1 /*! \file */ 2 /* 3 * kmp.h -- KPTS runtime header file. 4 */ 5 6 //===----------------------------------------------------------------------===// 7 // 8 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 9 // See https://llvm.org/LICENSE.txt for license information. 10 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 11 // 12 //===----------------------------------------------------------------------===// 13 14 #ifndef KMP_H 15 #define KMP_H 16 17 #include "kmp_config.h" 18 19 /* #define BUILD_PARALLEL_ORDERED 1 */ 20 21 /* This fix replaces gettimeofday with clock_gettime for better scalability on 22 the Altix. Requires user code to be linked with -lrt. */ 23 //#define FIX_SGI_CLOCK 24 25 /* Defines for OpenMP 3.0 tasking and auto scheduling */ 26 27 #ifndef KMP_STATIC_STEAL_ENABLED 28 #define KMP_STATIC_STEAL_ENABLED 1 29 #endif 30 31 #define TASK_CURRENT_NOT_QUEUED 0 32 #define TASK_CURRENT_QUEUED 1 33 34 #ifdef BUILD_TIED_TASK_STACK 35 #define TASK_STACK_EMPTY 0 // entries when the stack is empty 36 #define TASK_STACK_BLOCK_BITS 5 // Used in TASK_STACK_SIZE and TASK_STACK_MASK 37 // Number of entries in each task stack array 38 #define TASK_STACK_BLOCK_SIZE (1 << TASK_STACK_BLOCK_BITS) 39 // Mask for determining index into stack block 40 #define TASK_STACK_INDEX_MASK (TASK_STACK_BLOCK_SIZE - 1) 41 #endif // BUILD_TIED_TASK_STACK 42 43 #define TASK_NOT_PUSHED 1 44 #define TASK_SUCCESSFULLY_PUSHED 0 45 #define TASK_TIED 1 46 #define TASK_UNTIED 0 47 #define TASK_EXPLICIT 1 48 #define TASK_IMPLICIT 0 49 #define TASK_PROXY 1 50 #define TASK_FULL 0 51 #define TASK_DETACHABLE 1 52 #define TASK_UNDETACHABLE 0 53 54 #define KMP_CANCEL_THREADS 55 #define KMP_THREAD_ATTR 56 57 // Android does not have pthread_cancel. Undefine KMP_CANCEL_THREADS if being 58 // built on Android 59 #if defined(__ANDROID__) 60 #undef KMP_CANCEL_THREADS 61 #endif 62 63 #include <signal.h> 64 #include <stdarg.h> 65 #include <stddef.h> 66 #include <stdio.h> 67 #include <stdlib.h> 68 #include <string.h> 69 #include <limits> 70 #include <type_traits> 71 /* include <ctype.h> don't use; problems with /MD on Windows* OS NT due to bad 72 Microsoft library. Some macros provided below to replace these functions */ 73 #ifndef __ABSOFT_WIN 74 #include <sys/types.h> 75 #endif 76 #include <limits.h> 77 #include <time.h> 78 79 #include <errno.h> 80 81 #include "kmp_os.h" 82 83 #include "kmp_safe_c_api.h" 84 85 #if KMP_STATS_ENABLED 86 class kmp_stats_list; 87 #endif 88 89 #if KMP_USE_HIER_SCHED 90 // Only include hierarchical scheduling if affinity is supported 91 #undef KMP_USE_HIER_SCHED 92 #define KMP_USE_HIER_SCHED KMP_AFFINITY_SUPPORTED 93 #endif 94 95 #if KMP_USE_HWLOC && KMP_AFFINITY_SUPPORTED 96 #include "hwloc.h" 97 #ifndef HWLOC_OBJ_NUMANODE 98 #define HWLOC_OBJ_NUMANODE HWLOC_OBJ_NODE 99 #endif 100 #ifndef HWLOC_OBJ_PACKAGE 101 #define HWLOC_OBJ_PACKAGE HWLOC_OBJ_SOCKET 102 #endif 103 #endif 104 105 #if KMP_ARCH_X86 || KMP_ARCH_X86_64 106 #include <xmmintrin.h> 107 #endif 108 109 // The below has to be defined before including "kmp_barrier.h". 110 #define KMP_INTERNAL_MALLOC(sz) malloc(sz) 111 #define KMP_INTERNAL_FREE(p) free(p) 112 #define KMP_INTERNAL_REALLOC(p, sz) realloc((p), (sz)) 113 #define KMP_INTERNAL_CALLOC(n, sz) calloc((n), (sz)) 114 115 #include "kmp_debug.h" 116 #include "kmp_lock.h" 117 #include "kmp_version.h" 118 #include "kmp_barrier.h" 119 #if USE_DEBUGGER 120 #include "kmp_debugger.h" 121 #endif 122 #include "kmp_i18n.h" 123 124 #define KMP_HANDLE_SIGNALS (KMP_OS_UNIX || KMP_OS_WINDOWS) 125 126 #include "kmp_wrapper_malloc.h" 127 #if KMP_OS_UNIX 128 #include <unistd.h> 129 #if !defined NSIG && defined _NSIG 130 #define NSIG _NSIG 131 #endif 132 #endif 133 134 #if KMP_OS_LINUX 135 #pragma weak clock_gettime 136 #endif 137 138 #if OMPT_SUPPORT 139 #include "ompt-internal.h" 140 #endif 141 142 #if OMPD_SUPPORT 143 #include "ompd-specific.h" 144 #endif 145 146 #ifndef UNLIKELY 147 #define UNLIKELY(x) (x) 148 #endif 149 150 // Affinity format function 151 #include "kmp_str.h" 152 153 // 0 - no fast memory allocation, alignment: 8-byte on x86, 16-byte on x64. 154 // 3 - fast allocation using sync, non-sync free lists of any size, non-self 155 // free lists of limited size. 156 #ifndef USE_FAST_MEMORY 157 #define USE_FAST_MEMORY 3 158 #endif 159 160 #ifndef KMP_NESTED_HOT_TEAMS 161 #define KMP_NESTED_HOT_TEAMS 0 162 #define USE_NESTED_HOT_ARG(x) 163 #else 164 #if KMP_NESTED_HOT_TEAMS 165 #define USE_NESTED_HOT_ARG(x) , x 166 #else 167 #define USE_NESTED_HOT_ARG(x) 168 #endif 169 #endif 170 171 // Assume using BGET compare_exchange instruction instead of lock by default. 172 #ifndef USE_CMP_XCHG_FOR_BGET 173 #define USE_CMP_XCHG_FOR_BGET 1 174 #endif 175 176 // Test to see if queuing lock is better than bootstrap lock for bget 177 // #ifndef USE_QUEUING_LOCK_FOR_BGET 178 // #define USE_QUEUING_LOCK_FOR_BGET 179 // #endif 180 181 #define KMP_NSEC_PER_SEC 1000000000L 182 #define KMP_USEC_PER_SEC 1000000L 183 184 /*! 185 @ingroup BASIC_TYPES 186 @{ 187 */ 188 189 /*! 190 Values for bit flags used in the ident_t to describe the fields. 191 */ 192 enum { 193 /*! Use trampoline for internal microtasks */ 194 KMP_IDENT_IMB = 0x01, 195 /*! Use c-style ident structure */ 196 KMP_IDENT_KMPC = 0x02, 197 /* 0x04 is no longer used */ 198 /*! Entry point generated by auto-parallelization */ 199 KMP_IDENT_AUTOPAR = 0x08, 200 /*! Compiler generates atomic reduction option for kmpc_reduce* */ 201 KMP_IDENT_ATOMIC_REDUCE = 0x10, 202 /*! To mark a 'barrier' directive in user code */ 203 KMP_IDENT_BARRIER_EXPL = 0x20, 204 /*! To Mark implicit barriers. */ 205 KMP_IDENT_BARRIER_IMPL = 0x0040, 206 KMP_IDENT_BARRIER_IMPL_MASK = 0x01C0, 207 KMP_IDENT_BARRIER_IMPL_FOR = 0x0040, 208 KMP_IDENT_BARRIER_IMPL_SECTIONS = 0x00C0, 209 210 KMP_IDENT_BARRIER_IMPL_SINGLE = 0x0140, 211 KMP_IDENT_BARRIER_IMPL_WORKSHARE = 0x01C0, 212 213 /*! To mark a static loop in OMPT callbacks */ 214 KMP_IDENT_WORK_LOOP = 0x200, 215 /*! To mark a sections directive in OMPT callbacks */ 216 KMP_IDENT_WORK_SECTIONS = 0x400, 217 /*! To mark a distribute construct in OMPT callbacks */ 218 KMP_IDENT_WORK_DISTRIBUTE = 0x800, 219 /*! Atomic hint; bottom four bits as omp_sync_hint_t. Top four reserved and 220 not currently used. If one day we need more bits, then we can use 221 an invalid combination of hints to mean that another, larger field 222 should be used in a different flag. */ 223 KMP_IDENT_ATOMIC_HINT_MASK = 0xFF0000, 224 KMP_IDENT_ATOMIC_HINT_UNCONTENDED = 0x010000, 225 KMP_IDENT_ATOMIC_HINT_CONTENDED = 0x020000, 226 KMP_IDENT_ATOMIC_HINT_NONSPECULATIVE = 0x040000, 227 KMP_IDENT_ATOMIC_HINT_SPECULATIVE = 0x080000, 228 KMP_IDENT_OPENMP_SPEC_VERSION_MASK = 0xFF000000 229 }; 230 231 /*! 232 * The ident structure that describes a source location. 233 */ 234 typedef struct ident { 235 kmp_int32 reserved_1; /**< might be used in Fortran; see above */ 236 kmp_int32 flags; /**< also f.flags; KMP_IDENT_xxx flags; KMP_IDENT_KMPC 237 identifies this union member */ 238 kmp_int32 reserved_2; /**< not really used in Fortran any more; see above */ 239 #if USE_ITT_BUILD 240 /* but currently used for storing region-specific ITT */ 241 /* contextual information. */ 242 #endif /* USE_ITT_BUILD */ 243 kmp_int32 reserved_3; /**< source[4] in Fortran, do not use for C++ */ 244 char const *psource; /**< String describing the source location. 245 The string is composed of semi-colon separated fields 246 which describe the source file, the function and a pair 247 of line numbers that delimit the construct. */ 248 // Returns the OpenMP version in form major*10+minor (e.g., 50 for 5.0) 249 kmp_int32 get_openmp_version() { 250 return (((flags & KMP_IDENT_OPENMP_SPEC_VERSION_MASK) >> 24) & 0xFF); 251 } 252 } ident_t; 253 /*! 254 @} 255 */ 256 257 // Some forward declarations. 258 typedef union kmp_team kmp_team_t; 259 typedef struct kmp_taskdata kmp_taskdata_t; 260 typedef union kmp_task_team kmp_task_team_t; 261 typedef union kmp_team kmp_team_p; 262 typedef union kmp_info kmp_info_p; 263 typedef union kmp_root kmp_root_p; 264 265 template <bool C = false, bool S = true> class kmp_flag_32; 266 template <bool C = false, bool S = true> class kmp_flag_64; 267 template <bool C = false, bool S = true> class kmp_atomic_flag_64; 268 class kmp_flag_oncore; 269 270 #ifdef __cplusplus 271 extern "C" { 272 #endif 273 274 /* ------------------------------------------------------------------------ */ 275 276 /* Pack two 32-bit signed integers into a 64-bit signed integer */ 277 /* ToDo: Fix word ordering for big-endian machines. */ 278 #define KMP_PACK_64(HIGH_32, LOW_32) \ 279 ((kmp_int64)((((kmp_uint64)(HIGH_32)) << 32) | (kmp_uint64)(LOW_32))) 280 281 // Generic string manipulation macros. Assume that _x is of type char * 282 #define SKIP_WS(_x) \ 283 { \ 284 while (*(_x) == ' ' || *(_x) == '\t') \ 285 (_x)++; \ 286 } 287 #define SKIP_DIGITS(_x) \ 288 { \ 289 while (*(_x) >= '0' && *(_x) <= '9') \ 290 (_x)++; \ 291 } 292 #define SKIP_TOKEN(_x) \ 293 { \ 294 while ((*(_x) >= '0' && *(_x) <= '9') || (*(_x) >= 'a' && *(_x) <= 'z') || \ 295 (*(_x) >= 'A' && *(_x) <= 'Z') || *(_x) == '_') \ 296 (_x)++; \ 297 } 298 #define SKIP_TO(_x, _c) \ 299 { \ 300 while (*(_x) != '\0' && *(_x) != (_c)) \ 301 (_x)++; \ 302 } 303 304 /* ------------------------------------------------------------------------ */ 305 306 #define KMP_MAX(x, y) ((x) > (y) ? (x) : (y)) 307 #define KMP_MIN(x, y) ((x) < (y) ? (x) : (y)) 308 309 /* ------------------------------------------------------------------------ */ 310 /* Enumeration types */ 311 312 enum kmp_state_timer { 313 ts_stop, 314 ts_start, 315 ts_pause, 316 317 ts_last_state 318 }; 319 320 enum dynamic_mode { 321 dynamic_default, 322 #ifdef USE_LOAD_BALANCE 323 dynamic_load_balance, 324 #endif /* USE_LOAD_BALANCE */ 325 dynamic_random, 326 dynamic_thread_limit, 327 dynamic_max 328 }; 329 330 /* external schedule constants, duplicate enum omp_sched in omp.h in order to 331 * not include it here */ 332 #ifndef KMP_SCHED_TYPE_DEFINED 333 #define KMP_SCHED_TYPE_DEFINED 334 typedef enum kmp_sched { 335 kmp_sched_lower = 0, // lower and upper bounds are for routine parameter check 336 // Note: need to adjust __kmp_sch_map global array in case enum is changed 337 kmp_sched_static = 1, // mapped to kmp_sch_static_chunked (33) 338 kmp_sched_dynamic = 2, // mapped to kmp_sch_dynamic_chunked (35) 339 kmp_sched_guided = 3, // mapped to kmp_sch_guided_chunked (36) 340 kmp_sched_auto = 4, // mapped to kmp_sch_auto (38) 341 kmp_sched_upper_std = 5, // upper bound for standard schedules 342 kmp_sched_lower_ext = 100, // lower bound of Intel extension schedules 343 kmp_sched_trapezoidal = 101, // mapped to kmp_sch_trapezoidal (39) 344 #if KMP_STATIC_STEAL_ENABLED 345 kmp_sched_static_steal = 102, // mapped to kmp_sch_static_steal (44) 346 #endif 347 kmp_sched_upper, 348 kmp_sched_default = kmp_sched_static, // default scheduling 349 kmp_sched_monotonic = 0x80000000 350 } kmp_sched_t; 351 #endif 352 353 /*! 354 @ingroup WORK_SHARING 355 * Describes the loop schedule to be used for a parallel for loop. 356 */ 357 enum sched_type : kmp_int32 { 358 kmp_sch_lower = 32, /**< lower bound for unordered values */ 359 kmp_sch_static_chunked = 33, 360 kmp_sch_static = 34, /**< static unspecialized */ 361 kmp_sch_dynamic_chunked = 35, 362 kmp_sch_guided_chunked = 36, /**< guided unspecialized */ 363 kmp_sch_runtime = 37, 364 kmp_sch_auto = 38, /**< auto */ 365 kmp_sch_trapezoidal = 39, 366 367 /* accessible only through KMP_SCHEDULE environment variable */ 368 kmp_sch_static_greedy = 40, 369 kmp_sch_static_balanced = 41, 370 /* accessible only through KMP_SCHEDULE environment variable */ 371 kmp_sch_guided_iterative_chunked = 42, 372 kmp_sch_guided_analytical_chunked = 43, 373 /* accessible only through KMP_SCHEDULE environment variable */ 374 kmp_sch_static_steal = 44, 375 376 /* static with chunk adjustment (e.g., simd) */ 377 kmp_sch_static_balanced_chunked = 45, 378 kmp_sch_guided_simd = 46, /**< guided with chunk adjustment */ 379 kmp_sch_runtime_simd = 47, /**< runtime with chunk adjustment */ 380 381 /* accessible only through KMP_SCHEDULE environment variable */ 382 kmp_sch_upper, /**< upper bound for unordered values */ 383 384 kmp_ord_lower = 64, /**< lower bound for ordered values, must be power of 2 */ 385 kmp_ord_static_chunked = 65, 386 kmp_ord_static = 66, /**< ordered static unspecialized */ 387 kmp_ord_dynamic_chunked = 67, 388 kmp_ord_guided_chunked = 68, 389 kmp_ord_runtime = 69, 390 kmp_ord_auto = 70, /**< ordered auto */ 391 kmp_ord_trapezoidal = 71, 392 kmp_ord_upper, /**< upper bound for ordered values */ 393 394 /* Schedules for Distribute construct */ 395 kmp_distribute_static_chunked = 91, /**< distribute static chunked */ 396 kmp_distribute_static = 92, /**< distribute static unspecialized */ 397 398 /* For the "nomerge" versions, kmp_dispatch_next*() will always return a 399 single iteration/chunk, even if the loop is serialized. For the schedule 400 types listed above, the entire iteration vector is returned if the loop is 401 serialized. This doesn't work for gcc/gcomp sections. */ 402 kmp_nm_lower = 160, /**< lower bound for nomerge values */ 403 404 kmp_nm_static_chunked = 405 (kmp_sch_static_chunked - kmp_sch_lower + kmp_nm_lower), 406 kmp_nm_static = 162, /**< static unspecialized */ 407 kmp_nm_dynamic_chunked = 163, 408 kmp_nm_guided_chunked = 164, /**< guided unspecialized */ 409 kmp_nm_runtime = 165, 410 kmp_nm_auto = 166, /**< auto */ 411 kmp_nm_trapezoidal = 167, 412 413 /* accessible only through KMP_SCHEDULE environment variable */ 414 kmp_nm_static_greedy = 168, 415 kmp_nm_static_balanced = 169, 416 /* accessible only through KMP_SCHEDULE environment variable */ 417 kmp_nm_guided_iterative_chunked = 170, 418 kmp_nm_guided_analytical_chunked = 171, 419 kmp_nm_static_steal = 420 172, /* accessible only through OMP_SCHEDULE environment variable */ 421 422 kmp_nm_ord_static_chunked = 193, 423 kmp_nm_ord_static = 194, /**< ordered static unspecialized */ 424 kmp_nm_ord_dynamic_chunked = 195, 425 kmp_nm_ord_guided_chunked = 196, 426 kmp_nm_ord_runtime = 197, 427 kmp_nm_ord_auto = 198, /**< auto */ 428 kmp_nm_ord_trapezoidal = 199, 429 kmp_nm_upper, /**< upper bound for nomerge values */ 430 431 /* Support for OpenMP 4.5 monotonic and nonmonotonic schedule modifiers. Since 432 we need to distinguish the three possible cases (no modifier, monotonic 433 modifier, nonmonotonic modifier), we need separate bits for each modifier. 434 The absence of monotonic does not imply nonmonotonic, especially since 4.5 435 says that the behaviour of the "no modifier" case is implementation defined 436 in 4.5, but will become "nonmonotonic" in 5.0. 437 438 Since we're passing a full 32 bit value, we can use a couple of high bits 439 for these flags; out of paranoia we avoid the sign bit. 440 441 These modifiers can be or-ed into non-static schedules by the compiler to 442 pass the additional information. They will be stripped early in the 443 processing in __kmp_dispatch_init when setting up schedules, so most of the 444 code won't ever see schedules with these bits set. */ 445 kmp_sch_modifier_monotonic = 446 (1 << 29), /**< Set if the monotonic schedule modifier was present */ 447 kmp_sch_modifier_nonmonotonic = 448 (1 << 30), /**< Set if the nonmonotonic schedule modifier was present */ 449 450 #define SCHEDULE_WITHOUT_MODIFIERS(s) \ 451 (enum sched_type)( \ 452 (s) & ~(kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic)) 453 #define SCHEDULE_HAS_MONOTONIC(s) (((s)&kmp_sch_modifier_monotonic) != 0) 454 #define SCHEDULE_HAS_NONMONOTONIC(s) (((s)&kmp_sch_modifier_nonmonotonic) != 0) 455 #define SCHEDULE_HAS_NO_MODIFIERS(s) \ 456 (((s) & (kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic)) == 0) 457 #define SCHEDULE_GET_MODIFIERS(s) \ 458 ((enum sched_type)( \ 459 (s) & (kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic))) 460 #define SCHEDULE_SET_MODIFIERS(s, m) \ 461 (s = (enum sched_type)((kmp_int32)s | (kmp_int32)m)) 462 #define SCHEDULE_NONMONOTONIC 0 463 #define SCHEDULE_MONOTONIC 1 464 465 kmp_sch_default = kmp_sch_static /**< default scheduling algorithm */ 466 }; 467 468 // Apply modifiers on internal kind to standard kind 469 static inline void 470 __kmp_sched_apply_mods_stdkind(kmp_sched_t *kind, 471 enum sched_type internal_kind) { 472 if (SCHEDULE_HAS_MONOTONIC(internal_kind)) { 473 *kind = (kmp_sched_t)((int)*kind | (int)kmp_sched_monotonic); 474 } 475 } 476 477 // Apply modifiers on standard kind to internal kind 478 static inline void 479 __kmp_sched_apply_mods_intkind(kmp_sched_t kind, 480 enum sched_type *internal_kind) { 481 if ((int)kind & (int)kmp_sched_monotonic) { 482 *internal_kind = (enum sched_type)((int)*internal_kind | 483 (int)kmp_sch_modifier_monotonic); 484 } 485 } 486 487 // Get standard schedule without modifiers 488 static inline kmp_sched_t __kmp_sched_without_mods(kmp_sched_t kind) { 489 return (kmp_sched_t)((int)kind & ~((int)kmp_sched_monotonic)); 490 } 491 492 /* Type to keep runtime schedule set via OMP_SCHEDULE or omp_set_schedule() */ 493 typedef union kmp_r_sched { 494 struct { 495 enum sched_type r_sched_type; 496 int chunk; 497 }; 498 kmp_int64 sched; 499 } kmp_r_sched_t; 500 501 extern enum sched_type __kmp_sch_map[]; // map OMP 3.0 schedule types with our 502 // internal schedule types 503 504 enum library_type { 505 library_none, 506 library_serial, 507 library_turnaround, 508 library_throughput 509 }; 510 511 #if KMP_OS_LINUX 512 enum clock_function_type { 513 clock_function_gettimeofday, 514 clock_function_clock_gettime 515 }; 516 #endif /* KMP_OS_LINUX */ 517 518 #if KMP_MIC_SUPPORTED 519 enum mic_type { non_mic, mic1, mic2, mic3, dummy }; 520 #endif 521 522 /* -- fast reduction stuff ------------------------------------------------ */ 523 524 #undef KMP_FAST_REDUCTION_BARRIER 525 #define KMP_FAST_REDUCTION_BARRIER 1 526 527 #undef KMP_FAST_REDUCTION_CORE_DUO 528 #if KMP_ARCH_X86 || KMP_ARCH_X86_64 529 #define KMP_FAST_REDUCTION_CORE_DUO 1 530 #endif 531 532 enum _reduction_method { 533 reduction_method_not_defined = 0, 534 critical_reduce_block = (1 << 8), 535 atomic_reduce_block = (2 << 8), 536 tree_reduce_block = (3 << 8), 537 empty_reduce_block = (4 << 8) 538 }; 539 540 // Description of the packed_reduction_method variable: 541 // The packed_reduction_method variable consists of two enum types variables 542 // that are packed together into 0-th byte and 1-st byte: 543 // 0: (packed_reduction_method & 0x000000FF) is a 'enum barrier_type' value of 544 // barrier that will be used in fast reduction: bs_plain_barrier or 545 // bs_reduction_barrier 546 // 1: (packed_reduction_method & 0x0000FF00) is a reduction method that will 547 // be used in fast reduction; 548 // Reduction method is of 'enum _reduction_method' type and it's defined the way 549 // so that the bits of 0-th byte are empty, so no need to execute a shift 550 // instruction while packing/unpacking 551 552 #if KMP_FAST_REDUCTION_BARRIER 553 #define PACK_REDUCTION_METHOD_AND_BARRIER(reduction_method, barrier_type) \ 554 ((reduction_method) | (barrier_type)) 555 556 #define UNPACK_REDUCTION_METHOD(packed_reduction_method) \ 557 ((enum _reduction_method)((packed_reduction_method) & (0x0000FF00))) 558 559 #define UNPACK_REDUCTION_BARRIER(packed_reduction_method) \ 560 ((enum barrier_type)((packed_reduction_method) & (0x000000FF))) 561 #else 562 #define PACK_REDUCTION_METHOD_AND_BARRIER(reduction_method, barrier_type) \ 563 (reduction_method) 564 565 #define UNPACK_REDUCTION_METHOD(packed_reduction_method) \ 566 (packed_reduction_method) 567 568 #define UNPACK_REDUCTION_BARRIER(packed_reduction_method) (bs_plain_barrier) 569 #endif 570 571 #define TEST_REDUCTION_METHOD(packed_reduction_method, which_reduction_block) \ 572 ((UNPACK_REDUCTION_METHOD(packed_reduction_method)) == \ 573 (which_reduction_block)) 574 575 #if KMP_FAST_REDUCTION_BARRIER 576 #define TREE_REDUCE_BLOCK_WITH_REDUCTION_BARRIER \ 577 (PACK_REDUCTION_METHOD_AND_BARRIER(tree_reduce_block, bs_reduction_barrier)) 578 579 #define TREE_REDUCE_BLOCK_WITH_PLAIN_BARRIER \ 580 (PACK_REDUCTION_METHOD_AND_BARRIER(tree_reduce_block, bs_plain_barrier)) 581 #endif 582 583 typedef int PACKED_REDUCTION_METHOD_T; 584 585 /* -- end of fast reduction stuff ----------------------------------------- */ 586 587 #if KMP_OS_WINDOWS 588 #define USE_CBLKDATA 589 #if KMP_MSVC_COMPAT 590 #pragma warning(push) 591 #pragma warning(disable : 271 310) 592 #endif 593 #include <windows.h> 594 #if KMP_MSVC_COMPAT 595 #pragma warning(pop) 596 #endif 597 #endif 598 599 #if KMP_OS_UNIX 600 #include <dlfcn.h> 601 #include <pthread.h> 602 #endif 603 604 enum kmp_hw_t : int { 605 KMP_HW_UNKNOWN = -1, 606 KMP_HW_SOCKET = 0, 607 KMP_HW_PROC_GROUP, 608 KMP_HW_NUMA, 609 KMP_HW_DIE, 610 KMP_HW_LLC, 611 KMP_HW_L3, 612 KMP_HW_TILE, 613 KMP_HW_MODULE, 614 KMP_HW_L2, 615 KMP_HW_L1, 616 KMP_HW_CORE, 617 KMP_HW_THREAD, 618 KMP_HW_LAST 619 }; 620 621 typedef enum kmp_hw_core_type_t { 622 KMP_HW_CORE_TYPE_UNKNOWN = 0x0, 623 #if KMP_ARCH_X86 || KMP_ARCH_X86_64 624 KMP_HW_CORE_TYPE_ATOM = 0x20, 625 KMP_HW_CORE_TYPE_CORE = 0x40, 626 KMP_HW_MAX_NUM_CORE_TYPES = 3, 627 #else 628 KMP_HW_MAX_NUM_CORE_TYPES = 1, 629 #endif 630 } kmp_hw_core_type_t; 631 632 #define KMP_HW_MAX_NUM_CORE_EFFS 8 633 634 #define KMP_DEBUG_ASSERT_VALID_HW_TYPE(type) \ 635 KMP_DEBUG_ASSERT(type >= (kmp_hw_t)0 && type < KMP_HW_LAST) 636 #define KMP_ASSERT_VALID_HW_TYPE(type) \ 637 KMP_ASSERT(type >= (kmp_hw_t)0 && type < KMP_HW_LAST) 638 639 #define KMP_FOREACH_HW_TYPE(type) \ 640 for (kmp_hw_t type = (kmp_hw_t)0; type < KMP_HW_LAST; \ 641 type = (kmp_hw_t)((int)type + 1)) 642 643 const char *__kmp_hw_get_keyword(kmp_hw_t type, bool plural = false); 644 const char *__kmp_hw_get_catalog_string(kmp_hw_t type, bool plural = false); 645 const char *__kmp_hw_get_core_type_string(kmp_hw_core_type_t type); 646 647 /* Only Linux* OS and Windows* OS support thread affinity. */ 648 #if KMP_AFFINITY_SUPPORTED 649 650 // GROUP_AFFINITY is already defined for _MSC_VER>=1600 (VS2010 and later). 