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
16 int smpi_coll_tuned_allreduce_smp_rsag(void *send_buf, void *recv_buf,
17 int count, MPI_Datatype dtype, MPI_Op op,
22 int tag = COLL_TAG_ALLREDUCE;
25 if(smpi_comm_get_leaders_comm(comm)==MPI_COMM_NULL){
26 smpi_comm_init_smp(comm);
29 if (smpi_comm_is_uniform(comm)){
30 num_core = smpi_comm_size(smpi_comm_get_intra_comm(comm));
33 #ifdef MPICH2_REDUCTION
34 MPI_User_function * uop = MPIR_Op_table[op % 16 - 1];
36 MPI_User_function *uop;
37 struct MPIR_OP *op_ptr;
38 op_ptr = MPIR_ToPointer(op);
42 comm_size = smpi_comm_size(comm);
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;
57 //printf("intra com size = %d\n",num_core);
58 //printf("inter com size = %d\n",inter_comm_size);
62 smpi_mpi_sendrecv(send_buf, count, dtype, rank, tag,
63 recv_buf, count, dtype, rank, tag, comm, &status);
66 // SMP_binomial_reduce
68 while (mask < num_core) {
69 if ((mask & intra_rank) == 0) {
70 src = (inter_rank * num_core) + (intra_rank | mask);
71 // if (src < ((inter_rank + 1) * num_core)) {
72 if (src < comm_size) {
73 smpi_mpi_recv(tmp_buf, count, dtype, src, tag, comm, &status);
74 smpi_op_apply(op, tmp_buf, recv_buf, &count, &dtype);
75 //printf("Node %d recv from node %d when mask is %d\n", rank, src, mask);
79 dst = (inter_rank * num_core) + (intra_rank & (~mask));
80 smpi_mpi_send(recv_buf, count, dtype, dst, tag, comm);
81 //printf("Node %d send to node %d when mask is %d\n", rank, dst, mask);
89 // INTER: reduce-scatter
90 if (intra_rank == 0) {
91 int send_offset, recv_offset;
92 int seg_count = count / inter_comm_size;
93 int to = ((inter_rank + 1) % inter_comm_size) * num_core;
95 ((inter_rank + inter_comm_size - 1) % inter_comm_size) * num_core;
98 //printf("node %d to %d from %d\n",rank,to,from);
100 for (i = 0; i < (inter_comm_size - 1); i++) {
103 ((inter_rank - 1 - i +
104 inter_comm_size) % inter_comm_size) * seg_count * extent;
106 ((inter_rank - 2 - i +
107 inter_comm_size) % inter_comm_size) * seg_count * extent;
109 smpi_mpi_sendrecv((char *) recv_buf + send_offset, seg_count, dtype, to,
110 tag + i, tmp_buf, seg_count, dtype, from, tag + i, comm,
114 smpi_op_apply(op, tmp_buf, (char *) recv_buf + recv_offset, &seg_count,
119 for (i = 0; i < (inter_comm_size - 1); i++) {
123 inter_comm_size) % inter_comm_size) * seg_count * extent;
125 ((inter_rank - 1 - i +
126 inter_comm_size) % inter_comm_size) * seg_count * extent;
128 smpi_mpi_sendrecv((char *) recv_buf + send_offset, seg_count, dtype, to,
129 tag + i, (char *) recv_buf + recv_offset, seg_count, dtype,
130 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 smpi_mpi_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 smpi_mpi_send(recv_buf, count, dtype, dst, tag, comm);
156 //printf("Node %d send to node %d when mask is %d\n", rank, dst, mask);
165 // INTER_binomial_bcast
168 if (intra_rank == 0) {
170 while (mask < inter_comm_size) {
171 if (inter_rank & mask) {
172 src = (inter_rank - mask) * num_core;
173 //printf("Node %d recv from node %d when mask is %d\n", rank, src, mask);
174 smpi_mpi_recv(recv_buf, count, dtype, src, tag, comm, &status);
181 //printf("My rank = %d my mask = %d\n", rank,mask);
184 if (inter_rank < inter_comm_size) {
185 dst = (inter_rank + mask) * num_core;
186 if (dst < comm_size) {
187 //printf("Node %d send to node %d when mask is %d\n", rank, dst, mask);
188 smpi_mpi_send(recv_buf, count, dtype, dst, tag, comm);
197 // INTRA_binomial_bcast
199 int num_core_in_current_smp = num_core;
200 if (inter_rank == (inter_comm_size - 1)) {
201 num_core_in_current_smp = comm_size - (inter_rank * num_core);
203 // printf("Node %d num_core = %d\n",rank, num_core_in_current_smp);
205 while (mask < num_core_in_current_smp) {
206 if (intra_rank & mask) {
207 src = (inter_rank * num_core) + (intra_rank - mask);
208 //printf("Node %d recv from node %d when mask is %d\n", rank, src, mask);
209 smpi_mpi_recv(recv_buf, count, dtype, src, tag, comm, &status);
216 //printf("My rank = %d my mask = %d\n", rank,mask);
219 dst = (inter_rank * num_core) + (intra_rank + mask);
220 if (dst < comm_size) {
221 //printf("Node %d send to node %d when mask is %d\n", rank, dst, mask);
222 smpi_mpi_send(recv_buf, count, dtype, dst, tag, comm);
228 smpi_free_tmp_buffer(tmp_buf);