1 /* 2 * kmp_gsupport.cpp 3 */ 4 5 //===----------------------------------------------------------------------===// 6 // 7 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 8 // See https://llvm.org/LICENSE.txt for license information. 9 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 10 // 11 //===----------------------------------------------------------------------===// 12 13 #include "kmp.h" 14 #include "kmp_atomic.h" 15 16 #if OMPT_SUPPORT 17 #include "ompt-specific.h" 18 #endif 19 20 enum { 21 KMP_GOMP_TASK_UNTIED_FLAG = 1, 22 KMP_GOMP_TASK_FINAL_FLAG = 2, 23 KMP_GOMP_TASK_DEPENDS_FLAG = 8 24 }; 25 26 enum { 27 KMP_GOMP_DEPOBJ_IN = 1, 28 KMP_GOMP_DEPOBJ_OUT = 2, 29 KMP_GOMP_DEPOBJ_INOUT = 3, 30 KMP_GOMP_DEPOBJ_MTXINOUTSET = 4 31 }; 32 33 // This class helps convert gomp dependency info into 34 // kmp_depend_info_t structures 35 class kmp_gomp_depends_info_t { 36 void **depend; 37 kmp_int32 num_deps; 38 size_t num_out, num_mutexinout, num_in, num_depobj; 39 size_t offset; 40 41 public: 42 kmp_gomp_depends_info_t(void **depend) : depend(depend) { 43 size_t ndeps = (kmp_intptr_t)depend[0]; 44 // GOMP taskdep structure: 45 // if depend[0] != 0: 46 // depend = [ ndeps | nout | &out | ... | &out | &in | ... | &in ] 47 // 48 // if depend[0] == 0: 49 // depend = [ 0 | ndeps | nout | nmtx | nin | &out | ... | &out | &mtx | 50 // ... | &mtx | &in | ... | &in | &depobj | ... | &depobj ] 51 if (ndeps) { 52 num_out = (kmp_intptr_t)depend[1]; 53 num_in = ndeps - num_out; 54 num_mutexinout = num_depobj = 0; 55 offset = 2; 56 } else { 57 ndeps = (kmp_intptr_t)depend[1]; 58 num_out = (kmp_intptr_t)depend[2]; 59 num_mutexinout = (kmp_intptr_t)depend[3]; 60 num_in = (kmp_intptr_t)depend[4]; 61 num_depobj = ndeps - num_out - num_mutexinout - num_in; 62 KMP_ASSERT(num_depobj <= ndeps); 63 offset = 5; 64 } 65 num_deps = static_cast<kmp_int32>(ndeps); 66 } 67 kmp_int32 get_num_deps() const { return num_deps; } 68 kmp_depend_info_t get_kmp_depend(size_t index) const { 69 kmp_depend_info_t retval; 70 memset(&retval, '\0', sizeof(retval)); 71 KMP_ASSERT(index < (size_t)num_deps); 72 retval.len = 0; 73 // Because inout and out are logically equivalent, 74 // use inout and in dependency flags. GOMP does not provide a 75 // way to distinguish if user specified out vs. inout. 76 if (index < num_out) { 77 retval.flags.in = 1; 78 retval.flags.out = 1; 79 retval.base_addr = (kmp_intptr_t)depend[offset + index]; 80 } else if (index >= num_out && index < (num_out + num_mutexinout)) { 81 retval.flags.mtx = 1; 82 retval.base_addr = (kmp_intptr_t)depend[offset + index]; 83 } else if (index >= (num_out + num_mutexinout) && 84 index < (num_out + num_mutexinout + num_in)) { 85 retval.flags.in = 1; 86 retval.base_addr = (kmp_intptr_t)depend[offset + index]; 87 } else { 88 // depobj is a two element array (size of elements are size of pointer) 89 // depobj[0] = base_addr 90 // depobj[1] = type (in, out, inout, mutexinoutset, etc.) 91 kmp_intptr_t *depobj = (kmp_intptr_t *)depend[offset + index]; 92 retval.base_addr = depobj[0]; 93 switch (depobj[1]) { 94 case KMP_GOMP_DEPOBJ_IN: 95 retval.flags.in = 1; 96 break; 97 case KMP_GOMP_DEPOBJ_OUT: 98 retval.flags.out = 1; 99 break; 100 case KMP_GOMP_DEPOBJ_INOUT: 101 retval.flags.in = 1; 102 retval.flags.out = 1; 103 break; 104 case KMP_GOMP_DEPOBJ_MTXINOUTSET: 105 retval.flags.mtx = 1; 106 break; 107 default: 108 KMP_FATAL(GompFeatureNotSupported, "Unknown depobj type"); 109 } 110 } 111 return retval; 112 } 113 }; 114 115 #ifdef __cplusplus 116 extern "C" { 117 #endif // __cplusplus 118 119 #define MKLOC(loc, routine) \ 120 static ident_t loc = {0, KMP_IDENT_KMPC, 0, 0, ";unknown;unknown;0;0;;"}; 121 122 #include "kmp_ftn_os.h" 123 124 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_BARRIER)(void) { 125 int gtid = __kmp_entry_gtid(); 126 MKLOC(loc, "GOMP_barrier"); 127 KA_TRACE(20, ("GOMP_barrier: T#%d\n", gtid)); 128 #if OMPT_SUPPORT && OMPT_OPTIONAL 129 ompt_frame_t *ompt_frame; 130 if (ompt_enabled.enabled) { 131 __ompt_get_task_info_internal(0, NULL, NULL, &ompt_frame, NULL, NULL); 132 ompt_frame->enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0); 133 } 134 OMPT_STORE_RETURN_ADDRESS(gtid); 135 #endif 136 __kmpc_barrier(&loc, gtid); 137 #if OMPT_SUPPORT && OMPT_OPTIONAL 138 if (ompt_enabled.enabled) { 139 ompt_frame->enter_frame = ompt_data_none; 140 } 141 #endif 142 } 143 144 // Mutual exclusion 145 146 // The symbol that icc/ifort generates for unnamed for unnamed critical sections 147 // - .gomp_critical_user_ - is defined using .comm in any objects reference it. 148 // We can't reference it directly here in C code, as the symbol contains a ".". 149 // 150 // The RTL contains an assembly language definition of .gomp_critical_user_ 151 // with another symbol __kmp_unnamed_critical_addr initialized with it's 152 // address. 153 extern kmp_critical_name *__kmp_unnamed_critical_addr; 154 155 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_CRITICAL_START)(void) { 156 int gtid = __kmp_entry_gtid(); 157 MKLOC(loc, "GOMP_critical_start"); 158 KA_TRACE(20, ("GOMP_critical_start: T#%d\n", gtid)); 159 #if OMPT_SUPPORT && OMPT_OPTIONAL 160 OMPT_STORE_RETURN_ADDRESS(gtid); 161 #endif 162 __kmpc_critical(&loc, gtid, __kmp_unnamed_critical_addr); 163 } 164 165 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_CRITICAL_END)(void) { 166 int gtid = __kmp_get_gtid(); 167 MKLOC(loc, "GOMP_critical_end"); 168 KA_TRACE(20, ("GOMP_critical_end: T#%d\n", gtid)); 169 #if OMPT_SUPPORT && OMPT_OPTIONAL 170 OMPT_STORE_RETURN_ADDRESS(gtid); 171 #endif 172 __kmpc_end_critical(&loc, gtid, __kmp_unnamed_critical_addr); 173 } 174 175 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_CRITICAL_NAME_START)(void **pptr) { 176 int gtid = __kmp_entry_gtid(); 177 MKLOC(loc, "GOMP_critical_name_start"); 178 KA_TRACE(20, ("GOMP_critical_name_start: T#%d\n", gtid)); 179 __kmpc_critical(&loc, gtid, (kmp_critical_name *)pptr); 180 } 181 182 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_CRITICAL_NAME_END)(void **pptr) { 183 int gtid = __kmp_get_gtid(); 184 MKLOC(loc, "GOMP_critical_name_end"); 185 KA_TRACE(20, ("GOMP_critical_name_end: T#%d\n", gtid)); 186 __kmpc_end_critical(&loc, gtid, (kmp_critical_name *)pptr); 187 } 188 189 // The Gnu codegen tries to use locked operations to perform atomic updates 190 // inline. If it can't, then it calls GOMP_atomic_start() before performing 191 // the update and GOMP_atomic_end() afterward, regardless of the data type. 192 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_ATOMIC_START)(void) { 193 int gtid = __kmp_entry_gtid(); 194 KA_TRACE(20, ("GOMP_atomic_start: T#%d\n", gtid)); 195 196 #if OMPT_SUPPORT 197 __ompt_thread_assign_wait_id(0); 198 #endif 199 200 __kmp_acquire_atomic_lock(&__kmp_atomic_lock, gtid); 201 } 202 203 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_ATOMIC_END)(void) { 204 int gtid = __kmp_get_gtid(); 205 KA_TRACE(20, ("GOMP_atomic_end: T#%d\n", gtid)); 206 __kmp_release_atomic_lock(&__kmp_atomic_lock, gtid); 207 } 208 209 int KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SINGLE_START)(void) { 210 int gtid = __kmp_entry_gtid(); 211 MKLOC(loc, "GOMP_single_start"); 212 KA_TRACE(20, ("GOMP_single_start: T#%d\n", gtid)); 213 214 if (!TCR_4(__kmp_init_parallel)) 215 __kmp_parallel_initialize(); 216 __kmp_resume_if_soft_paused(); 217 218 // 3rd parameter == FALSE prevents kmp_enter_single from pushing a 219 // workshare when USE_CHECKS is defined. We need to avoid the push, 220 // as there is no corresponding GOMP_single_end() call. 221 kmp_int32 rc = __kmp_enter_single(gtid, &loc, FALSE); 222 223 #if OMPT_SUPPORT && OMPT_OPTIONAL 224 kmp_info_t *this_thr = __kmp_threads[gtid]; 225 kmp_team_t *team = this_thr->th.th_team; 226 int tid = __kmp_tid_from_gtid(gtid); 227 228 if (ompt_enabled.enabled) { 229 if (rc) { 230 if (ompt_enabled.ompt_callback_work) { 231 ompt_callbacks.ompt_callback(ompt_callback_work)( 232 ompt_work_single_executor, ompt_scope_begin, 233 &(team->t.ompt_team_info.parallel_data), 234 &(team->t.t_implicit_task_taskdata[tid].ompt_task_info.task_data), 235 1, OMPT_GET_RETURN_ADDRESS(0)); 236 } 237 } else { 238 if (ompt_enabled.ompt_callback_work) { 239 ompt_callbacks.ompt_callback(ompt_callback_work)( 240 ompt_work_single_other, ompt_scope_begin, 241 &(team->t.ompt_team_info.parallel_data), 242 &(team->t.t_implicit_task_taskdata[tid].ompt_task_info.task_data), 243 1, OMPT_GET_RETURN_ADDRESS(0)); 244 ompt_callbacks.ompt_callback(ompt_callback_work)( 245 ompt_work_single_other, ompt_scope_end, 246 &(team->t.ompt_team_info.parallel_data), 247 &(team->t.t_implicit_task_taskdata[tid].ompt_task_info.task_data), 248 1, OMPT_GET_RETURN_ADDRESS(0)); 249 } 250 } 251 } 252 #endif 253 254 return rc; 255 } 256 257 void *KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SINGLE_COPY_START)(void) { 258 void *retval; 259 int gtid = __kmp_entry_gtid(); 260 MKLOC(loc, "GOMP_single_copy_start"); 261 KA_TRACE(20, ("GOMP_single_copy_start: T#%d\n", gtid)); 262 263 if (!TCR_4(__kmp_init_parallel)) 264 __kmp_parallel_initialize(); 265 __kmp_resume_if_soft_paused(); 266 267 // If this is the first thread to enter, return NULL. The generated code will 268 // then call GOMP_single_copy_end() for this thread only, with the 269 // copyprivate data pointer as an argument. 270 if (__kmp_enter_single(gtid, &loc, FALSE)) 271 return NULL; 272 273 // Wait for the first thread to set the copyprivate data pointer, 274 // and for all other threads to reach this point. 275 276 #if OMPT_SUPPORT && OMPT_OPTIONAL 277 ompt_frame_t *ompt_frame; 278 if (ompt_enabled.enabled) { 279 __ompt_get_task_info_internal(0, NULL, NULL, &ompt_frame, NULL, NULL); 280 ompt_frame->enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0); 281 } 282 OMPT_STORE_RETURN_ADDRESS(gtid); 283 #endif 284 __kmp_barrier(bs_plain_barrier, gtid, FALSE, 0, NULL, NULL); 285 286 // Retrieve the value of the copyprivate data point, and wait for all 287 // threads to do likewise, then return. 288 retval = __kmp_team_from_gtid(gtid)->t.t_copypriv_data; 289 { 290 #if OMPT_SUPPORT && OMPT_OPTIONAL 291 OMPT_STORE_RETURN_ADDRESS(gtid); 292 #endif 293 __kmp_barrier(bs_plain_barrier, gtid, FALSE, 0, NULL, NULL); 294 } 295 #if OMPT_SUPPORT && OMPT_OPTIONAL 296 if (ompt_enabled.enabled) { 297 ompt_frame->enter_frame = ompt_data_none; 298 } 299 #endif 300 return retval; 301 } 302 303 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SINGLE_COPY_END)(void *data) { 304 int gtid = __kmp_get_gtid(); 305 KA_TRACE(20, ("GOMP_single_copy_end: T#%d\n", gtid)); 306 307 // Set the copyprivate data pointer fo the team, then hit the barrier so that 308 // the other threads will continue on and read it. Hit another barrier before 309 // continuing, so that the know that the copyprivate data pointer has been 310 // propagated to all threads before trying to reuse the t_copypriv_data field. 311 __kmp_team_from_gtid(gtid)->t.t_copypriv_data = data; 312 #if OMPT_SUPPORT && OMPT_OPTIONAL 313 ompt_frame_t *ompt_frame; 314 if (ompt_enabled.enabled) { 315 __ompt_get_task_info_internal(0, NULL, NULL, &ompt_frame, NULL, NULL); 316 ompt_frame->enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0); 317 } 318 OMPT_STORE_RETURN_ADDRESS(gtid); 319 #endif 320 __kmp_barrier(bs_plain_barrier, gtid, FALSE, 0, NULL, NULL); 321 { 322 #if OMPT_SUPPORT && OMPT_OPTIONAL 323 OMPT_STORE_RETURN_ADDRESS(gtid); 324 #endif 325 __kmp_barrier(bs_plain_barrier, gtid, FALSE, 0, NULL, NULL); 326 } 327 #if OMPT_SUPPORT && OMPT_OPTIONAL 328 if (ompt_enabled.enabled) { 329 ompt_frame->enter_frame = ompt_data_none; 330 } 331 #endif 332 } 333 334 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_ORDERED_START)(void) { 335 int gtid = __kmp_entry_gtid(); 336 MKLOC(loc, "GOMP_ordered_start"); 337 KA_TRACE(20, ("GOMP_ordered_start: T#%d\n", gtid)); 338 #if OMPT_SUPPORT && OMPT_OPTIONAL 339 OMPT_STORE_RETURN_ADDRESS(gtid); 340 #endif 341 __kmpc_ordered(&loc, gtid); 342 } 343 344 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_ORDERED_END)(void) { 345 int gtid = __kmp_get_gtid(); 346 MKLOC(loc, "GOMP_ordered_end"); 347 KA_TRACE(20, ("GOMP_ordered_start: T#%d\n", gtid)); 348 #if OMPT_SUPPORT && OMPT_OPTIONAL 349 OMPT_STORE_RETURN_ADDRESS(gtid); 350 #endif 351 __kmpc_end_ordered(&loc, gtid); 352 } 353 354 // Dispatch macro defs 355 // 356 // They come in two flavors: 64-bit unsigned, and either 32-bit signed 357 // (IA-32 architecture) or 64-bit signed (Intel(R) 64). 