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 
31 #define TASK_CURRENT_NOT_QUEUED 0
32 #define TASK_CURRENT_QUEUED 1
33 
34 #ifdef BUILD_TIED_TASK_STACK
35 #define TASK_STACK_EMPTY 0 // entries when the stack is empty
36 #define TASK_STACK_BLOCK_BITS 5 // Used in TASK_STACK_SIZE and TASK_STACK_MASK
37 // Number of entries in each task stack array
38 #define TASK_STACK_BLOCK_SIZE (1 << TASK_STACK_BLOCK_BITS)
39 // Mask for determining index into stack block
40 #define TASK_STACK_INDEX_MASK (TASK_STACK_BLOCK_SIZE - 1)
41 #endif // BUILD_TIED_TASK_STACK
42 
43 #define TASK_NOT_PUSHED 1
44 #define TASK_SUCCESSFULLY_PUSHED 0
45 #define TASK_TIED 1
46 #define TASK_UNTIED 0
47 #define TASK_EXPLICIT 1
48 #define TASK_IMPLICIT 0
49 #define TASK_PROXY 1
50 #define TASK_FULL 0
51 #define TASK_DETACHABLE 1
52 #define TASK_UNDETACHABLE 0
53 
54 #define KMP_CANCEL_THREADS
55 #define KMP_THREAD_ATTR
56 
57 // Android does not have pthread_cancel. Undefine KMP_CANCEL_THREADS if being
58 // built on Android
59 #if defined(__ANDROID__)
60 #undef KMP_CANCEL_THREADS
61 #endif
62 
63 #include <signal.h>
64 #include <stdarg.h>
65 #include <stddef.h>
66 #include <stdio.h>
67 #include <stdlib.h>
68 #include <string.h>
69 #include <limits>
70 #include <type_traits>
71 /* include <ctype.h> don't use; problems with /MD on Windows* OS NT due to bad
72  Microsoft library. Some macros provided below to replace these functions */
73 #ifndef __ABSOFT_WIN
74 #include <sys/types.h>
75 #endif
76 #include <limits.h>
77 #include <time.h>
78 
79 #include <errno.h>
80 
81 #include "kmp_os.h"
82 
83 #include "kmp_safe_c_api.h"
84 
85 #if KMP_STATS_ENABLED
86 class kmp_stats_list;
87 #endif
88 
89 #if KMP_USE_HIER_SCHED
90 // Only include hierarchical scheduling if affinity is supported
91 #undef KMP_USE_HIER_SCHED
92 #define KMP_USE_HIER_SCHED KMP_AFFINITY_SUPPORTED
93 #endif
94 
95 #if KMP_USE_HWLOC && KMP_AFFINITY_SUPPORTED
96 #include "hwloc.h"
97 #ifndef HWLOC_OBJ_NUMANODE
98 #define HWLOC_OBJ_NUMANODE HWLOC_OBJ_NODE
99 #endif
100 #ifndef HWLOC_OBJ_PACKAGE
101 #define HWLOC_OBJ_PACKAGE HWLOC_OBJ_SOCKET
102 #endif
103 #if HWLOC_API_VERSION >= 0x00020000
104 // hwloc 2.0 changed type of depth of object from unsigned to int
105 typedef int kmp_hwloc_depth_t;
106 #else
107 typedef unsigned int kmp_hwloc_depth_t;
108 #endif
109 #endif
110 
111 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
112 #include <xmmintrin.h>
113 #endif
114 
115 #include "kmp_debug.h"
116 #include "kmp_lock.h"
117 #include "kmp_version.h"
118 #include "kmp_barrier.h"
119 #if USE_DEBUGGER
120 #include "kmp_debugger.h"
121 #endif
122 #include "kmp_i18n.h"
123 
124 #define KMP_HANDLE_SIGNALS (KMP_OS_UNIX || KMP_OS_WINDOWS)
125 
126 #include "kmp_wrapper_malloc.h"
127 #if KMP_OS_UNIX
128 #include <unistd.h>
129 #if !defined NSIG && defined _NSIG
130 #define NSIG _NSIG
131 #endif
132 #endif
133 
134 #if KMP_OS_LINUX
135 #pragma weak clock_gettime
136 #endif
137 
138 #if OMPT_SUPPORT
139 #include "ompt-internal.h"
140 #endif
141 
142 #if OMPD_SUPPORT
143 #include "ompd-specific.h"
144 #endif
145 
146 #ifndef UNLIKELY
147 #define UNLIKELY(x) (x)
148 #endif
149 
150 // Affinity format function
151 #include "kmp_str.h"
152 
153 // 0 - no fast memory allocation, alignment: 8-byte on x86, 16-byte on x64.
154 // 3 - fast allocation using sync, non-sync free lists of any size, non-self
155 // free lists of limited size.
156 #ifndef USE_FAST_MEMORY
157 #define USE_FAST_MEMORY 3
158 #endif
159 
160 #ifndef KMP_NESTED_HOT_TEAMS
161 #define KMP_NESTED_HOT_TEAMS 0
162 #define USE_NESTED_HOT_ARG(x)
163 #else
164 #if KMP_NESTED_HOT_TEAMS
165 #define USE_NESTED_HOT_ARG(x) , x
166 #else
167 #define USE_NESTED_HOT_ARG(x)
168 #endif
169 #endif
170 
171 // Assume using BGET compare_exchange instruction instead of lock by default.
172 #ifndef USE_CMP_XCHG_FOR_BGET
173 #define USE_CMP_XCHG_FOR_BGET 1
174 #endif
175 
176 // Test to see if queuing lock is better than bootstrap lock for bget
177 // #ifndef USE_QUEUING_LOCK_FOR_BGET
178 // #define USE_QUEUING_LOCK_FOR_BGET
179 // #endif
180 
181 #define KMP_NSEC_PER_SEC 1000000000L
182 #define KMP_USEC_PER_SEC 1000000L
183 
192 enum {
197  /* 0x04 is no longer used */
206  KMP_IDENT_BARRIER_IMPL_MASK = 0x01C0,
207  KMP_IDENT_BARRIER_IMPL_FOR = 0x0040,
208  KMP_IDENT_BARRIER_IMPL_SECTIONS = 0x00C0,
209 
210  KMP_IDENT_BARRIER_IMPL_SINGLE = 0x0140,
211  KMP_IDENT_BARRIER_IMPL_WORKSHARE = 0x01C0,
212 
224  KMP_IDENT_ATOMIC_HINT_UNCONTENDED = 0x010000,
225  KMP_IDENT_ATOMIC_HINT_CONTENDED = 0x020000,
226  KMP_IDENT_ATOMIC_HINT_NONSPECULATIVE = 0x040000,
227  KMP_IDENT_ATOMIC_HINT_SPECULATIVE = 0x080000,
228  KMP_IDENT_OPENMP_SPEC_VERSION_MASK = 0xFF000000
229 };
230 
234 typedef struct ident {
235  kmp_int32 reserved_1;
236  kmp_int32 flags;
238  kmp_int32 reserved_2;
239 #if USE_ITT_BUILD
240 /* but currently used for storing region-specific ITT */
241 /* contextual information. */
242 #endif /* USE_ITT_BUILD */
243  kmp_int32 reserved_3;
244  char const *psource;
248  // Returns the OpenMP version in form major*10+minor (e.g., 50 for 5.0)
249  kmp_int32 get_openmp_version() {
250  return (((flags & KMP_IDENT_OPENMP_SPEC_VERSION_MASK) >> 24) & 0xFF);
251  }
257 // Some forward declarations.
258 typedef union kmp_team kmp_team_t;
259 typedef struct kmp_taskdata kmp_taskdata_t;
260 typedef union kmp_task_team kmp_task_team_t;
261 typedef union kmp_team kmp_team_p;
262 typedef union kmp_info kmp_info_p;
263 typedef union kmp_root kmp_root_p;
264 
265 template <bool C = false, bool S = true> class kmp_flag_32;
266 template <bool C = false, bool S = true> class kmp_flag_64;
267 template <bool C = false, bool S = true> class kmp_atomic_flag_64;
268 class kmp_flag_oncore;
269 
270 #ifdef __cplusplus
271 extern "C" {
272 #endif
273 
274 /* ------------------------------------------------------------------------ */
275 
276 /* Pack two 32-bit signed integers into a 64-bit signed integer */
277 /* ToDo: Fix word ordering for big-endian machines. */
278 #define KMP_PACK_64(HIGH_32, LOW_32) \
279  ((kmp_int64)((((kmp_uint64)(HIGH_32)) << 32) | (kmp_uint64)(LOW_32)))
280 
281 // Generic string manipulation macros. Assume that _x is of type char *
282 #define SKIP_WS(_x) \
283  { \
284  while (*(_x) == ' ' || *(_x) == '\t') \
285  (_x)++; \
286  }
287 #define SKIP_DIGITS(_x) \
288  { \
289  while (*(_x) >= '0' && *(_x) <= '9') \
290  (_x)++; \
291  }
292 #define SKIP_TOKEN(_x) \
293  { \
294  while ((*(_x) >= '0' && *(_x) <= '9') || (*(_x) >= 'a' && *(_x) <= 'z') || \
295  (*(_x) >= 'A' && *(_x) <= 'Z') || *(_x) == '_') \
296  (_x)++; \
297  }
298 #define SKIP_TO(_x, _c) \
299  { \
300  while (*(_x) != '\0' && *(_x) != (_c)) \
301  (_x)++; \
302  }
303 
304 /* ------------------------------------------------------------------------ */
305 
306 #define KMP_MAX(x, y) ((x) > (y) ? (x) : (y))
307 #define KMP_MIN(x, y) ((x) < (y) ? (x) : (y))
308 
309 /* ------------------------------------------------------------------------ */
310 /* Enumeration types */
311 
312 enum kmp_state_timer {
313  ts_stop,
314  ts_start,
315  ts_pause,
316 
317  ts_last_state
318 };
319 
320 enum dynamic_mode {
321  dynamic_default,
322 #ifdef USE_LOAD_BALANCE
323  dynamic_load_balance,
324 #endif /* USE_LOAD_BALANCE */
325  dynamic_random,
326  dynamic_thread_limit,
327  dynamic_max
328 };
329 
330 /* external schedule constants, duplicate enum omp_sched in omp.h in order to
331  * not include it here */
332 #ifndef KMP_SCHED_TYPE_DEFINED
333 #define KMP_SCHED_TYPE_DEFINED
334 typedef enum kmp_sched {
335  kmp_sched_lower = 0, // lower and upper bounds are for routine parameter check
336  // Note: need to adjust __kmp_sch_map global array in case enum is changed
337  kmp_sched_static = 1, // mapped to kmp_sch_static_chunked (33)
338  kmp_sched_dynamic = 2, // mapped to kmp_sch_dynamic_chunked (35)
339  kmp_sched_guided = 3, // mapped to kmp_sch_guided_chunked (36)
340  kmp_sched_auto = 4, // mapped to kmp_sch_auto (38)
341  kmp_sched_upper_std = 5, // upper bound for standard schedules
342  kmp_sched_lower_ext = 100, // lower bound of Intel extension schedules
343  kmp_sched_trapezoidal = 101, // mapped to kmp_sch_trapezoidal (39)
344 #if KMP_STATIC_STEAL_ENABLED
345  kmp_sched_static_steal = 102, // mapped to kmp_sch_static_steal (44)
346 #endif
347  kmp_sched_upper,
348  kmp_sched_default = kmp_sched_static, // default scheduling
349  kmp_sched_monotonic = 0x80000000
350 } kmp_sched_t;
351 #endif
352 
357 enum sched_type : kmp_int32 {
359  kmp_sch_static_chunked = 33,
361  kmp_sch_dynamic_chunked = 35,
363  kmp_sch_runtime = 37,
365  kmp_sch_trapezoidal = 39,
366 
367  /* accessible only through KMP_SCHEDULE environment variable */
368  kmp_sch_static_greedy = 40,
369  kmp_sch_static_balanced = 41,
370  /* accessible only through KMP_SCHEDULE environment variable */
371  kmp_sch_guided_iterative_chunked = 42,
372  kmp_sch_guided_analytical_chunked = 43,
373  /* accessible only through KMP_SCHEDULE environment variable */
374  kmp_sch_static_steal = 44,
375 
376  /* static with chunk adjustment (e.g., simd) */
377  kmp_sch_static_balanced_chunked = 45,
381  /* accessible only through KMP_SCHEDULE environment variable */
385  kmp_ord_static_chunked = 65,
387  kmp_ord_dynamic_chunked = 67,
388  kmp_ord_guided_chunked = 68,
389  kmp_ord_runtime = 69,
391  kmp_ord_trapezoidal = 71,
394  /* Schedules for Distribute construct */
398  /* For the "nomerge" versions, kmp_dispatch_next*() will always return a
399  single iteration/chunk, even if the loop is serialized. For the schedule
400  types listed above, the entire iteration vector is returned if the loop is
401  serialized. This doesn't work for gcc/gcomp sections. */
402  kmp_nm_lower = 160,
404  kmp_nm_static_chunked =
405  (kmp_sch_static_chunked - kmp_sch_lower + kmp_nm_lower),
407  kmp_nm_dynamic_chunked = 163,
409  kmp_nm_runtime = 165,
410  kmp_nm_auto = 166,
411  kmp_nm_trapezoidal = 167,
412 
413  /* accessible only through KMP_SCHEDULE environment variable */
414  kmp_nm_static_greedy = 168,
415  kmp_nm_static_balanced = 169,
416  /* accessible only through KMP_SCHEDULE environment variable */
417  kmp_nm_guided_iterative_chunked = 170,
418  kmp_nm_guided_analytical_chunked = 171,
419  kmp_nm_static_steal =
420  172, /* accessible only through OMP_SCHEDULE environment variable */
421 
422  kmp_nm_ord_static_chunked = 193,
424  kmp_nm_ord_dynamic_chunked = 195,
425  kmp_nm_ord_guided_chunked = 196,
426  kmp_nm_ord_runtime = 197,
428  kmp_nm_ord_trapezoidal = 199,
431  /* Support for OpenMP 4.5 monotonic and nonmonotonic schedule modifiers. Since
432  we need to distinguish the three possible cases (no modifier, monotonic
433  modifier, nonmonotonic modifier), we need separate bits for each modifier.
434  The absence of monotonic does not imply nonmonotonic, especially since 4.5
435  says that the behaviour of the "no modifier" case is implementation defined
436  in 4.5, but will become "nonmonotonic" in 5.0.
437 
438  Since we're passing a full 32 bit value, we can use a couple of high bits
439  for these flags; out of paranoia we avoid the sign bit.
440 
441  These modifiers can be or-ed into non-static schedules by the compiler to
442  pass the additional information. They will be stripped early in the
443  processing in __kmp_dispatch_init when setting up schedules, so most of the
444  code won't ever see schedules with these bits set. */
446  (1 << 29),
448  (1 << 30),
450 #define SCHEDULE_WITHOUT_MODIFIERS(s) \
451  (enum sched_type)( \
453 #define SCHEDULE_HAS_MONOTONIC(s) (((s)&kmp_sch_modifier_monotonic) != 0)
454 #define SCHEDULE_HAS_NONMONOTONIC(s) (((s)&kmp_sch_modifier_nonmonotonic) != 0)
455 #define SCHEDULE_HAS_NO_MODIFIERS(s) \
456  (((s) & (kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic)) == 0)
457 #define SCHEDULE_GET_MODIFIERS(s) \
458  ((enum sched_type)( \
459  (s) & (kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic)))
460 #define SCHEDULE_SET_MODIFIERS(s, m) \
461  (s = (enum sched_type)((kmp_int32)s | (kmp_int32)m))
462 #define SCHEDULE_NONMONOTONIC 0
463 #define SCHEDULE_MONOTONIC 1
464 
466 };
467 
468 // Apply modifiers on internal kind to standard kind
469 static inline void
470 __kmp_sched_apply_mods_stdkind(kmp_sched_t *kind,
471  enum sched_type internal_kind) {
472  if (SCHEDULE_HAS_MONOTONIC(internal_kind)) {
473  *kind = (kmp_sched_t)((int)*kind | (int)kmp_sched_monotonic);
474  }
475 }
476 
477 // Apply modifiers on standard kind to internal kind
478 static inline void
479 __kmp_sched_apply_mods_intkind(kmp_sched_t kind,
480  enum sched_type *internal_kind) {
481  if ((int)kind & (int)kmp_sched_monotonic) {
482  *internal_kind = (enum sched_type)((int)*internal_kind |
484  }
485 }
486 
487 // Get standard schedule without modifiers
488 static inline kmp_sched_t __kmp_sched_without_mods(kmp_sched_t kind) {
489  return (kmp_sched_t)((int)kind & ~((int)kmp_sched_monotonic));
490 }
491 
492 /* Type to keep runtime schedule set via OMP_SCHEDULE or omp_set_schedule() */
493 typedef union kmp_r_sched {
494  struct {
495  enum sched_type r_sched_type;
496  int chunk;
497  };
498  kmp_int64 sched;
499 } kmp_r_sched_t;
500 
501 extern enum sched_type __kmp_sch_map[]; // map OMP 3.0 schedule types with our
502 // internal schedule types
503 
504 enum library_type {
505  library_none,
506  library_serial,
507  library_turnaround,
508  library_throughput
509 };
510 
511 #if KMP_OS_LINUX
512 enum clock_function_type {
513  clock_function_gettimeofday,
514  clock_function_clock_gettime
515 };
516 #endif /* KMP_OS_LINUX */
517 
518 #if KMP_MIC_SUPPORTED
519 enum mic_type { non_mic, mic1, mic2, mic3, dummy };
520 #endif
521 
522 /* -- fast reduction stuff ------------------------------------------------ */
523 
524 #undef KMP_FAST_REDUCTION_BARRIER
525 #define KMP_FAST_REDUCTION_BARRIER 1
526 
527 #undef KMP_FAST_REDUCTION_CORE_DUO
528 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
529 #define KMP_FAST_REDUCTION_CORE_DUO 1
530 #endif
531 
532 enum _reduction_method {
533  reduction_method_not_defined = 0,
534  critical_reduce_block = (1 << 8),
535  atomic_reduce_block = (2 << 8),
536  tree_reduce_block = (3 << 8),
537  empty_reduce_block = (4 << 8)
538 };
539 
540 // Description of the packed_reduction_method variable:
541 // The packed_reduction_method variable consists of two enum types variables
542 // that are packed together into 0-th byte and 1-st byte:
543 // 0: (packed_reduction_method & 0x000000FF) is a 'enum barrier_type' value of
544 // barrier that will be used in fast reduction: bs_plain_barrier or
545 // bs_reduction_barrier
546 // 1: (packed_reduction_method & 0x0000FF00) is a reduction method that will
547 // be used in fast reduction;
548 // Reduction method is of 'enum _reduction_method' type and it's defined the way
549 // so that the bits of 0-th byte are empty, so no need to execute a shift
550 // instruction while packing/unpacking
551 
552 #if KMP_FAST_REDUCTION_BARRIER
553 #define PACK_REDUCTION_METHOD_AND_BARRIER(reduction_method, barrier_type) \
554  ((reduction_method) | (barrier_type))
555 
556 #define UNPACK_REDUCTION_METHOD(packed_reduction_method) \
557  ((enum _reduction_method)((packed_reduction_method) & (0x0000FF00)))
558 
559 #define UNPACK_REDUCTION_BARRIER(packed_reduction_method) \
560  ((enum barrier_type)((packed_reduction_method) & (0x000000FF)))
561 #else
562 #define PACK_REDUCTION_METHOD_AND_BARRIER(reduction_method, barrier_type) \
563  (reduction_method)
564 
565 #define UNPACK_REDUCTION_METHOD(packed_reduction_method) \
566  (packed_reduction_method)
567 
568 #define UNPACK_REDUCTION_BARRIER(packed_reduction_method) (bs_plain_barrier)
569 #endif
570 
571 #define TEST_REDUCTION_METHOD(packed_reduction_method, which_reduction_block) \
572  ((UNPACK_REDUCTION_METHOD(packed_reduction_method)) == \
573  (which_reduction_block))
574 
575 #if KMP_FAST_REDUCTION_BARRIER
576 #define TREE_REDUCE_BLOCK_WITH_REDUCTION_BARRIER \
577  (PACK_REDUCTION_METHOD_AND_BARRIER(tree_reduce_block, bs_reduction_barrier))
578 
579 #define TREE_REDUCE_BLOCK_WITH_PLAIN_BARRIER \
580  (PACK_REDUCTION_METHOD_AND_BARRIER(tree_reduce_block, bs_plain_barrier))
581 #endif
582 
583 typedef int PACKED_REDUCTION_METHOD_T;
584 
585 /* -- end of fast reduction stuff ----------------------------------------- */
586 
587 #if KMP_OS_WINDOWS
588 #define USE_CBLKDATA
589 #if KMP_MSVC_COMPAT
590 #pragma warning(push)
591 #pragma warning(disable : 271 310)
592 #endif
593 #include <windows.h>
594 #if KMP_MSVC_COMPAT
595 #pragma warning(pop)
596 #endif
597 #endif
598 
599 #if KMP_OS_UNIX
600 #include <dlfcn.h>
601 #include <pthread.h>
602 #endif
603 
604 enum kmp_hw_t : int {
605  KMP_HW_UNKNOWN = -1,
606  KMP_HW_SOCKET = 0,
607  KMP_HW_PROC_GROUP,
608  KMP_HW_NUMA,
609  KMP_HW_DIE,
610  KMP_HW_LLC,
611  KMP_HW_L3,
612  KMP_HW_TILE,
613  KMP_HW_MODULE,
614  KMP_HW_L2,
615  KMP_HW_L1,
616  KMP_HW_CORE,
617  KMP_HW_THREAD,
618  KMP_HW_LAST
619 };
620 
621 #define KMP_DEBUG_ASSERT_VALID_HW_TYPE(type) \
622  KMP_DEBUG_ASSERT(type >= (kmp_hw_t)0 && type < KMP_HW_LAST)
623 #define KMP_ASSERT_VALID_HW_TYPE(type) \
624  KMP_ASSERT(type >= (kmp_hw_t)0 && type < KMP_HW_LAST)
625 
626 #define KMP_FOREACH_HW_TYPE(type) \
627  for (kmp_hw_t type = (kmp_hw_t)0; type < KMP_HW_LAST; \
628  type = (kmp_hw_t)((int)type + 1))
629 
630 const char *__kmp_hw_get_keyword(kmp_hw_t type, bool plural = false);
631 const char *__kmp_hw_get_catalog_string(kmp_hw_t type, bool plural = false);
632 
633 /* Only Linux* OS and Windows* OS support thread affinity. */
634 #if KMP_AFFINITY_SUPPORTED
635 
636 // GROUP_AFFINITY is already defined for _MSC_VER>=1600 (VS2010 and later).
637 #if KMP_OS_WINDOWS
638 #if _MSC_VER < 1600 && KMP_MSVC_COMPAT
639 typedef struct GROUP_AFFINITY {
640  KAFFINITY Mask;
641  WORD Group;
642  WORD Reserved[3];
643 } GROUP_AFFINITY;
644 #endif /* _MSC_VER < 1600 */
645 #if KMP_GROUP_AFFINITY
646 extern int __kmp_num_proc_groups;
647 #else
648 static const int __kmp_num_proc_groups = 1;
649 #endif /* KMP_GROUP_AFFINITY */
650 typedef DWORD (*kmp_GetActiveProcessorCount_t)(WORD);
651 extern kmp_GetActiveProcessorCount_t __kmp_GetActiveProcessorCount;
652 
653 typedef WORD (*kmp_GetActiveProcessorGroupCount_t)(void);
654 extern kmp_GetActiveProcessorGroupCount_t __kmp_GetActiveProcessorGroupCount;
655 
656 typedef BOOL (*kmp_GetThreadGroupAffinity_t)(HANDLE, GROUP_AFFINITY *);
657 extern kmp_GetThreadGroupAffinity_t __kmp_GetThreadGroupAffinity;
658 
659 typedef BOOL (*kmp_SetThreadGroupAffinity_t)(HANDLE, const GROUP_AFFINITY *,
660  GROUP_AFFINITY *);
661 extern kmp_SetThreadGroupAffinity_t __kmp_SetThreadGroupAffinity;
662 #endif /* KMP_OS_WINDOWS */
663 
664 #if KMP_USE_HWLOC
665 extern hwloc_topology_t __kmp_hwloc_topology;
666 extern int __kmp_hwloc_error;
667 #endif
668 
669 extern size_t __kmp_affin_mask_size;
670 #define KMP_AFFINITY_CAPABLE() (__kmp_affin_mask_size > 0)
671 #define KMP_AFFINITY_DISABLE() (__kmp_affin_mask_size = 0)
672 #define KMP_AFFINITY_ENABLE(mask_size) (__kmp_affin_mask_size = mask_size)
673 #define KMP_CPU_SET_ITERATE(i, mask) \
674  for (i = (mask)->begin(); (int)i != (mask)->end(); i = (mask)->next(i))
675 #define KMP_CPU_SET(i, mask) (mask)->set(i)
676 #define KMP_CPU_ISSET(i, mask) (mask)->is_set(i)
677 #define KMP_CPU_CLR(i, mask) (mask)->clear(i)
678 #define KMP_CPU_ZERO(mask) (mask)->zero()
679 #define KMP_CPU_COPY(dest, src) (dest)->copy(src)
680 #define KMP_CPU_AND(dest, src) (dest)->bitwise_and(src)
681 #define KMP_CPU_COMPLEMENT(max_bit_number, mask) (mask)->bitwise_not()
682 #define KMP_CPU_UNION(dest, src) (dest)->bitwise_or(src)
683 #define KMP_CPU_ALLOC(ptr) (ptr = __kmp_affinity_dispatch->allocate_mask())
684 #define KMP_CPU_FREE(ptr) __kmp_affinity_dispatch->deallocate_mask(ptr)
685 #define KMP_CPU_ALLOC_ON_STACK(ptr) KMP_CPU_ALLOC(ptr)
686 #define KMP_CPU_FREE_FROM_STACK(ptr) KMP_CPU_FREE(ptr)
687 #define KMP_CPU_INTERNAL_ALLOC(ptr) KMP_CPU_ALLOC(ptr)
688 #define KMP_CPU_INTERNAL_FREE(ptr) KMP_CPU_FREE(ptr)
689 #define KMP_CPU_INDEX(arr, i) __kmp_affinity_dispatch->index_mask_array(arr, i)
690 #define KMP_CPU_ALLOC_ARRAY(arr, n) \
691  (arr = __kmp_affinity_dispatch->allocate_mask_array(n))
692 #define KMP_CPU_FREE_ARRAY(arr, n) \
693  __kmp_affinity_dispatch->deallocate_mask_array(arr)
694 #define KMP_CPU_INTERNAL_ALLOC_ARRAY(arr, n) KMP_CPU_ALLOC_ARRAY(arr, n)
695 #define KMP_CPU_INTERNAL_FREE_ARRAY(arr, n) KMP_CPU_FREE_ARRAY(arr, n)
696 #define __kmp_get_system_affinity(mask, abort_bool) \
697  (mask)->get_system_affinity(abort_bool)
698 #define __kmp_set_system_affinity(mask, abort_bool) \
699  (mask)->set_system_affinity(abort_bool)
700 #define __kmp_get_proc_group(mask) (mask)->get_proc_group()
701 
702 class KMPAffinity {
703 public:
704  class Mask {
705  public:
706  void *operator new(size_t n);
707  void operator delete(void *p);
708  void *operator new[](size_t n);
709  void operator delete[](void *p);
710  virtual ~Mask() {}
711  // Set bit i to 1
712  virtual void set(int i) {}
713  // Return bit i
714  virtual bool is_set(int i) const { return false; }
715  // Set bit i to 0
716  virtual void clear(int i) {}
717  // Zero out entire mask
718  virtual void zero() {}
719  // Copy src into this mask
720  virtual void copy(const Mask *src) {}
721  // this &= rhs
722  virtual void bitwise_and(const Mask *rhs) {}
723  // this |= rhs
724  virtual void bitwise_or(const Mask *rhs) {}
725  // this = ~this
726  virtual void bitwise_not() {}
727  // API for iterating over an affinity mask
728  // for (int i = mask->begin(); i != mask->end(); i = mask->next(i))
729  virtual int begin() const { return 0; }
730  virtual int end() const { return 0; }
731  virtual int next(int previous) const { return 0; }
732 #if KMP_OS_WINDOWS
733  virtual int set_process_affinity(bool abort_on_error) const { return -1; }
734 #endif
735  // Set the system's affinity to this affinity mask's value
736  virtual int set_system_affinity(bool abort_on_error) const { return -1; }
737  // Set this affinity mask to the current system affinity
738  virtual int get_system_affinity(bool abort_on_error) { return -1; }
739  // Only 1 DWORD in the mask should have any procs set.
