1 /* Copyright (c) 2013-2023. 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.hpp"
41 #define MV2_INTRA_SHMEM_REDUCE_MSG 2048
43 #define mv2_g_shmem_coll_max_msg_size (1 << 17)
44 #define SHMEM_COLL_BLOCK_SIZE (local_size * mv2_g_shmem_coll_max_msg_size)
45 #define mv2_use_knomial_reduce 1
47 #define MPIR_Reduce_inter_knomial_wrapper_MV2 reduce__mvapich2_knomial
48 #define MPIR_Reduce_intra_knomial_wrapper_MV2 reduce__mvapich2_knomial
49 #define MPIR_Reduce_binomial_MV2 reduce__binomial
50 #define MPIR_Reduce_redscat_gather_MV2 reduce__scatter_gather
51 #define MPIR_Reduce_shmem_MV2 reduce__ompi_basic_linear
53 extern int (*MV2_Reduce_function)( const void *sendbuf,
56 MPI_Datatype datatype,
61 extern int (*MV2_Reduce_intra_function)( const void *sendbuf,
64 MPI_Datatype datatype,
70 /*Fn pointers for collectives */
71 static int (*reduce_fn)(const void *sendbuf,
74 MPI_Datatype datatype,
75 MPI_Op op, int root, MPI_Comm comm);
76 namespace simgrid::smpi {
77 int reduce__mvapich2_two_level( const void *sendbuf,
80 MPI_Datatype datatype,
85 int mpi_errno = MPI_SUCCESS;
86 int my_rank, total_size, local_rank, local_size;
87 int leader_comm_rank = -1, leader_comm_size = 0;
88 MPI_Comm shmem_comm, leader_comm;
89 int leader_root, leader_of_root;
90 const unsigned char* in_buf = nullptr;
91 unsigned char *out_buf = nullptr, *tmp_buf = nullptr;
92 MPI_Aint true_lb, true_extent, extent;
94 int intra_node_root=0;
96 //if not set (use of the algo directly, without mvapich2 selector)
97 if (MV2_Reduce_function == nullptr)
98 MV2_Reduce_function = reduce__mpich;
99 if (MV2_Reduce_intra_function == nullptr)
100 MV2_Reduce_intra_function = reduce__mpich;
102 if(comm->get_leaders_comm()==MPI_COMM_NULL){
106 my_rank = comm->rank();
107 total_size = comm->size();
108 shmem_comm = comm->get_intra_comm();
109 local_rank = shmem_comm->rank();
110 local_size = shmem_comm->size();
112 leader_comm = comm->get_leaders_comm();
113 int* leaders_map = comm->get_leaders_map();
114 leader_of_root = comm->group()->rank(leaders_map[root]);
115 leader_root = leader_comm->group()->rank(leaders_map[root]);
117 bool is_commutative = (op == MPI_OP_NULL || op->is_commutative());
119 datatype->extent(&true_lb,
121 extent =datatype->get_extent();
122 stride = count * std::max(extent, true_extent);
124 if (local_size == total_size) {
125 /* First handle the case where there is only one node */
126 if (stride <= MV2_INTRA_SHMEM_REDUCE_MSG && is_commutative) {
127 if (local_rank == 0 ) {
128 tmp_buf = smpi_get_tmp_sendbuffer(count * std::max(extent, true_extent));
129 tmp_buf = tmp_buf - true_lb;
132 if (sendbuf != MPI_IN_PLACE) {
133 in_buf = static_cast<const unsigned char*>(sendbuf);
135 in_buf = static_cast<const unsigned char*>(recvbuf);
138 if (local_rank == 0) {
139 if( my_rank != root) {
142 out_buf = static_cast<unsigned char*>(recvbuf);
143 if (in_buf == out_buf) {
144 in_buf = static_cast<const unsigned char*>(MPI_IN_PLACE);
145 out_buf = static_cast<unsigned char*>(recvbuf);
149 in_buf = static_cast<const unsigned char*>(sendbuf);
153 if (count * (std::max(extent, true_extent)) < SHMEM_COLL_BLOCK_SIZE) {
154 mpi_errno = MPIR_Reduce_shmem_MV2(in_buf, out_buf, count, datatype, op, 0, shmem_comm);
156 mpi_errno = MPIR_Reduce_intra_knomial_wrapper_MV2(in_buf, out_buf, count, datatype, op, 0, shmem_comm);
159 if (local_rank == 0 && root != my_rank) {
160 Request::send(out_buf, count, datatype, root,
161 COLL_TAG_REDUCE+1, comm);
163 if ((local_rank != 0) && (root == my_rank)) {
