1 /* 2 * kmp_gsupport.cpp 3 */ 4 5 //===----------------------------------------------------------------------===// 6 // 7 // The LLVM Compiler Infrastructure 8 // 9 // This file is dual licensed under the MIT and the University of Illinois Open 10 // Source Licenses. See LICENSE.txt for details. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "kmp.h" 15 #include "kmp_atomic.h" 16 17 #if OMPT_SUPPORT 18 #include "ompt-specific.h" 19 #endif 20 21 #ifdef __cplusplus 22 extern "C" { 23 #endif // __cplusplus 24 25 #define MKLOC(loc, routine) \ 26 static ident_t(loc) = {0, KMP_IDENT_KMPC, 0, 0, ";unknown;unknown;0;0;;"}; 27 28 #include "kmp_ftn_os.h" 29 30 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_BARRIER)(void) { 31 int gtid = __kmp_entry_gtid(); 32 MKLOC(loc, "GOMP_barrier"); 33 KA_TRACE(20, ("GOMP_barrier: T#%d\n", gtid)); 34 #if OMPT_SUPPORT && OMPT_OPTIONAL 35 omp_frame_t *ompt_frame; 36 if (ompt_enabled.enabled) { 37 __ompt_get_task_info_internal(0, NULL, NULL, &ompt_frame, NULL, NULL); 38 ompt_frame->enter_frame = OMPT_GET_FRAME_ADDRESS(1); 39 OMPT_STORE_RETURN_ADDRESS(gtid); 40 } 41 #endif 42 __kmpc_barrier(&loc, gtid); 43 #if OMPT_SUPPORT && OMPT_OPTIONAL 44 if (ompt_enabled.enabled) { 45 ompt_frame->enter_frame = NULL; 46 } 47 #endif 48 } 49 50 // Mutual exclusion 51 52 // The symbol that icc/ifort generates for unnamed for unnamed critical sections 53 // - .gomp_critical_user_ - is defined using .comm in any objects reference it. 54 // We can't reference it directly here in C code, as the symbol contains a ".". 55 // 56 // The RTL contains an assembly language definition of .gomp_critical_user_ 57 // with another symbol __kmp_unnamed_critical_addr initialized with it's 58 // address. 59 extern kmp_critical_name *__kmp_unnamed_critical_addr; 60 61 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_CRITICAL_START)(void) { 62 int gtid = __kmp_entry_gtid(); 63 MKLOC(loc, "GOMP_critical_start"); 64 KA_TRACE(20, ("GOMP_critical_start: T#%d\n", gtid)); 65 #if OMPT_SUPPORT && OMPT_OPTIONAL 66 OMPT_STORE_RETURN_ADDRESS(gtid); 67 #endif 68 __kmpc_critical(&loc, gtid, __kmp_unnamed_critical_addr); 69 } 70 71 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_CRITICAL_END)(void) { 72 int gtid = __kmp_get_gtid(); 73 MKLOC(loc, "GOMP_critical_end"); 74 KA_TRACE(20, ("GOMP_critical_end: T#%d\n", gtid)); 75 #if OMPT_SUPPORT && OMPT_OPTIONAL 76 OMPT_STORE_RETURN_ADDRESS(gtid); 77 #endif 78 __kmpc_end_critical(&loc, gtid, __kmp_unnamed_critical_addr); 79 } 80 81 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_CRITICAL_NAME_START)(void **pptr) { 82 int gtid = __kmp_entry_gtid(); 83 MKLOC(loc, "GOMP_critical_name_start"); 84 KA_TRACE(20, ("GOMP_critical_name_start: T#%d\n", gtid)); 85 __kmpc_critical(&loc, gtid, (kmp_critical_name *)pptr); 86 } 87 88 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_CRITICAL_NAME_END)(void **pptr) { 89 int gtid = __kmp_get_gtid(); 90 MKLOC(loc, "GOMP_critical_name_end"); 91 KA_TRACE(20, ("GOMP_critical_name_end: T#%d\n", gtid)); 92 __kmpc_end_critical(&loc, gtid, (kmp_critical_name *)pptr); 93 } 94 95 // The Gnu codegen tries to use locked operations to perform atomic updates 96 // inline. If it can't, then it calls GOMP_atomic_start() before performing 97 // the update and GOMP_atomic_end() afterward, regardless of the data type. 98 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_ATOMIC_START)(void) { 99 int gtid = __kmp_entry_gtid(); 100 KA_TRACE(20, ("GOMP_atomic_start: T#%d\n", gtid)); 101 102 #if OMPT_SUPPORT 103 __ompt_thread_assign_wait_id(0); 104 #endif 105 106 __kmp_acquire_atomic_lock(&__kmp_atomic_lock, gtid); 107 } 108 109 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_ATOMIC_END)(void) { 110 int gtid = __kmp_get_gtid(); 111 KA_TRACE(20, ("GOMP_atomic_end: T#%d\n", gtid)); 112 __kmp_release_atomic_lock(&__kmp_atomic_lock, gtid); 113 } 114 115 int KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SINGLE_START)(void) { 116 int gtid = __kmp_entry_gtid(); 117 MKLOC(loc, "GOMP_single_start"); 118 KA_TRACE(20, ("GOMP_single_start: T#%d\n", gtid)); 119 120 if (!TCR_4(__kmp_init_parallel)) 121 __kmp_parallel_initialize(); 122 123 // 3rd parameter == FALSE prevents kmp_enter_single from pushing a 124 // workshare when USE_CHECKS is defined. We need to avoid the push, 125 // as there is no corresponding GOMP_single_end() call. 126 kmp_int32 rc = __kmp_enter_single(gtid, &loc, FALSE); 127 128 #if OMPT_SUPPORT && OMPT_OPTIONAL 129 kmp_info_t *this_thr = __kmp_threads[gtid]; 130 kmp_team_t *team = this_thr->th.th_team; 131 int tid = __kmp_tid_from_gtid(gtid); 132 133 if (ompt_enabled.enabled) { 134 if (rc) { 135 if (ompt_enabled.ompt_callback_work) { 136 ompt_callbacks.ompt_callback(ompt_callback_work)( 137 ompt_work_single_executor, ompt_scope_begin, 138 &(team->t.ompt_team_info.parallel_data), 139 &(team->t.t_implicit_task_taskdata[tid].ompt_task_info.task_data), 140 1, OMPT_GET_RETURN_ADDRESS(0)); 141 } 142 } else { 143 if (ompt_enabled.ompt_callback_work) { 144 ompt_callbacks.ompt_callback(ompt_callback_work)( 145 ompt_work_single_other, ompt_scope_begin, 146 &(team->t.ompt_team_info.parallel_data), 147 &(team->t.t_implicit_task_taskdata[tid].ompt_task_info.task_data), 148 1, OMPT_GET_RETURN_ADDRESS(0)); 149 ompt_callbacks.ompt_callback(ompt_callback_work)( 150 ompt_work_single_other, ompt_scope_end, 151 &(team->t.ompt_team_info.parallel_data), 152 &(team->t.t_implicit_task_taskdata[tid].ompt_task_info.task_data), 153 1, OMPT_GET_RETURN_ADDRESS(0)); 154 } 155 } 156 } 157 #endif 158 159 return rc; 160 } 161 162 void *KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SINGLE_COPY_START)(void) { 163 void *retval; 164 int gtid = __kmp_entry_gtid(); 165 MKLOC(loc, "GOMP_single_copy_start"); 166 KA_TRACE(20, ("GOMP_single_copy_start: T#%d\n", gtid)); 167 168 if (!TCR_4(__kmp_init_parallel)) 169 __kmp_parallel_initialize(); 170 171 // If this is the first thread to enter, return NULL. The generated code will 172 // then call GOMP_single_copy_end() for this thread only, with the 173 // copyprivate data pointer as an argument. 174 if (__kmp_enter_single(gtid, &loc, FALSE)) 175 return NULL; 176 177 // Wait for the first thread to set the copyprivate data pointer, 178 // and for all other threads to reach this point. 179 180 #if OMPT_SUPPORT && OMPT_OPTIONAL 181 omp_frame_t *ompt_frame; 182 if (ompt_enabled.enabled) { 183 __ompt_get_task_info_internal(0, NULL, NULL, &ompt_frame, NULL, NULL); 184 ompt_frame->enter_frame = OMPT_GET_FRAME_ADDRESS(1); 185 OMPT_STORE_RETURN_ADDRESS(gtid); 186 } 187 #endif 188 __kmp_barrier(bs_plain_barrier, gtid, FALSE, 0, NULL, NULL); 189 190 // Retrieve the value of the copyprivate data point, and wait for all 191 // threads to do likewise, then return. 192 retval = __kmp_team_from_gtid(gtid)->t.t_copypriv_data; 193 #if OMPT_SUPPORT && OMPT_OPTIONAL 194 if (ompt_enabled.enabled) { 195 OMPT_STORE_RETURN_ADDRESS(gtid); 196 } 197 #endif 198 __kmp_barrier(bs_plain_barrier, gtid, FALSE, 0, NULL, NULL); 199 #if OMPT_SUPPORT && OMPT_OPTIONAL 200 if (ompt_enabled.enabled) { 201 ompt_frame->enter_frame = NULL; 202 } 203 #endif 204 return retval; 205 } 206 207 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SINGLE_COPY_END)(void *data) { 208 int gtid = __kmp_get_gtid(); 209 KA_TRACE(20, ("GOMP_single_copy_end: T#%d\n", gtid)); 210 211 // Set the copyprivate data pointer fo the team, then hit the barrier so that 212 // the other threads will continue on and read it. Hit another barrier before 213 // continuing, so that the know that the copyprivate data pointer has been 214 // propagated to all threads before trying to reuse the t_copypriv_data field. 215 __kmp_team_from_gtid(gtid)->t.t_copypriv_data = data; 216 #if OMPT_SUPPORT && OMPT_OPTIONAL 217 omp_frame_t *ompt_frame; 218 if (ompt_enabled.enabled) { 219 __ompt_get_task_info_internal(0, NULL, NULL, &ompt_frame, NULL, NULL); 220 ompt_frame->enter_frame = OMPT_GET_FRAME_ADDRESS(1); 221 OMPT_STORE_RETURN_ADDRESS(gtid); 222 } 223 #endif 224 __kmp_barrier(bs_plain_barrier, gtid, FALSE, 0, NULL, NULL); 225 #if OMPT_SUPPORT && OMPT_OPTIONAL 226 if (ompt_enabled.enabled) { 227 OMPT_STORE_RETURN_ADDRESS(gtid); 228 } 229 #endif 230 __kmp_barrier(bs_plain_barrier, gtid, FALSE, 0, NULL, NULL); 231 #if OMPT_SUPPORT && OMPT_OPTIONAL 232 if (ompt_enabled.enabled) { 233 ompt_frame->enter_frame = NULL; 234 } 235 #endif 236 } 237 238 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_ORDERED_START)(void) { 239 int gtid = __kmp_entry_gtid(); 240 MKLOC(loc, "GOMP_ordered_start"); 241 KA_TRACE(20, ("GOMP_ordered_start: T#%d\n", gtid)); 242 #if OMPT_SUPPORT && OMPT_OPTIONAL 243 OMPT_STORE_RETURN_ADDRESS(gtid); 244 #endif 245 __kmpc_ordered(&loc, gtid); 246 } 247 248 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_ORDERED_END)(void) { 249 int gtid = __kmp_get_gtid(); 250 MKLOC(loc, "GOMP_ordered_end"); 251 KA_TRACE(20, ("GOMP_ordered_start: T#%d\n", gtid)); 252 #if OMPT_SUPPORT && OMPT_OPTIONAL 253 OMPT_STORE_RETURN_ADDRESS(gtid); 254 #endif 255 __kmpc_end_ordered(&loc, gtid); 256 } 257 258 // Dispatch macro defs 259 // 260 // They come in two flavors: 64-bit unsigned, and either 32-bit signed 261 // (IA-32 architecture) or 64-bit signed (Intel(R) 64). 