LLVM OpenMP* Runtime Library
kmp.h
1 
2 /*
3  * kmp.h -- KPTS runtime header file.
4  */
5 
6 //===----------------------------------------------------------------------===//
7 //
8 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
9 // See https://llvm.org/LICENSE.txt for license information.
10 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #ifndef KMP_H
15 #define KMP_H
16 
17 #include "kmp_config.h"
18 
19 /* #define BUILD_PARALLEL_ORDERED 1 */
20 
21 /* This fix replaces gettimeofday with clock_gettime for better scalability on
22  the Altix. Requires user code to be linked with -lrt. */
23 //#define FIX_SGI_CLOCK
24 
25 /* Defines for OpenMP 3.0 tasking and auto scheduling */
26 
27 #ifndef KMP_STATIC_STEAL_ENABLED
28 #define KMP_STATIC_STEAL_ENABLED 1
29 #endif
30 #define KMP_WEIGHTED_ITERATIONS_SUPPORTED \
31  (KMP_AFFINITY_SUPPORTED && KMP_STATIC_STEAL_ENABLED && \
32  (KMP_ARCH_X86 || KMP_ARCH_X86_64))
33 
34 #define TASK_CURRENT_NOT_QUEUED 0
35 #define TASK_CURRENT_QUEUED 1
36 
37 #define TASK_NOT_PUSHED 1
38 #define TASK_SUCCESSFULLY_PUSHED 0
39 #define TASK_TIED 1
40 #define TASK_UNTIED 0
41 #define TASK_EXPLICIT 1
42 #define TASK_IMPLICIT 0
43 #define TASK_PROXY 1
44 #define TASK_FULL 0
45 #define TASK_DETACHABLE 1
46 #define TASK_UNDETACHABLE 0
47 
48 #define KMP_CANCEL_THREADS
49 #define KMP_THREAD_ATTR
50 
51 // Android does not have pthread_cancel. Undefine KMP_CANCEL_THREADS if being
52 // built on Android
53 #if defined(__ANDROID__)
54 #undef KMP_CANCEL_THREADS
55 #endif
56 
57 // Some WASI targets (e.g., wasm32-wasi-threads) do not support thread
58 // cancellation.
59 #if KMP_OS_WASI
60 #undef KMP_CANCEL_THREADS
61 #endif
62 
63 #if !KMP_OS_WASI
64 #include <signal.h>
65 #endif
66 #include <stdarg.h>
67 #include <stddef.h>
68 #include <stdio.h>
69 #include <stdlib.h>
70 #include <string.h>
71 #include <limits>
72 #include <type_traits>
73 /* include <ctype.h> don't use; problems with /MD on Windows* OS NT due to bad
74  Microsoft library. Some macros provided below to replace these functions */
75 #ifndef __ABSOFT_WIN
76 #include <sys/types.h>
77 #endif
78 #include <limits.h>
79 #include <time.h>
80 
81 #include <errno.h>
82 
83 #include "kmp_os.h"
84 
85 #include "kmp_safe_c_api.h"
86 
87 #if KMP_STATS_ENABLED
88 class kmp_stats_list;
89 #endif
90 
91 #if KMP_USE_HIER_SCHED
92 // Only include hierarchical scheduling if affinity is supported
93 #undef KMP_USE_HIER_SCHED
94 #define KMP_USE_HIER_SCHED KMP_AFFINITY_SUPPORTED
95 #endif
96 
97 // OMPD_SKIP_HWLOC used in libompd/omp-icv.cpp to avoid OMPD depending on hwloc
98 #if KMP_USE_HWLOC && KMP_AFFINITY_SUPPORTED && !defined(OMPD_SKIP_HWLOC)
99 #include "hwloc.h"
100 #ifndef HWLOC_OBJ_NUMANODE
101 #define HWLOC_OBJ_NUMANODE HWLOC_OBJ_NODE
102 #endif
103 #ifndef HWLOC_OBJ_PACKAGE
104 #define HWLOC_OBJ_PACKAGE HWLOC_OBJ_SOCKET
105 #endif
106 #endif
107 
108 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
109 #include <xmmintrin.h>
110 #endif
111 
112 // The below has to be defined before including "kmp_barrier.h".
113 #define KMP_INTERNAL_MALLOC(sz) malloc(sz)
114 #define KMP_INTERNAL_FREE(p) free(p)
115 #define KMP_INTERNAL_REALLOC(p, sz) realloc((p), (sz))
116 #define KMP_INTERNAL_CALLOC(n, sz) calloc((n), (sz))
117 
118 #include "kmp_debug.h"
119 #include "kmp_lock.h"
120 #include "kmp_version.h"
121 #include "kmp_barrier.h"
122 #if USE_DEBUGGER
123 #include "kmp_debugger.h"
124 #endif
125 #include "kmp_i18n.h"
126 
127 #define KMP_HANDLE_SIGNALS ((KMP_OS_UNIX && !KMP_OS_WASI) || KMP_OS_WINDOWS)
128 
129 #include "kmp_wrapper_malloc.h"
130 #if KMP_OS_UNIX
131 #include <unistd.h>
132 #if !defined NSIG && defined _NSIG
133 #define NSIG _NSIG
134 #endif
135 #endif
136 
137 #if KMP_OS_LINUX
138 #pragma weak clock_gettime
139 #endif
140 
141 #if OMPT_SUPPORT
142 #include "ompt-internal.h"
143 #endif
144 
145 #if OMPD_SUPPORT
146 #include "ompd-specific.h"
147 #endif
148 
149 #ifndef UNLIKELY
150 #define UNLIKELY(x) (x)
151 #endif
152 
153 // Affinity format function
154 #include "kmp_str.h"
155 
156 // 0 - no fast memory allocation, alignment: 8-byte on x86, 16-byte on x64.
157 // 3 - fast allocation using sync, non-sync free lists of any size, non-self
158 // free lists of limited size.
159 #ifndef USE_FAST_MEMORY
160 #define USE_FAST_MEMORY 3
161 #endif
162 
163 // Assume using BGET compare_exchange instruction instead of lock by default.
164 #ifndef USE_CMP_XCHG_FOR_BGET
165 #define USE_CMP_XCHG_FOR_BGET 1
166 #endif
167 
168 // Test to see if queuing lock is better than bootstrap lock for bget
169 // #ifndef USE_QUEUING_LOCK_FOR_BGET
170 // #define USE_QUEUING_LOCK_FOR_BGET
171 // #endif
172 
173 #define KMP_NSEC_PER_SEC 1000000000L
174 #define KMP_USEC_PER_SEC 1000000L
175 #define KMP_NSEC_PER_USEC 1000L
176 
185 enum {
190  /* 0x04 is no longer used */
199  KMP_IDENT_BARRIER_IMPL_MASK = 0x01C0,
200  KMP_IDENT_BARRIER_IMPL_FOR = 0x0040,
201  KMP_IDENT_BARRIER_IMPL_SECTIONS = 0x00C0,
202 
203  KMP_IDENT_BARRIER_IMPL_SINGLE = 0x0140,
204  KMP_IDENT_BARRIER_IMPL_WORKSHARE = 0x01C0,
205 
217  KMP_IDENT_ATOMIC_HINT_UNCONTENDED = 0x010000,
218  KMP_IDENT_ATOMIC_HINT_CONTENDED = 0x020000,
219  KMP_IDENT_ATOMIC_HINT_NONSPECULATIVE = 0x040000,
220  KMP_IDENT_ATOMIC_HINT_SPECULATIVE = 0x080000,
221  KMP_IDENT_OPENMP_SPEC_VERSION_MASK = 0xFF000000
222 };
223 
227 typedef struct ident {
228  kmp_int32 reserved_1;
229  kmp_int32 flags;
231  kmp_int32 reserved_2;
232 #if USE_ITT_BUILD
233 /* but currently used for storing region-specific ITT */
234 /* contextual information. */
235 #endif /* USE_ITT_BUILD */
236  kmp_int32 reserved_3;
237  char const *psource;
241  // Returns the OpenMP version in form major*10+minor (e.g., 50 for 5.0)
242  kmp_int32 get_openmp_version() {
243  return (((flags & KMP_IDENT_OPENMP_SPEC_VERSION_MASK) >> 24) & 0xFF);
244  }
250 // Some forward declarations.
251 typedef union kmp_team kmp_team_t;
252 typedef struct kmp_taskdata kmp_taskdata_t;
253 typedef union kmp_task_team kmp_task_team_t;
254 typedef union kmp_team kmp_team_p;
255 typedef union kmp_info kmp_info_p;
256 typedef union kmp_root kmp_root_p;
257 
258 template <bool C = false, bool S = true> class kmp_flag_32;
259 template <bool C = false, bool S = true> class kmp_flag_64;
260 template <bool C = false, bool S = true> class kmp_atomic_flag_64;
261 class kmp_flag_oncore;
262 
263 #ifdef __cplusplus
264 extern "C" {
265 #endif
266 
267 /* ------------------------------------------------------------------------ */
268 
269 /* Pack two 32-bit signed integers into a 64-bit signed integer */
270 /* ToDo: Fix word ordering for big-endian machines. */
271 #define KMP_PACK_64(HIGH_32, LOW_32) \
272  ((kmp_int64)((((kmp_uint64)(HIGH_32)) << 32) | (kmp_uint64)(LOW_32)))
273 
274 // Generic string manipulation macros. Assume that _x is of type char *
275 #define SKIP_WS(_x) \
276  { \
277  while (*(_x) == ' ' || *(_x) == '\t') \
278  (_x)++; \
279  }
280 #define SKIP_DIGITS(_x) \
281  { \
282  while (*(_x) >= '0' && *(_x) <= '9') \
283  (_x)++; \
284  }
285 #define SKIP_TOKEN(_x) \
286  { \
287  while ((*(_x) >= '0' && *(_x) <= '9') || (*(_x) >= 'a' && *(_x) <= 'z') || \
288  (*(_x) >= 'A' && *(_x) <= 'Z') || *(_x) == '_') \
289  (_x)++; \
290  }
291 #define SKIP_TO(_x, _c) \
292  { \
293  while (*(_x) != '\0' && *(_x) != (_c)) \
294  (_x)++; \
295  }
296 
297 /* ------------------------------------------------------------------------ */
298 
299 #define KMP_MAX(x, y) ((x) > (y) ? (x) : (y))
300 #define KMP_MIN(x, y) ((x) < (y) ? (x) : (y))
301 
302 /* ------------------------------------------------------------------------ */
303 /* Enumeration types */
304 
305 enum kmp_state_timer {
306  ts_stop,
307  ts_start,
308  ts_pause,
309 
310  ts_last_state
311 };
312 
313 enum dynamic_mode {
314  dynamic_default,
315 #ifdef USE_LOAD_BALANCE
316  dynamic_load_balance,
317 #endif /* USE_LOAD_BALANCE */
318  dynamic_random,
319  dynamic_thread_limit,
320  dynamic_max
321 };
322 
323 /* external schedule constants, duplicate enum omp_sched in omp.h in order to
324  * not include it here */
325 #ifndef KMP_SCHED_TYPE_DEFINED
326 #define KMP_SCHED_TYPE_DEFINED
327 typedef enum kmp_sched {
328  kmp_sched_lower = 0, // lower and upper bounds are for routine parameter check
329  // Note: need to adjust __kmp_sch_map global array in case enum is changed
330  kmp_sched_static = 1, // mapped to kmp_sch_static_chunked (33)
331  kmp_sched_dynamic = 2, // mapped to kmp_sch_dynamic_chunked (35)
332  kmp_sched_guided = 3, // mapped to kmp_sch_guided_chunked (36)
333  kmp_sched_auto = 4, // mapped to kmp_sch_auto (38)
334  kmp_sched_upper_std = 5, // upper bound for standard schedules
335  kmp_sched_lower_ext = 100, // lower bound of Intel extension schedules
336  kmp_sched_trapezoidal = 101, // mapped to kmp_sch_trapezoidal (39)
337 #if KMP_STATIC_STEAL_ENABLED
338  kmp_sched_static_steal = 102, // mapped to kmp_sch_static_steal (44)
339 #endif
340  kmp_sched_upper,
341  kmp_sched_default = kmp_sched_static, // default scheduling
342  kmp_sched_monotonic = 0x80000000
343 } kmp_sched_t;
344 #endif
345 
350 enum sched_type : kmp_int32 {
352  kmp_sch_static_chunked = 33,
354  kmp_sch_dynamic_chunked = 35,
356  kmp_sch_runtime = 37,
358  kmp_sch_trapezoidal = 39,
359 
360  /* accessible only through KMP_SCHEDULE environment variable */
361  kmp_sch_static_greedy = 40,
362  kmp_sch_static_balanced = 41,
363  /* accessible only through KMP_SCHEDULE environment variable */
364  kmp_sch_guided_iterative_chunked = 42,
365  kmp_sch_guided_analytical_chunked = 43,
366  /* accessible only through KMP_SCHEDULE environment variable */
367  kmp_sch_static_steal = 44,
368 
369  /* static with chunk adjustment (e.g., simd) */
370  kmp_sch_static_balanced_chunked = 45,
374  /* accessible only through KMP_SCHEDULE environment variable */
378  kmp_ord_static_chunked = 65,
380  kmp_ord_dynamic_chunked = 67,
381  kmp_ord_guided_chunked = 68,
382  kmp_ord_runtime = 69,
384  kmp_ord_trapezoidal = 71,
387  /* Schedules for Distribute construct */
391  /* For the "nomerge" versions, kmp_dispatch_next*() will always return a
392  single iteration/chunk, even if the loop is serialized. For the schedule
393  types listed above, the entire iteration vector is returned if the loop is
394  serialized. This doesn't work for gcc/gcomp sections. */
395  kmp_nm_lower = 160,
397  kmp_nm_static_chunked =
398  (kmp_sch_static_chunked - kmp_sch_lower + kmp_nm_lower),
400  kmp_nm_dynamic_chunked = 163,
402  kmp_nm_runtime = 165,
403  kmp_nm_auto = 166,
404  kmp_nm_trapezoidal = 167,
405 
406  /* accessible only through KMP_SCHEDULE environment variable */
407  kmp_nm_static_greedy = 168,
408  kmp_nm_static_balanced = 169,
409  /* accessible only through KMP_SCHEDULE environment variable */
410  kmp_nm_guided_iterative_chunked = 170,
411  kmp_nm_guided_analytical_chunked = 171,
412  kmp_nm_static_steal =
413  172, /* accessible only through OMP_SCHEDULE environment variable */
414 
415  kmp_nm_ord_static_chunked = 193,
417  kmp_nm_ord_dynamic_chunked = 195,
418  kmp_nm_ord_guided_chunked = 196,
419  kmp_nm_ord_runtime = 197,
421  kmp_nm_ord_trapezoidal = 199,
424  /* Support for OpenMP 4.5 monotonic and nonmonotonic schedule modifiers. Since
425  we need to distinguish the three possible cases (no modifier, monotonic
426  modifier, nonmonotonic modifier), we need separate bits for each modifier.
427  The absence of monotonic does not imply nonmonotonic, especially since 4.5
428  says that the behaviour of the "no modifier" case is implementation defined
429  in 4.5, but will become "nonmonotonic" in 5.0.
430 
431  Since we're passing a full 32 bit value, we can use a couple of high bits
432  for these flags; out of paranoia we avoid the sign bit.
433 
434  These modifiers can be or-ed into non-static schedules by the compiler to
435  pass the additional information. They will be stripped early in the
436  processing in __kmp_dispatch_init when setting up schedules, so most of the
437  code won't ever see schedules with these bits set. */
439  (1 << 29),
441  (1 << 30),
443 #define SCHEDULE_WITHOUT_MODIFIERS(s) \
444  (enum sched_type)( \
446 #define SCHEDULE_HAS_MONOTONIC(s) (((s)&kmp_sch_modifier_monotonic) != 0)
447 #define SCHEDULE_HAS_NONMONOTONIC(s) (((s)&kmp_sch_modifier_nonmonotonic) != 0)
448 #define SCHEDULE_HAS_NO_MODIFIERS(s) \
449  (((s) & (kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic)) == 0)
450 #define SCHEDULE_GET_MODIFIERS(s) \
451  ((enum sched_type)( \
452  (s) & (kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic)))
453 #define SCHEDULE_SET_MODIFIERS(s, m) \
454  (s = (enum sched_type)((kmp_int32)s | (kmp_int32)m))
455 #define SCHEDULE_NONMONOTONIC 0
456 #define SCHEDULE_MONOTONIC 1
457 
459 };
460 
461 // Apply modifiers on internal kind to standard kind
462 static inline void
463 __kmp_sched_apply_mods_stdkind(kmp_sched_t *kind,
464  enum sched_type internal_kind) {
465  if (SCHEDULE_HAS_MONOTONIC(internal_kind)) {
466  *kind = (kmp_sched_t)((int)*kind | (int)kmp_sched_monotonic);
467  }
468 }
469 
470 // Apply modifiers on standard kind to internal kind
471 static inline void
472 __kmp_sched_apply_mods_intkind(kmp_sched_t kind,
473  enum sched_type *internal_kind) {
474  if ((int)kind & (int)kmp_sched_monotonic) {
475  *internal_kind = (enum sched_type)((int)*internal_kind |
477  }
478 }
479 
480 // Get standard schedule without modifiers
481 static inline kmp_sched_t __kmp_sched_without_mods(kmp_sched_t kind) {
482  return (kmp_sched_t)((int)kind & ~((int)kmp_sched_monotonic));
483 }
484 
485 /* Type to keep runtime schedule set via OMP_SCHEDULE or omp_set_schedule() */
486 typedef union kmp_r_sched {
487  struct {
488  enum sched_type r_sched_type;
489  int chunk;
490  };
491  kmp_int64 sched;
492 } kmp_r_sched_t;
493 
494 extern enum sched_type __kmp_sch_map[]; // map OMP 3.0 schedule types with our
495 // internal schedule types
496 
497 enum library_type {
498  library_none,
499  library_serial,
500  library_turnaround,
501  library_throughput
502 };
503 
504 #if KMP_MIC_SUPPORTED
505 enum mic_type { non_mic, mic1, mic2, mic3, dummy };
506 #endif
507 
508 // OpenMP 3.1 - Nested num threads array
509 typedef struct kmp_nested_nthreads_t {
510  int *nth;
511  int size;
512  int used;
513 } kmp_nested_nthreads_t;
514 
515 extern kmp_nested_nthreads_t __kmp_nested_nth;
516 
517 /* -- fast reduction stuff ------------------------------------------------ */
518 
519 #undef KMP_FAST_REDUCTION_BARRIER
520 #define KMP_FAST_REDUCTION_BARRIER 1
521 
522 #undef KMP_FAST_REDUCTION_CORE_DUO
523 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
524 #define KMP_FAST_REDUCTION_CORE_DUO 1
525 #endif
526 
527 enum _reduction_method {
528  reduction_method_not_defined = 0,
529  critical_reduce_block = (1 << 8),
530  atomic_reduce_block = (2 << 8),
531  tree_reduce_block = (3 << 8),
532  empty_reduce_block = (4 << 8)
533 };
534 
535 // Description of the packed_reduction_method variable:
536 // The packed_reduction_method variable consists of two enum types variables
537 // that are packed together into 0-th byte and 1-st byte:
538 // 0: (packed_reduction_method & 0x000000FF) is a 'enum barrier_type' value of
539 // barrier that will be used in fast reduction: bs_plain_barrier or
540 // bs_reduction_barrier
541 // 1: (packed_reduction_method & 0x0000FF00) is a reduction method that will
542 // be used in fast reduction;
543 // Reduction method is of 'enum _reduction_method' type and it's defined the way
544 // so that the bits of 0-th byte are empty, so no need to execute a shift
545 // instruction while packing/unpacking
546 
547 #if KMP_FAST_REDUCTION_BARRIER
548 #define PACK_REDUCTION_METHOD_AND_BARRIER(reduction_method, barrier_type) \
549  ((reduction_method) | (barrier_type))
550 
551 #define UNPACK_REDUCTION_METHOD(packed_reduction_method) \
552  ((enum _reduction_method)((packed_reduction_method) & (0x0000FF00)))
553 
554 #define UNPACK_REDUCTION_BARRIER(packed_reduction_method) \
555  ((enum barrier_type)((packed_reduction_method) & (0x000000FF)))
556 #else
557 #define PACK_REDUCTION_METHOD_AND_BARRIER(reduction_method, barrier_type) \
558  (reduction_method)
559 
560 #define UNPACK_REDUCTION_METHOD(packed_reduction_method) \
561  (packed_reduction_method)
562 
563 #define UNPACK_REDUCTION_BARRIER(packed_reduction_method) (bs_plain_barrier)
564 #endif
565 
566 #define TEST_REDUCTION_METHOD(packed_reduction_method, which_reduction_block) \
567  ((UNPACK_REDUCTION_METHOD(packed_reduction_method)) == \
568  (which_reduction_block))
569 
570 #if KMP_FAST_REDUCTION_BARRIER
571 #define TREE_REDUCE_BLOCK_WITH_REDUCTION_BARRIER \
572  (PACK_REDUCTION_METHOD_AND_BARRIER(tree_reduce_block, bs_reduction_barrier))
573 
574 #define TREE_REDUCE_BLOCK_WITH_PLAIN_BARRIER \
575  (PACK_REDUCTION_METHOD_AND_BARRIER(tree_reduce_block, bs_plain_barrier))
576 #endif
577 
578 typedef int PACKED_REDUCTION_METHOD_T;
579 
580 /* -- end of fast reduction stuff ----------------------------------------- */
581 
582 #if KMP_OS_WINDOWS
583 #define USE_CBLKDATA
584 #if KMP_MSVC_COMPAT
585 #pragma warning(push)
586 #pragma warning(disable : 271 310)
587 #endif
588 #include <windows.h>
589 #if KMP_MSVC_COMPAT
590 #pragma warning(pop)
591 #endif
592 #endif
593 
594 #if KMP_OS_UNIX
595 #if !KMP_OS_WASI
596 #include <dlfcn.h>
597 #endif
598 #include <pthread.h>
599 #endif
600 
601 enum kmp_hw_t : int {
602  KMP_HW_UNKNOWN = -1,
603  KMP_HW_SOCKET = 0,
604  KMP_HW_PROC_GROUP,
605  KMP_HW_NUMA,
606  KMP_HW_DIE,
607  KMP_HW_LLC,
608  KMP_HW_L3,
609  KMP_HW_TILE,
610  KMP_HW_MODULE,
611  KMP_HW_L2,
612  KMP_HW_L1,
613  KMP_HW_CORE,
614  KMP_HW_THREAD,
615  KMP_HW_LAST
616 };
617 
618 typedef enum kmp_hw_core_type_t {
619  KMP_HW_CORE_TYPE_UNKNOWN = 0x0,
620 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
621  KMP_HW_CORE_TYPE_ATOM = 0x20,
622  KMP_HW_CORE_TYPE_CORE = 0x40,
623  KMP_HW_MAX_NUM_CORE_TYPES = 3,
624 #else
625  KMP_HW_MAX_NUM_CORE_TYPES = 1,
626 #endif
627 } kmp_hw_core_type_t;
628 
629 #define KMP_HW_MAX_NUM_CORE_EFFS 8
630 
631 #define KMP_DEBUG_ASSERT_VALID_HW_TYPE(type) \
632  KMP_DEBUG_ASSERT(type >= (kmp_hw_t)0 && type < KMP_HW_LAST)
633 #define KMP_ASSERT_VALID_HW_TYPE(type) \
634  KMP_ASSERT(type >= (kmp_hw_t)0 && type < KMP_HW_LAST)
635 
636 #define KMP_FOREACH_HW_TYPE(type) \
637  for (kmp_hw_t type = (kmp_hw_t)0; type < KMP_HW_LAST; \
638  type = (kmp_hw_t)((int)type + 1))
639 
640 const char *__kmp_hw_get_keyword(kmp_hw_t type, bool plural = false);
641 const char *__kmp_hw_get_catalog_string(kmp_hw_t type, bool plural = false);
642 const char *__kmp_hw_get_core_type_string(kmp_hw_core_type_t type);
643 
644 /* Only Linux* OS and Windows* OS support thread affinity. */
645 #if KMP_AFFINITY_SUPPORTED
646 
647 // GROUP_AFFINITY is already defined for _MSC_VER>=1600 (VS2010 and later).
648 #if KMP_OS_WINDOWS
649 #if _MSC_VER < 1600 && KMP_MSVC_COMPAT
650 typedef struct GROUP_AFFINITY {
651  KAFFINITY Mask;
652  WORD Group;
653  WORD Reserved[3];
654 } GROUP_AFFINITY;
655 #endif /* _MSC_VER < 1600 */
656 #if KMP_GROUP_AFFINITY
657 extern int __kmp_num_proc_groups;
658 #else
659 static const int __kmp_num_proc_groups = 1;
660 #endif /* KMP_GROUP_AFFINITY */
661 typedef DWORD (*kmp_GetActiveProcessorCount_t)(WORD);
662 extern kmp_GetActiveProcessorCount_t __kmp_GetActiveProcessorCount;
663 
664 typedef WORD (*kmp_GetActiveProcessorGroupCount_t)(void);
665 extern kmp_GetActiveProcessorGroupCount_t __kmp_GetActiveProcessorGroupCount;
666 
667 typedef BOOL (*kmp_GetThreadGroupAffinity_t)(HANDLE, GROUP_AFFINITY *);
668 extern kmp_GetThreadGroupAffinity_t __kmp_GetThreadGroupAffinity;
669 
670 typedef BOOL (*kmp_SetThreadGroupAffinity_t)(HANDLE, const GROUP_AFFINITY *,
671  GROUP_AFFINITY *);
672 extern kmp_SetThreadGroupAffinity_t __kmp_SetThreadGroupAffinity;
673 #endif /* KMP_OS_WINDOWS */
674 
675 #if KMP_USE_HWLOC && !defined(OMPD_SKIP_HWLOC)
676 extern hwloc_topology_t __kmp_hwloc_topology;
677 extern int __kmp_hwloc_error;
678 #endif
679 
680 extern size_t __kmp_affin_mask_size;
681 #define KMP_AFFINITY_CAPABLE() (__kmp_affin_mask_size > 0)
682 #define KMP_AFFINITY_DISABLE() (__kmp_affin_mask_size = 0)
683 #define KMP_AFFINITY_ENABLE(mask_size) (__kmp_affin_mask_size = mask_size)
684 #define KMP_CPU_SET_ITERATE(i, mask) \
685  for (i = (mask)->begin(); (int)i != (mask)->end(); i = (mask)->next(i))
686 #define KMP_CPU_SET(i, mask) (mask)->set(i)
687 #define KMP_CPU_ISSET(i, mask) (mask)->is_set(i)
688 #define KMP_CPU_CLR(i, mask) (mask)->clear(i)
689 #define KMP_CPU_ZERO(mask) (mask)->zero()
690 #define KMP_CPU_ISEMPTY(mask) (mask)->empty()
691 #define KMP_CPU_COPY(dest, src) (dest)->copy(src)
692 #define KMP_CPU_AND(dest, src) (dest)->bitwise_and(src)
693 #define KMP_CPU_COMPLEMENT(max_bit_number, mask) (mask)->bitwise_not()
694 #define KMP_CPU_UNION(dest, src) (dest)->bitwise_or(src)
695 #define KMP_CPU_EQUAL(dest, src) (dest)->is_equal(src)
696 #define KMP_CPU_ALLOC(ptr) (ptr = __kmp_affinity_dispatch->allocate_mask())
697 #define KMP_CPU_FREE(ptr) __kmp_affinity_dispatch->deallocate_mask(ptr)
698 #define KMP_CPU_ALLOC_ON_STACK(ptr) KMP_CPU_ALLOC(ptr)
699 #define KMP_CPU_FREE_FROM_STACK(ptr) KMP_CPU_FREE(ptr)
700 #define KMP_CPU_INTERNAL_ALLOC(ptr) KMP_CPU_ALLOC(ptr)
701 #define KMP_CPU_INTERNAL_FREE(ptr) KMP_CPU_FREE(ptr)
702 #define KMP_CPU_INDEX(arr, i) __kmp_affinity_dispatch->index_mask_array(arr, i)
703 #define KMP_CPU_ALLOC_ARRAY(arr, n) \
704  (arr = __kmp_affinity_dispatch->allocate_mask_array(n))
705 #define KMP_CPU_FREE_ARRAY(arr, n) \
706  __kmp_affinity_dispatch->deallocate_mask_array(arr)
707 #define KMP_CPU_INTERNAL_ALLOC_ARRAY(arr, n) KMP_CPU_ALLOC_ARRAY(arr, n)
708 #define KMP_CPU_INTERNAL_FREE_ARRAY(arr, n) KMP_CPU_FREE_ARRAY(arr, n)
709 #define __kmp_get_system_affinity(mask, abort_bool) \
710  (mask)->get_system_affinity(abort_bool)
711 #define __kmp_set_system_affinity(mask, abort_bool) \
712  (mask)->set_system_affinity(abort_bool)
713 #define __kmp_get_proc_group(mask) (mask)->get_proc_group()
714 
715 class KMPAffinity {
716 public:
717  class Mask {
718  public:
719  void *operator new(size_t n);
720  void operator delete(void *p);
721  void *operator new[](size_t n);
722  void operator delete[](void *p);
723  virtual ~Mask() {}
724  // Set bit i to 1
725  virtual void set(int i) {}
726  // Return bit i
727  virtual bool is_set(int i) const { return false; }
728  // Set bit i to 0
729  virtual void clear(int i) {}
730  // Zero out entire mask
731  virtual void zero() {}
732  // Check whether mask is empty
733  virtual bool empty() const { return true; }
734  // Copy src into this mask
735  virtual void copy(const Mask *src) {}
736  // this &= rhs
737  virtual void bitwise_and(const Mask *rhs) {}
738  // this |= rhs
739  virtual void bitwise_or(const Mask *rhs) {}
740  // this = ~this
741  virtual void bitwise_not() {}
742  // this == rhs
743  virtual bool is_equal(const Mask *rhs) const { return false; }
744  // API for iterating over an affinity mask
745  // for (int i = mask->begin(); i != mask->end(); i = mask->next(i))
746  virtual int begin() const { return 0; }
747  virtual int end() const { return 0; }
748  virtual int next(int previous) const { return 0; }
749 #if KMP_OS_WINDOWS
750  virtual int set_process_affinity(bool abort_on_error) const { return -1; }
751 #endif
752  // Set the system's affinity to this affinity mask's value
753  virtual int set_system_affinity(bool abort_on_error) const { return -1; }
754  // Set this affinity mask to the current system affinity
755  virtual int get_system_affinity(bool abort_on_error) { return -1; }
756  // Only 1 DWORD in the mask should have any procs set.
757  // Return the appropriate index, or -1 for an invalid mask.
