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_lock_t *steal_lock; // lock used for chunk stealing
1679  // KMP_ALIGN(32) ensures (if the KMP_ALIGN macro is turned on)
1680  // a) parm3 is properly aligned and
1681  // b) all parm1-4 are on the same cache line.
1682  // Because of parm1-4 are used together, performance seems to be better
1683  // if they are on the same cache line (not measured though).
1684 
1685  struct KMP_ALIGN(32) { // AC: changed 16 to 32 in order to simplify template
1686  kmp_int32 parm1; // structures in kmp_dispatch.cpp. This should
1687  kmp_int32 parm2; // make no real change at least while padding is off.
1688  kmp_int32 parm3;
1689  kmp_int32 parm4;
1690  };
1691 
1692  kmp_uint32 ordered_lower;
1693  kmp_uint32 ordered_upper;
1694 #if KMP_OS_WINDOWS
1695  kmp_int32 last_upper;
1696 #endif /* KMP_OS_WINDOWS */
1697 } dispatch_private_info32_t;
1698 
1699 typedef struct KMP_ALIGN_CACHE dispatch_private_info64 {
1700  kmp_int64 count; // current chunk number for static & static-steal scheduling
1701  kmp_int64 ub; /* upper-bound */
1702  /* Adding KMP_ALIGN_CACHE here doesn't help / can hurt performance */
1703  kmp_int64 lb; /* lower-bound */
1704  kmp_int64 st; /* stride */
1705  kmp_int64 tc; /* trip count (number of iterations) */
1706  kmp_lock_t *steal_lock; // lock used for chunk stealing
1707  /* parm[1-4] are used in different ways by different scheduling algorithms */
1708 
1709  // KMP_ALIGN( 32 ) ensures ( if the KMP_ALIGN macro is turned on )
1710  // a) parm3 is properly aligned and
1711  // b) all parm1-4 are in the same cache line.
1712  // Because of parm1-4 are used together, performance seems to be better
1713  // if they are in the same line (not measured though).
1714 
1715  struct KMP_ALIGN(32) {
1716  kmp_int64 parm1;
1717  kmp_int64 parm2;
1718  kmp_int64 parm3;
1719  kmp_int64 parm4;
1720  };
1721 
1722  kmp_uint64 ordered_lower;
1723  kmp_uint64 ordered_upper;
1724 #if KMP_OS_WINDOWS
1725  kmp_int64 last_upper;
1726 #endif /* KMP_OS_WINDOWS */
1727 } dispatch_private_info64_t;
1728 #else /* KMP_STATIC_STEAL_ENABLED */
1729 typedef struct KMP_ALIGN_CACHE dispatch_private_info32 {
1730  kmp_int32 lb;
1731  kmp_int32 ub;
1732  kmp_int32 st;
1733  kmp_int32 tc;
1734 
1735  kmp_int32 parm1;
1736  kmp_int32 parm2;
1737  kmp_int32 parm3;
1738  kmp_int32 parm4;
1739 
1740  kmp_int32 count;
1741 
1742  kmp_uint32 ordered_lower;
1743  kmp_uint32 ordered_upper;
1744 #if KMP_OS_WINDOWS
1745  kmp_int32 last_upper;
1746 #endif /* KMP_OS_WINDOWS */
1747 } dispatch_private_info32_t;
1748 
1749 typedef struct KMP_ALIGN_CACHE dispatch_private_info64 {
1750  kmp_int64 lb; /* lower-bound */
1751  kmp_int64 ub; /* upper-bound */
1752  kmp_int64 st; /* stride */
1753  kmp_int64 tc; /* trip count (number of iterations) */
1754 
1755  /* parm[1-4] are used in different ways by different scheduling algorithms */
1756  kmp_int64 parm1;
1757  kmp_int64 parm2;
1758  kmp_int64 parm3;
1759  kmp_int64 parm4;
1760 
1761  kmp_int64 count; /* current chunk number for static scheduling */
1762 
1763  kmp_uint64 ordered_lower;
1764  kmp_uint64 ordered_upper;
1765 #if KMP_OS_WINDOWS
1766  kmp_int64 last_upper;
1767 #endif /* KMP_OS_WINDOWS */
1768 } dispatch_private_info64_t;
1769 #endif /* KMP_STATIC_STEAL_ENABLED */
1770 
1771 typedef struct KMP_ALIGN_CACHE dispatch_private_info {
1772  union private_info {
1773  dispatch_private_info32_t p32;
1774  dispatch_private_info64_t p64;
1775  } u;
1776  enum sched_type schedule; /* scheduling algorithm */
1777  kmp_sched_flags_t flags; /* flags (e.g., ordered, nomerge, etc.) */
1778  std::atomic<kmp_uint32> steal_flag; // static_steal only, state of a buffer
1779  kmp_int32 ordered_bumped;
1780  // Stack of buffers for nest of serial regions
1781  struct dispatch_private_info *next;
1782  kmp_int32 type_size; /* the size of types in private_info */
1783 #if KMP_USE_HIER_SCHED
1784  kmp_int32 hier_id;
1785  void *parent; /* hierarchical scheduling parent pointer */
1786 #endif
1787  enum cons_type pushed_ws;
1788 } dispatch_private_info_t;
1789 
1790 typedef struct dispatch_shared_info32 {
1791  /* chunk index under dynamic, number of idle threads under static-steal;
1792  iteration index otherwise */
1793  volatile kmp_uint32 iteration;
1794  volatile kmp_int32 num_done;
1795  volatile kmp_uint32 ordered_iteration;
1796  // Dummy to retain the structure size after making ordered_iteration scalar
1797  kmp_int32 ordered_dummy[KMP_MAX_ORDERED - 1];
1798 } dispatch_shared_info32_t;
1799 
1800 typedef struct dispatch_shared_info64 {
1801  /* chunk index under dynamic, number of idle threads under static-steal;
1802  iteration index otherwise */
1803  volatile kmp_uint64 iteration;
1804  volatile kmp_int64 num_done;
1805  volatile kmp_uint64 ordered_iteration;
1806  // Dummy to retain the structure size after making ordered_iteration scalar
1807  kmp_int64 ordered_dummy[KMP_MAX_ORDERED - 3];
1808 } dispatch_shared_info64_t;
1809 
1810 typedef struct dispatch_shared_info {
1811  union shared_info {
1812  dispatch_shared_info32_t s32;
1813  dispatch_shared_info64_t s64;
1814  } u;
1815  volatile kmp_uint32 buffer_index;
1816  volatile kmp_int32 doacross_buf_idx; // teamwise index
1817  volatile kmp_uint32 *doacross_flags; // shared array of iteration flags (0/1)
1818  kmp_int32 doacross_num_done; // count finished threads
1819 #if KMP_USE_HIER_SCHED
1820  void *hier;
1821 #endif
1822 #if KMP_USE_HWLOC
1823  // When linking with libhwloc, the ORDERED EPCC test slows down on big
1824  // machines (> 48 cores). Performance analysis showed that a cache thrash
1825  // was occurring and this padding helps alleviate the problem.
1826  char padding[64];
1827 #endif
1828 } dispatch_shared_info_t;
1829 
1830 typedef struct kmp_disp {
1831  /* Vector for ORDERED SECTION */
1832  void (*th_deo_fcn)(int *gtid, int *cid, ident_t *);
1833  /* Vector for END ORDERED SECTION */
1834  void (*th_dxo_fcn)(int *gtid, int *cid, ident_t *);
1835 
1836  dispatch_shared_info_t *th_dispatch_sh_current;
1837  dispatch_private_info_t *th_dispatch_pr_current;
1838 
1839  dispatch_private_info_t *th_disp_buffer;
1840  kmp_uint32 th_disp_index;
1841  kmp_int32 th_doacross_buf_idx; // thread's doacross buffer index
1842  volatile kmp_uint32 *th_doacross_flags; // pointer to shared array of flags
1843  kmp_int64 *th_doacross_info; // info on loop bounds
1844 #if KMP_USE_INTERNODE_ALIGNMENT
1845  char more_padding[INTERNODE_CACHE_LINE];
1846 #endif
1847 } kmp_disp_t;
1848 
1849 /* ------------------------------------------------------------------------ */
1850 /* Barrier stuff */
1851 
1852 /* constants for barrier state update */
1853 #define KMP_INIT_BARRIER_STATE 0 /* should probably start from zero */
1854 #define KMP_BARRIER_SLEEP_BIT 0 /* bit used for suspend/sleep part of state */
1855 #define KMP_BARRIER_UNUSED_BIT 1 // bit that must never be set for valid state
1856 #define KMP_BARRIER_BUMP_BIT 2 /* lsb used for bump of go/arrived state */
1857 
1858 #define KMP_BARRIER_SLEEP_STATE (1 << KMP_BARRIER_SLEEP_BIT)
1859 #define KMP_BARRIER_UNUSED_STATE (1 << KMP_BARRIER_UNUSED_BIT)
1860 #define KMP_BARRIER_STATE_BUMP (1 << KMP_BARRIER_BUMP_BIT)
1861 
1862 #if (KMP_BARRIER_SLEEP_BIT >= KMP_BARRIER_BUMP_BIT)
1863 #error "Barrier sleep bit must be smaller than barrier bump bit"
1864 #endif
1865 #if (KMP_BARRIER_UNUSED_BIT >= KMP_BARRIER_BUMP_BIT)
1866 #error "Barrier unused bit must be smaller than barrier bump bit"
1867 #endif
1868 
1869 // Constants for release barrier wait state: currently, hierarchical only
1870 #define KMP_BARRIER_NOT_WAITING 0 // Normal state; worker not in wait_sleep
1871 #define KMP_BARRIER_OWN_FLAG \
1872  1 // Normal state; worker waiting on own b_go flag in release
1873 #define KMP_BARRIER_PARENT_FLAG \
1874  2 // Special state; worker waiting on parent's b_go flag in release
1875 #define KMP_BARRIER_SWITCH_TO_OWN_FLAG \
1876  3 // Special state; tells worker to shift from parent to own b_go
1877 #define KMP_BARRIER_SWITCHING \
1878  4 // Special state; worker resets appropriate flag on wake-up
1879 
1880 #define KMP_NOT_SAFE_TO_REAP \
1881  0 // Thread th_reap_state: not safe to reap (tasking)
1882 #define KMP_SAFE_TO_REAP 1 // Thread th_reap_state: safe to reap (not tasking)
1883 
1884 // The flag_type describes the storage used for the flag.
1885 enum flag_type {
1886  flag32,
1887  flag64,
1888  atomic_flag64,
1889  flag_oncore,
1890  flag_unset
1891 };
1892 
1893 enum barrier_type {
1894  bs_plain_barrier = 0, /* 0, All non-fork/join barriers (except reduction
1895  barriers if enabled) */
1896  bs_forkjoin_barrier, /* 1, All fork/join (parallel region) barriers */
1897 #if KMP_FAST_REDUCTION_BARRIER
1898  bs_reduction_barrier, /* 2, All barriers that are used in reduction */
1899 #endif // KMP_FAST_REDUCTION_BARRIER
1900  bs_last_barrier /* Just a placeholder to mark the end */
1901 };
1902 
1903 // to work with reduction barriers just like with plain barriers
1904 #if !KMP_FAST_REDUCTION_BARRIER
1905 #define bs_reduction_barrier bs_plain_barrier
1906 #endif // KMP_FAST_REDUCTION_BARRIER
1907 
1908 typedef enum kmp_bar_pat { /* Barrier communication patterns */
1909  bp_linear_bar =
1910  0, /* Single level (degenerate) tree */
1911  bp_tree_bar =
1912  1, /* Balanced tree with branching factor 2^n */
1913  bp_hyper_bar = 2, /* Hypercube-embedded tree with min
1914  branching factor 2^n */
1915  bp_hierarchical_bar = 3, /* Machine hierarchy tree */
1916  bp_dist_bar = 4, /* Distributed barrier */
1917  bp_last_bar /* Placeholder to mark the end */
1918 } kmp_bar_pat_e;
1919 
1920 #define KMP_BARRIER_ICV_PUSH 1
1921 
1922 /* Record for holding the values of the internal controls stack records */
1923 typedef struct kmp_internal_control {
1924  int serial_nesting_level; /* corresponds to the value of the
1925  th_team_serialized field */
1926  kmp_int8 dynamic; /* internal control for dynamic adjustment of threads (per
1927  thread) */
1928  kmp_int8
1929  bt_set; /* internal control for whether blocktime is explicitly set */
1930  int blocktime; /* internal control for blocktime */
1931 #if KMP_USE_MONITOR
1932  int bt_intervals; /* internal control for blocktime intervals */
1933 #endif
1934  int nproc; /* internal control for #threads for next parallel region (per
1935  thread) */
1936  int thread_limit; /* internal control for thread-limit-var */
1937  int max_active_levels; /* internal control for max_active_levels */
1938  kmp_r_sched_t
1939  sched; /* internal control for runtime schedule {sched,chunk} pair */
1940  kmp_proc_bind_t proc_bind; /* internal control for affinity */
1941  kmp_int32 default_device; /* internal control for default device */
1942  struct kmp_internal_control *next;
1943 } kmp_internal_control_t;
1944 
1945 static inline void copy_icvs(kmp_internal_control_t *dst,
1946  kmp_internal_control_t *src) {
1947  *dst = *src;
1948 }
1949 
1950 /* Thread barrier needs volatile barrier fields */
1951 typedef struct KMP_ALIGN_CACHE kmp_bstate {
1952  // th_fixed_icvs is aligned by virtue of kmp_bstate being aligned (and all
1953  // uses of it). It is not explicitly aligned below, because we *don't* want
1954  // it to be padded -- instead, we fit b_go into the same cache line with
1955  // th_fixed_icvs, enabling NGO cache lines stores in the hierarchical barrier.
1956  kmp_internal_control_t th_fixed_icvs; // Initial ICVs for the thread
1957  // Tuck b_go into end of th_fixed_icvs cache line, so it can be stored with
1958  // same NGO store
1959  volatile kmp_uint64 b_go; // STATE => task should proceed (hierarchical)
1960  KMP_ALIGN_CACHE volatile kmp_uint64
1961  b_arrived; // STATE => task reached synch point.
1962  kmp_uint32 *skip_per_level;
1963  kmp_uint32 my_level;
1964  kmp_int32 parent_tid;
1965  kmp_int32 old_tid;
1966  kmp_uint32 depth;
1967  struct kmp_bstate *parent_bar;
1968  kmp_team_t *team;
1969  kmp_uint64 leaf_state;
1970  kmp_uint32 nproc;
1971  kmp_uint8 base_leaf_kids;
1972  kmp_uint8 leaf_kids;
1973  kmp_uint8 offset;
1974  kmp_uint8 wait_flag;
1975  kmp_uint8 use_oncore_barrier;
1976 #if USE_DEBUGGER
1977  // The following field is intended for the debugger solely. Only the worker
1978  // thread itself accesses this field: the worker increases it by 1 when it
1979  // arrives to a barrier.
1980  KMP_ALIGN_CACHE kmp_uint b_worker_arrived;
1981 #endif /* USE_DEBUGGER */
1982 } kmp_bstate_t;
1983 
1984 union KMP_ALIGN_CACHE kmp_barrier_union {
1985  double b_align; /* use worst case alignment */
1986  char b_pad[KMP_PAD(kmp_bstate_t, CACHE_LINE)];
1987  kmp_bstate_t bb;
1988 };
1989 
1990 typedef union kmp_barrier_union kmp_balign_t;
1991 
1992 /* Team barrier needs only non-volatile arrived counter */
1993 union KMP_ALIGN_CACHE kmp_barrier_team_union {
1994  double b_align; /* use worst case alignment */
1995  char b_pad[CACHE_LINE];
1996  struct {
1997  kmp_uint64 b_arrived; /* STATE => task reached synch point. */
1998 #if USE_DEBUGGER
1999  // The following two fields are indended for the debugger solely. Only
2000  // primary thread of the team accesses these fields: the first one is
2001  // increased by 1 when the primary thread arrives to a barrier, the second
2002  // one is increased by one when all the threads arrived.
2003  kmp_uint b_master_arrived;
2004  kmp_uint b_team_arrived;
2005 #endif
2006  };
2007 };
2008 
2009 typedef union kmp_barrier_team_union kmp_balign_team_t;
2010 
2011 /* Padding for Linux* OS pthreads condition variables and mutexes used to signal
2012  threads when a condition changes. This is to workaround an NPTL bug where
2013  padding was added to pthread_cond_t which caused the initialization routine
2014  to write outside of the structure if compiled on pre-NPTL threads. */
2015 #if KMP_OS_WINDOWS
2016 typedef struct kmp_win32_mutex {
2017  /* The Lock */
2018  CRITICAL_SECTION cs;
2019 } kmp_win32_mutex_t;
2020 
2021 typedef struct kmp_win32_cond {
2022  /* Count of the number of waiters. */
2023  int waiters_count_;
2024 
2025  /* Serialize access to <waiters_count_> */
2026  kmp_win32_mutex_t waiters_count_lock_;
2027 
2028  /* Number of threads to release via a <cond_broadcast> or a <cond_signal> */
2029  int release_count_;
2030 
2031  /* Keeps track of the current "generation" so that we don't allow */
2032  /* one thread to steal all the "releases" from the broadcast. */
2033  int wait_generation_count_;
2034 
2035  /* A manual-reset event that's used to block and release waiting threads. */
2036  HANDLE event_;
2037 } kmp_win32_cond_t;
2038 #endif
2039 
2040 #if KMP_OS_UNIX
2041 
2042 union KMP_ALIGN_CACHE kmp_cond_union {
2043  double c_align;
2044  char c_pad[CACHE_LINE];
2045  pthread_cond_t c_cond;
2046 };
2047 
2048 typedef union kmp_cond_union kmp_cond_align_t;
2049 
2050 union KMP_ALIGN_CACHE kmp_mutex_union {
2051  double m_align;
2052  char m_pad[CACHE_LINE];
2053  pthread_mutex_t m_mutex;
2054 };
2055 
2056 typedef union kmp_mutex_union kmp_mutex_align_t;
2057 
2058 #endif /* KMP_OS_UNIX */
2059 
2060 typedef struct kmp_desc_base {
2061  void *ds_stackbase;
2062  size_t ds_stacksize;
2063  int ds_stackgrow;
2064  kmp_thread_t ds_thread;
2065  volatile int ds_tid;
2066  int ds_gtid;
2067 #if KMP_OS_WINDOWS
2068  volatile int ds_alive;
2069  DWORD ds_thread_id;
2070 /* ds_thread keeps thread handle on Windows* OS. It is enough for RTL purposes.