262 263 #if KMP_ARCH_X86 || KMP_ARCH_ARM || KMP_ARCH_MIPS 264 #define KMP_DISPATCH_INIT __kmp_aux_dispatch_init_4 265 #define KMP_DISPATCH_FINI_CHUNK __kmp_aux_dispatch_fini_chunk_4 266 #define KMP_DISPATCH_NEXT __kmpc_dispatch_next_4 267 #else 268 #define KMP_DISPATCH_INIT __kmp_aux_dispatch_init_8 269 #define KMP_DISPATCH_FINI_CHUNK __kmp_aux_dispatch_fini_chunk_8 270 #define KMP_DISPATCH_NEXT __kmpc_dispatch_next_8 271 #endif /* KMP_ARCH_X86 */ 272 273 #define KMP_DISPATCH_INIT_ULL __kmp_aux_dispatch_init_8u 274 #define KMP_DISPATCH_FINI_CHUNK_ULL __kmp_aux_dispatch_fini_chunk_8u 275 #define KMP_DISPATCH_NEXT_ULL __kmpc_dispatch_next_8u 276 277 // The parallel contruct 278 279 #ifndef KMP_DEBUG 280 static 281 #endif /* KMP_DEBUG */ 282 void 283 __kmp_GOMP_microtask_wrapper(int *gtid, int *npr, void (*task)(void *), 284 void *data) { 285 #if OMPT_SUPPORT 286 kmp_info_t *thr; 287 omp_frame_t *ompt_frame; 288 omp_state_t enclosing_state; 289 290 if (ompt_enabled.enabled) { 291 // get pointer to thread data structure 292 thr = __kmp_threads[*gtid]; 293 294 // save enclosing task state; set current state for task 295 enclosing_state = thr->th.ompt_thread_info.state; 296 thr->th.ompt_thread_info.state = omp_state_work_parallel; 297 298 // set task frame 299 __ompt_get_task_info_internal(0, NULL, NULL, &ompt_frame, NULL, NULL); 300 ompt_frame->exit_frame = OMPT_GET_FRAME_ADDRESS(0); 301 } 302 #endif 303 304 task(data); 305 306 #if OMPT_SUPPORT 307 if (ompt_enabled.enabled) { 308 // clear task frame 309 ompt_frame->exit_frame = NULL; 310 311 // restore enclosing state 312 thr->th.ompt_thread_info.state = enclosing_state; 313 } 314 #endif 315 } 316 317 #ifndef KMP_DEBUG 318 static 319 #endif /* KMP_DEBUG */ 320 void 321 __kmp_GOMP_parallel_microtask_wrapper(int *gtid, int *npr, 322 void (*task)(void *), void *data, 323 unsigned num_threads, ident_t *loc, 324 enum sched_type schedule, long start, 325 long end, long incr, 326 long chunk_size) { 327 // Intialize the loop worksharing construct. 328 329 KMP_DISPATCH_INIT(loc, *gtid, schedule, start, end, incr, chunk_size, 330 schedule != kmp_sch_static); 331 332 #if OMPT_SUPPORT 333 kmp_info_t *thr; 334 omp_frame_t *ompt_frame; 335 omp_state_t enclosing_state; 336 337 if (ompt_enabled.enabled) { 338 thr = __kmp_threads[*gtid]; 339 // save enclosing task state; set current state for task 340 enclosing_state = thr->th.ompt_thread_info.state; 341 thr->th.ompt_thread_info.state = omp_state_work_parallel; 342 343 // set task frame 344 __ompt_get_task_info_internal(0, NULL, NULL, &ompt_frame, NULL, NULL); 345 ompt_frame->exit_frame = OMPT_GET_FRAME_ADDRESS(0); 346 } 347 #endif 348 349 // Now invoke the microtask. 350 task(data); 351 352 #if OMPT_SUPPORT 353 if (ompt_enabled.enabled) { 354 // clear task frame 355 ompt_frame->exit_frame = NULL; 356 357 // reset enclosing state 358 thr->th.ompt_thread_info.state = enclosing_state; 359 } 360 #endif 361 } 362 363 #ifndef KMP_DEBUG 364 static 365 #endif /* KMP_DEBUG */ 366 void 367 __kmp_GOMP_fork_call(ident_t *loc, int gtid, void (*unwrapped_task)(void *), 368 microtask_t wrapper, int argc, ...) { 369 int rc; 370 kmp_info_t *thr = __kmp_threads[gtid]; 371 kmp_team_t *team = thr->th.th_team; 372 int tid = __kmp_tid_from_gtid(gtid); 373 374 va_list ap; 375 va_start(ap, argc); 376 377 rc = __kmp_fork_call(loc, gtid, fork_context_gnu, argc, wrapper, 378 __kmp_invoke_task_func, 379 #if (KMP_ARCH_X86_64 || KMP_ARCH_ARM || KMP_ARCH_AARCH64) && KMP_OS_LINUX 380 &ap 381 #else 382 ap 383 #endif 384 ); 385 386 va_end(ap); 387 388 if (rc) { 389 __kmp_run_before_invoked_task(gtid, tid, thr, team); 390 } 391 392 #if OMPT_SUPPORT 393 int ompt_team_size; 394 if (ompt_enabled.enabled) { 395 ompt_team_info_t *team_info = __ompt_get_teaminfo(0, NULL); 396 ompt_task_info_t *task_info = __ompt_get_task_info_object(0); 397 398 // implicit task callback 399 if (ompt_enabled.ompt_callback_implicit_task) { 400 ompt_team_size = __kmp_team_from_gtid(gtid)->t.t_nproc; 401 ompt_callbacks.ompt_callback(ompt_callback_implicit_task)( 402 ompt_scope_begin, &(team_info->parallel_data), 403 &(task_info->task_data), ompt_team_size, __kmp_tid_from_gtid(gtid)); 404 task_info->thread_num = __kmp_tid_from_gtid(gtid); 405 } 406 thr->th.ompt_thread_info.state = omp_state_work_parallel; 407 } 408 #endif 409 } 410 411 static void __kmp_GOMP_serialized_parallel(ident_t *loc, kmp_int32 gtid, 412 void (*task)(void *)) { 413 #if OMPT_SUPPORT 414 OMPT_STORE_RETURN_ADDRESS(gtid); 415 #endif 416 __kmp_serialized_parallel(loc, gtid); 417 } 418 419 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_START)(void (*task)(void *), 420 void *data, 421 unsigned num_threads) { 422 int gtid = __kmp_entry_gtid(); 423 424 #if OMPT_SUPPORT 425 omp_frame_t *parent_frame, *frame; 426 427 if (ompt_enabled.enabled) { 428 __ompt_get_task_info_internal(0, NULL, NULL, &parent_frame, NULL, NULL); 429 parent_frame->enter_frame = OMPT_GET_FRAME_ADDRESS(1); 430 OMPT_STORE_RETURN_ADDRESS(gtid); 431 } 432 #endif 433 434 MKLOC(loc, "GOMP_parallel_start"); 435 KA_TRACE(20, ("GOMP_parallel_start: T#%d\n", gtid)); 436 437 if (__kmpc_ok_to_fork(&loc) && (num_threads != 1)) { 438 if (num_threads != 0) { 439 __kmp_push_num_threads(&loc, gtid, num_threads); 440 } 441 __kmp_GOMP_fork_call(&loc, gtid, task, 442 (microtask_t)__kmp_GOMP_microtask_wrapper, 2, task, 443 data); 444 } else { 445 __kmp_GOMP_serialized_parallel(&loc, gtid, task); 446 } 447 448 #if OMPT_SUPPORT 449 if (ompt_enabled.enabled) { 450 __ompt_get_task_info_internal(0, NULL, NULL, &frame, NULL, NULL); 451 frame->exit_frame = OMPT_GET_FRAME_ADDRESS(1); 452 } 453 #endif 454 } 455 456 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_END)(void) { 457 int gtid = __kmp_get_gtid(); 458 kmp_info_t *thr; 459 int ompt_team_size = __kmp_team_from_gtid(gtid)->t.t_nproc; 460 461 thr = __kmp_threads[gtid]; 462 463 MKLOC(loc, "GOMP_parallel_end"); 464 KA_TRACE(20, ("GOMP_parallel_end: T#%d\n", gtid)); 465 466 if (!thr->th.th_team->t.t_serialized) { 467 __kmp_run_after_invoked_task(gtid, __kmp_tid_from_gtid(gtid), thr, 468 thr->th.th_team); 469 470 #if OMPT_SUPPORT 471 if (ompt_enabled.enabled) { 472 // Implicit task is finished here, in the barrier we might schedule 473 // deferred tasks, 474 // these don't see the implicit task on the stack 475 OMPT_CUR_TASK_INFO(thr)->frame.exit_frame = NULL; 476 } 477 #endif 478 479 __kmp_join_call(&loc, gtid 480 #if OMPT_SUPPORT 481 , 482 fork_context_gnu 483 #endif 484 ); 485 } else { 486 __kmpc_end_serialized_parallel(&loc, gtid); 487 } 488 } 489 490 // Loop worksharing constructs 491 492 // The Gnu codegen passes in an exclusive upper bound for the overall range, 493 // but the libguide dispatch code expects an inclusive upper bound, hence the 494 // "end - incr" 5th argument to KMP_DISPATCH_INIT (and the " ub - str" 11th 495 // argument to __kmp_GOMP_fork_call). 496 // 497 // Conversely, KMP_DISPATCH_NEXT returns and inclusive upper bound in *p_ub, 498 // but the Gnu codegen expects an excluside upper bound, so the adjustment 499 // "*p_ub += stride" compenstates for the discrepancy. 500 // 501 // Correction: the gnu codegen always adjusts the upper bound by +-1, not the 502 // stride value. We adjust the dispatch parameters accordingly (by +-1), but 503 // we still adjust p_ub by the actual stride value. 504 // 505 // The "runtime" versions do not take a chunk_sz parameter. 506 // 507 // The profile lib cannot support construct checking of unordered loops that 508 // are predetermined by the compiler to be statically scheduled, as the gcc 509 // codegen will not always emit calls to GOMP_loop_static_next() to get the 510 // next iteration. Instead, it emits inline code to call omp_get_thread_num() 511 // num and calculate the iteration space using the result. It doesn't do this 512 // with ordered static loop, so they can be checked. 513 514 #if OMPT_SUPPORT 515 #define IF_OMPT_SUPPORT(code) code 516 #else 517 #define IF_OMPT_SUPPORT(code) 518 #endif 519 520 #define LOOP_START(func, schedule) \ 521 int func(long lb, long ub, long str, long chunk_sz, long *p_lb, \ 522 long *p_ub) { \ 523 int status; \ 524 long stride; \ 525 int gtid = __kmp_entry_gtid(); \ 526 MKLOC(loc, KMP_STR(func)); \ 527 KA_TRACE( \ 528 20, \ 529 (KMP_STR( \ 530 func) ": T#%d, lb 0x%lx, ub 0x%lx, str 0x%lx, chunk_sz 0x%lx\n", \ 531 gtid, lb, ub, str, chunk_sz)); \ 532 \ 533 if ((str > 0) ? (lb < ub) : (lb > ub)) { \ 534 IF_OMPT_SUPPORT(OMPT_STORE_RETURN_ADDRESS(gtid);) \ 535 KMP_DISPATCH_INIT(&loc, gtid, (schedule), lb, \ 536 (str > 0) ? (ub - 1) : (ub + 1), str, chunk_sz, \ 537 (schedule) != kmp_sch_static); \ 538 IF_OMPT_SUPPORT(OMPT_STORE_RETURN_ADDRESS(gtid);) \ 539 status = KMP_DISPATCH_NEXT(&loc, gtid, NULL, (kmp_int *)p_lb, \ 540 (kmp_int *)p_ub, (kmp_int *)&stride); \ 541 if (status) { \ 542 KMP_DEBUG_ASSERT(stride == str); \ 543 *p_ub += (str > 0) ? 