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 <ctype.h> don't use; problems with /MD on Windows* OS NT due to bad
70  Microsoft library. Some macros provided below to replace these functions */
71 #ifndef __ABSOFT_WIN
72 #include <sys/types.h>
73 #endif
74 #include <limits.h>
75 #include <time.h>
76 
77 #include <errno.h>
78 
79 #include "kmp_os.h"
80 
81 #include "kmp_safe_c_api.h"
82 
83 #if KMP_STATS_ENABLED
84 class kmp_stats_list;
85 #endif
86 
87 #if KMP_USE_HIER_SCHED
88 // Only include hierarchical scheduling if affinity is supported
89 #undef KMP_USE_HIER_SCHED
90 #define KMP_USE_HIER_SCHED KMP_AFFINITY_SUPPORTED
91 #endif
92 
93 #if KMP_USE_HWLOC && KMP_AFFINITY_SUPPORTED
94 #include "hwloc.h"
95 #ifndef HWLOC_OBJ_NUMANODE
96 #define HWLOC_OBJ_NUMANODE HWLOC_OBJ_NODE
97 #endif
98 #ifndef HWLOC_OBJ_PACKAGE
99 #define HWLOC_OBJ_PACKAGE HWLOC_OBJ_SOCKET
100 #endif
101 #if HWLOC_API_VERSION >= 0x00020000
102 // hwloc 2.0 changed type of depth of object from unsigned to int
103 typedef int kmp_hwloc_depth_t;
104 #else
105 typedef unsigned int kmp_hwloc_depth_t;
106 #endif
107 #endif
108 
109 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
110 #include <xmmintrin.h>
111 #endif
112 
113 #include "kmp_debug.h"
114 #include "kmp_lock.h"
115 #include "kmp_version.h"
116 #if USE_DEBUGGER
117 #include "kmp_debugger.h"
118 #endif
119 #include "kmp_i18n.h"
120 
121 #define KMP_HANDLE_SIGNALS (KMP_OS_UNIX || KMP_OS_WINDOWS)
122 
123 #include "kmp_wrapper_malloc.h"
124 #if KMP_OS_UNIX
125 #include <unistd.h>
126 #if !defined NSIG && defined _NSIG
127 #define NSIG _NSIG
128 #endif
129 #endif
130 
131 #if KMP_OS_LINUX
132 #pragma weak clock_gettime
133 #endif
134 
135 #if OMPT_SUPPORT
136 #include "ompt-internal.h"
137 #endif
138 
139 // Affinity format function
140 #include "kmp_str.h"
141 
142 // 0 - no fast memory allocation, alignment: 8-byte on x86, 16-byte on x64.
143 // 3 - fast allocation using sync, non-sync free lists of any size, non-self
144 // free lists of limited size.
145 #ifndef USE_FAST_MEMORY
146 #define USE_FAST_MEMORY 3
147 #endif
148 
149 #ifndef KMP_NESTED_HOT_TEAMS
150 #define KMP_NESTED_HOT_TEAMS 0
151 #define USE_NESTED_HOT_ARG(x)
152 #else
153 #if KMP_NESTED_HOT_TEAMS
154 #define USE_NESTED_HOT_ARG(x) , x
155 #else
156 #define USE_NESTED_HOT_ARG(x)
157 #endif
158 #endif
159 
160 // Assume using BGET compare_exchange instruction instead of lock by default.
161 #ifndef USE_CMP_XCHG_FOR_BGET
162 #define USE_CMP_XCHG_FOR_BGET 1
163 #endif
164 
165 // Test to see if queuing lock is better than bootstrap lock for bget
166 // #ifndef USE_QUEUING_LOCK_FOR_BGET
167 // #define USE_QUEUING_LOCK_FOR_BGET
168 // #endif
169 
170 #define KMP_NSEC_PER_SEC 1000000000L
171 #define KMP_USEC_PER_SEC 1000000L
172 
181 enum {
186  /* 0x04 is no longer used */
195  KMP_IDENT_BARRIER_IMPL_MASK = 0x01C0,
196  KMP_IDENT_BARRIER_IMPL_FOR = 0x0040,
197  KMP_IDENT_BARRIER_IMPL_SECTIONS = 0x00C0,
198 
199  KMP_IDENT_BARRIER_IMPL_SINGLE = 0x0140,
200  KMP_IDENT_BARRIER_IMPL_WORKSHARE = 0x01C0,
201 
213  KMP_IDENT_ATOMIC_HINT_UNCONTENDED = 0x010000,
214  KMP_IDENT_ATOMIC_HINT_CONTENDED = 0x020000,
215  KMP_IDENT_ATOMIC_HINT_NONSPECULATIVE = 0x040000,
216  KMP_IDENT_ATOMIC_HINT_SPECULATIVE = 0x080000,
217 };
218 
222 typedef struct ident {
223  kmp_int32 reserved_1;
224  kmp_int32 flags;
226  kmp_int32 reserved_2;
227 #if USE_ITT_BUILD
228 /* but currently used for storing region-specific ITT */
229 /* contextual information. */
230 #endif /* USE_ITT_BUILD */
231  kmp_int32 reserved_3;
232  char const *psource;
236 } ident_t;
241 // Some forward declarations.
242 typedef union kmp_team kmp_team_t;
243 typedef struct kmp_taskdata kmp_taskdata_t;
244 typedef union kmp_task_team kmp_task_team_t;
245 typedef union kmp_team kmp_team_p;
246 typedef union kmp_info kmp_info_p;
247 typedef union kmp_root kmp_root_p;
248 
249 #ifdef __cplusplus
250 extern "C" {
251 #endif
252 
253 /* ------------------------------------------------------------------------ */
254 
255 /* Pack two 32-bit signed integers into a 64-bit signed integer */
256 /* ToDo: Fix word ordering for big-endian machines. */
257 #define KMP_PACK_64(HIGH_32, LOW_32) \
258  ((kmp_int64)((((kmp_uint64)(HIGH_32)) << 32) | (kmp_uint64)(LOW_32)))
259 
260 // Generic string manipulation macros. Assume that _x is of type char *
261 #define SKIP_WS(_x) \
262  { \
263  while (*(_x) == ' ' || *(_x) == '\t') \
264  (_x)++; \
265  }
266 #define SKIP_DIGITS(_x) \
267  { \
268  while (*(_x) >= '0' && *(_x) <= '9') \
269  (_x)++; \
270  }
271 #define SKIP_TOKEN(_x) \
272  { \
273  while ((*(_x) >= '0' && *(_x) <= '9') || (*(_x) >= 'a' && *(_x) <= 'z') || \
274  (*(_x) >= 'A' && *(_x) <= 'Z') || *(_x) == '_') \
275  (_x)++; \
276  }
277 #define SKIP_TO(_x, _c) \
278  { \
279  while (*(_x) != '\0' && *(_x) != (_c)) \
280  (_x)++; \
281  }
282 
283 /* ------------------------------------------------------------------------ */
284 
285 #define KMP_MAX(x, y) ((x) > (y) ? (x) : (y))
286 #define KMP_MIN(x, y) ((x) < (y) ? (x) : (y))
287 
288 /* ------------------------------------------------------------------------ */
289 /* Enumeration types */
290 
291 enum kmp_state_timer {
292  ts_stop,
293  ts_start,
294  ts_pause,
295 
296  ts_last_state
297 };
298 
299 enum dynamic_mode {
300  dynamic_default,
301 #ifdef USE_LOAD_BALANCE
302  dynamic_load_balance,
303 #endif /* USE_LOAD_BALANCE */
304  dynamic_random,
305  dynamic_thread_limit,
306  dynamic_max
307 };
308 
309 /* external schedule constants, duplicate enum omp_sched in omp.h in order to
310  * not include it here */
311 #ifndef KMP_SCHED_TYPE_DEFINED
312 #define KMP_SCHED_TYPE_DEFINED
313 typedef enum kmp_sched {
314  kmp_sched_lower = 0, // lower and upper bounds are for routine parameter check
315  // Note: need to adjust __kmp_sch_map global array in case enum is changed
316  kmp_sched_static = 1, // mapped to kmp_sch_static_chunked (33)
317  kmp_sched_dynamic = 2, // mapped to kmp_sch_dynamic_chunked (35)
318  kmp_sched_guided = 3, // mapped to kmp_sch_guided_chunked (36)
319  kmp_sched_auto = 4, // mapped to kmp_sch_auto (38)
320  kmp_sched_upper_std = 5, // upper bound for standard schedules
321  kmp_sched_lower_ext = 100, // lower bound of Intel extension schedules
322  kmp_sched_trapezoidal = 101, // mapped to kmp_sch_trapezoidal (39)
323 #if KMP_STATIC_STEAL_ENABLED
324  kmp_sched_static_steal = 102, // mapped to kmp_sch_static_steal (44)
325 #endif
326  kmp_sched_upper,
327  kmp_sched_default = kmp_sched_static, // default scheduling
328  kmp_sched_monotonic = 0x80000000
329 } kmp_sched_t;
330 #endif
331 
336 enum sched_type : kmp_int32 {
338  kmp_sch_static_chunked = 33,
340  kmp_sch_dynamic_chunked = 35,
342  kmp_sch_runtime = 37,
344  kmp_sch_trapezoidal = 39,
345 
346  /* accessible only through KMP_SCHEDULE environment variable */
347  kmp_sch_static_greedy = 40,
348  kmp_sch_static_balanced = 41,
349  /* accessible only through KMP_SCHEDULE environment variable */
350  kmp_sch_guided_iterative_chunked = 42,
351  kmp_sch_guided_analytical_chunked = 43,
352  /* accessible only through KMP_SCHEDULE environment variable */
353  kmp_sch_static_steal = 44,
354 
355  /* static with chunk adjustment (e.g., simd) */
356  kmp_sch_static_balanced_chunked = 45,
360  /* accessible only through KMP_SCHEDULE environment variable */
364  kmp_ord_static_chunked = 65,
366  kmp_ord_dynamic_chunked = 67,
367  kmp_ord_guided_chunked = 68,
368  kmp_ord_runtime = 69,
370  kmp_ord_trapezoidal = 71,
373  /* Schedules for Distribute construct */
377  /* For the "nomerge" versions, kmp_dispatch_next*() will always return a
378  single iteration/chunk, even if the loop is serialized. For the schedule
379  types listed above, the entire iteration vector is returned if the loop is
380  serialized. This doesn't work for gcc/gcomp sections. */
381  kmp_nm_lower = 160,
383  kmp_nm_static_chunked =
384  (kmp_sch_static_chunked - kmp_sch_lower + kmp_nm_lower),
386  kmp_nm_dynamic_chunked = 163,
388  kmp_nm_runtime = 165,
389  kmp_nm_auto = 166,
390  kmp_nm_trapezoidal = 167,
391 
392  /* accessible only through KMP_SCHEDULE environment variable */
393  kmp_nm_static_greedy = 168,
394  kmp_nm_static_balanced = 169,
395  /* accessible only through KMP_SCHEDULE environment variable */
396  kmp_nm_guided_iterative_chunked = 170,
397  kmp_nm_guided_analytical_chunked = 171,
398  kmp_nm_static_steal =
399  172, /* accessible only through OMP_SCHEDULE environment variable */
400 
401  kmp_nm_ord_static_chunked = 193,
403  kmp_nm_ord_dynamic_chunked = 195,
404  kmp_nm_ord_guided_chunked = 196,
405  kmp_nm_ord_runtime = 197,
407  kmp_nm_ord_trapezoidal = 199,
410  /* Support for OpenMP 4.5 monotonic and nonmonotonic schedule modifiers. Since
411  we need to distinguish the three possible cases (no modifier, monotonic
412  modifier, nonmonotonic modifier), we need separate bits for each modifier.
413  The absence of monotonic does not imply nonmonotonic, especially since 4.5
414  says that the behaviour of the "no modifier" case is implementation defined
415  in 4.5, but will become "nonmonotonic" in 5.0.
416 
417  Since we're passing a full 32 bit value, we can use a couple of high bits
418  for these flags; out of paranoia we avoid the sign bit.
419 
420  These modifiers can be or-ed into non-static schedules by the compiler to
421  pass the additional information. They will be stripped early in the
422  processing in __kmp_dispatch_init when setting up schedules, so most of the
423  code won't ever see schedules with these bits set. */
425  (1 << 29),
427  (1 << 30),
429 #define SCHEDULE_WITHOUT_MODIFIERS(s) \
430  (enum sched_type)( \
432 #define SCHEDULE_HAS_MONOTONIC(s) (((s)&kmp_sch_modifier_monotonic) != 0)
433 #define SCHEDULE_HAS_NONMONOTONIC(s) (((s)&kmp_sch_modifier_nonmonotonic) != 0)
434 #define SCHEDULE_HAS_NO_MODIFIERS(s) \
435  (((s) & (kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic)) == 0)
436 #define SCHEDULE_GET_MODIFIERS(s) \
437  ((enum sched_type)( \
438  (s) & (kmp_sch_modifier_nonmonotonic | kmp_sch_modifier_monotonic)))
439 #define SCHEDULE_SET_MODIFIERS(s, m) \
440  (s = (enum sched_type)((kmp_int32)s | (kmp_int32)m))
441 #define SCHEDULE_NONMONOTONIC 0
442 #define SCHEDULE_MONOTONIC 1
443 
445 };
446 
447 // Apply modifiers on internal kind to standard kind
448 static inline void
449 __kmp_sched_apply_mods_stdkind(kmp_sched_t *kind,
450  enum sched_type internal_kind) {
451  if (SCHEDULE_HAS_MONOTONIC(internal_kind)) {
452  *kind = (kmp_sched_t)((int)*kind | (int)kmp_sched_monotonic);
453  }
454 }
455 
456 // Apply modifiers on standard kind to internal kind
457 static inline void
458 __kmp_sched_apply_mods_intkind(kmp_sched_t kind,
459  enum sched_type *internal_kind) {
460  if ((int)kind & (int)kmp_sched_monotonic) {
461  *internal_kind = (enum sched_type)((int)*internal_kind |
463  }
464 }
465 
466 // Get standard schedule without modifiers
467 static inline kmp_sched_t __kmp_sched_without_mods(kmp_sched_t kind) {
468  return (kmp_sched_t)((int)kind & ~((int)kmp_sched_monotonic));
469 }
470 
471 /* Type to keep runtime schedule set via OMP_SCHEDULE or omp_set_schedule() */
472 typedef union kmp_r_sched {
473  struct {
474  enum sched_type r_sched_type;
475  int chunk;
476  };
477  kmp_int64 sched;
478 } kmp_r_sched_t;
479 
480 extern enum sched_type __kmp_sch_map[]; // map OMP 3.0 schedule types with our
481 // internal schedule types
482 
483 enum library_type {
484  library_none,
485  library_serial,
486  library_turnaround,
487  library_throughput
488 };
489 
490 #if KMP_OS_LINUX
491 enum clock_function_type {
492  clock_function_gettimeofday,
493  clock_function_clock_gettime
494 };
495 #endif /* KMP_OS_LINUX */
496 
497 #if KMP_MIC_SUPPORTED
498 enum mic_type { non_mic, mic1, mic2, mic3, dummy };
499 #endif
500 
501 /* -- fast reduction stuff ------------------------------------------------ */
502 
503 #undef KMP_FAST_REDUCTION_BARRIER
504 #define KMP_FAST_REDUCTION_BARRIER 1
505 
506 #undef KMP_FAST_REDUCTION_CORE_DUO
507 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
508 #define KMP_FAST_REDUCTION_CORE_DUO 1
509 #endif
510 
511 enum _reduction_method {
512  reduction_method_not_defined = 0,
513  critical_reduce_block = (1 << 8),
514  atomic_reduce_block = (2 << 8),
515  tree_reduce_block = (3 << 8),
516  empty_reduce_block = (4 << 8)
517 };
518 
519 // Description of the packed_reduction_method variable:
520 // The packed_reduction_method variable consists of two enum types variables
521 // that are packed together into 0-th byte and 1-st byte:
522 // 0: (packed_reduction_method & 0x000000FF) is a 'enum barrier_type' value of
523 // barrier that will be used in fast reduction: bs_plain_barrier or
524 // bs_reduction_barrier
525 // 1: (packed_reduction_method & 0x0000FF00) is a reduction method that will
526 // be used in fast reduction;
527 // Reduction method is of 'enum _reduction_method' type and it's defined the way
528 // so that the bits of 0-th byte are empty, so no need to execute a shift
529 // instruction while packing/unpacking
530 
531 #if KMP_FAST_REDUCTION_BARRIER
532 #define PACK_REDUCTION_METHOD_AND_BARRIER(reduction_method, barrier_type) \
533  ((reduction_method) | (barrier_type))
534 
535 #define UNPACK_REDUCTION_METHOD(packed_reduction_method) \
536  ((enum _reduction_method)((packed_reduction_method) & (0x0000FF00)))
537 
538 #define UNPACK_REDUCTION_BARRIER(packed_reduction_method) \
539  ((enum barrier_type)((packed_reduction_method) & (0x000000FF)))
540 #else
541 #define PACK_REDUCTION_METHOD_AND_BARRIER(reduction_method, barrier_type) \
542  (reduction_method)
543 
544 #define UNPACK_REDUCTION_METHOD(packed_reduction_method) \
545  (packed_reduction_method)
546 
547 #define UNPACK_REDUCTION_BARRIER(packed_reduction_method) (bs_plain_barrier)
548 #endif
549 
550 #define TEST_REDUCTION_METHOD(packed_reduction_method, which_reduction_block) \
551  ((UNPACK_REDUCTION_METHOD(packed_reduction_method)) == \
552  (which_reduction_block))
553 
554 #if KMP_FAST_REDUCTION_BARRIER
555 #define TREE_REDUCE_BLOCK_WITH_REDUCTION_BARRIER \
556  (PACK_REDUCTION_METHOD_AND_BARRIER(tree_reduce_block, bs_reduction_barrier))
557 
558 #define TREE_REDUCE_BLOCK_WITH_PLAIN_BARRIER \
559  (PACK_REDUCTION_METHOD_AND_BARRIER(tree_reduce_block, bs_plain_barrier))
560 #endif
561 
562 typedef int PACKED_REDUCTION_METHOD_T;
563 
564 /* -- end of fast reduction stuff ----------------------------------------- */
565 
566 #if KMP_OS_WINDOWS
567 #define USE_CBLKDATA
568 #if KMP_MSVC_COMPAT
569 #pragma warning(push)
570 #pragma warning(disable : 271 310)
571 #endif
572 #include <windows.h>
573 #if KMP_MSVC_COMPAT
574 #pragma warning(pop)
575 #endif
576 #endif
577 
578 #if KMP_OS_UNIX
579 #include <dlfcn.h>
580 #include <pthread.h>
581 #endif
582 
583 /* Only Linux* OS and Windows* OS support thread affinity. */
584 #if KMP_AFFINITY_SUPPORTED
585 
586 // GROUP_AFFINITY is already defined for _MSC_VER>=1600 (VS2010 and later).
587 #if KMP_OS_WINDOWS
588 #if _MSC_VER < 1600 && KMP_MSVC_COMPAT
589 typedef struct GROUP_AFFINITY {
590  KAFFINITY Mask;
591  WORD Group;
592  WORD Reserved[3];
593 } GROUP_AFFINITY;
594 #endif /* _MSC_VER < 1600 */
595 #if KMP_GROUP_AFFINITY
596 extern int __kmp_num_proc_groups;
597 #else
598 static const int __kmp_num_proc_groups = 1;
599 #endif /* KMP_GROUP_AFFINITY */
600 typedef DWORD (*kmp_GetActiveProcessorCount_t)(WORD);
601 extern kmp_GetActiveProcessorCount_t __kmp_GetActiveProcessorCount;
602 
603 typedef WORD (*kmp_GetActiveProcessorGroupCount_t)(void);
604 extern kmp_GetActiveProcessorGroupCount_t __kmp_GetActiveProcessorGroupCount;
605 
606 typedef BOOL (*kmp_GetThreadGroupAffinity_t)(HANDLE, GROUP_AFFINITY *);
607 extern kmp_GetThreadGroupAffinity_t __kmp_GetThreadGroupAffinity;
608 
609 typedef BOOL (*kmp_SetThreadGroupAffinity_t)(HANDLE, const GROUP_AFFINITY *,
610  GROUP_AFFINITY *);
611 extern kmp_SetThreadGroupAffinity_t __kmp_SetThreadGroupAffinity;
612 #endif /* KMP_OS_WINDOWS */
613 
614 #if KMP_USE_HWLOC
615 extern hwloc_topology_t __kmp_hwloc_topology;
616 extern int __kmp_hwloc_error;
617 extern int __kmp_numa_detected;
618 extern int __kmp_tile_depth;
619 #endif
620 
621 extern size_t __kmp_affin_mask_size;
622 #define KMP_AFFINITY_CAPABLE() (__kmp_affin_mask_size > 0)
623 #define KMP_AFFINITY_DISABLE() (__kmp_affin_mask_size = 0)
624 #define KMP_AFFINITY_ENABLE(mask_size) (__kmp_affin_mask_size = mask_size)
625 #define KMP_CPU_SET_ITERATE(i, mask) \
626  for (i = (mask)->begin(); (int)i != (mask)->end(); i = (mask)->next(i))
627 #define KMP_CPU_SET(i, mask) (mask)->set(i)
628 #define KMP_CPU_ISSET(i, mask) (mask)->is_set(i)
629 #define KMP_CPU_CLR(i, mask) (mask)->clear(i)
630 #define KMP_CPU_ZERO(mask) (mask)->zero()
631 #define KMP_CPU_COPY(dest, src) (dest)->copy(src)
632 #define KMP_CPU_AND(dest, src) (dest)->bitwise_and(src)
633 #define KMP_CPU_COMPLEMENT(max_bit_number, mask) (mask)->bitwise_not()
634 #define KMP_CPU_UNION(dest, src) (dest)->bitwise_or(src)
635 #define KMP_CPU_ALLOC(ptr) (ptr = __kmp_affinity_dispatch->allocate_mask())
636 #define KMP_CPU_FREE(ptr) __kmp_affinity_dispatch->deallocate_mask(ptr)
637 #define KMP_CPU_ALLOC_ON_STACK(ptr) KMP_CPU_ALLOC(ptr)
638 #define KMP_CPU_FREE_FROM_STACK(ptr) KMP_CPU_FREE(ptr)
639 #define KMP_CPU_INTERNAL_ALLOC(ptr) KMP_CPU_ALLOC(ptr)
640 #define KMP_CPU_INTERNAL_FREE(ptr) KMP_CPU_FREE(ptr)
641 #define KMP_CPU_INDEX(arr, i) __kmp_affinity_dispatch->index_mask_array(arr, i)
642 #define KMP_CPU_ALLOC_ARRAY(arr, n) \
643  (arr = __kmp_affinity_dispatch->allocate_mask_array(n))
644 #define KMP_CPU_FREE_ARRAY(arr, n) \
645  __kmp_affinity_dispatch->deallocate_mask_array(arr)
646 #define KMP_CPU_INTERNAL_ALLOC_ARRAY(arr, n) KMP_CPU_ALLOC_ARRAY(arr, n)
647 #define KMP_CPU_INTERNAL_FREE_ARRAY(arr, n) KMP_CPU_FREE_ARRAY(arr, n)
648 #define __kmp_get_system_affinity(mask, abort_bool) \
649  (mask)->get_system_affinity(abort_bool)
650 #define __kmp_set_system_affinity(mask, abort_bool) \
651  (mask)->set_system_affinity(abort_bool)
652 #define __kmp_get_proc_group(mask) (mask)->get_proc_group()
653 
654 class KMPAffinity {
655 public:
656  class Mask {
657  public:
658  void *operator new(size_t n);
659  void operator delete(void *p);
660  void *operator new[](size_t n);
661  void operator delete[](void *p);
662  virtual ~Mask() {}
663  // Set bit i to 1
664  virtual void set(int i) {}
665  // Return bit i
666  virtual bool is_set(int i) const { return false; }
667  // Set bit i to 0
668  virtual void clear(int i) {}
669  // Zero out entire mask
670  virtual void zero() {}
671  // Copy src into this mask
672  virtual void copy(const Mask *src) {}
673  // this &= rhs
674  virtual void bitwise_and(const Mask *rhs) {}
675  // this |= rhs
676  virtual void bitwise_or(const Mask *rhs) {}
677  // this = ~this
678  virtual void bitwise_not() {}
679  // API for iterating over an affinity mask
680  // for (int i = mask->begin(); i != mask->end(); i = mask->next(i))
681  virtual int begin() const { return 0; }
682  virtual int end() const { return 0; }
683  virtual int next(int previous) const { return 0; }
684  // Set the system's affinity to this affinity mask's value
685  virtual int set_system_affinity(bool abort_on_error) const { return -1; }
686  // Set this affinity mask to the current system affinity
687  virtual int get_system_affinity(bool abort_on_error) { return -1; }
688  // Only 1 DWORD in the mask should have any procs set.