758  virtual int get_proc_group() const { return -1; }
759  int get_max_cpu() const {
760  int cpu;
761  int max_cpu = -1;
762  KMP_CPU_SET_ITERATE(cpu, this) {
763  if (cpu > max_cpu)
764  max_cpu = cpu;
765  }
766  return max_cpu;
767  }
768  };
769  void *operator new(size_t n);
770  void operator delete(void *p);
771  // Need virtual destructor
772  virtual ~KMPAffinity() = default;
773  // Determine if affinity is capable
774  virtual void determine_capable(const char *env_var) {}
775  // Bind the current thread to os proc
776  virtual void bind_thread(int proc) {}
777  // Factory functions to allocate/deallocate a mask
778  virtual Mask *allocate_mask() { return nullptr; }
779  virtual void deallocate_mask(Mask *m) {}
780  virtual Mask *allocate_mask_array(int num) { return nullptr; }
781  virtual void deallocate_mask_array(Mask *m) {}
782  virtual Mask *index_mask_array(Mask *m, int index) { return nullptr; }
783  static void pick_api();
784  static void destroy_api();
785  enum api_type {
786  NATIVE_OS
787 #if KMP_USE_HWLOC
788  ,
789  HWLOC
790 #endif
791  };
792  virtual api_type get_api_type() const {
793  KMP_ASSERT(0);
794  return NATIVE_OS;
795  }
796 
797 private:
798  static bool picked_api;
799 };
800 
801 typedef KMPAffinity::Mask kmp_affin_mask_t;
802 extern KMPAffinity *__kmp_affinity_dispatch;
803 
804 #if !KMP_OS_AIX
805 class kmp_affinity_raii_t {
806  kmp_affin_mask_t *mask;
807  bool restored;
808 
809 public:
810  kmp_affinity_raii_t(const kmp_affin_mask_t *new_mask = nullptr)
811  : mask(nullptr), restored(false) {
812  if (KMP_AFFINITY_CAPABLE()) {
813  KMP_CPU_ALLOC(mask);
814  KMP_ASSERT(mask != NULL);
815  __kmp_get_system_affinity(mask, /*abort_on_error=*/true);
816  if (new_mask)
817  __kmp_set_system_affinity(new_mask, /*abort_on_error=*/true);
818  }
819  }
820  void restore() {
821  if (mask && KMP_AFFINITY_CAPABLE() && !restored) {
822  __kmp_set_system_affinity(mask, /*abort_on_error=*/true);
823  KMP_CPU_FREE(mask);
824  }
825  restored = true;
826  }
827  ~kmp_affinity_raii_t() { restore(); }
828 };
829 #endif // !KMP_OS_AIX
830 
831 // Declare local char buffers with this size for printing debug and info
832 // messages, using __kmp_affinity_print_mask().
833 #define KMP_AFFIN_MASK_PRINT_LEN 1024
834 
835 enum affinity_type {
836  affinity_none = 0,
837  affinity_physical,
838  affinity_logical,
839  affinity_compact,
840  affinity_scatter,
841  affinity_explicit,
842  affinity_balanced,
843  affinity_disabled, // not used outsize the env var parser
844  affinity_default
845 };
846 
847 enum affinity_top_method {
848  affinity_top_method_all = 0, // try all (supported) methods, in order
849 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
850  affinity_top_method_apicid,
851  affinity_top_method_x2apicid,
852  affinity_top_method_x2apicid_1f,
853 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
854  affinity_top_method_cpuinfo, // KMP_CPUINFO_FILE is usable on Windows* OS, too
855 #if KMP_GROUP_AFFINITY
856  affinity_top_method_group,
857 #endif /* KMP_GROUP_AFFINITY */
858  affinity_top_method_flat,
859 #if KMP_USE_HWLOC
860  affinity_top_method_hwloc,
861 #endif
862  affinity_top_method_default
863 };
864 
865 #define affinity_respect_mask_default (2)
866 
867 typedef struct kmp_affinity_flags_t {
868  unsigned dups : 1;
869  unsigned verbose : 1;
870  unsigned warnings : 1;
871  unsigned respect : 2;
872  unsigned reset : 1;
873  unsigned initialized : 1;
874  unsigned core_types_gran : 1;
875  unsigned core_effs_gran : 1;
876  unsigned omp_places : 1;
877  unsigned reserved : 22;
878 } kmp_affinity_flags_t;
879 KMP_BUILD_ASSERT(sizeof(kmp_affinity_flags_t) == 4);
880 
881 typedef struct kmp_affinity_ids_t {
882  int os_id;
883  int ids[KMP_HW_LAST];
884 } kmp_affinity_ids_t;
885 
886 typedef struct kmp_affinity_attrs_t {
887  int core_type : 8;
888  int core_eff : 8;
889  unsigned valid : 1;
890  unsigned reserved : 15;
891 } kmp_affinity_attrs_t;
892 #define KMP_AFFINITY_ATTRS_UNKNOWN \
893  { KMP_HW_CORE_TYPE_UNKNOWN, kmp_hw_attr_t::UNKNOWN_CORE_EFF, 0, 0 }
894 
895 typedef struct kmp_affinity_t {
896  char *proclist;
897  enum affinity_type type;
898  kmp_hw_t gran;
899  int gran_levels;
900  kmp_affinity_attrs_t core_attr_gran;
901  int compact;
902  int offset;
903  kmp_affinity_flags_t flags;
904  unsigned num_masks;
905  kmp_affin_mask_t *masks;
906  kmp_affinity_ids_t *ids;
907  kmp_affinity_attrs_t *attrs;
908  unsigned num_os_id_masks;
909  kmp_affin_mask_t *os_id_masks;
910  const char *env_var;
911 } kmp_affinity_t;
912 
913 #define KMP_AFFINITY_INIT(env) \
914  { \
915  nullptr, affinity_default, KMP_HW_UNKNOWN, -1, KMP_AFFINITY_ATTRS_UNKNOWN, \
916  0, 0, \
917  {TRUE, FALSE, TRUE, affinity_respect_mask_default, FALSE, FALSE, \
918  FALSE, FALSE, FALSE}, \
919  0, nullptr, nullptr, nullptr, 0, nullptr, env \
920  }
921 
922 extern enum affinity_top_method __kmp_affinity_top_method;
923 extern kmp_affinity_t __kmp_affinity;
924 extern kmp_affinity_t __kmp_hh_affinity;
925 extern kmp_affinity_t *__kmp_affinities[2];
926 
927 extern void __kmp_affinity_bind_thread(int which);
928 
929 extern kmp_affin_mask_t *__kmp_affin_fullMask;
930 extern kmp_affin_mask_t *__kmp_affin_origMask;
931 extern char *__kmp_cpuinfo_file;
932 
933 #if KMP_WEIGHTED_ITERATIONS_SUPPORTED
934 extern int __kmp_first_osid_with_ecore;
935 #endif
936 
937 #endif /* KMP_AFFINITY_SUPPORTED */
938 
939 // This needs to be kept in sync with the values in omp.h !!!
940 typedef enum kmp_proc_bind_t {
941  proc_bind_false = 0,
942  proc_bind_true,
943  proc_bind_primary,
944  proc_bind_close,
945  proc_bind_spread,
946  proc_bind_intel, // use KMP_AFFINITY interface
947  proc_bind_default
948 } kmp_proc_bind_t;
949 
950 typedef struct kmp_nested_proc_bind_t {
951  kmp_proc_bind_t *bind_types;
952  int size;
953  int used;
954 } kmp_nested_proc_bind_t;
955 
956 extern kmp_nested_proc_bind_t __kmp_nested_proc_bind;
957 extern kmp_proc_bind_t __kmp_teams_proc_bind;
958 
959 extern int __kmp_display_affinity;
960 extern char *__kmp_affinity_format;
961 static const size_t KMP_AFFINITY_FORMAT_SIZE = 512;
962 #if OMPT_SUPPORT
963 extern int __kmp_tool;
964 extern char *__kmp_tool_libraries;
965 #endif // OMPT_SUPPORT
966 
967 #if KMP_AFFINITY_SUPPORTED
968 #define KMP_PLACE_ALL (-1)
969 #define KMP_PLACE_UNDEFINED (-2)
970 // Is KMP_AFFINITY is being used instead of OMP_PROC_BIND/OMP_PLACES?
971 #define KMP_AFFINITY_NON_PROC_BIND \
972  ((__kmp_nested_proc_bind.bind_types[0] == proc_bind_false || \
973  __kmp_nested_proc_bind.bind_types[0] == proc_bind_intel) && \
974  (__kmp_affinity.num_masks > 0 || __kmp_affinity.type == affinity_balanced))
975 #endif /* KMP_AFFINITY_SUPPORTED */
976 
977 extern int __kmp_affinity_num_places;
978 
979 typedef enum kmp_cancel_kind_t {
980  cancel_noreq = 0,
981  cancel_parallel = 1,
982  cancel_loop = 2,
983  cancel_sections = 3,
984  cancel_taskgroup = 4
985 } kmp_cancel_kind_t;
986 
987 // KMP_HW_SUBSET support:
988 typedef struct kmp_hws_item {
989  int num;
990  int offset;
991 } kmp_hws_item_t;
992 
993 extern kmp_hws_item_t __kmp_hws_socket;
994 extern kmp_hws_item_t __kmp_hws_die;
995 extern kmp_hws_item_t __kmp_hws_node;
996 extern kmp_hws_item_t __kmp_hws_tile;
997 extern kmp_hws_item_t __kmp_hws_core;
998 extern kmp_hws_item_t __kmp_hws_proc;
999 extern int __kmp_hws_requested;
1000 extern int __kmp_hws_abs_flag; // absolute or per-item number requested
1001 
1002 /* ------------------------------------------------------------------------ */
1003 
1004 #define KMP_PAD(type, sz) \
1005  (sizeof(type) + (sz - ((sizeof(type) - 1) % (sz)) - 1))
1006 
1007 // We need to avoid using -1 as a GTID as +1 is added to the gtid
1008 // when storing it in a lock, and the value 0 is reserved.
1009 #define KMP_GTID_DNE (-2) /* Does not exist */
1010 #define KMP_GTID_SHUTDOWN (-3) /* Library is shutting down */
1011 #define KMP_GTID_MONITOR (-4) /* Monitor thread ID */
1012 #define KMP_GTID_UNKNOWN (-5) /* Is not known */
1013 #define KMP_GTID_MIN (-6) /* Minimal gtid for low bound check in DEBUG */
1014 
1015 /* OpenMP 5.0 Memory Management support */
1016 
1017 #ifndef __OMP_H
1018 // Duplicate type definitions from omp.h
1019 typedef uintptr_t omp_uintptr_t;
1020 
1021 typedef enum {
1022  omp_atk_sync_hint = 1,
1023  omp_atk_alignment = 2,
1024  omp_atk_access = 3,
1025  omp_atk_pool_size = 4,
1026  omp_atk_fallback = 5,
1027  omp_atk_fb_data = 6,
1028  omp_atk_pinned = 7,
1029  omp_atk_partition = 8,
1030  omp_atk_pin_device = 9,
1031  omp_atk_preferred_device = 10,
1032  omp_atk_device_access = 11,
1033  omp_atk_target_access = 12,
1034  omp_atk_atomic_scope = 13,
1035  omp_atk_part_size = 14
1036 } omp_alloctrait_key_t;
1037 
1038 typedef enum {
1039  omp_atv_false = 0,
1040  omp_atv_true = 1,
1041  omp_atv_contended = 3,
1042  omp_atv_uncontended = 4,
1043  omp_atv_serialized = 5,
1044  omp_atv_sequential = omp_atv_serialized, // (deprecated)
1045  omp_atv_private = 6,
1046  omp_atv_device = 7,
1047  omp_atv_thread = 8,
1048  omp_atv_pteam = 9,
1049  omp_atv_cgroup = 10,
1050  omp_atv_default_mem_fb = 11,
1051  omp_atv_null_fb = 12,
1052  omp_atv_abort_fb = 13,
1053  omp_atv_allocator_fb = 14,
1054  omp_atv_environment = 15,
1055  omp_atv_nearest = 16,
1056  omp_atv_blocked = 17,
1057  omp_atv_interleaved = 18,
1058  omp_atv_all = 19,
1059  omp_atv_single = 20,
1060  omp_atv_multiple = 21,
1061  omp_atv_memspace = 22
1062 } omp_alloctrait_value_t;
1063 #define omp_atv_default ((omp_uintptr_t)-1)
1064 
1065 typedef void *omp_memspace_handle_t;
1066 extern omp_memspace_handle_t const omp_null_mem_space;
1067 extern omp_memspace_handle_t const omp_default_mem_space;
1068 extern omp_memspace_handle_t const omp_large_cap_mem_space;
1069 extern omp_memspace_handle_t const omp_const_mem_space;
1070 extern omp_memspace_handle_t const omp_high_bw_mem_space;
1071 extern omp_memspace_handle_t const omp_low_lat_mem_space;
1072 extern omp_memspace_handle_t const llvm_omp_target_host_mem_space;
1073 extern omp_memspace_handle_t const llvm_omp_target_shared_mem_space;
1074 extern omp_memspace_handle_t const llvm_omp_target_device_mem_space;
1075 extern omp_memspace_handle_t const kmp_max_mem_space;
1076 
1077 typedef struct {
1078  omp_alloctrait_key_t key;
1079  omp_uintptr_t value;
1080 } omp_alloctrait_t;
1081 
1082 typedef void *omp_allocator_handle_t;
1083 extern omp_allocator_handle_t const omp_null_allocator;
1084 extern omp_allocator_handle_t const omp_default_mem_alloc;
1085 extern omp_allocator_handle_t const omp_large_cap_mem_alloc;
1086 extern omp_allocator_handle_t const omp_const_mem_alloc;
1087 extern omp_allocator_handle_t const omp_high_bw_mem_alloc;
1088 extern omp_allocator_handle_t const omp_low_lat_mem_alloc;
1089 extern omp_allocator_handle_t const omp_cgroup_mem_alloc;
1090 extern omp_allocator_handle_t const omp_pteam_mem_alloc;
1091 extern omp_allocator_handle_t const omp_thread_mem_alloc;
1092 extern omp_allocator_handle_t const llvm_omp_target_host_mem_alloc;
1093 extern omp_allocator_handle_t const llvm_omp_target_shared_mem_alloc;
1094 extern omp_allocator_handle_t const llvm_omp_target_device_mem_alloc;
1095 extern omp_allocator_handle_t const kmp_max_mem_alloc;
1096 extern omp_allocator_handle_t __kmp_def_allocator;
1097 
1098 // end of duplicate type definitions from omp.h
1099 #endif
1100 
1101 extern int __kmp_memkind_available;
1102 extern bool __kmp_hwloc_available;
1103 
1105 typedef struct kmp_memspace_t {
1106  omp_memspace_handle_t memspace; // predefined input memory space
1107  int num_resources = 0; // number of available resources
1108  int *resources = nullptr; // available resources
1109  kmp_memspace_t *next = nullptr; // next memory space handle
1110 } kmp_memspace_t;
1111 
1113 typedef struct kmp_allocator_t {
1114  omp_memspace_handle_t memspace;
1115  void **memkind; // pointer to memkind
1116  size_t alignment;
1117  omp_alloctrait_value_t fb;
1118  kmp_allocator_t *fb_data;
1119  kmp_uint64 pool_size;
1120  kmp_uint64 pool_used;
1121  bool pinned;
1122  omp_alloctrait_value_t partition;
1123  int pin_device;
1124  int preferred_device;
1125  omp_alloctrait_value_t target_access;
1126  omp_alloctrait_value_t atomic_scope;
1127  size_t part_size;
1128 #if KMP_USE_HWLOC
1129  omp_alloctrait_value_t membind;
1130 #endif
1131 } kmp_allocator_t;
1132 
1133 extern omp_allocator_handle_t __kmpc_init_allocator(int gtid,
1134  omp_memspace_handle_t,
1135  int ntraits,
1136  omp_alloctrait_t traits[]);
1137 extern void __kmpc_destroy_allocator(int gtid, omp_allocator_handle_t al);
1138 extern void __kmpc_set_default_allocator(int gtid, omp_allocator_handle_t al);
1139 extern omp_allocator_handle_t __kmpc_get_default_allocator(int gtid);
1140 // external interfaces, may be used by compiler
1141 extern void *__kmpc_alloc(int gtid, size_t sz, omp_allocator_handle_t al);
1142 extern void *__kmpc_aligned_alloc(int gtid, size_t align, size_t sz,
1143  omp_allocator_handle_t al);
1144 extern void *__kmpc_calloc(int gtid, size_t nmemb, size_t sz,
1145  omp_allocator_handle_t al);
1146 extern void *__kmpc_realloc(int gtid, void *ptr, size_t sz,
1147  omp_allocator_handle_t al,
1148  omp_allocator_handle_t free_al);
1149 extern void __kmpc_free(int gtid, void *ptr, omp_allocator_handle_t al);
1150 // internal interfaces, contain real implementation
1151 extern void *__kmp_alloc(int gtid, size_t align, size_t sz,
1152  omp_allocator_handle_t al);
1153 extern void *__kmp_calloc(int gtid, size_t align, size_t nmemb, size_t sz,
1154  omp_allocator_handle_t al);
1155 extern void *__kmp_realloc(int gtid, void *ptr, size_t sz,
1156  omp_allocator_handle_t al,
1157  omp_allocator_handle_t free_al);
1158 extern void ___kmpc_free(int gtid, void *ptr, omp_allocator_handle_t al);
1159 
1160 extern void __kmp_init_memkind();
1161 extern void __kmp_fini_memkind();
1162 extern void __kmp_init_target_mem();
1163 extern void __kmp_fini_target_mem();
1164 
1165 // OpenMP 6.0 (TR11) Memory Management support
1166 extern omp_memspace_handle_t __kmp_get_devices_memspace(int ndevs,
1167  const int *devs,
1168  omp_memspace_handle_t,
1169  int host);
1170 extern omp_allocator_handle_t __kmp_get_devices_allocator(int ndevs,
1171  const int *devs,
1172  omp_memspace_handle_t,
1173  int host);
1174 extern int __kmp_get_memspace_num_resources(omp_memspace_handle_t memspace);
1175 extern omp_memspace_handle_t
1176 __kmp_get_submemspace(omp_memspace_handle_t memspace, int num_resources,
1177  int *resources);
1178 
1179 /* ------------------------------------------------------------------------ */
1180 
1181 #if ENABLE_LIBOMPTARGET
1182 extern void __kmp_init_target_task();
1183 #endif
1184 
1185 /* ------------------------------------------------------------------------ */
1186 
1187 #define KMP_UINT64_MAX \
1188  (~((kmp_uint64)1 << ((sizeof(kmp_uint64) * (1 << 3)) - 1)))
1189 
1190 #define KMP_MIN_NTH 1
1191 
1192 #ifndef KMP_MAX_NTH
1193 #if defined(PTHREAD_THREADS_MAX) && PTHREAD_THREADS_MAX < INT_MAX
1194 #define KMP_MAX_NTH PTHREAD_THREADS_MAX
1195 #else
1196 #ifdef __ve__
1197 // VE's pthread supports only up to 64 threads per a VE process.
1198 // Please check p. 14 of following documentation for more details.
1199 // https://sxauroratsubasa.sakura.ne.jp/documents/veos/en/VEOS_high_level_design.pdf
1200 #define KMP_MAX_NTH 64
1201 #else
1202 #define KMP_MAX_NTH INT_MAX
1203 #endif
1204 #endif
1205 #endif /* KMP_MAX_NTH */
1206 
1207 #ifdef PTHREAD_STACK_MIN
1208 #define KMP_MIN_STKSIZE ((size_t)PTHREAD_STACK_MIN)
1209 #else
1210 #define KMP_MIN_STKSIZE ((size_t)(32 * 1024))
1211 #endif
1212 
1213 #if KMP_OS_AIX && KMP_ARCH_PPC
1214 #define KMP_MAX_STKSIZE 0x10000000 /* 256Mb max size on 32-bit AIX */
1215 #else
1216 #define KMP_MAX_STKSIZE (~((size_t)1 << ((sizeof(size_t) * (1 << 3)) - 1)))
1217 #endif
1218 
1219 #if KMP_ARCH_X86
1220 #define KMP_DEFAULT_STKSIZE ((size_t)(2 * 1024 * 1024))
1221 #elif KMP_ARCH_X86_64
1222 #define KMP_DEFAULT_STKSIZE ((size_t)(4 * 1024 * 1024))
1223 #define KMP_BACKUP_STKSIZE ((size_t)(2 * 1024 * 1024))
1224 #elif KMP_ARCH_VE
1225 // Minimum stack size for pthread for VE is 4MB.
1226 // https://www.hpc.nec/documents/veos/en/glibc/Difference_Points_glibc.htm
1227 #define KMP_DEFAULT_STKSIZE ((size_t)(4 * 1024 * 1024))
1228 #elif KMP_OS_AIX
1229 // The default stack size for worker threads on AIX is 4MB.
1230 #define KMP_DEFAULT_STKSIZE ((size_t)(4 * 1024 * 1024))
1231 #else
1232 #define KMP_DEFAULT_STKSIZE ((size_t)(1024 * 1024))
1233 #endif
1234 
1235 #define KMP_DEFAULT_MALLOC_POOL_INCR ((size_t)(1024 * 1024))
1236 #define KMP_MIN_MALLOC_POOL_INCR ((size_t)(4 * 1024))
1237 #define KMP_MAX_MALLOC_POOL_INCR \
1238  (~((size_t)1 << ((sizeof(size_t) * (1 << 3)) - 1)))
1239 
1240 #define KMP_MIN_STKOFFSET (0)
1241 #define KMP_MAX_STKOFFSET KMP_MAX_STKSIZE
1242 #if KMP_OS_DARWIN
1243 #define KMP_DEFAULT_STKOFFSET KMP_MIN_STKOFFSET
1244 #else
1245 #define KMP_DEFAULT_STKOFFSET CACHE_LINE
1246 #endif
1247 
1248 #define KMP_MIN_STKPADDING (0)
1249 #define KMP_MAX_STKPADDING (2 * 1024 * 1024)
1250 
1251 #define KMP_BLOCKTIME_MULTIPLIER \
1252  (1000000) /* number of blocktime units per second */
1253 #define KMP_MIN_BLOCKTIME (0)
1254 #define KMP_MAX_BLOCKTIME \
1255  (INT_MAX) /* Must be this for "infinite" setting the work */
1256 
1257 /* __kmp_blocktime is in microseconds */
1258 #define KMP_DEFAULT_BLOCKTIME (__kmp_is_hybrid_cpu() ? (0) : (200000))
1259 
1260 #if KMP_USE_MONITOR
1261 #define KMP_DEFAULT_MONITOR_STKSIZE ((size_t)(64 * 1024))
1262 #define KMP_MIN_MONITOR_WAKEUPS (1) // min times monitor wakes up per second
1263 #define KMP_MAX_MONITOR_WAKEUPS (1000) // max times monitor can wake up per sec
1264 
1265 /* Calculate new number of monitor wakeups for a specific block time based on
1266  previous monitor_wakeups. Only allow increasing number of wakeups */
1267 #define KMP_WAKEUPS_FROM_BLOCKTIME(blocktime, monitor_wakeups) \
1268  (((blocktime) == KMP_MAX_BLOCKTIME) ? (monitor_wakeups) \
1269  : ((blocktime) == KMP_MIN_BLOCKTIME) ? KMP_MAX_MONITOR_WAKEUPS \
1270  : ((monitor_wakeups) > (KMP_BLOCKTIME_MULTIPLIER / (blocktime))) \
1271  ? (monitor_wakeups) \
1272  : (KMP_BLOCKTIME_MULTIPLIER) / (blocktime))
1273 
1274 /* Calculate number of intervals for a specific block time based on
1275  monitor_wakeups */
1276 #define KMP_INTERVALS_FROM_BLOCKTIME(blocktime, monitor_wakeups) \
1277  (((blocktime) + (KMP_BLOCKTIME_MULTIPLIER / (monitor_wakeups)) - 1) / \
1278  (KMP_BLOCKTIME_MULTIPLIER / (monitor_wakeups)))
1279 #else
1280 #define KMP_BLOCKTIME(team, tid) \
1281  (get__bt_set(team, tid) ? get__blocktime(team, tid) : __kmp_dflt_blocktime)
1282 #if KMP_OS_UNIX && (KMP_ARCH_X86 || KMP_ARCH_X86_64)
1283 // HW TSC is used to reduce overhead (clock tick instead of nanosecond).
1284 extern kmp_uint64 __kmp_ticks_per_msec;
1285 extern kmp_uint64 __kmp_ticks_per_usec;
1286 #if KMP_COMPILER_ICC || KMP_COMPILER_ICX
1287 #define KMP_NOW() ((kmp_uint64)_rdtsc())
1288 #else
1289 #define KMP_NOW() __kmp_hardware_timestamp()
1290 #endif
1291 #define KMP_BLOCKTIME_INTERVAL(team, tid) \
1292  ((kmp_uint64)KMP_BLOCKTIME(team, tid) * __kmp_ticks_per_usec)
1293 #define KMP_BLOCKING(goal, count) ((goal) > KMP_NOW())
1294 #else
1295 // System time is retrieved sporadically while blocking.
1296 extern kmp_uint64 __kmp_now_nsec();
1297 #define KMP_NOW() __kmp_now_nsec()
1298 #define KMP_BLOCKTIME_INTERVAL(team, tid) \
1299  ((kmp_uint64)KMP_BLOCKTIME(team, tid) * (kmp_uint64)KMP_NSEC_PER_USEC)
1300 #define KMP_BLOCKING(goal, count) ((count) % 1000 != 0 || (goal) > KMP_NOW())
1301 #endif
1302 #endif // KMP_USE_MONITOR
1303 
1304 #define KMP_MIN_STATSCOLS 40
1305 #define KMP_MAX_STATSCOLS 4096
1306 #define KMP_DEFAULT_STATSCOLS 80
1307 
1308 #define KMP_MIN_INTERVAL 0
1309 #define KMP_MAX_INTERVAL (INT_MAX - 1)
1310 #define KMP_DEFAULT_INTERVAL 0
1311 
1312 #define KMP_MIN_CHUNK 1
1313 #define KMP_MAX_CHUNK (INT_MAX - 1)
1314 #define KMP_DEFAULT_CHUNK 1
1315 
1316 #define KMP_MIN_DISP_NUM_BUFF 1
1317 #define KMP_DFLT_DISP_NUM_BUFF 7
1318 #define KMP_MAX_DISP_NUM_BUFF 4096
1319 
1320 #define KMP_MAX_ORDERED 8
1321 
1322 #define KMP_MAX_FIELDS 32
1323 
1324 #define KMP_MAX_BRANCH_BITS 31
1325 
1326 #define KMP_MAX_ACTIVE_LEVELS_LIMIT INT_MAX
1327 
1328 #define KMP_MAX_DEFAULT_DEVICE_LIMIT INT_MAX
1329 
1330 #define KMP_MAX_TASK_PRIORITY_LIMIT INT_MAX
1331 
1332 /* Minimum number of threads before switch to TLS gtid (experimentally
1333  determined) */
1334 /* josh TODO: what about OS X* tuning? */
1335 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
1336 #define KMP_TLS_GTID_MIN 5
1337 #else
1338 #define KMP_TLS_GTID_MIN INT_MAX
1339 #endif
1340 
1341 #define KMP_MASTER_TID(tid) (0 == (tid))
1342 #define KMP_WORKER_TID(tid) (0 != (tid))
1343 
1344 #define KMP_MASTER_GTID(gtid) (0 == __kmp_tid_from_gtid((gtid)))
1345 #define KMP_WORKER_GTID(gtid) (0 != __kmp_tid_from_gtid((gtid)))
1346 #define KMP_INITIAL_GTID(gtid) (0 == (gtid))
1347 
1348 #ifndef TRUE
1349 #define FALSE 0
1350 #define TRUE (!FALSE)
1351 #endif
1352 
1353 /* NOTE: all of the following constants must be even */
1354 
1355 #if KMP_OS_WINDOWS
1356 #define KMP_INIT_WAIT 64U /* initial number of spin-tests */
1357 #define KMP_NEXT_WAIT 32U /* susequent number of spin-tests */
1358 #elif KMP_OS_LINUX
1359 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1360 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1361 #elif KMP_OS_DARWIN
1362 /* TODO: tune for KMP_OS_DARWIN */
1363 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1364 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1365 #elif KMP_OS_DRAGONFLY
1366 /* TODO: tune for KMP_OS_DRAGONFLY */
1367 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1368 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1369 #elif KMP_OS_FREEBSD
1370 /* TODO: tune for KMP_OS_FREEBSD */
1371 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1372 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1373 #elif KMP_OS_NETBSD
1374 /* TODO: tune for KMP_OS_NETBSD */
1375 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1376 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1377 #elif KMP_OS_OPENBSD
1378 /* TODO: tune for KMP_OS_OPENBSD */
1379 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1380 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1381 #elif KMP_OS_HAIKU
1382 /* TODO: tune for KMP_OS_HAIKU */
1383 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1384 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1385 #elif KMP_OS_HURD
1386 /* TODO: tune for KMP_OS_HURD */
1387 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1388 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1389 #elif KMP_OS_SOLARIS
1390 /* TODO: tune for KMP_OS_SOLARIS */
1391 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1392 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1393 #elif KMP_OS_WASI
1394 /* TODO: tune for KMP_OS_WASI */
1395 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1396 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1397 #elif KMP_OS_AIX
1398 /* TODO: tune for KMP_OS_AIX */
1399 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1400 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1401 #endif
1402 
1403 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
1404 typedef struct kmp_cpuid {
1405  kmp_uint32 eax;
1406  kmp_uint32 ebx;
1407  kmp_uint32 ecx;
1408  kmp_uint32 edx;
1409 } kmp_cpuid_t;
1410 
1411 typedef struct kmp_cpuinfo_flags_t {
1412  unsigned sse2 : 1; // 0 if SSE2 instructions are not supported, 1 otherwise.
1413  unsigned rtm : 1; // 0 if RTM instructions are not supported, 1 otherwise.
1414  unsigned hybrid : 1;
1415  unsigned reserved : 29; // Ensure size of 32 bits
1416 } kmp_cpuinfo_flags_t;
1417 
1418 typedef struct kmp_cpuinfo {
1419  int initialized; // If 0, other fields are not initialized.
1420  int signature; // CPUID(1).EAX
1421  int family; // CPUID(1).EAX[27:20]+CPUID(1).EAX[11:8] (Extended Family+Family)
1422  int model; // ( CPUID(1).EAX[19:16] << 4 ) + CPUID(1).EAX[7:4] ( ( Extended
1423  // Model << 4 ) + Model)
1424  int stepping; // CPUID(1).EAX[3:0] ( Stepping )
1425  kmp_cpuinfo_flags_t flags;
1426  int apic_id;
1427  kmp_uint64 frequency; // Nominal CPU frequency in Hz.
