1 /*! \file */ 2 /* 3 * kmp.h -- KPTS runtime header file. 4 */ 5 6 //===----------------------------------------------------------------------===// 7 // 8 // The LLVM Compiler Infrastructure 9 // 10 // This file is dual licensed under the MIT and the University of Illinois Open 11 // Source Licenses. See LICENSE.txt for details. 12 // 13 //===----------------------------------------------------------------------===// 14 15 #ifndef KMP_H 16 #define KMP_H 17 18 #include "kmp_config.h" 19 20 /* #define BUILD_PARALLEL_ORDERED 1 */ 21 22 /* This fix replaces gettimeofday with clock_gettime for better scalability on 23 the Altix. Requires user code to be linked with -lrt. */ 24 //#define FIX_SGI_CLOCK 25 26 /* Defines for OpenMP 3.0 tasking and auto scheduling */ 27 28 #ifndef KMP_STATIC_STEAL_ENABLED 29 #define KMP_STATIC_STEAL_ENABLED 1 30 #endif 31 32 #define TASK_CURRENT_NOT_QUEUED 0 33 #define TASK_CURRENT_QUEUED 1 34 35 #ifdef BUILD_TIED_TASK_STACK 36 #define TASK_STACK_EMPTY 0 // entries when the stack is empty 37 #define TASK_STACK_BLOCK_BITS 5 // Used in TASK_STACK_SIZE and TASK_STACK_MASK 38 // Number of entries in each task stack array 39 #define TASK_STACK_BLOCK_SIZE (1 << TASK_STACK_BLOCK_BITS) 40 // Mask for determining index into stack block 41 #define TASK_STACK_INDEX_MASK (TASK_STACK_BLOCK_SIZE - 1) 42 #endif // BUILD_TIED_TASK_STACK 43 44 #define TASK_NOT_PUSHED 1 45 #define TASK_SUCCESSFULLY_PUSHED 0 46 #define TASK_TIED 1 47 #define TASK_UNTIED 0 48 #define TASK_EXPLICIT 1 49 #define TASK_IMPLICIT 0 50 #define TASK_PROXY 1 51 #define TASK_FULL 0 52 53 #define KMP_CANCEL_THREADS 54 #define KMP_THREAD_ATTR 55 56 // Android does not have pthread_cancel. Undefine KMP_CANCEL_THREADS if being 57 // built on Android 58 #if defined(__ANDROID__) 59 #undef KMP_CANCEL_THREADS 60 #endif 61 62 #include <signal.h> 63 #include <stdarg.h> 64 #include <stddef.h> 65 #include <stdio.h> 66 #include <stdlib.h> 67 #include <string.h> 68 /* include <ctype.h> don't use; problems with /MD on Windows* OS NT due to bad 69 Microsoft library. Some macros provided below to replace these functions */ 70 #ifndef __ABSOFT_WIN 71 #include <sys/types.h> 72 #endif 73 #include <limits.h> 74 #include <time.h> 75 76 #include <errno.h> 77 78 #include "kmp_os.h" 79 80 #include "kmp_safe_c_api.h" 81 82 #if KMP_STATS_ENABLED 83 class kmp_stats_list; 84 #endif 85 86 #if KMP_USE_HWLOC && KMP_AFFINITY_SUPPORTED 87 #include "hwloc.h" 88 #ifndef HWLOC_OBJ_NUMANODE 89 #define HWLOC_OBJ_NUMANODE HWLOC_OBJ_NODE 90 #endif 91 #ifndef HWLOC_OBJ_PACKAGE 92 #define HWLOC_OBJ_PACKAGE HWLOC_OBJ_SOCKET 93 #endif 94 #endif 95 96 #if KMP_ARCH_X86 || KMP_ARCH_X86_64 97 #include <xmmintrin.h> 98 #endif 99 100 #include "kmp_debug.h" 101 #include "kmp_lock.h" 102 #include "kmp_version.h" 103 #if USE_DEBUGGER 104 #include "kmp_debugger.h" 105 #endif 106 #include "kmp_i18n.h" 107 108 #define KMP_HANDLE_SIGNALS (KMP_OS_UNIX || KMP_OS_WINDOWS) 109 110 #include "kmp_wrapper_malloc.h" 111 #if KMP_OS_UNIX 112 #include <unistd.h> 113 #if !defined NSIG && defined _NSIG 114 #define NSIG _NSIG 115 #endif 116 #endif 117 118 #if KMP_OS_LINUX 119 #pragma weak clock_gettime 120 #endif 121 122 #if OMPT_SUPPORT 123 #include "ompt-internal.h" 124 #endif 125 126 /*Select data placement in NUMA memory */ 127 #define NO_FIRST_TOUCH 0 128 #define FIRST_TOUCH 1 /* Exploit SGI's first touch page placement algo */ 129 130 /* If not specified on compile command line, assume no first touch */ 131 #ifndef BUILD_MEMORY 132 #define BUILD_MEMORY NO_FIRST_TOUCH 133 #endif 134 135 // 0 - no fast memory allocation, alignment: 8-byte on x86, 16-byte on x64. 136 // 3 - fast allocation using sync, non-sync free lists of any size, non-self 137 // free lists of limited size. 138 #ifndef USE_FAST_MEMORY 139 #define USE_FAST_MEMORY 3 140 #endif 141 142 #ifndef KMP_NESTED_HOT_TEAMS 143 #define KMP_NESTED_HOT_TEAMS 0 144 #define USE_NESTED_HOT_ARG(x) 145 #else 146 #if KMP_NESTED_HOT_TEAMS 147 #if OMP_40_ENABLED 148 #define USE_NESTED_HOT_ARG(x) , x 149 #else 150 // Nested hot teams feature depends on omp 4.0, disable it for earlier versions 151 #undef KMP_NESTED_HOT_TEAMS 152 #define KMP_NESTED_HOT_TEAMS 0 153 #define USE_NESTED_HOT_ARG(x) 154 #endif 155 #else 156 #define USE_NESTED_HOT_ARG(x) 157 #endif 158 #endif 159 160 // Assume using BGET compare_exchange instruction instead of lock by default. 161 #ifndef USE_CMP_XCHG_FOR_BGET 162 #define USE_CMP_XCHG_FOR_BGET 1 163 #endif 164 165 // Test to see if queuing lock is better than bootstrap lock for bget 166 // #ifndef USE_QUEUING_LOCK_FOR_BGET 167 // #define USE_QUEUING_LOCK_FOR_BGET 168 // #endif 169 170 #define KMP_NSEC_PER_SEC 1000000000L 171 #define KMP_USEC_PER_SEC 1000000L 172 173 /*! 174 @ingroup BASIC_TYPES 175 @{ 176 */ 177 178 // FIXME DOXYGEN... need to group these flags somehow (Making them an anonymous 179 // enum would do it...) 180 /*! 181 Values for bit flags used in the ident_t to describe the fields. 182 */ 183 /*! Use trampoline for internal microtasks */ 184 #define KMP_IDENT_IMB 0x01 185 /*! Use c-style ident structure */ 186 #define KMP_IDENT_KMPC 0x02 187 /* 0x04 is no longer used */ 188 /*! Entry point generated by auto-parallelization */ 189 #define KMP_IDENT_AUTOPAR 0x08 190 /*! Compiler generates atomic reduction option for kmpc_reduce* */ 191 #define KMP_IDENT_ATOMIC_REDUCE 0x10 192 /*! To mark a 'barrier' directive in user code */ 193 #define KMP_IDENT_BARRIER_EXPL 0x20 194 /*! To Mark implicit barriers. */ 195 #define KMP_IDENT_BARRIER_IMPL 0x0040 196 #define KMP_IDENT_BARRIER_IMPL_MASK 0x01C0 197 #define KMP_IDENT_BARRIER_IMPL_FOR 0x0040 198 #define KMP_IDENT_BARRIER_IMPL_SECTIONS 0x00C0 199 200 #define KMP_IDENT_BARRIER_IMPL_SINGLE 0x0140 201 #define KMP_IDENT_BARRIER_IMPL_WORKSHARE 0x01C0 202 203 #define KMP_IDENT_WORK_LOOP 0x200 // static loop 204 #define KMP_IDENT_WORK_SECTIONS 0x400 // sections 205 #define KMP_IDENT_WORK_DISTRIBUTE 0x800 // distribute 206 207 /*! 208 * The ident structure that describes a source location. 209 */ 210 typedef struct ident { 211 kmp_int32 reserved_1; /**< might be used in Fortran; see above */ 212 kmp_int32 flags; /**< also f.flags; KMP_IDENT_xxx flags; KMP_IDENT_KMPC 213 identifies this union member */ 214 kmp_int32 reserved_2; /**< not really used in Fortran any more; see above */ 215 #if USE_ITT_BUILD 216 /* but currently used for storing region-specific ITT */ 217 /* contextual information. */ 218 #endif /* USE_ITT_BUILD */ 219 kmp_int32 reserved_3; /**< source[4] in Fortran, do not use for C++ */ 220 char const *psource; /**< String describing the source location. 221 The string is composed of semi-colon separated fields 222 which describe the source file, the function and a pair 223 of line numbers that delimit the construct. */ 224 } ident_t; 225 /*! 226 @} 227 */ 228 229 // Some forward declarations. 230 typedef union kmp_team kmp_team_t; 231 typedef struct kmp_taskdata kmp_taskdata_t; 232 typedef union kmp_task_team kmp_task_team_t; 233 typedef union kmp_team kmp_team_p; 234 typedef union kmp_info kmp_info_p; 235 typedef union kmp_root kmp_root_p; 236 237 #ifdef __cplusplus 238 extern "C" { 239 #endif 240 241 /* ------------------------------------------------------------------------ */ 242 243 /* Pack two 32-bit signed integers into a 64-bit signed integer */ 244 /* ToDo: Fix word ordering for big-endian machines. */ 245 #define KMP_PACK_64(HIGH_32, LOW_32) \ 246 ((kmp_int64)((((kmp_uint64)(HIGH_32)) << 32) | (kmp_uint64)(LOW_32))) 247 248 // Generic string manipulation macros. Assume that _x is of type char * 249 #define SKIP_WS(_x) \ 250 { \ 251 while (*(_x) == ' ' || *(_x) == '\t') \ 252 (_x)++; \ 253 } 254 #define SKIP_DIGITS(_x) \ 255 { \ 256 while (*(_x) >= '0' && *(_x) <= '9') \ 257 (_x)++; \ 258 } 259 #define SKIP_TO(_x, _c) \ 260 { \ 261 while (*(_x) != '\0' && *(_x) != (_c)) \ 262 (_x)++; \ 263 } 264 265 /* ------------------------------------------------------------------------ */ 266 267 #define KMP_MAX(x, y) ((x) > (y) ? (x) : (y)) 268 #define KMP_MIN(x, y) ((x) < (y) ? (x) : (y)) 269 270 /* ------------------------------------------------------------------------ */ 271 /* Enumeration types */ 272 273 enum kmp_state_timer { 274 ts_stop, 275 ts_start, 276 ts_pause, 277 278 ts_last_state 279 }; 280 281 enum dynamic_mode { 282 dynamic_default, 283 #ifdef USE_LOAD_BALANCE 284 dynamic_load_balance, 285 #endif /* USE_LOAD_BALANCE */ 286 dynamic_random, 287 dynamic_thread_limit, 288 dynamic_max 289 }; 290 291 /* external schedule constants, duplicate enum omp_sched in omp.h in order to 292 * not include it here */ 293 #ifndef KMP_SCHED_TYPE_DEFINED 294 #define KMP_SCHED_TYPE_DEFINED 295 typedef enum kmp_sched { 296 kmp_sched_lower = 0, // lower and upper bounds are for routine parameter check 297 // Note: need to adjust __kmp_sch_map global array in case enum is changed 298 kmp_sched_static = 1, // mapped to kmp_sch_static_chunked (33) 299 kmp_sched_dynamic = 2, // mapped to kmp_sch_dynamic_chunked (35) 300 kmp_sched_guided = 3, // mapped to kmp_sch_guided_chunked (36) 301 kmp_sched_auto = 4, // mapped to kmp_sch_auto (38) 302 kmp_sched_upper_std = 5, // upper bound for standard schedules 303 kmp_sched_lower_ext = 100, // lower bound of Intel extension schedules 304 kmp_sched_trapezoidal = 101, // mapped to kmp_sch_trapezoidal (39) 305 #if KMP_STATIC_STEAL_ENABLED 306 kmp_sched_static_steal = 102, // mapped to kmp_sch_static_steal (44) 307 #endif 308 kmp_sched_upper, 309 kmp_sched_default = kmp_sched_static // default scheduling 310 } kmp_sched_t; 311 #endif 312 313 /*! 314 @ingroup WORK_SHARING 315 * Describes the loop schedule to be used for a parallel for loop. 316 */ 317 enum sched_type : kmp_int32 { 318 kmp_sch_lower = 32, /**< lower bound for unordered values */ 319 kmp_sch_static_chunked = 33, 320 kmp_sch_static = 34, /**< static unspecialized */ 321 kmp_sch_dynamic_chunked = 35, 322 kmp_sch_guided_chunked = 36, /**< guided unspecialized */ 323 kmp_sch_runtime = 37, 324 kmp_sch_auto = 38, /**< auto */ 325 kmp_sch_trapezoidal = 39, 326 327 /* accessible only through KMP_SCHEDULE environment variable */ 328 kmp_sch_static_greedy = 40, 329 kmp_sch_static_balanced = 41, 330 /* accessible only through KMP_SCHEDULE environment variable */ 331 kmp_sch_guided_iterative_chunked = 42, 332 kmp_sch_guided_analytical_chunked = 43, 333 /* accessible only through KMP_SCHEDULE environment variable */ 334 kmp_sch_static_steal = 44, 335 336 #if OMP_45_ENABLED 337 /* static with chunk adjustment (e.g., simd) */ 338 kmp_sch_static_balanced_chunked = 45, 339 kmp_sch_guided_simd = 46, /**< guided with chunk adjustment */ 340 kmp_sch_runtime_simd = 47, /**< runtime with chunk adjustment */ 341 #endif 342 343 /* accessible only through KMP_SCHEDULE environment variable */ 344 kmp_sch_upper, /**< upper bound for unordered values */ 345 346 kmp_ord_lower = 64, /**< lower bound for ordered values, must be power of 2 */ 347 kmp_ord_static_chunked = 65, 348 kmp_ord_static = 66, /**< ordered static unspecialized */ 349 kmp_ord_dynamic_chunked = 67, 350 kmp_ord_guided_chunked = 68, 351 kmp_ord_runtime = 69, 352 kmp_ord_auto = 70, /**< ordered auto */ 353 kmp_ord_trapezoidal = 71, 354 kmp_ord_upper, /**< upper bound for ordered values */ 355 356 #if OMP_40_ENABLED 357 /* Schedules for Distribute construct */ 358 kmp_distribute_static_chunked = 91, /**< distribute static chunked */ 359 kmp_distribute_static = 92, /**< distribute static unspecialized */ 360 #endif 361 362 /* For the "nomerge" versions, kmp_dispatch_next*() will always return a 363 single iteration/chunk, even if the loop is serialized. For the schedule 364 types listed above, the entire iteration vector is returned if the loop is 365 serialized. This doesn't work for gcc/gcomp sections. */ 366 kmp_nm_lower = 160, /**< lower bound for nomerge values */ 367 368 kmp_nm_static_chunked = 369 (kmp_sch_static_chunked - kmp_sch_lower + kmp_nm_lower), 370 kmp_nm_static = 162, /**< static unspecialized */ 371 kmp_nm_dynamic_chunked = 163, 372 kmp_nm_guided_chunked = 164, /**< guided unspecialized */ 373 kmp_nm_runtime = 165, 374 kmp_nm_auto = 166, /**< auto */ 375 kmp_nm_trapezoidal = 167, 376 377 /* accessible only through KMP_SCHEDULE environment variable */ 378 kmp_nm_static_greedy = 168, 379 kmp_nm_static_balanced = 169, 380 /* accessible only through KMP_SCHEDULE environment variable */ 381 kmp_nm_guided_iterative_chunked = 170, 382 kmp_nm_guided_analytical_chunked = 171, 383 kmp_nm_static_steal = 384 172, /* accessible only through OMP_SCHEDULE environment variable */ 385 386 kmp_nm_ord_static_chunked = 193, 387 kmp_nm_ord_static = 194, /**< ordered static unspecialized */ 388 kmp_nm_ord_dynamic_chunked = 195, 389 kmp_nm_ord_guided_chunked = 196, 390 kmp_nm_ord_runtime = 197, 391 kmp_nm_ord_auto = 198, /**< auto */ 392 kmp_nm_ord_trapezoidal = 199, 393 kmp_nm_upper, /**< upper bound for nomerge values */ 394 395 #if OMP_45_ENABLED 396 /* Support for OpenMP 4.5 monotonic and nonmonotonic schedule modifiers. Since 397 we need to distinguish the three possible cases (no modifier, monotonic 398 modifier, nonmonotonic modifier), we need separate bits for each modifier. 399 The absence of monotonic does not imply nonmonotonic, especially since 4.5 400 says that the behaviour of the "no modifier" case is implementation defined 401 in 4.5, but will become "nonmonotonic" in 5.0. 402 403 Since we're passing a full 32 bit value, we can use a couple of high bits 404 for these flags; out of paranoia we avoid the sign bit. 405 406 These modifiers can be or-ed into non-static schedules by the compiler to 407 pass the additional information. They will be stripped early in the 408 processing in __kmp_dispatch_init when setting up schedules, so most of the 409 code won't ever see schedules with these bits set. */ 410 kmp_sch_modifier_monotonic = 411 (1 << 29), /**< Set if the monotonic schedule modifier was present */ 412 kmp_sch_modifier_nonmonotonic = 413 (1 << 30), /**< Set if the nonmonotonic schedule modifier was present */ 414 415 #define SCHEDULE_WITHOUT_MODIFIERS(s) \ 416 (enum sched_type)( \ 417 (s) & ~(kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic)) 418 #define SCHEDULE_HAS_MONOTONIC(s) (((s)&kmp_sch_modifier_monotonic) != 0) 419 #define SCHEDULE_HAS_NONMONOTONIC(s) (((s)&kmp_sch_modifier_nonmonotonic) != 0) 420 #define SCHEDULE_HAS_NO_MODIFIERS(s) \ 421 (((s) & (kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic)) == 0) 422 #else 423 /* By doing this we hope to avoid multiple tests on OMP_45_ENABLED. Compilers 424 can now eliminate tests on compile time constants and dead code that results 425 from them, so we can leave code guarded by such an if in place. */ 426 #define SCHEDULE_WITHOUT_MODIFIERS(s) (s) 427 #define SCHEDULE_HAS_MONOTONIC(s) false 428 #define SCHEDULE_HAS_NONMONOTONIC(s) false 429 #define SCHEDULE_HAS_NO_MODIFIERS(s) true 430 #endif 431 432 kmp_sch_default = kmp_sch_static /**< default scheduling algorithm */ 433 }; 434 435 /* Type to keep runtime schedule set via OMP_SCHEDULE or omp_set_schedule() */ 436 typedef union kmp_r_sched { 437 struct { 438 enum sched_type r_sched_type; 439 int chunk; 440 }; 441 kmp_int64 sched; 442 } kmp_r_sched_t; 443 444 extern enum sched_type __kmp_sch_map[]; // map OMP 3.0 schedule types with our 445 // internal schedule types 446 447 enum library_type { 448 library_none, 449 library_serial, 450 library_turnaround, 451 library_throughput 452 }; 453 454 #if KMP_OS_LINUX 455 enum clock_function_type { 456 clock_function_gettimeofday, 457 clock_function_clock_gettime 458 }; 459 #endif /* KMP_OS_LINUX */ 460 461 #if KMP_MIC_SUPPORTED 462 enum mic_type { non_mic, mic1, mic2, mic3, dummy }; 463 #endif 464 465 /* -- fast reduction stuff ------------------------------------------------ */ 466 467 #undef KMP_FAST_REDUCTION_BARRIER 468 #define KMP_FAST_REDUCTION_BARRIER 1 469 470 #undef KMP_FAST_REDUCTION_CORE_DUO 471 #if KMP_ARCH_X86 || KMP_ARCH_X86_64 472 #define KMP_FAST_REDUCTION_CORE_DUO 1 473 #endif 474 475 enum _reduction_method { 476 reduction_method_not_defined = 0, 477 critical_reduce_block = (1 << 8), 478 atomic_reduce_block = (2 << 8), 479 tree_reduce_block = (3 << 8), 480 empty_reduce_block = (4 << 8) 481 }; 482 483 // Description of the packed_reduction_method variable: 484 // The packed_reduction_method variable consists of two enum types variables 485 // that are packed together into 0-th byte and 1-st byte: 486 // 0: (packed_reduction_method & 0x000000FF) is a 'enum barrier_type' value of 487 // barrier that will be used in fast reduction: bs_plain_barrier or 488 // bs_reduction_barrier 489 // 1: (packed_reduction_method & 0x0000FF00) is a reduction method that will 490 // be used in fast reduction; 491 // Reduction method is of 'enum _reduction_method' type and it's defined the way 492 // so that the bits of 0-th byte are empty, so no need to execute a shift 493 // instruction while packing/unpacking 494 495 #if KMP_FAST_REDUCTION_BARRIER 496 #define PACK_REDUCTION_METHOD_AND_BARRIER(reduction_method, barrier_type) \ 497 ((reduction_method) | (barrier_type)) 498 499 #define UNPACK_REDUCTION_METHOD(packed_reduction_method) \ 500 ((enum _reduction_method)((packed_reduction_method) & (0x0000FF00))) 501 502 #define UNPACK_REDUCTION_BARRIER(packed_reduction_method) \ 503 ((enum barrier_type)((packed_reduction_method) & (0x000000FF))) 504 #else 505 #define PACK_REDUCTION_METHOD_AND_BARRIER(reduction_method, barrier_type) \ 506 (reduction_method) 507 508 #define UNPACK_REDUCTION_METHOD(packed_reduction_method) \ 509 (packed_reduction_method) 510 511 #define UNPACK_REDUCTION_BARRIER(packed_reduction_method) (bs_plain_barrier) 512 #endif 513 514 #define TEST_REDUCTION_METHOD(packed_reduction_method, which_reduction_block) \ 515 ((UNPACK_REDUCTION_METHOD(packed_reduction_method)) == \ 516 (which_reduction_block)) 517 518 #if KMP_FAST_REDUCTION_BARRIER 519 #define TREE_REDUCE_BLOCK_WITH_REDUCTION_BARRIER \ 520 (PACK_REDUCTION_METHOD_AND_BARRIER(tree_reduce_block, bs_reduction_barrier)) 521 522 #define TREE_REDUCE_BLOCK_WITH_PLAIN_BARRIER \ 523 (PACK_REDUCTION_METHOD_AND_BARRIER(tree_reduce_block, bs_plain_barrier)) 524 #endif 525 526 typedef int PACKED_REDUCTION_METHOD_T; 527 528 /* -- end of fast reduction stuff ----------------------------------------- */ 529 530 #if KMP_OS_WINDOWS 531 #define USE_CBLKDATA 532 #pragma warning(push) 533 #pragma warning(disable : 271 310) 534 #include <windows.h> 535 #pragma warning(pop) 536 #endif 537 538 #if KMP_OS_UNIX 539 #include <dlfcn.h> 540 #include <pthread.h> 541 #endif 542 543 /* Only Linux* OS and Windows* OS support thread affinity. */ 544 #if KMP_AFFINITY_SUPPORTED 545 546 // GROUP_AFFINITY is already defined for _MSC_VER>=1600 (VS2010 and later). 