+/* 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>
{
MPI_Status status;
MPI_Aint extent;
- int tag = 4321, rank, nprocs, send_size, newcnt, share;
- int pof2 = 1, mask, send_idx, recv_idx, dst, send_cnt, recv_cnt;
+ int tag = COLL_TAG_ALLREDUCE, send_size, newcnt, share;
+ unsigned int pof2 = 1, mask;
+ int send_idx, recv_idx, dst, send_cnt, recv_cnt;
void *recv, *tmp_buf;
- rank = smpi_comm_rank(comm);
- nprocs = smpi_comm_size(comm);
+ unsigned int rank = smpi_comm_rank(comm);
+ unsigned int nprocs = smpi_comm_size(comm);
+
+ if((nprocs&(nprocs-1)))
+ THROWF(arg_error,0, "allreduce rab1 algorithm can't be used with non power of two number of processes ! ");
extent = smpi_datatype_get_extent(dtype);
pof2 <<= 1;
pof2 >>= 1;
- mask = 1;
send_idx = recv_idx = 0;
// uneven count
send_size = (count + nprocs) / nprocs;
newcnt = send_size * nprocs;
- recv = (void *) xbt_malloc(extent * newcnt);
- tmp_buf = (void *) xbt_malloc(extent * newcnt);
+ recv = (void *) smpi_get_tmp_recvbuffer(extent * newcnt);
+ tmp_buf = (void *) smpi_get_tmp_sendbuffer(extent * newcnt);
memcpy(recv, sbuff, extent * count);
smpi_mpi_sendrecv((char *) recv + send_idx * extent, send_cnt, dtype, dst, tag,
tmp_buf, recv_cnt, dtype, dst, tag, comm, &status);
- star_reduction(op, tmp_buf, (char *) recv + recv_idx * extent, &recv_cnt,
+ smpi_op_apply(op, tmp_buf, (char *) recv + recv_idx * extent, &recv_cnt,
&dtype);
// update send_idx for next iteration
mpi_coll_allgather_fun(tmp_buf, recv_cnt, dtype, recv, recv_cnt, dtype, comm);
memcpy(rbuff, recv, count * extent);
- free(recv);
- free(tmp_buf);
+ smpi_free_tmp_buffer(recv);
+ smpi_free_tmp_buffer(tmp_buf);
}
else {
- tmp_buf = (void *) xbt_malloc(extent * count);
+ tmp_buf = (void *) smpi_get_tmp_sendbuffer(extent * count);
memcpy(rbuff, sbuff, count * extent);
mask = pof2 / 2;
share = count / pof2;
smpi_mpi_sendrecv((char *) rbuff + send_idx * extent, send_cnt, dtype, dst,
tag, tmp_buf, recv_cnt, dtype, dst, tag, comm, &status);
- star_reduction(op, tmp_buf, (char *) rbuff + recv_idx * extent, &recv_cnt,
+ smpi_op_apply(op, tmp_buf, (char *) rbuff + recv_idx * extent, &recv_cnt,
&dtype);
// update send_idx for next iteration
memcpy(tmp_buf, (char *) rbuff + recv_idx * extent, recv_cnt * extent);
mpi_coll_allgather_fun(tmp_buf, recv_cnt, dtype, rbuff, recv_cnt, dtype, comm);
- free(tmp_buf);
+ smpi_free_tmp_buffer(tmp_buf);
}
return MPI_SUCCESS;