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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343
|
/* -*- Mode: C; c-basic-offset:2 ; indent-tabs-mode:nil -*- */
/*
* Copyright (c) 2006 The Trustees of Indiana University and Indiana
* University Research and Technology
* Corporation. All rights reserved.
* Copyright (c) 2006 The Technical University of Chemnitz. All
* rights reserved.
* Copyright (c) 2014-2018 Research Organization for Information Science
* and Technology (RIST). All rights reserved.
* Copyright (c) 2015 Los Alamos National Security, LLC. All rights
* reserved.
* Copyright (c) 2017 IBM Corporation. All rights reserved.
* Copyright (c) 2018 FUJITSU LIMITED. All rights reserved.
* $COPYRIGHT$
*
* Additional copyrights may follow
*
* Author(s): Torsten Hoefler <htor@cs.indiana.edu>
*
*/
#include "opal/align.h"
#include "ompi/op/op.h"
#include "nbc_internal.h"
static inline int scan_sched_linear(
int rank, int comm_size, const void *sendbuf, void *recvbuf, int count,
MPI_Datatype datatype, MPI_Op op, char inplace, NBC_Schedule *schedule,
void *tmpbuf);
static inline int scan_sched_recursivedoubling(
int rank, int comm_size, const void *sendbuf, void *recvbuf,
int count, MPI_Datatype datatype, MPI_Op op, char inplace,
NBC_Schedule *schedule, void *tmpbuf1, void *tmpbuf2);
#ifdef NBC_CACHE_SCHEDULE
/* tree comparison function for schedule cache */
int NBC_Scan_args_compare(NBC_Scan_args *a, NBC_Scan_args *b, void *param) {
if ((a->sendbuf == b->sendbuf) &&
(a->recvbuf == b->recvbuf) &&
(a->count == b->count) &&
(a->datatype == b->datatype) &&
(a->op == b->op) ) {
return 0;
}
if (a->sendbuf < b->sendbuf) {
return -1;
}
return 1;
}
#endif
static int nbc_scan_init(const void* sendbuf, void* recvbuf, int count, MPI_Datatype datatype, MPI_Op op,
struct ompi_communicator_t *comm, ompi_request_t ** request,
mca_coll_base_module_t *module, bool persistent) {
int rank, p, res;
ptrdiff_t gap, span;
NBC_Schedule *schedule;
void *tmpbuf = NULL, *tmpbuf1 = NULL, *tmpbuf2 = NULL;
enum { NBC_SCAN_LINEAR, NBC_SCAN_RDBL } alg;
char inplace;
ompi_coll_libnbc_module_t *libnbc_module = (ompi_coll_libnbc_module_t*) module;
NBC_IN_PLACE(sendbuf, recvbuf, inplace);
rank = ompi_comm_rank (comm);
p = ompi_comm_size (comm);
if (count == 0) {
return nbc_get_noop_request(persistent, request);
}
span = opal_datatype_span(&datatype->super, count, &gap);
if (libnbc_iscan_algorithm == 2) {
alg = NBC_SCAN_RDBL;
ptrdiff_t span_align = OPAL_ALIGN(span, datatype->super.align, ptrdiff_t);
tmpbuf = malloc(span_align + span);
if (NULL == tmpbuf) { return OMPI_ERR_OUT_OF_RESOURCE; }
tmpbuf1 = (void *)(-gap);
tmpbuf2 = (char *)(span_align) - gap;
} else {
alg = NBC_SCAN_LINEAR;
if (rank > 0) {
tmpbuf = malloc(span);
if (NULL == tmpbuf) { return OMPI_ERR_OUT_OF_RESOURCE; }
}
}
#ifdef NBC_CACHE_SCHEDULE
NBC_Scan_args *args, *found, search;
/* search schedule in communicator specific tree */
search.sendbuf = sendbuf;
search.recvbuf = recvbuf;
search.count = count;
search.datatype = datatype;
search.op = op;
found = (NBC_Scan_args *) hb_tree_search ((hb_tree *) libnbc_module->NBC_Dict[NBC_SCAN], &search);
if (NULL == found) {
#endif
schedule = OBJ_NEW(NBC_Schedule);
if (OPAL_UNLIKELY(NULL == schedule)) {
free(tmpbuf);
return OMPI_ERR_OUT_OF_RESOURCE;
}
if (alg == NBC_SCAN_LINEAR) {
res = scan_sched_linear(rank, p, sendbuf, recvbuf, count, datatype,