547 #if KMP_OS_WINDOWS 548 #if _MSC_VER < 1600 549 typedef struct GROUP_AFFINITY { 550 KAFFINITY Mask; 551 WORD Group; 552 WORD Reserved[3]; 553 } GROUP_AFFINITY; 554 #endif /* _MSC_VER < 1600 */ 555 #if KMP_GROUP_AFFINITY 556 extern int __kmp_num_proc_groups; 557 #else 558 static const int __kmp_num_proc_groups = 1; 559 #endif /* KMP_GROUP_AFFINITY */ 560 typedef DWORD (*kmp_GetActiveProcessorCount_t)(WORD); 561 extern kmp_GetActiveProcessorCount_t __kmp_GetActiveProcessorCount; 562 563 typedef WORD (*kmp_GetActiveProcessorGroupCount_t)(void); 564 extern kmp_GetActiveProcessorGroupCount_t __kmp_GetActiveProcessorGroupCount; 565 566 typedef BOOL (*kmp_GetThreadGroupAffinity_t)(HANDLE, GROUP_AFFINITY *); 567 extern kmp_GetThreadGroupAffinity_t __kmp_GetThreadGroupAffinity; 568 569 typedef BOOL (*kmp_SetThreadGroupAffinity_t)(HANDLE, const GROUP_AFFINITY *, 570 GROUP_AFFINITY *); 571 extern kmp_SetThreadGroupAffinity_t __kmp_SetThreadGroupAffinity; 572 #endif /* KMP_OS_WINDOWS */ 573 574 #if KMP_USE_HWLOC 575 extern hwloc_topology_t __kmp_hwloc_topology; 576 extern int __kmp_hwloc_error; 577 extern int __kmp_numa_detected; 578 extern int __kmp_tile_depth; 579 #endif 580 581 extern size_t __kmp_affin_mask_size; 582 #define KMP_AFFINITY_CAPABLE() (__kmp_affin_mask_size > 0) 583 #define KMP_AFFINITY_DISABLE() (__kmp_affin_mask_size = 0) 584 #define KMP_AFFINITY_ENABLE(mask_size) (__kmp_affin_mask_size = mask_size) 585 #define KMP_CPU_SET_ITERATE(i, mask) \ 586 for (i = (mask)->begin(); i != (mask)->end(); i = (mask)->next(i)) 587 #define KMP_CPU_SET(i, mask) (mask)->set(i) 588 #define KMP_CPU_ISSET(i, mask) (mask)->is_set(i) 589 #define KMP_CPU_CLR(i, mask) (mask)->clear(i) 590 #define KMP_CPU_ZERO(mask) (mask)->zero() 591 #define KMP_CPU_COPY(dest, src) (dest)->copy(src) 592 #define KMP_CPU_AND(dest, src) (dest)->bitwise_and(src) 593 #define KMP_CPU_COMPLEMENT(max_bit_number, mask) (mask)->bitwise_not() 594 #define KMP_CPU_UNION(dest, src) (dest)->bitwise_or(src) 595 #define KMP_CPU_ALLOC(ptr) (ptr = __kmp_affinity_dispatch->allocate_mask()) 596 #define KMP_CPU_FREE(ptr) __kmp_affinity_dispatch->deallocate_mask(ptr) 597 #define KMP_CPU_ALLOC_ON_STACK(ptr) KMP_CPU_ALLOC(ptr) 598 #define KMP_CPU_FREE_FROM_STACK(ptr) KMP_CPU_FREE(ptr) 599 #define KMP_CPU_INTERNAL_ALLOC(ptr) KMP_CPU_ALLOC(ptr) 600 #define KMP_CPU_INTERNAL_FREE(ptr) KMP_CPU_FREE(ptr) 601 #define KMP_CPU_INDEX(arr, i) __kmp_affinity_dispatch->index_mask_array(arr, i) 602 #define KMP_CPU_ALLOC_ARRAY(arr, n) \ 603 (arr = __kmp_affinity_dispatch->allocate_mask_array(n)) 604 #define KMP_CPU_FREE_ARRAY(arr, n) \ 605 __kmp_affinity_dispatch->deallocate_mask_array(arr) 606 #define KMP_CPU_INTERNAL_ALLOC_ARRAY(arr, n) KMP_CPU_ALLOC_ARRAY(arr, n) 607 #define KMP_CPU_INTERNAL_FREE_ARRAY(arr, n) KMP_CPU_FREE_ARRAY(arr, n) 608 #define __kmp_get_system_affinity(mask, abort_bool) \ 609 (mask)->get_system_affinity(abort_bool) 610 #define __kmp_set_system_affinity(mask, abort_bool) \ 611 (mask)->set_system_affinity(abort_bool) 612 #define __kmp_get_proc_group(mask) (mask)->get_proc_group() 613 614 class KMPAffinity { 615 public: 616 class Mask { 617 public: 618 void *operator new(size_t n); 619 void operator delete(void *p); 620 void *operator new[](size_t n); 621 void operator delete[](void *p); 622 virtual ~Mask() {} 623 // Set bit i to 1 624 virtual void set(int i) {} 625 // Return bit i 626 virtual bool is_set(int i) const { return false; } 627 // Set bit i to 0 628 virtual void clear(int i) {} 629 // Zero out entire mask 630 virtual void zero() {} 631 // Copy src into this mask 632 virtual void copy(const Mask *src) {} 633 // this &= rhs 634 virtual void bitwise_and(const Mask *rhs) {} 635 // this |= rhs 636 virtual void bitwise_or(const Mask *rhs) {} 637 // this = ~this 638 virtual void bitwise_not() {} 639 // API for iterating over an affinity mask 640 // for (int i = mask->begin(); i != mask->end(); i = mask->next(i)) 641 virtual int begin() const { return 0; } 642 virtual int end() const { return 0; } 643 virtual int next(int previous) const { return 0; } 644 // Set the system's affinity to this affinity mask's value 645 virtual int set_system_affinity(bool abort_on_error) const { return -1; } 646 // Set this affinity mask to the current system affinity 647 virtual int get_system_affinity(bool abort_on_error) { return -1; } 648 // Only 1 DWORD in the mask should have any procs set. 649 // Return the appropriate index, or -1 for an invalid mask. 650 virtual int get_proc_group() const { return -1; } 651 }; 652 void *operator new(size_t n); 653 void operator delete(void *p); 654 // Need virtual destructor 655 virtual ~KMPAffinity() = default; 656 // Determine if affinity is capable 657 virtual void determine_capable(const char *env_var) {} 658 // Bind the current thread to os proc 659 virtual void bind_thread(int proc) {} 660 // Factory functions to allocate/deallocate a mask 661 virtual Mask *allocate_mask() { return nullptr; } 662 virtual void deallocate_mask(Mask *m) {} 663 virtual Mask *allocate_mask_array(int num) { return nullptr; } 664 virtual void deallocate_mask_array(Mask *m) {} 665 virtual Mask *index_mask_array(Mask *m, int index) { return nullptr; } 666 static void pick_api(); 667 static void destroy_api(); 668 enum api_type { 669 NATIVE_OS 670 #if KMP_USE_HWLOC 671 , 672 HWLOC 673 #endif 674 }; 675 virtual api_type get_api_type() const { 676 KMP_ASSERT(0); 677 return NATIVE_OS; 678 } 679 680 private: 681 static bool picked_api; 682 }; 683 684 typedef KMPAffinity::Mask kmp_affin_mask_t; 685 extern KMPAffinity *__kmp_affinity_dispatch; 686 687 // Declare local char buffers with this size for printing debug and info 688 // messages, using __kmp_affinity_print_mask(). 689 #define KMP_AFFIN_MASK_PRINT_LEN 1024 690 691 enum affinity_type { 692 affinity_none = 0, 693 affinity_physical, 694 affinity_logical, 695 affinity_compact, 696 affinity_scatter, 697 affinity_explicit, 698 affinity_balanced, 699 affinity_disabled, // not used outsize the env var parser 700 affinity_default 701 }; 702 703 enum affinity_gran { 704 affinity_gran_fine = 0, 705 affinity_gran_thread, 706 affinity_gran_core, 707 affinity_gran_tile, 708 affinity_gran_numa, 709 affinity_gran_package, 710 affinity_gran_node, 711 #if KMP_GROUP_AFFINITY 712 // The "group" granularity isn't necesssarily coarser than all of the 713 // other levels, but we put it last in the enum. 714 affinity_gran_group, 715 #endif /* KMP_GROUP_AFFINITY */ 716 affinity_gran_default 717 }; 718 719 enum affinity_top_method { 720 affinity_top_method_all = 0, // try all (supported) methods, in order 721 #if KMP_ARCH_X86 || KMP_ARCH_X86_64 722 affinity_top_method_apicid, 723 affinity_top_method_x2apicid, 724 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */ 725 affinity_top_method_cpuinfo, // KMP_CPUINFO_FILE is usable on Windows* OS, too 726 #if KMP_GROUP_AFFINITY 727 affinity_top_method_group, 728 #endif /* KMP_GROUP_AFFINITY */ 729 affinity_top_method_flat, 730 #if KMP_USE_HWLOC 731 affinity_top_method_hwloc, 732 #endif 733 affinity_top_method_default 734 }; 735 736 #define affinity_respect_mask_default (-1) 737 738 extern enum affinity_type __kmp_affinity_type; /* Affinity type */ 739 extern enum affinity_gran __kmp_affinity_gran; /* Affinity granularity */ 740 extern int __kmp_affinity_gran_levels; /* corresponding int value */ 741 extern int __kmp_affinity_dups; /* Affinity duplicate masks */ 742 extern enum affinity_top_method __kmp_affinity_top_method; 743 extern int __kmp_affinity_compact; /* Affinity 'compact' value */ 744 extern int __kmp_affinity_offset; /* Affinity offset value */ 745 extern int __kmp_affinity_verbose; /* Was verbose specified for KMP_AFFINITY? */ 746 extern int __kmp_affinity_warnings; /* KMP_AFFINITY warnings enabled ? */ 747 extern int __kmp_affinity_respect_mask; // Respect process' init affinity mask? 748 extern char *__kmp_affinity_proclist; /* proc ID list */ 749 extern kmp_affin_mask_t *__kmp_affinity_masks; 750 extern unsigned __kmp_affinity_num_masks; 751 extern void __kmp_affinity_bind_thread(int which); 752 753 extern kmp_affin_mask_t *__kmp_affin_fullMask; 754 extern char *__kmp_cpuinfo_file; 755 756 #endif /* KMP_AFFINITY_SUPPORTED */ 757 758 #if OMP_40_ENABLED 759 760 // This needs to be kept in sync with the values in omp.h !!! 761 typedef enum kmp_proc_bind_t { 762 proc_bind_false = 0, 763 proc_bind_true, 764 proc_bind_master, 765 proc_bind_close, 766 proc_bind_spread, 767 proc_bind_intel, // use KMP_AFFINITY interface 768 proc_bind_default 769 } kmp_proc_bind_t; 770 771 typedef struct kmp_nested_proc_bind_t { 772 kmp_proc_bind_t *bind_types; 773 int size; 774 int used; 775 } kmp_nested_proc_bind_t; 776 777 extern kmp_nested_proc_bind_t __kmp_nested_proc_bind; 778 779 #endif /* OMP_40_ENABLED */ 780 781 #if KMP_AFFINITY_SUPPORTED 782 #define KMP_PLACE_ALL (-1) 783 #define KMP_PLACE_UNDEFINED (-2) 784 #endif /* KMP_AFFINITY_SUPPORTED */ 785 786 extern int __kmp_affinity_num_places; 787 788 #if OMP_40_ENABLED 789 typedef enum kmp_cancel_kind_t { 790 cancel_noreq = 0, 791 cancel_parallel = 1, 792 cancel_loop = 2, 793 cancel_sections = 3, 794 cancel_taskgroup = 4 795 } kmp_cancel_kind_t; 796 #endif // OMP_40_ENABLED 797 798 // KMP_HW_SUBSET support: 799 typedef struct kmp_hws_item { 800 int num; 801 int offset; 802 } kmp_hws_item_t; 803 804 extern kmp_hws_item_t __kmp_hws_socket; 805 extern kmp_hws_item_t __kmp_hws_node; 806 extern kmp_hws_item_t __kmp_hws_tile; 807 extern kmp_hws_item_t __kmp_hws_core; 808 extern kmp_hws_item_t __kmp_hws_proc; 809 extern int __kmp_hws_requested; 810 extern int __kmp_hws_abs_flag; // absolute or per-item number requested 811 812 /* ------------------------------------------------------------------------ */ 813 814 #define KMP_PAD(type, sz) \ 815 (sizeof(type) + (sz - ((sizeof(type) - 1) % (sz)) - 1)) 816 817 // We need to avoid using -1 as a GTID as +1 is added to the gtid 818 // when storing it in a lock, and the value 0 is reserved. 819 #define KMP_GTID_DNE (-2) /* Does not exist */ 820 #define KMP_GTID_SHUTDOWN (-3) /* Library is shutting down */ 821 #define KMP_GTID_MONITOR (-4) /* Monitor thread ID */ 822 #define KMP_GTID_UNKNOWN (-5) /* Is not known */ 823 #define KMP_GTID_MIN (-6) /* Minimal gtid for low bound check in DEBUG */ 824 825 #define __kmp_get_gtid() __kmp_get_global_thread_id() 826 #define __kmp_entry_gtid() __kmp_get_global_thread_id_reg() 827 828 #define __kmp_tid_from_gtid(gtid) \ 829 (KMP_DEBUG_ASSERT((gtid) >= 0), __kmp_threads[(gtid)]->th.th_info.ds.ds_tid) 830 831 #define __kmp_get_tid() (__kmp_tid_from_gtid(__kmp_get_gtid())) 832 #define __kmp_gtid_from_tid(tid, team) \ 833 (KMP_DEBUG_ASSERT((tid) >= 0 && (team) != NULL), \ 834 team->t.t_threads[(tid)]->th.th_info.ds.ds_gtid) 835 836 #define __kmp_get_team() (__kmp_threads[(__kmp_get_gtid())]->th.th_team) 837 #define __kmp_team_from_gtid(gtid) \ 838 (KMP_DEBUG_ASSERT((gtid) >= 0), __kmp_threads[(gtid)]->th.th_team) 839 840 #define __kmp_thread_from_gtid(gtid) \ 841 (KMP_DEBUG_ASSERT((gtid) >= 0), __kmp_threads[(gtid)]) 842 #define __kmp_get_thread() (__kmp_thread_from_gtid(__kmp_get_gtid())) 843 844 // Returns current thread (pointer to kmp_info_t). In contrast to 845 // __kmp_get_thread(), it works with registered and not-yet-registered threads. 846 #define __kmp_gtid_from_thread(thr) \ 847 (KMP_DEBUG_ASSERT((thr) != NULL), (thr)->th.th_info.ds.ds_gtid) 848 849 // AT: Which way is correct? 850 // AT: 1. nproc = __kmp_threads[ ( gtid ) ] -> th.th_team -> t.t_nproc; 851 // AT: 2. nproc = __kmp_threads[ ( gtid ) ] -> th.th_team_nproc; 852 #define __kmp_get_team_num_threads(gtid) \ 853 (__kmp_threads[(gtid)]->th.th_team->t.t_nproc) 854 855 /* ------------------------------------------------------------------------ */ 856 857 #define KMP_UINT64_MAX \ 858 (~((kmp_uint64)1 << ((sizeof(kmp_uint64) * (1 << 3)) - 1))) 859 860 #define KMP_MIN_NTH 1 861 862 #ifndef KMP_MAX_NTH 863 #if defined(PTHREAD_THREADS_MAX) && PTHREAD_THREADS_MAX < INT_MAX 864 #define KMP_MAX_NTH PTHREAD_THREADS_MAX 865 #else 866 #define KMP_MAX_NTH INT_MAX 867 #endif 868 #endif /* KMP_MAX_NTH */ 869 870 #ifdef PTHREAD_STACK_MIN 871 #define KMP_MIN_STKSIZE PTHREAD_STACK_MIN 872 #else 873 #define KMP_MIN_STKSIZE ((size_t)(32 * 1024)) 874 #endif 875 876 #define KMP_MAX_STKSIZE (~((size_t)1 << ((sizeof(size_t) * (1 << 3)) - 1))) 877 878 #if KMP_ARCH_X86 879 #define KMP_DEFAULT_STKSIZE ((size_t)(2 * 1024 * 1024)) 880 #elif KMP_ARCH_X86_64 881 #define KMP_DEFAULT_STKSIZE ((size_t)(4 * 1024 * 1024)) 882 #define KMP_BACKUP_STKSIZE ((size_t)(2 * 1024 * 1024)) 883 #else 884 #define KMP_DEFAULT_STKSIZE ((size_t)(1024 * 1024)) 885 #endif 886 887 #define KMP_DEFAULT_MALLOC_POOL_INCR ((size_t)(1024 * 1024)) 888 #define KMP_MIN_MALLOC_POOL_INCR ((size_t)(4 * 1024)) 889 #define KMP_MAX_MALLOC_POOL_INCR \ 890 (~((size_t)1 << ((sizeof(size_t) * (1 << 3)) - 1))) 891 892 #define KMP_MIN_STKOFFSET (0) 893 #define KMP_MAX_STKOFFSET KMP_MAX_STKSIZE 894 #if KMP_OS_DARWIN 895 #define KMP_DEFAULT_STKOFFSET KMP_MIN_STKOFFSET 896 #else 897 #define KMP_DEFAULT_STKOFFSET CACHE_LINE 898 #endif 899 900 #define KMP_MIN_STKPADDING (0) 901 #define KMP_MAX_STKPADDING (2 * 1024 * 1024) 902 903 #define KMP_BLOCKTIME_MULTIPLIER \ 904 (1000) /* number of blocktime units per second */ 905 #define KMP_MIN_BLOCKTIME (0) 906 #define KMP_MAX_BLOCKTIME \ 907 (INT_MAX) /* Must be this for "infinite" setting the work */ 908 #define KMP_DEFAULT_BLOCKTIME (200) /* __kmp_blocktime is in milliseconds */ 909 910 #if KMP_USE_MONITOR 911 #define KMP_DEFAULT_MONITOR_STKSIZE ((size_t)(64 * 1024)) 912 #define KMP_MIN_MONITOR_WAKEUPS (1) // min times monitor wakes up per second 913 #define KMP_MAX_MONITOR_WAKEUPS (1000) // max times monitor can wake up per sec 914 915 /* Calculate new number of monitor wakeups for a specific block time based on 916 previous monitor_wakeups. Only allow increasing number of wakeups */ 917 #define KMP_WAKEUPS_FROM_BLOCKTIME(blocktime, monitor_wakeups) \ 918 (((blocktime) == KMP_MAX_BLOCKTIME) \ 919 ? (monitor_wakeups) \ 920 : ((blocktime) == KMP_MIN_BLOCKTIME) \ 921 ? KMP_MAX_MONITOR_WAKEUPS \ 922 : ((monitor_wakeups) > (KMP_BLOCKTIME_MULTIPLIER / (blocktime))) \ 923 ? (monitor_wakeups) \ 924 : (KMP_BLOCKTIME_MULTIPLIER) / (blocktime)) 925 926 /* Calculate number of intervals for a specific block time based on 927 monitor_wakeups */ 928 #define KMP_INTERVALS_FROM_BLOCKTIME(blocktime, monitor_wakeups) \ 929 (((blocktime) + (KMP_BLOCKTIME_MULTIPLIER / (monitor_wakeups)) - 1) / \ 930 (KMP_BLOCKTIME_MULTIPLIER / (monitor_wakeups))) 931 #else 932 #define KMP_BLOCKTIME(team, tid) \ 933 (get__bt_set(team, tid) ? get__blocktime(team, tid) : __kmp_dflt_blocktime) 934 #if KMP_OS_UNIX && (KMP_ARCH_X86 || KMP_ARCH_X86_64) 935 // HW TSC is used to reduce overhead (clock tick instead of nanosecond). 936 extern kmp_uint64 __kmp_ticks_per_msec; 937 #if KMP_COMPILER_ICC 938 #define KMP_NOW() _rdtsc() 939 #else 940 #define KMP_NOW() __kmp_hardware_timestamp() 941 #endif 942 #define KMP_NOW_MSEC() (KMP_NOW() / __kmp_ticks_per_msec) 943 #define KMP_BLOCKTIME_INTERVAL(team, tid) \ 944 (KMP_BLOCKTIME(team, tid) * __kmp_ticks_per_msec) 945 #define KMP_BLOCKING(goal, count) ((goal) > KMP_NOW()) 946 #else 947 // System time is retrieved sporadically while blocking. 948 extern kmp_uint64 __kmp_now_nsec(); 949 #define KMP_NOW() __kmp_now_nsec() 950 #define KMP_NOW_MSEC() (KMP_NOW() / KMP_USEC_PER_SEC) 951 #define KMP_BLOCKTIME_INTERVAL(team, tid) \ 952 (KMP_BLOCKTIME(team, tid) * KMP_USEC_PER_SEC) 953 #define KMP_BLOCKING(goal, count) ((count) % 1000 != 0 || (goal) > KMP_NOW()) 954 #endif 955 #define KMP_YIELD_NOW() \ 956 (KMP_NOW_MSEC() / KMP_MAX(__kmp_dflt_blocktime, 1) % \ 957 (__kmp_yield_on_count + __kmp_yield_off_count) < \ 958 (kmp_uint32)__kmp_yield_on_count) 959 #endif // KMP_USE_MONITOR 960 961 #define KMP_MIN_STATSCOLS 40 962 #define KMP_MAX_STATSCOLS 4096 963 #define KMP_DEFAULT_STATSCOLS 80 964 965 #define KMP_MIN_INTERVAL 0 966 #define KMP_MAX_INTERVAL (INT_MAX - 1) 967 #define KMP_DEFAULT_INTERVAL 0 968 969 #define KMP_MIN_CHUNK 1 970 #define KMP_MAX_CHUNK (INT_MAX - 1) 971 #define KMP_DEFAULT_CHUNK 1 972 973 #define KMP_MIN_INIT_WAIT 1 974 #define KMP_MAX_INIT_WAIT (INT_MAX / 2) 975 #define KMP_DEFAULT_INIT_WAIT 2048U 976 977 #define KMP_MIN_NEXT_WAIT 1 978 #define KMP_MAX_NEXT_WAIT (INT_MAX / 2) 979 #define KMP_DEFAULT_NEXT_WAIT 1024U 980 981 #define KMP_DFLT_DISP_NUM_BUFF 7 982 #define KMP_MAX_ORDERED 8 983 984 #define KMP_MAX_FIELDS 32 985 986 #define KMP_MAX_BRANCH_BITS 31 987 988 #define KMP_MAX_ACTIVE_LEVELS_LIMIT INT_MAX 989 990 #define KMP_MAX_DEFAULT_DEVICE_LIMIT INT_MAX 991 992 #define KMP_MAX_TASK_PRIORITY_LIMIT INT_MAX 993 994 /* Minimum number of threads before switch to TLS gtid (experimentally 995 determined) */ 996 /* josh TODO: what about OS X* tuning? */ 997 #if KMP_ARCH_X86 || KMP_ARCH_X86_64 998 #define KMP_TLS_GTID_MIN 5 999 #else 1000 #define KMP_TLS_GTID_MIN INT_MAX 1001 #endif 1002 1003 #define KMP_MASTER_TID(tid) ((tid) == 0) 1004 #define KMP_WORKER_TID(tid) ((tid) != 0) 1005 1006 #define KMP_MASTER_GTID(gtid) (__kmp_tid_from_gtid((gtid)) == 0) 1007 #define KMP_WORKER_GTID(gtid) (__kmp_tid_from_gtid((gtid)) != 0) 1008 #define KMP_UBER_GTID(gtid) \ 1009 (KMP_DEBUG_ASSERT((gtid) >= KMP_GTID_MIN), \ 1010 KMP_DEBUG_ASSERT((gtid) < __kmp_threads_capacity), \ 1011 (gtid) >= 0 && __kmp_root[(gtid)] && __kmp_threads[(gtid)] && \ 1012 (__kmp_threads[(gtid)] == __kmp_root[(gtid)]->r.r_uber_thread)) 1013 #define KMP_INITIAL_GTID(gtid) ((gtid) == 0) 1014 1015 #ifndef TRUE 1016 #define FALSE 0 1017 #define TRUE (!FALSE) 1018 #endif 1019 1020 /* NOTE: all of the following constants must be even */ 1021 1022 #if KMP_OS_WINDOWS 1023 #define KMP_INIT_WAIT 64U /* initial number of spin-tests */ 1024 #define KMP_NEXT_WAIT 32U /* susequent number of spin-tests */ 1025 #elif KMP_OS_CNK 1026 #define KMP_INIT_WAIT 16U /* initial number of spin-tests */ 1027 #define KMP_NEXT_WAIT 8U /* susequent number of spin-tests */ 1028 #elif KMP_OS_LINUX 1029 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */ 1030 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */ 1031 #elif KMP_OS_DARWIN 1032 /* TODO: tune for KMP_OS_DARWIN */ 1033 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */ 1034 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */ 1035 #elif KMP_OS_FREEBSD 1036 /* TODO: tune for KMP_OS_FREEBSD */ 1037 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */ 1038 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */ 1039 #elif KMP_OS_NETBSD 1040 /* TODO: tune for KMP_OS_NETBSD */ 1041 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */ 1042 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */ 1043 #endif 1044 1045 #if KMP_ARCH_X86 || KMP_ARCH_X86_64 1046 typedef struct kmp_cpuid { 1047 kmp_uint32 eax; 1048 kmp_uint32 ebx; 1049 kmp_uint32 ecx; 1050 kmp_uint32 edx; 1051 } kmp_cpuid_t; 1052 extern void __kmp_x86_cpuid(int mode, int mode2, struct kmp_cpuid *p); 1053 #if KMP_ARCH_X86 1054 extern void __kmp_x86_pause(void); 1055 #elif KMP_MIC 1056 // Performance testing on KNC (C0QS-7120 P/A/X/D, 61-core, 16 GB Memory) showed 1057 // regression after removal of extra PAUSE from KMP_YIELD_SPIN(). Changing 1058 // the delay from 100 to 300 showed even better performance than double PAUSE 1059 // on Spec OMP2001 and LCPC tasking tests, no regressions on EPCC. 1060 static void __kmp_x86_pause(void) { _mm_delay_32(300); } 1061 #else 1062 static void __kmp_x86_pause(void) { _mm_pause(); } 1063 #endif 1064 #define KMP_CPU_PAUSE() __kmp_x86_pause() 1065 #elif KMP_ARCH_PPC64 1066 #define KMP_PPC64_PRI_LOW() __asm__ volatile("or 1, 1, 1") 1067 #define KMP_PPC64_PRI_MED() __asm__ volatile("or 2, 2, 2") 1068 #define KMP_PPC64_PRI_LOC_MB() __asm__ volatile("" : : : "memory") 1069 #define KMP_CPU_PAUSE() \ 1070 do { \ 1071 KMP_PPC64_PRI_LOW(); \ 1072 KMP_PPC64_PRI_MED(); \ 1073 KMP_PPC64_PRI_LOC_MB(); \ 1074 } while (0) 1075 #else 1076 #define KMP_CPU_PAUSE() /* nothing to do */ 1077 #endif 1078 1079 #define KMP_INIT_YIELD(count) \ 1080 { (count) = __kmp_yield_init; } 1081 1082 #define KMP_YIELD(cond) \ 1083 { \ 1084 KMP_CPU_PAUSE(); \ 1085 __kmp_yield((cond)); \ 1086 } 1087 1088 // Note the decrement of 2 in the following Macros. With KMP_LIBRARY=turnaround, 1089 // there should be no yielding since initial value from KMP_INIT_YIELD() is odd. 1090 1091 #define KMP_YIELD_WHEN(cond, count) \ 1092 { \ 1093 KMP_CPU_PAUSE(); \ 1094 (count) -= 2; \ 1095 if (!