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 typedef enum kmp_hw_core_type_t {
622  KMP_HW_CORE_TYPE_UNKNOWN = 0x0,
623 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
624  KMP_HW_CORE_TYPE_ATOM = 0x20,
625  KMP_HW_CORE_TYPE_CORE = 0x40,
626  KMP_HW_MAX_NUM_CORE_TYPES = 3,
627 #else
628  KMP_HW_MAX_NUM_CORE_TYPES = 1,
629 #endif
630 } kmp_hw_core_type_t;
631 
632 #define KMP_HW_MAX_NUM_CORE_EFFS 8
633 
634 #define KMP_DEBUG_ASSERT_VALID_HW_TYPE(type) \
635  KMP_DEBUG_ASSERT(type >= (kmp_hw_t)0 && type < KMP_HW_LAST)
636 #define KMP_ASSERT_VALID_HW_TYPE(type) \
637  KMP_ASSERT(type >= (kmp_hw_t)0 && type < KMP_HW_LAST)
638 
639 #define KMP_FOREACH_HW_TYPE(type) \
640  for (kmp_hw_t type = (kmp_hw_t)0; type < KMP_HW_LAST; \
641  type = (kmp_hw_t)((int)type + 1))
642 
643 const char *__kmp_hw_get_keyword(kmp_hw_t type, bool plural = false);
644 const char *__kmp_hw_get_catalog_string(kmp_hw_t type, bool plural = false);
645 const char *__kmp_hw_get_core_type_string(kmp_hw_core_type_t type);
646 
647 /* Only Linux* OS and Windows* OS support thread affinity. */
648 #if KMP_AFFINITY_SUPPORTED
649 
650 // GROUP_AFFINITY is already defined for _MSC_VER>=1600 (VS2010 and later).
651 #if KMP_OS_WINDOWS
652 #if _MSC_VER < 1600 && KMP_MSVC_COMPAT
653 typedef struct GROUP_AFFINITY {
654  KAFFINITY Mask;
655  WORD Group;
656  WORD Reserved[3];
657 } GROUP_AFFINITY;
658 #endif /* _MSC_VER < 1600 */
659 #if KMP_GROUP_AFFINITY
660 extern int __kmp_num_proc_groups;
661 #else
662 static const int __kmp_num_proc_groups = 1;
663 #endif /* KMP_GROUP_AFFINITY */
664 typedef DWORD (*kmp_GetActiveProcessorCount_t)(WORD);
665 extern kmp_GetActiveProcessorCount_t __kmp_GetActiveProcessorCount;
666 
667 typedef WORD (*kmp_GetActiveProcessorGroupCount_t)(void);
668 extern kmp_GetActiveProcessorGroupCount_t __kmp_GetActiveProcessorGroupCount;
669 
670 typedef BOOL (*kmp_GetThreadGroupAffinity_t)(HANDLE, GROUP_AFFINITY *);
671 extern kmp_GetThreadGroupAffinity_t __kmp_GetThreadGroupAffinity;
672 
673 typedef BOOL (*kmp_SetThreadGroupAffinity_t)(HANDLE, const GROUP_AFFINITY *,
674  GROUP_AFFINITY *);
675 extern kmp_SetThreadGroupAffinity_t __kmp_SetThreadGroupAffinity;
676 #endif /* KMP_OS_WINDOWS */
677 
678 #if KMP_USE_HWLOC
679 extern hwloc_topology_t __kmp_hwloc_topology;
680 extern int __kmp_hwloc_error;
681 #endif
682 
683 extern size_t __kmp_affin_mask_size;
684 #define KMP_AFFINITY_CAPABLE() (__kmp_affin_mask_size > 0)
685 #define KMP_AFFINITY_DISABLE() (__kmp_affin_mask_size = 0)
686 #define KMP_AFFINITY_ENABLE(mask_size) (__kmp_affin_mask_size = mask_size)
687 #define KMP_CPU_SET_ITERATE(i, mask) \
688  for (i = (mask)->begin(); (int)i != (mask)->end(); i = (mask)->next(i))
689 #define KMP_CPU_SET(i, mask) (mask)->set(i)
690 #define KMP_CPU_ISSET(i, mask) (mask)->is_set(i)
691 #define KMP_CPU_CLR(i, mask) (mask)->clear(i)
692 #define KMP_CPU_ZERO(mask) (mask)->zero()
693 #define KMP_CPU_COPY(dest, src) (dest)->copy(src)
694 #define KMP_CPU_AND(dest, src) (dest)->bitwise_and(src)
695 #define KMP_CPU_COMPLEMENT(max_bit_number, mask) (mask)->bitwise_not()
696 #define KMP_CPU_UNION(dest, src) (dest)->bitwise_or(src)
697 #define KMP_CPU_ALLOC(ptr) (ptr = __kmp_affinity_dispatch->allocate_mask())
698 #define KMP_CPU_FREE(ptr) __kmp_affinity_dispatch->deallocate_mask(ptr)
699 #define KMP_CPU_ALLOC_ON_STACK(ptr) KMP_CPU_ALLOC(ptr)
700 #define KMP_CPU_FREE_FROM_STACK(ptr) KMP_CPU_FREE(ptr)
701 #define KMP_CPU_INTERNAL_ALLOC(ptr) KMP_CPU_ALLOC(ptr)
702 #define KMP_CPU_INTERNAL_FREE(ptr) KMP_CPU_FREE(ptr)
703 #define KMP_CPU_INDEX(arr, i) __kmp_affinity_dispatch->index_mask_array(arr, i)
704 #define KMP_CPU_ALLOC_ARRAY(arr, n) \
705  (arr = __kmp_affinity_dispatch->allocate_mask_array(n))
706 #define KMP_CPU_FREE_ARRAY(arr, n) \
707  __kmp_affinity_dispatch->deallocate_mask_array(arr)
708 #define KMP_CPU_INTERNAL_ALLOC_ARRAY(arr, n) KMP_CPU_ALLOC_ARRAY(arr, n)
709 #define KMP_CPU_INTERNAL_FREE_ARRAY(arr, n) KMP_CPU_FREE_ARRAY(arr, n)
710 #define __kmp_get_system_affinity(mask, abort_bool) \
711  (mask)->get_system_affinity(abort_bool)
712 #define __kmp_set_system_affinity(mask, abort_bool) \
713  (mask)->set_system_affinity(abort_bool)
714 #define __kmp_get_proc_group(mask) (mask)->get_proc_group()
715 
716 class KMPAffinity {
717 public:
718  class Mask {
719  public:
720  void *operator new(size_t n);
721  void operator delete(void *p);
722  void *operator new[](size_t n);
723  void operator delete[](void *p);
724  virtual ~Mask() {}
725  // Set bit i to 1
726  virtual void set(int i) {}
727  // Return bit i
728  virtual bool is_set(int i) const { return false; }
729  // Set bit i to 0
730  virtual void clear(int i) {}
731  // Zero out entire mask
732  virtual void zero() {}
733  // Copy src into this mask
734  virtual void copy(const Mask *src) {}
735  // this &= rhs
736  virtual void bitwise_and(const Mask *rhs) {}
737  // this |= rhs
738  virtual void bitwise_or(const Mask *rhs) {}
739  // this = ~this
740  virtual void bitwise_not() {}
741  // API for iterating over an affinity mask
742  // for (int i = mask->begin(); i != mask->end(); i = mask->next(i))
743  virtual int begin() const { return 0; }
744  virtual int end() const { return 0; }
745  virtual int next(int previous) const { return 0; }
746 #if KMP_OS_WINDOWS
747  virtual int set_process_affinity(bool abort_on_error) const { return -1; }
748 #endif
749  // Set the system's affinity to this affinity mask's value
750  virtual int set_system_affinity(bool abort_on_error) const { return -1; }
751  // Set this affinity mask to the current system affinity
752  virtual int get_system_affinity(bool abort_on_error) { return -1; }
753  // Only 1 DWORD in the mask should have any procs set.
754  // Return the appropriate index, or -1 for an invalid mask.
755  virtual int get_proc_group() const { return -1; }
756  };
757  void *operator new(size_t n);
758  void operator delete(void *p);
759  // Need virtual destructor
760  virtual ~KMPAffinity() = default;
761  // Determine if affinity is capable
762  virtual void determine_capable(const char *env_var) {}
763  // Bind the current thread to os proc
764  virtual void bind_thread(int proc) {}
765  // Factory functions to allocate/deallocate a mask
766  virtual Mask *allocate_mask() { return nullptr; }
767  virtual void deallocate_mask(Mask *m) {}
768  virtual Mask *allocate_mask_array(int num) { return nullptr; }
769  virtual void deallocate_mask_array(Mask *m) {}
770  virtual Mask *index_mask_array(Mask *m, int index) { return nullptr; }
771  static void pick_api();
772  static void destroy_api();
773  enum api_type {
774  NATIVE_OS
775 #if KMP_USE_HWLOC
776  ,
777  HWLOC
778 #endif
779  };
780  virtual api_type get_api_type() const {
781  KMP_ASSERT(0);
782  return NATIVE_OS;
783  }
784 
785 private:
786  static bool picked_api;
787 };
788 
789 typedef KMPAffinity::Mask kmp_affin_mask_t;
790 extern KMPAffinity *__kmp_affinity_dispatch;
791 
792 // Declare local char buffers with this size for printing debug and info
793 // messages, using __kmp_affinity_print_mask().
794 #define KMP_AFFIN_MASK_PRINT_LEN 1024
795 
796 enum affinity_type {
797  affinity_none = 0,
798  affinity_physical,
799  affinity_logical,
800  affinity_compact,
801  affinity_scatter,
802  affinity_explicit,
803  affinity_balanced,
804  affinity_disabled, // not used outsize the env var parser
805  affinity_default
806 };
807 
808 enum affinity_top_method {
809  affinity_top_method_all = 0, // try all (supported) methods, in order
810 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
811  affinity_top_method_apicid,
812  affinity_top_method_x2apicid,
813  affinity_top_method_x2apicid_1f,
814 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
815  affinity_top_method_cpuinfo, // KMP_CPUINFO_FILE is usable on Windows* OS, too
816 #if KMP_GROUP_AFFINITY
817  affinity_top_method_group,
818 #endif /* KMP_GROUP_AFFINITY */
819  affinity_top_method_flat,
820 #if KMP_USE_HWLOC
821  affinity_top_method_hwloc,
822 #endif
823  affinity_top_method_default
824 };
825 
826 #define affinity_respect_mask_default (-1)
827 
828 extern enum affinity_type __kmp_affinity_type; /* Affinity type */
829 extern kmp_hw_t __kmp_affinity_gran; /* Affinity granularity */
830 extern int __kmp_affinity_gran_levels; /* corresponding int value */
831 extern int __kmp_affinity_dups; /* Affinity duplicate masks */
832 extern enum affinity_top_method __kmp_affinity_top_method;
833 extern int __kmp_affinity_compact; /* Affinity 'compact' value */
834 extern int __kmp_affinity_offset; /* Affinity offset value */
835 extern int __kmp_affinity_verbose; /* Was verbose specified for KMP_AFFINITY? */
836 extern int __kmp_affinity_warnings; /* KMP_AFFINITY warnings enabled ? */
837 extern int __kmp_affinity_respect_mask; // Respect process' init affinity mask?
838 extern char *__kmp_affinity_proclist; /* proc ID list */
839 extern kmp_affin_mask_t *__kmp_affinity_masks;
840 extern unsigned __kmp_affinity_num_masks;
841 extern void __kmp_affinity_bind_thread(int which);
842 
843 extern kmp_affin_mask_t *__kmp_affin_fullMask;
844 extern char *__kmp_cpuinfo_file;
845 
846 #endif /* KMP_AFFINITY_SUPPORTED */
847 
848 // This needs to be kept in sync with the values in omp.h !!!
849 typedef enum kmp_proc_bind_t {
850  proc_bind_false = 0,
851  proc_bind_true,
852  proc_bind_primary,
853  proc_bind_close,
854  proc_bind_spread,
855  proc_bind_intel, // use KMP_AFFINITY interface
856  proc_bind_default
857 } kmp_proc_bind_t;
858 
859 typedef struct kmp_nested_proc_bind_t {
860  kmp_proc_bind_t *bind_types;
861  int size;
862  int used;
863 } kmp_nested_proc_bind_t;
864 
865 extern kmp_nested_proc_bind_t __kmp_nested_proc_bind;
866 extern kmp_proc_bind_t __kmp_teams_proc_bind;
867 
868 extern int __kmp_display_affinity;
869 extern char *__kmp_affinity_format;
870 static const size_t KMP_AFFINITY_FORMAT_SIZE = 512;
871 #if OMPT_SUPPORT
872 extern int __kmp_tool;
873 extern char *__kmp_tool_libraries;
874 #endif // OMPT_SUPPORT
875 
876 #if KMP_AFFINITY_SUPPORTED
877 #define KMP_PLACE_ALL (-1)
878 #define KMP_PLACE_UNDEFINED (-2)
879 // Is KMP_AFFINITY is being used instead of OMP_PROC_BIND/OMP_PLACES?
880 #define KMP_AFFINITY_NON_PROC_BIND \
881  ((__kmp_nested_proc_bind.bind_types[0] == proc_bind_false || \
882  __kmp_nested_proc_bind.bind_types[0] == proc_bind_intel) && \
883  (__kmp_affinity_num_masks > 0 || __kmp_affinity_type == affinity_balanced))
884 #endif /* KMP_AFFINITY_SUPPORTED */
885 
886 extern int __kmp_affinity_num_places;
887 
888 typedef enum kmp_cancel_kind_t {
889  cancel_noreq = 0,
890  cancel_parallel = 1,
891  cancel_loop = 2,
892  cancel_sections = 3,
893  cancel_taskgroup = 4
894 } kmp_cancel_kind_t;
895 
896 // KMP_HW_SUBSET support:
897 typedef struct kmp_hws_item {
898  int num;
899  int offset;
900 } kmp_hws_item_t;
901 
902 extern kmp_hws_item_t __kmp_hws_socket;
903 extern kmp_hws_item_t __kmp_hws_die;
904 extern kmp_hws_item_t __kmp_hws_node;
905 extern kmp_hws_item_t __kmp_hws_tile;
906 extern kmp_hws_item_t __kmp_hws_core;
907 extern kmp_hws_item_t __kmp_hws_proc;
908 extern int __kmp_hws_requested;
909 extern int __kmp_hws_abs_flag; // absolute or per-item number requested
910 
911 /* ------------------------------------------------------------------------ */
912 
913 #define KMP_PAD(type, sz) \
914  (sizeof(type) + (sz - ((sizeof(type) - 1) % (sz)) - 1))
915 
916 // We need to avoid using -1 as a GTID as +1 is added to the gtid
917 // when storing it in a lock, and the value 0 is reserved.
918 #define KMP_GTID_DNE (-2) /* Does not exist */
919 #define KMP_GTID_SHUTDOWN (-3) /* Library is shutting down */
920 #define KMP_GTID_MONITOR (-4) /* Monitor thread ID */
921 #define KMP_GTID_UNKNOWN (-5) /* Is not known */
922 #define KMP_GTID_MIN (-6) /* Minimal gtid for low bound check in DEBUG */
923 
924 /* OpenMP 5.0 Memory Management support */
925 
926 #ifndef __OMP_H
927 // Duplicate type definitions from omp.h
928 typedef uintptr_t omp_uintptr_t;
929 
930 typedef enum {
931  omp_atk_sync_hint = 1,
932  omp_atk_alignment = 2,
933  omp_atk_access = 3,
934  omp_atk_pool_size = 4,
935  omp_atk_fallback = 5,
936  omp_atk_fb_data = 6,
937  omp_atk_pinned = 7,
938  omp_atk_partition = 8
939 } omp_alloctrait_key_t;
940 
941 typedef enum {
942  omp_atv_false = 0,
943  omp_atv_true = 1,
944  omp_atv_contended = 3,
945  omp_atv_uncontended = 4,
946  omp_atv_serialized = 5,
947  omp_atv_sequential = omp_atv_serialized, // (deprecated)
948  omp_atv_private = 6,
949  omp_atv_all = 7,
950  omp_atv_thread = 8,
951  omp_atv_pteam = 9,
952  omp_atv_cgroup = 10,
953  omp_atv_default_mem_fb = 11,
954  omp_atv_null_fb = 12,
955  omp_atv_abort_fb = 13,
956  omp_atv_allocator_fb = 14,
957  omp_atv_environment = 15,
958  omp_atv_nearest = 16,
959  omp_atv_blocked = 17,
960  omp_atv_interleaved = 18
961 } omp_alloctrait_value_t;
962 #define omp_atv_default ((omp_uintptr_t)-1)
963 
964 typedef void *omp_memspace_handle_t;
965 extern omp_memspace_handle_t const omp_default_mem_space;
966 extern omp_memspace_handle_t const omp_large_cap_mem_space;
967 extern omp_memspace_handle_t const omp_const_mem_space;
968 extern omp_memspace_handle_t const omp_high_bw_mem_space;
969 extern omp_memspace_handle_t const omp_low_lat_mem_space;
970 // Preview of target memory support
971 extern omp_memspace_handle_t const llvm_omp_target_host_mem_space;
972 extern omp_memspace_handle_t const llvm_omp_target_shared_mem_space;
973 extern omp_memspace_handle_t const llvm_omp_target_device_mem_space;
974 
975 typedef struct {
976  omp_alloctrait_key_t key;
977  omp_uintptr_t value;
978 } omp_alloctrait_t;
979 
980 typedef void *omp_allocator_handle_t;
981 extern omp_allocator_handle_t const omp_null_allocator;
982 extern omp_allocator_handle_t const omp_default_mem_alloc;
983 extern omp_allocator_handle_t const omp_large_cap_mem_alloc;
984 extern omp_allocator_handle_t const omp_const_mem_alloc;
985 extern omp_allocator_handle_t const omp_high_bw_mem_alloc;
986 extern omp_allocator_handle_t const omp_low_lat_mem_alloc;
987 extern omp_allocator_handle_t const omp_cgroup_mem_alloc;
988 extern omp_allocator_handle_t const omp_pteam_mem_alloc;
989 extern omp_allocator_handle_t const omp_thread_mem_alloc;
990 // Preview of target memory support
991 extern omp_allocator_handle_t const llvm_omp_target_host_mem_alloc;
992 extern omp_allocator_handle_t const llvm_omp_target_shared_mem_alloc;
993 extern omp_allocator_handle_t const llvm_omp_target_device_mem_alloc;
994 extern omp_allocator_handle_t const kmp_max_mem_alloc;
995 extern omp_allocator_handle_t __kmp_def_allocator;
996 
997 // end of duplicate type definitions from omp.h
998 #endif
999 
1000 extern int __kmp_memkind_available;
1001 
1002 typedef omp_memspace_handle_t kmp_memspace_t; // placeholder
1003 
1004 typedef struct kmp_allocator_t {
1005  omp_memspace_handle_t memspace;
1006  void **memkind; // pointer to memkind
1007  size_t alignment;
1008  omp_alloctrait_value_t fb;
1009  kmp_allocator_t *fb_data;
1010  kmp_uint64 pool_size;
1011  kmp_uint64 pool_used;
1012 } kmp_allocator_t;
1013 
1014 extern omp_allocator_handle_t __kmpc_init_allocator(int gtid,
1015  omp_memspace_handle_t,
1016  int ntraits,
1017  omp_alloctrait_t traits[]);
1018 extern void __kmpc_destroy_allocator(int gtid, omp_allocator_handle_t al);
1019 extern void __kmpc_set_default_allocator(int gtid, omp_allocator_handle_t al);
1020 extern omp_allocator_handle_t __kmpc_get_default_allocator(int gtid);
1021 // external interfaces, may be used by compiler
1022 extern void *__kmpc_alloc(int gtid, size_t sz, omp_allocator_handle_t al);
1023 extern void *__kmpc_aligned_alloc(int gtid, size_t align, size_t sz,
1024  omp_allocator_handle_t al);
1025 extern void *__kmpc_calloc(int gtid, size_t nmemb, size_t sz,
1026  omp_allocator_handle_t al);
1027 extern void *__kmpc_realloc(int gtid, void *ptr, size_t sz,
1028  omp_allocator_handle_t al,
1029  omp_allocator_handle_t free_al);
1030 extern void __kmpc_free(int gtid, void *ptr, omp_allocator_handle_t al);
1031 // internal interfaces, contain real implementation
1032 extern void *__kmp_alloc(int gtid, size_t align, size_t sz,
1033  omp_allocator_handle_t al);
1034 extern void *__kmp_calloc(int gtid, size_t align, size_t nmemb, size_t sz,
1035  omp_allocator_handle_t al);
1036 extern void *__kmp_realloc(int gtid, void *ptr, size_t sz,
1037  omp_allocator_handle_t al,
1038  omp_allocator_handle_t free_al);
1039 extern void ___kmpc_free(int gtid, void *ptr, omp_allocator_handle_t al);
1040 
1041 extern void __kmp_init_memkind();
1042 extern void __kmp_fini_memkind();
1043 extern void __kmp_init_target_mem();
1044 
1045 /* ------------------------------------------------------------------------ */
1046 
1047 #define KMP_UINT64_MAX \
1048  (~((kmp_uint64)1 << ((sizeof(kmp_uint64) * (1 << 3)) - 1)))
1049 
1050 #define KMP_MIN_NTH 1
1051 
1052 #ifndef KMP_MAX_NTH
1053 #if defined(PTHREAD_THREADS_MAX) && PTHREAD_THREADS_MAX < INT_MAX
1054 #define KMP_MAX_NTH PTHREAD_THREADS_MAX
1055 #else
1056 #define KMP_MAX_NTH INT_MAX
1057 #endif
1058 #endif /* KMP_MAX_NTH */
1059 
1060 #ifdef PTHREAD_STACK_MIN
1061 #define KMP_MIN_STKSIZE PTHREAD_STACK_MIN
1062 #else
1063 #define KMP_MIN_STKSIZE ((size_t)(32 * 1024))
1064 #endif
1065 
1066 #define KMP_MAX_STKSIZE (~((size_t)1 << ((sizeof(size_t) * (1 << 3)) - 1)))
1067 
1068 #if KMP_ARCH_X86
1069 #define KMP_DEFAULT_STKSIZE ((size_t)(2 * 1024 * 1024))
1070 #elif KMP_ARCH_X86_64
1071 #define KMP_DEFAULT_STKSIZE ((size_t)(4 * 1024 * 1024))
1072 #define KMP_BACKUP_STKSIZE ((size_t)(2 * 1024 * 1024))
1073 #else
1074 #define KMP_DEFAULT_STKSIZE ((size_t)(1024 * 1024))
1075 #endif
1076 
1077 #define KMP_DEFAULT_MALLOC_POOL_INCR ((size_t)(1024 * 1024))
1078 #define KMP_MIN_MALLOC_POOL_INCR ((size_t)(4 * 1024))
1079 #define KMP_MAX_MALLOC_POOL_INCR \
1080  (~((size_t)1 << ((sizeof(size_t) * (1 << 3)) - 1)))
1081 
1082 #define KMP_MIN_STKOFFSET (0)
1083 #define KMP_MAX_STKOFFSET KMP_MAX_STKSIZE
1084 #if KMP_OS_DARWIN
1085 #define KMP_DEFAULT_STKOFFSET KMP_MIN_STKOFFSET
1086 #else
1087 #define KMP_DEFAULT_STKOFFSET CACHE_LINE
1088 #endif
1089 
1090 #define KMP_MIN_STKPADDING (0)
1091 #define KMP_MAX_STKPADDING (2 * 1024 * 1024)
1092 
1093 #define KMP_BLOCKTIME_MULTIPLIER \
1094  (1000) /* number of blocktime units per second */
1095 #define KMP_MIN_BLOCKTIME (0)
1096 #define KMP_MAX_BLOCKTIME \
1097  (INT_MAX) /* Must be this for "infinite" setting the work */
1098 
1099 /* __kmp_blocktime is in milliseconds */
1100 #define KMP_DEFAULT_BLOCKTIME (__kmp_is_hybrid_cpu() ? (0) : (200))
1101 
1102 #if KMP_USE_MONITOR
1103 #define KMP_DEFAULT_MONITOR_STKSIZE ((size_t)(64 * 1024))
1104 #define KMP_MIN_MONITOR_WAKEUPS (1) // min times monitor wakes up per second
1105 #define KMP_MAX_MONITOR_WAKEUPS (1000) // max times monitor can wake up per sec
1106 
1107 /* Calculate new number of monitor wakeups for a specific block time based on
1108  previous monitor_wakeups. Only allow increasing number of wakeups */
1109 #define KMP_WAKEUPS_FROM_BLOCKTIME(blocktime, monitor_wakeups) \
1110  (((blocktime) == KMP_MAX_BLOCKTIME) ? (monitor_wakeups) \
1111  : ((blocktime) == KMP_MIN_BLOCKTIME) ? KMP_MAX_MONITOR_WAKEUPS \
1112  : ((monitor_wakeups) > (KMP_BLOCKTIME_MULTIPLIER / (blocktime))) \
1113  ? (monitor_wakeups) \
1114  : (KMP_BLOCKTIME_MULTIPLIER) / (blocktime))
1115 
1116 /* Calculate number of intervals for a specific block time based on
1117  monitor_wakeups */
1118 #define KMP_INTERVALS_FROM_BLOCKTIME(blocktime, monitor_wakeups) \
1119  (((blocktime) + (KMP_BLOCKTIME_MULTIPLIER / (monitor_wakeups)) - 1) / \
1120  (KMP_BLOCKTIME_MULTIPLIER / (monitor_wakeups)))
1121 #else
1122 #define KMP_BLOCKTIME(team, tid) \
1123  (get__bt_set(team, tid) ? get__blocktime(team, tid) : __kmp_dflt_blocktime)
1124 #if KMP_OS_UNIX && (KMP_ARCH_X86 || KMP_ARCH_X86_64)
1125 // HW TSC is used to reduce overhead (clock tick instead of nanosecond).
