1 /* Copyright (c) 2013-2022. 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"
10 This function performs all-reduce operation as follow.
11 1) binomial_tree reduce inside each SMP node
12 2) reduce-scatter -inter between root of each SMP node
13 3) allgather - inter between root of each SMP node
14 4) binomial_tree bcast inside each SMP node
18 int allreduce__smp_rsag(const void *send_buf, void *recv_buf,
19 int count, MPI_Datatype dtype, MPI_Op op,
23 int tag = COLL_TAG_ALLREDUCE;
26 if(comm->get_leaders_comm()==MPI_COMM_NULL){
30 if (comm->is_uniform()){
31 num_core = comm->get_intra_comm()->size();
34 #ifdef MPICH2_REDUCTION
35 MPI_User_function * uop = MPIR_Op_table[op % 16 - 1];
37 MPI_User_function *uop;
39 op_ptr = MPIR_ToPointer(op);
43 comm_size = comm->size();
46 extent = dtype->get_extent();
47 unsigned char* tmp_buf = 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 Request::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 Request::recv(tmp_buf, count, dtype, src, tag, comm, &status);
75 if(op!=MPI_OP_NULL) op->apply( 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 Request::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 seg_count = count / inter_comm_size;
94 int to = ((inter_rank + 1) % inter_comm_size) * num_core;
96 ((inter_rank + inter_comm_size - 1) % inter_comm_size) * num_core;
99 //printf("node %d to %d from %d\n",rank,to,from);
101 for (i = 0; i < (inter_comm_size - 1); i++) {
104 ((inter_rank - 1 - i +
105 inter_comm_size) % inter_comm_size) * seg_count * extent;
107 ((inter_rank - 2 - i +
108 inter_comm_size) % inter_comm_size) * seg_count * extent;
110 Request::sendrecv((char *) recv_buf + send_offset, seg_count, dtype, to,
111 tag + i, tmp_buf, seg_count, dtype, from, tag + i, comm,
115 if(op!=MPI_OP_NULL) op->apply( tmp_buf, (char *) recv_buf + recv_offset, &seg_count,
120 for (i = 0; i < (inter_comm_size - 1); i++) {
124 inter_comm_size) % inter_comm_size) * seg_count * extent;
126 ((inter_rank - 1 - i +
127 inter_comm_size) % inter_comm_size) * seg_count * extent;
129 Request::sendrecv((char *) recv_buf + send_offset, seg_count, dtype, to,
130 tag + i, (char *) recv_buf + recv_offset, seg_count, dtype,
131 from, tag + i, comm, &status);
138 // INTER_binomial_reduce
140 // only root node for each SMP
141 // if (intra_rank == 0) {
144 // while (mask < inter_comm_size) {
145 // if ((mask & inter_rank) == 0) {
146 // src = (inter_rank | mask) * num_core;
147 // if (src < comm_size) {
148 // Request::recv(tmp_buf, count, dtype, src, tag, comm, &status);
149 // (* uop) (tmp_buf, recv_buf, &count, &dtype);
150 //printf("Node %d recv from node %d when mask is %d\n", rank, src, mask);
154 // dst = (inter_rank & (~mask)) * num_core;
155 // Request::send(recv_buf, count, dtype, dst, tag, comm);
156 //printf("Node %d send to node %d when mask is %d\n", rank, dst, mask);
163 // INTER_binomial_bcast
166 // if (intra_rank == 0) {
168 // while (mask < inter_comm_size) {
169 // if (inter_rank & mask) {
170 // src = (inter_rank - mask) * num_core;
171 //printf("Node %d recv from node %d when mask is %d\n", rank, src, mask);
172 // Request::recv(recv_buf, count, dtype, src, tag, comm, &status);
179 //printf("My rank = %d my mask = %d\n", rank,mask);
181 // while (mask > 0) {
182 // if (inter_rank < inter_comm_size) {
183 // dst = (inter_rank + mask) * num_core;
184 // if (dst < comm_size) {
185 //printf("Node %d send to node %d when mask is %d\n", rank, dst, mask);
186 // Request::send(recv_buf, count, dtype, dst, tag, comm);
194 // INTRA_binomial_bcast
196 int num_core_in_current_smp = num_core;
197 if (inter_rank == (inter_comm_size - 1)) {
198 num_core_in_current_smp = comm_size - (inter_rank * num_core);
200 // printf("Node %d num_core = %d\n",rank, num_core_in_current_smp);
202 while (mask < num_core_in_current_smp) {
203 if (intra_rank & mask) {
204 src = (inter_rank * num_core) + (intra_rank - mask);
205 //printf("Node %d recv from node %d when mask is %d\n", rank, src, mask);
206 Request::recv(recv_buf, count, dtype, src, tag, comm, &status);
213 //printf("My rank = %d my mask = %d\n", rank,mask);
216 dst = (inter_rank * num_core) + (intra_rank + mask);
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);
224 smpi_free_tmp_buffer(tmp_buf);