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 ompt_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_start: 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 ompt_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 ompt_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 ompt_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 ompt_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 ompt_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, #func); \ 527 KA_TRACE(20, \ 528 (#func ": T#%d, lb 0x%lx, ub 0x%lx, str 0x%lx, chunk_sz 0x%lx\n", \ 529 gtid, lb, ub, str, chunk_sz)); \ 530 \ 531 if ((str > 0) ? (lb < ub) : (lb > ub)) { \ 532 IF_OMPT_SUPPORT(OMPT_STORE_RETURN_ADDRESS(gtid);) \ 533 KMP_DISPATCH_INIT(&loc, gtid, (schedule), lb, \ 534 (str > 0) ? (ub - 1) : (ub + 1), str, chunk_sz, \ 535 (schedule) != kmp_sch_static); \ 536 IF_OMPT_SUPPORT(OMPT_STORE_RETURN_ADDRESS(gtid);) \ 537 status = KMP_DISPATCH_NEXT(&loc, gtid, NULL, (kmp_int *)p_lb, \ 538 (kmp_int *)p_ub, (kmp_int *)&stride); \ 539 if (status) { \ 540 KMP_DEBUG_ASSERT(stride == str); \ 541 *p_ub += (str > 0) ? 1 : -1; \ 542 } \ 543 } else { \ 544 status = 0; \ 545 } \ 546 \ 547 KA_TRACE(20, \ 548 (#func " exit: T#%d, *p_lb 0x%lx, *p_ub 0x%lx, returning %d\n", \ 549 gtid, *p_lb, *p_ub, status)); \ 550 return status; \ 551 } 552 553 #define LOOP_RUNTIME_START(func, schedule) \ 554 int func(long lb, long ub, long str, long *p_lb, long *p_ub) { \ 555 int status; \ 556 long stride; \ 557 long chunk_sz = 0; \ 558 int gtid = __kmp_entry_gtid(); \ 559 MKLOC(loc, #func); \ 560 KA_TRACE(20, \ 561 (#func ": T#%d, lb 0x%lx, ub 0x%lx, str 0x%lx, chunk_sz %d\n", \ 562 gtid, lb, ub, str, chunk_sz)); \ 563 \ 564 if ((str > 0) ? (lb < ub) : (lb > ub)) { \ 565 IF_OMPT_SUPPORT(OMPT_STORE_RETURN_ADDRESS(gtid);) \ 566 KMP_DISPATCH_INIT(&loc, gtid, (schedule), lb, \ 567 (str > 0) ? (ub - 1) : (ub + 1), str, chunk_sz, TRUE); \ 568 IF_OMPT_SUPPORT(OMPT_STORE_RETURN_ADDRESS(gtid);) \ 569 status = KMP_DISPATCH_NEXT(&loc, gtid, NULL, (kmp_int *)p_lb, \ 570 (kmp_int *)p_ub, (kmp_int *)&stride); \ 571 if (status) { \ 572 KMP_DEBUG_ASSERT(stride == str); \ 573 *p_ub += (str > 0) ? 1 : -1; \ 574 } \ 575 } else { \ 576 status = 0; \ 577 } \ 578 \ 579 KA_TRACE(20, \ 580 (#func " exit: T#%d, *p_lb 0x%lx, *p_ub 0x%lx, returning %d\n", \ 581 gtid, *p_lb, *p_ub, status)); \ 582 return status; \ 583 } 584 585 #define LOOP_NEXT(func, fini_code) \ 586 int func(long *p_lb, long *p_ub) { \ 587 int status; \ 588 long stride; \ 589 int gtid = __kmp_get_gtid(); \ 590 MKLOC(loc, #func); \ 591 KA_TRACE(20, (#func ": T#%d\n", gtid)); \ 592 \ 593 IF_OMPT_SUPPORT(OMPT_STORE_RETURN_ADDRESS(gtid);) \ 594 fini_code status = KMP_DISPATCH_NEXT(&loc, gtid, NULL, (kmp_int *)p_lb, \ 595 (kmp_int *)p_ub, (kmp_int *)&stride); \ 596 if (status) { \ 597 *p_ub += (stride > 0) ? 1 : -1; \ 598 } \ 599 \ 600 KA_TRACE(20, \ 601 (#func " exit: T#%d, *p_lb 0x%lx, *p_ub 0x%lx, stride 0x%lx, " \ 602 "returning %d\n", \ 603 gtid, *p_lb, *p_ub, stride, status)); \ 604 return status; \ 605 } 606 607 LOOP_START(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_STATIC_START), kmp_sch_static) 608 LOOP_NEXT(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_STATIC_NEXT), {}) 609 LOOP_START(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_DYNAMIC_START), 610 kmp_sch_dynamic_chunked) 611 LOOP_NEXT(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_DYNAMIC_NEXT), {}) 612 LOOP_START(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_GUIDED_START), 613 kmp_sch_guided_chunked) 614 LOOP_NEXT(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_GUIDED_NEXT), {}) 615 LOOP_RUNTIME_START(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_RUNTIME_START), 616 kmp_sch_runtime) 617 LOOP_NEXT(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_RUNTIME_NEXT), {}) 618 619 LOOP_START(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ORDERED_STATIC_START), 620 kmp_ord_static) 621 LOOP_NEXT(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ORDERED_STATIC_NEXT), 622 { KMP_DISPATCH_FINI_CHUNK(&loc, gtid); }) 623 LOOP_START(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ORDERED_DYNAMIC_START), 624 kmp_ord_dynamic_chunked) 625 LOOP_NEXT(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ORDERED_DYNAMIC_NEXT), 626 { KMP_DISPATCH_FINI_CHUNK(&loc, gtid); }) 627 LOOP_START(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ORDERED_GUIDED_START), 628 kmp_ord_guided_chunked) 629 LOOP_NEXT(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ORDERED_GUIDED_NEXT), 630 { KMP_DISPATCH_FINI_CHUNK(&loc, gtid); }) 631 LOOP_RUNTIME_START( 632 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ORDERED_RUNTIME_START), 633 kmp_ord_runtime) 634 LOOP_NEXT(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ORDERED_RUNTIME_NEXT), 635 { KMP_DISPATCH_FINI_CHUNK(&loc, gtid); }) 636 637 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_END)(void) { 638 int gtid = __kmp_get_gtid(); 639 KA_TRACE(20, ("GOMP_loop_end: T#%d\n", gtid)) 640 641 #if OMPT_SUPPORT && OMPT_OPTIONAL 642 ompt_frame_t *ompt_frame; 643 if (ompt_enabled.enabled) { 644 __ompt_get_task_info_internal(0, NULL, NULL, &ompt_frame, NULL, NULL); 645 ompt_frame->enter_frame = OMPT_GET_FRAME_ADDRESS(1); 646 OMPT_STORE_RETURN_ADDRESS(gtid); 647 } 648 #endif 649 __kmp_barrier(bs_plain_barrier, gtid, FALSE, 0, NULL, NULL); 650 #if OMPT_SUPPORT && OMPT_OPTIONAL 651 if (ompt_enabled.enabled) { 652 ompt_frame->enter_frame = NULL; 653 } 654 #endif 655 656 KA_TRACE(20, ("GOMP_loop_end exit: T#%d\n", gtid)) 657 } 658 659 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_END_NOWAIT)(void) { 660 KA_TRACE(20, ("GOMP_loop_end_nowait: T#%d\n", __kmp_get_gtid())) 661 } 662 663 // Unsigned long long loop worksharing constructs 664 // 665 // These are new with gcc 4.4 666 667 #define LOOP_START_ULL(func, schedule) \ 668 int func(int up, unsigned long long lb, unsigned long long ub, \ 669 unsigned long long str, unsigned long long chunk_sz, \ 670 unsigned long long *p_lb, unsigned long long *p_ub) { \ 671 int status; \ 672 long long str2 = up ? ((long long)str) : -((long long)str); \ 673 long long stride; \ 674 int gtid = __kmp_entry_gtid(); \ 675 MKLOC(loc, #func); \ 676 \ 677 KA_TRACE( \ 678 20, \ 679 (#func \ 680 ": T#%d, up %d, lb 0x%llx, ub 0x%llx, str 0x%llx, chunk_sz 0x%llx\n", \ 681 gtid, up, lb, ub, str, chunk_sz)); \ 682 \ 683 if ((str > 0) ? (lb < ub) : (lb > ub)) { \ 684 KMP_DISPATCH_INIT_ULL(&loc, gtid, (schedule), lb, \ 685 (str2 > 0) ? (ub - 1) : (ub + 1), str2, chunk_sz, \ 686 (schedule) != kmp_sch_static); \ 687 status = \ 688 KMP_DISPATCH_NEXT_ULL(&loc, gtid, NULL, (kmp_uint64 *)p_lb, \ 689 (kmp_uint64 *)p_ub, (kmp_int64 *)&stride); \ 690 if (status) { \ 691 KMP_DEBUG_ASSERT(stride == str2); \ 692 *p_ub += (str > 0) ? 1 : -1; \ 693 } \ 694 } else { \ 695 status = 0; \ 696 } \ 697 \ 698 KA_TRACE(20, \ 699 (#func " exit: T#%d, *p_lb 0x%llx, *p_ub 0x%llx, returning %d\n", \ 700 gtid, *p_lb, *p_ub, status)); \ 701 return status; \ 702 } 703 704 #define LOOP_RUNTIME_START_ULL(func, schedule) \ 705 int func(int up, unsigned long long lb, unsigned long long ub, \ 706 unsigned long long str, unsigned long long *p_lb, \ 707 unsigned long long *p_ub) { \ 708 int status; \ 709 long long str2 = up ? ((long long)str) : -((long long)str); \ 710 unsigned long long stride; \ 711 unsigned long long chunk_sz = 0; \ 712 int gtid = __kmp_entry_gtid(); \ 713 MKLOC(loc, #func); \ 714 \ 715 KA_TRACE( \ 716 20, \ 717 (#func \ 718 ": T#%d, up %d, lb 0x%llx, ub 0x%llx, str 0x%llx, chunk_sz 0x%llx\n", \ 719 gtid, up, lb, ub, str, chunk_sz)); \ 720 \ 721 if ((str > 0) ? (lb < ub) : (lb > ub)) { \ 722 KMP_DISPATCH_INIT_ULL(&loc, gtid, (schedule), lb, \ 723 (str2 > 0) ? (ub - 1) : (ub + 1), str2, chunk_sz, \ 724 TRUE); \ 725 status = \ 726 KMP_DISPATCH_NEXT_ULL(&loc, gtid, NULL, (kmp_uint64 *)p_lb, \ 727 (kmp_uint64 *)p_ub, (kmp_int64 *)&stride); \ 728 if (status) { \ 729 KMP_DEBUG_ASSERT((long long)stride == str2); \ 730 *p_ub += (str > 0) ? 1 : -1; \ 731 } \ 732 } else { \ 733 status = 0; \ 734 } \ 735 \ 736 KA_TRACE(20, \ 737 (#func " exit: T#%d, *p_lb 0x%llx, *p_ub 0x%llx, returning %d\n", \ 738 gtid, *p_lb, *p_ub, status)); \ 739 return status; \ 740 } 741 742 #define LOOP_NEXT_ULL(func, fini_code) \ 743 int func(unsigned long long *p_lb, unsigned long long *p_ub) { \ 744 int status; \ 745 long long stride; \ 746 int gtid = __kmp_get_gtid(); \ 747 MKLOC(loc, #func); \ 748 KA_TRACE(20, (#func ": T#%d\n", gtid)); \ 749 \ 750 fini_code status = \ 751 KMP_DISPATCH_NEXT_ULL(&loc, gtid, NULL, (kmp_uint64 *)p_lb, \ 752 (kmp_uint64 *)p_ub, (kmp_int64 *)&stride); \ 753 if (status) { \ 754 *p_ub += (stride > 0) ? 1 : -1; \ 755 } \ 756 \ 757 KA_TRACE(20, \ 758 (#func " exit: T#%d, *p_lb 0x%llx, *p_ub 0x%llx, stride 0x%llx, " \ 759 "returning %d\n", \ 760 gtid, *p_lb, *p_ub, stride, status)); \ 761 return status; \ 762 } 763 764 LOOP_START_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_STATIC_START), 765 kmp_sch_static) 766 LOOP_NEXT_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_STATIC_NEXT), {}) 767 LOOP_START_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_DYNAMIC_START), 768 kmp_sch_dynamic_chunked) 769 LOOP_NEXT_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_DYNAMIC_NEXT), {}) 770 LOOP_START_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_GUIDED_START), 771 kmp_sch_guided_chunked) 772 LOOP_NEXT_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_GUIDED_NEXT), {}) 773 LOOP_RUNTIME_START_ULL( 774 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_RUNTIME_START), kmp_sch_runtime) 775 LOOP_NEXT_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_RUNTIME_NEXT), {}) 776 777 LOOP_START_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_STATIC_START), 778 kmp_ord_static) 779 LOOP_NEXT_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_STATIC_NEXT), 780 { KMP_DISPATCH_FINI_CHUNK_ULL(&loc, gtid); }) 781 LOOP_START_ULL( 782 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_DYNAMIC_START), 783 kmp_ord_dynamic_chunked) 784 LOOP_NEXT_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_DYNAMIC_NEXT), 785 { KMP_DISPATCH_FINI_CHUNK_ULL(&loc, gtid); }) 786 LOOP_START_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_GUIDED_START), 787 kmp_ord_guided_chunked) 788 LOOP_NEXT_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_GUIDED_NEXT), 789 { KMP_DISPATCH_FINI_CHUNK_ULL(&loc, gtid); }) 790 LOOP_RUNTIME_START_ULL( 791 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_RUNTIME_START), 792 kmp_ord_runtime) 793 LOOP_NEXT_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_RUNTIME_NEXT), 794 { KMP_DISPATCH_FINI_CHUNK_ULL(&loc, gtid); }) 795 796 // Combined parallel / loop worksharing constructs 797 // 798 // There are no ull versions (yet). 