1126 extern kmp_uint64 __kmp_ticks_per_msec;
1127 #if KMP_COMPILER_ICC
1128 #define KMP_NOW() ((kmp_uint64)_rdtsc())
1129 #else
1130 #define KMP_NOW() __kmp_hardware_timestamp()
1131 #endif
1132 #define KMP_NOW_MSEC() (KMP_NOW() / __kmp_ticks_per_msec)
1133 #define KMP_BLOCKTIME_INTERVAL(team, tid) \
1134  (KMP_BLOCKTIME(team, tid) * __kmp_ticks_per_msec)
1135 #define KMP_BLOCKING(goal, count) ((goal) > KMP_NOW())
1136 #else
1137 // System time is retrieved sporadically while blocking.
1138 extern kmp_uint64 __kmp_now_nsec();
1139 #define KMP_NOW() __kmp_now_nsec()
1140 #define KMP_NOW_MSEC() (KMP_NOW() / KMP_USEC_PER_SEC)
1141 #define KMP_BLOCKTIME_INTERVAL(team, tid) \
1142  (KMP_BLOCKTIME(team, tid) * KMP_USEC_PER_SEC)
1143 #define KMP_BLOCKING(goal, count) ((count) % 1000 != 0 || (goal) > KMP_NOW())
1144 #endif
1145 #endif // KMP_USE_MONITOR
1146 
1147 #define KMP_MIN_STATSCOLS 40
1148 #define KMP_MAX_STATSCOLS 4096
1149 #define KMP_DEFAULT_STATSCOLS 80
1150 
1151 #define KMP_MIN_INTERVAL 0
1152 #define KMP_MAX_INTERVAL (INT_MAX - 1)
1153 #define KMP_DEFAULT_INTERVAL 0
1154 
1155 #define KMP_MIN_CHUNK 1
1156 #define KMP_MAX_CHUNK (INT_MAX - 1)
1157 #define KMP_DEFAULT_CHUNK 1
1158 
1159 #define KMP_MIN_DISP_NUM_BUFF 1
1160 #define KMP_DFLT_DISP_NUM_BUFF 7
1161 #define KMP_MAX_DISP_NUM_BUFF 4096
1162 
1163 #define KMP_MAX_ORDERED 8
1164 
1165 #define KMP_MAX_FIELDS 32
1166 
1167 #define KMP_MAX_BRANCH_BITS 31
1168 
1169 #define KMP_MAX_ACTIVE_LEVELS_LIMIT INT_MAX
1170 
1171 #define KMP_MAX_DEFAULT_DEVICE_LIMIT INT_MAX
1172 
1173 #define KMP_MAX_TASK_PRIORITY_LIMIT INT_MAX
1174 
1175 /* Minimum number of threads before switch to TLS gtid (experimentally
1176  determined) */
1177 /* josh TODO: what about OS X* tuning? */
1178 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
1179 #define KMP_TLS_GTID_MIN 5
1180 #else
1181 #define KMP_TLS_GTID_MIN INT_MAX
1182 #endif
1183 
1184 #define KMP_MASTER_TID(tid) (0 == (tid))
1185 #define KMP_WORKER_TID(tid) (0 != (tid))
1186 
1187 #define KMP_MASTER_GTID(gtid) (0 == __kmp_tid_from_gtid((gtid)))
1188 #define KMP_WORKER_GTID(gtid) (0 != __kmp_tid_from_gtid((gtid)))
1189 #define KMP_INITIAL_GTID(gtid) (0 == (gtid))
1190 
1191 #ifndef TRUE
1192 #define FALSE 0
1193 #define TRUE (!FALSE)
1194 #endif
1195 
1196 /* NOTE: all of the following constants must be even */
1197 
1198 #if KMP_OS_WINDOWS
1199 #define KMP_INIT_WAIT 64U /* initial number of spin-tests */
1200 #define KMP_NEXT_WAIT 32U /* susequent number of spin-tests */
1201 #elif KMP_OS_LINUX
1202 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1203 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1204 #elif KMP_OS_DARWIN
1205 /* TODO: tune for KMP_OS_DARWIN */
1206 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1207 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1208 #elif KMP_OS_DRAGONFLY
1209 /* TODO: tune for KMP_OS_DRAGONFLY */
1210 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1211 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1212 #elif KMP_OS_FREEBSD
1213 /* TODO: tune for KMP_OS_FREEBSD */
1214 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1215 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1216 #elif KMP_OS_NETBSD
1217 /* TODO: tune for KMP_OS_NETBSD */
1218 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1219 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1220 #elif KMP_OS_HURD
1221 /* TODO: tune for KMP_OS_HURD */
1222 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1223 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1224 #elif KMP_OS_OPENBSD
1225 /* TODO: tune for KMP_OS_OPENBSD */
1226 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1227 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1228 #endif
1229 
1230 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
1231 typedef struct kmp_cpuid {
1232  kmp_uint32 eax;
1233  kmp_uint32 ebx;
1234  kmp_uint32 ecx;
1235  kmp_uint32 edx;
1236 } kmp_cpuid_t;
1237 
1238 typedef struct kmp_cpuinfo_flags_t {
1239  unsigned sse2 : 1; // 0 if SSE2 instructions are not supported, 1 otherwise.
1240  unsigned rtm : 1; // 0 if RTM instructions are not supported, 1 otherwise.
1241  unsigned hybrid : 1;
1242  unsigned reserved : 29; // Ensure size of 32 bits
1243 } kmp_cpuinfo_flags_t;
1244 
1245 typedef struct kmp_cpuinfo {
1246  int initialized; // If 0, other fields are not initialized.
1247  int signature; // CPUID(1).EAX
1248  int family; // CPUID(1).EAX[27:20]+CPUID(1).EAX[11:8] (Extended Family+Family)
1249  int model; // ( CPUID(1).EAX[19:16] << 4 ) + CPUID(1).EAX[7:4] ( ( Extended
1250  // Model << 4 ) + Model)
1251  int stepping; // CPUID(1).EAX[3:0] ( Stepping )
1252  kmp_cpuinfo_flags_t flags;
1253  int apic_id;
1254  int physical_id;
1255  int logical_id;
1256  kmp_uint64 frequency; // Nominal CPU frequency in Hz.
1257  char name[3 * sizeof(kmp_cpuid_t)]; // CPUID(0x80000002,0x80000003,0x80000004)
1258 } kmp_cpuinfo_t;
1259 
1260 extern void __kmp_query_cpuid(kmp_cpuinfo_t *p);
1261 
1262 #if KMP_OS_UNIX
1263 // subleaf is only needed for cache and topology discovery and can be set to
1264 // zero in most cases
1265 static inline void __kmp_x86_cpuid(int leaf, int subleaf, struct kmp_cpuid *p) {
1266  __asm__ __volatile__("cpuid"
1267  : "=a"(p->eax), "=b"(p->ebx), "=c"(p->ecx), "=d"(p->edx)
1268  : "a"(leaf), "c"(subleaf));
1269 }
1270 // Load p into FPU control word
1271 static inline void __kmp_load_x87_fpu_control_word(const kmp_int16 *p) {
1272  __asm__ __volatile__("fldcw %0" : : "m"(*p));
1273 }
1274 // Store FPU control word into p
1275 static inline void __kmp_store_x87_fpu_control_word(kmp_int16 *p) {
1276  __asm__ __volatile__("fstcw %0" : "=m"(*p));
1277 }
1278 static inline void __kmp_clear_x87_fpu_status_word() {
1279 #if KMP_MIC
1280  // 32-bit protected mode x87 FPU state
1281  struct x87_fpu_state {
1282  unsigned cw;
1283  unsigned sw;
1284  unsigned tw;
1285  unsigned fip;
1286  unsigned fips;
1287  unsigned fdp;
1288  unsigned fds;
1289  };
1290  struct x87_fpu_state fpu_state = {0, 0, 0, 0, 0, 0, 0};
1291  __asm__ __volatile__("fstenv %0\n\t" // store FP env
1292  "andw $0x7f00, %1\n\t" // clear 0-7,15 bits of FP SW
1293  "fldenv %0\n\t" // load FP env back
1294  : "+m"(fpu_state), "+m"(fpu_state.sw));
1295 #else
1296  __asm__ __volatile__("fnclex");
1297 #endif // KMP_MIC
1298 }
1299 #if __SSE__
1300 static inline void __kmp_load_mxcsr(const kmp_uint32 *p) { _mm_setcsr(*p); }
1301 static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = _mm_getcsr(); }
1302 #else
1303 static inline void __kmp_load_mxcsr(const kmp_uint32 *p) {}
1304 static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = 0; }
1305 #endif
1306 #else
1307 // Windows still has these as external functions in assembly file
1308 extern void __kmp_x86_cpuid(int mode, int mode2, struct kmp_cpuid *p);
1309 extern void __kmp_load_x87_fpu_control_word(const kmp_int16 *p);
1310 extern void __kmp_store_x87_fpu_control_word(kmp_int16 *p);
1311 extern void __kmp_clear_x87_fpu_status_word();
1312 static inline void __kmp_load_mxcsr(const kmp_uint32 *p) { _mm_setcsr(*p); }
1313 static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = _mm_getcsr(); }
1314 #endif // KMP_OS_UNIX
1315 
1316 #define KMP_X86_MXCSR_MASK 0xffffffc0 /* ignore status flags (6 lsb) */
1317 
1318 #if KMP_ARCH_X86
1319 extern void __kmp_x86_pause(void);
1320 #elif KMP_MIC
1321 // Performance testing on KNC (C0QS-7120 P/A/X/D, 61-core, 16 GB Memory) showed
1322 // regression after removal of extra PAUSE from spin loops. Changing
1323 // the delay from 100 to 300 showed even better performance than double PAUSE
1324 // on Spec OMP2001 and LCPC tasking tests, no regressions on EPCC.
1325 static inline void __kmp_x86_pause(void) { _mm_delay_32(300); }
1326 #else
1327 static inline void __kmp_x86_pause(void) { _mm_pause(); }
1328 #endif
1329 #define KMP_CPU_PAUSE() __kmp_x86_pause()
1330 #elif KMP_ARCH_PPC64
1331 #define KMP_PPC64_PRI_LOW() __asm__ volatile("or 1, 1, 1")
1332 #define KMP_PPC64_PRI_MED() __asm__ volatile("or 2, 2, 2")
1333 #define KMP_PPC64_PRI_LOC_MB() __asm__ volatile("" : : : "memory")
1334 #define KMP_CPU_PAUSE() \
1335  do { \
1336  KMP_PPC64_PRI_LOW(); \
1337  KMP_PPC64_PRI_MED(); \
1338  KMP_PPC64_PRI_LOC_MB(); \
1339  } while (0)
1340 #else
1341 #define KMP_CPU_PAUSE() /* nothing to do */
1342 #endif
1343 
1344 #define KMP_INIT_YIELD(count) \
1345  { (count) = __kmp_yield_init; }
1346 
1347 #define KMP_OVERSUBSCRIBED \
1348  (TCR_4(__kmp_nth) > (__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc))
1349 
1350 #define KMP_TRY_YIELD \
1351  ((__kmp_use_yield == 1) || (__kmp_use_yield == 2 && (KMP_OVERSUBSCRIBED)))
1352 
1353 #define KMP_TRY_YIELD_OVERSUB \
1354  ((__kmp_use_yield == 1 || __kmp_use_yield == 2) && (KMP_OVERSUBSCRIBED))
1355 
1356 #define KMP_YIELD(cond) \
1357  { \
1358  KMP_CPU_PAUSE(); \
1359  if ((cond) && (KMP_TRY_YIELD)) \
1360  __kmp_yield(); \
1361  }
1362 
1363 #define KMP_YIELD_OVERSUB() \
1364  { \
1365  KMP_CPU_PAUSE(); \
1366  if ((KMP_TRY_YIELD_OVERSUB)) \
1367  __kmp_yield(); \
1368  }
1369 
1370 // Note the decrement of 2 in the following Macros. With KMP_LIBRARY=turnaround,
1371 // there should be no yielding since initial value from KMP_INIT_YIELD() is odd.
1372 #define KMP_YIELD_SPIN(count) \
1373  { \
1374  KMP_CPU_PAUSE(); \
1375  if (KMP_TRY_YIELD) { \
1376  (count) -= 2; \
1377  if (!(count)) { \
1378  __kmp_yield(); \
1379  (count) = __kmp_yield_next; \
1380  } \
1381  } \
1382  }
1383 
1384 #define KMP_YIELD_OVERSUB_ELSE_SPIN(count) \
1385  { \
1386  KMP_CPU_PAUSE(); \
1387  if ((KMP_TRY_YIELD_OVERSUB)) \
1388  __kmp_yield(); \
1389  else if (__kmp_use_yield == 1) { \
1390  (count) -= 2; \
1391  if (!(count)) { \
1392  __kmp_yield(); \
1393  (count) = __kmp_yield_next; \
1394  } \
1395  } \
1396  }
1397 
1398 // User-level Monitor/Mwait
1399 #if KMP_HAVE_UMWAIT
1400 // We always try for UMWAIT first
1401 #if KMP_HAVE_WAITPKG_INTRINSICS
1402 #if KMP_HAVE_IMMINTRIN_H
1403 #include <immintrin.h>
1404 #elif KMP_HAVE_INTRIN_H
1405 #include <intrin.h>
1406 #endif
1407 #endif // KMP_HAVE_WAITPKG_INTRINSICS
1408 KMP_ATTRIBUTE_TARGET_WAITPKG
1409 static inline int __kmp_tpause(uint32_t hint, uint64_t counter) {
1410 #if !KMP_HAVE_WAITPKG_INTRINSICS
1411  uint32_t timeHi = uint32_t(counter >> 32);
1412  uint32_t timeLo = uint32_t(counter & 0xffffffff);
1413  char flag;
1414  __asm__ volatile("#tpause\n.byte 0x66, 0x0F, 0xAE, 0xF1\n"
1415  "setb %0"
1416  : "=r"(flag)
1417  : "a"(timeLo), "d"(timeHi), "c"(hint)
1418  :);
1419  return flag;
1420 #else
1421  return _tpause(hint, counter);
1422 #endif
1423 }
1424 KMP_ATTRIBUTE_TARGET_WAITPKG
1425 static inline void __kmp_umonitor(void *cacheline) {
1426 #if !KMP_HAVE_WAITPKG_INTRINSICS
1427  __asm__ volatile("# umonitor\n.byte 0xF3, 0x0F, 0xAE, 0x01 "
1428  :
1429  : "a"(cacheline)
1430  :);
1431 #else
1432  _umonitor(cacheline);
1433 #endif
1434 }
1435 KMP_ATTRIBUTE_TARGET_WAITPKG
1436 static inline int __kmp_umwait(uint32_t hint, uint64_t counter) {
1437 #if !KMP_HAVE_WAITPKG_INTRINSICS
1438  uint32_t timeHi = uint32_t(counter >> 32);
1439  uint32_t timeLo = uint32_t(counter & 0xffffffff);
1440  char flag;
1441  __asm__ volatile("#umwait\n.byte 0xF2, 0x0F, 0xAE, 0xF1\n"
1442  "setb %0"
1443  : "=r"(flag)
1444  : "a"(timeLo), "d"(timeHi), "c"(hint)
1445  :);
1446  return flag;
1447 #else
1448  return _umwait(hint, counter);
1449 #endif
1450 }
1451 #elif KMP_HAVE_MWAIT
1452 #if KMP_OS_UNIX
1453 #include <pmmintrin.h>
1454 #else
1455 #include <intrin.h>
1456 #endif
1457 #if KMP_OS_UNIX
1458 __attribute__((target("sse3")))
1459 #endif
1460 static inline void
1461 __kmp_mm_monitor(void *cacheline, unsigned extensions, unsigned hints) {
1462  _mm_monitor(cacheline, extensions, hints);
1463 }
1464 #if KMP_OS_UNIX
1465 __attribute__((target("sse3")))
1466 #endif
1467 static inline void
1468 __kmp_mm_mwait(unsigned extensions, unsigned hints) {
1469  _mm_mwait(extensions, hints);
1470 }
1471 #endif // KMP_HAVE_UMWAIT
1472 
1473 /* ------------------------------------------------------------------------ */
1474 /* Support datatypes for the orphaned construct nesting checks. */
1475 /* ------------------------------------------------------------------------ */
1476 
1477 /* When adding to this enum, add its corresponding string in cons_text_c[]
1478  * array in kmp_error.cpp */
1479 enum cons_type {
1480  ct_none,
1481  ct_parallel,
1482  ct_pdo,
1483  ct_pdo_ordered,
1484  ct_psections,
1485  ct_psingle,
1486  ct_critical,
1487  ct_ordered_in_parallel,
1488  ct_ordered_in_pdo,
1489  ct_master,
1490  ct_reduce,
1491  ct_barrier,
1492  ct_masked
1493 };
1494 
1495 #define IS_CONS_TYPE_ORDERED(ct) ((ct) == ct_pdo_ordered)
1496 
1497 struct cons_data {
1498  ident_t const *ident;
1499  enum cons_type type;
1500  int prev;
1501  kmp_user_lock_p
1502  name; /* address exclusively for critical section name comparison */
1503 };
1504 
1505 struct cons_header {
1506  int p_top, w_top, s_top;
1507  int stack_size, stack_top;
1508  struct cons_data *stack_data;
1509 };
1510 
1511 struct kmp_region_info {
1512  char *text;
1513  int offset[KMP_MAX_FIELDS];
1514  int length[KMP_MAX_FIELDS];
1515 };
1516 
1517 /* ---------------------------------------------------------------------- */
1518 /* ---------------------------------------------------------------------- */
1519 
1520 #if KMP_OS_WINDOWS
1521 typedef HANDLE kmp_thread_t;
1522 typedef DWORD kmp_key_t;
1523 #endif /* KMP_OS_WINDOWS */
1524 
1525 #if KMP_OS_UNIX
1526 typedef pthread_t kmp_thread_t;
1527 typedef pthread_key_t kmp_key_t;
1528 #endif
1529 
1530 extern kmp_key_t __kmp_gtid_threadprivate_key;
1531 
1532 typedef struct kmp_sys_info {
1533  long maxrss; /* the maximum resident set size utilized (in kilobytes) */
1534  long minflt; /* the number of page faults serviced without any I/O */
1535  long majflt; /* the number of page faults serviced that required I/O */
1536  long nswap; /* the number of times a process was "swapped" out of memory */
1537  long inblock; /* the number of times the file system had to perform input */
1538  long oublock; /* the number of times the file system had to perform output */
1539  long nvcsw; /* the number of times a context switch was voluntarily */
1540  long nivcsw; /* the number of times a context switch was forced */
1541 } kmp_sys_info_t;
1542 
1543 #if USE_ITT_BUILD
1544 // We cannot include "kmp_itt.h" due to circular dependency. Declare the only
1545 // required type here. Later we will check the type meets requirements.