1 : -1; \ 544 } \ 545 } else { \ 546 status = 0; \ 547 } \ 548 \ 549 KA_TRACE( \ 550 20, \ 551 (KMP_STR( \ 552 func) " exit: T#%d, *p_lb 0x%lx, *p_ub 0x%lx, returning %d\n", \ 553 gtid, *p_lb, *p_ub, status)); \ 554 return status; \ 555 } 556 557 #define LOOP_RUNTIME_START(func, schedule) \ 558 int func(long lb, long ub, long str, long *p_lb, long *p_ub) { \ 559 int status; \ 560 long stride; \ 561 long chunk_sz = 0; \ 562 int gtid = __kmp_entry_gtid(); \ 563 MKLOC(loc, KMP_STR(func)); \ 564 KA_TRACE( \ 565 20, \ 566 (KMP_STR(func) ": T#%d, lb 0x%lx, ub 0x%lx, str 0x%lx, chunk_sz %d\n", \ 567 gtid, lb, ub, str, chunk_sz)); \ 568 \ 569 if ((str > 0) ? (lb < ub) : (lb > ub)) { \ 570 IF_OMPT_SUPPORT(OMPT_STORE_RETURN_ADDRESS(gtid);) \ 571 KMP_DISPATCH_INIT(&loc, gtid, (schedule), lb, \ 572 (str > 0) ? (ub - 1) : (ub + 1), str, chunk_sz, TRUE); \ 573 IF_OMPT_SUPPORT(OMPT_STORE_RETURN_ADDRESS(gtid);) \ 574 status = KMP_DISPATCH_NEXT(&loc, gtid, NULL, (kmp_int *)p_lb, \ 575 (kmp_int *)p_ub, (kmp_int *)&stride); \ 576 if (status) { \ 577 KMP_DEBUG_ASSERT(stride == str); \ 578 *p_ub += (str > 0) ? 1 : -1; \ 579 } \ 580 } else { \ 581 status = 0; \ 582 } \ 583 \ 584 KA_TRACE( \ 585 20, \ 586 (KMP_STR( \ 587 func) " exit: T#%d, *p_lb 0x%lx, *p_ub 0x%lx, returning %d\n", \ 588 gtid, *p_lb, *p_ub, status)); \ 589 return status; \ 590 } 591 592 #if OMP_45_ENABLED 593 #define KMP_DOACROSS_FINI(status, gtid) \ 594 if (!status && __kmp_threads[gtid]->th.th_dispatch->th_doacross_flags) { \ 595 __kmpc_doacross_fini(NULL, gtid); \ 596 } 597 #else 598 #define KMP_DOACROSS_FINI(status, gtid) /* Nothing */ 599 #endif 600 601 #define LOOP_NEXT(func, fini_code) \ 602 int func(long *p_lb, long *p_ub) { \ 603 int status; \ 604 long stride; \ 605 int gtid = __kmp_get_gtid(); \ 606 MKLOC(loc, KMP_STR(func)); \ 607 KA_TRACE(20, (KMP_STR(func) ": T#%d\n", gtid)); \ 608 \ 609 IF_OMPT_SUPPORT(OMPT_STORE_RETURN_ADDRESS(gtid);) \ 610 fini_code status = KMP_DISPATCH_NEXT(&loc, gtid, NULL, (kmp_int *)p_lb, \ 611 (kmp_int *)p_ub, (kmp_int *)&stride); \ 612 if (status) { \ 613 *p_ub += (stride > 0) ? 1 : -1; \ 614 } \ 615 KMP_DOACROSS_FINI(status, gtid) \ 616 \ 617 KA_TRACE( \ 618 20, \ 619 (KMP_STR(func) " exit: T#%d, *p_lb 0x%lx, *p_ub 0x%lx, stride 0x%lx, " \ 620 "returning %d\n", \ 621 gtid, *p_lb, *p_ub, stride, status)); \ 622 return status; \ 623 } 624 625 LOOP_START(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_STATIC_START), kmp_sch_static) 626 LOOP_NEXT(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_STATIC_NEXT), {}) 627 LOOP_START(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_DYNAMIC_START), 628 kmp_sch_dynamic_chunked) 629 LOOP_NEXT(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_DYNAMIC_NEXT), {}) 630 LOOP_START(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_GUIDED_START), 631 kmp_sch_guided_chunked) 632 LOOP_NEXT(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_GUIDED_NEXT), {}) 633 LOOP_RUNTIME_START(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_RUNTIME_START), 634 kmp_sch_runtime) 635 LOOP_NEXT(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_RUNTIME_NEXT), {}) 636 637 LOOP_START(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ORDERED_STATIC_START), 638 kmp_ord_static) 639 LOOP_NEXT(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ORDERED_STATIC_NEXT), 640 { KMP_DISPATCH_FINI_CHUNK(&loc, gtid); }) 641 LOOP_START(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ORDERED_DYNAMIC_START), 642 kmp_ord_dynamic_chunked) 643 LOOP_NEXT(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ORDERED_DYNAMIC_NEXT), 644 { KMP_DISPATCH_FINI_CHUNK(&loc, gtid); }) 645 LOOP_START(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ORDERED_GUIDED_START), 646 kmp_ord_guided_chunked) 647 LOOP_NEXT(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ORDERED_GUIDED_NEXT), 648 { KMP_DISPATCH_FINI_CHUNK(&loc, gtid); }) 649 LOOP_RUNTIME_START( 650 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ORDERED_RUNTIME_START), 651 kmp_ord_runtime) 652 LOOP_NEXT(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ORDERED_RUNTIME_NEXT), 653 { KMP_DISPATCH_FINI_CHUNK(&loc, gtid); }) 654 655 #if OMP_45_ENABLED 656 #define LOOP_DOACROSS_START(func, schedule) \ 657 bool func(unsigned ncounts, long *counts, long chunk_sz, long *p_lb, \ 658 long *p_ub) { \ 659 int status; \ 660 long stride, lb, ub, str; \ 661 int gtid = __kmp_entry_gtid(); \ 662 struct kmp_dim *dims = \ 663 (struct kmp_dim *)__kmp_allocate(sizeof(struct kmp_dim) * ncounts); \ 664 MKLOC(loc, KMP_STR(func)); \ 665 for (unsigned i = 0; i < ncounts; ++i) { \ 666 dims[i].lo = 0; \ 667 dims[i].up = counts[i] - 1; \ 668 dims[i].st = 1; \ 669 } \ 670 __kmpc_doacross_init(&loc, gtid, (int)ncounts, dims); \ 671 lb = 0; \ 672 ub = counts[0]; \ 673 str = 1; \ 674 KA_TRACE(20, (KMP_STR(func) ": T#%d, ncounts %u, lb 0x%lx, ub 0x%lx, str " \ 675 "0x%lx, chunk_sz " \ 676 "0x%lx\n", \ 677 gtid, ncounts, lb, ub, str, chunk_sz)); \ 678 \ 679 if ((str > 0) ? (lb < ub) : (lb > ub)) { \ 680 KMP_DISPATCH_INIT(&loc, gtid, (schedule), lb, \ 681 (str > 0) ? (ub - 1) : (ub + 1), str, chunk_sz, \ 682 (schedule) != kmp_sch_static); \ 683 status = KMP_DISPATCH_NEXT(&loc, gtid, NULL, (kmp_int *)p_lb, \ 684 (kmp_int *)p_ub, (kmp_int *)&stride); \ 685 if (status) { \ 686 KMP_DEBUG_ASSERT(stride == str); \ 687 *p_ub += (str > 0) ? 1 : -1; \ 688 } \ 689 } else { \ 690 status = 0; \ 691 } \ 692 KMP_DOACROSS_FINI(status, gtid); \ 693 \ 694 KA_TRACE( \ 695 20, \ 696 (KMP_STR( \ 697 func) " exit: T#%d, *p_lb 0x%lx, *p_ub 0x%lx, returning %d\n", \ 698 gtid, *p_lb, *p_ub, status)); \ 699 __kmp_free(dims); \ 700 return status; \ 701 } 702 703 #define LOOP_DOACROSS_RUNTIME_START(func, schedule) \ 704 int func(unsigned ncounts, long *counts, long *p_lb, long *p_ub) { \ 705 int status; \ 706 long stride, lb, ub, str; \ 707 long chunk_sz = 0; \ 708 int gtid = __kmp_entry_gtid(); \ 709 struct kmp_dim *dims = \ 710 (struct kmp_dim *)__kmp_allocate(sizeof(struct kmp_dim) * ncounts); \ 711 MKLOC(loc, KMP_STR(func)); \ 712 for (unsigned i = 0; i < ncounts; ++i) { \ 713 dims[i].lo = 0; \ 714 dims[i].up = counts[i] - 1; \ 715 dims[i].st = 1; \ 716 } \ 717 __kmpc_doacross_init(&loc, gtid, (int)ncounts, dims); \ 718 lb = 0; \ 719 ub = counts[0]; \ 720 str = 1; \ 721 KA_TRACE( \ 722 20, \ 723 (KMP_STR(func) ": T#%d, lb 0x%lx, ub 0x%lx, str 0x%lx, chunk_sz %d\n", \ 724 gtid, lb, ub, str, chunk_sz)); \ 725 \ 726 if ((str > 0) ? (lb < ub) : (lb > ub)) { \ 727 KMP_DISPATCH_INIT(&loc, gtid, (schedule), lb, \ 728 (str > 0) ? (ub - 1) : (ub + 1), str, chunk_sz, TRUE); \ 729 status = KMP_DISPATCH_NEXT(&loc, gtid, NULL, (kmp_int *)p_lb, \ 730 (kmp_int *)p_ub, (kmp_int *)&stride); \ 731 if (status) { \ 732 KMP_DEBUG_ASSERT(stride == str); \ 733 *p_ub += (str > 0) ? 1 : -1; \ 734 } \ 735 } else { \ 736 status = 0; \ 737 } \ 738 KMP_DOACROSS_FINI(status, gtid); \ 739 \ 740 KA_TRACE( \ 741 20, \ 742 (KMP_STR( \ 743 func) " exit: T#%d, *p_lb 0x%lx, *p_ub 0x%lx, returning %d\n", \ 744 gtid, *p_lb, *p_ub, status)); \ 745 __kmp_free(dims); \ 746 return status; \ 747 } 748 749 LOOP_DOACROSS_START( 750 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_DOACROSS_STATIC_START), 751 kmp_sch_static) 752 LOOP_DOACROSS_START( 753 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_DOACROSS_DYNAMIC_START), 754 kmp_sch_dynamic_chunked) 755 LOOP_DOACROSS_START( 756 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_DOACROSS_GUIDED_START), 757 kmp_sch_guided_chunked) 758 LOOP_DOACROSS_RUNTIME_START( 759 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_DOACROSS_RUNTIME_START), 760 kmp_sch_runtime) 761 #endif // OMP_45_ENABLED 762 763 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_END)(void) { 764 int gtid = __kmp_get_gtid(); 765 KA_TRACE(20, ("GOMP_loop_end: T#%d\n", gtid)) 766 767 #if OMPT_SUPPORT && OMPT_OPTIONAL 768 omp_frame_t *ompt_frame; 769 if (ompt_enabled.enabled) { 770 __ompt_get_task_info_internal(0, NULL, NULL, &ompt_frame, NULL, NULL); 771 ompt_frame->enter_frame = OMPT_GET_FRAME_ADDRESS(1); 772 OMPT_STORE_RETURN_ADDRESS(gtid); 773 } 774 #endif 775 __kmp_barrier(bs_plain_barrier, gtid, FALSE, 0, NULL, NULL); 776 #if OMPT_SUPPORT && OMPT_OPTIONAL 777 if (ompt_enabled.enabled) { 778 ompt_frame->enter_frame = NULL; 779 } 780 #endif 781 782 KA_TRACE(20, ("GOMP_loop_end exit: T#%d\n", gtid)) 783 } 784 785 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_END_NOWAIT)(void) { 786 KA_TRACE(20, ("GOMP_loop_end_nowait: T#%d\n", __kmp_get_gtid())) 787 } 788 789 // Unsigned long long loop worksharing constructs 790 // 791 // These are new with gcc 4.4 792 793 #define LOOP_START_ULL(func, schedule) \ 794 int func(int up, unsigned long long lb, unsigned long long ub, \ 795 unsigned long long str, unsigned long long chunk_sz, \ 796 unsigned long long *p_lb, unsigned long long *p_ub) { \ 797 int status; \ 798 long long str2 = up ? ((long long)str) : -((long long)str); \ 799 long long stride; \ 800 int gtid = __kmp_entry_gtid(); \ 801 MKLOC(loc, KMP_STR(func)); \ 802 \ 803 KA_TRACE(20, (KMP_STR(func) ": T#%d, up %d, lb 0x%llx, ub 0x%llx, str " \ 804 "0x%llx, chunk_sz 0x%llx\n", \ 805 gtid, up, lb, ub, str, chunk_sz)); \ 806 \ 807 if ((str > 0) ? (lb < ub) : (lb > ub)) { \ 808 KMP_DISPATCH_INIT_ULL(&loc, gtid, (schedule), lb, \ 809 (str2 > 0) ? (ub - 1) : (ub + 1), str2, chunk_sz, \ 810 (schedule) != kmp_sch_static); \ 811 status = \ 812 KMP_DISPATCH_NEXT_ULL(&loc, gtid, NULL, (kmp_uint64 *)p_lb, \ 813 (kmp_uint64 *)p_ub, (kmp_int64 *)&stride); \ 814 if (status) { \ 815 KMP_DEBUG_ASSERT(stride == str2); \ 816 *p_ub += (str > 0) ? 