799 800 #define PARALLEL_LOOP_START(func, schedule, ompt_pre, ompt_post) \ 801 void func(void (*task)(void *), void *data, unsigned num_threads, long lb, \ 802 long ub, long str, long chunk_sz) { \ 803 int gtid = __kmp_entry_gtid(); \ 804 MKLOC(loc, #func); \ 805 KA_TRACE(20, \ 806 (#func ": T#%d, lb 0x%lx, ub 0x%lx, str 0x%lx, chunk_sz 0x%lx\n", \ 807 gtid, lb, ub, str, chunk_sz)); \ 808 \ 809 ompt_pre(); \ 810 \ 811 if (__kmpc_ok_to_fork(&loc) && (num_threads != 1)) { \ 812 if (num_threads != 0) { \ 813 __kmp_push_num_threads(&loc, gtid, num_threads); \ 814 } \ 815 __kmp_GOMP_fork_call(&loc, gtid, task, \ 816 (microtask_t)__kmp_GOMP_parallel_microtask_wrapper, \ 817 9, task, data, num_threads, &loc, (schedule), lb, \ 818 (str > 0) ? (ub - 1) : (ub + 1), str, chunk_sz); \ 819 IF_OMPT_SUPPORT(OMPT_STORE_RETURN_ADDRESS(gtid)); \ 820 } else { \ 821 __kmp_GOMP_serialized_parallel(&loc, gtid, task); \ 822 IF_OMPT_SUPPORT(OMPT_STORE_RETURN_ADDRESS(gtid)); \ 823 } \ 824 \ 825 KMP_DISPATCH_INIT(&loc, gtid, (schedule), lb, \ 826 (str > 0) ? (ub - 1) : (ub + 1), str, chunk_sz, \ 827 (schedule) != kmp_sch_static); \ 828 \ 829 ompt_post(); \ 830 \ 831 KA_TRACE(20, (#func " exit: T#%d\n", gtid)); \ 832 } 833 834 #if OMPT_SUPPORT && OMPT_OPTIONAL 835 836 #define OMPT_LOOP_PRE() \ 837 ompt_frame_t *parent_frame; \ 838 if (ompt_enabled.enabled) { \ 839 __ompt_get_task_info_internal(0, NULL, NULL, &parent_frame, NULL, NULL); \ 840 parent_frame->enter_frame = OMPT_GET_FRAME_ADDRESS(1); \ 841 OMPT_STORE_RETURN_ADDRESS(gtid); \ 842 } 843 844 #define OMPT_LOOP_POST() \ 845 if (ompt_enabled.enabled) { \ 846 parent_frame->enter_frame = NULL; \ 847 } 848 849 #else 850 851 #define OMPT_LOOP_PRE() 852 853 #define OMPT_LOOP_POST() 854 855 #endif 856 857 PARALLEL_LOOP_START( 858 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_LOOP_STATIC_START), 859 kmp_sch_static, OMPT_LOOP_PRE, OMPT_LOOP_POST) 860 PARALLEL_LOOP_START( 861 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_LOOP_DYNAMIC_START), 862 kmp_sch_dynamic_chunked, OMPT_LOOP_PRE, OMPT_LOOP_POST) 863 PARALLEL_LOOP_START( 864 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_LOOP_GUIDED_START), 865 kmp_sch_guided_chunked, OMPT_LOOP_PRE, OMPT_LOOP_POST) 866 PARALLEL_LOOP_START( 867 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_LOOP_RUNTIME_START), 868 kmp_sch_runtime, OMPT_LOOP_PRE, OMPT_LOOP_POST) 869 870 // Tasking constructs 871 872 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASK)(void (*func)(void *), void *data, 873 void (*copy_func)(void *, void *), 874 long arg_size, long arg_align, 875 bool if_cond, unsigned gomp_flags 876 #if OMP_40_ENABLED 877 , 878 void **depend 879 #endif 880 ) { 881 MKLOC(loc, "GOMP_task"); 882 int gtid = __kmp_entry_gtid(); 883 kmp_int32 flags = 0; 884 kmp_tasking_flags_t *input_flags = (kmp_tasking_flags_t *)&flags; 885 886 KA_TRACE(20, ("GOMP_task: T#%d\n", gtid)); 887 888 // The low-order bit is the "untied" flag 889 if (!(gomp_flags & 1)) { 890 input_flags->tiedness = 1; 891 } 892 // The second low-order bit is the "final" flag 893 if (gomp_flags & 2) { 894 input_flags->final = 1; 895 } 896 input_flags->native = 1; 897 // __kmp_task_alloc() sets up all other flags 898 899 if (!if_cond) { 900 arg_size = 0; 901 } 902 903 kmp_task_t *task = __kmp_task_alloc( 904 &loc, gtid, input_flags, sizeof(kmp_task_t), 905 arg_size ? arg_size + arg_align - 1 : 0, (kmp_routine_entry_t)func); 906 907 if (arg_size > 0) { 908 if (arg_align > 0) { 909 task->shareds = (void *)((((size_t)task->shareds) + arg_align - 1) / 910 arg_align * arg_align); 911 } 912 // else error?? 913 914 if (copy_func) { 915 (*copy_func)(task->shareds, data); 916 } else { 917 KMP_MEMCPY(task->shareds, data, arg_size); 918 } 919 } 920 921 #if OMPT_SUPPORT 922 kmp_taskdata_t *current_task; 923 if (ompt_enabled.enabled) { 924 OMPT_STORE_RETURN_ADDRESS(gtid); 925 current_task = __kmp_threads[gtid]->th.th_current_task; 926 current_task->ompt_task_info.frame.enter_frame = OMPT_GET_FRAME_ADDRESS(1); 927 } 928 #endif 929 930 if (if_cond) { 931 #if OMP_40_ENABLED 932 if (gomp_flags & 8) { 933 KMP_ASSERT(depend); 934 const size_t ndeps = (kmp_intptr_t)depend[0]; 935 const size_t nout = (kmp_intptr_t)depend[1]; 936 kmp_depend_info_t dep_list[ndeps]; 937 938 for (size_t i = 0U; i < ndeps; i++) { 939 dep_list[i].base_addr = (kmp_intptr_t)depend[2U + i]; 940 dep_list[i].len = 0U; 941 dep_list[i].flags.in = 1; 942 dep_list[i].flags.out = (i < nout); 943 } 944 __kmpc_omp_task_with_deps(&loc, gtid, task, ndeps, dep_list, 0, NULL); 945 } else { 946 #endif 947 __kmpc_omp_task(&loc, gtid, task); 948 } 949 } else { 950 #if OMPT_SUPPORT 951 ompt_thread_info_t oldInfo; 952 kmp_info_t *thread; 953 kmp_taskdata_t *taskdata; 954 kmp_taskdata_t *current_task; 955 if (ompt_enabled.enabled) { 956 // Store the threads states and restore them after the task 957 thread = __kmp_threads[gtid]; 958 taskdata = KMP_TASK_TO_TASKDATA(task); 959 oldInfo = thread->th.ompt_thread_info; 960 thread->th.ompt_thread_info.wait_id = 0; 961 thread->th.ompt_thread_info.state = omp_state_work_parallel; 962 taskdata->ompt_task_info.frame.exit_frame = OMPT_GET_FRAME_ADDRESS(0); 963 OMPT_STORE_RETURN_ADDRESS(gtid); 964 } 965 #endif 966 967 __kmpc_omp_task_begin_if0(&loc, gtid, task); 968 func(data); 969 __kmpc_omp_task_complete_if0(&loc, gtid, task); 970 971 #if OMPT_SUPPORT 972 if (ompt_enabled.enabled) { 973 