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 #if HWLOC_API_VERSION >= 0x00020000 104 // hwloc 2.0 changed type of depth of object from unsigned to int 105 typedef int kmp_hwloc_depth_t; 106 #else 107 typedef unsigned int kmp_hwloc_depth_t; 108 #endif 109 #endif 110 111 #if KMP_ARCH_X86 || KMP_ARCH_X86_64 112 #include <xmmintrin.h> 113 #endif 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 #define KMP_DEBUG_ASSERT_VALID_HW_TYPE(type) \ 622 KMP_DEBUG_ASSERT(type >= (kmp_hw_t)0 && type < KMP_HW_LAST) 623 #define KMP_ASSERT_VALID_HW_TYPE(type) \ 624 KMP_ASSERT(type >= (kmp_hw_t)0 && type < KMP_HW_LAST) 625 626 #define KMP_FOREACH_HW_TYPE(type) \ 627 for (kmp_hw_t type = (kmp_hw_t)0; type < KMP_HW_LAST; \ 628 type = (kmp_hw_t)((int)type + 1)) 629 630 const char *__kmp_hw_get_keyword(kmp_hw_t type, bool plural = false); 631 const char *__kmp_hw_get_catalog_string(kmp_hw_t type, bool plural = false); 632 633 /* Only Linux* OS and Windows* OS support thread affinity. */ 634 #if KMP_AFFINITY_SUPPORTED 635 636 // GROUP_AFFINITY is already defined for _MSC_VER>=1600 (VS2010 and later). 637 #if KMP_OS_WINDOWS 638 #if _MSC_VER < 1600 && KMP_MSVC_COMPAT 639 typedef struct GROUP_AFFINITY { 640 KAFFINITY Mask; 641 WORD Group; 642 WORD Reserved[3]; 643 } GROUP_AFFINITY; 644 #endif /* _MSC_VER < 1600 */ 645 #if KMP_GROUP_AFFINITY 646 extern int __kmp_num_proc_groups; 647 #else 648 static const int __kmp_num_proc_groups = 1; 649 #endif /* KMP_GROUP_AFFINITY */ 650 typedef DWORD (*kmp_GetActiveProcessorCount_t)(WORD); 651 extern kmp_GetActiveProcessorCount_t __kmp_GetActiveProcessorCount; 652 653 typedef WORD (*kmp_GetActiveProcessorGroupCount_t)(void); 654 extern kmp_GetActiveProcessorGroupCount_t __kmp_GetActiveProcessorGroupCount; 655 656 typedef BOOL (*kmp_GetThreadGroupAffinity_t)(HANDLE, GROUP_AFFINITY *); 657 extern kmp_GetThreadGroupAffinity_t __kmp_GetThreadGroupAffinity; 658 659 typedef BOOL (*kmp_SetThreadGroupAffinity_t)(HANDLE, const GROUP_AFFINITY *, 660 GROUP_AFFINITY *); 661 extern kmp_SetThreadGroupAffinity_t __kmp_SetThreadGroupAffinity; 662 #endif /* KMP_OS_WINDOWS */ 663 664 #if KMP_USE_HWLOC 665 extern hwloc_topology_t __kmp_hwloc_topology; 666 extern int __kmp_hwloc_error; 667 #endif 668 669 extern size_t __kmp_affin_mask_size; 670 #define KMP_AFFINITY_CAPABLE() (__kmp_affin_mask_size > 0) 671 #define KMP_AFFINITY_DISABLE() (__kmp_affin_mask_size = 0) 672 #define KMP_AFFINITY_ENABLE(mask_size) (__kmp_affin_mask_size = mask_size) 673 #define KMP_CPU_SET_ITERATE(i, mask) \ 674 for (i = (mask)->begin(); (int)i != (mask)->end(); i = (mask)->next(i)) 675 #define KMP_CPU_SET(i, mask) (mask)->set(i) 676 #define KMP_CPU_ISSET(i, mask) (mask)->is_set(i) 677 #define KMP_CPU_CLR(i, mask) (mask)->clear(i) 678 #define KMP_CPU_ZERO(mask) (mask)->zero() 679 #define KMP_CPU_COPY(dest, src) (dest)->copy(src) 680 #define KMP_CPU_AND(dest, src) (dest)->bitwise_and(src) 681 #define KMP_CPU_COMPLEMENT(max_bit_number, mask) (mask)->bitwise_not() 682 #define KMP_CPU_UNION(dest, src) (dest)->bitwise_or(src) 683 #define KMP_CPU_ALLOC(ptr) (ptr = __kmp_affinity_dispatch->allocate_mask()) 684 #define KMP_CPU_FREE(ptr) __kmp_affinity_dispatch->deallocate_mask(ptr) 685 #define KMP_CPU_ALLOC_ON_STACK(ptr) KMP_CPU_ALLOC(ptr) 686 #define KMP_CPU_FREE_FROM_STACK(ptr) KMP_CPU_FREE(ptr) 687 #define KMP_CPU_INTERNAL_ALLOC(ptr) KMP_CPU_ALLOC(ptr) 688 #define KMP_CPU_INTERNAL_FREE(ptr) KMP_CPU_FREE(ptr) 689 #define KMP_CPU_INDEX(arr, i) __kmp_affinity_dispatch->index_mask_array(arr, i) 690 #define KMP_CPU_ALLOC_ARRAY(arr, n) \ 691 (arr = __kmp_affinity_dispatch->allocate_mask_array(n)) 692 #define KMP_CPU_FREE_ARRAY(arr, n) \ 693 __kmp_affinity_dispatch->deallocate_mask_array(arr) 694 #define KMP_CPU_INTERNAL_ALLOC_ARRAY(arr, n) KMP_CPU_ALLOC_ARRAY(arr, n) 695 #define KMP_CPU_INTERNAL_FREE_ARRAY(arr, n) KMP_CPU_FREE_ARRAY(arr, n) 696 #define __kmp_get_system_affinity(mask, abort_bool) \ 697 (mask)->get_system_affinity(abort_bool) 698 #define __kmp_set_system_affinity(mask, abort_bool) \ 699 (mask)->set_system_affinity(abort_bool) 700 #define __kmp_get_proc_group(mask) (mask)->get_proc_group() 701 702 class KMPAffinity { 703 public: 704 class Mask { 705 public: 706 void *operator new(size_t n); 707 void operator delete(void *p); 708 void *operator new[](size_t n); 709 void operator delete[](void *p); 710 virtual ~Mask() {} 711 // Set bit i to 1 712 virtual void set(int i) {} 713 // Return bit i 714 virtual bool is_set(int i) const { return false; } 715 // Set bit i to 0 716 virtual void clear(int i) {} 717 // Zero out entire mask 718 virtual void zero() {} 719 // Copy src into this mask 720 virtual void copy(const Mask *src) {} 721 // this &= rhs 722 virtual void bitwise_and(const Mask *rhs) {} 723 // this |= rhs 724 virtual void bitwise_or(const Mask *rhs) {} 725 // this = ~this 726 virtual void bitwise_not() {} 727 // API for iterating over an affinity mask 728 // for (int i = mask->begin(); i != mask->end(); i = mask->next(i)) 729 virtual int begin() const { return 0; } 730 virtual int end() const { return 0; } 731 virtual int next(int previous) const { return 0; } 732 #if KMP_OS_WINDOWS 733 virtual int set_process_affinity(bool abort_on_error) const { return -1; } 734 #endif 735 // Set the system's affinity to this affinity mask's value 736 virtual int set_system_affinity(bool abort_on_error) const { return -1; } 737 // Set this affinity mask to the current system affinity 738 virtual int get_system_affinity(bool abort_on_error) { return -1; } 739 // Only 1 DWORD in the mask should have any procs set. 740 // Return the appropriate index, or -1 for an invalid mask. 741 virtual int get_proc_group() const { return -1; } 742 }; 743 void *operator new(size_t n); 744 void operator delete(void *p); 745 // Need virtual destructor 746 virtual ~KMPAffinity() = default; 747 // Determine if affinity is capable 748 virtual void determine_capable(const char *env_var) {} 749 // Bind the current thread to os proc 750 virtual void bind_thread(int proc) {} 751 // Factory functions to allocate/deallocate a mask 752 virtual Mask *allocate_mask() { return nullptr; } 753 virtual void deallocate_mask(Mask *m) {} 754 virtual Mask *allocate_mask_array(int num) { return nullptr; } 755 virtual void deallocate_mask_array(Mask *m) {} 756 virtual Mask *index_mask_array(Mask *m, int index) { return nullptr; } 757 static void pick_api(); 758 static void destroy_api(); 759 enum api_type { 760 NATIVE_OS 761 #if KMP_USE_HWLOC 762 , 763 HWLOC 764 #endif 765 }; 766 virtual api_type get_api_type() const { 767 KMP_ASSERT(0); 768 return NATIVE_OS; 769 } 770 771 private: 772 static bool picked_api; 773 }; 774 775 typedef KMPAffinity::Mask kmp_affin_mask_t; 776 extern KMPAffinity *__kmp_affinity_dispatch; 777 778 // Declare local char buffers with this size for printing debug and info 779 // messages, using __kmp_affinity_print_mask(). 780 #define KMP_AFFIN_MASK_PRINT_LEN 1024 781 782 enum affinity_type { 783 affinity_none = 0, 784 affinity_physical, 785 affinity_logical, 786 affinity_compact, 787 affinity_scatter, 788 affinity_explicit, 789 affinity_balanced, 790 affinity_disabled, // not used outsize the env var parser 791 affinity_default 792 }; 793 794 enum affinity_top_method { 795 affinity_top_method_all = 0, // try all (supported) methods, in order 796 #if KMP_ARCH_X86 || KMP_ARCH_X86_64 797 affinity_top_method_apicid, 798 affinity_top_method_x2apicid, 799 affinity_top_method_x2apicid_1f, 800 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */ 801 affinity_top_method_cpuinfo, // KMP_CPUINFO_FILE is usable on Windows* OS, too 802 #if KMP_GROUP_AFFINITY 803 affinity_top_method_group, 804 #endif /* KMP_GROUP_AFFINITY */ 805 affinity_top_method_flat, 806 #if KMP_USE_HWLOC 807 affinity_top_method_hwloc, 808 #endif 809 affinity_top_method_default 810 }; 811 812 #define affinity_respect_mask_default (-1) 813 814 extern enum affinity_type __kmp_affinity_type; /* Affinity type */ 815 extern kmp_hw_t __kmp_affinity_gran; /* Affinity granularity */ 816 extern int __kmp_affinity_gran_levels; /* corresponding int value */ 817 extern int __kmp_affinity_dups; /* Affinity duplicate masks */ 818 extern enum affinity_top_method __kmp_affinity_top_method; 819 extern int __kmp_affinity_compact; /* Affinity 'compact' value */ 820 extern int __kmp_affinity_offset; /* Affinity offset value */ 821 extern int __kmp_affinity_verbose; /* Was verbose specified for KMP_AFFINITY? */ 822 extern int __kmp_affinity_warnings; /* KMP_AFFINITY warnings enabled ? */ 823 extern int __kmp_affinity_respect_mask; // Respect process' init affinity mask? 824 extern char *__kmp_affinity_proclist; /* proc ID list */ 825 extern kmp_affin_mask_t *__kmp_affinity_masks; 826 extern unsigned __kmp_affinity_num_masks; 827 extern void __kmp_affinity_bind_thread(int which); 828 829 extern kmp_affin_mask_t *__kmp_affin_fullMask; 830 extern char *__kmp_cpuinfo_file; 831 832 #endif /* KMP_AFFINITY_SUPPORTED */ 833 834 // This needs to be kept in sync with the values in omp.h !!! 835 typedef enum kmp_proc_bind_t { 836 proc_bind_false = 0, 837 proc_bind_true, 838 proc_bind_primary, 839 proc_bind_close, 840 proc_bind_spread, 841 proc_bind_intel, // use KMP_AFFINITY interface 842 proc_bind_default 843 } kmp_proc_bind_t; 844 845 typedef struct kmp_nested_proc_bind_t { 846 kmp_proc_bind_t *bind_types; 847 int size; 848 int used; 849 } kmp_nested_proc_bind_t; 850 851 extern kmp_nested_proc_bind_t __kmp_nested_proc_bind; 852 extern kmp_proc_bind_t __kmp_teams_proc_bind; 853 854 extern int __kmp_display_affinity; 855 extern char *__kmp_affinity_format; 856 static const size_t KMP_AFFINITY_FORMAT_SIZE = 512; 857 #if OMPT_SUPPORT 858 extern int __kmp_tool; 859 extern char *__kmp_tool_libraries; 860 #endif // OMPT_SUPPORT 861 862 #if KMP_AFFINITY_SUPPORTED 863 #define KMP_PLACE_ALL (-1) 864 #define KMP_PLACE_UNDEFINED (-2) 865 // Is KMP_AFFINITY is being used instead of OMP_PROC_BIND/OMP_PLACES? 866 #define KMP_AFFINITY_NON_PROC_BIND \ 867 ((__kmp_nested_proc_bind.bind_types[0] == proc_bind_false || \ 868 __kmp_nested_proc_bind.bind_types[0] == proc_bind_intel) && \ 869 (__kmp_affinity_num_masks > 0 || __kmp_affinity_type == affinity_balanced)) 870 #endif /* KMP_AFFINITY_SUPPORTED */ 871 872 extern int __kmp_affinity_num_places; 873 874 typedef enum kmp_cancel_kind_t { 875 cancel_noreq = 0, 876 cancel_parallel = 1, 877 cancel_loop = 2, 878 cancel_sections = 3, 879 cancel_taskgroup = 4 880 } kmp_cancel_kind_t; 881 882 // KMP_HW_SUBSET support: 883 typedef struct kmp_hws_item { 884 int num; 885 int offset; 886 } kmp_hws_item_t; 887 888 extern kmp_hws_item_t __kmp_hws_socket; 889 extern kmp_hws_item_t __kmp_hws_die; 890 extern kmp_hws_item_t __kmp_hws_node; 891 extern kmp_hws_item_t __kmp_hws_tile; 892 extern kmp_hws_item_t __kmp_hws_core; 893 extern kmp_hws_item_t __kmp_hws_proc; 894 extern int __kmp_hws_requested; 895 extern int __kmp_hws_abs_flag; // absolute or per-item number requested 896 897 /* ------------------------------------------------------------------------ */ 898 899 #define KMP_PAD(type, sz) \ 900 (sizeof(type) + (sz - ((sizeof(type) - 1) % (sz)) - 1)) 901 902 // We need to avoid using -1 as a GTID as +1 is added to the gtid 903 // when storing it in a lock, and the value 0 is reserved. 904 #define KMP_GTID_DNE (-2) /* Does not exist */ 905 #define KMP_GTID_SHUTDOWN (-3) /* Library is shutting down */ 906 #define KMP_GTID_MONITOR (-4) /* Monitor thread ID */ 907 #define KMP_GTID_UNKNOWN (-5) /* Is not known */ 908 #define KMP_GTID_MIN (-6) /* Minimal gtid for low bound check in DEBUG */ 909 910 /* OpenMP 5.0 Memory Management support */ 911 912 #ifndef __OMP_H 913 // Duplicate type definitions from omp.h 914 typedef uintptr_t omp_uintptr_t; 915 916 typedef enum { 917 omp_atk_sync_hint = 1, 918 omp_atk_alignment = 2, 919 omp_atk_access = 3, 920 omp_atk_pool_size = 4, 921 omp_atk_fallback = 5, 922 omp_atk_fb_data = 6, 923 omp_atk_pinned = 7, 924 omp_atk_partition = 8 925 } omp_alloctrait_key_t; 926 927 typedef enum { 928 omp_atv_false = 0, 929 omp_atv_true = 1, 930 omp_atv_contended = 3, 931 omp_atv_uncontended = 4, 932 omp_atv_serialized = 5, 933 omp_atv_sequential = omp_atv_serialized, // (deprecated) 934 omp_atv_private = 6, 935 omp_atv_all = 7, 936 omp_atv_thread = 8, 937 omp_atv_pteam = 9, 938 omp_atv_cgroup = 10, 939 omp_atv_default_mem_fb = 11, 940 omp_atv_null_fb = 12, 941 omp_atv_abort_fb = 13, 942 omp_atv_allocator_fb = 14, 943 omp_atv_environment = 15, 944 omp_atv_nearest = 16, 945 omp_atv_blocked = 17, 946 omp_atv_interleaved = 18 947 } omp_alloctrait_value_t; 948 #define omp_atv_default ((omp_uintptr_t)-1) 949 950 typedef void *omp_memspace_handle_t; 951 extern omp_memspace_handle_t const omp_default_mem_space; 952 extern omp_memspace_handle_t const omp_large_cap_mem_space; 953 extern omp_memspace_handle_t const omp_const_mem_space; 954 extern omp_memspace_handle_t const omp_high_bw_mem_space; 955 extern omp_memspace_handle_t const omp_low_lat_mem_space; 956 // Preview of target memory support 957 extern omp_memspace_handle_t const llvm_omp_target_host_mem_space; 958 extern omp_memspace_handle_t const llvm_omp_target_shared_mem_space; 959 extern omp_memspace_handle_t const llvm_omp_target_device_mem_space; 960 961 typedef struct { 962 omp_alloctrait_key_t key; 963 omp_uintptr_t value; 964 } omp_alloctrait_t; 965 966 typedef void *omp_allocator_handle_t; 967 extern omp_allocator_handle_t const omp_null_allocator; 968 extern omp_allocator_handle_t const omp_default_mem_alloc; 969 extern omp_allocator_handle_t const omp_large_cap_mem_alloc; 970 extern omp_allocator_handle_t const omp_const_mem_alloc; 971 extern omp_allocator_handle_t const omp_high_bw_mem_alloc; 972 extern omp_allocator_handle_t const omp_low_lat_mem_alloc; 973 extern omp_allocator_handle_t const omp_cgroup_mem_alloc; 974 extern omp_allocator_handle_t const omp_pteam_mem_alloc; 975 extern omp_allocator_handle_t const omp_thread_mem_alloc; 976 // Preview of target memory support 977 extern omp_allocator_handle_t const llvm_omp_target_host_mem_alloc; 978 extern omp_allocator_handle_t const llvm_omp_target_shared_mem_alloc; 979 extern omp_allocator_handle_t const llvm_omp_target_device_mem_alloc; 980 extern omp_allocator_handle_t const kmp_max_mem_alloc; 981 extern omp_allocator_handle_t __kmp_def_allocator; 982 983 // end of duplicate type definitions from omp.h 984 #endif 985 986 extern int __kmp_memkind_available; 987 988 typedef omp_memspace_handle_t kmp_memspace_t; // placeholder 989 990 typedef struct kmp_allocator_t { 991 omp_memspace_handle_t memspace; 992 void **memkind; // pointer to memkind 993 size_t alignment; 994 omp_alloctrait_value_t fb; 995 kmp_allocator_t *fb_data; 996 kmp_uint64 pool_size; 997 kmp_uint64 pool_used; 998 } kmp_allocator_t; 999 1000 extern omp_allocator_handle_t __kmpc_init_allocator(int gtid, 1001 omp_memspace_handle_t, 1002 int ntraits, 1003 omp_alloctrait_t traits[]); 1004 extern void __kmpc_destroy_allocator(int gtid, omp_allocator_handle_t al); 1005 extern void __kmpc_set_default_allocator(int gtid, omp_allocator_handle_t al); 1006 extern omp_allocator_handle_t __kmpc_get_default_allocator(int gtid); 1007 // external interfaces, may be used by compiler 1008 extern void *__kmpc_alloc(int gtid, size_t sz, omp_allocator_handle_t al); 1009 extern void *__kmpc_aligned_alloc(int gtid, size_t align, size_t sz, 1010 omp_allocator_handle_t al); 1011 extern void *__kmpc_calloc(int gtid, size_t nmemb, size_t sz, 1012 omp_allocator_handle_t al); 1013 extern void *__kmpc_realloc(int gtid, void *ptr, size_t sz, 1014 omp_allocator_handle_t al, 1015 omp_allocator_handle_t free_al); 1016 extern void __kmpc_free(int gtid, void *ptr, omp_allocator_handle_t al); 1017 // internal interfaces, contain real implementation 1018 extern void *__kmp_alloc(int gtid, size_t align, size_t sz, 1019 omp_allocator_handle_t al); 1020 extern void *__kmp_calloc(int gtid, size_t align, size_t nmemb, size_t sz, 1021 omp_allocator_handle_t al); 1022 extern void *__kmp_realloc(int gtid, void *ptr, size_t sz, 1023 omp_allocator_handle_t al, 1024 omp_allocator_handle_t free_al); 1025 extern void ___kmpc_free(int gtid, void *ptr, omp_allocator_handle_t al); 1026 1027 extern void __kmp_init_memkind(); 1028 extern void __kmp_fini_memkind(); 1029 extern void __kmp_init_target_mem(); 1030 1031 /* ------------------------------------------------------------------------ */ 1032 1033 #define KMP_UINT64_MAX \ 1034 (~((kmp_uint64)1 << ((sizeof(kmp_uint64) * (1 << 3)) - 1))) 1035 1036 #define KMP_MIN_NTH 1 1037 1038 #ifndef KMP_MAX_NTH 1039 #if defined(PTHREAD_THREADS_MAX) && PTHREAD_THREADS_MAX < INT_MAX 1040 #define KMP_MAX_NTH PTHREAD_THREADS_MAX 1041 #else 1042 #define KMP_MAX_NTH INT_MAX 1043 #endif 1044 #endif /* KMP_MAX_NTH */ 1045 1046 #ifdef PTHREAD_STACK_MIN 1047 #define KMP_MIN_STKSIZE PTHREAD_STACK_MIN 1048 #else 1049 #define KMP_MIN_STKSIZE ((size_t)(32 * 1024)) 1050 #endif 1051 1052 #define KMP_MAX_STKSIZE (~((size_t)1 << ((sizeof(size_t) * (1 << 3)) - 1))) 1053 1054 #if KMP_ARCH_X86 1055 #define KMP_DEFAULT_STKSIZE ((size_t)(2 * 1024 * 1024)) 1056 #elif KMP_ARCH_X86_64 1057 #define KMP_DEFAULT_STKSIZE ((size_t)(4 * 1024 * 1024)) 1058 #define KMP_BACKUP_STKSIZE ((size_t)(2 * 1024 * 1024)) 1059 #else 1060 #define KMP_DEFAULT_STKSIZE ((size_t)(1024 * 1024)) 1061 #endif 1062 1063 #define KMP_DEFAULT_MALLOC_POOL_INCR ((size_t)(1024 * 1024)) 1064 #define KMP_MIN_MALLOC_POOL_INCR ((size_t)(4 * 1024)) 1065 #define KMP_MAX_MALLOC_POOL_INCR \ 1066 (~((size_t)1 << ((sizeof(size_t) * (1 << 3)) - 1))) 1067 1068 #define KMP_MIN_STKOFFSET (0) 1069 #define KMP_MAX_STKOFFSET KMP_MAX_STKSIZE 1070 #if KMP_OS_DARWIN 1071 #define KMP_DEFAULT_STKOFFSET KMP_MIN_STKOFFSET 1072 #else 1073 #define KMP_DEFAULT_STKOFFSET CACHE_LINE 1074 #endif 1075 1076 #define KMP_MIN_STKPADDING (0) 1077 #define KMP_MAX_STKPADDING (2 * 1024 * 1024) 1078 1079 #define KMP_BLOCKTIME_MULTIPLIER \ 1080 (1000) /* number of blocktime units per second */ 1081 #define KMP_MIN_BLOCKTIME (0) 1082 #define KMP_MAX_BLOCKTIME \ 1083 (INT_MAX) /* Must be this for "infinite" setting the work */ 1084 1085 /* __kmp_blocktime is in milliseconds */ 1086 #define KMP_DEFAULT_BLOCKTIME (__kmp_is_hybrid_cpu() ? (0) : (200)) 1087 1088 #if KMP_USE_MONITOR 1089 #define KMP_DEFAULT_MONITOR_STKSIZE ((size_t)(64 * 1024)) 1090 #define KMP_MIN_MONITOR_WAKEUPS (1) // min times monitor wakes up per second 1091 #define KMP_MAX_MONITOR_WAKEUPS (1000) // max times monitor can wake up per sec 1092 1093 /* Calculate new number of monitor wakeups for a specific block time based on 1094 previous monitor_wakeups. Only allow increasing number of wakeups */ 1095 #define KMP_WAKEUPS_FROM_BLOCKTIME(blocktime, monitor_wakeups) \ 1096 (((blocktime) == KMP_MAX_BLOCKTIME) ? (monitor_wakeups) \ 1097 : ((blocktime) == KMP_MIN_BLOCKTIME) ? KMP_MAX_MONITOR_WAKEUPS \ 1098 : ((monitor_wakeups) > (KMP_BLOCKTIME_MULTIPLIER / (blocktime))) \ 1099 ? (monitor_wakeups) \ 1100 : (KMP_BLOCKTIME_MULTIPLIER) / (blocktime)) 1101 1102 /* Calculate number of intervals for a specific block time based on 1103 monitor_wakeups */ 1104 #define KMP_INTERVALS_FROM_BLOCKTIME(blocktime, monitor_wakeups) \ 1105 (((blocktime) + (KMP_BLOCKTIME_MULTIPLIER / (monitor_wakeups)) - 1) / \ 1106 (KMP_BLOCKTIME_MULTIPLIER / (monitor_wakeups))) 1107 #else 1108 #define KMP_BLOCKTIME(team, tid) \ 1109 (get__bt_set(team, tid) ? get__blocktime(team, tid) : __kmp_dflt_blocktime) 1110 #if KMP_OS_UNIX && (KMP_ARCH_X86 || KMP_ARCH_X86_64) 1111 // HW TSC is used to reduce overhead (clock tick instead of nanosecond). 