1 : -1; \ 817 } \ 818 } else { \ 819 status = 0; \ 820 } \ 821 \ 822 KA_TRACE( \ 823 20, \ 824 (KMP_STR( \ 825 func) " exit: T#%d, *p_lb 0x%llx, *p_ub 0x%llx, returning %d\n", \ 826 gtid, *p_lb, *p_ub, status)); \ 827 return status; \ 828 } 829 830 #define LOOP_RUNTIME_START_ULL(func, schedule) \ 831 int func(int up, unsigned long long lb, unsigned long long ub, \ 832 unsigned long long str, unsigned long long *p_lb, \ 833 unsigned long long *p_ub) { \ 834 int status; \ 835 long long str2 = up ? ((long long)str) : -((long long)str); \ 836 unsigned long long stride; \ 837 unsigned long long chunk_sz = 0; \ 838 int gtid = __kmp_entry_gtid(); \ 839 MKLOC(loc, KMP_STR(func)); \ 840 \ 841 KA_TRACE(20, (KMP_STR(func) ": T#%d, up %d, lb 0x%llx, ub 0x%llx, str " \ 842 "0x%llx, chunk_sz 0x%llx\n", \ 843 gtid, up, lb, ub, str, chunk_sz)); \ 844 \ 845 if ((str > 0) ? (lb < ub) : (lb > ub)) { \ 846 KMP_DISPATCH_INIT_ULL(&loc, gtid, (schedule), lb, \ 847 (str2 > 0) ? (ub - 1) : (ub + 1), str2, chunk_sz, \ 848 TRUE); \ 849 status = \ 850 KMP_DISPATCH_NEXT_ULL(&loc, gtid, NULL, (kmp_uint64 *)p_lb, \ 851 (kmp_uint64 *)p_ub, (kmp_int64 *)&stride); \ 852 if (status) { \ 853 KMP_DEBUG_ASSERT((long long)stride == str2); \ 854 *p_ub += (str > 0) ? 1 : -1; \ 855 } \ 856 } else { \ 857 status = 0; \ 858 } \ 859 \ 860 KA_TRACE( \ 861 20, \ 862 (KMP_STR( \ 863 func) " exit: T#%d, *p_lb 0x%llx, *p_ub 0x%llx, returning %d\n", \ 864 gtid, *p_lb, *p_ub, status)); \ 865 return status; \ 866 } 867 868 #define LOOP_NEXT_ULL(func, fini_code) \ 869 int func(unsigned long long *p_lb, unsigned long long *p_ub) { \ 870 int status; \ 871 long long stride; \ 872 int gtid = __kmp_get_gtid(); \ 873 MKLOC(loc, KMP_STR(func)); \ 874 KA_TRACE(20, (KMP_STR(func) ": T#%d\n", gtid)); \ 875 \ 876 fini_code status = \ 877 KMP_DISPATCH_NEXT_ULL(&loc, gtid, NULL, (kmp_uint64 *)p_lb, \ 878 (kmp_uint64 *)p_ub, (kmp_int64 *)&stride); \ 879 if (status) { \ 880 *p_ub += (stride > 0) ? 1 : -1; \ 881 } \ 882 \ 883 KA_TRACE( \ 884 20, \ 885 (KMP_STR( \ 886 func) " exit: T#%d, *p_lb 0x%llx, *p_ub 0x%llx, stride 0x%llx, " \ 887 "returning %d\n", \ 888 gtid, *p_lb, *p_ub, stride, status)); \ 889 return status; \ 890 } 891 892 LOOP_START_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_STATIC_START), 893 kmp_sch_static) 894 LOOP_NEXT_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_STATIC_NEXT), {}) 895 LOOP_START_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_DYNAMIC_START), 896 kmp_sch_dynamic_chunked) 897 LOOP_NEXT_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_DYNAMIC_NEXT), {}) 898 LOOP_START_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_GUIDED_START), 899 kmp_sch_guided_chunked) 900 LOOP_NEXT_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_GUIDED_NEXT), {}) 901 LOOP_RUNTIME_START_ULL( 902 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_RUNTIME_START), kmp_sch_runtime) 903 LOOP_NEXT_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_RUNTIME_NEXT), {}) 904 905 LOOP_START_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_STATIC_START), 906 kmp_ord_static) 907 LOOP_NEXT_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_STATIC_NEXT), 908 { KMP_DISPATCH_FINI_CHUNK_ULL(&loc, gtid); }) 909 LOOP_START_ULL( 910 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_DYNAMIC_START), 911 kmp_ord_dynamic_chunked) 912 LOOP_NEXT_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_DYNAMIC_NEXT), 913 { KMP_DISPATCH_FINI_CHUNK_ULL(&loc, gtid); }) 914 LOOP_START_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_GUIDED_START), 915 kmp_ord_guided_chunked) 916 LOOP_NEXT_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_GUIDED_NEXT), 917 { KMP_DISPATCH_FINI_CHUNK_ULL(&loc, gtid); }) 918 LOOP_RUNTIME_START_ULL( 919 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_RUNTIME_START), 920 kmp_ord_runtime) 921 LOOP_NEXT_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_RUNTIME_NEXT), 922 { KMP_DISPATCH_FINI_CHUNK_ULL(&loc, gtid); }) 923 924 #if OMP_45_ENABLED 925 #define LOOP_DOACROSS_START_ULL(func, schedule) \ 926 int func(unsigned ncounts, unsigned long long *counts, \ 927 unsigned long long chunk_sz, unsigned long long *p_lb, \ 928 unsigned long long *p_ub) { \ 929 int status; \ 930 long long stride, str, lb, ub; \ 931 int gtid = __kmp_entry_gtid(); \ 932 struct kmp_dim *dims = \ 933 (struct kmp_dim *)__kmp_allocate(sizeof(struct kmp_dim) * ncounts); \ 934 MKLOC(loc, KMP_STR(func)); \ 935 for (unsigned i = 0; i < ncounts; ++i) { \ 936 dims[i].lo = 0; \ 937 dims[i].up = counts[i] - 1; \ 938 dims[i].st = 1; \ 939 } \ 940 __kmpc_doacross_init(&loc, gtid, (int)ncounts, dims); \ 941 lb = 0; \ 942 ub = counts[0]; \ 943 str = 1; \ 944 \ 945 KA_TRACE(20, (KMP_STR(func) ": T#%d, lb 0x%llx, ub 0x%llx, str " \ 946 "0x%llx, chunk_sz 0x%llx\n", \ 947 gtid, lb, ub, str, chunk_sz)); \ 948 \ 949 if ((str > 0) ? (lb < ub) : (lb > ub)) { \ 950 KMP_DISPATCH_INIT_ULL(&loc, gtid, (schedule), lb, \ 951 (str > 0) ? (ub - 1) : (ub + 1), str, chunk_sz, \ 952 (schedule) != kmp_sch_static); \ 953 status = \ 954 KMP_DISPATCH_NEXT_ULL(&loc, gtid, NULL, (kmp_uint64 *)p_lb, \ 955 (kmp_uint64 *)p_ub, (kmp_int64 *)&stride); \ 956 if (status) { \ 957 KMP_DEBUG_ASSERT(stride == str); \ 958 *p_ub += (str > 0) ? 1 : -1; \ 959 } \ 960 } else { \ 961 status = 0; \ 962 } \ 963 KMP_DOACROSS_FINI(status, gtid); \ 964 \ 965 KA_TRACE( \ 966 20, \ 967 (KMP_STR( \ 968 func) " exit: T#%d, *p_lb 0x%llx, *p_ub 0x%llx, returning %d\n", \ 969 gtid, *p_lb, *p_ub, status)); \ 970 __kmp_free(dims); \ 971 return status; \ 972 } 973 974 #define LOOP_DOACROSS_RUNTIME_START_ULL(func, schedule) \ 975 int func(unsigned ncounts, unsigned long long *counts, \ 976 unsigned long long *p_lb, unsigned long long *p_ub) { \ 977 int status; \ 978 unsigned long long stride, str, lb, ub; \ 979 unsigned long long chunk_sz = 0; \ 980 int gtid = __kmp_entry_gtid(); \ 981 struct kmp_dim *dims = \ 982 (struct kmp_dim *)__kmp_allocate(sizeof(struct kmp_dim) * ncounts); \ 983 MKLOC(loc, KMP_STR(func)); \ 984 for (unsigned i = 0; i < ncounts; ++i) { \ 985 dims[i].lo = 0; \ 986 dims[i].up = counts[i] - 1; \ 987 dims[i].st = 1; \ 988 } \ 989 __kmpc_doacross_init(&loc, gtid, (int)ncounts, dims); \ 990 lb = 0; \ 991 ub = counts[0]; \ 992 str = 1; \ 993 KA_TRACE(20, (KMP_STR(func) ": T#%d, lb 0x%llx, ub 0x%llx, str " \ 994 "0x%llx, chunk_sz 0x%llx\n", \ 995 gtid, lb, ub, str, chunk_sz)); \ 996 \ 997 if ((str > 0) ? (lb < ub) : (lb > ub)) { \ 998 KMP_DISPATCH_INIT_ULL(&loc, gtid, (schedule), lb, \ 999 (str > 0) ? (ub - 1) : (ub + 1), str, chunk_sz, \ 1000 TRUE); \ 1001 status = \ 1002 KMP_DISPATCH_NEXT_ULL(&loc, gtid, NULL, (kmp_uint64 *)p_lb, \ 1003 (kmp_uint64 *)p_ub, (kmp_int64 *)&stride); \ 1004 if (status) { \ 1005 KMP_DEBUG_ASSERT((long long)stride == str); \ 1006 *p_ub += (str > 0) ? 1 : -1; \ 1007 } \ 1008 } else { \ 1009 status = 0; \ 1010 } \ 1011 KMP_DOACROSS_FINI(status, gtid); \ 1012 \ 1013 KA_TRACE( \ 1014 20, \ 1015 (KMP_STR( \ 1016 func) " exit: T#%d, *p_lb 0x%llx, *p_ub 0x%llx, returning %d\n", \ 1017 gtid, *p_lb, *p_ub, status)); \ 1018 __kmp_free(dims); \ 1019 return status; \ 1020 } 1021 1022 LOOP_DOACROSS_START_ULL( 1023 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_DOACROSS_STATIC_START), 1024 kmp_sch_static) 1025 LOOP_DOACROSS_START_ULL( 1026 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_DOACROSS_DYNAMIC_START), 1027 kmp_sch_dynamic_chunked) 1028 LOOP_DOACROSS_START_ULL( 1029 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_DOACROSS_GUIDED_START), 1030 kmp_sch_guided_chunked) 1031 LOOP_DOACROSS_RUNTIME_START_ULL( 1032 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_DOACROSS_RUNTIME_START), 1033 kmp_sch_runtime) 1034 #endif 1035 1036 // Combined parallel / loop worksharing constructs 1037 // 1038 // There are no ull versions (yet). 