358 359 #if KMP_ARCH_X86 || KMP_ARCH_ARM || KMP_ARCH_MIPS 360 #define KMP_DISPATCH_INIT __kmp_aux_dispatch_init_4 361 #define KMP_DISPATCH_FINI_CHUNK __kmp_aux_dispatch_fini_chunk_4 362 #define KMP_DISPATCH_NEXT __kmpc_dispatch_next_4 363 #else 364 #define KMP_DISPATCH_INIT __kmp_aux_dispatch_init_8 365 #define KMP_DISPATCH_FINI_CHUNK __kmp_aux_dispatch_fini_chunk_8 366 #define KMP_DISPATCH_NEXT __kmpc_dispatch_next_8 367 #endif /* KMP_ARCH_X86 */ 368 369 #define KMP_DISPATCH_INIT_ULL __kmp_aux_dispatch_init_8u 370 #define KMP_DISPATCH_FINI_CHUNK_ULL __kmp_aux_dispatch_fini_chunk_8u 371 #define KMP_DISPATCH_NEXT_ULL __kmpc_dispatch_next_8u 372 373 // The parallel construct 374 375 #ifndef KMP_DEBUG 376 static 377 #endif /* KMP_DEBUG */ 378 void 379 __kmp_GOMP_microtask_wrapper(int *gtid, int *npr, void (*task)(void *), 380 void *data) { 381 #if OMPT_SUPPORT 382 kmp_info_t *thr; 383 ompt_frame_t *ompt_frame; 384 ompt_state_t enclosing_state; 385 386 if (ompt_enabled.enabled) { 387 // get pointer to thread data structure 388 thr = __kmp_threads[*gtid]; 389 390 // save enclosing task state; set current state for task 391 enclosing_state = thr->th.ompt_thread_info.state; 392 thr->th.ompt_thread_info.state = ompt_state_work_parallel; 393 394 // set task frame 395 __ompt_get_task_info_internal(0, NULL, NULL, &ompt_frame, NULL, NULL); 396 ompt_frame->exit_frame.ptr = OMPT_GET_FRAME_ADDRESS(0); 397 } 398 #endif 399 400 task(data); 401 402 #if OMPT_SUPPORT 403 if (ompt_enabled.enabled) { 404 // clear task frame 405 ompt_frame->exit_frame = ompt_data_none; 406 407 // restore enclosing state 408 thr->th.ompt_thread_info.state = enclosing_state; 409 } 410 #endif 411 } 412 413 #ifndef KMP_DEBUG 414 static 415 #endif /* KMP_DEBUG */ 416 void 417 __kmp_GOMP_parallel_microtask_wrapper(int *gtid, int *npr, 418 void (*task)(void *), void *data, 419 unsigned num_threads, ident_t *loc, 420 enum sched_type schedule, long start, 421 long end, long incr, 422 long chunk_size) { 423 // Initialize the loop worksharing construct. 424 425 KMP_DISPATCH_INIT(loc, *gtid, schedule, start, end, incr, chunk_size, 426 schedule != kmp_sch_static); 427 428 #if OMPT_SUPPORT 429 kmp_info_t *thr; 430 ompt_frame_t *ompt_frame; 431 ompt_state_t enclosing_state; 432 433 if (ompt_enabled.enabled) { 434 thr = __kmp_threads[*gtid]; 435 // save enclosing task state; set current state for task 436 enclosing_state = thr->th.ompt_thread_info.state; 437 thr->th.ompt_thread_info.state = ompt_state_work_parallel; 438 439 // set task frame 440 __ompt_get_task_info_internal(0, NULL, NULL, &ompt_frame, NULL, NULL); 441 ompt_frame->exit_frame.ptr = OMPT_GET_FRAME_ADDRESS(0); 442 } 443 #endif 444 445 // Now invoke the microtask. 446 task(data); 447 448 #if OMPT_SUPPORT 449 if (ompt_enabled.enabled) { 450 // clear task frame 451 ompt_frame->exit_frame = ompt_data_none; 452 453 // reset enclosing state 454 thr->th.ompt_thread_info.state = enclosing_state; 455 } 456 #endif 457 } 458 459 static void __kmp_GOMP_fork_call(ident_t *loc, int gtid, unsigned num_threads, 460 unsigned flags, void (*unwrapped_task)(void *), 461 microtask_t wrapper, int argc, ...) { 462 int rc; 463 kmp_info_t *thr = __kmp_threads[gtid]; 464 kmp_team_t *team = thr->th.th_team; 465 int tid = __kmp_tid_from_gtid(gtid); 466 467 va_list ap; 468 va_start(ap, argc); 469 470 if (num_threads != 0) 471 __kmp_push_num_threads(loc, gtid, num_threads); 472 if (flags != 0) 473 __kmp_push_proc_bind(loc, gtid, (kmp_proc_bind_t)flags); 474 rc = __kmp_fork_call(loc, gtid, fork_context_gnu, argc, wrapper, 475 __kmp_invoke_task_func, kmp_va_addr_of(ap)); 476 477 va_end(ap); 478 479 if (rc) { 480 __kmp_run_before_invoked_task(gtid, tid, thr, team); 481 } 482 483 #if OMPT_SUPPORT 484 int ompt_team_size; 485 if (ompt_enabled.enabled) { 486 ompt_team_info_t *team_info = __ompt_get_teaminfo(0, NULL); 487 ompt_task_info_t *task_info = __ompt_get_task_info_object(0); 488 489 // implicit task callback 490 if (ompt_enabled.ompt_callback_implicit_task) { 491 ompt_team_size = __kmp_team_from_gtid(gtid)->t.t_nproc; 492 ompt_callbacks.ompt_callback(ompt_callback_implicit_task)( 493 ompt_scope_begin, &(team_info->parallel_data), 494 &(task_info->task_data), ompt_team_size, __kmp_tid_from_gtid(gtid), 495 ompt_task_implicit); // TODO: Can this be ompt_task_initial? 496 task_info->thread_num = __kmp_tid_from_gtid(gtid); 497 } 498 thr->th.ompt_thread_info.state = ompt_state_work_parallel; 499 } 500 #endif 501 } 502 503 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_START)(void (*task)(void *), 504 void *data, 505 unsigned num_threads) { 506 int gtid = __kmp_entry_gtid(); 507 508 #if OMPT_SUPPORT 509 ompt_frame_t *parent_frame, *frame; 510 511 if (ompt_enabled.enabled) { 512 __ompt_get_task_info_internal(0, NULL, NULL, &parent_frame, NULL, NULL); 513 parent_frame->enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0); 514 } 515 OMPT_STORE_RETURN_ADDRESS(gtid); 516 #endif 517 518 MKLOC(loc, "GOMP_parallel_start"); 519 KA_TRACE(20, ("GOMP_parallel_start: T#%d\n", gtid)); 520 __kmp_GOMP_fork_call(&loc, gtid, num_threads, 0u, task, 521 (microtask_t)__kmp_GOMP_microtask_wrapper, 2, task, 522 data); 523 #if OMPT_SUPPORT 524 if (ompt_enabled.enabled) { 525 __ompt_get_task_info_internal(0, NULL, NULL, &frame, NULL, NULL); 526 frame->exit_frame.ptr = OMPT_GET_FRAME_ADDRESS(0); 527 } 528 #endif 529 #if OMPD_SUPPORT 530 if (ompd_state & OMPD_ENABLE_BP) 531 ompd_bp_parallel_begin(); 532 #endif 533 } 534 535 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_END)(void) { 536 int gtid = __kmp_get_gtid(); 537 kmp_info_t *thr; 538 539 thr = __kmp_threads[gtid]; 540 541 MKLOC(loc, "GOMP_parallel_end"); 542 KA_TRACE(20, ("GOMP_parallel_end: T#%d\n", gtid)); 543 544 if (!thr->th.th_team->t.t_serialized) { 545 __kmp_run_after_invoked_task(gtid, __kmp_tid_from_gtid(gtid), thr, 546 thr->th.th_team); 547 } 548 #if OMPT_SUPPORT 549 if (ompt_enabled.enabled) { 550 // Implicit task is finished here, in the barrier we might schedule 551 // deferred tasks, 552 // these don't see the implicit task on the stack 553 OMPT_CUR_TASK_INFO(thr)->frame.exit_frame = ompt_data_none; 554 } 555 #endif 556 557 __kmp_join_call(&loc, gtid 558 #if OMPT_SUPPORT 559 , 560 fork_context_gnu 561 #endif 562 ); 563 #if OMPD_SUPPORT 564 if (ompd_state & OMPD_ENABLE_BP) 565 ompd_bp_parallel_end(); 566 #endif 567 } 568 569 // Loop worksharing constructs 570 571 // The Gnu codegen passes in an exclusive upper bound for the overall range, 572 // but the libguide dispatch code expects an inclusive upper bound, hence the 573 // "end - incr" 5th argument to KMP_DISPATCH_INIT (and the " ub - str" 11th 574 // argument to __kmp_GOMP_fork_call). 575 // 576 // Conversely, KMP_DISPATCH_NEXT returns and inclusive upper bound in *p_ub, 577 // but the Gnu codegen expects an exclusive upper bound, so the adjustment 578 // "*p_ub += stride" compensates for the discrepancy. 579 // 580 // Correction: the gnu codegen always adjusts the upper bound by +-1, not the 581 // stride value. We adjust the dispatch parameters accordingly (by +-1), but 582 // we still adjust p_ub by the actual stride value. 583 // 584 // The "runtime" versions do not take a chunk_sz parameter. 585 // 586 // The profile lib cannot support construct checking of unordered loops that 587 // are predetermined by the compiler to be statically scheduled, as the gcc 588 // codegen will not always emit calls to GOMP_loop_static_next() to get the 589 // next iteration. Instead, it emits inline code to call omp_get_thread_num() 590 // num and calculate the iteration space using the result. It doesn't do this 591 // with ordered static loop, so they can be checked. 592 593 #if OMPT_SUPPORT 594 #define IF_OMPT_SUPPORT(code) code 595 #else 596 #define IF_OMPT_SUPPORT(code) 597 #endif 598 599 #define LOOP_START(func, schedule) \ 600 int func(long lb, long ub, long str, long chunk_sz, long *p_lb, \ 601 long *p_ub) { \ 602 int status; \ 603 long stride; \ 604 int gtid = __kmp_entry_gtid(); \ 605 MKLOC(loc, KMP_STR(func)); \ 606 KA_TRACE( \ 607 20, \ 608 (KMP_STR( \ 609 func) ": T#%d, lb 0x%lx, ub 0x%lx, str 0x%lx, chunk_sz 0x%lx\n", \ 610 gtid, lb, ub, str, chunk_sz)); \ 611 \ 612 if ((str > 0) ? (lb < ub) : (lb > ub)) { \ 613 { \ 614 IF_OMPT_SUPPORT(OMPT_STORE_RETURN_ADDRESS(gtid);) \ 615 KMP_DISPATCH_INIT(&loc, gtid, (schedule), lb, \ 616 (str > 0) ? (ub - 1) : (ub + 1), str, chunk_sz, \ 617 (schedule) != kmp_sch_static); \ 618 } \ 619 { \ 620 IF_OMPT_SUPPORT(OMPT_STORE_RETURN_ADDRESS(gtid);) \ 621 status = KMP_DISPATCH_NEXT(&loc, gtid, NULL, (kmp_int *)p_lb, \ 622 (kmp_int *)p_ub, (kmp_int *)&stride); \ 623 } \ 624 if (status) { \ 625 KMP_DEBUG_ASSERT(stride == str); \ 626 *p_ub += (str > 0) ? 1 : -1; \ 627 } \ 628 } else { \ 629 status = 0; \ 630 } \ 631 \ 632 KA_TRACE( \ 633 20, \ 634 (KMP_STR( \ 635 func) " exit: T#%d, *p_lb 0x%lx, *p_ub 0x%lx, returning %d\n", \ 636 gtid, *p_lb, *p_ub, status)); \ 637 return status; \ 638 } 639 640 #define LOOP_RUNTIME_START(func, schedule) \ 641 int func(long lb, long ub, long str, long *p_lb, long *p_ub) { \ 642 int status; \ 643 long stride; \ 644 long chunk_sz = 0; \ 645 int gtid = __kmp_entry_gtid(); \ 646 MKLOC(loc, KMP_STR(func)); \ 647 KA_TRACE( \ 648 20, \ 649 (KMP_STR(func) ": T#%d, lb 0x%lx, ub 0x%lx, str 0x%lx, chunk_sz %d\n", \ 650 gtid, lb, ub, str, chunk_sz)); \ 651 \ 652 if ((str > 0) ? (lb < ub) : (lb > ub)) { \ 653 { \ 654 IF_OMPT_SUPPORT(OMPT_STORE_RETURN_ADDRESS(gtid);) \ 655 KMP_DISPATCH_INIT(&loc, gtid, (schedule), lb, \ 656 (str > 0) ? (ub - 1) : (ub + 1), str, chunk_sz, \ 657 TRUE); \ 658 } \ 659 { \ 660 IF_OMPT_SUPPORT(OMPT_STORE_RETURN_ADDRESS(gtid);) \ 661 status = KMP_DISPATCH_NEXT(&loc, gtid, NULL, (kmp_int *)p_lb, \ 662 (kmp_int *)p_ub, (kmp_int *)&stride); \ 663 } \ 664 if (status) { \ 665 KMP_DEBUG_ASSERT(stride == str); \ 666 *p_ub += (str > 0) ? 1 : -1; \ 667 } \ 668 } else { \ 669 status = 0; \ 670 } \ 671 \ 672 KA_TRACE( \ 673 20, \ 674 (KMP_STR( \ 675 func) " exit: T#%d, *p_lb 0x%lx, *p_ub 0x%lx, returning %d\n", \ 676 gtid, *p_lb, *p_ub, status)); \ 677 return status; \ 678 } 679 680 #define KMP_DOACROSS_FINI(status, gtid) \ 681 if (!status && __kmp_threads[gtid]->th.th_dispatch->th_doacross_flags) { \ 682 __kmpc_doacross_fini(NULL, gtid); \ 683 } 684 685 #define LOOP_NEXT(func, fini_code) \ 686 int func(long *p_lb, long *p_ub) { \ 687 int status; \ 688 long stride; \ 689 int gtid = __kmp_get_gtid(); \ 690 MKLOC(loc, KMP_STR(func)); \ 691 KA_TRACE(20, (KMP_STR(func) ": T#%d\n", gtid)); \ 692 \ 693 IF_OMPT_SUPPORT(OMPT_STORE_RETURN_ADDRESS(gtid);) \ 694 fini_code status = KMP_DISPATCH_NEXT(&loc, gtid, NULL, (kmp_int *)p_lb, \ 695 (kmp_int *)p_ub, (kmp_int *)&stride); \ 696 if (status) { \ 697 *p_ub += (stride > 0) ? 1 : -1; \ 698 } \ 699 KMP_DOACROSS_FINI(status, gtid) \ 700 \ 701 KA_TRACE( \ 702 20, \ 703 (KMP_STR(func) " exit: T#%d, *p_lb 0x%lx, *p_ub 0x%lx, stride 0x%lx, " \ 704 "returning %d\n", \ 705 gtid, *p_lb, *p_ub, stride, status)); \ 706 return status; \ 707 } 708 709 LOOP_START(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_STATIC_START), kmp_sch_static) 710 LOOP_NEXT(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_STATIC_NEXT), {}) 711 LOOP_START(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_DYNAMIC_START), 712 kmp_sch_dynamic_chunked) 713 LOOP_START(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_NONMONOTONIC_DYNAMIC_START), 714 kmp_sch_dynamic_chunked) 715 LOOP_NEXT(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_DYNAMIC_NEXT), {}) 716 LOOP_NEXT(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_NONMONOTONIC_DYNAMIC_NEXT), {}) 717 LOOP_START(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_GUIDED_START), 718 kmp_sch_guided_chunked) 719 LOOP_START(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_NONMONOTONIC_GUIDED_START), 720 kmp_sch_guided_chunked) 721 LOOP_NEXT(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_GUIDED_NEXT), {}) 722 LOOP_NEXT(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_NONMONOTONIC_GUIDED_NEXT), {}) 723 LOOP_RUNTIME_START(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_RUNTIME_START), 724 kmp_sch_runtime) 725 LOOP_NEXT(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_RUNTIME_NEXT), {}) 726 LOOP_RUNTIME_START( 727 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_MAYBE_NONMONOTONIC_RUNTIME_START), 728 kmp_sch_runtime) 729 LOOP_RUNTIME_START( 730 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_NONMONOTONIC_RUNTIME_START), 731 kmp_sch_runtime) 732 LOOP_NEXT( 733 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_MAYBE_NONMONOTONIC_RUNTIME_NEXT), {}) 734 LOOP_NEXT(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_NONMONOTONIC_RUNTIME_NEXT), {}) 735 736 LOOP_START(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ORDERED_STATIC_START), 737 kmp_ord_static) 738 