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"
10 This fucntion 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 Coll_allreduce_smp_rsag::allreduce(void *send_buf, void *recv_buf,
19 int count, MPI_Datatype dtype, MPI_Op op,
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;
39 struct MPIR_OP *op_ptr;
40 op_ptr = MPIR_ToPointer(op);
44 comm_size = comm->size();
47 extent = dtype->get_extent();
48 tmp_buf = (void *) 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 seg_count = count / inter_comm_size;
95 int to = ((inter_rank + 1) % inter_comm_size) * num_core;
97 ((inter_rank + inter_comm_size - 1) % inter_comm_size) * num_core;
100 //printf("node %d to %d from %d\n",rank,to,from);
102 for (i = 0; i < (inter_comm_size - 1); i++) {
105 ((inter_rank - 1 - i +
106 inter_comm_size) % inter_comm_size) * seg_count * extent;
108 ((inter_rank - 2 - i +
109 inter_comm_size) % inter_comm_size) * seg_count * extent;
111 Request::sendrecv((char *) recv_buf + send_offset, seg_count, dtype, to,
112 tag + i, tmp_buf, seg_count, dtype, from, tag + i, comm,
116 if(op!=MPI_OP_NULL) op->apply( tmp_buf, (char *) recv_buf + recv_offset, &seg_count,
121 for (i = 0; i < (inter_comm_size - 1); i++) {
125 inter_comm_size) % inter_comm_size) * seg_count * extent;
127 ((inter_rank - 1 - i +
128 inter_comm_size) % inter_comm_size) * seg_count * extent;
130 Request::sendrecv((char *) recv_buf + send_offset, seg_count, dtype, to,
131 tag + i, (char *) recv_buf + recv_offset, seg_count, dtype,
132 from, tag + i, comm, &status);
139 // INTER_binomial_reduce
141 // only root node for each SMP
142 // if (intra_rank == 0) {
145 // while (mask < inter_comm_size) {
146 // if ((mask & inter_rank) == 0) {
147 // src = (inter_rank | mask) * num_core;
148 // if (src < comm_size) {
149 // Request::recv(tmp_buf, count, dtype, src, tag, comm, &status);
150 // (* uop) (tmp_buf, recv_buf, &count, &dtype);
151 //printf("Node %d recv from node %d when mask is %d\n", rank, src, mask);
155 // dst = (inter_rank & (~mask)) * num_core;
156 // Request::send(recv_buf, count, dtype, dst, tag, comm);
157 //printf("Node %d send to node %d when mask is %d\n", rank, dst, mask);
164 // INTER_binomial_bcast
167 // if (intra_rank == 0) {
169 // while (mask < inter_comm_size) {
170 // if (inter_rank & mask) {
171 // src = (inter_rank - mask) * num_core;
172 //printf("Node %d recv from node %d when mask is %d\n", rank, src, mask);
173 // Request::recv(recv_buf, count, dtype, src, tag, comm, &status);
180 //printf("My rank = %d my mask = %d\n", rank,mask);
182 // while (mask > 0) {
183 // if (inter_rank < inter_comm_size) {
184 // dst = (inter_rank + mask) * num_core;
185 // if (dst < comm_size) {
186 //printf("Node %d send to node %d when mask is %d\n", rank, dst, mask);
187 // Request::send(recv_buf, count, dtype, dst, tag, comm);
195 // INTRA_binomial_bcast
197 int num_core_in_current_smp = num_core;
198 if (inter_rank == (inter_comm_size - 1)) {
199 num_core_in_current_smp = comm_size - (inter_rank * num_core);
201 // printf("Node %d num_core = %d\n",rank, num_core_in_current_smp);
203 while (mask < num_core_in_current_smp) {
204 if (intra_rank & mask) {
205 src = (inter_rank * num_core) + (intra_rank - mask);
206 //printf("Node %d recv from node %d when mask is %d\n", rank, src, mask);
207 Request::recv(recv_buf, count, dtype, src, tag, comm, &status);
214 //printf("My rank = %d my mask = %d\n", rank,mask);
217 dst = (inter_rank * num_core) + (intra_rank + mask);
218 if (dst < comm_size) {
219 //printf("Node %d send to node %d when mask is %d\n", rank, dst, mask);
220 Request::send(recv_buf, count, dtype, dst, tag, comm);
225 smpi_free_tmp_buffer(tmp_buf);