1428  char name[3 * sizeof(kmp_cpuid_t)]; // CPUID(0x80000002,0x80000003,0x80000004)
1429 } kmp_cpuinfo_t;
1430 
1431 extern void __kmp_query_cpuid(kmp_cpuinfo_t *p);
1432 
1433 #if KMP_OS_UNIX
1434 // subleaf is only needed for cache and topology discovery and can be set to
1435 // zero in most cases
1436 static inline void __kmp_x86_cpuid(int leaf, int subleaf, struct kmp_cpuid *p) {
1437  __asm__ __volatile__("cpuid"
1438  : "=a"(p->eax), "=b"(p->ebx), "=c"(p->ecx), "=d"(p->edx)
1439  : "a"(leaf), "c"(subleaf));
1440 }
1441 // Load p into FPU control word
1442 static inline void __kmp_load_x87_fpu_control_word(const kmp_int16 *p) {
1443  __asm__ __volatile__("fldcw %0" : : "m"(*p));
1444 }
1445 // Store FPU control word into p
1446 static inline void __kmp_store_x87_fpu_control_word(kmp_int16 *p) {
1447  __asm__ __volatile__("fstcw %0" : "=m"(*p));
1448 }
1449 static inline void __kmp_clear_x87_fpu_status_word() {
1450 #if KMP_MIC
1451  // 32-bit protected mode x87 FPU state
1452  struct x87_fpu_state {
1453  unsigned cw;
1454  unsigned sw;
1455  unsigned tw;
1456  unsigned fip;
1457  unsigned fips;
1458  unsigned fdp;
1459  unsigned fds;
1460  };
1461  struct x87_fpu_state fpu_state = {0, 0, 0, 0, 0, 0, 0};
1462  __asm__ __volatile__("fstenv %0\n\t" // store FP env
1463  "andw $0x7f00, %1\n\t" // clear 0-7,15 bits of FP SW
1464  "fldenv %0\n\t" // load FP env back
1465  : "+m"(fpu_state), "+m"(fpu_state.sw));
1466 #else
1467  __asm__ __volatile__("fnclex");
1468 #endif // KMP_MIC
1469 }
1470 #if __SSE__
1471 static inline void __kmp_load_mxcsr(const kmp_uint32 *p) { _mm_setcsr(*p); }
1472 static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = _mm_getcsr(); }
1473 #else
1474 static inline void __kmp_load_mxcsr(const kmp_uint32 *p) {}
1475 static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = 0; }
1476 #endif
1477 #else
1478 // Windows still has these as external functions in assembly file
1479 extern void __kmp_x86_cpuid(int mode, int mode2, struct kmp_cpuid *p);
1480 extern void __kmp_load_x87_fpu_control_word(const kmp_int16 *p);
1481 extern void __kmp_store_x87_fpu_control_word(kmp_int16 *p);
1482 extern void __kmp_clear_x87_fpu_status_word();
1483 static inline void __kmp_load_mxcsr(const kmp_uint32 *p) { _mm_setcsr(*p); }
1484 static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = _mm_getcsr(); }
1485 #endif // KMP_OS_UNIX
1486 
1487 #define KMP_X86_MXCSR_MASK 0xffffffc0 /* ignore status flags (6 lsb) */
1488 
1489 // User-level Monitor/Mwait
1490 #if KMP_HAVE_UMWAIT
1491 // We always try for UMWAIT first
1492 #if KMP_HAVE_WAITPKG_INTRINSICS
1493 #if KMP_HAVE_IMMINTRIN_H
1494 #include <immintrin.h>
1495 #elif KMP_HAVE_INTRIN_H
1496 #include <intrin.h>
1497 #endif
1498 #endif // KMP_HAVE_WAITPKG_INTRINSICS
1499 
1500 KMP_ATTRIBUTE_TARGET_WAITPKG
1501 static inline int __kmp_tpause(uint32_t hint, uint64_t counter) {
1502 #if !KMP_HAVE_WAITPKG_INTRINSICS
1503  uint32_t timeHi = uint32_t(counter >> 32);
1504  uint32_t timeLo = uint32_t(counter & 0xffffffff);
1505  char flag;
1506  __asm__ volatile("#tpause\n.byte 0x66, 0x0F, 0xAE, 0xF1\n"
1507  "setb %0"
1508  // The "=q" restraint means any register accessible as rl
1509  // in 32-bit mode: a, b, c, and d;
1510  // in 64-bit mode: any integer register
1511  : "=q"(flag)
1512  : "a"(timeLo), "d"(timeHi), "c"(hint)
1513  :);
1514  return flag;
1515 #else
1516  return _tpause(hint, counter);
1517 #endif
1518 }
1519 KMP_ATTRIBUTE_TARGET_WAITPKG
1520 static inline void __kmp_umonitor(void *cacheline) {
1521 #if !KMP_HAVE_WAITPKG_INTRINSICS
1522  __asm__ volatile("# umonitor\n.byte 0xF3, 0x0F, 0xAE, 0x01 "
1523  :
1524  : "a"(cacheline)
1525  :);
1526 #else
1527  _umonitor(cacheline);
1528 #endif
1529 }
1530 KMP_ATTRIBUTE_TARGET_WAITPKG
1531 static inline int __kmp_umwait(uint32_t hint, uint64_t counter) {
1532 #if !KMP_HAVE_WAITPKG_INTRINSICS
1533  uint32_t timeHi = uint32_t(counter >> 32);
1534  uint32_t timeLo = uint32_t(counter & 0xffffffff);
1535  char flag;
1536  __asm__ volatile("#umwait\n.byte 0xF2, 0x0F, 0xAE, 0xF1\n"
1537  "setb %0"
1538  // The "=q" restraint means any register accessible as rl
1539  // in 32-bit mode: a, b, c, and d;
1540  // in 64-bit mode: any integer register
1541  : "=q"(flag)
1542  : "a"(timeLo), "d"(timeHi), "c"(hint)
1543  :);
1544  return flag;
1545 #else
1546  return _umwait(hint, counter);
1547 #endif
1548 }
1549 #elif KMP_HAVE_MWAIT
1550 #if KMP_OS_UNIX
1551 #include <pmmintrin.h>
1552 #else
1553 #include <intrin.h>
1554 #endif
1555 #if KMP_OS_UNIX
1556 __attribute__((target("sse3")))
1557 #endif
1558 static inline void
1559 __kmp_mm_monitor(void *cacheline, unsigned extensions, unsigned hints) {
1560  _mm_monitor(cacheline, extensions, hints);
1561 }
1562 #if KMP_OS_UNIX
1563 __attribute__((target("sse3")))
1564 #endif
1565 static inline void
1566 __kmp_mm_mwait(unsigned extensions, unsigned hints) {
1567  _mm_mwait(extensions, hints);
1568 }
1569 #endif // KMP_HAVE_UMWAIT
1570 
1571 #if KMP_ARCH_X86
1572 extern void __kmp_x86_pause(void);
1573 #elif KMP_MIC
1574 // Performance testing on KNC (C0QS-7120 P/A/X/D, 61-core, 16 GB Memory) showed
1575 // regression after removal of extra PAUSE from spin loops. Changing
1576 // the delay from 100 to 300 showed even better performance than double PAUSE
1577 // on Spec OMP2001 and LCPC tasking tests, no regressions on EPCC.
1578 static inline void __kmp_x86_pause(void) { _mm_delay_32(300); }
1579 #else
1580 static inline void __kmp_x86_pause(void) { _mm_pause(); }
1581 #endif
1582 #define KMP_CPU_PAUSE() __kmp_x86_pause()
1583 #elif KMP_ARCH_PPC64
1584 #define KMP_PPC64_PRI_LOW() __asm__ volatile("or 1, 1, 1")
1585 #define KMP_PPC64_PRI_MED() __asm__ volatile("or 2, 2, 2")
1586 #define KMP_PPC64_PRI_LOC_MB() __asm__ volatile("" : : : "memory")
1587 #define KMP_CPU_PAUSE() \
1588  do { \
1589  KMP_PPC64_PRI_LOW(); \
1590  KMP_PPC64_PRI_MED(); \
1591  KMP_PPC64_PRI_LOC_MB(); \
1592  } while (0)
1593 #else
1594 #define KMP_CPU_PAUSE() /* nothing to do */
1595 #endif
1596 
1597 #define KMP_INIT_YIELD(count) \
1598  { (count) = __kmp_yield_init; }
1599 
1600 #define KMP_INIT_BACKOFF(time) \
1601  { (time) = __kmp_pause_init; }
1602 
1603 #define KMP_OVERSUBSCRIBED \
1604  (TCR_4(__kmp_nth) > (__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc))
1605 
1606 #define KMP_TRY_YIELD \
1607  ((__kmp_use_yield == 1) || (__kmp_use_yield == 2 && (KMP_OVERSUBSCRIBED)))
1608 
1609 #define KMP_TRY_YIELD_OVERSUB \
1610  ((__kmp_use_yield == 1 || __kmp_use_yield == 2) && (KMP_OVERSUBSCRIBED))
1611 
1612 #define KMP_YIELD(cond) \
1613  { \
1614  KMP_CPU_PAUSE(); \
1615  if ((cond) && (KMP_TRY_YIELD)) \
1616  __kmp_yield(); \
1617  }
1618 
1619 #define KMP_YIELD_OVERSUB() \
1620  { \
1621  KMP_CPU_PAUSE(); \
1622  if ((KMP_TRY_YIELD_OVERSUB)) \
1623  __kmp_yield(); \
1624  }
1625 
1626 // Note the decrement of 2 in the following Macros. With KMP_LIBRARY=turnaround,
1627 // there should be no yielding since initial value from KMP_INIT_YIELD() is odd.
1628 #define KMP_YIELD_SPIN(count) \
1629  { \
1630  KMP_CPU_PAUSE(); \
1631  if (KMP_TRY_YIELD) { \
1632  (count) -= 2; \
1633  if (!(count)) { \
1634  __kmp_yield(); \
1635  (count) = __kmp_yield_next; \
1636  } \
1637  } \
1638  }
1639 
1640 // If TPAUSE is available & enabled, use it. If oversubscribed, use the slower
1641 // (C0.2) state, which improves performance of other SMT threads on the same
1642 // core, otherwise, use the fast (C0.1) default state, or whatever the user has
1643 // requested. Uses a timed TPAUSE, and exponential backoff. If TPAUSE isn't
1644 // available, fall back to the regular CPU pause and yield combination.
1645 #if KMP_HAVE_UMWAIT
1646 #define KMP_TPAUSE_MAX_MASK ((kmp_uint64)0xFFFF)
1647 #define KMP_YIELD_OVERSUB_ELSE_SPIN(count, time) \
1648  { \
1649  if (__kmp_tpause_enabled) { \
1650  if (KMP_OVERSUBSCRIBED) { \
1651  __kmp_tpause(0, (time)); \
1652  } else { \
1653  __kmp_tpause(__kmp_tpause_hint, (time)); \
1654  } \
1655  (time) = (time << 1 | 1) & KMP_TPAUSE_MAX_MASK; \
1656  } else { \
1657  KMP_CPU_PAUSE(); \
1658  if ((KMP_TRY_YIELD_OVERSUB)) { \
1659  __kmp_yield(); \
1660  } else if (__kmp_use_yield == 1) { \
1661  (count) -= 2; \
1662  if (!(count)) { \
1663  __kmp_yield(); \
1664  (count) = __kmp_yield_next; \
1665  } \
1666  } \
1667  } \
1668  }
1669 #else
1670 #define KMP_YIELD_OVERSUB_ELSE_SPIN(count, time) \
1671  { \
1672  KMP_CPU_PAUSE(); \
1673  if ((KMP_TRY_YIELD_OVERSUB)) \
1674  __kmp_yield(); \
1675  else if (__kmp_use_yield == 1) { \
1676  (count) -= 2; \
1677  if (!(count)) { \
1678  __kmp_yield(); \
1679  (count) = __kmp_yield_next; \
1680  } \
1681  } \
1682  }
1683 #endif // KMP_HAVE_UMWAIT
1684 
1685 /* ------------------------------------------------------------------------ */
1686 /* Support datatypes for the orphaned construct nesting checks. */
1687 /* ------------------------------------------------------------------------ */
1688 
1689 /* When adding to this enum, add its corresponding string in cons_text_c[]
1690  * array in kmp_error.cpp */
1691 enum cons_type {
1692  ct_none,
1693  ct_parallel,
1694  ct_pdo,
1695  ct_pdo_ordered,
1696  ct_psections,
1697  ct_psingle,
1698  ct_critical,
1699  ct_ordered_in_parallel,
1700  ct_ordered_in_pdo,
1701  ct_master,
1702  ct_reduce,
1703  ct_barrier,
1704  ct_masked
1705 };
1706 
1707 #define IS_CONS_TYPE_ORDERED(ct) ((ct) == ct_pdo_ordered)
1708 
1709 struct cons_data {
1710  ident_t const *ident;
1711  enum cons_type type;
1712  int prev;
1713  kmp_user_lock_p
1714  name; /* address exclusively for critical section name comparison */
1715 };
1716 
1717 struct cons_header {
1718  int p_top, w_top, s_top;
1719  int stack_size, stack_top;
1720  struct cons_data *stack_data;
1721 };
1722 
1723 struct kmp_region_info {
1724  char *text;
1725  int offset[KMP_MAX_FIELDS];
1726  int length[KMP_MAX_FIELDS];
1727 };
1728 
1729 /* ---------------------------------------------------------------------- */
1730 /* ---------------------------------------------------------------------- */
1731 
1732 #if KMP_OS_WINDOWS
1733 typedef HANDLE kmp_thread_t;
1734 typedef DWORD kmp_key_t;
1735 #endif /* KMP_OS_WINDOWS */
1736 
1737 #if KMP_OS_UNIX
1738 typedef pthread_t kmp_thread_t;
1739 typedef pthread_key_t kmp_key_t;
1740 #endif
1741 
1742 extern kmp_key_t __kmp_gtid_threadprivate_key;
1743 
1744 typedef struct kmp_sys_info {
1745  long maxrss; /* the maximum resident set size utilized (in kilobytes) */
1746  long minflt; /* the number of page faults serviced without any I/O */
1747  long majflt; /* the number of page faults serviced that required I/O */
1748  long nswap; /* the number of times a process was "swapped" out of memory */
1749  long inblock; /* the number of times the file system had to perform input */
1750  long oublock; /* the number of times the file system had to perform output */
1751  long nvcsw; /* the number of times a context switch was voluntarily */
1752  long nivcsw; /* the number of times a context switch was forced */
1753 } kmp_sys_info_t;
1754 
1755 #if USE_ITT_BUILD
1756 // We cannot include "kmp_itt.h" due to circular dependency. Declare the only
1757 // required type here. Later we will check the type meets requirements.
1758 typedef int kmp_itt_mark_t;
1759 #define KMP_ITT_DEBUG 0
1760 #endif /* USE_ITT_BUILD */
1761 
1762 typedef kmp_int32 kmp_critical_name[8];
1763 
1773 typedef void (*kmpc_micro)(kmp_int32 *global_tid, kmp_int32 *bound_tid, ...);
1774 typedef void (*kmpc_micro_bound)(kmp_int32 *bound_tid, kmp_int32 *bound_nth,
1775  ...);
1776 
1781 /* ---------------------------------------------------------------------------
1782  */
1783 /* Threadprivate initialization/finalization function declarations */
1784 
1785 /* for non-array objects: __kmpc_threadprivate_register() */
1786 
1791 typedef void *(*kmpc_ctor)(void *);
1792 
1797 typedef void (*kmpc_dtor)(
1798  void * /*, size_t */); /* 2nd arg: magic number for KCC unused by Intel
1799  compiler */
1804 typedef void *(*kmpc_cctor)(void *, void *);
1805 
1806 /* for array objects: __kmpc_threadprivate_register_vec() */
1807 /* First arg: "this" pointer */
1808 /* Last arg: number of array elements */
1814 typedef void *(*kmpc_ctor_vec)(void *, size_t);
1820 typedef void (*kmpc_dtor_vec)(void *, size_t);
1826 typedef void *(*kmpc_cctor_vec)(void *, void *,
1827  size_t); /* function unused by compiler */
1828 
1833 /* keeps tracked of threadprivate cache allocations for cleanup later */
1834 typedef struct kmp_cached_addr {
1835  void **addr; /* address of allocated cache */
1836  void ***compiler_cache; /* pointer to compiler's cache */
1837  void *data; /* pointer to global data */
1838  struct kmp_cached_addr *next; /* pointer to next cached address */
1839 } kmp_cached_addr_t;
1840 
1841 struct private_data {
1842  struct private_data *next; /* The next descriptor in the list */
1843  void *data; /* The data buffer for this descriptor */
1844  int more; /* The repeat count for this descriptor */
1845  size_t size; /* The data size for this descriptor */
1846 };
1847 
1848 struct private_common {
1849  struct private_common *next;
1850  struct private_common *link;
1851  void *gbl_addr;
1852  void *par_addr; /* par_addr == gbl_addr for PRIMARY thread */
1853  size_t cmn_size;
1854 };
1855 
1856 struct shared_common {
1857  struct shared_common *next;
1858  struct private_data *pod_init;
1859  void *obj_init;
1860  void *gbl_addr;
1861  union {
1862  kmpc_ctor ctor;
1863  kmpc_ctor_vec ctorv;
1864  } ct;
1865  union {
1866  kmpc_cctor cctor;
1867  kmpc_cctor_vec cctorv;
1868  } cct;
1869  union {
1870  kmpc_dtor dtor;
1871  kmpc_dtor_vec dtorv;
1872  } dt;
1873  size_t vec_len;
1874  int is_vec;
1875  size_t cmn_size;
1876 };
1877 
1878 #define KMP_HASH_TABLE_LOG2 9 /* log2 of the hash table size */
1879 #define KMP_HASH_TABLE_SIZE \
1880  (1 << KMP_HASH_TABLE_LOG2) /* size of the hash table */
1881 #define KMP_HASH_SHIFT 3 /* throw away this many low bits from the address */
1882 #define KMP_HASH(x) \
1883  ((((kmp_uintptr_t)x) >> KMP_HASH_SHIFT) & (KMP_HASH_TABLE_SIZE - 1))
1884 
1885 struct common_table {
1886  struct private_common *data[KMP_HASH_TABLE_SIZE];
1887 };
1888 
1889 struct shared_table {
1890  struct shared_common *data[KMP_HASH_TABLE_SIZE];
1891 };
1892 
1893 /* ------------------------------------------------------------------------ */
1894 
1895 #if KMP_USE_HIER_SCHED
1896 // Shared barrier data that exists inside a single unit of the scheduling
1897 // hierarchy
1898 typedef struct kmp_hier_private_bdata_t {
1899  kmp_int32 num_active;
1900  kmp_uint64 index;
1901  kmp_uint64 wait_val[2];
1902 } kmp_hier_private_bdata_t;
1903 #endif
1904 
1905 typedef struct kmp_sched_flags {
1906  unsigned ordered : 1;
1907  unsigned nomerge : 1;
1908  unsigned contains_last : 1;
1909  unsigned use_hier : 1; // Used in KMP_USE_HIER_SCHED code
1910  unsigned use_hybrid : 1; // Used in KMP_WEIGHTED_ITERATIONS_SUPPORTED code
1911  unsigned unused : 27;
1912 } kmp_sched_flags_t;
1913 
1914 KMP_BUILD_ASSERT(sizeof(kmp_sched_flags_t) == 4);
1915 
1916 #if KMP_STATIC_STEAL_ENABLED
1917 typedef struct KMP_ALIGN_CACHE dispatch_private_info32 {
1918  kmp_int32 count;
1919  kmp_int32 ub;
1920  /* Adding KMP_ALIGN_CACHE here doesn't help / can hurt performance */
1921  kmp_int32 lb;
1922  kmp_int32 st;
1923  kmp_int32 tc;
1924  kmp_lock_t *steal_lock; // lock used for chunk stealing
1925 
1926  kmp_uint32 ordered_lower;
1927  kmp_uint32 ordered_upper;
1928 
1929  // KMP_ALIGN(32) ensures (if the KMP_ALIGN macro is turned on)
1930  // a) parm3 is properly aligned and
1931  // b) all parm1-4 are on the same cache line.
1932  // Because of parm1-4 are used together, performance seems to be better
1933  // if they are on the same cache line (not measured though).
1934 
1935  struct KMP_ALIGN(32) {
1936  kmp_int32 parm1;
1937  kmp_int32 parm2;
1938  kmp_int32 parm3;
1939  kmp_int32 parm4;
1940  };
1941 
1942 #if KMP_WEIGHTED_ITERATIONS_SUPPORTED
1943  kmp_uint32 pchunks;
1944  kmp_uint32 num_procs_with_pcore;
1945  kmp_int32 first_thread_with_ecore;
1946 #endif
1947 #if KMP_OS_WINDOWS
1948  kmp_int32 last_upper;
1949 #endif /* KMP_OS_WINDOWS */
1950 } dispatch_private_info32_t;
1951 
1952 #if CACHE_LINE <= 128
1953 KMP_BUILD_ASSERT(sizeof(dispatch_private_info32_t) <= 128);
1954 #endif
1955 
1956 typedef struct KMP_ALIGN_CACHE dispatch_private_info64 {
1957  kmp_int64 count; // current chunk number for static & static-steal scheduling
1958  kmp_int64 ub; /* upper-bound */
1959  /* Adding KMP_ALIGN_CACHE here doesn't help / can hurt performance */
1960  kmp_int64 lb; /* lower-bound */
1961  kmp_int64 st; /* stride */
1962  kmp_int64 tc; /* trip count (number of iterations) */
1963  kmp_lock_t *steal_lock; // lock used for chunk stealing
1964 
1965  kmp_uint64 ordered_lower;
1966  kmp_uint64 ordered_upper;
1967  /* parm[1-4] are used in different ways by different scheduling algorithms */
1968 
1969  // KMP_ALIGN(32) ensures ( if the KMP_ALIGN macro is turned on )
1970  // a) parm3 is properly aligned and
1971  // b) all parm1-4 are in the same cache line.
1972  // Because of parm1-4 are used together, performance seems to be better
1973  // if they are in the same line (not measured though).
1974  struct KMP_ALIGN(32) {
1975  kmp_int64 parm1;
1976  kmp_int64 parm2;
1977  kmp_int64 parm3;
1978  kmp_int64 parm4;
1979  };
1980 
1981 #if KMP_WEIGHTED_ITERATIONS_SUPPORTED
1982  kmp_uint64 pchunks;
1983  kmp_uint64 num_procs_with_pcore;
1984  kmp_int64 first_thread_with_ecore;
1985 #endif
1986 
1987 #if KMP_OS_WINDOWS
1988  kmp_int64 last_upper;
1989 #endif /* KMP_OS_WINDOWS */
1990 } dispatch_private_info64_t;
1991 
1992 #if CACHE_LINE <= 128
1993 KMP_BUILD_ASSERT(sizeof(dispatch_private_info64_t) <= 128);
1994 #endif
1995 
1996 #else /* KMP_STATIC_STEAL_ENABLED */
1997 typedef struct KMP_ALIGN_CACHE dispatch_private_info32 {
1998  kmp_int32 lb;
1999  kmp_int32 ub;
2000  kmp_int32 st;
2001  kmp_int32 tc;
2002 
2003  kmp_int32 parm1;
2004  kmp_int32 parm2;
2005  kmp_int32 parm3;
2006  kmp_int32 parm4;
2007 
2008  kmp_int32 count;
2009 
2010  kmp_uint32 ordered_lower;
2011  kmp_uint32 ordered_upper;
2012 #if KMP_OS_WINDOWS
2013  kmp_int32 last_upper;
2014 #endif /* KMP_OS_WINDOWS */
2015 } dispatch_private_info32_t;
2016 
2017 typedef struct KMP_ALIGN_CACHE dispatch_private_info64 {
2018  kmp_int64 lb; /* lower-bound */
2019  kmp_int64 ub; /* upper-bound */
2020  kmp_int64 st; /* stride */
2021  kmp_int64 tc; /* trip count (number of iterations) */
2022 
2023  /* parm[1-4] are used in different ways by different scheduling algorithms */
2024  kmp_int64 parm1;
2025  kmp_int64 parm2;
2026  kmp_int64 parm3;
2027  kmp_int64 parm4;
2028 
2029  kmp_int64 count; /* current chunk number for static scheduling */
2030 
2031  kmp_uint64 ordered_lower;
2032  kmp_uint64 ordered_upper;
2033 #if KMP_OS_WINDOWS
2034  kmp_int64 last_upper;
2035 #endif /* KMP_OS_WINDOWS */
2036 } dispatch_private_info64_t;
2037 #endif /* KMP_STATIC_STEAL_ENABLED */
2038 
2039 typedef struct KMP_ALIGN_CACHE dispatch_private_info {
2040  union private_info {
2041  dispatch_private_info32_t p32;
2042  dispatch_private_info64_t p64;
2043  } u;
2044  enum sched_type schedule; /* scheduling algorithm */
2045  kmp_sched_flags_t flags; /* flags (e.g., ordered, nomerge, etc.) */
2046  std::atomic<kmp_uint32> steal_flag; // static_steal only, state of a buffer
2047  kmp_int32 ordered_bumped;
2048  // Stack of buffers for nest of serial regions
2049  struct dispatch_private_info *next;
2050  kmp_int32 type_size; /* the size of types in private_info */
2051 #if KMP_USE_HIER_SCHED
2052  kmp_int32 hier_id;
2053  void *parent; /* hierarchical scheduling parent pointer */
2054 #endif
2055  enum cons_type pushed_ws;
2056 } dispatch_private_info_t;
2057 
2058 typedef struct dispatch_shared_info32 {
2059  /* chunk index under dynamic, number of idle threads under static-steal;
2060  iteration index otherwise */
2061  volatile kmp_uint32 iteration;
2062  volatile kmp_int32 num_done;
2063  volatile kmp_uint32 ordered_iteration;
2064  // Dummy to retain the structure size after making ordered_iteration scalar
2065  kmp_int32 ordered_dummy[KMP_MAX_ORDERED - 1];
2066 } dispatch_shared_info32_t;
2067 
2068 typedef struct dispatch_shared_info64 {
2069  /* chunk index under dynamic, number of idle threads under static-steal;
2070  iteration index otherwise */
2071  volatile kmp_uint64 iteration;
2072  volatile kmp_int64 num_done;
2073  volatile kmp_uint64 ordered_iteration;
2074  // Dummy to retain the structure size after making ordered_iteration scalar
2075  kmp_int64 ordered_dummy[KMP_MAX_ORDERED - 3];
2076 } dispatch_shared_info64_t;
2077 
2078 typedef struct dispatch_shared_info {
2079  union shared_info {
2080  dispatch_shared_info32_t s32;
2081  dispatch_shared_info64_t s64;
2082  } u;
2083  volatile kmp_uint32 buffer_index;
2084  volatile kmp_int32 doacross_buf_idx; // teamwise index
2085  volatile kmp_uint32 *doacross_flags; // shared array of iteration flags (0/1)
2086  kmp_int32 doacross_num_done; // count finished threads
2087 #if KMP_USE_HIER_SCHED
2088  void *hier;
2089 #endif
2090 #if KMP_USE_HWLOC
2091  // When linking with libhwloc, the ORDERED EPCC test slows down on big
2092  // machines (> 48 cores). Performance analysis showed that a cache thrash
2093  // was occurring and this padding helps alleviate the problem.
2094  char padding[64];
2095 #endif
2096 } dispatch_shared_info_t;
2097 
2098 typedef struct kmp_disp {
2099  /* Vector for ORDERED SECTION */
2100  void (*th_deo_fcn)(int *gtid, int *cid, ident_t *);
2101  /* Vector for END ORDERED SECTION */
2102  void (*th_dxo_fcn)(int *gtid, int *cid, ident_t *);
2103 
2104  dispatch_shared_info_t *th_dispatch_sh_current;
2105  dispatch_private_info_t *th_dispatch_pr_current;
2106 
2107  dispatch_private_info_t *th_disp_buffer;
2108  kmp_uint32 th_disp_index;
2109  kmp_int32 th_doacross_buf_idx; // thread's doacross buffer index
2110  volatile kmp_uint32 *th_doacross_flags; // pointer to shared array of flags
2111  kmp_int64 *th_doacross_info; // info on loop bounds
2112 #if KMP_USE_INTERNODE_ALIGNMENT
2113  char more_padding[INTERNODE_CACHE_LINE];
2114 #endif
2115 } kmp_disp_t;
2116 
2117 /* ------------------------------------------------------------------------ */
2118 /* Barrier stuff */
2119 
2120 /* constants for barrier state update */
2121 #define KMP_INIT_BARRIER_STATE 0 /* should probably start from zero */
2122 #define KMP_BARRIER_SLEEP_BIT 0 /* bit used for suspend/sleep part of state */
2123 #define KMP_BARRIER_UNUSED_BIT 1 // bit that must never be set for valid state
2124 #define KMP_BARRIER_BUMP_BIT 2 /* lsb used for bump of go/arrived state */
2125 
2126 #define KMP_BARRIER_SLEEP_STATE (1 << KMP_BARRIER_SLEEP_BIT)
2127 #define KMP_BARRIER_UNUSED_STATE (1 << KMP_BARRIER_UNUSED_BIT)
2128 #define KMP_BARRIER_STATE_BUMP (1 << KMP_BARRIER_BUMP_BIT)
2129 
2130 #if (KMP_BARRIER_SLEEP_BIT >= KMP_BARRIER_BUMP_BIT)
2131 #error "Barrier sleep bit must be smaller than barrier bump bit"
2132 #endif
2133 #if (KMP_BARRIER_UNUSED_BIT >= KMP_BARRIER_BUMP_BIT)
2134 #error "Barrier unused bit must be smaller than barrier bump bit"
2135 #endif
2136 
2137 // Constants for release barrier wait state: currently, hierarchical only
2138 #define KMP_BARRIER_NOT_WAITING 0 // Normal state; worker not in wait_sleep
2139 #define KMP_BARRIER_OWN_FLAG \
2140  1 // Normal state; worker waiting on own b_go flag in release
2141 #define KMP_BARRIER_PARENT_FLAG \
2142  2 // Special state; worker waiting on parent's b_go flag in release
2143 #define KMP_BARRIER_SWITCH_TO_OWN_FLAG \
2144  3 // Special state; tells worker to shift from parent to own b_go
2145 #define KMP_BARRIER_SWITCHING \
2146  4 // Special state; worker resets appropriate flag on wake-up
2147 
2148 #define KMP_NOT_SAFE_TO_REAP \
2149  0 // Thread th_reap_state: not safe to reap (tasking)
2150 #define KMP_SAFE_TO_REAP 1 // Thread th_reap_state: safe to reap (not tasking)
2151 
2152 // The flag_type describes the storage used for the flag.