689  // Return the appropriate index, or -1 for an invalid mask.
690  virtual int get_proc_group() const { return -1; }
691  };
692  void *operator new(size_t n);
693  void operator delete(void *p);
694  // Need virtual destructor
695  virtual ~KMPAffinity() = default;
696  // Determine if affinity is capable
697  virtual void determine_capable(const char *env_var) {}
698  // Bind the current thread to os proc
699  virtual void bind_thread(int proc) {}
700  // Factory functions to allocate/deallocate a mask
701  virtual Mask *allocate_mask() { return nullptr; }
702  virtual void deallocate_mask(Mask *m) {}
703  virtual Mask *allocate_mask_array(int num) { return nullptr; }
704  virtual void deallocate_mask_array(Mask *m) {}
705  virtual Mask *index_mask_array(Mask *m, int index) { return nullptr; }
706  static void pick_api();
707  static void destroy_api();
708  enum api_type {
709  NATIVE_OS
710 #if KMP_USE_HWLOC
711  ,
712  HWLOC
713 #endif
714  };
715  virtual api_type get_api_type() const {
716  KMP_ASSERT(0);
717  return NATIVE_OS;
718  }
719 
720 private:
721  static bool picked_api;
722 };
723 
724 typedef KMPAffinity::Mask kmp_affin_mask_t;
725 extern KMPAffinity *__kmp_affinity_dispatch;
726 
727 // Declare local char buffers with this size for printing debug and info
728 // messages, using __kmp_affinity_print_mask().
729 #define KMP_AFFIN_MASK_PRINT_LEN 1024
730 
731 enum affinity_type {
732  affinity_none = 0,
733  affinity_physical,
734  affinity_logical,
735  affinity_compact,
736  affinity_scatter,
737  affinity_explicit,
738  affinity_balanced,
739  affinity_disabled, // not used outsize the env var parser
740  affinity_default
741 };
742 
743 enum affinity_gran {
744  affinity_gran_fine = 0,
745  affinity_gran_thread,
746  affinity_gran_core,
747  affinity_gran_tile,
748  affinity_gran_numa,
749  affinity_gran_package,
750  affinity_gran_node,
751 #if KMP_GROUP_AFFINITY
752  // The "group" granularity isn't necesssarily coarser than all of the
753  // other levels, but we put it last in the enum.
754  affinity_gran_group,
755 #endif /* KMP_GROUP_AFFINITY */
756  affinity_gran_default
757 };
758 
759 enum affinity_top_method {
760  affinity_top_method_all = 0, // try all (supported) methods, in order
761 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
762  affinity_top_method_apicid,
763  affinity_top_method_x2apicid,
764 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
765  affinity_top_method_cpuinfo, // KMP_CPUINFO_FILE is usable on Windows* OS, too
766 #if KMP_GROUP_AFFINITY
767  affinity_top_method_group,
768 #endif /* KMP_GROUP_AFFINITY */
769  affinity_top_method_flat,
770 #if KMP_USE_HWLOC
771  affinity_top_method_hwloc,
772 #endif
773  affinity_top_method_default
774 };
775 
776 #define affinity_respect_mask_default (-1)
777 
778 extern enum affinity_type __kmp_affinity_type; /* Affinity type */
779 extern enum affinity_gran __kmp_affinity_gran; /* Affinity granularity */
780 extern int __kmp_affinity_gran_levels; /* corresponding int value */
781 extern int __kmp_affinity_dups; /* Affinity duplicate masks */
782 extern enum affinity_top_method __kmp_affinity_top_method;
783 extern int __kmp_affinity_compact; /* Affinity 'compact' value */
784 extern int __kmp_affinity_offset; /* Affinity offset value */
785 extern int __kmp_affinity_verbose; /* Was verbose specified for KMP_AFFINITY? */
786 extern int __kmp_affinity_warnings; /* KMP_AFFINITY warnings enabled ? */
787 extern int __kmp_affinity_respect_mask; // Respect process' init affinity mask?
788 extern char *__kmp_affinity_proclist; /* proc ID list */
789 extern kmp_affin_mask_t *__kmp_affinity_masks;
790 extern unsigned __kmp_affinity_num_masks;
791 extern void __kmp_affinity_bind_thread(int which);
792 
793 extern kmp_affin_mask_t *__kmp_affin_fullMask;
794 extern char *__kmp_cpuinfo_file;
795 
796 #endif /* KMP_AFFINITY_SUPPORTED */
797 
798 // This needs to be kept in sync with the values in omp.h !!!
799 typedef enum kmp_proc_bind_t {
800  proc_bind_false = 0,
801  proc_bind_true,
802  proc_bind_master,
803  proc_bind_close,
804  proc_bind_spread,
805  proc_bind_intel, // use KMP_AFFINITY interface
806  proc_bind_default
807 } kmp_proc_bind_t;
808 
809 typedef struct kmp_nested_proc_bind_t {
810  kmp_proc_bind_t *bind_types;
811  int size;
812  int used;
813 } kmp_nested_proc_bind_t;
814 
815 extern kmp_nested_proc_bind_t __kmp_nested_proc_bind;
816 
817 extern int __kmp_display_affinity;
818 extern char *__kmp_affinity_format;
819 static const size_t KMP_AFFINITY_FORMAT_SIZE = 512;
820 
821 #if KMP_AFFINITY_SUPPORTED
822 #define KMP_PLACE_ALL (-1)
823 #define KMP_PLACE_UNDEFINED (-2)
824 // Is KMP_AFFINITY is being used instead of OMP_PROC_BIND/OMP_PLACES?
825 #define KMP_AFFINITY_NON_PROC_BIND \
826  ((__kmp_nested_proc_bind.bind_types[0] == proc_bind_false || \
827  __kmp_nested_proc_bind.bind_types[0] == proc_bind_intel) && \
828  (__kmp_affinity_num_masks > 0 || __kmp_affinity_type == affinity_balanced))
829 #endif /* KMP_AFFINITY_SUPPORTED */
830 
831 extern int __kmp_affinity_num_places;
832 
833 typedef enum kmp_cancel_kind_t {
834  cancel_noreq = 0,
835  cancel_parallel = 1,
836  cancel_loop = 2,
837  cancel_sections = 3,
838  cancel_taskgroup = 4
839 } kmp_cancel_kind_t;
840 
841 // KMP_HW_SUBSET support:
842 typedef struct kmp_hws_item {
843  int num;
844  int offset;
845 } kmp_hws_item_t;
846 
847 extern kmp_hws_item_t __kmp_hws_socket;
848 extern kmp_hws_item_t __kmp_hws_node;
849 extern kmp_hws_item_t __kmp_hws_tile;
850 extern kmp_hws_item_t __kmp_hws_core;
851 extern kmp_hws_item_t __kmp_hws_proc;
852 extern int __kmp_hws_requested;
853 extern int __kmp_hws_abs_flag; // absolute or per-item number requested
854 
855 /* ------------------------------------------------------------------------ */
856 
857 #define KMP_PAD(type, sz) \
858  (sizeof(type) + (sz - ((sizeof(type) - 1) % (sz)) - 1))
859 
860 // We need to avoid using -1 as a GTID as +1 is added to the gtid
861 // when storing it in a lock, and the value 0 is reserved.
862 #define KMP_GTID_DNE (-2) /* Does not exist */
863 #define KMP_GTID_SHUTDOWN (-3) /* Library is shutting down */
864 #define KMP_GTID_MONITOR (-4) /* Monitor thread ID */
865 #define KMP_GTID_UNKNOWN (-5) /* Is not known */
866 #define KMP_GTID_MIN (-6) /* Minimal gtid for low bound check in DEBUG */
867 
868 /* OpenMP 5.0 Memory Management support */
869 
870 #ifndef __OMP_H
871 // Duplicate type definitios from omp.h
872 typedef uintptr_t omp_uintptr_t;
873 
874 typedef enum {
875  OMP_ATK_THREADMODEL = 1,
876  OMP_ATK_ALIGNMENT = 2,
877  OMP_ATK_ACCESS = 3,
878  OMP_ATK_POOL_SIZE = 4,
879  OMP_ATK_FALLBACK = 5,
880  OMP_ATK_FB_DATA = 6,
881  OMP_ATK_PINNED = 7,
882  OMP_ATK_PARTITION = 8
883 } omp_alloctrait_key_t;
884 
885 typedef enum {
886  OMP_ATV_FALSE = 0,
887  OMP_ATV_TRUE = 1,
888  OMP_ATV_DEFAULT = 2,
889  OMP_ATV_CONTENDED = 3,
890  OMP_ATV_UNCONTENDED = 4,
891  OMP_ATV_SEQUENTIAL = 5,
892  OMP_ATV_PRIVATE = 6,
893  OMP_ATV_ALL = 7,
894  OMP_ATV_THREAD = 8,
895  OMP_ATV_PTEAM = 9,
896  OMP_ATV_CGROUP = 10,
897  OMP_ATV_DEFAULT_MEM_FB = 11,
898  OMP_ATV_NULL_FB = 12,
899  OMP_ATV_ABORT_FB = 13,
900  OMP_ATV_ALLOCATOR_FB = 14,
901  OMP_ATV_ENVIRONMENT = 15,
902  OMP_ATV_NEAREST = 16,
903  OMP_ATV_BLOCKED = 17,
904  OMP_ATV_INTERLEAVED = 18
905 } omp_alloctrait_value_t;
906 
907 typedef void *omp_memspace_handle_t;
908 extern omp_memspace_handle_t const omp_default_mem_space;
909 extern omp_memspace_handle_t const omp_large_cap_mem_space;
910 extern omp_memspace_handle_t const omp_const_mem_space;
911 extern omp_memspace_handle_t const omp_high_bw_mem_space;
912 extern omp_memspace_handle_t const omp_low_lat_mem_space;
913 
914 typedef struct {
915  omp_alloctrait_key_t key;
916  omp_uintptr_t value;
917 } omp_alloctrait_t;
918 
919 typedef void *omp_allocator_handle_t;
920 extern omp_allocator_handle_t const omp_null_allocator;
921 extern omp_allocator_handle_t const omp_default_mem_alloc;
922 extern omp_allocator_handle_t const omp_large_cap_mem_alloc;
923 extern omp_allocator_handle_t const omp_const_mem_alloc;
924 extern omp_allocator_handle_t const omp_high_bw_mem_alloc;
925 extern omp_allocator_handle_t const omp_low_lat_mem_alloc;
926 extern omp_allocator_handle_t const omp_cgroup_mem_alloc;
927 extern omp_allocator_handle_t const omp_pteam_mem_alloc;
928 extern omp_allocator_handle_t const omp_thread_mem_alloc;
929 extern omp_allocator_handle_t const kmp_max_mem_alloc;
930 extern omp_allocator_handle_t __kmp_def_allocator;
931 
932 // end of duplicate type definitios from omp.h
933 #endif
934 
935 extern int __kmp_memkind_available;
936 
937 typedef omp_memspace_handle_t kmp_memspace_t; // placeholder
938 
939 typedef struct kmp_allocator_t {
940  omp_memspace_handle_t memspace;
941  void **memkind; // pointer to memkind
942  int alignment;
943  omp_alloctrait_value_t fb;
944  kmp_allocator_t *fb_data;
945  kmp_uint64 pool_size;
946  kmp_uint64 pool_used;
947 } kmp_allocator_t;
948 
949 extern omp_allocator_handle_t __kmpc_init_allocator(int gtid,
950  omp_memspace_handle_t,
951  int ntraits,
952  omp_alloctrait_t traits[]);
953 extern void __kmpc_destroy_allocator(int gtid, omp_allocator_handle_t al);
954 extern void __kmpc_set_default_allocator(int gtid, omp_allocator_handle_t al);
955 extern omp_allocator_handle_t __kmpc_get_default_allocator(int gtid);
956 extern void *__kmpc_alloc(int gtid, size_t sz, omp_allocator_handle_t al);
957 extern void __kmpc_free(int gtid, void *ptr, omp_allocator_handle_t al);
958 
959 extern void __kmp_init_memkind();
960 extern void __kmp_fini_memkind();
961 
962 /* ------------------------------------------------------------------------ */
963 
964 #define KMP_UINT64_MAX \
965  (~((kmp_uint64)1 << ((sizeof(kmp_uint64) * (1 << 3)) - 1)))
966 
967 #define KMP_MIN_NTH 1
968 
969 #ifndef KMP_MAX_NTH
970 #if defined(PTHREAD_THREADS_MAX) && PTHREAD_THREADS_MAX < INT_MAX
971 #define KMP_MAX_NTH PTHREAD_THREADS_MAX
972 #else
973 #define KMP_MAX_NTH INT_MAX
974 #endif
975 #endif /* KMP_MAX_NTH */
976 
977 #ifdef PTHREAD_STACK_MIN
978 #define KMP_MIN_STKSIZE PTHREAD_STACK_MIN
979 #else
980 #define KMP_MIN_STKSIZE ((size_t)(32 * 1024))
981 #endif
982 
983 #define KMP_MAX_STKSIZE (~((size_t)1 << ((sizeof(size_t) * (1 << 3)) - 1)))
984 
985 #if KMP_ARCH_X86
986 #define KMP_DEFAULT_STKSIZE ((size_t)(2 * 1024 * 1024))
987 #elif KMP_ARCH_X86_64
988 #define KMP_DEFAULT_STKSIZE ((size_t)(4 * 1024 * 1024))
989 #define KMP_BACKUP_STKSIZE ((size_t)(2 * 1024 * 1024))
990 #else
991 #define KMP_DEFAULT_STKSIZE ((size_t)(1024 * 1024))
992 #endif
993 
994 #define KMP_DEFAULT_MALLOC_POOL_INCR ((size_t)(1024 * 1024))
995 #define KMP_MIN_MALLOC_POOL_INCR ((size_t)(4 * 1024))
996 #define KMP_MAX_MALLOC_POOL_INCR \
997  (~((size_t)1 << ((sizeof(size_t) * (1 << 3)) - 1)))
998 
999 #define KMP_MIN_STKOFFSET (0)
1000 #define KMP_MAX_STKOFFSET KMP_MAX_STKSIZE
1001 #if KMP_OS_DARWIN
1002 #define KMP_DEFAULT_STKOFFSET KMP_MIN_STKOFFSET
1003 #else
1004 #define KMP_DEFAULT_STKOFFSET CACHE_LINE
1005 #endif
1006 
1007 #define KMP_MIN_STKPADDING (0)
1008 #define KMP_MAX_STKPADDING (2 * 1024 * 1024)
1009 
1010 #define KMP_BLOCKTIME_MULTIPLIER \
1011  (1000) /* number of blocktime units per second */
1012 #define KMP_MIN_BLOCKTIME (0)
1013 #define KMP_MAX_BLOCKTIME \
1014  (INT_MAX) /* Must be this for "infinite" setting the work */
1015 #define KMP_DEFAULT_BLOCKTIME (200) /* __kmp_blocktime is in milliseconds */
1016 
1017 #if KMP_USE_MONITOR
1018 #define KMP_DEFAULT_MONITOR_STKSIZE ((size_t)(64 * 1024))
1019 #define KMP_MIN_MONITOR_WAKEUPS (1) // min times monitor wakes up per second
1020 #define KMP_MAX_MONITOR_WAKEUPS (1000) // max times monitor can wake up per sec
1021 
1022 /* Calculate new number of monitor wakeups for a specific block time based on
1023  previous monitor_wakeups. Only allow increasing number of wakeups */
1024 #define KMP_WAKEUPS_FROM_BLOCKTIME(blocktime, monitor_wakeups) \
1025  (((blocktime) == KMP_MAX_BLOCKTIME) \
1026  ? (monitor_wakeups) \
1027  : ((blocktime) == KMP_MIN_BLOCKTIME) \
1028  ? KMP_MAX_MONITOR_WAKEUPS \
1029  : ((monitor_wakeups) > (KMP_BLOCKTIME_MULTIPLIER / (blocktime))) \
1030  ? (monitor_wakeups) \
1031  : (KMP_BLOCKTIME_MULTIPLIER) / (blocktime))
1032 
1033 /* Calculate number of intervals for a specific block time based on
1034  monitor_wakeups */
1035 #define KMP_INTERVALS_FROM_BLOCKTIME(blocktime, monitor_wakeups) \
1036  (((blocktime) + (KMP_BLOCKTIME_MULTIPLIER / (monitor_wakeups)) - 1) / \
1037  (KMP_BLOCKTIME_MULTIPLIER / (monitor_wakeups)))
1038 #else
1039 #define KMP_BLOCKTIME(team, tid) \
1040  (get__bt_set(team, tid) ? get__blocktime(team, tid) : __kmp_dflt_blocktime)
1041 #if KMP_OS_UNIX && (KMP_ARCH_X86 || KMP_ARCH_X86_64)
1042 // HW TSC is used to reduce overhead (clock tick instead of nanosecond).
1043 extern kmp_uint64 __kmp_ticks_per_msec;
1044 #if KMP_COMPILER_ICC
1045 #define KMP_NOW() ((kmp_uint64)_rdtsc())
1046 #else
1047 #define KMP_NOW() __kmp_hardware_timestamp()
1048 #endif
1049 #define KMP_NOW_MSEC() (KMP_NOW() / __kmp_ticks_per_msec)
1050 #define KMP_BLOCKTIME_INTERVAL(team, tid) \
1051  (KMP_BLOCKTIME(team, tid) * __kmp_ticks_per_msec)
1052 #define KMP_BLOCKING(goal, count) ((goal) > KMP_NOW())
1053 #else
1054 // System time is retrieved sporadically while blocking.
1055 extern kmp_uint64 __kmp_now_nsec();
1056 #define KMP_NOW() __kmp_now_nsec()
1057 #define KMP_NOW_MSEC() (KMP_NOW() / KMP_USEC_PER_SEC)
1058 #define KMP_BLOCKTIME_INTERVAL(team, tid) \
1059  (KMP_BLOCKTIME(team, tid) * KMP_USEC_PER_SEC)
1060 #define KMP_BLOCKING(goal, count) ((count) % 1000 != 0 || (goal) > KMP_NOW())
1061 #endif
1062 #endif // KMP_USE_MONITOR
1063 
1064 #define KMP_MIN_STATSCOLS 40
1065 #define KMP_MAX_STATSCOLS 4096
1066 #define KMP_DEFAULT_STATSCOLS 80
1067 
1068 #define KMP_MIN_INTERVAL 0
1069 #define KMP_MAX_INTERVAL (INT_MAX - 1)
1070 #define KMP_DEFAULT_INTERVAL 0
1071 
1072 #define KMP_MIN_CHUNK 1
1073 #define KMP_MAX_CHUNK (INT_MAX - 1)
1074 #define KMP_DEFAULT_CHUNK 1
1075 
1076 #define KMP_DFLT_DISP_NUM_BUFF 7
1077 #define KMP_MAX_ORDERED 8
1078 
1079 #define KMP_MAX_FIELDS 32
1080 
1081 #define KMP_MAX_BRANCH_BITS 31
1082 
1083 #define KMP_MAX_ACTIVE_LEVELS_LIMIT INT_MAX
1084 
1085 #define KMP_MAX_DEFAULT_DEVICE_LIMIT INT_MAX
1086 
1087 #define KMP_MAX_TASK_PRIORITY_LIMIT INT_MAX
1088 
1089 /* Minimum number of threads before switch to TLS gtid (experimentally
1090  determined) */
1091 /* josh TODO: what about OS X* tuning? */
1092 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
1093 #define KMP_TLS_GTID_MIN 5
1094 #else
1095 #define KMP_TLS_GTID_MIN INT_MAX
1096 #endif
1097 
1098 #define KMP_MASTER_TID(tid) ((tid) == 0)
1099 #define KMP_WORKER_TID(tid) ((tid) != 0)
1100 
1101 #define KMP_MASTER_GTID(gtid) (__kmp_tid_from_gtid((gtid)) == 0)
1102 #define KMP_WORKER_GTID(gtid) (__kmp_tid_from_gtid((gtid)) != 0)
1103 #define KMP_INITIAL_GTID(gtid) ((gtid) == 0)
1104 
1105 #ifndef TRUE
1106 #define FALSE 0
1107 #define TRUE (!FALSE)
1108 #endif
1109 
1110 /* NOTE: all of the following constants must be even */
1111 
1112 #if KMP_OS_WINDOWS
1113 #define KMP_INIT_WAIT 64U /* initial number of spin-tests */
1114 #define KMP_NEXT_WAIT 32U /* susequent number of spin-tests */
1115 #elif KMP_OS_CNK
1116 #define KMP_INIT_WAIT 16U /* initial number of spin-tests */
1117 #define KMP_NEXT_WAIT 8U /* susequent number of spin-tests */
1118 #elif KMP_OS_LINUX
1119 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1120 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1121 #elif KMP_OS_DARWIN
1122 /* TODO: tune for KMP_OS_DARWIN */
1123 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1124 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1125 #elif KMP_OS_DRAGONFLY
1126 /* TODO: tune for KMP_OS_DRAGONFLY */
1127 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1128 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1129 #elif KMP_OS_FREEBSD
1130 /* TODO: tune for KMP_OS_FREEBSD */
1131 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1132 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1133 #elif KMP_OS_NETBSD
1134 /* TODO: tune for KMP_OS_NETBSD */
1135 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1136 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1137 #elif KMP_OS_HURD
1138 /* TODO: tune for KMP_OS_HURD */
1139 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1140 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1141 #elif KMP_OS_OPENBSD
1142 /* TODO: tune for KMP_OS_OPENBSD */
1143 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1144 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1145 #elif KMP_OS_KFREEBSD
1146 /* TODO: tune for KMP_OS_KFREEBSD */
1147 #define KMP_INIT_WAIT 1024U /* initial number of spin-tests */
1148 #define KMP_NEXT_WAIT 512U /* susequent number of spin-tests */
1149 #endif
1150 
1151 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
1152 typedef struct kmp_cpuid {
1153  kmp_uint32 eax;
1154  kmp_uint32 ebx;
1155  kmp_uint32 ecx;
1156  kmp_uint32 edx;
1157 } kmp_cpuid_t;
1158 
1159 typedef struct kmp_cpuinfo {
1160  int initialized; // If 0, other fields are not initialized.
1161  int signature; // CPUID(1).EAX
1162  int family; // CPUID(1).EAX[27:20]+CPUID(1).EAX[11:8] (Extended Family+Family)
1163  int model; // ( CPUID(1).EAX[19:16] << 4 ) + CPUID(1).EAX[7:4] ( ( Extended
1164  // Model << 4 ) + Model)
1165  int stepping; // CPUID(1).EAX[3:0] ( Stepping )
1166  int sse2; // 0 if SSE2 instructions are not supported, 1 otherwise.
1167  int rtm; // 0 if RTM instructions are not supported, 1 otherwise.
