1 /* 2 * kmp_runtime.cpp -- KPTS runtime support library 3 */ 4 5 6 //===----------------------------------------------------------------------===// 7 // 8 // The LLVM Compiler Infrastructure 9 // 10 // This file is dual licensed under the MIT and the University of Illinois Open 11 // Source Licenses. See LICENSE.txt for details. 12 // 13 //===----------------------------------------------------------------------===// 14 15 16 #include "kmp.h" 17 #include "kmp_affinity.h" 18 #include "kmp_atomic.h" 19 #include "kmp_environment.h" 20 #include "kmp_error.h" 21 #include "kmp_i18n.h" 22 #include "kmp_io.h" 23 #include "kmp_itt.h" 24 #include "kmp_settings.h" 25 #include "kmp_stats.h" 26 #include "kmp_str.h" 27 #include "kmp_wait_release.h" 28 #include "kmp_wrapper_getpid.h" 29 30 #if OMPT_SUPPORT 31 #include "ompt-specific.h" 32 #endif 33 34 /* these are temporary issues to be dealt with */ 35 #define KMP_USE_PRCTL 0 36 37 #if KMP_OS_WINDOWS 38 #include <process.h> 39 #endif 40 41 #include "tsan_annotations.h" 42 43 #if defined(KMP_GOMP_COMPAT) 44 char const __kmp_version_alt_comp[] = 45 KMP_VERSION_PREFIX "alternative compiler support: yes"; 46 #endif /* defined(KMP_GOMP_COMPAT) */ 47 48 char const __kmp_version_omp_api[] = KMP_VERSION_PREFIX "API version: " 49 #if OMP_50_ENABLED 50 "5.0 (201611)"; 51 #elif OMP_45_ENABLED 52 "4.5 (201511)"; 53 #elif OMP_40_ENABLED 54 "4.0 (201307)"; 55 #else 56 "3.1 (201107)"; 57 #endif 58 59 #ifdef KMP_DEBUG 60 char const __kmp_version_lock[] = 61 KMP_VERSION_PREFIX "lock type: run time selectable"; 62 #endif /* KMP_DEBUG */ 63 64 #define KMP_MIN(x, y) ((x) < (y) ? (x) : (y)) 65 66 /* ------------------------------------------------------------------------ */ 67 68 kmp_info_t __kmp_monitor; 69 70 /* Forward declarations */ 71 72 void __kmp_cleanup(void); 73 74 static void __kmp_initialize_info(kmp_info_t *, kmp_team_t *, int tid, 75 int gtid); 76 static void __kmp_initialize_team(kmp_team_t *team, int new_nproc, 77 kmp_internal_control_t *new_icvs, 78 ident_t *loc); 79 #if OMP_40_ENABLED && KMP_AFFINITY_SUPPORTED 80 static void __kmp_partition_places(kmp_team_t *team, 81 int update_master_only = 0); 82 #endif 83 static void __kmp_do_serial_initialize(void); 84 void __kmp_fork_barrier(int gtid, int tid); 85 void __kmp_join_barrier(int gtid); 86 void __kmp_setup_icv_copy(kmp_team_t *team, int new_nproc, 87 kmp_internal_control_t *new_icvs, ident_t *loc); 88 89 #ifdef USE_LOAD_BALANCE 90 static int __kmp_load_balance_nproc(kmp_root_t *root, int set_nproc); 91 #endif 92 93 static int __kmp_expand_threads(int nWish, int nNeed); 94 #if KMP_OS_WINDOWS 95 static int __kmp_unregister_root_other_thread(int gtid); 96 #endif 97 static void __kmp_unregister_library(void); // called by __kmp_internal_end() 98 static void __kmp_reap_thread(kmp_info_t *thread, int is_root); 99 static kmp_info_t *__kmp_thread_pool_insert_pt = NULL; 100 101 /* Calculate the identifier of the current thread */ 102 /* fast (and somewhat portable) way to get unique identifier of executing 103 thread. Returns KMP_GTID_DNE if we haven't been assigned a gtid. */ 104 int __kmp_get_global_thread_id() { 105 int i; 106 kmp_info_t **other_threads; 107 size_t stack_data; 108 char *stack_addr; 109 size_t stack_size; 110 char *stack_base; 111 112 KA_TRACE( 113 1000, 114 ("*** __kmp_get_global_thread_id: entering, nproc=%d all_nproc=%d\n", 115 __kmp_nth, __kmp_all_nth)); 116 117 /* JPH - to handle the case where __kmpc_end(0) is called immediately prior to 118 a parallel region, made it return KMP_GTID_DNE to force serial_initialize 119 by caller. Had to handle KMP_GTID_DNE at all call-sites, or else guarantee 120 __kmp_init_gtid for this to work. */ 121 122 if (!TCR_4(__kmp_init_gtid)) 123 return KMP_GTID_DNE; 124 125 #ifdef KMP_TDATA_GTID 126 if (TCR_4(__kmp_gtid_mode) >= 3) { 127 KA_TRACE(1000, ("*** __kmp_get_global_thread_id: using TDATA\n")); 128 return __kmp_gtid; 129 } 130 #endif 131 if (TCR_4(__kmp_gtid_mode) >= 2) { 132 KA_TRACE(1000, ("*** __kmp_get_global_thread_id: using keyed TLS\n")); 133 return __kmp_gtid_get_specific(); 134 } 135 KA_TRACE(1000, ("*** __kmp_get_global_thread_id: using internal alg.\n")); 136 137 stack_addr = (char *)&stack_data; 138 other_threads = __kmp_threads; 139 140 /* ATT: The code below is a source of potential bugs due to unsynchronized 141 access to __kmp_threads array. For example: 142 1. Current thread loads other_threads[i] to thr and checks it, it is 143 non-NULL. 144 2. Current thread is suspended by OS. 145 3. Another thread unregisters and finishes (debug versions of free() 146 may fill memory with something like 0xEF). 147 4. Current thread is resumed. 148 5. Current thread reads junk from *thr. 149 TODO: Fix it. --ln */ 150 151 for (i = 0; i < __kmp_threads_capacity; i++) { 152 153 kmp_info_t *thr = (kmp_info_t *)TCR_SYNC_PTR(other_threads[i]); 154 if (!thr) 155 continue; 156 157 stack_size = (size_t)TCR_PTR(thr->th.th_info.ds.ds_stacksize); 158 stack_base = (char *)TCR_PTR(thr->th.th_info.ds.ds_stackbase); 159 160 /* stack grows down -- search through all of the active threads */ 161 162 if (stack_addr <= stack_base) { 163 size_t stack_diff = stack_base - stack_addr; 164 165 if (stack_diff <= stack_size) { 166 /* The only way we can be closer than the allocated */ 167 /* stack size is if we are running on this thread. */ 168 KMP_DEBUG_ASSERT(__kmp_gtid_get_specific() == i); 169 return i; 170 } 171 } 172 } 173 174 /* get specific to try and determine our gtid */ 175 KA_TRACE(1000, 176 ("*** __kmp_get_global_thread_id: internal alg. failed to find " 177 "thread, using TLS\n")); 178 i = __kmp_gtid_get_specific(); 179 180 /*fprintf( stderr, "=== %d\n", i ); */ /* GROO */ 181 182 /* if we havn't been assigned a gtid, then return code */ 183 if (i < 0) 184 return i; 185 186 /* dynamically updated stack window for uber threads to avoid get_specific 187 call */ 188 if (!TCR_4(other_threads[i]->th.th_info.ds.ds_stackgrow)) { 189 KMP_FATAL(StackOverflow, i); 190 } 191 192 stack_base = (char *)other_threads[i]->th.th_info.ds.ds_stackbase; 193 if (stack_addr > stack_base) { 194 TCW_PTR(other_threads[i]->th.th_info.ds.ds_stackbase, stack_addr); 195 TCW_PTR(other_threads[i]->th.th_info.ds.ds_stacksize, 196 other_threads[i]->th.th_info.ds.ds_stacksize + stack_addr - 197 stack_base); 198 } else { 199 TCW_PTR(other_threads[i]->th.th_info.ds.ds_stacksize, 200 stack_base - stack_addr); 201 } 202 203 /* Reprint stack bounds for ubermaster since they have been refined */ 204 if (__kmp_storage_map) { 205 char *stack_end = (char *)other_threads[i]->th.th_info.ds.ds_stackbase; 206 char *stack_beg = stack_end - other_threads[i]->th.th_info.ds.ds_stacksize; 207 __kmp_print_storage_map_gtid(i, stack_beg, stack_end, 208 other_threads[i]->th.th_info.ds.ds_stacksize, 209 "th_%d stack (refinement)", i); 210 } 211 return i; 212 } 213 214 int __kmp_get_global_thread_id_reg() { 215 int gtid; 216 217 if (!__kmp_init_serial) { 218 gtid = KMP_GTID_DNE; 219 } else 220 #ifdef KMP_TDATA_GTID 221 if (TCR_4(__kmp_gtid_mode) >= 3) { 222 KA_TRACE(1000, ("*** __kmp_get_global_thread_id_reg: using TDATA\n")); 223 gtid = __kmp_gtid; 224 } else 225 #endif 226 if (TCR_4(__kmp_gtid_mode) >= 2) { 227 KA_TRACE(1000, ("*** __kmp_get_global_thread_id_reg: using keyed TLS\n")); 228 gtid = __kmp_gtid_get_specific(); 229 } else { 230 KA_TRACE(1000, 231 ("*** __kmp_get_global_thread_id_reg: using internal alg.\n")); 232 gtid = __kmp_get_global_thread_id(); 233 } 234 235 /* we must be a new uber master sibling thread */ 236 if (gtid == KMP_GTID_DNE) { 237 KA_TRACE(10, 238 ("__kmp_get_global_thread_id_reg: Encountered new root thread. " 239 "Registering a new gtid.\n")); 240 __kmp_acquire_bootstrap_lock(&__kmp_initz_lock); 241 if (!__kmp_init_serial) { 242 __kmp_do_serial_initialize(); 243 gtid = __kmp_gtid_get_specific(); 244 } else { 245 gtid = __kmp_register_root(FALSE); 246 } 247 __kmp_release_bootstrap_lock(&__kmp_initz_lock); 248 /*__kmp_printf( "+++ %d\n", gtid ); */ /* GROO */ 249 } 250 251 KMP_DEBUG_ASSERT(gtid >= 0); 252 253 return gtid; 254 } 255 256 /* caller must hold forkjoin_lock */ 257 void __kmp_check_stack_overlap(kmp_info_t *th) { 258 int f; 259 char *stack_beg = NULL; 260 char *stack_end = NULL; 261 int gtid; 262 263 KA_TRACE(10, ("__kmp_check_stack_overlap: called\n")); 264 if (__kmp_storage_map) { 265 stack_end = (char *)th->th.th_info.ds.ds_stackbase; 266 stack_beg = stack_end - th->th.th_info.ds.ds_stacksize; 267 268 gtid = __kmp_gtid_from_thread(th); 269 270 if (gtid == KMP_GTID_MONITOR) { 271 __kmp_print_storage_map_gtid( 272 gtid, stack_beg, stack_end, th->th.th_info.ds.ds_stacksize, 273 "th_%s stack (%s)", "mon", 274 (th->th.th_info.ds.ds_stackgrow) ? "initial" : "actual"); 275 } else { 276 __kmp_print_storage_map_gtid( 277 gtid, stack_beg, stack_end, th->th.th_info.ds.ds_stacksize, 278 "th_%d stack (%s)", gtid, 279 (th->th.th_info.ds.ds_stackgrow) ? "initial" : "actual"); 280 } 281 } 282 283 /* No point in checking ubermaster threads since they use refinement and 284 * cannot overlap */ 285 gtid = __kmp_gtid_from_thread(th); 286 if (__kmp_env_checks == TRUE && !KMP_UBER_GTID(gtid)) { 287 KA_TRACE(10, 288 ("__kmp_check_stack_overlap: performing extensive checking\n")); 289 if (stack_beg == NULL) { 290 stack_end = (char *)th->th.th_info.ds.ds_stackbase; 291 stack_beg = stack_end - th->th.th_info.ds.ds_stacksize; 292 } 293 294 for (f = 0; f < __kmp_threads_capacity; f++) { 295 kmp_info_t *f_th = (kmp_info_t *)TCR_SYNC_PTR(__kmp_threads[f]); 296 297 if (f_th && f_th != th) { 298 char *other_stack_end = 299 (char *)TCR_PTR(f_th->th.th_info.ds.ds_stackbase); 300 char *other_stack_beg = 301 other_stack_end - (size_t)TCR_PTR(f_th->th.th_info.ds.ds_stacksize); 302 if ((stack_beg > other_stack_beg && stack_beg < other_stack_end) || 303 (stack_end > other_stack_beg && stack_end < other_stack_end)) { 304 305 /* Print the other stack values before the abort */ 306 if (__kmp_storage_map) 307 __kmp_print_storage_map_gtid( 308 -1, other_stack_beg, other_stack_end, 309 (size_t)TCR_PTR(f_th->th.th_info.ds.ds_stacksize), 310 "th_%d stack (overlapped)", __kmp_gtid_from_thread(f_th)); 311 312 __kmp_fatal(KMP_MSG(StackOverlap), KMP_HNT(ChangeStackLimit), 313 __kmp_msg_null); 314 } 315 } 316 } 317 } 318 KA_TRACE(10, ("__kmp_check_stack_overlap: returning\n")); 319 } 320 321 /* ------------------------------------------------------------------------ */ 322 323 void __kmp_infinite_loop(void) { 324 static int done = FALSE; 325 326 while (!done) { 327 KMP_YIELD(1); 328 } 329 } 330 331 #define MAX_MESSAGE 512 332 333 void __kmp_print_storage_map_gtid(int gtid, void *p1, void *p2, size_t size, 334 char const *format, ...) { 335 char buffer[MAX_MESSAGE]; 336 va_list ap; 337 338 va_start(ap, format); 339 KMP_SNPRINTF(buffer, sizeof(buffer), "OMP storage map: %p %p%8lu %s\n", p1, 340 p2, (unsigned long)size, format); 341 __kmp_acquire_bootstrap_lock(&__kmp_stdio_lock); 342 __kmp_vprintf(kmp_err, buffer, ap); 343 #if KMP_PRINT_DATA_PLACEMENT 344 int node; 345 if (gtid >= 0) { 346 if (p1 <= p2 && (char *)p2 - (char *)p1 == size) { 347 if (__kmp_storage_map_verbose) { 348 node = __kmp_get_host_node(p1); 349 if (node < 0) /* doesn't work, so don't try this next time */ 350 __kmp_storage_map_verbose = FALSE; 351 else { 352 char *last; 353 int lastNode; 354 int localProc = __kmp_get_cpu_from_gtid(gtid); 355 356 const int page_size = KMP_GET_PAGE_SIZE(); 357 358 p1 = (void *)((size_t)p1 & ~((size_t)page_size - 1)); 359 p2 = (void *)(((size_t)p2 - 1) & ~((size_t)page_size - 1)); 360 if (localProc >= 0) 361 __kmp_printf_no_lock(" GTID %d localNode %d\n", gtid, 362 localProc >> 1); 363 else 364 __kmp_printf_no_lock(" GTID %d\n", gtid); 365 #if KMP_USE_PRCTL 366 /* The more elaborate format is disabled for now because of the prctl 367 * hanging bug. */ 368 do { 369 last = p1; 370 lastNode = node; 371 /* This loop collates adjacent pages with the same host node. */ 372 do { 373 (char *)p1 += page_size; 374 } while (p1 <= p2 && (node = __kmp_get_host_node(p1)) == lastNode); 375 __kmp_printf_no_lock(" %p-%p memNode %d\n", last, (char *)p1 - 1, 376 lastNode); 377 } while (p1 <= p2); 378 #else 379 __kmp_printf_no_lock(" %p-%p memNode %d\n", p1, 380 (char *)p1 + (page_size - 1), 381 __kmp_get_host_node(p1)); 382 if (p1 < p2) { 383 __kmp_printf_no_lock(" %p-%p memNode %d\n", p2, 384 (char *)p2 + (page_size - 1), 385 __kmp_get_host_node(p2)); 386 } 387 #endif 388 } 389 } 390 } else 391 __kmp_printf_no_lock(" %s\n", KMP_I18N_STR(StorageMapWarning)); 392 } 393 #endif /* KMP_PRINT_DATA_PLACEMENT */ 394 __kmp_release_bootstrap_lock(&__kmp_stdio_lock); 395 } 396 397 void __kmp_warn(char const *format, ...) { 398 char buffer[MAX_MESSAGE]; 399 va_list ap; 400 401 if (__kmp_generate_warnings == kmp_warnings_off) { 402 return; 403 } 404 405 va_start(ap, format); 406 407 KMP_SNPRINTF(buffer, sizeof(buffer), "OMP warning: %s\n", format); 408 __kmp_acquire_bootstrap_lock(&__kmp_stdio_lock); 409 __kmp_vprintf(kmp_err, buffer, ap); 410 __kmp_release_bootstrap_lock(&__kmp_stdio_lock); 411 412 va_end(ap); 413 } 414 415 void __kmp_abort_process() { 416 // Later threads may stall here, but that's ok because abort() will kill them. 417 __kmp_acquire_bootstrap_lock(&__kmp_exit_lock); 418 419 if (__kmp_debug_buf) { 420 __kmp_dump_debug_buffer(); 421 }; // if 422 423 if (KMP_OS_WINDOWS) { 424 // Let other threads know of abnormal termination and prevent deadlock 425 // if abort happened during library initialization or shutdown 426 __kmp_global.g.g_abort = SIGABRT; 427 428 /* On Windows* OS by default abort() causes pop-up error box, which stalls 429 nightly testing. Unfortunately, we cannot reliably suppress pop-up error 430 boxes. _set_abort_behavior() works well, but this function is not 431 available in VS7 (this is not problem for DLL, but it is a problem for 432 static OpenMP RTL). SetErrorMode (and so, timelimit utility) does not 433 help, at least in some versions of MS C RTL. 434 435 It seems following sequence is the only way to simulate abort() and 436 avoid pop-up error box. */ 437 raise(SIGABRT); 438 _exit(3); // Just in case, if signal ignored, exit anyway. 439 } else { 440 abort(); 441 }; // if 442 443 __kmp_infinite_loop(); 444 __kmp_release_bootstrap_lock(&__kmp_exit_lock); 445 446 } // __kmp_abort_process 447 448 void __kmp_abort_thread(void) { 449 // TODO: Eliminate g_abort global variable and this function. 450 // In case of abort just call abort(), it will kill all the threads. 451 __kmp_infinite_loop(); 452 } // __kmp_abort_thread 453 454 /* Print out the storage map for the major kmp_info_t thread data structures 455 that are allocated together. */ 456 457 static void __kmp_print_thread_storage_map(kmp_info_t *thr, int gtid) { 458 __kmp_print_storage_map_gtid(gtid, thr, thr + 1, sizeof(kmp_info_t), "th_%d", 459 gtid); 460 461 __kmp_print_storage_map_gtid(gtid, &thr->th.th_info, &thr->th.th_team, 462 sizeof(kmp_desc_t), "th_%d.th_info", gtid); 463 464 __kmp_print_storage_map_gtid(gtid, &thr->th.th_local, &thr->th.th_pri_head, 465 sizeof(kmp_local_t), "th_%d.th_local", gtid); 466 467 __kmp_print_storage_map_gtid( 468 gtid, &thr->th.th_bar[0], &thr->th.th_bar[bs_last_barrier], 469 sizeof(kmp_balign_t) * bs_last_barrier, "th_%d.th_bar", gtid); 470 471 __kmp_print_storage_map_gtid(gtid, &thr->th.th_bar[bs_plain_barrier], 472 &thr->th.th_bar[bs_plain_barrier + 1], 473 sizeof(kmp_balign_t), "th_%d.th_bar[plain]", 474 gtid); 475 476 __kmp_print_storage_map_gtid(gtid, &thr->th.th_bar[bs_forkjoin_barrier], 477 &thr->th.th_bar[bs_forkjoin_barrier + 1], 478 sizeof(kmp_balign_t), "th_%d.th_bar[forkjoin]", 479 gtid); 480 481 #if KMP_FAST_REDUCTION_BARRIER 482 __kmp_print_storage_map_gtid(gtid, &thr->th.th_bar[bs_reduction_barrier], 483 &thr->th.th_bar[bs_reduction_barrier + 1], 484 sizeof(kmp_balign_t), "th_%d.th_bar[reduction]", 485 gtid); 486 #endif // KMP_FAST_REDUCTION_BARRIER 487 } 488 489 /* Print out the storage map for the major kmp_team_t team data structures 490 that are allocated together. */ 491 492 static void __kmp_print_team_storage_map(const char *header, kmp_team_t *team, 493 int team_id, int num_thr) { 494 int num_disp_buff = team->t.t_max_nproc > 1 ? __kmp_dispatch_num_buffers : 2; 495 __kmp_print_storage_map_gtid(-1, team, team + 1, sizeof(kmp_team_t), "%s_%d", 496 header, team_id); 497 498 __kmp_print_storage_map_gtid(-1, &team->t.t_bar[0], 499 &team->t.t_bar[bs_last_barrier], 500 sizeof(kmp_balign_team_t) * bs_last_barrier, 501 "%s_%d.t_bar", header, team_id); 502 503 __kmp_print_storage_map_gtid(-1, &team->t.t_bar[bs_plain_barrier], 504 &team->t.t_bar[bs_plain_barrier + 1], 505 sizeof(kmp_balign_team_t), "%s_%d.t_bar[plain]", 506 header, team_id); 507 508 __kmp_print_storage_map_gtid(-1, &team->t.t_bar[bs_forkjoin_barrier], 509 &team->t.t_bar[bs_forkjoin_barrier + 1], 510 sizeof(kmp_balign_team_t), 511 "%s_%d.t_bar[forkjoin]", header, team_id); 512 513 #if KMP_FAST_REDUCTION_BARRIER 514 __kmp_print_storage_map_gtid(-1, &team->t.t_bar[bs_reduction_barrier], 515 &team->t.t_bar[bs_reduction_barrier + 1], 516 sizeof(kmp_balign_team_t), 517 "%s_%d.t_bar[reduction]", header, team_id); 518 #endif // KMP_FAST_REDUCTION_BARRIER 519 520 __kmp_print_storage_map_gtid( 521 -1, &team->t.t_dispatch[0], &team->t.t_dispatch[num_thr], 522 sizeof(kmp_disp_t) * num_thr, "%s_%d.t_dispatch", header, team_id); 523 524 __kmp_print_storage_map_gtid( 525 -1, &team->t.t_threads[0], &team->t.t_threads[num_thr], 526 sizeof(kmp_info_t *) * num_thr, "%s_%d.t_threads", header, team_id); 527 528 __kmp_print_storage_map_gtid(-1, &team->t.t_disp_buffer[0], 529 &team->t.t_disp_buffer[num_disp_buff], 530 sizeof(dispatch_shared_info_t) * num_disp_buff, 531 "%s_%d.t_disp_buffer", header, team_id); 532 533 __kmp_print_storage_map_gtid(-1, &team->t.t_taskq, &team->t.t_copypriv_data, 534 sizeof(kmp_taskq_t), "%s_%d.t_taskq", header, 535 team_id); 536 } 537 538 static void __kmp_init_allocator() {} 539 static void __kmp_fini_allocator() {} 540 541 /* ------------------------------------------------------------------------ */ 542 543 #ifdef KMP_DYNAMIC_LIB 544 #if KMP_OS_WINDOWS 545 546 static void __kmp_reset_lock(kmp_bootstrap_lock_t *lck) { 547 // TODO: Change to __kmp_break_bootstrap_lock(). 548 __kmp_init_bootstrap_lock(lck); // make the lock released 549 } 550 551 static void __kmp_reset_locks_on_process_detach(int gtid_req) { 552 int i; 553 int thread_count; 554 555 // PROCESS_DETACH is expected to be called by a thread that executes 556 // ProcessExit() or FreeLibrary(). OS terminates other threads (except the one 557 // calling ProcessExit or FreeLibrary). So, it might be safe to access the 558 // __kmp_threads[] without taking the forkjoin_lock. However, in fact, some 559 // threads can be still alive here, although being about to be terminated. The 560 // threads in the array with ds_thread==0 are most suspicious. Actually, it 561 // can be not safe to access the __kmp_threads[]. 562 563 // TODO: does it make sense to check __kmp_roots[] ? 564 565 // Let's check that there are no other alive threads registered with the OMP 566 // lib. 567 while (1) { 568 thread_count = 0; 569 for (i = 0; i < __kmp_threads_capacity; ++i) { 570 if (!__kmp_threads) 571 continue; 572 kmp_info_t *th = __kmp_threads[i]; 573 if (th == NULL) 574 continue; 575 int gtid = th->th.th_info.ds.ds_gtid; 576 if (gtid == gtid_req) 577 continue; 578 if (gtid < 0) 579 continue; 580 DWORD exit_val; 581 int alive = __kmp_is_thread_alive(th, &exit_val); 582 if (alive) { 583 ++thread_count; 584 } 585 } 586 if (thread_count == 0) 587 break; // success 588 } 589 590 // Assume that I'm alone. Now it might be safe to check and reset locks. 591 // __kmp_forkjoin_lock and __kmp_stdio_lock are expected to be reset. 592 __kmp_reset_lock(&__kmp_forkjoin_lock); 593 #ifdef KMP_DEBUG 594 __kmp_reset_lock(&__kmp_stdio_lock); 595 #endif // KMP_DEBUG 596 } 597 598 BOOL WINAPI DllMain(HINSTANCE hInstDLL, DWORD fdwReason, LPVOID lpReserved) { 599 //__kmp_acquire_bootstrap_lock( &__kmp_initz_lock ); 600 601 switch (fdwReason) { 602 603 case DLL_PROCESS_ATTACH: 604 KA_TRACE(10, ("DllMain: PROCESS_ATTACH\n")); 605 606 return TRUE; 607 608 case DLL_PROCESS_DETACH: 609 KA_TRACE(10, ("DllMain: PROCESS_DETACH T#%d\n", __kmp_gtid_get_specific())); 610 611 if (lpReserved != NULL) { 612 // lpReserved is used for telling the difference: 613 // lpReserved == NULL when FreeLibrary() was called, 614 // lpReserved != NULL when the process terminates. 615 // When FreeLibrary() is called, worker threads remain alive. So they will 616 // release the forkjoin lock by themselves. When the process terminates, 617 // worker threads disappear triggering the problem of unreleased forkjoin 618 // lock as described below. 619 620 // A worker thread can take the forkjoin lock. The problem comes up if 621 // that worker thread becomes dead before it releases the forkjoin lock. 622 // The forkjoin lock remains taken, while the thread executing 623 // DllMain()->PROCESS_DETACH->__kmp_internal_end_library() below will try 624 // to take the forkjoin lock and will always fail, so that the application 625 // will never finish [normally]. This scenario is possible if 626 // __kmpc_end() has not been executed. It looks like it's not a corner 627 // case, but common cases: 628 // - the main function was compiled by an alternative compiler; 629 // - the main function was compiled by icl but without /Qopenmp 630 // (application with plugins); 631 // - application terminates by calling C exit(), Fortran CALL EXIT() or 632 // Fortran STOP. 633 // - alive foreign thread prevented __kmpc_end from doing cleanup. 634 // 635 // This is a hack to work around the problem. 636 // TODO: !!! figure out something better. 637 __kmp_reset_locks_on_process_detach(__kmp_gtid_get_specific()); 638 } 639 640 __kmp_internal_end_library(__kmp_gtid_get_specific()); 641 642 return TRUE; 643 644 case DLL_THREAD_ATTACH: 645 KA_TRACE(10, ("DllMain: THREAD_ATTACH\n")); 646 647 /* if we want to register new siblings all the time here call 648 * __kmp_get_gtid(); */ 649 return TRUE; 650 651 case DLL_THREAD_DETACH: 652 KA_TRACE(10, ("DllMain: THREAD_DETACH T#%d\n", __kmp_gtid_get_specific())); 653 654 __kmp_internal_end_thread(__kmp_gtid_get_specific()); 655 return TRUE; 656 } 657 658 return TRUE; 659 } 660 661 #endif /* KMP_OS_WINDOWS */ 662 #endif /* KMP_DYNAMIC_LIB */ 663 664 /* Change the library type to "status" and return the old type */ 665 /* called from within initialization routines where __kmp_initz_lock is held */ 666 int __kmp_change_library(int status) { 667 int old_status; 668 669 old_status = __kmp_yield_init & 670 1; // check whether KMP_LIBRARY=throughput (even init count) 671 672 if (status) { 673 __kmp_yield_init |= 1; // throughput => turnaround (odd init count) 674 } else { 675 __kmp_yield_init &= ~1; // turnaround => throughput (even init count) 676 } 677 678 return old_status; // return previous setting of whether 679 // KMP_LIBRARY=throughput 680 } 681 682 /* __kmp_parallel_deo -- Wait until it's our turn. */ 683 void __kmp_parallel_deo(int *gtid_ref, int *cid_ref, ident_t *loc_ref) { 684 int gtid = *gtid_ref; 685 #ifdef BUILD_PARALLEL_ORDERED 686 kmp_team_t *team = __kmp_team_from_gtid(gtid); 687 #endif /* BUILD_PARALLEL_ORDERED */ 688 689 if (__kmp_env_consistency_check) { 690 if (__kmp_threads[gtid]->th.th_root->r.r_active) 691 #if KMP_USE_DYNAMIC_LOCK 692 __kmp_push_sync(gtid, ct_ordered_in_parallel, loc_ref, NULL, 0); 693 #else 694 __kmp_push_sync(gtid, ct_ordered_in_parallel, loc_ref, NULL); 695 #endif 696 } 697 #ifdef BUILD_PARALLEL_ORDERED 698 if (!team->t.t_serialized) { 699 KMP_MB(); 700 KMP_WAIT_YIELD(&team->t.t_ordered.dt.t_value, __kmp_tid_from_gtid(gtid), 701 KMP_EQ, NULL); 702 KMP_MB(); 703 } 704 #endif /* BUILD_PARALLEL_ORDERED */ 705 } 706 707 /* __kmp_parallel_dxo -- Signal the next task. */ 708 void __kmp_parallel_dxo(int *gtid_ref, int *cid_ref, ident_t *loc_ref) { 709 int gtid = *gtid_ref; 710 #ifdef BUILD_PARALLEL_ORDERED 711 int tid = __kmp_tid_from_gtid(gtid); 712 kmp_team_t *team = __kmp_team_from_gtid(gtid); 713 #endif /* BUILD_PARALLEL_ORDERED */ 714 715 if (__kmp_env_consistency_check) { 716 if (__kmp_threads[gtid]->th.th_root->r.r_active) 717 __kmp_pop_sync(gtid, ct_ordered_in_parallel, loc_ref); 718 } 719 #ifdef BUILD_PARALLEL_ORDERED 720 if (!team->t.t_serialized) { 721 KMP_MB(); /* Flush all pending memory write invalidates. */ 722 723 /* use the tid of the next thread in this team */ 724 /* TODO replace with general release procedure */ 725 team->t.t_ordered.dt.t_value = ((tid + 1) % team->t.t_nproc); 726 727 #if OMPT_SUPPORT && OMPT_BLAME 728 if (ompt_enabled && 729 ompt_callbacks.ompt_callback(ompt_event_release_ordered)) { 730 /* accept blame for "ordered" waiting */ 731 kmp_info_t *this_thread = __kmp_threads[gtid]; 732 ompt_callbacks.ompt_callback(ompt_event_release_ordered)( 733 this_thread->th.ompt_thread_info.wait_id); 734 } 735 #endif 736 737 KMP_MB(); /* Flush all pending memory write invalidates. */ 738 } 739 #endif /* BUILD_PARALLEL_ORDERED */ 740 } 741 742 /* ------------------------------------------------------------------------ */ 743 /* The BARRIER for a SINGLE process section is always explicit */ 744 745 int __kmp_enter_single(int gtid, ident_t *id_ref, int push_ws) { 746 int status; 747 kmp_info_t *th; 748 kmp_team_t *team; 749 750 if (!TCR_4(__kmp_init_parallel)) 751 __kmp_parallel_initialize(); 752 753 th = __kmp_threads[gtid]; 754 team = th->th.th_team; 755 status = 0; 756 757 th->th.th_ident = id_ref; 758 759 if (team->t.t_serialized) { 760 status = 1; 761 } else { 762 kmp_int32 old_this = th->th.th_local.this_construct; 763 764 ++th->th.th_local.this_construct; 765 /* try to set team count to thread count--success means thread got the 766 single block */ 767 /* TODO: Should this be acquire or release? */ 768 if (team->t.t_construct == old_this) { 769 status = KMP_COMPARE_AND_STORE_ACQ32(&team->t.t_construct, old_this, 770 th->th.th_local.this_construct); 771 } 772 #if USE_ITT_BUILD 773 if (__itt_metadata_add_ptr && __kmp_forkjoin_frames_mode == 3 && 774 KMP_MASTER_GTID(gtid) && 775 #if OMP_40_ENABLED 776 th->th.th_teams_microtask == NULL && 777 #endif 778 team->t.t_active_level == 779 1) { // Only report metadata by master of active team at level 1 780 __kmp_itt_metadata_single(id_ref); 781 } 782 #endif /* USE_ITT_BUILD */ 783 } 784 785 if (__kmp_env_consistency_check) { 786 if (status && push_ws) { 787 __kmp_push_workshare(gtid, ct_psingle, id_ref); 788 } else { 789 __kmp_check_workshare(gtid, ct_psingle, id_ref); 790 } 791 } 792 #if USE_ITT_BUILD 793 if (status) { 794 __kmp_itt_single_start(gtid); 795 } 796 #endif /* USE_ITT_BUILD */ 797 return status; 798 } 799 800 void __kmp_exit_single(int gtid) { 801 #if USE_ITT_BUILD 802 __kmp_itt_single_end(gtid); 803 #endif /* USE_ITT_BUILD */ 804 if (__kmp_env_consistency_check) 805 __kmp_pop_workshare(gtid, ct_psingle, NULL); 806 } 807 808 /* determine if we can go parallel or must use a serialized parallel region and 809 * how many threads we can use 810 * set_nproc is the number of threads requested for the team 811 * returns 0 if we should serialize or only use one thread, 812 * otherwise the number of threads to use 813 * The forkjoin lock is held by the caller. */ 814 static int __kmp_reserve_threads(kmp_root_t *root, kmp_team_t *parent_team, 815 int master_tid, int set_nthreads 816 #if OMP_40_ENABLED 817 , 818 int enter_teams 819 #endif /* OMP_40_ENABLED */ 820 ) { 821 int capacity; 822 int new_nthreads; 823 KMP_DEBUG_ASSERT(__kmp_init_serial); 824 KMP_DEBUG_ASSERT(root && parent_team); 825 826 // If dyn-var is set, dynamically adjust the number of desired threads, 827 // according to the method specified by dynamic_mode. 828 new_nthreads = set_nthreads; 829 if (!get__dynamic_2(parent_team, master_tid)) { 830 ; 831 } 832 #ifdef USE_LOAD_BALANCE 833 else if (__kmp_global.g.g_dynamic_mode == dynamic_load_balance) { 834 new_nthreads = __kmp_load_balance_nproc(root, set_nthreads); 835 if (new_nthreads == 1) { 836 KC_TRACE(10, ("__kmp_reserve_threads: T#%d load balance reduced " 837 "reservation to 1 thread\n", 838 master_tid)); 839 return 1; 840 } 841 if (new_nthreads < set_nthreads) { 842 KC_TRACE(10, ("__kmp_reserve_threads: T#%d load balance reduced " 843 "reservation to %d threads\n", 844 master_tid, new_nthreads)); 845 } 846 } 847 #endif /* USE_LOAD_BALANCE */ 848 else if (__kmp_global.g.g_dynamic_mode == dynamic_thread_limit) { 849 new_nthreads = __kmp_avail_proc - __kmp_nth + 850 (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc); 851 if (new_nthreads <= 1) { 852 KC_TRACE(10, ("__kmp_reserve_threads: T#%d thread limit reduced " 853 "reservation to 1 thread\n", 854 master_tid)); 855 return 1; 856 } 857 if (new_nthreads < set_nthreads) { 858 KC_TRACE(10, ("__kmp_reserve_threads: T#%d thread limit reduced " 859 "reservation to %d threads\n", 860 master_tid, new_nthreads)); 861 } else { 862 new_nthreads = set_nthreads; 863 } 864 } else if (__kmp_global.g.g_dynamic_mode == dynamic_random) { 865 if (set_nthreads > 2) { 866 new_nthreads = __kmp_get_random(parent_team->t.t_threads[master_tid]); 867 new_nthreads = (new_nthreads % set_nthreads) + 1; 868 if (new_nthreads == 1) { 869 KC_TRACE(10, ("__kmp_reserve_threads: T#%d dynamic random reduced " 870 "reservation to 1 thread\n", 871 master_tid)); 872 return 1; 873 } 874 if (new_nthreads < set_nthreads) { 875 KC_TRACE(10, ("__kmp_reserve_threads: T#%d dynamic random reduced " 876 "reservation to %d threads\n", 877 master_tid, new_nthreads)); 878 } 879 } 880 } else { 881 KMP_ASSERT(0); 882 } 883 884 // Respect KMP_ALL_THREADS/KMP_DEVICE_THREAD_LIMIT. 885 if (__kmp_nth + new_nthreads - 886 (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc) > 887 __kmp_max_nth) { 888 int tl_nthreads = __kmp_max_nth - __kmp_nth + 889 (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc); 890 if (tl_nthreads <= 0) { 891 tl_nthreads = 1; 892 } 893 894 // If dyn-var is false, emit a 1-time warning. 895 if (!get__dynamic_2(parent_team, master_tid) && (!__kmp_reserve_warn)) { 896 __kmp_reserve_warn = 1; 897 __kmp_msg(kmp_ms_warning, 898 KMP_MSG(CantFormThrTeam, set_nthreads, tl_nthreads), 899 KMP_HNT(Unset_ALL_THREADS), __kmp_msg_null); 900 } 901 if (tl_nthreads == 1) { 902 KC_TRACE(10, ("__kmp_reserve_threads: T#%d KMP_DEVICE_THREAD_LIMIT " 903 "reduced reservation to 1 thread\n", 904 master_tid)); 905 return 1; 906 } 907 KC_TRACE(10, ("__kmp_reserve_threads: T#%d KMP_DEVICE_THREAD_LIMIT reduced " 908 "reservation to %d threads\n", 909 master_tid, tl_nthreads)); 910 new_nthreads = tl_nthreads; 911 } 912 913 // Respect OMP_THREAD_LIMIT 914 if (root->r.r_cg_nthreads + new_nthreads - 915 (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc) > 916 __kmp_cg_max_nth) { 917 int tl_nthreads = __kmp_cg_max_nth - root->r.r_cg_nthreads + 918 (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc); 919 if (tl_nthreads <= 0) { 920 tl_nthreads = 1; 921 } 922 923 // If dyn-var is false, emit a 1-time warning. 924 if (!get__dynamic_2(parent_team, master_tid) && (!__kmp_reserve_warn)) { 925 __kmp_reserve_warn = 1; 926 __kmp_msg(kmp_ms_warning, 927 KMP_MSG(CantFormThrTeam, set_nthreads, tl_nthreads), 928 KMP_HNT(Unset_ALL_THREADS), __kmp_msg_null); 929 } 930 if (tl_nthreads == 1) { 931 KC_TRACE(10, ("__kmp_reserve_threads: T#%d OMP_THREAD_LIMIT " 932 "reduced reservation to 1 thread\n", 933 master_tid)); 934 return 1; 935 } 936 KC_TRACE(10, ("__kmp_reserve_threads: T#%d OMP_THREAD_LIMIT reduced " 937 "reservation to %d threads\n", 938 master_tid, tl_nthreads)); 939 new_nthreads = tl_nthreads; 940 } 941 942 // Check if the threads array is large enough, or needs expanding. 943 // See comment in __kmp_register_root() about the adjustment if 944 // __kmp_threads[0] == NULL. 945 capacity = __kmp_threads_capacity; 946 if (TCR_PTR(__kmp_threads[0]) == NULL) { 947 --capacity; 948 } 949 if (__kmp_nth + new_nthreads - 950 (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc) > 951 capacity) { 952 // Expand the threads array. 953 int slotsRequired = __kmp_nth + new_nthreads - 954 (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc) - 955 capacity; 956 int slotsAdded = __kmp_expand_threads(slotsRequired, slotsRequired); 957 if (slotsAdded < slotsRequired) { 958 // The threads array was not expanded enough. 959 new_nthreads -= (slotsRequired - slotsAdded); 960 KMP_ASSERT(new_nthreads >= 1); 961 962 // If dyn-var is false, emit a 1-time warning. 963 if (!get__dynamic_2(parent_team, master_tid) && (!__kmp_reserve_warn)) { 964 __kmp_reserve_warn = 1; 965 if (__kmp_tp_cached) { 966 __kmp_msg(kmp_ms_warning, 967 KMP_MSG(CantFormThrTeam, set_nthreads, new_nthreads), 968 KMP_HNT(Set_ALL_THREADPRIVATE, __kmp_tp_capacity), 969 KMP_HNT(PossibleSystemLimitOnThreads), __kmp_msg_null); 970 } else { 971 __kmp_msg(kmp_ms_warning, 972 KMP_MSG(CantFormThrTeam, set_nthreads, new_nthreads), 973 KMP_HNT(SystemLimitOnThreads), __kmp_msg_null); 974 } 975 } 976 } 977 } 978 979 #ifdef KMP_DEBUG 980 if (new_nthreads == 1) { 981 KC_TRACE(10, 982 ("__kmp_reserve_threads: T#%d serializing team after reclaiming " 983 "dead roots and rechecking; requested %d threads\n", 984 __kmp_get_gtid(), set_nthreads)); 985 } else { 986 KC_TRACE(10, ("__kmp_reserve_threads: T#%d allocating %d threads; requested" 987 " %d threads\n", 988 __kmp_get_gtid(), new_nthreads, set_nthreads)); 989 } 990 #endif // KMP_DEBUG 991 return new_nthreads; 992 } 993 994 /* Allocate threads from the thread pool and assign them to the new team. We are 995 assured that there are enough threads available, because we checked on that 996 earlier within critical section forkjoin */ 997 static void __kmp_fork_team_threads(kmp_root_t *root, kmp_team_t *team, 998 kmp_info_t *master_th, int master_gtid) { 999 int i; 1000 int use_hot_team; 1001 1002 KA_TRACE(10, ("__kmp_fork_team_threads: new_nprocs = %d\n", team->t.t_nproc)); 1003 KMP_DEBUG_ASSERT(master_gtid == __kmp_get_gtid()); 1004 KMP_MB(); 1005 1006 /* first, let's setup the master thread */ 1007 master_th->th.th_info.ds.ds_tid = 0; 1008 master_th->th.th_team = team; 1009 master_th->th.th_team_nproc = team->t.t_nproc; 1010 master_th->th.th_team_master = master_th; 1011 master_th->th.th_team_serialized = FALSE; 1012 master_th->th.th_dispatch = &team->t.t_dispatch[0]; 1013 1014 /* make sure we are not the optimized hot team */ 1015 #if KMP_NESTED_HOT_TEAMS 1016 use_hot_team = 0; 1017 kmp_hot_team_ptr_t *hot_teams = master_th->th.th_hot_teams; 1018 if (hot_teams) { // hot teams array is not allocated if 1019 // KMP_HOT_TEAMS_MAX_LEVEL=0 1020 int level = team->t.t_active_level - 1; // index in array of hot teams 1021 if (master_th->th.th_teams_microtask) { // are we inside the teams? 