1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336
|
/* -*- Mode: C; c-basic-offset:4 ; indent-tabs-mode:nil -*- */
/*
* Copyright (c) 2008 The Trustees of Indiana University and Indiana
* University Research and Technology
* Corporation. All rights reserved.
* Copyright (c) 2009 Cisco Systems, Inc. All rights reserved.
* Copyright (c) 2011-2017 The University of Tennessee and The University
* of Tennessee Research Foundation. All rights
* reserved.
* Copyright (c) 2011-2013 Inria. All rights reserved.
* Copyright (c) 2011-2013 Université Bordeaux 1
* Copyright (c) 2014-2015 Research Organization for Information Science
* and Technology (RIST). All rights reserved.
* Copyright (c) 2016-2017 IBM Corporation. All rights reserved.
* Copyright (c) 2018 Triad National Security, LLC. All rights
* reserved.
*/
#include "ompi_config.h"
#include "ompi/communicator/communicator.h"
#include "ompi/info/info.h"
#include "ompi/mca/topo/base/base.h"
#include "ompi/datatype/ompi_datatype.h"
#include "ompi/mca/pml/pml.h"
#include "ompi/request/request.h"
#define IN_INDEX 0
#define OUT_INDEX 1
#define MCA_TOPO_BASE_TAG_DIST_EDGE_IN -50
#define MCA_TOPO_BASE_TAG_DIST_EDGE_OUT -51
typedef struct _dist_graph_elem {
int in;
int out;
} mca_topo_base_dist_graph_elem_t;
int mca_topo_base_dist_graph_distribute(mca_topo_base_module_t* module,
ompi_communicator_t *comm,
int n, const int nodes[],
const int degrees[], const int targets[],
const int weights[],
mca_topo_base_comm_dist_graph_2_2_0_t** ptopo)
{
int i, j, err, count, left_over, pending_reqs, current_pos, index, csize;
int *rin = NULL, *rout, *temp = NULL;
mca_topo_base_dist_graph_elem_t *pos, *cnt, *idx;
size_t int_size, how_much;
ompi_status_public_t status;
ompi_request_t **reqs = NULL;
mca_topo_base_comm_dist_graph_2_2_0_t* topo=NULL;
ompi_datatype_type_size( (ompi_datatype_t*)&ompi_mpi_int, &int_size);
csize = ompi_comm_size(comm);
/**
* We compress the counts: for each peer we maintain an in and an out.
* In addition we compute 3 arrays (that are allocated in one go):
* - cnt: the number of elements for a peer
* - pos: the position of the first element for a peer
* - idx: temporary indexes and message count after the reduce.
*/
cnt = (mca_topo_base_dist_graph_elem_t*)calloc(3 * csize, sizeof(mca_topo_base_dist_graph_elem_t));
if( NULL == cnt ) {
err = OMPI_ERR_OUT_OF_RESOURCE;
goto bail_out;
}
pos = cnt + csize;
idx = pos + csize;
for( index = i = 0; i < n; i++ ) {
cnt[nodes[i]].out += degrees[i];
for( j = 0; j < degrees[i]; ++j ) {
cnt[targets[index]].in++;
index++;
}
}
/**
* Prepare the positions array. The ith element is the corresponding
* starting position of the ith neighbor in the global array.
