), comm);
}
}
} else { // case last non-ROOT of each SMP
- if (((rank + 1) % NUM_CORE == 0) || (rank == (size - 1))) {
+ if (((rank + 1) % num_core == 0) || (rank == (size - 1))) {
for (i = 0; i < pipe_length; i++) {
request_array[i] = smpi_mpi_irecv((char *) buf + (i * increment), segment, datatype,
from_intra, (tag + i), comm);