651 #if KMP_OS_WINDOWS 652 #if _MSC_VER < 1600 && KMP_MSVC_COMPAT 653 typedef struct GROUP_AFFINITY { 654 KAFFINITY Mask; 655 WORD Group; 656 WORD Reserved[3]; 657 } GROUP_AFFINITY; 658 #endif /* _MSC_VER < 1600 */ 659 #if KMP_GROUP_AFFINITY 660 extern int __kmp_num_proc_groups; 661 #else 662 static const int __kmp_num_proc_groups = 1; 663 #endif /* KMP_GROUP_AFFINITY */ 664 typedef DWORD (*kmp_GetActiveProcessorCount_t)(WORD); 665 extern kmp_GetActiveProcessorCount_t __kmp_GetActiveProcessorCount; 666 667 typedef WORD (*kmp_GetActiveProcessorGroupCount_t)(void); 668 extern kmp_GetActiveProcessorGroupCount_t __kmp_GetActiveProcessorGroupCount; 669 670 typedef BOOL (*kmp_GetThreadGroupAffinity_t)(HANDLE, GROUP_AFFINITY *); 671 extern kmp_GetThreadGroupAffinity_t __kmp_GetThreadGroupAffinity; 672 673 typedef BOOL (*kmp_SetThreadGroupAffinity_t)(HANDLE, const GROUP_AFFINITY *, 674 GROUP_AFFINITY *); 675 extern kmp_SetThreadGroupAffinity_t __kmp_SetThreadGroupAffinity; 676 #endif /* KMP_OS_WINDOWS */ 677 678 #if KMP_USE_HWLOC 679 extern hwloc_topology_t __kmp_hwloc_topology; 680 extern int __kmp_hwloc_error; 681 #endif 682 683 extern size_t __kmp_affin_mask_size; 684 #define KMP_AFFINITY_CAPABLE() (__kmp_affin_mask_size > 0) 685 #define KMP_AFFINITY_DISABLE() (__kmp_affin_mask_size = 0) 686 #define KMP_AFFINITY_ENABLE(mask_size) (__kmp_affin_mask_size = mask_size) 687 #define KMP_CPU_SET_ITERATE(i, mask) \ 688 for (i = (mask)->begin(); (int)i != (mask)->end(); i = (mask)->next(i)) 689 #define KMP_CPU_SET(i, mask) (mask)->set(i) 690 #define KMP_CPU_ISSET(i, mask) (mask)->is_set(i) 691 #define KMP_CPU_CLR(i, mask) (mask)->clear(i) 692 #define KMP_CPU_ZERO(mask) (mask)->zero() 693 #define KMP_CPU_COPY(dest, src) (dest)->copy(src) 694 #define KMP_CPU_AND(dest, src) (dest)->bitwise_and(src) 695 #define KMP_CPU_COMPLEMENT(max_bit_number, mask) (mask)->bitwise_not() 696 #define KMP_CPU_UNION(dest, src) (dest)->bitwise_or(src) 697 #define KMP_CPU_ALLOC(ptr) (ptr = __kmp_affinity_dispatch->allocate_mask()) 698 #define KMP_CPU_FREE(ptr) __kmp_affinity_dispatch->deallocate_mask(ptr) 699 #define KMP_CPU_ALLOC_ON_STACK(ptr) KMP_CPU_ALLOC(ptr) 700 #define KMP_CPU_FREE_FROM_STACK(ptr) KMP_CPU_FREE(ptr) 701 #define KMP_CPU_INTERNAL_ALLOC(ptr) KMP_CPU_ALLOC(ptr) 702 #define KMP_CPU_INTERNAL_FREE(ptr) KMP_CPU_FREE(ptr) 703 #define KMP_CPU_INDEX(arr, i) __kmp_affinity_dispatch->index_mask_array(arr, i) 704 #define KMP_CPU_ALLOC_ARRAY(arr, n) \ 705 (arr = __kmp_affinity_dispatch->allocate_mask_array(n)) 706 #define KMP_CPU_FREE_ARRAY(arr, n) \ 707 __kmp_affinity_dispatch->deallocate_mask_array(arr) 708 #define KMP_CPU_INTERNAL_ALLOC_ARRAY(arr, n) KMP_CPU_ALLOC_ARRAY(arr, n) 709 #define KMP_CPU_INTERNAL_FREE_ARRAY(arr, n) KMP_CPU_FREE_ARRAY(arr, n) 710 #define __kmp_get_system_affinity(mask, abort_bool) \ 711 (mask)->get_system_affinity(abort_bool) 712 #define __kmp_set_system_affinity(mask, abort_bool) \ 713 (mask)->set_system_affinity(abort_bool) 714 #define __kmp_get_proc_group(mask) (mask)->get_proc_group() 715 716 class KMPAffinity { 717 public: 718 class Mask { 719 public: 720 void *operator new(size_t n); 721 void operator delete(void *p); 722 void *operator new[](size_t n); 723 void operator delete[](void *p); 724 virtual ~Mask() {} 725 // Set bit i to 1 726 virtual void set(int i) {} 727 // Return bit i 728 virtual bool is_set(int i) const { return false; } 729 // Set bit i to 0 730 virtual void clear(int i) {} 731 // Zero out entire mask 732 virtual void zero() {} 733 // Copy src into this mask 734 virtual void copy(const Mask *src) {} 735 // this &= rhs 736 virtual void bitwise_and(const Mask *rhs) {} 737 // this |= rhs 738 virtual void bitwise_or(const Mask *rhs) {} 739 // this = ~this 740 virtual void bitwise_not() {} 741 // API for iterating over an affinity mask 742 // for (int i = mask->begin(); i != mask->end(); i = mask->next(i)) 743 virtual int begin() const { return 0; } 744 virtual int end() const { return 0; } 745 virtual int next(int previous) const { return 0; } 746 #if KMP_OS_WINDOWS 747 virtual int set_process_affinity(bool abort_on_error) const { return -1; } 748 #endif 749 // Set the system's affinity to this affinity mask's value 750 virtual int set_system_affinity(bool abort_on_error) const { return -1; } 751 // Set this affinity mask to the current system affinity 752 virtual int get_system_affinity(bool abort_on_error) { return -1; } 753 // Only 1 DWORD in the mask should have any procs set. 754 // Return the appropriate index, or -1 for an invalid mask. 755 virtual int get_proc_group() const { return -1; } 756 int get_max_cpu() const { 757 int cpu; 758 int max_cpu = -1; 759 KMP_CPU_SET_ITERATE(cpu, this) { 760 if (cpu > max_cpu) 761 max_cpu = cpu; 762 } 763 return max_cpu; 764 } 765 }; 766 void *operator new(size_t n); 767 void operator delete(void *p); 768 // Need virtual destructor 769 virtual ~KMPAffinity() = default; 770 // Determine if affinity is capable 771 virtual void determine_capable(const char *env_var) {} 772 // Bind the current thread to os proc 773 virtual void bind_thread(int proc) {} 774 // Factory functions to allocate/deallocate a mask 775 virtual Mask *allocate_mask() { return nullptr; } 776 virtual void deallocate_mask(Mask *m) {} 777 virtual Mask *allocate_mask_array(int num) { return nullptr; } 778 virtual void deallocate_mask_array(Mask *m) {} 779 virtual Mask *index_mask_array(Mask *m, int index) { return nullptr; } 780 static void pick_api(); 781 static void destroy_api(); 782 enum api_type { 783 NATIVE_OS 784 #if KMP_USE_HWLOC 785 , 786 HWLOC 787 #endif 788 }; 789 virtual api_type get_api_type() const { 790 KMP_ASSERT(0); 791 return NATIVE_OS; 792 } 793 794 private: 795 static bool picked_api; 796 }; 797 798 typedef KMPAffinity::Mask kmp_affin_mask_t; 799 extern KMPAffinity *__kmp_affinity_dispatch; 800 801 class kmp_affinity_raii_t { 802 kmp_affin_mask_t *mask; 803 bool restored; 804 805 public: 806 kmp_affinity_raii_t(const kmp_affin_mask_t *new_mask = nullptr) 807 : restored(false) { 808 if (KMP_AFFINITY_CAPABLE()) { 809 KMP_CPU_ALLOC(mask); 810 KMP_ASSERT(mask != NULL); 811 __kmp_get_system_affinity(mask, /*abort_on_error=*/true); 812 if (new_mask) 813 __kmp_set_system_affinity(new_mask, /*abort_on_error=*/true); 814 } 815 } 816 void restore() { 817 if (!restored && KMP_AFFINITY_CAPABLE()) { 818 __kmp_set_system_affinity(mask, /*abort_on_error=*/true); 819 KMP_CPU_FREE(mask); 820 } 821 restored = true; 822 } 823 ~kmp_affinity_raii_t() { restore(); } 824 }; 825 826 // Declare local char buffers with this size for printing debug and info 827 // messages, using __kmp_affinity_print_mask(). 828 #define KMP_AFFIN_MASK_PRINT_LEN 1024 829 830 enum affinity_type { 831 affinity_none = 0, 832 affinity_physical, 833 affinity_logical, 834 affinity_compact, 835 affinity_scatter, 836 affinity_explicit, 837 affinity_balanced, 838 affinity_disabled, // not used outsize the env var parser 839 affinity_default 840 }; 841 842 enum affinity_top_method { 843 affinity_top_method_all = 0, // try all (supported) methods, in order 844 #if KMP_ARCH_X86 || KMP_ARCH_X86_64 845 affinity_top_method_apicid, 846 affinity_top_method_x2apicid, 847 affinity_top_method_x2apicid_1f, 848 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */ 849 affinity_top_method_cpuinfo, // KMP_CPUINFO_FILE is usable on Windows* OS, too 850 #if KMP_GROUP_AFFINITY 851 affinity_top_method_group, 852 #endif /* KMP_GROUP_AFFINITY */ 853 affinity_top_method_flat, 854 #if KMP_USE_HWLOC 855 affinity_top_method_hwloc, 856 #endif 857 affinity_top_method_default 858 }; 859 860 #define affinity_respect_mask_default (2) 861 862 typedef struct kmp_affinity_flags_t { 863 unsigned dups : 1; 864 unsigned verbose : 1; 865 unsigned warnings : 1; 866 unsigned respect : 2; 867 unsigned reset : 1; 868 unsigned initialized : 1; 869 unsigned reserved : 25; 870 } kmp_affinity_flags_t; 871 KMP_BUILD_ASSERT(sizeof(kmp_affinity_flags_t) == 4); 872 873 typedef struct kmp_affinity_ids_t { 874 int ids[KMP_HW_LAST]; 875 int operator[](size_t idx) const { return ids[idx]; } 876 int &operator[](size_t idx) { return ids[idx]; } 877 kmp_affinity_ids_t &operator=(const kmp_affinity_ids_t &rhs) { 878 for (int i = 0; i < KMP_HW_LAST; ++i) 879 ids[i] = rhs[i]; 880 return *this; 881 } 882 } kmp_affinity_ids_t; 883 884 typedef struct kmp_affinity_attrs_t { 885 int core_type : 8; 886 int core_eff : 8; 887 unsigned valid : 1; 888 unsigned reserved : 15; 889 } kmp_affinity_attrs_t; 890 #define KMP_AFFINITY_ATTRS_UNKNOWN \ 891 { KMP_HW_CORE_TYPE_UNKNOWN, kmp_hw_attr_t::UNKNOWN_CORE_EFF, 0, 0 } 892 893 typedef struct kmp_affinity_t { 894 char *proclist; 895 enum affinity_type type; 896 kmp_hw_t gran; 897 int gran_levels; 898 int compact; 899 int offset; 900 kmp_affinity_flags_t flags; 901 unsigned num_masks; 902 kmp_affin_mask_t *masks; 903 kmp_affinity_ids_t *ids; 904 kmp_affinity_attrs_t *attrs; 905 unsigned num_os_id_masks; 906 kmp_affin_mask_t *os_id_masks; 907 const char *env_var; 908 } kmp_affinity_t; 909 910 #define KMP_AFFINITY_INIT(env) \ 911 { \ 912 nullptr, affinity_default, KMP_HW_UNKNOWN, -1, 0, 0, \ 913 {TRUE, FALSE, TRUE, affinity_respect_mask_default, FALSE, FALSE}, 0, \ 914 nullptr, nullptr, nullptr, 0, nullptr, env \ 915 } 916 917 extern enum affinity_top_method __kmp_affinity_top_method; 918 extern kmp_affinity_t __kmp_affinity; 919 extern kmp_affinity_t __kmp_hh_affinity; 920 extern kmp_affinity_t *__kmp_affinities[2]; 921 922 extern void __kmp_affinity_bind_thread(int which); 923 924 extern kmp_affin_mask_t *__kmp_affin_fullMask; 925 extern kmp_affin_mask_t *__kmp_affin_origMask; 926 extern char *__kmp_cpuinfo_file; 927 928 #endif /* KMP_AFFINITY_SUPPORTED */ 929 930 // This needs to be kept in sync with the values in omp.h !!! 931 typedef enum kmp_proc_bind_t { 932 proc_bind_false = 0, 933 proc_bind_true, 934 proc_bind_primary, 935 proc_bind_close, 936 proc_bind_spread, 937 proc_bind_intel, // use KMP_AFFINITY interface 938 proc_bind_default 939 } kmp_proc_bind_t; 940 941 typedef struct kmp_nested_proc_bind_t { 942 kmp_proc_bind_t *bind_types; 943 int size; 944 int used; 945 } kmp_nested_proc_bind_t; 946 947 extern kmp_nested_proc_bind_t __kmp_nested_proc_bind; 948 extern kmp_proc_bind_t __kmp_teams_proc_bind; 949 950 extern int __kmp_display_affinity; 951 extern char *__kmp_affinity_format; 952 static const size_t KMP_AFFINITY_FORMAT_SIZE = 512; 953 #if OMPT_SUPPORT 954 extern int __kmp_tool; 955 extern char *__kmp_tool_libraries; 956 #endif // OMPT_SUPPORT 957 958 #if KMP_AFFINITY_SUPPORTED 959 #define KMP_PLACE_ALL (-1) 960 #define KMP_PLACE_UNDEFINED (-2) 961 // Is KMP_AFFINITY is being used instead of OMP_PROC_BIND/OMP_PLACES? 962 #define KMP_AFFINITY_NON_PROC_BIND \ 963 ((__kmp_nested_proc_bind.bind_types[0] == proc_bind_false || \ 964 __kmp_nested_proc_bind.bind_types[0] == proc_bind_intel) && \ 965 (__kmp_affinity.num_masks > 0 || __kmp_affinity.type == affinity_balanced)) 966 #endif /* KMP_AFFINITY_SUPPORTED */ 967 968 extern int __kmp_affinity_num_places; 969 970 typedef enum kmp_cancel_kind_t { 971 cancel_noreq = 0, 972 cancel_parallel = 1, 973 cancel_loop = 2, 974 cancel_sections = 3, 975 cancel_taskgroup = 4 976 } kmp_cancel_kind_t; 977 978 // KMP_HW_SUBSET support: 979 typedef struct kmp_hws_item { 980 int num; 981 int offset; 982 } kmp_hws_item_t; 983 984 extern kmp_hws_item_t __kmp_hws_socket; 985 extern kmp_hws_item_t __kmp_hws_die; 986 extern kmp_hws_item_t __kmp_hws_node; 987 extern kmp_hws_item_t __kmp_hws_tile; 988 extern kmp_hws_item_t __kmp_hws_core; 989 extern kmp_hws_item_t __kmp_hws_proc; 990 extern int __kmp_hws_requested; 991 extern int __kmp_hws_abs_flag; // absolute or per-item number requested 992 993 /* ------------------------------------------------------------------------ */ 994 995 #define KMP_PAD(type, sz) \ 996 (sizeof(type) + (sz - ((sizeof(type) - 1) % (sz)) - 1)) 997 998 // We need to avoid using -1 as a GTID as +1 is added to the gtid 999 // when storing it in a lock, and the value 0 is reserved. 1000 #define KMP_GTID_DNE (-2) /* Does not exist */ 1001 #define KMP_GTID_SHUTDOWN (-3) /* Library is shutting down */ 1002 #define KMP_GTID_MONITOR (-4) /* Monitor thread ID */ 1003 #define KMP_GTID_UNKNOWN (-5) /* Is not known */ 1004 #define KMP_GTID_MIN (-6) /* Minimal gtid for low bound check in DEBUG */ 1005 1006 /* OpenMP 5.0 Memory Management support */ 1007 1008 #ifndef __OMP_H 1009 // Duplicate type definitions from omp.h 1010 typedef uintptr_t omp_uintptr_t; 1011 1012 typedef enum { 1013 omp_atk_sync_hint = 1, 1014 omp_atk_alignment = 2, 1015 omp_atk_access = 3, 1016 omp_atk_pool_size = 4, 1017 omp_atk_fallback = 5, 1018 omp_atk_fb_data = 6, 1019 omp_atk_pinned = 7, 1020 omp_atk_partition = 8 1021 } omp_alloctrait_key_t; 1022 1023 typedef enum { 1024 omp_atv_false = 0, 1025 omp_atv_true = 1, 1026 omp_atv_contended = 3, 1027 omp_atv_uncontended = 4, 1028 omp_atv_serialized = 5, 1029 omp_atv_sequential = omp_atv_serialized, // (deprecated) 1030 omp_atv_private = 6, 1031 omp_atv_all = 7, 1032 omp_atv_thread = 8, 1033 omp_atv_pteam = 9, 1034 omp_atv_cgroup = 10, 1035 omp_atv_default_mem_fb = 11, 1036 omp_atv_null_fb = 12, 1037 omp_atv_abort_fb = 13, 1038 omp_atv_allocator_fb = 14, 1039 omp_atv_environment = 15, 1040 omp_atv_nearest = 16, 1041 omp_atv_blocked = 17, 1042 omp_atv_interleaved = 18 1043 } omp_alloctrait_value_t; 1044 #define omp_atv_default ((omp_uintptr_t)-1) 1045 1046 typedef void *omp_memspace_handle_t; 1047 extern omp_memspace_handle_t const omp_default_mem_space; 1048 extern omp_memspace_handle_t const omp_large_cap_mem_space; 1049 extern omp_memspace_handle_t const omp_const_mem_space; 1050 extern omp_memspace_handle_t const omp_high_bw_mem_space; 1051 extern omp_memspace_handle_t const omp_low_lat_mem_space; 1052 extern omp_memspace_handle_t const llvm_omp_target_host_mem_space; 1053 extern omp_memspace_handle_t const llvm_omp_target_shared_mem_space; 1054 extern omp_memspace_handle_t const llvm_omp_target_device_mem_space; 1055 1056 typedef struct { 1057 omp_alloctrait_key_t key; 1058 omp_uintptr_t value; 1059 } omp_alloctrait_t; 1060 1061 typedef void *omp_allocator_handle_t; 1062 extern omp_allocator_handle_t const omp_null_allocator; 1063 extern omp_allocator_handle_t const omp_default_mem_alloc; 1064 extern omp_allocator_handle_t const omp_large_cap_mem_alloc; 1065 extern omp_allocator_handle_t const omp_const_mem_alloc; 1066 extern omp_allocator_handle_t const omp_high_bw_mem_alloc; 1067 extern omp_allocator_handle_t const omp_low_lat_mem_alloc; 1068 extern omp_allocator_handle_t const omp_cgroup_mem_alloc; 1069 extern omp_allocator_handle_t const omp_pteam_mem_alloc; 1070 extern omp_allocator_handle_t const omp_thread_mem_alloc; 1071 extern omp_allocator_handle_t const llvm_omp_target_host_mem_alloc; 1072 extern omp_allocator_handle_t const llvm_omp_target_shared_mem_alloc; 1073 extern omp_allocator_handle_t const llvm_omp_target_device_mem_alloc; 1074 extern omp_allocator_handle_t const kmp_max_mem_alloc; 1075 extern omp_allocator_handle_t __kmp_def_allocator; 1076 1077 // end of duplicate type definitions from omp.h 1078 #endif 1079 1080 extern int __kmp_memkind_available; 1081 1082 typedef omp_memspace_handle_t kmp_memspace_t; // placeholder 1083 1084 typedef struct kmp_allocator_t { 1085 omp_memspace_handle_t memspace; 1086 void **memkind; // pointer to memkind 1087 size_t alignment; 1088 omp_alloctrait_value_t fb; 1089 kmp_allocator_t *fb_data; 1090 kmp_uint64 pool_size; 1091 kmp_uint64 pool_used; 1092 bool pinned; 1093 } kmp_allocator_t; 1094 1095 extern omp_allocator_handle_t __kmpc_init_allocator(int gtid, 1096 omp_memspace_handle_t, 1097 int ntraits, 1098 omp_alloctrait_t traits[]); 1099 extern void __kmpc_destroy_allocator(int gtid, omp_allocator_handle_t al); 1100 extern void __kmpc_set_default_allocator(int gtid, omp_allocator_handle_t al); 1101 extern omp_allocator_handle_t __kmpc_get_default_allocator(int gtid); 1102 // external interfaces, may be used by compiler 1103 extern void *__kmpc_alloc(int gtid, size_t sz, omp_allocator_handle_t al); 1104 extern void *__kmpc_aligned_alloc(int gtid, size_t align, size_t sz, 1105 omp_allocator_handle_t al); 1106 extern void *__kmpc_calloc(int gtid, size_t nmemb, size_t sz, 1107 omp_allocator_handle_t al); 1108 extern void *__kmpc_realloc(int gtid, void *ptr, size_t sz, 1109 omp_allocator_handle_t al, 1110 omp_allocator_handle_t free_al); 1111 extern void __kmpc_free(int gtid, void *ptr, omp_allocator_handle_t al); 1112 // internal interfaces, contain real implementation 1113 extern void *__kmp_alloc(int gtid, size_t align, size_t sz, 1114 omp_allocator_handle_t al); 1115 extern void *__kmp_calloc(int gtid, size_t align, size_t nmemb, size_t sz, 1116 omp_allocator_handle_t al); 1117 extern void *__kmp_realloc(int gtid, void *ptr, size_t sz, 1118 omp_allocator_handle_t al, 1119 omp_allocator_handle_t free_al); 1120 extern void ___kmpc_free(int gtid, void *ptr, omp_allocator_handle_t al); 1121 1122 extern void __kmp_init_memkind(); 1123 extern void __kmp_fini_memkind(); 1124 extern void __kmp_init_target_mem(); 1125 1126 /* ------------------------------------------------------------------------ */ 1127 1128 #if ENABLE_LIBOMPTARGET 1129 extern void __kmp_init_target_task(); 1130 #endif 1131 1132 /* ------------------------------------------------------------------------ */ 1133 1134 #define KMP_UINT64_MAX \ 1135 (~((kmp_uint64)1 << ((sizeof(kmp_uint64) * (1 << 3)) - 1))) 1136 1137 #define KMP_MIN_NTH 1 1138 1139 #ifndef KMP_MAX_NTH 1140 #if defined(PTHREAD_THREADS_MAX) && PTHREAD_THREADS_MAX < INT_MAX 1141 #define KMP_MAX_NTH PTHREAD_THREADS_MAX 1142 #else 1143 #define KMP_MAX_NTH INT_MAX 1144 #endif 1145 #endif /* KMP_MAX_NTH */ 1146 1147 #ifdef PTHREAD_STACK_MIN 1148 #define KMP_MIN_STKSIZE ((size_t)PTHREAD_STACK_MIN) 1149 #else 1150 #define KMP_MIN_STKSIZE ((size_t)(32 * 1024)) 1151 #endif 1152 1153 #define KMP_MAX_STKSIZE (~((size_t)1 << ((sizeof(size_t) * (1 << 3)) - 1))) 1154 1155 #if KMP_ARCH_X86 1156 #define KMP_DEFAULT_STKSIZE ((size_t)(2 * 1024 * 1024)) 1157 #elif KMP_ARCH_X86_64 1158 #define KMP_DEFAULT_STKSIZE ((size_t)(4 * 1024 * 1024)) 1159 #define KMP_BACKUP_STKSIZE ((size_t)(2 * 1024 * 1024)) 1160 #else 1161 #define KMP_DEFAULT_STKSIZE ((size_t)(1024 * 1024)) 1162 #endif 1163 1164 #define KMP_DEFAULT_MALLOC_POOL_INCR ((size_t)(1024 * 1024)) 1165 #define KMP_MIN_MALLOC_POOL_INCR ((size_t)(4 * 1024)) 1166 #define KMP_MAX_MALLOC_POOL_INCR \ 1167 (~((size_t)1 << ((sizeof(size_t) * (1 << 3)) - 1))) 1168 1169 #define KMP_MIN_STKOFFSET (0) 1170 #define KMP_MAX_STKOFFSET KMP_MAX_STKSIZE 1171 #if KMP_OS_DARWIN 1172 #define KMP_DEFAULT_STKOFFSET KMP_MIN_STKOFFSET 1173 #else 1174 #define KMP_DEFAULT_STKOFFSET CACHE_LINE 1175 #endif 1176 1177 #define KMP_MIN_STKPADDING (0) 1178 #define KMP_MAX_STKPADDING (2 * 1024 * 1024) 1179 1180 #define KMP_BLOCKTIME_MULTIPLIER \ 1181 (1000) /* number of blocktime units per second */ 1182 #define KMP_MIN_BLOCKTIME (0) 1183 #define KMP_MAX_BLOCKTIME \ 1184 (INT_MAX) /* Must be this for "infinite" setting the work */ 1185 1186 /* __kmp_blocktime is in milliseconds */ 1187 #define KMP_DEFAULT_BLOCKTIME (__kmp_is_hybrid_cpu() ? (0) : (200)) 1188 1189 #if KMP_USE_MONITOR 1190 #define KMP_DEFAULT_MONITOR_STKSIZE ((size_t)(64 * 1024)) 1191 #define KMP_MIN_MONITOR_WAKEUPS (1) // min times monitor wakes up per second 1192 #define KMP_MAX_MONITOR_WAKEUPS (1000) // max times monitor can wake up per sec 1193 1194 /* Calculate new number of monitor wakeups for a specific block time based on 1195 previous monitor_wakeups. Only allow increasing number of wakeups */ 1196 #define KMP_WAKEUPS_FROM_BLOCKTIME(blocktime, monitor_wakeups) \ 1197 (((blocktime) == KMP_MAX_BLOCKTIME) ? (monitor_wakeups) \ 1198 : ((blocktime) == KMP_MIN_BLOCKTIME) ? KMP_MAX_MONITOR_WAKEUPS \ 1199 : ((monitor_wakeups) > (KMP_BLOCKTIME_MULTIPLIER / (blocktime))) \ 1200 ? (monitor_wakeups) \ 1201 : (KMP_BLOCKTIME_MULTIPLIER) / (blocktime)) 1202 1203 /* Calculate number of intervals for a specific block time based on 1204 monitor_wakeups */ 1205 #define KMP_INTERVALS_FROM_BLOCKTIME(blocktime, monitor_wakeups) \ 1206 (((blocktime) + (KMP_BLOCKTIME_MULTIPLIER / (monitor_wakeups)) - 1) / \ 1207 (KMP_BLOCKTIME_MULTIPLIER / (monitor_wakeups))) 1208 #else 1209 #define KMP_BLOCKTIME(team, tid) \ 1210 (get__bt_set(team, tid) ? get__blocktime(team, tid) : __kmp_dflt_blocktime) 1211 #if KMP_OS_UNIX && (KMP_ARCH_X86 || KMP_ARCH_X86_64) 1212 // HW TSC is used to reduce overhead (clock tick instead of nanosecond). 1213 extern kmp_uint64 __kmp_ticks_per_msec; 1214 #if KMP_COMPILER_ICC || KMP_COMPILER_ICX 1215 #define KMP_NOW() ((kmp_uint64)_rdtsc()) 1216 #else 1217 #define KMP_NOW() __kmp_hardware_timestamp() 1218 #endif 1219 #define KMP_NOW_MSEC() (KMP_NOW() / __kmp_ticks_per_msec) 1220 #define KMP_BLOCKTIME_INTERVAL(team, tid) \ 1221 (KMP_BLOCKTIME(team, tid) * __kmp_ticks_per_msec) 1222 #define KMP_BLOCKING(goal, count) ((goal) > KMP_NOW()) 1223 #else 1224 // System time is retrieved sporadically while blocking. 1225 extern kmp_uint64 __kmp_now_nsec(); 1226 #define KMP_NOW() __kmp_now_nsec() 1227 #define KMP_NOW_MSEC() (KMP_NOW() / KMP_USEC_PER_SEC) 1228 #define KMP_BLOCKTIME_INTERVAL(team, tid) \ 1229 (KMP_BLOCKTIME(team, tid) * KMP_USEC_PER_SEC) 1230 #define KMP_BLOCKING(goal, count) ((count) % 1000 != 0 || (goal) > KMP_NOW()) 1231 #endif 1232 #endif // KMP_USE_MONITOR 1233 1234 #define KMP_MIN_STATSCOLS 40 1235 #define KMP_MAX_STATSCOLS 4096 1236 #define KMP_DEFAULT_STATSCOLS 80 1237 1238 #define KMP_MIN_INTERVAL 0 1239 #define KMP_MAX_INTERVAL (INT_MAX - 1) 1240 #define KMP_DEFAULT_INTERVAL 0 1241 1242 #define KMP_MIN_CHUNK 1 1243 #define KMP_MAX_CHUNK (INT_MAX - 1) 1244 #define KMP_DEFAULT_CHUNK 1 1245 1246 #define KMP_MIN_DISP_NUM_BUFF 1 1247 #define KMP_DFLT_DISP_NUM_BUFF 7 1248 #define KMP_MAX_DISP_NUM_BUFF 4096 1249 1250 #define KMP_MAX_ORDERED 8 1251 1252 #define KMP_MAX_FIELDS 32 1253 1254 #define KMP_MAX_BRANCH_BITS 31 1255 1256 #define KMP_MAX_ACTIVE_LEVELS_LIMIT INT_MAX 1257 1258 #define KMP_MAX_DEFAULT_DEVICE_LIMIT INT_MAX 1259 1260 #define KMP_MAX_TASK_PRIORITY_LIMIT INT_MAX 1261 1262 /* Minimum number of threads before switch to TLS gtid (experimentally 1263 determined) */ 1264 /* josh TODO: what about OS X* tuning? */ 1265 #if KMP_ARCH_X86 || KMP_ARCH_X86_64 1266 #define KMP_TLS_GTID_MIN 5 1267 #else 1268 #define KMP_TLS_GTID_MIN INT_MAX 1269 #endif 1270 1271 #define KMP_MASTER_TID(tid) (0 == (tid)) 1272 #define KMP_WORKER_TID(tid) (0 != (tid)) 1273 1274 #define KMP_MASTER_GTID(gtid) (0 == __kmp_tid_from_gtid((gtid))) 1275 #define KMP_WORKER_GTID(gtid) (0 != __kmp_tid_from_gtid((gtid))) 1276 #define KMP_INITIAL_GTID(gtid) (0 == (gtid)) 1277 1278 #ifndef TRUE 1279 #define FALSE 0 1280 #define TRUE (!FALSE) 1281 #endif 1282 1283 /* NOTE: all of the following constants must be even */ 1284 1285 #if KMP_OS_WINDOWS 1286 #define KMP_INIT_WAIT 64U /* initial number of spin-tests */ 1287 #define KMP_NEXT_WAIT 32U /* susequent number of spin-tests */ 1288 #elif KMP_OS_LINUX 1289 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */ 1290 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */ 1291 #elif KMP_OS_DARWIN 1292 /* TODO: tune for KMP_OS_DARWIN */ 1293 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */ 1294 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */ 1295 #elif KMP_OS_DRAGONFLY 1296 /* TODO: tune for KMP_OS_DRAGONFLY */ 1297 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */ 1298 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */ 1299 #elif KMP_OS_FREEBSD 1300 /* TODO: tune for KMP_OS_FREEBSD */ 1301 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */ 1302 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */ 1303 #elif KMP_OS_NETBSD 1304 /* TODO: tune for KMP_OS_NETBSD */ 1305 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */ 1306 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */ 1307 #elif KMP_OS_HURD 1308 /* TODO: tune for KMP_OS_HURD */ 1309 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */ 1310 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */ 1311 #elif KMP_OS_OPENBSD 1312 /* TODO: tune for KMP_OS_OPENBSD */ 1313 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */ 1314 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */ 1315 #endif 1316 1317 #if KMP_ARCH_X86 || KMP_ARCH_X86_64 1318 typedef struct kmp_cpuid { 1319 kmp_uint32 eax; 1320 kmp_uint32 ebx; 1321 kmp_uint32 ecx; 1322 kmp_uint32 edx; 1323 } kmp_cpuid_t; 1324 1325 typedef struct kmp_cpuinfo_flags_t { 1326 unsigned sse2 : 1; // 0 if SSE2 instructions are not supported, 1 otherwise. 