1168  int cpu_stackoffset;
1169  int apic_id;
1170  int physical_id;
1171  int logical_id;
1172  kmp_uint64 frequency; // Nominal CPU frequency in Hz.
1173  char name[3 * sizeof(kmp_cpuid_t)]; // CPUID(0x80000002,0x80000003,0x80000004)
1174 } kmp_cpuinfo_t;
1175 
1176 extern void __kmp_query_cpuid(kmp_cpuinfo_t *p);
1177 
1178 #if KMP_OS_UNIX
1179 // subleaf is only needed for cache and topology discovery and can be set to
1180 // zero in most cases
1181 static inline void __kmp_x86_cpuid(int leaf, int subleaf, struct kmp_cpuid *p) {
1182  __asm__ __volatile__("cpuid"
1183  : "=a"(p->eax), "=b"(p->ebx), "=c"(p->ecx), "=d"(p->edx)
1184  : "a"(leaf), "c"(subleaf));
1185 }
1186 // Load p into FPU control word
1187 static inline void __kmp_load_x87_fpu_control_word(const kmp_int16 *p) {
1188  __asm__ __volatile__("fldcw %0" : : "m"(*p));
1189 }
1190 // Store FPU control word into p
1191 static inline void __kmp_store_x87_fpu_control_word(kmp_int16 *p) {
1192  __asm__ __volatile__("fstcw %0" : "=m"(*p));
1193 }
1194 static inline void __kmp_clear_x87_fpu_status_word() {
1195 #if KMP_MIC
1196  // 32-bit protected mode x87 FPU state
1197  struct x87_fpu_state {
1198  unsigned cw;
1199  unsigned sw;
1200  unsigned tw;
1201  unsigned fip;
1202  unsigned fips;
1203  unsigned fdp;
1204  unsigned fds;
1205  };
1206  struct x87_fpu_state fpu_state = {0, 0, 0, 0, 0, 0, 0};
1207  __asm__ __volatile__("fstenv %0\n\t" // store FP env
1208  "andw $0x7f00, %1\n\t" // clear 0-7,15 bits of FP SW
1209  "fldenv %0\n\t" // load FP env back
1210  : "+m"(fpu_state), "+m"(fpu_state.sw));
1211 #else
1212  __asm__ __volatile__("fnclex");
1213 #endif // KMP_MIC
1214 }
1215 #if __SSE__
1216 static inline void __kmp_load_mxcsr(const kmp_uint32 *p) { _mm_setcsr(*p); }
1217 static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = _mm_getcsr(); }
1218 #else
1219 static inline void __kmp_load_mxcsr(const kmp_uint32 *p) {}
1220 static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = 0; }
1221 #endif
1222 #else
1223 // Windows still has these as external functions in assembly file
1224 extern void __kmp_x86_cpuid(int mode, int mode2, struct kmp_cpuid *p);
1225 extern void __kmp_load_x87_fpu_control_word(const kmp_int16 *p);
1226 extern void __kmp_store_x87_fpu_control_word(kmp_int16 *p);
1227 extern void __kmp_clear_x87_fpu_status_word();
1228 static inline void __kmp_load_mxcsr(const kmp_uint32 *p) { _mm_setcsr(*p); }
1229 static inline void __kmp_store_mxcsr(kmp_uint32 *p) { *p = _mm_getcsr(); }
1230 #endif // KMP_OS_UNIX
1231 
1232 #define KMP_X86_MXCSR_MASK 0xffffffc0 /* ignore status flags (6 lsb) */
1233 
1234 #if KMP_ARCH_X86
1235 extern void __kmp_x86_pause(void);
1236 #elif KMP_MIC
1237 // Performance testing on KNC (C0QS-7120 P/A/X/D, 61-core, 16 GB Memory) showed
1238 // regression after removal of extra PAUSE from spin loops. Changing
1239 // the delay from 100 to 300 showed even better performance than double PAUSE
1240 // on Spec OMP2001 and LCPC tasking tests, no regressions on EPCC.
1241 static inline void __kmp_x86_pause(void) { _mm_delay_32(300); }
1242 #else
1243 static inline void __kmp_x86_pause(void) { _mm_pause(); }
1244 #endif
1245 #define KMP_CPU_PAUSE() __kmp_x86_pause()
1246 #elif KMP_ARCH_PPC64
1247 #define KMP_PPC64_PRI_LOW() __asm__ volatile("or 1, 1, 1")
1248 #define KMP_PPC64_PRI_MED() __asm__ volatile("or 2, 2, 2")
1249 #define KMP_PPC64_PRI_LOC_MB() __asm__ volatile("" : : : "memory")
1250 #define KMP_CPU_PAUSE() \
1251  do { \
1252  KMP_PPC64_PRI_LOW(); \
1253  KMP_PPC64_PRI_MED(); \
1254  KMP_PPC64_PRI_LOC_MB(); \
1255  } while (0)
1256 #else
1257 #define KMP_CPU_PAUSE() /* nothing to do */
1258 #endif
1259 
1260 #define KMP_INIT_YIELD(count) \
1261  { (count) = __kmp_yield_init; }
1262 
1263 #define KMP_OVERSUBSCRIBED \
1264  (TCR_4(__kmp_nth) > (__kmp_avail_proc ? __kmp_avail_proc : __kmp_xproc))
1265 
1266 #define KMP_TRY_YIELD \
1267  ((__kmp_use_yield == 1) || (__kmp_use_yield == 2 && (KMP_OVERSUBSCRIBED)))
1268 
1269 #define KMP_TRY_YIELD_OVERSUB \
1270  ((__kmp_use_yield == 1 || __kmp_use_yield == 2) && (KMP_OVERSUBSCRIBED))
1271 
1272 #define KMP_YIELD(cond) \
1273  { \
1274  KMP_CPU_PAUSE(); \
1275  if ((cond) && (KMP_TRY_YIELD)) \
1276  __kmp_yield(); \
1277  }
1278 
1279 #define KMP_YIELD_OVERSUB() \
1280  { \
1281  KMP_CPU_PAUSE(); \
1282  if ((KMP_TRY_YIELD_OVERSUB)) \
1283  __kmp_yield(); \
1284  }
1285 
1286 // Note the decrement of 2 in the following Macros. With KMP_LIBRARY=turnaround,
1287 // there should be no yielding since initial value from KMP_INIT_YIELD() is odd.
1288 #define KMP_YIELD_SPIN(count) \
1289  { \
1290  KMP_CPU_PAUSE(); \
1291  if (KMP_TRY_YIELD) { \
1292  (count) -= 2; \
1293  if (!(count)) { \
1294  __kmp_yield(); \
1295  (count) = __kmp_yield_next; \
1296  } \
1297  } \
1298  }
1299 
1300 #define KMP_YIELD_OVERSUB_ELSE_SPIN(count) \
1301  { \
1302  KMP_CPU_PAUSE(); \
1303  if ((KMP_TRY_YIELD_OVERSUB)) \
1304  __kmp_yield(); \
1305  else if (__kmp_use_yield == 1) { \
1306  (count) -= 2; \
1307  if (!(count)) { \
1308  __kmp_yield(); \
1309  (count) = __kmp_yield_next; \
1310  } \
1311  } \
1312  }
1313 
1314 /* ------------------------------------------------------------------------ */
1315 /* Support datatypes for the orphaned construct nesting checks. */
1316 /* ------------------------------------------------------------------------ */
1317 
1318 enum cons_type {
1319  ct_none,
1320  ct_parallel,
1321  ct_pdo,
1322  ct_pdo_ordered,
1323  ct_psections,
1324  ct_psingle,
1325  ct_critical,
1326  ct_ordered_in_parallel,
1327  ct_ordered_in_pdo,
1328  ct_master,
1329  ct_reduce,
1330  ct_barrier
1331 };
1332 
1333 #define IS_CONS_TYPE_ORDERED(ct) ((ct) == ct_pdo_ordered)
1334 
1335 struct cons_data {
1336  ident_t const *ident;
1337  enum cons_type type;
1338  int prev;
1339  kmp_user_lock_p
1340  name; /* address exclusively for critical section name comparison */
1341 };
1342 
1343 struct cons_header {
1344  int p_top, w_top, s_top;
1345  int stack_size, stack_top;
1346  struct cons_data *stack_data;
1347 };
1348 
1349 struct kmp_region_info {
1350  char *text;
1351  int offset[KMP_MAX_FIELDS];
1352  int length[KMP_MAX_FIELDS];
1353 };
1354 
1355 /* ---------------------------------------------------------------------- */
1356 /* ---------------------------------------------------------------------- */
1357 
1358 #if KMP_OS_WINDOWS
1359 typedef HANDLE kmp_thread_t;
1360 typedef DWORD kmp_key_t;
1361 #endif /* KMP_OS_WINDOWS */
1362 
1363 #if KMP_OS_UNIX
1364 typedef pthread_t kmp_thread_t;
1365 typedef pthread_key_t kmp_key_t;
1366 #endif
1367 
1368 extern kmp_key_t __kmp_gtid_threadprivate_key;
1369 
1370 typedef struct kmp_sys_info {
1371  long maxrss; /* the maximum resident set size utilized (in kilobytes) */
1372  long minflt; /* the number of page faults serviced without any I/O */
1373  long majflt; /* the number of page faults serviced that required I/O */
1374  long nswap; /* the number of times a process was "swapped" out of memory */
1375  long inblock; /* the number of times the file system had to perform input */
1376  long oublock; /* the number of times the file system had to perform output */
1377  long nvcsw; /* the number of times a context switch was voluntarily */
1378  long nivcsw; /* the number of times a context switch was forced */
1379 } kmp_sys_info_t;
1380 
1381 #if USE_ITT_BUILD
1382 // We cannot include "kmp_itt.h" due to circular dependency. Declare the only
1383 // required type here. Later we will check the type meets requirements.
1384 typedef int kmp_itt_mark_t;
1385 #define KMP_ITT_DEBUG 0
1386 #endif /* USE_ITT_BUILD */
1387 
1388 typedef kmp_int32 kmp_critical_name[8];
1389 
1399 typedef void (*kmpc_micro)(kmp_int32 *global_tid, kmp_int32 *bound_tid, ...);
1400 typedef void (*kmpc_micro_bound)(kmp_int32 *bound_tid, kmp_int32 *bound_nth,
1401  ...);
1402 
1407 /* ---------------------------------------------------------------------------
1408  */
1409 /* Threadprivate initialization/finalization function declarations */
1410 
1411 /* for non-array objects: __kmpc_threadprivate_register() */
1412 
1417 typedef void *(*kmpc_ctor)(void *);
1418 
1423 typedef void (*kmpc_dtor)(
1424  void * /*, size_t */); /* 2nd arg: magic number for KCC unused by Intel
1425  compiler */
1430 typedef void *(*kmpc_cctor)(void *, void *);
1431 
1432 /* for array objects: __kmpc_threadprivate_register_vec() */
1433 /* First arg: "this" pointer */
1434 /* Last arg: number of array elements */
1440 typedef void *(*kmpc_ctor_vec)(void *, size_t);
1446 typedef void (*kmpc_dtor_vec)(void *, size_t);
1452 typedef void *(*kmpc_cctor_vec)(void *, void *,
1453  size_t); /* function unused by compiler */
1454 
1459 /* keeps tracked of threadprivate cache allocations for cleanup later */
1460 typedef struct kmp_cached_addr {
1461  void **addr; /* address of allocated cache */
1462  void ***compiler_cache; /* pointer to compiler's cache */
1463  void *data; /* pointer to global data */
1464  struct kmp_cached_addr *next; /* pointer to next cached address */
1465 } kmp_cached_addr_t;
1466 
1467 struct private_data {
1468  struct private_data *next; /* The next descriptor in the list */
1469  void *data; /* The data buffer for this descriptor */
1470  int more; /* The repeat count for this descriptor */
1471  size_t size; /* The data size for this descriptor */
1472 };
1473 
1474 struct private_common {
1475  struct private_common *next;
1476  struct private_common *link;
1477  void *gbl_addr;
1478  void *par_addr; /* par_addr == gbl_addr for MASTER thread */
1479  size_t cmn_size;
1480 };
1481 
1482 struct shared_common {
1483  struct shared_common *next;
1484  struct private_data *pod_init;
1485  void *obj_init;
1486  void *gbl_addr;
1487  union {
1488  kmpc_ctor ctor;
1489  kmpc_ctor_vec ctorv;
1490  } ct;
1491  union {
1492  kmpc_cctor cctor;
1493  kmpc_cctor_vec cctorv;
1494  } cct;
1495  union {
1496  kmpc_dtor dtor;
1497  kmpc_dtor_vec dtorv;
1498  } dt;
1499  size_t vec_len;
1500  int is_vec;
1501  size_t cmn_size;
1502 };
1503 
1504 #define KMP_HASH_TABLE_LOG2 9 /* log2 of the hash table size */
1505 #define KMP_HASH_TABLE_SIZE \
1506  (1 << KMP_HASH_TABLE_LOG2) /* size of the hash table */
1507 #define KMP_HASH_SHIFT 3 /* throw away this many low bits from the address */
1508 #define KMP_HASH(x) \
1509  ((((kmp_uintptr_t)x) >> KMP_HASH_SHIFT) & (KMP_HASH_TABLE_SIZE - 1))
1510 
1511 struct common_table {
1512  struct private_common *data[KMP_HASH_TABLE_SIZE];
1513 };
1514 
1515 struct shared_table {
1516  struct shared_common *data[KMP_HASH_TABLE_SIZE];
1517 };
1518 
1519 /* ------------------------------------------------------------------------ */
1520 
1521 #if KMP_USE_HIER_SCHED
1522 // Shared barrier data that exists inside a single unit of the scheduling
1523 // hierarchy
1524 typedef struct kmp_hier_private_bdata_t {
1525  kmp_int32 num_active;
1526  kmp_uint64 index;
1527  kmp_uint64 wait_val[2];
1528 } kmp_hier_private_bdata_t;
1529 #endif
1530 
1531 typedef struct kmp_sched_flags {
1532  unsigned ordered : 1;
1533  unsigned nomerge : 1;
1534  unsigned contains_last : 1;
1535 #if KMP_USE_HIER_SCHED
1536  unsigned use_hier : 1;
1537  unsigned unused : 28;
1538 #else
1539  unsigned unused : 29;
1540 #endif
1541 } kmp_sched_flags_t;
1542 
1543 KMP_BUILD_ASSERT(sizeof(kmp_sched_flags_t) == 4);
1544 
1545 #if KMP_STATIC_STEAL_ENABLED
1546 typedef struct KMP_ALIGN_CACHE dispatch_private_info32 {
1547  kmp_int32 count;
1548  kmp_int32 ub;
1549  /* Adding KMP_ALIGN_CACHE here doesn't help / can hurt performance */
1550  kmp_int32 lb;
1551  kmp_int32 st;
1552  kmp_int32 tc;
1553  kmp_int32 static_steal_counter; /* for static_steal only; maybe better to put
1554  after ub */
1555 
1556  // KMP_ALIGN( 16 ) ensures ( if the KMP_ALIGN macro is turned on )
1557  // a) parm3 is properly aligned and
1558  // b) all parm1-4 are in the same cache line.
1559  // Because of parm1-4 are used together, performance seems to be better
1560  // if they are in the same line (not measured though).
1561 
1562  struct KMP_ALIGN(32) { // AC: changed 16 to 32 in order to simplify template
1563  kmp_int32 parm1; // structures in kmp_dispatch.cpp. This should
1564  kmp_int32 parm2; // make no real change at least while padding is off.
1565  kmp_int32 parm3;
1566  kmp_int32 parm4;
1567  };
1568 
1569  kmp_uint32 ordered_lower;
1570  kmp_uint32 ordered_upper;
1571 #if KMP_OS_WINDOWS
1572  // This var can be placed in the hole between 'tc' and 'parm1', instead of
1573  // 'static_steal_counter'. It would be nice to measure execution times.
1574  // Conditional if/endif can be removed at all.
1575  kmp_int32 last_upper;
1576 #endif /* KMP_OS_WINDOWS */
1577 } dispatch_private_info32_t;
1578 
1579 typedef struct KMP_ALIGN_CACHE dispatch_private_info64 {
1580  kmp_int64 count; // current chunk number for static & static-steal scheduling
1581  kmp_int64 ub; /* upper-bound */
1582  /* Adding KMP_ALIGN_CACHE here doesn't help / can hurt performance */
1583  kmp_int64 lb; /* lower-bound */
1584  kmp_int64 st; /* stride */
1585  kmp_int64 tc; /* trip count (number of iterations) */
1586  kmp_int64 static_steal_counter; /* for static_steal only; maybe better to put
1587  after ub */
1588 
1589  /* parm[1-4] are used in different ways by different scheduling algorithms */
1590 
1591  // KMP_ALIGN( 32 ) ensures ( if the KMP_ALIGN macro is turned on )
1592  // a) parm3 is properly aligned and
1593  // b) all parm1-4 are in the same cache line.
1594  // Because of parm1-4 are used together, performance seems to be better
1595  // if they are in the same line (not measured though).
1596 
1597  struct KMP_ALIGN(32) {
1598  kmp_int64 parm1;
1599  kmp_int64 parm2;
1600  kmp_int64 parm3;
1601  kmp_int64 parm4;
1602  };
1603 
1604  kmp_uint64 ordered_lower;
1605  kmp_uint64 ordered_upper;
1606 #if KMP_OS_WINDOWS
1607  // This var can be placed in the hole between 'tc' and 'parm1', instead of
1608  // 'static_steal_counter'. It would be nice to measure execution times.
1609  // Conditional if/endif can be removed at all.
1610  kmp_int64 last_upper;
1611 #endif /* KMP_OS_WINDOWS */
1612 } dispatch_private_info64_t;
1613 #else /* KMP_STATIC_STEAL_ENABLED */
1614 typedef struct KMP_ALIGN_CACHE dispatch_private_info32 {
1615  kmp_int32 lb;
1616  kmp_int32 ub;
1617  kmp_int32 st;
1618  kmp_int32 tc;
1619 
1620  kmp_int32 parm1;
1621  kmp_int32 parm2;
1622  kmp_int32 parm3;
1623  kmp_int32 parm4;
1624 
1625  kmp_int32 count;
1626 
1627  kmp_uint32 ordered_lower;
1628  kmp_uint32 ordered_upper;
1629 #if KMP_OS_WINDOWS
1630  kmp_int32 last_upper;
1631 #endif /* KMP_OS_WINDOWS */
1632 } dispatch_private_info32_t;
1633 
1634 typedef struct KMP_ALIGN_CACHE dispatch_private_info64 {
1635  kmp_int64 lb; /* lower-bound */
1636  kmp_int64 ub; /* upper-bound */
1637  kmp_int64 st; /* stride */
1638  kmp_int64 tc; /* trip count (number of iterations) */
1639 
1640  /* parm[1-4] are used in different ways by different scheduling algorithms */
1641  kmp_int64 parm1;
1642  kmp_int64 parm2;
1643  kmp_int64 parm3;
1644  kmp_int64 parm4;
1645 
1646  kmp_int64 count; /* current chunk number for static scheduling */
1647 
1648  kmp_uint64 ordered_lower;
1649  kmp_uint64 ordered_upper;
1650 #if KMP_OS_WINDOWS
1651  kmp_int64 last_upper;
1652 #endif /* KMP_OS_WINDOWS */
1653 } dispatch_private_info64_t;
1654 #endif /* KMP_STATIC_STEAL_ENABLED */
1655 
1656 typedef struct KMP_ALIGN_CACHE dispatch_private_info {
1657  union private_info {
1658  dispatch_private_info32_t p32;
1659  dispatch_private_info64_t p64;
1660  } u;
1661  enum sched_type schedule; /* scheduling algorithm */
1662  kmp_sched_flags_t flags; /* flags (e.g., ordered, nomerge, etc.) */
1663  kmp_int32 ordered_bumped;
1664  // To retain the structure size after making ordered_iteration scalar
1665  kmp_int32 ordered_dummy[KMP_MAX_ORDERED - 3];
1666  // Stack of buffers for nest of serial regions
1667  struct dispatch_private_info *next;
1668  kmp_int32 type_size; /* the size of types in private_info */
1669 #if KMP_USE_HIER_SCHED
1670  kmp_int32 hier_id;
1671  void *parent; /* hierarchical scheduling parent pointer */
1672 #endif
1673  enum cons_type pushed_ws;
1674 } dispatch_private_info_t;
1675 
1676 typedef struct dispatch_shared_info32 {
1677  /* chunk index under dynamic, number of idle threads under static-steal;
1678  iteration index otherwise */
1679  volatile kmp_uint32 iteration;
1680  volatile kmp_uint32 num_done;
1681  volatile kmp_uint32 ordered_iteration;
1682  // Dummy to retain the structure size after making ordered_iteration scalar
1683  kmp_int32 ordered_dummy[KMP_MAX_ORDERED - 1];
1684 } dispatch_shared_info32_t;
1685 
1686 typedef struct dispatch_shared_info64 {
1687  /* chunk index under dynamic, number of idle threads under static-steal;
1688  iteration index otherwise */
1689  volatile kmp_uint64 iteration;
1690  volatile kmp_uint64 num_done;
1691  volatile kmp_uint64 ordered_iteration;
1692  // Dummy to retain the structure size after making ordered_iteration scalar
1693  kmp_int64 ordered_dummy[KMP_MAX_ORDERED - 3];
1694 } dispatch_shared_info64_t;
1695 
1696 typedef struct dispatch_shared_info {
1697  union shared_info {
1698  dispatch_shared_info32_t s32;
1699  dispatch_shared_info64_t s64;
1700  } u;
1701  volatile kmp_uint32 buffer_index;
1702  volatile kmp_int32 doacross_buf_idx; // teamwise index
1703  volatile kmp_uint32 *doacross_flags; // shared array of iteration flags (0/1)
1704  kmp_int32 doacross_num_done; // count finished threads
1705 #if KMP_USE_HIER_SCHED
1706  void *hier;
1707 #endif
1708 #if KMP_USE_HWLOC
1709  // When linking with libhwloc, the ORDERED EPCC test slows down on big
1710  // machines (> 48 cores). Performance analysis showed that a cache thrash
1711  // was occurring and this padding helps alleviate the problem.