2071  However, debugger support (libomp_db) cannot work with handles, because they
2072  uncomparable. For example, debugger requests info about thread with handle h.
2073  h is valid within debugger process, and meaningless within debugee process.
2074  Even if h is duped by call to DuplicateHandle(), so the result h' is valid
2075  within debugee process, but it is a *new* handle which does *not* equal to
2076  any other handle in debugee... The only way to compare handles is convert
2077  them to system-wide ids. GetThreadId() function is available only in
2078  Longhorn and Server 2003. :-( In contrast, GetCurrentThreadId() is available
2079  on all Windows* OS flavours (including Windows* 95). Thus, we have to get
2080  thread id by call to GetCurrentThreadId() from within the thread and save it
2081  to let libomp_db identify threads. */
2082 #endif /* KMP_OS_WINDOWS */
2083 } kmp_desc_base_t;
2084 
2085 typedef union KMP_ALIGN_CACHE kmp_desc {
2086  double ds_align; /* use worst case alignment */
2087  char ds_pad[KMP_PAD(kmp_desc_base_t, CACHE_LINE)];
2088  kmp_desc_base_t ds;
2089 } kmp_desc_t;
2090 
2091 typedef struct kmp_local {
2092  volatile int this_construct; /* count of single's encountered by thread */
2093  void *reduce_data;
2094 #if KMP_USE_BGET
2095  void *bget_data;
2096  void *bget_list;
2097 #if !USE_CMP_XCHG_FOR_BGET
2098 #ifdef USE_QUEUING_LOCK_FOR_BGET
2099  kmp_lock_t bget_lock; /* Lock for accessing bget free list */
2100 #else
2101  kmp_bootstrap_lock_t bget_lock; // Lock for accessing bget free list. Must be
2102 // bootstrap lock so we can use it at library
2103 // shutdown.
2104 #endif /* USE_LOCK_FOR_BGET */
2105 #endif /* ! USE_CMP_XCHG_FOR_BGET */
2106 #endif /* KMP_USE_BGET */
2107 
2108  PACKED_REDUCTION_METHOD_T
2109  packed_reduction_method; /* stored by __kmpc_reduce*(), used by
2110  __kmpc_end_reduce*() */
2111 
2112 } kmp_local_t;
2113 
2114 #define KMP_CHECK_UPDATE(a, b) \
2115  if ((a) != (b)) \
2116  (a) = (b)
2117 #define KMP_CHECK_UPDATE_SYNC(a, b) \
2118  if ((a) != (b)) \
2119  TCW_SYNC_PTR((a), (b))
2120 
2121 #define get__blocktime(xteam, xtid) \
2122  ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.blocktime)
2123 #define get__bt_set(xteam, xtid) \
2124  ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_set)
2125 #if KMP_USE_MONITOR
2126 #define get__bt_intervals(xteam, xtid) \
2127  ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_intervals)
2128 #endif
2129 
2130 #define get__dynamic_2(xteam, xtid) \
2131  ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.dynamic)
2132 #define get__nproc_2(xteam, xtid) \
2133  ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.nproc)
2134 #define get__sched_2(xteam, xtid) \
2135  ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.sched)
2136 
2137 #define set__blocktime_team(xteam, xtid, xval) \
2138  (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.blocktime) = \
2139  (xval))
2140 
2141 #if KMP_USE_MONITOR
2142 #define set__bt_intervals_team(xteam, xtid, xval) \
2143  (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_intervals) = \
2144  (xval))
2145 #endif
2146 
2147 #define set__bt_set_team(xteam, xtid, xval) \
2148  (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_set) = (xval))
2149 
2150 #define set__dynamic(xthread, xval) \
2151  (((xthread)->th.th_current_task->td_icvs.dynamic) = (xval))
2152 #define get__dynamic(xthread) \
2153  (((xthread)->th.th_current_task->td_icvs.dynamic) ? (FTN_TRUE) : (FTN_FALSE))
2154 
2155 #define set__nproc(xthread, xval) \
2156  (((xthread)->th.th_current_task->td_icvs.nproc) = (xval))
2157 
2158 #define set__thread_limit(xthread, xval) \
2159  (((xthread)->th.th_current_task->td_icvs.thread_limit) = (xval))
2160 
2161 #define set__max_active_levels(xthread, xval) \
2162  (((xthread)->th.th_current_task->td_icvs.max_active_levels) = (xval))
2163 
2164 #define get__max_active_levels(xthread) \
2165  ((xthread)->th.th_current_task->td_icvs.max_active_levels)
2166 
2167 #define set__sched(xthread, xval) \
2168  (((xthread)->th.th_current_task->td_icvs.sched) = (xval))
2169 
2170 #define set__proc_bind(xthread, xval) \
2171  (((xthread)->th.th_current_task->td_icvs.proc_bind) = (xval))
2172 #define get__proc_bind(xthread) \
2173  ((xthread)->th.th_current_task->td_icvs.proc_bind)
2174 
2175 // OpenMP tasking data structures
2176 
2177 typedef enum kmp_tasking_mode {
2178  tskm_immediate_exec = 0,
2179  tskm_extra_barrier = 1,
2180  tskm_task_teams = 2,
2181  tskm_max = 2
2182 } kmp_tasking_mode_t;
2183 
2184 extern kmp_tasking_mode_t
2185  __kmp_tasking_mode; /* determines how/when to execute tasks */
2186 extern int __kmp_task_stealing_constraint;
2187 extern int __kmp_enable_task_throttling;
2188 extern kmp_int32 __kmp_default_device; // Set via OMP_DEFAULT_DEVICE if
2189 // specified, defaults to 0 otherwise
2190 // Set via OMP_MAX_TASK_PRIORITY if specified, defaults to 0 otherwise
2191 extern kmp_int32 __kmp_max_task_priority;
2192 // Set via KMP_TASKLOOP_MIN_TASKS if specified, defaults to 0 otherwise
2193 extern kmp_uint64 __kmp_taskloop_min_tasks;
2194 
2195 /* NOTE: kmp_taskdata_t and kmp_task_t structures allocated in single block with
2196  taskdata first */
2197 #define KMP_TASK_TO_TASKDATA(task) (((kmp_taskdata_t *)task) - 1)
2198 #define KMP_TASKDATA_TO_TASK(taskdata) (kmp_task_t *)(taskdata + 1)
2199 
2200 // The tt_found_tasks flag is a signal to all threads in the team that tasks
2201 // were spawned and queued since the previous barrier release.
2202 #define KMP_TASKING_ENABLED(task_team) \
2203  (TRUE == TCR_SYNC_4((task_team)->tt.tt_found_tasks))
2211 typedef kmp_int32 (*kmp_routine_entry_t)(kmp_int32, void *);
2212 
2213 typedef union kmp_cmplrdata {
2214  kmp_int32 priority;
2215  kmp_routine_entry_t
2216  destructors; /* pointer to function to invoke deconstructors of
2217  firstprivate C++ objects */
2218  /* future data */
2219 } kmp_cmplrdata_t;
2220 
2221 /* sizeof_kmp_task_t passed as arg to kmpc_omp_task call */
2224 typedef struct kmp_task { /* GEH: Shouldn't this be aligned somehow? */
2225  void *shareds;
2226  kmp_routine_entry_t
2227  routine;
2228  kmp_int32 part_id;
2229  kmp_cmplrdata_t
2230  data1; /* Two known optional additions: destructors and priority */
2231  kmp_cmplrdata_t data2; /* Process destructors first, priority second */
2232  /* future data */
2233  /* private vars */
2234 } kmp_task_t;
2235 
2240 typedef struct kmp_taskgroup {
2241  std::atomic<kmp_int32> count; // number of allocated and incomplete tasks
2242  std::atomic<kmp_int32>
2243  cancel_request; // request for cancellation of this taskgroup
2244  struct kmp_taskgroup *parent; // parent taskgroup
2245  // Block of data to perform task reduction
2246  void *reduce_data; // reduction related info
2247  kmp_int32 reduce_num_data; // number of data items to reduce
2248  uintptr_t *gomp_data; // gomp reduction data
2249 } kmp_taskgroup_t;
2250 
2251 // forward declarations
2252 typedef union kmp_depnode kmp_depnode_t;
2253 typedef struct kmp_depnode_list kmp_depnode_list_t;
2254 typedef struct kmp_dephash_entry kmp_dephash_entry_t;
2255 
2256 #define KMP_DEP_IN 0x1
2257 #define KMP_DEP_OUT 0x2
2258 #define KMP_DEP_INOUT 0x3
2259 #define KMP_DEP_MTX 0x4
2260 #define KMP_DEP_SET 0x8
2261 // Compiler sends us this info:
2262 typedef struct kmp_depend_info {
2263  kmp_intptr_t base_addr;
2264  size_t len;
2265  union {
2266  kmp_uint8 flag;
2267  struct {
2268  unsigned in : 1;
2269  unsigned out : 1;
2270  unsigned mtx : 1;
2271  unsigned set : 1;
2272  } flags;
2273  };
2274 } kmp_depend_info_t;
2275 
2276 // Internal structures to work with task dependencies:
2277 struct kmp_depnode_list {
2278  kmp_depnode_t *node;
2279  kmp_depnode_list_t *next;
2280 };
2281 
2282 // Max number of mutexinoutset dependencies per node
2283 #define MAX_MTX_DEPS 4
2284 
2285 typedef struct kmp_base_depnode {
2286  kmp_depnode_list_t *successors; /* used under lock */
2287  kmp_task_t *task; /* non-NULL if depnode is active, used under lock */
2288  kmp_lock_t *mtx_locks[MAX_MTX_DEPS]; /* lock mutexinoutset dependent tasks */
2289  kmp_int32 mtx_num_locks; /* number of locks in mtx_locks array */
2290  kmp_lock_t lock; /* guards shared fields: task, successors */
2291 #if KMP_SUPPORT_GRAPH_OUTPUT
2292  kmp_uint32 id;
2293 #endif
2294  std::atomic<kmp_int32> npredecessors;
2295  std::atomic<kmp_int32> nrefs;
2296 } kmp_base_depnode_t;
2297 
2298 union KMP_ALIGN_CACHE kmp_depnode {
2299  double dn_align; /* use worst case alignment */
2300  char dn_pad[KMP_PAD(kmp_base_depnode_t, CACHE_LINE)];
2301  kmp_base_depnode_t dn;
2302 };
2303 
2304 struct kmp_dephash_entry {
2305  kmp_intptr_t addr;
2306  kmp_depnode_t *last_out;
2307  kmp_depnode_list_t *last_set;
2308  kmp_depnode_list_t *prev_set;
2309  kmp_uint8 last_flag;
2310  kmp_lock_t *mtx_lock; /* is referenced by depnodes w/mutexinoutset dep */
2311  kmp_dephash_entry_t *next_in_bucket;
2312 };
2313 
2314 typedef struct kmp_dephash {
2315  kmp_dephash_entry_t **buckets;
2316  size_t size;
2317  size_t generation;
2318  kmp_uint32 nelements;
2319  kmp_uint32 nconflicts;
2320 } kmp_dephash_t;
2321 
2322 typedef struct kmp_task_affinity_info {
2323  kmp_intptr_t base_addr;
2324  size_t len;
2325  struct {
2326  bool flag1 : 1;
2327  bool flag2 : 1;
2328  kmp_int32 reserved : 30;
2329  } flags;
2330 } kmp_task_affinity_info_t;
2331 
2332 typedef enum kmp_event_type_t {
2333  KMP_EVENT_UNINITIALIZED = 0,
2334  KMP_EVENT_ALLOW_COMPLETION = 1
2335 } kmp_event_type_t;
2336 
2337 typedef struct {
2338  kmp_event_type_t type;
2339  kmp_tas_lock_t lock;
2340  union {
2341  kmp_task_t *task;
2342  } ed;
2343 } kmp_event_t;
2344 
2345 #ifdef BUILD_TIED_TASK_STACK
2346 
2347 /* Tied Task stack definitions */
2348 typedef struct kmp_stack_block {
2349  kmp_taskdata_t *sb_block[TASK_STACK_BLOCK_SIZE];
2350  struct kmp_stack_block *sb_next;
2351  struct kmp_stack_block *sb_prev;
2352 } kmp_stack_block_t;
2353 
2354 typedef struct kmp_task_stack {
2355  kmp_stack_block_t ts_first_block; // first block of stack entries
2356  kmp_taskdata_t **ts_top; // pointer to the top of stack
2357  kmp_int32 ts_entries; // number of entries on the stack
2358 } kmp_task_stack_t;
2359 
2360 #endif // BUILD_TIED_TASK_STACK
2361 
2362 typedef struct kmp_tasking_flags { /* Total struct must be exactly 32 bits */
2363  /* Compiler flags */ /* Total compiler flags must be 16 bits */
2364  unsigned tiedness : 1; /* task is either tied (1) or untied (0) */
2365  unsigned final : 1; /* task is final(1) so execute immediately */
2366  unsigned merged_if0 : 1; /* no __kmpc_task_{begin/complete}_if0 calls in if0
2367  code path */
2368  unsigned destructors_thunk : 1; /* set if the compiler creates a thunk to
2369  invoke destructors from the runtime */
2370  unsigned proxy : 1; /* task is a proxy task (it will be executed outside the
2371  context of the RTL) */
2372  unsigned priority_specified : 1; /* set if the compiler provides priority
2373  setting for the task */
2374  unsigned detachable : 1; /* 1 == can detach */
2375  unsigned hidden_helper : 1; /* 1 == hidden helper task */
2376  unsigned reserved : 8; /* reserved for compiler use */
2377 
2378  /* Library flags */ /* Total library flags must be 16 bits */
2379  unsigned tasktype : 1; /* task is either explicit(1) or implicit (0) */
2380  unsigned task_serial : 1; // task is executed immediately (1) or deferred (0)
2381  unsigned tasking_ser : 1; // all tasks in team are either executed immediately
2382  // (1) or may be deferred (0)
2383  unsigned team_serial : 1; // entire team is serial (1) [1 thread] or parallel
2384  // (0) [>= 2 threads]
2385  /* If either team_serial or tasking_ser is set, task team may be NULL */
2386  /* Task State Flags: */
2387  unsigned started : 1; /* 1==started, 0==not started */
2388  unsigned executing : 1; /* 1==executing, 0==not executing */
2389  unsigned complete : 1; /* 1==complete, 0==not complete */
2390  unsigned freed : 1; /* 1==freed, 0==allocated */
2391  unsigned native : 1; /* 1==gcc-compiled task, 0==intel */
2392  unsigned reserved31 : 7; /* reserved for library use */
2393 
2394 } kmp_tasking_flags_t;
2395 
2396 struct kmp_taskdata { /* aligned during dynamic allocation */
2397  kmp_int32 td_task_id; /* id, assigned by debugger */
2398  kmp_tasking_flags_t td_flags; /* task flags */
2399  kmp_team_t *td_team; /* team for this task */
2400  kmp_info_p *td_alloc_thread; /* thread that allocated data structures */
2401  /* Currently not used except for perhaps IDB */
2402  kmp_taskdata_t *td_parent; /* parent task */
2403  kmp_int32 td_level; /* task nesting level */
2404  std::atomic<kmp_int32> td_untied_count; // untied task active parts counter
2405  ident_t *td_ident; /* task identifier */
2406  // Taskwait data.
2407  ident_t *td_taskwait_ident;
2408  kmp_uint32 td_taskwait_counter;
2409  kmp_int32 td_taskwait_thread; /* gtid + 1 of thread encountered taskwait */
2410  KMP_ALIGN_CACHE kmp_internal_control_t
2411  td_icvs; /* Internal control variables for the task */
2412  KMP_ALIGN_CACHE std::atomic<kmp_int32>
2413  td_allocated_child_tasks; /* Child tasks (+ current task) not yet
2414  deallocated */
2415  std::atomic<kmp_int32>
2416  td_incomplete_child_tasks; /* Child tasks not yet complete */
2417  kmp_taskgroup_t
2418  *td_taskgroup; // Each task keeps pointer to its current taskgroup
2419  kmp_dephash_t
2420  *td_dephash; // Dependencies for children tasks are tracked from here
2421  kmp_depnode_t
2422  *td_depnode; // Pointer to graph node if this task has dependencies
2423  kmp_task_team_t *td_task_team;
2424  // The global thread id of the encountering thread. We need it because when a
2425  // regular task depends on a hidden helper task, and the hidden helper task
2426  // is finished on a hidden helper thread, it will call __kmp_release_deps to
2427  // release all dependences. If now the task is a regular task, we need to pass
2428  // the encountering gtid such that the task will be picked up and executed by
2429  // its encountering team instead of hidden helper team.