1039 1040 #define PARALLEL_LOOP_START(func, schedule, ompt_pre, ompt_post) \ 1041 void func(void (*task)(void *), void *data, unsigned num_threads, long lb, \ 1042 long ub, long str, long chunk_sz) { \ 1043 int gtid = __kmp_entry_gtid(); \ 1044 MKLOC(loc, KMP_STR(func)); \ 1045 KA_TRACE( \ 1046 20, \ 1047 (KMP_STR( \ 1048 func) ": T#%d, lb 0x%lx, ub 0x%lx, str 0x%lx, chunk_sz 0x%lx\n", \ 1049 gtid, lb, ub, str, chunk_sz)); \ 1050 \ 1051 ompt_pre(); \ 1052 \ 1053 if (__kmpc_ok_to_fork(&loc) && (num_threads != 1)) { \ 1054 if (num_threads != 0) { \ 1055 __kmp_push_num_threads(&loc, gtid, num_threads); \ 1056 } \ 1057 __kmp_GOMP_fork_call(&loc, gtid, task, \ 1058 (microtask_t)__kmp_GOMP_parallel_microtask_wrapper, \ 1059 9, task, data, num_threads, &loc, (schedule), lb, \ 1060 (str > 0) ? (ub - 1) : (ub + 1), str, chunk_sz); \ 1061 IF_OMPT_SUPPORT(OMPT_STORE_RETURN_ADDRESS(gtid)); \ 1062 } else { \ 1063 __kmp_GOMP_serialized_parallel(&loc, gtid, task); \ 1064 IF_OMPT_SUPPORT(OMPT_STORE_RETURN_ADDRESS(gtid)); \ 1065 } \ 1066 \ 1067 KMP_DISPATCH_INIT(&loc, gtid, (schedule), lb, \ 1068 (str > 0) ? (ub - 1) : (ub + 1), str, chunk_sz, \ 1069 (schedule) != kmp_sch_static); \ 1070 \ 1071 ompt_post(); \ 1072 \ 1073 KA_TRACE(20, (KMP_STR(func) " exit: T#%d\n", gtid)); \ 1074 } 1075 1076 #if OMPT_SUPPORT && OMPT_OPTIONAL 1077 1078 #define OMPT_LOOP_PRE() \ 1079 omp_frame_t *parent_frame; \ 1080 if (ompt_enabled.enabled) { \ 1081 __ompt_get_task_info_internal(0, NULL, NULL, &parent_frame, NULL, NULL); \ 1082 parent_frame->enter_frame = OMPT_GET_FRAME_ADDRESS(1); \ 1083 OMPT_STORE_RETURN_ADDRESS(gtid); \ 1084 } 1085 1086 #define OMPT_LOOP_POST() \ 1087 if (ompt_enabled.enabled) { \ 1088 parent_frame->enter_frame = NULL; \ 1089 } 1090 1091 #else 1092 1093 #define OMPT_LOOP_PRE() 1094 1095 #define OMPT_LOOP_POST() 1096 1097 #endif 1098 1099 PARALLEL_LOOP_START( 1100 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_LOOP_STATIC_START), 1101 kmp_sch_static, OMPT_LOOP_PRE, OMPT_LOOP_POST) 1102 PARALLEL_LOOP_START( 1103 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_LOOP_DYNAMIC_START), 1104 kmp_sch_dynamic_chunked, OMPT_LOOP_PRE, OMPT_LOOP_POST) 1105 PARALLEL_LOOP_START( 1106 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_LOOP_GUIDED_START), 1107 kmp_sch_guided_chunked, OMPT_LOOP_PRE, OMPT_LOOP_POST) 1108 PARALLEL_LOOP_START( 1109 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_LOOP_RUNTIME_START), 1110 kmp_sch_runtime, OMPT_LOOP_PRE, OMPT_LOOP_POST) 1111 1112 // Tasking constructs 1113 1114 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASK)(void (*func)(void *), void *data, 1115 void (*copy_func)(void *, void *), 1116 long arg_size, long arg_align, 1117 bool if_cond, unsigned gomp_flags 1118 #if OMP_40_ENABLED 1119 , 1120 void **depend 1121 #endif 1122 ) { 1123 MKLOC(loc, "GOMP_task"); 1124 int gtid = __kmp_entry_gtid(); 1125 kmp_int32 flags = 0; 1126 kmp_tasking_flags_t *input_flags = (kmp_tasking_flags_t *)&flags; 1127 1128 KA_TRACE(20, ("GOMP_task: T#%d\n", gtid)); 1129 1130 // The low-order bit is the "untied" flag 1131 if (!(gomp_flags & 1)) { 1132 input_flags->tiedness = 1; 1133 } 1134 // The second low-order bit is the "final" flag 1135 if (gomp_flags & 2) { 1136 input_flags->final = 1; 1137 } 1138 input_flags->native = 1; 1139 // __kmp_task_alloc() sets up all other flags 1140 1141 if (!if_cond) { 1142 arg_size = 0; 1143 } 1144 1145 kmp_task_t *task = __kmp_task_alloc( 1146 &loc, gtid, input_flags, sizeof(kmp_task_t), 1147 arg_size ? arg_size + arg_align - 1 : 0, (kmp_routine_entry_t)func); 1148 1149 if (arg_size > 0) { 1150 if (arg_align > 0) { 1151 task->shareds = (void *)((((size_t)task->shareds) + arg_align - 1) / 1152 arg_align * arg_align); 1153 } 1154 // else error?? 1155 1156 if (copy_func) { 1157 (*copy_func)(task->shareds, data); 1158 } else { 1159 KMP_MEMCPY(task->shareds, data, arg_size); 1160 } 1161 } 1162 1163 #if OMPT_SUPPORT 1164 kmp_taskdata_t *current_task; 1165 if (ompt_enabled.enabled) { 1166 OMPT_STORE_RETURN_ADDRESS(gtid); 1167 current_task = __kmp_threads[gtid]->th.th_current_task; 1168 current_task->ompt_task_info.frame.enter_frame = OMPT_GET_FRAME_ADDRESS(1); 1169 } 1170 #endif 1171 1172 if (if_cond) { 1173 #if OMP_40_ENABLED 1174 if (gomp_flags & 8) { 1175 KMP_ASSERT(depend); 1176 const size_t ndeps = (kmp_intptr_t)depend[0]; 1177 const size_t nout = (kmp_intptr_t)depend[1]; 1178 kmp_depend_info_t dep_list[ndeps]; 1179 1180 for (size_t i = 0U; i < ndeps; i++) { 1181 dep_list[i].base_addr = (kmp_intptr_t)depend[2U + i]; 1182 dep_list[i].len = 0U; 1183 dep_list[i].flags.in = 1; 1184 dep_list[i].flags.out = (i < nout); 1185 } 1186 __kmpc_omp_task_with_deps(&loc, gtid, task, ndeps, dep_list, 0, NULL); 1187 } else { 1188 #endif 1189 __kmpc_omp_task(&loc, gtid, task); 1190 } 1191 } else { 1192 #if OMPT_SUPPORT 1193 ompt_thread_info_t oldInfo; 1194 kmp_info_t *thread; 1195 kmp_taskdata_t *taskdata; 1196 kmp_taskdata_t *current_task; 1197 if (ompt_enabled.enabled) { 1198 // Store the threads states and restore them after the task 1199 thread = __kmp_threads[gtid]; 1200 taskdata = KMP_TASK_TO_TASKDATA(task); 1201 oldInfo = thread->th.ompt_thread_info; 1202 thread->th.ompt_thread_info.wait_id = 0; 1203 thread->th.ompt_thread_info.state = omp_state_work_parallel; 1204 taskdata->ompt_task_info.frame.exit_frame = OMPT_GET_FRAME_ADDRESS(0); 1205 OMPT_STORE_RETURN_ADDRESS(gtid); 1206 } 1207 #endif 1208 1209 __kmpc_omp_task_begin_if0(&loc, gtid, task); 1210 func(data); 1211 __kmpc_omp_task_complete_if0(&loc, gtid, task); 1212 1213 #if OMPT_SUPPORT 1214 if (ompt_enabled.enabled) { 1215 thread->th.ompt_thread_info = oldInfo; 1216 taskdata->ompt_task_info.frame.exit_frame = NULL; 1217 } 1218 #endif 1219 } 1220 #if OMPT_SUPPORT 1221 if (ompt_enabled.enabled) { 1222 current_task->ompt_task_info.frame.enter_frame = NULL; 1223 } 1224 #endif 1225 1226 KA_TRACE(20, ("GOMP_task exit: T#%d\n", gtid)); 1227 } 1228 1229 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASKWAIT)(void) { 1230 MKLOC(loc, "GOMP_taskwait"); 1231 int gtid = __kmp_entry_gtid(); 1232 1233 #if OMPT_SUPPORT 1234 if (ompt_enabled.enabled) 1235 OMPT_STORE_RETURN_ADDRESS(gtid); 1236 #endif 1237 1238 KA_TRACE(20, ("GOMP_taskwait: T#%d\n", gtid)); 1239 1240 __kmpc_omp_taskwait(&loc, gtid); 1241 1242 KA_TRACE(20, ("GOMP_taskwait exit: T#%d\n", gtid)); 1243 } 1244 1245 // Sections worksharing constructs 1246 // 1247 // For the sections construct, we initialize a dynamically scheduled loop 1248 // worksharing construct with lb 1 and stride 1, and use the iteration #'s 1249 // that its returns as sections ids. 1250 // 1251 // There are no special entry points for ordered sections, so we always use 1252 // the dynamically scheduled workshare, even if the sections aren't ordered. 1253 1254 unsigned KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SECTIONS_START)(unsigned count) { 1255 int status; 1256 kmp_int lb, ub, stride; 1257 int gtid = __kmp_entry_gtid(); 1258 MKLOC(loc, "GOMP_sections_start"); 1259 KA_TRACE(20, ("GOMP_sections_start: T#%d\n", gtid)); 1260 1261 KMP_DISPATCH_INIT(&loc, gtid, kmp_nm_dynamic_chunked, 1, count, 1, 1, TRUE); 1262 1263 status = KMP_DISPATCH_NEXT(&loc, gtid, NULL, &lb, &ub, &stride); 1264 if (status) { 1265 KMP_DEBUG_ASSERT(stride == 1); 1266 KMP_DEBUG_ASSERT(lb > 0); 1267 KMP_ASSERT(lb == ub); 1268 } else { 1269 lb = 0; 1270 } 1271 1272 KA_TRACE(20, ("GOMP_sections_start exit: T#%d returning %u\n", gtid, 1273 (unsigned)lb)); 1274 return (unsigned)lb; 1275 } 1276 1277 unsigned KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SECTIONS_NEXT)(void) { 1278 int status; 1279 kmp_int lb, ub, stride; 1280 int gtid = __kmp_get_gtid(); 1281 MKLOC(loc, "GOMP_sections_next"); 1282 KA_TRACE(20, ("GOMP_sections_next: T#%d\n", gtid)); 1283 1284 #if OMPT_SUPPORT 1285 OMPT_STORE_RETURN_ADDRESS(gtid); 1286 #endif 1287 1288 status = KMP_DISPATCH_NEXT(&loc, gtid, NULL, &lb, &ub, &stride); 1289 if (status) { 1290 KMP_DEBUG_ASSERT(stride == 1); 1291 KMP_DEBUG_ASSERT(lb > 0); 1292 KMP_ASSERT(lb == ub); 1293 } else { 1294 lb = 0; 1295 } 1296 1297 KA_TRACE( 1298 20, ("GOMP_sections_next exit: T#%d returning %u\n", gtid, (unsigned)lb)); 1299 return (unsigned)lb; 1300 } 1301 1302 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_SECTIONS_START)( 1303 void (*task)(void *), void *data, unsigned num_threads, unsigned count) { 1304 int gtid = __kmp_entry_gtid(); 1305 1306 #if OMPT_SUPPORT 1307 omp_frame_t *parent_frame; 1308 1309 if (ompt_enabled.enabled) { 1310 __ompt_get_task_info_internal(0, NULL, NULL, &parent_frame, NULL, NULL); 1311 parent_frame->enter_frame = OMPT_GET_FRAME_ADDRESS(1); 1312 OMPT_STORE_RETURN_ADDRESS(gtid); 1313 } 1314 #endif 1315 1316 MKLOC(loc, "GOMP_parallel_sections_start"); 1317 KA_TRACE(20, ("GOMP_parallel_sections_start: T#%d\n", gtid)); 1318 1319 if (__kmpc_ok_to_fork(&loc) && (num_threads != 1)) { 1320 if (num_threads != 0) { 1321 __kmp_push_num_threads(&loc, gtid, num_threads); 1322 } 1323 __kmp_GOMP_fork_call(&loc, gtid, task, 1324 (microtask_t)__kmp_GOMP_parallel_microtask_wrapper, 9, 1325 task, data, num_threads, &loc, kmp_nm_dynamic_chunked, 1326 (kmp_int)1, (kmp_int)count, (kmp_int)1, (kmp_int)1); 1327 } else { 1328 __kmp_GOMP_serialized_parallel(&loc, gtid, task); 1329 } 1330 1331 #if OMPT_SUPPORT 1332 if (ompt_enabled.enabled) { 1333 parent_frame->enter_frame = NULL; 1334 } 1335 #endif 1336 1337 KMP_DISPATCH_INIT(&loc, gtid, kmp_nm_dynamic_chunked, 1, count, 1, 1, TRUE); 1338 1339 KA_TRACE(20, ("GOMP_parallel_sections_start exit: T#%d\n", gtid)); 1340 } 1341 1342 