1 /* 2 * kmp_gsupport.cpp 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 #include "kmp.h" 14 #include "kmp_atomic.h" 15 16 #if OMPT_SUPPORT 17 #include "ompt-specific.h" 18 #endif 19 20 #ifdef __cplusplus 21 extern "C" { 22 #endif // __cplusplus 23 24 #define MKLOC(loc, routine) \ 25 static ident_t(loc) = {0, KMP_IDENT_KMPC, 0, 0, ";unknown;unknown;0;0;;"}; 26 27 #include "kmp_ftn_os.h" 28 29 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_BARRIER)(void) { 30 int gtid = __kmp_entry_gtid(); 31 MKLOC(loc, "GOMP_barrier"); 32 KA_TRACE(20, ("GOMP_barrier: T#%d\n", gtid)); 33 #if OMPT_SUPPORT && OMPT_OPTIONAL 34 ompt_frame_t *ompt_frame; 35 if (ompt_enabled.enabled) { 36 __ompt_get_task_info_internal(0, NULL, NULL, &ompt_frame, NULL, NULL); 37 ompt_frame->enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0); 38 OMPT_STORE_RETURN_ADDRESS(gtid); 39 } 40 #endif 41 __kmpc_barrier(&loc, gtid); 42 #if OMPT_SUPPORT && OMPT_OPTIONAL 43 if (ompt_enabled.enabled) { 44 ompt_frame->enter_frame = ompt_data_none; 45 } 46 #endif 47 } 48 49 // Mutual exclusion 50 51 // The symbol that icc/ifort generates for unnamed for unnamed critical sections 52 // - .gomp_critical_user_ - is defined using .comm in any objects reference it. 53 // We can't reference it directly here in C code, as the symbol contains a ".". 54 // 55 // The RTL contains an assembly language definition of .gomp_critical_user_ 56 // with another symbol __kmp_unnamed_critical_addr initialized with it's 57 // address. 58 extern kmp_critical_name *__kmp_unnamed_critical_addr; 59 60 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_CRITICAL_START)(void) { 61 int gtid = __kmp_entry_gtid(); 62 MKLOC(loc, "GOMP_critical_start"); 63 KA_TRACE(20, ("GOMP_critical_start: T#%d\n", gtid)); 64 #if OMPT_SUPPORT && OMPT_OPTIONAL 65 OMPT_STORE_RETURN_ADDRESS(gtid); 66 #endif 67 __kmpc_critical(&loc, gtid, __kmp_unnamed_critical_addr); 68 } 69 70 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_CRITICAL_END)(void) { 71 int gtid = __kmp_get_gtid(); 72 MKLOC(loc, "GOMP_critical_end"); 73 KA_TRACE(20, ("GOMP_critical_end: T#%d\n", gtid)); 74 #if OMPT_SUPPORT && OMPT_OPTIONAL 75 OMPT_STORE_RETURN_ADDRESS(gtid); 76 #endif 77 __kmpc_end_critical(&loc, gtid, __kmp_unnamed_critical_addr); 78 } 79 80 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_CRITICAL_NAME_START)(void **pptr) { 81 int gtid = __kmp_entry_gtid(); 82 MKLOC(loc, "GOMP_critical_name_start"); 83 KA_TRACE(20, ("GOMP_critical_name_start: T#%d\n", gtid)); 84 __kmpc_critical(&loc, gtid, (kmp_critical_name *)pptr); 85 } 86 87 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_CRITICAL_NAME_END)(void **pptr) { 88 int gtid = __kmp_get_gtid(); 89 MKLOC(loc, "GOMP_critical_name_end"); 90 KA_TRACE(20, ("GOMP_critical_name_end: T#%d\n", gtid)); 91 __kmpc_end_critical(&loc, gtid, (kmp_critical_name *)pptr); 92 } 93 94 // The Gnu codegen tries to use locked operations to perform atomic updates 95 // inline. If it can't, then it calls GOMP_atomic_start() before performing 96 // the update and GOMP_atomic_end() afterward, regardless of the data type. 97 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_ATOMIC_START)(void) { 98 int gtid = __kmp_entry_gtid(); 99 KA_TRACE(20, ("GOMP_atomic_start: T#%d\n", gtid)); 100 101 #if OMPT_SUPPORT 102 __ompt_thread_assign_wait_id(0); 103 #endif 104 105 __kmp_acquire_atomic_lock(&__kmp_atomic_lock, gtid); 106 } 107 108 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_ATOMIC_END)(void) { 109 int gtid = __kmp_get_gtid(); 110 KA_TRACE(20, ("GOMP_atomic_end: T#%d\n", gtid)); 111 __kmp_release_atomic_lock(&__kmp_atomic_lock, gtid); 112 } 113 114 int KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SINGLE_START)(void) { 115 int gtid = __kmp_entry_gtid(); 116 MKLOC(loc, "GOMP_single_start"); 117 KA_TRACE(20, ("GOMP_single_start: T#%d\n", gtid)); 118 119 if (!TCR_4(__kmp_init_parallel)) 120 __kmp_parallel_initialize(); 121 122 #if OMP_50_ENABLED 123 __kmp_resume_if_soft_paused(); 124 #endif 125 126 // 3rd parameter == FALSE prevents kmp_enter_single from pushing a 127 // workshare when USE_CHECKS is defined. We need to avoid the push, 128 // as there is no corresponding GOMP_single_end() call. 129 kmp_int32 rc = __kmp_enter_single(gtid, &loc, FALSE); 130 131 #if OMPT_SUPPORT && OMPT_OPTIONAL 132 kmp_info_t *this_thr = __kmp_threads[gtid]; 133 kmp_team_t *team = this_thr->th.th_team; 134 int tid = __kmp_tid_from_gtid(gtid); 135 136 if (ompt_enabled.enabled) { 137 if (rc) { 138 if (ompt_enabled.ompt_callback_work) { 139 ompt_callbacks.ompt_callback(ompt_callback_work)( 140 ompt_work_single_executor, ompt_scope_begin, 141 &(team->t.ompt_team_info.parallel_data), 142 &(team->t.t_implicit_task_taskdata[tid].ompt_task_info.task_data), 143 1, OMPT_GET_RETURN_ADDRESS(0)); 144 } 145 } else { 146 if (ompt_enabled.ompt_callback_work) { 147 ompt_callbacks.ompt_callback(ompt_callback_work)( 148 ompt_work_single_other, ompt_scope_begin, 149 &(team->t.ompt_team_info.parallel_data), 150 &(team->t.t_implicit_task_taskdata[tid].ompt_task_info.task_data), 151 1, OMPT_GET_RETURN_ADDRESS(0)); 152 ompt_callbacks.ompt_callback(ompt_callback_work)( 153 ompt_work_single_other, ompt_scope_end, 154 &(team->t.ompt_team_info.parallel_data), 155 &(team->t.t_implicit_task_taskdata[tid].ompt_task_info.task_data), 156 1, OMPT_GET_RETURN_ADDRESS(0)); 157 } 158 } 159 } 160 #endif 161 162 return rc; 163 } 164 165 void *KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SINGLE_COPY_START)(void) { 166 void *retval; 167 int gtid = __kmp_entry_gtid(); 168 MKLOC(loc, "GOMP_single_copy_start"); 169 KA_TRACE(20, ("GOMP_single_copy_start: T#%d\n", gtid)); 170 171 if (!TCR_4(__kmp_init_parallel)) 172 __kmp_parallel_initialize(); 173 174 #if OMP_50_ENABLED 175 __kmp_resume_if_soft_paused(); 176 #endif 177 178 // If this is the first thread to enter, return NULL. The generated code will 179 // then call GOMP_single_copy_end() for this thread only, with the 180 // copyprivate data pointer as an argument. 181 if (__kmp_enter_single(gtid, &loc, FALSE)) 182 return NULL; 183 184 // Wait for the first thread to set the copyprivate data pointer, 185 // and for all other threads to reach this point. 186 187 #if OMPT_SUPPORT && OMPT_OPTIONAL 188 ompt_frame_t *ompt_frame; 189 if (ompt_enabled.enabled) { 190 __ompt_get_task_info_internal(0, NULL, NULL, &ompt_frame, NULL, NULL); 191 ompt_frame->enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0); 192 OMPT_STORE_RETURN_ADDRESS(gtid); 193 } 194 #endif 195 __kmp_barrier(bs_plain_barrier, gtid, FALSE, 0, NULL, NULL); 196 197 // Retrieve the value of the copyprivate data point, and wait for all 198 // threads to do likewise, then return. 199 retval = __kmp_team_from_gtid(gtid)->t.t_copypriv_data; 200 #if OMPT_SUPPORT && OMPT_OPTIONAL 201 if (ompt_enabled.enabled) { 202 OMPT_STORE_RETURN_ADDRESS(gtid); 203 } 204 #endif 205 __kmp_barrier(bs_plain_barrier, gtid, FALSE, 0, NULL, NULL); 206 #if OMPT_SUPPORT && OMPT_OPTIONAL 207 if (ompt_enabled.enabled) { 208 ompt_frame->enter_frame = ompt_data_none; 209 } 210 #endif 211 return retval; 212 } 213 214 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SINGLE_COPY_END)(void *data) { 215 int gtid = __kmp_get_gtid(); 216 KA_TRACE(20, ("GOMP_single_copy_end: T#%d\n", gtid)); 217 218 // Set the copyprivate data pointer fo the team, then hit the barrier so that 219 // the other threads will continue on and read it. Hit another barrier before 220 // continuing, so that the know that the copyprivate data pointer has been 221 // propagated to all threads before trying to reuse the t_copypriv_data field. 222 __kmp_team_from_gtid(gtid)->t.t_copypriv_data = data; 223 #if OMPT_SUPPORT && OMPT_OPTIONAL 224 ompt_frame_t *ompt_frame; 225 if (ompt_enabled.enabled) { 226 __ompt_get_task_info_internal(0, NULL, NULL, &ompt_frame, NULL, NULL); 227 ompt_frame->enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0); 228 OMPT_STORE_RETURN_ADDRESS(gtid); 229 } 230 #endif 231 __kmp_barrier(bs_plain_barrier, gtid, FALSE, 0, NULL, NULL); 232 #if OMPT_SUPPORT && OMPT_OPTIONAL 233 if (ompt_enabled.enabled) { 234 OMPT_STORE_RETURN_ADDRESS(gtid); 235 } 236 #endif 237 __kmp_barrier(bs_plain_barrier, gtid, FALSE, 0, NULL, NULL); 238 #if OMPT_SUPPORT && OMPT_OPTIONAL 239 if (ompt_enabled.enabled) { 240 ompt_frame->enter_frame = ompt_data_none; 241 } 242 #endif 243 } 244 245 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_ORDERED_START)(void) { 246 int gtid = __kmp_entry_gtid(); 247 MKLOC(loc, "GOMP_ordered_start"); 248 KA_TRACE(20, ("GOMP_ordered_start: T#%d\n", gtid)); 249 #if OMPT_SUPPORT && OMPT_OPTIONAL 250 OMPT_STORE_RETURN_ADDRESS(gtid); 251 #endif 252 __kmpc_ordered(&loc, gtid); 253 } 254 255 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_ORDERED_END)(void) { 256 int gtid = __kmp_get_gtid(); 257 MKLOC(loc, "GOMP_ordered_end"); 258 KA_TRACE(20, ("GOMP_ordered_start: T#%d\n", gtid)); 259 #if OMPT_SUPPORT && OMPT_OPTIONAL 260 OMPT_STORE_RETURN_ADDRESS(gtid); 261 #endif 262 __kmpc_end_ordered(&loc, gtid); 263 } 264 265 // Dispatch macro defs 266 // 267 // They come in two flavors: 64-bit unsigned, and either 32-bit signed 268 // (IA-32 architecture) or 64-bit signed (Intel(R) 64). 