1 /* Copyright (c) 2013-2020. 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);
78 int reduce__mvapich2_two_level( const void *sendbuf,
81 MPI_Datatype datatype,
86 int mpi_errno = MPI_SUCCESS;
87 int my_rank, total_size, local_rank, local_size;
88 int leader_comm_rank = -1, leader_comm_size = 0;
89 MPI_Comm shmem_comm, leader_comm;
90 int leader_root, leader_of_root;
91 const unsigned char* in_buf = nullptr;
92 unsigned char *out_buf = nullptr, *tmp_buf = nullptr;
93 MPI_Aint true_lb, true_extent, extent;
95 int intra_node_root=0;
97 //if not set (use of the algo directly, without mvapich2 selector)
98 if (MV2_Reduce_function == nullptr)
99 MV2_Reduce_function = reduce__mpich;
100 if (MV2_Reduce_intra_function == nullptr)
101 MV2_Reduce_intra_function = reduce__mpich;
103 if(comm->get_leaders_comm()==MPI_COMM_NULL){
107 my_rank = comm->rank();
108 total_size = comm->size();
109 shmem_comm = comm->get_intra_comm();
110 local_rank = shmem_comm->rank();
111 local_size = shmem_comm->size();
113 leader_comm = comm->get_leaders_comm();
114 int* leaders_map = comm->get_leaders_map();
115 leader_of_root = comm->group()->rank(leaders_map[root]);
116 leader_root = leader_comm->group()->rank(leaders_map[root]);
118 bool is_commutative = (op == MPI_OP_NULL || op->is_commutative());
120 datatype->extent(&true_lb,
122 extent =datatype->get_extent();
123 stride = count * std::max(extent, true_extent);
125 if (local_size == total_size) {
126 /* First handle the case where there is only one node */
127 if (stride <= MV2_INTRA_SHMEM_REDUCE_MSG && is_commutative) {
128 if (local_rank == 0 ) {
129 tmp_buf = smpi_get_tmp_sendbuffer(count * std::max(extent, true_extent));
130 tmp_buf = tmp_buf - true_lb;
133 if (sendbuf != MPI_IN_PLACE) {
134 in_buf = static_cast<const unsigned char*>(sendbuf);
136 in_buf = static_cast<const unsigned char*>(recvbuf);
139 if (local_rank == 0) {
140 if( my_rank != root) {
143 out_buf = static_cast<unsigned char*>(recvbuf);
144 if (in_buf == out_buf) {
145 in_buf = static_cast<const unsigned char*>(MPI_IN_PLACE);
146 out_buf = static_cast<unsigned char*>(recvbuf);
150 in_buf = static_cast<const unsigned char*>(sendbuf);
154 if (count * (std::max(extent, true_extent)) < SHMEM_COLL_BLOCK_SIZE) {
155 mpi_errno = MPIR_Reduce_shmem_MV2(in_buf, out_buf, count, datatype, op, 0, shmem_comm);
157 mpi_errno = MPIR_Reduce_intra_knomial_wrapper_MV2(in_buf, out_buf, count, datatype, op, 0, shmem_comm);
160 if (local_rank == 0 && root != my_rank) {
161 Request::send(out_buf, count, datatype, root,
162 COLL_TAG_REDUCE+1, comm);
164 if ((local_rank != 0) && (root == my_rank)) {
165 Request::recv(recvbuf, count, datatype,
166 leader_of_root, COLL_TAG_REDUCE+1, comm,
170 if(mv2_use_knomial_reduce == 1) {
171 reduce_fn = &MPIR_Reduce_intra_knomial_wrapper_MV2;
173 reduce_fn = &MPIR_Reduce_binomial_MV2;
175 mpi_errno = reduce_fn(sendbuf, recvbuf, count,
180 if (tmp_buf != nullptr)
181 smpi_free_tmp_buffer(tmp_buf + true_lb);
186 if (local_rank == 0) {
187 leader_comm = comm->get_leaders_comm();
188 if(leader_comm==MPI_COMM_NULL){
189 leader_comm = MPI_COMM_WORLD;
191 leader_comm_size = leader_comm->size();
192 leader_comm_rank = leader_comm->rank();
193 tmp_buf = smpi_get_tmp_sendbuffer(count * std::max(extent, true_extent));
194 tmp_buf = tmp_buf - true_lb;
196 if (sendbuf != MPI_IN_PLACE) {
197 in_buf = static_cast<const unsigned char*>(sendbuf);
199 in_buf = static_cast<const unsigned char*>(recvbuf);
201 if (local_rank == 0) {
202 out_buf = static_cast<unsigned char*>(tmp_buf);
209 /* Lets do the intra-node reduce operations, if we have more than one
210 * process in the node */
212 /*Fix the input and outbuf buffers for the intra-node reduce.