(count)) { \ 1096 __kmp_yield(cond); \ 1097 (count) = __kmp_yield_next; \ 1098 } \ 1099 } 1100 #define KMP_YIELD_SPIN(count) \ 1101 { \ 1102 KMP_CPU_PAUSE(); \ 1103 (count) -= 2; \ 1104 if (!(count)) { \ 1105 __kmp_yield(1); \ 1106 (count) = __kmp_yield_next; \ 1107 } \ 1108 } 1109 1110 /* ------------------------------------------------------------------------ */ 1111 /* Support datatypes for the orphaned construct nesting checks. */ 1112 /* ------------------------------------------------------------------------ */ 1113 1114 enum cons_type { 1115 ct_none, 1116 ct_parallel, 1117 ct_pdo, 1118 ct_pdo_ordered, 1119 ct_psections, 1120 ct_psingle, 1121 1122 /* the following must be left in order and not split up */ 1123 ct_taskq, 1124 ct_task, // really task inside non-ordered taskq, considered worksharing type 1125 ct_task_ordered, /* really task inside ordered taskq, considered a worksharing 1126 type */ 1127 /* the preceding must be left in order and not split up */ 1128 1129 ct_critical, 1130 ct_ordered_in_parallel, 1131 ct_ordered_in_pdo, 1132 ct_ordered_in_taskq, 1133 ct_master, 1134 ct_reduce, 1135 ct_barrier 1136 }; 1137 1138 /* test to see if we are in a taskq construct */ 1139 #define IS_CONS_TYPE_TASKQ(ct) \ 1140 (((int)(ct)) >= ((int)ct_taskq) && ((int)(ct)) <= ((int)ct_task_ordered)) 1141 #define IS_CONS_TYPE_ORDERED(ct) \ 1142 ((ct) == ct_pdo_ordered || (ct) == ct_task_ordered) 1143 1144 struct cons_data { 1145 ident_t const *ident; 1146 enum cons_type type; 1147 int prev; 1148 kmp_user_lock_p 1149 name; /* address exclusively for critical section name comparison */ 1150 }; 1151 1152 struct cons_header { 1153 int p_top, w_top, s_top; 1154 int stack_size, stack_top; 1155 struct cons_data *stack_data; 1156 }; 1157 1158 struct kmp_region_info { 1159 char *text; 1160 int offset[KMP_MAX_FIELDS]; 1161 int length[KMP_MAX_FIELDS]; 1162 }; 1163 1164 /* ---------------------------------------------------------------------- */ 1165 /* ---------------------------------------------------------------------- */ 1166 1167 #if KMP_OS_WINDOWS 1168 typedef HANDLE kmp_thread_t; 1169 typedef DWORD kmp_key_t; 1170 #endif /* KMP_OS_WINDOWS */ 1171 1172 #if KMP_OS_UNIX 1173 typedef pthread_t kmp_thread_t; 1174 typedef pthread_key_t kmp_key_t; 1175 #endif 1176 1177 extern kmp_key_t __kmp_gtid_threadprivate_key; 1178 1179 typedef struct kmp_sys_info { 1180 long maxrss; /* the maximum resident set size utilized (in kilobytes) */ 1181 long minflt; /* the number of page faults serviced without any I/O */ 1182 long majflt; /* the number of page faults serviced that required I/O */ 1183 long nswap; /* the number of times a process was "swapped" out of memory */ 1184 long inblock; /* the number of times the file system had to perform input */ 1185 long oublock; /* the number of times the file system had to perform output */ 1186 long nvcsw; /* the number of times a context switch was voluntarily */ 1187 long nivcsw; /* the number of times a context switch was forced */ 1188 } kmp_sys_info_t; 1189 1190 #if KMP_ARCH_X86 || KMP_ARCH_X86_64 1191 typedef struct kmp_cpuinfo { 1192 int initialized; // If 0, other fields are not initialized. 1193 int signature; // CPUID(1).EAX 1194 int family; // CPUID(1).EAX[27:20]+CPUID(1).EAX[11:8] (Extended Family+Family) 1195 int model; // ( CPUID(1).EAX[19:16] << 4 ) + CPUID(1).EAX[7:4] ( ( Extended 1196 // Model << 4 ) + Model) 1197 int stepping; // CPUID(1).EAX[3:0] ( Stepping ) 1198 int sse2; // 0 if SSE2 instructions are not supported, 1 otherwise. 1199 int rtm; // 0 if RTM instructions are not supported, 1 otherwise. 1200 int cpu_stackoffset; 1201 int apic_id; 1202 int physical_id; 1203 int logical_id; 1204 kmp_uint64 frequency; // Nominal CPU frequency in Hz. 1205 char name[3 * sizeof(kmp_cpuid_t)]; // CPUID(0x80000002,0x80000003,0x80000004) 1206 } kmp_cpuinfo_t; 1207 #endif 1208 1209 #if USE_ITT_BUILD 1210 // We cannot include "kmp_itt.h" due to circular dependency. Declare the only 1211 // required type here. Later we will check the type meets requirements. 1212 typedef int kmp_itt_mark_t; 1213 #define KMP_ITT_DEBUG 0 1214 #endif /* USE_ITT_BUILD */ 1215 1216 /* Taskq data structures */ 1217 1218 #define HIGH_WATER_MARK(nslots) (((nslots)*3) / 4) 1219 // num thunks that each thread can simultaneously execute from a task queue 1220 #define __KMP_TASKQ_THUNKS_PER_TH 1 1221 1222 /* flags for taskq_global_flags, kmp_task_queue_t tq_flags, kmpc_thunk_t 1223 th_flags */ 1224 1225 #define TQF_IS_ORDERED 0x0001 // __kmpc_taskq interface, taskq ordered 1226 // __kmpc_taskq interface, taskq with lastprivate list 1227 #define TQF_IS_LASTPRIVATE 0x0002 1228 #define TQF_IS_NOWAIT 0x0004 // __kmpc_taskq interface, end taskq nowait 1229 // __kmpc_taskq interface, use heuristics to decide task queue size 1230 #define TQF_HEURISTICS 0x0008 1231 1232 // __kmpc_taskq interface, reserved for future use 1233 #define TQF_INTERFACE_RESERVED1 0x0010 1234 // __kmpc_taskq interface, reserved for future use 1235 #define TQF_INTERFACE_RESERVED2 0x0020 1236 // __kmpc_taskq interface, reserved for future use 1237 #define TQF_INTERFACE_RESERVED3 0x0040 1238 // __kmpc_taskq interface, reserved for future use 1239 #define TQF_INTERFACE_RESERVED4 0x0080 1240 1241 #define TQF_INTERFACE_FLAGS 0x00ff // all the __kmpc_taskq interface flags 1242 // internal/read by instrumentation; only used with TQF_IS_LASTPRIVATE 1243 #define TQF_IS_LAST_TASK 0x0100 1244 // internal use only; this thunk->th_task is the taskq_task 1245 #define TQF_TASKQ_TASK 0x0200 1246 // internal use only; must release worker threads once ANY queued task 1247 // exists (global) 1248 #define TQF_RELEASE_WORKERS 0x0400 1249 // internal use only; notify workers that master has finished enqueuing tasks 1250 #define TQF_ALL_TASKS_QUEUED 0x0800 1251 // internal use only: this queue encountered in parallel context: not serialized 1252 #define TQF_PARALLEL_CONTEXT 0x1000 1253 // internal use only; this queue is on the freelist and not in use 1254 #define TQF_DEALLOCATED 0x2000 1255 1256 #define TQF_INTERNAL_FLAGS 0x3f00 // all the internal use only flags 1257 1258 typedef struct KMP_ALIGN_CACHE kmpc_aligned_int32_t { 1259 kmp_int32 ai_data; 1260 } kmpc_aligned_int32_t; 1261 1262 typedef struct KMP_ALIGN_CACHE kmpc_aligned_queue_slot_t { 1263 struct kmpc_thunk_t *qs_thunk; 1264 } kmpc_aligned_queue_slot_t; 1265 1266 typedef struct kmpc_task_queue_t { 1267 /* task queue linkage fields for n-ary tree of queues (locked with global 1268 taskq_tree_lck) */ 1269 kmp_lock_t tq_link_lck; /* lock for child link, child next/prev links and 1270 child ref counts */ 1271 union { 1272 struct kmpc_task_queue_t *tq_parent; // pointer to parent taskq, not locked 1273 // for taskq internal freelists, locked with global taskq_freelist_lck 1274 struct kmpc_task_queue_t *tq_next_free; 1275 } tq; 1276 // pointer to linked-list of children, locked by tq's tq_link_lck 1277 volatile struct kmpc_task_queue_t *tq_first_child; 1278 // next child in linked-list, locked by parent tq's tq_link_lck 1279 struct kmpc_task_queue_t *tq_next_child; 1280 // previous child in linked-list, locked by parent tq's tq_link_lck 1281 struct kmpc_task_queue_t *tq_prev_child; 1282 // reference count of threads with access to this task queue 1283 volatile kmp_int32 tq_ref_count; 1284 /* (other than the thread executing the kmpc_end_taskq call) */ 1285 /* locked by parent tq's tq_link_lck */ 1286 1287 /* shared data for task queue */ 1288 /* per-thread array of pointers to shared variable structures */ 1289 struct kmpc_aligned_shared_vars_t *tq_shareds; 1290 /* only one array element exists for all but outermost taskq */ 1291 1292 /* bookkeeping for ordered task queue */ 1293 kmp_uint32 tq_tasknum_queuing; // ordered task # assigned while queuing tasks 1294 // ordered number of next task to be served (executed) 1295 volatile kmp_uint32 tq_tasknum_serving; 1296 1297 /* thunk storage management for task queue */ 1298 kmp_lock_t tq_free_thunks_lck; /* lock for thunk freelist manipulation */ 1299 // thunk freelist, chained via th.th_next_free 1300 struct kmpc_thunk_t *tq_free_thunks; 1301 // space allocated for thunks for this task queue 1302 struct kmpc_thunk_t *tq_thunk_space; 1303 1304 /* data fields for queue itself */ 1305 kmp_lock_t tq_queue_lck; /* lock for [de]enqueue operations: tq_queue, 1306 tq_head, tq_tail, tq_nfull */ 1307 /* array of queue slots to hold thunks for tasks */ 1308 kmpc_aligned_queue_slot_t *tq_queue; 1309 volatile struct kmpc_thunk_t *tq_taskq_slot; /* special slot for taskq task 1310 thunk, occupied if not NULL */ 1311 kmp_int32 tq_nslots; /* # of tq_thunk_space thunks alloc'd (not incl. 1312 tq_taskq_slot space) */ 1313 kmp_int32 tq_head; // enqueue puts item here (index into tq_queue array) 1314 kmp_int32 tq_tail; // dequeue takes item from here (index into tq_queue array) 1315 volatile kmp_int32 tq_nfull; // # of occupied entries in task queue right now 1316 kmp_int32 tq_hiwat; /* high-water mark for tq_nfull and queue scheduling */ 1317 volatile kmp_int32 tq_flags; /* TQF_xxx */ 1318 1319 /* bookkeeping for outstanding thunks */ 1320 1321 /* per-thread array for # of regular thunks currently being executed */ 1322 struct kmpc_aligned_int32_t *tq_th_thunks; 1323 kmp_int32 tq_nproc; /* number of thunks in the th_thunks array */ 1324 1325 /* statistics library bookkeeping */ 1326 ident_t *tq_loc; /* source location information for taskq directive */ 1327 } kmpc_task_queue_t; 1328 1329 typedef void (*kmpc_task_t)(kmp_int32 global_tid, struct kmpc_thunk_t *thunk); 1330 1331 /* sizeof_shareds passed as arg to __kmpc_taskq call */ 1332 typedef struct kmpc_shared_vars_t { /* aligned during dynamic allocation */ 1333 kmpc_task_queue_t *sv_queue; /* (pointers to) shared vars */ 1334 } kmpc_shared_vars_t; 1335 1336 typedef struct KMP_ALIGN_CACHE kmpc_aligned_shared_vars_t { 1337 volatile struct kmpc_shared_vars_t *ai_data; 1338 } kmpc_aligned_shared_vars_t; 1339 1340 /* sizeof_thunk passed as arg to kmpc_taskq call */ 1341 typedef struct kmpc_thunk_t { /* aligned during dynamic allocation */ 1342 union { /* field used for internal freelists too */ 1343 kmpc_shared_vars_t *th_shareds; 1344 struct kmpc_thunk_t *th_next_free; /* freelist of individual thunks within 1345 queue, head at tq_free_thunks */ 1346 } th; 1347 kmpc_task_t th_task; /* taskq_task if flags & TQF_TASKQ_TASK */ 1348 struct kmpc_thunk_t *th_encl_thunk; /* pointer to dynamically enclosing thunk 1349 on this thread's call stack */ 1350 // TQF_xxx(tq_flags interface plus possible internal flags) 1351 kmp_int32 th_flags; 1352 1353 kmp_int32 th_status; 1354 kmp_uint32 th_tasknum; /* task number assigned in order of queuing, used for 1355 ordered sections */ 1356 /* private vars */ 1357 } kmpc_thunk_t; 1358 1359 typedef struct KMP_ALIGN_CACHE kmp_taskq { 1360 int tq_curr_thunk_capacity; 1361 1362 kmpc_task_queue_t *tq_root; 1363 kmp_int32 tq_global_flags; 1364 1365 kmp_lock_t tq_freelist_lck; 1366 kmpc_task_queue_t *tq_freelist; 1367 1368 kmpc_thunk_t **tq_curr_thunk; 1369 } kmp_taskq_t; 1370 1371 /* END Taskq data structures */ 1372 1373 typedef kmp_int32 kmp_critical_name[8]; 1374 1375 /*! 1376 @ingroup PARALLEL 1377 The type for a microtask which gets passed to @ref __kmpc_fork_call(). 1378 The arguments to the outlined function are 1379 @param global_tid the global thread identity of the thread executing the 1380 function. 1381 @param bound_tid the local identitiy of the thread executing the function 1382 @param ... pointers to shared variables accessed by the function. 1383 */ 1384 typedef void (*kmpc_micro)(kmp_int32 *global_tid, kmp_int32 *bound_tid, ...); 1385 typedef void (*kmpc_micro_bound)(kmp_int32 *bound_tid, kmp_int32 *bound_nth, 1386 ...); 1387 1388 /*! 1389 @ingroup THREADPRIVATE 1390 @{ 1391 */ 1392 /* --------------------------------------------------------------------------- 1393 */ 1394 /* Threadprivate initialization/finalization function declarations */ 1395 1396 /* for non-array objects: __kmpc_threadprivate_register() */ 1397 1398 /*! 1399 Pointer to the constructor function. 1400 The first argument is the <tt>this</tt> pointer 1401 */ 1402 typedef void *(*kmpc_ctor)(void *); 1403 1404 /*! 1405 Pointer to the destructor function. 1406 The first argument is the <tt>this</tt> pointer 1407 */ 1408 typedef void (*kmpc_dtor)( 1409 void * /*, size_t */); /* 2nd arg: magic number for KCC unused by Intel 1410 compiler */ 1411 /*! 1412 Pointer to an alternate constructor. 1413 The first argument is the <tt>this</tt> pointer. 1414 */ 1415 typedef void *(*kmpc_cctor)(void *, void *); 1416 1417 /* for array objects: __kmpc_threadprivate_register_vec() */ 1418 /* First arg: "this" pointer */ 1419 /* Last arg: number of array elements */ 1420 /*! 1421 Array constructor. 1422 First argument is the <tt>this</tt> pointer 1423 Second argument the number of array elements. 1424 */ 1425 typedef void *(*kmpc_ctor_vec)(void *, size_t); 1426 /*! 1427 Pointer to the array destructor function. 1428 The first argument is the <tt>this</tt> pointer 1429 Second argument the number of array elements. 1430 */ 1431 typedef void (*kmpc_dtor_vec)(void *, size_t); 1432 /*! 1433 Array constructor. 1434 First argument is the <tt>this</tt> pointer 1435 Third argument the number of array elements. 1436 */ 1437 typedef void *(*kmpc_cctor_vec)(void *, void *, 1438 size_t); /* function unused by compiler */ 1439 1440 /*! 1441 @} 1442 */ 1443 1444 /* keeps tracked of threadprivate cache allocations for cleanup later */ 1445 typedef struct kmp_cached_addr { 1446 void **addr; /* address of allocated cache */ 1447 struct kmp_cached_addr *next; /* pointer to next cached address */ 1448 } kmp_cached_addr_t; 1449 1450 struct private_data { 1451 struct private_data *next; /* The next descriptor in the list */ 1452 void *data; /* The data buffer for this descriptor */ 1453 int more; /* The repeat count for this descriptor */ 1454 size_t size; /* The data size for this descriptor */ 1455 }; 1456 1457 struct private_common { 1458 struct private_common *next; 1459 struct private_common *link; 1460 void *gbl_addr; 1461 void *par_addr; /* par_addr == gbl_addr for MASTER thread */ 1462 size_t cmn_size; 1463 }; 1464 1465 struct shared_common { 1466 struct shared_common *next; 1467 struct private_data *pod_init; 1468 void *obj_init; 1469 void *gbl_addr; 1470 union { 1471 kmpc_ctor ctor; 1472 kmpc_ctor_vec ctorv; 1473 } ct; 1474 union { 1475 kmpc_cctor cctor; 1476 kmpc_cctor_vec cctorv; 1477 } cct; 1478 union { 1479 kmpc_dtor dtor; 1480 kmpc_dtor_vec dtorv; 1481 } dt; 1482 size_t vec_len; 1483 int is_vec; 1484 size_t cmn_size; 1485 }; 1486 1487 #define KMP_HASH_TABLE_LOG2 9 /* log2 of the hash table size */ 1488 #define KMP_HASH_TABLE_SIZE \ 1489 (1 << KMP_HASH_TABLE_LOG2) /* size of the hash table */ 1490 #define KMP_HASH_SHIFT 3 /* throw away this many low bits from the address */ 1491 #define KMP_HASH(x) \ 1492 ((((kmp_uintptr_t)x) >> KMP_HASH_SHIFT) & (KMP_HASH_TABLE_SIZE - 1)) 1493 1494 struct common_table { 1495 struct private_common *data[KMP_HASH_TABLE_SIZE]; 1496 }; 1497 1498 struct shared_table { 1499 struct shared_common *data[KMP_HASH_TABLE_SIZE]; 1500 }; 1501 1502 /* ------------------------------------------------------------------------ */ 1503 1504 #if KMP_STATIC_STEAL_ENABLED 1505 typedef struct KMP_ALIGN_CACHE dispatch_private_info32 { 1506 kmp_int32 count; 1507 kmp_int32 ub; 1508 /* Adding KMP_ALIGN_CACHE here doesn't help / can hurt performance */ 1509 kmp_int32 lb; 1510 kmp_int32 st; 1511 kmp_int32 tc; 1512 kmp_int32 static_steal_counter; /* for static_steal only; maybe better to put 1513 after ub */ 1514 1515 // KMP_ALIGN( 16 ) ensures ( if the KMP_ALIGN macro is turned on ) 1516 // a) parm3 is properly aligned and 1517 // b) all parm1-4 are in the same cache line. 1518 // Because of parm1-4 are used together, performance seems to be better 1519 // if they are in the same line (not measured though). 1520 1521 struct KMP_ALIGN(32) { // AC: changed 16 to 32 in order to simplify template 1522 kmp_int32 parm1; // structures in kmp_dispatch.cpp. This should 1523 kmp_int32 parm2; // make no real change at least while padding is off. 1524 kmp_int32 parm3; 1525 kmp_int32 parm4; 1526 }; 1527 1528 kmp_uint32 ordered_lower; 1529 kmp_uint32 ordered_upper; 1530 #if KMP_OS_WINDOWS 1531 // This var can be placed in the hole between 'tc' and 'parm1', instead of 1532 // 'static_steal_counter'. It would be nice to measure execution times. 