thread->th.ompt_thread_info = oldInfo; 974 taskdata->ompt_task_info.frame.exit_frame = NULL; 975 } 976 #endif 977 } 978 #if OMPT_SUPPORT 979 if (ompt_enabled.enabled) { 980 current_task->ompt_task_info.frame.enter_frame = NULL; 981 } 982 #endif 983 984 KA_TRACE(20, ("GOMP_task exit: T#%d\n", gtid)); 985 } 986 987 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASKWAIT)(void) { 988 MKLOC(loc, "GOMP_taskwait"); 989 int gtid = __kmp_entry_gtid(); 990 991 #if OMPT_SUPPORT 992 if (ompt_enabled.enabled) 993 OMPT_STORE_RETURN_ADDRESS(gtid); 994 #endif 995 996 KA_TRACE(20, ("GOMP_taskwait: T#%d\n", gtid)); 997 998 __kmpc_omp_taskwait(&loc, gtid); 999 1000 KA_TRACE(20, ("GOMP_taskwait exit: T#%d\n", gtid)); 1001 } 1002 1003 // Sections worksharing constructs 1004 // 1005 // For the sections construct, we initialize a dynamically scheduled loop 1006 // worksharing construct with lb 1 and stride 1, and use the iteration #'s 1007 // that its returns as sections ids. 1008 // 1009 // There are no special entry points for ordered sections, so we always use 1010 // the dynamically scheduled workshare, even if the sections aren't ordered. 1011 1012 unsigned KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SECTIONS_START)(unsigned count) { 1013 int status; 1014 kmp_int lb, ub, stride; 1015 int gtid = __kmp_entry_gtid(); 1016 MKLOC(loc, "GOMP_sections_start"); 1017 KA_TRACE(20, ("GOMP_sections_start: T#%d\n", gtid)); 1018 1019 KMP_DISPATCH_INIT(&loc, gtid, kmp_nm_dynamic_chunked, 1, count, 1, 1, TRUE); 1020 1021 status = KMP_DISPATCH_NEXT(&loc, gtid, NULL, &lb, &ub, &stride); 1022 if (status) { 1023 KMP_DEBUG_ASSERT(stride == 1); 1024 KMP_DEBUG_ASSERT(lb > 0); 1025 KMP_ASSERT(lb == ub); 1026 } else { 1027 lb = 0; 1028 } 1029 1030 KA_TRACE(20, ("GOMP_sections_start exit: T#%d returning %u\n", gtid, 1031 (unsigned)lb)); 1032 return (unsigned)lb; 1033 } 1034 1035 unsigned KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SECTIONS_NEXT)(void) { 1036 int status; 1037 kmp_int lb, ub, stride; 1038 int gtid = __kmp_get_gtid(); 1039 MKLOC(loc, "GOMP_sections_next"); 1040 KA_TRACE(20, ("GOMP_sections_next: T#%d\n", gtid)); 1041 1042 #if OMPT_SUPPORT 1043 OMPT_STORE_RETURN_ADDRESS(gtid); 1044 #endif 1045 1046 status = KMP_DISPATCH_NEXT(&loc, gtid, NULL, &lb, &ub, &stride); 1047 if (status) { 1048 KMP_DEBUG_ASSERT(stride == 1); 1049 KMP_DEBUG_ASSERT(lb > 0); 1050 KMP_ASSERT(lb == ub); 1051 } else { 1052 lb = 0; 1053 } 1054 1055 KA_TRACE( 1056 20, ("GOMP_sections_next exit: T#%d returning %u\n", gtid, (unsigned)lb)); 1057 return (unsigned)lb; 1058 } 1059 1060 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_SECTIONS_START)( 1061 void (*task)(void *), void *data, unsigned num_threads, unsigned count) { 1062 int gtid = __kmp_entry_gtid(); 1063 1064 #if OMPT_SUPPORT 1065 ompt_frame_t *parent_frame; 1066 1067 if (ompt_enabled.enabled) { 1068 __ompt_get_task_info_internal(0, NULL, NULL, &parent_frame, NULL, NULL); 1069 parent_frame->enter_frame = OMPT_GET_FRAME_ADDRESS(1); 1070 OMPT_STORE_RETURN_ADDRESS(gtid); 1071 } 1072 #endif 1073 1074 MKLOC(loc, "GOMP_parallel_sections_start"); 1075 KA_TRACE(20, ("GOMP_parallel_sections_start: T#%d\n", gtid)); 1076 1077 if (__kmpc_ok_to_fork(&loc) && (num_threads != 1)) { 1078 if (num_threads != 0) { 1079 __kmp_push_num_threads(&loc, gtid, num_threads); 1080 } 1081 __kmp_GOMP_fork_call(&loc, gtid, task, 1082 (microtask_t)__kmp_GOMP_parallel_microtask_wrapper, 9, 1083 task, data, num_threads, &loc, kmp_nm_dynamic_chunked, 1084 (kmp_int)1, (kmp_int)count, (kmp_int)1, (kmp_int)1); 1085 } else { 1086 __kmp_GOMP_serialized_parallel(&loc, gtid, task); 1087 } 1088 1089 #if OMPT_SUPPORT 1090 if (ompt_enabled.enabled) { 1091 parent_frame->enter_frame = NULL; 1092 } 1093 #endif 1094 1095 KMP_DISPATCH_INIT(&loc, gtid, kmp_nm_dynamic_chunked, 1, count, 1, 1, TRUE); 1096 1097 KA_TRACE(20, ("GOMP_parallel_sections_start exit: T#%d\n", gtid)); 1098 } 1099 1100 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SECTIONS_END)(void) { 1101 int gtid = __kmp_get_gtid(); 1102 KA_TRACE(20, ("GOMP_sections_end: T#%d\n", gtid)) 1103 1104 #if OMPT_SUPPORT 1105 ompt_frame_t *ompt_frame; 1106 if (ompt_enabled.enabled) { 1107 __ompt_get_task_info_internal(0, NULL, NULL, &ompt_frame, NULL, NULL); 1108 ompt_frame->enter_frame = OMPT_GET_FRAME_ADDRESS(1); 1109 OMPT_STORE_RETURN_ADDRESS(gtid); 1110 } 1111 #endif 1112 __kmp_barrier(bs_plain_barrier, gtid, FALSE, 0, NULL, NULL); 1113 #if OMPT_SUPPORT 1114 if (ompt_enabled.enabled) { 1115 ompt_frame->enter_frame = NULL; 1116 } 1117 #endif 1118 1119 KA_TRACE(20, ("GOMP_sections_end exit: T#%d\n", gtid)) 1120 } 1121 1122 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SECTIONS_END_NOWAIT)(void) { 1123 KA_TRACE(20, ("GOMP_sections_end_nowait: T#%d\n", __kmp_get_gtid())) 1124 } 1125 1126 // libgomp has an empty function for GOMP_taskyield as of 2013-10-10 1127 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASKYIELD)(void) { 1128 KA_TRACE(20, ("GOMP_taskyield: T#%d\n", __kmp_get_gtid())) 1129 return; 1130 } 1131 1132 #if OMP_40_ENABLED // these are new GOMP_4.0 entry points 1133 1134 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL)(void (*task)(void *), 1135 void *data, 1136 unsigned num_threads, 1137 unsigned int flags) { 1138 int gtid = __kmp_entry_gtid(); 1139 MKLOC(loc, "GOMP_parallel"); 1140 KA_TRACE(20, ("GOMP_parallel: T#%d\n", gtid)); 1141 1142 #if OMPT_SUPPORT 