1712  char padding[64];
1713 #endif
1714 } dispatch_shared_info_t;
1715 
1716 typedef struct kmp_disp {
1717  /* Vector for ORDERED SECTION */
1718  void (*th_deo_fcn)(int *gtid, int *cid, ident_t *);
1719  /* Vector for END ORDERED SECTION */
1720  void (*th_dxo_fcn)(int *gtid, int *cid, ident_t *);
1721 
1722  dispatch_shared_info_t *th_dispatch_sh_current;
1723  dispatch_private_info_t *th_dispatch_pr_current;
1724 
1725  dispatch_private_info_t *th_disp_buffer;
1726  kmp_int32 th_disp_index;
1727  kmp_int32 th_doacross_buf_idx; // thread's doacross buffer index
1728  volatile kmp_uint32 *th_doacross_flags; // pointer to shared array of flags
1729  union { // we can use union here because doacross cannot be used in
1730  // nonmonotonic loops
1731  kmp_int64 *th_doacross_info; // info on loop bounds
1732  kmp_lock_t *th_steal_lock; // lock used for chunk stealing (8-byte variable)
1733  };
1734 #if KMP_USE_INTERNODE_ALIGNMENT
1735  char more_padding[INTERNODE_CACHE_LINE];
1736 #endif
1737 } kmp_disp_t;
1738 
1739 /* ------------------------------------------------------------------------ */
1740 /* Barrier stuff */
1741 
1742 /* constants for barrier state update */
1743 #define KMP_INIT_BARRIER_STATE 0 /* should probably start from zero */
1744 #define KMP_BARRIER_SLEEP_BIT 0 /* bit used for suspend/sleep part of state */
1745 #define KMP_BARRIER_UNUSED_BIT 1 // bit that must never be set for valid state
1746 #define KMP_BARRIER_BUMP_BIT 2 /* lsb used for bump of go/arrived state */
1747 
1748 #define KMP_BARRIER_SLEEP_STATE (1 << KMP_BARRIER_SLEEP_BIT)
1749 #define KMP_BARRIER_UNUSED_STATE (1 << KMP_BARRIER_UNUSED_BIT)
1750 #define KMP_BARRIER_STATE_BUMP (1 << KMP_BARRIER_BUMP_BIT)
1751 
1752 #if (KMP_BARRIER_SLEEP_BIT >= KMP_BARRIER_BUMP_BIT)
1753 #error "Barrier sleep bit must be smaller than barrier bump bit"
1754 #endif
1755 #if (KMP_BARRIER_UNUSED_BIT >= KMP_BARRIER_BUMP_BIT)
1756 #error "Barrier unused bit must be smaller than barrier bump bit"
1757 #endif
1758 
1759 // Constants for release barrier wait state: currently, hierarchical only
1760 #define KMP_BARRIER_NOT_WAITING 0 // Normal state; worker not in wait_sleep
1761 #define KMP_BARRIER_OWN_FLAG \
1762  1 // Normal state; worker waiting on own b_go flag in release
1763 #define KMP_BARRIER_PARENT_FLAG \
1764  2 // Special state; worker waiting on parent's b_go flag in release
1765 #define KMP_BARRIER_SWITCH_TO_OWN_FLAG \
1766  3 // Special state; tells worker to shift from parent to own b_go
1767 #define KMP_BARRIER_SWITCHING \
1768  4 // Special state; worker resets appropriate flag on wake-up
1769 
1770 #define KMP_NOT_SAFE_TO_REAP \
1771  0 // Thread th_reap_state: not safe to reap (tasking)
1772 #define KMP_SAFE_TO_REAP 1 // Thread th_reap_state: safe to reap (not tasking)
1773 
1774 enum barrier_type {
1775  bs_plain_barrier = 0, /* 0, All non-fork/join barriers (except reduction
1776  barriers if enabled) */
1777  bs_forkjoin_barrier, /* 1, All fork/join (parallel region) barriers */
1778 #if KMP_FAST_REDUCTION_BARRIER
1779  bs_reduction_barrier, /* 2, All barriers that are used in reduction */
1780 #endif // KMP_FAST_REDUCTION_BARRIER
1781  bs_last_barrier /* Just a placeholder to mark the end */
1782 };
1783 
1784 // to work with reduction barriers just like with plain barriers
1785 #if !KMP_FAST_REDUCTION_BARRIER
1786 #define bs_reduction_barrier bs_plain_barrier
1787 #endif // KMP_FAST_REDUCTION_BARRIER
1788 
1789 typedef enum kmp_bar_pat { /* Barrier communication patterns */
1790  bp_linear_bar =
1791  0, /* Single level (degenerate) tree */
1792  bp_tree_bar =
1793  1, /* Balanced tree with branching factor 2^n */
1794  bp_hyper_bar =
1795  2, /* Hypercube-embedded tree with min branching
1796  factor 2^n */
1797  bp_hierarchical_bar = 3, /* Machine hierarchy tree */
1798  bp_last_bar /* Placeholder to mark the end */
1799 } kmp_bar_pat_e;
1800 
1801 #define KMP_BARRIER_ICV_PUSH 1
1802 
1803 /* Record for holding the values of the internal controls stack records */
1804 typedef struct kmp_internal_control {
1805  int serial_nesting_level; /* corresponds to the value of the
1806  th_team_serialized field */
1807  kmp_int8 dynamic; /* internal control for dynamic adjustment of threads (per
1808  thread) */
1809  kmp_int8
1810  bt_set; /* internal control for whether blocktime is explicitly set */
1811  int blocktime; /* internal control for blocktime */
1812 #if KMP_USE_MONITOR
1813  int bt_intervals; /* internal control for blocktime intervals */
1814 #endif
1815  int nproc; /* internal control for #threads for next parallel region (per
1816  thread) */
1817  int thread_limit; /* internal control for thread-limit-var */
1818  int max_active_levels; /* internal control for max_active_levels */
1819  kmp_r_sched_t
1820  sched; /* internal control for runtime schedule {sched,chunk} pair */
1821  kmp_proc_bind_t proc_bind; /* internal control for affinity */
1822  kmp_int32 default_device; /* internal control for default device */
1823  struct kmp_internal_control *next;
1824 } kmp_internal_control_t;
1825 
1826 static inline void copy_icvs(kmp_internal_control_t *dst,
1827  kmp_internal_control_t *src) {
1828  *dst = *src;
1829 }
1830 
1831 /* Thread barrier needs volatile barrier fields */
1832 typedef struct KMP_ALIGN_CACHE kmp_bstate {
1833  // th_fixed_icvs is aligned by virtue of kmp_bstate being aligned (and all
1834  // uses of it). It is not explicitly aligned below, because we *don't* want
1835  // it to be padded -- instead, we fit b_go into the same cache line with
1836  // th_fixed_icvs, enabling NGO cache lines stores in the hierarchical barrier.
1837  kmp_internal_control_t th_fixed_icvs; // Initial ICVs for the thread
1838  // Tuck b_go into end of th_fixed_icvs cache line, so it can be stored with
1839  // same NGO store
1840  volatile kmp_uint64 b_go; // STATE => task should proceed (hierarchical)
1841  KMP_ALIGN_CACHE volatile kmp_uint64
1842  b_arrived; // STATE => task reached synch point.
1843  kmp_uint32 *skip_per_level;
1844  kmp_uint32 my_level;
1845  kmp_int32 parent_tid;
1846  kmp_int32 old_tid;
1847  kmp_uint32 depth;
1848  struct kmp_bstate *parent_bar;
1849  kmp_team_t *team;
1850  kmp_uint64 leaf_state;
1851  kmp_uint32 nproc;
1852  kmp_uint8 base_leaf_kids;
1853  kmp_uint8 leaf_kids;
1854  kmp_uint8 offset;
1855  kmp_uint8 wait_flag;
1856  kmp_uint8 use_oncore_barrier;
1857 #if USE_DEBUGGER
1858  // The following field is intended for the debugger solely. Only the worker
1859  // thread itself accesses this field: the worker increases it by 1 when it
1860  // arrives to a barrier.
1861  KMP_ALIGN_CACHE kmp_uint b_worker_arrived;
1862 #endif /* USE_DEBUGGER */
1863 } kmp_bstate_t;
1864 
1865 union KMP_ALIGN_CACHE kmp_barrier_union {
1866  double b_align; /* use worst case alignment */
1867  char b_pad[KMP_PAD(kmp_bstate_t, CACHE_LINE)];
1868  kmp_bstate_t bb;
1869 };
1870 
1871 typedef union kmp_barrier_union kmp_balign_t;
1872 
1873 /* Team barrier needs only non-volatile arrived counter */
1874 union KMP_ALIGN_CACHE kmp_barrier_team_union {
1875  double b_align; /* use worst case alignment */
1876  char b_pad[CACHE_LINE];
1877  struct {
1878  kmp_uint64 b_arrived; /* STATE => task reached synch point. */
1879 #if USE_DEBUGGER
1880  // The following two fields are indended for the debugger solely. Only
1881  // master of the team accesses these fields: the first one is increased by
1882  // 1 when master arrives to a barrier, the second one is increased by one
1883  // when all the threads arrived.
1884  kmp_uint b_master_arrived;
1885  kmp_uint b_team_arrived;
1886 #endif
1887  };
1888 };
1889 
1890 typedef union kmp_barrier_team_union kmp_balign_team_t;
1891 
1892 /* Padding for Linux* OS pthreads condition variables and mutexes used to signal
1893  threads when a condition changes. This is to workaround an NPTL bug where
1894  padding was added to pthread_cond_t which caused the initialization routine
1895  to write outside of the structure if compiled on pre-NPTL threads. */
1896 #if KMP_OS_WINDOWS
1897 typedef struct kmp_win32_mutex {
1898  /* The Lock */
1899  CRITICAL_SECTION cs;
1900 } kmp_win32_mutex_t;
1901 
1902 typedef struct kmp_win32_cond {
1903  /* Count of the number of waiters. */
1904  int waiters_count_;
1905 
1906  /* Serialize access to <waiters_count_> */
1907  kmp_win32_mutex_t waiters_count_lock_;
1908 
1909  /* Number of threads to release via a <cond_broadcast> or a <cond_signal> */
1910  int release_count_;
1911 
1912  /* Keeps track of the current "generation" so that we don't allow */
1913  /* one thread to steal all the "releases" from the broadcast. */
1914  int wait_generation_count_;
1915 
1916  /* A manual-reset event that's used to block and release waiting threads. */
1917  HANDLE event_;
1918 } kmp_win32_cond_t;
1919 #endif
1920 
1921 #if KMP_OS_UNIX
1922 
1923 union KMP_ALIGN_CACHE kmp_cond_union {
1924  double c_align;
1925  char c_pad[CACHE_LINE];
1926  pthread_cond_t c_cond;
1927 };
1928 
1929 typedef union kmp_cond_union kmp_cond_align_t;
1930 
1931 union KMP_ALIGN_CACHE kmp_mutex_union {
1932  double m_align;
1933  char m_pad[CACHE_LINE];
1934  pthread_mutex_t m_mutex;
1935 };
1936 
1937 typedef union kmp_mutex_union kmp_mutex_align_t;
1938 
1939 #endif /* KMP_OS_UNIX */
1940 
1941 typedef struct kmp_desc_base {
1942  void *ds_stackbase;
1943  size_t ds_stacksize;
1944  int ds_stackgrow;
1945  kmp_thread_t ds_thread;
1946  volatile int ds_tid;
1947  int ds_gtid;
1948 #if KMP_OS_WINDOWS
1949  volatile int ds_alive;
1950  DWORD ds_thread_id;
1951 /* ds_thread keeps thread handle on Windows* OS. It is enough for RTL purposes.
1952  However, debugger support (libomp_db) cannot work with handles, because they
1953  uncomparable. For example, debugger requests info about thread with handle h.
1954  h is valid within debugger process, and meaningless within debugee process.
1955  Even if h is duped by call to DuplicateHandle(), so the result h' is valid
1956  within debugee process, but it is a *new* handle which does *not* equal to
1957  any other handle in debugee... The only way to compare handles is convert
1958  them to system-wide ids. GetThreadId() function is available only in
1959  Longhorn and Server 2003. :-( In contrast, GetCurrentThreadId() is available
1960  on all Windows* OS flavours (including Windows* 95). Thus, we have to get
1961  thread id by call to GetCurrentThreadId() from within the thread and save it
1962  to let libomp_db identify threads. */
1963 #endif /* KMP_OS_WINDOWS */
1964 } kmp_desc_base_t;
1965 
1966 typedef union KMP_ALIGN_CACHE kmp_desc {
1967  double ds_align; /* use worst case alignment */
1968  char ds_pad[KMP_PAD(kmp_desc_base_t, CACHE_LINE)];
1969  kmp_desc_base_t ds;
1970 } kmp_desc_t;
1971 
1972 typedef struct kmp_local {
1973  volatile int this_construct; /* count of single's encountered by thread */
1974  void *reduce_data;
1975 #if KMP_USE_BGET
1976  void *bget_data;
1977  void *bget_list;
1978 #if !USE_CMP_XCHG_FOR_BGET
1979 #ifdef USE_QUEUING_LOCK_FOR_BGET
1980  kmp_lock_t bget_lock; /* Lock for accessing bget free list */
1981 #else
1982  kmp_bootstrap_lock_t bget_lock; // Lock for accessing bget free list. Must be
1983 // bootstrap lock so we can use it at library
1984 // shutdown.
1985 #endif /* USE_LOCK_FOR_BGET */
1986 #endif /* ! USE_CMP_XCHG_FOR_BGET */
1987 #endif /* KMP_USE_BGET */
1988 
1989  PACKED_REDUCTION_METHOD_T
1990  packed_reduction_method; /* stored by __kmpc_reduce*(), used by
1991  __kmpc_end_reduce*() */
1992 
1993 } kmp_local_t;
1994 
1995 #define KMP_CHECK_UPDATE(a, b) \
1996  if ((a) != (b)) \
1997  (a) = (b)
1998 #define KMP_CHECK_UPDATE_SYNC(a, b) \
1999  if ((a) != (b)) \
2000  TCW_SYNC_PTR((a), (b))
2001 
2002 #define get__blocktime(xteam, xtid) \
2003  ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.blocktime)
2004 #define get__bt_set(xteam, xtid) \
2005  ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_set)
2006 #if KMP_USE_MONITOR
2007 #define get__bt_intervals(xteam, xtid) \
2008  ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_intervals)
2009 #endif
2010 
2011 #define get__dynamic_2(xteam, xtid) \
2012  ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.dynamic)
2013 #define get__nproc_2(xteam, xtid) \
2014  ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.nproc)
2015 #define get__sched_2(xteam, xtid) \
2016  ((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.sched)
2017 
2018 #define set__blocktime_team(xteam, xtid, xval) \
2019  (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.blocktime) = \
2020  (xval))
2021 
2022 #if KMP_USE_MONITOR
2023 #define set__bt_intervals_team(xteam, xtid, xval) \
2024  (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_intervals) = \
2025  (xval))
2026 #endif
2027 
2028 #define set__bt_set_team(xteam, xtid, xval) \
2029  (((xteam)->t.t_threads[(xtid)]->th.th_current_task->td_icvs.bt_set) = (xval))
2030 
2031 #define set__dynamic(xthread, xval) \
2032  (((xthread)->th.th_current_task->td_icvs.dynamic) = (xval))
2033 #define get__dynamic(xthread) \
2034  (((xthread)->th.th_current_task->td_icvs.dynamic) ? (FTN_TRUE) : (FTN_FALSE))
2035 
2036 #define set__nproc(xthread, xval) \
2037  (((xthread)->th.th_current_task->td_icvs.nproc) = (xval))
2038 
2039 #define set__thread_limit(xthread, xval) \
2040  (((xthread)->th.th_current_task->td_icvs.thread_limit) = (xval))
2041 
2042 #define set__max_active_levels(xthread, xval) \
2043  (((xthread)->th.th_current_task->td_icvs.max_active_levels) = (xval))
2044 
2045 #define get__max_active_levels(xthread) \
2046  ((xthread)->th.th_current_task->td_icvs.max_active_levels)
2047 
2048 #define set__sched(xthread, xval) \
2049  (((xthread)->th.th_current_task->td_icvs.sched) = (xval))
2050 
2051 #define set__proc_bind(xthread, xval) \
2052  (((xthread)->th.th_current_task->td_icvs.proc_bind) = (xval))
2053 #define get__proc_bind(xthread) \
2054  ((xthread)->th.th_current_task->td_icvs.proc_bind)
2055 
2056 // OpenMP tasking data structures
2057 
2058 typedef enum kmp_tasking_mode {
2059  tskm_immediate_exec = 0,
2060  tskm_extra_barrier = 1,
2061  tskm_task_teams = 2,
2062  tskm_max = 2
2063 } kmp_tasking_mode_t;
2064 
2065 extern kmp_tasking_mode_t
2066  __kmp_tasking_mode; /* determines how/when to execute tasks */
2067 extern int __kmp_task_stealing_constraint;
2068 extern int __kmp_enable_task_throttling;
2069 extern kmp_int32 __kmp_default_device; // Set via OMP_DEFAULT_DEVICE if
2070 // specified, defaults to 0 otherwise
2071 // Set via OMP_MAX_TASK_PRIORITY if specified, defaults to 0 otherwise
2072 extern kmp_int32 __kmp_max_task_priority;
2073 // Set via KMP_TASKLOOP_MIN_TASKS if specified, defaults to 0 otherwise
2074 extern kmp_uint64 __kmp_taskloop_min_tasks;
2075 
2076 /* NOTE: kmp_taskdata_t and kmp_task_t structures allocated in single block with
2077  taskdata first */
2078 #define KMP_TASK_TO_TASKDATA(task) (((kmp_taskdata_t *)task) - 1)
2079 #define KMP_TASKDATA_TO_TASK(taskdata) (kmp_task_t *)(taskdata + 1)
2080 
2081 // The tt_found_tasks flag is a signal to all threads in the team that tasks
2082 // were spawned and queued since the previous barrier release.
2083 #define KMP_TASKING_ENABLED(task_team) \
2084  (TCR_SYNC_4((task_team)->tt.tt_found_tasks) == TRUE)
2085 
2092 typedef kmp_int32 (*kmp_routine_entry_t)(kmp_int32, void *);
2093 
2094 typedef union kmp_cmplrdata {
2095  kmp_int32 priority;
2096  kmp_routine_entry_t
2097  destructors; /* pointer to function to invoke deconstructors of
2098  firstprivate C++ objects */
2099  /* future data */
2100 } kmp_cmplrdata_t;
2101 
2102 /* sizeof_kmp_task_t passed as arg to kmpc_omp_task call */
2105 typedef struct kmp_task { /* GEH: Shouldn't this be aligned somehow? */
2106  void *shareds;
2107  kmp_routine_entry_t
2108  routine;
2109  kmp_int32 part_id;
2110  kmp_cmplrdata_t
2111  data1; /* Two known optional additions: destructors and priority */
2112  kmp_cmplrdata_t data2; /* Process destructors first, priority second */
2113  /* future data */
2114  /* private vars */
2115 } kmp_task_t;
2116 
2121 typedef struct kmp_taskgroup {
2122  std::atomic<kmp_int32> count; // number of allocated and incomplete tasks
2123  std::atomic<kmp_int32>
2124  cancel_request; // request for cancellation of this taskgroup
2125  struct kmp_taskgroup *parent; // parent taskgroup
2126  // Block of data to perform task reduction
2127  void *reduce_data; // reduction related info
2128  kmp_int32 reduce_num_data; // number of data items to reduce
2129 } kmp_taskgroup_t;
2130 
2131 // forward declarations
2132 typedef union kmp_depnode kmp_depnode_t;
2133 typedef struct kmp_depnode_list kmp_depnode_list_t;
2134 typedef struct kmp_dephash_entry kmp_dephash_entry_t;
2135 
2136 // Compiler sends us this info:
2137 typedef struct kmp_depend_info {
2138  kmp_intptr_t base_addr;
2139  size_t len;
2140  struct {
2141  bool in : 1;
2142  bool out : 1;
2143  bool mtx : 1;
2144  } flags;
2145 } kmp_depend_info_t;
2146 
2147 // Internal structures to work with task dependencies:
2148 struct kmp_depnode_list {
2149  kmp_depnode_t *node;
2150  kmp_depnode_list_t *next;
2151 };
2152 
2153 // Max number of mutexinoutset dependencies per node
2154 #define MAX_MTX_DEPS 4
2155 
2156 typedef struct kmp_base_depnode {
2157  kmp_depnode_list_t *successors; /* used under lock */
2158  kmp_task_t *task; /* non-NULL if depnode is active, used under lock */
2159  kmp_lock_t *mtx_locks[MAX_MTX_DEPS]; /* lock mutexinoutset dependent tasks */
2160  kmp_int32 mtx_num_locks; /* number of locks in mtx_locks array */
2161  kmp_lock_t lock; /* guards shared fields: task, successors */
2162 #if KMP_SUPPORT_GRAPH_OUTPUT
2163  kmp_uint32 id;
2164 #endif
2165  std::atomic<kmp_int32> npredecessors;
2166  std::atomic<kmp_int32> nrefs;
2167 } kmp_base_depnode_t;
2168 
2169 union KMP_ALIGN_CACHE kmp_depnode {
2170  double dn_align; /* use worst case alignment */
2171  char dn_pad[KMP_PAD(kmp_base_depnode_t, CACHE_LINE)];
2172  kmp_base_depnode_t dn;
2173 };
2174 
2175 struct kmp_dephash_entry {
2176  kmp_intptr_t addr;
2177  kmp_depnode_t *last_out;
2178  kmp_depnode_list_t *last_ins;
2179  kmp_depnode_list_t *last_mtxs;
2180  kmp_int32 last_flag;
2181  kmp_lock_t *mtx_lock; /* is referenced by depnodes w/mutexinoutset dep */
2182  kmp_dephash_entry_t *next_in_bucket;
2183 };
2184 
2185 typedef struct kmp_dephash {
2186  kmp_dephash_entry_t **buckets;
2187  size_t size;
2188 #ifdef KMP_DEBUG
2189  kmp_uint32 nelements;
2190  kmp_uint32 nconflicts;
2191 #endif
2192 } kmp_dephash_t;
2193 
2194 typedef struct kmp_task_affinity_info {
2195  kmp_intptr_t base_addr;
2196  size_t len;
2197  struct {
2198  bool flag1 : 1;
2199  bool flag2 : 1;
2200  kmp_int32 reserved : 30;
2201  } flags;
2202 } kmp_task_affinity_info_t;
2203 
2204 typedef enum kmp_event_type_t {
2205  KMP_EVENT_UNINITIALIZED = 0,
2206  KMP_EVENT_ALLOW_COMPLETION = 1
2207 } kmp_event_type_t;
2208 
2209 typedef struct {
2210  kmp_event_type_t type;
2211  kmp_tas_lock_t lock;
2212  union {
2213  kmp_task_t *task;
2214  } ed;
2215 } kmp_event_t;
2216 
2217 #ifdef BUILD_TIED_TASK_STACK
2218 
2219 /* Tied Task stack definitions */
2220 typedef struct kmp_stack_block {
2221  kmp_taskdata_t *sb_block[TASK_STACK_BLOCK_SIZE];
2222  struct kmp_stack_block *sb_next;
2223  struct kmp_stack_block *sb_prev;
2224 } kmp_stack_block_t;
2225 
2226 typedef struct kmp_task_stack {
2227  kmp_stack_block_t ts_first_block; // first block of stack entries
2228  kmp_taskdata_t **ts_top; // pointer to the top of stack
2229  kmp_int32 ts_entries; // number of entries on the stack
2230 } kmp_task_stack_t;
2231 
2232 #endif // BUILD_TIED_TASK_STACK
2233 
2234 typedef struct kmp_tasking_flags { /* Total struct must be exactly 32 bits */
2235  /* Compiler flags */ /* Total compiler flags must be 16 bits */
2236  unsigned tiedness : 1; /* task is either tied (1) or untied (0) */
2237  unsigned final : 1; /* task is final(1) so execute immediately */
2238  unsigned merged_if0 : 1; /* no __kmpc_task_{begin/complete}_if0 calls in if0
2239  code path */
2240  unsigned destructors_thunk : 1; /* set if the compiler creates a thunk to
2241  invoke destructors from the runtime */
2242  unsigned proxy : 1; /* task is a proxy task (it will be executed outside the
2243  context of the RTL) */
2244  unsigned priority_specified : 1; /* set if the compiler provides priority
2245  setting for the task */
2246  unsigned detachable : 1; /* 1 == can detach */
2247  unsigned reserved : 9; /* reserved for compiler use */
2248 
2249  /* Library flags */ /* Total library flags must be 16 bits */
2250  unsigned tasktype : 1; /* task is either explicit(1) or implicit (0) */
2251  unsigned task_serial : 1; // task is executed immediately (1) or deferred (0)
2252  unsigned tasking_ser : 1; // all tasks in team are either executed immediately
2253  // (1) or may be deferred (0)
2254  unsigned team_serial : 1; // entire team is serial (1) [1 thread] or parallel
2255  // (0) [>= 2 threads]
2256  /* If either team_serial or tasking_ser is set, task team may be NULL */
2257  /* Task State Flags: */
2258  unsigned started : 1; /* 1==started, 0==not started */
2259  unsigned executing : 1; /* 1==executing, 0==not executing */
2260  unsigned complete : 1; /* 1==complete, 0==not complete */
2261  unsigned freed : 1; /* 1==freed, 0==allocateed */
2262  unsigned native : 1; /* 1==gcc-compiled task, 0==intel */
2263  unsigned reserved31 : 7; /* reserved for library use */
2264 
2265 } kmp_tasking_flags_t;
2266 
2267 struct kmp_taskdata { /* aligned during dynamic allocation */
2268  kmp_int32 td_task_id; /* id, assigned by debugger */
2269  kmp_tasking_flags_t td_flags; /* task flags */
2270  kmp_team_t *td_team; /* team for this task */
2271  kmp_info_p *td_alloc_thread; /* thread that allocated data structures */
2272  /* Currently not used except for perhaps IDB */
2273  kmp_taskdata_t *td_parent; /* parent task */
2274  kmp_int32 td_level; /* task nesting level */
2275  std::atomic<kmp_int32> td_untied_count; // untied task active parts counter
2276  ident_t *td_ident; /* task identifier */
2277  // Taskwait data.