1546 typedef int kmp_itt_mark_t;
1547 #define KMP_ITT_DEBUG 0
1548 #endif /* USE_ITT_BUILD */
1549 
1550 typedef kmp_int32 kmp_critical_name[8];
1551 
1561 typedef void (*kmpc_micro)(kmp_int32 *global_tid, kmp_int32 *bound_tid, ...);
1562 typedef void (*kmpc_micro_bound)(kmp_int32 *bound_tid, kmp_int32 *bound_nth,
1563  ...);
1564 
1569 /* ---------------------------------------------------------------------------
1570  */
1571 /* Threadprivate initialization/finalization function declarations */
1572 
1573 /* for non-array objects: __kmpc_threadprivate_register() */
1574 
1579 typedef void *(*kmpc_ctor)(void *);
1580 
1585 typedef void (*kmpc_dtor)(
1586  void * /*, size_t */); /* 2nd arg: magic number for KCC unused by Intel
1587  compiler */
1592 typedef void *(*kmpc_cctor)(void *, void *);
1593 
1594 /* for array objects: __kmpc_threadprivate_register_vec() */
1595 /* First arg: "this" pointer */
1596 /* Last arg: number of array elements */
1602 typedef void *(*kmpc_ctor_vec)(void *, size_t);
1608 typedef void (*kmpc_dtor_vec)(void *, size_t);
1614 typedef void *(*kmpc_cctor_vec)(void *, void *,
1615  size_t); /* function unused by compiler */
1616 
1621 /* keeps tracked of threadprivate cache allocations for cleanup later */
1622 typedef struct kmp_cached_addr {
1623  void **addr; /* address of allocated cache */
1624  void ***compiler_cache; /* pointer to compiler's cache */
1625  void *data; /* pointer to global data */
1626  struct kmp_cached_addr *next; /* pointer to next cached address */
1627 } kmp_cached_addr_t;
1628 
1629 struct private_data {
1630  struct private_data *next; /* The next descriptor in the list */
1631  void *data; /* The data buffer for this descriptor */
1632  int more; /* The repeat count for this descriptor */
1633  size_t size; /* The data size for this descriptor */
1634 };
1635 
1636 struct private_common {
1637  struct private_common *next;
1638  struct private_common *link;
1639  void *gbl_addr;
1640  void *par_addr; /* par_addr == gbl_addr for PRIMARY thread */
1641  size_t cmn_size;
1642 };
1643 
1644 struct shared_common {
1645  struct shared_common *next;
1646  struct private_data *pod_init;
1647  void *obj_init;
1648  void *gbl_addr;
1649  union {
1650  kmpc_ctor ctor;
1651  kmpc_ctor_vec ctorv;
1652  } ct;
1653  union {
1654  kmpc_cctor cctor;
1655  kmpc_cctor_vec cctorv;
1656  } cct;
1657  union {
1658  kmpc_dtor dtor;
1659  kmpc_dtor_vec dtorv;
1660  } dt;
1661  size_t vec_len;
1662  int is_vec;
1663  size_t cmn_size;
1664 };
1665 
1666 #define KMP_HASH_TABLE_LOG2 9 /* log2 of the hash table size */
1667 #define KMP_HASH_TABLE_SIZE \
1668  (1 << KMP_HASH_TABLE_LOG2) /* size of the hash table */
1669 #define KMP_HASH_SHIFT 3 /* throw away this many low bits from the address */
1670 #define KMP_HASH(x) \
1671  ((((kmp_uintptr_t)x) >> KMP_HASH_SHIFT) & (KMP_HASH_TABLE_SIZE - 1))
1672 
1673 struct common_table {
1674  struct private_common *data[KMP_HASH_TABLE_SIZE];
1675 };
1676 
1677 struct shared_table {
1678  struct shared_common *data[KMP_HASH_TABLE_SIZE];
1679 };
1680 
1681 /* ------------------------------------------------------------------------ */
1682 
1683 #if KMP_USE_HIER_SCHED
1684 // Shared barrier data that exists inside a single unit of the scheduling
1685 // hierarchy
1686 typedef struct kmp_hier_private_bdata_t {
1687  kmp_int32 num_active;
1688  kmp_uint64 index;
1689  kmp_uint64 wait_val[2];
1690 } kmp_hier_private_bdata_t;
1691 #endif
1692 
1693 typedef struct kmp_sched_flags {
1694  unsigned ordered : 1;
1695  unsigned nomerge : 1;
1696  unsigned contains_last : 1;
1697 #if KMP_USE_HIER_SCHED
1698  unsigned use_hier : 1;
1699  unsigned unused : 28;
1700 #else
1701  unsigned unused : 29;
1702 #endif
1703 } kmp_sched_flags_t;
1704 
1705 KMP_BUILD_ASSERT(sizeof(kmp_sched_flags_t) == 4);
1706 
1707 #if KMP_STATIC_STEAL_ENABLED
1708 typedef struct KMP_ALIGN_CACHE dispatch_private_info32 {
1709  kmp_int32 count;
1710  kmp_int32 ub;
1711  /* Adding KMP_ALIGN_CACHE here doesn't help / can hurt performance */
1712  kmp_int32 lb;
1713  kmp_int32 st;
1714  kmp_int32 tc;
1715  kmp_lock_t *steal_lock; // lock used for chunk stealing
1716  // KMP_ALIGN(32) ensures (if the KMP_ALIGN macro is turned on)
1717  // a) parm3 is properly aligned and
1718  // b) all parm1-4 are on the same cache line.
1719  // Because of parm1-4 are used together, performance seems to be better
1720  // if they are on the same cache line (not measured though).
1721 
1722  struct KMP_ALIGN(32) { // AC: changed 16 to 32 in order to simplify template
1723  kmp_int32 parm1; // structures in kmp_dispatch.cpp. This should
1724  kmp_int32 parm2; // make no real change at least while padding is off.
1725  kmp_int32 parm3;
1726  kmp_int32 parm4;
1727  };
1728 
1729  kmp_uint32 ordered_lower;
1730  kmp_uint32 ordered_upper;
1731 #if KMP_OS_WINDOWS
1732  kmp_int32 last_upper;
1733 #endif /* KMP_OS_WINDOWS */
1734 } dispatch_private_info32_t;
1735 
1736 typedef struct KMP_ALIGN_CACHE dispatch_private_info64 {
1737  kmp_int64 count; // current chunk number for static & static-steal scheduling
1738  kmp_int64 ub; /* upper-bound */
1739  /* Adding KMP_ALIGN_CACHE here doesn't help / can hurt performance */
1740  kmp_int64 lb; /* lower-bound */
1741  kmp_int64 st; /* stride */
1742  kmp_int64 tc; /* trip count (number of iterations) */
1743  kmp_lock_t *steal_lock; // lock used for chunk stealing
1744  /* parm[1-4] are used in different ways by different scheduling algorithms */
1745 
1746  // KMP_ALIGN( 32 ) ensures ( if the KMP_ALIGN macro is turned on )
1747  // a) parm3 is properly aligned and
1748  // b) all parm1-4 are in the same cache line.
1749  // Because of parm1-4 are used together, performance seems to be better
1750  // if they are in the same line (not measured though).
1751 
1752  struct KMP_ALIGN(32) {
1753  kmp_int64 parm1;
1754  kmp_int64 parm2;
1755  kmp_int64 parm3;
1756  kmp_int64 parm4;
1757  };
1758 
1759  kmp_uint64 ordered_lower;
1760  kmp_uint64 ordered_upper;
1761 #if KMP_OS_WINDOWS
1762  kmp_int64 last_upper;
1763 #endif /* KMP_OS_WINDOWS */
1764 } dispatch_private_info64_t;
1765 #else /* KMP_STATIC_STEAL_ENABLED */
1766 typedef struct KMP_ALIGN_CACHE dispatch_private_info32 {
1767  kmp_int32 lb;
1768  kmp_int32 ub;
1769  kmp_int32 st;
1770  kmp_int32 tc;
1771 
1772  kmp_int32 parm1;
1773  kmp_int32 parm2;
1774  kmp_int32 parm3;
1775  kmp_int32 parm4;
1776 
1777  kmp_int32 count;
1778 
1779  kmp_uint32 ordered_lower;
1780  kmp_uint32 ordered_upper;
1781 #if KMP_OS_WINDOWS
1782  kmp_int32 last_upper;
1783 #endif /* KMP_OS_WINDOWS */
1784 } dispatch_private_info32_t;
1785 
1786 typedef struct KMP_ALIGN_CACHE dispatch_private_info64 {
1787  kmp_int64 lb; /* lower-bound */
1788  kmp_int64 ub; /* upper-bound */
1789  kmp_int64 st; /* stride */
1790  kmp_int64 tc; /* trip count (number of iterations) */
1791 
1792  /* parm[1-4] are used in different ways by different scheduling algorithms */
1793  kmp_int64 parm1;
1794  kmp_int64 parm2;
1795  kmp_int64 parm3;
1796  kmp_int64 parm4;
1797 
1798  kmp_int64 count; /* current chunk number for static scheduling */
1799 
1800  kmp_uint64 ordered_lower;
1801  kmp_uint64 ordered_upper;
1802 #if KMP_OS_WINDOWS
1803  kmp_int64 last_upper;
1804 #endif /* KMP_OS_WINDOWS */
1805 } dispatch_private_info64_t;
1806 #endif /* KMP_STATIC_STEAL_ENABLED */
1807 
1808 typedef struct KMP_ALIGN_CACHE dispatch_private_info {
1809  union private_info {
1810  dispatch_private_info32_t p32;
1811  dispatch_private_info64_t p64;
1812  } u;
1813  enum sched_type schedule; /* scheduling algorithm */
1814  kmp_sched_flags_t flags; /* flags (e.g., ordered, nomerge, etc.) */
1815  std::atomic<kmp_uint32> steal_flag; // static_steal only, state of a buffer
1816  kmp_int32 ordered_bumped;
1817  // Stack of buffers for nest of serial regions
1818  struct dispatch_private_info *next;
1819  kmp_int32 type_size; /* the size of types in private_info */
1820 #if KMP_USE_HIER_SCHED
1821  kmp_int32 hier_id;
1822  void *parent; /* hierarchical scheduling parent pointer */
1823 #endif
1824  enum cons_type pushed_ws;
1825 } dispatch_private_info_t;
1826 
1827 typedef struct dispatch_shared_info32 {
1828  /* chunk index under dynamic, number of idle threads under static-steal;
1829  iteration index otherwise */
1830  volatile kmp_uint32 iteration;
1831  volatile kmp_int32 num_done;
1832  volatile kmp_uint32 ordered_iteration;
1833  // Dummy to retain the structure size after making ordered_iteration scalar
1834  kmp_int32 ordered_dummy[KMP_MAX_ORDERED - 1];
1835 } dispatch_shared_info32_t;
1836 
1837 typedef struct dispatch_shared_info64 {
1838  /* chunk index under dynamic, number of idle threads under static-steal;
1839  iteration index otherwise */
1840  volatile kmp_uint64 iteration;
1841  volatile kmp_int64 num_done;
1842  volatile kmp_uint64 ordered_iteration;
1843  // Dummy to retain the structure size after making ordered_iteration scalar
1844  kmp_int64 ordered_dummy[KMP_MAX_ORDERED - 3];
1845 } dispatch_shared_info64_t;
1846 
1847 typedef struct dispatch_shared_info {
1848  union shared_info {
1849  dispatch_shared_info32_t s32;
1850  dispatch_shared_info64_t s64;
1851  } u;
1852  volatile kmp_uint32 buffer_index;
1853  volatile kmp_int32 doacross_buf_idx; // teamwise index
1854  volatile kmp_uint32 *doacross_flags; // shared array of iteration flags (0/1)
1855  kmp_int32 doacross_num_done; // count finished threads
1856 #if KMP_USE_HIER_SCHED
1857  void *hier;
1858 #endif
1859 #if KMP_USE_HWLOC
1860  // When linking with libhwloc, the ORDERED EPCC test slows down on big
1861  // machines (> 48 cores). Performance analysis showed that a cache thrash
1862  // was occurring and this padding helps alleviate the problem.
1863  char padding[64];
1864 #endif
1865 } dispatch_shared_info_t;
1866 
1867 typedef struct kmp_disp {
1868  /* Vector for ORDERED SECTION */
1869  void (*th_deo_fcn)(int *gtid, int *cid, ident_t *);
1870  /* Vector for END ORDERED SECTION */
1871  void (*th_dxo_fcn)(int *gtid, int *cid, ident_t *);
1872 
1873  dispatch_shared_info_t *th_dispatch_sh_current;
1874  dispatch_private_info_t *th_dispatch_pr_current;
1875 
1876  dispatch_private_info_t *th_disp_buffer;
1877  kmp_uint32 th_disp_index;
1878  kmp_int32 th_doacross_buf_idx; // thread's doacross buffer index
1879  volatile kmp_uint32 *th_doacross_flags; // pointer to shared array of flags
1880  kmp_int64 *th_doacross_info; // info on loop bounds
1881 #if KMP_USE_INTERNODE_ALIGNMENT
1882  char more_padding[INTERNODE_CACHE_LINE];
1883 #endif
1884 } kmp_disp_t;
1885 
1886 /* ------------------------------------------------------------------------ */
1887 /* Barrier stuff */
1888 
1889 /* constants for barrier state update */
1890 #define KMP_INIT_BARRIER_STATE 0 /* should probably start from zero */
1891 #define KMP_BARRIER_SLEEP_BIT 0 /* bit used for suspend/sleep part of state */
1892 #define KMP_BARRIER_UNUSED_BIT 1 // bit that must never be set for valid state
1893 #define KMP_BARRIER_BUMP_BIT 2 /* lsb used for bump of go/arrived state */
1894 
1895 #define KMP_BARRIER_SLEEP_STATE (1 << KMP_BARRIER_SLEEP_BIT)
1896 #define KMP_BARRIER_UNUSED_STATE (1 << KMP_BARRIER_UNUSED_BIT)
1897 #define KMP_BARRIER_STATE_BUMP (1 << KMP_BARRIER_BUMP_BIT)
1898 
1899 #if (KMP_BARRIER_SLEEP_BIT >= KMP_BARRIER_BUMP_BIT)
1900 #error "Barrier sleep bit must be smaller than barrier bump bit"
1901 #endif
1902 #if (KMP_BARRIER_UNUSED_BIT >= KMP_BARRIER_BUMP_BIT)
1903 #error "Barrier unused bit must be smaller than barrier bump bit"
1904 #endif
1905 
1906 // Constants for release barrier wait state: currently, hierarchical only
1907 #define KMP_BARRIER_NOT_WAITING 0 // Normal state; worker not in wait_sleep
1908 #define KMP_BARRIER_OWN_FLAG \
1909  1 // Normal state; worker waiting on own b_go flag in release
1910 #define KMP_BARRIER_PARENT_FLAG \
1911  2 // Special state; worker waiting on parent's b_go flag in release
1912 #define KMP_BARRIER_SWITCH_TO_OWN_FLAG \
1913  3 // Special state; tells worker to shift from parent to own b_go
1914 #define KMP_BARRIER_SWITCHING \
1915  4 // Special state; worker resets appropriate flag on wake-up
1916 
1917 #define KMP_NOT_SAFE_TO_REAP \
1918  0 // Thread th_reap_state: not safe to reap (tasking)
1919 #define KMP_SAFE_TO_REAP 1 // Thread th_reap_state: safe to reap (not tasking)
1920 
1921 // The flag_type describes the storage used for the flag.
1922 enum flag_type {
1923  flag32,
1924  flag64,
1925  atomic_flag64,
1926  flag_oncore,
1927  flag_unset
1928 };
1929 
1930 enum barrier_type {
1931  bs_plain_barrier = 0, /* 0, All non-fork/join barriers (except reduction
1932  barriers if enabled) */
1933  bs_forkjoin_barrier, /* 1, All fork/join (parallel region) barriers */
1934 #if KMP_FAST_REDUCTION_BARRIER
1935  bs_reduction_barrier, /* 2, All barriers that are used in reduction */
1936 #endif // KMP_FAST_REDUCTION_BARRIER
1937  bs_last_barrier /* Just a placeholder to mark the end */
1938 };
1939 
1940 // to work with reduction barriers just like with plain barriers
1941 #if !KMP_FAST_REDUCTION_BARRIER
1942 #define bs_reduction_barrier bs_plain_barrier
1943 #endif // KMP_FAST_REDUCTION_BARRIER
1944 
1945 typedef enum kmp_bar_pat { /* Barrier communication patterns */
1946  bp_linear_bar =
1947  0, /* Single level (degenerate) tree */
1948  bp_tree_bar =
1949  1, /* Balanced tree with branching factor 2^n */
1950  bp_hyper_bar = 2, /* Hypercube-embedded tree with min
1951  branching factor 2^n */
1952  bp_hierarchical_bar = 3, /* Machine hierarchy tree */
1953  bp_dist_bar = 4, /* Distributed barrier */
1954  bp_last_bar /* Placeholder to mark the end */
1955 } kmp_bar_pat_e;
1956 
1957 #define KMP_BARRIER_ICV_PUSH 1
1958 
1959 /* Record for holding the values of the internal controls stack records */
1960 typedef struct kmp_internal_control {
1961  int serial_nesting_level; /* corresponds to the value of the
1962  th_team_serialized field */
1963  kmp_int8 dynamic; /* internal control for dynamic adjustment of threads (per
1964  thread) */
1965  kmp_int8
1966  bt_set; /* internal control for whether blocktime is explicitly set */
1967  int blocktime; /* internal control for blocktime */
1968 #if KMP_USE_MONITOR
1969  int bt_intervals; /* internal control for blocktime intervals */
1970 #endif
1971  int nproc; /* internal control for #threads for next parallel region (per
1972  thread) */
1973  int thread_limit; /* internal control for thread-limit-var */
1974  int max_active_levels; /* internal control for max_active_levels */
1975  kmp_r_sched_t
1976  sched; /* internal control for runtime schedule {sched,chunk} pair */
1977  kmp_proc_bind_t proc_bind; /* internal control for affinity */
1978  kmp_int32 default_device; /* internal control for default device */
1979  struct kmp_internal_control *next;
1980 } kmp_internal_control_t;
1981 
1982 static inline void copy_icvs(kmp_internal_control_t *dst,
1983  kmp_internal_control_t *src) {
1984  *dst = *src;
1985 }
1986 
1987 /* Thread barrier needs volatile barrier fields */
1988 typedef struct KMP_ALIGN_CACHE kmp_bstate {
1989  // th_fixed_icvs is aligned by virtue of kmp_bstate being aligned (and all
1990  // uses of it). It is not explicitly aligned below, because we *don't* want
1991  // it to be padded -- instead, we fit b_go into the same cache line with
1992  // th_fixed_icvs, enabling NGO cache lines stores in the hierarchical barrier.
1993  kmp_internal_control_t th_fixed_icvs; // Initial ICVs for the thread
1994  // Tuck b_go into end of th_fixed_icvs cache line, so it can be stored with
1995  // same NGO store
1996  volatile kmp_uint64 b_go; // STATE => task should proceed (hierarchical)
1997  KMP_ALIGN_CACHE volatile kmp_uint64
1998  b_arrived; // STATE => task reached synch point.
1999  kmp_uint32 *skip_per_level;
2000  kmp_uint32 my_level;
2001  kmp_int32 parent_tid;
2002  kmp_int32 old_tid;
2003  kmp_uint32 depth;
2004  struct kmp_bstate *parent_bar;
2005  kmp_team_t *team;
2006  kmp_uint64 leaf_state;
2007  kmp_uint32 nproc;
2008  kmp_uint8 base_leaf_kids;
2009  kmp_uint8 leaf_kids;
2010  kmp_uint8 offset;
2011  kmp_uint8 wait_flag;
2012  kmp_uint8 use_oncore_barrier;
2013 #if USE_DEBUGGER
2014  // The following field is intended for the debugger solely. Only the worker
2015  // thread itself accesses this field: the worker increases it by 1 when it
2016  // arrives to a barrier.
2017  KMP_ALIGN_CACHE kmp_uint b_worker_arrived;
2018 #endif /* USE_DEBUGGER */
2019 } kmp_bstate_t;
2020 
2021 union KMP_ALIGN_CACHE kmp_barrier_union {
2022  double b_align; /* use worst case alignment */
2023  char b_pad[KMP_PAD(kmp_bstate_t, CACHE_LINE)];
2024  kmp_bstate_t bb;
2025 };
2026 
2027 typedef union kmp_barrier_union kmp_balign_t;
2028 
2029 /* Team barrier needs only non-volatile arrived counter */
2030 union KMP_ALIGN_CACHE kmp_barrier_team_union {
2031  double b_align; /* use worst case alignment */
2032  char b_pad[CACHE_LINE];
2033  struct {
2034  kmp_uint64 b_arrived; /* STATE => task reached synch point. */
2035 #if USE_DEBUGGER
2036  // The following two fields are indended for the debugger solely. Only
2037  // primary thread of the team accesses these fields: the first one is
2038  // increased by 1 when the primary thread arrives to a barrier, the second
2039  // one is increased by one when all the threads arrived.
2040  kmp_uint b_master_arrived;
2041  kmp_uint b_team_arrived;
2042 #endif
2043  };
2044 };
2045 
2046 typedef union kmp_barrier_team_union kmp_balign_team_t;
2047 
2048 /* Padding for Linux* OS pthreads condition variables and mutexes used to signal
2049  threads when a condition changes. This is to workaround an NPTL bug where
2050  padding was added to pthread_cond_t which caused the initialization routine
2051  to write outside of the structure if compiled on pre-NPTL threads. */
2052 #if KMP_OS_WINDOWS
2053 typedef struct kmp_win32_mutex {
2054  /* The Lock */
2055  CRITICAL_SECTION cs;
2056 } kmp_win32_mutex_t;
2057 
2058 typedef struct kmp_win32_cond {
2059  /* Count of the number of waiters. */
2060  int waiters_count_;
2061 
2062  /* Serialize access to <waiters_count_> */
2063  kmp_win32_mutex_t waiters_count_lock_;
2064 
2065  /* Number of threads to release via a <cond_broadcast> or a <cond_signal> */
2066  int release_count_;
2067 
2068  /* Keeps track of the current "generation" so that we don't allow */
2069  /* one thread to steal all the "releases" from the broadcast. */
2070  int wait_generation_count_;
2071 
2072  /* A manual-reset event that's used to block and release waiting threads. */
2073  HANDLE event_;
2074 } kmp_win32_cond_t;
2075 #endif
2076 
2077 #if KMP_OS_UNIX
2078 
2079 union KMP_ALIGN_CACHE kmp_cond_union {
2080  double c_align;
2081  char c_pad[CACHE_LINE];
2082  pthread_cond_t c_cond;
2083 };
2084 
2085 typedef union kmp_cond_union kmp_cond_align_t;
2086 
2087 union KMP_ALIGN_CACHE kmp_mutex_union {
2088  double m_align;
2089  char m_pad[CACHE_LINE];
2090  pthread_mutex_t m_mutex;
2091 };
2092 
2093 typedef union kmp_mutex_union kmp_mutex_align_t;
2094 
2095 #endif /* KMP_OS_UNIX */
2096 
2097 typedef struct kmp_desc_base {
2098  void *ds_stackbase;
2099  size_t ds_stacksize;
2100  int ds_stackgrow;
2101  kmp_thread_t ds_thread;
2102  volatile int ds_tid;
2103  int ds_gtid;
2104 #if KMP_OS_WINDOWS
2105  volatile int ds_alive;
2106  DWORD ds_thread_id;
2107 /* ds_thread keeps thread handle on Windows* OS. It is enough for RTL purposes.
2108  However, debugger support (libomp_db) cannot work with handles, because they
2109  uncomparable. For example, debugger requests info about thread with handle h.
2110  h is valid within debugger process, and meaningless within debugee process.
2111  Even if h is duped by call to DuplicateHandle(), so the result h' is valid
2112  within debugee process, but it is a *new* handle which does *not* equal to
2113  any other handle in debugee... The only way to compare handles is convert
2114  them to system-wide ids. GetThreadId() function is available only in
2115  Longhorn and Server 2003. :-( In contrast, GetCurrentThreadId() is available
2116  on all Windows* OS flavours (including Windows* 95). Thus, we have to get
2117  thread id by call to GetCurrentThreadId() from within the thread and save it
2118  to let libomp_db identify threads. */
2119 #endif /* KMP_OS_WINDOWS */
2120 } kmp_desc_base_t;
2121 
2122 typedef union KMP_ALIGN_CACHE kmp_desc {
2123  double ds_align; /* use worst case alignment */
2124  char ds_pad[KMP_PAD(kmp_desc_base_t, CACHE_LINE)];
2125  kmp_desc_base_t ds;
2126 } kmp_desc_t;
2127 
2128 typedef struct kmp_local {
2129  volatile int this_construct; /* count of single's encountered by thread */
2130  void *reduce_data;
2131 #if KMP_USE_BGET
2132  void *bget_data;
2133  void *bget_list;
2134 #if !USE_CMP_XCHG_FOR_BGET
2135 #ifdef USE_QUEUING_LOCK_FOR_BGET
2136  kmp_lock_t bget_lock; /* Lock for accessing bget free list */
2137 #else
2138  kmp_bootstrap_lock_t bget_lock; // Lock for accessing bget free list. Must be
2139 // bootstrap lock so we can use it at library
2140 // shutdown.