2430  kmp_int32 encountering_gtid;
2431  size_t td_size_alloc; // Size of task structure, including shareds etc.
2432 #if defined(KMP_GOMP_COMPAT)
2433  // 4 or 8 byte integers for the loop bounds in GOMP_taskloop
2434  kmp_int32 td_size_loop_bounds;
2435 #endif
2436  kmp_taskdata_t *td_last_tied; // keep tied task for task scheduling constraint
2437 #if defined(KMP_GOMP_COMPAT)
2438  // GOMP sends in a copy function for copy constructors
2439  void (*td_copy_func)(void *, void *);
2440 #endif
2441  kmp_event_t td_allow_completion_event;
2442 #if OMPT_SUPPORT
2443  ompt_task_info_t ompt_task_info;
2444 #endif
2445 }; // struct kmp_taskdata
2446 
2447 // Make sure padding above worked
2448 KMP_BUILD_ASSERT(sizeof(kmp_taskdata_t) % sizeof(void *) == 0);
2449 
2450 // Data for task team but per thread
2451 typedef struct kmp_base_thread_data {
2452  kmp_info_p *td_thr; // Pointer back to thread info
2453  // Used only in __kmp_execute_tasks_template, maybe not avail until task is
2454  // queued?
2455  kmp_bootstrap_lock_t td_deque_lock; // Lock for accessing deque
2456  kmp_taskdata_t *
2457  *td_deque; // Deque of tasks encountered by td_thr, dynamically allocated
2458  kmp_int32 td_deque_size; // Size of deck
2459  kmp_uint32 td_deque_head; // Head of deque (will wrap)
2460  kmp_uint32 td_deque_tail; // Tail of deque (will wrap)
2461  kmp_int32 td_deque_ntasks; // Number of tasks in deque
2462  // GEH: shouldn't this be volatile since used in while-spin?
2463  kmp_int32 td_deque_last_stolen; // Thread number of last successful steal
2464 #ifdef BUILD_TIED_TASK_STACK
2465  kmp_task_stack_t td_susp_tied_tasks; // Stack of suspended tied tasks for task
2466 // scheduling constraint
2467 #endif // BUILD_TIED_TASK_STACK
2468 } kmp_base_thread_data_t;
2469 
2470 #define TASK_DEQUE_BITS 8 // Used solely to define INITIAL_TASK_DEQUE_SIZE
2471 #define INITIAL_TASK_DEQUE_SIZE (1 << TASK_DEQUE_BITS)
2472 
2473 #define TASK_DEQUE_SIZE(td) ((td).td_deque_size)
2474 #define TASK_DEQUE_MASK(td) ((td).td_deque_size - 1)
2475 
2476 typedef union KMP_ALIGN_CACHE kmp_thread_data {
2477  kmp_base_thread_data_t td;
2478  double td_align; /* use worst case alignment */
2479  char td_pad[KMP_PAD(kmp_base_thread_data_t, CACHE_LINE)];
2480 } kmp_thread_data_t;
2481 
2482 // Data for task teams which are used when tasking is enabled for the team
2483 typedef struct kmp_base_task_team {
2484  kmp_bootstrap_lock_t
2485  tt_threads_lock; /* Lock used to allocate per-thread part of task team */
2486  /* must be bootstrap lock since used at library shutdown*/
2487  kmp_task_team_t *tt_next; /* For linking the task team free list */
2488  kmp_thread_data_t
2489  *tt_threads_data; /* Array of per-thread structures for task team */
2490  /* Data survives task team deallocation */
2491  kmp_int32 tt_found_tasks; /* Have we found tasks and queued them while
2492  executing this team? */
2493  /* TRUE means tt_threads_data is set up and initialized */
2494  kmp_int32 tt_nproc; /* #threads in team */
2495  kmp_int32 tt_max_threads; // # entries allocated for threads_data array
2496  kmp_int32 tt_found_proxy_tasks; // found proxy tasks since last barrier
2497  kmp_int32 tt_untied_task_encountered;
2498  // There is hidden helper thread encountered in this task team so that we must
2499  // wait when waiting on task team
2500  kmp_int32 tt_hidden_helper_task_encountered;
2501 
2502  KMP_ALIGN_CACHE
2503  std::atomic<kmp_int32> tt_unfinished_threads; /* #threads still active */
2504 
2505  KMP_ALIGN_CACHE
2506  volatile kmp_uint32
2507  tt_active; /* is the team still actively executing tasks */
2508 } kmp_base_task_team_t;
2509 
2510 union KMP_ALIGN_CACHE kmp_task_team {
2511  kmp_base_task_team_t tt;
2512  double tt_align; /* use worst case alignment */
2513  char tt_pad[KMP_PAD(kmp_base_task_team_t, CACHE_LINE)];
2514 };
2515 
2516 #if (USE_FAST_MEMORY == 3) || (USE_FAST_MEMORY == 5)
2517 // Free lists keep same-size free memory slots for fast memory allocation
2518 // routines
2519 typedef struct kmp_free_list {
2520  void *th_free_list_self; // Self-allocated tasks free list
2521  void *th_free_list_sync; // Self-allocated tasks stolen/returned by other
2522  // threads
2523  void *th_free_list_other; // Non-self free list (to be returned to owner's
2524  // sync list)
2525 } kmp_free_list_t;
2526 #endif
2527 #if KMP_NESTED_HOT_TEAMS
2528 // Hot teams array keeps hot teams and their sizes for given thread. Hot teams
2529 // are not put in teams pool, and they don't put threads in threads pool.
2530 typedef struct kmp_hot_team_ptr {
2531  kmp_team_p *hot_team; // pointer to hot_team of given nesting level
2532  kmp_int32 hot_team_nth; // number of threads allocated for the hot_team
2533 } kmp_hot_team_ptr_t;
2534 #endif
2535 typedef struct kmp_teams_size {
2536  kmp_int32 nteams; // number of teams in a league
2537  kmp_int32 nth; // number of threads in each team of the league
2538 } kmp_teams_size_t;
2539 
2540 // This struct stores a thread that acts as a "root" for a contention
2541 // group. Contention groups are rooted at kmp_root threads, but also at
2542 // each primary thread of each team created in the teams construct.
2543 // This struct therefore also stores a thread_limit associated with
2544 // that contention group, and a counter to track the number of threads
2545 // active in that contention group. Each thread has a list of these: CG
2546 // root threads have an entry in their list in which cg_root refers to
2547 // the thread itself, whereas other workers in the CG will have a
2548 // single entry where cg_root is same as the entry containing their CG
2549 // root. When a thread encounters a teams construct, it will add a new
2550 // entry to the front of its list, because it now roots a new CG.
2551 typedef struct kmp_cg_root {
2552  kmp_info_p *cg_root; // "root" thread for a contention group
2553  // The CG root's limit comes from OMP_THREAD_LIMIT for root threads, or
2554  // thread_limit clause for teams primary threads
2555  kmp_int32 cg_thread_limit;
2556  kmp_int32 cg_nthreads; // Count of active threads in CG rooted at cg_root
2557  struct kmp_cg_root *up; // pointer to higher level CG root in list
2558 } kmp_cg_root_t;
2559 
2560 // OpenMP thread data structures
2561 
2562 typedef struct KMP_ALIGN_CACHE kmp_base_info {
2563  /* Start with the readonly data which is cache aligned and padded. This is
2564  written before the thread starts working by the primary thread. Uber
2565  masters may update themselves later. Usage does not consider serialized
2566  regions. */
2567  kmp_desc_t th_info;
2568  kmp_team_p *th_team; /* team we belong to */
2569  kmp_root_p *th_root; /* pointer to root of task hierarchy */
2570  kmp_info_p *th_next_pool; /* next available thread in the pool */
2571  kmp_disp_t *th_dispatch; /* thread's dispatch data */
2572  int th_in_pool; /* in thread pool (32 bits for TCR/TCW) */
2573 
2574  /* The following are cached from the team info structure */
2575  /* TODO use these in more places as determined to be needed via profiling */
2576  int th_team_nproc; /* number of threads in a team */
2577  kmp_info_p *th_team_master; /* the team's primary thread */
2578  int th_team_serialized; /* team is serialized */
2579  microtask_t th_teams_microtask; /* save entry address for teams construct */
2580  int th_teams_level; /* save initial level of teams construct */
2581 /* it is 0 on device but may be any on host */
2582 
2583 /* The blocktime info is copied from the team struct to the thread struct */
2584 /* at the start of a barrier, and the values stored in the team are used */
2585 /* at points in the code where the team struct is no longer guaranteed */
2586 /* to exist (from the POV of worker threads). */
2587 #if KMP_USE_MONITOR
2588  int th_team_bt_intervals;
2589  int th_team_bt_set;
2590 #else
2591  kmp_uint64 th_team_bt_intervals;
2592 #endif
2593 
2594 #if KMP_AFFINITY_SUPPORTED
2595  kmp_affin_mask_t *th_affin_mask; /* thread's current affinity mask */
2596 #endif
2597  omp_allocator_handle_t th_def_allocator; /* default allocator */
2598  /* The data set by the primary thread at reinit, then R/W by the worker */
2599  KMP_ALIGN_CACHE int
2600  th_set_nproc; /* if > 0, then only use this request for the next fork */
2601 #if KMP_NESTED_HOT_TEAMS
2602  kmp_hot_team_ptr_t *th_hot_teams; /* array of hot teams */
2603 #endif
2604  kmp_proc_bind_t
2605  th_set_proc_bind; /* if != proc_bind_default, use request for next fork */
2606  kmp_teams_size_t
2607  th_teams_size; /* number of teams/threads in teams construct */
2608 #if KMP_AFFINITY_SUPPORTED
2609  int th_current_place; /* place currently bound to */
2610  int th_new_place; /* place to bind to in par reg */
2611  int th_first_place; /* first place in partition */
2612  int th_last_place; /* last place in partition */
2613 #endif
2614  int th_prev_level; /* previous level for affinity format */
2615  int th_prev_num_threads; /* previous num_threads for affinity format */
2616 #if USE_ITT_BUILD
2617  kmp_uint64 th_bar_arrive_time; /* arrival to barrier timestamp */
2618  kmp_uint64 th_bar_min_time; /* minimum arrival time at the barrier */
2619  kmp_uint64 th_frame_time; /* frame timestamp */
2620 #endif /* USE_ITT_BUILD */
2621  kmp_local_t th_local;
2622  struct private_common *th_pri_head;
2623 
2624  /* Now the data only used by the worker (after initial allocation) */
2625  /* TODO the first serial team should actually be stored in the info_t
2626  structure. this will help reduce initial allocation overhead */
2627  KMP_ALIGN_CACHE kmp_team_p
2628  *th_serial_team; /*serialized team held in reserve*/
2629 
2630 #if OMPT_SUPPORT
2631  ompt_thread_info_t ompt_thread_info;
2632 #endif
2633 
2634  /* The following are also read by the primary thread during reinit */
2635  struct common_table *th_pri_common;
2636 
2637  volatile kmp_uint32 th_spin_here; /* thread-local location for spinning */
2638  /* while awaiting queuing lock acquire */
2639 
2640  volatile void *th_sleep_loc; // this points at a kmp_flag<T>
2641  flag_type th_sleep_loc_type; // enum type of flag stored in th_sleep_loc
2642 
2643  ident_t *th_ident;
2644  unsigned th_x; // Random number generator data
2645  unsigned th_a; // Random number generator data
2646 
2647  /* Tasking-related data for the thread */
2648  kmp_task_team_t *th_task_team; // Task team struct
2649  kmp_taskdata_t *th_current_task; // Innermost Task being executed
2650  kmp_uint8 th_task_state; // alternating 0/1 for task team identification
2651  kmp_uint8 *th_task_state_memo_stack; // Stack holding memos of th_task_state
2652  // at nested levels
2653  kmp_uint32 th_task_state_top; // Top element of th_task_state_memo_stack
2654  kmp_uint32 th_task_state_stack_sz; // Size of th_task_state_memo_stack
2655  kmp_uint32 th_reap_state; // Non-zero indicates thread is not
2656  // tasking, thus safe to reap
2657 
2658  /* More stuff for keeping track of active/sleeping threads (this part is
2659  written by the worker thread) */
2660  kmp_uint8 th_active_in_pool; // included in count of #active threads in pool
2661  int th_active; // ! sleeping; 32 bits for TCR/TCW
2662  std::atomic<kmp_uint32> th_used_in_team; // Flag indicating use in team
2663  // 0 = not used in team; 1 = used in team;
2664  // 2 = transitioning to not used in team; 3 = transitioning to used in team
2665  struct cons_header *th_cons; // used for consistency check
2666 #if KMP_USE_HIER_SCHED
2667  // used for hierarchical scheduling
2668  kmp_hier_private_bdata_t *th_hier_bar_data;
2669 #endif
2670 
2671  /* Add the syncronizing data which is cache aligned and padded. */
2672  KMP_ALIGN_CACHE kmp_balign_t th_bar[bs_last_barrier];
2673 
2674  KMP_ALIGN_CACHE volatile kmp_int32
2675  th_next_waiting; /* gtid+1 of next thread on lock wait queue, 0 if none */
2676 
2677 #if (USE_FAST_MEMORY == 3) || (USE_FAST_MEMORY == 5)
2678 #define NUM_LISTS 4
2679  kmp_free_list_t th_free_lists[NUM_LISTS]; // Free lists for fast memory
2680 // allocation routines
2681 #endif
2682 
2683 #if KMP_OS_WINDOWS
2684  kmp_win32_cond_t th_suspend_cv;
2685  kmp_win32_mutex_t th_suspend_mx;
2686  std::atomic<int> th_suspend_init;
2687 #endif
2688 #if KMP_OS_UNIX
2689  kmp_cond_align_t th_suspend_cv;
2690  kmp_mutex_align_t th_suspend_mx;
2691  std::atomic<int> th_suspend_init_count;
2692 #endif
2693 
2694 #if USE_ITT_BUILD
2695  kmp_itt_mark_t th_itt_mark_single;
2696 // alignment ???
2697 #endif /* USE_ITT_BUILD */
2698 #if KMP_STATS_ENABLED
2699  kmp_stats_list *th_stats;
2700 #endif
2701 #if KMP_OS_UNIX
2702  std::atomic<bool> th_blocking;
2703 #endif
2704  kmp_cg_root_t *th_cg_roots; // list of cg_roots associated with this thread
2705 } kmp_base_info_t;
2706 
2707 typedef union KMP_ALIGN_CACHE kmp_info {
2708  double th_align; /* use worst case alignment */
2709  char th_pad[KMP_PAD(kmp_base_info_t, CACHE_LINE)];
2710  kmp_base_info_t th;
2711 } kmp_info_t;
2712 
2713 // OpenMP thread team data structures
2714 
2715 typedef struct kmp_base_data {
2716  volatile kmp_uint32 t_value;
2717 } kmp_base_data_t;
2718 
2719 typedef union KMP_ALIGN_CACHE kmp_sleep_team {
2720  double dt_align; /* use worst case alignment */
2721  char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
2722  kmp_base_data_t dt;
2723 } kmp_sleep_team_t;
2724 
2725 typedef union KMP_ALIGN_CACHE kmp_ordered_team {
2726  double dt_align; /* use worst case alignment */
2727  char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
2728  kmp_base_data_t dt;
2729 } kmp_ordered_team_t;
2730 
2731 typedef int (*launch_t)(int gtid);
2732 
2733 /* Minimum number of ARGV entries to malloc if necessary */
2734 #define KMP_MIN_MALLOC_ARGV_ENTRIES 100
2735 
2736 // Set up how many argv pointers will fit in cache lines containing
2737 // t_inline_argv. Historically, we have supported at least 96 bytes. Using a
2738 // larger value for more space between the primary write/worker read section and
2739 // read/write by all section seems to buy more performance on EPCC PARALLEL.