213 *Node leaders will have the reduced data in tmp_buf after
215 if (MV2_Reduce_intra_function == & MPIR_Reduce_shmem_MV2)
217 if (is_commutative && (count * (std::max(extent, true_extent)) < SHMEM_COLL_BLOCK_SIZE)) {
218 mpi_errno = MV2_Reduce_intra_function(in_buf, out_buf, count, datatype, op, intra_node_root, shmem_comm);
220 mpi_errno = MPIR_Reduce_intra_knomial_wrapper_MV2(in_buf, out_buf, count,
222 intra_node_root, shmem_comm);
226 mpi_errno = MV2_Reduce_intra_function(in_buf, out_buf, count,
228 intra_node_root, shmem_comm);
231 smpi_free_tmp_buffer(tmp_buf + true_lb);
232 tmp_buf = (unsigned char*)in_buf; // xxx
235 /* Now work on the inter-leader phase. Data is in tmp_buf */
236 if (local_rank == 0 && leader_comm_size > 1) {
237 /*The leader of root will have the global reduced data in tmp_buf
239 at the end of the reduce */
240 if (leader_comm_rank == leader_root) {
241 if (my_rank == root) {
242 /* I am the root of the leader-comm, and the
243 * root of the reduce op. So, I will write the
244 * final result directly into my recvbuf */
245 if(tmp_buf != recvbuf) {
247 out_buf = static_cast<unsigned char*>(recvbuf);
250 unsigned char* buf = smpi_get_tmp_sendbuffer(count * datatype->get_extent());
251 Datatype::copy(tmp_buf, count, datatype, buf, count, datatype);
252 // in_buf = MPI_IN_PLACE;
254 out_buf = static_cast<unsigned char*>(recvbuf);
257 unsigned char* buf = smpi_get_tmp_sendbuffer(count * datatype->get_extent());
258 Datatype::copy(tmp_buf, count, datatype, buf, count, datatype);
259 // in_buf = MPI_IN_PLACE;
268 /* inter-leader communication */
269 mpi_errno = MV2_Reduce_function(in_buf, out_buf, count,
271 leader_root, leader_comm);
275 if (local_size > 1) {
276 /* Send the message to the root if the leader is not the
277 * root of the reduce operation. The reduced data is in tmp_buf */
278 if ((local_rank == 0) && (root != my_rank) && (leader_root == leader_comm_rank)) {
279 Request::send(tmp_buf, count, datatype, root, COLL_TAG_REDUCE + 1, comm);
281 if ((local_rank != 0) && (root == my_rank)) {
282 Request::recv(recvbuf, count, datatype, leader_of_root, COLL_TAG_REDUCE + 1, comm, MPI_STATUS_IGNORE);
284 smpi_free_tmp_buffer(tmp_buf + true_lb);
286 if (leader_comm_rank == leader_root) {
287 if (my_rank != root || (my_rank == root && tmp_buf == recvbuf)) {
288 smpi_free_tmp_buffer(in_buf);