*/
pos[0].in = 0;
pos[0].out = 0;
for( i = 0; i < (csize - 1); i++ ) {
pos[i + 1].in = pos[i].in + cnt[i].in;
pos[i + 1].out = pos[i].out + cnt[i].out;
}
rin = (int*)calloc(2 * (pos[csize - 1].in + cnt[csize - 1].in +
pos[csize - 1].out + cnt[csize - 1].out), sizeof(int));
if( NULL == rin ) {
err = OMPI_ERR_OUT_OF_RESOURCE;
goto bail_out;
}
rout = &rin[2 * (pos[csize - 1].in + cnt[csize - 1].in)];
for( index = i = 0; i < n; ++i ) { /* for each of the nodes */
for( j = 0; j < degrees[i]; ++j ) { /* for each node's degree */
int position = pos[nodes[i]].out + idx[nodes[i]].out;
if( MPI_UNWEIGHTED != weights ) {
position *= 2;
rout[position + 1] = weights[index];
}
rout[position + 0] = targets[index];
idx[nodes[i]].out++;
position = pos[targets[index]].in + idx[targets[index]].in;
if( MPI_UNWEIGHTED != weights ) {
position *= 2;
rin[position + 1] = weights[index];
}
rin[position + 0] = nodes[i];
idx[targets[index]].in++;
index++;
}
}
err = comm->c_coll->coll_reduce_scatter_block( MPI_IN_PLACE, idx, 2,
(ompi_datatype_t*)&ompi_mpi_int, MPI_SUM, comm,
comm->c_coll->coll_reduce_scatter_block_module);
/**
* At this point in the indexes array we have:
* - idx[0].in total number of IN edges
* - idx[0].out total number of OUT edges
*/
topo = OBJ_NEW(mca_topo_base_comm_dist_graph_2_2_0_t);
if( NULL == topo ) {
err = OMPI_ERR_OUT_OF_RESOURCE;
goto bail_out;
}
topo->indegree = idx[0].in;
topo->outdegree = idx[0].out;
topo->weighted = (weights != MPI_UNWEIGHTED);
if (topo->indegree > 0) {
topo->in = (int*)malloc(sizeof(int) * topo->indegree);
if (NULL == topo->in) {
err = OMPI_ERR_OUT_OF_RESOURCE;
goto bail_out;
}
if (MPI_UNWEIGHTED != weights) {
topo->inw = (int*)malloc(sizeof(int) * topo->indegree);
if (NULL == topo->inw) {
err = OMPI_ERR_OUT_OF_RESOURCE;
goto bail_out;
}
}
}
if (topo->outdegree > 0) {
topo->out = (int*)malloc(sizeof(int) * topo->outdegree);
if (NULL == topo->out) {
err = OMPI_ERR_OUT_OF_RESOURCE;
goto bail_out;
}
if (MPI_UNWEIGHTED != weights) {
topo->outw = (int*)malloc(sizeof(int) * topo->outdegree);
if (NULL == topo->outw) {
err = OMPI_ERR_OUT_OF_RESOURCE;
goto bail_out;
}
}
}
reqs = (ompi_request_t**)malloc(sizeof(ompi_request_t*) * 2 * csize);
for (pending_reqs = i = 0; i < csize; ++i) {
int position;
if( 0 != (count = cnt[i].in) ) {
position = pos[i].in;
if (MPI_UNWEIGHTED != weights) {
count *= 2; /* don't forget the weights */
position *= 2;
}
err = MCA_PML_CALL(isend( &rin[position], count, (ompi_datatype_t*)&ompi_mpi_int,
i, MCA_TOPO_BASE_TAG_DIST_EDGE_IN, MCA_PML_BASE_SEND_STANDARD,
comm, &reqs[pending_reqs]));
pending_reqs++;
}
if( 0 != (count = cnt[i].out) ) {
position = pos[i].out;
if (MPI_UNWEIGHTED != weights) {
count *= 2; /* don't forget the weights */
position *= 2;
}
err = MCA_PML_CALL(isend(&rout[position], count, (ompi_datatype_t*)&ompi_mpi_int,
i, MCA_TOPO_BASE_TAG_DIST_EDGE_OUT, MCA_PML_BASE_SEND_STANDARD,
comm, &reqs[pending_reqs]));
pending_reqs++;
}
}
/**
* Now let's receive the input edges in a temporary array
* and then move them to their corresponding place.
*/
count = topo->indegree;
temp = topo->in;
if (MPI_UNWEIGHTED != weights) {
count *= 2; /* don't forget the weights */
if (count > 0) {
/* Allocate an array big enough to hold the edges and
their weights */
temp = (int*)malloc(count*sizeof(int));
if (NULL == temp) {
err = OMPI_ERR_OUT_OF_RESOURCE;
goto bail_out;
}
}
}
for( left_over = count, current_pos = i = 0; left_over > 0; i++ ) {
MCA_PML_CALL(recv( &temp[count - left_over], left_over, (ompi_datatype_t*)&ompi_mpi_int, /* keep receiving in the same buffer */
MPI_ANY_SOURCE, MCA_TOPO_BASE_TAG_DIST_EDGE_IN,
comm, &status ));
how_much = status._ucount / int_size;
if (MPI_UNWEIGHTED != weights) {
for( j = 0; j < ((int)how_much >> 1); j++, current_pos++ ) {
topo->in[current_pos] = temp[2 * j + 0 + (count - left_over)];
topo->inw[current_pos] = temp[2 * j + 1 + (count - left_over)];
}
}
left_over -= how_much;
}
if (MPI_UNWEIGHTED != weights) {
free(temp);
}
/**
* Now let's receive the output edges in a temporary array
* and then move them to their corresponding place.