2153 enum flag_type {
2154  flag32,
2155  flag64,
2156  atomic_flag64,
2157  flag_oncore,
2158  flag_unset
2159 };
2160 
2161 enum barrier_type {
2162  bs_plain_barrier = 0, /* 0, All non-fork/join barriers (except reduction
2163  barriers if enabled) */
2164  bs_forkjoin_barrier, /* 1, All fork/join (parallel region) barriers */
2165 #if KMP_FAST_REDUCTION_BARRIER
2166  bs_reduction_barrier, /* 2, All barriers that are used in reduction */
2167 #endif // KMP_FAST_REDUCTION_BARRIER
2168  bs_last_barrier /* Just a placeholder to mark the end */
2169 };
2170 
2171 // to work with reduction barriers just like with plain barriers
2172 #if !KMP_FAST_REDUCTION_BARRIER
2173 #define bs_reduction_barrier bs_plain_barrier
2174 #endif // KMP_FAST_REDUCTION_BARRIER
2175 
2176 typedef enum kmp_bar_pat { /* Barrier communication patterns */
2177  bp_linear_bar =
2178  0, /* Single level (degenerate) tree */
2179  bp_tree_bar =
2180  1, /* Balanced tree with branching factor 2^n */
2181  bp_hyper_bar = 2, /* Hypercube-embedded tree with min
2182  branching factor 2^n */
2183  bp_hierarchical_bar = 3, /* Machine hierarchy tree */
2184  bp_dist_bar = 4, /* Distributed barrier */
2185  bp_last_bar /* Placeholder to mark the end */
2186 } kmp_bar_pat_e;
2187 
2188 #define KMP_BARRIER_ICV_PUSH 1
2189 
2190 /* Record for holding the values of the internal controls stack records */
2191 typedef struct kmp_internal_control {
2192  int serial_nesting_level; /* corresponds to the value of the
2193  th_team_serialized field */
2194  kmp_int8 dynamic; /* internal control for dynamic adjustment of threads (per
2195  thread) */
2196  kmp_int8
2197  bt_set; /* internal control for whether blocktime is explicitly set */
2198  int blocktime; /* internal control for blocktime */
2199 #if KMP_USE_MONITOR
2200  int bt_intervals; /* internal control for blocktime intervals */
2201 #endif
2202  int nproc; /* internal control for #threads for next parallel region (per
2203  thread) */
2204  int thread_limit; /* internal control for thread-limit-var */
2205  int task_thread_limit; /* internal control for thread-limit-var of a task*/
2206  int max_active_levels; /* internal control for max_active_levels */
2207  kmp_r_sched_t
2208  sched; /* internal control for runtime schedule {sched,chunk} pair */
2209  kmp_proc_bind_t proc_bind; /* internal control for affinity */
2210  kmp_int32 default_device; /* internal control for default device */
2211  struct kmp_internal_control *next;
2212 } kmp_internal_control_t;
2213 
2214 static inline void copy_icvs(kmp_internal_control_t *dst,
2215  kmp_internal_control_t *src) {
2216  *dst = *src;
2217 }
2218 
2219 /* Thread barrier needs volatile barrier fields */
2220 typedef struct KMP_ALIGN_CACHE kmp_bstate {
2221  // th_fixed_icvs is aligned by virtue of kmp_bstate being aligned (and all
2222  // uses of it). It is not explicitly aligned below, because we *don't* want
2223  // it to be padded -- instead, we fit b_go into the same cache line with
2224  // th_fixed_icvs, enabling NGO cache lines stores in the hierarchical barrier.
2225  kmp_internal_control_t th_fixed_icvs; // Initial ICVs for the thread
2226  // Tuck b_go into end of th_fixed_icvs cache line, so it can be stored with
2227  // same NGO store
2228  volatile kmp_uint64 b_go; // STATE => task should proceed (hierarchical)
2229  KMP_ALIGN_CACHE volatile kmp_uint64
2230  b_arrived; // STATE => task reached synch point.
2231  kmp_uint32 *skip_per_level;
2232  kmp_uint32 my_level;
2233  kmp_int32 parent_tid;
2234  kmp_int32 old_tid;
2235  kmp_uint32 depth;
2236  struct kmp_bstate *parent_bar;
2237  kmp_team_t *team;
2238  kmp_uint64 leaf_state;
2239  kmp_uint32 nproc;
2240  kmp_uint8 base_leaf_kids;
2241  kmp_uint8 leaf_kids;
2242  kmp_uint8 offset;
2243  kmp_uint8 wait_flag;
2244  kmp_uint8 use_oncore_barrier;
2245 #if USE_DEBUGGER
2246  // The following field is intended for the debugger solely. Only the worker
2247  // thread itself accesses this field: the worker increases it by 1 when it
2248  // arrives to a barrier.
2249  KMP_ALIGN_CACHE kmp_uint b_worker_arrived;
2250 #endif /* USE_DEBUGGER */
2251 } kmp_bstate_t;
2252 
2253 union KMP_ALIGN_CACHE kmp_barrier_union {
2254  double b_align; /* use worst case alignment */
2255  char b_pad[KMP_PAD(kmp_bstate_t, CACHE_LINE)];
2256  kmp_bstate_t bb;
2257 };
2258 
2259 typedef union kmp_barrier_union kmp_balign_t;
2260 
2261 /* Team barrier needs only non-volatile arrived counter */
2262 union KMP_ALIGN_CACHE kmp_barrier_team_union {
2263  double b_align; /* use worst case alignment */
2264  char b_pad[CACHE_LINE];
2265  struct {
2266  kmp_uint64 b_arrived; /* STATE => task reached synch point. */
2267 #if USE_DEBUGGER
2268  // The following two fields are indended for the debugger solely. Only
2269  // primary thread of the team accesses these fields: the first one is
2270  // increased by 1 when the primary thread arrives to a barrier, the second
2271  // one is increased by one when all the threads arrived.
2272  kmp_uint b_master_arrived;
2273  kmp_uint b_team_arrived;
2274 #endif
2275  };
2276 };
2277 
2278 typedef union kmp_barrier_team_union kmp_balign_team_t;
2279 
2280 /* Padding for Linux* OS pthreads condition variables and mutexes used to signal
2281  threads when a condition changes. This is to workaround an NPTL bug where
2282  padding was added to pthread_cond_t which caused the initialization routine
2283  to write outside of the structure if compiled on pre-NPTL threads. */
2284 #if KMP_OS_WINDOWS
2285 typedef struct kmp_win32_mutex {
2286  /* The Lock */
2287  CRITICAL_SECTION cs;
2288 } kmp_win32_mutex_t;
2289 
2290 typedef struct kmp_win32_cond {
2291  /* Count of the number of waiters. */
2292  int waiters_count_;
2293 
2294  /* Serialize access to <waiters_count_> */
2295  kmp_win32_mutex_t waiters_count_lock_;
2296 
2297  /* Number of threads to release via a <cond_broadcast> or a <cond_signal> */
2298  int release_count_;
2299 
2300  /* Keeps track of the current "generation" so that we don't allow */
2301  /* one thread to steal all the "releases" from the broadcast. */
2302  int wait_generation_count_;
2303 
2304  /* A manual-reset event that's used to block and release waiting threads. */
2305  HANDLE event_;
2306 } kmp_win32_cond_t;
2307 #endif
2308 
2309 #if KMP_OS_UNIX
2310 
2311 union KMP_ALIGN_CACHE kmp_cond_union {
2312  double c_align;
2313  char c_pad[CACHE_LINE];
2314  pthread_cond_t c_cond;
2315 };
2316 
2317 typedef union kmp_cond_union kmp_cond_align_t;
2318 
2319 union KMP_ALIGN_CACHE kmp_mutex_union {
2320  double m_align;
2321  char m_pad[CACHE_LINE];
2322  pthread_mutex_t m_mutex;
2323 };
2324 
2325 typedef union kmp_mutex_union kmp_mutex_align_t;
2326 
2327 #endif /* KMP_OS_UNIX */
2328 
2329 typedef struct kmp_desc_base {
2330  void *ds_stackbase;
2331  size_t ds_stacksize;
2332  int ds_stackgrow;
2333  kmp_thread_t ds_thread;
2334  volatile int ds_tid;
2335  int ds_gtid;
2336 #if KMP_OS_WINDOWS
2337  volatile int ds_alive;
2338  DWORD ds_thread_id;
2339 /* ds_thread keeps thread handle on Windows* OS. It is enough for RTL purposes.
2340  However, debugger support (libomp_db) cannot work with handles, because they
2341  uncomparable. For example, debugger requests info about thread with handle h.
2342  h is valid within debugger process, and meaningless within debugee process.
2343  Even if h is duped by call to DuplicateHandle(), so the result h' is valid
2344  within debugee process, but it is a *new* handle which does *not* equal to
2345  any other handle in debugee... The only way to compare handles is convert
2346  them to system-wide ids. GetThreadId() function is available only in
2347  Longhorn and Server 2003. :-( In contrast, GetCurrentThreadId() is available
2348  on all Windows* OS flavours (including Windows* 95). Thus, we have to get
2349  thread id by call to GetCurrentThreadId() from within the thread and save it
2350  to let libomp_db identify threads. */
2351 #endif /* KMP_OS_WINDOWS */
2352 } kmp_desc_base_t;
2353 
2354 typedef union KMP_ALIGN_CACHE kmp_desc {
2355  double ds_align; /* use worst case alignment */
2356  char ds_pad[KMP_PAD(kmp_desc_base_t, CACHE_LINE)];
2357  kmp_desc_base_t ds;
2358 } kmp_desc_t;
2359 
2360 typedef struct kmp_local {
2361  volatile int this_construct; /* count of single's encountered by thread */
2362  void *reduce_data;
2363 #if KMP_USE_BGET
2364  void *bget_data;
2365  void *bget_list;
2366 #if !USE_CMP_XCHG_FOR_BGET
2367 #ifdef USE_QUEUING_LOCK_FOR_BGET
2368  kmp_lock_t bget_lock; /* Lock for accessing bget free list */
2369 #else
2370  kmp_bootstrap_lock_t bget_lock; // Lock for accessing bget free list. Must be
2371 // bootstrap lock so we can use it at library
2372 // shutdown.
2373 #endif /* USE_LOCK_FOR_BGET */
2374 #endif /* ! USE_CMP_XCHG_FOR_BGET */
2375 #endif /* KMP_USE_BGET */
2376 
2377  PACKED_REDUCTION_METHOD_T
2378  packed_reduction_method; /* stored by __kmpc_reduce*(), used by
2379  __kmpc_end_reduce*() */
2380 
2381 } kmp_local_t;
2382 
2383 #define KMP_CHECK_UPDATE(a, b) \
2384  if ((a) != (b)) \
2385  (a) = (b)
2386 #define KMP_CHECK_UPDATE_SYNC(a, b) \
2387  if ((a) != (b)) \
2388  TCW_SYNC_PTR((a), (b))
2389 
2390 #define get__blocktime(xteam, xtid) \
2391  ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.blocktime)
2392 #define get__bt_set(xteam, xtid) \
2393  ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_set)
2394 #if KMP_USE_MONITOR
2395 #define get__bt_intervals(xteam, xtid) \
2396  ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_intervals)
2397 #endif
2398 
2399 #define get__dynamic_2(xteam, xtid) \
2400  ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.dynamic)
2401 #define get__nproc_2(xteam, xtid) \
2402  ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.nproc)
2403 #define get__sched_2(xteam, xtid) \
2404  ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.sched)
2405 
2406 #define set__blocktime_team(xteam, xtid, xval) \
2407  (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.blocktime) = \
2408  (xval))
2409 
2410 #if KMP_USE_MONITOR
2411 #define set__bt_intervals_team(xteam, xtid, xval) \
2412  (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_intervals) = \
2413  (xval))
2414 #endif
2415 
2416 #define set__bt_set_team(xteam, xtid, xval) \
2417  (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_set) = (xval))
2418 
2419 #define set__dynamic(xthread, xval) \
2420  (((xthread)->th.th_current_task->td_icvs.dynamic) = (xval))
2421 #define get__dynamic(xthread) \
2422  (((xthread)->th.th_current_task->td_icvs.dynamic) ? (FTN_TRUE) : (FTN_FALSE))
2423 
2424 #define set__nproc(xthread, xval) \
2425  (((xthread)->th.th_current_task->td_icvs.nproc) = (xval))
2426 
2427 #define set__thread_limit(xthread, xval) \
2428  (((xthread)->th.th_current_task->td_icvs.thread_limit) = (xval))
2429 
2430 #define set__max_active_levels(xthread, xval) \
2431  (((xthread)->th.th_current_task->td_icvs.max_active_levels) = (xval))
2432 
2433 #define get__max_active_levels(xthread) \
2434  ((xthread)->th.th_current_task->td_icvs.max_active_levels)
2435 
2436 #define set__sched(xthread, xval) \
2437  (((xthread)->th.th_current_task->td_icvs.sched) = (xval))
2438 
2439 #define set__proc_bind(xthread, xval) \
2440  (((xthread)->th.th_current_task->td_icvs.proc_bind) = (xval))
2441 #define get__proc_bind(xthread) \
2442  ((xthread)->th.th_current_task->td_icvs.proc_bind)
2443 
2444 // OpenMP tasking data structures
2445 
2446 typedef enum kmp_tasking_mode {
2447  tskm_immediate_exec = 0,
2448  tskm_extra_barrier = 1,
2449  tskm_task_teams = 2,
2450  tskm_max = 2
2451 } kmp_tasking_mode_t;
2452 
2453 extern kmp_tasking_mode_t
2454  __kmp_tasking_mode; /* determines how/when to execute tasks */
2455 extern int __kmp_task_stealing_constraint;
2456 extern int __kmp_enable_task_throttling;
2457 extern kmp_int32 __kmp_default_device; // Set via OMP_DEFAULT_DEVICE if
2458 // specified, defaults to 0 otherwise
2459 // Set via OMP_MAX_TASK_PRIORITY if specified, defaults to 0 otherwise
2460 extern kmp_int32 __kmp_max_task_priority;
2461 // Set via KMP_TASKLOOP_MIN_TASKS if specified, defaults to 0 otherwise
2462 extern kmp_uint64 __kmp_taskloop_min_tasks;
2463 
2464 /* NOTE: kmp_taskdata_t and kmp_task_t structures allocated in single block with
2465  taskdata first */
2466 #define KMP_TASK_TO_TASKDATA(task) (((kmp_taskdata_t *)task) - 1)
2467 #define KMP_TASKDATA_TO_TASK(taskdata) (kmp_task_t *)(taskdata + 1)
2468 
2469 // The tt_found_tasks flag is a signal to all threads in the team that tasks
2470 // were spawned and queued since the previous barrier release.
2471 #define KMP_TASKING_ENABLED(task_team) \
2472  (TRUE == TCR_SYNC_4((task_team)->tt.tt_found_tasks))
2480 typedef kmp_int32 (*kmp_routine_entry_t)(kmp_int32, void *);
2481 
2482 typedef union kmp_cmplrdata {
2483  kmp_int32 priority;
2484  kmp_routine_entry_t
2485  destructors; /* pointer to function to invoke deconstructors of
2486  firstprivate C++ objects */
2487  /* future data */
2488 } kmp_cmplrdata_t;
2489 
2490 /* sizeof_kmp_task_t passed as arg to kmpc_omp_task call */
2493 typedef struct kmp_task { /* GEH: Shouldn't this be aligned somehow? */
2494  void *shareds;
2495  kmp_routine_entry_t
2496  routine;
2497  kmp_int32 part_id;
2498  kmp_cmplrdata_t
2499  data1; /* Two known optional additions: destructors and priority */
2500  kmp_cmplrdata_t data2; /* Process destructors first, priority second */
2501  /* future data */
2502  /* private vars */
2503 } kmp_task_t;
2504 
2509 typedef struct kmp_taskgroup {
2510  std::atomic<kmp_int32> count; // number of allocated and incomplete tasks
2511  std::atomic<kmp_int32>
2512  cancel_request; // request for cancellation of this taskgroup
2513  struct kmp_taskgroup *parent; // parent taskgroup
2514  // Block of data to perform task reduction
2515  void *reduce_data; // reduction related info
2516  kmp_int32 reduce_num_data; // number of data items to reduce
2517  uintptr_t *gomp_data; // gomp reduction data
2518 } kmp_taskgroup_t;
2519 
2520 // forward declarations
2521 typedef union kmp_depnode kmp_depnode_t;
2522 typedef struct kmp_depnode_list kmp_depnode_list_t;
2523 typedef struct kmp_dephash_entry kmp_dephash_entry_t;
2524 
2525 // macros for checking dep flag as an integer
2526 #define KMP_DEP_IN 0x1
2527 #define KMP_DEP_OUT 0x2
2528 #define KMP_DEP_INOUT 0x3
2529 #define KMP_DEP_MTX 0x4
2530 #define KMP_DEP_SET 0x8
2531 #define KMP_DEP_ALL 0x80
2532 // Compiler sends us this info. Note: some test cases contain an explicit copy
2533 // of this struct and should be in sync with any changes here.
2534 typedef struct kmp_depend_info {
2535  kmp_intptr_t base_addr;
2536  size_t len;
2537  union {
2538  kmp_uint8 flag; // flag as an unsigned char
2539  struct { // flag as a set of 8 bits
2540 #if defined(__BYTE_ORDER__) && (__BYTE_ORDER__ == __ORDER_BIG_ENDIAN__)
2541  /* Same fields as in the #else branch, but in reverse order */
2542  unsigned all : 1;
2543  unsigned unused : 3;
2544  unsigned set : 1;
2545  unsigned mtx : 1;
2546  unsigned out : 1;
2547  unsigned in : 1;
2548 #else
2549  unsigned in : 1;
2550  unsigned out : 1;
2551  unsigned mtx : 1;
2552  unsigned set : 1;
2553  unsigned unused : 3;
2554  unsigned all : 1;
2555 #endif
2556  } flags;
2557  };
2558 } kmp_depend_info_t;
2559 
2560 // Internal structures to work with task dependencies:
2561 struct kmp_depnode_list {
2562  kmp_depnode_t *node;
2563  kmp_depnode_list_t *next;
2564 };
2565 
2566 // Max number of mutexinoutset dependencies per node
2567 #define MAX_MTX_DEPS 4
2568 
2569 typedef struct kmp_base_depnode {
2570  kmp_depnode_list_t *successors; /* used under lock */
2571  kmp_task_t *task; /* non-NULL if depnode is active, used under lock */
2572  kmp_lock_t *mtx_locks[MAX_MTX_DEPS]; /* lock mutexinoutset dependent tasks */
2573  kmp_int32 mtx_num_locks; /* number of locks in mtx_locks array */
2574  kmp_lock_t lock; /* guards shared fields: task, successors */
2575 #if KMP_SUPPORT_GRAPH_OUTPUT
2576  kmp_uint32 id;
2577 #endif
2578  std::atomic<kmp_int32> npredecessors;
2579  std::atomic<kmp_int32> nrefs;
2580 } kmp_base_depnode_t;
2581 
2582 union KMP_ALIGN_CACHE kmp_depnode {
2583  double dn_align; /* use worst case alignment */
2584  char dn_pad[KMP_PAD(kmp_base_depnode_t, CACHE_LINE)];
2585  kmp_base_depnode_t dn;
2586 };
2587 
2588 struct kmp_dephash_entry {
2589  kmp_intptr_t addr;
2590  kmp_depnode_t *last_out;
2591  kmp_depnode_list_t *last_set;
2592  kmp_depnode_list_t *prev_set;
2593  kmp_uint8 last_flag;
2594  kmp_lock_t *mtx_lock; /* is referenced by depnodes w/mutexinoutset dep */
2595  kmp_dephash_entry_t *next_in_bucket;
2596 };
2597 
2598 typedef struct kmp_dephash {
2599  kmp_dephash_entry_t **buckets;
2600  size_t size;
2601  kmp_depnode_t *last_all;
2602  size_t generation;
2603  kmp_uint32 nelements;
2604  kmp_uint32 nconflicts;
2605 } kmp_dephash_t;
2606 
2607 typedef struct kmp_task_affinity_info {
2608  kmp_intptr_t base_addr;
2609  size_t len;
2610  struct {
2611  bool flag1 : 1;
2612  bool flag2 : 1;
2613  kmp_int32 reserved : 30;
2614  } flags;
2615 } kmp_task_affinity_info_t;
2616 
2617 typedef enum kmp_event_type_t {
2618  KMP_EVENT_UNINITIALIZED = 0,
2619  KMP_EVENT_ALLOW_COMPLETION = 1
2620 } kmp_event_type_t;
2621 
2622 typedef struct {
2623  kmp_event_type_t type;
2624  kmp_tas_lock_t lock;
2625  union {
2626  kmp_task_t *task;
2627  } ed;
2628 } kmp_event_t;
2629 
2630 #if OMPX_TASKGRAPH
2631 // Initial number of allocated nodes while recording
2632 #define INIT_MAPSIZE 50
2633 
2634 typedef struct kmp_taskgraph_flags { /*This needs to be exactly 32 bits */
2635  unsigned nowait : 1;
2636  unsigned re_record : 1;
2637  unsigned reserved : 30;
2638 } kmp_taskgraph_flags_t;
2639 
2641 typedef struct kmp_node_info {
2642  kmp_task_t *task; // Pointer to the actual task
2643  kmp_int32 *successors; // Array of the succesors ids
2644  kmp_int32 nsuccessors; // Number of succesors of the node
2645  std::atomic<kmp_int32>
2646  npredecessors_counter; // Number of predessors on the fly
2647  kmp_int32 npredecessors; // Total number of predecessors
2648  kmp_int32 successors_size; // Number of allocated succesors ids
2649  kmp_taskdata_t *parent_task; // Parent implicit task
2650 } kmp_node_info_t;
2651 
2653 typedef enum kmp_tdg_status {
2654  KMP_TDG_NONE = 0,
2655  KMP_TDG_RECORDING = 1,
2656  KMP_TDG_READY = 2
2657 } kmp_tdg_status_t;
2658 
2660 typedef struct kmp_tdg_info {
2661  kmp_int32 tdg_id; // Unique idenfifier of the TDG
2662  kmp_taskgraph_flags_t tdg_flags; // Flags related to a TDG
2663  kmp_int32 map_size; // Number of allocated TDG nodes
2664  kmp_int32 num_roots; // Number of roots tasks int the TDG
2665  kmp_int32 *root_tasks; // Array of tasks identifiers that are roots
2666  kmp_node_info_t *record_map; // Array of TDG nodes
2667  kmp_tdg_status_t tdg_status =
2668  KMP_TDG_NONE; // Status of the TDG (recording, ready...)
2669  std::atomic<kmp_int32> num_tasks; // Number of TDG nodes
2670  kmp_bootstrap_lock_t
2671  graph_lock; // Protect graph attributes when updated via taskloop_recur
2672  // Taskloop reduction related
2673  void *rec_taskred_data; // Data to pass to __kmpc_task_reduction_init or
2674  // __kmpc_taskred_init
2675  kmp_int32 rec_num_taskred;
2676 } kmp_tdg_info_t;
2677 
2678 extern int __kmp_tdg_dot;
2679 extern kmp_int32 __kmp_max_tdgs;
2680 extern kmp_tdg_info_t **__kmp_global_tdgs;
2681 extern kmp_int32 __kmp_curr_tdg_idx;
2682 extern kmp_int32 __kmp_successors_size;
2683 extern std::atomic<kmp_int32> __kmp_tdg_task_id;
2684 extern kmp_int32 __kmp_num_tdg;
2685 #endif
2686 
2687 typedef struct kmp_tasking_flags { /* Total struct must be exactly 32 bits */
2688 #if defined(__BYTE_ORDER__) && (__BYTE_ORDER__ == __ORDER_BIG_ENDIAN__)
2689  /* Same fields as in the #else branch, but in reverse order */
2690 #if OMPX_TASKGRAPH
2691  unsigned reserved31 : 4;
2692  unsigned onced : 1;
2693 #else
2694  unsigned reserved31 : 5;
2695 #endif
2696  unsigned hidden_helper : 1;
2697  unsigned target : 1;
2698  unsigned native : 1;
2699  unsigned freed : 1;
2700  unsigned complete : 1;
2701  unsigned executing : 1;
2702  unsigned started : 1;
2703  unsigned team_serial : 1;
2704  unsigned tasking_ser : 1;
2705  unsigned task_serial : 1;
2706  unsigned tasktype : 1;
2707  unsigned reserved : 8;
2708  unsigned free_agent_eligible : 1;
2709  unsigned detachable : 1;
2710  unsigned priority_specified : 1;
2711  unsigned proxy : 1;
2712  unsigned destructors_thunk : 1;
2713  unsigned merged_if0 : 1;
2714  unsigned final : 1;
2715  unsigned tiedness : 1;
2716 #else
2717  /* Compiler flags */ /* Total compiler flags must be 16 bits */
2718  unsigned tiedness : 1; /* task is either tied (1) or untied (0) */
2719  unsigned final : 1; /* task is final(1) so execute immediately */
2720  unsigned merged_if0 : 1; /* no __kmpc_task_{begin/complete}_if0 calls in if0
2721  code path */
2722  unsigned destructors_thunk : 1; /* set if the compiler creates a thunk to
2723  invoke destructors from the runtime */
2724  unsigned proxy : 1; /* task is a proxy task (it will be executed outside the
2725  context of the RTL) */
2726  unsigned priority_specified : 1; /* set if the compiler provides priority
2727  setting for the task */
2728  unsigned detachable : 1; /* 1 == can detach */
2729  unsigned free_agent_eligible : 1; /* set if task can be executed by a
2730  free-agent thread */
2731  unsigned reserved : 8; /* reserved for compiler use */
2732 
2733  /* Library flags */ /* Total library flags must be 16 bits */
2734  unsigned tasktype : 1; /* task is either explicit(1) or implicit (0) */
2735  unsigned task_serial : 1; // task is executed immediately (1) or deferred (0)
2736  unsigned tasking_ser : 1; // all tasks in team are either executed immediately
2737  // (1) or may be deferred (0)
2738  unsigned team_serial : 1; // entire team is serial (1) [1 thread] or parallel
2739  // (0) [>= 2 threads]
2740  /* If either team_serial or tasking_ser is set, task team may be NULL */
2741  /* Task State Flags: */
2742  unsigned started : 1; /* 1==started, 0==not started */
2743  unsigned executing : 1; /* 1==executing, 0==not executing */
2744  unsigned complete : 1; /* 1==complete, 0==not complete */
2745  unsigned freed : 1; /* 1==freed, 0==allocated */
2746  unsigned native : 1; /* 1==gcc-compiled task, 0==intel */
2747  unsigned target : 1;
2748  unsigned hidden_helper : 1; /* 1 == hidden helper task */
2749 #if OMPX_TASKGRAPH
2750  unsigned onced : 1; /* 1==ran once already, 0==never ran, record & replay purposes */
2751  unsigned reserved31 : 4; /* reserved for library use */
2752 #else
2753  unsigned reserved31 : 5; /* reserved for library use */
2754 #endif
2755 #endif
2756 } kmp_tasking_flags_t;
2757 
2758 typedef struct kmp_target_data {
2759  void *async_handle; // libomptarget async handle for task completion query
2760 } kmp_target_data_t;
2761 
2762 struct kmp_taskdata { /* aligned during dynamic allocation */
2763  kmp_int32 td_task_id; /* id, assigned by debugger */
2764  kmp_tasking_flags_t td_flags; /* task flags */
2765  kmp_team_t *td_team; /* team for this task */
2766  kmp_info_p *td_alloc_thread; /* thread that allocated data structures */
2767  /* Currently not used except for perhaps IDB */
2768  kmp_taskdata_t *td_parent; /* parent task */
2769  kmp_int32 td_level; /* task nesting level */
2770  std::atomic<kmp_int32> td_untied_count; // untied task active parts counter
2771  ident_t *td_ident; /* task identifier */
2772  // Taskwait data.
2773  ident_t *td_taskwait_ident;
2774  kmp_uint32 td_taskwait_counter;
2775  kmp_int32 td_taskwait_thread; /* gtid + 1 of thread encountered taskwait */
2776  KMP_ALIGN_CACHE kmp_internal_control_t
2777  td_icvs; /* Internal control variables for the task */
2778  KMP_ALIGN_CACHE std::atomic<kmp_int32>
2779  td_allocated_child_tasks; /* Child tasks (+ current task) not yet
2780  deallocated */
2781  std::atomic<kmp_int32>
2782  td_incomplete_child_tasks; /* Child tasks not yet complete */
2783  kmp_taskgroup_t
2784  *td_taskgroup; // Each task keeps pointer to its current taskgroup
2785  kmp_dephash_t
2786  *td_dephash; // Dependencies for children tasks are tracked from here
2787  kmp_depnode_t
2788  *td_depnode; // Pointer to graph node if this task has dependencies
2789  kmp_task_team_t *td_task_team;
2790  size_t td_size_alloc; // Size of task structure, including shareds etc.
2791 #if defined(KMP_GOMP_COMPAT)
2792  // 4 or 8 byte integers for the loop bounds in GOMP_taskloop
2793  kmp_int32 td_size_loop_bounds;
2794 #endif
2795  kmp_taskdata_t *td_last_tied; // keep tied task for task scheduling constraint
2796 #if defined(KMP_GOMP_COMPAT)
2797  // GOMP sends in a copy function for copy constructors
2798  void (*td_copy_func)(void *, void *);
2799 #endif
2800  kmp_event_t td_allow_completion_event;
2801 #if OMPT_SUPPORT
2802  ompt_task_info_t ompt_task_info;
2803 #endif
2804 #if OMPX_TASKGRAPH
2805  bool is_taskgraph = 0; // whether the task is within a TDG
2806  kmp_tdg_info_t *tdg; // used to associate task with a TDG
2807  kmp_int32 td_tdg_task_id; // local task id in its TDG
2808 #endif
2809  kmp_target_data_t td_target_data;
2810 }; // struct kmp_taskdata
2811 
2812 // Make sure padding above worked
2813 KMP_BUILD_ASSERT(sizeof(kmp_taskdata_t) % sizeof(void *) == 0);
2814 
2815 // Data for task team but per thread
2816 typedef struct kmp_base_thread_data {
2817  kmp_info_p *td_thr; // Pointer back to thread info
2818  // Used only in __kmp_execute_tasks_template, maybe not avail until task is
2819  // queued?
2820  kmp_bootstrap_lock_t td_deque_lock; // Lock for accessing deque
2821  kmp_taskdata_t *
2822  *td_deque; // Deque of tasks encountered by td_thr, dynamically allocated
2823  kmp_int32 td_deque_size; // Size of deck
2824  kmp_uint32 td_deque_head; // Head of deque (will wrap)
2825  kmp_uint32 td_deque_tail; // Tail of deque (will wrap)
2826  kmp_int32 td_deque_ntasks; // Number of tasks in deque
2827  // GEH: shouldn't this be volatile since used in while-spin?