1533 // Conditional if/endif can be removed at all. 1534 kmp_int32 last_upper; 1535 #endif /* KMP_OS_WINDOWS */ 1536 } dispatch_private_info32_t; 1537 1538 typedef struct KMP_ALIGN_CACHE dispatch_private_info64 { 1539 kmp_int64 count; // current chunk number for static & static-steal scheduling 1540 kmp_int64 ub; /* upper-bound */ 1541 /* Adding KMP_ALIGN_CACHE here doesn't help / can hurt performance */ 1542 kmp_int64 lb; /* lower-bound */ 1543 kmp_int64 st; /* stride */ 1544 kmp_int64 tc; /* trip count (number of iterations) */ 1545 kmp_int64 static_steal_counter; /* for static_steal only; maybe better to put 1546 after ub */ 1547 1548 /* parm[1-4] are used in different ways by different scheduling algorithms */ 1549 1550 // KMP_ALIGN( 32 ) ensures ( if the KMP_ALIGN macro is turned on ) 1551 // a) parm3 is properly aligned and 1552 // b) all parm1-4 are in the same cache line. 1553 // Because of parm1-4 are used together, performance seems to be better 1554 // if they are in the same line (not measured though). 1555 1556 struct KMP_ALIGN(32) { 1557 kmp_int64 parm1; 1558 kmp_int64 parm2; 1559 kmp_int64 parm3; 1560 kmp_int64 parm4; 1561 }; 1562 1563 kmp_uint64 ordered_lower; 1564 kmp_uint64 ordered_upper; 1565 #if KMP_OS_WINDOWS 1566 // This var can be placed in the hole between 'tc' and 'parm1', instead of 1567 // 'static_steal_counter'. It would be nice to measure execution times. 1568 // Conditional if/endif can be removed at all. 1569 kmp_int64 last_upper; 1570 #endif /* KMP_OS_WINDOWS */ 1571 } dispatch_private_info64_t; 1572 #else /* KMP_STATIC_STEAL_ENABLED */ 1573 typedef struct KMP_ALIGN_CACHE dispatch_private_info32 { 1574 kmp_int32 lb; 1575 kmp_int32 ub; 1576 kmp_int32 st; 1577 kmp_int32 tc; 1578 1579 kmp_int32 parm1; 1580 kmp_int32 parm2; 1581 kmp_int32 parm3; 1582 kmp_int32 parm4; 1583 1584 kmp_int32 count; 1585 1586 kmp_uint32 ordered_lower; 1587 kmp_uint32 ordered_upper; 1588 #if KMP_OS_WINDOWS 1589 kmp_int32 last_upper; 1590 #endif /* KMP_OS_WINDOWS */ 1591 } dispatch_private_info32_t; 1592 1593 typedef struct KMP_ALIGN_CACHE dispatch_private_info64 { 1594 kmp_int64 lb; /* lower-bound */ 1595 kmp_int64 ub; /* upper-bound */ 1596 kmp_int64 st; /* stride */ 1597 kmp_int64 tc; /* trip count (number of iterations) */ 1598 1599 /* parm[1-4] are used in different ways by different scheduling algorithms */ 1600 kmp_int64 parm1; 1601 kmp_int64 parm2; 1602 kmp_int64 parm3; 1603 kmp_int64 parm4; 1604 1605 kmp_int64 count; /* current chunk number for static scheduling */ 1606 1607 kmp_uint64 ordered_lower; 1608 kmp_uint64 ordered_upper; 1609 #if KMP_OS_WINDOWS 1610 kmp_int64 last_upper; 1611 #endif /* KMP_OS_WINDOWS */ 1612 } dispatch_private_info64_t; 1613 #endif /* KMP_STATIC_STEAL_ENABLED */ 1614 1615 typedef struct KMP_ALIGN_CACHE dispatch_private_info { 1616 union private_info { 1617 dispatch_private_info32_t p32; 1618 dispatch_private_info64_t p64; 1619 } u; 1620 enum sched_type schedule; /* scheduling algorithm */ 1621 kmp_int32 ordered; /* ordered clause specified */ 1622 kmp_int32 ordered_bumped; 1623 // To retain the structure size after making ordered_iteration scalar 1624 kmp_int32 ordered_dummy[KMP_MAX_ORDERED - 3]; 1625 // Stack of buffers for nest of serial regions 1626 struct dispatch_private_info *next; 1627 kmp_int32 nomerge; /* don't merge iters if serialized */ 1628 kmp_int32 type_size; /* the size of types in private_info */ 1629 enum cons_type pushed_ws; 1630 } dispatch_private_info_t; 1631 1632 typedef struct dispatch_shared_info32 { 1633 /* chunk index under dynamic, number of idle threads under static-steal; 1634 iteration index otherwise */ 1635 volatile kmp_uint32 iteration; 1636 volatile kmp_uint32 num_done; 1637 volatile kmp_uint32 ordered_iteration; 1638 // Dummy to retain the structure size after making ordered_iteration scalar 1639 kmp_int32 ordered_dummy[KMP_MAX_ORDERED - 1]; 1640 } dispatch_shared_info32_t; 1641 1642 typedef struct dispatch_shared_info64 { 1643 /* chunk index under dynamic, number of idle threads under static-steal; 1644 iteration index otherwise */ 1645 volatile kmp_uint64 iteration; 1646 volatile kmp_uint64 num_done; 1647 volatile kmp_uint64 ordered_iteration; 1648 // Dummy to retain the structure size after making ordered_iteration scalar 1649 kmp_int64 ordered_dummy[KMP_MAX_ORDERED - 3]; 1650 } dispatch_shared_info64_t; 1651 1652 typedef struct dispatch_shared_info { 1653 union shared_info { 1654 dispatch_shared_info32_t s32; 1655 dispatch_shared_info64_t s64; 1656 } u; 1657 volatile kmp_uint32 buffer_index; 1658 #if OMP_45_ENABLED 1659 volatile kmp_int32 doacross_buf_idx; // teamwise index 1660 volatile kmp_uint32 *doacross_flags; // shared array of iteration flags (0/1) 1661 kmp_int32 doacross_num_done; // count finished threads 1662 #endif 1663 #if KMP_USE_HWLOC 1664 // When linking with libhwloc, the ORDERED EPCC test slows down on big 1665 // machines (> 48 cores). Performance analysis showed that a cache thrash 1666 // was occurring and this padding helps alleviate the problem. 1667 char padding[64]; 1668 #endif 1669 } dispatch_shared_info_t; 1670 1671 typedef struct kmp_disp { 1672 /* Vector for ORDERED SECTION */ 1673 void (*th_deo_fcn)(int *gtid, int *cid, ident_t *); 1674 /* Vector for END ORDERED SECTION */ 1675 void (*th_dxo_fcn)(int *gtid, int *cid, ident_t *); 1676 1677 dispatch_shared_info_t *th_dispatch_sh_current; 1678 dispatch_private_info_t *th_dispatch_pr_current; 1679 1680 dispatch_private_info_t *th_disp_buffer; 1681 kmp_int32 th_disp_index; 1682 #if OMP_45_ENABLED 1683 kmp_int32 th_doacross_buf_idx; // thread's doacross buffer index 1684 volatile kmp_uint32 *th_doacross_flags; // pointer to shared array of flags 1685 union { // we can use union here because doacross cannot be used in 1686 // nonmonotonic loops 1687 kmp_int64 *th_doacross_info; // info on loop bounds 1688 kmp_lock_t *th_steal_lock; // lock used for chunk stealing (8-byte variable) 1689 }; 1690 #else 1691 #if KMP_STATIC_STEAL_ENABLED 1692 kmp_lock_t *th_steal_lock; // lock used for chunk stealing (8-byte variable) 1693 void *dummy_padding[1]; // make it 64 bytes on Intel(R) 64 1694 #else 1695 void *dummy_padding[2]; // make it 64 bytes on Intel(R) 64 1696 #endif 1697 #endif 1698 #if KMP_USE_INTERNODE_ALIGNMENT 1699 char more_padding[INTERNODE_CACHE_LINE]; 1700 #endif 1701 } kmp_disp_t; 1702 1703 /* ------------------------------------------------------------------------ */ 1704 /* Barrier stuff */ 1705 1706 /* constants for barrier state update */ 1707 #define KMP_INIT_BARRIER_STATE 0 /* should probably start from zero */ 1708 #define KMP_BARRIER_SLEEP_BIT 0 /* bit used for suspend/sleep part of state */ 1709 #define KMP_BARRIER_UNUSED_BIT 1 // bit that must never be set for valid state 1710 #define KMP_BARRIER_BUMP_BIT 2 /* lsb used for bump of go/arrived state */ 1711 1712 #define KMP_BARRIER_SLEEP_STATE (1 << KMP_BARRIER_SLEEP_BIT) 1713 #define KMP_BARRIER_UNUSED_STATE (1 << KMP_BARRIER_UNUSED_BIT) 1714 #define KMP_BARRIER_STATE_BUMP (1 << KMP_BARRIER_BUMP_BIT) 1715 1716 #if (KMP_BARRIER_SLEEP_BIT >= KMP_BARRIER_BUMP_BIT) 1717 #error "Barrier sleep bit must be smaller than barrier bump bit" 1718 #endif 1719 #if (KMP_BARRIER_UNUSED_BIT >= KMP_BARRIER_BUMP_BIT) 1720 #error "Barrier unused bit must be smaller than barrier bump bit" 1721 #endif 1722 1723 // Constants for release barrier wait state: currently, hierarchical only 1724 #define KMP_BARRIER_NOT_WAITING 0 // Normal state; worker not in wait_sleep 1725 #define KMP_BARRIER_OWN_FLAG \ 1726 1 // Normal state; worker waiting on own b_go flag in release 1727 #define KMP_BARRIER_PARENT_FLAG \ 1728 2 // Special state; worker waiting on parent's b_go flag in release 1729 #define KMP_BARRIER_SWITCH_TO_OWN_FLAG \ 1730 3 // Special state; tells worker to shift from parent to own b_go 1731 #define KMP_BARRIER_SWITCHING \ 1732 4 // Special state; worker resets appropriate flag on wake-up 1733 1734 #define KMP_NOT_SAFE_TO_REAP \ 1735 0 // Thread th_reap_state: not safe to reap (tasking) 1736 #define KMP_SAFE_TO_REAP 1 // Thread th_reap_state: safe to reap (not tasking) 1737 1738 enum barrier_type { 1739 bs_plain_barrier = 0, /* 0, All non-fork/join barriers (except reduction 1740 barriers if enabled) */ 1741 bs_forkjoin_barrier, /* 1, All fork/join (parallel region) barriers */ 1742 #if KMP_FAST_REDUCTION_BARRIER 1743 bs_reduction_barrier, /* 2, All barriers that are used in reduction */ 1744 #endif // KMP_FAST_REDUCTION_BARRIER 1745 bs_last_barrier /* Just a placeholder to mark the end */ 1746 }; 1747 1748 // to work with reduction barriers just like with plain barriers 1749 #if !KMP_FAST_REDUCTION_BARRIER 1750 #define bs_reduction_barrier bs_plain_barrier 1751 #endif // KMP_FAST_REDUCTION_BARRIER 1752 1753 typedef enum kmp_bar_pat { /* Barrier communication patterns */ 1754 bp_linear_bar = 1755 0, /* Single level (degenerate) tree */ 1756 bp_tree_bar = 1757 1, /* Balanced tree with branching factor 2^n */ 1758 bp_hyper_bar = 1759 2, /* Hypercube-embedded tree with min branching 1760 factor 2^n */ 1761 bp_hierarchical_bar = 3, /* Machine hierarchy tree */ 1762 bp_last_bar /* Placeholder to mark the end */ 1763 } kmp_bar_pat_e; 1764 1765 #define KMP_BARRIER_ICV_PUSH 1 1766 1767 /* Record for holding the values of the internal controls stack records */ 1768 typedef struct kmp_internal_control { 1769 int serial_nesting_level; /* corresponds to the value of the 1770 th_team_serialized field */ 1771 kmp_int8 nested; /* internal control for nested parallelism (per thread) */ 1772 kmp_int8 dynamic; /* internal control for dynamic adjustment of threads (per 1773 thread) */ 1774 kmp_int8 1775 bt_set; /* internal control for whether blocktime is explicitly set */ 1776 int blocktime; /* internal control for blocktime */ 1777 #if KMP_USE_MONITOR 1778 int bt_intervals; /* internal control for blocktime intervals */ 1779 #endif 1780 int nproc; /* internal control for #threads for next parallel region (per 1781 thread) */ 1782 int max_active_levels; /* internal control for max_active_levels */ 1783 kmp_r_sched_t 1784 sched; /* internal control for runtime schedule {sched,chunk} pair */ 1785 #if OMP_40_ENABLED 1786 kmp_proc_bind_t proc_bind; /* internal control for affinity */ 1787 kmp_int32 default_device; /* internal control for default device */ 1788 #endif // OMP_40_ENABLED 1789 struct kmp_internal_control *next; 1790 } kmp_internal_control_t; 1791 1792 static inline void copy_icvs(kmp_internal_control_t *dst, 1793 kmp_internal_control_t *src) { 1794 *dst = *src; 1795 } 1796 1797 /* Thread barrier needs volatile barrier fields */ 1798 typedef struct KMP_ALIGN_CACHE kmp_bstate { 1799 // th_fixed_icvs is aligned by virtue of kmp_bstate being aligned (and all 1800 // uses of it). It is not explicitly aligned below, because we *don't* want 1801 // it to be padded -- instead, we fit b_go into the same cache line with 1802 // th_fixed_icvs, enabling NGO cache lines stores in the hierarchical barrier. 1803 kmp_internal_control_t th_fixed_icvs; // Initial ICVs for the thread 1804 // Tuck b_go into end of th_fixed_icvs cache line, so it can be stored with 1805 // same NGO store 1806 volatile kmp_uint64 b_go; // STATE => task should proceed (hierarchical) 1807 KMP_ALIGN_CACHE volatile kmp_uint64 1808 b_arrived; // STATE => task reached synch point. 1809 kmp_uint32 *skip_per_level; 1810 kmp_uint32 my_level; 1811 kmp_int32 parent_tid; 1812 kmp_int32 old_tid; 1813 kmp_uint32 depth; 1814 struct kmp_bstate *parent_bar; 1815 kmp_team_t *team; 1816 kmp_uint64 leaf_state; 1817 kmp_uint32 nproc; 1818 kmp_uint8 base_leaf_kids; 1819 kmp_uint8 leaf_kids; 1820 kmp_uint8 offset; 1821 kmp_uint8 wait_flag; 1822 kmp_uint8 use_oncore_barrier; 1823 #if USE_DEBUGGER 1824 // The following field is intended for the debugger solely. Only the worker 1825 // thread itself accesses this field: the worker increases it by 1 when it 1826 // arrives to a barrier. 1827 KMP_ALIGN_CACHE kmp_uint b_worker_arrived; 1828 #endif /* USE_DEBUGGER */ 1829 } kmp_bstate_t; 1830 1831 union KMP_ALIGN_CACHE kmp_barrier_union { 1832 double b_align; /* use worst case alignment */ 1833 char b_pad[KMP_PAD(kmp_bstate_t, CACHE_LINE)]; 1834 kmp_bstate_t bb; 1835 }; 1836 1837 typedef union kmp_barrier_union kmp_balign_t; 1838 1839 /* Team barrier needs only non-volatile arrived counter */ 1840 union KMP_ALIGN_CACHE kmp_barrier_team_union { 1841 double b_align; /* use worst case alignment */ 1842 char b_pad[CACHE_LINE]; 1843 struct { 1844 kmp_uint64 b_arrived; /* STATE => task reached synch point. */ 1845 #if USE_DEBUGGER 1846 // The following two fields are indended for the debugger solely. Only 1847 // master of the team accesses these fields: the first one is increased by 1848 // 1 when master arrives to a barrier, the second one is increased by one 1849 // when all the threads arrived. 1850 kmp_uint b_master_arrived; 1851 kmp_uint b_team_arrived; 1852 #endif 1853 }; 1854 }; 1855 1856 typedef union kmp_barrier_team_union kmp_balign_team_t; 1857 1858 /* Padding for Linux* OS pthreads condition variables and mutexes used to signal 1859 threads when a condition changes. This is to workaround an NPTL bug where 1860 padding was added to pthread_cond_t which caused the initialization routine 1861 to write outside of the structure if compiled on pre-NPTL threads. */ 1862 #if KMP_OS_WINDOWS 1863 typedef struct kmp_win32_mutex { 1864 /* The Lock */ 1865 CRITICAL_SECTION cs; 1866 } kmp_win32_mutex_t; 1867 1868 typedef struct kmp_win32_cond { 1869 /* Count of the number of waiters. */ 1870 int waiters_count_; 1871 1872 /* Serialize access to <waiters_count_> */ 1873 kmp_win32_mutex_t waiters_count_lock_; 1874 1875 /* Number of threads to release via a <cond_broadcast> or a <cond_signal> */ 1876 int release_count_; 1877 1878 /* Keeps track of the current "generation" so that we don't allow */ 1879 /* one thread to steal all the "releases" from the broadcast. */ 1880 int wait_generation_count_; 1881 1882 /* A manual-reset event that's used to block and release waiting threads. */ 1883 HANDLE event_; 1884 } kmp_win32_cond_t; 1885 #endif 1886 1887 #if KMP_OS_UNIX 1888 1889 union KMP_ALIGN_CACHE kmp_cond_union { 1890 double c_align; 1891 char c_pad[CACHE_LINE]; 1892 pthread_cond_t c_cond; 1893 }; 1894 1895 typedef union kmp_cond_union kmp_cond_align_t; 1896 1897 union KMP_ALIGN_CACHE kmp_mutex_union { 1898 double m_align; 1899 char m_pad[CACHE_LINE]; 1900 pthread_mutex_t m_mutex; 1901 }; 1902 1903 typedef union kmp_mutex_union kmp_mutex_align_t; 1904 1905 #endif /* KMP_OS_UNIX */ 1906 1907 typedef struct kmp_desc_base { 1908 void *ds_stackbase; 1909 size_t ds_stacksize; 1910 int ds_stackgrow; 1911 kmp_thread_t ds_thread; 1912 volatile int ds_tid; 1913 int ds_gtid; 1914 #if KMP_OS_WINDOWS 1915 volatile int ds_alive; 1916 DWORD ds_thread_id; 1917 /* ds_thread keeps thread handle on Windows* OS. It is enough for RTL purposes. 1918 However, debugger support (libomp_db) cannot work with handles, because they 1919 uncomparable. For example, debugger requests info about thread with handle h. 1920 h is valid within debugger process, and meaningless within debugee process. 1921 Even if h is duped by call to DuplicateHandle(), so the result h' is valid 1922 within debugee process, but it is a *new* handle which does *not* equal to 1923 any other handle in debugee... The only way to compare handles is convert 1924 them to system-wide ids. GetThreadId() function is available only in 1925 Longhorn and Server 2003. :-( In contrast, GetCurrentThreadId() is available 1926 on all Windows* OS flavours (including Windows* 95). Thus, we have to get 1927 thread id by call to GetCurrentThreadId() from within the thread and save it 1928 to let libomp_db identify threads. */ 1929 #endif /* KMP_OS_WINDOWS */ 1930 } kmp_desc_base_t; 1931 1932 typedef union KMP_ALIGN_CACHE kmp_desc { 1933 double ds_align; /* use worst case alignment */ 1934 char ds_pad[KMP_PAD(kmp_desc_base_t, CACHE_LINE)]; 1935 kmp_desc_base_t ds; 1936 } kmp_desc_t; 1937 1938 typedef struct kmp_local { 1939 volatile int this_construct; /* count of single's encountered by thread */ 1940 void *reduce_data; 1941 #if KMP_USE_BGET 1942 void *bget_data; 1943 void *bget_list; 1944 #if !USE_CMP_XCHG_FOR_BGET 1945 #ifdef USE_QUEUING_LOCK_FOR_BGET 1946 kmp_lock_t bget_lock; /* Lock for accessing bget free list */ 1947 #else 1948 kmp_bootstrap_lock_t bget_lock; // Lock for accessing bget free list. Must be 1949 // bootstrap lock so we can use it at library 1950 // shutdown. 1951 #endif /* USE_LOCK_FOR_BGET */ 1952 #endif /* ! USE_CMP_XCHG_FOR_BGET */ 1953 #endif /* KMP_USE_BGET */ 1954 1955 PACKED_REDUCTION_METHOD_T 1956 packed_reduction_method; /* stored by __kmpc_reduce*(), used by 1957 __kmpc_end_reduce*() */ 1958 1959 } kmp_local_t; 1960 1961 #define KMP_CHECK_UPDATE(a, b) \ 1962 if ((a) != (b)) \ 1963 (a) = (b) 1964 #define KMP_CHECK_UPDATE_SYNC(a, b) \ 1965 if ((a) != (b)) \ 1966 TCW_SYNC_PTR((a), (b)) 1967 1968 #define get__blocktime(xteam, xtid) \ 1969 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.blocktime) 1970 #define get__bt_set(xteam, xtid) \ 1971 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_set) 1972 #if KMP_USE_MONITOR 1973 #define get__bt_intervals(xteam, xtid) \ 1974 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_intervals) 1975 #endif 1976 1977 #define get__nested_2(xteam, xtid) \ 1978 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.nested) 1979 #define get__dynamic_2(xteam, xtid) \ 1980 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.dynamic) 1981 #define get__nproc_2(xteam, xtid) \ 1982 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.nproc) 1983 #define get__sched_2(xteam, xtid) \ 1984 ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.sched) 1985 1986 #define set__blocktime_team(xteam, xtid, xval) \ 1987 (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.blocktime) = \ 1988 (xval)) 1989 1990 #if KMP_USE_MONITOR 1991 #define set__bt_intervals_team(xteam, xtid, xval) \ 1992 (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_intervals) = \ 1993 (xval)) 1994 #endif 1995 1996 #define set__bt_set_team(xteam, xtid, xval) \ 1997 (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_set) = (xval)) 1998 1999 #define set__nested(xthread, xval) \ 2000 (((xthread)->th.th_current_task->td_icvs.nested) = (xval)) 2001 #define get__nested(xthread) \ 2002 (((xthread)->th.th_current_task->td_icvs.nested) ? (FTN_TRUE) : (FTN_FALSE)) 2003 2004 #define set__dynamic(xthread, xval) \ 2005 (((xthread)->th.th_current_task->td_icvs.dynamic) = (xval)) 2006 #define get__dynamic(xthread) \ 2007 (((xthread)->th.th_current_task->td_icvs.dynamic) ? (FTN_TRUE) : (FTN_FALSE)) 2008 2009 #define set__nproc(xthread, xval) \ 2010 (((xthread)->th.th_current_task->td_icvs.nproc) = (xval)) 2011 2012 #define set__max_active_levels(xthread, xval) \ 2013 (((xthread)->th.th_current_task->td_icvs.max_active_levels) = (xval)) 2014 2015 #define set__sched(xthread, xval) \ 2016 (((xthread)->th.th_current_task->td_icvs.sched) = (xval)) 2017 2018 #if OMP_40_ENABLED 2019 2020 #define set__proc_bind(xthread, xval) \ 2021 (((xthread)->th.th_current_task->td_icvs.proc_bind) = (xval)) 2022 #define get__proc_bind(xthread) \ 2023 ((xthread)->th.th_current_task->td_icvs.proc_bind) 2024 2025 #endif /* OMP_40_ENABLED */ 2026 2027 // OpenMP tasking data structures 2028 2029 typedef enum kmp_tasking_mode { 2030 tskm_immediate_exec = 0, 2031 tskm_extra_barrier = 1, 2032 tskm_task_teams = 2, 2033 tskm_max = 2 2034 } kmp_tasking_mode_t; 2035 2036 extern kmp_tasking_mode_t 2037 __kmp_tasking_mode; /* determines how/when to execute tasks */ 2038 extern int __kmp_task_stealing_constraint; 2039 #if OMP_40_ENABLED 2040 extern kmp_int32 __kmp_default_device; // Set via OMP_DEFAULT_DEVICE if 2041 // specified, defaults to 0 otherwise 2042 #endif 2043 #if OMP_45_ENABLED 2044 // Set via OMP_MAX_TASK_PRIORITY if specified, defaults to 0 otherwise 2045 extern kmp_int32 __kmp_max_task_priority; 2046 // Set via KMP_TASKLOOP_MIN_TASKS if specified, defaults to 0 otherwise 2047 extern kmp_uint64 __kmp_taskloop_min_tasks; 2048 #endif 2049 2050 /* NOTE: kmp_taskdata_t and kmp_task_t structures allocated in single block with 2051 taskdata first */ 2052 #define KMP_TASK_TO_TASKDATA(task) (((kmp_taskdata_t *)task) - 1) 2053 #define KMP_TASKDATA_TO_TASK(taskdata) (kmp_task_t *)(taskdata + 1) 2054 2055 // The tt_found_tasks flag is a signal to all threads in the team that tasks 2056 // were spawned and queued since the previous barrier release. 