2740 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
2741 #define KMP_INLINE_ARGV_BYTES \
2742  (4 * CACHE_LINE - \
2743  ((3 * KMP_PTR_SKIP + 2 * sizeof(int) + 2 * sizeof(kmp_int8) + \
2744  sizeof(kmp_int16) + sizeof(kmp_uint32)) % \
2745  CACHE_LINE))
2746 #else
2747 #define KMP_INLINE_ARGV_BYTES \
2748  (2 * CACHE_LINE - ((3 * KMP_PTR_SKIP + 2 * sizeof(int)) % CACHE_LINE))
2749 #endif
2750 #define KMP_INLINE_ARGV_ENTRIES (int)(KMP_INLINE_ARGV_BYTES / KMP_PTR_SKIP)
2751 
2752 typedef struct KMP_ALIGN_CACHE kmp_base_team {
2753  // Synchronization Data
2754  // ---------------------------------------------------------------------------
2755  KMP_ALIGN_CACHE kmp_ordered_team_t t_ordered;
2756  kmp_balign_team_t t_bar[bs_last_barrier];
2757  std::atomic<int> t_construct; // count of single directive encountered by team
2758  char pad[sizeof(kmp_lock_t)]; // padding to maintain performance on big iron
2759 
2760  // [0] - parallel / [1] - worksharing task reduction data shared by taskgroups
2761  std::atomic<void *> t_tg_reduce_data[2]; // to support task modifier
2762  std::atomic<int> t_tg_fini_counter[2]; // sync end of task reductions
2763 
2764  // Primary thread only
2765  // ---------------------------------------------------------------------------
2766  KMP_ALIGN_CACHE int t_master_tid; // tid of primary thread in parent team
2767  int t_master_this_cons; // "this_construct" single counter of primary thread
2768  // in parent team
2769  ident_t *t_ident; // if volatile, have to change too much other crud to
2770  // volatile too
2771  kmp_team_p *t_parent; // parent team
2772  kmp_team_p *t_next_pool; // next free team in the team pool
2773  kmp_disp_t *t_dispatch; // thread's dispatch data
2774  kmp_task_team_t *t_task_team[2]; // Task team struct; switch between 2
2775  kmp_proc_bind_t t_proc_bind; // bind type for par region
2776 #if USE_ITT_BUILD
2777  kmp_uint64 t_region_time; // region begin timestamp
2778 #endif /* USE_ITT_BUILD */
2779 
2780  // Primary thread write, workers read
2781  // --------------------------------------------------------------------------
2782  KMP_ALIGN_CACHE void **t_argv;
2783  int t_argc;
2784  int t_nproc; // number of threads in team
2785  microtask_t t_pkfn;
2786  launch_t t_invoke; // procedure to launch the microtask
2787 
2788 #if OMPT_SUPPORT
2789  ompt_team_info_t ompt_team_info;
2790  ompt_lw_taskteam_t *ompt_serialized_team_info;
2791 #endif
2792 
2793 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
2794  kmp_int8 t_fp_control_saved;
2795  kmp_int8 t_pad2b;
2796  kmp_int16 t_x87_fpu_control_word; // FP control regs
2797  kmp_uint32 t_mxcsr;
2798 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
2799 
2800  void *t_inline_argv[KMP_INLINE_ARGV_ENTRIES];
2801 
2802  KMP_ALIGN_CACHE kmp_info_t **t_threads;
2803  kmp_taskdata_t
2804  *t_implicit_task_taskdata; // Taskdata for the thread's implicit task
2805  int t_level; // nested parallel level
2806 
2807  KMP_ALIGN_CACHE int t_max_argc;
2808  int t_max_nproc; // max threads this team can handle (dynamically expandable)
2809  int t_serialized; // levels deep of serialized teams
2810  dispatch_shared_info_t *t_disp_buffer; // buffers for dispatch system
2811  int t_id; // team's id, assigned by debugger.
2812  int t_active_level; // nested active parallel level
2813  kmp_r_sched_t t_sched; // run-time schedule for the team
2814 #if KMP_AFFINITY_SUPPORTED
2815  int t_first_place; // first & last place in parent thread's partition.
2816  int t_last_place; // Restore these values to primary thread after par region.
2817 #endif // KMP_AFFINITY_SUPPORTED
2818  int t_display_affinity;
2819  int t_size_changed; // team size was changed?: 0: no, 1: yes, -1: changed via
2820  // omp_set_num_threads() call
2821  omp_allocator_handle_t t_def_allocator; /* default allocator */
2822 
2823 // Read/write by workers as well
2824 #if (KMP_ARCH_X86 || KMP_ARCH_X86_64)
2825  // Using CACHE_LINE=64 reduces memory footprint, but causes a big perf
2826  // regression of epcc 'parallel' and 'barrier' on fxe256lin01. This extra
2827  // padding serves to fix the performance of epcc 'parallel' and 'barrier' when
2828  // CACHE_LINE=64. TODO: investigate more and get rid if this padding.
2829  char dummy_padding[1024];
2830 #endif
2831  // Internal control stack for additional nested teams.
2832  KMP_ALIGN_CACHE kmp_internal_control_t *t_control_stack_top;
2833  // for SERIALIZED teams nested 2 or more levels deep
2834  // typed flag to store request state of cancellation
2835  std::atomic<kmp_int32> t_cancel_request;
2836  int t_master_active; // save on fork, restore on join
2837  void *t_copypriv_data; // team specific pointer to copyprivate data array
2838 #if KMP_OS_WINDOWS
2839  std::atomic<kmp_uint32> t_copyin_counter;
2840 #endif
2841 #if USE_ITT_BUILD
2842  void *t_stack_id; // team specific stack stitching id (for ittnotify)
2843 #endif /* USE_ITT_BUILD */
2844  distributedBarrier *b; // Distributed barrier data associated with team
2845 } kmp_base_team_t;
2846 
2847 union KMP_ALIGN_CACHE kmp_team {
2848  kmp_base_team_t t;
2849  double t_align; /* use worst case alignment */
2850  char t_pad[KMP_PAD(kmp_base_team_t, CACHE_LINE)];
2851 };
2852 
2853 typedef union KMP_ALIGN_CACHE kmp_time_global {
2854  double dt_align; /* use worst case alignment */
2855  char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
2856  kmp_base_data_t dt;
2857 } kmp_time_global_t;
2858 
2859 typedef struct kmp_base_global {
2860  /* cache-aligned */
2861  kmp_time_global_t g_time;
2862 
2863  /* non cache-aligned */
2864  volatile int g_abort;
2865  volatile int g_done;
2866 
2867  int g_dynamic;
2868  enum dynamic_mode g_dynamic_mode;
2869 } kmp_base_global_t;
2870 
2871 typedef union KMP_ALIGN_CACHE kmp_global {
2872  kmp_base_global_t g;
2873  double g_align; /* use worst case alignment */
2874  char g_pad[KMP_PAD(kmp_base_global_t, CACHE_LINE)];
2875 } kmp_global_t;
2876 
2877 typedef struct kmp_base_root {
2878  // TODO: GEH - combine r_active with r_in_parallel then r_active ==
2879  // (r_in_parallel>= 0)
2880  // TODO: GEH - then replace r_active with t_active_levels if we can to reduce
2881  // the synch overhead or keeping r_active
2882  volatile int r_active; /* TRUE if some region in a nest has > 1 thread */
2883  // keeps a count of active parallel regions per root
2884  std::atomic<int> r_in_parallel;
2885  // GEH: This is misnamed, should be r_active_levels
2886  kmp_team_t *r_root_team;
2887  kmp_team_t *r_hot_team;
2888  kmp_info_t *r_uber_thread;
2889  kmp_lock_t r_begin_lock;
2890  volatile int r_begin;
2891  int r_blocktime; /* blocktime for this root and descendants */
2892 #if KMP_AFFINITY_SUPPORTED
2893  int r_affinity_assigned;
2894 #endif // KMP_AFFINITY_SUPPORTED
2895 } kmp_base_root_t;
2896 
2897 typedef union KMP_ALIGN_CACHE kmp_root {
2898  kmp_base_root_t r;
2899  double r_align; /* use worst case alignment */
2900  char r_pad[KMP_PAD(kmp_base_root_t, CACHE_LINE)];
2901 } kmp_root_t;
2902 
2903 struct fortran_inx_info {
2904  kmp_int32 data;
2905 };
2906 
2907 /* ------------------------------------------------------------------------ */
2908 
2909 extern int __kmp_settings;
2910 extern int __kmp_duplicate_library_ok;
2911 #if USE_ITT_BUILD
2912 extern int __kmp_forkjoin_frames;
2913 extern int __kmp_forkjoin_frames_mode;
2914 #endif
2915 extern PACKED_REDUCTION_METHOD_T __kmp_force_reduction_method;
2916 extern int __kmp_determ_red;
2917 
2918 #ifdef KMP_DEBUG
2919 extern int kmp_a_debug;
2920 extern int kmp_b_debug;
2921 extern int kmp_c_debug;
2922 extern int kmp_d_debug;
2923 extern int kmp_e_debug;
2924 extern int kmp_f_debug;
2925 #endif /* KMP_DEBUG */
2926 
2927 /* For debug information logging using rotating buffer */
2928 #define KMP_DEBUG_BUF_LINES_INIT 512
2929 #define KMP_DEBUG_BUF_LINES_MIN 1
2930 
2931 #define KMP_DEBUG_BUF_CHARS_INIT 128
2932 #define KMP_DEBUG_BUF_CHARS_MIN 2
2933 
2934 extern int
2935  __kmp_debug_buf; /* TRUE means use buffer, FALSE means print to stderr */
2936 extern int __kmp_debug_buf_lines; /* How many lines of debug stored in buffer */
2937 extern int
2938  __kmp_debug_buf_chars; /* How many characters allowed per line in buffer */
2939 extern int __kmp_debug_buf_atomic; /* TRUE means use atomic update of buffer
2940  entry pointer */
2941 
2942 extern char *__kmp_debug_buffer; /* Debug buffer itself */
2943 extern std::atomic<int> __kmp_debug_count; /* Counter for number of lines
2944  printed in buffer so far */
2945 extern int __kmp_debug_buf_warn_chars; /* Keep track of char increase
2946  recommended in warnings */
2947 /* end rotating debug buffer */
2948 
2949 #ifdef KMP_DEBUG
2950 extern int __kmp_par_range; /* +1 => only go par for constructs in range */
2951 
2952 #define KMP_PAR_RANGE_ROUTINE_LEN 1024
2953 extern char __kmp_par_range_routine[KMP_PAR_RANGE_ROUTINE_LEN];
2954 #define KMP_PAR_RANGE_FILENAME_LEN 1024
2955 extern char __kmp_par_range_filename[KMP_PAR_RANGE_FILENAME_LEN];
2956 extern int __kmp_par_range_lb;
2957 extern int __kmp_par_range_ub;
2958 #endif
2959 
2960 /* For printing out dynamic storage map for threads and teams */
2961 extern int
2962  __kmp_storage_map; /* True means print storage map for threads and teams */
2963 extern int __kmp_storage_map_verbose; /* True means storage map includes
2964  placement info */
2965 extern int __kmp_storage_map_verbose_specified;
2966 
2967 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
2968 extern kmp_cpuinfo_t __kmp_cpuinfo;
2969 #endif
2970 
2971 extern volatile int __kmp_init_serial;
2972 extern volatile int __kmp_init_gtid;
2973 extern volatile int __kmp_init_common;
2974 extern volatile int __kmp_init_middle;
2975 extern volatile int __kmp_init_parallel;
2976 #if KMP_USE_MONITOR
2977 extern volatile int __kmp_init_monitor;
2978 #endif
2979 extern volatile int __kmp_init_user_locks;
2980 extern volatile int __kmp_init_hidden_helper_threads;
2981 extern int __kmp_init_counter;
2982 extern int __kmp_root_counter;
2983 extern int __kmp_version;
2984 
2985 /* list of address of allocated caches for commons */
2986 extern kmp_cached_addr_t *__kmp_threadpriv_cache_list;
2987 
2988 /* Barrier algorithm types and options */
2989 extern kmp_uint32 __kmp_barrier_gather_bb_dflt;
2990 extern kmp_uint32 __kmp_barrier_release_bb_dflt;
2991 extern kmp_bar_pat_e __kmp_barrier_gather_pat_dflt;
2992 extern kmp_bar_pat_e __kmp_barrier_release_pat_dflt;
2993 extern kmp_uint32 __kmp_barrier_gather_branch_bits[bs_last_barrier];
2994 extern kmp_uint32 __kmp_barrier_release_branch_bits[bs_last_barrier];
2995 extern kmp_bar_pat_e __kmp_barrier_gather_pattern[bs_last_barrier];
2996 extern kmp_bar_pat_e __kmp_barrier_release_pattern[bs_last_barrier];
2997 extern char const *__kmp_barrier_branch_bit_env_name[bs_last_barrier];
2998 extern char const *__kmp_barrier_pattern_env_name[bs_last_barrier];
2999 extern char const *__kmp_barrier_type_name[bs_last_barrier];
3000 extern char const *__kmp_barrier_pattern_name[bp_last_bar];
3001 
3002 /* Global Locks */
3003 extern kmp_bootstrap_lock_t __kmp_initz_lock; /* control initialization */
3004 extern kmp_bootstrap_lock_t __kmp_forkjoin_lock; /* control fork/join access */
3005 extern kmp_bootstrap_lock_t __kmp_task_team_lock;
3006 extern kmp_bootstrap_lock_t
3007  __kmp_exit_lock; /* exit() is not always thread-safe */
3008 #if KMP_USE_MONITOR
3009 extern kmp_bootstrap_lock_t
3010  __kmp_monitor_lock; /* control monitor thread creation */
3011 #endif
3012 extern kmp_bootstrap_lock_t
3013  __kmp_tp_cached_lock; /* used for the hack to allow threadprivate cache and
3014  __kmp_threads expansion to co-exist */
3015 
3016 extern kmp_lock_t __kmp_global_lock; /* control OS/global access */
3017 extern kmp_queuing_lock_t __kmp_dispatch_lock; /* control dispatch access */
3018 extern kmp_lock_t __kmp_debug_lock; /* control I/O access for KMP_DEBUG */
3019 
3020 extern enum library_type __kmp_library;
3021 
3022 extern enum sched_type __kmp_sched; /* default runtime scheduling */
3023 extern enum sched_type __kmp_static; /* default static scheduling method */
3024 extern enum sched_type __kmp_guided; /* default guided scheduling method */
3025 extern enum sched_type __kmp_auto; /* default auto scheduling method */
3026 extern int __kmp_chunk; /* default runtime chunk size */
3027 extern int __kmp_force_monotonic; /* whether monotonic scheduling forced */
3028 
3029 extern size_t __kmp_stksize; /* stack size per thread */
3030 #if KMP_USE_MONITOR
3031 extern size_t __kmp_monitor_stksize; /* stack size for monitor thread */
3032 #endif
3033 extern size_t __kmp_stkoffset; /* stack offset per thread */
3034 extern int __kmp_stkpadding; /* Should we pad root thread(s) stack */
3035 
3036 extern size_t
3037  __kmp_malloc_pool_incr; /* incremental size of pool for kmp_malloc() */
3038 extern int __kmp_env_stksize; /* was KMP_STACKSIZE specified? */
3039 extern int __kmp_env_blocktime; /* was KMP_BLOCKTIME specified? */
3040 extern int __kmp_env_checks; /* was KMP_CHECKS specified? */
3041 extern int __kmp_env_consistency_check; // was KMP_CONSISTENCY_CHECK specified?