2828  kmp_int32 td_deque_last_stolen; // Thread number of last successful steal
2829 } kmp_base_thread_data_t;
2830 
2831 #define TASK_DEQUE_BITS 8 // Used solely to define INITIAL_TASK_DEQUE_SIZE
2832 #define INITIAL_TASK_DEQUE_SIZE (1 << TASK_DEQUE_BITS)
2833 
2834 #define TASK_DEQUE_SIZE(td) ((td).td_deque_size)
2835 #define TASK_DEQUE_MASK(td) ((td).td_deque_size - 1)
2836 
2837 typedef union KMP_ALIGN_CACHE kmp_thread_data {
2838  kmp_base_thread_data_t td;
2839  double td_align; /* use worst case alignment */
2840  char td_pad[KMP_PAD(kmp_base_thread_data_t, CACHE_LINE)];
2841 } kmp_thread_data_t;
2842 
2843 typedef struct kmp_task_pri {
2844  kmp_thread_data_t td;
2845  kmp_int32 priority;
2846  kmp_task_pri *next;
2847 } kmp_task_pri_t;
2848 
2849 // Data for task teams which are used when tasking is enabled for the team
2850 typedef struct kmp_base_task_team {
2851  kmp_bootstrap_lock_t
2852  tt_threads_lock; /* Lock used to allocate per-thread part of task team */
2853  /* must be bootstrap lock since used at library shutdown*/
2854 
2855  // TODO: check performance vs kmp_tas_lock_t
2856  kmp_bootstrap_lock_t tt_task_pri_lock; /* Lock to access priority tasks */
2857  kmp_task_pri_t *tt_task_pri_list;
2858 
2859  kmp_task_team_t *tt_next; /* For linking the task team free list */
2860  kmp_thread_data_t
2861  *tt_threads_data; /* Array of per-thread structures for task team */
2862  /* Data survives task team deallocation */
2863  kmp_int32 tt_found_tasks; /* Have we found tasks and queued them while
2864  executing this team? */
2865  /* TRUE means tt_threads_data is set up and initialized */
2866  kmp_int32 tt_nproc; /* #threads in team */
2867  kmp_int32 tt_max_threads; // # entries allocated for threads_data array
2868  kmp_int32 tt_found_proxy_tasks; // found proxy tasks since last barrier
2869  kmp_int32 tt_untied_task_encountered;
2870  std::atomic<kmp_int32> tt_num_task_pri; // number of priority tasks enqueued
2871  // There is hidden helper thread encountered in this task team so that we must
2872  // wait when waiting on task team
2873  kmp_int32 tt_hidden_helper_task_encountered;
2874 
2875  KMP_ALIGN_CACHE
2876  std::atomic<kmp_int32> tt_unfinished_threads; /* #threads still active */
2877 
2878  KMP_ALIGN_CACHE
2879  volatile kmp_uint32
2880  tt_active; /* is the team still actively executing tasks */
2881 } kmp_base_task_team_t;
2882 
2883 union KMP_ALIGN_CACHE kmp_task_team {
2884  kmp_base_task_team_t tt;
2885  double tt_align; /* use worst case alignment */
2886  char tt_pad[KMP_PAD(kmp_base_task_team_t, CACHE_LINE)];
2887 };
2888 
2889 typedef struct kmp_task_team_list_t {
2890  kmp_task_team_t *task_team;
2891  kmp_task_team_list_t *next;
2892 } kmp_task_team_list_t;
2893 
2894 #if (USE_FAST_MEMORY == 3) || (USE_FAST_MEMORY == 5)
2895 // Free lists keep same-size free memory slots for fast memory allocation
2896 // routines
2897 typedef struct kmp_free_list {
2898  void *th_free_list_self; // Self-allocated tasks free list
2899  void *th_free_list_sync; // Self-allocated tasks stolen/returned by other
2900  // threads
2901  void *th_free_list_other; // Non-self free list (to be returned to owner's
2902  // sync list)
2903 } kmp_free_list_t;
2904 #endif
2905 // Hot teams array keeps hot teams and their sizes for given thread. Hot teams
2906 // are not put in teams pool, and they don't put threads in threads pool.
2907 typedef struct kmp_hot_team_ptr {
2908  kmp_team_p *hot_team; // pointer to hot_team of given nesting level
2909  kmp_int32 hot_team_nth; // number of threads allocated for the hot_team
2910 } kmp_hot_team_ptr_t;
2911 typedef struct kmp_teams_size {
2912  kmp_int32 nteams; // number of teams in a league
2913  kmp_int32 nth; // number of threads in each team of the league
2914 } kmp_teams_size_t;
2915 
2916 // This struct stores a thread that acts as a "root" for a contention
2917 // group. Contention groups are rooted at kmp_root threads, but also at
2918 // each primary thread of each team created in the teams construct.
2919 // This struct therefore also stores a thread_limit associated with
2920 // that contention group, and a counter to track the number of threads
2921 // active in that contention group. Each thread has a list of these: CG
2922 // root threads have an entry in their list in which cg_root refers to
2923 // the thread itself, whereas other workers in the CG will have a
2924 // single entry where cg_root is same as the entry containing their CG
2925 // root. When a thread encounters a teams construct, it will add a new
2926 // entry to the front of its list, because it now roots a new CG.
2927 typedef struct kmp_cg_root {
2928  kmp_info_p *cg_root; // "root" thread for a contention group
2929  // The CG root's limit comes from OMP_THREAD_LIMIT for root threads, or
2930  // thread_limit clause for teams primary threads
2931  kmp_int32 cg_thread_limit;
2932  kmp_int32 cg_nthreads; // Count of active threads in CG rooted at cg_root
2933  struct kmp_cg_root *up; // pointer to higher level CG root in list
2934 } kmp_cg_root_t;
2935 
2936 // OpenMP thread data structures
2937 
2938 typedef struct KMP_ALIGN_CACHE kmp_base_info {
2939  /* Start with the readonly data which is cache aligned and padded. This is
2940  written before the thread starts working by the primary thread. Uber
2941  masters may update themselves later. Usage does not consider serialized
2942  regions. */
2943  kmp_desc_t th_info;
2944  kmp_team_p *th_team; /* team we belong to */
2945  kmp_root_p *th_root; /* pointer to root of task hierarchy */
2946  kmp_info_p *th_next_pool; /* next available thread in the pool */
2947  kmp_disp_t *th_dispatch; /* thread's dispatch data */
2948  int th_in_pool; /* in thread pool (32 bits for TCR/TCW) */
2949 
2950  /* The following are cached from the team info structure */
2951  /* TODO use these in more places as determined to be needed via profiling */
2952  int th_team_nproc; /* number of threads in a team */
2953  kmp_info_p *th_team_master; /* the team's primary thread */
2954  int th_team_serialized; /* team is serialized */
2955  microtask_t th_teams_microtask; /* save entry address for teams construct */
2956  int th_teams_level; /* save initial level of teams construct */
2957 /* it is 0 on device but may be any on host */
2958 
2959 /* The blocktime info is copied from the team struct to the thread struct */
2960 /* at the start of a barrier, and the values stored in the team are used */
2961 /* at points in the code where the team struct is no longer guaranteed */
2962 /* to exist (from the POV of worker threads). */
2963 #if KMP_USE_MONITOR
2964  int th_team_bt_intervals;
2965  int th_team_bt_set;
2966 #else
2967  kmp_uint64 th_team_bt_intervals;
2968 #endif
2969 
2970 #if KMP_AFFINITY_SUPPORTED
2971  kmp_affin_mask_t *th_affin_mask; /* thread's current affinity mask */
2972  kmp_affinity_ids_t th_topology_ids; /* thread's current topology ids */
2973  kmp_affinity_attrs_t th_topology_attrs; /* thread's current topology attrs */
2974 #endif
2975  omp_allocator_handle_t th_def_allocator; /* default allocator */
2976  /* The data set by the primary thread at reinit, then R/W by the worker */
2977  KMP_ALIGN_CACHE int
2978  th_set_nproc; /* if > 0, then only use this request for the next fork */
2979  int *th_set_nested_nth;
2980  bool th_nt_strict; // num_threads clause has strict modifier
2981  ident_t *th_nt_loc; // loc for strict modifier
2982  int th_nt_sev; // error severity for strict modifier
2983  const char *th_nt_msg; // error message for strict modifier
2984  int th_set_nested_nth_sz;
2985  kmp_hot_team_ptr_t *th_hot_teams; /* array of hot teams */
2986  kmp_proc_bind_t
2987  th_set_proc_bind; /* if != proc_bind_default, use request for next fork */
2988  kmp_teams_size_t
2989  th_teams_size; /* number of teams/threads in teams construct */
2990 #if KMP_AFFINITY_SUPPORTED
2991  int th_current_place; /* place currently bound to */
2992  int th_new_place; /* place to bind to in par reg */
2993  int th_first_place; /* first place in partition */
2994  int th_last_place; /* last place in partition */
2995 #endif
2996  int th_prev_level; /* previous level for affinity format */
2997  int th_prev_num_threads; /* previous num_threads for affinity format */
2998 #if USE_ITT_BUILD
2999  kmp_uint64 th_bar_arrive_time; /* arrival to barrier timestamp */
3000  kmp_uint64 th_bar_min_time; /* minimum arrival time at the barrier */
3001  kmp_uint64 th_frame_time; /* frame timestamp */
3002 #endif /* USE_ITT_BUILD */
3003  kmp_local_t th_local;
3004  struct private_common *th_pri_head;
3005 
3006  /* Now the data only used by the worker (after initial allocation) */
3007  /* TODO the first serial team should actually be stored in the info_t
3008  structure. this will help reduce initial allocation overhead */
3009  KMP_ALIGN_CACHE kmp_team_p
3010  *th_serial_team; /*serialized team held in reserve*/
3011 
3012 #if OMPT_SUPPORT
3013  ompt_thread_info_t ompt_thread_info;
3014 #endif
3015 
3016  /* The following are also read by the primary thread during reinit */
3017  struct common_table *th_pri_common;
3018 
3019  volatile kmp_uint32 th_spin_here; /* thread-local location for spinning */
3020  /* while awaiting queuing lock acquire */
3021 
3022  volatile void *th_sleep_loc; // this points at a kmp_flag<T>
3023  flag_type th_sleep_loc_type; // enum type of flag stored in th_sleep_loc
3024 
3025  ident_t *th_ident;
3026  unsigned th_x; // Random number generator data
3027  unsigned th_a; // Random number generator data
3028 
3029  /* Tasking-related data for the thread */
3030  kmp_task_team_t *th_task_team; // Task team struct
3031  kmp_taskdata_t *th_current_task; // Innermost Task being executed
3032  kmp_uint8 th_task_state; // alternating 0/1 for task team identification
3033  kmp_uint32 th_reap_state; // Non-zero indicates thread is not
3034  // tasking, thus safe to reap
3035 
3036  /* More stuff for keeping track of active/sleeping threads (this part is
3037  written by the worker thread) */
3038  kmp_uint8 th_active_in_pool; // included in count of #active threads in pool
3039  int th_active; // ! sleeping; 32 bits for TCR/TCW
3040  std::atomic<kmp_uint32> th_used_in_team; // Flag indicating use in team
3041  // 0 = not used in team; 1 = used in team;
3042  // 2 = transitioning to not used in team; 3 = transitioning to used in team
3043  struct cons_header *th_cons; // used for consistency check
3044 #if KMP_USE_HIER_SCHED
3045  // used for hierarchical scheduling
3046  kmp_hier_private_bdata_t *th_hier_bar_data;
3047 #endif
3048 
3049  /* Add the syncronizing data which is cache aligned and padded. */
3050  KMP_ALIGN_CACHE kmp_balign_t th_bar[bs_last_barrier];
3051 
3052  KMP_ALIGN_CACHE volatile kmp_int32
3053  th_next_waiting; /* gtid+1 of next thread on lock wait queue, 0 if none */
3054 
3055 #if (USE_FAST_MEMORY == 3) || (USE_FAST_MEMORY == 5)
3056 #define NUM_LISTS 4
3057  kmp_free_list_t th_free_lists[NUM_LISTS]; // Free lists for fast memory
3058 // allocation routines
3059 #endif
3060 
3061 #if KMP_OS_WINDOWS
3062  kmp_win32_cond_t th_suspend_cv;
3063  kmp_win32_mutex_t th_suspend_mx;
3064  std::atomic<int> th_suspend_init;
3065 #endif
3066 #if KMP_OS_UNIX
3067  kmp_cond_align_t th_suspend_cv;
3068  kmp_mutex_align_t th_suspend_mx;
3069  std::atomic<int> th_suspend_init_count;
3070 #endif
3071 
3072 #if USE_ITT_BUILD
3073  kmp_itt_mark_t th_itt_mark_single;
3074 // alignment ???
3075 #endif /* USE_ITT_BUILD */
3076 #if KMP_STATS_ENABLED
3077  kmp_stats_list *th_stats;
3078 #endif
3079 #if KMP_OS_UNIX
3080  std::atomic<bool> th_blocking;
3081 #endif
3082  kmp_cg_root_t *th_cg_roots; // list of cg_roots associated with this thread
3083 } kmp_base_info_t;
3084 
3085 typedef union KMP_ALIGN_CACHE kmp_info {
3086  double th_align; /* use worst case alignment */
3087  char th_pad[KMP_PAD(kmp_base_info_t, CACHE_LINE)];
3088  kmp_base_info_t th;
3089 } kmp_info_t;
3090 
3091 // OpenMP thread team data structures
3092 
3093 typedef struct kmp_base_data {
3094  volatile kmp_uint32 t_value;
3095 } kmp_base_data_t;
3096 
3097 typedef union KMP_ALIGN_CACHE kmp_sleep_team {
3098  double dt_align; /* use worst case alignment */
3099  char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
3100  kmp_base_data_t dt;
3101 } kmp_sleep_team_t;
3102 
3103 typedef union KMP_ALIGN_CACHE kmp_ordered_team {
3104  double dt_align; /* use worst case alignment */
3105  char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
3106  kmp_base_data_t dt;
3107 } kmp_ordered_team_t;
3108 
3109 typedef int (*launch_t)(int gtid);
3110 
3111 /* Minimum number of ARGV entries to malloc if necessary */
3112 #define KMP_MIN_MALLOC_ARGV_ENTRIES 100
3113 
3114 // Set up how many argv pointers will fit in cache lines containing
3115 // t_inline_argv. Historically, we have supported at least 96 bytes. Using a
3116 // larger value for more space between the primary write/worker read section and
3117 // read/write by all section seems to buy more performance on EPCC PARALLEL.
3118 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
3119 #define KMP_INLINE_ARGV_BYTES \
3120  (4 * CACHE_LINE - \
3121  ((3 * KMP_PTR_SKIP + 2 * sizeof(int) + 2 * sizeof(kmp_int8) + \
3122  sizeof(kmp_int16) + sizeof(kmp_uint32)) % \
3123  CACHE_LINE))
3124 #else
3125 #define KMP_INLINE_ARGV_BYTES \
3126  (2 * CACHE_LINE - ((3 * KMP_PTR_SKIP + 2 * sizeof(int)) % CACHE_LINE))
3127 #endif
3128 #define KMP_INLINE_ARGV_ENTRIES (int)(KMP_INLINE_ARGV_BYTES / KMP_PTR_SKIP)
3129 
3130 typedef struct KMP_ALIGN_CACHE kmp_base_team {
3131  // Synchronization Data
3132  // ---------------------------------------------------------------------------
3133  KMP_ALIGN_CACHE kmp_ordered_team_t t_ordered;
3134  kmp_balign_team_t t_bar[bs_last_barrier];
3135  std::atomic<int> t_construct; // count of single directive encountered by team
3136  char pad[sizeof(kmp_lock_t)]; // padding to maintain performance on big iron
3137 
3138  // [0] - parallel / [1] - worksharing task reduction data shared by taskgroups
3139  std::atomic<void *> t_tg_reduce_data[2]; // to support task modifier
3140  std::atomic<int> t_tg_fini_counter[2]; // sync end of task reductions
3141 
3142  // Primary thread only
3143  // ---------------------------------------------------------------------------
3144  KMP_ALIGN_CACHE int t_master_tid; // tid of primary thread in parent team
3145  int t_master_this_cons; // "this_construct" single counter of primary thread
3146  // in parent team
3147  ident_t *t_ident; // if volatile, have to change too much other crud to
3148  // volatile too
3149  kmp_team_p *t_parent; // parent team
3150  kmp_team_p *t_next_pool; // next free team in the team pool
3151  kmp_disp_t *t_dispatch; // thread's dispatch data
3152  kmp_task_team_t *t_task_team[2]; // Task team struct; switch between 2
3153  kmp_proc_bind_t t_proc_bind; // bind type for par region
3154  int t_primary_task_state; // primary thread's task state saved
3155 #if USE_ITT_BUILD
3156  kmp_uint64 t_region_time; // region begin timestamp
3157 #endif /* USE_ITT_BUILD */
3158 
3159  // Primary thread write, workers read
3160  // --------------------------------------------------------------------------
3161  KMP_ALIGN_CACHE void **t_argv;
3162  int t_argc;
3163  int t_nproc; // number of threads in team
3164  microtask_t t_pkfn;
3165  launch_t t_invoke; // procedure to launch the microtask
3166 
3167 #if OMPT_SUPPORT
3168  ompt_team_info_t ompt_team_info;
3169  ompt_lw_taskteam_t *ompt_serialized_team_info;
3170 #endif
3171 
3172 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
3173  kmp_int8 t_fp_control_saved;
3174  kmp_int8 t_pad2b;
3175  kmp_int16 t_x87_fpu_control_word; // FP control regs
3176  kmp_uint32 t_mxcsr;
3177 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
3178 
3179  void *t_inline_argv[KMP_INLINE_ARGV_ENTRIES];
3180 
3181  KMP_ALIGN_CACHE kmp_info_t **t_threads;
3182  kmp_taskdata_t
3183  *t_implicit_task_taskdata; // Taskdata for the thread's implicit task
3184  int t_level; // nested parallel level
3185 
3186  KMP_ALIGN_CACHE int t_max_argc;
3187  int t_max_nproc; // max threads this team can handle (dynamically expandable)
3188  int t_serialized; // levels deep of serialized teams
3189  dispatch_shared_info_t *t_disp_buffer; // buffers for dispatch system
3190  int t_id; // team's id, assigned by debugger.
3191  int t_active_level; // nested active parallel level
3192  kmp_r_sched_t t_sched; // run-time schedule for the team
3193 #if KMP_AFFINITY_SUPPORTED
3194  int t_first_place; // first & last place in parent thread's partition.
3195  int t_last_place; // Restore these values to primary thread after par region.
3196 #endif // KMP_AFFINITY_SUPPORTED
3197  int t_display_affinity;
3198  int t_size_changed; // team size was changed?: 0: no, 1: yes, -1: changed via
3199  // omp_set_num_threads() call
3200  omp_allocator_handle_t t_def_allocator; /* default allocator */
3201 
3202 // Read/write by workers as well
3203 #if (KMP_ARCH_X86 || KMP_ARCH_X86_64)
3204  // Using CACHE_LINE=64 reduces memory footprint, but causes a big perf
3205  // regression of epcc 'parallel' and 'barrier' on fxe256lin01. This extra
3206  // padding serves to fix the performance of epcc 'parallel' and 'barrier' when
3207  // CACHE_LINE=64. TODO: investigate more and get rid if this padding.
3208  char dummy_padding[1024];
3209 #endif
3210  // Internal control stack for additional nested teams.
3211  KMP_ALIGN_CACHE kmp_internal_control_t *t_control_stack_top;
3212  // for SERIALIZED teams nested 2 or more levels deep
3213  // typed flag to store request state of cancellation
3214  std::atomic<kmp_int32> t_cancel_request;
3215  int t_master_active; // save on fork, restore on join
3216  void *t_copypriv_data; // team specific pointer to copyprivate data array
3217 #if KMP_OS_WINDOWS
3218  std::atomic<kmp_uint32> t_copyin_counter;
3219 #endif
3220 #if USE_ITT_BUILD
3221  void *t_stack_id; // team specific stack stitching id (for ittnotify)
3222 #endif /* USE_ITT_BUILD */
3223  distributedBarrier *b; // Distributed barrier data associated with team
3224  kmp_nested_nthreads_t *t_nested_nth;
3225 } kmp_base_team_t;
3226 
3227 // Assert that the list structure fits and aligns within
3228 // the double task team pointer
3229 KMP_BUILD_ASSERT(sizeof(kmp_task_team_t *[2]) == sizeof(kmp_task_team_list_t));
3230 KMP_BUILD_ASSERT(alignof(kmp_task_team_t *[2]) ==
3231  alignof(kmp_task_team_list_t));
3232 
3233 union KMP_ALIGN_CACHE kmp_team {
3234  kmp_base_team_t t;
3235  double t_align; /* use worst case alignment */
3236  char t_pad[KMP_PAD(kmp_base_team_t, CACHE_LINE)];
3237 };
3238 
3239 typedef union KMP_ALIGN_CACHE kmp_time_global {
3240  double dt_align; /* use worst case alignment */
3241  char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
3242  kmp_base_data_t dt;
3243 } kmp_time_global_t;
3244 
3245 typedef struct kmp_base_global {
3246  /* cache-aligned */
3247  kmp_time_global_t g_time;
3248 
3249  /* non cache-aligned */
3250  volatile int g_abort;
3251  volatile int g_done;
3252 
3253  int g_dynamic;
3254  enum dynamic_mode g_dynamic_mode;
3255 } kmp_base_global_t;
3256 
3257 typedef union KMP_ALIGN_CACHE kmp_global {
3258  kmp_base_global_t g;
3259  double g_align; /* use worst case alignment */
3260  char g_pad[KMP_PAD(kmp_base_global_t, CACHE_LINE)];
3261 } kmp_global_t;
3262 
3263 typedef struct kmp_base_root {
3264  // TODO: GEH - combine r_active with r_in_parallel then r_active ==
3265  // (r_in_parallel>= 0)
3266  // TODO: GEH - then replace r_active with t_active_levels if we can to reduce
3267  // the synch overhead or keeping r_active
3268  volatile int r_active; /* TRUE if some region in a nest has > 1 thread */
3269  // keeps a count of active parallel regions per root
3270  std::atomic<int> r_in_parallel;
3271  // GEH: This is misnamed, should be r_active_levels
3272  kmp_team_t *r_root_team;
3273  kmp_team_t *r_hot_team;
3274  kmp_info_t *r_uber_thread;
3275  kmp_lock_t r_begin_lock;
3276  volatile int r_begin;
3277  int r_blocktime; /* blocktime for this root and descendants */
3278 #if KMP_AFFINITY_SUPPORTED
3279  int r_affinity_assigned;
3280 #endif // KMP_AFFINITY_SUPPORTED
3281 } kmp_base_root_t;
3282 
3283 typedef union KMP_ALIGN_CACHE kmp_root {
3284  kmp_base_root_t r;
3285  double r_align; /* use worst case alignment */
3286  char r_pad[KMP_PAD(kmp_base_root_t, CACHE_LINE)];
3287 } kmp_root_t;
3288 
3289 struct fortran_inx_info {
3290  kmp_int32 data;
3291 };
3292 
3293 // This list type exists to hold old __kmp_threads arrays so that
3294 // old references to them may complete while reallocation takes place when
3295 // expanding the array. The items in this list are kept alive until library
3296 // shutdown.
3297 typedef struct kmp_old_threads_list_t {
3298  kmp_info_t **threads;
3299  struct kmp_old_threads_list_t *next;
3300 } kmp_old_threads_list_t;
3301 
3302 /* ------------------------------------------------------------------------ */
3303 
3304 extern int __kmp_settings;
3305 extern int __kmp_duplicate_library_ok;
3306 #if USE_ITT_BUILD
3307 extern int __kmp_forkjoin_frames;
3308 extern int __kmp_forkjoin_frames_mode;
3309 #endif
3310 extern PACKED_REDUCTION_METHOD_T __kmp_force_reduction_method;
3311 extern int __kmp_determ_red;
3312 
3313 #ifdef KMP_DEBUG
3314 extern int kmp_a_debug;
3315 extern int kmp_b_debug;
3316 extern int kmp_c_debug;
3317 extern int kmp_d_debug;
3318 extern int kmp_e_debug;
3319 extern int kmp_f_debug;
3320 #endif /* KMP_DEBUG */
3321 
3322 /* For debug information logging using rotating buffer */
3323 #define KMP_DEBUG_BUF_LINES_INIT 512
3324 #define KMP_DEBUG_BUF_LINES_MIN 1
3325 
3326 #define KMP_DEBUG_BUF_CHARS_INIT 128
3327 #define KMP_DEBUG_BUF_CHARS_MIN 2
3328 
3329 extern int
3330  __kmp_debug_buf; /* TRUE means use buffer, FALSE means print to stderr */
3331 extern int __kmp_debug_buf_lines; /* How many lines of debug stored in buffer */
3332 extern int
3333  __kmp_debug_buf_chars; /* How many characters allowed per line in buffer */
3334 extern int __kmp_debug_buf_atomic; /* TRUE means use atomic update of buffer
3335  entry pointer */
3336 
3337 extern char *__kmp_debug_buffer; /* Debug buffer itself */
3338 extern std::atomic<int> __kmp_debug_count; /* Counter for number of lines
3339  printed in buffer so far */
3340 extern int __kmp_debug_buf_warn_chars; /* Keep track of char increase
3341  recommended in warnings */
3342 /* end rotating debug buffer */
3343 
3344 #ifdef KMP_DEBUG
3345 extern int __kmp_par_range; /* +1 => only go par for constructs in range */
3346 
3347 #define KMP_PAR_RANGE_ROUTINE_LEN 1024
3348 extern char __kmp_par_range_routine[KMP_PAR_RANGE_ROUTINE_LEN];
3349 #define KMP_PAR_RANGE_FILENAME_LEN 1024
3350 extern char __kmp_par_range_filename[KMP_PAR_RANGE_FILENAME_LEN];
3351 extern int __kmp_par_range_lb;
3352 extern int __kmp_par_range_ub;
3353 #endif
3354 
3355 /* For printing out dynamic storage map for threads and teams */
3356 extern int
3357  __kmp_storage_map; /* True means print storage map for threads and teams */
3358 extern int __kmp_storage_map_verbose; /* True means storage map includes
3359  placement info */
3360 extern int __kmp_storage_map_verbose_specified;
3361 
3362 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
3363 extern kmp_cpuinfo_t __kmp_cpuinfo;
3364 static inline bool __kmp_is_hybrid_cpu() { return __kmp_cpuinfo.flags.hybrid; }
3365 #elif KMP_OS_DARWIN && KMP_ARCH_AARCH64
3366 static inline bool __kmp_is_hybrid_cpu() { return true; }
3367 #else
3368 static inline bool __kmp_is_hybrid_cpu() { return false; }
3369 #endif
3370 
3371 extern volatile int __kmp_init_serial;
3372 extern volatile int __kmp_init_gtid;
3373 extern volatile int __kmp_init_common;
3374 extern volatile int __kmp_need_register_serial;
3375 extern volatile int __kmp_init_middle;
3376 extern volatile int __kmp_init_parallel;
3377 #if KMP_USE_MONITOR
3378 extern volatile int __kmp_init_monitor;
3379 #endif
3380 extern volatile int __kmp_init_user_locks;
3381 extern volatile int __kmp_init_hidden_helper_threads;
3382 extern int __kmp_init_counter;
3383 extern int __kmp_root_counter;
3384 extern int __kmp_version;
3385 
3386 /* list of address of allocated caches for commons */
3387 extern kmp_cached_addr_t *__kmp_threadpriv_cache_list;
3388 
3389 /* Barrier algorithm types and options */
3390 extern kmp_uint32 __kmp_barrier_gather_bb_dflt;
3391 extern kmp_uint32 __kmp_barrier_release_bb_dflt;
3392 extern kmp_bar_pat_e __kmp_barrier_gather_pat_dflt;
3393 extern kmp_bar_pat_e __kmp_barrier_release_pat_dflt;
3394 extern kmp_uint32 __kmp_barrier_gather_branch_bits[bs_last_barrier];
3395 extern kmp_uint32 __kmp_barrier_release_branch_bits[bs_last_barrier];
3396 extern kmp_bar_pat_e __kmp_barrier_gather_pattern[bs_last_barrier];
3397 extern kmp_bar_pat_e __kmp_barrier_release_pattern[bs_last_barrier];
3398 extern char const *__kmp_barrier_branch_bit_env_name[bs_last_barrier];
3399 extern char const *__kmp_barrier_pattern_env_name[bs_last_barrier];
3400 extern char const *__kmp_barrier_type_name[bs_last_barrier];
3401 extern char const *__kmp_barrier_pattern_name[bp_last_bar];
3402 
3403 /* Global Locks */
3404 extern kmp_bootstrap_lock_t __kmp_initz_lock; /* control initialization */
3405 extern kmp_bootstrap_lock_t __kmp_forkjoin_lock; /* control fork/join access */
3406 extern kmp_bootstrap_lock_t __kmp_task_team_lock;
3407 extern kmp_bootstrap_lock_t
3408  __kmp_exit_lock; /* exit() is not always thread-safe */
3409 #if KMP_USE_MONITOR
3410 extern kmp_bootstrap_lock_t
3411  __kmp_monitor_lock; /* control monitor thread creation */
3412 #endif
3413 extern kmp_bootstrap_lock_t
3414  __kmp_tp_cached_lock; /* used for the hack to allow threadprivate cache and
3415  __kmp_threads expansion to co-exist */
3416 
3417 extern kmp_lock_t __kmp_global_lock; /* control OS/global access */
3418 
3419 extern enum library_type __kmp_library;
3420 
3421 extern enum sched_type __kmp_sched; /* default runtime scheduling */
3422 extern enum sched_type __kmp_static; /* default static scheduling method */
3423 extern enum sched_type __kmp_guided; /* default guided scheduling method */
3424 extern enum sched_type __kmp_auto; /* default auto scheduling method */
3425 extern int __kmp_chunk; /* default runtime chunk size */
3426 extern int __kmp_force_monotonic; /* whether monotonic scheduling forced */
3427 
3428 extern size_t __kmp_stksize; /* stack size per thread */
3429 #if KMP_USE_MONITOR
3430 extern size_t __kmp_monitor_stksize; /* stack size for monitor thread */
3431 #endif
3432 extern size_t __kmp_stkoffset; /* stack offset per thread */
3433 extern int __kmp_stkpadding; /* Should we pad root thread(s) stack */
3434 
3435 extern size_t
3436  __kmp_malloc_pool_incr; /* incremental size of pool for kmp_malloc() */
3437 extern int __kmp_env_stksize; /* was KMP_STACKSIZE specified? */
3438 extern int __kmp_env_blocktime; /* was KMP_BLOCKTIME specified? */
3439 extern int __kmp_env_checks; /* was KMP_CHECKS specified? */
3440 extern int __kmp_env_consistency_check; // was KMP_CONSISTENCY_CHECK specified?