1327 unsigned rtm : 1; // 0 if RTM instructions are not supported, 1 otherwise. 1328 unsigned hybrid : 1; 1329 unsigned reserved : 29; // Ensure size of 32 bits 1330 } kmp_cpuinfo_flags_t; 1331 1332 typedef struct kmp_cpuinfo { 1333 int initialized; // If 0, other fields are not initialized. 1334 int signature; // CPUID(1).EAX 1335 int family; // CPUID(1).EAX[27:20]+CPUID(1).EAX[11:8] (Extended Family+Family) 1336 int model; // ( CPUID(1).EAX[19:16] << 4 ) + CPUID(1).EAX[7:4] ( ( Extended 1337 // Model << 4 ) + Model) 1338 int stepping; // CPUID(1).EAX[3:0] ( Stepping ) 1339 kmp_cpuinfo_flags_t flags; 1340 int apic_id; 1341 int physical_id; 1342 int logical_id; 1343 kmp_uint64 frequency; // Nominal CPU frequency in Hz. 1344 char name[3 * sizeof(kmp_cpuid_t)]; // CPUID(0x80000002,0x80000003,0x80000004) 1345 } kmp_cpuinfo_t; 1346 1347 extern void __kmp_query_cpuid(kmp_cpuinfo_t *p); 1348 1349 #if KMP_OS_UNIX 1350 // subleaf is only needed for cache and topology discovery and can be set to 1351 // zero in most cases 1352 static inline void __kmp_x86_cpuid(int leaf, int subleaf, struct kmp_cpuid *p) { 1353 __asm__ __volatile__("cpuid" 1354 : "=a"(p->eax), "=b"(p->ebx), "=c"(p->ecx), "=d"(p->edx) 1355 : "a"(leaf), "c"(subleaf)); 1356 } 1357 // Load p into FPU control word 1358 static inline void __kmp_load_x87_fpu_control_word(const kmp_int16 *p) { 1359 __asm__ __volatile__("fldcw %0" : : "m"(*p)); 1360 } 1361 // Store FPU control word into p 1362 static inline void __kmp_store_x87_fpu_control_word(kmp_int16 *p) { 1363 __asm__ __volatile__("fstcw %0" : "=m"(*p)); 1364 } 1365 static inline void __kmp_clear_x87_fpu_status_word() { 1366 #if KMP_MIC 1367 // 32-bit protected mode x87 FPU state 1368 struct x87_fpu_state { 1369 unsigned cw; 1370 unsigned sw; 1371 unsigned tw; 1372 unsigned fip; 1373 unsigned fips; 1374 unsigned fdp; 1375 unsigned fds; 1376 }; 1377 struct x87_fpu_state fpu_state = {0, 0, 0, 0, 0, 0, 0}; 1378 __asm__ __volatile__("fstenv %0\n\t" // store FP env 1379 "andw $0x7f00, %1\n\t" // clear 0-7,15 bits of FP SW 1380 "fldenv %0\n\t" // load FP env back 1381 : "+m"(fpu_state), "+m"(fpu_state.sw)); 1382 #else 1383 __asm__ __volatile__("fnclex"); 1384 #endif // KMP_MIC 1385 } 1386 #if __SSE__ 1387 static inline void __kmp_load_mxcsr(const kmp_uint32 *p) { _mm_setcsr(*p); } 1388 static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = _mm_getcsr(); } 1389 #else 1390 static inline void __kmp_load_mxcsr(const kmp_uint32 *p) {} 1391 static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = 0; } 1392 #endif 1393 #else 1394 // Windows still has these as external functions in assembly file 1395 extern void __kmp_x86_cpuid(int mode, int mode2, struct kmp_cpuid *p); 1396 extern void __kmp_load_x87_fpu_control_word(const kmp_int16 *p); 1397 extern void __kmp_store_x87_fpu_control_word(kmp_int16 *p); 1398 extern void __kmp_clear_x87_fpu_status_word(); 1399 static inline void __kmp_load_mxcsr(const kmp_uint32 *p) { _mm_setcsr(*p); } 1400 static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = _mm_getcsr(); } 1401 #endif // KMP_OS_UNIX 1402 1403 #define KMP_X86_MXCSR_MASK 0xffffffc0 /* ignore status flags (6 lsb) */ 1404 1405 // User-level Monitor/Mwait 1406 #if KMP_HAVE_UMWAIT 1407 // We always try for UMWAIT first 1408 #if KMP_HAVE_WAITPKG_INTRINSICS 1409 #if KMP_HAVE_IMMINTRIN_H 1410 #include <immintrin.h> 1411 #elif KMP_HAVE_INTRIN_H 1412 #include <intrin.h> 1413 #endif 1414 #endif // KMP_HAVE_WAITPKG_INTRINSICS 1415 1416 KMP_ATTRIBUTE_TARGET_WAITPKG 1417 static inline int __kmp_tpause(uint32_t hint, uint64_t counter) { 1418 #if !KMP_HAVE_WAITPKG_INTRINSICS 1419 uint32_t timeHi = uint32_t(counter >> 32); 1420 uint32_t timeLo = uint32_t(counter & 0xffffffff); 1421 char flag; 1422 __asm__ volatile("#tpause\n.byte 0x66, 0x0F, 0xAE, 0xF1\n" 1423 "setb %0" 1424 // The "=q" restraint means any register accessible as rl 1425 // in 32-bit mode: a, b, c, and d; 1426 // in 64-bit mode: any integer register 1427 : "=q"(flag) 1428 : "a"(timeLo), "d"(timeHi), "c"(hint) 1429 :); 1430 return flag; 1431 #else 1432 return _tpause(hint, counter); 1433 #endif 1434 } 1435 KMP_ATTRIBUTE_TARGET_WAITPKG 1436 static inline void __kmp_umonitor(void *cacheline) { 1437 #if !KMP_HAVE_WAITPKG_INTRINSICS 1438 __asm__ volatile("# umonitor\n.byte 0xF3, 0x0F, 0xAE, 0x01 " 1439 : 1440 : "a"(cacheline) 1441 :); 1442 #else 1443 _umonitor(cacheline); 1444 #endif 1445 } 1446 KMP_ATTRIBUTE_TARGET_WAITPKG 1447 static inline int __kmp_umwait(uint32_t hint, uint64_t counter) { 1448 #if !KMP_HAVE_WAITPKG_INTRINSICS 1449 uint32_t timeHi = uint32_t(counter >> 32); 1450 uint32_t timeLo = uint32_t(counter & 0xffffffff); 1451 char flag; 1452 __asm__ volatile("#umwait\n.byte 0xF2, 0x0F, 0xAE, 0xF1\n" 1453 "setb %0" 1454 // The "=q" restraint means any register accessible as rl 1455 // in 32-bit mode: a, b, c, and d; 1456 // in 64-bit mode: any integer register 1457 : "=q"(flag) 1458 : "a"(timeLo), "d"(timeHi), "c"(hint) 1459 :); 1460 return flag; 1461 #else 1462 return _umwait(hint, counter); 1463 #endif 1464 } 1465 #elif KMP_HAVE_MWAIT 1466 #if KMP_OS_UNIX 1467 #include <pmmintrin.h> 1468 #else 1469 #include <intrin.h> 1470 #endif 1471 #if KMP_OS_UNIX 1472 __attribute__((target("sse3"))) 1473 #endif 1474 static inline void 1475 __kmp_mm_monitor(void *cacheline, unsigned extensions, unsigned hints) { 1476 _mm_monitor(cacheline, extensions, hints); 1477 } 1478 #if KMP_OS_UNIX 1479 __attribute__((target("sse3"))) 1480 #endif 1481 static inline void 1482 __kmp_mm_mwait(unsigned extensions, unsigned hints) { 1483 _mm_mwait(extensions, hints); 1484 } 1485 #endif // KMP_HAVE_UMWAIT 1486 1487 #if KMP_ARCH_X86 1488 extern void __kmp_x86_pause(void); 1489 #elif KMP_MIC 1490 // Performance testing on KNC (C0QS-7120 P/A/X/D, 61-core, 16 GB Memory) showed 1491 // regression after removal of extra PAUSE from spin loops. Changing 1492 // the delay from 100 to 300 showed even better performance than double PAUSE 1493 // on Spec OMP2001 and LCPC tasking tests, no regressions on EPCC. 1494 static inline void __kmp_x86_pause(void) { _mm_delay_32(300); } 1495 #else 1496 static inline void __kmp_x86_pause(void) { _mm_pause(); } 1497 #endif 1498 #define KMP_CPU_PAUSE() __kmp_x86_pause() 1499 #elif KMP_ARCH_PPC64 1500 #define KMP_PPC64_PRI_LOW() __asm__ volatile("or 1, 1, 1") 1501 #define KMP_PPC64_PRI_MED() __asm__ volatile("or 2, 2, 2") 1502 #define KMP_PPC64_PRI_LOC_MB() __asm__ volatile("" : : : "memory") 1503 #define KMP_CPU_PAUSE() \ 1504 do { \ 1505 KMP_PPC64_PRI_LOW(); \ 1506 KMP_PPC64_PRI_MED(); \ 1507 KMP_PPC64_PRI_LOC_MB(); \ 1508 } while (0) 1509 #else 1510 #define KMP_CPU_PAUSE() /* nothing to do */ 1511 #endif 1512 1513 #define KMP_INIT_YIELD(count) \ 1514 { (count) = __kmp_yield_init; } 1515 1516 #define KMP_INIT_BACKOFF(time) \ 1517 { (time) = __kmp_pause_init; } 1518 1519 #define KMP_OVERSUBSCRIBED \ 1520 (TCR_4(__kmp_nth) > (__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc)) 1521 1522 #define KMP_TRY_YIELD \ 1523 ((__kmp_use_yield == 1) || (__kmp_use_yield == 2 && (KMP_OVERSUBSCRIBED))) 1524 1525 #define KMP_TRY_YIELD_OVERSUB \ 1526 ((__kmp_use_yield == 1 || __kmp_use_yield == 2) && (KMP_OVERSUBSCRIBED)) 1527 1528 #define KMP_YIELD(cond) \ 1529 { \ 1530 KMP_CPU_PAUSE(); \ 1531 if ((cond) && (KMP_TRY_YIELD)) \ 1532 __kmp_yield(); \ 1533 } 1534 1535 #define KMP_YIELD_OVERSUB() \ 1536 { \ 1537 KMP_CPU_PAUSE(); \ 1538 if ((KMP_TRY_YIELD_OVERSUB)) \ 1539 __kmp_yield(); \ 1540 } 1541 1542 // Note the decrement of 2 in the following Macros. With KMP_LIBRARY=turnaround, 1543 // there should be no yielding since initial value from KMP_INIT_YIELD() is odd. 1544 #define KMP_YIELD_SPIN(count) \ 1545 { \ 1546 KMP_CPU_PAUSE(); \ 1547 if (KMP_TRY_YIELD) { \ 1548 (count) -= 2; \ 1549 if (!(count)) { \ 1550 __kmp_yield(); \ 1551 (count) = __kmp_yield_next; \ 1552 } \ 1553 } \ 1554 } 1555 1556 // If TPAUSE is available & enabled, use it. If oversubscribed, use the slower 1557 // (C0.2) state, which improves performance of other SMT threads on the same 1558 // core, otherwise, use the fast (C0.1) default state, or whatever the user has 1559 // requested. Uses a timed TPAUSE, and exponential backoff. If TPAUSE isn't 1560 // available, fall back to the regular CPU pause and yield combination. 1561 #if KMP_HAVE_UMWAIT 1562 #define KMP_TPAUSE_MAX_MASK ((kmp_uint64)0xFFFF) 1563 #define KMP_YIELD_OVERSUB_ELSE_SPIN(count, time) \ 1564 { \ 1565 if (__kmp_tpause_enabled) { \ 1566 if (KMP_OVERSUBSCRIBED) { \ 1567 __kmp_tpause(0, (time)); \ 1568 } else { \ 1569 __kmp_tpause(__kmp_tpause_hint, (time)); \ 1570 } \ 1571 (time) = (time << 1 | 1) & KMP_TPAUSE_MAX_MASK; \ 1572 } else { \ 1573 KMP_CPU_PAUSE(); \ 1574 if ((KMP_TRY_YIELD_OVERSUB)) { \ 1575 __kmp_yield(); \ 1576 } else if (__kmp_use_yield == 1) { \ 1577 (count) -= 2; \ 1578 if (!(count)) { \ 1579 __kmp_yield(); \ 1580 (count) = __kmp_yield_next; \ 1581 } \ 1582 } \ 1583 } \ 1584 } 1585 #else 1586 #define KMP_YIELD_OVERSUB_ELSE_SPIN(count, time) \ 1587 { \ 1588 KMP_CPU_PAUSE(); \ 1589 if ((KMP_TRY_YIELD_OVERSUB)) \ 1590 __kmp_yield(); \ 1591 else if (__kmp_use_yield == 1) { \ 1592 (count) -= 2; \ 1593 if (!(count)) { \ 1594 __kmp_yield(); \ 1595 (count) = __kmp_yield_next; \ 1596 } \ 1597 } \ 1598 } 1599 #endif // KMP_HAVE_UMWAIT 1600 1601 /* ------------------------------------------------------------------------ */ 1602 /* Support datatypes for the orphaned construct nesting checks. */ 1603 /* ------------------------------------------------------------------------ */ 1604 1605 /* When adding to this enum, add its corresponding string in cons_text_c[] 1606 * array in kmp_error.cpp */ 1607 enum cons_type { 1608 ct_none, 1609 ct_parallel, 1610 ct_pdo, 1611 ct_pdo_ordered, 1612 ct_psections, 1613 ct_psingle, 1614 ct_critical, 1615 ct_ordered_in_parallel, 1616 ct_ordered_in_pdo, 1617 ct_master, 1618 ct_reduce, 1619 ct_barrier, 1620 ct_masked 1621 }; 1622 1623 #define IS_CONS_TYPE_ORDERED(ct) ((ct) == ct_pdo_ordered) 1624 1625 struct cons_data { 1626 ident_t const *ident; 1627 enum cons_type type; 1628 int prev; 1629 kmp_user_lock_p 1630 name; /* address exclusively for critical section name comparison */ 1631 }; 1632 1633 struct cons_header { 1634 int p_top, w_top, s_top; 1635 int stack_size, stack_top; 1636 struct cons_data *stack_data; 1637 }; 1638 1639 struct kmp_region_info { 1640 char *text; 1641 int offset[KMP_MAX_FIELDS]; 1642 int length[KMP_MAX_FIELDS]; 1643 }; 1644 1645 /* ---------------------------------------------------------------------- */ 1646 /* ---------------------------------------------------------------------- */ 1647 1648 #if KMP_OS_WINDOWS 1649 typedef HANDLE kmp_thread_t; 1650 typedef DWORD kmp_key_t; 1651 #endif /* KMP_OS_WINDOWS */ 1652 1653 #if KMP_OS_UNIX 1654 typedef pthread_t kmp_thread_t; 1655 typedef pthread_key_t kmp_key_t; 1656 #endif 1657 1658 extern kmp_key_t __kmp_gtid_threadprivate_key; 1659 1660 typedef struct kmp_sys_info { 1661 long maxrss; /* the maximum resident set size utilized (in kilobytes) */ 1662 long minflt; /* the number of page faults serviced without any I/O */ 1663 long majflt; /* the number of page faults serviced that required I/O */ 1664 long nswap; /* the number of times a process was "swapped" out of memory */ 1665 long inblock; /* the number of times the file system had to perform input */ 1666 long oublock; /* the number of times the file system had to perform output */ 1667 long nvcsw; /* the number of times a context switch was voluntarily */ 1668 long nivcsw; /* the number of times a context switch was forced */ 1669 } kmp_sys_info_t; 1670 1671 #if USE_ITT_BUILD 1672 // We cannot include "kmp_itt.h" due to circular dependency. Declare the only 1673 // required type here. Later we will check the type meets requirements. 1674 typedef int kmp_itt_mark_t; 1675 #define KMP_ITT_DEBUG 0 1676 #endif /* USE_ITT_BUILD */ 1677 1678 typedef kmp_int32 kmp_critical_name[8]; 1679 1680 /*! 1681 @ingroup PARALLEL 1682 The type for a microtask which gets passed to @ref __kmpc_fork_call(). 1683 The arguments to the outlined function are 1684 @param global_tid the global thread identity of the thread executing the 1685 function. 1686 @param bound_tid the local identity of the thread executing the function 1687 @param ... pointers to shared variables accessed by the function. 1688 */ 1689 typedef void (*kmpc_micro)(kmp_int32 *global_tid, kmp_int32 *bound_tid, ...); 1690 typedef void (*kmpc_micro_bound)(kmp_int32 *bound_tid, kmp_int32 *bound_nth, 1691 ...); 1692 1693 /*! 1694 @ingroup THREADPRIVATE 1695 @{ 1696 */ 1697 /* --------------------------------------------------------------------------- 1698 */ 1699 /* Threadprivate initialization/finalization function declarations */ 1700 1701 /* for non-array objects: __kmpc_threadprivate_register() */ 1702 1703 /*! 1704 Pointer to the constructor function. 1705 The first argument is the <tt>this</tt> pointer 1706 */ 1707 typedef void *(*kmpc_ctor)(void *); 1708 1709 /*! 1710 Pointer to the destructor function. 1711 The first argument is the <tt>this</tt> pointer 1712 */ 1713 typedef void (*kmpc_dtor)( 1714 void * /*, size_t */); /* 2nd arg: magic number for KCC unused by Intel 1715 compiler */ 1716 /*! 1717 Pointer to an alternate constructor. 1718 The first argument is the <tt>this</tt> pointer. 1719 */ 1720 typedef void *(*kmpc_cctor)(void *, void *); 1721 1722 /* for array objects: __kmpc_threadprivate_register_vec() */ 1723 /* First arg: "this" pointer */ 1724 /* Last arg: number of array elements */ 1725 /*! 1726 Array constructor. 1727 First argument is the <tt>this</tt> pointer 1728 Second argument the number of array elements. 1729 */ 1730 typedef void *(*kmpc_ctor_vec)(void *, size_t); 1731 /*! 1732 Pointer to the array destructor function. 1733 The first argument is the <tt>this</tt> pointer 1734 Second argument the number of array elements. 1735 */ 1736 typedef void (*kmpc_dtor_vec)(void *, size_t); 1737 /*! 1738 Array constructor. 1739 First argument is the <tt>this</tt> pointer 1740 Third argument the number of array elements. 1741 */ 1742 typedef void *(*kmpc_cctor_vec)(void *, void *, 1743 size_t); /* function unused by compiler */ 1744 1745 /*! 1746 @} 1747 */ 1748 1749 /* keeps tracked of threadprivate cache allocations for cleanup later */ 1750 typedef struct kmp_cached_addr { 1751 void **addr; /* address of allocated cache */ 1752 void ***compiler_cache; /* pointer to compiler's cache */ 1753 void *data; /* pointer to global data */ 1754 struct kmp_cached_addr *next; /* pointer to next cached address */ 1755 } kmp_cached_addr_t; 1756 1757 struct private_data { 1758 struct private_data *next; /* The next descriptor in the list */ 1759 void *data; /* The data buffer for this descriptor */ 1760 int more; /* The repeat count for this descriptor */ 1761 size_t size; /* The data size for this descriptor */ 1762 }; 1763 1764 struct private_common { 1765 struct private_common *next; 1766 struct private_common *link; 1767 void *gbl_addr; 1768 void *par_addr; /* par_addr == gbl_addr for PRIMARY thread */ 1769 size_t cmn_size; 1770 }; 1771 1772 struct shared_common { 1773 struct shared_common *next; 1774 struct private_data *pod_init; 1775 void *obj_init; 1776 void *gbl_addr; 1777 union { 1778 kmpc_ctor ctor; 1779 kmpc_ctor_vec ctorv; 1780 } ct; 1781 union { 1782 kmpc_cctor cctor; 1783 kmpc_cctor_vec cctorv; 1784 } cct; 1785 union { 1786 kmpc_dtor dtor; 1787 kmpc_dtor_vec dtorv; 1788 } dt; 1789 size_t vec_len; 1790 int is_vec; 1791 size_t cmn_size; 1792 }; 1793 1794 #define KMP_HASH_TABLE_LOG2 9 /* log2 of the hash table size */ 1795 #define KMP_HASH_TABLE_SIZE \ 1796 (1 << KMP_HASH_TABLE_LOG2) /* size of the hash table */ 1797 #define KMP_HASH_SHIFT 3 /* throw away this many low bits from the address */ 1798 #define KMP_HASH(x) \ 1799 ((((kmp_uintptr_t)x) >> KMP_HASH_SHIFT) & (KMP_HASH_TABLE_SIZE - 1)) 1800 1801 struct common_table { 1802 struct private_common *data[KMP_HASH_TABLE_SIZE]; 1803 }; 1804 1805 struct shared_table { 1806 struct shared_common *data[KMP_HASH_TABLE_SIZE]; 1807 }; 1808 1809 /* ------------------------------------------------------------------------ */ 1810 1811 #if KMP_USE_HIER_SCHED 1812 // Shared barrier data that exists inside a single unit of the scheduling 1813 // hierarchy 1814 typedef struct kmp_hier_private_bdata_t { 1815 kmp_int32 num_active; 1816 kmp_uint64 index; 1817 kmp_uint64 wait_val[2]; 1818 } kmp_hier_private_bdata_t; 1819 #endif 1820 1821 typedef struct kmp_sched_flags { 1822 unsigned ordered : 1; 1823 unsigned nomerge : 1; 1824 unsigned contains_last : 1; 1825 #if KMP_USE_HIER_SCHED 1826 unsigned use_hier : 1; 1827 unsigned unused : 28; 1828 #else 1829 unsigned unused : 29; 1830 #endif 1831 } kmp_sched_flags_t; 1832 1833 KMP_BUILD_ASSERT(sizeof(kmp_sched_flags_t) == 4); 1834 1835 #if KMP_STATIC_STEAL_ENABLED 1836 typedef struct KMP_ALIGN_CACHE dispatch_private_info32 { 1837 kmp_int32 count; 1838 kmp_int32 ub; 1839 /* Adding KMP_ALIGN_CACHE here doesn't help / can hurt performance */ 1840 kmp_int32 lb; 1841 kmp_int32 st; 1842 kmp_int32 tc; 1843 kmp_lock_t *steal_lock; // lock used for chunk stealing 1844 // KMP_ALIGN(32) ensures (if the KMP_ALIGN macro is turned on) 1845 // a) parm3 is properly aligned and 1846 // b) all parm1-4 are on the same cache line. 1847 // Because of parm1-4 are used together, performance seems to be better 1848 // if they are on the same cache line (not measured though). 1849 1850 struct KMP_ALIGN(32) { // AC: changed 16 to 32 in order to simplify template 1851 kmp_int32 parm1; // structures in kmp_dispatch.cpp. This should 1852 kmp_int32 parm2; // make no real change at least while padding is off. 1853 kmp_int32 parm3; 1854 kmp_int32 parm4; 1855 }; 1856 1857 kmp_uint32 ordered_lower; 1858 kmp_uint32 ordered_upper; 1859 #if KMP_OS_WINDOWS 1860 kmp_int32 last_upper; 1861 #endif /* KMP_OS_WINDOWS */ 1862 } dispatch_private_info32_t; 1863 1864 typedef struct KMP_ALIGN_CACHE dispatch_private_info64 { 1865 kmp_int64 count; // current chunk number for static & static-steal scheduling 1866 kmp_int64 ub; /* upper-bound */ 1867 /* Adding KMP_ALIGN_CACHE here doesn't help / can hurt performance */ 1868 kmp_int64 lb; /* lower-bound */ 1869 kmp_int64 st; /* stride */ 1870 kmp_int64 tc; /* trip count (number of iterations) */ 1871 kmp_lock_t *steal_lock; // lock used for chunk stealing 1872 /* parm[1-4] are used in different ways by different scheduling algorithms */ 1873 1874 // KMP_ALIGN( 32 ) ensures ( if the KMP_ALIGN macro is turned on ) 1875 // a) parm3 is properly aligned and 1876 // b) all parm1-4 are in the same cache line. 1877 // Because of parm1-4 are used together, performance seems to be better 1878 // if they are in the same line (not measured though). 1879 1880 struct KMP_ALIGN(32) { 1881 kmp_int64 parm1; 1882 kmp_int64 parm2; 1883 kmp_int64 parm3; 1884 kmp_int64 parm4; 1885 }; 1886 1887 kmp_uint64 ordered_lower; 1888 kmp_uint64 ordered_upper; 1889 #if KMP_OS_WINDOWS 1890 kmp_int64 last_upper; 1891 #endif /* KMP_OS_WINDOWS */ 1892 } dispatch_private_info64_t; 1893 #else /* KMP_STATIC_STEAL_ENABLED */ 1894 typedef struct KMP_ALIGN_CACHE dispatch_private_info32 { 1895 kmp_int32 lb; 1896 kmp_int32 ub; 1897 kmp_int32 st; 1898 kmp_int32 tc; 1899 1900 kmp_int32 parm1; 1901 kmp_int32 parm2; 1902 kmp_int32 parm3; 1903 kmp_int32 parm4; 1904 1905 kmp_int32 count; 1906 1907 kmp_uint32 ordered_lower; 1908 kmp_uint32 ordered_upper; 1909 #if KMP_OS_WINDOWS 1910 kmp_int32 last_upper; 1911 #endif /* KMP_OS_WINDOWS */ 1912 } dispatch_private_info32_t; 1913 1914 typedef struct KMP_ALIGN_CACHE dispatch_private_info64 { 1915 kmp_int64 lb; /* lower-bound */ 1916 kmp_int64 ub; /* upper-bound */ 1917 kmp_int64 st; /* stride */ 1918 kmp_int64 tc; /* trip count (number of iterations) */ 1919 1920 /* parm[1-4] are used in different ways by different scheduling algorithms */ 1921 kmp_int64 parm1; 1922 kmp_int64 parm2; 1923 kmp_int64 parm3; 1924 kmp_int64 parm4; 1925 1926 kmp_int64 count; /* current chunk number for static scheduling */ 1927 1928 kmp_uint64 ordered_lower; 1929 kmp_uint64 ordered_upper; 1930 #if KMP_OS_WINDOWS 1931 kmp_int64 last_upper; 1932 #endif /* KMP_OS_WINDOWS */ 1933 } dispatch_private_info64_t; 1934 #endif /* KMP_STATIC_STEAL_ENABLED */ 1935 1936 typedef struct KMP_ALIGN_CACHE dispatch_private_info { 1937 union private_info { 1938 dispatch_private_info32_t p32; 1939 dispatch_private_info64_t p64; 1940 } u; 1941 enum sched_type schedule; /* scheduling algorithm */ 1942 kmp_sched_flags_t flags; /* flags (e.g., ordered, nomerge, etc.) */ 1943 std::atomic<kmp_uint32> steal_flag; // static_steal only, state of a buffer 1944 kmp_int32 ordered_bumped; 1945 // Stack of buffers for nest of serial regions 1946 struct dispatch_private_info *next; 1947 kmp_int32 type_size; /* the size of types in private_info */ 1948 #if KMP_USE_HIER_SCHED 1949 kmp_int32 hier_id; 1950 void *parent; /* hierarchical scheduling parent pointer */ 1951 #endif 1952 enum cons_type pushed_ws; 1953 } dispatch_private_info_t; 1954 1955 typedef struct dispatch_shared_info32 { 1956 /* chunk index under dynamic, number of idle threads under static-steal; 1957 iteration index otherwise */ 1958 volatile kmp_uint32 iteration; 1959 volatile kmp_int32 num_done; 1960 volatile kmp_uint32 ordered_iteration; 1961 // Dummy to retain the structure size after making ordered_iteration scalar 1962 kmp_int32 ordered_dummy[KMP_MAX_ORDERED - 1]; 1963 } dispatch_shared_info32_t; 1964 1965 typedef struct dispatch_shared_info64 { 1966 /* chunk index under dynamic, number of idle threads under static-steal; 1967 iteration index otherwise */ 1968 volatile kmp_uint64 iteration; 1969 volatile kmp_int64 num_done; 1970 volatile kmp_uint64 ordered_iteration; 1971 // Dummy to retain the structure size after making ordered_iteration scalar 1972 kmp_int64 ordered_dummy[KMP_MAX_ORDERED - 3]; 1973 } dispatch_shared_info64_t; 1974 1975 typedef struct dispatch_shared_info { 1976 union shared_info { 1977 dispatch_shared_info32_t s32; 1978 dispatch_shared_info64_t s64; 1979 } u; 1980 volatile kmp_uint32 buffer_index; 1981 volatile kmp_int32 doacross_buf_idx; // teamwise index 1982 volatile kmp_uint32 *doacross_flags; // shared array of iteration flags (0/1) 1983 kmp_int32 doacross_num_done; // count finished threads 1984 #if KMP_USE_HIER_SCHED 1985 void *hier; 1986 #endif 1987 #if KMP_USE_HWLOC 1988 // When linking with libhwloc, the ORDERED EPCC test slows down on big 1989 // machines (> 48 cores). Performance analysis showed that a cache thrash 1990 // was occurring and this padding helps alleviate the problem. 