3042 extern int __kmp_generate_warnings; /* should we issue warnings? */
3043 extern int __kmp_reserve_warn; /* have we issued reserve_threads warning? */
3044 
3045 #ifdef DEBUG_SUSPEND
3046 extern int __kmp_suspend_count; /* count inside __kmp_suspend_template() */
3047 #endif
3048 
3049 extern kmp_int32 __kmp_use_yield;
3050 extern kmp_int32 __kmp_use_yield_exp_set;
3051 extern kmp_uint32 __kmp_yield_init;
3052 extern kmp_uint32 __kmp_yield_next;
3053 
3054 /* ------------------------------------------------------------------------- */
3055 extern int __kmp_allThreadsSpecified;
3056 
3057 extern size_t __kmp_align_alloc;
3058 /* following data protected by initialization routines */
3059 extern int __kmp_xproc; /* number of processors in the system */
3060 extern int __kmp_avail_proc; /* number of processors available to the process */
3061 extern size_t __kmp_sys_min_stksize; /* system-defined minimum stack size */
3062 extern int __kmp_sys_max_nth; /* system-imposed maximum number of threads */
3063 // maximum total number of concurrently-existing threads on device
3064 extern int __kmp_max_nth;
3065 // maximum total number of concurrently-existing threads in a contention group
3066 extern int __kmp_cg_max_nth;
3067 extern int __kmp_teams_max_nth; // max threads used in a teams construct
3068 extern int __kmp_threads_capacity; /* capacity of the arrays __kmp_threads and
3069  __kmp_root */
3070 extern int __kmp_dflt_team_nth; /* default number of threads in a parallel
3071  region a la OMP_NUM_THREADS */
3072 extern int __kmp_dflt_team_nth_ub; /* upper bound on "" determined at serial
3073  initialization */
3074 extern int __kmp_tp_capacity; /* capacity of __kmp_threads if threadprivate is
3075  used (fixed) */
3076 extern int __kmp_tp_cached; /* whether threadprivate cache has been created
3077  (__kmpc_threadprivate_cached()) */
3078 extern int __kmp_dflt_blocktime; /* number of milliseconds to wait before
3079  blocking (env setting) */
3080 #if KMP_USE_MONITOR
3081 extern int
3082  __kmp_monitor_wakeups; /* number of times monitor wakes up per second */
3083 extern int __kmp_bt_intervals; /* number of monitor timestamp intervals before
3084  blocking */
3085 #endif
3086 #ifdef KMP_ADJUST_BLOCKTIME
3087 extern int __kmp_zero_bt; /* whether blocktime has been forced to zero */
3088 #endif /* KMP_ADJUST_BLOCKTIME */
3089 #ifdef KMP_DFLT_NTH_CORES
3090 extern int __kmp_ncores; /* Total number of cores for threads placement */
3091 #endif
3092 /* Number of millisecs to delay on abort for Intel(R) VTune(TM) tools */
3093 extern int __kmp_abort_delay;
3094 
3095 extern int __kmp_need_register_atfork_specified;
3096 extern int __kmp_need_register_atfork; /* At initialization, call pthread_atfork
3097  to install fork handler */
3098 extern int __kmp_gtid_mode; /* Method of getting gtid, values:
3099  0 - not set, will be set at runtime
3100  1 - using stack search
3101  2 - dynamic TLS (pthread_getspecific(Linux* OS/OS
3102  X*) or TlsGetValue(Windows* OS))
3103  3 - static TLS (__declspec(thread) __kmp_gtid),
3104  Linux* OS .so only. */
3105 extern int
3106  __kmp_adjust_gtid_mode; /* If true, adjust method based on #threads */
3107 #ifdef KMP_TDATA_GTID
3108 extern KMP_THREAD_LOCAL int __kmp_gtid;
3109 #endif
3110 extern int __kmp_tls_gtid_min; /* #threads below which use sp search for gtid */
3111 extern int __kmp_foreign_tp; // If true, separate TP var for each foreign thread
3112 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
3113 extern int __kmp_inherit_fp_control; // copy fp creg(s) parent->workers at fork
3114 extern kmp_int16 __kmp_init_x87_fpu_control_word; // init thread's FP ctrl reg
3115 extern kmp_uint32 __kmp_init_mxcsr; /* init thread's mxscr */
3116 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
3117 
3118 // max_active_levels for nested parallelism enabled by default via
3119 // OMP_MAX_ACTIVE_LEVELS, OMP_NESTED, OMP_NUM_THREADS, and OMP_PROC_BIND
3120 extern int __kmp_dflt_max_active_levels;
3121 // Indicates whether value of __kmp_dflt_max_active_levels was already
3122 // explicitly set by OMP_MAX_ACTIVE_LEVELS or OMP_NESTED=false
3123 extern bool __kmp_dflt_max_active_levels_set;
3124 extern int __kmp_dispatch_num_buffers; /* max possible dynamic loops in
3125  concurrent execution per team */
3126 #if KMP_NESTED_HOT_TEAMS
3127 extern int __kmp_hot_teams_mode;
3128 extern int __kmp_hot_teams_max_level;
3129 #endif
3130 
3131 #if KMP_OS_LINUX
3132 extern enum clock_function_type __kmp_clock_function;
3133 extern int __kmp_clock_function_param;
3134 #endif /* KMP_OS_LINUX */
3135 
3136 #if KMP_MIC_SUPPORTED
3137 extern enum mic_type __kmp_mic_type;
3138 #endif
3139 
3140 #ifdef USE_LOAD_BALANCE
3141 extern double __kmp_load_balance_interval; // load balance algorithm interval
3142 #endif /* USE_LOAD_BALANCE */
3143 
3144 // OpenMP 3.1 - Nested num threads array
3145 typedef struct kmp_nested_nthreads_t {
3146  int *nth;
3147  int size;
3148  int used;
3149 } kmp_nested_nthreads_t;
3150 
3151 extern kmp_nested_nthreads_t __kmp_nested_nth;
3152 
3153 #if KMP_USE_ADAPTIVE_LOCKS
3154 
3155 // Parameters for the speculative lock backoff system.
3156 struct kmp_adaptive_backoff_params_t {
3157  // Number of soft retries before it counts as a hard retry.
3158  kmp_uint32 max_soft_retries;
3159  // Badness is a bit mask : 0,1,3,7,15,... on each hard failure we move one to
3160  // the right
3161  kmp_uint32 max_badness;
3162 };
3163 
3164 extern kmp_adaptive_backoff_params_t __kmp_adaptive_backoff_params;
3165 
3166 #if KMP_DEBUG_ADAPTIVE_LOCKS
3167 extern const char *__kmp_speculative_statsfile;
3168 #endif
3169 
3170 #endif // KMP_USE_ADAPTIVE_LOCKS
3171 
3172 extern int __kmp_display_env; /* TRUE or FALSE */
3173 extern int __kmp_display_env_verbose; /* TRUE if OMP_DISPLAY_ENV=VERBOSE */
3174 extern int __kmp_omp_cancellation; /* TRUE or FALSE */
3175 extern int __kmp_nteams;
3176 extern int __kmp_teams_thread_limit;
3177 
3178 /* ------------------------------------------------------------------------- */
3179 
3180 /* the following are protected by the fork/join lock */
3181 /* write: lock read: anytime */
3182 extern kmp_info_t **__kmp_threads; /* Descriptors for the threads */
3183 /* read/write: lock */
3184 extern volatile kmp_team_t *__kmp_team_pool;
3185 extern volatile kmp_info_t *__kmp_thread_pool;
3186 extern kmp_info_t *__kmp_thread_pool_insert_pt;
3187 
3188 // total num threads reachable from some root thread including all root threads
3189 extern volatile int __kmp_nth;
3190 /* total number of threads reachable from some root thread including all root
3191  threads, and those in the thread pool */
3192 extern volatile int __kmp_all_nth;
3193 extern std::atomic<int> __kmp_thread_pool_active_nth;
3194 
3195 extern kmp_root_t **__kmp_root; /* root of thread hierarchy */
3196 /* end data protected by fork/join lock */
3197 /* ------------------------------------------------------------------------- */
3198 
3199 #define __kmp_get_gtid() __kmp_get_global_thread_id()
3200 #define __kmp_entry_gtid() __kmp_get_global_thread_id_reg()
3201 #define __kmp_get_tid() (__kmp_tid_from_gtid(__kmp_get_gtid()))
3202 #define __kmp_get_team() (__kmp_threads[(__kmp_get_gtid())]->th.th_team)
3203 #define __kmp_get_thread() (__kmp_thread_from_gtid(__kmp_get_gtid()))
3204 
3205 // AT: Which way is correct?
3206 // AT: 1. nproc = __kmp_threads[ ( gtid ) ] -> th.th_team -> t.t_nproc;
3207 // AT: 2. nproc = __kmp_threads[ ( gtid ) ] -> th.th_team_nproc;
3208 #define __kmp_get_team_num_threads(gtid) \
3209  (__kmp_threads[(gtid)]->th.th_team->t.t_nproc)
3210 
3211 static inline bool KMP_UBER_GTID(int gtid) {
3212  KMP_DEBUG_ASSERT(gtid >= KMP_GTID_MIN);
3213  KMP_DEBUG_ASSERT(gtid < __kmp_threads_capacity);
3214  return (gtid >= 0 && __kmp_root[gtid] && __kmp_threads[gtid] &&
3215  __kmp_threads[gtid] == __kmp_root[gtid]->r.r_uber_thread);
3216 }
3217 
3218 static inline int __kmp_tid_from_gtid(int gtid) {
3219  KMP_DEBUG_ASSERT(gtid >= 0);
3220  return __kmp_threads[gtid]->th.th_info.ds.ds_tid;
3221 }
3222 
3223 static inline int __kmp_gtid_from_tid(int tid, const kmp_team_t *team) {
3224  KMP_DEBUG_ASSERT(tid >= 0 && team);
3225  return team->t.t_threads[tid]->th.th_info.ds.ds_gtid;
3226 }
3227 
3228 static inline int __kmp_gtid_from_thread(const kmp_info_t *thr) {
3229  KMP_DEBUG_ASSERT(thr);
3230  return thr->th.th_info.ds.ds_gtid;
3231 }
3232 
3233 static inline kmp_info_t *__kmp_thread_from_gtid(int gtid) {
3234  KMP_DEBUG_ASSERT(gtid >= 0);
3235  return __kmp_threads[gtid];
3236 }
3237 
3238 static inline kmp_team_t *__kmp_team_from_gtid(int gtid) {
3239  KMP_DEBUG_ASSERT(gtid >= 0);
3240  return __kmp_threads[gtid]->th.th_team;
3241 }
3242 
3243 static inline void __kmp_assert_valid_gtid(kmp_int32 gtid) {
3244  if (UNLIKELY(gtid < 0 || gtid >= __kmp_threads_capacity))
3245  KMP_FATAL(ThreadIdentInvalid);
3246 }
3247 
3248 #if KMP_HAVE_MWAIT || KMP_HAVE_UMWAIT
3249 extern int __kmp_user_level_mwait; // TRUE or FALSE; from KMP_USER_LEVEL_MWAIT
3250 extern int __kmp_umwait_enabled; // Runtime check if user-level mwait enabled
3251 extern int __kmp_mwait_enabled; // Runtime check if ring3 mwait is enabled
3252 extern int __kmp_mwait_hints; // Hints to pass in to mwait
3253 #endif
3254 
3255 /* ------------------------------------------------------------------------- */
3256 
3257 extern kmp_global_t __kmp_global; /* global status */
3258 
3259 extern kmp_info_t __kmp_monitor;
3260 // For Debugging Support Library
3261 extern std::atomic<kmp_int32> __kmp_team_counter;
3262 // For Debugging Support Library
3263 extern std::atomic<kmp_int32> __kmp_task_counter;
3264 
3265 #if USE_DEBUGGER
3266 #define _KMP_GEN_ID(counter) \
3267  (__kmp_debugging ? KMP_ATOMIC_INC(&counter) + 1 : ~0)
3268 #else
3269 #define _KMP_GEN_ID(counter) (~0)
3270 #endif /* USE_DEBUGGER */
3271 
3272 #define KMP_GEN_TASK_ID() _KMP_GEN_ID(__kmp_task_counter)
3273 #define KMP_GEN_TEAM_ID() _KMP_GEN_ID(__kmp_team_counter)
3274 
3275 /* ------------------------------------------------------------------------ */
3276 
3277 extern void __kmp_print_storage_map_gtid(int gtid, void *p1, void *p2,
3278  size_t size, char const *format, ...);
3279 
3280 extern void __kmp_serial_initialize(void);
3281 extern void __kmp_middle_initialize(void);
3282 extern void __kmp_parallel_initialize(void);
3283 
3284 extern void __kmp_internal_begin(void);
3285 extern void __kmp_internal_end_library(int gtid);
3286 extern void __kmp_internal_end_thread(int gtid);
3287 extern void __kmp_internal_end_atexit(void);
3288 extern void __kmp_internal_end_dtor(void);
3289 extern void __kmp_internal_end_dest(void *);
3290 
3291 extern int __kmp_register_root(int initial_thread);
3292 extern void __kmp_unregister_root(int gtid);
3293 extern void __kmp_unregister_library(void); // called by __kmp_internal_end()
3294 
3295 extern int __kmp_ignore_mppbeg(void);
3296 extern int __kmp_ignore_mppend(void);
3297 
3298 extern int __kmp_enter_single(int gtid, ident_t *id_ref, int push_ws);
3299 extern void __kmp_exit_single(int gtid);
3300 
3301 extern void __kmp_parallel_deo(int *gtid_ref, int *cid_ref, ident_t *loc_ref);
3302 extern void __kmp_parallel_dxo(int *gtid_ref, int *cid_ref, ident_t *loc_ref);
3303 
3304 #ifdef USE_LOAD_BALANCE
3305 extern int __kmp_get_load_balance(int);
3306 #endif
3307 
3308 extern int __kmp_get_global_thread_id(void);
3309 extern int __kmp_get_global_thread_id_reg(void);
3310 extern void __kmp_exit_thread(int exit_status);
3311 extern void __kmp_abort(char const *format, ...);
3312 extern void __kmp_abort_thread(void);
3313 KMP_NORETURN extern void __kmp_abort_process(void);
3314 extern void __kmp_warn(char const *format, ...);
3315 
3316 extern void __kmp_set_num_threads(int new_nth, int gtid);
3317 
3318 // Returns current thread (pointer to kmp_info_t). Current thread *must* be
3319 // registered.
3320 static inline kmp_info_t *__kmp_entry_thread() {
3321  int gtid = __kmp_entry_gtid();
3322 
3323  return __kmp_threads[gtid];
3324 }
3325 
3326 extern void __kmp_set_max_active_levels(int gtid, int new_max_active_levels);
3327 extern int __kmp_get_max_active_levels(int gtid);
3328 extern int __kmp_get_ancestor_thread_num(int gtid, int level);
3329 extern int __kmp_get_team_size(int gtid, int level);
3330 extern void __kmp_set_schedule(int gtid, kmp_sched_t new_sched, int chunk);
3331 extern void __kmp_get_schedule(int gtid, kmp_sched_t *sched, int *chunk);
3332 
3333 extern unsigned short __kmp_get_random(kmp_info_t *thread);
3334 extern void __kmp_init_random(kmp_info_t *thread);
3335 
3336 extern kmp_r_sched_t __kmp_get_schedule_global(void);
3337 extern void __kmp_adjust_num_threads(int new_nproc);
3338 extern void __kmp_check_stksize(size_t *val);
3339 
3340 extern void *___kmp_allocate(size_t size KMP_SRC_LOC_DECL);
3341 extern void *___kmp_page_allocate(size_t size KMP_SRC_LOC_DECL);
3342 extern void ___kmp_free(void *ptr KMP_SRC_LOC_DECL);
3343 #define __kmp_allocate(size) ___kmp_allocate((size)KMP_SRC_LOC_CURR)
3344 #define __kmp_page_allocate(size) ___kmp_page_allocate((size)KMP_SRC_LOC_CURR)
3345 #define __kmp_free(ptr) ___kmp_free((ptr)KMP_SRC_LOC_CURR)
3346 
3347 #if USE_FAST_MEMORY
3348 extern void *___kmp_fast_allocate(kmp_info_t *this_thr,
3349  size_t size KMP_SRC_LOC_DECL);
3350 extern void ___kmp_fast_free(kmp_info_t *this_thr, void *ptr KMP_SRC_LOC_DECL);
3351 extern void __kmp_free_fast_memory(kmp_info_t *this_thr);
3352 extern void __kmp_initialize_fast_memory(kmp_info_t *this_thr);
3353 #define __kmp_fast_allocate(this_thr, size) \
3354  ___kmp_fast_allocate((this_thr), (size)KMP_SRC_LOC_CURR)
3355 #define __kmp_fast_free(this_thr, ptr) \
3356  ___kmp_fast_free((this_thr), (ptr)KMP_SRC_LOC_CURR)
3357 #endif
3358 
3359 extern void *___kmp_thread_malloc(kmp_info_t *th, size_t size KMP_SRC_LOC_DECL);