2057 #define KMP_TASKING_ENABLED(task_team) \ 2058 (TCR_SYNC_4((task_team)->tt.tt_found_tasks) == TRUE) 2059 /*! 2060 @ingroup BASIC_TYPES 2061 @{ 2062 */ 2063 2064 /*! 2065 */ 2066 typedef kmp_int32 (*kmp_routine_entry_t)(kmp_int32, void *); 2067 2068 #if OMP_40_ENABLED || OMP_45_ENABLED 2069 typedef union kmp_cmplrdata { 2070 #if OMP_45_ENABLED 2071 kmp_int32 priority; /**< priority specified by user for the task */ 2072 #endif // OMP_45_ENABLED 2073 #if OMP_40_ENABLED 2074 kmp_routine_entry_t 2075 destructors; /* pointer to function to invoke deconstructors of 2076 firstprivate C++ objects */ 2077 #endif // OMP_40_ENABLED 2078 /* future data */ 2079 } kmp_cmplrdata_t; 2080 #endif 2081 2082 /* sizeof_kmp_task_t passed as arg to kmpc_omp_task call */ 2083 /*! 2084 */ 2085 typedef struct kmp_task { /* GEH: Shouldn't this be aligned somehow? */ 2086 void *shareds; /**< pointer to block of pointers to shared vars */ 2087 kmp_routine_entry_t 2088 routine; /**< pointer to routine to call for executing task */ 2089 kmp_int32 part_id; /**< part id for the task */ 2090 #if OMP_40_ENABLED || OMP_45_ENABLED 2091 kmp_cmplrdata_t 2092 data1; /* Two known optional additions: destructors and priority */ 2093 kmp_cmplrdata_t data2; /* Process destructors first, priority second */ 2094 /* future data */ 2095 #endif 2096 /* private vars */ 2097 } kmp_task_t; 2098 2099 /*! 2100 @} 2101 */ 2102 2103 #if OMP_40_ENABLED 2104 typedef struct kmp_taskgroup { 2105 kmp_int32 count; // number of allocated and not yet complete tasks 2106 kmp_int32 cancel_request; // request for cancellation of this taskgroup 2107 struct kmp_taskgroup *parent; // parent taskgroup 2108 // TODO: change to OMP_50_ENABLED, need to change build tools for this to work 2109 #if OMP_45_ENABLED 2110 // Block of data to perform task reduction 2111 void *reduce_data; // reduction related info 2112 kmp_int32 reduce_num_data; // number of data items to reduce 2113 #endif 2114 } kmp_taskgroup_t; 2115 2116 // forward declarations 2117 typedef union kmp_depnode kmp_depnode_t; 2118 typedef struct kmp_depnode_list kmp_depnode_list_t; 2119 typedef struct kmp_dephash_entry kmp_dephash_entry_t; 2120 2121 typedef struct kmp_depend_info { 2122 kmp_intptr_t base_addr; 2123 size_t len; 2124 struct { 2125 bool in : 1; 2126 bool out : 1; 2127 } flags; 2128 } kmp_depend_info_t; 2129 2130 struct kmp_depnode_list { 2131 kmp_depnode_t *node; 2132 kmp_depnode_list_t *next; 2133 }; 2134 2135 typedef struct kmp_base_depnode { 2136 kmp_depnode_list_t *successors; 2137 kmp_task_t *task; 2138 2139 kmp_lock_t lock; 2140 2141 #if KMP_SUPPORT_GRAPH_OUTPUT 2142 kmp_uint32 id; 2143 #endif 2144 2145 volatile kmp_int32 npredecessors; 2146 volatile kmp_int32 nrefs; 2147 } kmp_base_depnode_t; 2148 2149 union KMP_ALIGN_CACHE kmp_depnode { 2150 double dn_align; /* use worst case alignment */ 2151 char dn_pad[KMP_PAD(kmp_base_depnode_t, CACHE_LINE)]; 2152 kmp_base_depnode_t dn; 2153 }; 2154 2155 struct kmp_dephash_entry { 2156 kmp_intptr_t addr; 2157 kmp_depnode_t *last_out; 2158 kmp_depnode_list_t *last_ins; 2159 kmp_dephash_entry_t *next_in_bucket; 2160 }; 2161 2162 typedef struct kmp_dephash { 2163 kmp_dephash_entry_t **buckets; 2164 size_t size; 2165 #ifdef KMP_DEBUG 2166 kmp_uint32 nelements; 2167 kmp_uint32 nconflicts; 2168 #endif 2169 } kmp_dephash_t; 2170 2171 #endif 2172 2173 #ifdef BUILD_TIED_TASK_STACK 2174 2175 /* Tied Task stack definitions */ 2176 typedef struct kmp_stack_block { 2177 kmp_taskdata_t *sb_block[TASK_STACK_BLOCK_SIZE]; 2178 struct kmp_stack_block *sb_next; 2179 struct kmp_stack_block *sb_prev; 2180 } kmp_stack_block_t; 2181 2182 typedef struct kmp_task_stack { 2183 kmp_stack_block_t ts_first_block; // first block of stack entries 2184 kmp_taskdata_t **ts_top; // pointer to the top of stack 2185 kmp_int32 ts_entries; // number of entries on the stack 2186 } kmp_task_stack_t; 2187 2188 #endif // BUILD_TIED_TASK_STACK 2189 2190 typedef struct kmp_tasking_flags { /* Total struct must be exactly 32 bits */ 2191 /* Compiler flags */ /* Total compiler flags must be 16 bits */ 2192 unsigned tiedness : 1; /* task is either tied (1) or untied (0) */ 2193 unsigned final : 1; /* task is final(1) so execute immediately */ 2194 unsigned merged_if0 : 1; /* no __kmpc_task_{begin/complete}_if0 calls in if0 2195 code path */ 2196 #if OMP_40_ENABLED 2197 unsigned destructors_thunk : 1; /* set if the compiler creates a thunk to 2198 invoke destructors from the runtime */ 2199 #if OMP_45_ENABLED 2200 unsigned proxy : 1; /* task is a proxy task (it will be executed outside the 2201 context of the RTL) */ 2202 unsigned priority_specified : 1; /* set if the compiler provides priority 2203 setting for the task */ 2204 unsigned reserved : 10; /* reserved for compiler use */ 2205 #else 2206 unsigned reserved : 12; /* reserved for compiler use */ 2207 #endif 2208 #else // OMP_40_ENABLED 2209 unsigned reserved : 13; /* reserved for compiler use */ 2210 #endif // OMP_40_ENABLED 2211 2212 /* Library flags */ /* Total library flags must be 16 bits */ 2213 unsigned tasktype : 1; /* task is either explicit(1) or implicit (0) */ 2214 unsigned task_serial : 1; // task is executed immediately (1) or deferred (0) 2215 unsigned tasking_ser : 1; // all tasks in team are either executed immediately 2216 // (1) or may be deferred (0) 2217 unsigned team_serial : 1; // entire team is serial (1) [1 thread] or parallel 2218 // (0) [>= 2 threads] 2219 /* If either team_serial or tasking_ser is set, task team may be NULL */ 2220 /* Task State Flags: */ 2221 unsigned started : 1; /* 1==started, 0==not started */ 2222 unsigned executing : 1; /* 1==executing, 0==not executing */ 2223 unsigned complete : 1; /* 1==complete, 0==not complete */ 2224 unsigned freed : 1; /* 1==freed, 0==allocateed */ 2225 unsigned native : 1; /* 1==gcc-compiled task, 0==intel */ 2226 unsigned reserved31 : 7; /* reserved for library use */ 2227 2228 } kmp_tasking_flags_t; 2229 2230 struct kmp_taskdata { /* aligned during dynamic allocation */ 2231 kmp_int32 td_task_id; /* id, assigned by debugger */ 2232 kmp_tasking_flags_t td_flags; /* task flags */ 2233 kmp_team_t *td_team; /* team for this task */ 2234 kmp_info_p *td_alloc_thread; /* thread that allocated data structures */ 2235 /* Currently not used except for perhaps IDB */ 2236 kmp_taskdata_t *td_parent; /* parent task */ 2237 kmp_int32 td_level; /* task nesting level */ 2238 kmp_int32 td_untied_count; /* untied task active parts counter */ 2239 ident_t *td_ident; /* task identifier */ 2240 // Taskwait data. 2241 ident_t *td_taskwait_ident; 2242 kmp_uint32 td_taskwait_counter; 2243 kmp_int32 td_taskwait_thread; /* gtid + 1 of thread encountered taskwait */ 2244 KMP_ALIGN_CACHE kmp_internal_control_t 2245 td_icvs; /* Internal control variables for the task */ 2246 KMP_ALIGN_CACHE volatile kmp_int32 2247 td_allocated_child_tasks; /* Child tasks (+ current task) not yet 2248 deallocated */ 2249 volatile kmp_int32 2250 td_incomplete_child_tasks; /* Child tasks not yet complete */ 2251 #if OMP_40_ENABLED 2252 kmp_taskgroup_t 2253 *td_taskgroup; // Each task keeps pointer to its current taskgroup 2254 kmp_dephash_t 2255 *td_dephash; // Dependencies for children tasks are tracked from here 2256 kmp_depnode_t 2257 *td_depnode; // Pointer to graph node if this task has dependencies 2258 #endif // OMP_40_ENABLED 2259 #if OMP_45_ENABLED 2260 kmp_task_team_t *td_task_team; 2261 kmp_int32 td_size_alloc; // The size of task structure, including shareds etc. 2262 #endif // OMP_45_ENABLED 2263 kmp_taskdata_t *td_last_tied; // keep tied task for task scheduling constraint 2264 #if OMPT_SUPPORT 2265 ompt_task_info_t ompt_task_info; 2266 #endif 2267 }; // struct kmp_taskdata 2268 2269 // Make sure padding above worked 2270 KMP_BUILD_ASSERT(sizeof(kmp_taskdata_t) % sizeof(void *) == 0); 2271 2272 // Data for task team but per thread 2273 typedef struct kmp_base_thread_data { 2274 kmp_info_p *td_thr; // Pointer back to thread info 2275 // Used only in __kmp_execute_tasks_template, maybe not avail until task is 2276 // queued? 2277 kmp_bootstrap_lock_t td_deque_lock; // Lock for accessing deque 2278 kmp_taskdata_t * 2279 *td_deque; // Deque of tasks encountered by td_thr, dynamically allocated 2280 kmp_int32 td_deque_size; // Size of deck 2281 kmp_uint32 td_deque_head; // Head of deque (will wrap) 2282 kmp_uint32 td_deque_tail; // Tail of deque (will wrap) 2283 kmp_int32 td_deque_ntasks; // Number of tasks in deque 2284 // GEH: shouldn't this be volatile since used in while-spin? 2285 kmp_int32 td_deque_last_stolen; // Thread number of last successful steal 2286 #ifdef BUILD_TIED_TASK_STACK 2287 kmp_task_stack_t td_susp_tied_tasks; // Stack of suspended tied tasks for task 2288 // scheduling constraint 2289 #endif // BUILD_TIED_TASK_STACK 2290 } kmp_base_thread_data_t; 2291 2292 #define TASK_DEQUE_BITS 8 // Used solely to define INITIAL_TASK_DEQUE_SIZE 2293 #define INITIAL_TASK_DEQUE_SIZE (1 << TASK_DEQUE_BITS) 2294 2295 #define TASK_DEQUE_SIZE(td) ((td).td_deque_size) 2296 #define TASK_DEQUE_MASK(td) ((td).td_deque_size - 1) 2297 2298 typedef union KMP_ALIGN_CACHE kmp_thread_data { 2299 kmp_base_thread_data_t td; 2300 double td_align; /* use worst case alignment */ 2301 char td_pad[KMP_PAD(kmp_base_thread_data_t, CACHE_LINE)]; 2302 } kmp_thread_data_t; 2303 2304 // Data for task teams which are used when tasking is enabled for the team 2305 typedef struct kmp_base_task_team { 2306 kmp_bootstrap_lock_t 2307 tt_threads_lock; /* Lock used to allocate per-thread part of task team */ 2308 /* must be bootstrap lock since used at library shutdown*/ 2309 kmp_task_team_t *tt_next; /* For linking the task team free list */ 2310 kmp_thread_data_t 2311 *tt_threads_data; /* Array of per-thread structures for task team */ 2312 /* Data survives task team deallocation */ 2313 kmp_int32 tt_found_tasks; /* Have we found tasks and queued them while 2314 executing this team? */ 2315 /* TRUE means tt_threads_data is set up and initialized */ 2316 kmp_int32 tt_nproc; /* #threads in team */ 2317 kmp_int32 2318 tt_max_threads; /* number of entries allocated for threads_data array */ 2319 #if OMP_45_ENABLED 2320 kmp_int32 2321 tt_found_proxy_tasks; /* Have we found proxy tasks since last barrier */ 2322 #endif 2323 kmp_int32 tt_untied_task_encountered; 2324 2325 KMP_ALIGN_CACHE 2326 volatile kmp_int32 tt_unfinished_threads; /* #threads still active */ 2327 2328 KMP_ALIGN_CACHE 2329 volatile kmp_uint32 2330 tt_active; /* is the team still actively executing tasks */ 2331 } kmp_base_task_team_t; 2332 2333 union KMP_ALIGN_CACHE kmp_task_team { 2334 kmp_base_task_team_t tt; 2335 double tt_align; /* use worst case alignment */ 2336 char tt_pad[KMP_PAD(kmp_base_task_team_t, CACHE_LINE)]; 2337 }; 2338 2339 #if (USE_FAST_MEMORY == 3) || (USE_FAST_MEMORY == 5) 2340 // Free lists keep same-size free memory slots for fast memory allocation 2341 // routines 2342 typedef struct kmp_free_list { 2343 void *th_free_list_self; // Self-allocated tasks free list 2344 void *th_free_list_sync; // Self-allocated tasks stolen/returned by other 2345 // threads 2346 void *th_free_list_other; // Non-self free list (to be returned to owner's 2347 // sync list) 2348 } kmp_free_list_t; 2349 #endif 2350 #if KMP_NESTED_HOT_TEAMS 2351 // Hot teams array keeps hot teams and their sizes for given thread. Hot teams 2352 // are not put in teams pool, and they don't put threads in threads pool. 2353 typedef struct kmp_hot_team_ptr { 2354 kmp_team_p *hot_team; // pointer to hot_team of given nesting level 2355 kmp_int32 hot_team_nth; // number of threads allocated for the hot_team 2356 } kmp_hot_team_ptr_t; 2357 #endif 2358 #if OMP_40_ENABLED 2359 typedef struct kmp_teams_size { 2360 kmp_int32 nteams; // number of teams in a league 2361 kmp_int32 nth; // number of threads in each team of the league 2362 } kmp_teams_size_t; 2363 #endif 2364 2365 // OpenMP thread data structures 2366 2367 typedef struct KMP_ALIGN_CACHE kmp_base_info { 2368 /* Start with the readonly data which is cache aligned and padded. This is 2369 written before the thread starts working by the master. Uber masters may 2370 update themselves later. Usage does not consider serialized regions. */ 2371 kmp_desc_t th_info; 2372 kmp_team_p *th_team; /* team we belong to */ 2373 kmp_root_p *th_root; /* pointer to root of task hierarchy */ 2374 kmp_info_p *th_next_pool; /* next available thread in the pool */ 2375 kmp_disp_t *th_dispatch; /* thread's dispatch data */ 2376 int th_in_pool; /* in thread pool (32 bits for TCR/TCW) */ 2377 2378 /* The following are cached from the team info structure */ 2379 /* TODO use these in more places as determined to be needed via profiling */ 2380 int th_team_nproc; /* number of threads in a team */ 2381 kmp_info_p *th_team_master; /* the team's master thread */ 2382 int th_team_serialized; /* team is serialized */ 2383 #if OMP_40_ENABLED 2384 microtask_t th_teams_microtask; /* save entry address for teams construct */ 2385 int th_teams_level; /* save initial level of teams construct */ 2386 /* it is 0 on device but may be any on host */ 2387 #endif 2388 2389 /* The blocktime info is copied from the team struct to the thread sruct */ 2390 /* at the start of a barrier, and the values stored in the team are used */ 2391 /* at points in the code where the team struct is no longer guaranteed */ 2392 /* to exist (from the POV of worker threads). */ 2393 #if KMP_USE_MONITOR 2394 int th_team_bt_intervals; 2395 int th_team_bt_set; 2396 #else 2397 kmp_uint64 th_team_bt_intervals; 2398 #endif 2399 2400 #if KMP_AFFINITY_SUPPORTED 2401 kmp_affin_mask_t *th_affin_mask; /* thread's current affinity mask */ 2402 #endif 2403 2404 /* The data set by the master at reinit, then R/W by the worker */ 2405 KMP_ALIGN_CACHE int 2406 th_set_nproc; /* if > 0, then only use this request for the next fork */ 2407 #if KMP_NESTED_HOT_TEAMS 2408 kmp_hot_team_ptr_t *th_hot_teams; /* array of hot teams */ 2409 #endif 2410 #if OMP_40_ENABLED 2411 kmp_proc_bind_t 2412 th_set_proc_bind; /* if != proc_bind_default, use request for next fork */ 2413 kmp_teams_size_t 2414 th_teams_size; /* number of teams/threads in teams construct */ 2415 #if KMP_AFFINITY_SUPPORTED 2416 int th_current_place; /* place currently bound to */ 2417 int th_new_place; /* place to bind to in par reg */ 2418 int th_first_place; /* first place in partition */ 2419 int th_last_place; /* last place in partition */ 2420 #endif 2421 #endif 2422 #if USE_ITT_BUILD 2423 kmp_uint64 th_bar_arrive_time; /* arrival to barrier timestamp */ 2424 kmp_uint64 th_bar_min_time; /* minimum arrival time at the barrier */ 2425 kmp_uint64 th_frame_time; /* frame timestamp */ 2426 #endif /* USE_ITT_BUILD */ 2427 kmp_local_t th_local; 2428 struct private_common *th_pri_head; 2429 2430 /* Now the data only used by the worker (after initial allocation) */ 2431 /* TODO the first serial team should actually be stored in the info_t 2432 structure. this will help reduce initial allocation overhead */ 2433 KMP_ALIGN_CACHE kmp_team_p 2434 *th_serial_team; /*serialized team held in reserve*/ 2435 2436 #if OMPT_SUPPORT 2437 ompt_thread_info_t ompt_thread_info; 2438 #endif 2439 2440 /* The following are also read by the master during reinit */ 2441 struct common_table *th_pri_common; 2442 2443 volatile kmp_uint32 th_spin_here; /* thread-local location for spinning */ 2444 /* while awaiting queuing lock acquire */ 2445 2446 volatile void *th_sleep_loc; // this points at a kmp_flag<T> 2447 2448 ident_t *th_ident; 2449 unsigned th_x; // Random number generator data 2450 unsigned th_a; // Random number generator data 2451 2452 /* Tasking-related data for the thread */ 2453 kmp_task_team_t *th_task_team; // Task team struct 2454 kmp_taskdata_t *th_current_task; // Innermost Task being executed 2455 kmp_uint8 th_task_state; // alternating 0/1 for task team identification 2456 kmp_uint8 *th_task_state_memo_stack; // Stack holding memos of th_task_state 2457 // at nested levels 2458 kmp_uint32 th_task_state_top; // Top element of th_task_state_memo_stack 2459 kmp_uint32 th_task_state_stack_sz; // Size of th_task_state_memo_stack 2460 kmp_uint32 th_reap_state; // Non-zero indicates thread is not 2461 // tasking, thus safe to reap 2462 2463 /* More stuff for keeping track of active/sleeping threads (this part is 2464 written by the worker thread) */ 2465 kmp_uint8 th_active_in_pool; // included in count of #active threads in pool 2466 int th_active; // ! sleeping; 32 bits for TCR/TCW 2467 struct cons_header *th_cons; // used for consistency check 2468 2469 /* Add the syncronizing data which is cache aligned and padded. */ 2470 KMP_ALIGN_CACHE kmp_balign_t th_bar[bs_last_barrier]; 2471 2472 KMP_ALIGN_CACHE volatile kmp_int32 2473 th_next_waiting; /* gtid+1 of next thread on lock wait queue, 0 if none */ 2474 2475 #if (USE_FAST_MEMORY == 3) || (USE_FAST_MEMORY == 5) 2476 #define NUM_LISTS 4 2477 kmp_free_list_t th_free_lists[NUM_LISTS]; // Free lists for fast memory 2478 // allocation routines 2479 #endif 2480 2481 #if KMP_OS_WINDOWS 2482 kmp_win32_cond_t th_suspend_cv; 2483 kmp_win32_mutex_t th_suspend_mx; 2484 int th_suspend_init; 2485 #endif 2486 #if KMP_OS_UNIX 2487 kmp_cond_align_t th_suspend_cv; 2488 kmp_mutex_align_t th_suspend_mx; 2489 int th_suspend_init_count; 2490 #endif 2491 2492 #if USE_ITT_BUILD 2493 kmp_itt_mark_t th_itt_mark_single; 2494 // alignment ??? 