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"
9 /* change number of core per smp-node
10 we assume that number of core per process will be the same for all implementations */
16 This fucntion performs all-reduce operation as follow.
17 1) binomial_tree reduce inside each SMP node
18 2) reduce-scatter -inter between root of each SMP node
19 3) allgather - inter between root of each SMP node
20 4) binomial_tree bcast inside each SMP node
22 int smpi_coll_tuned_allreduce_smp_rsag(void *send_buf, void *recv_buf,
23 int count, MPI_Datatype dtype, MPI_Op op,
28 int tag = COLL_TAG_ALLREDUCE;
31 int num_core = simcall_host_get_core(SIMIX_host_self());
32 // do we use the default one or the number of cores in the platform ?
33 // if the number of cores is one, the platform may be simulated with 1 node = 1 core
34 if (num_core == 1) num_core = NUM_CORE;
36 #ifdef MPICH2_REDUCTION
37 MPI_User_function * uop = MPIR_Op_table[op % 16 - 1];
39 MPI_User_function *uop;
40 struct MPIR_OP *op_ptr;
41 op_ptr = MPIR_ToPointer(op);
45 comm_size = smpi_comm_size(comm);
46 rank = smpi_comm_rank(comm);
48 extent = smpi_datatype_get_extent(dtype);
49 tmp_buf = (void *) xbt_malloc(count * extent);
51 int intra_rank, inter_rank;
52 intra_rank = rank % num_core;
53 inter_rank = rank / num_core;
55 //printf("node %d intra_rank = %d, inter_rank = %d\n", rank, intra_rank, inter_rank);
57 int inter_comm_size = (comm_size + num_core - 1) / num_core;
60 //printf("intra com size = %d\n",num_core);
61 //printf("inter com size = %d\n",inter_comm_size);
65 smpi_mpi_sendrecv(send_buf, count, dtype, rank, tag,
66 recv_buf, count, dtype, rank, tag, comm, &status);
69 // SMP_binomial_reduce
71 while (mask < num_core) {
72 if ((mask & intra_rank) == 0) {
73 src = (inter_rank * num_core) + (intra_rank | mask);
74 // if (src < ((inter_rank + 1) * num_core)) {
75 if (src < comm_size) {
76 smpi_mpi_recv(tmp_buf, count, dtype, src, tag, comm, &status);
77 smpi_op_apply(op, tmp_buf, recv_buf, &count, &dtype);
78 //printf("Node %d recv from node %d when mask is %d\n", rank, src, mask);
82 dst = (inter_rank * num_core) + (intra_rank & (~mask));
83 smpi_mpi_send(recv_buf, count, dtype, dst, tag, comm);
84 //printf("Node %d send to node %d when mask is %d\n", rank, dst, mask);
92 // INTER: reduce-scatter
93 if (intra_rank == 0) {
94 int send_offset, recv_offset;
95 int seg_count = count / inter_comm_size;
96 int to = ((inter_rank + 1) % inter_comm_size) * num_core;
98 ((inter_rank + inter_comm_size - 1) % inter_comm_size) * num_core;
101 //printf("node %d to %d from %d\n",rank,to,from);
103 for (i = 0; i < (inter_comm_size - 1); i++) {
106 ((inter_rank - 1 - i +
107 inter_comm_size) % inter_comm_size) * seg_count * extent;
109 ((inter_rank - 2 - i +
110 inter_comm_size) % inter_comm_size) * seg_count * extent;
112 smpi_mpi_sendrecv((char *) recv_buf + send_offset, seg_count, dtype, to,
113 tag + i, tmp_buf, seg_count, dtype, from, tag + i, comm,
117 smpi_op_apply(op, tmp_buf, (char *) recv_buf + recv_offset, &seg_count,
122 for (i = 0; i < (inter_comm_size - 1); i++) {
126 inter_comm_size) % inter_comm_size) * seg_count * extent;
128 ((inter_rank - 1 - i +
129 inter_comm_size) % inter_comm_size) * seg_count * extent;
131 smpi_mpi_sendrecv((char *) recv_buf + send_offset, seg_count, dtype, to,
132 tag + i, (char *) recv_buf + recv_offset, seg_count, dtype,
133 from, tag + i, comm, &status);
141 // INTER_binomial_reduce
143 // only root node for each SMP
144 if (intra_rank == 0) {
147 while (mask < inter_comm_size) {
148 if ((mask & inter_rank) == 0) {
149 src = (inter_rank | mask) * num_core;
150 if (src < comm_size) {
151 smpi_mpi_recv(tmp_buf, count, dtype, src, tag, comm, &status);
152 (* uop) (tmp_buf, recv_buf, &count, &dtype);
153 //printf("Node %d recv from node %d when mask is %d\n", rank, src, mask);
157 dst = (inter_rank & (~mask)) * num_core;
158 smpi_mpi_send(recv_buf, count, dtype, dst, tag, comm);
159 //printf("Node %d send to node %d when mask is %d\n", rank, dst, mask);
168 // INTER_binomial_bcast
171 if (intra_rank == 0) {
173 while (mask < inter_comm_size) {
174 if (inter_rank & mask) {
175 src = (inter_rank - mask) * num_core;
176 //printf("Node %d recv from node %d when mask is %d\n", rank, src, mask);
177 smpi_mpi_recv(recv_buf, count, dtype, src, tag, comm, &status);
184 //printf("My rank = %d my mask = %d\n", rank,mask);
187 if (inter_rank < inter_comm_size) {
188 dst = (inter_rank + mask) * num_core;
189 if (dst < comm_size) {
190 //printf("Node %d send to node %d when mask is %d\n", rank, dst, mask);
191 smpi_mpi_send(recv_buf, count, dtype, dst, tag, comm);
200 // INTRA_binomial_bcast
202 int num_core_in_current_smp = num_core;
203 if (inter_rank == (inter_comm_size - 1)) {
204 num_core_in_current_smp = comm_size - (inter_rank * num_core);
206 // printf("Node %d num_core = %d\n",rank, num_core_in_current_smp);
208 while (mask < num_core_in_current_smp) {
209 if (intra_rank & mask) {
210 src = (inter_rank * num_core) + (intra_rank - mask);
211 //printf("Node %d recv from node %d when mask is %d\n", rank, src, mask);
212 smpi_mpi_recv(recv_buf, count, dtype, src, tag, comm, &status);
219 //printf("My rank = %d my mask = %d\n", rank,mask);
222 dst = (inter_rank * num_core) + (intra_rank + mask);
223 if (dst < comm_size) {
224 //printf("Node %d send to node %d when mask is %d\n", rank, dst, mask);
225 smpi_mpi_send(recv_buf, count, dtype, dst, tag, comm);