1112 extern kmp_uint64 __kmp_ticks_per_msec; 1113 #if KMP_COMPILER_ICC 1114 #define KMP_NOW() ((kmp_uint64)_rdtsc()) 1115 #else 1116 #define KMP_NOW() __kmp_hardware_timestamp() 1117 #endif 1118 #define KMP_NOW_MSEC() (KMP_NOW() / __kmp_ticks_per_msec) 1119 #define KMP_BLOCKTIME_INTERVAL(team, tid) \ 1120 (KMP_BLOCKTIME(team, tid) * __kmp_ticks_per_msec) 1121 #define KMP_BLOCKING(goal, count) ((goal) > KMP_NOW()) 1122 #else 1123 // System time is retrieved sporadically while blocking. 1124 extern kmp_uint64 __kmp_now_nsec(); 1125 #define KMP_NOW() __kmp_now_nsec() 1126 #define KMP_NOW_MSEC() (KMP_NOW() / KMP_USEC_PER_SEC) 1127 #define KMP_BLOCKTIME_INTERVAL(team, tid) \ 1128 (KMP_BLOCKTIME(team, tid) * KMP_USEC_PER_SEC) 1129 #define KMP_BLOCKING(goal, count) ((count) % 1000 != 0 || (goal) > KMP_NOW()) 1130 #endif 1131 #endif // KMP_USE_MONITOR 1132 1133 #define KMP_MIN_STATSCOLS 40 1134 #define KMP_MAX_STATSCOLS 4096 1135 #define KMP_DEFAULT_STATSCOLS 80 1136 1137 #define KMP_MIN_INTERVAL 0 1138 #define KMP_MAX_INTERVAL (INT_MAX - 1) 1139 #define KMP_DEFAULT_INTERVAL 0 1140 1141 #define KMP_MIN_CHUNK 1 1142 #define KMP_MAX_CHUNK (INT_MAX - 1) 1143 #define KMP_DEFAULT_CHUNK 1 1144 1145 #define KMP_MIN_DISP_NUM_BUFF 1 1146 #define KMP_DFLT_DISP_NUM_BUFF 7 1147 #define KMP_MAX_DISP_NUM_BUFF 4096 1148 1149 #define KMP_MAX_ORDERED 8 1150 1151 #define KMP_MAX_FIELDS 32 1152 1153 #define KMP_MAX_BRANCH_BITS 31 1154 1155 #define KMP_MAX_ACTIVE_LEVELS_LIMIT INT_MAX 1156 1157 #define KMP_MAX_DEFAULT_DEVICE_LIMIT INT_MAX 1158 1159 #define KMP_MAX_TASK_PRIORITY_LIMIT INT_MAX 1160 1161 /* Minimum number of threads before switch to TLS gtid (experimentally 1162 determined) */ 1163 /* josh TODO: what about OS X* tuning? */ 1164 #if KMP_ARCH_X86 || KMP_ARCH_X86_64 1165 #define KMP_TLS_GTID_MIN 5 1166 #else 1167 #define KMP_TLS_GTID_MIN INT_MAX 1168 #endif 1169 1170 #define KMP_MASTER_TID(tid) (0 == (tid)) 1171 #define KMP_WORKER_TID(tid) (0 != (tid)) 1172 1173 #define KMP_MASTER_GTID(gtid) (0 == __kmp_tid_from_gtid((gtid))) 1174 #define KMP_WORKER_GTID(gtid) (0 != __kmp_tid_from_gtid((gtid))) 1175 #define KMP_INITIAL_GTID(gtid) (0 == (gtid)) 1176 1177 #ifndef TRUE 1178 #define FALSE 0 1179 #define TRUE (!FALSE) 1180 #endif 1181 1182 /* NOTE: all of the following constants must be even */ 1183 1184 #if KMP_OS_WINDOWS 1185 #define KMP_INIT_WAIT 64U /* initial number of spin-tests */ 1186 #define KMP_NEXT_WAIT 32U /* susequent number of spin-tests */ 1187 #elif KMP_OS_LINUX 1188 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */ 1189 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */ 1190 #elif KMP_OS_DARWIN 1191 /* TODO: tune for KMP_OS_DARWIN */ 1192 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */ 1193 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */ 1194 #elif KMP_OS_DRAGONFLY 1195 /* TODO: tune for KMP_OS_DRAGONFLY */ 1196 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */ 1197 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */ 1198 #elif KMP_OS_FREEBSD 1199 /* TODO: tune for KMP_OS_FREEBSD */ 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_NETBSD 1203 /* TODO: tune for KMP_OS_NETBSD */ 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_HURD 1207 /* TODO: tune for KMP_OS_HURD */ 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_OPENBSD 1211 /* TODO: tune for KMP_OS_OPENBSD */ 1212 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */ 1213 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */ 1214 #endif 1215 1216 #if KMP_ARCH_X86 || KMP_ARCH_X86_64 1217 typedef struct kmp_cpuid { 1218 kmp_uint32 eax; 1219 kmp_uint32 ebx; 1220 kmp_uint32 ecx; 1221 kmp_uint32 edx; 1222 } kmp_cpuid_t; 1223 1224 typedef struct kmp_cpuinfo_flags_t { 1225 unsigned sse2 : 1; // 0 if SSE2 instructions are not supported, 1 otherwise. 1226 unsigned rtm : 1; // 0 if RTM instructions are not supported, 1 otherwise. 1227 unsigned hybrid : 1; 1228 unsigned reserved : 29; // Ensure size of 32 bits 1229 } kmp_cpuinfo_flags_t; 1230 1231 typedef struct kmp_cpuinfo { 1232 int initialized; // If 0, other fields are not initialized. 1233 int signature; // CPUID(1).EAX 1234 int family; // CPUID(1).EAX[27:20]+CPUID(1).EAX[11:8] (Extended Family+Family) 1235 int model; // ( CPUID(1).EAX[19:16] << 4 ) + CPUID(1).EAX[7:4] ( ( Extended 1236 // Model << 4 ) + Model) 1237 int stepping; // CPUID(1).EAX[3:0] ( Stepping ) 1238 kmp_cpuinfo_flags_t flags; 1239 int apic_id; 1240 int physical_id; 1241 int logical_id; 1242 kmp_uint64 frequency; // Nominal CPU frequency in Hz. 1243 char name[3 * sizeof(kmp_cpuid_t)]; // CPUID(0x80000002,0x80000003,0x80000004) 1244 } kmp_cpuinfo_t; 1245 1246 extern void __kmp_query_cpuid(kmp_cpuinfo_t *p); 1247 1248 #if KMP_OS_UNIX 1249 // subleaf is only needed for cache and topology discovery and can be set to 1250 // zero in most cases 1251 static inline void __kmp_x86_cpuid(int leaf, int subleaf, struct kmp_cpuid *p) { 1252 __asm__ __volatile__("cpuid" 1253 : "=a"(p->eax), "=b"(p->ebx), "=c"(p->ecx), "=d"(p->edx) 1254 : "a"(leaf), "c"(subleaf)); 1255 } 1256 // Load p into FPU control word 1257 static inline void __kmp_load_x87_fpu_control_word(const kmp_int16 *p) { 1258 __asm__ __volatile__("fldcw %0" : : "m"(*p)); 1259 } 1260 // Store FPU control word into p 1261 static inline void __kmp_store_x87_fpu_control_word(kmp_int16 *p) { 1262 __asm__ __volatile__("fstcw %0" : "=m"(*p)); 1263 } 1264 static inline void __kmp_clear_x87_fpu_status_word() { 1265 #if KMP_MIC 1266 // 32-bit protected mode x87 FPU state 1267 struct x87_fpu_state { 1268 unsigned cw; 1269 unsigned sw; 1270 unsigned tw; 1271 unsigned fip; 1272 unsigned fips; 1273 unsigned fdp; 1274 unsigned fds; 1275 }; 1276 struct x87_fpu_state fpu_state = {0, 0, 0, 0, 0, 0, 0}; 1277 __asm__ __volatile__("fstenv %0\n\t" // store FP env 1278 "andw $0x7f00, %1\n\t" // clear 0-7,15 bits of FP SW 1279 "fldenv %0\n\t" // load FP env back 1280 : "+m"(fpu_state), "+m"(fpu_state.sw)); 1281 #else 1282 __asm__ __volatile__("fnclex"); 1283 #endif // KMP_MIC 1284 } 1285 #if __SSE__ 1286 static inline void __kmp_load_mxcsr(const kmp_uint32 *p) { _mm_setcsr(*p); } 1287 static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = _mm_getcsr(); } 1288 #else 1289 static inline void __kmp_load_mxcsr(const kmp_uint32 *p) {} 1290 static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = 0; } 1291 #endif 1292 #else 1293 // Windows still has these as external functions in assembly file 1294 extern void __kmp_x86_cpuid(int mode, int mode2, struct kmp_cpuid *p); 1295 extern void __kmp_load_x87_fpu_control_word(const kmp_int16 *p); 1296 extern void __kmp_store_x87_fpu_control_word(kmp_int16 *p); 1297 extern void __kmp_clear_x87_fpu_status_word(); 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 #endif // KMP_OS_UNIX 1301 1302 #define KMP_X86_MXCSR_MASK 0xffffffc0 /* ignore status flags (6 lsb) */ 1303 1304 #if KMP_ARCH_X86 1305 extern void __kmp_x86_pause(void); 1306 #elif KMP_MIC 1307 // Performance testing on KNC (C0QS-7120 P/A/X/D, 61-core, 16 GB Memory) showed 1308 // regression after removal of extra PAUSE from spin loops. Changing 1309 // the delay from 100 to 300 showed even better performance than double PAUSE 1310 // on Spec OMP2001 and LCPC tasking tests, no regressions on EPCC. 1311 static inline void __kmp_x86_pause(void) { _mm_delay_32(300); } 1312 #else 1313 static inline void __kmp_x86_pause(void) { _mm_pause(); } 1314 #endif 1315 #define KMP_CPU_PAUSE() __kmp_x86_pause() 1316 #elif KMP_ARCH_PPC64 1317 #define KMP_PPC64_PRI_LOW() __asm__ volatile("or 1, 1, 1") 1318 #define KMP_PPC64_PRI_MED() __asm__ volatile("or 2, 2, 2") 1319 #define KMP_PPC64_PRI_LOC_MB() __asm__ volatile("" : : : "memory") 1320 #define KMP_CPU_PAUSE() \ 1321 do { \ 1322 KMP_PPC64_PRI_LOW(); \ 1323 KMP_PPC64_PRI_MED(); \ 1324 KMP_PPC64_PRI_LOC_MB(); \ 1325 } while (0) 1326 #else 1327 #define KMP_CPU_PAUSE() /* nothing to do */ 1328 #endif 1329 1330 #define KMP_INIT_YIELD(count) \ 1331 { (count) = __kmp_yield_init; } 1332 1333 #define KMP_OVERSUBSCRIBED \ 1334 (TCR_4(__kmp_nth) > (__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc)) 1335 1336 #define KMP_TRY_YIELD \ 1337 ((__kmp_use_yield == 1) || (__kmp_use_yield == 2 && (KMP_OVERSUBSCRIBED))) 1338 1339 #define KMP_TRY_YIELD_OVERSUB \ 1340 ((__kmp_use_yield == 1 || __kmp_use_yield == 2) && (KMP_OVERSUBSCRIBED)) 1341 1342 #define KMP_YIELD(cond) \ 1343 { \ 1344 KMP_CPU_PAUSE(); \ 1345 if ((cond) && (KMP_TRY_YIELD)) \ 1346 __kmp_yield(); \ 1347 } 1348 1349 #define KMP_YIELD_OVERSUB() \ 1350 { \ 1351 KMP_CPU_PAUSE(); \ 1352 if ((KMP_TRY_YIELD_OVERSUB)) \ 1353 __kmp_yield(); \ 1354 } 1355 1356 // Note the decrement of 2 in the following Macros. With KMP_LIBRARY=turnaround, 1357 // there should be no yielding since initial value from KMP_INIT_YIELD() is odd. 1358 #define KMP_YIELD_SPIN(count) \ 1359 { \ 1360 KMP_CPU_PAUSE(); \ 1361 if (KMP_TRY_YIELD) { \ 1362 (count) -= 2; \ 1363 if (!(count)) { \ 1364 __kmp_yield(); \ 1365 (count) = __kmp_yield_next; \ 1366 } \ 1367 } \ 1368 } 1369 1370 #define KMP_YIELD_OVERSUB_ELSE_SPIN(count) \ 1371 { \ 1372 KMP_CPU_PAUSE(); \ 1373 if ((KMP_TRY_YIELD_OVERSUB)) \ 1374 __kmp_yield(); \ 1375 else if (__kmp_use_yield == 1) { \ 1376 (count) -= 2; \ 1377 if (!(count)) { \ 1378 __kmp_yield(); \ 1379 (count) = __kmp_yield_next; \ 1380 } \ 1381 } \ 1382 } 1383 1384 // User-level Monitor/Mwait 1385 #if KMP_HAVE_UMWAIT 1386 // We always try for UMWAIT first 1387 #if KMP_HAVE_WAITPKG_INTRINSICS 1388 #if KMP_HAVE_IMMINTRIN_H 1389 #include <immintrin.h> 1390 #elif KMP_HAVE_INTRIN_H 1391 #include <intrin.h> 1392 #endif 1393 #endif // KMP_HAVE_WAITPKG_INTRINSICS 1394 KMP_ATTRIBUTE_TARGET_WAITPKG 1395 static inline int __kmp_tpause(uint32_t hint, uint64_t counter) { 1396 #if !KMP_HAVE_WAITPKG_INTRINSICS 1397 uint32_t timeHi = uint32_t(counter >> 32); 1398 uint32_t timeLo = uint32_t(counter & 0xffffffff); 1399 char flag; 1400 __asm__ volatile("#tpause\n.byte 0x66, 0x0F, 0xAE, 0xF1\n" 1401 "setb %0" 1402 : "=r"(flag) 1403 : "a"(timeLo), "d"(timeHi), "c"(hint) 1404 :); 1405 return flag; 1406 #else 1407 return _tpause(hint, counter); 1408 #endif 1409 } 1410 KMP_ATTRIBUTE_TARGET_WAITPKG 1411 static inline void __kmp_umonitor(void *cacheline) { 1412 #if !KMP_HAVE_WAITPKG_INTRINSICS 1413 __asm__ volatile("# umonitor\n.byte 0xF3, 0x0F, 0xAE, 0x01 " 1414 : 1415 : "a"(cacheline) 1416 :); 1417 #else 1418 _umonitor(cacheline); 1419 #endif 1420 } 1421 KMP_ATTRIBUTE_TARGET_WAITPKG 1422 static inline int __kmp_umwait(uint32_t hint, uint64_t counter) { 1423 #if !KMP_HAVE_WAITPKG_INTRINSICS 1424 uint32_t timeHi = uint32_t(counter >> 32); 1425 uint32_t timeLo = uint32_t(counter & 0xffffffff); 1426 char flag; 1427 __asm__ volatile("#umwait\n.byte 0xF2, 0x0F, 0xAE, 0xF1\n" 1428 "setb %0" 1429 : "=r"(flag) 1430 : "a"(timeLo), "d"(timeHi), "c"(hint) 1431 :); 1432 return flag; 1433 #else 1434 return _umwait(hint, counter); 1435 #endif 1436 } 1437 #elif KMP_HAVE_MWAIT 1438 #if KMP_OS_UNIX 1439 #include <pmmintrin.h> 1440 #else 1441 #include <intrin.h> 1442 #endif 1443 #if KMP_OS_UNIX 1444 __attribute__((target("sse3"))) 1445 #endif 1446 static inline void 1447 __kmp_mm_monitor(void *cacheline, unsigned extensions, unsigned hints) { 1448 _mm_monitor(cacheline, extensions, hints); 1449 } 1450 #if KMP_OS_UNIX 1451 __attribute__((target("sse3"))) 1452 #endif 1453 static inline void 1454 __kmp_mm_mwait(unsigned extensions, unsigned hints) { 1455 _mm_mwait(extensions, hints); 1456 } 1457 #endif // KMP_HAVE_UMWAIT 1458 1459 /* ------------------------------------------------------------------------ */ 1460 /* Support datatypes for the orphaned construct nesting checks. */ 1461 /* ------------------------------------------------------------------------ */ 1462 1463 /* When adding to this enum, add its corresponding string in cons_text_c[] 1464 * array in kmp_error.cpp */ 1465 enum cons_type { 1466 ct_none, 1467 ct_parallel, 1468 ct_pdo, 1469 ct_pdo_ordered, 1470 ct_psections, 1471 ct_psingle, 1472 ct_critical, 1473 ct_ordered_in_parallel, 1474 ct_ordered_in_pdo, 1475 ct_master, 1476 ct_reduce, 1477 ct_barrier, 1478 ct_masked 1479 }; 1480 1481 #define IS_CONS_TYPE_ORDERED(ct) ((ct) == ct_pdo_ordered) 1482 1483 struct cons_data { 1484 ident_t const *ident; 1485 enum cons_type type; 1486 int prev; 1487 kmp_user_lock_p 1488 name; /* address exclusively for critical section name comparison */ 1489 }; 1490 1491 struct cons_header { 1492 int p_top, w_top, s_top; 1493 int stack_size, stack_top; 1494 struct cons_data *stack_data; 1495 }; 1496 1497 struct kmp_region_info { 1498 char *text; 1499 int offset[KMP_MAX_FIELDS]; 1500 int length[KMP_MAX_FIELDS]; 1501 }; 1502 1503 /* ---------------------------------------------------------------------- */ 1504 /* ---------------------------------------------------------------------- */ 1505 1506 #if KMP_OS_WINDOWS 1507 typedef HANDLE kmp_thread_t; 1508 typedef DWORD kmp_key_t; 1509 #endif /* KMP_OS_WINDOWS */ 1510 1511 #if KMP_OS_UNIX 1512 typedef pthread_t kmp_thread_t; 1513 typedef pthread_key_t kmp_key_t; 1514 #endif 1515 1516 extern kmp_key_t __kmp_gtid_threadprivate_key; 1517 1518 typedef struct kmp_sys_info { 1519 long maxrss; /* the maximum resident set size utilized (in kilobytes) */ 1520 long minflt; /* the number of page faults serviced without any I/O */ 1521 long majflt; /* the number of page faults serviced that required I/O */ 1522 long nswap; /* the number of times a process was "swapped" out of memory */ 1523 long inblock; /* the number of times the file system had to perform input */ 1524 long oublock; /* the number of times the file system had to perform output */ 1525 long nvcsw; /* the number of times a context switch was voluntarily */ 1526 long nivcsw; /* the number of times a context switch was forced */ 1527 } kmp_sys_info_t; 1528 1529 #if USE_ITT_BUILD 1530 // We cannot include "kmp_itt.h" due to circular dependency. Declare the only 1531 // required type here. Later we will check the type meets requirements. 1532 typedef int kmp_itt_mark_t; 1533 #define KMP_ITT_DEBUG 0 1534 #endif /* USE_ITT_BUILD */ 1535 1536 typedef kmp_int32 kmp_critical_name[8]; 1537 1538 /*! 1539 @ingroup PARALLEL 1540 The type for a microtask which gets passed to @ref __kmpc_fork_call(). 1541 The arguments to the outlined function are 1542 @param global_tid the global thread identity of the thread executing the 1543 function. 1544 @param bound_tid the local identity of the thread executing the function 1545 @param ... pointers to shared variables accessed by the function. 1546 */ 1547 typedef void (*kmpc_micro)(kmp_int32 *global_tid, kmp_int32 *bound_tid, ...); 1548 typedef void (*kmpc_micro_bound)(kmp_int32 *bound_tid, kmp_int32 *bound_nth, 1549 ...); 1550 1551 /*! 1552 @ingroup THREADPRIVATE 1553 @{ 1554 */ 1555 /* --------------------------------------------------------------------------- 1556 */ 1557 /* Threadprivate initialization/finalization function declarations */ 1558 1559 /* for non-array objects: __kmpc_threadprivate_register() */ 1560 1561 /*! 1562 Pointer to the constructor function. 1563 The first argument is the <tt>this</tt> pointer 1564 */ 1565 typedef void *(*kmpc_ctor)(void *); 1566 1567 /*! 1568 Pointer to the destructor function. 1569 The first argument is the <tt>this</tt> pointer 1570 */ 1571 typedef void (*kmpc_dtor)( 1572 void * /*, size_t */); /* 2nd arg: magic number for KCC unused by Intel 1573 compiler */ 1574 /*! 1575 Pointer to an alternate constructor. 1576 The first argument is the <tt>this</tt> pointer. 1577 */ 1578 typedef void *(*kmpc_cctor)(void *, void *); 1579 1580 /* for array objects: __kmpc_threadprivate_register_vec() */ 1581 /* First arg: "this" pointer */ 1582 /* Last arg: number of array elements */ 1583 /*! 1584 Array constructor. 1585 First argument is the <tt>this</tt> pointer 1586 Second argument the number of array elements. 1587 */ 1588 typedef void *(*kmpc_ctor_vec)(void *, size_t); 1589 /*! 1590 Pointer to the array destructor function. 1591 The first argument is the <tt>this</tt> pointer 1592 Second argument the number of array elements. 1593 */ 1594 typedef void (*kmpc_dtor_vec)(void *, size_t); 1595 /*! 1596 Array constructor. 1597 First argument is the <tt>this</tt> pointer 1598 Third argument the number of array elements. 1599 */ 1600 typedef void *(*kmpc_cctor_vec)(void *, void *, 1601 size_t); /* function unused by compiler */ 1602 1603 /*! 1604 @} 1605 */ 1606 1607 /* keeps tracked of threadprivate cache allocations for cleanup later */ 1608 typedef struct kmp_cached_addr { 1609 void **addr; /* address of allocated cache */ 1610 void ***compiler_cache; /* pointer to compiler's cache */ 1611 void *data; /* pointer to global data */ 1612 struct kmp_cached_addr *next; /* pointer to next cached address */ 1613 } kmp_cached_addr_t; 1614 1615 struct private_data { 1616 struct private_data *next; /* The next descriptor in the list */ 1617 void *data; /* The data buffer for this descriptor */ 1618 int more; /* The repeat count for this descriptor */ 1619 size_t size; /* The data size for this descriptor */ 1620 }; 1621 1622 struct private_common { 1623 struct private_common *next; 1624 struct private_common *link; 1625 void *gbl_addr; 1626 void *par_addr; /* par_addr == gbl_addr for PRIMARY thread */ 1627 size_t cmn_size; 1628 }; 1629 1630 struct shared_common { 1631 struct shared_common *next; 1632 struct private_data *pod_init; 1633 void *obj_init; 1634 void *gbl_addr; 1635 union { 1636 kmpc_ctor ctor; 1637 kmpc_ctor_vec ctorv; 1638 } ct; 1639 union { 1640 kmpc_cctor cctor; 1641 kmpc_cctor_vec cctorv; 1642 } cct; 1643 union { 1644 kmpc_dtor dtor; 1645 kmpc_dtor_vec dtorv; 1646 } dt; 1647 size_t vec_len; 1648 int is_vec; 1649 size_t cmn_size; 1650 }; 1651 1652 #define KMP_HASH_TABLE_LOG2 9 /* log2 of the hash table size */ 1653 #define KMP_HASH_TABLE_SIZE \ 1654 (1 << KMP_HASH_TABLE_LOG2) /* size of the hash table */ 1655 #define KMP_HASH_SHIFT 3 /* throw away this many low bits from the address */ 1656 #define KMP_HASH(x) \ 1657 ((((kmp_uintptr_t)x) >> KMP_HASH_SHIFT) & (KMP_HASH_TABLE_SIZE - 1)) 1658 1659 struct common_table { 1660 struct private_common *data[KMP_HASH_TABLE_SIZE]; 1661 }; 1662 1663 struct shared_table { 1664 struct shared_common *data[KMP_HASH_TABLE_SIZE]; 1665 }; 1666 1667 /* ------------------------------------------------------------------------ */ 1668 1669 #if KMP_USE_HIER_SCHED 1670 // Shared barrier data that exists inside a single unit of the scheduling 1671 // hierarchy 1672 typedef struct kmp_hier_private_bdata_t { 1673 kmp_int32 num_active; 1674 kmp_uint64 index; 1675 kmp_uint64 wait_val[2]; 1676 } kmp_hier_private_bdata_t; 1677 #endif 1678 1679 typedef struct kmp_sched_flags { 1680 unsigned ordered : 1; 1681 unsigned nomerge : 1; 1682 unsigned contains_last : 1; 1683 #if KMP_USE_HIER_SCHED 1684 unsigned use_hier : 1; 1685 unsigned unused : 28; 1686 #else 1687 unsigned unused : 29; 1688 #endif 1689 } kmp_sched_flags_t; 1690 1691 KMP_BUILD_ASSERT(sizeof(kmp_sched_flags_t) == 4); 1692 1693 #if KMP_STATIC_STEAL_ENABLED 1694 typedef struct KMP_ALIGN_CACHE dispatch_private_info32 { 1695 kmp_int32 count; 1696 kmp_int32 ub; 1697 /* Adding KMP_ALIGN_CACHE here doesn't help / can hurt performance */ 1698 kmp_int32 lb; 1699 kmp_int32 st; 1700 kmp_int32 tc; 1701 kmp_lock_t *steal_lock; // lock used for chunk stealing 1702 // KMP_ALIGN(32) ensures (if the KMP_ALIGN macro is turned on) 1703 // a) parm3 is properly aligned and 1704 // b) all parm1-4 are on the same cache line. 1705 // Because of parm1-4 are used together, performance seems to be better 1706 // if they are on the same cache line (not measured though). 