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