3360 extern void *___kmp_thread_calloc(kmp_info_t *th, size_t nelem,
3361  size_t elsize KMP_SRC_LOC_DECL);
3362 extern void *___kmp_thread_realloc(kmp_info_t *th, void *ptr,
3363  size_t size KMP_SRC_LOC_DECL);
3364 extern void ___kmp_thread_free(kmp_info_t *th, void *ptr KMP_SRC_LOC_DECL);
3365 #define __kmp_thread_malloc(th, size) \
3366  ___kmp_thread_malloc((th), (size)KMP_SRC_LOC_CURR)
3367 #define __kmp_thread_calloc(th, nelem, elsize) \
3368  ___kmp_thread_calloc((th), (nelem), (elsize)KMP_SRC_LOC_CURR)
3369 #define __kmp_thread_realloc(th, ptr, size) \
3370  ___kmp_thread_realloc((th), (ptr), (size)KMP_SRC_LOC_CURR)
3371 #define __kmp_thread_free(th, ptr) \
3372  ___kmp_thread_free((th), (ptr)KMP_SRC_LOC_CURR)
3373 
3374 #define KMP_INTERNAL_MALLOC(sz) malloc(sz)
3375 #define KMP_INTERNAL_FREE(p) free(p)
3376 #define KMP_INTERNAL_REALLOC(p, sz) realloc((p), (sz))
3377 #define KMP_INTERNAL_CALLOC(n, sz) calloc((n), (sz))
3378 
3379 extern void __kmp_push_num_threads(ident_t *loc, int gtid, int num_threads);
3380 
3381 extern void __kmp_push_proc_bind(ident_t *loc, int gtid,
3382  kmp_proc_bind_t proc_bind);
3383 extern void __kmp_push_num_teams(ident_t *loc, int gtid, int num_teams,
3384  int num_threads);
3385 extern void __kmp_push_num_teams_51(ident_t *loc, int gtid, int num_teams_lb,
3386  int num_teams_ub, int num_threads);
3387 
3388 extern void __kmp_yield();
3389 
3390 extern void __kmpc_dispatch_init_4(ident_t *loc, kmp_int32 gtid,
3391  enum sched_type schedule, kmp_int32 lb,
3392  kmp_int32 ub, kmp_int32 st, kmp_int32 chunk);
3393 extern void __kmpc_dispatch_init_4u(ident_t *loc, kmp_int32 gtid,
3394  enum sched_type schedule, kmp_uint32 lb,
3395  kmp_uint32 ub, kmp_int32 st,
3396  kmp_int32 chunk);
3397 extern void __kmpc_dispatch_init_8(ident_t *loc, kmp_int32 gtid,
3398  enum sched_type schedule, kmp_int64 lb,
3399  kmp_int64 ub, kmp_int64 st, kmp_int64 chunk);
3400 extern void __kmpc_dispatch_init_8u(ident_t *loc, kmp_int32 gtid,
3401  enum sched_type schedule, kmp_uint64 lb,
3402  kmp_uint64 ub, kmp_int64 st,
3403  kmp_int64 chunk);
3404 
3405 extern int __kmpc_dispatch_next_4(ident_t *loc, kmp_int32 gtid,
3406  kmp_int32 *p_last, kmp_int32 *p_lb,
3407  kmp_int32 *p_ub, kmp_int32 *p_st);
3408 extern int __kmpc_dispatch_next_4u(ident_t *loc, kmp_int32 gtid,
3409  kmp_int32 *p_last, kmp_uint32 *p_lb,
3410  kmp_uint32 *p_ub, kmp_int32 *p_st);
3411 extern int __kmpc_dispatch_next_8(ident_t *loc, kmp_int32 gtid,
3412  kmp_int32 *p_last, kmp_int64 *p_lb,
3413  kmp_int64 *p_ub, kmp_int64 *p_st);
3414 extern int __kmpc_dispatch_next_8u(ident_t *loc, kmp_int32 gtid,
3415  kmp_int32 *p_last, kmp_uint64 *p_lb,
3416  kmp_uint64 *p_ub, kmp_int64 *p_st);
3417 
3418 extern void __kmpc_dispatch_fini_4(ident_t *loc, kmp_int32 gtid);
3419 extern void __kmpc_dispatch_fini_8(ident_t *loc, kmp_int32 gtid);
3420 extern void __kmpc_dispatch_fini_4u(ident_t *loc, kmp_int32 gtid);
3421 extern void __kmpc_dispatch_fini_8u(ident_t *loc, kmp_int32 gtid);
3422 
3423 #ifdef KMP_GOMP_COMPAT
3424 
3425 extern void __kmp_aux_dispatch_init_4(ident_t *loc, kmp_int32 gtid,
3426  enum sched_type schedule, kmp_int32 lb,
3427  kmp_int32 ub, kmp_int32 st,
3428  kmp_int32 chunk, int push_ws);
3429 extern void __kmp_aux_dispatch_init_4u(ident_t *loc, kmp_int32 gtid,
3430  enum sched_type schedule, kmp_uint32 lb,
3431  kmp_uint32 ub, kmp_int32 st,
3432  kmp_int32 chunk, int push_ws);
3433 extern void __kmp_aux_dispatch_init_8(ident_t *loc, kmp_int32 gtid,
3434  enum sched_type schedule, kmp_int64 lb,
3435  kmp_int64 ub, kmp_int64 st,
3436  kmp_int64 chunk, int push_ws);
3437 extern void __kmp_aux_dispatch_init_8u(ident_t *loc, kmp_int32 gtid,
3438  enum sched_type schedule, kmp_uint64 lb,
3439  kmp_uint64 ub, kmp_int64 st,
3440  kmp_int64 chunk, int push_ws);
3441 extern void __kmp_aux_dispatch_fini_chunk_4(ident_t *loc, kmp_int32 gtid);
3442 extern void __kmp_aux_dispatch_fini_chunk_8(ident_t *loc, kmp_int32 gtid);
3443 extern void __kmp_aux_dispatch_fini_chunk_4u(ident_t *loc, kmp_int32 gtid);
3444 extern void __kmp_aux_dispatch_fini_chunk_8u(ident_t *loc, kmp_int32 gtid);
3445 
3446 #endif /* KMP_GOMP_COMPAT */
3447 
3448 extern kmp_uint32 __kmp_eq_4(kmp_uint32 value, kmp_uint32 checker);
3449 extern kmp_uint32 __kmp_neq_4(kmp_uint32 value, kmp_uint32 checker);
3450 extern kmp_uint32 __kmp_lt_4(kmp_uint32 value, kmp_uint32 checker);
3451 extern kmp_uint32 __kmp_ge_4(kmp_uint32 value, kmp_uint32 checker);
3452 extern kmp_uint32 __kmp_le_4(kmp_uint32 value, kmp_uint32 checker);
3453 extern kmp_uint32 __kmp_wait_4(kmp_uint32 volatile *spinner, kmp_uint32 checker,
3454  kmp_uint32 (*pred)(kmp_uint32, kmp_uint32),
3455  void *obj);
3456 extern void __kmp_wait_4_ptr(void *spinner, kmp_uint32 checker,
3457  kmp_uint32 (*pred)(void *, kmp_uint32), void *obj);
3458 
3459 extern void __kmp_wait_64(kmp_info_t *this_thr, kmp_flag_64<> *flag,
3460  int final_spin
3461 #if USE_ITT_BUILD
3462  ,
3463  void *itt_sync_obj
3464 #endif
3465 );
3466 extern void __kmp_release_64(kmp_flag_64<> *flag);
3467 
3468 extern void __kmp_infinite_loop(void);
3469 
3470 extern void __kmp_cleanup(void);
3471 
3472 #if KMP_HANDLE_SIGNALS
3473 extern int __kmp_handle_signals;
3474 extern void __kmp_install_signals(int parallel_init);
3475 extern void __kmp_remove_signals(void);
3476 #endif
3477 
3478 extern void __kmp_clear_system_time(void);
3479 extern void __kmp_read_system_time(double *delta);
3480 
3481 extern void __kmp_check_stack_overlap(kmp_info_t *thr);
3482 
3483 extern void __kmp_expand_host_name(char *buffer, size_t size);
3484 extern void __kmp_expand_file_name(char *result, size_t rlen, char *pattern);
3485 
3486 #if KMP_ARCH_X86 || KMP_ARCH_X86_64 || (KMP_OS_WINDOWS && KMP_ARCH_AARCH64)
3487 extern void
3488 __kmp_initialize_system_tick(void); /* Initialize timer tick value */
3489 #endif
3490 
3491 extern void
3492 __kmp_runtime_initialize(void); /* machine specific initialization */
3493 extern void __kmp_runtime_destroy(void);
3494 
3495 #if KMP_AFFINITY_SUPPORTED
3496 extern char *__kmp_affinity_print_mask(char *buf, int buf_len,
3497  kmp_affin_mask_t *mask);
3498 extern kmp_str_buf_t *__kmp_affinity_str_buf_mask(kmp_str_buf_t *buf,
3499  kmp_affin_mask_t *mask);
3500 extern void __kmp_affinity_initialize(void);
3501 extern void __kmp_affinity_uninitialize(void);
3502 extern void __kmp_affinity_set_init_mask(
3503  int gtid, int isa_root); /* set affinity according to KMP_AFFINITY */
3504 extern void __kmp_affinity_set_place(int gtid);
3505 extern void __kmp_affinity_determine_capable(const char *env_var);
3506 extern int __kmp_aux_set_affinity(void **mask);
3507 extern int __kmp_aux_get_affinity(void **mask);
3508 extern int __kmp_aux_get_affinity_max_proc();
3509 extern int __kmp_aux_set_affinity_mask_proc(int proc, void **mask);
3510 extern int __kmp_aux_unset_affinity_mask_proc(int proc, void **mask);
3511 extern int __kmp_aux_get_affinity_mask_proc(int proc, void **mask);
3512 extern void __kmp_balanced_affinity(kmp_info_t *th, int team_size);
3513 #if KMP_OS_LINUX || KMP_OS_FREEBSD
3514 extern int kmp_set_thread_affinity_mask_initial(void);
3515 #endif
3516 static inline void __kmp_assign_root_init_mask() {
3517  int gtid = __kmp_entry_gtid();
3518  kmp_root_t *r = __kmp_threads[gtid]->th.th_root;
3519  if (r->r.r_uber_thread == __kmp_threads[gtid] && !r->r.r_affinity_assigned) {
3520  __kmp_affinity_set_init_mask(gtid, TRUE);
3521  r->r.r_affinity_assigned = TRUE;
3522  }
3523 }
3524 #else /* KMP_AFFINITY_SUPPORTED */
3525 #define __kmp_assign_root_init_mask() /* Nothing */
3526 #endif /* KMP_AFFINITY_SUPPORTED */
3527 // No need for KMP_AFFINITY_SUPPORTED guard as only one field in the
3528 // format string is for affinity, so platforms that do not support
3529 // affinity can still use the other fields, e.g., %n for num_threads
3530 extern size_t __kmp_aux_capture_affinity(int gtid, const char *format,
3531  kmp_str_buf_t *buffer);
3532 extern void __kmp_aux_display_affinity(int gtid, const char *format);
3533 
3534 extern void __kmp_cleanup_hierarchy();
3535 extern void __kmp_get_hierarchy(kmp_uint32 nproc, kmp_bstate_t *thr_bar);
3536 
3537 #if KMP_USE_FUTEX
3538 
3539 extern int __kmp_futex_determine_capable(void);
3540 
3541 #endif // KMP_USE_FUTEX
3542 
3543 extern void __kmp_gtid_set_specific(int gtid);
3544 extern int __kmp_gtid_get_specific(void);
3545 
3546 extern double __kmp_read_cpu_time(void);
3547 
3548 extern int __kmp_read_system_info(struct kmp_sys_info *info);
3549 
3550 #if KMP_USE_MONITOR
3551 extern void __kmp_create_monitor(kmp_info_t *th);
3552 #endif
3553 
3554 extern void *__kmp_launch_thread(kmp_info_t *thr);
3555 
3556 extern void __kmp_create_worker(int gtid, kmp_info_t *th, size_t stack_size);
3557 
3558 #if KMP_OS_WINDOWS
3559 extern int __kmp_still_running(kmp_info_t *th);
3560 extern int __kmp_is_thread_alive(kmp_info_t *th, DWORD *exit_val);
3561 extern void __kmp_free_handle(kmp_thread_t tHandle);
3562 #endif
3563 
3564 #if KMP_USE_MONITOR
3565 extern void __kmp_reap_monitor(kmp_info_t *th);
3566 #endif
3567 extern void __kmp_reap_worker(kmp_info_t *th);
3568 extern void __kmp_terminate_thread(int gtid);
3569 
3570 extern int __kmp_try_suspend_mx(kmp_info_t *th);
3571 extern void __kmp_lock_suspend_mx(kmp_info_t *th);
3572 extern void __kmp_unlock_suspend_mx(kmp_info_t *th);
3573 
3574 extern void __kmp_elapsed(double *);
3575 extern void __kmp_elapsed_tick(double *);
3576 
3577 extern void __kmp_enable(int old_state);
3578 extern void __kmp_disable(int *old_state);
3579 
3580 extern void __kmp_thread_sleep(int millis);
3581 
3582 extern void __kmp_common_initialize(void);
3583 extern void __kmp_common_destroy(void);
3584 extern void __kmp_common_destroy_gtid(int gtid);
3585 
3586 #if KMP_OS_UNIX
3587 extern void __kmp_register_atfork(void);
3588 #endif
3589 extern void __kmp_suspend_initialize(void);
3590 extern void __kmp_suspend_initialize_thread(kmp_info_t *th);
3591 extern void __kmp_suspend_uninitialize_thread(kmp_info_t *th);
3592 
3593 extern kmp_info_t *__kmp_allocate_thread(kmp_root_t *root, kmp_team_t *team,
3594  int tid);
3595 extern kmp_team_t *
3596 __kmp_allocate_team(kmp_root_t *root, int new_nproc, int max_nproc,
3597 #if OMPT_SUPPORT
3598  ompt_data_t ompt_parallel_data,
3599 #endif
3600  kmp_proc_bind_t proc_bind, kmp_internal_control_t *new_icvs,
3601  int argc USE_NESTED_HOT_ARG(kmp_info_t *thr));
3602 extern void __kmp_free_thread(kmp_info_t *);
3603 extern void __kmp_free_team(kmp_root_t *,
3604  kmp_team_t *USE_NESTED_HOT_ARG(kmp_info_t *));
3605 extern kmp_team_t *__kmp_reap_team(kmp_team_t *);
3606 
3607 /* ------------------------------------------------------------------------ */
3608 
3609 extern void __kmp_initialize_bget(kmp_info_t *th);
3610 extern void __kmp_finalize_bget(kmp_info_t *th);
3611 
3612 KMP_EXPORT void *kmpc_malloc(size_t size);
3613 KMP_EXPORT void *kmpc_aligned_malloc(size_t size, size_t alignment);
3614 KMP_EXPORT void *kmpc_calloc(size_t nelem, size_t elsize);
3615 KMP_EXPORT void *kmpc_realloc(void *ptr, size_t size);
3616 KMP_EXPORT void kmpc_free(void *ptr);
3617 
3618 /* declarations for internal use */
3619 
3620 extern int __kmp_barrier(enum barrier_type bt, int gtid, int is_split,
3621  size_t reduce_size, void *reduce_data,
3622  void (*reduce)(void *, void *));
3623 extern void __kmp_end_split_barrier(enum barrier_type bt, int gtid);
3624 extern int __kmp_barrier_gomp_cancel(int gtid);
3625 
3630 enum fork_context_e {
3631  fork_context_gnu,
3633  fork_context_intel,
3634  fork_context_last
3635 };
3636 extern int __kmp_fork_call(ident_t *loc, int gtid,
3637  enum fork_context_e fork_context, kmp_int32 argc,
3638  microtask_t microtask, launch_t invoker,
3639  kmp_va_list ap);
3640 
3641 extern void __kmp_join_call(ident_t *loc, int gtid
3642 #if OMPT_SUPPORT
3643  ,
3644  enum fork_context_e fork_context
3645 #endif
3646  ,
3647  int exit_teams = 0);
3648 
3649 extern void __kmp_serialized_parallel(ident_t *id, kmp_int32 gtid);
3650 extern void __kmp_internal_fork(ident_t *id, int gtid, kmp_team_t *team);
3651 extern void __kmp_internal_join(ident_t *id, int gtid, kmp_team_t *team);
3652 extern int __kmp_invoke_task_func(int gtid);
3653 extern void __kmp_run_before_invoked_task(int gtid, int tid,
3654  kmp_info_t *this_thr,
3655  kmp_team_t *team);
3656 extern void __kmp_run_after_invoked_task(int gtid, int tid,
3657  kmp_info_t *this_thr,
3658  kmp_team_t *team);
3659 
3660 // should never have been exported
3661 KMP_EXPORT int __kmpc_invoke_task_func(int gtid);
3662 extern int __kmp_invoke_teams_master(int gtid);
3663 extern void __kmp_teams_master(int gtid);
3664 extern int __kmp_aux_get_team_num();
3665 extern int __kmp_aux_get_num_teams();
3666 extern void __kmp_save_internal_controls(kmp_info_t *thread);
3667 extern void __kmp_user_set_library(enum library_type arg);
3668 extern void __kmp_aux_set_library(enum library_type arg);
3669 extern void __kmp_aux_set_stacksize(size_t arg);
3670 extern void __kmp_aux_set_blocktime(int arg, kmp_info_t *thread, int tid);
3671 extern void __kmp_aux_set_defaults(char const *str, size_t len);
3672 
3673 /* Functions called from __kmp_aux_env_initialize() in kmp_settings.cpp */
3674 void kmpc_set_blocktime(int arg);
3675 void ompc_set_nested(int flag);
3676 void ompc_set_dynamic(int flag);
3677 void ompc_set_num_threads(int arg);
3678 
3679 extern void __kmp_push_current_task_to_thread(kmp_info_t *this_thr,
3680  kmp_team_t *team, int tid);
3681 extern void __kmp_pop_current_task_from_thread(kmp_info_t *this_thr);
3682 extern kmp_task_t *__kmp_task_alloc(ident_t *loc_ref, kmp_int32 gtid,
3683  kmp_tasking_flags_t *flags,
3684  size_t sizeof_kmp_task_t,
3685  size_t sizeof_shareds,
3686  kmp_routine_entry_t task_entry);
3687 extern void __kmp_init_implicit_task(ident_t *loc_ref, kmp_info_t *this_thr,
3688  kmp_team_t *team, int tid,