1991 char padding[64]; 1992 #endif 1993 } dispatch_shared_info_t; 1994 1995 typedef struct kmp_disp { 1996 /* Vector for ORDERED SECTION */ 1997 void (*th_deo_fcn)(int *gtid, int *cid, ident_t *); 1998 /* Vector for END ORDERED SECTION */ 1999 void (*th_dxo_fcn)(int *gtid, int *cid, ident_t *); 2000 2001 dispatch_shared_info_t *th_dispatch_sh_current; 2002 dispatch_private_info_t *th_dispatch_pr_current; 2003 2004 dispatch_private_info_t *th_disp_buffer; 2005 kmp_uint32 th_disp_index; 2006 kmp_int32 th_doacross_buf_idx; // thread's doacross buffer index 2007 volatile kmp_uint32 *th_doacross_flags; // pointer to shared array of flags 2008 kmp_int64 *th_doacross_info; // info on loop bounds 2009 #if KMP_USE_INTERNODE_ALIGNMENT 2010 char more_padding[INTERNODE_CACHE_LINE]; 2011 #endif 2012 } kmp_disp_t; 2013 2014 /* ------------------------------------------------------------------------ */ 2015 /* Barrier stuff */ 2016 2017 /* constants for barrier state update */ 2018 #define KMP_INIT_BARRIER_STATE 0 /* should probably start from zero */ 2019 #define KMP_BARRIER_SLEEP_BIT 0 /* bit used for suspend/sleep part of state */ 2020 #define KMP_BARRIER_UNUSED_BIT 1 // bit that must never be set for valid state 2021 #define KMP_BARRIER_BUMP_BIT 2 /* lsb used for bump of go/arrived state */ 2022 2023 #define KMP_BARRIER_SLEEP_STATE (1 << KMP_BARRIER_SLEEP_BIT) 2024 #define KMP_BARRIER_UNUSED_STATE (1 << KMP_BARRIER_UNUSED_BIT) 2025 #define KMP_BARRIER_STATE_BUMP (1 << KMP_BARRIER_BUMP_BIT) 2026 2027 #if (KMP_BARRIER_SLEEP_BIT >= KMP_BARRIER_BUMP_BIT) 2028 #error "Barrier sleep bit must be smaller than barrier bump bit" 2029 #endif 2030 #if (KMP_BARRIER_UNUSED_BIT >= KMP_BARRIER_BUMP_BIT) 2031 #error "Barrier unused bit must be smaller than barrier bump bit" 2032 #endif 2033 2034 // Constants for release barrier wait state: currently, hierarchical only 2035 #define KMP_BARRIER_NOT_WAITING 0 // Normal state; worker not in wait_sleep 2036 #define KMP_BARRIER_OWN_FLAG \ 2037 1 // Normal state; worker waiting on own b_go flag in release 2038 #define KMP_BARRIER_PARENT_FLAG \ 2039 2 // Special state; worker waiting on parent's b_go flag in release 2040 #define KMP_BARRIER_SWITCH_TO_OWN_FLAG \ 2041 3 // Special state; tells worker to shift from parent to own b_go 2042 #define KMP_BARRIER_SWITCHING \ 2043 4 // Special state; worker resets appropriate flag on wake-up 2044 2045 #define KMP_NOT_SAFE_TO_REAP \ 2046 0 // Thread th_reap_state: not safe to reap (tasking) 2047 #define KMP_SAFE_TO_REAP 1 // Thread th_reap_state: safe to reap (not tasking) 2048 2049 // The flag_type describes the storage used for the flag. 2050 enum flag_type { 2051 flag32, /**< atomic 32 bit flags */ 2052 flag64, /**< 64 bit flags */ 2053 atomic_flag64, /**< atomic 64 bit flags */ 2054 flag_oncore, /**< special 64-bit flag for on-core barrier (hierarchical) */ 2055 flag_unset 2056 }; 2057 2058 enum barrier_type { 2059 bs_plain_barrier = 0, /* 0, All non-fork/join barriers (except reduction 2060 barriers if enabled) */ 2061 bs_forkjoin_barrier, /* 1, All fork/join (parallel region) barriers */ 2062 #if KMP_FAST_REDUCTION_BARRIER 2063 bs_reduction_barrier, /* 2, All barriers that are used in reduction */ 2064 #endif // KMP_FAST_REDUCTION_BARRIER 2065 bs_last_barrier /* Just a placeholder to mark the end */ 2066 }; 2067 2068 // to work with reduction barriers just like with plain barriers 2069 #if !KMP_FAST_REDUCTION_BARRIER 2070 #define bs_reduction_barrier bs_plain_barrier 2071 #endif // KMP_FAST_REDUCTION_BARRIER 2072 2073 typedef enum kmp_bar_pat { /* Barrier communication patterns */ 2074 bp_linear_bar = 2075 0, /* Single level (degenerate) tree */ 2076 bp_tree_bar = 2077 1, /* Balanced tree with branching factor 2^n */ 2078 bp_hyper_bar = 2, /* Hypercube-embedded tree with min 2079 branching factor 2^n */ 2080 bp_hierarchical_bar = 3, /* Machine hierarchy tree */ 2081 bp_dist_bar = 4, /* Distributed barrier */ 2082 bp_last_bar /* Placeholder to mark the end */ 2083 } kmp_bar_pat_e; 2084 2085 #define KMP_BARRIER_ICV_PUSH 1 2086 2087 /* Record for holding the values of the internal controls stack records */ 2088 typedef struct kmp_internal_control { 2089 int serial_nesting_level; /* corresponds to the value of the 2090 th_team_serialized field */ 2091 kmp_int8 dynamic; /* internal control for dynamic adjustment of threads (per 2092 thread) */ 2093 kmp_int8 2094 bt_set; /* internal control for whether blocktime is explicitly set */ 2095 int blocktime; /* internal control for blocktime */ 2096 #if KMP_USE_MONITOR 2097 int bt_intervals; /* internal control for blocktime intervals */ 2098 #endif 2099 int nproc; /* internal control for #threads for next parallel region (per 2100 thread) */ 2101 int thread_limit; /* internal control for thread-limit-var */ 2102 int max_active_levels; /* internal control for max_active_levels */ 2103 kmp_r_sched_t 2104 sched; /* internal control for runtime schedule {sched,chunk} pair */ 2105 kmp_proc_bind_t proc_bind; /* internal control for affinity */ 2106 kmp_int32 default_device; /* internal control for default device */ 2107 struct kmp_internal_control *next; 2108 } kmp_internal_control_t; 2109 2110 static inline void copy_icvs(kmp_internal_control_t *dst, 2111 kmp_internal_control_t *src) { 2112 *dst = *src; 2113 } 2114 2115 /* Thread barrier needs volatile barrier fields */ 2116 typedef struct KMP_ALIGN_CACHE kmp_bstate { 2117 // th_fixed_icvs is aligned by virtue of kmp_bstate being aligned (and all 2118 // uses of it). It is not explicitly aligned below, because we *don't* want 2119 // it to be padded -- instead, we fit b_go into the same cache line with 2120 // th_fixed_icvs, enabling NGO cache lines stores in the hierarchical barrier. 2121 kmp_internal_control_t th_fixed_icvs; // Initial ICVs for the thread 2122 // Tuck b_go into end of th_fixed_icvs cache line, so it can be stored with 2123 // same NGO store 2124 volatile kmp_uint64 b_go; // STATE => task should proceed (hierarchical) 2125 KMP_ALIGN_CACHE volatile kmp_uint64 2126 b_arrived; // STATE => task reached synch point. 2127 kmp_uint32 *skip_per_level; 2128 kmp_uint32 my_level; 2129 kmp_int32 parent_tid; 2130 kmp_int32 old_tid; 2131 kmp_uint32 depth; 2132 struct kmp_bstate *parent_bar; 2133 kmp_team_t *team; 2134 kmp_uint64 leaf_state; 2135 kmp_uint32 nproc; 2136 kmp_uint8 base_leaf_kids; 2137 kmp_uint8 leaf_kids; 2138 kmp_uint8 offset; 2139 kmp_uint8 wait_flag; 2140 kmp_uint8 use_oncore_barrier; 2141 #if USE_DEBUGGER 2142 // The following field is intended for the debugger solely. Only the worker 2143 // thread itself accesses this field: the worker increases it by 1 when it 2144 // arrives to a barrier. 2145 KMP_ALIGN_CACHE kmp_uint b_worker_arrived; 2146 #endif /* USE_DEBUGGER */ 2147 } kmp_bstate_t; 2148 2149 union KMP_ALIGN_CACHE kmp_barrier_union { 2150 double b_align; /* use worst case alignment */ 2151 char b_pad[KMP_PAD(kmp_bstate_t, CACHE_LINE)]; 2152 kmp_bstate_t bb; 2153 }; 2154 2155 typedef union kmp_barrier_union kmp_balign_t; 2156 2157 /* Team barrier needs only non-volatile arrived counter */ 2158 union KMP_ALIGN_CACHE kmp_barrier_team_union { 2159 double b_align; /* use worst case alignment */ 2160 char b_pad[CACHE_LINE]; 2161 struct { 2162 kmp_uint64 b_arrived; /* STATE => task reached synch point. */ 2163 #if USE_DEBUGGER 2164 // The following two fields are indended for the debugger solely. Only 2165 // primary thread of the team accesses these fields: the first one is 2166 // increased by 1 when the primary thread arrives to a barrier, the second 2167 // one is increased by one when all the threads arrived. 2168 kmp_uint b_master_arrived; 2169 kmp_uint b_team_arrived; 2170 #endif 2171 }; 2172 }; 2173 2174 typedef union kmp_barrier_team_union kmp_balign_team_t; 2175 2176 /* Padding for Linux* OS pthreads condition variables and mutexes used to signal 2177 threads when a condition changes. This is to workaround an NPTL bug where 2178 padding was added to pthread_cond_t which caused the initialization routine 2179 to write outside of the structure if compiled on pre-NPTL threads. */ 2180 #if KMP_OS_WINDOWS 2181 typedef struct kmp_win32_mutex { 2182 /* The Lock */ 2183 CRITICAL_SECTION cs; 2184 } kmp_win32_mutex_t; 2185 2186 typedef struct kmp_win32_cond { 2187 /* Count of the number of waiters. */ 2188 int waiters_count_; 2189 2190 /* Serialize access to <waiters_count_> */ 2191 kmp_win32_mutex_t waiters_count_lock_; 2192 2193 /* Number of threads to release via a <cond_broadcast> or a <cond_signal> */ 2194 int release_count_; 2195 2196 /* Keeps track of the current "generation" so that we don't allow */ 2197 /* one thread to steal all the "releases" from the broadcast. */ 2198 int wait_generation_count_; 2199 2200 /* A manual-reset event that's used to block and release waiting threads. */ 2201 HANDLE event_; 2202 } kmp_win32_cond_t; 2203 #endif 2204 2205 #if KMP_OS_UNIX 2206 2207 union KMP_ALIGN_CACHE kmp_cond_union { 2208 double c_align; 2209 char c_pad[CACHE_LINE]; 2210 pthread_cond_t c_cond; 2211 }; 2212 2213 typedef union kmp_cond_union kmp_cond_align_t; 2214 2215 union KMP_ALIGN_CACHE kmp_mutex_union { 2216 double m_align; 2217 char m_pad[CACHE_LINE]; 2218 pthread_mutex_t m_mutex; 2219 }; 2220 2221 typedef union kmp_mutex_union kmp_mutex_align_t; 2222 2223 #endif /* KMP_OS_UNIX */ 2224 2225 typedef struct kmp_desc_base { 2226 void *ds_stackbase; 2227 size_t ds_stacksize; 2228 int ds_stackgrow; 2229 kmp_thread_t ds_thread; 2230 volatile int ds_tid; 2231 int ds_gtid; 2232 #if KMP_OS_WINDOWS 2233 volatile int ds_alive; 2234 DWORD ds_thread_id; 2235 /* ds_thread keeps thread handle on Windows* OS. It is enough for RTL purposes. 2236 However, debugger support (libomp_db) cannot work with handles, because they 2237 uncomparable. For example, debugger requests info about thread with handle h. 2238 h is valid within debugger process, and meaningless within debugee process. 2239 Even if h is duped by call to DuplicateHandle(), so the result h' is valid 2240 within debugee process, but it is a *new* handle which does *not* equal to 2241 any other handle in debugee... The only way to compare handles is convert 2242 them to system-wide ids. GetThreadId() function is available only in 2243 Longhorn and Server 2003. :-( In contrast, GetCurrentThreadId() is available 2244 on all Windows* OS flavours (including Windows* 95). Thus, we have to get 2245 thread id by call to GetCurrentThreadId() from within the thread and save it 2246 to let libomp_db identify threads. */ 2247 #endif /* KMP_OS_WINDOWS */ 2248 } kmp_desc_base_t; 2249 2250 typedef union KMP_ALIGN_CACHE kmp_desc { 2251 double ds_align; /* use worst case alignment */ 2252 char ds_pad[KMP_PAD(kmp_desc_base_t, CACHE_LINE)]; 2253 kmp_desc_base_t ds; 2254 } kmp_desc_t; 2255 2256 typedef struct kmp_local { 2257 volatile int this_construct; /* count of single's encountered by thread */ 2258 void *reduce_data; 2259 #if KMP_USE_BGET 2260 void *bget_data; 2261 void *bget_list; 2262 #if !USE_CMP_XCHG_FOR_BGET 2263 #ifdef USE_QUEUING_LOCK_FOR_BGET 2264 kmp_lock_t bget_lock; /* Lock for accessing bget free list */ 2265 #else 2266 kmp_bootstrap_lock_t bget_lock; // Lock for accessing bget free list. Must be 2267 // bootstrap lock so we can use it at library 2268 // shutdown. 2269 #endif /* USE_LOCK_FOR_BGET */ 2270 #endif /* ! USE_CMP_XCHG_FOR_BGET */ 2271 #endif /* KMP_USE_BGET */ 2272 2273 PACKED_REDUCTION_METHOD_T 2274 packed_reduction_method; /* stored by __kmpc_reduce*(), used by 2275 __kmpc_end_reduce*() */ 2276 2277 } kmp_local_t; 2278 2279 #define KMP_CHECK_UPDATE(a, b) \ 2280 if ((a) != (b)) \ 2281 (a) = (b) 2282 #define KMP_CHECK_UPDATE_SYNC(a, b) \ 2283 if ((a) != (b)) \ 2284 TCW_SYNC_PTR((a), (b)) 2285 2286 #define get__blocktime(xteam, xtid) \ 2287 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.blocktime) 2288 #define get__bt_set(xteam, xtid) \ 2289 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_set) 2290 #if KMP_USE_MONITOR 2291 #define get__bt_intervals(xteam, xtid) \ 2292 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_intervals) 2293 #endif 2294 2295 #define get__dynamic_2(xteam, xtid) \ 2296 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.dynamic) 2297 #define get__nproc_2(xteam, xtid) \ 2298 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.nproc) 2299 #define get__sched_2(xteam, xtid) \ 2300 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.sched) 2301 2302 #define set__blocktime_team(xteam, xtid, xval) \ 2303 (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.blocktime) = \ 2304 (xval)) 2305 2306 #if KMP_USE_MONITOR 2307 #define set__bt_intervals_team(xteam, xtid, xval) \ 2308 (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_intervals) = \ 2309 (xval)) 2310 #endif 2311 2312 #define set__bt_set_team(xteam, xtid, xval) \ 2313 (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_set) = (xval)) 2314 2315 #define set__dynamic(xthread, xval) \ 2316 (((xthread)->th.th_current_task->td_icvs.dynamic) = (xval)) 2317 #define get__dynamic(xthread) \ 2318 (((xthread)->th.th_current_task->td_icvs.dynamic) ? (FTN_TRUE) : (FTN_FALSE)) 2319 2320 #define set__nproc(xthread, xval) \ 2321 (((xthread)->th.th_current_task->td_icvs.nproc) = (xval)) 2322 2323 #define set__thread_limit(xthread, xval) \ 2324 (((xthread)->th.th_current_task->td_icvs.thread_limit) = (xval)) 2325 2326 #define set__max_active_levels(xthread, xval) \ 2327 (((xthread)->th.th_current_task->td_icvs.max_active_levels) = (xval)) 2328 2329 #define get__max_active_levels(xthread) \ 2330 ((xthread)->th.th_current_task->td_icvs.max_active_levels) 2331 2332 #define set__sched(xthread, xval) \ 2333 (((xthread)->th.th_current_task->td_icvs.sched) = (xval)) 2334 2335 #define set__proc_bind(xthread, xval) \ 2336 (((xthread)->th.th_current_task->td_icvs.proc_bind) = (xval)) 2337 #define get__proc_bind(xthread) \ 2338 ((xthread)->th.th_current_task->td_icvs.proc_bind) 2339 2340 // OpenMP tasking data structures 2341 2342 typedef enum kmp_tasking_mode { 2343 tskm_immediate_exec = 0, 2344 tskm_extra_barrier = 1, 2345 tskm_task_teams = 2, 2346 tskm_max = 2 2347 } kmp_tasking_mode_t; 2348 2349 extern kmp_tasking_mode_t 2350 __kmp_tasking_mode; /* determines how/when to execute tasks */ 2351 extern int __kmp_task_stealing_constraint; 2352 extern int __kmp_enable_task_throttling; 2353 extern kmp_int32 __kmp_default_device; // Set via OMP_DEFAULT_DEVICE if 2354 // specified, defaults to 0 otherwise 2355 // Set via OMP_MAX_TASK_PRIORITY if specified, defaults to 0 otherwise 2356 extern kmp_int32 __kmp_max_task_priority; 2357 // Set via KMP_TASKLOOP_MIN_TASKS if specified, defaults to 0 otherwise 2358 extern kmp_uint64 __kmp_taskloop_min_tasks; 2359 2360 /* NOTE: kmp_taskdata_t and kmp_task_t structures allocated in single block with 2361 taskdata first */ 2362 #define KMP_TASK_TO_TASKDATA(task) (((kmp_taskdata_t *)task) - 1) 2363 #define KMP_TASKDATA_TO_TASK(taskdata) (kmp_task_t *)(taskdata + 1) 2364 2365 // The tt_found_tasks flag is a signal to all threads in the team that tasks 2366 // were spawned and queued since the previous barrier release. 2367 #define KMP_TASKING_ENABLED(task_team) \ 2368 (TRUE == TCR_SYNC_4((task_team)->tt.tt_found_tasks)) 2369 /*! 2370 @ingroup BASIC_TYPES 2371 @{ 2372 */ 2373 2374 /*! 2375 */ 2376 typedef kmp_int32 (*kmp_routine_entry_t)(kmp_int32, void *); 2377 2378 typedef union kmp_cmplrdata { 2379 kmp_int32 priority; /**< priority specified by user for the task */ 2380 kmp_routine_entry_t 2381 destructors; /* pointer to function to invoke deconstructors of 2382 firstprivate C++ objects */ 2383 /* future data */ 2384 } kmp_cmplrdata_t; 2385 2386 /* sizeof_kmp_task_t passed as arg to kmpc_omp_task call */ 2387 /*! 2388 */ 2389 typedef struct kmp_task { /* GEH: Shouldn't this be aligned somehow? */ 2390 void *shareds; /**< pointer to block of pointers to shared vars */ 2391 kmp_routine_entry_t 2392 routine; /**< pointer to routine to call for executing task */ 2393 kmp_int32 part_id; /**< part id for the task */ 2394 kmp_cmplrdata_t 2395 data1; /* Two known optional additions: destructors and priority */ 2396 kmp_cmplrdata_t data2; /* Process destructors first, priority second */ 2397 /* future data */ 2398 /* private vars */ 2399 } kmp_task_t; 2400 2401 /*! 2402 @} 2403 */ 2404 2405 typedef struct kmp_taskgroup { 2406 std::atomic<kmp_int32> count; // number of allocated and incomplete tasks 2407 std::atomic<kmp_int32> 2408 cancel_request; // request for cancellation of this taskgroup 2409 struct kmp_taskgroup *parent; // parent taskgroup 2410 // Block of data to perform task reduction 2411 void *reduce_data; // reduction related info 2412 kmp_int32 reduce_num_data; // number of data items to reduce 2413 uintptr_t *gomp_data; // gomp reduction data 2414 } kmp_taskgroup_t; 2415 2416 // forward declarations 2417 typedef union kmp_depnode kmp_depnode_t; 2418 typedef struct kmp_depnode_list kmp_depnode_list_t; 2419 typedef struct kmp_dephash_entry kmp_dephash_entry_t; 2420 2421 // macros for checking dep flag as an integer 2422 #define KMP_DEP_IN 0x1 2423 #define KMP_DEP_OUT 0x2 2424 #define KMP_DEP_INOUT 0x3 2425 #define KMP_DEP_MTX 0x4 2426 #define KMP_DEP_SET 0x8 2427 #define KMP_DEP_ALL 0x80 2428 // Compiler sends us this info: 2429 typedef struct kmp_depend_info { 2430 kmp_intptr_t base_addr; 2431 size_t len; 2432 union { 2433 kmp_uint8 flag; // flag as an unsigned char 2434 struct { // flag as a set of 8 bits 2435 unsigned in : 1; 2436 unsigned out : 1; 2437 unsigned mtx : 1; 2438 unsigned set : 1; 2439 unsigned unused : 3; 2440 unsigned all : 1; 2441 } flags; 2442 }; 2443 } kmp_depend_info_t; 2444 2445 // Internal structures to work with task dependencies: 2446 struct kmp_depnode_list { 2447 kmp_depnode_t *node; 2448 kmp_depnode_list_t *next; 2449 }; 2450 2451 // Max number of mutexinoutset dependencies per node 2452 #define MAX_MTX_DEPS 4 2453 2454 typedef struct kmp_base_depnode { 2455 kmp_depnode_list_t *successors; /* used under lock */ 2456 kmp_task_t *task; /* non-NULL if depnode is active, used under lock */ 2457 kmp_lock_t *mtx_locks[MAX_MTX_DEPS]; /* lock mutexinoutset dependent tasks */ 2458 kmp_int32 mtx_num_locks; /* number of locks in mtx_locks array */ 2459 kmp_lock_t lock; /* guards shared fields: task, successors */ 2460 #if KMP_SUPPORT_GRAPH_OUTPUT 2461 kmp_uint32 id; 2462 #endif 2463 std::atomic<kmp_int32> npredecessors; 2464 std::atomic<kmp_int32> nrefs; 2465 } kmp_base_depnode_t; 2466 2467 union KMP_ALIGN_CACHE kmp_depnode { 2468 double dn_align; /* use worst case alignment */ 2469 char dn_pad[KMP_PAD(kmp_base_depnode_t, CACHE_LINE)]; 2470 kmp_base_depnode_t dn; 2471 }; 2472 2473 struct kmp_dephash_entry { 2474 kmp_intptr_t addr; 2475 kmp_depnode_t *last_out; 2476 kmp_depnode_list_t *last_set; 2477 kmp_depnode_list_t *prev_set; 2478 kmp_uint8 last_flag; 2479 kmp_lock_t *mtx_lock; /* is referenced by depnodes w/mutexinoutset dep */ 2480 kmp_dephash_entry_t *next_in_bucket; 2481 }; 2482 2483 typedef struct kmp_dephash { 2484 kmp_dephash_entry_t **buckets; 2485 size_t size; 2486 kmp_depnode_t *last_all; 2487 size_t generation; 2488 kmp_uint32 nelements; 2489 kmp_uint32 nconflicts; 2490 } kmp_dephash_t; 2491 2492 typedef struct kmp_task_affinity_info { 2493 kmp_intptr_t base_addr; 2494 size_t len; 2495 struct { 2496 bool flag1 : 1; 2497 bool flag2 : 1; 2498 kmp_int32 reserved : 30; 2499 } flags; 2500 } kmp_task_affinity_info_t; 2501 2502 typedef enum kmp_event_type_t { 2503 KMP_EVENT_UNINITIALIZED = 0, 2504 KMP_EVENT_ALLOW_COMPLETION = 1 2505 } kmp_event_type_t; 2506 2507 typedef struct { 2508 kmp_event_type_t type; 2509 kmp_tas_lock_t lock; 2510 union { 2511 kmp_task_t *task; 2512 } ed; 2513 } kmp_event_t; 2514 2515 #if OMPX_TASKGRAPH 2516 // Initial number of allocated nodes while recording 2517 #define INIT_MAPSIZE 50 2518 2519 typedef struct kmp_taskgraph_flags { /*This needs to be exactly 32 bits */ 2520 unsigned nowait : 1; 2521 unsigned re_record : 1; 2522 unsigned reserved : 30; 2523 } kmp_taskgraph_flags_t; 2524 2525 /// Represents a TDG node 2526 typedef struct kmp_node_info { 2527 kmp_task_t *task; // Pointer to the actual task 2528 kmp_int32 *successors; // Array of the succesors ids 2529 kmp_int32 nsuccessors; // Number of succesors of the node 2530 std::atomic<kmp_int32> 2531 npredecessors_counter; // Number of predessors on the fly 2532 kmp_int32 npredecessors; // Total number of predecessors 2533 kmp_int32 successors_size; // Number of allocated succesors ids 2534 kmp_taskdata_t *parent_task; // Parent implicit task 2535 } kmp_node_info_t; 2536 2537 /// Represent a TDG's current status 2538 typedef enum kmp_tdg_status { 2539 KMP_TDG_NONE = 0, 2540 KMP_TDG_RECORDING = 1, 2541 KMP_TDG_READY = 2 2542 } kmp_tdg_status_t; 2543 2544 /// Structure that contains a TDG 2545 typedef struct kmp_tdg_info { 2546 kmp_int32 tdg_id; // Unique idenfifier of the TDG 2547 kmp_taskgraph_flags_t tdg_flags; // Flags related to a TDG 2548 kmp_int32 map_size; // Number of allocated TDG nodes 2549 kmp_int32 num_roots; // Number of roots tasks int the TDG 2550 kmp_int32 *root_tasks; // Array of tasks identifiers that are roots 2551 kmp_node_info_t *record_map; // Array of TDG nodes 2552 kmp_tdg_status_t tdg_status = 2553 KMP_TDG_NONE; // Status of the TDG (recording, ready...) 