1 /* Copyright (c) 2013-2019. 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. */
7 #include "../colls_private.hpp"
8 //#include <star-reduction.c>
11 This fucntion performs all-reduce operation as follow.
12 1) binomial_tree reduce inside each SMP node
13 2) reduce-scatter -inter between root of each SMP node
14 3) allgather - inter between root of each SMP node
15 4) binomial_tree bcast inside each SMP node
19 int Coll_allreduce_smp_rsag_lr::allreduce(const void *send_buf, void *recv_buf,
20 int count, MPI_Datatype dtype,
21 MPI_Op op, MPI_Comm comm)
24 int tag = COLL_TAG_ALLREDUCE;
27 if(comm->get_leaders_comm()==MPI_COMM_NULL){
31 if (comm->is_uniform()){
32 num_core = comm->get_intra_comm()->size();
35 #ifdef MPICH2_REDUCTION
36 MPI_User_function * uop = MPIR_Op_table[op % 16 - 1];
38 MPI_User_function *uop;
40 op_ptr = MPIR_ToPointer(op);
44 comm_size = comm->size();
47 extent = dtype->get_extent();
48 unsigned char* tmp_buf = smpi_get_tmp_sendbuffer(count * extent);
50 int intra_rank, inter_rank;
51 intra_rank = rank % num_core;
52 inter_rank = rank / num_core;
54 //printf("node %d intra_rank = %d, inter_rank = %d\n", rank, intra_rank, inter_rank);
56 int inter_comm_size = (comm_size + num_core - 1) / num_core;
59 //printf("intra com size = %d\n",num_core);
60 //printf("inter com size = %d\n",inter_comm_size);
64 Request::sendrecv(send_buf, count, dtype, rank, tag,
65 recv_buf, count, dtype, rank, tag, comm, &status);
68 // SMP_binomial_reduce
70 while (mask < num_core) {
71 if ((mask & intra_rank) == 0) {
72 src = (inter_rank * num_core) + (intra_rank | mask);
73 // if (src < ((inter_rank + 1) * num_core)) {
74 if (src < comm_size) {
75 Request::recv(tmp_buf, count, dtype, src, tag, comm, &status);
76 if(op!=MPI_OP_NULL) op->apply( tmp_buf, recv_buf, &count, dtype);
77 //printf("Node %d recv from node %d when mask is %d\n", rank, src, mask);
81 dst = (inter_rank * num_core) + (intra_rank & (~mask));
82 Request::send(recv_buf, count, dtype, dst, tag, comm);
83 //printf("Node %d send to node %d when mask is %d\n", rank, dst, mask);
91 // INTER: reduce-scatter
92 if (intra_rank == 0) {
93 int send_offset, recv_offset;
94 int send_count, recv_count;
95 int curr_size = count / inter_comm_size;
96 int curr_remainder = count % inter_comm_size;
98 int to = ((inter_rank + 1) % inter_comm_size) * num_core;
100 ((inter_rank + inter_comm_size - 1) % inter_comm_size) * num_core;
103 //printf("node %d to %d from %d\n",rank,to,from);
105 /* last segment may have a larger size since it also include the remainder */
106 int last_segment_ptr =
107 (inter_comm_size - 1) * (count / inter_comm_size) * extent;
109 for (i = 0; i < (inter_comm_size - 1); i++) {
112 ((inter_rank - 1 - i +
113 inter_comm_size) % inter_comm_size) * curr_size * extent;
115 ((inter_rank - 2 - i +
116 inter_comm_size) % inter_comm_size) * curr_size * extent;
119 if (send_offset != last_segment_ptr)
120 send_count = curr_size;
122 send_count = curr_size + curr_remainder;
124 if (recv_offset != last_segment_ptr)
125 recv_count = curr_size;
127 recv_count = curr_size + curr_remainder;
129 Request::sendrecv((char *) recv_buf + send_offset, send_count, dtype, to,
130 tag + i, tmp_buf, recv_count, dtype, from, tag + i, comm,
134 if(op!=MPI_OP_NULL) op->apply( tmp_buf, (char *) recv_buf + recv_offset, &recv_count,
139 for (i = 0; i < (inter_comm_size - 1); i++) {
143 inter_comm_size) % inter_comm_size) * curr_size * extent;
145 ((inter_rank - 1 - i +
146 inter_comm_size) % inter_comm_size) * curr_size * extent;
149 if (send_offset != last_segment_ptr)
150 send_count = curr_size;
152 send_count = curr_size + curr_remainder;
154 if (recv_offset != last_segment_ptr)
155 recv_count = curr_size;
157 recv_count = curr_size + curr_remainder;
159 Request::sendrecv((char *) recv_buf + send_offset, send_count, dtype, to,
160 tag + i, (char *) recv_buf + recv_offset, recv_count, dtype,
161 from, tag + i, comm, &status);
168 // INTER_binomial_reduce
170 // only root node for each SMP
172 if (intra_rank == 0) {
175 while (mask < inter_comm_size) {
176 if ((mask & inter_rank) == 0) {
177 src = (inter_rank | mask) * num_core;
178 if (src < comm_size) {
179 Request::recv(tmp_buf, count, dtype, src, tag, comm, &status);
180 (* uop) (tmp_buf, recv_buf, &count, &dtype);
181 //printf("Node %d recv from node %d when mask is %d\n", rank, src, mask);
185 dst = (inter_rank & (~mask)) * num_core;
186 Request::send(recv_buf, count, dtype, dst, tag, comm);
187 //printf("Node %d send to node %d when mask is %d\n", rank, dst, mask);
195 // INTER_binomial_bcast
198 // if (intra_rank == 0) {
200 // while (mask < inter_comm_size) {
201 // if (inter_rank & mask) {
202 // src = (inter_rank - mask) * num_core;
203 //printf("Node %d recv from node %d when mask is %d\n", rank, src, mask);
204 // Request::recv(recv_buf, count, dtype, src, tag, comm, &status);
211 //printf("My rank = %d my mask = %d\n", rank,mask);
213 // while (mask > 0) {
214 // if (inter_rank < inter_comm_size) {
215 // dst = (inter_rank + mask) * num_core;
216 // if (dst < comm_size) {
217 // //printf("Node %d send to node %d when mask is %d\n", rank, dst, mask);
218 // Request::send(recv_buf, count, dtype, dst, tag, comm);
225 // INTRA_binomial_bcast
227 int num_core_in_current_smp = num_core;
228 if (inter_rank == (inter_comm_size - 1)) {
229 num_core_in_current_smp = comm_size - (inter_rank * num_core);
231 // printf("Node %d num_core = %d\n",rank, num_core_in_current_smp);
233 while (mask < num_core_in_current_smp) {
234 if (intra_rank & mask) {
235 src = (inter_rank * num_core) + (intra_rank - mask);
236 //printf("Node %d recv from node %d when mask is %d\n", rank, src, mask);
237 Request::recv(recv_buf, count, dtype, src, tag, comm, &status);
244 //printf("My rank = %d my mask = %d\n", rank,mask);
247 dst = (inter_rank * num_core) + (intra_rank + mask);
248 if (dst < comm_size) {
249 //printf("Node %d send to node %d when mask is %d\n", rank, dst, mask);
250 Request::send(recv_buf, count, dtype, dst, tag, comm);
256 smpi_free_tmp_buffer(tmp_buf);