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