1707 1708 struct KMP_ALIGN(32) { // AC: changed 16 to 32 in order to simplify template 1709 kmp_int32 parm1; // structures in kmp_dispatch.cpp. This should 1710 kmp_int32 parm2; // make no real change at least while padding is off. 1711 kmp_int32 parm3; 1712 kmp_int32 parm4; 1713 }; 1714 1715 kmp_uint32 ordered_lower; 1716 kmp_uint32 ordered_upper; 1717 #if KMP_OS_WINDOWS 1718 kmp_int32 last_upper; 1719 #endif /* KMP_OS_WINDOWS */ 1720 } dispatch_private_info32_t; 1721 1722 typedef struct KMP_ALIGN_CACHE dispatch_private_info64 { 1723 kmp_int64 count; // current chunk number for static & static-steal scheduling 1724 kmp_int64 ub; /* upper-bound */ 1725 /* Adding KMP_ALIGN_CACHE here doesn't help / can hurt performance */ 1726 kmp_int64 lb; /* lower-bound */ 1727 kmp_int64 st; /* stride */ 1728 kmp_int64 tc; /* trip count (number of iterations) */ 1729 kmp_lock_t *steal_lock; // lock used for chunk stealing 1730 /* parm[1-4] are used in different ways by different scheduling algorithms */ 1731 1732 // KMP_ALIGN( 32 ) ensures ( if the KMP_ALIGN macro is turned on ) 1733 // a) parm3 is properly aligned and 1734 // b) all parm1-4 are in the same cache line. 1735 // Because of parm1-4 are used together, performance seems to be better 1736 // if they are in the same line (not measured though). 1737 1738 struct KMP_ALIGN(32) { 1739 kmp_int64 parm1; 1740 kmp_int64 parm2; 1741 kmp_int64 parm3; 1742 kmp_int64 parm4; 1743 }; 1744 1745 kmp_uint64 ordered_lower; 1746 kmp_uint64 ordered_upper; 1747 #if KMP_OS_WINDOWS 1748 kmp_int64 last_upper; 1749 #endif /* KMP_OS_WINDOWS */ 1750 } dispatch_private_info64_t; 1751 #else /* KMP_STATIC_STEAL_ENABLED */ 1752 typedef struct KMP_ALIGN_CACHE dispatch_private_info32 { 1753 kmp_int32 lb; 1754 kmp_int32 ub; 1755 kmp_int32 st; 1756 kmp_int32 tc; 1757 1758 kmp_int32 parm1; 1759 kmp_int32 parm2; 1760 kmp_int32 parm3; 1761 kmp_int32 parm4; 1762 1763 kmp_int32 count; 1764 1765 kmp_uint32 ordered_lower; 1766 kmp_uint32 ordered_upper; 1767 #if KMP_OS_WINDOWS 1768 kmp_int32 last_upper; 1769 #endif /* KMP_OS_WINDOWS */ 1770 } dispatch_private_info32_t; 1771 1772 typedef struct KMP_ALIGN_CACHE dispatch_private_info64 { 1773 kmp_int64 lb; /* lower-bound */ 1774 kmp_int64 ub; /* upper-bound */ 1775 kmp_int64 st; /* stride */ 1776 kmp_int64 tc; /* trip count (number of iterations) */ 1777 1778 /* parm[1-4] are used in different ways by different scheduling algorithms */ 1779 kmp_int64 parm1; 1780 kmp_int64 parm2; 1781 kmp_int64 parm3; 1782 kmp_int64 parm4; 1783 1784 kmp_int64 count; /* current chunk number for static scheduling */ 1785 1786 kmp_uint64 ordered_lower; 1787 kmp_uint64 ordered_upper; 1788 #if KMP_OS_WINDOWS 1789 kmp_int64 last_upper; 1790 #endif /* KMP_OS_WINDOWS */ 1791 } dispatch_private_info64_t; 1792 #endif /* KMP_STATIC_STEAL_ENABLED */ 1793 1794 typedef struct KMP_ALIGN_CACHE dispatch_private_info { 1795 union private_info { 1796 dispatch_private_info32_t p32; 1797 dispatch_private_info64_t p64; 1798 } u; 1799 enum sched_type schedule; /* scheduling algorithm */ 1800 kmp_sched_flags_t flags; /* flags (e.g., ordered, nomerge, etc.) */ 1801 std::atomic<kmp_uint32> steal_flag; // static_steal only, state of a buffer 1802 kmp_int32 ordered_bumped; 1803 // Stack of buffers for nest of serial regions 1804 struct dispatch_private_info *next; 1805 kmp_int32 type_size; /* the size of types in private_info */ 1806 #if KMP_USE_HIER_SCHED 1807 kmp_int32 hier_id; 1808 void *parent; /* hierarchical scheduling parent pointer */ 1809 #endif 1810 enum cons_type pushed_ws; 1811 } dispatch_private_info_t; 1812 1813 typedef struct dispatch_shared_info32 { 1814 /* chunk index under dynamic, number of idle threads under static-steal; 1815 iteration index otherwise */ 1816 volatile kmp_uint32 iteration; 1817 volatile kmp_int32 num_done; 1818 volatile kmp_uint32 ordered_iteration; 1819 // Dummy to retain the structure size after making ordered_iteration scalar 1820 kmp_int32 ordered_dummy[KMP_MAX_ORDERED - 1]; 1821 } dispatch_shared_info32_t; 1822 1823 typedef struct dispatch_shared_info64 { 1824 /* chunk index under dynamic, number of idle threads under static-steal; 1825 iteration index otherwise */ 1826 volatile kmp_uint64 iteration; 1827 volatile kmp_int64 num_done; 1828 volatile kmp_uint64 ordered_iteration; 1829 // Dummy to retain the structure size after making ordered_iteration scalar 1830 kmp_int64 ordered_dummy[KMP_MAX_ORDERED - 3]; 1831 } dispatch_shared_info64_t; 1832 1833 typedef struct dispatch_shared_info { 1834 union shared_info { 1835 dispatch_shared_info32_t s32; 1836 dispatch_shared_info64_t s64; 1837 } u; 1838 volatile kmp_uint32 buffer_index; 1839 volatile kmp_int32 doacross_buf_idx; // teamwise index 1840 volatile kmp_uint32 *doacross_flags; // shared array of iteration flags (0/1) 1841 kmp_int32 doacross_num_done; // count finished threads 1842 #if KMP_USE_HIER_SCHED 1843 void *hier; 1844 #endif 1845 #if KMP_USE_HWLOC 1846 // When linking with libhwloc, the ORDERED EPCC test slows down on big 1847 // machines (> 48 cores). Performance analysis showed that a cache thrash 1848 // was occurring and this padding helps alleviate the problem. 1849 char padding[64]; 1850 #endif 1851 } dispatch_shared_info_t; 1852 1853 typedef struct kmp_disp { 1854 /* Vector for ORDERED SECTION */ 1855 void (*th_deo_fcn)(int *gtid, int *cid, ident_t *); 1856 /* Vector for END ORDERED SECTION */ 1857 void (*th_dxo_fcn)(int *gtid, int *cid, ident_t *); 1858 1859 dispatch_shared_info_t *th_dispatch_sh_current; 1860 dispatch_private_info_t *th_dispatch_pr_current; 1861 1862 dispatch_private_info_t *th_disp_buffer; 1863 kmp_uint32 th_disp_index; 1864 kmp_int32 th_doacross_buf_idx; // thread's doacross buffer index 1865 volatile kmp_uint32 *th_doacross_flags; // pointer to shared array of flags 1866 kmp_int64 *th_doacross_info; // info on loop bounds 1867 #if KMP_USE_INTERNODE_ALIGNMENT 1868 char more_padding[INTERNODE_CACHE_LINE]; 1869 #endif 1870 } kmp_disp_t; 1871 1872 /* ------------------------------------------------------------------------ */ 1873 /* Barrier stuff */ 1874 1875 /* constants for barrier state update */ 1876 #define KMP_INIT_BARRIER_STATE 0 /* should probably start from zero */ 1877 #define KMP_BARRIER_SLEEP_BIT 0 /* bit used for suspend/sleep part of state */ 1878 #define KMP_BARRIER_UNUSED_BIT 1 // bit that must never be set for valid state 1879 #define KMP_BARRIER_BUMP_BIT 2 /* lsb used for bump of go/arrived state */ 1880 1881 #define KMP_BARRIER_SLEEP_STATE (1 << KMP_BARRIER_SLEEP_BIT) 1882 #define KMP_BARRIER_UNUSED_STATE (1 << KMP_BARRIER_UNUSED_BIT) 1883 #define KMP_BARRIER_STATE_BUMP (1 << KMP_BARRIER_BUMP_BIT) 1884 1885 #if (KMP_BARRIER_SLEEP_BIT >= KMP_BARRIER_BUMP_BIT) 1886 #error "Barrier sleep bit must be smaller than barrier bump bit" 1887 #endif 1888 #if (KMP_BARRIER_UNUSED_BIT >= KMP_BARRIER_BUMP_BIT) 1889 #error "Barrier unused bit must be smaller than barrier bump bit" 1890 #endif 1891 1892 // Constants for release barrier wait state: currently, hierarchical only 1893 #define KMP_BARRIER_NOT_WAITING 0 // Normal state; worker not in wait_sleep 1894 #define KMP_BARRIER_OWN_FLAG \ 1895 1 // Normal state; worker waiting on own b_go flag in release 1896 #define KMP_BARRIER_PARENT_FLAG \ 1897 2 // Special state; worker waiting on parent's b_go flag in release 1898 #define KMP_BARRIER_SWITCH_TO_OWN_FLAG \ 1899 3 // Special state; tells worker to shift from parent to own b_go 1900 #define KMP_BARRIER_SWITCHING \ 1901 4 // Special state; worker resets appropriate flag on wake-up 1902 1903 #define KMP_NOT_SAFE_TO_REAP \ 1904 0 // Thread th_reap_state: not safe to reap (tasking) 1905 #define KMP_SAFE_TO_REAP 1 // Thread th_reap_state: safe to reap (not tasking) 1906 1907 // The flag_type describes the storage used for the flag. 1908 enum flag_type { 1909 flag32, /**< atomic 32 bit flags */ 1910 flag64, /**< 64 bit flags */ 1911 atomic_flag64, /**< atomic 64 bit flags */ 1912 flag_oncore, /**< special 64-bit flag for on-core barrier (hierarchical) */ 1913 flag_unset 1914 }; 1915 1916 enum barrier_type { 1917 bs_plain_barrier = 0, /* 0, All non-fork/join barriers (except reduction 1918 barriers if enabled) */ 1919 bs_forkjoin_barrier, /* 1, All fork/join (parallel region) barriers */ 1920 #if KMP_FAST_REDUCTION_BARRIER 1921 bs_reduction_barrier, /* 2, All barriers that are used in reduction */ 1922 #endif // KMP_FAST_REDUCTION_BARRIER 1923 bs_last_barrier /* Just a placeholder to mark the end */ 1924 }; 1925 1926 // to work with reduction barriers just like with plain barriers 1927 #if !KMP_FAST_REDUCTION_BARRIER 1928 #define bs_reduction_barrier bs_plain_barrier 1929 #endif // KMP_FAST_REDUCTION_BARRIER 1930 1931 typedef enum kmp_bar_pat { /* Barrier communication patterns */ 1932 bp_linear_bar = 1933 0, /* Single level (degenerate) tree */ 1934 bp_tree_bar = 1935 1, /* Balanced tree with branching factor 2^n */ 1936 bp_hyper_bar = 2, /* Hypercube-embedded tree with min 1937 branching factor 2^n */ 1938 bp_hierarchical_bar = 3, /* Machine hierarchy tree */ 1939 bp_dist_bar = 4, /* Distributed barrier */ 1940 bp_last_bar /* Placeholder to mark the end */ 1941 } kmp_bar_pat_e; 1942 1943 #define KMP_BARRIER_ICV_PUSH 1 1944 1945 /* Record for holding the values of the internal controls stack records */ 1946 typedef struct kmp_internal_control { 1947 int serial_nesting_level; /* corresponds to the value of the 1948 th_team_serialized field */ 1949 kmp_int8 dynamic; /* internal control for dynamic adjustment of threads (per 1950 thread) */ 1951 kmp_int8 1952 bt_set; /* internal control for whether blocktime is explicitly set */ 1953 int blocktime; /* internal control for blocktime */ 1954 #if KMP_USE_MONITOR 1955 int bt_intervals; /* internal control for blocktime intervals */ 1956 #endif 1957 int nproc; /* internal control for #threads for next parallel region (per 1958 thread) */ 1959 int thread_limit; /* internal control for thread-limit-var */ 1960 int max_active_levels; /* internal control for max_active_levels */ 1961 kmp_r_sched_t 1962 sched; /* internal control for runtime schedule {sched,chunk} pair */ 1963 kmp_proc_bind_t proc_bind; /* internal control for affinity */ 1964 kmp_int32 default_device; /* internal control for default device */ 1965 struct kmp_internal_control *next; 1966 } kmp_internal_control_t; 1967 1968 static inline void copy_icvs(kmp_internal_control_t *dst, 1969 kmp_internal_control_t *src) { 1970 *dst = *src; 1971 } 1972 1973 /* Thread barrier needs volatile barrier fields */ 1974 typedef struct KMP_ALIGN_CACHE kmp_bstate { 1975 // th_fixed_icvs is aligned by virtue of kmp_bstate being aligned (and all 1976 // uses of it). It is not explicitly aligned below, because we *don't* want 1977 // it to be padded -- instead, we fit b_go into the same cache line with 1978 // th_fixed_icvs, enabling NGO cache lines stores in the hierarchical barrier. 1979 kmp_internal_control_t th_fixed_icvs; // Initial ICVs for the thread 1980 // Tuck b_go into end of th_fixed_icvs cache line, so it can be stored with 1981 // same NGO store 1982 volatile kmp_uint64 b_go; // STATE => task should proceed (hierarchical) 1983 KMP_ALIGN_CACHE volatile kmp_uint64 1984 b_arrived; // STATE => task reached synch point. 1985 kmp_uint32 *skip_per_level; 1986 kmp_uint32 my_level; 1987 kmp_int32 parent_tid; 1988 kmp_int32 old_tid; 1989 kmp_uint32 depth; 1990 struct kmp_bstate *parent_bar; 1991 kmp_team_t *team; 1992 kmp_uint64 leaf_state; 1993 kmp_uint32 nproc; 1994 kmp_uint8 base_leaf_kids; 1995 kmp_uint8 leaf_kids; 1996 kmp_uint8 offset; 1997 kmp_uint8 wait_flag; 1998 kmp_uint8 use_oncore_barrier; 1999 #if USE_DEBUGGER 2000 // The following field is intended for the debugger solely. Only the worker 2001 // thread itself accesses this field: the worker increases it by 1 when it 2002 // arrives to a barrier. 2003 KMP_ALIGN_CACHE kmp_uint b_worker_arrived; 2004 #endif /* USE_DEBUGGER */ 2005 } kmp_bstate_t; 2006 2007 union KMP_ALIGN_CACHE kmp_barrier_union { 2008 double b_align; /* use worst case alignment */ 2009 char b_pad[KMP_PAD(kmp_bstate_t, CACHE_LINE)]; 2010 kmp_bstate_t bb; 2011 }; 2012 2013 typedef union kmp_barrier_union kmp_balign_t; 2014 2015 /* Team barrier needs only non-volatile arrived counter */ 2016 union KMP_ALIGN_CACHE kmp_barrier_team_union { 2017 double b_align; /* use worst case alignment */ 2018 char b_pad[CACHE_LINE]; 2019 struct { 2020 kmp_uint64 b_arrived; /* STATE => task reached synch point. */ 2021 #if USE_DEBUGGER 2022 // The following two fields are indended for the debugger solely. Only 2023 // primary thread of the team accesses these fields: the first one is 2024 // increased by 1 when the primary thread arrives to a barrier, the second 2025 // one is increased by one when all the threads arrived. 2026 kmp_uint b_master_arrived; 2027 kmp_uint b_team_arrived; 2028 #endif 2029 }; 2030 }; 2031 2032 typedef union kmp_barrier_team_union kmp_balign_team_t; 2033 2034 /* Padding for Linux* OS pthreads condition variables and mutexes used to signal 2035 threads when a condition changes. This is to workaround an NPTL bug where 2036 padding was added to pthread_cond_t which caused the initialization routine 2037 to write outside of the structure if compiled on pre-NPTL threads. */ 2038 #if KMP_OS_WINDOWS 2039 typedef struct kmp_win32_mutex { 2040 /* The Lock */ 2041 CRITICAL_SECTION cs; 2042 } kmp_win32_mutex_t; 2043 2044 typedef struct kmp_win32_cond { 2045 /* Count of the number of waiters. */ 2046 int waiters_count_; 2047 2048 /* Serialize access to <waiters_count_> */ 2049 kmp_win32_mutex_t waiters_count_lock_; 2050 2051 /* Number of threads to release via a <cond_broadcast> or a <cond_signal> */ 2052 int release_count_; 2053 2054 /* Keeps track of the current "generation" so that we don't allow */ 2055 /* one thread to steal all the "releases" from the broadcast. */ 2056 int wait_generation_count_; 2057 2058 /* A manual-reset event that's used to block and release waiting threads. */ 2059 HANDLE event_; 2060 } kmp_win32_cond_t; 2061 #endif 2062 2063 #if KMP_OS_UNIX 2064 2065 union KMP_ALIGN_CACHE kmp_cond_union { 2066 double c_align; 2067 char c_pad[CACHE_LINE]; 2068 pthread_cond_t c_cond; 2069 }; 2070 2071 typedef union kmp_cond_union kmp_cond_align_t; 2072 2073 union KMP_ALIGN_CACHE kmp_mutex_union { 2074 double m_align; 2075 char m_pad[CACHE_LINE]; 2076 pthread_mutex_t m_mutex; 2077 }; 2078 2079 typedef union kmp_mutex_union kmp_mutex_align_t; 2080 2081 #endif /* KMP_OS_UNIX */ 2082 2083 typedef struct kmp_desc_base { 2084 void *ds_stackbase; 2085 size_t ds_stacksize; 2086 int ds_stackgrow; 2087 kmp_thread_t ds_thread; 2088 volatile int ds_tid; 2089 int ds_gtid; 2090 #if KMP_OS_WINDOWS 2091 volatile int ds_alive; 2092 DWORD ds_thread_id; 2093 /* ds_thread keeps thread handle on Windows* OS. It is enough for RTL purposes. 2094 However, debugger support (libomp_db) cannot work with handles, because they 2095 uncomparable. For example, debugger requests info about thread with handle h. 2096 h is valid within debugger process, and meaningless within debugee process. 2097 Even if h is duped by call to DuplicateHandle(), so the result h' is valid 2098 within debugee process, but it is a *new* handle which does *not* equal to 2099 any other handle in debugee... The only way to compare handles is convert 2100 them to system-wide ids. GetThreadId() function is available only in 2101 Longhorn and Server 2003. :-( In contrast, GetCurrentThreadId() is available 2102 on all Windows* OS flavours (including Windows* 95). Thus, we have to get 2103 thread id by call to GetCurrentThreadId() from within the thread and save it 2104 to let libomp_db identify threads. */ 2105 #endif /* KMP_OS_WINDOWS */ 2106 } kmp_desc_base_t; 2107 2108 typedef union KMP_ALIGN_CACHE kmp_desc { 2109 double ds_align; /* use worst case alignment */ 2110 char ds_pad[KMP_PAD(kmp_desc_base_t, CACHE_LINE)]; 2111 kmp_desc_base_t ds; 2112 } kmp_desc_t; 2113 2114 typedef struct kmp_local { 2115 volatile int this_construct; /* count of single's encountered by thread */ 2116 void *reduce_data; 2117 #if KMP_USE_BGET 2118 void *bget_data; 2119 void *bget_list; 2120 #if !USE_CMP_XCHG_FOR_BGET 2121 #ifdef USE_QUEUING_LOCK_FOR_BGET 2122 kmp_lock_t bget_lock; /* Lock for accessing bget free list */ 2123 #else 2124 kmp_bootstrap_lock_t bget_lock; // Lock for accessing bget free list. Must be 2125 // bootstrap lock so we can use it at library 2126 // shutdown. 