2278  ident_t *td_taskwait_ident;
2279  kmp_uint32 td_taskwait_counter;
2280  kmp_int32 td_taskwait_thread; /* gtid + 1 of thread encountered taskwait */
2281  KMP_ALIGN_CACHE kmp_internal_control_t
2282  td_icvs; /* Internal control variables for the task */
2283  KMP_ALIGN_CACHE std::atomic<kmp_int32>
2284  td_allocated_child_tasks; /* Child tasks (+ current task) not yet
2285  deallocated */
2286  std::atomic<kmp_int32>
2287  td_incomplete_child_tasks; /* Child tasks not yet complete */
2288  kmp_taskgroup_t
2289  *td_taskgroup; // Each task keeps pointer to its current taskgroup
2290  kmp_dephash_t
2291  *td_dephash; // Dependencies for children tasks are tracked from here
2292  kmp_depnode_t
2293  *td_depnode; // Pointer to graph node if this task has dependencies
2294  kmp_task_team_t *td_task_team;
2295  kmp_int32 td_size_alloc; // The size of task structure, including shareds etc.
2296 #if defined(KMP_GOMP_COMPAT)
2297  // 4 or 8 byte integers for the loop bounds in GOMP_taskloop
2298  kmp_int32 td_size_loop_bounds;
2299 #endif
2300  kmp_taskdata_t *td_last_tied; // keep tied task for task scheduling constraint
2301 #if defined(KMP_GOMP_COMPAT)
2302  // GOMP sends in a copy function for copy constructors
2303  void (*td_copy_func)(void *, void *);
2304 #endif
2305  kmp_event_t td_allow_completion_event;
2306 #if OMPT_SUPPORT
2307  ompt_task_info_t ompt_task_info;
2308 #endif
2309 }; // struct kmp_taskdata
2310 
2311 // Make sure padding above worked
2312 KMP_BUILD_ASSERT(sizeof(kmp_taskdata_t) % sizeof(void *) == 0);
2313 
2314 // Data for task team but per thread
2315 typedef struct kmp_base_thread_data {
2316  kmp_info_p *td_thr; // Pointer back to thread info
2317  // Used only in __kmp_execute_tasks_template, maybe not avail until task is
2318  // queued?
2319  kmp_bootstrap_lock_t td_deque_lock; // Lock for accessing deque
2320  kmp_taskdata_t *
2321  *td_deque; // Deque of tasks encountered by td_thr, dynamically allocated
2322  kmp_int32 td_deque_size; // Size of deck
2323  kmp_uint32 td_deque_head; // Head of deque (will wrap)
2324  kmp_uint32 td_deque_tail; // Tail of deque (will wrap)
2325  kmp_int32 td_deque_ntasks; // Number of tasks in deque
2326  // GEH: shouldn't this be volatile since used in while-spin?
2327  kmp_int32 td_deque_last_stolen; // Thread number of last successful steal
2328 #ifdef BUILD_TIED_TASK_STACK
2329  kmp_task_stack_t td_susp_tied_tasks; // Stack of suspended tied tasks for task
2330 // scheduling constraint
2331 #endif // BUILD_TIED_TASK_STACK
2332 } kmp_base_thread_data_t;
2333 
2334 #define TASK_DEQUE_BITS 8 // Used solely to define INITIAL_TASK_DEQUE_SIZE
2335 #define INITIAL_TASK_DEQUE_SIZE (1 << TASK_DEQUE_BITS)
2336 
2337 #define TASK_DEQUE_SIZE(td) ((td).td_deque_size)
2338 #define TASK_DEQUE_MASK(td) ((td).td_deque_size - 1)
2339 
2340 typedef union KMP_ALIGN_CACHE kmp_thread_data {
2341  kmp_base_thread_data_t td;
2342  double td_align; /* use worst case alignment */
2343  char td_pad[KMP_PAD(kmp_base_thread_data_t, CACHE_LINE)];
2344 } kmp_thread_data_t;
2345 
2346 // Data for task teams which are used when tasking is enabled for the team
2347 typedef struct kmp_base_task_team {
2348  kmp_bootstrap_lock_t
2349  tt_threads_lock; /* Lock used to allocate per-thread part of task team */
2350  /* must be bootstrap lock since used at library shutdown*/
2351  kmp_task_team_t *tt_next; /* For linking the task team free list */
2352  kmp_thread_data_t
2353  *tt_threads_data; /* Array of per-thread structures for task team */
2354  /* Data survives task team deallocation */
2355  kmp_int32 tt_found_tasks; /* Have we found tasks and queued them while
2356  executing this team? */
2357  /* TRUE means tt_threads_data is set up and initialized */
2358  kmp_int32 tt_nproc; /* #threads in team */
2359  kmp_int32 tt_max_threads; // # entries allocated for threads_data array
2360  kmp_int32 tt_found_proxy_tasks; // found proxy tasks since last barrier
2361  kmp_int32 tt_untied_task_encountered;
2362 
2363  KMP_ALIGN_CACHE
2364  std::atomic<kmp_int32> tt_unfinished_threads; /* #threads still active */
2365 
2366  KMP_ALIGN_CACHE
2367  volatile kmp_uint32
2368  tt_active; /* is the team still actively executing tasks */
2369 } kmp_base_task_team_t;
2370 
2371 union KMP_ALIGN_CACHE kmp_task_team {
2372  kmp_base_task_team_t tt;
2373  double tt_align; /* use worst case alignment */
2374  char tt_pad[KMP_PAD(kmp_base_task_team_t, CACHE_LINE)];
2375 };
2376 
2377 #if (USE_FAST_MEMORY == 3) || (USE_FAST_MEMORY == 5)
2378 // Free lists keep same-size free memory slots for fast memory allocation
2379 // routines
2380 typedef struct kmp_free_list {
2381  void *th_free_list_self; // Self-allocated tasks free list
2382  void *th_free_list_sync; // Self-allocated tasks stolen/returned by other
2383  // threads
2384  void *th_free_list_other; // Non-self free list (to be returned to owner's
2385  // sync list)
2386 } kmp_free_list_t;
2387 #endif
2388 #if KMP_NESTED_HOT_TEAMS
2389 // Hot teams array keeps hot teams and their sizes for given thread. Hot teams
2390 // are not put in teams pool, and they don't put threads in threads pool.
2391 typedef struct kmp_hot_team_ptr {
2392  kmp_team_p *hot_team; // pointer to hot_team of given nesting level
2393  kmp_int32 hot_team_nth; // number of threads allocated for the hot_team
2394 } kmp_hot_team_ptr_t;
2395 #endif
2396 typedef struct kmp_teams_size {
2397  kmp_int32 nteams; // number of teams in a league
2398  kmp_int32 nth; // number of threads in each team of the league
2399 } kmp_teams_size_t;
2400 
2401 // This struct stores a thread that acts as a "root" for a contention
2402 // group. Contention groups are rooted at kmp_root threads, but also at
2403 // each master thread of each team created in the teams construct.
2404 // This struct therefore also stores a thread_limit associated with
2405 // that contention group, and a counter to track the number of threads
2406 // active in that contention group. Each thread has a list of these: CG
2407 // root threads have an entry in their list in which cg_root refers to
2408 // the thread itself, whereas other workers in the CG will have a
2409 // single entry where cg_root is same as the entry containing their CG
2410 // root. When a thread encounters a teams construct, it will add a new
2411 // entry to the front of its list, because it now roots a new CG.
2412 typedef struct kmp_cg_root {
2413  kmp_info_p *cg_root; // "root" thread for a contention group
2414  // The CG root's limit comes from OMP_THREAD_LIMIT for root threads, or
2415  // thread_limit clause for teams masters
2416  kmp_int32 cg_thread_limit;
2417  kmp_int32 cg_nthreads; // Count of active threads in CG rooted at cg_root
2418  struct kmp_cg_root *up; // pointer to higher level CG root in list
2419 } kmp_cg_root_t;
2420 
2421 // OpenMP thread data structures
2422 
2423 typedef struct KMP_ALIGN_CACHE kmp_base_info {
2424  /* Start with the readonly data which is cache aligned and padded. This is
2425  written before the thread starts working by the master. Uber masters may
2426  update themselves later. Usage does not consider serialized regions. */
2427  kmp_desc_t th_info;
2428  kmp_team_p *th_team; /* team we belong to */
2429  kmp_root_p *th_root; /* pointer to root of task hierarchy */
2430  kmp_info_p *th_next_pool; /* next available thread in the pool */
2431  kmp_disp_t *th_dispatch; /* thread's dispatch data */
2432  int th_in_pool; /* in thread pool (32 bits for TCR/TCW) */
2433 
2434  /* The following are cached from the team info structure */
2435  /* TODO use these in more places as determined to be needed via profiling */
2436  int th_team_nproc; /* number of threads in a team */
2437  kmp_info_p *th_team_master; /* the team's master thread */
2438  int th_team_serialized; /* team is serialized */
2439  microtask_t th_teams_microtask; /* save entry address for teams construct */
2440  int th_teams_level; /* save initial level of teams construct */
2441 /* it is 0 on device but may be any on host */
2442 
2443 /* The blocktime info is copied from the team struct to the thread sruct */
2444 /* at the start of a barrier, and the values stored in the team are used */
2445 /* at points in the code where the team struct is no longer guaranteed */
2446 /* to exist (from the POV of worker threads). */
2447 #if KMP_USE_MONITOR
2448  int th_team_bt_intervals;
2449  int th_team_bt_set;
2450 #else
2451  kmp_uint64 th_team_bt_intervals;
2452 #endif
2453 
2454 #if KMP_AFFINITY_SUPPORTED
2455  kmp_affin_mask_t *th_affin_mask; /* thread's current affinity mask */
2456 #endif
2457  omp_allocator_handle_t th_def_allocator; /* default allocator */
2458  /* The data set by the master at reinit, then R/W by the worker */
2459  KMP_ALIGN_CACHE int
2460  th_set_nproc; /* if > 0, then only use this request for the next fork */
2461 #if KMP_NESTED_HOT_TEAMS
2462  kmp_hot_team_ptr_t *th_hot_teams; /* array of hot teams */
2463 #endif
2464  kmp_proc_bind_t
2465  th_set_proc_bind; /* if != proc_bind_default, use request for next fork */
2466  kmp_teams_size_t
2467  th_teams_size; /* number of teams/threads in teams construct */
2468 #if KMP_AFFINITY_SUPPORTED
2469  int th_current_place; /* place currently bound to */
2470  int th_new_place; /* place to bind to in par reg */
2471  int th_first_place; /* first place in partition */
2472  int th_last_place; /* last place in partition */
2473 #endif
2474  int th_prev_level; /* previous level for affinity format */
2475  int th_prev_num_threads; /* previous num_threads for affinity format */
2476 #if USE_ITT_BUILD
2477  kmp_uint64 th_bar_arrive_time; /* arrival to barrier timestamp */
2478  kmp_uint64 th_bar_min_time; /* minimum arrival time at the barrier */
2479  kmp_uint64 th_frame_time; /* frame timestamp */
2480 #endif /* USE_ITT_BUILD */
2481  kmp_local_t th_local;
2482  struct private_common *th_pri_head;
2483 
2484  /* Now the data only used by the worker (after initial allocation) */
2485  /* TODO the first serial team should actually be stored in the info_t
2486  structure. this will help reduce initial allocation overhead */
2487  KMP_ALIGN_CACHE kmp_team_p
2488  *th_serial_team; /*serialized team held in reserve*/
2489 
2490 #if OMPT_SUPPORT
2491  ompt_thread_info_t ompt_thread_info;
2492 #endif
2493 
2494  /* The following are also read by the master during reinit */
2495  struct common_table *th_pri_common;
2496 
2497  volatile kmp_uint32 th_spin_here; /* thread-local location for spinning */
2498  /* while awaiting queuing lock acquire */
2499 
2500  volatile void *th_sleep_loc; // this points at a kmp_flag<T>
2501 
2502  ident_t *th_ident;
2503  unsigned th_x; // Random number generator data
2504  unsigned th_a; // Random number generator data
2505 
2506  /* Tasking-related data for the thread */
2507  kmp_task_team_t *th_task_team; // Task team struct
2508  kmp_taskdata_t *th_current_task; // Innermost Task being executed
2509  kmp_uint8 th_task_state; // alternating 0/1 for task team identification
2510  kmp_uint8 *th_task_state_memo_stack; // Stack holding memos of th_task_state
2511  // at nested levels
2512  kmp_uint32 th_task_state_top; // Top element of th_task_state_memo_stack
2513  kmp_uint32 th_task_state_stack_sz; // Size of th_task_state_memo_stack
2514  kmp_uint32 th_reap_state; // Non-zero indicates thread is not
2515  // tasking, thus safe to reap
2516 
2517  /* More stuff for keeping track of active/sleeping threads (this part is
2518  written by the worker thread) */
2519  kmp_uint8 th_active_in_pool; // included in count of #active threads in pool
2520  int th_active; // ! sleeping; 32 bits for TCR/TCW
2521  struct cons_header *th_cons; // used for consistency check
2522 #if KMP_USE_HIER_SCHED
2523  // used for hierarchical scheduling
2524  kmp_hier_private_bdata_t *th_hier_bar_data;
2525 #endif
2526 
2527  /* Add the syncronizing data which is cache aligned and padded. */
2528  KMP_ALIGN_CACHE kmp_balign_t th_bar[bs_last_barrier];
2529 
2530  KMP_ALIGN_CACHE volatile kmp_int32
2531  th_next_waiting; /* gtid+1 of next thread on lock wait queue, 0 if none */
2532 
2533 #if (USE_FAST_MEMORY == 3) || (USE_FAST_MEMORY == 5)
2534 #define NUM_LISTS 4
2535  kmp_free_list_t th_free_lists[NUM_LISTS]; // Free lists for fast memory
2536 // allocation routines
2537 #endif
2538 
2539 #if KMP_OS_WINDOWS
2540  kmp_win32_cond_t th_suspend_cv;
2541  kmp_win32_mutex_t th_suspend_mx;
2542  std::atomic<int> th_suspend_init;
2543 #endif
2544 #if KMP_OS_UNIX
2545  kmp_cond_align_t th_suspend_cv;
2546  kmp_mutex_align_t th_suspend_mx;
2547  std::atomic<int> th_suspend_init_count;
2548 #endif
2549 
2550 #if USE_ITT_BUILD
2551  kmp_itt_mark_t th_itt_mark_single;
2552 // alignment ???
2553 #endif /* USE_ITT_BUILD */
2554 #if KMP_STATS_ENABLED
2555  kmp_stats_list *th_stats;
2556 #endif
2557 #if KMP_OS_UNIX
2558  std::atomic<bool> th_blocking;
2559 #endif
2560  kmp_cg_root_t *th_cg_roots; // list of cg_roots associated with this thread
2561 } kmp_base_info_t;
2562 
2563 typedef union KMP_ALIGN_CACHE kmp_info {
2564  double th_align; /* use worst case alignment */
2565  char th_pad[KMP_PAD(kmp_base_info_t, CACHE_LINE)];
2566  kmp_base_info_t th;
2567 } kmp_info_t;
2568 
2569 // OpenMP thread team data structures
2570 
2571 typedef struct kmp_base_data { volatile kmp_uint32 t_value; } kmp_base_data_t;
2572 
2573 typedef union KMP_ALIGN_CACHE kmp_sleep_team {
2574  double dt_align; /* use worst case alignment */
2575  char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
2576  kmp_base_data_t dt;
2577 } kmp_sleep_team_t;
2578 
2579 typedef union KMP_ALIGN_CACHE kmp_ordered_team {
2580  double dt_align; /* use worst case alignment */
2581  char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
2582  kmp_base_data_t dt;
2583 } kmp_ordered_team_t;
2584 
2585 typedef int (*launch_t)(int gtid);
2586 
2587 /* Minimum number of ARGV entries to malloc if necessary */
2588 #define KMP_MIN_MALLOC_ARGV_ENTRIES 100
2589 
2590 // Set up how many argv pointers will fit in cache lines containing
2591 // t_inline_argv. Historically, we have supported at least 96 bytes. Using a
2592 // larger value for more space between the master write/worker read section and
2593 // read/write by all section seems to buy more performance on EPCC PARALLEL.
2594 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
2595 #define KMP_INLINE_ARGV_BYTES \
2596  (4 * CACHE_LINE - \
2597  ((3 * KMP_PTR_SKIP + 2 * sizeof(int) + 2 * sizeof(kmp_int8) + \
2598  sizeof(kmp_int16) + sizeof(kmp_uint32)) % \
2599  CACHE_LINE))
2600 #else
2601 #define KMP_INLINE_ARGV_BYTES \
2602  (2 * CACHE_LINE - ((3 * KMP_PTR_SKIP + 2 * sizeof(int)) % CACHE_LINE))
2603 #endif
2604 #define KMP_INLINE_ARGV_ENTRIES (int)(KMP_INLINE_ARGV_BYTES / KMP_PTR_SKIP)
2605 
2606 typedef struct KMP_ALIGN_CACHE kmp_base_team {
2607  // Synchronization Data
2608  // ---------------------------------------------------------------------------
2609  KMP_ALIGN_CACHE kmp_ordered_team_t t_ordered;
2610  kmp_balign_team_t t_bar[bs_last_barrier];
2611  std::atomic<int> t_construct; // count of single directive encountered by team
2612  char pad[sizeof(kmp_lock_t)]; // padding to maintain performance on big iron
2613 
2614  // [0] - parallel / [1] - worksharing task reduction data shared by taskgroups
2615  std::atomic<void *> t_tg_reduce_data[2]; // to support task modifier
2616  std::atomic<int> t_tg_fini_counter[2]; // sync end of task reductions
2617 
2618  // Master only
2619  // ---------------------------------------------------------------------------
2620  KMP_ALIGN_CACHE int t_master_tid; // tid of master in parent team
2621  int t_master_this_cons; // "this_construct" single counter of master in parent
2622  // team
2623  ident_t *t_ident; // if volatile, have to change too much other crud to
2624  // volatile too
2625  kmp_team_p *t_parent; // parent team
2626  kmp_team_p *t_next_pool; // next free team in the team pool
2627  kmp_disp_t *t_dispatch; // thread's dispatch data
2628  kmp_task_team_t *t_task_team[2]; // Task team struct; switch between 2
2629  kmp_proc_bind_t t_proc_bind; // bind type for par region
2630 #if USE_ITT_BUILD
2631  kmp_uint64 t_region_time; // region begin timestamp
2632 #endif /* USE_ITT_BUILD */
2633 
2634  // Master write, workers read
2635  // --------------------------------------------------------------------------
2636  KMP_ALIGN_CACHE void **t_argv;
2637  int t_argc;
2638  int t_nproc; // number of threads in team
2639  microtask_t t_pkfn;
2640  launch_t t_invoke; // procedure to launch the microtask
2641 
2642 #if OMPT_SUPPORT
2643  ompt_team_info_t ompt_team_info;
2644  ompt_lw_taskteam_t *ompt_serialized_team_info;
2645 #endif
2646 
2647 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
2648  kmp_int8 t_fp_control_saved;
2649  kmp_int8 t_pad2b;
2650  kmp_int16 t_x87_fpu_control_word; // FP control regs
2651  kmp_uint32 t_mxcsr;
2652 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
2653 
2654  void *t_inline_argv[KMP_INLINE_ARGV_ENTRIES];
2655 
2656  KMP_ALIGN_CACHE kmp_info_t **t_threads;
2657  kmp_taskdata_t
2658  *t_implicit_task_taskdata; // Taskdata for the thread's implicit task
2659  int t_level; // nested parallel level
2660 
2661  KMP_ALIGN_CACHE int t_max_argc;
2662  int t_max_nproc; // max threads this team can handle (dynamicly expandable)
2663  int t_serialized; // levels deep of serialized teams
2664  dispatch_shared_info_t *t_disp_buffer; // buffers for dispatch system
2665  int t_id; // team's id, assigned by debugger.
2666  int t_active_level; // nested active parallel level
2667  kmp_r_sched_t t_sched; // run-time schedule for the team
2668 #if KMP_AFFINITY_SUPPORTED
2669  int t_first_place; // first & last place in parent thread's partition.
2670  int t_last_place; // Restore these values to master after par region.
2671 #endif // KMP_AFFINITY_SUPPORTED
2672  int t_display_affinity;
2673  int t_size_changed; // team size was changed?: 0: no, 1: yes, -1: changed via
2674  // omp_set_num_threads() call
2675  omp_allocator_handle_t t_def_allocator; /* default allocator */
2676 
2677 // Read/write by workers as well
2678 #if (KMP_ARCH_X86 || KMP_ARCH_X86_64)
2679  // Using CACHE_LINE=64 reduces memory footprint, but causes a big perf
2680  // regression of epcc 'parallel' and 'barrier' on fxe256lin01. This extra
2681  // padding serves to fix the performance of epcc 'parallel' and 'barrier' when
2682  // CACHE_LINE=64. TODO: investigate more and get rid if this padding.
2683  char dummy_padding[1024];
2684 #endif
2685  // Internal control stack for additional nested teams.