2554 std::atomic<kmp_int32> num_tasks; // Number of TDG nodes 2555 kmp_bootstrap_lock_t 2556 graph_lock; // Protect graph attributes when updated via taskloop_recur 2557 // Taskloop reduction related 2558 void *rec_taskred_data; // Data to pass to __kmpc_task_reduction_init or 2559 // __kmpc_taskred_init 2560 kmp_int32 rec_num_taskred; 2561 } kmp_tdg_info_t; 2562 2563 extern int __kmp_tdg_dot; 2564 extern kmp_int32 __kmp_max_tdgs; 2565 extern kmp_tdg_info_t **__kmp_global_tdgs; 2566 extern kmp_int32 __kmp_curr_tdg_idx; 2567 extern kmp_int32 __kmp_successors_size; 2568 extern std::atomic<kmp_int32> __kmp_tdg_task_id; 2569 extern kmp_int32 __kmp_num_tdg; 2570 #endif 2571 2572 #ifdef BUILD_TIED_TASK_STACK 2573 2574 /* Tied Task stack definitions */ 2575 typedef struct kmp_stack_block { 2576 kmp_taskdata_t *sb_block[TASK_STACK_BLOCK_SIZE]; 2577 struct kmp_stack_block *sb_next; 2578 struct kmp_stack_block *sb_prev; 2579 } kmp_stack_block_t; 2580 2581 typedef struct kmp_task_stack { 2582 kmp_stack_block_t ts_first_block; // first block of stack entries 2583 kmp_taskdata_t **ts_top; // pointer to the top of stack 2584 kmp_int32 ts_entries; // number of entries on the stack 2585 } kmp_task_stack_t; 2586 2587 #endif // BUILD_TIED_TASK_STACK 2588 2589 typedef struct kmp_tasking_flags { /* Total struct must be exactly 32 bits */ 2590 /* Compiler flags */ /* Total compiler flags must be 16 bits */ 2591 unsigned tiedness : 1; /* task is either tied (1) or untied (0) */ 2592 unsigned final : 1; /* task is final(1) so execute immediately */ 2593 unsigned merged_if0 : 1; /* no __kmpc_task_{begin/complete}_if0 calls in if0 2594 code path */ 2595 unsigned destructors_thunk : 1; /* set if the compiler creates a thunk to 2596 invoke destructors from the runtime */ 2597 unsigned proxy : 1; /* task is a proxy task (it will be executed outside the 2598 context of the RTL) */ 2599 unsigned priority_specified : 1; /* set if the compiler provides priority 2600 setting for the task */ 2601 unsigned detachable : 1; /* 1 == can detach */ 2602 unsigned hidden_helper : 1; /* 1 == hidden helper task */ 2603 unsigned reserved : 8; /* reserved for compiler use */ 2604 2605 /* Library flags */ /* Total library flags must be 16 bits */ 2606 unsigned tasktype : 1; /* task is either explicit(1) or implicit (0) */ 2607 unsigned task_serial : 1; // task is executed immediately (1) or deferred (0) 2608 unsigned tasking_ser : 1; // all tasks in team are either executed immediately 2609 // (1) or may be deferred (0) 2610 unsigned team_serial : 1; // entire team is serial (1) [1 thread] or parallel 2611 // (0) [>= 2 threads] 2612 /* If either team_serial or tasking_ser is set, task team may be NULL */ 2613 /* Task State Flags: */ 2614 unsigned started : 1; /* 1==started, 0==not started */ 2615 unsigned executing : 1; /* 1==executing, 0==not executing */ 2616 unsigned complete : 1; /* 1==complete, 0==not complete */ 2617 unsigned freed : 1; /* 1==freed, 0==allocated */ 2618 unsigned native : 1; /* 1==gcc-compiled task, 0==intel */ 2619 #if OMPX_TASKGRAPH 2620 unsigned onced : 1; /* 1==ran once already, 0==never ran, record & replay purposes */ 2621 unsigned reserved31 : 6; /* reserved for library use */ 2622 #else 2623 unsigned reserved31 : 7; /* reserved for library use */ 2624 #endif 2625 2626 } kmp_tasking_flags_t; 2627 2628 typedef struct kmp_target_data { 2629 void *async_handle; // libomptarget async handle for task completion query 2630 } kmp_target_data_t; 2631 2632 struct kmp_taskdata { /* aligned during dynamic allocation */ 2633 kmp_int32 td_task_id; /* id, assigned by debugger */ 2634 kmp_tasking_flags_t td_flags; /* task flags */ 2635 kmp_team_t *td_team; /* team for this task */ 2636 kmp_info_p *td_alloc_thread; /* thread that allocated data structures */ 2637 /* Currently not used except for perhaps IDB */ 2638 kmp_taskdata_t *td_parent; /* parent task */ 2639 kmp_int32 td_level; /* task nesting level */ 2640 std::atomic<kmp_int32> td_untied_count; // untied task active parts counter 2641 ident_t *td_ident; /* task identifier */ 2642 // Taskwait data. 2643 ident_t *td_taskwait_ident; 2644 kmp_uint32 td_taskwait_counter; 2645 kmp_int32 td_taskwait_thread; /* gtid + 1 of thread encountered taskwait */ 2646 KMP_ALIGN_CACHE kmp_internal_control_t 2647 td_icvs; /* Internal control variables for the task */ 2648 KMP_ALIGN_CACHE std::atomic<kmp_int32> 2649 td_allocated_child_tasks; /* Child tasks (+ current task) not yet 2650 deallocated */ 2651 std::atomic<kmp_int32> 2652 td_incomplete_child_tasks; /* Child tasks not yet complete */ 2653 kmp_taskgroup_t 2654 *td_taskgroup; // Each task keeps pointer to its current taskgroup 2655 kmp_dephash_t 2656 *td_dephash; // Dependencies for children tasks are tracked from here 2657 kmp_depnode_t 2658 *td_depnode; // Pointer to graph node if this task has dependencies 2659 kmp_task_team_t *td_task_team; 2660 size_t td_size_alloc; // Size of task structure, including shareds etc. 2661 #if defined(KMP_GOMP_COMPAT) 2662 // 4 or 8 byte integers for the loop bounds in GOMP_taskloop 2663 kmp_int32 td_size_loop_bounds; 2664 #endif 2665 kmp_taskdata_t *td_last_tied; // keep tied task for task scheduling constraint 2666 #if defined(KMP_GOMP_COMPAT) 2667 // GOMP sends in a copy function for copy constructors 2668 void (*td_copy_func)(void *, void *); 2669 #endif 2670 kmp_event_t td_allow_completion_event; 2671 #if OMPT_SUPPORT 2672 ompt_task_info_t ompt_task_info; 2673 #endif 2674 #if OMPX_TASKGRAPH 2675 bool is_taskgraph = 0; // whether the task is within a TDG 2676 kmp_tdg_info_t *tdg; // used to associate task with a TDG 2677 #endif 2678 kmp_target_data_t td_target_data; 2679 }; // struct kmp_taskdata 2680 2681 // Make sure padding above worked 2682 KMP_BUILD_ASSERT(sizeof(kmp_taskdata_t) % sizeof(void *) == 0); 2683 2684 // Data for task team but per thread 2685 typedef struct kmp_base_thread_data { 2686 kmp_info_p *td_thr; // Pointer back to thread info 2687 // Used only in __kmp_execute_tasks_template, maybe not avail until task is 2688 // queued? 2689 kmp_bootstrap_lock_t td_deque_lock; // Lock for accessing deque 2690 kmp_taskdata_t * 2691 *td_deque; // Deque of tasks encountered by td_thr, dynamically allocated 2692 kmp_int32 td_deque_size; // Size of deck 2693 kmp_uint32 td_deque_head; // Head of deque (will wrap) 2694 kmp_uint32 td_deque_tail; // Tail of deque (will wrap) 2695 kmp_int32 td_deque_ntasks; // Number of tasks in deque 2696 // GEH: shouldn't this be volatile since used in while-spin? 2697 kmp_int32 td_deque_last_stolen; // Thread number of last successful steal 2698 #ifdef BUILD_TIED_TASK_STACK 2699 kmp_task_stack_t td_susp_tied_tasks; // Stack of suspended tied tasks for task 2700 // scheduling constraint 2701 #endif // BUILD_TIED_TASK_STACK 2702 } kmp_base_thread_data_t; 2703 2704 #define TASK_DEQUE_BITS 8 // Used solely to define INITIAL_TASK_DEQUE_SIZE 2705 #define INITIAL_TASK_DEQUE_SIZE (1 << TASK_DEQUE_BITS) 2706 2707 #define TASK_DEQUE_SIZE(td) ((td).td_deque_size) 2708 #define TASK_DEQUE_MASK(td) ((td).td_deque_size - 1) 2709 2710 typedef union KMP_ALIGN_CACHE kmp_thread_data { 2711 kmp_base_thread_data_t td; 2712 double td_align; /* use worst case alignment */ 2713 char td_pad[KMP_PAD(kmp_base_thread_data_t, CACHE_LINE)]; 2714 } kmp_thread_data_t; 2715 2716 typedef struct kmp_task_pri { 2717 kmp_thread_data_t td; 2718 kmp_int32 priority; 2719 kmp_task_pri *next; 2720 } kmp_task_pri_t; 2721 2722 // Data for task teams which are used when tasking is enabled for the team 2723 typedef struct kmp_base_task_team { 2724 kmp_bootstrap_lock_t 2725 tt_threads_lock; /* Lock used to allocate per-thread part of task team */ 2726 /* must be bootstrap lock since used at library shutdown*/ 2727 2728 // TODO: check performance vs kmp_tas_lock_t 2729 kmp_bootstrap_lock_t tt_task_pri_lock; /* Lock to access priority tasks */ 2730 kmp_task_pri_t *tt_task_pri_list; 2731 2732 kmp_task_team_t *tt_next; /* For linking the task team free list */ 2733 kmp_thread_data_t 2734 *tt_threads_data; /* Array of per-thread structures for task team */ 2735 /* Data survives task team deallocation */ 2736 kmp_int32 tt_found_tasks; /* Have we found tasks and queued them while 2737 executing this team? */ 2738 /* TRUE means tt_threads_data is set up and initialized */ 2739 kmp_int32 tt_nproc; /* #threads in team */ 2740 kmp_int32 tt_max_threads; // # entries allocated for threads_data array 2741 kmp_int32 tt_found_proxy_tasks; // found proxy tasks since last barrier 2742 kmp_int32 tt_untied_task_encountered; 2743 std::atomic<kmp_int32> tt_num_task_pri; // number of priority tasks enqueued 2744 // There is hidden helper thread encountered in this task team so that we must 2745 // wait when waiting on task team 2746 kmp_int32 tt_hidden_helper_task_encountered; 2747 2748 KMP_ALIGN_CACHE 2749 std::atomic<kmp_int32> tt_unfinished_threads; /* #threads still active */ 2750 2751 KMP_ALIGN_CACHE 2752 volatile kmp_uint32 2753 tt_active; /* is the team still actively executing tasks */ 2754 } kmp_base_task_team_t; 2755 2756 union KMP_ALIGN_CACHE kmp_task_team { 2757 kmp_base_task_team_t tt; 2758 double tt_align; /* use worst case alignment */ 2759 char tt_pad[KMP_PAD(kmp_base_task_team_t, CACHE_LINE)]; 2760 }; 2761 2762 #if (USE_FAST_MEMORY == 3) || (USE_FAST_MEMORY == 5) 2763 // Free lists keep same-size free memory slots for fast memory allocation 2764 // routines 2765 typedef struct kmp_free_list { 2766 void *th_free_list_self; // Self-allocated tasks free list 2767 void *th_free_list_sync; // Self-allocated tasks stolen/returned by other 2768 // threads 2769 void *th_free_list_other; // Non-self free list (to be returned to owner's 2770 // sync list) 2771 } kmp_free_list_t; 2772 #endif 2773 #if KMP_NESTED_HOT_TEAMS 2774 // Hot teams array keeps hot teams and their sizes for given thread. Hot teams 2775 // are not put in teams pool, and they don't put threads in threads pool. 2776 typedef struct kmp_hot_team_ptr { 2777 kmp_team_p *hot_team; // pointer to hot_team of given nesting level 2778 kmp_int32 hot_team_nth; // number of threads allocated for the hot_team 2779 } kmp_hot_team_ptr_t; 2780 #endif 2781 typedef struct kmp_teams_size { 2782 kmp_int32 nteams; // number of teams in a league 2783 kmp_int32 nth; // number of threads in each team of the league 2784 } kmp_teams_size_t; 2785 2786 // This struct stores a thread that acts as a "root" for a contention 2787 // group. Contention groups are rooted at kmp_root threads, but also at 2788 // each primary thread of each team created in the teams construct. 2789 // This struct therefore also stores a thread_limit associated with 2790 // that contention group, and a counter to track the number of threads 2791 // active in that contention group. Each thread has a list of these: CG 2792 // root threads have an entry in their list in which cg_root refers to 2793 // the thread itself, whereas other workers in the CG will have a 2794 // single entry where cg_root is same as the entry containing their CG 2795 // root. When a thread encounters a teams construct, it will add a new 2796 // entry to the front of its list, because it now roots a new CG. 2797 typedef struct kmp_cg_root { 2798 kmp_info_p *cg_root; // "root" thread for a contention group 2799 // The CG root's limit comes from OMP_THREAD_LIMIT for root threads, or 2800 // thread_limit clause for teams primary threads 2801 kmp_int32 cg_thread_limit; 2802 kmp_int32 cg_nthreads; // Count of active threads in CG rooted at cg_root 2803 struct kmp_cg_root *up; // pointer to higher level CG root in list 2804 } kmp_cg_root_t; 2805 2806 // OpenMP thread data structures 2807 2808 typedef struct KMP_ALIGN_CACHE kmp_base_info { 2809 /* Start with the readonly data which is cache aligned and padded. This is 2810 written before the thread starts working by the primary thread. Uber 2811 masters may update themselves later. Usage does not consider serialized 2812 regions. */ 2813 kmp_desc_t th_info; 2814 kmp_team_p *th_team; /* team we belong to */ 2815 kmp_root_p *th_root; /* pointer to root of task hierarchy */ 2816 kmp_info_p *th_next_pool; /* next available thread in the pool */ 2817 kmp_disp_t *th_dispatch; /* thread's dispatch data */ 2818 int th_in_pool; /* in thread pool (32 bits for TCR/TCW) */ 2819 2820 /* The following are cached from the team info structure */ 2821 /* TODO use these in more places as determined to be needed via profiling */ 2822 int th_team_nproc; /* number of threads in a team */ 2823 kmp_info_p *th_team_master; /* the team's primary thread */ 2824 int th_team_serialized; /* team is serialized */ 2825 microtask_t th_teams_microtask; /* save entry address for teams construct */ 2826 int th_teams_level; /* save initial level of teams construct */ 2827 /* it is 0 on device but may be any on host */ 2828 2829 /* The blocktime info is copied from the team struct to the thread struct */ 2830 /* at the start of a barrier, and the values stored in the team are used */ 2831 /* at points in the code where the team struct is no longer guaranteed */ 2832 /* to exist (from the POV of worker threads). */ 2833 #if KMP_USE_MONITOR 2834 int th_team_bt_intervals; 2835 int th_team_bt_set; 2836 #else 2837 kmp_uint64 th_team_bt_intervals; 2838 #endif 2839 2840 #if KMP_AFFINITY_SUPPORTED 2841 kmp_affin_mask_t *th_affin_mask; /* thread's current affinity mask */ 2842 kmp_affinity_ids_t th_topology_ids; /* thread's current topology ids */ 2843 kmp_affinity_attrs_t th_topology_attrs; /* thread's current topology attrs */ 2844 #endif 2845 omp_allocator_handle_t th_def_allocator; /* default allocator */ 2846 /* The data set by the primary thread at reinit, then R/W by the worker */ 2847 KMP_ALIGN_CACHE int 2848 th_set_nproc; /* if > 0, then only use this request for the next fork */ 2849 #if KMP_NESTED_HOT_TEAMS 2850 kmp_hot_team_ptr_t *th_hot_teams; /* array of hot teams */ 2851 #endif 2852 kmp_proc_bind_t 2853 th_set_proc_bind; /* if != proc_bind_default, use request for next fork */ 2854 kmp_teams_size_t 2855 th_teams_size; /* number of teams/threads in teams construct */ 2856 #if KMP_AFFINITY_SUPPORTED 2857 int th_current_place; /* place currently bound to */ 2858 int th_new_place; /* place to bind to in par reg */ 2859 int th_first_place; /* first place in partition */ 2860 int th_last_place; /* last place in partition */ 2861 #endif 2862 int th_prev_level; /* previous level for affinity format */ 2863 int th_prev_num_threads; /* previous num_threads for affinity format */ 2864 #if USE_ITT_BUILD 2865 kmp_uint64 th_bar_arrive_time; /* arrival to barrier timestamp */ 2866 kmp_uint64 th_bar_min_time; /* minimum arrival time at the barrier */ 2867 kmp_uint64 th_frame_time; /* frame timestamp */ 2868 #endif /* USE_ITT_BUILD */ 2869 kmp_local_t th_local; 2870 struct private_common *th_pri_head; 2871 2872 /* Now the data only used by the worker (after initial allocation) */ 2873 /* TODO the first serial team should actually be stored in the info_t 2874 structure. this will help reduce initial allocation overhead */ 2875 KMP_ALIGN_CACHE kmp_team_p 2876 *th_serial_team; /*serialized team held in reserve*/ 2877 2878 #if OMPT_SUPPORT 2879 ompt_thread_info_t ompt_thread_info; 2880 #endif 2881 2882 /* The following are also read by the primary thread during reinit */ 2883 struct common_table *th_pri_common; 2884 2885 volatile kmp_uint32 th_spin_here; /* thread-local location for spinning */ 2886 /* while awaiting queuing lock acquire */ 2887 2888 volatile void *th_sleep_loc; // this points at a kmp_flag<T> 2889 flag_type th_sleep_loc_type; // enum type of flag stored in th_sleep_loc 2890 2891 ident_t *th_ident; 2892 unsigned th_x; // Random number generator data 2893 unsigned th_a; // Random number generator data 2894 2895 /* Tasking-related data for the thread */ 2896 kmp_task_team_t *th_task_team; // Task team struct 2897 kmp_taskdata_t *th_current_task; // Innermost Task being executed 2898 kmp_uint8 th_task_state; // alternating 0/1 for task team identification 2899 kmp_uint8 *th_task_state_memo_stack; // Stack holding memos of th_task_state 2900 // at nested levels 2901 kmp_uint32 th_task_state_top; // Top element of th_task_state_memo_stack 2902 kmp_uint32 th_task_state_stack_sz; // Size of th_task_state_memo_stack 2903 kmp_uint32 th_reap_state; // Non-zero indicates thread is not 2904 // tasking, thus safe to reap 2905 2906 /* More stuff for keeping track of active/sleeping threads (this part is 2907 written by the worker thread) */ 2908 kmp_uint8 th_active_in_pool; // included in count of #active threads in pool 2909 int th_active; // ! sleeping; 32 bits for TCR/TCW 2910 std::atomic<kmp_uint32> th_used_in_team; // Flag indicating use in team 2911 // 0 = not used in team; 1 = used in team; 2912 // 2 = transitioning to not used in team; 3 = transitioning to used in team 2913 struct cons_header *th_cons; // used for consistency check 2914 #if KMP_USE_HIER_SCHED 2915 // used for hierarchical scheduling 2916 kmp_hier_private_bdata_t *th_hier_bar_data; 2917 #endif 2918 2919 /* Add the syncronizing data which is cache aligned and padded. */ 2920 KMP_ALIGN_CACHE kmp_balign_t th_bar[bs_last_barrier]; 2921 2922 KMP_ALIGN_CACHE volatile kmp_int32 2923 th_next_waiting; /* gtid+1 of next thread on lock wait queue, 0 if none */ 2924 2925 #if (USE_FAST_MEMORY == 3) || (USE_FAST_MEMORY == 5) 2926 #define NUM_LISTS 4 2927 kmp_free_list_t th_free_lists[NUM_LISTS]; // Free lists for fast memory 2928 // allocation routines 2929 #endif 2930 2931 #if KMP_OS_WINDOWS 2932 kmp_win32_cond_t th_suspend_cv; 2933 kmp_win32_mutex_t th_suspend_mx; 2934 std::atomic<int> th_suspend_init; 2935 #endif 2936 #if KMP_OS_UNIX 2937 kmp_cond_align_t th_suspend_cv; 2938 kmp_mutex_align_t th_suspend_mx; 2939 std::atomic<int> th_suspend_init_count; 2940 #endif 2941 2942 #if USE_ITT_BUILD 2943 kmp_itt_mark_t th_itt_mark_single; 2944 // alignment ??? 2945 #endif /* USE_ITT_BUILD */ 2946 #if KMP_STATS_ENABLED 2947 kmp_stats_list *th_stats; 2948 #endif 2949 #if KMP_OS_UNIX 2950 std::atomic<bool> th_blocking; 2951 #endif 2952 kmp_cg_root_t *th_cg_roots; // list of cg_roots associated with this thread 2953 } kmp_base_info_t; 2954 2955 typedef union KMP_ALIGN_CACHE kmp_info { 2956 double th_align; /* use worst case alignment */ 2957 char th_pad[KMP_PAD(kmp_base_info_t, CACHE_LINE)]; 2958 kmp_base_info_t th; 2959 } kmp_info_t; 2960 2961 // OpenMP thread team data structures 2962 2963 typedef struct kmp_base_data { 2964 volatile kmp_uint32 t_value; 2965 } kmp_base_data_t; 2966 2967 typedef union KMP_ALIGN_CACHE kmp_sleep_team { 2968 double dt_align; /* use worst case alignment */ 2969 char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)]; 2970 kmp_base_data_t dt; 2971 } kmp_sleep_team_t; 2972 2973 typedef union KMP_ALIGN_CACHE kmp_ordered_team { 2974 double dt_align; /* use worst case alignment */ 2975 char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)]; 2976 kmp_base_data_t dt; 2977 } kmp_ordered_team_t; 2978 2979 typedef int (*launch_t)(int gtid); 2980 2981 /* Minimum number of ARGV entries to malloc if necessary */ 2982 #define KMP_MIN_MALLOC_ARGV_ENTRIES 100 2983 2984 // Set up how many argv pointers will fit in cache lines containing 2985 // t_inline_argv. Historically, we have supported at least 96 bytes. Using a 2986 // larger value for more space between the primary write/worker read section and 2987 // read/write by all section seems to buy more performance on EPCC PARALLEL. 2988 #if KMP_ARCH_X86 || KMP_ARCH_X86_64 2989 #define KMP_INLINE_ARGV_BYTES \ 2990 (4 * CACHE_LINE - \ 2991 ((3 * KMP_PTR_SKIP + 2 * sizeof(int) + 2 * sizeof(kmp_int8) + \ 2992 sizeof(kmp_int16) + sizeof(kmp_uint32)) % \ 2993 CACHE_LINE)) 2994 #else 2995 #define KMP_INLINE_ARGV_BYTES \ 2996 (2 * CACHE_LINE - ((3 * KMP_PTR_SKIP + 2 * sizeof(int)) % CACHE_LINE)) 2997 #endif 2998 #define KMP_INLINE_ARGV_ENTRIES (int)(KMP_INLINE_ARGV_BYTES / KMP_PTR_SKIP) 2999 3000 typedef struct KMP_ALIGN_CACHE kmp_base_team { 3001 // Synchronization Data 3002 // --------------------------------------------------------------------------- 3003 KMP_ALIGN_CACHE kmp_ordered_team_t t_ordered; 3004 kmp_balign_team_t t_bar[bs_last_barrier]; 3005 std::atomic<int> t_construct; // count of single directive encountered by team 3006 char pad[sizeof(kmp_lock_t)]; // padding to maintain performance on big iron 3007 3008 // [0] - parallel / [1] - worksharing task reduction data shared by taskgroups 3009 std::atomic<void *> t_tg_reduce_data[2]; // to support task modifier 3010 std::atomic<int> t_tg_fini_counter[2]; // sync end of task reductions 3011 3012 // Primary thread only 3013 // --------------------------------------------------------------------------- 3014 KMP_ALIGN_CACHE int t_master_tid; // tid of primary thread in parent team 3015 int t_master_this_cons; // "this_construct" single counter of primary thread 3016 // in parent team 3017 ident_t *t_ident; // if volatile, have to change too much other crud to 3018 // volatile too 3019 kmp_team_p *t_parent; // parent team 3020 kmp_team_p *t_next_pool; // next free team in the team pool 3021 kmp_disp_t *t_dispatch; // thread's dispatch data 3022 kmp_task_team_t *t_task_team[2]; // Task team struct; switch between 2 3023 kmp_proc_bind_t t_proc_bind; // bind type for par region 3024 #if USE_ITT_BUILD 3025 kmp_uint64 t_region_time; // region begin timestamp 3026 #endif /* USE_ITT_BUILD */ 3027 3028 // Primary thread write, workers read 3029 // -------------------------------------------------------------------------- 3030 KMP_ALIGN_CACHE void **t_argv; 3031 int t_argc; 3032 int t_nproc; // number of threads in team 3033 microtask_t t_pkfn; 3034 launch_t t_invoke; // procedure to launch the microtask 3035 3036 #if OMPT_SUPPORT 3037 ompt_team_info_t ompt_team_info; 3038 ompt_lw_taskteam_t *ompt_serialized_team_info; 3039 #endif 3040 3041 #if KMP_ARCH_X86 || KMP_ARCH_X86_64 3042 kmp_int8 t_fp_control_saved; 3043 kmp_int8 t_pad2b; 3044 kmp_int16 t_x87_fpu_control_word; // FP control regs 3045 kmp_uint32 t_mxcsr; 3046 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */ 3047 3048 void *t_inline_argv[KMP_INLINE_ARGV_ENTRIES]; 3049 3050 KMP_ALIGN_CACHE kmp_info_t **t_threads; 3051 kmp_taskdata_t 3052 *t_implicit_task_taskdata; // Taskdata for the thread's implicit task 3053 int t_level; // nested parallel level 3054 3055 KMP_ALIGN_CACHE int t_max_argc; 3056 int t_max_nproc; // max threads this team can handle (dynamically expandable) 3057 int t_serialized; // levels deep of serialized teams 3058 dispatch_shared_info_t *t_disp_buffer; // buffers for dispatch system 3059 int t_id; // team's id, assigned by debugger. 3060 int t_active_level; // nested active parallel level 3061 kmp_r_sched_t t_sched; // run-time schedule for the team 3062 #if KMP_AFFINITY_SUPPORTED 3063 int t_first_place; // first & last place in parent thread's partition. 3064 int t_last_place; // Restore these values to primary thread after par region. 3065 #endif // KMP_AFFINITY_SUPPORTED 3066 int t_display_affinity; 3067 int t_size_changed; // team size was changed?: 0: no, 1: yes, -1: changed via 3068 // omp_set_num_threads() call 3069 omp_allocator_handle_t t_def_allocator; /* default allocator */ 3070 3071 // Read/write by workers as well 3072 #if (KMP_ARCH_X86 || KMP_ARCH_X86_64) 3073 // Using CACHE_LINE=64 reduces memory footprint, but causes a big perf 3074 // regression of epcc 'parallel' and 'barrier' on fxe256lin01. This extra 3075 // padding serves to fix the performance of epcc 'parallel' and 'barrier' when 3076 // CACHE_LINE=64. TODO: investigate more and get rid if this padding. 3077 char dummy_padding[1024]; 3078 #endif 3079 // Internal control stack for additional nested teams. 