269 270 #if KMP_ARCH_X86 || KMP_ARCH_ARM || KMP_ARCH_MIPS 271 #define KMP_DISPATCH_INIT __kmp_aux_dispatch_init_4 272 #define KMP_DISPATCH_FINI_CHUNK __kmp_aux_dispatch_fini_chunk_4 273 #define KMP_DISPATCH_NEXT __kmpc_dispatch_next_4 274 #else 275 #define KMP_DISPATCH_INIT __kmp_aux_dispatch_init_8 276 #define KMP_DISPATCH_FINI_CHUNK __kmp_aux_dispatch_fini_chunk_8 277 #define KMP_DISPATCH_NEXT __kmpc_dispatch_next_8 278 #endif /* KMP_ARCH_X86 */ 279 280 #define KMP_DISPATCH_INIT_ULL __kmp_aux_dispatch_init_8u 281 #define KMP_DISPATCH_FINI_CHUNK_ULL __kmp_aux_dispatch_fini_chunk_8u 282 #define KMP_DISPATCH_NEXT_ULL __kmpc_dispatch_next_8u 283 284 // The parallel contruct 285 286 #ifndef KMP_DEBUG 287 static 288 #endif /* KMP_DEBUG */ 289 void 290 __kmp_GOMP_microtask_wrapper(int *gtid, int *npr, void (*task)(void *), 291 void *data) { 292 #if OMPT_SUPPORT 293 kmp_info_t *thr; 294 ompt_frame_t *ompt_frame; 295 ompt_state_t enclosing_state; 296 297 if (ompt_enabled.enabled) { 298 // get pointer to thread data structure 299 thr = __kmp_threads[*gtid]; 300 301 // save enclosing task state; set current state for task 302 enclosing_state = thr->th.ompt_thread_info.state; 303 thr->th.ompt_thread_info.state = ompt_state_work_parallel; 304 305 // set task frame 306 __ompt_get_task_info_internal(0, NULL, NULL, &ompt_frame, NULL, NULL); 307 ompt_frame->exit_frame.ptr = OMPT_GET_FRAME_ADDRESS(0); 308 } 309 #endif 310 311 task(data); 312 313 #if OMPT_SUPPORT 314 if (ompt_enabled.enabled) { 315 // clear task frame 316 ompt_frame->exit_frame = ompt_data_none; 317 318 // restore enclosing state 319 thr->th.ompt_thread_info.state = enclosing_state; 320 } 321 #endif 322 } 323 324 #ifndef KMP_DEBUG 325 static 326 #endif /* KMP_DEBUG */ 327 void 328 __kmp_GOMP_parallel_microtask_wrapper(int *gtid, int *npr, 329 void (*task)(void *), void *data, 330 unsigned num_threads, ident_t *loc, 331 enum sched_type schedule, long start, 332 long end, long incr, 333 long chunk_size) { 334 // Intialize the loop worksharing construct. 335 336 KMP_DISPATCH_INIT(loc, *gtid, schedule, start, end, incr, chunk_size, 337 schedule != kmp_sch_static); 338 339 #if OMPT_SUPPORT 340 kmp_info_t *thr; 341 ompt_frame_t *ompt_frame; 342 ompt_state_t enclosing_state; 343 344 if (ompt_enabled.enabled) { 345 thr = __kmp_threads[*gtid]; 346 // save enclosing task state; set current state for task 347 enclosing_state = thr->th.ompt_thread_info.state; 348 thr->th.ompt_thread_info.state = ompt_state_work_parallel; 349 350 // set task frame 351 __ompt_get_task_info_internal(0, NULL, NULL, &ompt_frame, NULL, NULL); 352 ompt_frame->exit_frame.ptr = OMPT_GET_FRAME_ADDRESS(0); 353 } 354 #endif 355 356 // Now invoke the microtask. 357 task(data); 358 359 #if OMPT_SUPPORT 360 if (ompt_enabled.enabled) { 361 // clear task frame 362 ompt_frame->exit_frame = ompt_data_none; 363 364 // reset enclosing state 365 thr->th.ompt_thread_info.state = enclosing_state; 366 } 367 #endif 368 } 369 370 #ifndef KMP_DEBUG 371 static 372 #endif /* KMP_DEBUG */ 373 void 374 __kmp_GOMP_fork_call(ident_t *loc, int gtid, void (*unwrapped_task)(void *), 375 microtask_t wrapper, int argc, ...) { 376 int rc; 377 kmp_info_t *thr = __kmp_threads[gtid]; 378 kmp_team_t *team = thr->th.th_team; 379 int tid = __kmp_tid_from_gtid(gtid); 380 381 va_list ap; 382 va_start(ap, argc); 383 384 rc = __kmp_fork_call(loc, gtid, fork_context_gnu, argc, wrapper, 385 __kmp_invoke_task_func, 386 #if (KMP_ARCH_X86_64 || KMP_ARCH_ARM || KMP_ARCH_AARCH64) && KMP_OS_LINUX 387 &ap 388 #else 389 ap 390 #endif 391 ); 392 393 va_end(ap); 394 395 if (rc) { 396 __kmp_run_before_invoked_task(gtid, tid, thr, team); 397 } 398 399 #if OMPT_SUPPORT 400 int ompt_team_size; 401 if (ompt_enabled.enabled) { 402 ompt_team_info_t *team_info = __ompt_get_teaminfo(0, NULL); 403 ompt_task_info_t *task_info = __ompt_get_task_info_object(0); 404 405 // implicit task callback 406 if (ompt_enabled.ompt_callback_implicit_task) { 407 ompt_team_size = __kmp_team_from_gtid(gtid)->t.t_nproc; 408 ompt_callbacks.ompt_callback(ompt_callback_implicit_task)( 409 ompt_scope_begin, &(team_info->parallel_data), 410 &(task_info->task_data), ompt_team_size, __kmp_tid_from_gtid(gtid), ompt_task_implicit); // TODO: Can this be ompt_task_initial? 411 task_info->thread_num = __kmp_tid_from_gtid(gtid); 412 } 413 thr->th.ompt_thread_info.state = ompt_state_work_parallel; 414 } 415 #endif 416 } 417 418 static void __kmp_GOMP_serialized_parallel(ident_t *loc, kmp_int32 gtid, 419 void (*task)(void *)) { 420 #if OMPT_SUPPORT 421 OMPT_STORE_RETURN_ADDRESS(gtid); 422 #endif 423 __kmp_serialized_parallel(loc, gtid); 424 } 425 426 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_START)(void (*task)(void *), 427 void *data, 428 unsigned num_threads) { 429 int gtid = __kmp_entry_gtid(); 430 431 #if OMPT_SUPPORT 432 ompt_frame_t *parent_frame, *frame; 433 434 if (ompt_enabled.enabled) { 435 __ompt_get_task_info_internal(0, NULL, NULL, &parent_frame, NULL, NULL); 436 parent_frame->enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0); 437 OMPT_STORE_RETURN_ADDRESS(gtid); 438 } 439 #endif 440 441 MKLOC(loc, "GOMP_parallel_start"); 442 KA_TRACE(20, ("GOMP_parallel_start: T#%d\n", gtid)); 443 444 if (__kmpc_ok_to_fork(&loc) && (num_threads != 1)) { 445 if (num_threads != 0) { 446 __kmp_push_num_threads(&loc, gtid, num_threads); 447 } 448 __kmp_GOMP_fork_call(&loc, gtid, task, 449 (microtask_t)__kmp_GOMP_microtask_wrapper, 2, task, 450 data); 451 } else { 452 __kmp_GOMP_serialized_parallel(&loc, gtid, task); 453 } 454 455 #if OMPT_SUPPORT 456 if (ompt_enabled.enabled) { 457 __ompt_get_task_info_internal(0, NULL, NULL, &frame, NULL, NULL); 458 frame->exit_frame.ptr = OMPT_GET_FRAME_ADDRESS(0); 459 } 460 #endif 461 } 462 463 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_END)(void) { 464 int gtid = __kmp_get_gtid(); 465 kmp_info_t *thr; 466 467 thr = __kmp_threads[gtid]; 468 469 MKLOC(loc, "GOMP_parallel_end"); 470 KA_TRACE(20, ("GOMP_parallel_end: T#%d\n", gtid)); 471 472 if (!thr->th.th_team->t.t_serialized) { 473 __kmp_run_after_invoked_task(gtid, __kmp_tid_from_gtid(gtid), thr, 474 thr->th.th_team); 475 476 #if OMPT_SUPPORT 477 if (ompt_enabled.enabled) { 478 // Implicit task is finished here, in the barrier we might schedule 479 // deferred tasks, 480 // these don't see the implicit task on the stack 481 OMPT_CUR_TASK_INFO(thr)->frame.exit_frame = ompt_data_none; 482 } 483 #endif 484 485 __kmp_join_call(&loc, gtid 486 #if OMPT_SUPPORT 487 , 488 fork_context_gnu 489 #endif 490 ); 491 } else { 492 __kmpc_end_serialized_parallel(&loc, gtid); 493 } 494 } 495 496 // Loop worksharing constructs 497 498 // The Gnu codegen passes in an exclusive upper bound for the overall range, 499 // but the libguide dispatch code expects an inclusive upper bound, hence the 500 // "end - incr" 5th argument to KMP_DISPATCH_INIT (and the " ub - str" 11th 501 // argument to __kmp_GOMP_fork_call). 502 // 503 // Conversely, KMP_DISPATCH_NEXT returns and inclusive upper bound in *p_ub, 504 // but the Gnu codegen expects an excluside upper bound, so the adjustment 505 // "*p_ub += stride" compenstates for the discrepancy. 506 // 507 // Correction: the gnu codegen always adjusts the upper bound by +-1, not the 508 // stride value. We adjust the dispatch parameters accordingly (by +-1), but 509 // we still adjust p_ub by the actual stride value. 510 // 511 // The "runtime" versions do not take a chunk_sz parameter. 512 // 513 // The profile lib cannot support construct checking of unordered loops that 514 // are predetermined by the compiler to be statically scheduled, as the gcc 515 // codegen will not always emit calls to GOMP_loop_static_next() to get the 516 // next iteration. Instead, it emits inline code to call omp_get_thread_num() 517 // num and calculate the iteration space using the result. It doesn't do this 518 // with ordered static loop, so they can be checked. 519 520 #if OMPT_SUPPORT 521 #define IF_OMPT_SUPPORT(code) code 522 #else 523 #define IF_OMPT_SUPPORT(code) 524 #endif 525 526 #define LOOP_START(func, schedule) \ 527 int func(long lb, long ub, long str, long chunk_sz, long *p_lb, \ 528 long *p_ub) { \ 529 int status; \ 530 long stride; \ 531 int gtid = __kmp_entry_gtid(); \ 532 MKLOC(loc, KMP_STR(func)); \ 533 KA_TRACE( \ 534 20, \ 535 (KMP_STR( \ 536 func) ": T#%d, lb 0x%lx, ub 0x%lx, str 0x%lx, chunk_sz 0x%lx\n", \ 537 gtid, lb, ub, str, chunk_sz)); \ 538 \ 539 if ((str > 0) ? (lb < ub) : (lb > ub)) { \ 540 IF_OMPT_SUPPORT(OMPT_STORE_RETURN_ADDRESS(gtid);) \ 541 KMP_DISPATCH_INIT(&loc, gtid, (schedule), lb, \ 542 (str > 0) ? (ub - 1) : (ub + 1), str, chunk_sz, \ 543 (schedule) != kmp_sch_static); \ 544 IF_OMPT_SUPPORT(OMPT_STORE_RETURN_ADDRESS(gtid);) \ 545 status = KMP_DISPATCH_NEXT(&loc, gtid, NULL, (kmp_int *)p_lb, \ 546 (kmp_int *)p_ub, (kmp_int *)&stride); \ 547 if (status) { \ 548 KMP_DEBUG_ASSERT(stride == str); \ 549 *p_ub += (str > 0) ? 