1022 if (master_th->th.th_teams_size.nteams > 1) { 1023 ++level; // level was not increased in teams construct for 1024 // team_of_masters 1025 } 1026 if (team->t.t_pkfn != (microtask_t)__kmp_teams_master && 1027 master_th->th.th_teams_level == team->t.t_level) { 1028 ++level; // level was not increased in teams construct for 1029 // team_of_workers before the parallel 1030 } // team->t.t_level will be increased inside parallel 1031 } 1032 if (level < __kmp_hot_teams_max_level) { 1033 if (hot_teams[level].hot_team) { 1034 // hot team has already been allocated for given level 1035 KMP_DEBUG_ASSERT(hot_teams[level].hot_team == team); 1036 use_hot_team = 1; // the team is ready to use 1037 } else { 1038 use_hot_team = 0; // AC: threads are not allocated yet 1039 hot_teams[level].hot_team = team; // remember new hot team 1040 hot_teams[level].hot_team_nth = team->t.t_nproc; 1041 } 1042 } else { 1043 use_hot_team = 0; 1044 } 1045 } 1046 #else 1047 use_hot_team = team == root->r.r_hot_team; 1048 #endif 1049 if (!use_hot_team) { 1050 1051 /* install the master thread */ 1052 team->t.t_threads[0] = master_th; 1053 __kmp_initialize_info(master_th, team, 0, master_gtid); 1054 1055 /* now, install the worker threads */ 1056 for (i = 1; i < team->t.t_nproc; i++) { 1057 1058 /* fork or reallocate a new thread and install it in team */ 1059 kmp_info_t *thr = __kmp_allocate_thread(root, team, i); 1060 team->t.t_threads[i] = thr; 1061 KMP_DEBUG_ASSERT(thr); 1062 KMP_DEBUG_ASSERT(thr->th.th_team == team); 1063 /* align team and thread arrived states */ 1064 KA_TRACE(20, ("__kmp_fork_team_threads: T#%d(%d:%d) init arrived " 1065 "T#%d(%d:%d) join =%llu, plain=%llu\n", 1066 __kmp_gtid_from_tid(0, team), team->t.t_id, 0, 1067 __kmp_gtid_from_tid(i, team), team->t.t_id, i, 1068 team->t.t_bar[bs_forkjoin_barrier].b_arrived, 1069 team->t.t_bar[bs_plain_barrier].b_arrived)); 1070 #if OMP_40_ENABLED 1071 thr->th.th_teams_microtask = master_th->th.th_teams_microtask; 1072 thr->th.th_teams_level = master_th->th.th_teams_level; 1073 thr->th.th_teams_size = master_th->th.th_teams_size; 1074 #endif 1075 { // Initialize threads' barrier data. 1076 int b; 1077 kmp_balign_t *balign = team->t.t_threads[i]->th.th_bar; 1078 for (b = 0; b < bs_last_barrier; ++b) { 1079 balign[b].bb.b_arrived = team->t.t_bar[b].b_arrived; 1080 KMP_DEBUG_ASSERT(balign[b].bb.wait_flag != KMP_BARRIER_PARENT_FLAG); 1081 #if USE_DEBUGGER 1082 balign[b].bb.b_worker_arrived = team->t.t_bar[b].b_team_arrived; 1083 #endif 1084 }; // for b 1085 } 1086 } 1087 1088 #if OMP_40_ENABLED && KMP_AFFINITY_SUPPORTED 1089 __kmp_partition_places(team); 1090 #endif 1091 } 1092 1093 KMP_MB(); 1094 } 1095 1096 #if KMP_ARCH_X86 || KMP_ARCH_X86_64 1097 // Propagate any changes to the floating point control registers out to the team 1098 // We try to avoid unnecessary writes to the relevant cache line in the team 1099 // structure, so we don't make changes unless they are needed. 1100 inline static void propagateFPControl(kmp_team_t *team) { 1101 if (__kmp_inherit_fp_control) { 1102 kmp_int16 x87_fpu_control_word; 1103 kmp_uint32 mxcsr; 1104 1105 // Get master values of FPU control flags (both X87 and vector) 1106 __kmp_store_x87_fpu_control_word(&x87_fpu_control_word); 1107 __kmp_store_mxcsr(&mxcsr); 1108 mxcsr &= KMP_X86_MXCSR_MASK; 1109 1110 // There is no point looking at t_fp_control_saved here. 1111 // If it is TRUE, we still have to update the values if they are different from 1112 // those we now have. 1113 // If it is FALSE we didn't save anything yet, but our objective is the same. We 1114 // have to ensure that the values in the team are the same as those we have. 1115 // So, this code achieves what we need whether or not t_fp_control_saved is 1116 // true. By checking whether the value needs updating we avoid unnecessary 1117 // writes that would put the cache-line into a written state, causing all 1118 // threads in the team to have to read it again. 1119 KMP_CHECK_UPDATE(team->t.t_x87_fpu_control_word, x87_fpu_control_word); 1120 KMP_CHECK_UPDATE(team->t.t_mxcsr, mxcsr); 1121 // Although we don't use this value, other code in the runtime wants to know 1122 // whether it should restore them. So we must ensure it is correct. 1123 KMP_CHECK_UPDATE(team->t.t_fp_control_saved, TRUE); 1124 } else { 1125 // Similarly here. Don't write to this cache-line in the team structure 1126 // unless we have to. 1127 KMP_CHECK_UPDATE(team->t.t_fp_control_saved, FALSE); 1128 } 1129 } 1130 1131 // Do the opposite, setting the hardware registers to the updated values from 1132 // the team. 1133 inline static void updateHWFPControl(kmp_team_t *team) { 1134 if (__kmp_inherit_fp_control && team->t.t_fp_control_saved) { 1135 // Only reset the fp control regs if they have been changed in the team. 1136 // the parallel region that we are exiting. 1137 kmp_int16 x87_fpu_control_word; 1138 kmp_uint32 mxcsr; 1139 __kmp_store_x87_fpu_control_word(&x87_fpu_control_word); 1140 __kmp_store_mxcsr(&mxcsr); 1141 mxcsr &= KMP_X86_MXCSR_MASK; 1142 1143 if (team->t.t_x87_fpu_control_word != x87_fpu_control_word) { 1144 __kmp_clear_x87_fpu_status_word(); 1145 __kmp_load_x87_fpu_control_word(&team->t.t_x87_fpu_control_word); 1146 } 1147 1148 if (team->t.t_mxcsr != mxcsr) { 1149 __kmp_load_mxcsr(&team->t.t_mxcsr); 1150 } 1151 } 1152 } 1153 #else 1154 #define propagateFPControl(x) ((void)0) 1155 #define updateHWFPControl(x) ((void)0) 1156 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */ 1157 1158 static void __kmp_alloc_argv_entries(int argc, kmp_team_t *team, 1159 int realloc); // forward declaration 1160 1161 /* Run a parallel region that has been serialized, so runs only in a team of the 1162 single master thread. */ 1163 void __kmp_serialized_parallel(ident_t *loc, kmp_int32 global_tid) { 1164 kmp_info_t *this_thr; 1165 kmp_team_t *serial_team; 1166 1167 KC_TRACE(10, ("__kmpc_serialized_parallel: called by T#%d\n", global_tid)); 1168 1169 /* Skip all this code for autopar serialized loops since it results in 1170 unacceptable overhead */ 1171 if (loc != NULL && (loc->flags & KMP_IDENT_AUTOPAR)) 1172 return; 1173 1174 if (!TCR_4(__kmp_init_parallel)) 1175 __kmp_parallel_initialize(); 1176 1177 this_thr = __kmp_threads[global_tid]; 1178 serial_team = this_thr->th.th_serial_team; 1179 1180 /* utilize the serialized team held by this thread */ 1181 KMP_DEBUG_ASSERT(serial_team); 1182 KMP_MB(); 1183 1184 if (__kmp_tasking_mode != tskm_immediate_exec) { 1185 KMP_DEBUG_ASSERT( 1186 this_thr->th.th_task_team == 1187 this_thr->th.th_team->t.t_task_team[this_thr->th.th_task_state]); 1188 KMP_DEBUG_ASSERT(serial_team->t.t_task_team[this_thr->th.th_task_state] == 1189 NULL); 1190 KA_TRACE(20, ("__kmpc_serialized_parallel: T#%d pushing task_team %p / " 1191 "team %p, new task_team = NULL\n", 1192 global_tid, this_thr->th.th_task_team, this_thr->th.th_team)); 1193 this_thr->th.th_task_team = NULL; 1194 } 1195 1196 #if OMP_40_ENABLED 1197 kmp_proc_bind_t proc_bind = this_thr->th.th_set_proc_bind; 1198 if (this_thr->th.th_current_task->td_icvs.proc_bind == proc_bind_false) { 1199 proc_bind = proc_bind_false; 1200 } else if (proc_bind == proc_bind_default) { 1201 // No proc_bind clause was specified, so use the current value 1202 // of proc-bind-var for this parallel region. 1203 proc_bind = this_thr->th.th_current_task->td_icvs.proc_bind; 1204 } 1205 // Reset for next parallel region 1206 this_thr->th.th_set_proc_bind = proc_bind_default; 1207 #endif /* OMP_40_ENABLED */ 1208 1209 if (this_thr->th.th_team != serial_team) { 1210 // Nested level will be an index in the nested nthreads array 1211 int level = this_thr->th.th_team->t.t_level; 1212 1213 if (serial_team->t.t_serialized) { 1214 /* this serial team was already used 1215 TODO increase performance by making this locks more specific */ 1216 kmp_team_t *new_team; 1217 1218 __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock); 1219 1220 #if OMPT_SUPPORT 1221 ompt_parallel_id_t ompt_parallel_id = __ompt_parallel_id_new(global_tid); 1222 #endif 1223 1224 new_team = __kmp_allocate_team(this_thr->th.th_root, 1, 1, 1225 #if OMPT_SUPPORT 1226 ompt_parallel_id, 1227 #endif 1228 #if OMP_40_ENABLED 1229 proc_bind, 1230 #endif 1231 &this_thr->th.th_current_task->td_icvs, 1232 0 USE_NESTED_HOT_ARG(NULL)); 1233 __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock); 1234 KMP_ASSERT(new_team); 1235 1236 /* setup new serialized team and install it */ 1237 new_team->t.t_threads[0] = this_thr; 1238 new_team->t.t_parent = this_thr->th.th_team; 1239 serial_team = new_team; 1240 this_thr->th.th_serial_team = serial_team; 1241 1242 KF_TRACE( 1243 10, 1244 ("__kmpc_serialized_parallel: T#%d allocated new serial team %p\n", 1245 global_tid, serial_team)); 1246 1247 /* TODO the above breaks the requirement that if we run out of resources, 1248 then we can still guarantee that serialized teams are ok, since we may 1249 need to allocate a new one */ 1250 } else { 1251 KF_TRACE( 1252 10, 1253 ("__kmpc_serialized_parallel: T#%d reusing cached serial team %p\n", 1254 global_tid, serial_team)); 1255 } 1256 1257 /* we have to initialize this serial team */ 1258 KMP_DEBUG_ASSERT(serial_team->t.t_threads); 1259 KMP_DEBUG_ASSERT(serial_team->t.t_threads[0] == this_thr); 1260 KMP_DEBUG_ASSERT(this_thr->th.th_team != serial_team); 1261 serial_team->t.t_ident = loc; 1262 serial_team->t.t_serialized = 1; 1263 serial_team->t.t_nproc = 1; 1264 serial_team->t.t_parent = this_thr->th.th_team; 1265 serial_team->t.t_sched = this_thr->th.th_team->t.t_sched; 1266 this_thr->th.th_team = serial_team; 1267 serial_team->t.t_master_tid = this_thr->th.th_info.ds.ds_tid; 1268 1269 KF_TRACE(10, ("__kmpc_serialized_parallel: T#d curtask=%p\n", global_tid, 1270 this_thr->th.th_current_task)); 1271 KMP_ASSERT(this_thr->th.th_current_task->td_flags.executing == 1); 1272 this_thr->th.th_current_task->td_flags.executing = 0; 1273 1274 __kmp_push_current_task_to_thread(this_thr, serial_team, 0); 1275 1276 /* TODO: GEH: do ICVs work for nested serialized teams? Don't we need an 1277 implicit task for each serialized task represented by 1278 team->t.t_serialized? */ 1279 copy_icvs(&this_thr->th.th_current_task->td_icvs, 1280 &this_thr->th.th_current_task->td_parent->td_icvs); 1281 1282 // Thread value exists in the nested nthreads array for the next nested 1283 // level 1284 if (__kmp_nested_nth.used && (level + 1 < __kmp_nested_nth.used)) { 1285 this_thr->th.th_current_task->td_icvs.nproc = 1286 __kmp_nested_nth.nth[level + 1]; 1287 } 1288 1289 #if OMP_40_ENABLED 1290 if (__kmp_nested_proc_bind.used && 1291 (level + 1 < __kmp_nested_proc_bind.used)) { 1292 this_thr->th.th_current_task->td_icvs.proc_bind = 1293 __kmp_nested_proc_bind.bind_types[level + 1]; 1294 } 1295 #endif /* OMP_40_ENABLED */ 1296 1297 #if USE_DEBUGGER 1298 serial_team->t.t_pkfn = (microtask_t)(~0); // For the debugger. 1299 #endif 1300 this_thr->th.th_info.ds.ds_tid = 0; 1301 1302 /* set thread cache values */ 1303 this_thr->th.th_team_nproc = 1; 1304 this_thr->th.th_team_master = this_thr; 1305 this_thr->th.th_team_serialized = 1; 1306 1307 serial_team->t.t_level = serial_team->t.t_parent->t.t_level + 1; 1308 serial_team->t.t_active_level = serial_team->t.t_parent->t.t_active_level; 1309 1310 propagateFPControl(serial_team); 1311 1312 /* check if we need to allocate dispatch buffers stack */ 1313 KMP_DEBUG_ASSERT(serial_team->t.t_dispatch); 1314 if (!serial_team->t.t_dispatch->th_disp_buffer) { 1315 serial_team->t.t_dispatch->th_disp_buffer = 1316 (dispatch_private_info_t *)__kmp_allocate( 1317 sizeof(dispatch_private_info_t)); 1318 } 1319 this_thr->th.th_dispatch = serial_team->t.t_dispatch; 1320 1321 #if OMPT_SUPPORT 1322 ompt_parallel_id_t ompt_parallel_id = __ompt_parallel_id_new(global_tid); 1323 __ompt_team_assign_id(serial_team, ompt_parallel_id); 1324 #endif 1325 1326 KMP_MB(); 1327 1328 } else { 1329 /* this serialized team is already being used, 1330 * that's fine, just add another nested level */ 1331 KMP_DEBUG_ASSERT(this_thr->th.th_team == serial_team); 1332 KMP_DEBUG_ASSERT(serial_team->t.t_threads); 1333 KMP_DEBUG_ASSERT(serial_team->t.t_threads[0] == this_thr); 1334 ++serial_team->t.t_serialized; 1335 this_thr->th.th_team_serialized = serial_team->t.t_serialized; 1336 1337 // Nested level will be an index in the nested nthreads array 1338 int level = this_thr->th.th_team->t.t_level; 1339 // Thread value exists in the nested nthreads array for the next nested 1340 // level 1341 if (__kmp_nested_nth.used && (level + 1 < __kmp_nested_nth.used)) { 1342 this_thr->th.th_current_task->td_icvs.nproc = 1343 __kmp_nested_nth.nth[level + 1]; 1344 } 1345 serial_team->t.t_level++; 1346 KF_TRACE(10, ("__kmpc_serialized_parallel: T#%d increasing nesting level " 1347 "of serial team %p to %d\n", 1348 global_tid, serial_team, serial_team->t.t_level)); 1349 1350 /* allocate/push dispatch buffers stack */ 1351 KMP_DEBUG_ASSERT(serial_team->t.t_dispatch); 1352 { 1353 dispatch_private_info_t *disp_buffer = 1354 (dispatch_private_info_t *)__kmp_allocate( 1355 sizeof(dispatch_private_info_t)); 1356 disp_buffer->next = serial_team->t.t_dispatch->th_disp_buffer; 1357 serial_team->t.t_dispatch->th_disp_buffer = disp_buffer; 1358 } 1359 this_thr->th.th_dispatch = serial_team->t.t_dispatch; 1360 1361 KMP_MB(); 1362 } 1363 #if OMP_40_ENABLED 1364 KMP_CHECK_UPDATE(serial_team->t.t_cancel_request, cancel_noreq); 1365 #endif 1366 1367 if (__kmp_env_consistency_check) 1368 __kmp_push_parallel(global_tid, NULL); 1369 } 1370 1371 /* most of the work for a fork */ 1372 /* return true if we really went parallel, false if serialized */ 1373 int __kmp_fork_call(ident_t *loc, int gtid, 1374 enum fork_context_e call_context, // Intel, GNU, ... 1375 kmp_int32 argc, 1376 #if OMPT_SUPPORT 1377 void *unwrapped_task, 1378 #endif 1379 microtask_t microtask, launch_t invoker, 1380 /* TODO: revert workaround for Intel(R) 64 tracker #96 */ 1381 #if (KMP_ARCH_X86_64 || KMP_ARCH_ARM || KMP_ARCH_AARCH64) && KMP_OS_LINUX 1382 va_list *ap 1383 #else 1384 va_list ap 1385 #endif 1386 ) { 1387 void **argv; 1388 int i; 1389 int master_tid; 1390 int master_this_cons; 1391 kmp_team_t *team; 1392 kmp_team_t *parent_team; 1393 kmp_info_t *master_th; 1394 kmp_root_t *root; 1395 int nthreads; 1396 int master_active; 1397 int master_set_numthreads; 1398 int level; 1399 #if OMP_40_ENABLED 1400 int active_level; 1401 int teams_level; 1402 #endif 1403 #if KMP_NESTED_HOT_TEAMS 1404 kmp_hot_team_ptr_t **p_hot_teams; 1405 #endif 1406 { // KMP_TIME_BLOCK 1407 KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(KMP_fork_call); 1408 KMP_COUNT_VALUE(OMP_PARALLEL_args, argc); 1409 1410 KA_TRACE(20, ("__kmp_fork_call: enter T#%d\n", gtid)); 1411 if (__kmp_stkpadding > 0 && __kmp_root[gtid] != NULL) { 1412 /* Some systems prefer the stack for the root thread(s) to start with */ 1413 /* some gap from the parent stack to prevent false sharing. */ 1414 void *dummy = KMP_ALLOCA(__kmp_stkpadding); 1415 /* These 2 lines below are so this does not get optimized out */ 1416 if (__kmp_stkpadding > KMP_MAX_STKPADDING) 1417 __kmp_stkpadding += (short)((kmp_int64)dummy); 1418 } 1419 1420 /* initialize if needed */ 1421 KMP_DEBUG_ASSERT( 1422 __kmp_init_serial); // AC: potentially unsafe, not in sync with shutdown 1423 if (!TCR_4(__kmp_init_parallel)) 1424 __kmp_parallel_initialize(); 1425 1426 /* setup current data */ 1427 master_th = __kmp_threads[gtid]; // AC: potentially unsafe, not in sync with 1428 // shutdown 1429 parent_team = master_th->th.th_team; 1430 master_tid = master_th->th.th_info.ds.ds_tid; 1431 master_this_cons = master_th->th.th_local.this_construct; 1432 root = master_th->th.th_root; 1433 master_active = root->r.r_active; 1434 master_set_numthreads = master_th->th.th_set_nproc; 1435 1436 #if OMPT_SUPPORT 1437 ompt_parallel_id_t ompt_parallel_id; 1438 ompt_task_id_t ompt_task_id; 1439 ompt_frame_t *ompt_frame; 1440 ompt_task_id_t my_task_id; 1441 ompt_parallel_id_t my_parallel_id; 1442 1443 if (ompt_enabled) { 1444 ompt_parallel_id = __ompt_parallel_id_new(gtid); 1445 ompt_task_id = __ompt_get_task_id_internal(0); 1446 ompt_frame = __ompt_get_task_frame_internal(0); 1447 } 1448 #endif 1449 1450 // Nested level will be an index in the nested nthreads array 1451 level = parent_team->t.t_level; 1452 // used to launch non-serial teams even if nested is not allowed 1453 active_level = parent_team->t.t_active_level; 1454 #if OMP_40_ENABLED 1455 // needed to check nesting inside the teams 1456 teams_level = master_th->th.th_teams_level; 1457 #endif 1458 #if KMP_NESTED_HOT_TEAMS 1459 p_hot_teams = &master_th->th.th_hot_teams; 1460 if (*p_hot_teams == NULL && __kmp_hot_teams_max_level > 0) { 1461 *p_hot_teams = (kmp_hot_team_ptr_t *)__kmp_allocate( 1462 sizeof(kmp_hot_team_ptr_t) * __kmp_hot_teams_max_level); 1463 (*p_hot_teams)[0].hot_team = root->r.r_hot_team; 1464 // it is either actual or not needed (when active_level > 0) 1465 (*p_hot_teams)[0].hot_team_nth = 1; 1466 } 1467 #endif 1468 1469 #if OMPT_SUPPORT 1470 if (ompt_enabled && 1471 ompt_callbacks.ompt_callback(ompt_event_parallel_begin)) { 1472 int team_size = master_set_numthreads; 1473 1474 ompt_callbacks.ompt_callback(ompt_event_parallel_begin)( 1475 ompt_task_id, ompt_frame, ompt_parallel_id, team_size, unwrapped_task, 1476 OMPT_INVOKER(call_context)); 1477 } 1478 #endif 1479 1480 master_th->th.th_ident = loc; 1481 1482 #if OMP_40_ENABLED 1483 if (master_th->th.th_teams_microtask && ap && 1484 microtask != (microtask_t)__kmp_teams_master && level == teams_level) { 1485 // AC: This is start of parallel that is nested inside teams construct. 1486 // The team is actual (hot), all workers are ready at the fork barrier. 1487 // No lock needed to initialize the team a bit, then free workers. 1488 parent_team->t.t_ident = loc; 1489 __kmp_alloc_argv_entries(argc, parent_team, TRUE); 1490 parent_team->t.t_argc = argc; 1491 argv = (void **)parent_team->t.t_argv; 1492 for (i = argc - 1; i >= 0; --i) 1493 /* TODO: revert workaround for Intel(R) 64 tracker #96 */ 1494 #if (KMP_ARCH_X86_64 || KMP_ARCH_ARM || KMP_ARCH_AARCH64) && KMP_OS_LINUX 1495 *argv++ = va_arg(*ap, void *); 1496 #else 1497 *argv++ = va_arg(ap, void *); 1498 #endif 1499 // Increment our nested depth levels, but not increase the serialization 1500 if (parent_team == master_th->th.th_serial_team) { 1501 // AC: we are in serialized parallel 1502 __kmpc_serialized_parallel(loc, gtid); 1503 KMP_DEBUG_ASSERT(parent_team->t.t_serialized > 1); 1504 // AC: need this in order enquiry functions work 1505 // correctly, will restore at join time 1506 parent_team->t.t_serialized--; 1507 #if OMPT_SUPPORT 1508 void *dummy; 1509 void **exit_runtime_p; 1510 1511 ompt_lw_taskteam_t lw_taskteam; 1512 1513 if (ompt_enabled) { 1514 __ompt_lw_taskteam_init(&lw_taskteam, master_th, gtid, unwrapped_task, 1515 ompt_parallel_id); 1516 lw_taskteam.ompt_task_info.task_id = __ompt_task_id_new(gtid); 1517 exit_runtime_p = 1518 &(lw_taskteam.ompt_task_info.frame.exit_runtime_frame); 1519 1520 __ompt_lw_taskteam_link(&lw_taskteam, master_th); 1521 1522 #if OMPT_TRACE 1523 /* OMPT implicit task begin */ 1524 my_task_id = lw_taskteam.ompt_task_info.task_id; 1525 my_parallel_id = parent_team->t.ompt_team_info.parallel_id; 1526 if (ompt_callbacks.ompt_callback(ompt_event_implicit_task_begin)) { 1527 ompt_callbacks.ompt_callback(ompt_event_implicit_task_begin)( 1528 my_parallel_id, my_task_id); 1529 } 1530 #endif 1531 1532 /* OMPT state */ 1533 master_th->th.ompt_thread_info.state = ompt_state_work_parallel; 1534 } else { 1535 exit_runtime_p = &dummy; 1536 } 1537 #endif 1538 1539 { 1540 KMP_TIME_PARTITIONED_BLOCK(OMP_parallel); 1541 KMP_SET_THREAD_STATE_BLOCK(IMPLICIT_TASK); 1542 __kmp_invoke_microtask(microtask, gtid, 0, argc, parent_team->t.t_argv 1543 #if OMPT_SUPPORT 1544 , 1545 exit_runtime_p 1546 #endif 1547 ); 1548 } 1549 1550 #if OMPT_SUPPORT 1551 *exit_runtime_p = NULL; 1552 if (ompt_enabled) { 1553 #if OMPT_TRACE 1554 lw_taskteam.ompt_task_info.frame.exit_runtime_frame = NULL; 1555 1556 if (ompt_callbacks.ompt_callback(ompt_event_implicit_task_end)) { 1557 ompt_callbacks.ompt_callback(ompt_event_implicit_task_end)( 1558 ompt_parallel_id, ompt_task_id); 1559 } 1560 1561 __ompt_lw_taskteam_unlink(master_th); 1562 // reset clear the task id only after unlinking the task 1563 lw_taskteam.ompt_task_info.task_id = ompt_task_id_none; 1564 #endif 1565 1566 if (ompt_callbacks.ompt_callback(ompt_event_parallel_end)) { 1567 ompt_callbacks.ompt_callback(ompt_event_parallel_end)( 1568 ompt_parallel_id, ompt_task_id, OMPT_INVOKER(call_context)); 1569 } 1570 master_th->th.ompt_thread_info.state = ompt_state_overhead; 1571 } 1572 #endif 1573 return TRUE; 1574 } 1575 1576 parent_team->t.t_pkfn = microtask; 1577 #if OMPT_SUPPORT 1578 parent_team->t.ompt_team_info.microtask = unwrapped_task; 1579 #endif 1580 parent_team->t.t_invoke = invoker; 1581 KMP_TEST_THEN_INC32((kmp_int32 *)&root->r.r_in_parallel); 1582 parent_team->t.t_active_level++; 1583 parent_team->t.t_level++; 1584 1585 /* Change number of threads in the team if requested */ 1586 if (master_set_numthreads) { // The parallel has num_threads clause 1587 if (master_set_numthreads < master_th->th.th_teams_size.nth) { 1588 // AC: only can reduce number of threads dynamically, can't increase 1589 kmp_info_t **other_threads = parent_team->t.t_threads; 1590 parent_team->t.t_nproc = master_set_numthreads; 1591 for (i = 0; i < master_set_numthreads; ++i) { 1592 other_threads[i]->th.th_team_nproc = master_set_numthreads; 1593 } 1594 // Keep extra threads hot in the team for possible next parallels 1595 } 1596 master_th->th.th_set_nproc = 0; 1597 } 1598 1599 #if USE_DEBUGGER 1600 if (__kmp_debugging) { // Let debugger override number of threads. 1601 int nth = __kmp_omp_num_threads(loc); 1602 if (nth > 0) { // 0 means debugger doesn't want to change num threads 1603 master_set_numthreads = nth; 1604 }; // if 1605 }; // if 1606 #endif 1607 1608 KF_TRACE(10, ("__kmp_fork_call: before internal fork: root=%p, team=%p, " 1609 "master_th=%p, gtid=%d\n", 1610 root, parent_team, master_th, gtid)); 1611 __kmp_internal_fork(loc, gtid, parent_team); 1612 KF_TRACE(10, ("__kmp_fork_call: after internal fork: root=%p, team=%p, " 1613 "master_th=%p, gtid=%d\n", 1614 root, parent_team, master_th, gtid)); 1615 1616 /* Invoke microtask for MASTER thread */ 1617 KA_TRACE(20, ("__kmp_fork_call: T#%d(%d:0) invoke microtask = %p\n", gtid, 1618 parent_team->t.t_id, parent_team->t.t_pkfn)); 1619 1620 { 1621 KMP_TIME_PARTITIONED_BLOCK(OMP_parallel); 1622 KMP_SET_THREAD_STATE_BLOCK(IMPLICIT_TASK); 1623 if (!parent_team->t.t_invoke(gtid)) { 1624 KMP_ASSERT2(0, "cannot invoke microtask for MASTER thread"); 1625 } 1626 } 1627 KA_TRACE(20, ("__kmp_fork_call: T#%d(%d:0) done microtask = %p\n", gtid, 1628 parent_team->t.t_id, parent_team->t.t_pkfn)); 1629 KMP_MB(); /* Flush all pending memory write invalidates. */ 1630 1631 KA_TRACE(20, ("__kmp_fork_call: parallel exit T#%d\n", gtid)); 1632 1633 return TRUE; 1634 } // Parallel closely nested in teams construct 1635 #endif /* OMP_40_ENABLED */ 1636 1637 #if KMP_DEBUG 1638 if (__kmp_tasking_mode != tskm_immediate_exec) { 1639 KMP_DEBUG_ASSERT(master_th->th.th_task_team == 1640 parent_team->t.t_task_team[master_th->th.th_task_state]); 1641 } 1642 #endif 1643 1644 if (parent_team->t.t_active_level >= 1645 master_th->th.th_current_task->td_icvs.max_active_levels) { 1646 nthreads = 1; 1647 } else { 1648 #if OMP_40_ENABLED 1649 int enter_teams = ((ap == NULL && active_level == 0) || 1650 (ap && teams_level > 0 && teams_level == level)); 1651 #endif 1652 nthreads = 1653 master_set_numthreads 1654 ? master_set_numthreads 1655 : get__nproc_2( 1656 parent_team, 1657 master_tid); // TODO: get nproc directly from current task 1658 1659 // Check if we need to take forkjoin lock? (no need for serialized 1660 // parallel out of teams construct). This code moved here from 1661 // __kmp_reserve_threads() to speedup nested serialized parallels. 1662 if (nthreads > 1) { 1663 if ((!get__nested(master_th) && (root->r.r_in_parallel 1664 #if OMP_40_ENABLED 1665 && !enter_teams 1666 #endif /* OMP_40_ENABLED */ 1667 )) || 1668 (__kmp_library == library_serial)) { 1669 KC_TRACE(10, ("__kmp_fork_call: T#%d serializing team; requested %d" 1670 " threads\n", 1671 gtid, nthreads)); 1672 nthreads = 1; 1673 } 1674 } 1675 if (nthreads > 1) { 1676 /* determine how many new threads we can use */ 1677 __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock); 1678 nthreads = __kmp_reserve_threads( 1679 root, parent_team, master_tid, nthreads 1680 #if OMP_40_ENABLED 1681 /* AC: If we execute teams from parallel region (on host), then 1682 teams should be created but each can only have 1 thread if 1683 nesting is disabled. If teams called from serial region, then 1684 teams and their threads should be created regardless of the 1685 nesting setting. */ 1686 , 1687 enter_teams 1688 #endif /* OMP_40_ENABLED */ 1689 ); 1690 if (nthreads == 1) { 1691 // Free lock for single thread execution here; for multi-thread 1692 // execution it will be freed later after team of threads created 1693 // and initialized 1694 __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock); 1695 } 1696 } 1697 } 1698 KMP_DEBUG_ASSERT(nthreads > 0); 1699 1700 // If we temporarily changed the set number of threads then restore it now 1701 master_th->th.th_set_nproc = 0; 1702 1703 /* create a serialized parallel region? */ 1704 if (nthreads == 1) { 1705 /* josh todo: hypothetical question: what do we do for OS X*? */ 1706 #if KMP_OS_LINUX && \ 1707 (KMP_ARCH_X86 || KMP_ARCH_X86_64 || KMP_ARCH_ARM || KMP_ARCH_AARCH64) 1708 void *args[argc]; 1709 #else 1710 void **args = (void **)KMP_ALLOCA(argc * sizeof(void *)); 1711 #endif /* KMP_OS_LINUX && ( KMP_ARCH_X86 || KMP_ARCH_X86_64 || KMP_ARCH_ARM || \ 1712 KMP_ARCH_AARCH64) */ 1713 1714 KA_TRACE(20, 1715 ("__kmp_fork_call: T#%d serializing parallel region\n", gtid)); 1716 1717 __kmpc_serialized_parallel(loc, gtid); 1718 1719 if (call_context == fork_context_intel) { 1720 /* TODO this sucks, use the compiler itself to pass args! :) */ 1721 master_th->th.th_serial_team->t.t_ident = loc; 1722 #if OMP_40_ENABLED 1723 if (!ap) { 1724 // revert change made in __kmpc_serialized_parallel() 1725 master_th->th.th_serial_team->t.t_level--; 1726 // Get args from parent team for teams construct 1727 1728 #if OMPT_SUPPORT 1729 void *dummy; 1730 void **exit_runtime_p; 1731 1732 ompt_lw_taskteam_t lw_taskteam; 1733 1734 if (ompt_enabled) { 1735 __ompt_lw_taskteam_init(&lw_taskteam, master_th, gtid, 1736 unwrapped_task, ompt_parallel_id); 1737 lw_taskteam.ompt_task_info.task_id = __ompt_task_id_new(gtid); 1738 exit_runtime_p = 1739 &(lw_taskteam.ompt_task_info.frame.exit_runtime_frame); 1740 1741 __ompt_lw_taskteam_link(&lw_taskteam, master_th); 1742 1743 #if OMPT_TRACE 1744 my_task_id = lw_taskteam.ompt_task_info.task_id; 1745 if (ompt_callbacks.ompt_callback(ompt_event_implicit_task_begin)) { 1746 ompt_callbacks.ompt_callback(ompt_event_implicit_task_begin)( 1747 ompt_parallel_id, my_task_id); 1748 } 1749 #endif 1750 1751 /* OMPT state */ 1752 master_th->th.ompt_thread_info.state = ompt_state_work_parallel; 1753 } else { 1754 exit_runtime_p = &dummy; 1755 } 1756 #endif 1757 1758 { 1759 KMP_TIME_PARTITIONED_BLOCK(OMP_parallel); 1760 KMP_SET_THREAD_STATE_BLOCK(IMPLICIT_TASK); 1761 __kmp_invoke_microtask(microtask, gtid, 0, argc, 1762 parent_team->t.t_argv 1763 #if OMPT_SUPPORT 1764 , 1765 exit_runtime_p 1766 #endif 1767 ); 1768 } 1769 1770 #if OMPT_SUPPORT 1771 *exit_runtime_p = NULL; 1772 if (ompt_enabled) { 1773 lw_taskteam.ompt_task_info.frame.exit_runtime_frame = NULL; 1774 1775 #if OMPT_TRACE 1776 if (ompt_callbacks.ompt_callback(ompt_event_implicit_task_end)) { 1777 ompt_callbacks.ompt_callback(ompt_event_implicit_task_end)( 1778 ompt_parallel_id, ompt_task_id); 1779 } 1780 #endif 1781 1782 __ompt_lw_taskteam_unlink(master_th); 1783 // reset clear the task id only after unlinking the task 1784 lw_taskteam.ompt_task_info.task_id = ompt_task_id_none; 1785 1786 if (ompt_callbacks.ompt_callback(ompt_event_parallel_end)) { 1787 ompt_callbacks.ompt_callback(ompt_event_parallel_end)( 1788 ompt_parallel_id, ompt_task_id, OMPT_INVOKER(call_context)); 1789 } 1790 master_th->th.ompt_thread_info.state = ompt_state_overhead; 1791 } 1792 #endif 1793 } else if (microtask == (microtask_t)__kmp_teams_master) { 1794 KMP_DEBUG_ASSERT(master_th->th.th_team == 1795 master_th->th.th_serial_team); 1796 team = master_th->th.th_team; 1797 // team->t.t_pkfn = microtask; 1798 team->t.t_invoke = invoker; 1799 __kmp_alloc_argv_entries(argc, team, TRUE); 1800 team->t.t_argc = argc; 1801 argv = (void **)team->t.t_argv; 1802 if (ap) { 1803 for (i = argc - 1; i >= 0; --i) 1804 // TODO: revert workaround for Intel(R) 64 tracker #96 1805 #if (KMP_ARCH_X86_64 || KMP_ARCH_ARM || KMP_ARCH_AARCH64) && KMP_OS_LINUX 1806 *argv++ = va_arg(*ap, void *); 1807 #else 1808 *argv++ = va_arg(ap, void *); 1809 #endif 1810 } else { 1811 for (i = 0; i < argc; ++i) 1812 // Get args from parent team for teams construct 1813 argv[i] = parent_team->t.t_argv[i]; 1814 } 1815 // AC: revert change made in __kmpc_serialized_parallel() 1816 // because initial code in teams should have level=0 1817 team->t.t_level--; 1818 // AC: call special invoker for outer "parallel" of teams construct 1819 { 1820 KMP_TIME_PARTITIONED_BLOCK(OMP_parallel); 1821 KMP_SET_THREAD_STATE_BLOCK(IMPLICIT_TASK); 1822 invoker(gtid); 1823 } 1824 } else { 1825 #endif /* OMP_40_ENABLED */ 1826 argv = args; 1827 for (i = argc - 1; i >= 0; --i) 1828 // TODO: revert workaround for Intel(R) 64 tracker #96 1829 #if (KMP_ARCH_X86_64 || KMP_ARCH_ARM || KMP_ARCH_AARCH64) && KMP_OS_LINUX 1830 *argv++ = va_arg(*ap, void *); 1831 #else 1832 *argv++ = va_arg(ap, void *); 1833 #endif 1834 KMP_MB(); 1835 1836 #if OMPT_SUPPORT 1837 void *dummy; 1838 void **exit_runtime_p; 1839 1840 ompt_lw_taskteam_t lw_taskteam; 1841 1842 if (ompt_enabled) { 1843 __ompt_lw_taskteam_init(&lw_taskteam, master_th, gtid, 1844 unwrapped_task, ompt_parallel_id); 1845 lw_taskteam.ompt_task_info.task_id = __ompt_task_id_new(gtid); 1846 exit_runtime_p = 1847 &(lw_taskteam.ompt_task_info.frame.exit_runtime_frame); 1848 1849 __ompt_lw_taskteam_link(&lw_taskteam, master_th); 1850 1851 #if OMPT_TRACE 1852 /* OMPT implicit task begin */ 1853 my_task_id = lw_taskteam.ompt_task_info.task_id; 1854 my_parallel_id = ompt_parallel_id; 1855 if (ompt_callbacks.ompt_callback(ompt_event_implicit_task_begin)) { 1856 ompt_callbacks.ompt_callback(ompt_event_implicit_task_begin)( 1857 my_parallel_id, my_task_id); 1858 } 1859 #endif 1860 1861 /* OMPT state */ 1862 master_th->th.ompt_thread_info.state = ompt_state_work_parallel; 1863 } else { 1864 exit_runtime_p = &dummy; 1865 } 1866 #endif 1867 1868 { 1869 KMP_TIME_PARTITIONED_BLOCK(OMP_parallel); 1870 KMP_SET_THREAD_STATE_BLOCK(IMPLICIT_TASK); 1871 __kmp_invoke_microtask(microtask, gtid, 0, argc, args 1872 #if OMPT_SUPPORT 1873 , 1874 exit_runtime_p 1875 #endif 1876 ); 1877 } 1878 1879 #if OMPT_SUPPORT 1880 *exit_runtime_p = NULL; 1881 if (ompt_enabled) { 1882 #if OMPT_TRACE 1883 lw_taskteam.ompt_task_info.frame.exit_runtime_frame = NULL; 1884 1885 if (ompt_callbacks.ompt_callback(ompt_event_implicit_task_end)) { 1886 ompt_callbacks.ompt_callback(ompt_event_implicit_task_end)( 1887 my_parallel_id, my_task_id); 1888 } 1889 #endif 1890 1891 __ompt_lw_taskteam_unlink(master_th); 1892 // reset clear the task id only after unlinking the task 1893 lw_taskteam.ompt_task_info.task_id = ompt_task_id_none; 1894 1895 if (ompt_callbacks.ompt_callback(ompt_event_parallel_end)) { 1896 ompt_callbacks.ompt_callback(ompt_event_parallel_end)( 1897 ompt_parallel_id, ompt_task_id, OMPT_INVOKER(call_context)); 1898 } 1899 master_th->th.ompt_thread_info.state = ompt_state_overhead; 1900 } 1901 #endif 1902 #if OMP_40_ENABLED 1903 } 1904 #endif /* OMP_40_ENABLED */ 1905 } else if (call_context == fork_context_gnu) { 1906 #if OMPT_SUPPORT 1907 ompt_lw_taskteam_t *lwt = 1908 (ompt_lw_taskteam_t *)__kmp_allocate(sizeof(ompt_lw_taskteam_t)); 1909 __ompt_lw_taskteam_init(lwt, master_th, gtid, unwrapped_task, 1910 ompt_parallel_id); 1911 1912 lwt->ompt_task_info.task_id = __ompt_task_id_new(gtid); 1913 lwt->ompt_task_info.frame.exit_runtime_frame = NULL; 1914 __ompt_lw_taskteam_link(lwt, master_th); 1915 #endif 1916 1917 // we were called from GNU native code 1918 KA_TRACE(20, ("__kmp_fork_call: T#%d serial exit\n", gtid)); 1919 return FALSE; 1920 } else { 1921 KMP_ASSERT2(call_context < fork_context_last, 1922 "__kmp_fork_call: unknown fork_context parameter"); 1923 } 1924 1925 KA_TRACE(20, ("__kmp_fork_call: T#%d serial exit\n", gtid)); 1926 KMP_MB(); 1927 return FALSE; 1928 } 1929 1930 // GEH: only modify the executing flag in the case when not serialized 1931 // serialized case is handled in kmpc_serialized_parallel 1932 KF_TRACE(10, ("__kmp_fork_call: parent_team_aclevel=%d, master_th=%p, " 1933 "curtask=%p, curtask_max_aclevel=%d\n", 1934 parent_team->t.t_active_level, master_th, 1935 master_th->th.th_current_task, 1936 master_th->th.th_current_task->td_icvs.max_active_levels)); 1937 // TODO: GEH - cannot do this assertion because root thread not set up as 1938 // executing 1939 // KMP_ASSERT( master_th->th.th_current_task->td_flags.executing == 1 ); 1940 master_th->th.th_current_task->td_flags.executing = 0; 1941 1942 #if OMP_40_ENABLED 1943 if (!master_th->th.th_teams_microtask || level > teams_level) 1944 #endif /* OMP_40_ENABLED */ 1945 { 1946 /* Increment our nested depth level */ 1947 KMP_TEST_THEN_INC32((kmp_int32 *)&root->r.r_in_parallel); 1948 } 1949 1950 // See if we need to make a copy of the ICVs. 1951 int nthreads_icv = master_th->th.th_current_task->td_icvs.nproc; 1952 if ((level + 1 < __kmp_nested_nth.used) && 1953 (__kmp_nested_nth.nth[level + 1] != nthreads_icv)) { 1954 nthreads_icv = __kmp_nested_nth.nth[level + 1]; 1955 } else { 1956 nthreads_icv = 0; // don't update 1957 } 1958 1959 #if OMP_40_ENABLED 1960 // Figure out the proc_bind_policy for the new team. 1961 kmp_proc_bind_t proc_bind = master_th->th.th_set_proc_bind; 1962 kmp_proc_bind_t proc_bind_icv = 1963 proc_bind_default; // proc_bind_default means don't update 1964 if (master_th->th.th_current_task->td_icvs.proc_bind == proc_bind_false) { 1965 proc_bind = proc_bind_false; 1966 } else { 1967 if (proc_bind == proc_bind_default) { 1968 // No proc_bind clause specified; use current proc-bind-var for this 1969 // parallel region 1970 proc_bind = master_th->th.th_current_task->td_icvs.proc_bind; 1971 } 1972 /* else: The proc_bind policy was specified explicitly on parallel clause. 