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SECTIONS_END)(void) { 1343 int gtid = __kmp_get_gtid(); 1344 KA_TRACE(20, ("GOMP_sections_end: T#%d\n", gtid)) 1345 1346 #if OMPT_SUPPORT 1347 omp_frame_t *ompt_frame; 1348 if (ompt_enabled.enabled) { 1349 __ompt_get_task_info_internal(0, NULL, NULL, &ompt_frame, NULL, NULL); 1350 ompt_frame->enter_frame = OMPT_GET_FRAME_ADDRESS(1); 1351 OMPT_STORE_RETURN_ADDRESS(gtid); 1352 } 1353 #endif 1354 __kmp_barrier(bs_plain_barrier, gtid, FALSE, 0, NULL, NULL); 1355 #if OMPT_SUPPORT 1356 if (ompt_enabled.enabled) { 1357 ompt_frame->enter_frame = NULL; 1358 } 1359 #endif 1360 1361 KA_TRACE(20, ("GOMP_sections_end exit: T#%d\n", gtid)) 1362 } 1363 1364 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SECTIONS_END_NOWAIT)(void) { 1365 KA_TRACE(20, ("GOMP_sections_end_nowait: T#%d\n", __kmp_get_gtid())) 1366 } 1367 1368 // libgomp has an empty function for GOMP_taskyield as of 2013-10-10 1369 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASKYIELD)(void) { 1370 KA_TRACE(20, ("GOMP_taskyield: T#%d\n", __kmp_get_gtid())) 1371 return; 1372 } 1373 1374 #if OMP_40_ENABLED // these are new GOMP_4.0 entry points 1375 1376 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL)(void (*task)(void *), 1377 void *data, 1378 unsigned num_threads, 1379 unsigned int flags) { 1380 int gtid = __kmp_entry_gtid(); 1381 MKLOC(loc, "GOMP_parallel"); 1382 KA_TRACE(20, ("GOMP_parallel: T#%d\n", gtid)); 1383 1384 #if OMPT_SUPPORT 1385 ompt_task_info_t *parent_task_info, *task_info; 1386 if (ompt_enabled.enabled) { 1387 parent_task_info = __ompt_get_task_info_object(0); 1388 parent_task_info->frame.enter_frame = OMPT_GET_FRAME_ADDRESS(1); 1389 OMPT_STORE_RETURN_ADDRESS(gtid); 1390 } 1391 #endif 1392 if (__kmpc_ok_to_fork(&loc) && (num_threads != 1)) { 1393 if (num_threads != 0) { 1394 __kmp_push_num_threads(&loc, gtid, num_threads); 1395 } 1396 if (flags != 0) { 1397 __kmp_push_proc_bind(&loc, gtid, (kmp_proc_bind_t)flags); 1398 } 1399 __kmp_GOMP_fork_call(&loc, gtid, task, 1400 (microtask_t)__kmp_GOMP_microtask_wrapper, 2, task, 1401 data); 1402 } else { 1403 __kmp_GOMP_serialized_parallel(&loc, gtid, task); 1404 } 1405 #if OMPT_SUPPORT 1406 if (ompt_enabled.enabled) { 1407 task_info = __ompt_get_task_info_object(0); 1408 task_info->frame.exit_frame = OMPT_GET_FRAME_ADDRESS(0); 1409 } 1410 #endif 1411 task(data); 1412 #if OMPT_SUPPORT 1413 if (ompt_enabled.enabled) { 1414 OMPT_STORE_RETURN_ADDRESS(gtid); 1415 } 1416 #endif 1417 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_END)(); 1418 #if OMPT_SUPPORT 1419 if (ompt_enabled.enabled) { 1420 task_info->frame.exit_frame = NULL; 1421 parent_task_info->frame.enter_frame = NULL; 1422 } 1423 #endif 1424 } 1425 1426 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_SECTIONS)(void (*task)(void *), 1427 void *data, 1428 unsigned num_threads, 1429 unsigned count, 1430 unsigned flags) { 1431 int gtid = __kmp_entry_gtid(); 1432 MKLOC(loc, "GOMP_parallel_sections"); 1433 KA_TRACE(20, ("GOMP_parallel_sections: T#%d\n", gtid)); 1434 1435 #if OMPT_SUPPORT 1436 OMPT_STORE_RETURN_ADDRESS(gtid); 1437 #endif 1438 1439 if (__kmpc_ok_to_fork(&loc) && (num_threads != 1)) { 1440 if (num_threads != 0) { 1441 __kmp_push_num_threads(&loc, gtid, num_threads); 1442 } 1443 if (flags != 0) { 1444 __kmp_push_proc_bind(&loc, gtid, (kmp_proc_bind_t)flags); 1445 } 1446 __kmp_GOMP_fork_call(&loc, gtid, task, 1447 (microtask_t)__kmp_GOMP_parallel_microtask_wrapper, 9, 1448 task, data, num_threads, &loc, kmp_nm_dynamic_chunked, 1449 (kmp_int)1, (kmp_int)count, (kmp_int)1, (kmp_int)1); 1450 } else { 1451 __kmp_GOMP_serialized_parallel(&loc, gtid, task); 1452 } 1453 1454 #if OMPT_SUPPORT 1455 OMPT_STORE_RETURN_ADDRESS(gtid); 1456 #endif 1457 1458 KMP_DISPATCH_INIT(&loc, gtid, kmp_nm_dynamic_chunked, 1, count, 1, 1, TRUE); 1459 1460 task(data); 1461 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_END)(); 1462 KA_TRACE(20, ("GOMP_parallel_sections exit: T#%d\n", gtid)); 1463 } 1464 1465 #define PARALLEL_LOOP(func, schedule, ompt_pre, ompt_post) \ 1466 void func(void (*task)(void *), void *data, unsigned num_threads, long lb, \ 1467 long ub, long str, long chunk_sz, unsigned flags) { \ 1468 int gtid = __kmp_entry_gtid(); \ 1469 MKLOC(loc, KMP_STR(func)); \ 1470 KA_TRACE( \ 1471 20, \ 1472 (KMP_STR( \ 1473 func) ": T#%d, lb 0x%lx, ub 0x%lx, str 0x%lx, chunk_sz 0x%lx\n", \ 1474 gtid, lb, ub, str, chunk_sz)); \ 1475 \ 1476 ompt_pre(); \ 1477 if (__kmpc_ok_to_fork(&loc) && (num_threads != 1)) { \ 1478 if (num_threads != 0) { \ 1479 __kmp_push_num_threads(&loc, gtid, num_threads); \ 1480 } \ 1481 if (flags != 0) { \ 1482 __kmp_push_proc_bind(&loc, gtid, (kmp_proc_bind_t)flags); \ 1483 } \ 1484 __kmp_GOMP_fork_call(&loc, gtid, task, \ 1485 (microtask_t)__kmp_GOMP_parallel_microtask_wrapper, \ 1486 9, task, data, num_threads, &loc, (schedule), lb, \ 1487 (str > 0) ? (ub - 1) : (ub + 1), str, chunk_sz); \ 1488 } else { \ 1489 __kmp_GOMP_serialized_parallel(&loc, gtid, task); \ 1490 } \ 1491 \ 1492 IF_OMPT_SUPPORT(OMPT_STORE_RETURN_ADDRESS(gtid);) \ 1493 KMP_DISPATCH_INIT(&loc, gtid, (schedule), lb, \ 1494 (str > 0) ? (ub - 1) : (ub + 1), str, chunk_sz, \ 1495 (schedule) != kmp_sch_static); \ 1496 task(data); \ 1497 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_END)(); \ 1498 ompt_post(); \ 1499 \ 1500 KA_TRACE(20, (KMP_STR(func) " exit: T#%d\n", gtid)); \ 1501 } 1502 1503 PARALLEL_LOOP(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_LOOP_STATIC), 1504 kmp_sch_static, OMPT_LOOP_PRE, OMPT_LOOP_POST) 1505 PARALLEL_LOOP(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_LOOP_DYNAMIC), 1506 kmp_sch_dynamic_chunked, OMPT_LOOP_PRE, OMPT_LOOP_POST) 1507 PARALLEL_LOOP(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_LOOP_GUIDED), 1508 kmp_sch_guided_chunked, OMPT_LOOP_PRE, OMPT_LOOP_POST) 1509 PARALLEL_LOOP(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_LOOP_RUNTIME), 1510 kmp_sch_runtime, OMPT_LOOP_PRE, OMPT_LOOP_POST) 1511 1512 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASKGROUP_START)(void) { 1513 int gtid = __kmp_entry_gtid(); 1514 MKLOC(loc, "GOMP_taskgroup_start"); 1515 KA_TRACE(20, ("GOMP_taskgroup_start: T#%d\n", gtid)); 1516 1517 #if OMPT_SUPPORT 1518 if (ompt_enabled.enabled) 1519 OMPT_STORE_RETURN_ADDRESS(gtid); 1520 #endif 1521 1522 __kmpc_taskgroup(&loc, gtid); 1523 1524 return; 1525 } 1526 1527 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASKGROUP_END)(void) { 1528 int gtid = __kmp_get_gtid(); 1529 MKLOC(loc, "GOMP_taskgroup_end"); 1530 KA_TRACE(20, ("GOMP_taskgroup_end: T#%d\n", gtid)); 1531 1532 #if OMPT_SUPPORT 1533 if (ompt_enabled.enabled) 1534 OMPT_STORE_RETURN_ADDRESS(gtid); 1535 #endif 1536 1537 __kmpc_end_taskgroup(&loc, gtid); 1538 1539 return; 1540 } 1541 1542 #ifndef KMP_DEBUG 1543 static 1544 #endif /* KMP_DEBUG */ 1545 kmp_int32 1546 __kmp_gomp_to_omp_cancellation_kind(int gomp_kind) { 1547 kmp_int32 cncl_kind = 0; 1548 switch (gomp_kind) { 1549 case 1: 1550 cncl_kind = cancel_parallel; 1551 break; 1552 case 2: 1553 cncl_kind = cancel_loop; 1554 break; 1555 case 4: 1556 cncl_kind = cancel_sections; 1557 break; 1558 case 8: 1559 cncl_kind = cancel_taskgroup; 1560 break; 1561 } 1562 return cncl_kind; 1563 } 1564 1565 bool KMP_EXPAND_NAME(KMP_API_NAME_GOMP_CANCELLATION_POINT)(int which) { 1566 if (__kmp_omp_cancellation) { 1567 KMP_FATAL(NoGompCancellation); 1568 } 1569 int gtid = __kmp_get_gtid(); 1570 MKLOC(loc, "GOMP_cancellation_point"); 1571 KA_TRACE(20, ("GOMP_cancellation_point: T#%d\n", gtid)); 1572 1573 kmp_int32 cncl_kind = __kmp_gomp_to_omp_cancellation_kind(which); 1574 1575 return __kmpc_cancellationpoint(&loc, gtid, cncl_kind); 1576 } 1577 1578 bool KMP_EXPAND_NAME(KMP_API_NAME_GOMP_BARRIER_CANCEL)(void) { 1579 if (__kmp_omp_cancellation) { 1580 KMP_FATAL(NoGompCancellation); 1581 } 1582 KMP_FATAL(NoGompCancellation); 1583 int gtid = __kmp_get_gtid(); 1584 MKLOC(loc, "GOMP_barrier_cancel"); 1585 KA_TRACE(20, ("GOMP_barrier_cancel: T#%d\n", gtid)); 1586 1587 return __kmpc_cancel_barrier(&loc, gtid); 1588 } 1589 1590 bool KMP_EXPAND_NAME(KMP_API_NAME_GOMP_CANCEL)(int which, bool do_cancel) { 1591 if (__kmp_omp_cancellation) { 1592 KMP_FATAL(NoGompCancellation); 1593 } else { 1594 return FALSE; 1595 } 1596 1597 int gtid = __kmp_get_gtid(); 1598 MKLOC(loc, "GOMP_cancel"); 1599 KA_TRACE(20, ("GOMP_cancel: T#%d\n", gtid)); 1600 1601 kmp_int32 cncl_kind = __kmp_gomp_to_omp_cancellation_kind(which); 1602 1603 if (do_cancel == FALSE) { 1604 return KMP_EXPAND_NAME(KMP_API_NAME_GOMP_CANCELLATION_POINT)(which); 1605 } else { 1606 return __kmpc_cancel(&loc, gtid, cncl_kind); 1607 } 1608 } 1609 1610 bool KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SECTIONS_END_CANCEL)(void) { 1611 if (__kmp_omp_cancellation) { 1612 KMP_FATAL(NoGompCancellation); 1613 } 1614 int gtid = __kmp_get_gtid(); 1615 MKLOC(loc, "GOMP_sections_end_cancel"); 1616 KA_TRACE(20, ("GOMP_sections_end_cancel: T#%d\n", gtid)); 1617 1618 return __kmpc_cancel_barrier(&loc, gtid); 1619 } 1620 