1 : -1; \ 550 } \ 551 } else { \ 552 status = 0; \ 553 } \ 554 \ 555 KA_TRACE( \ 556 20, \ 557 (KMP_STR( \ 558 func) " exit: T#%d, *p_lb 0x%lx, *p_ub 0x%lx, returning %d\n", \ 559 gtid, *p_lb, *p_ub, status)); \ 560 return status; \ 561 } 562 563 #define LOOP_RUNTIME_START(func, schedule) \ 564 int func(long lb, long ub, long str, long *p_lb, long *p_ub) { \ 565 int status; \ 566 long stride; \ 567 long chunk_sz = 0; \ 568 int gtid = __kmp_entry_gtid(); \ 569 MKLOC(loc, KMP_STR(func)); \ 570 KA_TRACE( \ 571 20, \ 572 (KMP_STR(func) ": T#%d, lb 0x%lx, ub 0x%lx, str 0x%lx, chunk_sz %d\n", \ 573 gtid, lb, ub, str, chunk_sz)); \ 574 \ 575 if ((str > 0) ? (lb < ub) : (lb > ub)) { \ 576 IF_OMPT_SUPPORT(OMPT_STORE_RETURN_ADDRESS(gtid);) \ 577 KMP_DISPATCH_INIT(&loc, gtid, (schedule), lb, \ 578 (str > 0) ? (ub - 1) : (ub + 1), str, chunk_sz, TRUE); \ 579 IF_OMPT_SUPPORT(OMPT_STORE_RETURN_ADDRESS(gtid);) \ 580 status = KMP_DISPATCH_NEXT(&loc, gtid, NULL, (kmp_int *)p_lb, \ 581 (kmp_int *)p_ub, (kmp_int *)&stride); \ 582 if (status) { \ 583 KMP_DEBUG_ASSERT(stride == str); \ 584 *p_ub += (str > 0) ? 1 : -1; \ 585 } \ 586 } else { \ 587 status = 0; \ 588 } \ 589 \ 590 KA_TRACE( \ 591 20, \ 592 (KMP_STR( \ 593 func) " exit: T#%d, *p_lb 0x%lx, *p_ub 0x%lx, returning %d\n", \ 594 gtid, *p_lb, *p_ub, status)); \ 595 return status; \ 596 } 597 598 #if OMP_45_ENABLED 599 #define KMP_DOACROSS_FINI(status, gtid) \ 600 if (!status && __kmp_threads[gtid]->th.th_dispatch->th_doacross_flags) { \ 601 __kmpc_doacross_fini(NULL, gtid); \ 602 } 603 #else 604 #define KMP_DOACROSS_FINI(status, gtid) /* Nothing */ 605 #endif 606 607 #define LOOP_NEXT(func, fini_code) \ 608 int func(long *p_lb, long *p_ub) { \ 609 int status; \ 610 long stride; \ 611 int gtid = __kmp_get_gtid(); \ 612 MKLOC(loc, KMP_STR(func)); \ 613 KA_TRACE(20, (KMP_STR(func) ": T#%d\n", gtid)); \ 614 \ 615 IF_OMPT_SUPPORT(OMPT_STORE_RETURN_ADDRESS(gtid);) \ 616 fini_code status = KMP_DISPATCH_NEXT(&loc, gtid, NULL, (kmp_int *)p_lb, \ 617 (kmp_int *)p_ub, (kmp_int *)&stride); \ 618 if (status) { \ 619 *p_ub += (stride > 0) ? 1 : -1; \ 620 } \ 621 KMP_DOACROSS_FINI(status, gtid) \ 622 \ 623 KA_TRACE( \ 624 20, \ 625 (KMP_STR(func) " exit: T#%d, *p_lb 0x%lx, *p_ub 0x%lx, stride 0x%lx, " \ 626 "returning %d\n", \ 627 gtid, *p_lb, *p_ub, stride, status)); \ 628 return status; \ 629 } 630 631 LOOP_START(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_STATIC_START), kmp_sch_static) 632 LOOP_NEXT(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_STATIC_NEXT), {}) 633 LOOP_START(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_DYNAMIC_START), 634 kmp_sch_dynamic_chunked) 635 LOOP_NEXT(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_DYNAMIC_NEXT), {}) 636 LOOP_START(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_GUIDED_START), 637 kmp_sch_guided_chunked) 638 LOOP_NEXT(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_GUIDED_NEXT), {}) 639 LOOP_RUNTIME_START(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_RUNTIME_START), 640 kmp_sch_runtime) 641 LOOP_NEXT(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_RUNTIME_NEXT), {}) 642 643 LOOP_START(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ORDERED_STATIC_START), 644 kmp_ord_static) 645 LOOP_NEXT(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ORDERED_STATIC_NEXT), 646 { KMP_DISPATCH_FINI_CHUNK(&loc, gtid); }) 647 LOOP_START(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ORDERED_DYNAMIC_START), 648 kmp_ord_dynamic_chunked) 649 LOOP_NEXT(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ORDERED_DYNAMIC_NEXT), 650 { KMP_DISPATCH_FINI_CHUNK(&loc, gtid); }) 651 LOOP_START(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ORDERED_GUIDED_START), 652 kmp_ord_guided_chunked) 653 LOOP_NEXT(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ORDERED_GUIDED_NEXT), 654 { KMP_DISPATCH_FINI_CHUNK(&loc, gtid); }) 655 LOOP_RUNTIME_START( 656 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ORDERED_RUNTIME_START), 657 kmp_ord_runtime) 658 LOOP_NEXT(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ORDERED_RUNTIME_NEXT), 659 { KMP_DISPATCH_FINI_CHUNK(&loc, gtid); }) 660 661 #if OMP_45_ENABLED 662 #define LOOP_DOACROSS_START(func, schedule) \ 663 bool func(unsigned ncounts, long *counts, long chunk_sz, long *p_lb, \ 664 long *p_ub) { \ 665 int status; \ 666 long stride, lb, ub, str; \ 667 int gtid = __kmp_entry_gtid(); \ 668 struct kmp_dim *dims = \ 669 (struct kmp_dim *)__kmp_allocate(sizeof(struct kmp_dim) * ncounts); \ 670 MKLOC(loc, KMP_STR(func)); \ 671 for (unsigned i = 0; i < ncounts; ++i) { \ 672 dims[i].lo = 0; \ 673 dims[i].up = counts[i] - 1; \ 674 dims[i].st = 1; \ 675 } \ 676 __kmpc_doacross_init(&loc, gtid, (int)ncounts, dims); \ 677 lb = 0; \ 678 ub = counts[0]; \ 679 str = 1; \ 680 KA_TRACE(20, (KMP_STR(func) ": T#%d, ncounts %u, lb 0x%lx, ub 0x%lx, str " \ 681 "0x%lx, chunk_sz " \ 682 "0x%lx\n", \ 683 gtid, ncounts, lb, ub, str, chunk_sz)); \ 684 \ 685 if ((str > 0) ? (lb < ub) : (lb > ub)) { \ 686 KMP_DISPATCH_INIT(&loc, gtid, (schedule), lb, \ 687 (str > 0) ? (ub - 1) : (ub + 1), str, chunk_sz, \ 688 (schedule) != kmp_sch_static); \ 689 status = KMP_DISPATCH_NEXT(&loc, gtid, NULL, (kmp_int *)p_lb, \ 690 (kmp_int *)p_ub, (kmp_int *)&stride); \ 691 if (status) { \ 692 KMP_DEBUG_ASSERT(stride == str); \ 693 *p_ub += (str > 0) ? 1 : -1; \ 694 } \ 695 } else { \ 696 status = 0; \ 697 } \ 698 KMP_DOACROSS_FINI(status, gtid); \ 699 \ 700 KA_TRACE( \ 701 20, \ 702 (KMP_STR( \ 703 func) " exit: T#%d, *p_lb 0x%lx, *p_ub 0x%lx, returning %d\n", \ 704 gtid, *p_lb, *p_ub, status)); \ 705 __kmp_free(dims); \ 706 return status; \ 707 } 708 709 #define LOOP_DOACROSS_RUNTIME_START(func, schedule) \ 710 int func(unsigned ncounts, long *counts, long *p_lb, long *p_ub) { \ 711 int status; \ 712 long stride, lb, ub, str; \ 713 long chunk_sz = 0; \ 714 int gtid = __kmp_entry_gtid(); \ 715 struct kmp_dim *dims = \ 716 (struct kmp_dim *)__kmp_allocate(sizeof(struct kmp_dim) * ncounts); \ 717 MKLOC(loc, KMP_STR(func)); \ 718 for (unsigned i = 0; i < ncounts; ++i) { \ 719 dims[i].lo = 0; \ 720 dims[i].up = counts[i] - 1; \ 721 dims[i].st = 1; \ 722 } \ 723 __kmpc_doacross_init(&loc, gtid, (int)ncounts, dims); \ 724 lb = 0; \ 725 ub = counts[0]; \ 726 str = 1; \ 727 KA_TRACE( \ 728 20, \ 729 (KMP_STR(func) ": T#%d, lb 0x%lx, ub 0x%lx, str 0x%lx, chunk_sz %d\n", \ 730 gtid, lb, ub, str, chunk_sz)); \ 731 \ 732 if ((str > 0) ? (lb < ub) : (lb > ub)) { \ 733 KMP_DISPATCH_INIT(&loc, gtid, (schedule), lb, \ 734 (str > 0) ? (ub - 1) : (ub + 1), str, chunk_sz, TRUE); \ 735 status = KMP_DISPATCH_NEXT(&loc, gtid, NULL, (kmp_int *)p_lb, \ 736 (kmp_int *)p_ub, (kmp_int *)&stride); \ 737 if (status) { \ 738 KMP_DEBUG_ASSERT(stride == str); \ 739 *p_ub += (str > 0) ? 1 : -1; \ 740 } \ 741 } else { \ 742 status = 0; \ 743 } \ 744 KMP_DOACROSS_FINI(status, gtid); \ 745 \ 746 KA_TRACE( \ 747 20, \ 748 (KMP_STR( \ 749 func) " exit: T#%d, *p_lb 0x%lx, *p_ub 0x%lx, returning %d\n", \ 750 gtid, *p_lb, *p_ub, status)); \ 751 __kmp_free(dims); \ 752 return status; \ 753 } 754 755 LOOP_DOACROSS_START( 756 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_DOACROSS_STATIC_START), 757 kmp_sch_static) 758 LOOP_DOACROSS_START( 759 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_DOACROSS_DYNAMIC_START), 760 kmp_sch_dynamic_chunked) 761 LOOP_DOACROSS_START( 762 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_DOACROSS_GUIDED_START), 763 kmp_sch_guided_chunked) 764 LOOP_DOACROSS_RUNTIME_START( 765 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_DOACROSS_RUNTIME_START), 766 kmp_sch_runtime) 767 #endif // OMP_45_ENABLED 768 769 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_END)(void) { 770 int gtid = __kmp_get_gtid(); 771 KA_TRACE(20, ("GOMP_loop_end: T#%d\n", gtid)) 772 773 #if OMPT_SUPPORT && OMPT_OPTIONAL 774 ompt_frame_t *ompt_frame; 775 if (ompt_enabled.enabled) { 776 __ompt_get_task_info_internal(0, NULL, NULL, &ompt_frame, NULL, NULL); 777 ompt_frame->enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0); 778 OMPT_STORE_RETURN_ADDRESS(gtid); 779 } 780 #endif 781 __kmp_barrier(bs_plain_barrier, gtid, FALSE, 0, NULL, NULL); 782 #if OMPT_SUPPORT && OMPT_OPTIONAL 783 if (ompt_enabled.enabled) { 784 ompt_frame->enter_frame = ompt_data_none; 785 } 786 #endif 787 788 KA_TRACE(20, ("GOMP_loop_end exit: T#%d\n", gtid)) 789 } 790 791 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_END_NOWAIT)(void) { 792 KA_TRACE(20, ("GOMP_loop_end_nowait: T#%d\n", __kmp_get_gtid())) 793 } 794 795 // Unsigned long long loop worksharing constructs 796 // 797 // These are new with gcc 4.4 798 799 #define LOOP_START_ULL(func, schedule) \ 800 int func(int up, unsigned long long lb, unsigned long long ub, \ 801 unsigned long long str, unsigned long long chunk_sz, \ 802 unsigned long long *p_lb, unsigned long long *p_ub) { \ 803 int status; \ 804 long long str2 = up ? ((long long)str) : -((long long)str); \ 805 long long stride; \ 806 int gtid = __kmp_entry_gtid(); \ 807 MKLOC(loc, KMP_STR(func)); \ 808 \ 809 KA_TRACE(20, (KMP_STR(func) ": T#%d, up %d, lb 0x%llx, ub 0x%llx, str " \ 810 "0x%llx, chunk_sz 0x%llx\n", \ 811 gtid, up, lb, ub, str, chunk_sz)); \ 812 \ 813 if ((str > 0) ? (lb < ub) : (lb > ub)) { \ 814 KMP_DISPATCH_INIT_ULL(&loc, gtid, (schedule), lb, \ 815 (str2 > 0) ? (ub - 1) : (ub + 1), str2, chunk_sz, \ 816 (schedule) != kmp_sch_static); \ 817 status = \ 818 KMP_DISPATCH_NEXT_ULL(&loc, gtid, NULL, (kmp_uint64 *)p_lb, \ 819 (kmp_uint64 *)p_ub, (kmp_int64 *)&stride); \ 820 if (status) { \ 821 KMP_DEBUG_ASSERT(stride == str2); \ 822 *p_ub += (str > 0) ? 