LOOP_NEXT(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ORDERED_STATIC_NEXT), 739 { KMP_DISPATCH_FINI_CHUNK(&loc, gtid); }) 740 LOOP_START(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ORDERED_DYNAMIC_START), 741 kmp_ord_dynamic_chunked) 742 LOOP_NEXT(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ORDERED_DYNAMIC_NEXT), 743 { KMP_DISPATCH_FINI_CHUNK(&loc, gtid); }) 744 LOOP_START(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ORDERED_GUIDED_START), 745 kmp_ord_guided_chunked) 746 LOOP_NEXT(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ORDERED_GUIDED_NEXT), 747 { KMP_DISPATCH_FINI_CHUNK(&loc, gtid); }) 748 LOOP_RUNTIME_START( 749 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ORDERED_RUNTIME_START), 750 kmp_ord_runtime) 751 LOOP_NEXT(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ORDERED_RUNTIME_NEXT), 752 { KMP_DISPATCH_FINI_CHUNK(&loc, gtid); }) 753 754 #define LOOP_DOACROSS_START(func, schedule) \ 755 bool func(unsigned ncounts, long *counts, long chunk_sz, long *p_lb, \ 756 long *p_ub) { \ 757 int status; \ 758 long stride, lb, ub, str; \ 759 int gtid = __kmp_entry_gtid(); \ 760 struct kmp_dim *dims = \ 761 (struct kmp_dim *)__kmp_allocate(sizeof(struct kmp_dim) * ncounts); \ 762 MKLOC(loc, KMP_STR(func)); \ 763 for (unsigned i = 0; i < ncounts; ++i) { \ 764 dims[i].lo = 0; \ 765 dims[i].up = counts[i] - 1; \ 766 dims[i].st = 1; \ 767 } \ 768 __kmpc_doacross_init(&loc, gtid, (int)ncounts, dims); \ 769 lb = 0; \ 770 ub = counts[0]; \ 771 str = 1; \ 772 KA_TRACE(20, (KMP_STR(func) ": T#%d, ncounts %u, lb 0x%lx, ub 0x%lx, str " \ 773 "0x%lx, chunk_sz " \ 774 "0x%lx\n", \ 775 gtid, ncounts, lb, ub, str, chunk_sz)); \ 776 \ 777 if ((str > 0) ? (lb < ub) : (lb > ub)) { \ 778 KMP_DISPATCH_INIT(&loc, gtid, (schedule), lb, \ 779 (str > 0) ? (ub - 1) : (ub + 1), str, chunk_sz, \ 780 (schedule) != kmp_sch_static); \ 781 status = KMP_DISPATCH_NEXT(&loc, gtid, NULL, (kmp_int *)p_lb, \ 782 (kmp_int *)p_ub, (kmp_int *)&stride); \ 783 if (status) { \ 784 KMP_DEBUG_ASSERT(stride == str); \ 785 *p_ub += (str > 0) ? 1 : -1; \ 786 } \ 787 } else { \ 788 status = 0; \ 789 } \ 790 KMP_DOACROSS_FINI(status, gtid); \ 791 \ 792 KA_TRACE( \ 793 20, \ 794 (KMP_STR( \ 795 func) " exit: T#%d, *p_lb 0x%lx, *p_ub 0x%lx, returning %d\n", \ 796 gtid, *p_lb, *p_ub, status)); \ 797 __kmp_free(dims); \ 798 return status; \ 799 } 800 801 #define LOOP_DOACROSS_RUNTIME_START(func, schedule) \ 802 int func(unsigned ncounts, long *counts, long *p_lb, long *p_ub) { \ 803 int status; \ 804 long stride, lb, ub, str; \ 805 long chunk_sz = 0; \ 806 int gtid = __kmp_entry_gtid(); \ 807 struct kmp_dim *dims = \ 808 (struct kmp_dim *)__kmp_allocate(sizeof(struct kmp_dim) * ncounts); \ 809 MKLOC(loc, KMP_STR(func)); \ 810 for (unsigned i = 0; i < ncounts; ++i) { \ 811 dims[i].lo = 0; \ 812 dims[i].up = counts[i] - 1; \ 813 dims[i].st = 1; \ 814 } \ 815 __kmpc_doacross_init(&loc, gtid, (int)ncounts, dims); \ 816 lb = 0; \ 817 ub = counts[0]; \ 818 str = 1; \ 819 KA_TRACE( \ 820 20, \ 821 (KMP_STR(func) ": T#%d, lb 0x%lx, ub 0x%lx, str 0x%lx, chunk_sz %d\n", \ 822 gtid, lb, ub, str, chunk_sz)); \ 823 \ 824 if ((str > 0) ? (lb < ub) : (lb > ub)) { \ 825 KMP_DISPATCH_INIT(&loc, gtid, (schedule), lb, \ 826 (str > 0) ? (ub - 1) : (ub + 1), str, chunk_sz, TRUE); \ 827 status = KMP_DISPATCH_NEXT(&loc, gtid, NULL, (kmp_int *)p_lb, \ 828 (kmp_int *)p_ub, (kmp_int *)&stride); \ 829 if (status) { \ 830 KMP_DEBUG_ASSERT(stride == str); \ 831 *p_ub += (str > 0) ? 1 : -1; \ 832 } \ 833 } else { \ 834 status = 0; \ 835 } \ 836 KMP_DOACROSS_FINI(status, gtid); \ 837 \ 838 KA_TRACE( \ 839 20, \ 840 (KMP_STR( \ 841 func) " exit: T#%d, *p_lb 0x%lx, *p_ub 0x%lx, returning %d\n", \ 842 gtid, *p_lb, *p_ub, status)); \ 843 __kmp_free(dims); \ 844 return status; \ 845 } 846 847 LOOP_DOACROSS_START( 848 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_DOACROSS_STATIC_START), 849 kmp_sch_static) 850 LOOP_DOACROSS_START( 851 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_DOACROSS_DYNAMIC_START), 852 kmp_sch_dynamic_chunked) 853 LOOP_DOACROSS_START( 854 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_DOACROSS_GUIDED_START), 855 kmp_sch_guided_chunked) 856 LOOP_DOACROSS_RUNTIME_START( 857 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_DOACROSS_RUNTIME_START), 858 kmp_sch_runtime) 859 860 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_END)(void) { 861 int gtid = __kmp_get_gtid(); 862 KA_TRACE(20, ("GOMP_loop_end: T#%d\n", gtid)) 863 864 #if OMPT_SUPPORT && OMPT_OPTIONAL 865 ompt_frame_t *ompt_frame; 866 if (ompt_enabled.enabled) { 867 __ompt_get_task_info_internal(0, NULL, NULL, &ompt_frame, NULL, NULL); 868 ompt_frame->enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0); 869 OMPT_STORE_RETURN_ADDRESS(gtid); 870 } 871 #endif 872 __kmp_barrier(bs_plain_barrier, gtid, FALSE, 0, NULL, NULL); 873 #if OMPT_SUPPORT && OMPT_OPTIONAL 874 if (ompt_enabled.enabled) { 875 ompt_frame->enter_frame = ompt_data_none; 876 } 877 #endif 878 879 KA_TRACE(20, ("GOMP_loop_end exit: T#%d\n", gtid)) 880 } 881 882 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_END_NOWAIT)(void) { 883 KA_TRACE(20, ("GOMP_loop_end_nowait: T#%d\n", __kmp_get_gtid())) 884 } 885 886 // Unsigned long long loop worksharing constructs 887 // 888 // These are new with gcc 4.4 889 890 #define LOOP_START_ULL(func, schedule) \ 891 int func(int up, unsigned long long lb, unsigned long long ub, \ 892 unsigned long long str, unsigned long long chunk_sz, \ 893 unsigned long long *p_lb, unsigned long long *p_ub) { \ 894 int status; \ 895 long long str2 = up ? ((long long)str) : -((long long)str); \ 896 long long stride; \ 897 int gtid = __kmp_entry_gtid(); \ 898 MKLOC(loc, KMP_STR(func)); \ 899 \ 900 KA_TRACE(20, (KMP_STR(func) ": T#%d, up %d, lb 0x%llx, ub 0x%llx, str " \ 901 "0x%llx, chunk_sz 0x%llx\n", \ 902 gtid, up, lb, ub, str, chunk_sz)); \ 903 \ 904 if ((str > 0) ? (lb < ub) : (lb > ub)) { \ 905 KMP_DISPATCH_INIT_ULL(&loc, gtid, (schedule), lb, \ 906 (str2 > 0) ? (ub - 1) : (ub + 1), str2, chunk_sz, \ 907 (schedule) != kmp_sch_static); \ 908 status = \ 909 KMP_DISPATCH_NEXT_ULL(&loc, gtid, NULL, (kmp_uint64 *)p_lb, \ 910 (kmp_uint64 *)p_ub, (kmp_int64 *)&stride); \ 911 if (status) { \ 912 KMP_DEBUG_ASSERT(stride == str2); \ 913 *p_ub += (str > 0) ? 1 : -1; \ 914 } \ 915 } else { \ 916 status = 0; \ 917 } \ 918 \ 919 KA_TRACE( \ 920 20, \ 921 (KMP_STR( \ 922 func) " exit: T#%d, *p_lb 0x%llx, *p_ub 0x%llx, returning %d\n", \ 923 gtid, *p_lb, *p_ub, status)); \ 924 return status; \ 925 } 926 927 #define LOOP_RUNTIME_START_ULL(func, schedule) \ 928 int func(int up, unsigned long long lb, unsigned long long ub, \ 929 unsigned long long str, unsigned long long *p_lb, \ 930 unsigned long long *p_ub) { \ 931 int status; \ 932 long long str2 = up ? ((long long)str) : -((long long)str); \ 933 unsigned long long stride; \ 934 unsigned long long chunk_sz = 0; \ 935 int gtid = __kmp_entry_gtid(); \ 936 MKLOC(loc, KMP_STR(func)); \ 937 \ 938 KA_TRACE(20, (KMP_STR(func) ": T#%d, up %d, lb 0x%llx, ub 0x%llx, str " \ 939 "0x%llx, chunk_sz 0x%llx\n", \ 940 gtid, up, lb, ub, str, chunk_sz)); \ 941 \ 942 if ((str > 0) ? (lb < ub) : (lb > ub)) { \ 943 KMP_DISPATCH_INIT_ULL(&loc, gtid, (schedule), lb, \ 944 (str2 > 0) ? (ub - 1) : (ub + 1), str2, chunk_sz, \ 945 TRUE); \ 946 status = \ 947 KMP_DISPATCH_NEXT_ULL(&loc, gtid, NULL, (kmp_uint64 *)p_lb, \ 948 (kmp_uint64 *)p_ub, (kmp_int64 *)&stride); \ 949 if (status) { \ 950 KMP_DEBUG_ASSERT((long long)stride == str2); \ 951 *p_ub += (str > 0) ? 1 : -1; \ 952 } \ 953 } else { \ 954 status = 0; \ 955 } \ 956 \ 957 KA_TRACE( \ 958 20, \ 959 (KMP_STR( \ 960 func) " exit: T#%d, *p_lb 0x%llx, *p_ub 0x%llx, returning %d\n", \ 961 gtid, *p_lb, *p_ub, status)); \ 962 return status; \ 963 } 964 965 #define LOOP_NEXT_ULL(func, fini_code) \ 966 int func(unsigned long long *p_lb, unsigned long long *p_ub) { \ 967 int status; \ 968 long long stride; \ 969 int gtid = __kmp_get_gtid(); \ 970 MKLOC(loc, KMP_STR(func)); \ 971 KA_TRACE(20, (KMP_STR(func) ": T#%d\n", gtid)); \ 972 \ 973 fini_code status = \ 974 KMP_DISPATCH_NEXT_ULL(&loc, gtid, NULL, (kmp_uint64 *)p_lb, \ 975 (kmp_uint64 *)p_ub, (kmp_int64 *)&stride); \ 976 if (status) { \ 977 *p_ub += (stride > 0) ? 1 : -1; \ 978 } \ 979 \ 980 KA_TRACE( \ 981 20, \ 982 (KMP_STR( \ 983 func) " exit: T#%d, *p_lb 0x%llx, *p_ub 0x%llx, stride 0x%llx, " \ 984 "returning %d\n", \ 985 gtid, *p_lb, *p_ub, stride, status)); \ 986 return status; \ 987 } 988 989 LOOP_START_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_STATIC_START), 990 kmp_sch_static) 991 LOOP_NEXT_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_STATIC_NEXT), {}) 992 LOOP_START_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_DYNAMIC_START), 993 kmp_sch_dynamic_chunked) 994 LOOP_NEXT_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_DYNAMIC_NEXT), {}) 995 LOOP_START_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_GUIDED_START), 996 kmp_sch_guided_chunked) 997 LOOP_NEXT_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_GUIDED_NEXT), {}) 998 LOOP_START_ULL( 999 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_NONMONOTONIC_DYNAMIC_START), 1000 kmp_sch_dynamic_chunked) 1001 LOOP_NEXT_ULL( 1002 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_NONMONOTONIC_DYNAMIC_NEXT), {}) 1003 LOOP_START_ULL( 1004 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_NONMONOTONIC_GUIDED_START), 1005 kmp_sch_guided_chunked) 1006 LOOP_NEXT_ULL( 1007 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_NONMONOTONIC_GUIDED_NEXT), {}) 1008 LOOP_RUNTIME_START_ULL( 1009 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_RUNTIME_START), kmp_sch_runtime) 1010 LOOP_NEXT_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_RUNTIME_NEXT), {}) 1011 LOOP_RUNTIME_START_ULL( 1012 KMP_EXPAND_NAME( 1013 KMP_API_NAME_GOMP_LOOP_ULL_MAYBE_NONMONOTONIC_RUNTIME_START), 1014 kmp_sch_runtime) 1015 LOOP_RUNTIME_START_ULL( 1016 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_NONMONOTONIC_RUNTIME_START), 1017 kmp_sch_runtime) 1018 LOOP_NEXT_ULL( 1019 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_MAYBE_NONMONOTONIC_RUNTIME_NEXT), 1020 {}) 1021 LOOP_NEXT_ULL( 1022 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_NONMONOTONIC_RUNTIME_NEXT), {}) 1023 1024 LOOP_START_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_STATIC_START), 1025 kmp_ord_static) 1026 LOOP_NEXT_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_STATIC_NEXT), 1027 { KMP_DISPATCH_FINI_CHUNK_ULL(&loc, gtid); }) 1028 LOOP_START_ULL( 1029 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_DYNAMIC_START), 1030 kmp_ord_dynamic_chunked) 1031 LOOP_NEXT_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_DYNAMIC_NEXT), 1032 { KMP_DISPATCH_FINI_CHUNK_ULL(&loc, gtid); }) 1033 LOOP_START_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_GUIDED_START), 1034 kmp_ord_guided_chunked) 1035 LOOP_NEXT_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_GUIDED_NEXT), 1036 { KMP_DISPATCH_FINI_CHUNK_ULL(&loc, gtid); }) 1037 LOOP_RUNTIME_START_ULL( 1038 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_RUNTIME_START), 1039 kmp_ord_runtime) 1040 LOOP_NEXT_ULL(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_RUNTIME_NEXT), 1041 { KMP_DISPATCH_FINI_CHUNK_ULL(&loc, gtid); }) 1042 1043 #define LOOP_DOACROSS_START_ULL(func, schedule) \ 1044 int func(unsigned ncounts, unsigned long long *counts, \ 1045 unsigned long long chunk_sz, unsigned long long *p_lb, \ 1046 unsigned long long *p_ub) { \ 1047 int status; \ 1048 long long stride, str, lb, ub; \ 1049 int gtid = __kmp_entry_gtid(); \ 1050 struct kmp_dim *dims = \ 1051 (struct kmp_dim *)__kmp_allocate(sizeof(struct kmp_dim) * ncounts); \ 1052 MKLOC(loc, KMP_STR(func)); \ 1053 for (unsigned i = 0; i < ncounts; ++i) { \ 1054 dims[i].lo = 0; \ 1055 dims[i].up = counts[i] - 1; \ 1056 dims[i].st = 1; \ 1057 } \ 1058 __kmpc_doacross_init(&loc, gtid, (int)ncounts, dims); \ 1059 lb = 0; \ 1060 ub = counts[0]; \ 1061 str = 1; \ 1062 \ 1063 KA_TRACE(20, (KMP_STR(func) ": T#%d, lb 0x%llx, ub 0x%llx, str " \ 1064 "0x%llx, chunk_sz 0x%llx\n", \ 1065 gtid, lb, ub, str, chunk_sz)); \ 1066 \ 1067 if ((str > 0) ? (lb < ub) : (lb > ub)) { \ 1068 KMP_DISPATCH_INIT_ULL(&loc, gtid, (schedule), lb, \ 1069 (str > 0) ? (ub - 1) : (ub + 1), str, chunk_sz, \ 1070 (schedule) != kmp_sch_static); \ 1071 status = \ 1072 KMP_DISPATCH_NEXT_ULL(&loc, gtid, NULL, (kmp_uint64 *)p_lb, \ 1073 (kmp_uint64 *)p_ub, (kmp_int64 *)&stride); \ 1074 if (status) { \ 1075 KMP_DEBUG_ASSERT(stride == str); \ 1076 *p_ub += (str > 0) ? 