740  // Return the appropriate index, or -1 for an invalid mask.
741  virtual int get_proc_group() const { return -1; }
742  };
743  void *operator new(size_t n);
744  void operator delete(void *p);
745  // Need virtual destructor
746  virtual ~KMPAffinity() = default;
747  // Determine if affinity is capable
748  virtual void determine_capable(const char *env_var) {}
749  // Bind the current thread to os proc
750  virtual void bind_thread(int proc) {}
751  // Factory functions to allocate/deallocate a mask
752  virtual Mask *allocate_mask() { return nullptr; }
753  virtual void deallocate_mask(Mask *m) {}
754  virtual Mask *allocate_mask_array(int num) { return nullptr; }
755  virtual void deallocate_mask_array(Mask *m) {}
756  virtual Mask *index_mask_array(Mask *m, int index) { return nullptr; }
757  static void pick_api();
758  static void destroy_api();
759  enum api_type {
760  NATIVE_OS
761 #if KMP_USE_HWLOC
762  ,
763  HWLOC
764 #endif
765  };
766  virtual api_type get_api_type() const {
767  KMP_ASSERT(0);
768  return NATIVE_OS;
769  }
770 
771 private:
772  static bool picked_api;
773 };
774 
775 typedef KMPAffinity::Mask kmp_affin_mask_t;
776 extern KMPAffinity *__kmp_affinity_dispatch;
777 
778 // Declare local char buffers with this size for printing debug and info
779 // messages, using __kmp_affinity_print_mask().
780 #define KMP_AFFIN_MASK_PRINT_LEN 1024
781 
782 enum affinity_type {
783  affinity_none = 0,
784  affinity_physical,
785  affinity_logical,
786  affinity_compact,
787  affinity_scatter,
788  affinity_explicit,
789  affinity_balanced,
790  affinity_disabled, // not used outsize the env var parser
791  affinity_default
792 };
793 
794 enum affinity_top_method {
795  affinity_top_method_all = 0, // try all (supported) methods, in order
796 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
797  affinity_top_method_apicid,
798  affinity_top_method_x2apicid,
799  affinity_top_method_x2apicid_1f,
800 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
801  affinity_top_method_cpuinfo, // KMP_CPUINFO_FILE is usable on Windows* OS, too
802 #if KMP_GROUP_AFFINITY
803  affinity_top_method_group,
804 #endif /* KMP_GROUP_AFFINITY */
805  affinity_top_method_flat,
806 #if KMP_USE_HWLOC
807  affinity_top_method_hwloc,
808 #endif
809  affinity_top_method_default
810 };
811 
812 #define affinity_respect_mask_default (-1)
813 
814 extern enum affinity_type __kmp_affinity_type; /* Affinity type */
815 extern kmp_hw_t __kmp_affinity_gran; /* Affinity granularity */
816 extern int __kmp_affinity_gran_levels; /* corresponding int value */
817 extern int __kmp_affinity_dups; /* Affinity duplicate masks */
818 extern enum affinity_top_method __kmp_affinity_top_method;
819 extern int __kmp_affinity_compact; /* Affinity 'compact' value */
820 extern int __kmp_affinity_offset; /* Affinity offset value */
821 extern int __kmp_affinity_verbose; /* Was verbose specified for KMP_AFFINITY? */
822 extern int __kmp_affinity_warnings; /* KMP_AFFINITY warnings enabled ? */
823 extern int __kmp_affinity_respect_mask; // Respect process' init affinity mask?
824 extern char *__kmp_affinity_proclist; /* proc ID list */
825 extern kmp_affin_mask_t *__kmp_affinity_masks;
826 extern unsigned __kmp_affinity_num_masks;
827 extern void __kmp_affinity_bind_thread(int which);
828 
829 extern kmp_affin_mask_t *__kmp_affin_fullMask;
830 extern char *__kmp_cpuinfo_file;
831 
832 #endif /* KMP_AFFINITY_SUPPORTED */
833 
834 // This needs to be kept in sync with the values in omp.h !!!
835 typedef enum kmp_proc_bind_t {
836  proc_bind_false = 0,
837  proc_bind_true,
838  proc_bind_primary,
839  proc_bind_close,
840  proc_bind_spread,
841  proc_bind_intel, // use KMP_AFFINITY interface
842  proc_bind_default
843 } kmp_proc_bind_t;
844 
845 typedef struct kmp_nested_proc_bind_t {
846  kmp_proc_bind_t *bind_types;
847  int size;
848  int used;
849 } kmp_nested_proc_bind_t;
850 
851 extern kmp_nested_proc_bind_t __kmp_nested_proc_bind;
852 
853 extern int __kmp_display_affinity;
854 extern char *__kmp_affinity_format;
855 static const size_t KMP_AFFINITY_FORMAT_SIZE = 512;
856 #if OMPT_SUPPORT
857 extern int __kmp_tool;
858 extern char *__kmp_tool_libraries;
859 #endif // OMPT_SUPPORT
860 
861 #if KMP_AFFINITY_SUPPORTED
862 #define KMP_PLACE_ALL (-1)
863 #define KMP_PLACE_UNDEFINED (-2)
864 // Is KMP_AFFINITY is being used instead of OMP_PROC_BIND/OMP_PLACES?
865 #define KMP_AFFINITY_NON_PROC_BIND \
866  ((__kmp_nested_proc_bind.bind_types[0] == proc_bind_false || \
867  __kmp_nested_proc_bind.bind_types[0] == proc_bind_intel) && \
868  (__kmp_affinity_num_masks > 0 || __kmp_affinity_type == affinity_balanced))
869 #endif /* KMP_AFFINITY_SUPPORTED */
870 
871 extern int __kmp_affinity_num_places;
872 
873 typedef enum kmp_cancel_kind_t {
874  cancel_noreq = 0,
875  cancel_parallel = 1,
876  cancel_loop = 2,
877  cancel_sections = 3,
878  cancel_taskgroup = 4
879 } kmp_cancel_kind_t;
880 
881 // KMP_HW_SUBSET support:
882 typedef struct kmp_hws_item {
883  int num;
884  int offset;
885 } kmp_hws_item_t;
886 
887 extern kmp_hws_item_t __kmp_hws_socket;
888 extern kmp_hws_item_t __kmp_hws_die;
889 extern kmp_hws_item_t __kmp_hws_node;
890 extern kmp_hws_item_t __kmp_hws_tile;
891 extern kmp_hws_item_t __kmp_hws_core;
892 extern kmp_hws_item_t __kmp_hws_proc;
893 extern int __kmp_hws_requested;
894 extern int __kmp_hws_abs_flag; // absolute or per-item number requested
895 
896 /* ------------------------------------------------------------------------ */
897 
898 #define KMP_PAD(type, sz) \
899  (sizeof(type) + (sz - ((sizeof(type) - 1) % (sz)) - 1))
900 
901 // We need to avoid using -1 as a GTID as +1 is added to the gtid
902 // when storing it in a lock, and the value 0 is reserved.
903 #define KMP_GTID_DNE (-2) /* Does not exist */
904 #define KMP_GTID_SHUTDOWN (-3) /* Library is shutting down */
905 #define KMP_GTID_MONITOR (-4) /* Monitor thread ID */
906 #define KMP_GTID_UNKNOWN (-5) /* Is not known */
907 #define KMP_GTID_MIN (-6) /* Minimal gtid for low bound check in DEBUG */
908 
909 /* OpenMP 5.0 Memory Management support */
910 
911 #ifndef __OMP_H
912 // Duplicate type definitions from omp.h
913 typedef uintptr_t omp_uintptr_t;
914 
915 typedef enum {
916  omp_atk_sync_hint = 1,
917  omp_atk_alignment = 2,
918  omp_atk_access = 3,
919  omp_atk_pool_size = 4,
920  omp_atk_fallback = 5,
921  omp_atk_fb_data = 6,
922  omp_atk_pinned = 7,
923  omp_atk_partition = 8
924 } omp_alloctrait_key_t;
925 
926 typedef enum {
927  omp_atv_false = 0,
928  omp_atv_true = 1,
929  omp_atv_contended = 3,
930  omp_atv_uncontended = 4,
931  omp_atv_serialized = 5,
932  omp_atv_sequential = omp_atv_serialized, // (deprecated)
933  omp_atv_private = 6,
934  omp_atv_all = 7,
935  omp_atv_thread = 8,
936  omp_atv_pteam = 9,
937  omp_atv_cgroup = 10,
938  omp_atv_default_mem_fb = 11,
939  omp_atv_null_fb = 12,
940  omp_atv_abort_fb = 13,
941  omp_atv_allocator_fb = 14,
942  omp_atv_environment = 15,
943  omp_atv_nearest = 16,
944  omp_atv_blocked = 17,
945  omp_atv_interleaved = 18
946 } omp_alloctrait_value_t;
947 #define omp_atv_default ((omp_uintptr_t)-1)
948 
949 typedef void *omp_memspace_handle_t;
950 extern omp_memspace_handle_t const omp_default_mem_space;
951 extern omp_memspace_handle_t const omp_large_cap_mem_space;
952 extern omp_memspace_handle_t const omp_const_mem_space;
953 extern omp_memspace_handle_t const omp_high_bw_mem_space;
954 extern omp_memspace_handle_t const omp_low_lat_mem_space;
955 // Preview of target memory support
956 extern omp_memspace_handle_t const llvm_omp_target_host_mem_space;
957 extern omp_memspace_handle_t const llvm_omp_target_shared_mem_space;
958 extern omp_memspace_handle_t const llvm_omp_target_device_mem_space;
959 
960 typedef struct {
961  omp_alloctrait_key_t key;
962  omp_uintptr_t value;
963 } omp_alloctrait_t;
964 
965 typedef void *omp_allocator_handle_t;
966 extern omp_allocator_handle_t const omp_null_allocator;
967 extern omp_allocator_handle_t const omp_default_mem_alloc;
968 extern omp_allocator_handle_t const omp_large_cap_mem_alloc;
969 extern omp_allocator_handle_t const omp_const_mem_alloc;
970 extern omp_allocator_handle_t const omp_high_bw_mem_alloc;
971 extern omp_allocator_handle_t const omp_low_lat_mem_alloc;
972 extern omp_allocator_handle_t const omp_cgroup_mem_alloc;
973 extern omp_allocator_handle_t const omp_pteam_mem_alloc;
974 extern omp_allocator_handle_t const omp_thread_mem_alloc;
975 // Preview of target memory support
976 extern omp_allocator_handle_t const llvm_omp_target_host_mem_alloc;
977 extern omp_allocator_handle_t const llvm_omp_target_shared_mem_alloc;
978 extern omp_allocator_handle_t const llvm_omp_target_device_mem_alloc;
979 extern omp_allocator_handle_t const kmp_max_mem_alloc;
980 extern omp_allocator_handle_t __kmp_def_allocator;
981 
982 // end of duplicate type definitions from omp.h
983 #endif
984 
985 extern int __kmp_memkind_available;
986 
987 typedef omp_memspace_handle_t kmp_memspace_t; // placeholder
988 
989 typedef struct kmp_allocator_t {
990  omp_memspace_handle_t memspace;
991  void **memkind; // pointer to memkind
992  int alignment;
993  omp_alloctrait_value_t fb;
994  kmp_allocator_t *fb_data;
995  kmp_uint64 pool_size;
996  kmp_uint64 pool_used;
997 } kmp_allocator_t;
998 
999 extern omp_allocator_handle_t __kmpc_init_allocator(int gtid,
1000  omp_memspace_handle_t,
1001  int ntraits,
1002  omp_alloctrait_t traits[]);
1003 extern void __kmpc_destroy_allocator(int gtid, omp_allocator_handle_t al);
1004 extern void __kmpc_set_default_allocator(int gtid, omp_allocator_handle_t al);
1005 extern omp_allocator_handle_t __kmpc_get_default_allocator(int gtid);
1006 extern void *__kmpc_alloc(int gtid, size_t sz, omp_allocator_handle_t al);
1007 extern void *__kmpc_calloc(int gtid, size_t nmemb, size_t sz,
1008  omp_allocator_handle_t al);
1009 extern void *__kmpc_realloc(int gtid, void *ptr, size_t sz,
1010  omp_allocator_handle_t al,
1011  omp_allocator_handle_t free_al);
1012 extern void __kmpc_free(int gtid, void *ptr, omp_allocator_handle_t al);
1013 
1014 extern void __kmp_init_memkind();
1015 extern void __kmp_fini_memkind();
1016 extern void __kmp_init_target_mem();
1017 
1018 /* ------------------------------------------------------------------------ */
1019 
1020 #define KMP_UINT64_MAX \
1021  (~((kmp_uint64)1 << ((sizeof(kmp_uint64) * (1 << 3)) - 1)))
1022 
1023 #define KMP_MIN_NTH 1
1024 
1025 #ifndef KMP_MAX_NTH
1026 #if defined(PTHREAD_THREADS_MAX) && PTHREAD_THREADS_MAX < INT_MAX
1027 #define KMP_MAX_NTH PTHREAD_THREADS_MAX
1028 #else
1029 #define KMP_MAX_NTH INT_MAX
1030 #endif
1031 #endif /* KMP_MAX_NTH */
1032 
1033 #ifdef PTHREAD_STACK_MIN
1034 #define KMP_MIN_STKSIZE PTHREAD_STACK_MIN
1035 #else
1036 #define KMP_MIN_STKSIZE ((size_t)(32 * 1024))
1037 #endif
1038 
1039 #define KMP_MAX_STKSIZE (~((size_t)1 << ((sizeof(size_t) * (1 << 3)) - 1)))
1040 
1041 #if KMP_ARCH_X86
1042 #define KMP_DEFAULT_STKSIZE ((size_t)(2 * 1024 * 1024))
1043 #elif KMP_ARCH_X86_64
1044 #define KMP_DEFAULT_STKSIZE ((size_t)(4 * 1024 * 1024))
1045 #define KMP_BACKUP_STKSIZE ((size_t)(2 * 1024 * 1024))
1046 #else
1047 #define KMP_DEFAULT_STKSIZE ((size_t)(1024 * 1024))
1048 #endif
1049 
1050 #define KMP_DEFAULT_MALLOC_POOL_INCR ((size_t)(1024 * 1024))
1051 #define KMP_MIN_MALLOC_POOL_INCR ((size_t)(4 * 1024))
1052 #define KMP_MAX_MALLOC_POOL_INCR \
1053  (~((size_t)1 << ((sizeof(size_t) * (1 << 3)) - 1)))
1054 
1055 #define KMP_MIN_STKOFFSET (0)
1056 #define KMP_MAX_STKOFFSET KMP_MAX_STKSIZE
1057 #if KMP_OS_DARWIN
1058 #define KMP_DEFAULT_STKOFFSET KMP_MIN_STKOFFSET
1059 #else
1060 #define KMP_DEFAULT_STKOFFSET CACHE_LINE
1061 #endif
1062 
1063 #define KMP_MIN_STKPADDING (0)
1064 #define KMP_MAX_STKPADDING (2 * 1024 * 1024)
1065 
1066 #define KMP_BLOCKTIME_MULTIPLIER \
1067  (1000) /* number of blocktime units per second */
1068 #define KMP_MIN_BLOCKTIME (0)
1069 #define KMP_MAX_BLOCKTIME \
1070  (INT_MAX) /* Must be this for "infinite" setting the work */
1071 #define KMP_DEFAULT_BLOCKTIME (200) /* __kmp_blocktime is in milliseconds */
1072 
1073 #if KMP_USE_MONITOR
1074 #define KMP_DEFAULT_MONITOR_STKSIZE ((size_t)(64 * 1024))
1075 #define KMP_MIN_MONITOR_WAKEUPS (1) // min times monitor wakes up per second
1076 #define KMP_MAX_MONITOR_WAKEUPS (1000) // max times monitor can wake up per sec
1077 
1078 /* Calculate new number of monitor wakeups for a specific block time based on
1079  previous monitor_wakeups. Only allow increasing number of wakeups */
1080 #define KMP_WAKEUPS_FROM_BLOCKTIME(blocktime, monitor_wakeups) \
1081  (((blocktime) == KMP_MAX_BLOCKTIME) ? (monitor_wakeups) \
1082  : ((blocktime) == KMP_MIN_BLOCKTIME) ? KMP_MAX_MONITOR_WAKEUPS \
1083  : ((monitor_wakeups) > (KMP_BLOCKTIME_MULTIPLIER / (blocktime))) \
1084  ? (monitor_wakeups) \
1085  : (KMP_BLOCKTIME_MULTIPLIER) / (blocktime))
1086 
1087 /* Calculate number of intervals for a specific block time based on
1088  monitor_wakeups */
1089 #define KMP_INTERVALS_FROM_BLOCKTIME(blocktime, monitor_wakeups) \
1090  (((blocktime) + (KMP_BLOCKTIME_MULTIPLIER / (monitor_wakeups)) - 1) / \
1091  (KMP_BLOCKTIME_MULTIPLIER / (monitor_wakeups)))
1092 #else
1093 #define KMP_BLOCKTIME(team, tid) \
1094  (get__bt_set(team, tid) ? get__blocktime(team, tid) : __kmp_dflt_blocktime)
1095 #if KMP_OS_UNIX && (KMP_ARCH_X86 || KMP_ARCH_X86_64)
1096 // HW TSC is used to reduce overhead (clock tick instead of nanosecond).
1097 extern kmp_uint64 __kmp_ticks_per_msec;
1098 #if KMP_COMPILER_ICC
1099 #define KMP_NOW() ((kmp_uint64)_rdtsc())
1100 #else
1101 #define KMP_NOW() __kmp_hardware_timestamp()
1102 #endif
1103 #define KMP_NOW_MSEC() (KMP_NOW() / __kmp_ticks_per_msec)
1104 #define KMP_BLOCKTIME_INTERVAL(team, tid) \
1105  (KMP_BLOCKTIME(team, tid) * __kmp_ticks_per_msec)
1106 #define KMP_BLOCKING(goal, count) ((goal) > KMP_NOW())
1107 #else
1108 // System time is retrieved sporadically while blocking.
1109 extern kmp_uint64 __kmp_now_nsec();
1110 #define KMP_NOW() __kmp_now_nsec()
1111 #define KMP_NOW_MSEC() (KMP_NOW() / KMP_USEC_PER_SEC)
1112 #define KMP_BLOCKTIME_INTERVAL(team, tid) \
1113  (KMP_BLOCKTIME(team, tid) * KMP_USEC_PER_SEC)
1114 #define KMP_BLOCKING(goal, count) ((count) % 1000 != 0 || (goal) > KMP_NOW())
1115 #endif
1116 #endif // KMP_USE_MONITOR
1117 
1118 #define KMP_MIN_STATSCOLS 40
1119 #define KMP_MAX_STATSCOLS 4096
1120 #define KMP_DEFAULT_STATSCOLS 80
1121 
1122 #define KMP_MIN_INTERVAL 0
1123 #define KMP_MAX_INTERVAL (INT_MAX - 1)
1124 #define KMP_DEFAULT_INTERVAL 0
1125 
1126 #define KMP_MIN_CHUNK 1
1127 #define KMP_MAX_CHUNK (INT_MAX - 1)
1128 #define KMP_DEFAULT_CHUNK 1
1129 
1130 #define KMP_MIN_DISP_NUM_BUFF 1
1131 #define KMP_DFLT_DISP_NUM_BUFF 7
1132 #define KMP_MAX_DISP_NUM_BUFF 4096
1133 
1134 #define KMP_MAX_ORDERED 8
1135 
1136 #define KMP_MAX_FIELDS 32
1137 
1138 #define KMP_MAX_BRANCH_BITS 31
1139 
1140 #define KMP_MAX_ACTIVE_LEVELS_LIMIT INT_MAX
1141 
1142 #define KMP_MAX_DEFAULT_DEVICE_LIMIT INT_MAX
1143 
1144 #define KMP_MAX_TASK_PRIORITY_LIMIT INT_MAX
1145 
1146 /* Minimum number of threads before switch to TLS gtid (experimentally
1147  determined) */
1148 /* josh TODO: what about OS X* tuning? */
1149 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
1150 #define KMP_TLS_GTID_MIN 5
1151 #else
1152 #define KMP_TLS_GTID_MIN INT_MAX
1153 #endif
1154 
1155 #define KMP_MASTER_TID(tid) (0 == (tid))
1156 #define KMP_WORKER_TID(tid) (0 != (tid))
1157 
1158 #define KMP_MASTER_GTID(gtid) (0 == __kmp_tid_from_gtid((gtid)))
1159 #define KMP_WORKER_GTID(gtid) (0 != __kmp_tid_from_gtid((gtid)))
1160 #define KMP_INITIAL_GTID(gtid) (0 == (gtid))
1161 
1162 #ifndef TRUE
1163 #define FALSE 0
1164 #define TRUE (!FALSE)
1165 #endif
1166 
1167 /* NOTE: all of the following constants must be even */
1168 
1169 #if KMP_OS_WINDOWS
1170 #define KMP_INIT_WAIT 64U /* initial number of spin-tests */
1171 #define KMP_NEXT_WAIT 32U /* susequent number of spin-tests */
1172 #elif KMP_OS_LINUX
1173 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1174 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1175 #elif KMP_OS_DARWIN
1176 /* TODO: tune for KMP_OS_DARWIN */
1177 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1178 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1179 #elif KMP_OS_DRAGONFLY
1180 /* TODO: tune for KMP_OS_DRAGONFLY */
1181 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1182 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1183 #elif KMP_OS_FREEBSD
1184 /* TODO: tune for KMP_OS_FREEBSD */
1185 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1186 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1187 #elif KMP_OS_NETBSD
1188 /* TODO: tune for KMP_OS_NETBSD */
1189 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1190 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1191 #elif KMP_OS_HURD
1192 /* TODO: tune for KMP_OS_HURD */
1193 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1194 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1195 #elif KMP_OS_OPENBSD
1196 /* TODO: tune for KMP_OS_OPENBSD */
1197 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1198 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1199 #endif
1200 
1201 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
1202 typedef struct kmp_cpuid {
1203  kmp_uint32 eax;
1204  kmp_uint32 ebx;
1205  kmp_uint32 ecx;
1206  kmp_uint32 edx;
1207 } kmp_cpuid_t;
1208 
1209 typedef struct kmp_cpuinfo {
1210  int initialized; // If 0, other fields are not initialized.
1211  int signature; // CPUID(1).EAX
1212  int family; // CPUID(1).EAX[27:20]+CPUID(1).EAX[11:8] (Extended Family+Family)
1213  int model; // ( CPUID(1).EAX[19:16] << 4 ) + CPUID(1).EAX[7:4] ( ( Extended
1214  // Model << 4 ) + Model)
1215  int stepping; // CPUID(1).EAX[3:0] ( Stepping )
1216  int sse2; // 0 if SSE2 instructions are not supported, 1 otherwise.
1217  int rtm; // 0 if RTM instructions are not supported, 1 otherwise.
1218  int apic_id;
1219  int physical_id;
1220  int logical_id;
1221  kmp_uint64 frequency; // Nominal CPU frequency in Hz.