1143 ompt_task_info_t *parent_task_info, *task_info; 1144 if (ompt_enabled.enabled) { 1145 parent_task_info = __ompt_get_task_info_object(0); 1146 parent_task_info->frame.enter_frame = OMPT_GET_FRAME_ADDRESS(1); 1147 OMPT_STORE_RETURN_ADDRESS(gtid); 1148 } 1149 #endif 1150 if (__kmpc_ok_to_fork(&loc) && (num_threads != 1)) { 1151 if (num_threads != 0) { 1152 __kmp_push_num_threads(&loc, gtid, num_threads); 1153 } 1154 if (flags != 0) { 1155 __kmp_push_proc_bind(&loc, gtid, (kmp_proc_bind_t)flags); 1156 } 1157 __kmp_GOMP_fork_call(&loc, gtid, task, 1158 (microtask_t)__kmp_GOMP_microtask_wrapper, 2, task, 1159 data); 1160 } else { 1161 __kmp_GOMP_serialized_parallel(&loc, gtid, task); 1162 } 1163 #if OMPT_SUPPORT 1164 if (ompt_enabled.enabled) { 1165 task_info = __ompt_get_task_info_object(0); 1166 task_info->frame.exit_frame = OMPT_GET_FRAME_ADDRESS(0); 1167 } 1168 #endif 1169 task(data); 1170 #if OMPT_SUPPORT 1171 if (ompt_enabled.enabled) { 1172 OMPT_STORE_RETURN_ADDRESS(gtid); 1173 } 1174 #endif 1175 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_END)(); 1176 #if OMPT_SUPPORT 1177 if (ompt_enabled.enabled) { 1178 task_info->frame.exit_frame = NULL; 1179 parent_task_info->frame.enter_frame = NULL; 1180 } 1181 #endif 1182 } 1183 1184 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_SECTIONS)(void (*task)(void *), 1185 void *data, 1186 unsigned num_threads, 1187 unsigned count, 1188 unsigned flags) { 1189 int gtid = __kmp_entry_gtid(); 1190 MKLOC(loc, "GOMP_parallel_sections"); 1191 KA_TRACE(20, ("GOMP_parallel_sections: T#%d\n", gtid)); 1192 1193 #if OMPT_SUPPORT 1194 OMPT_STORE_RETURN_ADDRESS(gtid); 1195 #endif 1196 1197 if (__kmpc_ok_to_fork(&loc) && (num_threads != 1)) { 1198 if (num_threads != 0) { 1199 __kmp_push_num_threads(&loc, gtid, num_threads); 1200 } 1201 if (flags != 0) { 1202 __kmp_push_proc_bind(&loc, gtid, (kmp_proc_bind_t)flags); 1203 } 1204 __kmp_GOMP_fork_call(&loc, gtid, task, 1205 (microtask_t)__kmp_GOMP_parallel_microtask_wrapper, 9, 1206 task, data, num_threads, &loc, kmp_nm_dynamic_chunked, 1207 (kmp_int)1, (kmp_int)count, (kmp_int)1, (kmp_int)1); 1208 } else { 1209 __kmp_GOMP_serialized_parallel(&loc, gtid, task); 1210 } 1211 1212 #if OMPT_SUPPORT 1213 OMPT_STORE_RETURN_ADDRESS(gtid); 1214 #endif 1215 1216 KMP_DISPATCH_INIT(&loc, gtid, kmp_nm_dynamic_chunked, 1, count, 1, 1, TRUE); 1217 1218 task(data); 1219 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_END)(); 1220 KA_TRACE(20, ("GOMP_parallel_sections exit: T#%d\n", gtid)); 1221 } 1222 1223 #define PARALLEL_LOOP(func, schedule, ompt_pre, ompt_post) \ 1224 void func(void (*task)(void *), void *data, unsigned num_threads, long lb, \ 1225 long ub, long str, long chunk_sz, unsigned flags) { \ 1226 int gtid = __kmp_entry_gtid(); \ 1227 MKLOC(loc, #func); \ 1228 KA_TRACE(20, \ 1229 (#func ": T#%d, lb 0x%lx, ub 0x%lx, str 0x%lx, chunk_sz 0x%lx\n", \ 1230 gtid, lb, ub, str, chunk_sz)); \ 1231 \ 1232 ompt_pre(); \ 1233 if (__kmpc_ok_to_fork(&loc) && (num_threads != 1)) { \ 1234 if (num_threads != 0) { \ 1235 __kmp_push_num_threads(&loc, gtid, num_threads); \ 1236 } \ 1237 if (flags != 0) { \ 1238 __kmp_push_proc_bind(&loc, gtid, (kmp_proc_bind_t)flags); \ 1239 } \ 1240 __kmp_GOMP_fork_call(&loc, gtid, task, \ 1241 (microtask_t)__kmp_GOMP_parallel_microtask_wrapper, \ 1242 9, task, data, num_threads, &loc, (schedule), lb, \ 1243 (str > 0) ? (ub - 1) : (ub + 1), str, chunk_sz); \ 1244 } else { \ 1245 __kmp_GOMP_serialized_parallel(&loc, gtid, task); \ 1246 } \ 1247 \ 1248 IF_OMPT_SUPPORT(OMPT_STORE_RETURN_ADDRESS(gtid);) \ 1249 KMP_DISPATCH_INIT(&loc, gtid, (schedule), lb, \ 1250 (str > 0) ? (ub - 1) : (ub + 1), str, chunk_sz, \ 1251 (schedule) != kmp_sch_static); \ 1252 task(data); \ 1253 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_END)(); \ 1254 ompt_post(); \ 1255 \ 1256 KA_TRACE(20, (#func " exit: T#%d\n", gtid)); \ 1257 } 1258 1259 PARALLEL_LOOP(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_LOOP_STATIC), 1260 kmp_sch_static, OMPT_LOOP_PRE, OMPT_LOOP_POST) 1261 PARALLEL_LOOP(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_LOOP_DYNAMIC), 1262 kmp_sch_dynamic_chunked, OMPT_LOOP_PRE, OMPT_LOOP_POST) 1263 PARALLEL_LOOP(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_LOOP_GUIDED), 1264 kmp_sch_guided_chunked, OMPT_LOOP_PRE, OMPT_LOOP_POST) 1265 PARALLEL_LOOP(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_LOOP_RUNTIME), 1266 kmp_sch_runtime, OMPT_LOOP_PRE, OMPT_LOOP_POST) 1267 1268 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASKGROUP_START)(void) { 1269 int gtid = __kmp_entry_gtid(); 1270 MKLOC(loc, "GOMP_taskgroup_start"); 1271 KA_TRACE(20, ("GOMP_taskgroup_start: T#%d\n", gtid)); 1272 1273 #if OMPT_SUPPORT 1274 if (ompt_enabled.enabled) 1275 OMPT_STORE_RETURN_ADDRESS(gtid); 1276 #endif 1277 1278 __kmpc_taskgroup(&loc, gtid); 1279 1280 return; 1281 } 1282 1283 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASKGROUP_END)(void) { 1284 int gtid = __kmp_get_gtid(); 1285 MKLOC(loc, "GOMP_taskgroup_end"); 1286 KA_TRACE(20, ("GOMP_taskgroup_end: T#%d\n", gtid)); 1287 1288 #if OMPT_SUPPORT 1289 if (ompt_enabled.enabled) 1290 OMPT_STORE_RETURN_ADDRESS(gtid); 1291 #endif 1292 1293 __kmpc_end_taskgroup(&loc, gtid); 1294 1295 return; 1296 } 1297 1298 #ifndef KMP_DEBUG 1299 static 1300 #endif /* KMP_DEBUG */ 1301 kmp_int32 1302 __kmp_gomp_to_omp_cancellation_kind(int gomp_kind) { 