1 : -1; \ 823 } \ 824 } else { \ 825 status = 0; \ 826 } \ 827 \ 828 KA_TRACE( \ 829 20, \ 830 (KMP_STR( \ 831 func) " exit: T#%d, *p_lb 0x%llx, *p_ub 0x%llx, returning %d\n", \ 832 gtid, *p_lb, *p_ub, status)); \ 833 return status; \ 834 } 835 836 #define LOOP_RUNTIME_START_ULL(func, schedule) \ 837 int func(int up, unsigned long long lb, unsigned long long ub, \ 838 unsigned long long str, unsigned long long *p_lb, \ 839 unsigned long long *p_ub) { \ 840 int status; \ 841 long long str2 = up ? ((long long)str) : -((long long)str); \ 842 unsigned long long stride; \ 843 unsigned long long chunk_sz = 0; \ 844 int gtid = __kmp_entry_gtid(); \ 845 MKLOC(loc, KMP_STR(func)); \ 846 \ 847 KA_TRACE(20, (KMP_STR(func) ": T#%d, up %d, lb 0x%llx, ub 0x%llx, str " \ 848 "0x%llx, chunk_sz 0x%llx\n", \ 849 gtid, up, lb, ub, str, chunk_sz)); \ 850 \ 851 if ((str > 0) ? (lb < ub) : (lb > ub)) { \ 852 KMP_DISPATCH_INIT_ULL(&loc, gtid, (schedule), lb, \ 853 (str2 > 0) ? (ub - 1) : (ub + 1), str2, chunk_sz, \ 854 TRUE); \ 855 status = \ 856 KMP_DISPATCH_NEXT_ULL(&loc, gtid, NULL, (kmp_uint64 *)p_lb, \ 857 (kmp_uint64 *)p_ub, (kmp_int64 *)&stride); \ 858 if (status) { \ 859 KMP_DEBUG_ASSERT((long long)stride == str2); \ 860 *p_ub += (str > 0) ? 1 : -1; \ 861 } \ 862 } else { \ 863 status = 0; \ 864 } \ 865 \ 866 KA_TRACE( \ 867 20, \ 868 (KMP_STR( \ 869 func) " exit: T#%d, *p_lb 0x%llx, *p_ub 0x%llx, returning %d\n", \ 870 gtid, *p_lb, *p_ub, status)); \ 871 return status; \ 872 } 873 874 #define LOOP_NEXT_ULL(func, fini_code) \ 875 int func(unsigned long long *p_lb, unsigned long long *p_ub) { \ 876 int status; \ 877 long long stride; \ 878 int gtid = __kmp_get_gtid(); \ 879 MKLOC(loc, KMP_STR(func)); \ 880 KA_TRACE(20, (KMP_STR(func) ": T#%d\n", gtid)); \ 881 \ 882 fini_code status = \ 883 KMP_DISPATCH_NEXT_ULL(&loc, gtid, NULL, (kmp_uint64 *)p_lb, \ 884 (kmp_uint64 *)p_ub, (kmp_int64 *)&stride); \ 885 if (status) { \ 886 *p_ub += (stride > 0) ? 1 : -1; \ 887 } \ 888 \ 889 KA_TRACE( \ 890 20, \ 891 (KMP_STR( \ 892 func) " exit: T#%d, *p_lb 0x%llx, *p_ub 0x%llx, stride 0x%llx, " \ 893 "returning %d\n", \ 894 gtid, *p_lb, *p_ub, stride, status)); \ 895 return status; \ 896 } 897 898 LOOP_START_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_STATIC_START), 899 kmp_sch_static) 900 LOOP_NEXT_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_STATIC_NEXT), {}) 901 LOOP_START_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_DYNAMIC_START), 902 kmp_sch_dynamic_chunked) 903 LOOP_NEXT_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_DYNAMIC_NEXT), {}) 904 LOOP_START_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_GUIDED_START), 905 kmp_sch_guided_chunked) 906 LOOP_NEXT_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_GUIDED_NEXT), {}) 907 LOOP_RUNTIME_START_ULL( 908 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_RUNTIME_START), kmp_sch_runtime) 909 LOOP_NEXT_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_RUNTIME_NEXT), {}) 910 911 LOOP_START_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_STATIC_START), 912 kmp_ord_static) 913 LOOP_NEXT_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_STATIC_NEXT), 914 { KMP_DISPATCH_FINI_CHUNK_ULL(&loc, gtid); }) 915 LOOP_START_ULL( 916 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_DYNAMIC_START), 917 kmp_ord_dynamic_chunked) 918 LOOP_NEXT_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_DYNAMIC_NEXT), 919 { KMP_DISPATCH_FINI_CHUNK_ULL(&loc, gtid); }) 920 LOOP_START_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_GUIDED_START), 921 kmp_ord_guided_chunked) 922 LOOP_NEXT_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_GUIDED_NEXT), 923 { KMP_DISPATCH_FINI_CHUNK_ULL(&loc, gtid); }) 924 LOOP_RUNTIME_START_ULL( 925 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_RUNTIME_START), 926 kmp_ord_runtime) 927 LOOP_NEXT_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_RUNTIME_NEXT), 928 { KMP_DISPATCH_FINI_CHUNK_ULL(&loc, gtid); }) 929 930 #if OMP_45_ENABLED 931 #define LOOP_DOACROSS_START_ULL(func, schedule) \ 932 int func(unsigned ncounts, unsigned long long *counts, \ 933 unsigned long long chunk_sz, unsigned long long *p_lb, \ 934 unsigned long long *p_ub) { \ 935 int status; \ 936 long long stride, str, lb, ub; \ 937 int gtid = __kmp_entry_gtid(); \ 938 struct kmp_dim *dims = \ 939 (struct kmp_dim *)__kmp_allocate(sizeof(struct kmp_dim) * ncounts); \ 940 MKLOC(loc, KMP_STR(func)); \ 941 for (unsigned i = 0; i < ncounts; ++i) { \ 942 dims[i].lo = 0; \ 943 dims[i].up = counts[i] - 1; \ 944 dims[i].st = 1; \ 945 } \ 946 __kmpc_doacross_init(&loc, gtid, (int)ncounts, dims); \ 947 lb = 0; \ 948 ub = counts[0]; \ 949 str = 1; \ 950 \ 951 KA_TRACE(20, (KMP_STR(func) ": T#%d, lb 0x%llx, ub 0x%llx, str " \ 952 "0x%llx, chunk_sz 0x%llx\n", \ 953 gtid, lb, ub, str, chunk_sz)); \ 954 \ 955 if ((str > 0) ? (lb < ub) : (lb > ub)) { \ 956 KMP_DISPATCH_INIT_ULL(&loc, gtid, (schedule), lb, \ 957 (str > 0) ? (ub - 1) : (ub + 1), str, chunk_sz, \ 958 (schedule) != kmp_sch_static); \ 959 status = \ 960 KMP_DISPATCH_NEXT_ULL(&loc, gtid, NULL, (kmp_uint64 *)p_lb, \ 961 (kmp_uint64 *)p_ub, (kmp_int64 *)&stride); \ 962 if (status) { \ 963 KMP_DEBUG_ASSERT(stride == str); \ 964 *p_ub += (str > 0) ? 1 : -1; \ 965 } \ 966 } else { \ 967 status = 0; \ 968 } \ 969 KMP_DOACROSS_FINI(status, gtid); \ 970 \ 971 KA_TRACE( \ 972 20, \ 973 (KMP_STR( \ 974 func) " exit: T#%d, *p_lb 0x%llx, *p_ub 0x%llx, returning %d\n", \ 975 gtid, *p_lb, *p_ub, status)); \ 976 __kmp_free(dims); \ 977 return status; \ 978 } 979 980 #define LOOP_DOACROSS_RUNTIME_START_ULL(func, schedule) \ 981 int func(unsigned ncounts, unsigned long long *counts, \ 982 unsigned long long *p_lb, unsigned long long *p_ub) { \ 983 int status; \ 984 unsigned long long stride, str, lb, ub; \ 985 unsigned long long chunk_sz = 0; \ 986 int gtid = __kmp_entry_gtid(); \ 987 struct kmp_dim *dims = \ 988 (struct kmp_dim *)__kmp_allocate(sizeof(struct kmp_dim) * ncounts); \ 989 MKLOC(loc, KMP_STR(func)); \ 990 for (unsigned i = 0; i < ncounts; ++i) { \ 991 dims[i].lo = 0; \ 992 dims[i].up = counts[i] - 1; \ 993 dims[i].st = 1; \ 994 } \ 995 __kmpc_doacross_init(&loc, gtid, (int)ncounts, dims); \ 996 lb = 0; \ 997 ub = counts[0]; \ 998 str = 1; \ 999 KA_TRACE(20, (KMP_STR(func) ": T#%d, lb 0x%llx, ub 0x%llx, str " \ 1000 "0x%llx, chunk_sz 0x%llx\n", \ 1001 gtid, lb, ub, str, chunk_sz)); \ 1002 \ 1003 if ((str > 0) ? (lb < ub) : (lb > ub)) { \ 1004 KMP_DISPATCH_INIT_ULL(&loc, gtid, (schedule), lb, \ 1005 (str > 0) ? (ub - 1) : (ub + 1), str, chunk_sz, \ 1006 TRUE); \ 1007 status = \ 1008 KMP_DISPATCH_NEXT_ULL(&loc, gtid, NULL, (kmp_uint64 *)p_lb, \ 1009 (kmp_uint64 *)p_ub, (kmp_int64 *)&stride); \ 1010 if (status) { \ 1011 KMP_DEBUG_ASSERT(stride == str); \ 1012 *p_ub += (str > 0) ? 1 : -1; \ 1013 } \ 1014 } else { \ 1015 status = 0; \ 1016 } \ 1017 KMP_DOACROSS_FINI(status, gtid); \ 1018 \ 1019 KA_TRACE( \ 1020 20, \ 1021 (KMP_STR( \ 1022 func) " exit: T#%d, *p_lb 0x%llx, *p_ub 0x%llx, returning %d\n", \ 1023 gtid, *p_lb, *p_ub, status)); \ 1024 __kmp_free(dims); \ 1025 return status; \ 1026 } 1027 1028 LOOP_DOACROSS_START_ULL( 1029 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_DOACROSS_STATIC_START), 1030 kmp_sch_static) 1031 LOOP_DOACROSS_START_ULL( 1032 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_DOACROSS_DYNAMIC_START), 1033 kmp_sch_dynamic_chunked) 1034 LOOP_DOACROSS_START_ULL( 1035 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_DOACROSS_GUIDED_START), 1036 kmp_sch_guided_chunked) 1037 LOOP_DOACROSS_RUNTIME_START_ULL( 1038 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_DOACROSS_RUNTIME_START), 1039 kmp_sch_runtime) 1040 #endif 1041 1042 // Combined parallel / loop worksharing constructs 1043 // 1044 // There are no ull versions (yet). 