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. */
8 * Copyright (c) 2004-2005 The Trustees of Indiana University and Indiana
9 * University Research and Technology
10 * Corporation. All rights reserved.
11 * Copyright (c) 2004-2009 The University of Tennessee and The University
12 * of Tennessee Research Foundation. All rights
14 * Copyright (c) 2004-2005 High Performance Computing Center Stuttgart,
15 * University of Stuttgart. All rights reserved.
16 * Copyright (c) 2004-2005 The Regents of the University of California.
17 * All rights reserved.
19 * Additional copyrights may follow
21 /* -*- Mode: C; c-basic-offset:4 ; -*- */
22 /* Copyright (c) 2001-2014, The Ohio State University. All rights
25 * This file is part of the MVAPICH2 software package developed by the
26 * team members of The Ohio State University's Network-Based Computing
27 * Laboratory (NBCL), headed by Professor Dhabaleswar K. (DK) Panda.
29 * For detailed copyright and licensing information, please refer to the
30 * copyright file COPYRIGHT in the top level MVAPICH2 directory.
34 * (C) 2001 by Argonne National Laboratory.
35 * See COPYRIGHT in top-level directory.
38 #include "colls_private.h"
39 #include "src/smpi/smpi_group.hpp"
40 #define MV2_INTRA_SHMEM_REDUCE_MSG 2048
42 #define mv2_g_shmem_coll_max_msg_size (1 << 17)
43 #define SHMEM_COLL_BLOCK_SIZE (local_size * mv2_g_shmem_coll_max_msg_size)
44 #define mv2_use_knomial_reduce 1
46 #define MPIR_Reduce_inter_knomial_wrapper_MV2 smpi_coll_tuned_reduce_mvapich2_knomial
47 #define MPIR_Reduce_intra_knomial_wrapper_MV2 smpi_coll_tuned_reduce_mvapich2_knomial
48 #define MPIR_Reduce_binomial_MV2 smpi_coll_tuned_reduce_binomial
49 #define MPIR_Reduce_redscat_gather_MV2 smpi_coll_tuned_reduce_scatter_gather
50 #define MPIR_Reduce_shmem_MV2 smpi_coll_tuned_reduce_ompi_basic_linear
52 extern int (*MV2_Reduce_function)( void *sendbuf,
55 MPI_Datatype datatype,
60 extern int (*MV2_Reduce_intra_function)( void *sendbuf,
63 MPI_Datatype datatype,
69 /*Fn pointers for collectives */
70 static int (*reduce_fn)(void *sendbuf,
73 MPI_Datatype datatype,
74 MPI_Op op, int root, MPI_Comm comm);
76 int smpi_coll_tuned_reduce_mvapich2_two_level( void *sendbuf,
79 MPI_Datatype datatype,
84 int mpi_errno = MPI_SUCCESS;
85 int my_rank, total_size, local_rank, local_size;
86 int leader_comm_rank = -1, leader_comm_size = 0;
87 MPI_Comm shmem_comm, leader_comm;
88 int leader_root, leader_of_root;
89 void *in_buf = NULL, *out_buf = NULL, *tmp_buf = NULL;
90 MPI_Aint true_lb, true_extent, extent;
91 int is_commutative = 0, stride = 0;
92 int intra_node_root=0;
94 //if not set (use of the algo directly, without mvapich2 selector)
95 if(MV2_Reduce_function==NULL)
96 MV2_Reduce_function=smpi_coll_tuned_reduce_mpich;
97 if(MV2_Reduce_intra_function==NULL)
98 MV2_Reduce_intra_function=smpi_coll_tuned_reduce_mpich;
100 if(smpi_comm_get_leaders_comm(comm)==MPI_COMM_NULL){
101 smpi_comm_init_smp(comm);
104 my_rank = smpi_comm_rank(comm);
105 total_size = smpi_comm_size(comm);
106 shmem_comm = smpi_comm_get_intra_comm(comm);
107 local_rank = smpi_comm_rank(shmem_comm);
108 local_size = smpi_comm_size(shmem_comm);
110 leader_comm = smpi_comm_get_leaders_comm(comm);
111 int* leaders_map = smpi_comm_get_leaders_map(comm);
112 leader_of_root = smpi_comm_group(comm)->rank(leaders_map[root]);
113 leader_root = smpi_comm_group(leader_comm)->rank(leaders_map[root]);
115 is_commutative=smpi_op_is_commute(op);
117 smpi_datatype_extent(datatype, &true_lb,
119 extent =smpi_datatype_get_extent(datatype);
120 stride = count * MAX(extent, true_extent);
122 if (local_size == total_size) {
123 /* First handle the case where there is only one node */
124 if (stride <= MV2_INTRA_SHMEM_REDUCE_MSG &&
125 is_commutative == 1) {
126 if (local_rank == 0 ) {
127 tmp_buf=(void *)smpi_get_tmp_sendbuffer( count *
128 (MAX(extent, true_extent)));