2495 #endif /* USE_ITT_BUILD */ 2496 #if KMP_STATS_ENABLED 2497 kmp_stats_list *th_stats; 2498 #endif 2499 } kmp_base_info_t; 2500 2501 typedef union KMP_ALIGN_CACHE kmp_info { 2502 double th_align; /* use worst case alignment */ 2503 char th_pad[KMP_PAD(kmp_base_info_t, CACHE_LINE)]; 2504 kmp_base_info_t th; 2505 } kmp_info_t; 2506 2507 // OpenMP thread team data structures 2508 2509 typedef struct kmp_base_data { volatile kmp_uint32 t_value; } kmp_base_data_t; 2510 2511 typedef union KMP_ALIGN_CACHE kmp_sleep_team { 2512 double dt_align; /* use worst case alignment */ 2513 char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)]; 2514 kmp_base_data_t dt; 2515 } kmp_sleep_team_t; 2516 2517 typedef union KMP_ALIGN_CACHE kmp_ordered_team { 2518 double dt_align; /* use worst case alignment */ 2519 char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)]; 2520 kmp_base_data_t dt; 2521 } kmp_ordered_team_t; 2522 2523 typedef int (*launch_t)(int gtid); 2524 2525 /* Minimum number of ARGV entries to malloc if necessary */ 2526 #define KMP_MIN_MALLOC_ARGV_ENTRIES 100 2527 2528 // Set up how many argv pointers will fit in cache lines containing 2529 // t_inline_argv. Historically, we have supported at least 96 bytes. Using a 2530 // larger value for more space between the master write/worker read section and 2531 // read/write by all section seems to buy more performance on EPCC PARALLEL. 2532 #if KMP_ARCH_X86 || KMP_ARCH_X86_64 2533 #define KMP_INLINE_ARGV_BYTES \ 2534 (4 * CACHE_LINE - \ 2535 ((3 * KMP_PTR_SKIP + 2 * sizeof(int) + 2 * sizeof(kmp_int8) + \ 2536 sizeof(kmp_int16) + sizeof(kmp_uint32)) % \ 2537 CACHE_LINE)) 2538 #else 2539 #define KMP_INLINE_ARGV_BYTES \ 2540 (2 * CACHE_LINE - ((3 * KMP_PTR_SKIP + 2 * sizeof(int)) % CACHE_LINE)) 2541 #endif 2542 #define KMP_INLINE_ARGV_ENTRIES (int)(KMP_INLINE_ARGV_BYTES / KMP_PTR_SKIP) 2543 2544 typedef struct KMP_ALIGN_CACHE kmp_base_team { 2545 // Synchronization Data 2546 // --------------------------------------------------------------------------- 2547 KMP_ALIGN_CACHE kmp_ordered_team_t t_ordered; 2548 kmp_balign_team_t t_bar[bs_last_barrier]; 2549 volatile int t_construct; // count of single directive encountered by team 2550 char pad[sizeof(kmp_lock_t)]; // padding to maintain performance on big iron 2551 2552 // Master only 2553 // --------------------------------------------------------------------------- 2554 KMP_ALIGN_CACHE int t_master_tid; // tid of master in parent team 2555 int t_master_this_cons; // "this_construct" single counter of master in parent 2556 // team 2557 ident_t *t_ident; // if volatile, have to change too much other crud to 2558 // volatile too 2559 kmp_team_p *t_parent; // parent team 2560 kmp_team_p *t_next_pool; // next free team in the team pool 2561 kmp_disp_t *t_dispatch; // thread's dispatch data 2562 kmp_task_team_t *t_task_team[2]; // Task team struct; switch between 2 2563 #if OMP_40_ENABLED 2564 kmp_proc_bind_t t_proc_bind; // bind type for par region 2565 #endif // OMP_40_ENABLED 2566 #if USE_ITT_BUILD 2567 kmp_uint64 t_region_time; // region begin timestamp 2568 #endif /* USE_ITT_BUILD */ 2569 2570 // Master write, workers read 2571 // -------------------------------------------------------------------------- 2572 KMP_ALIGN_CACHE void **t_argv; 2573 int t_argc; 2574 int t_nproc; // number of threads in team 2575 microtask_t t_pkfn; 2576 launch_t t_invoke; // procedure to launch the microtask 2577 2578 #if OMPT_SUPPORT 2579 ompt_team_info_t ompt_team_info; 2580 ompt_lw_taskteam_t *ompt_serialized_team_info; 2581 #endif 2582 2583 #if KMP_ARCH_X86 || KMP_ARCH_X86_64 2584 kmp_int8 t_fp_control_saved; 2585 kmp_int8 t_pad2b; 2586 kmp_int16 t_x87_fpu_control_word; // FP control regs 2587 kmp_uint32 t_mxcsr; 2588 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */ 2589 2590 void *t_inline_argv[KMP_INLINE_ARGV_ENTRIES]; 2591 2592 KMP_ALIGN_CACHE kmp_info_t **t_threads; 2593 kmp_taskdata_t 2594 *t_implicit_task_taskdata; // Taskdata for the thread's implicit task 2595 int t_level; // nested parallel level 2596 2597 KMP_ALIGN_CACHE int t_max_argc; 2598 int t_max_nproc; // max threads this team can handle (dynamicly expandable) 2599 int t_serialized; // levels deep of serialized teams 2600 dispatch_shared_info_t *t_disp_buffer; // buffers for dispatch system 2601 int t_id; // team's id, assigned by debugger. 2602 int t_active_level; // nested active parallel level 2603 kmp_r_sched_t t_sched; // run-time schedule for the team 2604 #if OMP_40_ENABLED && KMP_AFFINITY_SUPPORTED 2605 int t_first_place; // first & last place in parent thread's partition. 2606 int t_last_place; // Restore these values to master after par region. 2607 #endif // OMP_40_ENABLED && KMP_AFFINITY_SUPPORTED 2608 int t_size_changed; // team size was changed?: 0: no, 1: yes, -1: changed via 2609 // omp_set_num_threads() call 2610 2611 // Read/write by workers as well 2612 #if (KMP_ARCH_X86 || KMP_ARCH_X86_64) 2613 // Using CACHE_LINE=64 reduces memory footprint, but causes a big perf 2614 // regression of epcc 'parallel' and 'barrier' on fxe256lin01. This extra 2615 // padding serves to fix the performance of epcc 'parallel' and 'barrier' when 2616 // CACHE_LINE=64. TODO: investigate more and get rid if this padding. 2617 char dummy_padding[1024]; 2618 #endif 2619 // Internal control stack for additional nested teams. 2620 KMP_ALIGN_CACHE kmp_internal_control_t *t_control_stack_top; 2621 // for SERIALIZED teams nested 2 or more levels deep 2622 #if OMP_40_ENABLED 2623 // typed flag to store request state of cancellation 2624 kmp_int32 t_cancel_request; 2625 #endif 2626 int t_master_active; // save on fork, restore on join 2627 kmp_taskq_t t_taskq; // this team's task queue 2628 void *t_copypriv_data; // team specific pointer to copyprivate data array 2629 kmp_uint32 t_copyin_counter; 2630 #if USE_ITT_BUILD 2631 void *t_stack_id; // team specific stack stitching id (for ittnotify) 2632 #endif /* USE_ITT_BUILD */ 2633 } kmp_base_team_t; 2634 2635 union KMP_ALIGN_CACHE kmp_team { 2636 kmp_base_team_t t; 2637 double t_align; /* use worst case alignment */ 2638 char t_pad[KMP_PAD(kmp_base_team_t, CACHE_LINE)]; 2639 }; 2640 2641 typedef union KMP_ALIGN_CACHE kmp_time_global { 2642 double dt_align; /* use worst case alignment */ 2643 char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)]; 2644 kmp_base_data_t dt; 2645 } kmp_time_global_t; 2646 2647 typedef struct kmp_base_global { 2648 /* cache-aligned */ 2649 kmp_time_global_t g_time; 2650 2651 /* non cache-aligned */ 2652 volatile int g_abort; 2653 volatile int g_done; 2654 2655 int g_dynamic; 2656 enum dynamic_mode g_dynamic_mode; 2657 } kmp_base_global_t; 2658 2659 typedef union KMP_ALIGN_CACHE kmp_global { 2660 kmp_base_global_t g; 2661 double g_align; /* use worst case alignment */ 2662 char g_pad[KMP_PAD(kmp_base_global_t, CACHE_LINE)]; 2663 } kmp_global_t; 2664 2665 typedef struct kmp_base_root { 2666 // TODO: GEH - combine r_active with r_in_parallel then r_active == 2667 // (r_in_parallel>= 0) 2668 // TODO: GEH - then replace r_active with t_active_levels if we can to reduce 2669 // the synch overhead or keeping r_active 2670 volatile int r_active; /* TRUE if some region in a nest has > 1 thread */ 2671 // GEH: This is misnamed, should be r_in_parallel 2672 volatile int r_nested; // TODO: GEH - This is unused, just remove it entirely. 2673 int r_in_parallel; /* keeps a count of active parallel regions per root */ 2674 // GEH: This is misnamed, should be r_active_levels 2675 kmp_team_t *r_root_team; 2676 kmp_team_t *r_hot_team; 2677 kmp_info_t *r_uber_thread; 2678 kmp_lock_t r_begin_lock; 2679 volatile int r_begin; 2680 int r_blocktime; /* blocktime for this root and descendants */ 2681 int r_cg_nthreads; // count of active threads in a contention group 2682 } kmp_base_root_t; 2683 2684 typedef union KMP_ALIGN_CACHE kmp_root { 2685 kmp_base_root_t r; 2686 double r_align; /* use worst case alignment */ 2687 char r_pad[KMP_PAD(kmp_base_root_t, CACHE_LINE)]; 2688 } kmp_root_t; 2689 2690 struct fortran_inx_info { 2691 kmp_int32 data; 2692 }; 2693 2694 /* ------------------------------------------------------------------------ */ 2695 2696 extern int __kmp_settings; 2697 extern int __kmp_duplicate_library_ok; 2698 #if USE_ITT_BUILD 2699 extern int __kmp_forkjoin_frames; 2700 extern int __kmp_forkjoin_frames_mode; 2701 #endif 2702 extern PACKED_REDUCTION_METHOD_T __kmp_force_reduction_method; 2703 extern int __kmp_determ_red; 2704 2705 #ifdef KMP_DEBUG 2706 extern int kmp_a_debug; 2707 extern int kmp_b_debug; 2708 extern int kmp_c_debug; 2709 extern int kmp_d_debug; 2710 extern int kmp_e_debug; 2711 extern int kmp_f_debug; 2712 #endif /* KMP_DEBUG */ 2713 2714 /* For debug information logging using rotating buffer */ 2715 #define KMP_DEBUG_BUF_LINES_INIT 512 2716 #define KMP_DEBUG_BUF_LINES_MIN 1 2717 2718 #define KMP_DEBUG_BUF_CHARS_INIT 128 2719 #define KMP_DEBUG_BUF_CHARS_MIN 2 2720 2721 extern int 2722 __kmp_debug_buf; /* TRUE means use buffer, FALSE means print to stderr */ 2723 extern int __kmp_debug_buf_lines; /* How many lines of debug stored in buffer */ 2724 extern int 2725 __kmp_debug_buf_chars; /* How many characters allowed per line in buffer */ 2726 extern int __kmp_debug_buf_atomic; /* TRUE means use atomic update of buffer 2727 entry pointer */ 2728 2729 extern char *__kmp_debug_buffer; /* Debug buffer itself */ 2730 extern int __kmp_debug_count; /* Counter for number of lines printed in buffer 2731 so far */ 2732 extern int __kmp_debug_buf_warn_chars; /* Keep track of char increase 2733 recommended in warnings */ 2734 /* end rotating debug buffer */ 2735 2736 #ifdef KMP_DEBUG 2737 extern int __kmp_par_range; /* +1 => only go par for constructs in range */ 2738 2739 #define KMP_PAR_RANGE_ROUTINE_LEN 1024 2740 extern char __kmp_par_range_routine[KMP_PAR_RANGE_ROUTINE_LEN]; 2741 #define KMP_PAR_RANGE_FILENAME_LEN 1024 2742 extern char __kmp_par_range_filename[KMP_PAR_RANGE_FILENAME_LEN]; 2743 extern int __kmp_par_range_lb; 2744 extern int __kmp_par_range_ub; 2745 #endif 2746 2747 /* For printing out dynamic storage map for threads and teams */ 2748 extern int 2749 __kmp_storage_map; /* True means print storage map for threads and teams */ 2750 extern int __kmp_storage_map_verbose; /* True means storage map includes 2751 placement info */ 2752 extern int __kmp_storage_map_verbose_specified; 2753 2754 #if KMP_ARCH_X86 || KMP_ARCH_X86_64 2755 extern kmp_cpuinfo_t __kmp_cpuinfo; 2756 #endif 2757 2758 extern volatile int __kmp_init_serial; 2759 extern volatile int __kmp_init_gtid; 2760 extern volatile int __kmp_init_common; 2761 extern volatile int __kmp_init_middle; 2762 extern volatile int __kmp_init_parallel; 2763 #if KMP_USE_MONITOR 2764 extern volatile int __kmp_init_monitor; 2765 #endif 2766 extern volatile int __kmp_init_user_locks; 2767 extern int __kmp_init_counter; 2768 extern int __kmp_root_counter; 2769 extern int __kmp_version; 2770 2771 /* list of address of allocated caches for commons */ 2772 extern kmp_cached_addr_t *__kmp_threadpriv_cache_list; 2773 2774 /* Barrier algorithm types and options */ 2775 extern kmp_uint32 __kmp_barrier_gather_bb_dflt; 2776 extern kmp_uint32 __kmp_barrier_release_bb_dflt; 2777 extern kmp_bar_pat_e __kmp_barrier_gather_pat_dflt; 2778 extern kmp_bar_pat_e __kmp_barrier_release_pat_dflt; 2779 extern kmp_uint32 __kmp_barrier_gather_branch_bits[bs_last_barrier]; 2780 extern kmp_uint32 __kmp_barrier_release_branch_bits[bs_last_barrier]; 2781 extern kmp_bar_pat_e __kmp_barrier_gather_pattern[bs_last_barrier]; 2782 extern kmp_bar_pat_e __kmp_barrier_release_pattern[bs_last_barrier]; 2783 extern char const *__kmp_barrier_branch_bit_env_name[bs_last_barrier]; 2784 extern char const *__kmp_barrier_pattern_env_name[bs_last_barrier]; 2785 extern char const *__kmp_barrier_type_name[bs_last_barrier]; 2786 extern char const *__kmp_barrier_pattern_name[bp_last_bar]; 2787 2788 /* Global Locks */ 2789 extern kmp_bootstrap_lock_t __kmp_initz_lock; /* control initialization */ 2790 extern kmp_bootstrap_lock_t __kmp_forkjoin_lock; /* control fork/join access */ 2791 extern kmp_bootstrap_lock_t 2792 __kmp_exit_lock; /* exit() is not always thread-safe */ 2793 #if KMP_USE_MONITOR 2794 extern kmp_bootstrap_lock_t 2795 __kmp_monitor_lock; /* control monitor thread creation */ 2796 #endif 2797 extern kmp_bootstrap_lock_t 2798 __kmp_tp_cached_lock; /* used for the hack to allow threadprivate cache and 2799 __kmp_threads expansion to co-exist */ 2800 2801 extern kmp_lock_t __kmp_global_lock; /* control OS/global access */ 2802 extern kmp_queuing_lock_t __kmp_dispatch_lock; /* control dispatch access */ 2803 extern kmp_lock_t __kmp_debug_lock; /* control I/O access for KMP_DEBUG */ 2804 2805 /* used for yielding spin-waits */ 2806 extern unsigned int __kmp_init_wait; /* initial number of spin-tests */ 2807 extern unsigned int __kmp_next_wait; /* susequent number of spin-tests */ 2808 2809 extern enum library_type __kmp_library; 2810 2811 extern enum sched_type __kmp_sched; /* default runtime scheduling */ 2812 extern enum sched_type __kmp_static; /* default static scheduling method */ 2813 extern enum sched_type __kmp_guided; /* default guided scheduling method */ 2814 extern enum sched_type __kmp_auto; /* default auto scheduling method */ 2815 extern int __kmp_chunk; /* default runtime chunk size */ 2816 2817 extern size_t __kmp_stksize; /* stack size per thread */ 2818 #if KMP_USE_MONITOR 2819 extern size_t __kmp_monitor_stksize; /* stack size for monitor thread */ 2820 #endif 2821 extern size_t __kmp_stkoffset; /* stack offset per thread */ 2822 extern int __kmp_stkpadding; /* Should we pad root thread(s) stack */ 2823 2824 extern size_t 2825 __kmp_malloc_pool_incr; /* incremental size of pool for kmp_malloc() */ 2826 extern int __kmp_env_stksize; /* was KMP_STACKSIZE specified? */ 2827 extern int __kmp_env_blocktime; /* was KMP_BLOCKTIME specified? */ 2828 extern int __kmp_env_checks; /* was KMP_CHECKS specified? */ 2829 extern int __kmp_env_consistency_check; // was KMP_CONSISTENCY_CHECK specified? 2830 extern int __kmp_generate_warnings; /* should we issue warnings? */ 2831 extern int __kmp_reserve_warn; /* have we issued reserve_threads warning? */ 2832 2833 #ifdef DEBUG_SUSPEND 2834 extern int __kmp_suspend_count; /* count inside __kmp_suspend_template() */ 2835 #endif 2836 2837 extern kmp_uint32 __kmp_yield_init; 2838 extern kmp_uint32 __kmp_yield_next; 2839 2840 #if KMP_USE_MONITOR 2841 extern kmp_uint32 __kmp_yielding_on; 2842 #endif 2843 extern kmp_uint32 __kmp_yield_cycle; 2844 extern kmp_int32 __kmp_yield_on_count; 2845 extern kmp_int32 __kmp_yield_off_count; 2846 2847 /* ------------------------------------------------------------------------- */ 2848 extern int __kmp_allThreadsSpecified; 2849 2850 extern size_t __kmp_align_alloc; 2851 /* following data protected by initialization routines */ 2852 extern int __kmp_xproc; /* number of processors in the system */ 2853 extern int __kmp_avail_proc; /* number of processors available to the process */ 2854 extern size_t __kmp_sys_min_stksize; /* system-defined minimum stack size */ 2855 extern int __kmp_sys_max_nth; /* system-imposed maximum number of threads */ 2856 // maximum total number of concurrently-existing threads on device 2857 extern int __kmp_max_nth; 2858 // maximum total number of concurrently-existing threads in a contention group 2859 extern int __kmp_cg_max_nth; 2860 extern int __kmp_teams_max_nth; // max threads used in a teams construct 2861 extern int __kmp_threads_capacity; /* capacity of the arrays __kmp_threads and 2862 __kmp_root */ 2863 extern int __kmp_dflt_team_nth; /* default number of threads in a parallel 2864 region a la OMP_NUM_THREADS */ 2865 extern int __kmp_dflt_team_nth_ub; /* upper bound on "" determined at serial 2866 initialization */ 2867 extern int __kmp_tp_capacity; /* capacity of __kmp_threads if threadprivate is 2868 used (fixed) */ 2869 extern int __kmp_tp_cached; /* whether threadprivate cache has been created 2870 (__kmpc_threadprivate_cached()) */ 2871 extern int __kmp_dflt_nested; /* nested parallelism enabled by default a la 2872 OMP_NESTED */ 2873 extern int __kmp_dflt_blocktime; /* number of milliseconds to wait before 2874 blocking (env setting) */ 2875 #if KMP_USE_MONITOR 2876 extern int 2877 __kmp_monitor_wakeups; /* number of times monitor wakes up per second */ 2878 extern int __kmp_bt_intervals; /* number of monitor timestamp intervals before 2879 blocking */ 2880 #endif 2881 #ifdef KMP_ADJUST_BLOCKTIME 2882 extern int __kmp_zero_bt; /* whether blocktime has been forced to zero */ 2883 #endif /* KMP_ADJUST_BLOCKTIME */ 2884 #ifdef KMP_DFLT_NTH_CORES 2885 extern int __kmp_ncores; /* Total number of cores for threads placement */ 2886 #endif 2887 /* Number of millisecs to delay on abort for Intel(R) VTune(TM) tools */ 2888 extern int __kmp_abort_delay; 2889 2890 extern int __kmp_need_register_atfork_specified; 2891 extern int 2892 __kmp_need_register_atfork; /* At initialization, call pthread_atfork to 2893 install fork handler */ 2894 extern int __kmp_gtid_mode; /* Method of getting gtid, values: 2895 0 - not set, will be set at runtime 2896 1 - using stack search 2897 2 - dynamic TLS (pthread_getspecific(Linux* OS/OS 2898 X*) or TlsGetValue(Windows* OS)) 2899 3 - static TLS (__declspec(thread) __kmp_gtid), 2900 Linux* OS .so only. */ 2901 extern int 2902 __kmp_adjust_gtid_mode; /* If true, adjust method based on #threads */ 2903 #ifdef KMP_TDATA_GTID 2904 extern KMP_THREAD_LOCAL int __kmp_gtid; 2905 #endif 2906 extern int __kmp_tls_gtid_min; /* #threads below which use sp search for gtid */ 2907 extern int __kmp_foreign_tp; // If true, separate TP var for each foreign thread 2908 #if KMP_ARCH_X86 || KMP_ARCH_X86_64 2909 extern int __kmp_inherit_fp_control; // copy fp creg(s) parent->workers at fork 2910 extern kmp_int16 __kmp_init_x87_fpu_control_word; // init thread's FP ctrl reg 2911 extern kmp_uint32 __kmp_init_mxcsr; /* init thread's mxscr */ 2912 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */ 2913 2914 extern int __kmp_dflt_max_active_levels; /* max_active_levels for nested 2915 parallelism enabled by default via 2916 OMP_MAX_ACTIVE_LEVELS */ 2917 extern int __kmp_dispatch_num_buffers; /* max possible dynamic loops in 2918 concurrent execution per team */ 2919 #if KMP_NESTED_HOT_TEAMS 2920 extern int __kmp_hot_teams_mode; 2921 extern int __kmp_hot_teams_max_level; 2922 #endif 2923 2924 #if KMP_OS_LINUX 2925 extern enum clock_function_type __kmp_clock_function; 2926 extern int __kmp_clock_function_param; 2927 #endif /* KMP_OS_LINUX */ 2928 2929 #if KMP_MIC_SUPPORTED 2930 extern enum mic_type __kmp_mic_type; 2931 #endif 2932 2933 #ifdef USE_LOAD_BALANCE 2934 extern double __kmp_load_balance_interval; // load balance algorithm interval 2935 #endif /* USE_LOAD_BALANCE */ 2936 2937 // OpenMP 3.1 - Nested num threads array 2938 typedef struct kmp_nested_nthreads_t { 2939 int *nth; 2940 int size; 2941 int used; 2942 } kmp_nested_nthreads_t; 2943 2944 extern kmp_nested_nthreads_t __kmp_nested_nth; 2945 2946 #if KMP_USE_ADAPTIVE_LOCKS 2947 2948 // Parameters for the speculative lock backoff system. 