1045 1046 #define PARALLEL_LOOP_START(func, schedule, ompt_pre, ompt_post) \ 1047 void func(void (*task)(void *), void *data, unsigned num_threads, long lb, \ 1048 long ub, long str, long chunk_sz) { \ 1049 int gtid = __kmp_entry_gtid(); \ 1050 MKLOC(loc, KMP_STR(func)); \ 1051 KA_TRACE( \ 1052 20, \ 1053 (KMP_STR( \ 1054 func) ": T#%d, lb 0x%lx, ub 0x%lx, str 0x%lx, chunk_sz 0x%lx\n", \ 1055 gtid, lb, ub, str, chunk_sz)); \ 1056 \ 1057 ompt_pre(); \ 1058 \ 1059 if (__kmpc_ok_to_fork(&loc) && (num_threads != 1)) { \ 1060 if (num_threads != 0) { \ 1061 __kmp_push_num_threads(&loc, gtid, num_threads); \ 1062 } \ 1063 __kmp_GOMP_fork_call(&loc, gtid, task, \ 1064 (microtask_t)__kmp_GOMP_parallel_microtask_wrapper, \ 1065 9, task, data, num_threads, &loc, (schedule), lb, \ 1066 (str > 0) ? (ub - 1) : (ub + 1), str, chunk_sz); \ 1067 IF_OMPT_SUPPORT(OMPT_STORE_RETURN_ADDRESS(gtid)); \ 1068 } else { \ 1069 __kmp_GOMP_serialized_parallel(&loc, gtid, task); \ 1070 IF_OMPT_SUPPORT(OMPT_STORE_RETURN_ADDRESS(gtid)); \ 1071 } \ 1072 \ 1073 KMP_DISPATCH_INIT(&loc, gtid, (schedule), lb, \ 1074 (str > 0) ? (ub - 1) : (ub + 1), str, chunk_sz, \ 1075 (schedule) != kmp_sch_static); \ 1076 \ 1077 ompt_post(); \ 1078 \ 1079 KA_TRACE(20, (KMP_STR(func) " exit: T#%d\n", gtid)); \ 1080 } 1081 1082 #if OMPT_SUPPORT && OMPT_OPTIONAL 1083 1084 #define OMPT_LOOP_PRE() \ 1085 ompt_frame_t *parent_frame; \ 1086 if (ompt_enabled.enabled) { \ 1087 __ompt_get_task_info_internal(0, NULL, NULL, &parent_frame, NULL, NULL); \ 1088 parent_frame->enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0); \ 1089 OMPT_STORE_RETURN_ADDRESS(gtid); \ 1090 } 1091 1092 #define OMPT_LOOP_POST() \ 1093 if (ompt_enabled.enabled) { \ 1094 parent_frame->enter_frame = ompt_data_none; \ 1095 } 1096 1097 #else 1098 1099 #define OMPT_LOOP_PRE() 1100 1101 #define OMPT_LOOP_POST() 1102 1103 #endif 1104 1105 PARALLEL_LOOP_START( 1106 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_LOOP_STATIC_START), 1107 kmp_sch_static, OMPT_LOOP_PRE, OMPT_LOOP_POST) 1108 PARALLEL_LOOP_START( 1109 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_LOOP_DYNAMIC_START), 1110 kmp_sch_dynamic_chunked, OMPT_LOOP_PRE, OMPT_LOOP_POST) 1111 PARALLEL_LOOP_START( 1112 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_LOOP_GUIDED_START), 1113 kmp_sch_guided_chunked, OMPT_LOOP_PRE, OMPT_LOOP_POST) 1114 PARALLEL_LOOP_START( 1115 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_LOOP_RUNTIME_START), 1116 kmp_sch_runtime, OMPT_LOOP_PRE, OMPT_LOOP_POST) 1117 1118 // Tasking constructs 1119 1120 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASK)(void (*func)(void *), void *data, 1121 void (*copy_func)(void *, void *), 1122 long arg_size, long arg_align, 1123 bool if_cond, unsigned gomp_flags 1124 #if OMP_40_ENABLED 1125 , 1126 void **depend 1127 #endif 1128 ) { 1129 MKLOC(loc, "GOMP_task"); 1130 int gtid = __kmp_entry_gtid(); 1131 kmp_int32 flags = 0; 1132 kmp_tasking_flags_t *input_flags = (kmp_tasking_flags_t *)&flags; 1133 1134 KA_TRACE(20, ("GOMP_task: T#%d\n", gtid)); 1135 1136 // The low-order bit is the "untied" flag 1137 if (!(gomp_flags & 1)) { 1138 input_flags->tiedness = 1; 1139 } 1140 // The second low-order bit is the "final" flag 1141 if (gomp_flags & 2) { 1142 input_flags->final = 1; 1143 } 1144 input_flags->native = 1; 1145 // __kmp_task_alloc() sets up all other flags 1146 1147 if (!if_cond) { 1148 arg_size = 0; 1149 } 1150 1151 kmp_task_t *task = __kmp_task_alloc( 1152 &loc, gtid, input_flags, sizeof(kmp_task_t), 1153 arg_size ? arg_size + arg_align - 1 : 0, (kmp_routine_entry_t)func); 1154 1155 if (arg_size > 0) { 1156 if (arg_align > 0) { 1157 task->shareds = (void *)((((size_t)task->shareds) + arg_align - 1) / 1158 arg_align * arg_align); 1159 } 1160 // else error?? 1161 1162 if (copy_func) { 1163 (*copy_func)(task->shareds, data); 1164 } else { 1165 KMP_MEMCPY(task->shareds, data, arg_size); 1166 } 1167 } 1168 1169 #if OMPT_SUPPORT 1170 kmp_taskdata_t *current_task; 1171 if (ompt_enabled.enabled) { 1172 OMPT_STORE_RETURN_ADDRESS(gtid); 1173 current_task = __kmp_threads[gtid]->th.th_current_task; 1174 current_task->ompt_task_info.frame.enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0); 1175 } 1176 #endif 1177 1178 if (if_cond) { 1179 #if OMP_40_ENABLED 1180 if (gomp_flags & 8) { 1181 KMP_ASSERT(depend); 1182 const size_t ndeps = (kmp_intptr_t)depend[0]; 1183 const size_t nout = (kmp_intptr_t)depend[1]; 1184 kmp_depend_info_t dep_list[ndeps]; 1185 1186 for (size_t i = 0U; i < ndeps; i++) { 1187 dep_list[i].base_addr = (kmp_intptr_t)depend[2U + i]; 1188 dep_list[i].len = 0U; 1189 dep_list[i].flags.in = 1; 1190 dep_list[i].flags.out = (i < nout); 1191 } 1192 __kmpc_omp_task_with_deps(&loc, gtid, task, ndeps, dep_list, 0, NULL); 1193 } else { 1194 #endif 1195 __kmpc_omp_task(&loc, gtid, task); 1196 } 1197 } else { 1198 #if OMPT_SUPPORT 1199 ompt_thread_info_t oldInfo; 1200 kmp_info_t *thread; 1201 kmp_taskdata_t *taskdata; 1202 if (ompt_enabled.enabled) { 1203 // Store the threads states and restore them after the task 1204 thread = __kmp_threads[gtid]; 1205 taskdata = KMP_TASK_TO_TASKDATA(task); 1206 oldInfo = thread->th.ompt_thread_info; 1207 thread->th.ompt_thread_info.wait_id = 0; 1208 thread->th.ompt_thread_info.state = ompt_state_work_parallel; 1209 taskdata->ompt_task_info.frame.exit_frame.ptr = OMPT_GET_FRAME_ADDRESS(0); 1210 OMPT_STORE_RETURN_ADDRESS(gtid); 1211 } 1212 #endif 1213 1214 __kmpc_omp_task_begin_if0(&loc, gtid, task); 1215 func(data); 1216 __kmpc_omp_task_complete_if0(&loc, gtid, task); 1217 1218 #if OMPT_SUPPORT 1219 if (ompt_enabled.enabled) { 1220 thread->th.ompt_thread_info = oldInfo; 1221 taskdata->ompt_task_info.frame.exit_frame = ompt_data_none; 1222 } 1223 #endif 1224 } 1225 #if OMPT_SUPPORT 1226 if (ompt_enabled.enabled) { 1227 current_task->ompt_task_info.frame.enter_frame = ompt_data_none; 1228 } 1229 #endif 1230 1231 KA_TRACE(20, ("GOMP_task exit: T#%d\n", gtid)); 1232 } 1233 1234 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASKWAIT)(void) { 1235 MKLOC(loc, "GOMP_taskwait"); 1236 int gtid = __kmp_entry_gtid(); 1237 1238 #if OMPT_SUPPORT 1239 if (ompt_enabled.enabled) 1240 OMPT_STORE_RETURN_ADDRESS(gtid); 1241 #endif 1242 1243 KA_TRACE(20, ("GOMP_taskwait: T#%d\n", gtid)); 1244 1245 __kmpc_omp_taskwait(&loc, gtid); 1246 1247 KA_TRACE(20, ("GOMP_taskwait exit: T#%d\n", gtid)); 1248 } 1249 1250 // Sections worksharing constructs 1251 // 1252 // For the sections construct, we initialize a dynamically scheduled loop 1253 // worksharing construct with lb 1 and stride 1, and use the iteration #'s 1254 // that its returns as sections ids. 1255 // 1256 // There are no special entry points for ordered sections, so we always use 1257 // the dynamically scheduled workshare, even if the sections aren't ordered. 1258 1259 unsigned KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SECTIONS_START)(unsigned count) { 1260 int status; 1261 kmp_int lb, ub, stride; 1262 int gtid = __kmp_entry_gtid(); 1263 MKLOC(loc, "GOMP_sections_start"); 1264 KA_TRACE(20, ("GOMP_sections_start: T#%d\n", gtid)); 1265 1266 KMP_DISPATCH_INIT(&loc, gtid, kmp_nm_dynamic_chunked, 1, count, 1, 1, TRUE); 1267 1268 status = KMP_DISPATCH_NEXT(&loc, gtid, NULL, &lb, &ub, &stride); 1269 if (status) { 1270 KMP_DEBUG_ASSERT(stride == 1); 1271 KMP_DEBUG_ASSERT(lb > 0); 1272 KMP_ASSERT(lb == ub); 1273 } else { 1274 lb = 0; 1275 } 1276 1277 KA_TRACE(20, ("GOMP_sections_start exit: T#%d returning %u\n", gtid, 1278 (unsigned)lb)); 1279 return (unsigned)lb; 1280 } 1281 1282 unsigned KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SECTIONS_NEXT)(void) { 1283 int status; 1284 kmp_int lb, ub, stride; 1285 int gtid = __kmp_get_gtid(); 1286 MKLOC(loc, "GOMP_sections_next"); 1287 KA_TRACE(20, ("GOMP_sections_next: T#%d\n", gtid)); 1288 1289 #if OMPT_SUPPORT 1290 OMPT_STORE_RETURN_ADDRESS(gtid); 1291 #endif 1292 1293 status = KMP_DISPATCH_NEXT(&loc, gtid, NULL, &lb, &ub, &stride); 1294 if (status) { 1295 KMP_DEBUG_ASSERT(stride == 1); 1296 KMP_DEBUG_ASSERT(lb > 0); 1297 KMP_ASSERT(lb == ub); 1298 } else { 1299 lb = 0; 1300 } 1301 1302 KA_TRACE( 1303 20, ("GOMP_sections_next exit: T#%d returning %u\n", gtid, (unsigned)lb)); 1304 return (unsigned)lb; 1305 } 1306 1307 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_SECTIONS_START)( 1308 void (*task)(void *), void *data, unsigned num_threads, unsigned count) { 1309 int gtid = __kmp_entry_gtid(); 1310 1311 #if OMPT_SUPPORT 1312 ompt_frame_t *parent_frame; 1313 1314 if (ompt_enabled.enabled) { 1315 __ompt_get_task_info_internal(0, NULL, NULL, &parent_frame, NULL, NULL); 1316 parent_frame->enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0); 1317 OMPT_STORE_RETURN_ADDRESS(gtid); 1318 } 1319 #endif 1320 1321 MKLOC(loc, "GOMP_parallel_sections_start"); 1322 KA_TRACE(20, ("GOMP_parallel_sections_start: T#%d\n", gtid)); 1323 1324 if (__kmpc_ok_to_fork(&loc) && (num_threads != 1)) { 1325 if (num_threads != 0) { 1326 __kmp_push_num_threads(&loc, gtid, num_threads); 1327 } 1328 __kmp_GOMP_fork_call(&loc, gtid, task, 1329 (microtask_t)__kmp_GOMP_parallel_microtask_wrapper, 9, 1330 task, data, num_threads, &loc, kmp_nm_dynamic_chunked, 1331 (kmp_int)1, (kmp_int)count, (kmp_int)1, (kmp_int)1); 1332 } else { 1333 __kmp_GOMP_serialized_parallel(&loc, gtid, task); 1334 } 1335 1336 #if OMPT_SUPPORT 1337 if (ompt_enabled.enabled) { 1338 parent_frame->enter_frame = ompt_data_none; 1339 } 1340 #endif 1341 1342 