129 tmp_buf = (void *) ((char *) tmp_buf - true_lb);
132 if (sendbuf != MPI_IN_PLACE) {
133 in_buf = (void *)sendbuf;
138 if (local_rank == 0) {
139 if( my_rank != root) {
143 if(in_buf == out_buf) {
144 in_buf = MPI_IN_PLACE;
149 in_buf = (void *)sendbuf;
153 if (count * (MAX(extent, true_extent)) < SHMEM_COLL_BLOCK_SIZE) {
154 mpi_errno = MPIR_Reduce_shmem_MV2(in_buf, out_buf, count,
159 mpi_errno = MPIR_Reduce_intra_knomial_wrapper_MV2(in_buf, out_buf, count,
164 if (local_rank == 0 && root != my_rank) {
165 smpi_mpi_send(out_buf, count, datatype, root,
166 COLL_TAG_REDUCE+1, comm);
168 if ((local_rank != 0) && (root == my_rank)) {
169 smpi_mpi_recv(recvbuf, count, datatype,
170 leader_of_root, COLL_TAG_REDUCE+1, comm,
174 if(mv2_use_knomial_reduce == 1) {
175 reduce_fn = &MPIR_Reduce_intra_knomial_wrapper_MV2;
177 reduce_fn = &MPIR_Reduce_binomial_MV2;
179 mpi_errno = reduce_fn(sendbuf, recvbuf, count,
185 smpi_free_tmp_buffer((void *) ((char *) tmp_buf + true_lb));
190 if (local_rank == 0) {
191 leader_comm = smpi_comm_get_leaders_comm(comm);
192 if(leader_comm==MPI_COMM_NULL){
193 leader_comm = MPI_COMM_WORLD;
195 leader_comm_size = smpi_comm_size(leader_comm);
196 leader_comm_rank = smpi_comm_rank(leader_comm);
197 tmp_buf=(void *)smpi_get_tmp_sendbuffer(count *
198 (MAX(extent, true_extent)));
199 tmp_buf = (void *) ((char *) tmp_buf - true_lb);
201 if (sendbuf != MPI_IN_PLACE) {
202 in_buf = (void *)sendbuf;
206 if (local_rank == 0) {
214 /* Lets do the intra-node reduce operations, if we have more than one
215 * process in the node */
217 /*Fix the input and outbuf buffers for the intra-node reduce.
218 *Node leaders will have the reduced data in tmp_buf after
220 if (MV2_Reduce_intra_function == & MPIR_Reduce_shmem_MV2)
222 if (is_commutative == 1
223 && (count * (MAX(extent, true_extent)) < SHMEM_COLL_BLOCK_SIZE)) {
224 mpi_errno = MV2_Reduce_intra_function(in_buf, out_buf, count,
226 intra_node_root, shmem_comm);
228 mpi_errno = MPIR_Reduce_intra_knomial_wrapper_MV2(in_buf, out_buf, count,
230 intra_node_root, shmem_comm);
234 mpi_errno = MV2_Reduce_intra_function(in_buf, out_buf, count,
236 intra_node_root, shmem_comm);
239 smpi_free_tmp_buffer((void *) ((char *) tmp_buf + true_lb));
243 /* Now work on the inter-leader phase. Data is in tmp_buf */
244 if (local_rank == 0 && leader_comm_size > 1) {
245 /*The leader of root will have the global reduced data in tmp_buf
247 at the end of the reduce */
248 if (leader_comm_rank == leader_root) {
249 if (my_rank == root) {
250 /* I am the root of the leader-comm, and the
251 * root of the reduce op. So, I will write the
252 * final result directly into my recvbuf */
253 if(tmp_buf != recvbuf) {
258 in_buf = (char *)smpi_get_tmp_sendbuffer(count*
259 smpi_datatype_get_extent(datatype));
260 smpi_datatype_copy(tmp_buf, count, datatype,
261 in_buf, count, datatype);
262 //in_buf = MPI_IN_PLACE;
266 in_buf = (char *)smpi_get_tmp_sendbuffer(count*
267 smpi_datatype_get_extent(datatype));
268 smpi_datatype_copy(tmp_buf, count, datatype,
269 in_buf, count, datatype);
270 //in_buf = MPI_IN_PLACE;
278 /* inter-leader communication */
279 mpi_errno = MV2_Reduce_function(in_buf, out_buf, count,
281 leader_root, leader_comm);
285 if (local_size > 1) {
286 /* Send the message to the root if the leader is not the
287 * root of the reduce operation. The reduced data is in tmp_buf */
288 if ((local_rank == 0) && (root != my_rank)
289 && (leader_root == leader_comm_rank)) {
290 smpi_mpi_send(tmp_buf, count, datatype, root,
291 COLL_TAG_REDUCE+1, comm);
293 if ((local_rank != 0) && (root == my_rank)) {
294 smpi_mpi_recv(recvbuf, count, datatype,
296 COLL_TAG_REDUCE+1, comm,
299 smpi_free_tmp_buffer((void *) ((char *) tmp_buf + true_lb));
301 if (leader_comm_rank == leader_root) {
302 if (my_rank != root || (my_rank == root && tmp_buf == recvbuf)) {
303 smpi_free_tmp_buffer(in_buf);