3080 KMP_ALIGN_CACHE kmp_internal_control_t *t_control_stack_top; 3081 // for SERIALIZED teams nested 2 or more levels deep 3082 // typed flag to store request state of cancellation 3083 std::atomic<kmp_int32> t_cancel_request; 3084 int t_master_active; // save on fork, restore on join 3085 void *t_copypriv_data; // team specific pointer to copyprivate data array 3086 #if KMP_OS_WINDOWS 3087 std::atomic<kmp_uint32> t_copyin_counter; 3088 #endif 3089 #if USE_ITT_BUILD 3090 void *t_stack_id; // team specific stack stitching id (for ittnotify) 3091 #endif /* USE_ITT_BUILD */ 3092 distributedBarrier *b; // Distributed barrier data associated with team 3093 } kmp_base_team_t; 3094 3095 union KMP_ALIGN_CACHE kmp_team { 3096 kmp_base_team_t t; 3097 double t_align; /* use worst case alignment */ 3098 char t_pad[KMP_PAD(kmp_base_team_t, CACHE_LINE)]; 3099 }; 3100 3101 typedef union KMP_ALIGN_CACHE kmp_time_global { 3102 double dt_align; /* use worst case alignment */ 3103 char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)]; 3104 kmp_base_data_t dt; 3105 } kmp_time_global_t; 3106 3107 typedef struct kmp_base_global { 3108 /* cache-aligned */ 3109 kmp_time_global_t g_time; 3110 3111 /* non cache-aligned */ 3112 volatile int g_abort; 3113 volatile int g_done; 3114 3115 int g_dynamic; 3116 enum dynamic_mode g_dynamic_mode; 3117 } kmp_base_global_t; 3118 3119 typedef union KMP_ALIGN_CACHE kmp_global { 3120 kmp_base_global_t g; 3121 double g_align; /* use worst case alignment */ 3122 char g_pad[KMP_PAD(kmp_base_global_t, CACHE_LINE)]; 3123 } kmp_global_t; 3124 3125 typedef struct kmp_base_root { 3126 // TODO: GEH - combine r_active with r_in_parallel then r_active == 3127 // (r_in_parallel>= 0) 3128 // TODO: GEH - then replace r_active with t_active_levels if we can to reduce 3129 // the synch overhead or keeping r_active 3130 volatile int r_active; /* TRUE if some region in a nest has > 1 thread */ 3131 // keeps a count of active parallel regions per root 3132 std::atomic<int> r_in_parallel; 3133 // GEH: This is misnamed, should be r_active_levels 3134 kmp_team_t *r_root_team; 3135 kmp_team_t *r_hot_team; 3136 kmp_info_t *r_uber_thread; 3137 kmp_lock_t r_begin_lock; 3138 volatile int r_begin; 3139 int r_blocktime; /* blocktime for this root and descendants */ 3140 #if KMP_AFFINITY_SUPPORTED 3141 int r_affinity_assigned; 3142 #endif // KMP_AFFINITY_SUPPORTED 3143 } kmp_base_root_t; 3144 3145 typedef union KMP_ALIGN_CACHE kmp_root { 3146 kmp_base_root_t r; 3147 double r_align; /* use worst case alignment */ 3148 char r_pad[KMP_PAD(kmp_base_root_t, CACHE_LINE)]; 3149 } kmp_root_t; 3150 3151 struct fortran_inx_info { 3152 kmp_int32 data; 3153 }; 3154 3155 // This list type exists to hold old __kmp_threads arrays so that 3156 // old references to them may complete while reallocation takes place when 3157 // expanding the array. The items in this list are kept alive until library 3158 // shutdown. 3159 typedef struct kmp_old_threads_list_t { 3160 kmp_info_t **threads; 3161 struct kmp_old_threads_list_t *next; 3162 } kmp_old_threads_list_t; 3163 3164 /* ------------------------------------------------------------------------ */ 3165 3166 extern int __kmp_settings; 3167 extern int __kmp_duplicate_library_ok; 3168 #if USE_ITT_BUILD 3169 extern int __kmp_forkjoin_frames; 3170 extern int __kmp_forkjoin_frames_mode; 3171 #endif 3172 extern PACKED_REDUCTION_METHOD_T __kmp_force_reduction_method; 3173 extern int __kmp_determ_red; 3174 3175 #ifdef KMP_DEBUG 3176 extern int kmp_a_debug; 3177 extern int kmp_b_debug; 3178 extern int kmp_c_debug; 3179 extern int kmp_d_debug; 3180 extern int kmp_e_debug; 3181 extern int kmp_f_debug; 3182 #endif /* KMP_DEBUG */ 3183 3184 /* For debug information logging using rotating buffer */ 3185 #define KMP_DEBUG_BUF_LINES_INIT 512 3186 #define KMP_DEBUG_BUF_LINES_MIN 1 3187 3188 #define KMP_DEBUG_BUF_CHARS_INIT 128 3189 #define KMP_DEBUG_BUF_CHARS_MIN 2 3190 3191 extern int 3192 __kmp_debug_buf; /* TRUE means use buffer, FALSE means print to stderr */ 3193 extern int __kmp_debug_buf_lines; /* How many lines of debug stored in buffer */ 3194 extern int 3195 __kmp_debug_buf_chars; /* How many characters allowed per line in buffer */ 3196 extern int __kmp_debug_buf_atomic; /* TRUE means use atomic update of buffer 3197 entry pointer */ 3198 3199 extern char *__kmp_debug_buffer; /* Debug buffer itself */ 3200 extern std::atomic<int> __kmp_debug_count; /* Counter for number of lines 3201 printed in buffer so far */ 3202 extern int __kmp_debug_buf_warn_chars; /* Keep track of char increase 3203 recommended in warnings */ 3204 /* end rotating debug buffer */ 3205 3206 #ifdef KMP_DEBUG 3207 extern int __kmp_par_range; /* +1 => only go par for constructs in range */ 3208 3209 #define KMP_PAR_RANGE_ROUTINE_LEN 1024 3210 extern char __kmp_par_range_routine[KMP_PAR_RANGE_ROUTINE_LEN]; 3211 #define KMP_PAR_RANGE_FILENAME_LEN 1024 3212 extern char __kmp_par_range_filename[KMP_PAR_RANGE_FILENAME_LEN]; 3213 extern int __kmp_par_range_lb; 3214 extern int __kmp_par_range_ub; 3215 #endif 3216 3217 /* For printing out dynamic storage map for threads and teams */ 3218 extern int 3219 __kmp_storage_map; /* True means print storage map for threads and teams */ 3220 extern int __kmp_storage_map_verbose; /* True means storage map includes 3221 placement info */ 3222 extern int __kmp_storage_map_verbose_specified; 3223 3224 #if KMP_ARCH_X86 || KMP_ARCH_X86_64 3225 extern kmp_cpuinfo_t __kmp_cpuinfo; 3226 static inline bool __kmp_is_hybrid_cpu() { return __kmp_cpuinfo.flags.hybrid; } 3227 #elif KMP_OS_DARWIN && KMP_ARCH_AARCH64 3228 static inline bool __kmp_is_hybrid_cpu() { return true; } 3229 #else 3230 static inline bool __kmp_is_hybrid_cpu() { return false; } 3231 #endif 3232 3233 extern volatile int __kmp_init_serial; 3234 extern volatile int __kmp_init_gtid; 3235 extern volatile int __kmp_init_common; 3236 extern volatile int __kmp_need_register_serial; 3237 extern volatile int __kmp_init_middle; 3238 extern volatile int __kmp_init_parallel; 3239 #if KMP_USE_MONITOR 3240 extern volatile int __kmp_init_monitor; 3241 #endif 3242 extern volatile int __kmp_init_user_locks; 3243 extern volatile int __kmp_init_hidden_helper_threads; 3244 extern int __kmp_init_counter; 3245 extern int __kmp_root_counter; 3246 extern int __kmp_version; 3247 3248 /* list of address of allocated caches for commons */ 3249 extern kmp_cached_addr_t *__kmp_threadpriv_cache_list; 3250 3251 /* Barrier algorithm types and options */ 3252 extern kmp_uint32 __kmp_barrier_gather_bb_dflt; 3253 extern kmp_uint32 __kmp_barrier_release_bb_dflt; 3254 extern kmp_bar_pat_e __kmp_barrier_gather_pat_dflt; 3255 extern kmp_bar_pat_e __kmp_barrier_release_pat_dflt; 3256 extern kmp_uint32 __kmp_barrier_gather_branch_bits[bs_last_barrier]; 3257 extern kmp_uint32 __kmp_barrier_release_branch_bits[bs_last_barrier]; 3258 extern kmp_bar_pat_e __kmp_barrier_gather_pattern[bs_last_barrier]; 3259 extern kmp_bar_pat_e __kmp_barrier_release_pattern[bs_last_barrier]; 3260 extern char const *__kmp_barrier_branch_bit_env_name[bs_last_barrier]; 3261 extern char const *__kmp_barrier_pattern_env_name[bs_last_barrier]; 3262 extern char const *__kmp_barrier_type_name[bs_last_barrier]; 3263 extern char const *__kmp_barrier_pattern_name[bp_last_bar]; 3264 3265 /* Global Locks */ 3266 extern kmp_bootstrap_lock_t __kmp_initz_lock; /* control initialization */ 3267 extern kmp_bootstrap_lock_t __kmp_forkjoin_lock; /* control fork/join access */ 3268 extern kmp_bootstrap_lock_t __kmp_task_team_lock; 3269 extern kmp_bootstrap_lock_t 3270 __kmp_exit_lock; /* exit() is not always thread-safe */ 3271 #if KMP_USE_MONITOR 3272 extern kmp_bootstrap_lock_t 3273 __kmp_monitor_lock; /* control monitor thread creation */ 3274 #endif 3275 extern kmp_bootstrap_lock_t 3276 __kmp_tp_cached_lock; /* used for the hack to allow threadprivate cache and 3277 __kmp_threads expansion to co-exist */ 3278 3279 extern kmp_lock_t __kmp_global_lock; /* control OS/global access */ 3280 extern kmp_queuing_lock_t __kmp_dispatch_lock; /* control dispatch access */ 3281 extern kmp_lock_t __kmp_debug_lock; /* control I/O access for KMP_DEBUG */ 3282 3283 extern enum library_type __kmp_library; 3284 3285 extern enum sched_type __kmp_sched; /* default runtime scheduling */ 3286 extern enum sched_type __kmp_static; /* default static scheduling method */ 3287 extern enum sched_type __kmp_guided; /* default guided scheduling method */ 3288 extern enum sched_type __kmp_auto; /* default auto scheduling method */ 3289 extern int __kmp_chunk; /* default runtime chunk size */ 3290 extern int __kmp_force_monotonic; /* whether monotonic scheduling forced */ 3291 3292 extern size_t __kmp_stksize; /* stack size per thread */ 3293 #if KMP_USE_MONITOR 3294 extern size_t __kmp_monitor_stksize; /* stack size for monitor thread */ 3295 #endif 3296 extern size_t __kmp_stkoffset; /* stack offset per thread */ 3297 extern int __kmp_stkpadding; /* Should we pad root thread(s) stack */ 3298 3299 extern size_t 3300 __kmp_malloc_pool_incr; /* incremental size of pool for kmp_malloc() */ 3301 extern int __kmp_env_stksize; /* was KMP_STACKSIZE specified? */ 3302 extern int __kmp_env_blocktime; /* was KMP_BLOCKTIME specified? */ 3303 extern int __kmp_env_checks; /* was KMP_CHECKS specified? */ 3304 extern int __kmp_env_consistency_check; // was KMP_CONSISTENCY_CHECK specified? 3305 extern int __kmp_generate_warnings; /* should we issue warnings? */ 3306 extern int __kmp_reserve_warn; /* have we issued reserve_threads warning? */ 3307 3308 #ifdef DEBUG_SUSPEND 3309 extern int __kmp_suspend_count; /* count inside __kmp_suspend_template() */ 3310 #endif 3311 3312 extern kmp_int32 __kmp_use_yield; 3313 extern kmp_int32 __kmp_use_yield_exp_set; 3314 extern kmp_uint32 __kmp_yield_init; 3315 extern kmp_uint32 __kmp_yield_next; 3316 extern kmp_uint64 __kmp_pause_init; 3317 3318 /* ------------------------------------------------------------------------- */ 3319 extern int __kmp_allThreadsSpecified; 3320 3321 extern size_t __kmp_align_alloc; 3322 /* following data protected by initialization routines */ 3323 extern int __kmp_xproc; /* number of processors in the system */ 3324 extern int __kmp_avail_proc; /* number of processors available to the process */ 3325 extern size_t __kmp_sys_min_stksize; /* system-defined minimum stack size */ 3326 extern int __kmp_sys_max_nth; /* system-imposed maximum number of threads */ 3327 // maximum total number of concurrently-existing threads on device 3328 extern int __kmp_max_nth; 3329 // maximum total number of concurrently-existing threads in a contention group 3330 extern int __kmp_cg_max_nth; 3331 extern int __kmp_teams_max_nth; // max threads used in a teams construct 3332 extern int __kmp_threads_capacity; /* capacity of the arrays __kmp_threads and 3333 __kmp_root */ 3334 extern int __kmp_dflt_team_nth; /* default number of threads in a parallel 3335 region a la OMP_NUM_THREADS */ 3336 extern int __kmp_dflt_team_nth_ub; /* upper bound on "" determined at serial 3337 initialization */ 3338 extern int __kmp_tp_capacity; /* capacity of __kmp_threads if threadprivate is 3339 used (fixed) */ 3340 extern int __kmp_tp_cached; /* whether threadprivate cache has been created 3341 (__kmpc_threadprivate_cached()) */ 3342 extern int __kmp_dflt_blocktime; /* number of milliseconds to wait before 3343 blocking (env setting) */ 3344 extern bool __kmp_wpolicy_passive; /* explicitly set passive wait policy */ 3345 #if KMP_USE_MONITOR 3346 extern int 3347 __kmp_monitor_wakeups; /* number of times monitor wakes up per second */ 3348 extern int __kmp_bt_intervals; /* number of monitor timestamp intervals before 3349 blocking */ 3350 #endif 3351 #ifdef KMP_ADJUST_BLOCKTIME 3352 extern int __kmp_zero_bt; /* whether blocktime has been forced to zero */ 3353 #endif /* KMP_ADJUST_BLOCKTIME */ 3354 #ifdef KMP_DFLT_NTH_CORES 3355 extern int __kmp_ncores; /* Total number of cores for threads placement */ 3356 #endif 3357 /* Number of millisecs to delay on abort for Intel(R) VTune(TM) tools */ 3358 extern int __kmp_abort_delay; 3359 3360 extern int __kmp_need_register_atfork_specified; 3361 extern int __kmp_need_register_atfork; /* At initialization, call pthread_atfork 3362 to install fork handler */ 3363 extern int __kmp_gtid_mode; /* Method of getting gtid, values: 3364 0 - not set, will be set at runtime 3365 1 - using stack search 3366 2 - dynamic TLS (pthread_getspecific(Linux* OS/OS 3367 X*) or TlsGetValue(Windows* OS)) 3368 3 - static TLS (__declspec(thread) __kmp_gtid), 3369 Linux* OS .so only. */ 3370 extern int 3371 __kmp_adjust_gtid_mode; /* If true, adjust method based on #threads */ 3372 #ifdef KMP_TDATA_GTID 3373 extern KMP_THREAD_LOCAL int __kmp_gtid; 3374 #endif 3375 extern int __kmp_tls_gtid_min; /* #threads below which use sp search for gtid */ 3376 extern int __kmp_foreign_tp; // If true, separate TP var for each foreign thread 3377 #if KMP_ARCH_X86 || KMP_ARCH_X86_64 3378 extern int __kmp_inherit_fp_control; // copy fp creg(s) parent->workers at fork 3379 extern kmp_int16 __kmp_init_x87_fpu_control_word; // init thread's FP ctrl reg 3380 extern kmp_uint32 __kmp_init_mxcsr; /* init thread's mxscr */ 3381 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */ 3382 3383 // max_active_levels for nested parallelism enabled by default via 3384 // OMP_MAX_ACTIVE_LEVELS, OMP_NESTED, OMP_NUM_THREADS, and OMP_PROC_BIND 3385 extern int __kmp_dflt_max_active_levels; 3386 // Indicates whether value of __kmp_dflt_max_active_levels was already 3387 // explicitly set by OMP_MAX_ACTIVE_LEVELS or OMP_NESTED=false 3388 extern bool __kmp_dflt_max_active_levels_set; 3389 extern int __kmp_dispatch_num_buffers; /* max possible dynamic loops in 3390 concurrent execution per team */ 3391 #if KMP_NESTED_HOT_TEAMS 3392 extern int __kmp_hot_teams_mode; 3393 extern int __kmp_hot_teams_max_level; 3394 #endif 3395 3396 #if KMP_OS_LINUX 3397 extern enum clock_function_type __kmp_clock_function; 3398 extern int __kmp_clock_function_param; 3399 #endif /* KMP_OS_LINUX */ 3400 3401 #if KMP_MIC_SUPPORTED 3402 extern enum mic_type __kmp_mic_type; 3403 #endif 3404 3405 #ifdef USE_LOAD_BALANCE 3406 extern double __kmp_load_balance_interval; // load balance algorithm interval 3407 #endif /* USE_LOAD_BALANCE */ 3408 3409 // OpenMP 3.1 - Nested num threads array 3410 typedef struct kmp_nested_nthreads_t { 3411 int *nth; 3412 int size; 3413 int used; 3414 } kmp_nested_nthreads_t; 3415 3416 extern kmp_nested_nthreads_t __kmp_nested_nth; 3417 3418 #if KMP_USE_ADAPTIVE_LOCKS 3419 3420 // Parameters for the speculative lock backoff system. 3421 struct kmp_adaptive_backoff_params_t { 3422 // Number of soft retries before it counts as a hard retry. 3423 kmp_uint32 max_soft_retries; 3424 // Badness is a bit mask : 0,1,3,7,15,... on each hard failure we move one to 3425 // the right 3426 kmp_uint32 max_badness; 3427 }; 3428 3429 extern kmp_adaptive_backoff_params_t __kmp_adaptive_backoff_params; 3430 3431 #if KMP_DEBUG_ADAPTIVE_LOCKS 3432 extern const char *__kmp_speculative_statsfile; 3433 #endif 3434 3435 #endif // KMP_USE_ADAPTIVE_LOCKS 3436 3437 extern int __kmp_display_env; /* TRUE or FALSE */ 3438 extern int __kmp_display_env_verbose; /* TRUE if OMP_DISPLAY_ENV=VERBOSE */ 3439 extern int __kmp_omp_cancellation; /* TRUE or FALSE */ 3440 extern int __kmp_nteams; 3441 extern int __kmp_teams_thread_limit; 3442 3443 /* ------------------------------------------------------------------------- */ 3444 3445 /* the following are protected by the fork/join lock */ 3446 /* write: lock read: anytime */ 3447 extern kmp_info_t **__kmp_threads; /* Descriptors for the threads */ 3448 /* Holds old arrays of __kmp_threads until library shutdown */ 3449 extern kmp_old_threads_list_t *__kmp_old_threads_list; 3450 /* read/write: lock */ 3451 extern volatile kmp_team_t *__kmp_team_pool; 3452 extern volatile kmp_info_t *__kmp_thread_pool; 3453 extern kmp_info_t *__kmp_thread_pool_insert_pt; 3454 3455 // total num threads reachable from some root thread including all root threads 3456 extern volatile int __kmp_nth; 3457 /* total number of threads reachable from some root thread including all root 3458 threads, and those in the thread pool */ 3459 extern volatile int __kmp_all_nth; 3460 extern std::atomic<int> __kmp_thread_pool_active_nth; 3461 3462 extern kmp_root_t **__kmp_root; /* root of thread hierarchy */ 3463 /* end data protected by fork/join lock */ 3464 /* ------------------------------------------------------------------------- */ 3465 3466 #define __kmp_get_gtid() __kmp_get_global_thread_id() 3467 #define __kmp_entry_gtid() __kmp_get_global_thread_id_reg() 3468 #define __kmp_get_tid() (__kmp_tid_from_gtid(__kmp_get_gtid())) 3469 #define __kmp_get_team() (__kmp_threads[(__kmp_get_gtid())]->th.th_team) 3470 #define __kmp_get_thread() (__kmp_thread_from_gtid(__kmp_get_gtid())) 3471 3472 // AT: Which way is correct? 3473 // AT: 1. nproc = __kmp_threads[ ( gtid ) ] -> th.th_team -> t.t_nproc; 3474 // AT: 2. nproc = __kmp_threads[ ( gtid ) ] -> th.th_team_nproc; 3475 #define __kmp_get_team_num_threads(gtid) \ 3476 (__kmp_threads[(gtid)]->th.th_team->t.t_nproc) 3477 3478 static inline bool KMP_UBER_GTID(int gtid) { 3479 KMP_DEBUG_ASSERT(gtid >= KMP_GTID_MIN); 3480 KMP_DEBUG_ASSERT(gtid < __kmp_threads_capacity); 3481 return (gtid >= 0 && __kmp_root[gtid] && __kmp_threads[gtid] && 3482 __kmp_threads[gtid] == __kmp_root[gtid]->r.r_uber_thread); 3483 } 3484 3485 static inline int __kmp_tid_from_gtid(int gtid) { 3486 KMP_DEBUG_ASSERT(gtid >= 0); 3487 return __kmp_threads[gtid]->th.th_info.ds.ds_tid; 3488 } 3489 3490 static inline int __kmp_gtid_from_tid(int tid, const kmp_team_t *team) { 3491 KMP_DEBUG_ASSERT(tid >= 0 && team); 3492 return team->t.t_threads[tid]->th.th_info.ds.ds_gtid; 3493 } 3494 3495 static inline int __kmp_gtid_from_thread(const kmp_info_t *thr) { 3496 KMP_DEBUG_ASSERT(thr); 3497 return thr->th.th_info.ds.ds_gtid; 3498 } 3499 3500 static inline kmp_info_t *__kmp_thread_from_gtid(int gtid) { 3501 KMP_DEBUG_ASSERT(gtid >= 0); 3502 return __kmp_threads[gtid]; 3503 } 3504 3505 static inline kmp_team_t *__kmp_team_from_gtid(int gtid) { 3506 KMP_DEBUG_ASSERT(gtid >= 0); 3507 return __kmp_threads[gtid]->th.th_team; 3508 } 3509 3510 static inline void __kmp_assert_valid_gtid(kmp_int32 gtid) { 3511 if (UNLIKELY(gtid < 0 || gtid >= __kmp_threads_capacity)) 3512 KMP_FATAL(ThreadIdentInvalid); 3513 } 3514 3515 #if KMP_HAVE_MWAIT || KMP_HAVE_UMWAIT 3516 extern int __kmp_user_level_mwait; // TRUE or FALSE; from KMP_USER_LEVEL_MWAIT 3517 extern int __kmp_umwait_enabled; // Runtime check if user-level mwait enabled 3518 extern int __kmp_mwait_enabled; // Runtime check if ring3 mwait is enabled 3519 extern int __kmp_mwait_hints; // Hints to pass in to mwait 3520 #endif 3521 3522 #if KMP_HAVE_UMWAIT 3523 extern int __kmp_waitpkg_enabled; // Runtime check if waitpkg exists 3524 extern int __kmp_tpause_state; // 0 (default), 1=C0.1, 2=C0.2; from KMP_TPAUSE 3525 extern int __kmp_tpause_hint; // 1=C0.1 (default), 0=C0.2; from KMP_TPAUSE 3526 extern int __kmp_tpause_enabled; // 0 (default), 1 (KMP_TPAUSE is non-zero) 3527 #endif 3528 3529 /* ------------------------------------------------------------------------- */ 3530 3531 extern kmp_global_t __kmp_global; /* global status */ 3532 3533 extern kmp_info_t __kmp_monitor; 3534 // For Debugging Support Library 3535 extern std::atomic<kmp_int32> __kmp_team_counter; 3536 // For Debugging Support Library 3537 extern std::atomic<kmp_int32> __kmp_task_counter; 3538 3539 #if USE_DEBUGGER 3540 #define _KMP_GEN_ID(counter) \ 3541 (__kmp_debugging ? KMP_ATOMIC_INC(&counter) + 1 : ~0) 3542 #else 3543 #define _KMP_GEN_ID(counter) (~0) 3544 #endif /* USE_DEBUGGER */ 3545 3546 #define KMP_GEN_TASK_ID() _KMP_GEN_ID(__kmp_task_counter) 3547 #define KMP_GEN_TEAM_ID() _KMP_GEN_ID(__kmp_team_counter) 3548 3549 /* ------------------------------------------------------------------------ */ 3550 3551 extern void __kmp_print_storage_map_gtid(int gtid, void *p1, void *p2, 3552 size_t size, char const *format, ...); 3553 3554 extern void __kmp_serial_initialize(void); 3555 extern void __kmp_middle_initialize(void); 3556 extern void __kmp_parallel_initialize(void); 3557 3558 extern void __kmp_internal_begin(void); 3559 extern void __kmp_internal_end_library(int gtid); 3560 extern void __kmp_internal_end_thread(int gtid); 3561 extern void __kmp_internal_end_atexit(void); 3562 extern void __kmp_internal_end_dtor(void); 3563 extern void __kmp_internal_end_dest(void *); 3564 3565 extern int __kmp_register_root(int initial_thread); 3566 extern void __kmp_unregister_root(int gtid); 3567 extern void __kmp_unregister_library(void); // called by __kmp_internal_end() 3568 3569 extern int __kmp_ignore_mppbeg(void); 3570 extern int __kmp_ignore_mppend(void); 3571 3572 extern int __kmp_enter_single(int gtid, ident_t *id_ref, int push_ws); 3573 extern void __kmp_exit_single(int gtid); 3574 3575 extern void __kmp_parallel_deo(int *gtid_ref, int *cid_ref, ident_t *loc_ref); 3576 extern void __kmp_parallel_dxo(int *gtid_ref, int *cid_ref, ident_t *loc_ref); 3577 3578 #ifdef USE_LOAD_BALANCE 3579 extern int __kmp_get_load_balance(int); 3580 #endif 3581 3582 extern int __kmp_get_global_thread_id(void); 3583 extern int __kmp_get_global_thread_id_reg(void); 3584 extern void __kmp_exit_thread(int exit_status); 3585 extern void __kmp_abort(char const *format, ...); 3586 extern void __kmp_abort_thread(void); 3587 KMP_NORETURN extern void __kmp_abort_process(void); 3588 extern void __kmp_warn(char const *format, ...); 3589 3590 extern void __kmp_set_num_threads(int new_nth, int gtid); 3591 3592 // Returns current thread (pointer to kmp_info_t). Current thread *must* be 3593 // registered. 