1222  char name[3 * sizeof(kmp_cpuid_t)]; // CPUID(0x80000002,0x80000003,0x80000004)
1223 } kmp_cpuinfo_t;
1224 
1225 extern void __kmp_query_cpuid(kmp_cpuinfo_t *p);
1226 
1227 #if KMP_OS_UNIX
1228 // subleaf is only needed for cache and topology discovery and can be set to
1229 // zero in most cases
1230 static inline void __kmp_x86_cpuid(int leaf, int subleaf, struct kmp_cpuid *p) {
1231  __asm__ __volatile__("cpuid"
1232  : "=a"(p->eax), "=b"(p->ebx), "=c"(p->ecx), "=d"(p->edx)
1233  : "a"(leaf), "c"(subleaf));
1234 }
1235 // Load p into FPU control word
1236 static inline void __kmp_load_x87_fpu_control_word(const kmp_int16 *p) {
1237  __asm__ __volatile__("fldcw %0" : : "m"(*p));
1238 }
1239 // Store FPU control word into p
1240 static inline void __kmp_store_x87_fpu_control_word(kmp_int16 *p) {
1241  __asm__ __volatile__("fstcw %0" : "=m"(*p));
1242 }
1243 static inline void __kmp_clear_x87_fpu_status_word() {
1244 #if KMP_MIC
1245  // 32-bit protected mode x87 FPU state
1246  struct x87_fpu_state {
1247  unsigned cw;
1248  unsigned sw;
1249  unsigned tw;
1250  unsigned fip;
1251  unsigned fips;
1252  unsigned fdp;
1253  unsigned fds;
1254  };
1255  struct x87_fpu_state fpu_state = {0, 0, 0, 0, 0, 0, 0};
1256  __asm__ __volatile__("fstenv %0\n\t" // store FP env
1257  "andw $0x7f00, %1\n\t" // clear 0-7,15 bits of FP SW
1258  "fldenv %0\n\t" // load FP env back
1259  : "+m"(fpu_state), "+m"(fpu_state.sw));
1260 #else
1261  __asm__ __volatile__("fnclex");
1262 #endif // KMP_MIC
1263 }
1264 #if __SSE__
1265 static inline void __kmp_load_mxcsr(const kmp_uint32 *p) { _mm_setcsr(*p); }
1266 static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = _mm_getcsr(); }
1267 #else
1268 static inline void __kmp_load_mxcsr(const kmp_uint32 *p) {}
1269 static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = 0; }
1270 #endif
1271 #else
1272 // Windows still has these as external functions in assembly file
1273 extern void __kmp_x86_cpuid(int mode, int mode2, struct kmp_cpuid *p);
1274 extern void __kmp_load_x87_fpu_control_word(const kmp_int16 *p);
1275 extern void __kmp_store_x87_fpu_control_word(kmp_int16 *p);
1276 extern void __kmp_clear_x87_fpu_status_word();
1277 static inline void __kmp_load_mxcsr(const kmp_uint32 *p) { _mm_setcsr(*p); }
1278 static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = _mm_getcsr(); }
1279 #endif // KMP_OS_UNIX
1280 
1281 #define KMP_X86_MXCSR_MASK 0xffffffc0 /* ignore status flags (6 lsb) */
1282 
1283 #if KMP_ARCH_X86
1284 extern void __kmp_x86_pause(void);
1285 #elif KMP_MIC
1286 // Performance testing on KNC (C0QS-7120 P/A/X/D, 61-core, 16 GB Memory) showed
1287 // regression after removal of extra PAUSE from spin loops. Changing
1288 // the delay from 100 to 300 showed even better performance than double PAUSE
1289 // on Spec OMP2001 and LCPC tasking tests, no regressions on EPCC.
1290 static inline void __kmp_x86_pause(void) { _mm_delay_32(300); }
1291 #else
1292 static inline void __kmp_x86_pause(void) { _mm_pause(); }
1293 #endif
1294 #define KMP_CPU_PAUSE() __kmp_x86_pause()
1295 #elif KMP_ARCH_PPC64
1296 #define KMP_PPC64_PRI_LOW() __asm__ volatile("or 1, 1, 1")
1297 #define KMP_PPC64_PRI_MED() __asm__ volatile("or 2, 2, 2")
1298 #define KMP_PPC64_PRI_LOC_MB() __asm__ volatile("" : : : "memory")
1299 #define KMP_CPU_PAUSE() \
1300  do { \
1301  KMP_PPC64_PRI_LOW(); \
1302  KMP_PPC64_PRI_MED(); \
1303  KMP_PPC64_PRI_LOC_MB(); \
1304  } while (0)
1305 #else
1306 #define KMP_CPU_PAUSE() /* nothing to do */
1307 #endif
1308 
1309 #define KMP_INIT_YIELD(count) \
1310  { (count) = __kmp_yield_init; }
1311 
1312 #define KMP_OVERSUBSCRIBED \
1313  (TCR_4(__kmp_nth) > (__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc))
1314 
1315 #define KMP_TRY_YIELD \
1316  ((__kmp_use_yield == 1) || (__kmp_use_yield == 2 && (KMP_OVERSUBSCRIBED)))
1317 
1318 #define KMP_TRY_YIELD_OVERSUB \
1319  ((__kmp_use_yield == 1 || __kmp_use_yield == 2) && (KMP_OVERSUBSCRIBED))
1320 
1321 #define KMP_YIELD(cond) \
1322  { \
1323  KMP_CPU_PAUSE(); \
1324  if ((cond) && (KMP_TRY_YIELD)) \
1325  __kmp_yield(); \
1326  }
1327 
1328 #define KMP_YIELD_OVERSUB() \
1329  { \
1330  KMP_CPU_PAUSE(); \
1331  if ((KMP_TRY_YIELD_OVERSUB)) \
1332  __kmp_yield(); \
1333  }
1334 
1335 // Note the decrement of 2 in the following Macros. With KMP_LIBRARY=turnaround,
1336 // there should be no yielding since initial value from KMP_INIT_YIELD() is odd.
1337 #define KMP_YIELD_SPIN(count) \
1338  { \
1339  KMP_CPU_PAUSE(); \
1340  if (KMP_TRY_YIELD) { \
1341  (count) -= 2; \
1342  if (!(count)) { \
1343  __kmp_yield(); \
1344  (count) = __kmp_yield_next; \
1345  } \
1346  } \
1347  }
1348 
1349 #define KMP_YIELD_OVERSUB_ELSE_SPIN(count) \
1350  { \
1351  KMP_CPU_PAUSE(); \
1352  if ((KMP_TRY_YIELD_OVERSUB)) \
1353  __kmp_yield(); \
1354  else if (__kmp_use_yield == 1) { \
1355  (count) -= 2; \
1356  if (!(count)) { \
1357  __kmp_yield(); \
1358  (count) = __kmp_yield_next; \
1359  } \
1360  } \
1361  }
1362 
1363 // User-level Monitor/Mwait
1364 #if KMP_HAVE_UMWAIT
1365 // We always try for UMWAIT first
1366 #if KMP_HAVE_WAITPKG_INTRINSICS
1367 #if KMP_HAVE_IMMINTRIN_H
1368 #include <immintrin.h>
1369 #elif KMP_HAVE_INTRIN_H
1370 #include <intrin.h>
1371 #endif
1372 #endif // KMP_HAVE_WAITPKG_INTRINSICS
1373 KMP_ATTRIBUTE_TARGET_WAITPKG
1374 static inline int __kmp_tpause(uint32_t hint, uint64_t counter) {
1375 #if !KMP_HAVE_WAITPKG_INTRINSICS
1376  uint32_t timeHi = uint32_t(counter >> 32);
1377  uint32_t timeLo = uint32_t(counter & 0xffffffff);
1378  char flag;
1379  __asm__ volatile("#tpause\n.byte 0x66, 0x0F, 0xAE, 0xF1\n"
1380  "setb %0"
1381  : "=r"(flag)
1382  : "a"(timeLo), "d"(timeHi), "c"(hint)
1383  :);
1384  return flag;
1385 #else
1386  return _tpause(hint, counter);
1387 #endif
1388 }
1389 KMP_ATTRIBUTE_TARGET_WAITPKG
1390 static inline void __kmp_umonitor(void *cacheline) {
1391 #if !KMP_HAVE_WAITPKG_INTRINSICS
1392  __asm__ volatile("# umonitor\n.byte 0xF3, 0x0F, 0xAE, 0x01 "
1393  :
1394  : "a"(cacheline)
1395  :);
1396 #else
1397  _umonitor(cacheline);
1398 #endif
1399 }
1400 KMP_ATTRIBUTE_TARGET_WAITPKG
1401 static inline int __kmp_umwait(uint32_t hint, uint64_t counter) {
1402 #if !KMP_HAVE_WAITPKG_INTRINSICS
1403  uint32_t timeHi = uint32_t(counter >> 32);
1404  uint32_t timeLo = uint32_t(counter & 0xffffffff);
1405  char flag;
1406  __asm__ volatile("#umwait\n.byte 0xF2, 0x0F, 0xAE, 0xF1\n"
1407  "setb %0"
1408  : "=r"(flag)
1409  : "a"(timeLo), "d"(timeHi), "c"(hint)
1410  :);
1411  return flag;
1412 #else
1413  return _umwait(hint, counter);
1414 #endif
1415 }
1416 #elif KMP_HAVE_MWAIT
1417 #if KMP_OS_UNIX
1418 #include <pmmintrin.h>
1419 #else
1420 #include <intrin.h>
1421 #endif
1422 #if KMP_OS_UNIX
1423 __attribute__((target("sse3")))
1424 #endif
1425 static inline void
1426 __kmp_mm_monitor(void *cacheline, unsigned extensions, unsigned hints) {
1427  _mm_monitor(cacheline, extensions, hints);
1428 }
1429 #if KMP_OS_UNIX
1430 __attribute__((target("sse3")))
1431 #endif
1432 static inline void
1433 __kmp_mm_mwait(unsigned extensions, unsigned hints) {
1434  _mm_mwait(extensions, hints);
1435 }
1436 #endif // KMP_HAVE_UMWAIT
1437 
1438 /* ------------------------------------------------------------------------ */
1439 /* Support datatypes for the orphaned construct nesting checks. */
1440 /* ------------------------------------------------------------------------ */
1441 
1442 enum cons_type {
1443  ct_none,
1444  ct_parallel,
1445  ct_pdo,
1446  ct_pdo_ordered,
1447  ct_psections,
1448  ct_psingle,
1449  ct_critical,
1450  ct_ordered_in_parallel,
1451  ct_ordered_in_pdo,
1452  ct_master,
1453  ct_reduce,
1454  ct_barrier,
1455  ct_masked
1456 };
1457 
1458 #define IS_CONS_TYPE_ORDERED(ct) ((ct) == ct_pdo_ordered)
1459 
1460 struct cons_data {
1461  ident_t const *ident;
1462  enum cons_type type;
1463  int prev;
1464  kmp_user_lock_p
1465  name; /* address exclusively for critical section name comparison */
1466 };
1467 
1468 struct cons_header {
1469  int p_top, w_top, s_top;
1470  int stack_size, stack_top;
1471  struct cons_data *stack_data;
1472 };
1473 
1474 struct kmp_region_info {
1475  char *text;
1476  int offset[KMP_MAX_FIELDS];
1477  int length[KMP_MAX_FIELDS];
1478 };
1479 
1480 /* ---------------------------------------------------------------------- */
1481 /* ---------------------------------------------------------------------- */
1482 
1483 #if KMP_OS_WINDOWS
1484 typedef HANDLE kmp_thread_t;
1485 typedef DWORD kmp_key_t;
1486 #endif /* KMP_OS_WINDOWS */
1487 
1488 #if KMP_OS_UNIX
1489 typedef pthread_t kmp_thread_t;
1490 typedef pthread_key_t kmp_key_t;
1491 #endif
1492 
1493 extern kmp_key_t __kmp_gtid_threadprivate_key;
1494 
1495 typedef struct kmp_sys_info {
1496  long maxrss; /* the maximum resident set size utilized (in kilobytes) */
1497  long minflt; /* the number of page faults serviced without any I/O */
1498  long majflt; /* the number of page faults serviced that required I/O */
1499  long nswap; /* the number of times a process was "swapped" out of memory */
1500  long inblock; /* the number of times the file system had to perform input */
1501  long oublock; /* the number of times the file system had to perform output */
1502  long nvcsw; /* the number of times a context switch was voluntarily */
1503  long nivcsw; /* the number of times a context switch was forced */
1504 } kmp_sys_info_t;
1505 
1506 #if USE_ITT_BUILD
1507 // We cannot include "kmp_itt.h" due to circular dependency. Declare the only
1508 // required type here. Later we will check the type meets requirements.
1509 typedef int kmp_itt_mark_t;
1510 #define KMP_ITT_DEBUG 0
1511 #endif /* USE_ITT_BUILD */
1512 
1513 typedef kmp_int32 kmp_critical_name[8];
1514 
1524 typedef void (*kmpc_micro)(kmp_int32 *global_tid, kmp_int32 *bound_tid, ...);
1525 typedef void (*kmpc_micro_bound)(kmp_int32 *bound_tid, kmp_int32 *bound_nth,
1526  ...);
1527 
1532 /* ---------------------------------------------------------------------------
1533  */
1534 /* Threadprivate initialization/finalization function declarations */
1535 
1536 /* for non-array objects: __kmpc_threadprivate_register() */
1537 
1542 typedef void *(*kmpc_ctor)(void *);
1543 
1548 typedef void (*kmpc_dtor)(
1549  void * /*, size_t */); /* 2nd arg: magic number for KCC unused by Intel
1550  compiler */
1555 typedef void *(*kmpc_cctor)(void *, void *);
1556 
1557 /* for array objects: __kmpc_threadprivate_register_vec() */
1558 /* First arg: "this" pointer */
1559 /* Last arg: number of array elements */
1565 typedef void *(*kmpc_ctor_vec)(void *, size_t);
1571 typedef void (*kmpc_dtor_vec)(void *, size_t);
1577 typedef void *(*kmpc_cctor_vec)(void *, void *,
1578  size_t); /* function unused by compiler */
1579 
1584 /* keeps tracked of threadprivate cache allocations for cleanup later */
1585 typedef struct kmp_cached_addr {
1586  void **addr; /* address of allocated cache */
1587  void ***compiler_cache; /* pointer to compiler's cache */
1588  void *data; /* pointer to global data */
1589  struct kmp_cached_addr *next; /* pointer to next cached address */
1590 } kmp_cached_addr_t;
1591 
1592 struct private_data {
1593  struct private_data *next; /* The next descriptor in the list */
1594  void *data; /* The data buffer for this descriptor */
1595  int more; /* The repeat count for this descriptor */
1596  size_t size; /* The data size for this descriptor */
1597 };
1598 
1599 struct private_common {
1600  struct private_common *next;
1601  struct private_common *link;
1602  void *gbl_addr;
1603  void *par_addr; /* par_addr == gbl_addr for PRIMARY thread */
1604  size_t cmn_size;
1605 };
1606 
1607 struct shared_common {
1608  struct shared_common *next;
1609  struct private_data *pod_init;
1610  void *obj_init;
1611  void *gbl_addr;
1612  union {
1613  kmpc_ctor ctor;
1614  kmpc_ctor_vec ctorv;
1615  } ct;
1616  union {
1617  kmpc_cctor cctor;
1618  kmpc_cctor_vec cctorv;
1619  } cct;
1620  union {
1621  kmpc_dtor dtor;
1622  kmpc_dtor_vec dtorv;
1623  } dt;
1624  size_t vec_len;
1625  int is_vec;
1626  size_t cmn_size;
1627 };
1628 
1629 #define KMP_HASH_TABLE_LOG2 9 /* log2 of the hash table size */
1630 #define KMP_HASH_TABLE_SIZE \
1631  (1 << KMP_HASH_TABLE_LOG2) /* size of the hash table */
1632 #define KMP_HASH_SHIFT 3 /* throw away this many low bits from the address */
1633 #define KMP_HASH(x) \
1634  ((((kmp_uintptr_t)x) >> KMP_HASH_SHIFT) & (KMP_HASH_TABLE_SIZE - 1))
1635 
1636 struct common_table {
1637  struct private_common *data[KMP_HASH_TABLE_SIZE];
1638 };
1639 
1640 struct shared_table {
1641  struct shared_common *data[KMP_HASH_TABLE_SIZE];
1642 };
1643 
1644 /* ------------------------------------------------------------------------ */
1645 
1646 #if KMP_USE_HIER_SCHED
1647 // Shared barrier data that exists inside a single unit of the scheduling
1648 // hierarchy
1649 typedef struct kmp_hier_private_bdata_t {
1650  kmp_int32 num_active;
1651  kmp_uint64 index;
1652  kmp_uint64 wait_val[2];
1653 } kmp_hier_private_bdata_t;
1654 #endif
1655 
1656 typedef struct kmp_sched_flags {
1657  unsigned ordered : 1;
1658  unsigned nomerge : 1;
1659  unsigned contains_last : 1;
1660 #if KMP_USE_HIER_SCHED
1661  unsigned use_hier : 1;
1662  unsigned unused : 28;
1663 #else
1664  unsigned unused : 29;
1665 #endif
1666 } kmp_sched_flags_t;
1667 
1668 KMP_BUILD_ASSERT(sizeof(kmp_sched_flags_t) == 4);
1669 
1670 #if KMP_STATIC_STEAL_ENABLED
1671 typedef struct KMP_ALIGN_CACHE dispatch_private_info32 {
1672  kmp_int32 count;
1673  kmp_int32 ub;
1674  /* Adding KMP_ALIGN_CACHE here doesn't help / can hurt performance */
1675  kmp_int32 lb;
1676  kmp_int32 st;
1677  kmp_int32 tc;
1678  kmp_int32 static_steal_counter; /* for static_steal only; maybe better to put
1679  after ub */
1680  kmp_lock_t *th_steal_lock; // lock used for chunk stealing
1681  // KMP_ALIGN( 16 ) ensures ( if the KMP_ALIGN macro is turned on )
1682  // a) parm3 is properly aligned and
1683  // b) all parm1-4 are in the same cache line.
1684  // Because of parm1-4 are used together, performance seems to be better
1685  // if they are in the same line (not measured though).
1686 
1687  struct KMP_ALIGN(32) { // AC: changed 16 to 32 in order to simplify template
1688  kmp_int32 parm1; // structures in kmp_dispatch.cpp. This should
1689  kmp_int32 parm2; // make no real change at least while padding is off.
1690  kmp_int32 parm3;
1691  kmp_int32 parm4;
1692  };
1693 
1694  kmp_uint32 ordered_lower;
1695  kmp_uint32 ordered_upper;
1696 #if KMP_OS_WINDOWS
1697  // This var can be placed in the hole between 'tc' and 'parm1', instead of
1698  // 'static_steal_counter'. It would be nice to measure execution times.
1699  // Conditional if/endif can be removed at all.
1700  kmp_int32 last_upper;
1701 #endif /* KMP_OS_WINDOWS */
1702 } dispatch_private_info32_t;
1703 
1704 typedef struct KMP_ALIGN_CACHE dispatch_private_info64 {
1705  kmp_int64 count; // current chunk number for static & static-steal scheduling
1706  kmp_int64 ub; /* upper-bound */
1707  /* Adding KMP_ALIGN_CACHE here doesn't help / can hurt performance */
1708  kmp_int64 lb; /* lower-bound */
1709  kmp_int64 st; /* stride */
1710  kmp_int64 tc; /* trip count (number of iterations) */
1711  kmp_int64 static_steal_counter; /* for static_steal only; maybe better to put
1712  after ub */
1713  kmp_lock_t *th_steal_lock; // lock used for chunk stealing
1714  /* parm[1-4] are used in different ways by different scheduling algorithms */
1715 
1716  // KMP_ALIGN( 32 ) ensures ( if the KMP_ALIGN macro is turned on )
1717  // a) parm3 is properly aligned and
1718  // b) all parm1-4 are in the same cache line.
1719  // Because of parm1-4 are used together, performance seems to be better
1720  // if they are in the same line (not measured though).
1721 
1722  struct KMP_ALIGN(32) {
1723  kmp_int64 parm1;
1724  kmp_int64 parm2;
1725  kmp_int64 parm3;
1726  kmp_int64 parm4;
1727  };
1728 
1729  kmp_uint64 ordered_lower;
1730  kmp_uint64 ordered_upper;
1731 #if KMP_OS_WINDOWS
1732  // This var can be placed in the hole between 'tc' and 'parm1', instead of
1733  // 'static_steal_counter'. It would be nice to measure execution times.
1734  // Conditional if/endif can be removed at all.
1735  kmp_int64 last_upper;
1736 #endif /* KMP_OS_WINDOWS */
1737 } dispatch_private_info64_t;
1738 #else /* KMP_STATIC_STEAL_ENABLED */
1739 typedef struct KMP_ALIGN_CACHE dispatch_private_info32 {
1740  kmp_int32 lb;
1741  kmp_int32 ub;
1742  kmp_int32 st;
1743  kmp_int32 tc;
1744 
1745  kmp_int32 parm1;
1746  kmp_int32 parm2;
1747  kmp_int32 parm3;
1748  kmp_int32 parm4;
1749 
1750  kmp_int32 count;
1751 
1752  kmp_uint32 ordered_lower;
1753  kmp_uint32 ordered_upper;
1754 #if KMP_OS_WINDOWS
1755  kmp_int32 last_upper;
1756 #endif /* KMP_OS_WINDOWS */
1757 } dispatch_private_info32_t;
1758 
1759 typedef struct KMP_ALIGN_CACHE dispatch_private_info64 {
1760  kmp_int64 lb; /* lower-bound */
1761  kmp_int64 ub; /* upper-bound */
1762  kmp_int64 st; /* stride */
1763  kmp_int64 tc; /* trip count (number of iterations) */
1764 
1765  /* parm[1-4] are used in different ways by different scheduling algorithms */
1766  kmp_int64 parm1;
1767  kmp_int64 parm2;
1768  kmp_int64 parm3;
1769  kmp_int64 parm4;
1770 
1771  kmp_int64 count; /* current chunk number for static scheduling */
1772 
1773  kmp_uint64 ordered_lower;
1774  kmp_uint64 ordered_upper;
1775 #if KMP_OS_WINDOWS
1776  kmp_int64 last_upper;
1777 #endif /* KMP_OS_WINDOWS */
1778 } dispatch_private_info64_t;
1779 #endif /* KMP_STATIC_STEAL_ENABLED */
1780 
1781 typedef struct KMP_ALIGN_CACHE dispatch_private_info {
1782  union private_info {
1783  dispatch_private_info32_t p32;
1784  dispatch_private_info64_t p64;
1785  } u;
1786  enum sched_type schedule; /* scheduling algorithm */
1787  kmp_sched_flags_t flags; /* flags (e.g., ordered, nomerge, etc.) */
1788  kmp_int32 ordered_bumped;
1789  // To retain the structure size after making ordered_iteration scalar
1790  kmp_int32 ordered_dummy[KMP_MAX_ORDERED - 3];
1791  // Stack of buffers for nest of serial regions
1792  struct dispatch_private_info *next;
1793  kmp_int32 type_size; /* the size of types in private_info */
1794 #if KMP_USE_HIER_SCHED
1795  kmp_int32 hier_id;
1796  void *parent; /* hierarchical scheduling parent pointer */
1797 #endif
1798  enum cons_type pushed_ws;
1799 } dispatch_private_info_t;
1800 
1801 typedef struct dispatch_shared_info32 {
1802  /* chunk index under dynamic, number of idle threads under static-steal;
1803  iteration index otherwise */
1804  volatile kmp_uint32 iteration;
1805  volatile kmp_uint32 num_done;
1806  volatile kmp_uint32 ordered_iteration;
1807  // Dummy to retain the structure size after making ordered_iteration scalar
1808  kmp_int32 ordered_dummy[KMP_MAX_ORDERED - 1];
1809 } dispatch_shared_info32_t;
1810 
1811 typedef struct dispatch_shared_info64 {
1812  /* chunk index under dynamic, number of idle threads under static-steal;
1813  iteration index otherwise */
1814  volatile kmp_uint64 iteration;
1815  volatile kmp_uint64 num_done;
1816  volatile kmp_uint64 ordered_iteration;
1817  // Dummy to retain the structure size after making ordered_iteration scalar
1818  kmp_int64 ordered_dummy[KMP_MAX_ORDERED - 3];
1819 } dispatch_shared_info64_t;
1820 
1821 typedef struct dispatch_shared_info {
1822  union shared_info {
1823  dispatch_shared_info32_t s32;
1824  dispatch_shared_info64_t s64;
1825  } u;
1826  volatile kmp_uint32 buffer_index;
1827  volatile kmp_int32 doacross_buf_idx; // teamwise index
1828  volatile kmp_uint32 *doacross_flags; // shared array of iteration flags (0/1)
1829  kmp_int32 doacross_num_done; // count finished threads
1830 #if KMP_USE_HIER_SCHED
1831  void *hier;
1832 #endif
1833 #if KMP_USE_HWLOC
1834  // When linking with libhwloc, the ORDERED EPCC test slows down on big
1835  // machines (> 48 cores). Performance analysis showed that a cache thrash
1836  // was occurring and this padding helps alleviate the problem.
1837  char padding[64];
1838 #endif
1839 } dispatch_shared_info_t;
1840 
1841 typedef struct kmp_disp {
1842  /* Vector for ORDERED SECTION */
1843  void (*th_deo_fcn)(int *gtid, int *cid, ident_t *);
1844  /* Vector for END ORDERED SECTION */
1845  void (*th_dxo_fcn)(int *gtid, int *cid, ident_t *);
1846 
1847  dispatch_shared_info_t *th_dispatch_sh_current;
1848  dispatch_private_info_t *th_dispatch_pr_current;
1849 
1850  dispatch_private_info_t *th_disp_buffer;
1851  kmp_int32 th_disp_index;
1852  kmp_int32 th_doacross_buf_idx; // thread's doacross buffer index
1853  volatile kmp_uint32 *th_doacross_flags; // pointer to shared array of flags
1854  kmp_int64 *th_doacross_info; // info on loop bounds
1855 #if KMP_USE_INTERNODE_ALIGNMENT
1856  char more_padding[INTERNODE_CACHE_LINE];
1857 #endif
1858 } kmp_disp_t;
1859 
1860 /* ------------------------------------------------------------------------ */
1861 /* Barrier stuff */
1862 
1863 /* constants for barrier state update */
1864 #define KMP_INIT_BARRIER_STATE 0 /* should probably start from zero */
1865 #define KMP_BARRIER_SLEEP_BIT 0 /* bit used for suspend/sleep part of state */
1866 #define KMP_BARRIER_UNUSED_BIT 1 // bit that must never be set for valid state
1867 #define KMP_BARRIER_BUMP_BIT 2 /* lsb used for bump of go/arrived state */
1868 
1869 #define KMP_BARRIER_SLEEP_STATE (1 << KMP_BARRIER_SLEEP_BIT)
1870 #define KMP_BARRIER_UNUSED_STATE (1 << KMP_BARRIER_UNUSED_BIT)
1871 #define KMP_BARRIER_STATE_BUMP (1 << KMP_BARRIER_BUMP_BIT)
1872 
1873 #if (KMP_BARRIER_SLEEP_BIT >= KMP_BARRIER_BUMP_BIT)
1874 #error "Barrier sleep bit must be smaller than barrier bump bit"
1875 #endif
1876 #if (KMP_BARRIER_UNUSED_BIT >= KMP_BARRIER_BUMP_BIT)
1877 #error "Barrier unused bit must be smaller than barrier bump bit"
1878 #endif
1879 
1880 // Constants for release barrier wait state: currently, hierarchical only
1881 #define KMP_BARRIER_NOT_WAITING 0 // Normal state; worker not in wait_sleep
1882 #define KMP_BARRIER_OWN_FLAG \
1883  1 // Normal state; worker waiting on own b_go flag in release
1884 #define KMP_BARRIER_PARENT_FLAG \
1885  2 // Special state; worker waiting on parent's b_go flag in release
1886 #define KMP_BARRIER_SWITCH_TO_OWN_FLAG \
1887  3 // Special state; tells worker to shift from parent to own b_go
1888 #define KMP_BARRIER_SWITCHING \
1889  4 // Special state; worker resets appropriate flag on wake-up
1890 
1891 #define KMP_NOT_SAFE_TO_REAP \
1892  0 // Thread th_reap_state: not safe to reap (tasking)
1893 #define KMP_SAFE_TO_REAP 1 // Thread th_reap_state: safe to reap (not tasking)
1894 
1895 // The flag_type describes the storage used for the flag.