2686  KMP_ALIGN_CACHE kmp_internal_control_t *t_control_stack_top;
2687  // for SERIALIZED teams nested 2 or more levels deep
2688  // typed flag to store request state of cancellation
2689  std::atomic<kmp_int32> t_cancel_request;
2690  int t_master_active; // save on fork, restore on join
2691  void *t_copypriv_data; // team specific pointer to copyprivate data array
2692 #if KMP_OS_WINDOWS
2693  std::atomic<kmp_uint32> t_copyin_counter;
2694 #endif
2695 #if USE_ITT_BUILD
2696  void *t_stack_id; // team specific stack stitching id (for ittnotify)
2697 #endif /* USE_ITT_BUILD */
2698 } kmp_base_team_t;
2699 
2700 union KMP_ALIGN_CACHE kmp_team {
2701  kmp_base_team_t t;
2702  double t_align; /* use worst case alignment */
2703  char t_pad[KMP_PAD(kmp_base_team_t, CACHE_LINE)];
2704 };
2705 
2706 typedef union KMP_ALIGN_CACHE kmp_time_global {
2707  double dt_align; /* use worst case alignment */
2708  char dt_pad[KMP_PAD(kmp_base_data_t, CACHE_LINE)];
2709  kmp_base_data_t dt;
2710 } kmp_time_global_t;
2711 
2712 typedef struct kmp_base_global {
2713  /* cache-aligned */
2714  kmp_time_global_t g_time;
2715 
2716  /* non cache-aligned */
2717  volatile int g_abort;
2718  volatile int g_done;
2719 
2720  int g_dynamic;
2721  enum dynamic_mode g_dynamic_mode;
2722 } kmp_base_global_t;
2723 
2724 typedef union KMP_ALIGN_CACHE kmp_global {
2725  kmp_base_global_t g;
2726  double g_align; /* use worst case alignment */
2727  char g_pad[KMP_PAD(kmp_base_global_t, CACHE_LINE)];
2728 } kmp_global_t;
2729 
2730 typedef struct kmp_base_root {
2731  // TODO: GEH - combine r_active with r_in_parallel then r_active ==
2732  // (r_in_parallel>= 0)
2733  // TODO: GEH - then replace r_active with t_active_levels if we can to reduce
2734  // the synch overhead or keeping r_active
2735  volatile int r_active; /* TRUE if some region in a nest has > 1 thread */
2736  // keeps a count of active parallel regions per root
2737  std::atomic<int> r_in_parallel;
2738  // GEH: This is misnamed, should be r_active_levels
2739  kmp_team_t *r_root_team;
2740  kmp_team_t *r_hot_team;
2741  kmp_info_t *r_uber_thread;
2742  kmp_lock_t r_begin_lock;
2743  volatile int r_begin;
2744  int r_blocktime; /* blocktime for this root and descendants */
2745 } kmp_base_root_t;
2746 
2747 typedef union KMP_ALIGN_CACHE kmp_root {
2748  kmp_base_root_t r;
2749  double r_align; /* use worst case alignment */
2750  char r_pad[KMP_PAD(kmp_base_root_t, CACHE_LINE)];
2751 } kmp_root_t;
2752 
2753 struct fortran_inx_info {
2754  kmp_int32 data;
2755 };
2756 
2757 /* ------------------------------------------------------------------------ */
2758 
2759 extern int __kmp_settings;
2760 extern int __kmp_duplicate_library_ok;
2761 #if USE_ITT_BUILD
2762 extern int __kmp_forkjoin_frames;
2763 extern int __kmp_forkjoin_frames_mode;
2764 #endif
2765 extern PACKED_REDUCTION_METHOD_T __kmp_force_reduction_method;
2766 extern int __kmp_determ_red;
2767 
2768 #ifdef KMP_DEBUG
2769 extern int kmp_a_debug;
2770 extern int kmp_b_debug;
2771 extern int kmp_c_debug;
2772 extern int kmp_d_debug;
2773 extern int kmp_e_debug;
2774 extern int kmp_f_debug;
2775 #endif /* KMP_DEBUG */
2776 
2777 /* For debug information logging using rotating buffer */
2778 #define KMP_DEBUG_BUF_LINES_INIT 512
2779 #define KMP_DEBUG_BUF_LINES_MIN 1
2780 
2781 #define KMP_DEBUG_BUF_CHARS_INIT 128
2782 #define KMP_DEBUG_BUF_CHARS_MIN 2
2783 
2784 extern int
2785  __kmp_debug_buf; /* TRUE means use buffer, FALSE means print to stderr */
2786 extern int __kmp_debug_buf_lines; /* How many lines of debug stored in buffer */
2787 extern int
2788  __kmp_debug_buf_chars; /* How many characters allowed per line in buffer */
2789 extern int __kmp_debug_buf_atomic; /* TRUE means use atomic update of buffer
2790  entry pointer */
2791 
2792 extern char *__kmp_debug_buffer; /* Debug buffer itself */
2793 extern std::atomic<int> __kmp_debug_count; /* Counter for number of lines
2794  printed in buffer so far */
2795 extern int __kmp_debug_buf_warn_chars; /* Keep track of char increase
2796  recommended in warnings */
2797 /* end rotating debug buffer */
2798 
2799 #ifdef KMP_DEBUG
2800 extern int __kmp_par_range; /* +1 => only go par for constructs in range */
2801 
2802 #define KMP_PAR_RANGE_ROUTINE_LEN 1024
2803 extern char __kmp_par_range_routine[KMP_PAR_RANGE_ROUTINE_LEN];
2804 #define KMP_PAR_RANGE_FILENAME_LEN 1024
2805 extern char __kmp_par_range_filename[KMP_PAR_RANGE_FILENAME_LEN];
2806 extern int __kmp_par_range_lb;
2807 extern int __kmp_par_range_ub;
2808 #endif
2809 
2810 /* For printing out dynamic storage map for threads and teams */
2811 extern int
2812  __kmp_storage_map; /* True means print storage map for threads and teams */
2813 extern int __kmp_storage_map_verbose; /* True means storage map includes
2814  placement info */
2815 extern int __kmp_storage_map_verbose_specified;
2816 
2817 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
2818 extern kmp_cpuinfo_t __kmp_cpuinfo;
2819 #endif
2820 
2821 extern volatile int __kmp_init_serial;
2822 extern volatile int __kmp_init_gtid;
2823 extern volatile int __kmp_init_common;
2824 extern volatile int __kmp_init_middle;
2825 extern volatile int __kmp_init_parallel;
2826 #if KMP_USE_MONITOR
2827 extern volatile int __kmp_init_monitor;
2828 #endif
2829 extern volatile int __kmp_init_user_locks;
2830 extern int __kmp_init_counter;
2831 extern int __kmp_root_counter;
2832 extern int __kmp_version;
2833 
2834 /* list of address of allocated caches for commons */
2835 extern kmp_cached_addr_t *__kmp_threadpriv_cache_list;
2836 
2837 /* Barrier algorithm types and options */
2838 extern kmp_uint32 __kmp_barrier_gather_bb_dflt;
2839 extern kmp_uint32 __kmp_barrier_release_bb_dflt;
2840 extern kmp_bar_pat_e __kmp_barrier_gather_pat_dflt;
2841 extern kmp_bar_pat_e __kmp_barrier_release_pat_dflt;
2842 extern kmp_uint32 __kmp_barrier_gather_branch_bits[bs_last_barrier];
2843 extern kmp_uint32 __kmp_barrier_release_branch_bits[bs_last_barrier];
2844 extern kmp_bar_pat_e __kmp_barrier_gather_pattern[bs_last_barrier];
2845 extern kmp_bar_pat_e __kmp_barrier_release_pattern[bs_last_barrier];
2846 extern char const *__kmp_barrier_branch_bit_env_name[bs_last_barrier];
2847 extern char const *__kmp_barrier_pattern_env_name[bs_last_barrier];
2848 extern char const *__kmp_barrier_type_name[bs_last_barrier];
2849 extern char const *__kmp_barrier_pattern_name[bp_last_bar];
2850 
2851 /* Global Locks */
2852 extern kmp_bootstrap_lock_t __kmp_initz_lock; /* control initialization */
2853 extern kmp_bootstrap_lock_t __kmp_forkjoin_lock; /* control fork/join access */
2854 extern kmp_bootstrap_lock_t __kmp_task_team_lock;
2855 extern kmp_bootstrap_lock_t
2856  __kmp_exit_lock; /* exit() is not always thread-safe */
2857 #if KMP_USE_MONITOR
2858 extern kmp_bootstrap_lock_t
2859  __kmp_monitor_lock; /* control monitor thread creation */
2860 #endif
2861 extern kmp_bootstrap_lock_t
2862  __kmp_tp_cached_lock; /* used for the hack to allow threadprivate cache and
2863  __kmp_threads expansion to co-exist */
2864 
2865 extern kmp_lock_t __kmp_global_lock; /* control OS/global access */
2866 extern kmp_queuing_lock_t __kmp_dispatch_lock; /* control dispatch access */
2867 extern kmp_lock_t __kmp_debug_lock; /* control I/O access for KMP_DEBUG */
2868 
2869 extern enum library_type __kmp_library;
2870 
2871 extern enum sched_type __kmp_sched; /* default runtime scheduling */
2872 extern enum sched_type __kmp_static; /* default static scheduling method */
2873 extern enum sched_type __kmp_guided; /* default guided scheduling method */
2874 extern enum sched_type __kmp_auto; /* default auto scheduling method */
2875 extern int __kmp_chunk; /* default runtime chunk size */
2876 
2877 extern size_t __kmp_stksize; /* stack size per thread */
2878 #if KMP_USE_MONITOR
2879 extern size_t __kmp_monitor_stksize; /* stack size for monitor thread */
2880 #endif
2881 extern size_t __kmp_stkoffset; /* stack offset per thread */
2882 extern int __kmp_stkpadding; /* Should we pad root thread(s) stack */
2883 
2884 extern size_t
2885  __kmp_malloc_pool_incr; /* incremental size of pool for kmp_malloc() */
2886 extern int __kmp_env_stksize; /* was KMP_STACKSIZE specified? */
2887 extern int __kmp_env_blocktime; /* was KMP_BLOCKTIME specified? */
2888 extern int __kmp_env_checks; /* was KMP_CHECKS specified? */
2889 extern int __kmp_env_consistency_check; // was KMP_CONSISTENCY_CHECK specified?
2890 extern int __kmp_generate_warnings; /* should we issue warnings? */
2891 extern int __kmp_reserve_warn; /* have we issued reserve_threads warning? */
2892 
2893 #ifdef DEBUG_SUSPEND
2894 extern int __kmp_suspend_count; /* count inside __kmp_suspend_template() */
2895 #endif
2896 
2897 extern kmp_int32 __kmp_use_yield;
2898 extern kmp_int32 __kmp_use_yield_exp_set;
2899 extern kmp_uint32 __kmp_yield_init;
2900 extern kmp_uint32 __kmp_yield_next;
2901 
2902 /* ------------------------------------------------------------------------- */
2903 extern int __kmp_allThreadsSpecified;
2904 
2905 extern size_t __kmp_align_alloc;
2906 /* following data protected by initialization routines */
2907 extern int __kmp_xproc; /* number of processors in the system */
2908 extern int __kmp_avail_proc; /* number of processors available to the process */
2909 extern size_t __kmp_sys_min_stksize; /* system-defined minimum stack size */
2910 extern int __kmp_sys_max_nth; /* system-imposed maximum number of threads */
2911 // maximum total number of concurrently-existing threads on device
2912 extern int __kmp_max_nth;
2913 // maximum total number of concurrently-existing threads in a contention group
2914 extern int __kmp_cg_max_nth;
2915 extern int __kmp_teams_max_nth; // max threads used in a teams construct
2916 extern int __kmp_threads_capacity; /* capacity of the arrays __kmp_threads and
2917  __kmp_root */
2918 extern int __kmp_dflt_team_nth; /* default number of threads in a parallel
2919  region a la OMP_NUM_THREADS */
2920 extern int __kmp_dflt_team_nth_ub; /* upper bound on "" determined at serial
2921  initialization */
2922 extern int __kmp_tp_capacity; /* capacity of __kmp_threads if threadprivate is
2923  used (fixed) */
2924 extern int __kmp_tp_cached; /* whether threadprivate cache has been created
2925  (__kmpc_threadprivate_cached()) */
2926 extern int __kmp_dflt_blocktime; /* number of milliseconds to wait before
2927  blocking (env setting) */
2928 #if KMP_USE_MONITOR
2929 extern int
2930  __kmp_monitor_wakeups; /* number of times monitor wakes up per second */
2931 extern int __kmp_bt_intervals; /* number of monitor timestamp intervals before
2932  blocking */
2933 #endif
2934 #ifdef KMP_ADJUST_BLOCKTIME
2935 extern int __kmp_zero_bt; /* whether blocktime has been forced to zero */
2936 #endif /* KMP_ADJUST_BLOCKTIME */
2937 #ifdef KMP_DFLT_NTH_CORES
2938 extern int __kmp_ncores; /* Total number of cores for threads placement */
2939 #endif
2940 /* Number of millisecs to delay on abort for Intel(R) VTune(TM) tools */
2941 extern int __kmp_abort_delay;
2942 
2943 extern int __kmp_need_register_atfork_specified;
2944 extern int
2945  __kmp_need_register_atfork; /* At initialization, call pthread_atfork to
2946  install fork handler */
2947 extern int __kmp_gtid_mode; /* Method of getting gtid, values:
2948  0 - not set, will be set at runtime
2949  1 - using stack search
2950  2 - dynamic TLS (pthread_getspecific(Linux* OS/OS
2951  X*) or TlsGetValue(Windows* OS))
2952  3 - static TLS (__declspec(thread) __kmp_gtid),
2953  Linux* OS .so only. */
2954 extern int
2955  __kmp_adjust_gtid_mode; /* If true, adjust method based on #threads */
2956 #ifdef KMP_TDATA_GTID
2957 extern KMP_THREAD_LOCAL int __kmp_gtid;
2958 #endif
2959 extern int __kmp_tls_gtid_min; /* #threads below which use sp search for gtid */
2960 extern int __kmp_foreign_tp; // If true, separate TP var for each foreign thread
2961 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
2962 extern int __kmp_inherit_fp_control; // copy fp creg(s) parent->workers at fork
2963 extern kmp_int16 __kmp_init_x87_fpu_control_word; // init thread's FP ctrl reg
2964 extern kmp_uint32 __kmp_init_mxcsr; /* init thread's mxscr */
2965 #endif /* KMP_ARCH_X86 || KMP_ARCH_X86_64 */
2966 
2967 // max_active_levels for nested parallelism enabled by default via
2968 // OMP_MAX_ACTIVE_LEVELS, OMP_NESTED, OMP_NUM_THREADS, and OMP_PROC_BIND
2969 extern int __kmp_dflt_max_active_levels;
2970 // Indicates whether value of __kmp_dflt_max_active_levels was already
2971 // explicitly set by OMP_MAX_ACTIVE_LEVELS or OMP_NESTED=false
2972 extern bool __kmp_dflt_max_active_levels_set;
2973 extern int __kmp_dispatch_num_buffers; /* max possible dynamic loops in
2974  concurrent execution per team */
2975 #if KMP_NESTED_HOT_TEAMS
2976 extern int __kmp_hot_teams_mode;
2977 extern int __kmp_hot_teams_max_level;
2978 #endif
2979 
2980 #if KMP_OS_LINUX
2981 extern enum clock_function_type __kmp_clock_function;
2982 extern int __kmp_clock_function_param;
2983 #endif /* KMP_OS_LINUX */
2984 
2985 #if KMP_MIC_SUPPORTED
2986 extern enum mic_type __kmp_mic_type;
2987 #endif
2988 
2989 #ifdef USE_LOAD_BALANCE
2990 extern double __kmp_load_balance_interval; // load balance algorithm interval
2991 #endif /* USE_LOAD_BALANCE */
2992 
2993 // OpenMP 3.1 - Nested num threads array
2994 typedef struct kmp_nested_nthreads_t {
2995  int *nth;
2996  int size;
2997  int used;
2998 } kmp_nested_nthreads_t;
2999 
3000 extern kmp_nested_nthreads_t __kmp_nested_nth;
3001 
3002 #if KMP_USE_ADAPTIVE_LOCKS
3003 
3004 // Parameters for the speculative lock backoff system.
3005 struct kmp_adaptive_backoff_params_t {
3006  // Number of soft retries before it counts as a hard retry.
3007  kmp_uint32 max_soft_retries;
3008  // Badness is a bit mask : 0,1,3,7,15,... on each hard failure we move one to
3009  // the right
3010  kmp_uint32 max_badness;
3011 };
3012 
3013 extern kmp_adaptive_backoff_params_t __kmp_adaptive_backoff_params;
3014 
3015 #if KMP_DEBUG_ADAPTIVE_LOCKS
3016 extern const char *__kmp_speculative_statsfile;
3017 #endif
3018 
3019 #endif // KMP_USE_ADAPTIVE_LOCKS
3020 
3021 extern int __kmp_display_env; /* TRUE or FALSE */
3022 extern int __kmp_display_env_verbose; /* TRUE if OMP_DISPLAY_ENV=VERBOSE */
3023 extern int __kmp_omp_cancellation; /* TRUE or FALSE */
3024 
3025 /* ------------------------------------------------------------------------- */
3026 
3027 /* the following are protected by the fork/join lock */
3028 /* write: lock read: anytime */
3029 extern kmp_info_t **__kmp_threads; /* Descriptors for the threads */
3030 /* read/write: lock */
3031 extern volatile kmp_team_t *__kmp_team_pool;
3032 extern volatile kmp_info_t *__kmp_thread_pool;
3033 extern kmp_info_t *__kmp_thread_pool_insert_pt;
3034 
3035 // total num threads reachable from some root thread including all root threads
3036 extern volatile int __kmp_nth;
3037 /* total number of threads reachable from some root thread including all root
3038  threads, and those in the thread pool */
3039 extern volatile int __kmp_all_nth;
3040 extern std::atomic<int> __kmp_thread_pool_active_nth;
3041 
3042 extern kmp_root_t **__kmp_root; /* root of thread hierarchy */
3043 /* end data protected by fork/join lock */
3044 /* ------------------------------------------------------------------------- */
3045 
3046 #define __kmp_get_gtid() __kmp_get_global_thread_id()
3047 #define __kmp_entry_gtid() __kmp_get_global_thread_id_reg()
3048 #define __kmp_get_tid() (__kmp_tid_from_gtid(__kmp_get_gtid()))
3049 #define __kmp_get_team() (__kmp_threads[(__kmp_get_gtid())]->th.th_team)
3050 #define __kmp_get_thread() (__kmp_thread_from_gtid(__kmp_get_gtid()))
3051 
3052 // AT: Which way is correct?
3053 // AT: 1. nproc = __kmp_threads[ ( gtid ) ] -> th.th_team -> t.t_nproc;
3054 // AT: 2. nproc = __kmp_threads[ ( gtid ) ] -> th.th_team_nproc;
3055 #define __kmp_get_team_num_threads(gtid) \
3056  (__kmp_threads[(gtid)]->th.th_team->t.t_nproc)
3057 
3058 static inline bool KMP_UBER_GTID(int gtid) {
3059  KMP_DEBUG_ASSERT(gtid >= KMP_GTID_MIN);
3060  KMP_DEBUG_ASSERT(gtid < __kmp_threads_capacity);
3061  return (gtid >= 0 && __kmp_root[gtid] && __kmp_threads[gtid] &&
3062  __kmp_threads[gtid] == __kmp_root[gtid]->r.r_uber_thread);
3063 }
3064 
3065 static inline int __kmp_tid_from_gtid(int gtid) {
3066  KMP_DEBUG_ASSERT(gtid >= 0);
3067  return __kmp_threads[gtid]->th.th_info.ds.ds_tid;
3068 }
3069 
3070 static inline int __kmp_gtid_from_tid(int tid, const kmp_team_t *team) {
3071  KMP_DEBUG_ASSERT(tid >= 0 && team);
3072  return team->t.t_threads[tid]->th.th_info.ds.ds_gtid;
3073 }
3074 
3075 static inline int __kmp_gtid_from_thread(const kmp_info_t *thr) {
3076  KMP_DEBUG_ASSERT(thr);
3077  return thr->th.th_info.ds.ds_gtid;
3078 }
3079 
3080 static inline kmp_info_t *__kmp_thread_from_gtid(int gtid) {
3081  KMP_DEBUG_ASSERT(gtid >= 0);
3082  return __kmp_threads[gtid];
3083 }
3084 
3085 static inline kmp_team_t *__kmp_team_from_gtid(int gtid) {
3086  KMP_DEBUG_ASSERT(gtid >= 0);
3087  return __kmp_threads[gtid]->th.th_team;
3088 }
3089 
3090 /* ------------------------------------------------------------------------- */
3091 
3092 extern kmp_global_t __kmp_global; /* global status */
3093 
3094 extern kmp_info_t __kmp_monitor;
3095 // For Debugging Support Library
3096 extern std::atomic<kmp_int32> __kmp_team_counter;
3097 // For Debugging Support Library
3098 extern std::atomic<kmp_int32> __kmp_task_counter;
3099 
3100 #if USE_DEBUGGER
3101 #define _KMP_GEN_ID(counter) \
3102  (__kmp_debugging ? KMP_ATOMIC_INC(&counter) + 1 : ~0)
3103 #else
3104 #define _KMP_GEN_ID(counter) (~0)
3105 #endif /* USE_DEBUGGER */
3106 
3107 #define KMP_GEN_TASK_ID() _KMP_GEN_ID(__kmp_task_counter)
3108 #define KMP_GEN_TEAM_ID() _KMP_GEN_ID(__kmp_team_counter)
3109 
3110 /* ------------------------------------------------------------------------ */
3111 
3112 extern void __kmp_print_storage_map_gtid(int gtid, void *p1, void *p2,
3113  size_t size, char const *format, ...);
3114 
3115 extern void __kmp_serial_initialize(void);
3116 extern void __kmp_middle_initialize(void);
3117 extern void __kmp_parallel_initialize(void);
3118 
3119 extern void __kmp_internal_begin(void);
3120 extern void __kmp_internal_end_library(int gtid);
3121 extern void __kmp_internal_end_thread(int gtid);
3122 extern void __kmp_internal_end_atexit(void);
3123 extern void __kmp_internal_end_fini(void);
3124 extern void __kmp_internal_end_dtor(void);
3125 extern void __kmp_internal_end_dest(void *);
3126 
3127 extern int __kmp_register_root(int initial_thread);
3128 extern void __kmp_unregister_root(int gtid);
3129 
3130 extern int __kmp_ignore_mppbeg(void);
3131 extern int __kmp_ignore_mppend(void);
3132 
3133 extern int __kmp_enter_single(int gtid, ident_t *id_ref, int push_ws);
3134 extern void __kmp_exit_single(int gtid);
3135 
3136 extern void __kmp_parallel_deo(int *gtid_ref, int *cid_ref, ident_t *loc_ref);
3137 extern void __kmp_parallel_dxo(int *gtid_ref, int *cid_ref, ident_t *loc_ref);
3138 
3139 #ifdef USE_LOAD_BALANCE
3140 extern int __kmp_get_load_balance(int);
3141 #endif
3142 
3143 extern int __kmp_get_global_thread_id(void);
3144 extern int __kmp_get_global_thread_id_reg(void);
3145 extern void __kmp_exit_thread(int exit_status);
3146 extern void __kmp_abort(char const *format, ...);
3147 extern void __kmp_abort_thread(void);
3148 KMP_NORETURN extern void __kmp_abort_process(void);
3149 extern void __kmp_warn(char const *format, ...);
3150 
3151 extern void __kmp_set_num_threads(int new_nth, int gtid);
3152 
3153 // Returns current thread (pointer to kmp_info_t). Current thread *must* be
3154 // registered.
