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. */
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
21 int smpi_coll_tuned_allreduce_smp_rsag_rab(void *sbuf, void *rbuf, int count,
22 MPI_Datatype dtype, MPI_Op op,
27 int tag = COLL_TAG_ALLREDUCE;
30 if(smpi_comm_get_leaders_comm(comm)==MPI_COMM_NULL){
31 smpi_comm_init_smp(comm);
34 if (smpi_comm_is_uniform(comm)){
35 num_core = smpi_comm_size(smpi_comm_get_intra_comm(comm));
38 comm_size = smpi_comm_size(comm);
40 if((comm_size&(comm_size-1)))
41 THROWF(arg_error,0, "allreduce smp rsag rab algorithm can't be used with non power of two number of processes ! ");
43 rank = smpi_comm_rank(comm);
45 extent = smpi_datatype_get_extent(dtype);
46 tmp_buf = (void *) smpi_get_tmp_sendbuffer(count * extent);
48 int intra_rank, inter_rank;
49 intra_rank = rank % num_core;
50 inter_rank = rank / num_core;
52 //printf("node %d intra_rank = %d, inter_rank = %d\n", rank, intra_rank, inter_rank);
54 int inter_comm_size = (comm_size + num_core - 1) / num_core;
56 smpi_mpi_sendrecv(sbuf, count, dtype, rank, tag,
57 rbuf, count, dtype, rank, tag, comm, &status);
59 // SMP_binomial_reduce
61 while (mask < num_core) {
62 if ((mask & intra_rank) == 0) {
63 src = (inter_rank * num_core) + (intra_rank | mask);
64 // if (src < ((inter_rank + 1) * num_core)) {
65 if (src < comm_size) {
66 smpi_mpi_recv(tmp_buf, count, dtype, src, tag, comm, &status);
67 smpi_op_apply(op, tmp_buf, rbuf, &count, &dtype);
68 //printf("Node %d recv from node %d when mask is %d\n", rank, src, mask);
72 dst = (inter_rank * num_core) + (intra_rank & (~mask));
73 smpi_mpi_send(rbuf, count, dtype, dst, tag, comm);
74 //printf("Node %d send to node %d when mask is %d\n", rank, dst, mask);
81 // INTER: reduce-scatter
82 if (intra_rank == 0) {
84 int dst, base_offset, send_base_offset, recv_base_offset, recv_chunk;
85 int curr_count, i, recv_offset, send_offset;
89 recv_chunk = extent * count / (comm_size / num_core);
92 curr_count = count / 2;
96 while (mask < (comm_size / num_core)) {
97 dst = inter_rank ^ mask;
101 if (inter_rank & mask) {
102 recv_base_offset = base_offset + curr_count;
103 send_base_offset = base_offset;
104 base_offset = recv_base_offset;
108 recv_base_offset = base_offset;
109 send_base_offset = base_offset + curr_count;
111 send_offset = send_base_offset * extent;
112 recv_offset = recv_base_offset * extent;
115 // 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);
117 smpi_mpi_sendrecv((char *)rbuf + send_offset, curr_count, dtype, (dst * num_core), tag,
118 tmp_buf, curr_count, dtype, (dst * num_core), tag,
121 smpi_op_apply(op, tmp_buf, (char *)rbuf + recv_offset, &curr_count, &dtype);
131 int size = (comm_size / num_core) / 2;
134 while (mask < (comm_size / num_core)) {
135 if (inter_rank & mask) {
147 // destination pair for both send and recv
148 dst = inter_rank ^ mask;
151 send_base_offset = base_offset;
152 if (inter_rank & mask) {
153 recv_base_offset = base_offset - i;
156 recv_base_offset = base_offset + i;
158 send_offset = send_base_offset * recv_chunk;
159 recv_offset = recv_base_offset * recv_chunk;
162 //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);
164 smpi_mpi_sendrecv((char *)rbuf + send_offset, curr_count, dtype, (dst * num_core), tag,
165 (char *)rbuf + recv_offset, curr_count, dtype, (dst * num_core), tag,
178 // intra SMP binomial bcast
180 int num_core_in_current_smp = num_core;
181 if (inter_rank == (inter_comm_size - 1)) {
182 num_core_in_current_smp = comm_size - (inter_rank * num_core);
184 // printf("Node %d num_core = %d\n",rank, num_core_in_current_smp);
186 while (mask < num_core_in_current_smp) {
187 if (intra_rank & mask) {
188 src = (inter_rank * num_core) + (intra_rank - mask);
189 //printf("Node %d recv from node %d when mask is %d\n", rank, src, mask);
190 smpi_mpi_recv(rbuf, count, dtype, src, tag, comm, &status);
197 //printf("My rank = %d my mask = %d\n", rank,mask);
200 dst = (inter_rank * num_core) + (intra_rank + mask);
201 if (dst < comm_size) {
202 //printf("Node %d send to node %d when mask is %d\n", rank, dst, mask);
203 smpi_mpi_send(rbuf, count, dtype, dst, tag, comm);
209 smpi_free_tmp_buffer(tmp_buf);