1 /* 2 * ompt-specific.cpp -- OMPT internal functions 3 */ 4 5 //===----------------------------------------------------------------------===// 6 // 7 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 8 // See https://llvm.org/LICENSE.txt for license information. 9 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 10 // 11 //===----------------------------------------------------------------------===// 12 13 //****************************************************************************** 14 // include files 15 //****************************************************************************** 16 17 #include "kmp.h" 18 #include "ompt-specific.h" 19 20 #if KMP_OS_UNIX 21 #include <dlfcn.h> 22 #endif 23 24 #if KMP_OS_WINDOWS 25 #define THREAD_LOCAL __declspec(thread) 26 #else 27 #define THREAD_LOCAL __thread 28 #endif 29 30 #define OMPT_WEAK_ATTRIBUTE KMP_WEAK_ATTRIBUTE_INTERNAL 31 32 //****************************************************************************** 33 // macros 34 //****************************************************************************** 35 36 #define LWT_FROM_TEAM(team) (team)->t.ompt_serialized_team_info 37 38 #define OMPT_THREAD_ID_BITS 16 39 40 //****************************************************************************** 41 // private operations 42 //****************************************************************************** 43 44 //---------------------------------------------------------- 45 // traverse the team and task hierarchy 46 // note: __ompt_get_teaminfo and __ompt_get_task_info_object 47 // traverse the hierarchy similarly and need to be 48 // kept consistent 49 //---------------------------------------------------------- 50 51 ompt_team_info_t *__ompt_get_teaminfo(int depth, int *size) { 52 kmp_info_t *thr = ompt_get_thread(); 53 54 if (thr) { 55 kmp_team *team = thr->th.th_team; 56 if (team == NULL) 57 return NULL; 58 59 ompt_lw_taskteam_t *next_lwt = LWT_FROM_TEAM(team), *lwt = NULL; 60 61 while (depth > 0) { 62 // next lightweight team (if any) 63 if (lwt) 64 lwt = lwt->parent; 65 66 // next heavyweight team (if any) after 67 // lightweight teams are exhausted 68 if (!lwt && team) { 69 if (next_lwt) { 70 lwt = next_lwt; 71 next_lwt = NULL; 72 } else { 73 team = team->t.t_parent; 74 if (team) { 75 next_lwt = LWT_FROM_TEAM(team); 76 } 77 } 78 } 79 80 depth--; 81 } 82 83 if (lwt) { 84 // lightweight teams have one task 85 if (size) 86 *size = 1; 87 88 // return team info for lightweight team 89 return &lwt->ompt_team_info; 90 } else if (team) { 91 // extract size from heavyweight team 92 if (size) 93 *size = team->t.t_nproc; 94 95 // return team info for heavyweight team 96 return &team->t.ompt_team_info; 97 } 98 } 99 100 return NULL; 101 } 102 103 ompt_task_info_t *__ompt_get_task_info_object(int depth) { 104 ompt_task_info_t *info = NULL; 105 kmp_info_t *thr = ompt_get_thread(); 106 107 if (thr) { 108 kmp_taskdata_t *taskdata = thr->th.th_current_task; 109 ompt_lw_taskteam_t *lwt = NULL, 110 *next_lwt = LWT_FROM_TEAM(taskdata->td_team); 111 112 while (depth > 0) { 113 // next lightweight team (if any) 114 if (lwt) 115 lwt = lwt->parent; 116 117 // next heavyweight team (if any) after 118 // lightweight teams are exhausted 119 if (!lwt && taskdata) { 120 if (next_lwt) { 121 lwt = next_lwt; 122 next_lwt = NULL; 123 } else { 124 taskdata = taskdata->td_parent; 125 if (taskdata) { 126 next_lwt = LWT_FROM_TEAM(taskdata->td_team); 127 } 128 } 129 } 130 depth--; 131 } 132 133 if (lwt) { 134 info = &lwt->ompt_task_info; 135 } else if (taskdata) { 136 info = &taskdata->ompt_task_info; 137 } 138 } 139 140 return info; 141 } 142 143 ompt_task_info_t *__ompt_get_scheduling_taskinfo(int depth) { 144 ompt_task_info_t *info = NULL; 145 kmp_info_t *thr = ompt_get_thread(); 146 147 if (thr) { 148 kmp_taskdata_t *taskdata = thr->th.th_current_task; 149 150 ompt_lw_taskteam_t *lwt = NULL, 151 *next_lwt = LWT_FROM_TEAM(taskdata->td_team); 152 153 while (depth > 0) { 154 // next lightweight team (if any) 155 if (lwt) 156 lwt = lwt->parent; 157 158 // next heavyweight team (if any) after 159 // lightweight teams are exhausted 160 if (!lwt && taskdata) { 161 // first try scheduling parent (for explicit task scheduling) 162 if (taskdata->ompt_task_info.scheduling_parent) { 163 taskdata = taskdata->ompt_task_info.scheduling_parent; 164 } else if (next_lwt) { 165 lwt = next_lwt; 166 next_lwt = NULL; 167 } else { 168 // then go for implicit tasks 169 taskdata = taskdata->td_parent; 170 if (taskdata) { 171 next_lwt = LWT_FROM_TEAM(taskdata->td_team); 172 } 173 } 174 } 175 depth--; 176 } 177 178 if (lwt) { 179 info = &lwt->ompt_task_info; 180 } else if (taskdata) { 181 info = &taskdata->ompt_task_info; 182 } 183 } 184 185 return info; 186 } 187 188 //****************************************************************************** 189 // interface operations 190 //****************************************************************************** 191 //---------------------------------------------------------- 192 // initialization support 193 //---------------------------------------------------------- 194 195 void __ompt_force_initialization() { __kmp_serial_initialize(); } 196 197 //---------------------------------------------------------- 198 // thread support 199 //---------------------------------------------------------- 200 201 ompt_data_t *__ompt_get_thread_data_internal() { 202 if (__kmp_get_gtid() >= 0) { 203 kmp_info_t *thread = ompt_get_thread(); 204 if (thread == NULL) 205 return NULL; 206 return &(thread->th.ompt_thread_info.thread_data); 207 } 208 return NULL; 209 } 210 211 //---------------------------------------------------------- 212 // state support 213 //---------------------------------------------------------- 214 215 void __ompt_thread_assign_wait_id(void *variable) { 216 kmp_info_t *ti = ompt_get_thread(); 217 218 if (ti) 219 ti->th.ompt_thread_info.wait_id = (ompt_wait_id_t)(uintptr_t)variable; 220 } 221 222 int __ompt_get_state_internal(ompt_wait_id_t *omp_wait_id) { 223 kmp_info_t *ti = ompt_get_thread(); 224 225 if (ti) { 226 if (omp_wait_id) 227 *omp_wait_id = ti->th.ompt_thread_info.wait_id; 228 return ti->th.ompt_thread_info.state; 229 } 230 return ompt_state_undefined; 231 } 232 233 //---------------------------------------------------------- 234 // parallel region support 235 //---------------------------------------------------------- 236 237 int __ompt_get_parallel_info_internal(int ancestor_level, 238 ompt_data_t **parallel_data, 239 int *team_size) { 240 if (__kmp_get_gtid() >= 0) { 241 ompt_team_info_t *info; 242 if (team_size) { 243 info = __ompt_get_teaminfo(ancestor_level, team_size); 244 } else { 245 info = __ompt_get_teaminfo(ancestor_level, NULL); 246 } 247 if (parallel_data) { 248 *parallel_data = info ? &(info->parallel_data) : NULL; 249 } 250 return info ? 2 : 0; 251 } else { 252 return 0; 253 } 254 } 255 256 //---------------------------------------------------------- 257 // lightweight task team support 258 //---------------------------------------------------------- 259 260 void __ompt_lw_taskteam_init(ompt_lw_taskteam_t *lwt, kmp_info_t *thr, int gtid, 261 ompt_data_t *ompt_pid, void *codeptr) { 262 // initialize parallel_data with input, return address to parallel_data on 263 // exit 264 lwt->ompt_team_info.parallel_data = *ompt_pid; 265 lwt->ompt_team_info.master_return_address = codeptr; 266 lwt->ompt_task_info.task_data.value = 0; 267 lwt->ompt_task_info.frame.enter_frame = ompt_data_none; 268 lwt->ompt_task_info.frame.exit_frame = ompt_data_none; 269 lwt->ompt_task_info.scheduling_parent = NULL; 270 lwt->heap = 0; 271 lwt->parent = 0; 272 } 273 274 void __ompt_lw_taskteam_link(ompt_lw_taskteam_t *lwt, kmp_info_t *thr, 275 int on_heap, bool always) { 