1896 enum flag_type {
1897  flag32,
1898  flag64,
1899  atomic_flag64,
1900  flag_oncore,
1901  flag_unset
1902 };
1903 
1904 enum barrier_type {
1905  bs_plain_barrier = 0, /* 0, All non-fork/join barriers (except reduction
1906  barriers if enabled) */
1907  bs_forkjoin_barrier, /* 1, All fork/join (parallel region) barriers */
1908 #if KMP_FAST_REDUCTION_BARRIER
1909  bs_reduction_barrier, /* 2, All barriers that are used in reduction */
1910 #endif // KMP_FAST_REDUCTION_BARRIER
1911  bs_last_barrier /* Just a placeholder to mark the end */
1912 };
1913 
1914 // to work with reduction barriers just like with plain barriers
1915 #if !KMP_FAST_REDUCTION_BARRIER
1916 #define bs_reduction_barrier bs_plain_barrier
1917 #endif // KMP_FAST_REDUCTION_BARRIER
1918 
1919 typedef enum kmp_bar_pat { /* Barrier communication patterns */
1920  bp_linear_bar =
1921  0, /* Single level (degenerate) tree */
1922  bp_tree_bar =
1923  1, /* Balanced tree with branching factor 2^n */
1924  bp_hyper_bar = 2, /* Hypercube-embedded tree with min
1925  branching factor 2^n */
1926  bp_hierarchical_bar = 3, /* Machine hierarchy tree */
1927  bp_dist_bar = 4, /* Distributed barrier */
1928  bp_last_bar /* Placeholder to mark the end */
1929 } kmp_bar_pat_e;
1930 
1931 #define KMP_BARRIER_ICV_PUSH 1
1932 
1933 /* Record for holding the values of the internal controls stack records */
1934 typedef struct kmp_internal_control {
1935  int serial_nesting_level; /* corresponds to the value of the
1936  th_team_serialized field */
1937  kmp_int8 dynamic; /* internal control for dynamic adjustment of threads (per
1938  thread) */
1939  kmp_int8
1940  bt_set; /* internal control for whether blocktime is explicitly set */
1941  int blocktime; /* internal control for blocktime */
1942 #if KMP_USE_MONITOR
1943  int bt_intervals; /* internal control for blocktime intervals */
1944 #endif
1945  int nproc; /* internal control for #threads for next parallel region (per
1946  thread) */
1947  int thread_limit; /* internal control for thread-limit-var */
1948  int max_active_levels; /* internal control for max_active_levels */
1949  kmp_r_sched_t
1950  sched; /* internal control for runtime schedule {sched,chunk} pair */
1951  kmp_proc_bind_t proc_bind; /* internal control for affinity */
1952  kmp_int32 default_device; /* internal control for default device */
1953  struct kmp_internal_control *next;
1954 } kmp_internal_control_t;
1955 
1956 static inline void copy_icvs(kmp_internal_control_t *dst,
1957  kmp_internal_control_t *src) {
1958  *dst = *src;
1959 }
1960 
1961 /* Thread barrier needs volatile barrier fields */
1962 typedef struct KMP_ALIGN_CACHE kmp_bstate {
1963  // th_fixed_icvs is aligned by virtue of kmp_bstate being aligned (and all
1964  // uses of it). It is not explicitly aligned below, because we *don't* want
1965  // it to be padded -- instead, we fit b_go into the same cache line with
1966  // th_fixed_icvs, enabling NGO cache lines stores in the hierarchical barrier.
1967  kmp_internal_control_t th_fixed_icvs; // Initial ICVs for the thread
1968  // Tuck b_go into end of th_fixed_icvs cache line, so it can be stored with
1969  // same NGO store
1970  volatile kmp_uint64 b_go; // STATE => task should proceed (hierarchical)
1971  KMP_ALIGN_CACHE volatile kmp_uint64
1972  b_arrived; // STATE => task reached synch point.
1973  kmp_uint32 *skip_per_level;
1974  kmp_uint32 my_level;
1975  kmp_int32 parent_tid;
1976  kmp_int32 old_tid;
1977  kmp_uint32 depth;
1978  struct kmp_bstate *parent_bar;
1979  kmp_team_t *team;
1980  kmp_uint64 leaf_state;
1981  kmp_uint32 nproc;
1982  kmp_uint8 base_leaf_kids;
1983  kmp_uint8 leaf_kids;
1984  kmp_uint8 offset;
1985  kmp_uint8 wait_flag;
1986  kmp_uint8 use_oncore_barrier;
1987 #if USE_DEBUGGER
1988  // The following field is intended for the debugger solely. Only the worker
1989  // thread itself accesses this field: the worker increases it by 1 when it
1990  // arrives to a barrier.
1991  KMP_ALIGN_CACHE kmp_uint b_worker_arrived;
1992 #endif /* USE_DEBUGGER */
1993 } kmp_bstate_t;
1994 
1995 union KMP_ALIGN_CACHE kmp_barrier_union {
1996  double b_align; /* use worst case alignment */
1997  char b_pad[KMP_PAD(kmp_bstate_t, CACHE_LINE)];
1998  kmp_bstate_t bb;
1999 };
2000 
2001 typedef union kmp_barrier_union kmp_balign_t;
2002 
2003 /* Team barrier needs only non-volatile arrived counter */
2004 union KMP_ALIGN_CACHE kmp_barrier_team_union {
2005  double b_align; /* use worst case alignment */
2006  char b_pad[CACHE_LINE];
2007  struct {
2008  kmp_uint64 b_arrived; /* STATE => task reached synch point. */
2009 #if USE_DEBUGGER
2010  // The following two fields are indended for the debugger solely. Only
2011  // primary thread of the team accesses these fields: the first one is
2012  // increased by 1 when the primary thread arrives to a barrier, the second
2013  // one is increased by one when all the threads arrived.
2014  kmp_uint b_master_arrived;
2015  kmp_uint b_team_arrived;
2016 #endif
2017  };
2018 };
2019 
2020 typedef union kmp_barrier_team_union kmp_balign_team_t;
2021 
2022 /* Padding for Linux* OS pthreads condition variables and mutexes used to signal
2023  threads when a condition changes. This is to workaround an NPTL bug where
2024  padding was added to pthread_cond_t which caused the initialization routine
2025  to write outside of the structure if compiled on pre-NPTL threads. */
2026 #if KMP_OS_WINDOWS
2027 typedef struct kmp_win32_mutex {
2028  /* The Lock */
2029  CRITICAL_SECTION cs;
2030 } kmp_win32_mutex_t;
2031 
2032 typedef struct kmp_win32_cond {
2033  /* Count of the number of waiters. */
2034  int waiters_count_;
2035 
2036  /* Serialize access to <waiters_count_> */
2037  kmp_win32_mutex_t waiters_count_lock_;
2038 
2039  /* Number of threads to release via a <cond_broadcast> or a <cond_signal> */
2040  int release_count_;
2041 
2042  /* Keeps track of the current "generation" so that we don't allow */
2043  /* one thread to steal all the "releases" from the broadcast. */
2044  int wait_generation_count_;
2045 
2046  /* A manual-reset event that's used to block and release waiting threads. */
2047  HANDLE event_;
2048 } kmp_win32_cond_t;
2049 #endif
2050 
2051 #if KMP_OS_UNIX
2052 
2053 union KMP_ALIGN_CACHE kmp_cond_union {
2054  double c_align;
2055  char c_pad[CACHE_LINE];
2056  pthread_cond_t c_cond;
2057 };
2058 
2059 typedef union kmp_cond_union kmp_cond_align_t;
2060 
2061 union KMP_ALIGN_CACHE kmp_mutex_union {
2062  double m_align;
2063  char m_pad[CACHE_LINE];
2064  pthread_mutex_t m_mutex;
2065 };
2066 
2067 typedef union kmp_mutex_union kmp_mutex_align_t;
2068 
2069 #endif /* KMP_OS_UNIX */
2070 
2071 typedef struct kmp_desc_base {
2072  void *ds_stackbase;
2073  size_t ds_stacksize;
2074  int ds_stackgrow;
2075  kmp_thread_t ds_thread;
2076  volatile int ds_tid;
2077  int ds_gtid;
2078 #if KMP_OS_WINDOWS
2079  volatile int ds_alive;
2080  DWORD ds_thread_id;
2081 /* ds_thread keeps thread handle on Windows* OS. It is enough for RTL purposes.
2082  However, debugger support (libomp_db) cannot work with handles, because they
2083  uncomparable. For example, debugger requests info about thread with handle h.
2084  h is valid within debugger process, and meaningless within debugee process.
2085  Even if h is duped by call to DuplicateHandle(), so the result h' is valid
2086  within debugee process, but it is a *new* handle which does *not* equal to
2087  any other handle in debugee... The only way to compare handles is convert
2088  them to system-wide ids. GetThreadId() function is available only in
2089  Longhorn and Server 2003. :-( In contrast, GetCurrentThreadId() is available
2090  on all Windows* OS flavours (including Windows* 95). Thus, we have to get
2091  thread id by call to GetCurrentThreadId() from within the thread and save it
2092  to let libomp_db identify threads. */
2093 #endif /* KMP_OS_WINDOWS */
2094 } kmp_desc_base_t;
2095 
2096 typedef union KMP_ALIGN_CACHE kmp_desc {
2097  double ds_align; /* use worst case alignment */
2098  char ds_pad[KMP_PAD(kmp_desc_base_t, CACHE_LINE)];
2099  kmp_desc_base_t ds;
2100 } kmp_desc_t;
2101 
2102 typedef struct kmp_local {
2103  volatile int this_construct; /* count of single's encountered by thread */
2104  void *reduce_data;
2105 #if KMP_USE_BGET
2106  void *bget_data;
2107  void *bget_list;
2108 #if !USE_CMP_XCHG_FOR_BGET
2109 #ifdef USE_QUEUING_LOCK_FOR_BGET
2110  kmp_lock_t bget_lock; /* Lock for accessing bget free list */
2111 #else
2112  kmp_bootstrap_lock_t bget_lock; // Lock for accessing bget free list. Must be
2113 // bootstrap lock so we can use it at library
2114 // shutdown.
2115 #endif /* USE_LOCK_FOR_BGET */
2116 #endif /* ! USE_CMP_XCHG_FOR_BGET */
2117 #endif /* KMP_USE_BGET */
2118 
2119  PACKED_REDUCTION_METHOD_T
2120  packed_reduction_method; /* stored by __kmpc_reduce*(), used by
2121  __kmpc_end_reduce*() */
2122 
2123 } kmp_local_t;
2124 
2125 #define KMP_CHECK_UPDATE(a, b) \
2126  if ((a) != (b)) \
2127  (a) = (b)
2128 #define KMP_CHECK_UPDATE_SYNC(a, b) \
2129  if ((a) != (b)) \
2130  TCW_SYNC_PTR((a), (b))
2131 
2132 #define get__blocktime(xteam, xtid) \
2133  ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.blocktime)
2134 #define get__bt_set(xteam, xtid) \
2135  ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_set)
2136 #if KMP_USE_MONITOR
2137 #define get__bt_intervals(xteam, xtid) \
2138  ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_intervals)
2139 #endif
2140 
2141 #define get__dynamic_2(xteam, xtid) \
2142  ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.dynamic)
2143 #define get__nproc_2(xteam, xtid) \
2144  ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.nproc)
2145 #define get__sched_2(xteam, xtid) \
2146  ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.sched)
2147 
2148 #define set__blocktime_team(xteam, xtid, xval) \
2149  (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.blocktime) = \
2150  (xval))
2151 
2152 #if KMP_USE_MONITOR
2153 #define set__bt_intervals_team(xteam, xtid, xval) \
2154  (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_intervals) = \
2155  (xval))
2156 #endif
2157 
2158 #define set__bt_set_team(xteam, xtid, xval) \
2159  (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_set) = (xval))
2160 
2161 #define set__dynamic(xthread, xval) \
2162  (((xthread)->th.th_current_task->td_icvs.dynamic) = (xval))
2163 #define get__dynamic(xthread) \
2164  (((xthread)->th.th_current_task->td_icvs.dynamic) ? (FTN_TRUE) : (FTN_FALSE))
2165 
2166 #define set__nproc(xthread, xval) \
2167  (((xthread)->th.th_current_task->td_icvs.nproc) = (xval))
2168 
2169 #define set__thread_limit(xthread, xval) \
2170  (((xthread)->th.th_current_task->td_icvs.thread_limit) = (xval))
2171 
2172 #define set__max_active_levels(xthread, xval) \
2173  (((xthread)->th.th_current_task->td_icvs.max_active_levels) = (xval))
2174 
2175 #define get__max_active_levels(xthread) \
2176  ((xthread)->th.th_current_task->td_icvs.max_active_levels)
2177 
2178 #define set__sched(xthread, xval) \
2179  (((xthread)->th.th_current_task->td_icvs.sched) = (xval))
2180 
2181 #define set__proc_bind(xthread, xval) \
2182  (((xthread)->th.th_current_task->td_icvs.proc_bind) = (xval))
2183 #define get__proc_bind(xthread) \
2184  ((xthread)->th.th_current_task->td_icvs.proc_bind)
2185 
2186 // OpenMP tasking data structures
2187 
2188 typedef enum kmp_tasking_mode {
2189  tskm_immediate_exec = 0,
2190  tskm_extra_barrier = 1,
2191  tskm_task_teams = 2,
2192  tskm_max = 2
2193 } kmp_tasking_mode_t;
2194 
2195 extern kmp_tasking_mode_t
2196  __kmp_tasking_mode; /* determines how/when to execute tasks */
2197 extern int __kmp_task_stealing_constraint;
2198 extern int __kmp_enable_task_throttling;
2199 extern kmp_int32 __kmp_default_device; // Set via OMP_DEFAULT_DEVICE if
2200 // specified, defaults to 0 otherwise
2201 // Set via OMP_MAX_TASK_PRIORITY if specified, defaults to 0 otherwise
2202 extern kmp_int32 __kmp_max_task_priority;
2203 // Set via KMP_TASKLOOP_MIN_TASKS if specified, defaults to 0 otherwise
2204 extern kmp_uint64 __kmp_taskloop_min_tasks;
2205 
2206 /* NOTE: kmp_taskdata_t and kmp_task_t structures allocated in single block with
2207  taskdata first */
2208 #define KMP_TASK_TO_TASKDATA(task) (((kmp_taskdata_t *)task) - 1)
2209 #define KMP_TASKDATA_TO_TASK(taskdata) (kmp_task_t *)(taskdata + 1)
2210 
2211 // The tt_found_tasks flag is a signal to all threads in the team that tasks
2212 // were spawned and queued since the previous barrier release.
2213 #define KMP_TASKING_ENABLED(task_team) \
2214  (TRUE == TCR_SYNC_4((task_team)->tt.tt_found_tasks))
2222 typedef kmp_int32 (*kmp_routine_entry_t)(kmp_int32, void *);
2223 
2224 typedef union kmp_cmplrdata {
2225  kmp_int32 priority;
2226  kmp_routine_entry_t
2227  destructors; /* pointer to function to invoke deconstructors of
2228  firstprivate C++ objects */
2229  /* future data */
2230 } kmp_cmplrdata_t;
2231 
2232 /* sizeof_kmp_task_t passed as arg to kmpc_omp_task call */
2235 typedef struct kmp_task { /* GEH: Shouldn't this be aligned somehow? */
2236  void *shareds;
2237  kmp_routine_entry_t
2238  routine;
2239  kmp_int32 part_id;
2240  kmp_cmplrdata_t
2241  data1; /* Two known optional additions: destructors and priority */
2242  kmp_cmplrdata_t data2; /* Process destructors first, priority second */
2243  /* future data */
2244  /* private vars */
2245 } kmp_task_t;
2246 
2251 typedef struct kmp_taskgroup {
2252  std::atomic<kmp_int32> count; // number of allocated and incomplete tasks
2253  std::atomic<kmp_int32>
2254  cancel_request; // request for cancellation of this taskgroup
2255  struct kmp_taskgroup *parent; // parent taskgroup
2256  // Block of data to perform task reduction
2257  void *reduce_data; // reduction related info
2258  kmp_int32 reduce_num_data; // number of data items to reduce
2259  uintptr_t *gomp_data; // gomp reduction data
2260 } kmp_taskgroup_t;
2261 
2262 // forward declarations
2263 typedef union kmp_depnode kmp_depnode_t;
2264 typedef struct kmp_depnode_list kmp_depnode_list_t;
2265 typedef struct kmp_dephash_entry kmp_dephash_entry_t;
2266 
2267 #define KMP_DEP_IN 0x1
2268 #define KMP_DEP_OUT 0x2
2269 #define KMP_DEP_INOUT 0x3
2270 #define KMP_DEP_MTX 0x4
2271 #define KMP_DEP_SET 0x8
2272 // Compiler sends us this info:
2273 typedef struct kmp_depend_info {
2274  kmp_intptr_t base_addr;
2275  size_t len;
2276  union {
2277  kmp_uint8 flag;
2278  struct {
2279  unsigned in : 1;
2280  unsigned out : 1;
2281  unsigned mtx : 1;
2282  unsigned set : 1;
2283  } flags;
2284  };
2285 } kmp_depend_info_t;
2286 
2287 // Internal structures to work with task dependencies:
2288 struct kmp_depnode_list {
2289  kmp_depnode_t *node;
2290  kmp_depnode_list_t *next;
2291 };
2292 
2293 // Max number of mutexinoutset dependencies per node
2294 #define MAX_MTX_DEPS 4
2295 
2296 typedef struct kmp_base_depnode {
2297  kmp_depnode_list_t *successors; /* used under lock */
2298  kmp_task_t *task; /* non-NULL if depnode is active, used under lock */
2299  kmp_lock_t *mtx_locks[MAX_MTX_DEPS]; /* lock mutexinoutset dependent tasks */
2300  kmp_int32 mtx_num_locks; /* number of locks in mtx_locks array */
2301  kmp_lock_t lock; /* guards shared fields: task, successors */
2302 #if KMP_SUPPORT_GRAPH_OUTPUT
2303  kmp_uint32 id;
2304 #endif
2305  std::atomic<kmp_int32> npredecessors;
2306  std::atomic<kmp_int32> nrefs;
2307 } kmp_base_depnode_t;
2308 
2309 union KMP_ALIGN_CACHE kmp_depnode {
2310  double dn_align; /* use worst case alignment */
2311  char dn_pad[KMP_PAD(kmp_base_depnode_t, CACHE_LINE)];
2312  kmp_base_depnode_t dn;
2313 };
2314 
2315 struct kmp_dephash_entry {
2316  kmp_intptr_t addr;
2317  kmp_depnode_t *last_out;
2318  kmp_depnode_list_t *last_set;
2319  kmp_depnode_list_t *prev_set;
2320  kmp_uint8 last_flag;
2321  kmp_lock_t *mtx_lock; /* is referenced by depnodes w/mutexinoutset dep */
2322  kmp_dephash_entry_t *next_in_bucket;
2323 };
2324 
2325 typedef struct kmp_dephash {
2326  kmp_dephash_entry_t **buckets;
2327  size_t size;
2328  size_t generation;
2329  kmp_uint32 nelements;
2330  kmp_uint32 nconflicts;
2331 } kmp_dephash_t;
2332 
2333 typedef struct kmp_task_affinity_info {
2334  kmp_intptr_t base_addr;
2335  size_t len;
2336  struct {
2337  bool flag1 : 1;
2338  bool flag2 : 1;
2339  kmp_int32 reserved : 30;
2340  } flags;
2341 } kmp_task_affinity_info_t;
2342 
2343 typedef enum kmp_event_type_t {
2344  KMP_EVENT_UNINITIALIZED = 0,
2345  KMP_EVENT_ALLOW_COMPLETION = 1
2346 } kmp_event_type_t;
2347 
2348 typedef struct {
2349  kmp_event_type_t type;
2350  kmp_tas_lock_t lock;
2351  union {
2352  kmp_task_t *task;
2353  } ed;
2354 } kmp_event_t;
2355 
2356 #ifdef BUILD_TIED_TASK_STACK
2357 
2358 /* Tied Task stack definitions */
2359 typedef struct kmp_stack_block {
2360  kmp_taskdata_t *sb_block[TASK_STACK_BLOCK_SIZE];
2361  struct kmp_stack_block *sb_next;
2362  struct kmp_stack_block *sb_prev;
2363 } kmp_stack_block_t;
2364 
2365 typedef struct kmp_task_stack {
2366  kmp_stack_block_t ts_first_block; // first block of stack entries
2367  kmp_taskdata_t **ts_top; // pointer to the top of stack
2368  kmp_int32 ts_entries; // number of entries on the stack
2369 } kmp_task_stack_t;
2370 
2371 #endif // BUILD_TIED_TASK_STACK
2372 
2373 typedef struct kmp_tasking_flags { /* Total struct must be exactly 32 bits */
2374  /* Compiler flags */ /* Total compiler flags must be 16 bits */
2375  unsigned tiedness : 1; /* task is either tied (1) or untied (0) */
2376  unsigned final : 1; /* task is final(1) so execute immediately */
2377  unsigned merged_if0 : 1; /* no __kmpc_task_{begin/complete}_if0 calls in if0
2378  code path */
2379  unsigned destructors_thunk : 1; /* set if the compiler creates a thunk to
2380  invoke destructors from the runtime */
2381  unsigned proxy : 1; /* task is a proxy task (it will be executed outside the
2382  context of the RTL) */
2383  unsigned priority_specified : 1; /* set if the compiler provides priority
2384  setting for the task */
2385  unsigned detachable : 1; /* 1 == can detach */
2386  unsigned hidden_helper : 1; /* 1 == hidden helper task */
2387  unsigned reserved : 8; /* reserved for compiler use */
2388 
2389  /* Library flags */ /* Total library flags must be 16 bits */
2390  unsigned tasktype : 1; /* task is either explicit(1) or implicit (0) */
2391  unsigned task_serial : 1; // task is executed immediately (1) or deferred (0)
2392  unsigned tasking_ser : 1; // all tasks in team are either executed immediately
2393  // (1) or may be deferred (0)
2394  unsigned team_serial : 1; // entire team is serial (1) [1 thread] or parallel
2395  // (0) [>= 2 threads]
2396  /* If either team_serial or tasking_ser is set, task team may be NULL */
2397  /* Task State Flags: */
2398  unsigned started : 1; /* 1==started, 0==not started */
2399  unsigned executing : 1; /* 1==executing, 0==not executing */
2400  unsigned complete : 1; /* 1==complete, 0==not complete */
2401  unsigned freed : 1; /* 1==freed, 0==allocated */
2402  unsigned native : 1; /* 1==gcc-compiled task, 0==intel */
2403  unsigned reserved31 : 7; /* reserved for library use */
2404 
2405 } kmp_tasking_flags_t;
2406 
2407 struct kmp_taskdata { /* aligned during dynamic allocation */
2408  kmp_int32 td_task_id; /* id, assigned by debugger */
2409  kmp_tasking_flags_t td_flags; /* task flags */
2410  kmp_team_t *td_team; /* team for this task */
2411  kmp_info_p *td_alloc_thread; /* thread that allocated data structures */
2412  /* Currently not used except for perhaps IDB */
2413  kmp_taskdata_t *td_parent; /* parent task */
2414  kmp_int32 td_level; /* task nesting level */
2415  std::atomic<kmp_int32> td_untied_count; // untied task active parts counter
2416  ident_t *td_ident; /* task identifier */
2417  // Taskwait data.
2418  ident_t *td_taskwait_ident;
2419  kmp_uint32 td_taskwait_counter;
2420  kmp_int32 td_taskwait_thread; /* gtid + 1 of thread encountered taskwait */
2421  KMP_ALIGN_CACHE kmp_internal_control_t
2422  td_icvs; /* Internal control variables for the task */
2423  KMP_ALIGN_CACHE std::atomic<kmp_int32>
2424  td_allocated_child_tasks; /* Child tasks (+ current task) not yet
2425  deallocated */
2426  std::atomic<kmp_int32>
2427  td_incomplete_child_tasks; /* Child tasks not yet complete */
2428  kmp_taskgroup_t
2429  *td_taskgroup; // Each task keeps pointer to its current taskgroup
2430  kmp_dephash_t
2431  *td_dephash; // Dependencies for children tasks are tracked from here
2432  kmp_depnode_t
2433  *td_depnode; // Pointer to graph node if this task has dependencies
2434  kmp_task_team_t *td_task_team;
2435  // The global thread id of the encountering thread. We need it because when a
2436  // regular task depends on a hidden helper task, and the hidden helper task
2437  // is finished on a hidden helper thread, it will call __kmp_release_deps to
2438  // release all dependences. If now the task is a regular task, we need to pass
2439  // the encountering gtid such that the task will be picked up and executed by
2440  // its encountering team instead of hidden helper team.
2441  kmp_int32 encountering_gtid;
2442  size_t td_size_alloc; // Size of task structure, including shareds etc.