2127 #endif /* USE_LOCK_FOR_BGET */ 2128 #endif /* ! USE_CMP_XCHG_FOR_BGET */ 2129 #endif /* KMP_USE_BGET */ 2130 2131 PACKED_REDUCTION_METHOD_T 2132 packed_reduction_method; /* stored by __kmpc_reduce*(), used by 2133 __kmpc_end_reduce*() */ 2134 2135 } kmp_local_t; 2136 2137 #define KMP_CHECK_UPDATE(a, b) \ 2138 if ((a) != (b)) \ 2139 (a) = (b) 2140 #define KMP_CHECK_UPDATE_SYNC(a, b) \ 2141 if ((a) != (b)) \ 2142 TCW_SYNC_PTR((a), (b)) 2143 2144 #define get__blocktime(xteam, xtid) \ 2145 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.blocktime) 2146 #define get__bt_set(xteam, xtid) \ 2147 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_set) 2148 #if KMP_USE_MONITOR 2149 #define get__bt_intervals(xteam, xtid) \ 2150 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_intervals) 2151 #endif 2152 2153 #define get__dynamic_2(xteam, xtid) \ 2154 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.dynamic) 2155 #define get__nproc_2(xteam, xtid) \ 2156 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.nproc) 2157 #define get__sched_2(xteam, xtid) \ 2158 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.sched) 2159 2160 #define set__blocktime_team(xteam, xtid, xval) \ 2161 (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.blocktime) = \ 2162 (xval)) 2163 2164 #if KMP_USE_MONITOR 2165 #define set__bt_intervals_team(xteam, xtid, xval) \ 2166 (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_intervals) = \ 2167 (xval)) 2168 #endif 2169 2170 #define set__bt_set_team(xteam, xtid, xval) \ 2171 (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_set) = (xval)) 2172 2173 #define set__dynamic(xthread, xval) \ 2174 (((xthread)->th.th_current_task->td_icvs.dynamic) = (xval)) 2175 #define get__dynamic(xthread) \ 2176 (((xthread)->th.th_current_task->td_icvs.dynamic) ? (FTN_TRUE) : (FTN_FALSE)) 2177 2178 #define set__nproc(xthread, xval) \ 2179 (((xthread)->th.th_current_task->td_icvs.nproc) = (xval)) 2180 2181 #define set__thread_limit(xthread, xval) \ 2182 (((xthread)->th.th_current_task->td_icvs.thread_limit) = (xval)) 2183 2184 #define set__max_active_levels(xthread, xval) \ 2185 (((xthread)->th.th_current_task->td_icvs.max_active_levels) = (xval)) 2186 2187 #define get__max_active_levels(xthread) \ 2188 ((xthread)->th.th_current_task->td_icvs.max_active_levels) 2189 2190 #define set__sched(xthread, xval) \ 2191 (((xthread)->th.th_current_task->td_icvs.sched) = (xval)) 2192 2193 #define set__proc_bind(xthread, xval) \ 2194 (((xthread)->th.th_current_task->td_icvs.proc_bind) = (xval)) 2195 #define get__proc_bind(xthread) \ 2196 ((xthread)->th.th_current_task->td_icvs.proc_bind) 2197 2198 // OpenMP tasking data structures 2199 2200 typedef enum kmp_tasking_mode { 2201 tskm_immediate_exec = 0, 2202 tskm_extra_barrier = 1, 2203 tskm_task_teams = 2, 2204 tskm_max = 2 2205 } kmp_tasking_mode_t; 2206 2207 extern kmp_tasking_mode_t 2208 __kmp_tasking_mode; /* determines how/when to execute tasks */ 2209 extern int __kmp_task_stealing_constraint; 2210 extern int __kmp_enable_task_throttling; 2211 extern kmp_int32 __kmp_default_device; // Set via OMP_DEFAULT_DEVICE if 2212 // specified, defaults to 0 otherwise 2213 // Set via OMP_MAX_TASK_PRIORITY if specified, defaults to 0 otherwise 2214 extern kmp_int32 __kmp_max_task_priority; 2215 // Set via KMP_TASKLOOP_MIN_TASKS if specified, defaults to 0 otherwise 2216 extern kmp_uint64 __kmp_taskloop_min_tasks; 2217 2218 /* NOTE: kmp_taskdata_t and kmp_task_t structures allocated in single block with 2219 taskdata first */ 2220 #define KMP_TASK_TO_TASKDATA(task) (((kmp_taskdata_t *)task) - 1) 2221 #define KMP_TASKDATA_TO_TASK(taskdata) (kmp_task_t *)(taskdata + 1) 2222 2223 // The tt_found_tasks flag is a signal to all threads in the team that tasks 2224 // were spawned and queued since the previous barrier release. 2225 #define KMP_TASKING_ENABLED(task_team) \ 2226 (TRUE == TCR_SYNC_4((task_team)->tt.tt_found_tasks)) 2227 /*! 2228 @ingroup BASIC_TYPES 2229 @{ 2230 */ 2231 2232 /*! 2233 */ 2234 typedef kmp_int32 (*kmp_routine_entry_t)(kmp_int32, void *); 2235 2236 typedef union kmp_cmplrdata { 2237 kmp_int32 priority; /**< priority specified by user for the task */ 2238 kmp_routine_entry_t 2239 destructors; /* pointer to function to invoke deconstructors of 2240 firstprivate C++ objects */ 2241 /* future data */ 2242 } kmp_cmplrdata_t; 2243 2244 /* sizeof_kmp_task_t passed as arg to kmpc_omp_task call */ 2245 /*! 2246 */ 2247 typedef struct kmp_task { /* GEH: Shouldn't this be aligned somehow? */ 2248 void *shareds; /**< pointer to block of pointers to shared vars */ 2249 kmp_routine_entry_t 2250 routine; /**< pointer to routine to call for executing task */ 2251 kmp_int32 part_id; /**< part id for the task */ 2252 kmp_cmplrdata_t 2253 data1; /* Two known optional additions: destructors and priority */ 2254 kmp_cmplrdata_t data2; /* Process destructors first, priority second */ 2255 /* future data */ 2256 /* private vars */ 2257 } kmp_task_t; 2258 2259 /*! 2260 @} 2261 */ 2262 2263 typedef struct kmp_taskgroup { 2264 std::atomic<kmp_int32> count; // number of allocated and incomplete tasks 2265 std::atomic<kmp_int32> 2266 cancel_request; // request for cancellation of this taskgroup 2267 struct kmp_taskgroup *parent; // parent taskgroup 2268 // Block of data to perform task reduction 2269 void *reduce_data; // reduction related info 2270 kmp_int32 reduce_num_data; // number of data items to reduce 2271 uintptr_t *gomp_data; // gomp reduction data 2272 } kmp_taskgroup_t; 2273 2274 // forward declarations 2275 typedef union kmp_depnode kmp_depnode_t; 2276 typedef struct kmp_depnode_list kmp_depnode_list_t; 2277 typedef struct kmp_dephash_entry kmp_dephash_entry_t; 2278 2279 // macros for checking dep flag as an integer 2280 #define KMP_DEP_IN 0x1 2281 #define KMP_DEP_OUT 0x2 2282 #define KMP_DEP_INOUT 0x3 2283 #define KMP_DEP_MTX 0x4 2284 #define KMP_DEP_SET 0x8 2285 #define KMP_DEP_ALL 0x80 2286 // Compiler sends us this info: 2287 typedef struct kmp_depend_info { 2288 kmp_intptr_t base_addr; 2289 size_t len; 2290 union { 2291 kmp_uint8 flag; // flag as an unsigned char 2292 struct { // flag as a set of 8 bits 2293 unsigned in : 1; 2294 unsigned out : 1; 2295 unsigned mtx : 1; 2296 unsigned set : 1; 2297 unsigned unused : 3; 2298 unsigned all : 1; 2299 } flags; 2300 }; 2301 } kmp_depend_info_t; 2302 2303 // Internal structures to work with task dependencies: 2304 struct kmp_depnode_list { 2305 kmp_depnode_t *node; 2306 kmp_depnode_list_t *next; 2307 }; 2308 2309 // Max number of mutexinoutset dependencies per node 2310 #define MAX_MTX_DEPS 4 2311 2312 typedef struct kmp_base_depnode { 2313 kmp_depnode_list_t *successors; /* used under lock */ 2314 kmp_task_t *task; /* non-NULL if depnode is active, used under lock */ 2315 kmp_lock_t *mtx_locks[MAX_MTX_DEPS]; /* lock mutexinoutset dependent tasks */ 2316 kmp_int32 mtx_num_locks; /* number of locks in mtx_locks array */ 2317 kmp_lock_t lock; /* guards shared fields: task, successors */ 2318 #if KMP_SUPPORT_GRAPH_OUTPUT 2319 kmp_uint32 id; 2320 #endif 2321 std::atomic<kmp_int32> npredecessors; 2322 std::atomic<kmp_int32> nrefs; 2323 } kmp_base_depnode_t; 2324 2325 union KMP_ALIGN_CACHE kmp_depnode { 2326 double dn_align; /* use worst case alignment */ 2327 char dn_pad[KMP_PAD(kmp_base_depnode_t, CACHE_LINE)]; 2328 kmp_base_depnode_t dn; 2329 }; 2330 2331 struct kmp_dephash_entry { 2332 kmp_intptr_t addr; 2333 kmp_depnode_t *last_out; 2334 kmp_depnode_list_t *last_set; 2335 kmp_depnode_list_t *prev_set; 2336 kmp_uint8 last_flag; 2337 kmp_lock_t *mtx_lock; /* is referenced by depnodes w/mutexinoutset dep */ 2338 kmp_dephash_entry_t *next_in_bucket; 2339 }; 2340 2341 typedef struct kmp_dephash { 2342 kmp_dephash_entry_t **buckets; 2343 size_t size; 2344 kmp_depnode_t *last_all; 2345 size_t generation; 2346 kmp_uint32 nelements; 2347 kmp_uint32 nconflicts; 2348 } kmp_dephash_t; 2349 2350 typedef struct kmp_task_affinity_info { 2351 kmp_intptr_t base_addr; 2352 size_t len; 2353 struct { 2354 bool flag1 : 1; 2355 bool flag2 : 1; 2356 kmp_int32 reserved : 30; 2357 } flags; 2358 } kmp_task_affinity_info_t; 2359 2360 typedef enum kmp_event_type_t { 2361 KMP_EVENT_UNINITIALIZED = 0, 2362 KMP_EVENT_ALLOW_COMPLETION = 1 2363 } kmp_event_type_t; 2364 2365 typedef struct { 2366 kmp_event_type_t type; 2367 kmp_tas_lock_t lock; 2368 union { 2369 kmp_task_t *task; 2370 } ed; 2371 } kmp_event_t; 2372 2373 #ifdef BUILD_TIED_TASK_STACK 2374 2375 /* Tied Task stack definitions */ 2376 typedef struct kmp_stack_block { 2377 kmp_taskdata_t *sb_block[TASK_STACK_BLOCK_SIZE]; 2378 struct kmp_stack_block *sb_next; 2379 struct kmp_stack_block *sb_prev; 2380 } kmp_stack_block_t; 2381 2382 typedef struct kmp_task_stack { 2383 kmp_stack_block_t ts_first_block; // first block of stack entries 2384 kmp_taskdata_t **ts_top; // pointer to the top of stack 2385 kmp_int32 ts_entries; // number of entries on the stack 2386 } kmp_task_stack_t; 2387 2388 #endif // BUILD_TIED_TASK_STACK 2389 2390 typedef struct kmp_tasking_flags { /* Total struct must be exactly 32 bits */ 2391 /* Compiler flags */ /* Total compiler flags must be 16 bits */ 2392 unsigned tiedness : 1; /* task is either tied (1) or untied (0) */ 2393 unsigned final : 1; /* task is final(1) so execute immediately */ 2394 unsigned merged_if0 : 1; /* no __kmpc_task_{begin/complete}_if0 calls in if0 2395 code path */ 2396 unsigned destructors_thunk : 1; /* set if the compiler creates a thunk to 2397 invoke destructors from the runtime */ 2398 unsigned proxy : 1; /* task is a proxy task (it will be executed outside the 2399 context of the RTL) */ 2400 unsigned priority_specified : 1; /* set if the compiler provides priority 2401 setting for the task */ 2402 unsigned detachable : 1; /* 1 == can detach */ 2403 unsigned hidden_helper : 1; /* 1 == hidden helper task */ 2404 unsigned reserved : 8; /* reserved for compiler use */ 2405 2406 /* Library flags */ /* Total library flags must be 16 bits */ 2407 unsigned tasktype : 1; /* task is either explicit(1) or implicit (0) */ 2408 unsigned task_serial : 1; // task is executed immediately (1) or deferred (0) 2409 unsigned tasking_ser : 1; // all tasks in team are either executed immediately 2410 // (1) or may be deferred (0) 2411 unsigned team_serial : 1; // entire team is serial (1) [1 thread] or parallel 2412 // (0) [>= 2 threads] 2413 /* If either team_serial or tasking_ser is set, task team may be NULL */ 2414 /* Task State Flags: */ 2415 unsigned started : 1; /* 1==started, 0==not started */ 2416 unsigned executing : 1; /* 1==executing, 0==not executing */ 2417 unsigned complete : 1; /* 1==complete, 0==not complete */ 2418 unsigned freed : 1; /* 1==freed, 0==allocated */ 2419 unsigned native : 1; /* 1==gcc-compiled task, 0==intel */ 2420 unsigned reserved31 : 7; /* reserved for library use */ 2421 2422 } kmp_tasking_flags_t; 2423 2424 struct kmp_taskdata { /* aligned during dynamic allocation */ 2425 kmp_int32 td_task_id; /* id, assigned by debugger */ 2426 kmp_tasking_flags_t td_flags; /* task flags */ 2427 kmp_team_t *td_team; /* team for this task */ 2428 kmp_info_p *td_alloc_thread; /* thread that allocated data structures */ 2429 /* Currently not used except for perhaps IDB */ 2430 kmp_taskdata_t *td_parent; /* parent task */ 2431 kmp_int32 td_level; /* task nesting level */ 2432 std::atomic<kmp_int32> td_untied_count; // untied task active parts counter 2433 ident_t *td_ident; /* task identifier */ 2434 // Taskwait data. 2435 ident_t *td_taskwait_ident; 2436 kmp_uint32 td_taskwait_counter; 2437 kmp_int32 td_taskwait_thread; /* gtid + 1 of thread encountered taskwait */ 2438 KMP_ALIGN_CACHE kmp_internal_control_t 2439 td_icvs; /* Internal control variables for the task */ 2440 KMP_ALIGN_CACHE std::atomic<kmp_int32> 2441 td_allocated_child_tasks; /* Child tasks (+ current task) not yet 2442 deallocated */ 2443 std::atomic<kmp_int32> 2444 td_incomplete_child_tasks; /* Child tasks not yet complete */ 2445 kmp_taskgroup_t 2446 *td_taskgroup; // Each task keeps pointer to its current taskgroup 2447 kmp_dephash_t 2448 *td_dephash; // Dependencies for children tasks are tracked from here 2449 kmp_depnode_t 2450 *td_depnode; // Pointer to graph node if this task has dependencies 2451 kmp_task_team_t *td_task_team; 2452 size_t td_size_alloc; // Size of task structure, including shareds etc. 2453 #if defined(KMP_GOMP_COMPAT) 2454 // 4 or 8 byte integers for the loop bounds in GOMP_taskloop 2455 kmp_int32 td_size_loop_bounds; 2456 #endif 2457 kmp_taskdata_t *td_last_tied; // keep tied task for task scheduling constraint 2458 #if defined(KMP_GOMP_COMPAT) 2459 // GOMP sends in a copy function for copy constructors 2460 void (*td_copy_func)(void *, void *); 2461 #endif 2462 kmp_event_t td_allow_completion_event; 2463 #if OMPT_SUPPORT 2464 ompt_task_info_t ompt_task_info; 2465 #endif 2466 }; // struct kmp_taskdata 2467 2468 // Make sure padding above worked 2469 KMP_BUILD_ASSERT(sizeof(kmp_taskdata_t) % sizeof(void *) == 0); 2470 2471 // Data for task team but per thread 2472 typedef struct kmp_base_thread_data { 2473 kmp_info_p *td_thr; // Pointer back to thread info 2474 // Used only in __kmp_execute_tasks_template, maybe not avail until task is 2475 // queued? 2476 kmp_bootstrap_lock_t td_deque_lock; // Lock for accessing deque 2477 kmp_taskdata_t * 2478 *td_deque; // Deque of tasks encountered by td_thr, dynamically allocated 2479 kmp_int32 td_deque_size; // Size of deck 2480 kmp_uint32 td_deque_head; // Head of deque (will wrap) 2481 kmp_uint32 td_deque_tail; // Tail of deque (will wrap) 2482 kmp_int32 td_deque_ntasks; // Number of tasks in deque 2483 // GEH: shouldn't this be volatile since used in while-spin? 2484 kmp_int32 td_deque_last_stolen; // Thread number of last successful steal 2485 #ifdef BUILD_TIED_TASK_STACK 2486 kmp_task_stack_t td_susp_tied_tasks; // Stack of suspended tied tasks for task 2487 // scheduling constraint 2488 #endif // BUILD_TIED_TASK_STACK 2489 } kmp_base_thread_data_t; 2490 2491 #define TASK_DEQUE_BITS 8 // Used solely to define INITIAL_TASK_DEQUE_SIZE 2492 #define INITIAL_TASK_DEQUE_SIZE (1 << TASK_DEQUE_BITS) 2493 2494 #define TASK_DEQUE_SIZE(td) ((td).td_deque_size) 2495 #define TASK_DEQUE_MASK(td) ((td).td_deque_size - 1) 2496 2497 typedef union KMP_ALIGN_CACHE kmp_thread_data { 2498 kmp_base_thread_data_t td; 2499 double td_align; /* use worst case alignment */ 2500 char td_pad[KMP_PAD(kmp_base_thread_data_t, CACHE_LINE)]; 2501 } kmp_thread_data_t; 2502 2503 // Data for task teams which are used when tasking is enabled for the team 2504 typedef struct kmp_base_task_team { 2505 kmp_bootstrap_lock_t 2506 tt_threads_lock; /* Lock used to allocate per-thread part of task team */ 2507 /* must be bootstrap lock since used at library shutdown*/ 2508 kmp_task_team_t *tt_next; /* For linking the task team free list */ 2509 kmp_thread_data_t 2510 *tt_threads_data; /* Array of per-thread structures for task team */ 2511 /* Data survives task team deallocation */ 2512 kmp_int32 tt_found_tasks; /* Have we found tasks and queued them while 2513 executing this team? */ 2514 /* TRUE means tt_threads_data is set up and initialized */ 2515 kmp_int32 tt_nproc; /* #threads in team */ 2516 kmp_int32 tt_max_threads; // # entries allocated for threads_data array 2517 kmp_int32 tt_found_proxy_tasks; // found proxy tasks since last barrier 2518 kmp_int32 tt_untied_task_encountered; 2519 // There is hidden helper thread encountered in this task team so that we must 2520 // wait when waiting on task team 2521 kmp_int32 tt_hidden_helper_task_encountered; 2522 2523 KMP_ALIGN_CACHE 2524 std::atomic<kmp_int32> tt_unfinished_threads; /* #threads still active */ 2525 2526 KMP_ALIGN_CACHE 2527 volatile kmp_uint32 2528 tt_active; /* is the team still actively executing tasks */ 2529 } kmp_base_task_team_t; 2530 2531 union KMP_ALIGN_CACHE kmp_task_team { 2532 kmp_base_task_team_t tt; 2533 double tt_align; /* use worst case alignment */ 2534 char tt_pad[KMP_PAD(kmp_base_task_team_t, CACHE_LINE)]; 2535 }; 2536 2537 #if (USE_FAST_MEMORY == 3) || (USE_FAST_MEMORY == 5) 2538 // Free lists keep same-size free memory slots for fast memory allocation 2539 // routines 2540 typedef struct kmp_free_list { 2541 void *th_free_list_self; // Self-allocated tasks free list 2542 void *th_free_list_sync; // Self-allocated tasks stolen/returned by other 2543 // threads 2544 void *th_free_list_other; // Non-self free list (to be returned to owner's 2545 // sync list) 2546 } kmp_free_list_t; 2547 #endif 2548 #if KMP_NESTED_HOT_TEAMS 2549 // Hot teams array keeps hot teams and their sizes for given thread. Hot teams 2550 // are not put in teams pool, and they don't put threads in threads pool. 2551 typedef struct kmp_hot_team_ptr { 2552 kmp_team_p *hot_team; // pointer to hot_team of given nesting level 2553 kmp_int32 hot_team_nth; // number of threads allocated for the hot_team 2554 } kmp_hot_team_ptr_t; 2555 #endif 2556 typedef struct kmp_teams_size { 2557 kmp_int32 nteams; // number of teams in a league 2558 kmp_int32 nth; // number of threads in each team of the league 2559 } kmp_teams_size_t; 2560 2561 // This struct stores a thread that acts as a "root" for a contention 2562 // group. Contention groups are rooted at kmp_root threads, but also at 2563 // each primary thread of each team created in the teams construct. 2564 // This struct therefore also stores a thread_limit associated with 2565 // that contention group, and a counter to track the number of threads 2566 // active in that contention group. Each thread has a list of these: CG 2567 // root threads have an entry in their list in which cg_root refers to 2568 // the thread itself, whereas other workers in the CG will have a 2569 // single entry where cg_root is same as the entry containing their CG 2570 // root. When a thread encounters a teams construct, it will add a new 2571 // entry to the front of its list, because it now roots a new CG. 2572 typedef struct kmp_cg_root { 2573 kmp_info_p *cg_root; // "root" thread for a contention group 2574 // The CG root's limit comes from OMP_THREAD_LIMIT for root threads, or 2575 // thread_limit clause for teams primary threads 2576 kmp_int32 cg_thread_limit; 2577 kmp_int32 cg_nthreads; // Count of active threads in CG rooted at cg_root 2578 struct kmp_cg_root *up; // pointer to higher level CG root in list 2579 } kmp_cg_root_t; 2580 2581 // OpenMP thread data structures 2582 2583 typedef struct KMP_ALIGN_CACHE kmp_base_info { 2584 /* Start with the readonly data which is cache aligned and padded. This is 2585 written before the thread starts working by the primary thread. Uber 2586 masters may update themselves later. Usage does not consider serialized 2587 regions. */ 2588 kmp_desc_t th_info; 2589 kmp_team_p *th_team; /* team we belong to */ 2590 kmp_root_p *th_root; /* pointer to root of task hierarchy */ 2591 kmp_info_p *th_next_pool; /* next available thread in the pool */ 2592 kmp_disp_t *th_dispatch; /* thread's dispatch data */ 2593 int th_in_pool; /* in thread pool (32 bits for TCR/TCW) */ 2594 2595 /* The following are cached from the team info structure */ 2596 /* TODO use these in more places as determined to be needed via profiling */ 2597 int th_team_nproc; /* number of threads in a team */ 2598 kmp_info_p *th_team_master; /* the team's primary thread */ 2599 int th_team_serialized; /* team is serialized */ 2600 microtask_t th_teams_microtask; /* save entry address for teams construct */ 2601 int th_teams_level; /* save initial level of teams construct */ 2602 /* it is 0 on device but may be any on host */ 2603 2604 /* The blocktime info is copied from the team struct to the thread struct */ 2605 /* at the start of a barrier, and the values stored in the team are used */ 2606 /* at points in the code where the team struct is no longer guaranteed */ 2607 /* to exist (from the POV of worker threads). */ 2608 #if KMP_USE_MONITOR 2609 int th_team_bt_intervals; 2610 int th_team_bt_set; 2611 #else 2612 kmp_uint64 th_team_bt_intervals; 2613 #endif 2614 2615 #if KMP_AFFINITY_SUPPORTED 2616 kmp_affin_mask_t *th_affin_mask; /* thread's current affinity mask */ 2617 #endif 2618 omp_allocator_handle_t th_def_allocator; /* default allocator */ 2619 /* The data set by the primary thread at reinit, then R/W by the worker */ 2620 KMP_ALIGN_CACHE int 2621 th_set_nproc; /* if > 0, then only use this request for the next fork */ 2622 #if KMP_NESTED_HOT_TEAMS 2623 kmp_hot_team_ptr_t *th_hot_teams; /* array of hot teams */ 2624 #endif 2625 kmp_proc_bind_t 2626 th_set_proc_bind; /* if != proc_bind_default, use request for next fork */ 2627 kmp_teams_size_t 2628 th_teams_size; /* number of teams/threads in teams construct */ 2629 #if KMP_AFFINITY_SUPPORTED 2630 int th_current_place; /* place currently bound to */ 2631 int th_new_place; /* place to bind to in par reg */ 2632 int th_first_place; /* first place in partition */ 2633 int th_last_place; /* last place in partition */ 2634 #endif 2635 int th_prev_level; /* previous level for affinity format */ 2636 int th_prev_num_threads; /* previous num_threads for affinity format */ 2637 #if USE_ITT_BUILD 2638 kmp_uint64 th_bar_arrive_time; /* arrival to barrier timestamp */ 2639 kmp_uint64 th_bar_min_time; /* minimum arrival time at the barrier */ 2640 kmp_uint64 th_frame_time; /* frame timestamp */ 2641 #endif /* USE_ITT_BUILD */ 2642 kmp_local_t th_local; 2643 struct private_common *th_pri_head; 2644 2645 /* Now the data only used by the worker (after initial allocation) */ 2646 /* TODO the first serial team should actually be stored in the info_t 2647 structure. this will help reduce initial allocation overhead */ 2648 KMP_ALIGN_CACHE kmp_team_p 2649 *th_serial_team; /*serialized team held in reserve*/ 2650 2651 #if OMPT_SUPPORT 2652 ompt_thread_info_t ompt_thread_info; 2653 #endif 2654 2655 /* The following are also read by the primary thread during reinit */ 2656 struct common_table *th_pri_common; 2657 2658 volatile kmp_uint32 th_spin_here; /* thread-local location for spinning */ 2659 /* while awaiting queuing lock acquire */ 2660 2661 volatile void *th_sleep_loc; // this points at a kmp_flag<T> 2662 flag_type th_sleep_loc_type; // enum type of flag stored in th_sleep_loc 2663 2664 ident_t *th_ident; 2665 unsigned th_x; // Random number generator data 2666 unsigned th_a; // Random number generator data 2667 2668 /* Tasking-related data for the thread */ 2669 kmp_task_team_t *th_task_team; // Task team struct 2670 kmp_taskdata_t *th_current_task; // Innermost Task being executed 2671 kmp_uint8 th_task_state; // alternating 0/1 for task team identification 2672 kmp_uint8 *th_task_state_memo_stack; // Stack holding memos of th_task_state 2673 // at nested levels 2674 kmp_uint32 th_task_state_top; // Top element of th_task_state_memo_stack 2675 kmp_uint32 th_task_state_stack_sz; // Size of th_task_state_memo_stack 2676 kmp_uint32 th_reap_state; // Non-zero indicates thread is not 2677 // tasking, thus safe to reap 2678 2679 /* More stuff for keeping track of active/sleeping threads (this part is 2680 written by the worker thread) */ 2681 kmp_uint8 th_active_in_pool; // included in count of #active threads in pool 2682 int th_active; // ! sleeping; 32 bits for TCR/TCW 2683 std::atomic<kmp_uint32> th_used_in_team; // Flag indicating use in team 2684 // 0 = not used in team; 1 = used in team; 2685 // 2 = transitioning to not used in team; 3 = transitioning to used in team 2686 struct cons_header *th_cons; // used for consistency check 2687 #if KMP_USE_HIER_SCHED 2688 // used for hierarchical scheduling 2689 kmp_hier_private_bdata_t *th_hier_bar_data; 2690 #endif 2691 2692 /* Add the syncronizing data which is cache aligned and padded. */ 2693 KMP_ALIGN_CACHE kmp_balign_t th_bar[bs_last_barrier]; 2694 2695 KMP_ALIGN_CACHE volatile kmp_int32 2696 th_next_waiting; /* gtid+1 of next thread on lock wait queue, 0 if none */ 2697 2698 #if (USE_FAST_MEMORY == 3) || (USE_FAST_MEMORY == 5) 2699 #define NUM_LISTS 4 2700 kmp_free_list_t th_free_lists[NUM_LISTS]; // Free lists for fast memory 2701 // allocation routines 2702 #endif 2703 2704 #if KMP_OS_WINDOWS 2705 kmp_win32_cond_t th_suspend_cv; 2706 kmp_win32_mutex_t th_suspend_mx; 2707 std::atomic<int> th_suspend_init; 2708 #endif 2709 #if KMP_OS_UNIX 2710 kmp_cond_align_t th_suspend_cv; 2711 kmp_mutex_align_t th_suspend_mx; 2712 std::atomic<int> th_suspend_init_count; 2713 #endif 2714 2715 #if USE_ITT_BUILD 2716 kmp_itt_mark_t th_itt_mark_single; 2717 // alignment ??? 2718 #endif /* USE_ITT_BUILD */ 2719 #if KMP_STATS_ENABLED 2720 kmp_stats_list *th_stats; 2721 #endif 2722 #if KMP_OS_UNIX 2723 std::atomic<bool> th_blocking; 2724 #endif 2725 kmp_cg_root_t *th_cg_roots; // list of cg_roots associated with this thread 2726 } kmp_base_info_t; 2727 2728 typedef union KMP_ALIGN_CACHE kmp_info { 2729 double th_align; /* use worst case alignment */ 2730 char th_pad[KMP_PAD(kmp_base_info_t, CACHE_LINE)]; 2731 kmp_base_info_t th; 2732 } kmp_info_t; 2733 2734 // OpenMP thread team data structures 2735 2736 typedef struct kmp_base_data { 2737 volatile kmp_uint32 t_value; 2738 } kmp_base_data_t; 2739 2740 typedef union KMP_ALIGN_CACHE kmp_sleep_team { 2741 double dt_align; /* use worst case alignment */ 2742 char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)]; 2743 kmp_base_data_t dt; 2744 } kmp_sleep_team_t; 2745 2746 typedef union KMP_ALIGN_CACHE kmp_ordered_team { 2747 double dt_align; /* use worst case alignment */ 2748 char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)]; 2749 kmp_base_data_t dt; 2750 } kmp_ordered_team_t; 2751 2752 typedef int (*launch_t)(int gtid); 2753 2754 /* Minimum number of ARGV entries to malloc if necessary */ 2755 #define KMP_MIN_MALLOC_ARGV_ENTRIES 100 2756 2757 // Set up how many argv pointers will fit in cache lines containing 2758 // t_inline_argv. Historically, we have supported at least 96 bytes. Using a 2759 // larger value for more space between the primary write/worker read section and 2760 // read/write by all section seems to buy more performance on EPCC PARALLEL. 2761 #if KMP_ARCH_X86 || KMP_ARCH_X86_64 2762 #define KMP_INLINE_ARGV_BYTES \ 2763 (4 * CACHE_LINE - \ 2764 ((3 * KMP_PTR_SKIP + 2 * sizeof(int) + 2 * sizeof(kmp_int8) + \ 2765 sizeof(kmp_int16) + sizeof(kmp_uint32)) % \ 2766 CACHE_LINE)) 2767 #else 2768 #define KMP_INLINE_ARGV_BYTES \ 2769 (2 * CACHE_LINE - ((3 * KMP_PTR_SKIP + 2 * sizeof(int)) % CACHE_LINE)) 2770 #endif 2771 #define KMP_INLINE_ARGV_ENTRIES (int)(KMP_INLINE_ARGV_BYTES / KMP_PTR_SKIP) 2772 2773 typedef struct KMP_ALIGN_CACHE kmp_base_team { 2774 // Synchronization Data 2775 // --------------------------------------------------------------------------- 2776 KMP_ALIGN_CACHE kmp_ordered_team_t t_ordered; 2777 kmp_balign_team_t t_bar[bs_last_barrier]; 2778 std::atomic<int> t_construct; // count of single directive encountered by team 2779 char pad[sizeof(kmp_lock_t)]; // padding to maintain performance on big iron 2780 2781 // [0] - parallel / [1] - worksharing task reduction data shared by taskgroups 2782 std::atomic<void *> t_tg_reduce_data[2]; // to support task modifier 2783 std::atomic<int> t_tg_fini_counter[2]; // sync end of task reductions 2784 2785 // Primary thread only 2786 // --------------------------------------------------------------------------- 2787 KMP_ALIGN_CACHE int t_master_tid; // tid of primary thread in parent team 2788 int t_master_this_cons; // "this_construct" single counter of primary thread 2789 // in parent team 2790 ident_t *t_ident; // if volatile, have to change too much other crud to 2791 // volatile too 2792 kmp_team_p *t_parent; // parent team 2793 kmp_team_p *t_next_pool; // next free team in the team pool 2794 kmp_disp_t *t_dispatch; // thread's dispatch data 2795 kmp_task_team_t *t_task_team[2]; // Task team struct; switch between 2 2796 kmp_proc_bind_t t_proc_bind; // bind type for par region 2797 #if USE_ITT_BUILD 2798 kmp_uint64 t_region_time; // region begin timestamp 2799 #endif /* USE_ITT_BUILD */ 2800 2801 // Primary thread write, workers read 2802 // -------------------------------------------------------------------------- 2803 KMP_ALIGN_CACHE void **t_argv; 2804 int t_argc; 2805 int t_nproc; // number of threads in team 2806 microtask_t t_pkfn; 2807 launch_t t_invoke; // procedure to launch the microtask 2808 2809 #if OMPT_SUPPORT 2810 ompt_team_info_t ompt_team_info; 2811 ompt_lw_taskteam_t *ompt_serialized_team_info; 2812 #endif 2813 2814 #if KMP_ARCH_X86 || KMP_ARCH_X86_64 2815 kmp_int8 t_fp_control_saved; 2816 kmp_int8 t_pad2b; 2817 kmp_int16 t_x87_fpu_control_word; // FP control regs 2818 kmp_uint32 t_mxcsr; 2819 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */ 2820 2821 void *t_inline_argv[KMP_INLINE_ARGV_ENTRIES]; 2822 2823 KMP_ALIGN_CACHE kmp_info_t **t_threads; 2824 kmp_taskdata_t 2825 *t_implicit_task_taskdata; // Taskdata for the thread's implicit task 2826 int t_level; // nested parallel level 2827 2828 KMP_ALIGN_CACHE int t_max_argc; 2829 int t_max_nproc; // max threads this team can handle (dynamically expandable) 2830 int t_serialized; // levels deep of serialized teams 2831 dispatch_shared_info_t *t_disp_buffer; // buffers for dispatch system 2832 int t_id; // team's id, assigned by debugger. 2833 int t_active_level; // nested active parallel level 2834 kmp_r_sched_t t_sched; // run-time schedule for the team 2835 #if KMP_AFFINITY_SUPPORTED 2836 int t_first_place; // first & last place in parent thread's partition. 2837 int t_last_place; // Restore these values to primary thread after par region. 2838 #endif // KMP_AFFINITY_SUPPORTED 2839 int t_display_affinity; 2840 int t_size_changed; // team size was changed?: 0: no, 1: yes, -1: changed via 2841 // omp_set_num_threads() call 2842 omp_allocator_handle_t t_def_allocator; /* default allocator */ 2843 2844 // Read/write by workers as well 2845 #if (KMP_ARCH_X86 || KMP_ARCH_X86_64) 2846 // Using CACHE_LINE=64 reduces memory footprint, but causes a big perf 2847 // regression of epcc 'parallel' and 'barrier' on fxe256lin01. This extra 2848 // padding serves to fix the performance of epcc 'parallel' and 'barrier' when 2849 // CACHE_LINE=64. TODO: investigate more and get rid if this padding. 2850 char dummy_padding[1024]; 2851 #endif 2852 // Internal control stack for additional nested teams. 2853 KMP_ALIGN_CACHE kmp_internal_control_t *t_control_stack_top; 2854 // for SERIALIZED teams nested 2 or more levels deep 2855 // typed flag to store request state of cancellation 2856 std::atomic<kmp_int32> t_cancel_request; 2857 int t_master_active; // save on fork, restore on join 2858 void *t_copypriv_data; // team specific pointer to copyprivate data array 2859 #if KMP_OS_WINDOWS 2860 std::atomic<kmp_uint32> t_copyin_counter; 2861 #endif 2862 #if USE_ITT_BUILD 2863 void *t_stack_id; // team specific stack stitching id (for ittnotify) 2864 #endif /* USE_ITT_BUILD */ 2865 distributedBarrier *b; // Distributed barrier data associated with team 2866 } kmp_base_team_t; 2867 2868 union KMP_ALIGN_CACHE kmp_team { 2869 kmp_base_team_t t; 2870 double t_align; /* use worst case alignment */ 2871 char t_pad[KMP_PAD(kmp_base_team_t, CACHE_LINE)]; 2872 }; 2873 2874 typedef union KMP_ALIGN_CACHE kmp_time_global { 2875 double dt_align; /* use worst case alignment */ 2876 char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)]; 2877 kmp_base_data_t dt; 2878 } kmp_time_global_t; 2879 2880 typedef struct kmp_base_global { 2881 /* cache-aligned */ 2882 kmp_time_global_t g_time; 2883 2884 /* non cache-aligned */ 2885 volatile int g_abort; 2886 volatile int g_done; 2887 2888 int g_dynamic; 2889 enum dynamic_mode g_dynamic_mode; 2890 } kmp_base_global_t; 2891 2892 typedef union KMP_ALIGN_CACHE kmp_global { 2893 kmp_base_global_t g; 2894 double g_align; /* use worst case alignment */ 2895 char g_pad[KMP_PAD(kmp_base_global_t, CACHE_LINE)]; 2896 } kmp_global_t; 2897 2898 typedef struct kmp_base_root { 2899 // TODO: GEH - combine r_active with r_in_parallel then r_active == 2900 // (r_in_parallel>= 0) 2901 // TODO: GEH - then replace r_active with t_active_levels if we can to reduce 2902 // the synch overhead or keeping r_active 2903 volatile int r_active; /* TRUE if some region in a nest has > 1 thread */ 2904 // keeps a count of active parallel regions per root 2905 std::atomic<int> r_in_parallel; 2906 // GEH: This is misnamed, should be r_active_levels 2907 kmp_team_t *r_root_team; 2908 kmp_team_t *r_hot_team; 2909 kmp_info_t *r_uber_thread; 2910 kmp_lock_t r_begin_lock; 2911 volatile int r_begin; 2912 int r_blocktime; /* blocktime for this root and descendants */ 2913 #if KMP_AFFINITY_SUPPORTED 2914 int r_affinity_assigned; 2915 #endif // KMP_AFFINITY_SUPPORTED 2916 } kmp_base_root_t; 2917 2918 typedef union KMP_ALIGN_CACHE kmp_root { 2919 kmp_base_root_t r; 2920 double r_align; /* use worst case alignment */ 2921 char r_pad[KMP_PAD(kmp_base_root_t, CACHE_LINE)]; 2922 } kmp_root_t; 2923 2924 struct fortran_inx_info { 2925 kmp_int32 data; 2926 }; 2927 2928 /* ------------------------------------------------------------------------ */ 2929 2930 extern int __kmp_settings; 2931 extern int __kmp_duplicate_library_ok; 2932 #if USE_ITT_BUILD 2933 extern int __kmp_forkjoin_frames; 2934 extern int __kmp_forkjoin_frames_mode; 2935 #endif 2936 extern PACKED_REDUCTION_METHOD_T __kmp_force_reduction_method; 2937 extern int __kmp_determ_red; 2938 2939 #ifdef KMP_DEBUG 2940 extern int kmp_a_debug; 2941 extern int kmp_b_debug; 2942 extern int kmp_c_debug; 2943 extern int kmp_d_debug; 2944 extern int kmp_e_debug; 2945 extern int kmp_f_debug; 2946 #endif /* KMP_DEBUG */ 2947 2948 /* For debug information logging using rotating buffer */ 2949 #define KMP_DEBUG_BUF_LINES_INIT 512 2950 #define KMP_DEBUG_BUF_LINES_MIN 1 2951 2952 #define KMP_DEBUG_BUF_CHARS_INIT 128 2953 #define KMP_DEBUG_BUF_CHARS_MIN 2 2954 2955 extern int 2956 __kmp_debug_buf; /* TRUE means use buffer, FALSE means print to stderr */ 2957 extern int __kmp_debug_buf_lines; /* How many lines of debug stored in buffer */ 2958 extern int 2959 __kmp_debug_buf_chars; /* How many characters allowed per line in buffer */ 2960 extern int __kmp_debug_buf_atomic; /* TRUE means use atomic update of buffer 2961 entry pointer */ 2962 2963 extern char *__kmp_debug_buffer; /* Debug buffer itself */ 2964 extern std::atomic<int> __kmp_debug_count; /* Counter for number of lines 2965 printed in buffer so far */ 2966 extern int __kmp_debug_buf_warn_chars; /* Keep track of char increase 2967 recommended in warnings */ 2968 /* end rotating debug buffer */ 2969 2970 #ifdef KMP_DEBUG 2971 extern int __kmp_par_range; /* +1 => only go par for constructs in range */ 2972 2973 #define KMP_PAR_RANGE_ROUTINE_LEN 1024 2974 extern char __kmp_par_range_routine[KMP_PAR_RANGE_ROUTINE_LEN]; 2975 #define KMP_PAR_RANGE_FILENAME_LEN 1024 2976 extern char __kmp_par_range_filename[KMP_PAR_RANGE_FILENAME_LEN]; 2977 extern int __kmp_par_range_lb; 2978 extern int __kmp_par_range_ub; 2979 #endif 2980 2981 /* For printing out dynamic storage map for threads and teams */ 2982 extern int 2983 __kmp_storage_map; /* True means print storage map for threads and teams */ 2984 extern int __kmp_storage_map_verbose; /* True means storage map includes 2985 placement info */ 2986 extern int __kmp_storage_map_verbose_specified; 2987 2988 #if KMP_ARCH_X86 || KMP_ARCH_X86_64 2989 extern kmp_cpuinfo_t __kmp_cpuinfo; 2990 static inline bool __kmp_is_hybrid_cpu() { return __kmp_cpuinfo.flags.hybrid; } 2991 #else 2992 static inline bool __kmp_is_hybrid_cpu() { return false; } 2993 #endif 2994 2995 extern volatile int __kmp_init_serial; 2996 extern volatile int __kmp_init_gtid; 2997 extern volatile int __kmp_init_common; 2998 extern volatile int __kmp_init_middle; 2999 extern volatile int __kmp_init_parallel; 3000 #if KMP_USE_MONITOR 3001 extern volatile int __kmp_init_monitor; 3002 #endif 3003 extern volatile int __kmp_init_user_locks; 3004 extern volatile int __kmp_init_hidden_helper_threads; 3005 extern int __kmp_init_counter; 3006 extern int __kmp_root_counter; 3007 extern int __kmp_version; 3008 3009 /* list of address of allocated caches for commons */ 3010 extern kmp_cached_addr_t *__kmp_threadpriv_cache_list; 3011 3012 /* Barrier algorithm types and options */ 3013 extern kmp_uint32 __kmp_barrier_gather_bb_dflt; 3014 extern kmp_uint32 __kmp_barrier_release_bb_dflt; 3015 extern kmp_bar_pat_e __kmp_barrier_gather_pat_dflt; 3016 extern kmp_bar_pat_e __kmp_barrier_release_pat_dflt; 3017 extern kmp_uint32 __kmp_barrier_gather_branch_bits[bs_last_barrier]; 3018 extern kmp_uint32 __kmp_barrier_release_branch_bits[bs_last_barrier]; 3019 extern kmp_bar_pat_e __kmp_barrier_gather_pattern[bs_last_barrier]; 3020 extern kmp_bar_pat_e __kmp_barrier_release_pattern[bs_last_barrier]; 3021 extern char const *__kmp_barrier_branch_bit_env_name[bs_last_barrier]; 3022 extern char const *__kmp_barrier_pattern_env_name[bs_last_barrier]; 3023 extern char const *__kmp_barrier_type_name[bs_last_barrier]; 3024 extern char const *__kmp_barrier_pattern_name[bp_last_bar]; 3025 3026 /* Global Locks */ 3027 extern kmp_bootstrap_lock_t __kmp_initz_lock; /* control initialization */ 3028 extern kmp_bootstrap_lock_t __kmp_forkjoin_lock; /* control fork/join access */ 3029 extern kmp_bootstrap_lock_t __kmp_task_team_lock; 3030 extern kmp_bootstrap_lock_t 3031 __kmp_exit_lock; /* exit() is not always thread-safe */ 3032 #if KMP_USE_MONITOR 3033 extern kmp_bootstrap_lock_t 3034 __kmp_monitor_lock; /* control monitor thread creation */ 3035 #endif 3036 extern kmp_bootstrap_lock_t 3037 __kmp_tp_cached_lock; /* used for the hack to allow threadprivate cache and 3038 __kmp_threads expansion to co-exist */ 3039 3040 extern kmp_lock_t __kmp_global_lock; /* control OS/global access */ 3041 extern kmp_queuing_lock_t __kmp_dispatch_lock; /* control dispatch access */ 3042 extern kmp_lock_t __kmp_debug_lock; /* control I/O access for KMP_DEBUG */ 3043 3044 extern enum library_type __kmp_library; 3045 3046 extern enum sched_type __kmp_sched; /* default runtime scheduling */ 3047 extern enum sched_type __kmp_static; /* default static scheduling method */ 3048 extern enum sched_type __kmp_guided; /* default guided scheduling method */ 3049 extern enum sched_type __kmp_auto; /* default auto scheduling method */ 3050 extern int __kmp_chunk; /* default runtime chunk size */ 3051 extern int __kmp_force_monotonic; /* whether monotonic scheduling forced */ 3052 3053 extern size_t __kmp_stksize; /* stack size per thread */ 3054 #if KMP_USE_MONITOR 3055 extern size_t __kmp_monitor_stksize; /* stack size for monitor thread */ 3056 #endif 3057 extern size_t __kmp_stkoffset; /* stack offset per thread */ 3058 extern int __kmp_stkpadding; /* Should we pad root thread(s) stack */ 3059 3060 extern size_t 3061 __kmp_malloc_pool_incr; /* incremental size of pool for kmp_malloc() */ 3062 extern int __kmp_env_stksize; /* was KMP_STACKSIZE specified? */ 3063 extern int __kmp_env_blocktime; /* was KMP_BLOCKTIME specified? */ 3064 extern int __kmp_env_checks; /* was KMP_CHECKS specified? */ 3065 extern int __kmp_env_consistency_check; // was KMP_CONSISTENCY_CHECK specified? 