1 : -1; \ 1077 } \ 1078 } else { \ 1079 status = 0; \ 1080 } \ 1081 KMP_DOACROSS_FINI(status, gtid); \ 1082 \ 1083 KA_TRACE( \ 1084 20, \ 1085 (KMP_STR( \ 1086 func) " exit: T#%d, *p_lb 0x%llx, *p_ub 0x%llx, returning %d\n", \ 1087 gtid, *p_lb, *p_ub, status)); \ 1088 __kmp_free(dims); \ 1089 return status; \ 1090 } 1091 1092 #define LOOP_DOACROSS_RUNTIME_START_ULL(func, schedule) \ 1093 int func(unsigned ncounts, unsigned long long *counts, \ 1094 unsigned long long *p_lb, unsigned long long *p_ub) { \ 1095 int status; \ 1096 unsigned long long stride, str, lb, ub; \ 1097 unsigned long long chunk_sz = 0; \ 1098 int gtid = __kmp_entry_gtid(); \ 1099 struct kmp_dim *dims = \ 1100 (struct kmp_dim *)__kmp_allocate(sizeof(struct kmp_dim) * ncounts); \ 1101 MKLOC(loc, KMP_STR(func)); \ 1102 for (unsigned i = 0; i < ncounts; ++i) { \ 1103 dims[i].lo = 0; \ 1104 dims[i].up = counts[i] - 1; \ 1105 dims[i].st = 1; \ 1106 } \ 1107 __kmpc_doacross_init(&loc, gtid, (int)ncounts, dims); \ 1108 lb = 0; \ 1109 ub = counts[0]; \ 1110 str = 1; \ 1111 KA_TRACE(20, (KMP_STR(func) ": T#%d, lb 0x%llx, ub 0x%llx, str " \ 1112 "0x%llx, chunk_sz 0x%llx\n", \ 1113 gtid, lb, ub, str, chunk_sz)); \ 1114 \ 1115 if ((str > 0) ? (lb < ub) : (lb > ub)) { \ 1116 KMP_DISPATCH_INIT_ULL(&loc, gtid, (schedule), lb, \ 1117 (str > 0) ? (ub - 1) : (ub + 1), str, chunk_sz, \ 1118 TRUE); \ 1119 status = \ 1120 KMP_DISPATCH_NEXT_ULL(&loc, gtid, NULL, (kmp_uint64 *)p_lb, \ 1121 (kmp_uint64 *)p_ub, (kmp_int64 *)&stride); \ 1122 if (status) { \ 1123 KMP_DEBUG_ASSERT(stride == str); \ 1124 *p_ub += (str > 0) ? 1 : -1; \ 1125 } \ 1126 } else { \ 1127 status = 0; \ 1128 } \ 1129 KMP_DOACROSS_FINI(status, gtid); \ 1130 \ 1131 KA_TRACE( \ 1132 20, \ 1133 (KMP_STR( \ 1134 func) " exit: T#%d, *p_lb 0x%llx, *p_ub 0x%llx, returning %d\n", \ 1135 gtid, *p_lb, *p_ub, status)); \ 1136 __kmp_free(dims); \ 1137 return status; \ 1138 } 1139 1140 LOOP_DOACROSS_START_ULL( 1141 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_DOACROSS_STATIC_START), 1142 kmp_sch_static) 1143 LOOP_DOACROSS_START_ULL( 1144 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_DOACROSS_DYNAMIC_START), 1145 kmp_sch_dynamic_chunked) 1146 LOOP_DOACROSS_START_ULL( 1147 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_DOACROSS_GUIDED_START), 1148 kmp_sch_guided_chunked) 1149 LOOP_DOACROSS_RUNTIME_START_ULL( 1150 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_DOACROSS_RUNTIME_START), 1151 kmp_sch_runtime) 1152 1153 // Combined parallel / loop worksharing constructs 1154 // 1155 // There are no ull versions (yet). 1156 1157 #define PARALLEL_LOOP_START(func, schedule, ompt_pre, ompt_post) \ 1158 void func(void (*task)(void *), void *data, unsigned num_threads, long lb, \ 1159 long ub, long str, long chunk_sz) { \ 1160 int gtid = __kmp_entry_gtid(); \ 1161 MKLOC(loc, KMP_STR(func)); \ 1162 KA_TRACE( \ 1163 20, \ 1164 (KMP_STR( \ 1165 func) ": T#%d, lb 0x%lx, ub 0x%lx, str 0x%lx, chunk_sz 0x%lx\n", \ 1166 gtid, lb, ub, str, chunk_sz)); \ 1167 \ 1168 ompt_pre(); \ 1169 \ 1170 __kmp_GOMP_fork_call(&loc, gtid, num_threads, 0u, task, \ 1171 (microtask_t)__kmp_GOMP_parallel_microtask_wrapper, \ 1172 9, task, data, num_threads, &loc, (schedule), lb, \ 1173 (str > 0) ? (ub - 1) : (ub + 1), str, chunk_sz); \ 1174 IF_OMPT_SUPPORT(OMPT_STORE_RETURN_ADDRESS(gtid)); \ 1175 \ 1176 KMP_DISPATCH_INIT(&loc, gtid, (schedule), lb, \ 1177 (str > 0) ? (ub - 1) : (ub + 1), str, chunk_sz, \ 1178 (schedule) != kmp_sch_static); \ 1179 \ 1180 ompt_post(); \ 1181 \ 1182 KA_TRACE(20, (KMP_STR(func) " exit: T#%d\n", gtid)); \ 1183 } 1184 1185 #if OMPT_SUPPORT && OMPT_OPTIONAL 1186 1187 #define OMPT_LOOP_PRE() \ 1188 ompt_frame_t *parent_frame; \ 1189 if (ompt_enabled.enabled) { \ 1190 __ompt_get_task_info_internal(0, NULL, NULL, &parent_frame, NULL, NULL); \ 1191 parent_frame->enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0); \ 1192 OMPT_STORE_RETURN_ADDRESS(gtid); \ 1193 } 1194 1195 #define OMPT_LOOP_POST() \ 1196 if (ompt_enabled.enabled) { \ 1197 parent_frame->enter_frame = ompt_data_none; \ 1198 } 1199 1200 #else 1201 1202 #define OMPT_LOOP_PRE() 1203 1204 #define OMPT_LOOP_POST() 1205 1206 #endif 1207 1208 PARALLEL_LOOP_START( 1209 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_LOOP_STATIC_START), 1210 kmp_sch_static, OMPT_LOOP_PRE, OMPT_LOOP_POST) 1211 PARALLEL_LOOP_START( 1212 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_LOOP_DYNAMIC_START), 1213 kmp_sch_dynamic_chunked, OMPT_LOOP_PRE, OMPT_LOOP_POST) 1214 PARALLEL_LOOP_START( 1215 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_LOOP_GUIDED_START), 1216 kmp_sch_guided_chunked, OMPT_LOOP_PRE, OMPT_LOOP_POST) 1217 PARALLEL_LOOP_START( 1218 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_LOOP_RUNTIME_START), 1219 kmp_sch_runtime, OMPT_LOOP_PRE, OMPT_LOOP_POST) 1220 1221 // Tasking constructs 1222 1223 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASK)(void (*func)(void *), void *data, 1224 void (*copy_func)(void *, void *), 1225 long arg_size, long arg_align, 1226 bool if_cond, unsigned gomp_flags, 1227 void **depend) { 1228 MKLOC(loc, "GOMP_task"); 1229 int gtid = __kmp_entry_gtid(); 1230 kmp_int32 flags = 0; 1231 kmp_tasking_flags_t *input_flags = (kmp_tasking_flags_t *)&flags; 1232 1233 KA_TRACE(20, ("GOMP_task: T#%d\n", gtid)); 1234 1235 // The low-order bit is the "untied" flag 1236 if (!(gomp_flags & KMP_GOMP_TASK_UNTIED_FLAG)) { 1237 input_flags->tiedness = TASK_TIED; 1238 } 1239 // The second low-order bit is the "final" flag 1240 if (gomp_flags & KMP_GOMP_TASK_FINAL_FLAG) { 1241 input_flags->final = 1; 1242 } 1243 input_flags->native = 1; 1244 // __kmp_task_alloc() sets up all other flags 1245 1246 if (!if_cond) { 1247 arg_size = 0; 1248 } 1249 1250 kmp_task_t *task = __kmp_task_alloc( 1251 &loc, gtid, input_flags, sizeof(kmp_task_t), 1252 arg_size ? arg_size + arg_align - 1 : 0, (kmp_routine_entry_t)func); 1253 1254 if (arg_size > 0) { 1255 if (arg_align > 0) { 1256 task->shareds = (void *)((((size_t)task->shareds) + arg_align - 1) / 1257 arg_align * arg_align); 1258 } 1259 // else error?? 1260 1261 if (copy_func) { 1262 (*copy_func)(task->shareds, data); 1263 } else { 1264 KMP_MEMCPY(task->shareds, data, arg_size); 1265 } 1266 } 1267 1268 #if OMPT_SUPPORT 1269 kmp_taskdata_t *current_task; 1270 if (ompt_enabled.enabled) { 1271 current_task = __kmp_threads[gtid]->th.th_current_task; 1272 current_task->ompt_task_info.frame.enter_frame.ptr = 1273 OMPT_GET_FRAME_ADDRESS(0); 1274 } 1275 OMPT_STORE_RETURN_ADDRESS(gtid); 1276 #endif 1277 1278 if (if_cond) { 1279 if (gomp_flags & KMP_GOMP_TASK_DEPENDS_FLAG) { 1280 KMP_ASSERT(depend); 1281 kmp_gomp_depends_info_t gomp_depends(depend); 1282 kmp_int32 ndeps = gomp_depends.get_num_deps(); 1283 kmp_depend_info_t dep_list[ndeps]; 1284 for (kmp_int32 i = 0; i < ndeps; i++) 1285 dep_list[i] = gomp_depends.get_kmp_depend(i); 1286 kmp_int32 ndeps_cnv; 1287 __kmp_type_convert(ndeps, &ndeps_cnv); 1288 __kmpc_omp_task_with_deps(&loc, gtid, task, ndeps_cnv, dep_list, 0, NULL); 1289 } else { 1290 __kmpc_omp_task(&loc, gtid, task); 1291 } 1292 } else { 1293 #if OMPT_SUPPORT 1294 ompt_thread_info_t oldInfo; 1295 kmp_info_t *thread; 1296 kmp_taskdata_t *taskdata; 1297 if (ompt_enabled.enabled) { 1298 // Store the threads states and restore them after the task 1299 thread = __kmp_threads[gtid]; 1300 taskdata = KMP_TASK_TO_TASKDATA(task); 1301 oldInfo = thread->th.ompt_thread_info; 1302 thread->th.ompt_thread_info.wait_id = 0; 1303 thread->th.ompt_thread_info.state = ompt_state_work_parallel; 1304 taskdata->ompt_task_info.frame.exit_frame.ptr = OMPT_GET_FRAME_ADDRESS(0); 1305 } 1306 OMPT_STORE_RETURN_ADDRESS(gtid); 1307 #endif 1308 if (gomp_flags & KMP_GOMP_TASK_DEPENDS_FLAG) { 1309 KMP_ASSERT(depend); 1310 kmp_gomp_depends_info_t gomp_depends(depend); 1311 kmp_int32 ndeps = gomp_depends.get_num_deps(); 1312 kmp_depend_info_t dep_list[ndeps]; 1313 for (kmp_int32 i = 0; i < ndeps; i++) 1314 dep_list[i] = gomp_depends.get_kmp_depend(i); 1315 __kmpc_omp_wait_deps(&loc, gtid, ndeps, dep_list, 0, NULL); 1316 } 1317 1318 __kmpc_omp_task_begin_if0(&loc, gtid, task); 1319 func(data); 1320 __kmpc_omp_task_complete_if0(&loc, gtid, task); 1321 1322 #if OMPT_SUPPORT 1323 if (ompt_enabled.enabled) { 1324 thread->th.ompt_thread_info = oldInfo; 1325 taskdata->ompt_task_info.frame.exit_frame = ompt_data_none; 1326 } 1327 #endif 1328 } 1329 #if OMPT_SUPPORT 1330 if (ompt_enabled.enabled) { 1331 current_task->ompt_task_info.frame.enter_frame = ompt_data_none; 1332 } 1333 #endif 1334 1335 KA_TRACE(20, ("GOMP_task exit: T#%d\n", gtid)); 1336 } 1337 1338 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASKWAIT)(void) { 1339 MKLOC(loc, "GOMP_taskwait"); 1340 int gtid = __kmp_entry_gtid(); 1341 1342 #if OMPT_SUPPORT 1343 OMPT_STORE_RETURN_ADDRESS(gtid); 1344 #endif 1345 1346 KA_TRACE(20, ("GOMP_taskwait: T#%d\n", gtid)); 1347 1348 __kmpc_omp_taskwait(&loc, gtid); 1349 1350 KA_TRACE(20, ("GOMP_taskwait exit: T#%d\n", gtid)); 1351 } 1352 1353 // Sections worksharing constructs 1354 // 1355 // For the sections construct, we initialize a dynamically scheduled loop 1356 // worksharing construct with lb 1 and stride 1, and use the iteration #'s 1357 // that its returns as sections ids. 1358 // 1359 // There are no special entry points for ordered sections, so we always use 1360 // the dynamically scheduled workshare, even if the sections aren't ordered. 1361 1362 unsigned KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SECTIONS_START)(unsigned count) { 1363 int status; 1364 kmp_int lb, ub, stride; 1365 int gtid = __kmp_entry_gtid(); 1366 MKLOC(loc, "GOMP_sections_start"); 1367 KA_TRACE(20, ("GOMP_sections_start: T#%d\n", gtid)); 1368 1369 KMP_DISPATCH_INIT(&loc, gtid, kmp_nm_dynamic_chunked, 1, count, 1, 1, TRUE); 1370 1371 status = KMP_DISPATCH_NEXT(&loc, gtid, NULL, &lb, &ub, &stride); 1372 if (status) { 1373 KMP_DEBUG_ASSERT(stride == 1); 1374 KMP_DEBUG_ASSERT(lb > 0); 1375 KMP_ASSERT(lb == ub); 1376 } else { 1377 lb = 0; 1378 } 1379 1380 KA_TRACE(20, ("GOMP_sections_start exit: T#%d returning %u\n", gtid, 1381 (unsigned)lb)); 1382 return (unsigned)lb; 1383 } 1384 1385 unsigned KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SECTIONS_NEXT)(void) { 1386 int status; 1387 kmp_int lb, ub, stride; 1388 int gtid = __kmp_get_gtid(); 1389 MKLOC(loc, "GOMP_sections_next"); 1390 KA_TRACE(20, ("GOMP_sections_next: T#%d\n", gtid)); 1391 1392 #if OMPT_SUPPORT 1393 OMPT_STORE_RETURN_ADDRESS(gtid); 1394 #endif 1395 1396 status = KMP_DISPATCH_NEXT(&loc, gtid, NULL, &lb, &ub, &stride); 1397 if (status) { 1398 KMP_DEBUG_ASSERT(stride == 1); 1399 KMP_DEBUG_ASSERT(lb > 0); 1400 KMP_ASSERT(lb == ub); 1401 } else { 1402 lb = 0; 1403 } 1404 1405 KA_TRACE( 1406 20, ("GOMP_sections_next exit: T#%d returning %u\n", gtid, (unsigned)lb)); 1407 return (unsigned)lb; 1408 } 1409 1410 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_SECTIONS_START)( 1411 void (*task)(void *), void *data, unsigned num_threads, unsigned count) { 1412 int gtid = __kmp_entry_gtid(); 1413 1414 #if OMPT_SUPPORT 1415 ompt_frame_t *parent_frame; 1416 1417 if (ompt_enabled.enabled) { 1418 __ompt_get_task_info_internal(0, NULL, NULL, &parent_frame, NULL, NULL); 1419 parent_frame->enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0); 1420 } 1421 OMPT_STORE_RETURN_ADDRESS(gtid); 1422 #endif 1423 1424 MKLOC(loc, "GOMP_parallel_sections_start"); 1425 KA_TRACE(20, ("GOMP_parallel_sections_start: T#%d\n", gtid)); 1426 1427 __kmp_GOMP_fork_call(&loc, gtid, num_threads, 0u, task, 1428 (microtask_t)__kmp_GOMP_parallel_microtask_wrapper, 9, 1429 task, data, num_threads, &loc, kmp_nm_dynamic_chunked, 1430 (kmp_int)1, (kmp_int)count, (kmp_int)1, (kmp_int)1); 1431 1432 #if OMPT_SUPPORT 1433 if (ompt_enabled.enabled) { 1434 parent_frame->enter_frame = ompt_data_none; 1435 } 1436 #endif 1437 1438 KMP_DISPATCH_INIT(&loc, gtid, kmp_nm_dynamic_chunked, 1, count, 1, 1, TRUE); 1439 1440 KA_TRACE(20, ("GOMP_parallel_sections_start exit: T#%d\n", gtid)); 1441 } 1442 1443 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SECTIONS_END)(void) { 1444 int gtid = __kmp_get_gtid(); 1445 KA_TRACE(20, ("GOMP_sections_end: T#%d\n", gtid)) 1446 1447 #if OMPT_SUPPORT 1448 ompt_frame_t *ompt_frame; 1449 if (ompt_enabled.enabled) { 1450 __ompt_get_task_info_internal(0, NULL, NULL, &ompt_frame, NULL, NULL); 1451 ompt_frame->enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0); 1452 } 1453 OMPT_STORE_RETURN_ADDRESS(gtid); 1454 #endif 1455 __kmp_barrier(bs_plain_barrier, gtid, FALSE, 0, NULL, NULL); 1456 #if OMPT_SUPPORT 1457 if (ompt_enabled.enabled) { 1458 ompt_frame->enter_frame = ompt_data_none; 1459 } 1460 #endif 1461 1462 KA_TRACE(20, ("GOMP_sections_end