2443 #if defined(KMP_GOMP_COMPAT)
2444  // 4 or 8 byte integers for the loop bounds in GOMP_taskloop
2445  kmp_int32 td_size_loop_bounds;
2446 #endif
2447  kmp_taskdata_t *td_last_tied; // keep tied task for task scheduling constraint
2448 #if defined(KMP_GOMP_COMPAT)
2449  // GOMP sends in a copy function for copy constructors
2450  void (*td_copy_func)(void *, void *);
2451 #endif
2452  kmp_event_t td_allow_completion_event;
2453 #if OMPT_SUPPORT
2454  ompt_task_info_t ompt_task_info;
2455 #endif
2456 }; // struct kmp_taskdata
2457 
2458 // Make sure padding above worked
2459 KMP_BUILD_ASSERT(sizeof(kmp_taskdata_t) % sizeof(void *) == 0);
2460 
2461 // Data for task team but per thread
2462 typedef struct kmp_base_thread_data {
2463  kmp_info_p *td_thr; // Pointer back to thread info
2464  // Used only in __kmp_execute_tasks_template, maybe not avail until task is
2465  // queued?
2466  kmp_bootstrap_lock_t td_deque_lock; // Lock for accessing deque
2467  kmp_taskdata_t *
2468  *td_deque; // Deque of tasks encountered by td_thr, dynamically allocated
2469  kmp_int32 td_deque_size; // Size of deck
2470  kmp_uint32 td_deque_head; // Head of deque (will wrap)
2471  kmp_uint32 td_deque_tail; // Tail of deque (will wrap)
2472  kmp_int32 td_deque_ntasks; // Number of tasks in deque
2473  // GEH: shouldn't this be volatile since used in while-spin?
2474  kmp_int32 td_deque_last_stolen; // Thread number of last successful steal
2475 #ifdef BUILD_TIED_TASK_STACK
2476  kmp_task_stack_t td_susp_tied_tasks; // Stack of suspended tied tasks for task
2477 // scheduling constraint
2478 #endif // BUILD_TIED_TASK_STACK
2479 } kmp_base_thread_data_t;
2480 
2481 #define TASK_DEQUE_BITS 8 // Used solely to define INITIAL_TASK_DEQUE_SIZE
2482 #define INITIAL_TASK_DEQUE_SIZE (1 << TASK_DEQUE_BITS)
2483 
2484 #define TASK_DEQUE_SIZE(td) ((td).td_deque_size)
2485 #define TASK_DEQUE_MASK(td) ((td).td_deque_size - 1)
2486 
2487 typedef union KMP_ALIGN_CACHE kmp_thread_data {
2488  kmp_base_thread_data_t td;
2489  double td_align; /* use worst case alignment */
2490  char td_pad[KMP_PAD(kmp_base_thread_data_t, CACHE_LINE)];
2491 } kmp_thread_data_t;
2492 
2493 // Data for task teams which are used when tasking is enabled for the team
2494 typedef struct kmp_base_task_team {
2495  kmp_bootstrap_lock_t
2496  tt_threads_lock; /* Lock used to allocate per-thread part of task team */
2497  /* must be bootstrap lock since used at library shutdown*/
2498  kmp_task_team_t *tt_next; /* For linking the task team free list */
2499  kmp_thread_data_t
2500  *tt_threads_data; /* Array of per-thread structures for task team */
2501  /* Data survives task team deallocation */
2502  kmp_int32 tt_found_tasks; /* Have we found tasks and queued them while
2503  executing this team? */
2504  /* TRUE means tt_threads_data is set up and initialized */
2505  kmp_int32 tt_nproc; /* #threads in team */
2506  kmp_int32 tt_max_threads; // # entries allocated for threads_data array
2507  kmp_int32 tt_found_proxy_tasks; // found proxy tasks since last barrier
2508  kmp_int32 tt_untied_task_encountered;
2509  // There is hidden helper thread encountered in this task team so that we must
2510  // wait when waiting on task team
2511  kmp_int32 tt_hidden_helper_task_encountered;
2512 
2513  KMP_ALIGN_CACHE
2514  std::atomic<kmp_int32> tt_unfinished_threads; /* #threads still active */
2515 
2516  KMP_ALIGN_CACHE
2517  volatile kmp_uint32
2518  tt_active; /* is the team still actively executing tasks */
2519 } kmp_base_task_team_t;
2520 
2521 union KMP_ALIGN_CACHE kmp_task_team {
2522  kmp_base_task_team_t tt;
2523  double tt_align; /* use worst case alignment */
2524  char tt_pad[KMP_PAD(kmp_base_task_team_t, CACHE_LINE)];
2525 };
2526 
2527 #if (USE_FAST_MEMORY == 3) || (USE_FAST_MEMORY == 5)
2528 // Free lists keep same-size free memory slots for fast memory allocation
2529 // routines
2530 typedef struct kmp_free_list {
2531  void *th_free_list_self; // Self-allocated tasks free list
2532  void *th_free_list_sync; // Self-allocated tasks stolen/returned by other
2533  // threads
2534  void *th_free_list_other; // Non-self free list (to be returned to owner's
2535  // sync list)
2536 } kmp_free_list_t;
2537 #endif
2538 #if KMP_NESTED_HOT_TEAMS
2539 // Hot teams array keeps hot teams and their sizes for given thread. Hot teams
2540 // are not put in teams pool, and they don't put threads in threads pool.
2541 typedef struct kmp_hot_team_ptr {
2542  kmp_team_p *hot_team; // pointer to hot_team of given nesting level
2543  kmp_int32 hot_team_nth; // number of threads allocated for the hot_team
2544 } kmp_hot_team_ptr_t;
2545 #endif
2546 typedef struct kmp_teams_size {
2547  kmp_int32 nteams; // number of teams in a league
2548  kmp_int32 nth; // number of threads in each team of the league
2549 } kmp_teams_size_t;
2550 
2551 // This struct stores a thread that acts as a "root" for a contention
2552 // group. Contention groups are rooted at kmp_root threads, but also at
2553 // each primary thread of each team created in the teams construct.
2554 // This struct therefore also stores a thread_limit associated with
2555 // that contention group, and a counter to track the number of threads
2556 // active in that contention group. Each thread has a list of these: CG
2557 // root threads have an entry in their list in which cg_root refers to
2558 // the thread itself, whereas other workers in the CG will have a
2559 // single entry where cg_root is same as the entry containing their CG
2560 // root. When a thread encounters a teams construct, it will add a new
2561 // entry to the front of its list, because it now roots a new CG.
2562 typedef struct kmp_cg_root {
2563  kmp_info_p *cg_root; // "root" thread for a contention group
2564  // The CG root's limit comes from OMP_THREAD_LIMIT for root threads, or
2565  // thread_limit clause for teams primary threads
2566  kmp_int32 cg_thread_limit;
2567  kmp_int32 cg_nthreads; // Count of active threads in CG rooted at cg_root
2568  struct kmp_cg_root *up; // pointer to higher level CG root in list
2569 } kmp_cg_root_t;
2570 
2571 // OpenMP thread data structures
2572 
2573 typedef struct KMP_ALIGN_CACHE kmp_base_info {
2574  /* Start with the readonly data which is cache aligned and padded. This is
2575  written before the thread starts working by the primary thread. Uber
2576  masters may update themselves later. Usage does not consider serialized
2577  regions. */
2578  kmp_desc_t th_info;
2579  kmp_team_p *th_team; /* team we belong to */
2580  kmp_root_p *th_root; /* pointer to root of task hierarchy */
2581  kmp_info_p *th_next_pool; /* next available thread in the pool */
2582  kmp_disp_t *th_dispatch; /* thread's dispatch data */
2583  int th_in_pool; /* in thread pool (32 bits for TCR/TCW) */
2584 
2585  /* The following are cached from the team info structure */
2586  /* TODO use these in more places as determined to be needed via profiling */
2587  int th_team_nproc; /* number of threads in a team */
2588  kmp_info_p *th_team_master; /* the team's primary thread */
2589  int th_team_serialized; /* team is serialized */
2590  microtask_t th_teams_microtask; /* save entry address for teams construct */
2591  int th_teams_level; /* save initial level of teams construct */
2592 /* it is 0 on device but may be any on host */
2593 
2594 /* The blocktime info is copied from the team struct to the thread struct */
2595 /* at the start of a barrier, and the values stored in the team are used */
2596 /* at points in the code where the team struct is no longer guaranteed */
2597 /* to exist (from the POV of worker threads). */
2598 #if KMP_USE_MONITOR
2599  int th_team_bt_intervals;
2600  int th_team_bt_set;
2601 #else
2602  kmp_uint64 th_team_bt_intervals;
2603 #endif
2604 
2605 #if KMP_AFFINITY_SUPPORTED
2606  kmp_affin_mask_t *th_affin_mask; /* thread's current affinity mask */
2607 #endif
2608  omp_allocator_handle_t th_def_allocator; /* default allocator */
2609  /* The data set by the primary thread at reinit, then R/W by the worker */
2610  KMP_ALIGN_CACHE int
2611  th_set_nproc; /* if > 0, then only use this request for the next fork */
2612 #if KMP_NESTED_HOT_TEAMS
2613  kmp_hot_team_ptr_t *th_hot_teams; /* array of hot teams */
2614 #endif
2615  kmp_proc_bind_t
2616  th_set_proc_bind; /* if != proc_bind_default, use request for next fork */
2617  kmp_teams_size_t
2618  th_teams_size; /* number of teams/threads in teams construct */
2619 #if KMP_AFFINITY_SUPPORTED
2620  int th_current_place; /* place currently bound to */
2621  int th_new_place; /* place to bind to in par reg */
2622  int th_first_place; /* first place in partition */
2623  int th_last_place; /* last place in partition */
2624 #endif
2625  int th_prev_level; /* previous level for affinity format */
2626  int th_prev_num_threads; /* previous num_threads for affinity format */
2627 #if USE_ITT_BUILD
2628  kmp_uint64 th_bar_arrive_time; /* arrival to barrier timestamp */
2629  kmp_uint64 th_bar_min_time; /* minimum arrival time at the barrier */
2630  kmp_uint64 th_frame_time; /* frame timestamp */
2631 #endif /* USE_ITT_BUILD */
2632  kmp_local_t th_local;
2633  struct private_common *th_pri_head;
2634 
2635  /* Now the data only used by the worker (after initial allocation) */
2636  /* TODO the first serial team should actually be stored in the info_t
2637  structure. this will help reduce initial allocation overhead */
2638  KMP_ALIGN_CACHE kmp_team_p
2639  *th_serial_team; /*serialized team held in reserve*/
2640 
2641 #if OMPT_SUPPORT
2642  ompt_thread_info_t ompt_thread_info;
2643 #endif
2644 
2645  /* The following are also read by the primary thread during reinit */
2646  struct common_table *th_pri_common;
2647 
2648  volatile kmp_uint32 th_spin_here; /* thread-local location for spinning */
2649  /* while awaiting queuing lock acquire */
2650 
2651  volatile void *th_sleep_loc; // this points at a kmp_flag<T>
2652  flag_type th_sleep_loc_type; // enum type of flag stored in th_sleep_loc
2653 
2654  ident_t *th_ident;
2655  unsigned th_x; // Random number generator data
2656  unsigned th_a; // Random number generator data
2657 
2658  /* Tasking-related data for the thread */
2659  kmp_task_team_t *th_task_team; // Task team struct
2660  kmp_taskdata_t *th_current_task; // Innermost Task being executed
2661  kmp_uint8 th_task_state; // alternating 0/1 for task team identification
2662  kmp_uint8 *th_task_state_memo_stack; // Stack holding memos of th_task_state
2663  // at nested levels
2664  kmp_uint32 th_task_state_top; // Top element of th_task_state_memo_stack
2665  kmp_uint32 th_task_state_stack_sz; // Size of th_task_state_memo_stack
2666  kmp_uint32 th_reap_state; // Non-zero indicates thread is not
2667  // tasking, thus safe to reap
2668 
2669  /* More stuff for keeping track of active/sleeping threads (this part is
2670  written by the worker thread) */
2671  kmp_uint8 th_active_in_pool; // included in count of #active threads in pool
2672  int th_active; // ! sleeping; 32 bits for TCR/TCW
2673  std::atomic<kmp_uint32> th_used_in_team; // Flag indicating use in team
2674  // 0 = not used in team; 1 = used in team;
2675  // 2 = transitioning to not used in team; 3 = transitioning to used in team
2676  struct cons_header *th_cons; // used for consistency check
2677 #if KMP_USE_HIER_SCHED
2678  // used for hierarchical scheduling
2679  kmp_hier_private_bdata_t *th_hier_bar_data;
2680 #endif
2681 
2682  /* Add the syncronizing data which is cache aligned and padded. */
2683  KMP_ALIGN_CACHE kmp_balign_t th_bar[bs_last_barrier];
2684 
2685  KMP_ALIGN_CACHE volatile kmp_int32
2686  th_next_waiting; /* gtid+1 of next thread on lock wait queue, 0 if none */
2687 
2688 #if (USE_FAST_MEMORY == 3) || (USE_FAST_MEMORY == 5)
2689 #define NUM_LISTS 4
2690  kmp_free_list_t th_free_lists[NUM_LISTS]; // Free lists for fast memory
2691 // allocation routines
2692 #endif
2693 
2694 #if KMP_OS_WINDOWS
2695  kmp_win32_cond_t th_suspend_cv;
2696  kmp_win32_mutex_t th_suspend_mx;
2697  std::atomic<int> th_suspend_init;
2698 #endif
2699 #if KMP_OS_UNIX
2700  kmp_cond_align_t th_suspend_cv;
2701  kmp_mutex_align_t th_suspend_mx;
2702  std::atomic<int> th_suspend_init_count;
2703 #endif
2704 
2705 #if USE_ITT_BUILD
2706  kmp_itt_mark_t th_itt_mark_single;
2707 // alignment ???
2708 #endif /* USE_ITT_BUILD */
2709 #if KMP_STATS_ENABLED
2710  kmp_stats_list *th_stats;
2711 #endif
2712 #if KMP_OS_UNIX
2713  std::atomic<bool> th_blocking;
2714 #endif
2715  kmp_cg_root_t *th_cg_roots; // list of cg_roots associated with this thread
2716 } kmp_base_info_t;
2717 
2718 typedef union KMP_ALIGN_CACHE kmp_info {
2719  double th_align; /* use worst case alignment */
2720  char th_pad[KMP_PAD(kmp_base_info_t, CACHE_LINE)];
2721  kmp_base_info_t th;
2722 } kmp_info_t;
2723 
2724 // OpenMP thread team data structures
2725 
2726 typedef struct kmp_base_data {
2727  volatile kmp_uint32 t_value;
2728 } kmp_base_data_t;
2729 
2730 typedef union KMP_ALIGN_CACHE kmp_sleep_team {
2731  double dt_align; /* use worst case alignment */
2732  char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
2733  kmp_base_data_t dt;
2734 } kmp_sleep_team_t;
2735 
2736 typedef union KMP_ALIGN_CACHE kmp_ordered_team {
2737  double dt_align; /* use worst case alignment */
2738  char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
2739  kmp_base_data_t dt;
2740 } kmp_ordered_team_t;
2741 
2742 typedef int (*launch_t)(int gtid);
2743 
2744 /* Minimum number of ARGV entries to malloc if necessary */
2745 #define KMP_MIN_MALLOC_ARGV_ENTRIES 100
2746 
2747 // Set up how many argv pointers will fit in cache lines containing
2748 // t_inline_argv. Historically, we have supported at least 96 bytes. Using a
2749 // larger value for more space between the primary write/worker read section and
2750 // read/write by all section seems to buy more performance on EPCC PARALLEL.
2751 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
2752 #define KMP_INLINE_ARGV_BYTES \
2753  (4 * CACHE_LINE - \
2754  ((3 * KMP_PTR_SKIP + 2 * sizeof(int) + 2 * sizeof(kmp_int8) + \
2755  sizeof(kmp_int16) + sizeof(kmp_uint32)) % \
2756  CACHE_LINE))
2757 #else
2758 #define KMP_INLINE_ARGV_BYTES \
2759  (2 * CACHE_LINE - ((3 * KMP_PTR_SKIP + 2 * sizeof(int)) % CACHE_LINE))
2760 #endif
2761 #define KMP_INLINE_ARGV_ENTRIES (int)(KMP_INLINE_ARGV_BYTES / KMP_PTR_SKIP)
2762 
2763 typedef struct KMP_ALIGN_CACHE kmp_base_team {
2764  // Synchronization Data
2765  // ---------------------------------------------------------------------------
2766  KMP_ALIGN_CACHE kmp_ordered_team_t t_ordered;
2767  kmp_balign_team_t t_bar[bs_last_barrier];
2768  std::atomic<int> t_construct; // count of single directive encountered by team
2769  char pad[sizeof(kmp_lock_t)]; // padding to maintain performance on big iron
2770 
2771  // [0] - parallel / [1] - worksharing task reduction data shared by taskgroups
2772  std::atomic<void *> t_tg_reduce_data[2]; // to support task modifier
2773  std::atomic<int> t_tg_fini_counter[2]; // sync end of task reductions
2774 
2775  // Primary thread only
2776  // ---------------------------------------------------------------------------
2777  KMP_ALIGN_CACHE int t_master_tid; // tid of primary thread in parent team
2778  int t_master_this_cons; // "this_construct" single counter of primary thread
2779  // in parent team
2780  ident_t *t_ident; // if volatile, have to change too much other crud to
2781  // volatile too
2782  kmp_team_p *t_parent; // parent team
2783  kmp_team_p *t_next_pool; // next free team in the team pool
2784  kmp_disp_t *t_dispatch; // thread's dispatch data
2785  kmp_task_team_t *t_task_team[2]; // Task team struct; switch between 2
2786  kmp_proc_bind_t t_proc_bind; // bind type for par region
2787 #if USE_ITT_BUILD
2788  kmp_uint64 t_region_time; // region begin timestamp
2789 #endif /* USE_ITT_BUILD */
2790 
2791  // Primary thread write, workers read
2792  // --------------------------------------------------------------------------
2793  KMP_ALIGN_CACHE void **t_argv;
2794  int t_argc;
2795  int t_nproc; // number of threads in team
2796  microtask_t t_pkfn;
2797  launch_t t_invoke; // procedure to launch the microtask
2798 
2799 #if OMPT_SUPPORT
2800  ompt_team_info_t ompt_team_info;
2801  ompt_lw_taskteam_t *ompt_serialized_team_info;
2802 #endif
2803 
2804 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
2805  kmp_int8 t_fp_control_saved;
2806  kmp_int8 t_pad2b;
2807  kmp_int16 t_x87_fpu_control_word; // FP control regs
2808  kmp_uint32 t_mxcsr;
2809 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
2810 
2811  void *t_inline_argv[KMP_INLINE_ARGV_ENTRIES];
2812 
2813  KMP_ALIGN_CACHE kmp_info_t **t_threads;
2814  kmp_taskdata_t
2815  *t_implicit_task_taskdata; // Taskdata for the thread's implicit task
2816  int t_level; // nested parallel level
2817 
2818  KMP_ALIGN_CACHE int t_max_argc;
2819  int t_max_nproc; // max threads this team can handle (dynamically expandable)
2820  int t_serialized; // levels deep of serialized teams
2821  dispatch_shared_info_t *t_disp_buffer; // buffers for dispatch system
2822  int t_id; // team's id, assigned by debugger.
2823  int t_active_level; // nested active parallel level
2824  kmp_r_sched_t t_sched; // run-time schedule for the team
2825 #if KMP_AFFINITY_SUPPORTED
2826  int t_first_place; // first & last place in parent thread's partition.
2827  int t_last_place; // Restore these values to primary thread after par region.
2828 #endif // KMP_AFFINITY_SUPPORTED
2829  int t_display_affinity;
2830  int t_size_changed; // team size was changed?: 0: no, 1: yes, -1: changed via
2831  // omp_set_num_threads() call
2832  omp_allocator_handle_t t_def_allocator; /* default allocator */
2833 
2834 // Read/write by workers as well
2835 #if (KMP_ARCH_X86 || KMP_ARCH_X86_64)
2836  // Using CACHE_LINE=64 reduces memory footprint, but causes a big perf
2837  // regression of epcc 'parallel' and 'barrier' on fxe256lin01. This extra
2838  // padding serves to fix the performance of epcc 'parallel' and 'barrier' when
2839  // CACHE_LINE=64. TODO: investigate more and get rid if this padding.
2840  char dummy_padding[1024];
2841 #endif
2842  // Internal control stack for additional nested teams.
2843  KMP_ALIGN_CACHE kmp_internal_control_t *t_control_stack_top;
2844  // for SERIALIZED teams nested 2 or more levels deep
2845  // typed flag to store request state of cancellation
2846  std::atomic<kmp_int32> t_cancel_request;
2847  int t_master_active; // save on fork, restore on join
2848  void *t_copypriv_data; // team specific pointer to copyprivate data array
2849 #if KMP_OS_WINDOWS
2850  std::atomic<kmp_uint32> t_copyin_counter;
2851 #endif
2852 #if USE_ITT_BUILD
2853  void *t_stack_id; // team specific stack stitching id (for ittnotify)
2854 #endif /* USE_ITT_BUILD */
2855  distributedBarrier *b; // Distributed barrier data associated with team
2856 } kmp_base_team_t;
2857 
2858 union KMP_ALIGN_CACHE kmp_team {
2859  kmp_base_team_t t;
2860  double t_align; /* use worst case alignment */
2861  char t_pad[KMP_PAD(kmp_base_team_t, CACHE_LINE)];
2862 };
2863 
2864 typedef union KMP_ALIGN_CACHE kmp_time_global {
2865  double dt_align; /* use worst case alignment */
2866  char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
2867  kmp_base_data_t dt;
2868 } kmp_time_global_t;
2869 
2870 typedef struct kmp_base_global {
2871  /* cache-aligned */
2872  kmp_time_global_t g_time;
2873 
2874  /* non cache-aligned */
2875  volatile int g_abort;
2876  volatile int g_done;
2877 
2878  int g_dynamic;
2879  enum dynamic_mode g_dynamic_mode;
2880 } kmp_base_global_t;
2881 
2882 typedef union KMP_ALIGN_CACHE kmp_global {
2883  kmp_base_global_t g;
2884  double g_align; /* use worst case alignment */
2885  char g_pad[KMP_PAD(kmp_base_global_t, CACHE_LINE)];
2886 } kmp_global_t;
2887 
2888 typedef struct kmp_base_root {
2889  // TODO: GEH - combine r_active with r_in_parallel then r_active ==
2890  // (r_in_parallel>= 0)
2891  // TODO: GEH - then replace r_active with t_active_levels if we can to reduce
2892  // the synch overhead or keeping r_active
2893  volatile int r_active; /* TRUE if some region in a nest has > 1 thread */
2894  // keeps a count of active parallel regions per root
2895  std::atomic<int> r_in_parallel;
2896  // GEH: This is misnamed, should be r_active_levels
2897  kmp_team_t *r_root_team;
2898  kmp_team_t *r_hot_team;
2899  kmp_info_t *r_uber_thread;
2900  kmp_lock_t r_begin_lock;
2901  volatile int r_begin;
2902  int r_blocktime; /* blocktime for this root and descendants */
2903 #if KMP_AFFINITY_SUPPORTED
2904  int r_affinity_assigned;
2905 #endif // KMP_AFFINITY_SUPPORTED
2906 } kmp_base_root_t;
2907 
2908 typedef union KMP_ALIGN_CACHE kmp_root {
2909  kmp_base_root_t r;
2910  double r_align; /* use worst case alignment */
2911  char r_pad[KMP_PAD(kmp_base_root_t, CACHE_LINE)];
2912 } kmp_root_t;
2913 
2914 struct fortran_inx_info {
2915  kmp_int32 data;
2916 };
2917 
2918 /* ------------------------------------------------------------------------ */
2919 
2920 extern int __kmp_settings;
2921 extern int __kmp_duplicate_library_ok;
2922 #if USE_ITT_BUILD
2923 extern int __kmp_forkjoin_frames;
2924 extern int __kmp_forkjoin_frames_mode;
2925 #endif
2926 extern PACKED_REDUCTION_METHOD_T __kmp_force_reduction_method;
2927 extern int __kmp_determ_red;
2928 
2929 #ifdef KMP_DEBUG
2930 extern int kmp_a_debug;
2931 extern int kmp_b_debug;
2932 extern int kmp_c_debug;
2933 extern int kmp_d_debug;
2934 extern int kmp_e_debug;
2935 extern int kmp_f_debug;
2936 #endif /* KMP_DEBUG */
2937 
2938 /* For debug information logging using rotating buffer */
2939 #define KMP_DEBUG_BUF_LINES_INIT 512
2940 #define KMP_DEBUG_BUF_LINES_MIN 1
2941 
2942 #define KMP_DEBUG_BUF_CHARS_INIT 128
2943 #define KMP_DEBUG_BUF_CHARS_MIN 2
2944 
2945 extern int
2946  __kmp_debug_buf; /* TRUE means use buffer, FALSE means print to stderr */
2947 extern int __kmp_debug_buf_lines; /* How many lines of debug stored in buffer */
2948 extern int
2949  __kmp_debug_buf_chars; /* How many characters allowed per line in buffer */
2950 extern int __kmp_debug_buf_atomic; /* TRUE means use atomic update of buffer
2951  entry pointer */
2952 
2953 extern char *__kmp_debug_buffer; /* Debug buffer itself */
2954 extern std::atomic<int> __kmp_debug_count; /* Counter for number of lines
2955  printed in buffer so far */
2956 extern int __kmp_debug_buf_warn_chars; /* Keep track of char increase
2957  recommended in warnings */
2958 /* end rotating debug buffer */
2959 
2960 #ifdef KMP_DEBUG
2961 extern int __kmp_par_range; /* +1 => only go par for constructs in range */
2962 
2963 #define KMP_PAR_RANGE_ROUTINE_LEN 1024
2964 extern char __kmp_par_range_routine[KMP_PAR_RANGE_ROUTINE_LEN];
2965 #define KMP_PAR_RANGE_FILENAME_LEN 1024
2966 extern char __kmp_par_range_filename[KMP_PAR_RANGE_FILENAME_LEN];
2967 extern int __kmp_par_range_lb;
2968 extern int __kmp_par_range_ub;
2969 #endif
2970 
2971 /* For printing out dynamic storage map for threads and teams */
2972 extern int
2973  __kmp_storage_map; /* True means print storage map for threads and teams */
2974 extern int __kmp_storage_map_verbose; /* True means storage map includes
2975  placement info */
2976 extern int __kmp_storage_map_verbose_specified;
2977 
2978 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
2979 extern kmp_cpuinfo_t __kmp_cpuinfo;
2980 #endif
2981 
2982 extern volatile int __kmp_init_serial;
2983 extern volatile int __kmp_init_gtid;
2984 extern volatile int __kmp_init_common;
2985 extern volatile int __kmp_init_middle;
2986 extern volatile int __kmp_init_parallel;
2987 #if KMP_USE_MONITOR
2988 extern volatile int __kmp_init_monitor;
2989 #endif
2990 extern volatile int __kmp_init_user_locks;
2991 extern volatile int __kmp_init_hidden_helper_threads;
2992 extern int __kmp_init_counter;
2993 extern int __kmp_root_counter;
2994 extern int __kmp_version;
2995 
2996 /* list of address of allocated caches for commons */
2997 extern kmp_cached_addr_t *__kmp_threadpriv_cache_list;
2998 
2999 /* Barrier algorithm types and options */
3000 extern kmp_uint32 __kmp_barrier_gather_bb_dflt;
3001 extern kmp_uint32 __kmp_barrier_release_bb_dflt;
3002 extern kmp_bar_pat_e __kmp_barrier_gather_pat_dflt;
3003 extern kmp_bar_pat_e __kmp_barrier_release_pat_dflt;
3004 extern kmp_uint32 __kmp_barrier_gather_branch_bits[bs_last_barrier];
3005 extern kmp_uint32 __kmp_barrier_release_branch_bits[bs_last_barrier];
3006 extern kmp_bar_pat_e __kmp_barrier_gather_pattern[bs_last_barrier];
3007 extern kmp_bar_pat_e __kmp_barrier_release_pattern[bs_last_barrier];
3008 extern char const *__kmp_barrier_branch_bit_env_name[bs_last_barrier];
3009 extern char const *__kmp_barrier_pattern_env_name[bs_last_barrier];
3010 extern char const *__kmp_barrier_type_name[bs_last_barrier];
3011 extern char const *__kmp_barrier_pattern_name[bp_last_bar];
3012 
3013 /* Global Locks */
3014 extern kmp_bootstrap_lock_t __kmp_initz_lock; /* control initialization */
3015 extern kmp_bootstrap_lock_t __kmp_forkjoin_lock; /* control fork/join access */
3016 extern kmp_bootstrap_lock_t __kmp_task_team_lock;
3017 extern kmp_bootstrap_lock_t
3018  __kmp_exit_lock; /* exit() is not always thread-safe */
3019 #if KMP_USE_MONITOR
3020 extern kmp_bootstrap_lock_t
3021  __kmp_monitor_lock; /* control monitor thread creation */
3022 #endif
3023 extern kmp_bootstrap_lock_t
3024  __kmp_tp_cached_lock; /* used for the hack to allow threadprivate cache and
3025  __kmp_threads expansion to co-exist */
3026 
3027 extern kmp_lock_t __kmp_global_lock; /* control OS/global access */
3028 extern kmp_queuing_lock_t __kmp_dispatch_lock; /* control dispatch access */
3029 extern kmp_lock_t __kmp_debug_lock; /* control I/O access for KMP_DEBUG */
3030 
3031 extern enum library_type __kmp_library;
3032 
3033 extern enum sched_type __kmp_sched; /* default runtime scheduling */
3034 extern enum sched_type __kmp_static; /* default static scheduling method */
3035 extern enum sched_type __kmp_guided; /* default guided scheduling method */
3036 extern enum sched_type __kmp_auto; /* default auto scheduling method */
3037 extern int __kmp_chunk; /* default runtime chunk size */
3038 extern int __kmp_force_monotonic; /* whether monotonic scheduling forced */
3039 
3040 extern size_t __kmp_stksize; /* stack size per thread */
3041 #if KMP_USE_MONITOR
3042 extern size_t __kmp_monitor_stksize; /* stack size for monitor thread */
3043 #endif
3044 extern size_t __kmp_stkoffset; /* stack offset per thread */
3045 extern int __kmp_stkpadding; /* Should we pad root thread(s) stack */
3046 
3047 extern size_t
3048  __kmp_malloc_pool_incr; /* incremental size of pool for kmp_malloc() */
3049 extern int __kmp_env_stksize; /* was KMP_STACKSIZE specified? */
3050 extern int __kmp_env_blocktime; /* was KMP_BLOCKTIME specified? */
3051 extern int __kmp_env_checks; /* was KMP_CHECKS specified? */
3052 extern int __kmp_env_consistency_check; // was KMP_CONSISTENCY_CHECK specified?
