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"
10 /*****************************************************************************
12 * Function: alltoall_3dmesh_shoot
17 send_buff: send input buffer
18 send_count: number of elements to send
19 send_type: data type of elements being sent
20 recv_buff: receive output buffer
21 recv_count: number of elements to received
22 recv_type: data type of elements being received
25 * Descrp: Function realizes the alltoall operation using the 3dmesh
26 algorithm. It actually performs allgather operation in x dimension,
27 y dimension, then in z dimension. Each node then extracts the
28 needed data. The communication in all dimension is simple.
31 ****************************************************************************/
33 static int alltoall_check_is_3dmesh(int num, int *i, int *j, int *k)
39 if ((num % (x * x)) == 0) {
49 int smpi_coll_tuned_alltoall_3dmesh(void *send_buff, int send_count,
50 MPI_Datatype send_type,
51 void *recv_buff, int recv_count,
52 MPI_Datatype recv_type, MPI_Comm comm)
54 MPI_Request *reqs, *req_ptr;
56 MPI_Status status, *statuses;
57 int i, j, src, dst, rank, num_procs, num_reqs, X, Y, Z, block_size, count;
58 int my_z, two_dsize, my_row_base, my_col_base, my_z_base, src_row_base;
59 int src_z_base, send_offset, recv_offset, tag = COLL_TAG_ALLTOALL;
61 char *tmp_buff1, *tmp_buff2;
63 rank = smpi_comm_rank(comm);
64 num_procs = smpi_comm_size(comm);
65 extent = smpi_datatype_get_extent(send_type);
67 if (!alltoall_check_is_3dmesh(num_procs, &X, &Y, &Z))
77 my_z = rank / two_dsize;
79 my_row_base = (rank / X) * X;
80 my_col_base = (rank % Y) + (my_z * two_dsize);
81 my_z_base = my_z * two_dsize;
83 block_size = extent * send_count;
85 tmp_buff1 = (char *) xbt_malloc(block_size * num_procs * two_dsize);
86 tmp_buff2 = (char *) xbt_malloc(block_size * two_dsize);
88 statuses = (MPI_Status *) xbt_malloc(num_reqs * sizeof(MPI_Status));
89 reqs = (MPI_Request *) xbt_malloc(num_reqs * sizeof(MPI_Request));
93 send_offset = recv_offset = (rank % two_dsize) * block_size * num_procs;
95 smpi_mpi_sendrecv(send_buff, send_count * num_procs, send_type, rank, tag,
96 tmp_buff1 + recv_offset, num_procs * recv_count,
97 recv_type, rank, tag, comm, &status);
99 count = send_count * num_procs;
101 for (i = 0; i < Y; i++) {
102 src = i + my_row_base;
105 recv_offset = (src % two_dsize) * block_size * num_procs;
106 *(req_ptr++) = smpi_mpi_irecv(tmp_buff1 + recv_offset, count, recv_type, src, tag, comm);
109 for (i = 0; i < Y; i++) {
110 dst = i + my_row_base;
113 smpi_mpi_send(send_buff, count, send_type, dst, tag, comm);
116 smpi_mpi_waitall(Y - 1, reqs, statuses);
120 for (i = 0; i < X; i++) {
121 src = (i * Y + my_col_base);
125 src_row_base = (src / X) * X;
127 recv_offset = (src_row_base % two_dsize) * block_size * num_procs;
128 *(req_ptr++) = smpi_mpi_irecv(tmp_buff1 + recv_offset, recv_count * num_procs * Y,
129 recv_type, src, tag, comm);
132 send_offset = (my_row_base % two_dsize) * block_size * num_procs;
133 for (i = 0; i < X; i++) {
134 dst = (i * Y + my_col_base);
137 smpi_mpi_send(tmp_buff1 + send_offset, send_count * num_procs * Y, send_type,
141 smpi_mpi_waitall(X - 1, reqs, statuses);
144 for (i = 0; i < two_dsize; i++) {
145 send_offset = (rank * block_size) + (i * block_size * num_procs);
146 recv_offset = (my_z_base * block_size) + (i * block_size);
147 smpi_mpi_sendrecv(tmp_buff1 + send_offset, send_count, send_type, rank, tag,
148 (char *) recv_buff + recv_offset, recv_count, recv_type,
149 rank, tag, comm, &status);
152 for (i = 1; i < Z; i++) {
153 src = (rank + i * two_dsize) % num_procs;
154 src_z_base = (src / two_dsize) * two_dsize;
156 recv_offset = (src_z_base * block_size);
158 *(req_ptr++) = smpi_mpi_irecv((char *) recv_buff + recv_offset, recv_count * two_dsize,
159 recv_type, src, tag, comm);
162 for (i = 1; i < Z; i++) {
163 dst = (rank + i * two_dsize) % num_procs;
166 for (j = 0; j < two_dsize; j++) {
167 send_offset = (dst + j * num_procs) * block_size;
168 smpi_mpi_sendrecv(tmp_buff1 + send_offset, send_count, send_type,
169 rank, tag, tmp_buff2 + recv_offset, recv_count,
170 recv_type, rank, tag, comm, &status);
172 recv_offset += block_size;
175 smpi_mpi_send(tmp_buff2, send_count * two_dsize, send_type, dst, tag, comm);
179 smpi_mpi_waitall(Z - 1, reqs, statuses);