2141 #endif /* USE_LOCK_FOR_BGET */
2142 #endif /* ! USE_CMP_XCHG_FOR_BGET */
2143 #endif /* KMP_USE_BGET */
2144 
2145  PACKED_REDUCTION_METHOD_T
2146  packed_reduction_method; /* stored by __kmpc_reduce*(), used by
2147  __kmpc_end_reduce*() */
2148 
2149 } kmp_local_t;
2150 
2151 #define KMP_CHECK_UPDATE(a, b) \
2152  if ((a) != (b)) \
2153  (a) = (b)
2154 #define KMP_CHECK_UPDATE_SYNC(a, b) \
2155  if ((a) != (b)) \
2156  TCW_SYNC_PTR((a), (b))
2157 
2158 #define get__blocktime(xteam, xtid) \
2159  ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.blocktime)
2160 #define get__bt_set(xteam, xtid) \
2161  ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_set)
2162 #if KMP_USE_MONITOR
2163 #define get__bt_intervals(xteam, xtid) \
2164  ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_intervals)
2165 #endif
2166 
2167 #define get__dynamic_2(xteam, xtid) \
2168  ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.dynamic)
2169 #define get__nproc_2(xteam, xtid) \
2170  ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.nproc)
2171 #define get__sched_2(xteam, xtid) \
2172  ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.sched)
2173 
2174 #define set__blocktime_team(xteam, xtid, xval) \
2175  (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.blocktime) = \
2176  (xval))
2177 
2178 #if KMP_USE_MONITOR
2179 #define set__bt_intervals_team(xteam, xtid, xval) \
2180  (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_intervals) = \
2181  (xval))
2182 #endif
2183 
2184 #define set__bt_set_team(xteam, xtid, xval) \
2185  (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_set) = (xval))
2186 
2187 #define set__dynamic(xthread, xval) \
2188  (((xthread)->th.th_current_task->td_icvs.dynamic) = (xval))
2189 #define get__dynamic(xthread) \
2190  (((xthread)->th.th_current_task->td_icvs.dynamic) ? (FTN_TRUE) : (FTN_FALSE))
2191 
2192 #define set__nproc(xthread, xval) \
2193  (((xthread)->th.th_current_task->td_icvs.nproc) = (xval))
2194 
2195 #define set__thread_limit(xthread, xval) \
2196  (((xthread)->th.th_current_task->td_icvs.thread_limit) = (xval))
2197 
2198 #define set__max_active_levels(xthread, xval) \
2199  (((xthread)->th.th_current_task->td_icvs.max_active_levels) = (xval))
2200 
2201 #define get__max_active_levels(xthread) \
2202  ((xthread)->th.th_current_task->td_icvs.max_active_levels)
2203 
2204 #define set__sched(xthread, xval) \
2205  (((xthread)->th.th_current_task->td_icvs.sched) = (xval))
2206 
2207 #define set__proc_bind(xthread, xval) \
2208  (((xthread)->th.th_current_task->td_icvs.proc_bind) = (xval))
2209 #define get__proc_bind(xthread) \
2210  ((xthread)->th.th_current_task->td_icvs.proc_bind)
2211 
2212 // OpenMP tasking data structures
2213 
2214 typedef enum kmp_tasking_mode {
2215  tskm_immediate_exec = 0,
2216  tskm_extra_barrier = 1,
2217  tskm_task_teams = 2,
2218  tskm_max = 2
2219 } kmp_tasking_mode_t;
2220 
2221 extern kmp_tasking_mode_t
2222  __kmp_tasking_mode; /* determines how/when to execute tasks */
2223 extern int __kmp_task_stealing_constraint;
2224 extern int __kmp_enable_task_throttling;
2225 extern kmp_int32 __kmp_default_device; // Set via OMP_DEFAULT_DEVICE if
2226 // specified, defaults to 0 otherwise
2227 // Set via OMP_MAX_TASK_PRIORITY if specified, defaults to 0 otherwise
2228 extern kmp_int32 __kmp_max_task_priority;
2229 // Set via KMP_TASKLOOP_MIN_TASKS if specified, defaults to 0 otherwise
2230 extern kmp_uint64 __kmp_taskloop_min_tasks;
2231 
2232 /* NOTE: kmp_taskdata_t and kmp_task_t structures allocated in single block with
2233  taskdata first */
2234 #define KMP_TASK_TO_TASKDATA(task) (((kmp_taskdata_t *)task) - 1)
2235 #define KMP_TASKDATA_TO_TASK(taskdata) (kmp_task_t *)(taskdata + 1)
2236 
2237 // The tt_found_tasks flag is a signal to all threads in the team that tasks
2238 // were spawned and queued since the previous barrier release.
2239 #define KMP_TASKING_ENABLED(task_team) \
2240  (TRUE == TCR_SYNC_4((task_team)->tt.tt_found_tasks))
2248 typedef kmp_int32 (*kmp_routine_entry_t)(kmp_int32, void *);
2249 
2250 typedef union kmp_cmplrdata {
2251  kmp_int32 priority;
2252  kmp_routine_entry_t
2253  destructors; /* pointer to function to invoke deconstructors of
2254  firstprivate C++ objects */
2255  /* future data */
2256 } kmp_cmplrdata_t;
2257 
2258 /* sizeof_kmp_task_t passed as arg to kmpc_omp_task call */
2261 typedef struct kmp_task { /* GEH: Shouldn't this be aligned somehow? */
2262  void *shareds;
2263  kmp_routine_entry_t
2264  routine;
2265  kmp_int32 part_id;
2266  kmp_cmplrdata_t
2267  data1; /* Two known optional additions: destructors and priority */
2268  kmp_cmplrdata_t data2; /* Process destructors first, priority second */
2269  /* future data */
2270  /* private vars */
2271 } kmp_task_t;
2272 
2277 typedef struct kmp_taskgroup {
2278  std::atomic<kmp_int32> count; // number of allocated and incomplete tasks
2279  std::atomic<kmp_int32>
2280  cancel_request; // request for cancellation of this taskgroup
2281  struct kmp_taskgroup *parent; // parent taskgroup
2282  // Block of data to perform task reduction
2283  void *reduce_data; // reduction related info
2284  kmp_int32 reduce_num_data; // number of data items to reduce
2285  uintptr_t *gomp_data; // gomp reduction data
2286 } kmp_taskgroup_t;
2287 
2288 // forward declarations
2289 typedef union kmp_depnode kmp_depnode_t;
2290 typedef struct kmp_depnode_list kmp_depnode_list_t;
2291 typedef struct kmp_dephash_entry kmp_dephash_entry_t;
2292 
2293 // macros for checking dep flag as an integer
2294 #define KMP_DEP_IN 0x1
2295 #define KMP_DEP_OUT 0x2
2296 #define KMP_DEP_INOUT 0x3
2297 #define KMP_DEP_MTX 0x4
2298 #define KMP_DEP_SET 0x8
2299 #define KMP_DEP_ALL 0x80
2300 // Compiler sends us this info:
2301 typedef struct kmp_depend_info {
2302  kmp_intptr_t base_addr;
2303  size_t len;
2304  union {
2305  kmp_uint8 flag; // flag as an unsigned char
2306  struct { // flag as a set of 8 bits
2307  unsigned in : 1;
2308  unsigned out : 1;
2309  unsigned mtx : 1;
2310  unsigned set : 1;
2311  unsigned unused : 3;
2312  unsigned all : 1;
2313  } flags;
2314  };
2315 } kmp_depend_info_t;
2316 
2317 // Internal structures to work with task dependencies:
2318 struct kmp_depnode_list {
2319  kmp_depnode_t *node;
2320  kmp_depnode_list_t *next;
2321 };
2322 
2323 // Max number of mutexinoutset dependencies per node
2324 #define MAX_MTX_DEPS 4
2325 
2326 typedef struct kmp_base_depnode {
2327  kmp_depnode_list_t *successors; /* used under lock */
2328  kmp_task_t *task; /* non-NULL if depnode is active, used under lock */
2329  kmp_lock_t *mtx_locks[MAX_MTX_DEPS]; /* lock mutexinoutset dependent tasks */
2330  kmp_int32 mtx_num_locks; /* number of locks in mtx_locks array */
2331  kmp_lock_t lock; /* guards shared fields: task, successors */
2332 #if KMP_SUPPORT_GRAPH_OUTPUT
2333  kmp_uint32 id;
2334 #endif
2335  std::atomic<kmp_int32> npredecessors;
2336  std::atomic<kmp_int32> nrefs;
2337 } kmp_base_depnode_t;
2338 
2339 union KMP_ALIGN_CACHE kmp_depnode {
2340  double dn_align; /* use worst case alignment */
2341  char dn_pad[KMP_PAD(kmp_base_depnode_t, CACHE_LINE)];
2342  kmp_base_depnode_t dn;
2343 };
2344 
2345 struct kmp_dephash_entry {
2346  kmp_intptr_t addr;
2347  kmp_depnode_t *last_out;
2348  kmp_depnode_list_t *last_set;
2349  kmp_depnode_list_t *prev_set;
2350  kmp_uint8 last_flag;
2351  kmp_lock_t *mtx_lock; /* is referenced by depnodes w/mutexinoutset dep */
2352  kmp_dephash_entry_t *next_in_bucket;
2353 };
2354 
2355 typedef struct kmp_dephash {
2356  kmp_dephash_entry_t **buckets;
2357  size_t size;
2358  kmp_depnode_t *last_all;
2359  size_t generation;
2360  kmp_uint32 nelements;
2361  kmp_uint32 nconflicts;
2362 } kmp_dephash_t;
2363 
2364 typedef struct kmp_task_affinity_info {
2365  kmp_intptr_t base_addr;
2366  size_t len;
2367  struct {
2368  bool flag1 : 1;
2369  bool flag2 : 1;
2370  kmp_int32 reserved : 30;
2371  } flags;
2372 } kmp_task_affinity_info_t;
2373 
2374 typedef enum kmp_event_type_t {
2375  KMP_EVENT_UNINITIALIZED = 0,
2376  KMP_EVENT_ALLOW_COMPLETION = 1
2377 } kmp_event_type_t;
2378 
2379 typedef struct {
2380  kmp_event_type_t type;
2381  kmp_tas_lock_t lock;
2382  union {
2383  kmp_task_t *task;
2384  } ed;
2385 } kmp_event_t;
2386 
2387 #ifdef BUILD_TIED_TASK_STACK
2388 
2389 /* Tied Task stack definitions */
2390 typedef struct kmp_stack_block {
2391  kmp_taskdata_t *sb_block[TASK_STACK_BLOCK_SIZE];
2392  struct kmp_stack_block *sb_next;
2393  struct kmp_stack_block *sb_prev;
2394 } kmp_stack_block_t;
2395 
2396 typedef struct kmp_task_stack {
2397  kmp_stack_block_t ts_first_block; // first block of stack entries
2398  kmp_taskdata_t **ts_top; // pointer to the top of stack
2399  kmp_int32 ts_entries; // number of entries on the stack
2400 } kmp_task_stack_t;
2401 
2402 #endif // BUILD_TIED_TASK_STACK
2403 
2404 typedef struct kmp_tasking_flags { /* Total struct must be exactly 32 bits */
2405  /* Compiler flags */ /* Total compiler flags must be 16 bits */
2406  unsigned tiedness : 1; /* task is either tied (1) or untied (0) */
2407  unsigned final : 1; /* task is final(1) so execute immediately */
2408  unsigned merged_if0 : 1; /* no __kmpc_task_{begin/complete}_if0 calls in if0
2409  code path */
2410  unsigned destructors_thunk : 1; /* set if the compiler creates a thunk to
2411  invoke destructors from the runtime */
2412  unsigned proxy : 1; /* task is a proxy task (it will be executed outside the
2413  context of the RTL) */
2414  unsigned priority_specified : 1; /* set if the compiler provides priority
2415  setting for the task */
2416  unsigned detachable : 1; /* 1 == can detach */
2417  unsigned hidden_helper : 1; /* 1 == hidden helper task */
2418  unsigned reserved : 8; /* reserved for compiler use */
2419 
2420  /* Library flags */ /* Total library flags must be 16 bits */
2421  unsigned tasktype : 1; /* task is either explicit(1) or implicit (0) */
2422  unsigned task_serial : 1; // task is executed immediately (1) or deferred (0)
2423  unsigned tasking_ser : 1; // all tasks in team are either executed immediately
2424  // (1) or may be deferred (0)
2425  unsigned team_serial : 1; // entire team is serial (1) [1 thread] or parallel
2426  // (0) [>= 2 threads]
2427  /* If either team_serial or tasking_ser is set, task team may be NULL */
2428  /* Task State Flags: */
2429  unsigned started : 1; /* 1==started, 0==not started */
2430  unsigned executing : 1; /* 1==executing, 0==not executing */
2431  unsigned complete : 1; /* 1==complete, 0==not complete */
2432  unsigned freed : 1; /* 1==freed, 0==allocated */
2433  unsigned native : 1; /* 1==gcc-compiled task, 0==intel */
2434  unsigned reserved31 : 7; /* reserved for library use */
2435 
2436 } kmp_tasking_flags_t;
2437 
2438 struct kmp_taskdata { /* aligned during dynamic allocation */
2439  kmp_int32 td_task_id; /* id, assigned by debugger */
2440  kmp_tasking_flags_t td_flags; /* task flags */
2441  kmp_team_t *td_team; /* team for this task */
2442  kmp_info_p *td_alloc_thread; /* thread that allocated data structures */
2443  /* Currently not used except for perhaps IDB */
2444  kmp_taskdata_t *td_parent; /* parent task */
2445  kmp_int32 td_level; /* task nesting level */
2446  std::atomic<kmp_int32> td_untied_count; // untied task active parts counter
2447  ident_t *td_ident; /* task identifier */
2448  // Taskwait data.
2449  ident_t *td_taskwait_ident;
2450  kmp_uint32 td_taskwait_counter;
2451  kmp_int32 td_taskwait_thread; /* gtid + 1 of thread encountered taskwait */
2452  KMP_ALIGN_CACHE kmp_internal_control_t
2453  td_icvs; /* Internal control variables for the task */
2454  KMP_ALIGN_CACHE std::atomic<kmp_int32>
2455  td_allocated_child_tasks; /* Child tasks (+ current task) not yet
2456  deallocated */
2457  std::atomic<kmp_int32>
2458  td_incomplete_child_tasks; /* Child tasks not yet complete */
2459  kmp_taskgroup_t
2460  *td_taskgroup; // Each task keeps pointer to its current taskgroup
2461  kmp_dephash_t
2462  *td_dephash; // Dependencies for children tasks are tracked from here
2463  kmp_depnode_t
2464  *td_depnode; // Pointer to graph node if this task has dependencies
2465  kmp_task_team_t *td_task_team;
2466  size_t td_size_alloc; // Size of task structure, including shareds etc.
2467 #if defined(KMP_GOMP_COMPAT)
2468  // 4 or 8 byte integers for the loop bounds in GOMP_taskloop
2469  kmp_int32 td_size_loop_bounds;
2470 #endif
2471  kmp_taskdata_t *td_last_tied; // keep tied task for task scheduling constraint
2472 #if defined(KMP_GOMP_COMPAT)
2473  // GOMP sends in a copy function for copy constructors
2474  void (*td_copy_func)(void *, void *);
2475 #endif
2476  kmp_event_t td_allow_completion_event;
2477 #if OMPT_SUPPORT
2478  ompt_task_info_t ompt_task_info;
2479 #endif
2480 }; // struct kmp_taskdata
2481 
2482 // Make sure padding above worked
2483 KMP_BUILD_ASSERT(sizeof(kmp_taskdata_t) % sizeof(void *) == 0);
2484 
2485 // Data for task team but per thread
2486 typedef struct kmp_base_thread_data {
2487  kmp_info_p *td_thr; // Pointer back to thread info
2488  // Used only in __kmp_execute_tasks_template, maybe not avail until task is
2489  // queued?
2490  kmp_bootstrap_lock_t td_deque_lock; // Lock for accessing deque
2491  kmp_taskdata_t *
2492  *td_deque; // Deque of tasks encountered by td_thr, dynamically allocated
2493  kmp_int32 td_deque_size; // Size of deck
2494  kmp_uint32 td_deque_head; // Head of deque (will wrap)
2495  kmp_uint32 td_deque_tail; // Tail of deque (will wrap)
2496  kmp_int32 td_deque_ntasks; // Number of tasks in deque
2497  // GEH: shouldn't this be volatile since used in while-spin?
2498  kmp_int32 td_deque_last_stolen; // Thread number of last successful steal
2499 #ifdef BUILD_TIED_TASK_STACK
2500  kmp_task_stack_t td_susp_tied_tasks; // Stack of suspended tied tasks for task
2501 // scheduling constraint
2502 #endif // BUILD_TIED_TASK_STACK
2503 } kmp_base_thread_data_t;
2504 
2505 #define TASK_DEQUE_BITS 8 // Used solely to define INITIAL_TASK_DEQUE_SIZE
2506 #define INITIAL_TASK_DEQUE_SIZE (1 << TASK_DEQUE_BITS)
2507 
2508 #define TASK_DEQUE_SIZE(td) ((td).td_deque_size)
2509 #define TASK_DEQUE_MASK(td) ((td).td_deque_size - 1)
2510 
2511 typedef union KMP_ALIGN_CACHE kmp_thread_data {
2512  kmp_base_thread_data_t td;
2513  double td_align; /* use worst case alignment */
2514  char td_pad[KMP_PAD(kmp_base_thread_data_t, CACHE_LINE)];
2515 } kmp_thread_data_t;
2516 
2517 // Data for task teams which are used when tasking is enabled for the team
2518 typedef struct kmp_base_task_team {
2519  kmp_bootstrap_lock_t
2520  tt_threads_lock; /* Lock used to allocate per-thread part of task team */
2521  /* must be bootstrap lock since used at library shutdown*/
2522  kmp_task_team_t *tt_next; /* For linking the task team free list */
2523  kmp_thread_data_t
2524  *tt_threads_data; /* Array of per-thread structures for task team */
2525  /* Data survives task team deallocation */
2526  kmp_int32 tt_found_tasks; /* Have we found tasks and queued them while
2527  executing this team? */
2528  /* TRUE means tt_threads_data is set up and initialized */
2529  kmp_int32 tt_nproc; /* #threads in team */
2530  kmp_int32 tt_max_threads; // # entries allocated for threads_data array
2531  kmp_int32 tt_found_proxy_tasks; // found proxy tasks since last barrier
2532  kmp_int32 tt_untied_task_encountered;
2533  // There is hidden helper thread encountered in this task team so that we must
2534  // wait when waiting on task team
2535  kmp_int32 tt_hidden_helper_task_encountered;
2536 
2537  KMP_ALIGN_CACHE
2538  std::atomic<kmp_int32> tt_unfinished_threads; /* #threads still active */
2539 
2540  KMP_ALIGN_CACHE
2541  volatile kmp_uint32
2542  tt_active; /* is the team still actively executing tasks */
2543 } kmp_base_task_team_t;
2544 
2545 union KMP_ALIGN_CACHE kmp_task_team {
2546  kmp_base_task_team_t tt;
2547  double tt_align; /* use worst case alignment */
2548  char tt_pad[KMP_PAD(kmp_base_task_team_t, CACHE_LINE)];
2549 };
2550 
2551 #if (USE_FAST_MEMORY == 3) || (USE_FAST_MEMORY == 5)
2552 // Free lists keep same-size free memory slots for fast memory allocation
2553 // routines
2554 typedef struct kmp_free_list {
2555  void *th_free_list_self; // Self-allocated tasks free list
2556  void *th_free_list_sync; // Self-allocated tasks stolen/returned by other
2557  // threads
2558  void *th_free_list_other; // Non-self free list (to be returned to owner's
2559  // sync list)
2560 } kmp_free_list_t;
2561 #endif
2562 #if KMP_NESTED_HOT_TEAMS
2563 // Hot teams array keeps hot teams and their sizes for given thread. Hot teams
2564 // are not put in teams pool, and they don't put threads in threads pool.
2565 typedef struct kmp_hot_team_ptr {
2566  kmp_team_p *hot_team; // pointer to hot_team of given nesting level
2567  kmp_int32 hot_team_nth; // number of threads allocated for the hot_team
2568 } kmp_hot_team_ptr_t;
2569 #endif
2570 typedef struct kmp_teams_size {
2571  kmp_int32 nteams; // number of teams in a league
2572  kmp_int32 nth; // number of threads in each team of the league
2573 } kmp_teams_size_t;
2574 
2575 // This struct stores a thread that acts as a "root" for a contention
2576 // group. Contention groups are rooted at kmp_root threads, but also at
2577 // each primary thread of each team created in the teams construct.
2578 // This struct therefore also stores a thread_limit associated with
2579 // that contention group, and a counter to track the number of threads
2580 // active in that contention group. Each thread has a list of these: CG
2581 // root threads have an entry in their list in which cg_root refers to
2582 // the thread itself, whereas other workers in the CG will have a
2583 // single entry where cg_root is same as the entry containing their CG
2584 // root. When a thread encounters a teams construct, it will add a new
2585 // entry to the front of its list, because it now roots a new CG.