276 ompt_lw_taskteam_t *link_lwt = lwt; 277 if (always || 278 thr->th.th_team->t.t_serialized > 279 1) { // we already have a team, so link the new team and swap values 280 if (on_heap) { // the lw_taskteam cannot stay on stack, allocate it on heap 281 link_lwt = 282 (ompt_lw_taskteam_t *)__kmp_allocate(sizeof(ompt_lw_taskteam_t)); 283 } 284 link_lwt->heap = on_heap; 285 286 // would be swap in the (on_stack) case. 287 ompt_team_info_t tmp_team = lwt->ompt_team_info; 288 link_lwt->ompt_team_info = *OMPT_CUR_TEAM_INFO(thr); 289 *OMPT_CUR_TEAM_INFO(thr) = tmp_team; 290 291 // link the taskteam into the list of taskteams: 292 ompt_lw_taskteam_t *my_parent = 293 thr->th.th_team->t.ompt_serialized_team_info; 294 link_lwt->parent = my_parent; 295 thr->th.th_team->t.ompt_serialized_team_info = link_lwt; 296 #if OMPD_SUPPORT 297 if (ompd_state & OMPD_ENABLE_BP) { 298 ompd_bp_parallel_begin(); 299 } 300 #endif 301 302 ompt_task_info_t tmp_task = lwt->ompt_task_info; 303 link_lwt->ompt_task_info = *OMPT_CUR_TASK_INFO(thr); 304 *OMPT_CUR_TASK_INFO(thr) = tmp_task; 305 } else { 306 // this is the first serialized team, so we just store the values in the 307 // team and drop the taskteam-object 308 *OMPT_CUR_TEAM_INFO(thr) = lwt->ompt_team_info; 309 #if OMPD_SUPPORT 310 if (ompd_state & OMPD_ENABLE_BP) { 311 ompd_bp_parallel_begin(); 312 } 313 #endif 314 *OMPT_CUR_TASK_INFO(thr) = lwt->ompt_task_info; 315 } 316 } 317 318 void __ompt_lw_taskteam_unlink(kmp_info_t *thr) { 319 ompt_lw_taskteam_t *lwtask = thr->th.th_team->t.ompt_serialized_team_info; 320 if (lwtask) { 321 ompt_task_info_t tmp_task = lwtask->ompt_task_info; 322 lwtask->ompt_task_info = *OMPT_CUR_TASK_INFO(thr); 323 *OMPT_CUR_TASK_INFO(thr) = tmp_task; 324 #if OMPD_SUPPORT 325 if (ompd_state & OMPD_ENABLE_BP) { 326 ompd_bp_parallel_end(); 327 } 328 #endif 329 thr->th.th_team->t.ompt_serialized_team_info = lwtask->parent; 330 331 ompt_team_info_t tmp_team = lwtask->ompt_team_info; 332 lwtask->ompt_team_info = *OMPT_CUR_TEAM_INFO(thr); 333 *OMPT_CUR_TEAM_INFO(thr) = tmp_team; 334 335 if (lwtask->heap) { 336 __kmp_free(lwtask); 337 lwtask = NULL; 338 } 339 } 340 // return lwtask; 341 } 342 343 //---------------------------------------------------------- 344 // task support 345 //---------------------------------------------------------- 346 347 int __ompt_get_task_info_internal(int ancestor_level, int *type, 348 ompt_data_t **task_data, 349 ompt_frame_t **task_frame, 350 ompt_data_t **parallel_data, 351 int *thread_num) { 352 if (__kmp_get_gtid() < 0) 353 return 0; 354 355 if (ancestor_level < 0) 356 return 0; 357 358 // copied from __ompt_get_scheduling_taskinfo 359 ompt_task_info_t *info = NULL; 360 ompt_team_info_t *team_info = NULL; 361 kmp_info_t *thr = ompt_get_thread(); 362 int level = ancestor_level; 363 364 if (thr) { 365 kmp_taskdata_t *taskdata = thr->th.th_current_task; 366 if (taskdata == NULL) 367 return 0; 368 kmp_team *team = thr->th.th_team, *prev_team = NULL; 369 if (team == NULL) 370 return 0; 371 ompt_lw_taskteam_t *lwt = NULL, 372 *next_lwt = LWT_FROM_TEAM(taskdata->td_team); 373 374 while (ancestor_level > 0) { 375 // next lightweight team (if any) 376 if (lwt) 377 lwt = lwt->parent; 378 379 // next heavyweight team (if any) after 380 // lightweight teams are exhausted 381 if (!lwt && taskdata) { 382 // first try scheduling parent (for explicit task scheduling) 383 if (taskdata->ompt_task_info.scheduling_parent) { 384 taskdata = taskdata->ompt_task_info.scheduling_parent; 385 } else if (next_lwt) { 386 lwt = next_lwt; 387 next_lwt = NULL; 388 } else { 389 // then go for implicit tasks 390 taskdata = taskdata->td_parent; 391 if (team == NULL) 392 return 0; 393 prev_team = team; 394 team = team->t.t_parent; 395 if (taskdata) { 396 next_lwt = LWT_FROM_TEAM(taskdata->td_team); 397 } 398 } 399 } 400 ancestor_level--; 401 } 402 403 if (lwt) { 404 