3689  int set_curr_task);
3690 extern void __kmp_finish_implicit_task(kmp_info_t *this_thr);
3691 extern void __kmp_free_implicit_task(kmp_info_t *this_thr);
3692 
3693 extern kmp_event_t *__kmpc_task_allow_completion_event(ident_t *loc_ref,
3694  int gtid,
3695  kmp_task_t *task);
3696 extern void __kmp_fulfill_event(kmp_event_t *event);
3697 
3698 extern void __kmp_free_task_team(kmp_info_t *thread,
3699  kmp_task_team_t *task_team);
3700 extern void __kmp_reap_task_teams(void);
3701 extern void __kmp_wait_to_unref_task_teams(void);
3702 extern void __kmp_task_team_setup(kmp_info_t *this_thr, kmp_team_t *team,
3703  int always);
3704 extern void __kmp_task_team_sync(kmp_info_t *this_thr, kmp_team_t *team);
3705 extern void __kmp_task_team_wait(kmp_info_t *this_thr, kmp_team_t *team
3706 #if USE_ITT_BUILD
3707  ,
3708  void *itt_sync_obj
3709 #endif /* USE_ITT_BUILD */
3710  ,
3711  int wait = 1);
3712 extern void __kmp_tasking_barrier(kmp_team_t *team, kmp_info_t *thread,
3713  int gtid);
3714 
3715 extern int __kmp_is_address_mapped(void *addr);
3716 extern kmp_uint64 __kmp_hardware_timestamp(void);
3717 
3718 #if KMP_OS_UNIX
3719 extern int __kmp_read_from_file(char const *path, char const *format, ...);
3720 #endif
3721 
3722 /* ------------------------------------------------------------------------ */
3723 //
3724 // Assembly routines that have no compiler intrinsic replacement
3725 //
3726 
3727 extern int __kmp_invoke_microtask(microtask_t pkfn, int gtid, int npr, int argc,
3728  void *argv[]
3729 #if OMPT_SUPPORT
3730  ,
3731  void **exit_frame_ptr
3732 #endif
3733 );
3734 
3735 /* ------------------------------------------------------------------------ */
3736 
3737 KMP_EXPORT void __kmpc_begin(ident_t *, kmp_int32 flags);
3738 KMP_EXPORT void __kmpc_end(ident_t *);
3739 
3740 KMP_EXPORT void __kmpc_threadprivate_register_vec(ident_t *, void *data,
3741  kmpc_ctor_vec ctor,
3742  kmpc_cctor_vec cctor,
3743  kmpc_dtor_vec dtor,
3744  size_t vector_length);
3745 KMP_EXPORT void __kmpc_threadprivate_register(ident_t *, void *data,
3746  kmpc_ctor ctor, kmpc_cctor cctor,
3747  kmpc_dtor dtor);
3748 KMP_EXPORT void *__kmpc_threadprivate(ident_t *, kmp_int32 global_tid,
3749  void *data, size_t size);
3750 
3751 KMP_EXPORT kmp_int32 __kmpc_global_thread_num(ident_t *);
3752 KMP_EXPORT kmp_int32 __kmpc_global_num_threads(ident_t *);
3753 KMP_EXPORT kmp_int32 __kmpc_bound_thread_num(ident_t *);
3754 KMP_EXPORT kmp_int32 __kmpc_bound_num_threads(ident_t *);
3755 
3756 KMP_EXPORT kmp_int32 __kmpc_ok_to_fork(ident_t *);
3757 KMP_EXPORT void __kmpc_fork_call(ident_t *, kmp_int32 nargs,
3758  kmpc_micro microtask, ...);
3759 
3760 KMP_EXPORT void __kmpc_serialized_parallel(ident_t *, kmp_int32 global_tid);
3761 KMP_EXPORT void __kmpc_end_serialized_parallel(ident_t *, kmp_int32 global_tid);
3762 
3763 KMP_EXPORT void __kmpc_flush(ident_t *);
3764 KMP_EXPORT void __kmpc_barrier(ident_t *, kmp_int32 global_tid);
3765 KMP_EXPORT kmp_int32 __kmpc_master(ident_t *, kmp_int32 global_tid);
3766 KMP_EXPORT void __kmpc_end_master(ident_t *, kmp_int32 global_tid);
3767 KMP_EXPORT kmp_int32 __kmpc_masked(ident_t *, kmp_int32 global_tid,
3768  kmp_int32 filter);
3769 KMP_EXPORT void __kmpc_end_masked(ident_t *, kmp_int32 global_tid);
3770 KMP_EXPORT void __kmpc_ordered(ident_t *, kmp_int32 global_tid);
3771 KMP_EXPORT void __kmpc_end_ordered(ident_t *, kmp_int32 global_tid);
3772 KMP_EXPORT void __kmpc_critical(ident_t *, kmp_int32 global_tid,
3773  kmp_critical_name *);
3774 KMP_EXPORT void __kmpc_end_critical(ident_t *, kmp_int32 global_tid,
3775  kmp_critical_name *);
3776 KMP_EXPORT void __kmpc_critical_with_hint(ident_t *, kmp_int32 global_tid,
3777  kmp_critical_name *, uint32_t hint);
3778 
3779 KMP_EXPORT kmp_int32 __kmpc_barrier_master(ident_t *, kmp_int32 global_tid);
3780 KMP_EXPORT void __kmpc_end_barrier_master(ident_t *, kmp_int32 global_tid);
3781 
3782 KMP_EXPORT kmp_int32 __kmpc_barrier_master_nowait(ident_t *,
3783  kmp_int32 global_tid);
3784 
3785 KMP_EXPORT kmp_int32 __kmpc_single(ident_t *, kmp_int32 global_tid);
3786 KMP_EXPORT void __kmpc_end_single(ident_t *, kmp_int32 global_tid);
3787 
3788 KMP_EXPORT void KMPC_FOR_STATIC_INIT(ident_t *loc, kmp_int32 global_tid,
3789  kmp_int32 schedtype, kmp_int32 *plastiter,
3790  kmp_int *plower, kmp_int *pupper,
3791  kmp_int *pstride, kmp_int incr,
3792  kmp_int chunk);
3793 
3794 KMP_EXPORT void __kmpc_for_static_fini(ident_t *loc, kmp_int32 global_tid);
3795 
3796 KMP_EXPORT void __kmpc_copyprivate(ident_t *loc, kmp_int32 global_tid,
3797  size_t cpy_size, void *cpy_data,
3798  void (*cpy_func)(void *, void *),
3799  kmp_int32 didit);
3800 
3801 extern void KMPC_SET_NUM_THREADS(int arg);
3802 extern void KMPC_SET_DYNAMIC(int flag);
3803 extern void KMPC_SET_NESTED(int flag);
3804 
3805 /* OMP 3.0 tasking interface routines */
3806 KMP_EXPORT kmp_int32 __kmpc_omp_task(ident_t *loc_ref, kmp_int32 gtid,
3807  kmp_task_t *new_task);
3808 KMP_EXPORT kmp_task_t *__kmpc_omp_task_alloc(ident_t *loc_ref, kmp_int32 gtid,
3809  kmp_int32 flags,
3810  size_t sizeof_kmp_task_t,
3811  size_t sizeof_shareds,
3812  kmp_routine_entry_t task_entry);
3813 KMP_EXPORT kmp_task_t *__kmpc_omp_target_task_alloc(
3814  ident_t *loc_ref, kmp_int32 gtid, kmp_int32 flags, size_t sizeof_kmp_task_t,
3815  size_t sizeof_shareds, kmp_routine_entry_t task_entry, kmp_int64 device_id);
3816 KMP_EXPORT void __kmpc_omp_task_begin_if0(ident_t *loc_ref, kmp_int32 gtid,
3817  kmp_task_t *task);
3818 KMP_EXPORT void __kmpc_omp_task_complete_if0(ident_t *loc_ref, kmp_int32 gtid,
3819  kmp_task_t *task);
3820 KMP_EXPORT kmp_int32 __kmpc_omp_task_parts(ident_t *loc_ref, kmp_int32 gtid,
3821  kmp_task_t *new_task);
3822 KMP_EXPORT kmp_int32 __kmpc_omp_taskwait(ident_t *loc_ref, kmp_int32 gtid);
3823 
3824 KMP_EXPORT kmp_int32 __kmpc_omp_taskyield(ident_t *loc_ref, kmp_int32 gtid,
3825  int end_part);
3826 
3827 #if TASK_UNUSED
3828 void __kmpc_omp_task_begin(ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *task);
3829 void __kmpc_omp_task_complete(ident_t *loc_ref, kmp_int32 gtid,
3830  kmp_task_t *task);
3831 #endif // TASK_UNUSED
3832 
3833 /* ------------------------------------------------------------------------ */
3834 
3835 KMP_EXPORT void __kmpc_taskgroup(ident_t *loc, int gtid);
3836 KMP_EXPORT void __kmpc_end_taskgroup(ident_t *loc, int gtid);
3837 
3838 KMP_EXPORT kmp_int32 __kmpc_omp_task_with_deps(
3839  ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 ndeps,
3840  kmp_depend_info_t *dep_list, kmp_int32 ndeps_noalias,
3841  kmp_depend_info_t *noalias_dep_list);
3842 KMP_EXPORT void __kmpc_omp_wait_deps(ident_t *loc_ref, kmp_int32 gtid,
3843  kmp_int32 ndeps,
3844  kmp_depend_info_t *dep_list,
3845  kmp_int32 ndeps_noalias,
3846  kmp_depend_info_t *noalias_dep_list);
3847 extern kmp_int32 __kmp_omp_task(kmp_int32 gtid, kmp_task_t *new_task,
3848  bool serialize_immediate);
3849 
3850 KMP_EXPORT kmp_int32 __kmpc_cancel(ident_t *loc_ref, kmp_int32 gtid,
3851  kmp_int32 cncl_kind);
3852 KMP_EXPORT kmp_int32 __kmpc_cancellationpoint(ident_t *loc_ref, kmp_int32 gtid,
3853  kmp_int32 cncl_kind);
3854 KMP_EXPORT kmp_int32 __kmpc_cancel_barrier(ident_t *loc_ref, kmp_int32 gtid);
3855 KMP_EXPORT int __kmp_get_cancellation_status(int cancel_kind);
3856 
3857 KMP_EXPORT void __kmpc_proxy_task_completed(kmp_int32 gtid, kmp_task_t *ptask);
3858 KMP_EXPORT void __kmpc_proxy_task_completed_ooo(kmp_task_t *ptask);
3859 KMP_EXPORT void __kmpc_taskloop(ident_t *loc, kmp_int32 gtid, kmp_task_t *task,
3860  kmp_int32 if_val, kmp_uint64 *lb,
3861  kmp_uint64 *ub, kmp_int64 st, kmp_int32 nogroup,
3862  kmp_int32 sched, kmp_uint64 grainsize,
3863  void *task_dup);
3864 KMP_EXPORT void __kmpc_taskloop_5(ident_t *loc, kmp_int32 gtid,
3865  kmp_task_t *task, kmp_int32 if_val,
3866  kmp_uint64 *lb, kmp_uint64 *ub, kmp_int64 st,
3867  kmp_int32 nogroup, kmp_int32 sched,
3868  kmp_uint64 grainsize, kmp_int32 modifier,
3869  void *task_dup);
3870 KMP_EXPORT void *__kmpc_task_reduction_init(int gtid, int num_data, void *data);
3871 KMP_EXPORT void *__kmpc_taskred_init(int gtid, int num_data, void *data);
3872 KMP_EXPORT void *__kmpc_task_reduction_get_th_data(int gtid, void *tg, void *d);
3873 KMP_EXPORT void *__kmpc_task_reduction_modifier_init(ident_t *loc, int gtid,
3874  int is_ws, int num,
3875  void *data);
3876 KMP_EXPORT void *__kmpc_taskred_modifier_init(ident_t *loc, int gtid, int is_ws,
3877  int num, void *data);
3878 KMP_EXPORT void __kmpc_task_reduction_modifier_fini(ident_t *loc, int gtid,
3879  int is_ws);
3880 KMP_EXPORT kmp_int32 __kmpc_omp_reg_task_with_affinity(
3881  ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 naffins,
3882  kmp_task_affinity_info_t *affin_list);
3883 KMP_EXPORT void __kmp_set_num_teams(int num_teams);
3884 KMP_EXPORT int __kmp_get_max_teams(void);
3885 KMP_EXPORT void __kmp_set_teams_thread_limit(int limit);
3886 KMP_EXPORT int __kmp_get_teams_thread_limit(void);
3887 
3888 /* Lock interface routines (fast versions with gtid passed in) */
3889 KMP_EXPORT void __kmpc_init_lock(ident_t *loc, kmp_int32 gtid,
3890  void **user_lock);
3891 KMP_EXPORT void __kmpc_init_nest_lock(ident_t *loc, kmp_int32 gtid,
3892  void **user_lock);
3893 KMP_EXPORT void __kmpc_destroy_lock(ident_t *loc, kmp_int32 gtid,
3894  void **user_lock);
3895 KMP_EXPORT void __kmpc_destroy_nest_lock(ident_t *loc, kmp_int32 gtid,
3896  void **user_lock);
3897 KMP_EXPORT void __kmpc_set_lock(ident_t *loc, kmp_int32 gtid, void **user_lock);
3898 KMP_EXPORT void __kmpc_set_nest_lock(ident_t *loc, kmp_int32 gtid,
3899  void **user_lock);
3900 KMP_EXPORT void __kmpc_unset_lock(ident_t *loc, kmp_int32 gtid,
3901  void **user_lock);
3902 KMP_EXPORT void __kmpc_unset_nest_lock(ident_t *loc, kmp_int32 gtid,
3903  void **user_lock);
3904 KMP_EXPORT int __kmpc_test_lock(ident_t *loc, kmp_int32 gtid, void **user_lock);
3905 KMP_EXPORT int __kmpc_test_nest_lock(ident_t *loc, kmp_int32 gtid,
3906  void **user_lock);
3907 
3908 KMP_EXPORT void __kmpc_init_lock_with_hint(ident_t *loc, kmp_int32 gtid,
3909  void **user_lock, uintptr_t hint);
3910 KMP_EXPORT void __kmpc_init_nest_lock_with_hint(ident_t *loc, kmp_int32 gtid,
3911  void **user_lock,
3912  uintptr_t hint);
3913 
3914 /* Interface to fast scalable reduce methods routines */
3915 
3916 KMP_EXPORT kmp_int32 __kmpc_reduce_nowait(
3917  ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
3918  void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
3919  kmp_critical_name *lck);
3920 KMP_EXPORT void __kmpc_end_reduce_nowait(ident_t *loc, kmp_int32 global_tid,
3921  kmp_critical_name *lck);
3922 KMP_EXPORT kmp_int32 __kmpc_reduce(
3923  ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
3924  void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
3925  kmp_critical_name *lck);
3926 KMP_EXPORT void __kmpc_end_reduce(ident_t *loc, kmp_int32 global_tid,
3927  kmp_critical_name *lck);
3928 
3929 /* Internal fast reduction routines */
3930 
3931 extern PACKED_REDUCTION_METHOD_T __kmp_determine_reduction_method(
3932  ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
3933  void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
3934  kmp_critical_name *lck);
3935 
3936 // this function is for testing set/get/determine reduce method
3937 KMP_EXPORT kmp_int32 __kmp_get_reduce_method(void);
3938 
3939 KMP_EXPORT kmp_uint64 __kmpc_get_taskid();
3940 KMP_EXPORT kmp_uint64 __kmpc_get_parent_taskid();
3941 
3942 // C++ port
3943 // missing 'extern "C"' declarations
3944 
3945 KMP_EXPORT kmp_int32 __kmpc_in_parallel(ident_t *loc);
3946 KMP_EXPORT void __kmpc_pop_num_threads(ident_t *loc, kmp_int32 global_tid);
3947 KMP_EXPORT void __kmpc_push_num_threads(ident_t *loc, kmp_int32 global_tid,
3948  kmp_int32 num_threads);
3949 
3950 KMP_EXPORT void __kmpc_push_proc_bind(ident_t *loc, kmp_int32 global_tid,
3951  int proc_bind);
3952 KMP_EXPORT void __kmpc_push_num_teams(ident_t *loc, kmp_int32 global_tid,
3953  kmp_int32 num_teams,
3954  kmp_int32 num_threads);
3955 /* Function for OpenMP 5.1 num_teams clause */
3956 KMP_EXPORT void __kmpc_push_num_teams_51(ident_t *loc, kmp_int32 global_tid,
3957  kmp_int32 num_teams_lb,
3958  kmp_int32 num_teams_ub,
3959  kmp_int32 num_threads);
3960 KMP_EXPORT void __kmpc_fork_teams(ident_t *loc, kmp_int32 argc,
3961  kmpc_micro microtask, ...);
3962 struct kmp_dim { // loop bounds info casted to kmp_int64
3963  kmp_int64 lo; // lower
3964  kmp_int64 up; // upper
3965  kmp_int64 st; // stride
3966 };
3967 KMP_EXPORT void __kmpc_doacross_init(ident_t *loc, kmp_int32 gtid,
3968  kmp_int32 num_dims,
3969  const struct kmp_dim *dims);
3970 KMP_EXPORT void __kmpc_doacross_wait(ident_t *loc, kmp_int32 gtid,
3971  const kmp_int64 *vec);
3972 KMP_EXPORT void __kmpc_doacross_post(ident_t *loc, kmp_int32 gtid,
3973  const kmp_int64 *vec);
3974 KMP_EXPORT void __kmpc_doacross_fini(ident_t *loc, kmp_int32 gtid);
3975 
3976 KMP_EXPORT void *__kmpc_threadprivate_cached(ident_t *loc, kmp_int32 global_tid,
3977  void *data, size_t size,
3978  void ***cache);
3979 
3980 // Symbols for MS mutual detection.
3981 extern int _You_must_link_with_exactly_one_OpenMP_library;
3982 extern int _You_must_link_with_Intel_OpenMP_library;
3983 #if KMP_OS_WINDOWS && (KMP_VERSION_MAJOR > 4)
3984 extern int _You_must_link_with_Microsoft_OpenMP_library;
3985 #endif
3986 
3987 // The routines below are not exported.
3988 // Consider making them 'static' in corresponding source files.
3989 void kmp_threadprivate_insert_private_data(int gtid, void *pc_addr,
3990  void *data_addr, size_t pc_size);
3991 struct private_common *kmp_threadprivate_insert(int gtid, void *pc_addr,
3992  void *data_addr,
3993  size_t pc_size);
3994 void __kmp_threadprivate_resize_cache(int newCapacity);
3995 void __kmp_cleanup_threadprivate_caches();
3996 
3997 // ompc_, kmpc_ entries moved from omp.h.
