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 /*****************************************************************************
53 num: the number of processors in a communicator
57 * Descp: takes a number and tries to find a factoring of x, y mesh out of it
60 ****************************************************************************/
63 static int is_2dmesh(int num, int *i, int *j)
86 /*****************************************************************************
87 * Function: allgather_2dmesh_shoot
89 * send_buff: send input buffer
90 * send_count: number of elements to send
91 * send_type: data type of elements being sent
92 * recv_buff: receive output buffer
93 * recv_count: number of elements to received
94 * recv_type: data type of elements being received
96 * Descrp: Function realizes the allgather operation using the 2dmesh
97 * algorithm. Allgather ommunication occurs first in the x dimension then in
98 * the y dimension. The communication in each dimension follows
100 * Auther: Ahmad Faraj
101 ****************************************************************************/
103 smpi_coll_tuned_allgather_2dmesh(void *send_buff, int send_count, MPI_Datatype
104 send_type, void *recv_buff, int recv_count,
105 MPI_Datatype recv_type, MPI_Comm comm)
107 MPI_Request *req, *req_ptr;
110 int i, src, dst, rank, num_procs;
111 int X, Y, send_offset, recv_offset;
112 int my_row_base, my_col_base, src_row_base, block_size, num_reqs;
115 rank = smpi_comm_rank(comm);
116 num_procs = smpi_comm_size(comm);
118 extent = smpi_datatype_get_extent(send_type);
120 block_size = extent * send_count;
122 if (!is_2dmesh(num_procs, &X, &Y))
125 my_row_base = (rank / Y) * Y;
126 my_col_base = rank % Y;
132 req = (MPI_Request *) xbt_malloc(num_reqs * sizeof(MPI_Request));
136 // do local allgather/local copy
137 recv_offset = rank * block_size;
138 smpi_datatype_copy(send_buff, send_count, send_type, (char *)recv_buff + recv_offset,
139 recv_count, recv_type);
142 for (i = 0; i < Y; i++) {
143 src = i + my_row_base;
146 recv_offset = src * block_size;
147 *(req_ptr++) = smpi_mpi_irecv((char *)recv_buff + recv_offset, recv_count, recv_type, src, tag,
152 for (i = 0; i < Y; i++) {
153 dst = i + my_row_base;
156 smpi_mpi_send(send_buff, send_count, send_type, dst, tag, comm);
159 smpi_mpi_waitall(Y - 1, req, MPI_STATUSES_IGNORE);
164 for (i = 0; i < X; i++) {
165 src = (i * Y + my_col_base);
168 src_row_base = (src / Y) * Y;
169 recv_offset = src_row_base * block_size;
170 *(req_ptr++) = smpi_mpi_irecv((char *)recv_buff + recv_offset, recv_count * Y, recv_type, src, tag,
174 for (i = 0; i < X; i++) {
175 dst = (i * Y + my_col_base);
178 send_offset = my_row_base * block_size;
179 smpi_mpi_send((char *)recv_buff + send_offset, send_count * Y, send_type, dst, tag,
183 smpi_mpi_waitall(X - 1, req, MPI_STATUSES_IGNORE);