2949 struct kmp_adaptive_backoff_params_t { 2950 // Number of soft retries before it counts as a hard retry. 2951 kmp_uint32 max_soft_retries; 2952 // Badness is a bit mask : 0,1,3,7,15,... on each hard failure we move one to 2953 // the right 2954 kmp_uint32 max_badness; 2955 }; 2956 2957 extern kmp_adaptive_backoff_params_t __kmp_adaptive_backoff_params; 2958 2959 #if KMP_DEBUG_ADAPTIVE_LOCKS 2960 extern char *__kmp_speculative_statsfile; 2961 #endif 2962 2963 #endif // KMP_USE_ADAPTIVE_LOCKS 2964 2965 #if OMP_40_ENABLED 2966 extern int __kmp_display_env; /* TRUE or FALSE */ 2967 extern int __kmp_display_env_verbose; /* TRUE if OMP_DISPLAY_ENV=VERBOSE */ 2968 extern int __kmp_omp_cancellation; /* TRUE or FALSE */ 2969 #endif 2970 2971 /* ------------------------------------------------------------------------- */ 2972 2973 /* the following are protected by the fork/join lock */ 2974 /* write: lock read: anytime */ 2975 extern kmp_info_t **__kmp_threads; /* Descriptors for the threads */ 2976 /* read/write: lock */ 2977 extern volatile kmp_team_t *__kmp_team_pool; 2978 extern volatile kmp_info_t *__kmp_thread_pool; 2979 2980 // total num threads reachable from some root thread including all root threads 2981 extern volatile int __kmp_nth; 2982 /* total number of threads reachable from some root thread including all root 2983 threads, and those in the thread pool */ 2984 extern volatile int __kmp_all_nth; 2985 extern int __kmp_thread_pool_nth; 2986 extern volatile int __kmp_thread_pool_active_nth; 2987 2988 extern kmp_root_t **__kmp_root; /* root of thread hierarchy */ 2989 /* end data protected by fork/join lock */ 2990 /* ------------------------------------------------------------------------- */ 2991 2992 extern kmp_global_t __kmp_global; /* global status */ 2993 2994 extern kmp_info_t __kmp_monitor; 2995 extern volatile kmp_uint32 __kmp_team_counter; // For Debugging Support Library 2996 extern volatile kmp_uint32 __kmp_task_counter; // For Debugging Support Library 2997 2998 #if USE_DEBUGGER 2999 3000 #define _KMP_GEN_ID(counter) \ 3001 (__kmp_debugging ? KMP_TEST_THEN_INC32((volatile kmp_int32 *)&counter) + 1 \ 3002 : ~0) 3003 #else 3004 #define _KMP_GEN_ID(counter) (~0) 3005 #endif /* USE_DEBUGGER */ 3006 3007 #define KMP_GEN_TASK_ID() _KMP_GEN_ID(__kmp_task_counter) 3008 #define KMP_GEN_TEAM_ID() _KMP_GEN_ID(__kmp_team_counter) 3009 3010 /* ------------------------------------------------------------------------ */ 3011 3012 extern void __kmp_print_storage_map_gtid(int gtid, void *p1, void *p2, 3013 size_t size, char const *format, ...); 3014 3015 extern void __kmp_serial_initialize(void); 3016 extern void __kmp_middle_initialize(void); 3017 extern void __kmp_parallel_initialize(void); 3018 3019 extern void __kmp_internal_begin(void); 3020 extern void __kmp_internal_end_library(int gtid); 3021 extern void __kmp_internal_end_thread(int gtid); 3022 extern void __kmp_internal_end_atexit(void); 3023 extern void __kmp_internal_end_fini(void); 3024 extern void __kmp_internal_end_dtor(void); 3025 extern void __kmp_internal_end_dest(void *); 3026 3027 extern int __kmp_register_root(int initial_thread); 3028 extern void __kmp_unregister_root(int gtid); 3029 3030 extern int __kmp_ignore_mppbeg(void); 3031 extern int __kmp_ignore_mppend(void); 3032 3033 extern int __kmp_enter_single(int gtid, ident_t *id_ref, int push_ws); 3034 extern void __kmp_exit_single(int gtid); 3035 3036 extern void __kmp_parallel_deo(int *gtid_ref, int *cid_ref, ident_t *loc_ref); 3037 extern void __kmp_parallel_dxo(int *gtid_ref, int *cid_ref, ident_t *loc_ref); 3038 3039 #ifdef USE_LOAD_BALANCE 3040 extern int __kmp_get_load_balance(int); 3041 #endif 3042 3043 extern int __kmp_get_global_thread_id(void); 3044 extern int __kmp_get_global_thread_id_reg(void); 3045 extern void __kmp_exit_thread(int exit_status); 3046 extern void __kmp_abort(char const *format, ...); 3047 extern void __kmp_abort_thread(void); 3048 KMP_NORETURN extern void __kmp_abort_process(void); 3049 extern void __kmp_warn(char const *format, ...); 3050 3051 extern void __kmp_set_num_threads(int new_nth, int gtid); 3052 3053 // Returns current thread (pointer to kmp_info_t). Current thread *must* be 3054 // registered. 3055 static inline kmp_info_t *__kmp_entry_thread() { 3056 int gtid = __kmp_entry_gtid(); 3057 3058 return __kmp_threads[gtid]; 3059 } 3060 3061 extern void __kmp_set_max_active_levels(int gtid, int new_max_active_levels); 3062 extern int __kmp_get_max_active_levels(int gtid); 3063 extern int __kmp_get_ancestor_thread_num(int gtid, int level); 3064 extern int __kmp_get_team_size(int gtid, int level); 3065 extern void __kmp_set_schedule(int gtid, kmp_sched_t new_sched, int chunk); 3066 extern void __kmp_get_schedule(int gtid, kmp_sched_t *sched, int *chunk); 3067 3068 extern unsigned short __kmp_get_random(kmp_info_t *thread); 3069 extern void __kmp_init_random(kmp_info_t *thread); 3070 3071 extern kmp_r_sched_t __kmp_get_schedule_global(void); 3072 extern void __kmp_adjust_num_threads(int new_nproc); 3073 3074 extern void *___kmp_allocate(size_t size KMP_SRC_LOC_DECL); 3075 extern void *___kmp_page_allocate(size_t size KMP_SRC_LOC_DECL); 3076 extern void ___kmp_free(void *ptr KMP_SRC_LOC_DECL); 3077 #define __kmp_allocate(size) ___kmp_allocate((size)KMP_SRC_LOC_CURR) 3078 #define __kmp_page_allocate(size) ___kmp_page_allocate((size)KMP_SRC_LOC_CURR) 3079 #define __kmp_free(ptr) ___kmp_free((ptr)KMP_SRC_LOC_CURR) 3080 3081 #if USE_FAST_MEMORY 3082 extern void *___kmp_fast_allocate(kmp_info_t *this_thr, 3083 size_t size KMP_SRC_LOC_DECL); 3084 extern void ___kmp_fast_free(kmp_info_t *this_thr, void *ptr KMP_SRC_LOC_DECL); 3085 extern void __kmp_free_fast_memory(kmp_info_t *this_thr); 3086 extern void __kmp_initialize_fast_memory(kmp_info_t *this_thr); 3087 #define __kmp_fast_allocate(this_thr, size) \ 3088 ___kmp_fast_allocate((this_thr), (size)KMP_SRC_LOC_CURR) 3089 #define __kmp_fast_free(this_thr, ptr) \ 3090 ___kmp_fast_free((this_thr), (ptr)KMP_SRC_LOC_CURR) 3091 #endif 3092 3093 extern void *___kmp_thread_malloc(kmp_info_t *th, size_t size KMP_SRC_LOC_DECL); 3094 extern void *___kmp_thread_calloc(kmp_info_t *th, size_t nelem, 3095 size_t elsize KMP_SRC_LOC_DECL); 3096 extern void *___kmp_thread_realloc(kmp_info_t *th, void *ptr, 3097 size_t size KMP_SRC_LOC_DECL); 3098 extern void ___kmp_thread_free(kmp_info_t *th, void *ptr KMP_SRC_LOC_DECL); 3099 #define __kmp_thread_malloc(th, size) \ 3100 ___kmp_thread_malloc((th), (size)KMP_SRC_LOC_CURR) 3101 #define __kmp_thread_calloc(th, nelem, elsize) \ 3102 ___kmp_thread_calloc((th), (nelem), (elsize)KMP_SRC_LOC_CURR) 3103 #define __kmp_thread_realloc(th, ptr, size) \ 3104 ___kmp_thread_realloc((th), (ptr), (size)KMP_SRC_LOC_CURR) 3105 #define __kmp_thread_free(th, ptr) \ 3106 ___kmp_thread_free((th), (ptr)KMP_SRC_LOC_CURR) 3107 3108 #define KMP_INTERNAL_MALLOC(sz) malloc(sz) 3109 #define KMP_INTERNAL_FREE(p) free(p) 3110 #define KMP_INTERNAL_REALLOC(p, sz) realloc((p), (sz)) 3111 #define KMP_INTERNAL_CALLOC(n, sz) calloc((n), (sz)) 3112 3113 extern void __kmp_push_num_threads(ident_t *loc, int gtid, int num_threads); 3114 3115 #if OMP_40_ENABLED 3116 extern void __kmp_push_proc_bind(ident_t *loc, int gtid, 3117 kmp_proc_bind_t proc_bind); 3118 extern void __kmp_push_num_teams(ident_t *loc, int gtid, int num_teams, 3119 int num_threads); 3120 #endif 3121 3122 extern void __kmp_yield(int cond); 3123 3124 extern void __kmpc_dispatch_init_4(ident_t *loc, kmp_int32 gtid, 3125 enum sched_type schedule, kmp_int32 lb, 3126 kmp_int32 ub, kmp_int32 st, kmp_int32 chunk); 3127 extern void __kmpc_dispatch_init_4u(ident_t *loc, kmp_int32 gtid, 3128 enum sched_type schedule, kmp_uint32 lb, 3129 kmp_uint32 ub, kmp_int32 st, 3130 kmp_int32 chunk); 3131 extern void __kmpc_dispatch_init_8(ident_t *loc, kmp_int32 gtid, 3132 enum sched_type schedule, kmp_int64 lb, 3133 kmp_int64 ub, kmp_int64 st, kmp_int64 chunk); 3134 extern void __kmpc_dispatch_init_8u(ident_t *loc, kmp_int32 gtid, 3135 enum sched_type schedule, kmp_uint64 lb, 3136 kmp_uint64 ub, kmp_int64 st, 3137 kmp_int64 chunk); 3138 3139 extern int __kmpc_dispatch_next_4(ident_t *loc, kmp_int32 gtid, 3140 kmp_int32 *p_last, kmp_int32 *p_lb, 3141 kmp_int32 *p_ub, kmp_int32 *p_st); 3142 extern int __kmpc_dispatch_next_4u(ident_t *loc, kmp_int32 gtid, 3143 kmp_int32 *p_last, kmp_uint32 *p_lb, 3144 kmp_uint32 *p_ub, kmp_int32 *p_st); 3145 extern int __kmpc_dispatch_next_8(ident_t *loc, kmp_int32 gtid, 3146 kmp_int32 *p_last, kmp_int64 *p_lb, 3147 kmp_int64 *p_ub, kmp_int64 *p_st); 3148 extern int __kmpc_dispatch_next_8u(ident_t *loc, kmp_int32 gtid, 3149 kmp_int32 *p_last, kmp_uint64 *p_lb, 3150 kmp_uint64 *p_ub, kmp_int64 *p_st); 3151 3152 extern void __kmpc_dispatch_fini_4(ident_t *loc, kmp_int32 gtid); 3153 extern void __kmpc_dispatch_fini_8(ident_t *loc, kmp_int32 gtid); 3154 extern void __kmpc_dispatch_fini_4u(ident_t *loc, kmp_int32 gtid); 3155 extern void __kmpc_dispatch_fini_8u(ident_t *loc, kmp_int32 gtid); 3156 3157 #ifdef KMP_GOMP_COMPAT 3158 3159 extern void __kmp_aux_dispatch_init_4(ident_t *loc, kmp_int32 gtid, 3160 enum sched_type schedule, kmp_int32 lb, 3161 kmp_int32 ub, kmp_int32 st, 3162 kmp_int32 chunk, int push_ws); 3163 extern void __kmp_aux_dispatch_init_4u(ident_t *loc, kmp_int32 gtid, 3164 enum sched_type schedule, kmp_uint32 lb, 3165 kmp_uint32 ub, kmp_int32 st, 3166 kmp_int32 chunk, int push_ws); 3167 extern void __kmp_aux_dispatch_init_8(ident_t *loc, kmp_int32 gtid, 3168 enum sched_type schedule, kmp_int64 lb, 3169 kmp_int64 ub, kmp_int64 st, 3170 kmp_int64 chunk, int push_ws); 3171 extern void __kmp_aux_dispatch_init_8u(ident_t *loc, kmp_int32 gtid, 3172 enum sched_type schedule, kmp_uint64 lb, 3173 kmp_uint64 ub, kmp_int64 st, 3174 kmp_int64 chunk, int push_ws); 3175 extern void __kmp_aux_dispatch_fini_chunk_4(ident_t *loc, kmp_int32 gtid); 3176 extern void __kmp_aux_dispatch_fini_chunk_8(ident_t *loc, kmp_int32 gtid); 3177 extern void __kmp_aux_dispatch_fini_chunk_4u(ident_t *loc, kmp_int32 gtid); 3178 extern void __kmp_aux_dispatch_fini_chunk_8u(ident_t *loc, kmp_int32 gtid); 3179 3180 #endif /* KMP_GOMP_COMPAT */ 3181 3182 extern kmp_uint32 __kmp_eq_4(kmp_uint32 value, kmp_uint32 checker); 3183 extern kmp_uint32 __kmp_neq_4(kmp_uint32 value, kmp_uint32 checker); 3184 extern kmp_uint32 __kmp_lt_4(kmp_uint32 value, kmp_uint32 checker); 3185 extern kmp_uint32 __kmp_ge_4(kmp_uint32 value, kmp_uint32 checker); 3186 extern kmp_uint32 __kmp_le_4(kmp_uint32 value, kmp_uint32 checker); 3187 extern kmp_uint32 __kmp_wait_yield_4(kmp_uint32 volatile *spinner, 3188 kmp_uint32 checker, 3189 kmp_uint32 (*pred)(kmp_uint32, kmp_uint32), 3190 void *obj); 3191 extern void __kmp_wait_yield_4_ptr(void *spinner, kmp_uint32 checker, 3192 kmp_uint32 (*pred)(void *, kmp_uint32), 3193 void *obj); 3194 3195 class kmp_flag_32; 3196 class kmp_flag_64; 3197 class kmp_flag_oncore; 3198 extern void __kmp_wait_64(kmp_info_t *this_thr, kmp_flag_64 *flag, 3199 int final_spin 3200 #if USE_ITT_BUILD 3201 , 3202 void *itt_sync_obj 3203 #endif 3204 ); 3205 extern void __kmp_release_64(kmp_flag_64 *flag); 3206 3207 extern void __kmp_infinite_loop(void); 3208 3209 extern void __kmp_cleanup(void); 3210 3211 #if KMP_HANDLE_SIGNALS 3212 extern int __kmp_handle_signals; 3213 extern void __kmp_install_signals(int parallel_init); 3214 extern void __kmp_remove_signals(void); 3215 #endif 3216 3217 extern void __kmp_clear_system_time(void); 3218 extern void __kmp_read_system_time(double *delta); 3219 3220 extern void __kmp_check_stack_overlap(kmp_info_t *thr); 3221 3222 extern void __kmp_expand_host_name(char *buffer, size_t size); 3223 extern void __kmp_expand_file_name(char *result, size_t rlen, char *pattern); 3224 3225 #if KMP_ARCH_X86 || KMP_ARCH_X86_64 3226 extern void 3227 __kmp_initialize_system_tick(void); /* Initialize timer tick value */ 3228 #endif 3229 3230 extern void 3231 __kmp_runtime_initialize(void); /* machine specific initialization */ 3232 extern void __kmp_runtime_destroy(void); 3233 3234 #if KMP_AFFINITY_SUPPORTED 3235 extern char *__kmp_affinity_print_mask(char *buf, int buf_len, 3236 kmp_affin_mask_t *mask); 3237 extern void __kmp_affinity_initialize(void); 3238 extern void __kmp_affinity_uninitialize(void); 3239 extern void __kmp_affinity_set_init_mask( 3240 int gtid, int isa_root); /* set affinity according to KMP_AFFINITY */ 3241 #if OMP_40_ENABLED 3242 extern void __kmp_affinity_set_place(int gtid); 3243 #endif 3244 extern void __kmp_affinity_determine_capable(const char *env_var); 3245 extern int __kmp_aux_set_affinity(void **mask); 3246 extern int __kmp_aux_get_affinity(void **mask); 3247 extern int __kmp_aux_get_affinity_max_proc(); 3248 extern int __kmp_aux_set_affinity_mask_proc(int proc, void **mask); 3249 extern int __kmp_aux_unset_affinity_mask_proc(int proc, void **mask); 3250 extern int __kmp_aux_get_affinity_mask_proc(int proc, void **mask); 3251 extern void __kmp_balanced_affinity(int tid, int team_size); 3252 #if KMP_OS_LINUX 3253 extern int kmp_set_thread_affinity_mask_initial(void); 3254 #endif 3255 #endif /* KMP_AFFINITY_SUPPORTED */ 3256 3257 extern void __kmp_cleanup_hierarchy(); 3258 extern void __kmp_get_hierarchy(kmp_uint32 nproc, kmp_bstate_t *thr_bar); 3259 3260 #if KMP_USE_FUTEX 3261 3262 extern int __kmp_futex_determine_capable(void); 3263 3264 #endif // KMP_USE_FUTEX 3265 3266 extern void __kmp_gtid_set_specific(int gtid); 3267 extern int __kmp_gtid_get_specific(void); 3268 3269 extern double __kmp_read_cpu_time(void); 3270 3271 extern int __kmp_read_system_info(struct kmp_sys_info *info); 3272 3273 #if KMP_USE_MONITOR 3274 extern void __kmp_create_monitor(kmp_info_t *th); 3275 #endif 3276 3277 extern void *__kmp_launch_thread(kmp_info_t *thr); 3278 3279 extern void __kmp_create_worker(int gtid, kmp_info_t *th, size_t stack_size); 3280 3281 #if KMP_OS_WINDOWS 3282 extern int __kmp_still_running(kmp_info_t *th); 3283 extern int __kmp_is_thread_alive(kmp_info_t *th, DWORD *exit_val); 3284 extern void __kmp_free_handle(kmp_thread_t tHandle); 3285 #endif 3286 3287 #if KMP_USE_MONITOR 3288 extern void __kmp_reap_monitor(kmp_info_t *th); 3289 #endif 3290 extern void __kmp_reap_worker(kmp_info_t *th); 3291 extern void __kmp_terminate_thread(int gtid); 3292 3293 extern void __kmp_suspend_32(int th_gtid, kmp_flag_32 *flag); 3294 extern void __kmp_suspend_64(int th_gtid, kmp_flag_64 *flag); 3295 extern void __kmp_suspend_oncore(int th_gtid, kmp_flag_oncore *flag); 3296 extern void __kmp_resume_32(int target_gtid, kmp_flag_32 *flag); 3297 extern void __kmp_resume_64(int target_gtid, kmp_flag_64 *flag); 3298 extern void __kmp_resume_oncore(int target_gtid, kmp_flag_oncore *flag); 3299 3300 extern void __kmp_elapsed(double *); 3301 extern void __kmp_elapsed_tick(double *); 3302 3303 extern void __kmp_enable(int old_state); 3304 extern void __kmp_disable(int *old_state); 3305 3306 extern void __kmp_thread_sleep(int millis); 3307 3308 extern void __kmp_common_initialize(void); 3309 extern void __kmp_common_destroy(void); 3310 extern void __kmp_common_destroy_gtid(int gtid); 3311 3312 #if KMP_OS_UNIX 3313 extern void __kmp_register_atfork(void); 3314 #endif 3315 extern void __kmp_suspend_initialize(void); 3316 extern void __kmp_suspend_uninitialize_thread(kmp_info_t *th); 3317 3318 extern kmp_info_t *__kmp_allocate_thread(kmp_root_t *root, kmp_team_t *team, 3319 int tid); 3320 #if OMP_40_ENABLED 3321 extern kmp_team_t * 3322 __kmp_allocate_team(kmp_root_t *root, int new_nproc, int max_nproc, 3323 #if OMPT_SUPPORT 3324 ompt_data_t ompt_parallel_data, 3325 #endif 3326 kmp_proc_bind_t proc_bind, kmp_internal_control_t *new_icvs, 3327 int argc USE_NESTED_HOT_ARG(kmp_info_t *thr)); 3328 #else 3329 extern kmp_team_t * 3330 __kmp_allocate_team(kmp_root_t *root, int new_nproc, int max_nproc, 3331 #if OMPT_SUPPORT 3332 ompt_id_t ompt_parallel_id, 3333 #endif 3334 kmp_internal_control_t *new_icvs, 3335 int argc USE_NESTED_HOT_ARG(kmp_info_t *thr)); 3336 #endif // OMP_40_ENABLED 3337 extern void __kmp_free_thread(kmp_info_t *); 3338 extern void __kmp_free_team(kmp_root_t *, 3339 kmp_team_t *USE_NESTED_HOT_ARG(kmp_info_t *)); 3340 extern kmp_team_t *__kmp_reap_team(kmp_team_t *); 3341 3342 /* ------------------------------------------------------------------------ */ 3343 3344 extern void __kmp_initialize_bget(kmp_info_t *th); 3345 extern void __kmp_finalize_bget(kmp_info_t *th); 3346 3347 KMP_EXPORT void *kmpc_malloc(size_t size); 3348 KMP_EXPORT void *kmpc_aligned_malloc(size_t size, size_t alignment); 3349 KMP_EXPORT void *kmpc_calloc(size_t nelem, size_t elsize); 3350 KMP_EXPORT void *kmpc_realloc(void *ptr, size_t size); 3351 KMP_EXPORT void kmpc_free(void *ptr); 3352 3353 /* declarations for internal use */ 3354 3355 extern int __kmp_barrier(enum barrier_type bt, int gtid, int is_split, 3356 size_t reduce_size, void *reduce_data, 3357 void (*reduce)(void *, void *)); 3358 extern void __kmp_end_split_barrier(enum barrier_type bt, int gtid); 3359 3360 /*! 