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>
13 /* change number of core per smp-node
14 we assume that number of core per process will be the same for all implementations */
20 This fucntion performs all-reduce operation as follow.
21 1) binomial_tree reduce inside each SMP node
22 2) reduce-scatter -inter between root of each SMP node
23 3) allgather - inter between root of each SMP node
24 4) binomial_tree bcast inside each SMP node
26 int smpi_coll_tuned_allreduce_smp_rsag_rab(void *sbuf, void *rbuf, int count,
27 MPI_Datatype dtype, MPI_Op op,
32 int tag = COLL_TAG_ALLREDUCE;
35 int num_core = simcall_host_get_core(SIMIX_host_self());
36 // do we use the default one or the number of cores in the platform ?
37 // if the number of cores is one, the platform may be simulated with 1 node = 1 core
38 if (num_core == 1) num_core = NUM_CORE;
40 comm_size = smpi_comm_size(comm);
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 ! ");
45 rank = smpi_comm_rank(comm);
47 extent = smpi_datatype_get_extent(dtype);
48 tmp_buf = (void *) xbt_malloc(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 smpi_mpi_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 smpi_mpi_recv(tmp_buf, count, dtype, src, tag, comm, &status);
69 smpi_op_apply(op, 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 smpi_mpi_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);
95 curr_count = count / 2;
101 while (mask < (comm_size / num_core)) {
102 dst = inter_rank ^ mask;
105 send_base_offset = base_offset;
108 if (inter_rank & mask) {
109 recv_base_offset = base_offset + curr_count;
110 send_base_offset = base_offset;
111 base_offset = recv_base_offset;
115 recv_base_offset = base_offset;
116 send_base_offset = base_offset + curr_count;
118 send_offset = send_base_offset * extent;
119 recv_offset = recv_base_offset * extent;
122 // 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);
124 smpi_mpi_sendrecv((char *)rbuf + send_offset, curr_count, dtype, (dst * num_core), tag,
125 tmp_buf, curr_count, dtype, (dst * num_core), tag,
128 smpi_op_apply(op, tmp_buf, (char *)rbuf + recv_offset, &curr_count, &dtype);
138 int size = (comm_size / num_core) / 2;
141 while (mask < (comm_size / num_core)) {
142 if (inter_rank & mask) {
154 // destination pair for both send and recv
155 dst = inter_rank ^ mask;
158 send_base_offset = base_offset;
159 if (inter_rank & mask) {
160 recv_base_offset = base_offset - i;
163 recv_base_offset = base_offset + i;
165 send_offset = send_base_offset * recv_chunk;
166 recv_offset = recv_base_offset * recv_chunk;
169 //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);
171 smpi_mpi_sendrecv((char *)rbuf + send_offset, curr_count, dtype, (dst * num_core), tag,
172 (char *)rbuf + recv_offset, curr_count, dtype, (dst * num_core), tag,
185 // intra SMP binomial bcast
187 int num_core_in_current_smp = num_core;
188 if (inter_rank == (inter_comm_size - 1)) {
189 num_core_in_current_smp = comm_size - (inter_rank * num_core);
191 // printf("Node %d num_core = %d\n",rank, num_core_in_current_smp);
193 while (mask < num_core_in_current_smp) {
194 if (intra_rank & mask) {
195 src = (inter_rank * num_core) + (intra_rank - mask);
196 //printf("Node %d recv from node %d when mask is %d\n", rank, src, mask);
197 smpi_mpi_recv(rbuf, count, dtype, src, tag, comm, &status);
204 //printf("My rank = %d my mask = %d\n", rank,mask);
207 dst = (inter_rank * num_core) + (intra_rank + mask);
208 if (dst < comm_size) {
209 //printf("Node %d send to node %d when mask is %d\n", rank, dst, mask);
210 smpi_mpi_send(rbuf, count, dtype, dst, tag, comm);