3155 static inline kmp_info_t *__kmp_entry_thread() {
3156  int gtid = __kmp_entry_gtid();
3157 
3158  return __kmp_threads[gtid];
3159 }
3160 
3161 extern void __kmp_set_max_active_levels(int gtid, int new_max_active_levels);
3162 extern int __kmp_get_max_active_levels(int gtid);
3163 extern int __kmp_get_ancestor_thread_num(int gtid, int level);
3164 extern int __kmp_get_team_size(int gtid, int level);
3165 extern void __kmp_set_schedule(int gtid, kmp_sched_t new_sched, int chunk);
3166 extern void __kmp_get_schedule(int gtid, kmp_sched_t *sched, int *chunk);
3167 
3168 extern unsigned short __kmp_get_random(kmp_info_t *thread);
3169 extern void __kmp_init_random(kmp_info_t *thread);
3170 
3171 extern kmp_r_sched_t __kmp_get_schedule_global(void);
3172 extern void __kmp_adjust_num_threads(int new_nproc);
3173 extern void __kmp_check_stksize(size_t *val);
3174 
3175 extern void *___kmp_allocate(size_t size KMP_SRC_LOC_DECL);
3176 extern void *___kmp_page_allocate(size_t size KMP_SRC_LOC_DECL);
3177 extern void ___kmp_free(void *ptr KMP_SRC_LOC_DECL);
3178 #define __kmp_allocate(size) ___kmp_allocate((size)KMP_SRC_LOC_CURR)
3179 #define __kmp_page_allocate(size) ___kmp_page_allocate((size)KMP_SRC_LOC_CURR)
3180 #define __kmp_free(ptr) ___kmp_free((ptr)KMP_SRC_LOC_CURR)
3181 
3182 #if USE_FAST_MEMORY
3183 extern void *___kmp_fast_allocate(kmp_info_t *this_thr,
3184  size_t size KMP_SRC_LOC_DECL);
3185 extern void ___kmp_fast_free(kmp_info_t *this_thr, void *ptr KMP_SRC_LOC_DECL);
3186 extern void __kmp_free_fast_memory(kmp_info_t *this_thr);
3187 extern void __kmp_initialize_fast_memory(kmp_info_t *this_thr);
3188 #define __kmp_fast_allocate(this_thr, size) \
3189  ___kmp_fast_allocate((this_thr), (size)KMP_SRC_LOC_CURR)
3190 #define __kmp_fast_free(this_thr, ptr) \
3191  ___kmp_fast_free((this_thr), (ptr)KMP_SRC_LOC_CURR)
3192 #endif
3193 
3194 extern void *___kmp_thread_malloc(kmp_info_t *th, size_t size KMP_SRC_LOC_DECL);
3195 extern void *___kmp_thread_calloc(kmp_info_t *th, size_t nelem,
3196  size_t elsize KMP_SRC_LOC_DECL);
3197 extern void *___kmp_thread_realloc(kmp_info_t *th, void *ptr,
3198  size_t size KMP_SRC_LOC_DECL);
3199 extern void ___kmp_thread_free(kmp_info_t *th, void *ptr KMP_SRC_LOC_DECL);
3200 #define __kmp_thread_malloc(th, size) \
3201  ___kmp_thread_malloc((th), (size)KMP_SRC_LOC_CURR)
3202 #define __kmp_thread_calloc(th, nelem, elsize) \
3203  ___kmp_thread_calloc((th), (nelem), (elsize)KMP_SRC_LOC_CURR)
3204 #define __kmp_thread_realloc(th, ptr, size) \
3205  ___kmp_thread_realloc((th), (ptr), (size)KMP_SRC_LOC_CURR)
3206 #define __kmp_thread_free(th, ptr) \
3207  ___kmp_thread_free((th), (ptr)KMP_SRC_LOC_CURR)
3208 
3209 #define KMP_INTERNAL_MALLOC(sz) malloc(sz)
3210 #define KMP_INTERNAL_FREE(p) free(p)
3211 #define KMP_INTERNAL_REALLOC(p, sz) realloc((p), (sz))
3212 #define KMP_INTERNAL_CALLOC(n, sz) calloc((n), (sz))
3213 
3214 extern void __kmp_push_num_threads(ident_t *loc, int gtid, int num_threads);
3215 
3216 extern void __kmp_push_proc_bind(ident_t *loc, int gtid,
3217  kmp_proc_bind_t proc_bind);
3218 extern void __kmp_push_num_teams(ident_t *loc, int gtid, int num_teams,
3219  int num_threads);
3220 
3221 extern void __kmp_yield();
3222 
3223 extern void __kmpc_dispatch_init_4(ident_t *loc, kmp_int32 gtid,
3224  enum sched_type schedule, kmp_int32 lb,
3225  kmp_int32 ub, kmp_int32 st, kmp_int32 chunk);
3226 extern void __kmpc_dispatch_init_4u(ident_t *loc, kmp_int32 gtid,
3227  enum sched_type schedule, kmp_uint32 lb,
3228  kmp_uint32 ub, kmp_int32 st,
3229  kmp_int32 chunk);
3230 extern void __kmpc_dispatch_init_8(ident_t *loc, kmp_int32 gtid,
3231  enum sched_type schedule, kmp_int64 lb,
3232  kmp_int64 ub, kmp_int64 st, kmp_int64 chunk);
3233 extern void __kmpc_dispatch_init_8u(ident_t *loc, kmp_int32 gtid,
3234  enum sched_type schedule, kmp_uint64 lb,
3235  kmp_uint64 ub, kmp_int64 st,
3236  kmp_int64 chunk);
3237 
3238 extern int __kmpc_dispatch_next_4(ident_t *loc, kmp_int32 gtid,
3239  kmp_int32 *p_last, kmp_int32 *p_lb,
3240  kmp_int32 *p_ub, kmp_int32 *p_st);
3241 extern int __kmpc_dispatch_next_4u(ident_t *loc, kmp_int32 gtid,
3242  kmp_int32 *p_last, kmp_uint32 *p_lb,
3243  kmp_uint32 *p_ub, kmp_int32 *p_st);
3244 extern int __kmpc_dispatch_next_8(ident_t *loc, kmp_int32 gtid,
3245  kmp_int32 *p_last, kmp_int64 *p_lb,
3246  kmp_int64 *p_ub, kmp_int64 *p_st);
3247 extern int __kmpc_dispatch_next_8u(ident_t *loc, kmp_int32 gtid,
3248  kmp_int32 *p_last, kmp_uint64 *p_lb,
3249  kmp_uint64 *p_ub, kmp_int64 *p_st);
3250 
3251 extern void __kmpc_dispatch_fini_4(ident_t *loc, kmp_int32 gtid);
3252 extern void __kmpc_dispatch_fini_8(ident_t *loc, kmp_int32 gtid);
3253 extern void __kmpc_dispatch_fini_4u(ident_t *loc, kmp_int32 gtid);
3254 extern void __kmpc_dispatch_fini_8u(ident_t *loc, kmp_int32 gtid);
3255 
3256 #ifdef KMP_GOMP_COMPAT
3257 
3258 extern void __kmp_aux_dispatch_init_4(ident_t *loc, kmp_int32 gtid,
3259  enum sched_type schedule, kmp_int32 lb,
3260  kmp_int32 ub, kmp_int32 st,
3261  kmp_int32 chunk, int push_ws);
3262 extern void __kmp_aux_dispatch_init_4u(ident_t *loc, kmp_int32 gtid,
3263  enum sched_type schedule, kmp_uint32 lb,
3264  kmp_uint32 ub, kmp_int32 st,
3265  kmp_int32 chunk, int push_ws);
3266 extern void __kmp_aux_dispatch_init_8(ident_t *loc, kmp_int32 gtid,
3267  enum sched_type schedule, kmp_int64 lb,
3268  kmp_int64 ub, kmp_int64 st,
3269  kmp_int64 chunk, int push_ws);
3270 extern void __kmp_aux_dispatch_init_8u(ident_t *loc, kmp_int32 gtid,
3271  enum sched_type schedule, kmp_uint64 lb,
3272  kmp_uint64 ub, kmp_int64 st,
3273  kmp_int64 chunk, int push_ws);
3274 extern void __kmp_aux_dispatch_fini_chunk_4(ident_t *loc, kmp_int32 gtid);
3275 extern void __kmp_aux_dispatch_fini_chunk_8(ident_t *loc, kmp_int32 gtid);
3276 extern void __kmp_aux_dispatch_fini_chunk_4u(ident_t *loc, kmp_int32 gtid);
3277 extern void __kmp_aux_dispatch_fini_chunk_8u(ident_t *loc, kmp_int32 gtid);
3278 
3279 #endif /* KMP_GOMP_COMPAT */
3280 
3281 extern kmp_uint32 __kmp_eq_4(kmp_uint32 value, kmp_uint32 checker);
3282 extern kmp_uint32 __kmp_neq_4(kmp_uint32 value, kmp_uint32 checker);
3283 extern kmp_uint32 __kmp_lt_4(kmp_uint32 value, kmp_uint32 checker);
3284 extern kmp_uint32 __kmp_ge_4(kmp_uint32 value, kmp_uint32 checker);
3285 extern kmp_uint32 __kmp_le_4(kmp_uint32 value, kmp_uint32 checker);
3286 extern kmp_uint32 __kmp_wait_4(kmp_uint32 volatile *spinner, kmp_uint32 checker,
3287  kmp_uint32 (*pred)(kmp_uint32, kmp_uint32),
3288  void *obj);
3289 extern void __kmp_wait_4_ptr(void *spinner, kmp_uint32 checker,
3290  kmp_uint32 (*pred)(void *, kmp_uint32), void *obj);
3291 
3292 class kmp_flag_32;
3293 class kmp_flag_64;
3294 class kmp_flag_oncore;
3295 extern void __kmp_wait_64(kmp_info_t *this_thr, kmp_flag_64 *flag,
3296  int final_spin
3297 #if USE_ITT_BUILD
3298  ,
3299  void *itt_sync_obj
3300 #endif
3301  );
3302 extern void __kmp_release_64(kmp_flag_64 *flag);
3303 
3304 extern void __kmp_infinite_loop(void);
3305 
3306 extern void __kmp_cleanup(void);
3307 
3308 #if KMP_HANDLE_SIGNALS
3309 extern int __kmp_handle_signals;
3310 extern void __kmp_install_signals(int parallel_init);
3311 extern void __kmp_remove_signals(void);
3312 #endif
3313 
3314 extern void __kmp_clear_system_time(void);
3315 extern void __kmp_read_system_time(double *delta);
3316 
3317 extern void __kmp_check_stack_overlap(kmp_info_t *thr);
3318 
3319 extern void __kmp_expand_host_name(char *buffer, size_t size);
3320 extern void __kmp_expand_file_name(char *result, size_t rlen, char *pattern);
3321 
3322 #if KMP_ARCH_X86 || KMP_ARCH_X86_64
3323 extern void
3324 __kmp_initialize_system_tick(void); /* Initialize timer tick value */
3325 #endif
3326 
3327 extern void
3328 __kmp_runtime_initialize(void); /* machine specific initialization */
3329 extern void __kmp_runtime_destroy(void);
3330 
3331 #if KMP_AFFINITY_SUPPORTED
3332 extern char *__kmp_affinity_print_mask(char *buf, int buf_len,
3333  kmp_affin_mask_t *mask);
3334 extern kmp_str_buf_t *__kmp_affinity_str_buf_mask(kmp_str_buf_t *buf,
3335  kmp_affin_mask_t *mask);
3336 extern void __kmp_affinity_initialize(void);
3337 extern void __kmp_affinity_uninitialize(void);
3338 extern void __kmp_affinity_set_init_mask(
3339  int gtid, int isa_root); /* set affinity according to KMP_AFFINITY */
3340 extern void __kmp_affinity_set_place(int gtid);
3341 extern void __kmp_affinity_determine_capable(const char *env_var);
3342 extern int __kmp_aux_set_affinity(void **mask);
3343 extern int __kmp_aux_get_affinity(void **mask);
3344 extern int __kmp_aux_get_affinity_max_proc();
3345 extern int __kmp_aux_set_affinity_mask_proc(int proc, void **mask);
3346 extern int __kmp_aux_unset_affinity_mask_proc(int proc, void **mask);
3347 extern int __kmp_aux_get_affinity_mask_proc(int proc, void **mask);
3348 extern void __kmp_balanced_affinity(kmp_info_t *th, int team_size);
3349 #if KMP_OS_LINUX
3350 extern int kmp_set_thread_affinity_mask_initial(void);
3351 #endif
3352 #endif /* KMP_AFFINITY_SUPPORTED */
3353 // No need for KMP_AFFINITY_SUPPORTED guard as only one field in the
3354 // format string is for affinity, so platforms that do not support
3355 // affinity can still use the other fields, e.g., %n for num_threads
3356 extern size_t __kmp_aux_capture_affinity(int gtid, const char *format,
3357  kmp_str_buf_t *buffer);
3358 extern void __kmp_aux_display_affinity(int gtid, const char *format);
3359 
3360 extern void __kmp_cleanup_hierarchy();
3361 extern void __kmp_get_hierarchy(kmp_uint32 nproc, kmp_bstate_t *thr_bar);
3362 
3363 #if KMP_USE_FUTEX
3364 
3365 extern int __kmp_futex_determine_capable(void);
3366 
3367 #endif // KMP_USE_FUTEX
3368 
3369 extern void __kmp_gtid_set_specific(int gtid);
3370 extern int __kmp_gtid_get_specific(void);
3371 
3372 extern double __kmp_read_cpu_time(void);
3373 
3374 extern int __kmp_read_system_info(struct kmp_sys_info *info);
3375 
3376 #if KMP_USE_MONITOR
3377 extern void __kmp_create_monitor(kmp_info_t *th);
3378 #endif
3379 
3380 extern void *__kmp_launch_thread(kmp_info_t *thr);
3381 
3382 extern void __kmp_create_worker(int gtid, kmp_info_t *th, size_t stack_size);
3383 
3384 #if KMP_OS_WINDOWS
3385 extern int __kmp_still_running(kmp_info_t *th);
3386 extern int __kmp_is_thread_alive(kmp_info_t *th, DWORD *exit_val);
3387 extern void __kmp_free_handle(kmp_thread_t tHandle);
3388 #endif
3389 
3390 #if KMP_USE_MONITOR
3391 extern void __kmp_reap_monitor(kmp_info_t *th);
3392 #endif
3393 extern void __kmp_reap_worker(kmp_info_t *th);
3394 extern void __kmp_terminate_thread(int gtid);
3395 
3396 extern int __kmp_try_suspend_mx(kmp_info_t *th);
3397 extern void __kmp_lock_suspend_mx(kmp_info_t *th);
3398 extern void __kmp_unlock_suspend_mx(kmp_info_t *th);
3399 
3400 extern void __kmp_suspend_32(int th_gtid, kmp_flag_32 *flag);
3401 extern void __kmp_suspend_64(int th_gtid, kmp_flag_64 *flag);
3402 extern void __kmp_suspend_oncore(int th_gtid, kmp_flag_oncore *flag);
3403 extern void __kmp_resume_32(int target_gtid, kmp_flag_32 *flag);
3404 extern void __kmp_resume_64(int target_gtid, kmp_flag_64 *flag);
3405 extern void __kmp_resume_oncore(int target_gtid, kmp_flag_oncore *flag);
3406 
3407 extern void __kmp_elapsed(double *);
3408 extern void __kmp_elapsed_tick(double *);
3409 
3410 extern void __kmp_enable(int old_state);
3411 extern void __kmp_disable(int *old_state);
3412 
3413 extern void __kmp_thread_sleep(int millis);
3414 
3415 extern void __kmp_common_initialize(void);
3416 extern void __kmp_common_destroy(void);
3417 extern void __kmp_common_destroy_gtid(int gtid);
3418 
3419 #if KMP_OS_UNIX
3420 extern void __kmp_register_atfork(void);
3421 #endif
3422 extern void __kmp_suspend_initialize(void);
3423 extern void __kmp_suspend_initialize_thread(kmp_info_t *th);
3424 extern void __kmp_suspend_uninitialize_thread(kmp_info_t *th);
3425 
3426 extern kmp_info_t *__kmp_allocate_thread(kmp_root_t *root, kmp_team_t *team,
3427  int tid);
3428 extern kmp_team_t *
3429 __kmp_allocate_team(kmp_root_t *root, int new_nproc, int max_nproc,
3430 #if OMPT_SUPPORT
3431  ompt_data_t ompt_parallel_data,
3432 #endif
3433  kmp_proc_bind_t proc_bind, kmp_internal_control_t *new_icvs,
3434  int argc USE_NESTED_HOT_ARG(kmp_info_t *thr));
3435 extern void __kmp_free_thread(kmp_info_t *);
3436 extern void __kmp_free_team(kmp_root_t *,
3437  kmp_team_t *USE_NESTED_HOT_ARG(kmp_info_t *));
3438 extern kmp_team_t *__kmp_reap_team(kmp_team_t *);
3439 
3440 /* ------------------------------------------------------------------------ */
3441 
3442 extern void __kmp_initialize_bget(kmp_info_t *th);
3443 extern void __kmp_finalize_bget(kmp_info_t *th);
3444 
3445 KMP_EXPORT void *kmpc_malloc(size_t size);
3446 KMP_EXPORT void *kmpc_aligned_malloc(size_t size, size_t alignment);
3447 KMP_EXPORT void *kmpc_calloc(size_t nelem, size_t elsize);
3448 KMP_EXPORT void *kmpc_realloc(void *ptr, size_t size);
3449 KMP_EXPORT void kmpc_free(void *ptr);
3450 
3451 /* declarations for internal use */
3452 
3453 extern int __kmp_barrier(enum barrier_type bt, int gtid, int is_split,
3454  size_t reduce_size, void *reduce_data,
3455  void (*reduce)(void *, void *));
3456 extern void __kmp_end_split_barrier(enum barrier_type bt, int gtid);
3457 extern int __kmp_barrier_gomp_cancel(int gtid);
3458 
3463 enum fork_context_e {
3464  fork_context_gnu,
3466  fork_context_intel,
3467  fork_context_last
3468 };
3469 extern int __kmp_fork_call(ident_t *loc, int gtid,
3470  enum fork_context_e fork_context, kmp_int32 argc,
3471  microtask_t microtask, launch_t invoker,
3472 /* TODO: revert workaround for Intel(R) 64 tracker #96 */
3473 #if (KMP_ARCH_ARM || KMP_ARCH_X86_64 || KMP_ARCH_AARCH64) && KMP_OS_LINUX
3474  va_list *ap
3475 #else
3476  va_list ap
3477 #endif
3478  );
3479 
3480 extern void __kmp_join_call(ident_t *loc, int gtid
3481 #if OMPT_SUPPORT
3482  ,
3483  enum fork_context_e fork_context
3484 #endif
3485  ,
3486  int exit_teams = 0);
3487 
3488 extern void __kmp_serialized_parallel(ident_t *id, kmp_int32 gtid);
3489 extern void __kmp_internal_fork(ident_t *id, int gtid, kmp_team_t *team);
3490 extern void __kmp_internal_join(ident_t *id, int gtid, kmp_team_t *team);
3491 extern int __kmp_invoke_task_func(int gtid);
3492 extern void __kmp_run_before_invoked_task(int gtid, int tid,
3493  kmp_info_t *this_thr,
3494  kmp_team_t *team);
3495 extern void __kmp_run_after_invoked_task(int gtid, int tid,
3496  kmp_info_t *this_thr,
3497  kmp_team_t *team);
3498 
3499 // should never have been exported
3500 KMP_EXPORT int __kmpc_invoke_task_func(int gtid);
3501 extern int __kmp_invoke_teams_master(int gtid);
3502 extern void __kmp_teams_master(int gtid);
3503 extern int __kmp_aux_get_team_num();
3504 extern int __kmp_aux_get_num_teams();
3505 extern void __kmp_save_internal_controls(kmp_info_t *thread);
3506 extern void __kmp_user_set_library(enum library_type arg);
3507 extern void __kmp_aux_set_library(enum library_type arg);
3508 extern void __kmp_aux_set_stacksize(size_t arg);
3509 extern void __kmp_aux_set_blocktime(int arg, kmp_info_t *thread, int tid);
3510 extern void __kmp_aux_set_defaults(char const *str, int len);
3511 
3512 /* Functions called from __kmp_aux_env_initialize() in kmp_settings.cpp */
3513 void kmpc_set_blocktime(int arg);
3514 void ompc_set_nested(int flag);
3515 void ompc_set_dynamic(int flag);
3516 void ompc_set_num_threads(int arg);
3517 
3518 extern void __kmp_push_current_task_to_thread(kmp_info_t *this_thr,
3519  kmp_team_t *team, int tid);
3520 extern void __kmp_pop_current_task_from_thread(kmp_info_t *this_thr);
3521 extern kmp_task_t *__kmp_task_alloc(ident_t *loc_ref, kmp_int32 gtid,
3522  kmp_tasking_flags_t *flags,
3523  size_t sizeof_kmp_task_t,
3524  size_t sizeof_shareds,
3525  kmp_routine_entry_t task_entry);
3526 extern void __kmp_init_implicit_task(ident_t *loc_ref, kmp_info_t *this_thr,
3527  kmp_team_t *team, int tid,
3528  int set_curr_task);
3529 extern void __kmp_finish_implicit_task(kmp_info_t *this_thr);
3530 extern void __kmp_free_implicit_task(kmp_info_t *this_thr);
3531 
3532 extern kmp_event_t *__kmpc_task_allow_completion_event(ident_t *loc_ref,
3533  int gtid,
3534  kmp_task_t *task);
3535 extern void __kmp_fulfill_event(kmp_event_t *event);
3536 
3537 int __kmp_execute_tasks_32(kmp_info_t *thread, kmp_int32 gtid,
3538  kmp_flag_32 *flag, int final_spin,
3539  int *thread_finished,
3540 #if USE_ITT_BUILD
3541  void *itt_sync_obj,
3542 #endif /* USE_ITT_BUILD */
3543  kmp_int32 is_constrained);
3544 int __kmp_execute_tasks_64(kmp_info_t *thread, kmp_int32 gtid,
3545  kmp_flag_64 *flag, int final_spin,
3546  int *thread_finished,
3547 #if USE_ITT_BUILD
3548  void *itt_sync_obj,
3549 #endif /* USE_ITT_BUILD */
3550  kmp_int32 is_constrained);
3551 int __kmp_execute_tasks_oncore(kmp_info_t *thread, kmp_int32 gtid,
3552  kmp_flag_oncore *flag, int final_spin,
3553  int *thread_finished,
3554 #if USE_ITT_BUILD
3555  void *itt_sync_obj,
3556 #endif /* USE_ITT_BUILD */
3557  kmp_int32 is_constrained);
3558 
3559 extern void __kmp_free_task_team(kmp_info_t *thread,
3560  kmp_task_team_t *task_team);
3561 extern void __kmp_reap_task_teams(void);
3562 extern void __kmp_wait_to_unref_task_teams(void);
3563 extern void __kmp_task_team_setup(kmp_info_t *this_thr, kmp_team_t *team,
3564  int always);
3565 extern void __kmp_task_team_sync(kmp_info_t *this_thr, kmp_team_t *team);
3566 extern void __kmp_task_team_wait(kmp_info_t *this_thr, kmp_team_t *team
3567 #if USE_ITT_BUILD
3568  ,
3569  void *itt_sync_obj
3570 #endif /* USE_ITT_BUILD */
3571  ,
3572  int wait = 1);
3573 extern void __kmp_tasking_barrier(kmp_team_t *team, kmp_info_t *thread,
3574  int gtid);
3575 
3576 extern int __kmp_is_address_mapped(void *addr);
3577 extern kmp_uint64 __kmp_hardware_timestamp(void);
3578 
3579 #if KMP_OS_UNIX
3580 extern int __kmp_read_from_file(char const *path, char const *format, ...);
3581 #endif
3582 
3583 /* ------------------------------------------------------------------------ */
3584 //
3585 // Assembly routines that have no compiler intrinsic replacement
3586 //
3587 
3588 extern int __kmp_invoke_microtask(microtask_t pkfn, int gtid, int npr, int argc,
3589  void *argv[]
3590 #if OMPT_SUPPORT
3591  ,
3592  void **exit_frame_ptr
3593 #endif
3594  );
3595 
3596 /* ------------------------------------------------------------------------ */
3597 
3598 KMP_EXPORT void __kmpc_begin(ident_t *, kmp_int32 flags);
3599 KMP_EXPORT void __kmpc_end(ident_t *);
3600 
3601 KMP_EXPORT void __kmpc_threadprivate_register_vec(ident_t *, void *data,
3602  kmpc_ctor_vec ctor,
3603  kmpc_cctor_vec cctor,
3604  kmpc_dtor_vec dtor,
3605  size_t vector_length);
3606 KMP_EXPORT void __kmpc_threadprivate_register(ident_t *, void *data,
3607  kmpc_ctor ctor, kmpc_cctor cctor,
3608  kmpc_dtor dtor);
3609 KMP_EXPORT void *__kmpc_threadprivate(ident_t *, kmp_int32 global_tid,
3610  void *data, size_t size);
3611 