3441 extern int __kmp_generate_warnings; /* should we issue warnings? */
3442 extern int __kmp_reserve_warn; /* have we issued reserve_threads warning? */
3443 
3444 #ifdef DEBUG_SUSPEND
3445 extern int __kmp_suspend_count; /* count inside __kmp_suspend_template() */
3446 #endif
3447 
3448 extern kmp_int32 __kmp_use_yield;
3449 extern kmp_int32 __kmp_use_yield_exp_set;
3450 extern kmp_uint32 __kmp_yield_init;
3451 extern kmp_uint32 __kmp_yield_next;
3452 extern kmp_uint64 __kmp_pause_init;
3453 
3454 /* ------------------------------------------------------------------------- */
3455 extern int __kmp_allThreadsSpecified;
3456 
3457 extern size_t __kmp_align_alloc;
3458 /* following data protected by initialization routines */
3459 extern int __kmp_xproc; /* number of processors in the system */
3460 extern int __kmp_avail_proc; /* number of processors available to the process */
3461 extern size_t __kmp_sys_min_stksize; /* system-defined minimum stack size */
3462 extern int __kmp_sys_max_nth; /* system-imposed maximum number of threads */
3463 // maximum total number of concurrently-existing threads on device
3464 extern int __kmp_max_nth;
3465 // maximum total number of concurrently-existing threads in a contention group
3466 extern int __kmp_cg_max_nth;
3467 extern int __kmp_task_max_nth; // max threads used in a task
3468 extern int __kmp_teams_max_nth; // max threads used in a teams construct
3469 extern int __kmp_threads_capacity; /* capacity of the arrays __kmp_threads and
3470  __kmp_root */
3471 extern int __kmp_dflt_team_nth; /* default number of threads in a parallel
3472  region a la OMP_NUM_THREADS */
3473 extern int __kmp_dflt_team_nth_ub; /* upper bound on "" determined at serial
3474  initialization */
3475 extern int __kmp_tp_capacity; /* capacity of __kmp_threads if threadprivate is
3476  used (fixed) */
3477 extern int __kmp_tp_cached; /* whether threadprivate cache has been created
3478  (__kmpc_threadprivate_cached()) */
3479 extern int __kmp_dflt_blocktime; /* number of microseconds to wait before
3480  blocking (env setting) */
3481 extern char __kmp_blocktime_units; /* 'm' or 'u' to note units specified */
3482 extern bool __kmp_wpolicy_passive; /* explicitly set passive wait policy */
3483 
3484 // Convert raw blocktime from ms to us if needed.
3485 static inline void __kmp_aux_convert_blocktime(int *bt) {
3486  if (__kmp_blocktime_units == 'm') {
3487  if (*bt > INT_MAX / 1000) {
3488  *bt = INT_MAX / 1000;
3489  KMP_INFORM(MaxValueUsing, "kmp_set_blocktime(ms)", bt);
3490  }
3491  *bt = *bt * 1000;
3492  }
3493 }
3494 
3495 #if KMP_USE_MONITOR
3496 extern int
3497  __kmp_monitor_wakeups; /* number of times monitor wakes up per second */
3498 extern int __kmp_bt_intervals; /* number of monitor timestamp intervals before
3499  blocking */
3500 #endif
3501 #ifdef KMP_ADJUST_BLOCKTIME
3502 extern int __kmp_zero_bt; /* whether blocktime has been forced to zero */
3503 #endif /* KMP_ADJUST_BLOCKTIME */
3504 #ifdef KMP_DFLT_NTH_CORES
3505 extern int __kmp_ncores; /* Total number of cores for threads placement */
3506 #endif
3507 /* Number of millisecs to delay on abort for Intel(R) VTune(TM) tools */
3508 extern int __kmp_abort_delay;
3509 
3510 extern int __kmp_need_register_atfork_specified;
3511 extern int __kmp_need_register_atfork; /* At initialization, call pthread_atfork
3512  to install fork handler */
3513 extern int __kmp_gtid_mode; /* Method of getting gtid, values:
3514  0 - not set, will be set at runtime
3515  1 - using stack search
3516  2 - dynamic TLS (pthread_getspecific(Linux* OS/OS
3517  X*) or TlsGetValue(Windows* OS))
3518  3 - static TLS (__declspec(thread) __kmp_gtid),
3519  Linux* OS .so only. */
3520 extern int
3521  __kmp_adjust_gtid_mode; /* If true, adjust method based on #threads */
3522 #ifdef KMP_TDATA_GTID
3523 extern KMP_THREAD_LOCAL int __kmp_gtid;
3524 #endif
3525 extern int __kmp_tls_gtid_min; /* #threads below which use sp search for gtid */
3526 extern int __kmp_foreign_tp; // If true, separate TP var for each foreign thread
3527 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
3528 extern int __kmp_inherit_fp_control; // copy fp creg(s) parent->workers at fork
3529 extern kmp_int16 __kmp_init_x87_fpu_control_word; // init thread's FP ctrl reg
3530 extern kmp_uint32 __kmp_init_mxcsr; /* init thread's mxscr */
3531 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
3532 
3533 // max_active_levels for nested parallelism enabled by default via
3534 // OMP_MAX_ACTIVE_LEVELS, OMP_NESTED, OMP_NUM_THREADS, and OMP_PROC_BIND
3535 extern int __kmp_dflt_max_active_levels;
3536 // Indicates whether value of __kmp_dflt_max_active_levels was already
3537 // explicitly set by OMP_MAX_ACTIVE_LEVELS or OMP_NESTED=false
3538 extern bool __kmp_dflt_max_active_levels_set;
3539 extern int __kmp_dispatch_num_buffers; /* max possible dynamic loops in
3540  concurrent execution per team */
3541 extern int __kmp_hot_teams_mode;
3542 extern int __kmp_hot_teams_max_level;
3543 
3544 #if KMP_MIC_SUPPORTED
3545 extern enum mic_type __kmp_mic_type;
3546 #endif
3547 
3548 #ifdef USE_LOAD_BALANCE
3549 extern double __kmp_load_balance_interval; // load balance algorithm interval
3550 #endif /* USE_LOAD_BALANCE */
3551 
3552 #if KMP_USE_ADAPTIVE_LOCKS
3553 
3554 // Parameters for the speculative lock backoff system.
3555 struct kmp_adaptive_backoff_params_t {
3556  // Number of soft retries before it counts as a hard retry.
3557  kmp_uint32 max_soft_retries;
3558  // Badness is a bit mask : 0,1,3,7,15,... on each hard failure we move one to
3559  // the right
3560  kmp_uint32 max_badness;
3561 };
3562 
3563 extern kmp_adaptive_backoff_params_t __kmp_adaptive_backoff_params;
3564 
3565 #if KMP_DEBUG_ADAPTIVE_LOCKS
3566 extern const char *__kmp_speculative_statsfile;
3567 #endif
3568 
3569 #endif // KMP_USE_ADAPTIVE_LOCKS
3570 
3571 extern int __kmp_display_env; /* TRUE or FALSE */
3572 extern int __kmp_display_env_verbose; /* TRUE if OMP_DISPLAY_ENV=VERBOSE */
3573 extern int __kmp_omp_cancellation; /* TRUE or FALSE */
3574 extern int __kmp_nteams;
3575 extern int __kmp_teams_thread_limit;
3576 
3577 /* ------------------------------------------------------------------------- */
3578 
3579 /* the following are protected by the fork/join lock */
3580 /* write: lock read: anytime */
3581 extern kmp_info_t **__kmp_threads; /* Descriptors for the threads */
3582 /* Holds old arrays of __kmp_threads until library shutdown */
3583 extern kmp_old_threads_list_t *__kmp_old_threads_list;
3584 /* read/write: lock */
3585 extern volatile kmp_team_t *__kmp_team_pool;
3586 extern volatile kmp_info_t *__kmp_thread_pool;
3587 extern kmp_info_t *__kmp_thread_pool_insert_pt;
3588 
3589 // total num threads reachable from some root thread including all root threads
3590 extern volatile int __kmp_nth;
3591 /* total number of threads reachable from some root thread including all root
3592  threads, and those in the thread pool */
3593 extern volatile int __kmp_all_nth;
3594 extern std::atomic<int> __kmp_thread_pool_active_nth;
3595 
3596 extern kmp_root_t **__kmp_root; /* root of thread hierarchy */
3597 /* end data protected by fork/join lock */
3598 /* ------------------------------------------------------------------------- */
3599 
3600 #define __kmp_get_gtid() __kmp_get_global_thread_id()
3601 #define __kmp_entry_gtid() __kmp_get_global_thread_id_reg()
3602 #define __kmp_get_tid() (__kmp_tid_from_gtid(__kmp_get_gtid()))
3603 #define __kmp_get_team() (__kmp_threads[(__kmp_get_gtid())]->th.th_team)
3604 #define __kmp_get_thread() (__kmp_thread_from_gtid(__kmp_get_gtid()))
3605 
3606 // AT: Which way is correct?
3607 // AT: 1. nproc = __kmp_threads[ ( gtid ) ] -> th.th_team -> t.t_nproc;
3608 // AT: 2. nproc = __kmp_threads[ ( gtid ) ] -> th.th_team_nproc;
3609 #define __kmp_get_team_num_threads(gtid) \
3610  (__kmp_threads[(gtid)]->th.th_team->t.t_nproc)
3611 
3612 static inline bool KMP_UBER_GTID(int gtid) {
3613  KMP_DEBUG_ASSERT(gtid >= KMP_GTID_MIN);
3614  KMP_DEBUG_ASSERT(gtid < __kmp_threads_capacity);
3615  return (gtid >= 0 && __kmp_root[gtid] && __kmp_threads[gtid] &&
3616  __kmp_threads[gtid] == __kmp_root[gtid]->r.r_uber_thread);
3617 }
3618 
3619 static inline int __kmp_tid_from_gtid(int gtid) {
3620  KMP_DEBUG_ASSERT(gtid >= 0);
3621  return __kmp_threads[gtid]->th.th_info.ds.ds_tid;
3622 }
3623 
3624 static inline int __kmp_gtid_from_tid(int tid, const kmp_team_t *team) {
3625  KMP_DEBUG_ASSERT(tid >= 0 && team);
3626  return team->t.t_threads[tid]->th.th_info.ds.ds_gtid;
3627 }
3628 
3629 static inline int __kmp_gtid_from_thread(const kmp_info_t *thr) {
3630  KMP_DEBUG_ASSERT(thr);
3631  return thr->th.th_info.ds.ds_gtid;
3632 }
3633 
3634 static inline kmp_info_t *__kmp_thread_from_gtid(int gtid) {
3635  KMP_DEBUG_ASSERT(gtid >= 0);
3636  return __kmp_threads[gtid];
3637 }
3638 
3639 static inline kmp_team_t *__kmp_team_from_gtid(int gtid) {
3640  KMP_DEBUG_ASSERT(gtid >= 0);
3641  return __kmp_threads[gtid]->th.th_team;
3642 }
3643 
3644 static inline void __kmp_assert_valid_gtid(kmp_int32 gtid) {
3645  if (UNLIKELY(gtid < 0 || gtid >= __kmp_threads_capacity))
3646  KMP_FATAL(ThreadIdentInvalid);
3647 }
3648 
3649 #if KMP_HAVE_MWAIT || KMP_HAVE_UMWAIT
3650 extern int __kmp_user_level_mwait; // TRUE or FALSE; from KMP_USER_LEVEL_MWAIT
3651 extern int __kmp_umwait_enabled; // Runtime check if user-level mwait enabled
3652 extern int __kmp_mwait_enabled; // Runtime check if ring3 mwait is enabled
3653 extern int __kmp_mwait_hints; // Hints to pass in to mwait
3654 #endif
3655 
3656 #if KMP_HAVE_UMWAIT
3657 extern int __kmp_waitpkg_enabled; // Runtime check if waitpkg exists
3658 extern int __kmp_tpause_state; // 0 (default), 1=C0.1, 2=C0.2; from KMP_TPAUSE
3659 extern int __kmp_tpause_hint; // 1=C0.1 (default), 0=C0.2; from KMP_TPAUSE
3660 extern int __kmp_tpause_enabled; // 0 (default), 1 (KMP_TPAUSE is non-zero)
3661 #endif
3662 
3663 /* ------------------------------------------------------------------------- */
3664 
3665 extern kmp_global_t __kmp_global; /* global status */
3666 
3667 extern kmp_info_t __kmp_monitor;
3668 // For Debugging Support Library
3669 extern std::atomic<kmp_int32> __kmp_team_counter;
3670 // For Debugging Support Library
3671 extern std::atomic<kmp_int32> __kmp_task_counter;
3672 
3673 #if USE_DEBUGGER
3674 #define _KMP_GEN_ID(counter) \
3675  (__kmp_debugging ? KMP_ATOMIC_INC(&counter) + 1 : ~0)
3676 #else
3677 #define _KMP_GEN_ID(counter) (~0)
3678 #endif /* USE_DEBUGGER */
3679 
3680 #define KMP_GEN_TASK_ID() _KMP_GEN_ID(__kmp_task_counter)
3681 #define KMP_GEN_TEAM_ID() _KMP_GEN_ID(__kmp_team_counter)
3682 
3683 /* ------------------------------------------------------------------------ */
3684 
3685 extern void __kmp_print_storage_map_gtid(int gtid, void *p1, void *p2,
3686  size_t size, char const *format, ...);
3687 
3688 extern void __kmp_serial_initialize(void);
3689 extern void __kmp_middle_initialize(void);
3690 extern void __kmp_parallel_initialize(void);
3691 
3692 extern void __kmp_internal_begin(void);
3693 extern void __kmp_internal_end_library(int gtid);
3694 extern void __kmp_internal_end_thread(int gtid);
3695 extern void __kmp_internal_end_atexit(void);
3696 extern void __kmp_internal_end_dtor(void);
3697 extern void __kmp_internal_end_dest(void *);
3698 
3699 extern int __kmp_register_root(int initial_thread);
3700 extern void __kmp_unregister_root(int gtid);
3701 extern void __kmp_unregister_library(void); // called by __kmp_internal_end()
3702 
3703 extern int __kmp_ignore_mppbeg(void);
3704 extern int __kmp_ignore_mppend(void);
3705 
3706 extern int __kmp_enter_single(int gtid, ident_t *id_ref, int push_ws);
3707 extern void __kmp_exit_single(int gtid);
3708 
3709 extern void __kmp_parallel_deo(int *gtid_ref, int *cid_ref, ident_t *loc_ref);
3710 extern void __kmp_parallel_dxo(int *gtid_ref, int *cid_ref, ident_t *loc_ref);
3711 
3712 #ifdef USE_LOAD_BALANCE
3713 extern int __kmp_get_load_balance(int);
3714 #endif
3715 
3716 extern int __kmp_get_global_thread_id(void);
3717 extern int __kmp_get_global_thread_id_reg(void);
3718 extern void __kmp_exit_thread(int exit_status);
3719 extern void __kmp_abort(char const *format, ...);
3720 extern void __kmp_abort_thread(void);
3721 KMP_NORETURN extern void __kmp_abort_process(void);
3722 extern void __kmp_warn(char const *format, ...);
3723 
3724 extern void __kmp_set_num_threads(int new_nth, int gtid);
3725 
3726 extern bool __kmp_detect_shm();
3727 extern bool __kmp_detect_tmp();
3728 
3729 // Returns current thread (pointer to kmp_info_t). Current thread *must* be
3730 // registered.
3731 static inline kmp_info_t *__kmp_entry_thread() {
3732  int gtid = __kmp_entry_gtid();
3733 
3734  return __kmp_threads[gtid];
3735 }
3736 
3737 extern void __kmp_set_max_active_levels(int gtid, int new_max_active_levels);
3738 extern int __kmp_get_max_active_levels(int gtid);
3739 extern int __kmp_get_ancestor_thread_num(int gtid, int level);
3740 extern int __kmp_get_team_size(int gtid, int level);
3741 extern void __kmp_set_schedule(int gtid, kmp_sched_t new_sched, int chunk);
3742 extern void __kmp_get_schedule(int gtid, kmp_sched_t *sched, int *chunk);
3743 
3744 extern unsigned short __kmp_get_random(kmp_info_t *thread);
3745 extern void __kmp_init_random(kmp_info_t *thread);
3746 
3747 extern kmp_r_sched_t __kmp_get_schedule_global(void);
3748 extern void __kmp_adjust_num_threads(int new_nproc);
3749 extern void __kmp_check_stksize(size_t *val);
3750 
3751 extern void *___kmp_allocate(size_t size KMP_SRC_LOC_DECL);
3752 extern void *___kmp_page_allocate(size_t size KMP_SRC_LOC_DECL);
3753 extern void ___kmp_free(void *ptr KMP_SRC_LOC_DECL);
3754 #define __kmp_allocate(size) ___kmp_allocate((size)KMP_SRC_LOC_CURR)
3755 #define __kmp_page_allocate(size) ___kmp_page_allocate((size)KMP_SRC_LOC_CURR)
3756 #define __kmp_free(ptr) ___kmp_free((ptr)KMP_SRC_LOC_CURR)
3757 
3758 #if USE_FAST_MEMORY
3759 extern void *___kmp_fast_allocate(kmp_info_t *this_thr,
3760  size_t size KMP_SRC_LOC_DECL);
3761 extern void ___kmp_fast_free(kmp_info_t *this_thr, void *ptr KMP_SRC_LOC_DECL);
3762 extern void __kmp_free_fast_memory(kmp_info_t *this_thr);
3763 extern void __kmp_initialize_fast_memory(kmp_info_t *this_thr);
3764 #define __kmp_fast_allocate(this_thr, size) \
3765  ___kmp_fast_allocate((this_thr), (size)KMP_SRC_LOC_CURR)
3766 #define __kmp_fast_free(this_thr, ptr) \
3767  ___kmp_fast_free((this_thr), (ptr)KMP_SRC_LOC_CURR)
3768 #endif
3769 
3770 extern void *___kmp_thread_malloc(kmp_info_t *th, size_t size KMP_SRC_LOC_DECL);
3771 extern void *___kmp_thread_calloc(kmp_info_t *th, size_t nelem,
3772  size_t elsize KMP_SRC_LOC_DECL);
3773 extern void *___kmp_thread_realloc(kmp_info_t *th, void *ptr,
3774  size_t size KMP_SRC_LOC_DECL);
3775 extern void ___kmp_thread_free(kmp_info_t *th, void *ptr KMP_SRC_LOC_DECL);
3776 #define __kmp_thread_malloc(th, size) \
3777  ___kmp_thread_malloc((th), (size)KMP_SRC_LOC_CURR)
3778 #define __kmp_thread_calloc(th, nelem, elsize) \
3779  ___kmp_thread_calloc((th), (nelem), (elsize)KMP_SRC_LOC_CURR)
3780 #define __kmp_thread_realloc(th, ptr, size) \
3781  ___kmp_thread_realloc((th), (ptr), (size)KMP_SRC_LOC_CURR)
3782 #define __kmp_thread_free(th, ptr) \
3783  ___kmp_thread_free((th), (ptr)KMP_SRC_LOC_CURR)
3784 
3785 extern void __kmp_push_num_threads(ident_t *loc, int gtid, int num_threads);
3786 extern void __kmp_push_num_threads_list(ident_t *loc, int gtid,
3787  kmp_uint32 list_length,
3788  int *num_threads_list);
3789 extern void __kmp_set_strict_num_threads(ident_t *loc, int gtid, int sev,
3790  const char *msg);
3791 
3792 extern void __kmp_push_proc_bind(ident_t *loc, int gtid,
3793  kmp_proc_bind_t proc_bind);
3794 extern void __kmp_push_num_teams(ident_t *loc, int gtid, int num_teams,
3795  int num_threads);
3796 extern void __kmp_push_num_teams_51(ident_t *loc, int gtid, int num_teams_lb,
3797  int num_teams_ub, int num_threads);
3798 
3799 extern void __kmp_yield();
3800 
3801 extern void __kmpc_dispatch_init_4(ident_t *loc, kmp_int32 gtid,
3802  enum sched_type schedule, kmp_int32 lb,
3803  kmp_int32 ub, kmp_int32 st, kmp_int32 chunk);
3804 extern void __kmpc_dispatch_init_4u(ident_t *loc, kmp_int32 gtid,
3805  enum sched_type schedule, kmp_uint32 lb,
3806  kmp_uint32 ub, kmp_int32 st,
3807  kmp_int32 chunk);
3808 extern void __kmpc_dispatch_init_8(ident_t *loc, kmp_int32 gtid,
3809  enum sched_type schedule, kmp_int64 lb,
3810  kmp_int64 ub, kmp_int64 st, kmp_int64 chunk);
3811 extern void __kmpc_dispatch_init_8u(ident_t *loc, kmp_int32 gtid,
3812  enum sched_type schedule, kmp_uint64 lb,
3813  kmp_uint64 ub, kmp_int64 st,
3814  kmp_int64 chunk);
3815 
3816 extern int __kmpc_dispatch_next_4(ident_t *loc, kmp_int32 gtid,
3817  kmp_int32 *p_last, kmp_int32 *p_lb,
3818  kmp_int32 *p_ub, kmp_int32 *p_st);
3819 extern int __kmpc_dispatch_next_4u(ident_t *loc, kmp_int32 gtid,
3820  kmp_int32 *p_last, kmp_uint32 *p_lb,
3821  kmp_uint32 *p_ub, kmp_int32 *p_st);
3822 extern int __kmpc_dispatch_next_8(ident_t *loc, kmp_int32 gtid,
3823  kmp_int32 *p_last, kmp_int64 *p_lb,
3824  kmp_int64 *p_ub, kmp_int64 *p_st);
3825 extern int __kmpc_dispatch_next_8u(ident_t *loc, kmp_int32 gtid,
3826  kmp_int32 *p_last, kmp_uint64 *p_lb,
3827  kmp_uint64 *p_ub, kmp_int64 *p_st);
3828 
3829 extern void __kmpc_dispatch_fini_4(ident_t *loc, kmp_int32 gtid);
3830 extern void __kmpc_dispatch_fini_8(ident_t *loc, kmp_int32 gtid);
3831 extern void __kmpc_dispatch_fini_4u(ident_t *loc, kmp_int32 gtid);
3832 extern void __kmpc_dispatch_fini_8u(ident_t *loc, kmp_int32 gtid);
3833 
3834 extern void __kmpc_dispatch_deinit(ident_t *loc, kmp_int32 gtid);
3835 
3836 #ifdef KMP_GOMP_COMPAT
3837 
3838 extern void __kmp_aux_dispatch_init_4(ident_t *loc, kmp_int32 gtid,
3839  enum sched_type schedule, kmp_int32 lb,
3840  kmp_int32 ub, kmp_int32 st,
3841  kmp_int32 chunk, int push_ws);
3842 extern void __kmp_aux_dispatch_init_4u(ident_t *loc, kmp_int32 gtid,
3843  enum sched_type schedule, kmp_uint32 lb,
3844  kmp_uint32 ub, kmp_int32 st,
3845  kmp_int32 chunk, int push_ws);
3846 extern void __kmp_aux_dispatch_init_8(ident_t *loc, kmp_int32 gtid,
3847  enum sched_type schedule, kmp_int64 lb,
3848  kmp_int64 ub, kmp_int64 st,
3849  kmp_int64 chunk, int push_ws);
3850 extern void __kmp_aux_dispatch_init_8u(ident_t *loc, kmp_int32 gtid,
3851  enum sched_type schedule, kmp_uint64 lb,
3852  kmp_uint64 ub, kmp_int64 st,
3853  kmp_int64 chunk, int push_ws);
3854 extern void __kmp_aux_dispatch_fini_chunk_4(ident_t *loc, kmp_int32 gtid);
3855 extern void __kmp_aux_dispatch_fini_chunk_8(ident_t *loc, kmp_int32 gtid);
3856 extern void __kmp_aux_dispatch_fini_chunk_4u(ident_t *loc, kmp_int32 gtid);
3857 extern void __kmp_aux_dispatch_fini_chunk_8u(ident_t *loc, kmp_int32 gtid);
3858 
3859 #endif /* KMP_GOMP_COMPAT */
3860 
3861 extern kmp_uint32 __kmp_eq_4(kmp_uint32 value, kmp_uint32 checker);
3862 extern kmp_uint32 __kmp_neq_4(kmp_uint32 value, kmp_uint32 checker);
3863 extern kmp_uint32 __kmp_lt_4(kmp_uint32 value, kmp_uint32 checker);
3864 extern kmp_uint32 __kmp_ge_4(kmp_uint32 value, kmp_uint32 checker);
3865 extern kmp_uint32 __kmp_le_4(kmp_uint32 value, kmp_uint32 checker);
3866 extern kmp_uint32 __kmp_wait_4(kmp_uint32 volatile *spinner, kmp_uint32 checker,
3867  kmp_uint32 (*pred)(kmp_uint32, kmp_uint32),
3868  void *obj);
3869 extern void __kmp_wait_4_ptr(void *spinner, kmp_uint32 checker,
3870  kmp_uint32 (*pred)(void *, kmp_uint32), void *obj);
3871 
3872 extern void __kmp_wait_64(kmp_info_t *this_thr, kmp_flag_64<> *flag,
3873  int final_spin
3874 #if USE_ITT_BUILD
3875  ,
3876  void *itt_sync_obj
3877 #endif
3878 );
3879 extern void __kmp_release_64(kmp_flag_64<> *flag);
3880 
3881 extern void __kmp_infinite_loop(void);
3882 
3883 extern void __kmp_cleanup(void);
3884 
3885 #if KMP_HANDLE_SIGNALS
3886 extern int __kmp_handle_signals;
3887 extern void __kmp_install_signals(int parallel_init);
3888 extern void __kmp_remove_signals(void);
3889 #endif
3890 
3891 extern void __kmp_clear_system_time(void);
3892 extern void __kmp_read_system_time(double *delta);
3893 
3894 extern void __kmp_check_stack_overlap(kmp_info_t *thr);
3895 
3896 extern void __kmp_expand_host_name(char *buffer, size_t size);
3897 extern void __kmp_expand_file_name(char *result, size_t rlen, char *pattern);
3898 
3899 #if KMP_ARCH_X86 || KMP_ARCH_X86_64 || (KMP_OS_WINDOWS && (KMP_ARCH_AARCH64 || KMP_ARCH_ARM))
3900 extern void
3901 __kmp_initialize_system_tick(void); /* Initialize timer tick value */
3902 #endif
3903 
3904 extern void
3905 __kmp_runtime_initialize(void); /* machine specific initialization */
3906 extern void __kmp_runtime_destroy(void);
3907 
3908 #if KMP_AFFINITY_SUPPORTED
3909 extern char *__kmp_affinity_print_mask(char *buf, int buf_len,
3910  kmp_affin_mask_t *mask);
3911 extern kmp_str_buf_t *__kmp_affinity_str_buf_mask(kmp_str_buf_t *buf,
3912  kmp_affin_mask_t *mask);
3913 extern void __kmp_affinity_initialize(kmp_affinity_t &affinity);
3914 extern void __kmp_affinity_uninitialize(void);
3915 extern void __kmp_affinity_set_init_mask(
3916  int gtid, int isa_root); /* set affinity according to KMP_AFFINITY */
3917 void __kmp_affinity_bind_init_mask(int gtid);
3918 extern void __kmp_affinity_bind_place(int gtid);
3919 extern void __kmp_affinity_determine_capable(const char *env_var);
3920 extern int __kmp_aux_set_affinity(void **mask);
3921 extern int __kmp_aux_get_affinity(void **mask);
3922 extern int __kmp_aux_get_affinity_max_proc();
3923 extern int __kmp_aux_set_affinity_mask_proc(int proc, void **mask);
3924 extern int __kmp_aux_unset_affinity_mask_proc(int proc, void **mask);
3925 extern int __kmp_aux_get_affinity_mask_proc(int proc, void **mask);
3926 extern void __kmp_balanced_affinity(kmp_info_t *th, int team_size);
3927 #if KMP_WEIGHTED_ITERATIONS_SUPPORTED
3928 extern int __kmp_get_first_osid_with_ecore(void);
3929 #endif
3930 #if KMP_OS_LINUX || KMP_OS_FREEBSD || KMP_OS_NETBSD || KMP_OS_DRAGONFLY || \
3931  KMP_OS_AIX
3932 extern int kmp_set_thread_affinity_mask_initial(void);
3933 #endif
3934 static inline void __kmp_assign_root_init_mask() {
3935  int gtid = __kmp_entry_gtid();
3936  kmp_root_t *r = __kmp_threads[gtid]->th.th_root;
3937  if (r->r.r_uber_thread == __kmp_threads[gtid] && !r->r.r_affinity_assigned) {
3938  __kmp_affinity_set_init_mask(gtid, /*isa_root=*/TRUE);
3939  __kmp_affinity_bind_init_mask(gtid);
3940  r->r.r_affinity_assigned = TRUE;
3941  }
3942 }
3943 static inline void __kmp_reset_root_init_mask(int gtid) {
3944  if (!KMP_AFFINITY_CAPABLE())
3945  return;
3946  kmp_info_t *th = __kmp_threads[gtid];
3947  kmp_root_t *r = th->th.th_root;
3948  if (r->r.r_uber_thread == th && r->r.r_affinity_assigned) {
3949  __kmp_set_system_affinity(__kmp_affin_origMask, FALSE);
3950  KMP_CPU_COPY(th->th.th_affin_mask, __kmp_affin_origMask);
3951  r->r.r_affinity_assigned = FALSE;
3952  }
3953 }
3954 #else /* KMP_AFFINITY_SUPPORTED */
3955 #define __kmp_assign_root_init_mask() /* Nothing */
3956 static inline void __kmp_reset_root_init_mask(int gtid) {}
3957 #endif /* KMP_AFFINITY_SUPPORTED */
3958 // No need for KMP_AFFINITY_SUPPORTED guard as only one field in the
3959 // format string is for affinity, so platforms that do not support
3960 // affinity can still use the other fields, e.g., %n for num_threads
3961 extern size_t __kmp_aux_capture_affinity(int gtid, const char *format,
3962  kmp_str_buf_t *buffer);
3963 extern void __kmp_aux_display_affinity(int gtid, const char *format);
3964 
3965 extern void __kmp_cleanup_hierarchy();
3966 extern void __kmp_get_hierarchy(kmp_uint32 nproc, kmp_bstate_t *thr_bar);
3967 
3968 #if KMP_USE_FUTEX
3969 
3970 extern int __kmp_futex_determine_capable(void);
3971 
3972 #endif // KMP_USE_FUTEX
3973 
3974 extern void __kmp_gtid_set_specific(int gtid);
3975 extern int __kmp_gtid_get_specific(void);
3976 
3977 extern double __kmp_read_cpu_time(void);
3978 
3979 extern int __kmp_read_system_info(struct kmp_sys_info *info);