1303 kmp_int32 cncl_kind = 0; 1304 switch (gomp_kind) { 1305 case 1: 1306 cncl_kind = cancel_parallel; 1307 break; 1308 case 2: 1309 cncl_kind = cancel_loop; 1310 break; 1311 case 4: 1312 cncl_kind = cancel_sections; 1313 break; 1314 case 8: 1315 cncl_kind = cancel_taskgroup; 1316 break; 1317 } 1318 return cncl_kind; 1319 } 1320 1321 bool KMP_EXPAND_NAME(KMP_API_NAME_GOMP_CANCELLATION_POINT)(int which) { 1322 if (__kmp_omp_cancellation) { 1323 KMP_FATAL(NoGompCancellation); 1324 } 1325 int gtid = __kmp_get_gtid(); 1326 MKLOC(loc, "GOMP_cancellation_point"); 1327 KA_TRACE(20, ("GOMP_cancellation_point: T#%d\n", gtid)); 1328 1329 kmp_int32 cncl_kind = __kmp_gomp_to_omp_cancellation_kind(which); 1330 1331 return __kmpc_cancellationpoint(&loc, gtid, cncl_kind); 1332 } 1333 1334 bool KMP_EXPAND_NAME(KMP_API_NAME_GOMP_BARRIER_CANCEL)(void) { 1335 if (__kmp_omp_cancellation) { 1336 KMP_FATAL(NoGompCancellation); 1337 } 1338 KMP_FATAL(NoGompCancellation); 1339 int gtid = __kmp_get_gtid(); 1340 MKLOC(loc, "GOMP_barrier_cancel"); 1341 KA_TRACE(20, ("GOMP_barrier_cancel: T#%d\n", gtid)); 1342 1343 return __kmpc_cancel_barrier(&loc, gtid); 1344 } 1345 1346 bool KMP_EXPAND_NAME(KMP_API_NAME_GOMP_CANCEL)(int which, bool do_cancel) { 1347 if (__kmp_omp_cancellation) { 1348 KMP_FATAL(NoGompCancellation); 1349 } else { 1350 return FALSE; 1351 } 1352 1353 int gtid = __kmp_get_gtid(); 1354 MKLOC(loc, "GOMP_cancel"); 1355 KA_TRACE(20, ("GOMP_cancel: T#%d\n", gtid)); 1356 1357 kmp_int32 cncl_kind = __kmp_gomp_to_omp_cancellation_kind(which); 1358 1359 if (do_cancel == FALSE) { 1360 return KMP_EXPAND_NAME(KMP_API_NAME_GOMP_CANCELLATION_POINT)(which); 1361 } else { 1362 return __kmpc_cancel(&loc, gtid, cncl_kind); 1363 } 1364 } 1365 1366 bool KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SECTIONS_END_CANCEL)(void) { 1367 if (__kmp_omp_cancellation) { 1368 KMP_FATAL(NoGompCancellation); 1369 } 1370 int gtid = __kmp_get_gtid(); 1371 MKLOC(loc, "GOMP_sections_end_cancel"); 1372 KA_TRACE(20, ("GOMP_sections_end_cancel: T#%d\n", gtid)); 1373 1374 return __kmpc_cancel_barrier(&loc, gtid); 1375 } 1376 1377 bool KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_END_CANCEL)(void) { 1378 if (__kmp_omp_cancellation) { 1379 KMP_FATAL(NoGompCancellation); 1380 } 1381 int gtid = __kmp_get_gtid(); 1382 MKLOC(loc, "GOMP_loop_end_cancel"); 1383 KA_TRACE(20, ("GOMP_loop_end_cancel: T#%d\n", gtid)); 1384 1385 return __kmpc_cancel_barrier(&loc, gtid); 1386 } 1387 1388 // All target functions are empty as of 2014-05-29 1389 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TARGET)(int device, void (*fn)(void *), 1390 const void *openmp_target, 1391 size_t mapnum, void **hostaddrs, 1392 size_t *sizes, 1393 unsigned char *kinds) { 1394 return; 1395 } 1396 1397 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TARGET_DATA)( 1398 int device, const void *openmp_target, size_t mapnum, void **hostaddrs, 1399 size_t *sizes, unsigned char *kinds) { 1400 return; 1401 } 1402 1403 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TARGET_END_DATA)(void) { return; } 1404 1405 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TARGET_UPDATE)( 1406 int device, const void *openmp_target, size_t mapnum, void **hostaddrs, 1407 size_t *sizes, unsigned char *kinds) { 1408 return; 1409 } 1410 1411 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TEAMS)(unsigned int num_teams, 1412 unsigned int thread_limit) { 1413 return; 1414 } 1415 #endif // OMP_40_ENABLED 1416 1417 #if OMP_45_ENABLED 1418 1419 // Task duplication function which copies src to dest (both are 1420 // preallocated task structures) 1421 static void __kmp_gomp_task_dup(kmp_task_t *dest, kmp_task_t *src, 1422 kmp_int32 last_private) { 1423 kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(src); 1424 if (taskdata->td_copy_func) { 1425 (taskdata->td_copy_func)(dest->shareds, src->shareds); 1426 } 1427 } 1428 1429 #ifdef __cplusplus 1430 } // extern "C" 1431 #endif 1432 1433 template <typename T> 1434 void __GOMP_taskloop(void (*func)(void *), void *data, 1435 void (*copy_func)(void *, void *), long arg_size, 1436 long arg_align, unsigned gomp_flags, 1437 unsigned long num_tasks, int priority, T start, T end, 1438 T step) { 1439 typedef void (*p_task_dup_t)(kmp_task_t *, kmp_task_t *, kmp_int32); 1440 MKLOC(loc, "GOMP_taskloop"); 1441 int sched; 1442 T *loop_bounds; 1443 int gtid = __kmp_entry_gtid(); 1444 kmp_int32 flags = 0; 1445 int if_val = gomp_flags & (1u << 10); 1446 int nogroup = gomp_flags & (1u << 11); 1447 int up = gomp_flags & (1u << 8); 1448 p_task_dup_t task_dup = NULL; 1449 kmp_tasking_flags_t *input_flags = (kmp_tasking_flags_t *)&flags; 1450 #ifdef KMP_DEBUG 1451 { 1452 char *buff; 1453 buff = __kmp_str_format( 1454 "GOMP_taskloop: T#%%d: func:%%p data:%%p copy_func:%%p " 1455 "arg_size:%%ld arg_align:%%ld gomp_flags:0x%%x num_tasks:%%lu " 1456 "priority:%%d start:%%%s end:%%%s step:%%%s\n", 1457 traits_t<T>::spec, traits_t<T>::spec, traits_t<T>::spec); 1458 KA_TRACE(20, (buff, gtid, func, data, copy_func, arg_size, arg_align, 1459 gomp_flags, num_tasks, priority, start, end, step)); 1460 __kmp_str_free(&buff); 1461 } 1462 #endif 1463 KMP_ASSERT((size_t)arg_size >= 2 * sizeof(T)); 1464 KMP_ASSERT(arg_align > 0); 1465 // The low-order bit is the "untied" flag 1466 if (!