3053 extern int __kmp_generate_warnings; /* should we issue warnings? */
3054 extern int __kmp_reserve_warn; /* have we issued reserve_threads warning? */
3055 
3056 #ifdef DEBUG_SUSPEND
3057 extern int __kmp_suspend_count; /* count inside __kmp_suspend_template() */
3058 #endif
3059 
3060 extern kmp_int32 __kmp_use_yield;
3061 extern kmp_int32 __kmp_use_yield_exp_set;
3062 extern kmp_uint32 __kmp_yield_init;
3063 extern kmp_uint32 __kmp_yield_next;
3064 
3065 /* ------------------------------------------------------------------------- */
3066 extern int __kmp_allThreadsSpecified;
3067 
3068 extern size_t __kmp_align_alloc;
3069 /* following data protected by initialization routines */
3070 extern int __kmp_xproc; /* number of processors in the system */
3071 extern int __kmp_avail_proc; /* number of processors available to the process */
3072 extern size_t __kmp_sys_min_stksize; /* system-defined minimum stack size */
3073 extern int __kmp_sys_max_nth; /* system-imposed maximum number of threads */
3074 // maximum total number of concurrently-existing threads on device
3075 extern int __kmp_max_nth;
3076 // maximum total number of concurrently-existing threads in a contention group
3077 extern int __kmp_cg_max_nth;
3078 extern int __kmp_teams_max_nth; // max threads used in a teams construct
3079 extern int __kmp_threads_capacity; /* capacity of the arrays __kmp_threads and
3080  __kmp_root */
3081 extern int __kmp_dflt_team_nth; /* default number of threads in a parallel
3082  region a la OMP_NUM_THREADS */
3083 extern int __kmp_dflt_team_nth_ub; /* upper bound on "" determined at serial
3084  initialization */
3085 extern int __kmp_tp_capacity; /* capacity of __kmp_threads if threadprivate is
3086  used (fixed) */
3087 extern int __kmp_tp_cached; /* whether threadprivate cache has been created
3088  (__kmpc_threadprivate_cached()) */
3089 extern int __kmp_dflt_blocktime; /* number of milliseconds to wait before
3090  blocking (env setting) */
3091 #if KMP_USE_MONITOR
3092 extern int
3093  __kmp_monitor_wakeups; /* number of times monitor wakes up per second */
3094 extern int __kmp_bt_intervals; /* number of monitor timestamp intervals before
3095  blocking */
3096 #endif
3097 #ifdef KMP_ADJUST_BLOCKTIME
3098 extern int __kmp_zero_bt; /* whether blocktime has been forced to zero */
3099 #endif /* KMP_ADJUST_BLOCKTIME */
3100 #ifdef KMP_DFLT_NTH_CORES
3101 extern int __kmp_ncores; /* Total number of cores for threads placement */
3102 #endif
3103 /* Number of millisecs to delay on abort for Intel(R) VTune(TM) tools */
3104 extern int __kmp_abort_delay;
3105 
3106 extern int __kmp_need_register_atfork_specified;
3107 extern int __kmp_need_register_atfork; /* At initialization, call pthread_atfork
3108  to install fork handler */
3109 extern int __kmp_gtid_mode; /* Method of getting gtid, values:
3110  0 - not set, will be set at runtime
3111  1 - using stack search
3112  2 - dynamic TLS (pthread_getspecific(Linux* OS/OS
3113  X*) or TlsGetValue(Windows* OS))
3114  3 - static TLS (__declspec(thread) __kmp_gtid),
3115  Linux* OS .so only. */
3116 extern int
3117  __kmp_adjust_gtid_mode; /* If true, adjust method based on #threads */
3118 #ifdef KMP_TDATA_GTID
3119 extern KMP_THREAD_LOCAL int __kmp_gtid;
3120 #endif
3121 extern int __kmp_tls_gtid_min; /* #threads below which use sp search for gtid */
3122 extern int __kmp_foreign_tp; // If true, separate TP var for each foreign thread
3123 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
3124 extern int __kmp_inherit_fp_control; // copy fp creg(s) parent->workers at fork
3125 extern kmp_int16 __kmp_init_x87_fpu_control_word; // init thread's FP ctrl reg
3126 extern kmp_uint32 __kmp_init_mxcsr; /* init thread's mxscr */
3127 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
3128 
3129 // max_active_levels for nested parallelism enabled by default via
3130 // OMP_MAX_ACTIVE_LEVELS, OMP_NESTED, OMP_NUM_THREADS, and OMP_PROC_BIND
3131 extern int __kmp_dflt_max_active_levels;
3132 // Indicates whether value of __kmp_dflt_max_active_levels was already
3133 // explicitly set by OMP_MAX_ACTIVE_LEVELS or OMP_NESTED=false
3134 extern bool __kmp_dflt_max_active_levels_set;
3135 extern int __kmp_dispatch_num_buffers; /* max possible dynamic loops in
3136  concurrent execution per team */
3137 #if KMP_NESTED_HOT_TEAMS
3138 extern int __kmp_hot_teams_mode;
3139 extern int __kmp_hot_teams_max_level;
3140 #endif
3141 
3142 #if KMP_OS_LINUX
3143 extern enum clock_function_type __kmp_clock_function;
3144 extern int __kmp_clock_function_param;
3145 #endif /* KMP_OS_LINUX */
3146 
3147 #if KMP_MIC_SUPPORTED
3148 extern enum mic_type __kmp_mic_type;
3149 #endif
3150 
3151 #ifdef USE_LOAD_BALANCE
3152 extern double __kmp_load_balance_interval; // load balance algorithm interval
3153 #endif /* USE_LOAD_BALANCE */
3154 
3155 // OpenMP 3.1 - Nested num threads array
3156 typedef struct kmp_nested_nthreads_t {
3157  int *nth;
3158  int size;
3159  int used;
3160 } kmp_nested_nthreads_t;
3161 
3162 extern kmp_nested_nthreads_t __kmp_nested_nth;
3163 
3164 #if KMP_USE_ADAPTIVE_LOCKS
3165 
3166 // Parameters for the speculative lock backoff system.
3167 struct kmp_adaptive_backoff_params_t {
3168  // Number of soft retries before it counts as a hard retry.
3169  kmp_uint32 max_soft_retries;
3170  // Badness is a bit mask : 0,1,3,7,15,... on each hard failure we move one to
3171  // the right
3172  kmp_uint32 max_badness;
3173 };
3174 
3175 extern kmp_adaptive_backoff_params_t __kmp_adaptive_backoff_params;
3176 
3177 #if KMP_DEBUG_ADAPTIVE_LOCKS
3178 extern const char *__kmp_speculative_statsfile;
3179 #endif
3180 
3181 #endif // KMP_USE_ADAPTIVE_LOCKS
3182 
3183 extern int __kmp_display_env; /* TRUE or FALSE */
3184 extern int __kmp_display_env_verbose; /* TRUE if OMP_DISPLAY_ENV=VERBOSE */
3185 extern int __kmp_omp_cancellation; /* TRUE or FALSE */
3186 extern int __kmp_nteams;
3187 extern int __kmp_teams_thread_limit;
3188 
3189 /* ------------------------------------------------------------------------- */
3190 
3191 /* the following are protected by the fork/join lock */
3192 /* write: lock read: anytime */
3193 extern kmp_info_t **__kmp_threads; /* Descriptors for the threads */
3194 /* read/write: lock */
3195 extern volatile kmp_team_t *__kmp_team_pool;
3196 extern volatile kmp_info_t *__kmp_thread_pool;
3197 extern kmp_info_t *__kmp_thread_pool_insert_pt;
3198 
3199 // total num threads reachable from some root thread including all root threads
3200 extern volatile int __kmp_nth;
3201 /* total number of threads reachable from some root thread including all root
3202  threads, and those in the thread pool */
3203 extern volatile int __kmp_all_nth;
3204 extern std::atomic<int> __kmp_thread_pool_active_nth;
3205 
3206 extern kmp_root_t **__kmp_root; /* root of thread hierarchy */
3207 /* end data protected by fork/join lock */
3208 /* ------------------------------------------------------------------------- */
3209 
3210 #define __kmp_get_gtid() __kmp_get_global_thread_id()
3211 #define __kmp_entry_gtid() __kmp_get_global_thread_id_reg()
3212 #define __kmp_get_tid() (__kmp_tid_from_gtid(__kmp_get_gtid()))
3213 #define __kmp_get_team() (__kmp_threads[(__kmp_get_gtid())]->th.th_team)
3214 #define __kmp_get_thread() (__kmp_thread_from_gtid(__kmp_get_gtid()))
3215 
3216 // AT: Which way is correct?
3217 // AT: 1. nproc = __kmp_threads[ ( gtid ) ] -> th.th_team -> t.t_nproc;
3218 // AT: 2. nproc = __kmp_threads[ ( gtid ) ] -> th.th_team_nproc;
3219 #define __kmp_get_team_num_threads(gtid) \
3220  (__kmp_threads[(gtid)]->th.th_team->t.t_nproc)
3221 
3222 static inline bool KMP_UBER_GTID(int gtid) {
3223  KMP_DEBUG_ASSERT(gtid >= KMP_GTID_MIN);
3224  KMP_DEBUG_ASSERT(gtid < __kmp_threads_capacity);
3225  return (gtid >= 0 && __kmp_root[gtid] && __kmp_threads[gtid] &&
3226  __kmp_threads[gtid] == __kmp_root[gtid]->r.r_uber_thread);
3227 }
3228 
3229 static inline int __kmp_tid_from_gtid(int gtid) {
3230  KMP_DEBUG_ASSERT(gtid >= 0);
3231  return __kmp_threads[gtid]->th.th_info.ds.ds_tid;
3232 }
3233 
3234 static inline int __kmp_gtid_from_tid(int tid, const kmp_team_t *team) {
3235  KMP_DEBUG_ASSERT(tid >= 0 && team);
3236  return team->t.t_threads[tid]->th.th_info.ds.ds_gtid;
3237 }
3238 
3239 static inline int __kmp_gtid_from_thread(const kmp_info_t *thr) {
3240  KMP_DEBUG_ASSERT(thr);
3241  return thr->th.th_info.ds.ds_gtid;
3242 }
3243 
3244 static inline kmp_info_t *__kmp_thread_from_gtid(int gtid) {
3245  KMP_DEBUG_ASSERT(gtid >= 0);
3246  return __kmp_threads[gtid];
3247 }
3248 
3249 static inline kmp_team_t *__kmp_team_from_gtid(int gtid) {
3250  KMP_DEBUG_ASSERT(gtid >= 0);
3251  return __kmp_threads[gtid]->th.th_team;
3252 }
3253 
3254 static inline void __kmp_assert_valid_gtid(kmp_int32 gtid) {
3255  if (UNLIKELY(gtid < 0 || gtid >= __kmp_threads_capacity))
3256  KMP_FATAL(ThreadIdentInvalid);
3257 }
3258 
3259 #if KMP_HAVE_MWAIT || KMP_HAVE_UMWAIT
3260 extern int __kmp_user_level_mwait; // TRUE or FALSE; from KMP_USER_LEVEL_MWAIT
3261 extern int __kmp_umwait_enabled; // Runtime check if user-level mwait enabled
3262 extern int __kmp_mwait_enabled; // Runtime check if ring3 mwait is enabled
3263 extern int __kmp_mwait_hints; // Hints to pass in to mwait
3264 #endif
3265 
3266 /* ------------------------------------------------------------------------- */
3267 
3268 extern kmp_global_t __kmp_global; /* global status */
3269 
3270 extern kmp_info_t __kmp_monitor;
3271 // For Debugging Support Library
3272 extern std::atomic<kmp_int32> __kmp_team_counter;
3273 // For Debugging Support Library
3274 extern std::atomic<kmp_int32> __kmp_task_counter;
3275 
3276 #if USE_DEBUGGER
3277 #define _KMP_GEN_ID(counter) \
3278  (__kmp_debugging ? KMP_ATOMIC_INC(&counter) + 1 : ~0)
3279 #else
3280 #define _KMP_GEN_ID(counter) (~0)
3281 #endif /* USE_DEBUGGER */
3282 
3283 #define KMP_GEN_TASK_ID() _KMP_GEN_ID(__kmp_task_counter)
3284 #define KMP_GEN_TEAM_ID() _KMP_GEN_ID(__kmp_team_counter)
3285 
3286 /* ------------------------------------------------------------------------ */
3287 
3288 extern void __kmp_print_storage_map_gtid(int gtid, void *p1, void *p2,
3289  size_t size, char const *format, ...);
3290 
3291 extern void __kmp_serial_initialize(void);
3292 extern void __kmp_middle_initialize(void);
3293 extern void __kmp_parallel_initialize(void);
3294 
3295 extern void __kmp_internal_begin(void);
3296 extern void __kmp_internal_end_library(int gtid);
3297 extern void __kmp_internal_end_thread(int gtid);
3298 extern void __kmp_internal_end_atexit(void);
3299 extern void __kmp_internal_end_dtor(void);
3300 extern void __kmp_internal_end_dest(void *);
3301 
3302 extern int __kmp_register_root(int initial_thread);
3303 extern void __kmp_unregister_root(int gtid);
3304 extern void __kmp_unregister_library(void); // called by __kmp_internal_end()
3305 
3306 extern int __kmp_ignore_mppbeg(void);
3307 extern int __kmp_ignore_mppend(void);
3308 
3309 extern int __kmp_enter_single(int gtid, ident_t *id_ref, int push_ws);
3310 extern void __kmp_exit_single(int gtid);
3311 
3312 extern void __kmp_parallel_deo(int *gtid_ref, int *cid_ref, ident_t *loc_ref);
3313 extern void __kmp_parallel_dxo(int *gtid_ref, int *cid_ref, ident_t *loc_ref);
3314 
3315 #ifdef USE_LOAD_BALANCE
3316 extern int __kmp_get_load_balance(int);
3317 #endif
3318 
3319 extern int __kmp_get_global_thread_id(void);
3320 extern int __kmp_get_global_thread_id_reg(void);
3321 extern void __kmp_exit_thread(int exit_status);
3322 extern void __kmp_abort(char const *format, ...);
3323 extern void __kmp_abort_thread(void);
3324 KMP_NORETURN extern void __kmp_abort_process(void);
3325 extern void __kmp_warn(char const *format, ...);
3326 
3327 extern void __kmp_set_num_threads(int new_nth, int gtid);
3328 
3329 // Returns current thread (pointer to kmp_info_t). Current thread *must* be
3330 // registered.