exit: T#%d\n", gtid)) 1463 } 1464 1465 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SECTIONS_END_NOWAIT)(void) { 1466 KA_TRACE(20, ("GOMP_sections_end_nowait: T#%d\n", __kmp_get_gtid())) 1467 } 1468 1469 // libgomp has an empty function for GOMP_taskyield as of 2013-10-10 1470 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASKYIELD)(void) { 1471 KA_TRACE(20, ("GOMP_taskyield: T#%d\n", __kmp_get_gtid())) 1472 return; 1473 } 1474 1475 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL)(void (*task)(void *), 1476 void *data, 1477 unsigned num_threads, 1478 unsigned int flags) { 1479 int gtid = __kmp_entry_gtid(); 1480 MKLOC(loc, "GOMP_parallel"); 1481 KA_TRACE(20, ("GOMP_parallel: T#%d\n", gtid)); 1482 1483 #if OMPT_SUPPORT 1484 ompt_task_info_t *parent_task_info, *task_info; 1485 if (ompt_enabled.enabled) { 1486 parent_task_info = __ompt_get_task_info_object(0); 1487 parent_task_info->frame.enter_frame.ptr = OMPT_GET_FRAME_ADDRESS(0); 1488 } 1489 OMPT_STORE_RETURN_ADDRESS(gtid); 1490 #endif 1491 __kmp_GOMP_fork_call(&loc, gtid, num_threads, flags, task, 1492 (microtask_t)__kmp_GOMP_microtask_wrapper, 2, task, 1493 data); 1494 #if OMPT_SUPPORT 1495 if (ompt_enabled.enabled) { 1496 task_info = __ompt_get_task_info_object(0); 1497 task_info->frame.exit_frame.ptr = OMPT_GET_FRAME_ADDRESS(0); 1498 } 1499 #endif 1500 task(data); 1501 { 1502 #if OMPT_SUPPORT 1503 OMPT_STORE_RETURN_ADDRESS(gtid); 1504 #endif 1505 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_END)(); 1506 } 1507 #if OMPT_SUPPORT 1508 if (ompt_enabled.enabled) { 1509 task_info->frame.exit_frame = ompt_data_none; 1510 parent_task_info->frame.enter_frame = ompt_data_none; 1511 } 1512 #endif 1513 } 1514 1515 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_SECTIONS)(void (*task)(void *), 1516 void *data, 1517 unsigned num_threads, 1518 unsigned count, 1519 unsigned flags) { 1520 int gtid = __kmp_entry_gtid(); 1521 MKLOC(loc, "GOMP_parallel_sections"); 1522 KA_TRACE(20, ("GOMP_parallel_sections: T#%d\n", gtid)); 1523 1524 #if OMPT_SUPPORT 1525 OMPT_STORE_RETURN_ADDRESS(gtid); 1526 #endif 1527 1528 __kmp_GOMP_fork_call(&loc, gtid, num_threads, flags, task, 1529 (microtask_t)__kmp_GOMP_parallel_microtask_wrapper, 9, 1530 task, data, num_threads, &loc, kmp_nm_dynamic_chunked, 1531 (kmp_int)1, (kmp_int)count, (kmp_int)1, (kmp_int)1); 1532 1533 { 1534 #if OMPT_SUPPORT 1535 OMPT_STORE_RETURN_ADDRESS(gtid); 1536 #endif 1537 1538 KMP_DISPATCH_INIT(&loc, gtid, kmp_nm_dynamic_chunked, 1, count, 1, 1, TRUE); 1539 } 1540 task(data); 1541 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_END)(); 1542 KA_TRACE(20, ("GOMP_parallel_sections exit: T#%d\n", gtid)); 1543 } 1544 1545 #define PARALLEL_LOOP(func, schedule, ompt_pre, ompt_post) \ 1546 void func(void (*task)(void *), void *data, unsigned num_threads, long lb, \ 1547 long ub, long str, long chunk_sz, unsigned flags) { \ 1548 int gtid = __kmp_entry_gtid(); \ 1549 MKLOC(loc, KMP_STR(func)); \ 1550 KA_TRACE( \ 1551 20, \ 1552 (KMP_STR( \ 1553 func) ": T#%d, lb 0x%lx, ub 0x%lx, str 0x%lx, chunk_sz 0x%lx\n", \ 1554 gtid, lb, ub, str, chunk_sz)); \ 1555 \ 1556 ompt_pre(); \ 1557 IF_OMPT_SUPPORT(OMPT_STORE_RETURN_ADDRESS(gtid);) \ 1558 __kmp_GOMP_fork_call(&loc, gtid, num_threads, flags, task, \ 1559 (microtask_t)__kmp_GOMP_parallel_microtask_wrapper, \ 1560 9, task, data, num_threads, &loc, (schedule), lb, \ 1561 (str > 0) ? (ub - 1) : (ub + 1), str, chunk_sz); \ 1562 \ 1563 { \ 1564 IF_OMPT_SUPPORT(OMPT_STORE_RETURN_ADDRESS(gtid);) \ 1565 KMP_DISPATCH_INIT(&loc, gtid, (schedule), lb, \ 1566 (str > 0) ? (ub - 1) : (ub + 1), str, chunk_sz, \ 1567 (schedule) != kmp_sch_static); \ 1568 } \ 1569 task(data); \ 1570 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_END)(); \ 1571 ompt_post(); \ 1572 \ 1573 KA_TRACE(20, (KMP_STR(func) " exit: T#%d\n", gtid)); \ 1574 } 1575 1576 PARALLEL_LOOP(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_LOOP_STATIC), 1577 kmp_sch_static, OMPT_LOOP_PRE, OMPT_LOOP_POST) 1578 PARALLEL_LOOP(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_LOOP_DYNAMIC), 1579 kmp_sch_dynamic_chunked, OMPT_LOOP_PRE, OMPT_LOOP_POST) 1580 PARALLEL_LOOP( 1581 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_LOOP_NONMONOTONIC_GUIDED), 1582 kmp_sch_guided_chunked, OMPT_LOOP_PRE, OMPT_LOOP_POST) 1583 PARALLEL_LOOP( 1584 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_LOOP_NONMONOTONIC_DYNAMIC), 1585 kmp_sch_dynamic_chunked, OMPT_LOOP_PRE, OMPT_LOOP_POST) 1586 PARALLEL_LOOP(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_LOOP_GUIDED), 1587 kmp_sch_guided_chunked, OMPT_LOOP_PRE, OMPT_LOOP_POST) 1588 PARALLEL_LOOP(KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_LOOP_RUNTIME), 1589 kmp_sch_runtime, OMPT_LOOP_PRE, OMPT_LOOP_POST) 1590 PARALLEL_LOOP( 1591 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_LOOP_MAYBE_NONMONOTONIC_RUNTIME), 1592 kmp_sch_runtime, OMPT_LOOP_PRE, OMPT_LOOP_POST) 1593 PARALLEL_LOOP( 1594 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_LOOP_NONMONOTONIC_RUNTIME), 1595 kmp_sch_runtime, OMPT_LOOP_PRE, OMPT_LOOP_POST) 1596 1597 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASKGROUP_START)(void) { 1598 int gtid = __kmp_entry_gtid(); 1599 MKLOC(loc, "GOMP_taskgroup_start"); 1600 KA_TRACE(20, ("GOMP_taskgroup_start: T#%d\n", gtid)); 1601 1602 #if OMPT_SUPPORT 1603 OMPT_STORE_RETURN_ADDRESS(gtid); 1604 #endif 1605 1606 __kmpc_taskgroup(&loc, gtid); 1607 1608 return; 1609 } 1610 1611 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASKGROUP_END)(void) { 1612 int gtid = __kmp_get_gtid(); 1613 MKLOC(loc, "GOMP_taskgroup_end"); 1614 KA_TRACE(20, ("GOMP_taskgroup_end: T#%d\n", gtid)); 1615 1616 #if OMPT_SUPPORT 1617 OMPT_STORE_RETURN_ADDRESS(gtid); 1618 #endif 1619 1620 __kmpc_end_taskgroup(&loc, gtid); 1621 1622 return; 1623 } 1624 1625 static kmp_int32 __kmp_gomp_to_omp_cancellation_kind(int gomp_kind) { 1626 kmp_int32 cncl_kind = 0; 1627 switch (gomp_kind) { 1628 case 1: 1629 cncl_kind = cancel_parallel; 1630 break; 1631 case 2: 1632 cncl_kind = cancel_loop; 1633 break; 1634 case 4: 1635 cncl_kind = cancel_sections; 1636 break; 1637 case 8: 1638 cncl_kind = cancel_taskgroup; 1639 break; 1640 } 1641 return cncl_kind; 1642 } 1643 1644 // Return true if cancellation should take place, false otherwise 1645 bool KMP_EXPAND_NAME(KMP_API_NAME_GOMP_CANCELLATION_POINT)(int which) { 1646 int gtid = __kmp_get_gtid(); 1647 MKLOC(loc, "GOMP_cancellation_point"); 1648 KA_TRACE(20, ("GOMP_cancellation_point: T#%d which:%d\n", gtid, which)); 1649 kmp_int32 cncl_kind = __kmp_gomp_to_omp_cancellation_kind(which); 1650 return __kmpc_cancellationpoint(&loc, gtid, cncl_kind); 1651 } 1652 1653 // Return true if cancellation should take place, false otherwise 1654 bool KMP_EXPAND_NAME(KMP_API_NAME_GOMP_CANCEL)(int which, bool do_cancel) { 1655 int gtid = __kmp_get_gtid(); 1656 MKLOC(loc, "GOMP_cancel"); 1657 KA_TRACE(20, ("GOMP_cancel: T#%d which:%d do_cancel:%d\n", gtid, which, 1658 (int)do_cancel)); 1659 kmp_int32 cncl_kind = __kmp_gomp_to_omp_cancellation_kind(which); 1660 1661 if (do_cancel == FALSE) { 1662 return __kmpc_cancellationpoint(&loc, gtid, cncl_kind); 1663 } else { 1664 return __kmpc_cancel(&loc, gtid, cncl_kind); 1665 } 1666 } 1667 1668 // Return true if cancellation should take place, false otherwise 1669 bool KMP_EXPAND_NAME(KMP_API_NAME_GOMP_BARRIER_CANCEL)(void) { 1670 int gtid = __kmp_get_gtid(); 1671 KA_TRACE(20, ("GOMP_barrier_cancel: T#%d\n", gtid)); 1672 return __kmp_barrier_gomp_cancel(gtid); 1673 } 1674 1675 // Return true if cancellation should take place, false otherwise 1676 bool KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SECTIONS_END_CANCEL)(void) { 1677 int gtid = __kmp_get_gtid(); 1678 KA_TRACE(20, ("GOMP_sections_end_cancel: T#%d\n", gtid)); 1679 return __kmp_barrier_gomp_cancel(gtid); 1680 } 1681 1682 // Return true if cancellation should take place, false otherwise 1683 bool KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_END_CANCEL)(void) { 1684 int gtid = __kmp_get_gtid(); 1685 KA_TRACE(20, ("GOMP_loop_end_cancel: T#%d\n", gtid)); 1686 return __kmp_barrier_gomp_cancel(gtid); 1687 } 1688 1689 // All target functions are empty as of 2014-05-29 1690 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TARGET)(int device, void (*fn)(void *), 1691 const void *openmp_target, 1692 size_t mapnum, void **hostaddrs, 1693 size_t *sizes, 1694 unsigned char *kinds) { 1695 return; 1696 } 1697 1698 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TARGET_DATA)( 1699 int device, const void *openmp_target, size_t mapnum, void **hostaddrs, 1700 size_t *sizes, unsigned char *kinds) { 1701 return; 1702 } 1703 1704 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TARGET_END_DATA)(void) { return; } 1705 1706 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TARGET_UPDATE)( 1707 int device, const void *openmp_target, size_t mapnum, void **hostaddrs, 1708 size_t *sizes, unsigned char *kinds) { 1709 return; 1710 } 1711 1712 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TEAMS)(unsigned int num_teams, 1713 unsigned int thread_limit) { 1714 return; 1715 } 1716 1717 // Task duplication function which copies src to dest (both are 1718 // preallocated task structures) 1719 static void __kmp_gomp_task_dup(kmp_task_t *dest, kmp_task_t *src, 1720 kmp_int32 last_private) { 1721 kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(src); 1722 if (taskdata->td_copy_func) { 1723 (taskdata->td_copy_func)(dest->shareds, src->shareds); 1724 } 1725 } 1726 1727 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASKGROUP_REDUCTION_REGISTER)( 1728 uintptr_t *); 1729 1730 #ifdef __cplusplus 1731 } // extern "C" 1732 #endif 1733 1734 template <typename T> 1735 void __GOMP_taskloop(void (*func)(void *), void *data, 1736 void (*copy_func)(void *, void *), long arg_size, 1737 long arg_align, unsigned gomp_flags, 1738 unsigned long num_tasks, int priority, T start, T end, 1739 T step) { 1740 typedef void (*p_task_dup_t)(kmp_task_t *, kmp_task_t *, kmp_int32); 1741 MKLOC(loc, "GOMP_taskloop"); 1742 int sched; 1743 T *loop_bounds; 1744 int gtid = __kmp_entry_gtid(); 1745 kmp_int32 flags = 0; 1746 int if_val = gomp_flags & (1u << 10); 1747 int nogroup = gomp_flags & (1u << 11); 1748 int up = gomp_flags & (1u << 8); 1749 int reductions = gomp_flags & (1u << 12); 1750 p_task_dup_t task_dup = NULL; 1751 kmp_tasking_flags_t *input_flags = (kmp_tasking_flags_t *)&flags; 1752 #ifdef KMP_DEBUG 1753 { 1754 char *buff; 1755 buff = __kmp_str_format( 1756 "GOMP_taskloop: T#%%d: func:%%p data:%%p copy_func:%%p " 1757 "arg_size:%%ld arg_align:%%ld gomp_flags:0x%%x num_tasks:%%lu " 1758 "priority:%%d start:%%%s end:%%%s step:%%%s\n", 1759 traits_t<T>::spec, traits_t<T>::spec, traits_t<T>::spec); 1760 KA_TRACE(20, (buff, gtid, func, data, copy_func, arg_size, arg_align, 1761 gomp_flags, num_tasks, priority, start, end, step)); 1762 __kmp_str_free(&buff); 1763 } 1764 #endif 1765 KMP_ASSERT((size_t)arg_size >= 2 * sizeof(T)); 1766 KMP_ASSERT(arg_align > 0); 1767 // The low-order bit is the "untied" flag 1768 if (!