(gomp_flags & 1)) { 1467 input_flags->tiedness = 1; 1468 } 1469 // The second low-order bit is the "final" flag 1470 if (gomp_flags & 2) { 1471 input_flags->final = 1; 1472 } 1473 // Negative step flag 1474 if (!up) { 1475 // If step is flagged as negative, but isn't properly sign extended 1476 // Then manually sign extend it. Could be a short, int, char embedded 1477 // in a long. So cannot assume any cast. 1478 if (step > 0) { 1479 for (int i = sizeof(T) * CHAR_BIT - 1; i >= 0L; --i) { 1480 // break at the first 1 bit 1481 if (step & ((T)1 << i)) 1482 break; 1483 step |= ((T)1 << i); 1484 } 1485 } 1486 } 1487 input_flags->native = 1; 1488 // Figure out if none/grainsize/num_tasks clause specified 1489 if (num_tasks > 0) { 1490 if (gomp_flags & (1u << 9)) 1491 sched = 1; // grainsize specified 1492 else 1493 sched = 2; // num_tasks specified 1494 // neither grainsize nor num_tasks specified 1495 } else { 1496 sched = 0; 1497 } 1498 1499 // __kmp_task_alloc() sets up all other flags 1500 kmp_task_t *task = 1501 __kmp_task_alloc(&loc, gtid, input_flags, sizeof(kmp_task_t), 1502 arg_size + arg_align - 1, (kmp_routine_entry_t)func); 1503 kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task); 1504 taskdata->td_copy_func = copy_func; 1505 taskdata->td_size_loop_bounds = sizeof(T); 1506 1507 // re-align shareds if needed and setup firstprivate copy constructors 1508 // through the task_dup mechanism 1509 task->shareds = (void *)((((size_t)task->shareds) + arg_align - 1) / 1510 arg_align * arg_align); 1511 if (copy_func) { 1512 task_dup = __kmp_gomp_task_dup; 1513 } 1514 KMP_MEMCPY(task->shareds, data, arg_size); 1515 1516 loop_bounds = (T *)task->shareds; 1517 loop_bounds[0] = start; 1518 loop_bounds[1] = end + (up ? -1 : 1); 1519 __kmpc_taskloop(&loc, gtid, task, if_val, (kmp_uint64 *)&(loop_bounds[0]), 1520 (kmp_uint64 *)&(loop_bounds[1]), (kmp_int64)step, nogroup, 1521 sched, (kmp_uint64)num_tasks, (void *)task_dup); 1522 } 1523 1524 #ifdef __cplusplus 1525 extern "C" { 1526 #endif // __cplusplus 1527 1528 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASKLOOP)( 1529 void (*func)(void *), void *data, void (*copy_func)(void *, void *), 1530 long arg_size, long arg_align, unsigned gomp_flags, unsigned long num_tasks, 1531 int priority, long start, long end, long step) { 1532 __GOMP_taskloop<long>(func, data, copy_func, arg_size, arg_align, gomp_flags, 1533 num_tasks, priority, start, end, step); 1534 } 1535 1536 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASKLOOP_ULL)( 1537 void (*func)(void *), void *data, void (*copy_func)(void *, void *), 1538 long arg_size, long arg_align, unsigned gomp_flags, unsigned long num_tasks, 1539 int priority, unsigned long long start, unsigned long long end, 1540 unsigned long long step) { 1541 __GOMP_taskloop<unsigned long long>(func, data, copy_func, arg_size, 1542 arg_align, gomp_flags, num_tasks, 1543 priority, start, end, step); 1544 } 1545 1546 #endif 1547 1548 /* The following sections of code create aliases for the GOMP_* functions, then 1549 create versioned symbols using the assembler directive .symver. This is only 1550 pertinent for ELF .so library. The KMP_VERSION_SYMBOL macro is defined in 1551 kmp_os.h */ 1552 1553 #ifdef KMP_USE_VERSION_SYMBOLS 1554 // GOMP_1.0 versioned symbols 1555 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_ATOMIC_END, 10, "GOMP_1.0"); 1556 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_ATOMIC_START, 10, "GOMP_1.0"); 1557 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_BARRIER, 10, "GOMP_1.0"); 1558 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_CRITICAL_END, 10, "GOMP_1.0"); 1559 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_CRITICAL_NAME_END, 10, "GOMP_1.0"); 1560 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_CRITICAL_NAME_START, 10, "GOMP_1.0"); 1561 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_CRITICAL_START, 10, "GOMP_1.0"); 1562 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_DYNAMIC_NEXT, 10, "GOMP_1.0"); 1563 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_DYNAMIC_START, 10, "GOMP_1.0"); 1564 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_END, 10, "GOMP_1.0"); 1565 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_END_NOWAIT, 10, "GOMP_1.0"); 1566 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_GUIDED_NEXT, 10, "GOMP_1.0"); 1567 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_GUIDED_START, 10, "GOMP_1.0"); 1568 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ORDERED_DYNAMIC_NEXT, 10, "GOMP_1.0"); 1569 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ORDERED_DYNAMIC_START, 10, 1570 "GOMP_1.0"); 1571 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ORDERED_GUIDED_NEXT, 10, "GOMP_1.0"); 1572 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ORDERED_GUIDED_START, 10, "GOMP_1.0"); 1573 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ORDERED_RUNTIME_NEXT, 10, "GOMP_1.0"); 1574 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ORDERED_RUNTIME_START, 10, 1575 "GOMP_1.0"); 1576 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ORDERED_STATIC_NEXT, 10, "GOMP_1.0"); 1577 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ORDERED_STATIC_START, 10, "GOMP_1.0"); 1578 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_RUNTIME_NEXT, 10, "GOMP_1.0"); 1579 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_RUNTIME_START, 10, "GOMP_1.0"); 1580 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_STATIC_NEXT, 