op, inplace, schedule, tmpbuf);
} else {
res = scan_sched_recursivedoubling(rank, p, sendbuf, recvbuf, count,
datatype, op, inplace, schedule, tmpbuf1, tmpbuf2);
}
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
OBJ_RELEASE(schedule);
free(tmpbuf);
return res;
}
res = NBC_Sched_commit(schedule);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
OBJ_RELEASE(schedule);
free(tmpbuf);
return res;
}
#ifdef NBC_CACHE_SCHEDULE
/* save schedule to tree */
args = (NBC_Scan_args *) malloc (sizeof (args));
if (NULL != args) {
args->sendbuf = sendbuf;
args->recvbuf = recvbuf;
args->count = count;
args->datatype = datatype;
args->op = op;
args->schedule = schedule;
res = hb_tree_insert ((hb_tree *) libnbc_module->NBC_Dict[NBC_SCAN], args, args, 0);
if (0 == res) {
OBJ_RETAIN(schedule);
/* increase number of elements for A2A */
if (++libnbc_module->NBC_Dict_size[NBC_SCAN] > NBC_SCHED_DICT_UPPER) {
NBC_SchedCache_dictwipe ((hb_tree *) libnbc_module->NBC_Dict[NBC_SCAN],
&libnbc_module->NBC_Dict_size[NBC_SCAN]);
}
} else {
NBC_Error("error in dict_insert() (%i)", res);
free (args);
}
}
} else {
/* found schedule */
schedule = found->schedule;
OBJ_RETAIN(schedule);
}
#endif
res = NBC_Schedule_request(schedule, comm, libnbc_module, persistent, request, tmpbuf);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
OBJ_RELEASE(schedule);
free(tmpbuf);
return res;
}
return OMPI_SUCCESS;
}
/*
* scan_sched_linear:
*
* Function: Linear algorithm for inclusive scan.
* Accepts: Same as MPI_Iscan
* Returns: MPI_SUCCESS or error code
*
* Working principle:
* 1. Each process (but process 0) receives from left neighbor
* 2. Performs op
* 3. All but rank p-1 do sends to it's right neighbor and exits
*
* Schedule length: O(1)
*/
static inline int scan_sched_linear(
int rank, int comm_size, const void *sendbuf, void *recvbuf, int count,
MPI_Datatype datatype, MPI_Op op, char inplace, NBC_Schedule *schedule,
void *tmpbuf)
{
int res = OMPI_SUCCESS;
if (!inplace) {
/* Copy data to recvbuf */
res = NBC_Sched_copy((void *)sendbuf, false, count, datatype,
recvbuf, false, count, datatype, schedule, false);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { goto cleanup_and_return; }
}
if (rank > 0) {
ptrdiff_t gap;
opal_datatype_span(&datatype->super, count, &gap);
/* We have to wait until we have the data */
res = NBC_Sched_recv((void *)(-gap), true, count, datatype, rank - 1, schedule, true);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { goto cleanup_and_return; }
/* Perform the reduce in my local buffer */
/* this cannot be done until tmpbuf is unused :-( so barrier after the op */
res = NBC_Sched_op((void *)(-gap), true, recvbuf, false, count, datatype, op, schedule,
true);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { goto cleanup_and_return; }
}
if (rank != comm_size - 1) {
res = NBC_Sched_send(recvbuf, false, count, datatype, rank + 1, schedule, false);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { goto cleanup_and_return; }
}
cleanup_and_return:
return res;
}
/*
* scan_sched_recursivedoubling:
*
* Function: Recursive doubling algorithm for inclusive scan.
* Accepts: Same as MPI_Iscan
* Returns: MPI_SUCCESS or error code
*
* Description: Implements recursive doubling algorithm for MPI_Iscan.
* The algorithm preserves order of operations so it can
* be used both by commutative and non-commutative operations.