(gomp_flags & 1)) { 1769 input_flags->tiedness = TASK_TIED; 1770 } 1771 // The second low-order bit is the "final" flag 1772 if (gomp_flags & 2) { 1773 input_flags->final = 1; 1774 } 1775 // Negative step flag 1776 if (!up) { 1777 // If step is flagged as negative, but isn't properly sign extended 1778 // Then manually sign extend it. Could be a short, int, char embedded 1779 // in a long. So cannot assume any cast. 1780 if (step > 0) { 1781 for (int i = sizeof(T) * CHAR_BIT - 1; i >= 0L; --i) { 1782 // break at the first 1 bit 1783 if (step & ((T)1 << i)) 1784 break; 1785 step |= ((T)1 << i); 1786 } 1787 } 1788 } 1789 input_flags->native = 1; 1790 // Figure out if none/grainsize/num_tasks clause specified 1791 if (num_tasks > 0) { 1792 if (gomp_flags & (1u << 9)) 1793 sched = 1; // grainsize specified 1794 else 1795 sched = 2; // num_tasks specified 1796 // neither grainsize nor num_tasks specified 1797 } else { 1798 sched = 0; 1799 } 1800 1801 // __kmp_task_alloc() sets up all other flags 1802 kmp_task_t *task = 1803 __kmp_task_alloc(&loc, gtid, input_flags, sizeof(kmp_task_t), 1804 arg_size + arg_align - 1, (kmp_routine_entry_t)func); 1805 kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task); 1806 taskdata->td_copy_func = copy_func; 1807 taskdata->td_size_loop_bounds = sizeof(T); 1808 1809 // re-align shareds if needed and setup firstprivate copy constructors 1810 // through the task_dup mechanism 1811 task->shareds = (void *)((((size_t)task->shareds) + arg_align - 1) / 1812 arg_align * arg_align); 1813 if (copy_func) { 1814 task_dup = __kmp_gomp_task_dup; 1815 } 1816 KMP_MEMCPY(task->shareds, data, arg_size); 1817 1818 loop_bounds = (T *)task->shareds; 1819 loop_bounds[0] = start; 1820 loop_bounds[1] = end + (up ? -1 : 1); 1821 1822 if (!nogroup) { 1823 #if OMPT_SUPPORT && OMPT_OPTIONAL 1824 OMPT_STORE_RETURN_ADDRESS(gtid); 1825 #endif 1826 __kmpc_taskgroup(&loc, gtid); 1827 if (reductions) { 1828 // The data pointer points to lb, ub, then reduction data 1829 struct data_t { 1830 T a, b; 1831 uintptr_t *d; 1832 }; 1833 uintptr_t *d = ((data_t *)data)->d; 1834 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASKGROUP_REDUCTION_REGISTER)(d); 1835 } 1836 } 1837 __kmpc_taskloop(&loc, gtid, task, if_val, (kmp_uint64 *)&(loop_bounds[0]), 1838 (kmp_uint64 *)&(loop_bounds[1]), (kmp_int64)step, 1, sched, 1839 (kmp_uint64)num_tasks, (void *)task_dup); 1840 if (!nogroup) { 1841 #if OMPT_SUPPORT && OMPT_OPTIONAL 1842 OMPT_STORE_RETURN_ADDRESS(gtid); 1843 #endif 1844 __kmpc_end_taskgroup(&loc, gtid); 1845 } 1846 } 1847 1848 // 4 byte version of GOMP_doacross_post 1849 // This verison needs to create a temporary array which converts 4 byte 1850 // integers into 8 byte integers 1851 template <typename T, bool need_conversion = (sizeof(long) == 4)> 1852 void __kmp_GOMP_doacross_post(T *count); 1853 1854 template <> void __kmp_GOMP_doacross_post<long, true>(long *count) { 1855 int gtid = __kmp_entry_gtid(); 1856 kmp_info_t *th = __kmp_threads[gtid]; 1857 MKLOC(loc, "GOMP_doacross_post"); 1858 kmp_int64 num_dims = th->th.th_dispatch->th_doacross_info[0]; 1859 kmp_int64 *vec = (kmp_int64 *)__kmp_thread_malloc( 1860 th, (size_t)(sizeof(kmp_int64) * num_dims)); 1861 for (kmp_int64 i = 0; i < num_dims; ++i) { 1862 vec[i] = (kmp_int64)count[i]; 1863 } 1864 __kmpc_doacross_post(&loc, gtid, vec); 1865 __kmp_thread_free(th, vec); 1866 } 1867 1868 // 8 byte versions of GOMP_doacross_post 1869 // This version can just pass in the count array directly instead of creating 1870 // a temporary array 1871 template <> void __kmp_GOMP_doacross_post<long, false>(long *count) { 1872 int gtid = __kmp_entry_gtid(); 1873 MKLOC(loc, "GOMP_doacross_post"); 1874 __kmpc_doacross_post(&loc, gtid, RCAST(kmp_int64 *, count)); 1875 } 1876 1877 template <typename T> void __kmp_GOMP_doacross_wait(T first, va_list args) { 1878 int gtid = __kmp_entry_gtid(); 1879 kmp_info_t *th = __kmp_threads[gtid]; 1880 MKLOC(loc, "GOMP_doacross_wait"); 1881 kmp_int64 num_dims = th->th.th_dispatch->th_doacross_info[0]; 1882 kmp_int64 *vec = (kmp_int64 *)__kmp_thread_malloc( 1883 th, (size_t)(sizeof(kmp_int64) * num_dims)); 1884 vec[0] = (kmp_int64)first; 1885 for (kmp_int64 i = 1; i < num_dims; ++i) { 1886 T item = va_arg(args, T); 1887 vec[i] = (kmp_int64)item; 1888 } 1889 __kmpc_doacross_wait(&loc, gtid, vec); 1890 __kmp_thread_free(th, vec); 1891 return; 1892 } 1893 1894 #ifdef __cplusplus 1895 extern "C" { 1896 #endif // __cplusplus 1897 1898 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASKLOOP)( 1899 void (*func)(void *), void *data, void (*copy_func)(void *, void *), 1900 long arg_size, long arg_align, unsigned gomp_flags, unsigned long num_tasks, 1901 int priority, long start, long end, long step) { 1902 __GOMP_taskloop<long>(func, data, copy_func, arg_size, arg_align, gomp_flags, 1903 num_tasks, priority, start, end, step); 1904 } 1905 1906 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASKLOOP_ULL)( 1907 void (*func)(void *), void *data, void (*copy_func)(void *, void *), 1908 long arg_size, long arg_align, unsigned gomp_flags, unsigned long num_tasks, 1909 int priority, unsigned long long start, unsigned long long end, 1910 unsigned long long step) { 1911 __GOMP_taskloop<unsigned long long>(func, data, copy_func, arg_size, 1912 arg_align, gomp_flags, num_tasks, 1913 priority, start, end, step); 1914 } 1915 1916 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_DOACROSS_POST)(long *count) { 1917 __kmp_GOMP_doacross_post(count); 1918 } 1919 1920 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_DOACROSS_WAIT)(long first, ...) { 1921 va_list args; 1922 va_start(args, first); 1923 __kmp_GOMP_doacross_wait<long>(first, args); 1924 va_end(args); 1925 } 1926 1927 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_DOACROSS_ULL_POST)( 1928 unsigned long long *count) { 1929 int gtid = __kmp_entry_gtid(); 1930 MKLOC(loc, "GOMP_doacross_ull_post"); 1931 __kmpc_doacross_post(&loc, gtid, RCAST(kmp_int64 *, count)); 1932 } 1933 1934 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_DOACROSS_ULL_WAIT)( 1935 unsigned long long first, ...) { 1936 va_list args; 1937 va_start(args, first); 1938 __kmp_GOMP_doacross_wait<unsigned long long>(first, args); 1939 va_end(args); 1940 } 1941 1942 // fn: the function each primary thread of new team will call 1943 // data: argument to fn 1944 // num_teams, thread_limit: max bounds on respective ICV 1945 // flags: unused 1946 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TEAMS_REG)(void (*fn)(void *), 1947 void *data, 1948 unsigned num_teams, 1949 unsigned thread_limit, 1950 unsigned flags) { 1951 MKLOC(loc, "GOMP_teams_reg"); 1952 int gtid = __kmp_entry_gtid(); 1953 KA_TRACE(20, ("GOMP_teams_reg: T#%d num_teams=%u thread_limit=%u flag=%u\n", 1954 gtid, num_teams, thread_limit, flags)); 1955 __kmpc_push_num_teams(&loc, gtid, num_teams, thread_limit); 1956 __kmpc_fork_teams(&loc, 2, (microtask_t)__kmp_GOMP_microtask_wrapper, fn, 1957 data); 1958 KA_TRACE(20, ("GOMP_teams_reg exit: T#%d\n", gtid)); 1959 } 1960 1961 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASKWAIT_DEPEND)(void **depend) { 1962 MKLOC(loc, "GOMP_taskwait_depend"); 1963 int gtid = __kmp_entry_gtid(); 1964 KA_TRACE(20, ("GOMP_taskwait_depend: T#%d\n", gtid)); 1965 kmp_gomp_depends_info_t gomp_depends(depend); 1966 kmp_int32 ndeps = gomp_depends.get_num_deps(); 1967 kmp_depend_info_t dep_list[ndeps]; 1968 for (kmp_int32 i = 0; i < ndeps; i++) 1969 dep_list[i] = gomp_depends.get_kmp_depend(i); 1970 #if OMPT_SUPPORT 1971 OMPT_STORE_RETURN_ADDRESS(gtid); 1972 #endif 1973 __kmpc_omp_wait_deps(&loc, gtid, ndeps, dep_list, 0, NULL); 1974 KA_TRACE(20, ("GOMP_taskwait_depend exit: T#%d\n", gtid)); 1975 } 1976 1977 static inline void 1978 __kmp_GOMP_taskgroup_reduction_register(uintptr_t *data, kmp_taskgroup_t *tg, 1979 int nthreads, 1980 uintptr_t *allocated = nullptr) { 1981 KMP_ASSERT(data); 1982 KMP_ASSERT(nthreads > 0); 1983 // Have private copy pointers point to previously allocated 1984 // reduction data or allocate new data here 1985 if (allocated) { 1986 data[2] = allocated[2]; 1987 data[6] = allocated[6]; 1988 } else { 1989 data[2] = (uintptr_t)__kmp_allocate(nthreads * data[1]); 1990 data[6] = data[2] + (nthreads * data[1]); 1991 } 1992 if (tg) 1993 tg->gomp_data = data; 1994 } 1995 1996 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASKGROUP_REDUCTION_REGISTER)( 1997 uintptr_t *data) { 1998 int gtid = __kmp_entry_gtid(); 1999 KA_TRACE(20, ("GOMP_taskgroup_reduction_register: T#%d\n", gtid)); 2000 kmp_info_t *thread = __kmp_threads[gtid]; 2001 kmp_taskgroup_t *tg = thread->th.th_current_task->td_taskgroup; 2002 int nthreads = thread->th.th_team_nproc; 2003 __kmp_GOMP_taskgroup_reduction_register(data, tg, nthreads); 2004 } 2005 2006 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASKGROUP_REDUCTION_UNREGISTER)( 2007 uintptr_t *data) { 2008 KA_TRACE(20, 2009 ("GOMP_taskgroup_reduction_unregister: T#%d\n", __kmp_get_gtid())); 2010 KMP_ASSERT(data && data[2]); 2011 __kmp_free((void *)data[2]); 2012 } 2013 2014 // Search through reduction data and set ptrs[] elements 2015 // to proper privatized copy address 2016 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASK_REDUCTION_REMAP)(size_t cnt, 2017 size_t cntorig, 2018 void **ptrs) { 2019 int gtid = __kmp_entry_gtid(); 2020 KA_TRACE(20, ("GOMP_task_reduction_remap: T#%d\n", gtid)); 2021 kmp_info_t *thread = __kmp_threads[gtid]; 2022 kmp_int32 tid = __kmp_get_tid(); 2023 for (size_t i = 0; i < cnt; ++i) { 2024 uintptr_t address = (uintptr_t)ptrs[i]; 2025 void *propagated_address = NULL; 2026 void *mapped_address = NULL; 2027 // Check taskgroups reduce data 2028 kmp_taskgroup_t *tg = thread->th.th_current_task->td_taskgroup; 2029 while (tg) { 2030 uintptr_t *gomp_data = tg->gomp_data; 2031 if (!gomp_data) { 2032 tg = tg->parent; 2033 continue; 2034 } 2035 // Check the shared addresses list 2036 size_t num_vars = (size_t)gomp_data[0]; 2037 uintptr_t per_thread_size = gomp_data[1]; 2038 uintptr_t reduce_data = gomp_data[2]; 2039 uintptr_t end_reduce_data = gomp_data[6]; 2040 for (size_t j = 0; j < num_vars; ++j) { 2041 uintptr_t *entry = gomp_data + 7 + 3 * j; 2042 if (entry[0] == address) { 2043 uintptr_t offset = entry[1]; 2044 mapped_address = 2045 (void *)(reduce_data + tid * per_thread_size + offset); 2046 if (i < cntorig) 2047 propagated_address = (void *)entry[0]; 2048 break; 2049 } 2050 } 2051 if (mapped_address) 2052 break; 2053 // Check if address is within privatized copies range 2054 if (!mapped_address && address >= reduce_data && 2055 address < end_reduce_data) { 2056 uintptr_t offset = (address - reduce_data) % per_thread_size; 2057 mapped_address = (void *)(reduce_data + tid * per_thread_size + offset); 2058 if (i < cntorig) { 2059 for (size_t j = 0; j < num_vars; ++j) { 2060 uintptr_t *entry = gomp_data + 7 + 3 * j; 2061 if (entry[1] == offset) { 2062 propagated_address = (void *)entry[0]; 2063 break; 2064 } 2065 } 2066 } 2067 } 2068 if (mapped_address) 2069 break; 2070 tg = tg->parent; 2071 } 2072 KMP_ASSERT(mapped_address); 2073 ptrs[i] = mapped_address; 2074 if (i < cntorig) { 2075 KMP_ASSERT(propagated_address); 2076 ptrs[cnt + i] = propagated_address; 2077 } 2078 } 2079 } 2080 2081 static void __kmp_GOMP_init_reductions(int gtid, uintptr_t *data, int is_ws) { 2082 kmp_info_t *thr = __kmp_threads[gtid]; 2083 kmp_team_t *team = thr->th.th_team; 2084 // First start a taskgroup 2085 __kmpc_taskgroup(NULL, gtid); 2086 // Then setup reduction data 2087 void *reduce_data = KMP_ATOMIC_LD_RLX(&team->t.t_tg_reduce_data[is_ws]); 2088 if (reduce_data == NULL && 2089 __kmp_atomic_compare_store(&team->t.t_tg_reduce_data[is_ws], reduce_data, 2090 (void *)1)) { 2091 // Single thread enters this block to initialize common reduction data 2092 KMP_DEBUG_ASSERT(reduce_data == NULL); 2093 __kmp_GOMP_taskgroup_reduction_register(data, NULL, thr->th.th_team_nproc); 2094 KMP_ATOMIC_ST_REL(&team->t.t_tg_fini_counter[is_ws], 0); 2095 KMP_ATOMIC_ST_REL(&team->t.t_tg_reduce_data[is_ws], (void *)data); 2096 } else { 2097 // Wait for task reduction initialization 2098 while ((reduce_data = KMP_ATOMIC_LD_ACQ( 2099 &team->t.t_tg_reduce_data[is_ws])) == (void *)1) { 2100 KMP_CPU_PAUSE(); 2101 } 2102 KMP_DEBUG_ASSERT(reduce_data > (void *)1); // should be valid pointer here 2103 } 2104 // For worksharing constructs, each thread has its own reduction structure. 2105 // Have each reduction structure point to same privatized copies of vars. 2106 // For parallel, each thread points to same reduction structure and privatized 2107 // copies of vars 2108 if (is_ws) { 2109 __kmp_GOMP_taskgroup_reduction_register( 2110 data, NULL, thr->th.th_team_nproc, 2111 (uintptr_t *)KMP_ATOMIC_LD_ACQ(&team->t.t_tg_reduce_data[is_ws])); 2112 } 2113 kmp_taskgroup_t *tg = thr->th.th_current_task->td_taskgroup; 2114 tg->gomp_data = data; 2115 } 2116 2117 static unsigned 2118 __kmp_GOMP_par_reductions_microtask_wrapper(int *gtid, int *npr, 2119 void (*task)(void *), void *data) { 2120 kmp_info_t *thr = __kmp_threads[*gtid]; 2121 kmp_team_t *team = thr->th.th_team; 2122 uintptr_t *reduce_data = *(uintptr_t **)data; 2123 __kmp_GOMP_init_reductions(*gtid, reduce_data, 0); 2124 2125 #if OMPT_SUPPORT 2126 ompt_frame_t *ompt_frame; 2127 ompt_state_t enclosing_state; 2128 2129 if (ompt_enabled.enabled) { 2130 // save enclosing task state; set current state for task 2131 enclosing_state = thr->th.ompt_thread_info.state; 2132 thr->th.ompt_thread_info.state = ompt_state_work_parallel; 2133 2134 // set task frame 2135 __ompt_get_task_info_internal(0, NULL, NULL, &ompt_frame, NULL, NULL); 2136 ompt_frame->exit_frame.ptr = OMPT_GET_FRAME_ADDRESS(0); 2137 } 2138 #endif 2139 2140 task(data); 2141 2142 #if OMPT_SUPPORT 2143 if (ompt_enabled.enabled) { 2144 // clear task frame 2145 ompt_frame->exit_frame = ompt_data_none; 2146 2147 // restore enclosing state 2148 thr->th.ompt_thread_info.state = enclosing_state; 2149 } 2150 #endif 2151 __kmpc_end_taskgroup(NULL, *gtid); 2152 // if last thread out, then reset the team's reduce data 2153 // the GOMP_taskgroup_reduction_unregister() function will deallocate 2154 // private copies after reduction calculations take place. 