-#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
{
int comm_size, rank;
void *tmp_buf;
- int tag = 50;
+ int tag = COLL_TAG_ALLREDUCE;
int mask, src, dst;
- 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;
MPI_Status status;
- /*
- #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 = MPIR_ToPointer(op);
- uop = op_ptr->op;
- #endif
- */
-
- MPI_Comm_size(comm, &comm_size);
- MPI_Comm_rank(comm, &rank);
- MPI_Aint extent;
- MPI_Type_extent(dtype, &extent);
- tmp_buf = (void *) malloc(count * extent);
+
+ comm_size=smpi_comm_size(comm);
+ rank=smpi_comm_rank(comm);
+ MPI_Aint extent, lb;
+ smpi_datatype_extent(dtype, &lb, &extent);
+ tmp_buf = (void *) xbt_malloc(count * extent);
/* compute intra and inter ranking */
int intra_rank, inter_rank;
int inter_comm_size = (comm_size + num_core - 1) / num_core;
/* copy input buffer to output buffer */
- MPI_Sendrecv(send_buf, count, dtype, rank, tag,
+ smpi_mpi_sendrecv(send_buf, count, dtype, rank, tag,
recv_buf, count, dtype, rank, tag, comm, &status);
/* start binomial reduce intra communication inside each SMP node */
if ((mask & intra_rank) == 0) {
src = (inter_rank * num_core) + (intra_rank | mask);
if (src < comm_size) {
- MPI_Recv(tmp_buf, count, dtype, src, tag, comm, &status);
- star_reduction(op, tmp_buf, recv_buf, &count, &dtype);
+ smpi_mpi_recv(tmp_buf, count, dtype, src, tag, comm, &status);
+ smpi_op_apply(op, tmp_buf, recv_buf, &count, &dtype);
}
} else {
dst = (inter_rank * num_core) + (intra_rank & (~mask));
- MPI_Send(recv_buf, count, dtype, dst, tag, comm);
+ smpi_mpi_send(recv_buf, count, dtype, dst, tag, comm);
break;
}
mask <<= 1;
if ((mask & inter_rank) == 0) {
src = (inter_rank | mask) * num_core;
if (src < comm_size) {
- MPI_Recv(tmp_buf, count, dtype, src, tag, comm, &status);
- star_reduction(op, tmp_buf, recv_buf, &count, &dtype);
+ smpi_mpi_recv(tmp_buf, count, dtype, src, tag, comm, &status);
+ smpi_op_apply(op, tmp_buf, recv_buf, &count, &dtype);
}
} else {
dst = (inter_rank & (~mask)) * num_core;
- MPI_Send(recv_buf, count, dtype, dst, tag, comm);
+ smpi_mpi_send(recv_buf, count, dtype, dst, tag, comm);
break;
}
mask <<= 1;
while (mask < inter_comm_size) {
if (inter_rank & mask) {
src = (inter_rank - mask) * num_core;
- MPI_Recv(recv_buf, count, dtype, src, tag, comm, &status);
+ smpi_mpi_recv(recv_buf, count, dtype, src, tag, comm, &status);
break;
}
mask <<= 1;
if (inter_rank < inter_comm_size) {
dst = (inter_rank + mask) * num_core;
if (dst < comm_size) {
- MPI_Send(recv_buf, count, dtype, dst, tag, comm);
+ smpi_mpi_send(recv_buf, count, dtype, dst, tag, comm);
}
}
mask >>= 1;
while (mask < num_core_in_current_smp) {
if (intra_rank & mask) {
src = (inter_rank * num_core) + (intra_rank - mask);
- MPI_Recv(recv_buf, count, dtype, src, tag, comm, &status);
+ smpi_mpi_recv(recv_buf, count, dtype, src, tag, comm, &status);
break;
}
mask <<= 1;
while (mask > 0) {
dst = (inter_rank * num_core) + (intra_rank + mask);
if (dst < comm_size) {
- MPI_Send(recv_buf, count, dtype, dst, tag, comm);
+ smpi_mpi_send(recv_buf, count, dtype, dst, tag, comm);
}
mask >>= 1;
}