3331 static inline kmp_info_t *__kmp_entry_thread() {
3332  int gtid = __kmp_entry_gtid();
3333 
3334  return __kmp_threads[gtid];
3335 }
3336 
3337 extern void __kmp_set_max_active_levels(int gtid, int new_max_active_levels);
3338 extern int __kmp_get_max_active_levels(int gtid);
3339 extern int __kmp_get_ancestor_thread_num(int gtid, int level);
3340 extern int __kmp_get_team_size(int gtid, int level);
3341 extern void __kmp_set_schedule(int gtid, kmp_sched_t new_sched, int chunk);
3342 extern void __kmp_get_schedule(int gtid, kmp_sched_t *sched, int *chunk);
3343 
3344 extern unsigned short __kmp_get_random(kmp_info_t *thread);
3345 extern void __kmp_init_random(kmp_info_t *thread);
3346 
3347 extern kmp_r_sched_t __kmp_get_schedule_global(void);
3348 extern void __kmp_adjust_num_threads(int new_nproc);
3349 extern void __kmp_check_stksize(size_t *val);
3350 
3351 extern void *___kmp_allocate(size_t size KMP_SRC_LOC_DECL);
3352 extern void *___kmp_page_allocate(size_t size KMP_SRC_LOC_DECL);
3353 extern void ___kmp_free(void *ptr KMP_SRC_LOC_DECL);
3354 #define __kmp_allocate(size) ___kmp_allocate((size)KMP_SRC_LOC_CURR)
3355 #define __kmp_page_allocate(size) ___kmp_page_allocate((size)KMP_SRC_LOC_CURR)
3356 #define __kmp_free(ptr) ___kmp_free((ptr)KMP_SRC_LOC_CURR)
3357 
3358 #if USE_FAST_MEMORY
3359 extern void *___kmp_fast_allocate(kmp_info_t *this_thr,
3360  size_t size KMP_SRC_LOC_DECL);
3361 extern void ___kmp_fast_free(kmp_info_t *this_thr, void *ptr KMP_SRC_LOC_DECL);
3362 extern void __kmp_free_fast_memory(kmp_info_t *this_thr);
3363 extern void __kmp_initialize_fast_memory(kmp_info_t *this_thr);
3364 #define __kmp_fast_allocate(this_thr, size) \
3365  ___kmp_fast_allocate((this_thr), (size)KMP_SRC_LOC_CURR)
3366 #define __kmp_fast_free(this_thr, ptr) \
3367  ___kmp_fast_free((this_thr), (ptr)KMP_SRC_LOC_CURR)
3368 #endif
3369 
3370 extern void *___kmp_thread_malloc(kmp_info_t *th, size_t size KMP_SRC_LOC_DECL);
3371 extern void *___kmp_thread_calloc(kmp_info_t *th, size_t nelem,
3372  size_t elsize KMP_SRC_LOC_DECL);
3373 extern void *___kmp_thread_realloc(kmp_info_t *th, void *ptr,
3374  size_t size KMP_SRC_LOC_DECL);
3375 extern void ___kmp_thread_free(kmp_info_t *th, void *ptr KMP_SRC_LOC_DECL);
3376 #define __kmp_thread_malloc(th, size) \
3377  ___kmp_thread_malloc((th), (size)KMP_SRC_LOC_CURR)
3378 #define __kmp_thread_calloc(th, nelem, elsize) \
3379  ___kmp_thread_calloc((th), (nelem), (elsize)KMP_SRC_LOC_CURR)
3380 #define __kmp_thread_realloc(th, ptr, size) \
3381  ___kmp_thread_realloc((th), (ptr), (size)KMP_SRC_LOC_CURR)
3382 #define __kmp_thread_free(th, ptr) \
3383  ___kmp_thread_free((th), (ptr)KMP_SRC_LOC_CURR)
3384 
3385 #define KMP_INTERNAL_MALLOC(sz) malloc(sz)
3386 #define KMP_INTERNAL_FREE(p) free(p)
3387 #define KMP_INTERNAL_REALLOC(p, sz) realloc((p), (sz))
3388 #define KMP_INTERNAL_CALLOC(n, sz) calloc((n), (sz))
3389 
3390 extern void __kmp_push_num_threads(ident_t *loc, int gtid, int num_threads);
3391 
3392 extern void __kmp_push_proc_bind(ident_t *loc, int gtid,
3393  kmp_proc_bind_t proc_bind);
3394 extern void __kmp_push_num_teams(ident_t *loc, int gtid, int num_teams,
3395  int num_threads);
3396 extern void __kmp_push_num_teams_51(ident_t *loc, int gtid, int num_teams_lb,
3397  int num_teams_ub, int num_threads);
3398 
3399 extern void __kmp_yield();
3400 
3401 extern void __kmpc_dispatch_init_4(ident_t *loc, kmp_int32 gtid,
3402  enum sched_type schedule, kmp_int32 lb,
3403  kmp_int32 ub, kmp_int32 st, kmp_int32 chunk);
3404 extern void __kmpc_dispatch_init_4u(ident_t *loc, kmp_int32 gtid,
3405  enum sched_type schedule, kmp_uint32 lb,
3406  kmp_uint32 ub, kmp_int32 st,
3407  kmp_int32 chunk);
3408 extern void __kmpc_dispatch_init_8(ident_t *loc, kmp_int32 gtid,
3409  enum sched_type schedule, kmp_int64 lb,
3410  kmp_int64 ub, kmp_int64 st, kmp_int64 chunk);
3411 extern void __kmpc_dispatch_init_8u(ident_t *loc, kmp_int32 gtid,
3412  enum sched_type schedule, kmp_uint64 lb,
3413  kmp_uint64 ub, kmp_int64 st,
3414  kmp_int64 chunk);
3415 
3416 extern int __kmpc_dispatch_next_4(ident_t *loc, kmp_int32 gtid,
3417  kmp_int32 *p_last, kmp_int32 *p_lb,
3418  kmp_int32 *p_ub, kmp_int32 *p_st);
3419 extern int __kmpc_dispatch_next_4u(ident_t *loc, kmp_int32 gtid,
3420  kmp_int32 *p_last, kmp_uint32 *p_lb,
3421  kmp_uint32 *p_ub, kmp_int32 *p_st);
3422 extern int __kmpc_dispatch_next_8(ident_t *loc, kmp_int32 gtid,
3423  kmp_int32 *p_last, kmp_int64 *p_lb,
3424  kmp_int64 *p_ub, kmp_int64 *p_st);
3425 extern int __kmpc_dispatch_next_8u(ident_t *loc, kmp_int32 gtid,
3426  kmp_int32 *p_last, kmp_uint64 *p_lb,
3427  kmp_uint64 *p_ub, kmp_int64 *p_st);
3428 
3429 extern void __kmpc_dispatch_fini_4(ident_t *loc, kmp_int32 gtid);
3430 extern void __kmpc_dispatch_fini_8(ident_t *loc, kmp_int32 gtid);
3431 extern void __kmpc_dispatch_fini_4u(ident_t *loc, kmp_int32 gtid);
3432 extern void __kmpc_dispatch_fini_8u(ident_t *loc, kmp_int32 gtid);
3433 
3434 #ifdef KMP_GOMP_COMPAT
3435 
3436 extern void __kmp_aux_dispatch_init_4(ident_t *loc, kmp_int32 gtid,
3437  enum sched_type schedule, kmp_int32 lb,
3438  kmp_int32 ub, kmp_int32 st,
3439  kmp_int32 chunk, int push_ws);
3440 extern void __kmp_aux_dispatch_init_4u(ident_t *loc, kmp_int32 gtid,
3441  enum sched_type schedule, kmp_uint32 lb,
3442  kmp_uint32 ub, kmp_int32 st,
3443  kmp_int32 chunk, int push_ws);
3444 extern void __kmp_aux_dispatch_init_8(ident_t *loc, kmp_int32 gtid,
3445  enum sched_type schedule, kmp_int64 lb,
3446  kmp_int64 ub, kmp_int64 st,
3447  kmp_int64 chunk, int push_ws);
3448 extern void __kmp_aux_dispatch_init_8u(ident_t *loc, kmp_int32 gtid,
3449  enum sched_type schedule, kmp_uint64 lb,
3450  kmp_uint64 ub, kmp_int64 st,
3451  kmp_int64 chunk, int push_ws);
3452 extern void __kmp_aux_dispatch_fini_chunk_4(ident_t *loc, kmp_int32 gtid);
3453 extern void __kmp_aux_dispatch_fini_chunk_8(ident_t *loc, kmp_int32 gtid);
3454 extern void __kmp_aux_dispatch_fini_chunk_4u(ident_t *loc, kmp_int32 gtid);
3455 extern void __kmp_aux_dispatch_fini_chunk_8u(ident_t *loc, kmp_int32 gtid);
3456 
3457 #endif /* KMP_GOMP_COMPAT */
3458 
3459 extern kmp_uint32 __kmp_eq_4(kmp_uint32 value, kmp_uint32 checker);
3460 extern kmp_uint32 __kmp_neq_4(kmp_uint32 value, kmp_uint32 checker);
3461 extern kmp_uint32 __kmp_lt_4(kmp_uint32 value, kmp_uint32 checker);
3462 extern kmp_uint32 __kmp_ge_4(kmp_uint32 value, kmp_uint32 checker);
3463 extern kmp_uint32 __kmp_le_4(kmp_uint32 value, kmp_uint32 checker);
3464 extern kmp_uint32 __kmp_wait_4(kmp_uint32 volatile *spinner, kmp_uint32 checker,
3465  kmp_uint32 (*pred)(kmp_uint32, kmp_uint32),
3466  void *obj);
3467 extern void __kmp_wait_4_ptr(void *spinner, kmp_uint32 checker,
3468  kmp_uint32 (*pred)(void *, kmp_uint32), void *obj);
3469 
3470 extern void __kmp_wait_64(kmp_info_t *this_thr, kmp_flag_64<> *flag,
3471  int final_spin
3472 #if USE_ITT_BUILD
3473  ,
3474  void *itt_sync_obj
3475 #endif
3476 );
3477 extern void __kmp_release_64(kmp_flag_64<> *flag);
3478 
3479 extern void __kmp_infinite_loop(void);
3480 
3481 extern void __kmp_cleanup(void);
3482 
3483 #if KMP_HANDLE_SIGNALS
3484 extern int __kmp_handle_signals;
3485 extern void __kmp_install_signals(int parallel_init);
3486 extern void __kmp_remove_signals(void);
3487 #endif
3488 
3489 extern void __kmp_clear_system_time(void);
3490 extern void __kmp_read_system_time(double *delta);
3491 
3492 extern void __kmp_check_stack_overlap(kmp_info_t *thr);
3493 
3494 extern void __kmp_expand_host_name(char *buffer, size_t size);
3495 extern void __kmp_expand_file_name(char *result, size_t rlen, char *pattern);
3496 
3497 #if KMP_ARCH_X86 || KMP_ARCH_X86_64 || (KMP_OS_WINDOWS && KMP_ARCH_AARCH64)
3498 extern void
3499 __kmp_initialize_system_tick(void); /* Initialize timer tick value */
3500 #endif
3501 
3502 extern void
3503 __kmp_runtime_initialize(void); /* machine specific initialization */
3504 extern void __kmp_runtime_destroy(void);
3505 
3506 #if KMP_AFFINITY_SUPPORTED
3507 extern char *__kmp_affinity_print_mask(char *buf, int buf_len,
3508  kmp_affin_mask_t *mask);
3509 extern kmp_str_buf_t *__kmp_affinity_str_buf_mask(kmp_str_buf_t *buf,
3510  kmp_affin_mask_t *mask);
3511 extern void __kmp_affinity_initialize(void);
3512 extern void __kmp_affinity_uninitialize(void);
3513 extern void __kmp_affinity_set_init_mask(
3514  int gtid, int isa_root); /* set affinity according to KMP_AFFINITY */
3515 extern void __kmp_affinity_set_place(int gtid);
3516 extern void __kmp_affinity_determine_capable(const char *env_var);
3517 extern int __kmp_aux_set_affinity(void **mask);
3518 extern int __kmp_aux_get_affinity(void **mask);
3519 extern int __kmp_aux_get_affinity_max_proc();
3520 extern int __kmp_aux_set_affinity_mask_proc(int proc, void **mask);
3521 extern int __kmp_aux_unset_affinity_mask_proc(int proc, void **mask);
3522 extern int __kmp_aux_get_affinity_mask_proc(int proc, void **mask);
3523 extern void __kmp_balanced_affinity(kmp_info_t *th, int team_size);
3524 #if KMP_OS_LINUX || KMP_OS_FREEBSD
3525 extern int kmp_set_thread_affinity_mask_initial(void);
3526 #endif
3527 static inline void __kmp_assign_root_init_mask() {
3528  int gtid = __kmp_entry_gtid();
3529  kmp_root_t *r = __kmp_threads[gtid]->th.th_root;
3530  if (r->r.r_uber_thread == __kmp_threads[gtid] && !r->r.r_affinity_assigned) {
3531  __kmp_affinity_set_init_mask(gtid, TRUE);
3532  r->r.r_affinity_assigned = TRUE;
3533  }
3534 }
3535 #else /* KMP_AFFINITY_SUPPORTED */
3536 #define __kmp_assign_root_init_mask() /* Nothing */
3537 #endif /* KMP_AFFINITY_SUPPORTED */
3538 // No need for KMP_AFFINITY_SUPPORTED guard as only one field in the
3539 // format string is for affinity, so platforms that do not support
3540 // affinity can still use the other fields, e.g., %n for num_threads
3541 extern size_t __kmp_aux_capture_affinity(int gtid, const char *format,
3542  kmp_str_buf_t *buffer);
3543 extern void __kmp_aux_display_affinity(int gtid, const char *format);
3544 
3545 extern void __kmp_cleanup_hierarchy();
3546 extern void __kmp_get_hierarchy(kmp_uint32 nproc, kmp_bstate_t *thr_bar);
3547 
3548 #if KMP_USE_FUTEX
3549 
3550 extern int __kmp_futex_determine_capable(void);
3551 
3552 #endif // KMP_USE_FUTEX
3553 
3554 extern void __kmp_gtid_set_specific(int gtid);
3555 extern int __kmp_gtid_get_specific(void);
3556 
3557 extern double __kmp_read_cpu_time(void);
3558 
3559 extern int __kmp_read_system_info(struct kmp_sys_info *info);
3560 
3561 #if KMP_USE_MONITOR
3562 extern void __kmp_create_monitor(kmp_info_t *th);
3563 #endif
3564 
3565 extern void *__kmp_launch_thread(kmp_info_t *thr);
3566 
3567 extern void __kmp_create_worker(int gtid, kmp_info_t *th, size_t stack_size);
3568 
3569 #if KMP_OS_WINDOWS
3570 extern int __kmp_still_running(kmp_info_t *th);
3571 extern int __kmp_is_thread_alive(kmp_info_t *th, DWORD *exit_val);
3572 extern void __kmp_free_handle(kmp_thread_t tHandle);
3573 #endif
3574 
3575 #if KMP_USE_MONITOR
3576 extern void __kmp_reap_monitor(kmp_info_t *th);
3577 #endif
3578 extern void __kmp_reap_worker(kmp_info_t *th);
3579 extern void __kmp_terminate_thread(int gtid);
3580 
3581 extern int __kmp_try_suspend_mx(kmp_info_t *th);
3582 extern void __kmp_lock_suspend_mx(kmp_info_t *th);
3583 extern void __kmp_unlock_suspend_mx(kmp_info_t *th);
3584 
3585 extern void __kmp_elapsed(double *);
3586 extern void __kmp_elapsed_tick(double *);
3587 
3588 extern void __kmp_enable(int old_state);
3589 extern void __kmp_disable(int *old_state);
3590 
3591 extern void __kmp_thread_sleep(int millis);
3592 
3593 extern void __kmp_common_initialize(void);
3594 extern void __kmp_common_destroy(void);
3595 extern void __kmp_common_destroy_gtid(int gtid);
3596 
3597 #if KMP_OS_UNIX
3598 extern void __kmp_register_atfork(void);
3599 #endif
3600 extern void __kmp_suspend_initialize(void);
3601 extern void __kmp_suspend_initialize_thread(kmp_info_t *th);
3602 extern void __kmp_suspend_uninitialize_thread(kmp_info_t *th);
3603 
3604 extern kmp_info_t *__kmp_allocate_thread(kmp_root_t *root, kmp_team_t *team,
3605  int tid);
3606 extern kmp_team_t *
3607 __kmp_allocate_team(kmp_root_t *root, int new_nproc, int max_nproc,
3608 #if OMPT_SUPPORT
3609  ompt_data_t ompt_parallel_data,
3610 #endif
3611  kmp_proc_bind_t proc_bind, kmp_internal_control_t *new_icvs,
3612  int argc USE_NESTED_HOT_ARG(kmp_info_t *thr));
3613 extern void __kmp_free_thread(kmp_info_t *);
3614 extern void __kmp_free_team(kmp_root_t *,
3615  kmp_team_t *USE_NESTED_HOT_ARG(kmp_info_t *));
3616 extern kmp_team_t *__kmp_reap_team(kmp_team_t *);
3617 
3618 /* ------------------------------------------------------------------------ */
3619 
3620 extern void __kmp_initialize_bget(kmp_info_t *th);
3621 extern void __kmp_finalize_bget(kmp_info_t *th);
3622 
3623 KMP_EXPORT void *kmpc_malloc(size_t size);
3624 KMP_EXPORT void *kmpc_aligned_malloc(size_t size, size_t alignment);
3625 KMP_EXPORT void *kmpc_calloc(size_t nelem, size_t elsize);
3626 KMP_EXPORT void *kmpc_realloc(void *ptr, size_t size);
3627 KMP_EXPORT void kmpc_free(void *ptr);
3628 
3629 /* declarations for internal use */
3630 
3631 extern int __kmp_barrier(enum barrier_type bt, int gtid, int is_split,
3632  size_t reduce_size, void *reduce_data,
3633  void (*reduce)(void *, void *));
3634 extern void __kmp_end_split_barrier(enum barrier_type bt, int gtid);
3635 extern int __kmp_barrier_gomp_cancel(int gtid);
3636 
3641 enum fork_context_e {
3642  fork_context_gnu,
3644  fork_context_intel,
3645  fork_context_last
3646 };
3647 extern int __kmp_fork_call(ident_t *loc, int gtid,
3648  enum fork_context_e fork_context, kmp_int32 argc,
3649  microtask_t microtask, launch_t invoker,
3650  kmp_va_list ap);
3651 
3652 extern void __kmp_join_call(ident_t *loc, int gtid
3653 #if OMPT_SUPPORT
3654  ,
3655  enum fork_context_e fork_context
3656 #endif
3657  ,
3658  int exit_teams = 0);
3659 
3660 extern void __kmp_serialized_parallel(ident_t *id, kmp_int32 gtid);
3661 extern void __kmp_internal_fork(ident_t *id, int gtid, kmp_team_t *team);
3662 extern void __kmp_internal_join(ident_t *id, int gtid, kmp_team_t *team);
3663 extern int __kmp_invoke_task_func(int gtid);
3664 extern void __kmp_run_before_invoked_task(int gtid, int tid,
3665  kmp_info_t *this_thr,
3666  kmp_team_t *team);
3667 extern void __kmp_run_after_invoked_task(int gtid, int tid,
3668  kmp_info_t *this_thr,
3669  kmp_team_t *team);
3670 
3671 // should never have been exported
3672 KMP_EXPORT int __kmpc_invoke_task_func(int gtid);
3673 extern int __kmp_invoke_teams_master(int gtid);
3674 extern void __kmp_teams_master(int gtid);
3675 extern int __kmp_aux_get_team_num();
3676 extern int __kmp_aux_get_num_teams();
3677 extern void __kmp_save_internal_controls(kmp_info_t *thread);
3678 extern void __kmp_user_set_library(enum library_type arg);
3679 extern void __kmp_aux_set_library(enum library_type arg);
3680 extern void __kmp_aux_set_stacksize(size_t arg);
3681 extern void __kmp_aux_set_blocktime(int arg, kmp_info_t *thread, int tid);
3682 extern void __kmp_aux_set_defaults(char const *str, size_t len);
3683 
3684 /* Functions called from __kmp_aux_env_initialize() in kmp_settings.cpp */
3685 void kmpc_set_blocktime(int arg);
3686 void ompc_set_nested(int flag);
3687 void ompc_set_dynamic(int flag);
3688 void ompc_set_num_threads(int arg);
3689 
3690 extern void __kmp_push_current_task_to_thread(kmp_info_t *this_thr,
3691  kmp_team_t *team, int tid);
3692 extern void __kmp_pop_current_task_from_thread(kmp_info_t *this_thr);
3693 extern kmp_task_t *__kmp_task_alloc(ident_t *loc_ref, kmp_int32 gtid,
3694  kmp_tasking_flags_t *flags,
3695  size_t sizeof_kmp_task_t,
3696  size_t sizeof_shareds,
3697  kmp_routine_entry_t task_entry);
3698 extern void __kmp_init_implicit_task(ident_t *loc_ref, kmp_info_t *this_thr,
3699  kmp_team_t *team, int tid,
3700  int set_curr_task);
3701 extern void __kmp_finish_implicit_task(kmp_info_t *this_thr);
3702 extern void __kmp_free_implicit_task(kmp_info_t *this_thr);
3703 
3704 extern kmp_event_t *__kmpc_task_allow_completion_event(ident_t *loc_ref,
3705  int gtid,
3706  kmp_task_t *task);
3707 extern void __kmp_fulfill_event(kmp_event_t *event);
3708 
3709 extern void __kmp_free_task_team(kmp_info_t *thread,
3710  kmp_task_team_t *task_team);
3711 extern void __kmp_reap_task_teams(void);
3712 extern void __kmp_wait_to_unref_task_teams(void);
3713 extern void __kmp_task_team_setup(kmp_info_t *this_thr, kmp_team_t *team,
3714  int always);
3715 extern void __kmp_task_team_sync(kmp_info_t *this_thr, kmp_team_t *team);
3716 extern void __kmp_task_team_wait(kmp_info_t *this_thr, kmp_team_t *team
3717 #if USE_ITT_BUILD
3718  ,
3719  void *itt_sync_obj
3720 #endif /* USE_ITT_BUILD */
3721  ,
3722  int wait = 1);
3723 extern void __kmp_tasking_barrier(kmp_team_t *team, kmp_info_t *thread,
3724  int gtid);
3725 
3726 extern int __kmp_is_address_mapped(void *addr);
3727 extern kmp_uint64 __kmp_hardware_timestamp(void);
3728 
3729 #if KMP_OS_UNIX
3730 extern int __kmp_read_from_file(char const *path, char const *format, ...);
3731 #endif
3732 
3733 /* ------------------------------------------------------------------------ */
3734 //
3735 // Assembly routines that have no compiler intrinsic replacement
3736 //
3737 
3738 extern int __kmp_invoke_microtask(microtask_t pkfn, int gtid, int npr, int argc,
3739  void *argv[]
3740 #if OMPT_SUPPORT
3741  ,
3742  void **exit_frame_ptr
3743 #endif
3744 );
3745 
3746 /* ------------------------------------------------------------------------ */
3747 
3748 KMP_EXPORT void __kmpc_begin(ident_t *, kmp_int32 flags);
3749 KMP_EXPORT void __kmpc_end(ident_t *);
3750 
3751 KMP_EXPORT void __kmpc_threadprivate_register_vec(ident_t *, void *data,
3752  kmpc_ctor_vec ctor,
3753  kmpc_cctor_vec cctor,
3754  kmpc_dtor_vec dtor,
3755  size_t vector_length);
3756 KMP_EXPORT void __kmpc_threadprivate_register(ident_t *, void *data,
3757  kmpc_ctor ctor, kmpc_cctor cctor,
3758  kmpc_dtor dtor);
3759 KMP_EXPORT void *__kmpc_threadprivate(ident_t *, kmp_int32 global_tid,
3760  void *data, size_t size);
3761 
3762 KMP_EXPORT kmp_int32 __kmpc_global_thread_num(ident_t *);
3763 KMP_EXPORT kmp_int32 __kmpc_global_num_threads(ident_t *);
3764 KMP_EXPORT kmp_int32 __kmpc_bound_thread_num(ident_t *);
3765 KMP_EXPORT kmp_int32 __kmpc_bound_num_threads(ident_t *);
3766 
3767 KMP_EXPORT kmp_int32 __kmpc_ok_to_fork(ident_t *);
3768 KMP_EXPORT void __kmpc_fork_call(ident_t *, kmp_int32 nargs,
3769  kmpc_micro microtask, ...);
3770 
3771 KMP_EXPORT void __kmpc_serialized_parallel(ident_t *, kmp_int32 global_tid);
3772 KMP_EXPORT void __kmpc_end_serialized_parallel(ident_t *, kmp_int32 global_tid);
3773 
3774 KMP_EXPORT void __kmpc_flush(ident_t *);
3775 KMP_EXPORT void __kmpc_barrier(ident_t *, kmp_int32 global_tid);
3776 KMP_EXPORT kmp_int32 __kmpc_master(ident_t *, kmp_int32 global_tid);
3777 KMP_EXPORT void __kmpc_end_master(ident_t *, kmp_int32 global_tid);
3778 KMP_EXPORT kmp_int32 __kmpc_masked(ident_t *, kmp_int32 global_tid,
3779  kmp_int32 filter);
3780 KMP_EXPORT void __kmpc_end_masked(ident_t *, kmp_int32 global_tid);
3781 KMP_EXPORT void __kmpc_ordered(ident_t *, kmp_int32 global_tid);
3782 KMP_EXPORT void __kmpc_end_ordered(ident_t *, kmp_int32 global_tid);
3783 KMP_EXPORT void __kmpc_critical(ident_t *, kmp_int32 global_tid,
3784  kmp_critical_name *);
3785 KMP_EXPORT void __kmpc_end_critical(ident_t *, kmp_int32 global_tid,
3786  kmp_critical_name *);
3787 KMP_EXPORT void __kmpc_critical_with_hint(ident_t *, kmp_int32 global_tid,
3788  kmp_critical_name *, uint32_t hint);
3789 
3790 KMP_EXPORT kmp_int32 __kmpc_barrier_master(ident_t *, kmp_int32 global_tid);
3791 KMP_EXPORT void __kmpc_end_barrier_master(ident_t *, kmp_int32 global_tid);
3792 
3793 KMP_EXPORT kmp_int32 __kmpc_barrier_master_nowait(ident_t *,
3794  kmp_int32 global_tid);
3795 
3796 KMP_EXPORT kmp_int32 __kmpc_single(ident_t *, kmp_int32 global_tid);
3797 KMP_EXPORT void __kmpc_end_single(ident_t *, kmp_int32 global_tid);
3798 
3799 KMP_EXPORT void KMPC_FOR_STATIC_INIT(ident_t *loc, kmp_int32 global_tid,
3800  kmp_int32 schedtype, kmp_int32 *plastiter,
3801  kmp_int *plower, kmp_int *pupper,
3802  kmp_int *pstride, kmp_int incr,
3803  kmp_int chunk);
3804 
3805 KMP_EXPORT void __kmpc_for_static_fini(ident_t *loc, kmp_int32 global_tid);
3806 
3807 KMP_EXPORT void __kmpc_copyprivate(ident_t *loc, kmp_int32 global_tid,
3808  size_t cpy_size, void *cpy_data,
3809  void (*cpy_func)(void *, void *),
3810  kmp_int32 didit);
3811 
3812 extern void KMPC_SET_NUM_THREADS(int arg);
3813 extern void KMPC_SET_DYNAMIC(int flag);
3814 extern void KMPC_SET_NESTED(int flag);
3815 
3816 /* OMP 3.0 tasking interface routines */
3817 KMP_EXPORT kmp_int32 __kmpc_omp_task(ident_t *loc_ref, kmp_int32 gtid,
3818  kmp_task_t *new_task);
3819 KMP_EXPORT kmp_task_t *__kmpc_omp_task_alloc(ident_t *loc_ref, kmp_int32 gtid,
3820  kmp_int32 flags,
3821  size_t sizeof_kmp_task_t,
3822  size_t sizeof_shareds,
3823  kmp_routine_entry_t task_entry);
3824 KMP_EXPORT kmp_task_t *__kmpc_omp_target_task_alloc(
3825  ident_t *loc_ref, kmp_int32 gtid, kmp_int32 flags, size_t sizeof_kmp_task_t,
3826  size_t sizeof_shareds, kmp_routine_entry_t task_entry, kmp_int64 device_id);
3827 KMP_EXPORT void __kmpc_omp_task_begin_if0(ident_t *loc_ref, kmp_int32 gtid,
3828  kmp_task_t *task);
3829 KMP_EXPORT void __kmpc_omp_task_complete_if0(ident_t *loc_ref, kmp_int32 gtid,
3830  kmp_task_t *task);
3831 KMP_EXPORT kmp_int32 __kmpc_omp_task_parts(ident_t *loc_ref, kmp_int32 gtid,
3832  kmp_task_t *new_task);
3833 KMP_EXPORT kmp_int32 __kmpc_omp_taskwait(ident_t *loc_ref, kmp_int32 gtid);
3834 
3835 KMP_EXPORT kmp_int32 __kmpc_omp_taskyield(ident_t *loc_ref, kmp_int32 gtid,
3836  int end_part);
3837 
3838 #if TASK_UNUSED
3839 void __kmpc_omp_task_begin(ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *task);
3840 void __kmpc_omp_task_complete(ident_t *loc_ref, kmp_int32 gtid,
3841  kmp_task_t *task);
3842 #endif // TASK_UNUSED
3843 
3844 /* ------------------------------------------------------------------------ */
3845 
3846 KMP_EXPORT void __kmpc_taskgroup(ident_t *loc, int gtid);
3847 KMP_EXPORT void __kmpc_end_taskgroup(ident_t *loc, int gtid);
3848 
3849 KMP_EXPORT kmp_int32 __kmpc_omp_task_with_deps(
3850  ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 ndeps,
3851  kmp_depend_info_t *dep_list, kmp_int32 ndeps_noalias,
3852  kmp_depend_info_t *noalias_dep_list);
3853 KMP_EXPORT void __kmpc_omp_wait_deps(ident_t *loc_ref, kmp_int32 gtid,
3854  kmp_int32 ndeps,
3855  kmp_depend_info_t *dep_list,
3856  kmp_int32 ndeps_noalias,
3857  kmp_depend_info_t *noalias_dep_list);
3858 extern kmp_int32 __kmp_omp_task(kmp_int32 gtid, kmp_task_t *new_task,
3859  bool serialize_immediate);
3860 
3861 KMP_EXPORT kmp_int32 __kmpc_cancel(ident_t *loc_ref, kmp_int32 gtid,
3862  kmp_int32 cncl_kind);
3863 KMP_EXPORT kmp_int32 __kmpc_cancellationpoint(ident_t *loc_ref, kmp_int32 gtid,
3864  kmp_int32 cncl_kind);
3865 KMP_EXPORT kmp_int32 __kmpc_cancel_barrier(ident_t *loc_ref, kmp_int32 gtid);
3866 KMP_EXPORT int __kmp_get_cancellation_status(int cancel_kind);
3867 
3868 KMP_EXPORT void __kmpc_proxy_task_completed(kmp_int32 gtid, kmp_task_t *ptask);
3869 KMP_EXPORT void __kmpc_proxy_task_completed_ooo(kmp_task_t *ptask);
3870 KMP_EXPORT void __kmpc_taskloop(ident_t *loc, kmp_int32 gtid, kmp_task_t *task,
3871  kmp_int32 if_val, kmp_uint64 *lb,
3872  kmp_uint64 *ub, kmp_int64 st, kmp_int32 nogroup,
3873  kmp_int32 sched, kmp_uint64 grainsize,
3874  void *task_dup);
3875 KMP_EXPORT void __kmpc_taskloop_5(ident_t *loc, kmp_int32 gtid,
3876  kmp_task_t *task, kmp_int32 if_val,
3877  kmp_uint64 *lb, kmp_uint64 *ub, kmp_int64 st,
3878  kmp_int32 nogroup, kmp_int32 sched,
3879  kmp_uint64 grainsize, kmp_int32 modifier,
3880  void *task_dup);
3881 KMP_EXPORT void *__kmpc_task_reduction_init(int gtid, int num_data, void *data);
3882 KMP_EXPORT void *__kmpc_taskred_init(int gtid, int num_data, void *data);
3883 KMP_EXPORT void *__kmpc_task_reduction_get_th_data(int gtid, void *tg, void *d);
3884 KMP_EXPORT void *__kmpc_task_reduction_modifier_init(ident_t *loc, int gtid,
3885  int is_ws, int num,
3886  void *data);
3887 KMP_EXPORT void *__kmpc_taskred_modifier_init(ident_t *loc, int gtid, int is_ws,
3888  int num, void *data);
3889 KMP_EXPORT void __kmpc_task_reduction_modifier_fini(ident_t *loc, int gtid,
3890  int is_ws);
3891 KMP_EXPORT kmp_int32 __kmpc_omp_reg_task_with_affinity(
3892  ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 naffins,
3893  kmp_task_affinity_info_t *affin_list);
3894 KMP_EXPORT void __kmp_set_num_teams(int num_teams);
3895 KMP_EXPORT int __kmp_get_max_teams(void);
3896 KMP_EXPORT void __kmp_set_teams_thread_limit(int limit);
3897 KMP_EXPORT int __kmp_get_teams_thread_limit(void);
3898 
3899 /* Lock interface routines (fast versions with gtid passed in) */
3900 KMP_EXPORT void __kmpc_init_lock(ident_t *loc, kmp_int32 gtid,
3901  void **user_lock);
3902 KMP_EXPORT void __kmpc_init_nest_lock(ident_t *loc, kmp_int32 gtid,
3903  void **user_lock);
3904 KMP_EXPORT void __kmpc_destroy_lock(ident_t *loc, kmp_int32 gtid,
3905  void **user_lock);
3906 KMP_EXPORT void __kmpc_destroy_nest_lock(ident_t *loc, kmp_int32 gtid,
3907  void **user_lock);
3908 KMP_EXPORT void __kmpc_set_lock(ident_t *loc, kmp_int32 gtid, void **user_lock);
3909 KMP_EXPORT void __kmpc_set_nest_lock(ident_t *loc, kmp_int32 gtid,
3910  void **user_lock);
3911 KMP_EXPORT void __kmpc_unset_lock(ident_t *loc, kmp_int32 gtid,
3912  void **user_lock);
3913 KMP_EXPORT void __kmpc_unset_nest_lock(ident_t *loc, kmp_int32 gtid,
3914  void **user_lock);
3915 KMP_EXPORT int __kmpc_test_lock(ident_t *loc, kmp_int32 gtid, void **user_lock);
3916 KMP_EXPORT int __kmpc_test_nest_lock(ident_t *loc, kmp_int32 gtid,
3917  void **user_lock);
3918 
3919 KMP_EXPORT void __kmpc_init_lock_with_hint(ident_t *loc, kmp_int32 gtid,
3920  void **user_lock, uintptr_t hint);
3921 KMP_EXPORT void __kmpc_init_nest_lock_with_hint(ident_t *loc, kmp_int32 gtid,
3922  void **user_lock,
3923  uintptr_t hint);
3924 
3925 /* Interface to fast scalable reduce methods routines */
3926 
3927 KMP_EXPORT kmp_int32 __kmpc_reduce_nowait(
3928  ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
3929  void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
3930  kmp_critical_name *lck);
3931 KMP_EXPORT void __kmpc_end_reduce_nowait(ident_t *loc, kmp_int32 global_tid,
3932  kmp_critical_name *lck);
3933 KMP_EXPORT kmp_int32 __kmpc_reduce(
3934  ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
3935  void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
3936  kmp_critical_name *lck);
3937 KMP_EXPORT void __kmpc_end_reduce(ident_t *loc, kmp_int32 global_tid,
3938  kmp_critical_name *lck);
3939 
3940 /* Internal fast reduction routines */
3941 
3942 extern PACKED_REDUCTION_METHOD_T __kmp_determine_reduction_method(
3943  ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
3944  void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
3945  kmp_critical_name *lck);
3946 
3947 // this function is for testing set/get/determine reduce method
3948 KMP_EXPORT kmp_int32 __kmp_get_reduce_method(void);
3949 
3950 KMP_EXPORT kmp_uint64 __kmpc_get_taskid();
3951 KMP_EXPORT kmp_uint64 __kmpc_get_parent_taskid();
3952 
3953 // C++ port
3954 // missing 'extern "C"' declarations
3955 
3956 KMP_EXPORT kmp_int32 __kmpc_in_parallel(ident_t *loc);
3957 KMP_EXPORT void __kmpc_pop_num_threads(ident_t *loc, kmp_int32 global_tid);
3958 KMP_EXPORT void __kmpc_push_num_threads(ident_t *loc, kmp_int32 global_tid,
3959  kmp_int32 num_threads);
3960 
3961 KMP_EXPORT void __kmpc_push_proc_bind(ident_t *loc, kmp_int32 global_tid,
3962  int proc_bind);
3963 KMP_EXPORT void __kmpc_push_num_teams(ident_t *loc, kmp_int32 global_tid,
3964  kmp_int32 num_teams,
3965  kmp_int32 num_threads);
3966 /* Function for OpenMP 5.1 num_teams clause */
3967 KMP_EXPORT void __kmpc_push_num_teams_51(ident_t *loc, kmp_int32 global_tid,
3968  kmp_int32 num_teams_lb,
3969  kmp_int32 num_teams_ub,
3970  kmp_int32 num_threads);
3971 KMP_EXPORT void __kmpc_fork_teams(ident_t *loc, kmp_int32 argc,
3972  kmpc_micro microtask, ...);
3973 struct kmp_dim { // loop bounds info casted to kmp_int64
3974  kmp_int64 lo; // lower
3975  kmp_int64 up; // upper
3976  kmp_int64 st; // stride
3977 };
3978 KMP_EXPORT void __kmpc_doacross_init(ident_t *loc, kmp_int32 gtid,
3979  kmp_int32 num_dims,
3980  const struct kmp_dim *dims);
3981 KMP_EXPORT void __kmpc_doacross_wait(ident_t *loc, kmp_int32 gtid,
3982  const kmp_int64 *vec);
3983 KMP_EXPORT void __kmpc_doacross_post(ident_t *loc, kmp_int32 gtid,
3984  const kmp_int64 *vec);
3985 KMP_EXPORT void __kmpc_doacross_fini(ident_t *loc, kmp_int32 gtid);
3986 
3987 KMP_EXPORT void *__kmpc_threadprivate_cached(ident_t *loc, kmp_int32 global_tid,
3988  void *data, size_t size,
3989  void ***cache);
3990 
3991 // Symbols for MS mutual detection.