2155 int count = KMP_ATOMIC_INC(&team->t.t_tg_fini_counter[0]); 2156 if (count == thr->th.th_team_nproc - 1) { 2157 KMP_ATOMIC_ST_REL(&team->t.t_tg_reduce_data[0], NULL); 2158 KMP_ATOMIC_ST_REL(&team->t.t_tg_fini_counter[0], 0); 2159 } 2160 return (unsigned)thr->th.th_team_nproc; 2161 } 2162 2163 unsigned KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_REDUCTIONS)( 2164 void (*task)(void *), void *data, unsigned num_threads, 2165 unsigned int flags) { 2166 MKLOC(loc, "GOMP_parallel_reductions"); 2167 int gtid = __kmp_entry_gtid(); 2168 KA_TRACE(20, ("GOMP_parallel_reductions: T#%d\n", gtid)); 2169 __kmp_GOMP_fork_call(&loc, gtid, num_threads, flags, task, 2170 (microtask_t)__kmp_GOMP_par_reductions_microtask_wrapper, 2171 2, task, data); 2172 unsigned retval = 2173 __kmp_GOMP_par_reductions_microtask_wrapper(>id, NULL, task, data); 2174 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_PARALLEL_END)(); 2175 KA_TRACE(20, ("GOMP_parallel_reductions exit: T#%d\n", gtid)); 2176 return retval; 2177 } 2178 2179 bool KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_START)( 2180 long start, long end, long incr, long sched, long chunk_size, long *istart, 2181 long *iend, uintptr_t *reductions, void **mem) { 2182 int status = 0; 2183 int gtid = __kmp_entry_gtid(); 2184 KA_TRACE(20, ("GOMP_loop_start: T#%d, reductions: %p\n", gtid, reductions)); 2185 if (reductions) 2186 __kmp_GOMP_init_reductions(gtid, reductions, 1); 2187 if (mem) 2188 KMP_FATAL(GompFeatureNotSupported, "scan"); 2189 if (istart == NULL) 2190 return true; 2191 const long MONOTONIC_FLAG = (long)(kmp_sched_monotonic); 2192 long monotonic = sched & MONOTONIC_FLAG; 2193 sched &= ~MONOTONIC_FLAG; 2194 if (sched == 0) { 2195 if (monotonic) 2196 status = KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_RUNTIME_START)( 2197 start, end, incr, istart, iend); 2198 else 2199 status = KMP_EXPAND_NAME( 2200 KMP_API_NAME_GOMP_LOOP_MAYBE_NONMONOTONIC_RUNTIME_START)( 2201 start, end, incr, istart, iend); 2202 } else if (sched == 1) { 2203 status = KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_STATIC_START)( 2204 start, end, incr, chunk_size, istart, iend); 2205 } else if (sched == 2) { 2206 if (monotonic) 2207 status = KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_DYNAMIC_START)( 2208 start, end, incr, chunk_size, istart, iend); 2209 else 2210 status = 2211 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_NONMONOTONIC_DYNAMIC_START)( 2212 start, end, incr, chunk_size, istart, iend); 2213 } else if (sched == 3) { 2214 if (monotonic) 2215 status = KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_GUIDED_START)( 2216 start, end, incr, chunk_size, istart, iend); 2217 else 2218 status = 2219 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_NONMONOTONIC_GUIDED_START)( 2220 start, end, incr, chunk_size, istart, iend); 2221 } else if (sched == 4) { 2222 status = KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_NONMONOTONIC_RUNTIME_START)( 2223 start, end, incr, istart, iend); 2224 } else { 2225 KMP_ASSERT(0); 2226 } 2227 return status; 2228 } 2229 2230 bool KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_START)( 2231 bool up, unsigned long long start, unsigned long long end, 2232 unsigned long long incr, long sched, unsigned long long chunk_size, 2233 unsigned long long *istart, unsigned long long *iend, uintptr_t *reductions, 2234 void **mem) { 2235 int status = 0; 2236 int gtid = __kmp_entry_gtid(); 2237 KA_TRACE(20, 2238 ("GOMP_loop_ull_start: T#%d, reductions: %p\n", gtid, reductions)); 2239 if (reductions) 2240 __kmp_GOMP_init_reductions(gtid, reductions, 1); 2241 if (mem) 2242 KMP_FATAL(GompFeatureNotSupported, "scan"); 2243 if (istart == NULL) 2244 return true; 2245 const long MONOTONIC_FLAG = (long)(kmp_sched_monotonic); 2246 long monotonic = sched & MONOTONIC_FLAG; 2247 sched &= ~MONOTONIC_FLAG; 2248 if (sched == 0) { 2249 if (monotonic) 2250 status = KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_RUNTIME_START)( 2251 up, start, end, incr, istart, iend); 2252 else 2253 status = KMP_EXPAND_NAME( 2254 KMP_API_NAME_GOMP_LOOP_ULL_MAYBE_NONMONOTONIC_RUNTIME_START)( 2255 up, start, end, incr, istart, iend); 2256 } else if (sched == 1) { 2257 status = KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_STATIC_START)( 2258 up, start, end, incr, chunk_size, istart, iend); 2259 } else if (sched == 2) { 2260 if (monotonic) 2261 status = KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_DYNAMIC_START)( 2262 up, start, end, incr, chunk_size, istart, iend); 2263 else 2264 status = KMP_EXPAND_NAME( 2265 KMP_API_NAME_GOMP_LOOP_ULL_NONMONOTONIC_DYNAMIC_START)( 2266 up, start, end, incr, chunk_size, istart, iend); 2267 } else if (sched == 3) { 2268 if (monotonic) 2269 status = KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_GUIDED_START)( 2270 up, start, end, incr, chunk_size, istart, iend); 2271 else 2272 status = 2273 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_NONMONOTONIC_GUIDED_START)( 2274 up, start, end, incr, chunk_size, istart, iend); 2275 } else if (sched == 4) { 2276 status = 2277 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_NONMONOTONIC_RUNTIME_START)( 2278 up, start, end, incr, istart, iend); 2279 } else { 2280 KMP_ASSERT(0); 2281 } 2282 return status; 2283 } 2284 2285 bool KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_DOACROSS_START)( 2286 unsigned ncounts, long *counts, long sched, long chunk_size, long *istart, 2287 long *iend, uintptr_t *reductions, void **mem) { 2288 int status = 0; 2289 int gtid = __kmp_entry_gtid(); 2290 KA_TRACE(20, ("GOMP_loop_doacross_start: T#%d, reductions: %p\n", gtid, 2291 reductions)); 2292 if (reductions) 2293 __kmp_GOMP_init_reductions(gtid, reductions, 1); 2294 if (mem) 2295 KMP_FATAL(GompFeatureNotSupported, "scan"); 2296 if (istart == NULL) 2297 return true; 2298 // Ignore any monotonic flag 2299 const long MONOTONIC_FLAG = (long)(kmp_sched_monotonic); 2300 sched &= ~MONOTONIC_FLAG; 2301 if (sched == 0) { 2302 status = KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_DOACROSS_RUNTIME_START)( 2303 ncounts, counts, istart, iend); 2304 } else if (sched == 1) { 2305 status = KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_DOACROSS_STATIC_START)( 2306 ncounts, counts, chunk_size, istart, iend); 2307 } else if (sched == 2) { 2308 status = KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_DOACROSS_DYNAMIC_START)( 2309 ncounts, counts, chunk_size, istart, iend); 2310 } else if (sched == 3) { 2311 status = KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_DOACROSS_GUIDED_START)( 2312 ncounts, counts, chunk_size, istart, iend); 2313 } else { 2314 KMP_ASSERT(0); 2315 } 2316 return status; 2317 } 2318 2319 bool KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_DOACROSS_START)( 2320 unsigned ncounts, unsigned long long *counts, long sched, 2321 unsigned long long chunk_size, unsigned long long *istart, 2322 unsigned long long *iend, uintptr_t *reductions, void **mem) { 2323 int status = 0; 2324 int gtid = __kmp_entry_gtid(); 2325 KA_TRACE(20, ("GOMP_loop_ull_doacross_start: T#%d, reductions: %p\n", gtid, 2326 reductions)); 2327 if (reductions) 2328 __kmp_GOMP_init_reductions(gtid, reductions, 1); 2329 if (mem) 2330 KMP_FATAL(GompFeatureNotSupported, "scan"); 2331 if (istart == NULL) 2332 return true; 2333 // Ignore any monotonic flag 2334 const long MONOTONIC_FLAG = (long)(kmp_sched_monotonic); 2335 sched &= ~MONOTONIC_FLAG; 2336 if (sched == 0) { 2337 status = KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_DOACROSS_RUNTIME_START)( 2338 ncounts, counts, istart, iend); 2339 } else if (sched == 1) { 2340 status = KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_DOACROSS_STATIC_START)( 2341 ncounts, counts, chunk_size, istart, iend); 2342 } else if (sched == 2) { 2343 status = KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_DOACROSS_DYNAMIC_START)( 2344 ncounts, counts, chunk_size, istart, iend); 2345 } else if (sched == 3) { 2346 status = KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_DOACROSS_GUIDED_START)( 2347 ncounts, counts, chunk_size, istart, iend); 2348 } else { 2349 KMP_ASSERT(0); 2350 } 2351 return status; 2352 } 2353 2354 bool KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ORDERED_START)( 2355 long start, long end, long incr, long sched, long chunk_size, long *istart, 2356 long *iend, uintptr_t *reductions, void **mem) { 2357 int status = 0; 2358 int gtid = __kmp_entry_gtid(); 2359 KA_TRACE(20, ("GOMP_loop_ordered_start: T#%d, reductions: %p\n", gtid, 2360 reductions)); 2361 if (reductions) 2362 __kmp_GOMP_init_reductions(gtid, reductions, 1); 2363 if (mem) 2364 KMP_FATAL(GompFeatureNotSupported, "scan"); 2365 if (istart == NULL) 2366 return true; 2367 // Ignore any monotonic flag 2368 const long MONOTONIC_FLAG = (long)(kmp_sched_monotonic); 2369 sched &= ~MONOTONIC_FLAG; 2370 if (sched == 0) { 2371 status = KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ORDERED_RUNTIME_START)( 2372 start, end, incr, istart, iend); 2373 } else if (sched == 1) { 2374 status = KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ORDERED_STATIC_START)( 2375 start, end, incr, chunk_size, istart, iend); 2376 } else if (sched == 2) { 2377 status = KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ORDERED_DYNAMIC_START)( 2378 start, end, incr, chunk_size, istart, iend); 2379 } else if (sched == 3) { 2380 status = KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ORDERED_GUIDED_START)( 2381 start, end, incr, chunk_size, istart, iend); 2382 } else { 2383 KMP_ASSERT(0); 2384 } 2385 return status; 2386 } 2387 2388 bool KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_START)( 2389 bool up, unsigned long long start, unsigned long long end, 2390 unsigned long long incr, long sched, unsigned long long chunk_size, 2391 unsigned long long *istart, unsigned long long *iend, uintptr_t *reductions, 2392 void **mem) { 2393 int status = 0; 2394 int gtid = __kmp_entry_gtid(); 2395 KA_TRACE(20, ("GOMP_loop_ull_ordered_start: T#%d, reductions: %p\n", gtid, 2396 reductions)); 2397 if (reductions) 2398 __kmp_GOMP_init_reductions(gtid, reductions, 1); 2399 if (mem) 2400 KMP_FATAL(GompFeatureNotSupported, "scan"); 2401 if (istart == NULL) 2402 return true; 2403 // Ignore any monotonic flag 2404 const long MONOTONIC_FLAG = (long)(kmp_sched_monotonic); 2405 sched &= ~MONOTONIC_FLAG; 2406 if (sched == 0) { 2407 status = KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_RUNTIME_START)( 2408 up, start, end, incr, istart, iend); 2409 } else if (sched == 1) { 2410 status = KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_STATIC_START)( 2411 up, start, end, incr, chunk_size, istart, iend); 2412 } else if (sched == 2) { 2413 status = KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_DYNAMIC_START)( 2414 up, start, end, incr, chunk_size, istart, iend); 2415 } else if (sched == 3) { 2416 status = KMP_EXPAND_NAME(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_GUIDED_START)( 2417 up, start, end, incr, chunk_size, istart, iend); 2418 } else { 2419 KMP_ASSERT(0); 2420 } 2421 return status; 2422 } 2423 2424 unsigned KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SECTIONS2_START)( 2425 unsigned count, uintptr_t *reductions, void **mem) { 2426 int gtid = __kmp_entry_gtid(); 2427 KA_TRACE(20, 2428 ("GOMP_sections2_start: T#%d, reductions: %p\n", gtid, reductions)); 2429 if (reductions) 2430 __kmp_GOMP_init_reductions(gtid, reductions, 1); 2431 if (mem) 2432 KMP_FATAL(GompFeatureNotSupported, "scan"); 2433 return KMP_EXPAND_NAME(KMP_API_NAME_GOMP_SECTIONS_START)(count); 2434 } 2435 2436 void KMP_EXPAND_NAME(KMP_API_NAME_GOMP_WORKSHARE_TASK_REDUCTION_UNREGISTER)( 2437 bool cancelled) { 2438 int gtid = __kmp_get_gtid(); 2439 MKLOC(loc, "GOMP_workshare_task_reduction_unregister"); 2440 KA_TRACE(20, ("GOMP_workshare_task_reduction_unregister: T#%d\n", gtid)); 2441 kmp_info_t *thr = __kmp_threads[gtid]; 2442 kmp_team_t *team = thr->th.th_team; 2443 __kmpc_end_taskgroup(NULL, gtid); 2444 // If last thread out of workshare, then reset the team's reduce data 2445 // the GOMP_taskgroup_reduction_unregister() function will deallocate 2446 // private copies after reduction calculations take place. 