2586 typedef struct kmp_cg_root {
2587  kmp_info_p *cg_root; // "root" thread for a contention group
2588  // The CG root's limit comes from OMP_THREAD_LIMIT for root threads, or
2589  // thread_limit clause for teams primary threads
2590  kmp_int32 cg_thread_limit;
2591  kmp_int32 cg_nthreads; // Count of active threads in CG rooted at cg_root
2592  struct kmp_cg_root *up; // pointer to higher level CG root in list
2593 } kmp_cg_root_t;
2594 
2595 // OpenMP thread data structures
2596 
2597 typedef struct KMP_ALIGN_CACHE kmp_base_info {
2598  /* Start with the readonly data which is cache aligned and padded. This is
2599  written before the thread starts working by the primary thread. Uber
2600  masters may update themselves later. Usage does not consider serialized
2601  regions. */
2602  kmp_desc_t th_info;
2603  kmp_team_p *th_team; /* team we belong to */
2604  kmp_root_p *th_root; /* pointer to root of task hierarchy */
2605  kmp_info_p *th_next_pool; /* next available thread in the pool */
2606  kmp_disp_t *th_dispatch; /* thread's dispatch data */
2607  int th_in_pool; /* in thread pool (32 bits for TCR/TCW) */
2608 
2609  /* The following are cached from the team info structure */
2610  /* TODO use these in more places as determined to be needed via profiling */
2611  int th_team_nproc; /* number of threads in a team */
2612  kmp_info_p *th_team_master; /* the team's primary thread */
2613  int th_team_serialized; /* team is serialized */
2614  microtask_t th_teams_microtask; /* save entry address for teams construct */
2615  int th_teams_level; /* save initial level of teams construct */
2616 /* it is 0 on device but may be any on host */
2617 
2618 /* The blocktime info is copied from the team struct to the thread struct */
2619 /* at the start of a barrier, and the values stored in the team are used */
2620 /* at points in the code where the team struct is no longer guaranteed */
2621 /* to exist (from the POV of worker threads). */
2622 #if KMP_USE_MONITOR
2623  int th_team_bt_intervals;
2624  int th_team_bt_set;
2625 #else
2626  kmp_uint64 th_team_bt_intervals;
2627 #endif
2628 
2629 #if KMP_AFFINITY_SUPPORTED
2630  kmp_affin_mask_t *th_affin_mask; /* thread's current affinity mask */
2631 #endif
2632  omp_allocator_handle_t th_def_allocator; /* default allocator */
2633  /* The data set by the primary thread at reinit, then R/W by the worker */
2634  KMP_ALIGN_CACHE int
2635  th_set_nproc; /* if > 0, then only use this request for the next fork */
2636 #if KMP_NESTED_HOT_TEAMS
2637  kmp_hot_team_ptr_t *th_hot_teams; /* array of hot teams */
2638 #endif
2639  kmp_proc_bind_t
2640  th_set_proc_bind; /* if != proc_bind_default, use request for next fork */
2641  kmp_teams_size_t
2642  th_teams_size; /* number of teams/threads in teams construct */
2643 #if KMP_AFFINITY_SUPPORTED
2644  int th_current_place; /* place currently bound to */
2645  int th_new_place; /* place to bind to in par reg */
2646  int th_first_place; /* first place in partition */
2647  int th_last_place; /* last place in partition */
2648 #endif
2649  int th_prev_level; /* previous level for affinity format */
2650  int th_prev_num_threads; /* previous num_threads for affinity format */
2651 #if USE_ITT_BUILD
2652  kmp_uint64 th_bar_arrive_time; /* arrival to barrier timestamp */
2653  kmp_uint64 th_bar_min_time; /* minimum arrival time at the barrier */
2654  kmp_uint64 th_frame_time; /* frame timestamp */
2655 #endif /* USE_ITT_BUILD */
2656  kmp_local_t th_local;
2657  struct private_common *th_pri_head;
2658 
2659  /* Now the data only used by the worker (after initial allocation) */
2660  /* TODO the first serial team should actually be stored in the info_t
2661  structure. this will help reduce initial allocation overhead */
2662  KMP_ALIGN_CACHE kmp_team_p
2663  *th_serial_team; /*serialized team held in reserve*/
2664 
2665 #if OMPT_SUPPORT
2666  ompt_thread_info_t ompt_thread_info;
2667 #endif
2668 
2669  /* The following are also read by the primary thread during reinit */
2670  struct common_table *th_pri_common;
2671 
2672  volatile kmp_uint32 th_spin_here; /* thread-local location for spinning */
2673  /* while awaiting queuing lock acquire */
2674 
2675  volatile void *th_sleep_loc; // this points at a kmp_flag<T>
2676  flag_type th_sleep_loc_type; // enum type of flag stored in th_sleep_loc
2677 
2678  ident_t *th_ident;
2679  unsigned th_x; // Random number generator data
2680  unsigned th_a; // Random number generator data
2681 
2682  /* Tasking-related data for the thread */
2683  kmp_task_team_t *th_task_team; // Task team struct
2684  kmp_taskdata_t *th_current_task; // Innermost Task being executed
2685  kmp_uint8 th_task_state; // alternating 0/1 for task team identification
2686  kmp_uint8 *th_task_state_memo_stack; // Stack holding memos of th_task_state
2687  // at nested levels
2688  kmp_uint32 th_task_state_top; // Top element of th_task_state_memo_stack
2689  kmp_uint32 th_task_state_stack_sz; // Size of th_task_state_memo_stack
2690  kmp_uint32 th_reap_state; // Non-zero indicates thread is not
2691  // tasking, thus safe to reap
2692 
2693  /* More stuff for keeping track of active/sleeping threads (this part is
2694  written by the worker thread) */
2695  kmp_uint8 th_active_in_pool; // included in count of #active threads in pool
2696  int th_active; // ! sleeping; 32 bits for TCR/TCW
2697  std::atomic<kmp_uint32> th_used_in_team; // Flag indicating use in team
2698  // 0 = not used in team; 1 = used in team;
2699  // 2 = transitioning to not used in team; 3 = transitioning to used in team
2700  struct cons_header *th_cons; // used for consistency check
2701 #if KMP_USE_HIER_SCHED
2702  // used for hierarchical scheduling
2703  kmp_hier_private_bdata_t *th_hier_bar_data;
2704 #endif
2705 
2706  /* Add the syncronizing data which is cache aligned and padded. */
2707  KMP_ALIGN_CACHE kmp_balign_t th_bar[bs_last_barrier];
2708 
2709  KMP_ALIGN_CACHE volatile kmp_int32
2710  th_next_waiting; /* gtid+1 of next thread on lock wait queue, 0 if none */
2711 
2712 #if (USE_FAST_MEMORY == 3) || (USE_FAST_MEMORY == 5)
2713 #define NUM_LISTS 4
2714  kmp_free_list_t th_free_lists[NUM_LISTS]; // Free lists for fast memory
2715 // allocation routines
2716 #endif
2717 
2718 #if KMP_OS_WINDOWS
2719  kmp_win32_cond_t th_suspend_cv;
2720  kmp_win32_mutex_t th_suspend_mx;
2721  std::atomic<int> th_suspend_init;
2722 #endif
2723 #if KMP_OS_UNIX
2724  kmp_cond_align_t th_suspend_cv;
2725  kmp_mutex_align_t th_suspend_mx;
2726  std::atomic<int> th_suspend_init_count;
2727 #endif
2728 
2729 #if USE_ITT_BUILD
2730  kmp_itt_mark_t th_itt_mark_single;
2731 // alignment ???
2732 #endif /* USE_ITT_BUILD */
2733 #if KMP_STATS_ENABLED
2734  kmp_stats_list *th_stats;
2735 #endif
2736 #if KMP_OS_UNIX
2737  std::atomic<bool> th_blocking;
2738 #endif
2739  kmp_cg_root_t *th_cg_roots; // list of cg_roots associated with this thread
2740 } kmp_base_info_t;
2741 
2742 typedef union KMP_ALIGN_CACHE kmp_info {
2743  double th_align; /* use worst case alignment */
2744  char th_pad[KMP_PAD(kmp_base_info_t, CACHE_LINE)];
2745  kmp_base_info_t th;
2746 } kmp_info_t;
2747 
2748 // OpenMP thread team data structures
2749 
2750 typedef struct kmp_base_data {
2751  volatile kmp_uint32 t_value;
2752 } kmp_base_data_t;
2753 
2754 typedef union KMP_ALIGN_CACHE kmp_sleep_team {
2755  double dt_align; /* use worst case alignment */
2756  char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
2757  kmp_base_data_t dt;
2758 } kmp_sleep_team_t;
2759 
2760 typedef union KMP_ALIGN_CACHE kmp_ordered_team {
2761  double dt_align; /* use worst case alignment */
2762  char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
2763  kmp_base_data_t dt;
2764 } kmp_ordered_team_t;
2765 
2766 typedef int (*launch_t)(int gtid);
2767 
2768 /* Minimum number of ARGV entries to malloc if necessary */
2769 #define KMP_MIN_MALLOC_ARGV_ENTRIES 100
2770 
2771 // Set up how many argv pointers will fit in cache lines containing
2772 // t_inline_argv. Historically, we have supported at least 96 bytes. Using a
2773 // larger value for more space between the primary write/worker read section and
2774 // read/write by all section seems to buy more performance on EPCC PARALLEL.
2775 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
2776 #define KMP_INLINE_ARGV_BYTES \
2777  (4 * CACHE_LINE - \
2778  ((3 * KMP_PTR_SKIP + 2 * sizeof(int) + 2 * sizeof(kmp_int8) + \
2779  sizeof(kmp_int16) + sizeof(kmp_uint32)) % \
2780  CACHE_LINE))
2781 #else
2782 #define KMP_INLINE_ARGV_BYTES \
2783  (2 * CACHE_LINE - ((3 * KMP_PTR_SKIP + 2 * sizeof(int)) % CACHE_LINE))
2784 #endif
2785 #define KMP_INLINE_ARGV_ENTRIES (int)(KMP_INLINE_ARGV_BYTES / KMP_PTR_SKIP)
2786 
2787 typedef struct KMP_ALIGN_CACHE kmp_base_team {
2788  // Synchronization Data
2789  // ---------------------------------------------------------------------------
2790  KMP_ALIGN_CACHE kmp_ordered_team_t t_ordered;
2791  kmp_balign_team_t t_bar[bs_last_barrier];
2792  std::atomic<int> t_construct; // count of single directive encountered by team
2793  char pad[sizeof(kmp_lock_t)]; // padding to maintain performance on big iron
2794 
2795  // [0] - parallel / [1] - worksharing task reduction data shared by taskgroups
2796  std::atomic<void *> t_tg_reduce_data[2]; // to support task modifier
2797  std::atomic<int> t_tg_fini_counter[2]; // sync end of task reductions
2798 
2799  // Primary thread only
2800  // ---------------------------------------------------------------------------
2801  KMP_ALIGN_CACHE int t_master_tid; // tid of primary thread in parent team
2802  int t_master_this_cons; // "this_construct" single counter of primary thread
2803  // in parent team
2804  ident_t *t_ident; // if volatile, have to change too much other crud to
2805  // volatile too
2806  kmp_team_p *t_parent; // parent team
2807  kmp_team_p *t_next_pool; // next free team in the team pool
2808  kmp_disp_t *t_dispatch; // thread's dispatch data
2809  kmp_task_team_t *t_task_team[2]; // Task team struct; switch between 2
2810  kmp_proc_bind_t t_proc_bind; // bind type for par region
2811 #if USE_ITT_BUILD
2812  kmp_uint64 t_region_time; // region begin timestamp
2813 #endif /* USE_ITT_BUILD */
2814 
2815  // Primary thread write, workers read
2816  // --------------------------------------------------------------------------
2817  KMP_ALIGN_CACHE void **t_argv;
2818  int t_argc;
2819  int t_nproc; // number of threads in team
2820  microtask_t t_pkfn;
2821  launch_t t_invoke; // procedure to launch the microtask
2822 
2823 #if OMPT_SUPPORT
2824  ompt_team_info_t ompt_team_info;
2825  ompt_lw_taskteam_t *ompt_serialized_team_info;
2826 #endif
2827 
2828 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
2829  kmp_int8 t_fp_control_saved;
2830  kmp_int8 t_pad2b;
2831  kmp_int16 t_x87_fpu_control_word; // FP control regs
2832  kmp_uint32 t_mxcsr;
2833 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
2834 
2835  void *t_inline_argv[KMP_INLINE_ARGV_ENTRIES];
2836 
2837  KMP_ALIGN_CACHE kmp_info_t **t_threads;
2838  kmp_taskdata_t
2839  *t_implicit_task_taskdata; // Taskdata for the thread's implicit task
2840  int t_level; // nested parallel level
2841 
2842  KMP_ALIGN_CACHE int t_max_argc;
2843  int t_max_nproc; // max threads this team can handle (dynamically expandable)
2844  int t_serialized; // levels deep of serialized teams
2845  dispatch_shared_info_t *t_disp_buffer; // buffers for dispatch system
2846  int t_id; // team's id, assigned by debugger.
2847  int t_active_level; // nested active parallel level
2848  kmp_r_sched_t t_sched; // run-time schedule for the team
2849 #if KMP_AFFINITY_SUPPORTED
2850  int t_first_place; // first & last place in parent thread's partition.
2851  int t_last_place; // Restore these values to primary thread after par region.
2852 #endif // KMP_AFFINITY_SUPPORTED
2853  int t_display_affinity;
2854  int t_size_changed; // team size was changed?: 0: no, 1: yes, -1: changed via
2855  // omp_set_num_threads() call
2856  omp_allocator_handle_t t_def_allocator; /* default allocator */
2857 
2858 // Read/write by workers as well
2859 #if (KMP_ARCH_X86 || KMP_ARCH_X86_64)
2860  // Using CACHE_LINE=64 reduces memory footprint, but causes a big perf
2861  // regression of epcc 'parallel' and 'barrier' on fxe256lin01. This extra
2862  // padding serves to fix the performance of epcc 'parallel' and 'barrier' when
2863  // CACHE_LINE=64. TODO: investigate more and get rid if this padding.
2864  char dummy_padding[1024];
2865 #endif
2866  // Internal control stack for additional nested teams.
2867  KMP_ALIGN_CACHE kmp_internal_control_t *t_control_stack_top;
2868  // for SERIALIZED teams nested 2 or more levels deep
2869  // typed flag to store request state of cancellation
2870  std::atomic<kmp_int32> t_cancel_request;
2871  int t_master_active; // save on fork, restore on join
2872  void *t_copypriv_data; // team specific pointer to copyprivate data array
2873 #if KMP_OS_WINDOWS
2874  std::atomic<kmp_uint32> t_copyin_counter;
2875 #endif
2876 #if USE_ITT_BUILD
2877  void *t_stack_id; // team specific stack stitching id (for ittnotify)
2878 #endif /* USE_ITT_BUILD */
2879  distributedBarrier *b; // Distributed barrier data associated with team
2880 } kmp_base_team_t;
2881 
2882 union KMP_ALIGN_CACHE kmp_team {
2883  kmp_base_team_t t;
2884  double t_align; /* use worst case alignment */
2885  char t_pad[KMP_PAD(kmp_base_team_t, CACHE_LINE)];
2886 };
2887 
2888 typedef union KMP_ALIGN_CACHE kmp_time_global {
2889  double dt_align; /* use worst case alignment */
2890  char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
2891  kmp_base_data_t dt;
2892 } kmp_time_global_t;
2893 
2894 typedef struct kmp_base_global {
2895  /* cache-aligned */
2896  kmp_time_global_t g_time;
2897 
2898  /* non cache-aligned */
2899  volatile int g_abort;
2900  volatile int g_done;
2901 
2902  int g_dynamic;
2903  enum dynamic_mode g_dynamic_mode;
2904 } kmp_base_global_t;
2905 
2906 typedef union KMP_ALIGN_CACHE kmp_global {
2907  kmp_base_global_t g;
2908  double g_align; /* use worst case alignment */
2909  char g_pad[KMP_PAD(kmp_base_global_t, CACHE_LINE)];
2910 } kmp_global_t;
2911 
2912 typedef struct kmp_base_root {
2913  // TODO: GEH - combine r_active with r_in_parallel then r_active ==
2914  // (r_in_parallel>= 0)
2915  // TODO: GEH - then replace r_active with t_active_levels if we can to reduce
2916  // the synch overhead or keeping r_active
2917  volatile int r_active; /* TRUE if some region in a nest has > 1 thread */
2918  // keeps a count of active parallel regions per root
2919  std::atomic<int> r_in_parallel;
2920  // GEH: This is misnamed, should be r_active_levels
2921  kmp_team_t *r_root_team;
2922  kmp_team_t *r_hot_team;
2923  kmp_info_t *r_uber_thread;
2924  kmp_lock_t r_begin_lock;
2925  volatile int r_begin;
2926  int r_blocktime; /* blocktime for this root and descendants */
2927 #if KMP_AFFINITY_SUPPORTED
2928  int r_affinity_assigned;
2929 #endif // KMP_AFFINITY_SUPPORTED
2930 } kmp_base_root_t;
2931 
2932 typedef union KMP_ALIGN_CACHE kmp_root {
2933  kmp_base_root_t r;
2934  double r_align; /* use worst case alignment */
2935  char r_pad[KMP_PAD(kmp_base_root_t, CACHE_LINE)];
2936 } kmp_root_t;
2937 
2938 struct fortran_inx_info {
2939  kmp_int32 data;
2940 };
2941 
2942 /* ------------------------------------------------------------------------ */
2943 
2944 extern int __kmp_settings;
2945 extern int __kmp_duplicate_library_ok;
2946 #if USE_ITT_BUILD
2947 extern int __kmp_forkjoin_frames;
2948 extern int __kmp_forkjoin_frames_mode;
2949 #endif
2950 extern PACKED_REDUCTION_METHOD_T __kmp_force_reduction_method;
2951 extern int __kmp_determ_red;
2952 
2953 #ifdef KMP_DEBUG
2954 extern int kmp_a_debug;
2955 extern int kmp_b_debug;
2956 extern int kmp_c_debug;
2957 extern int kmp_d_debug;
2958 extern int kmp_e_debug;
2959 extern int kmp_f_debug;
2960 #endif /* KMP_DEBUG */
2961 
2962 /* For debug information logging using rotating buffer */
2963 #define KMP_DEBUG_BUF_LINES_INIT 512
2964 #define KMP_DEBUG_BUF_LINES_MIN 1
2965 
2966 #define KMP_DEBUG_BUF_CHARS_INIT 128
2967 #define KMP_DEBUG_BUF_CHARS_MIN 2
2968 
2969 extern int
2970  __kmp_debug_buf; /* TRUE means use buffer, FALSE means print to stderr */
2971 extern int __kmp_debug_buf_lines; /* How many lines of debug stored in buffer */
2972 extern int
2973  __kmp_debug_buf_chars; /* How many characters allowed per line in buffer */
2974 extern int __kmp_debug_buf_atomic; /* TRUE means use atomic update of buffer
2975  entry pointer */
2976 
2977 extern char *__kmp_debug_buffer; /* Debug buffer itself */
2978 extern std::atomic<int> __kmp_debug_count; /* Counter for number of lines
2979  printed in buffer so far */
2980 extern int __kmp_debug_buf_warn_chars; /* Keep track of char increase
2981  recommended in warnings */
2982 /* end rotating debug buffer */
2983 
2984 #ifdef KMP_DEBUG
2985 extern int __kmp_par_range; /* +1 => only go par for constructs in range */
2986 
2987 #define KMP_PAR_RANGE_ROUTINE_LEN 1024
2988 extern char __kmp_par_range_routine[KMP_PAR_RANGE_ROUTINE_LEN];
2989 #define KMP_PAR_RANGE_FILENAME_LEN 1024
2990 extern char __kmp_par_range_filename[KMP_PAR_RANGE_FILENAME_LEN];
2991 extern int __kmp_par_range_lb;
2992 extern int __kmp_par_range_ub;
2993 #endif
2994 
2995 /* For printing out dynamic storage map for threads and teams */
2996 extern int
2997  __kmp_storage_map; /* True means print storage map for threads and teams */
2998 extern int __kmp_storage_map_verbose; /* True means storage map includes
2999  placement info */
3000 extern int __kmp_storage_map_verbose_specified;
3001 
3002 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
3003 extern kmp_cpuinfo_t __kmp_cpuinfo;
3004 static inline bool __kmp_is_hybrid_cpu() { return __kmp_cpuinfo.flags.hybrid; }
3005 #else
3006 static inline bool __kmp_is_hybrid_cpu() { return false; }
3007 #endif
3008 
3009 extern volatile int __kmp_init_serial;
3010 extern volatile int __kmp_init_gtid;
3011 extern volatile int __kmp_init_common;
3012 extern volatile int __kmp_init_middle;
3013 extern volatile int __kmp_init_parallel;
3014 #if KMP_USE_MONITOR
3015 extern volatile int __kmp_init_monitor;
3016 #endif
3017 extern volatile int __kmp_init_user_locks;
3018 extern volatile int __kmp_init_hidden_helper_threads;
3019 extern int __kmp_init_counter;
3020 extern int __kmp_root_counter;
3021 extern int __kmp_version;
3022 
3023 /* list of address of allocated caches for commons */
3024 extern kmp_cached_addr_t *__kmp_threadpriv_cache_list;
3025 
3026 /* Barrier algorithm types and options */
3027 extern kmp_uint32 __kmp_barrier_gather_bb_dflt;
3028 extern kmp_uint32 __kmp_barrier_release_bb_dflt;
3029 extern kmp_bar_pat_e __kmp_barrier_gather_pat_dflt;
3030 extern kmp_bar_pat_e __kmp_barrier_release_pat_dflt;
3031 extern kmp_uint32 __kmp_barrier_gather_branch_bits[bs_last_barrier];
3032 extern kmp_uint32 __kmp_barrier_release_branch_bits[bs_last_barrier];
3033 extern kmp_bar_pat_e __kmp_barrier_gather_pattern[bs_last_barrier];
3034 extern kmp_bar_pat_e __kmp_barrier_release_pattern[bs_last_barrier];
3035 extern char const *__kmp_barrier_branch_bit_env_name[bs_last_barrier];
3036 extern char const *__kmp_barrier_pattern_env_name[bs_last_barrier];
3037 extern char const *__kmp_barrier_type_name[bs_last_barrier];
3038 extern char const *__kmp_barrier_pattern_name[bp_last_bar];
3039 
3040 /* Global Locks */
3041 extern kmp_bootstrap_lock_t __kmp_initz_lock; /* control initialization */
3042 extern kmp_bootstrap_lock_t __kmp_forkjoin_lock; /* control fork/join access */
3043 extern kmp_bootstrap_lock_t __kmp_task_team_lock;
3044 extern kmp_bootstrap_lock_t
3045  __kmp_exit_lock; /* exit() is not always thread-safe */
3046 #if KMP_USE_MONITOR
3047 extern kmp_bootstrap_lock_t
3048  __kmp_monitor_lock; /* control monitor thread creation */
3049 #endif
3050 extern kmp_bootstrap_lock_t
3051  __kmp_tp_cached_lock; /* used for the hack to allow threadprivate cache and
3052  __kmp_threads expansion to co-exist */
3053 
3054 extern kmp_lock_t __kmp_global_lock; /* control OS/global access */
3055 extern kmp_queuing_lock_t __kmp_dispatch_lock; /* control dispatch access */
3056 extern kmp_lock_t __kmp_debug_lock; /* control I/O access for KMP_DEBUG */
3057 
3058 extern enum library_type __kmp_library;
3059 
3060 extern enum sched_type __kmp_sched; /* default runtime scheduling */
3061 extern enum sched_type __kmp_static; /* default static scheduling method */
3062 extern enum sched_type __kmp_guided; /* default guided scheduling method */
3063 extern enum sched_type __kmp_auto; /* default auto scheduling method */
3064 extern int __kmp_chunk; /* default runtime chunk size */
3065 extern int __kmp_force_monotonic; /* whether monotonic scheduling forced */
3066 
3067 extern size_t __kmp_stksize; /* stack size per thread */
3068 #if KMP_USE_MONITOR
3069 extern size_t __kmp_monitor_stksize; /* stack size for monitor thread */
3070 #endif
3071 extern size_t __kmp_stkoffset; /* stack offset per thread */
3072 extern int __kmp_stkpadding; /* Should we pad root thread(s) stack */
3073 
3074 extern size_t
3075  __kmp_malloc_pool_incr; /* incremental size of pool for kmp_malloc() */
3076 extern int __kmp_env_stksize; /* was KMP_STACKSIZE specified? */
3077 extern int __kmp_env_blocktime; /* was KMP_BLOCKTIME specified? */
3078 extern int __kmp_env_checks; /* was KMP_CHECKS specified? */
3079 extern int __kmp_env_consistency_check; // was KMP_CONSISTENCY_CHECK specified?