KMP_DISPATCH_INIT(&loc, gtid, kmp_nm_dynamic_chunked, 1, count, 1, 1, TRUE); 1343 1344 KA_TRACE(20, ("GOMP_parallel_sections_start exit: T#%d\n", gtid)); 1345 } 1346 1347 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SECTIONS_END)(void) { 1348 int gtid = __kmp_get_gtid(); 1349 KA_TRACE(20, ("GOMP_sections_end: T#%d\n", gtid)) 1350 1351 #if OMPT_SUPPORT 1352 ompt_frame_t *ompt_frame; 1353 if (ompt_enabled.enabled) { 1354 __ompt_get_task_info_internal(0, NULL, NULL, &ompt_frame, NULL, NULL); 1355 ompt_frame->enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0); 1356 OMPT_STORE_RETURN_ADDRESS(gtid); 1357 } 1358 #endif 1359 __kmp_barrier(bs_plain_barrier, gtid, FALSE, 0, NULL, NULL); 1360 #if OMPT_SUPPORT 1361 if (ompt_enabled.enabled) { 1362 ompt_frame->enter_frame = ompt_data_none; 1363 } 1364 #endif 1365 1366 KA_TRACE(20, ("GOMP_sections_end exit: T#%d\n", gtid)) 1367 } 1368 1369 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SECTIONS_END_NOWAIT)(void) { 1370 KA_TRACE(20, ("GOMP_sections_end_nowait: T#%d\n", __kmp_get_gtid())) 1371 } 1372 1373 // libgomp has an empty function for GOMP_taskyield as of 2013-10-10 1374 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASKYIELD)(void) { 1375 KA_TRACE(20, ("GOMP_taskyield: T#%d\n", __kmp_get_gtid())) 1376 return; 1377 } 1378 1379 #if OMP_40_ENABLED // these are new GOMP_4.0 entry points 1380 1381 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL)(void (*task)(void *), 1382 void *data, 1383 unsigned num_threads, 1384 unsigned int flags) { 1385 int gtid = __kmp_entry_gtid(); 1386 MKLOC(loc, "GOMP_parallel"); 1387 KA_TRACE(20, ("GOMP_parallel: T#%d\n", gtid)); 1388 1389 #if OMPT_SUPPORT 1390 ompt_task_info_t *parent_task_info, *task_info; 1391 if (ompt_enabled.enabled) { 1392 parent_task_info = __ompt_get_task_info_object(0); 1393 parent_task_info->frame.enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0); 1394 OMPT_STORE_RETURN_ADDRESS(gtid); 1395 } 1396 #endif 1397 if (__kmpc_ok_to_fork(&loc) && (num_threads != 1)) { 1398 if (num_threads != 0) { 1399 __kmp_push_num_threads(&loc, gtid, num_threads); 1400 } 1401 if (flags != 0) { 1402 __kmp_push_proc_bind(&loc, gtid, (kmp_proc_bind_t)flags); 1403 } 1404 __kmp_GOMP_fork_call(&loc, gtid, task, 1405 (microtask_t)__kmp_GOMP_microtask_wrapper, 2, task, 1406 data); 1407 } else { 1408 __kmp_GOMP_serialized_parallel(&loc, gtid, task); 1409 } 1410 #if OMPT_SUPPORT 1411 if (ompt_enabled.enabled) { 1412 task_info = __ompt_get_task_info_object(0); 1413 task_info->frame.exit_frame.ptr = OMPT_GET_FRAME_ADDRESS(0); 1414 } 1415 #endif 1416 task(data); 1417 #if OMPT_SUPPORT 1418 if (ompt_enabled.enabled) { 1419 OMPT_STORE_RETURN_ADDRESS(gtid); 1420 } 1421 #endif 1422 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_END)(); 1423 #if OMPT_SUPPORT 1424 if (ompt_enabled.enabled) { 1425 task_info->frame.exit_frame = ompt_data_none; 1426 parent_task_info->frame.enter_frame = ompt_data_none; 1427 } 1428 #endif 1429 } 1430 1431 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_SECTIONS)(void (*task)(void *), 1432 void *data, 1433 unsigned num_threads, 1434 unsigned count, 1435 unsigned flags) { 1436 int gtid = __kmp_entry_gtid(); 1437 MKLOC(loc, "GOMP_parallel_sections"); 1438 KA_TRACE(20, ("GOMP_parallel_sections: T#%d\n", gtid)); 1439 1440 #if OMPT_SUPPORT 1441 OMPT_STORE_RETURN_ADDRESS(gtid); 1442 #endif 1443 1444 if (__kmpc_ok_to_fork(&loc) && (num_threads != 1)) { 1445 if (num_threads != 0) { 1446 __kmp_push_num_threads(&loc, gtid, num_threads); 1447 } 1448 if (flags != 0) { 1449 __kmp_push_proc_bind(&loc, gtid, (kmp_proc_bind_t)flags); 1450 } 1451 __kmp_GOMP_fork_call(&loc, gtid, task, 1452 (microtask_t)__kmp_GOMP_parallel_microtask_wrapper, 9, 1453 task, data, num_threads, &loc, kmp_nm_dynamic_chunked, 1454 (kmp_int)1, (kmp_int)count, (kmp_int)1, (kmp_int)1); 1455 } else { 1456 __kmp_GOMP_serialized_parallel(&loc, gtid, task); 1457 } 1458 1459 #if OMPT_SUPPORT 1460 OMPT_STORE_RETURN_ADDRESS(gtid); 1461 #endif 1462 1463 KMP_DISPATCH_INIT(&loc, gtid, kmp_nm_dynamic_chunked, 1, count, 1, 1, TRUE); 1464 1465 task(data); 1466 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_END)(); 1467 KA_TRACE(20, ("GOMP_parallel_sections exit: T#%d\n", gtid)); 1468 } 1469 1470 #define PARALLEL_LOOP(func, schedule, ompt_pre, ompt_post) \ 1471 void func(void (*task)(void *), void *data, unsigned num_threads, long lb, \ 1472 long ub, long str, long chunk_sz, unsigned flags) { \ 1473 int gtid = __kmp_entry_gtid(); \ 1474 MKLOC(loc, KMP_STR(func)); \ 1475 KA_TRACE( \ 1476 20, \ 1477 (KMP_STR( \ 1478 func) ": T#%d, lb 0x%lx, ub 0x%lx, str 0x%lx, chunk_sz 0x%lx\n", \ 1479 gtid, lb, ub, str, chunk_sz)); \ 1480 \ 1481 ompt_pre(); \ 1482 if (__kmpc_ok_to_fork(&loc) && (num_threads != 1)) { \ 1483 if (num_threads != 0) { \ 1484 __kmp_push_num_threads(&loc, gtid, num_threads); \ 1485 } \ 1486 if (flags != 0) { \ 1487 __kmp_push_proc_bind(&loc, gtid, (kmp_proc_bind_t)flags); \ 1488 } \ 1489 __kmp_GOMP_fork_call(&loc, gtid, task, \ 1490 (microtask_t)__kmp_GOMP_parallel_microtask_wrapper, \ 1491 9, task, data, num_threads, &loc, (schedule), lb, \ 1492 (str > 0) ? (ub - 1) : (ub + 1), str, chunk_sz); \ 1493 } else { \ 1494 __kmp_GOMP_serialized_parallel(&loc, gtid, task); \ 1495 } \ 1496 \ 1497 IF_OMPT_SUPPORT(OMPT_STORE_RETURN_ADDRESS(gtid);) \ 1498 KMP_DISPATCH_INIT(&loc, gtid, (schedule), lb, \ 1499 (str > 0) ? (ub - 1) : (ub + 1), str, chunk_sz, \ 1500 (schedule) != kmp_sch_static); \ 1501 task(data); \ 1502 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_END)(); \ 1503 ompt_post(); \ 1504 \ 1505 KA_TRACE(20, (KMP_STR(func) " exit: T#%d\n", gtid)); \ 1506 } 1507 1508 PARALLEL_LOOP(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_LOOP_STATIC), 1509 kmp_sch_static, OMPT_LOOP_PRE, OMPT_LOOP_POST) 1510 PARALLEL_LOOP(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_LOOP_DYNAMIC), 1511 kmp_sch_dynamic_chunked, OMPT_LOOP_PRE, OMPT_LOOP_POST) 1512 PARALLEL_LOOP(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_LOOP_GUIDED), 1513 kmp_sch_guided_chunked, OMPT_LOOP_PRE, OMPT_LOOP_POST) 1514 PARALLEL_LOOP(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_LOOP_RUNTIME), 1515 kmp_sch_runtime, OMPT_LOOP_PRE, OMPT_LOOP_POST) 1516 1517 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASKGROUP_START)(void) { 1518 int gtid = __kmp_entry_gtid(); 1519 MKLOC(loc, "GOMP_taskgroup_start"); 1520 KA_TRACE(20, ("GOMP_taskgroup_start: T#%d\n", gtid)); 1521 1522 #if OMPT_SUPPORT 1523 if (ompt_enabled.enabled) 1524 OMPT_STORE_RETURN_ADDRESS(gtid); 1525 #endif 1526 1527 __kmpc_taskgroup(&loc, gtid); 1528 1529 return; 1530 } 1531 1532 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASKGROUP_END)(void) { 1533 int gtid = __kmp_get_gtid(); 1534 MKLOC(loc, "GOMP_taskgroup_end"); 1535 KA_TRACE(20, ("GOMP_taskgroup_end: T#%d\n", gtid)); 1536 1537 #if OMPT_SUPPORT 1538 if (ompt_enabled.enabled) 1539 OMPT_STORE_RETURN_ADDRESS(gtid); 1540 #endif 1541 1542 __kmpc_end_taskgroup(&loc, gtid); 1543 1544 return; 1545 } 1546 1547 static kmp_int32 __kmp_gomp_to_omp_cancellation_kind(int gomp_kind) { 1548 kmp_int32 cncl_kind = 0; 1549 switch (gomp_kind) { 1550 case 1: 1551 cncl_kind = cancel_parallel; 1552 break; 1553 case 2: 1554 cncl_kind = cancel_loop; 1555 break; 1556 case 4: 1557 cncl_kind = cancel_sections; 1558 break; 1559 case 8: 1560 cncl_kind = cancel_taskgroup; 1561 break; 1562 } 1563 return cncl_kind; 1564 } 1565 1566 // Return true if cancellation should take place, false otherwise 1567 bool KMP_EXPAND_NAME(KMP_API_NAME_GOMP_CANCELLATION_POINT)(int which) { 1568 int gtid = __kmp_get_gtid(); 1569 MKLOC(loc, "GOMP_cancellation_point"); 1570 KA_TRACE(20, ("GOMP_cancellation_point: T#%d which:%d\n", gtid, which)); 1571 kmp_int32 cncl_kind = __kmp_gomp_to_omp_cancellation_kind(which); 1572 return __kmpc_cancellationpoint(&loc, gtid, cncl_kind); 1573 } 1574 1575 // Return true if cancellation should take place, false otherwise 1576 bool KMP_EXPAND_NAME(KMP_API_NAME_GOMP_CANCEL)(int which, bool do_cancel) { 1577 int gtid = __kmp_get_gtid(); 1578 MKLOC(loc, "GOMP_cancel"); 1579 KA_TRACE(20, ("GOMP_cancel: T#%d which:%d do_cancel:%d\n", gtid, which, 1580 (int)do_cancel)); 1581 kmp_int32 cncl_kind = __kmp_gomp_to_omp_cancellation_kind(which); 1582 1583 if (do_cancel == FALSE) { 1584 return __kmpc_cancellationpoint(&loc, gtid, cncl_kind); 1585 } else { 1586 return __kmpc_cancel(&loc, gtid, cncl_kind); 1587 } 1588 } 1589 1590 // Return true if cancellation should take place, false otherwise 1591 bool KMP_EXPAND_NAME(KMP_API_NAME_GOMP_BARRIER_CANCEL)(void) { 1592 int gtid = __kmp_get_gtid(); 1593 KA_TRACE(20, ("GOMP_barrier_cancel: T#%d\n", gtid)); 1594 return __kmp_barrier_gomp_cancel(gtid); 1595 } 1596 1597 // Return true if cancellation should take place, false otherwise 1598 bool KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SECTIONS_END_CANCEL)(void) { 1599 int gtid = __kmp_get_gtid(); 1600 KA_TRACE(20, ("GOMP_sections_end_cancel: T#%d\n", gtid)); 1601 return __kmp_barrier_gomp_cancel(gtid); 1602 } 1603 1604 // Return true if cancellation should take place, false otherwise 1605 bool KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_END_CANCEL)(void) { 1606 int gtid = __kmp_get_gtid(); 1607 KA_TRACE(20, ("GOMP_loop_end_cancel: T#%d\n", gtid)); 1608 return __kmp_barrier_gomp_cancel(gtid); 1609 } 1610 1611 // All target functions are empty as of 2014-05-29 1612 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TARGET)(int device, void (*fn)(void *), 1613 const void *openmp_target, 1614 size_t mapnum, void **hostaddrs, 1615 size_t *sizes, 1616 unsigned char *kinds) { 1617 return; 1618 } 1619 1620 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TARGET_DATA)( 1621 int device, const void *openmp_target, size_t mapnum, void **hostaddrs, 1622 size_t *sizes, unsigned char *kinds) { 1623 return; 1624 } 1625 1626 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TARGET_END_DATA)(void) { return; } 1627 1628 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TARGET_UPDATE)( 1629 int device, const void *openmp_target, size_t mapnum, void **hostaddrs, 1630 size_t *sizes, unsigned char *kinds) { 1631 return; 1632 } 1633 1634 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TEAMS)(unsigned int num_teams, 1635 unsigned int thread_limit) { 1636 return; 1637 } 1638 #endif // OMP_40_ENABLED 1639 1640 #if OMP_45_ENABLED 1641 1642 // Task duplication function which copies src to dest (both are 1643 // preallocated task structures) 1644 static void __kmp_gomp_task_dup(kmp_task_t *dest, kmp_task_t *src, 1645 kmp_int32 last_private) { 1646 kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(src); 1647 if (taskdata->td_copy_func) { 1648 (taskdata->td_copy_func)(dest->shareds, src->shareds); 1649 } 1650 } 1651 1652 #ifdef __cplusplus 1653 } // extern "C" 1654 #endif 1655 1656 template <typename T> 1657 void __GOMP_taskloop(void (*func)(void *), void *data, 1658 void (*copy_func)(void *, void *), long arg_size, 1659 long arg_align, unsigned gomp_flags, 1660 unsigned long num_tasks, int priority, T start, T end, 1661 T step) { 1662 typedef void (*p_task_dup_t)(kmp_task_t *, kmp_task_t *, kmp_int32); 1663 MKLOC(loc, "GOMP_taskloop"); 1664 int sched; 1665 T *loop_bounds; 1666 int gtid = __kmp_entry_gtid(); 1667 kmp_int32 flags = 0; 1668 int if_val = gomp_flags & (1u << 10); 1669 int nogroup = gomp_flags & (1u << 11); 1670 int up = gomp_flags & (1u << 8); 1671 p_task_dup_t task_dup = NULL; 1672 kmp_tasking_flags_t *input_flags = (kmp_tasking_flags_t *)&flags; 1673 #ifdef KMP_DEBUG 1674 { 1675 char *buff; 1676 buff = __kmp_str_format( 1677 "GOMP_taskloop: T#%%d: func:%%p data:%%p copy_func:%%p " 1678 "arg_size:%%ld arg_align:%%ld gomp_flags:0x%%x num_tasks:%%lu " 1679 "priority:%%d start:%%%s end:%%%s step:%%%s\n", 1680 traits_t<T>::spec, traits_t<T>::spec, traits_t<T>::spec); 1681 KA_TRACE(20, (buff, gtid, func, data, copy_func, arg_size, arg_align, 1682 gomp_flags, num_tasks, priority, start, end, step)); 1683 __kmp_str_free(&buff); 1684 } 1685 #endif 1686 KMP_ASSERT((size_t)arg_size >= 2 * sizeof(T)); 1687 KMP_ASSERT(arg_align > 0); 1688 // The low-order bit is the "untied" flag 1689 if (!(gomp_flags & 1)) { 1690 input_flags->tiedness = 1; 1691 } 1692 // The second low-order bit is the "final" flag 1693 if (gomp_flags & 2) { 1694 input_flags->final = 1; 1695 } 1696 // Negative step flag 1697 if (!up) { 1698 // If step is flagged as negative, but isn't properly sign extended 1699 // Then manually sign extend it. Could be a short, int, char embedded 1700 // in a long. So cannot assume any cast. 1701 if (step > 0) { 1702 for (int i = sizeof(T) * CHAR_BIT - 1; i >= 0L; --i) { 1703 // break at the first 1 bit 1704 if (step & ((T)1 << i)) 1705 break; 1706 step |= ((T)1 << i); 1707 } 1708 } 1709 } 1710 input_flags->native = 1; 1711 // Figure out if none/grainsize/num_tasks clause specified 1712 if (num_tasks > 0) { 1713 if (gomp_flags & (1u << 9)) 1714 sched = 1; // grainsize specified 1715 else 1716 sched = 2; // num_tasks specified 1717 // neither grainsize nor num_tasks specified 1718 } else { 1719 sched = 0; 1720 } 1721 1722 // __kmp_task_alloc() sets up all other flags 1723 kmp_task_t *task = 1724 __kmp_task_alloc(&loc, gtid, input_flags, sizeof(kmp_task_t), 1725 arg_size + arg_align - 1, (kmp_routine_entry_t)func); 1726 kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task); 1727 taskdata->td_copy_func = copy_func; 1728 taskdata->td_size_loop_bounds = sizeof(T); 1729 1730 // re-align shareds if needed and setup firstprivate copy constructors 1731 // through the task_dup mechanism 1732 task->shareds = (void *)((((size_t)task->shareds) + arg_align - 1) / 1733 arg_align * arg_align); 1734 if (copy_func) { 1735 task_dup = __kmp_gomp_task_dup; 1736 } 1737 KMP_MEMCPY(task->shareds, data, arg_size); 1738 1739 loop_bounds = (T *)task->shareds; 1740 loop_bounds[0] = start; 1741 loop_bounds[1] = end + (up ? -1 : 1); 1742 __kmpc_taskloop(&loc, gtid, task, if_val, (kmp_uint64 *)&(loop_bounds[0]), 1743 (kmp_uint64 *)&(loop_bounds[1]), (kmp_int64)step, nogroup, 1744 sched, (kmp_uint64)num_tasks, (void *)task_dup); 1745 } 1746 1747 // 4 byte version of GOMP_doacross_post 1748 // This verison needs to create a temporary array which converts 4 byte 1749 // integers into 8 byte integeres 1750 template <typename T, bool need_conversion = (sizeof(long) == 4)> 1751 void __kmp_GOMP_doacross_post(T *count); 1752 1753 template <> void __kmp_GOMP_doacross_post<long, true>(long *count) { 1754 int gtid = __kmp_entry_gtid(); 1755 kmp_info_t *th = __kmp_threads[gtid]; 1756 MKLOC(loc, "GOMP_doacross_post"); 1757 kmp_int64 num_dims = th->th.th_dispatch->th_doacross_info[0]; 1758 kmp_int64 *vec = 1759 (kmp_int64 *)__kmp_thread_malloc(th, sizeof(kmp_int64) * num_dims); 1760 for (kmp_int64 i = 0; i < num_dims; ++i) { 1761 vec[i] = (kmp_int64)count[i]; 1762 } 1763 __kmpc_doacross_post(&loc, gtid, vec); 1764 __kmp_thread_free(th, vec); 1765 } 1766 1767 // 8 byte versions of GOMP_doacross_post 1768 // This version can just pass in the count array directly instead of creating 1769 // a temporary array 1770 template <> void __kmp_GOMP_doacross_post<long, false>(long *count) { 1771 int gtid = __kmp_entry_gtid(); 1772 MKLOC(loc, "GOMP_doacross_post"); 1773 __kmpc_doacross_post(&loc, gtid, RCAST(kmp_int64 *, count)); 1774 } 1775 1776 template <typename T> void __kmp_GOMP_doacross_wait(T first, va_list args) { 1777 int gtid = __kmp_entry_gtid(); 1778 kmp_info_t *th = __kmp_threads[gtid]; 1779 MKLOC(loc, "GOMP_doacross_wait"); 1780 kmp_int64 num_dims = th->th.th_dispatch->th_doacross_info[0]; 1781 kmp_int64 *vec = 1782 (kmp_int64 *)__kmp_thread_malloc(th, sizeof(kmp_int64) * num_dims); 1783 vec[0] = (kmp_int64)first; 1784 for (kmp_int64 i = 1; i < num_dims; ++i) { 1785 T item = va_arg(args, T); 1786 vec[i] = (kmp_int64)item; 1787 } 1788 __kmpc_doacross_wait(&loc, gtid, vec); 1789 __kmp_thread_free(th, vec); 1790 return; 1791 } 1792 1793 #ifdef __cplusplus 1794 extern "C" { 1795 #endif // __cplusplus 1796 1797 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASKLOOP)( 1798 void (*func)(void *), void *data, void (*copy_func)(void *, void *), 1799 long arg_size, long arg_align, unsigned gomp_flags, unsigned long num_tasks, 1800 int priority, long start, long end, long step) { 1801 __GOMP_taskloop<long>(func, data, copy_func, arg_size, arg_align, gomp_flags, 1802 num_tasks, priority, start, end, step); 1803 } 1804 1805 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASKLOOP_ULL)( 1806 void (*func)(void *), void *data, void (*copy_func)(void *, void *), 1807 long arg_size, long arg_align, unsigned gomp_flags, unsigned long num_tasks, 1808 int priority, unsigned long long start, unsigned long long end, 1809 unsigned long long step) { 1810 __GOMP_taskloop<unsigned long long>(func, data, copy_func, arg_size, 1811 arg_align, gomp_flags, num_tasks, 1812 priority, start, end, step); 1813 } 1814 1815 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_DOACROSS_POST)(long *count) { 1816 __kmp_GOMP_doacross_post(count); 1817 } 1818 1819 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_DOACROSS_WAIT)(long first, ...) { 1820 va_list args; 1821 va_start(args, first); 1822 __kmp_GOMP_doacross_wait<long>(first, args); 1823 va_end(args); 1824 } 1825 1826 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_DOACROSS_ULL_POST)( 1827 unsigned long long *count) { 1828 int gtid = __kmp_entry_gtid(); 1829 MKLOC(loc, "GOMP_doacross_ull_post"); 1830 __kmpc_doacross_post(&loc, gtid, RCAST(kmp_int64 *, count)); 1831 } 1832 1833 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_DOACROSS_ULL_WAIT)( 1834 unsigned long long first, ...) { 1835 va_list args; 1836 va_start(args, first); 1837 __kmp_GOMP_doacross_wait<unsigned long long>(first, args); 1838 va_end(args); 1839 } 1840 1841 #endif // OMP_45_ENABLED 1842 1843 /* The following sections of code create aliases for the GOMP_* functions, then 1844 create versioned symbols using the assembler directive .symver. This is only 1845 