2447 int count = KMP_ATOMIC_INC(&team->t.t_tg_fini_counter[1]); 2448 if (count == thr->th.th_team_nproc - 1) { 2449 KMP_EXPAND_NAME(KMP_API_NAME_GOMP_TASKGROUP_REDUCTION_UNREGISTER) 2450 ((uintptr_t *)KMP_ATOMIC_LD_RLX(&team->t.t_tg_reduce_data[1])); 2451 KMP_ATOMIC_ST_REL(&team->t.t_tg_reduce_data[1], NULL); 2452 KMP_ATOMIC_ST_REL(&team->t.t_tg_fini_counter[1], 0); 2453 } 2454 if (!cancelled) { 2455 __kmpc_barrier(&loc, gtid); 2456 } 2457 } 2458 2459 /* The following sections of code create aliases for the GOMP_* functions, then 2460 create versioned symbols using the assembler directive .symver. This is only 2461 pertinent for ELF .so library. The KMP_VERSION_SYMBOL macro is defined in 2462 kmp_os.h */ 2463 2464 #ifdef KMP_USE_VERSION_SYMBOLS 2465 // GOMP_1.0 versioned symbols 2466 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_ATOMIC_END, 10, "GOMP_1.0"); 2467 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_ATOMIC_START, 10, "GOMP_1.0"); 2468 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_BARRIER, 10, "GOMP_1.0"); 2469 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_CRITICAL_END, 10, "GOMP_1.0"); 2470 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_CRITICAL_NAME_END, 10, "GOMP_1.0"); 2471 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_CRITICAL_NAME_START, 10, "GOMP_1.0"); 2472 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_CRITICAL_START, 10, "GOMP_1.0"); 2473 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_DYNAMIC_NEXT, 10, "GOMP_1.0"); 2474 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_DYNAMIC_START, 10, "GOMP_1.0"); 2475 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_END, 10, "GOMP_1.0"); 2476 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_END_NOWAIT, 10, "GOMP_1.0"); 2477 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_GUIDED_NEXT, 10, "GOMP_1.0"); 2478 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_GUIDED_START, 10, "GOMP_1.0"); 2479 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ORDERED_DYNAMIC_NEXT, 10, "GOMP_1.0"); 2480 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ORDERED_DYNAMIC_START, 10, 2481 "GOMP_1.0"); 2482 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ORDERED_GUIDED_NEXT, 10, "GOMP_1.0"); 2483 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ORDERED_GUIDED_START, 10, "GOMP_1.0"); 2484 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ORDERED_RUNTIME_NEXT, 10, "GOMP_1.0"); 2485 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ORDERED_RUNTIME_START, 10, 2486 "GOMP_1.0"); 2487 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ORDERED_STATIC_NEXT, 10, "GOMP_1.0"); 2488 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ORDERED_STATIC_START, 10, "GOMP_1.0"); 2489 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_RUNTIME_NEXT, 10, "GOMP_1.0"); 2490 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_RUNTIME_START, 10, "GOMP_1.0"); 2491 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_STATIC_NEXT, 10, "GOMP_1.0"); 2492 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_STATIC_START, 10, "GOMP_1.0"); 2493 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_ORDERED_END, 10, "GOMP_1.0"); 2494 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_ORDERED_START, 10, "GOMP_1.0"); 2495 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_END, 10, "GOMP_1.0"); 2496 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_LOOP_DYNAMIC_START, 10, 2497 "GOMP_1.0"); 2498 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_LOOP_GUIDED_START, 10, 2499 "GOMP_1.0"); 2500 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_LOOP_RUNTIME_START, 10, 2501 "GOMP_1.0"); 2502 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_LOOP_STATIC_START, 10, 2503 "GOMP_1.0"); 2504 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_SECTIONS_START, 10, "GOMP_1.0"); 2505 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_START, 10, "GOMP_1.0"); 2506 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_SECTIONS_END, 10, "GOMP_1.0"); 2507 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_SECTIONS_END_NOWAIT, 10, "GOMP_1.0"); 2508 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_SECTIONS_NEXT, 10, "GOMP_1.0"); 2509 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_SECTIONS_START, 10, "GOMP_1.0"); 2510 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_SINGLE_COPY_END, 10, "GOMP_1.0"); 2511 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_SINGLE_COPY_START, 10, "GOMP_1.0"); 2512 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_SINGLE_START, 10, "GOMP_1.0"); 2513 2514 // GOMP_2.0 versioned symbols 2515 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TASK, 20, "GOMP_2.0"); 2516 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TASKWAIT, 20, "GOMP_2.0"); 2517 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_DYNAMIC_NEXT, 20, "GOMP_2.0"); 2518 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_DYNAMIC_START, 20, "GOMP_2.0"); 2519 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_GUIDED_NEXT, 20, "GOMP_2.0"); 2520 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_GUIDED_START, 20, "GOMP_2.0"); 2521 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_DYNAMIC_NEXT, 20, 2522 "GOMP_2.0"); 2523 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_DYNAMIC_START, 20, 2524 "GOMP_2.0"); 2525 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_GUIDED_NEXT, 20, 2526 "GOMP_2.0"); 2527 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_GUIDED_START, 20, 2528 "GOMP_2.0"); 2529 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_RUNTIME_NEXT, 20, 2530 "GOMP_2.0"); 2531 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_RUNTIME_START, 20, 2532 "GOMP_2.0"); 2533 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_STATIC_NEXT, 20, 2534 "GOMP_2.0"); 2535 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_STATIC_START, 20, 2536 "GOMP_2.0"); 2537 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_RUNTIME_NEXT, 20, "GOMP_2.0"); 2538 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_RUNTIME_START, 20, "GOMP_2.0"); 2539 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_STATIC_NEXT, 20, "GOMP_2.0"); 2540 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_STATIC_START, 20, "GOMP_2.0"); 2541 2542 // GOMP_3.0 versioned symbols 2543 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TASKYIELD, 30, "GOMP_3.0"); 2544 2545 // GOMP_4.0 versioned symbols 2546 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL, 40, "GOMP_4.0"); 2547 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_SECTIONS, 40, "GOMP_4.0"); 2548 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_LOOP_DYNAMIC, 40, "GOMP_4.0"); 2549 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_LOOP_GUIDED, 40, "GOMP_4.0"); 2550 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_LOOP_RUNTIME, 40, "GOMP_4.0"); 2551 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_LOOP_STATIC, 40, "GOMP_4.0"); 2552 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TASKGROUP_START, 40, "GOMP_4.0"); 2553 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TASKGROUP_END, 40, "GOMP_4.0"); 2554 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_BARRIER_CANCEL, 40, "GOMP_4.0"); 2555 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_CANCEL, 40, "GOMP_4.0"); 2556 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_CANCELLATION_POINT, 40, "GOMP_4.0"); 2557 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_END_CANCEL, 40, "GOMP_4.0"); 2558 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_SECTIONS_END_CANCEL, 40, "GOMP_4.0"); 2559 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TARGET, 40, "GOMP_4.0"); 2560 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TARGET_DATA, 40, "GOMP_4.0"); 2561 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TARGET_END_DATA, 40, "GOMP_4.0"); 2562 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TARGET_UPDATE, 40, "GOMP_4.0"); 2563 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TEAMS, 40, "GOMP_4.0"); 2564 2565 // GOMP_4.5 versioned symbols 2566 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TASKLOOP, 45, "GOMP_4.5"); 2567 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TASKLOOP_ULL, 45, "GOMP_4.5"); 2568 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_DOACROSS_POST, 45, "GOMP_4.5"); 2569 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_DOACROSS_WAIT, 45, "GOMP_4.5"); 2570 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_DOACROSS_STATIC_START, 45, 2571 "GOMP_4.5"); 2572 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_DOACROSS_DYNAMIC_START, 45, 2573 "GOMP_4.5"); 2574 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_DOACROSS_GUIDED_START, 45, 2575 "GOMP_4.5"); 2576 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_DOACROSS_RUNTIME_START, 45, 2577 "GOMP_4.5"); 2578 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_DOACROSS_ULL_POST, 45, "GOMP_4.5"); 2579 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_DOACROSS_ULL_WAIT, 45, "GOMP_4.5"); 2580 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_DOACROSS_STATIC_START, 45, 2581 "GOMP_4.5"); 2582 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_DOACROSS_DYNAMIC_START, 45, 2583 "GOMP_4.5"); 2584 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_DOACROSS_GUIDED_START, 45, 2585 "GOMP_4.5"); 2586 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_DOACROSS_RUNTIME_START, 45, 2587 "GOMP_4.5"); 2588 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_NONMONOTONIC_DYNAMIC_START, 45, 2589 "GOMP_4.5"); 2590 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_NONMONOTONIC_DYNAMIC_NEXT, 45, 2591 "GOMP_4.5"); 2592 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_NONMONOTONIC_GUIDED_START, 45, 2593 "GOMP_4.5"); 2594 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_NONMONOTONIC_GUIDED_NEXT, 45, 2595 "GOMP_4.5"); 2596 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_NONMONOTONIC_DYNAMIC_START, 45, 2597 "GOMP_4.5"); 2598 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_NONMONOTONIC_DYNAMIC_NEXT, 45, 2599 "GOMP_4.5"); 2600 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_NONMONOTONIC_GUIDED_START, 45, 2601 "GOMP_4.5"); 2602 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_NONMONOTONIC_GUIDED_NEXT, 45, 2603 "GOMP_4.5"); 2604 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_LOOP_NONMONOTONIC_DYNAMIC, 45, 2605 "GOMP_4.5"); 2606 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_LOOP_NONMONOTONIC_GUIDED, 45, 2607 "GOMP_4.5"); 2608 2609 // GOMP_5.0 versioned symbols 2610 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_MAYBE_NONMONOTONIC_RUNTIME_NEXT, 50, 2611 "GOMP_5.0"); 2612 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_MAYBE_NONMONOTONIC_RUNTIME_START, 50, 2613 "GOMP_5.0"); 2614 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_NONMONOTONIC_RUNTIME_NEXT, 50, 2615 "GOMP_5.0"); 2616 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_NONMONOTONIC_RUNTIME_START, 50, 2617 "GOMP_5.0"); 2618 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_MAYBE_NONMONOTONIC_RUNTIME_NEXT, 2619 50, "GOMP_5.0"); 2620 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_MAYBE_NONMONOTONIC_RUNTIME_START, 2621 50, "GOMP_5.0"); 2622 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_NONMONOTONIC_RUNTIME_NEXT, 50, 2623 "GOMP_5.0"); 2624 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_NONMONOTONIC_RUNTIME_START, 50, 2625 "GOMP_5.0"); 2626 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_LOOP_NONMONOTONIC_RUNTIME, 50, 2627 "GOMP_5.0"); 2628 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_LOOP_MAYBE_NONMONOTONIC_RUNTIME, 2629 50, "GOMP_5.0"); 2630 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TEAMS_REG, 50, "GOMP_5.0"); 2631 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TASKWAIT_DEPEND, 50, "GOMP_5.0"); 2632 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TASKGROUP_REDUCTION_REGISTER, 50, 2633 "GOMP_5.0"); 2634 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TASKGROUP_REDUCTION_UNREGISTER, 50, 2635 "GOMP_5.0"); 2636 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_TASK_REDUCTION_REMAP, 50, "GOMP_5.0"); 2637 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_PARALLEL_REDUCTIONS, 50, "GOMP_5.0"); 2638 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_START, 50, "GOMP_5.0"); 2639 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_START, 50, "GOMP_5.0"); 2640 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_DOACROSS_START, 50, "GOMP_5.0"); 2641 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_DOACROSS_START, 50, "GOMP_5.0"); 2642 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ORDERED_START, 50, "GOMP_5.0"); 2643 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_LOOP_ULL_ORDERED_START, 50, "GOMP_5.0"); 2644 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_SECTIONS2_START, 50, "GOMP_5.0"); 2645 KMP_VERSION_SYMBOL(KMP_API_NAME_GOMP_WORKSHARE_TASK_REDUCTION_UNREGISTER, 50, 2646 "GOMP_5.0"); 2647 #endif // KMP_USE_VERSION_SYMBOLS 2648 2649 #ifdef __cplusplus 2650 } // extern "C" 2651 #endif // __cplusplus 2652