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"
8 //#include <star-reduction.c>
10 /* change number of core per smp-node
11 we assume that number of core per process will be the same for all implementations */
17 This fucntion performs all-reduce operation as follow.
18 1) binomial_tree reduce inside each SMP node
19 2) reduce-scatter -inter between root of each SMP node
20 3) allgather - inter between root of each SMP node
21 4) binomial_tree bcast inside each SMP node
23 int smpi_coll_tuned_allreduce_smp_rsag_lr(void *send_buf, void *recv_buf,
24 int count, MPI_Datatype dtype,
25 MPI_Op op, MPI_Comm comm)
29 int tag = COLL_TAG_ALLREDUCE;
32 int num_core = simcall_host_get_core(SIMIX_host_self());
33 // do we use the default one or the number of cores in the platform ?
34 // if the number of cores is one, the platform may be simulated with 1 node = 1 core
35 if (num_core == 1) num_core = NUM_CORE;
37 #ifdef MPICH2_REDUCTION
38 MPI_User_function * uop = MPIR_Op_table[op % 16 - 1];
40 MPI_User_function *uop;
41 struct MPIR_OP *op_ptr;
42 op_ptr = MPIR_ToPointer(op);
46 comm_size = smpi_comm_size(comm);
47 rank = smpi_comm_rank(comm);
49 extent = smpi_datatype_get_extent(dtype);
50 tmp_buf = (void *) xbt_malloc(count * extent);
52 int intra_rank, inter_rank;
53 intra_rank = rank % num_core;
54 inter_rank = rank / num_core;
56 //printf("node %d intra_rank = %d, inter_rank = %d\n", rank, intra_rank, inter_rank);
58 int inter_comm_size = (comm_size + num_core - 1) / num_core;
61 //printf("intra com size = %d\n",num_core);
62 //printf("inter com size = %d\n",inter_comm_size);
66 smpi_mpi_sendrecv(send_buf, count, dtype, rank, tag,
67 recv_buf, count, dtype, rank, tag, comm, &status);
70 // SMP_binomial_reduce
72 while (mask < num_core) {
73 if ((mask & intra_rank) == 0) {
74 src = (inter_rank * num_core) + (intra_rank | mask);
75 // if (src < ((inter_rank + 1) * num_core)) {
76 if (src < comm_size) {
77 smpi_mpi_recv(tmp_buf, count, dtype, src, tag, comm, &status);
78 smpi_op_apply(op, tmp_buf, recv_buf, &count, &dtype);
79 //printf("Node %d recv from node %d when mask is %d\n", rank, src, mask);
83 dst = (inter_rank * num_core) + (intra_rank & (~mask));
84 smpi_mpi_send(recv_buf, count, dtype, dst, tag, comm);
85 //printf("Node %d send to node %d when mask is %d\n", rank, dst, mask);
93 // INTER: reduce-scatter
94 if (intra_rank == 0) {
95 int send_offset, recv_offset;
96 int send_count, recv_count;
97 int curr_size = count / inter_comm_size;
98 int curr_remainder = count % inter_comm_size;
100 int to = ((inter_rank + 1) % inter_comm_size) * num_core;
102 ((inter_rank + inter_comm_size - 1) % inter_comm_size) * num_core;
105 //printf("node %d to %d from %d\n",rank,to,from);
107 /* last segment may have a larger size since it also include the remainder */
108 int last_segment_ptr =
109 (inter_comm_size - 1) * (count / inter_comm_size) * extent;
111 for (i = 0; i < (inter_comm_size - 1); i++) {
114 ((inter_rank - 1 - i +
115 inter_comm_size) % inter_comm_size) * curr_size * extent;
117 ((inter_rank - 2 - i +
118 inter_comm_size) % inter_comm_size) * curr_size * extent;
121 if (send_offset != last_segment_ptr)
122 send_count = curr_size;
124 send_count = curr_size + curr_remainder;
126 if (recv_offset != last_segment_ptr)
127 recv_count = curr_size;
129 recv_count = curr_size + curr_remainder;
131 smpi_mpi_sendrecv((char *) recv_buf + send_offset, send_count, dtype, to,
132 tag + i, tmp_buf, recv_count, dtype, from, tag + i, comm,
136 smpi_op_apply(op, tmp_buf, (char *) recv_buf + recv_offset, &recv_count,
141 for (i = 0; i < (inter_comm_size - 1); i++) {
145 inter_comm_size) % inter_comm_size) * curr_size * extent;
147 ((inter_rank - 1 - i +
148 inter_comm_size) % inter_comm_size) * curr_size * extent;
151 if (send_offset != last_segment_ptr)
152 send_count = curr_size;
154 send_count = curr_size + curr_remainder;
156 if (recv_offset != last_segment_ptr)
157 recv_count = curr_size;
159 recv_count = curr_size + curr_remainder;
161 smpi_mpi_sendrecv((char *) recv_buf + send_offset, send_count, dtype, to,
162 tag + i, (char *) recv_buf + recv_offset, recv_count, dtype,
163 from, tag + i, comm, &status);
171 // INTER_binomial_reduce
173 // only root node for each SMP
174 if (intra_rank == 0) {
177 while (mask < inter_comm_size) {
178 if ((mask & inter_rank) == 0) {
179 src = (inter_rank | mask) * num_core;
180 if (src < comm_size) {
181 smpi_mpi_recv(tmp_buf, count, dtype, src, tag, comm, &status);
182 (* uop) (tmp_buf, recv_buf, &count, &dtype);
183 //printf("Node %d recv from node %d when mask is %d\n", rank, src, mask);
187 dst = (inter_rank & (~mask)) * num_core;
188 smpi_mpi_send(recv_buf, count, dtype, dst, tag, comm);
189 //printf("Node %d send to node %d when mask is %d\n", rank, dst, mask);
198 // INTER_binomial_bcast
201 if (intra_rank == 0) {
203 while (mask < inter_comm_size) {
204 if (inter_rank & mask) {
205 src = (inter_rank - mask) * num_core;
206 //printf("Node %d recv from node %d when mask is %d\n", rank, src, mask);
207 smpi_mpi_recv(recv_buf, count, dtype, src, tag, comm, &status);
214 //printf("My rank = %d my mask = %d\n", rank,mask);
217 if (inter_rank < inter_comm_size) {
218 dst = (inter_rank + mask) * num_core;
219 if (dst < comm_size) {
220 //printf("Node %d send to node %d when mask is %d\n", rank, dst, mask);
221 smpi_mpi_send(recv_buf, count, dtype, dst, tag, comm);
230 // INTRA_binomial_bcast
232 int num_core_in_current_smp = num_core;
233 if (inter_rank == (inter_comm_size - 1)) {
234 num_core_in_current_smp = comm_size - (inter_rank * num_core);
236 // printf("Node %d num_core = %d\n",rank, num_core_in_current_smp);
238 while (mask < num_core_in_current_smp) {
239 if (intra_rank & mask) {
240 src = (inter_rank * num_core) + (intra_rank - mask);
241 //printf("Node %d recv from node %d when mask is %d\n", rank, src, mask);
242 smpi_mpi_recv(recv_buf, count, dtype, src, tag, comm, &status);
249 //printf("My rank = %d my mask = %d\n", rank,mask);
252 dst = (inter_rank * num_core) + (intra_rank + mask);
253 if (dst < comm_size) {
254 //printf("Node %d send to node %d when mask is %d\n", rank, dst, mask);
255 smpi_mpi_send(recv_buf, count, dtype, dst, tag, comm);