pertinent for ELF .so library. The KMP_VERSION_SYMBOL macro is defined in 1846 kmp_os.h */ 1847 1848 #ifdef KMP_USE_VERSION_SYMBOLS 1849 // GOMP_1.0 versioned symbols 1850 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_ATOMIC_END, 10, "GOMP_1.0"); 1851 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_ATOMIC_START, 10, "GOMP_1.0"); 1852 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_BARRIER, 10, "GOMP_1.0"); 1853 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_CRITICAL_END, 10, "GOMP_1.0"); 1854 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_CRITICAL_NAME_END, 10, "GOMP_1.0"); 1855 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_CRITICAL_NAME_START, 10, "GOMP_1.0"); 1856 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_CRITICAL_START, 10, "GOMP_1.0"); 1857 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_DYNAMIC_NEXT, 10, "GOMP_1.0"); 1858 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_DYNAMIC_START, 10, "GOMP_1.0"); 1859 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_END, 10, "GOMP_1.0"); 1860 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_END_NOWAIT, 10, "GOMP_1.0"); 1861 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_GUIDED_NEXT, 10, "GOMP_1.0"); 1862 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_GUIDED_START, 10, "GOMP_1.0"); 1863 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ORDERED_DYNAMIC_NEXT, 10, "GOMP_1.0"); 1864 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ORDERED_DYNAMIC_START, 10, 1865 "GOMP_1.0"); 1866 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ORDERED_GUIDED_NEXT, 10, "GOMP_1.0"); 1867 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ORDERED_GUIDED_START, 10, "GOMP_1.0"); 1868 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ORDERED_RUNTIME_NEXT, 10, "GOMP_1.0"); 1869 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ORDERED_RUNTIME_START, 10, 1870 "GOMP_1.0"); 1871 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ORDERED_STATIC_NEXT, 10, "GOMP_1.0"); 1872 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ORDERED_STATIC_START, 10, "GOMP_1.0"); 1873 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_RUNTIME_NEXT, 10, "GOMP_1.0"); 1874 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_RUNTIME_START, 10, "GOMP_1.0"); 1875 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_STATIC_NEXT, 10, "GOMP_1.0"); 1876 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_STATIC_START, 10, "GOMP_1.0"); 1877 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_ORDERED_END, 10, "GOMP_1.0"); 1878 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_ORDERED_START, 10, "GOMP_1.0"); 1879 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_END, 10, "GOMP_1.0"); 1880 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_LOOP_DYNAMIC_START, 10, 1881 "GOMP_1.0"); 1882 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_LOOP_GUIDED_START, 10, 1883 "GOMP_1.0"); 1884 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_LOOP_RUNTIME_START, 10, 1885 "GOMP_1.0"); 1886 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_LOOP_STATIC_START, 10, 1887 "GOMP_1.0"); 1888 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_SECTIONS_START, 10, "GOMP_1.0"); 1889 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_START, 10, "GOMP_1.0"); 1890 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_SECTIONS_END, 10, "GOMP_1.0"); 1891 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_SECTIONS_END_NOWAIT, 10, "GOMP_1.0"); 1892 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_SECTIONS_NEXT, 10, "GOMP_1.0"); 1893 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_SECTIONS_START, 10, "GOMP_1.0"); 1894 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_SINGLE_COPY_END, 10, "GOMP_1.0"); 1895 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_SINGLE_COPY_START, 10, "GOMP_1.0"); 1896 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_SINGLE_START, 10, "GOMP_1.0"); 1897 1898 // GOMP_2.0 versioned symbols 1899 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TASK, 20, "GOMP_2.0"); 1900 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TASKWAIT, 20, "GOMP_2.0"); 1901 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_DYNAMIC_NEXT, 20, "GOMP_2.0"); 1902 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_DYNAMIC_START, 20, "GOMP_2.0"); 1903 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_GUIDED_NEXT, 20, "GOMP_2.0"); 1904 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_GUIDED_START, 20, "GOMP_2.0"); 1905 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_DYNAMIC_NEXT, 20, 1906 "GOMP_2.0"); 1907 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_DYNAMIC_START, 20, 1908 "GOMP_2.0"); 1909 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_GUIDED_NEXT, 20, 1910 "GOMP_2.0"); 1911 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_GUIDED_START, 20, 1912 "GOMP_2.0"); 1913 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_RUNTIME_NEXT, 20, 1914 "GOMP_2.0"); 1915 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_RUNTIME_START, 20, 1916 "GOMP_2.0"); 1917 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_STATIC_NEXT, 20, 1918 "GOMP_2.0"); 1919 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_STATIC_START, 20, 1920 "GOMP_2.0"); 1921 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_RUNTIME_NEXT, 20, "GOMP_2.0"); 1922 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_RUNTIME_START, 20, "GOMP_2.0"); 1923 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_STATIC_NEXT, 20, "GOMP_2.0"); 1924 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_STATIC_START, 20, "GOMP_2.0"); 1925 1926 // GOMP_3.0 versioned symbols 1927 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TASKYIELD, 30, "GOMP_3.0"); 1928 1929 // GOMP_4.0 versioned symbols 1930 #if OMP_40_ENABLED 1931 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL, 40, "GOMP_4.0"); 1932 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_SECTIONS, 40, "GOMP_4.0"); 1933 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_LOOP_DYNAMIC, 40, "GOMP_4.0"); 1934 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_LOOP_GUIDED, 40, "GOMP_4.0"); 1935 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_LOOP_RUNTIME, 40, "GOMP_4.0"); 1936 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_LOOP_STATIC, 40, "GOMP_4.0"); 1937 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TASKGROUP_START, 40, "GOMP_4.0"); 1938 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TASKGROUP_END, 40, "GOMP_4.0"); 1939 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_BARRIER_CANCEL, 40, "GOMP_4.0"); 1940 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_CANCEL, 40, "GOMP_4.0"); 1941 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_CANCELLATION_POINT, 40, "GOMP_4.0"); 1942 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_END_CANCEL, 40, "GOMP_4.0"); 1943 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_SECTIONS_END_CANCEL, 40, "GOMP_4.0"); 1944 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TARGET, 40, "GOMP_4.0"); 1945 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TARGET_DATA, 40, "GOMP_4.0"); 1946 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TARGET_END_DATA, 40, "GOMP_4.0"); 1947 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TARGET_UPDATE, 40, "GOMP_4.0"); 1948 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TEAMS, 40, "GOMP_4.0"); 1949 #endif 1950 1951 // GOMP_4.5 versioned symbols 1952 #if OMP_45_ENABLED 1953 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TASKLOOP, 45, "GOMP_4.5"); 1954 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TASKLOOP_ULL, 45, "GOMP_4.5"); 1955 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_DOACROSS_POST, 45, "GOMP_4.5"); 1956 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_DOACROSS_WAIT, 45, "GOMP_4.5"); 1957 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_DOACROSS_STATIC_START, 45, 1958 "GOMP_4.5"); 1959 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_DOACROSS_DYNAMIC_START, 45, 1960 "GOMP_4.5"); 1961 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_DOACROSS_GUIDED_START, 45, 1962 "GOMP_4.5"); 1963 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_DOACROSS_RUNTIME_START, 45, 1964 "GOMP_4.5"); 1965 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_DOACROSS_ULL_POST, 45, "GOMP_4.5"); 1966 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_DOACROSS_ULL_WAIT, 45, "GOMP_4.5"); 1967 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_DOACROSS_STATIC_START, 45, 1968 "GOMP_4.5"); 1969 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_DOACROSS_DYNAMIC_START, 45, 1970 "GOMP_4.5"); 1971 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_DOACROSS_GUIDED_START, 45, 1972 "GOMP_4.5"); 1973 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_DOACROSS_RUNTIME_START, 45, 1974 "GOMP_4.5"); 1975 #endif 1976 1977 #endif // KMP_USE_VERSION_SYMBOLS 1978 1979 #ifdef __cplusplus 1980 } // extern "C" 1981 #endif // __cplusplus 1982