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. */
8 * implemented by Pitch Patarasuk, 07/01/2007
10 #include "../colls_private.h"
11 //#include <star-reduction.c>
15 This fucntion performs all-reduce operation as follow.
16 1) binomial_tree reduce inside each SMP node
17 2) reduce-scatter -inter between root of each SMP node
18 3) allgather - inter between root of each SMP node
19 4) binomial_tree bcast inside each SMP node
23 int Coll_allreduce_smp_rsag_rab::allreduce(void *sbuf, void *rbuf, int count,
24 MPI_Datatype dtype, MPI_Op op,
29 int tag = COLL_TAG_ALLREDUCE;
32 if(comm->get_leaders_comm()==MPI_COMM_NULL){
36 if (comm->is_uniform()){
37 num_core = comm->get_intra_comm()->size();
40 comm_size = comm->size();
42 if((comm_size&(comm_size-1)))
43 THROWF(arg_error,0, "allreduce smp rsag rab algorithm can't be used with non power of two number of processes ! ");
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;
58 Request::sendrecv(sbuf, count, dtype, rank, tag,
59 rbuf, count, dtype, rank, tag, comm, &status);
61 // SMP_binomial_reduce
63 while (mask < num_core) {
64 if ((mask & intra_rank) == 0) {
65 src = (inter_rank * num_core) + (intra_rank | mask);
66 // if (src < ((inter_rank + 1) * num_core)) {
67 if (src < comm_size) {
68 Request::recv(tmp_buf, count, dtype, src, tag, comm, &status);
69 if(op!=MPI_OP_NULL) op->apply( tmp_buf, rbuf, &count, dtype);
70 //printf("Node %d recv from node %d when mask is %d\n", rank, src, mask);
74 dst = (inter_rank * num_core) + (intra_rank & (~mask));
75 Request::send(rbuf, count, dtype, dst, tag, comm);
76 //printf("Node %d send to node %d when mask is %d\n", rank, dst, mask);
83 // INTER: reduce-scatter
84 if (intra_rank == 0) {
86 int dst, base_offset, send_base_offset, recv_base_offset, recv_chunk;
87 int curr_count, i, recv_offset, send_offset;
91 recv_chunk = extent * count / (comm_size / num_core);
94 curr_count = count / 2;
98 while (mask < (comm_size / num_core)) {
99 dst = inter_rank ^ mask;
103 if (inter_rank & mask) {
104 recv_base_offset = base_offset + curr_count;
105 send_base_offset = base_offset;
106 base_offset = recv_base_offset;
110 recv_base_offset = base_offset;
111 send_base_offset = base_offset + curr_count;
113 send_offset = send_base_offset * extent;
114 recv_offset = recv_base_offset * extent;
117 // printf("node %d send to %d in phase %d s_offset = %d r_offset = %d count = %d\n",rank,dst,phase, send_offset, recv_offset, curr_count);
119 Request::sendrecv((char *)rbuf + send_offset, curr_count, dtype, (dst * num_core), tag,
120 tmp_buf, curr_count, dtype, (dst * num_core), tag,
123 if(op!=MPI_OP_NULL) op->apply( tmp_buf, (char *)rbuf + recv_offset, &curr_count, dtype);
133 int size = (comm_size / num_core) / 2;
136 while (mask < (comm_size / num_core)) {
137 if (inter_rank & mask) {
149 // destination pair for both send and recv
150 dst = inter_rank ^ mask;
153 send_base_offset = base_offset;
154 if (inter_rank & mask) {
155 recv_base_offset = base_offset - i;
158 recv_base_offset = base_offset + i;
160 send_offset = send_base_offset * recv_chunk;
161 recv_offset = recv_base_offset * recv_chunk;
164 //printf("node %d send to %d in phase %d s_offset = %d r_offset = %d count = %d\n",rank,dst,phase, send_offset, recv_offset, curr_count);
166 Request::sendrecv((char *)rbuf + send_offset, curr_count, dtype, (dst * num_core), tag,
167 (char *)rbuf + recv_offset, curr_count, dtype, (dst * num_core), tag,
180 // intra SMP binomial bcast
182 int num_core_in_current_smp = num_core;
183 if (inter_rank == (inter_comm_size - 1)) {
184 num_core_in_current_smp = comm_size - (inter_rank * num_core);
186 // printf("Node %d num_core = %d\n",rank, num_core_in_current_smp);
188 while (mask < num_core_in_current_smp) {
189 if (intra_rank & mask) {
190 src = (inter_rank * num_core) + (intra_rank - mask);
191 //printf("Node %d recv from node %d when mask is %d\n", rank, src, mask);
192 Request::recv(rbuf, count, dtype, src, tag, comm, &status);
199 //printf("My rank = %d my mask = %d\n", rank,mask);
202 dst = (inter_rank * num_core) + (intra_rank + mask);
203 if (dst < comm_size) {
204 //printf("Node %d send to node %d when mask is %d\n", rank, dst, mask);
205 Request::send(rbuf, count, dtype, dst, tag, comm);
211 smpi_free_tmp_buffer(tmp_buf);