info = &lwt->ompt_task_info; 405 team_info = &lwt->ompt_team_info; 406 if (type) { 407 *type = ompt_task_implicit; 408 } 409 } else if (taskdata) { 410 info = &taskdata->ompt_task_info; 411 team_info = &team->t.ompt_team_info; 412 if (type) { 413 if (taskdata->td_parent) { 414 *type = (taskdata->td_flags.tasktype ? ompt_task_explicit 415 : ompt_task_implicit) | 416 TASK_TYPE_DETAILS_FORMAT(taskdata); 417 } else { 418 *type = ompt_task_initial; 419 } 420 } 421 } 422 if (task_data) { 423 *task_data = info ? &info->task_data : NULL; 424 } 425 if (task_frame) { 426 // OpenMP spec asks for the scheduling task to be returned. 427 *task_frame = info ? &info->frame : NULL; 428 } 429 if (parallel_data) { 430 *parallel_data = team_info ? &(team_info->parallel_data) : NULL; 431 } 432 if (thread_num) { 433 if (level == 0) 434 *thread_num = __kmp_get_tid(); 435 else if (lwt) 436 *thread_num = 0; 437 else if (!prev_team) { 438 // The innermost parallel region contains at least one explicit task. 439 // The task at level > 0 is either an implicit task that 440 // corresponds to the mentioned region or one of the explicit tasks 441 // nested inside the same region. Note that the task isn't the 442 // innermost explicit tasks (because of condition level > 0). 443 // Since the task at this level still belongs to the innermost parallel 444 // region, thread_num is determined the same way as for level==0. 445 *thread_num = __kmp_get_tid(); 446 } else 447 *thread_num = prev_team->t.t_master_tid; 448 // *thread_num = team->t.t_master_tid; 449 } 450 return info ? 2 : 0; 451 } 452 return 0; 453 } 454 455 int __ompt_get_task_memory_internal(void **addr, size_t *size, int blocknum) { 456 if (blocknum != 0) 457 return 0; // support only a single block 458 459 kmp_info_t *thr = ompt_get_thread(); 460 if (!thr) 461 return 0; 462 463 kmp_taskdata_t *taskdata = thr->th.th_current_task; 464 kmp_task_t *task = KMP_TASKDATA_TO_TASK(taskdata); 465 466 if (taskdata->td_flags.tasktype != TASK_EXPLICIT) 467 return 0; // support only explicit task 468 469 void *ret_addr; 470 int64_t ret_size = taskdata->td_size_alloc - sizeof(kmp_taskdata_t); 471 472 // kmp_task_t->data1 is an optional member 473 if (taskdata->td_flags.destructors_thunk) 474 ret_addr = &task->data1 + 1; 475 else 476 ret_addr = &task->part_id + 1; 477 478 ret_size -= (char *)(ret_addr) - (char *)(task); 479 if (ret_size < 0) 480 return 0; 481 482 *addr = ret_addr; 483 *size = (size_t)ret_size; 484 return 1; 485 } 486 487 //---------------------------------------------------------- 488 // team support 489 //---------------------------------------------------------- 490 491 void __ompt_team_assign_id(kmp_team_t *team, ompt_data_t ompt_pid) { 492 team->t.ompt_team_info.parallel_data = ompt_pid; 493 } 494 495 //---------------------------------------------------------- 496 // misc 497 //---------------------------------------------------------- 498 499 static uint64_t __ompt_get_unique_id_internal() { 500 static uint64_t thread = 1; 501 static THREAD_LOCAL uint64_t ID = 0; 502 if (ID == 0) { 503 uint64_t new_thread = KMP_TEST_THEN_INC64((kmp_int64 *)&thread); 504 ID = new_thread << (sizeof(uint64_t) * 8 - OMPT_THREAD_ID_BITS); 505 } 506 return ++ID; 507 } 508 509 ompt_sync_region_t __ompt_get_barrier_kind(enum barrier_type bt, 510 kmp_info_t *thr) { 511 if (bt == bs_forkjoin_barrier) 512 return ompt_sync_region_barrier_implicit; 513 514 if (bt != bs_plain_barrier) 515 return ompt_sync_region_barrier_implementation; 516 517 if (!thr->th.th_ident) 518 return ompt_sync_region_barrier; 519 520 kmp_int32 flags = thr->th.th_ident->flags; 521 522 if ((flags & KMP_IDENT_BARRIER_EXPL) != 0) 523 return ompt_sync_region_barrier_explicit; 524 525 if ((flags & KMP_IDENT_BARRIER_IMPL) != 0) 526 return ompt_sync_region_barrier_implicit; 527 528 return ompt_sync_region_barrier_implementation; 529 } 530