*
* Example for 5 processes and commutative operation MPI_SUM:
* Process: 0 1 2 3 4
* recvbuf: [0] [1] [2] [3] [4]
* psend: [0] [1] [2] [3] [4]
*
* Step 1:
* recvbuf: [0] [0+1] [2] [2+3] [4]
* psend: [1+0] [0+1] [3+2] [2+3] [4]
*
* Step 2:
* recvbuf: [0] [0+1] [(1+0)+2] [(1+0)+(2+3)] [4]
* psend: [(3+2)+(1+0)] [(2+3)+(0+1)] [(1+0)+(3+2)] [(1+0)+(2+3)] [4]
*
* Step 3:
* recvbuf: [0] [0+1] [(1+0)+2] [(1+0)+(2+3)] [((3+2)+(1+0))+4]
* psend: [4+((3+2)+(1+0))] [((3+2)+(1+0))+4]
*
* Time complexity (worst case): \ceil(\log_2(p))(2\alpha + 2m\beta + 2m\gamma)
* Memory requirements (per process): 2 * count * typesize = O(count)
* Limitations: intra-communicators only
* Schedule length: O(log(p))
*/
static inline int scan_sched_recursivedoubling(
int rank, int comm_size, const void *sendbuf, void *recvbuf, int count,
MPI_Datatype datatype, MPI_Op op, char inplace,
NBC_Schedule *schedule, void *tmpbuf1, void *tmpbuf2)
{
int res = OMPI_SUCCESS;
if (!inplace) {
res = NBC_Sched_copy((void *)sendbuf, false, count, datatype,
recvbuf, false, count, datatype, schedule, true);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { goto cleanup_and_return; }
}
if (comm_size < 2)
goto cleanup_and_return;
char *psend = (char *)tmpbuf1;
char *precv = (char *)tmpbuf2;
res = NBC_Sched_copy(recvbuf, false, count, datatype,
psend, true, count, datatype, schedule, true);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { goto cleanup_and_return; }
int is_commute = ompi_op_is_commute(op);
for (int mask = 1; mask < comm_size; mask <<= 1) {
int remote = rank ^ mask;
if (remote < comm_size) {
res = NBC_Sched_send(psend, true, count, datatype, remote, schedule, false);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { goto cleanup_and_return; }
res = NBC_Sched_recv(precv, true, count, datatype, remote, schedule, true);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { goto cleanup_and_return; }
if (rank > remote) {
/* Accumulate prefix reduction: recvbuf = precv <op> recvbuf */
res = NBC_Sched_op(precv, true, recvbuf, false, count,
datatype, op, schedule, false);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { goto cleanup_and_return; }
/* Partial result: psend = precv <op> psend */
res = NBC_Sched_op(precv, true, psend, true, count,
datatype, op, schedule, true);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { goto cleanup_and_return; }
} else {
if (is_commute) {
/* psend = precv <op> psend */
res = NBC_Sched_op(precv, true, psend, true, count,
datatype, op, schedule, true);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { goto cleanup_and_return; }
} else {
/* precv = psend <op> precv */
res = NBC_Sched_op(psend, true, precv, true, count,
datatype, op, schedule, true);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) { goto cleanup_and_return; }
char *tmp = psend;
psend = precv;
precv = tmp;
}
}
}
}
cleanup_and_return:
return res;
}
int ompi_coll_libnbc_iscan(const void* sendbuf, void* recvbuf, int count, MPI_Datatype datatype, MPI_Op op,
struct ompi_communicator_t *comm, ompi_request_t ** request,
mca_coll_base_module_t *module) {
int res = nbc_scan_init(sendbuf, recvbuf, count, datatype, op,
comm, request, module, false);
if (OPAL_LIKELY(OMPI_SUCCESS != res)) {
return res;
}
res = NBC_Start(*(ompi_coll_libnbc_request_t **)request);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
NBC_Return_handle (*(ompi_coll_libnbc_request_t **)request);
*request = &ompi_request_null.request;
return res;
}
return OMPI_SUCCESS;
}
int ompi_coll_libnbc_scan_init(const void* sendbuf, void* recvbuf, int count, MPI_Datatype datatype, MPI_Op op,
struct ompi_communicator_t *comm, MPI_Info info, ompi_request_t ** request,
mca_coll_base_module_t *module) {
int res = nbc_scan_init(sendbuf, recvbuf, count, datatype, op,
comm, request, module, true);
if (OPAL_UNLIKELY(OMPI_SUCCESS != res)) {
return res;
}
return OMPI_SUCCESS;
}
|