3066 extern int __kmp_generate_warnings; /* should we issue warnings? */ 3067 extern int __kmp_reserve_warn; /* have we issued reserve_threads warning? */ 3068 3069 #ifdef DEBUG_SUSPEND 3070 extern int __kmp_suspend_count; /* count inside __kmp_suspend_template() */ 3071 #endif 3072 3073 extern kmp_int32 __kmp_use_yield; 3074 extern kmp_int32 __kmp_use_yield_exp_set; 3075 extern kmp_uint32 __kmp_yield_init; 3076 extern kmp_uint32 __kmp_yield_next; 3077 3078 /* ------------------------------------------------------------------------- */ 3079 extern int __kmp_allThreadsSpecified; 3080 3081 extern size_t __kmp_align_alloc; 3082 /* following data protected by initialization routines */ 3083 extern int __kmp_xproc; /* number of processors in the system */ 3084 extern int __kmp_avail_proc; /* number of processors available to the process */ 3085 extern size_t __kmp_sys_min_stksize; /* system-defined minimum stack size */ 3086 extern int __kmp_sys_max_nth; /* system-imposed maximum number of threads */ 3087 // maximum total number of concurrently-existing threads on device 3088 extern int __kmp_max_nth; 3089 // maximum total number of concurrently-existing threads in a contention group 3090 extern int __kmp_cg_max_nth; 3091 extern int __kmp_teams_max_nth; // max threads used in a teams construct 3092 extern int __kmp_threads_capacity; /* capacity of the arrays __kmp_threads and 3093 __kmp_root */ 3094 extern int __kmp_dflt_team_nth; /* default number of threads in a parallel 3095 region a la OMP_NUM_THREADS */ 3096 extern int __kmp_dflt_team_nth_ub; /* upper bound on "" determined at serial 3097 initialization */ 3098 extern int __kmp_tp_capacity; /* capacity of __kmp_threads if threadprivate is 3099 used (fixed) */ 3100 extern int __kmp_tp_cached; /* whether threadprivate cache has been created 3101 (__kmpc_threadprivate_cached()) */ 3102 extern int __kmp_dflt_blocktime; /* number of milliseconds to wait before 3103 blocking (env setting) */ 3104 #if KMP_USE_MONITOR 3105 extern int 3106 __kmp_monitor_wakeups; /* number of times monitor wakes up per second */ 3107 extern int __kmp_bt_intervals; /* number of monitor timestamp intervals before 3108 blocking */ 3109 #endif 3110 #ifdef KMP_ADJUST_BLOCKTIME 3111 extern int __kmp_zero_bt; /* whether blocktime has been forced to zero */ 3112 #endif /* KMP_ADJUST_BLOCKTIME */ 3113 #ifdef KMP_DFLT_NTH_CORES 3114 extern int __kmp_ncores; /* Total number of cores for threads placement */ 3115 #endif 3116 /* Number of millisecs to delay on abort for Intel(R) VTune(TM) tools */ 3117 extern int __kmp_abort_delay; 3118 3119 extern int __kmp_need_register_atfork_specified; 3120 extern int __kmp_need_register_atfork; /* At initialization, call pthread_atfork 3121 to install fork handler */ 3122 extern int __kmp_gtid_mode; /* Method of getting gtid, values: 3123 0 - not set, will be set at runtime 3124 1 - using stack search 3125 2 - dynamic TLS (pthread_getspecific(Linux* OS/OS 3126 X*) or TlsGetValue(Windows* OS)) 3127 3 - static TLS (__declspec(thread) __kmp_gtid), 3128 Linux* OS .so only. */ 3129 extern int 3130 __kmp_adjust_gtid_mode; /* If true, adjust method based on #threads */ 3131 #ifdef KMP_TDATA_GTID 3132 extern KMP_THREAD_LOCAL int __kmp_gtid; 3133 #endif 3134 extern int __kmp_tls_gtid_min; /* #threads below which use sp search for gtid */ 3135 extern int __kmp_foreign_tp; // If true, separate TP var for each foreign thread 3136 #if KMP_ARCH_X86 || KMP_ARCH_X86_64 3137 extern int __kmp_inherit_fp_control; // copy fp creg(s) parent->workers at fork 3138 extern kmp_int16 __kmp_init_x87_fpu_control_word; // init thread's FP ctrl reg 3139 extern kmp_uint32 __kmp_init_mxcsr; /* init thread's mxscr */ 3140 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */ 3141 3142 // max_active_levels for nested parallelism enabled by default via 3143 // OMP_MAX_ACTIVE_LEVELS, OMP_NESTED, OMP_NUM_THREADS, and OMP_PROC_BIND 3144 extern int __kmp_dflt_max_active_levels; 3145 // Indicates whether value of __kmp_dflt_max_active_levels was already 3146 // explicitly set by OMP_MAX_ACTIVE_LEVELS or OMP_NESTED=false 3147 extern bool __kmp_dflt_max_active_levels_set; 3148 extern int __kmp_dispatch_num_buffers; /* max possible dynamic loops in 3149 concurrent execution per team */ 3150 #if KMP_NESTED_HOT_TEAMS 3151 extern int __kmp_hot_teams_mode; 3152 extern int __kmp_hot_teams_max_level; 3153 #endif 3154 3155 #if KMP_OS_LINUX 3156 extern enum clock_function_type __kmp_clock_function; 3157 extern int __kmp_clock_function_param; 3158 #endif /* KMP_OS_LINUX */ 3159 3160 #if KMP_MIC_SUPPORTED 3161 extern enum mic_type __kmp_mic_type; 3162 #endif 3163 3164 #ifdef USE_LOAD_BALANCE 3165 extern double __kmp_load_balance_interval; // load balance algorithm interval 3166 #endif /* USE_LOAD_BALANCE */ 3167 3168 // OpenMP 3.1 - Nested num threads array 3169 typedef struct kmp_nested_nthreads_t { 3170 int *nth; 3171 int size; 3172 int used; 3173 } kmp_nested_nthreads_t; 3174 3175 extern kmp_nested_nthreads_t __kmp_nested_nth; 3176 3177 #if KMP_USE_ADAPTIVE_LOCKS 3178 3179 // Parameters for the speculative lock backoff system. 3180 struct kmp_adaptive_backoff_params_t { 3181 // Number of soft retries before it counts as a hard retry. 3182 kmp_uint32 max_soft_retries; 3183 // Badness is a bit mask : 0,1,3,7,15,... on each hard failure we move one to 3184 // the right 3185 kmp_uint32 max_badness; 3186 }; 3187 3188 extern kmp_adaptive_backoff_params_t __kmp_adaptive_backoff_params; 3189 3190 #if KMP_DEBUG_ADAPTIVE_LOCKS 3191 extern const char *__kmp_speculative_statsfile; 3192 #endif 3193 3194 #endif // KMP_USE_ADAPTIVE_LOCKS 3195 3196 extern int __kmp_display_env; /* TRUE or FALSE */ 3197 extern int __kmp_display_env_verbose; /* TRUE if OMP_DISPLAY_ENV=VERBOSE */ 3198 extern int __kmp_omp_cancellation; /* TRUE or FALSE */ 3199 extern int __kmp_nteams; 3200 extern int __kmp_teams_thread_limit; 3201 3202 /* ------------------------------------------------------------------------- */ 3203 3204 /* the following are protected by the fork/join lock */ 3205 /* write: lock read: anytime */ 3206 extern kmp_info_t **__kmp_threads; /* Descriptors for the threads */ 3207 /* read/write: lock */ 3208 extern volatile kmp_team_t *__kmp_team_pool; 3209 extern volatile kmp_info_t *__kmp_thread_pool; 3210 extern kmp_info_t *__kmp_thread_pool_insert_pt; 3211 3212 // total num threads reachable from some root thread including all root threads 3213 extern volatile int __kmp_nth; 3214 /* total number of threads reachable from some root thread including all root 3215 threads, and those in the thread pool */ 3216 extern volatile int __kmp_all_nth; 3217 extern std::atomic<int> __kmp_thread_pool_active_nth; 3218 3219 extern kmp_root_t **__kmp_root; /* root of thread hierarchy */ 3220 /* end data protected by fork/join lock */ 3221 /* ------------------------------------------------------------------------- */ 3222 3223 #define __kmp_get_gtid() __kmp_get_global_thread_id() 3224 #define __kmp_entry_gtid() __kmp_get_global_thread_id_reg() 3225 #define __kmp_get_tid() (__kmp_tid_from_gtid(__kmp_get_gtid())) 3226 #define __kmp_get_team() (__kmp_threads[(__kmp_get_gtid())]->th.th_team) 3227 #define __kmp_get_thread() (__kmp_thread_from_gtid(__kmp_get_gtid())) 3228 3229 // AT: Which way is correct? 3230 // AT: 1. nproc = __kmp_threads[ ( gtid ) ] -> th.th_team -> t.t_nproc; 3231 // AT: 2. nproc = __kmp_threads[ ( gtid ) ] -> th.th_team_nproc; 3232 #define __kmp_get_team_num_threads(gtid) \ 3233 (__kmp_threads[(gtid)]->th.th_team->t.t_nproc) 3234 3235 static inline bool KMP_UBER_GTID(int gtid) { 3236 KMP_DEBUG_ASSERT(gtid >= KMP_GTID_MIN); 3237 KMP_DEBUG_ASSERT(gtid < __kmp_threads_capacity); 3238 return (gtid >= 0 && __kmp_root[gtid] && __kmp_threads[gtid] && 3239 __kmp_threads[gtid] == __kmp_root[gtid]->r.r_uber_thread); 3240 } 3241 3242 static inline int __kmp_tid_from_gtid(int gtid) { 3243 KMP_DEBUG_ASSERT(gtid >= 0); 3244 return __kmp_threads[gtid]->th.th_info.ds.ds_tid; 3245 } 3246 3247 static inline int __kmp_gtid_from_tid(int tid, const kmp_team_t *team) { 3248 KMP_DEBUG_ASSERT(tid >= 0 && team); 3249 return team->t.t_threads[tid]->th.th_info.ds.ds_gtid; 3250 } 3251 3252 static inline int __kmp_gtid_from_thread(const kmp_info_t *thr) { 3253 KMP_DEBUG_ASSERT(thr); 3254 return thr->th.th_info.ds.ds_gtid; 3255 } 3256 3257 static inline kmp_info_t *__kmp_thread_from_gtid(int gtid) { 3258 KMP_DEBUG_ASSERT(gtid >= 0); 3259 return __kmp_threads[gtid]; 3260 } 3261 3262 static inline kmp_team_t *__kmp_team_from_gtid(int gtid) { 3263 KMP_DEBUG_ASSERT(gtid >= 0); 3264 return __kmp_threads[gtid]->th.th_team; 3265 } 3266 3267 static inline void __kmp_assert_valid_gtid(kmp_int32 gtid) { 3268 if (UNLIKELY(gtid < 0 || gtid >= __kmp_threads_capacity)) 3269 KMP_FATAL(ThreadIdentInvalid); 3270 } 3271 3272 #if KMP_HAVE_MWAIT || KMP_HAVE_UMWAIT 3273 extern int __kmp_user_level_mwait; // TRUE or FALSE; from KMP_USER_LEVEL_MWAIT 3274 extern int __kmp_umwait_enabled; // Runtime check if user-level mwait enabled 3275 extern int __kmp_mwait_enabled; // Runtime check if ring3 mwait is enabled 3276 extern int __kmp_mwait_hints; // Hints to pass in to mwait 3277 #endif 3278 3279 /* ------------------------------------------------------------------------- */ 3280 3281 extern kmp_global_t __kmp_global; /* global status */ 3282 3283 extern kmp_info_t __kmp_monitor; 3284 // For Debugging Support Library 3285 extern std::atomic<kmp_int32> __kmp_team_counter; 3286 // For Debugging Support Library 3287 extern std::atomic<kmp_int32> __kmp_task_counter; 3288 3289 #if USE_DEBUGGER 3290 #define _KMP_GEN_ID(counter) \ 3291 (__kmp_debugging ? KMP_ATOMIC_INC(&counter) + 1 : ~0) 3292 #else 3293 #define _KMP_GEN_ID(counter) (~0) 3294 #endif /* USE_DEBUGGER */ 3295 3296 #define KMP_GEN_TASK_ID() _KMP_GEN_ID(__kmp_task_counter) 3297 #define KMP_GEN_TEAM_ID() _KMP_GEN_ID(__kmp_team_counter) 3298 3299 /* ------------------------------------------------------------------------ */ 3300 3301 extern void __kmp_print_storage_map_gtid(int gtid, void *p1, void *p2, 3302 size_t size, char const *format, ...); 3303 3304 extern void __kmp_serial_initialize(void); 3305 extern void __kmp_middle_initialize(void); 3306 extern void __kmp_parallel_initialize(void); 3307 3308 extern void __kmp_internal_begin(void); 3309 extern void __kmp_internal_end_library(int gtid); 3310 extern void __kmp_internal_end_thread(int gtid); 3311 extern void __kmp_internal_end_atexit(void); 3312 extern void __kmp_internal_end_dtor(void); 3313 extern void __kmp_internal_end_dest(void *); 3314 3315 extern int __kmp_register_root(int initial_thread); 3316 extern void __kmp_unregister_root(int gtid); 3317 extern void __kmp_unregister_library(void); // called by __kmp_internal_end() 3318 3319 extern int __kmp_ignore_mppbeg(void); 3320 extern int __kmp_ignore_mppend(void); 3321 3322 extern int __kmp_enter_single(int gtid, ident_t *id_ref, int push_ws); 3323 extern void __kmp_exit_single(int gtid); 3324 3325 extern void __kmp_parallel_deo(int *gtid_ref, int *cid_ref, ident_t *loc_ref); 3326 extern void __kmp_parallel_dxo(int *gtid_ref, int *cid_ref, ident_t *loc_ref); 3327 3328 #ifdef USE_LOAD_BALANCE 3329 extern int __kmp_get_load_balance(int); 3330 #endif 3331 3332 extern int __kmp_get_global_thread_id(void); 3333 extern int __kmp_get_global_thread_id_reg(void); 3334 extern void __kmp_exit_thread(int exit_status); 3335 extern void __kmp_abort(char const *format, ...); 3336 extern void __kmp_abort_thread(void); 3337 KMP_NORETURN extern void __kmp_abort_process(void); 3338 extern void __kmp_warn(char const *format, ...); 3339 3340 extern void __kmp_set_num_threads(int new_nth, int gtid); 3341 3342 // Returns current thread (pointer to kmp_info_t). Current thread *must* be 3343 // registered. 3344 static inline kmp_info_t *__kmp_entry_thread() { 3345 int gtid = __kmp_entry_gtid(); 3346 3347 return __kmp_threads[gtid]; 3348 } 3349 3350 extern void __kmp_set_max_active_levels(int gtid, int new_max_active_levels); 3351 extern int __kmp_get_max_active_levels(int gtid); 3352 extern int __kmp_get_ancestor_thread_num(int gtid, int level); 3353 extern int __kmp_get_team_size(int gtid, int level); 3354 extern void __kmp_set_schedule(int gtid, kmp_sched_t new_sched, int chunk); 3355 extern void __kmp_get_schedule(int gtid, kmp_sched_t *sched, int *chunk); 3356 3357 extern unsigned short __kmp_get_random(kmp_info_t *thread); 3358 extern void __kmp_init_random(kmp_info_t *thread); 3359 3360 extern kmp_r_sched_t __kmp_get_schedule_global(void); 3361 extern void __kmp_adjust_num_threads(int new_nproc); 3362 extern void __kmp_check_stksize(size_t *val); 3363 3364 extern void *___kmp_allocate(size_t size KMP_SRC_LOC_DECL); 3365 extern void *___kmp_page_allocate(size_t size KMP_SRC_LOC_DECL); 3366 extern void ___kmp_free(void *ptr KMP_SRC_LOC_DECL); 3367 #define __kmp_allocate(size) ___kmp_allocate((size)KMP_SRC_LOC_CURR) 3368 #define __kmp_page_allocate(size) ___kmp_page_allocate((size)KMP_SRC_LOC_CURR) 3369 #define __kmp_free(ptr) ___kmp_free((ptr)KMP_SRC_LOC_CURR) 3370 3371 #if USE_FAST_MEMORY 3372 extern void *___kmp_fast_allocate(kmp_info_t *this_thr, 3373 size_t size KMP_SRC_LOC_DECL); 3374 extern void ___kmp_fast_free(kmp_info_t *this_thr, void *ptr KMP_SRC_LOC_DECL); 3375 extern void __kmp_free_fast_memory(kmp_info_t *this_thr); 3376 extern void __kmp_initialize_fast_memory(kmp_info_t *this_thr); 3377 #define __kmp_fast_allocate(this_thr, size) \ 3378 ___kmp_fast_allocate((this_thr), (size)KMP_SRC_LOC_CURR) 3379 #define __kmp_fast_free(this_thr, ptr) \ 3380 ___kmp_fast_free((this_thr), (ptr)KMP_SRC_LOC_CURR) 3381 #endif 3382 3383 extern void *___kmp_thread_malloc(kmp_info_t *th, size_t size KMP_SRC_LOC_DECL); 3384 extern void *___kmp_thread_calloc(kmp_info_t *th, size_t nelem, 3385 size_t elsize KMP_SRC_LOC_DECL); 3386 extern void *___kmp_thread_realloc(kmp_info_t *th, void *ptr, 3387 size_t size KMP_SRC_LOC_DECL); 3388 extern void ___kmp_thread_free(kmp_info_t *th, void *ptr KMP_SRC_LOC_DECL); 3389 #define __kmp_thread_malloc(th, size) \ 3390 ___kmp_thread_malloc((th), (size)KMP_SRC_LOC_CURR) 3391 #define __kmp_thread_calloc(th, nelem, elsize) \ 3392 ___kmp_thread_calloc((th), (nelem), (elsize)KMP_SRC_LOC_CURR) 3393 #define __kmp_thread_realloc(th, ptr, size) \ 3394 ___kmp_thread_realloc((th), (ptr), (size)KMP_SRC_LOC_CURR) 3395 #define __kmp_thread_free(th, ptr) \ 3396 ___kmp_thread_free((th), (ptr)KMP_SRC_LOC_CURR) 3397 3398 #define KMP_INTERNAL_MALLOC(sz) malloc(sz) 3399 #define KMP_INTERNAL_FREE(p) free(p) 3400 #define KMP_INTERNAL_REALLOC(p, sz) realloc((p), (sz)) 3401 #define KMP_INTERNAL_CALLOC(n, sz) calloc((n), (sz)) 3402 3403 extern void __kmp_push_num_threads(ident_t *loc, int gtid, int num_threads); 3404 3405 extern void __kmp_push_proc_bind(ident_t *loc, int gtid, 3406 kmp_proc_bind_t proc_bind); 3407 extern void __kmp_push_num_teams(ident_t *loc, int gtid, int num_teams, 3408 int num_threads); 3409 extern void __kmp_push_num_teams_51(ident_t *loc, int gtid, int num_teams_lb, 3410 int num_teams_ub, int num_threads); 3411 3412 extern void __kmp_yield(); 3413 3414 extern void __kmpc_dispatch_init_4(ident_t *loc, kmp_int32 gtid, 3415 enum sched_type schedule, kmp_int32 lb, 3416 kmp_int32 ub, kmp_int32 st, kmp_int32 chunk); 3417 extern void __kmpc_dispatch_init_4u(ident_t *loc, kmp_int32 gtid, 3418 enum sched_type schedule, kmp_uint32 lb, 3419 kmp_uint32 ub, kmp_int32 st, 3420 kmp_int32 chunk); 3421 extern void __kmpc_dispatch_init_8(ident_t *loc, kmp_int32 gtid, 3422 enum sched_type schedule, kmp_int64 lb, 3423 kmp_int64 ub, kmp_int64 st, kmp_int64 chunk); 3424 extern void __kmpc_dispatch_init_8u(ident_t *loc, kmp_int32 gtid, 3425 enum sched_type schedule, kmp_uint64 lb, 3426 kmp_uint64 ub, kmp_int64 st, 3427 kmp_int64 chunk); 3428 3429 extern int __kmpc_dispatch_next_4(ident_t *loc, kmp_int32 gtid, 3430 kmp_int32 *p_last, kmp_int32 *p_lb, 3431 kmp_int32 *p_ub, kmp_int32 *p_st); 3432 extern int __kmpc_dispatch_next_4u(ident_t *loc, kmp_int32 gtid, 3433 kmp_int32 *p_last, kmp_uint32 *p_lb, 3434 kmp_uint32 *p_ub, kmp_int32 *p_st); 3435 extern int __kmpc_dispatch_next_8(ident_t *loc, kmp_int32 gtid, 3436 kmp_int32 *p_last, kmp_int64 *p_lb, 3437 kmp_int64 *p_ub, kmp_int64 *p_st); 3438 extern int __kmpc_dispatch_next_8u(ident_t *loc, kmp_int32 gtid, 3439 kmp_int32 *p_last, kmp_uint64 *p_lb, 3440 kmp_uint64 *p_ub, kmp_int64 *p_st); 3441 3442 extern void __kmpc_dispatch_fini_4(ident_t *loc, kmp_int32 gtid); 3443 extern void __kmpc_dispatch_fini_8(ident_t *loc, kmp_int32 gtid); 3444 extern void __kmpc_dispatch_fini_4u(ident_t *loc, kmp_int32 gtid); 3445 extern void __kmpc_dispatch_fini_8u(ident_t *loc, kmp_int32 gtid); 3446 3447 #ifdef KMP_GOMP_COMPAT 3448 3449 extern void __kmp_aux_dispatch_init_4(ident_t *loc, kmp_int32 gtid, 3450 enum sched_type schedule, kmp_int32 lb, 3451 kmp_int32 ub, kmp_int32 st, 3452 kmp_int32 chunk, int push_ws); 3453 extern void __kmp_aux_dispatch_init_4u(ident_t *loc, kmp_int32 gtid, 3454 enum sched_type schedule, kmp_uint32 lb, 3455 kmp_uint32 ub, kmp_int32 st, 3456 kmp_int32 chunk, int push_ws); 3457 extern void __kmp_aux_dispatch_init_8(ident_t *loc, kmp_int32 gtid, 3458 enum sched_type schedule, kmp_int64 lb, 3459 kmp_int64 ub, kmp_int64 st, 3460 kmp_int64 chunk, int push_ws); 3461 extern void __kmp_aux_dispatch_init_8u(ident_t *loc, kmp_int32 gtid, 3462 enum sched_type schedule, kmp_uint64 lb, 3463 kmp_uint64 ub, kmp_int64 st, 3464 kmp_int64 chunk, int push_ws); 3465 extern void __kmp_aux_dispatch_fini_chunk_4(ident_t *loc, kmp_int32 gtid); 3466 extern void __kmp_aux_dispatch_fini_chunk_8(ident_t *loc, kmp_int32 gtid); 3467 extern void __kmp_aux_dispatch_fini_chunk_4u(ident_t *loc, kmp_int32 gtid); 3468 extern void __kmp_aux_dispatch_fini_chunk_8u(ident_t *loc, kmp_int32 gtid); 3469 3470 #endif /* KMP_GOMP_COMPAT */ 3471 3472 extern kmp_uint32 __kmp_eq_4(kmp_uint32 value, kmp_uint32 checker); 3473 extern kmp_uint32 __kmp_neq_4(kmp_uint32 value, kmp_uint32 checker); 3474 extern kmp_uint32 __kmp_lt_4(kmp_uint32 value, kmp_uint32 checker); 3475 extern kmp_uint32 __kmp_ge_4(kmp_uint32 value, kmp_uint32 checker); 3476 extern kmp_uint32 __kmp_le_4(kmp_uint32 value, kmp_uint32 checker); 3477 extern kmp_uint32 __kmp_wait_4(kmp_uint32 volatile *spinner, kmp_uint32 checker, 3478 kmp_uint32 (*pred)(kmp_uint32, kmp_uint32), 3479 void *obj); 3480 extern void __kmp_wait_4_ptr(void *spinner, kmp_uint32 checker, 3481 kmp_uint32 (*pred)(void *, kmp_uint32), void *obj); 3482 3483 extern void __kmp_wait_64(kmp_info_t *this_thr, kmp_flag_64<> *flag, 3484 int final_spin 3485 #if USE_ITT_BUILD 3486 , 3487 void *itt_sync_obj 3488 #endif 3489 ); 3490 extern void __kmp_release_64(kmp_flag_64<> *flag); 3491 3492 extern void __kmp_infinite_loop(void); 3493 3494 extern void __kmp_cleanup(void); 3495 3496 #if KMP_HANDLE_SIGNALS 3497 extern int __kmp_handle_signals; 3498 extern void __kmp_install_signals(int parallel_init); 3499 extern void __kmp_remove_signals(void); 3500 #endif 3501 3502 extern void __kmp_clear_system_time(void); 