3998 #if KMP_OS_WINDOWS
3999 #define KMPC_CONVENTION __cdecl
4000 #else
4001 #define KMPC_CONVENTION
4002 #endif
4003 
4004 #ifndef __OMP_H
4005 typedef enum omp_sched_t {
4006  omp_sched_static = 1,
4007  omp_sched_dynamic = 2,
4008  omp_sched_guided = 3,
4009  omp_sched_auto = 4
4010 } omp_sched_t;
4011 typedef void *kmp_affinity_mask_t;
4012 #endif
4013 
4014 KMP_EXPORT void KMPC_CONVENTION ompc_set_max_active_levels(int);
4015 KMP_EXPORT void KMPC_CONVENTION ompc_set_schedule(omp_sched_t, int);
4016 KMP_EXPORT int KMPC_CONVENTION ompc_get_ancestor_thread_num(int);
4017 KMP_EXPORT int KMPC_CONVENTION ompc_get_team_size(int);
4018 KMP_EXPORT int KMPC_CONVENTION
4019 kmpc_set_affinity_mask_proc(int, kmp_affinity_mask_t *);
4020 KMP_EXPORT int KMPC_CONVENTION
4021 kmpc_unset_affinity_mask_proc(int, kmp_affinity_mask_t *);
4022 KMP_EXPORT int KMPC_CONVENTION
4023 kmpc_get_affinity_mask_proc(int, kmp_affinity_mask_t *);
4024 
4025 KMP_EXPORT void KMPC_CONVENTION kmpc_set_stacksize(int);
4026 KMP_EXPORT void KMPC_CONVENTION kmpc_set_stacksize_s(size_t);
4027 KMP_EXPORT void KMPC_CONVENTION kmpc_set_library(int);
4028 KMP_EXPORT void KMPC_CONVENTION kmpc_set_defaults(char const *);
4029 KMP_EXPORT void KMPC_CONVENTION kmpc_set_disp_num_buffers(int);
4030 void KMP_EXPAND_NAME(ompc_set_affinity_format)(char const *format);
4031 size_t KMP_EXPAND_NAME(ompc_get_affinity_format)(char *buffer, size_t size);
4032 void KMP_EXPAND_NAME(ompc_display_affinity)(char const *format);
4033 size_t KMP_EXPAND_NAME(ompc_capture_affinity)(char *buffer, size_t buf_size,
4034  char const *format);
4035 
4036 enum kmp_target_offload_kind {
4037  tgt_disabled = 0,
4038  tgt_default = 1,
4039  tgt_mandatory = 2
4040 };
4041 typedef enum kmp_target_offload_kind kmp_target_offload_kind_t;
4042 // Set via OMP_TARGET_OFFLOAD if specified, defaults to tgt_default otherwise
4043 extern kmp_target_offload_kind_t __kmp_target_offload;
4044 extern int __kmpc_get_target_offload();
4045 
4046 // Constants used in libomptarget
4047 #define KMP_DEVICE_DEFAULT -1 // This is libomptarget's default device.
4048 #define KMP_DEVICE_ALL -11 // This is libomptarget's "all devices".
4049 
4050 // OMP Pause Resource
4051 
4052 // The following enum is used both to set the status in __kmp_pause_status, and
4053 // as the internal equivalent of the externally-visible omp_pause_resource_t.
4054 typedef enum kmp_pause_status_t {
4055  kmp_not_paused = 0, // status is not paused, or, requesting resume
4056  kmp_soft_paused = 1, // status is soft-paused, or, requesting soft pause
4057  kmp_hard_paused = 2 // status is hard-paused, or, requesting hard pause
4058 } kmp_pause_status_t;
4059 
4060 // This stores the pause state of the runtime
4061 extern kmp_pause_status_t __kmp_pause_status;
4062 extern int __kmpc_pause_resource(kmp_pause_status_t level);
4063 extern int __kmp_pause_resource(kmp_pause_status_t level);
4064 // Soft resume sets __kmp_pause_status, and wakes up all threads.
4065 extern void __kmp_resume_if_soft_paused();
4066 // Hard resume simply resets the status to not paused. Library will appear to
4067 // be uninitialized after hard pause. Let OMP constructs trigger required
4068 // initializations.
4069 static inline void __kmp_resume_if_hard_paused() {
4070  if (__kmp_pause_status == kmp_hard_paused) {
4071  __kmp_pause_status = kmp_not_paused;
4072  }
4073 }
4074 
4075 extern void __kmp_omp_display_env(int verbose);
4076 
4077 // 1: it is initializing hidden helper team
4078 extern volatile int __kmp_init_hidden_helper;
4079 // 1: the hidden helper team is done
4080 extern volatile int __kmp_hidden_helper_team_done;
4081 // 1: enable hidden helper task
4082 extern kmp_int32 __kmp_enable_hidden_helper;
4083 // Main thread of hidden helper team
4084 extern kmp_info_t *__kmp_hidden_helper_main_thread;
4085 // Descriptors for the hidden helper threads
4086 extern kmp_info_t **__kmp_hidden_helper_threads;
4087 // Number of hidden helper threads
4088 extern kmp_int32 __kmp_hidden_helper_threads_num;
4089 // Number of hidden helper tasks that have not been executed yet
4090 extern std::atomic<kmp_int32> __kmp_unexecuted_hidden_helper_tasks;
4091 
4092 extern void __kmp_hidden_helper_initialize();
4093 extern void __kmp_hidden_helper_threads_initz_routine();
4094 extern void __kmp_do_initialize_hidden_helper_threads();
4095 extern void __kmp_hidden_helper_threads_initz_wait();
4096 extern void __kmp_hidden_helper_initz_release();
4097 extern void __kmp_hidden_helper_threads_deinitz_wait();
4098 extern void __kmp_hidden_helper_threads_deinitz_release();
4099 extern void __kmp_hidden_helper_main_thread_wait();
4100 extern void __kmp_hidden_helper_worker_thread_wait();
4101 extern void __kmp_hidden_helper_worker_thread_signal();
4102 extern void __kmp_hidden_helper_main_thread_release();
4103 
4104 // Check whether a given thread is a hidden helper thread
4105 #define KMP_HIDDEN_HELPER_THREAD(gtid) \
4106  ((gtid) >= 1 && (gtid) <= __kmp_hidden_helper_threads_num)
4107 
4108 #define KMP_HIDDEN_HELPER_WORKER_THREAD(gtid) \
4109  ((gtid) > 1 && (gtid) <= __kmp_hidden_helper_threads_num)
4110 
4111 #define KMP_HIDDEN_HELPER_TEAM(team) \
4112  (team->t.t_threads[0] == __kmp_hidden_helper_main_thread)
4113 
4114 // Map a gtid to a hidden helper thread. The first hidden helper thread, a.k.a
4115 // main thread, is skipped.
4116 #define KMP_GTID_TO_SHADOW_GTID(gtid) \
4117  ((gtid) % (__kmp_hidden_helper_threads_num - 1) + 2)
4118 
4119 // Return the adjusted gtid value by subtracting from gtid the number
4120 // of hidden helper threads. This adjusted value is the gtid the thread would
4121 // have received if there were no hidden helper threads.
4122 static inline int __kmp_adjust_gtid_for_hidden_helpers(int gtid) {
4123  int adjusted_gtid = gtid;
4124  if (__kmp_hidden_helper_threads_num > 0 && gtid > 0 &&
4125  gtid - __kmp_hidden_helper_threads_num >= 0) {
4126  adjusted_gtid -= __kmp_hidden_helper_threads_num;
4127  }
4128  return adjusted_gtid;
4129 }
4130 
4131 // Support for error directive
4132 typedef enum kmp_severity_t {
4133  severity_warning = 1,
4134  severity_fatal = 2
4135 } kmp_severity_t;
4136 extern void __kmpc_error(ident_t *loc, int severity, const char *message);
4137 
4138 #ifdef __cplusplus
4139 }
4140 #endif
4141 
4142 template <bool C, bool S>
4143 extern void __kmp_suspend_32(int th_gtid, kmp_flag_32<C, S> *flag);
4144 template <bool C, bool S>
4145 extern void __kmp_suspend_64(int th_gtid, kmp_flag_64<C, S> *flag);
4146 template <bool C, bool S>
4147 extern void __kmp_atomic_suspend_64(int th_gtid,
4148  kmp_atomic_flag_64<C, S> *flag);
4149 extern void __kmp_suspend_oncore(int th_gtid, kmp_flag_oncore *flag);
4150 #if KMP_HAVE_MWAIT || KMP_HAVE_UMWAIT
4151 template <bool C, bool S>
4152 extern void __kmp_mwait_32(int th_gtid, kmp_flag_32<C, S> *flag);
4153 template <bool C, bool S>
4154 extern void __kmp_mwait_64(int th_gtid, kmp_flag_64<C, S> *flag);
4155 template <bool C, bool S>
4156 extern void __kmp_atomic_mwait_64(int th_gtid, kmp_atomic_flag_64<C, S> *flag);
4157 extern void __kmp_mwait_oncore(int th_gtid, kmp_flag_oncore *flag);
4158 #endif
4159 template <bool C, bool S>
4160 extern void __kmp_resume_32(int target_gtid, kmp_flag_32<C, S> *flag);
4161 template <bool C, bool S>
4162 extern void __kmp_resume_64(int target_gtid, kmp_flag_64<C, S> *flag);
4163 template <bool C, bool S>
4164 extern void __kmp_atomic_resume_64(int target_gtid,
4165  kmp_atomic_flag_64<C, S> *flag);
4166 extern void __kmp_resume_oncore(int target_gtid, kmp_flag_oncore *flag);
4167 
4168 template <bool C, bool S>
4169 int __kmp_execute_tasks_32(kmp_info_t *thread, kmp_int32 gtid,
4170  kmp_flag_32<C, S> *flag, int final_spin,
4171  int *thread_finished,
4172 #if USE_ITT_BUILD
4173  void *itt_sync_obj,
4174 #endif /* USE_ITT_BUILD */
4175  kmp_int32 is_constrained);
4176 template <bool C, bool S>
4177 int __kmp_execute_tasks_64(kmp_info_t *thread, kmp_int32 gtid,
4178  kmp_flag_64<C, S> *flag, int final_spin,
4179  int *thread_finished,
4180 #if USE_ITT_BUILD
4181  void *itt_sync_obj,
4182 #endif /* USE_ITT_BUILD */
4183  kmp_int32 is_constrained);
4184 template <bool C, bool S>
4185 int __kmp_atomic_execute_tasks_64(kmp_info_t *thread, kmp_int32 gtid,
4186  kmp_atomic_flag_64<C, S> *flag,
4187  int final_spin, int *thread_finished,
4188 #if USE_ITT_BUILD
4189  void *itt_sync_obj,
4190 #endif /* USE_ITT_BUILD */
4191  kmp_int32 is_constrained);
4192 int __kmp_execute_tasks_oncore(kmp_info_t *thread, kmp_int32 gtid,
4193  kmp_flag_oncore *flag, int final_spin,
4194  int *thread_finished,
4195 #if USE_ITT_BUILD
4196  void *itt_sync_obj,
4197 #endif /* USE_ITT_BUILD */
4198  kmp_int32 is_constrained);
4199 
4200 extern int __kmp_nesting_mode;
4201 extern int __kmp_nesting_mode_nlevels;
4202 extern int *__kmp_nesting_nth_level;
4203 extern void __kmp_init_nesting_mode();
4204 extern void __kmp_set_nesting_mode_threads();
4205 
4213  FILE *f;
4214 
4215  void close() {
4216  if (f && f != stdout && f != stderr) {
4217  fclose(f);
4218  f = nullptr;
4219  }
4220  }
4221 
4222 public:
4223  kmp_safe_raii_file_t() : f(nullptr) {}
4224  kmp_safe_raii_file_t(const char *filename, const char *mode,
4225  const char *env_var = nullptr)
4226  : f(nullptr) {
4227  open(filename, mode, env_var);
4228  }
4229  ~kmp_safe_raii_file_t() { close(); }
4230 
4234  void open(const char *filename, const char *mode,
4235  const char *env_var = nullptr) {
4236  KMP_ASSERT(!f);
4237  f = fopen(filename, mode);
4238  if (!f) {
4239  int code = errno;
4240  if (env_var) {
4241  __kmp_fatal(KMP_MSG(CantOpenFileForReading, filename), KMP_ERR(code),
4242  KMP_HNT(CheckEnvVar, env_var, filename), __kmp_msg_null);
4243  } else {
4244  __kmp_fatal(KMP_MSG(CantOpenFileForReading, filename), KMP_ERR(code),
4245  __kmp_msg_null);
4246  }
4247  }
4248  }
4251  void set_stdout() {
4252  KMP_ASSERT(!f);
4253  f = stdout;
4254  }
4257  void set_stderr() {
4258  KMP_ASSERT(!f);
4259  f = stderr;
4260  }
4261  operator bool() { return bool(f); }
4262  operator FILE *() { return f; }
4263 };
4264 
4265 template <typename SourceType, typename TargetType,
4266  bool isSourceSmaller = (sizeof(SourceType) < sizeof(TargetType)),
4267  bool isSourceEqual = (sizeof(SourceType) == sizeof(TargetType)),
4268  bool isSourceSigned = std::is_signed<SourceType>::value,
4269  bool isTargetSigned = std::is_signed<TargetType>::value>
4270 struct kmp_convert {};
4271 
4272 // Both types are signed; Source smaller
4273 template <typename SourceType, typename TargetType>
4274 struct kmp_convert<SourceType, TargetType, true, false, true, true> {
4275  static TargetType to(SourceType src) { return (TargetType)src; }
4276 };
4277 // Source equal
4278 template <typename SourceType, typename TargetType>
4279 struct kmp_convert<SourceType, TargetType, false, true, true, true> {
4280  static TargetType to(SourceType src) { return src; }
4281 };
4282 // Source bigger
4283 template <typename SourceType, typename TargetType>
4284 struct kmp_convert<SourceType, TargetType, false, false, true, true> {
4285  static TargetType to(SourceType src) {
4286  KMP_ASSERT(src <= static_cast<SourceType>(
4287  (std::numeric_limits<TargetType>::max)()));
4288  KMP_ASSERT(src >= static_cast<SourceType>(
4289  (std::numeric_limits<TargetType>::min)()));
4290  return (TargetType)src;
4291  }
4292 };
4293 
4294 // Source signed, Target unsigned
4295 // Source smaller
4296 template <typename SourceType, typename TargetType>
4297 struct kmp_convert<SourceType, TargetType, true, false, true, false> {
4298  static TargetType to(SourceType src) {
4299  KMP_ASSERT(src >= 0);
4300  return (TargetType)src;
4301  }
4302 };
4303 // Source equal
4304 template <typename SourceType, typename TargetType>
4305 struct kmp_convert<SourceType, TargetType, false, true, true, false> {
4306  static TargetType to(SourceType src) {
4307  KMP_ASSERT(src >= 0);
4308  return (TargetType)src;
4309  }
4310 };
4311 // Source bigger
4312 template <typename SourceType, typename TargetType>
4313 struct kmp_convert<SourceType, TargetType, false, false, true, false> {
4314  static TargetType to(SourceType src) {
4315  KMP_ASSERT(src >= 0);
4316  KMP_ASSERT(src <= static_cast<SourceType>(
4317  (std::numeric_limits<TargetType>::max)()));
4318  return (TargetType)src;
4319  }
4320 };
4321 
4322 // Source unsigned, Target signed
4323 // Source smaller
4324 template <typename SourceType, typename TargetType>
4325 struct kmp_convert<SourceType, TargetType, true, false, false, true> {
4326  static TargetType to(SourceType src) { return (TargetType)src; }
4327 };
4328 // Source equal
4329 template <typename SourceType, typename TargetType>
4330 struct kmp_convert<SourceType, TargetType, false, true, false, true> {
4331  static TargetType to(SourceType src) {
4332  KMP_ASSERT(src <= static_cast<SourceType>(
4333  (std::numeric_limits<TargetType>::max)()));
4334  return (TargetType)src;
4335  }
4336 };
4337 // Source bigger
4338 template <typename SourceType, typename TargetType>
4339 struct kmp_convert<SourceType, TargetType, false, false, false, true> {
4340  static TargetType to(SourceType src) {
4341  KMP_ASSERT(src <= static_cast<SourceType>(
4342  (std::numeric_limits<TargetType>::max)()));
4343  return (TargetType)src;
4344  }
4345 };
4346 
4347 // Source unsigned, Target unsigned
4348 // Source smaller
4349 template <typename SourceType, typename TargetType>
4350 struct kmp_convert<SourceType, TargetType, true, false, false, false> {
4351  static TargetType to(SourceType src) { return (TargetType)src; }
4352 };
4353 // Source equal
4354 template <typename SourceType, typename TargetType>
4355 struct kmp_convert<SourceType, TargetType, false, true, false, false> {
4356  static TargetType to(SourceType src) { return src; }
4357 };
4358 // Source bigger
4359 template <typename SourceType, typename TargetType>
4360 struct kmp_convert<SourceType, TargetType, false, false, false, false> {
4361  static TargetType to(SourceType src) {
4362  KMP_ASSERT(src <= static_cast<SourceType>(
4363  (std::numeric_limits<TargetType>::max)()));
4364  return (TargetType)src;
4365  }
4366 };
4367 
4368 template <typename T1, typename T2>
4369 static inline void __kmp_type_convert(T1 src, T2 *dest) {
4370  *dest = kmp_convert<T1, T2>::to(src);
4371 }
4372 
4373 #endif /* KMP_H */
void set_stdout()
Definition: kmp.h:4251
void set_stderr()
Definition: kmp.h:4257
void open(const char *filename, const char *mode, const char *env_var=nullptr)
Definition: kmp.h:4234
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