3980 
3981 #if KMP_USE_MONITOR
3982 extern void __kmp_create_monitor(kmp_info_t *th);
3983 #endif
3984 
3985 extern void *__kmp_launch_thread(kmp_info_t *thr);
3986 
3987 extern void __kmp_create_worker(int gtid, kmp_info_t *th, size_t stack_size);
3988 
3989 #if KMP_OS_WINDOWS
3990 extern int __kmp_still_running(kmp_info_t *th);
3991 extern int __kmp_is_thread_alive(kmp_info_t *th, DWORD *exit_val);
3992 extern void __kmp_free_handle(kmp_thread_t tHandle);
3993 #endif
3994 
3995 #if KMP_USE_MONITOR
3996 extern void __kmp_reap_monitor(kmp_info_t *th);
3997 #endif
3998 extern void __kmp_reap_worker(kmp_info_t *th);
3999 extern void __kmp_terminate_thread(int gtid);
4000 
4001 extern int __kmp_try_suspend_mx(kmp_info_t *th);
4002 extern void __kmp_lock_suspend_mx(kmp_info_t *th);
4003 extern void __kmp_unlock_suspend_mx(kmp_info_t *th);
4004 
4005 extern void __kmp_elapsed(double *);
4006 extern void __kmp_elapsed_tick(double *);
4007 
4008 extern void __kmp_enable(int old_state);
4009 extern void __kmp_disable(int *old_state);
4010 
4011 extern void __kmp_thread_sleep(int millis);
4012 
4013 extern void __kmp_common_initialize(void);
4014 extern void __kmp_common_destroy(void);
4015 extern void __kmp_common_destroy_gtid(int gtid);
4016 
4017 #if KMP_OS_UNIX
4018 extern void __kmp_register_atfork(void);
4019 #endif
4020 extern void __kmp_suspend_initialize(void);
4021 extern void __kmp_suspend_initialize_thread(kmp_info_t *th);
4022 extern void __kmp_suspend_uninitialize_thread(kmp_info_t *th);
4023 
4024 extern kmp_info_t *__kmp_allocate_thread(kmp_root_t *root, kmp_team_t *team,
4025  int tid);
4026 extern kmp_team_t *__kmp_allocate_team(kmp_root_t *root, int new_nproc,
4027  int max_nproc,
4028 #if OMPT_SUPPORT
4029  ompt_data_t ompt_parallel_data,
4030 #endif
4031  kmp_proc_bind_t proc_bind,
4032  kmp_internal_control_t *new_icvs,
4033  int argc, kmp_info_t *thr);
4034 extern void __kmp_free_thread(kmp_info_t *);
4035 extern void __kmp_free_team(kmp_root_t *, kmp_team_t *, kmp_info_t *);
4036 extern kmp_team_t *__kmp_reap_team(kmp_team_t *);
4037 
4038 /* ------------------------------------------------------------------------ */
4039 
4040 extern void __kmp_initialize_bget(kmp_info_t *th);
4041 extern void __kmp_finalize_bget(kmp_info_t *th);
4042 
4043 KMP_EXPORT void *kmpc_malloc(size_t size);
4044 KMP_EXPORT void *kmpc_aligned_malloc(size_t size, size_t alignment);
4045 KMP_EXPORT void *kmpc_calloc(size_t nelem, size_t elsize);
4046 KMP_EXPORT void *kmpc_realloc(void *ptr, size_t size);
4047 KMP_EXPORT void kmpc_free(void *ptr);
4048 
4049 /* declarations for internal use */
4050 
4051 extern int __kmp_barrier(enum barrier_type bt, int gtid, int is_split,
4052  size_t reduce_size, void *reduce_data,
4053  void (*reduce)(void *, void *));
4054 extern void __kmp_end_split_barrier(enum barrier_type bt, int gtid);
4055 extern int __kmp_barrier_gomp_cancel(int gtid);
4056 
4061 enum fork_context_e {
4062  fork_context_gnu,
4064  fork_context_intel,
4065  fork_context_last
4066 };
4067 extern int __kmp_fork_call(ident_t *loc, int gtid,
4068  enum fork_context_e fork_context, kmp_int32 argc,
4069  microtask_t microtask, launch_t invoker,
4070  kmp_va_list ap);
4071 
4072 extern void __kmp_join_call(ident_t *loc, int gtid
4073 #if OMPT_SUPPORT
4074  ,
4075  enum fork_context_e fork_context
4076 #endif
4077  ,
4078  int exit_teams = 0);
4079 
4080 extern void __kmp_serialized_parallel(ident_t *id, kmp_int32 gtid);
4081 extern void __kmp_internal_fork(ident_t *id, int gtid, kmp_team_t *team);
4082 extern void __kmp_internal_join(ident_t *id, int gtid, kmp_team_t *team);
4083 extern int __kmp_invoke_task_func(int gtid);
4084 extern void __kmp_run_before_invoked_task(int gtid, int tid,
4085  kmp_info_t *this_thr,
4086  kmp_team_t *team);
4087 extern void __kmp_run_after_invoked_task(int gtid, int tid,
4088  kmp_info_t *this_thr,
4089  kmp_team_t *team);
4090 
4091 // should never have been exported
4092 KMP_EXPORT int __kmpc_invoke_task_func(int gtid);
4093 extern int __kmp_invoke_teams_master(int gtid);
4094 extern void __kmp_teams_master(int gtid);
4095 extern int __kmp_aux_get_team_num();
4096 extern int __kmp_aux_get_num_teams();
4097 extern void __kmp_save_internal_controls(kmp_info_t *thread);
4098 extern void __kmp_user_set_library(enum library_type arg);
4099 extern void __kmp_aux_set_library(enum library_type arg);
4100 extern void __kmp_aux_set_stacksize(size_t arg);
4101 extern void __kmp_aux_set_blocktime(int arg, kmp_info_t *thread, int tid);
4102 extern void __kmp_aux_set_defaults(char const *str, size_t len);
4103 
4104 /* Functions called from __kmp_aux_env_initialize() in kmp_settings.cpp */
4105 void kmpc_set_blocktime(int arg);
4106 void ompc_set_nested(int flag);
4107 void ompc_set_dynamic(int flag);
4108 void ompc_set_num_threads(int arg);
4109 
4110 extern void __kmp_push_current_task_to_thread(kmp_info_t *this_thr,
4111  kmp_team_t *team, int tid);
4112 extern void __kmp_pop_current_task_from_thread(kmp_info_t *this_thr);
4113 extern kmp_task_t *__kmp_task_alloc(ident_t *loc_ref, kmp_int32 gtid,
4114  kmp_tasking_flags_t *flags,
4115  size_t sizeof_kmp_task_t,
4116  size_t sizeof_shareds,
4117  kmp_routine_entry_t task_entry);
4118 extern void __kmp_init_implicit_task(ident_t *loc_ref, kmp_info_t *this_thr,
4119  kmp_team_t *team, int tid,
4120  int set_curr_task);
4121 extern void __kmp_finish_implicit_task(kmp_info_t *this_thr);
4122 extern void __kmp_free_implicit_task(kmp_info_t *this_thr);
4123 
4124 extern kmp_event_t *__kmpc_task_allow_completion_event(ident_t *loc_ref,
4125  int gtid,
4126  kmp_task_t *task);
4127 extern void __kmp_fulfill_event(kmp_event_t *event);
4128 
4129 extern void __kmp_free_task_team(kmp_info_t *thread,
4130  kmp_task_team_t *task_team);
4131 extern void __kmp_reap_task_teams(void);
4132 extern void __kmp_push_task_team_node(kmp_info_t *thread, kmp_team_t *team);
4133 extern void __kmp_pop_task_team_node(kmp_info_t *thread, kmp_team_t *team);
4134 extern void __kmp_wait_to_unref_task_teams(void);
4135 extern void __kmp_task_team_setup(kmp_info_t *this_thr, kmp_team_t *team);
4136 extern void __kmp_task_team_sync(kmp_info_t *this_thr, kmp_team_t *team);
4137 extern void __kmp_task_team_wait(kmp_info_t *this_thr, kmp_team_t *team
4138 #if USE_ITT_BUILD
4139  ,
4140  void *itt_sync_obj
4141 #endif /* USE_ITT_BUILD */
4142  ,
4143  int wait = 1);
4144 extern void __kmp_tasking_barrier(kmp_team_t *team, kmp_info_t *thread,
4145  int gtid);
4146 #if KMP_DEBUG
4147 #define KMP_DEBUG_ASSERT_TASKTEAM_INVARIANT(team, thr) \
4148  KMP_DEBUG_ASSERT( \
4149  __kmp_tasking_mode != tskm_task_teams || team->t.t_nproc == 1 || \
4150  thr->th.th_task_team == team->t.t_task_team[thr->th.th_task_state])
4151 #else
4152 #define KMP_DEBUG_ASSERT_TASKTEAM_INVARIANT(team, thr) /* Nothing */
4153 #endif
4154 
4155 extern int __kmp_is_address_mapped(void *addr);
4156 extern kmp_uint64 __kmp_hardware_timestamp(void);
4157 
4158 #if KMP_OS_UNIX
4159 extern int __kmp_read_from_file(char const *path, char const *format, ...);
4160 #endif
4161 
4162 /* ------------------------------------------------------------------------ */
4163 //
4164 // Assembly routines that have no compiler intrinsic replacement
4165 //
4166 
4167 extern int __kmp_invoke_microtask(microtask_t pkfn, int gtid, int npr, int argc,
4168  void *argv[]
4169 #if OMPT_SUPPORT
4170  ,
4171  void **exit_frame_ptr
4172 #endif
4173 );
4174 
4175 /* ------------------------------------------------------------------------ */
4176 
4177 KMP_EXPORT void __kmpc_begin(ident_t *, kmp_int32 flags);
4178 KMP_EXPORT void __kmpc_end(ident_t *);
4179 
4180 KMP_EXPORT void __kmpc_threadprivate_register_vec(ident_t *, void *data,
4181  kmpc_ctor_vec ctor,
4182  kmpc_cctor_vec cctor,
4183  kmpc_dtor_vec dtor,
4184  size_t vector_length);
4185 KMP_EXPORT void __kmpc_threadprivate_register(ident_t *, void *data,
4186  kmpc_ctor ctor, kmpc_cctor cctor,
4187  kmpc_dtor dtor);
4188 KMP_EXPORT void *__kmpc_threadprivate(ident_t *, kmp_int32 global_tid,
4189  void *data, size_t size);
4190 
4191 KMP_EXPORT kmp_int32 __kmpc_global_thread_num(ident_t *);
4192 KMP_EXPORT kmp_int32 __kmpc_global_num_threads(ident_t *);
4193 KMP_EXPORT kmp_int32 __kmpc_bound_thread_num(ident_t *);
4194 KMP_EXPORT kmp_int32 __kmpc_bound_num_threads(ident_t *);
4195 
4196 KMP_EXPORT kmp_int32 __kmpc_ok_to_fork(ident_t *);
4197 KMP_EXPORT void __kmpc_fork_call(ident_t *, kmp_int32 nargs,
4198  kmpc_micro microtask, ...);
4199 KMP_EXPORT void __kmpc_fork_call_if(ident_t *loc, kmp_int32 nargs,
4200  kmpc_micro microtask, kmp_int32 cond,
4201  void *args);
4202 
4203 KMP_EXPORT void __kmpc_serialized_parallel(ident_t *, kmp_int32 global_tid);
4204 KMP_EXPORT void __kmpc_end_serialized_parallel(ident_t *, kmp_int32 global_tid);
4205 
4206 KMP_EXPORT void __kmpc_flush(ident_t *);
4207 KMP_EXPORT void __kmpc_barrier(ident_t *, kmp_int32 global_tid);
4208 KMP_EXPORT kmp_int32 __kmpc_master(ident_t *, kmp_int32 global_tid);
4209 KMP_EXPORT void __kmpc_end_master(ident_t *, kmp_int32 global_tid);
4210 KMP_EXPORT kmp_int32 __kmpc_masked(ident_t *, kmp_int32 global_tid,
4211  kmp_int32 filter);
4212 KMP_EXPORT void __kmpc_end_masked(ident_t *, kmp_int32 global_tid);
4213 KMP_EXPORT void __kmpc_ordered(ident_t *, kmp_int32 global_tid);
4214 KMP_EXPORT void __kmpc_end_ordered(ident_t *, kmp_int32 global_tid);
4215 KMP_EXPORT void __kmpc_critical(ident_t *, kmp_int32 global_tid,
4216  kmp_critical_name *);
4217 KMP_EXPORT void __kmpc_end_critical(ident_t *, kmp_int32 global_tid,
4218  kmp_critical_name *);
4219 KMP_EXPORT void __kmpc_critical_with_hint(ident_t *, kmp_int32 global_tid,
4220  kmp_critical_name *, uint32_t hint);
4221 
4222 KMP_EXPORT kmp_int32 __kmpc_barrier_master(ident_t *, kmp_int32 global_tid);
4223 KMP_EXPORT void __kmpc_end_barrier_master(ident_t *, kmp_int32 global_tid);
4224 
4225 KMP_EXPORT kmp_int32 __kmpc_barrier_master_nowait(ident_t *,
4226  kmp_int32 global_tid);
4227 
4228 KMP_EXPORT kmp_int32 __kmpc_single(ident_t *, kmp_int32 global_tid);
4229 KMP_EXPORT void __kmpc_end_single(ident_t *, kmp_int32 global_tid);
4230 
4231 KMP_EXPORT kmp_int32 __kmpc_sections_init(ident_t *loc, kmp_int32 global_tid);
4232 KMP_EXPORT kmp_int32 __kmpc_next_section(ident_t *loc, kmp_int32 global_tid,
4233  kmp_int32 numberOfSections);
4234 KMP_EXPORT void __kmpc_end_sections(ident_t *loc, kmp_int32 global_tid);
4235 
4236 KMP_EXPORT void KMPC_FOR_STATIC_INIT(ident_t *loc, kmp_int32 global_tid,
4237  kmp_int32 schedtype, kmp_int32 *plastiter,
4238  kmp_int *plower, kmp_int *pupper,
4239  kmp_int *pstride, kmp_int incr,
4240  kmp_int chunk);
4241 
4242 KMP_EXPORT void __kmpc_for_static_fini(ident_t *loc, kmp_int32 global_tid);
4243 
4244 KMP_EXPORT void __kmpc_copyprivate(ident_t *loc, kmp_int32 global_tid,
4245  size_t cpy_size, void *cpy_data,
4246  void (*cpy_func)(void *, void *),
4247  kmp_int32 didit);
4248 
4249 KMP_EXPORT void *__kmpc_copyprivate_light(ident_t *loc, kmp_int32 gtid,
4250  void *cpy_data);
4251 
4252 extern void KMPC_SET_NUM_THREADS(int arg);
4253 extern void KMPC_SET_DYNAMIC(int flag);
4254 extern void KMPC_SET_NESTED(int flag);
4255 
4256 /* OMP 3.0 tasking interface routines */
4257 KMP_EXPORT kmp_int32 __kmpc_omp_task(ident_t *loc_ref, kmp_int32 gtid,
4258  kmp_task_t *new_task);
4259 KMP_EXPORT kmp_task_t *__kmpc_omp_task_alloc(ident_t *loc_ref, kmp_int32 gtid,
4260  kmp_int32 flags,
4261  size_t sizeof_kmp_task_t,
4262  size_t sizeof_shareds,
4263  kmp_routine_entry_t task_entry);
4264 KMP_EXPORT kmp_task_t *__kmpc_omp_target_task_alloc(
4265  ident_t *loc_ref, kmp_int32 gtid, kmp_int32 flags, size_t sizeof_kmp_task_t,
4266  size_t sizeof_shareds, kmp_routine_entry_t task_entry, kmp_int64 device_id);
4267 KMP_EXPORT void __kmpc_omp_task_begin_if0(ident_t *loc_ref, kmp_int32 gtid,
4268  kmp_task_t *task);
4269 KMP_EXPORT void __kmpc_omp_task_complete_if0(ident_t *loc_ref, kmp_int32 gtid,
4270  kmp_task_t *task);
4271 KMP_EXPORT kmp_int32 __kmpc_omp_task_parts(ident_t *loc_ref, kmp_int32 gtid,
4272  kmp_task_t *new_task);
4273 KMP_EXPORT kmp_int32 __kmpc_omp_taskwait(ident_t *loc_ref, kmp_int32 gtid);
4274 KMP_EXPORT kmp_int32 __kmpc_omp_taskyield(ident_t *loc_ref, kmp_int32 gtid,
4275  int end_part);
4276 
4277 #if TASK_UNUSED
4278 void __kmpc_omp_task_begin(ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *task);
4279 void __kmpc_omp_task_complete(ident_t *loc_ref, kmp_int32 gtid,
4280  kmp_task_t *task);
4281 #endif // TASK_UNUSED
4282 
4283 /* ------------------------------------------------------------------------ */
4284 
4285 KMP_EXPORT void __kmpc_taskgroup(ident_t *loc, int gtid);
4286 KMP_EXPORT void __kmpc_end_taskgroup(ident_t *loc, int gtid);
4287 
4288 KMP_EXPORT kmp_int32 __kmpc_omp_task_with_deps(
4289  ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 ndeps,
4290  kmp_depend_info_t *dep_list, kmp_int32 ndeps_noalias,
4291  kmp_depend_info_t *noalias_dep_list);
4292 
4293 KMP_EXPORT kmp_base_depnode_t *__kmpc_task_get_depnode(kmp_task_t *task);
4294 
4295 KMP_EXPORT kmp_depnode_list_t *__kmpc_task_get_successors(kmp_task_t *task);
4296 
4297 KMP_EXPORT void __kmpc_omp_wait_deps(ident_t *loc_ref, kmp_int32 gtid,
4298  kmp_int32 ndeps,
4299  kmp_depend_info_t *dep_list,
4300  kmp_int32 ndeps_noalias,
4301  kmp_depend_info_t *noalias_dep_list);
4302 /* __kmpc_omp_taskwait_deps_51 : Function for OpenMP 5.1 nowait clause.
4303  * Placeholder for taskwait with nowait clause.*/
4304 KMP_EXPORT void __kmpc_omp_taskwait_deps_51(ident_t *loc_ref, kmp_int32 gtid,
4305  kmp_int32 ndeps,
4306  kmp_depend_info_t *dep_list,
4307  kmp_int32 ndeps_noalias,
4308  kmp_depend_info_t *noalias_dep_list,
4309  kmp_int32 has_no_wait);
4310 
4311 extern kmp_int32 __kmp_omp_task(kmp_int32 gtid, kmp_task_t *new_task,
4312  bool serialize_immediate);
4313 
4314 KMP_EXPORT kmp_int32 __kmpc_cancel(ident_t *loc_ref, kmp_int32 gtid,
4315  kmp_int32 cncl_kind);
4316 KMP_EXPORT kmp_int32 __kmpc_cancellationpoint(ident_t *loc_ref, kmp_int32 gtid,
4317  kmp_int32 cncl_kind);
4318 KMP_EXPORT kmp_int32 __kmpc_cancel_barrier(ident_t *loc_ref, kmp_int32 gtid);
4319 KMP_EXPORT int __kmp_get_cancellation_status(int cancel_kind);
4320 
4321 KMP_EXPORT void __kmpc_proxy_task_completed(kmp_int32 gtid, kmp_task_t *ptask);
4322 KMP_EXPORT void __kmpc_proxy_task_completed_ooo(kmp_task_t *ptask);
4323 KMP_EXPORT void __kmpc_taskloop(ident_t *loc, kmp_int32 gtid, kmp_task_t *task,
4324  kmp_int32 if_val, kmp_uint64 *lb,
4325  kmp_uint64 *ub, kmp_int64 st, kmp_int32 nogroup,
4326  kmp_int32 sched, kmp_uint64 grainsize,
4327  void *task_dup);
4328 KMP_EXPORT void __kmpc_taskloop_5(ident_t *loc, kmp_int32 gtid,
4329  kmp_task_t *task, kmp_int32 if_val,
4330  kmp_uint64 *lb, kmp_uint64 *ub, kmp_int64 st,
4331  kmp_int32 nogroup, kmp_int32 sched,
4332  kmp_uint64 grainsize, kmp_int32 modifier,
4333  void *task_dup);
4334 KMP_EXPORT void *__kmpc_task_reduction_init(int gtid, int num_data, void *data);
4335 KMP_EXPORT void *__kmpc_taskred_init(int gtid, int num_data, void *data);
4336 KMP_EXPORT void *__kmpc_task_reduction_get_th_data(int gtid, void *tg, void *d);
4337 KMP_EXPORT void *__kmpc_task_reduction_modifier_init(ident_t *loc, int gtid,
4338  int is_ws, int num,
4339  void *data);
4340 KMP_EXPORT void *__kmpc_taskred_modifier_init(ident_t *loc, int gtid, int is_ws,
4341  int num, void *data);
4342 KMP_EXPORT void __kmpc_task_reduction_modifier_fini(ident_t *loc, int gtid,
4343  int is_ws);
4344 KMP_EXPORT kmp_int32 __kmpc_omp_reg_task_with_affinity(
4345  ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 naffins,
4346  kmp_task_affinity_info_t *affin_list);
4347 KMP_EXPORT void __kmp_set_num_teams(int num_teams);
4348 KMP_EXPORT int __kmp_get_max_teams(void);
4349 KMP_EXPORT void __kmp_set_teams_thread_limit(int limit);
4350 KMP_EXPORT int __kmp_get_teams_thread_limit(void);
4351 
4352 /* Interface target task integration */
4353 KMP_EXPORT void **__kmpc_omp_get_target_async_handle_ptr(kmp_int32 gtid);
4354 KMP_EXPORT bool __kmpc_omp_has_task_team(kmp_int32 gtid);
4355 
4356 /* Lock interface routines (fast versions with gtid passed in) */
4357 KMP_EXPORT void __kmpc_init_lock(ident_t *loc, kmp_int32 gtid,
4358  void **user_lock);
4359 KMP_EXPORT void __kmpc_init_nest_lock(ident_t *loc, kmp_int32 gtid,
4360  void **user_lock);
4361 KMP_EXPORT void __kmpc_destroy_lock(ident_t *loc, kmp_int32 gtid,
4362  void **user_lock);
4363 KMP_EXPORT void __kmpc_destroy_nest_lock(ident_t *loc, kmp_int32 gtid,
4364  void **user_lock);
4365 KMP_EXPORT void __kmpc_set_lock(ident_t *loc, kmp_int32 gtid, void **user_lock);
4366 KMP_EXPORT void __kmpc_set_nest_lock(ident_t *loc, kmp_int32 gtid,
4367  void **user_lock);
4368 KMP_EXPORT void __kmpc_unset_lock(ident_t *loc, kmp_int32 gtid,
4369  void **user_lock);
4370 KMP_EXPORT void __kmpc_unset_nest_lock(ident_t *loc, kmp_int32 gtid,
4371  void **user_lock);
4372 KMP_EXPORT int __kmpc_test_lock(ident_t *loc, kmp_int32 gtid, void **user_lock);
4373 KMP_EXPORT int __kmpc_test_nest_lock(ident_t *loc, kmp_int32 gtid,
4374  void **user_lock);
4375 
4376 KMP_EXPORT void __kmpc_init_lock_with_hint(ident_t *loc, kmp_int32 gtid,
4377  void **user_lock, uintptr_t hint);
4378 KMP_EXPORT void __kmpc_init_nest_lock_with_hint(ident_t *loc, kmp_int32 gtid,
4379  void **user_lock,
4380  uintptr_t hint);
4381 
4382 #if OMPX_TASKGRAPH
4383 // Taskgraph's Record & Replay mechanism
4384 // __kmp_tdg_is_recording: check whether a given TDG is recording
4385 // status: the tdg's current status
4386 static inline bool __kmp_tdg_is_recording(kmp_tdg_status_t status) {
4387  return status == KMP_TDG_RECORDING;
4388 }
4389 
4390 KMP_EXPORT kmp_int32 __kmpc_start_record_task(ident_t *loc, kmp_int32 gtid,
4391  kmp_int32 input_flags,
4392  kmp_int32 tdg_id);
4393 KMP_EXPORT void __kmpc_end_record_task(ident_t *loc, kmp_int32 gtid,
4394  kmp_int32 input_flags, kmp_int32 tdg_id);
4395 #endif
4396 /* Interface to fast scalable reduce methods routines */
4397 
4398 KMP_EXPORT kmp_int32 __kmpc_reduce_nowait(
4399  ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
4400  void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
4401  kmp_critical_name *lck);
4402 KMP_EXPORT void __kmpc_end_reduce_nowait(ident_t *loc, kmp_int32 global_tid,
4403  kmp_critical_name *lck);
4404 KMP_EXPORT kmp_int32 __kmpc_reduce(
4405  ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
4406  void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
4407  kmp_critical_name *lck);
4408 KMP_EXPORT void __kmpc_end_reduce(ident_t *loc, kmp_int32 global_tid,
4409  kmp_critical_name *lck);
4410 
4411 /* Internal fast reduction routines */
4412 
4413 extern PACKED_REDUCTION_METHOD_T __kmp_determine_reduction_method(
4414  ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
4415  void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
4416  kmp_critical_name *lck);
4417 
4418 // this function is for testing set/get/determine reduce method
4419 KMP_EXPORT kmp_int32 __kmp_get_reduce_method(void);
4420 
4421 KMP_EXPORT kmp_uint64 __kmpc_get_taskid();
4422 KMP_EXPORT kmp_uint64 __kmpc_get_parent_taskid();
4423 
4424 // C++ port
4425 // missing 'extern "C"' declarations
4426 
4427 KMP_EXPORT kmp_int32 __kmpc_in_parallel(ident_t *loc);
4428 KMP_EXPORT void __kmpc_pop_num_threads(ident_t *loc, kmp_int32 global_tid);
4429 KMP_EXPORT void __kmpc_push_num_threads(ident_t *loc, kmp_int32 global_tid,
4430  kmp_int32 num_threads);
4431 KMP_EXPORT void __kmpc_push_num_threads_strict(ident_t *loc,
4432  kmp_int32 global_tid,
4433  kmp_int32 num_threads,
4434  int severity,
4435  const char *message);
4436 
4437 KMP_EXPORT void __kmpc_push_num_threads_list(ident_t *loc, kmp_int32 global_tid,
4438  kmp_uint32 list_length,
4439  kmp_int32 *num_threads_list);
4440 KMP_EXPORT void __kmpc_push_num_threads_list_strict(
4441  ident_t *loc, kmp_int32 global_tid, kmp_uint32 list_length,
4442  kmp_int32 *num_threads_list, int severity, const char *message);
4443 
4444 KMP_EXPORT void __kmpc_push_proc_bind(ident_t *loc, kmp_int32 global_tid,
4445  int proc_bind);
4446 KMP_EXPORT void __kmpc_push_num_teams(ident_t *loc, kmp_int32 global_tid,
4447  kmp_int32 num_teams,
4448  kmp_int32 num_threads);
4449 KMP_EXPORT void __kmpc_set_thread_limit(ident_t *loc, kmp_int32 global_tid,
4450  kmp_int32 thread_limit);
4451 /* Function for OpenMP 5.1 num_teams clause */
4452 KMP_EXPORT void __kmpc_push_num_teams_51(ident_t *loc, kmp_int32 global_tid,
4453  kmp_int32 num_teams_lb,
4454  kmp_int32 num_teams_ub,
4455  kmp_int32 num_threads);
4456 KMP_EXPORT void __kmpc_fork_teams(ident_t *loc, kmp_int32 argc,
4457  kmpc_micro microtask, ...);
4458 struct kmp_dim { // loop bounds info casted to kmp_int64
4459  kmp_int64 lo; // lower
4460  kmp_int64 up; // upper
4461  kmp_int64 st; // stride
4462 };
4463 KMP_EXPORT void __kmpc_doacross_init(ident_t *loc, kmp_int32 gtid,
4464  kmp_int32 num_dims,
4465  const struct kmp_dim *dims);
4466 KMP_EXPORT void __kmpc_doacross_wait(ident_t *loc, kmp_int32 gtid,
4467  const kmp_int64 *vec);
4468 KMP_EXPORT void __kmpc_doacross_post(ident_t *loc, kmp_int32 gtid,
4469  const kmp_int64 *vec);
4470 KMP_EXPORT void __kmpc_doacross_fini(ident_t *loc, kmp_int32 gtid);
4471 
4472 KMP_EXPORT void *__kmpc_threadprivate_cached(ident_t *loc, kmp_int32 global_tid,
4473  void *data, size_t size,
4474  void ***cache);
4475 
4476 // The routines below are not exported.
4477 // Consider making them 'static' in corresponding source files.
4478 void kmp_threadprivate_insert_private_data(int gtid, void *pc_addr,
4479  void *data_addr, size_t pc_size);
4480 struct private_common *kmp_threadprivate_insert(int gtid, void *pc_addr,
4481  void *data_addr,
4482  size_t pc_size);
4483 void __kmp_threadprivate_resize_cache(int newCapacity);
4484 void __kmp_cleanup_threadprivate_caches();
4485 
4486 // ompc_, kmpc_ entries moved from omp.h.
4487 #if KMP_OS_WINDOWS
4488 #define KMPC_CONVENTION __cdecl
4489 #else
4490 #define KMPC_CONVENTION
4491 #endif
4492 
4493 #ifndef __OMP_H
4494 typedef enum omp_sched_t {
4495  omp_sched_static = 1,
4496  omp_sched_dynamic = 2,
4497  omp_sched_guided = 3,
4498  omp_sched_auto = 4
4499 } omp_sched_t;
4500 typedef void *kmp_affinity_mask_t;
4501 #endif
4502 
4503 KMP_EXPORT void KMPC_CONVENTION ompc_set_max_active_levels(int);
4504 KMP_EXPORT void KMPC_CONVENTION ompc_set_schedule(omp_sched_t, int);
4505 KMP_EXPORT int KMPC_CONVENTION ompc_get_ancestor_thread_num(int);
4506 KMP_EXPORT int KMPC_CONVENTION ompc_get_team_size(int);
4507 KMP_EXPORT int KMPC_CONVENTION
4508 kmpc_set_affinity_mask_proc(int, kmp_affinity_mask_t *);
4509 KMP_EXPORT int KMPC_CONVENTION
4510 kmpc_unset_affinity_mask_proc(int, kmp_affinity_mask_t *);
4511 KMP_EXPORT int KMPC_CONVENTION
4512 kmpc_get_affinity_mask_proc(int, kmp_affinity_mask_t *);
4513 
4514 KMP_EXPORT void KMPC_CONVENTION kmpc_set_stacksize(int);
4515 KMP_EXPORT void KMPC_CONVENTION kmpc_set_stacksize_s(size_t);
4516 KMP_EXPORT void KMPC_CONVENTION kmpc_set_library(int);
4517 KMP_EXPORT void KMPC_CONVENTION kmpc_set_defaults(char const *);
4518 KMP_EXPORT void KMPC_CONVENTION kmpc_set_disp_num_buffers(int);
4519 void KMP_EXPAND_NAME(ompc_set_affinity_format)(char const *format);
4520 size_t KMP_EXPAND_NAME(ompc_get_affinity_format)(char *buffer, size_t size);
4521 void KMP_EXPAND_NAME(ompc_display_affinity)(char const *format);
4522 size_t KMP_EXPAND_NAME(ompc_capture_affinity)(char *buffer, size_t buf_size,
4523  char const *format);
4524 
4525 enum kmp_target_offload_kind {
4526  tgt_disabled = 0,
4527  tgt_default = 1,
4528  tgt_mandatory = 2
4529 };
4530 typedef enum kmp_target_offload_kind kmp_target_offload_kind_t;
4531 // Set via OMP_TARGET_OFFLOAD if specified, defaults to tgt_default otherwise
4532 extern kmp_target_offload_kind_t __kmp_target_offload;
4533 extern int __kmpc_get_target_offload();
4534 
4535 // Constants used in libomptarget
4536 #define KMP_DEVICE_DEFAULT -1 // This is libomptarget's default device.