1621 bool KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_END_CANCEL)(void) { 1622 if (__kmp_omp_cancellation) { 1623 KMP_FATAL(NoGompCancellation); 1624 } 1625 int gtid = __kmp_get_gtid(); 1626 MKLOC(loc, "GOMP_loop_end_cancel"); 1627 KA_TRACE(20, ("GOMP_loop_end_cancel: T#%d\n", gtid)); 1628 1629 return __kmpc_cancel_barrier(&loc, gtid); 1630 } 1631 1632 // All target functions are empty as of 2014-05-29 1633 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TARGET)(int device, void (*fn)(void *), 1634 const void *openmp_target, 1635 size_t mapnum, void **hostaddrs, 1636 size_t *sizes, 1637 unsigned char *kinds) { 1638 return; 1639 } 1640 1641 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TARGET_DATA)( 1642 int device, const void *openmp_target, size_t mapnum, void **hostaddrs, 1643 size_t *sizes, unsigned char *kinds) { 1644 return; 1645 } 1646 1647 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TARGET_END_DATA)(void) { return; } 1648 1649 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TARGET_UPDATE)( 1650 int device, const void *openmp_target, size_t mapnum, void **hostaddrs, 1651 size_t *sizes, unsigned char *kinds) { 1652 return; 1653 } 1654 1655 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TEAMS)(unsigned int num_teams, 1656 unsigned int thread_limit) { 1657 return; 1658 } 1659 #endif // OMP_40_ENABLED 1660 1661 #if OMP_45_ENABLED 1662 1663 // Task duplication function which copies src to dest (both are 1664 // preallocated task structures) 1665 static void __kmp_gomp_task_dup(kmp_task_t *dest, kmp_task_t *src, 1666 kmp_int32 last_private) { 1667 kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(src); 1668 if (taskdata->td_copy_func) { 1669 (taskdata->td_copy_func)(dest->shareds, src->shareds); 1670 } 1671 } 1672 1673 #ifdef __cplusplus 1674 } // extern "C" 1675 #endif 1676 1677 template <typename T> 1678 void __GOMP_taskloop(void (*func)(void *), void *data, 1679 void (*copy_func)(void *, void *), long arg_size, 1680 long arg_align, unsigned gomp_flags, 1681 unsigned long num_tasks, int priority, T start, T end, 1682 T step) { 1683 typedef void (*p_task_dup_t)(kmp_task_t *, kmp_task_t *, kmp_int32); 1684 MKLOC(loc, "GOMP_taskloop"); 1685 int sched; 1686 T *loop_bounds; 1687 int gtid = __kmp_entry_gtid(); 1688 kmp_int32 flags = 0; 1689 int if_val = gomp_flags & (1u << 10); 1690 int nogroup = gomp_flags & (1u << 11); 1691 int up = gomp_flags & (1u << 8); 1692 p_task_dup_t task_dup = NULL; 1693 kmp_tasking_flags_t *input_flags = (kmp_tasking_flags_t *)&flags; 1694 #ifdef KMP_DEBUG 1695 { 1696 char *buff; 1697 buff = __kmp_str_format( 1698 "GOMP_taskloop: T#%%d: func:%%p data:%%p copy_func:%%p " 1699 "arg_size:%%ld arg_align:%%ld gomp_flags:0x%%x num_tasks:%%lu " 1700 "priority:%%d start:%%%s end:%%%s step:%%%s\n", 1701 traits_t<T>::spec, traits_t<T>::spec, traits_t<T>::spec); 1702 KA_TRACE(20, (buff, gtid, func, data, copy_func, arg_size, arg_align, 1703 gomp_flags, num_tasks, priority, start, end, step)); 1704 __kmp_str_free(&buff); 1705 } 1706 #endif 1707 KMP_ASSERT((size_t)arg_size >= 2 * sizeof(T)); 1708 KMP_ASSERT(arg_align > 0); 1709 // The low-order bit is the "untied" flag 1710 if (!(gomp_flags & 1)) { 1711 input_flags->tiedness = 1; 1712 } 1713 // The second low-order bit is the "final" flag 1714 if (gomp_flags & 2) { 1715 input_flags->final = 1; 1716 } 1717 // Negative step flag 1718 if (!up) { 1719 // If step is flagged as negative, but isn't properly sign extended 1720 // Then manually sign extend it. Could be a short, int, char embedded 1721 // in a long. So cannot assume any cast. 1722 if (step > 0) { 1723 for (int i = sizeof(T) * CHAR_BIT - 1; i >= 0L; --i) { 1724 // break at the first 1 bit 1725 if (step & ((T)1 << i)) 1726 break; 1727 step |= ((T)1 << i); 1728 } 1729 } 1730 } 1731 input_flags->native = 1; 1732 // Figure out if none/grainsize/num_tasks clause specified 1733 if (num_tasks > 0) { 1734 if (gomp_flags & (1u << 9)) 1735 sched = 1; // grainsize specified 1736 else 1737 sched = 2; // num_tasks specified 1738 // neither grainsize nor num_tasks specified 1739 } else { 1740 sched = 0; 1741 } 1742 1743 // __kmp_task_alloc() sets up all other flags 1744 kmp_task_t *task = 1745 __kmp_task_alloc(&loc, gtid, input_flags, sizeof(kmp_task_t), 1746 arg_size + arg_align - 1, (kmp_routine_entry_t)func); 1747 kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task); 1748 taskdata->td_copy_func = copy_func; 1749 taskdata->td_size_loop_bounds = sizeof(T); 1750 1751 // re-align shareds if needed and setup firstprivate copy constructors 1752 // through the task_dup mechanism 1753 task->shareds = (void *)((((size_t)task->shareds) + arg_align - 1) / 1754 arg_align * arg_align); 1755 if (copy_func) { 1756 task_dup = __kmp_gomp_task_dup; 1757 } 1758 KMP_MEMCPY(task->shareds, data, arg_size); 1759 1760 loop_bounds = (T *)task->shareds; 1761 loop_bounds[0] = start; 1762 loop_bounds[1] = end + (up ? -1 : 1); 1763 __kmpc_taskloop(&loc, gtid, task, if_val, (kmp_uint64 *)&(loop_bounds[0]), 1764 (kmp_uint64 *)&(loop_bounds[1]), (kmp_int64)step, nogroup, 1765 sched, (kmp_uint64)num_tasks, (void *)task_dup); 1766 } 1767 1768 // 4 byte version of GOMP_doacross_post 1769 // This verison needs to create a temporary array which converts 4 byte 1770 // integers into 8 byte integeres 1771 template <typename T, bool need_conversion = (sizeof(long) == 4)> 1772 void __kmp_GOMP_doacross_post(T *count); 1773 1774 template <> void __kmp_GOMP_doacross_post<long, true>(long *count) { 1775 int gtid = __kmp_entry_gtid(); 1776 kmp_info_t *th = __kmp_threads[gtid]; 1777 MKLOC(loc, "GOMP_doacross_post"); 1778 kmp_int64 num_dims = th->th.th_dispatch->th_doacross_info[0]; 1779 kmp_int64 *vec = 1780 (kmp_int64 *)__kmp_thread_malloc(th, sizeof(kmp_int64) * num_dims); 1781 for (kmp_int64 i = 0; i < num_dims; ++i) { 1782 vec[i] = (kmp_int64)count[i]; 1783 } 1784 __kmpc_doacross_post(&loc, gtid, vec); 1785 __kmp_thread_free(th, vec); 1786 } 1787 1788 // 8 byte versions of GOMP_doacross_post 1789 // This version can just pass in the count array directly instead of creating 1790 // a temporary array 1791 template <> void __kmp_GOMP_doacross_post<long, false>(long *count) { 1792 int gtid = __kmp_entry_gtid(); 1793 MKLOC(loc, "GOMP_doacross_post"); 1794 __kmpc_doacross_post(&loc, gtid, RCAST(kmp_int64 *, count)); 1795 } 1796 1797 template <typename T> void __kmp_GOMP_doacross_wait(T first, va_list args) { 1798 int gtid = __kmp_entry_gtid(); 1799 kmp_info_t *th = __kmp_threads[gtid]; 1800 MKLOC(loc, "GOMP_doacross_wait"); 1801 kmp_int64 num_dims = th->th.th_dispatch->th_doacross_info[0]; 1802 kmp_int64 *vec = 1803 (kmp_int64 *)__kmp_thread_malloc(th, sizeof(kmp_int64) * num_dims); 1804 vec[0] = (kmp_int64)first; 1805 for (kmp_int64 i = 1; i < num_dims; ++i) { 1806 T item = va_arg(args, T); 1807 vec[i] = (kmp_int64)item; 1808 } 1809 __kmpc_doacross_wait(&loc, gtid, vec); 1810 __kmp_thread_free(th, vec); 1811 return; 1812 } 1813 1814 #ifdef __cplusplus 1815 extern "C" { 1816 #endif // __cplusplus 1817 1818 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASKLOOP)( 1819 void (*func)(void *), void *data, void (*copy_func)(void *, void *), 1820 long arg_size, long arg_align, unsigned gomp_flags, unsigned long num_tasks, 1821 int priority, long start, long end, long step) { 1822 __GOMP_taskloop<long>(func, data, copy_func, arg_size, arg_align, gomp_flags, 1823 num_tasks, priority, start, end, step); 1824 } 1825 1826 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASKLOOP_ULL)( 1827 void (*func)(void *), void *data, void (*copy_func)(void *, void *), 1828 long arg_size, long arg_align, unsigned gomp_flags, unsigned long num_tasks, 1829 int priority, unsigned long long start, unsigned long long end, 1830 unsigned long long step) { 1831 __GOMP_taskloop<unsigned long long>(func, data, copy_func, arg_size, 1832 arg_align, gomp_flags, num_tasks, 1833 priority, start, end, step); 1834 } 1835 1836 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_DOACROSS_POST)(long *count) { 1837 __kmp_GOMP_doacross_post(count); 1838 } 1839 1840 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_DOACROSS_WAIT)(long first, ...) { 1841 va_list args; 1842 va_start(args, first); 1843 __kmp_GOMP_doacross_wait<long>(first, args); 1844 va_end(args); 1845 } 1846 1847 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_DOACROSS_ULL_POST)( 1848 unsigned long long *count) { 1849 int gtid = __kmp_entry_gtid(); 1850 MKLOC(loc, "GOMP_doacross_ull_post"); 1851 __kmpc_doacross_post(&loc, gtid, RCAST(kmp_int64 *, count)); 1852 } 1853 1854 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_DOACROSS_ULL_WAIT)( 1855 unsigned long long first, ...) { 1856 va_list args; 1857 va_start(args, first); 1858 __kmp_GOMP_doacross_wait<unsigned long long>(first, args); 1859 va_end(args); 1860 } 1861 1862 #endif // OMP_45_ENABLED 1863 1864 /* The following sections of code create aliases for the GOMP_* functions, then 1865 create versioned symbols using the assembler directive .symver. This is