*/
count = topo->outdegree;
temp = topo->out;
if (MPI_UNWEIGHTED != weights) {
count *= 2; /* don't forget the weights */
if (count > 0) {
/* Allocate an array big enough to hold the edges and
their weights */
temp = (int*)malloc(count*sizeof(int));
if (NULL == temp) {
err = OMPI_ERR_OUT_OF_RESOURCE;
goto bail_out;
}
}
}
for( left_over = count, current_pos = i = 0; left_over > 0; i++ ) {
MCA_PML_CALL(recv( &temp[count - left_over], left_over, (ompi_datatype_t*)&ompi_mpi_int, /* keep receiving in the same buffer */
MPI_ANY_SOURCE, MCA_TOPO_BASE_TAG_DIST_EDGE_OUT,
comm, &status ));
how_much = status._ucount / int_size;
if (MPI_UNWEIGHTED != weights) {
for( j = 0; j < ((int)how_much >> 1); j++, current_pos++ ) {
topo->out[current_pos] = temp[2 * j + 0 + (count - left_over)];
topo->outw[current_pos] = temp[2 * j + 1 + (count - left_over)];
}
}
left_over -= how_much;
}
if (MPI_UNWEIGHTED != weights) {
free(temp);
}
err = ompi_request_wait_all(pending_reqs, reqs, MPI_STATUSES_IGNORE);
*ptopo = topo;
topo = NULL; /* don't free it below */
bail_out:
if( NULL != reqs ) {
free(reqs);
}
if( NULL != rin ) {
free(rin);
}
if( NULL != cnt ) {
free(cnt);
}
if( NULL != topo ) {
OBJ_RELEASE(topo);
}
return err;
}
int mca_topo_base_dist_graph_create(mca_topo_base_module_t* module,
ompi_communicator_t *comm_old,
int n, const int nodes[],
const int degrees[], const int targets[],
const int weights[],
opal_info_t *info, int reorder,
ompi_communicator_t **newcomm)
{
int err;
if (OMPI_SUCCESS != (err = ompi_comm_dup_with_info (comm_old, info, newcomm))) {
OBJ_RELEASE(module);
return err;
}
assert(NULL == (*newcomm)->c_topo);
(*newcomm)->c_topo = module;
(*newcomm)->c_topo->reorder = reorder;
(*newcomm)->c_flags |= OMPI_COMM_DIST_GRAPH;
err = mca_topo_base_dist_graph_distribute(module,
*newcomm,
n, nodes,
degrees, targets,
weights,
&((*newcomm)->c_topo->mtc.dist_graph));
if( OMPI_SUCCESS != err ) {
ompi_comm_free(newcomm);
}
return err;
}
static void mca_topo_base_comm_dist_graph_2_2_0_construct(mca_topo_base_comm_dist_graph_2_2_0_t * dist_graph) {
dist_graph->in = NULL;
dist_graph->inw = NULL;
dist_graph->out = NULL;
dist_graph->outw = NULL;
dist_graph->indegree = 0;
dist_graph->outdegree = 0;
dist_graph->weighted = false;
}
static void mca_topo_base_comm_dist_graph_2_2_0_destruct(mca_topo_base_comm_dist_graph_2_2_0_t * dist_graph) {
free(dist_graph->in);
free(dist_graph->inw);
free(dist_graph->out);
free(dist_graph->outw);
}
OBJ_CLASS_INSTANCE(mca_topo_base_comm_dist_graph_2_2_0_t, opal_object_t,
mca_topo_base_comm_dist_graph_2_2_0_construct,
mca_topo_base_comm_dist_graph_2_2_0_destruct);
|