4537 #define KMP_DEVICE_ALL -11 // This is libomptarget's "all devices".
4538 
4539 // OMP Pause Resource
4540 
4541 // The following enum is used both to set the status in __kmp_pause_status, and
4542 // as the internal equivalent of the externally-visible omp_pause_resource_t.
4543 typedef enum kmp_pause_status_t {
4544  kmp_not_paused = 0, // status is not paused, or, requesting resume
4545  kmp_soft_paused = 1, // status is soft-paused, or, requesting soft pause
4546  kmp_hard_paused = 2, // status is hard-paused, or, requesting hard pause
4547  kmp_stop_tool_paused = 3 // requesting stop_tool pause
4548 } kmp_pause_status_t;
4549 
4550 // This stores the pause state of the runtime
4551 extern kmp_pause_status_t __kmp_pause_status;
4552 extern int __kmpc_pause_resource(kmp_pause_status_t level);
4553 extern int __kmp_pause_resource(kmp_pause_status_t level);
4554 // Soft resume sets __kmp_pause_status, and wakes up all threads.
4555 extern void __kmp_resume_if_soft_paused();
4556 // Hard resume simply resets the status to not paused. Library will appear to
4557 // be uninitialized after hard pause. Let OMP constructs trigger required
4558 // initializations.
4559 static inline void __kmp_resume_if_hard_paused() {
4560  if (__kmp_pause_status == kmp_hard_paused) {
4561  __kmp_pause_status = kmp_not_paused;
4562  }
4563 }
4564 
4565 extern void __kmp_omp_display_env(int verbose);
4566 
4567 // 1: it is initializing hidden helper team
4568 extern volatile int __kmp_init_hidden_helper;
4569 // 1: the hidden helper team is done
4570 extern volatile int __kmp_hidden_helper_team_done;
4571 // 1: enable hidden helper task
4572 extern kmp_int32 __kmp_enable_hidden_helper;
4573 // Main thread of hidden helper team
4574 extern kmp_info_t *__kmp_hidden_helper_main_thread;
4575 // Descriptors for the hidden helper threads
4576 extern kmp_info_t **__kmp_hidden_helper_threads;
4577 // Number of hidden helper threads
4578 extern kmp_int32 __kmp_hidden_helper_threads_num;
4579 // Number of hidden helper tasks that have not been executed yet
4580 extern std::atomic<kmp_int32> __kmp_unexecuted_hidden_helper_tasks;
4581 
4582 extern void __kmp_hidden_helper_initialize();
4583 extern void __kmp_hidden_helper_threads_initz_routine();
4584 extern void __kmp_do_initialize_hidden_helper_threads();
4585 extern void __kmp_hidden_helper_threads_initz_wait();
4586 extern void __kmp_hidden_helper_initz_release();
4587 extern void __kmp_hidden_helper_threads_deinitz_wait();
4588 extern void __kmp_hidden_helper_threads_deinitz_release();
4589 extern void __kmp_hidden_helper_main_thread_wait();
4590 extern void __kmp_hidden_helper_worker_thread_wait();
4591 extern void __kmp_hidden_helper_worker_thread_signal();
4592 extern void __kmp_hidden_helper_main_thread_release();
4593 
4594 // Check whether a given thread is a hidden helper thread
4595 #define KMP_HIDDEN_HELPER_THREAD(gtid) \
4596  ((gtid) >= 1 && (gtid) <= __kmp_hidden_helper_threads_num)
4597 
4598 #define KMP_HIDDEN_HELPER_WORKER_THREAD(gtid) \
4599  ((gtid) > 1 && (gtid) <= __kmp_hidden_helper_threads_num)
4600 
4601 #define KMP_HIDDEN_HELPER_MAIN_THREAD(gtid) \
4602  ((gtid) == 1 && (gtid) <= __kmp_hidden_helper_threads_num)
4603 
4604 #define KMP_HIDDEN_HELPER_TEAM(team) \
4605  (team->t.t_threads[0] == __kmp_hidden_helper_main_thread)
4606 
4607 // Map a gtid to a hidden helper thread. The first hidden helper thread, a.k.a
4608 // main thread, is skipped.
4609 #define KMP_GTID_TO_SHADOW_GTID(gtid) \
4610  ((gtid) % (__kmp_hidden_helper_threads_num - 1) + 2)
4611 
4612 // Return the adjusted gtid value by subtracting from gtid the number
4613 // of hidden helper threads. This adjusted value is the gtid the thread would
4614 // have received if there were no hidden helper threads.
4615 static inline int __kmp_adjust_gtid_for_hidden_helpers(int gtid) {
4616  int adjusted_gtid = gtid;
4617  if (__kmp_hidden_helper_threads_num > 0 && gtid > 0 &&
4618  gtid - __kmp_hidden_helper_threads_num >= 0) {
4619  adjusted_gtid -= __kmp_hidden_helper_threads_num;
4620  }
4621  return adjusted_gtid;
4622 }
4623 
4624 // Support for error directive
4625 typedef enum kmp_severity_t {
4626  severity_warning = 1,
4627  severity_fatal = 2
4628 } kmp_severity_t;
4629 extern void __kmpc_error(ident_t *loc, int severity, const char *message);
4630 
4631 // Support for scope directive
4632 KMP_EXPORT void __kmpc_scope(ident_t *loc, kmp_int32 gtid, void *reserved);
4633 KMP_EXPORT void __kmpc_end_scope(ident_t *loc, kmp_int32 gtid, void *reserved);
4634 
4635 #ifdef __cplusplus
4636 }
4637 #endif
4638 
4639 template <bool C, bool S>
4640 extern void __kmp_suspend_32(int th_gtid, kmp_flag_32<C, S> *flag);
4641 template <bool C, bool S>
4642 extern void __kmp_suspend_64(int th_gtid, kmp_flag_64<C, S> *flag);
4643 template <bool C, bool S>
4644 extern void __kmp_atomic_suspend_64(int th_gtid,
4645  kmp_atomic_flag_64<C, S> *flag);
4646 extern void __kmp_suspend_oncore(int th_gtid, kmp_flag_oncore *flag);
4647 #if KMP_HAVE_MWAIT || KMP_HAVE_UMWAIT
4648 template <bool C, bool S>
4649 extern void __kmp_mwait_32(int th_gtid, kmp_flag_32<C, S> *flag);
4650 template <bool C, bool S>
4651 extern void __kmp_mwait_64(int th_gtid, kmp_flag_64<C, S> *flag);
4652 template <bool C, bool S>
4653 extern void __kmp_atomic_mwait_64(int th_gtid, kmp_atomic_flag_64<C, S> *flag);
4654 extern void __kmp_mwait_oncore(int th_gtid, kmp_flag_oncore *flag);
4655 #endif
4656 template <bool C, bool S>
4657 extern void __kmp_resume_32(int target_gtid, kmp_flag_32<C, S> *flag);
4658 template <bool C, bool S>
4659 extern void __kmp_resume_64(int target_gtid, kmp_flag_64<C, S> *flag);
4660 template <bool C, bool S>
4661 extern void __kmp_atomic_resume_64(int target_gtid,
4662  kmp_atomic_flag_64<C, S> *flag);
4663 extern void __kmp_resume_oncore(int target_gtid, kmp_flag_oncore *flag);
4664 
4665 template <bool C, bool S>
4666 int __kmp_execute_tasks_32(kmp_info_t *thread, kmp_int32 gtid,
4667  kmp_flag_32<C, S> *flag, int final_spin,
4668  int *thread_finished,
4669 #if USE_ITT_BUILD
4670  void *itt_sync_obj,
4671 #endif /* USE_ITT_BUILD */
4672  kmp_int32 is_constrained);
4673 template <bool C, bool S>
4674 int __kmp_execute_tasks_64(kmp_info_t *thread, kmp_int32 gtid,
4675  kmp_flag_64<C, S> *flag, int final_spin,
4676  int *thread_finished,
4677 #if USE_ITT_BUILD
4678  void *itt_sync_obj,
4679 #endif /* USE_ITT_BUILD */
4680  kmp_int32 is_constrained);
4681 template <bool C, bool S>
4682 int __kmp_atomic_execute_tasks_64(kmp_info_t *thread, kmp_int32 gtid,
4683  kmp_atomic_flag_64<C, S> *flag,
4684  int final_spin, int *thread_finished,
4685 #if USE_ITT_BUILD
4686  void *itt_sync_obj,
4687 #endif /* USE_ITT_BUILD */
4688  kmp_int32 is_constrained);
4689 int __kmp_execute_tasks_oncore(kmp_info_t *thread, kmp_int32 gtid,
4690  kmp_flag_oncore *flag, int final_spin,
4691  int *thread_finished,
4692 #if USE_ITT_BUILD
4693  void *itt_sync_obj,
4694 #endif /* USE_ITT_BUILD */
4695  kmp_int32 is_constrained);
4696 
4697 extern int __kmp_nesting_mode;
4698 extern int __kmp_nesting_mode_nlevels;
4699 extern int *__kmp_nesting_nth_level;
4700 extern void __kmp_init_nesting_mode();
4701 extern void __kmp_set_nesting_mode_threads();
4702 
4710  FILE *f;
4711 
4712  void close() {
4713  if (f && f != stdout && f != stderr) {
4714  fclose(f);
4715  f = nullptr;
4716  }
4717  }
4718 
4719 public:
4720  kmp_safe_raii_file_t() : f(nullptr) {}
4721  kmp_safe_raii_file_t(const char *filename, const char *mode,
4722  const char *env_var = nullptr)
4723  : f(nullptr) {
4724  open(filename, mode, env_var);
4725  }
4726  kmp_safe_raii_file_t(const kmp_safe_raii_file_t &other) = delete;
4727  kmp_safe_raii_file_t &operator=(const kmp_safe_raii_file_t &other) = delete;
4728  ~kmp_safe_raii_file_t() { close(); }
4729 
4733  void open(const char *filename, const char *mode,
4734  const char *env_var = nullptr) {
4735  KMP_ASSERT(!f);
4736  f = fopen(filename, mode);
4737  if (!f) {
4738  int code = errno;
4739  if (env_var) {
4740  __kmp_fatal(KMP_MSG(CantOpenFileForReading, filename), KMP_ERR(code),
4741  KMP_HNT(CheckEnvVar, env_var, filename), __kmp_msg_null);
4742  } else {
4743  __kmp_fatal(KMP_MSG(CantOpenFileForReading, filename), KMP_ERR(code),
4744  __kmp_msg_null);
4745  }
4746  }
4747  }
4750  int try_open(const char *filename, const char *mode) {
4751  KMP_ASSERT(!f);
4752  f = fopen(filename, mode);
4753  if (!f)
4754  return errno;
4755  return 0;
4756  }
4759  void set_stdout() {
4760  KMP_ASSERT(!f);
4761  f = stdout;
4762  }
4765  void set_stderr() {
4766  KMP_ASSERT(!f);
4767  f = stderr;
4768  }
4769  operator bool() { return bool(f); }
4770  operator FILE *() { return f; }
4771 };
4772 
4773 template <typename SourceType, typename TargetType,
4774  bool isSourceSmaller = (sizeof(SourceType) < sizeof(TargetType)),
4775  bool isSourceEqual = (sizeof(SourceType) == sizeof(TargetType)),
4776  bool isSourceSigned = std::is_signed<SourceType>::value,
4777  bool isTargetSigned = std::is_signed<TargetType>::value>
4778 struct kmp_convert {};
4779 
4780 // Both types are signed; Source smaller
4781 template <typename SourceType, typename TargetType>
4782 struct kmp_convert<SourceType, TargetType, true, false, true, true> {
4783  static TargetType to(SourceType src) { return (TargetType)src; }
4784 };
4785 // Source equal
4786 template <typename SourceType, typename TargetType>
4787 struct kmp_convert<SourceType, TargetType, false, true, true, true> {
4788  static TargetType to(SourceType src) { return src; }
4789 };
4790 // Source bigger
4791 template <typename SourceType, typename TargetType>
4792 struct kmp_convert<SourceType, TargetType, false, false, true, true> {
4793  static TargetType to(SourceType src) {
4794  KMP_ASSERT(src <= static_cast<SourceType>(
4795  (std::numeric_limits<TargetType>::max)()));
4796  KMP_ASSERT(src >= static_cast<SourceType>(
4797  (std::numeric_limits<TargetType>::min)()));
4798  return (TargetType)src;
4799  }
4800 };
4801 
4802 // Source signed, Target unsigned
4803 // Source smaller
4804 template <typename SourceType, typename TargetType>
4805 struct kmp_convert<SourceType, TargetType, true, false, true, false> {
4806  static TargetType to(SourceType src) {
4807  KMP_ASSERT(src >= 0);
4808  return (TargetType)src;
4809  }
4810 };
4811 // Source equal
4812 template <typename SourceType, typename TargetType>
4813 struct kmp_convert<SourceType, TargetType, false, true, true, false> {
4814  static TargetType to(SourceType src) {
4815  KMP_ASSERT(src >= 0);
4816  return (TargetType)src;
4817  }
4818 };
4819 // Source bigger
4820 template <typename SourceType, typename TargetType>
4821 struct kmp_convert<SourceType, TargetType, false, false, true, false> {
4822  static TargetType to(SourceType src) {
4823  KMP_ASSERT(src >= 0);
4824  KMP_ASSERT(src <= static_cast<SourceType>(
4825  (std::numeric_limits<TargetType>::max)()));
4826  return (TargetType)src;
4827  }
4828 };
4829 
4830 // Source unsigned, Target signed
4831 // Source smaller
4832 template <typename SourceType, typename TargetType>
4833 struct kmp_convert<SourceType, TargetType, true, false, false, true> {
4834  static TargetType to(SourceType src) { return (TargetType)src; }
4835 };
4836 // Source equal
4837 template <typename SourceType, typename TargetType>
4838 struct kmp_convert<SourceType, TargetType, false, true, false, true> {
4839  static TargetType to(SourceType src) {
4840  KMP_ASSERT(src <= static_cast<SourceType>(
4841  (std::numeric_limits<TargetType>::max)()));
4842  return (TargetType)src;
4843  }
4844 };
4845 // Source bigger
4846 template <typename SourceType, typename TargetType>
4847 struct kmp_convert<SourceType, TargetType, false, false, false, true> {
4848  static TargetType to(SourceType src) {
4849  KMP_ASSERT(src <= static_cast<SourceType>(
4850  (std::numeric_limits<TargetType>::max)()));
4851  return (TargetType)src;
4852  }
4853 };
4854 
4855 // Source unsigned, Target unsigned
4856 // Source smaller
4857 template <typename SourceType, typename TargetType>
4858 struct kmp_convert<SourceType, TargetType, true, false, false, false> {
4859  static TargetType to(SourceType src) { return (TargetType)src; }
4860 };
4861 // Source equal
4862 template <typename SourceType, typename TargetType>
4863 struct kmp_convert<SourceType, TargetType, false, true, false, false> {
4864  static TargetType to(SourceType src) { return src; }
4865 };
4866 // Source bigger
4867 template <typename SourceType, typename TargetType>
4868 struct kmp_convert<SourceType, TargetType, false, false, false, false> {
4869  static TargetType to(SourceType src) {
4870  KMP_ASSERT(src <= static_cast<SourceType>(
4871  (std::numeric_limits<TargetType>::max)()));
4872  return (TargetType)src;
4873  }
4874 };
4875 
4876 template <typename T1, typename T2>
4877 static inline void __kmp_type_convert(T1 src, T2 *dest) {
4878  *dest = kmp_convert<T1, T2>::to(src);
4879 }
4880 
4881 #endif /* KMP_H */
void set_stdout()
Definition: kmp.h:4759
void set_stderr()
Definition: kmp.h:4765
int try_open(const char *filename, const char *mode)
Definition: kmp.h:4750
void open(const char *filename, const char *mode, const char *env_var=nullptr)
Definition: kmp.h:4733
struct ident ident_t
@ KMP_IDENT_KMPC
Definition: kmp.h:189
@ KMP_IDENT_IMB
Definition: kmp.h:187
@ KMP_IDENT_WORK_LOOP
Definition: kmp.h:207
@ KMP_IDENT_BARRIER_IMPL
Definition: kmp.h:198
@ KMP_IDENT_WORK_SECTIONS
Definition: kmp.h:209
@ KMP_IDENT_AUTOPAR
Definition: kmp.h:192
@ KMP_IDENT_ATOMIC_HINT_MASK
Definition: kmp.h:216
@ KMP_IDENT_WORK_DISTRIBUTE
Definition: kmp.h:211
@ KMP_IDENT_BARRIER_EXPL
Definition: kmp.h:196
@ KMP_IDENT_ATOMIC_REDUCE
Definition: kmp.h:194
KMP_EXPORT kmp_int32 __kmpc_ok_to_fork(ident_t *)
KMP_EXPORT void __kmpc_fork_teams(ident_t *loc, kmp_int32 argc, kmpc_micro microtask,...)
KMP_EXPORT void __kmpc_fork_call_if(ident_t *loc, kmp_int32 nargs, kmpc_micro microtask, kmp_int32 cond, void *args)
KMP_EXPORT void __kmpc_push_num_threads(ident_t *loc, kmp_int32 global_tid, kmp_int32 num_threads)
KMP_EXPORT void __kmpc_set_thread_limit(ident_t *loc, kmp_int32 global_tid, kmp_int32 thread_limit)
KMP_EXPORT void __kmpc_serialized_parallel(ident_t *, kmp_int32 global_tid)
KMP_EXPORT void __kmpc_push_num_threads_list(ident_t *loc, kmp_int32 global_tid, kmp_uint32 list_length, kmp_int32 *num_threads_list)
KMP_EXPORT void __kmpc_push_num_teams(ident_t *loc, kmp_int32 global_tid, kmp_int32 num_teams, kmp_int32 num_threads)
KMP_EXPORT void __kmpc_fork_call(ident_t *, kmp_int32 nargs, kmpc_micro microtask,...)
KMP_EXPORT void __kmpc_end_serialized_parallel(ident_t *, kmp_int32 global_tid)
void(* kmpc_micro)(kmp_int32 *global_tid, kmp_int32 *bound_tid,...)
Definition: kmp.h:1773
KMP_EXPORT void __kmpc_push_num_teams_51(ident_t *loc, kmp_int32 global_tid, kmp_int32 num_teams_lb, kmp_int32 num_teams_ub, kmp_int32 num_threads)
KMP_EXPORT void __kmpc_begin(ident_t *, kmp_int32 flags)
KMP_EXPORT void __kmpc_end(ident_t *)
KMP_EXPORT void __kmpc_end_reduce(ident_t *loc, kmp_int32 global_tid, kmp_critical_name *lck)
KMP_EXPORT void __kmpc_end_barrier_master(ident_t *, kmp_int32 global_tid)
KMP_EXPORT kmp_int32 __kmpc_barrier_master_nowait(ident_t *, kmp_int32 global_tid)
KMP_EXPORT void __kmpc_end_reduce_nowait(ident_t *loc, kmp_int32 global_tid, kmp_critical_name *lck)
KMP_EXPORT kmp_int32 __kmpc_reduce(ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size, void *reduce_data, void(*reduce_func)(void *lhs_data, void *rhs_data), kmp_critical_name *lck)
KMP_EXPORT void __kmpc_barrier(ident_t *, kmp_int32 global_tid)
KMP_EXPORT void __kmpc_flush(ident_t *)
KMP_EXPORT kmp_int32 __kmpc_barrier_master(ident_t *, kmp_int32 global_tid)
KMP_EXPORT kmp_int32 __kmpc_reduce_nowait(ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size, void *reduce_data, void(*reduce_func)(void *lhs_data, void *rhs_data), kmp_critical_name *lck)
void __kmpc_taskloop(ident_t *loc, int gtid, kmp_task_t *task, int if_val, kmp_uint64 *lb, kmp_uint64 *ub, kmp_int64 st, int nogroup, int sched, kmp_uint64 grainsize, void *task_dup)
KMP_EXPORT void * __kmpc_taskred_modifier_init(ident_t *loc, int gtid, int is_ws, int num, void *data)
KMP_EXPORT void * __kmpc_taskred_init(int gtid, int num_data, void *data)
KMP_EXPORT void * __kmpc_task_reduction_init(int gtid, int num_data, void *data)
KMP_EXPORT bool __kmpc_omp_has_task_team(kmp_int32 gtid)
KMP_EXPORT void __kmpc_proxy_task_completed_ooo(kmp_task_t *ptask)
KMP_EXPORT void __kmpc_task_reduction_modifier_fini(ident_t *loc, int gtid, int is_ws)
KMP_EXPORT kmp_int32 __kmpc_omp_task_with_deps(ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 ndeps, kmp_depend_info_t *dep_list, kmp_int32 ndeps_noalias, kmp_depend_info_t *noalias_dep_list)
KMP_EXPORT kmp_int32 __kmpc_omp_reg_task_with_affinity(ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 naffins, kmp_task_affinity_info_t *affin_list)
KMP_EXPORT void * __kmpc_task_reduction_modifier_init(ident_t *loc, int gtid, int is_ws, int num, void *data)
void __kmpc_taskloop_5(ident_t *loc, int gtid, kmp_task_t *task, int if_val, kmp_uint64 *lb, kmp_uint64 *ub, kmp_int64 st, int nogroup, int sched, kmp_uint64 grainsize, int modifier, void *task_dup)
KMP_EXPORT void __kmpc_proxy_task_completed(kmp_int32 gtid, kmp_task_t *ptask)
KMP_EXPORT void __kmpc_omp_wait_deps(ident_t *loc_ref, kmp_int32 gtid, kmp_int32 ndeps, kmp_depend_info_t *dep_list, kmp_int32 ndeps_noalias, kmp_depend_info_t *noalias_dep_list)
KMP_EXPORT void * __kmpc_task_reduction_get_th_data(int gtid, void *tg, void *d)
KMP_EXPORT void ** __kmpc_omp_get_target_async_handle_ptr(kmp_int32 gtid)
void(* kmpc_dtor)(void *)
Definition: kmp.h:1797
KMP_EXPORT void __kmpc_threadprivate_register(ident_t *, void *data, kmpc_ctor ctor, kmpc_cctor cctor, kmpc_dtor dtor)
KMP_EXPORT void __kmpc_copyprivate(ident_t *loc, kmp_int32 global_tid, size_t cpy_size, void *cpy_data, void(*cpy_func)(void *, void *), kmp_int32 didit)
void *(* kmpc_ctor)(void *)
Definition: kmp.h:1791
void *(* kmpc_ctor_vec)(void *, size_t)
Definition: kmp.h:1814
void *(* kmpc_cctor)(void *, void *)
Definition: kmp.h:1804
KMP_EXPORT void * __kmpc_threadprivate_cached(ident_t *loc, kmp_int32 global_tid, void *data, size_t size, void ***cache)
void *(* kmpc_cctor_vec)(void *, void *, size_t)
Definition: kmp.h:1826
void(* kmpc_dtor_vec)(void *, size_t)
Definition: kmp.h:1820
KMP_EXPORT void __kmpc_threadprivate_register_vec(ident_t *, void *data, kmpc_ctor_vec ctor, kmpc_cctor_vec cctor, kmpc_dtor_vec dtor, size_t vector_length)
KMP_EXPORT void * __kmpc_copyprivate_light(ident_t *loc, kmp_int32 gtid, void *cpy_data)
KMP_EXPORT kmp_int32 __kmpc_global_num_threads(ident_t *)
KMP_EXPORT kmp_int32 __kmpc_global_thread_num(ident_t *)
KMP_EXPORT kmp_int32 __kmpc_in_parallel(ident_t *loc)
KMP_EXPORT kmp_int32 __kmpc_bound_thread_num(ident_t *)
KMP_EXPORT kmp_int32 __kmpc_bound_num_threads(ident_t *)
KMP_EXPORT void __kmpc_end_ordered(ident_t *, kmp_int32 global_tid)
KMP_EXPORT void __kmpc_end_critical(ident_t *, kmp_int32 global_tid, kmp_critical_name *)
KMP_EXPORT void __kmpc_for_static_fini(ident_t *loc, kmp_int32 global_tid)
int __kmpc_dispatch_next_4(ident_t *loc, kmp_int32 gtid, kmp_int32 *p_last, kmp_int32 *p_lb, kmp_int32 *p_ub, kmp_int32 *p_st)
sched_type
Definition: kmp.h:350
KMP_EXPORT void __kmpc_end_masked(ident_t *, kmp_int32 global_tid)
void __kmpc_dispatch_fini_4(ident_t *loc, kmp_int32 gtid)
void __kmpc_dispatch_deinit(ident_t *loc, kmp_int32 gtid)
KMP_EXPORT kmp_int32 __kmpc_master(ident_t *, kmp_int32 global_tid)
KMP_EXPORT void __kmpc_critical_with_hint(ident_t *, kmp_int32 global_tid, kmp_critical_name *, uint32_t hint)
KMP_EXPORT kmp_int32 __kmpc_single(ident_t *, kmp_int32 global_tid)
KMP_EXPORT kmp_int32 __kmpc_next_section(ident_t *loc, kmp_int32 global_tid, kmp_int32 numberOfSections)
void __kmpc_doacross_init(ident_t *loc, int gtid, int num_dims, const struct kmp_dim *dims)
int __kmpc_dispatch_next_4u(ident_t *loc, kmp_int32 gtid, kmp_int32 *p_last, kmp_uint32 *p_lb, kmp_uint32 *p_ub, kmp_int32 *p_st)
KMP_EXPORT void __kmpc_end_master(ident_t *, kmp_int32 global_tid)
KMP_EXPORT void __kmpc_end_sections(ident_t *loc, kmp_int32 global_tid)
KMP_EXPORT void __kmpc_end_single(ident_t *, kmp_int32 global_tid)
int __kmpc_dispatch_next_8u(ident_t *loc, kmp_int32 gtid, kmp_int32 *p_last, kmp_uint64 *p_lb, kmp_uint64 *p_ub, kmp_int64 *p_st)
void __kmpc_dispatch_fini_8(ident_t *loc, kmp_int32 gtid)
KMP_EXPORT kmp_int32 __kmpc_sections_init(ident_t *loc, kmp_int32 global_tid)
int __kmpc_dispatch_next_8(ident_t *loc, kmp_int32 gtid, kmp_int32 *p_last, kmp_int64 *p_lb, kmp_int64 *p_ub, kmp_int64 *p_st)
void __kmpc_dispatch_fini_8u(ident_t *loc, kmp_int32 gtid)
KMP_EXPORT void __kmpc_ordered(ident_t *, kmp_int32 global_tid)
KMP_EXPORT kmp_int32 __kmpc_masked(ident_t *, kmp_int32 global_tid, kmp_int32 filter)
void __kmpc_dispatch_init_4(ident_t *loc, kmp_int32 gtid, enum sched_type schedule, kmp_int32 lb, kmp_int32 ub, kmp_int32 st, kmp_int32 chunk)
void __kmpc_dispatch_init_4u(ident_t *loc, kmp_int32 gtid, enum sched_type schedule, kmp_uint32 lb, kmp_uint32 ub, kmp_int32 st, kmp_int32 chunk)
void __kmpc_dispatch_init_8u(ident_t *loc, kmp_int32 gtid, enum sched_type schedule, kmp_uint64 lb, kmp_uint64 ub, kmp_int64 st, kmp_int64 chunk)
void __kmpc_dispatch_fini_4u(ident_t *loc, kmp_int32 gtid)
void __kmpc_dispatch_init_8(ident_t *loc, kmp_int32 gtid, enum sched_type schedule, kmp_int64 lb, kmp_int64 ub, kmp_int64 st, kmp_int64 chunk)
KMP_EXPORT void __kmpc_critical(ident_t *, kmp_int32 global_tid, kmp_critical_name *)
@ kmp_nm_guided_chunked
Definition: kmp.h:401
@ kmp_sch_runtime_simd
Definition: kmp.h:372
@ kmp_nm_ord_auto
Definition: kmp.h:420
@ kmp_sch_auto
Definition: kmp.h:357
@ kmp_nm_auto
Definition: kmp.h:403
@ kmp_distribute_static_chunked
Definition: kmp.h:388
@ kmp_sch_static
Definition: kmp.h:353
@ kmp_sch_guided_simd
Definition: kmp.h:371
@ kmp_sch_modifier_monotonic
Definition: kmp.h:438
@ kmp_sch_default
Definition: kmp.h:458
@ kmp_sch_modifier_nonmonotonic
Definition: kmp.h:440
@ kmp_nm_ord_static
Definition: kmp.h:416
@ kmp_distribute_static
Definition: kmp.h:389
@ kmp_sch_guided_chunked
Definition: kmp.h:355
@ kmp_nm_static
Definition: kmp.h:399
@ kmp_sch_lower
Definition: kmp.h:351
@ kmp_nm_upper
Definition: kmp.h:422
@ kmp_ord_lower
Definition: kmp.h:377
@ kmp_ord_static
Definition: kmp.h:379
@ kmp_sch_upper
Definition: kmp.h:375
@ kmp_ord_upper
Definition: kmp.h:385
@ kmp_nm_lower
Definition: kmp.h:395
@ kmp_ord_auto
Definition: kmp.h:383
Definition: kmp.h:227
kmp_int32 reserved_1
Definition: kmp.h:228
char const * psource
Definition: kmp.h:237
kmp_int32 reserved_2
Definition: kmp.h:231
kmp_int32 reserved_3
Definition: kmp.h:236
kmp_int32 flags
Definition: kmp.h:229
Memory allocator information is shared with offload runtime.
Definition: kmp.h:1113
Memory space informaition is shared with offload runtime.
Definition: kmp.h:1105