1 #include "colls_private.h"
3 /*****************************************************************************
5 Copyright (c) 2006, Ahmad Faraj & Xin Yuan,
8 Redistribution and use in source and binary forms, with or without
9 modification, are permitted provided that the following conditions are met:
11 * Redistributions of source code must retain the above copyright notice,
12 this list of conditions and the following disclaimer.
14 * Redistributions in binary form must reproduce the above copyright notice,
15 this list of conditions and the following disclaimer in the documentation
16 and/or other materials provided with the distribution.
18 * Neither the name of the Florida State University nor the names of its
19 contributors may be used to endorse or promote products derived from this
20 software without specific prior written permission.
22 THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
23 ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
24 WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
25 DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
26 ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
27 (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
28 LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON
29 ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
30 (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
31 SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
33 *************************************************************************
34 * Any results obtained from executing this software require the *
35 * acknowledgment and citation of the software and its owners. *
36 * The full citation is given below: *
38 * A. Faraj and X. Yuan. "Automatic Generation and Tuning of MPI *
39 * Collective Communication Routines." The 19th ACM International *
40 * Conference on Supercomputing (ICS), Cambridge, Massachusetts, *
42 *************************************************************************
44 *****************************************************************************/
46 /*****************************************************************************
49 * num: the number of processors in a communicator
53 * descp: takes a number and tries to find a factoring of x*y*z mesh out of it
54 ****************************************************************************/
57 static int is_3dmesh(int num, int *i, int *j, int *k)
63 if ((num % (x * x)) == 0) {
73 /*****************************************************************************
74 * Function: allgather_3dmesh_shoot
76 * send_buff: send input buffer
77 * send_count: number of elements to send
78 * send_type: data type of elements being sent
79 * recv_buff: receive output buffer
80 * recv_count: number of elements to received
81 * recv_type: data type of elements being received
83 * Descrp: Function realizes the allgather operation using the 2dmesh
84 * algorithm. Allgather ommunication occurs first in the x dimension, y
85 * dimension, and then in the z dimension. Communication in each dimension
88 ****************************************************************************/
89 int smpi_coll_tuned_allgather_3dmesh(void *send_buff, int send_count,
90 MPI_Datatype send_type, void *recv_buff,
91 int recv_count, MPI_Datatype recv_type,
94 MPI_Request *req, *req_ptr;
97 int i, src, dst, rank, num_procs, block_size, my_z_base;
98 int my_z, X, Y, Z, send_offset, recv_offset;
99 int two_dsize, my_row_base, my_col_base, src_row_base, src_z_base, num_reqs;
100 int tag = COLL_TAG_ALLGATHER;
102 rank = smpi_comm_rank(comm);
103 num_procs = smpi_comm_size(comm);
104 extent = smpi_datatype_get_extent(send_type);
106 is_3dmesh(num_procs, &X, &Y, &Z);
116 my_z = rank / two_dsize;
118 my_row_base = (rank / X) * X;
119 my_col_base = (rank % Y) + (my_z * two_dsize);
120 my_z_base = my_z * two_dsize;
122 block_size = extent * send_count;
124 req = (MPI_Request *) xbt_malloc(num_reqs * sizeof(MPI_Request));
128 // do local allgather/local copy
129 recv_offset = rank * block_size;
130 smpi_datatype_copy(send_buff, send_count, send_type, (char *)recv_buff + recv_offset,
131 recv_count, recv_type);
134 for (i = 0; i < Y; i++) {
135 src = i + my_row_base;
138 recv_offset = src * block_size;
139 *(req_ptr++) = smpi_mpi_irecv((char *)recv_buff + recv_offset, send_count, recv_type, src, tag,
143 for (i = 0; i < Y; i++) {
144 dst = i + my_row_base;
147 smpi_mpi_send(send_buff, send_count, send_type, dst, tag, comm);
150 smpi_mpi_waitall(Y - 1, req, MPI_STATUSES_IGNORE);
153 // do colwise comm, it does not matter here if i*X or i *Y since X == Y
155 for (i = 0; i < X; i++) {
156 src = (i * Y + my_col_base);
160 src_row_base = (src / X) * X;
161 recv_offset = src_row_base * block_size;
162 *(req_ptr++) = smpi_mpi_irecv((char *)recv_buff + recv_offset, recv_count * Y, recv_type, src, tag,
166 send_offset = my_row_base * block_size;
168 for (i = 0; i < X; i++) {
169 dst = (i * Y + my_col_base);
172 smpi_mpi_send((char *)recv_buff + send_offset, send_count * Y, send_type, dst, tag,
176 smpi_mpi_waitall(X - 1, req, MPI_STATUSES_IGNORE);
179 for (i = 1; i < Z; i++) {
180 src = (rank + i * two_dsize) % num_procs;
181 src_z_base = (src / two_dsize) * two_dsize;
183 recv_offset = (src_z_base * block_size);
185 *(req_ptr++) = smpi_mpi_irecv((char *)recv_buff + recv_offset, recv_count * two_dsize, recv_type,
189 for (i = 1; i < Z; i++) {
190 dst = (rank + i * two_dsize) % num_procs;
191 send_offset = my_z_base * block_size;
192 smpi_mpi_send((char *)recv_buff + send_offset, send_count * two_dsize, send_type,
195 smpi_mpi_waitall(Z - 1, req, MPI_STATUSES_IGNORE);