3594 static inline kmp_info_t *__kmp_entry_thread() { 3595 int gtid = __kmp_entry_gtid(); 3596 3597 return __kmp_threads[gtid]; 3598 } 3599 3600 extern void __kmp_set_max_active_levels(int gtid, int new_max_active_levels); 3601 extern int __kmp_get_max_active_levels(int gtid); 3602 extern int __kmp_get_ancestor_thread_num(int gtid, int level); 3603 extern int __kmp_get_team_size(int gtid, int level); 3604 extern void __kmp_set_schedule(int gtid, kmp_sched_t new_sched, int chunk); 3605 extern void __kmp_get_schedule(int gtid, kmp_sched_t *sched, int *chunk); 3606 3607 extern unsigned short __kmp_get_random(kmp_info_t *thread); 3608 extern void __kmp_init_random(kmp_info_t *thread); 3609 3610 extern kmp_r_sched_t __kmp_get_schedule_global(void); 3611 extern void __kmp_adjust_num_threads(int new_nproc); 3612 extern void __kmp_check_stksize(size_t *val); 3613 3614 extern void *___kmp_allocate(size_t size KMP_SRC_LOC_DECL); 3615 extern void *___kmp_page_allocate(size_t size KMP_SRC_LOC_DECL); 3616 extern void ___kmp_free(void *ptr KMP_SRC_LOC_DECL); 3617 #define __kmp_allocate(size) ___kmp_allocate((size)KMP_SRC_LOC_CURR) 3618 #define __kmp_page_allocate(size) ___kmp_page_allocate((size)KMP_SRC_LOC_CURR) 3619 #define __kmp_free(ptr) ___kmp_free((ptr)KMP_SRC_LOC_CURR) 3620 3621 #if USE_FAST_MEMORY 3622 extern void *___kmp_fast_allocate(kmp_info_t *this_thr, 3623 size_t size KMP_SRC_LOC_DECL); 3624 extern void ___kmp_fast_free(kmp_info_t *this_thr, void *ptr KMP_SRC_LOC_DECL); 3625 extern void __kmp_free_fast_memory(kmp_info_t *this_thr); 3626 extern void __kmp_initialize_fast_memory(kmp_info_t *this_thr); 3627 #define __kmp_fast_allocate(this_thr, size) \ 3628 ___kmp_fast_allocate((this_thr), (size)KMP_SRC_LOC_CURR) 3629 #define __kmp_fast_free(this_thr, ptr) \ 3630 ___kmp_fast_free((this_thr), (ptr)KMP_SRC_LOC_CURR) 3631 #endif 3632 3633 extern void *___kmp_thread_malloc(kmp_info_t *th, size_t size KMP_SRC_LOC_DECL); 3634 extern void *___kmp_thread_calloc(kmp_info_t *th, size_t nelem, 3635 size_t elsize KMP_SRC_LOC_DECL); 3636 extern void *___kmp_thread_realloc(kmp_info_t *th, void *ptr, 3637 size_t size KMP_SRC_LOC_DECL); 3638 extern void ___kmp_thread_free(kmp_info_t *th, void *ptr KMP_SRC_LOC_DECL); 3639 #define __kmp_thread_malloc(th, size) \ 3640 ___kmp_thread_malloc((th), (size)KMP_SRC_LOC_CURR) 3641 #define __kmp_thread_calloc(th, nelem, elsize) \ 3642 ___kmp_thread_calloc((th), (nelem), (elsize)KMP_SRC_LOC_CURR) 3643 #define __kmp_thread_realloc(th, ptr, size) \ 3644 ___kmp_thread_realloc((th), (ptr), (size)KMP_SRC_LOC_CURR) 3645 #define __kmp_thread_free(th, ptr) \ 3646 ___kmp_thread_free((th), (ptr)KMP_SRC_LOC_CURR) 3647 3648 extern void __kmp_push_num_threads(ident_t *loc, int gtid, int num_threads); 3649 3650 extern void __kmp_push_proc_bind(ident_t *loc, int gtid, 3651 kmp_proc_bind_t proc_bind); 3652 extern void __kmp_push_num_teams(ident_t *loc, int gtid, int num_teams, 3653 int num_threads); 3654 extern void __kmp_push_num_teams_51(ident_t *loc, int gtid, int num_teams_lb, 3655 int num_teams_ub, int num_threads); 3656 3657 extern void __kmp_yield(); 3658 3659 extern void __kmpc_dispatch_init_4(ident_t *loc, kmp_int32 gtid, 3660 enum sched_type schedule, kmp_int32 lb, 3661 kmp_int32 ub, kmp_int32 st, kmp_int32 chunk); 3662 extern void __kmpc_dispatch_init_4u(ident_t *loc, kmp_int32 gtid, 3663 enum sched_type schedule, kmp_uint32 lb, 3664 kmp_uint32 ub, kmp_int32 st, 3665 kmp_int32 chunk); 3666 extern void __kmpc_dispatch_init_8(ident_t *loc, kmp_int32 gtid, 3667 enum sched_type schedule, kmp_int64 lb, 3668 kmp_int64 ub, kmp_int64 st, kmp_int64 chunk); 3669 extern void __kmpc_dispatch_init_8u(ident_t *loc, kmp_int32 gtid, 3670 enum sched_type schedule, kmp_uint64 lb, 3671 kmp_uint64 ub, kmp_int64 st, 3672 kmp_int64 chunk); 3673 3674 extern int __kmpc_dispatch_next_4(ident_t *loc, kmp_int32 gtid, 3675 kmp_int32 *p_last, kmp_int32 *p_lb, 3676 kmp_int32 *p_ub, kmp_int32 *p_st); 3677 extern int __kmpc_dispatch_next_4u(ident_t *loc, kmp_int32 gtid, 3678 kmp_int32 *p_last, kmp_uint32 *p_lb, 3679 kmp_uint32 *p_ub, kmp_int32 *p_st); 3680 extern int __kmpc_dispatch_next_8(ident_t *loc, kmp_int32 gtid, 3681 kmp_int32 *p_last, kmp_int64 *p_lb, 3682 kmp_int64 *p_ub, kmp_int64 *p_st); 3683 extern int __kmpc_dispatch_next_8u(ident_t *loc, kmp_int32 gtid, 3684 kmp_int32 *p_last, kmp_uint64 *p_lb, 3685 kmp_uint64 *p_ub, kmp_int64 *p_st); 3686 3687 extern void __kmpc_dispatch_fini_4(ident_t *loc, kmp_int32 gtid); 3688 extern void __kmpc_dispatch_fini_8(ident_t *loc, kmp_int32 gtid); 3689 extern void __kmpc_dispatch_fini_4u(ident_t *loc, kmp_int32 gtid); 3690 extern void __kmpc_dispatch_fini_8u(ident_t *loc, kmp_int32 gtid); 3691 3692 #ifdef KMP_GOMP_COMPAT 3693 3694 extern void __kmp_aux_dispatch_init_4(ident_t *loc, kmp_int32 gtid, 3695 enum sched_type schedule, kmp_int32 lb, 3696 kmp_int32 ub, kmp_int32 st, 3697 kmp_int32 chunk, int push_ws); 3698 extern void __kmp_aux_dispatch_init_4u(ident_t *loc, kmp_int32 gtid, 3699 enum sched_type schedule, kmp_uint32 lb, 3700 kmp_uint32 ub, kmp_int32 st, 3701 kmp_int32 chunk, int push_ws); 3702 extern void __kmp_aux_dispatch_init_8(ident_t *loc, kmp_int32 gtid, 3703 enum sched_type schedule, kmp_int64 lb, 3704 kmp_int64 ub, kmp_int64 st, 3705 kmp_int64 chunk, int push_ws); 3706 extern void __kmp_aux_dispatch_init_8u(ident_t *loc, kmp_int32 gtid, 3707 enum sched_type schedule, kmp_uint64 lb, 3708 kmp_uint64 ub, kmp_int64 st, 3709 kmp_int64 chunk, int push_ws); 3710 extern void __kmp_aux_dispatch_fini_chunk_4(ident_t *loc, kmp_int32 gtid); 3711 extern void __kmp_aux_dispatch_fini_chunk_8(ident_t *loc, kmp_int32 gtid); 3712 extern void __kmp_aux_dispatch_fini_chunk_4u(ident_t *loc, kmp_int32 gtid); 3713 extern void __kmp_aux_dispatch_fini_chunk_8u(ident_t *loc, kmp_int32 gtid); 3714 3715 #endif /* KMP_GOMP_COMPAT */ 3716 3717 extern kmp_uint32 __kmp_eq_4(kmp_uint32 value, kmp_uint32 checker); 3718 extern kmp_uint32 __kmp_neq_4(kmp_uint32 value, kmp_uint32 checker); 3719 extern kmp_uint32 __kmp_lt_4(kmp_uint32 value, kmp_uint32 checker); 3720 extern kmp_uint32 __kmp_ge_4(kmp_uint32 value, kmp_uint32 checker); 3721 extern kmp_uint32 __kmp_le_4(kmp_uint32 value, kmp_uint32 checker); 3722 extern kmp_uint32 __kmp_wait_4(kmp_uint32 volatile *spinner, kmp_uint32 checker, 3723 kmp_uint32 (*pred)(kmp_uint32, kmp_uint32), 3724 void *obj); 3725 extern void __kmp_wait_4_ptr(void *spinner, kmp_uint32 checker, 3726 kmp_uint32 (*pred)(void *, kmp_uint32), void *obj); 3727 3728 extern void __kmp_wait_64(kmp_info_t *this_thr, kmp_flag_64<> *flag, 3729 int final_spin 3730 #if USE_ITT_BUILD 3731 , 3732 void *itt_sync_obj 3733 #endif 3734 ); 3735 extern void __kmp_release_64(kmp_flag_64<> *flag); 3736 3737 extern void __kmp_infinite_loop(void); 3738 3739 extern void __kmp_cleanup(void); 3740 3741 #if KMP_HANDLE_SIGNALS 3742 extern int __kmp_handle_signals; 3743 extern void __kmp_install_signals(int parallel_init); 3744 extern void __kmp_remove_signals(void); 3745 #endif 3746 3747 extern void __kmp_clear_system_time(void); 3748 extern void __kmp_read_system_time(double *delta); 3749 3750 extern void __kmp_check_stack_overlap(kmp_info_t *thr); 3751 3752 extern void __kmp_expand_host_name(char *buffer, size_t size); 3753 extern void __kmp_expand_file_name(char *result, size_t rlen, char *pattern); 3754 3755 #if KMP_ARCH_X86 || KMP_ARCH_X86_64 || (KMP_OS_WINDOWS && (KMP_ARCH_AARCH64 || KMP_ARCH_ARM)) 3756 extern void 3757 __kmp_initialize_system_tick(void); /* Initialize timer tick value */ 3758 #endif 3759 3760 extern void 3761 __kmp_runtime_initialize(void); /* machine specific initialization */ 3762 extern void __kmp_runtime_destroy(void); 3763 3764 #if KMP_AFFINITY_SUPPORTED 3765 extern char *__kmp_affinity_print_mask(char *buf, int buf_len, 3766 kmp_affin_mask_t *mask); 3767 extern kmp_str_buf_t *__kmp_affinity_str_buf_mask(kmp_str_buf_t *buf, 3768 kmp_affin_mask_t *mask); 3769 extern void __kmp_affinity_initialize(kmp_affinity_t &affinity); 3770 extern void __kmp_affinity_uninitialize(void); 3771 extern void __kmp_affinity_set_init_mask( 3772 int gtid, int isa_root); /* set affinity according to KMP_AFFINITY */ 3773 extern void __kmp_affinity_set_place(int gtid); 3774 extern void __kmp_affinity_determine_capable(const char *env_var); 3775 extern int __kmp_aux_set_affinity(void **mask); 3776 extern int __kmp_aux_get_affinity(void **mask); 3777 extern int __kmp_aux_get_affinity_max_proc(); 3778 extern int __kmp_aux_set_affinity_mask_proc(int proc, void **mask); 3779 extern int __kmp_aux_unset_affinity_mask_proc(int proc, void **mask); 3780 extern int __kmp_aux_get_affinity_mask_proc(int proc, void **mask); 3781 extern void __kmp_balanced_affinity(kmp_info_t *th, int team_size); 3782 #if KMP_OS_LINUX || KMP_OS_FREEBSD 3783 extern int kmp_set_thread_affinity_mask_initial(void); 3784 #endif 3785 static inline void __kmp_assign_root_init_mask() { 3786 int gtid = __kmp_entry_gtid(); 3787 kmp_root_t *r = __kmp_threads[gtid]->th.th_root; 3788 if (r->r.r_uber_thread == __kmp_threads[gtid] && !r->r.r_affinity_assigned) { 3789 __kmp_affinity_set_init_mask(gtid, TRUE); 3790 r->r.r_affinity_assigned = TRUE; 3791 } 3792 } 3793 static inline void __kmp_reset_root_init_mask(int gtid) { 3794 if (!KMP_AFFINITY_CAPABLE()) 3795 return; 3796 kmp_info_t *th = __kmp_threads[gtid]; 3797 kmp_root_t *r = th->th.th_root; 3798 if (r->r.r_uber_thread == th && r->r.r_affinity_assigned) { 3799 __kmp_set_system_affinity(__kmp_affin_origMask, FALSE); 3800 KMP_CPU_COPY(th->th.th_affin_mask, __kmp_affin_origMask); 3801 r->r.r_affinity_assigned = FALSE; 3802 } 3803 } 3804 #else /* KMP_AFFINITY_SUPPORTED */ 3805 #define __kmp_assign_root_init_mask() /* Nothing */ 3806 static inline void __kmp_reset_root_init_mask(int gtid) {} 3807 #endif /* KMP_AFFINITY_SUPPORTED */ 3808 // No need for KMP_AFFINITY_SUPPORTED guard as only one field in the 3809 // format string is for affinity, so platforms that do not support 3810 // affinity can still use the other fields, e.g., %n for num_threads 3811 extern size_t __kmp_aux_capture_affinity(int gtid, const char *format, 3812 kmp_str_buf_t *buffer); 3813 extern void __kmp_aux_display_affinity(int gtid, const char *format); 3814 3815 extern void __kmp_cleanup_hierarchy(); 3816 extern void __kmp_get_hierarchy(kmp_uint32 nproc, kmp_bstate_t *thr_bar); 3817 3818 #if KMP_USE_FUTEX 3819 3820 extern int __kmp_futex_determine_capable(void); 3821 3822 #endif // KMP_USE_FUTEX 3823 3824 extern void __kmp_gtid_set_specific(int gtid); 3825 extern int __kmp_gtid_get_specific(void); 3826 3827 extern double __kmp_read_cpu_time(void); 3828 3829 extern int __kmp_read_system_info(struct kmp_sys_info *info); 3830 3831 #if KMP_USE_MONITOR 3832 extern void __kmp_create_monitor(kmp_info_t *th); 3833 #endif 3834 3835 extern void *__kmp_launch_thread(kmp_info_t *thr); 3836 3837 extern void __kmp_create_worker(int gtid, kmp_info_t *th, size_t stack_size); 3838 3839 #if KMP_OS_WINDOWS 3840 extern int __kmp_still_running(kmp_info_t *th); 3841 extern int __kmp_is_thread_alive(kmp_info_t *th, DWORD *exit_val); 3842 extern void __kmp_free_handle(kmp_thread_t tHandle); 3843 #endif 3844 3845 #if KMP_USE_MONITOR 3846 extern void __kmp_reap_monitor(kmp_info_t *th); 3847 #endif 3848 extern void __kmp_reap_worker(kmp_info_t *th); 3849 extern void __kmp_terminate_thread(int gtid); 3850 3851 extern int __kmp_try_suspend_mx(kmp_info_t *th); 3852 extern void __kmp_lock_suspend_mx(kmp_info_t *th); 3853 extern void __kmp_unlock_suspend_mx(kmp_info_t *th); 3854 3855 extern void __kmp_elapsed(double *); 3856 extern void __kmp_elapsed_tick(double *); 3857 3858 extern void __kmp_enable(int old_state); 3859 extern void __kmp_disable(int *old_state); 3860 3861 extern void __kmp_thread_sleep(int millis); 3862 3863 extern void __kmp_common_initialize(void); 3864 extern void __kmp_common_destroy(void); 3865 extern void __kmp_common_destroy_gtid(int gtid); 3866 3867 #if KMP_OS_UNIX 3868 extern void __kmp_register_atfork(void); 3869 #endif 3870 extern void __kmp_suspend_initialize(void); 3871 extern void __kmp_suspend_initialize_thread(kmp_info_t *th); 3872 extern void __kmp_suspend_uninitialize_thread(kmp_info_t *th); 3873 3874 extern kmp_info_t *__kmp_allocate_thread(kmp_root_t *root, kmp_team_t *team, 3875 int tid); 3876 extern kmp_team_t * 3877 __kmp_allocate_team(kmp_root_t *root, int new_nproc, int max_nproc, 3878 #if OMPT_SUPPORT 3879 ompt_data_t ompt_parallel_data, 3880 #endif 3881 kmp_proc_bind_t proc_bind, kmp_internal_control_t *new_icvs, 3882 int argc USE_NESTED_HOT_ARG(kmp_info_t *thr)); 3883 extern void __kmp_free_thread(kmp_info_t *); 3884 extern void __kmp_free_team(kmp_root_t *, 3885 kmp_team_t *USE_NESTED_HOT_ARG(kmp_info_t *)); 3886 extern kmp_team_t *__kmp_reap_team(kmp_team_t *); 3887 3888 /* ------------------------------------------------------------------------ */ 3889 3890 extern void __kmp_initialize_bget(kmp_info_t *th); 3891 extern void __kmp_finalize_bget(kmp_info_t *th); 3892 3893 KMP_EXPORT void *kmpc_malloc(size_t size); 3894 KMP_EXPORT void *kmpc_aligned_malloc(size_t size, size_t alignment); 3895 KMP_EXPORT void *kmpc_calloc(size_t nelem, size_t elsize); 3896 KMP_EXPORT void *kmpc_realloc(void *ptr, size_t size); 3897 KMP_EXPORT void kmpc_free(void *ptr); 3898 3899 /* declarations for internal use */ 3900 3901 extern int __kmp_barrier(enum barrier_type bt, int gtid, int is_split, 3902 size_t reduce_size, void *reduce_data, 3903 void (*reduce)(void *, void *)); 3904 extern void __kmp_end_split_barrier(enum barrier_type bt, int gtid); 3905 extern int __kmp_barrier_gomp_cancel(int gtid); 3906 3907 /*! 3908 * Tell the fork call which compiler generated the fork call, and therefore how 3909 * to deal with the call. 3910 */ 3911 enum fork_context_e { 3912 fork_context_gnu, /**< Called from GNU generated code, so must not invoke the 3913 microtask internally. */ 3914 fork_context_intel, /**< Called from Intel generated code. */ 3915 fork_context_last 3916 }; 3917 extern int __kmp_fork_call(ident_t *loc, int gtid, 3918 enum fork_context_e fork_context, kmp_int32 argc, 3919 microtask_t microtask, launch_t invoker, 3920 kmp_va_list ap); 3921 3922 extern void __kmp_join_call(ident_t *loc, int gtid 3923 #if OMPT_SUPPORT 3924 , 3925 enum fork_context_e fork_context 3926 #endif 3927 , 3928 int exit_teams = 0); 3929 3930 extern void __kmp_serialized_parallel(ident_t *id, kmp_int32 gtid); 3931 extern void __kmp_internal_fork(ident_t *id, int gtid, kmp_team_t *team); 3932 extern void __kmp_internal_join(ident_t *id, int gtid, kmp_team_t *team); 3933 extern int __kmp_invoke_task_func(int gtid); 3934 extern void __kmp_run_before_invoked_task(int gtid, int tid, 3935 kmp_info_t *this_thr, 3936 kmp_team_t *team); 3937 extern void __kmp_run_after_invoked_task(int gtid, int tid, 3938 kmp_info_t *this_thr, 3939 kmp_team_t *team); 3940 3941 // should never have been exported 3942 KMP_EXPORT int __kmpc_invoke_task_func(int gtid); 3943 extern int __kmp_invoke_teams_master(int gtid); 3944 extern void __kmp_teams_master(int gtid); 3945 extern int __kmp_aux_get_team_num(); 3946 extern int __kmp_aux_get_num_teams(); 3947 extern void __kmp_save_internal_controls(kmp_info_t *thread); 3948 extern void __kmp_user_set_library(enum library_type arg); 3949 extern void __kmp_aux_set_library(enum library_type arg); 3950 extern void __kmp_aux_set_stacksize(size_t arg); 3951 extern void __kmp_aux_set_blocktime(int arg, kmp_info_t *thread, int tid); 3952 extern void __kmp_aux_set_defaults(char const *str, size_t len); 3953 3954 /* Functions called from __kmp_aux_env_initialize() in kmp_settings.cpp */ 3955 void kmpc_set_blocktime(int arg); 3956 void ompc_set_nested(int flag); 3957 void ompc_set_dynamic(int flag); 3958 void ompc_set_num_threads(int arg); 3959 3960 extern void __kmp_push_current_task_to_thread(kmp_info_t *this_thr, 3961 kmp_team_t *team, int tid); 3962 extern void __kmp_pop_current_task_from_thread(kmp_info_t *this_thr); 3963 extern kmp_task_t *__kmp_task_alloc(ident_t *loc_ref, kmp_int32 gtid, 3964 kmp_tasking_flags_t *flags, 3965 size_t sizeof_kmp_task_t, 3966 size_t sizeof_shareds, 3967 kmp_routine_entry_t task_entry); 3968 extern void __kmp_init_implicit_task(ident_t *loc_ref, kmp_info_t *this_thr, 3969 kmp_team_t *team, int tid, 3970 int set_curr_task); 3971 extern void __kmp_finish_implicit_task(kmp_info_t *this_thr); 3972 extern void __kmp_free_implicit_task(kmp_info_t *this_thr); 3973 3974 extern kmp_event_t *__kmpc_task_allow_completion_event(ident_t *loc_ref, 3975 int gtid, 3976 kmp_task_t *task); 3977 extern void __kmp_fulfill_event(kmp_event_t *event); 3978 3979 extern void __kmp_free_task_team(kmp_info_t *thread, 3980 kmp_task_team_t *task_team); 3981 extern void __kmp_reap_task_teams(void); 3982 extern void __kmp_wait_to_unref_task_teams(void); 3983 extern void __kmp_task_team_setup(kmp_info_t *this_thr, kmp_team_t *team, 3984 int always); 3985 extern void __kmp_task_team_sync(kmp_info_t *this_thr, kmp_team_t *team); 3986 extern void __kmp_task_team_wait(kmp_info_t *this_thr, kmp_team_t *team 3987 #if USE_ITT_BUILD 3988 , 3989 void *itt_sync_obj 3990 #endif /* USE_ITT_BUILD */ 3991 , 3992 int wait = 1); 3993 extern void __kmp_tasking_barrier(kmp_team_t *team, kmp_info_t *thread, 3994 int gtid); 3995 3996 extern int __kmp_is_address_mapped(void *addr); 3997 extern kmp_uint64 __kmp_hardware_timestamp(void); 3998 3999 #if KMP_OS_UNIX 4000 extern int __kmp_read_from_file(char const *path, char const *format, ...); 4001 #endif 4002 4003 /* ------------------------------------------------------------------------ */ 4004 // 4005 // Assembly routines that have no compiler intrinsic replacement 4006 // 4007 4008 extern int __kmp_invoke_microtask(microtask_t pkfn, int gtid, int npr, int argc, 4009 void *argv[] 4010 #if OMPT_SUPPORT 4011 , 4012 void **exit_frame_ptr 4013 #endif 4014 ); 4015 4016 /* ------------------------------------------------------------------------ */ 4017 4018 KMP_EXPORT void __kmpc_begin(ident_t *, kmp_int32 flags); 4019 KMP_EXPORT void __kmpc_end(ident_t *); 4020 4021 KMP_EXPORT void __kmpc_threadprivate_register_vec(ident_t *, void *data, 4022 kmpc_ctor_vec ctor, 4023 kmpc_cctor_vec cctor, 4024 kmpc_dtor_vec dtor, 4025 size_t vector_length); 4026 KMP_EXPORT void __kmpc_threadprivate_register(ident_t *, void *data, 4027 kmpc_ctor ctor, kmpc_cctor cctor, 4028 kmpc_dtor dtor); 4029 KMP_EXPORT void *__kmpc_threadprivate(ident_t *, kmp_int32 global_tid, 4030 void *data, size_t size); 4031 4032 KMP_EXPORT kmp_int32 __kmpc_global_thread_num(ident_t *); 4033 KMP_EXPORT kmp_int32 __kmpc_global_num_threads(ident_t *); 4034 KMP_EXPORT kmp_int32 __kmpc_bound_thread_num(ident_t *); 4035 KMP_EXPORT kmp_int32 __kmpc_bound_num_threads(ident_t *); 4036 4037 KMP_EXPORT kmp_int32 __kmpc_ok_to_fork(ident_t *); 4038 KMP_EXPORT void __kmpc_fork_call(ident_t *, kmp_int32 nargs, 4039 kmpc_micro microtask, ...); 4040 KMP_EXPORT void __kmpc_fork_call_if(ident_t *loc, kmp_int32 nargs, 4041 kmpc_micro microtask, kmp_int32 cond, 4042 void *args); 4043 4044 KMP_EXPORT void __kmpc_serialized_parallel(ident_t *, kmp_int32 global_tid); 4045 KMP_EXPORT void __kmpc_end_serialized_parallel(ident_t *, kmp_int32 global_tid); 4046 4047 KMP_EXPORT void __kmpc_flush(ident_t *); 4048 KMP_EXPORT void __kmpc_barrier(ident_t *, kmp_int32 global_tid); 4049 KMP_EXPORT kmp_int32 __kmpc_master(ident_t *, kmp_int32 global_tid); 4050 KMP_EXPORT void __kmpc_end_master(ident_t *, kmp_int32 global_tid); 4051 KMP_EXPORT kmp_int32 __kmpc_masked(ident_t *, kmp_int32 global_tid, 4052 kmp_int32 filter); 4053 KMP_EXPORT void __kmpc_end_masked(ident_t *, kmp_int32 global_tid); 4054 KMP_EXPORT void __kmpc_ordered(ident_t *, kmp_int32 global_tid); 4055 KMP_EXPORT void __kmpc_end_ordered(ident_t *, kmp_int32 global_tid); 4056 KMP_EXPORT void __kmpc_critical(ident_t *, kmp_int32 global_tid, 4057 kmp_critical_name *); 4058 KMP_EXPORT void __kmpc_end_critical(ident_t *, kmp_int32 global_tid, 4059 kmp_critical_name *); 4060 KMP_EXPORT void __kmpc_critical_with_hint(ident_t *, kmp_int32 global_tid, 4061 kmp_critical_name *, uint32_t hint); 4062 4063 KMP_EXPORT kmp_int32 __kmpc_barrier_master(ident_t *, kmp_int32 global_tid); 4064 KMP_EXPORT void __kmpc_end_barrier_master(ident_t *, kmp_int32 global_tid); 4065 4066 KMP_EXPORT kmp_int32 __kmpc_barrier_master_nowait(ident_t *, 4067 kmp_int32 global_tid); 4068 4069 KMP_EXPORT kmp_int32 __kmpc_single(ident_t *, kmp_int32 global_tid); 4070 KMP_EXPORT void __kmpc_end_single(ident_t *, kmp_int32 global_tid); 4071 4072 KMP_EXPORT kmp_int32 __kmpc_sections_init(ident_t *loc, kmp_int32 global_tid); 4073 KMP_EXPORT kmp_int32 __kmpc_next_section(ident_t *loc, kmp_int32 global_tid, 4074 kmp_int32 numberOfSections); 4075 KMP_EXPORT void __kmpc_end_sections(ident_t *loc, kmp_int32 global_tid); 4076 4077 KMP_EXPORT void KMPC_FOR_STATIC_INIT(ident_t *loc, kmp_int32 global_tid, 4078 kmp_int32 schedtype, kmp_int32 *plastiter, 4079 kmp_int *plower, kmp_int *pupper, 4080 kmp_int *pstride, kmp_int incr, 4081 kmp_int chunk); 4082 4083 KMP_EXPORT void __kmpc_for_static_fini(ident_t *loc, kmp_int32 global_tid); 4084 4085 KMP_EXPORT void __kmpc_copyprivate(ident_t *loc, kmp_int32 global_tid, 4086 size_t cpy_size, void *cpy_data, 4087 void (*cpy_func)(void *, void *), 4088 kmp_int32 didit); 4089 4090 KMP_EXPORT void *__kmpc_copyprivate_light(ident_t *loc, kmp_int32 gtid, 4091 void *cpy_data); 4092 4093 extern void KMPC_SET_NUM_THREADS(int arg); 4094 extern void KMPC_SET_DYNAMIC(int flag); 4095 extern void KMPC_SET_NESTED(int flag); 4096 4097 /* OMP 3.0 tasking interface routines */ 4098 KMP_EXPORT kmp_int32 __kmpc_omp_task(ident_t *loc_ref, kmp_int32 gtid, 4099 kmp_task_t *new_task); 4100 KMP_EXPORT kmp_task_t *__kmpc_omp_task_alloc(ident_t *loc_ref, kmp_int32 gtid, 4101 kmp_int32 flags, 4102 size_t sizeof_kmp_task_t, 4103 size_t sizeof_shareds, 4104 kmp_routine_entry_t task_entry); 4105 KMP_EXPORT kmp_task_t *__kmpc_omp_target_task_alloc( 4106 ident_t *loc_ref, kmp_int32 gtid, kmp_int32 flags, size_t sizeof_kmp_task_t, 4107 size_t sizeof_shareds, kmp_routine_entry_t task_entry, kmp_int64 device_id); 4108 KMP_EXPORT void __kmpc_omp_task_begin_if0(ident_t *loc_ref, kmp_int32 gtid, 4109 kmp_task_t *task); 4110 KMP_EXPORT void __kmpc_omp_task_complete_if0(ident_t *loc_ref, kmp_int32 gtid, 4111 kmp_task_t *task); 4112 KMP_EXPORT kmp_int32 __kmpc_omp_task_parts(ident_t *loc_ref, kmp_int32 gtid, 4113 kmp_task_t *new_task); 4114 KMP_EXPORT kmp_int32 __kmpc_omp_taskwait(ident_t *loc_ref, kmp_int32 gtid); 4115 KMP_EXPORT kmp_int32 __kmpc_omp_taskyield(ident_t *loc_ref, kmp_int32 gtid, 4116 int end_part); 4117 4118 #if TASK_UNUSED 4119 void __kmpc_omp_task_begin(ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *task); 4120 void __kmpc_omp_task_complete(ident_t *loc_ref, kmp_int32 gtid, 4121 kmp_task_t *task); 4122 #endif // TASK_UNUSED 4123 4124 /* ------------------------------------------------------------------------ */ 4125 4126 KMP_EXPORT void __kmpc_taskgroup(ident_t *loc, int gtid); 4127 KMP_EXPORT void __kmpc_end_taskgroup(ident_t *loc, int gtid); 4128 4129 KMP_EXPORT kmp_int32 __kmpc_omp_task_with_deps( 4130 ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 ndeps, 4131 kmp_depend_info_t *dep_list, kmp_int32 ndeps_noalias, 4132 kmp_depend_info_t *noalias_dep_list); 4133 KMP_EXPORT void __kmpc_omp_wait_deps(ident_t *loc_ref, kmp_int32 gtid, 4134 kmp_int32 ndeps, 4135 kmp_depend_info_t *dep_list, 4136 kmp_int32 ndeps_noalias, 4137 kmp_depend_info_t *noalias_dep_list); 4138 /* __kmpc_omp_taskwait_deps_51 : Function for OpenMP 5.1 nowait clause. 4139 * Placeholder for taskwait with nowait clause.