1973 This overrides proc-bind-var for this parallel region, but does not 1974 change proc-bind-var. */ 1975 // Figure the value of proc-bind-var for the child threads. 1976 if ((level + 1 < __kmp_nested_proc_bind.used) && 1977 (__kmp_nested_proc_bind.bind_types[level + 1] != 1978 master_th->th.th_current_task->td_icvs.proc_bind)) { 1979 proc_bind_icv = __kmp_nested_proc_bind.bind_types[level + 1]; 1980 } 1981 } 1982 1983 // Reset for next parallel region 1984 master_th->th.th_set_proc_bind = proc_bind_default; 1985 #endif /* OMP_40_ENABLED */ 1986 1987 if ((nthreads_icv > 0) 1988 #if OMP_40_ENABLED 1989 || (proc_bind_icv != proc_bind_default) 1990 #endif /* OMP_40_ENABLED */ 1991 ) { 1992 kmp_internal_control_t new_icvs; 1993 copy_icvs(&new_icvs, &master_th->th.th_current_task->td_icvs); 1994 new_icvs.next = NULL; 1995 if (nthreads_icv > 0) { 1996 new_icvs.nproc = nthreads_icv; 1997 } 1998 1999 #if OMP_40_ENABLED 2000 if (proc_bind_icv != proc_bind_default) { 2001 new_icvs.proc_bind = proc_bind_icv; 2002 } 2003 #endif /* OMP_40_ENABLED */ 2004 2005 /* allocate a new parallel team */ 2006 KF_TRACE(10, ("__kmp_fork_call: before __kmp_allocate_team\n")); 2007 team = __kmp_allocate_team(root, nthreads, nthreads, 2008 #if OMPT_SUPPORT 2009 ompt_parallel_id, 2010 #endif 2011 #if OMP_40_ENABLED 2012 proc_bind, 2013 #endif 2014 &new_icvs, argc USE_NESTED_HOT_ARG(master_th)); 2015 } else { 2016 /* allocate a new parallel team */ 2017 KF_TRACE(10, ("__kmp_fork_call: before __kmp_allocate_team\n")); 2018 team = __kmp_allocate_team(root, nthreads, nthreads, 2019 #if OMPT_SUPPORT 2020 ompt_parallel_id, 2021 #endif 2022 #if OMP_40_ENABLED 2023 proc_bind, 2024 #endif 2025 &master_th->th.th_current_task->td_icvs, 2026 argc USE_NESTED_HOT_ARG(master_th)); 2027 } 2028 KF_TRACE( 2029 10, ("__kmp_fork_call: after __kmp_allocate_team - team = %p\n", team)); 2030 2031 /* setup the new team */ 2032 KMP_CHECK_UPDATE(team->t.t_master_tid, master_tid); 2033 KMP_CHECK_UPDATE(team->t.t_master_this_cons, master_this_cons); 2034 KMP_CHECK_UPDATE(team->t.t_ident, loc); 2035 KMP_CHECK_UPDATE(team->t.t_parent, parent_team); 2036 KMP_CHECK_UPDATE_SYNC(team->t.t_pkfn, microtask); 2037 #if OMPT_SUPPORT 2038 KMP_CHECK_UPDATE_SYNC(team->t.ompt_team_info.microtask, unwrapped_task); 2039 #endif 2040 KMP_CHECK_UPDATE(team->t.t_invoke, invoker); // TODO move to root, maybe 2041 // TODO: parent_team->t.t_level == INT_MAX ??? 2042 #if OMP_40_ENABLED 2043 if (!master_th->th.th_teams_microtask || level > teams_level) { 2044 #endif /* OMP_40_ENABLED */ 2045 int new_level = parent_team->t.t_level + 1; 2046 KMP_CHECK_UPDATE(team->t.t_level, new_level); 2047 new_level = parent_team->t.t_active_level + 1; 2048 KMP_CHECK_UPDATE(team->t.t_active_level, new_level); 2049 #if OMP_40_ENABLED 2050 } else { 2051 // AC: Do not increase parallel level at start of the teams construct 2052 int new_level = parent_team->t.t_level; 2053 KMP_CHECK_UPDATE(team->t.t_level, new_level); 2054 new_level = parent_team->t.t_active_level; 2055 KMP_CHECK_UPDATE(team->t.t_active_level, new_level); 2056 } 2057 #endif /* OMP_40_ENABLED */ 2058 kmp_r_sched_t new_sched = get__sched_2(parent_team, master_tid); 2059 if (team->t.t_sched.r_sched_type != new_sched.r_sched_type || 2060 team->t.t_sched.chunk != new_sched.chunk) 2061 team->t.t_sched = 2062 new_sched; // set master's schedule as new run-time schedule 2063 2064 #if OMP_40_ENABLED 2065 KMP_CHECK_UPDATE(team->t.t_cancel_request, cancel_noreq); 2066 #endif 2067 2068 // Update the floating point rounding in the team if required. 2069 propagateFPControl(team); 2070 2071 if (__kmp_tasking_mode != tskm_immediate_exec) { 2072 // Set master's task team to team's task team. Unless this is hot team, it 2073 // should be NULL. 2074 KMP_DEBUG_ASSERT(master_th->th.th_task_team == 2075 parent_team->t.t_task_team[master_th->th.th_task_state]); 2076 KA_TRACE(20, ("__kmp_fork_call: Master T#%d pushing task_team %p / team " 2077 "%p, new task_team %p / team %p\n", 2078 __kmp_gtid_from_thread(master_th), 2079 master_th->th.th_task_team, parent_team, 2080 team->t.t_task_team[master_th->th.th_task_state], team)); 2081 2082 if (active_level || master_th->th.th_task_team) { 2083 // Take a memo of master's task_state 2084 KMP_DEBUG_ASSERT(master_th->th.th_task_state_memo_stack); 2085 if (master_th->th.th_task_state_top >= 2086 master_th->th.th_task_state_stack_sz) { // increase size 2087 kmp_uint32 new_size = 2 * master_th->th.th_task_state_stack_sz; 2088 kmp_uint8 *old_stack, *new_stack; 2089 kmp_uint32 i; 2090 new_stack = (kmp_uint8 *)__kmp_allocate(new_size); 2091 for (i = 0; i < master_th->th.th_task_state_stack_sz; ++i) { 2092 new_stack[i] = master_th->th.th_task_state_memo_stack[i]; 2093 } 2094 for (i = master_th->th.th_task_state_stack_sz; i < new_size; 2095 ++i) { // zero-init rest of stack 2096 new_stack[i] = 0; 2097 } 2098 old_stack = master_th->th.th_task_state_memo_stack; 2099 master_th->th.th_task_state_memo_stack = new_stack; 2100 master_th->th.th_task_state_stack_sz = new_size; 2101 __kmp_free(old_stack); 2102 } 2103 // Store master's task_state on stack 2104 master_th->th 2105 .th_task_state_memo_stack[master_th->th.th_task_state_top] = 2106 master_th->th.th_task_state; 2107 master_th->th.th_task_state_top++; 2108 #if KMP_NESTED_HOT_TEAMS 2109 if (team == master_th->th.th_hot_teams[active_level].hot_team) { 2110 // Restore master's nested state if nested hot team 2111 master_th->th.th_task_state = 2112 master_th->th 2113 .th_task_state_memo_stack[master_th->th.th_task_state_top]; 2114 } else { 2115 #endif 2116 master_th->th.th_task_state = 0; 2117 #if KMP_NESTED_HOT_TEAMS 2118 } 2119 #endif 2120 } 2121 #if !KMP_NESTED_HOT_TEAMS 2122 KMP_DEBUG_ASSERT((master_th->th.th_task_team == NULL) || 2123 (team == root->r.r_hot_team)); 2124 #endif 2125 } 2126 2127 KA_TRACE( 2128 20, 2129 ("__kmp_fork_call: T#%d(%d:%d)->(%d:0) created a team of %d threads\n", 2130 gtid, parent_team->t.t_id, team->t.t_master_tid, team->t.t_id, 2131 team->t.t_nproc)); 2132 KMP_DEBUG_ASSERT(team != root->r.r_hot_team || 2133 (team->t.t_master_tid == 0 && 2134 (team->t.t_parent == root->r.r_root_team || 2135 team->t.t_parent->t.t_serialized))); 2136 KMP_MB(); 2137 2138 /* now, setup the arguments */ 2139 argv = (void **)team->t.t_argv; 2140 #if OMP_40_ENABLED 2141 if (ap) { 2142 #endif /* OMP_40_ENABLED */ 2143 for (i = argc - 1; i >= 0; --i) { 2144 // TODO: revert workaround for Intel(R) 64 tracker #96 2145 #if (KMP_ARCH_X86_64 || KMP_ARCH_ARM || KMP_ARCH_AARCH64) && KMP_OS_LINUX 2146 void *new_argv = va_arg(*ap, void *); 2147 #else 2148 void *new_argv = va_arg(ap, void *); 2149 #endif 2150 KMP_CHECK_UPDATE(*argv, new_argv); 2151 argv++; 2152 } 2153 #if OMP_40_ENABLED 2154 } else { 2155 for (i = 0; i < argc; ++i) { 2156 // Get args from parent team for teams construct 2157 KMP_CHECK_UPDATE(argv[i], team->t.t_parent->t.t_argv[i]); 2158 } 2159 } 2160 #endif /* OMP_40_ENABLED */ 2161 2162 /* now actually fork the threads */ 2163 KMP_CHECK_UPDATE(team->t.t_master_active, master_active); 2164 if (!root->r.r_active) // Only do assignment if it prevents cache ping-pong 2165 root->r.r_active = TRUE; 2166 2167 __kmp_fork_team_threads(root, team, master_th, gtid); 2168 __kmp_setup_icv_copy(team, nthreads, 2169 &master_th->th.th_current_task->td_icvs, loc); 2170 2171 #if OMPT_SUPPORT 2172 master_th->th.ompt_thread_info.state = ompt_state_work_parallel; 2173 #endif 2174 2175 __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock); 2176 2177 #if USE_ITT_BUILD 2178 if (team->t.t_active_level == 1 // only report frames at level 1 2179 #if OMP_40_ENABLED 2180 && !master_th->th.th_teams_microtask // not in teams construct 2181 #endif /* OMP_40_ENABLED */ 2182 ) { 2183 #if USE_ITT_NOTIFY 2184 if ((__itt_frame_submit_v3_ptr || KMP_ITT_DEBUG) && 2185 (__kmp_forkjoin_frames_mode == 3 || 2186 __kmp_forkjoin_frames_mode == 1)) { 2187 kmp_uint64 tmp_time = 0; 2188 if (__itt_get_timestamp_ptr) 2189 tmp_time = __itt_get_timestamp(); 2190 // Internal fork - report frame begin 2191 master_th->th.th_frame_time = tmp_time; 2192 if (__kmp_forkjoin_frames_mode == 3) 2193 team->t.t_region_time = tmp_time; 2194 } else 2195 // only one notification scheme (either "submit" or "forking/joined", not both) 2196 #endif /* USE_ITT_NOTIFY */ 2197 if ((__itt_frame_begin_v3_ptr || KMP_ITT_DEBUG) && 2198 __kmp_forkjoin_frames && !__kmp_forkjoin_frames_mode) { 2199 // Mark start of "parallel" region for VTune. 2200 __kmp_itt_region_forking(gtid, team->t.t_nproc, 0); 2201 } 2202 } 2203 #endif /* USE_ITT_BUILD */ 2204 2205 /* now go on and do the work */ 2206 KMP_DEBUG_ASSERT(team == __kmp_threads[gtid]->th.th_team); 2207 KMP_MB(); 2208 KF_TRACE(10, 2209 ("__kmp_internal_fork : root=%p, team=%p, master_th=%p, gtid=%d\n", 2210 root, team, master_th, gtid)); 2211 2212 #if USE_ITT_BUILD 2213 if (__itt_stack_caller_create_ptr) { 2214 team->t.t_stack_id = 2215 __kmp_itt_stack_caller_create(); // create new stack stitching id 2216 // before entering fork barrier 2217 } 2218 #endif /* USE_ITT_BUILD */ 2219 2220 #if OMP_40_ENABLED 2221 // AC: skip __kmp_internal_fork at teams construct, let only master 2222 // threads execute 2223 if (ap) 2224 #endif /* OMP_40_ENABLED */ 2225 { 2226 __kmp_internal_fork(loc, gtid, team); 2227 KF_TRACE(10, ("__kmp_internal_fork : after : root=%p, team=%p, " 2228 "master_th=%p, gtid=%d\n", 2229 root, team, master_th, gtid)); 2230 } 2231 2232 if (call_context == fork_context_gnu) { 2233 KA_TRACE(20, ("__kmp_fork_call: parallel exit T#%d\n", gtid)); 2234 return TRUE; 2235 } 2236 2237 /* Invoke microtask for MASTER thread */ 2238 KA_TRACE(20, ("__kmp_fork_call: T#%d(%d:0) invoke microtask = %p\n", gtid, 2239 team->t.t_id, team->t.t_pkfn)); 2240 } // END of timer KMP_fork_call block 2241 2242 { 2243 KMP_TIME_PARTITIONED_BLOCK(OMP_parallel); 2244 KMP_SET_THREAD_STATE_BLOCK(IMPLICIT_TASK); 2245 if (!team->t.t_invoke(gtid)) { 2246 KMP_ASSERT2(0, "cannot invoke microtask for MASTER thread"); 2247 } 2248 } 2249 KA_TRACE(20, ("__kmp_fork_call: T#%d(%d:0) done microtask = %p\n", gtid, 2250 team->t.t_id, team->t.t_pkfn)); 2251 KMP_MB(); /* Flush all pending memory write invalidates. */ 2252 2253 KA_TRACE(20, ("__kmp_fork_call: parallel exit T#%d\n", gtid)); 2254 2255 #if OMPT_SUPPORT 2256 if (ompt_enabled) { 2257 master_th->th.ompt_thread_info.state = ompt_state_overhead; 2258 } 2259 #endif 2260 2261 return TRUE; 2262 } 2263 2264 #if OMPT_SUPPORT 2265 static inline void __kmp_join_restore_state(kmp_info_t *thread, 2266 kmp_team_t *team) { 2267 // restore state outside the region 2268 thread->th.ompt_thread_info.state = 2269 ((team->t.t_serialized) ? ompt_state_work_serial 2270 : ompt_state_work_parallel); 2271 } 2272 2273 static inline void __kmp_join_ompt(kmp_info_t *thread, kmp_team_t *team, 2274 ompt_parallel_id_t parallel_id, 2275 fork_context_e fork_context) { 2276 ompt_task_info_t *task_info = __ompt_get_taskinfo(0); 2277 if (ompt_callbacks.ompt_callback(ompt_event_parallel_end)) { 2278 ompt_callbacks.ompt_callback(ompt_event_parallel_end)( 2279 parallel_id, task_info->task_id, OMPT_INVOKER(fork_context)); 2280 } 2281 2282 task_info->frame.reenter_runtime_frame = NULL; 2283 __kmp_join_restore_state(thread, team); 2284 } 2285 #endif 2286 2287 void __kmp_join_call(ident_t *loc, int gtid 2288 #if OMPT_SUPPORT 2289 , 2290 enum fork_context_e fork_context 2291 #endif 2292 #if OMP_40_ENABLED 2293 , 2294 int exit_teams 2295 #endif /* OMP_40_ENABLED */ 2296 ) { 2297 KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(KMP_join_call); 2298 kmp_team_t *team; 2299 kmp_team_t *parent_team; 2300 kmp_info_t *master_th; 2301 kmp_root_t *root; 2302 int master_active; 2303 int i; 2304 2305 KA_TRACE(20, ("__kmp_join_call: enter T#%d\n", gtid)); 2306 2307 /* setup current data */ 2308 master_th = __kmp_threads[gtid]; 2309 root = master_th->th.th_root; 2310 team = master_th->th.th_team; 2311 parent_team = team->t.t_parent; 2312 2313 master_th->th.th_ident = loc; 2314 2315 #if OMPT_SUPPORT 2316 if (ompt_enabled) { 2317 master_th->th.ompt_thread_info.state = ompt_state_overhead; 2318 } 2319 #endif 2320 2321 #if KMP_DEBUG 2322 if (__kmp_tasking_mode != tskm_immediate_exec && !exit_teams) { 2323 KA_TRACE(20, ("__kmp_join_call: T#%d, old team = %p old task_team = %p, " 2324 "th_task_team = %p\n", 2325 __kmp_gtid_from_thread(master_th), team, 2326 team->t.t_task_team[master_th->th.th_task_state], 2327 master_th->th.th_task_team)); 2328 KMP_DEBUG_ASSERT(master_th->th.th_task_team == 2329 team->t.t_task_team[master_th->th.th_task_state]); 2330 } 2331 #endif 2332 2333 if (team->t.t_serialized) { 2334 #if OMP_40_ENABLED 2335 if (master_th->th.th_teams_microtask) { 2336 // We are in teams construct 2337 int level = team->t.t_level; 2338 int tlevel = master_th->th.th_teams_level; 2339 if (level == tlevel) { 2340 // AC: we haven't incremented it earlier at start of teams construct, 2341 // so do it here - at the end of teams construct 2342 team->t.t_level++; 2343 } else if (level == tlevel + 1) { 2344 // AC: we are exiting parallel inside teams, need to increment 2345 // serialization in order to restore it in the next call to 2346 // __kmpc_end_serialized_parallel 2347 team->t.t_serialized++; 2348 } 2349 } 2350 #endif /* OMP_40_ENABLED */ 2351 __kmpc_end_serialized_parallel(loc, gtid); 2352 2353 #if OMPT_SUPPORT 2354 if (ompt_enabled) { 2355 __kmp_join_restore_state(master_th, parent_team); 2356 } 2357 #endif 2358 2359 return; 2360 } 2361 2362 master_active = team->t.t_master_active; 2363 2364 #if OMP_40_ENABLED 2365 if (!exit_teams) 2366 #endif /* OMP_40_ENABLED */ 2367 { 2368 // AC: No barrier for internal teams at exit from teams construct. 2369 // But there is barrier for external team (league). 2370 __kmp_internal_join(loc, gtid, team); 2371 } 2372 #if OMP_40_ENABLED 2373 else { 2374 master_th->th.th_task_state = 2375 0; // AC: no tasking in teams (out of any parallel) 2376 } 2377 #endif /* OMP_40_ENABLED */ 2378 2379 KMP_MB(); 2380 2381 #if OMPT_SUPPORT 2382 ompt_parallel_id_t parallel_id = team->t.ompt_team_info.parallel_id; 2383 #endif 2384 2385 #if USE_ITT_BUILD 2386 if (__itt_stack_caller_create_ptr) { 2387 __kmp_itt_stack_caller_destroy( 2388 (__itt_caller)team->t 2389 .t_stack_id); // destroy the stack stitching id after join barrier 2390 } 2391 2392 // Mark end of "parallel" region for VTune. 2393 if (team->t.t_active_level == 1 2394 #if OMP_40_ENABLED 2395 && !master_th->th.th_teams_microtask /* not in teams construct */ 2396 #endif /* OMP_40_ENABLED */ 2397 ) { 2398 master_th->th.th_ident = loc; 2399 // only one notification scheme (either "submit" or "forking/joined", not 2400 // both) 2401 if ((__itt_frame_submit_v3_ptr || KMP_ITT_DEBUG) && 2402 __kmp_forkjoin_frames_mode == 3) 2403 __kmp_itt_frame_submit(gtid, team->t.t_region_time, 2404 master_th->th.th_frame_time, 0, loc, 2405 master_th->th.th_team_nproc, 1); 2406 else if ((__itt_frame_end_v3_ptr || KMP_ITT_DEBUG) && 2407 !__kmp_forkjoin_frames_mode && __kmp_forkjoin_frames) 2408 __kmp_itt_region_joined(gtid); 2409 } // active_level == 1 2410 #endif /* USE_ITT_BUILD */ 2411 2412 #if OMP_40_ENABLED 2413 if (master_th->th.th_teams_microtask && !exit_teams && 2414 team->t.t_pkfn != (microtask_t)__kmp_teams_master && 2415 team->t.t_level == master_th->th.th_teams_level + 1) { 2416 // AC: We need to leave the team structure intact at the end of parallel 2417 // inside the teams construct, so that at the next parallel same (hot) team 2418 // works, only adjust nesting levels 2419 2420 /* Decrement our nested depth level */ 2421 team->t.t_level--; 2422 team->t.t_active_level--; 2423 KMP_TEST_THEN_DEC32((kmp_int32 *)&root->r.r_in_parallel); 2424 2425 /* Restore number of threads in the team if needed */ 2426 if (master_th->th.th_team_nproc < master_th->th.th_teams_size.nth) { 2427 int old_num = master_th->th.th_team_nproc; 2428 int new_num = master_th->th.th_teams_size.nth; 2429 kmp_info_t **other_threads = team->t.t_threads; 2430 team->t.t_nproc = new_num; 2431 for (i = 0; i < old_num; ++i) { 2432 other_threads[i]->th.th_team_nproc = new_num; 2433 } 2434 // Adjust states of non-used threads of the team 2435 for (i = old_num; i < new_num; ++i) { 2436 // Re-initialize thread's barrier data. 2437 int b; 2438 kmp_balign_t *balign = other_threads[i]->th.th_bar; 2439 for (b = 0; b < bs_last_barrier; ++b) { 2440 balign[b].bb.b_arrived = team->t.t_bar[b].b_arrived; 2441 KMP_DEBUG_ASSERT(balign[b].bb.wait_flag != KMP_BARRIER_PARENT_FLAG); 2442 #if USE_DEBUGGER 2443 balign[b].bb.b_worker_arrived = team->t.t_bar[b].b_team_arrived; 2444 #endif 2445 } 2446 if (__kmp_tasking_mode != tskm_immediate_exec) { 2447 // Synchronize thread's task state 2448 other_threads[i]->th.th_task_state = master_th->th.th_task_state; 2449 } 2450 } 2451 } 2452 2453 #if OMPT_SUPPORT 2454 if (ompt_enabled) { 2455 __kmp_join_ompt(master_th, parent_team, parallel_id, fork_context); 2456 } 2457 #endif 2458 2459 return; 2460 } 2461 #endif /* OMP_40_ENABLED */ 2462 2463 /* do cleanup and restore the parent team */ 2464 master_th->th.th_info.ds.ds_tid = team->t.t_master_tid; 2465 master_th->th.th_local.this_construct = team->t.t_master_this_cons; 2466 2467 master_th->th.th_dispatch = &parent_team->t.t_dispatch[team->t.t_master_tid]; 2468 2469 /* jc: The following lock has instructions with REL and ACQ semantics, 2470 separating the parallel user code called in this parallel region 2471 from the serial user code called after this function returns. */ 2472 __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock); 2473 2474 #if OMP_40_ENABLED 2475 if (!master_th->th.th_teams_microtask || 2476 team->t.t_level > master_th->th.th_teams_level) 2477 #endif /* OMP_40_ENABLED */ 2478 { 2479 /* Decrement our nested depth level */ 2480 KMP_TEST_THEN_DEC32((kmp_int32 *)&root->r.r_in_parallel); 2481 } 2482 KMP_DEBUG_ASSERT(root->r.r_in_parallel >= 0); 2483 2484 #if OMPT_SUPPORT && OMPT_TRACE 2485 if (ompt_enabled) { 2486 ompt_task_info_t *task_info = __ompt_get_taskinfo(0); 2487 if (ompt_callbacks.ompt_callback(ompt_event_implicit_task_end)) { 2488 ompt_callbacks.ompt_callback(ompt_event_implicit_task_end)( 2489 parallel_id, task_info->task_id); 2490 } 2491 task_info->frame.exit_runtime_frame = NULL; 2492 task_info->task_id = 0; 2493 } 2494 #endif 2495 2496 KF_TRACE(10, ("__kmp_join_call1: T#%d, this_thread=%p team=%p\n", 0, 2497 master_th, team)); 2498 __kmp_pop_current_task_from_thread(master_th); 2499 2500 #if OMP_40_ENABLED && KMP_AFFINITY_SUPPORTED 2501 // Restore master thread's partition. 2502 master_th->th.th_first_place = team->t.t_first_place; 2503 master_th->th.th_last_place = team->t.t_last_place; 2504 #endif /* OMP_40_ENABLED */ 2505 2506 updateHWFPControl(team); 2507 2508 if (root->r.r_active != master_active) 2509 root->r.r_active = master_active; 2510 2511 __kmp_free_team(root, team USE_NESTED_HOT_ARG( 2512 master_th)); // this will free worker threads 2513 2514 /* this race was fun to find. make sure the following is in the critical 2515 region otherwise assertions may fail occasionally since the old team may be 2516 reallocated and the hierarchy appears inconsistent. it is actually safe to 2517 run and won't cause any bugs, but will cause those assertion failures. it's 2518 only one deref&assign so might as well put this in the critical region */ 2519 master_th->th.th_team = parent_team; 2520 master_th->th.th_team_nproc = parent_team->t.t_nproc; 2521 master_th->th.th_team_master = parent_team->t.t_threads[0]; 2522 master_th->th.th_team_serialized = parent_team->t.t_serialized; 2523 2524 /* restore serialized team, if need be */ 2525 if (parent_team->t.t_serialized && 2526 parent_team != master_th->th.th_serial_team && 2527 parent_team != root->r.r_root_team) { 2528 __kmp_free_team(root, 2529 master_th->th.th_serial_team USE_NESTED_HOT_ARG(NULL)); 2530 master_th->th.th_serial_team = parent_team; 2531 } 2532 2533 if (__kmp_tasking_mode != tskm_immediate_exec) { 2534 if (master_th->th.th_task_state_top > 2535 0) { // Restore task state from memo stack 2536 KMP_DEBUG_ASSERT(master_th->th.th_task_state_memo_stack); 2537 // Remember master's state if we re-use this nested hot team 2538 master_th->th.th_task_state_memo_stack[master_th->th.th_task_state_top] = 2539 master_th->th.th_task_state; 2540 --master_th->th.th_task_state_top; // pop 2541 // Now restore state at this level 2542 master_th->th.th_task_state = 2543 master_th->th 2544 .th_task_state_memo_stack[master_th->th.th_task_state_top]; 2545 } 2546 // Copy the task team from the parent team to the master thread 2547 master_th->th.th_task_team = 2548 parent_team->t.t_task_team[master_th->th.th_task_state]; 2549 KA_TRACE(20, 2550 ("__kmp_join_call: Master T#%d restoring task_team %p / team %p\n", 2551 __kmp_gtid_from_thread(master_th), master_th->th.th_task_team, 2552 parent_team)); 2553 } 2554 2555 // TODO: GEH - cannot do this assertion because root thread not set up as 2556 // executing 2557 // KMP_ASSERT( master_th->th.th_current_task->td_flags.executing == 0 ); 2558 master_th->th.th_current_task->td_flags.executing = 1; 2559 2560 __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock); 2561 2562 #if OMPT_SUPPORT 2563 if (ompt_enabled) { 2564 __kmp_join_ompt(master_th, parent_team, parallel_id, fork_context); 2565 } 2566 #endif 2567 2568 KMP_MB(); 2569 KA_TRACE(20, ("__kmp_join_call: exit T#%d\n", gtid)); 2570 } 2571 2572 /* Check whether we should push an internal control record onto the 2573 serial team stack. If so, do it. */ 2574 void __kmp_save_internal_controls(kmp_info_t *thread) { 2575 2576 if (thread->th.th_team != thread->th.th_serial_team) { 2577 return; 2578 } 2579 if (thread->th.th_team->t.t_serialized > 1) { 2580 int push = 0; 2581 2582 if (thread->th.th_team->t.t_control_stack_top == NULL) { 2583 push = 1; 2584 } else { 2585 if (thread->th.th_team->t.t_control_stack_top->serial_nesting_level != 2586 thread->th.th_team->t.t_serialized) { 2587 push = 1; 2588 } 2589 } 2590 if (push) { /* push a record on the serial team's stack */ 2591 kmp_internal_control_t *control = 2592 (kmp_internal_control_t *)__kmp_allocate( 2593 sizeof(kmp_internal_control_t)); 2594 2595 copy_icvs(control, &thread->th.th_current_task->td_icvs); 2596 2597 control->serial_nesting_level = thread->th.th_team->t.t_serialized; 2598 2599 control->next = thread->th.th_team->t.t_control_stack_top; 2600 thread->th.th_team->t.t_control_stack_top = control; 2601 } 2602 } 2603 } 2604 2605 /* Changes set_nproc */ 2606 void __kmp_set_num_threads(int new_nth, int gtid) { 2607 kmp_info_t *thread; 2608 kmp_root_t *root; 2609 2610 KF_TRACE(10, ("__kmp_set_num_threads: new __kmp_nth = %d\n", new_nth)); 2611 KMP_DEBUG_ASSERT(__kmp_init_serial); 2612 2613 if (new_nth < 1) 2614 new_nth = 1; 2615 else if (new_nth > __kmp_max_nth) 2616 new_nth = __kmp_max_nth; 2617 2618 KMP_COUNT_VALUE(OMP_set_numthreads, new_nth); 2619 thread = __kmp_threads[gtid]; 2620 2621 __kmp_save_internal_controls(thread); 2622 2623 set__nproc(thread, new_nth); 2624 2625 // If this omp_set_num_threads() call will cause the hot team size to be 2626 // reduced (in the absence of a num_threads clause), then reduce it now, 2627 // rather than waiting for the next parallel region. 2628 root = thread->th.th_root; 2629 if (__kmp_init_parallel && (!root->r.r_active) && 2630 (root->r.r_hot_team->t.t_nproc > new_nth) 2631 #if KMP_NESTED_HOT_TEAMS 2632 && __kmp_hot_teams_max_level && !__kmp_hot_teams_mode 2633 #endif 2634 ) { 2635 kmp_team_t *hot_team = root->r.r_hot_team; 2636 int f; 2637 2638 __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock); 2639 2640 // Release the extra threads we don't need any more. 2641 for (f = new_nth; f < hot_team->t.t_nproc; f++) { 2642 KMP_DEBUG_ASSERT(hot_team->t.t_threads[f] != NULL); 2643 if (__kmp_tasking_mode != tskm_immediate_exec) { 2644 // When decreasing team size, threads no longer in the team should unref 2645 // task team. 2646 hot_team->t.t_threads[f]->th.th_task_team = NULL; 2647 } 2648 __kmp_free_thread(hot_team->t.t_threads[f]); 2649 hot_team->t.t_threads[f] = NULL; 2650 } 2651 hot_team->t.t_nproc = new_nth; 2652 #if KMP_NESTED_HOT_TEAMS 2653 if (thread->th.th_hot_teams) { 2654 KMP_DEBUG_ASSERT(hot_team == thread->th.th_hot_teams[0].hot_team); 2655 thread->th.th_hot_teams[0].hot_team_nth = new_nth; 2656 } 2657 #endif 2658 2659 __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock); 2660 2661 // Update the t_nproc field in the threads that are still active. 2662 for (f = 0; f < new_nth; f++) { 2663 KMP_DEBUG_ASSERT(hot_team->t.t_threads[f] != NULL); 2664 hot_team->t.t_threads[f]->th.th_team_nproc = new_nth; 2665 } 2666 // Special flag in case omp_set_num_threads() call 2667 hot_team->t.t_size_changed = -1; 2668 } 2669 } 2670 2671 /* Changes max_active_levels */ 2672 void __kmp_set_max_active_levels(int gtid, int max_active_levels) { 2673 kmp_info_t *thread; 2674 2675 KF_TRACE(10, ("__kmp_set_max_active_levels: new max_active_levels for thread " 2676 "%d = (%d)\n", 2677 gtid, max_active_levels)); 2678 KMP_DEBUG_ASSERT(__kmp_init_serial); 2679 2680 // validate max_active_levels 2681 if (max_active_levels < 0) { 2682 KMP_WARNING(ActiveLevelsNegative, max_active_levels); 2683 // We ignore this call if the user has specified a negative value. 2684 // The current setting won't be changed. The last valid setting will be 2685 // used. A warning will be issued (if warnings are allowed as controlled by 2686 // the KMP_WARNINGS env var). 2687 KF_TRACE(10, ("__kmp_set_max_active_levels: the call is ignored: new " 2688 "max_active_levels for thread %d = (%d)\n", 2689 gtid, max_active_levels)); 2690 return; 2691 } 2692 if (max_active_levels <= KMP_MAX_ACTIVE_LEVELS_LIMIT) { 2693 // it's OK, the max_active_levels is within the valid range: [ 0; 2694 // KMP_MAX_ACTIVE_LEVELS_LIMIT ] 2695 // We allow a zero value. (implementation defined behavior) 2696 } else { 2697 KMP_WARNING(ActiveLevelsExceedLimit, max_active_levels, 2698 KMP_MAX_ACTIVE_LEVELS_LIMIT); 2699 max_active_levels = KMP_MAX_ACTIVE_LEVELS_LIMIT; 2700 // Current upper limit is MAX_INT. (implementation defined behavior) 2701 // If the input exceeds the upper limit, we correct the input to be the 2702 // upper limit. (implementation defined behavior) 2703 // Actually, the flow should never get here until we use MAX_INT limit. 2704 } 2705 KF_TRACE(10, ("__kmp_set_max_active_levels: after validation: new " 2706 "max_active_levels for thread %d = (%d)\n", 2707 gtid, max_active_levels)); 2708 2709 thread = __kmp_threads[gtid]; 2710 2711 __kmp_save_internal_controls(thread); 2712 2713 set__max_active_levels(thread, max_active_levels); 2714 } 2715 2716 /* Gets max_active_levels */ 2717 int __kmp_get_max_active_levels(int gtid) { 2718 kmp_info_t *thread; 2719 2720 KF_TRACE(10, ("__kmp_get_max_active_levels: thread %d\n", gtid)); 2721 KMP_DEBUG_ASSERT(__kmp_init_serial); 2722 2723 thread = __kmp_threads[gtid]; 2724 KMP_DEBUG_ASSERT(thread->th.th_current_task); 2725 KF_TRACE(10, ("__kmp_get_max_active_levels: thread %d, curtask=%p, " 2726 "curtask_maxaclevel=%d\n", 2727 gtid, thread->th.th_current_task, 2728 thread->th.th_current_task->td_icvs.max_active_levels)); 2729 return thread->th.th_current_task->td_icvs.max_active_levels; 2730 } 2731 2732 /* Changes def_sched_var ICV values (run-time schedule kind and chunk) */ 2733 void __kmp_set_schedule(int gtid, kmp_sched_t kind, int chunk) { 2734 kmp_info_t *thread; 2735 // kmp_team_t *team; 2736 2737 KF_TRACE(10, ("__kmp_set_schedule: new schedule for thread %d = (%d, %d)\n", 2738 gtid, (int)kind, chunk)); 2739 KMP_DEBUG_ASSERT(__kmp_init_serial); 2740 2741 // Check if the kind parameter is valid, correct if needed. 2742 // Valid parameters should fit in one of two intervals - standard or extended: 2743 // <lower>, <valid>, <upper_std>, <lower_ext>, <valid>, <upper> 2744 // 2008-01-25: 0, 1 - 4, 5, 100, 101 - 102, 103 2745 if (kind <= kmp_sched_lower || kind >= kmp_sched_upper || 2746 (kind <= kmp_sched_lower_ext && kind >= kmp_sched_upper_std)) { 2747 // TODO: Hint needs attention in case we change the default schedule. 