1 /* Copyright (c) 2013-2017. 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.h"
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(void *send_buf, void *recv_buf,
20 int count, MPI_Datatype dtype,
21 MPI_Op op, MPI_Comm comm)
25 int tag = COLL_TAG_ALLREDUCE;
28 if(comm->get_leaders_comm()==MPI_COMM_NULL){
32 if (comm->is_uniform()){
33 num_core = comm->get_intra_comm()->size();
36 #ifdef MPICH2_REDUCTION
37 MPI_User_function * uop = MPIR_Op_table[op % 16 - 1];
39 MPI_User_function *uop;
40 struct MPIR_OP *op_ptr;
41 op_ptr = MPIR_ToPointer(op);
45 comm_size = comm->size();
48 extent = dtype->get_extent();
49 tmp_buf = (void *) smpi_get_tmp_sendbuffer(count * extent);
51 int intra_rank, inter_rank;
52 intra_rank = rank % num_core;
53 inter_rank = rank / num_core;
55 //printf("node %d intra_rank = %d, inter_rank = %d\n", rank, intra_rank, inter_rank);
57 int inter_comm_size = (comm_size + num_core - 1) / num_core;
60 //printf("intra com size = %d\n",num_core);
61 //printf("inter com size = %d\n",inter_comm_size);
65 Request::sendrecv(send_buf, count, dtype, rank, tag,
66 recv_buf, count, dtype, rank, tag, comm, &status);
69 // SMP_binomial_reduce
71 while (mask < num_core) {
72 if ((mask & intra_rank) == 0) {
73 src = (inter_rank * num_core) + (intra_rank | mask);
74 // if (src < ((inter_rank + 1) * num_core)) {
75 if (src < comm_size) {
76 Request::recv(tmp_buf, count, dtype, src, tag, comm, &status);
77 if(op!=MPI_OP_NULL) op->apply( tmp_buf, recv_buf, &count, dtype);
78 //printf("Node %d recv from node %d when mask is %d\n", rank, src, mask);
82 dst = (inter_rank * num_core) + (intra_rank & (~mask));
83 Request::send(recv_buf, count, dtype, dst, tag, comm);
84 //printf("Node %d send to node %d when mask is %d\n", rank, dst, mask);
92 // INTER: reduce-scatter
93 if (intra_rank == 0) {
94 int send_offset, recv_offset;
95 int send_count, recv_count;
96 int curr_size = count / inter_comm_size;
97 int curr_remainder = count % inter_comm_size;
99 int to = ((inter_rank + 1) % inter_comm_size) * num_core;
101 ((inter_rank + inter_comm_size - 1) % inter_comm_size) * num_core;
104 //printf("node %d to %d from %d\n",rank,to,from);
106 /* last segment may have a larger size since it also include the remainder */
107 int last_segment_ptr =
108 (inter_comm_size - 1) * (count / inter_comm_size) * extent;
110 for (i = 0; i < (inter_comm_size - 1); i++) {
113 ((inter_rank - 1 - i +
114 inter_comm_size) % inter_comm_size) * curr_size * extent;
116 ((inter_rank - 2 - i +
117 inter_comm_size) % inter_comm_size) * curr_size * extent;
120 if (send_offset != last_segment_ptr)
121 send_count = curr_size;
123 send_count = curr_size + curr_remainder;
125 if (recv_offset != last_segment_ptr)
126 recv_count = curr_size;
128 recv_count = curr_size + curr_remainder;
130 Request::sendrecv((char *) recv_buf + send_offset, send_count, dtype, to,
131 tag + i, tmp_buf, recv_count, dtype, from, tag + i, comm,
135 if(op!=MPI_OP_NULL) op->apply( tmp_buf, (char *) recv_buf + recv_offset, &recv_count,
140 for (i = 0; i < (inter_comm_size - 1); i++) {
144 inter_comm_size) % inter_comm_size) * curr_size * extent;
146 ((inter_rank - 1 - i +
147 inter_comm_size) % inter_comm_size) * curr_size * extent;
150 if (send_offset != last_segment_ptr)
151 send_count = curr_size;
153 send_count = curr_size + curr_remainder;
155 if (recv_offset != last_segment_ptr)
156 recv_count = curr_size;
158 recv_count = curr_size + curr_remainder;
160 Request::sendrecv((char *) recv_buf + send_offset, send_count, dtype, to,
161 tag + i, (char *) recv_buf + recv_offset, recv_count, dtype,
162 from, tag + i, comm, &status);
169 // INTER_binomial_reduce
171 // only root node for each SMP
173 if (intra_rank == 0) {
176 while (mask < inter_comm_size) {
177 if ((mask & inter_rank) == 0) {
178 src = (inter_rank | mask) * num_core;
179 if (src < comm_size) {
180 Request::recv(tmp_buf, count, dtype, src, tag, comm, &status);
181 (* uop) (tmp_buf, recv_buf, &count, &dtype);
182 //printf("Node %d recv from node %d when mask is %d\n", rank, src, mask);
186 dst = (inter_rank & (~mask)) * num_core;
187 Request::send(recv_buf, count, dtype, dst, tag, comm);
188 //printf("Node %d send to node %d when mask is %d\n", rank, dst, mask);
196 // INTER_binomial_bcast
199 // if (intra_rank == 0) {
201 // while (mask < inter_comm_size) {
202 // if (inter_rank & mask) {
203 // src = (inter_rank - mask) * num_core;
204 //printf("Node %d recv from node %d when mask is %d\n", rank, src, mask);
205 // Request::recv(recv_buf, count, dtype, src, tag, comm, &status);
212 //printf("My rank = %d my mask = %d\n", rank,mask);
214 // while (mask > 0) {
215 // if (inter_rank < inter_comm_size) {
216 // dst = (inter_rank + mask) * num_core;
217 // if (dst < comm_size) {
218 // //printf("Node %d send to node %d when mask is %d\n", rank, dst, mask);
219 // Request::send(recv_buf, count, dtype, dst, tag, comm);
226 // INTRA_binomial_bcast
228 int num_core_in_current_smp = num_core;
229 if (inter_rank == (inter_comm_size - 1)) {
230 num_core_in_current_smp = comm_size - (inter_rank * num_core);
232 // printf("Node %d num_core = %d\n",rank, num_core_in_current_smp);
234 while (mask < num_core_in_current_smp) {
235 if (intra_rank & mask) {
236 src = (inter_rank * num_core) + (intra_rank - mask);
237 //printf("Node %d recv from node %d when mask is %d\n", rank, src, mask);
238 Request::recv(recv_buf, count, dtype, src, tag, comm, &status);
245 //printf("My rank = %d my mask = %d\n", rank,mask);
248 dst = (inter_rank * num_core) + (intra_rank + mask);
249 if (dst < comm_size) {
250 //printf("Node %d send to node %d when mask is %d\n", rank, dst, mask);
251 Request::send(recv_buf, count, dtype, dst, tag, comm);
257 smpi_free_tmp_buffer(tmp_buf);