3080 extern int __kmp_generate_warnings; /* should we issue warnings? */
3081 extern int __kmp_reserve_warn; /* have we issued reserve_threads warning? */
3082 
3083 #ifdef DEBUG_SUSPEND
3084 extern int __kmp_suspend_count; /* count inside __kmp_suspend_template() */
3085 #endif
3086 
3087 extern kmp_int32 __kmp_use_yield;
3088 extern kmp_int32 __kmp_use_yield_exp_set;
3089 extern kmp_uint32 __kmp_yield_init;
3090 extern kmp_uint32 __kmp_yield_next;
3091 
3092 /* ------------------------------------------------------------------------- */
3093 extern int __kmp_allThreadsSpecified;
3094 
3095 extern size_t __kmp_align_alloc;
3096 /* following data protected by initialization routines */
3097 extern int __kmp_xproc; /* number of processors in the system */
3098 extern int __kmp_avail_proc; /* number of processors available to the process */
3099 extern size_t __kmp_sys_min_stksize; /* system-defined minimum stack size */
3100 extern int __kmp_sys_max_nth; /* system-imposed maximum number of threads */
3101 // maximum total number of concurrently-existing threads on device
3102 extern int __kmp_max_nth;
3103 // maximum total number of concurrently-existing threads in a contention group
3104 extern int __kmp_cg_max_nth;
3105 extern int __kmp_teams_max_nth; // max threads used in a teams construct
3106 extern int __kmp_threads_capacity; /* capacity of the arrays __kmp_threads and
3107  __kmp_root */
3108 extern int __kmp_dflt_team_nth; /* default number of threads in a parallel
3109  region a la OMP_NUM_THREADS */
3110 extern int __kmp_dflt_team_nth_ub; /* upper bound on "" determined at serial
3111  initialization */
3112 extern int __kmp_tp_capacity; /* capacity of __kmp_threads if threadprivate is
3113  used (fixed) */
3114 extern int __kmp_tp_cached; /* whether threadprivate cache has been created
3115  (__kmpc_threadprivate_cached()) */
3116 extern int __kmp_dflt_blocktime; /* number of milliseconds to wait before
3117  blocking (env setting) */
3118 #if KMP_USE_MONITOR
3119 extern int
3120  __kmp_monitor_wakeups; /* number of times monitor wakes up per second */
3121 extern int __kmp_bt_intervals; /* number of monitor timestamp intervals before
3122  blocking */
3123 #endif
3124 #ifdef KMP_ADJUST_BLOCKTIME
3125 extern int __kmp_zero_bt; /* whether blocktime has been forced to zero */
3126 #endif /* KMP_ADJUST_BLOCKTIME */
3127 #ifdef KMP_DFLT_NTH_CORES
3128 extern int __kmp_ncores; /* Total number of cores for threads placement */
3129 #endif
3130 /* Number of millisecs to delay on abort for Intel(R) VTune(TM) tools */
3131 extern int __kmp_abort_delay;
3132 
3133 extern int __kmp_need_register_atfork_specified;
3134 extern int __kmp_need_register_atfork; /* At initialization, call pthread_atfork
3135  to install fork handler */
3136 extern int __kmp_gtid_mode; /* Method of getting gtid, values:
3137  0 - not set, will be set at runtime
3138  1 - using stack search
3139  2 - dynamic TLS (pthread_getspecific(Linux* OS/OS
3140  X*) or TlsGetValue(Windows* OS))
3141  3 - static TLS (__declspec(thread) __kmp_gtid),
3142  Linux* OS .so only. */
3143 extern int
3144  __kmp_adjust_gtid_mode; /* If true, adjust method based on #threads */
3145 #ifdef KMP_TDATA_GTID
3146 extern KMP_THREAD_LOCAL int __kmp_gtid;
3147 #endif
3148 extern int __kmp_tls_gtid_min; /* #threads below which use sp search for gtid */
3149 extern int __kmp_foreign_tp; // If true, separate TP var for each foreign thread
3150 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
3151 extern int __kmp_inherit_fp_control; // copy fp creg(s) parent->workers at fork
3152 extern kmp_int16 __kmp_init_x87_fpu_control_word; // init thread's FP ctrl reg
3153 extern kmp_uint32 __kmp_init_mxcsr; /* init thread's mxscr */
3154 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
3155 
3156 // max_active_levels for nested parallelism enabled by default via
3157 // OMP_MAX_ACTIVE_LEVELS, OMP_NESTED, OMP_NUM_THREADS, and OMP_PROC_BIND
3158 extern int __kmp_dflt_max_active_levels;
3159 // Indicates whether value of __kmp_dflt_max_active_levels was already
3160 // explicitly set by OMP_MAX_ACTIVE_LEVELS or OMP_NESTED=false
3161 extern bool __kmp_dflt_max_active_levels_set;
3162 extern int __kmp_dispatch_num_buffers; /* max possible dynamic loops in
3163  concurrent execution per team */
3164 #if KMP_NESTED_HOT_TEAMS
3165 extern int __kmp_hot_teams_mode;
3166 extern int __kmp_hot_teams_max_level;
3167 #endif
3168 
3169 #if KMP_OS_LINUX
3170 extern enum clock_function_type __kmp_clock_function;
3171 extern int __kmp_clock_function_param;
3172 #endif /* KMP_OS_LINUX */
3173 
3174 #if KMP_MIC_SUPPORTED
3175 extern enum mic_type __kmp_mic_type;
3176 #endif
3177 
3178 #ifdef USE_LOAD_BALANCE
3179 extern double __kmp_load_balance_interval; // load balance algorithm interval
3180 #endif /* USE_LOAD_BALANCE */
3181 
3182 // OpenMP 3.1 - Nested num threads array
3183 typedef struct kmp_nested_nthreads_t {
3184  int *nth;
3185  int size;
3186  int used;
3187 } kmp_nested_nthreads_t;
3188 
3189 extern kmp_nested_nthreads_t __kmp_nested_nth;
3190 
3191 #if KMP_USE_ADAPTIVE_LOCKS
3192 
3193 // Parameters for the speculative lock backoff system.
3194 struct kmp_adaptive_backoff_params_t {
3195  // Number of soft retries before it counts as a hard retry.
3196  kmp_uint32 max_soft_retries;
3197  // Badness is a bit mask : 0,1,3,7,15,... on each hard failure we move one to
3198  // the right
3199  kmp_uint32 max_badness;
3200 };
3201 
3202 extern kmp_adaptive_backoff_params_t __kmp_adaptive_backoff_params;
3203 
3204 #if KMP_DEBUG_ADAPTIVE_LOCKS
3205 extern const char *__kmp_speculative_statsfile;
3206 #endif
3207 
3208 #endif // KMP_USE_ADAPTIVE_LOCKS
3209 
3210 extern int __kmp_display_env; /* TRUE or FALSE */
3211 extern int __kmp_display_env_verbose; /* TRUE if OMP_DISPLAY_ENV=VERBOSE */
3212 extern int __kmp_omp_cancellation; /* TRUE or FALSE */
3213 extern int __kmp_nteams;
3214 extern int __kmp_teams_thread_limit;
3215 
3216 /* ------------------------------------------------------------------------- */
3217 
3218 /* the following are protected by the fork/join lock */
3219 /* write: lock read: anytime */
3220 extern kmp_info_t **__kmp_threads; /* Descriptors for the threads */
3221 /* read/write: lock */
3222 extern volatile kmp_team_t *__kmp_team_pool;
3223 extern volatile kmp_info_t *__kmp_thread_pool;
3224 extern kmp_info_t *__kmp_thread_pool_insert_pt;
3225 
3226 // total num threads reachable from some root thread including all root threads
3227 extern volatile int __kmp_nth;
3228 /* total number of threads reachable from some root thread including all root
3229  threads, and those in the thread pool */
3230 extern volatile int __kmp_all_nth;
3231 extern std::atomic<int> __kmp_thread_pool_active_nth;
3232 
3233 extern kmp_root_t **__kmp_root; /* root of thread hierarchy */
3234 /* end data protected by fork/join lock */
3235 /* ------------------------------------------------------------------------- */
3236 
3237 #define __kmp_get_gtid() __kmp_get_global_thread_id()
3238 #define __kmp_entry_gtid() __kmp_get_global_thread_id_reg()
3239 #define __kmp_get_tid() (__kmp_tid_from_gtid(__kmp_get_gtid()))
3240 #define __kmp_get_team() (__kmp_threads[(__kmp_get_gtid())]->th.th_team)
3241 #define __kmp_get_thread() (__kmp_thread_from_gtid(__kmp_get_gtid()))
3242 
3243 // AT: Which way is correct?
3244 // AT: 1. nproc = __kmp_threads[ ( gtid ) ] -> th.th_team -> t.t_nproc;
3245 // AT: 2. nproc = __kmp_threads[ ( gtid ) ] -> th.th_team_nproc;
3246 #define __kmp_get_team_num_threads(gtid) \
3247  (__kmp_threads[(gtid)]->th.th_team->t.t_nproc)
3248 
3249 static inline bool KMP_UBER_GTID(int gtid) {
3250  KMP_DEBUG_ASSERT(gtid >= KMP_GTID_MIN);
3251  KMP_DEBUG_ASSERT(gtid < __kmp_threads_capacity);
3252  return (gtid >= 0 && __kmp_root[gtid] && __kmp_threads[gtid] &&
3253  __kmp_threads[gtid] == __kmp_root[gtid]->r.r_uber_thread);
3254 }
3255 
3256 static inline int __kmp_tid_from_gtid(int gtid) {
3257  KMP_DEBUG_ASSERT(gtid >= 0);
3258  return __kmp_threads[gtid]->th.th_info.ds.ds_tid;
3259 }
3260 
3261 static inline int __kmp_gtid_from_tid(int tid, const kmp_team_t *team) {
3262  KMP_DEBUG_ASSERT(tid >= 0 && team);
3263  return team->t.t_threads[tid]->th.th_info.ds.ds_gtid;
3264 }
3265 
3266 static inline int __kmp_gtid_from_thread(const kmp_info_t *thr) {
3267  KMP_DEBUG_ASSERT(thr);
3268  return thr->th.th_info.ds.ds_gtid;
3269 }
3270 
3271 static inline kmp_info_t *__kmp_thread_from_gtid(int gtid) {
3272  KMP_DEBUG_ASSERT(gtid >= 0);
3273  return __kmp_threads[gtid];
3274 }
3275 
3276 static inline kmp_team_t *__kmp_team_from_gtid(int gtid) {
3277  KMP_DEBUG_ASSERT(gtid >= 0);
3278  return __kmp_threads[gtid]->th.th_team;
3279 }
3280 
3281 static inline void __kmp_assert_valid_gtid(kmp_int32 gtid) {
3282  if (UNLIKELY(gtid < 0 || gtid >= __kmp_threads_capacity))
3283  KMP_FATAL(ThreadIdentInvalid);
3284 }
3285 
3286 #if KMP_HAVE_MWAIT || KMP_HAVE_UMWAIT
3287 extern int __kmp_user_level_mwait; // TRUE or FALSE; from KMP_USER_LEVEL_MWAIT
3288 extern int __kmp_umwait_enabled; // Runtime check if user-level mwait enabled
3289 extern int __kmp_mwait_enabled; // Runtime check if ring3 mwait is enabled
3290 extern int __kmp_mwait_hints; // Hints to pass in to mwait
3291 #endif
3292 
3293 /* ------------------------------------------------------------------------- */
3294 
3295 extern kmp_global_t __kmp_global; /* global status */
3296 
3297 extern kmp_info_t __kmp_monitor;
3298 // For Debugging Support Library
3299 extern std::atomic<kmp_int32> __kmp_team_counter;
3300 // For Debugging Support Library
3301 extern std::atomic<kmp_int32> __kmp_task_counter;
3302 
3303 #if USE_DEBUGGER
3304 #define _KMP_GEN_ID(counter) \
3305  (__kmp_debugging ? KMP_ATOMIC_INC(&counter) + 1 : ~0)
3306 #else
3307 #define _KMP_GEN_ID(counter) (~0)
3308 #endif /* USE_DEBUGGER */
3309 
3310 #define KMP_GEN_TASK_ID() _KMP_GEN_ID(__kmp_task_counter)
3311 #define KMP_GEN_TEAM_ID() _KMP_GEN_ID(__kmp_team_counter)
3312 
3313 /* ------------------------------------------------------------------------ */
3314 
3315 extern void __kmp_print_storage_map_gtid(int gtid, void *p1, void *p2,
3316  size_t size, char const *format, ...);
3317 
3318 extern void __kmp_serial_initialize(void);
3319 extern void __kmp_middle_initialize(void);
3320 extern void __kmp_parallel_initialize(void);
3321 
3322 extern void __kmp_internal_begin(void);
3323 extern void __kmp_internal_end_library(int gtid);
3324 extern void __kmp_internal_end_thread(int gtid);
3325 extern void __kmp_internal_end_atexit(void);
3326 extern void __kmp_internal_end_dtor(void);
3327 extern void __kmp_internal_end_dest(void *);
3328 
3329 extern int __kmp_register_root(int initial_thread);
3330 extern void __kmp_unregister_root(int gtid);
3331 extern void __kmp_unregister_library(void); // called by __kmp_internal_end()
3332 
3333 extern int __kmp_ignore_mppbeg(void);
3334 extern int __kmp_ignore_mppend(void);
3335 
3336 extern int __kmp_enter_single(int gtid, ident_t *id_ref, int push_ws);
3337 extern void __kmp_exit_single(int gtid);
3338 
3339 extern void __kmp_parallel_deo(int *gtid_ref, int *cid_ref, ident_t *loc_ref);
3340 extern void __kmp_parallel_dxo(int *gtid_ref, int *cid_ref, ident_t *loc_ref);
3341 
3342 #ifdef USE_LOAD_BALANCE
3343 extern int __kmp_get_load_balance(int);
3344 #endif
3345 
3346 extern int __kmp_get_global_thread_id(void);
3347 extern int __kmp_get_global_thread_id_reg(void);
3348 extern void __kmp_exit_thread(int exit_status);
3349 extern void __kmp_abort(char const *format, ...);
3350 extern void __kmp_abort_thread(void);
3351 KMP_NORETURN extern void __kmp_abort_process(void);
3352 extern void __kmp_warn(char const *format, ...);
3353 
3354 extern void __kmp_set_num_threads(int new_nth, int gtid);
3355 
3356 // Returns current thread (pointer to kmp_info_t). Current thread *must* be
3357 // registered.
3358 static inline kmp_info_t *__kmp_entry_thread() {
3359  int gtid = __kmp_entry_gtid();
3360 
3361  return __kmp_threads[gtid];
3362 }
3363 
3364 extern void __kmp_set_max_active_levels(int gtid, int new_max_active_levels);
3365 extern int __kmp_get_max_active_levels(int gtid);
3366 extern int __kmp_get_ancestor_thread_num(int gtid, int level);
3367 extern int __kmp_get_team_size(int gtid, int level);
3368 extern void __kmp_set_schedule(int gtid, kmp_sched_t new_sched, int chunk);
3369 extern void __kmp_get_schedule(int gtid, kmp_sched_t *sched, int *chunk);
3370 
3371 extern unsigned short __kmp_get_random(kmp_info_t *thread);
3372 extern void __kmp_init_random(kmp_info_t *thread);
3373 
3374 extern kmp_r_sched_t __kmp_get_schedule_global(void);
3375 extern void __kmp_adjust_num_threads(int new_nproc);
3376 extern void __kmp_check_stksize(size_t *val);
3377 
3378 extern void *___kmp_allocate(size_t size KMP_SRC_LOC_DECL);
3379 extern void *___kmp_page_allocate(size_t size KMP_SRC_LOC_DECL);
3380 extern void ___kmp_free(void *ptr KMP_SRC_LOC_DECL);
3381 #define __kmp_allocate(size) ___kmp_allocate((size)KMP_SRC_LOC_CURR)
3382 #define __kmp_page_allocate(size) ___kmp_page_allocate((size)KMP_SRC_LOC_CURR)
3383 #define __kmp_free(ptr) ___kmp_free((ptr)KMP_SRC_LOC_CURR)
3384 
3385 #if USE_FAST_MEMORY
3386 extern void *___kmp_fast_allocate(kmp_info_t *this_thr,
3387  size_t size KMP_SRC_LOC_DECL);
3388 extern void ___kmp_fast_free(kmp_info_t *this_thr, void *ptr KMP_SRC_LOC_DECL);
3389 extern void __kmp_free_fast_memory(kmp_info_t *this_thr);
3390 extern void __kmp_initialize_fast_memory(kmp_info_t *this_thr);
3391 #define __kmp_fast_allocate(this_thr, size) \
3392  ___kmp_fast_allocate((this_thr), (size)KMP_SRC_LOC_CURR)
3393 #define __kmp_fast_free(this_thr, ptr) \
3394  ___kmp_fast_free((this_thr), (ptr)KMP_SRC_LOC_CURR)
3395 #endif
3396 
3397 extern void *___kmp_thread_malloc(kmp_info_t *th, size_t size KMP_SRC_LOC_DECL);
3398 extern void *___kmp_thread_calloc(kmp_info_t *th, size_t nelem,
3399  size_t elsize KMP_SRC_LOC_DECL);
3400 extern void *___kmp_thread_realloc(kmp_info_t *th, void *ptr,
3401  size_t size KMP_SRC_LOC_DECL);
3402 extern void ___kmp_thread_free(kmp_info_t *th, void *ptr KMP_SRC_LOC_DECL);
3403 #define __kmp_thread_malloc(th, size) \
3404  ___kmp_thread_malloc((th), (size)KMP_SRC_LOC_CURR)
3405 #define __kmp_thread_calloc(th, nelem, elsize) \
3406  ___kmp_thread_calloc((th), (nelem), (elsize)KMP_SRC_LOC_CURR)
3407 #define __kmp_thread_realloc(th, ptr, size) \
3408  ___kmp_thread_realloc((th), (ptr), (size)KMP_SRC_LOC_CURR)
3409 #define __kmp_thread_free(th, ptr) \
3410  ___kmp_thread_free((th), (ptr)KMP_SRC_LOC_CURR)
3411 
3412 #define KMP_INTERNAL_MALLOC(sz) malloc(sz)
3413 #define KMP_INTERNAL_FREE(p) free(p)
3414 #define KMP_INTERNAL_REALLOC(p, sz) realloc((p), (sz))
3415 #define KMP_INTERNAL_CALLOC(n, sz) calloc((n), (sz))
3416 
3417 extern void __kmp_push_num_threads(ident_t *loc, int gtid, int num_threads);
3418 
3419 extern void __kmp_push_proc_bind(ident_t *loc, int gtid,
3420  kmp_proc_bind_t proc_bind);
3421 extern void __kmp_push_num_teams(ident_t *loc, int gtid, int num_teams,
3422  int num_threads);
3423 extern void __kmp_push_num_teams_51(ident_t *loc, int gtid, int num_teams_lb,
3424  int num_teams_ub, int num_threads);
3425 
3426 extern void __kmp_yield();
3427 
3428 extern void __kmpc_dispatch_init_4(ident_t *loc, kmp_int32 gtid,
3429  enum sched_type schedule, kmp_int32 lb,
3430  kmp_int32 ub, kmp_int32 st, kmp_int32 chunk);
3431 extern void __kmpc_dispatch_init_4u(ident_t *loc, kmp_int32 gtid,
3432  enum sched_type schedule, kmp_uint32 lb,
3433  kmp_uint32 ub, kmp_int32 st,
3434  kmp_int32 chunk);
3435 extern void __kmpc_dispatch_init_8(ident_t *loc, kmp_int32 gtid,
3436  enum sched_type schedule, kmp_int64 lb,
3437  kmp_int64 ub, kmp_int64 st, kmp_int64 chunk);
3438 extern void __kmpc_dispatch_init_8u(ident_t *loc, kmp_int32 gtid,
3439  enum sched_type schedule, kmp_uint64 lb,
3440  kmp_uint64 ub, kmp_int64 st,
3441  kmp_int64 chunk);
3442 
3443 extern int __kmpc_dispatch_next_4(ident_t *loc, kmp_int32 gtid,
3444  kmp_int32 *p_last, kmp_int32 *p_lb,
3445  kmp_int32 *p_ub, kmp_int32 *p_st);
3446 extern int __kmpc_dispatch_next_4u(ident_t *loc, kmp_int32 gtid,
3447  kmp_int32 *p_last, kmp_uint32 *p_lb,
3448  kmp_uint32 *p_ub, kmp_int32 *p_st);
3449 extern int __kmpc_dispatch_next_8(ident_t *loc, kmp_int32 gtid,
3450  kmp_int32 *p_last, kmp_int64 *p_lb,
3451  kmp_int64 *p_ub, kmp_int64 *p_st);
3452 extern int __kmpc_dispatch_next_8u(ident_t *loc, kmp_int32 gtid,
3453  kmp_int32 *p_last, kmp_uint64 *p_lb,
3454  kmp_uint64 *p_ub, kmp_int64 *p_st);
3455 
3456 extern void __kmpc_dispatch_fini_4(ident_t *loc, kmp_int32 gtid);
3457 extern void __kmpc_dispatch_fini_8(ident_t *loc, kmp_int32 gtid);
3458 extern void __kmpc_dispatch_fini_4u(ident_t *loc, kmp_int32 gtid);
3459 extern void __kmpc_dispatch_fini_8u(ident_t *loc, kmp_int32 gtid);
3460 
3461 #ifdef KMP_GOMP_COMPAT
3462 
3463 extern void __kmp_aux_dispatch_init_4(ident_t *loc, kmp_int32 gtid,
3464  enum sched_type schedule, kmp_int32 lb,
3465  kmp_int32 ub, kmp_int32 st,
3466  kmp_int32 chunk, int push_ws);
3467 extern void __kmp_aux_dispatch_init_4u(ident_t *loc, kmp_int32 gtid,
3468  enum sched_type schedule, kmp_uint32 lb,
3469  kmp_uint32 ub, kmp_int32 st,
3470  kmp_int32 chunk, int push_ws);
3471 extern void __kmp_aux_dispatch_init_8(ident_t *loc, kmp_int32 gtid,
3472  enum sched_type schedule, kmp_int64 lb,
3473  kmp_int64 ub, kmp_int64 st,
3474  kmp_int64 chunk, int push_ws);
3475 extern void __kmp_aux_dispatch_init_8u(ident_t *loc, kmp_int32 gtid,
3476  enum sched_type schedule, kmp_uint64 lb,
3477  kmp_uint64 ub, kmp_int64 st,
3478  kmp_int64 chunk, int push_ws);
3479 extern void __kmp_aux_dispatch_fini_chunk_4(ident_t *loc, kmp_int32 gtid);
3480 extern void __kmp_aux_dispatch_fini_chunk_8(ident_t *loc, kmp_int32 gtid);
3481 extern void __kmp_aux_dispatch_fini_chunk_4u(ident_t *loc, kmp_int32 gtid);
3482 extern void __kmp_aux_dispatch_fini_chunk_8u(ident_t *loc, kmp_int32 gtid);
3483 
3484 #endif /* KMP_GOMP_COMPAT */
3485 
3486 extern kmp_uint32 __kmp_eq_4(kmp_uint32 value, kmp_uint32 checker);
3487 extern kmp_uint32 __kmp_neq_4(kmp_uint32 value, kmp_uint32 checker);
3488 extern kmp_uint32 __kmp_lt_4(kmp_uint32 value, kmp_uint32 checker);
3489 extern kmp_uint32 __kmp_ge_4(kmp_uint32 value, kmp_uint32 checker);
3490 extern kmp_uint32 __kmp_le_4(kmp_uint32 value, kmp_uint32 checker);
3491 extern kmp_uint32 __kmp_wait_4(kmp_uint32 volatile *spinner, kmp_uint32 checker,
3492  kmp_uint32 (*pred)(kmp_uint32, kmp_uint32),
3493  void *obj);
3494 extern void __kmp_wait_4_ptr(void *spinner, kmp_uint32 checker,
3495  kmp_uint32 (*pred)(void *, kmp_uint32), void *obj);
3496 
3497 extern void __kmp_wait_64(kmp_info_t *this_thr, kmp_flag_64<> *flag,
3498  int final_spin
3499 #if USE_ITT_BUILD
3500  ,
3501  void *itt_sync_obj
3502 #endif
3503 );
3504 extern void __kmp_release_64(kmp_flag_64<> *flag);
3505 
3506 extern void __kmp_infinite_loop(void);
3507 
3508 extern void __kmp_cleanup(void);
3509 
3510 #if KMP_HANDLE_SIGNALS
3511 extern int __kmp_handle_signals;
3512 extern void __kmp_install_signals(int parallel_init);
3513 extern void __kmp_remove_signals(void);
3514 #endif
3515 
3516 extern void __kmp_clear_system_time(void);
3517 extern void __kmp_read_system_time(double *delta);
3518 
3519 extern void __kmp_check_stack_overlap(kmp_info_t *thr);
3520 
3521 extern void __kmp_expand_host_name(char *buffer, size_t size);
3522 extern void __kmp_expand_file_name(char *result, size_t rlen, char *pattern);
3523 
3524 #if KMP_ARCH_X86 || KMP_ARCH_X86_64 || (KMP_OS_WINDOWS && KMP_ARCH_AARCH64)
3525 extern void
3526 __kmp_initialize_system_tick(void); /* Initialize timer tick value */
3527 #endif
3528 
3529 extern void
3530 __kmp_runtime_initialize(void); /* machine specific initialization */
3531 extern void __kmp_runtime_destroy(void);
3532 
3533 #if KMP_AFFINITY_SUPPORTED
3534 extern char *__kmp_affinity_print_mask(char *buf, int buf_len,
3535  kmp_affin_mask_t *mask);
3536 extern kmp_str_buf_t *__kmp_affinity_str_buf_mask(kmp_str_buf_t *buf,
3537  kmp_affin_mask_t *mask);
3538 extern void __kmp_affinity_initialize(void);
3539 extern void __kmp_affinity_uninitialize(void);
3540 extern void __kmp_affinity_set_init_mask(
3541  int gtid, int isa_root); /* set affinity according to KMP_AFFINITY */
3542 extern void __kmp_affinity_set_place(int gtid);
3543 extern void __kmp_affinity_determine_capable(const char *env_var);