2748 __kmp_msg(kmp_ms_warning, KMP_MSG(ScheduleKindOutOfRange, kind), 2749 KMP_HNT(DefaultScheduleKindUsed, "static, no chunk"), 2750 __kmp_msg_null); 2751 kind = kmp_sched_default; 2752 chunk = 0; // ignore chunk value in case of bad kind 2753 } 2754 2755 thread = __kmp_threads[gtid]; 2756 2757 __kmp_save_internal_controls(thread); 2758 2759 if (kind < kmp_sched_upper_std) { 2760 if (kind == kmp_sched_static && chunk < KMP_DEFAULT_CHUNK) { 2761 // differ static chunked vs. unchunked: chunk should be invalid to 2762 // indicate unchunked schedule (which is the default) 2763 thread->th.th_current_task->td_icvs.sched.r_sched_type = kmp_sch_static; 2764 } else { 2765 thread->th.th_current_task->td_icvs.sched.r_sched_type = 2766 __kmp_sch_map[kind - kmp_sched_lower - 1]; 2767 } 2768 } else { 2769 // __kmp_sch_map[ kind - kmp_sched_lower_ext + kmp_sched_upper_std - 2770 // kmp_sched_lower - 2 ]; 2771 thread->th.th_current_task->td_icvs.sched.r_sched_type = 2772 __kmp_sch_map[kind - kmp_sched_lower_ext + kmp_sched_upper_std - 2773 kmp_sched_lower - 2]; 2774 } 2775 if (kind == kmp_sched_auto || chunk < 1) { 2776 // ignore parameter chunk for schedule auto 2777 thread->th.th_current_task->td_icvs.sched.chunk = KMP_DEFAULT_CHUNK; 2778 } else { 2779 thread->th.th_current_task->td_icvs.sched.chunk = chunk; 2780 } 2781 } 2782 2783 /* Gets def_sched_var ICV values */ 2784 void __kmp_get_schedule(int gtid, kmp_sched_t *kind, int *chunk) { 2785 kmp_info_t *thread; 2786 enum sched_type th_type; 2787 2788 KF_TRACE(10, ("__kmp_get_schedule: thread %d\n", gtid)); 2789 KMP_DEBUG_ASSERT(__kmp_init_serial); 2790 2791 thread = __kmp_threads[gtid]; 2792 2793 th_type = thread->th.th_current_task->td_icvs.sched.r_sched_type; 2794 2795 switch (th_type) { 2796 case kmp_sch_static: 2797 case kmp_sch_static_greedy: 2798 case kmp_sch_static_balanced: 2799 *kind = kmp_sched_static; 2800 *chunk = 0; // chunk was not set, try to show this fact via zero value 2801 return; 2802 case kmp_sch_static_chunked: 2803 *kind = kmp_sched_static; 2804 break; 2805 case kmp_sch_dynamic_chunked: 2806 *kind = kmp_sched_dynamic; 2807 break; 2808 case kmp_sch_guided_chunked: 2809 case kmp_sch_guided_iterative_chunked: 2810 case kmp_sch_guided_analytical_chunked: 2811 *kind = kmp_sched_guided; 2812 break; 2813 case kmp_sch_auto: 2814 *kind = kmp_sched_auto; 2815 break; 2816 case kmp_sch_trapezoidal: 2817 *kind = kmp_sched_trapezoidal; 2818 break; 2819 #if KMP_STATIC_STEAL_ENABLED 2820 case kmp_sch_static_steal: 2821 *kind = kmp_sched_static_steal; 2822 break; 2823 #endif 2824 default: 2825 KMP_FATAL(UnknownSchedulingType, th_type); 2826 } 2827 2828 *chunk = thread->th.th_current_task->td_icvs.sched.chunk; 2829 } 2830 2831 int __kmp_get_ancestor_thread_num(int gtid, int level) { 2832 2833 int ii, dd; 2834 kmp_team_t *team; 2835 kmp_info_t *thr; 2836 2837 KF_TRACE(10, ("__kmp_get_ancestor_thread_num: thread %d %d\n", gtid, level)); 2838 KMP_DEBUG_ASSERT(__kmp_init_serial); 2839 2840 // validate level 2841 if (level == 0) 2842 return 0; 2843 if (level < 0) 2844 return -1; 2845 thr = __kmp_threads[gtid]; 2846 team = thr->th.th_team; 2847 ii = team->t.t_level; 2848 if (level > ii) 2849 return -1; 2850 2851 #if OMP_40_ENABLED 2852 if (thr->th.th_teams_microtask) { 2853 // AC: we are in teams region where multiple nested teams have same level 2854 int tlevel = thr->th.th_teams_level; // the level of the teams construct 2855 if (level <= 2856 tlevel) { // otherwise usual algorithm works (will not touch the teams) 2857 KMP_DEBUG_ASSERT(ii >= tlevel); 2858 // AC: As we need to pass by the teams league, we need to artificially 2859 // increase ii 2860 if (ii == tlevel) { 2861 ii += 2; // three teams have same level 2862 } else { 2863 ii++; // two teams have same level 2864 } 2865 } 2866 } 2867 #endif 2868 2869 if (ii == level) 2870 return __kmp_tid_from_gtid(gtid); 2871 2872 dd = team->t.t_serialized; 2873 level++; 2874 while (ii > level) { 2875 for (dd = team->t.t_serialized; (dd > 0) && (ii > level); dd--, ii--) { 2876 } 2877 if ((team->t.t_serialized) && (!dd)) { 2878 team = team->t.t_parent; 2879 continue; 2880 } 2881 if (ii > level) { 2882 team = team->t.t_parent; 2883 dd = team->t.t_serialized; 2884 ii--; 2885 } 2886 } 2887 2888 return (dd > 1) ? (0) : (team->t.t_master_tid); 2889 } 2890 2891 int __kmp_get_team_size(int gtid, int level) { 2892 2893 int ii, dd; 2894 kmp_team_t *team; 2895 kmp_info_t *thr; 2896 2897 KF_TRACE(10, ("__kmp_get_team_size: thread %d %d\n", gtid, level)); 2898 KMP_DEBUG_ASSERT(__kmp_init_serial); 2899 2900 // validate level 2901 if (level == 0) 2902 return 1; 2903 if (level < 0) 2904 return -1; 2905 thr = __kmp_threads[gtid]; 2906 team = thr->th.th_team; 2907 ii = team->t.t_level; 2908 if (level > ii) 2909 return -1; 2910 2911 #if OMP_40_ENABLED 2912 if (thr->th.th_teams_microtask) { 2913 // AC: we are in teams region where multiple nested teams have same level 2914 int tlevel = thr->th.th_teams_level; // the level of the teams construct 2915 if (level <= 2916 tlevel) { // otherwise usual algorithm works (will not touch the teams) 2917 KMP_DEBUG_ASSERT(ii >= tlevel); 2918 // AC: As we need to pass by the teams league, we need to artificially 2919 // increase ii 2920 if (ii == tlevel) { 2921 ii += 2; // three teams have same level 2922 } else { 2923 ii++; // two teams have same level 2924 } 2925 } 2926 } 2927 #endif 2928 2929 while (ii > level) { 2930 for (dd = team->t.t_serialized; (dd > 0) && (ii > level); dd--, ii--) { 2931 } 2932 if (team->t.t_serialized && (!dd)) { 2933 team = team->t.t_parent; 2934 continue; 2935 } 2936 if (ii > level) { 2937 team = team->t.t_parent; 2938 ii--; 2939 } 2940 } 2941 2942 return team->t.t_nproc; 2943 } 2944 2945 kmp_r_sched_t __kmp_get_schedule_global() { 2946 // This routine created because pairs (__kmp_sched, __kmp_chunk) and 2947 // (__kmp_static, __kmp_guided) may be changed by kmp_set_defaults 2948 // independently. So one can get the updated schedule here. 2949 2950 kmp_r_sched_t r_sched; 2951 2952 // create schedule from 4 globals: __kmp_sched, __kmp_chunk, __kmp_static, 2953 // __kmp_guided. __kmp_sched should keep original value, so that user can set 2954 // KMP_SCHEDULE multiple times, and thus have different run-time schedules in 2955 // different roots (even in OMP 2.5) 2956 if (__kmp_sched == kmp_sch_static) { 2957 r_sched.r_sched_type = __kmp_static; // replace STATIC with more detailed 2958 // schedule (balanced or greedy) 2959 } else if (__kmp_sched == kmp_sch_guided_chunked) { 2960 r_sched.r_sched_type = __kmp_guided; // replace GUIDED with more detailed 2961 // schedule (iterative or analytical) 2962 } else { 2963 r_sched.r_sched_type = 2964 __kmp_sched; // (STATIC_CHUNKED), or (DYNAMIC_CHUNKED), or other 2965 } 2966 2967 if (__kmp_chunk < KMP_DEFAULT_CHUNK) { // __kmp_chunk may be wrong here (if it 2968 // was not ever set) 2969 r_sched.chunk = KMP_DEFAULT_CHUNK; 2970 } else { 2971 r_sched.chunk = __kmp_chunk; 2972 } 2973 2974 return r_sched; 2975 } 2976 2977 /* Allocate (realloc == FALSE) * or reallocate (realloc == TRUE) 2978 at least argc number of *t_argv entries for the requested team. */ 2979 static void __kmp_alloc_argv_entries(int argc, kmp_team_t *team, int realloc) { 2980 2981 KMP_DEBUG_ASSERT(team); 2982 if (!realloc || argc > team->t.t_max_argc) { 2983 2984 KA_TRACE(100, ("__kmp_alloc_argv_entries: team %d: needed entries=%d, " 2985 "current entries=%d\n", 2986 team->t.t_id, argc, (realloc) ? team->t.t_max_argc : 0)); 2987 /* if previously allocated heap space for args, free them */ 2988 if (realloc && team->t.t_argv != &team->t.t_inline_argv[0]) 2989 __kmp_free((void *)team->t.t_argv); 2990 2991 if (argc <= KMP_INLINE_ARGV_ENTRIES) { 2992 /* use unused space in the cache line for arguments */ 2993 team->t.t_max_argc = KMP_INLINE_ARGV_ENTRIES; 2994 KA_TRACE(100, ("__kmp_alloc_argv_entries: team %d: inline allocate %d " 2995 "argv entries\n", 2996 team->t.t_id, team->t.t_max_argc)); 2997 team->t.t_argv = &team->t.t_inline_argv[0]; 2998 if (__kmp_storage_map) { 2999 __kmp_print_storage_map_gtid( 3000 -1, &team->t.t_inline_argv[0], 3001 &team->t.t_inline_argv[KMP_INLINE_ARGV_ENTRIES], 3002 (sizeof(void *) * KMP_INLINE_ARGV_ENTRIES), "team_%d.t_inline_argv", 3003 team->t.t_id); 3004 } 3005 } else { 3006 /* allocate space for arguments in the heap */ 3007 team->t.t_max_argc = (argc <= (KMP_MIN_MALLOC_ARGV_ENTRIES >> 1)) 3008 ? KMP_MIN_MALLOC_ARGV_ENTRIES 3009 : 2 * argc; 3010 KA_TRACE(100, ("__kmp_alloc_argv_entries: team %d: dynamic allocate %d " 3011 "argv entries\n", 3012 team->t.t_id, team->t.t_max_argc)); 3013 team->t.t_argv = 3014 (void **)__kmp_page_allocate(sizeof(void *) * team->t.t_max_argc); 3015 if (__kmp_storage_map) { 3016 __kmp_print_storage_map_gtid(-1, &team->t.t_argv[0], 3017 &team->t.t_argv[team->t.t_max_argc], 3018 sizeof(void *) * team->t.t_max_argc, 3019 "team_%d.t_argv", team->t.t_id); 3020 } 3021 } 3022 } 3023 } 3024 3025 static void __kmp_allocate_team_arrays(kmp_team_t *team, int max_nth) { 3026 int i; 3027 int num_disp_buff = max_nth > 1 ? __kmp_dispatch_num_buffers : 2; 3028 team->t.t_threads = 3029 (kmp_info_t **)__kmp_allocate(sizeof(kmp_info_t *) * max_nth); 3030 team->t.t_disp_buffer = (dispatch_shared_info_t *)__kmp_allocate( 3031 sizeof(dispatch_shared_info_t) * num_disp_buff); 3032 team->t.t_dispatch = 3033 (kmp_disp_t *)__kmp_allocate(sizeof(kmp_disp_t) * max_nth); 3034 team->t.t_implicit_task_taskdata = 3035 (kmp_taskdata_t *)__kmp_allocate(sizeof(kmp_taskdata_t) * max_nth); 3036 team->t.t_max_nproc = max_nth; 3037 3038 /* setup dispatch buffers */ 3039 for (i = 0; i < num_disp_buff; ++i) { 3040 team->t.t_disp_buffer[i].buffer_index = i; 3041 #if OMP_45_ENABLED 3042 team->t.t_disp_buffer[i].doacross_buf_idx = i; 3043 #endif 3044 } 3045 } 3046 3047 static void __kmp_free_team_arrays(kmp_team_t *team) { 3048 /* Note: this does not free the threads in t_threads (__kmp_free_threads) */ 3049 int i; 3050 for (i = 0; i < team->t.t_max_nproc; ++i) { 3051 if (team->t.t_dispatch[i].th_disp_buffer != NULL) { 3052 __kmp_free(team->t.t_dispatch[i].th_disp_buffer); 3053 team->t.t_dispatch[i].th_disp_buffer = NULL; 3054 }; // if 3055 }; // for 3056 __kmp_free(team->t.t_threads); 3057 __kmp_free(team->t.t_disp_buffer); 3058 __kmp_free(team->t.t_dispatch); 3059 __kmp_free(team->t.t_implicit_task_taskdata); 3060 team->t.t_threads = NULL; 3061 team->t.t_disp_buffer = NULL; 3062 team->t.t_dispatch = NULL; 3063 team->t.t_implicit_task_taskdata = 0; 3064 } 3065 3066 static void __kmp_reallocate_team_arrays(kmp_team_t *team, int max_nth) { 3067 kmp_info_t **oldThreads = team->t.t_threads; 3068 3069 __kmp_free(team->t.t_disp_buffer); 3070 __kmp_free(team->t.t_dispatch); 3071 __kmp_free(team->t.t_implicit_task_taskdata); 3072 __kmp_allocate_team_arrays(team, max_nth); 3073 3074 KMP_MEMCPY(team->t.t_threads, oldThreads, 3075 team->t.t_nproc * sizeof(kmp_info_t *)); 3076 3077 __kmp_free(oldThreads); 3078 } 3079 3080 static kmp_internal_control_t __kmp_get_global_icvs(void) { 3081 3082 kmp_r_sched_t r_sched = 3083 __kmp_get_schedule_global(); // get current state of scheduling globals 3084 3085 #if OMP_40_ENABLED 3086 KMP_DEBUG_ASSERT(__kmp_nested_proc_bind.used > 0); 3087 #endif /* OMP_40_ENABLED */ 3088 3089 kmp_internal_control_t g_icvs = { 3090 0, // int serial_nesting_level; //corresponds to value of th_team_serialized 3091 (kmp_int8)__kmp_dflt_nested, // int nested; //internal control 3092 // for nested parallelism (per thread) 3093 (kmp_int8)__kmp_global.g.g_dynamic, // internal control for dynamic 3094 // adjustment of threads (per thread) 3095 (kmp_int8)__kmp_env_blocktime, // int bt_set; //internal control for 3096 // whether blocktime is explicitly set 3097 __kmp_dflt_blocktime, // int blocktime; //internal control for blocktime 3098 #if KMP_USE_MONITOR 3099 __kmp_bt_intervals, // int bt_intervals; //internal control for blocktime 3100 // intervals 3101 #endif 3102 __kmp_dflt_team_nth, // int nproc; //internal control for # of threads for 3103 // next parallel region (per thread) 3104 // (use a max ub on value if __kmp_parallel_initialize not called yet) 3105 __kmp_dflt_max_active_levels, // int max_active_levels; //internal control 3106 // for max_active_levels 3107 r_sched, // kmp_r_sched_t sched; //internal control for runtime schedule 3108 // {sched,chunk} pair 3109 #if OMP_40_ENABLED 3110 __kmp_nested_proc_bind.bind_types[0], 3111 __kmp_default_device, 3112 #endif /* OMP_40_ENABLED */ 3113 NULL // struct kmp_internal_control *next; 3114 }; 3115 3116 return g_icvs; 3117 } 3118 3119 static kmp_internal_control_t __kmp_get_x_global_icvs(const kmp_team_t *team) { 3120 3121 kmp_internal_control_t gx_icvs; 3122 gx_icvs.serial_nesting_level = 3123 0; // probably =team->t.t_serial like in save_inter_controls 3124 copy_icvs(&gx_icvs, &team->t.t_threads[0]->th.th_current_task->td_icvs); 3125 gx_icvs.next = NULL; 3126 3127 return gx_icvs; 3128 } 3129 3130 static void __kmp_initialize_root(kmp_root_t *root) { 3131 int f; 3132 kmp_team_t *root_team; 3133 kmp_team_t *hot_team; 3134 int hot_team_max_nth; 3135 kmp_r_sched_t r_sched = 3136 __kmp_get_schedule_global(); // get current state of scheduling globals 3137 kmp_internal_control_t r_icvs = __kmp_get_global_icvs(); 3138 KMP_DEBUG_ASSERT(root); 3139 KMP_ASSERT(!root->r.r_begin); 3140 3141 /* setup the root state structure */ 3142 __kmp_init_lock(&root->r.r_begin_lock); 3143 root->r.r_begin = FALSE; 3144 root->r.r_active = FALSE; 3145 root->r.r_in_parallel = 0; 3146 root->r.r_blocktime = __kmp_dflt_blocktime; 3147 root->r.r_nested = __kmp_dflt_nested; 3148 root->r.r_cg_nthreads = 1; 3149 3150 /* setup the root team for this task */ 3151 /* allocate the root team structure */ 3152 KF_TRACE(10, ("__kmp_initialize_root: before root_team\n")); 3153 3154 root_team = 3155 __kmp_allocate_team(root, 3156 1, // new_nproc 3157 1, // max_nproc 3158 #if OMPT_SUPPORT 3159 0, // root parallel id 3160 #endif 3161 #if OMP_40_ENABLED 3162 __kmp_nested_proc_bind.bind_types[0], 3163 #endif 3164 &r_icvs, 3165 0 // argc 3166 USE_NESTED_HOT_ARG(NULL) // master thread is unknown 3167 ); 3168 #if USE_DEBUGGER 3169 // Non-NULL value should be assigned to make the debugger display the root 3170 // team. 3171 TCW_SYNC_PTR(root_team->t.t_pkfn, (microtask_t)(~0)); 3172 #endif 3173 3174 KF_TRACE(10, ("__kmp_initialize_root: after root_team = %p\n", root_team)); 3175 3176 root->r.r_root_team = root_team; 3177 root_team->t.t_control_stack_top = NULL; 3178 3179 /* initialize root team */ 3180 root_team->t.t_threads[0] = NULL; 3181 root_team->t.t_nproc = 1; 3182 root_team->t.t_serialized = 1; 3183 // TODO???: root_team->t.t_max_active_levels = __kmp_dflt_max_active_levels; 3184 root_team->t.t_sched.r_sched_type = r_sched.r_sched_type; 3185 root_team->t.t_sched.chunk = r_sched.chunk; 3186 KA_TRACE( 3187 20, 3188 ("__kmp_initialize_root: init root team %d arrived: join=%u, plain=%u\n", 3189 root_team->t.t_id, KMP_INIT_BARRIER_STATE, KMP_INIT_BARRIER_STATE)); 3190 3191 /* setup the hot team for this task */ 3192 /* allocate the hot team structure */ 3193 KF_TRACE(10, ("__kmp_initialize_root: before hot_team\n")); 3194 3195 hot_team = 3196 __kmp_allocate_team(root, 3197 1, // new_nproc 3198 __kmp_dflt_team_nth_ub * 2, // max_nproc 3199 #if OMPT_SUPPORT 3200 0, // root parallel id 3201 #endif 3202 #if OMP_40_ENABLED 3203 __kmp_nested_proc_bind.bind_types[0], 3204 #endif 3205 &r_icvs, 3206 0 // argc 3207 USE_NESTED_HOT_ARG(NULL) // master thread is unknown 3208 ); 3209 KF_TRACE(10, ("__kmp_initialize_root: after hot_team = %p\n", hot_team)); 3210 3211 root->r.r_hot_team = hot_team; 3212 root_team->t.t_control_stack_top = NULL; 3213 3214 /* first-time initialization */ 3215 hot_team->t.t_parent = root_team; 3216 3217 /* initialize hot team */ 3218 hot_team_max_nth = hot_team->t.t_max_nproc; 3219 for (f = 0; f < hot_team_max_nth; ++f) { 3220 hot_team->t.t_threads[f] = NULL; 3221 }; // for 3222 hot_team->t.t_nproc = 1; 3223 // TODO???: hot_team->t.t_max_active_levels = __kmp_dflt_max_active_levels; 3224 hot_team->t.t_sched.r_sched_type = r_sched.r_sched_type; 3225 hot_team->t.t_sched.chunk = r_sched.chunk; 3226 hot_team->t.t_size_changed = 0; 3227 } 3228 3229 #ifdef KMP_DEBUG 3230 3231 typedef struct kmp_team_list_item { 3232 kmp_team_p const *entry; 3233 struct kmp_team_list_item *next; 3234 } kmp_team_list_item_t; 3235 typedef kmp_team_list_item_t *kmp_team_list_t; 3236 3237 static void __kmp_print_structure_team_accum( // Add team to list of teams. 3238 kmp_team_list_t list, // List of teams. 3239 kmp_team_p const *team // Team to add. 3240 ) { 3241 3242 // List must terminate with item where both entry and next are NULL. 3243 // Team is added to the list only once. 3244 // List is sorted in ascending order by team id. 3245 // Team id is *not* a key. 3246 3247 kmp_team_list_t l; 3248 3249 KMP_DEBUG_ASSERT(list != NULL); 3250 if (team == NULL) { 3251 return; 3252 }; // if 3253 3254 __kmp_print_structure_team_accum(list, team->t.t_parent); 3255 __kmp_print_structure_team_accum(list, team->t.t_next_pool); 3256 3257 // Search list for the team. 3258 l = list; 3259 while (l->next != NULL && l->entry != team) { 3260 l = l->next; 3261 }; // while 3262 if (l->next != NULL) { 3263 return; // Team has been added before, exit. 3264 }; // if 3265 3266 // Team is not found. Search list again for insertion point. 3267 l = list; 3268 while (l->next != NULL && l->entry->t.t_id <= team->t.t_id) { 3269 l = l->next; 3270 }; // while 3271 3272 // Insert team. 3273 { 3274 kmp_team_list_item_t *item = (kmp_team_list_item_t *)KMP_INTERNAL_MALLOC( 3275 sizeof(kmp_team_list_item_t)); 3276 *item = *l; 3277 l->entry = team; 3278 l->next = item; 3279 } 3280 } 3281 3282 static void __kmp_print_structure_team(char const *title, kmp_team_p const *team 3283 3284 ) { 3285 __kmp_printf("%s", title); 3286 if (team != NULL) { 3287 __kmp_printf("%2x %p\n", team->t.t_id, team); 3288 } else { 3289 __kmp_printf(" - (nil)\n"); 3290 }; // if 3291 } 3292 3293 static void __kmp_print_structure_thread(char const *title, 3294 kmp_info_p const *thread) { 3295 __kmp_printf("%s", title); 3296 if (thread != NULL) { 3297 __kmp_printf("%2d %p\n", thread->th.th_info.ds.ds_gtid, thread); 3298 } else { 3299 __kmp_printf(" - (nil)\n"); 3300 }; // if 3301 } 3302 3303 void __kmp_print_structure(void) { 3304 3305 kmp_team_list_t list; 3306 3307 // Initialize list of teams. 3308 list = 3309 (kmp_team_list_item_t *)KMP_INTERNAL_MALLOC(sizeof(kmp_team_list_item_t)); 3310 list->entry = NULL; 3311 list->next = NULL; 3312 3313 __kmp_printf("\n------------------------------\nGlobal Thread " 3314 "Table\n------------------------------\n"); 3315 { 3316 int gtid; 3317 for (gtid = 0; gtid < __kmp_threads_capacity; ++gtid) { 3318 __kmp_printf("%2d", gtid); 3319 if (__kmp_threads != NULL) { 3320 __kmp_printf(" %p", __kmp_threads[gtid]); 3321 }; // if 3322 if (__kmp_root != NULL) { 3323 __kmp_printf(" %p", __kmp_root[gtid]); 3324 }; // if 3325 __kmp_printf("\n"); 3326 }; // for gtid 3327 } 3328 3329 // Print out __kmp_threads array. 3330 __kmp_printf("\n------------------------------\nThreads\n--------------------" 3331 "----------\n"); 3332 if (__kmp_threads != NULL) { 3333 int gtid; 3334 for (gtid = 0; gtid < __kmp_threads_capacity; ++gtid) { 3335 kmp_info_t const *thread = __kmp_threads[gtid]; 3336 if (thread != NULL) { 3337 __kmp_printf("GTID %2d %p:\n", gtid, thread); 3338 __kmp_printf(" Our Root: %p\n", thread->th.th_root); 3339 __kmp_print_structure_team(" Our Team: ", thread->th.th_team); 3340 __kmp_print_structure_team(" Serial Team: ", 3341 thread->th.th_serial_team); 3342 __kmp_printf(" Threads: %2d\n", thread->th.th_team_nproc); 3343 __kmp_print_structure_thread(" Master: ", 3344 thread->th.th_team_master); 3345 __kmp_printf(" Serialized?: %2d\n", thread->th.th_team_serialized); 3346 __kmp_printf(" Set NProc: %2d\n", thread->th.th_set_nproc); 3347 #if OMP_40_ENABLED 3348 __kmp_printf(" Set Proc Bind: %2d\n", thread->th.th_set_proc_bind); 3349 #endif 3350 __kmp_print_structure_thread(" Next in pool: ", 3351 thread->th.th_next_pool); 3352 __kmp_printf("\n"); 3353 __kmp_print_structure_team_accum(list, thread->th.th_team); 3354 __kmp_print_structure_team_accum(list, thread->th.th_serial_team); 3355 }; // if 3356 }; // for gtid 3357 } else { 3358 __kmp_printf("Threads array is not allocated.\n"); 3359 }; // if 3360 3361 // Print out __kmp_root array. 3362 __kmp_printf("\n------------------------------\nUbers\n----------------------" 3363 "--------\n"); 3364 if (__kmp_root != NULL) { 3365 int gtid; 3366 for (gtid = 0; gtid < __kmp_threads_capacity; ++gtid) { 3367 kmp_root_t const *root = __kmp_root[gtid]; 3368 if (root != NULL) { 3369 __kmp_printf("GTID %2d %p:\n", gtid, root); 3370 __kmp_print_structure_team(" Root Team: ", root->r.r_root_team); 3371 __kmp_print_structure_team(" Hot Team: ", root->r.r_hot_team); 3372 __kmp_print_structure_thread(" Uber Thread: ", 3373 root->r.r_uber_thread); 3374 __kmp_printf(" Active?: %2d\n", root->r.r_active); 3375 __kmp_printf(" Nested?: %2d\n", root->r.r_nested); 3376 __kmp_printf(" In Parallel: %2d\n", root->r.r_in_parallel); 3377 __kmp_printf("\n"); 3378 __kmp_print_structure_team_accum(list, root->r.r_root_team); 3379 __kmp_print_structure_team_accum(list, root->r.r_hot_team); 3380 }; // if 3381 }; // for gtid 3382 } else { 3383 __kmp_printf("Ubers array is not allocated.\n"); 3384 }; // if 3385 3386 __kmp_printf("\n------------------------------\nTeams\n----------------------" 3387 "--------\n"); 3388 while (list->next != NULL) { 3389 kmp_team_p const *team = list->entry; 3390 int i; 3391 __kmp_printf("Team %2x %p:\n", team->t.t_id, team); 3392 __kmp_print_structure_team(" Parent Team: ", team->t.t_parent); 3393 __kmp_printf(" Master TID: %2d\n", team->t.t_master_tid); 3394 __kmp_printf(" Max threads: %2d\n", team->t.t_max_nproc); 3395 __kmp_printf(" Levels of serial: %2d\n", team->t.t_serialized); 3396 __kmp_printf(" Number threads: %2d\n", team->t.t_nproc); 3397 for (i = 0; i < team->t.t_nproc; ++i) { 3398 __kmp_printf(" Thread %2d: ", i); 3399 __kmp_print_structure_thread("", team->t.t_threads[i]); 3400 }; // for i 3401 __kmp_print_structure_team(" Next in pool: ", team->t.t_next_pool); 3402 __kmp_printf("\n"); 3403 list = list->next; 3404 }; // while 3405 3406 // Print out __kmp_thread_pool and __kmp_team_pool. 3407 __kmp_printf("\n------------------------------\nPools\n----------------------" 3408 "--------\n"); 3409 __kmp_print_structure_thread("Thread pool: ", 3410 CCAST(kmp_info_t *, __kmp_thread_pool)); 3411 __kmp_print_structure_team("Team pool: ", 3412 CCAST(kmp_team_t *, __kmp_team_pool)); 3413 __kmp_printf("\n"); 3414 3415 // Free team list. 3416 while (list != NULL) { 3417 kmp_team_list_item_t *item = list; 3418 list = list->next; 3419 KMP_INTERNAL_FREE(item); 3420 }; // while 3421 } 3422 3423 #endif 3424 3425 //--------------------------------------------------------------------------- 3426 // Stuff for per-thread fast random number generator 3427 // Table of primes 3428 static const unsigned __kmp_primes[] = { 3429 0x9e3779b1, 0xffe6cc59, 0x2109f6dd, 0x43977ab5, 0xba5703f5, 0xb495a877, 3430 0xe1626741, 0x79695e6b, 0xbc98c09f, 0xd5bee2b3, 0x287488f9, 0x3af18231, 3431 0x9677cd4d, 0xbe3a6929, 0xadc6a877, 0xdcf0674b, 0xbe4d6fe9, 0x5f15e201, 3432 0x99afc3fd, 0xf3f16801, 0xe222cfff, 0x24ba5fdb, 0x0620452d, 0x79f149e3, 3433 0xc8b93f49, 0x972702cd, 0xb07dd827, 0x6c97d5ed, 0x085a3d61, 0x46eb5ea7, 3434 0x3d9910ed, 0x2e687b5b, 0x29609227, 0x6eb081f1, 0x0954c4e1, 0x9d114db9, 3435 0x542acfa9, 0xb3e6bd7b, 0x0742d917, 0xe9f3ffa7, 0x54581edb, 0xf2480f45, 3436 0x0bb9288f, 0xef1affc7, 0x85fa0ca7, 0x3ccc14db, 0xe6baf34b, 0x343377f7, 3437 0x5ca19031, 0xe6d9293b, 0xf0a9f391, 0x5d2e980b, 0xfc411073, 0xc3749363, 3438 0xb892d829, 0x3549366b, 0x629750ad, 0xb98294e5, 0x892d9483, 0xc235baf3, 3439 0x3d2402a3, 0x6bdef3c9, 0xbec333cd, 0x40c9520f}; 3440 3441 //--------------------------------------------------------------------------- 3442 // __kmp_get_random: Get a random number using a linear congruential method. 3443 unsigned short __kmp_get_random(kmp_info_t *thread) { 3444 unsigned x = thread->th.th_x; 3445 unsigned short r = x >> 16; 3446 3447 thread->th.th_x = x * thread->th.th_a + 1; 3448 3449 KA_TRACE(30, ("__kmp_get_random: THREAD: %d, RETURN: %u\n", 3450 thread->th.th_info.ds.ds_tid, r)); 3451 3452 return r; 3453 } 3454 //-------------------------------------------------------- 3455 // __kmp_init_random: Initialize a random number generator 3456 void __kmp_init_random(kmp_info_t *thread) { 3457 unsigned seed = thread->th.th_info.ds.ds_tid; 3458 3459 thread->th.th_a = 3460 __kmp_primes[seed % (sizeof(__kmp_primes) / sizeof(__kmp_primes[0]))]; 3461 thread->th.th_x = (seed + 1) * thread->th.th_a + 1; 3462 KA_TRACE(30, 3463 ("__kmp_init_random: THREAD: %u; A: %u\n", seed, thread->th.th_a)); 3464 } 3465 3466 #if KMP_OS_WINDOWS 3467 /* reclaim array entries for root threads that are already dead, returns number 3468 * reclaimed */ 3469 static int __kmp_reclaim_dead_roots(void) { 3470 int i, r = 0; 3471 3472 for (i = 0; i < __kmp_threads_capacity; ++i) { 3473 if (KMP_UBER_GTID(i) && 3474 !__kmp_still_running((kmp_info_t *)TCR_SYNC_PTR(__kmp_threads[i])) && 3475 !__kmp_root[i] 3476 ->r.r_active) { // AC: reclaim only roots died in non-active state 3477 r += __kmp_unregister_root_other_thread(i); 3478 } 3479 } 3480 return r; 3481 } 3482 #endif 3483 3484 /* This function attempts to create free entries in __kmp_threads and 3485 __kmp_root, and returns the number of free entries generated. 3486 3487 For Windows* OS static library, the first mechanism used is to reclaim array 3488 entries for root threads that are already dead. 3489 3490 On all platforms, expansion is attempted on the arrays __kmp_threads_ and 3491 __kmp_root, with appropriate update to __kmp_threads_capacity. Array 3492 capacity is increased by doubling with clipping to __kmp_tp_capacity, if 3493 threadprivate cache array has been created. Synchronization with 3494 __kmpc_threadprivate_cached is done using __kmp_tp_cached_lock. 3495 3496 After any dead root reclamation, if the clipping value allows array expansion 3497 to result in the generation of a total of nWish free slots, the function does 3498 that expansion. If not, but the clipping value allows array expansion to 3499 result in the generation of a total of nNeed free slots, the function does 3500 that expansion. Otherwise, nothing is done beyond the possible initial root 3501 thread reclamation. However, if nNeed is zero, a best-effort attempt is made 3502 to fulfil nWish as far as possible, i.e. the function will attempt to create 3503 as many free slots as possible up to nWish. 3504 3505 If any argument is negative, the behavior is undefined. */ 3506 static int __kmp_expand_threads(int nWish, int nNeed) { 3507 int added = 0; 3508 int old_tp_cached; 3509 int __kmp_actual_max_nth; 3510 3511 if (nNeed > nWish) /* normalize the arguments */ 3512 nWish = nNeed; 3513 #if KMP_OS_WINDOWS && !defined KMP_DYNAMIC_LIB 3514 /* only for Windows static library */ 3515 /* reclaim array entries for root threads that are already dead */ 3516 added = __kmp_reclaim_dead_roots(); 3517 3518 if (nNeed) { 3519 nNeed -= added; 3520 if (nNeed < 0) 3521 nNeed = 0; 3522 } 3523 if (nWish) { 3524 nWish -= added; 3525 if (nWish < 0) 3526 nWish = 0; 3527 } 3528 #endif 3529 if (nWish <= 0) 3530 return added; 3531 3532 while (1) { 3533 int nTarget; 3534 int minimumRequiredCapacity; 3535 int newCapacity; 3536 kmp_info_t **newThreads; 3537 kmp_root_t **newRoot; 3538 3539 // Note that __kmp_threads_capacity is not bounded by __kmp_max_nth. If 3540 // __kmp_max_nth is set to some value less than __kmp_sys_max_nth by the 3541 // user via KMP_DEVICE_THREAD_LIMIT, then __kmp_threads_capacity may become 3542 // > __kmp_max_nth in one of two ways: 3543 // 3544 // 1) The initialization thread (gtid = 0) exits. __kmp_threads[0] 3545 // may not be resused by another thread, so we may need to increase 3546 // __kmp_threads_capacity to __kmp_max_nth + 1. 3547 // 3548 // 2) New foreign root(s) are encountered. We always register new foreign 3549 // roots. This may cause a smaller # of threads to be allocated at 3550 // subsequent parallel regions, but the worker threads hang around (and 3551 // eventually go to sleep) and need slots in the __kmp_threads[] array. 3552 // 3553 // Anyway, that is the reason for moving the check to see if 3554 // __kmp_max_nth was exceeded into __kmp_reserve_threads() 3555 // instead of having it performed here. -BB 3556 old_tp_cached = __kmp_tp_cached; 3557 __kmp_actual_max_nth = 3558 old_tp_cached ? __kmp_tp_capacity : __kmp_sys_max_nth; 3559 KMP_DEBUG_ASSERT(__kmp_actual_max_nth >= __kmp_threads_capacity); 3560 3561 /* compute expansion headroom to check if we can expand and whether to aim 3562 for nWish or nNeed */ 3563 nTarget = nWish; 3564 if (__kmp_actual_max_nth - __kmp_threads_capacity < nTarget) { 3565 /* can't fulfil nWish, so try nNeed */ 3566 if (nNeed) { 3567 nTarget = nNeed; 3568 if (__kmp_actual_max_nth - __kmp_threads_capacity < nTarget) { 3569 /* possible expansion too small -- give up */ 3570 break; 3571 } 3572 } else { 3573 /* best-effort */ 3574 nTarget = __kmp_actual_max_nth - __kmp_threads_capacity; 3575 if (!nTarget) { 3576 /* can expand at all -- give up */ 3577 break; 3578 } 3579 } 3580 } 3581 minimumRequiredCapacity = __kmp_threads_capacity + nTarget; 3582 3583 newCapacity = __kmp_threads_capacity; 3584 do { 3585 newCapacity = newCapacity <= (__kmp_actual_max_nth >> 1) 3586 ? (newCapacity << 1) 3587 : __kmp_actual_max_nth; 3588 } while (newCapacity < minimumRequiredCapacity); 3589 newThreads = (kmp_info_t **)__kmp_allocate( 3590 (sizeof(kmp_info_t *) + sizeof(kmp_root_t *)) * newCapacity + 3591 CACHE_LINE); 3592 newRoot = (kmp_root_t **)((char *)newThreads + 3593 sizeof(kmp_info_t *) * newCapacity); 3594 KMP_MEMCPY(newThreads, __kmp_threads, 3595 __kmp_threads_capacity * sizeof(kmp_info_t *)); 3596 KMP_MEMCPY(newRoot, __kmp_root, 3597 __kmp_threads_capacity * sizeof(kmp_root_t *)); 3598 memset(newThreads + __kmp_threads_capacity, 0, 3599 (newCapacity - __kmp_threads_capacity) * sizeof(kmp_info_t *)); 3600 memset(newRoot + __kmp_threads_capacity, 0, 3601 (newCapacity - __kmp_threads_capacity) * sizeof(kmp_root_t *)); 3602 3603 if (!old_tp_cached && __kmp_tp_cached && newCapacity > __kmp_tp_capacity) { 3604 /* __kmp_tp_cached has changed, i.e. __kmpc_threadprivate_cached has 3605 allocated a threadprivate cache while we were allocating the expanded 3606 array, and our new capacity is larger than the threadprivate cache 3607 capacity, so we should deallocate the expanded arrays and try again. 3608 This is the first check of a double-check pair. */ 3609 __kmp_free(newThreads); 3610 continue; /* start over and try again */ 3611 } 3612 __kmp_acquire_bootstrap_lock(&__kmp_tp_cached_lock); 3613 if (!old_tp_cached && __kmp_tp_cached && newCapacity > __kmp_tp_capacity) { 3614 /* Same check as above, but this time with the lock so we can be sure if 3615 we can succeed. */ 3616 __kmp_release_bootstrap_lock(&__kmp_tp_cached_lock); 3617 __kmp_free(newThreads); 3618 continue; /* start over and try again */ 3619 } else { 3620 /* success */ 3621 // __kmp_free( __kmp_threads ); // ATT: It leads to crash. Need to be 3622 // investigated. 3623 *(kmp_info_t * *volatile *)&__kmp_threads = newThreads; 3624 *(kmp_root_t * *volatile *)&__kmp_root = newRoot; 3625 added += newCapacity - __kmp_threads_capacity; 3626 *(volatile int *)&__kmp_threads_capacity = newCapacity; 3627 __kmp_release_bootstrap_lock(&__kmp_tp_cached_lock); 3628 break; /* succeeded, so we can exit the loop */ 3629 } 3630 } 3631 return added; 3632 } 3633 3634 /* Register the current thread as a root thread and obtain our gtid. We must 3635 have the __kmp_initz_lock held at this point. Argument TRUE only if are the 3636 thread that calls from __kmp_do_serial_initialize() */ 3637 int __kmp_register_root(int initial_thread) { 3638 kmp_info_t *root_thread; 3639 kmp_root_t *root; 3640 int gtid; 3641 int capacity; 3642 __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock); 3643 KA_TRACE(20, ("__kmp_register_root: entered\n")); 3644 KMP_MB(); 3645 3646 /* 2007-03-02: 3647 If initial thread did not invoke OpenMP RTL yet, and this thread is not an 3648 initial one, "__kmp_all_nth >= __kmp_threads_capacity" condition does not 3649 work as expected -- it may return false (that means there is at least one 3650 empty slot in __kmp_threads array), but it is possible the only free slot 3651 is #0, which is reserved for initial thread and so cannot be used for this 3652 one. Following code workarounds this bug. 3653 3654 However, right solution seems to be not reserving slot #0 for initial 3655 thread because: 3656 (1) there is no magic in slot #0, 3657 (2) we cannot detect initial thread reliably (the first thread which does 3658 serial initialization may be not a real initial thread). 3659 */ 3660 capacity = __kmp_threads_capacity; 3661 if (!initial_thread && TCR_PTR(__kmp_threads[0]) == NULL) { 3662 --capacity; 3663 }; // if 3664 3665 /* see if there are too many threads */ 3666 if (__kmp_all_nth >= capacity && !__kmp_expand_threads(1, 1)) { 3667 if (__kmp_tp_cached) { 3668 __kmp_fatal(KMP_MSG(CantRegisterNewThread), 3669 KMP_HNT(Set_ALL_THREADPRIVATE, __kmp_tp_capacity), 3670 KMP_HNT(PossibleSystemLimitOnThreads), __kmp_msg_null); 3671 } else { 3672 __kmp_fatal(KMP_MSG(CantRegisterNewThread), KMP_HNT(SystemLimitOnThreads), 3673 __kmp_msg_null); 3674 } 3675 }; // if 3676 3677 /* find an available thread slot */ 3678 /* Don't reassign the zero slot since we need that to only be used by initial 3679 thread */ 3680 for (gtid = (initial_thread ? 0 : 1); TCR_PTR(__kmp_threads[gtid]) != NULL; 3681 gtid++) 3682 ; 3683 KA_TRACE(1, 3684 ("__kmp_register_root: found slot in threads array: T#%d\n", gtid)); 3685 KMP_ASSERT(gtid < __kmp_threads_capacity); 3686 3687 /* update global accounting */ 3688 __kmp_all_nth++; 3689 TCW_4(__kmp_nth, __kmp_nth + 1); 3690 3691 // if __kmp_adjust_gtid_mode is set, then we use method #1 (sp search) for low 3692 // numbers of procs, and method #2 (keyed API call) for higher numbers. 3693 if (__kmp_adjust_gtid_mode) { 3694 if (__kmp_all_nth >= __kmp_tls_gtid_min) { 3695 if (TCR_4(__kmp_gtid_mode) != 2) { 3696 TCW_4(__kmp_gtid_mode, 2); 3697 } 3698 } else { 3699 if (TCR_4(__kmp_gtid_mode) != 1) { 3700 TCW_4(__kmp_gtid_mode, 1); 3701 } 3702 } 3703 } 3704 3705 #ifdef KMP_ADJUST_BLOCKTIME 3706 /* Adjust blocktime to zero if necessary */ 3707 /* Middle initialization might not have occurred yet */ 3708 if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) { 3709 if (__kmp_nth > __kmp_avail_proc) { 3710 __kmp_zero_bt = TRUE; 3711 } 3712 } 3713 #endif /* KMP_ADJUST_BLOCKTIME */ 3714 3715 /* setup this new hierarchy */ 3716 if (!(root = __kmp_root[gtid])) { 3717 root = __kmp_root[gtid] = (kmp_root_t *)__kmp_allocate(sizeof(kmp_root_t)); 3718 KMP_DEBUG_ASSERT(!root->r.r_root_team); 3719 } 3720 3721 #if KMP_STATS_ENABLED 3722 // Initialize stats as soon as possible (right after gtid assignment). 3723 __kmp_stats_thread_ptr = __kmp_stats_list->push_back(gtid); 3724 KMP_START_EXPLICIT_TIMER(OMP_worker_thread_life); 3725 KMP_SET_THREAD_STATE(SERIAL_REGION); 3726 KMP_INIT_PARTITIONED_TIMERS(OMP_serial); 3727 #endif 3728 __kmp_initialize_root(root); 3729 3730 /* setup new root thread structure */ 3731 if (root->r.r_uber_thread) { 3732 root_thread = root->r.r_uber_thread; 3733 } else { 3734 root_thread = (kmp_info_t *)__kmp_allocate(sizeof(kmp_info_t)); 3735 if (__kmp_storage_map) { 3736 __kmp_print_thread_storage_map(root_thread, gtid); 3737 } 3738 root_thread->th.th_info.ds.ds_gtid = gtid; 3739 root_thread->th.th_root = root; 3740 if (__kmp_env_consistency_check) { 3741 root_thread->th.th_cons = __kmp_allocate_cons_stack(gtid); 3742 } 3743 #if USE_FAST_MEMORY 3744 __kmp_initialize_fast_memory(root_thread); 3745 #endif /* USE_FAST_MEMORY */ 3746 3747 #if KMP_USE_BGET 3748 KMP_DEBUG_ASSERT(root_thread->th.th_local.bget_data == NULL); 3749 __kmp_initialize_bget(root_thread); 3750 #endif 3751 __kmp_init_random(root_thread); // Initialize random number generator 3752 } 3753 3754 /* setup the serial team held in reserve by the root thread */ 3755 if (!root_thread->th.th_serial_team) { 3756 kmp_internal_control_t r_icvs = __kmp_get_global_icvs(); 3757 KF_TRACE(10, ("__kmp_register_root: before serial_team\n")); 3758 root_thread->th.th_serial_team = 3759 __kmp_allocate_team(root, 1, 1, 3760 #if OMPT_SUPPORT 3761 0, // root parallel id 3762 #endif 3763 #if OMP_40_ENABLED 3764 proc_bind_default, 3765 #endif 3766 &r_icvs, 0 USE_NESTED_HOT_ARG(NULL)); 3767 } 3768 KMP_ASSERT(root_thread->th.th_serial_team); 3769 KF_TRACE(10, ("__kmp_register_root: after serial_team = %p\n", 3770 root_thread->th.th_serial_team)); 3771 3772 /* drop root_thread into place */ 3773 TCW_SYNC_PTR(__kmp_threads[gtid], root_thread); 3774 3775 root->r.r_root_team->t.t_threads[0] = root_thread; 3776 root->r.r_hot_team->t.t_threads[0] = root_thread; 3777 root_thread->th.th_serial_team->t.t_threads[0] = root_thread; 3778 // AC: the team created in reserve, not for execution (it is unused for now). 3779 root_thread->th.th_serial_team->t.t_serialized = 0; 3780 root->r.r_uber_thread = root_thread; 3781 3782 /* initialize the thread, get it ready to go */ 3783 __kmp_initialize_info(root_thread, root->r.r_root_team, 0, gtid); 3784 TCW_4(__kmp_init_gtid, TRUE); 3785 3786 /* prepare the master thread for get_gtid() */ 3787 __kmp_gtid_set_specific(gtid); 3788 3789 #if USE_ITT_BUILD 3790 __kmp_itt_thread_name(gtid); 3791 #endif /* USE_ITT_BUILD */ 3792 3793 #ifdef KMP_TDATA_GTID 3794 __kmp_gtid = gtid; 3795 #endif 3796 __kmp_create_worker(gtid, root_thread, __kmp_stksize); 3797 KMP_DEBUG_ASSERT(__kmp_gtid_get_specific() == gtid); 3798 3799 KA_TRACE(20, ("__kmp_register_root: T#%d init T#%d(%d:%d) arrived: join=%u, " 3800 "plain=%u\n", 3801 gtid, __kmp_gtid_from_tid(0, root->r.r_hot_team), 3802 root->r.r_hot_team->t.t_id, 0, KMP_INIT_BARRIER_STATE, 3803 KMP_INIT_BARRIER_STATE)); 3804 { // Initialize barrier data. 3805 int b; 3806 for (b = 0; b < bs_last_barrier; ++b) { 3807 root_thread->th.th_bar[b].bb.b_arrived = KMP_INIT_BARRIER_STATE; 3808 #if USE_DEBUGGER 3809 root_thread->th.th_bar[b].bb.b_worker_arrived = 0; 3810 #endif 3811 }; // for 3812 } 3813 KMP_DEBUG_ASSERT(root->r.r_hot_team->t.t_bar[bs_forkjoin_barrier].b_arrived == 3814 KMP_INIT_BARRIER_STATE); 3815 3816 #if KMP_AFFINITY_SUPPORTED 3817 #if OMP_40_ENABLED 3818 root_thread->th.th_current_place = KMP_PLACE_UNDEFINED; 3819 root_thread->th.th_new_place = KMP_PLACE_UNDEFINED; 3820 root_thread->th.th_first_place = KMP_PLACE_UNDEFINED; 3821 root_thread->th.th_last_place = KMP_PLACE_UNDEFINED; 3822 #endif 3823 3824 if (TCR_4(__kmp_init_middle)) { 3825 __kmp_affinity_set_init_mask(gtid, TRUE); 3826 } 3827 #endif /* KMP_AFFINITY_SUPPORTED */ 3828 3829 __kmp_root_counter++; 3830 3831 KMP_MB(); 3832 __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock); 3833 3834 return gtid; 3835 } 3836 3837 #if KMP_NESTED_HOT_TEAMS 3838 static int __kmp_free_hot_teams(kmp_root_t *root, kmp_info_t *thr, int level, 3839 const int max_level) { 3840 int i, n, nth; 3841 kmp_hot_team_ptr_t *hot_teams = thr->th.th_hot_teams; 3842 if (!hot_teams || !hot_teams[level].hot_team) { 3843 return 0; 3844 } 3845 KMP_DEBUG_ASSERT(level < max_level); 3846 kmp_team_t *team = hot_teams[level].hot_team; 3847 nth = hot_teams[level].hot_team_nth; 3848 n = nth - 1; // master is not freed 3849 if (level < max_level - 1) { 3850 for (i = 0; i < nth; ++i) { 3851 kmp_info_t *th = team->t.t_threads[i]; 3852 n += __kmp_free_hot_teams(root, th, level + 1, max_level); 3853 if (i > 0 && th->th.th_hot_teams) { 3854 __kmp_free(th->th.th_hot_teams); 3855 th->th.th_hot_teams = NULL; 3856 } 3857 } 3858 } 3859 __kmp_free_team(root, team, NULL); 3860 return n; 3861 } 3862 #endif 3863 3864 // Resets a root thread and clear its root and hot teams. 