10, "GOMP_1.0"); 1581 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_STATIC_START, 10, "GOMP_1.0"); 1582 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_ORDERED_END, 10, "GOMP_1.0"); 1583 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_ORDERED_START, 10, "GOMP_1.0"); 1584 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_END, 10, "GOMP_1.0"); 1585 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_LOOP_DYNAMIC_START, 10, 1586 "GOMP_1.0"); 1587 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_LOOP_GUIDED_START, 10, 1588 "GOMP_1.0"); 1589 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_LOOP_RUNTIME_START, 10, 1590 "GOMP_1.0"); 1591 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_LOOP_STATIC_START, 10, 1592 "GOMP_1.0"); 1593 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_SECTIONS_START, 10, "GOMP_1.0"); 1594 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_START, 10, "GOMP_1.0"); 1595 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_SECTIONS_END, 10, "GOMP_1.0"); 1596 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_SECTIONS_END_NOWAIT, 10, "GOMP_1.0"); 1597 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_SECTIONS_NEXT, 10, "GOMP_1.0"); 1598 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_SECTIONS_START, 10, "GOMP_1.0"); 1599 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_SINGLE_COPY_END, 10, "GOMP_1.0"); 1600 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_SINGLE_COPY_START, 10, "GOMP_1.0"); 1601 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_SINGLE_START, 10, "GOMP_1.0"); 1602 1603 // GOMP_2.0 versioned symbols 1604 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TASK, 20, "GOMP_2.0"); 1605 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TASKWAIT, 20, "GOMP_2.0"); 1606 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_DYNAMIC_NEXT, 20, "GOMP_2.0"); 1607 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_DYNAMIC_START, 20, "GOMP_2.0"); 1608 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_GUIDED_NEXT, 20, "GOMP_2.0"); 1609 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_GUIDED_START, 20, "GOMP_2.0"); 1610 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_DYNAMIC_NEXT, 20, 1611 "GOMP_2.0"); 1612 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_DYNAMIC_START, 20, 1613 "GOMP_2.0"); 1614 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_GUIDED_NEXT, 20, 1615 "GOMP_2.0"); 1616 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_GUIDED_START, 20, 1617 "GOMP_2.0"); 1618 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_RUNTIME_NEXT, 20, 1619 "GOMP_2.0"); 1620 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_RUNTIME_START, 20, 1621 "GOMP_2.0"); 1622 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_STATIC_NEXT, 20, 1623 "GOMP_2.0"); 1624 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_STATIC_START, 20, 1625 "GOMP_2.0"); 1626 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_RUNTIME_NEXT, 20, "GOMP_2.0"); 1627 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_RUNTIME_START, 20, "GOMP_2.0"); 1628 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_STATIC_NEXT, 20, "GOMP_2.0"); 1629 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_STATIC_START, 20, "GOMP_2.0"); 1630 1631 // GOMP_3.0 versioned symbols 1632 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TASKYIELD, 30, "GOMP_3.0"); 1633 1634 // GOMP_4.0 versioned symbols 1635 #if OMP_40_ENABLED 1636 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL, 40, "GOMP_4.0"); 1637 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_SECTIONS, 40, "GOMP_4.0"); 1638 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_LOOP_DYNAMIC, 40, "GOMP_4.0"); 1639 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_LOOP_GUIDED, 40, "GOMP_4.0"); 1640 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_LOOP_RUNTIME, 40, "GOMP_4.0"); 1641 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_LOOP_STATIC, 40, "GOMP_4.0"); 1642 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TASKGROUP_START, 40, "GOMP_4.0"); 1643 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TASKGROUP_END, 40, "GOMP_4.0"); 1644 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_BARRIER_CANCEL, 40, "GOMP_4.0"); 1645 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_CANCEL, 40, "GOMP_4.0"); 1646 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_CANCELLATION_POINT, 40, "GOMP_4.0"); 1647 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_END_CANCEL, 40, "GOMP_4.0"); 1648 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_SECTIONS_END_CANCEL, 40, "GOMP_4.0"); 1649 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TARGET, 40, "GOMP_4.0"); 1650 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TARGET_DATA, 40, "GOMP_4.0"); 1651 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TARGET_END_DATA, 40, "GOMP_4.0"); 1652 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TARGET_UPDATE, 40, "GOMP_4.0"); 1653 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TEAMS, 40, "GOMP_4.0"); 1654 #endif 1655 1656 #if OMP_45_ENABLED 1657 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TASKLOOP, 45, "GOMP_4.5"); 1658 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TASKLOOP_ULL, 45, "GOMP_4.5"); 1659 #endif 1660 1661 #endif // KMP_USE_VERSION_SYMBOLS 1662 1663 #ifdef __cplusplus 1664 } // extern "C" 1665 #endif // __cplusplus 1666