3612 KMP_EXPORT kmp_int32 __kmpc_global_thread_num(ident_t *);
3613 KMP_EXPORT kmp_int32 __kmpc_global_num_threads(ident_t *);
3614 KMP_EXPORT kmp_int32 __kmpc_bound_thread_num(ident_t *);
3615 KMP_EXPORT kmp_int32 __kmpc_bound_num_threads(ident_t *);
3616 
3617 KMP_EXPORT kmp_int32 __kmpc_ok_to_fork(ident_t *);
3618 KMP_EXPORT void __kmpc_fork_call(ident_t *, kmp_int32 nargs,
3619  kmpc_micro microtask, ...);
3620 
3621 KMP_EXPORT void __kmpc_serialized_parallel(ident_t *, kmp_int32 global_tid);
3622 KMP_EXPORT void __kmpc_end_serialized_parallel(ident_t *, kmp_int32 global_tid);
3623 
3624 KMP_EXPORT void __kmpc_flush(ident_t *);
3625 KMP_EXPORT void __kmpc_barrier(ident_t *, kmp_int32 global_tid);
3626 KMP_EXPORT kmp_int32 __kmpc_master(ident_t *, kmp_int32 global_tid);
3627 KMP_EXPORT void __kmpc_end_master(ident_t *, kmp_int32 global_tid);
3628 KMP_EXPORT void __kmpc_ordered(ident_t *, kmp_int32 global_tid);
3629 KMP_EXPORT void __kmpc_end_ordered(ident_t *, kmp_int32 global_tid);
3630 KMP_EXPORT void __kmpc_critical(ident_t *, kmp_int32 global_tid,
3631  kmp_critical_name *);
3632 KMP_EXPORT void __kmpc_end_critical(ident_t *, kmp_int32 global_tid,
3633  kmp_critical_name *);
3634 KMP_EXPORT void __kmpc_critical_with_hint(ident_t *, kmp_int32 global_tid,
3635  kmp_critical_name *, uint32_t hint);
3636 
3637 KMP_EXPORT kmp_int32 __kmpc_barrier_master(ident_t *, kmp_int32 global_tid);
3638 KMP_EXPORT void __kmpc_end_barrier_master(ident_t *, kmp_int32 global_tid);
3639 
3640 KMP_EXPORT kmp_int32 __kmpc_barrier_master_nowait(ident_t *,
3641  kmp_int32 global_tid);
3642 
3643 KMP_EXPORT kmp_int32 __kmpc_single(ident_t *, kmp_int32 global_tid);
3644 KMP_EXPORT void __kmpc_end_single(ident_t *, kmp_int32 global_tid);
3645 
3646 KMP_EXPORT void KMPC_FOR_STATIC_INIT(ident_t *loc, kmp_int32 global_tid,
3647  kmp_int32 schedtype, kmp_int32 *plastiter,
3648  kmp_int *plower, kmp_int *pupper,
3649  kmp_int *pstride, kmp_int incr,
3650  kmp_int chunk);
3651 
3652 KMP_EXPORT void __kmpc_for_static_fini(ident_t *loc, kmp_int32 global_tid);
3653 
3654 KMP_EXPORT void __kmpc_copyprivate(ident_t *loc, kmp_int32 global_tid,
3655  size_t cpy_size, void *cpy_data,
3656  void (*cpy_func)(void *, void *),
3657  kmp_int32 didit);
3658 
3659 extern void KMPC_SET_NUM_THREADS(int arg);
3660 extern void KMPC_SET_DYNAMIC(int flag);
3661 extern void KMPC_SET_NESTED(int flag);
3662 
3663 /* OMP 3.0 tasking interface routines */
3664 KMP_EXPORT kmp_int32 __kmpc_omp_task(ident_t *loc_ref, kmp_int32 gtid,
3665  kmp_task_t *new_task);
3666 KMP_EXPORT kmp_task_t *__kmpc_omp_task_alloc(ident_t *loc_ref, kmp_int32 gtid,
3667  kmp_int32 flags,
3668  size_t sizeof_kmp_task_t,
3669  size_t sizeof_shareds,
3670  kmp_routine_entry_t task_entry);
3671 KMP_EXPORT kmp_task_t *__kmpc_omp_target_task_alloc(ident_t *loc_ref, kmp_int32 gtid,
3672  kmp_int32 flags,
3673  size_t sizeof_kmp_task_t,
3674  size_t sizeof_shareds,
3675  kmp_routine_entry_t task_entry,
3676  kmp_int64 device_id);
3677 KMP_EXPORT void __kmpc_omp_task_begin_if0(ident_t *loc_ref, kmp_int32 gtid,
3678  kmp_task_t *task);
3679 KMP_EXPORT void __kmpc_omp_task_complete_if0(ident_t *loc_ref, kmp_int32 gtid,
3680  kmp_task_t *task);
3681 KMP_EXPORT kmp_int32 __kmpc_omp_task_parts(ident_t *loc_ref, kmp_int32 gtid,
3682  kmp_task_t *new_task);
3683 KMP_EXPORT kmp_int32 __kmpc_omp_taskwait(ident_t *loc_ref, kmp_int32 gtid);
3684 
3685 KMP_EXPORT kmp_int32 __kmpc_omp_taskyield(ident_t *loc_ref, kmp_int32 gtid,
3686  int end_part);
3687 
3688 #if TASK_UNUSED
3689 void __kmpc_omp_task_begin(ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *task);
3690 void __kmpc_omp_task_complete(ident_t *loc_ref, kmp_int32 gtid,
3691  kmp_task_t *task);
3692 #endif // TASK_UNUSED
3693 
3694 /* ------------------------------------------------------------------------ */
3695 
3696 KMP_EXPORT void __kmpc_taskgroup(ident_t *loc, int gtid);
3697 KMP_EXPORT void __kmpc_end_taskgroup(ident_t *loc, int gtid);
3698 
3699 KMP_EXPORT kmp_int32 __kmpc_omp_task_with_deps(
3700  ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 ndeps,
3701  kmp_depend_info_t *dep_list, kmp_int32 ndeps_noalias,
3702  kmp_depend_info_t *noalias_dep_list);
3703 KMP_EXPORT void __kmpc_omp_wait_deps(ident_t *loc_ref, kmp_int32 gtid,
3704  kmp_int32 ndeps,
3705  kmp_depend_info_t *dep_list,
3706  kmp_int32 ndeps_noalias,
3707  kmp_depend_info_t *noalias_dep_list);
3708 extern kmp_int32 __kmp_omp_task(kmp_int32 gtid, kmp_task_t *new_task,
3709  bool serialize_immediate);
3710 
3711 KMP_EXPORT kmp_int32 __kmpc_cancel(ident_t *loc_ref, kmp_int32 gtid,
3712  kmp_int32 cncl_kind);
3713 KMP_EXPORT kmp_int32 __kmpc_cancellationpoint(ident_t *loc_ref, kmp_int32 gtid,
3714  kmp_int32 cncl_kind);
3715 KMP_EXPORT kmp_int32 __kmpc_cancel_barrier(ident_t *loc_ref, kmp_int32 gtid);
3716 KMP_EXPORT int __kmp_get_cancellation_status(int cancel_kind);
3717 
3718 KMP_EXPORT void __kmpc_proxy_task_completed(kmp_int32 gtid, kmp_task_t *ptask);
3719 KMP_EXPORT void __kmpc_proxy_task_completed_ooo(kmp_task_t *ptask);
3720 KMP_EXPORT void __kmpc_taskloop(ident_t *loc, kmp_int32 gtid, kmp_task_t *task,
3721  kmp_int32 if_val, kmp_uint64 *lb,
3722  kmp_uint64 *ub, kmp_int64 st, kmp_int32 nogroup,
3723  kmp_int32 sched, kmp_uint64 grainsize,
3724  void *task_dup);
3725 KMP_EXPORT void *__kmpc_task_reduction_init(int gtid, int num_data, void *data);
3726 KMP_EXPORT void *__kmpc_taskred_init(int gtid, int num_data, void *data);
3727 KMP_EXPORT void *__kmpc_task_reduction_get_th_data(int gtid, void *tg, void *d);
3728 KMP_EXPORT void *__kmpc_task_reduction_modifier_init(ident_t *loc, int gtid,
3729  int is_ws, int num,
3730  void *data);
3731 KMP_EXPORT void *__kmpc_taskred_modifier_init(ident_t *loc, int gtid, int is_ws,
3732  int num, void *data);
3733 KMP_EXPORT void __kmpc_task_reduction_modifier_fini(ident_t *loc, int gtid,
3734  int is_ws);
3735 KMP_EXPORT kmp_int32 __kmpc_omp_reg_task_with_affinity(
3736  ident_t *loc_ref, kmp_int32 gtid, kmp_task_t *new_task, kmp_int32 naffins,
3737  kmp_task_affinity_info_t *affin_list);
3738 
3739 /* Lock interface routines (fast versions with gtid passed in) */
3740 KMP_EXPORT void __kmpc_init_lock(ident_t *loc, kmp_int32 gtid,
3741  void **user_lock);
3742 KMP_EXPORT void __kmpc_init_nest_lock(ident_t *loc, kmp_int32 gtid,
3743  void **user_lock);
3744 KMP_EXPORT void __kmpc_destroy_lock(ident_t *loc, kmp_int32 gtid,
3745  void **user_lock);
3746 KMP_EXPORT void __kmpc_destroy_nest_lock(ident_t *loc, kmp_int32 gtid,
3747  void **user_lock);
3748 KMP_EXPORT void __kmpc_set_lock(ident_t *loc, kmp_int32 gtid, void **user_lock);
3749 KMP_EXPORT void __kmpc_set_nest_lock(ident_t *loc, kmp_int32 gtid,
3750  void **user_lock);
3751 KMP_EXPORT void __kmpc_unset_lock(ident_t *loc, kmp_int32 gtid,
3752  void **user_lock);
3753 KMP_EXPORT void __kmpc_unset_nest_lock(ident_t *loc, kmp_int32 gtid,
3754  void **user_lock);
3755 KMP_EXPORT int __kmpc_test_lock(ident_t *loc, kmp_int32 gtid, void **user_lock);
3756 KMP_EXPORT int __kmpc_test_nest_lock(ident_t *loc, kmp_int32 gtid,
3757  void **user_lock);
3758 
3759 KMP_EXPORT void __kmpc_init_lock_with_hint(ident_t *loc, kmp_int32 gtid,
3760  void **user_lock, uintptr_t hint);
3761 KMP_EXPORT void __kmpc_init_nest_lock_with_hint(ident_t *loc, kmp_int32 gtid,
3762  void **user_lock,
3763  uintptr_t hint);
3764 
3765 /* Interface to fast scalable reduce methods routines */
3766 
3767 KMP_EXPORT kmp_int32 __kmpc_reduce_nowait(
3768  ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
3769  void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
3770  kmp_critical_name *lck);
3771 KMP_EXPORT void __kmpc_end_reduce_nowait(ident_t *loc, kmp_int32 global_tid,
3772  kmp_critical_name *lck);
3773 KMP_EXPORT kmp_int32 __kmpc_reduce(
3774  ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
3775  void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
3776  kmp_critical_name *lck);
3777 KMP_EXPORT void __kmpc_end_reduce(ident_t *loc, kmp_int32 global_tid,
3778  kmp_critical_name *lck);
3779 
3780 /* Internal fast reduction routines */
3781 
3782 extern PACKED_REDUCTION_METHOD_T __kmp_determine_reduction_method(
3783  ident_t *loc, kmp_int32 global_tid, kmp_int32 num_vars, size_t reduce_size,
3784  void *reduce_data, void (*reduce_func)(void *lhs_data, void *rhs_data),
3785  kmp_critical_name *lck);
3786 
3787 // this function is for testing set/get/determine reduce method
3788 KMP_EXPORT kmp_int32 __kmp_get_reduce_method(void);
3789 
3790 KMP_EXPORT kmp_uint64 __kmpc_get_taskid();
3791 KMP_EXPORT kmp_uint64 __kmpc_get_parent_taskid();
3792 
3793 // C++ port
3794 // missing 'extern "C"' declarations
3795 
3796 KMP_EXPORT kmp_int32 __kmpc_in_parallel(ident_t *loc);
3797 KMP_EXPORT void __kmpc_pop_num_threads(ident_t *loc, kmp_int32 global_tid);
3798 KMP_EXPORT void __kmpc_push_num_threads(ident_t *loc, kmp_int32 global_tid,
3799  kmp_int32 num_threads);
3800 
3801 KMP_EXPORT void __kmpc_push_proc_bind(ident_t *loc, kmp_int32 global_tid,
3802  int proc_bind);
3803 KMP_EXPORT void __kmpc_push_num_teams(ident_t *loc, kmp_int32 global_tid,
3804  kmp_int32 num_teams,
3805  kmp_int32 num_threads);
3806 KMP_EXPORT void __kmpc_fork_teams(ident_t *loc, kmp_int32 argc,
3807  kmpc_micro microtask, ...);
3808 struct kmp_dim { // loop bounds info casted to kmp_int64
3809  kmp_int64 lo; // lower
3810  kmp_int64 up; // upper
3811  kmp_int64 st; // stride
3812 };
3813 KMP_EXPORT void __kmpc_doacross_init(ident_t *loc, kmp_int32 gtid,
3814  kmp_int32 num_dims,
3815  const struct kmp_dim *dims);
3816 KMP_EXPORT void __kmpc_doacross_wait(ident_t *loc, kmp_int32 gtid,
3817  const kmp_int64 *vec);
3818 KMP_EXPORT void __kmpc_doacross_post(ident_t *loc, kmp_int32 gtid,
3819  const kmp_int64 *vec);
3820 KMP_EXPORT void __kmpc_doacross_fini(ident_t *loc, kmp_int32 gtid);
3821 
3822 KMP_EXPORT void *__kmpc_threadprivate_cached(ident_t *loc, kmp_int32 global_tid,
3823  void *data, size_t size,
3824  void ***cache);
3825 
3826 // Symbols for MS mutual detection.
3827 extern int _You_must_link_with_exactly_one_OpenMP_library;
3828 extern int _You_must_link_with_Intel_OpenMP_library;
3829 #if KMP_OS_WINDOWS && (KMP_VERSION_MAJOR > 4)
3830 extern int _You_must_link_with_Microsoft_OpenMP_library;
3831 #endif
3832 
3833 // The routines below are not exported.
3834 // Consider making them 'static' in corresponding source files.
3835 void kmp_threadprivate_insert_private_data(int gtid, void *pc_addr,
3836  void *data_addr, size_t pc_size);
3837 struct private_common *kmp_threadprivate_insert(int gtid, void *pc_addr,
3838  void *data_addr,
3839  size_t pc_size);
3840 void __kmp_threadprivate_resize_cache(int newCapacity);
3841 void __kmp_cleanup_threadprivate_caches();
3842 
3843 // ompc_, kmpc_ entries moved from omp.h.
3844 #if KMP_OS_WINDOWS
3845 #define KMPC_CONVENTION __cdecl
3846 #else
3847 #define KMPC_CONVENTION
3848 #endif
3849 
3850 #ifndef __OMP_H
3851 typedef enum omp_sched_t {
3852  omp_sched_static = 1,
3853  omp_sched_dynamic = 2,
3854  omp_sched_guided = 3,
3855  omp_sched_auto = 4
3856 } omp_sched_t;
3857 typedef void *kmp_affinity_mask_t;
3858 #endif
3859 
3860 KMP_EXPORT void KMPC_CONVENTION ompc_set_max_active_levels(int);
3861 KMP_EXPORT void KMPC_CONVENTION ompc_set_schedule(omp_sched_t, int);
3862 KMP_EXPORT int KMPC_CONVENTION ompc_get_ancestor_thread_num(int);
3863 KMP_EXPORT int KMPC_CONVENTION ompc_get_team_size(int);
3864 KMP_EXPORT int KMPC_CONVENTION
3865 kmpc_set_affinity_mask_proc(int, kmp_affinity_mask_t *);
3866 KMP_EXPORT int KMPC_CONVENTION
3867 kmpc_unset_affinity_mask_proc(int, kmp_affinity_mask_t *);
3868 KMP_EXPORT int KMPC_CONVENTION
3869 kmpc_get_affinity_mask_proc(int, kmp_affinity_mask_t *);
3870 
3871 KMP_EXPORT void KMPC_CONVENTION kmpc_set_stacksize(int);
3872 KMP_EXPORT void KMPC_CONVENTION kmpc_set_stacksize_s(size_t);
3873 KMP_EXPORT void KMPC_CONVENTION kmpc_set_library(int);
3874 KMP_EXPORT void KMPC_CONVENTION kmpc_set_defaults(char const *);
3875 KMP_EXPORT void KMPC_CONVENTION kmpc_set_disp_num_buffers(int);
3876 
3877 enum kmp_target_offload_kind {
3878  tgt_disabled = 0,
3879  tgt_default = 1,
3880  tgt_mandatory = 2
3881 };
3882 typedef enum kmp_target_offload_kind kmp_target_offload_kind_t;
3883 // Set via OMP_TARGET_OFFLOAD if specified, defaults to tgt_default otherwise
3884 extern kmp_target_offload_kind_t __kmp_target_offload;
3885 extern int __kmpc_get_target_offload();
3886 
3887 // Constants used in libomptarget
3888 #define KMP_DEVICE_DEFAULT -1 // This is libomptarget's default device.
3889 #define KMP_HOST_DEVICE -10 // This is what it is in libomptarget, go figure.
3890 #define KMP_DEVICE_ALL -11 // This is libomptarget's "all devices".
3891 
3892 // OMP Pause Resource
3893 
3894 // The following enum is used both to set the status in __kmp_pause_status, and
3895 // as the internal equivalent of the externally-visible omp_pause_resource_t.
3896 typedef enum kmp_pause_status_t {
3897  kmp_not_paused = 0, // status is not paused, or, requesting resume
3898  kmp_soft_paused = 1, // status is soft-paused, or, requesting soft pause
3899  kmp_hard_paused = 2 // status is hard-paused, or, requesting hard pause
3900 } kmp_pause_status_t;
3901 
3902 // This stores the pause state of the runtime
3903 extern kmp_pause_status_t __kmp_pause_status;
3904 extern int __kmpc_pause_resource(kmp_pause_status_t level);
3905 extern int __kmp_pause_resource(kmp_pause_status_t level);
3906 // Soft resume sets __kmp_pause_status, and wakes up all threads.
3907 extern void __kmp_resume_if_soft_paused();
3908 // Hard resume simply resets the status to not paused. Library will appear to
3909 // be uninitialized after hard pause. Let OMP constructs trigger required
3910 // initializations.
3911 static inline void __kmp_resume_if_hard_paused() {
3912  if (__kmp_pause_status == kmp_hard_paused) {
3913  __kmp_pause_status = kmp_not_paused;
3914  }
3915 }
3916 
3917 #ifdef __cplusplus
3918 }
3919 #endif
3920 
3921 #endif /* KMP_H */
KMP_EXPORT void __kmpc_task_reduction_modifier_fini(ident_t *loc, int gtid, int is_ws)
KMP_EXPORT void * __kmpc_task_reduction_modifier_init(ident_t *loc, int gtid, int is_ws, int num, void *data)
KMP_EXPORT kmp_int32 __kmpc_master(ident_t *, kmp_int32 global_tid)
KMP_EXPORT kmp_int32 __kmpc_barrier_master(ident_t *, kmp_int32 global_tid)
kmp_int32 reserved_2
Definition: kmp.h:226
void __kmpc_dispatch_fini_4(ident_t *loc, kmp_int32 gtid)
KMP_EXPORT void __kmpc_end_single(ident_t *, kmp_int32 global_tid)
void(* kmpc_dtor)(void *)
Definition: kmp.h:1423
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)
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 kmp_int32 __kmpc_global_thread_num(ident_t *)
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)
void(* kmpc_dtor_vec)(void *, size_t)
Definition: kmp.h:1446
KMP_EXPORT void __kmpc_for_static_fini(ident_t *loc, kmp_int32 global_tid)
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_int32 reserved_1
Definition: kmp.h:223
void *(* kmpc_ctor_vec)(void *, size_t)
Definition: kmp.h:1440
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_threadprivate_cached(ident_t *loc, kmp_int32 global_tid, void *data, size_t size, void ***cache)
kmp_int32 reserved_3
Definition: kmp.h:231
void *(* kmpc_cctor_vec)(void *, void *, size_t)
Definition: kmp.h:1452
KMP_EXPORT void __kmpc_flush(ident_t *)
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)
KMP_EXPORT kmp_int32 __kmpc_single(ident_t *, kmp_int32 global_tid)
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)
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)
KMP_EXPORT void __kmpc_end(ident_t *)
KMP_EXPORT void __kmpc_end_ordered(ident_t *, kmp_int32 global_tid)
KMP_EXPORT void __kmpc_end_serialized_parallel(ident_t *, kmp_int32 global_tid)
void *(* kmpc_cctor)(void *, void *)
Definition: kmp.h:1430
void __kmpc_doacross_init(ident_t *loc, int gtid, int num_dims, const struct kmp_dim *dims)
KMP_EXPORT void __kmpc_threadprivate_register(ident_t *, void *data, kmpc_ctor ctor, kmpc_cctor cctor, kmpc_dtor dtor)
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 void __kmpc_begin(ident_t *, kmp_int32 flags)
KMP_EXPORT void * __kmpc_taskred_modifier_init(ident_t *loc, int gtid, int is_ws, int num, void *data)
KMP_EXPORT kmp_int32 __kmpc_bound_thread_num(ident_t *)
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)
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)
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)
KMP_EXPORT void __kmpc_proxy_task_completed_ooo(kmp_task_t *ptask)
KMP_EXPORT void __kmpc_ordered(ident_t *, kmp_int32 global_tid)
KMP_EXPORT void * __kmpc_taskred_init(int gtid, int num_data, void *data)
KMP_EXPORT void * __kmpc_task_reduction_get_th_data(int gtid, void *tg, void *d)
sched_type
Definition: kmp.h:336
KMP_EXPORT void __kmpc_critical(ident_t *, kmp_int32 global_tid, kmp_critical_name *)
Definition: kmp.h:222
KMP_EXPORT void __kmpc_end_barrier_master(ident_t *, kmp_int32 global_tid)
KMP_EXPORT void __kmpc_proxy_task_completed(kmp_int32 gtid, kmp_task_t *ptask)
KMP_EXPORT void __kmpc_end_master(ident_t *, kmp_int32 global_tid)
KMP_EXPORT void __kmpc_push_num_threads(ident_t *loc, kmp_int32 global_tid, kmp_int32 num_threads)
KMP_EXPORT void __kmpc_fork_teams(ident_t *loc, kmp_int32 argc, kmpc_micro microtask,...)
KMP_EXPORT kmp_int32 __kmpc_in_parallel(ident_t *loc)
KMP_EXPORT kmp_int32 __kmpc_ok_to_fork(ident_t *)
KMP_EXPORT kmp_int32 __kmpc_global_num_threads(ident_t *)
void __kmpc_dispatch_fini_8u(ident_t *loc, kmp_int32 gtid)
KMP_EXPORT kmp_int32 __kmpc_bound_num_threads(ident_t *)
KMP_EXPORT void __kmpc_end_reduce(ident_t *loc, kmp_int32 global_tid, kmp_critical_name *lck)
void __kmpc_dispatch_fini_4u(ident_t *loc, kmp_int32 gtid)
KMP_EXPORT void __kmpc_barrier(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 void __kmpc_end_critical(ident_t *, kmp_int32 global_tid, kmp_critical_name *)
KMP_EXPORT void * __kmpc_task_reduction_init(int gtid, int num_data, void *data)
void *(* kmpc_ctor)(void *)
Definition: kmp.h:1417
KMP_EXPORT void __kmpc_push_num_teams(ident_t *loc, kmp_int32 global_tid, kmp_int32 num_teams, kmp_int32 num_threads)
void __kmpc_dispatch_fini_8(ident_t *loc, kmp_int32 gtid)
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)
KMP_EXPORT void __kmpc_critical_with_hint(ident_t *, kmp_int32 global_tid, kmp_critical_name *, uint32_t hint)
void(* kmpc_micro)(kmp_int32 *global_tid, kmp_int32 *bound_tid,...)
Definition: kmp.h:1399
KMP_EXPORT kmp_int32 __kmpc_barrier_master_nowait(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)
KMP_EXPORT void __kmpc_serialized_parallel(ident_t *, kmp_int32 global_tid)
KMP_EXPORT void __kmpc_fork_call(ident_t *, kmp_int32 nargs, kmpc_micro microtask,...)
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)
char const * psource
Definition: kmp.h:232
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_int32 flags
Definition: kmp.h:224
struct ident ident_t