3865 // Returns the number of __kmp_threads entries directly and indirectly freed. 3866 static int __kmp_reset_root(int gtid, kmp_root_t *root) { 3867 kmp_team_t *root_team = root->r.r_root_team; 3868 kmp_team_t *hot_team = root->r.r_hot_team; 3869 int n = hot_team->t.t_nproc; 3870 int i; 3871 3872 KMP_DEBUG_ASSERT(!root->r.r_active); 3873 3874 root->r.r_root_team = NULL; 3875 root->r.r_hot_team = NULL; 3876 // __kmp_free_team() does not free hot teams, so we have to clear r_hot_team 3877 // before call to __kmp_free_team(). 3878 __kmp_free_team(root, root_team USE_NESTED_HOT_ARG(NULL)); 3879 #if KMP_NESTED_HOT_TEAMS 3880 if (__kmp_hot_teams_max_level > 3881 0) { // need to free nested hot teams and their threads if any 3882 for (i = 0; i < hot_team->t.t_nproc; ++i) { 3883 kmp_info_t *th = hot_team->t.t_threads[i]; 3884 if (__kmp_hot_teams_max_level > 1) { 3885 n += __kmp_free_hot_teams(root, th, 1, __kmp_hot_teams_max_level); 3886 } 3887 if (th->th.th_hot_teams) { 3888 __kmp_free(th->th.th_hot_teams); 3889 th->th.th_hot_teams = NULL; 3890 } 3891 } 3892 } 3893 #endif 3894 __kmp_free_team(root, hot_team USE_NESTED_HOT_ARG(NULL)); 3895 3896 // Before we can reap the thread, we need to make certain that all other 3897 // threads in the teams that had this root as ancestor have stopped trying to 3898 // steal tasks. 3899 if (__kmp_tasking_mode != tskm_immediate_exec) { 3900 __kmp_wait_to_unref_task_teams(); 3901 } 3902 3903 #if KMP_OS_WINDOWS 3904 /* Close Handle of root duplicated in __kmp_create_worker (tr #62919) */ 3905 KA_TRACE( 3906 10, ("__kmp_reset_root: free handle, th = %p, handle = %" KMP_UINTPTR_SPEC 3907 "\n", 3908 (LPVOID) & (root->r.r_uber_thread->th), 3909 root->r.r_uber_thread->th.th_info.ds.ds_thread)); 3910 __kmp_free_handle(root->r.r_uber_thread->th.th_info.ds.ds_thread); 3911 #endif /* KMP_OS_WINDOWS */ 3912 3913 #if OMPT_SUPPORT 3914 if (ompt_enabled && ompt_callbacks.ompt_callback(ompt_event_thread_end)) { 3915 int gtid = __kmp_get_gtid(); 3916 __ompt_thread_end(ompt_thread_initial, gtid); 3917 } 3918 #endif 3919 3920 TCW_4(__kmp_nth, 3921 __kmp_nth - 1); // __kmp_reap_thread will decrement __kmp_all_nth. 3922 root->r.r_cg_nthreads--; 3923 3924 __kmp_reap_thread(root->r.r_uber_thread, 1); 3925 3926 // We canot put root thread to __kmp_thread_pool, so we have to reap it istead 3927 // of freeing. 3928 root->r.r_uber_thread = NULL; 3929 /* mark root as no longer in use */ 3930 root->r.r_begin = FALSE; 3931 3932 return n; 3933 } 3934 3935 void __kmp_unregister_root_current_thread(int gtid) { 3936 KA_TRACE(1, ("__kmp_unregister_root_current_thread: enter T#%d\n", gtid)); 3937 /* this lock should be ok, since unregister_root_current_thread is never 3938 called during an abort, only during a normal close. furthermore, if you 3939 have the forkjoin lock, you should never try to get the initz lock */ 3940 __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock); 3941 if (TCR_4(__kmp_global.g.g_done) || !__kmp_init_serial) { 3942 KC_TRACE(10, ("__kmp_unregister_root_current_thread: already finished, " 3943 "exiting T#%d\n", 3944 gtid)); 3945 __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock); 3946 return; 3947 } 3948 kmp_root_t *root = __kmp_root[gtid]; 3949 3950 KMP_DEBUG_ASSERT(__kmp_threads && __kmp_threads[gtid]); 3951 KMP_ASSERT(KMP_UBER_GTID(gtid)); 3952 KMP_ASSERT(root == __kmp_threads[gtid]->th.th_root); 3953 KMP_ASSERT(root->r.r_active == FALSE); 3954 3955 KMP_MB(); 3956 3957 #if OMP_45_ENABLED 3958 kmp_info_t *thread = __kmp_threads[gtid]; 3959 kmp_team_t *team = thread->th.th_team; 3960 kmp_task_team_t *task_team = thread->th.th_task_team; 3961 3962 // we need to wait for the proxy tasks before finishing the thread 3963 if (task_team != NULL && task_team->tt.tt_found_proxy_tasks) { 3964 #if OMPT_SUPPORT 3965 // the runtime is shutting down so we won't report any events 3966 thread->th.ompt_thread_info.state = ompt_state_undefined; 3967 #endif 3968 __kmp_task_team_wait(thread, team USE_ITT_BUILD_ARG(NULL)); 3969 } 3970 #endif 3971 3972 __kmp_reset_root(gtid, root); 3973 3974 /* free up this thread slot */ 3975 __kmp_gtid_set_specific(KMP_GTID_DNE); 3976 #ifdef KMP_TDATA_GTID 3977 __kmp_gtid = KMP_GTID_DNE; 3978 #endif 3979 3980 KMP_MB(); 3981 KC_TRACE(10, 3982 ("__kmp_unregister_root_current_thread: T#%d unregistered\n", gtid)); 3983 3984 __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock); 3985 } 3986 3987 #if KMP_OS_WINDOWS 3988 /* __kmp_forkjoin_lock must be already held 3989 Unregisters a root thread that is not the current thread. Returns the number 3990 of __kmp_threads entries freed as a result. */ 3991 static int __kmp_unregister_root_other_thread(int gtid) { 3992 kmp_root_t *root = __kmp_root[gtid]; 3993 int r; 3994 3995 KA_TRACE(1, ("__kmp_unregister_root_other_thread: enter T#%d\n", gtid)); 3996 KMP_DEBUG_ASSERT(__kmp_threads && __kmp_threads[gtid]); 3997 KMP_ASSERT(KMP_UBER_GTID(gtid)); 3998 KMP_ASSERT(root == __kmp_threads[gtid]->th.th_root); 3999 KMP_ASSERT(root->r.r_active == FALSE); 4000 4001 r = __kmp_reset_root(gtid, root); 4002 KC_TRACE(10, 4003 ("__kmp_unregister_root_other_thread: T#%d unregistered\n", gtid)); 4004 return r; 4005 } 4006 #endif 4007 4008 #if KMP_DEBUG 4009 void __kmp_task_info() { 4010 4011 kmp_int32 gtid = __kmp_entry_gtid(); 4012 kmp_int32 tid = __kmp_tid_from_gtid(gtid); 4013 kmp_info_t *this_thr = __kmp_threads[gtid]; 4014 kmp_team_t *steam = this_thr->th.th_serial_team; 4015 kmp_team_t *team = this_thr->th.th_team; 4016 4017 __kmp_printf("__kmp_task_info: gtid=%d tid=%d t_thread=%p team=%p curtask=%p " 4018 "ptask=%p\n", 4019 gtid, tid, this_thr, team, this_thr->th.th_current_task, 4020 team->t.t_implicit_task_taskdata[tid].td_parent); 4021 } 4022 #endif // KMP_DEBUG 4023 4024 /* TODO optimize with one big memclr, take out what isn't needed, split 4025 responsibility to workers as much as possible, and delay initialization of 4026 features as much as possible */ 4027 static void __kmp_initialize_info(kmp_info_t *this_thr, kmp_team_t *team, 4028 int tid, int gtid) { 4029 /* this_thr->th.th_info.ds.ds_gtid is setup in 4030 kmp_allocate_thread/create_worker. 4031 this_thr->th.th_serial_team is setup in __kmp_allocate_thread */ 4032 kmp_info_t *master = team->t.t_threads[0]; 4033 KMP_DEBUG_ASSERT(this_thr != NULL); 4034 KMP_DEBUG_ASSERT(this_thr->th.th_serial_team); 4035 KMP_DEBUG_ASSERT(team); 4036 KMP_DEBUG_ASSERT(team->t.t_threads); 4037 KMP_DEBUG_ASSERT(team->t.t_dispatch); 4038 KMP_DEBUG_ASSERT(master); 4039 KMP_DEBUG_ASSERT(master->th.th_root); 4040 4041 KMP_MB(); 4042 4043 TCW_SYNC_PTR(this_thr->th.th_team, team); 4044 4045 this_thr->th.th_info.ds.ds_tid = tid; 4046 this_thr->th.th_set_nproc = 0; 4047 if (__kmp_tasking_mode != tskm_immediate_exec) 4048 // When tasking is possible, threads are not safe to reap until they are 4049 // done tasking; this will be set when tasking code is exited in wait 4050 this_thr->th.th_reap_state = KMP_NOT_SAFE_TO_REAP; 4051 else // no tasking --> always safe to reap 4052 this_thr->th.th_reap_state = KMP_SAFE_TO_REAP; 4053 #if OMP_40_ENABLED 4054 this_thr->th.th_set_proc_bind = proc_bind_default; 4055 #if KMP_AFFINITY_SUPPORTED 4056 this_thr->th.th_new_place = this_thr->th.th_current_place; 4057 #endif 4058 #endif 4059 this_thr->th.th_root = master->th.th_root; 4060 4061 /* setup the thread's cache of the team structure */ 4062 this_thr->th.th_team_nproc = team->t.t_nproc; 4063 this_thr->th.th_team_master = master; 4064 this_thr->th.th_team_serialized = team->t.t_serialized; 4065 TCW_PTR(this_thr->th.th_sleep_loc, NULL); 4066 4067 KMP_DEBUG_ASSERT(team->t.t_implicit_task_taskdata); 4068 4069 KF_TRACE(10, ("__kmp_initialize_info1: T#%d:%d this_thread=%p curtask=%p\n", 4070 tid, gtid, this_thr, this_thr->th.th_current_task)); 4071 4072 __kmp_init_implicit_task(this_thr->th.th_team_master->th.th_ident, this_thr, 4073 team, tid, TRUE); 4074 4075 KF_TRACE(10, ("__kmp_initialize_info2: T#%d:%d this_thread=%p curtask=%p\n", 4076 tid, gtid, this_thr, this_thr->th.th_current_task)); 4077 // TODO: Initialize ICVs from parent; GEH - isn't that already done in 4078 // __kmp_initialize_team()? 4079 4080 /* TODO no worksharing in speculative threads */ 4081 this_thr->th.th_dispatch = &team->t.t_dispatch[tid]; 4082 4083 this_thr->th.th_local.this_construct = 0; 4084 4085 if (!this_thr->th.th_pri_common) { 4086 this_thr->th.th_pri_common = 4087 (struct common_table *)__kmp_allocate(sizeof(struct common_table)); 4088 if (__kmp_storage_map) { 4089 __kmp_print_storage_map_gtid( 4090 gtid, this_thr->th.th_pri_common, this_thr->th.th_pri_common + 1, 4091 sizeof(struct common_table), "th_%d.th_pri_common\n", gtid); 4092 }; // if 4093 this_thr->th.th_pri_head = NULL; 4094 }; // if 4095 4096 /* Initialize dynamic dispatch */ 4097 { 4098 volatile kmp_disp_t *dispatch = this_thr->th.th_dispatch; 4099 // Use team max_nproc since this will never change for the team. 4100 size_t disp_size = 4101 sizeof(dispatch_private_info_t) * 4102 (team->t.t_max_nproc == 1 ? 1 : __kmp_dispatch_num_buffers); 4103 KD_TRACE(10, ("__kmp_initialize_info: T#%d max_nproc: %d\n", gtid, 4104 team->t.t_max_nproc)); 4105 KMP_ASSERT(dispatch); 4106 KMP_DEBUG_ASSERT(team->t.t_dispatch); 4107 KMP_DEBUG_ASSERT(dispatch == &team->t.t_dispatch[tid]); 4108 4109 dispatch->th_disp_index = 0; 4110 #if OMP_45_ENABLED 4111 dispatch->th_doacross_buf_idx = 0; 4112 #endif 4113 if (!dispatch->th_disp_buffer) { 4114 dispatch->th_disp_buffer = 4115 (dispatch_private_info_t *)__kmp_allocate(disp_size); 4116 4117 if (__kmp_storage_map) { 4118 __kmp_print_storage_map_gtid( 4119 gtid, &dispatch->th_disp_buffer[0], 4120 &dispatch->th_disp_buffer[team->t.t_max_nproc == 1 4121 ? 1 4122 : __kmp_dispatch_num_buffers], 4123 disp_size, "th_%d.th_dispatch.th_disp_buffer " 4124 "(team_%d.t_dispatch[%d].th_disp_buffer)", 4125 gtid, team->t.t_id, gtid); 4126 } 4127 } else { 4128 memset(&dispatch->th_disp_buffer[0], '\0', disp_size); 4129 } 4130 4131 dispatch->th_dispatch_pr_current = 0; 4132 dispatch->th_dispatch_sh_current = 0; 4133 4134 dispatch->th_deo_fcn = 0; /* ORDERED */ 4135 dispatch->th_dxo_fcn = 0; /* END ORDERED */ 4136 } 4137 4138 this_thr->th.th_next_pool = NULL; 4139 4140 if (!this_thr->th.th_task_state_memo_stack) { 4141 size_t i; 4142 this_thr->th.th_task_state_memo_stack = 4143 (kmp_uint8 *)__kmp_allocate(4 * sizeof(kmp_uint8)); 4144 this_thr->th.th_task_state_top = 0; 4145 this_thr->th.th_task_state_stack_sz = 4; 4146 for (i = 0; i < this_thr->th.th_task_state_stack_sz; 4147 ++i) // zero init the stack 4148 this_thr->th.th_task_state_memo_stack[i] = 0; 4149 } 4150 4151 KMP_DEBUG_ASSERT(!this_thr->th.th_spin_here); 4152 KMP_DEBUG_ASSERT(this_thr->th.th_next_waiting == 0); 4153 4154 KMP_MB(); 4155 } 4156 4157 /* allocate a new thread for the requesting team. this is only called from 4158 within a forkjoin critical section. we will first try to get an available 4159 thread from the thread pool. if none is available, we will fork a new one 4160 assuming we are able to create a new one. this should be assured, as the 4161 caller should check on this first. */ 4162 kmp_info_t *__kmp_allocate_thread(kmp_root_t *root, kmp_team_t *team, 4163 int new_tid) { 4164 kmp_team_t *serial_team; 4165 kmp_info_t *new_thr; 4166 int new_gtid; 4167 4168 KA_TRACE(20, ("__kmp_allocate_thread: T#%d\n", __kmp_get_gtid())); 4169 KMP_DEBUG_ASSERT(root && team); 4170 #if !KMP_NESTED_HOT_TEAMS 4171 KMP_DEBUG_ASSERT(KMP_MASTER_GTID(__kmp_get_gtid())); 4172 #endif 4173 KMP_MB(); 4174 4175 /* first, try to get one from the thread pool */ 4176 if (__kmp_thread_pool) { 4177 4178 new_thr = CCAST(kmp_info_t *, __kmp_thread_pool); 4179 __kmp_thread_pool = (volatile kmp_info_t *)new_thr->th.th_next_pool; 4180 if (new_thr == __kmp_thread_pool_insert_pt) { 4181 __kmp_thread_pool_insert_pt = NULL; 4182 } 4183 TCW_4(new_thr->th.th_in_pool, FALSE); 4184 // Don't touch th_active_in_pool or th_active. 4185 // The worker thread adjusts those flags as it sleeps/awakens. 4186 __kmp_thread_pool_nth--; 4187 4188 KA_TRACE(20, ("__kmp_allocate_thread: T#%d using thread T#%d\n", 4189 __kmp_get_gtid(), new_thr->th.th_info.ds.ds_gtid)); 4190 KMP_ASSERT(!new_thr->th.th_team); 4191 KMP_DEBUG_ASSERT(__kmp_nth < __kmp_threads_capacity); 4192 KMP_DEBUG_ASSERT(__kmp_thread_pool_nth >= 0); 4193 4194 /* setup the thread structure */ 4195 __kmp_initialize_info(new_thr, team, new_tid, 4196 new_thr->th.th_info.ds.ds_gtid); 4197 KMP_DEBUG_ASSERT(new_thr->th.th_serial_team); 4198 4199 TCW_4(__kmp_nth, __kmp_nth + 1); 4200 root->r.r_cg_nthreads++; 4201 4202 new_thr->th.th_task_state = 0; 4203 new_thr->th.th_task_state_top = 0; 4204 new_thr->th.th_task_state_stack_sz = 4; 4205 4206 #ifdef KMP_ADJUST_BLOCKTIME 4207 /* Adjust blocktime back to zero if necessary */ 4208 /* Middle initialization might not have occurred yet */ 4209 if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) { 4210 if (__kmp_nth > __kmp_avail_proc) { 4211 __kmp_zero_bt = TRUE; 4212 } 4213 } 4214 #endif /* KMP_ADJUST_BLOCKTIME */ 4215 4216 #if KMP_DEBUG 4217 // If thread entered pool via __kmp_free_thread, wait_flag should != 4218 // KMP_BARRIER_PARENT_FLAG. 4219 int b; 4220 kmp_balign_t *balign = new_thr->th.th_bar; 4221 for (b = 0; b < bs_last_barrier; ++b) 4222 KMP_DEBUG_ASSERT(balign[b].bb.wait_flag != KMP_BARRIER_PARENT_FLAG); 4223 #endif 4224 4225 KF_TRACE(10, ("__kmp_allocate_thread: T#%d using thread %p T#%d\n", 4226 __kmp_get_gtid(), new_thr, new_thr->th.th_info.ds.ds_gtid)); 4227 4228 KMP_MB(); 4229 return new_thr; 4230 } 4231 4232 /* no, well fork a new one */ 4233 KMP_ASSERT(__kmp_nth == __kmp_all_nth); 4234 KMP_ASSERT(__kmp_all_nth < __kmp_threads_capacity); 4235 4236 #if KMP_USE_MONITOR 4237 // If this is the first worker thread the RTL is creating, then also 4238 // launch the monitor thread. We try to do this as early as possible. 4239 if (!TCR_4(__kmp_init_monitor)) { 4240 __kmp_acquire_bootstrap_lock(&__kmp_monitor_lock); 4241 if (!TCR_4(__kmp_init_monitor)) { 4242 KF_TRACE(10, ("before __kmp_create_monitor\n")); 4243 TCW_4(__kmp_init_monitor, 1); 4244 __kmp_create_monitor(&__kmp_monitor); 4245 KF_TRACE(10, ("after __kmp_create_monitor\n")); 4246 #if KMP_OS_WINDOWS 4247 // AC: wait until monitor has started. This is a fix for CQ232808. 4248 // The reason is that if the library is loaded/unloaded in a loop with 4249 // small (parallel) work in between, then there is high probability that 4250 // monitor thread started after the library shutdown. At shutdown it is 4251 // too late to cope with the problem, because when the master is in 4252 // DllMain (process detach) the monitor has no chances to start (it is 4253 // blocked), and master has no means to inform the monitor that the 4254 // library has gone, because all the memory which the monitor can access 4255 // is going to be released/reset. 4256 while (TCR_4(__kmp_init_monitor) < 2) { 4257 KMP_YIELD(TRUE); 4258 } 4259 KF_TRACE(10, ("after monitor thread has started\n")); 4260 #endif 4261 } 4262 __kmp_release_bootstrap_lock(&__kmp_monitor_lock); 4263 } 4264 #endif 4265 4266 KMP_MB(); 4267 for (new_gtid = 1; TCR_PTR(__kmp_threads[new_gtid]) != NULL; ++new_gtid) { 4268 KMP_DEBUG_ASSERT(new_gtid < __kmp_threads_capacity); 4269 } 4270 4271 /* allocate space for it. */ 4272 new_thr = (kmp_info_t *)__kmp_allocate(sizeof(kmp_info_t)); 4273 4274 TCW_SYNC_PTR(__kmp_threads[new_gtid], new_thr); 4275 4276 if (__kmp_storage_map) { 4277 __kmp_print_thread_storage_map(new_thr, new_gtid); 4278 } 4279 4280 // add the reserve serialized team, initialized from the team's master thread 4281 { 4282 kmp_internal_control_t r_icvs = __kmp_get_x_global_icvs(team); 4283 KF_TRACE(10, ("__kmp_allocate_thread: before th_serial/serial_team\n")); 4284 new_thr->th.th_serial_team = serial_team = 4285 (kmp_team_t *)__kmp_allocate_team(root, 1, 1, 4286 #if OMPT_SUPPORT 4287 0, // root parallel id 4288 #endif 4289 #if OMP_40_ENABLED 4290 proc_bind_default, 4291 #endif 4292 &r_icvs, 0 USE_NESTED_HOT_ARG(NULL)); 4293 } 4294 KMP_ASSERT(serial_team); 4295 serial_team->t.t_serialized = 0; // AC: the team created in reserve, not for 4296 // execution (it is unused for now). 4297 serial_team->t.t_threads[0] = new_thr; 4298 KF_TRACE(10, 4299 ("__kmp_allocate_thread: after th_serial/serial_team : new_thr=%p\n", 4300 new_thr)); 4301 4302 /* setup the thread structures */ 4303 __kmp_initialize_info(new_thr, team, new_tid, new_gtid); 4304 4305 #if USE_FAST_MEMORY 4306 __kmp_initialize_fast_memory(new_thr); 4307 #endif /* USE_FAST_MEMORY */ 4308 4309 #if KMP_USE_BGET 4310 KMP_DEBUG_ASSERT(new_thr->th.th_local.bget_data == NULL); 4311 __kmp_initialize_bget(new_thr); 4312 #endif 4313 4314 __kmp_init_random(new_thr); // Initialize random number generator 4315 4316 /* Initialize these only once when thread is grabbed for a team allocation */ 4317 KA_TRACE(20, 4318 ("__kmp_allocate_thread: T#%d init go fork=%u, plain=%u\n", 4319 __kmp_get_gtid(), KMP_INIT_BARRIER_STATE, KMP_INIT_BARRIER_STATE)); 4320 4321 int b; 4322 kmp_balign_t *balign = new_thr->th.th_bar; 4323 for (b = 0; b < bs_last_barrier; ++b) { 4324 balign[b].bb.b_go = KMP_INIT_BARRIER_STATE; 4325 balign[b].bb.team = NULL; 4326 balign[b].bb.wait_flag = KMP_BARRIER_NOT_WAITING; 4327 balign[b].bb.use_oncore_barrier = 0; 4328 } 4329 4330 new_thr->th.th_spin_here = FALSE; 4331 new_thr->th.th_next_waiting = 0; 4332 4333 #if OMP_40_ENABLED && KMP_AFFINITY_SUPPORTED 4334 new_thr->th.th_current_place = KMP_PLACE_UNDEFINED; 4335 new_thr->th.th_new_place = KMP_PLACE_UNDEFINED; 4336 new_thr->th.th_first_place = KMP_PLACE_UNDEFINED; 4337 new_thr->th.th_last_place = KMP_PLACE_UNDEFINED; 4338 #endif 4339 4340 TCW_4(new_thr->th.th_in_pool, FALSE); 4341 new_thr->th.th_active_in_pool = FALSE; 4342 TCW_4(new_thr->th.th_active, TRUE); 4343 4344 /* adjust the global counters */ 4345 __kmp_all_nth++; 4346 __kmp_nth++; 4347 4348 root->r.r_cg_nthreads++; 4349 4350 // if __kmp_adjust_gtid_mode is set, then we use method #1 (sp search) for low 4351 // numbers of procs, and method #2 (keyed API call) for higher numbers. 4352 if (__kmp_adjust_gtid_mode) { 4353 if (__kmp_all_nth >= __kmp_tls_gtid_min) { 4354 if (TCR_4(__kmp_gtid_mode) != 2) { 4355 TCW_4(__kmp_gtid_mode, 2); 4356 } 4357 } else { 4358 if (TCR_4(__kmp_gtid_mode) != 1) { 4359 TCW_4(__kmp_gtid_mode, 1); 4360 } 4361 } 4362 } 4363 4364 #ifdef KMP_ADJUST_BLOCKTIME 4365 /* Adjust blocktime back to zero if necessary */ 4366 /* Middle initialization might not have occurred yet */ 4367 if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) { 4368 if (__kmp_nth > __kmp_avail_proc) { 4369 __kmp_zero_bt = TRUE; 4370 } 4371 } 4372 #endif /* KMP_ADJUST_BLOCKTIME */ 4373 4374 /* actually fork it and create the new worker thread */ 4375 KF_TRACE( 4376 10, ("__kmp_allocate_thread: before __kmp_create_worker: %p\n", new_thr)); 4377 __kmp_create_worker(new_gtid, new_thr, __kmp_stksize); 4378 KF_TRACE(10, 4379 ("__kmp_allocate_thread: after __kmp_create_worker: %p\n", new_thr)); 4380 4381 KA_TRACE(20, ("__kmp_allocate_thread: T#%d forked T#%d\n", __kmp_get_gtid(), 4382 new_gtid)); 4383 KMP_MB(); 4384 return new_thr; 4385 } 4386 4387 /* Reinitialize team for reuse. 4388 The hot team code calls this case at every fork barrier, so EPCC barrier 4389 test are extremely sensitive to changes in it, esp. writes to the team 4390 struct, which cause a cache invalidation in all threads. 4391 IF YOU TOUCH THIS ROUTINE, RUN EPCC C SYNCBENCH ON A BIG-IRON MACHINE!!! */ 4392 static void __kmp_reinitialize_team(kmp_team_t *team, 4393 kmp_internal_control_t *new_icvs, 4394 ident_t *loc) { 4395 KF_TRACE(10, ("__kmp_reinitialize_team: enter this_thread=%p team=%p\n", 4396 team->t.t_threads[0], team)); 4397 KMP_DEBUG_ASSERT(team && new_icvs); 4398 KMP_DEBUG_ASSERT((!TCR_4(__kmp_init_parallel)) || new_icvs->nproc); 4399 KMP_CHECK_UPDATE(team->t.t_ident, loc); 4400 4401 KMP_CHECK_UPDATE(team->t.t_id, KMP_GEN_TEAM_ID()); 4402 // Copy ICVs to the master thread's implicit taskdata 4403 __kmp_init_implicit_task(loc, team->t.t_threads[0], team, 0, FALSE); 4404 copy_icvs(&team->t.t_implicit_task_taskdata[0].td_icvs, new_icvs); 4405 4406 KF_TRACE(10, ("__kmp_reinitialize_team: exit this_thread=%p team=%p\n", 4407 team->t.t_threads[0], team)); 4408 } 4409 4410 /* Initialize the team data structure. 4411 This assumes the t_threads and t_max_nproc are already set. 4412 Also, we don't touch the arguments */ 4413 static void __kmp_initialize_team(kmp_team_t *team, int new_nproc, 4414 kmp_internal_control_t *new_icvs, 4415 ident_t *loc) { 4416 KF_TRACE(10, ("__kmp_initialize_team: enter: team=%p\n", team)); 4417 4418 /* verify */ 4419 KMP_DEBUG_ASSERT(team); 4420 KMP_DEBUG_ASSERT(new_nproc <= team->t.t_max_nproc); 4421 KMP_DEBUG_ASSERT(team->t.t_threads); 4422 KMP_MB(); 4423 4424 team->t.t_master_tid = 0; /* not needed */ 4425 /* team->t.t_master_bar; not needed */ 4426 team->t.t_serialized = new_nproc > 1 ? 0 : 1; 4427 team->t.t_nproc = new_nproc; 4428 4429 /* team->t.t_parent = NULL; TODO not needed & would mess up hot team */ 4430 team->t.t_next_pool = NULL; 4431 /* memset( team->t.t_threads, 0, sizeof(kmp_info_t*)*new_nproc ); would mess 4432 * up hot team */ 4433 4434 TCW_SYNC_PTR(team->t.t_pkfn, NULL); /* not needed */ 4435 team->t.t_invoke = NULL; /* not needed */ 4436 4437 // TODO???: team->t.t_max_active_levels = new_max_active_levels; 4438 team->t.t_sched = new_icvs->sched; 4439 4440 #if KMP_ARCH_X86 || KMP_ARCH_X86_64 4441 team->t.t_fp_control_saved = FALSE; /* not needed */ 4442 team->t.t_x87_fpu_control_word = 0; /* not needed */ 4443 team->t.t_mxcsr = 0; /* not needed */ 4444 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */ 4445 4446 team->t.t_construct = 0; 4447 4448 team->t.t_ordered.dt.t_value = 0; 4449 team->t.t_master_active = FALSE; 4450 4451 memset(&team->t.t_taskq, '\0', sizeof(kmp_taskq_t)); 4452 4453 #ifdef KMP_DEBUG 4454 team->t.t_copypriv_data = NULL; /* not necessary, but nice for debugging */ 4455 #endif 4456 team->t.t_copyin_counter = 0; /* for barrier-free copyin implementation */ 4457 4458 team->t.t_control_stack_top = NULL; 4459 4460 __kmp_reinitialize_team(team, new_icvs, loc); 4461 4462 KMP_MB(); 4463 KF_TRACE(10, ("__kmp_initialize_team: exit: team=%p\n", team)); 4464 } 4465 4466 #if KMP_OS_LINUX && KMP_AFFINITY_SUPPORTED 4467 /* Sets full mask for thread and returns old mask, no changes to structures. */ 4468 static void 4469 __kmp_set_thread_affinity_mask_full_tmp(kmp_affin_mask_t *old_mask) { 4470 if (KMP_AFFINITY_CAPABLE()) { 4471 int status; 4472 if (old_mask != NULL) { 4473 status = __kmp_get_system_affinity(old_mask, TRUE); 4474 int error = errno; 4475 if (status != 0) { 4476 __kmp_fatal(KMP_MSG(ChangeThreadAffMaskError), KMP_ERR(error), 4477 __kmp_msg_null); 4478 } 4479 } 4480 __kmp_set_system_affinity(__kmp_affin_fullMask, TRUE); 4481 } 4482 } 4483 #endif 4484 4485 #if OMP_40_ENABLED && KMP_AFFINITY_SUPPORTED 4486 4487 // __kmp_partition_places() is the heart of the OpenMP 4.0 affinity mechanism. 4488 // It calculats the worker + master thread's partition based upon the parent 4489 // thread's partition, and binds each worker to a thread in their partition. 4490 // The master thread's partition should already include its current binding. 4491 static void __kmp_partition_places(kmp_team_t *team, int update_master_only) { 4492 // Copy the master thread's place partion to the team struct 4493 kmp_info_t *master_th = team->t.t_threads[0]; 4494 KMP_DEBUG_ASSERT(master_th != NULL); 4495 kmp_proc_bind_t proc_bind = team->t.t_proc_bind; 4496 int first_place = master_th->th.th_first_place; 4497 int last_place = master_th->th.th_last_place; 4498 int masters_place = master_th->th.th_current_place; 4499 team->t.t_first_place = first_place; 4500 team->t.t_last_place = last_place; 4501 4502 KA_TRACE(20, ("__kmp_partition_places: enter: proc_bind = %d T#%d(%d:0) " 4503 "bound to place %d partition = [%d,%d]\n", 4504 proc_bind, __kmp_gtid_from_thread(team->t.t_threads[0]), 4505 team->t.t_id, masters_place, first_place, last_place)); 4506 4507 switch (proc_bind) { 4508 4509 case proc_bind_default: 4510 // serial teams might have the proc_bind policy set to proc_bind_default. It 4511 // doesn't matter, as we don't rebind master thread for any proc_bind policy 4512 KMP_DEBUG_ASSERT(team->t.t_nproc == 1); 4513 break; 4514 4515 case proc_bind_master: { 4516 int f; 4517 int n_th = team->t.t_nproc; 4518 for (f = 1; f < n_th; f++) { 4519 kmp_info_t *th = team->t.t_threads[f]; 4520 KMP_DEBUG_ASSERT(th != NULL); 4521 th->th.th_first_place = first_place; 4522 th->th.th_last_place = last_place; 4523 th->th.th_new_place = masters_place; 4524 4525 KA_TRACE(100, ("__kmp_partition_places: master: T#%d(%d:%d) place %d " 4526 "partition = [%d,%d]\n", 4527 __kmp_gtid_from_thread(team->t.t_threads[f]), team->t.t_id, 4528 f, masters_place, first_place, last_place)); 4529 } 4530 } break; 4531 4532 case proc_bind_close: { 4533 int f; 4534 int n_th = team->t.t_nproc; 4535 int n_places; 4536 if (first_place <= last_place) { 4537 n_places = last_place - first_place + 1; 4538 } else { 4539 n_places = __kmp_affinity_num_masks - first_place + last_place + 1; 4540 } 4541 if (n_th <= n_places) { 4542 int place = masters_place; 4543 for (f = 1; f < n_th; f++) { 4544 kmp_info_t *th = team->t.t_threads[f]; 4545 KMP_DEBUG_ASSERT(th != NULL); 4546 4547 if (place == last_place) { 4548 place = first_place; 4549 } else if (place == (int)(__kmp_affinity_num_masks - 1)) { 4550 place = 0; 4551 } else { 4552 place++; 4553 } 4554 th->th.th_first_place = first_place; 4555 th->th.th_last_place = last_place; 4556 th->th.th_new_place = place; 4557 4558 KA_TRACE(100, ("__kmp_partition_places: close: T#%d(%d:%d) place %d " 4559 "partition = [%d,%d]\n", 4560 __kmp_gtid_from_thread(team->t.t_threads[f]), 4561 team->t.t_id, f, place, first_place, last_place)); 4562 } 4563 } else { 4564 int S, rem, gap, s_count; 4565 S = n_th / n_places; 4566 s_count = 0; 4567 rem = n_th - (S * n_places); 4568 gap = rem > 0 ? n_places / rem : n_places; 4569 int place = masters_place; 4570 int gap_ct = gap; 4571 for (f = 0; f < n_th; f++) { 4572 kmp_info_t *th = team->t.t_threads[f]; 4573 KMP_DEBUG_ASSERT(th != NULL); 4574 4575 th->th.th_first_place = first_place; 4576 th->th.th_last_place = last_place; 4577 th->th.th_new_place = place; 4578 s_count++; 4579 4580 if ((s_count == S) && rem && (gap_ct == gap)) { 4581 // do nothing, add an extra thread to place on next iteration 4582 } else if ((s_count == S + 1) && rem && (gap_ct == gap)) { 4583 // we added an extra thread to this place; move to next place 4584 if (place == last_place) { 4585 place = first_place; 4586 } else if (place == (int)(__kmp_affinity_num_masks - 1)) { 4587 place = 0; 4588 } else { 4589 place++; 4590 } 4591 s_count = 0; 4592 gap_ct = 1; 4593 rem--; 4594 } else if (s_count == S) { // place full; don't add extra 4595 if (place == last_place) { 4596 place = first_place; 4597 } else if (place == (int)(__kmp_affinity_num_masks - 1)) { 4598 place = 0; 4599 } else { 4600 place++; 4601 } 4602 gap_ct++; 4603 s_count = 0; 4604 } 4605 4606 KA_TRACE(100, 4607 ("__kmp_partition_places: close: T#%d(%d:%d) place %d " 4608 "partition = [%d,%d]\n", 4609 __kmp_gtid_from_thread(team->t.t_threads[f]), team->t.t_id, f, 4610 th->th.th_new_place, first_place, last_place)); 4611 } 4612 KMP_DEBUG_ASSERT(place == masters_place); 4613 } 4614 } break; 4615 4616 case proc_bind_spread: { 4617 int f; 4618 int n_th = team->t.t_nproc; 4619 int n_places; 4620 int thidx; 4621 if (first_place <= last_place) { 4622 n_places = last_place - first_place + 1; 4623 } else { 4624 n_places = __kmp_affinity_num_masks - first_place + last_place + 1; 4625 } 4626 if (n_th <= n_places) { 4627 int place = -1; 4628 4629 if (n_places != static_cast<int>(__kmp_affinity_num_masks)) { 4630 int S = n_places / n_th; 4631 int s_count, rem, gap, gap_ct; 4632 4633 place = masters_place; 4634 rem = n_places - n_th * S; 4635 gap = rem ? n_th / rem : 1; 4636 gap_ct = gap; 4637 thidx = n_th; 4638 if (update_master_only == 1) 4639 thidx = 1; 4640 for (f = 0; f < thidx; f++) { 4641 kmp_info_t *th = team->t.t_threads[f]; 4642 KMP_DEBUG_ASSERT(th != NULL); 4643 4644 th->th.th_first_place = place; 4645 th->th.th_new_place = place; 4646 s_count = 1; 4647 while (s_count < S) { 4648 if (place == last_place) { 4649 place = first_place; 4650 } else if (place == (int)(__kmp_affinity_num_masks - 1)) { 4651 place = 0; 4652 } else { 4653 place++; 4654 } 4655 s_count++; 4656 } 4657 if (rem && (gap_ct == gap)) { 4658 if (place == last_place) { 4659 place = first_place; 4660 } else if (place == (int)(__kmp_affinity_num_masks - 1)) { 4661 place = 0; 4662 } else { 4663 place++; 4664 } 4665 rem--; 4666 gap_ct = 0; 4667 } 4668 th->th.th_last_place = place; 4669 gap_ct++; 4670 4671 if (place == last_place) { 4672 place = first_place; 4673 } else if (place == (int)(__kmp_affinity_num_masks - 1)) { 4674 place = 0; 4675 } else { 4676 place++; 4677 } 4678 4679 KA_TRACE(100, ("__kmp_partition_places: spread: T#%d(%d:%d) place %d " 4680 "partition = [%d,%d], __kmp_affinity_num_masks: %u\n", 4681 __kmp_gtid_from_thread(team->t.t_threads[f]), 4682 team->t.t_id, f, th->th.th_new_place, 4683 th->th.th_first_place, th->th.th_last_place, 4684 __kmp_affinity_num_masks)); 4685 } 4686 } else { 4687 /* Having uniform space of available computation places I can create 4688 T partitions of round(P/T) size and put threads into the first 4689 place of each partition. */ 4690 double current = static_cast<double>(masters_place); 4691 double spacing = 4692 (static_cast<double>(n_places + 1) / static_cast<double>(n_th)); 4693 int first, last; 4694 kmp_info_t *th; 4695 4696 thidx = n_th + 1; 4697 if (update_master_only == 1) 4698 thidx = 1; 4699 for (f = 0; f < thidx; f++) { 4700 first = static_cast<int>(current); 4701 last = static_cast<int>(current + spacing) - 1; 4702 KMP_DEBUG_ASSERT(last >= first); 4703 if (first >= n_places) { 4704 if (masters_place) { 4705 first -= n_places; 4706 last -= n_places; 4707 if (first == (masters_place + 1)) { 4708 KMP_DEBUG_ASSERT(f == n_th); 4709 first--; 4710 } 4711 if (last == masters_place) { 4712 KMP_DEBUG_ASSERT(f == (n_th - 1)); 4713 last--; 4714 } 4715 } else { 4716 KMP_DEBUG_ASSERT(f == n_th); 4717 first = 0; 4718 last = 0; 4719 } 4720 } 4721 if (last >= n_places) { 4722 last = (n_places - 1); 4723 } 4724 place = first; 4725 current += spacing; 4726 if (f < n_th) { 4727 KMP_DEBUG_ASSERT(0 <= first); 4728 KMP_DEBUG_ASSERT(n_places > first); 4729 KMP_DEBUG_ASSERT(0 <= last); 4730 KMP_DEBUG_ASSERT(n_places > last); 4731 KMP_DEBUG_ASSERT(last_place >= first_place); 4732 th = team->t.t_threads[f]; 4733 KMP_DEBUG_ASSERT(th); 4734 th->th.th_first_place = first; 4735 th->th.th_new_place = place; 4736 th->th.th_last_place = last; 4737 4738 KA_TRACE(100, ("__kmp_partition_places: spread: T#%d(%d:%d) place %d " 4739 "partition = [%d,%d], spacing = %.4f\n", 4740 __kmp_gtid_from_thread(team->t.t_threads[f]), 4741 team->t.t_id, f, th->th.th_new_place, 4742 th->th.th_first_place, th->th.th_last_place, 4743 spacing)); 4744 } 4745 } 4746 } 4747 KMP_DEBUG_ASSERT(update_master_only || place == masters_place); 4748 } else { 4749 int S, rem, gap, s_count; 4750 S = n_th / n_places; 4751 s_count = 0; 4752 rem = n_th - (S * n_places); 4753 gap = rem > 0 ? n_places / rem : n_places; 4754 int place = masters_place; 4755 int gap_ct = gap; 4756 thidx = n_th; 4757 if (update_master_only == 1) 4758 thidx = 1; 4759 for (f = 0; f < thidx; f++) { 4760 kmp_info_t *th = team->t.t_threads[f]; 4761 KMP_DEBUG_ASSERT(th != NULL); 4762 4763 th->th.th_first_place = place; 4764 th->th.th_last_place = place; 4765 th->th.th_new_place = place; 4766 s_count++; 4767 4768 if ((s_count == S) && rem && (gap_ct == gap)) { 4769 // do nothing, add an extra thread to place on next iteration 4770 } else if ((s_count == S + 1) && rem && (gap_ct == gap)) { 4771 // we added an extra thread to this place; move on to next place 4772 if (place == last_place) { 4773 place = first_place; 4774 } else if (place == (int)(__kmp_affinity_num_masks - 1)) { 4775 place = 0; 4776 } else { 4777 place++; 4778 } 4779 s_count = 0; 4780 gap_ct = 1; 4781 rem--; 4782 } else if (s_count == S) { // place is full; don't add extra thread 4783 if (place == last_place) { 4784 place = first_place; 4785 } else if (place == (int)(__kmp_affinity_num_masks - 1)) { 4786 place = 0; 4787 } else { 4788 place++; 4789 } 4790 gap_ct++; 4791 s_count = 0; 4792 } 4793 4794 KA_TRACE(100, ("__kmp_partition_places: spread: T#%d(%d:%d) place %d " 4795 "partition = [%d,%d]\n", 4796 __kmp_gtid_from_thread(team->t.t_threads[f]), 4797 team->t.t_id, f, th->th.th_new_place, 4798 th->th.th_first_place, th->th.th_last_place)); 4799 } 4800 KMP_DEBUG_ASSERT(update_master_only || place == masters_place); 4801 } 4802 } break; 4803 4804 default: 4805 break; 4806 } 4807 4808 KA_TRACE(20, ("__kmp_partition_places: exit T#%d\n", team->t.t_id)); 4809 } 4810 4811 #endif /* OMP_40_ENABLED && KMP_AFFINITY_SUPPORTED */ 4812 4813 /* allocate a new team data structure to use. take one off of the free pool if 4814 available */ 4815 kmp_team_t * 4816 __kmp_allocate_team(kmp_root_t *root, int new_nproc, int max_nproc, 4817 #if OMPT_SUPPORT 4818 ompt_parallel_id_t ompt_parallel_id, 4819 #endif 4820 #if OMP_40_ENABLED 4821 kmp_proc_bind_t new_proc_bind, 4822 #endif 4823 kmp_internal_control_t *new_icvs, 4824 int argc USE_NESTED_HOT_ARG(kmp_info_t *master)) { 4825 KMP_TIME_DEVELOPER_PARTITIONED_BLOCK(KMP_allocate_team); 4826 int f; 4827 kmp_team_t *team; 4828 int use_hot_team = !root->r.r_active; 4829 int level = 0; 4830 4831 KA_TRACE(20, ("__kmp_allocate_team: called\n")); 4832 KMP_DEBUG_ASSERT(new_nproc >= 1 && argc >= 0); 4833 KMP_DEBUG_ASSERT(max_nproc >= new_nproc); 4834 KMP_MB(); 4835 4836 #if KMP_NESTED_HOT_TEAMS 4837 kmp_hot_team_ptr_t *hot_teams; 4838 if (master) { 4839 team = master->th.th_team; 4840 level = team->t.t_active_level; 4841 if (master->th.th_teams_microtask) { // in teams construct? 