-#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"
//#include <star-reduction.c>
int smpi_coll_tuned_allreduce_rdb(void *sbuff, void *rbuff, int count,
MPI_Datatype dtype, MPI_Op op, MPI_Comm comm)
{
- int nprocs, rank, type_size, tag = 543;
+ int nprocs, rank, tag = COLL_TAG_ALLREDUCE;
int mask, dst, pof2, newrank, rem, newdst;
- MPI_Aint extent;
+ MPI_Aint extent, lb;
MPI_Status status;
void *tmp_buf = NULL;
/*
uop = op_ptr->op;
#endif
*/
- MPI_Comm_size(comm, &nprocs);
- MPI_Comm_rank(comm, &rank);
-
- MPI_Type_extent(dtype, &extent);
- tmp_buf = (void *) malloc(count * extent);
- if (!tmp_buf) {
- printf("Could not allocate memory for tmp_buf\n");
- return 1;
- }
+ nprocs=smpi_comm_size(comm);
+ rank=smpi_comm_rank(comm);
- MPI_Sendrecv(sbuff, count, dtype, rank, 500,
- rbuff, count, dtype, rank, 500, comm, &status);
+ smpi_datatype_extent(dtype, &lb, &extent);
+ tmp_buf = (void *) smpi_get_tmp_sendbuffer(count * extent);
- MPI_Type_size(dtype, &type_size);
+ smpi_mpi_sendrecv(sbuff, count, dtype, rank, 500,
+ rbuff, count, dtype, rank, 500, comm, &status);
// find nearest power-of-two less than or equal to comm_size
pof2 = 1;
// even
if (rank % 2 == 0) {
- MPI_Send(rbuff, count, dtype, rank + 1, tag, comm);
+ smpi_mpi_send(rbuff, count, dtype, rank + 1, tag, comm);
// temporarily set the rank to -1 so that this
// process does not pariticipate in recursive
newrank = -1;
} else // odd
{
- MPI_Recv(tmp_buf, count, dtype, rank - 1, tag, comm, &status);
+ smpi_mpi_recv(tmp_buf, count, dtype, rank - 1, tag, comm, &status);
// do the reduction on received data. since the
// ordering is right, it doesn't matter whether
// the operation is commutative or not.
- star_reduction(op, tmp_buf, rbuff, &count, &dtype);
+ smpi_op_apply(op, tmp_buf, rbuff, &count, &dtype);
// change the rank
newrank = rank / 2;
// Send the most current data, which is in recvbuf. Recv
// into tmp_buf
- MPI_Sendrecv(rbuff, count, dtype, dst, tag, tmp_buf, count, dtype,
+ smpi_mpi_sendrecv(rbuff, count, dtype, dst, tag, tmp_buf, count, dtype,
dst, tag, comm, &status);
// tmp_buf contains data received in this step.
// we assume it is commuttive op
// if (op -> op_commute || (dst < rank))
if ((dst < rank)) {
- star_reduction(op, tmp_buf, rbuff, &count, &dtype);
+ smpi_op_apply(op, tmp_buf, rbuff, &count, &dtype);
} else // op is noncommutative and the order is not right
{
- star_reduction(op, rbuff, tmp_buf, &count, &dtype);
+ smpi_op_apply(op, rbuff, tmp_buf, &count, &dtype);
// copy result back into recvbuf
- MPI_Sendrecv(tmp_buf, count, dtype, rank, tag, rbuff, count,
+ smpi_mpi_sendrecv(tmp_buf, count, dtype, rank, tag, rbuff, count,
dtype, rank, tag, comm, &status);
}
mask <<= 1;
if (rank < 2 * rem) {
if (rank % 2) // odd
- MPI_Send(rbuff, count, dtype, rank - 1, tag, comm);
+ smpi_mpi_send(rbuff, count, dtype, rank - 1, tag, comm);
else // even
- MPI_Recv(rbuff, count, dtype, rank + 1, tag, comm, &status);
+ smpi_mpi_recv(rbuff, count, dtype, rank + 1, tag, comm, &status);
}
- free(tmp_buf);
- return 0;
+ smpi_free_tmp_buffer(tmp_buf);
+ return MPI_SUCCESS;
}