1 /* Copyright (c) 2013-2014. The SimGrid Team.
2 * All rights reserved. */
4 /* This program is free software; you can redistribute it and/or modify it
5 * under the terms of the license (GNU LGPL) which comes with this package. */
7 #include "colls_private.h"
8 //#include <star-reduction.c>
10 // NP pow of 2 for now
11 int smpi_coll_tuned_allreduce_rab1(void *sbuff, void *rbuff,
12 int count, MPI_Datatype dtype,
13 MPI_Op op, MPI_Comm comm)
17 int tag = COLL_TAG_ALLREDUCE, rank, nprocs, send_size, newcnt, share;
18 int pof2 = 1, mask, send_idx, recv_idx, dst, send_cnt, recv_cnt;
22 rank = smpi_comm_rank(comm);
23 nprocs = smpi_comm_size(comm);
25 if((nprocs&(nprocs-1)))
26 THROWF(arg_error,0, "allreduce rab1 algorithm can't be used with non power of two number of processes ! ");
28 extent = smpi_datatype_get_extent(dtype);
31 while (pof2 <= nprocs)
36 send_idx = recv_idx = 0;
39 if ((count % nprocs)) {
40 send_size = (count + nprocs) / nprocs;
41 newcnt = send_size * nprocs;
43 recv = (void *) xbt_malloc(extent * newcnt);
44 tmp_buf = (void *) xbt_malloc(extent * newcnt);
45 memcpy(recv, sbuff, extent * count);
49 share = newcnt / pof2;
52 send_cnt = recv_cnt = newcnt / (pof2 / mask);
55 send_idx = recv_idx + (mask * share);
57 recv_idx = send_idx + (mask * share);
59 smpi_mpi_sendrecv((char *) recv + send_idx * extent, send_cnt, dtype, dst, tag,
60 tmp_buf, recv_cnt, dtype, dst, tag, comm, &status);
62 smpi_op_apply(op, tmp_buf, (char *) recv + recv_idx * extent, &recv_cnt,
65 // update send_idx for next iteration
70 memcpy(tmp_buf, (char *) recv + recv_idx * extent, recv_cnt * extent);
71 mpi_coll_allgather_fun(tmp_buf, recv_cnt, dtype, recv, recv_cnt, dtype, comm);
73 memcpy(rbuff, recv, count * extent);
80 tmp_buf = (void *) xbt_malloc(extent * count);
81 memcpy(rbuff, sbuff, count * extent);
86 send_cnt = recv_cnt = count / (pof2 / mask);
89 send_idx = recv_idx + (mask * share);
91 recv_idx = send_idx + (mask * share);
93 smpi_mpi_sendrecv((char *) rbuff + send_idx * extent, send_cnt, dtype, dst,
94 tag, tmp_buf, recv_cnt, dtype, dst, tag, comm, &status);
96 smpi_op_apply(op, tmp_buf, (char *) rbuff + recv_idx * extent, &recv_cnt,
99 // update send_idx for next iteration
104 memcpy(tmp_buf, (char *) rbuff + recv_idx * extent, recv_cnt * extent);
105 mpi_coll_allgather_fun(tmp_buf, recv_cnt, dtype, rbuff, recv_cnt, dtype, comm);