4842 if (master->th.th_teams_size.nteams > 1 && 4843 ( // #teams > 1 4844 team->t.t_pkfn == 4845 (microtask_t)__kmp_teams_master || // inner fork of the teams 4846 master->th.th_teams_level < 4847 team->t.t_level)) { // or nested parallel inside the teams 4848 ++level; // not increment if #teams==1, or for outer fork of the teams; 4849 // increment otherwise 4850 } 4851 } 4852 hot_teams = master->th.th_hot_teams; 4853 if (level < __kmp_hot_teams_max_level && hot_teams && 4854 hot_teams[level] 4855 .hot_team) { // hot team has already been allocated for given level 4856 use_hot_team = 1; 4857 } else { 4858 use_hot_team = 0; 4859 } 4860 } 4861 #endif 4862 // Optimization to use a "hot" team 4863 if (use_hot_team && new_nproc > 1) { 4864 KMP_DEBUG_ASSERT(new_nproc == max_nproc); 4865 #if KMP_NESTED_HOT_TEAMS 4866 team = hot_teams[level].hot_team; 4867 #else 4868 team = root->r.r_hot_team; 4869 #endif 4870 #if KMP_DEBUG 4871 if (__kmp_tasking_mode != tskm_immediate_exec) { 4872 KA_TRACE(20, ("__kmp_allocate_team: hot team task_team[0] = %p " 4873 "task_team[1] = %p before reinit\n", 4874 team->t.t_task_team[0], team->t.t_task_team[1])); 4875 } 4876 #endif 4877 4878 // Has the number of threads changed? 4879 /* Let's assume the most common case is that the number of threads is 4880 unchanged, and put that case first. */ 4881 if (team->t.t_nproc == new_nproc) { // Check changes in number of threads 4882 KA_TRACE(20, ("__kmp_allocate_team: reusing hot team\n")); 4883 // This case can mean that omp_set_num_threads() was called and the hot 4884 // team size was already reduced, so we check the special flag 4885 if (team->t.t_size_changed == -1) { 4886 team->t.t_size_changed = 1; 4887 } else { 4888 KMP_CHECK_UPDATE(team->t.t_size_changed, 0); 4889 } 4890 4891 // TODO???: team->t.t_max_active_levels = new_max_active_levels; 4892 kmp_r_sched_t new_sched = new_icvs->sched; 4893 if (team->t.t_sched.r_sched_type != new_sched.r_sched_type || 4894 team->t.t_sched.chunk != new_sched.chunk) 4895 team->t.t_sched = 4896 new_sched; // set master's schedule as new run-time schedule 4897 4898 __kmp_reinitialize_team(team, new_icvs, 4899 root->r.r_uber_thread->th.th_ident); 4900 4901 KF_TRACE(10, ("__kmp_allocate_team2: T#%d, this_thread=%p team=%p\n", 0, 4902 team->t.t_threads[0], team)); 4903 __kmp_push_current_task_to_thread(team->t.t_threads[0], team, 0); 4904 4905 #if OMP_40_ENABLED 4906 #if KMP_AFFINITY_SUPPORTED 4907 if ((team->t.t_size_changed == 0) && 4908 (team->t.t_proc_bind == new_proc_bind)) { 4909 if (new_proc_bind == proc_bind_spread) { 4910 __kmp_partition_places( 4911 team, 1); // add flag to update only master for spread 4912 } 4913 KA_TRACE(200, ("__kmp_allocate_team: reusing hot team #%d bindings: " 4914 "proc_bind = %d, partition = [%d,%d]\n", 4915 team->t.t_id, new_proc_bind, team->t.t_first_place, 4916 team->t.t_last_place)); 4917 } else { 4918 KMP_CHECK_UPDATE(team->t.t_proc_bind, new_proc_bind); 4919 __kmp_partition_places(team); 4920 } 4921 #else 4922 KMP_CHECK_UPDATE(team->t.t_proc_bind, new_proc_bind); 4923 #endif /* KMP_AFFINITY_SUPPORTED */ 4924 #endif /* OMP_40_ENABLED */ 4925 } else if (team->t.t_nproc > new_nproc) { 4926 KA_TRACE(20, 4927 ("__kmp_allocate_team: decreasing hot team thread count to %d\n", 4928 new_nproc)); 4929 4930 team->t.t_size_changed = 1; 4931 #if KMP_NESTED_HOT_TEAMS 4932 if (__kmp_hot_teams_mode == 0) { 4933 // AC: saved number of threads should correspond to team's value in this 4934 // mode, can be bigger in mode 1, when hot team has threads in reserve 4935 KMP_DEBUG_ASSERT(hot_teams[level].hot_team_nth == team->t.t_nproc); 4936 hot_teams[level].hot_team_nth = new_nproc; 4937 #endif // KMP_NESTED_HOT_TEAMS 4938 /* release the extra threads we don't need any more */ 4939 for (f = new_nproc; f < team->t.t_nproc; f++) { 4940 KMP_DEBUG_ASSERT(team->t.t_threads[f]); 4941 if (__kmp_tasking_mode != tskm_immediate_exec) { 4942 // When decreasing team size, threads no longer in the team should 4943 // unref task team. 4944 team->t.t_threads[f]->th.th_task_team = NULL; 4945 } 4946 __kmp_free_thread(team->t.t_threads[f]); 4947 team->t.t_threads[f] = NULL; 4948 } 4949 #if KMP_NESTED_HOT_TEAMS 4950 } // (__kmp_hot_teams_mode == 0) 4951 else { 4952 // When keeping extra threads in team, switch threads to wait on own 4953 // b_go flag 4954 for (f = new_nproc; f < team->t.t_nproc; ++f) { 4955 KMP_DEBUG_ASSERT(team->t.t_threads[f]); 4956 kmp_balign_t *balign = team->t.t_threads[f]->th.th_bar; 4957 for (int b = 0; b < bs_last_barrier; ++b) { 4958 if (balign[b].bb.wait_flag == KMP_BARRIER_PARENT_FLAG) { 4959 balign[b].bb.wait_flag = KMP_BARRIER_SWITCH_TO_OWN_FLAG; 4960 } 4961 KMP_CHECK_UPDATE(balign[b].bb.leaf_kids, 0); 4962 } 4963 } 4964 } 4965 #endif // KMP_NESTED_HOT_TEAMS 4966 team->t.t_nproc = new_nproc; 4967 // TODO???: team->t.t_max_active_levels = new_max_active_levels; 4968 if (team->t.t_sched.r_sched_type != new_icvs->sched.r_sched_type || 4969 team->t.t_sched.chunk != new_icvs->sched.chunk) 4970 team->t.t_sched = new_icvs->sched; 4971 __kmp_reinitialize_team(team, new_icvs, 4972 root->r.r_uber_thread->th.th_ident); 4973 4974 /* update the remaining threads */ 4975 for (f = 0; f < new_nproc; ++f) { 4976 team->t.t_threads[f]->th.th_team_nproc = new_nproc; 4977 } 4978 // restore the current task state of the master thread: should be the 4979 // implicit task 4980 KF_TRACE(10, ("__kmp_allocate_team: T#%d, this_thread=%p team=%p\n", 0, 4981 team->t.t_threads[0], team)); 4982 4983 __kmp_push_current_task_to_thread(team->t.t_threads[0], team, 0); 4984 4985 #ifdef KMP_DEBUG 4986 for (f = 0; f < team->t.t_nproc; f++) { 4987 KMP_DEBUG_ASSERT(team->t.t_threads[f] && 4988 team->t.t_threads[f]->th.th_team_nproc == 4989 team->t.t_nproc); 4990 } 4991 #endif 4992 4993 #if OMP_40_ENABLED 4994 KMP_CHECK_UPDATE(team->t.t_proc_bind, new_proc_bind); 4995 #if KMP_AFFINITY_SUPPORTED 4996 __kmp_partition_places(team); 4997 #endif 4998 #endif 4999 } else { // team->t.t_nproc < new_nproc 5000 #if KMP_OS_LINUX && KMP_AFFINITY_SUPPORTED 5001 kmp_affin_mask_t *old_mask; 5002 if (KMP_AFFINITY_CAPABLE()) { 5003 KMP_CPU_ALLOC(old_mask); 5004 } 5005 #endif 5006 5007 KA_TRACE(20, 5008 ("__kmp_allocate_team: increasing hot team thread count to %d\n", 5009 new_nproc)); 5010 5011 team->t.t_size_changed = 1; 5012 5013 #if KMP_NESTED_HOT_TEAMS 5014 int avail_threads = hot_teams[level].hot_team_nth; 5015 if (new_nproc < avail_threads) 5016 avail_threads = new_nproc; 5017 kmp_info_t **other_threads = team->t.t_threads; 5018 for (f = team->t.t_nproc; f < avail_threads; ++f) { 5019 // Adjust barrier data of reserved threads (if any) of the team 5020 // Other data will be set in __kmp_initialize_info() below. 5021 int b; 5022 kmp_balign_t *balign = other_threads[f]->th.th_bar; 5023 for (b = 0; b < bs_last_barrier; ++b) { 5024 balign[b].bb.b_arrived = team->t.t_bar[b].b_arrived; 5025 KMP_DEBUG_ASSERT(balign[b].bb.wait_flag != KMP_BARRIER_PARENT_FLAG); 5026 #if USE_DEBUGGER 5027 balign[b].bb.b_worker_arrived = team->t.t_bar[b].b_team_arrived; 5028 #endif 5029 } 5030 } 5031 if (hot_teams[level].hot_team_nth >= new_nproc) { 5032 // we have all needed threads in reserve, no need to allocate any 5033 // this only possible in mode 1, cannot have reserved threads in mode 0 5034 KMP_DEBUG_ASSERT(__kmp_hot_teams_mode == 1); 5035 team->t.t_nproc = new_nproc; // just get reserved threads involved 5036 } else { 5037 // we may have some threads in reserve, but not enough 5038 team->t.t_nproc = 5039 hot_teams[level] 5040 .hot_team_nth; // get reserved threads involved if any 5041 hot_teams[level].hot_team_nth = new_nproc; // adjust hot team max size 5042 #endif // KMP_NESTED_HOT_TEAMS 5043 if (team->t.t_max_nproc < new_nproc) { 5044 /* reallocate larger arrays */ 5045 __kmp_reallocate_team_arrays(team, new_nproc); 5046 __kmp_reinitialize_team(team, new_icvs, NULL); 5047 } 5048 5049 #if KMP_OS_LINUX && KMP_AFFINITY_SUPPORTED 5050 /* Temporarily set full mask for master thread before creation of 5051 workers. The reason is that workers inherit the affinity from master, 5052 so if a lot of workers are created on the single core quickly, they 5053 don't get a chance to set their own affinity for a long time. */ 5054 __kmp_set_thread_affinity_mask_full_tmp(old_mask); 5055 #endif 5056 5057 /* allocate new threads for the hot team */ 5058 for (f = team->t.t_nproc; f < new_nproc; f++) { 5059 kmp_info_t *new_worker = __kmp_allocate_thread(root, team, f); 5060 KMP_DEBUG_ASSERT(new_worker); 5061 team->t.t_threads[f] = new_worker; 5062 5063 KA_TRACE(20, 5064 ("__kmp_allocate_team: team %d init T#%d arrived: " 5065 "join=%llu, plain=%llu\n", 5066 team->t.t_id, __kmp_gtid_from_tid(f, team), team->t.t_id, f, 5067 team->t.t_bar[bs_forkjoin_barrier].b_arrived, 5068 team->t.t_bar[bs_plain_barrier].b_arrived)); 5069 5070 { // Initialize barrier data for new threads. 5071 int b; 5072 kmp_balign_t *balign = new_worker->th.th_bar; 5073 for (b = 0; b < bs_last_barrier; ++b) { 5074 balign[b].bb.b_arrived = team->t.t_bar[b].b_arrived; 5075 KMP_DEBUG_ASSERT(balign[b].bb.wait_flag != 5076 KMP_BARRIER_PARENT_FLAG); 5077 #if USE_DEBUGGER 5078 balign[b].bb.b_worker_arrived = team->t.t_bar[b].b_team_arrived; 5079 #endif 5080 } 5081 } 5082 } 5083 5084 #if KMP_OS_LINUX && KMP_AFFINITY_SUPPORTED 5085 if (KMP_AFFINITY_CAPABLE()) { 5086 /* Restore initial master thread's affinity mask */ 5087 __kmp_set_system_affinity(old_mask, TRUE); 5088 KMP_CPU_FREE(old_mask); 5089 } 5090 #endif 5091 #if KMP_NESTED_HOT_TEAMS 5092 } // end of check of t_nproc vs. new_nproc vs. hot_team_nth 5093 #endif // KMP_NESTED_HOT_TEAMS 5094 /* make sure everyone is syncronized */ 5095 int old_nproc = team->t.t_nproc; // save old value and use to update only 5096 // new threads below 5097 __kmp_initialize_team(team, new_nproc, new_icvs, 5098 root->r.r_uber_thread->th.th_ident); 5099 5100 /* reinitialize the threads */ 5101 KMP_DEBUG_ASSERT(team->t.t_nproc == new_nproc); 5102 for (f = 0; f < team->t.t_nproc; ++f) 5103 __kmp_initialize_info(team->t.t_threads[f], team, f, 5104 __kmp_gtid_from_tid(f, team)); 5105 if (level) { // set th_task_state for new threads in nested hot team 5106 // __kmp_initialize_info() no longer zeroes th_task_state, so we should 5107 // only need to set the th_task_state for the new threads. th_task_state 5108 // for master thread will not be accurate until after this in 5109 // __kmp_fork_call(), so we look to the master's memo_stack to get the 5110 // correct value. 5111 for (f = old_nproc; f < team->t.t_nproc; ++f) 5112 team->t.t_threads[f]->th.th_task_state = 5113 team->t.t_threads[0]->th.th_task_state_memo_stack[level]; 5114 } else { // set th_task_state for new threads in non-nested hot team 5115 int old_state = 5116 team->t.t_threads[0]->th.th_task_state; // copy master's state 5117 for (f = old_nproc; f < team->t.t_nproc; ++f) 5118 team->t.t_threads[f]->th.th_task_state = old_state; 5119 } 5120 5121 #ifdef KMP_DEBUG 5122 for (f = 0; f < team->t.t_nproc; ++f) { 5123 KMP_DEBUG_ASSERT(team->t.t_threads[f] && 5124 team->t.t_threads[f]->th.th_team_nproc == 5125 team->t.t_nproc); 5126 } 5127 #endif 5128 5129 #if OMP_40_ENABLED 5130 KMP_CHECK_UPDATE(team->t.t_proc_bind, new_proc_bind); 5131 #if KMP_AFFINITY_SUPPORTED 5132 __kmp_partition_places(team); 5133 #endif 5134 #endif 5135 } // Check changes in number of threads 5136 5137 #if OMP_40_ENABLED 5138 kmp_info_t *master = team->t.t_threads[0]; 5139 if (master->th.th_teams_microtask) { 5140 for (f = 1; f < new_nproc; ++f) { 5141 // propagate teams construct specific info to workers 5142 kmp_info_t *thr = team->t.t_threads[f]; 5143 thr->th.th_teams_microtask = master->th.th_teams_microtask; 5144 thr->th.th_teams_level = master->th.th_teams_level; 5145 thr->th.th_teams_size = master->th.th_teams_size; 5146 } 5147 } 5148 #endif /* OMP_40_ENABLED */ 5149 #if KMP_NESTED_HOT_TEAMS 5150 if (level) { 5151 // Sync barrier state for nested hot teams, not needed for outermost hot 5152 // team. 5153 for (f = 1; f < new_nproc; ++f) { 5154 kmp_info_t *thr = team->t.t_threads[f]; 5155 int b; 5156 kmp_balign_t *balign = thr->th.th_bar; 5157 for (b = 0; b < bs_last_barrier; ++b) { 5158 balign[b].bb.b_arrived = team->t.t_bar[b].b_arrived; 5159 KMP_DEBUG_ASSERT(balign[b].bb.wait_flag != KMP_BARRIER_PARENT_FLAG); 5160 #if USE_DEBUGGER 5161 balign[b].bb.b_worker_arrived = team->t.t_bar[b].b_team_arrived; 5162 #endif 5163 } 5164 } 5165 } 5166 #endif // KMP_NESTED_HOT_TEAMS 5167 5168 /* reallocate space for arguments if necessary */ 5169 __kmp_alloc_argv_entries(argc, team, TRUE); 5170 KMP_CHECK_UPDATE(team->t.t_argc, argc); 5171 // The hot team re-uses the previous task team, 5172 // if untouched during the previous release->gather phase. 5173 5174 KF_TRACE(10, (" hot_team = %p\n", team)); 5175 5176 #if KMP_DEBUG 5177 if (__kmp_tasking_mode != tskm_immediate_exec) { 5178 KA_TRACE(20, ("__kmp_allocate_team: hot team task_team[0] = %p " 5179 "task_team[1] = %p after reinit\n", 5180 team->t.t_task_team[0], team->t.t_task_team[1])); 5181 } 5182 #endif 5183 5184 #if OMPT_SUPPORT 5185 __ompt_team_assign_id(team, ompt_parallel_id); 5186 #endif 5187 5188 KMP_MB(); 5189 5190 return team; 5191 } 5192 5193 /* next, let's try to take one from the team pool */ 5194 KMP_MB(); 5195 for (team = CCAST(kmp_team_t *, __kmp_team_pool); (team);) { 5196 /* TODO: consider resizing undersized teams instead of reaping them, now 5197 that we have a resizing mechanism */ 5198 if (team->t.t_max_nproc >= max_nproc) { 5199 /* take this team from the team pool */ 5200 __kmp_team_pool = team->t.t_next_pool; 5201 5202 /* setup the team for fresh use */ 5203 __kmp_initialize_team(team, new_nproc, new_icvs, NULL); 5204 5205 KA_TRACE(20, ("__kmp_allocate_team: setting task_team[0] %p and " 5206 "task_team[1] %p to NULL\n", 5207 &team->t.t_task_team[0], &team->t.t_task_team[1])); 5208 team->t.t_task_team[0] = NULL; 5209 team->t.t_task_team[1] = NULL; 5210 5211 /* reallocate space for arguments if necessary */ 5212 __kmp_alloc_argv_entries(argc, team, TRUE); 5213 KMP_CHECK_UPDATE(team->t.t_argc, argc); 5214 5215 KA_TRACE( 5216 20, ("__kmp_allocate_team: team %d init arrived: join=%u, plain=%u\n", 5217 team->t.t_id, KMP_INIT_BARRIER_STATE, KMP_INIT_BARRIER_STATE)); 5218 { // Initialize barrier data. 5219 int b; 5220 for (b = 0; b < bs_last_barrier; ++b) { 5221 team->t.t_bar[b].b_arrived = KMP_INIT_BARRIER_STATE; 5222 #if USE_DEBUGGER 5223 team->t.t_bar[b].b_master_arrived = 0; 5224 team->t.t_bar[b].b_team_arrived = 0; 5225 #endif 5226 } 5227 } 5228 5229 #if OMP_40_ENABLED 5230 team->t.t_proc_bind = new_proc_bind; 5231 #endif 5232 5233 KA_TRACE(20, ("__kmp_allocate_team: using team from pool %d.\n", 5234 team->t.t_id)); 5235 5236 #if OMPT_SUPPORT 5237 __ompt_team_assign_id(team, ompt_parallel_id); 5238 #endif 5239 5240 KMP_MB(); 5241 5242 return team; 5243 } 5244 5245 /* reap team if it is too small, then loop back and check the next one */ 5246 // not sure if this is wise, but, will be redone during the hot-teams rewrite. 5247 /* TODO: Use technique to find the right size hot-team, don't reap them */ 5248 team = __kmp_reap_team(team); 5249 __kmp_team_pool = team; 5250 } 5251 5252 /* nothing available in the pool, no matter, make a new team! */ 5253 KMP_MB(); 5254 team = (kmp_team_t *)__kmp_allocate(sizeof(kmp_team_t)); 5255 5256 /* and set it up */ 5257 team->t.t_max_nproc = max_nproc; 5258 /* NOTE well, for some reason allocating one big buffer and dividing it up 5259 seems to really hurt performance a lot on the P4, so, let's not use this */ 5260 __kmp_allocate_team_arrays(team, max_nproc); 5261 5262 KA_TRACE(20, ("__kmp_allocate_team: making a new team\n")); 5263 __kmp_initialize_team(team, new_nproc, new_icvs, NULL); 5264 5265 KA_TRACE(20, ("__kmp_allocate_team: setting task_team[0] %p and task_team[1] " 5266 "%p to NULL\n", 5267 &team->t.t_task_team[0], &team->t.t_task_team[1])); 5268 team->t.t_task_team[0] = NULL; // to be removed, as __kmp_allocate zeroes 5269 // memory, no need to duplicate 5270 team->t.t_task_team[1] = NULL; // to be removed, as __kmp_allocate zeroes 5271 // memory, no need to duplicate 5272 5273 if (__kmp_storage_map) { 5274 __kmp_print_team_storage_map("team", team, team->t.t_id, new_nproc); 5275 } 5276 5277 /* allocate space for arguments */ 5278 __kmp_alloc_argv_entries(argc, team, FALSE); 5279 team->t.t_argc = argc; 5280 5281 KA_TRACE(20, 5282 ("__kmp_allocate_team: team %d init arrived: join=%u, plain=%u\n", 5283 team->t.t_id, KMP_INIT_BARRIER_STATE, KMP_INIT_BARRIER_STATE)); 5284 { // Initialize barrier data. 5285 int b; 5286 for (b = 0; b < bs_last_barrier; ++b) { 5287 team->t.t_bar[b].b_arrived = KMP_INIT_BARRIER_STATE; 5288 #if USE_DEBUGGER 5289 team->t.t_bar[b].b_master_arrived = 0; 5290 team->t.t_bar[b].b_team_arrived = 0; 5291 #endif 5292 } 5293 } 5294 5295 #if OMP_40_ENABLED 5296 team->t.t_proc_bind = new_proc_bind; 5297 #endif 5298 5299 #if OMPT_SUPPORT 5300 __ompt_team_assign_id(team, ompt_parallel_id); 5301 team->t.ompt_serialized_team_info = NULL; 5302 #endif 5303 5304 KMP_MB(); 5305 5306 KA_TRACE(20, ("__kmp_allocate_team: done creating a new team %d.\n", 5307 team->t.t_id)); 5308 5309 return team; 5310 } 5311 5312 /* TODO implement hot-teams at all levels */ 5313 /* TODO implement lazy thread release on demand (disband request) */ 5314 5315 /* free the team. return it to the team pool. release all the threads 5316 * associated with it */ 5317 void __kmp_free_team(kmp_root_t *root, 5318 kmp_team_t *team USE_NESTED_HOT_ARG(kmp_info_t *master)) { 5319 int f; 5320 KA_TRACE(20, ("__kmp_free_team: T#%d freeing team %d\n", __kmp_get_gtid(), 5321 team->t.t_id)); 5322 5323 /* verify state */ 5324 KMP_DEBUG_ASSERT(root); 5325 KMP_DEBUG_ASSERT(team); 5326 KMP_DEBUG_ASSERT(team->t.t_nproc <= team->t.t_max_nproc); 5327 KMP_DEBUG_ASSERT(team->t.t_threads); 5328 5329 int use_hot_team = team == root->r.r_hot_team; 5330 #if KMP_NESTED_HOT_TEAMS 5331 int level; 5332 kmp_hot_team_ptr_t *hot_teams; 5333 if (master) { 5334 level = team->t.t_active_level - 1; 5335 if (master->th.th_teams_microtask) { // in teams construct? 5336 if (master->th.th_teams_size.nteams > 1) { 5337 ++level; // level was not increased in teams construct for 5338 // team_of_masters 5339 } 5340 if (team->t.t_pkfn != (microtask_t)__kmp_teams_master && 5341 master->th.th_teams_level == team->t.t_level) { 5342 ++level; // level was not increased in teams construct for 5343 // team_of_workers before the parallel 5344 } // team->t.t_level will be increased inside parallel 5345 } 5346 hot_teams = master->th.th_hot_teams; 5347 if (level < __kmp_hot_teams_max_level) { 5348 KMP_DEBUG_ASSERT(team == hot_teams[level].hot_team); 5349 use_hot_team = 1; 5350 } 5351 } 5352 #endif // KMP_NESTED_HOT_TEAMS 5353 5354 /* team is done working */ 5355 TCW_SYNC_PTR(team->t.t_pkfn, 5356 NULL); // Important for Debugging Support Library. 5357 team->t.t_copyin_counter = 0; // init counter for possible reuse 5358 // Do not reset pointer to parent team to NULL for hot teams. 5359 5360 /* if we are non-hot team, release our threads */ 5361 if (!use_hot_team) { 5362 if (__kmp_tasking_mode != tskm_immediate_exec) { 5363 // Wait for threads to reach reapable state 5364 for (f = 1; f < team->t.t_nproc; ++f) { 5365 KMP_DEBUG_ASSERT(team->t.t_threads[f]); 5366 kmp_info_t *th = team->t.t_threads[f]; 5367 volatile kmp_uint32 *state = &th->th.th_reap_state; 5368 while (*state != KMP_SAFE_TO_REAP) { 5369 #if KMP_OS_WINDOWS 5370 // On Windows a thread can be killed at any time, check this 5371 DWORD ecode; 5372 if (!__kmp_is_thread_alive(th, &ecode)) { 5373 *state = KMP_SAFE_TO_REAP; // reset the flag for dead thread 5374 break; 5375 } 5376 #endif 5377 // first check if thread is sleeping 5378 kmp_flag_64 fl(&th->th.th_bar[bs_forkjoin_barrier].bb.b_go, th); 5379 if (fl.is_sleeping()) 5380 fl.resume(__kmp_gtid_from_thread(th)); 5381 KMP_CPU_PAUSE(); 5382 } 5383 } 5384 5385 // Delete task teams 5386 int tt_idx; 5387 for (tt_idx = 0; tt_idx < 2; ++tt_idx) { 5388 kmp_task_team_t *task_team = team->t.t_task_team[tt_idx]; 5389 if (task_team != NULL) { 5390 for (f = 0; f < team->t.t_nproc; 5391 ++f) { // Have all threads unref task teams 5392 team->t.t_threads[f]->th.th_task_team = NULL; 5393 } 5394 KA_TRACE( 5395 20, 5396 ("__kmp_free_team: T#%d deactivating task_team %p on team %d\n", 5397 __kmp_get_gtid(), task_team, team->t.t_id)); 5398 #if KMP_NESTED_HOT_TEAMS 5399 __kmp_free_task_team(master, task_team); 5400 #endif 5401 team->t.t_task_team[tt_idx] = NULL; 5402 } 5403 } 5404 } 5405 5406 // Reset pointer to parent team only for non-hot teams. 5407 team->t.t_parent = NULL; 5408 team->t.t_level = 0; 5409 team->t.t_active_level = 0; 5410 5411 /* free the worker threads */ 5412 for (f = 1; f < team->t.t_nproc; ++f) { 5413 KMP_DEBUG_ASSERT(team->t.t_threads[f]); 5414 __kmp_free_thread(team->t.t_threads[f]); 5415 team->t.t_threads[f] = NULL; 5416 } 5417 5418 /* put the team back in the team pool */ 5419 /* TODO limit size of team pool, call reap_team if pool too large */ 5420 team->t.t_next_pool = CCAST(kmp_team_t *, __kmp_team_pool); 5421 __kmp_team_pool = (volatile kmp_team_t *)team; 5422 } 5423 5424 KMP_MB(); 5425 } 5426 5427 /* reap the team. destroy it, reclaim all its resources and free its memory */ 5428 kmp_team_t *__kmp_reap_team(kmp_team_t *team) { 5429 kmp_team_t *next_pool = team->t.t_next_pool; 5430 5431 KMP_DEBUG_ASSERT(team); 5432 KMP_DEBUG_ASSERT(team->t.t_dispatch); 5433 KMP_DEBUG_ASSERT(team->t.t_disp_buffer); 5434 KMP_DEBUG_ASSERT(team->t.t_threads); 5435 KMP_DEBUG_ASSERT(team->t.t_argv); 5436 5437 /* TODO clean the threads that are a part of this? */ 5438 5439 /* free stuff */ 5440 __kmp_free_team_arrays(team); 5441 if (team->t.t_argv != &team->t.t_inline_argv[0]) 5442 __kmp_free((void *)team->t.t_argv); 5443 __kmp_free(team); 5444 5445 KMP_MB(); 5446 return next_pool; 5447 } 5448 5449 // Free the thread. Don't reap it, just place it on the pool of available 5450 // threads. 5451 // 5452 // Changes for Quad issue 527845: We need a predictable OMP tid <-> gtid 5453 // binding for the affinity mechanism to be useful. 5454 // 5455 // Now, we always keep the free list (__kmp_thread_pool) sorted by gtid. 5456 // However, we want to avoid a potential performance problem by always 5457 // scanning through the list to find the correct point at which to insert 5458 // the thread (potential N**2 behavior). To do this we keep track of the 5459 // last place a thread struct was inserted (__kmp_thread_pool_insert_pt). 5460 // With single-level parallelism, threads will always be added to the tail 5461 // of the list, kept track of by __kmp_thread_pool_insert_pt. With nested 5462 // parallelism, all bets are off and we may need to scan through the entire 5463 // free list. 5464 // 5465 // This change also has a potentially large performance benefit, for some 5466 // applications. Previously, as threads were freed from the hot team, they 5467 // would be placed back on the free list in inverse order. If the hot team 5468 // grew back to it's original size, then the freed thread would be placed 5469 // back on the hot team in reverse order. This could cause bad cache 5470 // locality problems on programs where the size of the hot team regularly 5471 // grew and shrunk. 5472 // 5473 // Now, for single-level parallelism, the OMP tid is alway == gtid. 5474 void __kmp_free_thread(kmp_info_t *this_th) { 5475 int gtid; 5476 kmp_info_t **scan; 5477 kmp_root_t *root = this_th->th.th_root; 5478 5479 KA_TRACE(20, ("__kmp_free_thread: T#%d putting T#%d back on free pool.\n", 5480 __kmp_get_gtid(), this_th->th.th_info.ds.ds_gtid)); 5481 5482 KMP_DEBUG_ASSERT(this_th); 5483 5484 // When moving thread to pool, switch thread to wait on own b_go flag, and 5485 // uninitialized (NULL team). 5486 int b; 5487 kmp_balign_t *balign = this_th->th.th_bar; 5488 for (b = 0; b < bs_last_barrier; ++b) { 5489 if (balign[b].bb.wait_flag == KMP_BARRIER_PARENT_FLAG) 5490 balign[b].bb.wait_flag = KMP_BARRIER_SWITCH_TO_OWN_FLAG; 5491 balign[b].bb.team = NULL; 5492 balign[b].bb.leaf_kids = 0; 5493 } 5494 this_th->th.th_task_state = 0; 5495 5496 /* put thread back on the free pool */ 5497 TCW_PTR(this_th->th.th_team, NULL); 5498 TCW_PTR(this_th->th.th_root, NULL); 5499 TCW_PTR(this_th->th.th_dispatch, NULL); /* NOT NEEDED */ 5500 5501 // If the __kmp_thread_pool_insert_pt is already past the new insert 5502 // point, then we need to re-scan the entire list. 5503 gtid = this_th->th.th_info.ds.ds_gtid; 5504 if (__kmp_thread_pool_insert_pt != NULL) { 5505 KMP_DEBUG_ASSERT(__kmp_thread_pool != NULL); 5506 if (__kmp_thread_pool_insert_pt->th.th_info.ds.ds_gtid > gtid) { 5507 __kmp_thread_pool_insert_pt = NULL; 5508 } 5509 } 5510 5511 // Scan down the list to find the place to insert the thread. 5512 // scan is the address of a link in the list, possibly the address of 5513 // __kmp_thread_pool itself. 5514 // 5515 // In the absence of nested parallism, the for loop will have 0 iterations. 5516 if (__kmp_thread_pool_insert_pt != NULL) { 5517 scan = &(__kmp_thread_pool_insert_pt->th.th_next_pool); 5518 } else { 5519 scan = CCAST(kmp_info_t **, &__kmp_thread_pool); 5520 } 5521 for (; (*scan != NULL) && ((*scan)->th.th_info.ds.ds_gtid < gtid); 5522 scan = &((*scan)->th.th_next_pool)) 5523 ; 5524 5525 // Insert the new element on the list, and set __kmp_thread_pool_insert_pt 5526 // to its address. 5527 TCW_PTR(this_th->th.th_next_pool, *scan); 5528 __kmp_thread_pool_insert_pt = *scan = this_th; 5529 KMP_DEBUG_ASSERT((this_th->th.th_next_pool == NULL) || 5530 (this_th->th.th_info.ds.ds_gtid < 5531 this_th->th.th_next_pool->th.th_info.ds.ds_gtid)); 5532 TCW_4(this_th->th.th_in_pool, TRUE); 5533 __kmp_thread_pool_nth++; 5534 5535 TCW_4(__kmp_nth, __kmp_nth - 1); 5536 root->r.r_cg_nthreads--; 5537 5538 #ifdef KMP_ADJUST_BLOCKTIME 5539 /* Adjust blocktime back to user setting or default if necessary */ 5540 /* Middle initialization might never have occurred */ 5541 if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) { 5542 KMP_DEBUG_ASSERT(__kmp_avail_proc > 0); 5543 if (__kmp_nth <= __kmp_avail_proc) { 5544 __kmp_zero_bt = FALSE; 5545 } 5546 } 5547 #endif /* KMP_ADJUST_BLOCKTIME */ 5548 5549 KMP_MB(); 5550 } 5551 5552 /* ------------------------------------------------------------------------ */ 5553 5554 void *__kmp_launch_thread(kmp_info_t *this_thr) { 5555 int gtid = this_thr->th.th_info.ds.ds_gtid; 5556 /* void *stack_data;*/ 5557 kmp_team_t *(*volatile pteam); 5558 5559 KMP_MB(); 5560 KA_TRACE(10, ("__kmp_launch_thread: T#%d start\n", gtid)); 5561 5562 if (__kmp_env_consistency_check) { 5563 this_thr->th.th_cons = __kmp_allocate_cons_stack(gtid); // ATT: Memory leak? 5564 } 5565 5566 #if OMPT_SUPPORT 5567 if (ompt_enabled) { 5568 this_thr->th.ompt_thread_info.state = ompt_state_overhead; 5569 this_thr->th.ompt_thread_info.wait_id = 0; 5570 this_thr->th.ompt_thread_info.idle_frame = __builtin_frame_address(0); 5571 if (ompt_callbacks.ompt_callback(ompt_event_thread_begin)) { 5572 __ompt_thread_begin(ompt_thread_worker, gtid); 5573 } 5574 } 5575 #endif 5576 5577 /* This is the place where threads wait for work */ 5578 while (!TCR_4(__kmp_global.g.g_done)) { 5579 KMP_DEBUG_ASSERT(this_thr == __kmp_threads[gtid]); 5580 KMP_MB(); 5581 5582 /* wait for work to do */ 5583 KA_TRACE(20, ("__kmp_launch_thread: T#%d waiting for work\n", gtid)); 5584 5585 #if OMPT_SUPPORT 5586 if (ompt_enabled) { 5587 this_thr->th.ompt_thread_info.state = ompt_state_idle; 5588 } 5589 #endif 5590 5591 /* No tid yet since not part of a team */ 5592 __kmp_fork_barrier(gtid, KMP_GTID_DNE); 5593 5594 #if OMPT_SUPPORT 5595 if (ompt_enabled) { 5596 this_thr->th.ompt_thread_info.state = ompt_state_overhead; 5597 } 5598 #endif 5599 5600 pteam = (kmp_team_t * (*))(&this_thr->th.th_team); 5601 5602 /* have we been allocated? */ 5603 if (TCR_SYNC_PTR(*pteam) && !TCR_4(__kmp_global.g.g_done)) { 5604 #if OMPT_SUPPORT 5605 ompt_task_info_t *task_info; 5606 ompt_parallel_id_t my_parallel_id; 5607 if (ompt_enabled) { 5608 task_info = __ompt_get_taskinfo(0); 5609 my_parallel_id = (*pteam)->t.ompt_team_info.parallel_id; 5610 } 5611 #endif 5612 /* we were just woken up, so run our new task */ 5613 if (TCR_SYNC_PTR((*pteam)->t.t_pkfn) != NULL) { 5614 int rc; 5615 KA_TRACE(20, 5616 ("__kmp_launch_thread: T#%d(%d:%d) invoke microtask = %p\n", 5617 gtid, (*pteam)->t.t_id, __kmp_tid_from_gtid(gtid), 5618 (*pteam)->t.t_pkfn)); 5619 5620 updateHWFPControl(*pteam); 5621 5622 #if OMPT_SUPPORT 5623 if (ompt_enabled) { 5624 this_thr->th.ompt_thread_info.state = ompt_state_work_parallel; 5625 // Initialize OMPT task id for implicit task. 5626 int tid = __kmp_tid_from_gtid(gtid); 5627 task_info->task_id = __ompt_task_id_new(tid); 5628 } 5629 #endif 5630 5631 { 5632 KMP_TIME_PARTITIONED_BLOCK(OMP_parallel); 5633 KMP_SET_THREAD_STATE_BLOCK(IMPLICIT_TASK); 5634 rc = (*pteam)->t.t_invoke(gtid); 5635 } 5636 KMP_ASSERT(rc); 5637 5638 #if OMPT_SUPPORT 5639 if (ompt_enabled) { 5640 /* no frame set while outside task */ 5641 task_info->frame.exit_runtime_frame = NULL; 5642 5643 this_thr->th.ompt_thread_info.state = ompt_state_overhead; 5644 } 5645 #endif 5646 KMP_MB(); 5647 KA_TRACE(20, ("__kmp_launch_thread: T#%d(%d:%d) done microtask = %p\n", 5648 gtid, (*pteam)->t.t_id, __kmp_tid_from_gtid(gtid), 5649 (*pteam)->t.t_pkfn)); 5650 } 5651 /* join barrier after parallel region */ 5652 __kmp_join_barrier(gtid); 5653 #if OMPT_SUPPORT && OMPT_TRACE 5654 if (ompt_enabled) { 5655 if (ompt_callbacks.ompt_callback(ompt_event_implicit_task_end)) { 5656 // don't access *pteam here: it may have already been freed 5657 // by the master thread behind the barrier (possible race) 5658 ompt_callbacks.ompt_callback(ompt_event_implicit_task_end)( 5659 my_parallel_id, task_info->task_id); 5660 } 5661 task_info->frame.exit_runtime_frame = NULL; 5662 task_info->task_id = 0; 5663 } 5664 #endif 5665 } 5666 } 5667 TCR_SYNC_PTR((intptr_t)__kmp_global.g.g_done); 5668 5669 #if OMPT_SUPPORT 5670 if (ompt_enabled && ompt_callbacks.ompt_callback(ompt_event_thread_end)) { 5671 __ompt_thread_end(ompt_thread_worker, gtid); 5672 } 5673 #endif 5674 5675 this_thr->th.th_task_team = NULL; 5676 /* run the destructors for the threadprivate data for this thread */ 5677 __kmp_common_destroy_gtid(gtid); 5678 5679 KA_TRACE(10, ("__kmp_launch_thread: T#%d done\n", gtid)); 5680 KMP_MB(); 5681 return this_thr; 5682 } 5683 5684 /* ------------------------------------------------------------------------ */ 5685 5686 void __kmp_internal_end_dest(void *specific_gtid) { 5687 #if KMP_COMPILER_ICC 5688 #pragma warning(push) 5689 #pragma warning(disable : 810) // conversion from "void *" to "int" may lose 5690 // significant bits 5691 #endif 5692 // Make sure no significant bits are lost 5693 int gtid = (kmp_intptr_t)specific_gtid - 1; 5694 #if KMP_COMPILER_ICC 5695 #pragma warning(pop) 5696 #endif 5697 5698 KA_TRACE(30, ("__kmp_internal_end_dest: T#%d\n", gtid)); 5699 /* NOTE: the gtid is stored as gitd+1 in the thread-local-storage 5700 * this is because 0 is reserved for the nothing-stored case */ 5701 5702 /* josh: One reason for setting the gtid specific data even when it is being 5703 destroyed by pthread is to allow gtid lookup through thread specific data 5704 (__kmp_gtid_get_specific). Some of the code, especially stat code, 5705 that gets executed in the call to __kmp_internal_end_thread, actually 5706 gets the gtid through the thread specific data. Setting it here seems 5707 rather inelegant and perhaps wrong, but allows __kmp_internal_end_thread 5708 to run smoothly. 5709 todo: get rid of this after we remove the dependence on 5710 __kmp_gtid_get_specific */ 5711 if (gtid >= 0 && KMP_UBER_GTID(gtid)) 5712 __kmp_gtid_set_specific(gtid); 5713 #ifdef KMP_TDATA_GTID 5714 __kmp_gtid = gtid; 5715 #endif 5716 __kmp_internal_end_thread(gtid); 5717 } 5718 5719 #if KMP_OS_UNIX && KMP_DYNAMIC_LIB 5720 5721 // 2009-09-08 (lev): It looks the destructor does not work. In simple test cases 5722 // destructors work perfectly, but in real libomp.so I have no evidence it is 5723 // ever called. However, -fini linker option in makefile.mk works fine. 5724 5725 __attribute__((destructor)) void __kmp_internal_end_dtor(void) { 5726 __kmp_internal_end_atexit(); 5727 } 5728 5729 void __kmp_internal_end_fini(void) { __kmp_internal_end_atexit(); } 5730 5731 #endif 5732 5733 /* [Windows] josh: when the atexit handler is called, there may still be more 5734 than one thread alive */ 5735 void __kmp_internal_end_atexit(void) { 5736 KA_TRACE(30, ("__kmp_internal_end_atexit\n")); 5737 /* [Windows] 5738 josh: ideally, we want to completely shutdown the library in this atexit 5739 handler, but stat code that depends on thread specific data for gtid fails 5740 because that data becomes unavailable at some point during the shutdown, so 5741 we call __kmp_internal_end_thread instead. We should eventually remove the 5742 dependency on __kmp_get_specific_gtid in the stat code and use 5743 __kmp_internal_end_library to cleanly shutdown the library. 5744 5745 // TODO: Can some of this comment about GVS be removed? 