3992 extern int _You_must_link_with_exactly_one_OpenMP_library;
3993 extern int _You_must_link_with_Intel_OpenMP_library;
3994 #if KMP_OS_WINDOWS && (KMP_VERSION_MAJOR > 4)
3995 extern int _You_must_link_with_Microsoft_OpenMP_library;
3996 #endif
3997 
3998 // The routines below are not exported.
3999 // Consider making them 'static' in corresponding source files.
4000 void kmp_threadprivate_insert_private_data(int gtid, void *pc_addr,
4001  void *data_addr, size_t pc_size);
4002 struct private_common *kmp_threadprivate_insert(int gtid, void *pc_addr,
4003  void *data_addr,
4004  size_t pc_size);
4005 void __kmp_threadprivate_resize_cache(int newCapacity);
4006 void __kmp_cleanup_threadprivate_caches();
4007 
4008 // ompc_, kmpc_ entries moved from omp.h.
4009 #if KMP_OS_WINDOWS
4010 #define KMPC_CONVENTION __cdecl
4011 #else
4012 #define KMPC_CONVENTION
4013 #endif
4014 
4015 #ifndef __OMP_H
4016 typedef enum omp_sched_t {
4017  omp_sched_static = 1,
4018  omp_sched_dynamic = 2,
4019  omp_sched_guided = 3,
4020  omp_sched_auto = 4
4021 } omp_sched_t;
4022 typedef void *kmp_affinity_mask_t;
4023 #endif
4024 
4025 KMP_EXPORT void KMPC_CONVENTION ompc_set_max_active_levels(int);
4026 KMP_EXPORT void KMPC_CONVENTION ompc_set_schedule(omp_sched_t, int);
4027 KMP_EXPORT int KMPC_CONVENTION ompc_get_ancestor_thread_num(int);
4028 KMP_EXPORT int KMPC_CONVENTION ompc_get_team_size(int);
4029 KMP_EXPORT int KMPC_CONVENTION
4030 kmpc_set_affinity_mask_proc(int, kmp_affinity_mask_t *);
4031 KMP_EXPORT int KMPC_CONVENTION
4032 kmpc_unset_affinity_mask_proc(int, kmp_affinity_mask_t *);
4033 KMP_EXPORT int KMPC_CONVENTION
4034 kmpc_get_affinity_mask_proc(int, kmp_affinity_mask_t *);
4035 
4036 KMP_EXPORT void KMPC_CONVENTION kmpc_set_stacksize(int);
4037 KMP_EXPORT void KMPC_CONVENTION kmpc_set_stacksize_s(size_t);
4038 KMP_EXPORT void KMPC_CONVENTION kmpc_set_library(int);
4039 KMP_EXPORT void KMPC_CONVENTION kmpc_set_defaults(char const *);
4040 KMP_EXPORT void KMPC_CONVENTION kmpc_set_disp_num_buffers(int);
4041 void KMP_EXPAND_NAME(ompc_set_affinity_format)(char const *format);
4042 size_t KMP_EXPAND_NAME(ompc_get_affinity_format)(char *buffer, size_t size);
4043 void KMP_EXPAND_NAME(ompc_display_affinity)(char const *format);
4044 size_t KMP_EXPAND_NAME(ompc_capture_affinity)(char *buffer, size_t buf_size,
4045  char const *format);
4046 
4047 enum kmp_target_offload_kind {
4048  tgt_disabled = 0,
4049  tgt_default = 1,
4050  tgt_mandatory = 2
4051 };
4052 typedef enum kmp_target_offload_kind kmp_target_offload_kind_t;
4053 // Set via OMP_TARGET_OFFLOAD if specified, defaults to tgt_default otherwise
4054 extern kmp_target_offload_kind_t __kmp_target_offload;
4055 extern int __kmpc_get_target_offload();
4056 
4057 // Constants used in libomptarget
4058 #define KMP_DEVICE_DEFAULT -1 // This is libomptarget's default device.
4059 #define KMP_DEVICE_ALL -11 // This is libomptarget's "all devices".
4060 
4061 // OMP Pause Resource
4062 
4063 // The following enum is used both to set the status in __kmp_pause_status, and
4064 // as the internal equivalent of the externally-visible omp_pause_resource_t.
4065 typedef enum kmp_pause_status_t {
4066  kmp_not_paused = 0, // status is not paused, or, requesting resume
4067  kmp_soft_paused = 1, // status is soft-paused, or, requesting soft pause
4068  kmp_hard_paused = 2 // status is hard-paused, or, requesting hard pause
4069 } kmp_pause_status_t;
4070 
4071 // This stores the pause state of the runtime
4072 extern kmp_pause_status_t __kmp_pause_status;
4073 extern int __kmpc_pause_resource(kmp_pause_status_t level);
4074 extern int __kmp_pause_resource(kmp_pause_status_t level);
4075 // Soft resume sets __kmp_pause_status, and wakes up all threads.
4076 extern void __kmp_resume_if_soft_paused();
4077 // Hard resume simply resets the status to not paused. Library will appear to
4078 // be uninitialized after hard pause. Let OMP constructs trigger required
4079 // initializations.
4080 static inline void __kmp_resume_if_hard_paused() {
4081  if (__kmp_pause_status == kmp_hard_paused) {
4082  __kmp_pause_status = kmp_not_paused;
4083  }
4084 }
4085 
4086 extern void __kmp_omp_display_env(int verbose);
4087 
4088 // 1: it is initializing hidden helper team
4089 extern volatile int __kmp_init_hidden_helper;
4090 // 1: the hidden helper team is done
4091 extern volatile int __kmp_hidden_helper_team_done;
4092 // 1: enable hidden helper task
4093 extern kmp_int32 __kmp_enable_hidden_helper;
4094 // Main thread of hidden helper team
4095 extern kmp_info_t *__kmp_hidden_helper_main_thread;
4096 // Descriptors for the hidden helper threads
4097 extern kmp_info_t **__kmp_hidden_helper_threads;
4098 // Number of hidden helper threads
4099 extern kmp_int32 __kmp_hidden_helper_threads_num;
4100 // Number of hidden helper tasks that have not been executed yet
4101 extern std::atomic<kmp_int32> __kmp_unexecuted_hidden_helper_tasks;
4102 
4103 extern void __kmp_hidden_helper_initialize();
4104 extern void __kmp_hidden_helper_threads_initz_routine();
4105 extern void __kmp_do_initialize_hidden_helper_threads();
4106 extern void __kmp_hidden_helper_threads_initz_wait();
4107 extern void __kmp_hidden_helper_initz_release();
4108 extern void __kmp_hidden_helper_threads_deinitz_wait();
4109 extern void __kmp_hidden_helper_threads_deinitz_release();
4110 extern void __kmp_hidden_helper_main_thread_wait();
4111 extern void __kmp_hidden_helper_worker_thread_wait();
4112 extern void __kmp_hidden_helper_worker_thread_signal();
4113 extern void __kmp_hidden_helper_main_thread_release();
4114 
4115 // Check whether a given thread is a hidden helper thread
4116 #define KMP_HIDDEN_HELPER_THREAD(gtid) \
4117  ((gtid) >= 1 && (gtid) <= __kmp_hidden_helper_threads_num)
4118 
4119 #define KMP_HIDDEN_HELPER_WORKER_THREAD(gtid) \
4120  ((gtid) > 1 && (gtid) <= __kmp_hidden_helper_threads_num)
4121 
4122 #define KMP_HIDDEN_HELPER_TEAM(team) \
4123  (team->t.t_threads[0] == __kmp_hidden_helper_main_thread)
4124 
4125 // Map a gtid to a hidden helper thread. The first hidden helper thread, a.k.a
4126 // main thread, is skipped.
4127 #define KMP_GTID_TO_SHADOW_GTID(gtid) \
4128  ((gtid) % (__kmp_hidden_helper_threads_num - 1) + 2)
4129 
4130 // Return the adjusted gtid value by subtracting from gtid the number
4131 // of hidden helper threads. This adjusted value is the gtid the thread would
4132 // have received if there were no hidden helper threads.
4133 static inline int __kmp_adjust_gtid_for_hidden_helpers(int gtid) {
4134  int adjusted_gtid = gtid;
4135  if (__kmp_hidden_helper_threads_num > 0 && gtid > 0 &&
4136  gtid - __kmp_hidden_helper_threads_num >= 0) {
4137  adjusted_gtid -= __kmp_hidden_helper_threads_num;
4138  }
4139  return adjusted_gtid;
4140 }
4141 
4142 // Support for error directive
4143 typedef enum kmp_severity_t {
4144  severity_warning = 1,
4145  severity_fatal = 2
4146 } kmp_severity_t;
4147 extern void __kmpc_error(ident_t *loc, int severity, const char *message);
4148 
4149 #ifdef __cplusplus
4150 }
4151 #endif
4152 
4153 template <bool C, bool S>
4154 extern void __kmp_suspend_32(int th_gtid, kmp_flag_32<C, S> *flag);
4155 template <bool C, bool S>
4156 extern void __kmp_suspend_64(int th_gtid, kmp_flag_64<C, S> *flag);
4157 template <bool C, bool S>
4158 extern void __kmp_atomic_suspend_64(int th_gtid,
4159  kmp_atomic_flag_64<C, S> *flag);
4160 extern void __kmp_suspend_oncore(int th_gtid, kmp_flag_oncore *flag);
4161 #if KMP_HAVE_MWAIT || KMP_HAVE_UMWAIT
4162 template <bool C, bool S>
4163 extern void __kmp_mwait_32(int th_gtid, kmp_flag_32<C, S> *flag);
4164 template <bool C, bool S>
4165 extern void __kmp_mwait_64(int th_gtid, kmp_flag_64<C, S> *flag);
4166 template <bool C, bool S>
4167 extern void __kmp_atomic_mwait_64(int th_gtid, kmp_atomic_flag_64<C, S> *flag);
4168 extern void __kmp_mwait_oncore(int th_gtid, kmp_flag_oncore *flag);
4169 #endif
4170 template <bool C, bool S>
4171 extern void __kmp_resume_32(int target_gtid, kmp_flag_32<C, S> *flag);
4172 template <bool C, bool S>
4173 extern void __kmp_resume_64(int target_gtid, kmp_flag_64<C, S> *flag);
4174 template <bool C, bool S>
4175 extern void __kmp_atomic_resume_64(int target_gtid,
4176  kmp_atomic_flag_64<C, S> *flag);
4177 extern void __kmp_resume_oncore(int target_gtid, kmp_flag_oncore *flag);
4178 
4179 template <bool C, bool S>
4180 int __kmp_execute_tasks_32(kmp_info_t *thread, kmp_int32 gtid,
4181  kmp_flag_32<C, S> *flag, int final_spin,
4182  int *thread_finished,
4183 #if USE_ITT_BUILD
4184  void *itt_sync_obj,
4185 #endif /* USE_ITT_BUILD */
4186  kmp_int32 is_constrained);
4187 template <bool C, bool S>
4188 int __kmp_execute_tasks_64(kmp_info_t *thread, kmp_int32 gtid,
4189  kmp_flag_64<C, S> *flag, int final_spin,
4190  int *thread_finished,
4191 #if USE_ITT_BUILD
4192  void *itt_sync_obj,
4193 #endif /* USE_ITT_BUILD */
4194  kmp_int32 is_constrained);
4195 template <bool C, bool S>
4196 int __kmp_atomic_execute_tasks_64(kmp_info_t *thread, kmp_int32 gtid,
4197  kmp_atomic_flag_64<C, S> *flag,
4198  int final_spin, int *thread_finished,
4199 #if USE_ITT_BUILD
4200  void *itt_sync_obj,
4201 #endif /* USE_ITT_BUILD */
4202  kmp_int32 is_constrained);
4203 int __kmp_execute_tasks_oncore(kmp_info_t *thread, kmp_int32 gtid,
4204  kmp_flag_oncore *flag, int final_spin,
4205  int *thread_finished,
4206 #if USE_ITT_BUILD
4207  void *itt_sync_obj,
4208 #endif /* USE_ITT_BUILD */
4209  kmp_int32 is_constrained);
4210 
4211 extern int __kmp_nesting_mode;
4212 extern int __kmp_nesting_mode_nlevels;
4213 extern int *__kmp_nesting_nth_level;
4214 extern void __kmp_init_nesting_mode();
4215 extern void __kmp_set_nesting_mode_threads();
4216 
4224  FILE *f;
4225 
4226  void close() {
4227  if (f && f != stdout && f != stderr) {
4228  fclose(f);
4229  f = nullptr;
4230  }
4231  }
4232 
4233 public:
4234  kmp_safe_raii_file_t() : f(nullptr) {}
4235  kmp_safe_raii_file_t(const char *filename, const char *mode,
4236  const char *env_var = nullptr)
4237  : f(nullptr) {
4238  open(filename, mode, env_var);
4239  }
4240  ~kmp_safe_raii_file_t() { close(); }
4241 
4245  void open(const char *filename, const char *mode,
4246  const char *env_var = nullptr) {
4247  KMP_ASSERT(!f);
4248  f = fopen(filename, mode);
4249  if (!f) {
4250  int code = errno;
4251  if (env_var) {
4252  __kmp_fatal(KMP_MSG(CantOpenFileForReading, filename), KMP_ERR(code),
4253  KMP_HNT(CheckEnvVar, env_var, filename), __kmp_msg_null);
4254  } else {
4255  __kmp_fatal(KMP_MSG(CantOpenFileForReading, filename), KMP_ERR(code),
4256  __kmp_msg_null);
4257  }
4258  }
4259  }
4262  void set_stdout() {
4263  KMP_ASSERT(!f);
4264  f = stdout;
4265  }
4268  void set_stderr() {
4269  KMP_ASSERT(!f);
4270  f = stderr;
4271  }
4272  operator bool() { return bool(f); }
4273  operator FILE *() { return f; }
4274 };
4275 
4276 template <typename SourceType, typename TargetType,
4277  bool isSourceSmaller = (sizeof(SourceType) < sizeof(TargetType)),
4278  bool isSourceEqual = (sizeof(SourceType) == sizeof(TargetType)),
4279  bool isSourceSigned = std::is_signed<SourceType>::value,
4280  bool isTargetSigned = std::is_signed<TargetType>::value>
4281 struct kmp_convert {};
4282 
4283 // Both types are signed; Source smaller
4284 template <typename SourceType, typename TargetType>
4285 struct kmp_convert<SourceType, TargetType, true, false, true, true> {
4286  static TargetType to(SourceType src) { return (TargetType)src; }
4287 };
4288 // Source equal
4289 template <typename SourceType, typename TargetType>
4290 struct kmp_convert<SourceType, TargetType, false, true, true, true> {
4291  static TargetType to(SourceType src) { return src; }
4292 };
4293 // Source bigger
4294 template <typename SourceType, typename TargetType>
4295 struct kmp_convert<SourceType, TargetType, false, false, true, true> {
4296  static TargetType to(SourceType src) {
4297  KMP_ASSERT(src <= static_cast<SourceType>(
4298  (std::numeric_limits<TargetType>::max)()));
4299  KMP_ASSERT(src >= static_cast<SourceType>(
4300  (std::numeric_limits<TargetType>::min)()));
4301  return (TargetType)src;
4302  }
4303 };
4304 
4305 // Source signed, Target unsigned
4306 // Source smaller
4307 template <typename SourceType, typename TargetType>
4308 struct kmp_convert<SourceType, TargetType, true, false, true, false> {
4309  static TargetType to(SourceType src) {
4310  KMP_ASSERT(src >= 0);
4311  return (TargetType)src;
4312  }
4313 };
4314 // Source equal
4315 template <typename SourceType, typename TargetType>
4316 struct kmp_convert<SourceType, TargetType, false, true, true, false> {
4317  static TargetType to(SourceType src) {
4318  KMP_ASSERT(src >= 0);
4319  return (TargetType)src;
4320  }
4321 };
4322 // Source bigger
4323 template <typename SourceType, typename TargetType>
4324 struct kmp_convert<SourceType, TargetType, false, false, true, false> {
4325  static TargetType to(SourceType src) {
4326  KMP_ASSERT(src >= 0);
4327  KMP_ASSERT(src <= static_cast<SourceType>(
4328  (std::numeric_limits<TargetType>::max)()));
4329  return (TargetType)src;
4330  }
4331 };
4332 
4333 // Source unsigned, Target signed
4334 // Source smaller
4335 template <typename SourceType, typename TargetType>
4336 struct kmp_convert<SourceType, TargetType, true, false, false, true> {
4337  static TargetType to(SourceType src) { return (TargetType)src; }
4338 };
4339 // Source equal
4340 template <typename SourceType, typename TargetType>
4341 struct kmp_convert<SourceType, TargetType, false, true, false, true> {
4342  static TargetType to(SourceType src) {
4343  KMP_ASSERT(src <= static_cast<SourceType>(
4344  (std::numeric_limits<TargetType>::max)()));
4345  return (TargetType)src;
4346  }
4347 };
4348 // Source bigger
4349 template <typename SourceType, typename TargetType>
4350 struct kmp_convert<SourceType, TargetType, false, false, false, true> {
4351  static TargetType to(SourceType src) {
4352  KMP_ASSERT(src <= static_cast<SourceType>(
4353  (std::numeric_limits<TargetType>::max)()));
4354  return (TargetType)src;
4355  }
4356 };
4357 
4358 // Source unsigned, Target unsigned
4359 // Source smaller
4360 template <typename SourceType, typename TargetType>
4361 struct kmp_convert<SourceType, TargetType, true, false, false, false> {
4362  static TargetType to(SourceType src) { return (TargetType)src; }
4363 };
4364 // Source equal
4365 template <typename SourceType, typename TargetType>
4366 struct kmp_convert<SourceType, TargetType, false, true, false, false> {
4367  static TargetType to(SourceType src) { return src; }
4368 };
4369 // Source bigger
4370 template <typename SourceType, typename TargetType>
4371 struct kmp_convert<SourceType, TargetType, false, false, false, false> {
4372  static TargetType to(SourceType src) {
4373  KMP_ASSERT(src <= static_cast<SourceType>(
4374  (std::numeric_limits<TargetType>::max)()));
4375  return (TargetType)src;
4376  }
4377 };
4378 
4379 template <typename T1, typename T2>
4380 static inline void __kmp_type_convert(T1 src, T2 *dest) {
4381  *dest = kmp_convert<T1, T2>::to(src);
4382 }
4383 
4384 #endif /* KMP_H */
void set_stdout()
Definition: kmp.h:4262
void set_stderr()
Definition: kmp.h:4268
void open(const char *filename, const char *mode, const char *env_var=nullptr)
Definition: kmp.h:4245
struct ident ident_t
@ KMP_IDENT_KMPC
Definition: kmp.h:196
@ KMP_IDENT_IMB
Definition: kmp.h:194
@ KMP_IDENT_WORK_LOOP
Definition: kmp.h:214
@ KMP_IDENT_BARRIER_IMPL
Definition: kmp.h:205
@ KMP_IDENT_WORK_SECTIONS
Definition: kmp.h:216
@ KMP_IDENT_AUTOPAR
Definition: kmp.h:199
@ KMP_IDENT_ATOMIC_HINT_MASK
Definition: kmp.h:223
@ KMP_IDENT_WORK_DISTRIBUTE
Definition: kmp.h:218
@ KMP_IDENT_BARRIER_EXPL
Definition: kmp.h:203
@ KMP_IDENT_ATOMIC_REDUCE
Definition: kmp.h:201
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_push_num_threads(ident_t *loc, kmp_int32 global_tid, kmp_int32 num_threads)
KMP_EXPORT void __kmpc_serialized_parallel(ident_t *, kmp_int32 global_tid)
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:1524
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 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)
void(* kmpc_dtor)(void *)
Definition: kmp.h:1548
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:1542
void *(* kmpc_ctor_vec)(void *, size_t)
Definition: kmp.h:1565
void *(* kmpc_cctor)(void *, void *)
Definition: kmp.h:1555
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:1577
void(* kmpc_dtor_vec)(void *, size_t)
Definition: kmp.h:1571
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 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:357
KMP_EXPORT void __kmpc_end_masked(ident_t *, kmp_int32 global_tid)
void __kmpc_dispatch_fini_4(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)
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_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)
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:408
@ kmp_sch_runtime_simd
Definition: kmp.h:379
@ kmp_nm_ord_auto
Definition: kmp.h:427
@ kmp_sch_auto
Definition: kmp.h:364
@ kmp_nm_auto
Definition: kmp.h:410
@ kmp_distribute_static_chunked
Definition: kmp.h:395
@ kmp_sch_static
Definition: kmp.h:360
@ kmp_sch_guided_simd
Definition: kmp.h:378
@ kmp_sch_modifier_monotonic
Definition: kmp.h:445
@ kmp_sch_default
Definition: kmp.h:465
@ kmp_sch_modifier_nonmonotonic
Definition: kmp.h:447
@ kmp_nm_ord_static
Definition: kmp.h:423
@ kmp_distribute_static
Definition: kmp.h:396
@ kmp_sch_guided_chunked
Definition: kmp.h:362
@ kmp_nm_static
Definition: kmp.h:406
@ kmp_sch_lower
Definition: kmp.h:358
@ kmp_nm_upper
Definition: kmp.h:429
@ kmp_ord_lower
Definition: kmp.h:384
@ kmp_ord_static
Definition: kmp.h:386
@ kmp_sch_upper
Definition: kmp.h:382
@ kmp_ord_upper
Definition: kmp.h:392
@ kmp_nm_lower
Definition: kmp.h:402
@ kmp_ord_auto
Definition: kmp.h:390
Definition: kmp.h:234
kmp_int32 reserved_1
Definition: kmp.h:235
char const * psource
Definition: kmp.h:244
kmp_int32 reserved_2
Definition: kmp.h:238
kmp_int32 reserved_3
Definition: kmp.h:243
kmp_int32 flags
Definition: kmp.h:236