3503 extern void __kmp_read_system_time(double *delta); 3504 3505 extern void __kmp_check_stack_overlap(kmp_info_t *thr); 3506 3507 extern void __kmp_expand_host_name(char *buffer, size_t size); 3508 extern void __kmp_expand_file_name(char *result, size_t rlen, char *pattern); 3509 3510 #if KMP_ARCH_X86 || KMP_ARCH_X86_64 || (KMP_OS_WINDOWS && KMP_ARCH_AARCH64) 3511 extern void 3512 __kmp_initialize_system_tick(void); /* Initialize timer tick value */ 3513 #endif 3514 3515 extern void 3516 __kmp_runtime_initialize(void); /* machine specific initialization */ 3517 extern void __kmp_runtime_destroy(void); 3518 3519 #if KMP_AFFINITY_SUPPORTED 3520 extern char *__kmp_affinity_print_mask(char *buf, int buf_len, 3521 kmp_affin_mask_t *mask); 3522 extern kmp_str_buf_t *__kmp_affinity_str_buf_mask(kmp_str_buf_t *buf, 3523 kmp_affin_mask_t *mask); 3524 extern void __kmp_affinity_initialize(void); 3525 extern void __kmp_affinity_uninitialize(void); 3526 extern void __kmp_affinity_set_init_mask( 3527 int gtid, int isa_root); /* set affinity according to KMP_AFFINITY */ 3528 extern void __kmp_affinity_set_place(int gtid); 3529 extern void __kmp_affinity_determine_capable(const char *env_var); 3530 extern int __kmp_aux_set_affinity(void **mask); 3531 extern int __kmp_aux_get_affinity(void **mask); 3532 extern int __kmp_aux_get_affinity_max_proc(); 3533 extern int __kmp_aux_set_affinity_mask_proc(int proc, void **mask); 3534 extern int __kmp_aux_unset_affinity_mask_proc(int proc, void **mask); 3535 extern int __kmp_aux_get_affinity_mask_proc(int proc, void **mask); 3536 extern void __kmp_balanced_affinity(kmp_info_t *th, int team_size); 3537 #if KMP_OS_LINUX || KMP_OS_FREEBSD 3538 extern int kmp_set_thread_affinity_mask_initial(void); 3539 #endif 3540 static inline void __kmp_assign_root_init_mask() { 3541 int gtid = __kmp_entry_gtid(); 3542 kmp_root_t *r = __kmp_threads[gtid]->th.th_root; 3543 if (r->r.r_uber_thread == __kmp_threads[gtid] && !r->r.r_affinity_assigned) { 3544 __kmp_affinity_set_init_mask(gtid, TRUE); 3545 r->r.r_affinity_assigned = TRUE; 3546 } 3547 } 3548 #else /* KMP_AFFINITY_SUPPORTED */ 3549 #define __kmp_assign_root_init_mask() /* Nothing */ 3550 #endif /* KMP_AFFINITY_SUPPORTED */ 3551 // No need for KMP_AFFINITY_SUPPORTED guard as only one field in the 3552 // format string is for affinity, so platforms that do not support 3553 // affinity can still use the other fields, e.g., %n for num_threads 3554 extern size_t __kmp_aux_capture_affinity(int gtid, const char *format, 3555 kmp_str_buf_t *buffer); 3556 extern void __kmp_aux_display_affinity(int gtid, const char *format); 3557 3558 extern void __kmp_cleanup_hierarchy(); 3559 extern void __kmp_get_hierarchy(kmp_uint32 nproc, kmp_bstate_t *thr_bar); 3560 3561 #if KMP_USE_FUTEX 3562 3563 extern int __kmp_futex_determine_capable(void); 3564 3565 #endif // KMP_USE_FUTEX 3566 3567 extern void __kmp_gtid_set_specific(int gtid); 3568 extern int __kmp_gtid_get_specific(void); 3569 3570 extern double __kmp_read_cpu_time(void); 3571 3572 extern int __kmp_read_system_info(struct kmp_sys_info *info); 3573 3574 #if KMP_USE_MONITOR 3575 extern void __kmp_create_monitor(kmp_info_t *th); 3576 #endif 3577 3578 extern void *__kmp_launch_thread(kmp_info_t *thr); 3579 3580 extern void __kmp_create_worker(int gtid, kmp_info_t *th, size_t stack_size); 3581 3582 #if KMP_OS_WINDOWS 3583 extern int __kmp_still_running(kmp_info_t *th); 3584 extern int __kmp_is_thread_alive(kmp_info_t *th, DWORD *exit_val); 3585 extern void __kmp_free_handle(kmp_thread_t tHandle); 3586 #endif 3587 3588 #if KMP_USE_MONITOR 3589 extern void __kmp_reap_monitor(kmp_info_t *th); 3590 #endif 3591 extern void __kmp_reap_worker(kmp_info_t *th); 3592 extern void __kmp_terminate_thread(int gtid); 3593 3594 extern int __kmp_try_suspend_mx(kmp_info_t *th); 3595 extern void __kmp_lock_suspend_mx(kmp_info_t *th); 3596 extern void __kmp_unlock_suspend_mx(kmp_info_t *th); 3597 3598 extern void __kmp_elapsed(double *); 3599 extern void __kmp_elapsed_tick(double *); 3600 3601 extern void __kmp_enable(int old_state); 3602 extern void __kmp_disable(int *old_state); 3603 3604 extern void __kmp_thread_sleep(int millis); 3605 3606 extern void __kmp_common_initialize(void); 3607 extern void __kmp_common_destroy(void); 3608 extern void __kmp_common_destroy_gtid(int gtid); 3609 3610 #if KMP_OS_UNIX 3611 extern void __kmp_register_atfork(void); 3612 #endif 3613 extern void __kmp_suspend_initialize(void); 3614 extern void __kmp_suspend_initialize_thread(kmp_info_t *th); 3615 extern void __kmp_suspend_uninitialize_thread(kmp_info_t *th); 3616 3617 extern kmp_info_t *__kmp_allocate_thread(kmp_root_t *root, kmp_team_t *team, 3618 int tid); 3619 extern kmp_team_t * 3620 __kmp_allocate_team(kmp_root_t *root, int new_nproc, int max_nproc, 3621 #if OMPT_SUPPORT 3622 ompt_data_t ompt_parallel_data, 3623 #endif 3624 kmp_proc_bind_t proc_bind, kmp_internal_control_t *new_icvs, 3625 int argc USE_NESTED_HOT_ARG(kmp_info_t *thr)); 3626 extern void __kmp_free_thread(kmp_info_t *); 3627 extern void __kmp_free_team(kmp_root_t *, 3628 kmp_team_t *USE_NESTED_HOT_ARG(kmp_info_t *)); 3629 extern kmp_team_t *__kmp_reap_team(kmp_team_t *); 3630 3631 /* ------------------------------------------------------------------------ */ 3632 3633 extern void __kmp_initialize_bget(kmp_info_t *th); 3634 extern void __kmp_finalize_bget(kmp_info_t *th); 3635 3636 KMP_EXPORT void *kmpc_malloc(size_t size); 3637 KMP_EXPORT void *kmpc_aligned_malloc(size_t size, size_t alignment); 3638 KMP_EXPORT void *kmpc_calloc(size_t nelem, size_t elsize); 3639 KMP_EXPORT void *kmpc_realloc(void *ptr, size_t size); 3640 KMP_EXPORT void kmpc_free(void *ptr); 3641 3642 /* declarations for internal use */ 3643 3644 extern int __kmp_barrier(enum barrier_type bt, int gtid, int is_split, 3645 size_t reduce_size, void *reduce_data, 3646 void (*reduce)(void *, void *)); 3647 extern void __kmp_end_split_barrier(enum barrier_type bt, int gtid); 3648 extern int __kmp_barrier_gomp_cancel(int gtid); 3649 3650 /*! 3651 * Tell the fork call which compiler generated the fork call, and therefore how 3652 * to deal with the call. 3653 */ 3654 enum fork_context_e { 3655 fork_context_gnu, /**< Called from GNU generated code, so must not invoke the 3656 microtask internally. */ 3657 fork_context_intel, /**< Called from Intel generated code. */ 3658 fork_context_last 3659 }; 3660 extern int __kmp_fork_call(ident_t *loc, int gtid, 3661 enum fork_context_e fork_context, kmp_int32 argc, 3662 microtask_t microtask, launch_t invoker, 3663 kmp_va_list ap); 3664 3665 extern void __kmp_join_call(ident_t *loc, int gtid 3666 #if OMPT_SUPPORT 3667 , 3668 enum fork_context_e fork_context 3669 #endif 3670 , 3671 int exit_teams = 0); 3672 3673 extern void __kmp_serialized_parallel(ident_t *id, kmp_int32 gtid); 3674 extern void __kmp_internal_fork(ident_t *id, int gtid, kmp_team_t *team); 3675 extern void __kmp_internal_join(ident_t *id, int gtid, kmp_team_t *team); 3676 extern int __kmp_invoke_task_func(int gtid); 3677 extern void __kmp_run_before_invoked_task(int gtid, int tid, 3678 kmp_info_t *this_thr, 3679 kmp_team_t *team); 3680 extern void __kmp_run_after_invoked_task(int gtid, int tid, 3681 kmp_info_t *this_thr, 3682 kmp_team_t *team); 3683 3684 // should never have been exported 3685 KMP_EXPORT int __kmpc_invoke_task_func(int gtid); 3686 extern int __kmp_invoke_teams_master(int gtid); 3687 extern void __kmp_teams_master(int gtid); 3688 extern int __kmp_aux_get_team_num(); 3689 extern int __kmp_aux_get_num_teams(); 3690 extern void __kmp_save_internal_controls(kmp_info_t *thread); 3691 extern void __kmp_user_set_library(enum library_type arg); 3692 extern void __kmp_aux_set_library(enum library_type arg); 3693 extern void __kmp_aux_set_stacksize(size_t arg); 3694 extern void __kmp_aux_set_blocktime(int arg, kmp_info_t *thread, int tid); 3695 extern void __kmp_aux_set_defaults(char const *str, size_t len); 3696 3697 /* Functions called from __kmp_aux_env_initialize() in kmp_settings.cpp */ 3698 void kmpc_set_blocktime(int arg); 3699 void ompc_set_nested(int flag); 3700 void ompc_set_dynamic(int flag); 3701 void ompc_set_num_threads(int arg); 3702 3703 extern void __kmp_push_current_task_to_thread(kmp_info_t *this_thr, 3704 kmp_team_t *team, int tid); 3705 extern void __kmp_pop_current_task_from_thread(kmp_info_t *this_thr); 3706 extern kmp_task_t *__kmp_task_alloc(ident_t *loc_ref, kmp_int32 gtid, 3707 kmp_tasking_flags_t *flags, 3708 size_t sizeof_kmp_task_t, 3709 size_t sizeof_shareds, 3710 kmp_routine_entry_t task_entry); 3711 extern void __kmp_init_implicit_task(ident_t *loc_ref, kmp_info_t *this_thr, 3712 kmp_team_t *team, int tid, 3713 int set_curr_task); 3714 extern void __kmp_finish_implicit_task(kmp_info_t *this_thr); 3715 extern void __kmp_free_implicit_task(kmp_info_t *this_thr); 3716 3717 extern kmp_event_t *__kmpc_task_allow_completion_event(ident_t *loc_ref, 3718 int gtid, 3719 kmp_task_t *task); 3720 extern void __kmp_fulfill_event(kmp_event_t *event); 3721 3722 extern void __kmp_free_task_team(kmp_info_t *thread, 3723 kmp_task_team_t *task_team); 3724 extern void __kmp_reap_task_teams(void); 3725 extern void __kmp_wait_to_unref_task_teams(void); 3726 extern void __kmp_task_team_setup(kmp_info_t *this_thr, kmp_team_t *team, 3727 int always); 3728 extern void __kmp_task_team_sync(kmp_info_t *this_thr, kmp_team_t *team); 3729 extern void __kmp_task_team_wait(kmp_info_t *this_thr, kmp_team_t *team 3730 #if USE_ITT_BUILD 3731 , 3732 void *itt_sync_obj 3733 #endif /* USE_ITT_BUILD */ 3734 , 3735 int wait = 1); 3736 extern void __kmp_tasking_barrier(kmp_team_t *team, kmp_info_t *thread, 3737 int gtid); 3738 3739 extern int __kmp_is_address_mapped(void *addr); 3740 extern kmp_uint64 __kmp_hardware_timestamp(void); 3741 3742 #if KMP_OS_UNIX 3743 extern int __kmp_read_from_file(char const *path, char const *format, ...); 3744 #endif 3745 3746 /* ------------------------------------------------------------------------ */ 3747 // 3748 // Assembly routines that have no compiler intrinsic replacement 3749 // 3750 3751 extern int __kmp_invoke_microtask(microtask_t pkfn, int gtid, int npr, int argc, 3752 void *argv[] 3753 #if OMPT_SUPPORT 3754 , 3755 void **exit_frame_ptr 3756 #endif 3757 ); 3758 3759 /* ------------------------------------------------------------------------ */ 3760 3761 KMP_EXPORT void __kmpc_begin(ident_t *, kmp_int32 flags); 3762 KMP_EXPORT void __kmpc_end(ident_t *); 3763 3764 KMP_EXPORT void __kmpc_threadprivate_register_vec(ident_t *, void *data, 3765 kmpc_ctor_vec ctor, 3766 kmpc_cctor_vec cctor, 3767 kmpc_dtor_vec dtor, 3768 size_t vector_length); 3769 KMP_EXPORT void __kmpc_threadprivate_register(ident_t *, void *data, 3770 kmpc_ctor ctor, kmpc_cctor cctor, 3771 kmpc_dtor dtor); 3772 KMP_EXPORT void *__kmpc_threadprivate(ident_t *, kmp_int32 global_tid, 3773 void *data, size_t size); 3774 3775 KMP_EXPORT kmp_int32 __kmpc_global_thread_num(ident_t *); 3776 KMP_EXPORT kmp_int32 __kmpc_global_num_threads(ident_t *); 3777 KMP_EXPORT kmp_int32 __kmpc_bound_thread_num(ident_t *); 3778 KMP_EXPORT kmp_int32 __kmpc_bound_num_threads(ident_t *); 3779 3780 KMP_EXPORT kmp_int32 __kmpc_ok_to_fork(ident_t *); 3781 KMP_EXPORT void __kmpc_fork_call(ident_t *, kmp_int32 nargs, 3782 kmpc_micro microtask, ...); 3783 3784 KMP_EXPORT void __kmpc_serialized_parallel(ident_t *, kmp_int32 global_tid); 3785 KMP_EXPORT void __kmpc_end_serialized_parallel(ident_t *, kmp_int32 global_tid); 3786 3787 KMP_EXPORT void __kmpc_flush(ident_t *); 3788 KMP_EXPORT void __kmpc_barrier(ident_t *, kmp_int32 global_tid); 3789 KMP_EXPORT kmp_int32 __kmpc_master(ident_t *, kmp_int32 global_tid); 3790 KMP_EXPORT void __kmpc_end_master(ident_t *, kmp_int32 global_tid); 3791 KMP_EXPORT kmp_int32 __kmpc_masked(ident_t *, kmp_int32 global_tid, 3792 kmp_int32 filter); 3793 KMP_EXPORT void __kmpc_end_masked(ident_t *, kmp_int32 global_tid); 3794 KMP_EXPORT void __kmpc_ordered(ident_t *, kmp_int32 global_tid); 3795 KMP_EXPORT void __kmpc_end_ordered(ident_t *, kmp_int32 global_tid); 3796 KMP_EXPORT void __kmpc_critical(ident_t *, kmp_int32 global_tid, 3797 kmp_critical_name *); 3798 KMP_EXPORT void __kmpc_end_critical(ident_t *, kmp_int32 global_tid, 3799 kmp_critical_name *); 3800 KMP_EXPORT void __kmpc_critical_with_hint(ident_t *, kmp_int32 global_tid, 3801 kmp_critical_name *, uint32_t hint); 3802 3803 KMP_EXPORT kmp_int32 __kmpc_barrier_master(ident_t *, kmp_int32 global_tid); 3804 KMP_EXPORT void __kmpc_end_barrier_master(ident_t *, kmp_int32 global_tid); 3805 3806 KMP_EXPORT kmp_int32 __kmpc_barrier_master_nowait(ident_t *, 3807 kmp_int32 global_tid); 3808 3809 KMP_EXPORT kmp_int32 __kmpc_single(ident_t *, kmp_int32 global_tid); 3810 KMP_EXPORT void __kmpc_end_single(ident_t *, kmp_int32 global_tid); 3811 3812 KMP_EXPORT void KMPC_FOR_STATIC_INIT(ident_t *loc, kmp_int32 global_tid, 3813 kmp_int32 schedtype, kmp_int32 *plastiter, 3814 kmp_int *plower, kmp_int *pupper, 3815 kmp_int *pstride, kmp_int incr, 3816 kmp_int chunk); 3817 3818 KMP_EXPORT void __kmpc_for_static_fini(ident_t *loc, kmp_int32 global_tid); 3819 3820 KMP_EXPORT void __kmpc_copyprivate(ident_t *loc, kmp_int32 global_tid, 3821 size_t cpy_size, void *cpy_data, 3822 void (*cpy_func)(void *, void *), 3823 kmp_int32 didit); 3824 3825 extern void KMPC_SET_NUM_THREADS(int arg); 3826 extern void KMPC_SET_DYNAMIC(int flag); 3827 extern void KMPC_SET_NESTED(int flag); 3828 3829 /* OMP 3.0 tasking interface routines */ 3830 KMP_EXPORT kmp_int32 __kmpc_omp_task(ident_t *loc_ref, kmp_int32 gtid, 3831 kmp_task_t *new_task); 3832 KMP_EXPORT kmp_task_t *__kmpc_omp_task_alloc(ident_t *loc_ref, kmp_int32 gtid, 3833 kmp_int32 flags, 3834 size_t sizeof_kmp_task_t, 3835 size_t sizeof_shareds, 3836 kmp_routine_entry_t task_entry); 3837 KMP_EXPORT kmp_task_t *__kmpc_omp_target_task_alloc( 3838 ident_t *loc_ref, kmp_int32 gtid, kmp_int32 flags, size_t sizeof_kmp_task_t, 3839 size_t sizeof_shareds, kmp_routine_entry_t task_entry, kmp_int64 device_id); 3840 KMP_EXPORT void __kmpc_omp_task_begin_if0(ident_t *loc_ref, kmp_int32 gtid, 3841 kmp_task_t *task); 3842 KMP_EXPORT void __kmpc_omp_task_complete_if0(ident_t *loc_ref, kmp_int32 gtid, 3843 kmp_task_t *task); 3844 KMP_EXPORT kmp_int32 __kmpc_omp_task_parts(ident_t *loc_ref, kmp_int32 gtid, 3845 kmp_task_t *new_task); 3846 KMP_EXPORT kmp_int32 __kmpc_omp_taskwait(ident_t *loc_ref, kmp_int32 gtid); 3847 3848 KMP_EXPORT kmp_int32 __kmpc_omp_taskyield(ident_t *loc_ref, kmp_int32 gtid, 3849 int end_part); 3850 3851 #if TASK_UNUSED 3852 void __kmpc_omp_task_begin(ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *task); 3853 void __kmpc_omp_task_complete(ident_t *loc_ref, kmp_int32 gtid, 3854 kmp_task_t *task); 3855 #endif // TASK_UNUSED 3856 3857 /* ------------------------------------------------------------------------ */ 3858 3859 KMP_EXPORT void __kmpc_taskgroup(ident_t *loc, int gtid); 3860 KMP_EXPORT void __kmpc_end_taskgroup(ident_t *loc, int gtid); 3861 3862 KMP_EXPORT kmp_int32 __kmpc_omp_task_with_deps( 3863 ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 ndeps, 3864 kmp_depend_info_t *dep_list, kmp_int32 ndeps_noalias, 3865 kmp_depend_info_t *noalias_dep_list); 3866 KMP_EXPORT void __kmpc_omp_wait_deps(ident_t *loc_ref, kmp_int32 gtid, 3867 kmp_int32 ndeps, 3868 kmp_depend_info_t *dep_list, 3869 kmp_int32 ndeps_noalias, 3870 kmp_depend_info_t *noalias_dep_list); 3871 extern kmp_int32 __kmp_omp_task(kmp_int32 gtid, kmp_task_t *new_task, 3872 bool serialize_immediate); 3873 3874 KMP_EXPORT kmp_int32 __kmpc_cancel(ident_t *loc_ref, kmp_int32 gtid, 3875 kmp_int32 cncl_kind); 3876 KMP_EXPORT kmp_int32 __kmpc_cancellationpoint(ident_t *loc_ref, kmp_int32 gtid, 3877 kmp_int32 cncl_kind); 3878 KMP_EXPORT kmp_int32 __kmpc_cancel_barrier(ident_t *loc_ref, kmp_int32 gtid); 3879 KMP_EXPORT int __kmp_get_cancellation_status(int cancel_kind); 3880 3881 KMP_EXPORT void __kmpc_proxy_task_completed(kmp_int32 gtid, kmp_task_t *ptask); 3882 KMP_EXPORT void __kmpc_proxy_task_completed_ooo(kmp_task_t *ptask); 3883 KMP_EXPORT void __kmpc_taskloop(ident_t *loc, kmp_int32 gtid, kmp_task_t *task, 3884 kmp_int32 if_val, kmp_uint64 *lb, 3885 kmp_uint64 *ub, kmp_int64 st, kmp_int32 nogroup, 3886 kmp_int32 sched, kmp_uint64 grainsize, 3887 void *task_dup); 3888 KMP_EXPORT void __kmpc_taskloop_5(ident_t *loc, kmp_int32 gtid, 3889 kmp_task_t *task, kmp_int32 if_val, 3890 kmp_uint64 *lb, kmp_uint64 *ub, kmp_int64 st, 3891 kmp_int32 nogroup, kmp_int32 sched, 3892 kmp_uint64 grainsize, kmp_int32 modifier, 3893 void *task_dup); 3894 KMP_EXPORT void *__kmpc_task_reduction_init(int gtid, int num_data, void *data); 3895 KMP_EXPORT void *__kmpc_taskred_init(int gtid, int num_data, void *data); 3896 KMP_EXPORT void *__kmpc_task_reduction_get_th_data(int gtid, void *tg, void *d); 3897 KMP_EXPORT void *__kmpc_task_reduction_modifier_init(ident_t *loc, int gtid, 3898 int is_ws, int num, 3899 void *data); 3900 KMP_EXPORT void *__kmpc_taskred_modifier_init(ident_t *loc, int gtid, int is_ws, 3901 int num, void *data); 3902 KMP_EXPORT void __kmpc_task_reduction_modifier_fini(ident_t *loc, int gtid, 3903 int is_ws); 3904 KMP_EXPORT kmp_int32 __kmpc_omp_reg_task_with_affinity( 3905 ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 naffins, 3906 kmp_task_affinity_info_t *affin_list); 3907 KMP_EXPORT void __kmp_set_num_teams(int num_teams); 3908 KMP_EXPORT int __kmp_get_max_teams(void); 3909 KMP_EXPORT void __kmp_set_teams_thread_limit(int limit); 3910 KMP_EXPORT int __kmp_get_teams_thread_limit(void); 3911 3912 /* Lock interface routines (fast versions with gtid passed in) */ 3913 KMP_EXPORT void __kmpc_init_lock(ident_t *loc, kmp_int32 gtid, 3914 void **user_lock); 3915 KMP_EXPORT void __kmpc_init_nest_lock(ident_t *loc, kmp_int32 gtid, 3916 void **user_lock); 3917 KMP_EXPORT void __kmpc_destroy_lock(ident_t *loc, kmp_int32 gtid, 3918 void **user_lock); 3919 KMP_EXPORT void __kmpc_destroy_nest_lock(ident_t *loc, kmp_int32 gtid, 3920 void **user_lock); 3921 KMP_EXPORT void __kmpc_set_lock(ident_t *loc, kmp_int32 gtid, void **user_lock); 3922 KMP_EXPORT void __kmpc_set_nest_lock(ident_t *loc, kmp_int32 gtid, 3923 void **user_lock); 3924 KMP_EXPORT void __kmpc_unset_lock(ident_t *loc, kmp_int32 gtid, 3925 void **user_lock); 3926 KMP_EXPORT void __kmpc_unset_nest_lock(ident_t *loc, kmp_int32 gtid, 3927 void **user_lock); 3928 KMP_EXPORT int __kmpc_test_lock(ident_t *loc, kmp_int32 gtid, void **user_lock); 3929 KMP_EXPORT int __kmpc_test_nest_lock(ident_t *loc, kmp_int32 gtid, 3930 void **user_lock); 3931 3932 KMP_EXPORT void __kmpc_init_lock_with_hint(ident_t *loc, kmp_int32 gtid, 3933 void **user_lock, uintptr_t hint); 3934 KMP_EXPORT void __kmpc_init_nest_lock_with_hint(ident_t *loc, kmp_int32 gtid, 3935 void **user_lock, 3936 uintptr_t hint); 3937 3938 /* Interface to fast scalable reduce methods routines */ 3939 3940 KMP_EXPORT kmp_int32 __kmpc_reduce_nowait( 3941 ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size, 3942 void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data), 3943 kmp_critical_name *lck); 3944 KMP_EXPORT void __kmpc_end_reduce_nowait(ident_t *loc, kmp_int32 global_tid, 3945 kmp_critical_name *lck); 3946 KMP_EXPORT kmp_int32 __kmpc_reduce( 3947 ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size, 3948 void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data), 3949 kmp_critical_name *lck); 3950 KMP_EXPORT void __kmpc_end_reduce(ident_t *loc, kmp_int32 global_tid, 3951 kmp_critical_name *lck); 3952 3953 /* Internal fast reduction routines */ 3954 3955 extern PACKED_REDUCTION_METHOD_T __kmp_determine_reduction_method( 3956 ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size, 3957 void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data), 3958 kmp_critical_name *lck); 3959 3960 // this function is for testing set/get/determine reduce method 3961 KMP_EXPORT kmp_int32 __kmp_get_reduce_method(void); 3962 3963 KMP_EXPORT kmp_uint64 __kmpc_get_taskid(); 3964 KMP_EXPORT kmp_uint64 __kmpc_get_parent_taskid(); 3965 3966 // C++ port 3967 // missing 'extern "C"' declarations 3968 3969 KMP_EXPORT kmp_int32 __kmpc_in_parallel(ident_t *loc); 3970 KMP_EXPORT void __kmpc_pop_num_threads(ident_t *loc, kmp_int32 global_tid); 3971 KMP_EXPORT void __kmpc_push_num_threads(ident_t *loc, kmp_int32 global_tid, 3972 kmp_int32 num_threads); 3973 3974 KMP_EXPORT void __kmpc_push_proc_bind(ident_t *loc, kmp_int32 global_tid, 3975 int proc_bind); 3976 KMP_EXPORT void __kmpc_push_num_teams(ident_t *loc, kmp_int32 global_tid, 3977 kmp_int32 num_teams, 3978 kmp_int32 num_threads); 3979 /* Function for OpenMP 5.1 num_teams clause */ 3980 KMP_EXPORT void __kmpc_push_num_teams_51(ident_t *loc, kmp_int32 global_tid, 3981 kmp_int32 num_teams_lb, 3982 kmp_int32 num_teams_ub, 3983 kmp_int32 num_threads); 3984 KMP_EXPORT void __kmpc_fork_teams(ident_t *loc, kmp_int32 argc, 3985 kmpc_micro microtask, ...); 3986 struct kmp_dim { // loop bounds info casted to kmp_int64 3987 kmp_int64 lo; // lower 3988 kmp_int64 up; // upper 3989 kmp_int64 st; // stride 3990 }; 3991 KMP_EXPORT void __kmpc_doacross_init(ident_t *loc, kmp_int32 gtid, 3992 kmp_int32 num_dims, 3993 const struct kmp_dim *dims); 3994 KMP_EXPORT void __kmpc_doacross_wait(ident_t *loc, kmp_int32 gtid, 3995 const kmp_int64 *vec); 3996 KMP_EXPORT void __kmpc_doacross_post(ident_t *loc, kmp_int32 gtid, 3997 const kmp_int64 *vec); 3998 KMP_EXPORT void __kmpc_doacross_fini(ident_t *loc, kmp_int32 gtid); 3999 4000 KMP_EXPORT void *__kmpc_threadprivate_cached(ident_t *loc, kmp_int32 global_tid, 4001 void *data, size_t size, 4002 void ***cache); 4003 4004 // Symbols for MS mutual detection. 