5746 I suspect that the offending stat code is executed when the calling thread 5747 tries to clean up a dead root thread's data structures, resulting in GVS 5748 code trying to close the GVS structures for that thread, but since the stat 5749 code uses __kmp_get_specific_gtid to get the gtid with the assumption that 5750 the calling thread is cleaning up itself instead of another thread, it get 5751 confused. This happens because allowing a thread to unregister and cleanup 5752 another thread is a recent modification for addressing an issue. 5753 Based on the current design (20050722), a thread may end up 5754 trying to unregister another thread only if thread death does not trigger 5755 the calling of __kmp_internal_end_thread. For Linux* OS, there is the 5756 thread specific data destructor function to detect thread death. For 5757 Windows dynamic, there is DllMain(THREAD_DETACH). For Windows static, there 5758 is nothing. Thus, the workaround is applicable only for Windows static 5759 stat library. */ 5760 __kmp_internal_end_library(-1); 5761 #if KMP_OS_WINDOWS 5762 __kmp_close_console(); 5763 #endif 5764 } 5765 5766 static void __kmp_reap_thread(kmp_info_t *thread, int is_root) { 5767 // It is assumed __kmp_forkjoin_lock is acquired. 5768 5769 int gtid; 5770 5771 KMP_DEBUG_ASSERT(thread != NULL); 5772 5773 gtid = thread->th.th_info.ds.ds_gtid; 5774 5775 if (!is_root) { 5776 5777 if (__kmp_dflt_blocktime != KMP_MAX_BLOCKTIME) { 5778 /* Assume the threads are at the fork barrier here */ 5779 KA_TRACE( 5780 20, ("__kmp_reap_thread: releasing T#%d from fork barrier for reap\n", 5781 gtid)); 5782 /* Need release fence here to prevent seg faults for tree forkjoin barrier 5783 * (GEH) */ 5784 ANNOTATE_HAPPENS_BEFORE(thread); 5785 kmp_flag_64 flag(&thread->th.th_bar[bs_forkjoin_barrier].bb.b_go, thread); 5786 __kmp_release_64(&flag); 5787 }; // if 5788 5789 // Terminate OS thread. 5790 __kmp_reap_worker(thread); 5791 5792 // The thread was killed asynchronously. If it was actively 5793 // spinning in the thread pool, decrement the global count. 5794 // 5795 // There is a small timing hole here - if the worker thread was just waking 5796 // up after sleeping in the pool, had reset it's th_active_in_pool flag but 5797 // not decremented the global counter __kmp_thread_pool_active_nth yet, then 5798 // the global counter might not get updated. 5799 // 5800 // Currently, this can only happen as the library is unloaded, 5801 // so there are no harmful side effects. 5802 if (thread->th.th_active_in_pool) { 5803 thread->th.th_active_in_pool = FALSE; 5804 KMP_TEST_THEN_DEC32(&__kmp_thread_pool_active_nth); 5805 KMP_DEBUG_ASSERT(TCR_4(__kmp_thread_pool_active_nth) >= 0); 5806 } 5807 5808 // Decrement # of [worker] threads in the pool. 5809 KMP_DEBUG_ASSERT(__kmp_thread_pool_nth > 0); 5810 --__kmp_thread_pool_nth; 5811 }; // if 5812 5813 __kmp_free_implicit_task(thread); 5814 5815 // Free the fast memory for tasking 5816 #if USE_FAST_MEMORY 5817 __kmp_free_fast_memory(thread); 5818 #endif /* USE_FAST_MEMORY */ 5819 5820 __kmp_suspend_uninitialize_thread(thread); 5821 5822 KMP_DEBUG_ASSERT(__kmp_threads[gtid] == thread); 5823 TCW_SYNC_PTR(__kmp_threads[gtid], NULL); 5824 5825 --__kmp_all_nth; 5826 // __kmp_nth was decremented when thread is added to the pool. 5827 5828 #ifdef KMP_ADJUST_BLOCKTIME 5829 /* Adjust blocktime back to user setting or default if necessary */ 5830 /* Middle initialization might never have occurred */ 5831 if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) { 5832 KMP_DEBUG_ASSERT(__kmp_avail_proc > 0); 5833 if (__kmp_nth <= __kmp_avail_proc) { 5834 __kmp_zero_bt = FALSE; 5835 } 5836 } 5837 #endif /* KMP_ADJUST_BLOCKTIME */ 5838 5839 /* free the memory being used */ 5840 if (__kmp_env_consistency_check) { 5841 if (thread->th.th_cons) { 5842 __kmp_free_cons_stack(thread->th.th_cons); 5843 thread->th.th_cons = NULL; 5844 }; // if 5845 } 5846 5847 if (thread->th.th_pri_common != NULL) { 5848 __kmp_free(thread->th.th_pri_common); 5849 thread->th.th_pri_common = NULL; 5850 }; // if 5851 5852 if (thread->th.th_task_state_memo_stack != NULL) { 5853 __kmp_free(thread->th.th_task_state_memo_stack); 5854 thread->th.th_task_state_memo_stack = NULL; 5855 } 5856 5857 #if KMP_USE_BGET 5858 if (thread->th.th_local.bget_data != NULL) { 5859 __kmp_finalize_bget(thread); 5860 }; // if 5861 #endif 5862 5863 #if KMP_AFFINITY_SUPPORTED 5864 if (thread->th.th_affin_mask != NULL) { 5865 KMP_CPU_FREE(thread->th.th_affin_mask); 5866 thread->th.th_affin_mask = NULL; 5867 }; // if 5868 #endif /* KMP_AFFINITY_SUPPORTED */ 5869 5870 __kmp_reap_team(thread->th.th_serial_team); 5871 thread->th.th_serial_team = NULL; 5872 __kmp_free(thread); 5873 5874 KMP_MB(); 5875 5876 } // __kmp_reap_thread 5877 5878 static void __kmp_internal_end(void) { 5879 int i; 5880 5881 /* First, unregister the library */ 5882 __kmp_unregister_library(); 5883 5884 #if KMP_OS_WINDOWS 5885 /* In Win static library, we can't tell when a root actually dies, so we 5886 reclaim the data structures for any root threads that have died but not 5887 unregistered themselves, in order to shut down cleanly. 5888 In Win dynamic library we also can't tell when a thread dies. */ 5889 __kmp_reclaim_dead_roots(); // AC: moved here to always clean resources of 5890 // dead roots 5891 #endif 5892 5893 for (i = 0; i < __kmp_threads_capacity; i++) 5894 if (__kmp_root[i]) 5895 if (__kmp_root[i]->r.r_active) 5896 break; 5897 KMP_MB(); /* Flush all pending memory write invalidates. */ 5898 TCW_SYNC_4(__kmp_global.g.g_done, TRUE); 5899 5900 if (i < __kmp_threads_capacity) { 5901 #if KMP_USE_MONITOR 5902 // 2009-09-08 (lev): Other alive roots found. Why do we kill the monitor?? 5903 KMP_MB(); /* Flush all pending memory write invalidates. */ 5904 5905 // Need to check that monitor was initialized before reaping it. If we are 5906 // called form __kmp_atfork_child (which sets __kmp_init_parallel = 0), then 5907 // __kmp_monitor will appear to contain valid data, but it is only valid in the 5908 // parent process, not the child. 5909 // New behavior (201008): instead of keying off of the flag 5910 // __kmp_init_parallel, the monitor thread creation is keyed off 5911 // of the new flag __kmp_init_monitor. 5912 __kmp_acquire_bootstrap_lock(&__kmp_monitor_lock); 5913 if (TCR_4(__kmp_init_monitor)) { 5914 __kmp_reap_monitor(&__kmp_monitor); 5915 TCW_4(__kmp_init_monitor, 0); 5916 } 5917 __kmp_release_bootstrap_lock(&__kmp_monitor_lock); 5918 KA_TRACE(10, ("__kmp_internal_end: monitor reaped\n")); 5919 #endif // KMP_USE_MONITOR 5920 } else { 5921 /* TODO move this to cleanup code */ 5922 #ifdef KMP_DEBUG 5923 /* make sure that everything has properly ended */ 5924 for (i = 0; i < __kmp_threads_capacity; i++) { 5925 if (__kmp_root[i]) { 5926 // KMP_ASSERT( ! KMP_UBER_GTID( i ) ); // AC: 5927 // there can be uber threads alive here 5928 KMP_ASSERT(!__kmp_root[i]->r.r_active); // TODO: can they be active? 5929 } 5930 } 5931 #endif 5932 5933 KMP_MB(); 5934 5935 // Reap the worker threads. 5936 // This is valid for now, but be careful if threads are reaped sooner. 5937 while (__kmp_thread_pool != NULL) { // Loop thru all the thread in the pool. 5938 // Get the next thread from the pool. 5939 kmp_info_t *thread = CCAST(kmp_info_t *, __kmp_thread_pool); 5940 __kmp_thread_pool = thread->th.th_next_pool; 5941 // Reap it. 5942 KMP_DEBUG_ASSERT(thread->th.th_reap_state == KMP_SAFE_TO_REAP); 5943 thread->th.th_next_pool = NULL; 5944 thread->th.th_in_pool = FALSE; 5945 __kmp_reap_thread(thread, 0); 5946 }; // while 5947 __kmp_thread_pool_insert_pt = NULL; 5948 5949 // Reap teams. 5950 while (__kmp_team_pool != NULL) { // Loop thru all the teams in the pool. 5951 // Get the next team from the pool. 5952 kmp_team_t *team = CCAST(kmp_team_t *, __kmp_team_pool); 5953 __kmp_team_pool = team->t.t_next_pool; 5954 // Reap it. 5955 team->t.t_next_pool = NULL; 5956 __kmp_reap_team(team); 5957 }; // while 5958 5959 __kmp_reap_task_teams(); 5960 5961 for (i = 0; i < __kmp_threads_capacity; ++i) { 5962 // TBD: Add some checking... 5963 // Something like KMP_DEBUG_ASSERT( __kmp_thread[ i ] == NULL ); 5964 } 5965 5966 /* Make sure all threadprivate destructors get run by joining with all 5967 worker threads before resetting this flag */ 5968 TCW_SYNC_4(__kmp_init_common, FALSE); 5969 5970 KA_TRACE(10, ("__kmp_internal_end: all workers reaped\n")); 5971 KMP_MB(); 5972 5973 #if KMP_USE_MONITOR 5974 // See note above: One of the possible fixes for CQ138434 / CQ140126 5975 // 5976 // FIXME: push both code fragments down and CSE them? 5977 // push them into __kmp_cleanup() ? 5978 __kmp_acquire_bootstrap_lock(&__kmp_monitor_lock); 5979 if (TCR_4(__kmp_init_monitor)) { 5980 __kmp_reap_monitor(&__kmp_monitor); 5981 TCW_4(__kmp_init_monitor, 0); 5982 } 5983 __kmp_release_bootstrap_lock(&__kmp_monitor_lock); 5984 KA_TRACE(10, ("__kmp_internal_end: monitor reaped\n")); 5985 #endif 5986 } /* else !__kmp_global.t_active */ 5987 TCW_4(__kmp_init_gtid, FALSE); 5988 KMP_MB(); /* Flush all pending memory write invalidates. */ 5989 5990 __kmp_cleanup(); 5991 #if OMPT_SUPPORT 5992 ompt_fini(); 5993 #endif 5994 } 5995 5996 void __kmp_internal_end_library(int gtid_req) { 5997 /* if we have already cleaned up, don't try again, it wouldn't be pretty */ 5998 /* this shouldn't be a race condition because __kmp_internal_end() is the 5999 only place to clear __kmp_serial_init */ 6000 /* we'll check this later too, after we get the lock */ 6001 // 2009-09-06: We do not set g_abort without setting g_done. This check looks 6002 // redundaant, because the next check will work in any case. 6003 if (__kmp_global.g.g_abort) { 6004 KA_TRACE(11, ("__kmp_internal_end_library: abort, exiting\n")); 6005 /* TODO abort? */ 6006 return; 6007 } 6008 if (TCR_4(__kmp_global.g.g_done) || !__kmp_init_serial) { 6009 KA_TRACE(10, ("__kmp_internal_end_library: already finished\n")); 6010 return; 6011 } 6012 6013 KMP_MB(); /* Flush all pending memory write invalidates. */ 6014 6015 /* find out who we are and what we should do */ 6016 { 6017 int gtid = (gtid_req >= 0) ? gtid_req : __kmp_gtid_get_specific(); 6018 KA_TRACE( 6019 10, ("__kmp_internal_end_library: enter T#%d (%d)\n", gtid, gtid_req)); 6020 if (gtid == KMP_GTID_SHUTDOWN) { 6021 KA_TRACE(10, ("__kmp_internal_end_library: !__kmp_init_runtime, system " 6022 "already shutdown\n")); 6023 return; 6024 } else if (gtid == KMP_GTID_MONITOR) { 6025 KA_TRACE(10, ("__kmp_internal_end_library: monitor thread, gtid not " 6026 "registered, or system shutdown\n")); 6027 return; 6028 } else if (gtid == KMP_GTID_DNE) { 6029 KA_TRACE(10, ("__kmp_internal_end_library: gtid not registered or system " 6030 "shutdown\n")); 6031 /* we don't know who we are, but we may still shutdown the library */ 6032 } else if (KMP_UBER_GTID(gtid)) { 6033 /* unregister ourselves as an uber thread. gtid is no longer valid */ 6034 if (__kmp_root[gtid]->r.r_active) { 6035 __kmp_global.g.g_abort = -1; 6036 TCW_SYNC_4(__kmp_global.g.g_done, TRUE); 6037 KA_TRACE(10, 6038 ("__kmp_internal_end_library: root still active, abort T#%d\n", 6039 gtid)); 6040 return; 6041 } else { 6042 KA_TRACE( 6043 10, 6044 ("__kmp_internal_end_library: unregistering sibling T#%d\n", gtid)); 6045 __kmp_unregister_root_current_thread(gtid); 6046 } 6047 } else { 6048 /* worker threads may call this function through the atexit handler, if they 6049 * call exit() */ 6050 /* For now, skip the usual subsequent processing and just dump the debug buffer. 6051 TODO: do a thorough shutdown instead */ 6052 #ifdef DUMP_DEBUG_ON_EXIT 6053 if (__kmp_debug_buf) 6054 __kmp_dump_debug_buffer(); 6055 #endif 6056 return; 6057 } 6058 } 6059 /* synchronize the termination process */ 6060 __kmp_acquire_bootstrap_lock(&__kmp_initz_lock); 6061 6062 /* have we already finished */ 6063 if (__kmp_global.g.g_abort) { 6064 KA_TRACE(10, ("__kmp_internal_end_library: abort, exiting\n")); 6065 /* TODO abort? */ 6066 __kmp_release_bootstrap_lock(&__kmp_initz_lock); 6067 return; 6068 } 6069 if (TCR_4(__kmp_global.g.g_done) || !__kmp_init_serial) { 6070 __kmp_release_bootstrap_lock(&__kmp_initz_lock); 6071 return; 6072 } 6073 6074 /* We need this lock to enforce mutex between this reading of 6075 __kmp_threads_capacity and the writing by __kmp_register_root. 6076 Alternatively, we can use a counter of roots that is atomically updated by 6077 __kmp_get_global_thread_id_reg, __kmp_do_serial_initialize and 6078 __kmp_internal_end_*. */ 6079 __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock); 6080 6081 /* now we can safely conduct the actual termination */ 6082 __kmp_internal_end(); 6083 6084 __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock); 6085 __kmp_release_bootstrap_lock(&__kmp_initz_lock); 6086 6087 KA_TRACE(10, ("__kmp_internal_end_library: exit\n")); 6088 6089 #ifdef DUMP_DEBUG_ON_EXIT 6090 if (__kmp_debug_buf) 6091 __kmp_dump_debug_buffer(); 6092 #endif 6093 6094 #if KMP_OS_WINDOWS 6095 __kmp_close_console(); 6096 #endif 6097 6098 __kmp_fini_allocator(); 6099 6100 } // __kmp_internal_end_library 6101 6102 void __kmp_internal_end_thread(int gtid_req) { 6103 int i; 6104 6105 /* if we have already cleaned up, don't try again, it wouldn't be pretty */ 6106 /* this shouldn't be a race condition because __kmp_internal_end() is the 6107 * only place to clear __kmp_serial_init */ 6108 /* we'll check this later too, after we get the lock */ 6109 // 2009-09-06: We do not set g_abort without setting g_done. This check looks 6110 // redundant, because the next check will work in any case. 6111 if (__kmp_global.g.g_abort) { 6112 KA_TRACE(11, ("__kmp_internal_end_thread: abort, exiting\n")); 6113 /* TODO abort? */ 6114 return; 6115 } 6116 if (TCR_4(__kmp_global.g.g_done) || !__kmp_init_serial) { 6117 KA_TRACE(10, ("__kmp_internal_end_thread: already finished\n")); 6118 return; 6119 } 6120 6121 KMP_MB(); /* Flush all pending memory write invalidates. */ 6122 6123 /* find out who we are and what we should do */ 6124 { 6125 int gtid = (gtid_req >= 0) ? gtid_req : __kmp_gtid_get_specific(); 6126 KA_TRACE(10, 6127 ("__kmp_internal_end_thread: enter T#%d (%d)\n", gtid, gtid_req)); 6128 if (gtid == KMP_GTID_SHUTDOWN) { 6129 KA_TRACE(10, ("__kmp_internal_end_thread: !__kmp_init_runtime, system " 6130 "already shutdown\n")); 6131 return; 6132 } else if (gtid == KMP_GTID_MONITOR) { 6133 KA_TRACE(10, ("__kmp_internal_end_thread: monitor thread, gtid not " 6134 "registered, or system shutdown\n")); 6135 return; 6136 } else if (gtid == KMP_GTID_DNE) { 6137 KA_TRACE(10, ("__kmp_internal_end_thread: gtid not registered or system " 6138 "shutdown\n")); 6139 return; 6140 /* we don't know who we are */ 6141 } else if (KMP_UBER_GTID(gtid)) { 6142 /* unregister ourselves as an uber thread. gtid is no longer valid */ 6143 if (__kmp_root[gtid]->r.r_active) { 6144 __kmp_global.g.g_abort = -1; 6145 TCW_SYNC_4(__kmp_global.g.g_done, TRUE); 6146 KA_TRACE(10, 6147 ("__kmp_internal_end_thread: root still active, abort T#%d\n", 6148 gtid)); 6149 return; 6150 } else { 6151 KA_TRACE(10, ("__kmp_internal_end_thread: unregistering sibling T#%d\n", 6152 gtid)); 6153 __kmp_unregister_root_current_thread(gtid); 6154 } 6155 } else { 6156 /* just a worker thread, let's leave */ 6157 KA_TRACE(10, ("__kmp_internal_end_thread: worker thread T#%d\n", gtid)); 6158 6159 if (gtid >= 0) { 6160 __kmp_threads[gtid]->th.th_task_team = NULL; 6161 } 6162 6163 KA_TRACE(10, 6164 ("__kmp_internal_end_thread: worker thread done, exiting T#%d\n", 6165 gtid)); 6166 return; 6167 } 6168 } 6169 #if defined KMP_DYNAMIC_LIB 6170 // AC: lets not shutdown the Linux* OS dynamic library at the exit of uber 6171 // thread, because we will better shutdown later in the library destructor. 6172 // The reason of this change is performance problem when non-openmp thread in 6173 // a loop forks and joins many openmp threads. We can save a lot of time 6174 // keeping worker threads alive until the program shutdown. 6175 // OM: Removed Linux* OS restriction to fix the crash on OS X* (DPD200239966) 6176 // and Windows(DPD200287443) that occurs when using critical sections from 6177 // foreign threads. 6178 KA_TRACE(10, ("__kmp_internal_end_thread: exiting T#%d\n", gtid_req)); 6179 return; 6180 #endif 6181 /* synchronize the termination process */ 6182 __kmp_acquire_bootstrap_lock(&__kmp_initz_lock); 6183 6184 /* have we already finished */ 6185 if (__kmp_global.g.g_abort) { 6186 KA_TRACE(10, ("__kmp_internal_end_thread: abort, exiting\n")); 6187 /* TODO abort? */ 6188 __kmp_release_bootstrap_lock(&__kmp_initz_lock); 6189 return; 6190 } 6191 if (TCR_4(__kmp_global.g.g_done) || !__kmp_init_serial) { 6192 __kmp_release_bootstrap_lock(&__kmp_initz_lock); 6193 return; 6194 } 6195 6196 /* We need this lock to enforce mutex between this reading of 6197 __kmp_threads_capacity and the writing by __kmp_register_root. 6198 Alternatively, we can use a counter of roots that is atomically updated by 6199 __kmp_get_global_thread_id_reg, __kmp_do_serial_initialize and 6200 __kmp_internal_end_*. */ 6201 6202 /* should we finish the run-time? are all siblings done? */ 6203 __kmp_acquire_bootstrap_lock(&__kmp_forkjoin_lock); 6204 6205 for (i = 0; i < __kmp_threads_capacity; ++i) { 6206 if (KMP_UBER_GTID(i)) { 6207 KA_TRACE( 6208 10, 6209 ("__kmp_internal_end_thread: remaining sibling task: gtid==%d\n", i)); 6210 __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock); 6211 __kmp_release_bootstrap_lock(&__kmp_initz_lock); 6212 return; 6213 }; 6214 } 6215 6216 /* now we can safely conduct the actual termination */ 6217 6218 __kmp_internal_end(); 6219 6220 __kmp_release_bootstrap_lock(&__kmp_forkjoin_lock); 6221 __kmp_release_bootstrap_lock(&__kmp_initz_lock); 6222 6223 KA_TRACE(10, ("__kmp_internal_end_thread: exit T#%d\n", gtid_req)); 6224 6225 #ifdef DUMP_DEBUG_ON_EXIT 6226 if (__kmp_debug_buf) 6227 __kmp_dump_debug_buffer(); 6228 #endif 6229 } // __kmp_internal_end_thread 6230 6231 // ----------------------------------------------------------------------------- 6232 // Library registration stuff. 6233 6234 static long __kmp_registration_flag = 0; 6235 // Random value used to indicate library initialization. 6236 static char *__kmp_registration_str = NULL; 6237 // Value to be saved in env var __KMP_REGISTERED_LIB_<pid>. 6238 6239 static inline char *__kmp_reg_status_name() { 6240 /* On RHEL 3u5 if linked statically, getpid() returns different values in 6241 each thread. If registration and unregistration go in different threads 6242 (omp_misc_other_root_exit.cpp test case), the name of registered_lib_env 6243 env var can not be found, because the name will contain different pid. */ 6244 return __kmp_str_format("__KMP_REGISTERED_LIB_%d", (int)getpid()); 6245 } // __kmp_reg_status_get 6246 6247 void __kmp_register_library_startup(void) { 6248 6249 char *name = __kmp_reg_status_name(); // Name of the environment variable. 6250 int done = 0; 6251 union { 6252 double dtime; 6253 long ltime; 6254 } time; 6255 #if KMP_ARCH_X86 || KMP_ARCH_X86_64 6256 __kmp_initialize_system_tick(); 6257 #endif 6258 __kmp_read_system_time(&time.dtime); 6259 __kmp_registration_flag = 0xCAFE0000L | (time.ltime & 0x0000FFFFL); 6260 __kmp_registration_str = 6261 __kmp_str_format("%p-%lx-%s", &__kmp_registration_flag, 6262 __kmp_registration_flag, KMP_LIBRARY_FILE); 6263 6264 KA_TRACE(50, ("__kmp_register_library_startup: %s=\"%s\"\n", name, 6265 __kmp_registration_str)); 6266 6267 while (!done) { 6268 6269 char *value = NULL; // Actual value of the environment variable. 6270 6271 // Set environment variable, but do not overwrite if it is exist. 6272 __kmp_env_set(name, __kmp_registration_str, 0); 6273 // Check the variable is written. 6274 value = __kmp_env_get(name); 6275 if (value != NULL && strcmp(value, __kmp_registration_str) == 0) { 6276 6277 done = 1; // Ok, environment variable set successfully, exit the loop. 6278 6279 } else { 6280 6281 // Oops. Write failed. Another copy of OpenMP RTL is in memory. 6282 // Check whether it alive or dead. 6283 int neighbor = 0; // 0 -- unknown status, 1 -- alive, 2 -- dead. 6284 char *tail = value; 6285 char *flag_addr_str = NULL; 6286 char *flag_val_str = NULL; 6287 char const *file_name = NULL; 6288 __kmp_str_split(tail, '-', &flag_addr_str, &tail); 6289 __kmp_str_split(tail, '-', &flag_val_str, &tail); 6290 file_name = tail; 6291 if (tail != NULL) { 6292 long *flag_addr = 0; 6293 long flag_val = 0; 6294 KMP_SSCANF(flag_addr_str, "%p", &flag_addr); 6295 KMP_SSCANF(flag_val_str, "%lx", &flag_val); 6296 if (flag_addr != 0 && flag_val != 0 && strcmp(file_name, "") != 0) { 6297 // First, check whether environment-encoded address is mapped into 6298 // addr space. 6299 // If so, dereference it to see if it still has the right value. 6300 if (__kmp_is_address_mapped(flag_addr) && *flag_addr == flag_val) { 6301 neighbor = 1; 6302 } else { 6303 // If not, then we know the other copy of the library is no longer 6304 // running. 6305 neighbor = 2; 6306 }; // if 6307 }; // if 6308 }; // if 6309 switch (neighbor) { 6310 case 0: // Cannot parse environment variable -- neighbor status unknown. 6311 // Assume it is the incompatible format of future version of the 6312 // library. Assume the other library is alive. 6313 // WARN( ... ); // TODO: Issue a warning. 6314 file_name = "unknown library"; 6315 // Attention! Falling to the next case. That's intentional. 6316 case 1: { // Neighbor is alive. 6317 // Check it is allowed. 6318 char *duplicate_ok = __kmp_env_get("KMP_DUPLICATE_LIB_OK"); 6319 if (!__kmp_str_match_true(duplicate_ok)) { 6320 // That's not allowed. Issue fatal error. 6321 __kmp_fatal(KMP_MSG(DuplicateLibrary, KMP_LIBRARY_FILE, file_name), 6322 KMP_HNT(DuplicateLibrary), __kmp_msg_null); 6323 }; // if 6324 KMP_INTERNAL_FREE(duplicate_ok); 6325 __kmp_duplicate_library_ok = 1; 6326 done = 1; // Exit the loop. 6327 } break; 6328 case 2: { // Neighbor is dead. 6329 // Clear the variable and try to register library again. 6330 __kmp_env_unset(name); 6331 } break; 6332 default: { KMP_DEBUG_ASSERT(0); } break; 6333 }; // switch 6334 6335 }; // if 6336 KMP_INTERNAL_FREE((void *)value); 6337 6338 }; // while 6339 KMP_INTERNAL_FREE((void *)name); 6340 6341 } // func __kmp_register_library_startup 6342 6343 void __kmp_unregister_library(void) { 6344 6345 char *name = __kmp_reg_status_name(); 6346 char *value = __kmp_env_get(name); 6347 6348 KMP_DEBUG_ASSERT(__kmp_registration_flag != 0); 6349 KMP_DEBUG_ASSERT(__kmp_registration_str != NULL); 6350 if (value != NULL && strcmp(value, __kmp_registration_str) == 0) { 6351 // Ok, this is our variable. Delete it. 6352 __kmp_env_unset(name); 6353 }; // if 6354 6355 KMP_INTERNAL_FREE(__kmp_registration_str); 6356 KMP_INTERNAL_FREE(value); 6357 KMP_INTERNAL_FREE(name); 6358 6359 __kmp_registration_flag = 0; 6360 __kmp_registration_str = NULL; 6361 6362 } // __kmp_unregister_library 6363 6364 // End of Library registration stuff. 6365 // ----------------------------------------------------------------------------- 6366 6367 #if KMP_MIC_SUPPORTED 6368 6369 static void __kmp_check_mic_type() { 6370 kmp_cpuid_t cpuid_state = {0}; 6371 kmp_cpuid_t *cs_p = &cpuid_state; 6372 __kmp_x86_cpuid(1, 0, cs_p); 6373 // We don't support mic1 at the moment 6374 if ((cs_p->eax & 0xff0) == 0xB10) { 6375 __kmp_mic_type = mic2; 6376 } else if ((cs_p->eax & 0xf0ff0) == 0x50670) { 6377 __kmp_mic_type = mic3; 6378 } else { 6379 __kmp_mic_type = non_mic; 6380 } 6381 } 6382 6383 #endif /* KMP_MIC_SUPPORTED */ 6384 6385 static void __kmp_do_serial_initialize(void) { 6386 int i, gtid; 6387 int size; 6388 6389 KA_TRACE(10, ("__kmp_do_serial_initialize: enter\n")); 6390 6391 KMP_DEBUG_ASSERT(sizeof(kmp_int32) == 4); 6392 KMP_DEBUG_ASSERT(sizeof(kmp_uint32) == 4); 6393 KMP_DEBUG_ASSERT(sizeof(kmp_int64) == 8); 6394 KMP_DEBUG_ASSERT(sizeof(kmp_uint64) == 8); 6395 KMP_DEBUG_ASSERT(sizeof(kmp_intptr_t) == sizeof(void *)); 6396 6397 #if OMPT_SUPPORT 6398 ompt_pre_init(); 6399 #endif 6400 6401 __kmp_validate_locks(); 6402 6403 /* Initialize internal memory allocator */ 6404 __kmp_init_allocator(); 6405 6406 /* Register the library startup via an environment variable and check to see 6407 whether another copy of the library is already registered. */ 6408 6409 __kmp_register_library_startup(); 6410 6411 /* TODO reinitialization of library */ 6412 if (TCR_4(__kmp_global.g.g_done)) { 6413 KA_TRACE(10, ("__kmp_do_serial_initialize: reinitialization of library\n")); 6414 } 6415 6416 __kmp_global.g.g_abort = 0; 6417 TCW_SYNC_4(__kmp_global.g.g_done, FALSE); 6418 6419 /* initialize the locks */ 6420 #if KMP_USE_ADAPTIVE_LOCKS 6421 #if KMP_DEBUG_ADAPTIVE_LOCKS 6422 __kmp_init_speculative_stats(); 6423 #endif 6424 #endif 6425 #if KMP_STATS_ENABLED 6426 __kmp_stats_init(); 6427 #endif 6428 __kmp_init_lock(&__kmp_global_lock); 6429 __kmp_init_queuing_lock(&__kmp_dispatch_lock); 6430 __kmp_init_lock(&__kmp_debug_lock); 6431 __kmp_init_atomic_lock(&__kmp_atomic_lock); 6432 __kmp_init_atomic_lock(&__kmp_atomic_lock_1i); 6433 __kmp_init_atomic_lock(&__kmp_atomic_lock_2i); 6434 __kmp_init_atomic_lock(&__kmp_atomic_lock_4i); 6435 __kmp_init_atomic_lock(&__kmp_atomic_lock_4r); 6436 __kmp_init_atomic_lock(&__kmp_atomic_lock_8i); 6437 __kmp_init_atomic_lock(&__kmp_atomic_lock_8r); 6438 __kmp_init_atomic_lock(&__kmp_atomic_lock_8c); 6439 __kmp_init_atomic_lock(&__kmp_atomic_lock_10r); 6440 __kmp_init_atomic_lock(&__kmp_atomic_lock_16r); 6441 __kmp_init_atomic_lock(&__kmp_atomic_lock_16c); 6442 __kmp_init_atomic_lock(&__kmp_atomic_lock_20c); 6443 __kmp_init_atomic_lock(&__kmp_atomic_lock_32c); 6444 __kmp_init_bootstrap_lock(&__kmp_forkjoin_lock); 6445 __kmp_init_bootstrap_lock(&__kmp_exit_lock); 6446 #if KMP_USE_MONITOR 6447 __kmp_init_bootstrap_lock(&__kmp_monitor_lock); 6448 #endif 6449 __kmp_init_bootstrap_lock(&__kmp_tp_cached_lock); 6450 6451 /* conduct initialization and initial setup of configuration */ 6452 6453 __kmp_runtime_initialize(); 6454 6455 #if KMP_MIC_SUPPORTED 6456 __kmp_check_mic_type(); 6457 #endif 6458 6459 // Some global variable initialization moved here from kmp_env_initialize() 6460 #ifdef KMP_DEBUG 6461 kmp_diag = 0; 6462 #endif 6463 __kmp_abort_delay = 0; 6464 6465 // From __kmp_init_dflt_team_nth() 6466 /* assume the entire machine will be used */ 6467 __kmp_dflt_team_nth_ub = __kmp_xproc; 6468 if (__kmp_dflt_team_nth_ub < KMP_MIN_NTH) { 6469 __kmp_dflt_team_nth_ub = KMP_MIN_NTH; 6470 } 6471 if (__kmp_dflt_team_nth_ub > __kmp_sys_max_nth) { 6472 __kmp_dflt_team_nth_ub = __kmp_sys_max_nth; 6473 } 6474 __kmp_max_nth = __kmp_sys_max_nth; 6475 __kmp_cg_max_nth = __kmp_sys_max_nth; 6476 __kmp_teams_max_nth = __kmp_xproc; // set a "reasonable" default 6477 if (__kmp_teams_max_nth > __kmp_sys_max_nth) { 6478 __kmp_teams_max_nth = __kmp_sys_max_nth; 6479 } 6480 6481 // Three vars below moved here from __kmp_env_initialize() "KMP_BLOCKTIME" 6482 // part 6483 __kmp_dflt_blocktime = KMP_DEFAULT_BLOCKTIME; 6484 #if KMP_USE_MONITOR 6485 __kmp_monitor_wakeups = 6486 KMP_WAKEUPS_FROM_BLOCKTIME(__kmp_dflt_blocktime, __kmp_monitor_wakeups); 6487 __kmp_bt_intervals = 6488 KMP_INTERVALS_FROM_BLOCKTIME(__kmp_dflt_blocktime, __kmp_monitor_wakeups); 6489 #endif 6490 // From "KMP_LIBRARY" part of __kmp_env_initialize() 6491 __kmp_library = library_throughput; 6492 // From KMP_SCHEDULE initialization 6493 __kmp_static = kmp_sch_static_balanced; 6494 // AC: do not use analytical here, because it is non-monotonous 6495 //__kmp_guided = kmp_sch_guided_iterative_chunked; 6496 //__kmp_auto = kmp_sch_guided_analytical_chunked; // AC: it is the default, no 6497 // need to repeat assignment 6498 // Barrier initialization. Moved here from __kmp_env_initialize() Barrier branch 6499 // bit control and barrier method control parts 6500 #if KMP_FAST_REDUCTION_BARRIER 6501 #define kmp_reduction_barrier_gather_bb ((int)1) 6502 #define kmp_reduction_barrier_release_bb ((int)1) 6503 #define kmp_reduction_barrier_gather_pat bp_hyper_bar 6504 #define kmp_reduction_barrier_release_pat bp_hyper_bar 6505 #endif // KMP_FAST_REDUCTION_BARRIER 6506 for (i = bs_plain_barrier; i < bs_last_barrier; i++) { 6507 __kmp_barrier_gather_branch_bits[i] = __kmp_barrier_gather_bb_dflt; 6508 __kmp_barrier_release_branch_bits[i] = __kmp_barrier_release_bb_dflt; 6509 __kmp_barrier_gather_pattern[i] = __kmp_barrier_gather_pat_dflt; 6510 __kmp_barrier_release_pattern[i] = __kmp_barrier_release_pat_dflt; 6511 #if KMP_FAST_REDUCTION_BARRIER 6512 if (i == bs_reduction_barrier) { // tested and confirmed on ALTIX only ( 6513 // lin_64 ): hyper,1 6514 __kmp_barrier_gather_branch_bits[i] = kmp_reduction_barrier_gather_bb; 6515 __kmp_barrier_release_branch_bits[i] = kmp_reduction_barrier_release_bb; 6516 __kmp_barrier_gather_pattern[i] = kmp_reduction_barrier_gather_pat; 6517 __kmp_barrier_release_pattern[i] = kmp_reduction_barrier_release_pat; 6518 } 6519 #endif // KMP_FAST_REDUCTION_BARRIER 6520 } 6521 #if KMP_FAST_REDUCTION_BARRIER 6522 #undef kmp_reduction_barrier_release_pat 6523 #undef kmp_reduction_barrier_gather_pat 6524 #undef kmp_reduction_barrier_release_bb 6525 #undef kmp_reduction_barrier_gather_bb 6526 #endif // KMP_FAST_REDUCTION_BARRIER 6527 #if KMP_MIC_SUPPORTED 6528 if (__kmp_mic_type == mic2) { // KNC 6529 // AC: plane=3,2, forkjoin=2,1 are optimal for 240 threads on KNC 6530 __kmp_barrier_gather_branch_bits[bs_plain_barrier] = 3; // plain gather 6531 __kmp_barrier_release_branch_bits[bs_forkjoin_barrier] = 6532 1; // forkjoin release 6533 __kmp_barrier_gather_pattern[bs_forkjoin_barrier] = bp_hierarchical_bar; 6534 __kmp_barrier_release_pattern[bs_forkjoin_barrier] = bp_hierarchical_bar; 6535 } 6536 #if KMP_FAST_REDUCTION_BARRIER 6537 if (__kmp_mic_type == mic2) { // KNC 6538 __kmp_barrier_gather_pattern[bs_reduction_barrier] = bp_hierarchical_bar; 6539 __kmp_barrier_release_pattern[bs_reduction_barrier] = bp_hierarchical_bar; 6540 } 6541 #endif // KMP_FAST_REDUCTION_BARRIER 6542 #endif // KMP_MIC_SUPPORTED 6543 6544 // From KMP_CHECKS initialization 6545 #ifdef KMP_DEBUG 6546 __kmp_env_checks = TRUE; /* development versions have the extra checks */ 6547 #else 6548 __kmp_env_checks = FALSE; /* port versions do not have the extra checks */ 6549 #endif 6550 6551 // From "KMP_FOREIGN_THREADS_THREADPRIVATE" initialization 6552 __kmp_foreign_tp = TRUE; 6553 6554 __kmp_global.g.g_dynamic = FALSE; 6555 __kmp_global.g.g_dynamic_mode = dynamic_default; 6556 6557 __kmp_env_initialize(NULL); 6558 6559 // Print all messages in message catalog for testing purposes. 6560 #ifdef KMP_DEBUG 6561 char const *val = __kmp_env_get("KMP_DUMP_CATALOG"); 6562 if (__kmp_str_match_true(val)) { 6563 kmp_str_buf_t buffer; 6564 __kmp_str_buf_init(&buffer); 6565 __kmp_i18n_dump_catalog(&buffer); 6566 __kmp_printf("%s", buffer.str); 6567 __kmp_str_buf_free(&buffer); 6568 }; // if 6569 __kmp_env_free(&val); 6570 #endif 6571 6572 __kmp_threads_capacity = 6573 __kmp_initial_threads_capacity(__kmp_dflt_team_nth_ub); 6574 // Moved here from __kmp_env_initialize() "KMP_ALL_THREADPRIVATE" part 6575 __kmp_tp_capacity = __kmp_default_tp_capacity( 6576 __kmp_dflt_team_nth_ub, __kmp_max_nth, __kmp_allThreadsSpecified); 6577 6578 // If the library is shut down properly, both pools must be NULL. Just in 6579 // case, set them to NULL -- some memory may leak, but subsequent code will 6580 // work even if pools are not freed. 6581 KMP_DEBUG_ASSERT(__kmp_thread_pool == NULL); 6582 KMP_DEBUG_ASSERT(__kmp_thread_pool_insert_pt == NULL); 6583 KMP_DEBUG_ASSERT(__kmp_team_pool == NULL); 6584 __kmp_thread_pool = NULL; 6585 __kmp_thread_pool_insert_pt = NULL; 6586 __kmp_team_pool = NULL; 6587 6588 /* Allocate all of the variable sized records */ 6589 /* NOTE: __kmp_threads_capacity entries are allocated, but the arrays are 6590 * expandable */ 6591 /* Since allocation is cache-aligned, just add extra padding at the end */ 6592 size = 6593 (sizeof(kmp_info_t *) + sizeof(kmp_root_t *)) * __kmp_threads_capacity + 6594 CACHE_LINE; 6595 __kmp_threads = (kmp_info_t **)__kmp_allocate(size); 6596 __kmp_root = (kmp_root_t **)((char *)__kmp_threads + 6597 sizeof(kmp_info_t *) * __kmp_threads_capacity); 6598 6599 /* init thread counts */ 6600 KMP_DEBUG_ASSERT(__kmp_all_nth == 6601 0); // Asserts fail if the library is reinitializing and 6602 KMP_DEBUG_ASSERT(__kmp_nth == 0); // something was wrong in termination. 6603 __kmp_all_nth = 0; 6604 __kmp_nth = 0; 6605 6606 /* setup the uber master thread and hierarchy */ 6607 gtid = __kmp_register_root(TRUE); 6608 KA_TRACE(10, ("__kmp_do_serial_initialize T#%d\n", gtid)); 6609 KMP_ASSERT(KMP_UBER_GTID(gtid)); 6610 KMP_ASSERT(KMP_INITIAL_GTID(gtid)); 6611 6612 KMP_MB(); /* Flush all pending memory write invalidates. */ 6613 6614 __kmp_common_initialize(); 6615 6616 #if KMP_OS_UNIX 6617 /* invoke the child fork handler */ 6618 __kmp_register_atfork(); 6619 #endif 6620 6621 #if !defined KMP_DYNAMIC_LIB 6622 { 6623 /* Invoke the exit handler when the program finishes, only for static 6624 library. For dynamic library, we already have _fini and DllMain. */ 6625 int rc = atexit(__kmp_internal_end_atexit); 6626 if (rc != 0) { 6627 __kmp_fatal(KMP_MSG(FunctionError, "atexit()"), KMP_ERR(rc), 6628 __kmp_msg_null); 6629 }; // if 6630 } 6631 #endif 6632 6633 #if KMP_HANDLE_SIGNALS 6634 #if KMP_OS_UNIX 6635 /* NOTE: make sure that this is called before the user installs their own 6636 signal handlers so that the user handlers are called first. this way they 6637 can return false, not call our handler, avoid terminating the library, and 6638 continue execution where they left off. */ 6639 __kmp_install_signals(FALSE); 6640 #endif /* KMP_OS_UNIX */ 6641 #if KMP_OS_WINDOWS 6642 __kmp_install_signals(TRUE); 6643 #endif /* KMP_OS_WINDOWS */ 6644 #endif 6645 6646 /* we have finished the serial initialization */ 6647 __kmp_init_counter++; 6648 6649 __kmp_init_serial = TRUE; 6650 6651 if (__kmp_settings) { 6652 __kmp_env_print(); 6653 } 6654 6655 #if OMP_40_ENABLED 6656 if (__kmp_display_env || __kmp_display_env_verbose) { 6657 __kmp_env_print_2(); 6658 } 6659 #endif // OMP_40_ENABLED 6660 6661 #if OMPT_SUPPORT 6662 ompt_post_init(); 6663 #endif 6664 6665 KMP_MB(); 6666 6667 KA_TRACE(10, ("__kmp_do_serial_initialize: exit\n")); 6668 } 6669 6670 void __kmp_serial_initialize(void) { 6671 if (__kmp_init_serial) { 6672 return; 6673 } 6674 __kmp_acquire_bootstrap_lock(&__kmp_initz_lock); 6675 if (__kmp_init_serial) { 6676 __kmp_release_bootstrap_lock(&__kmp_initz_lock); 6677 return; 6678 } 6679 __kmp_do_serial_initialize(); 6680 __kmp_release_bootstrap_lock(&__kmp_initz_lock); 6681 } 6682 6683 static void __kmp_do_middle_initialize(void) { 6684 int i, j; 6685 int prev_dflt_team_nth; 6686 6687 if (!