3544 extern int __kmp_aux_set_affinity(void **mask);
3545 extern int __kmp_aux_get_affinity(void **mask);
3546 extern int __kmp_aux_get_affinity_max_proc();
3547 extern int __kmp_aux_set_affinity_mask_proc(int proc, void **mask);
3548 extern int __kmp_aux_unset_affinity_mask_proc(int proc, void **mask);
3549 extern int __kmp_aux_get_affinity_mask_proc(int proc, void **mask);
3550 extern void __kmp_balanced_affinity(kmp_info_t *th, int team_size);
3551 #if KMP_OS_LINUX || KMP_OS_FREEBSD
3552 extern int kmp_set_thread_affinity_mask_initial(void);
3553 #endif
3554 static inline void __kmp_assign_root_init_mask() {
3555  int gtid = __kmp_entry_gtid();
3556  kmp_root_t *r = __kmp_threads[gtid]->th.th_root;
3557  if (r->r.r_uber_thread == __kmp_threads[gtid] && !r->r.r_affinity_assigned) {
3558  __kmp_affinity_set_init_mask(gtid, TRUE);
3559  r->r.r_affinity_assigned = TRUE;
3560  }
3561 }
3562 #else /* KMP_AFFINITY_SUPPORTED */
3563 #define __kmp_assign_root_init_mask() /* Nothing */
3564 #endif /* KMP_AFFINITY_SUPPORTED */
3565 // No need for KMP_AFFINITY_SUPPORTED guard as only one field in the
3566 // format string is for affinity, so platforms that do not support
3567 // affinity can still use the other fields, e.g., %n for num_threads
3568 extern size_t __kmp_aux_capture_affinity(int gtid, const char *format,
3569  kmp_str_buf_t *buffer);
3570 extern void __kmp_aux_display_affinity(int gtid, const char *format);
3571 
3572 extern void __kmp_cleanup_hierarchy();
3573 extern void __kmp_get_hierarchy(kmp_uint32 nproc, kmp_bstate_t *thr_bar);
3574 
3575 #if KMP_USE_FUTEX
3576 
3577 extern int __kmp_futex_determine_capable(void);
3578 
3579 #endif // KMP_USE_FUTEX
3580 
3581 extern void __kmp_gtid_set_specific(int gtid);
3582 extern int __kmp_gtid_get_specific(void);
3583 
3584 extern double __kmp_read_cpu_time(void);
3585 
3586 extern int __kmp_read_system_info(struct kmp_sys_info *info);
3587 
3588 #if KMP_USE_MONITOR
3589 extern void __kmp_create_monitor(kmp_info_t *th);
3590 #endif
3591 
3592 extern void *__kmp_launch_thread(kmp_info_t *thr);
3593 
3594 extern void __kmp_create_worker(int gtid, kmp_info_t *th, size_t stack_size);
3595 
3596 #if KMP_OS_WINDOWS
3597 extern int __kmp_still_running(kmp_info_t *th);
3598 extern int __kmp_is_thread_alive(kmp_info_t *th, DWORD *exit_val);
3599 extern void __kmp_free_handle(kmp_thread_t tHandle);
3600 #endif
3601 
3602 #if KMP_USE_MONITOR
3603 extern void __kmp_reap_monitor(kmp_info_t *th);
3604 #endif
3605 extern void __kmp_reap_worker(kmp_info_t *th);
3606 extern void __kmp_terminate_thread(int gtid);
3607 
3608 extern int __kmp_try_suspend_mx(kmp_info_t *th);
3609 extern void __kmp_lock_suspend_mx(kmp_info_t *th);
3610 extern void __kmp_unlock_suspend_mx(kmp_info_t *th);
3611 
3612 extern void __kmp_elapsed(double *);
3613 extern void __kmp_elapsed_tick(double *);
3614 
3615 extern void __kmp_enable(int old_state);
3616 extern void __kmp_disable(int *old_state);
3617 
3618 extern void __kmp_thread_sleep(int millis);
3619 
3620 extern void __kmp_common_initialize(void);
3621 extern void __kmp_common_destroy(void);
3622 extern void __kmp_common_destroy_gtid(int gtid);
3623 
3624 #if KMP_OS_UNIX
3625 extern void __kmp_register_atfork(void);
3626 #endif
3627 extern void __kmp_suspend_initialize(void);
3628 extern void __kmp_suspend_initialize_thread(kmp_info_t *th);
3629 extern void __kmp_suspend_uninitialize_thread(kmp_info_t *th);
3630 
3631 extern kmp_info_t *__kmp_allocate_thread(kmp_root_t *root, kmp_team_t *team,
3632  int tid);
3633 extern kmp_team_t *
3634 __kmp_allocate_team(kmp_root_t *root, int new_nproc, int max_nproc,
3635 #if OMPT_SUPPORT
3636  ompt_data_t ompt_parallel_data,
3637 #endif
3638  kmp_proc_bind_t proc_bind, kmp_internal_control_t *new_icvs,
3639  int argc USE_NESTED_HOT_ARG(kmp_info_t *thr));
3640 extern void __kmp_free_thread(kmp_info_t *);
3641 extern void __kmp_free_team(kmp_root_t *,
3642  kmp_team_t *USE_NESTED_HOT_ARG(kmp_info_t *));
3643 extern kmp_team_t *__kmp_reap_team(kmp_team_t *);
3644 
3645 /* ------------------------------------------------------------------------ */
3646 
3647 extern void __kmp_initialize_bget(kmp_info_t *th);
3648 extern void __kmp_finalize_bget(kmp_info_t *th);
3649 
3650 KMP_EXPORT void *kmpc_malloc(size_t size);
3651 KMP_EXPORT void *kmpc_aligned_malloc(size_t size, size_t alignment);
3652 KMP_EXPORT void *kmpc_calloc(size_t nelem, size_t elsize);
3653 KMP_EXPORT void *kmpc_realloc(void *ptr, size_t size);
3654 KMP_EXPORT void kmpc_free(void *ptr);
3655 
3656 /* declarations for internal use */
3657 
3658 extern int __kmp_barrier(enum barrier_type bt, int gtid, int is_split,
3659  size_t reduce_size, void *reduce_data,
3660  void (*reduce)(void *, void *));
3661 extern void __kmp_end_split_barrier(enum barrier_type bt, int gtid);
3662 extern int __kmp_barrier_gomp_cancel(int gtid);
3663 
3668 enum fork_context_e {
3669  fork_context_gnu,
3671  fork_context_intel,
3672  fork_context_last
3673 };
3674 extern int __kmp_fork_call(ident_t *loc, int gtid,
3675  enum fork_context_e fork_context, kmp_int32 argc,
3676  microtask_t microtask, launch_t invoker,
3677  kmp_va_list ap);
3678 
3679 extern void __kmp_join_call(ident_t *loc, int gtid
3680 #if OMPT_SUPPORT
3681  ,
3682  enum fork_context_e fork_context
3683 #endif
3684  ,
3685  int exit_teams = 0);
3686 
3687 extern void __kmp_serialized_parallel(ident_t *id, kmp_int32 gtid);
3688 extern void __kmp_internal_fork(ident_t *id, int gtid, kmp_team_t *team);
3689 extern void __kmp_internal_join(ident_t *id, int gtid, kmp_team_t *team);
3690 extern int __kmp_invoke_task_func(int gtid);
3691 extern void __kmp_run_before_invoked_task(int gtid, int tid,
3692  kmp_info_t *this_thr,
3693  kmp_team_t *team);
3694 extern void __kmp_run_after_invoked_task(int gtid, int tid,
3695  kmp_info_t *this_thr,
3696  kmp_team_t *team);
3697 
3698 // should never have been exported
3699 KMP_EXPORT int __kmpc_invoke_task_func(int gtid);
3700 extern int __kmp_invoke_teams_master(int gtid);
3701 extern void __kmp_teams_master(int gtid);
3702 extern int __kmp_aux_get_team_num();
3703 extern int __kmp_aux_get_num_teams();
3704 extern void __kmp_save_internal_controls(kmp_info_t *thread);
3705 extern void __kmp_user_set_library(enum library_type arg);
3706 extern void __kmp_aux_set_library(enum library_type arg);
3707 extern void __kmp_aux_set_stacksize(size_t arg);
3708 extern void __kmp_aux_set_blocktime(int arg, kmp_info_t *thread, int tid);
3709 extern void __kmp_aux_set_defaults(char const *str, size_t len);
3710 
3711 /* Functions called from __kmp_aux_env_initialize() in kmp_settings.cpp */
3712 void kmpc_set_blocktime(int arg);
3713 void ompc_set_nested(int flag);
3714 void ompc_set_dynamic(int flag);
3715 void ompc_set_num_threads(int arg);
3716 
3717 extern void __kmp_push_current_task_to_thread(kmp_info_t *this_thr,
3718  kmp_team_t *team, int tid);
3719 extern void __kmp_pop_current_task_from_thread(kmp_info_t *this_thr);
3720 extern kmp_task_t *__kmp_task_alloc(ident_t *loc_ref, kmp_int32 gtid,
3721  kmp_tasking_flags_t *flags,
3722  size_t sizeof_kmp_task_t,
3723  size_t sizeof_shareds,
3724  kmp_routine_entry_t task_entry);
3725 extern void __kmp_init_implicit_task(ident_t *loc_ref, kmp_info_t *this_thr,
3726  kmp_team_t *team, int tid,
3727  int set_curr_task);
3728 extern void __kmp_finish_implicit_task(kmp_info_t *this_thr);
3729 extern void __kmp_free_implicit_task(kmp_info_t *this_thr);
3730 
3731 extern kmp_event_t *__kmpc_task_allow_completion_event(ident_t *loc_ref,
3732  int gtid,
3733  kmp_task_t *task);
3734 extern void __kmp_fulfill_event(kmp_event_t *event);
3735 
3736 extern void __kmp_free_task_team(kmp_info_t *thread,
3737  kmp_task_team_t *task_team);
3738 extern void __kmp_reap_task_teams(void);
3739 extern void __kmp_wait_to_unref_task_teams(void);
3740 extern void __kmp_task_team_setup(kmp_info_t *this_thr, kmp_team_t *team,
3741  int always);
3742 extern void __kmp_task_team_sync(kmp_info_t *this_thr, kmp_team_t *team);
3743 extern void __kmp_task_team_wait(kmp_info_t *this_thr, kmp_team_t *team
3744 #if USE_ITT_BUILD
3745  ,
3746  void *itt_sync_obj
3747 #endif /* USE_ITT_BUILD */
3748  ,
3749  int wait = 1);
3750 extern void __kmp_tasking_barrier(kmp_team_t *team, kmp_info_t *thread,
3751  int gtid);
3752 
3753 extern int __kmp_is_address_mapped(void *addr);
3754 extern kmp_uint64 __kmp_hardware_timestamp(void);
3755 
3756 #if KMP_OS_UNIX
3757 extern int __kmp_read_from_file(char const *path, char const *format, ...);
3758 #endif
3759 
3760 /* ------------------------------------------------------------------------ */
3761 //
3762 // Assembly routines that have no compiler intrinsic replacement
3763 //
3764 
3765 extern int __kmp_invoke_microtask(microtask_t pkfn, int gtid, int npr, int argc,
3766  void *argv[]
3767 #if OMPT_SUPPORT
3768  ,
3769  void **exit_frame_ptr
3770 #endif
3771 );
3772 
3773 /* ------------------------------------------------------------------------ */
3774 
3775 KMP_EXPORT void __kmpc_begin(ident_t *, kmp_int32 flags);
3776 KMP_EXPORT void __kmpc_end(ident_t *);
3777 
3778 KMP_EXPORT void __kmpc_threadprivate_register_vec(ident_t *, void *data,
3779  kmpc_ctor_vec ctor,
3780  kmpc_cctor_vec cctor,
3781  kmpc_dtor_vec dtor,
3782  size_t vector_length);
3783 KMP_EXPORT void __kmpc_threadprivate_register(ident_t *, void *data,
3784  kmpc_ctor ctor, kmpc_cctor cctor,
3785  kmpc_dtor dtor);
3786 KMP_EXPORT void *__kmpc_threadprivate(ident_t *, kmp_int32 global_tid,
3787  void *data, size_t size);
3788 
3789 KMP_EXPORT kmp_int32 __kmpc_global_thread_num(ident_t *);
3790 KMP_EXPORT kmp_int32 __kmpc_global_num_threads(ident_t *);
3791 KMP_EXPORT kmp_int32 __kmpc_bound_thread_num(ident_t *);
3792 KMP_EXPORT kmp_int32 __kmpc_bound_num_threads(ident_t *);
3793 
3794 KMP_EXPORT kmp_int32 __kmpc_ok_to_fork(ident_t *);
3795 KMP_EXPORT void __kmpc_fork_call(ident_t *, kmp_int32 nargs,
3796  kmpc_micro microtask, ...);
3797 
3798 KMP_EXPORT void __kmpc_serialized_parallel(ident_t *, kmp_int32 global_tid);
3799 KMP_EXPORT void __kmpc_end_serialized_parallel(ident_t *, kmp_int32 global_tid);
3800 
3801 KMP_EXPORT void __kmpc_flush(ident_t *);
3802 KMP_EXPORT void __kmpc_barrier(ident_t *, kmp_int32 global_tid);
3803 KMP_EXPORT kmp_int32 __kmpc_master(ident_t *, kmp_int32 global_tid);
3804 KMP_EXPORT void __kmpc_end_master(ident_t *, kmp_int32 global_tid);
3805 KMP_EXPORT kmp_int32 __kmpc_masked(ident_t *, kmp_int32 global_tid,
3806  kmp_int32 filter);
3807 KMP_EXPORT void __kmpc_end_masked(ident_t *, kmp_int32 global_tid);
3808 KMP_EXPORT void __kmpc_ordered(ident_t *, kmp_int32 global_tid);
3809 KMP_EXPORT void __kmpc_end_ordered(ident_t *, kmp_int32 global_tid);
3810 KMP_EXPORT void __kmpc_critical(ident_t *, kmp_int32 global_tid,
3811  kmp_critical_name *);
3812 KMP_EXPORT void __kmpc_end_critical(ident_t *, kmp_int32 global_tid,
3813  kmp_critical_name *);
3814 KMP_EXPORT void __kmpc_critical_with_hint(ident_t *, kmp_int32 global_tid,
3815  kmp_critical_name *, uint32_t hint);
3816 
3817 KMP_EXPORT kmp_int32 __kmpc_barrier_master(ident_t *, kmp_int32 global_tid);
3818 KMP_EXPORT void __kmpc_end_barrier_master(ident_t *, kmp_int32 global_tid);
3819 
3820 KMP_EXPORT kmp_int32 __kmpc_barrier_master_nowait(ident_t *,
3821  kmp_int32 global_tid);
3822 
3823 KMP_EXPORT kmp_int32 __kmpc_single(ident_t *, kmp_int32 global_tid);
3824 KMP_EXPORT void __kmpc_end_single(ident_t *, kmp_int32 global_tid);
3825 
3826 KMP_EXPORT void KMPC_FOR_STATIC_INIT(ident_t *loc, kmp_int32 global_tid,
3827  kmp_int32 schedtype, kmp_int32 *plastiter,
3828  kmp_int *plower, kmp_int *pupper,
3829  kmp_int *pstride, kmp_int incr,
3830  kmp_int chunk);
3831 
3832 KMP_EXPORT void __kmpc_for_static_fini(ident_t *loc, kmp_int32 global_tid);
3833 
3834 KMP_EXPORT void __kmpc_copyprivate(ident_t *loc, kmp_int32 global_tid,
3835  size_t cpy_size, void *cpy_data,
3836  void (*cpy_func)(void *, void *),
3837  kmp_int32 didit);
3838 
3839 extern void KMPC_SET_NUM_THREADS(int arg);
3840 extern void KMPC_SET_DYNAMIC(int flag);
3841 extern void KMPC_SET_NESTED(int flag);
3842 
3843 /* OMP 3.0 tasking interface routines */
3844 KMP_EXPORT kmp_int32 __kmpc_omp_task(ident_t *loc_ref, kmp_int32 gtid,
3845  kmp_task_t *new_task);
3846 KMP_EXPORT kmp_task_t *__kmpc_omp_task_alloc(ident_t *loc_ref, kmp_int32 gtid,
3847  kmp_int32 flags,
3848  size_t sizeof_kmp_task_t,
3849  size_t sizeof_shareds,
3850  kmp_routine_entry_t task_entry);
3851 KMP_EXPORT kmp_task_t *__kmpc_omp_target_task_alloc(
3852  ident_t *loc_ref, kmp_int32 gtid, kmp_int32 flags, size_t sizeof_kmp_task_t,
3853  size_t sizeof_shareds, kmp_routine_entry_t task_entry, kmp_int64 device_id);
3854 KMP_EXPORT void __kmpc_omp_task_begin_if0(ident_t *loc_ref, kmp_int32 gtid,
3855  kmp_task_t *task);
3856 KMP_EXPORT void __kmpc_omp_task_complete_if0(ident_t *loc_ref, kmp_int32 gtid,
3857  kmp_task_t *task);
3858 KMP_EXPORT kmp_int32 __kmpc_omp_task_parts(ident_t *loc_ref, kmp_int32 gtid,
3859  kmp_task_t *new_task);
3860 KMP_EXPORT kmp_int32 __kmpc_omp_taskwait(ident_t *loc_ref, kmp_int32 gtid);
3861 
3862 KMP_EXPORT kmp_int32 __kmpc_omp_taskyield(ident_t *loc_ref, kmp_int32 gtid,
3863  int end_part);
3864 
3865 #if TASK_UNUSED
3866 void __kmpc_omp_task_begin(ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *task);
3867 void __kmpc_omp_task_complete(ident_t *loc_ref, kmp_int32 gtid,
3868  kmp_task_t *task);
3869 #endif // TASK_UNUSED
3870 
3871 /* ------------------------------------------------------------------------ */
3872 
3873 KMP_EXPORT void __kmpc_taskgroup(ident_t *loc, int gtid);
3874 KMP_EXPORT void __kmpc_end_taskgroup(ident_t *loc, int gtid);
3875 
3876 KMP_EXPORT kmp_int32 __kmpc_omp_task_with_deps(
3877  ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 ndeps,
3878  kmp_depend_info_t *dep_list, kmp_int32 ndeps_noalias,
3879  kmp_depend_info_t *noalias_dep_list);
3880 KMP_EXPORT void __kmpc_omp_wait_deps(ident_t *loc_ref, kmp_int32 gtid,
3881  kmp_int32 ndeps,
3882  kmp_depend_info_t *dep_list,
3883  kmp_int32 ndeps_noalias,
3884  kmp_depend_info_t *noalias_dep_list);
3885 extern kmp_int32 __kmp_omp_task(kmp_int32 gtid, kmp_task_t *new_task,
3886  bool serialize_immediate);
3887 
3888 KMP_EXPORT kmp_int32 __kmpc_cancel(ident_t *loc_ref, kmp_int32 gtid,
3889  kmp_int32 cncl_kind);
3890 KMP_EXPORT kmp_int32 __kmpc_cancellationpoint(ident_t *loc_ref, kmp_int32 gtid,
3891  kmp_int32 cncl_kind);
3892 KMP_EXPORT kmp_int32 __kmpc_cancel_barrier(ident_t *loc_ref, kmp_int32 gtid);
3893 KMP_EXPORT int __kmp_get_cancellation_status(int cancel_kind);
3894 
3895 KMP_EXPORT void __kmpc_proxy_task_completed(kmp_int32 gtid, kmp_task_t *ptask);
3896 KMP_EXPORT void __kmpc_proxy_task_completed_ooo(kmp_task_t *ptask);
3897 KMP_EXPORT void __kmpc_taskloop(ident_t *loc, kmp_int32 gtid, kmp_task_t *task,
3898  kmp_int32 if_val, kmp_uint64 *lb,
3899  kmp_uint64 *ub, kmp_int64 st, kmp_int32 nogroup,
3900  kmp_int32 sched, kmp_uint64 grainsize,
3901  void *task_dup);
3902 KMP_EXPORT void __kmpc_taskloop_5(ident_t *loc, kmp_int32 gtid,
3903  kmp_task_t *task, kmp_int32 if_val,
3904  kmp_uint64 *lb, kmp_uint64 *ub, kmp_int64 st,
3905  kmp_int32 nogroup, kmp_int32 sched,
3906  kmp_uint64 grainsize, kmp_int32 modifier,
3907  void *task_dup);
3908 KMP_EXPORT void *__kmpc_task_reduction_init(int gtid, int num_data, void *data);
3909 KMP_EXPORT void *__kmpc_taskred_init(int gtid, int num_data, void *data);
3910 KMP_EXPORT void *__kmpc_task_reduction_get_th_data(int gtid, void *tg, void *d);
3911 KMP_EXPORT void *__kmpc_task_reduction_modifier_init(ident_t *loc, int gtid,
3912  int is_ws, int num,
3913  void *data);
3914 KMP_EXPORT void *__kmpc_taskred_modifier_init(ident_t *loc, int gtid, int is_ws,
3915  int num, void *data);
3916 KMP_EXPORT void __kmpc_task_reduction_modifier_fini(ident_t *loc, int gtid,
3917  int is_ws);
3918 KMP_EXPORT kmp_int32 __kmpc_omp_reg_task_with_affinity(
3919  ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 naffins,
3920  kmp_task_affinity_info_t *affin_list);
3921 KMP_EXPORT void __kmp_set_num_teams(int num_teams);
3922 KMP_EXPORT int __kmp_get_max_teams(void);
3923 KMP_EXPORT void __kmp_set_teams_thread_limit(int limit);
3924 KMP_EXPORT int __kmp_get_teams_thread_limit(void);
3925 
3926 /* Lock interface routines (fast versions with gtid passed in) */
3927 KMP_EXPORT void __kmpc_init_lock(ident_t *loc, kmp_int32 gtid,
3928  void **user_lock);
3929 KMP_EXPORT void __kmpc_init_nest_lock(ident_t *loc, kmp_int32 gtid,
3930  void **user_lock);
3931 KMP_EXPORT void __kmpc_destroy_lock(ident_t *loc, kmp_int32 gtid,
3932  void **user_lock);
3933 KMP_EXPORT void __kmpc_destroy_nest_lock(ident_t *loc, kmp_int32 gtid,
3934  void **user_lock);
3935 KMP_EXPORT void __kmpc_set_lock(ident_t *loc, kmp_int32 gtid, void **user_lock);
3936 KMP_EXPORT void __kmpc_set_nest_lock(ident_t *loc, kmp_int32 gtid,
3937  void **user_lock);
3938 KMP_EXPORT void __kmpc_unset_lock(ident_t *loc, kmp_int32 gtid,
3939  void **user_lock);
3940 KMP_EXPORT void __kmpc_unset_nest_lock(ident_t *loc, kmp_int32 gtid,
3941  void **user_lock);
3942 KMP_EXPORT int __kmpc_test_lock(ident_t *loc, kmp_int32 gtid, void **user_lock);
3943 KMP_EXPORT int __kmpc_test_nest_lock(ident_t *loc, kmp_int32 gtid,
3944  void **user_lock);
3945 
3946 KMP_EXPORT void __kmpc_init_lock_with_hint(ident_t *loc, kmp_int32 gtid,
3947  void **user_lock, uintptr_t hint);
3948 KMP_EXPORT void __kmpc_init_nest_lock_with_hint(ident_t *loc, kmp_int32 gtid,
3949  void **user_lock,
3950  uintptr_t hint);
3951 
3952 /* Interface to fast scalable reduce methods routines */
3953 
3954 KMP_EXPORT kmp_int32 __kmpc_reduce_nowait(
3955  ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
3956  void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
3957  kmp_critical_name *lck);
3958 KMP_EXPORT void __kmpc_end_reduce_nowait(ident_t *loc, kmp_int32 global_tid,
3959  kmp_critical_name *lck);
3960 KMP_EXPORT kmp_int32 __kmpc_reduce(
3961  ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
3962  void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
3963  kmp_critical_name *lck);
3964 KMP_EXPORT void __kmpc_end_reduce(ident_t *loc, kmp_int32 global_tid,
3965  kmp_critical_name *lck);
3966 
3967 /* Internal fast reduction routines */
3968 
3969 extern PACKED_REDUCTION_METHOD_T __kmp_determine_reduction_method(
3970  ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
3971  void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
3972  kmp_critical_name *lck);
3973 
3974 // this function is for testing set/get/determine reduce method
3975 KMP_EXPORT kmp_int32 __kmp_get_reduce_method(void);
3976 
3977 KMP_EXPORT kmp_uint64 __kmpc_get_taskid();
3978 KMP_EXPORT kmp_uint64 __kmpc_get_parent_taskid();
3979 
3980 // C++ port
3981 // missing 'extern "C"' declarations
3982 
3983 KMP_EXPORT kmp_int32 __kmpc_in_parallel(ident_t *loc);
3984 KMP_EXPORT void __kmpc_pop_num_threads(ident_t *loc, kmp_int32 global_tid);
3985 KMP_EXPORT void __kmpc_push_num_threads(ident_t *loc, kmp_int32 global_tid,
3986  kmp_int32 num_threads);
3987 
3988 KMP_EXPORT void __kmpc_push_proc_bind(ident_t *loc, kmp_int32 global_tid,
3989  int proc_bind);
3990 KMP_EXPORT void __kmpc_push_num_teams(ident_t *loc, kmp_int32 global_tid,
3991  kmp_int32 num_teams,
3992  kmp_int32 num_threads);
3993 /* Function for OpenMP 5.1 num_teams clause */
3994 KMP_EXPORT void __kmpc_push_num_teams_51(ident_t *loc, kmp_int32 global_tid,
3995  kmp_int32 num_teams_lb,
3996  kmp_int32 num_teams_ub,
3997  kmp_int32 num_threads);
3998 KMP_EXPORT void __kmpc_fork_teams(ident_t *loc, kmp_int32 argc,
3999  kmpc_micro microtask, ...);
4000 struct kmp_dim { // loop bounds info casted to kmp_int64
4001  kmp_int64 lo; // lower
4002  kmp_int64 up; // upper
4003  kmp_int64 st; // stride
4004 };
4005 KMP_EXPORT void __kmpc_doacross_init(ident_t *loc, kmp_int32 gtid,
4006  kmp_int32 num_dims,
4007  const struct kmp_dim *dims);
4008 KMP_EXPORT void __kmpc_doacross_wait(ident_t *loc, kmp_int32 gtid,
4009  const kmp_int64 *vec);
4010 KMP_EXPORT void __kmpc_doacross_post(ident_t *loc, kmp_int32 gtid,
4011  const kmp_int64 *vec);
4012 KMP_EXPORT void __kmpc_doacross_fini(ident_t *loc, kmp_int32 gtid);
4013 
4014 KMP_EXPORT void *__kmpc_threadprivate_cached(ident_t *loc, kmp_int32 global_tid,
4015  void *data, size_t size,
4016  void ***cache);
4017 
4018 // Symbols for MS mutual detection.