*/ 4140 KMP_EXPORT void __kmpc_omp_taskwait_deps_51(ident_t *loc_ref, kmp_int32 gtid, 4141 kmp_int32 ndeps, 4142 kmp_depend_info_t *dep_list, 4143 kmp_int32 ndeps_noalias, 4144 kmp_depend_info_t *noalias_dep_list, 4145 kmp_int32 has_no_wait); 4146 4147 extern kmp_int32 __kmp_omp_task(kmp_int32 gtid, kmp_task_t *new_task, 4148 bool serialize_immediate); 4149 4150 KMP_EXPORT kmp_int32 __kmpc_cancel(ident_t *loc_ref, kmp_int32 gtid, 4151 kmp_int32 cncl_kind); 4152 KMP_EXPORT kmp_int32 __kmpc_cancellationpoint(ident_t *loc_ref, kmp_int32 gtid, 4153 kmp_int32 cncl_kind); 4154 KMP_EXPORT kmp_int32 __kmpc_cancel_barrier(ident_t *loc_ref, kmp_int32 gtid); 4155 KMP_EXPORT int __kmp_get_cancellation_status(int cancel_kind); 4156 4157 KMP_EXPORT void __kmpc_proxy_task_completed(kmp_int32 gtid, kmp_task_t *ptask); 4158 KMP_EXPORT void __kmpc_proxy_task_completed_ooo(kmp_task_t *ptask); 4159 KMP_EXPORT void __kmpc_taskloop(ident_t *loc, kmp_int32 gtid, kmp_task_t *task, 4160 kmp_int32 if_val, kmp_uint64 *lb, 4161 kmp_uint64 *ub, kmp_int64 st, kmp_int32 nogroup, 4162 kmp_int32 sched, kmp_uint64 grainsize, 4163 void *task_dup); 4164 KMP_EXPORT void __kmpc_taskloop_5(ident_t *loc, kmp_int32 gtid, 4165 kmp_task_t *task, kmp_int32 if_val, 4166 kmp_uint64 *lb, kmp_uint64 *ub, kmp_int64 st, 4167 kmp_int32 nogroup, kmp_int32 sched, 4168 kmp_uint64 grainsize, kmp_int32 modifier, 4169 void *task_dup); 4170 KMP_EXPORT void *__kmpc_task_reduction_init(int gtid, int num_data, void *data); 4171 KMP_EXPORT void *__kmpc_taskred_init(int gtid, int num_data, void *data); 4172 KMP_EXPORT void *__kmpc_task_reduction_get_th_data(int gtid, void *tg, void *d); 4173 KMP_EXPORT void *__kmpc_task_reduction_modifier_init(ident_t *loc, int gtid, 4174 int is_ws, int num, 4175 void *data); 4176 KMP_EXPORT void *__kmpc_taskred_modifier_init(ident_t *loc, int gtid, int is_ws, 4177 int num, void *data); 4178 KMP_EXPORT void __kmpc_task_reduction_modifier_fini(ident_t *loc, int gtid, 4179 int is_ws); 4180 KMP_EXPORT kmp_int32 __kmpc_omp_reg_task_with_affinity( 4181 ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 naffins, 4182 kmp_task_affinity_info_t *affin_list); 4183 KMP_EXPORT void __kmp_set_num_teams(int num_teams); 4184 KMP_EXPORT int __kmp_get_max_teams(void); 4185 KMP_EXPORT void __kmp_set_teams_thread_limit(int limit); 4186 KMP_EXPORT int __kmp_get_teams_thread_limit(void); 4187 4188 /* Interface target task integration */ 4189 KMP_EXPORT void **__kmpc_omp_get_target_async_handle_ptr(kmp_int32 gtid); 4190 KMP_EXPORT bool __kmpc_omp_has_task_team(kmp_int32 gtid); 4191 4192 /* Lock interface routines (fast versions with gtid passed in) */ 4193 KMP_EXPORT void __kmpc_init_lock(ident_t *loc, kmp_int32 gtid, 4194 void **user_lock); 4195 KMP_EXPORT void __kmpc_init_nest_lock(ident_t *loc, kmp_int32 gtid, 4196 void **user_lock); 4197 KMP_EXPORT void __kmpc_destroy_lock(ident_t *loc, kmp_int32 gtid, 4198 void **user_lock); 4199 KMP_EXPORT void __kmpc_destroy_nest_lock(ident_t *loc, kmp_int32 gtid, 4200 void **user_lock); 4201 KMP_EXPORT void __kmpc_set_lock(ident_t *loc, kmp_int32 gtid, void **user_lock); 4202 KMP_EXPORT void __kmpc_set_nest_lock(ident_t *loc, kmp_int32 gtid, 4203 void **user_lock); 4204 KMP_EXPORT void __kmpc_unset_lock(ident_t *loc, kmp_int32 gtid, 4205 void **user_lock); 4206 KMP_EXPORT void __kmpc_unset_nest_lock(ident_t *loc, kmp_int32 gtid, 4207 void **user_lock); 4208 KMP_EXPORT int __kmpc_test_lock(ident_t *loc, kmp_int32 gtid, void **user_lock); 4209 KMP_EXPORT int __kmpc_test_nest_lock(ident_t *loc, kmp_int32 gtid, 4210 void **user_lock); 4211 4212 KMP_EXPORT void __kmpc_init_lock_with_hint(ident_t *loc, kmp_int32 gtid, 4213 void **user_lock, uintptr_t hint); 4214 KMP_EXPORT void __kmpc_init_nest_lock_with_hint(ident_t *loc, kmp_int32 gtid, 4215 void **user_lock, 4216 uintptr_t hint); 4217 4218 #if OMPX_TASKGRAPH 4219 // Taskgraph's Record & Replay mechanism 4220 // __kmp_tdg_is_recording: check whether a given TDG is recording 4221 // status: the tdg's current status 4222 static inline bool __kmp_tdg_is_recording(kmp_tdg_status_t status) { 4223 return status == KMP_TDG_RECORDING; 4224 } 4225 4226 KMP_EXPORT kmp_int32 __kmpc_start_record_task(ident_t *loc, kmp_int32 gtid, 4227 kmp_int32 input_flags, 4228 kmp_int32 tdg_id); 4229 KMP_EXPORT void __kmpc_end_record_task(ident_t *loc, kmp_int32 gtid, 4230 kmp_int32 input_flags, kmp_int32 tdg_id); 4231 #endif 4232 /* Interface to fast scalable reduce methods routines */ 4233 4234 KMP_EXPORT kmp_int32 __kmpc_reduce_nowait( 4235 ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size, 4236 void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data), 4237 kmp_critical_name *lck); 4238 KMP_EXPORT void __kmpc_end_reduce_nowait(ident_t *loc, kmp_int32 global_tid, 4239 kmp_critical_name *lck); 4240 KMP_EXPORT kmp_int32 __kmpc_reduce( 4241 ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size, 4242 void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data), 4243 kmp_critical_name *lck); 4244 KMP_EXPORT void __kmpc_end_reduce(ident_t *loc, kmp_int32 global_tid, 4245 kmp_critical_name *lck); 4246 4247 /* Internal fast reduction routines */ 4248 4249 extern PACKED_REDUCTION_METHOD_T __kmp_determine_reduction_method( 4250 ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size, 4251 void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data), 4252 kmp_critical_name *lck); 4253 4254 // this function is for testing set/get/determine reduce method 4255 KMP_EXPORT kmp_int32 __kmp_get_reduce_method(void); 4256 4257 KMP_EXPORT kmp_uint64 __kmpc_get_taskid(); 4258 KMP_EXPORT kmp_uint64 __kmpc_get_parent_taskid(); 4259 4260 // C++ port 4261 // missing 'extern "C"' declarations 4262 4263 KMP_EXPORT kmp_int32 __kmpc_in_parallel(ident_t *loc); 4264 KMP_EXPORT void __kmpc_pop_num_threads(ident_t *loc, kmp_int32 global_tid); 4265 KMP_EXPORT void __kmpc_push_num_threads(ident_t *loc, kmp_int32 global_tid, 4266 kmp_int32 num_threads); 4267 4268 KMP_EXPORT void __kmpc_push_proc_bind(ident_t *loc, kmp_int32 global_tid, 4269 int proc_bind); 4270 KMP_EXPORT void __kmpc_push_num_teams(ident_t *loc, kmp_int32 global_tid, 4271 kmp_int32 num_teams, 4272 kmp_int32 num_threads); 4273 /* Function for OpenMP 5.1 num_teams clause */ 4274 KMP_EXPORT void __kmpc_push_num_teams_51(ident_t *loc, kmp_int32 global_tid, 4275 kmp_int32 num_teams_lb, 4276 kmp_int32 num_teams_ub, 4277 kmp_int32 num_threads); 4278 KMP_EXPORT void __kmpc_fork_teams(ident_t *loc, kmp_int32 argc, 4279 kmpc_micro microtask, ...); 4280 struct kmp_dim { // loop bounds info casted to kmp_int64 4281 kmp_int64 lo; // lower 4282 kmp_int64 up; // upper 4283 kmp_int64 st; // stride 4284 }; 4285 KMP_EXPORT void __kmpc_doacross_init(ident_t *loc, kmp_int32 gtid, 4286 kmp_int32 num_dims, 4287 const struct kmp_dim *dims); 4288 KMP_EXPORT void __kmpc_doacross_wait(ident_t *loc, kmp_int32 gtid, 4289 const kmp_int64 *vec); 4290 KMP_EXPORT void __kmpc_doacross_post(ident_t *loc, kmp_int32 gtid, 4291 const kmp_int64 *vec); 4292 KMP_EXPORT void __kmpc_doacross_fini(ident_t *loc, kmp_int32 gtid); 4293 4294 KMP_EXPORT void *__kmpc_threadprivate_cached(ident_t *loc, kmp_int32 global_tid, 4295 void *data, size_t size, 4296 void ***cache); 4297 4298 // The routines below are not exported. 4299 // Consider making them 'static' in corresponding source files. 4300 void kmp_threadprivate_insert_private_data(int gtid, void *pc_addr, 4301 void *data_addr, size_t pc_size); 4302 struct private_common *kmp_threadprivate_insert(int gtid, void *pc_addr, 4303 void *data_addr, 4304 size_t pc_size); 4305 void __kmp_threadprivate_resize_cache(int newCapacity); 4306 void __kmp_cleanup_threadprivate_caches(); 4307 4308 // ompc_, kmpc_ entries moved from omp.h. 4309 #if KMP_OS_WINDOWS 4310 #define KMPC_CONVENTION __cdecl 4311 #else 4312 #define KMPC_CONVENTION 4313 #endif 4314 4315 #ifndef __OMP_H 4316 typedef enum omp_sched_t { 4317 omp_sched_static = 1, 4318 omp_sched_dynamic = 2, 4319 omp_sched_guided = 3, 4320 omp_sched_auto = 4 4321 } omp_sched_t; 4322 typedef void *kmp_affinity_mask_t; 4323 #endif 4324 4325 KMP_EXPORT void KMPC_CONVENTION ompc_set_max_active_levels(int); 4326 KMP_EXPORT void KMPC_CONVENTION ompc_set_schedule(omp_sched_t, int); 4327 KMP_EXPORT int KMPC_CONVENTION ompc_get_ancestor_thread_num(int); 4328 KMP_EXPORT int KMPC_CONVENTION ompc_get_team_size(int); 4329 KMP_EXPORT int KMPC_CONVENTION 4330 kmpc_set_affinity_mask_proc(int, kmp_affinity_mask_t *); 4331 KMP_EXPORT int KMPC_CONVENTION 4332 kmpc_unset_affinity_mask_proc(int, kmp_affinity_mask_t *); 4333 KMP_EXPORT int KMPC_CONVENTION 4334 kmpc_get_affinity_mask_proc(int, kmp_affinity_mask_t *); 4335 4336 KMP_EXPORT void KMPC_CONVENTION kmpc_set_stacksize(int); 4337 KMP_EXPORT void KMPC_CONVENTION kmpc_set_stacksize_s(size_t); 4338 KMP_EXPORT void KMPC_CONVENTION kmpc_set_library(int); 4339 KMP_EXPORT void KMPC_CONVENTION kmpc_set_defaults(char const *); 4340 KMP_EXPORT void KMPC_CONVENTION kmpc_set_disp_num_buffers(int); 4341 void KMP_EXPAND_NAME(ompc_set_affinity_format)(char const *format); 4342 size_t KMP_EXPAND_NAME(ompc_get_affinity_format)(char *buffer, size_t size); 4343 void KMP_EXPAND_NAME(ompc_display_affinity)(char const *format); 4344 size_t KMP_EXPAND_NAME(ompc_capture_affinity)(char *buffer, size_t buf_size, 4345 char const *format); 4346 4347 enum kmp_target_offload_kind { 4348 tgt_disabled = 0, 4349 tgt_default = 1, 4350 tgt_mandatory = 2 4351 }; 4352 typedef enum kmp_target_offload_kind kmp_target_offload_kind_t; 4353 // Set via OMP_TARGET_OFFLOAD if specified, defaults to tgt_default otherwise 4354 extern kmp_target_offload_kind_t __kmp_target_offload; 4355 extern int __kmpc_get_target_offload(); 4356 4357 // Constants used in libomptarget 4358 #define KMP_DEVICE_DEFAULT -1 // This is libomptarget's default device. 4359 #define KMP_DEVICE_ALL -11 // This is libomptarget's "all devices". 4360 4361 // OMP Pause Resource 4362 4363 // The following enum is used both to set the status in __kmp_pause_status, and 4364 // as the internal equivalent of the externally-visible omp_pause_resource_t. 4365 typedef enum kmp_pause_status_t { 4366 kmp_not_paused = 0, // status is not paused, or, requesting resume 4367 kmp_soft_paused = 1, // status is soft-paused, or, requesting soft pause 4368 kmp_hard_paused = 2 // status is hard-paused, or, requesting hard pause 4369 } kmp_pause_status_t; 4370 4371 // This stores the pause state of the runtime 4372 extern kmp_pause_status_t __kmp_pause_status; 4373 extern int __kmpc_pause_resource(kmp_pause_status_t level); 4374 extern int __kmp_pause_resource(kmp_pause_status_t level); 4375 // Soft resume sets __kmp_pause_status, and wakes up all threads. 4376 extern void __kmp_resume_if_soft_paused(); 4377 // Hard resume simply resets the status to not paused. Library will appear to 4378 // be uninitialized after hard pause. Let OMP constructs trigger required 4379 // initializations. 4380 static inline void __kmp_resume_if_hard_paused() { 4381 if (__kmp_pause_status == kmp_hard_paused) { 4382 __kmp_pause_status = kmp_not_paused; 4383 } 4384 } 4385 4386 extern void __kmp_omp_display_env(int verbose); 4387 4388 // 1: it is initializing hidden helper team 4389 extern volatile int __kmp_init_hidden_helper; 4390 // 1: the hidden helper team is done 4391 extern volatile int __kmp_hidden_helper_team_done; 4392 // 1: enable hidden helper task 4393 extern kmp_int32 __kmp_enable_hidden_helper; 4394 // Main thread of hidden helper team 4395 extern kmp_info_t *__kmp_hidden_helper_main_thread; 4396 // Descriptors for the hidden helper threads 4397 extern kmp_info_t **__kmp_hidden_helper_threads; 4398 // Number of hidden helper threads 4399 extern kmp_int32 __kmp_hidden_helper_threads_num; 4400 // Number of hidden helper tasks that have not been executed yet 4401 extern std::atomic<kmp_int32> __kmp_unexecuted_hidden_helper_tasks; 4402 4403 extern void __kmp_hidden_helper_initialize(); 4404 extern void __kmp_hidden_helper_threads_initz_routine(); 4405 extern void __kmp_do_initialize_hidden_helper_threads(); 4406 extern void __kmp_hidden_helper_threads_initz_wait(); 4407 extern void __kmp_hidden_helper_initz_release(); 4408 extern void __kmp_hidden_helper_threads_deinitz_wait(); 4409 extern void __kmp_hidden_helper_threads_deinitz_release(); 4410 extern void __kmp_hidden_helper_main_thread_wait(); 4411 extern void __kmp_hidden_helper_worker_thread_wait(); 4412 extern void __kmp_hidden_helper_worker_thread_signal(); 4413 extern void __kmp_hidden_helper_main_thread_release(); 4414 4415 // Check whether a given thread is a hidden helper thread 4416 #define KMP_HIDDEN_HELPER_THREAD(gtid) \ 4417 ((gtid) >= 1 && (gtid) <= __kmp_hidden_helper_threads_num) 4418 4419 #define KMP_HIDDEN_HELPER_WORKER_THREAD(gtid) \ 4420 ((gtid) > 1 && (gtid) <= __kmp_hidden_helper_threads_num) 4421 4422 #define KMP_HIDDEN_HELPER_MAIN_THREAD(gtid) \ 4423 ((gtid) == 1 && (gtid) <= __kmp_hidden_helper_threads_num) 4424 4425 #define KMP_HIDDEN_HELPER_TEAM(team) \ 4426 (team->t.t_threads[0] == __kmp_hidden_helper_main_thread) 4427 4428 // Map a gtid to a hidden helper thread. The first hidden helper thread, a.k.a 4429 // main thread, is skipped. 4430 #define KMP_GTID_TO_SHADOW_GTID(gtid) \ 4431 ((gtid) % (__kmp_hidden_helper_threads_num - 1) + 2) 4432 4433 // Return the adjusted gtid value by subtracting from gtid the number 4434 // of hidden helper threads. This adjusted value is the gtid the thread would 4435 // have received if there were no hidden helper threads. 4436 static inline int __kmp_adjust_gtid_for_hidden_helpers(int gtid) { 4437 int adjusted_gtid = gtid; 4438 if (__kmp_hidden_helper_threads_num > 0 && gtid > 0 && 4439 gtid - __kmp_hidden_helper_threads_num >= 0) { 4440 adjusted_gtid -= __kmp_hidden_helper_threads_num; 4441 } 4442 return adjusted_gtid; 4443 } 4444 4445 // Support for error directive 4446 typedef enum kmp_severity_t { 4447 severity_warning = 1, 4448 severity_fatal = 2 4449 } kmp_severity_t; 4450 extern void __kmpc_error(ident_t *loc, int severity, const char *message); 4451 4452 // Support for scope directive 4453 KMP_EXPORT void __kmpc_scope(ident_t *loc, kmp_int32 gtid, void *reserved); 4454 KMP_EXPORT void __kmpc_end_scope(ident_t *loc, kmp_int32 gtid, void *reserved); 4455 4456 #ifdef __cplusplus 4457 } 4458 #endif 4459 4460 template <bool C, bool S> 4461 extern void __kmp_suspend_32(int th_gtid, kmp_flag_32<C, S> *flag); 4462 template <bool C, bool S> 4463 extern void __kmp_suspend_64(int th_gtid, kmp_flag_64<C, S> *flag); 4464 template <bool C, bool S> 4465 extern void __kmp_atomic_suspend_64(int th_gtid, 4466 kmp_atomic_flag_64<C, S> *flag); 4467 extern void __kmp_suspend_oncore(int th_gtid, kmp_flag_oncore *flag); 4468 #if KMP_HAVE_MWAIT || KMP_HAVE_UMWAIT 4469 template <bool C, bool S> 4470 extern void __kmp_mwait_32(int th_gtid, kmp_flag_32<C, S> *flag); 4471 template <bool C, bool S> 4472 extern void __kmp_mwait_64(int th_gtid, kmp_flag_64<C, S> *flag); 4473 template <bool C, bool S> 4474 extern void __kmp_atomic_mwait_64(int th_gtid, kmp_atomic_flag_64<C, S> *flag); 4475 extern void __kmp_mwait_oncore(int th_gtid, kmp_flag_oncore *flag); 4476 #endif 4477 template <bool C, bool S> 4478 extern void __kmp_resume_32(int target_gtid, kmp_flag_32<C, S> *flag); 4479 template <bool C, bool S> 4480 extern void __kmp_resume_64(int target_gtid, kmp_flag_64<C, S> *flag); 4481 template <bool C, bool S> 4482 extern void __kmp_atomic_resume_64(int target_gtid, 4483 kmp_atomic_flag_64<C, S> *flag); 4484 extern void __kmp_resume_oncore(int target_gtid, kmp_flag_oncore *flag); 4485 4486 template <bool C, bool S> 4487 int __kmp_execute_tasks_32(kmp_info_t *thread, kmp_int32 gtid, 4488 kmp_flag_32<C, S> *flag, int final_spin, 4489 int *thread_finished, 4490 #if USE_ITT_BUILD 4491 void *itt_sync_obj, 4492 #endif /* USE_ITT_BUILD */ 4493 kmp_int32 is_constrained); 4494 template <bool C, bool S> 4495 int __kmp_execute_tasks_64(kmp_info_t *thread, kmp_int32 gtid, 4496 kmp_flag_64<C, S> *flag, int final_spin, 4497 int *thread_finished, 4498 #if USE_ITT_BUILD 4499 void *itt_sync_obj, 4500 #endif /* USE_ITT_BUILD */ 4501 kmp_int32 is_constrained); 4502 template <bool C, bool S> 4503 int __kmp_atomic_execute_tasks_64(kmp_info_t *thread, kmp_int32 gtid, 4504 kmp_atomic_flag_64<C, S> *flag, 4505 int final_spin, int *thread_finished, 4506 #if USE_ITT_BUILD 4507 void *itt_sync_obj, 4508 #endif /* USE_ITT_BUILD */ 4509 kmp_int32 is_constrained); 4510 int __kmp_execute_tasks_oncore(kmp_info_t *thread, kmp_int32 gtid, 4511 kmp_flag_oncore *flag, int final_spin, 4512 int *thread_finished, 4513 #if USE_ITT_BUILD 4514 void *itt_sync_obj, 4515 #endif /* USE_ITT_BUILD */ 4516 kmp_int32 is_constrained); 4517 4518 extern int __kmp_nesting_mode; 4519 extern int __kmp_nesting_mode_nlevels; 4520 extern int *__kmp_nesting_nth_level; 4521 extern void __kmp_init_nesting_mode(); 4522 extern void __kmp_set_nesting_mode_threads(); 4523 4524 /// This class safely opens and closes a C-style FILE* object using RAII 4525 /// semantics. There are also methods which allow using stdout or stderr as 4526 /// the underlying FILE* object. With the implicit conversion operator to 4527 /// FILE*, an object with this type can be used in any function which takes 4528 /// a FILE* object e.g., fprintf(). 4529 /// No close method is needed at use sites. 4530 class kmp_safe_raii_file_t { 4531 FILE *f; 4532 4533 void close() { 4534 if (f && f != stdout && f != stderr) { 4535 fclose(f); 4536 f = nullptr; 4537 } 4538 } 4539 4540 public: 4541 kmp_safe_raii_file_t() : f(nullptr) {} 4542 kmp_safe_raii_file_t(const char *filename, const char *mode, 4543 const char *env_var = nullptr) 4544 : f(nullptr) { 4545 open(filename, mode, env_var); 4546 } 4547 ~kmp_safe_raii_file_t() { close(); } 4548 4549 /// Open filename using mode. This is automatically closed in the destructor. 4550 /// The env_var parameter indicates the environment variable the filename 4551 /// came from if != nullptr. 4552 void open(const char *filename, const char *mode, 4553 const char *env_var = nullptr) { 4554 KMP_ASSERT(!f); 4555 f = fopen(filename, mode); 4556 if (!f) { 4557 int code = errno; 4558 if (env_var) { 4559 __kmp_fatal(KMP_MSG(CantOpenFileForReading, filename), KMP_ERR(code), 4560 KMP_HNT(CheckEnvVar, env_var, filename), __kmp_msg_null); 4561 } else { 4562 __kmp_fatal(KMP_MSG(CantOpenFileForReading, filename), KMP_ERR(code), 4563 __kmp_msg_null); 4564 } 4565 } 4566 } 4567 /// Instead of erroring out, return non-zero when 4568 /// unsuccessful fopen() for any reason 4569 int try_open(const char *filename, const char *mode) { 4570 KMP_ASSERT(!f); 4571 f = fopen(filename, mode); 4572 if (!f) 4573 return errno; 4574 return 0; 4575 } 4576 /// Set the FILE* object to stdout and output there 4577 /// No open call should happen before this call. 4578 void set_stdout() { 4579 KMP_ASSERT(!f); 4580 f = stdout; 4581 } 4582 /// Set the FILE* object to stderr and output there 4583 /// No open call should happen before this call. 4584 void set_stderr() { 4585 KMP_ASSERT(!f); 4586 f = stderr; 4587 } 4588 operator bool() { return bool(f); } 4589 operator FILE *() { return f; } 4590 }; 4591 4592 template <typename SourceType, typename TargetType, 4593 bool isSourceSmaller = (sizeof(SourceType) < sizeof(TargetType)), 4594 bool isSourceEqual = (sizeof(SourceType) == sizeof(TargetType)), 4595 bool isSourceSigned = std::is_signed<SourceType>::value, 4596 bool isTargetSigned = std::is_signed<TargetType>::value> 4597 struct kmp_convert {}; 4598 4599 // Both types are signed; Source smaller 4600 template <typename SourceType, typename TargetType> 4601 struct kmp_convert<SourceType, TargetType, true, false, true, true> { 4602 static TargetType to(SourceType src) { return (TargetType)src; } 4603 }; 4604 // Source equal 4605 template <typename SourceType, typename TargetType> 4606 struct kmp_convert<SourceType, TargetType, false, true, true, true> { 4607 static TargetType to(SourceType src) { return src; } 4608 }; 4609 // Source bigger 4610 template <typename SourceType, typename TargetType> 4611 struct kmp_convert<SourceType, TargetType, false, false, true, true> { 4612 static TargetType to(SourceType src) { 4613 KMP_ASSERT(src <= static_cast<SourceType>( 4614 (std::numeric_limits<TargetType>::max)())); 4615 KMP_ASSERT(src >= static_cast<SourceType>( 4616 (std::numeric_limits<TargetType>::min)())); 4617 return (TargetType)src; 4618 } 4619 }; 4620 4621 // Source signed, Target unsigned 4622 // Source smaller 4623 template <typename SourceType, typename TargetType> 4624 struct kmp_convert<SourceType, TargetType, true, false, true, false> { 4625 static TargetType to(SourceType src) { 4626 KMP_ASSERT(src >= 0); 4627 return (TargetType)src; 4628 } 4629 }; 4630 // Source equal 4631 template <typename SourceType, typename TargetType> 4632 struct kmp_convert<SourceType, TargetType, false, true, true, false> { 4633 static TargetType to(SourceType src) { 4634 KMP_ASSERT(src >= 0); 4635 return (TargetType)src; 4636 } 4637 }; 4638 // Source bigger 4639 template <typename SourceType, typename TargetType> 4640 struct kmp_convert<SourceType, TargetType, false, false, true, false> { 4641 static TargetType to(SourceType src) { 4642 KMP_ASSERT(src >= 0); 4643 KMP_ASSERT(src <= static_cast<SourceType>( 4644 (std::numeric_limits<TargetType>::max)())); 4645 return (TargetType)src; 4646 } 4647 }; 4648 4649 // Source unsigned, Target signed 4650 // Source smaller 4651 template <typename SourceType, typename TargetType> 4652 struct kmp_convert<SourceType, TargetType, true, false, false, true> { 4653 static TargetType to(SourceType src) { return (TargetType)src; } 4654 }; 4655 // Source equal 4656 template <typename SourceType, typename TargetType> 4657 struct kmp_convert<SourceType, TargetType, false, true, false, true> { 4658 static TargetType to(SourceType src) { 4659 KMP_ASSERT(src <= static_cast<SourceType>( 4660 (std::numeric_limits<TargetType>::max)())); 4661 return (TargetType)src; 4662 } 4663 }; 4664 // Source bigger 4665 template <typename SourceType, typename TargetType> 4666 struct kmp_convert<SourceType, TargetType, false, false, false, true> { 4667 static TargetType to(SourceType src) { 4668 KMP_ASSERT(src <= static_cast<SourceType>( 4669 (std::numeric_limits<TargetType>::max)())); 4670 return (TargetType)src; 4671 } 4672 }; 4673 4674 // Source unsigned, Target unsigned 4675 // Source smaller 4676 template <typename SourceType, typename TargetType> 4677 struct kmp_convert<SourceType, TargetType, true, false, false, false> { 4678 static TargetType to(SourceType src) { return (TargetType)src; } 4679 }; 4680 // Source equal 4681 template <typename SourceType, typename TargetType> 4682 struct kmp_convert<SourceType, TargetType, false, true, false, false> { 4683 static TargetType to(SourceType src) { return src; } 4684 }; 4685 // Source bigger 4686 template <typename SourceType, typename TargetType> 4687 struct kmp_convert<SourceType, TargetType, false, false, false, false> { 4688 static TargetType to(SourceType src) { 4689 KMP_ASSERT(src <= static_cast<SourceType>( 4690 (std::numeric_limits<TargetType>::max)())); 4691 return (TargetType)src; 4692 } 4693 }; 4694 4695 template <typename T1, typename T2> 4696 static inline void __kmp_type_convert(T1 src, T2 *dest) { 4697 *dest = kmp_convert<T1, T2>::to(src); 4698 } 4699 4700 #endif /* KMP_H */ 4701