__kmp_init_serial) { 6688 __kmp_do_serial_initialize(); 6689 } 6690 6691 KA_TRACE(10, ("__kmp_middle_initialize: enter\n")); 6692 6693 // Save the previous value for the __kmp_dflt_team_nth so that 6694 // we can avoid some reinitialization if it hasn't changed. 6695 prev_dflt_team_nth = __kmp_dflt_team_nth; 6696 6697 #if KMP_AFFINITY_SUPPORTED 6698 // __kmp_affinity_initialize() will try to set __kmp_ncores to the 6699 // number of cores on the machine. 6700 __kmp_affinity_initialize(); 6701 6702 // Run through the __kmp_threads array and set the affinity mask 6703 // for each root thread that is currently registered with the RTL. 6704 for (i = 0; i < __kmp_threads_capacity; i++) { 6705 if (TCR_PTR(__kmp_threads[i]) != NULL) { 6706 __kmp_affinity_set_init_mask(i, TRUE); 6707 } 6708 } 6709 #endif /* KMP_AFFINITY_SUPPORTED */ 6710 6711 KMP_ASSERT(__kmp_xproc > 0); 6712 if (__kmp_avail_proc == 0) { 6713 __kmp_avail_proc = __kmp_xproc; 6714 } 6715 6716 // If there were empty places in num_threads list (OMP_NUM_THREADS=,,2,3), 6717 // correct them now 6718 j = 0; 6719 while ((j < __kmp_nested_nth.used) && !__kmp_nested_nth.nth[j]) { 6720 __kmp_nested_nth.nth[j] = __kmp_dflt_team_nth = __kmp_dflt_team_nth_ub = 6721 __kmp_avail_proc; 6722 j++; 6723 } 6724 6725 if (__kmp_dflt_team_nth == 0) { 6726 #ifdef KMP_DFLT_NTH_CORES 6727 // Default #threads = #cores 6728 __kmp_dflt_team_nth = __kmp_ncores; 6729 KA_TRACE(20, ("__kmp_middle_initialize: setting __kmp_dflt_team_nth = " 6730 "__kmp_ncores (%d)\n", 6731 __kmp_dflt_team_nth)); 6732 #else 6733 // Default #threads = #available OS procs 6734 __kmp_dflt_team_nth = __kmp_avail_proc; 6735 KA_TRACE(20, ("__kmp_middle_initialize: setting __kmp_dflt_team_nth = " 6736 "__kmp_avail_proc(%d)\n", 6737 __kmp_dflt_team_nth)); 6738 #endif /* KMP_DFLT_NTH_CORES */ 6739 } 6740 6741 if (__kmp_dflt_team_nth < KMP_MIN_NTH) { 6742 __kmp_dflt_team_nth = KMP_MIN_NTH; 6743 } 6744 if (__kmp_dflt_team_nth > __kmp_sys_max_nth) { 6745 __kmp_dflt_team_nth = __kmp_sys_max_nth; 6746 } 6747 6748 // There's no harm in continuing if the following check fails, 6749 // but it indicates an error in the previous logic. 6750 KMP_DEBUG_ASSERT(__kmp_dflt_team_nth <= __kmp_dflt_team_nth_ub); 6751 6752 if (__kmp_dflt_team_nth != prev_dflt_team_nth) { 6753 // Run through the __kmp_threads array and set the num threads icv for each 6754 // root thread that is currently registered with the RTL (which has not 6755 // already explicitly set its nthreads-var with a call to 6756 // omp_set_num_threads()). 6757 for (i = 0; i < __kmp_threads_capacity; i++) { 6758 kmp_info_t *thread = __kmp_threads[i]; 6759 if (thread == NULL) 6760 continue; 6761 if (thread->th.th_current_task->td_icvs.nproc != 0) 6762 continue; 6763 6764 set__nproc(__kmp_threads[i], __kmp_dflt_team_nth); 6765 } 6766 } 6767 KA_TRACE( 6768 20, 6769 ("__kmp_middle_initialize: final value for __kmp_dflt_team_nth = %d\n", 6770 __kmp_dflt_team_nth)); 6771 6772 #ifdef KMP_ADJUST_BLOCKTIME 6773 /* Adjust blocktime to zero if necessary now that __kmp_avail_proc is set */ 6774 if (!__kmp_env_blocktime && (__kmp_avail_proc > 0)) { 6775 KMP_DEBUG_ASSERT(__kmp_avail_proc > 0); 6776 if (__kmp_nth > __kmp_avail_proc) { 6777 __kmp_zero_bt = TRUE; 6778 } 6779 } 6780 #endif /* KMP_ADJUST_BLOCKTIME */ 6781 6782 /* we have finished middle initialization */ 6783 TCW_SYNC_4(__kmp_init_middle, TRUE); 6784 6785 KA_TRACE(10, ("__kmp_do_middle_initialize: exit\n")); 6786 } 6787 6788 void __kmp_middle_initialize(void) { 6789 if (__kmp_init_middle) { 6790 return; 6791 } 6792 __kmp_acquire_bootstrap_lock(&__kmp_initz_lock); 6793 if (__kmp_init_middle) { 6794 __kmp_release_bootstrap_lock(&__kmp_initz_lock); 6795 return; 6796 } 6797 __kmp_do_middle_initialize(); 6798 __kmp_release_bootstrap_lock(&__kmp_initz_lock); 6799 } 6800 6801 void __kmp_parallel_initialize(void) { 6802 int gtid = __kmp_entry_gtid(); // this might be a new root 6803 6804 /* synchronize parallel initialization (for sibling) */ 6805 if (TCR_4(__kmp_init_parallel)) 6806 return; 6807 __kmp_acquire_bootstrap_lock(&__kmp_initz_lock); 6808 if (TCR_4(__kmp_init_parallel)) { 6809 __kmp_release_bootstrap_lock(&__kmp_initz_lock); 6810 return; 6811 } 6812 6813 /* TODO reinitialization after we have already shut down */ 6814 if (TCR_4(__kmp_global.g.g_done)) { 6815 KA_TRACE( 6816 10, 6817 ("__kmp_parallel_initialize: attempt to init while shutting down\n")); 6818 __kmp_infinite_loop(); 6819 } 6820 6821 /* jc: The lock __kmp_initz_lock is already held, so calling 6822 __kmp_serial_initialize would cause a deadlock. So we call 6823 __kmp_do_serial_initialize directly. */ 6824 if (!__kmp_init_middle) { 6825 __kmp_do_middle_initialize(); 6826 } 6827 6828 /* begin initialization */ 6829 KA_TRACE(10, ("__kmp_parallel_initialize: enter\n")); 6830 KMP_ASSERT(KMP_UBER_GTID(gtid)); 6831 6832 #if KMP_ARCH_X86 || KMP_ARCH_X86_64 6833 // Save the FP control regs. 6834 // Worker threads will set theirs to these values at thread startup. 6835 __kmp_store_x87_fpu_control_word(&__kmp_init_x87_fpu_control_word); 6836 __kmp_store_mxcsr(&__kmp_init_mxcsr); 6837 __kmp_init_mxcsr &= KMP_X86_MXCSR_MASK; 6838 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */ 6839 6840 #if KMP_OS_UNIX 6841 #if KMP_HANDLE_SIGNALS 6842 /* must be after __kmp_serial_initialize */ 6843 __kmp_install_signals(TRUE); 6844 #endif 6845 #endif 6846 6847 __kmp_suspend_initialize(); 6848 6849 #if defined(USE_LOAD_BALANCE) 6850 if (__kmp_global.g.g_dynamic_mode == dynamic_default) { 6851 __kmp_global.g.g_dynamic_mode = dynamic_load_balance; 6852 } 6853 #else 6854 if (__kmp_global.g.g_dynamic_mode == dynamic_default) { 6855 __kmp_global.g.g_dynamic_mode = dynamic_thread_limit; 6856 } 6857 #endif 6858 6859 if (__kmp_version) { 6860 __kmp_print_version_2(); 6861 } 6862 6863 /* we have finished parallel initialization */ 6864 TCW_SYNC_4(__kmp_init_parallel, TRUE); 6865 6866 KMP_MB(); 6867 KA_TRACE(10, ("__kmp_parallel_initialize: exit\n")); 6868 6869 __kmp_release_bootstrap_lock(&__kmp_initz_lock); 6870 } 6871 6872 /* ------------------------------------------------------------------------ */ 6873 6874 void __kmp_run_before_invoked_task(int gtid, int tid, kmp_info_t *this_thr, 6875 kmp_team_t *team) { 6876 kmp_disp_t *dispatch; 6877 6878 KMP_MB(); 6879 6880 /* none of the threads have encountered any constructs, yet. */ 6881 this_thr->th.th_local.this_construct = 0; 6882 #if KMP_CACHE_MANAGE 6883 KMP_CACHE_PREFETCH(&this_thr->th.th_bar[bs_forkjoin_barrier].bb.b_arrived); 6884 #endif /* KMP_CACHE_MANAGE */ 6885 dispatch = (kmp_disp_t *)TCR_PTR(this_thr->th.th_dispatch); 6886 KMP_DEBUG_ASSERT(dispatch); 6887 KMP_DEBUG_ASSERT(team->t.t_dispatch); 6888 // KMP_DEBUG_ASSERT( this_thr->th.th_dispatch == &team->t.t_dispatch[ 6889 // this_thr->th.th_info.ds.ds_tid ] ); 6890 6891 dispatch->th_disp_index = 0; /* reset the dispatch buffer counter */ 6892 #if OMP_45_ENABLED 6893 dispatch->th_doacross_buf_idx = 6894 0; /* reset the doacross dispatch buffer counter */ 6895 #endif 6896 if (__kmp_env_consistency_check) 6897 __kmp_push_parallel(gtid, team->t.t_ident); 6898 6899 KMP_MB(); /* Flush all pending memory write invalidates. */ 6900 } 6901 6902 void __kmp_run_after_invoked_task(int gtid, int tid, kmp_info_t *this_thr, 6903 kmp_team_t *team) { 6904 if (__kmp_env_consistency_check) 6905 __kmp_pop_parallel(gtid, team->t.t_ident); 6906 6907 __kmp_finish_implicit_task(this_thr); 6908 } 6909 6910 int __kmp_invoke_task_func(int gtid) { 6911 int rc; 6912 int tid = __kmp_tid_from_gtid(gtid); 6913 kmp_info_t *this_thr = __kmp_threads[gtid]; 6914 kmp_team_t *team = this_thr->th.th_team; 6915 6916 __kmp_run_before_invoked_task(gtid, tid, this_thr, team); 6917 #if USE_ITT_BUILD 6918 if (__itt_stack_caller_create_ptr) { 6919 __kmp_itt_stack_callee_enter( 6920 (__itt_caller) 6921 team->t.t_stack_id); // inform ittnotify about entering user's code 6922 } 6923 #endif /* USE_ITT_BUILD */ 6924 #if INCLUDE_SSC_MARKS 6925 SSC_MARK_INVOKING(); 6926 #endif 6927 6928 #if OMPT_SUPPORT 6929 void *dummy; 6930 void **exit_runtime_p; 6931 ompt_task_id_t my_task_id; 6932 ompt_parallel_id_t my_parallel_id; 6933 6934 if (ompt_enabled) { 6935 exit_runtime_p = &(team->t.t_implicit_task_taskdata[tid] 6936 .ompt_task_info.frame.exit_runtime_frame); 6937 } else { 6938 exit_runtime_p = &dummy; 6939 } 6940 6941 #if OMPT_TRACE 6942 my_task_id = team->t.t_implicit_task_taskdata[tid].ompt_task_info.task_id; 6943 my_parallel_id = team->t.ompt_team_info.parallel_id; 6944 if (ompt_enabled && 6945 ompt_callbacks.ompt_callback(ompt_event_implicit_task_begin)) { 6946 ompt_callbacks.ompt_callback(ompt_event_implicit_task_begin)(my_parallel_id, 6947 my_task_id); 6948 } 6949 #endif 6950 #endif 6951 6952 { 6953 KMP_TIME_PARTITIONED_BLOCK(OMP_parallel); 6954 KMP_SET_THREAD_STATE_BLOCK(IMPLICIT_TASK); 6955 rc = 6956 __kmp_invoke_microtask((microtask_t)TCR_SYNC_PTR(team->t.t_pkfn), gtid, 6957 tid, (int)team->t.t_argc, (void **)team->t.t_argv 6958 #if OMPT_SUPPORT 6959 , 6960 exit_runtime_p 6961 #endif 6962 ); 6963 #if OMPT_SUPPORT 6964 *exit_runtime_p = NULL; 6965 #endif 6966 } 6967 6968 #if USE_ITT_BUILD 6969 if (__itt_stack_caller_create_ptr) { 6970 __kmp_itt_stack_callee_leave( 6971 (__itt_caller) 6972 team->t.t_stack_id); // inform ittnotify about leaving user's code 6973 } 6974 #endif /* USE_ITT_BUILD */ 6975 __kmp_run_after_invoked_task(gtid, tid, this_thr, team); 6976 6977 return rc; 6978 } 6979 6980 #if OMP_40_ENABLED 6981 void __kmp_teams_master(int gtid) { 6982 // This routine is called by all master threads in teams construct 6983 kmp_info_t *thr = __kmp_threads[gtid]; 6984 kmp_team_t *team = thr->th.th_team; 6985 ident_t *loc = team->t.t_ident; 6986 thr->th.th_set_nproc = thr->th.th_teams_size.nth; 6987 KMP_DEBUG_ASSERT(thr->th.th_teams_microtask); 6988 KMP_DEBUG_ASSERT(thr->th.th_set_nproc); 6989 KA_TRACE(20, ("__kmp_teams_master: T#%d, Tid %d, microtask %p\n", gtid, 6990 __kmp_tid_from_gtid(gtid), thr->th.th_teams_microtask)); 6991 // Launch league of teams now, but not let workers execute 6992 // (they hang on fork barrier until next parallel) 6993 #if INCLUDE_SSC_MARKS 6994 SSC_MARK_FORKING(); 6995 #endif 6996 __kmp_fork_call(loc, gtid, fork_context_intel, team->t.t_argc, 6997 #if OMPT_SUPPORT 6998 (void *)thr->th.th_teams_microtask, // "unwrapped" task 6999 #endif 7000 (microtask_t)thr->th.th_teams_microtask, // "wrapped" task 7001 VOLATILE_CAST(launch_t) __kmp_invoke_task_func, NULL); 7002 #if INCLUDE_SSC_MARKS 7003 SSC_MARK_JOINING(); 7004 #endif 7005 7006 // AC: last parameter "1" eliminates join barrier which won't work because 7007 // worker threads are in a fork barrier waiting for more parallel regions 7008 __kmp_join_call(loc, gtid 7009 #if OMPT_SUPPORT 7010 , 7011 fork_context_intel 7012 #endif 7013 , 7014 1); 7015 } 7016 7017 int __kmp_invoke_teams_master(int gtid) { 7018 kmp_info_t *this_thr = __kmp_threads[gtid]; 7019 kmp_team_t *team = this_thr->th.th_team; 7020 #if KMP_DEBUG 7021 if (!__kmp_threads[gtid]->th.th_team->t.t_serialized) 7022 KMP_DEBUG_ASSERT((void *)__kmp_threads[gtid]->th.th_team->t.t_pkfn == 7023 (void *)__kmp_teams_master); 7024 #endif 7025 __kmp_run_before_invoked_task(gtid, 0, this_thr, team); 7026 __kmp_teams_master(gtid); 7027 __kmp_run_after_invoked_task(gtid, 0, this_thr, team); 7028 return 1; 7029 } 7030 #endif /* OMP_40_ENABLED */ 7031 7032 /* this sets the requested number of threads for the next parallel region 7033 encountered by this team. since this should be enclosed in the forkjoin 7034 critical section it should avoid race conditions with assymmetrical nested 7035 parallelism */ 7036 7037 void __kmp_push_num_threads(ident_t *id, int gtid, int num_threads) { 7038 kmp_info_t *thr = __kmp_threads[gtid]; 7039 7040 if (num_threads > 0) 7041 thr->th.th_set_nproc = num_threads; 7042 } 7043 7044 #if OMP_40_ENABLED 7045 7046 /* this sets the requested number of teams for the teams region and/or 7047 the number of threads for the next parallel region encountered */ 7048 void __kmp_push_num_teams(ident_t *id, int gtid, int num_teams, 7049 int num_threads) { 7050 kmp_info_t *thr = __kmp_threads[gtid]; 7051 KMP_DEBUG_ASSERT(num_teams >= 0); 7052 KMP_DEBUG_ASSERT(num_threads >= 0); 7053 7054 if (num_teams == 0) 7055 num_teams = 1; // default number of teams is 1. 7056 if (num_teams > __kmp_teams_max_nth) { // if too many teams requested? 7057 if (!__kmp_reserve_warn) { 7058 __kmp_reserve_warn = 1; 7059 __kmp_msg(kmp_ms_warning, 7060 KMP_MSG(CantFormThrTeam, num_teams, __kmp_teams_max_nth), 7061 KMP_HNT(Unset_ALL_THREADS), __kmp_msg_null); 7062 } 7063 num_teams = __kmp_teams_max_nth; 7064 } 7065 // Set number of teams (number of threads in the outer "parallel" of the 7066 // teams) 7067 thr->th.th_set_nproc = thr->th.th_teams_size.nteams = num_teams; 7068 7069 // Remember the number of threads for inner parallel regions 7070 if (num_threads == 0) { 7071 if (!TCR_4(__kmp_init_middle)) 7072 __kmp_middle_initialize(); // get __kmp_avail_proc calculated 7073 num_threads = __kmp_avail_proc / num_teams; 7074 if (num_teams * num_threads > __kmp_teams_max_nth) { 7075 // adjust num_threads w/o warning as it is not user setting 7076 num_threads = __kmp_teams_max_nth / num_teams; 7077 } 7078 } else { 7079 if (num_teams * num_threads > __kmp_teams_max_nth) { 7080 int new_threads = __kmp_teams_max_nth / num_teams; 7081 if (!__kmp_reserve_warn) { // user asked for too many threads 7082 __kmp_reserve_warn = 1; // that conflicts with KMP_TEAMS_THREAD_LIMIT 7083 __kmp_msg(kmp_ms_warning, 7084 KMP_MSG(CantFormThrTeam, num_threads, new_threads), 7085 KMP_HNT(Unset_ALL_THREADS), __kmp_msg_null); 7086 } 7087 num_threads = new_threads; 7088 } 7089 } 7090 thr->th.th_teams_size.nth = num_threads; 7091 } 7092 7093 // Set the proc_bind var to use in the following parallel region. 7094 void __kmp_push_proc_bind(ident_t *id, int gtid, kmp_proc_bind_t proc_bind) { 7095 kmp_info_t *thr = __kmp_threads[gtid]; 7096 thr->th.th_set_proc_bind = proc_bind; 7097 } 7098 7099 #endif /* OMP_40_ENABLED */ 7100 7101 /* Launch the worker threads into the microtask. */ 7102 7103 void __kmp_internal_fork(ident_t *id, int gtid, kmp_team_t *team) { 7104 kmp_info_t *this_thr = __kmp_threads[gtid]; 7105 7106 #ifdef KMP_DEBUG 7107 int f; 7108 #endif /* KMP_DEBUG */ 7109 7110 KMP_DEBUG_ASSERT(team); 7111 KMP_DEBUG_ASSERT(this_thr->th.th_team == team); 7112 KMP_ASSERT(KMP_MASTER_GTID(gtid)); 7113 KMP_MB(); /* Flush all pending memory write invalidates. */ 7114 7115 team->t.t_construct = 0; /* no single directives seen yet */ 7116 team->t.t_ordered.dt.t_value = 7117 0; /* thread 0 enters the ordered section first */ 7118 7119 /* Reset the identifiers on the dispatch buffer */ 7120 KMP_DEBUG_ASSERT(team->t.t_disp_buffer); 7121 if (team->t.t_max_nproc > 1) { 7122 int i; 7123 for (i = 0; i < __kmp_dispatch_num_buffers; ++i) { 7124 team->t.t_disp_buffer[i].buffer_index = i; 7125 #if OMP_45_ENABLED 7126 team->t.t_disp_buffer[i].doacross_buf_idx = i; 7127 #endif 7128 } 7129 } else { 7130 team->t.t_disp_buffer[0].buffer_index = 0; 7131 #if OMP_45_ENABLED 7132 team->t.t_disp_buffer[0].doacross_buf_idx = 0; 7133 #endif 7134 } 7135 7136 KMP_MB(); /* Flush all pending memory write invalidates. */ 7137 KMP_ASSERT(this_thr->th.th_team == team); 7138 7139 #ifdef KMP_DEBUG 7140 for (f = 0; f < team->t.t_nproc; f++) { 7141 KMP_DEBUG_ASSERT(team->t.t_threads[f] && 7142 team->t.t_threads[f]->th.th_team_nproc == team->t.t_nproc); 7143 } 7144 #endif /* KMP_DEBUG */ 7145 7146 /* release the worker threads so they may begin working */ 7147 __kmp_fork_barrier(gtid, 0); 7148 } 7149 7150 void __kmp_internal_join(ident_t *id, int gtid, kmp_team_t *team) { 7151 kmp_info_t *this_thr = __kmp_threads[gtid]; 7152 7153 KMP_DEBUG_ASSERT(team); 7154 KMP_DEBUG_ASSERT(this_thr->th.th_team == team); 7155 KMP_ASSERT(KMP_MASTER_GTID(gtid)); 7156 KMP_MB(); /* Flush all pending memory write invalidates. */ 7157 7158 /* Join barrier after fork */ 7159 7160 #ifdef KMP_DEBUG 7161 if (__kmp_threads[gtid] && 7162 __kmp_threads[gtid]->th.th_team_nproc != team->t.t_nproc) { 7163 __kmp_printf("GTID: %d, __kmp_threads[%d]=%p\n", gtid, gtid, 7164 __kmp_threads[gtid]); 7165 __kmp_printf("__kmp_threads[%d]->th.th_team_nproc=%d, TEAM: %p, " 7166 "team->t.t_nproc=%d\n", 7167 gtid, __kmp_threads[gtid]->th.th_team_nproc, team, 7168 team->t.t_nproc); 7169 __kmp_print_structure(); 7170 } 7171 KMP_DEBUG_ASSERT(__kmp_threads[gtid] && 7172 __kmp_threads[gtid]->th.th_team_nproc == team->t.t_nproc); 7173 #endif /* KMP_DEBUG */ 7174 7175 __kmp_join_barrier(gtid); /* wait for everyone */ 7176 7177 KMP_MB(); /* Flush all pending memory write invalidates. */ 7178 KMP_ASSERT(this_thr->th.th_team == team); 7179 } 7180 7181 /* ------------------------------------------------------------------------ */ 7182 7183 #ifdef USE_LOAD_BALANCE 7184 7185 // Return the worker threads actively spinning in the hot team, if we 7186 // are at the outermost level of parallelism. Otherwise, return 0. 7187 static int __kmp_active_hot_team_nproc(kmp_root_t *root) { 7188 int i; 7189 int retval; 7190 kmp_team_t *hot_team; 7191 7192 if (root->r.r_active) { 7193 return 0; 7194 } 7195 hot_team = root->r.r_hot_team; 7196 if (__kmp_dflt_blocktime == KMP_MAX_BLOCKTIME) { 7197 return hot_team->t.t_nproc - 1; // Don't count master thread 7198 } 7199 7200 // Skip the master thread - it is accounted for elsewhere. 7201 retval = 0; 7202 for (i = 1; i < hot_team->t.t_nproc; i++) { 7203 if (hot_team->t.t_threads[i]->th.th_active) { 7204 retval++; 7205 } 7206 } 7207 return retval; 7208 } 7209 7210 // Perform an automatic adjustment to the number of 7211 // threads used by the next parallel region. 7212 static int __kmp_load_balance_nproc(kmp_root_t *root, int set_nproc) { 7213 int retval; 7214 int pool_active; 7215 int hot_team_active; 7216 int team_curr_active; 7217 int system_active; 7218 7219 KB_TRACE(20, ("__kmp_load_balance_nproc: called root:%p set_nproc:%d\n", root, 7220 set_nproc)); 7221 KMP_DEBUG_ASSERT(root); 7222 KMP_DEBUG_ASSERT(root->r.r_root_team->t.t_threads[0] 7223 ->th.th_current_task->td_icvs.dynamic == TRUE); 7224 KMP_DEBUG_ASSERT(set_nproc > 1); 7225 7226 if (set_nproc == 1) { 7227 KB_TRACE(20, ("__kmp_load_balance_nproc: serial execution.\n")); 7228 return 1; 7229 } 7230 7231 // Threads that are active in the thread pool, active in the hot team for this 7232 // particular root (if we are at the outer par level), and the currently 7233 // executing thread (to become the master) are available to add to the new 7234 // team, but are currently contributing to the system load, and must be 7235 // accounted for. 7236 pool_active = TCR_4(__kmp_thread_pool_active_nth); 7237 hot_team_active = __kmp_active_hot_team_nproc(root); 7238 team_curr_active = pool_active + hot_team_active + 1; 7239 7240 // Check the system load. 7241 system_active = __kmp_get_load_balance(__kmp_avail_proc + team_curr_active); 7242 KB_TRACE(30, ("__kmp_load_balance_nproc: system active = %d pool active = %d " 7243 "hot team active = %d\n", 7244 system_active, pool_active, hot_team_active)); 7245 7246 if (system_active < 0) { 7247 // There was an error reading the necessary info from /proc, so use the 7248 // thread limit algorithm instead. Once we set __kmp_global.g.g_dynamic_mode 7249 // = dynamic_thread_limit, we shouldn't wind up getting back here. 7250 __kmp_global.g.g_dynamic_mode = dynamic_thread_limit; 7251 KMP_WARNING(CantLoadBalUsing, "KMP_DYNAMIC_MODE=thread limit"); 7252 7253 // Make this call behave like the thread limit algorithm. 7254 retval = __kmp_avail_proc - __kmp_nth + 7255 (root->r.r_active ? 1 : root->r.r_hot_team->t.t_nproc); 7256 if (retval > set_nproc) { 7257 retval = set_nproc; 7258 } 7259 if (retval < KMP_MIN_NTH) { 7260 retval = KMP_MIN_NTH; 7261 } 7262 7263 KB_TRACE(20, ("__kmp_load_balance_nproc: thread limit exit. retval:%d\n", 7264 retval)); 7265 return retval; 7266 } 7267 7268 // There is a slight delay in the load balance algorithm in detecting new 7269 // running procs. The real system load at this instant should be at least as 7270 // large as the #active omp thread that are available to add to the team. 7271 if (system_active < team_curr_active) { 7272 system_active = team_curr_active; 7273 } 7274 retval = __kmp_avail_proc - system_active + team_curr_active; 7275 if (retval > set_nproc) { 7276 retval = set_nproc; 7277 } 7278 if (retval < KMP_MIN_NTH) { 7279 retval = KMP_MIN_NTH; 7280 } 7281 7282 KB_TRACE(20, ("__kmp_load_balance_nproc: exit. retval:%d\n", retval)); 7283 return retval; 7284 } // __kmp_load_balance_nproc() 7285 7286 #endif /* USE_LOAD_BALANCE */ 7287 7288 /* ------------------------------------------------------------------------ */ 7289 7290 /* NOTE: this is called with the __kmp_init_lock held */ 7291 void __kmp_cleanup(void) { 7292 int f; 7293 7294 KA_TRACE(10, ("__kmp_cleanup: enter\n")); 7295 7296 if (TCR_4(__kmp_init_parallel)) { 7297 #if KMP_HANDLE_SIGNALS 7298 __kmp_remove_signals(); 7299 #endif 7300 TCW_4(__kmp_init_parallel, FALSE); 7301 } 7302 7303 if (TCR_4(__kmp_init_middle)) { 7304 #if KMP_AFFINITY_SUPPORTED 7305 __kmp_affinity_uninitialize(); 7306 #endif /* KMP_AFFINITY_SUPPORTED */ 7307 __kmp_cleanup_hierarchy(); 7308 TCW_4(__kmp_init_middle, FALSE); 7309 } 7310 7311 KA_TRACE(10, ("__kmp_cleanup: go serial cleanup\n")); 7312 7313 if (__kmp_init_serial) { 7314 __kmp_runtime_destroy(); 7315 __kmp_init_serial = FALSE; 7316 } 7317 7318 for (f = 0; f < __kmp_threads_capacity; f++) { 7319 if (__kmp_root[f] != NULL) { 7320 __kmp_free(__kmp_root[f]); 7321 __kmp_root[f] = NULL; 7322 } 7323 } 7324 __kmp_free(__kmp_threads); 7325 // __kmp_threads and __kmp_root were allocated at once, as single block, so 7326 // there is no need in freeing __kmp_root. 7327 __kmp_threads = NULL; 7328 __kmp_root = NULL; 7329 __kmp_threads_capacity = 0; 7330 7331 #if KMP_USE_DYNAMIC_LOCK 7332 __kmp_cleanup_indirect_user_locks(); 7333 #else 7334 __kmp_cleanup_user_locks(); 7335 #endif 7336 7337 #if KMP_AFFINITY_SUPPORTED 7338 KMP_INTERNAL_FREE(CCAST(char *, __kmp_cpuinfo_file)); 7339 __kmp_cpuinfo_file = NULL; 7340 #endif /* KMP_AFFINITY_SUPPORTED */ 7341 7342 #if KMP_USE_ADAPTIVE_LOCKS 7343 #if KMP_DEBUG_ADAPTIVE_LOCKS 7344 __kmp_print_speculative_stats(); 7345 #endif 7346 #endif 7347 KMP_INTERNAL_FREE(__kmp_nested_nth.nth); 7348 __kmp_nested_nth.nth = NULL; 7349 __kmp_nested_nth.size = 0; 7350 __kmp_nested_nth.used = 0; 7351 KMP_INTERNAL_FREE(__kmp_nested_proc_bind.bind_types); 7352 __kmp_nested_proc_bind.bind_types = NULL; 7353 __kmp_nested_proc_bind.size = 0; 7354 __kmp_nested_proc_bind.used = 0; 7355 7356 __kmp_i18n_catclose(); 7357 7358 #if KMP_STATS_ENABLED 7359 __kmp_stats_fini(); 7360 #endif 7361 7362 KA_TRACE(10, ("__kmp_cleanup: exit\n")); 7363 } 7364 7365 /* ------------------------------------------------------------------------ */ 7366 7367 int __kmp_ignore_mppbeg(void) { 7368 char *env; 7369 7370 if ((env = getenv("KMP_IGNORE_MPPBEG")) != NULL) { 7371 if (__kmp_str_match_false(env)) 7372 return FALSE; 7373 } 7374 // By default __kmpc_begin() is no-op. 7375 return TRUE; 7376 } 7377 7378 int __kmp_ignore_mppend(void) { 7379 char *env; 7380 7381 if ((env = getenv("KMP_IGNORE_MPPEND")) != NULL) { 7382 if (__kmp_str_match_false(env)) 7383 return FALSE; 7384 } 7385 // By default __kmpc_end() is no-op. 7386 return TRUE; 7387 } 7388 7389 void __kmp_internal_begin(void) { 7390 int gtid; 7391 kmp_root_t *root; 7392 7393 /* this is a very important step as it will register new sibling threads 7394 and assign these new uber threads a new gtid */ 7395 gtid = __kmp_entry_gtid(); 7396 root = __kmp_threads[gtid]->th.th_root; 7397 KMP_ASSERT(KMP_UBER_GTID(gtid)); 7398 7399 if (root->r.r_begin) 7400 return; 7401 __kmp_acquire_lock(&root->r.r_begin_lock, gtid); 7402 if (root->r.r_begin) { 7403 __kmp_release_lock(&root->r.r_begin_lock, gtid); 7404 return; 7405 } 7406 7407 root->r.r_begin = TRUE; 7408 7409 __kmp_release_lock(&root->r.r_begin_lock, gtid); 7410 } 7411 7412 /* ------------------------------------------------------------------------ */ 7413 7414 void __kmp_user_set_library(enum library_type arg) { 7415 int gtid; 7416 kmp_root_t *root; 7417 kmp_info_t *thread; 7418 7419 /* first, make sure we are initialized so we can get our gtid */ 7420 7421 gtid = __kmp_entry_gtid(); 7422 thread = __kmp_threads[gtid]; 7423 7424 root = thread->th.th_root; 7425 7426 KA_TRACE(20, ("__kmp_user_set_library: enter T#%d, arg: %d, %d\n", gtid, arg, 7427 library_serial)); 7428 if (root->r.r_in_parallel) { /* Must be called in serial section of top-level 7429 thread */ 7430 KMP_WARNING(SetLibraryIncorrectCall); 7431 return; 7432 } 7433 7434 switch (arg) { 7435 case library_serial: 7436 thread->th.th_set_nproc = 0; 7437 set__nproc(thread, 1); 7438 break; 7439 case library_turnaround: 7440 thread->th.th_set_nproc = 0; 7441 set__nproc(thread, __kmp_dflt_team_nth ? __kmp_dflt_team_nth 7442 : __kmp_dflt_team_nth_ub); 7443 break; 7444 case library_throughput: 7445 thread->th.th_set_nproc = 0; 7446 set__nproc(thread, __kmp_dflt_team_nth ? __kmp_dflt_team_nth 7447 : __kmp_dflt_team_nth_ub); 7448 break; 7449 default: 7450 KMP_FATAL(UnknownLibraryType, arg); 7451 } 7452 7453 __kmp_aux_set_library(arg); 7454 } 7455 7456 void __kmp_aux_set_stacksize(size_t arg) { 7457 if (!__kmp_init_serial) 7458 __kmp_serial_initialize(); 7459 7460 #if KMP_OS_DARWIN 7461 if (arg & (0x1000 - 1)) { 7462 arg &= ~(0x1000 - 1); 7463 if (arg + 0x1000) /* check for overflow if we round up */ 7464 arg += 0x1000; 7465 } 7466 #endif 7467 __kmp_acquire_bootstrap_lock(&__kmp_initz_lock); 7468 7469 /* only change the default stacksize before the first parallel region */ 7470 if (!TCR_4(__kmp_init_parallel)) { 7471 size_t value = arg; /* argument is in bytes */ 7472 7473 if (value < __kmp_sys_min_stksize) 7474 value = __kmp_sys_min_stksize; 7475 else if (value > KMP_MAX_STKSIZE) 7476 value = KMP_MAX_STKSIZE; 7477 7478 __kmp_stksize = value; 7479 7480 __kmp_env_stksize = TRUE; /* was KMP_STACKSIZE specified? */ 7481 } 7482 7483 __kmp_release_bootstrap_lock(&__kmp_initz_lock); 7484 } 7485 7486 /* set the behaviour of the runtime library */ 7487 /* TODO this can cause some odd behaviour with sibling parallelism... */ 7488 void __kmp_aux_set_library(enum library_type arg) { 7489 __kmp_library = arg; 7490 7491 switch (__kmp_library) { 7492 case library_serial: { 7493 KMP_INFORM(LibraryIsSerial); 7494 (void)__kmp_change_library(TRUE); 7495 } break; 7496 case library_turnaround: 7497 (void)__kmp_change_library(TRUE); 7498 break; 7499 case library_throughput: 7500 (void)__kmp_change_library(FALSE); 7501 break; 7502 default: 7503 KMP_FATAL(UnknownLibraryType, arg); 7504 } 7505 } 7506 7507 /* ------------------------------------------------------------------------ */ 7508 7509 void __kmp_aux_set_blocktime(int arg, kmp_info_t *thread, int tid) { 7510 int blocktime = arg; /* argument is in milliseconds */ 7511 #if KMP_USE_MONITOR 7512 int bt_intervals; 7513 #endif 7514 int bt_set; 7515 7516 __kmp_save_internal_controls(thread); 7517 7518 /* Normalize and set blocktime for the teams */ 7519 if (blocktime < KMP_MIN_BLOCKTIME) 7520 blocktime = KMP_MIN_BLOCKTIME; 7521 else if (blocktime > KMP_MAX_BLOCKTIME) 7522 blocktime = KMP_MAX_BLOCKTIME; 7523 7524 set__blocktime_team(thread->th.th_team, tid, blocktime); 7525 set__blocktime_team(thread->th.th_serial_team, 0, blocktime); 7526 7527 #if KMP_USE_MONITOR 7528 /* Calculate and set blocktime intervals for the teams */ 7529 bt_intervals = KMP_INTERVALS_FROM_BLOCKTIME(blocktime, __kmp_monitor_wakeups); 7530 7531 set__bt_intervals_team(thread->th.th_team, tid, bt_intervals); 7532 set__bt_intervals_team(thread->th.th_serial_team, 0, bt_intervals); 7533 #endif 7534 7535 /* Set whether blocktime has been set to "TRUE" */ 7536 bt_set = TRUE; 7537 7538 set__bt_set_team(thread->th.th_team, tid, bt_set); 7539 set__bt_set_team(thread->th.th_serial_team, 0, bt_set); 7540 #if KMP_USE_MONITOR 7541 KF_TRACE(10, ("kmp_set_blocktime: T#%d(%d:%d), blocktime=%d, " 7542 "bt_intervals=%d, monitor_updates=%d\n", 7543 __kmp_gtid_from_tid(tid, thread->th.th_team), 7544 thread->th.th_team->t.t_id, tid, blocktime, bt_intervals, 7545 __kmp_monitor_wakeups)); 7546 #else 7547 KF_TRACE(10, ("kmp_set_blocktime: T#%d(%d:%d), blocktime=%d\n", 7548 __kmp_gtid_from_tid(tid, thread->th.th_team), 7549 thread->th.th_team->t.t_id, tid, blocktime)); 7550 #endif 7551 } 7552 7553 void __kmp_aux_set_defaults(char const *str, int len) { 7554 if (!__kmp_init_serial) { 7555 __kmp_serial_initialize(); 7556 }; 7557 __kmp_env_initialize(str); 7558 7559 if (__kmp_settings 7560 #if OMP_40_ENABLED 7561 || __kmp_display_env || __kmp_display_env_verbose 7562 #endif // OMP_40_ENABLED 7563 ) { 7564 __kmp_env_print(); 7565 } 7566 } // __kmp_aux_set_defaults 7567 7568 /* ------------------------------------------------------------------------ */ 7569 /* internal fast reduction routines */ 7570 7571 PACKED_REDUCTION_METHOD_T 7572 __kmp_determine_reduction_method( 7573 ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size, 7574 void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data), 7575 kmp_critical_name *lck) { 7576 7577 // Default reduction method: critical construct ( lck != NULL, like in current 7578 // PAROPT ) 7579 // If ( reduce_data!=NULL && reduce_func!=NULL ): the tree-reduction method 7580 // can be selected by RTL 7581 // If loc->flags contains KMP_IDENT_ATOMIC_REDUCE, the atomic reduce method 7582 // can be selected by RTL 7583 // Finally, it's up to OpenMP RTL to make a decision on which method to select 7584 // among generated by PAROPT. 7585 7586 PACKED_REDUCTION_METHOD_T retval; 7587 7588 int team_size; 7589 7590 KMP_DEBUG_ASSERT(loc); // it would be nice to test ( loc != 0 ) 7591 KMP_DEBUG_ASSERT(lck); // it would be nice to test ( lck != 0 ) 7592 7593 #define FAST_REDUCTION_ATOMIC_METHOD_GENERATED \ 7594 ((loc->flags & (KMP_IDENT_ATOMIC_REDUCE)) == (KMP_IDENT_ATOMIC_REDUCE)) 7595 #define FAST_REDUCTION_TREE_METHOD_GENERATED ((reduce_data) && (reduce_func)) 7596 7597 retval = critical_reduce_block; 7598 7599 // another choice of getting a team size (with 1 dynamic deference) is slower 7600 team_size = __kmp_get_team_num_threads(global_tid); 7601 if (team_size == 1) { 7602 7603 retval = empty_reduce_block; 7604 7605 } else { 7606 7607 int atomic_available = FAST_REDUCTION_ATOMIC_METHOD_GENERATED; 7608 int tree_available = FAST_REDUCTION_TREE_METHOD_GENERATED; 7609 7610 #if KMP_ARCH_X86_64 || KMP_ARCH_PPC64 || KMP_ARCH_AARCH64 || KMP_ARCH_MIPS64 7611 7612 #if KMP_OS_LINUX || KMP_OS_FREEBSD || KMP_OS_NETBSD || KMP_OS_WINDOWS || \ 7613 KMP_OS_DARWIN 7614 7615 int teamsize_cutoff = 4; 7616 7617 #if KMP_MIC_SUPPORTED 7618 if (__kmp_mic_type != non_mic) { 7619 teamsize_cutoff = 8; 7620 } 7621 #endif 7622 if (tree_available) { 7623 if (team_size <= teamsize_cutoff) { 7624 if (atomic_available) { 7625 retval = atomic_reduce_block; 7626 } 7627 } else { 7628 retval = TREE_REDUCE_BLOCK_WITH_REDUCTION_BARRIER; 7629 } 7630 } else if (atomic_available) { 7631 retval = atomic_reduce_block; 7632 } 7633 #else 7634 #error "Unknown or unsupported OS" 7635 #endif // KMP_OS_LINUX || KMP_OS_FREEBSD || KMP_OS_NETBSD || KMP_OS_WINDOWS || 7636 // KMP_OS_DARWIN 7637 7638 #elif KMP_ARCH_X86 || KMP_ARCH_ARM || KMP_ARCH_AARCH || KMP_ARCH_MIPS 7639 7640 #if KMP_OS_LINUX || KMP_OS_WINDOWS 7641 7642 // basic tuning 7643 7644 if (atomic_available) { 7645 if (num_vars <= 2) { // && ( team_size <= 8 ) due to false-sharing ??? 7646 retval = atomic_reduce_block; 7647 } 7648 } // otherwise: use critical section 7649 7650 #elif KMP_OS_DARWIN 7651 7652 if (atomic_available && (num_vars <= 3)) { 7653 retval = atomic_reduce_block; 7654 } else if (tree_available) { 7655 if ((reduce_size > (9 * sizeof(kmp_real64))) && 7656 (reduce_size < (2000 * sizeof(kmp_real64)))) { 7657 retval = TREE_REDUCE_BLOCK_WITH_PLAIN_BARRIER; 7658 } 7659 } // otherwise: use critical section 7660 7661 #else 7662 #error "Unknown or unsupported OS" 7663 #endif 7664 7665 #else 7666 #error "Unknown or unsupported architecture" 7667 #endif 7668 } 7669 7670 // KMP_FORCE_REDUCTION 7671 7672 // If the team is serialized (team_size == 1), ignore the forced reduction 7673 // method and stay with the unsynchronized method (empty_reduce_block) 7674 if (__kmp_force_reduction_method != reduction_method_not_defined && 7675 team_size != 1) { 7676 7677 PACKED_REDUCTION_METHOD_T forced_retval = critical_reduce_block; 7678 7679 int atomic_available, tree_available; 7680 7681 switch ((forced_retval = __kmp_force_reduction_method)) { 7682 case critical_reduce_block: 7683 KMP_ASSERT(lck); // lck should be != 0 7684 break; 7685 7686 case atomic_reduce_block: 7687 atomic_available = FAST_REDUCTION_ATOMIC_METHOD_GENERATED; 7688 if (!atomic_available) { 7689 KMP_WARNING(RedMethodNotSupported, "atomic"); 7690 forced_retval = critical_reduce_block; 7691 } 7692 break; 7693 7694 case tree_reduce_block: 7695 tree_available = FAST_REDUCTION_TREE_METHOD_GENERATED; 7696 if (!tree_available) { 7697 KMP_WARNING(RedMethodNotSupported, "tree"); 7698 forced_retval = critical_reduce_block; 7699 } else { 7700 #if KMP_FAST_REDUCTION_BARRIER 7701 forced_retval = TREE_REDUCE_BLOCK_WITH_REDUCTION_BARRIER; 7702 #endif 7703 } 7704 break; 7705 7706 default: 7707 KMP_ASSERT(0); // "unsupported method specified" 7708 } 7709 7710 retval = forced_retval; 7711 } 7712 7713 KA_TRACE(10, ("reduction method selected=%08x\n", retval)); 7714 7715 #undef FAST_REDUCTION_TREE_METHOD_GENERATED 7716 #undef FAST_REDUCTION_ATOMIC_METHOD_GENERATED 7717 7718 return (retval); 7719 } 7720 7721 // this function is for testing set/get/determine reduce method 7722 kmp_int32 __kmp_get_reduce_method(void) { 7723 return ((__kmp_entry_thread()->th.th_local.packed_reduction_method) >> 8); 7724 } 7725