164 Request::recv(recvbuf, count, datatype,
165 leader_of_root, COLL_TAG_REDUCE+1, comm,
169 if(mv2_use_knomial_reduce == 1) {
170 reduce_fn = &MPIR_Reduce_intra_knomial_wrapper_MV2;
172 reduce_fn = &MPIR_Reduce_binomial_MV2;
174 mpi_errno = reduce_fn(sendbuf, recvbuf, count,
179 if (tmp_buf != nullptr)
180 smpi_free_tmp_buffer(tmp_buf + true_lb);
185 if (local_rank == 0) {
186 leader_comm = comm->get_leaders_comm();
187 if(leader_comm==MPI_COMM_NULL){
188 leader_comm = MPI_COMM_WORLD;
190 leader_comm_size = leader_comm->size();
191 leader_comm_rank = leader_comm->rank();
192 tmp_buf = smpi_get_tmp_sendbuffer(count * std::max(extent, true_extent));
193 tmp_buf = tmp_buf - true_lb;
195 if (sendbuf != MPI_IN_PLACE) {
196 in_buf = static_cast<const unsigned char*>(sendbuf);
198 in_buf = static_cast<const unsigned char*>(recvbuf);
200 if (local_rank == 0) {
201 out_buf = static_cast<unsigned char*>(tmp_buf);
208 /* Lets do the intra-node reduce operations, if we have more than one
209 * process in the node */
211 /*Fix the input and outbuf buffers for the intra-node reduce.
212 *Node leaders will have the reduced data in tmp_buf after
214 if (MV2_Reduce_intra_function == & MPIR_Reduce_shmem_MV2)
216 if (is_commutative && (count * (std::max(extent, true_extent)) < SHMEM_COLL_BLOCK_SIZE)) {
217 mpi_errno = MV2_Reduce_intra_function(in_buf, out_buf, count, datatype, op, intra_node_root, shmem_comm);
219 mpi_errno = MPIR_Reduce_intra_knomial_wrapper_MV2(in_buf, out_buf, count,
221 intra_node_root, shmem_comm);
225 mpi_errno = MV2_Reduce_intra_function(in_buf, out_buf, count,
227 intra_node_root, shmem_comm);
230 smpi_free_tmp_buffer(tmp_buf + true_lb);
231 tmp_buf = (unsigned char*)in_buf; // xxx
234 /* Now work on the inter-leader phase. Data is in tmp_buf */
235 if (local_rank == 0 && leader_comm_size > 1) {
236 /*The leader of root will have the global reduced data in tmp_buf
238 at the end of the reduce */
239 if (leader_comm_rank == leader_root) {
240 if (my_rank == root) {
241 /* I am the root of the leader-comm, and the
242 * root of the reduce op. So, I will write the
243 * final result directly into my recvbuf */
244 if(tmp_buf != recvbuf) {
246 out_buf = static_cast<unsigned char*>(recvbuf);
249 unsigned char* buf = smpi_get_tmp_sendbuffer(count * datatype->get_extent());
250 Datatype::copy(tmp_buf, count, datatype, buf, count, datatype);
251 // in_buf = MPI_IN_PLACE;
253 out_buf = static_cast<unsigned char*>(recvbuf);
256 unsigned char* buf = smpi_get_tmp_sendbuffer(count * datatype->get_extent());
257 Datatype::copy(tmp_buf, count, datatype, buf, count, datatype);
258 // in_buf = MPI_IN_PLACE;
267 /* inter-leader communication */
268 mpi_errno = MV2_Reduce_function(in_buf, out_buf, count,
270 leader_root, leader_comm);
274 if (local_size > 1) {
275 /* Send the message to the root if the leader is not the
276 * root of the reduce operation. The reduced data is in tmp_buf */
277 if ((local_rank == 0) && (root != my_rank) && (leader_root == leader_comm_rank)) {
278 Request::send(tmp_buf, count, datatype, root, COLL_TAG_REDUCE + 1, comm);
280 if ((local_rank != 0) && (root == my_rank)) {
281 Request::recv(recvbuf, count, datatype, leader_of_root, COLL_TAG_REDUCE + 1, comm, MPI_STATUS_IGNORE);
283 smpi_free_tmp_buffer(tmp_buf + true_lb);
285 if (leader_comm_rank == leader_root) {
286 if (my_rank != root || (my_rank == root && tmp_buf == recvbuf)) {
287 smpi_free_tmp_buffer(in_buf);
297 } // namespace simgrid::smpi