4005 extern int _You_must_link_with_exactly_one_OpenMP_library; 4006 extern int _You_must_link_with_Intel_OpenMP_library; 4007 #if KMP_OS_WINDOWS && (KMP_VERSION_MAJOR > 4) 4008 extern int _You_must_link_with_Microsoft_OpenMP_library; 4009 #endif 4010 4011 // The routines below are not exported. 4012 // Consider making them 'static' in corresponding source files. 4013 void kmp_threadprivate_insert_private_data(int gtid, void *pc_addr, 4014 void *data_addr, size_t pc_size); 4015 struct private_common *kmp_threadprivate_insert(int gtid, void *pc_addr, 4016 void *data_addr, 4017 size_t pc_size); 4018 void __kmp_threadprivate_resize_cache(int newCapacity); 4019 void __kmp_cleanup_threadprivate_caches(); 4020 4021 // ompc_, kmpc_ entries moved from omp.h. 4022 #if KMP_OS_WINDOWS 4023 #define KMPC_CONVENTION __cdecl 4024 #else 4025 #define KMPC_CONVENTION 4026 #endif 4027 4028 #ifndef __OMP_H 4029 typedef enum omp_sched_t { 4030 omp_sched_static = 1, 4031 omp_sched_dynamic = 2, 4032 omp_sched_guided = 3, 4033 omp_sched_auto = 4 4034 } omp_sched_t; 4035 typedef void *kmp_affinity_mask_t; 4036 #endif 4037 4038 KMP_EXPORT void KMPC_CONVENTION ompc_set_max_active_levels(int); 4039 KMP_EXPORT void KMPC_CONVENTION ompc_set_schedule(omp_sched_t, int); 4040 KMP_EXPORT int KMPC_CONVENTION ompc_get_ancestor_thread_num(int); 4041 KMP_EXPORT int KMPC_CONVENTION ompc_get_team_size(int); 4042 KMP_EXPORT int KMPC_CONVENTION 4043 kmpc_set_affinity_mask_proc(int, kmp_affinity_mask_t *); 4044 KMP_EXPORT int KMPC_CONVENTION 4045 kmpc_unset_affinity_mask_proc(int, kmp_affinity_mask_t *); 4046 KMP_EXPORT int KMPC_CONVENTION 4047 kmpc_get_affinity_mask_proc(int, kmp_affinity_mask_t *); 4048 4049 KMP_EXPORT void KMPC_CONVENTION kmpc_set_stacksize(int); 4050 KMP_EXPORT void KMPC_CONVENTION kmpc_set_stacksize_s(size_t); 4051 KMP_EXPORT void KMPC_CONVENTION kmpc_set_library(int); 4052 KMP_EXPORT void KMPC_CONVENTION kmpc_set_defaults(char const *); 4053 KMP_EXPORT void KMPC_CONVENTION kmpc_set_disp_num_buffers(int); 4054 void KMP_EXPAND_NAME(ompc_set_affinity_format)(char const *format); 4055 size_t KMP_EXPAND_NAME(ompc_get_affinity_format)(char *buffer, size_t size); 4056 void KMP_EXPAND_NAME(ompc_display_affinity)(char const *format); 4057 size_t KMP_EXPAND_NAME(ompc_capture_affinity)(char *buffer, size_t buf_size, 4058 char const *format); 4059 4060 enum kmp_target_offload_kind { 4061 tgt_disabled = 0, 4062 tgt_default = 1, 4063 tgt_mandatory = 2 4064 }; 4065 typedef enum kmp_target_offload_kind kmp_target_offload_kind_t; 4066 // Set via OMP_TARGET_OFFLOAD if specified, defaults to tgt_default otherwise 4067 extern kmp_target_offload_kind_t __kmp_target_offload; 4068 extern int __kmpc_get_target_offload(); 4069 4070 // Constants used in libomptarget 4071 #define KMP_DEVICE_DEFAULT -1 // This is libomptarget's default device. 4072 #define KMP_DEVICE_ALL -11 // This is libomptarget's "all devices". 4073 4074 // OMP Pause Resource 4075 4076 // The following enum is used both to set the status in __kmp_pause_status, and 4077 // as the internal equivalent of the externally-visible omp_pause_resource_t. 4078 typedef enum kmp_pause_status_t { 4079 kmp_not_paused = 0, // status is not paused, or, requesting resume 4080 kmp_soft_paused = 1, // status is soft-paused, or, requesting soft pause 4081 kmp_hard_paused = 2 // status is hard-paused, or, requesting hard pause 4082 } kmp_pause_status_t; 4083 4084 // This stores the pause state of the runtime 4085 extern kmp_pause_status_t __kmp_pause_status; 4086 extern int __kmpc_pause_resource(kmp_pause_status_t level); 4087 extern int __kmp_pause_resource(kmp_pause_status_t level); 4088 // Soft resume sets __kmp_pause_status, and wakes up all threads. 4089 extern void __kmp_resume_if_soft_paused(); 4090 // Hard resume simply resets the status to not paused. Library will appear to 4091 // be uninitialized after hard pause. Let OMP constructs trigger required 4092 // initializations. 4093 static inline void __kmp_resume_if_hard_paused() { 4094 if (__kmp_pause_status == kmp_hard_paused) { 4095 __kmp_pause_status = kmp_not_paused; 4096 } 4097 } 4098 4099 extern void __kmp_omp_display_env(int verbose); 4100 4101 // 1: it is initializing hidden helper team 4102 extern volatile int __kmp_init_hidden_helper; 4103 // 1: the hidden helper team is done 4104 extern volatile int __kmp_hidden_helper_team_done; 4105 // 1: enable hidden helper task 4106 extern kmp_int32 __kmp_enable_hidden_helper; 4107 // Main thread of hidden helper team 4108 extern kmp_info_t *__kmp_hidden_helper_main_thread; 4109 // Descriptors for the hidden helper threads 4110 extern kmp_info_t **__kmp_hidden_helper_threads; 4111 // Number of hidden helper threads 4112 extern kmp_int32 __kmp_hidden_helper_threads_num; 4113 // Number of hidden helper tasks that have not been executed yet 4114 extern std::atomic<kmp_int32> __kmp_unexecuted_hidden_helper_tasks; 4115 4116 extern void __kmp_hidden_helper_initialize(); 4117 extern void __kmp_hidden_helper_threads_initz_routine(); 4118 extern void __kmp_do_initialize_hidden_helper_threads(); 4119 extern void __kmp_hidden_helper_threads_initz_wait(); 4120 extern void __kmp_hidden_helper_initz_release(); 4121 extern void __kmp_hidden_helper_threads_deinitz_wait(); 4122 extern void __kmp_hidden_helper_threads_deinitz_release(); 4123 extern void __kmp_hidden_helper_main_thread_wait(); 4124 extern void __kmp_hidden_helper_worker_thread_wait(); 4125 extern void __kmp_hidden_helper_worker_thread_signal(); 4126 extern void __kmp_hidden_helper_main_thread_release(); 4127 4128 // Check whether a given thread is a hidden helper thread 4129 #define KMP_HIDDEN_HELPER_THREAD(gtid) \ 4130 ((gtid) >= 1 && (gtid) <= __kmp_hidden_helper_threads_num) 4131 4132 #define KMP_HIDDEN_HELPER_WORKER_THREAD(gtid) \ 4133 ((gtid) > 1 && (gtid) <= __kmp_hidden_helper_threads_num) 4134 4135 #define KMP_HIDDEN_HELPER_TEAM(team) \ 4136 (team->t.t_threads[0] == __kmp_hidden_helper_main_thread) 4137 4138 // Map a gtid to a hidden helper thread. The first hidden helper thread, a.k.a 4139 // main thread, is skipped. 4140 #define KMP_GTID_TO_SHADOW_GTID(gtid) \ 4141 ((gtid) % (__kmp_hidden_helper_threads_num - 1) + 2) 4142 4143 // Return the adjusted gtid value by subtracting from gtid the number 4144 // of hidden helper threads. This adjusted value is the gtid the thread would 4145 // have received if there were no hidden helper threads. 4146 static inline int __kmp_adjust_gtid_for_hidden_helpers(int gtid) { 4147 int adjusted_gtid = gtid; 4148 if (__kmp_hidden_helper_threads_num > 0 && gtid > 0 && 4149 gtid - __kmp_hidden_helper_threads_num >= 0) { 4150 adjusted_gtid -= __kmp_hidden_helper_threads_num; 4151 } 4152 return adjusted_gtid; 4153 } 4154 4155 // Support for error directive 4156 typedef enum kmp_severity_t { 4157 severity_warning = 1, 4158 severity_fatal = 2 4159 } kmp_severity_t; 4160 extern void __kmpc_error(ident_t *loc, int severity, const char *message); 4161 4162 // Support for scope directive 4163 KMP_EXPORT void __kmpc_scope(ident_t *loc, kmp_int32 gtid, void *reserved); 4164 KMP_EXPORT void __kmpc_end_scope(ident_t *loc, kmp_int32 gtid, void *reserved); 4165 4166 #ifdef __cplusplus 4167 } 4168 #endif 4169 4170 template <bool C, bool S> 4171 extern void __kmp_suspend_32(int th_gtid, kmp_flag_32<C, S> *flag); 4172 template <bool C, bool S> 4173 extern void __kmp_suspend_64(int th_gtid, kmp_flag_64<C, S> *flag); 4174 template <bool C, bool S> 4175 extern void __kmp_atomic_suspend_64(int th_gtid, 4176 kmp_atomic_flag_64<C, S> *flag); 4177 extern void __kmp_suspend_oncore(int th_gtid, kmp_flag_oncore *flag); 4178 #if KMP_HAVE_MWAIT || KMP_HAVE_UMWAIT 4179 template <bool C, bool S> 4180 extern void __kmp_mwait_32(int th_gtid, kmp_flag_32<C, S> *flag); 4181 template <bool C, bool S> 4182 extern void __kmp_mwait_64(int th_gtid, kmp_flag_64<C, S> *flag); 4183 template <bool C, bool S> 4184 extern void __kmp_atomic_mwait_64(int th_gtid, kmp_atomic_flag_64<C, S> *flag); 4185 extern void __kmp_mwait_oncore(int th_gtid, kmp_flag_oncore *flag); 4186 #endif 4187 template <bool C, bool S> 4188 extern void __kmp_resume_32(int target_gtid, kmp_flag_32<C, S> *flag); 4189 template <bool C, bool S> 4190 extern void __kmp_resume_64(int target_gtid, kmp_flag_64<C, S> *flag); 4191 template <bool C, bool S> 4192 extern void __kmp_atomic_resume_64(int target_gtid, 4193 kmp_atomic_flag_64<C, S> *flag); 4194 extern void __kmp_resume_oncore(int target_gtid, kmp_flag_oncore *flag); 4195 4196 template <bool C, bool S> 4197 int __kmp_execute_tasks_32(kmp_info_t *thread, kmp_int32 gtid, 4198 kmp_flag_32<C, S> *flag, int final_spin, 4199 int *thread_finished, 4200 #if USE_ITT_BUILD 4201 void *itt_sync_obj, 4202 #endif /* USE_ITT_BUILD */ 4203 kmp_int32 is_constrained); 4204 template <bool C, bool S> 4205 int __kmp_execute_tasks_64(kmp_info_t *thread, kmp_int32 gtid, 4206 kmp_flag_64<C, S> *flag, int final_spin, 4207 int *thread_finished, 4208 #if USE_ITT_BUILD 4209 void *itt_sync_obj, 4210 #endif /* USE_ITT_BUILD */ 4211 kmp_int32 is_constrained); 4212 template <bool C, bool S> 4213 int __kmp_atomic_execute_tasks_64(kmp_info_t *thread, kmp_int32 gtid, 4214 kmp_atomic_flag_64<C, S> *flag, 4215 int final_spin, int *thread_finished, 4216 #if USE_ITT_BUILD 4217 void *itt_sync_obj, 4218 #endif /* USE_ITT_BUILD */ 4219 kmp_int32 is_constrained); 4220 int __kmp_execute_tasks_oncore(kmp_info_t *thread, kmp_int32 gtid, 4221 kmp_flag_oncore *flag, int final_spin, 4222 int *thread_finished, 4223 #if USE_ITT_BUILD 4224 void *itt_sync_obj, 4225 #endif /* USE_ITT_BUILD */ 4226 kmp_int32 is_constrained); 4227 4228 extern int __kmp_nesting_mode; 4229 extern int __kmp_nesting_mode_nlevels; 4230 extern int *__kmp_nesting_nth_level; 4231 extern void __kmp_init_nesting_mode(); 4232 extern void __kmp_set_nesting_mode_threads(); 4233 4234 /// This class safely opens and closes a C-style FILE* object using RAII 4235 /// semantics. There are also methods which allow using stdout or stderr as 4236 /// the underlying FILE* object. With the implicit conversion operator to 4237 /// FILE*, an object with this type can be used in any function which takes 4238 /// a FILE* object e.g., fprintf(). 4239 /// No close method is needed at use sites. 4240 class kmp_safe_raii_file_t { 4241 FILE *f; 4242 4243 void close() { 4244 if (f && f != stdout && f != stderr) { 4245 fclose(f); 4246 f = nullptr; 4247 } 4248 } 4249 4250 public: 4251 kmp_safe_raii_file_t() : f(nullptr) {} 4252 kmp_safe_raii_file_t(const char *filename, const char *mode, 4253 const char *env_var = nullptr) 4254 : f(nullptr) { 4255 open(filename, mode, env_var); 4256 } 4257 ~kmp_safe_raii_file_t() { close(); } 4258 4259 /// Open filename using mode. This is automatically closed in the destructor. 4260 /// The env_var parameter indicates the environment variable the filename 4261 /// came from if != nullptr. 4262 void open(const char *filename, const char *mode, 4263 const char *env_var = nullptr) { 4264 KMP_ASSERT(!f); 4265 f = fopen(filename, mode); 4266 if (!f) { 4267 int code = errno; 4268 if (env_var) { 4269 __kmp_fatal(KMP_MSG(CantOpenFileForReading, filename), KMP_ERR(code), 4270 KMP_HNT(CheckEnvVar, env_var, filename), __kmp_msg_null); 4271 } else { 4272 __kmp_fatal(KMP_MSG(CantOpenFileForReading, filename), KMP_ERR(code), 4273 __kmp_msg_null); 4274 } 4275 } 4276 } 4277 /// Instead of erroring out, return non-zero when 4278 /// unsuccessful fopen() for any reason 4279 int try_open(const char *filename, const char *mode) { 4280 KMP_ASSERT(!f); 4281 f = fopen(filename, mode); 4282 if (!f) 4283 return errno; 4284 return 0; 4285 } 4286 /// Set the FILE* object to stdout and output there 4287 /// No open call should happen before this call. 4288 void set_stdout() { 4289 KMP_ASSERT(!f); 4290 f = stdout; 4291 } 4292 /// Set the FILE* object to stderr and output there 4293 /// No open call should happen before this call. 4294 void set_stderr() { 4295 KMP_ASSERT(!f); 4296 f = stderr; 4297 } 4298 operator bool() { return bool(f); } 4299 operator FILE *() { return f; } 4300 }; 4301 4302 template <typename SourceType, typename TargetType, 4303 bool isSourceSmaller = (sizeof(SourceType) < sizeof(TargetType)), 4304 bool isSourceEqual = (sizeof(SourceType) == sizeof(TargetType)), 4305 bool isSourceSigned = std::is_signed<SourceType>::value, 4306 bool isTargetSigned = std::is_signed<TargetType>::value> 4307 struct kmp_convert {}; 4308 4309 // Both types are signed; Source smaller 4310 template <typename SourceType, typename TargetType> 4311 struct kmp_convert<SourceType, TargetType, true, false, true, true> { 4312 static TargetType to(SourceType src) { return (TargetType)src; } 4313 }; 4314 // Source equal 4315 template <typename SourceType, typename TargetType> 4316 struct kmp_convert<SourceType, TargetType, false, true, true, true> { 4317 static TargetType to(SourceType src) { return src; } 4318 }; 4319 // Source bigger 4320 template <typename SourceType, typename TargetType> 4321 struct kmp_convert<SourceType, TargetType, false, false, true, true> { 4322 static TargetType to(SourceType src) { 4323 KMP_ASSERT(src <= static_cast<SourceType>( 4324 (std::numeric_limits<TargetType>::max)())); 4325 KMP_ASSERT(src >= static_cast<SourceType>( 4326 (std::numeric_limits<TargetType>::min)())); 4327 return (TargetType)src; 4328 } 4329 }; 4330 4331 // Source signed, Target unsigned 4332 // Source smaller 4333 template <typename SourceType, typename TargetType> 4334 struct kmp_convert<SourceType, TargetType, true, false, true, false> { 4335 static TargetType to(SourceType src) { 4336 KMP_ASSERT(src >= 0); 4337 return (TargetType)src; 4338 } 4339 }; 4340 // Source equal 4341 template <typename SourceType, typename TargetType> 4342 struct kmp_convert<SourceType, TargetType, false, true, true, false> { 4343 static TargetType to(SourceType src) { 4344 KMP_ASSERT(src >= 0); 4345 return (TargetType)src; 4346 } 4347 }; 4348 // Source bigger 4349 template <typename SourceType, typename TargetType> 4350 struct kmp_convert<SourceType, TargetType, false, false, true, false> { 4351 static TargetType to(SourceType src) { 4352 KMP_ASSERT(src >= 0); 4353 KMP_ASSERT(src <= static_cast<SourceType>( 4354 (std::numeric_limits<TargetType>::max)())); 4355 return (TargetType)src; 4356 } 4357 }; 4358 4359 // Source unsigned, Target signed 4360 // Source smaller 4361 template <typename SourceType, typename TargetType> 4362 struct kmp_convert<SourceType, TargetType, true, false, false, true> { 4363 static TargetType to(SourceType src) { return (TargetType)src; } 4364 }; 4365 // Source equal 4366 template <typename SourceType, typename TargetType> 4367 struct kmp_convert<SourceType, TargetType, false, true, false, true> { 4368 static TargetType to(SourceType src) { 4369 KMP_ASSERT(src <= static_cast<SourceType>( 4370 (std::numeric_limits<TargetType>::max)())); 4371 return (TargetType)src; 4372 } 4373 }; 4374 // Source bigger 4375 template <typename SourceType, typename TargetType> 4376 struct kmp_convert<SourceType, TargetType, false, false, false, true> { 4377 static TargetType to(SourceType src) { 4378 KMP_ASSERT(src <= static_cast<SourceType>( 4379 (std::numeric_limits<TargetType>::max)())); 4380 return (TargetType)src; 4381 } 4382 }; 4383 4384 // Source unsigned, Target unsigned 4385 // Source smaller 4386 template <typename SourceType, typename TargetType> 4387 struct kmp_convert<SourceType, TargetType, true, false, false, false> { 4388 static TargetType to(SourceType src) { return (TargetType)src; } 4389 }; 4390 // Source equal 4391 template <typename SourceType, typename TargetType> 4392 struct kmp_convert<SourceType, TargetType, false, true, false, false> { 4393 static TargetType to(SourceType src) { return src; } 4394 }; 4395 // Source bigger 4396 template <typename SourceType, typename TargetType> 4397 struct kmp_convert<SourceType, TargetType, false, false, false, false> { 4398 static TargetType to(SourceType src) { 4399 KMP_ASSERT(src <= static_cast<SourceType>( 4400 (std::numeric_limits<TargetType>::max)())); 4401 return (TargetType)src; 4402 } 4403 }; 4404 4405 template <typename T1, typename T2> 4406 static inline void __kmp_type_convert(T1 src, T2 *dest) { 4407 *dest = kmp_convert<T1, T2>::to(src); 4408 } 4409 4410 #endif /* KMP_H */ 4411