only 1866 pertinent for ELF .so library. The KMP_VERSION_SYMBOL macro is defined in 1867 kmp_os.h */ 1868 1869 #ifdef KMP_USE_VERSION_SYMBOLS 1870 // GOMP_1.0 versioned symbols 1871 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_ATOMIC_END, 10, "GOMP_1.0"); 1872 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_ATOMIC_START, 10, "GOMP_1.0"); 1873 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_BARRIER, 10, "GOMP_1.0"); 1874 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_CRITICAL_END, 10, "GOMP_1.0"); 1875 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_CRITICAL_NAME_END, 10, "GOMP_1.0"); 1876 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_CRITICAL_NAME_START, 10, "GOMP_1.0"); 1877 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_CRITICAL_START, 10, "GOMP_1.0"); 1878 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_DYNAMIC_NEXT, 10, "GOMP_1.0"); 1879 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_DYNAMIC_START, 10, "GOMP_1.0"); 1880 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_END, 10, "GOMP_1.0"); 1881 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_END_NOWAIT, 10, "GOMP_1.0"); 1882 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_GUIDED_NEXT, 10, "GOMP_1.0"); 1883 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_GUIDED_START, 10, "GOMP_1.0"); 1884 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ORDERED_DYNAMIC_NEXT, 10, "GOMP_1.0"); 1885 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ORDERED_DYNAMIC_START, 10, 1886 "GOMP_1.0"); 1887 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ORDERED_GUIDED_NEXT, 10, "GOMP_1.0"); 1888 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ORDERED_GUIDED_START, 10, "GOMP_1.0"); 1889 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ORDERED_RUNTIME_NEXT, 10, "GOMP_1.0"); 1890 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ORDERED_RUNTIME_START, 10, 1891 "GOMP_1.0"); 1892 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ORDERED_STATIC_NEXT, 10, "GOMP_1.0"); 1893 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ORDERED_STATIC_START, 10, "GOMP_1.0"); 1894 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_RUNTIME_NEXT, 10, "GOMP_1.0"); 1895 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_RUNTIME_START, 10, "GOMP_1.0"); 1896 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_STATIC_NEXT, 10, "GOMP_1.0"); 1897 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_STATIC_START, 10, "GOMP_1.0"); 1898 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_ORDERED_END, 10, "GOMP_1.0"); 1899 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_ORDERED_START, 10, "GOMP_1.0"); 1900 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_END, 10, "GOMP_1.0"); 1901 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_LOOP_DYNAMIC_START, 10, 1902 "GOMP_1.0"); 1903 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_LOOP_GUIDED_START, 10, 1904 "GOMP_1.0"); 1905 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_LOOP_RUNTIME_START, 10, 1906 "GOMP_1.0"); 1907 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_LOOP_STATIC_START, 10, 1908 "GOMP_1.0"); 1909 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_SECTIONS_START, 10, "GOMP_1.0"); 1910 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_START, 10, "GOMP_1.0"); 1911 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_SECTIONS_END, 10, "GOMP_1.0"); 1912 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_SECTIONS_END_NOWAIT, 10, "GOMP_1.0"); 1913 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_SECTIONS_NEXT, 10, "GOMP_1.0"); 1914 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_SECTIONS_START, 10, "GOMP_1.0"); 1915 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_SINGLE_COPY_END, 10, "GOMP_1.0"); 1916 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_SINGLE_COPY_START, 10, "GOMP_1.0"); 1917 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_SINGLE_START, 10, "GOMP_1.0"); 1918 1919 // GOMP_2.0 versioned symbols 1920 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TASK, 20, "GOMP_2.0"); 1921 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TASKWAIT, 20, "GOMP_2.0"); 1922 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_DYNAMIC_NEXT, 20, "GOMP_2.0"); 1923 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_DYNAMIC_START, 20, "GOMP_2.0"); 1924 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_GUIDED_NEXT, 20, "GOMP_2.0"); 1925 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_GUIDED_START, 20, "GOMP_2.0"); 1926 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_DYNAMIC_NEXT, 20, 1927 "GOMP_2.0"); 1928 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_DYNAMIC_START, 20, 1929 "GOMP_2.0"); 1930 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_GUIDED_NEXT, 20, 1931 "GOMP_2.0"); 1932 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_GUIDED_START, 20, 1933 "GOMP_2.0"); 1934 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_RUNTIME_NEXT, 20, 1935 "GOMP_2.0"); 1936 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_RUNTIME_START, 20, 1937 "GOMP_2.0"); 1938 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_STATIC_NEXT, 20, 1939 "GOMP_2.0"); 1940 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_STATIC_START, 20, 1941 "GOMP_2.0"); 1942 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_RUNTIME_NEXT, 20, "GOMP_2.0"); 1943 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_RUNTIME_START, 20, "GOMP_2.0"); 1944 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_STATIC_NEXT, 20, "GOMP_2.0"); 1945 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_STATIC_START, 20, "GOMP_2.0"); 1946 1947 // GOMP_3.0 versioned symbols 1948 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TASKYIELD, 30, "GOMP_3.0"); 1949 1950 // GOMP_4.0 versioned symbols 1951 #if OMP_40_ENABLED 1952 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL, 40, "GOMP_4.0"); 1953 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_SECTIONS, 40, "GOMP_4.0"); 1954 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_LOOP_DYNAMIC, 40, "GOMP_4.0"); 1955 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_LOOP_GUIDED, 40, "GOMP_4.0"); 1956 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_LOOP_RUNTIME, 40, "GOMP_4.0"); 1957 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_LOOP_STATIC, 40, "GOMP_4.0"); 1958 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TASKGROUP_START, 40, "GOMP_4.0"); 1959 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TASKGROUP_END, 40, "GOMP_4.0"); 1960 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_BARRIER_CANCEL, 40, "GOMP_4.0"); 1961 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_CANCEL, 40, "GOMP_4.0"); 1962 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_CANCELLATION_POINT, 40, "GOMP_4.0"); 1963 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_END_CANCEL, 40, "GOMP_4.0"); 1964 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_SECTIONS_END_CANCEL, 40, "GOMP_4.0"); 1965 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TARGET, 40, "GOMP_4.0"); 1966 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TARGET_DATA, 40, "GOMP_4.0"); 1967 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TARGET_END_DATA, 40, "GOMP_4.0"); 1968 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TARGET_UPDATE, 40, "GOMP_4.0"); 1969 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TEAMS, 40, "GOMP_4.0"); 1970 #endif 1971 1972 // GOMP_4.5 versioned symbols 1973 #if OMP_45_ENABLED 1974 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TASKLOOP, 45, "GOMP_4.5"); 1975 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TASKLOOP_ULL, 45, "GOMP_4.5"); 1976 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_DOACROSS_POST, 45, "GOMP_4.5"); 1977 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_DOACROSS_WAIT, 45, "GOMP_4.5"); 1978 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_DOACROSS_STATIC_START, 45, 1979 "GOMP_4.5"); 1980 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_DOACROSS_DYNAMIC_START, 45, 1981 "GOMP_4.5"); 1982 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_DOACROSS_GUIDED_START, 45, 1983 "GOMP_4.5"); 1984 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_DOACROSS_RUNTIME_START, 45, 1985 "GOMP_4.5"); 1986 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_DOACROSS_ULL_POST, 45, "GOMP_4.5"); 1987 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_DOACROSS_ULL_WAIT, 45, "GOMP_4.5"); 1988 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_DOACROSS_STATIC_START, 45, 1989 "GOMP_4.5"); 1990 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_DOACROSS_DYNAMIC_START, 45, 1991 "GOMP_4.5"); 1992 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_DOACROSS_GUIDED_START, 45, 1993 "GOMP_4.5"); 1994 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_DOACROSS_RUNTIME_START, 45, 1995 "GOMP_4.5"); 1996 #endif 1997 1998 #endif // KMP_USE_VERSION_SYMBOLS 1999 2000 #ifdef __cplusplus 2001 } // extern "C" 2002 #endif // __cplusplus 2003