3361 * Tell the fork call which compiler generated the fork call, and therefore how 3362 * to deal with the call. 3363 */ 3364 enum fork_context_e { 3365 fork_context_gnu, /**< Called from GNU generated code, so must not invoke the 3366 microtask internally. */ 3367 fork_context_intel, /**< Called from Intel generated code. */ 3368 fork_context_last 3369 }; 3370 extern int __kmp_fork_call(ident_t *loc, int gtid, 3371 enum fork_context_e fork_context, kmp_int32 argc, 3372 microtask_t microtask, launch_t invoker, 3373 /* TODO: revert workaround for Intel(R) 64 tracker #96 */ 3374 #if (KMP_ARCH_ARM || KMP_ARCH_X86_64 || KMP_ARCH_AARCH64) && KMP_OS_LINUX 3375 va_list *ap 3376 #else 3377 va_list ap 3378 #endif 3379 ); 3380 3381 extern void __kmp_join_call(ident_t *loc, int gtid 3382 #if OMPT_SUPPORT 3383 , 3384 enum fork_context_e fork_context 3385 #endif 3386 #if OMP_40_ENABLED 3387 , 3388 int exit_teams = 0 3389 #endif 3390 ); 3391 3392 extern void __kmp_serialized_parallel(ident_t *id, kmp_int32 gtid); 3393 extern void __kmp_internal_fork(ident_t *id, int gtid, kmp_team_t *team); 3394 extern void __kmp_internal_join(ident_t *id, int gtid, kmp_team_t *team); 3395 extern int __kmp_invoke_task_func(int gtid); 3396 extern void __kmp_run_before_invoked_task(int gtid, int tid, 3397 kmp_info_t *this_thr, 3398 kmp_team_t *team); 3399 extern void __kmp_run_after_invoked_task(int gtid, int tid, 3400 kmp_info_t *this_thr, 3401 kmp_team_t *team); 3402 3403 // should never have been exported 3404 KMP_EXPORT int __kmpc_invoke_task_func(int gtid); 3405 #if OMP_40_ENABLED 3406 extern int __kmp_invoke_teams_master(int gtid); 3407 extern void __kmp_teams_master(int gtid); 3408 #endif 3409 extern void __kmp_save_internal_controls(kmp_info_t *thread); 3410 extern void __kmp_user_set_library(enum library_type arg); 3411 extern void __kmp_aux_set_library(enum library_type arg); 3412 extern void __kmp_aux_set_stacksize(size_t arg); 3413 extern void __kmp_aux_set_blocktime(int arg, kmp_info_t *thread, int tid); 3414 extern void __kmp_aux_set_defaults(char const *str, int len); 3415 3416 /* Functions called from __kmp_aux_env_initialize() in kmp_settings.cpp */ 3417 void kmpc_set_blocktime(int arg); 3418 void ompc_set_nested(int flag); 3419 void ompc_set_dynamic(int flag); 3420 void ompc_set_num_threads(int arg); 3421 3422 extern void __kmp_push_current_task_to_thread(kmp_info_t *this_thr, 3423 kmp_team_t *team, int tid); 3424 extern void __kmp_pop_current_task_from_thread(kmp_info_t *this_thr); 3425 extern kmp_task_t *__kmp_task_alloc(ident_t *loc_ref, kmp_int32 gtid, 3426 kmp_tasking_flags_t *flags, 3427 size_t sizeof_kmp_task_t, 3428 size_t sizeof_shareds, 3429 kmp_routine_entry_t task_entry); 3430 extern void __kmp_init_implicit_task(ident_t *loc_ref, kmp_info_t *this_thr, 3431 kmp_team_t *team, int tid, 3432 int set_curr_task); 3433 extern void __kmp_finish_implicit_task(kmp_info_t *this_thr); 3434 extern void __kmp_free_implicit_task(kmp_info_t *this_thr); 3435 int __kmp_execute_tasks_32(kmp_info_t *thread, kmp_int32 gtid, 3436 kmp_flag_32 *flag, int final_spin, 3437 int *thread_finished, 3438 #if USE_ITT_BUILD 3439 void *itt_sync_obj, 3440 #endif /* USE_ITT_BUILD */ 3441 kmp_int32 is_constrained); 3442 int __kmp_execute_tasks_64(kmp_info_t *thread, kmp_int32 gtid, 3443 kmp_flag_64 *flag, int final_spin, 3444 int *thread_finished, 3445 #if USE_ITT_BUILD 3446 void *itt_sync_obj, 3447 #endif /* USE_ITT_BUILD */ 3448 kmp_int32 is_constrained); 3449 int __kmp_execute_tasks_oncore(kmp_info_t *thread, kmp_int32 gtid, 3450 kmp_flag_oncore *flag, int final_spin, 3451 int *thread_finished, 3452 #if USE_ITT_BUILD 3453 void *itt_sync_obj, 3454 #endif /* USE_ITT_BUILD */ 3455 kmp_int32 is_constrained); 3456 3457 extern void __kmp_free_task_team(kmp_info_t *thread, 3458 kmp_task_team_t *task_team); 3459 extern void __kmp_reap_task_teams(void); 3460 extern void __kmp_wait_to_unref_task_teams(void); 3461 extern void __kmp_task_team_setup(kmp_info_t *this_thr, kmp_team_t *team, 3462 int always); 3463 extern void __kmp_task_team_sync(kmp_info_t *this_thr, kmp_team_t *team); 3464 extern void __kmp_task_team_wait(kmp_info_t *this_thr, kmp_team_t *team 3465 #if USE_ITT_BUILD 3466 , 3467 void *itt_sync_obj 3468 #endif /* USE_ITT_BUILD */ 3469 , 3470 int wait = 1); 3471 extern void __kmp_tasking_barrier(kmp_team_t *team, kmp_info_t *thread, 3472 int gtid); 3473 3474 extern int __kmp_is_address_mapped(void *addr); 3475 extern kmp_uint64 __kmp_hardware_timestamp(void); 3476 3477 #if KMP_OS_UNIX 3478 extern int __kmp_read_from_file(char const *path, char const *format, ...); 3479 #endif 3480 3481 /* ------------------------------------------------------------------------ */ 3482 // 3483 // Assembly routines that have no compiler intrinsic replacement 3484 // 3485 3486 #if KMP_ARCH_X86 || KMP_ARCH_X86_64 3487 3488 extern void __kmp_query_cpuid(kmp_cpuinfo_t *p); 3489 3490 #define __kmp_load_mxcsr(p) _mm_setcsr(*(p)) 3491 static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = _mm_getcsr(); } 3492 3493 extern void __kmp_load_x87_fpu_control_word(kmp_int16 *p); 3494 extern void __kmp_store_x87_fpu_control_word(kmp_int16 *p); 3495 extern void __kmp_clear_x87_fpu_status_word(); 3496 #define KMP_X86_MXCSR_MASK 0xffffffc0 /* ignore status flags (6 lsb) */ 3497 3498 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */ 3499 3500 extern int __kmp_invoke_microtask(microtask_t pkfn, int gtid, int npr, int argc, 3501 void *argv[] 3502 #if OMPT_SUPPORT 3503 , 3504 void **exit_frame_ptr 3505 #endif 3506 ); 3507 3508 /* ------------------------------------------------------------------------ */ 3509 3510 KMP_EXPORT void __kmpc_begin(ident_t *, kmp_int32 flags); 3511 KMP_EXPORT void __kmpc_end(ident_t *); 3512 3513 KMP_EXPORT void __kmpc_threadprivate_register_vec(ident_t *, void *data, 3514 kmpc_ctor_vec ctor, 3515 kmpc_cctor_vec cctor, 3516 kmpc_dtor_vec dtor, 3517 size_t vector_length); 3518 KMP_EXPORT void __kmpc_threadprivate_register(ident_t *, void *data, 3519 kmpc_ctor ctor, kmpc_cctor cctor, 3520 kmpc_dtor dtor); 3521 KMP_EXPORT void *__kmpc_threadprivate(ident_t *, kmp_int32 global_tid, 3522 void *data, size_t size); 3523 3524 KMP_EXPORT kmp_int32 __kmpc_global_thread_num(ident_t *); 3525 KMP_EXPORT kmp_int32 __kmpc_global_num_threads(ident_t *); 3526 KMP_EXPORT kmp_int32 __kmpc_bound_thread_num(ident_t *); 3527 KMP_EXPORT kmp_int32 __kmpc_bound_num_threads(ident_t *); 3528 3529 KMP_EXPORT kmp_int32 __kmpc_ok_to_fork(ident_t *); 3530 KMP_EXPORT void __kmpc_fork_call(ident_t *, kmp_int32 nargs, 3531 kmpc_micro microtask, ...); 3532 3533 KMP_EXPORT void __kmpc_serialized_parallel(ident_t *, kmp_int32 global_tid); 3534 KMP_EXPORT void __kmpc_end_serialized_parallel(ident_t *, kmp_int32 global_tid); 3535 3536 KMP_EXPORT void __kmpc_flush(ident_t *); 3537 KMP_EXPORT void __kmpc_barrier(ident_t *, kmp_int32 global_tid); 3538 KMP_EXPORT kmp_int32 __kmpc_master(ident_t *, kmp_int32 global_tid); 3539 KMP_EXPORT void __kmpc_end_master(ident_t *, kmp_int32 global_tid); 3540 KMP_EXPORT void __kmpc_ordered(ident_t *, kmp_int32 global_tid); 3541 KMP_EXPORT void __kmpc_end_ordered(ident_t *, kmp_int32 global_tid); 3542 KMP_EXPORT void __kmpc_critical(ident_t *, kmp_int32 global_tid, 3543 kmp_critical_name *); 3544 KMP_EXPORT void __kmpc_end_critical(ident_t *, kmp_int32 global_tid, 3545 kmp_critical_name *); 3546 3547 #if OMP_45_ENABLED 3548 KMP_EXPORT void __kmpc_critical_with_hint(ident_t *, kmp_int32 global_tid, 3549 kmp_critical_name *, uintptr_t hint); 3550 #endif 3551 3552 KMP_EXPORT kmp_int32 __kmpc_barrier_master(ident_t *, kmp_int32 global_tid); 3553 KMP_EXPORT void __kmpc_end_barrier_master(ident_t *, kmp_int32 global_tid); 3554 3555 KMP_EXPORT kmp_int32 __kmpc_barrier_master_nowait(ident_t *, 3556 kmp_int32 global_tid); 3557 3558 KMP_EXPORT kmp_int32 __kmpc_single(ident_t *, kmp_int32 global_tid); 3559 KMP_EXPORT void __kmpc_end_single(ident_t *, kmp_int32 global_tid); 3560 3561 KMP_EXPORT void KMPC_FOR_STATIC_INIT(ident_t *loc, kmp_int32 global_tid, 3562 kmp_int32 schedtype, kmp_int32 *plastiter, 3563 kmp_int *plower, kmp_int *pupper, 3564 kmp_int *pstride, kmp_int incr, 3565 kmp_int chunk); 3566 3567 KMP_EXPORT void __kmpc_for_static_fini(ident_t *loc, kmp_int32 global_tid); 3568 3569 KMP_EXPORT void __kmpc_copyprivate(ident_t *loc, kmp_int32 global_tid, 3570 size_t cpy_size, void *cpy_data, 3571 void (*cpy_func)(void *, void *), 3572 kmp_int32 didit); 3573 3574 extern void KMPC_SET_NUM_THREADS(int arg); 3575 extern void KMPC_SET_DYNAMIC(int flag); 3576 extern void KMPC_SET_NESTED(int flag); 3577 3578 /* Taskq interface routines */ 3579 KMP_EXPORT kmpc_thunk_t *__kmpc_taskq(ident_t *loc, kmp_int32 global_tid, 3580 kmpc_task_t taskq_task, 3581 size_t sizeof_thunk, 3582 size_t sizeof_shareds, kmp_int32 flags, 3583 kmpc_shared_vars_t **shareds); 3584 KMP_EXPORT void __kmpc_end_taskq(ident_t *loc, kmp_int32 global_tid, 3585 kmpc_thunk_t *thunk); 3586 KMP_EXPORT kmp_int32 __kmpc_task(ident_t *loc, kmp_int32 global_tid, 3587 kmpc_thunk_t *thunk); 3588 KMP_EXPORT void __kmpc_taskq_task(ident_t *loc, kmp_int32 global_tid, 3589 kmpc_thunk_t *thunk, kmp_int32 status); 3590 KMP_EXPORT void __kmpc_end_taskq_task(ident_t *loc, kmp_int32 global_tid, 3591 kmpc_thunk_t *thunk); 3592 KMP_EXPORT kmpc_thunk_t *__kmpc_task_buffer(ident_t *loc, kmp_int32 global_tid, 3593 kmpc_thunk_t *taskq_thunk, 3594 kmpc_task_t task); 3595 3596 /* OMP 3.0 tasking interface routines */ 3597 KMP_EXPORT kmp_int32 __kmpc_omp_task(ident_t *loc_ref, kmp_int32 gtid, 3598 kmp_task_t *new_task); 3599 KMP_EXPORT kmp_task_t *__kmpc_omp_task_alloc(ident_t *loc_ref, kmp_int32 gtid, 3600 kmp_int32 flags, 3601 size_t sizeof_kmp_task_t, 3602 size_t sizeof_shareds, 3603 kmp_routine_entry_t task_entry); 3604 KMP_EXPORT void __kmpc_omp_task_begin_if0(ident_t *loc_ref, kmp_int32 gtid, 3605 kmp_task_t *task); 3606 KMP_EXPORT void __kmpc_omp_task_complete_if0(ident_t *loc_ref, kmp_int32 gtid, 3607 kmp_task_t *task); 3608 KMP_EXPORT kmp_int32 __kmpc_omp_task_parts(ident_t *loc_ref, kmp_int32 gtid, 3609 kmp_task_t *new_task); 3610 KMP_EXPORT kmp_int32 __kmpc_omp_taskwait(ident_t *loc_ref, kmp_int32 gtid); 3611 3612 KMP_EXPORT kmp_int32 __kmpc_omp_taskyield(ident_t *loc_ref, kmp_int32 gtid, 3613 int end_part); 3614 3615 #if TASK_UNUSED 3616 void __kmpc_omp_task_begin(ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *task); 3617 void __kmpc_omp_task_complete(ident_t *loc_ref, kmp_int32 gtid, 3618 kmp_task_t *task); 3619 #endif // TASK_UNUSED 3620 3621 /* ------------------------------------------------------------------------ */ 3622 3623 #if OMP_40_ENABLED 3624 3625 KMP_EXPORT void __kmpc_taskgroup(ident_t *loc, int gtid); 3626 KMP_EXPORT void __kmpc_end_taskgroup(ident_t *loc, int gtid); 3627 3628 KMP_EXPORT kmp_int32 __kmpc_omp_task_with_deps( 3629 ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 ndeps, 3630 kmp_depend_info_t *dep_list, kmp_int32 ndeps_noalias, 3631 kmp_depend_info_t *noalias_dep_list); 3632 KMP_EXPORT void __kmpc_omp_wait_deps(ident_t *loc_ref, kmp_int32 gtid, 3633 kmp_int32 ndeps, 3634 kmp_depend_info_t *dep_list, 3635 kmp_int32 ndeps_noalias, 3636 kmp_depend_info_t *noalias_dep_list); 3637 extern void __kmp_release_deps(kmp_int32 gtid, kmp_taskdata_t *task); 3638 extern void __kmp_dephash_free_entries(kmp_info_t *thread, kmp_dephash_t *h); 3639 extern void __kmp_dephash_free(kmp_info_t *thread, kmp_dephash_t *h); 3640 3641 extern kmp_int32 __kmp_omp_task(kmp_int32 gtid, kmp_task_t *new_task, 3642 bool serialize_immediate); 3643 3644 KMP_EXPORT kmp_int32 __kmpc_cancel(ident_t *loc_ref, kmp_int32 gtid, 3645 kmp_int32 cncl_kind); 3646 KMP_EXPORT kmp_int32 __kmpc_cancellationpoint(ident_t *loc_ref, kmp_int32 gtid, 3647 kmp_int32 cncl_kind); 3648 KMP_EXPORT kmp_int32 __kmpc_cancel_barrier(ident_t *loc_ref, kmp_int32 gtid); 3649 KMP_EXPORT int __kmp_get_cancellation_status(int cancel_kind); 3650 3651 #if OMP_45_ENABLED 3652 3653 KMP_EXPORT void __kmpc_proxy_task_completed(kmp_int32 gtid, kmp_task_t *ptask); 3654 KMP_EXPORT void __kmpc_proxy_task_completed_ooo(kmp_task_t *ptask); 3655 KMP_EXPORT void __kmpc_taskloop(ident_t *loc, kmp_int32 gtid, kmp_task_t *task, 3656 kmp_int32 if_val, kmp_uint64 *lb, 3657 kmp_uint64 *ub, kmp_int64 st, kmp_int32 nogroup, 3658 kmp_int32 sched, kmp_uint64 grainsize, 3659 void *task_dup); 3660 #endif 3661 // TODO: change to OMP_50_ENABLED, need to change build tools for this to work 3662 #if OMP_45_ENABLED 3663 KMP_EXPORT void *__kmpc_task_reduction_init(int gtid, int num_data, void *data); 3664 KMP_EXPORT void *__kmpc_task_reduction_get_th_data(int gtid, void *tg, void *d); 3665 #endif 3666 3667 #endif 3668 3669 /* Lock interface routines (fast versions with gtid passed in) */ 3670 KMP_EXPORT void __kmpc_init_lock(ident_t *loc, kmp_int32 gtid, 3671 void **user_lock); 3672 KMP_EXPORT void __kmpc_init_nest_lock(ident_t *loc, kmp_int32 gtid, 3673 void **user_lock); 3674 KMP_EXPORT void __kmpc_destroy_lock(ident_t *loc, kmp_int32 gtid, 3675 void **user_lock); 3676 KMP_EXPORT void __kmpc_destroy_nest_lock(ident_t *loc, kmp_int32 gtid, 3677 void **user_lock); 3678 KMP_EXPORT void __kmpc_set_lock(ident_t *loc, kmp_int32 gtid, void **user_lock); 3679 KMP_EXPORT void __kmpc_set_nest_lock(ident_t *loc, kmp_int32 gtid, 3680 void **user_lock); 3681 KMP_EXPORT void __kmpc_unset_lock(ident_t *loc, kmp_int32 gtid, 3682 void **user_lock); 3683 KMP_EXPORT void __kmpc_unset_nest_lock(ident_t *loc, kmp_int32 gtid, 3684 void **user_lock); 3685 KMP_EXPORT int __kmpc_test_lock(ident_t *loc, kmp_int32 gtid, void **user_lock); 3686 KMP_EXPORT int __kmpc_test_nest_lock(ident_t *loc, kmp_int32 gtid, 3687 void **user_lock); 3688 3689 #if OMP_45_ENABLED 3690 KMP_EXPORT void __kmpc_init_lock_with_hint(ident_t *loc, kmp_int32 gtid, 3691 void **user_lock, uintptr_t hint); 3692 KMP_EXPORT void __kmpc_init_nest_lock_with_hint(ident_t *loc, kmp_int32 gtid, 3693 void **user_lock, 3694 uintptr_t hint); 3695 #endif 3696 3697 /* Interface to fast scalable reduce methods routines */ 3698 3699 KMP_EXPORT kmp_int32 __kmpc_reduce_nowait( 3700 ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size, 3701 void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data), 3702 kmp_critical_name *lck); 3703 KMP_EXPORT void __kmpc_end_reduce_nowait(ident_t *loc, kmp_int32 global_tid, 3704 kmp_critical_name *lck); 3705 KMP_EXPORT kmp_int32 __kmpc_reduce( 3706 ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size, 3707 void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data), 3708 kmp_critical_name *lck); 3709 KMP_EXPORT void __kmpc_end_reduce(ident_t *loc, kmp_int32 global_tid, 3710 kmp_critical_name *lck); 3711 3712 /* Internal fast reduction routines */ 3713 3714 extern PACKED_REDUCTION_METHOD_T __kmp_determine_reduction_method( 3715 ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size, 3716 void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data), 3717 kmp_critical_name *lck); 3718 3719 // this function is for testing set/get/determine reduce method 3720 KMP_EXPORT kmp_int32 __kmp_get_reduce_method(void); 3721 3722 KMP_EXPORT kmp_uint64 __kmpc_get_taskid(); 3723 KMP_EXPORT kmp_uint64 __kmpc_get_parent_taskid(); 3724 3725 // C++ port 3726 // missing 'extern "C"' declarations 3727 3728 KMP_EXPORT kmp_int32 __kmpc_in_parallel(ident_t *loc); 3729 KMP_EXPORT void __kmpc_pop_num_threads(ident_t *loc, kmp_int32 global_tid); 3730 KMP_EXPORT void __kmpc_push_num_threads(ident_t *loc, kmp_int32 global_tid, 3731 kmp_int32 num_threads); 3732 3733 #if OMP_40_ENABLED 3734 KMP_EXPORT void __kmpc_push_proc_bind(ident_t *loc, kmp_int32 global_tid, 3735 int proc_bind); 3736 KMP_EXPORT void __kmpc_push_num_teams(ident_t *loc, kmp_int32 global_tid, 3737 kmp_int32 num_teams, 3738 kmp_int32 num_threads); 3739 KMP_EXPORT void __kmpc_fork_teams(ident_t *loc, kmp_int32 argc, 3740 kmpc_micro microtask, ...); 3741 #endif 3742 #if OMP_45_ENABLED 3743 struct kmp_dim { // loop bounds info casted to kmp_int64 3744 kmp_int64 lo; // lower 3745 kmp_int64 up; // upper 3746 kmp_int64 st; // stride 3747 }; 3748 KMP_EXPORT void __kmpc_doacross_init(ident_t *loc, kmp_int32 gtid, 3749 kmp_int32 num_dims, struct kmp_dim *dims); 3750 KMP_EXPORT void __kmpc_doacross_wait(ident_t *loc, kmp_int32 gtid, 3751 kmp_int64 *vec); 3752 KMP_EXPORT void __kmpc_doacross_post(ident_t *loc, kmp_int32 gtid, 3753 kmp_int64 *vec); 3754 KMP_EXPORT void __kmpc_doacross_fini(ident_t *loc, kmp_int32 gtid); 3755 #endif 3756 3757 KMP_EXPORT void *__kmpc_threadprivate_cached(ident_t *loc, kmp_int32 global_tid, 3758 void *data, size_t size, 3759 void ***cache); 3760 3761 // Symbols for MS mutual detection. 3762 extern int _You_must_link_with_exactly_one_OpenMP_library; 3763 extern int _You_must_link_with_Intel_OpenMP_library; 3764 #if KMP_OS_WINDOWS && (KMP_VERSION_MAJOR > 4) 3765 extern int _You_must_link_with_Microsoft_OpenMP_library; 3766 #endif 3767 3768 // The routines below are not exported. 3769 // Consider making them 'static' in corresponding source files. 3770 void kmp_threadprivate_insert_private_data(int gtid, void *pc_addr, 3771 void *data_addr, size_t pc_size); 3772 struct private_common *kmp_threadprivate_insert(int gtid, void *pc_addr, 3773 void *data_addr, 3774 size_t pc_size); 3775 3776 // ompc_, kmpc_ entries moved from omp.h. 3777 #if KMP_OS_WINDOWS 3778 #define KMPC_CONVENTION __cdecl 3779 #else 3780 #define KMPC_CONVENTION 3781 #endif 3782 3783 #ifndef __OMP_H 3784 typedef enum omp_sched_t { 3785 omp_sched_static = 1, 3786 omp_sched_dynamic = 2, 3787 omp_sched_guided = 3, 3788 omp_sched_auto = 4 3789 } omp_sched_t; 3790 typedef void *kmp_affinity_mask_t; 3791 #endif 3792 3793 KMP_EXPORT void KMPC_CONVENTION ompc_set_max_active_levels(int); 3794 KMP_EXPORT void KMPC_CONVENTION ompc_set_schedule(omp_sched_t, int); 3795 KMP_EXPORT int KMPC_CONVENTION ompc_get_ancestor_thread_num(int); 3796 KMP_EXPORT int KMPC_CONVENTION ompc_get_team_size(int); 3797 KMP_EXPORT int KMPC_CONVENTION 3798 kmpc_set_affinity_mask_proc(int, kmp_affinity_mask_t *); 3799 KMP_EXPORT int KMPC_CONVENTION 3800 kmpc_unset_affinity_mask_proc(int, kmp_affinity_mask_t *); 3801 KMP_EXPORT int KMPC_CONVENTION 3802 kmpc_get_affinity_mask_proc(int, kmp_affinity_mask_t *); 3803 3804 KMP_EXPORT void KMPC_CONVENTION kmpc_set_stacksize(int); 3805 KMP_EXPORT void KMPC_CONVENTION kmpc_set_stacksize_s(size_t); 3806 KMP_EXPORT void KMPC_CONVENTION kmpc_set_library(int); 3807 KMP_EXPORT void KMPC_CONVENTION kmpc_set_defaults(char const *); 3808 KMP_EXPORT void KMPC_CONVENTION kmpc_set_disp_num_buffers(int); 3809 3810 #ifdef __cplusplus 3811 } 3812 #endif 3813 3814 #endif /* KMP_H */ 3815