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. */
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
17 int smpi_coll_tuned_allreduce_smp_rsag_lr(void *send_buf, void *recv_buf,
18 int count, MPI_Datatype dtype,
19 MPI_Op op, MPI_Comm comm)
23 int tag = COLL_TAG_ALLREDUCE;
26 if(smpi_comm_get_leaders_comm(comm)==MPI_COMM_NULL){
27 smpi_comm_init_smp(comm);
30 if (smpi_comm_is_uniform(comm)){
31 num_core = smpi_comm_size(smpi_comm_get_intra_comm(comm));
34 #ifdef MPICH2_REDUCTION
35 MPI_User_function * uop = MPIR_Op_table[op % 16 - 1];
37 MPI_User_function *uop;
38 struct MPIR_OP *op_ptr;
39 op_ptr = MPIR_ToPointer(op);
43 comm_size = smpi_comm_size(comm);
44 rank = smpi_comm_rank(comm);
46 extent = smpi_datatype_get_extent(dtype);
47 tmp_buf = (void *) smpi_get_tmp_sendbuffer(count * extent);
49 int intra_rank, inter_rank;
50 intra_rank = rank % num_core;
51 inter_rank = rank / num_core;
53 //printf("node %d intra_rank = %d, inter_rank = %d\n", rank, intra_rank, inter_rank);
55 int inter_comm_size = (comm_size + num_core - 1) / num_core;
58 //printf("intra com size = %d\n",num_core);
59 //printf("inter com size = %d\n",inter_comm_size);
63 smpi_mpi_sendrecv(send_buf, count, dtype, rank, tag,
64 recv_buf, count, dtype, rank, tag, comm, &status);
67 // SMP_binomial_reduce
69 while (mask < num_core) {
70 if ((mask & intra_rank) == 0) {
71 src = (inter_rank * num_core) + (intra_rank | mask);
72 // if (src < ((inter_rank + 1) * num_core)) {
73 if (src < comm_size) {
74 smpi_mpi_recv(tmp_buf, count, dtype, src, tag, comm, &status);
75 smpi_op_apply(op, tmp_buf, recv_buf, &count, &dtype);
76 //printf("Node %d recv from node %d when mask is %d\n", rank, src, mask);
80 dst = (inter_rank * num_core) + (intra_rank & (~mask));
81 smpi_mpi_send(recv_buf, count, dtype, dst, tag, comm);
82 //printf("Node %d send to node %d when mask is %d\n", rank, dst, mask);
90 // INTER: reduce-scatter
91 if (intra_rank == 0) {
92 int send_offset, recv_offset;
93 int send_count, recv_count;
94 int curr_size = count / inter_comm_size;
95 int curr_remainder = count % inter_comm_size;
97 int to = ((inter_rank + 1) % inter_comm_size) * num_core;
99 ((inter_rank + inter_comm_size - 1) % inter_comm_size) * num_core;
102 //printf("node %d to %d from %d\n",rank,to,from);
104 /* last segment may have a larger size since it also include the remainder */
105 int last_segment_ptr =
106 (inter_comm_size - 1) * (count / inter_comm_size) * extent;
108 for (i = 0; i < (inter_comm_size - 1); i++) {
111 ((inter_rank - 1 - i +
112 inter_comm_size) % inter_comm_size) * curr_size * extent;
114 ((inter_rank - 2 - i +
115 inter_comm_size) % inter_comm_size) * curr_size * extent;
118 if (send_offset != last_segment_ptr)
119 send_count = curr_size;
121 send_count = curr_size + curr_remainder;
123 if (recv_offset != last_segment_ptr)
124 recv_count = curr_size;
126 recv_count = curr_size + curr_remainder;
128 smpi_mpi_sendrecv((char *) recv_buf + send_offset, send_count, dtype, to,
129 tag + i, tmp_buf, recv_count, dtype, from, tag + i, comm,
133 smpi_op_apply(op, tmp_buf, (char *) recv_buf + recv_offset, &recv_count,
138 for (i = 0; i < (inter_comm_size - 1); i++) {
142 inter_comm_size) % inter_comm_size) * curr_size * extent;
144 ((inter_rank - 1 - i +
145 inter_comm_size) % inter_comm_size) * curr_size * extent;
148 if (send_offset != last_segment_ptr)
149 send_count = curr_size;
151 send_count = curr_size + curr_remainder;
153 if (recv_offset != last_segment_ptr)
154 recv_count = curr_size;
156 recv_count = curr_size + curr_remainder;
158 smpi_mpi_sendrecv((char *) recv_buf + send_offset, send_count, dtype, to,
159 tag + i, (char *) recv_buf + recv_offset, recv_count, dtype,
160 from, tag + i, comm, &status);
168 // INTER_binomial_reduce
170 // only root node for each SMP
171 if (intra_rank == 0) {
174 while (mask < inter_comm_size) {
175 if ((mask & inter_rank) == 0) {
176 src = (inter_rank | mask) * num_core;
177 if (src < comm_size) {
178 smpi_mpi_recv(tmp_buf, count, dtype, src, tag, comm, &status);
179 (* uop) (tmp_buf, recv_buf, &count, &dtype);
180 //printf("Node %d recv from node %d when mask is %d\n", rank, src, mask);
184 dst = (inter_rank & (~mask)) * num_core;
185 smpi_mpi_send(recv_buf, count, dtype, dst, tag, comm);
186 //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 smpi_mpi_recv(recv_buf, count, dtype, src, tag, comm, &status);
211 //printf("My rank = %d my mask = %d\n", rank,mask);
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 smpi_mpi_send(recv_buf, count, dtype, dst, tag, comm);
227 // INTRA_binomial_bcast
229 int num_core_in_current_smp = num_core;
230 if (inter_rank == (inter_comm_size - 1)) {
231 num_core_in_current_smp = comm_size - (inter_rank * num_core);
233 // printf("Node %d num_core = %d\n",rank, num_core_in_current_smp);
235 while (mask < num_core_in_current_smp) {
236 if (intra_rank & mask) {
237 src = (inter_rank * num_core) + (intra_rank - mask);
238 //printf("Node %d recv from node %d when mask is %d\n", rank, src, mask);
239 smpi_mpi_recv(recv_buf, count, dtype, src, tag, comm, &status);
246 //printf("My rank = %d my mask = %d\n", rank,mask);
249 dst = (inter_rank * num_core) + (intra_rank + mask);
250 if (dst < comm_size) {
251 //printf("Node %d send to node %d when mask is %d\n", rank, dst, mask);
252 smpi_mpi_send(recv_buf, count, dtype, dst, tag, comm);
258 smpi_free_tmp_buffer(tmp_buf);