4019 extern int _You_must_link_with_exactly_one_OpenMP_library;
4020 extern int _You_must_link_with_Intel_OpenMP_library;
4021 #if KMP_OS_WINDOWS && (KMP_VERSION_MAJOR > 4)
4022 extern int _You_must_link_with_Microsoft_OpenMP_library;
4023 #endif
4024 
4025 // The routines below are not exported.
4026 // Consider making them 'static' in corresponding source files.
4027 void kmp_threadprivate_insert_private_data(int gtid, void *pc_addr,
4028  void *data_addr, size_t pc_size);
4029 struct private_common *kmp_threadprivate_insert(int gtid, void *pc_addr,
4030  void *data_addr,
4031  size_t pc_size);
4032 void __kmp_threadprivate_resize_cache(int newCapacity);
4033 void __kmp_cleanup_threadprivate_caches();
4034 
4035 // ompc_, kmpc_ entries moved from omp.h.
4036 #if KMP_OS_WINDOWS
4037 #define KMPC_CONVENTION __cdecl
4038 #else
4039 #define KMPC_CONVENTION
4040 #endif
4041 
4042 #ifndef __OMP_H
4043 typedef enum omp_sched_t {
4044  omp_sched_static = 1,
4045  omp_sched_dynamic = 2,
4046  omp_sched_guided = 3,
4047  omp_sched_auto = 4
4048 } omp_sched_t;
4049 typedef void *kmp_affinity_mask_t;
4050 #endif
4051 
4052 KMP_EXPORT void KMPC_CONVENTION ompc_set_max_active_levels(int);
4053 KMP_EXPORT void KMPC_CONVENTION ompc_set_schedule(omp_sched_t, int);
4054 KMP_EXPORT int KMPC_CONVENTION ompc_get_ancestor_thread_num(int);
4055 KMP_EXPORT int KMPC_CONVENTION ompc_get_team_size(int);
4056 KMP_EXPORT int KMPC_CONVENTION
4057 kmpc_set_affinity_mask_proc(int, kmp_affinity_mask_t *);
4058 KMP_EXPORT int KMPC_CONVENTION
4059 kmpc_unset_affinity_mask_proc(int, kmp_affinity_mask_t *);
4060 KMP_EXPORT int KMPC_CONVENTION
4061 kmpc_get_affinity_mask_proc(int, kmp_affinity_mask_t *);
4062 
4063 KMP_EXPORT void KMPC_CONVENTION kmpc_set_stacksize(int);
4064 KMP_EXPORT void KMPC_CONVENTION kmpc_set_stacksize_s(size_t);
4065 KMP_EXPORT void KMPC_CONVENTION kmpc_set_library(int);
4066 KMP_EXPORT void KMPC_CONVENTION kmpc_set_defaults(char const *);
4067 KMP_EXPORT void KMPC_CONVENTION kmpc_set_disp_num_buffers(int);
4068 void KMP_EXPAND_NAME(ompc_set_affinity_format)(char const *format);
4069 size_t KMP_EXPAND_NAME(ompc_get_affinity_format)(char *buffer, size_t size);
4070 void KMP_EXPAND_NAME(ompc_display_affinity)(char const *format);
4071 size_t KMP_EXPAND_NAME(ompc_capture_affinity)(char *buffer, size_t buf_size,
4072  char const *format);
4073 
4074 enum kmp_target_offload_kind {
4075  tgt_disabled = 0,
4076  tgt_default = 1,
4077  tgt_mandatory = 2
4078 };
4079 typedef enum kmp_target_offload_kind kmp_target_offload_kind_t;
4080 // Set via OMP_TARGET_OFFLOAD if specified, defaults to tgt_default otherwise
4081 extern kmp_target_offload_kind_t __kmp_target_offload;
4082 extern int __kmpc_get_target_offload();
4083 
4084 // Constants used in libomptarget
4085 #define KMP_DEVICE_DEFAULT -1 // This is libomptarget's default device.
4086 #define KMP_DEVICE_ALL -11 // This is libomptarget's "all devices".
4087 
4088 // OMP Pause Resource
4089 
4090 // The following enum is used both to set the status in __kmp_pause_status, and
4091 // as the internal equivalent of the externally-visible omp_pause_resource_t.
4092 typedef enum kmp_pause_status_t {
4093  kmp_not_paused = 0, // status is not paused, or, requesting resume
4094  kmp_soft_paused = 1, // status is soft-paused, or, requesting soft pause
4095  kmp_hard_paused = 2 // status is hard-paused, or, requesting hard pause
4096 } kmp_pause_status_t;
4097 
4098 // This stores the pause state of the runtime
4099 extern kmp_pause_status_t __kmp_pause_status;
4100 extern int __kmpc_pause_resource(kmp_pause_status_t level);
4101 extern int __kmp_pause_resource(kmp_pause_status_t level);
4102 // Soft resume sets __kmp_pause_status, and wakes up all threads.
4103 extern void __kmp_resume_if_soft_paused();
4104 // Hard resume simply resets the status to not paused. Library will appear to
4105 // be uninitialized after hard pause. Let OMP constructs trigger required
4106 // initializations.
4107 static inline void __kmp_resume_if_hard_paused() {
4108  if (__kmp_pause_status == kmp_hard_paused) {
4109  __kmp_pause_status = kmp_not_paused;
4110  }
4111 }
4112 
4113 extern void __kmp_omp_display_env(int verbose);
4114 
4115 // 1: it is initializing hidden helper team
4116 extern volatile int __kmp_init_hidden_helper;
4117 // 1: the hidden helper team is done
4118 extern volatile int __kmp_hidden_helper_team_done;
4119 // 1: enable hidden helper task
4120 extern kmp_int32 __kmp_enable_hidden_helper;
4121 // Main thread of hidden helper team
4122 extern kmp_info_t *__kmp_hidden_helper_main_thread;
4123 // Descriptors for the hidden helper threads
4124 extern kmp_info_t **__kmp_hidden_helper_threads;
4125 // Number of hidden helper threads
4126 extern kmp_int32 __kmp_hidden_helper_threads_num;
4127 // Number of hidden helper tasks that have not been executed yet
4128 extern std::atomic<kmp_int32> __kmp_unexecuted_hidden_helper_tasks;
4129 
4130 extern void __kmp_hidden_helper_initialize();
4131 extern void __kmp_hidden_helper_threads_initz_routine();
4132 extern void __kmp_do_initialize_hidden_helper_threads();
4133 extern void __kmp_hidden_helper_threads_initz_wait();
4134 extern void __kmp_hidden_helper_initz_release();
4135 extern void __kmp_hidden_helper_threads_deinitz_wait();
4136 extern void __kmp_hidden_helper_threads_deinitz_release();
4137 extern void __kmp_hidden_helper_main_thread_wait();
4138 extern void __kmp_hidden_helper_worker_thread_wait();
4139 extern void __kmp_hidden_helper_worker_thread_signal();
4140 extern void __kmp_hidden_helper_main_thread_release();
4141 
4142 // Check whether a given thread is a hidden helper thread
4143 #define KMP_HIDDEN_HELPER_THREAD(gtid) \
4144  ((gtid) >= 1 && (gtid) <= __kmp_hidden_helper_threads_num)
4145 
4146 #define KMP_HIDDEN_HELPER_WORKER_THREAD(gtid) \
4147  ((gtid) > 1 && (gtid) <= __kmp_hidden_helper_threads_num)
4148 
4149 #define KMP_HIDDEN_HELPER_TEAM(team) \
4150  (team->t.t_threads[0] == __kmp_hidden_helper_main_thread)
4151 
4152 // Map a gtid to a hidden helper thread. The first hidden helper thread, a.k.a
4153 // main thread, is skipped.
4154 #define KMP_GTID_TO_SHADOW_GTID(gtid) \
4155  ((gtid) % (__kmp_hidden_helper_threads_num - 1) + 2)
4156 
4157 // Return the adjusted gtid value by subtracting from gtid the number
4158 // of hidden helper threads. This adjusted value is the gtid the thread would
4159 // have received if there were no hidden helper threads.
4160 static inline int __kmp_adjust_gtid_for_hidden_helpers(int gtid) {
4161  int adjusted_gtid = gtid;
4162  if (__kmp_hidden_helper_threads_num > 0 && gtid > 0 &&
4163  gtid - __kmp_hidden_helper_threads_num >= 0) {
4164  adjusted_gtid -= __kmp_hidden_helper_threads_num;
4165  }
4166  return adjusted_gtid;
4167 }
4168 
4169 // Support for error directive
4170 typedef enum kmp_severity_t {
4171  severity_warning = 1,
4172  severity_fatal = 2
4173 } kmp_severity_t;
4174 extern void __kmpc_error(ident_t *loc, int severity, const char *message);
4175 
4176 // Support for scope directive
4177 KMP_EXPORT void __kmpc_scope(ident_t *loc, kmp_int32 gtid, void *reserved);
4178 KMP_EXPORT void __kmpc_end_scope(ident_t *loc, kmp_int32 gtid, void *reserved);
4179 
4180 #ifdef __cplusplus
4181 }
4182 #endif
4183 
4184 template <bool C, bool S>
4185 extern void __kmp_suspend_32(int th_gtid, kmp_flag_32<C, S> *flag);
4186 template <bool C, bool S>
4187 extern void __kmp_suspend_64(int th_gtid, kmp_flag_64<C, S> *flag);
4188 template <bool C, bool S>
4189 extern void __kmp_atomic_suspend_64(int th_gtid,
4190  kmp_atomic_flag_64<C, S> *flag);
4191 extern void __kmp_suspend_oncore(int th_gtid, kmp_flag_oncore *flag);
4192 #if KMP_HAVE_MWAIT || KMP_HAVE_UMWAIT
4193 template <bool C, bool S>
4194 extern void __kmp_mwait_32(int th_gtid, kmp_flag_32<C, S> *flag);
4195 template <bool C, bool S>
4196 extern void __kmp_mwait_64(int th_gtid, kmp_flag_64<C, S> *flag);
4197 template <bool C, bool S>
4198 extern void __kmp_atomic_mwait_64(int th_gtid, kmp_atomic_flag_64<C, S> *flag);
4199 extern void __kmp_mwait_oncore(int th_gtid, kmp_flag_oncore *flag);
4200 #endif
4201 template <bool C, bool S>
4202 extern void __kmp_resume_32(int target_gtid, kmp_flag_32<C, S> *flag);
4203 template <bool C, bool S>
4204 extern void __kmp_resume_64(int target_gtid, kmp_flag_64<C, S> *flag);
4205 template <bool C, bool S>
4206 extern void __kmp_atomic_resume_64(int target_gtid,
4207  kmp_atomic_flag_64<C, S> *flag);
4208 extern void __kmp_resume_oncore(int target_gtid, kmp_flag_oncore *flag);
4209 
4210 template <bool C, bool S>
4211 int __kmp_execute_tasks_32(kmp_info_t *thread, kmp_int32 gtid,
4212  kmp_flag_32<C, S> *flag, int final_spin,
4213  int *thread_finished,
4214 #if USE_ITT_BUILD
4215  void *itt_sync_obj,
4216 #endif /* USE_ITT_BUILD */
4217  kmp_int32 is_constrained);
4218 template <bool C, bool S>
4219 int __kmp_execute_tasks_64(kmp_info_t *thread, kmp_int32 gtid,
4220  kmp_flag_64<C, S> *flag, int final_spin,
4221  int *thread_finished,
4222 #if USE_ITT_BUILD
4223  void *itt_sync_obj,
4224 #endif /* USE_ITT_BUILD */
4225  kmp_int32 is_constrained);
4226 template <bool C, bool S>
4227 int __kmp_atomic_execute_tasks_64(kmp_info_t *thread, kmp_int32 gtid,
4228  kmp_atomic_flag_64<C, S> *flag,
4229  int final_spin, int *thread_finished,
4230 #if USE_ITT_BUILD
4231  void *itt_sync_obj,
4232 #endif /* USE_ITT_BUILD */
4233  kmp_int32 is_constrained);
4234 int __kmp_execute_tasks_oncore(kmp_info_t *thread, kmp_int32 gtid,
4235  kmp_flag_oncore *flag, int final_spin,
4236  int *thread_finished,
4237 #if USE_ITT_BUILD
4238  void *itt_sync_obj,
4239 #endif /* USE_ITT_BUILD */
4240  kmp_int32 is_constrained);
4241 
4242 extern int __kmp_nesting_mode;
4243 extern int __kmp_nesting_mode_nlevels;
4244 extern int *__kmp_nesting_nth_level;
4245 extern void __kmp_init_nesting_mode();
4246 extern void __kmp_set_nesting_mode_threads();
4247 
4255  FILE *f;
4256 
4257  void close() {
4258  if (f && f != stdout && f != stderr) {
4259  fclose(f);
4260  f = nullptr;
4261  }
4262  }
4263 
4264 public:
4265  kmp_safe_raii_file_t() : f(nullptr) {}
4266  kmp_safe_raii_file_t(const char *filename, const char *mode,
4267  const char *env_var = nullptr)
4268  : f(nullptr) {
4269  open(filename, mode, env_var);
4270  }
4271  ~kmp_safe_raii_file_t() { close(); }
4272 
4276  void open(const char *filename, const char *mode,
4277  const char *env_var = nullptr) {
4278  KMP_ASSERT(!f);
4279  f = fopen(filename, mode);
4280  if (!f) {
4281  int code = errno;
4282  if (env_var) {
4283  __kmp_fatal(KMP_MSG(CantOpenFileForReading, filename), KMP_ERR(code),
4284  KMP_HNT(CheckEnvVar, env_var, filename), __kmp_msg_null);
4285  } else {
4286  __kmp_fatal(KMP_MSG(CantOpenFileForReading, filename), KMP_ERR(code),
4287  __kmp_msg_null);
4288  }
4289  }
4290  }
4293  int try_open(const char *filename, const char *mode) {
4294  KMP_ASSERT(!f);
4295  f = fopen(filename, mode);
4296  if (!f)
4297  return errno;
4298  return 0;
4299  }
4302  void set_stdout() {
4303  KMP_ASSERT(!f);
4304  f = stdout;
4305  }
4308  void set_stderr() {
4309  KMP_ASSERT(!f);
4310  f = stderr;
4311  }
4312  operator bool() { return bool(f); }
4313  operator FILE *() { return f; }
4314 };
4315 
4316 template <typename SourceType, typename TargetType,
4317  bool isSourceSmaller = (sizeof(SourceType) < sizeof(TargetType)),
4318  bool isSourceEqual = (sizeof(SourceType) == sizeof(TargetType)),
4319  bool isSourceSigned = std::is_signed<SourceType>::value,
4320  bool isTargetSigned = std::is_signed<TargetType>::value>
4321 struct kmp_convert {};
4322 
4323 // Both types are signed; Source smaller
4324 template <typename SourceType, typename TargetType>
4325 struct kmp_convert<SourceType, TargetType, true, false, true, 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, true, true> {
4331  static TargetType to(SourceType src) { return src; }
4332 };
4333 // Source bigger
4334 template <typename SourceType, typename TargetType>
4335 struct kmp_convert<SourceType, TargetType, false, false, true, true> {
4336  static TargetType to(SourceType src) {
4337  KMP_ASSERT(src <= static_cast<SourceType>(
4338  (std::numeric_limits<TargetType>::max)()));
4339  KMP_ASSERT(src >= static_cast<SourceType>(
4340  (std::numeric_limits<TargetType>::min)()));
4341  return (TargetType)src;
4342  }
4343 };
4344 
4345 // Source signed, Target unsigned
4346 // Source smaller
4347 template <typename SourceType, typename TargetType>
4348 struct kmp_convert<SourceType, TargetType, true, false, true, false> {
4349  static TargetType to(SourceType src) {
4350  KMP_ASSERT(src >= 0);
4351  return (TargetType)src;
4352  }
4353 };
4354 // Source equal
4355 template <typename SourceType, typename TargetType>
4356 struct kmp_convert<SourceType, TargetType, false, true, true, false> {
4357  static TargetType to(SourceType src) {
4358  KMP_ASSERT(src >= 0);
4359  return (TargetType)src;
4360  }
4361 };
4362 // Source bigger
4363 template <typename SourceType, typename TargetType>
4364 struct kmp_convert<SourceType, TargetType, false, false, true, false> {
4365  static TargetType to(SourceType src) {
4366  KMP_ASSERT(src >= 0);
4367  KMP_ASSERT(src <= static_cast<SourceType>(
4368  (std::numeric_limits<TargetType>::max)()));
4369  return (TargetType)src;
4370  }
4371 };
4372 
4373 // Source unsigned, Target signed
4374 // Source smaller
4375 template <typename SourceType, typename TargetType>
4376 struct kmp_convert<SourceType, TargetType, true, false, false, true> {
4377  static TargetType to(SourceType src) { return (TargetType)src; }
4378 };
4379 // Source equal
4380 template <typename SourceType, typename TargetType>
4381 struct kmp_convert<SourceType, TargetType, false, true, false, true> {
4382  static TargetType to(SourceType src) {
4383  KMP_ASSERT(src <= static_cast<SourceType>(
4384  (std::numeric_limits<TargetType>::max)()));
4385  return (TargetType)src;
4386  }
4387 };
4388 // Source bigger
4389 template <typename SourceType, typename TargetType>
4390 struct kmp_convert<SourceType, TargetType, false, false, false, true> {
4391  static TargetType to(SourceType src) {
4392  KMP_ASSERT(src <= static_cast<SourceType>(
4393  (std::numeric_limits<TargetType>::max)()));
4394  return (TargetType)src;
4395  }
4396 };
4397 
4398 // Source unsigned, Target unsigned
4399 // Source smaller
4400 template <typename SourceType, typename TargetType>
4401 struct kmp_convert<SourceType, TargetType, true, false, false, false> {
4402  static TargetType to(SourceType src) { return (TargetType)src; }
4403 };
4404 // Source equal
4405 template <typename SourceType, typename TargetType>
4406 struct kmp_convert<SourceType, TargetType, false, true, false, false> {
4407  static TargetType to(SourceType src) { return src; }
4408 };
4409 // Source bigger
4410 template <typename SourceType, typename TargetType>
4411 struct kmp_convert<SourceType, TargetType, false, false, false, false> {
4412  static TargetType to(SourceType src) {
4413  KMP_ASSERT(src <= static_cast<SourceType>(
4414  (std::numeric_limits<TargetType>::max)()));
4415  return (TargetType)src;
4416  }
4417 };
4418 
4419 template <typename T1, typename T2>
4420 static inline void __kmp_type_convert(T1 src, T2 *dest) {
4421  *dest = kmp_convert<T1, T2>::to(src);
4422 }
4423 
4424 #endif /* KMP_H */
void set_stdout()
Definition: kmp.h:4302
void set_stderr()
Definition: kmp.h:4308
int try_open(const char *filename, const char *mode)
Definition: kmp.h:4293
void open(const char *filename, const char *mode, const char *env_var=nullptr)
Definition: kmp.h:4276
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:1561
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:1585
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:1579
void *(* kmpc_ctor_vec)(void *, size_t)
Definition: kmp.h:1602
void *(* kmpc_cctor)(void *, void *)
Definition: kmp.h:1592
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:1614
void(* kmpc_dtor_vec)(void *, size_t)
Definition: kmp.h:1608
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